xref: /f-stack/freebsd/netinet/tcp_stacks/rack.c (revision 77ab9738) 
1 /*-
2  * Copyright (c) 2016-2020 Netflix, Inc.
3  *
4  * Redistribution and use in source and binary forms, with or without
5  * modification, are permitted provided that the following conditions
6  * are met:
7  * 1. Redistributions of source code must retain the above copyright
8  *    notice, this list of conditions and the following disclaimer.
9  * 2. Redistributions in binary form must reproduce the above copyright
10  *    notice, this list of conditions and the following disclaimer in the
11  *    documentation and/or other materials provided with the distribution.
12  *
13  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
14  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
15  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
16  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
17  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
18  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
19  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
20  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
21  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
22  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
23  * SUCH DAMAGE.
24  *
25  */
26 
27 #include <sys/cdefs.h>
28 __FBSDID("$FreeBSD$");
29 
30 #include "opt_inet.h"
31 #include "opt_inet6.h"
32 #include "opt_ipsec.h"
33 #include "opt_tcpdebug.h"
34 #include "opt_ratelimit.h"
35 #include <sys/param.h>
36 #include <sys/arb.h>
37 #include <sys/module.h>
38 #include <sys/kernel.h>
39 #ifdef TCP_HHOOK
40 #include <sys/hhook.h>
41 #endif
42 #include <sys/lock.h>
43 #include <sys/malloc.h>
44 #include <sys/lock.h>
45 #include <sys/mutex.h>
46 #include <sys/mbuf.h>
47 #include <sys/proc.h>		/* for proc0 declaration */
48 #include <sys/socket.h>
49 #include <sys/socketvar.h>
50 #include <sys/sysctl.h>
51 #include <sys/systm.h>
52 #ifdef STATS
53 #include <sys/qmath.h>
54 #include <sys/tree.h>
55 #include <sys/stats.h> /* Must come after qmath.h and tree.h */
56 #else
57 #include <sys/tree.h>
58 #endif
59 #include <sys/refcount.h>
60 #include <sys/queue.h>
61 #include <sys/tim_filter.h>
62 #include <sys/smp.h>
63 #include <sys/kthread.h>
64 #include <sys/kern_prefetch.h>
65 #include <sys/protosw.h>
66 
67 #include <vm/uma.h>
68 
69 #include <net/route.h>
70 #include <net/route/nhop.h>
71 #include <net/vnet.h>
72 
73 #define TCPSTATES		/* for logging */
74 
75 #include <netinet/in.h>
76 #include <netinet/in_kdtrace.h>
77 #include <netinet/in_pcb.h>
78 #include <netinet/ip.h>
79 #include <netinet/ip_icmp.h>	/* required for icmp_var.h */
80 #include <netinet/icmp_var.h>	/* for ICMP_BANDLIM */
81 #include <netinet/ip_var.h>
82 #include <netinet/ip6.h>
83 #include <netinet6/in6_pcb.h>
84 #include <netinet6/ip6_var.h>
85 #include <netinet/tcp.h>
86 #define	TCPOUTFLAGS
87 #include <netinet/tcp_fsm.h>
88 #include <netinet/tcp_log_buf.h>
89 #include <netinet/tcp_seq.h>
90 #include <netinet/tcp_timer.h>
91 #include <netinet/tcp_var.h>
92 #include <netinet/tcp_hpts.h>
93 #include <netinet/tcp_ratelimit.h>
94 #include <netinet/tcpip.h>
95 #include <netinet/cc/cc.h>
96 #include <netinet/tcp_fastopen.h>
97 #include <netinet/tcp_lro.h>
98 #ifdef NETFLIX_SHARED_CWND
99 #include <netinet/tcp_shared_cwnd.h>
100 #endif
101 #ifdef TCPDEBUG
102 #include <netinet/tcp_debug.h>
103 #endif				/* TCPDEBUG */
104 #ifdef TCP_OFFLOAD
105 #include <netinet/tcp_offload.h>
106 #endif
107 #ifdef INET6
108 #include <netinet6/tcp6_var.h>
109 #endif
110 
111 #include <netipsec/ipsec_support.h>
112 
113 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
114 #include <netipsec/ipsec.h>
115 #include <netipsec/ipsec6.h>
116 #endif				/* IPSEC */
117 
118 #include <netinet/udp.h>
119 #include <netinet/udp_var.h>
120 #include <machine/in_cksum.h>
121 
122 #ifdef MAC
123 #include <security/mac/mac_framework.h>
124 #endif
125 #include "sack_filter.h"
126 #include "tcp_rack.h"
127 #include "rack_bbr_common.h"
128 
129 uma_zone_t rack_zone;
130 uma_zone_t rack_pcb_zone;
131 
132 #ifndef TICKS2SBT
133 #define	TICKS2SBT(__t)	(tick_sbt * ((sbintime_t)(__t)))
134 #endif
135 
136 struct sysctl_ctx_list rack_sysctl_ctx;
137 struct sysctl_oid *rack_sysctl_root;
138 
139 #define CUM_ACKED 1
140 #define SACKED 2
141 
142 #ifdef FSTACK
143 #define MODNAME tcp_rack
144 #define STACKNAME rack
145 #endif
146 
147 /*
148  * The RACK module incorporates a number of
149  * TCP ideas that have been put out into the IETF
150  * over the last few years:
151  * - Matt Mathis's Rate Halving which slowly drops
152  *    the congestion window so that the ack clock can
153  *    be maintained during a recovery.
154  * - Yuchung Cheng's RACK TCP (for which its named) that
155  *    will stop us using the number of dup acks and instead
156  *    use time as the gage of when we retransmit.
157  * - Reorder Detection of RFC4737 and the Tail-Loss probe draft
158  *    of Dukkipati et.al.
159  * RACK depends on SACK, so if an endpoint arrives that
160  * cannot do SACK the state machine below will shuttle the
161  * connection back to using the "default" TCP stack that is
162  * in FreeBSD.
163  *
164  * To implement RACK the original TCP stack was first decomposed
165  * into a functional state machine with individual states
166  * for each of the possible TCP connection states. The do_segement
167  * functions role in life is to mandate the connection supports SACK
168  * initially and then assure that the RACK state matches the conenction
169  * state before calling the states do_segment function. Each
170  * state is simplified due to the fact that the original do_segment
171  * has been decomposed and we *know* what state we are in (no
172  * switches on the state) and all tests for SACK are gone. This
173  * greatly simplifies what each state does.
174  *
175  * TCP output is also over-written with a new version since it
176  * must maintain the new rack scoreboard.
177  *
178  */
179 static int32_t rack_tlp_thresh = 1;
180 static int32_t rack_tlp_limit = 2;	/* No more than 2 TLPs w-out new data */
181 static int32_t rack_tlp_use_greater = 1;
182 static int32_t rack_reorder_thresh = 2;
183 static int32_t rack_reorder_fade = 60000;	/* 0 - never fade, def 60,000
184 						 * - 60 seconds */
185 /* Attack threshold detections */
186 static uint32_t rack_highest_sack_thresh_seen = 0;
187 static uint32_t rack_highest_move_thresh_seen = 0;
188 
189 static int32_t rack_pkt_delay = 1;
190 static int32_t rack_early_recovery = 1;
191 static int32_t rack_send_a_lot_in_prr = 1;
192 static int32_t rack_min_to = 1;	/* Number of ms minimum timeout */
193 static int32_t rack_verbose_logging = 0;
194 static int32_t rack_ignore_data_after_close = 1;
195 static int32_t rack_enable_shared_cwnd = 0;
196 static int32_t rack_limits_scwnd = 1;
197 static int32_t rack_enable_mqueue_for_nonpaced = 0;
198 static int32_t rack_disable_prr = 0;
199 static int32_t use_rack_rr = 1;
200 static int32_t rack_non_rxt_use_cr = 0; /* does a non-rxt in recovery use the configured rate (ss/ca)? */
201 static int32_t rack_persist_min = 250;	/* 250ms */
202 static int32_t rack_persist_max = 2000;	/* 2 Second */
203 static int32_t rack_sack_not_required = 0;	/* set to one to allow non-sack to use rack */
204 static int32_t rack_default_init_window = 0; 	/* Use system default */
205 static int32_t rack_limit_time_with_srtt = 0;
206 static int32_t rack_hw_pace_adjust = 0;
207 /*
208  * Currently regular tcp has a rto_min of 30ms
209  * the backoff goes 12 times so that ends up
210  * being a total of 122.850 seconds before a
211  * connection is killed.
212  */
213 static uint32_t rack_def_data_window = 20;
214 static uint32_t rack_goal_bdp = 2;
215 static uint32_t rack_min_srtts = 1;
216 static uint32_t rack_min_measure_usec = 0;
217 static int32_t rack_tlp_min = 10;
218 static int32_t rack_rto_min = 30;	/* 30ms same as main freebsd */
219 static int32_t rack_rto_max = 4000;	/* 4 seconds */
220 static const int32_t rack_free_cache = 2;
221 static int32_t rack_hptsi_segments = 40;
222 static int32_t rack_rate_sample_method = USE_RTT_LOW;
223 static int32_t rack_pace_every_seg = 0;
224 static int32_t rack_delayed_ack_time = 200;	/* 200ms */
225 static int32_t rack_slot_reduction = 4;
226 static int32_t rack_wma_divisor = 8;		/* For WMA calculation */
227 static int32_t rack_cwnd_block_ends_measure = 0;
228 static int32_t rack_rwnd_block_ends_measure = 0;
229 
230 static int32_t rack_lower_cwnd_at_tlp = 0;
231 static int32_t rack_use_proportional_reduce = 0;
232 static int32_t rack_proportional_rate = 10;
233 static int32_t rack_tlp_max_resend = 2;
234 static int32_t rack_limited_retran = 0;
235 static int32_t rack_always_send_oldest = 0;
236 static int32_t rack_tlp_threshold_use = TLP_USE_TWO_ONE;
237 
238 static uint16_t rack_per_of_gp_ss = 250;	/* 250 % slow-start */
239 static uint16_t rack_per_of_gp_ca = 200;	/* 200 % congestion-avoidance */
240 static uint16_t rack_per_of_gp_rec = 200;	/* 200 % of bw */
241 
242 /* Probertt */
243 static uint16_t rack_per_of_gp_probertt = 60;	/* 60% of bw */
244 static uint16_t rack_per_of_gp_lowthresh = 40;	/* 40% is bottom */
245 static uint16_t rack_per_of_gp_probertt_reduce = 10; /* 10% reduction */
246 static uint16_t rack_atexit_prtt_hbp = 130;	/* Clamp to 130% on exit prtt if highly buffered path */
247 static uint16_t rack_atexit_prtt = 130;	/* Clamp to 100% on exit prtt if non highly buffered path */
248 
249 static uint32_t rack_max_drain_wait = 2;	/* How man gp srtt's before we give up draining */
250 static uint32_t rack_must_drain = 1;		/* How many GP srtt's we *must* wait */
251 static uint32_t rack_probertt_use_min_rtt_entry = 1;	/* Use the min to calculate the goal else gp_srtt */
252 static uint32_t rack_probertt_use_min_rtt_exit = 0;
253 static uint32_t rack_probe_rtt_sets_cwnd = 0;
254 static uint32_t rack_probe_rtt_safety_val = 2000000;	/* No more than 2 sec in probe-rtt */
255 static uint32_t rack_time_between_probertt = 9600000;	/* 9.6 sec in us */
256 static uint32_t rack_probertt_gpsrtt_cnt_mul = 0;	/* How many srtt periods does probe-rtt last top fraction */
257 static uint32_t rack_probertt_gpsrtt_cnt_div = 0;	/* How many srtt periods does probe-rtt last bottom fraction  */
258 static uint32_t rack_min_probertt_hold = 200000;	/* Equal to delayed ack time */
259 static uint32_t rack_probertt_filter_life = 10000000;
260 static uint32_t rack_probertt_lower_within = 10;
261 static uint32_t rack_min_rtt_movement = 250;	/* Must move at least 250 useconds to count as a lowering */
262 static int32_t rack_pace_one_seg = 0;		/* Shall we pace for less than 1.4Meg 1MSS at a time */
263 static int32_t rack_probertt_clear_is = 1;
264 static int32_t rack_max_drain_hbp = 1;		/* Extra drain times gpsrtt for highly buffered paths */
265 static int32_t rack_hbp_thresh = 3;		/* what is the divisor max_rtt/min_rtt to decided a hbp */
266 
267 /* Part of pacing */
268 static int32_t rack_max_per_above = 30;		/* When we go to increment stop if above 100+this% */
269 
270 /* Timely information */
271 /* Combine these two gives the range of 'no change' to bw */
272 /* ie the up/down provide the upper and lower bound  */
273 static int32_t rack_gp_per_bw_mul_up = 2;	/* 2% */
274 static int32_t rack_gp_per_bw_mul_down = 4;	/* 4% */
275 static int32_t rack_gp_rtt_maxmul = 3;		/* 3 x maxmin */
276 static int32_t rack_gp_rtt_minmul = 1;		/* minrtt + (minrtt/mindiv) is lower rtt */
277 static int32_t rack_gp_rtt_mindiv = 4;		/* minrtt + (minrtt * minmul/mindiv) is lower rtt */
278 static int32_t rack_gp_decrease_per = 20;	/* 20% decrease in multipler */
279 static int32_t rack_gp_increase_per = 2;	/* 2% increase in multipler */
280 static int32_t rack_per_lower_bound = 50;	/* Don't allow to drop below this multiplier */
281 static int32_t rack_per_upper_bound_ss = 0;	/* Don't allow SS to grow above this */
282 static int32_t rack_per_upper_bound_ca = 0;	/* Don't allow CA to grow above this */
283 static int32_t rack_do_dyn_mul = 0;		/* Are the rack gp multipliers dynamic */
284 static int32_t rack_gp_no_rec_chg = 1;		/* Prohibit recovery from reducing it's multiplier */
285 static int32_t rack_timely_dec_clear = 6;	/* Do we clear decrement count at a value (6)? */
286 static int32_t rack_timely_max_push_rise = 3;	/* One round of pushing */
287 static int32_t rack_timely_max_push_drop = 3;	/* Three round of pushing */
288 static int32_t rack_timely_min_segs = 4;	/* 4 segment minimum */
289 static int32_t rack_use_max_for_nobackoff = 0;
290 static int32_t rack_timely_int_timely_only = 0;	/* do interim timely's only use the timely algo (no b/w changes)? */
291 static int32_t rack_timely_no_stopping = 0;
292 static int32_t rack_down_raise_thresh = 100;
293 static int32_t rack_req_segs = 1;
294 
295 /* Weird delayed ack mode */
296 static int32_t rack_use_imac_dack = 0;
297 /* Rack specific counters */
298 counter_u64_t rack_badfr;
299 counter_u64_t rack_badfr_bytes;
300 counter_u64_t rack_rtm_prr_retran;
301 counter_u64_t rack_rtm_prr_newdata;
302 counter_u64_t rack_timestamp_mismatch;
303 counter_u64_t rack_reorder_seen;
304 counter_u64_t rack_paced_segments;
305 counter_u64_t rack_unpaced_segments;
306 counter_u64_t rack_calc_zero;
307 counter_u64_t rack_calc_nonzero;
308 counter_u64_t rack_saw_enobuf;
309 counter_u64_t rack_saw_enetunreach;
310 counter_u64_t rack_per_timer_hole;
311 
312 /* Tail loss probe counters */
313 counter_u64_t rack_tlp_tot;
314 counter_u64_t rack_tlp_newdata;
315 counter_u64_t rack_tlp_retran;
316 counter_u64_t rack_tlp_retran_bytes;
317 counter_u64_t rack_tlp_retran_fail;
318 counter_u64_t rack_to_tot;
319 counter_u64_t rack_to_arm_rack;
320 counter_u64_t rack_to_arm_tlp;
321 counter_u64_t rack_to_alloc;
322 counter_u64_t rack_to_alloc_hard;
323 counter_u64_t rack_to_alloc_emerg;
324 counter_u64_t rack_to_alloc_limited;
325 counter_u64_t rack_alloc_limited_conns;
326 counter_u64_t rack_split_limited;
327 
328 counter_u64_t rack_sack_proc_all;
329 counter_u64_t rack_sack_proc_short;
330 counter_u64_t rack_sack_proc_restart;
331 counter_u64_t rack_sack_attacks_detected;
332 counter_u64_t rack_sack_attacks_reversed;
333 counter_u64_t rack_sack_used_next_merge;
334 counter_u64_t rack_sack_splits;
335 counter_u64_t rack_sack_used_prev_merge;
336 counter_u64_t rack_sack_skipped_acked;
337 counter_u64_t rack_ack_total;
338 counter_u64_t rack_express_sack;
339 counter_u64_t rack_sack_total;
340 counter_u64_t rack_move_none;
341 counter_u64_t rack_move_some;
342 
343 counter_u64_t rack_used_tlpmethod;
344 counter_u64_t rack_used_tlpmethod2;
345 counter_u64_t rack_enter_tlp_calc;
346 counter_u64_t rack_input_idle_reduces;
347 counter_u64_t rack_collapsed_win;
348 counter_u64_t rack_tlp_does_nada;
349 counter_u64_t rack_try_scwnd;
350 
351 /* Temp CPU counters */
352 counter_u64_t rack_find_high;
353 
354 counter_u64_t rack_progress_drops;
355 counter_u64_t rack_out_size[TCP_MSS_ACCT_SIZE];
356 counter_u64_t rack_opts_arry[RACK_OPTS_SIZE];
357 
358 static void
359 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick,  int event, int line);
360 
361 static int
362 rack_process_ack(struct mbuf *m, struct tcphdr *th,
363     struct socket *so, struct tcpcb *tp, struct tcpopt *to,
364     uint32_t tiwin, int32_t tlen, int32_t * ofia, int32_t thflags, int32_t * ret_val);
365 static int
366 rack_process_data(struct mbuf *m, struct tcphdr *th,
367     struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
368     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt);
369 static void
370 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack,
371     struct tcphdr *th, uint16_t nsegs, uint16_t type, int32_t recovery);
372 static struct rack_sendmap *rack_alloc(struct tcp_rack *rack);
373 static struct rack_sendmap *rack_alloc_limit(struct tcp_rack *rack,
374     uint8_t limit_type);
375 static struct rack_sendmap *
376 rack_check_recovery_mode(struct tcpcb *tp,
377     uint32_t tsused);
378 static void
379 rack_cong_signal(struct tcpcb *tp, struct tcphdr *th,
380     uint32_t type);
381 static void rack_counter_destroy(void);
382 static int
383 rack_ctloutput(struct socket *so, struct sockopt *sopt,
384     struct inpcb *inp, struct tcpcb *tp);
385 static int32_t rack_ctor(void *mem, int32_t size, void *arg, int32_t how);
386 static void
387 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line);
388 static void
389 rack_do_segment(struct mbuf *m, struct tcphdr *th,
390     struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
391     uint8_t iptos);
392 static void rack_dtor(void *mem, int32_t size, void *arg);
393 static void
394 rack_earlier_retran(struct tcpcb *tp, struct rack_sendmap *rsm,
395     uint32_t t, uint32_t cts);
396 static void
397 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
398     uint32_t flex1, uint32_t flex2,
399     uint32_t flex3, uint32_t flex4,
400     uint32_t flex5, uint32_t flex6,
401     uint16_t flex7, uint8_t mod);
402 static void
403 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
404    uint64_t bw_est, uint64_t bw, uint64_t len_time, int method, int line, struct rack_sendmap *rsm);
405 static struct rack_sendmap *
406 rack_find_high_nonack(struct tcp_rack *rack,
407     struct rack_sendmap *rsm);
408 static struct rack_sendmap *rack_find_lowest_rsm(struct tcp_rack *rack);
409 static void rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm);
410 static void rack_fini(struct tcpcb *tp, int32_t tcb_is_purged);
411 static int
412 rack_get_sockopt(struct socket *so, struct sockopt *sopt,
413     struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack);
414 static void
415 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
416 			    tcp_seq th_ack, int line);
417 static uint32_t
418 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss);
419 static int32_t rack_handoff_ok(struct tcpcb *tp);
420 static int32_t rack_init(struct tcpcb *tp);
421 static void rack_init_sysctls(void);
422 static void
423 rack_log_ack(struct tcpcb *tp, struct tcpopt *to,
424     struct tcphdr *th);
425 static void
426 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
427     uint32_t seq_out, uint8_t th_flags, int32_t err, uint32_t ts,
428     uint8_t pass, struct rack_sendmap *hintrsm, uint32_t us_cts);
429 static void
430 rack_log_sack_passed(struct tcpcb *tp, struct tcp_rack *rack,
431     struct rack_sendmap *rsm);
432 static void rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm);
433 static int32_t rack_output(struct tcpcb *tp);
434 
435 static uint32_t
436 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack,
437     struct sackblk *sack, struct tcpopt *to, struct rack_sendmap **prsm,
438     uint32_t cts, int *moved_two);
439 static void rack_post_recovery(struct tcpcb *tp, struct tcphdr *th);
440 static void rack_remxt_tmr(struct tcpcb *tp);
441 static int
442 rack_set_sockopt(struct socket *so, struct sockopt *sopt,
443     struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack);
444 static void rack_set_state(struct tcpcb *tp, struct tcp_rack *rack);
445 static int32_t rack_stopall(struct tcpcb *tp);
446 static void
447 rack_timer_activate(struct tcpcb *tp, uint32_t timer_type,
448     uint32_t delta);
449 static int32_t rack_timer_active(struct tcpcb *tp, uint32_t timer_type);
450 static void rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line);
451 static void rack_timer_stop(struct tcpcb *tp, uint32_t timer_type);
452 static uint32_t
453 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
454     struct rack_sendmap *rsm, uint32_t ts, int32_t * lenp);
455 static void
456 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
457     struct rack_sendmap *rsm, uint32_t ts);
458 static int
459 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
460     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack);
461 static int32_t tcp_addrack(module_t mod, int32_t type, void *data);
462 static int
463 rack_do_close_wait(struct mbuf *m, struct tcphdr *th,
464     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
465     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
466 static int
467 rack_do_closing(struct mbuf *m, struct tcphdr *th,
468     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
469     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
470 static int
471 rack_do_established(struct mbuf *m, struct tcphdr *th,
472     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
473     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
474 static int
475 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th,
476     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
477     int32_t tlen, uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos);
478 static int
479 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th,
480     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
481     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
482 static int
483 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th,
484     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
485     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
486 static int
487 rack_do_lastack(struct mbuf *m, struct tcphdr *th,
488     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
489     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
490 static int
491 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th,
492     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
493     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
494 static int
495 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th,
496     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
497     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
498 struct rack_sendmap *
499 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack,
500     uint32_t tsused);
501 static void tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt,
502     uint32_t len, uint32_t us_tim, int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt);
503 static void
504      tcp_rack_partialack(struct tcpcb *tp, struct tcphdr *th);
505 
506 int32_t rack_clear_counter=0;
507 
508 static int
sysctl_rack_clear(SYSCTL_HANDLER_ARGS)509 sysctl_rack_clear(SYSCTL_HANDLER_ARGS)
510 {
511 	uint32_t stat;
512 	int32_t error;
513 
514 	error = SYSCTL_OUT(req, &rack_clear_counter, sizeof(uint32_t));
515 	if (error || req->newptr == NULL)
516 		return error;
517 
518 	error = SYSCTL_IN(req, &stat, sizeof(uint32_t));
519 	if (error)
520 		return (error);
521 	if (stat == 1) {
522 #ifdef INVARIANTS
523 		printf("Clearing RACK counters\n");
524 #endif
525 		counter_u64_zero(rack_badfr);
526 		counter_u64_zero(rack_badfr_bytes);
527 		counter_u64_zero(rack_rtm_prr_retran);
528 		counter_u64_zero(rack_rtm_prr_newdata);
529 		counter_u64_zero(rack_timestamp_mismatch);
530 		counter_u64_zero(rack_reorder_seen);
531 		counter_u64_zero(rack_tlp_tot);
532 		counter_u64_zero(rack_tlp_newdata);
533 		counter_u64_zero(rack_tlp_retran);
534 		counter_u64_zero(rack_tlp_retran_bytes);
535 		counter_u64_zero(rack_tlp_retran_fail);
536 		counter_u64_zero(rack_to_tot);
537 		counter_u64_zero(rack_to_arm_rack);
538 		counter_u64_zero(rack_to_arm_tlp);
539 		counter_u64_zero(rack_paced_segments);
540 		counter_u64_zero(rack_calc_zero);
541 		counter_u64_zero(rack_calc_nonzero);
542 		counter_u64_zero(rack_unpaced_segments);
543 		counter_u64_zero(rack_saw_enobuf);
544 		counter_u64_zero(rack_saw_enetunreach);
545 		counter_u64_zero(rack_per_timer_hole);
546 		counter_u64_zero(rack_to_alloc_hard);
547 		counter_u64_zero(rack_to_alloc_emerg);
548 		counter_u64_zero(rack_sack_proc_all);
549 		counter_u64_zero(rack_sack_proc_short);
550 		counter_u64_zero(rack_sack_proc_restart);
551 		counter_u64_zero(rack_to_alloc);
552 		counter_u64_zero(rack_to_alloc_limited);
553 		counter_u64_zero(rack_alloc_limited_conns);
554 		counter_u64_zero(rack_split_limited);
555 		counter_u64_zero(rack_find_high);
556 		counter_u64_zero(rack_sack_attacks_detected);
557 		counter_u64_zero(rack_sack_attacks_reversed);
558 		counter_u64_zero(rack_sack_used_next_merge);
559 		counter_u64_zero(rack_sack_used_prev_merge);
560 		counter_u64_zero(rack_sack_splits);
561 		counter_u64_zero(rack_sack_skipped_acked);
562 		counter_u64_zero(rack_ack_total);
563 		counter_u64_zero(rack_express_sack);
564 		counter_u64_zero(rack_sack_total);
565 		counter_u64_zero(rack_move_none);
566 		counter_u64_zero(rack_move_some);
567 		counter_u64_zero(rack_used_tlpmethod);
568 		counter_u64_zero(rack_used_tlpmethod2);
569 		counter_u64_zero(rack_enter_tlp_calc);
570 		counter_u64_zero(rack_progress_drops);
571 		counter_u64_zero(rack_tlp_does_nada);
572 		counter_u64_zero(rack_try_scwnd);
573 		counter_u64_zero(rack_collapsed_win);
574 	}
575 	rack_clear_counter = 0;
576 	return (0);
577 }
578 
579 static void
rack_init_sysctls(void)580 rack_init_sysctls(void)
581 {
582 	struct sysctl_oid *rack_counters;
583 	struct sysctl_oid *rack_attack;
584 	struct sysctl_oid *rack_pacing;
585 	struct sysctl_oid *rack_timely;
586 	struct sysctl_oid *rack_timers;
587 	struct sysctl_oid *rack_tlp;
588 	struct sysctl_oid *rack_misc;
589 	struct sysctl_oid *rack_measure;
590 	struct sysctl_oid *rack_probertt;
591 
592 	rack_attack = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
593 	    SYSCTL_CHILDREN(rack_sysctl_root),
594 	    OID_AUTO,
595 	    "sack_attack",
596 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
597 	    "Rack Sack Attack Counters and Controls");
598 	rack_counters = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
599 	    SYSCTL_CHILDREN(rack_sysctl_root),
600 	    OID_AUTO,
601 	    "stats",
602 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
603 	    "Rack Counters");
604 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
605 	    SYSCTL_CHILDREN(rack_sysctl_root),
606 	    OID_AUTO, "rate_sample_method", CTLFLAG_RW,
607 	    &rack_rate_sample_method , USE_RTT_LOW,
608 	    "What method should we use for rate sampling 0=high, 1=low ");
609 	/* Probe rtt related controls */
610 	rack_probertt = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
611 	    SYSCTL_CHILDREN(rack_sysctl_root),
612 	    OID_AUTO,
613 	    "probertt",
614 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
615 	    "ProbeRTT related Controls");
616 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
617 	    SYSCTL_CHILDREN(rack_probertt),
618 	    OID_AUTO, "exit_per_hpb", CTLFLAG_RW,
619 	    &rack_atexit_prtt_hbp, 130,
620 	    "What percentage above goodput do we clamp CA/SS to at exit on high-BDP path 110%");
621 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
622 	    SYSCTL_CHILDREN(rack_probertt),
623 	    OID_AUTO, "exit_per_nonhpb", CTLFLAG_RW,
624 	    &rack_atexit_prtt, 130,
625 	    "What percentage above goodput do we clamp CA/SS to at exit on a non high-BDP path 100%");
626 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
627 	    SYSCTL_CHILDREN(rack_probertt),
628 	    OID_AUTO, "gp_per_mul", CTLFLAG_RW,
629 	    &rack_per_of_gp_probertt, 60,
630 	    "What percentage of goodput do we pace at in probertt");
631 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
632 	    SYSCTL_CHILDREN(rack_probertt),
633 	    OID_AUTO, "gp_per_reduce", CTLFLAG_RW,
634 	    &rack_per_of_gp_probertt_reduce, 10,
635 	    "What percentage of goodput do we reduce every gp_srtt");
636 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
637 	    SYSCTL_CHILDREN(rack_probertt),
638 	    OID_AUTO, "gp_per_low", CTLFLAG_RW,
639 	    &rack_per_of_gp_lowthresh, 40,
640 	    "What percentage of goodput do we allow the multiplier to fall to");
641 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
642 	    SYSCTL_CHILDREN(rack_probertt),
643 	    OID_AUTO, "time_between", CTLFLAG_RW,
644 	    & rack_time_between_probertt, 96000000,
645 	    "How many useconds between the lowest rtt falling must past before we enter probertt");
646 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
647 	    SYSCTL_CHILDREN(rack_probertt),
648 	    OID_AUTO, "safety", CTLFLAG_RW,
649 	    &rack_probe_rtt_safety_val, 2000000,
650 	    "If not zero, provides a maximum usecond that you can stay in probertt (2sec = 2000000)");
651 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
652 	    SYSCTL_CHILDREN(rack_probertt),
653 	    OID_AUTO, "sets_cwnd", CTLFLAG_RW,
654 	    &rack_probe_rtt_sets_cwnd, 0,
655 	    "Do we set the cwnd too (if always_lower is on)");
656 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
657 	    SYSCTL_CHILDREN(rack_probertt),
658 	    OID_AUTO, "maxdrainsrtts", CTLFLAG_RW,
659 	    &rack_max_drain_wait, 2,
660 	    "Maximum number of gp_srtt's to hold in drain waiting for flight to reach goal");
661 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
662 	    SYSCTL_CHILDREN(rack_probertt),
663 	    OID_AUTO, "mustdrainsrtts", CTLFLAG_RW,
664 	    &rack_must_drain, 1,
665 	    "We must drain this many gp_srtt's waiting for flight to reach goal");
666 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
667 	    SYSCTL_CHILDREN(rack_probertt),
668 	    OID_AUTO, "goal_use_min_entry", CTLFLAG_RW,
669 	    &rack_probertt_use_min_rtt_entry, 1,
670 	    "Should we use the min-rtt to calculate the goal rtt (else gp_srtt) at entry");
671 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
672 	    SYSCTL_CHILDREN(rack_probertt),
673 	    OID_AUTO, "goal_use_min_exit", CTLFLAG_RW,
674 	    &rack_probertt_use_min_rtt_exit, 0,
675 	    "How to set cwnd at exit, 0 - dynamic, 1 - use min-rtt, 2 - use curgprtt, 3 - entry gp-rtt");
676 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
677 	    SYSCTL_CHILDREN(rack_probertt),
678 	    OID_AUTO, "length_div", CTLFLAG_RW,
679 	    &rack_probertt_gpsrtt_cnt_div, 0,
680 	    "How many recent goodput srtt periods plus hold tim does probertt last (bottom of fraction)");
681 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
682 	    SYSCTL_CHILDREN(rack_probertt),
683 	    OID_AUTO, "length_mul", CTLFLAG_RW,
684 	    &rack_probertt_gpsrtt_cnt_mul, 0,
685 	    "How many recent goodput srtt periods plus hold tim does probertt last (top of fraction)");
686 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
687 	    SYSCTL_CHILDREN(rack_probertt),
688 	    OID_AUTO, "holdtim_at_target", CTLFLAG_RW,
689 	    &rack_min_probertt_hold, 200000,
690 	    "What is the minimum time we hold probertt at target");
691 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
692 	    SYSCTL_CHILDREN(rack_probertt),
693 	    OID_AUTO, "filter_life", CTLFLAG_RW,
694 	    &rack_probertt_filter_life, 10000000,
695 	    "What is the time for the filters life in useconds");
696 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
697 	    SYSCTL_CHILDREN(rack_probertt),
698 	    OID_AUTO, "lower_within", CTLFLAG_RW,
699 	    &rack_probertt_lower_within, 10,
700 	    "If the rtt goes lower within this percentage of the time, go into probe-rtt");
701 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
702 	    SYSCTL_CHILDREN(rack_probertt),
703 	    OID_AUTO, "must_move", CTLFLAG_RW,
704 	    &rack_min_rtt_movement, 250,
705 	    "How much is the minimum movement in rtt to count as a drop for probertt purposes");
706 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
707 	    SYSCTL_CHILDREN(rack_probertt),
708 	    OID_AUTO, "clear_is_cnts", CTLFLAG_RW,
709 	    &rack_probertt_clear_is, 1,
710 	    "Do we clear I/S counts on exiting probe-rtt");
711 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
712 	    SYSCTL_CHILDREN(rack_probertt),
713 	    OID_AUTO, "hbp_extra_drain", CTLFLAG_RW,
714 	    &rack_max_drain_hbp, 1,
715 	    "How many extra drain gpsrtt's do we get in highly buffered paths");
716 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
717 	    SYSCTL_CHILDREN(rack_probertt),
718 	    OID_AUTO, "hbp_threshold", CTLFLAG_RW,
719 	    &rack_hbp_thresh, 3,
720 	    "We are highly buffered if min_rtt_seen / max_rtt_seen > this-threshold");
721 	/* Pacing related sysctls */
722 	rack_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
723 	    SYSCTL_CHILDREN(rack_sysctl_root),
724 	    OID_AUTO,
725 	    "pacing",
726 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
727 	    "Pacing related Controls");
728 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
729 	    SYSCTL_CHILDREN(rack_pacing),
730 	    OID_AUTO, "max_pace_over", CTLFLAG_RW,
731 	    &rack_max_per_above, 30,
732 	    "What is the maximum allowable percentage that we can pace above (so 30 = 130% of our goal)");
733 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
734 	    SYSCTL_CHILDREN(rack_pacing),
735 	    OID_AUTO, "pace_to_one", CTLFLAG_RW,
736 	    &rack_pace_one_seg, 0,
737 	    "Do we allow low b/w pacing of 1MSS instead of two");
738 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
739 	    SYSCTL_CHILDREN(rack_pacing),
740 	    OID_AUTO, "limit_wsrtt", CTLFLAG_RW,
741 	    &rack_limit_time_with_srtt, 0,
742 	    "Do we limit pacing time based on srtt");
743 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
744 	    SYSCTL_CHILDREN(rack_pacing),
745 	    OID_AUTO, "init_win", CTLFLAG_RW,
746 	    &rack_default_init_window, 0,
747 	    "Do we have a rack initial window 0 = system default");
748 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
749 	    SYSCTL_CHILDREN(rack_pacing),
750 	    OID_AUTO, "hw_pacing_adjust", CTLFLAG_RW,
751 	    &rack_hw_pace_adjust, 0,
752 	    "What percentage do we raise the MSS by (11 = 1.1%)");
753 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
754 	    SYSCTL_CHILDREN(rack_pacing),
755 	    OID_AUTO, "gp_per_ss", CTLFLAG_RW,
756 	    &rack_per_of_gp_ss, 250,
757 	    "If non zero, what percentage of goodput to pace at in slow start");
758 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
759 	    SYSCTL_CHILDREN(rack_pacing),
760 	    OID_AUTO, "gp_per_ca", CTLFLAG_RW,
761 	    &rack_per_of_gp_ca, 150,
762 	    "If non zero, what percentage of goodput to pace at in congestion avoidance");
763 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
764 	    SYSCTL_CHILDREN(rack_pacing),
765 	    OID_AUTO, "gp_per_rec", CTLFLAG_RW,
766 	    &rack_per_of_gp_rec, 200,
767 	    "If non zero, what percentage of goodput to pace at in recovery");
768 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
769 	    SYSCTL_CHILDREN(rack_pacing),
770 	    OID_AUTO, "pace_max_seg", CTLFLAG_RW,
771 	    &rack_hptsi_segments, 40,
772 	    "What size is the max for TSO segments in pacing and burst mitigation");
773 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
774 	    SYSCTL_CHILDREN(rack_pacing),
775 	    OID_AUTO, "burst_reduces", CTLFLAG_RW,
776 	    &rack_slot_reduction, 4,
777 	    "When doing only burst mitigation what is the reduce divisor");
778 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
779 	    SYSCTL_CHILDREN(rack_sysctl_root),
780 	    OID_AUTO, "use_pacing", CTLFLAG_RW,
781 	    &rack_pace_every_seg, 0,
782 	    "If set we use pacing, if clear we use only the original burst mitigation");
783 
784 	rack_timely = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
785 	    SYSCTL_CHILDREN(rack_sysctl_root),
786 	    OID_AUTO,
787 	    "timely",
788 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
789 	    "Rack Timely RTT Controls");
790 	/* Timely based GP dynmics */
791 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
792 	    SYSCTL_CHILDREN(rack_timely),
793 	    OID_AUTO, "upper", CTLFLAG_RW,
794 	    &rack_gp_per_bw_mul_up, 2,
795 	    "Rack timely upper range for equal b/w (in percentage)");
796 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
797 	    SYSCTL_CHILDREN(rack_timely),
798 	    OID_AUTO, "lower", CTLFLAG_RW,
799 	    &rack_gp_per_bw_mul_down, 4,
800 	    "Rack timely lower range for equal b/w (in percentage)");
801 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
802 	    SYSCTL_CHILDREN(rack_timely),
803 	    OID_AUTO, "rtt_max_mul", CTLFLAG_RW,
804 	    &rack_gp_rtt_maxmul, 3,
805 	    "Rack timely multipler of lowest rtt for rtt_max");
806 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
807 	    SYSCTL_CHILDREN(rack_timely),
808 	    OID_AUTO, "rtt_min_div", CTLFLAG_RW,
809 	    &rack_gp_rtt_mindiv, 4,
810 	    "Rack timely divisor used for rtt + (rtt * mul/divisor) for check for lower rtt");
811 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
812 	    SYSCTL_CHILDREN(rack_timely),
813 	    OID_AUTO, "rtt_min_mul", CTLFLAG_RW,
814 	    &rack_gp_rtt_minmul, 1,
815 	    "Rack timely multiplier used for rtt + (rtt * mul/divisor) for check for lower rtt");
816 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
817 	    SYSCTL_CHILDREN(rack_timely),
818 	    OID_AUTO, "decrease", CTLFLAG_RW,
819 	    &rack_gp_decrease_per, 20,
820 	    "Rack timely decrease percentage of our GP multiplication factor");
821 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
822 	    SYSCTL_CHILDREN(rack_timely),
823 	    OID_AUTO, "increase", CTLFLAG_RW,
824 	    &rack_gp_increase_per, 2,
825 	    "Rack timely increase perentage of our GP multiplication factor");
826 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
827 	    SYSCTL_CHILDREN(rack_timely),
828 	    OID_AUTO, "lowerbound", CTLFLAG_RW,
829 	    &rack_per_lower_bound, 50,
830 	    "Rack timely lowest percentage we allow GP multiplier to fall to");
831 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
832 	    SYSCTL_CHILDREN(rack_timely),
833 	    OID_AUTO, "upperboundss", CTLFLAG_RW,
834 	    &rack_per_upper_bound_ss, 0,
835 	    "Rack timely higest percentage we allow GP multiplier in SS to raise to (0 is no upperbound)");
836 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
837 	    SYSCTL_CHILDREN(rack_timely),
838 	    OID_AUTO, "upperboundca", CTLFLAG_RW,
839 	    &rack_per_upper_bound_ca, 0,
840 	    "Rack timely higest percentage we allow GP multiplier to CA raise to (0 is no upperbound)");
841 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
842 	    SYSCTL_CHILDREN(rack_timely),
843 	    OID_AUTO, "dynamicgp", CTLFLAG_RW,
844 	    &rack_do_dyn_mul, 0,
845 	    "Rack timely do we enable dynmaic timely goodput by default");
846 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
847 	    SYSCTL_CHILDREN(rack_timely),
848 	    OID_AUTO, "no_rec_red", CTLFLAG_RW,
849 	    &rack_gp_no_rec_chg, 1,
850 	    "Rack timely do we prohibit the recovery multiplier from being lowered");
851 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
852 	    SYSCTL_CHILDREN(rack_timely),
853 	    OID_AUTO, "red_clear_cnt", CTLFLAG_RW,
854 	    &rack_timely_dec_clear, 6,
855 	    "Rack timely what threshold do we count to before another boost during b/w decent");
856 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
857 	    SYSCTL_CHILDREN(rack_timely),
858 	    OID_AUTO, "max_push_rise", CTLFLAG_RW,
859 	    &rack_timely_max_push_rise, 3,
860 	    "Rack timely how many times do we push up with b/w increase");
861 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
862 	    SYSCTL_CHILDREN(rack_timely),
863 	    OID_AUTO, "max_push_drop", CTLFLAG_RW,
864 	    &rack_timely_max_push_drop, 3,
865 	    "Rack timely how many times do we push back on b/w decent");
866 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
867 	    SYSCTL_CHILDREN(rack_timely),
868 	    OID_AUTO, "min_segs", CTLFLAG_RW,
869 	    &rack_timely_min_segs, 4,
870 	    "Rack timely when setting the cwnd what is the min num segments");
871 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
872 	    SYSCTL_CHILDREN(rack_timely),
873 	    OID_AUTO, "noback_max", CTLFLAG_RW,
874 	    &rack_use_max_for_nobackoff, 0,
875 	    "Rack timely when deciding if to backoff on a loss, do we use under max rtt else min");
876 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
877 	    SYSCTL_CHILDREN(rack_timely),
878 	    OID_AUTO, "interim_timely_only", CTLFLAG_RW,
879 	    &rack_timely_int_timely_only, 0,
880 	    "Rack timely when doing interim timely's do we only do timely (no b/w consideration)");
881 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
882 	    SYSCTL_CHILDREN(rack_timely),
883 	    OID_AUTO, "nonstop", CTLFLAG_RW,
884 	    &rack_timely_no_stopping, 0,
885 	    "Rack timely don't stop increase");
886 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
887 	    SYSCTL_CHILDREN(rack_timely),
888 	    OID_AUTO, "dec_raise_thresh", CTLFLAG_RW,
889 	    &rack_down_raise_thresh, 100,
890 	    "If the CA or SS is below this threshold raise on the first 3 b/w lowers (0=always)");
891 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
892 	    SYSCTL_CHILDREN(rack_timely),
893 	    OID_AUTO, "bottom_drag_segs", CTLFLAG_RW,
894 	    &rack_req_segs, 1,
895 	    "Bottom dragging if not these many segments outstanding and room");
896 
897 	/* TLP and Rack related parameters */
898 	rack_tlp = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
899 	    SYSCTL_CHILDREN(rack_sysctl_root),
900 	    OID_AUTO,
901 	    "tlp",
902 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
903 	    "TLP and Rack related Controls");
904 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
905 	    SYSCTL_CHILDREN(rack_tlp),
906 	    OID_AUTO, "use_rrr", CTLFLAG_RW,
907 	    &use_rack_rr, 1,
908 	    "Do we use Rack Rapid Recovery");
909 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
910 	    SYSCTL_CHILDREN(rack_tlp),
911 	    OID_AUTO, "nonrxt_use_cr", CTLFLAG_RW,
912 	    &rack_non_rxt_use_cr, 0,
913 	    "Do we use ss/ca rate if in recovery we are transmitting a new data chunk");
914 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
915 	    SYSCTL_CHILDREN(rack_tlp),
916 	    OID_AUTO, "tlpmethod", CTLFLAG_RW,
917 	    &rack_tlp_threshold_use, TLP_USE_TWO_ONE,
918 	    "What method do we do for TLP time calc 0=no-de-ack-comp, 1=ID, 2=2.1, 3=2.2");
919 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
920 	    SYSCTL_CHILDREN(rack_tlp),
921 	    OID_AUTO, "limit", CTLFLAG_RW,
922 	    &rack_tlp_limit, 2,
923 	    "How many TLP's can be sent without sending new data");
924 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
925 	    SYSCTL_CHILDREN(rack_tlp),
926 	    OID_AUTO, "use_greater", CTLFLAG_RW,
927 	    &rack_tlp_use_greater, 1,
928 	    "Should we use the rack_rtt time if its greater than srtt");
929 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
930 	    SYSCTL_CHILDREN(rack_tlp),
931 	    OID_AUTO, "tlpminto", CTLFLAG_RW,
932 	    &rack_tlp_min, 10,
933 	    "TLP minimum timeout per the specification (10ms)");
934 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
935 	    SYSCTL_CHILDREN(rack_tlp),
936 	    OID_AUTO, "send_oldest", CTLFLAG_RW,
937 	    &rack_always_send_oldest, 0,
938 	    "Should we always send the oldest TLP and RACK-TLP");
939 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
940 	    SYSCTL_CHILDREN(rack_tlp),
941 	    OID_AUTO, "rack_tlimit", CTLFLAG_RW,
942 	    &rack_limited_retran, 0,
943 	    "How many times can a rack timeout drive out sends");
944 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
945 	    SYSCTL_CHILDREN(rack_tlp),
946 	    OID_AUTO, "tlp_retry", CTLFLAG_RW,
947 	    &rack_tlp_max_resend, 2,
948 	    "How many times does TLP retry a single segment or multiple with no ACK");
949 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
950 	    SYSCTL_CHILDREN(rack_tlp),
951 	    OID_AUTO, "tlp_cwnd_flag", CTLFLAG_RW,
952 	    &rack_lower_cwnd_at_tlp, 0,
953 	    "When a TLP completes a retran should we enter recovery");
954 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
955 	    SYSCTL_CHILDREN(rack_tlp),
956 	    OID_AUTO, "reorder_thresh", CTLFLAG_RW,
957 	    &rack_reorder_thresh, 2,
958 	    "What factor for rack will be added when seeing reordering (shift right)");
959 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
960 	    SYSCTL_CHILDREN(rack_tlp),
961 	    OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW,
962 	    &rack_tlp_thresh, 1,
963 	    "What divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)");
964 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
965 	    SYSCTL_CHILDREN(rack_tlp),
966 	    OID_AUTO, "reorder_fade", CTLFLAG_RW,
967 	    &rack_reorder_fade, 0,
968 	    "Does reorder detection fade, if so how many ms (0 means never)");
969 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
970 	    SYSCTL_CHILDREN(rack_tlp),
971 	    OID_AUTO, "pktdelay", CTLFLAG_RW,
972 	    &rack_pkt_delay, 1,
973 	    "Extra RACK time (in ms) besides reordering thresh");
974 
975 	/* Timer related controls */
976 	rack_timers = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
977 	    SYSCTL_CHILDREN(rack_sysctl_root),
978 	    OID_AUTO,
979 	    "timers",
980 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
981 	    "Timer related controls");
982 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
983 	    SYSCTL_CHILDREN(rack_timers),
984 	    OID_AUTO, "persmin", CTLFLAG_RW,
985 	    &rack_persist_min, 250,
986 	    "What is the minimum time in milliseconds between persists");
987 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
988 	    SYSCTL_CHILDREN(rack_timers),
989 	    OID_AUTO, "persmax", CTLFLAG_RW,
990 	    &rack_persist_max, 2000,
991 	    "What is the largest delay in milliseconds between persists");
992 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
993 	    SYSCTL_CHILDREN(rack_timers),
994 	    OID_AUTO, "delayed_ack", CTLFLAG_RW,
995 	    &rack_delayed_ack_time, 200,
996 	    "Delayed ack time (200ms)");
997 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
998 	    SYSCTL_CHILDREN(rack_timers),
999 	    OID_AUTO, "minrto", CTLFLAG_RW,
1000 	    &rack_rto_min, 0,
1001 	    "Minimum RTO in ms -- set with caution below 1000 due to TLP");
1002 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1003 	    SYSCTL_CHILDREN(rack_timers),
1004 	    OID_AUTO, "maxrto", CTLFLAG_RW,
1005 	    &rack_rto_max, 0,
1006 	    "Maxiumum RTO in ms -- should be at least as large as min_rto");
1007 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1008 	    SYSCTL_CHILDREN(rack_timers),
1009 	    OID_AUTO, "minto", CTLFLAG_RW,
1010 	    &rack_min_to, 1,
1011 	    "Minimum rack timeout in milliseconds");
1012 	/* Measure controls */
1013 	rack_measure = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1014 	    SYSCTL_CHILDREN(rack_sysctl_root),
1015 	    OID_AUTO,
1016 	    "measure",
1017 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1018 	    "Measure related controls");
1019 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1020 	    SYSCTL_CHILDREN(rack_measure),
1021 	    OID_AUTO, "wma_divisor", CTLFLAG_RW,
1022 	    &rack_wma_divisor, 8,
1023 	    "When doing b/w calculation what is the  divisor for the WMA");
1024 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1025 	    SYSCTL_CHILDREN(rack_measure),
1026 	    OID_AUTO, "end_cwnd", CTLFLAG_RW,
1027 	    &rack_cwnd_block_ends_measure, 0,
1028 	    "Does a cwnd just-return end the measurement window (app limited)");
1029 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1030 	    SYSCTL_CHILDREN(rack_measure),
1031 	    OID_AUTO, "end_rwnd", CTLFLAG_RW,
1032 	    &rack_rwnd_block_ends_measure, 0,
1033 	    "Does an rwnd just-return end the measurement window (app limited -- not persists)");
1034 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1035 	    SYSCTL_CHILDREN(rack_measure),
1036 	    OID_AUTO, "min_target", CTLFLAG_RW,
1037 	    &rack_def_data_window, 20,
1038 	    "What is the minimum target window (in mss) for a GP measurements");
1039 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1040 	    SYSCTL_CHILDREN(rack_measure),
1041 	    OID_AUTO, "goal_bdp", CTLFLAG_RW,
1042 	    &rack_goal_bdp, 2,
1043 	    "What is the goal BDP to measure");
1044 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1045 	    SYSCTL_CHILDREN(rack_measure),
1046 	    OID_AUTO, "min_srtts", CTLFLAG_RW,
1047 	    &rack_min_srtts, 1,
1048 	    "What is the goal BDP to measure");
1049 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1050 	    SYSCTL_CHILDREN(rack_measure),
1051 	    OID_AUTO, "min_measure_tim", CTLFLAG_RW,
1052 	    &rack_min_measure_usec, 0,
1053 	    "What is the Minimum time time for a measurement if 0, this is off");
1054 	/* Misc rack controls */
1055 	rack_misc = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1056 	    SYSCTL_CHILDREN(rack_sysctl_root),
1057 	    OID_AUTO,
1058 	    "misc",
1059 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1060 	    "Misc related controls");
1061 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1062 	    SYSCTL_CHILDREN(rack_misc),
1063 	    OID_AUTO, "shared_cwnd", CTLFLAG_RW,
1064 	    &rack_enable_shared_cwnd, 0,
1065 	    "Should RACK try to use the shared cwnd on connections where allowed");
1066 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1067 	    SYSCTL_CHILDREN(rack_misc),
1068 	    OID_AUTO, "limits_on_scwnd", CTLFLAG_RW,
1069 	    &rack_limits_scwnd, 1,
1070 	    "Should RACK place low end time limits on the shared cwnd feature");
1071 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1072 	    SYSCTL_CHILDREN(rack_misc),
1073 	    OID_AUTO, "non_paced_lro_queue", CTLFLAG_RW,
1074 	    &rack_enable_mqueue_for_nonpaced, 0,
1075 	    "Should RACK use mbuf queuing for non-paced connections");
1076 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1077 	    SYSCTL_CHILDREN(rack_misc),
1078 	    OID_AUTO, "iMac_dack", CTLFLAG_RW,
1079 	    &rack_use_imac_dack, 0,
1080 	    "Should RACK try to emulate iMac delayed ack");
1081 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1082 	    SYSCTL_CHILDREN(rack_misc),
1083 	    OID_AUTO, "no_prr", CTLFLAG_RW,
1084 	    &rack_disable_prr, 0,
1085 	    "Should RACK not use prr and only pace (must have pacing on)");
1086 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1087 	    SYSCTL_CHILDREN(rack_misc),
1088 	    OID_AUTO, "bb_verbose", CTLFLAG_RW,
1089 	    &rack_verbose_logging, 0,
1090 	    "Should RACK black box logging be verbose");
1091 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1092 	    SYSCTL_CHILDREN(rack_misc),
1093 	    OID_AUTO, "data_after_close", CTLFLAG_RW,
1094 	    &rack_ignore_data_after_close, 1,
1095 	    "Do we hold off sending a RST until all pending data is ack'd");
1096 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1097 	    SYSCTL_CHILDREN(rack_misc),
1098 	    OID_AUTO, "no_sack_needed", CTLFLAG_RW,
1099 	    &rack_sack_not_required, 0,
1100 	    "Do we allow rack to run on connections not supporting SACK");
1101 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1102 	    SYSCTL_CHILDREN(rack_misc),
1103 	    OID_AUTO, "recovery_loss_prop", CTLFLAG_RW,
1104 	    &rack_use_proportional_reduce, 0,
1105 	    "Should we proportionaly reduce cwnd based on the number of losses ");
1106 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1107 	    SYSCTL_CHILDREN(rack_misc),
1108 	    OID_AUTO, "recovery_prop", CTLFLAG_RW,
1109 	    &rack_proportional_rate, 10,
1110 	    "What percent reduction per loss");
1111 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1112 	    SYSCTL_CHILDREN(rack_misc),
1113 	    OID_AUTO, "prr_sendalot", CTLFLAG_RW,
1114 	    &rack_send_a_lot_in_prr, 1,
1115 	    "Send a lot in prr");
1116 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1117 	    SYSCTL_CHILDREN(rack_misc),
1118 	    OID_AUTO, "earlyrecovery", CTLFLAG_RW,
1119 	    &rack_early_recovery, 1,
1120 	    "Do we do early recovery with rack");
1121 	/* Sack Attacker detection stuff */
1122 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1123 	    SYSCTL_CHILDREN(rack_attack),
1124 	    OID_AUTO, "detect_highsackratio", CTLFLAG_RW,
1125 	    &rack_highest_sack_thresh_seen, 0,
1126 	    "Highest sack to ack ratio seen");
1127 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1128 	    SYSCTL_CHILDREN(rack_attack),
1129 	    OID_AUTO, "detect_highmoveratio", CTLFLAG_RW,
1130 	    &rack_highest_move_thresh_seen, 0,
1131 	    "Highest move to non-move ratio seen");
1132 	rack_ack_total = counter_u64_alloc(M_WAITOK);
1133 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1134 	    SYSCTL_CHILDREN(rack_attack),
1135 	    OID_AUTO, "acktotal", CTLFLAG_RD,
1136 	    &rack_ack_total,
1137 	    "Total number of Ack's");
1138 	rack_express_sack = counter_u64_alloc(M_WAITOK);
1139 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1140 	    SYSCTL_CHILDREN(rack_attack),
1141 	    OID_AUTO, "exp_sacktotal", CTLFLAG_RD,
1142 	    &rack_express_sack,
1143 	    "Total expresss number of Sack's");
1144 	rack_sack_total = counter_u64_alloc(M_WAITOK);
1145 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1146 	    SYSCTL_CHILDREN(rack_attack),
1147 	    OID_AUTO, "sacktotal", CTLFLAG_RD,
1148 	    &rack_sack_total,
1149 	    "Total number of SACKs");
1150 	rack_move_none = counter_u64_alloc(M_WAITOK);
1151 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1152 	    SYSCTL_CHILDREN(rack_attack),
1153 	    OID_AUTO, "move_none", CTLFLAG_RD,
1154 	    &rack_move_none,
1155 	    "Total number of SACK index reuse of postions under threshold");
1156 	rack_move_some = counter_u64_alloc(M_WAITOK);
1157 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1158 	    SYSCTL_CHILDREN(rack_attack),
1159 	    OID_AUTO, "move_some", CTLFLAG_RD,
1160 	    &rack_move_some,
1161 	    "Total number of SACK index reuse of postions over threshold");
1162 	rack_sack_attacks_detected = counter_u64_alloc(M_WAITOK);
1163 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1164 	    SYSCTL_CHILDREN(rack_attack),
1165 	    OID_AUTO, "attacks", CTLFLAG_RD,
1166 	    &rack_sack_attacks_detected,
1167 	    "Total number of SACK attackers that had sack disabled");
1168 	rack_sack_attacks_reversed = counter_u64_alloc(M_WAITOK);
1169 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1170 	    SYSCTL_CHILDREN(rack_attack),
1171 	    OID_AUTO, "reversed", CTLFLAG_RD,
1172 	    &rack_sack_attacks_reversed,
1173 	    "Total number of SACK attackers that were later determined false positive");
1174 	rack_sack_used_next_merge = counter_u64_alloc(M_WAITOK);
1175 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1176 	    SYSCTL_CHILDREN(rack_attack),
1177 	    OID_AUTO, "nextmerge", CTLFLAG_RD,
1178 	    &rack_sack_used_next_merge,
1179 	    "Total number of times we used the next merge");
1180 	rack_sack_used_prev_merge = counter_u64_alloc(M_WAITOK);
1181 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1182 	    SYSCTL_CHILDREN(rack_attack),
1183 	    OID_AUTO, "prevmerge", CTLFLAG_RD,
1184 	    &rack_sack_used_prev_merge,
1185 	    "Total number of times we used the prev merge");
1186 	/* Counters */
1187 	rack_badfr = counter_u64_alloc(M_WAITOK);
1188 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1189 	    SYSCTL_CHILDREN(rack_counters),
1190 	    OID_AUTO, "badfr", CTLFLAG_RD,
1191 	    &rack_badfr, "Total number of bad FRs");
1192 	rack_badfr_bytes = counter_u64_alloc(M_WAITOK);
1193 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1194 	    SYSCTL_CHILDREN(rack_counters),
1195 	    OID_AUTO, "badfr_bytes", CTLFLAG_RD,
1196 	    &rack_badfr_bytes, "Total number of bad FRs");
1197 	rack_rtm_prr_retran = counter_u64_alloc(M_WAITOK);
1198 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1199 	    SYSCTL_CHILDREN(rack_counters),
1200 	    OID_AUTO, "prrsndret", CTLFLAG_RD,
1201 	    &rack_rtm_prr_retran,
1202 	    "Total number of prr based retransmits");
1203 	rack_rtm_prr_newdata = counter_u64_alloc(M_WAITOK);
1204 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1205 	    SYSCTL_CHILDREN(rack_counters),
1206 	    OID_AUTO, "prrsndnew", CTLFLAG_RD,
1207 	    &rack_rtm_prr_newdata,
1208 	    "Total number of prr based new transmits");
1209 	rack_timestamp_mismatch = counter_u64_alloc(M_WAITOK);
1210 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1211 	    SYSCTL_CHILDREN(rack_counters),
1212 	    OID_AUTO, "tsnf", CTLFLAG_RD,
1213 	    &rack_timestamp_mismatch,
1214 	    "Total number of timestamps that we could not find the reported ts");
1215 	rack_find_high = counter_u64_alloc(M_WAITOK);
1216 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1217 	    SYSCTL_CHILDREN(rack_counters),
1218 	    OID_AUTO, "findhigh", CTLFLAG_RD,
1219 	    &rack_find_high,
1220 	    "Total number of FIN causing find-high");
1221 	rack_reorder_seen = counter_u64_alloc(M_WAITOK);
1222 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1223 	    SYSCTL_CHILDREN(rack_counters),
1224 	    OID_AUTO, "reordering", CTLFLAG_RD,
1225 	    &rack_reorder_seen,
1226 	    "Total number of times we added delay due to reordering");
1227 	rack_tlp_tot = counter_u64_alloc(M_WAITOK);
1228 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1229 	    SYSCTL_CHILDREN(rack_counters),
1230 	    OID_AUTO, "tlp_to_total", CTLFLAG_RD,
1231 	    &rack_tlp_tot,
1232 	    "Total number of tail loss probe expirations");
1233 	rack_tlp_newdata = counter_u64_alloc(M_WAITOK);
1234 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1235 	    SYSCTL_CHILDREN(rack_counters),
1236 	    OID_AUTO, "tlp_new", CTLFLAG_RD,
1237 	    &rack_tlp_newdata,
1238 	    "Total number of tail loss probe sending new data");
1239 	rack_tlp_retran = counter_u64_alloc(M_WAITOK);
1240 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1241 	    SYSCTL_CHILDREN(rack_counters),
1242 	    OID_AUTO, "tlp_retran", CTLFLAG_RD,
1243 	    &rack_tlp_retran,
1244 	    "Total number of tail loss probe sending retransmitted data");
1245 	rack_tlp_retran_bytes = counter_u64_alloc(M_WAITOK);
1246 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1247 	    SYSCTL_CHILDREN(rack_counters),
1248 	    OID_AUTO, "tlp_retran_bytes", CTLFLAG_RD,
1249 	    &rack_tlp_retran_bytes,
1250 	    "Total bytes of tail loss probe sending retransmitted data");
1251 	rack_tlp_retran_fail = counter_u64_alloc(M_WAITOK);
1252 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1253 	    SYSCTL_CHILDREN(rack_counters),
1254 	    OID_AUTO, "tlp_retran_fail", CTLFLAG_RD,
1255 	    &rack_tlp_retran_fail,
1256 	    "Total number of tail loss probe sending retransmitted data that failed (wait for t3)");
1257 	rack_to_tot = counter_u64_alloc(M_WAITOK);
1258 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1259 	    SYSCTL_CHILDREN(rack_counters),
1260 	    OID_AUTO, "rack_to_tot", CTLFLAG_RD,
1261 	    &rack_to_tot,
1262 	    "Total number of times the rack to expired");
1263 	rack_to_arm_rack = counter_u64_alloc(M_WAITOK);
1264 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1265 	    SYSCTL_CHILDREN(rack_counters),
1266 	    OID_AUTO, "arm_rack", CTLFLAG_RD,
1267 	    &rack_to_arm_rack,
1268 	    "Total number of times the rack timer armed");
1269 	rack_to_arm_tlp = counter_u64_alloc(M_WAITOK);
1270 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1271 	    SYSCTL_CHILDREN(rack_counters),
1272 	    OID_AUTO, "arm_tlp", CTLFLAG_RD,
1273 	    &rack_to_arm_tlp,
1274 	    "Total number of times the tlp timer armed");
1275 	rack_calc_zero = counter_u64_alloc(M_WAITOK);
1276 	rack_calc_nonzero = counter_u64_alloc(M_WAITOK);
1277 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1278 	    SYSCTL_CHILDREN(rack_counters),
1279 	    OID_AUTO, "calc_zero", CTLFLAG_RD,
1280 	    &rack_calc_zero,
1281 	    "Total number of times pacing time worked out to zero");
1282 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1283 	    SYSCTL_CHILDREN(rack_counters),
1284 	    OID_AUTO, "calc_nonzero", CTLFLAG_RD,
1285 	    &rack_calc_nonzero,
1286 	    "Total number of times pacing time worked out to non-zero");
1287 	rack_paced_segments = counter_u64_alloc(M_WAITOK);
1288 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1289 	    SYSCTL_CHILDREN(rack_counters),
1290 	    OID_AUTO, "paced", CTLFLAG_RD,
1291 	    &rack_paced_segments,
1292 	    "Total number of times a segment send caused hptsi");
1293 	rack_unpaced_segments = counter_u64_alloc(M_WAITOK);
1294 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1295 	    SYSCTL_CHILDREN(rack_counters),
1296 	    OID_AUTO, "unpaced", CTLFLAG_RD,
1297 	    &rack_unpaced_segments,
1298 	    "Total number of times a segment did not cause hptsi");
1299 	rack_saw_enobuf = counter_u64_alloc(M_WAITOK);
1300 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1301 	    SYSCTL_CHILDREN(rack_counters),
1302 	    OID_AUTO, "saw_enobufs", CTLFLAG_RD,
1303 	    &rack_saw_enobuf,
1304 	    "Total number of times a segment did not cause hptsi");
1305 	rack_saw_enetunreach = counter_u64_alloc(M_WAITOK);
1306 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1307 	    SYSCTL_CHILDREN(rack_counters),
1308 	    OID_AUTO, "saw_enetunreach", CTLFLAG_RD,
1309 	    &rack_saw_enetunreach,
1310 	    "Total number of times a segment did not cause hptsi");
1311 	rack_to_alloc = counter_u64_alloc(M_WAITOK);
1312 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1313 	    SYSCTL_CHILDREN(rack_counters),
1314 	    OID_AUTO, "allocs", CTLFLAG_RD,
1315 	    &rack_to_alloc,
1316 	    "Total allocations of tracking structures");
1317 	rack_to_alloc_hard = counter_u64_alloc(M_WAITOK);
1318 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1319 	    SYSCTL_CHILDREN(rack_counters),
1320 	    OID_AUTO, "allochard", CTLFLAG_RD,
1321 	    &rack_to_alloc_hard,
1322 	    "Total allocations done with sleeping the hard way");
1323 	rack_to_alloc_emerg = counter_u64_alloc(M_WAITOK);
1324 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1325 	    SYSCTL_CHILDREN(rack_counters),
1326 	    OID_AUTO, "allocemerg", CTLFLAG_RD,
1327 	    &rack_to_alloc_emerg,
1328 	    "Total allocations done from emergency cache");
1329 	rack_to_alloc_limited = counter_u64_alloc(M_WAITOK);
1330 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1331 	    SYSCTL_CHILDREN(rack_counters),
1332 	    OID_AUTO, "alloc_limited", CTLFLAG_RD,
1333 	    &rack_to_alloc_limited,
1334 	    "Total allocations dropped due to limit");
1335 	rack_alloc_limited_conns = counter_u64_alloc(M_WAITOK);
1336 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1337 	    SYSCTL_CHILDREN(rack_counters),
1338 	    OID_AUTO, "alloc_limited_conns", CTLFLAG_RD,
1339 	    &rack_alloc_limited_conns,
1340 	    "Connections with allocations dropped due to limit");
1341 	rack_split_limited = counter_u64_alloc(M_WAITOK);
1342 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1343 	    SYSCTL_CHILDREN(rack_counters),
1344 	    OID_AUTO, "split_limited", CTLFLAG_RD,
1345 	    &rack_split_limited,
1346 	    "Split allocations dropped due to limit");
1347 	rack_sack_proc_all = counter_u64_alloc(M_WAITOK);
1348 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1349 	    SYSCTL_CHILDREN(rack_counters),
1350 	    OID_AUTO, "sack_long", CTLFLAG_RD,
1351 	    &rack_sack_proc_all,
1352 	    "Total times we had to walk whole list for sack processing");
1353 	rack_sack_proc_restart = counter_u64_alloc(M_WAITOK);
1354 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1355 	    SYSCTL_CHILDREN(rack_counters),
1356 	    OID_AUTO, "sack_restart", CTLFLAG_RD,
1357 	    &rack_sack_proc_restart,
1358 	    "Total times we had to walk whole list due to a restart");
1359 	rack_sack_proc_short = counter_u64_alloc(M_WAITOK);
1360 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1361 	    SYSCTL_CHILDREN(rack_counters),
1362 	    OID_AUTO, "sack_short", CTLFLAG_RD,
1363 	    &rack_sack_proc_short,
1364 	    "Total times we took shortcut for sack processing");
1365 	rack_enter_tlp_calc = counter_u64_alloc(M_WAITOK);
1366 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1367 	    SYSCTL_CHILDREN(rack_counters),
1368 	    OID_AUTO, "tlp_calc_entered", CTLFLAG_RD,
1369 	    &rack_enter_tlp_calc,
1370 	    "Total times we called calc-tlp");
1371 	rack_used_tlpmethod = counter_u64_alloc(M_WAITOK);
1372 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1373 	    SYSCTL_CHILDREN(rack_counters),
1374 	    OID_AUTO, "hit_tlp_method", CTLFLAG_RD,
1375 	    &rack_used_tlpmethod,
1376 	    "Total number of runt sacks");
1377 	rack_used_tlpmethod2 = counter_u64_alloc(M_WAITOK);
1378 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1379 	    SYSCTL_CHILDREN(rack_counters),
1380 	    OID_AUTO, "hit_tlp_method2", CTLFLAG_RD,
1381 	    &rack_used_tlpmethod2,
1382 	    "Total number of times we hit TLP method 2");
1383 	rack_sack_skipped_acked = counter_u64_alloc(M_WAITOK);
1384 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1385 	    SYSCTL_CHILDREN(rack_attack),
1386 	    OID_AUTO, "skipacked", CTLFLAG_RD,
1387 	    &rack_sack_skipped_acked,
1388 	    "Total number of times we skipped previously sacked");
1389 	rack_sack_splits = counter_u64_alloc(M_WAITOK);
1390 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1391 	    SYSCTL_CHILDREN(rack_attack),
1392 	    OID_AUTO, "ofsplit", CTLFLAG_RD,
1393 	    &rack_sack_splits,
1394 	    "Total number of times we did the old fashion tree split");
1395 	rack_progress_drops = counter_u64_alloc(M_WAITOK);
1396 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1397 	    SYSCTL_CHILDREN(rack_counters),
1398 	    OID_AUTO, "prog_drops", CTLFLAG_RD,
1399 	    &rack_progress_drops,
1400 	    "Total number of progress drops");
1401 	rack_input_idle_reduces = counter_u64_alloc(M_WAITOK);
1402 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1403 	    SYSCTL_CHILDREN(rack_counters),
1404 	    OID_AUTO, "idle_reduce_oninput", CTLFLAG_RD,
1405 	    &rack_input_idle_reduces,
1406 	    "Total number of idle reductions on input");
1407 	rack_collapsed_win = counter_u64_alloc(M_WAITOK);
1408 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1409 	    SYSCTL_CHILDREN(rack_counters),
1410 	    OID_AUTO, "collapsed_win", CTLFLAG_RD,
1411 	    &rack_collapsed_win,
1412 	    "Total number of collapsed windows");
1413 	rack_tlp_does_nada = counter_u64_alloc(M_WAITOK);
1414 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1415 	    SYSCTL_CHILDREN(rack_counters),
1416 	    OID_AUTO, "tlp_nada", CTLFLAG_RD,
1417 	    &rack_tlp_does_nada,
1418 	    "Total number of nada tlp calls");
1419 	rack_try_scwnd = counter_u64_alloc(M_WAITOK);
1420 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1421 	    SYSCTL_CHILDREN(rack_counters),
1422 	    OID_AUTO, "tried_scwnd", CTLFLAG_RD,
1423 	    &rack_try_scwnd,
1424 	    "Total number of scwnd attempts");
1425 
1426 	rack_per_timer_hole = counter_u64_alloc(M_WAITOK);
1427 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1428 	    SYSCTL_CHILDREN(rack_counters),
1429 	    OID_AUTO, "timer_hole", CTLFLAG_RD,
1430 	    &rack_per_timer_hole,
1431 	    "Total persists start in timer hole");
1432 	COUNTER_ARRAY_ALLOC(rack_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK);
1433 	SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1434 	    OID_AUTO, "outsize", CTLFLAG_RD,
1435 	    rack_out_size, TCP_MSS_ACCT_SIZE, "MSS send sizes");
1436 	COUNTER_ARRAY_ALLOC(rack_opts_arry, RACK_OPTS_SIZE, M_WAITOK);
1437 	SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1438 	    OID_AUTO, "opts", CTLFLAG_RD,
1439 	    rack_opts_arry, RACK_OPTS_SIZE, "RACK Option Stats");
1440 	SYSCTL_ADD_PROC(&rack_sysctl_ctx,
1441 	    SYSCTL_CHILDREN(rack_sysctl_root),
1442 	    OID_AUTO, "clear", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE,
1443 	    &rack_clear_counter, 0, sysctl_rack_clear, "IU", "Clear counters");
1444 }
1445 
1446 static __inline int
rb_map_cmp(struct rack_sendmap * b,struct rack_sendmap * a)1447 rb_map_cmp(struct rack_sendmap *b, struct rack_sendmap *a)
1448 {
1449 	if (SEQ_GEQ(b->r_start, a->r_start) &&
1450 	    SEQ_LT(b->r_start, a->r_end)) {
1451 		/*
1452 		 * The entry b is within the
1453 		 * block a. i.e.:
1454 		 * a --   |-------------|
1455 		 * b --   |----|
1456 		 * <or>
1457 		 * b --       |------|
1458 		 * <or>
1459 		 * b --       |-----------|
1460 		 */
1461 		return (0);
1462 	} else if (SEQ_GEQ(b->r_start, a->r_end)) {
1463 		/*
1464 		 * b falls as either the next
1465 		 * sequence block after a so a
1466 		 * is said to be smaller than b.
1467 		 * i.e:
1468 		 * a --   |------|
1469 		 * b --          |--------|
1470 		 * or
1471 		 * b --              |-----|
1472 		 */
1473 		return (1);
1474 	}
1475 	/*
1476 	 * Whats left is where a is
1477 	 * larger than b. i.e:
1478 	 * a --         |-------|
1479 	 * b --  |---|
1480 	 * or even possibly
1481 	 * b --   |--------------|
1482 	 */
1483 	return (-1);
1484 }
1485 
1486 RB_PROTOTYPE(rack_rb_tree_head, rack_sendmap, r_next, rb_map_cmp);
1487 RB_GENERATE(rack_rb_tree_head, rack_sendmap, r_next, rb_map_cmp);
1488 
1489 static uint32_t
rc_init_window(struct tcp_rack * rack)1490 rc_init_window(struct tcp_rack *rack)
1491 {
1492 	uint32_t win;
1493 
1494 	if (rack->rc_init_win == 0) {
1495 		/*
1496 		 * Nothing set by the user, use the system stack
1497 		 * default.
1498 		 */
1499 		return(tcp_compute_initwnd(tcp_maxseg(rack->rc_tp)));
1500 	}
1501 	win = ctf_fixed_maxseg(rack->rc_tp) * rack->rc_init_win;
1502 	return(win);
1503 }
1504 
1505 static uint64_t
rack_get_fixed_pacing_bw(struct tcp_rack * rack)1506 rack_get_fixed_pacing_bw(struct tcp_rack *rack)
1507 {
1508 	if (IN_RECOVERY(rack->rc_tp->t_flags))
1509 		return (rack->r_ctl.rc_fixed_pacing_rate_rec);
1510 	else if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
1511 		return (rack->r_ctl.rc_fixed_pacing_rate_ss);
1512 	else
1513 		return (rack->r_ctl.rc_fixed_pacing_rate_ca);
1514 }
1515 
1516 static uint64_t
rack_get_bw(struct tcp_rack * rack)1517 rack_get_bw(struct tcp_rack *rack)
1518 {
1519 	if (rack->use_fixed_rate) {
1520 		/* Return the fixed pacing rate */
1521 		return (rack_get_fixed_pacing_bw(rack));
1522 	}
1523 	if (rack->r_ctl.gp_bw == 0) {
1524 		/*
1525 		 * We have yet no b/w measurement,
1526 		 * if we have a user set initial bw
1527 		 * return it. If we don't have that and
1528 		 * we have an srtt, use the tcp IW (10) to
1529 		 * calculate a fictional b/w over the SRTT
1530 		 * which is more or less a guess. Note
1531 		 * we don't use our IW from rack on purpose
1532 		 * so if we have like IW=30, we are not
1533 		 * calculating a "huge" b/w.
1534 		 */
1535 		uint64_t bw, srtt;
1536 		if (rack->r_ctl.init_rate)
1537 			return (rack->r_ctl.init_rate);
1538 
1539 		/* Has the user set a max peak rate? */
1540 #ifdef NETFLIX_PEAKRATE
1541 		if (rack->rc_tp->t_maxpeakrate)
1542 			return (rack->rc_tp->t_maxpeakrate);
1543 #endif
1544 		/* Ok lets come up with the IW guess, if we have a srtt */
1545 		if (rack->rc_tp->t_srtt == 0) {
1546 			/*
1547 			 * Go with old pacing method
1548 			 * i.e. burst mitigation only.
1549 			 */
1550 			return (0);
1551 		}
1552 		/* Ok lets get the initial TCP win (not racks) */
1553 		bw = tcp_compute_initwnd(tcp_maxseg(rack->rc_tp));
1554 		srtt = ((uint64_t)TICKS_2_USEC(rack->rc_tp->t_srtt) >> TCP_RTT_SHIFT);
1555 		bw *= (uint64_t)USECS_IN_SECOND;
1556 		bw /= srtt;
1557 		return (bw);
1558 	} else {
1559 		uint64_t bw;
1560 
1561 		if(rack->r_ctl.num_avg >= RACK_REQ_AVG) {
1562 			/* Averaging is done, we can return the value */
1563 			bw = rack->r_ctl.gp_bw;
1564 		} else {
1565 			/* Still doing initial average must calculate */
1566 			bw = rack->r_ctl.gp_bw / rack->r_ctl.num_avg;
1567 		}
1568 #ifdef NETFLIX_PEAKRATE
1569 		if ((rack->rc_tp->t_maxpeakrate) &&
1570 		    (bw > rack->rc_tp->t_maxpeakrate)) {
1571 			/* The user has set a peak rate to pace at
1572 			 * don't allow us to pace faster than that.
1573 			 */
1574 			return (rack->rc_tp->t_maxpeakrate);
1575 		}
1576 #endif
1577 		return (bw);
1578 	}
1579 }
1580 
1581 static uint16_t
rack_get_output_gain(struct tcp_rack * rack,struct rack_sendmap * rsm)1582 rack_get_output_gain(struct tcp_rack *rack, struct rack_sendmap *rsm)
1583 {
1584 	if (rack->use_fixed_rate) {
1585 		return (100);
1586 	} else if (rack->in_probe_rtt && (rsm == NULL))
1587 		return(rack->r_ctl.rack_per_of_gp_probertt);
1588 	else if ((IN_RECOVERY(rack->rc_tp->t_flags) &&
1589 		  rack->r_ctl.rack_per_of_gp_rec)) {
1590 		if (rsm) {
1591 			/* a retransmission always use the recovery rate */
1592 			return(rack->r_ctl.rack_per_of_gp_rec);
1593 		} else if (rack->rack_rec_nonrxt_use_cr) {
1594 			/* Directed to use the configured rate */
1595 			goto configured_rate;
1596 		} else if (rack->rack_no_prr &&
1597 			   (rack->r_ctl.rack_per_of_gp_rec > 100)) {
1598 			/* No PRR, lets just use the b/w estimate only */
1599 			return(100);
1600 		} else {
1601 			/*
1602 			 * Here we may have a non-retransmit but we
1603 			 * have no overrides, so just use the recovery
1604 			 * rate (prr is in effect).
1605 			 */
1606 			return(rack->r_ctl.rack_per_of_gp_rec);
1607 		}
1608 	}
1609 configured_rate:
1610 	/* For the configured rate we look at our cwnd vs the ssthresh */
1611 	if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
1612 		return (rack->r_ctl.rack_per_of_gp_ss);
1613 	else
1614 		return(rack->r_ctl.rack_per_of_gp_ca);
1615 }
1616 
1617 static uint64_t
rack_get_output_bw(struct tcp_rack * rack,uint64_t bw,struct rack_sendmap * rsm)1618 rack_get_output_bw(struct tcp_rack *rack, uint64_t bw, struct rack_sendmap *rsm)
1619 {
1620 	/*
1621 	 * We allow rack_per_of_gp_xx to dictate our bw rate we want.
1622 	 */
1623 	uint64_t bw_est;
1624 	uint64_t gain;
1625 
1626 	gain = (uint64_t)rack_get_output_gain(rack, rsm);
1627 	bw_est = bw * gain;
1628 	bw_est /= (uint64_t)100;
1629 	/* Never fall below the minimum (def 64kbps) */
1630 	if (bw_est < RACK_MIN_BW)
1631 		bw_est = RACK_MIN_BW;
1632 	return (bw_est);
1633 }
1634 
1635 static void
rack_log_retran_reason(struct tcp_rack * rack,struct rack_sendmap * rsm,uint32_t tsused,uint32_t thresh,int mod)1636 rack_log_retran_reason(struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t tsused, uint32_t thresh, int mod)
1637 {
1638 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1639 		union tcp_log_stackspecific log;
1640 		struct timeval tv;
1641 
1642 		if ((mod != 1) && (rack_verbose_logging == 0)) {
1643 			/*
1644 			 * We get 3 values currently for mod
1645 			 * 1 - We are retransmitting and this tells the reason.
1646 			 * 2 - We are clearing a dup-ack count.
1647 			 * 3 - We are incrementing a dup-ack count.
1648 			 *
1649 			 * The clear/increment are only logged
1650 			 * if you have BBverbose on.
1651 			 */
1652 			return;
1653 		}
1654 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
1655 		log.u_bbr.flex1 = tsused;
1656 		log.u_bbr.flex2 = thresh;
1657 		log.u_bbr.flex3 = rsm->r_flags;
1658 		log.u_bbr.flex4 = rsm->r_dupack;
1659 		log.u_bbr.flex5 = rsm->r_start;
1660 		log.u_bbr.flex6 = rsm->r_end;
1661 		log.u_bbr.flex8 = mod;
1662 		log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
1663 		log.u_bbr.ininput = rack->rc_inp->inp_in_input;
1664 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
1665 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
1666 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
1667 		    &rack->rc_inp->inp_socket->so_rcv,
1668 		    &rack->rc_inp->inp_socket->so_snd,
1669 		    BBR_LOG_SETTINGS_CHG, 0,
1670 		    0, &log, false, &tv);
1671 	}
1672 }
1673 
1674 static void
rack_log_to_start(struct tcp_rack * rack,uint32_t cts,uint32_t to,int32_t slot,uint8_t which)1675 rack_log_to_start(struct tcp_rack *rack, uint32_t cts, uint32_t to, int32_t slot, uint8_t which)
1676 {
1677 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1678 		union tcp_log_stackspecific log;
1679 		struct timeval tv;
1680 
1681 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
1682 		log.u_bbr.flex1 = TICKS_2_MSEC(rack->rc_tp->t_srtt >> TCP_RTT_SHIFT);
1683 		log.u_bbr.flex2 = to * 1000;
1684 		log.u_bbr.flex3 = rack->r_ctl.rc_hpts_flags;
1685 		log.u_bbr.flex4 = slot;
1686 		log.u_bbr.flex5 = rack->rc_inp->inp_hptsslot;
1687 		log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
1688 		log.u_bbr.flex7 = rack->rc_in_persist;
1689 		log.u_bbr.flex8 = which;
1690 		if (rack->rack_no_prr)
1691 			log.u_bbr.pkts_out = 0;
1692 		else
1693 			log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
1694 		log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
1695 		log.u_bbr.ininput = rack->rc_inp->inp_in_input;
1696 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
1697 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
1698 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
1699 		    &rack->rc_inp->inp_socket->so_rcv,
1700 		    &rack->rc_inp->inp_socket->so_snd,
1701 		    BBR_LOG_TIMERSTAR, 0,
1702 		    0, &log, false, &tv);
1703 	}
1704 }
1705 
1706 static void
rack_log_to_event(struct tcp_rack * rack,int32_t to_num,struct rack_sendmap * rsm)1707 rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm)
1708 {
1709 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1710 		union tcp_log_stackspecific log;
1711 		struct timeval tv;
1712 
1713 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
1714 		log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
1715 		log.u_bbr.ininput = rack->rc_inp->inp_in_input;
1716 		log.u_bbr.flex8 = to_num;
1717 		log.u_bbr.flex1 = rack->r_ctl.rc_rack_min_rtt;
1718 		log.u_bbr.flex2 = rack->rc_rack_rtt;
1719 		if (rsm == NULL)
1720 			log.u_bbr.flex3 = 0;
1721 		else
1722 			log.u_bbr.flex3 = rsm->r_end - rsm->r_start;
1723 		if (rack->rack_no_prr)
1724 			log.u_bbr.flex5 = 0;
1725 		else
1726 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
1727 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
1728 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
1729 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
1730 		    &rack->rc_inp->inp_socket->so_rcv,
1731 		    &rack->rc_inp->inp_socket->so_snd,
1732 		    BBR_LOG_RTO, 0,
1733 		    0, &log, false, &tv);
1734 	}
1735 }
1736 
1737 static void
rack_log_rtt_upd(struct tcpcb * tp,struct tcp_rack * rack,uint32_t t,uint32_t len,struct rack_sendmap * rsm,int conf)1738 rack_log_rtt_upd(struct tcpcb *tp, struct tcp_rack *rack, uint32_t t, uint32_t len,
1739 		 struct rack_sendmap *rsm, int conf)
1740 {
1741 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
1742 		union tcp_log_stackspecific log;
1743 		struct timeval tv;
1744 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
1745 		log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
1746 		log.u_bbr.ininput = rack->rc_inp->inp_in_input;
1747 		log.u_bbr.flex1 = t;
1748 		log.u_bbr.flex2 = len;
1749 		log.u_bbr.flex3 = rack->r_ctl.rc_rack_min_rtt * HPTS_USEC_IN_MSEC;
1750 		log.u_bbr.flex4 = rack->r_ctl.rack_rs.rs_rtt_lowest * HPTS_USEC_IN_MSEC;
1751 		log.u_bbr.flex5 = rack->r_ctl.rack_rs.rs_rtt_highest * HPTS_USEC_IN_MSEC;
1752 		log.u_bbr.flex6 = rack->r_ctl.rack_rs.rs_rtt_cnt;
1753 		log.u_bbr.flex7 = conf;
1754 		log.u_bbr.rttProp = (uint64_t)rack->r_ctl.rack_rs.rs_rtt_tot * (uint64_t)HPTS_USEC_IN_MSEC;
1755 		log.u_bbr.flex8 = rack->r_ctl.rc_rate_sample_method;
1756 		if (rack->rack_no_prr)
1757 			log.u_bbr.pkts_out = 0;
1758 		else
1759 			log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
1760 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
1761 		log.u_bbr.delivered = rack->r_ctl.rack_rs.rs_us_rtt;
1762 		log.u_bbr.pkts_out = rack->r_ctl.rack_rs.rs_flags;
1763 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
1764 		if (rsm) {
1765 			log.u_bbr.pkt_epoch = rsm->r_start;
1766 			log.u_bbr.lost = rsm->r_end;
1767 			log.u_bbr.cwnd_gain = rsm->r_rtr_cnt;
1768 		} else {
1769 			/* Its a SYN */
1770 			log.u_bbr.pkt_epoch = rack->rc_tp->iss;
1771 			log.u_bbr.lost = 0;
1772 			log.u_bbr.cwnd_gain = 0;
1773 		}
1774 		/* Write out general bits of interest rrs here */
1775 		log.u_bbr.use_lt_bw = rack->rc_highly_buffered;
1776 		log.u_bbr.use_lt_bw <<= 1;
1777 		log.u_bbr.use_lt_bw |= rack->forced_ack;
1778 		log.u_bbr.use_lt_bw <<= 1;
1779 		log.u_bbr.use_lt_bw |= rack->rc_gp_dyn_mul;
1780 		log.u_bbr.use_lt_bw <<= 1;
1781 		log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
1782 		log.u_bbr.use_lt_bw <<= 1;
1783 		log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
1784 		log.u_bbr.use_lt_bw <<= 1;
1785 		log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
1786 		log.u_bbr.use_lt_bw <<= 1;
1787 		log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
1788 		log.u_bbr.use_lt_bw <<= 1;
1789 		log.u_bbr.use_lt_bw |= rack->rc_dragged_bottom;
1790 		log.u_bbr.applimited = rack->r_ctl.rc_target_probertt_flight;
1791 		log.u_bbr.epoch = rack->r_ctl.rc_time_probertt_starts;
1792 		log.u_bbr.lt_epoch = rack->r_ctl.rc_time_probertt_entered;
1793 		log.u_bbr.cur_del_rate = rack->r_ctl.rc_lower_rtt_us_cts;
1794 		log.u_bbr.delRate = rack->r_ctl.rc_gp_srtt;
1795 		TCP_LOG_EVENTP(tp, NULL,
1796 		    &rack->rc_inp->inp_socket->so_rcv,
1797 		    &rack->rc_inp->inp_socket->so_snd,
1798 		    BBR_LOG_BBRRTT, 0,
1799 		    0, &log, false, &tv);
1800 	}
1801 }
1802 
1803 static void
rack_log_rtt_sample(struct tcp_rack * rack,uint32_t rtt)1804 rack_log_rtt_sample(struct tcp_rack *rack, uint32_t rtt)
1805 {
1806 	/*
1807 	 * Log the rtt sample we are
1808 	 * applying to the srtt algorithm in
1809 	 * useconds.
1810 	 */
1811 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1812 		union tcp_log_stackspecific log;
1813 		struct timeval tv;
1814 
1815 		/* Convert our ms to a microsecond */
1816 		memset(&log, 0, sizeof(log));
1817 		log.u_bbr.flex1 = rtt * 1000;
1818 		log.u_bbr.flex2 = rack->r_ctl.ack_count;
1819 		log.u_bbr.flex3 = rack->r_ctl.sack_count;
1820 		log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move;
1821 		log.u_bbr.flex5 = rack->r_ctl.sack_moved_extra;
1822 		log.u_bbr.flex8 = rack->sack_attack_disable;
1823 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
1824 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
1825 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
1826 		    &rack->rc_inp->inp_socket->so_rcv,
1827 		    &rack->rc_inp->inp_socket->so_snd,
1828 		    TCP_LOG_RTT, 0,
1829 		    0, &log, false, &tv);
1830 	}
1831 }
1832 
1833 static inline void
rack_log_progress_event(struct tcp_rack * rack,struct tcpcb * tp,uint32_t tick,int event,int line)1834 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick,  int event, int line)
1835 {
1836 	if (rack_verbose_logging && (tp->t_logstate != TCP_LOG_STATE_OFF)) {
1837 		union tcp_log_stackspecific log;
1838 		struct timeval tv;
1839 
1840 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
1841 		log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
1842 		log.u_bbr.ininput = rack->rc_inp->inp_in_input;
1843 		log.u_bbr.flex1 = line;
1844 		log.u_bbr.flex2 = tick;
1845 		log.u_bbr.flex3 = tp->t_maxunacktime;
1846 		log.u_bbr.flex4 = tp->t_acktime;
1847 		log.u_bbr.flex8 = event;
1848 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
1849 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
1850 		TCP_LOG_EVENTP(tp, NULL,
1851 		    &rack->rc_inp->inp_socket->so_rcv,
1852 		    &rack->rc_inp->inp_socket->so_snd,
1853 		    BBR_LOG_PROGRESS, 0,
1854 		    0, &log, false, &tv);
1855 	}
1856 }
1857 
1858 static void
rack_log_type_bbrsnd(struct tcp_rack * rack,uint32_t len,uint32_t slot,uint32_t cts,struct timeval * tv)1859 rack_log_type_bbrsnd(struct tcp_rack *rack, uint32_t len, uint32_t slot, uint32_t cts, struct timeval *tv)
1860 {
1861 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1862 		union tcp_log_stackspecific log;
1863 
1864 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
1865 		log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
1866 		log.u_bbr.ininput = rack->rc_inp->inp_in_input;
1867 		log.u_bbr.flex1 = slot;
1868 		if (rack->rack_no_prr)
1869 			log.u_bbr.flex2 = 0;
1870 		else
1871 			log.u_bbr.flex2 = rack->r_ctl.rc_prr_sndcnt;
1872 		log.u_bbr.flex7 = (0x0000ffff & rack->r_ctl.rc_hpts_flags);
1873 		log.u_bbr.flex8 = rack->rc_in_persist;
1874 		log.u_bbr.timeStamp = cts;
1875 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
1876 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
1877 		    &rack->rc_inp->inp_socket->so_rcv,
1878 		    &rack->rc_inp->inp_socket->so_snd,
1879 		    BBR_LOG_BBRSND, 0,
1880 		    0, &log, false, tv);
1881 	}
1882 }
1883 
1884 static void
rack_log_doseg_done(struct tcp_rack * rack,uint32_t cts,int32_t nxt_pkt,int32_t did_out,int way_out)1885 rack_log_doseg_done(struct tcp_rack *rack, uint32_t cts, int32_t nxt_pkt, int32_t did_out, int way_out)
1886 {
1887 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1888 		union tcp_log_stackspecific log;
1889 		struct timeval tv;
1890 
1891 		memset(&log, 0, sizeof(log));
1892 		log.u_bbr.flex1 = did_out;
1893 		log.u_bbr.flex2 = nxt_pkt;
1894 		log.u_bbr.flex3 = way_out;
1895 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
1896 		if (rack->rack_no_prr)
1897 			log.u_bbr.flex5 = 0;
1898 		else
1899 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
1900 		log.u_bbr.applimited = rack->r_ctl.rc_pace_min_segs;
1901 		log.u_bbr.flex7 = rack->r_wanted_output;
1902 		log.u_bbr.flex8 = rack->rc_in_persist;
1903 		log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
1904 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
1905 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
1906 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
1907 		    &rack->rc_inp->inp_socket->so_rcv,
1908 		    &rack->rc_inp->inp_socket->so_snd,
1909 		    BBR_LOG_DOSEG_DONE, 0,
1910 		    0, &log, false, &tv);
1911 	}
1912 }
1913 
1914 static void
rack_log_type_hrdwtso(struct tcpcb * tp,struct tcp_rack * rack,int len,int mod,int32_t orig_len,int frm)1915 rack_log_type_hrdwtso(struct tcpcb *tp, struct tcp_rack *rack, int len, int mod, int32_t orig_len, int frm)
1916 {
1917 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
1918 		union tcp_log_stackspecific log;
1919 		struct timeval tv;
1920 		uint32_t cts;
1921 
1922 		memset(&log, 0, sizeof(log));
1923 		cts = tcp_get_usecs(&tv);
1924 		log.u_bbr.flex1 = rack->r_ctl.rc_pace_min_segs;
1925 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
1926 		log.u_bbr.flex4 = len;
1927 		log.u_bbr.flex5 = orig_len;
1928 		log.u_bbr.flex6 = rack->r_ctl.rc_sacked;
1929 		log.u_bbr.flex7 = mod;
1930 		log.u_bbr.flex8 = frm;
1931 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
1932 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
1933 		TCP_LOG_EVENTP(tp, NULL,
1934 		    &tp->t_inpcb->inp_socket->so_rcv,
1935 		    &tp->t_inpcb->inp_socket->so_snd,
1936 		    TCP_HDWR_TLS, 0,
1937 		    0, &log, false, &tv);
1938 	}
1939 }
1940 
1941 static void
rack_log_type_just_return(struct tcp_rack * rack,uint32_t cts,uint32_t tlen,uint32_t slot,uint8_t hpts_calling,int reason,uint32_t cwnd_to_use)1942 rack_log_type_just_return(struct tcp_rack *rack, uint32_t cts, uint32_t tlen, uint32_t slot,
1943 			  uint8_t hpts_calling, int reason, uint32_t cwnd_to_use)
1944 {
1945 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1946 		union tcp_log_stackspecific log;
1947 		struct timeval tv;
1948 
1949 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
1950 		log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
1951 		log.u_bbr.ininput = rack->rc_inp->inp_in_input;
1952 		log.u_bbr.flex1 = slot;
1953 		log.u_bbr.flex2 = rack->r_ctl.rc_hpts_flags;
1954 		log.u_bbr.flex4 = reason;
1955 		if (rack->rack_no_prr)
1956 			log.u_bbr.flex5 = 0;
1957 		else
1958 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
1959 		log.u_bbr.flex7 = hpts_calling;
1960 		log.u_bbr.flex8 = rack->rc_in_persist;
1961 		log.u_bbr.lt_epoch = cwnd_to_use;
1962 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
1963 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
1964 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
1965 		    &rack->rc_inp->inp_socket->so_rcv,
1966 		    &rack->rc_inp->inp_socket->so_snd,
1967 		    BBR_LOG_JUSTRET, 0,
1968 		    tlen, &log, false, &tv);
1969 	}
1970 }
1971 
1972 static void
rack_log_to_cancel(struct tcp_rack * rack,int32_t hpts_removed,int line,uint32_t us_cts,struct timeval * tv,uint32_t flags_on_entry)1973 rack_log_to_cancel(struct tcp_rack *rack, int32_t hpts_removed, int line, uint32_t us_cts,
1974 		   struct timeval *tv, uint32_t flags_on_entry)
1975 {
1976 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1977 		union tcp_log_stackspecific log;
1978 
1979 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
1980 		log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
1981 		log.u_bbr.ininput = rack->rc_inp->inp_in_input;
1982 		log.u_bbr.flex1 = line;
1983 		log.u_bbr.flex2 = rack->r_ctl.rc_last_output_to;
1984 		log.u_bbr.flex3 = flags_on_entry;
1985 		log.u_bbr.flex4 = us_cts;
1986 		if (rack->rack_no_prr)
1987 			log.u_bbr.flex5 = 0;
1988 		else
1989 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
1990 		log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
1991 		log.u_bbr.flex7 = hpts_removed;
1992 		log.u_bbr.flex8 = 1;
1993 		log.u_bbr.applimited = rack->r_ctl.rc_hpts_flags;
1994 		log.u_bbr.timeStamp = us_cts;
1995 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
1996 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
1997 		    &rack->rc_inp->inp_socket->so_rcv,
1998 		    &rack->rc_inp->inp_socket->so_snd,
1999 		    BBR_LOG_TIMERCANC, 0,
2000 		    0, &log, false, tv);
2001 	}
2002 }
2003 
2004 static void
rack_log_alt_to_to_cancel(struct tcp_rack * rack,uint32_t flex1,uint32_t flex2,uint32_t flex3,uint32_t flex4,uint32_t flex5,uint32_t flex6,uint16_t flex7,uint8_t mod)2005 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
2006 			  uint32_t flex1, uint32_t flex2,
2007 			  uint32_t flex3, uint32_t flex4,
2008 			  uint32_t flex5, uint32_t flex6,
2009 			  uint16_t flex7, uint8_t mod)
2010 {
2011 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2012 		union tcp_log_stackspecific log;
2013 		struct timeval tv;
2014 
2015 		if (mod == 1) {
2016 			/* No you can't use 1, its for the real to cancel */
2017 			return;
2018 		}
2019 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2020 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2021 		log.u_bbr.flex1 = flex1;
2022 		log.u_bbr.flex2 = flex2;
2023 		log.u_bbr.flex3 = flex3;
2024 		log.u_bbr.flex4 = flex4;
2025 		log.u_bbr.flex5 = flex5;
2026 		log.u_bbr.flex6 = flex6;
2027 		log.u_bbr.flex7 = flex7;
2028 		log.u_bbr.flex8 =  mod;
2029 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2030 		    &rack->rc_inp->inp_socket->so_rcv,
2031 		    &rack->rc_inp->inp_socket->so_snd,
2032 		    BBR_LOG_TIMERCANC, 0,
2033 		    0, &log, false, &tv);
2034 	}
2035 }
2036 
2037 static void
rack_log_to_processing(struct tcp_rack * rack,uint32_t cts,int32_t ret,int32_t timers)2038 rack_log_to_processing(struct tcp_rack *rack, uint32_t cts, int32_t ret, int32_t timers)
2039 {
2040 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2041 		union tcp_log_stackspecific log;
2042 		struct timeval tv;
2043 
2044 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2045 		log.u_bbr.flex1 = timers;
2046 		log.u_bbr.flex2 = ret;
2047 		log.u_bbr.flex3 = rack->r_ctl.rc_timer_exp;
2048 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
2049 		log.u_bbr.flex5 = cts;
2050 		if (rack->rack_no_prr)
2051 			log.u_bbr.flex6 = 0;
2052 		else
2053 			log.u_bbr.flex6 = rack->r_ctl.rc_prr_sndcnt;
2054 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2055 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2056 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2057 		    &rack->rc_inp->inp_socket->so_rcv,
2058 		    &rack->rc_inp->inp_socket->so_snd,
2059 		    BBR_LOG_TO_PROCESS, 0,
2060 		    0, &log, false, &tv);
2061 	}
2062 }
2063 
2064 static void
rack_log_to_prr(struct tcp_rack * rack,int frm,int orig_cwnd)2065 rack_log_to_prr(struct tcp_rack *rack, int frm, int orig_cwnd)
2066 {
2067 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2068 		union tcp_log_stackspecific log;
2069 		struct timeval tv;
2070 
2071 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2072 		log.u_bbr.flex1 = rack->r_ctl.rc_prr_out;
2073 		log.u_bbr.flex2 = rack->r_ctl.rc_prr_recovery_fs;
2074 		if (rack->rack_no_prr)
2075 			log.u_bbr.flex3 = 0;
2076 		else
2077 			log.u_bbr.flex3 = rack->r_ctl.rc_prr_sndcnt;
2078 		log.u_bbr.flex4 = rack->r_ctl.rc_prr_delivered;
2079 		log.u_bbr.flex5 = rack->r_ctl.rc_sacked;
2080 		log.u_bbr.flex6 = rack->r_ctl.rc_holes_rxt;
2081 		log.u_bbr.flex8 = frm;
2082 		log.u_bbr.pkts_out = orig_cwnd;
2083 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2084 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2085 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2086 		    &rack->rc_inp->inp_socket->so_rcv,
2087 		    &rack->rc_inp->inp_socket->so_snd,
2088 		    BBR_LOG_BBRUPD, 0,
2089 		    0, &log, false, &tv);
2090 	}
2091 }
2092 
2093 #ifdef NETFLIX_EXP_DETECTION
2094 static void
rack_log_sad(struct tcp_rack * rack,int event)2095 rack_log_sad(struct tcp_rack *rack, int event)
2096 {
2097 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2098 		union tcp_log_stackspecific log;
2099 		struct timeval tv;
2100 
2101 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2102 		log.u_bbr.flex1 = rack->r_ctl.sack_count;
2103 		log.u_bbr.flex2 = rack->r_ctl.ack_count;
2104 		log.u_bbr.flex3 = rack->r_ctl.sack_moved_extra;
2105 		log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move;
2106 		log.u_bbr.flex5 = rack->r_ctl.rc_num_maps_alloced;
2107 		log.u_bbr.flex6 = tcp_sack_to_ack_thresh;
2108 		log.u_bbr.pkts_out = tcp_sack_to_move_thresh;
2109 		log.u_bbr.lt_epoch = (tcp_force_detection << 8);
2110 		log.u_bbr.lt_epoch |= rack->do_detection;
2111 		log.u_bbr.applimited = tcp_map_minimum;
2112 		log.u_bbr.flex7 = rack->sack_attack_disable;
2113 		log.u_bbr.flex8 = event;
2114 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2115 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2116 		log.u_bbr.delivered = tcp_sad_decay_val;
2117 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2118 		    &rack->rc_inp->inp_socket->so_rcv,
2119 		    &rack->rc_inp->inp_socket->so_snd,
2120 		    TCP_SAD_DETECTION, 0,
2121 		    0, &log, false, &tv);
2122 	}
2123 }
2124 #endif
2125 
2126 static void
rack_counter_destroy(void)2127 rack_counter_destroy(void)
2128 {
2129 	counter_u64_free(rack_ack_total);
2130 	counter_u64_free(rack_express_sack);
2131 	counter_u64_free(rack_sack_total);
2132 	counter_u64_free(rack_move_none);
2133 	counter_u64_free(rack_move_some);
2134 	counter_u64_free(rack_sack_attacks_detected);
2135 	counter_u64_free(rack_sack_attacks_reversed);
2136 	counter_u64_free(rack_sack_used_next_merge);
2137 	counter_u64_free(rack_sack_used_prev_merge);
2138 	counter_u64_free(rack_badfr);
2139 	counter_u64_free(rack_badfr_bytes);
2140 	counter_u64_free(rack_rtm_prr_retran);
2141 	counter_u64_free(rack_rtm_prr_newdata);
2142 	counter_u64_free(rack_timestamp_mismatch);
2143 	counter_u64_free(rack_find_high);
2144 	counter_u64_free(rack_reorder_seen);
2145 	counter_u64_free(rack_tlp_tot);
2146 	counter_u64_free(rack_tlp_newdata);
2147 	counter_u64_free(rack_tlp_retran);
2148 	counter_u64_free(rack_tlp_retran_bytes);
2149 	counter_u64_free(rack_tlp_retran_fail);
2150 	counter_u64_free(rack_to_tot);
2151 	counter_u64_free(rack_to_arm_rack);
2152 	counter_u64_free(rack_to_arm_tlp);
2153 	counter_u64_free(rack_calc_zero);
2154 	counter_u64_free(rack_calc_nonzero);
2155 	counter_u64_free(rack_paced_segments);
2156 	counter_u64_free(rack_unpaced_segments);
2157 	counter_u64_free(rack_saw_enobuf);
2158 	counter_u64_free(rack_saw_enetunreach);
2159 	counter_u64_free(rack_to_alloc);
2160 	counter_u64_free(rack_to_alloc_hard);
2161 	counter_u64_free(rack_to_alloc_emerg);
2162 	counter_u64_free(rack_to_alloc_limited);
2163 	counter_u64_free(rack_alloc_limited_conns);
2164 	counter_u64_free(rack_split_limited);
2165 	counter_u64_free(rack_sack_proc_all);
2166 	counter_u64_free(rack_sack_proc_restart);
2167 	counter_u64_free(rack_sack_proc_short);
2168 	counter_u64_free(rack_enter_tlp_calc);
2169 	counter_u64_free(rack_used_tlpmethod);
2170 	counter_u64_free(rack_used_tlpmethod2);
2171 	counter_u64_free(rack_sack_skipped_acked);
2172 	counter_u64_free(rack_sack_splits);
2173 	counter_u64_free(rack_progress_drops);
2174 	counter_u64_free(rack_input_idle_reduces);
2175 	counter_u64_free(rack_collapsed_win);
2176 	counter_u64_free(rack_tlp_does_nada);
2177 	counter_u64_free(rack_try_scwnd);
2178 	counter_u64_free(rack_per_timer_hole);
2179 	COUNTER_ARRAY_FREE(rack_out_size, TCP_MSS_ACCT_SIZE);
2180 	COUNTER_ARRAY_FREE(rack_opts_arry, RACK_OPTS_SIZE);
2181 }
2182 
2183 static struct rack_sendmap *
rack_alloc(struct tcp_rack * rack)2184 rack_alloc(struct tcp_rack *rack)
2185 {
2186 	struct rack_sendmap *rsm;
2187 
2188 	rsm = uma_zalloc(rack_zone, M_NOWAIT);
2189 	if (rsm) {
2190 		rack->r_ctl.rc_num_maps_alloced++;
2191 		counter_u64_add(rack_to_alloc, 1);
2192 		return (rsm);
2193 	}
2194 	if (rack->rc_free_cnt) {
2195 		counter_u64_add(rack_to_alloc_emerg, 1);
2196 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
2197 		TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
2198 		rack->rc_free_cnt--;
2199 		return (rsm);
2200 	}
2201 	return (NULL);
2202 }
2203 
2204 static struct rack_sendmap *
rack_alloc_full_limit(struct tcp_rack * rack)2205 rack_alloc_full_limit(struct tcp_rack *rack)
2206 {
2207 	if ((V_tcp_map_entries_limit > 0) &&
2208 	    (rack->do_detection == 0) &&
2209 	    (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
2210 		counter_u64_add(rack_to_alloc_limited, 1);
2211 		if (!rack->alloc_limit_reported) {
2212 			rack->alloc_limit_reported = 1;
2213 			counter_u64_add(rack_alloc_limited_conns, 1);
2214 		}
2215 		return (NULL);
2216 	}
2217 	return (rack_alloc(rack));
2218 }
2219 
2220 /* wrapper to allocate a sendmap entry, subject to a specific limit */
2221 static struct rack_sendmap *
rack_alloc_limit(struct tcp_rack * rack,uint8_t limit_type)2222 rack_alloc_limit(struct tcp_rack *rack, uint8_t limit_type)
2223 {
2224 	struct rack_sendmap *rsm;
2225 
2226 	if (limit_type) {
2227 		/* currently there is only one limit type */
2228 		if (V_tcp_map_split_limit > 0 &&
2229 		    (rack->do_detection == 0) &&
2230 		    rack->r_ctl.rc_num_split_allocs >= V_tcp_map_split_limit) {
2231 			counter_u64_add(rack_split_limited, 1);
2232 			if (!rack->alloc_limit_reported) {
2233 				rack->alloc_limit_reported = 1;
2234 				counter_u64_add(rack_alloc_limited_conns, 1);
2235 			}
2236 			return (NULL);
2237 		}
2238 	}
2239 
2240 	/* allocate and mark in the limit type, if set */
2241 	rsm = rack_alloc(rack);
2242 	if (rsm != NULL && limit_type) {
2243 		rsm->r_limit_type = limit_type;
2244 		rack->r_ctl.rc_num_split_allocs++;
2245 	}
2246 	return (rsm);
2247 }
2248 
2249 static void
rack_free(struct tcp_rack * rack,struct rack_sendmap * rsm)2250 rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm)
2251 {
2252 	if (rsm->r_flags & RACK_APP_LIMITED) {
2253 		if (rack->r_ctl.rc_app_limited_cnt > 0) {
2254 			rack->r_ctl.rc_app_limited_cnt--;
2255 		}
2256 	}
2257 	if (rsm->r_limit_type) {
2258 		/* currently there is only one limit type */
2259 		rack->r_ctl.rc_num_split_allocs--;
2260 	}
2261 	if (rsm == rack->r_ctl.rc_first_appl) {
2262 		if (rack->r_ctl.rc_app_limited_cnt == 0)
2263 			rack->r_ctl.rc_first_appl = NULL;
2264 		else {
2265 			/* Follow the next one out */
2266 			struct rack_sendmap fe;
2267 
2268 			fe.r_start = rsm->r_nseq_appl;
2269 			rack->r_ctl.rc_first_appl = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
2270 		}
2271 	}
2272 	if (rsm == rack->r_ctl.rc_resend)
2273 		rack->r_ctl.rc_resend = NULL;
2274 	if (rsm == rack->r_ctl.rc_rsm_at_retran)
2275 		rack->r_ctl.rc_rsm_at_retran = NULL;
2276 	if (rsm == rack->r_ctl.rc_end_appl)
2277 		rack->r_ctl.rc_end_appl = NULL;
2278 	if (rack->r_ctl.rc_tlpsend == rsm)
2279 		rack->r_ctl.rc_tlpsend = NULL;
2280 	if (rack->r_ctl.rc_sacklast == rsm)
2281 		rack->r_ctl.rc_sacklast = NULL;
2282 	if (rack->rc_free_cnt < rack_free_cache) {
2283 		memset(rsm, 0, sizeof(struct rack_sendmap));
2284 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_tnext);
2285 		rsm->r_limit_type = 0;
2286 		rack->rc_free_cnt++;
2287 		return;
2288 	}
2289 	rack->r_ctl.rc_num_maps_alloced--;
2290 	uma_zfree(rack_zone, rsm);
2291 }
2292 
2293 static uint32_t
rack_get_measure_window(struct tcpcb * tp,struct tcp_rack * rack)2294 rack_get_measure_window(struct tcpcb *tp, struct tcp_rack *rack)
2295 {
2296 	uint64_t srtt, bw, len, tim;
2297 	uint32_t segsiz, def_len, minl;
2298 
2299 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
2300 	def_len = rack_def_data_window * segsiz;
2301 	if (rack->rc_gp_filled == 0) {
2302 		/*
2303 		 * We have no measurement (IW is in flight?) so
2304 		 * we can only guess using our data_window sysctl
2305 		 * value (usually 100MSS).
2306 		 */
2307 		return (def_len);
2308 	}
2309 	/*
2310 	 * Now we have a number of factors to consider.
2311 	 *
2312 	 * 1) We have a desired BDP which is usually
2313 	 *    at least 2.
2314 	 * 2) We have a minimum number of rtt's usually 1 SRTT
2315 	 *    but we allow it too to be more.
2316 	 * 3) We want to make sure a measurement last N useconds (if
2317 	 *    we have set rack_min_measure_usec.
2318 	 *
2319 	 * We handle the first concern here by trying to create a data
2320 	 * window of max(rack_def_data_window, DesiredBDP). The
2321 	 * second concern we handle in not letting the measurement
2322 	 * window end normally until at least the required SRTT's
2323 	 * have gone by which is done further below in
2324 	 * rack_enough_for_measurement(). Finally the third concern
2325 	 * we also handle here by calculating how long that time
2326 	 * would take at the current BW and then return the
2327 	 * max of our first calculation and that length. Note
2328 	 * that if rack_min_measure_usec is 0, we don't deal
2329 	 * with concern 3. Also for both Concern 1 and 3 an
2330 	 * application limited period could end the measurement
2331 	 * earlier.
2332 	 *
2333 	 * So lets calculate the BDP with the "known" b/w using
2334 	 * the SRTT has our rtt and then multiply it by the
2335 	 * goal.
2336 	 */
2337 	bw = rack_get_bw(rack);
2338 	srtt = ((uint64_t)TICKS_2_USEC(tp->t_srtt) >> TCP_RTT_SHIFT);
2339 	len = bw * srtt;
2340 	len /= (uint64_t)HPTS_USEC_IN_SEC;
2341 	len *= max(1, rack_goal_bdp);
2342         /* Now we need to round up to the nearest MSS */
2343 	len = roundup(len, segsiz);
2344 	if (rack_min_measure_usec) {
2345 		/* Now calculate our min length for this b/w */
2346 		tim = rack_min_measure_usec;
2347 		minl = (tim * bw) / (uint64_t)HPTS_USEC_IN_SEC;
2348 		if (minl == 0)
2349 			minl = 1;
2350 		minl = roundup(minl, segsiz);
2351 		if (len < minl)
2352 			len = minl;
2353 	}
2354 	/*
2355 	 * Now if we have a very small window we want
2356 	 * to attempt to get the window that is
2357 	 * as small as possible. This happens on
2358 	 * low b/w connections and we don't want to
2359 	 * span huge numbers of rtt's between measurements.
2360 	 *
2361 	 * We basically include 2 over our "MIN window" so
2362 	 * that the measurement can be shortened (possibly) by
2363 	 * an ack'ed packet.
2364 	 */
2365 	if (len < def_len)
2366 		return (max((uint32_t)len, ((MIN_GP_WIN+2) * segsiz)));
2367 	else
2368 		return (max((uint32_t)len, def_len));
2369 
2370 }
2371 
2372 static int
rack_enough_for_measurement(struct tcpcb * tp,struct tcp_rack * rack,tcp_seq th_ack)2373 rack_enough_for_measurement(struct tcpcb *tp, struct tcp_rack *rack, tcp_seq th_ack)
2374 {
2375 	uint32_t tim, srtts, segsiz;
2376 
2377 	/*
2378 	 * Has enough time passed for the GP measurement to be valid?
2379 	 */
2380 	if ((tp->snd_max == tp->snd_una) ||
2381 	    (th_ack == tp->snd_max)){
2382 		/* All is acked */
2383 		return (1);
2384 	}
2385 	if (SEQ_LT(th_ack, tp->gput_seq)) {
2386 		/* Not enough bytes yet */
2387 		return (0);
2388 	}
2389 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
2390 	if (SEQ_LT(th_ack, tp->gput_ack) &&
2391 	    ((th_ack - tp->gput_seq) < max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
2392 		/* Not enough bytes yet */
2393 		return (0);
2394 	}
2395 	if (rack->r_ctl.rc_first_appl &&
2396 	    (rack->r_ctl.rc_first_appl->r_start == th_ack)) {
2397 		/*
2398 		 * We are up to the app limited point
2399 		 * we have to measure irrespective of the time..
2400 		 */
2401 		return (1);
2402 	}
2403 	/* Now what about time? */
2404 	srtts = (rack->r_ctl.rc_gp_srtt * rack_min_srtts);
2405 	tim = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - tp->gput_ts;
2406 	if (tim >= srtts) {
2407 		return (1);
2408 	}
2409 	/* Nope not even a full SRTT has passed */
2410 	return (0);
2411 }
2412 
2413 static void
rack_log_timely(struct tcp_rack * rack,uint32_t logged,uint64_t cur_bw,uint64_t low_bnd,uint64_t up_bnd,int line,uint8_t method)2414 rack_log_timely(struct tcp_rack *rack,
2415 		uint32_t logged, uint64_t cur_bw, uint64_t low_bnd,
2416 		uint64_t up_bnd, int line, uint8_t method)
2417 {
2418 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2419 		union tcp_log_stackspecific log;
2420 		struct timeval tv;
2421 
2422 		memset(&log, 0, sizeof(log));
2423 		log.u_bbr.flex1 = logged;
2424 		log.u_bbr.flex2 = rack->rc_gp_timely_inc_cnt;
2425 		log.u_bbr.flex2 <<= 4;
2426 		log.u_bbr.flex2 |= rack->rc_gp_timely_dec_cnt;
2427 		log.u_bbr.flex2 <<= 4;
2428 		log.u_bbr.flex2 |= rack->rc_gp_incr;
2429 		log.u_bbr.flex2 <<= 4;
2430 		log.u_bbr.flex2 |= rack->rc_gp_bwred;
2431 		log.u_bbr.flex3 = rack->rc_gp_incr;
2432 		log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
2433 		log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ca;
2434 		log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_rec;
2435 		log.u_bbr.flex7 = rack->rc_gp_bwred;
2436 		log.u_bbr.flex8 = method;
2437 		log.u_bbr.cur_del_rate = cur_bw;
2438 		log.u_bbr.delRate = low_bnd;
2439 		log.u_bbr.bw_inuse = up_bnd;
2440 		log.u_bbr.rttProp = rack_get_bw(rack);
2441 		log.u_bbr.pkt_epoch = line;
2442 		log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
2443 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2444 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2445 		log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
2446 		log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
2447 		log.u_bbr.cwnd_gain = rack->rc_dragged_bottom;
2448 		log.u_bbr.cwnd_gain <<= 1;
2449 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_rec;
2450 		log.u_bbr.cwnd_gain <<= 1;
2451 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
2452 		log.u_bbr.cwnd_gain <<= 1;
2453 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
2454 		log.u_bbr.lost = rack->r_ctl.rc_loss_count;
2455 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2456 		    &rack->rc_inp->inp_socket->so_rcv,
2457 		    &rack->rc_inp->inp_socket->so_snd,
2458 		    TCP_TIMELY_WORK, 0,
2459 		    0, &log, false, &tv);
2460 	}
2461 }
2462 
2463 static int
rack_bw_can_be_raised(struct tcp_rack * rack,uint64_t cur_bw,uint64_t last_bw_est,uint16_t mult)2464 rack_bw_can_be_raised(struct tcp_rack *rack, uint64_t cur_bw, uint64_t last_bw_est, uint16_t mult)
2465 {
2466 	/*
2467 	 * Before we increase we need to know if
2468 	 * the estimate just made was less than
2469 	 * our pacing goal (i.e. (cur_bw * mult) > last_bw_est)
2470 	 *
2471 	 * If we already are pacing at a fast enough
2472 	 * rate to push us faster there is no sense of
2473 	 * increasing.
2474 	 *
2475 	 * We first caculate our actual pacing rate (ss or ca multipler
2476 	 * times our cur_bw).
2477 	 *
2478 	 * Then we take the last measured rate and multipy by our
2479 	 * maximum pacing overage to give us a max allowable rate.
2480 	 *
2481 	 * If our act_rate is smaller than our max_allowable rate
2482 	 * then we should increase. Else we should hold steady.
2483 	 *
2484 	 */
2485 	uint64_t act_rate, max_allow_rate;
2486 
2487 	if (rack_timely_no_stopping)
2488 		return (1);
2489 
2490 	if ((cur_bw == 0) || (last_bw_est == 0)) {
2491 		/*
2492 		 * Initial startup case or
2493 		 * everything is acked case.
2494 		 */
2495 		rack_log_timely(rack,  mult, cur_bw, 0, 0,
2496 				__LINE__, 9);
2497 		return (1);
2498 	}
2499 	if (mult <= 100) {
2500 		/*
2501 		 * We can always pace at or slightly above our rate.
2502 		 */
2503 		rack_log_timely(rack,  mult, cur_bw, 0, 0,
2504 				__LINE__, 9);
2505 		return (1);
2506 	}
2507 	act_rate = cur_bw * (uint64_t)mult;
2508 	act_rate /= 100;
2509 	max_allow_rate = last_bw_est * ((uint64_t)rack_max_per_above + (uint64_t)100);
2510 	max_allow_rate /= 100;
2511 	if (act_rate < max_allow_rate) {
2512 		/*
2513 		 * Here the rate we are actually pacing at
2514 		 * is smaller than 10% above our last measurement.
2515 		 * This means we are pacing below what we would
2516 		 * like to try to achieve (plus some wiggle room).
2517 		 */
2518 		rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
2519 				__LINE__, 9);
2520 		return (1);
2521 	} else {
2522 		/*
2523 		 * Here we are already pacing at least rack_max_per_above(10%)
2524 		 * what we are getting back. This indicates most likely
2525 		 * that we are being limited (cwnd/rwnd/app) and can't
2526 		 * get any more b/w. There is no sense of trying to
2527 		 * raise up the pacing rate its not speeding us up
2528 		 * and we already are pacing faster than we are getting.
2529 		 */
2530 		rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
2531 				__LINE__, 8);
2532 		return (0);
2533 	}
2534 }
2535 
2536 static void
rack_validate_multipliers_at_or_above100(struct tcp_rack * rack)2537 rack_validate_multipliers_at_or_above100(struct tcp_rack *rack)
2538 {
2539 	/*
2540 	 * When we drag bottom, we want to assure
2541 	 * that no multiplier is below 1.0, if so
2542 	 * we want to restore it to at least that.
2543 	 */
2544 	if (rack->r_ctl.rack_per_of_gp_rec  < 100) {
2545 		/* This is unlikely we usually do not touch recovery */
2546 		rack->r_ctl.rack_per_of_gp_rec = 100;
2547 	}
2548 	if (rack->r_ctl.rack_per_of_gp_ca < 100) {
2549 		rack->r_ctl.rack_per_of_gp_ca = 100;
2550 	}
2551 	if (rack->r_ctl.rack_per_of_gp_ss < 100) {
2552 		rack->r_ctl.rack_per_of_gp_ss = 100;
2553 	}
2554 }
2555 
2556 static void
rack_validate_multipliers_at_or_below_100(struct tcp_rack * rack)2557 rack_validate_multipliers_at_or_below_100(struct tcp_rack *rack)
2558 {
2559 	if (rack->r_ctl.rack_per_of_gp_ca > 100) {
2560 		rack->r_ctl.rack_per_of_gp_ca = 100;
2561 	}
2562 	if (rack->r_ctl.rack_per_of_gp_ss > 100) {
2563 		rack->r_ctl.rack_per_of_gp_ss = 100;
2564 	}
2565 }
2566 
2567 static void
rack_increase_bw_mul(struct tcp_rack * rack,int timely_says,uint64_t cur_bw,uint64_t last_bw_est,int override)2568 rack_increase_bw_mul(struct tcp_rack *rack, int timely_says, uint64_t cur_bw, uint64_t last_bw_est, int override)
2569 {
2570 	int32_t  calc, logged, plus;
2571 
2572 	logged = 0;
2573 
2574 	if (override) {
2575 		/*
2576 		 * override is passed when we are
2577 		 * loosing b/w and making one last
2578 		 * gasp at trying to not loose out
2579 		 * to a new-reno flow.
2580 		 */
2581 		goto extra_boost;
2582 	}
2583 	/* In classic timely we boost by 5x if we have 5 increases in a row, lets not */
2584 	if (rack->rc_gp_incr &&
2585 	    ((rack->rc_gp_timely_inc_cnt + 1) >= RACK_TIMELY_CNT_BOOST)) {
2586 		/*
2587 		 * Reset and get 5 strokes more before the boost. Note
2588 		 * that the count is 0 based so we have to add one.
2589 		 */
2590 extra_boost:
2591 		plus = (uint32_t)rack_gp_increase_per * RACK_TIMELY_CNT_BOOST;
2592 		rack->rc_gp_timely_inc_cnt = 0;
2593 	} else
2594 		plus = (uint32_t)rack_gp_increase_per;
2595 	/* Must be at least 1% increase for true timely increases */
2596 	if ((plus < 1) &&
2597 	    ((rack->r_ctl.rc_rtt_diff <= 0) || (timely_says <= 0)))
2598 		plus = 1;
2599 	if (rack->rc_gp_saw_rec &&
2600 	    (rack->rc_gp_no_rec_chg == 0) &&
2601 	    rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
2602 				  rack->r_ctl.rack_per_of_gp_rec)) {
2603 		/* We have been in recovery ding it too */
2604 		calc = rack->r_ctl.rack_per_of_gp_rec + plus;
2605 		if (calc > 0xffff)
2606 			calc = 0xffff;
2607 		logged |= 1;
2608 		rack->r_ctl.rack_per_of_gp_rec = (uint16_t)calc;
2609 		if (rack_per_upper_bound_ss &&
2610 		    (rack->rc_dragged_bottom == 0) &&
2611 		    (rack->r_ctl.rack_per_of_gp_rec > rack_per_upper_bound_ss))
2612 			rack->r_ctl.rack_per_of_gp_rec = rack_per_upper_bound_ss;
2613 	}
2614 	if (rack->rc_gp_saw_ca &&
2615 	    (rack->rc_gp_saw_ss == 0) &&
2616 	    rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
2617 				  rack->r_ctl.rack_per_of_gp_ca)) {
2618 		/* In CA */
2619 		calc = rack->r_ctl.rack_per_of_gp_ca + plus;
2620 		if (calc > 0xffff)
2621 			calc = 0xffff;
2622 		logged |= 2;
2623 		rack->r_ctl.rack_per_of_gp_ca = (uint16_t)calc;
2624 		if (rack_per_upper_bound_ca &&
2625 		    (rack->rc_dragged_bottom == 0) &&
2626 		    (rack->r_ctl.rack_per_of_gp_ca > rack_per_upper_bound_ca))
2627 			rack->r_ctl.rack_per_of_gp_ca = rack_per_upper_bound_ca;
2628 	}
2629 	if (rack->rc_gp_saw_ss &&
2630 	    rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
2631 				  rack->r_ctl.rack_per_of_gp_ss)) {
2632 		/* In SS */
2633 		calc = rack->r_ctl.rack_per_of_gp_ss + plus;
2634 		if (calc > 0xffff)
2635 			calc = 0xffff;
2636 		rack->r_ctl.rack_per_of_gp_ss = (uint16_t)calc;
2637 		if (rack_per_upper_bound_ss &&
2638 		    (rack->rc_dragged_bottom == 0) &&
2639 		    (rack->r_ctl.rack_per_of_gp_ss > rack_per_upper_bound_ss))
2640 			rack->r_ctl.rack_per_of_gp_ss = rack_per_upper_bound_ss;
2641 		logged |= 4;
2642 	}
2643 	if (logged &&
2644 	    (rack->rc_gp_incr == 0)){
2645 		/* Go into increment mode */
2646 		rack->rc_gp_incr = 1;
2647 		rack->rc_gp_timely_inc_cnt = 0;
2648 	}
2649 	if (rack->rc_gp_incr &&
2650 	    logged &&
2651 	    (rack->rc_gp_timely_inc_cnt < RACK_TIMELY_CNT_BOOST)) {
2652 		rack->rc_gp_timely_inc_cnt++;
2653 	}
2654 	rack_log_timely(rack,  logged, plus, 0, 0,
2655 			__LINE__, 1);
2656 }
2657 
2658 static uint32_t
rack_get_decrease(struct tcp_rack * rack,uint32_t curper,int32_t rtt_diff)2659 rack_get_decrease(struct tcp_rack *rack, uint32_t curper, int32_t rtt_diff)
2660 {
2661 	/*
2662 	 * norm_grad = rtt_diff / minrtt;
2663 	 * new_per = curper  * (1 - B * norm_grad)
2664 	 *
2665 	 * B = rack_gp_decrease_per (default 10%)
2666 	 * rtt_dif = input var current rtt-diff
2667 	 * curper = input var current percentage
2668 	 * minrtt = from rack filter
2669 	 *
2670 	 */
2671 	uint64_t perf;
2672 
2673 	perf = (((uint64_t)curper * ((uint64_t)1000000 -
2674 		    ((uint64_t)rack_gp_decrease_per * (uint64_t)10000 *
2675 		     (((uint64_t)rtt_diff * (uint64_t)1000000)/
2676 		      (uint64_t)get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt)))/
2677 		     (uint64_t)1000000)) /
2678 		(uint64_t)1000000);
2679 	if (perf > curper) {
2680 		/* TSNH */
2681 		perf = curper - 1;
2682 	}
2683 	return ((uint32_t)perf);
2684 }
2685 
2686 static uint32_t
rack_decrease_highrtt(struct tcp_rack * rack,uint32_t curper,uint32_t rtt)2687 rack_decrease_highrtt(struct tcp_rack *rack, uint32_t curper, uint32_t rtt)
2688 {
2689 	/*
2690 	 *                                   highrttthresh
2691 	 * result = curper * (1 - (B * ( 1 -  ------          ))
2692 	 *                                     gp_srtt
2693 	 *
2694 	 * B = rack_gp_decrease_per (default 10%)
2695 	 * highrttthresh = filter_min * rack_gp_rtt_maxmul
2696 	 */
2697 	uint64_t perf;
2698 	uint32_t highrttthresh;
2699 
2700 	highrttthresh = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
2701 
2702 	perf =  (((uint64_t)curper * ((uint64_t)1000000 -
2703 				    ((uint64_t)rack_gp_decrease_per * ((uint64_t)1000000 -
2704 					((uint64_t)highrttthresh * (uint64_t)1000000) /
2705 						    (uint64_t)rtt)) / 100)) /(uint64_t)1000000);
2706 	return (perf);
2707 }
2708 
2709 static void
rack_decrease_bw_mul(struct tcp_rack * rack,int timely_says,uint32_t rtt,int32_t rtt_diff)2710 rack_decrease_bw_mul(struct tcp_rack *rack, int timely_says, uint32_t rtt, int32_t rtt_diff)
2711 {
2712 	uint64_t logvar, logvar2, logvar3;
2713 	uint32_t logged, new_per, ss_red, ca_red, rec_red, alt, val;
2714 
2715 	if (rack->rc_gp_incr) {
2716 		/* Turn off increment counting  */
2717 		rack->rc_gp_incr = 0;
2718 		rack->rc_gp_timely_inc_cnt = 0;
2719 	}
2720 	ss_red = ca_red = rec_red = 0;
2721 	logged = 0;
2722 	/* Calculate the reduction value */
2723 	if (rtt_diff < 0) {
2724 		rtt_diff *= -1;
2725 	}
2726 	/* Must be at least 1% reduction */
2727 	if (rack->rc_gp_saw_rec && (rack->rc_gp_no_rec_chg == 0)) {
2728 		/* We have been in recovery ding it too */
2729 		if (timely_says == 2) {
2730 			new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_rec, rtt);
2731 			alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
2732 			if (alt < new_per)
2733 				val = alt;
2734 			else
2735 				val = new_per;
2736 		} else
2737 			 val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
2738 		if (rack->r_ctl.rack_per_of_gp_rec > val) {
2739 			rec_red = (rack->r_ctl.rack_per_of_gp_rec - val);
2740 			rack->r_ctl.rack_per_of_gp_rec = (uint16_t)val;
2741 		} else {
2742 			rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
2743 			rec_red = 0;
2744 		}
2745 		if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_rec)
2746 			rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
2747 		logged |= 1;
2748 	}
2749 	if (rack->rc_gp_saw_ss) {
2750 		/* Sent in SS */
2751 		if (timely_says == 2) {
2752 			new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ss, rtt);
2753 			alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
2754 			if (alt < new_per)
2755 				val = alt;
2756 			else
2757 				val = new_per;
2758 		} else
2759 			val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ss, rtt_diff);
2760 		if (rack->r_ctl.rack_per_of_gp_ss > new_per) {
2761 			ss_red = rack->r_ctl.rack_per_of_gp_ss - val;
2762 			rack->r_ctl.rack_per_of_gp_ss = (uint16_t)val;
2763 		} else {
2764 			ss_red = new_per;
2765 			rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
2766 			logvar = new_per;
2767 			logvar <<= 32;
2768 			logvar |= alt;
2769 			logvar2 = (uint32_t)rtt;
2770 			logvar2 <<= 32;
2771 			logvar2 |= (uint32_t)rtt_diff;
2772 			logvar3 = rack_gp_rtt_maxmul;
2773 			logvar3 <<= 32;
2774 			logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
2775 			rack_log_timely(rack, timely_says,
2776 					logvar2, logvar3,
2777 					logvar, __LINE__, 10);
2778 		}
2779 		if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ss)
2780 			rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
2781 		logged |= 4;
2782 	} else 	if (rack->rc_gp_saw_ca) {
2783 		/* Sent in CA */
2784 		if (timely_says == 2) {
2785 			new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ca, rtt);
2786 			alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
2787 			if (alt < new_per)
2788 				val = alt;
2789 			else
2790 				val = new_per;
2791 		} else
2792 			val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ca, rtt_diff);
2793 		if (rack->r_ctl.rack_per_of_gp_ca > val) {
2794 			ca_red = rack->r_ctl.rack_per_of_gp_ca - val;
2795 			rack->r_ctl.rack_per_of_gp_ca = (uint16_t)val;
2796 		} else {
2797 			rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
2798 			ca_red = 0;
2799 			logvar = new_per;
2800 			logvar <<= 32;
2801 			logvar |= alt;
2802 			logvar2 = (uint32_t)rtt;
2803 			logvar2 <<= 32;
2804 			logvar2 |= (uint32_t)rtt_diff;
2805 			logvar3 = rack_gp_rtt_maxmul;
2806 			logvar3 <<= 32;
2807 			logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
2808 			rack_log_timely(rack, timely_says,
2809 					logvar2, logvar3,
2810 					logvar, __LINE__, 10);
2811 		}
2812 		if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ca)
2813 			rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
2814 		logged |= 2;
2815 	}
2816 	if (rack->rc_gp_timely_dec_cnt < 0x7) {
2817 		rack->rc_gp_timely_dec_cnt++;
2818 		if (rack_timely_dec_clear &&
2819 		    (rack->rc_gp_timely_dec_cnt == rack_timely_dec_clear))
2820 			rack->rc_gp_timely_dec_cnt = 0;
2821 	}
2822 	logvar = ss_red;
2823 	logvar <<= 32;
2824 	logvar |= ca_red;
2825 	rack_log_timely(rack,  logged, rec_red, rack_per_lower_bound, logvar,
2826 			__LINE__, 2);
2827 }
2828 
2829 static void
rack_log_rtt_shrinks(struct tcp_rack * rack,uint32_t us_cts,uint32_t rtt,uint32_t line,uint8_t reas)2830 rack_log_rtt_shrinks(struct tcp_rack *rack, uint32_t us_cts,
2831 		     uint32_t rtt, uint32_t line, uint8_t reas)
2832 {
2833 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2834 		union tcp_log_stackspecific log;
2835 		struct timeval tv;
2836 
2837 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2838 		log.u_bbr.flex1 = line;
2839 		log.u_bbr.flex2 = rack->r_ctl.rc_time_probertt_starts;
2840 		log.u_bbr.flex3 = rack->r_ctl.rc_lower_rtt_us_cts;
2841 		log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
2842 		log.u_bbr.flex5 = rtt;
2843 		log.u_bbr.flex6 = rack->rc_highly_buffered;
2844 		log.u_bbr.flex6 <<= 1;
2845 		log.u_bbr.flex6 |= rack->forced_ack;
2846 		log.u_bbr.flex6 <<= 1;
2847 		log.u_bbr.flex6 |= rack->rc_gp_dyn_mul;
2848 		log.u_bbr.flex6 <<= 1;
2849 		log.u_bbr.flex6 |= rack->in_probe_rtt;
2850 		log.u_bbr.flex6 <<= 1;
2851 		log.u_bbr.flex6 |= rack->measure_saw_probe_rtt;
2852 		log.u_bbr.flex7 = rack->r_ctl.rack_per_of_gp_probertt;
2853 		log.u_bbr.pacing_gain = rack->r_ctl.rack_per_of_gp_ca;
2854 		log.u_bbr.cwnd_gain = rack->r_ctl.rack_per_of_gp_rec;
2855 		log.u_bbr.flex8 = reas;
2856 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2857 		log.u_bbr.delRate = rack_get_bw(rack);
2858 		log.u_bbr.cur_del_rate = rack->r_ctl.rc_highest_us_rtt;
2859 		log.u_bbr.cur_del_rate <<= 32;
2860 		log.u_bbr.cur_del_rate |= rack->r_ctl.rc_lowest_us_rtt;
2861 		log.u_bbr.applimited = rack->r_ctl.rc_time_probertt_entered;
2862 		log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
2863 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2864 		log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
2865 		log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
2866 		log.u_bbr.pkt_epoch = rack->r_ctl.rc_lower_rtt_us_cts;
2867 		log.u_bbr.delivered = rack->r_ctl.rc_target_probertt_flight;
2868 		log.u_bbr.lost = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
2869 		log.u_bbr.rttProp = us_cts;
2870 		log.u_bbr.rttProp <<= 32;
2871 		log.u_bbr.rttProp |= rack->r_ctl.rc_entry_gp_rtt;
2872 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2873 		    &rack->rc_inp->inp_socket->so_rcv,
2874 		    &rack->rc_inp->inp_socket->so_snd,
2875 		    BBR_LOG_RTT_SHRINKS, 0,
2876 		    0, &log, false, &rack->r_ctl.act_rcv_time);
2877 	}
2878 }
2879 
2880 static void
rack_set_prtt_target(struct tcp_rack * rack,uint32_t segsiz,uint32_t rtt)2881 rack_set_prtt_target(struct tcp_rack *rack, uint32_t segsiz, uint32_t rtt)
2882 {
2883 	uint64_t bwdp;
2884 
2885 	bwdp = rack_get_bw(rack);
2886 	bwdp *= (uint64_t)rtt;
2887 	bwdp /= (uint64_t)HPTS_USEC_IN_SEC;
2888 	rack->r_ctl.rc_target_probertt_flight = roundup((uint32_t)bwdp, segsiz);
2889 	if (rack->r_ctl.rc_target_probertt_flight < (segsiz * rack_timely_min_segs)) {
2890 		/*
2891 		 * A window protocol must be able to have 4 packets
2892 		 * outstanding as the floor in order to function
2893 		 * (especially considering delayed ack :D).
2894 		 */
2895 		rack->r_ctl.rc_target_probertt_flight = (segsiz * rack_timely_min_segs);
2896 	}
2897 }
2898 
2899 static void
rack_enter_probertt(struct tcp_rack * rack,uint32_t us_cts)2900 rack_enter_probertt(struct tcp_rack *rack, uint32_t us_cts)
2901 {
2902 	/**
2903 	 * ProbeRTT is a bit different in rack_pacing than in
2904 	 * BBR. It is like BBR in that it uses the lowering of
2905 	 * the RTT as a signal that we saw something new and
2906 	 * counts from there for how long between. But it is
2907 	 * different in that its quite simple. It does not
2908 	 * play with the cwnd and wait until we get down
2909 	 * to N segments outstanding and hold that for
2910 	 * 200ms. Instead it just sets the pacing reduction
2911 	 * rate to a set percentage (70 by default) and hold
2912 	 * that for a number of recent GP Srtt's.
2913 	 */
2914 	uint32_t segsiz;
2915 
2916 	if (rack->rc_gp_dyn_mul == 0)
2917 		return;
2918 
2919 	if (rack->rc_tp->snd_max == rack->rc_tp->snd_una) {
2920 		/* We are idle */
2921 		return;
2922 	}
2923 	if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
2924 	    SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
2925 		/*
2926 		 * Stop the goodput now, the idea here is
2927 		 * that future measurements with in_probe_rtt
2928 		 * won't register if they are not greater so
2929 		 * we want to get what info (if any) is available
2930 		 * now.
2931 		 */
2932 		rack_do_goodput_measurement(rack->rc_tp, rack,
2933 					    rack->rc_tp->snd_una, __LINE__);
2934 	}
2935 	rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
2936 	rack->r_ctl.rc_time_probertt_entered = us_cts;
2937 	segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
2938 		     rack->r_ctl.rc_pace_min_segs);
2939 	rack->in_probe_rtt = 1;
2940 	rack->measure_saw_probe_rtt = 1;
2941 	rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
2942 	rack->r_ctl.rc_time_probertt_starts = 0;
2943 	rack->r_ctl.rc_entry_gp_rtt = rack->r_ctl.rc_gp_srtt;
2944 	if (rack_probertt_use_min_rtt_entry)
2945 		rack_set_prtt_target(rack, segsiz, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
2946 	else
2947 		rack_set_prtt_target(rack, segsiz, rack->r_ctl.rc_gp_srtt);
2948 	rack_log_rtt_shrinks(rack,  us_cts,  get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
2949 			     __LINE__, RACK_RTTS_ENTERPROBE);
2950 }
2951 
2952 static void
rack_exit_probertt(struct tcp_rack * rack,uint32_t us_cts)2953 rack_exit_probertt(struct tcp_rack *rack, uint32_t us_cts)
2954 {
2955 	struct rack_sendmap *rsm;
2956 	uint32_t segsiz;
2957 
2958 	segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
2959 		     rack->r_ctl.rc_pace_min_segs);
2960 	rack->in_probe_rtt = 0;
2961 	if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
2962 	    SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
2963 		/*
2964 		 * Stop the goodput now, the idea here is
2965 		 * that future measurements with in_probe_rtt
2966 		 * won't register if they are not greater so
2967 		 * we want to get what info (if any) is available
2968 		 * now.
2969 		 */
2970 		rack_do_goodput_measurement(rack->rc_tp, rack,
2971 					    rack->rc_tp->snd_una, __LINE__);
2972 	} else if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
2973 		/*
2974 		 * We don't have enough data to make a measurement.
2975 		 * So lets just stop and start here after exiting
2976 		 * probe-rtt. We probably are not interested in
2977 		 * the results anyway.
2978 		 */
2979 		rack->rc_tp->t_flags &= ~TF_GPUTINPROG;
2980 	}
2981 	/*
2982 	 * Measurements through the current snd_max are going
2983 	 * to be limited by the slower pacing rate.
2984 	 *
2985 	 * We need to mark these as app-limited so we
2986 	 * don't collapse the b/w.
2987 	 */
2988 	rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
2989 	if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
2990 		if (rack->r_ctl.rc_app_limited_cnt == 0)
2991 			rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
2992 		else {
2993 			/*
2994 			 * Go out to the end app limited and mark
2995 			 * this new one as next and move the end_appl up
2996 			 * to this guy.
2997 			 */
2998 			if (rack->r_ctl.rc_end_appl)
2999 				rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
3000 			rack->r_ctl.rc_end_appl = rsm;
3001 		}
3002 		rsm->r_flags |= RACK_APP_LIMITED;
3003 		rack->r_ctl.rc_app_limited_cnt++;
3004 	}
3005 	/*
3006 	 * Now, we need to examine our pacing rate multipliers.
3007 	 * If its under 100%, we need to kick it back up to
3008 	 * 100%. We also don't let it be over our "max" above
3009 	 * the actual rate i.e. 100% + rack_clamp_atexit_prtt.
3010 	 * Note setting clamp_atexit_prtt to 0 has the effect
3011 	 * of setting CA/SS to 100% always at exit (which is
3012 	 * the default behavior).
3013 	 */
3014 	if (rack_probertt_clear_is) {
3015 		rack->rc_gp_incr = 0;
3016 		rack->rc_gp_bwred = 0;
3017 		rack->rc_gp_timely_inc_cnt = 0;
3018 		rack->rc_gp_timely_dec_cnt = 0;
3019 	}
3020 	/* Do we do any clamping at exit? */
3021 	if (rack->rc_highly_buffered && rack_atexit_prtt_hbp) {
3022 		rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt_hbp;
3023 		rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt_hbp;
3024 	}
3025 	if ((rack->rc_highly_buffered == 0) && rack_atexit_prtt) {
3026 		rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt;
3027 		rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt;
3028 	}
3029 	/*
3030 	 * Lets set rtt_diff to 0, so that we will get a "boost"
3031 	 * after exiting.
3032 	 */
3033 	rack->r_ctl.rc_rtt_diff = 0;
3034 
3035 	/* Clear all flags so we start fresh */
3036 	rack->rc_tp->t_bytes_acked = 0;
3037 	rack->rc_tp->ccv->flags &= ~CCF_ABC_SENTAWND;
3038 	/*
3039 	 * If configured to, set the cwnd and ssthresh to
3040 	 * our targets.
3041 	 */
3042 	if (rack_probe_rtt_sets_cwnd) {
3043 		uint64_t ebdp;
3044 		uint32_t setto;
3045 
3046 		/* Set ssthresh so we get into CA once we hit our target */
3047 		if (rack_probertt_use_min_rtt_exit == 1) {
3048 			/* Set to min rtt */
3049 			rack_set_prtt_target(rack, segsiz,
3050 					     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
3051 		} else if (rack_probertt_use_min_rtt_exit == 2) {
3052 			/* Set to current gp rtt */
3053 			rack_set_prtt_target(rack, segsiz,
3054 					     rack->r_ctl.rc_gp_srtt);
3055 		} else if (rack_probertt_use_min_rtt_exit == 3) {
3056 			/* Set to entry gp rtt */
3057 			rack_set_prtt_target(rack, segsiz,
3058 					     rack->r_ctl.rc_entry_gp_rtt);
3059 		} else  {
3060 			uint64_t sum;
3061 			uint32_t setval;
3062 
3063 			sum = rack->r_ctl.rc_entry_gp_rtt;
3064 			sum *= 10;
3065 			sum /= (uint64_t)(max(1, rack->r_ctl.rc_gp_srtt));
3066 			if (sum >= 20) {
3067 				/*
3068 				 * A highly buffered path needs
3069 				 * cwnd space for timely to work.
3070 				 * Lets set things up as if
3071 				 * we are heading back here again.
3072 				 */
3073 				setval = rack->r_ctl.rc_entry_gp_rtt;
3074 			} else if (sum >= 15) {
3075 				/*
3076 				 * Lets take the smaller of the
3077 				 * two since we are just somewhat
3078 				 * buffered.
3079 				 */
3080 				setval = rack->r_ctl.rc_gp_srtt;
3081 				if (setval > rack->r_ctl.rc_entry_gp_rtt)
3082 					setval = rack->r_ctl.rc_entry_gp_rtt;
3083 			} else {
3084 				/*
3085 				 * Here we are not highly buffered
3086 				 * and should pick the min we can to
3087 				 * keep from causing loss.
3088 				 */
3089 				setval = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3090 			}
3091 			rack_set_prtt_target(rack, segsiz,
3092 					     setval);
3093 		}
3094 		if (rack_probe_rtt_sets_cwnd > 1) {
3095 			/* There is a percentage here to boost */
3096 			ebdp = rack->r_ctl.rc_target_probertt_flight;
3097 			ebdp *= rack_probe_rtt_sets_cwnd;
3098 			ebdp /= 100;
3099 			setto = rack->r_ctl.rc_target_probertt_flight + ebdp;
3100 		} else
3101 			setto = rack->r_ctl.rc_target_probertt_flight;
3102 		rack->rc_tp->snd_cwnd = roundup(setto, segsiz);
3103 		if (rack->rc_tp->snd_cwnd < (segsiz * rack_timely_min_segs)) {
3104 			/* Enforce a min */
3105 			rack->rc_tp->snd_cwnd = segsiz * rack_timely_min_segs;
3106 		}
3107 		/* If we set in the cwnd also set the ssthresh point so we are in CA */
3108 		rack->rc_tp->snd_ssthresh = (rack->rc_tp->snd_cwnd - 1);
3109 	}
3110 	rack_log_rtt_shrinks(rack,  us_cts,
3111 			     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3112 			     __LINE__, RACK_RTTS_EXITPROBE);
3113 	/* Clear times last so log has all the info */
3114 	rack->r_ctl.rc_probertt_sndmax_atexit = rack->rc_tp->snd_max;
3115 	rack->r_ctl.rc_time_probertt_entered = us_cts;
3116 	rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
3117 	rack->r_ctl.rc_time_of_last_probertt = us_cts;
3118 }
3119 
3120 static void
rack_check_probe_rtt(struct tcp_rack * rack,uint32_t us_cts)3121 rack_check_probe_rtt(struct tcp_rack *rack, uint32_t us_cts)
3122 {
3123 	/* Check in on probe-rtt */
3124 	if (rack->rc_gp_filled == 0) {
3125 		/* We do not do p-rtt unless we have gp measurements */
3126 		return;
3127 	}
3128 	if (rack->in_probe_rtt) {
3129 		uint64_t no_overflow;
3130 		uint32_t endtime, must_stay;
3131 
3132 		if (rack->r_ctl.rc_went_idle_time &&
3133 		    ((us_cts - rack->r_ctl.rc_went_idle_time) > rack_min_probertt_hold)) {
3134 			/*
3135 			 * We went idle during prtt, just exit now.
3136 			 */
3137 			rack_exit_probertt(rack, us_cts);
3138 		} else if (rack_probe_rtt_safety_val &&
3139 		    TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered) &&
3140 		    ((us_cts - rack->r_ctl.rc_time_probertt_entered) > rack_probe_rtt_safety_val)) {
3141 			/*
3142 			 * Probe RTT safety value triggered!
3143 			 */
3144 			rack_log_rtt_shrinks(rack,  us_cts,
3145 					     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3146 					     __LINE__, RACK_RTTS_SAFETY);
3147 			rack_exit_probertt(rack, us_cts);
3148 		}
3149 		/* Calculate the max we will wait */
3150 		endtime = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_max_drain_wait);
3151 		if (rack->rc_highly_buffered)
3152 			endtime += (rack->r_ctl.rc_gp_srtt * rack_max_drain_hbp);
3153 		/* Calculate the min we must wait */
3154 		must_stay = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_must_drain);
3155 		if ((ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.rc_target_probertt_flight) &&
3156 		    TSTMP_LT(us_cts, endtime)) {
3157 			uint32_t calc;
3158 			/* Do we lower more? */
3159 no_exit:
3160 			if (TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered))
3161 				calc = us_cts - rack->r_ctl.rc_time_probertt_entered;
3162 			else
3163 				calc = 0;
3164 			calc /= max(rack->r_ctl.rc_gp_srtt, 1);
3165 			if (calc) {
3166 				/* Maybe */
3167 				calc *= rack_per_of_gp_probertt_reduce;
3168 				rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt - calc;
3169 				/* Limit it too */
3170 				if (rack->r_ctl.rack_per_of_gp_probertt < rack_per_of_gp_lowthresh)
3171 					rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_lowthresh;
3172 			}
3173 			/* We must reach target or the time set */
3174 			return;
3175 		}
3176 		if (rack->r_ctl.rc_time_probertt_starts == 0) {
3177 			if ((TSTMP_LT(us_cts, must_stay) &&
3178 			     rack->rc_highly_buffered) ||
3179 			     (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) >
3180 			      rack->r_ctl.rc_target_probertt_flight)) {
3181 				/* We are not past the must_stay time */
3182 				goto no_exit;
3183 			}
3184 			rack_log_rtt_shrinks(rack,  us_cts,
3185 					     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3186 					     __LINE__, RACK_RTTS_REACHTARGET);
3187 			rack->r_ctl.rc_time_probertt_starts = us_cts;
3188 			if (rack->r_ctl.rc_time_probertt_starts == 0)
3189 				rack->r_ctl.rc_time_probertt_starts = 1;
3190 			/* Restore back to our rate we want to pace at in prtt */
3191 			rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
3192 		}
3193 		/*
3194 		 * Setup our end time, some number of gp_srtts plus 200ms.
3195 		 */
3196 		no_overflow = ((uint64_t)rack->r_ctl.rc_gp_srtt *
3197 			       (uint64_t)rack_probertt_gpsrtt_cnt_mul);
3198 		if (rack_probertt_gpsrtt_cnt_div)
3199 			endtime = (uint32_t)(no_overflow / (uint64_t)rack_probertt_gpsrtt_cnt_div);
3200 		else
3201 			endtime = 0;
3202 		endtime += rack_min_probertt_hold;
3203 		endtime += rack->r_ctl.rc_time_probertt_starts;
3204 		if (TSTMP_GEQ(us_cts,  endtime)) {
3205 			/* yes, exit probertt  */
3206 			rack_exit_probertt(rack, us_cts);
3207  		}
3208 
3209 	} else 	if((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= rack_time_between_probertt) {
3210 		/* Go into probertt, its been too long since we went lower  */
3211 		rack_enter_probertt(rack, us_cts);
3212 	}
3213 }
3214 
3215 static void
rack_update_multiplier(struct tcp_rack * rack,int32_t timely_says,uint64_t last_bw_est,uint32_t rtt,int32_t rtt_diff)3216 rack_update_multiplier(struct tcp_rack *rack, int32_t timely_says, uint64_t last_bw_est,
3217 		       uint32_t rtt, int32_t rtt_diff)
3218 {
3219 	uint64_t cur_bw, up_bnd, low_bnd, subfr;
3220 	uint32_t losses;
3221 
3222 	if ((rack->rc_gp_dyn_mul == 0) ||
3223 	    (rack->use_fixed_rate) ||
3224 	    (rack->in_probe_rtt) ||
3225 	    (rack->rc_always_pace == 0)) {
3226 		/* No dynamic GP multipler in play */
3227 		return;
3228 	}
3229 	losses = rack->r_ctl.rc_loss_count - rack->r_ctl.rc_loss_at_start;
3230 	cur_bw = rack_get_bw(rack);
3231 	/* Calculate our up and down range */
3232 	up_bnd = rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_up;
3233 	up_bnd /= 100;
3234 	up_bnd += rack->r_ctl.last_gp_comp_bw;
3235 
3236 	subfr = (uint64_t)rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_down;
3237 	subfr /= 100;
3238 	low_bnd = rack->r_ctl.last_gp_comp_bw - subfr;
3239 	if ((timely_says == 2) && (rack->r_ctl.rc_no_push_at_mrtt)) {
3240 		/*
3241 		 * This is the case where our RTT is above
3242 		 * the max target and we have been configured
3243 		 * to just do timely no bonus up stuff in that case.
3244 		 *
3245 		 * There are two configurations, set to 1, and we
3246 		 * just do timely if we are over our max. If its
3247 		 * set above 1 then we slam the multipliers down
3248 		 * to 100 and then decrement per timely.
3249 		 */
3250 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
3251 				__LINE__, 3);
3252 		if (rack->r_ctl.rc_no_push_at_mrtt > 1)
3253 			rack_validate_multipliers_at_or_below_100(rack);
3254 		rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
3255 	} else if ((last_bw_est < low_bnd) && !losses) {
3256 		/*
3257 		 * We are decreasing this is a bit complicated this
3258 		 * means we are loosing ground. This could be
3259 		 * because another flow entered and we are competing
3260 		 * for b/w with it. This will push the RTT up which
3261 		 * makes timely unusable unless we want to get shoved
3262 		 * into a corner and just be backed off (the age
3263 		 * old problem with delay based CC).
3264 		 *
3265 		 * On the other hand if it was a route change we
3266 		 * would like to stay somewhat contained and not
3267 		 * blow out the buffers.
3268 		 */
3269 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
3270 				__LINE__, 3);
3271 		rack->r_ctl.last_gp_comp_bw = cur_bw;
3272 		if (rack->rc_gp_bwred == 0) {
3273 			/* Go into reduction counting */
3274 			rack->rc_gp_bwred = 1;
3275 			rack->rc_gp_timely_dec_cnt = 0;
3276 		}
3277 		if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) ||
3278 		    (timely_says == 0)) {
3279 			/*
3280 			 * Push another time with a faster pacing
3281 			 * to try to gain back (we include override to
3282 			 * get a full raise factor).
3283 			 */
3284 			if ((rack->rc_gp_saw_ca && rack->r_ctl.rack_per_of_gp_ca <= rack_down_raise_thresh) ||
3285 			    (rack->rc_gp_saw_ss && rack->r_ctl.rack_per_of_gp_ss <= rack_down_raise_thresh) ||
3286 			    (timely_says == 0) ||
3287 			    (rack_down_raise_thresh == 0)) {
3288 				/*
3289 				 * Do an override up in b/w if we were
3290 				 * below the threshold or if the threshold
3291 				 * is zero we always do the raise.
3292 				 */
3293 				rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 1);
3294 			} else {
3295 				/* Log it stays the same */
3296 				rack_log_timely(rack,  0, last_bw_est, low_bnd, 0,
3297 						__LINE__, 11);
3298 			}
3299 			rack->rc_gp_timely_dec_cnt++;
3300 			/* We are not incrementing really no-count */
3301 			rack->rc_gp_incr = 0;
3302 			rack->rc_gp_timely_inc_cnt = 0;
3303 		} else {
3304 			/*
3305 			 * Lets just use the RTT
3306 			 * information and give up
3307 			 * pushing.
3308 			 */
3309 			goto use_timely;
3310 		}
3311 	}  else if ((timely_says != 2) &&
3312 		    !losses &&
3313 		    (last_bw_est > up_bnd)) {
3314 		/*
3315 		 * We are increasing b/w lets keep going, updating
3316 		 * our b/w and ignoring any timely input, unless
3317 		 * of course we are at our max raise (if there is one).
3318 		 */
3319 
3320 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
3321 				__LINE__, 3);
3322 		rack->r_ctl.last_gp_comp_bw = cur_bw;
3323 		if (rack->rc_gp_saw_ss &&
3324 		    rack_per_upper_bound_ss &&
3325 		     (rack->r_ctl.rack_per_of_gp_ss == rack_per_upper_bound_ss)) {
3326 			    /*
3327 			     * In cases where we can't go higher
3328 			     * we should just use timely.
3329 			     */
3330 			    goto use_timely;
3331 		}
3332 		if (rack->rc_gp_saw_ca &&
3333 		    rack_per_upper_bound_ca &&
3334 		    (rack->r_ctl.rack_per_of_gp_ca == rack_per_upper_bound_ca)) {
3335 			    /*
3336 			     * In cases where we can't go higher
3337 			     * we should just use timely.
3338 			     */
3339 			    goto use_timely;
3340 		}
3341 		rack->rc_gp_bwred = 0;
3342 		rack->rc_gp_timely_dec_cnt = 0;
3343 		/* You get a set number of pushes if timely is trying to reduce  */
3344 		if ((rack->rc_gp_incr < rack_timely_max_push_rise) || (timely_says == 0)) {
3345 			rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
3346 		} else {
3347  			/* Log it stays the same */
3348 			rack_log_timely(rack,  0, last_bw_est, up_bnd, 0,
3349 			    __LINE__, 12);
3350 		}
3351 		return;
3352 	} else {
3353 		/*
3354 		 * We are staying between the lower and upper range bounds
3355 		 * so use timely to decide.
3356 		 */
3357 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
3358 				__LINE__, 3);
3359 use_timely:
3360 		if (timely_says) {
3361 			rack->rc_gp_incr = 0;
3362 			rack->rc_gp_timely_inc_cnt = 0;
3363 			if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) &&
3364 			    !losses &&
3365 			    (last_bw_est < low_bnd)) {
3366 				/* We are loosing ground */
3367 				rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
3368 				rack->rc_gp_timely_dec_cnt++;
3369 				/* We are not incrementing really no-count */
3370 				rack->rc_gp_incr = 0;
3371 				rack->rc_gp_timely_inc_cnt = 0;
3372 			} else
3373 				rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
3374 		} else  {
3375 			rack->rc_gp_bwred = 0;
3376 			rack->rc_gp_timely_dec_cnt = 0;
3377 			rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
3378 		}
3379 	}
3380 }
3381 
3382 static int32_t
rack_make_timely_judgement(struct tcp_rack * rack,uint32_t rtt,int32_t rtt_diff,uint32_t prev_rtt)3383 rack_make_timely_judgement(struct tcp_rack *rack, uint32_t rtt, int32_t rtt_diff, uint32_t prev_rtt)
3384 {
3385 	int32_t timely_says;
3386 	uint64_t log_mult, log_rtt_a_diff;
3387 
3388 	log_rtt_a_diff = rtt;
3389 	log_rtt_a_diff <<= 32;
3390 	log_rtt_a_diff |= (uint32_t)rtt_diff;
3391 	if (rtt >= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) *
3392 		    rack_gp_rtt_maxmul)) {
3393 		/* Reduce the b/w multipler */
3394 		timely_says = 2;
3395 		log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
3396 		log_mult <<= 32;
3397 		log_mult |= prev_rtt;
3398 		rack_log_timely(rack,  timely_says, log_mult,
3399 				get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3400 				log_rtt_a_diff, __LINE__, 4);
3401 	} else if (rtt <= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
3402 			   ((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
3403 			    max(rack_gp_rtt_mindiv , 1)))) {
3404 		/* Increase the b/w multipler */
3405 		log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
3406 			((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
3407 			 max(rack_gp_rtt_mindiv , 1));
3408 		log_mult <<= 32;
3409 		log_mult |= prev_rtt;
3410 		timely_says = 0;
3411 		rack_log_timely(rack,  timely_says, log_mult ,
3412 				get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3413 				log_rtt_a_diff, __LINE__, 5);
3414 	} else {
3415 		/*
3416 		 * Use a gradient to find it the timely gradient
3417 		 * is:
3418 		 * grad = rc_rtt_diff / min_rtt;
3419 		 *
3420 		 * anything below or equal to 0 will be
3421 		 * a increase indication. Anything above
3422 		 * zero is a decrease. Note we take care
3423 		 * of the actual gradient calculation
3424 		 * in the reduction (its not needed for
3425 		 * increase).
3426 		 */
3427 		log_mult = prev_rtt;
3428 		if (rtt_diff <= 0) {
3429 			/*
3430 			 * Rttdiff is less than zero, increase the
3431 			 * b/w multipler (its 0 or negative)
3432 			 */
3433 			timely_says = 0;
3434 			rack_log_timely(rack,  timely_says, log_mult,
3435 					get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 6);
3436 		} else {
3437 			/* Reduce the b/w multipler */
3438 			timely_says = 1;
3439 			rack_log_timely(rack,  timely_says, log_mult,
3440 					get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 7);
3441 		}
3442 	}
3443 	return (timely_says);
3444 }
3445 
3446 static void
rack_do_goodput_measurement(struct tcpcb * tp,struct tcp_rack * rack,tcp_seq th_ack,int line)3447 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
3448 			    tcp_seq th_ack, int line)
3449 {
3450 	uint64_t tim, bytes_ps, ltim, stim, utim;
3451 	uint32_t segsiz, bytes, reqbytes, us_cts;
3452 	int32_t gput, new_rtt_diff, timely_says;
3453 
3454 	us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
3455 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
3456 	if (TSTMP_GEQ(us_cts, tp->gput_ts))
3457 		tim = us_cts - tp->gput_ts;
3458 	else
3459 		tim = 0;
3460 
3461 	if (TSTMP_GT(rack->r_ctl.rc_gp_cumack_ts, rack->r_ctl.rc_gp_output_ts))
3462 		stim = rack->r_ctl.rc_gp_cumack_ts - rack->r_ctl.rc_gp_output_ts;
3463 	else
3464 		stim = 0;
3465 	/*
3466 	 * Use the larger of the send time or ack time. This prevents us
3467 	 * from being influenced by ack artifacts to come up with too
3468 	 * high of measurement. Note that since we are spanning over many more
3469 	 * bytes in most of our measurements hopefully that is less likely to
3470 	 * occur.
3471 	 */
3472 	if (tim > stim)
3473 		utim = max(tim, 1);
3474 	else
3475 		utim = max(stim, 1);
3476 	/* Lets validate utim */
3477 	ltim = max(1, (utim/HPTS_USEC_IN_MSEC));
3478 	gput = (((uint64_t) (th_ack - tp->gput_seq)) << 3) / ltim;
3479 	reqbytes = min(rc_init_window(rack), (MIN_GP_WIN * segsiz));
3480 	if ((tim == 0) && (stim == 0)) {
3481 		/*
3482 		 * Invalid measurement time, maybe
3483 		 * all on one ack/one send?
3484 		 */
3485 		bytes = 0;
3486 		bytes_ps = 0;
3487 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
3488 					   0, 0, 0, 10, __LINE__, NULL);
3489 		goto skip_measurement;
3490 	}
3491 	if (rack->r_ctl.rc_gp_lowrtt == 0xffffffff) {
3492 		/* We never made a us_rtt measurement? */
3493 		bytes = 0;
3494 		bytes_ps = 0;
3495 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
3496 					   0, 0, 0, 10, __LINE__, NULL);
3497 		goto skip_measurement;
3498 	}
3499 	/*
3500 	 * Calculate the maximum possible b/w this connection
3501 	 * could have. We base our calculation on the lowest
3502 	 * rtt we have seen during the measurement and the
3503 	 * largest rwnd the client has given us in that time. This
3504 	 * forms a BDP that is the maximum that we could ever
3505 	 * get to the client. Anything larger is not valid.
3506 	 *
3507 	 * I originally had code here that rejected measurements
3508 	 * where the time was less than 1/2 the latest us_rtt.
3509 	 * But after thinking on that I realized its wrong since
3510 	 * say you had a 150Mbps or even 1Gbps link, and you
3511 	 * were a long way away.. example I am in Europe (100ms rtt)
3512 	 * talking to my 1Gbps link in S.C. Now measuring say 150,000
3513 	 * bytes my time would be 1.2ms, and yet my rtt would say
3514 	 * the measurement was invalid the time was < 50ms. The
3515 	 * same thing is true for 150Mb (8ms of time).
3516 	 *
3517 	 * A better way I realized is to look at what the maximum
3518 	 * the connection could possibly do. This is gated on
3519 	 * the lowest RTT we have seen and the highest rwnd.
3520 	 * We should in theory never exceed that, if we are
3521 	 * then something on the path is storing up packets
3522 	 * and then feeding them all at once to our endpoint
3523 	 * messing up our measurement.
3524 	 */
3525 	rack->r_ctl.last_max_bw = rack->r_ctl.rc_gp_high_rwnd;
3526 	rack->r_ctl.last_max_bw *= HPTS_USEC_IN_SEC;
3527 	rack->r_ctl.last_max_bw /= rack->r_ctl.rc_gp_lowrtt;
3528 	if (SEQ_LT(th_ack, tp->gput_seq)) {
3529 		/* No measurement can be made */
3530 		bytes = 0;
3531 		bytes_ps = 0;
3532 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
3533 					   0, 0, 0, 10, __LINE__, NULL);
3534 		goto skip_measurement;
3535 	} else
3536 		bytes = (th_ack - tp->gput_seq);
3537 	bytes_ps = (uint64_t)bytes;
3538 	/*
3539 	 * Don't measure a b/w for pacing unless we have gotten at least
3540 	 * an initial windows worth of data in this measurement interval.
3541 	 *
3542 	 * Small numbers of bytes get badly influenced by delayed ack and
3543 	 * other artifacts. Note we take the initial window or our
3544 	 * defined minimum GP (defaulting to 10 which hopefully is the
3545 	 * IW).
3546 	 */
3547 	if (rack->rc_gp_filled == 0) {
3548 		/*
3549 		 * The initial estimate is special. We
3550 		 * have blasted out an IW worth of packets
3551 		 * without a real valid ack ts results. We
3552 		 * then setup the app_limited_needs_set flag,
3553 		 * this should get the first ack in (probably 2
3554 		 * MSS worth) to be recorded as the timestamp.
3555 		 * We thus allow a smaller number of bytes i.e.
3556 		 * IW - 2MSS.
3557 		 */
3558 		reqbytes -= (2 * segsiz);
3559 		/* Also lets fill previous for our first measurement to be neutral */
3560 		rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
3561 	}
3562 	if ((bytes_ps < reqbytes) || rack->app_limited_needs_set) {
3563 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
3564 					   rack->r_ctl.rc_app_limited_cnt,
3565 					   0, 0, 10, __LINE__, NULL);
3566 		goto skip_measurement;
3567 	}
3568 	/*
3569 	 * We now need to calculate the Timely like status so
3570 	 * we can update (possibly) the b/w multipliers.
3571 	 */
3572 	new_rtt_diff = (int32_t)rack->r_ctl.rc_gp_srtt - (int32_t)rack->r_ctl.rc_prev_gp_srtt;
3573 	if (rack->rc_gp_filled == 0) {
3574 		/* No previous reading */
3575 		rack->r_ctl.rc_rtt_diff = new_rtt_diff;
3576 	} else {
3577 		if (rack->measure_saw_probe_rtt == 0) {
3578 			/*
3579 			 * We don't want a probertt to be counted
3580 			 * since it will be negative incorrectly. We
3581 			 * expect to be reducing the RTT when we
3582 			 * pace at a slower rate.
3583 			 */
3584 			rack->r_ctl.rc_rtt_diff -= (rack->r_ctl.rc_rtt_diff / 8);
3585 			rack->r_ctl.rc_rtt_diff += (new_rtt_diff / 8);
3586 		}
3587 	}
3588 	timely_says = rack_make_timely_judgement(rack,
3589 		rack->r_ctl.rc_gp_srtt,
3590 		rack->r_ctl.rc_rtt_diff,
3591 	        rack->r_ctl.rc_prev_gp_srtt
3592 		);
3593 	bytes_ps *= HPTS_USEC_IN_SEC;
3594 	bytes_ps /= utim;
3595 	if (bytes_ps > rack->r_ctl.last_max_bw) {
3596 		/*
3597 		 * Something is on path playing
3598 		 * since this b/w is not possible based
3599 		 * on our BDP (highest rwnd and lowest rtt
3600 		 * we saw in the measurement window).
3601 		 *
3602 		 * Another option here would be to
3603 		 * instead skip the measurement.
3604 		 */
3605 		rack_log_pacing_delay_calc(rack, bytes, reqbytes,
3606 					   bytes_ps, rack->r_ctl.last_max_bw, 0,
3607 					   11, __LINE__, NULL);
3608 		bytes_ps = rack->r_ctl.last_max_bw;
3609 	}
3610 	/* We store gp for b/w in bytes per second  */
3611 	if (rack->rc_gp_filled == 0) {
3612 		/* Initial measurment */
3613 		if (bytes_ps) {
3614 			rack->r_ctl.gp_bw = bytes_ps;
3615 			rack->rc_gp_filled = 1;
3616 			rack->r_ctl.num_avg = 1;
3617 			rack_set_pace_segments(rack->rc_tp, rack, __LINE__);
3618 		} else {
3619 			rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
3620 						   rack->r_ctl.rc_app_limited_cnt,
3621 						   0, 0, 10, __LINE__, NULL);
3622 		}
3623 		if (rack->rc_inp->inp_in_hpts &&
3624 		    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
3625 			/*
3626 			 * Ok we can't trust the pacer in this case
3627 			 * where we transition from un-paced to paced.
3628 			 * Or for that matter when the burst mitigation
3629 			 * was making a wild guess and got it wrong.
3630 			 * Stop the pacer and clear up all the aggregate
3631 			 * delays etc.
3632 			 */
3633 			tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT);
3634 			rack->r_ctl.rc_hpts_flags = 0;
3635 			rack->r_ctl.rc_last_output_to = 0;
3636 		}
3637 	} else if (rack->r_ctl.num_avg < RACK_REQ_AVG) {
3638 		/* Still a small number run an average */
3639 		rack->r_ctl.gp_bw += bytes_ps;
3640 		rack->r_ctl.num_avg++;
3641 		if (rack->r_ctl.num_avg >= RACK_REQ_AVG) {
3642 			/* We have collected enought to move forward */
3643 			rack->r_ctl.gp_bw /= (uint64_t)rack->r_ctl.num_avg;
3644 		}
3645 	} else {
3646 		/*
3647 		 * We want to take 1/wma of the goodput and add in to 7/8th
3648 		 * of the old value weighted by the srtt. So if your measurement
3649 		 * period is say 2 SRTT's long you would get 1/4 as the
3650 		 * value, if it was like 1/2 SRTT then you would get 1/16th.
3651 		 *
3652 		 * But we must be careful not to take too much i.e. if the
3653 		 * srtt is say 20ms and the measurement is taken over
3654 		 * 400ms our weight would be 400/20 i.e. 20. On the
3655 		 * other hand if we get a measurement over 1ms with a
3656 		 * 10ms rtt we only want to take a much smaller portion.
3657 		 */
3658 		uint64_t  resid_bw, subpart, addpart, srtt;
3659 
3660 		srtt = ((uint64_t)TICKS_2_USEC(tp->t_srtt) >> TCP_RTT_SHIFT);
3661 		if (srtt == 0) {
3662 			/*
3663 			 * Strange why did t_srtt go back to zero?
3664 			 */
3665 			if (rack->r_ctl.rc_rack_min_rtt)
3666 				srtt = (rack->r_ctl.rc_rack_min_rtt * HPTS_USEC_IN_MSEC);
3667 			else
3668 				srtt = HPTS_USEC_IN_MSEC;
3669 		}
3670 		/*
3671 		 * XXXrrs: Note for reviewers, in playing with
3672 		 * dynamic pacing I discovered this GP calculation
3673 		 * as done originally leads to some undesired results.
3674 		 * Basically you can get longer measurements contributing
3675 		 * too much to the WMA. Thus I changed it if you are doing
3676 		 * dynamic adjustments to only do the aportioned adjustment
3677 		 * if we have a very small (time wise) measurement. Longer
3678 		 * measurements just get there weight (defaulting to 1/8)
3679 		 * add to the WMA. We may want to think about changing
3680 		 * this to always do that for both sides i.e. dynamic
3681 		 * and non-dynamic... but considering lots of folks
3682 		 * were playing with this I did not want to change the
3683 		 * calculation per.se. without your thoughts.. Lawerence?
3684 		 * Peter??
3685 		 */
3686 		if (rack->rc_gp_dyn_mul == 0) {
3687 			subpart = rack->r_ctl.gp_bw * utim;
3688 			subpart /= (srtt * 8);
3689 			if (subpart < (rack->r_ctl.gp_bw / 2)) {
3690 				/*
3691 				 * The b/w update takes no more
3692 				 * away then 1/2 our running total
3693 				 * so factor it in.
3694 				 */
3695 				addpart = bytes_ps * utim;
3696 				addpart /= (srtt * 8);
3697 			} else {
3698 				/*
3699 				 * Don't allow a single measurement
3700 				 * to account for more than 1/2 of the
3701 				 * WMA. This could happen on a retransmission
3702 				 * where utim becomes huge compared to
3703 				 * srtt (multiple retransmissions when using
3704 				 * the sending rate which factors in all the
3705 				 * transmissions from the first one).
3706 				 */
3707 				subpart = rack->r_ctl.gp_bw / 2;
3708 				addpart = bytes_ps / 2;
3709 			}
3710 			resid_bw = rack->r_ctl.gp_bw - subpart;
3711 			rack->r_ctl.gp_bw = resid_bw + addpart;
3712 		} else {
3713 			if ((utim / srtt) <= 1) {
3714 				/*
3715 				 * The b/w update was over a small period
3716 				 * of time. The idea here is to prevent a small
3717 				 * measurement time period from counting
3718 				 * too much. So we scale it based on the
3719 				 * time so it attributes less than 1/rack_wma_divisor
3720 				 * of its measurement.
3721 				 */
3722 				subpart = rack->r_ctl.gp_bw * utim;
3723 				subpart /= (srtt * rack_wma_divisor);
3724 				addpart = bytes_ps * utim;
3725 				addpart /= (srtt * rack_wma_divisor);
3726 			} else {
3727 				/*
3728 				 * The scaled measurement was long
3729 				 * enough so lets just add in the
3730 				 * portion of the measurment i.e. 1/rack_wma_divisor
3731 				 */
3732 				subpart = rack->r_ctl.gp_bw / rack_wma_divisor;
3733 				addpart = bytes_ps / rack_wma_divisor;
3734 			}
3735 			if ((rack->measure_saw_probe_rtt == 0) ||
3736 		            (bytes_ps > rack->r_ctl.gp_bw)) {
3737 				/*
3738 				 * For probe-rtt we only add it in
3739 				 * if its larger, all others we just
3740 				 * add in.
3741 				 */
3742 				resid_bw = rack->r_ctl.gp_bw - subpart;
3743 				rack->r_ctl.gp_bw = resid_bw + addpart;
3744 			}
3745 		}
3746 	}
3747 	/* We do not update any multipliers if we are in or have seen a probe-rtt */
3748 	if ((rack->measure_saw_probe_rtt == 0) && rack->rc_gp_rtt_set)
3749 		rack_update_multiplier(rack, timely_says, bytes_ps,
3750 				       rack->r_ctl.rc_gp_srtt,
3751 				       rack->r_ctl.rc_rtt_diff);
3752 	rack_log_pacing_delay_calc(rack, bytes, tim, bytes_ps, stim,
3753 				   rack_get_bw(rack), 3, line, NULL);
3754 	/* reset the gp srtt and setup the new prev */
3755 	rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
3756 	/* Record the lost count for the next measurement */
3757 	rack->r_ctl.rc_loss_at_start = rack->r_ctl.rc_loss_count;
3758 	/*
3759 	 * We restart our diffs based on the gpsrtt in the
3760 	 * measurement window.
3761 	 */
3762 	rack->rc_gp_rtt_set = 0;
3763 	rack->rc_gp_saw_rec = 0;
3764 	rack->rc_gp_saw_ca = 0;
3765 	rack->rc_gp_saw_ss = 0;
3766 	rack->rc_dragged_bottom = 0;
3767 skip_measurement:
3768 
3769 #ifdef STATS
3770 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT,
3771 				 gput);
3772 	/*
3773 	 * XXXLAS: This is a temporary hack, and should be
3774 	 * chained off VOI_TCP_GPUT when stats(9) grows an
3775 	 * API to deal with chained VOIs.
3776 	 */
3777 	if (tp->t_stats_gput_prev > 0)
3778 		stats_voi_update_abs_s32(tp->t_stats,
3779 					 VOI_TCP_GPUT_ND,
3780 					 ((gput - tp->t_stats_gput_prev) * 100) /
3781 					 tp->t_stats_gput_prev);
3782 #endif
3783 	tp->t_flags &= ~TF_GPUTINPROG;
3784 	tp->t_stats_gput_prev = gput;
3785 	/*
3786 	 * Now are we app limited now and there is space from where we
3787 	 * were to where we want to go?
3788 	 *
3789 	 * We don't do the other case i.e. non-applimited here since
3790 	 * the next send will trigger us picking up the missing data.
3791 	 */
3792 	if (rack->r_ctl.rc_first_appl &&
3793 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
3794 	    rack->r_ctl.rc_app_limited_cnt &&
3795 	    (SEQ_GT(rack->r_ctl.rc_first_appl->r_start, th_ack)) &&
3796 	    ((rack->r_ctl.rc_first_appl->r_start - th_ack) >
3797 	     max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
3798 		/*
3799 		 * Yep there is enough outstanding to make a measurement here.
3800 		 */
3801 		struct rack_sendmap *rsm, fe;
3802 
3803 		tp->t_flags |= TF_GPUTINPROG;
3804 		rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
3805 		rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
3806 		tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
3807 		rack->app_limited_needs_set = 0;
3808 		tp->gput_seq = th_ack;
3809 		if (rack->in_probe_rtt)
3810 			rack->measure_saw_probe_rtt = 1;
3811 		else if ((rack->measure_saw_probe_rtt) &&
3812 			 (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
3813 			rack->measure_saw_probe_rtt = 0;
3814 		if ((rack->r_ctl.rc_first_appl->r_start - th_ack) >= rack_get_measure_window(tp, rack)) {
3815 			/* There is a full window to gain info from */
3816 			tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
3817 		} else {
3818 			/* We can only measure up to the applimited point */
3819 			tp->gput_ack = tp->gput_seq + (rack->r_ctl.rc_first_appl->r_start - th_ack);
3820 		}
3821 		/*
3822 		 * Now we need to find the timestamp of the send at tp->gput_seq
3823 		 * for the send based measurement.
3824 		 */
3825 		fe.r_start = tp->gput_seq;
3826 		rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
3827 		if (rsm) {
3828 			/* Ok send-based limit is set */
3829 			if (SEQ_LT(rsm->r_start, tp->gput_seq)) {
3830 				/*
3831 				 * Move back to include the earlier part
3832 				 * so our ack time lines up right (this may
3833 				 * make an overlapping measurement but thats
3834 				 * ok).
3835 				 */
3836 				tp->gput_seq = rsm->r_start;
3837 			}
3838 			if (rsm->r_flags & RACK_ACKED)
3839 				tp->gput_ts = rsm->r_ack_arrival;
3840 			else
3841 				rack->app_limited_needs_set = 1;
3842 			rack->r_ctl.rc_gp_output_ts = rsm->usec_orig_send;
3843 		} else {
3844 			/*
3845 			 * If we don't find the rsm due to some
3846 			 * send-limit set the current time, which
3847 			 * basically disables the send-limit.
3848 			 */
3849 			rack->r_ctl.rc_gp_output_ts = tcp_get_usecs(NULL);
3850 		}
3851 		rack_log_pacing_delay_calc(rack,
3852 					   tp->gput_seq,
3853 					   tp->gput_ack,
3854 					   (uint64_t)rsm,
3855 					   tp->gput_ts,
3856 					   rack->r_ctl.rc_app_limited_cnt,
3857 					   9,
3858 					   __LINE__, NULL);
3859 	}
3860 }
3861 
3862 /*
3863  * CC wrapper hook functions
3864  */
3865 static void
rack_ack_received(struct tcpcb * tp,struct tcp_rack * rack,struct tcphdr * th,uint16_t nsegs,uint16_t type,int32_t recovery)3866 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack, struct tcphdr *th, uint16_t nsegs,
3867     uint16_t type, int32_t recovery)
3868 {
3869 	INP_WLOCK_ASSERT(tp->t_inpcb);
3870 	tp->ccv->nsegs = nsegs;
3871 	tp->ccv->bytes_this_ack = BYTES_THIS_ACK(tp, th);
3872 	if ((recovery) && (rack->r_ctl.rc_early_recovery_segs)) {
3873 		uint32_t max;
3874 
3875 		max = rack->r_ctl.rc_early_recovery_segs * ctf_fixed_maxseg(tp);
3876 		if (tp->ccv->bytes_this_ack > max) {
3877 			tp->ccv->bytes_this_ack = max;
3878 		}
3879 	}
3880 	if (rack->r_ctl.cwnd_to_use <= tp->snd_wnd)
3881 		tp->ccv->flags |= CCF_CWND_LIMITED;
3882 	else
3883 		tp->ccv->flags &= ~CCF_CWND_LIMITED;
3884 #ifdef STATS
3885 	stats_voi_update_abs_s32(tp->t_stats, VOI_TCP_CALCFRWINDIFF,
3886 	    ((int32_t)rack->r_ctl.cwnd_to_use) - tp->snd_wnd);
3887 #endif
3888 	if ((tp->t_flags & TF_GPUTINPROG) &&
3889 	    rack_enough_for_measurement(tp, rack, th->th_ack)) {
3890 		/* Measure the Goodput */
3891 		rack_do_goodput_measurement(tp, rack, th->th_ack, __LINE__);
3892 #ifdef NETFLIX_PEAKRATE
3893 		if ((type == CC_ACK) &&
3894 		    (tp->t_maxpeakrate)) {
3895 			/*
3896 			 * We update t_peakrate_thr. This gives us roughly
3897 			 * one update per round trip time. Note
3898 			 * it will only be used if pace_always is off i.e
3899 			 * we don't do this for paced flows.
3900 			 */
3901 			tcp_update_peakrate_thr(tp);
3902 		}
3903 #endif
3904 	}
3905 	if (rack->r_ctl.cwnd_to_use > tp->snd_ssthresh) {
3906 		tp->t_bytes_acked += tp->ccv->bytes_this_ack;
3907 		if (tp->t_bytes_acked >= rack->r_ctl.cwnd_to_use) {
3908 			tp->t_bytes_acked -= rack->r_ctl.cwnd_to_use;
3909 			tp->ccv->flags |= CCF_ABC_SENTAWND;
3910 		}
3911 	} else {
3912 		tp->ccv->flags &= ~CCF_ABC_SENTAWND;
3913 		tp->t_bytes_acked = 0;
3914 	}
3915 	if (CC_ALGO(tp)->ack_received != NULL) {
3916 		/* XXXLAS: Find a way to live without this */
3917 		tp->ccv->curack = th->th_ack;
3918 		CC_ALGO(tp)->ack_received(tp->ccv, type);
3919 	}
3920 #ifdef STATS
3921 	stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_LCWIN, rack->r_ctl.cwnd_to_use);
3922 #endif
3923 	if (rack->r_ctl.rc_rack_largest_cwnd < rack->r_ctl.cwnd_to_use) {
3924 		rack->r_ctl.rc_rack_largest_cwnd = rack->r_ctl.cwnd_to_use;
3925 	}
3926 #ifdef NETFLIX_PEAKRATE
3927 	/* we enforce max peak rate if it is set and we are not pacing */
3928 	if ((rack->rc_always_pace == 0) &&
3929 	    tp->t_peakrate_thr &&
3930 	    (tp->snd_cwnd > tp->t_peakrate_thr)) {
3931 		tp->snd_cwnd = tp->t_peakrate_thr;
3932 	}
3933 #endif
3934 }
3935 
3936 static void
tcp_rack_partialack(struct tcpcb * tp,struct tcphdr * th)3937 tcp_rack_partialack(struct tcpcb *tp, struct tcphdr *th)
3938 {
3939 	struct tcp_rack *rack;
3940 
3941 	rack = (struct tcp_rack *)tp->t_fb_ptr;
3942 	INP_WLOCK_ASSERT(tp->t_inpcb);
3943 	/*
3944 	 * If we are doing PRR and have enough
3945 	 * room to send <or> we are pacing and prr
3946 	 * is disabled we will want to see if we
3947 	 * can send data (by setting r_wanted_output to
3948 	 * true).
3949 	 */
3950 	if ((rack->r_ctl.rc_prr_sndcnt > 0) ||
3951 	    rack->rack_no_prr)
3952 		rack->r_wanted_output = 1;
3953 }
3954 
3955 static void
rack_post_recovery(struct tcpcb * tp,struct tcphdr * th)3956 rack_post_recovery(struct tcpcb *tp, struct tcphdr *th)
3957 {
3958 	struct tcp_rack *rack;
3959 	uint32_t orig_cwnd;
3960 
3961 	orig_cwnd = tp->snd_cwnd;
3962 	INP_WLOCK_ASSERT(tp->t_inpcb);
3963 	rack = (struct tcp_rack *)tp->t_fb_ptr;
3964 	if (rack->rc_not_backing_off == 0) {
3965 		/* only alert CC if we alerted when we entered */
3966 		if (CC_ALGO(tp)->post_recovery != NULL) {
3967 			tp->ccv->curack = th->th_ack;
3968 			CC_ALGO(tp)->post_recovery(tp->ccv);
3969 		}
3970 		if (tp->snd_cwnd > tp->snd_ssthresh) {
3971 			/* Drop us down to the ssthresh (1/2 cwnd at loss) */
3972 			tp->snd_cwnd = tp->snd_ssthresh;
3973 		}
3974 	}
3975 	if ((rack->rack_no_prr == 0) &&
3976 	    (rack->r_ctl.rc_prr_sndcnt > 0)) {
3977 		/* Suck the next prr cnt back into cwnd */
3978 		tp->snd_cwnd += rack->r_ctl.rc_prr_sndcnt;
3979 		rack->r_ctl.rc_prr_sndcnt = 0;
3980 		rack_log_to_prr(rack, 1, 0);
3981 	}
3982 	rack_log_to_prr(rack, 14, orig_cwnd);
3983 	tp->snd_recover = tp->snd_una;
3984 	EXIT_RECOVERY(tp->t_flags);
3985 }
3986 
3987 static void
rack_cong_signal(struct tcpcb * tp,struct tcphdr * th,uint32_t type)3988 rack_cong_signal(struct tcpcb *tp, struct tcphdr *th, uint32_t type)
3989 {
3990 	struct tcp_rack *rack;
3991 
3992 	INP_WLOCK_ASSERT(tp->t_inpcb);
3993 
3994 	rack = (struct tcp_rack *)tp->t_fb_ptr;
3995 	switch (type) {
3996 	case CC_NDUPACK:
3997 		tp->t_flags &= ~TF_WASFRECOVERY;
3998 		tp->t_flags &= ~TF_WASCRECOVERY;
3999 		if (!IN_FASTRECOVERY(tp->t_flags)) {
4000 			rack->r_ctl.rc_prr_delivered = 0;
4001 			rack->r_ctl.rc_prr_out = 0;
4002 			if (rack->rack_no_prr == 0) {
4003 				rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
4004 				rack_log_to_prr(rack, 2, 0);
4005 			}
4006 			rack->r_ctl.rc_prr_recovery_fs = tp->snd_max - tp->snd_una;
4007 			tp->snd_recover = tp->snd_max;
4008 			if (tp->t_flags2 & TF2_ECN_PERMIT)
4009 				tp->t_flags2 |= TF2_ECN_SND_CWR;
4010 		}
4011 		break;
4012 	case CC_ECN:
4013 		if (!IN_CONGRECOVERY(tp->t_flags) ||
4014 		    /*
4015 		     * Allow ECN reaction on ACK to CWR, if
4016 		     * that data segment was also CE marked.
4017 		     */
4018 		    SEQ_GEQ(th->th_ack, tp->snd_recover)) {
4019 			EXIT_CONGRECOVERY(tp->t_flags);
4020 			KMOD_TCPSTAT_INC(tcps_ecn_rcwnd);
4021 			tp->snd_recover = tp->snd_max + 1;
4022 			if (tp->t_flags2 & TF2_ECN_PERMIT)
4023 				tp->t_flags2 |= TF2_ECN_SND_CWR;
4024 		}
4025 		break;
4026 	case CC_RTO:
4027 		tp->t_dupacks = 0;
4028 		tp->t_bytes_acked = 0;
4029 		EXIT_RECOVERY(tp->t_flags);
4030 		tp->snd_ssthresh = max(2, min(tp->snd_wnd, rack->r_ctl.cwnd_to_use) / 2 /
4031 		    ctf_fixed_maxseg(tp)) * ctf_fixed_maxseg(tp);
4032 		tp->snd_cwnd = ctf_fixed_maxseg(tp);
4033 		if (tp->t_flags2 & TF2_ECN_PERMIT)
4034 			tp->t_flags2 |= TF2_ECN_SND_CWR;
4035 		break;
4036 	case CC_RTO_ERR:
4037 		KMOD_TCPSTAT_INC(tcps_sndrexmitbad);
4038 		/* RTO was unnecessary, so reset everything. */
4039 		tp->snd_cwnd = tp->snd_cwnd_prev;
4040 		tp->snd_ssthresh = tp->snd_ssthresh_prev;
4041 		tp->snd_recover = tp->snd_recover_prev;
4042 		if (tp->t_flags & TF_WASFRECOVERY) {
4043 			ENTER_FASTRECOVERY(tp->t_flags);
4044 			tp->t_flags &= ~TF_WASFRECOVERY;
4045 		}
4046 		if (tp->t_flags & TF_WASCRECOVERY) {
4047 			ENTER_CONGRECOVERY(tp->t_flags);
4048 			tp->t_flags &= ~TF_WASCRECOVERY;
4049 		}
4050 		tp->snd_nxt = tp->snd_max;
4051 		tp->t_badrxtwin = 0;
4052 		break;
4053 	}
4054 	/*
4055 	 * If we are below our max rtt, don't
4056 	 * signal the CC control to change things.
4057 	 * instead set it up so that we are in
4058 	 * recovery but not going to back off.
4059 	 */
4060 
4061 	if (rack->rc_highly_buffered) {
4062 		/*
4063 		 * Do we use the higher rtt for
4064 		 * our threshold to not backoff (like CDG)?
4065 		 */
4066 		uint32_t rtt_mul, rtt_div;
4067 
4068 		if (rack_use_max_for_nobackoff) {
4069 			rtt_mul = (rack_gp_rtt_maxmul - 1);
4070 			rtt_div = 1;
4071 		} else {
4072 			rtt_mul = rack_gp_rtt_minmul;
4073 			rtt_div = max(rack_gp_rtt_mindiv , 1);
4074 		}
4075 		if (rack->r_ctl.rc_gp_srtt <= (rack->r_ctl.rc_lowest_us_rtt +
4076 					       ((rack->r_ctl.rc_lowest_us_rtt * rtt_mul) /
4077 						rtt_div))) {
4078 			/* below our min threshold */
4079 			rack->rc_not_backing_off = 1;
4080 			ENTER_RECOVERY(rack->rc_tp->t_flags);
4081 			rack_log_rtt_shrinks(rack, 0,
4082 					     rtt_mul,
4083 					     rtt_div,
4084 					     RACK_RTTS_NOBACKOFF);
4085 			return;
4086 		}
4087 	}
4088 	rack->rc_not_backing_off = 0;
4089 	if (CC_ALGO(tp)->cong_signal != NULL) {
4090 		if (th != NULL)
4091 			tp->ccv->curack = th->th_ack;
4092 		CC_ALGO(tp)->cong_signal(tp->ccv, type);
4093 	}
4094 }
4095 
4096 static inline void
rack_cc_after_idle(struct tcp_rack * rack,struct tcpcb * tp)4097 rack_cc_after_idle(struct tcp_rack *rack, struct tcpcb *tp)
4098 {
4099 	uint32_t i_cwnd;
4100 
4101 	INP_WLOCK_ASSERT(tp->t_inpcb);
4102 
4103 #ifdef NETFLIX_STATS
4104 	KMOD_TCPSTAT_INC(tcps_idle_restarts);
4105 	if (tp->t_state == TCPS_ESTABLISHED)
4106 		KMOD_TCPSTAT_INC(tcps_idle_estrestarts);
4107 #endif
4108 	if (CC_ALGO(tp)->after_idle != NULL)
4109 		CC_ALGO(tp)->after_idle(tp->ccv);
4110 
4111 	if (tp->snd_cwnd == 1)
4112 		i_cwnd = tp->t_maxseg;		/* SYN(-ACK) lost */
4113 	else
4114 		i_cwnd = rc_init_window(rack);
4115 
4116 	/*
4117 	 * Being idle is no differnt than the initial window. If the cc
4118 	 * clamps it down below the initial window raise it to the initial
4119 	 * window.
4120 	 */
4121 	if (tp->snd_cwnd < i_cwnd) {
4122 		tp->snd_cwnd = i_cwnd;
4123 	}
4124 }
4125 
4126 /*
4127  * Indicate whether this ack should be delayed.  We can delay the ack if
4128  * following conditions are met:
4129  *	- There is no delayed ack timer in progress.
4130  *	- Our last ack wasn't a 0-sized window. We never want to delay
4131  *	  the ack that opens up a 0-sized window.
4132  *	- LRO wasn't used for this segment. We make sure by checking that the
4133  *	  segment size is not larger than the MSS.
4134  *	- Delayed acks are enabled or this is a half-synchronized T/TCP
4135  *	  connection.
4136  */
4137 #define DELAY_ACK(tp, tlen)			 \
4138 	(((tp->t_flags & TF_RXWIN0SENT) == 0) && \
4139 	((tp->t_flags & TF_DELACK) == 0) && 	 \
4140 	(tlen <= tp->t_maxseg) &&		 \
4141 	(tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN)))
4142 
4143 static struct rack_sendmap *
rack_find_lowest_rsm(struct tcp_rack * rack)4144 rack_find_lowest_rsm(struct tcp_rack *rack)
4145 {
4146 	struct rack_sendmap *rsm;
4147 
4148 	/*
4149 	 * Walk the time-order transmitted list looking for an rsm that is
4150 	 * not acked. This will be the one that was sent the longest time
4151 	 * ago that is still outstanding.
4152 	 */
4153 	TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
4154 		if (rsm->r_flags & RACK_ACKED) {
4155 			continue;
4156 		}
4157 		goto finish;
4158 	}
4159 finish:
4160 	return (rsm);
4161 }
4162 
4163 static struct rack_sendmap *
rack_find_high_nonack(struct tcp_rack * rack,struct rack_sendmap * rsm)4164 rack_find_high_nonack(struct tcp_rack *rack, struct rack_sendmap *rsm)
4165 {
4166 	struct rack_sendmap *prsm;
4167 
4168 	/*
4169 	 * Walk the sequence order list backward until we hit and arrive at
4170 	 * the highest seq not acked. In theory when this is called it
4171 	 * should be the last segment (which it was not).
4172 	 */
4173 	counter_u64_add(rack_find_high, 1);
4174 	prsm = rsm;
4175 	RB_FOREACH_REVERSE_FROM(prsm, rack_rb_tree_head, rsm) {
4176 		if (prsm->r_flags & (RACK_ACKED | RACK_HAS_FIN)) {
4177 			continue;
4178 		}
4179 		return (prsm);
4180 	}
4181 	return (NULL);
4182 }
4183 
4184 static uint32_t
rack_calc_thresh_rack(struct tcp_rack * rack,uint32_t srtt,uint32_t cts)4185 rack_calc_thresh_rack(struct tcp_rack *rack, uint32_t srtt, uint32_t cts)
4186 {
4187 	int32_t lro;
4188 	uint32_t thresh;
4189 
4190 	/*
4191 	 * lro is the flag we use to determine if we have seen reordering.
4192 	 * If it gets set we have seen reordering. The reorder logic either
4193 	 * works in one of two ways:
4194 	 *
4195 	 * If reorder-fade is configured, then we track the last time we saw
4196 	 * re-ordering occur. If we reach the point where enough time as
4197 	 * passed we no longer consider reordering has occuring.
4198 	 *
4199 	 * Or if reorder-face is 0, then once we see reordering we consider
4200 	 * the connection to alway be subject to reordering and just set lro
4201 	 * to 1.
4202 	 *
4203 	 * In the end if lro is non-zero we add the extra time for
4204 	 * reordering in.
4205 	 */
4206 	if (srtt == 0)
4207 		srtt = 1;
4208 	if (rack->r_ctl.rc_reorder_ts) {
4209 		if (rack->r_ctl.rc_reorder_fade) {
4210 			if (SEQ_GEQ(cts, rack->r_ctl.rc_reorder_ts)) {
4211 				lro = cts - rack->r_ctl.rc_reorder_ts;
4212 				if (lro == 0) {
4213 					/*
4214 					 * No time as passed since the last
4215 					 * reorder, mark it as reordering.
4216 					 */
4217 					lro = 1;
4218 				}
4219 			} else {
4220 				/* Negative time? */
4221 				lro = 0;
4222 			}
4223 			if (lro > rack->r_ctl.rc_reorder_fade) {
4224 				/* Turn off reordering seen too */
4225 				rack->r_ctl.rc_reorder_ts = 0;
4226 				lro = 0;
4227 			}
4228 		} else {
4229 			/* Reodering does not fade */
4230 			lro = 1;
4231 		}
4232 	} else {
4233 		lro = 0;
4234 	}
4235 	thresh = srtt + rack->r_ctl.rc_pkt_delay;
4236 	if (lro) {
4237 		/* It must be set, if not you get 1/4 rtt */
4238 		if (rack->r_ctl.rc_reorder_shift)
4239 			thresh += (srtt >> rack->r_ctl.rc_reorder_shift);
4240 		else
4241 			thresh += (srtt >> 2);
4242 	} else {
4243 		thresh += 1;
4244 	}
4245 	/* We don't let the rack timeout be above a RTO */
4246 	if (thresh > TICKS_2_MSEC(rack->rc_tp->t_rxtcur)) {
4247 		thresh = TICKS_2_MSEC(rack->rc_tp->t_rxtcur);
4248 	}
4249 	/* And we don't want it above the RTO max either */
4250 	if (thresh > rack_rto_max) {
4251 		thresh = rack_rto_max;
4252 	}
4253 	return (thresh);
4254 }
4255 
4256 static uint32_t
rack_calc_thresh_tlp(struct tcpcb * tp,struct tcp_rack * rack,struct rack_sendmap * rsm,uint32_t srtt)4257 rack_calc_thresh_tlp(struct tcpcb *tp, struct tcp_rack *rack,
4258 		     struct rack_sendmap *rsm, uint32_t srtt)
4259 {
4260 	struct rack_sendmap *prsm;
4261 	uint32_t thresh, len;
4262 	int segsiz;
4263 
4264 	if (srtt == 0)
4265 		srtt = 1;
4266 	if (rack->r_ctl.rc_tlp_threshold)
4267 		thresh = srtt + (srtt / rack->r_ctl.rc_tlp_threshold);
4268 	else
4269 		thresh = (srtt * 2);
4270 
4271 	/* Get the previous sent packet, if any  */
4272 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
4273 	counter_u64_add(rack_enter_tlp_calc, 1);
4274 	len = rsm->r_end - rsm->r_start;
4275 	if (rack->rack_tlp_threshold_use == TLP_USE_ID) {
4276 		/* Exactly like the ID */
4277 		if (((tp->snd_max - tp->snd_una) - rack->r_ctl.rc_sacked + rack->r_ctl.rc_holes_rxt) <= segsiz) {
4278 			uint32_t alt_thresh;
4279 			/*
4280 			 * Compensate for delayed-ack with the d-ack time.
4281 			 */
4282 			counter_u64_add(rack_used_tlpmethod, 1);
4283 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
4284 			if (alt_thresh > thresh)
4285 				thresh = alt_thresh;
4286 		}
4287 	} else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_ONE) {
4288 		/* 2.1 behavior */
4289 		prsm = TAILQ_PREV(rsm, rack_head, r_tnext);
4290 		if (prsm && (len <= segsiz)) {
4291 			/*
4292 			 * Two packets outstanding, thresh should be (2*srtt) +
4293 			 * possible inter-packet delay (if any).
4294 			 */
4295 			uint32_t inter_gap = 0;
4296 			int idx, nidx;
4297 
4298 			counter_u64_add(rack_used_tlpmethod, 1);
4299 			idx = rsm->r_rtr_cnt - 1;
4300 			nidx = prsm->r_rtr_cnt - 1;
4301 			if (TSTMP_GEQ(rsm->r_tim_lastsent[nidx], prsm->r_tim_lastsent[idx])) {
4302 				/* Yes it was sent later (or at the same time) */
4303 				inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx];
4304 			}
4305 			thresh += inter_gap;
4306 		} else 	if (len <= segsiz) {
4307 			/*
4308 			 * Possibly compensate for delayed-ack.
4309 			 */
4310 			uint32_t alt_thresh;
4311 
4312 			counter_u64_add(rack_used_tlpmethod2, 1);
4313 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
4314 			if (alt_thresh > thresh)
4315 				thresh = alt_thresh;
4316 		}
4317 	} else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_TWO) {
4318 		/* 2.2 behavior */
4319 		if (len <= segsiz) {
4320 			uint32_t alt_thresh;
4321 			/*
4322 			 * Compensate for delayed-ack with the d-ack time.
4323 			 */
4324 			counter_u64_add(rack_used_tlpmethod, 1);
4325 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
4326 			if (alt_thresh > thresh)
4327 				thresh = alt_thresh;
4328 		}
4329 	}
4330  	/* Not above an RTO */
4331 	if (thresh > TICKS_2_MSEC(tp->t_rxtcur)) {
4332 		thresh = TICKS_2_MSEC(tp->t_rxtcur);
4333 	}
4334 	/* Not above a RTO max */
4335 	if (thresh > rack_rto_max) {
4336 		thresh = rack_rto_max;
4337 	}
4338 	/* Apply user supplied min TLP */
4339 	if (thresh < rack_tlp_min) {
4340 		thresh = rack_tlp_min;
4341 	}
4342 	return (thresh);
4343 }
4344 
4345 static uint32_t
rack_grab_rtt(struct tcpcb * tp,struct tcp_rack * rack)4346 rack_grab_rtt(struct tcpcb *tp, struct tcp_rack *rack)
4347 {
4348 	/*
4349 	 * We want the rack_rtt which is the
4350 	 * last rtt we measured. However if that
4351 	 * does not exist we fallback to the srtt (which
4352 	 * we probably will never do) and then as a last
4353 	 * resort we use RACK_INITIAL_RTO if no srtt is
4354 	 * yet set.
4355 	 */
4356 	if (rack->rc_rack_rtt)
4357 		return(rack->rc_rack_rtt);
4358 	else if (tp->t_srtt == 0)
4359 		return(RACK_INITIAL_RTO);
4360 	return (TICKS_2_MSEC(tp->t_srtt >> TCP_RTT_SHIFT));
4361 }
4362 
4363 static struct rack_sendmap *
rack_check_recovery_mode(struct tcpcb * tp,uint32_t tsused)4364 rack_check_recovery_mode(struct tcpcb *tp, uint32_t tsused)
4365 {
4366 	/*
4367 	 * Check to see that we don't need to fall into recovery. We will
4368 	 * need to do so if our oldest transmit is past the time we should
4369 	 * have had an ack.
4370 	 */
4371 	struct tcp_rack *rack;
4372 	struct rack_sendmap *rsm;
4373 	int32_t idx;
4374 	uint32_t srtt, thresh;
4375 
4376 	rack = (struct tcp_rack *)tp->t_fb_ptr;
4377 	if (RB_EMPTY(&rack->r_ctl.rc_mtree)) {
4378 		return (NULL);
4379 	}
4380 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
4381 	if (rsm == NULL)
4382 		return (NULL);
4383 
4384 	if (rsm->r_flags & RACK_ACKED) {
4385 		rsm = rack_find_lowest_rsm(rack);
4386 		if (rsm == NULL)
4387 			return (NULL);
4388 	}
4389 	idx = rsm->r_rtr_cnt - 1;
4390 	srtt = rack_grab_rtt(tp, rack);
4391 	thresh = rack_calc_thresh_rack(rack, srtt, tsused);
4392 	if (TSTMP_LT(tsused, rsm->r_tim_lastsent[idx])) {
4393 		return (NULL);
4394 	}
4395 	if ((tsused - rsm->r_tim_lastsent[idx]) < thresh) {
4396 		return (NULL);
4397 	}
4398 	/* Ok if we reach here we are over-due and this guy can be sent */
4399 	if (IN_RECOVERY(tp->t_flags) == 0) {
4400 		/*
4401 		 * For the one that enters us into recovery record undo
4402 		 * info.
4403 		 */
4404 		rack->r_ctl.rc_rsm_start = rsm->r_start;
4405 		rack->r_ctl.rc_cwnd_at = tp->snd_cwnd;
4406 		rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh;
4407 	}
4408 	rack_cong_signal(tp, NULL, CC_NDUPACK);
4409 	return (rsm);
4410 }
4411 
4412 static uint32_t
rack_get_persists_timer_val(struct tcpcb * tp,struct tcp_rack * rack)4413 rack_get_persists_timer_val(struct tcpcb *tp, struct tcp_rack *rack)
4414 {
4415 	int32_t t;
4416 	int32_t tt;
4417 	uint32_t ret_val;
4418 
4419 	t = TICKS_2_MSEC((tp->t_srtt >> TCP_RTT_SHIFT) + ((tp->t_rttvar * 4) >> TCP_RTT_SHIFT));
4420 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
4421 	    rack_persist_min, rack_persist_max);
4422 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
4423 		tp->t_rxtshift++;
4424 	rack->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT;
4425 	ret_val = (uint32_t)tt;
4426 	return (ret_val);
4427 }
4428 
4429 static uint32_t
rack_timer_start(struct tcpcb * tp,struct tcp_rack * rack,uint32_t cts,int sup_rack)4430 rack_timer_start(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int sup_rack)
4431 {
4432 	/*
4433 	 * Start the FR timer, we do this based on getting the first one in
4434 	 * the rc_tmap. Note that if its NULL we must stop the timer. in all
4435 	 * events we need to stop the running timer (if its running) before
4436 	 * starting the new one.
4437 	 */
4438 	uint32_t thresh, exp, to, srtt, time_since_sent, tstmp_touse;
4439 	uint32_t srtt_cur;
4440 	int32_t idx;
4441 	int32_t is_tlp_timer = 0;
4442 	struct rack_sendmap *rsm;
4443 
4444 	if (rack->t_timers_stopped) {
4445 		/* All timers have been stopped none are to run */
4446 		return (0);
4447 	}
4448 	if (rack->rc_in_persist) {
4449 		/* We can't start any timer in persists */
4450 		return (rack_get_persists_timer_val(tp, rack));
4451 	}
4452 	rack->rc_on_min_to = 0;
4453 	if ((tp->t_state < TCPS_ESTABLISHED) ||
4454 	    ((tp->t_flags & TF_SACK_PERMIT) == 0))
4455 		goto activate_rxt;
4456 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
4457 	if ((rsm == NULL) || sup_rack) {
4458 		/* Nothing on the send map */
4459 activate_rxt:
4460 		time_since_sent = 0;
4461 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
4462 		if (rsm) {
4463 			idx = rsm->r_rtr_cnt - 1;
4464 			if (TSTMP_GEQ(rsm->r_tim_lastsent[idx], rack->r_ctl.rc_tlp_rxt_last_time))
4465 				tstmp_touse = rsm->r_tim_lastsent[idx];
4466 			else
4467 				tstmp_touse = rack->r_ctl.rc_tlp_rxt_last_time;
4468 			if (TSTMP_GT(cts, tstmp_touse))
4469 			    time_since_sent = cts - tstmp_touse;
4470 		}
4471 		if (SEQ_LT(tp->snd_una, tp->snd_max) || sbavail(&(tp->t_inpcb->inp_socket->so_snd))) {
4472 			rack->r_ctl.rc_hpts_flags |= PACE_TMR_RXT;
4473 			to = TICKS_2_MSEC(tp->t_rxtcur);
4474 			if (to > time_since_sent)
4475 				to -= time_since_sent;
4476 			else
4477 				to = rack->r_ctl.rc_min_to;
4478 			if (to == 0)
4479 				to = 1;
4480 			return (to);
4481 		}
4482 		return (0);
4483 	}
4484 	if (rsm->r_flags & RACK_ACKED) {
4485 		rsm = rack_find_lowest_rsm(rack);
4486 		if (rsm == NULL) {
4487 			/* No lowest? */
4488 			goto activate_rxt;
4489 		}
4490 	}
4491 	if (rack->sack_attack_disable) {
4492 		/*
4493 		 * We don't want to do
4494 		 * any TLP's if you are an attacker.
4495 		 * Though if you are doing what
4496 		 * is expected you may still have
4497 		 * SACK-PASSED marks.
4498 		 */
4499 		goto activate_rxt;
4500 	}
4501 	/* Convert from ms to usecs */
4502 	if ((rsm->r_flags & RACK_SACK_PASSED) || (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
4503 		if ((tp->t_flags & TF_SENTFIN) &&
4504 		    ((tp->snd_max - tp->snd_una) == 1) &&
4505 		    (rsm->r_flags & RACK_HAS_FIN)) {
4506 			/*
4507 			 * We don't start a rack timer if all we have is a
4508 			 * FIN outstanding.
4509 			 */
4510 			goto activate_rxt;
4511 		}
4512 		if ((rack->use_rack_rr == 0) &&
4513 		    (IN_RECOVERY(tp->t_flags)) &&
4514 		    (rack->rack_no_prr == 0) &&
4515 		     (rack->r_ctl.rc_prr_sndcnt  < ctf_fixed_maxseg(tp))) {
4516 			/*
4517 			 * We are not cheating, in recovery  and
4518 			 * not enough ack's to yet get our next
4519 			 * retransmission out.
4520 			 *
4521 			 * Note that classified attackers do not
4522 			 * get to use the rack-cheat.
4523 			 */
4524 			goto activate_tlp;
4525 		}
4526 		srtt = rack_grab_rtt(tp, rack);
4527 		thresh = rack_calc_thresh_rack(rack, srtt, cts);
4528 		idx = rsm->r_rtr_cnt - 1;
4529 		exp = rsm->r_tim_lastsent[idx] + thresh;
4530 		if (SEQ_GEQ(exp, cts)) {
4531 			to = exp - cts;
4532 			if (to < rack->r_ctl.rc_min_to) {
4533 				to = rack->r_ctl.rc_min_to;
4534 				if (rack->r_rr_config == 3)
4535 					rack->rc_on_min_to = 1;
4536 			}
4537 		} else {
4538 			to = rack->r_ctl.rc_min_to;
4539 			if (rack->r_rr_config == 3)
4540 				rack->rc_on_min_to = 1;
4541 		}
4542 	} else {
4543 		/* Ok we need to do a TLP not RACK */
4544 activate_tlp:
4545 		if ((rack->rc_tlp_in_progress != 0) &&
4546 		    (rack->r_ctl.rc_tlp_cnt_out >= rack_tlp_limit)) {
4547 			/*
4548 			 * The previous send was a TLP and we have sent
4549 			 * N TLP's without sending new data.
4550 			 */
4551 			goto activate_rxt;
4552 		}
4553 		rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
4554 		if (rsm == NULL) {
4555 			/* We found no rsm to TLP with. */
4556 			goto activate_rxt;
4557 		}
4558 		if (rsm->r_flags & RACK_HAS_FIN) {
4559 			/* If its a FIN we dont do TLP */
4560 			rsm = NULL;
4561 			goto activate_rxt;
4562 		}
4563 		idx = rsm->r_rtr_cnt - 1;
4564 		time_since_sent = 0;
4565 		if (TSTMP_GEQ(rsm->r_tim_lastsent[idx], rack->r_ctl.rc_tlp_rxt_last_time))
4566 			tstmp_touse = rsm->r_tim_lastsent[idx];
4567 		else
4568 			tstmp_touse = rack->r_ctl.rc_tlp_rxt_last_time;
4569 		if (TSTMP_GT(cts, tstmp_touse))
4570 		    time_since_sent = cts - tstmp_touse;
4571 		is_tlp_timer = 1;
4572 		if (tp->t_srtt) {
4573 			srtt_cur = (tp->t_srtt >> TCP_RTT_SHIFT);
4574 			srtt = TICKS_2_MSEC(srtt_cur);
4575 		} else
4576 			srtt = RACK_INITIAL_RTO;
4577 		/*
4578 		 * If the SRTT is not keeping up and the
4579 		 * rack RTT has spiked we want to use
4580 		 * the last RTT not the smoothed one.
4581 		 */
4582 		if (rack_tlp_use_greater && (srtt < rack_grab_rtt(tp, rack)))
4583 			srtt = rack_grab_rtt(tp, rack);
4584 		thresh = rack_calc_thresh_tlp(tp, rack, rsm, srtt);
4585 		if (thresh > time_since_sent)
4586 			to = thresh - time_since_sent;
4587 		else {
4588 			to = rack->r_ctl.rc_min_to;
4589 			rack_log_alt_to_to_cancel(rack,
4590 						  thresh,		/* flex1 */
4591 						  time_since_sent,	/* flex2 */
4592 						  tstmp_touse,		/* flex3 */
4593 						  rack->r_ctl.rc_tlp_rxt_last_time, /* flex4 */
4594 						  rsm->r_tim_lastsent[idx],
4595 						  srtt,
4596 						  idx, 99);
4597 		}
4598 		if (to > TCPTV_REXMTMAX) {
4599 			/*
4600 			 * If the TLP time works out to larger than the max
4601 			 * RTO lets not do TLP.. just RTO.
4602 			 */
4603 			goto activate_rxt;
4604 		}
4605 	}
4606 	if (is_tlp_timer == 0) {
4607 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_RACK;
4608 	} else {
4609 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_TLP;
4610 	}
4611 	if (to == 0)
4612 		to = 1;
4613 	return (to);
4614 }
4615 
4616 static void
rack_enter_persist(struct tcpcb * tp,struct tcp_rack * rack,uint32_t cts)4617 rack_enter_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
4618 {
4619 	if (rack->rc_in_persist == 0) {
4620 		if (tp->t_flags & TF_GPUTINPROG) {
4621 			/*
4622 			 * Stop the goodput now, the calling of the
4623 			 * measurement function clears the flag.
4624 			 */
4625 			rack_do_goodput_measurement(tp, rack, tp->snd_una, __LINE__);
4626 		}
4627 #ifdef NETFLIX_SHARED_CWND
4628 		if (rack->r_ctl.rc_scw) {
4629 			tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
4630 			rack->rack_scwnd_is_idle = 1;
4631 		}
4632 #endif
4633 		rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
4634 		if (rack->r_ctl.rc_went_idle_time == 0)
4635 			rack->r_ctl.rc_went_idle_time = 1;
4636 		rack_timer_cancel(tp, rack, cts, __LINE__);
4637 		tp->t_rxtshift = 0;
4638 		rack->rc_in_persist = 1;
4639 	}
4640 }
4641 
4642 static void
rack_exit_persist(struct tcpcb * tp,struct tcp_rack * rack,uint32_t cts)4643 rack_exit_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
4644 {
4645 	if (rack->rc_inp->inp_in_hpts)  {
4646 		tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT);
4647 		rack->r_ctl.rc_hpts_flags  = 0;
4648 	}
4649 #ifdef NETFLIX_SHARED_CWND
4650 	if (rack->r_ctl.rc_scw) {
4651 		tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
4652 		rack->rack_scwnd_is_idle = 0;
4653 	}
4654 #endif
4655 	if (rack->rc_gp_dyn_mul &&
4656 	    (rack->use_fixed_rate == 0) &&
4657 	    (rack->rc_always_pace)) {
4658 		/*
4659 		 * Do we count this as if a probe-rtt just
4660 		 * finished?
4661 		 */
4662 		uint32_t time_idle, idle_min;
4663 
4664 		time_idle = tcp_get_usecs(NULL) - rack->r_ctl.rc_went_idle_time;
4665 		idle_min = rack_min_probertt_hold;
4666 		if (rack_probertt_gpsrtt_cnt_div) {
4667 			uint64_t extra;
4668 			extra = (uint64_t)rack->r_ctl.rc_gp_srtt *
4669 				(uint64_t)rack_probertt_gpsrtt_cnt_mul;
4670 			extra /= (uint64_t)rack_probertt_gpsrtt_cnt_div;
4671 			idle_min += (uint32_t)extra;
4672 		}
4673 		if (time_idle >= idle_min)  {
4674 			/* Yes, we count it as a probe-rtt. */
4675 			uint32_t us_cts;
4676 
4677 			us_cts = tcp_get_usecs(NULL);
4678 			if (rack->in_probe_rtt == 0) {
4679 				rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
4680 				rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
4681 				rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
4682 				rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
4683 			} else {
4684 				rack_exit_probertt(rack, us_cts);
4685 			}
4686 		}
4687 	}
4688 	rack->rc_in_persist = 0;
4689 	rack->r_ctl.rc_went_idle_time = 0;
4690 	tp->t_rxtshift = 0;
4691  	rack->r_ctl.rc_agg_delayed = 0;
4692 	rack->r_early = 0;
4693 	rack->r_late = 0;
4694 	rack->r_ctl.rc_agg_early = 0;
4695 }
4696 
4697 static void
rack_log_hpts_diag(struct tcp_rack * rack,uint32_t cts,struct hpts_diag * diag,struct timeval * tv)4698 rack_log_hpts_diag(struct tcp_rack *rack, uint32_t cts,
4699 		   struct hpts_diag *diag, struct timeval *tv)
4700 {
4701 	if (rack_verbose_logging && rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
4702 		union tcp_log_stackspecific log;
4703 
4704 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
4705 		log.u_bbr.flex1 = diag->p_nxt_slot;
4706 		log.u_bbr.flex2 = diag->p_cur_slot;
4707 		log.u_bbr.flex3 = diag->slot_req;
4708 		log.u_bbr.flex4 = diag->inp_hptsslot;
4709 		log.u_bbr.flex5 = diag->slot_remaining;
4710 		log.u_bbr.flex6 = diag->need_new_to;
4711 		log.u_bbr.flex7 = diag->p_hpts_active;
4712 		log.u_bbr.flex8 = diag->p_on_min_sleep;
4713 		/* Hijack other fields as needed  */
4714 		log.u_bbr.epoch = diag->have_slept;
4715 		log.u_bbr.lt_epoch = diag->yet_to_sleep;
4716 		log.u_bbr.pkts_out = diag->co_ret;
4717 		log.u_bbr.applimited = diag->hpts_sleep_time;
4718 		log.u_bbr.delivered = diag->p_prev_slot;
4719 		log.u_bbr.inflight = diag->p_runningtick;
4720 		log.u_bbr.bw_inuse = diag->wheel_tick;
4721 		log.u_bbr.rttProp = diag->wheel_cts;
4722 		log.u_bbr.timeStamp = cts;
4723 		log.u_bbr.delRate = diag->maxticks;
4724 		log.u_bbr.cur_del_rate = diag->p_curtick;
4725 		log.u_bbr.cur_del_rate <<= 32;
4726 		log.u_bbr.cur_del_rate |= diag->p_lasttick;
4727 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
4728 		    &rack->rc_inp->inp_socket->so_rcv,
4729 		    &rack->rc_inp->inp_socket->so_snd,
4730 		    BBR_LOG_HPTSDIAG, 0,
4731 		    0, &log, false, tv);
4732 	}
4733 
4734 }
4735 
4736 static void
rack_start_hpts_timer(struct tcp_rack * rack,struct tcpcb * tp,uint32_t cts,int32_t slot,uint32_t tot_len_this_send,int sup_rack)4737 rack_start_hpts_timer(struct tcp_rack *rack, struct tcpcb *tp, uint32_t cts,
4738       int32_t slot, uint32_t tot_len_this_send, int sup_rack)
4739 {
4740 	struct hpts_diag diag;
4741 	struct inpcb *inp;
4742 	struct timeval tv;
4743 	uint32_t delayed_ack = 0;
4744 	uint32_t hpts_timeout;
4745 	uint8_t stopped;
4746 	uint32_t left = 0;
4747 	uint32_t us_cts;
4748 
4749 	inp = tp->t_inpcb;
4750 	if ((tp->t_state == TCPS_CLOSED) ||
4751 	    (tp->t_state == TCPS_LISTEN)) {
4752 		return;
4753 	}
4754 	if (inp->inp_in_hpts) {
4755 		/* Already on the pacer */
4756 		return;
4757 	}
4758 	stopped = rack->rc_tmr_stopped;
4759 	if (stopped && TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) {
4760 		left = rack->r_ctl.rc_timer_exp - cts;
4761 	}
4762 	rack->r_ctl.rc_timer_exp = 0;
4763 	rack->r_ctl.rc_hpts_flags = 0;
4764 	us_cts = tcp_get_usecs(&tv);
4765 	/* Now early/late accounting */
4766 	if (rack->r_early) {
4767 		/*
4768 		 * We have a early carry over set,
4769 		 * we can always add more time so we
4770 		 * can always make this compensation.
4771 		 */
4772 		slot += rack->r_ctl.rc_agg_early;
4773 		rack->r_early = 0;
4774 		rack->r_ctl.rc_agg_early = 0;
4775 	}
4776 	if (rack->r_late) {
4777 		/*
4778 		 * This is harder, we can
4779 		 * compensate some but it
4780 		 * really depends on what
4781 		 * the current pacing time is.
4782 		 */
4783 		if (rack->r_ctl.rc_agg_delayed >= slot) {
4784 			/*
4785 			 * We can't compensate for it all.
4786 			 * And we have to have some time
4787 			 * on the clock. We always have a min
4788 			 * 10 slots (10 x 10 i.e. 100 usecs).
4789 			 */
4790 			if (slot <= HPTS_TICKS_PER_USEC) {
4791 				/* We gain delay */
4792 				rack->r_ctl.rc_agg_delayed += (HPTS_TICKS_PER_USEC - slot);
4793 				slot = HPTS_TICKS_PER_USEC;
4794 			} else {
4795 				/* We take off some */
4796 				rack->r_ctl.rc_agg_delayed -= (slot - HPTS_TICKS_PER_USEC);
4797 				slot = HPTS_TICKS_PER_USEC;
4798 			}
4799 		} else {
4800 			slot -= rack->r_ctl.rc_agg_delayed;
4801 			rack->r_ctl.rc_agg_delayed = 0;
4802 			/* Make sure we have 100 useconds at minimum */
4803 			if (slot < HPTS_TICKS_PER_USEC) {
4804 				rack->r_ctl.rc_agg_delayed = HPTS_TICKS_PER_USEC - slot;
4805 				slot = HPTS_TICKS_PER_USEC;
4806 			}
4807 			if (rack->r_ctl.rc_agg_delayed == 0)
4808 				rack->r_late = 0;
4809 		}
4810 	}
4811 	if (slot) {
4812 		/* We are pacing too */
4813 		rack->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT;
4814 	}
4815 	hpts_timeout = rack_timer_start(tp, rack, cts, sup_rack);
4816 #ifdef NETFLIX_EXP_DETECTION
4817 	if (rack->sack_attack_disable &&
4818 	    (slot < tcp_sad_pacing_interval)) {
4819 		/*
4820 		 * We have a potential attacker on
4821 		 * the line. We have possibly some
4822 		 * (or now) pacing time set. We want to
4823 		 * slow down the processing of sacks by some
4824 		 * amount (if it is an attacker). Set the default
4825 		 * slot for attackers in place (unless the orginal
4826 		 * interval is longer). Its stored in
4827 		 * micro-seconds, so lets convert to msecs.
4828 		 */
4829 		slot = tcp_sad_pacing_interval;
4830 	}
4831 #endif
4832 	if (tp->t_flags & TF_DELACK) {
4833 		delayed_ack = TICKS_2_MSEC(tcp_delacktime);
4834 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_DELACK;
4835 	}
4836 	if (delayed_ack && ((hpts_timeout == 0) ||
4837 			    (delayed_ack < hpts_timeout)))
4838 		hpts_timeout = delayed_ack;
4839 	else
4840 		rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
4841 	/*
4842 	 * If no timers are going to run and we will fall off the hptsi
4843 	 * wheel, we resort to a keep-alive timer if its configured.
4844 	 */
4845 	if ((hpts_timeout == 0) &&
4846 	    (slot == 0)) {
4847 		if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
4848 		    (tp->t_state <= TCPS_CLOSING)) {
4849 			/*
4850 			 * Ok we have no timer (persists, rack, tlp, rxt  or
4851 			 * del-ack), we don't have segments being paced. So
4852 			 * all that is left is the keepalive timer.
4853 			 */
4854 			if (TCPS_HAVEESTABLISHED(tp->t_state)) {
4855 				/* Get the established keep-alive time */
4856 				hpts_timeout = TP_KEEPIDLE(tp);
4857 			} else {
4858 				/* Get the initial setup keep-alive time */
4859 				hpts_timeout = TP_KEEPINIT(tp);
4860 			}
4861 			rack->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP;
4862 			if (rack->in_probe_rtt) {
4863 				/*
4864 				 * We want to instead not wake up a long time from
4865 				 * now but to wake up about the time we would
4866 				 * exit probe-rtt and initiate a keep-alive ack.
4867 				 * This will get us out of probe-rtt and update
4868 				 * our min-rtt.
4869 				 */
4870 				hpts_timeout = (rack_min_probertt_hold / HPTS_USEC_IN_MSEC);
4871 			}
4872 		}
4873 	}
4874 	if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) ==
4875 	    (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) {
4876 		/*
4877 		 * RACK, TLP, persists and RXT timers all are restartable
4878 		 * based on actions input .. i.e we received a packet (ack
4879 		 * or sack) and that changes things (rw, or snd_una etc).
4880 		 * Thus we can restart them with a new value. For
4881 		 * keep-alive, delayed_ack we keep track of what was left
4882 		 * and restart the timer with a smaller value.
4883 		 */
4884 		if (left < hpts_timeout)
4885 			hpts_timeout = left;
4886 	}
4887 	if (hpts_timeout) {
4888 		/*
4889 		 * Hack alert for now we can't time-out over 2,147,483
4890 		 * seconds (a bit more than 596 hours), which is probably ok
4891 		 * :).
4892 		 */
4893 		if (hpts_timeout > 0x7ffffffe)
4894 			hpts_timeout = 0x7ffffffe;
4895 		rack->r_ctl.rc_timer_exp = cts + hpts_timeout;
4896 	}
4897 	if ((rack->rc_gp_filled == 0) &&
4898 	    (hpts_timeout < slot) &&
4899 	    (rack->r_ctl.rc_hpts_flags & (PACE_TMR_TLP|PACE_TMR_RXT))) {
4900 		/*
4901 		 * We have no good estimate yet for the
4902 		 * old clunky burst mitigation or the
4903 		 * real pacing. And the tlp or rxt is smaller
4904 		 * than the pacing calculation. Lets not
4905 		 * pace that long since we know the calculation
4906 		 * so far is not accurate.
4907 		 */
4908 		slot = hpts_timeout;
4909 	}
4910 	rack->r_ctl.last_pacing_time = slot;
4911 	if (slot) {
4912 		rack->r_ctl.rc_last_output_to = us_cts + slot;
4913 		if (rack->rc_always_pace || rack->r_mbuf_queue) {
4914 			if ((rack->rc_gp_filled == 0) ||
4915 			    rack->pacing_longer_than_rtt) {
4916 				inp->inp_flags2 &= ~(INP_DONT_SACK_QUEUE|INP_MBUF_QUEUE_READY);
4917 			} else {
4918 				inp->inp_flags2 |= INP_MBUF_QUEUE_READY;
4919 				if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK) &&
4920 				    (rack->r_rr_config != 3))
4921 					inp->inp_flags2 |= INP_DONT_SACK_QUEUE;
4922 				else
4923 					inp->inp_flags2 &= ~INP_DONT_SACK_QUEUE;
4924 			}
4925 		}
4926 		if ((rack->use_rack_rr) &&
4927 		    (rack->r_rr_config < 2) &&
4928 		    ((hpts_timeout) && ((hpts_timeout * HPTS_USEC_IN_MSEC) < slot))) {
4929 			/*
4930 			 * Arrange for the hpts to kick back in after the
4931 			 * t-o if the t-o does not cause a send.
4932 			 */
4933 			(void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_MS_TO_SLOTS(hpts_timeout),
4934 						   __LINE__, &diag);
4935 			rack_log_hpts_diag(rack, us_cts, &diag, &tv);
4936 			rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
4937 		} else {
4938 			(void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(slot),
4939 						   __LINE__, &diag);
4940 			rack_log_hpts_diag(rack, us_cts, &diag, &tv);
4941 			rack_log_to_start(rack, cts, hpts_timeout, slot, 1);
4942 		}
4943 	} else if (hpts_timeout) {
4944 		if (rack->rc_always_pace || rack->r_mbuf_queue) {
4945 			if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK)  {
4946 				/* For a rack timer, don't wake us */
4947 				inp->inp_flags2 |= INP_MBUF_QUEUE_READY;
4948 				if  (rack->r_rr_config != 3)
4949 					inp->inp_flags2 |= INP_DONT_SACK_QUEUE;
4950 				else
4951 					inp->inp_flags2 &= ~INP_DONT_SACK_QUEUE;
4952 			} else {
4953 				/* All other timers wake us up */
4954 				inp->inp_flags2 &= ~INP_MBUF_QUEUE_READY;
4955 				inp->inp_flags2 &= ~INP_DONT_SACK_QUEUE;
4956 			}
4957 		}
4958 		(void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_MS_TO_SLOTS(hpts_timeout),
4959 					   __LINE__, &diag);
4960 		rack_log_hpts_diag(rack, us_cts, &diag, &tv);
4961 		rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
4962 	} else {
4963 		/* No timer starting */
4964 #ifdef INVARIANTS
4965 		if (SEQ_GT(tp->snd_max, tp->snd_una)) {
4966 			panic("tp:%p rack:%p tlts:%d cts:%u slot:%u pto:%u -- no timer started?",
4967 			    tp, rack, tot_len_this_send, cts, slot, hpts_timeout);
4968 		}
4969 #endif
4970 	}
4971 	rack->rc_tmr_stopped = 0;
4972 	if (slot)
4973 		rack_log_type_bbrsnd(rack, tot_len_this_send, slot, us_cts, &tv);
4974 }
4975 
4976 /*
4977  * RACK Timer, here we simply do logging and house keeping.
4978  * the normal rack_output() function will call the
4979  * appropriate thing to check if we need to do a RACK retransmit.
4980  * We return 1, saying don't proceed with rack_output only
4981  * when all timers have been stopped (destroyed PCB?).
4982  */
4983 static int
rack_timeout_rack(struct tcpcb * tp,struct tcp_rack * rack,uint32_t cts)4984 rack_timeout_rack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
4985 {
4986 	/*
4987 	 * This timer simply provides an internal trigger to send out data.
4988 	 * The check_recovery_mode call will see if there are needed
4989 	 * retransmissions, if so we will enter fast-recovery. The output
4990 	 * call may or may not do the same thing depending on sysctl
4991 	 * settings.
4992 	 */
4993 	struct rack_sendmap *rsm;
4994 	int32_t recovery;
4995 
4996 	if (tp->t_timers->tt_flags & TT_STOPPED) {
4997 		return (1);
4998 	}
4999 	recovery = IN_RECOVERY(tp->t_flags);
5000 	counter_u64_add(rack_to_tot, 1);
5001 	if (rack->r_state && (rack->r_state != tp->t_state))
5002 		rack_set_state(tp, rack);
5003 	rack->rc_on_min_to = 0;
5004 	rsm = rack_check_recovery_mode(tp, cts);
5005 	rack_log_to_event(rack, RACK_TO_FRM_RACK, rsm);
5006 	if (rsm) {
5007 		uint32_t rtt;
5008 
5009 		rack->r_ctl.rc_resend = rsm;
5010 		if (rack->use_rack_rr) {
5011 			/*
5012 			 * Don't accumulate extra pacing delay
5013 			 * we are allowing the rack timer to
5014 			 * over-ride pacing i.e. rrr takes precedence
5015 			 * if the pacing interval is longer than the rrr
5016 			 * time (in other words we get the min pacing
5017 			 * time versus rrr pacing time).
5018 			 */
5019 			rack->r_timer_override = 1;
5020 			rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
5021 		}
5022 		rtt = rack->rc_rack_rtt;
5023 		if (rtt == 0)
5024 			rtt = 1;
5025 		if (rack->rack_no_prr == 0) {
5026 			if ((recovery == 0) &&
5027 			    (rack->r_ctl.rc_prr_sndcnt < ctf_fixed_maxseg(tp))) {
5028 				/*
5029 				 * The rack-timeout that enter's us into recovery
5030 				 * will force out one MSS and set us up so that we
5031 				 * can do one more send in 2*rtt (transitioning the
5032 				 * rack timeout into a rack-tlp).
5033 				 */
5034 				rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
5035 				rack->r_timer_override = 1;
5036 				rack_log_to_prr(rack, 3, 0);
5037 			} else if ((rack->r_ctl.rc_prr_sndcnt < (rsm->r_end - rsm->r_start)) &&
5038 				   rack->use_rack_rr) {
5039 				/*
5040 				 * When a rack timer goes, if the rack rr is
5041 				 * on, arrange it so we can send a full segment
5042 				 * overriding prr (though we pay a price for this
5043 				 * for future new sends).
5044 				 */
5045 				rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
5046 				rack_log_to_prr(rack, 4, 0);
5047 			}
5048 		}
5049 	}
5050 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK;
5051 	if (rsm == NULL) {
5052 		/* restart a timer and return 1 */
5053 		rack_start_hpts_timer(rack, tp, cts,
5054 				      0, 0, 0);
5055 		return (1);
5056 	}
5057 	return (0);
5058 }
5059 
5060 static __inline void
rack_clone_rsm(struct tcp_rack * rack,struct rack_sendmap * nrsm,struct rack_sendmap * rsm,uint32_t start)5061 rack_clone_rsm(struct tcp_rack *rack, struct rack_sendmap *nrsm,
5062 	       struct rack_sendmap *rsm, uint32_t start)
5063 {
5064 	int idx;
5065 
5066 	nrsm->r_start = start;
5067 	nrsm->r_end = rsm->r_end;
5068 	nrsm->r_rtr_cnt = rsm->r_rtr_cnt;
5069 	nrsm->r_flags = rsm->r_flags;
5070 	nrsm->r_dupack = rsm->r_dupack;
5071 	nrsm->usec_orig_send = rsm->usec_orig_send;
5072 	nrsm->r_rtr_bytes = 0;
5073 	rsm->r_end = nrsm->r_start;
5074 	nrsm->r_just_ret = rsm->r_just_ret;
5075 	for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) {
5076 		nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx];
5077 	}
5078 }
5079 
5080 static struct rack_sendmap *
rack_merge_rsm(struct tcp_rack * rack,struct rack_sendmap * l_rsm,struct rack_sendmap * r_rsm)5081 rack_merge_rsm(struct tcp_rack *rack,
5082 	       struct rack_sendmap *l_rsm,
5083 	       struct rack_sendmap *r_rsm)
5084 {
5085 	/*
5086 	 * We are merging two ack'd RSM's,
5087 	 * the l_rsm is on the left (lower seq
5088 	 * values) and the r_rsm is on the right
5089 	 * (higher seq value). The simplest way
5090 	 * to merge these is to move the right
5091 	 * one into the left. I don't think there
5092 	 * is any reason we need to try to find
5093 	 * the oldest (or last oldest retransmitted).
5094 	 */
5095 	struct rack_sendmap *rm;
5096 
5097 	l_rsm->r_end = r_rsm->r_end;
5098 	if (l_rsm->r_dupack < r_rsm->r_dupack)
5099 		l_rsm->r_dupack = r_rsm->r_dupack;
5100 	if (r_rsm->r_rtr_bytes)
5101 		l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes;
5102 	if (r_rsm->r_in_tmap) {
5103 		/* This really should not happen */
5104 		TAILQ_REMOVE(&rack->r_ctl.rc_tmap, r_rsm, r_tnext);
5105 		r_rsm->r_in_tmap = 0;
5106 	}
5107 
5108 	/* Now the flags */
5109 	if (r_rsm->r_flags & RACK_HAS_FIN)
5110 		l_rsm->r_flags |= RACK_HAS_FIN;
5111 	if (r_rsm->r_flags & RACK_TLP)
5112 		l_rsm->r_flags |= RACK_TLP;
5113 	if (r_rsm->r_flags & RACK_RWND_COLLAPSED)
5114 		l_rsm->r_flags |= RACK_RWND_COLLAPSED;
5115 	if ((r_rsm->r_flags & RACK_APP_LIMITED)  &&
5116 	    ((l_rsm->r_flags & RACK_APP_LIMITED) == 0)) {
5117 		/*
5118 		 * If both are app-limited then let the
5119 		 * free lower the count. If right is app
5120 		 * limited and left is not, transfer.
5121 		 */
5122 		l_rsm->r_flags |= RACK_APP_LIMITED;
5123 		r_rsm->r_flags &= ~RACK_APP_LIMITED;
5124 		if (r_rsm == rack->r_ctl.rc_first_appl)
5125 			rack->r_ctl.rc_first_appl = l_rsm;
5126 	}
5127 	rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, r_rsm);
5128 #ifdef INVARIANTS
5129 	if (rm != r_rsm) {
5130 		panic("removing head in rack:%p rsm:%p rm:%p",
5131 		      rack, r_rsm, rm);
5132 	}
5133 #endif
5134 	if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) {
5135 		/* Transfer the split limit to the map we free */
5136 		r_rsm->r_limit_type = l_rsm->r_limit_type;
5137 		l_rsm->r_limit_type = 0;
5138 	}
5139 	rack_free(rack, r_rsm);
5140 	return(l_rsm);
5141 }
5142 
5143 /*
5144  * TLP Timer, here we simply setup what segment we want to
5145  * have the TLP expire on, the normal rack_output() will then
5146  * send it out.
5147  *
5148  * We return 1, saying don't proceed with rack_output only
5149  * when all timers have been stopped (destroyed PCB?).
5150  */
5151 static int
rack_timeout_tlp(struct tcpcb * tp,struct tcp_rack * rack,uint32_t cts)5152 rack_timeout_tlp(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5153 {
5154 	/*
5155 	 * Tail Loss Probe.
5156 	 */
5157 	struct rack_sendmap *rsm = NULL;
5158 	struct rack_sendmap *insret;
5159 	struct socket *so;
5160 	uint32_t amm, old_prr_snd = 0;
5161 	uint32_t out, avail;
5162 	int collapsed_win = 0;
5163 
5164 	if (tp->t_timers->tt_flags & TT_STOPPED) {
5165 		return (1);
5166 	}
5167 	if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
5168 		/* Its not time yet */
5169 		return (0);
5170 	}
5171 	if (ctf_progress_timeout_check(tp, true)) {
5172 		rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
5173 		tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
5174 		return (1);
5175 	}
5176 	/*
5177 	 * A TLP timer has expired. We have been idle for 2 rtts. So we now
5178 	 * need to figure out how to force a full MSS segment out.
5179 	 */
5180 	rack_log_to_event(rack, RACK_TO_FRM_TLP, NULL);
5181 	counter_u64_add(rack_tlp_tot, 1);
5182 	if (rack->r_state && (rack->r_state != tp->t_state))
5183 		rack_set_state(tp, rack);
5184 	so = tp->t_inpcb->inp_socket;
5185 	avail = sbavail(&so->so_snd);
5186 	out = tp->snd_max - tp->snd_una;
5187 	if (out > tp->snd_wnd) {
5188 		/* special case, we need a retransmission */
5189 		collapsed_win = 1;
5190 		goto need_retran;
5191 	}
5192 	/*
5193 	 * Check our send oldest always settings, and if
5194 	 * there is an oldest to send jump to the need_retran.
5195 	 */
5196 	if (rack_always_send_oldest && (TAILQ_EMPTY(&rack->r_ctl.rc_tmap) == 0))
5197 		goto need_retran;
5198 
5199 	if (avail > out) {
5200 		/* New data is available */
5201 		amm = avail - out;
5202 		if (amm > ctf_fixed_maxseg(tp)) {
5203 			amm = ctf_fixed_maxseg(tp);
5204 			if ((amm + out) > tp->snd_wnd) {
5205 				/* We are rwnd limited */
5206 				goto need_retran;
5207 			}
5208 		} else if (amm < ctf_fixed_maxseg(tp)) {
5209 			/* not enough to fill a MTU */
5210 			goto need_retran;
5211 		}
5212 		if (IN_RECOVERY(tp->t_flags)) {
5213 			/* Unlikely */
5214 			if (rack->rack_no_prr == 0) {
5215 				old_prr_snd = rack->r_ctl.rc_prr_sndcnt;
5216 				if (out + amm <= tp->snd_wnd) {
5217 					rack->r_ctl.rc_prr_sndcnt = amm;
5218 					rack_log_to_prr(rack, 4, 0);
5219 				}
5220 			} else
5221 				goto need_retran;
5222 		} else {
5223 			/* Set the send-new override */
5224 			if (out + amm <= tp->snd_wnd)
5225 				rack->r_ctl.rc_tlp_new_data = amm;
5226 			else
5227 				goto need_retran;
5228 		}
5229 		rack->r_ctl.rc_tlpsend = NULL;
5230 		counter_u64_add(rack_tlp_newdata, 1);
5231 		goto send;
5232 	}
5233 need_retran:
5234 	/*
5235 	 * Ok we need to arrange the last un-acked segment to be re-sent, or
5236 	 * optionally the first un-acked segment.
5237 	 */
5238 	if (collapsed_win == 0) {
5239 		if (rack_always_send_oldest)
5240 			rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5241 		else {
5242 			rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
5243 			if (rsm && (rsm->r_flags & (RACK_ACKED | RACK_HAS_FIN))) {
5244 				rsm = rack_find_high_nonack(rack, rsm);
5245 			}
5246 		}
5247 		if (rsm == NULL) {
5248 			counter_u64_add(rack_tlp_does_nada, 1);
5249 #ifdef TCP_BLACKBOX
5250 			tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
5251 #endif
5252 			goto out;
5253 		}
5254 	} else {
5255 		/*
5256 		 * We must find the last segment
5257 		 * that was acceptable by the client.
5258 		 */
5259 		RB_FOREACH_REVERSE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
5260 			if ((rsm->r_flags & RACK_RWND_COLLAPSED) == 0) {
5261 				/* Found one */
5262 				break;
5263 			}
5264 		}
5265 		if (rsm == NULL) {
5266 			/* None? if so send the first */
5267 			rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
5268 			if (rsm == NULL) {
5269 				counter_u64_add(rack_tlp_does_nada, 1);
5270 #ifdef TCP_BLACKBOX
5271 				tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
5272 #endif
5273 				goto out;
5274 			}
5275 		}
5276 	}
5277 	if ((rsm->r_end - rsm->r_start) > ctf_fixed_maxseg(tp)) {
5278 		/*
5279 		 * We need to split this the last segment in two.
5280 		 */
5281 		struct rack_sendmap *nrsm;
5282 
5283 		nrsm = rack_alloc_full_limit(rack);
5284 		if (nrsm == NULL) {
5285 			/*
5286 			 * No memory to split, we will just exit and punt
5287 			 * off to the RXT timer.
5288 			 */
5289 			counter_u64_add(rack_tlp_does_nada, 1);
5290 			goto out;
5291 		}
5292 		rack_clone_rsm(rack, nrsm, rsm,
5293 			       (rsm->r_end - ctf_fixed_maxseg(tp)));
5294 		insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
5295 #ifdef INVARIANTS
5296 		if (insret != NULL) {
5297 			panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
5298 			      nrsm, insret, rack, rsm);
5299 		}
5300 #endif
5301 		if (rsm->r_in_tmap) {
5302 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
5303 			nrsm->r_in_tmap = 1;
5304 		}
5305 		rsm->r_flags &= (~RACK_HAS_FIN);
5306 		rsm = nrsm;
5307 	}
5308 	rack->r_ctl.rc_tlpsend = rsm;
5309 send:
5310 	rack->r_timer_override = 1;
5311 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
5312 	return (0);
5313 out:
5314 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
5315 	return (0);
5316 }
5317 
5318 /*
5319  * Delayed ack Timer, here we simply need to setup the
5320  * ACK_NOW flag and remove the DELACK flag. From there
5321  * the output routine will send the ack out.
5322  *
5323  * We only return 1, saying don't proceed, if all timers
5324  * are stopped (destroyed PCB?).
5325  */
5326 static int
rack_timeout_delack(struct tcpcb * tp,struct tcp_rack * rack,uint32_t cts)5327 rack_timeout_delack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5328 {
5329 	if (tp->t_timers->tt_flags & TT_STOPPED) {
5330 		return (1);
5331 	}
5332 	rack_log_to_event(rack, RACK_TO_FRM_DELACK, NULL);
5333 	tp->t_flags &= ~TF_DELACK;
5334 	tp->t_flags |= TF_ACKNOW;
5335 	KMOD_TCPSTAT_INC(tcps_delack);
5336 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
5337 	return (0);
5338 }
5339 
5340 /*
5341  * Persists timer, here we simply send the
5342  * same thing as a keepalive will.
5343  * the one byte send.
5344  *
5345  * We only return 1, saying don't proceed, if all timers
5346  * are stopped (destroyed PCB?).
5347  */
5348 static int
rack_timeout_persist(struct tcpcb * tp,struct tcp_rack * rack,uint32_t cts)5349 rack_timeout_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5350 {
5351 	struct tcptemp *t_template;
5352 	struct inpcb *inp;
5353 	int32_t retval = 1;
5354 
5355 	inp = tp->t_inpcb;
5356 
5357 	if (tp->t_timers->tt_flags & TT_STOPPED) {
5358 		return (1);
5359 	}
5360 	if (rack->rc_in_persist == 0)
5361 		return (0);
5362 	if (ctf_progress_timeout_check(tp, false)) {
5363 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
5364 		rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
5365 		tcp_set_inp_to_drop(inp, ETIMEDOUT);
5366 		return (1);
5367 	}
5368 	KASSERT(inp != NULL, ("%s: tp %p tp->t_inpcb == NULL", __func__, tp));
5369 	/*
5370 	 * Persistence timer into zero window. Force a byte to be output, if
5371 	 * possible.
5372 	 */
5373 	KMOD_TCPSTAT_INC(tcps_persisttimeo);
5374 	/*
5375 	 * Hack: if the peer is dead/unreachable, we do not time out if the
5376 	 * window is closed.  After a full backoff, drop the connection if
5377 	 * the idle time (no responses to probes) reaches the maximum
5378 	 * backoff that we would use if retransmitting.
5379 	 */
5380 	if (tp->t_rxtshift == TCP_MAXRXTSHIFT &&
5381 	    (ticks - tp->t_rcvtime >= tcp_maxpersistidle ||
5382 	    ticks - tp->t_rcvtime >= TCP_REXMTVAL(tp) * tcp_totbackoff)) {
5383 		KMOD_TCPSTAT_INC(tcps_persistdrop);
5384 		retval = 1;
5385 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
5386 		tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT);
5387 		goto out;
5388 	}
5389 	if ((sbavail(&rack->rc_inp->inp_socket->so_snd) == 0) &&
5390 	    tp->snd_una == tp->snd_max)
5391 		rack_exit_persist(tp, rack, cts);
5392 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT;
5393 	/*
5394 	 * If the user has closed the socket then drop a persisting
5395 	 * connection after a much reduced timeout.
5396 	 */
5397 	if (tp->t_state > TCPS_CLOSE_WAIT &&
5398 	    (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) {
5399 		retval = 1;
5400 		KMOD_TCPSTAT_INC(tcps_persistdrop);
5401 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
5402 		tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT);
5403 		goto out;
5404 	}
5405 	t_template = tcpip_maketemplate(rack->rc_inp);
5406 	if (t_template) {
5407 		/* only set it if we were answered */
5408 		if (rack->forced_ack == 0) {
5409 			rack->forced_ack = 1;
5410 			rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
5411 		}
5412 		tcp_respond(tp, t_template->tt_ipgen,
5413 			    &t_template->tt_t, (struct mbuf *)NULL,
5414 			    tp->rcv_nxt, tp->snd_una - 1, 0);
5415 		/* This sends an ack */
5416 		if (tp->t_flags & TF_DELACK)
5417 			tp->t_flags &= ~TF_DELACK;
5418 		free(t_template, M_TEMP);
5419 	}
5420 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
5421 		tp->t_rxtshift++;
5422 out:
5423 	rack_log_to_event(rack, RACK_TO_FRM_PERSIST, NULL);
5424 	rack_start_hpts_timer(rack, tp, cts,
5425 			      0, 0, 0);
5426 	return (retval);
5427 }
5428 
5429 /*
5430  * If a keepalive goes off, we had no other timers
5431  * happening. We always return 1 here since this
5432  * routine either drops the connection or sends
5433  * out a segment with respond.
5434  */
5435 static int
rack_timeout_keepalive(struct tcpcb * tp,struct tcp_rack * rack,uint32_t cts)5436 rack_timeout_keepalive(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5437 {
5438 	struct tcptemp *t_template;
5439 	struct inpcb *inp;
5440 
5441 	if (tp->t_timers->tt_flags & TT_STOPPED) {
5442 		return (1);
5443 	}
5444 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP;
5445 	inp = tp->t_inpcb;
5446 	rack_log_to_event(rack, RACK_TO_FRM_KEEP, NULL);
5447 	/*
5448 	 * Keep-alive timer went off; send something or drop connection if
5449 	 * idle for too long.
5450 	 */
5451 	KMOD_TCPSTAT_INC(tcps_keeptimeo);
5452 	if (tp->t_state < TCPS_ESTABLISHED)
5453 		goto dropit;
5454 	if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
5455 	    tp->t_state <= TCPS_CLOSING) {
5456 		if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp))
5457 			goto dropit;
5458 		/*
5459 		 * Send a packet designed to force a response if the peer is
5460 		 * up and reachable: either an ACK if the connection is
5461 		 * still alive, or an RST if the peer has closed the
5462 		 * connection due to timeout or reboot. Using sequence
5463 		 * number tp->snd_una-1 causes the transmitted zero-length
5464 		 * segment to lie outside the receive window; by the
5465 		 * protocol spec, this requires the correspondent TCP to
5466 		 * respond.
5467 		 */
5468 		KMOD_TCPSTAT_INC(tcps_keepprobe);
5469 		t_template = tcpip_maketemplate(inp);
5470 		if (t_template) {
5471 			if (rack->forced_ack == 0) {
5472 				rack->forced_ack = 1;
5473 				rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
5474 			}
5475 			tcp_respond(tp, t_template->tt_ipgen,
5476 			    &t_template->tt_t, (struct mbuf *)NULL,
5477 			    tp->rcv_nxt, tp->snd_una - 1, 0);
5478 			free(t_template, M_TEMP);
5479 		}
5480 	}
5481 	rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
5482 	return (1);
5483 dropit:
5484 	KMOD_TCPSTAT_INC(tcps_keepdrops);
5485 	tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
5486 	tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT);
5487 	return (1);
5488 }
5489 
5490 /*
5491  * Retransmit helper function, clear up all the ack
5492  * flags and take care of important book keeping.
5493  */
5494 static void
rack_remxt_tmr(struct tcpcb * tp)5495 rack_remxt_tmr(struct tcpcb *tp)
5496 {
5497 	/*
5498 	 * The retransmit timer went off, all sack'd blocks must be
5499 	 * un-acked.
5500 	 */
5501 	struct rack_sendmap *rsm, *trsm = NULL;
5502 	struct tcp_rack *rack;
5503 	int32_t cnt = 0;
5504 
5505 	rack = (struct tcp_rack *)tp->t_fb_ptr;
5506 	rack_timer_cancel(tp, rack, tcp_ts_getticks(), __LINE__);
5507 	rack_log_to_event(rack, RACK_TO_FRM_TMR, NULL);
5508 	if (rack->r_state && (rack->r_state != tp->t_state))
5509 		rack_set_state(tp, rack);
5510 	/*
5511 	 * Ideally we would like to be able to
5512 	 * mark SACK-PASS on anything not acked here.
5513 	 * However, if we do that we would burst out
5514 	 * all that data 1ms apart. This would be unwise,
5515 	 * so for now we will just let the normal rxt timer
5516 	 * and tlp timer take care of it.
5517 	 */
5518 	RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
5519 		if (rsm->r_flags & RACK_ACKED) {
5520 			cnt++;
5521 			rsm->r_dupack = 0;
5522 			rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
5523 			if (rsm->r_in_tmap == 0) {
5524 				/* We must re-add it back to the tlist */
5525 				if (trsm == NULL) {
5526 					TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
5527 				} else {
5528 					TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, trsm, rsm, r_tnext);
5529 				}
5530 				rsm->r_in_tmap = 1;
5531 			}
5532 		}
5533 		trsm = rsm;
5534 		if (rsm->r_flags & RACK_ACKED)
5535 			rsm->r_flags |= RACK_WAS_ACKED;
5536 		rsm->r_flags &= ~(RACK_ACKED | RACK_SACK_PASSED | RACK_WAS_SACKPASS);
5537 	}
5538 	/* Clear the count (we just un-acked them) */
5539 	rack->r_ctl.rc_sacked = 0;
5540 	rack->r_ctl.rc_agg_delayed = 0;
5541 	rack->r_early = 0;
5542 	rack->r_ctl.rc_agg_early = 0;
5543 	rack->r_late = 0;
5544 	/* Clear the tlp rtx mark */
5545 	rack->r_ctl.rc_resend = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
5546 	rack->r_ctl.rc_prr_sndcnt = 0;
5547 	rack_log_to_prr(rack, 6, 0);
5548 	rack->r_timer_override = 1;
5549 }
5550 
5551 static void
rack_cc_conn_init(struct tcpcb * tp)5552 rack_cc_conn_init(struct tcpcb *tp)
5553 {
5554 	struct tcp_rack *rack;
5555 
5556 	rack = (struct tcp_rack *)tp->t_fb_ptr;
5557 	cc_conn_init(tp);
5558 	/*
5559 	 * We want a chance to stay in slowstart as
5560 	 * we create a connection. TCP spec says that
5561 	 * initially ssthresh is infinite. For our
5562 	 * purposes that is the snd_wnd.
5563 	 */
5564 	if (tp->snd_ssthresh < tp->snd_wnd) {
5565 		tp->snd_ssthresh = tp->snd_wnd;
5566 	}
5567 	/*
5568 	 * We also want to assure a IW worth of
5569 	 * data can get inflight.
5570 	 */
5571 	if (rc_init_window(rack) < tp->snd_cwnd)
5572 		tp->snd_cwnd = rc_init_window(rack);
5573 }
5574 
5575 /*
5576  * Re-transmit timeout! If we drop the PCB we will return 1, otherwise
5577  * we will setup to retransmit the lowest seq number outstanding.
5578  */
5579 static int
rack_timeout_rxt(struct tcpcb * tp,struct tcp_rack * rack,uint32_t cts)5580 rack_timeout_rxt(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5581 {
5582 	int32_t rexmt;
5583 	struct inpcb *inp;
5584 	int32_t retval = 0;
5585 	bool isipv6;
5586 
5587 	inp = tp->t_inpcb;
5588 	if (tp->t_timers->tt_flags & TT_STOPPED) {
5589 		return (1);
5590 	}
5591 	if (ctf_progress_timeout_check(tp, false)) {
5592 		tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
5593 		rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
5594 		tcp_set_inp_to_drop(inp, ETIMEDOUT);
5595 		return (1);
5596 	}
5597 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT;
5598 	if (TCPS_HAVEESTABLISHED(tp->t_state) &&
5599 	    (tp->snd_una == tp->snd_max)) {
5600 		/* Nothing outstanding .. nothing to do */
5601 		return (0);
5602 	}
5603 	/*
5604 	 * Retransmission timer went off.  Message has not been acked within
5605 	 * retransmit interval.  Back off to a longer retransmit interval
5606 	 * and retransmit one segment.
5607 	 */
5608 	rack_remxt_tmr(tp);
5609 	if ((rack->r_ctl.rc_resend == NULL) ||
5610 	    ((rack->r_ctl.rc_resend->r_flags & RACK_RWND_COLLAPSED) == 0)) {
5611 		/*
5612 		 * If the rwnd collapsed on
5613 		 * the one we are retransmitting
5614 		 * it does not count against the
5615 		 * rxt count.
5616 		 */
5617 		tp->t_rxtshift++;
5618 	}
5619 	if (tp->t_rxtshift > TCP_MAXRXTSHIFT) {
5620 		tp->t_rxtshift = TCP_MAXRXTSHIFT;
5621 		KMOD_TCPSTAT_INC(tcps_timeoutdrop);
5622 		retval = 1;
5623 		tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
5624 		tcp_set_inp_to_drop(rack->rc_inp,
5625 		    (tp->t_softerror ? (uint16_t) tp->t_softerror : ETIMEDOUT));
5626 		goto out;
5627 	}
5628 	if (tp->t_state == TCPS_SYN_SENT) {
5629 		/*
5630 		 * If the SYN was retransmitted, indicate CWND to be limited
5631 		 * to 1 segment in cc_conn_init().
5632 		 */
5633 		tp->snd_cwnd = 1;
5634 	} else if (tp->t_rxtshift == 1) {
5635 		/*
5636 		 * first retransmit; record ssthresh and cwnd so they can be
5637 		 * recovered if this turns out to be a "bad" retransmit. A
5638 		 * retransmit is considered "bad" if an ACK for this segment
5639 		 * is received within RTT/2 interval; the assumption here is
5640 		 * that the ACK was already in flight.  See "On Estimating
5641 		 * End-to-End Network Path Properties" by Allman and Paxson
5642 		 * for more details.
5643 		 */
5644 		tp->snd_cwnd_prev = tp->snd_cwnd;
5645 		tp->snd_ssthresh_prev = tp->snd_ssthresh;
5646 		tp->snd_recover_prev = tp->snd_recover;
5647 		if (IN_FASTRECOVERY(tp->t_flags))
5648 			tp->t_flags |= TF_WASFRECOVERY;
5649 		else
5650 			tp->t_flags &= ~TF_WASFRECOVERY;
5651 		if (IN_CONGRECOVERY(tp->t_flags))
5652 			tp->t_flags |= TF_WASCRECOVERY;
5653 		else
5654 			tp->t_flags &= ~TF_WASCRECOVERY;
5655 		tp->t_badrxtwin = ticks + (tp->t_srtt >> (TCP_RTT_SHIFT + 1));
5656 		tp->t_flags |= TF_PREVVALID;
5657 	} else
5658 		tp->t_flags &= ~TF_PREVVALID;
5659 	KMOD_TCPSTAT_INC(tcps_rexmttimeo);
5660 	if ((tp->t_state == TCPS_SYN_SENT) ||
5661 	    (tp->t_state == TCPS_SYN_RECEIVED))
5662 		rexmt = MSEC_2_TICKS(RACK_INITIAL_RTO * tcp_backoff[tp->t_rxtshift]);
5663 	else
5664 		rexmt = TCP_REXMTVAL(tp) * tcp_backoff[tp->t_rxtshift];
5665 	TCPT_RANGESET(tp->t_rxtcur, rexmt,
5666 	   max(MSEC_2_TICKS(rack_rto_min), rexmt),
5667 	   MSEC_2_TICKS(rack_rto_max));
5668 	/*
5669 	 * We enter the path for PLMTUD if connection is established or, if
5670 	 * connection is FIN_WAIT_1 status, reason for the last is that if
5671 	 * amount of data we send is very small, we could send it in couple
5672 	 * of packets and process straight to FIN. In that case we won't
5673 	 * catch ESTABLISHED state.
5674 	 */
5675 #ifdef INET6
5676 	isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) ? true : false;
5677 #else
5678 	isipv6 = false;
5679 #endif
5680 	if (((V_tcp_pmtud_blackhole_detect == 1) ||
5681 	    (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) ||
5682 	    (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) &&
5683 	    ((tp->t_state == TCPS_ESTABLISHED) ||
5684 	    (tp->t_state == TCPS_FIN_WAIT_1))) {
5685 		/*
5686 		 * Idea here is that at each stage of mtu probe (usually,
5687 		 * 1448 -> 1188 -> 524) should be given 2 chances to recover
5688 		 * before further clamping down. 'tp->t_rxtshift % 2 == 0'
5689 		 * should take care of that.
5690 		 */
5691 		if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) ==
5692 		    (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) &&
5693 		    (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 &&
5694 		    tp->t_rxtshift % 2 == 0)) {
5695 			/*
5696 			 * Enter Path MTU Black-hole Detection mechanism: -
5697 			 * Disable Path MTU Discovery (IP "DF" bit). -
5698 			 * Reduce MTU to lower value than what we negotiated
5699 			 * with peer.
5700 			 */
5701 			if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) {
5702 				/* Record that we may have found a black hole. */
5703 				tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE;
5704 				/* Keep track of previous MSS. */
5705 				tp->t_pmtud_saved_maxseg = tp->t_maxseg;
5706 			}
5707 
5708 			/*
5709 			 * Reduce the MSS to blackhole value or to the
5710 			 * default in an attempt to retransmit.
5711 			 */
5712 #ifdef INET6
5713 			if (isipv6 &&
5714 			    tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) {
5715 				/* Use the sysctl tuneable blackhole MSS. */
5716 				tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss;
5717 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
5718 			} else if (isipv6) {
5719 				/* Use the default MSS. */
5720 				tp->t_maxseg = V_tcp_v6mssdflt;
5721 				/*
5722 				 * Disable Path MTU Discovery when we switch
5723 				 * to minmss.
5724 				 */
5725 				tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
5726 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
5727 			}
5728 #endif
5729 #if defined(INET6) && defined(INET)
5730 			else
5731 #endif
5732 #ifdef INET
5733 			if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) {
5734 				/* Use the sysctl tuneable blackhole MSS. */
5735 				tp->t_maxseg = V_tcp_pmtud_blackhole_mss;
5736 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
5737 			} else {
5738 				/* Use the default MSS. */
5739 				tp->t_maxseg = V_tcp_mssdflt;
5740 				/*
5741 				 * Disable Path MTU Discovery when we switch
5742 				 * to minmss.
5743 				 */
5744 				tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
5745 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
5746 			}
5747 #endif
5748 		} else {
5749 			/*
5750 			 * If further retransmissions are still unsuccessful
5751 			 * with a lowered MTU, maybe this isn't a blackhole
5752 			 * and we restore the previous MSS and blackhole
5753 			 * detection flags. The limit '6' is determined by
5754 			 * giving each probe stage (1448, 1188, 524) 2
5755 			 * chances to recover.
5756 			 */
5757 			if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) &&
5758 			    (tp->t_rxtshift >= 6)) {
5759 				tp->t_flags2 |= TF2_PLPMTU_PMTUD;
5760 				tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE;
5761 				tp->t_maxseg = tp->t_pmtud_saved_maxseg;
5762 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_failed);
5763 			}
5764 		}
5765 	}
5766 	/*
5767 	 * If we backed off this far, our srtt estimate is probably bogus.
5768 	 * Clobber it so we'll take the next rtt measurement as our srtt;
5769 	 * move the current srtt into rttvar to keep the current retransmit
5770 	 * times until then.
5771 	 */
5772 	if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) {
5773 #ifdef INET6
5774 		if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0)
5775 			in6_losing(tp->t_inpcb);
5776 		else
5777 #endif
5778 			in_losing(tp->t_inpcb);
5779 		tp->t_rttvar += (tp->t_srtt >> TCP_RTT_SHIFT);
5780 		tp->t_srtt = 0;
5781 	}
5782 	sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
5783 	tp->snd_recover = tp->snd_max;
5784 	tp->t_flags |= TF_ACKNOW;
5785 	tp->t_rtttime = 0;
5786 	rack_cong_signal(tp, NULL, CC_RTO);
5787 out:
5788 	return (retval);
5789 }
5790 
5791 static int
rack_process_timers(struct tcpcb * tp,struct tcp_rack * rack,uint32_t cts,uint8_t hpts_calling)5792 rack_process_timers(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t hpts_calling)
5793 {
5794 	int32_t ret = 0;
5795 	int32_t timers = (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK);
5796 
5797 	if (timers == 0) {
5798 		return (0);
5799 	}
5800 	if (tp->t_state == TCPS_LISTEN) {
5801 		/* no timers on listen sockets */
5802 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)
5803 			return (0);
5804 		return (1);
5805 	}
5806 	if ((timers & PACE_TMR_RACK) &&
5807 	    rack->rc_on_min_to) {
5808 		/*
5809 		 * For the rack timer when we
5810 		 * are on a min-timeout (which means rrr_conf = 3)
5811 		 * we don't want to check the timer. It may
5812 		 * be going off for a pace and thats ok we
5813 		 * want to send the retransmit (if its ready).
5814 		 *
5815 		 * If its on a normal rack timer (non-min) then
5816 		 * we will check if its expired.
5817 		 */
5818 		goto skip_time_check;
5819 	}
5820 	if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
5821 		uint32_t left;
5822 
5823 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
5824 			ret = -1;
5825 			rack_log_to_processing(rack, cts, ret, 0);
5826 			return (0);
5827 		}
5828 		if (hpts_calling == 0) {
5829 			/*
5830 			 * A user send or queued mbuf (sack) has called us? We
5831 			 * return 0 and let the pacing guards
5832 			 * deal with it if they should or
5833 			 * should not cause a send.
5834 			 */
5835 			ret = -2;
5836 			rack_log_to_processing(rack, cts, ret, 0);
5837 			return (0);
5838 		}
5839 		/*
5840 		 * Ok our timer went off early and we are not paced false
5841 		 * alarm, go back to sleep.
5842 		 */
5843 		ret = -3;
5844 		left = rack->r_ctl.rc_timer_exp - cts;
5845 		tcp_hpts_insert(tp->t_inpcb, HPTS_MS_TO_SLOTS(left));
5846 		rack_log_to_processing(rack, cts, ret, left);
5847 		return (1);
5848 	}
5849 skip_time_check:
5850 	rack->rc_tmr_stopped = 0;
5851 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK;
5852 	if (timers & PACE_TMR_DELACK) {
5853 		ret = rack_timeout_delack(tp, rack, cts);
5854 	} else if (timers & PACE_TMR_RACK) {
5855 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
5856 		ret = rack_timeout_rack(tp, rack, cts);
5857 	} else if (timers & PACE_TMR_TLP) {
5858 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
5859 		ret = rack_timeout_tlp(tp, rack, cts);
5860 	} else if (timers & PACE_TMR_RXT) {
5861 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
5862 		ret = rack_timeout_rxt(tp, rack, cts);
5863 	} else if (timers & PACE_TMR_PERSIT) {
5864 		ret = rack_timeout_persist(tp, rack, cts);
5865 	} else if (timers & PACE_TMR_KEEP) {
5866 		ret = rack_timeout_keepalive(tp, rack, cts);
5867 	}
5868 	rack_log_to_processing(rack, cts, ret, timers);
5869 	return (ret);
5870 }
5871 
5872 static void
rack_timer_cancel(struct tcpcb * tp,struct tcp_rack * rack,uint32_t cts,int line)5873 rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line)
5874 {
5875 	struct timeval tv;
5876 	uint32_t us_cts, flags_on_entry;
5877 	uint8_t hpts_removed = 0;
5878 
5879 	flags_on_entry = rack->r_ctl.rc_hpts_flags;
5880 	us_cts = tcp_get_usecs(&tv);
5881 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
5882 	    ((TSTMP_GEQ(us_cts, rack->r_ctl.rc_last_output_to)) ||
5883 	     ((tp->snd_max - tp->snd_una) == 0))) {
5884 		tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT);
5885 		hpts_removed = 1;
5886 		/* If we were not delayed cancel out the flag. */
5887 		if ((tp->snd_max - tp->snd_una) == 0)
5888 			rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
5889 		rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
5890 	}
5891 	if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
5892 		rack->rc_tmr_stopped = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
5893 		if (rack->rc_inp->inp_in_hpts &&
5894 		    ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)) {
5895 			/*
5896 			 * Canceling timer's when we have no output being
5897 			 * paced. We also must remove ourselves from the
5898 			 * hpts.
5899 			 */
5900 			tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT);
5901 			hpts_removed = 1;
5902 		}
5903 		rack->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK);
5904 	}
5905 	if (hpts_removed == 0)
5906 		rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
5907 }
5908 
5909 static void
rack_timer_stop(struct tcpcb * tp,uint32_t timer_type)5910 rack_timer_stop(struct tcpcb *tp, uint32_t timer_type)
5911 {
5912 	return;
5913 }
5914 
5915 static int
rack_stopall(struct tcpcb * tp)5916 rack_stopall(struct tcpcb *tp)
5917 {
5918 	struct tcp_rack *rack;
5919 	rack = (struct tcp_rack *)tp->t_fb_ptr;
5920 	rack->t_timers_stopped = 1;
5921 	return (0);
5922 }
5923 
5924 static void
rack_timer_activate(struct tcpcb * tp,uint32_t timer_type,uint32_t delta)5925 rack_timer_activate(struct tcpcb *tp, uint32_t timer_type, uint32_t delta)
5926 {
5927 	return;
5928 }
5929 
5930 static int
rack_timer_active(struct tcpcb * tp,uint32_t timer_type)5931 rack_timer_active(struct tcpcb *tp, uint32_t timer_type)
5932 {
5933 	return (0);
5934 }
5935 
5936 static void
rack_stop_all_timers(struct tcpcb * tp)5937 rack_stop_all_timers(struct tcpcb *tp)
5938 {
5939 	struct tcp_rack *rack;
5940 
5941 	/*
5942 	 * Assure no timers are running.
5943 	 */
5944 	if (tcp_timer_active(tp, TT_PERSIST)) {
5945 		/* We enter in persists, set the flag appropriately */
5946 		rack = (struct tcp_rack *)tp->t_fb_ptr;
5947 		rack->rc_in_persist = 1;
5948 	}
5949 	tcp_timer_suspend(tp, TT_PERSIST);
5950 	tcp_timer_suspend(tp, TT_REXMT);
5951 	tcp_timer_suspend(tp, TT_KEEP);
5952 	tcp_timer_suspend(tp, TT_DELACK);
5953 }
5954 
5955 static void
rack_update_rsm(struct tcpcb * tp,struct tcp_rack * rack,struct rack_sendmap * rsm,uint32_t ts)5956 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
5957     struct rack_sendmap *rsm, uint32_t ts)
5958 {
5959 	int32_t idx;
5960 
5961 	rsm->r_rtr_cnt++;
5962 	rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
5963 	rsm->r_dupack = 0;
5964 	if (rsm->r_rtr_cnt > RACK_NUM_OF_RETRANS) {
5965 		rsm->r_rtr_cnt = RACK_NUM_OF_RETRANS;
5966 		rsm->r_flags |= RACK_OVERMAX;
5967 	}
5968 	if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & RACK_TLP) == 0)) {
5969 		rack->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start);
5970 		rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start);
5971 	}
5972 	idx = rsm->r_rtr_cnt - 1;
5973 	rsm->r_tim_lastsent[idx] = ts;
5974 	if (rsm->r_flags & RACK_ACKED) {
5975 		/* Problably MTU discovery messing with us */
5976 		rsm->r_flags &= ~RACK_ACKED;
5977 		rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
5978 	}
5979 	if (rsm->r_in_tmap) {
5980 		TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
5981 		rsm->r_in_tmap = 0;
5982 	}
5983 	TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
5984 	rsm->r_in_tmap = 1;
5985 	if (rsm->r_flags & RACK_SACK_PASSED) {
5986 		/* We have retransmitted due to the SACK pass */
5987 		rsm->r_flags &= ~RACK_SACK_PASSED;
5988 		rsm->r_flags |= RACK_WAS_SACKPASS;
5989 	}
5990 }
5991 
5992 static uint32_t
rack_update_entry(struct tcpcb * tp,struct tcp_rack * rack,struct rack_sendmap * rsm,uint32_t ts,int32_t * lenp)5993 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
5994     struct rack_sendmap *rsm, uint32_t ts, int32_t *lenp)
5995 {
5996 	/*
5997 	 * We (re-)transmitted starting at rsm->r_start for some length
5998 	 * (possibly less than r_end.
5999 	 */
6000 	struct rack_sendmap *nrsm, *insret;
6001 	uint32_t c_end;
6002 	int32_t len;
6003 
6004 	len = *lenp;
6005 	c_end = rsm->r_start + len;
6006 	if (SEQ_GEQ(c_end, rsm->r_end)) {
6007 		/*
6008 		 * We retransmitted the whole piece or more than the whole
6009 		 * slopping into the next rsm.
6010 		 */
6011 		rack_update_rsm(tp, rack, rsm, ts);
6012 		if (c_end == rsm->r_end) {
6013 			*lenp = 0;
6014 			return (0);
6015 		} else {
6016 			int32_t act_len;
6017 
6018 			/* Hangs over the end return whats left */
6019 			act_len = rsm->r_end - rsm->r_start;
6020 			*lenp = (len - act_len);
6021 			return (rsm->r_end);
6022 		}
6023 		/* We don't get out of this block. */
6024 	}
6025 	/*
6026 	 * Here we retransmitted less than the whole thing which means we
6027 	 * have to split this into what was transmitted and what was not.
6028 	 */
6029 	nrsm = rack_alloc_full_limit(rack);
6030 	if (nrsm == NULL) {
6031 		/*
6032 		 * We can't get memory, so lets not proceed.
6033 		 */
6034 		*lenp = 0;
6035 		return (0);
6036 	}
6037 	/*
6038 	 * So here we are going to take the original rsm and make it what we
6039 	 * retransmitted. nrsm will be the tail portion we did not
6040 	 * retransmit. For example say the chunk was 1, 11 (10 bytes). And
6041 	 * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to
6042 	 * 1, 6 and the new piece will be 6, 11.
6043 	 */
6044 	rack_clone_rsm(rack, nrsm, rsm, c_end);
6045 	nrsm->r_dupack = 0;
6046 	rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
6047 	insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
6048 #ifdef INVARIANTS
6049 	if (insret != NULL) {
6050 		panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
6051 		      nrsm, insret, rack, rsm);
6052 	}
6053 #endif
6054 	if (rsm->r_in_tmap) {
6055 		TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
6056 		nrsm->r_in_tmap = 1;
6057 	}
6058 	rsm->r_flags &= (~RACK_HAS_FIN);
6059 	rack_update_rsm(tp, rack, rsm, ts);
6060 	*lenp = 0;
6061 	return (0);
6062 }
6063 
6064 static void
rack_log_output(struct tcpcb * tp,struct tcpopt * to,int32_t len,uint32_t seq_out,uint8_t th_flags,int32_t err,uint32_t ts,uint8_t pass,struct rack_sendmap * hintrsm,uint32_t us_cts)6065 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
6066     uint32_t seq_out, uint8_t th_flags, int32_t err, uint32_t ts,
6067     uint8_t pass, struct rack_sendmap *hintrsm, uint32_t us_cts)
6068 {
6069 	struct tcp_rack *rack;
6070 	struct rack_sendmap *rsm, *nrsm, *insret, fe;
6071 	register uint32_t snd_max, snd_una;
6072 
6073 	/*
6074 	 * Add to the RACK log of packets in flight or retransmitted. If
6075 	 * there is a TS option we will use the TS echoed, if not we will
6076 	 * grab a TS.
6077 	 *
6078 	 * Retransmissions will increment the count and move the ts to its
6079 	 * proper place. Note that if options do not include TS's then we
6080 	 * won't be able to effectively use the ACK for an RTT on a retran.
6081 	 *
6082 	 * Notes about r_start and r_end. Lets consider a send starting at
6083 	 * sequence 1 for 10 bytes. In such an example the r_start would be
6084 	 * 1 (starting sequence) but the r_end would be r_start+len i.e. 11.
6085 	 * This means that r_end is actually the first sequence for the next
6086 	 * slot (11).
6087 	 *
6088 	 */
6089 	/*
6090 	 * If err is set what do we do XXXrrs? should we not add the thing?
6091 	 * -- i.e. return if err != 0 or should we pretend we sent it? --
6092 	 * i.e. proceed with add ** do this for now.
6093 	 */
6094 	INP_WLOCK_ASSERT(tp->t_inpcb);
6095 	if (err)
6096 		/*
6097 		 * We don't log errors -- we could but snd_max does not
6098 		 * advance in this case either.
6099 		 */
6100 		return;
6101 
6102 	if (th_flags & TH_RST) {
6103 		/*
6104 		 * We don't log resets and we return immediately from
6105 		 * sending
6106 		 */
6107 		return;
6108 	}
6109 	rack = (struct tcp_rack *)tp->t_fb_ptr;
6110 	snd_una = tp->snd_una;
6111 	if (SEQ_LEQ((seq_out + len), snd_una)) {
6112 		/* Are sending an old segment to induce an ack (keep-alive)? */
6113 		return;
6114 	}
6115 	if (SEQ_LT(seq_out, snd_una)) {
6116 		/* huh? should we panic? */
6117 		uint32_t end;
6118 
6119 		end = seq_out + len;
6120 		seq_out = snd_una;
6121 		if (SEQ_GEQ(end, seq_out))
6122 			len = end - seq_out;
6123 		else
6124 			len = 0;
6125 	}
6126 	snd_max = tp->snd_max;
6127 	if (th_flags & (TH_SYN | TH_FIN)) {
6128 		/*
6129 		 * The call to rack_log_output is made before bumping
6130 		 * snd_max. This means we can record one extra byte on a SYN
6131 		 * or FIN if seq_out is adding more on and a FIN is present
6132 		 * (and we are not resending).
6133 		 */
6134 		if ((th_flags & TH_SYN) && (seq_out == tp->iss))
6135 			len++;
6136 		if (th_flags & TH_FIN)
6137 			len++;
6138 		if (SEQ_LT(snd_max, tp->snd_nxt)) {
6139 			/*
6140 			 * The add/update as not been done for the FIN/SYN
6141 			 * yet.
6142 			 */
6143 			snd_max = tp->snd_nxt;
6144 		}
6145 	}
6146 	if (len == 0) {
6147 		/* We don't log zero window probes */
6148 		return;
6149 	}
6150 	rack->r_ctl.rc_time_last_sent = ts;
6151 	if (IN_RECOVERY(tp->t_flags)) {
6152 		rack->r_ctl.rc_prr_out += len;
6153 	}
6154 	/* First question is it a retransmission or new? */
6155 	if (seq_out == snd_max) {
6156 		/* Its new */
6157 again:
6158 		rsm = rack_alloc(rack);
6159 		if (rsm == NULL) {
6160 			/*
6161 			 * Hmm out of memory and the tcb got destroyed while
6162 			 * we tried to wait.
6163 			 */
6164 			return;
6165 		}
6166 		if (th_flags & TH_FIN) {
6167 			rsm->r_flags = RACK_HAS_FIN;
6168 		} else {
6169 			rsm->r_flags = 0;
6170 		}
6171 		rsm->r_tim_lastsent[0] = ts;
6172 		rsm->r_rtr_cnt = 1;
6173 		rsm->r_rtr_bytes = 0;
6174 		rsm->usec_orig_send = us_cts;
6175 		if (th_flags & TH_SYN) {
6176 			/* The data space is one beyond snd_una */
6177 			rsm->r_flags |= RACK_HAS_SIN;
6178 			rsm->r_start = seq_out + 1;
6179 			rsm->r_end = rsm->r_start + (len - 1);
6180 		} else {
6181 			/* Normal case */
6182 			rsm->r_start = seq_out;
6183 			rsm->r_end = rsm->r_start + len;
6184 		}
6185 		rsm->r_dupack = 0;
6186 		rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
6187 		insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
6188 #ifdef INVARIANTS
6189 		if (insret != NULL) {
6190 			panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
6191 			      nrsm, insret, rack, rsm);
6192 		}
6193 #endif
6194 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
6195 		rsm->r_in_tmap = 1;
6196 		/*
6197 		 * Special case detection, is there just a single
6198 		 * packet outstanding when we are not in recovery?
6199 		 *
6200 		 * If this is true mark it so.
6201 		 */
6202 		if ((IN_RECOVERY(tp->t_flags) == 0) &&
6203 		    (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) == ctf_fixed_maxseg(tp))) {
6204 			struct rack_sendmap *prsm;
6205 
6206 			prsm = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
6207 			if (prsm)
6208 				prsm->r_one_out_nr = 1;
6209 		}
6210 		return;
6211 	}
6212 	/*
6213 	 * If we reach here its a retransmission and we need to find it.
6214 	 */
6215 	memset(&fe, 0, sizeof(fe));
6216 more:
6217 	if (hintrsm && (hintrsm->r_start == seq_out)) {
6218 		rsm = hintrsm;
6219 		hintrsm = NULL;
6220 	} else {
6221 		/* No hints sorry */
6222 		rsm = NULL;
6223 	}
6224 	if ((rsm) && (rsm->r_start == seq_out)) {
6225 		seq_out = rack_update_entry(tp, rack, rsm, ts, &len);
6226 		if (len == 0) {
6227 			return;
6228 		} else {
6229 			goto more;
6230 		}
6231 	}
6232 	/* Ok it was not the last pointer go through it the hard way. */
6233 refind:
6234 	fe.r_start = seq_out;
6235 	rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
6236 	if (rsm) {
6237 		if (rsm->r_start == seq_out) {
6238 			seq_out = rack_update_entry(tp, rack, rsm, ts, &len);
6239 			if (len == 0) {
6240 				return;
6241 			} else {
6242 				goto refind;
6243 			}
6244 		}
6245 		if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) {
6246 			/* Transmitted within this piece */
6247 			/*
6248 			 * Ok we must split off the front and then let the
6249 			 * update do the rest
6250 			 */
6251 			nrsm = rack_alloc_full_limit(rack);
6252 			if (nrsm == NULL) {
6253 				rack_update_rsm(tp, rack, rsm, ts);
6254 				return;
6255 			}
6256 			/*
6257 			 * copy rsm to nrsm and then trim the front of rsm
6258 			 * to not include this part.
6259 			 */
6260 			rack_clone_rsm(rack, nrsm, rsm, seq_out);
6261 			insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
6262 #ifdef INVARIANTS
6263 			if (insret != NULL) {
6264 				panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
6265 				      nrsm, insret, rack, rsm);
6266 			}
6267 #endif
6268 			if (rsm->r_in_tmap) {
6269 				TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
6270 				nrsm->r_in_tmap = 1;
6271 			}
6272 			rsm->r_flags &= (~RACK_HAS_FIN);
6273 			seq_out = rack_update_entry(tp, rack, nrsm, ts, &len);
6274 			if (len == 0) {
6275 				return;
6276 			} else if (len > 0)
6277 				goto refind;
6278 		}
6279 	}
6280 	/*
6281 	 * Hmm not found in map did they retransmit both old and on into the
6282 	 * new?
6283 	 */
6284 	if (seq_out == tp->snd_max) {
6285 		goto again;
6286 	} else if (SEQ_LT(seq_out, tp->snd_max)) {
6287 #ifdef INVARIANTS
6288 		printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n",
6289 		    seq_out, len, tp->snd_una, tp->snd_max);
6290 		printf("Starting Dump of all rack entries\n");
6291 		RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
6292 			printf("rsm:%p start:%u end:%u\n",
6293 			    rsm, rsm->r_start, rsm->r_end);
6294 		}
6295 		printf("Dump complete\n");
6296 		panic("seq_out not found rack:%p tp:%p",
6297 		    rack, tp);
6298 #endif
6299 	} else {
6300 #ifdef INVARIANTS
6301 		/*
6302 		 * Hmm beyond sndmax? (only if we are using the new rtt-pack
6303 		 * flag)
6304 		 */
6305 		panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p",
6306 		    seq_out, len, tp->snd_max, tp);
6307 #endif
6308 	}
6309 }
6310 
6311 /*
6312  * Record one of the RTT updates from an ack into
6313  * our sample structure.
6314  */
6315 
6316 static void
tcp_rack_xmit_timer(struct tcp_rack * rack,int32_t rtt,uint32_t len,uint32_t us_rtt,int confidence,struct rack_sendmap * rsm,uint16_t rtrcnt)6317 tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt, uint32_t len, uint32_t us_rtt,
6318 		    int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt)
6319 {
6320 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
6321 	    (rack->r_ctl.rack_rs.rs_rtt_lowest > rtt)) {
6322 		rack->r_ctl.rack_rs.rs_rtt_lowest = rtt;
6323 	}
6324 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
6325 	    (rack->r_ctl.rack_rs.rs_rtt_highest < rtt)) {
6326 		rack->r_ctl.rack_rs.rs_rtt_highest = rtt;
6327 	}
6328 	if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
6329 	    if (us_rtt < rack->r_ctl.rc_gp_lowrtt)
6330 		rack->r_ctl.rc_gp_lowrtt = us_rtt;
6331 	    if (rack->rc_tp->snd_wnd > rack->r_ctl.rc_gp_high_rwnd)
6332 		    rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
6333 	}
6334 	if ((confidence == 1) &&
6335 	    ((rsm == NULL) ||
6336 	     (rsm->r_just_ret) ||
6337 	     (rsm->r_one_out_nr &&
6338 	      len < (ctf_fixed_maxseg(rack->rc_tp) * 2)))) {
6339 		/*
6340 		 * If the rsm had a just return
6341 		 * hit it then we can't trust the
6342 		 * rtt measurement for buffer deterimination
6343 		 * Note that a confidence of 2, indicates
6344 		 * SACK'd which overrides the r_just_ret or
6345 		 * the r_one_out_nr. If it was a CUM-ACK and
6346 		 * we had only two outstanding, but get an
6347 		 * ack for only 1. Then that also lowers our
6348 		 * confidence.
6349 		 */
6350 		confidence = 0;
6351 	}
6352 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
6353 	    (rack->r_ctl.rack_rs.rs_us_rtt > us_rtt)) {
6354 		if (rack->r_ctl.rack_rs.confidence == 0) {
6355 			/*
6356 			 * We take anything with no current confidence
6357 			 * saved.
6358 			 */
6359 			rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
6360 			rack->r_ctl.rack_rs.confidence = confidence;
6361 			rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
6362 		} else if (confidence || rack->r_ctl.rack_rs.confidence) {
6363 			/*
6364 			 * Once we have a confident number,
6365 			 * we can update it with a smaller
6366 			 * value since this confident number
6367 			 * may include the DSACK time until
6368 			 * the next segment (the second one) arrived.
6369 			 */
6370 			rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
6371 			rack->r_ctl.rack_rs.confidence = confidence;
6372 			rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
6373 		}
6374 	}
6375 	rack_log_rtt_upd(rack->rc_tp, rack, us_rtt, len, rsm, confidence);
6376 	rack->r_ctl.rack_rs.rs_flags = RACK_RTT_VALID;
6377 	rack->r_ctl.rack_rs.rs_rtt_tot += rtt;
6378 	rack->r_ctl.rack_rs.rs_rtt_cnt++;
6379 }
6380 
6381 /*
6382  * Collect new round-trip time estimate
6383  * and update averages and current timeout.
6384  */
6385 static void
tcp_rack_xmit_timer_commit(struct tcp_rack * rack,struct tcpcb * tp)6386 tcp_rack_xmit_timer_commit(struct tcp_rack *rack, struct tcpcb *tp)
6387 {
6388 	int32_t delta;
6389 	uint32_t o_srtt, o_var;
6390 	int32_t hrtt_up = 0;
6391 	int32_t rtt;
6392 
6393 	if (rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY)
6394 		/* No valid sample */
6395 		return;
6396 	if (rack->r_ctl.rc_rate_sample_method == USE_RTT_LOW) {
6397 		/* We are to use the lowest RTT seen in a single ack */
6398 		rtt = rack->r_ctl.rack_rs.rs_rtt_lowest;
6399 	} else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_HIGH) {
6400 		/* We are to use the highest RTT seen in a single ack */
6401 		rtt = rack->r_ctl.rack_rs.rs_rtt_highest;
6402 	} else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_AVG) {
6403 		/* We are to use the average RTT seen in a single ack */
6404 		rtt = (int32_t)(rack->r_ctl.rack_rs.rs_rtt_tot /
6405 				(uint64_t)rack->r_ctl.rack_rs.rs_rtt_cnt);
6406 	} else {
6407 #ifdef INVARIANTS
6408 		panic("Unknown rtt variant %d", rack->r_ctl.rc_rate_sample_method);
6409 #endif
6410 		return;
6411 	}
6412 	if (rtt == 0)
6413 		rtt = 1;
6414 	if (rack->rc_gp_rtt_set == 0) {
6415 		/*
6416 		 * With no RTT we have to accept
6417 		 * even one we are not confident of.
6418 		 */
6419 		rack->r_ctl.rc_gp_srtt = rack->r_ctl.rack_rs.rs_us_rtt;
6420 		rack->rc_gp_rtt_set = 1;
6421 	} else if (rack->r_ctl.rack_rs.confidence) {
6422 		/* update the running gp srtt */
6423 		rack->r_ctl.rc_gp_srtt -= (rack->r_ctl.rc_gp_srtt/8);
6424 		rack->r_ctl.rc_gp_srtt += rack->r_ctl.rack_rs.rs_us_rtt / 8;
6425 	}
6426 	if (rack->r_ctl.rack_rs.confidence) {
6427 		/*
6428 		 * record the low and high for highly buffered path computation,
6429 		 * we only do this if we are confident (not a retransmission).
6430 		 */
6431 		if (rack->r_ctl.rc_highest_us_rtt < rack->r_ctl.rack_rs.rs_us_rtt) {
6432 			rack->r_ctl.rc_highest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
6433 			hrtt_up = 1;
6434 		}
6435 		if (rack->rc_highly_buffered == 0) {
6436 			/*
6437 			 * Currently once we declare a path has
6438 			 * highly buffered there is no going
6439 			 * back, which may be a problem...
6440 			 */
6441 			if ((rack->r_ctl.rc_highest_us_rtt / rack->r_ctl.rc_lowest_us_rtt) > rack_hbp_thresh) {
6442 				rack_log_rtt_shrinks(rack, rack->r_ctl.rack_rs.rs_us_rtt,
6443 						     rack->r_ctl.rc_highest_us_rtt,
6444 						     rack->r_ctl.rc_lowest_us_rtt,
6445 						     RACK_RTTS_SEEHBP);
6446 				rack->rc_highly_buffered = 1;
6447 			}
6448 		}
6449 	}
6450 	if ((rack->r_ctl.rack_rs.confidence) ||
6451 	    (rack->r_ctl.rack_rs.rs_us_rtrcnt == 1)) {
6452 		/*
6453 		 * If we are highly confident of it <or> it was
6454 		 * never retransmitted we accept it as the last us_rtt.
6455 		 */
6456 		rack->r_ctl.rc_last_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
6457 		/* The lowest rtt can be set if its was not retransmited */
6458 		if (rack->r_ctl.rc_lowest_us_rtt > rack->r_ctl.rack_rs.rs_us_rtt) {
6459 			rack->r_ctl.rc_lowest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
6460 			if (rack->r_ctl.rc_lowest_us_rtt == 0)
6461 				rack->r_ctl.rc_lowest_us_rtt = 1;
6462 		}
6463 	}
6464 	rack_log_rtt_sample(rack, rtt);
6465 	o_srtt = tp->t_srtt;
6466 	o_var = tp->t_rttvar;
6467 	rack = (struct tcp_rack *)tp->t_fb_ptr;
6468 	if (tp->t_srtt != 0) {
6469 		/*
6470 		 * srtt is stored as fixed point with 5 bits after the
6471 		 * binary point (i.e., scaled by 8).  The following magic is
6472 		 * equivalent to the smoothing algorithm in rfc793 with an
6473 		 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point).
6474 		 * Adjust rtt to origin 0.
6475 		 */
6476 		delta = ((rtt - 1) << TCP_DELTA_SHIFT)
6477 		    - (tp->t_srtt >> (TCP_RTT_SHIFT - TCP_DELTA_SHIFT));
6478 
6479 		tp->t_srtt += delta;
6480 		if (tp->t_srtt <= 0)
6481 			tp->t_srtt = 1;
6482 
6483 		/*
6484 		 * We accumulate a smoothed rtt variance (actually, a
6485 		 * smoothed mean difference), then set the retransmit timer
6486 		 * to smoothed rtt + 4 times the smoothed variance. rttvar
6487 		 * is stored as fixed point with 4 bits after the binary
6488 		 * point (scaled by 16).  The following is equivalent to
6489 		 * rfc793 smoothing with an alpha of .75 (rttvar =
6490 		 * rttvar*3/4 + |delta| / 4).  This replaces rfc793's
6491 		 * wired-in beta.
6492 		 */
6493 		if (delta < 0)
6494 			delta = -delta;
6495 		delta -= tp->t_rttvar >> (TCP_RTTVAR_SHIFT - TCP_DELTA_SHIFT);
6496 		tp->t_rttvar += delta;
6497 		if (tp->t_rttvar <= 0)
6498 			tp->t_rttvar = 1;
6499 		if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar)
6500 			tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
6501 	} else {
6502 		/*
6503 		 * No rtt measurement yet - use the unsmoothed rtt. Set the
6504 		 * variance to half the rtt (so our first retransmit happens
6505 		 * at 3*rtt).
6506 		 */
6507 		tp->t_srtt = rtt << TCP_RTT_SHIFT;
6508 		tp->t_rttvar = rtt << (TCP_RTTVAR_SHIFT - 1);
6509 		tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
6510 	}
6511 	KMOD_TCPSTAT_INC(tcps_rttupdated);
6512 	tp->t_rttupdated++;
6513 #ifdef STATS
6514 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt));
6515 #endif
6516 	tp->t_rxtshift = 0;
6517 
6518 	/*
6519 	 * the retransmit should happen at rtt + 4 * rttvar. Because of the
6520 	 * way we do the smoothing, srtt and rttvar will each average +1/2
6521 	 * tick of bias.  When we compute the retransmit timer, we want 1/2
6522 	 * tick of rounding and 1 extra tick because of +-1/2 tick
6523 	 * uncertainty in the firing of the timer.  The bias will give us
6524 	 * exactly the 1.5 tick we need.  But, because the bias is
6525 	 * statistical, we have to test that we don't drop below the minimum
6526 	 * feasible timer (which is 2 ticks).
6527 	 */
6528 	TCPT_RANGESET(tp->t_rxtcur, TCP_REXMTVAL(tp),
6529 	   max(MSEC_2_TICKS(rack_rto_min), rtt + 2), MSEC_2_TICKS(rack_rto_max));
6530 	tp->t_softerror = 0;
6531 }
6532 
6533 static void
rack_earlier_retran(struct tcpcb * tp,struct rack_sendmap * rsm,uint32_t t,uint32_t cts)6534 rack_earlier_retran(struct tcpcb *tp, struct rack_sendmap *rsm,
6535     uint32_t t, uint32_t cts)
6536 {
6537 	/*
6538 	 * For this RSM, we acknowledged the data from a previous
6539 	 * transmission, not the last one we made. This means we did a false
6540 	 * retransmit.
6541 	 */
6542 	struct tcp_rack *rack;
6543 
6544 	if (rsm->r_flags & RACK_HAS_FIN) {
6545 		/*
6546 		 * The sending of the FIN often is multiple sent when we
6547 		 * have everything outstanding ack'd. We ignore this case
6548 		 * since its over now.
6549 		 */
6550 		return;
6551 	}
6552 	if (rsm->r_flags & RACK_TLP) {
6553 		/*
6554 		 * We expect TLP's to have this occur.
6555 		 */
6556 		return;
6557 	}
6558 	rack = (struct tcp_rack *)tp->t_fb_ptr;
6559 	/* should we undo cc changes and exit recovery? */
6560 	if (IN_RECOVERY(tp->t_flags)) {
6561 		if (rack->r_ctl.rc_rsm_start == rsm->r_start) {
6562 			/*
6563 			 * Undo what we ratched down and exit recovery if
6564 			 * possible
6565 			 */
6566 			EXIT_RECOVERY(tp->t_flags);
6567 			tp->snd_recover = tp->snd_una;
6568 			if (rack->r_ctl.rc_cwnd_at > tp->snd_cwnd)
6569 				tp->snd_cwnd = rack->r_ctl.rc_cwnd_at;
6570 			if (rack->r_ctl.rc_ssthresh_at > tp->snd_ssthresh)
6571 				tp->snd_ssthresh = rack->r_ctl.rc_ssthresh_at;
6572 		}
6573 	}
6574 	if (rsm->r_flags & RACK_WAS_SACKPASS) {
6575 		/*
6576 		 * We retransmitted based on a sack and the earlier
6577 		 * retransmission ack'd it - re-ordering is occuring.
6578 		 */
6579 		counter_u64_add(rack_reorder_seen, 1);
6580 		rack->r_ctl.rc_reorder_ts = cts;
6581 	}
6582 	counter_u64_add(rack_badfr, 1);
6583 	counter_u64_add(rack_badfr_bytes, (rsm->r_end - rsm->r_start));
6584 }
6585 
6586 static void
rack_apply_updated_usrtt(struct tcp_rack * rack,uint32_t us_rtt,uint32_t us_cts)6587 rack_apply_updated_usrtt(struct tcp_rack *rack, uint32_t us_rtt, uint32_t us_cts)
6588 {
6589 	/*
6590 	 * Apply to filter the inbound us-rtt at us_cts.
6591 	 */
6592 	uint32_t old_rtt;
6593 
6594 	old_rtt = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
6595 	apply_filter_min_small(&rack->r_ctl.rc_gp_min_rtt,
6596 			       us_rtt, us_cts);
6597 	if (rack->r_ctl.last_pacing_time &&
6598 	    rack->rc_gp_dyn_mul &&
6599 	    (rack->r_ctl.last_pacing_time > us_rtt))
6600 		rack->pacing_longer_than_rtt = 1;
6601 	else
6602 		rack->pacing_longer_than_rtt = 0;
6603 	if (old_rtt > us_rtt) {
6604 		/* We just hit a new lower rtt time */
6605 		rack_log_rtt_shrinks(rack,  us_cts,  old_rtt,
6606 				     __LINE__, RACK_RTTS_NEWRTT);
6607 		/*
6608 		 * Only count it if its lower than what we saw within our
6609 		 * calculated range.
6610 		 */
6611 		if ((old_rtt - us_rtt) > rack_min_rtt_movement) {
6612 			if (rack_probertt_lower_within &&
6613 			    rack->rc_gp_dyn_mul &&
6614 			    (rack->use_fixed_rate == 0) &&
6615 			    (rack->rc_always_pace)) {
6616 				/*
6617 				 * We are seeing a new lower rtt very close
6618 				 * to the time that we would have entered probe-rtt.
6619 				 * This is probably due to the fact that a peer flow
6620 				 * has entered probe-rtt. Lets go in now too.
6621 				 */
6622 				uint32_t val;
6623 
6624 				val = rack_probertt_lower_within * rack_time_between_probertt;
6625 				val /= 100;
6626 				if ((rack->in_probe_rtt == 0)  &&
6627 				    ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= (rack_time_between_probertt - val)))	{
6628 					rack_enter_probertt(rack, us_cts);
6629 				}
6630 			}
6631 			rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
6632 		}
6633 	}
6634 }
6635 
6636 static int
rack_update_rtt(struct tcpcb * tp,struct tcp_rack * rack,struct rack_sendmap * rsm,struct tcpopt * to,uint32_t cts,int32_t ack_type,tcp_seq th_ack)6637 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
6638     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack)
6639 {
6640 	int32_t i;
6641 	uint32_t t, len_acked;
6642 
6643 	if ((rsm->r_flags & RACK_ACKED) ||
6644 	    (rsm->r_flags & RACK_WAS_ACKED))
6645 		/* Already done */
6646 		return (0);
6647 
6648 	if (ack_type == CUM_ACKED) {
6649 		if (SEQ_GT(th_ack, rsm->r_end))
6650 			len_acked = rsm->r_end - rsm->r_start;
6651 		else
6652 			len_acked = th_ack - rsm->r_start;
6653 	} else
6654 		len_acked = rsm->r_end - rsm->r_start;
6655 	if (rsm->r_rtr_cnt == 1) {
6656 		uint32_t us_rtt;
6657 
6658 		t = cts - rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
6659 		if ((int)t <= 0)
6660 			t = 1;
6661 		if (!tp->t_rttlow || tp->t_rttlow > t)
6662 			tp->t_rttlow = t;
6663 		if (!rack->r_ctl.rc_rack_min_rtt ||
6664 		    SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
6665 			rack->r_ctl.rc_rack_min_rtt = t;
6666 			if (rack->r_ctl.rc_rack_min_rtt == 0) {
6667 				rack->r_ctl.rc_rack_min_rtt = 1;
6668 			}
6669 		}
6670 		us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - rsm->usec_orig_send;
6671 		if (us_rtt == 0)
6672 			us_rtt = 1;
6673 		rack_apply_updated_usrtt(rack, us_rtt, tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time));
6674 		if (ack_type == SACKED)
6675 			tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt, 2 , rsm, rsm->r_rtr_cnt);
6676 		else {
6677 			/*
6678 			 * For cum-ack we are only confident if what
6679 			 * is being acked is included in a measurement.
6680 			 * Otherwise it could be an idle period that
6681 			 * includes Delayed-ack time.
6682 			 */
6683 			tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt,
6684 					    (rack->app_limited_needs_set ? 0 : 1), rsm, rsm->r_rtr_cnt);
6685 		}
6686 		if ((rsm->r_flags & RACK_TLP) &&
6687 		    (!IN_RECOVERY(tp->t_flags))) {
6688 			/* Segment was a TLP and our retrans matched */
6689 			if (rack->r_ctl.rc_tlp_cwnd_reduce) {
6690 				rack->r_ctl.rc_rsm_start = tp->snd_max;
6691 				rack->r_ctl.rc_cwnd_at = tp->snd_cwnd;
6692 				rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh;
6693 				rack_cong_signal(tp, NULL, CC_NDUPACK);
6694 				/*
6695 				 * When we enter recovery we need to assure
6696 				 * we send one packet.
6697 				 */
6698 				if (rack->rack_no_prr == 0) {
6699 					rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
6700 					rack_log_to_prr(rack, 7, 0);
6701 				}
6702 			}
6703 		}
6704 		if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) {
6705 			/* New more recent rack_tmit_time */
6706 			rack->r_ctl.rc_rack_tmit_time = rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
6707 			rack->rc_rack_rtt = t;
6708 		}
6709 		return (1);
6710 	}
6711 	/*
6712 	 * We clear the soft/rxtshift since we got an ack.
6713 	 * There is no assurance we will call the commit() function
6714 	 * so we need to clear these to avoid incorrect handling.
6715 	 */
6716 	tp->t_rxtshift = 0;
6717 	tp->t_softerror = 0;
6718 	if ((to->to_flags & TOF_TS) &&
6719 	    (ack_type == CUM_ACKED) &&
6720 	    (to->to_tsecr) &&
6721 	    ((rsm->r_flags & RACK_OVERMAX) == 0)) {
6722 		/*
6723 		 * Now which timestamp does it match? In this block the ACK
6724 		 * must be coming from a previous transmission.
6725 		 */
6726 		for (i = 0; i < rsm->r_rtr_cnt; i++) {
6727 			if (rsm->r_tim_lastsent[i] == to->to_tsecr) {
6728 				t = cts - rsm->r_tim_lastsent[i];
6729 				if ((int)t <= 0)
6730 					t = 1;
6731 				if ((i + 1) < rsm->r_rtr_cnt) {
6732 					/* Likely */
6733 					rack_earlier_retran(tp, rsm, t, cts);
6734 				}
6735 				if (!tp->t_rttlow || tp->t_rttlow > t)
6736 					tp->t_rttlow = t;
6737 				if (!rack->r_ctl.rc_rack_min_rtt || SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
6738 					rack->r_ctl.rc_rack_min_rtt = t;
6739 					if (rack->r_ctl.rc_rack_min_rtt == 0) {
6740 						rack->r_ctl.rc_rack_min_rtt = 1;
6741 					}
6742 				}
6743 				if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
6744 				    rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) {
6745 					/* New more recent rack_tmit_time */
6746 					rack->r_ctl.rc_rack_tmit_time = rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
6747 					rack->rc_rack_rtt = t;
6748 				}
6749 				tcp_rack_xmit_timer(rack, t + 1, len_acked, (t * HPTS_USEC_IN_MSEC), 0, rsm,
6750 						    rsm->r_rtr_cnt);
6751 				return (1);
6752 			}
6753 		}
6754 		goto ts_not_found;
6755 	} else {
6756 		/*
6757 		 * Ok its a SACK block that we retransmitted. or a windows
6758 		 * machine without timestamps. We can tell nothing from the
6759 		 * time-stamp since its not there or the time the peer last
6760 		 * recieved a segment that moved forward its cum-ack point.
6761 		 */
6762 ts_not_found:
6763 		i = rsm->r_rtr_cnt - 1;
6764 		t = cts - rsm->r_tim_lastsent[i];
6765 		if ((int)t <= 0)
6766 			t = 1;
6767 		if (rack->r_ctl.rc_rack_min_rtt && SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
6768 			/*
6769 			 * We retransmitted and the ack came back in less
6770 			 * than the smallest rtt we have observed. We most
6771 			 * likey did an improper retransmit as outlined in
6772 			 * 4.2 Step 3 point 2 in the rack-draft.
6773 			 */
6774 			i = rsm->r_rtr_cnt - 2;
6775 			t = cts - rsm->r_tim_lastsent[i];
6776 			rack_earlier_retran(tp, rsm, t, cts);
6777 		} else if (rack->r_ctl.rc_rack_min_rtt) {
6778 			/*
6779 			 * We retransmitted it and the retransmit did the
6780 			 * job.
6781 			 */
6782 			if (!rack->r_ctl.rc_rack_min_rtt ||
6783 			    SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
6784 				rack->r_ctl.rc_rack_min_rtt = t;
6785 				if (rack->r_ctl.rc_rack_min_rtt == 0) {
6786 					rack->r_ctl.rc_rack_min_rtt = 1;
6787 				}
6788 			}
6789 			if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, rsm->r_tim_lastsent[i])) {
6790 				/* New more recent rack_tmit_time */
6791 				rack->r_ctl.rc_rack_tmit_time = rsm->r_tim_lastsent[i];
6792 				rack->rc_rack_rtt = t;
6793 			}
6794 			return (1);
6795 		}
6796 	}
6797 	return (0);
6798 }
6799 
6800 /*
6801  * Mark the SACK_PASSED flag on all entries prior to rsm send wise.
6802  */
6803 static void
rack_log_sack_passed(struct tcpcb * tp,struct tcp_rack * rack,struct rack_sendmap * rsm)6804 rack_log_sack_passed(struct tcpcb *tp,
6805     struct tcp_rack *rack, struct rack_sendmap *rsm)
6806 {
6807 	struct rack_sendmap *nrsm;
6808 
6809 	nrsm = rsm;
6810 	TAILQ_FOREACH_REVERSE_FROM(nrsm, &rack->r_ctl.rc_tmap,
6811 	    rack_head, r_tnext) {
6812 		if (nrsm == rsm) {
6813 			/* Skip orginal segment he is acked */
6814 			continue;
6815 		}
6816 		if (nrsm->r_flags & RACK_ACKED) {
6817 			/*
6818 			 * Skip ack'd segments, though we
6819 			 * should not see these, since tmap
6820 			 * should not have ack'd segments.
6821 			 */
6822 			continue;
6823 		}
6824 		if (nrsm->r_flags & RACK_SACK_PASSED) {
6825 			/*
6826 			 * We found one that is already marked
6827 			 * passed, we have been here before and
6828 			 * so all others below this are marked.
6829 			 */
6830 			break;
6831 		}
6832 		nrsm->r_flags |= RACK_SACK_PASSED;
6833 		nrsm->r_flags &= ~RACK_WAS_SACKPASS;
6834 	}
6835 }
6836 
6837 static void
rack_need_set_test(struct tcpcb * tp,struct tcp_rack * rack,struct rack_sendmap * rsm,tcp_seq th_ack,int line,int use_which)6838 rack_need_set_test(struct tcpcb *tp,
6839 		   struct tcp_rack *rack,
6840 		   struct rack_sendmap *rsm,
6841 		   tcp_seq th_ack,
6842 		   int line,
6843 		   int use_which)
6844 {
6845 
6846 	if ((tp->t_flags & TF_GPUTINPROG) &&
6847 	    SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
6848 		/*
6849 		 * We were app limited, and this ack
6850 		 * butts up or goes beyond the point where we want
6851 		 * to start our next measurement. We need
6852 		 * to record the new gput_ts as here and
6853 		 * possibly update the start sequence.
6854 		 */
6855 		uint32_t seq, ts;
6856 
6857 		if (rsm->r_rtr_cnt > 1) {
6858 			/*
6859 			 * This is a retransmit, can we
6860 			 * really make any assessment at this
6861 			 * point?  We are not really sure of
6862 			 * the timestamp, is it this or the
6863 			 * previous transmission?
6864 			 *
6865 			 * Lets wait for something better that
6866 			 * is not retransmitted.
6867 			 */
6868 			return;
6869 		}
6870 		seq = tp->gput_seq;
6871 		ts = tp->gput_ts;
6872 		rack->app_limited_needs_set = 0;
6873 		tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
6874 		/* Do we start at a new end? */
6875 		if ((use_which == RACK_USE_BEG) &&
6876 		    SEQ_GEQ(rsm->r_start, tp->gput_seq)) {
6877 			/*
6878 			 * When we get an ACK that just eats
6879 			 * up some of the rsm, we set RACK_USE_BEG
6880 			 * since whats at r_start (i.e. th_ack)
6881 			 * is left unacked and thats where the
6882 			 * measurement not starts.
6883 			 */
6884 			tp->gput_seq = rsm->r_start;
6885 			rack->r_ctl.rc_gp_output_ts = rsm->usec_orig_send;
6886 		}
6887 		if ((use_which == RACK_USE_END) &&
6888 		    SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
6889 			    /*
6890 			     * We use the end when the cumack
6891 			     * is moving forward and completely
6892 			     * deleting the rsm passed so basically
6893 			     * r_end holds th_ack.
6894 			     *
6895 			     * For SACK's we also want to use the end
6896 			     * since this piece just got sacked and
6897 			     * we want to target anything after that
6898 			     * in our measurement.
6899 			     */
6900 			    tp->gput_seq = rsm->r_end;
6901 			    rack->r_ctl.rc_gp_output_ts = rsm->usec_orig_send;
6902 		}
6903 		if (use_which == RACK_USE_END_OR_THACK) {
6904 			/*
6905 			 * special case for ack moving forward,
6906 			 * not a sack, we need to move all the
6907 			 * way up to where this ack cum-ack moves
6908 			 * to.
6909 			 */
6910 			if (SEQ_GT(th_ack, rsm->r_end))
6911 				tp->gput_seq = th_ack;
6912 			else
6913 				tp->gput_seq = rsm->r_end;
6914 			rack->r_ctl.rc_gp_output_ts = rsm->usec_orig_send;
6915 		}
6916 		if (SEQ_GT(tp->gput_seq, tp->gput_ack)) {
6917 			/*
6918 			 * We moved beyond this guy's range, re-calculate
6919 			 * the new end point.
6920 			 */
6921 			if (rack->rc_gp_filled == 0) {
6922 				tp->gput_ack = tp->gput_seq + max(rc_init_window(rack), (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
6923 			} else {
6924 				tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
6925 			}
6926 		}
6927 		/*
6928 		 * We are moving the goal post, we may be able to clear the
6929 		 * measure_saw_probe_rtt flag.
6930 		 */
6931 		if ((rack->in_probe_rtt == 0) &&
6932 		    (rack->measure_saw_probe_rtt) &&
6933 		    (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
6934 			rack->measure_saw_probe_rtt = 0;
6935 		rack_log_pacing_delay_calc(rack, ts, tp->gput_ts,
6936 					   seq, tp->gput_seq, 0, 5, line, NULL);
6937 		if (rack->rc_gp_filled &&
6938 		    ((tp->gput_ack - tp->gput_seq) <
6939 		     max(rc_init_window(rack), (MIN_GP_WIN *
6940 						ctf_fixed_maxseg(tp))))) {
6941 			/*
6942 			 * There is no sense of continuing this measurement
6943 			 * because its too small to gain us anything we
6944 			 * trust. Skip it and that way we can start a new
6945 			 * measurement quicker.
6946 			 */
6947 			rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
6948 						   0, 0, 0, 6, __LINE__, NULL);
6949 			tp->t_flags &= ~TF_GPUTINPROG;
6950 		}
6951 	}
6952 }
6953 
6954 static uint32_t
rack_proc_sack_blk(struct tcpcb * tp,struct tcp_rack * rack,struct sackblk * sack,struct tcpopt * to,struct rack_sendmap ** prsm,uint32_t cts,int * moved_two)6955 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack, struct sackblk *sack,
6956 		   struct tcpopt *to, struct rack_sendmap **prsm, uint32_t cts, int *moved_two)
6957 {
6958 	uint32_t start, end, changed = 0;
6959 	struct rack_sendmap stack_map;
6960 	struct rack_sendmap *rsm, *nrsm, fe, *insret, *prev, *next;
6961 	int32_t used_ref = 1;
6962 	int moved = 0;
6963 
6964 	start = sack->start;
6965 	end = sack->end;
6966 	rsm = *prsm;
6967 	memset(&fe, 0, sizeof(fe));
6968 do_rest_ofb:
6969 	if ((rsm == NULL) ||
6970 	    (SEQ_LT(end, rsm->r_start)) ||
6971 	    (SEQ_GEQ(start, rsm->r_end)) ||
6972 	    (SEQ_LT(start, rsm->r_start))) {
6973 		/*
6974 		 * We are not in the right spot,
6975 		 * find the correct spot in the tree.
6976 		 */
6977 		used_ref = 0;
6978 		fe.r_start = start;
6979 		rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
6980 		moved++;
6981 	}
6982 	if (rsm == NULL) {
6983 		/* TSNH */
6984 		goto out;
6985 	}
6986 	/* Ok we have an ACK for some piece of this rsm */
6987 	if (rsm->r_start != start) {
6988 		if ((rsm->r_flags & RACK_ACKED) == 0) {
6989 			/**
6990 			 * Need to split this in two pieces the before and after,
6991 			 * the before remains in the map, the after must be
6992 			 * added. In other words we have:
6993 			 * rsm        |--------------|
6994 			 * sackblk        |------->
6995 			 * rsm will become
6996 			 *     rsm    |---|
6997 			 * and nrsm will be  the sacked piece
6998 			 *     nrsm       |----------|
6999 			 *
7000 			 * But before we start down that path lets
7001 			 * see if the sack spans over on top of
7002 			 * the next guy and it is already sacked.
7003 			 */
7004 			next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7005 			if (next && (next->r_flags & RACK_ACKED) &&
7006 			    SEQ_GEQ(end, next->r_start)) {
7007 				/**
7008 				 * So the next one is already acked, and
7009 				 * we can thus by hookery use our stack_map
7010 				 * to reflect the piece being sacked and
7011 				 * then adjust the two tree entries moving
7012 				 * the start and ends around. So we start like:
7013 				 *  rsm     |------------|             (not-acked)
7014 				 *  next                 |-----------| (acked)
7015 				 *  sackblk        |-------->
7016 				 *  We want to end like so:
7017 				 *  rsm     |------|                   (not-acked)
7018 				 *  next           |-----------------| (acked)
7019 				 *  nrsm           |-----|
7020 				 * Where nrsm is a temporary stack piece we
7021 				 * use to update all the gizmos.
7022 				 */
7023 				/* Copy up our fudge block */
7024 				nrsm = &stack_map;
7025 				memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
7026 				/* Now adjust our tree blocks */
7027 				rsm->r_end = start;
7028 				next->r_start = start;
7029 				/* Clear out the dup ack count of the remainder */
7030 				rsm->r_dupack = 0;
7031 				rsm->r_just_ret = 0;
7032 				rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
7033 				/* Now lets make sure our fudge block is right */
7034 				nrsm->r_start = start;
7035 				/* Now lets update all the stats and such */
7036 				rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
7037 				if (rack->app_limited_needs_set)
7038 					rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
7039 				changed += (nrsm->r_end - nrsm->r_start);
7040 				rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
7041 				if (nrsm->r_flags & RACK_SACK_PASSED) {
7042 					counter_u64_add(rack_reorder_seen, 1);
7043 					rack->r_ctl.rc_reorder_ts = cts;
7044 				}
7045 				/*
7046 				 * Now we want to go up from rsm (the
7047 				 * one left un-acked) to the next one
7048 				 * in the tmap. We do this so when
7049 				 * we walk backwards we include marking
7050 				 * sack-passed on rsm (The one passed in
7051 				 * is skipped since it is generally called
7052 				 * on something sacked before removing it
7053 				 * from the tmap).
7054 				 */
7055 				if (rsm->r_in_tmap) {
7056 					nrsm = TAILQ_NEXT(rsm, r_tnext);
7057 					/*
7058 					 * Now that we have the next
7059 					 * one walk backwards from there.
7060 					 */
7061 					if (nrsm && nrsm->r_in_tmap)
7062 						rack_log_sack_passed(tp, rack, nrsm);
7063 				}
7064 				/* Now are we done? */
7065 				if (SEQ_LT(end, next->r_end) ||
7066 				    (end == next->r_end)) {
7067 					/* Done with block */
7068 					goto out;
7069 				}
7070 				counter_u64_add(rack_sack_used_next_merge, 1);
7071 				/* Postion for the next block */
7072 				start = next->r_end;
7073 				rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, next);
7074 				if (rsm == NULL)
7075 					goto out;
7076 			} else {
7077 				/**
7078 				 * We can't use any hookery here, so we
7079 				 * need to split the map. We enter like
7080 				 * so:
7081 				 *  rsm      |--------|
7082 				 *  sackblk       |----->
7083 				 * We will add the new block nrsm and
7084 				 * that will be the new portion, and then
7085 				 * fall through after reseting rsm. So we
7086 				 * split and look like this:
7087 				 *  rsm      |----|
7088 				 *  sackblk       |----->
7089 				 *  nrsm          |---|
7090 				 * We then fall through reseting
7091 				 * rsm to nrsm, so the next block
7092 				 * picks it up.
7093 				 */
7094 				nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
7095 				if (nrsm == NULL) {
7096 					/*
7097 					 * failed XXXrrs what can we do but loose the sack
7098 					 * info?
7099 					 */
7100 					goto out;
7101 				}
7102 				counter_u64_add(rack_sack_splits, 1);
7103 				rack_clone_rsm(rack, nrsm, rsm, start);
7104 				rsm->r_just_ret = 0;
7105 				insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7106 #ifdef INVARIANTS
7107 				if (insret != NULL) {
7108 					panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7109 					      nrsm, insret, rack, rsm);
7110 				}
7111 #endif
7112 				if (rsm->r_in_tmap) {
7113 					TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7114 					nrsm->r_in_tmap = 1;
7115 				}
7116 				rsm->r_flags &= (~RACK_HAS_FIN);
7117 				/* Position us to point to the new nrsm that starts the sack blk */
7118 				rsm = nrsm;
7119 			}
7120 		} else {
7121 			/* Already sacked this piece */
7122 			counter_u64_add(rack_sack_skipped_acked, 1);
7123 			moved++;
7124 			if (end == rsm->r_end) {
7125 				/* Done with block */
7126 				rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7127 				goto out;
7128 			} else if (SEQ_LT(end, rsm->r_end)) {
7129 				/* A partial sack to a already sacked block */
7130 				moved++;
7131 				rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7132 				goto out;
7133 			} else {
7134 				/*
7135 				 * The end goes beyond this guy
7136 				 * repostion the start to the
7137 				 * next block.
7138 				 */
7139 				start = rsm->r_end;
7140 				rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7141 				if (rsm == NULL)
7142 					goto out;
7143 			}
7144 		}
7145 	}
7146 	if (SEQ_GEQ(end, rsm->r_end)) {
7147 		/**
7148 		 * The end of this block is either beyond this guy or right
7149 		 * at this guy. I.e.:
7150 		 *  rsm ---                 |-----|
7151 		 *  end                     |-----|
7152 		 *  <or>
7153 		 *  end                     |---------|
7154 		 */
7155 		if ((rsm->r_flags & RACK_ACKED) == 0) {
7156 			rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
7157 			changed += (rsm->r_end - rsm->r_start);
7158 			rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
7159 			if (rsm->r_in_tmap) /* should be true */
7160 				rack_log_sack_passed(tp, rack, rsm);
7161 			/* Is Reordering occuring? */
7162 			if (rsm->r_flags & RACK_SACK_PASSED) {
7163 				rsm->r_flags &= ~RACK_SACK_PASSED;
7164 				counter_u64_add(rack_reorder_seen, 1);
7165 				rack->r_ctl.rc_reorder_ts = cts;
7166 			}
7167 			if (rack->app_limited_needs_set)
7168 				rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
7169 			rsm->r_ack_arrival = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
7170 			rsm->r_flags |= RACK_ACKED;
7171 			rsm->r_flags &= ~RACK_TLP;
7172 			if (rsm->r_in_tmap) {
7173 				TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7174 				rsm->r_in_tmap = 0;
7175 			}
7176 		} else {
7177 			counter_u64_add(rack_sack_skipped_acked, 1);
7178 			moved++;
7179 		}
7180 		if (end == rsm->r_end) {
7181 			/* This block only - done, setup for next  */
7182 			goto out;
7183 		}
7184 		/*
7185 		 * There is more not coverend by this rsm move on
7186 		 * to the next block in the RB tree.
7187 		 */
7188 		nrsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7189 		start = rsm->r_end;
7190 		rsm = nrsm;
7191 		if (rsm == NULL)
7192 			goto out;
7193 		goto do_rest_ofb;
7194 	}
7195 	/**
7196 	 * The end of this sack block is smaller than
7197 	 * our rsm i.e.:
7198 	 *  rsm ---                 |-----|
7199 	 *  end                     |--|
7200 	 */
7201 	if ((rsm->r_flags & RACK_ACKED) == 0) {
7202 		prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7203 		if (prev && (prev->r_flags & RACK_ACKED)) {
7204 			/**
7205 			 * Goal, we want the right remainder of rsm to shrink
7206 			 * in place and span from (rsm->r_start = end) to rsm->r_end.
7207 			 * We want to expand prev to go all the way
7208 			 * to prev->r_end <- end.
7209 			 * so in the tree we have before:
7210 			 *   prev     |--------|         (acked)
7211 			 *   rsm               |-------| (non-acked)
7212 			 *   sackblk           |-|
7213 			 * We churn it so we end up with
7214 			 *   prev     |----------|       (acked)
7215 			 *   rsm                 |-----| (non-acked)
7216 			 *   nrsm              |-| (temporary)
7217 			 */
7218 			nrsm = &stack_map;
7219 			memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
7220 			prev->r_end = end;
7221 			rsm->r_start = end;
7222 			/* Now adjust nrsm (stack copy) to be
7223 			 * the one that is the small
7224 			 * piece that was "sacked".
7225 			 */
7226 			nrsm->r_end = end;
7227 			rsm->r_dupack = 0;
7228 			rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
7229 			/*
7230 			 * Now nrsm is our new little piece
7231 			 * that is acked (which was merged
7232 			 * to prev). Update the rtt and changed
7233 			 * based on that. Also check for reordering.
7234 			 */
7235 			rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
7236 			if (rack->app_limited_needs_set)
7237 				rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
7238 			changed += (nrsm->r_end - nrsm->r_start);
7239 			rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
7240 			if (nrsm->r_flags & RACK_SACK_PASSED) {
7241 				counter_u64_add(rack_reorder_seen, 1);
7242 				rack->r_ctl.rc_reorder_ts = cts;
7243 			}
7244 			rsm = prev;
7245 			counter_u64_add(rack_sack_used_prev_merge, 1);
7246 		} else {
7247 			/**
7248 			 * This is the case where our previous
7249 			 * block is not acked either, so we must
7250 			 * split the block in two.
7251 			 */
7252 			nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
7253 			if (nrsm == NULL) {
7254 				/* failed rrs what can we do but loose the sack info? */
7255 				goto out;
7256 			}
7257 			/**
7258 			 * In this case nrsm becomes
7259 			 * nrsm->r_start = end;
7260 			 * nrsm->r_end = rsm->r_end;
7261 			 * which is un-acked.
7262 			 * <and>
7263 			 * rsm->r_end = nrsm->r_start;
7264 			 * i.e. the remaining un-acked
7265 			 * piece is left on the left
7266 			 * hand side.
7267 			 *
7268 			 * So we start like this
7269 			 * rsm      |----------| (not acked)
7270 			 * sackblk  |---|
7271 			 * build it so we have
7272 			 * rsm      |---|         (acked)
7273 			 * nrsm         |------|  (not acked)
7274 			 */
7275 			counter_u64_add(rack_sack_splits, 1);
7276 			rack_clone_rsm(rack, nrsm, rsm, end);
7277 			rsm->r_flags &= (~RACK_HAS_FIN);
7278 			rsm->r_just_ret = 0;
7279 			insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7280 #ifdef INVARIANTS
7281 			if (insret != NULL) {
7282 				panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7283 				      nrsm, insret, rack, rsm);
7284 			}
7285 #endif
7286 			if (rsm->r_in_tmap) {
7287 				TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7288 				nrsm->r_in_tmap = 1;
7289 			}
7290 			nrsm->r_dupack = 0;
7291 			rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
7292 			rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
7293 			changed += (rsm->r_end - rsm->r_start);
7294 			rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
7295 			if (rsm->r_in_tmap) /* should be true */
7296 				rack_log_sack_passed(tp, rack, rsm);
7297 			/* Is Reordering occuring? */
7298 			if (rsm->r_flags & RACK_SACK_PASSED) {
7299 				rsm->r_flags &= ~RACK_SACK_PASSED;
7300 				counter_u64_add(rack_reorder_seen, 1);
7301 				rack->r_ctl.rc_reorder_ts = cts;
7302 			}
7303 			if (rack->app_limited_needs_set)
7304 				rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
7305 			rsm->r_ack_arrival = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
7306 			rsm->r_flags |= RACK_ACKED;
7307 			rsm->r_flags &= ~RACK_TLP;
7308 			if (rsm->r_in_tmap) {
7309 				TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7310 				rsm->r_in_tmap = 0;
7311 			}
7312 		}
7313 	} else if (start != end){
7314 		/*
7315 		 * The block was already acked.
7316 		 */
7317 		counter_u64_add(rack_sack_skipped_acked, 1);
7318 		moved++;
7319 	}
7320 out:
7321 	if (rsm && (rsm->r_flags & RACK_ACKED)) {
7322 		/*
7323 		 * Now can we merge where we worked
7324 		 * with either the previous or
7325 		 * next block?
7326 		 */
7327 		next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7328 		while (next) {
7329 		    if (next->r_flags & RACK_ACKED) {
7330 			/* yep this and next can be merged */
7331 			rsm = rack_merge_rsm(rack, rsm, next);
7332 			next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7333 		    } else
7334 			    break;
7335 		}
7336 		/* Now what about the previous? */
7337 		prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7338 		while (prev) {
7339 		    if (prev->r_flags & RACK_ACKED) {
7340 			/* yep the previous and this can be merged */
7341 			rsm = rack_merge_rsm(rack, prev, rsm);
7342 			prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7343 		    } else
7344 			    break;
7345 		}
7346 	}
7347 	if (used_ref == 0) {
7348 		counter_u64_add(rack_sack_proc_all, 1);
7349 	} else {
7350 		counter_u64_add(rack_sack_proc_short, 1);
7351 	}
7352 	/* Save off the next one for quick reference. */
7353 	if (rsm)
7354 		nrsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7355 	else
7356 		nrsm = NULL;
7357 	*prsm = rack->r_ctl.rc_sacklast = nrsm;
7358 	/* Pass back the moved. */
7359 	*moved_two = moved;
7360 	return (changed);
7361 }
7362 
7363 static void inline
rack_peer_reneges(struct tcp_rack * rack,struct rack_sendmap * rsm,tcp_seq th_ack)7364 rack_peer_reneges(struct tcp_rack *rack, struct rack_sendmap *rsm, tcp_seq th_ack)
7365 {
7366 	struct rack_sendmap *tmap;
7367 
7368 	tmap = NULL;
7369 	while (rsm && (rsm->r_flags & RACK_ACKED)) {
7370 		/* Its no longer sacked, mark it so */
7371 		rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
7372 #ifdef INVARIANTS
7373 		if (rsm->r_in_tmap) {
7374 			panic("rack:%p rsm:%p flags:0x%x in tmap?",
7375 			      rack, rsm, rsm->r_flags);
7376 		}
7377 #endif
7378 		rsm->r_flags &= ~(RACK_ACKED|RACK_SACK_PASSED|RACK_WAS_SACKPASS);
7379 		/* Rebuild it into our tmap */
7380 		if (tmap == NULL) {
7381 			TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7382 			tmap = rsm;
7383 		} else {
7384 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, tmap, rsm, r_tnext);
7385 			tmap = rsm;
7386 		}
7387 		tmap->r_in_tmap = 1;
7388 		rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7389 	}
7390 	/*
7391 	 * Now lets possibly clear the sack filter so we start
7392 	 * recognizing sacks that cover this area.
7393 	 */
7394 	sack_filter_clear(&rack->r_ctl.rack_sf, th_ack);
7395 
7396 }
7397 
7398 static void
rack_do_decay(struct tcp_rack * rack)7399 rack_do_decay(struct tcp_rack *rack)
7400 {
7401 	struct timeval res;
7402 
7403 #define	timersub(tvp, uvp, vvp)						\
7404 	do {								\
7405 		(vvp)->tv_sec = (tvp)->tv_sec - (uvp)->tv_sec;		\
7406 		(vvp)->tv_usec = (tvp)->tv_usec - (uvp)->tv_usec;	\
7407 		if ((vvp)->tv_usec < 0) {				\
7408 			(vvp)->tv_sec--;				\
7409 			(vvp)->tv_usec += 1000000;			\
7410 		}							\
7411 	} while (0)
7412 
7413 	timersub(&rack->r_ctl.act_rcv_time, &rack->r_ctl.rc_last_time_decay, &res);
7414 #undef timersub
7415 
7416 	rack->r_ctl.input_pkt++;
7417 	if ((rack->rc_in_persist) ||
7418 	    (res.tv_sec >= 1) ||
7419 	    (rack->rc_tp->snd_max == rack->rc_tp->snd_una)) {
7420 		/*
7421 		 * Check for decay of non-SAD,
7422 		 * we want all SAD detection metrics to
7423 		 * decay 1/4 per second (or more) passed.
7424 		 */
7425 		uint32_t pkt_delta;
7426 
7427 		pkt_delta = rack->r_ctl.input_pkt - rack->r_ctl.saved_input_pkt;
7428 		/* Update our saved tracking values */
7429 		rack->r_ctl.saved_input_pkt = rack->r_ctl.input_pkt;
7430 		rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
7431 		/* Now do we escape without decay? */
7432 #ifdef NETFLIX_EXP_DETECTION
7433 		if (rack->rc_in_persist ||
7434 		    (rack->rc_tp->snd_max == rack->rc_tp->snd_una) ||
7435 		    (pkt_delta < tcp_sad_low_pps)){
7436 			/*
7437 			 * We don't decay idle connections
7438 			 * or ones that have a low input pps.
7439 			 */
7440 			return;
7441 		}
7442 		/* Decay the counters */
7443 		rack->r_ctl.ack_count = ctf_decay_count(rack->r_ctl.ack_count,
7444 							tcp_sad_decay_val);
7445 		rack->r_ctl.sack_count = ctf_decay_count(rack->r_ctl.sack_count,
7446 							 tcp_sad_decay_val);
7447 		rack->r_ctl.sack_moved_extra = ctf_decay_count(rack->r_ctl.sack_moved_extra,
7448 							       tcp_sad_decay_val);
7449 		rack->r_ctl.sack_noextra_move = ctf_decay_count(rack->r_ctl.sack_noextra_move,
7450 								tcp_sad_decay_val);
7451 #endif
7452 	}
7453 }
7454 
7455 static void
rack_log_ack(struct tcpcb * tp,struct tcpopt * to,struct tcphdr * th)7456 rack_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th)
7457 {
7458 	uint32_t changed, entered_recovery = 0;
7459 	struct tcp_rack *rack;
7460 	struct rack_sendmap *rsm, *rm;
7461 	struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1];
7462 	register uint32_t th_ack;
7463 	int32_t i, j, k, num_sack_blks = 0;
7464 	uint32_t cts, acked, ack_point, sack_changed = 0;
7465 	int loop_start = 0, moved_two = 0;
7466 	uint32_t tsused;
7467 
7468 	INP_WLOCK_ASSERT(tp->t_inpcb);
7469 	if (th->th_flags & TH_RST) {
7470 		/* We don't log resets */
7471 		return;
7472 	}
7473 	rack = (struct tcp_rack *)tp->t_fb_ptr;
7474 	cts = tcp_ts_getticks();
7475 	rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
7476 	changed = 0;
7477 	th_ack = th->th_ack;
7478 	if (rack->sack_attack_disable == 0)
7479 		rack_do_decay(rack);
7480 	if (BYTES_THIS_ACK(tp, th) >= ctf_fixed_maxseg(rack->rc_tp)) {
7481 		/*
7482 		 * You only get credit for
7483 		 * MSS and greater (and you get extra
7484 		 * credit for larger cum-ack moves).
7485 		 */
7486 		int ac;
7487 
7488 		ac = BYTES_THIS_ACK(tp, th) / ctf_fixed_maxseg(rack->rc_tp);
7489 		rack->r_ctl.ack_count += ac;
7490 		counter_u64_add(rack_ack_total, ac);
7491 	}
7492 	if (rack->r_ctl.ack_count > 0xfff00000) {
7493 		/*
7494 		 * reduce the number to keep us under
7495 		 * a uint32_t.
7496 		 */
7497 		rack->r_ctl.ack_count /= 2;
7498 		rack->r_ctl.sack_count /= 2;
7499 	}
7500 	if (SEQ_GT(th_ack, tp->snd_una)) {
7501 		rack_log_progress_event(rack, tp, ticks, PROGRESS_UPDATE, __LINE__);
7502 		tp->t_acktime = ticks;
7503 	}
7504 	if (rsm && SEQ_GT(th_ack, rsm->r_start))
7505 		changed = th_ack - rsm->r_start;
7506 	if (changed) {
7507 		/*
7508 		 * The ACK point is advancing to th_ack, we must drop off
7509 		 * the packets in the rack log and calculate any eligble
7510 		 * RTT's.
7511 		 */
7512 		rack->r_wanted_output = 1;
7513 more:
7514 		rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
7515 		if (rsm == NULL) {
7516 			if ((th_ack - 1) == tp->iss) {
7517 				/*
7518 				 * For the SYN incoming case we will not
7519 				 * have called tcp_output for the sending of
7520 				 * the SYN, so there will be no map. All
7521 				 * other cases should probably be a panic.
7522 				 */
7523 				goto proc_sack;
7524 			}
7525 			if (tp->t_flags & TF_SENTFIN) {
7526 				/* if we send a FIN we will not hav a map */
7527 				goto proc_sack;
7528 			}
7529 #ifdef INVARIANTS
7530 			panic("No rack map tp:%p for th:%p state:%d rack:%p snd_una:%u snd_max:%u snd_nxt:%u chg:%d\n",
7531 			      tp,
7532 			      th, tp->t_state, rack,
7533 			      tp->snd_una, tp->snd_max, tp->snd_nxt, changed);
7534 #endif
7535 			goto proc_sack;
7536 		}
7537 		if (SEQ_LT(th_ack, rsm->r_start)) {
7538 			/* Huh map is missing this */
7539 #ifdef INVARIANTS
7540 			printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d\n",
7541 			       rsm->r_start,
7542 			       th_ack, tp->t_state, rack->r_state);
7543 #endif
7544 			goto proc_sack;
7545 		}
7546 		rack_update_rtt(tp, rack, rsm, to, cts, CUM_ACKED, th_ack);
7547 		/* Now do we consume the whole thing? */
7548 		if (SEQ_GEQ(th_ack, rsm->r_end)) {
7549 			/* Its all consumed. */
7550 			uint32_t left;
7551 			uint8_t newly_acked;
7552 
7553 			rack->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
7554 			rsm->r_rtr_bytes = 0;
7555 			/* Record the time of highest cumack sent */
7556 			rack->r_ctl.rc_gp_cumack_ts = rsm->usec_orig_send;
7557 			rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7558 #ifdef INVARIANTS
7559 			if (rm != rsm) {
7560 				panic("removing head in rack:%p rsm:%p rm:%p",
7561 				      rack, rsm, rm);
7562 			}
7563 #endif
7564 			if (rsm->r_in_tmap) {
7565 				TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7566 				rsm->r_in_tmap = 0;
7567 			}
7568 			newly_acked = 1;
7569 			if (rsm->r_flags & RACK_ACKED) {
7570 				/*
7571 				 * It was acked on the scoreboard -- remove
7572 				 * it from total
7573 				 */
7574 				rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
7575 				newly_acked = 0;
7576 			} else if (rsm->r_flags & RACK_SACK_PASSED) {
7577 				/*
7578 				 * There are segments ACKED on the
7579 				 * scoreboard further up. We are seeing
7580 				 * reordering.
7581 				 */
7582 				rsm->r_flags &= ~RACK_SACK_PASSED;
7583 				counter_u64_add(rack_reorder_seen, 1);
7584 				rsm->r_ack_arrival = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
7585 				rsm->r_flags |= RACK_ACKED;
7586 				rack->r_ctl.rc_reorder_ts = cts;
7587 			}
7588 			left = th_ack - rsm->r_end;
7589 			if (rack->app_limited_needs_set && newly_acked)
7590 				rack_need_set_test(tp, rack, rsm, th_ack, __LINE__, RACK_USE_END_OR_THACK);
7591 			/* Free back to zone */
7592 			rack_free(rack, rsm);
7593 			if (left) {
7594 				goto more;
7595 			}
7596 			goto proc_sack;
7597 		}
7598 		if (rsm->r_flags & RACK_ACKED) {
7599 			/*
7600 			 * It was acked on the scoreboard -- remove it from
7601 			 * total for the part being cum-acked.
7602 			 */
7603 			rack->r_ctl.rc_sacked -= (th_ack - rsm->r_start);
7604 		}
7605 		/*
7606 		 * Clear the dup ack count for
7607 		 * the piece that remains.
7608 		 */
7609 		rsm->r_dupack = 0;
7610 		rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
7611 		if (rsm->r_rtr_bytes) {
7612 			/*
7613 			 * It was retransmitted adjust the
7614 			 * sack holes for what was acked.
7615 			 */
7616 			int ack_am;
7617 
7618 			ack_am = (th_ack - rsm->r_start);
7619 			if (ack_am >= rsm->r_rtr_bytes) {
7620 				rack->r_ctl.rc_holes_rxt -= ack_am;
7621 				rsm->r_rtr_bytes -= ack_am;
7622 			}
7623 		}
7624 		/*
7625 		 * Update where the piece starts and record
7626 		 * the time of send of highest cumack sent.
7627 		 */
7628 		rack->r_ctl.rc_gp_cumack_ts = rsm->usec_orig_send;
7629 		rsm->r_start = th_ack;
7630 		if (rack->app_limited_needs_set)
7631 			rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_BEG);
7632 	}
7633 proc_sack:
7634 	/* Check for reneging */
7635 	rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
7636 	if (rsm && (rsm->r_flags & RACK_ACKED) && (th_ack == rsm->r_start)) {
7637 		/*
7638 		 * The peer has moved snd_una up to
7639 		 * the edge of this send, i.e. one
7640 		 * that it had previously acked. The only
7641 		 * way that can be true if the peer threw
7642 		 * away data (space issues) that it had
7643 		 * previously sacked (else it would have
7644 		 * given us snd_una up to (rsm->r_end).
7645 		 * We need to undo the acked markings here.
7646 		 *
7647 		 * Note we have to look to make sure th_ack is
7648 		 * our rsm->r_start in case we get an old ack
7649 		 * where th_ack is behind snd_una.
7650 		 */
7651 		rack_peer_reneges(rack, rsm, th->th_ack);
7652 	}
7653 	if ((to->to_flags & TOF_SACK) == 0) {
7654 		/* We are done nothing left */
7655 		goto out;
7656 	}
7657 	/* Sack block processing */
7658 	if (SEQ_GT(th_ack, tp->snd_una))
7659 		ack_point = th_ack;
7660 	else
7661 		ack_point = tp->snd_una;
7662 	for (i = 0; i < to->to_nsacks; i++) {
7663 		bcopy((to->to_sacks + i * TCPOLEN_SACK),
7664 		      &sack, sizeof(sack));
7665 		sack.start = ntohl(sack.start);
7666 		sack.end = ntohl(sack.end);
7667 		if (SEQ_GT(sack.end, sack.start) &&
7668 		    SEQ_GT(sack.start, ack_point) &&
7669 		    SEQ_LT(sack.start, tp->snd_max) &&
7670 		    SEQ_GT(sack.end, ack_point) &&
7671 		    SEQ_LEQ(sack.end, tp->snd_max)) {
7672 			sack_blocks[num_sack_blks] = sack;
7673 			num_sack_blks++;
7674 #ifdef NETFLIX_STATS
7675 		} else if (SEQ_LEQ(sack.start, th_ack) &&
7676 			   SEQ_LEQ(sack.end, th_ack)) {
7677 			/*
7678 			 * Its a D-SACK block.
7679 			 */
7680 			tcp_record_dsack(sack.start, sack.end);
7681 #endif
7682 		}
7683 	}
7684 	/*
7685 	 * Sort the SACK blocks so we can update the rack scoreboard with
7686 	 * just one pass.
7687 	 */
7688 	num_sack_blks = sack_filter_blks(&rack->r_ctl.rack_sf, sack_blocks,
7689 					 num_sack_blks, th->th_ack);
7690 	ctf_log_sack_filter(rack->rc_tp, num_sack_blks, sack_blocks);
7691 	if (num_sack_blks == 0)  {
7692 		/* Nothing to sack (DSACKs?) */
7693 		goto out_with_totals;
7694 	}
7695 	if (num_sack_blks < 2) {
7696 		/* Only one, we don't need to sort */
7697 		goto do_sack_work;
7698 	}
7699 	/* Sort the sacks */
7700 	for (i = 0; i < num_sack_blks; i++) {
7701 		for (j = i + 1; j < num_sack_blks; j++) {
7702 			if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
7703 				sack = sack_blocks[i];
7704 				sack_blocks[i] = sack_blocks[j];
7705 				sack_blocks[j] = sack;
7706 			}
7707 		}
7708 	}
7709 	/*
7710 	 * Now are any of the sack block ends the same (yes some
7711 	 * implementations send these)?
7712 	 */
7713 again:
7714 	if (num_sack_blks == 0)
7715 		goto out_with_totals;
7716 	if (num_sack_blks > 1) {
7717 		for (i = 0; i < num_sack_blks; i++) {
7718 			for (j = i + 1; j < num_sack_blks; j++) {
7719 				if (sack_blocks[i].end == sack_blocks[j].end) {
7720 					/*
7721 					 * Ok these two have the same end we
7722 					 * want the smallest end and then
7723 					 * throw away the larger and start
7724 					 * again.
7725 					 */
7726 					if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) {
7727 						/*
7728 						 * The second block covers
7729 						 * more area use that
7730 						 */
7731 						sack_blocks[i].start = sack_blocks[j].start;
7732 					}
7733 					/*
7734 					 * Now collapse out the dup-sack and
7735 					 * lower the count
7736 					 */
7737 					for (k = (j + 1); k < num_sack_blks; k++) {
7738 						sack_blocks[j].start = sack_blocks[k].start;
7739 						sack_blocks[j].end = sack_blocks[k].end;
7740 						j++;
7741 					}
7742 					num_sack_blks--;
7743 					goto again;
7744 				}
7745 			}
7746 		}
7747 	}
7748 do_sack_work:
7749 	/*
7750 	 * First lets look to see if
7751 	 * we have retransmitted and
7752 	 * can use the transmit next?
7753 	 */
7754 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
7755 	if (rsm &&
7756 	    SEQ_GT(sack_blocks[0].end, rsm->r_start) &&
7757 	    SEQ_LT(sack_blocks[0].start, rsm->r_end)) {
7758 		/*
7759 		 * We probably did the FR and the next
7760 		 * SACK in continues as we would expect.
7761 		 */
7762 		acked = rack_proc_sack_blk(tp, rack, &sack_blocks[0], to, &rsm, cts, &moved_two);
7763 		if (acked) {
7764 			rack->r_wanted_output = 1;
7765 			changed += acked;
7766 			sack_changed += acked;
7767 		}
7768 		if (num_sack_blks == 1) {
7769 			/*
7770 			 * This is what we would expect from
7771 			 * a normal implementation to happen
7772 			 * after we have retransmitted the FR,
7773 			 * i.e the sack-filter pushes down
7774 			 * to 1 block and the next to be retransmitted
7775 			 * is the sequence in the sack block (has more
7776 			 * are acked). Count this as ACK'd data to boost
7777 			 * up the chances of recovering any false positives.
7778 			 */
7779 			rack->r_ctl.ack_count += (acked / ctf_fixed_maxseg(rack->rc_tp));
7780 			counter_u64_add(rack_ack_total, (acked / ctf_fixed_maxseg(rack->rc_tp)));
7781 			counter_u64_add(rack_express_sack, 1);
7782 			if (rack->r_ctl.ack_count > 0xfff00000) {
7783 				/*
7784 				 * reduce the number to keep us under
7785 				 * a uint32_t.
7786 				 */
7787 				rack->r_ctl.ack_count /= 2;
7788 				rack->r_ctl.sack_count /= 2;
7789 			}
7790 			goto out_with_totals;
7791 		} else {
7792 			/*
7793 			 * Start the loop through the
7794 			 * rest of blocks, past the first block.
7795 			 */
7796 			moved_two = 0;
7797 			loop_start = 1;
7798 		}
7799 	}
7800 	/* Its a sack of some sort */
7801 	rack->r_ctl.sack_count++;
7802 	if (rack->r_ctl.sack_count > 0xfff00000) {
7803 		/*
7804 		 * reduce the number to keep us under
7805 		 * a uint32_t.
7806 		 */
7807 		rack->r_ctl.ack_count /= 2;
7808 		rack->r_ctl.sack_count /= 2;
7809 	}
7810 	counter_u64_add(rack_sack_total, 1);
7811 	if (rack->sack_attack_disable) {
7812 		/* An attacker disablement is in place */
7813 		if (num_sack_blks > 1) {
7814 			rack->r_ctl.sack_count += (num_sack_blks - 1);
7815 			rack->r_ctl.sack_moved_extra++;
7816 			counter_u64_add(rack_move_some, 1);
7817 			if (rack->r_ctl.sack_moved_extra > 0xfff00000) {
7818 				rack->r_ctl.sack_moved_extra /= 2;
7819 				rack->r_ctl.sack_noextra_move /= 2;
7820 			}
7821 		}
7822 		goto out;
7823 	}
7824 	rsm = rack->r_ctl.rc_sacklast;
7825 	for (i = loop_start; i < num_sack_blks; i++) {
7826 		acked = rack_proc_sack_blk(tp, rack, &sack_blocks[i], to, &rsm, cts, &moved_two);
7827 		if (acked) {
7828 			rack->r_wanted_output = 1;
7829 			changed += acked;
7830 			sack_changed += acked;
7831 		}
7832 		if (moved_two) {
7833 			/*
7834 			 * If we did not get a SACK for at least a MSS and
7835 			 * had to move at all, or if we moved more than our
7836 			 * threshold, it counts against the "extra" move.
7837 			 */
7838 			rack->r_ctl.sack_moved_extra += moved_two;
7839 			counter_u64_add(rack_move_some, 1);
7840 		} else {
7841 			/*
7842 			 * else we did not have to move
7843 			 * any more than we would expect.
7844 			 */
7845 			rack->r_ctl.sack_noextra_move++;
7846 			counter_u64_add(rack_move_none, 1);
7847 		}
7848 		if (moved_two && (acked < ctf_fixed_maxseg(rack->rc_tp))) {
7849 			/*
7850 			 * If the SACK was not a full MSS then
7851 			 * we add to sack_count the number of
7852 			 * MSS's (or possibly more than
7853 			 * a MSS if its a TSO send) we had to skip by.
7854 			 */
7855 			rack->r_ctl.sack_count += moved_two;
7856 			counter_u64_add(rack_sack_total, moved_two);
7857 		}
7858 		/*
7859 		 * Now we need to setup for the next
7860 		 * round. First we make sure we won't
7861 		 * exceed the size of our uint32_t on
7862 		 * the various counts, and then clear out
7863 		 * moved_two.
7864 		 */
7865 		if ((rack->r_ctl.sack_moved_extra > 0xfff00000) ||
7866 		    (rack->r_ctl.sack_noextra_move > 0xfff00000)) {
7867 			rack->r_ctl.sack_moved_extra /= 2;
7868 			rack->r_ctl.sack_noextra_move /= 2;
7869 		}
7870 		if (rack->r_ctl.sack_count > 0xfff00000) {
7871 			rack->r_ctl.ack_count /= 2;
7872 			rack->r_ctl.sack_count /= 2;
7873 		}
7874 		moved_two = 0;
7875 	}
7876 out_with_totals:
7877 	if (num_sack_blks > 1) {
7878 		/*
7879 		 * You get an extra stroke if
7880 		 * you have more than one sack-blk, this
7881 		 * could be where we are skipping forward
7882 		 * and the sack-filter is still working, or
7883 		 * it could be an attacker constantly
7884 		 * moving us.
7885 		 */
7886 		rack->r_ctl.sack_moved_extra++;
7887 		counter_u64_add(rack_move_some, 1);
7888 	}
7889 out:
7890 #ifdef NETFLIX_EXP_DETECTION
7891 	if ((rack->do_detection || tcp_force_detection) &&
7892 	    tcp_sack_to_ack_thresh &&
7893 	    tcp_sack_to_move_thresh &&
7894 	    ((rack->r_ctl.rc_num_maps_alloced > tcp_map_minimum) || rack->sack_attack_disable)) {
7895 		/*
7896 		 * We have thresholds set to find
7897 		 * possible attackers and disable sack.
7898 		 * Check them.
7899 		 */
7900 		uint64_t ackratio, moveratio, movetotal;
7901 
7902 		/* Log detecting */
7903 		rack_log_sad(rack, 1);
7904 		ackratio = (uint64_t)(rack->r_ctl.sack_count);
7905 		ackratio *= (uint64_t)(1000);
7906 		if (rack->r_ctl.ack_count)
7907 			ackratio /= (uint64_t)(rack->r_ctl.ack_count);
7908 		else {
7909 			/* We really should not hit here */
7910 			ackratio = 1000;
7911 		}
7912 		if ((rack->sack_attack_disable  == 0) &&
7913 		    (ackratio > rack_highest_sack_thresh_seen))
7914 			rack_highest_sack_thresh_seen = (uint32_t)ackratio;
7915 		movetotal = rack->r_ctl.sack_moved_extra;
7916 		movetotal += rack->r_ctl.sack_noextra_move;
7917 		moveratio = rack->r_ctl.sack_moved_extra;
7918 		moveratio *= (uint64_t)1000;
7919 		if (movetotal)
7920 			moveratio /= movetotal;
7921 		else {
7922 			/* No moves, thats pretty good */
7923 			moveratio = 0;
7924 		}
7925 		if ((rack->sack_attack_disable == 0) &&
7926 		    (moveratio > rack_highest_move_thresh_seen))
7927 			rack_highest_move_thresh_seen = (uint32_t)moveratio;
7928 		if (rack->sack_attack_disable == 0) {
7929 			if ((ackratio > tcp_sack_to_ack_thresh) &&
7930 			    (moveratio > tcp_sack_to_move_thresh)) {
7931 				/* Disable sack processing */
7932 				rack->sack_attack_disable = 1;
7933 				if (rack->r_rep_attack == 0) {
7934 					rack->r_rep_attack = 1;
7935 					counter_u64_add(rack_sack_attacks_detected, 1);
7936 				}
7937 				if (tcp_attack_on_turns_on_logging) {
7938 					/*
7939 					 * Turn on logging, used for debugging
7940 					 * false positives.
7941 					 */
7942 					rack->rc_tp->t_logstate = tcp_attack_on_turns_on_logging;
7943 				}
7944 				/* Clamp the cwnd at flight size */
7945 				rack->r_ctl.rc_saved_cwnd = rack->rc_tp->snd_cwnd;
7946 				rack->rc_tp->snd_cwnd = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
7947 				rack_log_sad(rack, 2);
7948 			}
7949 		} else {
7950 			/* We are sack-disabled check for false positives */
7951 			if ((ackratio <= tcp_restoral_thresh) ||
7952 			    (rack->r_ctl.rc_num_maps_alloced  < tcp_map_minimum)) {
7953 				rack->sack_attack_disable  = 0;
7954 				rack_log_sad(rack, 3);
7955 				/* Restart counting */
7956 				rack->r_ctl.sack_count = 0;
7957 				rack->r_ctl.sack_moved_extra = 0;
7958 				rack->r_ctl.sack_noextra_move = 1;
7959 				rack->r_ctl.ack_count = max(1,
7960 				      (BYTES_THIS_ACK(tp, th)/ctf_fixed_maxseg(rack->rc_tp)));
7961 
7962 				if (rack->r_rep_reverse == 0) {
7963 					rack->r_rep_reverse = 1;
7964 					counter_u64_add(rack_sack_attacks_reversed, 1);
7965 				}
7966 				/* Restore the cwnd */
7967 				if (rack->r_ctl.rc_saved_cwnd > rack->rc_tp->snd_cwnd)
7968 					rack->rc_tp->snd_cwnd = rack->r_ctl.rc_saved_cwnd;
7969 			}
7970 		}
7971 	}
7972 #endif
7973 	if (changed) {
7974 		/* Something changed cancel the rack timer */
7975 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
7976 	}
7977 	tsused = tcp_ts_getticks();
7978 	rsm = tcp_rack_output(tp, rack, tsused);
7979 	if ((!IN_RECOVERY(tp->t_flags)) &&
7980 	    rsm) {
7981 		/* Enter recovery */
7982 		rack->r_ctl.rc_rsm_start = rsm->r_start;
7983 		rack->r_ctl.rc_cwnd_at = tp->snd_cwnd;
7984 		rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh;
7985 		entered_recovery = 1;
7986 		rack_cong_signal(tp, NULL, CC_NDUPACK);
7987 		/*
7988 		 * When we enter recovery we need to assure we send
7989 		 * one packet.
7990 		 */
7991 		if (rack->rack_no_prr == 0) {
7992 			rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
7993 			rack_log_to_prr(rack, 8, 0);
7994 		}
7995 		rack->r_timer_override = 1;
7996 		rack->r_early = 0;
7997 		rack->r_ctl.rc_agg_early = 0;
7998 	} else if (IN_RECOVERY(tp->t_flags) &&
7999 		   rsm &&
8000  		   (rack->r_rr_config == 3)) {
8001 		/*
8002 		 * Assure we can output and we get no
8003 		 * remembered pace time except the retransmit.
8004 		 */
8005 		rack->r_timer_override = 1;
8006 		rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
8007 		rack->r_ctl.rc_resend = rsm;
8008 	}
8009 	if (IN_RECOVERY(tp->t_flags) &&
8010 	    (rack->rack_no_prr == 0) &&
8011 	    (entered_recovery == 0)) {
8012 		/* Deal with PRR here (in recovery only) */
8013 		uint32_t pipe, snd_una;
8014 
8015 		rack->r_ctl.rc_prr_delivered += changed;
8016 		/* Compute prr_sndcnt */
8017 		if (SEQ_GT(tp->snd_una, th_ack)) {
8018 			snd_una = tp->snd_una;
8019 		} else {
8020 			snd_una = th_ack;
8021 		}
8022 		pipe = ((tp->snd_max - snd_una) - rack->r_ctl.rc_sacked) + rack->r_ctl.rc_holes_rxt;
8023 		if (pipe > tp->snd_ssthresh) {
8024 			long sndcnt;
8025 
8026 			sndcnt = rack->r_ctl.rc_prr_delivered * tp->snd_ssthresh;
8027 			if (rack->r_ctl.rc_prr_recovery_fs > 0)
8028 				sndcnt /= (long)rack->r_ctl.rc_prr_recovery_fs;
8029 			else {
8030 				rack->r_ctl.rc_prr_sndcnt = 0;
8031 				rack_log_to_prr(rack, 9, 0);
8032 				sndcnt = 0;
8033 			}
8034 			sndcnt++;
8035 			if (sndcnt > (long)rack->r_ctl.rc_prr_out)
8036 				sndcnt -= rack->r_ctl.rc_prr_out;
8037 			else
8038 				sndcnt = 0;
8039 			rack->r_ctl.rc_prr_sndcnt = sndcnt;
8040 			rack_log_to_prr(rack, 10, 0);
8041 		} else {
8042 			uint32_t limit;
8043 
8044 			if (rack->r_ctl.rc_prr_delivered > rack->r_ctl.rc_prr_out)
8045 				limit = (rack->r_ctl.rc_prr_delivered - rack->r_ctl.rc_prr_out);
8046 			else
8047 				limit = 0;
8048 			if (changed > limit)
8049 				limit = changed;
8050 			limit += ctf_fixed_maxseg(tp);
8051 			if (tp->snd_ssthresh > pipe) {
8052 				rack->r_ctl.rc_prr_sndcnt = min((tp->snd_ssthresh - pipe), limit);
8053 				rack_log_to_prr(rack, 11, 0);
8054 			} else {
8055 				rack->r_ctl.rc_prr_sndcnt = min(0, limit);
8056 				rack_log_to_prr(rack, 12, 0);
8057 			}
8058 		}
8059 		if ((rsm && (rack->r_ctl.rc_prr_sndcnt >= ctf_fixed_maxseg(tp)) &&
8060 		     ((rack->rc_inp->inp_in_hpts == 0) &&
8061 		      ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)))) {
8062 			/*
8063 			 * If you are pacing output you don't want
8064 			 * to override.
8065 			 */
8066 			rack->r_early = 0;
8067 			rack->r_ctl.rc_agg_early = 0;
8068 			rack->r_timer_override = 1;
8069 		}
8070 	}
8071 }
8072 
8073 static void
rack_strike_dupack(struct tcp_rack * rack)8074 rack_strike_dupack(struct tcp_rack *rack)
8075 {
8076 	struct rack_sendmap *rsm;
8077 
8078 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
8079 	if (rsm && (rsm->r_dupack < 0xff)) {
8080 		rsm->r_dupack++;
8081 		if (rsm->r_dupack >= DUP_ACK_THRESHOLD) {
8082 			rack->r_wanted_output = 1;
8083 			rack->r_timer_override = 1;
8084 			rack_log_retran_reason(rack, rsm, __LINE__, 1, 3);
8085 		} else {
8086 			rack_log_retran_reason(rack, rsm, __LINE__, 0, 3);
8087 		}
8088 	}
8089 }
8090 
8091 static void
rack_check_bottom_drag(struct tcpcb * tp,struct tcp_rack * rack,struct socket * so,int32_t acked)8092 rack_check_bottom_drag(struct tcpcb *tp,
8093 		       struct tcp_rack *rack,
8094 		       struct socket *so, int32_t acked)
8095 {
8096 	uint32_t segsiz, minseg;
8097 
8098 	segsiz = ctf_fixed_maxseg(tp);
8099 	minseg = segsiz;
8100 
8101 	if (tp->snd_max == tp->snd_una) {
8102 		/*
8103 		 * We are doing dynamic pacing and we are way
8104 		 * under. Basically everything got acked while
8105 		 * we were still waiting on the pacer to expire.
8106 		 *
8107 		 * This means we need to boost the b/w in
8108 		 * addition to any earlier boosting of
8109 		 * the multipler.
8110 		 */
8111 		rack->rc_dragged_bottom = 1;
8112 		rack_validate_multipliers_at_or_above100(rack);
8113 		/*
8114 		 * Lets use the segment bytes acked plus
8115 		 * the lowest RTT seen as the basis to
8116 		 * form a b/w estimate. This will be off
8117 		 * due to the fact that the true estimate
8118 		 * should be around 1/2 the time of the RTT
8119 		 * but we can settle for that.
8120 		 */
8121 		if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_VALID) &&
8122 		    acked) {
8123 			uint64_t bw, calc_bw, rtt;
8124 
8125 			rtt = rack->r_ctl.rack_rs.rs_us_rtt;
8126 			bw = acked;
8127 			calc_bw = bw * 1000000;
8128 			calc_bw /= rtt;
8129 			if (rack->r_ctl.last_max_bw &&
8130 			    (rack->r_ctl.last_max_bw < calc_bw)) {
8131 				/*
8132 				 * If we have a last calculated max bw
8133 				 * enforce it.
8134 				 */
8135 				calc_bw = rack->r_ctl.last_max_bw;
8136 			}
8137 			/* now plop it in */
8138 			if (rack->rc_gp_filled == 0) {
8139 				if (calc_bw > ONE_POINT_TWO_MEG) {
8140 					/*
8141 					 * If we have no measurement
8142 					 * don't let us set in more than
8143 					 * 1.2Mbps. If we are still too
8144 					 * low after pacing with this we
8145 					 * will hopefully have a max b/w
8146 					 * available to sanity check things.
8147 					 */
8148 					calc_bw = ONE_POINT_TWO_MEG;
8149 				}
8150 				rack->r_ctl.rc_rtt_diff = 0;
8151 				rack->r_ctl.gp_bw = calc_bw;
8152 				rack->rc_gp_filled = 1;
8153 				rack->r_ctl.num_avg = RACK_REQ_AVG;
8154 				rack_set_pace_segments(rack->rc_tp, rack, __LINE__);
8155 			} else if (calc_bw > rack->r_ctl.gp_bw) {
8156 				rack->r_ctl.rc_rtt_diff = 0;
8157 				rack->r_ctl.num_avg = RACK_REQ_AVG;
8158 				rack->r_ctl.gp_bw = calc_bw;
8159 				rack_set_pace_segments(rack->rc_tp, rack, __LINE__);
8160 			} else
8161 				rack_increase_bw_mul(rack, -1, 0, 0, 1);
8162 			/*
8163 			 * For acks over 1mss we do a extra boost to simulate
8164 			 * where we would get 2 acks (we want 110 for the mul).
8165 			 */
8166 			if (acked > segsiz)
8167 				rack_increase_bw_mul(rack, -1, 0, 0, 1);
8168 		} else {
8169 			/*
8170 			 * Huh, this should not be, settle
8171 			 * for just an old increase.
8172 			 */
8173 			rack_increase_bw_mul(rack, -1, 0, 0, 1);
8174 		}
8175 	} else if ((IN_RECOVERY(tp->t_flags) == 0) &&
8176 		   (sbavail(&so->so_snd) > max((segsiz * (4 + rack_req_segs)),
8177 					       minseg)) &&
8178 		   (rack->r_ctl.cwnd_to_use > max((segsiz * (rack_req_segs + 2)), minseg)) &&
8179 		   (tp->snd_wnd > max((segsiz * (rack_req_segs + 2)), minseg)) &&
8180 		   (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) <=
8181 		    (segsiz * rack_req_segs))) {
8182 		/*
8183 		 * We are doing dynamic GP pacing and
8184 		 * we have everything except 1MSS or less
8185 		 * bytes left out. We are still pacing away.
8186 		 * And there is data that could be sent, This
8187 		 * means we are inserting delayed ack time in
8188 		 * our measurements because we are pacing too slow.
8189 		 */
8190 		rack_validate_multipliers_at_or_above100(rack);
8191 		rack->rc_dragged_bottom = 1;
8192 		rack_increase_bw_mul(rack, -1, 0, 0, 1);
8193 	}
8194 }
8195 
8196 /*
8197  * Return value of 1, we do not need to call rack_process_data().
8198  * return value of 0, rack_process_data can be called.
8199  * For ret_val if its 0 the TCP is locked, if its non-zero
8200  * its unlocked and probably unsafe to touch the TCB.
8201  */
8202 static int
rack_process_ack(struct mbuf * m,struct tcphdr * th,struct socket * so,struct tcpcb * tp,struct tcpopt * to,uint32_t tiwin,int32_t tlen,int32_t * ofia,int32_t thflags,int32_t * ret_val)8203 rack_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so,
8204     struct tcpcb *tp, struct tcpopt *to,
8205     uint32_t tiwin, int32_t tlen,
8206     int32_t * ofia, int32_t thflags, int32_t * ret_val)
8207 {
8208 	int32_t ourfinisacked = 0;
8209 	int32_t nsegs, acked_amount;
8210 	int32_t acked;
8211 	struct mbuf *mfree;
8212 	struct tcp_rack *rack;
8213 	int32_t under_pacing = 0;
8214 	int32_t recovery = 0;
8215 
8216 	rack = (struct tcp_rack *)tp->t_fb_ptr;
8217 	if (SEQ_GT(th->th_ack, tp->snd_max)) {
8218 		ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val);
8219 		rack->r_wanted_output = 1;
8220 		return (1);
8221 	}
8222 	if (rack->rc_gp_filled &&
8223 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
8224 		under_pacing = 1;
8225 	}
8226 	if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) {
8227 		if (rack->rc_in_persist)
8228 			tp->t_rxtshift = 0;
8229 		if ((th->th_ack == tp->snd_una) && (tiwin == tp->snd_wnd))
8230 			rack_strike_dupack(rack);
8231 		rack_log_ack(tp, to, th);
8232 	}
8233 	if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
8234 		/*
8235 		 * Old ack, behind (or duplicate to) the last one rcv'd
8236 		 * Note: Should mark reordering is occuring! We should also
8237 		 * look for sack blocks arriving e.g. ack 1, 4-4 then ack 1,
8238 		 * 3-3, 4-4 would be reording. As well as ack 1, 3-3 <no
8239 		 * retran and> ack 3
8240 		 */
8241 		return (0);
8242 	}
8243 	/*
8244 	 * If we reach this point, ACK is not a duplicate, i.e., it ACKs
8245 	 * something we sent.
8246 	 */
8247 	if (tp->t_flags & TF_NEEDSYN) {
8248 		/*
8249 		 * T/TCP: Connection was half-synchronized, and our SYN has
8250 		 * been ACK'd (so connection is now fully synchronized).  Go
8251 		 * to non-starred state, increment snd_una for ACK of SYN,
8252 		 * and check if we can do window scaling.
8253 		 */
8254 		tp->t_flags &= ~TF_NEEDSYN;
8255 		tp->snd_una++;
8256 		/* Do window scaling? */
8257 		if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
8258 		    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
8259 			tp->rcv_scale = tp->request_r_scale;
8260 			/* Send window already scaled. */
8261 		}
8262 	}
8263 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
8264 	INP_WLOCK_ASSERT(tp->t_inpcb);
8265 
8266 	acked = BYTES_THIS_ACK(tp, th);
8267 	KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
8268 	KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
8269 	/*
8270 	 * If we just performed our first retransmit, and the ACK arrives
8271 	 * within our recovery window, then it was a mistake to do the
8272 	 * retransmit in the first place.  Recover our original cwnd and
8273 	 * ssthresh, and proceed to transmit where we left off.
8274 	 */
8275 	if (tp->t_flags & TF_PREVVALID) {
8276 		tp->t_flags &= ~TF_PREVVALID;
8277 		if (tp->t_rxtshift == 1 &&
8278 		    (int)(ticks - tp->t_badrxtwin) < 0)
8279 			rack_cong_signal(tp, th, CC_RTO_ERR);
8280 	}
8281 	if (acked) {
8282 		/* assure we are not backed off */
8283 		tp->t_rxtshift = 0;
8284 		rack->rc_tlp_in_progress = 0;
8285 		rack->r_ctl.rc_tlp_cnt_out = 0;
8286 		/*
8287 		 * If it is the RXT timer we want to
8288 		 * stop it, so we can restart a TLP.
8289 		 */
8290 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
8291 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
8292 #ifdef NETFLIX_HTTP_LOGGING
8293 		tcp_http_check_for_comp(rack->rc_tp, th->th_ack);
8294 #endif
8295 	}
8296 	/*
8297 	 * If we have a timestamp reply, update smoothed round trip time. If
8298 	 * no timestamp is present but transmit timer is running and timed
8299 	 * sequence number was acked, update smoothed round trip time. Since
8300 	 * we now have an rtt measurement, cancel the timer backoff (cf.,
8301 	 * Phil Karn's retransmit alg.). Recompute the initial retransmit
8302 	 * timer.
8303 	 *
8304 	 * Some boxes send broken timestamp replies during the SYN+ACK
8305 	 * phase, ignore timestamps of 0 or we could calculate a huge RTT
8306 	 * and blow up the retransmit timer.
8307 	 */
8308 	/*
8309 	 * If all outstanding data is acked, stop retransmit timer and
8310 	 * remember to restart (more output or persist). If there is more
8311 	 * data to be acked, restart retransmit timer, using current
8312 	 * (possibly backed-off) value.
8313 	 */
8314 	if (acked == 0) {
8315 		if (ofia)
8316 			*ofia = ourfinisacked;
8317 		return (0);
8318 	}
8319 	if (rack->r_ctl.rc_early_recovery) {
8320 		if (IN_RECOVERY(tp->t_flags)) {
8321 			if (SEQ_LT(th->th_ack, tp->snd_recover) &&
8322 			    (SEQ_LT(th->th_ack, tp->snd_max))) {
8323 				tcp_rack_partialack(tp, th);
8324 			} else {
8325 				rack_post_recovery(tp, th);
8326 				recovery = 1;
8327 			}
8328 		}
8329 	}
8330 	/*
8331 	 * Let the congestion control algorithm update congestion control
8332 	 * related information. This typically means increasing the
8333 	 * congestion window.
8334 	 */
8335 	rack_ack_received(tp, rack, th, nsegs, CC_ACK, recovery);
8336 	SOCKBUF_LOCK(&so->so_snd);
8337 	acked_amount = min(acked, (int)sbavail(&so->so_snd));
8338 	tp->snd_wnd -= acked_amount;
8339 	mfree = sbcut_locked(&so->so_snd, acked_amount);
8340 	if ((sbused(&so->so_snd) == 0) &&
8341 	    (acked > acked_amount) &&
8342 	    (tp->t_state >= TCPS_FIN_WAIT_1) &&
8343 	    (tp->t_flags & TF_SENTFIN)) {
8344 		/*
8345 		 * We must be sure our fin
8346 		 * was sent and acked (we can be
8347 		 * in FIN_WAIT_1 without having
8348 		 * sent the fin).
8349 		 */
8350 		ourfinisacked = 1;
8351 	}
8352 	SOCKBUF_UNLOCK(&so->so_snd);
8353 	tp->t_flags |= TF_WAKESOW;
8354 	m_freem(mfree);
8355 	if (rack->r_ctl.rc_early_recovery == 0) {
8356 		if (IN_RECOVERY(tp->t_flags)) {
8357 			if (SEQ_LT(th->th_ack, tp->snd_recover) &&
8358 			    (SEQ_LT(th->th_ack, tp->snd_max))) {
8359 				tcp_rack_partialack(tp, th);
8360 			} else {
8361 				rack_post_recovery(tp, th);
8362 			}
8363 		}
8364 	}
8365 	tp->snd_una = th->th_ack;
8366 	if (SEQ_GT(tp->snd_una, tp->snd_recover))
8367 		tp->snd_recover = tp->snd_una;
8368 
8369 	if (SEQ_LT(tp->snd_nxt, tp->snd_una)) {
8370 		tp->snd_nxt = tp->snd_una;
8371 	}
8372 	if (under_pacing &&
8373 	    (rack->use_fixed_rate == 0) &&
8374 	    (rack->in_probe_rtt == 0) &&
8375 	    rack->rc_gp_dyn_mul &&
8376 	    rack->rc_always_pace) {
8377 		/* Check if we are dragging bottom */
8378 		rack_check_bottom_drag(tp, rack, so, acked);
8379 	}
8380 	if (tp->snd_una == tp->snd_max) {
8381 		/* Nothing left outstanding */
8382 		rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
8383 		if (rack->r_ctl.rc_went_idle_time == 0)
8384 			rack->r_ctl.rc_went_idle_time = 1;
8385 		rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
8386 		if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0)
8387 			tp->t_acktime = 0;
8388 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
8389 		/* Set need output so persist might get set */
8390 		rack->r_wanted_output = 1;
8391 		sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
8392 		if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
8393 		    (sbavail(&so->so_snd) == 0) &&
8394 		    (tp->t_flags2 & TF2_DROP_AF_DATA)) {
8395 			/*
8396 			 * The socket was gone and the
8397 			 * peer sent data, time to
8398 			 * reset him.
8399 			 */
8400 			*ret_val = 1;
8401 			/* tcp_close will kill the inp pre-log the Reset */
8402 			tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
8403 			tp = tcp_close(tp);
8404 			ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen);
8405 			return (1);
8406 		}
8407 	}
8408 	if (ofia)
8409 		*ofia = ourfinisacked;
8410 	return (0);
8411 }
8412 
8413 static void
rack_collapsed_window(struct tcp_rack * rack)8414 rack_collapsed_window(struct tcp_rack *rack)
8415 {
8416 	/*
8417 	 * Now we must walk the
8418 	 * send map and divide the
8419 	 * ones left stranded. These
8420 	 * guys can't cause us to abort
8421 	 * the connection and are really
8422 	 * "unsent". However if a buggy
8423 	 * client actually did keep some
8424 	 * of the data i.e. collapsed the win
8425 	 * and refused to ack and then opened
8426 	 * the win and acked that data. We would
8427 	 * get into an ack war, the simplier
8428 	 * method then of just pretending we
8429 	 * did not send those segments something
8430 	 * won't work.
8431 	 */
8432 	struct rack_sendmap *rsm, *nrsm, fe, *insret;
8433 	tcp_seq max_seq;
8434 
8435 	max_seq = rack->rc_tp->snd_una + rack->rc_tp->snd_wnd;
8436 	memset(&fe, 0, sizeof(fe));
8437 	fe.r_start = max_seq;
8438 	/* Find the first seq past or at maxseq */
8439 	rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
8440 	if (rsm == NULL) {
8441 		/* Nothing to do strange */
8442 		rack->rc_has_collapsed = 0;
8443 		return;
8444 	}
8445 	/*
8446 	 * Now do we need to split at
8447 	 * the collapse point?
8448 	 */
8449 	if (SEQ_GT(max_seq, rsm->r_start)) {
8450 		nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
8451 		if (nrsm == NULL) {
8452 			/* We can't get a rsm, mark all? */
8453 			nrsm = rsm;
8454 			goto no_split;
8455 		}
8456 		/* Clone it */
8457 		rack_clone_rsm(rack, nrsm, rsm, max_seq);
8458 		insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8459 #ifdef INVARIANTS
8460 		if (insret != NULL) {
8461 			panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
8462 			      nrsm, insret, rack, rsm);
8463 		}
8464 #endif
8465 		if (rsm->r_in_tmap) {
8466 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8467 			nrsm->r_in_tmap = 1;
8468 		}
8469 		/*
8470 		 * Set in the new RSM as the
8471 		 * collapsed starting point
8472 		 */
8473 		rsm = nrsm;
8474 	}
8475 no_split:
8476 	counter_u64_add(rack_collapsed_win, 1);
8477 	RB_FOREACH_FROM(nrsm, rack_rb_tree_head, rsm) {
8478 		nrsm->r_flags |= RACK_RWND_COLLAPSED;
8479 		rack->rc_has_collapsed = 1;
8480 	}
8481 }
8482 
8483 static void
rack_un_collapse_window(struct tcp_rack * rack)8484 rack_un_collapse_window(struct tcp_rack *rack)
8485 {
8486 	struct rack_sendmap *rsm;
8487 
8488 	RB_FOREACH_REVERSE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
8489 		if (rsm->r_flags & RACK_RWND_COLLAPSED)
8490 			rsm->r_flags &= ~RACK_RWND_COLLAPSED;
8491 		else
8492 			break;
8493 	}
8494 	rack->rc_has_collapsed = 0;
8495 }
8496 
8497 static void
rack_handle_delayed_ack(struct tcpcb * tp,struct tcp_rack * rack,int32_t tlen,int32_t tfo_syn)8498 rack_handle_delayed_ack(struct tcpcb *tp, struct tcp_rack *rack,
8499 			int32_t tlen, int32_t tfo_syn)
8500 {
8501 	if (DELAY_ACK(tp, tlen) || tfo_syn) {
8502 		if (rack->rc_dack_mode &&
8503 		    (tlen > 500) &&
8504 		    (rack->rc_dack_toggle == 1)) {
8505 			goto no_delayed_ack;
8506 		}
8507 		rack_timer_cancel(tp, rack,
8508 				  rack->r_ctl.rc_rcvtime, __LINE__);
8509 		tp->t_flags |= TF_DELACK;
8510 	} else {
8511 no_delayed_ack:
8512 		rack->r_wanted_output = 1;
8513 		tp->t_flags |= TF_ACKNOW;
8514 		if (rack->rc_dack_mode) {
8515 			if (tp->t_flags & TF_DELACK)
8516 				rack->rc_dack_toggle = 1;
8517 			else
8518 				rack->rc_dack_toggle = 0;
8519 		}
8520 	}
8521 }
8522 /*
8523  * Return value of 1, the TCB is unlocked and most
8524  * likely gone, return value of 0, the TCP is still
8525  * locked.
8526  */
8527 static int
rack_process_data(struct mbuf * m,struct tcphdr * th,struct socket * so,struct tcpcb * tp,int32_t drop_hdrlen,int32_t tlen,uint32_t tiwin,int32_t thflags,int32_t nxt_pkt)8528 rack_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so,
8529     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
8530     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
8531 {
8532 	/*
8533 	 * Update window information. Don't look at window if no ACK: TAC's
8534 	 * send garbage on first SYN.
8535 	 */
8536 	int32_t nsegs;
8537 	int32_t tfo_syn;
8538 	struct tcp_rack *rack;
8539 
8540 	rack = (struct tcp_rack *)tp->t_fb_ptr;
8541 	INP_WLOCK_ASSERT(tp->t_inpcb);
8542 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
8543 	if ((thflags & TH_ACK) &&
8544 	    (SEQ_LT(tp->snd_wl1, th->th_seq) ||
8545 	    (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) ||
8546 	    (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) {
8547 		/* keep track of pure window updates */
8548 		if (tlen == 0 &&
8549 		    tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd)
8550 			KMOD_TCPSTAT_INC(tcps_rcvwinupd);
8551 		tp->snd_wnd = tiwin;
8552 		tp->snd_wl1 = th->th_seq;
8553 		tp->snd_wl2 = th->th_ack;
8554 		if (tp->snd_wnd > tp->max_sndwnd)
8555 			tp->max_sndwnd = tp->snd_wnd;
8556 		rack->r_wanted_output = 1;
8557 	} else if (thflags & TH_ACK) {
8558 		if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) {
8559 			tp->snd_wnd = tiwin;
8560 			tp->snd_wl1 = th->th_seq;
8561 			tp->snd_wl2 = th->th_ack;
8562 		}
8563 	}
8564 	if (tp->snd_wnd < ctf_outstanding(tp))
8565 		/* The peer collapsed the window */
8566 		rack_collapsed_window(rack);
8567 	else if (rack->rc_has_collapsed)
8568 		rack_un_collapse_window(rack);
8569 	/* Was persist timer active and now we have window space? */
8570 	if ((rack->rc_in_persist != 0) &&
8571 	    (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
8572 				rack->r_ctl.rc_pace_min_segs))) {
8573 		rack_exit_persist(tp, rack, rack->r_ctl.rc_rcvtime);
8574 		tp->snd_nxt = tp->snd_max;
8575 		/* Make sure we output to start the timer */
8576 		rack->r_wanted_output = 1;
8577 	}
8578 	/* Do we enter persists? */
8579 	if ((rack->rc_in_persist == 0) &&
8580 	    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
8581 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
8582 	    (tp->snd_max == tp->snd_una) &&
8583 	    sbavail(&tp->t_inpcb->inp_socket->so_snd) &&
8584 	    (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) {
8585 		/*
8586 		 * Here the rwnd is less than
8587 		 * the pacing size, we are established,
8588 		 * nothing is outstanding, and there is
8589 		 * data to send. Enter persists.
8590 		 */
8591 		tp->snd_nxt = tp->snd_una;
8592 		rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
8593 	}
8594 	if (tp->t_flags2 & TF2_DROP_AF_DATA) {
8595 		m_freem(m);
8596 		return (0);
8597 	}
8598 	/*
8599 	 * don't process the URG bit, ignore them drag
8600 	 * along the up.
8601 	 */
8602 	tp->rcv_up = tp->rcv_nxt;
8603 	INP_WLOCK_ASSERT(tp->t_inpcb);
8604 
8605 	/*
8606 	 * Process the segment text, merging it into the TCP sequencing
8607 	 * queue, and arranging for acknowledgment of receipt if necessary.
8608 	 * This process logically involves adjusting tp->rcv_wnd as data is
8609 	 * presented to the user (this happens in tcp_usrreq.c, case
8610 	 * PRU_RCVD).  If a FIN has already been received on this connection
8611 	 * then we just ignore the text.
8612 	 */
8613 	tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) &&
8614 		   IS_FASTOPEN(tp->t_flags));
8615 	if ((tlen || (thflags & TH_FIN) || (tfo_syn && tlen > 0)) &&
8616 	    TCPS_HAVERCVDFIN(tp->t_state) == 0) {
8617 		tcp_seq save_start = th->th_seq;
8618 		tcp_seq save_rnxt  = tp->rcv_nxt;
8619 		int     save_tlen  = tlen;
8620 
8621 		m_adj(m, drop_hdrlen);	/* delayed header drop */
8622 		/*
8623 		 * Insert segment which includes th into TCP reassembly
8624 		 * queue with control block tp.  Set thflags to whether
8625 		 * reassembly now includes a segment with FIN.  This handles
8626 		 * the common case inline (segment is the next to be
8627 		 * received on an established connection, and the queue is
8628 		 * empty), avoiding linkage into and removal from the queue
8629 		 * and repetition of various conversions. Set DELACK for
8630 		 * segments received in order, but ack immediately when
8631 		 * segments are out of order (so fast retransmit can work).
8632 		 */
8633 		if (th->th_seq == tp->rcv_nxt &&
8634 		    SEGQ_EMPTY(tp) &&
8635 		    (TCPS_HAVEESTABLISHED(tp->t_state) ||
8636 		    tfo_syn)) {
8637 #ifdef NETFLIX_SB_LIMITS
8638 			u_int mcnt, appended;
8639 
8640 			if (so->so_rcv.sb_shlim) {
8641 				mcnt = m_memcnt(m);
8642 				appended = 0;
8643 				if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
8644 				    CFO_NOSLEEP, NULL) == false) {
8645 					counter_u64_add(tcp_sb_shlim_fails, 1);
8646 					m_freem(m);
8647 					return (0);
8648 				}
8649 			}
8650 #endif
8651 			rack_handle_delayed_ack(tp, rack, tlen, tfo_syn);
8652 			tp->rcv_nxt += tlen;
8653 			if (tlen &&
8654 			    ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
8655 			    (tp->t_fbyte_in == 0)) {
8656 				tp->t_fbyte_in = ticks;
8657 				if (tp->t_fbyte_in == 0)
8658 					tp->t_fbyte_in = 1;
8659 				if (tp->t_fbyte_out && tp->t_fbyte_in)
8660 					tp->t_flags2 |= TF2_FBYTES_COMPLETE;
8661 			}
8662 			thflags = th->th_flags & TH_FIN;
8663 			KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
8664 			KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
8665 			SOCKBUF_LOCK(&so->so_rcv);
8666 			if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
8667 				m_freem(m);
8668 			} else
8669 #ifdef NETFLIX_SB_LIMITS
8670 				appended =
8671 #endif
8672 					sbappendstream_locked(&so->so_rcv, m, 0);
8673 			SOCKBUF_UNLOCK(&so->so_rcv);
8674 			tp->t_flags |= TF_WAKESOR;
8675 #ifdef NETFLIX_SB_LIMITS
8676 			if (so->so_rcv.sb_shlim && appended != mcnt)
8677 				counter_fo_release(so->so_rcv.sb_shlim,
8678 				    mcnt - appended);
8679 #endif
8680 		} else {
8681 			/*
8682 			 * XXX: Due to the header drop above "th" is
8683 			 * theoretically invalid by now.  Fortunately
8684 			 * m_adj() doesn't actually frees any mbufs when
8685 			 * trimming from the head.
8686 			 */
8687 			tcp_seq temp = save_start;
8688 			thflags = tcp_reass(tp, th, &temp, &tlen, m);
8689 			tp->t_flags |= TF_ACKNOW;
8690 		}
8691                 if ((tp->t_flags & TF_SACK_PERMIT) && (save_tlen > 0)) {
8692                         if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) {
8693                                 /*
8694                                  * DSACK actually handled in the fastpath
8695                                  * above.
8696                                  */
8697 				RACK_OPTS_INC(tcp_sack_path_1);
8698                                 tcp_update_sack_list(tp, save_start,
8699                                     save_start + save_tlen);
8700                         } else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) {
8701                                 if ((tp->rcv_numsacks >= 1) &&
8702                                     (tp->sackblks[0].end == save_start)) {
8703                                         /*
8704                                          * Partial overlap, recorded at todrop
8705                                          * above.
8706                                          */
8707 					RACK_OPTS_INC(tcp_sack_path_2a);
8708                                         tcp_update_sack_list(tp,
8709                                             tp->sackblks[0].start,
8710                                             tp->sackblks[0].end);
8711                                 } else {
8712 					RACK_OPTS_INC(tcp_sack_path_2b);
8713                                         tcp_update_dsack_list(tp, save_start,
8714                                             save_start + save_tlen);
8715                                 }
8716                         } else if (tlen >= save_tlen) {
8717                                 /* Update of sackblks. */
8718 				RACK_OPTS_INC(tcp_sack_path_3);
8719                                 tcp_update_dsack_list(tp, save_start,
8720                                     save_start + save_tlen);
8721                         } else if (tlen > 0) {
8722 				RACK_OPTS_INC(tcp_sack_path_4);
8723                                 tcp_update_dsack_list(tp, save_start,
8724                                     save_start + tlen);
8725                         }
8726                 }
8727 	} else {
8728 		m_freem(m);
8729 		thflags &= ~TH_FIN;
8730 	}
8731 
8732 	/*
8733 	 * If FIN is received ACK the FIN and let the user know that the
8734 	 * connection is closing.
8735 	 */
8736 	if (thflags & TH_FIN) {
8737 		if (TCPS_HAVERCVDFIN(tp->t_state) == 0) {
8738 			socantrcvmore(so);
8739 			/* The socket upcall is handled by socantrcvmore. */
8740 			tp->t_flags &= ~TF_WAKESOR;
8741 			/*
8742 			 * If connection is half-synchronized (ie NEEDSYN
8743 			 * flag on) then delay ACK, so it may be piggybacked
8744 			 * when SYN is sent. Otherwise, since we received a
8745 			 * FIN then no more input can be expected, send ACK
8746 			 * now.
8747 			 */
8748 			if (tp->t_flags & TF_NEEDSYN) {
8749 				rack_timer_cancel(tp, rack,
8750 				    rack->r_ctl.rc_rcvtime, __LINE__);
8751 				tp->t_flags |= TF_DELACK;
8752 			} else {
8753 				tp->t_flags |= TF_ACKNOW;
8754 			}
8755 			tp->rcv_nxt++;
8756 		}
8757 		switch (tp->t_state) {
8758 			/*
8759 			 * In SYN_RECEIVED and ESTABLISHED STATES enter the
8760 			 * CLOSE_WAIT state.
8761 			 */
8762 		case TCPS_SYN_RECEIVED:
8763 			tp->t_starttime = ticks;
8764 			/* FALLTHROUGH */
8765 		case TCPS_ESTABLISHED:
8766 			rack_timer_cancel(tp, rack,
8767 			    rack->r_ctl.rc_rcvtime, __LINE__);
8768 			tcp_state_change(tp, TCPS_CLOSE_WAIT);
8769 			break;
8770 
8771 			/*
8772 			 * If still in FIN_WAIT_1 STATE FIN has not been
8773 			 * acked so enter the CLOSING state.
8774 			 */
8775 		case TCPS_FIN_WAIT_1:
8776 			rack_timer_cancel(tp, rack,
8777 			    rack->r_ctl.rc_rcvtime, __LINE__);
8778 			tcp_state_change(tp, TCPS_CLOSING);
8779 			break;
8780 
8781 			/*
8782 			 * In FIN_WAIT_2 state enter the TIME_WAIT state,
8783 			 * starting the time-wait timer, turning off the
8784 			 * other standard timers.
8785 			 */
8786 		case TCPS_FIN_WAIT_2:
8787 			rack_timer_cancel(tp, rack,
8788 			    rack->r_ctl.rc_rcvtime, __LINE__);
8789 			tcp_twstart(tp);
8790 			return (1);
8791 		}
8792 	}
8793 	/*
8794 	 * Return any desired output.
8795 	 */
8796 	if ((tp->t_flags & TF_ACKNOW) ||
8797 	    (sbavail(&so->so_snd) > (tp->snd_max - tp->snd_una))) {
8798 		rack->r_wanted_output = 1;
8799 	}
8800 	INP_WLOCK_ASSERT(tp->t_inpcb);
8801 	return (0);
8802 }
8803 
8804 /*
8805  * Here nothing is really faster, its just that we
8806  * have broken out the fast-data path also just like
8807  * the fast-ack.
8808  */
8809 static int
rack_do_fastnewdata(struct mbuf * m,struct tcphdr * th,struct socket * so,struct tcpcb * tp,struct tcpopt * to,int32_t drop_hdrlen,int32_t tlen,uint32_t tiwin,int32_t nxt_pkt,uint8_t iptos)8810 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so,
8811     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
8812     uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos)
8813 {
8814 	int32_t nsegs;
8815 	int32_t newsize = 0;	/* automatic sockbuf scaling */
8816 	struct tcp_rack *rack;
8817 #ifdef NETFLIX_SB_LIMITS
8818 	u_int mcnt, appended;
8819 #endif
8820 #ifdef TCPDEBUG
8821 	/*
8822 	 * The size of tcp_saveipgen must be the size of the max ip header,
8823 	 * now IPv6.
8824 	 */
8825 	u_char tcp_saveipgen[IP6_HDR_LEN];
8826 	struct tcphdr tcp_savetcp;
8827 	short ostate = 0;
8828 
8829 #endif
8830 	/*
8831 	 * If last ACK falls within this segment's sequence numbers, record
8832 	 * the timestamp. NOTE that the test is modified according to the
8833 	 * latest proposal of the [email protected] list (Braden 1993/04/26).
8834 	 */
8835 	if (__predict_false(th->th_seq != tp->rcv_nxt)) {
8836 		return (0);
8837 	}
8838 	if (__predict_false(tp->snd_nxt != tp->snd_max)) {
8839 		return (0);
8840 	}
8841 	if (tiwin && tiwin != tp->snd_wnd) {
8842 		return (0);
8843 	}
8844 	if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) {
8845 		return (0);
8846 	}
8847 	if (__predict_false((to->to_flags & TOF_TS) &&
8848 	    (TSTMP_LT(to->to_tsval, tp->ts_recent)))) {
8849 		return (0);
8850 	}
8851 	if (__predict_false((th->th_ack != tp->snd_una))) {
8852 		return (0);
8853 	}
8854 	if (__predict_false(tlen > sbspace(&so->so_rcv))) {
8855 		return (0);
8856 	}
8857 	if ((to->to_flags & TOF_TS) != 0 &&
8858 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
8859 		tp->ts_recent_age = tcp_ts_getticks();
8860 		tp->ts_recent = to->to_tsval;
8861 	}
8862 	rack = (struct tcp_rack *)tp->t_fb_ptr;
8863 	/*
8864 	 * This is a pure, in-sequence data packet with nothing on the
8865 	 * reassembly queue and we have enough buffer space to take it.
8866 	 */
8867 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
8868 
8869 #ifdef NETFLIX_SB_LIMITS
8870 	if (so->so_rcv.sb_shlim) {
8871 		mcnt = m_memcnt(m);
8872 		appended = 0;
8873 		if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
8874 		    CFO_NOSLEEP, NULL) == false) {
8875 			counter_u64_add(tcp_sb_shlim_fails, 1);
8876 			m_freem(m);
8877 			return (1);
8878 		}
8879 	}
8880 #endif
8881 	/* Clean receiver SACK report if present */
8882 	if (tp->rcv_numsacks)
8883 		tcp_clean_sackreport(tp);
8884 	KMOD_TCPSTAT_INC(tcps_preddat);
8885 	tp->rcv_nxt += tlen;
8886 	if (tlen &&
8887 	    ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
8888 	    (tp->t_fbyte_in == 0)) {
8889 		tp->t_fbyte_in = ticks;
8890 		if (tp->t_fbyte_in == 0)
8891 			tp->t_fbyte_in = 1;
8892 		if (tp->t_fbyte_out && tp->t_fbyte_in)
8893 			tp->t_flags2 |= TF2_FBYTES_COMPLETE;
8894 	}
8895 	/*
8896 	 * Pull snd_wl1 up to prevent seq wrap relative to th_seq.
8897 	 */
8898 	tp->snd_wl1 = th->th_seq;
8899 	/*
8900 	 * Pull rcv_up up to prevent seq wrap relative to rcv_nxt.
8901 	 */
8902 	tp->rcv_up = tp->rcv_nxt;
8903 	KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
8904 	KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
8905 #ifdef TCPDEBUG
8906 	if (so->so_options & SO_DEBUG)
8907 		tcp_trace(TA_INPUT, ostate, tp,
8908 		    (void *)tcp_saveipgen, &tcp_savetcp, 0);
8909 #endif
8910 	newsize = tcp_autorcvbuf(m, th, so, tp, tlen);
8911 
8912 	/* Add data to socket buffer. */
8913 	SOCKBUF_LOCK(&so->so_rcv);
8914 	if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
8915 		m_freem(m);
8916 	} else {
8917 		/*
8918 		 * Set new socket buffer size. Give up when limit is
8919 		 * reached.
8920 		 */
8921 		if (newsize)
8922 			if (!sbreserve_locked(&so->so_rcv,
8923 			    newsize, so, NULL))
8924 				so->so_rcv.sb_flags &= ~SB_AUTOSIZE;
8925 		m_adj(m, drop_hdrlen);	/* delayed header drop */
8926 #ifdef NETFLIX_SB_LIMITS
8927 		appended =
8928 #endif
8929 			sbappendstream_locked(&so->so_rcv, m, 0);
8930 		ctf_calc_rwin(so, tp);
8931 	}
8932 	SOCKBUF_UNLOCK(&so->so_rcv);
8933 	tp->t_flags |= TF_WAKESOR;
8934 #ifdef NETFLIX_SB_LIMITS
8935 	if (so->so_rcv.sb_shlim && mcnt != appended)
8936 		counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended);
8937 #endif
8938 	rack_handle_delayed_ack(tp, rack, tlen, 0);
8939 	if (tp->snd_una == tp->snd_max)
8940 		sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
8941 	return (1);
8942 }
8943 
8944 /*
8945  * This subfunction is used to try to highly optimize the
8946  * fast path. We again allow window updates that are
8947  * in sequence to remain in the fast-path. We also add
8948  * in the __predict's to attempt to help the compiler.
8949  * Note that if we return a 0, then we can *not* process
8950  * it and the caller should push the packet into the
8951  * slow-path.
8952  */
8953 static int
rack_fastack(struct mbuf * m,struct tcphdr * th,struct socket * so,struct tcpcb * tp,struct tcpopt * to,int32_t drop_hdrlen,int32_t tlen,uint32_t tiwin,int32_t nxt_pkt,uint32_t cts)8954 rack_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
8955     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
8956     uint32_t tiwin, int32_t nxt_pkt, uint32_t cts)
8957 {
8958 	int32_t acked;
8959 	int32_t nsegs;
8960 #ifdef TCPDEBUG
8961 	/*
8962 	 * The size of tcp_saveipgen must be the size of the max ip header,
8963 	 * now IPv6.
8964 	 */
8965 	u_char tcp_saveipgen[IP6_HDR_LEN];
8966 	struct tcphdr tcp_savetcp;
8967 	short ostate = 0;
8968 #endif
8969 	int32_t under_pacing = 0;
8970 	struct tcp_rack *rack;
8971 
8972 	if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
8973 		/* Old ack, behind (or duplicate to) the last one rcv'd */
8974 		return (0);
8975 	}
8976 	if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) {
8977 		/* Above what we have sent? */
8978 		return (0);
8979 	}
8980 	if (__predict_false(tp->snd_nxt != tp->snd_max)) {
8981 		/* We are retransmitting */
8982 		return (0);
8983 	}
8984 	if (__predict_false(tiwin == 0)) {
8985 		/* zero window */
8986 		return (0);
8987 	}
8988 	if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) {
8989 		/* We need a SYN or a FIN, unlikely.. */
8990 		return (0);
8991 	}
8992 	if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) {
8993 		/* Timestamp is behind .. old ack with seq wrap? */
8994 		return (0);
8995 	}
8996 	if (__predict_false(IN_RECOVERY(tp->t_flags))) {
8997 		/* Still recovering */
8998 		return (0);
8999 	}
9000 	rack = (struct tcp_rack *)tp->t_fb_ptr;
9001 	if (rack->r_ctl.rc_sacked) {
9002 		/* We have sack holes on our scoreboard */
9003 		return (0);
9004 	}
9005 	/* Ok if we reach here, we can process a fast-ack */
9006 	if (rack->rc_gp_filled &&
9007 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
9008 		under_pacing = 1;
9009 	}
9010 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
9011 	rack_log_ack(tp, to, th);
9012 	/* Did the window get updated? */
9013 	if (tiwin != tp->snd_wnd) {
9014 		tp->snd_wnd = tiwin;
9015 		tp->snd_wl1 = th->th_seq;
9016 		if (tp->snd_wnd > tp->max_sndwnd)
9017 			tp->max_sndwnd = tp->snd_wnd;
9018 	}
9019 	/* Do we exit persists? */
9020 	if ((rack->rc_in_persist != 0) &&
9021 	    (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
9022 			       rack->r_ctl.rc_pace_min_segs))) {
9023 		rack_exit_persist(tp, rack, cts);
9024 	}
9025 	/* Do we enter persists? */
9026 	if ((rack->rc_in_persist == 0) &&
9027 	    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
9028 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
9029 	    (tp->snd_max == tp->snd_una) &&
9030 	    sbavail(&tp->t_inpcb->inp_socket->so_snd) &&
9031 	    (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) {
9032 		/*
9033 		 * Here the rwnd is less than
9034 		 * the pacing size, we are established,
9035 		 * nothing is outstanding, and there is
9036 		 * data to send. Enter persists.
9037 		 */
9038 		tp->snd_nxt = tp->snd_una;
9039 		rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
9040 	}
9041 	/*
9042 	 * If last ACK falls within this segment's sequence numbers, record
9043 	 * the timestamp. NOTE that the test is modified according to the
9044 	 * latest proposal of the [email protected] list (Braden 1993/04/26).
9045 	 */
9046 	if ((to->to_flags & TOF_TS) != 0 &&
9047 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
9048 		tp->ts_recent_age = tcp_ts_getticks();
9049 		tp->ts_recent = to->to_tsval;
9050 	}
9051 	/*
9052 	 * This is a pure ack for outstanding data.
9053 	 */
9054 	KMOD_TCPSTAT_INC(tcps_predack);
9055 
9056 	/*
9057 	 * "bad retransmit" recovery.
9058 	 */
9059 	if (tp->t_flags & TF_PREVVALID) {
9060 		tp->t_flags &= ~TF_PREVVALID;
9061 		if (tp->t_rxtshift == 1 &&
9062 		    (int)(ticks - tp->t_badrxtwin) < 0)
9063 			rack_cong_signal(tp, th, CC_RTO_ERR);
9064 	}
9065 	/*
9066 	 * Recalculate the transmit timer / rtt.
9067 	 *
9068 	 * Some boxes send broken timestamp replies during the SYN+ACK
9069 	 * phase, ignore timestamps of 0 or we could calculate a huge RTT
9070 	 * and blow up the retransmit timer.
9071 	 */
9072 	acked = BYTES_THIS_ACK(tp, th);
9073 
9074 #ifdef TCP_HHOOK
9075 	/* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */
9076 	hhook_run_tcp_est_in(tp, th, to);
9077 #endif
9078 
9079 	KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
9080 	KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
9081 	sbdrop(&so->so_snd, acked);
9082 	if (acked) {
9083 		/* assure we are not backed off */
9084 		tp->t_rxtshift = 0;
9085 		rack->rc_tlp_in_progress = 0;
9086 		rack->r_ctl.rc_tlp_cnt_out = 0;
9087 		/*
9088 		 * If it is the RXT timer we want to
9089 		 * stop it, so we can restart a TLP.
9090 		 */
9091 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
9092 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
9093 #ifdef NETFLIX_HTTP_LOGGING
9094 		tcp_http_check_for_comp(rack->rc_tp, th->th_ack);
9095 #endif
9096 	}
9097 	/*
9098 	 * Let the congestion control algorithm update congestion control
9099 	 * related information. This typically means increasing the
9100 	 * congestion window.
9101 	 */
9102 	rack_ack_received(tp, rack, th, nsegs, CC_ACK, 0);
9103 
9104 	tp->snd_una = th->th_ack;
9105 	if (tp->snd_wnd < ctf_outstanding(tp)) {
9106 		/* The peer collapsed the window */
9107 		rack_collapsed_window(rack);
9108 	} else if (rack->rc_has_collapsed)
9109 		rack_un_collapse_window(rack);
9110 
9111 	/*
9112 	 * Pull snd_wl2 up to prevent seq wrap relative to th_ack.
9113 	 */
9114 	tp->snd_wl2 = th->th_ack;
9115 	tp->t_dupacks = 0;
9116 	m_freem(m);
9117 	/* ND6_HINT(tp);	 *//* Some progress has been made. */
9118 
9119 	/*
9120 	 * If all outstanding data are acked, stop retransmit timer,
9121 	 * otherwise restart timer using current (possibly backed-off)
9122 	 * value. If process is waiting for space, wakeup/selwakeup/signal.
9123 	 * If data are ready to send, let tcp_output decide between more
9124 	 * output or persist.
9125 	 */
9126 #ifdef TCPDEBUG
9127 	if (so->so_options & SO_DEBUG)
9128 		tcp_trace(TA_INPUT, ostate, tp,
9129 		    (void *)tcp_saveipgen,
9130 		    &tcp_savetcp, 0);
9131 #endif
9132 	if (under_pacing &&
9133 	    (rack->use_fixed_rate == 0) &&
9134 	    (rack->in_probe_rtt == 0) &&
9135 	    rack->rc_gp_dyn_mul &&
9136 	    rack->rc_always_pace) {
9137 		/* Check if we are dragging bottom */
9138 		rack_check_bottom_drag(tp, rack, so, acked);
9139 	}
9140 	if (tp->snd_una == tp->snd_max) {
9141 		rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
9142 		if (rack->r_ctl.rc_went_idle_time == 0)
9143 			rack->r_ctl.rc_went_idle_time = 1;
9144 		rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
9145 		if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0)
9146 			tp->t_acktime = 0;
9147 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
9148 	}
9149 	/* Wake up the socket if we have room to write more */
9150 	tp->t_flags |= TF_WAKESOW;
9151 	if (sbavail(&so->so_snd)) {
9152 		rack->r_wanted_output = 1;
9153 	}
9154 	return (1);
9155 }
9156 
9157 /*
9158  * Return value of 1, the TCB is unlocked and most
9159  * likely gone, return value of 0, the TCP is still
9160  * locked.
9161  */
9162 static int
rack_do_syn_sent(struct mbuf * m,struct tcphdr * th,struct socket * so,struct tcpcb * tp,struct tcpopt * to,int32_t drop_hdrlen,int32_t tlen,uint32_t tiwin,int32_t thflags,int32_t nxt_pkt,uint8_t iptos)9163 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so,
9164     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9165     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9166 {
9167 	int32_t ret_val = 0;
9168 	int32_t todrop;
9169 	int32_t ourfinisacked = 0;
9170 	struct tcp_rack *rack;
9171 
9172 	ctf_calc_rwin(so, tp);
9173 	/*
9174 	 * If the state is SYN_SENT: if seg contains an ACK, but not for our
9175 	 * SYN, drop the input. if seg contains a RST, then drop the
9176 	 * connection. if seg does not contain SYN, then drop it. Otherwise
9177 	 * this is an acceptable SYN segment initialize tp->rcv_nxt and
9178 	 * tp->irs if seg contains ack then advance tp->snd_una if seg
9179 	 * contains an ECE and ECN support is enabled, the stream is ECN
9180 	 * capable. if SYN has been acked change to ESTABLISHED else
9181 	 * SYN_RCVD state arrange for segment to be acked (eventually)
9182 	 * continue processing rest of data/controls.
9183 	 */
9184 	if ((thflags & TH_ACK) &&
9185 	    (SEQ_LEQ(th->th_ack, tp->iss) ||
9186 	    SEQ_GT(th->th_ack, tp->snd_max))) {
9187 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
9188 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9189 		return (1);
9190 	}
9191 	if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) {
9192 		TCP_PROBE5(connect__refused, NULL, tp,
9193 		    mtod(m, const char *), tp, th);
9194 		tp = tcp_drop(tp, ECONNREFUSED);
9195 		ctf_do_drop(m, tp);
9196 		return (1);
9197 	}
9198 	if (thflags & TH_RST) {
9199 		ctf_do_drop(m, tp);
9200 		return (1);
9201 	}
9202 	if (!(thflags & TH_SYN)) {
9203 		ctf_do_drop(m, tp);
9204 		return (1);
9205 	}
9206 	tp->irs = th->th_seq;
9207 	tcp_rcvseqinit(tp);
9208 	rack = (struct tcp_rack *)tp->t_fb_ptr;
9209 	if (thflags & TH_ACK) {
9210 		int tfo_partial = 0;
9211 
9212 		KMOD_TCPSTAT_INC(tcps_connects);
9213 		soisconnected(so);
9214 #ifdef MAC
9215 		mac_socketpeer_set_from_mbuf(m, so);
9216 #endif
9217 		/* Do window scaling on this connection? */
9218 		if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
9219 		    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
9220 			tp->rcv_scale = tp->request_r_scale;
9221 		}
9222 		tp->rcv_adv += min(tp->rcv_wnd,
9223 		    TCP_MAXWIN << tp->rcv_scale);
9224 		/*
9225 		 * If not all the data that was sent in the TFO SYN
9226 		 * has been acked, resend the remainder right away.
9227 		 */
9228 		if (IS_FASTOPEN(tp->t_flags) &&
9229 		    (tp->snd_una != tp->snd_max)) {
9230 			tp->snd_nxt = th->th_ack;
9231 			tfo_partial = 1;
9232 		}
9233 		/*
9234 		 * If there's data, delay ACK; if there's also a FIN ACKNOW
9235 		 * will be turned on later.
9236 		 */
9237 		if (DELAY_ACK(tp, tlen) && tlen != 0 && !tfo_partial) {
9238 			rack_timer_cancel(tp, rack,
9239 					  rack->r_ctl.rc_rcvtime, __LINE__);
9240 			tp->t_flags |= TF_DELACK;
9241 		} else {
9242 			rack->r_wanted_output = 1;
9243 			tp->t_flags |= TF_ACKNOW;
9244 			rack->rc_dack_toggle = 0;
9245 		}
9246 		if (((thflags & (TH_CWR | TH_ECE)) == TH_ECE) &&
9247 		    (V_tcp_do_ecn == 1)) {
9248 			tp->t_flags2 |= TF2_ECN_PERMIT;
9249 			KMOD_TCPSTAT_INC(tcps_ecn_shs);
9250 		}
9251 		if (SEQ_GT(th->th_ack, tp->snd_una)) {
9252 			/*
9253 			 * We advance snd_una for the
9254 			 * fast open case. If th_ack is
9255 			 * acknowledging data beyond
9256 			 * snd_una we can't just call
9257 			 * ack-processing since the
9258 			 * data stream in our send-map
9259 			 * will start at snd_una + 1 (one
9260 			 * beyond the SYN). If its just
9261 			 * equal we don't need to do that
9262 			 * and there is no send_map.
9263 			 */
9264 			tp->snd_una++;
9265 		}
9266 		/*
9267 		 * Received <SYN,ACK> in SYN_SENT[*] state. Transitions:
9268 		 * SYN_SENT  --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1
9269 		 */
9270 		tp->t_starttime = ticks;
9271 		if (tp->t_flags & TF_NEEDFIN) {
9272 			tcp_state_change(tp, TCPS_FIN_WAIT_1);
9273 			tp->t_flags &= ~TF_NEEDFIN;
9274 			thflags &= ~TH_SYN;
9275 		} else {
9276 			tcp_state_change(tp, TCPS_ESTABLISHED);
9277 			TCP_PROBE5(connect__established, NULL, tp,
9278 			    mtod(m, const char *), tp, th);
9279 			rack_cc_conn_init(tp);
9280 		}
9281 	} else {
9282 		/*
9283 		 * Received initial SYN in SYN-SENT[*] state => simultaneous
9284 		 * open.  If segment contains CC option and there is a
9285 		 * cached CC, apply TAO test. If it succeeds, connection is *
9286 		 * half-synchronized. Otherwise, do 3-way handshake:
9287 		 * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If
9288 		 * there was no CC option, clear cached CC value.
9289 		 */
9290 		tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN);
9291 		tcp_state_change(tp, TCPS_SYN_RECEIVED);
9292 	}
9293 	INP_WLOCK_ASSERT(tp->t_inpcb);
9294 	/*
9295 	 * Advance th->th_seq to correspond to first data byte. If data,
9296 	 * trim to stay within window, dropping FIN if necessary.
9297 	 */
9298 	th->th_seq++;
9299 	if (tlen > tp->rcv_wnd) {
9300 		todrop = tlen - tp->rcv_wnd;
9301 		m_adj(m, -todrop);
9302 		tlen = tp->rcv_wnd;
9303 		thflags &= ~TH_FIN;
9304 		KMOD_TCPSTAT_INC(tcps_rcvpackafterwin);
9305 		KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop);
9306 	}
9307 	tp->snd_wl1 = th->th_seq - 1;
9308 	tp->rcv_up = th->th_seq;
9309 	/*
9310 	 * Client side of transaction: already sent SYN and data. If the
9311 	 * remote host used T/TCP to validate the SYN, our data will be
9312 	 * ACK'd; if so, enter normal data segment processing in the middle
9313 	 * of step 5, ack processing. Otherwise, goto step 6.
9314 	 */
9315 	if (thflags & TH_ACK) {
9316 		/* For syn-sent we need to possibly update the rtt */
9317 		if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
9318 			uint32_t t;
9319 
9320 			t = tcp_ts_getticks() - to->to_tsecr;
9321 			if (!tp->t_rttlow || tp->t_rttlow > t)
9322 				tp->t_rttlow = t;
9323 			tcp_rack_xmit_timer(rack, t + 1, 1, (t * HPTS_USEC_IN_MSEC), 0, NULL, 2);
9324 			tcp_rack_xmit_timer_commit(rack, tp);
9325 		}
9326 		if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val))
9327 			return (ret_val);
9328 		/* We may have changed to FIN_WAIT_1 above */
9329 		if (tp->t_state == TCPS_FIN_WAIT_1) {
9330 			/*
9331 			 * In FIN_WAIT_1 STATE in addition to the processing
9332 			 * for the ESTABLISHED state if our FIN is now
9333 			 * acknowledged then enter FIN_WAIT_2.
9334 			 */
9335 			if (ourfinisacked) {
9336 				/*
9337 				 * If we can't receive any more data, then
9338 				 * closing user can proceed. Starting the
9339 				 * timer is contrary to the specification,
9340 				 * but if we don't get a FIN we'll hang
9341 				 * forever.
9342 				 *
9343 				 * XXXjl: we should release the tp also, and
9344 				 * use a compressed state.
9345 				 */
9346 				if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
9347 					soisdisconnected(so);
9348 					tcp_timer_activate(tp, TT_2MSL,
9349 					    (tcp_fast_finwait2_recycle ?
9350 					    tcp_finwait2_timeout :
9351 					    TP_MAXIDLE(tp)));
9352 				}
9353 				tcp_state_change(tp, TCPS_FIN_WAIT_2);
9354 			}
9355 		}
9356 	}
9357 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
9358 	   tiwin, thflags, nxt_pkt));
9359 }
9360 
9361 /*
9362  * Return value of 1, the TCB is unlocked and most
9363  * likely gone, return value of 0, the TCP is still
9364  * locked.
9365  */
9366 static int
rack_do_syn_recv(struct mbuf * m,struct tcphdr * th,struct socket * so,struct tcpcb * tp,struct tcpopt * to,int32_t drop_hdrlen,int32_t tlen,uint32_t tiwin,int32_t thflags,int32_t nxt_pkt,uint8_t iptos)9367 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so,
9368     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9369     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9370 {
9371 	struct tcp_rack *rack;
9372 	int32_t ret_val = 0;
9373 	int32_t ourfinisacked = 0;
9374 
9375 	ctf_calc_rwin(so, tp);
9376 	if ((thflags & TH_ACK) &&
9377 	    (SEQ_LEQ(th->th_ack, tp->snd_una) ||
9378 	    SEQ_GT(th->th_ack, tp->snd_max))) {
9379 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
9380 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9381 		return (1);
9382 	}
9383 	rack = (struct tcp_rack *)tp->t_fb_ptr;
9384 	if (IS_FASTOPEN(tp->t_flags)) {
9385 		/*
9386 		 * When a TFO connection is in SYN_RECEIVED, the
9387 		 * only valid packets are the initial SYN, a
9388 		 * retransmit/copy of the initial SYN (possibly with
9389 		 * a subset of the original data), a valid ACK, a
9390 		 * FIN, or a RST.
9391 		 */
9392 		if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) {
9393 			tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
9394 			ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9395 			return (1);
9396 		} else if (thflags & TH_SYN) {
9397 			/* non-initial SYN is ignored */
9398 			if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT) ||
9399 			    (rack->r_ctl.rc_hpts_flags & PACE_TMR_TLP) ||
9400 			    (rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) {
9401 				ctf_do_drop(m, NULL);
9402 				return (0);
9403 			}
9404 		} else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) {
9405 			ctf_do_drop(m, NULL);
9406 			return (0);
9407 		}
9408 	}
9409 	if ((thflags & TH_RST) ||
9410 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
9411 		return (ctf_process_rst(m, th, so, tp));
9412 	/*
9413 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9414 	 * it's less than ts_recent, drop it.
9415 	 */
9416 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9417 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9418 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9419 			return (ret_val);
9420 	}
9421 	/*
9422 	 * In the SYN-RECEIVED state, validate that the packet belongs to
9423 	 * this connection before trimming the data to fit the receive
9424 	 * window.  Check the sequence number versus IRS since we know the
9425 	 * sequence numbers haven't wrapped.  This is a partial fix for the
9426 	 * "LAND" DoS attack.
9427 	 */
9428 	if (SEQ_LT(th->th_seq, tp->irs)) {
9429 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
9430 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9431 		return (1);
9432 	}
9433 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9434 		return (ret_val);
9435 	}
9436 	/*
9437 	 * If last ACK falls within this segment's sequence numbers, record
9438 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
9439 	 * from the latest proposal of the [email protected] list (Braden
9440 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
9441 	 * with our earlier PAWS tests, so this check should be solely
9442 	 * predicated on the sequence space of this segment. 3) That we
9443 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9444 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9445 	 * SEG.Len, This modified check allows us to overcome RFC1323's
9446 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9447 	 * p.869. In such cases, we can still calculate the RTT correctly
9448 	 * when RCV.NXT == Last.ACK.Sent.
9449 	 */
9450 	if ((to->to_flags & TOF_TS) != 0 &&
9451 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9452 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9453 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9454 		tp->ts_recent_age = tcp_ts_getticks();
9455 		tp->ts_recent = to->to_tsval;
9456 	}
9457 	tp->snd_wnd = tiwin;
9458 	/*
9459 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9460 	 * is on (half-synchronized state), then queue data for later
9461 	 * processing; else drop segment and return.
9462 	 */
9463 	if ((thflags & TH_ACK) == 0) {
9464 		if (IS_FASTOPEN(tp->t_flags)) {
9465 			rack_cc_conn_init(tp);
9466 		}
9467 		return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
9468 		    tiwin, thflags, nxt_pkt));
9469 	}
9470 	KMOD_TCPSTAT_INC(tcps_connects);
9471 	soisconnected(so);
9472 	/* Do window scaling? */
9473 	if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
9474 	    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
9475 		tp->rcv_scale = tp->request_r_scale;
9476 	}
9477 	/*
9478 	 * Make transitions: SYN-RECEIVED  -> ESTABLISHED SYN-RECEIVED* ->
9479 	 * FIN-WAIT-1
9480 	 */
9481 	tp->t_starttime = ticks;
9482 	if (IS_FASTOPEN(tp->t_flags) && tp->t_tfo_pending) {
9483 		tcp_fastopen_decrement_counter(tp->t_tfo_pending);
9484 		tp->t_tfo_pending = NULL;
9485 	}
9486 	if (tp->t_flags & TF_NEEDFIN) {
9487 		tcp_state_change(tp, TCPS_FIN_WAIT_1);
9488 		tp->t_flags &= ~TF_NEEDFIN;
9489 	} else {
9490 		tcp_state_change(tp, TCPS_ESTABLISHED);
9491 		TCP_PROBE5(accept__established, NULL, tp,
9492 		    mtod(m, const char *), tp, th);
9493 		/*
9494 		 * TFO connections call cc_conn_init() during SYN
9495 		 * processing.  Calling it again here for such connections
9496 		 * is not harmless as it would undo the snd_cwnd reduction
9497 		 * that occurs when a TFO SYN|ACK is retransmitted.
9498 		 */
9499 		if (!IS_FASTOPEN(tp->t_flags))
9500 			rack_cc_conn_init(tp);
9501 	}
9502 	/*
9503 	 * Account for the ACK of our SYN prior to
9504 	 * regular ACK processing below, except for
9505 	 * simultaneous SYN, which is handled later.
9506 	 */
9507 	if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN))
9508 		tp->snd_una++;
9509 	/*
9510 	 * If segment contains data or ACK, will call tcp_reass() later; if
9511 	 * not, do so now to pass queued data to user.
9512 	 */
9513 	if (tlen == 0 && (thflags & TH_FIN) == 0)
9514 		(void) tcp_reass(tp, (struct tcphdr *)0, NULL, 0,
9515 		    (struct mbuf *)0);
9516 	tp->snd_wl1 = th->th_seq - 1;
9517 	/* For syn-recv we need to possibly update the rtt */
9518 	if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
9519 		uint32_t t;
9520 
9521 		t = tcp_ts_getticks() - to->to_tsecr;
9522 		if (!tp->t_rttlow || tp->t_rttlow > t)
9523 			tp->t_rttlow = t;
9524 		tcp_rack_xmit_timer(rack, t + 1, 1, (t * HPTS_USEC_IN_MSEC), 0, NULL, 2);
9525 		tcp_rack_xmit_timer_commit(rack, tp);
9526 	}
9527 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
9528 		return (ret_val);
9529 	}
9530 	if (tp->t_state == TCPS_FIN_WAIT_1) {
9531 		/* We could have went to FIN_WAIT_1 (or EST) above */
9532 		/*
9533 		 * In FIN_WAIT_1 STATE in addition to the processing for the
9534 		 * ESTABLISHED state if our FIN is now acknowledged then
9535 		 * enter FIN_WAIT_2.
9536 		 */
9537 		if (ourfinisacked) {
9538 			/*
9539 			 * If we can't receive any more data, then closing
9540 			 * user can proceed. Starting the timer is contrary
9541 			 * to the specification, but if we don't get a FIN
9542 			 * we'll hang forever.
9543 			 *
9544 			 * XXXjl: we should release the tp also, and use a
9545 			 * compressed state.
9546 			 */
9547 			if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
9548 				soisdisconnected(so);
9549 				tcp_timer_activate(tp, TT_2MSL,
9550 				    (tcp_fast_finwait2_recycle ?
9551 				    tcp_finwait2_timeout :
9552 				    TP_MAXIDLE(tp)));
9553 			}
9554 			tcp_state_change(tp, TCPS_FIN_WAIT_2);
9555 		}
9556 	}
9557 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
9558 	    tiwin, thflags, nxt_pkt));
9559 }
9560 
9561 /*
9562  * Return value of 1, the TCB is unlocked and most
9563  * likely gone, return value of 0, the TCP is still
9564  * locked.
9565  */
9566 static int
rack_do_established(struct mbuf * m,struct tcphdr * th,struct socket * so,struct tcpcb * tp,struct tcpopt * to,int32_t drop_hdrlen,int32_t tlen,uint32_t tiwin,int32_t thflags,int32_t nxt_pkt,uint8_t iptos)9567 rack_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so,
9568     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9569     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9570 {
9571 	int32_t ret_val = 0;
9572 	struct tcp_rack *rack;
9573 
9574 	/*
9575 	 * Header prediction: check for the two common cases of a
9576 	 * uni-directional data xfer.  If the packet has no control flags,
9577 	 * is in-sequence, the window didn't change and we're not
9578 	 * retransmitting, it's a candidate.  If the length is zero and the
9579 	 * ack moved forward, we're the sender side of the xfer.  Just free
9580 	 * the data acked & wake any higher level process that was blocked
9581 	 * waiting for space.  If the length is non-zero and the ack didn't
9582 	 * move, we're the receiver side.  If we're getting packets in-order
9583 	 * (the reassembly queue is empty), add the data toc The socket
9584 	 * buffer and note that we need a delayed ack. Make sure that the
9585 	 * hidden state-flags are also off. Since we check for
9586 	 * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN.
9587 	 */
9588 	rack = (struct tcp_rack *)tp->t_fb_ptr;
9589 	if (__predict_true(((to->to_flags & TOF_SACK) == 0)) &&
9590 	    __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_ACK)) == TH_ACK) &&
9591 	    __predict_true(SEGQ_EMPTY(tp)) &&
9592 	    __predict_true(th->th_seq == tp->rcv_nxt)) {
9593 		if (tlen == 0) {
9594 			if (rack_fastack(m, th, so, tp, to, drop_hdrlen, tlen,
9595 			    tiwin, nxt_pkt, rack->r_ctl.rc_rcvtime)) {
9596 				return (0);
9597 			}
9598 		} else {
9599 			if (rack_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen,
9600 			    tiwin, nxt_pkt, iptos)) {
9601 				return (0);
9602 			}
9603 		}
9604 	}
9605 	ctf_calc_rwin(so, tp);
9606 
9607 	if ((thflags & TH_RST) ||
9608 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
9609 		return (ctf_process_rst(m, th, so, tp));
9610 
9611 	/*
9612 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9613 	 * synchronized state.
9614 	 */
9615 	if (thflags & TH_SYN) {
9616 		ctf_challenge_ack(m, th, tp, &ret_val);
9617 		return (ret_val);
9618 	}
9619 	/*
9620 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9621 	 * it's less than ts_recent, drop it.
9622 	 */
9623 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9624 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9625 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9626 			return (ret_val);
9627 	}
9628 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9629 		return (ret_val);
9630 	}
9631 	/*
9632 	 * If last ACK falls within this segment's sequence numbers, record
9633 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
9634 	 * from the latest proposal of the [email protected] list (Braden
9635 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
9636 	 * with our earlier PAWS tests, so this check should be solely
9637 	 * predicated on the sequence space of this segment. 3) That we
9638 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9639 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9640 	 * SEG.Len, This modified check allows us to overcome RFC1323's
9641 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9642 	 * p.869. In such cases, we can still calculate the RTT correctly
9643 	 * when RCV.NXT == Last.ACK.Sent.
9644 	 */
9645 	if ((to->to_flags & TOF_TS) != 0 &&
9646 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9647 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9648 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9649 		tp->ts_recent_age = tcp_ts_getticks();
9650 		tp->ts_recent = to->to_tsval;
9651 	}
9652 	/*
9653 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9654 	 * is on (half-synchronized state), then queue data for later
9655 	 * processing; else drop segment and return.
9656 	 */
9657 	if ((thflags & TH_ACK) == 0) {
9658 		if (tp->t_flags & TF_NEEDSYN) {
9659 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
9660 			    tiwin, thflags, nxt_pkt));
9661 
9662 		} else if (tp->t_flags & TF_ACKNOW) {
9663 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
9664 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output= 1;
9665 			return (ret_val);
9666 		} else {
9667 			ctf_do_drop(m, NULL);
9668 			return (0);
9669 		}
9670 	}
9671 	/*
9672 	 * Ack processing.
9673 	 */
9674 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
9675 		return (ret_val);
9676 	}
9677 	if (sbavail(&so->so_snd)) {
9678 		if (ctf_progress_timeout_check(tp, true)) {
9679 			rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
9680 			tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
9681 			ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9682 			return (1);
9683 		}
9684 	}
9685 	/* State changes only happen in rack_process_data() */
9686 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
9687 	    tiwin, thflags, nxt_pkt));
9688 }
9689 
9690 /*
9691  * Return value of 1, the TCB is unlocked and most
9692  * likely gone, return value of 0, the TCP is still
9693  * locked.
9694  */
9695 static int
rack_do_close_wait(struct mbuf * m,struct tcphdr * th,struct socket * so,struct tcpcb * tp,struct tcpopt * to,int32_t drop_hdrlen,int32_t tlen,uint32_t tiwin,int32_t thflags,int32_t nxt_pkt,uint8_t iptos)9696 rack_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so,
9697     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9698     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9699 {
9700 	int32_t ret_val = 0;
9701 
9702 	ctf_calc_rwin(so, tp);
9703 	if ((thflags & TH_RST) ||
9704 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
9705 		return (ctf_process_rst(m, th, so, tp));
9706 	/*
9707 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9708 	 * synchronized state.
9709 	 */
9710 	if (thflags & TH_SYN) {
9711 		ctf_challenge_ack(m, th, tp, &ret_val);
9712 		return (ret_val);
9713 	}
9714 	/*
9715 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9716 	 * it's less than ts_recent, drop it.
9717 	 */
9718 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9719 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9720 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9721 			return (ret_val);
9722 	}
9723 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9724 		return (ret_val);
9725 	}
9726 	/*
9727 	 * If last ACK falls within this segment's sequence numbers, record
9728 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
9729 	 * from the latest proposal of the [email protected] list (Braden
9730 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
9731 	 * with our earlier PAWS tests, so this check should be solely
9732 	 * predicated on the sequence space of this segment. 3) That we
9733 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9734 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9735 	 * SEG.Len, This modified check allows us to overcome RFC1323's
9736 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9737 	 * p.869. In such cases, we can still calculate the RTT correctly
9738 	 * when RCV.NXT == Last.ACK.Sent.
9739 	 */
9740 	if ((to->to_flags & TOF_TS) != 0 &&
9741 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9742 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9743 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9744 		tp->ts_recent_age = tcp_ts_getticks();
9745 		tp->ts_recent = to->to_tsval;
9746 	}
9747 	/*
9748 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9749 	 * is on (half-synchronized state), then queue data for later
9750 	 * processing; else drop segment and return.
9751 	 */
9752 	if ((thflags & TH_ACK) == 0) {
9753 		if (tp->t_flags & TF_NEEDSYN) {
9754 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
9755 			    tiwin, thflags, nxt_pkt));
9756 
9757 		} else if (tp->t_flags & TF_ACKNOW) {
9758 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
9759 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
9760 			return (ret_val);
9761 		} else {
9762 			ctf_do_drop(m, NULL);
9763 			return (0);
9764 		}
9765 	}
9766 	/*
9767 	 * Ack processing.
9768 	 */
9769 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
9770 		return (ret_val);
9771 	}
9772 	if (sbavail(&so->so_snd)) {
9773 		if (ctf_progress_timeout_check(tp, true)) {
9774 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
9775 						tp, tick, PROGRESS_DROP, __LINE__);
9776 			tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
9777 			ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9778 			return (1);
9779 		}
9780 	}
9781 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
9782 	    tiwin, thflags, nxt_pkt));
9783 }
9784 
9785 static int
rack_check_data_after_close(struct mbuf * m,struct tcpcb * tp,int32_t * tlen,struct tcphdr * th,struct socket * so)9786 rack_check_data_after_close(struct mbuf *m,
9787     struct tcpcb *tp, int32_t *tlen, struct tcphdr *th, struct socket *so)
9788 {
9789 	struct tcp_rack *rack;
9790 
9791 	rack = (struct tcp_rack *)tp->t_fb_ptr;
9792 	if (rack->rc_allow_data_af_clo == 0) {
9793 	close_now:
9794 		tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
9795 		/* tcp_close will kill the inp pre-log the Reset */
9796 		tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
9797 		tp = tcp_close(tp);
9798 		KMOD_TCPSTAT_INC(tcps_rcvafterclose);
9799 		ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen));
9800 		return (1);
9801 	}
9802 	if (sbavail(&so->so_snd) == 0)
9803 		goto close_now;
9804 	/* Ok we allow data that is ignored and a followup reset */
9805 	tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
9806 	tp->rcv_nxt = th->th_seq + *tlen;
9807 	tp->t_flags2 |= TF2_DROP_AF_DATA;
9808 	rack->r_wanted_output = 1;
9809 	*tlen = 0;
9810 	return (0);
9811 }
9812 
9813 /*
9814  * Return value of 1, the TCB is unlocked and most
9815  * likely gone, return value of 0, the TCP is still
9816  * locked.
9817  */
9818 static int
rack_do_fin_wait_1(struct mbuf * m,struct tcphdr * th,struct socket * so,struct tcpcb * tp,struct tcpopt * to,int32_t drop_hdrlen,int32_t tlen,uint32_t tiwin,int32_t thflags,int32_t nxt_pkt,uint8_t iptos)9819 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so,
9820     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9821     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9822 {
9823 	int32_t ret_val = 0;
9824 	int32_t ourfinisacked = 0;
9825 
9826 	ctf_calc_rwin(so, tp);
9827 
9828 	if ((thflags & TH_RST) ||
9829 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
9830 		return (ctf_process_rst(m, th, so, tp));
9831 	/*
9832 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9833 	 * synchronized state.
9834 	 */
9835 	if (thflags & TH_SYN) {
9836 		ctf_challenge_ack(m, th, tp, &ret_val);
9837 		return (ret_val);
9838 	}
9839 	/*
9840 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9841 	 * it's less than ts_recent, drop it.
9842 	 */
9843 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9844 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9845 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9846 			return (ret_val);
9847 	}
9848 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9849 		return (ret_val);
9850 	}
9851 	/*
9852 	 * If new data are received on a connection after the user processes
9853 	 * are gone, then RST the other end.
9854 	 */
9855 	if ((so->so_state & SS_NOFDREF) && tlen) {
9856 		if (rack_check_data_after_close(m, tp, &tlen, th, so))
9857 			return (1);
9858 	}
9859 	/*
9860 	 * If last ACK falls within this segment's sequence numbers, record
9861 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
9862 	 * from the latest proposal of the [email protected] list (Braden
9863 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
9864 	 * with our earlier PAWS tests, so this check should be solely
9865 	 * predicated on the sequence space of this segment. 3) That we
9866 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9867 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9868 	 * SEG.Len, This modified check allows us to overcome RFC1323's
9869 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9870 	 * p.869. In such cases, we can still calculate the RTT correctly
9871 	 * when RCV.NXT == Last.ACK.Sent.
9872 	 */
9873 	if ((to->to_flags & TOF_TS) != 0 &&
9874 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9875 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9876 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9877 		tp->ts_recent_age = tcp_ts_getticks();
9878 		tp->ts_recent = to->to_tsval;
9879 	}
9880 	/*
9881 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9882 	 * is on (half-synchronized state), then queue data for later
9883 	 * processing; else drop segment and return.
9884 	 */
9885 	if ((thflags & TH_ACK) == 0) {
9886 		if (tp->t_flags & TF_NEEDSYN) {
9887 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
9888 			    tiwin, thflags, nxt_pkt));
9889 		} else if (tp->t_flags & TF_ACKNOW) {
9890 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
9891 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
9892 			return (ret_val);
9893 		} else {
9894 			ctf_do_drop(m, NULL);
9895 			return (0);
9896 		}
9897 	}
9898 	/*
9899 	 * Ack processing.
9900 	 */
9901 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
9902 		return (ret_val);
9903 	}
9904 	if (ourfinisacked) {
9905 		/*
9906 		 * If we can't receive any more data, then closing user can
9907 		 * proceed. Starting the timer is contrary to the
9908 		 * specification, but if we don't get a FIN we'll hang
9909 		 * forever.
9910 		 *
9911 		 * XXXjl: we should release the tp also, and use a
9912 		 * compressed state.
9913 		 */
9914 		if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
9915 			soisdisconnected(so);
9916 			tcp_timer_activate(tp, TT_2MSL,
9917 			    (tcp_fast_finwait2_recycle ?
9918 			    tcp_finwait2_timeout :
9919 			    TP_MAXIDLE(tp)));
9920 		}
9921 		tcp_state_change(tp, TCPS_FIN_WAIT_2);
9922 	}
9923 	if (sbavail(&so->so_snd)) {
9924 		if (ctf_progress_timeout_check(tp, true)) {
9925 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
9926 						tp, tick, PROGRESS_DROP, __LINE__);
9927 			tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
9928 			ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9929 			return (1);
9930 		}
9931 	}
9932 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
9933 	    tiwin, thflags, nxt_pkt));
9934 }
9935 
9936 /*
9937  * Return value of 1, the TCB is unlocked and most
9938  * likely gone, return value of 0, the TCP is still
9939  * locked.
9940  */
9941 static int
rack_do_closing(struct mbuf * m,struct tcphdr * th,struct socket * so,struct tcpcb * tp,struct tcpopt * to,int32_t drop_hdrlen,int32_t tlen,uint32_t tiwin,int32_t thflags,int32_t nxt_pkt,uint8_t iptos)9942 rack_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so,
9943     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9944     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9945 {
9946 	int32_t ret_val = 0;
9947 	int32_t ourfinisacked = 0;
9948 
9949 	ctf_calc_rwin(so, tp);
9950 
9951 	if ((thflags & TH_RST) ||
9952 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
9953 		return (ctf_process_rst(m, th, so, tp));
9954 	/*
9955 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9956 	 * synchronized state.
9957 	 */
9958 	if (thflags & TH_SYN) {
9959 		ctf_challenge_ack(m, th, tp, &ret_val);
9960 		return (ret_val);
9961 	}
9962 	/*
9963 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9964 	 * it's less than ts_recent, drop it.
9965 	 */
9966 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9967 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9968 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9969 			return (ret_val);
9970 	}
9971 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9972 		return (ret_val);
9973 	}
9974 	/*
9975 	 * If new data are received on a connection after the user processes
9976 	 * are gone, then RST the other end.
9977 	 */
9978 	if ((so->so_state & SS_NOFDREF) && tlen) {
9979 		if (rack_check_data_after_close(m, tp, &tlen, th, so))
9980 			return (1);
9981 	}
9982 	/*
9983 	 * If last ACK falls within this segment's sequence numbers, record
9984 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
9985 	 * from the latest proposal of the [email protected] list (Braden
9986 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
9987 	 * with our earlier PAWS tests, so this check should be solely
9988 	 * predicated on the sequence space of this segment. 3) That we
9989 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9990 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9991 	 * SEG.Len, This modified check allows us to overcome RFC1323's
9992 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9993 	 * p.869. In such cases, we can still calculate the RTT correctly
9994 	 * when RCV.NXT == Last.ACK.Sent.
9995 	 */
9996 	if ((to->to_flags & TOF_TS) != 0 &&
9997 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9998 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9999 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
10000 		tp->ts_recent_age = tcp_ts_getticks();
10001 		tp->ts_recent = to->to_tsval;
10002 	}
10003 	/*
10004 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
10005 	 * is on (half-synchronized state), then queue data for later
10006 	 * processing; else drop segment and return.
10007 	 */
10008 	if ((thflags & TH_ACK) == 0) {
10009 		if (tp->t_flags & TF_NEEDSYN) {
10010 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
10011 			    tiwin, thflags, nxt_pkt));
10012 		} else if (tp->t_flags & TF_ACKNOW) {
10013 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
10014 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output= 1;
10015 			return (ret_val);
10016 		} else {
10017 			ctf_do_drop(m, NULL);
10018 			return (0);
10019 		}
10020 	}
10021 	/*
10022 	 * Ack processing.
10023 	 */
10024 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
10025 		return (ret_val);
10026 	}
10027 	if (ourfinisacked) {
10028 		tcp_twstart(tp);
10029 		m_freem(m);
10030 		return (1);
10031 	}
10032 	if (sbavail(&so->so_snd)) {
10033 		if (ctf_progress_timeout_check(tp, true)) {
10034 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
10035 						tp, tick, PROGRESS_DROP, __LINE__);
10036 			tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
10037 			ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
10038 			return (1);
10039 		}
10040 	}
10041 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
10042 	    tiwin, thflags, nxt_pkt));
10043 }
10044 
10045 /*
10046  * Return value of 1, the TCB is unlocked and most
10047  * likely gone, return value of 0, the TCP is still
10048  * locked.
10049  */
10050 static int
rack_do_lastack(struct mbuf * m,struct tcphdr * th,struct socket * so,struct tcpcb * tp,struct tcpopt * to,int32_t drop_hdrlen,int32_t tlen,uint32_t tiwin,int32_t thflags,int32_t nxt_pkt,uint8_t iptos)10051 rack_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
10052     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
10053     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
10054 {
10055 	int32_t ret_val = 0;
10056 	int32_t ourfinisacked = 0;
10057 
10058 	ctf_calc_rwin(so, tp);
10059 
10060 	if ((thflags & TH_RST) ||
10061 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
10062 		return (ctf_process_rst(m, th, so, tp));
10063 	/*
10064 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
10065 	 * synchronized state.
10066 	 */
10067 	if (thflags & TH_SYN) {
10068 		ctf_challenge_ack(m, th, tp, &ret_val);
10069 		return (ret_val);
10070 	}
10071 	/*
10072 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
10073 	 * it's less than ts_recent, drop it.
10074 	 */
10075 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
10076 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
10077 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
10078 			return (ret_val);
10079 	}
10080 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
10081 		return (ret_val);
10082 	}
10083 	/*
10084 	 * If new data are received on a connection after the user processes
10085 	 * are gone, then RST the other end.
10086 	 */
10087 	if ((so->so_state & SS_NOFDREF) && tlen) {
10088 		if (rack_check_data_after_close(m, tp, &tlen, th, so))
10089 			return (1);
10090 	}
10091 	/*
10092 	 * If last ACK falls within this segment's sequence numbers, record
10093 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
10094 	 * from the latest proposal of the [email protected] list (Braden
10095 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
10096 	 * with our earlier PAWS tests, so this check should be solely
10097 	 * predicated on the sequence space of this segment. 3) That we
10098 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
10099 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
10100 	 * SEG.Len, This modified check allows us to overcome RFC1323's
10101 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
10102 	 * p.869. In such cases, we can still calculate the RTT correctly
10103 	 * when RCV.NXT == Last.ACK.Sent.
10104 	 */
10105 	if ((to->to_flags & TOF_TS) != 0 &&
10106 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
10107 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
10108 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
10109 		tp->ts_recent_age = tcp_ts_getticks();
10110 		tp->ts_recent = to->to_tsval;
10111 	}
10112 	/*
10113 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
10114 	 * is on (half-synchronized state), then queue data for later
10115 	 * processing; else drop segment and return.
10116 	 */
10117 	if ((thflags & TH_ACK) == 0) {
10118 		if (tp->t_flags & TF_NEEDSYN) {
10119 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
10120 			    tiwin, thflags, nxt_pkt));
10121 		} else if (tp->t_flags & TF_ACKNOW) {
10122 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
10123 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
10124 			return (ret_val);
10125 		} else {
10126 			ctf_do_drop(m, NULL);
10127 			return (0);
10128 		}
10129 	}
10130 	/*
10131 	 * case TCPS_LAST_ACK: Ack processing.
10132 	 */
10133 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
10134 		return (ret_val);
10135 	}
10136 	if (ourfinisacked) {
10137 		tp = tcp_close(tp);
10138 		ctf_do_drop(m, tp);
10139 		return (1);
10140 	}
10141 	if (sbavail(&so->so_snd)) {
10142 		if (ctf_progress_timeout_check(tp, true)) {
10143 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
10144 						tp, tick, PROGRESS_DROP, __LINE__);
10145 			tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
10146 			ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
10147 			return (1);
10148 		}
10149 	}
10150 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
10151 	    tiwin, thflags, nxt_pkt));
10152 }
10153 
10154 /*
10155  * Return value of 1, the TCB is unlocked and most
10156  * likely gone, return value of 0, the TCP is still
10157  * locked.
10158  */
10159 static int
rack_do_fin_wait_2(struct mbuf * m,struct tcphdr * th,struct socket * so,struct tcpcb * tp,struct tcpopt * to,int32_t drop_hdrlen,int32_t tlen,uint32_t tiwin,int32_t thflags,int32_t nxt_pkt,uint8_t iptos)10160 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so,
10161     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
10162     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
10163 {
10164 	int32_t ret_val = 0;
10165 	int32_t ourfinisacked = 0;
10166 
10167 	ctf_calc_rwin(so, tp);
10168 
10169 	/* Reset receive buffer auto scaling when not in bulk receive mode. */
10170 	if ((thflags & TH_RST) ||
10171 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
10172 		return (ctf_process_rst(m, th, so, tp));
10173 	/*
10174 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
10175 	 * synchronized state.
10176 	 */
10177 	if (thflags & TH_SYN) {
10178 		ctf_challenge_ack(m, th, tp, &ret_val);
10179 		return (ret_val);
10180 	}
10181 	/*
10182 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
10183 	 * it's less than ts_recent, drop it.
10184 	 */
10185 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
10186 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
10187 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
10188 			return (ret_val);
10189 	}
10190 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
10191 		return (ret_val);
10192 	}
10193 	/*
10194 	 * If new data are received on a connection after the user processes
10195 	 * are gone, then RST the other end.
10196 	 */
10197 	if ((so->so_state & SS_NOFDREF) &&
10198 	    tlen) {
10199 		if (rack_check_data_after_close(m, tp, &tlen, th, so))
10200 			return (1);
10201 	}
10202 	/*
10203 	 * If last ACK falls within this segment's sequence numbers, record
10204 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
10205 	 * from the latest proposal of the [email protected] list (Braden
10206 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
10207 	 * with our earlier PAWS tests, so this check should be solely
10208 	 * predicated on the sequence space of this segment. 3) That we
10209 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
10210 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
10211 	 * SEG.Len, This modified check allows us to overcome RFC1323's
10212 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
10213 	 * p.869. In such cases, we can still calculate the RTT correctly
10214 	 * when RCV.NXT == Last.ACK.Sent.
10215 	 */
10216 	if ((to->to_flags & TOF_TS) != 0 &&
10217 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
10218 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
10219 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
10220 		tp->ts_recent_age = tcp_ts_getticks();
10221 		tp->ts_recent = to->to_tsval;
10222 	}
10223 	/*
10224 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
10225 	 * is on (half-synchronized state), then queue data for later
10226 	 * processing; else drop segment and return.
10227 	 */
10228 	if ((thflags & TH_ACK) == 0) {
10229 		if (tp->t_flags & TF_NEEDSYN) {
10230 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
10231 			    tiwin, thflags, nxt_pkt));
10232 		} else if (tp->t_flags & TF_ACKNOW) {
10233 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
10234 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
10235 			return (ret_val);
10236 		} else {
10237 			ctf_do_drop(m, NULL);
10238 			return (0);
10239 		}
10240 	}
10241 	/*
10242 	 * Ack processing.
10243 	 */
10244 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
10245 		return (ret_val);
10246 	}
10247 	if (sbavail(&so->so_snd)) {
10248 		if (ctf_progress_timeout_check(tp, true)) {
10249 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
10250 						tp, tick, PROGRESS_DROP, __LINE__);
10251 			tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
10252 			ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
10253 			return (1);
10254 		}
10255 	}
10256 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
10257 	    tiwin, thflags, nxt_pkt));
10258 }
10259 
10260 static void inline
rack_clear_rate_sample(struct tcp_rack * rack)10261 rack_clear_rate_sample(struct tcp_rack *rack)
10262 {
10263 	rack->r_ctl.rack_rs.rs_flags = RACK_RTT_EMPTY;
10264 	rack->r_ctl.rack_rs.rs_rtt_cnt = 0;
10265 	rack->r_ctl.rack_rs.rs_rtt_tot = 0;
10266 }
10267 
10268 static void
rack_set_pace_segments(struct tcpcb * tp,struct tcp_rack * rack,uint32_t line)10269 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line)
10270 {
10271 	uint64_t bw_est, rate_wanted;
10272 	int chged = 0;
10273 	uint32_t user_max;
10274 
10275 	user_max = ctf_fixed_maxseg(tp) * rack->rc_user_set_max_segs;
10276 	if (ctf_fixed_maxseg(tp) != rack->r_ctl.rc_pace_min_segs)
10277 		chged = 1;
10278 	rack->r_ctl.rc_pace_min_segs = ctf_fixed_maxseg(tp);
10279 	if (rack->use_fixed_rate || rack->rc_force_max_seg) {
10280 		if (user_max != rack->r_ctl.rc_pace_max_segs)
10281 			chged = 1;
10282 	}
10283 	if (rack->rc_force_max_seg) {
10284 		rack->r_ctl.rc_pace_max_segs = user_max;
10285 	} else if (rack->use_fixed_rate) {
10286 		bw_est = rack_get_bw(rack);
10287 		if ((rack->r_ctl.crte == NULL) ||
10288 		    (bw_est != rack->r_ctl.crte->rate))  {
10289 			rack->r_ctl.rc_pace_max_segs = user_max;
10290 		} else {
10291 			/* We are pacing right at the hardware rate */
10292 			uint32_t segsiz;
10293 
10294 			segsiz = min(ctf_fixed_maxseg(tp),
10295 				     rack->r_ctl.rc_pace_min_segs);
10296 			rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(
10297 				                           bw_est, segsiz, 0,
10298 							   rack->r_ctl.crte, NULL);
10299 		}
10300 	} else if (rack->rc_always_pace) {
10301 		if (rack->r_ctl.gp_bw ||
10302 #ifdef NETFLIX_PEAKRATE
10303 		    rack->rc_tp->t_maxpeakrate ||
10304 #endif
10305 		    rack->r_ctl.init_rate) {
10306 			/* We have a rate of some sort set */
10307 			uint32_t  orig;
10308 
10309 			bw_est = rack_get_bw(rack);
10310 			orig = rack->r_ctl.rc_pace_max_segs;
10311 			rate_wanted = rack_get_output_bw(rack, bw_est, NULL);
10312 			if (rate_wanted) {
10313 				/* We have something */
10314 				rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack,
10315 										   rate_wanted,
10316 										   ctf_fixed_maxseg(rack->rc_tp));
10317 			} else
10318 				rack->r_ctl.rc_pace_max_segs = rack->r_ctl.rc_pace_min_segs;
10319 			if (orig != rack->r_ctl.rc_pace_max_segs)
10320 				chged = 1;
10321 		} else if ((rack->r_ctl.gp_bw == 0) &&
10322 			   (rack->r_ctl.rc_pace_max_segs == 0)) {
10323 			/*
10324 			 * If we have nothing limit us to bursting
10325 			 * out IW sized pieces.
10326 			 */
10327 			chged = 1;
10328 			rack->r_ctl.rc_pace_max_segs = rc_init_window(rack);
10329 		}
10330 	}
10331 	if (rack->r_ctl.rc_pace_max_segs > PACE_MAX_IP_BYTES) {
10332 		chged = 1;
10333 		rack->r_ctl.rc_pace_max_segs = PACE_MAX_IP_BYTES;
10334 	}
10335 	if (chged)
10336 		rack_log_type_hrdwtso(tp, rack, 0, rack->rc_inp->inp_socket->so_snd.sb_flags, line, 2);
10337 }
10338 
10339 static int
rack_init(struct tcpcb * tp)10340 rack_init(struct tcpcb *tp)
10341 {
10342 	struct tcp_rack *rack = NULL;
10343 	struct rack_sendmap *insret;
10344 	uint32_t iwin, snt, us_cts;
10345 
10346 	tp->t_fb_ptr = uma_zalloc(rack_pcb_zone, M_NOWAIT);
10347 	if (tp->t_fb_ptr == NULL) {
10348 		/*
10349 		 * We need to allocate memory but cant. The INP and INP_INFO
10350 		 * locks and they are recusive (happens during setup. So a
10351 		 * scheme to drop the locks fails :(
10352 		 *
10353 		 */
10354 		return (ENOMEM);
10355 	}
10356 	memset(tp->t_fb_ptr, 0, sizeof(struct tcp_rack));
10357 
10358 	rack = (struct tcp_rack *)tp->t_fb_ptr;
10359 	RB_INIT(&rack->r_ctl.rc_mtree);
10360 	TAILQ_INIT(&rack->r_ctl.rc_free);
10361 	TAILQ_INIT(&rack->r_ctl.rc_tmap);
10362 	rack->rc_tp = tp;
10363 	if (tp->t_inpcb) {
10364 		rack->rc_inp = tp->t_inpcb;
10365 	}
10366 	/* Probably not needed but lets be sure */
10367 	rack_clear_rate_sample(rack);
10368 	rack->r_ctl.rc_reorder_fade = rack_reorder_fade;
10369 	rack->rc_allow_data_af_clo = rack_ignore_data_after_close;
10370 	rack->r_ctl.rc_tlp_threshold = rack_tlp_thresh;
10371 	if (use_rack_rr)
10372 		rack->use_rack_rr = 1;
10373 	if (V_tcp_delack_enabled)
10374 		tp->t_delayed_ack = 1;
10375 	else
10376 		tp->t_delayed_ack = 0;
10377 	if (rack_enable_shared_cwnd)
10378 		rack->rack_enable_scwnd = 1;
10379 	rack->rc_user_set_max_segs = rack_hptsi_segments;
10380 	rack->rc_force_max_seg = 0;
10381 	if (rack_use_imac_dack)
10382 		rack->rc_dack_mode = 1;
10383 	rack->r_ctl.rc_reorder_shift = rack_reorder_thresh;
10384 	rack->r_ctl.rc_pkt_delay = rack_pkt_delay;
10385 	rack->r_ctl.rc_prop_reduce = rack_use_proportional_reduce;
10386 	rack->r_ctl.rc_prop_rate = rack_proportional_rate;
10387 	rack->r_ctl.rc_tlp_cwnd_reduce = rack_lower_cwnd_at_tlp;
10388 	rack->r_ctl.rc_early_recovery = rack_early_recovery;
10389 	rack->r_ctl.rc_lowest_us_rtt = 0xffffffff;
10390 	rack->r_ctl.rc_highest_us_rtt = 0;
10391 	if (rack_disable_prr)
10392 		rack->rack_no_prr = 1;
10393 	if (rack_gp_no_rec_chg)
10394 		rack->rc_gp_no_rec_chg = 1;
10395 	rack->rc_always_pace = rack_pace_every_seg;
10396 	if (rack_enable_mqueue_for_nonpaced)
10397 		rack->r_mbuf_queue = 1;
10398 	else
10399 		rack->r_mbuf_queue = 0;
10400 	if  (rack->r_mbuf_queue || rack->rc_always_pace)
10401 		tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
10402 	else
10403 		tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
10404 	rack_set_pace_segments(tp, rack, __LINE__);
10405 	if (rack_limits_scwnd)
10406 		rack->r_limit_scw  = 1;
10407 	else
10408 		rack->r_limit_scw  = 0;
10409 	rack->r_ctl.rc_high_rwnd = tp->snd_wnd;
10410 	rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
10411 	rack->r_ctl.rc_rate_sample_method = rack_rate_sample_method;
10412 	rack->rack_tlp_threshold_use = rack_tlp_threshold_use;
10413 	rack->r_ctl.rc_prr_sendalot = rack_send_a_lot_in_prr;
10414 	rack->r_ctl.rc_min_to = rack_min_to;
10415 	microuptime(&rack->r_ctl.act_rcv_time);
10416 	rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
10417 	rack->r_running_late = 0;
10418 	rack->r_running_early = 0;
10419 	rack->rc_init_win = rack_default_init_window;
10420 	rack->r_ctl.rack_per_of_gp_ss = rack_per_of_gp_ss;
10421 	if (rack_do_dyn_mul) {
10422 		/* When dynamic adjustment is on CA needs to start at 100% */
10423 		rack->rc_gp_dyn_mul = 1;
10424 		if (rack_do_dyn_mul >= 100)
10425 			rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
10426 	} else
10427 		rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
10428 	rack->r_ctl.rack_per_of_gp_rec = rack_per_of_gp_rec;
10429 	rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
10430 	rack->r_ctl.rc_tlp_rxt_last_time = tcp_tv_to_mssectick(&rack->r_ctl.act_rcv_time);
10431 	setup_time_filter_small(&rack->r_ctl.rc_gp_min_rtt, FILTER_TYPE_MIN,
10432 				rack_probertt_filter_life);
10433 	us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
10434 	rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
10435 	rack->r_ctl.rc_time_of_last_probertt = us_cts;
10436 	rack->r_ctl.rc_time_probertt_starts = 0;
10437 	/* Do we force on detection? */
10438 #ifdef NETFLIX_EXP_DETECTION
10439 	if (tcp_force_detection)
10440 		rack->do_detection = 1;
10441 	else
10442 #endif
10443 		rack->do_detection = 0;
10444 	if (rack_non_rxt_use_cr)
10445 		rack->rack_rec_nonrxt_use_cr = 1;
10446 	if (tp->snd_una != tp->snd_max) {
10447 		/* Create a send map for the current outstanding data */
10448 		struct rack_sendmap *rsm;
10449 
10450 		rsm = rack_alloc(rack);
10451 		if (rsm == NULL) {
10452 			uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
10453 			tp->t_fb_ptr = NULL;
10454 			return (ENOMEM);
10455 		}
10456 		rsm->r_flags = RACK_OVERMAX;
10457 		rsm->r_tim_lastsent[0] = rack->r_ctl.rc_tlp_rxt_last_time;
10458 		rsm->r_rtr_cnt = 1;
10459 		rsm->r_rtr_bytes = 0;
10460 		rsm->r_start = tp->snd_una;
10461 		if (tp->t_flags & TF_SENTFIN) {
10462 			rsm->r_end = tp->snd_max - 1;
10463 			rsm->r_flags |= RACK_HAS_FIN;
10464 		} else {
10465 			rsm->r_end = tp->snd_max;
10466 		}
10467 		rsm->usec_orig_send = us_cts;
10468 		rsm->r_dupack = 0;
10469 		insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
10470 #ifdef INVARIANTS
10471 		if (insret != NULL) {
10472 			panic("Insert in rb tree fails ret:%p rack:%p rsm:%p",
10473 			      insret, rack, rsm);
10474 		}
10475 #endif
10476 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
10477 		rsm->r_in_tmap = 1;
10478 	}
10479 	/* Cancel the GP measurement in progress */
10480 	tp->t_flags &= ~TF_GPUTINPROG;
10481 	if (SEQ_GT(tp->snd_max, tp->iss))
10482 		snt = tp->snd_max - tp->iss;
10483 	else
10484 		snt = 0;
10485 	iwin = rc_init_window(rack);
10486 	if (snt < iwin) {
10487 		/* We are not past the initial window
10488 		 * so we need to make sure cwnd is
10489 		 * correct.
10490 		 */
10491 		if (tp->snd_cwnd < iwin)
10492 			tp->snd_cwnd = iwin;
10493 		/*
10494 		 * If we are within the initial window
10495 		 * we want ssthresh to be unlimited. Setting
10496 		 * it to the rwnd (which the default stack does
10497 		 * and older racks) is not really a good idea
10498 		 * since we want to be in SS and grow both the
10499 		 * cwnd and the rwnd (via dynamic rwnd growth). If
10500 		 * we set it to the rwnd then as the peer grows its
10501 		 * rwnd we will be stuck in CA and never hit SS.
10502 		 *
10503 		 * Its far better to raise it up high (this takes the
10504 		 * risk that there as been a loss already, probably
10505 		 * we should have an indicator in all stacks of loss
10506 		 * but we don't), but considering the normal use this
10507 		 * is a risk worth taking. The consequences of not
10508 		 * hitting SS are far worse than going one more time
10509 		 * into it early on (before we have sent even a IW).
10510 		 * It is highly unlikely that we will have had a loss
10511 		 * before getting the IW out.
10512 		 */
10513 		tp->snd_ssthresh = 0xffffffff;
10514 	}
10515 	rack_stop_all_timers(tp);
10516 	rack_start_hpts_timer(rack, tp, tcp_ts_getticks(), 0, 0, 0);
10517 	rack_log_rtt_shrinks(rack,  us_cts,  0,
10518 			     __LINE__, RACK_RTTS_INIT);
10519 	return (0);
10520 }
10521 
10522 static int
rack_handoff_ok(struct tcpcb * tp)10523 rack_handoff_ok(struct tcpcb *tp)
10524 {
10525 	if ((tp->t_state == TCPS_CLOSED) ||
10526 	    (tp->t_state == TCPS_LISTEN)) {
10527 		/* Sure no problem though it may not stick */
10528 		return (0);
10529 	}
10530 	if ((tp->t_state == TCPS_SYN_SENT) ||
10531 	    (tp->t_state == TCPS_SYN_RECEIVED)) {
10532 		/*
10533 		 * We really don't know if you support sack,
10534 		 * you have to get to ESTAB or beyond to tell.
10535 		 */
10536 		return (EAGAIN);
10537 	}
10538 	if ((tp->t_flags & TF_SENTFIN) && ((tp->snd_max - tp->snd_una) > 1)) {
10539 		/*
10540 		 * Rack will only send a FIN after all data is acknowledged.
10541 		 * So in this case we have more data outstanding. We can't
10542 		 * switch stacks until either all data and only the FIN
10543 		 * is left (in which case rack_init() now knows how
10544 		 * to deal with that) <or> all is acknowledged and we
10545 		 * are only left with incoming data, though why you
10546 		 * would want to switch to rack after all data is acknowledged
10547 		 * I have no idea (rrs)!
10548 		 */
10549 		return (EAGAIN);
10550 	}
10551 	if ((tp->t_flags & TF_SACK_PERMIT) || rack_sack_not_required){
10552 		return (0);
10553 	}
10554 	/*
10555 	 * If we reach here we don't do SACK on this connection so we can
10556 	 * never do rack.
10557 	 */
10558 	return (EINVAL);
10559 }
10560 
10561 static void
rack_fini(struct tcpcb * tp,int32_t tcb_is_purged)10562 rack_fini(struct tcpcb *tp, int32_t tcb_is_purged)
10563 {
10564 	if (tp->t_fb_ptr) {
10565 		struct tcp_rack *rack;
10566 		struct rack_sendmap *rsm, *nrsm, *rm;
10567 
10568 		rack = (struct tcp_rack *)tp->t_fb_ptr;
10569 #ifdef NETFLIX_SHARED_CWND
10570 		if (rack->r_ctl.rc_scw) {
10571 			uint32_t limit;
10572 
10573 			if (rack->r_limit_scw)
10574 				limit = max(1, rack->r_ctl.rc_lowest_us_rtt);
10575 			else
10576 				limit = 0;
10577 			tcp_shared_cwnd_free_full(tp, rack->r_ctl.rc_scw,
10578 						  rack->r_ctl.rc_scw_index,
10579 						  limit);
10580 			rack->r_ctl.rc_scw = NULL;
10581 		}
10582 #endif
10583 		/* rack does not use force data but other stacks may clear it */
10584 		tp->t_flags &= ~TF_FORCEDATA;
10585 		if (tp->t_inpcb) {
10586 			tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
10587 			tp->t_inpcb->inp_flags2 &= ~INP_MBUF_QUEUE_READY;
10588 			tp->t_inpcb->inp_flags2 &= ~INP_DONT_SACK_QUEUE;
10589 		}
10590 #ifdef TCP_BLACKBOX
10591 		tcp_log_flowend(tp);
10592 #endif
10593 		RB_FOREACH_SAFE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm) {
10594 			rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
10595 #ifdef INVARIANTS
10596 			if (rm != rsm) {
10597 				panic("At fini, rack:%p rsm:%p rm:%p",
10598 				      rack, rsm, rm);
10599 			}
10600 #endif
10601 			uma_zfree(rack_zone, rsm);
10602 		}
10603 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
10604 		while (rsm) {
10605 			TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
10606 			uma_zfree(rack_zone, rsm);
10607 			rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
10608 		}
10609 		rack->rc_free_cnt = 0;
10610 		uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
10611 		tp->t_fb_ptr = NULL;
10612 	}
10613 	/* Cancel the GP measurement in progress */
10614 	tp->t_flags &= ~TF_GPUTINPROG;
10615 	/* Make sure snd_nxt is correctly set */
10616 	tp->snd_nxt = tp->snd_max;
10617 }
10618 
10619 static void
rack_set_state(struct tcpcb * tp,struct tcp_rack * rack)10620 rack_set_state(struct tcpcb *tp, struct tcp_rack *rack)
10621 {
10622 	switch (tp->t_state) {
10623 	case TCPS_SYN_SENT:
10624 		rack->r_state = TCPS_SYN_SENT;
10625 		rack->r_substate = rack_do_syn_sent;
10626 		break;
10627 	case TCPS_SYN_RECEIVED:
10628 		rack->r_state = TCPS_SYN_RECEIVED;
10629 		rack->r_substate = rack_do_syn_recv;
10630 		break;
10631 	case TCPS_ESTABLISHED:
10632 		rack_set_pace_segments(tp, rack, __LINE__);
10633 		rack->r_state = TCPS_ESTABLISHED;
10634 		rack->r_substate = rack_do_established;
10635 		break;
10636 	case TCPS_CLOSE_WAIT:
10637 		rack->r_state = TCPS_CLOSE_WAIT;
10638 		rack->r_substate = rack_do_close_wait;
10639 		break;
10640 	case TCPS_FIN_WAIT_1:
10641 		rack->r_state = TCPS_FIN_WAIT_1;
10642 		rack->r_substate = rack_do_fin_wait_1;
10643 		break;
10644 	case TCPS_CLOSING:
10645 		rack->r_state = TCPS_CLOSING;
10646 		rack->r_substate = rack_do_closing;
10647 		break;
10648 	case TCPS_LAST_ACK:
10649 		rack->r_state = TCPS_LAST_ACK;
10650 		rack->r_substate = rack_do_lastack;
10651 		break;
10652 	case TCPS_FIN_WAIT_2:
10653 		rack->r_state = TCPS_FIN_WAIT_2;
10654 		rack->r_substate = rack_do_fin_wait_2;
10655 		break;
10656 	case TCPS_LISTEN:
10657 	case TCPS_CLOSED:
10658 	case TCPS_TIME_WAIT:
10659 	default:
10660 		break;
10661 	};
10662 }
10663 
10664 static void
rack_timer_audit(struct tcpcb * tp,struct tcp_rack * rack,struct sockbuf * sb)10665 rack_timer_audit(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb)
10666 {
10667 	/*
10668 	 * We received an ack, and then did not
10669 	 * call send or were bounced out due to the
10670 	 * hpts was running. Now a timer is up as well, is
10671 	 * it the right timer?
10672 	 */
10673 	struct rack_sendmap *rsm;
10674 	int tmr_up;
10675 
10676 	tmr_up = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
10677 	if (rack->rc_in_persist && (tmr_up == PACE_TMR_PERSIT))
10678 		return;
10679 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
10680 	if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) &&
10681 	    (tmr_up == PACE_TMR_RXT)) {
10682 		/* Should be an RXT */
10683 		return;
10684 	}
10685 	if (rsm == NULL) {
10686 		/* Nothing outstanding? */
10687 		if (tp->t_flags & TF_DELACK) {
10688 			if (tmr_up == PACE_TMR_DELACK)
10689 				/* We are supposed to have delayed ack up and we do */
10690 				return;
10691 		} else if (sbavail(&tp->t_inpcb->inp_socket->so_snd) && (tmr_up == PACE_TMR_RXT)) {
10692 			/*
10693 			 * if we hit enobufs then we would expect the possiblity
10694 			 * of nothing outstanding and the RXT up (and the hptsi timer).
10695 			 */
10696 			return;
10697 		} else if (((V_tcp_always_keepalive ||
10698 			     rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
10699 			    (tp->t_state <= TCPS_CLOSING)) &&
10700 			   (tmr_up == PACE_TMR_KEEP) &&
10701 			   (tp->snd_max == tp->snd_una)) {
10702 			/* We should have keep alive up and we do */
10703 			return;
10704 		}
10705 	}
10706 	if (SEQ_GT(tp->snd_max, tp->snd_una) &&
10707 		   ((tmr_up == PACE_TMR_TLP) ||
10708 		    (tmr_up == PACE_TMR_RACK) ||
10709 		    (tmr_up == PACE_TMR_RXT))) {
10710 		/*
10711 		 * Either a Rack, TLP or RXT is fine if  we
10712 		 * have outstanding data.
10713 		 */
10714 		return;
10715 	} else if (tmr_up == PACE_TMR_DELACK) {
10716 		/*
10717 		 * If the delayed ack was going to go off
10718 		 * before the rtx/tlp/rack timer were going to
10719 		 * expire, then that would be the timer in control.
10720 		 * Note we don't check the time here trusting the
10721 		 * code is correct.
10722 		 */
10723 		return;
10724 	}
10725 	/*
10726 	 * Ok the timer originally started is not what we want now.
10727 	 * We will force the hpts to be stopped if any, and restart
10728 	 * with the slot set to what was in the saved slot.
10729 	 */
10730 	if (rack->rc_inp->inp_in_hpts) {
10731 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
10732 			uint32_t us_cts;
10733 
10734 			us_cts = tcp_get_usecs(NULL);
10735 			if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
10736 				rack->r_early = 1;
10737 				rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
10738 			}
10739 			rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
10740 		}
10741 		tcp_hpts_remove(tp->t_inpcb, HPTS_REMOVE_OUTPUT);
10742 	}
10743 	rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
10744 	rack_start_hpts_timer(rack, tp, tcp_ts_getticks(), 0, 0, 0);
10745 }
10746 
10747 static int
rack_do_segment_nounlock(struct mbuf * m,struct tcphdr * th,struct socket * so,struct tcpcb * tp,int32_t drop_hdrlen,int32_t tlen,uint8_t iptos,int32_t nxt_pkt,struct timeval * tv)10748 rack_do_segment_nounlock(struct mbuf *m, struct tcphdr *th, struct socket *so,
10749     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos,
10750     int32_t nxt_pkt, struct timeval *tv)
10751 {
10752 	int32_t thflags, retval, did_out = 0;
10753 	int32_t way_out = 0;
10754 	uint32_t cts;
10755 	uint32_t tiwin;
10756 	struct timespec ts;
10757 	struct tcpopt to;
10758 	struct tcp_rack *rack;
10759 	struct rack_sendmap *rsm;
10760 	int32_t prev_state = 0;
10761 	uint32_t us_cts;
10762 	/*
10763 	 * tv passed from common code is from either M_TSTMP_LRO or
10764 	 * tcp_get_usecs() if no LRO m_pkthdr timestamp is present. The
10765 	 * rack_pacing stack assumes tv always refers to 'now', so we overwrite
10766 	 * tv here to guarantee that.
10767 	 */
10768 	if (m->m_flags & M_TSTMP_LRO)
10769 		tcp_get_usecs(tv);
10770 
10771 	cts = tcp_tv_to_mssectick(tv);
10772 	rack = (struct tcp_rack *)tp->t_fb_ptr;
10773 
10774 	if ((m->m_flags & M_TSTMP) ||
10775 	    (m->m_flags & M_TSTMP_LRO)) {
10776 		mbuf_tstmp2timespec(m, &ts);
10777 		rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
10778 		rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
10779 	} else
10780 		rack->r_ctl.act_rcv_time = *tv;
10781 	kern_prefetch(rack, &prev_state);
10782 	prev_state = 0;
10783 	thflags = th->th_flags;
10784 
10785 	NET_EPOCH_ASSERT();
10786 	INP_WLOCK_ASSERT(tp->t_inpcb);
10787 	KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN",
10788 	    __func__));
10789 	KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT",
10790 	    __func__));
10791 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
10792 		union tcp_log_stackspecific log;
10793 		struct timeval ltv;
10794 #ifdef NETFLIX_HTTP_LOGGING
10795 		struct http_sendfile_track *http_req;
10796 
10797 		if (SEQ_GT(th->th_ack, tp->snd_una)) {
10798 			http_req = tcp_http_find_req_for_seq(tp, (th->th_ack-1));
10799 		} else {
10800 			http_req = tcp_http_find_req_for_seq(tp, th->th_ack);
10801 		}
10802 #endif
10803 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
10804 		log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
10805 		log.u_bbr.ininput = rack->rc_inp->inp_in_input;
10806 		if (rack->rack_no_prr == 0)
10807 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
10808 		else
10809 			log.u_bbr.flex1 = 0;
10810 		log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
10811 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
10812 		log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
10813 		log.u_bbr.flex3 = m->m_flags;
10814 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
10815 		if (m->m_flags & M_TSTMP) {
10816 			/* Record the hardware timestamp if present */
10817 			mbuf_tstmp2timespec(m, &ts);
10818 			ltv.tv_sec = ts.tv_sec;
10819 			ltv.tv_usec = ts.tv_nsec / 1000;
10820 			log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
10821 		} else if (m->m_flags & M_TSTMP_LRO) {
10822 			/* Record the LRO the arrival timestamp */
10823 			mbuf_tstmp2timespec(m, &ts);
10824 			ltv.tv_sec = ts.tv_sec;
10825 			ltv.tv_usec = ts.tv_nsec / 1000;
10826 			log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
10827 		}
10828 		log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
10829 		/* Log the rcv time */
10830 		log.u_bbr.delRate = m->m_pkthdr.rcv_tstmp;
10831 #ifdef NETFLIX_HTTP_LOGGING
10832 		log.u_bbr.applimited = tp->t_http_closed;
10833 		log.u_bbr.applimited <<= 8;
10834 		log.u_bbr.applimited |= tp->t_http_open;
10835 		log.u_bbr.applimited <<= 8;
10836 		log.u_bbr.applimited |= tp->t_http_req;
10837 		if (http_req) {
10838 			/* Copy out any client req info */
10839 			/* seconds */
10840 			log.u_bbr.pkt_epoch = (http_req->localtime / HPTS_USEC_IN_SEC);
10841 			/* useconds */
10842 			log.u_bbr.delivered = (http_req->localtime % HPTS_USEC_IN_SEC);
10843 			log.u_bbr.rttProp = http_req->timestamp;
10844 			log.u_bbr.cur_del_rate = http_req->start;
10845 			if (http_req->flags & TCP_HTTP_TRACK_FLG_OPEN) {
10846 				log.u_bbr.flex8 |= 1;
10847 			} else {
10848 				log.u_bbr.flex8 |= 2;
10849 				log.u_bbr.bw_inuse = http_req->end;
10850 			}
10851 			log.u_bbr.flex6 = http_req->start_seq;
10852 			if (http_req->flags & TCP_HTTP_TRACK_FLG_COMP) {
10853 				log.u_bbr.flex8 |= 4;
10854 				log.u_bbr.epoch = http_req->end_seq;
10855 			}
10856 		}
10857 #endif
10858 		TCP_LOG_EVENTP(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_IN, 0,
10859 		    tlen, &log, true, &ltv);
10860 	}
10861 	if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) {
10862 		way_out = 4;
10863 		retval = 0;
10864 		goto done_with_input;
10865 	}
10866 	/*
10867 	 * If a segment with the ACK-bit set arrives in the SYN-SENT state
10868 	 * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9.
10869 	 */
10870 	if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) &&
10871 	    (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) {
10872 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
10873 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
10874 		return(1);
10875 	}
10876 
10877 	/*
10878 	 * Parse options on any incoming segment.
10879 	 */
10880 	tcp_dooptions(&to, (u_char *)(th + 1),
10881 	    (th->th_off << 2) - sizeof(struct tcphdr),
10882 	    (thflags & TH_SYN) ? TO_SYN : 0);
10883 
10884 	/*
10885 	 * If timestamps were negotiated during SYN/ACK and a
10886 	 * segment without a timestamp is received, silently drop
10887 	 * the segment, unless it is a RST segment or missing timestamps are
10888 	 * tolerated.
10889 	 * See section 3.2 of RFC 7323.
10890 	 */
10891 	if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS) &&
10892 	    ((thflags & TH_RST) == 0) && (V_tcp_tolerate_missing_ts == 0)) {
10893 		way_out = 5;
10894 		retval = 0;
10895 		goto done_with_input;
10896 	}
10897 
10898 	/*
10899 	 * Segment received on connection. Reset idle time and keep-alive
10900 	 * timer. XXX: This should be done after segment validation to
10901 	 * ignore broken/spoofed segs.
10902 	 */
10903 	if  (tp->t_idle_reduce &&
10904 	     (tp->snd_max == tp->snd_una) &&
10905 	     ((ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
10906 		counter_u64_add(rack_input_idle_reduces, 1);
10907 		rack_cc_after_idle(rack, tp);
10908 	}
10909 	tp->t_rcvtime = ticks;
10910 	/*
10911 	 * Unscale the window into a 32-bit value. For the SYN_SENT state
10912 	 * the scale is zero.
10913 	 */
10914 	tiwin = th->th_win << tp->snd_scale;
10915 #ifdef STATS
10916 	stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin);
10917 #endif
10918 	if (tiwin > rack->r_ctl.rc_high_rwnd)
10919 		rack->r_ctl.rc_high_rwnd = tiwin;
10920 	/*
10921 	 * TCP ECN processing. XXXJTL: If we ever use ECN, we need to move
10922 	 * this to occur after we've validated the segment.
10923 	 */
10924 	if (tp->t_flags2 & TF2_ECN_PERMIT) {
10925 		if (thflags & TH_CWR) {
10926 			tp->t_flags2 &= ~TF2_ECN_SND_ECE;
10927 			tp->t_flags |= TF_ACKNOW;
10928 		}
10929 		switch (iptos & IPTOS_ECN_MASK) {
10930 		case IPTOS_ECN_CE:
10931 			tp->t_flags2 |= TF2_ECN_SND_ECE;
10932 			KMOD_TCPSTAT_INC(tcps_ecn_ce);
10933 			break;
10934 		case IPTOS_ECN_ECT0:
10935 			KMOD_TCPSTAT_INC(tcps_ecn_ect0);
10936 			break;
10937 		case IPTOS_ECN_ECT1:
10938 			KMOD_TCPSTAT_INC(tcps_ecn_ect1);
10939 			break;
10940 		}
10941 
10942 		/* Process a packet differently from RFC3168. */
10943 		cc_ecnpkt_handler(tp, th, iptos);
10944 
10945 		/* Congestion experienced. */
10946 		if (thflags & TH_ECE) {
10947 			rack_cong_signal(tp, th, CC_ECN);
10948 		}
10949 	}
10950 
10951 	/*
10952 	 * If echoed timestamp is later than the current time, fall back to
10953 	 * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
10954 	 * were used when this connection was established.
10955 	 */
10956 	if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) {
10957 		to.to_tsecr -= tp->ts_offset;
10958 		if (TSTMP_GT(to.to_tsecr, cts))
10959 			to.to_tsecr = 0;
10960 	}
10961 
10962 	/*
10963 	 * If its the first time in we need to take care of options and
10964 	 * verify we can do SACK for rack!
10965 	 */
10966 	if (rack->r_state == 0) {
10967 		/* Should be init'd by rack_init() */
10968 		KASSERT(rack->rc_inp != NULL,
10969 		    ("%s: rack->rc_inp unexpectedly NULL", __func__));
10970 		if (rack->rc_inp == NULL) {
10971 			rack->rc_inp = tp->t_inpcb;
10972 		}
10973 
10974 		/*
10975 		 * Process options only when we get SYN/ACK back. The SYN
10976 		 * case for incoming connections is handled in tcp_syncache.
10977 		 * According to RFC1323 the window field in a SYN (i.e., a
10978 		 * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX
10979 		 * this is traditional behavior, may need to be cleaned up.
10980 		 */
10981 		if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) {
10982 			/* Handle parallel SYN for ECN */
10983 			if (!(thflags & TH_ACK) &&
10984 			    ((thflags & (TH_CWR | TH_ECE)) == (TH_CWR | TH_ECE)) &&
10985 			    ((V_tcp_do_ecn == 1) || (V_tcp_do_ecn == 2))) {
10986 				tp->t_flags2 |= TF2_ECN_PERMIT;
10987 				tp->t_flags2 |= TF2_ECN_SND_ECE;
10988 				TCPSTAT_INC(tcps_ecn_shs);
10989 			}
10990 			if ((to.to_flags & TOF_SCALE) &&
10991 			    (tp->t_flags & TF_REQ_SCALE)) {
10992 				tp->t_flags |= TF_RCVD_SCALE;
10993 				tp->snd_scale = to.to_wscale;
10994 			} else
10995 				tp->t_flags &= ~TF_REQ_SCALE;
10996 			/*
10997 			 * Initial send window.  It will be updated with the
10998 			 * next incoming segment to the scaled value.
10999 			 */
11000 			tp->snd_wnd = th->th_win;
11001 			if ((to.to_flags & TOF_TS) &&
11002 			    (tp->t_flags & TF_REQ_TSTMP)) {
11003 				tp->t_flags |= TF_RCVD_TSTMP;
11004 				tp->ts_recent = to.to_tsval;
11005 				tp->ts_recent_age = cts;
11006 			} else
11007 				tp->t_flags &= ~TF_REQ_TSTMP;
11008 			if (to.to_flags & TOF_MSS)
11009 				tcp_mss(tp, to.to_mss);
11010 			if ((tp->t_flags & TF_SACK_PERMIT) &&
11011 			    (to.to_flags & TOF_SACKPERM) == 0)
11012 				tp->t_flags &= ~TF_SACK_PERMIT;
11013 			if (IS_FASTOPEN(tp->t_flags)) {
11014 				if (to.to_flags & TOF_FASTOPEN) {
11015 					uint16_t mss;
11016 
11017 					if (to.to_flags & TOF_MSS)
11018 						mss = to.to_mss;
11019 					else
11020 						if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0)
11021 							mss = TCP6_MSS;
11022 						else
11023 							mss = TCP_MSS;
11024 					tcp_fastopen_update_cache(tp, mss,
11025 					    to.to_tfo_len, to.to_tfo_cookie);
11026 				} else
11027 					tcp_fastopen_disable_path(tp);
11028 			}
11029 		}
11030 		/*
11031 		 * At this point we are at the initial call. Here we decide
11032 		 * if we are doing RACK or not. We do this by seeing if
11033 		 * TF_SACK_PERMIT is set and the sack-not-required is clear.
11034 		 * The code now does do dup-ack counting so if you don't
11035 		 * switch back you won't get rack & TLP, but you will still
11036 		 * get this stack.
11037 		 */
11038 
11039 		if ((rack_sack_not_required == 0) &&
11040 		    ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
11041 			tcp_switch_back_to_default(tp);
11042 			(*tp->t_fb->tfb_tcp_do_segment) (m, th, so, tp, drop_hdrlen,
11043 			    tlen, iptos);
11044 			return (1);
11045 		}
11046 		/* Set the flag */
11047 		rack->r_is_v6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
11048 		tcp_set_hpts(tp->t_inpcb);
11049 		sack_filter_clear(&rack->r_ctl.rack_sf, th->th_ack);
11050 	}
11051 	if (thflags & TH_FIN)
11052 		tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN);
11053 	us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
11054 	if ((rack->rc_gp_dyn_mul) &&
11055 	    (rack->use_fixed_rate == 0) &&
11056 	    (rack->rc_always_pace)) {
11057 		/* Check in on probertt */
11058 		rack_check_probe_rtt(rack, us_cts);
11059 	}
11060 	if (rack->forced_ack) {
11061 		uint32_t us_rtt;
11062 
11063 		/*
11064 		 * A persist or keep-alive was forced out, update our
11065 		 * min rtt time. Note we do not worry about lost
11066 		 * retransmissions since KEEP-ALIVES and persists
11067 		 * are usually way long on times of sending (though
11068 		 * if we were really paranoid or worried we could
11069 		 * at least use timestamps if available to validate).
11070 		 */
11071 		rack->forced_ack = 0;
11072 		us_rtt = us_cts - rack->r_ctl.forced_ack_ts;
11073 		if (us_rtt == 0)
11074 			us_rtt = 1;
11075 		rack_log_rtt_upd(tp, rack, us_rtt, 0, NULL, 3);
11076 		rack_apply_updated_usrtt(rack, us_rtt, us_cts);
11077 	}
11078 	/*
11079 	 * This is the one exception case where we set the rack state
11080 	 * always. All other times (timers etc) we must have a rack-state
11081 	 * set (so we assure we have done the checks above for SACK).
11082 	 */
11083 	rack->r_ctl.rc_rcvtime = cts;
11084 	if (rack->r_state != tp->t_state)
11085 		rack_set_state(tp, rack);
11086 	if (SEQ_GT(th->th_ack, tp->snd_una) &&
11087 	    (rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree)) != NULL)
11088 		kern_prefetch(rsm, &prev_state);
11089 	prev_state = rack->r_state;
11090 	rack_clear_rate_sample(rack);
11091 	retval = (*rack->r_substate) (m, th, so,
11092 	    tp, &to, drop_hdrlen,
11093 	    tlen, tiwin, thflags, nxt_pkt, iptos);
11094 #ifdef INVARIANTS
11095 	if ((retval == 0) &&
11096 	    (tp->t_inpcb == NULL)) {
11097 		panic("retval:%d tp:%p t_inpcb:NULL state:%d",
11098 		    retval, tp, prev_state);
11099 	}
11100 #endif
11101 	if (retval == 0) {
11102 		/*
11103 		 * If retval is 1 the tcb is unlocked and most likely the tp
11104 		 * is gone.
11105 		 */
11106 		INP_WLOCK_ASSERT(tp->t_inpcb);
11107 		if ((rack->rc_gp_dyn_mul) &&
11108 		    (rack->rc_always_pace) &&
11109 		    (rack->use_fixed_rate == 0) &&
11110 		    rack->in_probe_rtt &&
11111 		    (rack->r_ctl.rc_time_probertt_starts == 0)) {
11112 			/*
11113 			 * If we are going for target, lets recheck before
11114 			 * we output.
11115 			 */
11116 			rack_check_probe_rtt(rack, us_cts);
11117 		}
11118 		if (rack->set_pacing_done_a_iw == 0) {
11119 			/* How much has been acked? */
11120 			if ((tp->snd_una - tp->iss) > (ctf_fixed_maxseg(tp) * 10)) {
11121 				/* We have enough to set in the pacing segment size */
11122 				rack->set_pacing_done_a_iw = 1;
11123 				rack_set_pace_segments(tp, rack, __LINE__);
11124 			}
11125 		}
11126 		tcp_rack_xmit_timer_commit(rack, tp);
11127 		if (nxt_pkt == 0) {
11128 			if (rack->r_wanted_output != 0) {
11129 do_output_now:
11130 				did_out = 1;
11131 				(void)tp->t_fb->tfb_tcp_output(tp);
11132 			}
11133 			rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
11134 		}
11135 		if ((nxt_pkt == 0) &&
11136 		    ((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) &&
11137 		    (SEQ_GT(tp->snd_max, tp->snd_una) ||
11138 		     (tp->t_flags & TF_DELACK) ||
11139 		     ((V_tcp_always_keepalive || rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
11140 		      (tp->t_state <= TCPS_CLOSING)))) {
11141 			/* We could not send (probably in the hpts but stopped the timer earlier)? */
11142 			if ((tp->snd_max == tp->snd_una) &&
11143 			    ((tp->t_flags & TF_DELACK) == 0) &&
11144 			    (rack->rc_inp->inp_in_hpts) &&
11145 			    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
11146 				/* keep alive not needed if we are hptsi output yet */
11147 				;
11148 			} else {
11149 				int late = 0;
11150 				if (rack->rc_inp->inp_in_hpts) {
11151 					if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
11152 						us_cts = tcp_get_usecs(NULL);
11153 						if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
11154 							rack->r_early = 1;
11155 							rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
11156 						} else
11157 							late = 1;
11158 						rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
11159 					}
11160 					tcp_hpts_remove(tp->t_inpcb, HPTS_REMOVE_OUTPUT);
11161 				}
11162 				if (late && (did_out == 0)) {
11163 					/*
11164 					 * We are late in the sending
11165 					 * and we did not call the output
11166 					 * (this probably should not happen).
11167 					 */
11168 					goto do_output_now;
11169 				}
11170 				rack_start_hpts_timer(rack, tp, tcp_ts_getticks(), 0, 0, 0);
11171 			}
11172 			way_out = 1;
11173 		} else if (nxt_pkt == 0) {
11174 			/* Do we have the correct timer running? */
11175 			rack_timer_audit(tp, rack, &so->so_snd);
11176 			way_out = 2;
11177 		}
11178 	done_with_input:
11179 		rack_log_doseg_done(rack, cts, nxt_pkt, did_out, way_out);
11180 		if (did_out)
11181 			rack->r_wanted_output = 0;
11182 #ifdef INVARIANTS
11183 		if (tp->t_inpcb == NULL) {
11184 			panic("OP:%d retval:%d tp:%p t_inpcb:NULL state:%d",
11185 			      did_out,
11186 			      retval, tp, prev_state);
11187 		}
11188 #endif
11189 	}
11190 	return (retval);
11191 }
11192 
11193 void
rack_do_segment(struct mbuf * m,struct tcphdr * th,struct socket * so,struct tcpcb * tp,int32_t drop_hdrlen,int32_t tlen,uint8_t iptos)11194 rack_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so,
11195     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos)
11196 {
11197 	struct timeval tv;
11198 
11199 	/* First lets see if we have old packets */
11200 	if (tp->t_in_pkt) {
11201 		if (ctf_do_queued_segments(so, tp, 1)) {
11202 			m_freem(m);
11203 			return;
11204 		}
11205 	}
11206 	if (m->m_flags & M_TSTMP_LRO) {
11207 		tv.tv_sec = m->m_pkthdr.rcv_tstmp /1000000000;
11208 		tv.tv_usec = (m->m_pkthdr.rcv_tstmp % 1000000000)/1000;
11209 	} else {
11210 		/* Should not be should we kassert instead? */
11211 		tcp_get_usecs(&tv);
11212 	}
11213 	if(rack_do_segment_nounlock(m, th, so, tp,
11214 				    drop_hdrlen, tlen, iptos, 0, &tv) == 0) {
11215 		tcp_handle_wakeup(tp, so);
11216 		INP_WUNLOCK(tp->t_inpcb);
11217 	}
11218 }
11219 
11220 struct rack_sendmap *
tcp_rack_output(struct tcpcb * tp,struct tcp_rack * rack,uint32_t tsused)11221 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tsused)
11222 {
11223 	struct rack_sendmap *rsm = NULL;
11224 	int32_t idx;
11225 	uint32_t srtt = 0, thresh = 0, ts_low = 0;
11226 
11227 	/* Return the next guy to be re-transmitted */
11228 	if (RB_EMPTY(&rack->r_ctl.rc_mtree)) {
11229 		return (NULL);
11230 	}
11231 	if (tp->t_flags & TF_SENTFIN) {
11232 		/* retran the end FIN? */
11233 		return (NULL);
11234 	}
11235 	/* ok lets look at this one */
11236 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
11237 	if (rsm && ((rsm->r_flags & RACK_ACKED) == 0)) {
11238 		goto check_it;
11239 	}
11240 	rsm = rack_find_lowest_rsm(rack);
11241 	if (rsm == NULL) {
11242 		return (NULL);
11243 	}
11244 check_it:
11245 	if (rsm->r_flags & RACK_ACKED) {
11246 		return (NULL);
11247 	}
11248 	if (((rsm->r_flags & RACK_SACK_PASSED) == 0) &&
11249 	    (rsm->r_dupack < DUP_ACK_THRESHOLD)) {
11250 		/* Its not yet ready */
11251 		return (NULL);
11252 	}
11253 	srtt = rack_grab_rtt(tp, rack);
11254 	idx = rsm->r_rtr_cnt - 1;
11255 	ts_low = rsm->r_tim_lastsent[idx];
11256 	thresh = rack_calc_thresh_rack(rack, srtt, tsused);
11257 	if ((tsused == ts_low) ||
11258 	    (TSTMP_LT(tsused, ts_low))) {
11259 		/* No time since sending */
11260 		return (NULL);
11261 	}
11262 	if ((tsused - ts_low) < thresh) {
11263 		/* It has not been long enough yet */
11264 		return (NULL);
11265 	}
11266 	if ((rsm->r_dupack >= DUP_ACK_THRESHOLD) ||
11267 	    ((rsm->r_flags & RACK_SACK_PASSED) &&
11268 	     (rack->sack_attack_disable == 0))) {
11269 		/*
11270 		 * We have passed the dup-ack threshold <or>
11271 		 * a SACK has indicated this is missing.
11272 		 * Note that if you are a declared attacker
11273 		 * it is only the dup-ack threshold that
11274 		 * will cause retransmits.
11275 		 */
11276 		/* log retransmit reason */
11277 		rack_log_retran_reason(rack, rsm, (tsused - ts_low), thresh, 1);
11278 		return (rsm);
11279 	}
11280 	return (NULL);
11281 }
11282 
11283 static void
rack_log_pacing_delay_calc(struct tcp_rack * rack,uint32_t len,uint32_t slot,uint64_t bw_est,uint64_t bw,uint64_t len_time,int method,int line,struct rack_sendmap * rsm)11284 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
11285 			   uint64_t bw_est, uint64_t bw, uint64_t len_time, int method,
11286 			   int line, struct rack_sendmap *rsm)
11287 {
11288 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
11289 		union tcp_log_stackspecific log;
11290 		struct timeval tv;
11291 
11292 		memset(&log, 0, sizeof(log));
11293 		log.u_bbr.flex1 = slot;
11294 		log.u_bbr.flex2 = len;
11295 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_min_segs;
11296 		log.u_bbr.flex4 = rack->r_ctl.rc_pace_max_segs;
11297 		log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ss;
11298 		log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_ca;
11299 		log.u_bbr.use_lt_bw = rack->app_limited_needs_set;
11300 		log.u_bbr.use_lt_bw <<= 1;
11301 		log.u_bbr.use_lt_bw = rack->rc_gp_filled;
11302 		log.u_bbr.use_lt_bw <<= 1;
11303 		log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
11304 		log.u_bbr.use_lt_bw <<= 1;
11305 		log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
11306 		log.u_bbr.pkt_epoch = line;
11307 		log.u_bbr.applimited = rack->r_ctl.rack_per_of_gp_rec;
11308 		log.u_bbr.bw_inuse = bw_est;
11309 		log.u_bbr.delRate = bw;
11310 		if (rack->r_ctl.gp_bw == 0)
11311 			log.u_bbr.cur_del_rate = 0;
11312 		else
11313 			log.u_bbr.cur_del_rate = rack_get_bw(rack);
11314 		log.u_bbr.rttProp = len_time;
11315 		log.u_bbr.pkts_out = rack->r_ctl.rc_rack_min_rtt;
11316 		log.u_bbr.lost = rack->r_ctl.rc_probertt_sndmax_atexit;
11317 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
11318 		if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh) {
11319 			/* We are in slow start */
11320 			log.u_bbr.flex7 = 1;
11321 		} else {
11322 			/* we are on congestion avoidance */
11323 			log.u_bbr.flex7 = 0;
11324 		}
11325 		log.u_bbr.flex8 = method;
11326 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
11327 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
11328 		log.u_bbr.cwnd_gain = rack->rc_gp_saw_rec;
11329 		log.u_bbr.cwnd_gain <<= 1;
11330 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
11331 		log.u_bbr.cwnd_gain <<= 1;
11332 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
11333 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
11334 		    &rack->rc_inp->inp_socket->so_rcv,
11335 		    &rack->rc_inp->inp_socket->so_snd,
11336 		    BBR_LOG_HPTSI_CALC, 0,
11337 		    0, &log, false, &tv);
11338 	}
11339 }
11340 
11341 static uint32_t
rack_get_pacing_len(struct tcp_rack * rack,uint64_t bw,uint32_t mss)11342 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss)
11343 {
11344 	uint32_t new_tso, user_max;
11345 
11346 	user_max = rack->rc_user_set_max_segs * mss;
11347 	if (rack->rc_force_max_seg) {
11348 		return (user_max);
11349 	}
11350 	if (rack->use_fixed_rate &&
11351 	    ((rack->r_ctl.crte == NULL) ||
11352 	     (bw != rack->r_ctl.crte->rate))) {
11353 		/* Use the user mss since we are not exactly matched */
11354 		return (user_max);
11355 	}
11356 	new_tso = tcp_get_pacing_burst_size(bw, mss, rack_pace_one_seg, rack->r_ctl.crte, NULL);
11357 	if (new_tso > user_max)
11358 		new_tso = user_max;
11359 	return(new_tso);
11360 }
11361 
11362 static void
rack_log_hdwr_pacing(struct tcp_rack * rack,const struct ifnet * ifp,uint64_t rate,uint64_t hw_rate,int line,int error)11363 rack_log_hdwr_pacing(struct tcp_rack *rack, const struct ifnet *ifp,
11364 		     uint64_t rate, uint64_t hw_rate, int line,
11365 		     int error)
11366 {
11367 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
11368 		union tcp_log_stackspecific log;
11369 		struct timeval tv;
11370 
11371 		memset(&log, 0, sizeof(log));
11372 		log.u_bbr.flex1 = ((hw_rate >> 32) & 0x00000000ffffffff);
11373 		log.u_bbr.flex2 = (hw_rate & 0x00000000ffffffff);
11374 		log.u_bbr.flex3 = (((uint64_t)ifp  >> 32) & 0x00000000ffffffff);
11375 		log.u_bbr.flex4 = ((uint64_t)ifp & 0x00000000ffffffff);
11376 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
11377 		log.u_bbr.bw_inuse = rate;
11378 		log.u_bbr.flex5 = line;
11379 		log.u_bbr.flex6 = error;
11380 		log.u_bbr.applimited = rack->r_ctl.rc_pace_max_segs;
11381 		log.u_bbr.flex8 = rack->use_fixed_rate;
11382 		log.u_bbr.flex8 <<= 1;
11383 		log.u_bbr.flex8 |= rack->rack_hdrw_pacing;
11384 		log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
11385 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
11386 		    &rack->rc_inp->inp_socket->so_rcv,
11387 		    &rack->rc_inp->inp_socket->so_snd,
11388 		    BBR_LOG_HDWR_PACE, 0,
11389 		    0, &log, false, &tv);
11390 	}
11391 }
11392 
11393 static int32_t
pace_to_fill_cwnd(struct tcp_rack * rack,int32_t slot,uint32_t len,uint32_t segsiz)11394 pace_to_fill_cwnd(struct tcp_rack *rack, int32_t slot, uint32_t len, uint32_t segsiz)
11395 {
11396 	uint64_t lentim, fill_bw;
11397 
11398 	/* Lets first see if we are full, if so continue with normal rate */
11399 	if (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.cwnd_to_use)
11400 		return (slot);
11401 	if ((ctf_outstanding(rack->rc_tp) + (segsiz-1)) > rack->rc_tp->snd_wnd)
11402 		return (slot);
11403 	if (rack->r_ctl.rc_last_us_rtt == 0)
11404 		return (slot);
11405 	if (rack->rc_pace_fill_if_rttin_range &&
11406 	    (rack->r_ctl.rc_last_us_rtt >=
11407 	     (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack->rtt_limit_mul))) {
11408 		/* The rtt is huge, N * smallest, lets not fill */
11409 		return (slot);
11410 	}
11411 	/*
11412 	 * first lets calculate the b/w based on the last us-rtt
11413 	 * and the sndwnd.
11414 	 */
11415 	fill_bw = rack->r_ctl.cwnd_to_use;
11416 	/* Take the rwnd if its smaller */
11417 	if (fill_bw > rack->rc_tp->snd_wnd)
11418 		fill_bw = rack->rc_tp->snd_wnd;
11419 	fill_bw *= (uint64_t)HPTS_USEC_IN_SEC;
11420 	fill_bw /= (uint64_t)rack->r_ctl.rc_last_us_rtt;
11421 	/* We are below the min b/w */
11422 	if (fill_bw < RACK_MIN_BW)
11423 		return (slot);
11424 	/*
11425 	 * Ok fill_bw holds our mythical b/w to fill the cwnd
11426 	 * in a rtt, what does that time wise equate too?
11427 	 */
11428 	lentim = (uint64_t)(len) * (uint64_t)HPTS_USEC_IN_SEC;
11429 	lentim /= fill_bw;
11430 	if (lentim < slot) {
11431 		rack_log_pacing_delay_calc(rack, len, slot, fill_bw,
11432 					   0, lentim, 12, __LINE__, NULL);
11433 		return ((int32_t)lentim);
11434 	} else
11435 		return (slot);
11436 }
11437 
11438 static int32_t
rack_get_pacing_delay(struct tcp_rack * rack,struct tcpcb * tp,uint32_t len,struct rack_sendmap * rsm,uint32_t segsiz)11439 rack_get_pacing_delay(struct tcp_rack *rack, struct tcpcb *tp, uint32_t len, struct rack_sendmap *rsm, uint32_t segsiz)
11440 {
11441 	struct rack_sendmap *lrsm;
11442 	int32_t slot = 0;
11443 	int err;
11444 
11445 	if (rack->rc_always_pace == 0) {
11446 		/*
11447 		 * We use the most optimistic possible cwnd/srtt for
11448 		 * sending calculations. This will make our
11449 		 * calculation anticipate getting more through
11450 		 * quicker then possible. But thats ok we don't want
11451 		 * the peer to have a gap in data sending.
11452 		 */
11453 		uint32_t srtt, cwnd, tr_perms = 0;
11454 		int32_t reduce = 0;
11455 
11456 	old_method:
11457 		/*
11458 		 * We keep no precise pacing with the old method
11459 		 * instead we use the pacer to mitigate bursts.
11460 		 */
11461 		rack->r_ctl.rc_agg_delayed = 0;
11462 		rack->r_early = 0;
11463 		rack->r_late = 0;
11464 		rack->r_ctl.rc_agg_early = 0;
11465 		if (rack->r_ctl.rc_rack_min_rtt)
11466 			srtt = rack->r_ctl.rc_rack_min_rtt;
11467 		else
11468 			srtt = TICKS_2_MSEC((tp->t_srtt >> TCP_RTT_SHIFT));
11469 		if (rack->r_ctl.rc_rack_largest_cwnd)
11470 			cwnd = rack->r_ctl.rc_rack_largest_cwnd;
11471 		else
11472 			cwnd = rack->r_ctl.cwnd_to_use;
11473 		tr_perms = cwnd / srtt;
11474 		if (tr_perms == 0) {
11475 			tr_perms = ctf_fixed_maxseg(tp);
11476 		}
11477 		/*
11478 		 * Calculate how long this will take to drain, if
11479 		 * the calculation comes out to zero, thats ok we
11480 		 * will use send_a_lot to possibly spin around for
11481 		 * more increasing tot_len_this_send to the point
11482 		 * that its going to require a pace, or we hit the
11483 		 * cwnd. Which in that case we are just waiting for
11484 		 * a ACK.
11485 		 */
11486 		slot = len / tr_perms;
11487 		/* Now do we reduce the time so we don't run dry? */
11488 		if (slot && rack_slot_reduction) {
11489 			reduce = (slot / rack_slot_reduction);
11490 			if (reduce < slot) {
11491 				slot -= reduce;
11492 			} else
11493 				slot = 0;
11494 		}
11495 		slot *=  HPTS_USEC_IN_MSEC;
11496 		if (rsm == NULL) {
11497 			/*
11498 			 * We always consider ourselves app limited with old style
11499 			 * that are not retransmits. This could be the initial
11500 			 * measurement, but thats ok its all setup and specially
11501 			 * handled. If another send leaks out, then that too will
11502 			 * be mark app-limited.
11503 			 */
11504 			lrsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
11505 			if (lrsm && ((lrsm->r_flags & RACK_APP_LIMITED) == 0)) {
11506 				rack->r_ctl.rc_first_appl = lrsm;
11507 				lrsm->r_flags |= RACK_APP_LIMITED;
11508 				rack->r_ctl.rc_app_limited_cnt++;
11509 			}
11510 		}
11511 		rack_log_pacing_delay_calc(rack, len, slot, tr_perms, reduce, 0, 7, __LINE__, NULL);
11512 	} else {
11513 		uint64_t bw_est, res, lentim, rate_wanted;
11514 		uint32_t orig_val, srtt, segs, oh;
11515 
11516 		if ((rack->r_rr_config == 1) && rsm) {
11517 			return (rack->r_ctl.rc_min_to * HPTS_USEC_IN_MSEC);
11518 		}
11519 		if (rack->use_fixed_rate) {
11520 			rate_wanted = bw_est = rack_get_fixed_pacing_bw(rack);
11521 		} else if ((rack->r_ctl.init_rate == 0) &&
11522 #ifdef NETFLIX_PEAKRATE
11523 			   (rack->rc_tp->t_maxpeakrate == 0) &&
11524 #endif
11525 			   (rack->r_ctl.gp_bw == 0)) {
11526 			/* no way to yet do an estimate */
11527 			bw_est = rate_wanted = 0;
11528 		} else {
11529 			bw_est = rack_get_bw(rack);
11530 			rate_wanted = rack_get_output_bw(rack, bw_est, rsm);
11531 		}
11532 		if ((bw_est == 0) || (rate_wanted == 0)) {
11533 			/*
11534 			 * No way yet to make a b/w estimate or
11535 			 * our raise is set incorrectly.
11536 			 */
11537 			goto old_method;
11538 		}
11539 		/* We need to account for all the overheads */
11540 		segs = (len + segsiz - 1) / segsiz;
11541 		/*
11542 		 * We need the diff between 1514 bytes (e-mtu with e-hdr)
11543 		 * and how much data we put in each packet. Yes this
11544 		 * means we may be off if we are larger than 1500 bytes
11545 		 * or smaller. But this just makes us more conservative.
11546 		 */
11547 		if (ETHERNET_SEGMENT_SIZE > segsiz)
11548 			oh = ETHERNET_SEGMENT_SIZE - segsiz;
11549 		else
11550 			oh = 0;
11551 		segs *= oh;
11552 		lentim = (uint64_t)(len + segs)  * (uint64_t)HPTS_USEC_IN_SEC;
11553 		res = lentim / rate_wanted;
11554 		slot = (uint32_t)res;
11555 		orig_val = rack->r_ctl.rc_pace_max_segs;
11556 		rack_set_pace_segments(rack->rc_tp, rack, __LINE__);
11557 		/* Did we change the TSO size, if so log it */
11558 		if (rack->r_ctl.rc_pace_max_segs != orig_val)
11559 			rack_log_pacing_delay_calc(rack, len, slot, orig_val, 0, 0, 15, __LINE__, NULL);
11560 		if ((rack->rc_pace_to_cwnd) &&
11561 		    (rack->in_probe_rtt == 0) &&
11562 		    (IN_RECOVERY(rack->rc_tp->t_flags) == 0)) {
11563 			/*
11564 			 * We want to pace at our rate *or* faster to
11565 			 * fill the cwnd to the max if its not full.
11566 			 */
11567 			slot = pace_to_fill_cwnd(rack, slot, (len+segs), segsiz);
11568 		}
11569 		if ((rack->rc_inp->inp_route.ro_nh != NULL) &&
11570 		    (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
11571 			if ((rack->rack_hdw_pace_ena) &&
11572 			    (rack->rack_hdrw_pacing == 0) &&
11573 			    (rack->rack_attempt_hdwr_pace == 0)) {
11574 				/*
11575 				 * Lets attempt to turn on hardware pacing
11576 				 * if we can.
11577 				 */
11578 				rack->rack_attempt_hdwr_pace = 1;
11579 				rack->r_ctl.crte = tcp_set_pacing_rate(rack->rc_tp,
11580 								       rack->rc_inp->inp_route.ro_nh->nh_ifp,
11581 								       rate_wanted,
11582 								       RS_PACING_GEQ,
11583 								       &err);
11584 				if (rack->r_ctl.crte) {
11585 					rack->rack_hdrw_pacing = 1;
11586 					rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(rate_wanted, segsiz,
11587 												 0, rack->r_ctl.crte,
11588 												 NULL);
11589 					rack_log_hdwr_pacing(rack, rack->rc_inp->inp_route.ro_nh->nh_ifp,
11590 							     rate_wanted, rack->r_ctl.crte->rate, __LINE__,
11591 							     err);
11592 				}
11593 			} else if (rack->rack_hdrw_pacing &&
11594 				   (rack->r_ctl.crte->rate != rate_wanted)) {
11595 				/* Do we need to adjust our rate? */
11596 				const struct tcp_hwrate_limit_table *nrte;
11597 
11598 				nrte = tcp_chg_pacing_rate(rack->r_ctl.crte,
11599 							   rack->rc_tp,
11600 							   rack->rc_inp->inp_route.ro_nh->nh_ifp,
11601 							   rate_wanted,
11602 							   RS_PACING_GEQ,
11603 							   &err);
11604 				if (nrte == NULL) {
11605 					/* Lost the rate */
11606 					rack->rack_hdrw_pacing = 0;
11607 					rack_set_pace_segments(rack->rc_tp, rack, __LINE__);
11608 				} else if (nrte != rack->r_ctl.crte) {
11609 					rack->r_ctl.crte = nrte;
11610 					rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(rate_wanted,
11611 												 segsiz, 0,
11612 												 rack->r_ctl.crte,
11613 												 NULL);
11614 					rack_log_hdwr_pacing(rack, rack->rc_inp->inp_route.ro_nh->nh_ifp,
11615 							     rate_wanted, rack->r_ctl.crte->rate, __LINE__,
11616 							     err);
11617 				}
11618 			}
11619 		}
11620 		if (rack_limit_time_with_srtt &&
11621 		    (rack->use_fixed_rate == 0) &&
11622 #ifdef NETFLIX_PEAKRATE
11623 		    (rack->rc_tp->t_maxpeakrate == 0) &&
11624 #endif
11625 		    (rack->rack_hdrw_pacing == 0)) {
11626 			/*
11627 			 * Sanity check, we do not allow the pacing delay
11628 			 * to be longer than the SRTT of the path. If it is
11629 			 * a slow path, then adding a packet should increase
11630 			 * the RTT and compensate for this i.e. the srtt will
11631 			 * be greater so the allowed pacing time will be greater.
11632 			 *
11633 			 * Note this restriction is not for where a peak rate
11634 			 * is set, we are doing fixed pacing or hardware pacing.
11635 			 */
11636 			if (rack->rc_tp->t_srtt)
11637 				srtt = (TICKS_2_USEC(rack->rc_tp->t_srtt) >> TCP_RTT_SHIFT);
11638 			else
11639 				srtt = RACK_INITIAL_RTO * HPTS_USEC_IN_MSEC;	/* its in ms convert */
11640 			if (srtt < slot) {
11641 				rack_log_pacing_delay_calc(rack, srtt, slot, rate_wanted, bw_est, lentim, 99, __LINE__, NULL);
11642 				slot = srtt;
11643 			}
11644 		}
11645 		rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, bw_est, lentim, 2, __LINE__, rsm);
11646 	}
11647 	if (slot)
11648 		counter_u64_add(rack_calc_nonzero, 1);
11649 	else
11650 		counter_u64_add(rack_calc_zero, 1);
11651 	return (slot);
11652 }
11653 
11654 static void
rack_start_gp_measurement(struct tcpcb * tp,struct tcp_rack * rack,tcp_seq startseq,uint32_t sb_offset)11655 rack_start_gp_measurement(struct tcpcb *tp, struct tcp_rack *rack,
11656     tcp_seq startseq, uint32_t sb_offset)
11657 {
11658 	struct rack_sendmap *my_rsm = NULL;
11659 	struct rack_sendmap fe;
11660 
11661 	if (tp->t_state < TCPS_ESTABLISHED) {
11662 		/*
11663 		 * We don't start any measurements if we are
11664 		 * not at least established.
11665 		 */
11666 		return;
11667 	}
11668 	tp->t_flags |= TF_GPUTINPROG;
11669 	rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
11670 	rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
11671 	tp->gput_seq = startseq;
11672 	rack->app_limited_needs_set = 0;
11673 	if (rack->in_probe_rtt)
11674 		rack->measure_saw_probe_rtt = 1;
11675 	else if ((rack->measure_saw_probe_rtt) &&
11676 		 (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
11677 		rack->measure_saw_probe_rtt = 0;
11678 	if (rack->rc_gp_filled)
11679 		tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
11680 	else {
11681 		/* Special case initial measurement */
11682 		rack->r_ctl.rc_gp_output_ts = tp->gput_ts = tcp_get_usecs(NULL);
11683 	}
11684 	/*
11685 	 * We take a guess out into the future,
11686 	 * if we have no measurement and no
11687 	 * initial rate, we measure the first
11688 	 * initial-windows worth of data to
11689 	 * speed up getting some GP measurement and
11690 	 * thus start pacing.
11691 	 */
11692 	if ((rack->rc_gp_filled == 0) && (rack->r_ctl.init_rate == 0)) {
11693 		rack->app_limited_needs_set = 1;
11694 		tp->gput_ack = startseq + max(rc_init_window(rack),
11695 					      (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
11696 		rack_log_pacing_delay_calc(rack,
11697 					   tp->gput_seq,
11698 					   tp->gput_ack,
11699 					   0,
11700 					   tp->gput_ts,
11701 					   rack->r_ctl.rc_app_limited_cnt,
11702 					   9,
11703 					   __LINE__, NULL);
11704 		return;
11705 	}
11706 	if (sb_offset) {
11707 		/*
11708 		 * We are out somewhere in the sb
11709 		 * can we use the already outstanding data?
11710 		 */
11711 
11712 		if (rack->r_ctl.rc_app_limited_cnt == 0) {
11713 			/*
11714 			 * Yes first one is good and in this case
11715 			 * the tp->gput_ts is correctly set based on
11716 			 * the last ack that arrived (no need to
11717 			 * set things up when an ack comes in).
11718 			 */
11719 			my_rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
11720 			if ((my_rsm == NULL) ||
11721 			    (my_rsm->r_rtr_cnt != 1)) {
11722 				/* retransmission? */
11723 				goto use_latest;
11724 			}
11725 		} else {
11726 			if (rack->r_ctl.rc_first_appl == NULL) {
11727 				/*
11728 				 * If rc_first_appl is NULL
11729 				 * then the cnt should be 0.
11730 				 * This is probably an error, maybe
11731 				 * a KASSERT would be approprate.
11732 				 */
11733 				goto use_latest;
11734 			}
11735 			/*
11736 			 * If we have a marker pointer to the last one that is
11737 			 * app limited we can use that, but we need to set
11738 			 * things up so that when it gets ack'ed we record
11739 			 * the ack time (if its not already acked).
11740 			 */
11741 			rack->app_limited_needs_set = 1;
11742 			/*
11743 			 * We want to get to the rsm that is either
11744 			 * next with space i.e. over 1 MSS or the one
11745 			 * after that (after the app-limited).
11746 			 */
11747 			my_rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
11748 					 rack->r_ctl.rc_first_appl);
11749 			if (my_rsm) {
11750 				if ((my_rsm->r_end - my_rsm->r_start) <= ctf_fixed_maxseg(tp))
11751 					/* Have to use the next one */
11752 					my_rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
11753 							 my_rsm);
11754 				else {
11755 					/* Use after the first MSS of it is acked */
11756 					tp->gput_seq = my_rsm->r_start + ctf_fixed_maxseg(tp);
11757 					goto start_set;
11758 				}
11759 			}
11760 			if ((my_rsm == NULL) ||
11761 			    (my_rsm->r_rtr_cnt != 1)) {
11762 				/*
11763 				 * Either its a retransmit or
11764 				 * the last is the app-limited one.
11765 				 */
11766 				goto use_latest;
11767 			}
11768 		}
11769 		tp->gput_seq = my_rsm->r_start;
11770 start_set:
11771 		if (my_rsm->r_flags & RACK_ACKED) {
11772 			/*
11773 			 * This one has been acked use the arrival ack time
11774 			 */
11775 			tp->gput_ts = my_rsm->r_ack_arrival;
11776 			rack->app_limited_needs_set = 0;
11777 		}
11778 		rack->r_ctl.rc_gp_output_ts = my_rsm->usec_orig_send;
11779 		tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
11780 		rack_log_pacing_delay_calc(rack,
11781 					   tp->gput_seq,
11782 					   tp->gput_ack,
11783 					   (uint64_t)my_rsm,
11784 					   tp->gput_ts,
11785 					   rack->r_ctl.rc_app_limited_cnt,
11786 					   9,
11787 					   __LINE__, NULL);
11788 		return;
11789 	}
11790 
11791 use_latest:
11792 	/*
11793 	 * We don't know how long we may have been
11794 	 * idle or if this is the first-send. Lets
11795 	 * setup the flag so we will trim off
11796 	 * the first ack'd data so we get a true
11797 	 * measurement.
11798 	 */
11799 	rack->app_limited_needs_set = 1;
11800 	tp->gput_ack = startseq + rack_get_measure_window(tp, rack);
11801 	/* Find this guy so we can pull the send time */
11802 	fe.r_start = startseq;
11803 	my_rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
11804 	if (my_rsm) {
11805 		rack->r_ctl.rc_gp_output_ts = my_rsm->usec_orig_send;
11806 		if (my_rsm->r_flags & RACK_ACKED) {
11807 			/*
11808 			 * Unlikely since its probably what was
11809 			 * just transmitted (but I am paranoid).
11810 			 */
11811 			tp->gput_ts = my_rsm->r_ack_arrival;
11812 			rack->app_limited_needs_set = 0;
11813 		}
11814 		if (SEQ_LT(my_rsm->r_start, tp->gput_seq)) {
11815 			/* This also is unlikely */
11816 			tp->gput_seq = my_rsm->r_start;
11817 		}
11818 	} else {
11819 		/*
11820 		 * TSNH unless we have some send-map limit,
11821 		 * and even at that it should not be hitting
11822 		 * that limit (we should have stopped sending).
11823 		 */
11824 		rack->r_ctl.rc_gp_output_ts = tcp_get_usecs(NULL);
11825 	}
11826 	rack_log_pacing_delay_calc(rack,
11827 				   tp->gput_seq,
11828 				   tp->gput_ack,
11829 				   (uint64_t)my_rsm,
11830 				   tp->gput_ts,
11831 				   rack->r_ctl.rc_app_limited_cnt,
11832 				   9, __LINE__, NULL);
11833 }
11834 
11835 static inline uint32_t
rack_what_can_we_send(struct tcpcb * tp,struct tcp_rack * rack,uint32_t cwnd_to_use,uint32_t avail,int32_t sb_offset)11836 rack_what_can_we_send(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t cwnd_to_use,
11837     uint32_t avail, int32_t sb_offset)
11838 {
11839 	uint32_t len;
11840 	uint32_t sendwin;
11841 
11842 	if (tp->snd_wnd > cwnd_to_use)
11843 		sendwin = cwnd_to_use;
11844 	else
11845 		sendwin = tp->snd_wnd;
11846 	if (ctf_outstanding(tp) >= tp->snd_wnd) {
11847 		/* We never want to go over our peers rcv-window */
11848 		len = 0;
11849 	} else {
11850 		uint32_t flight;
11851 
11852 		flight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
11853 		if (flight >= sendwin) {
11854 			/*
11855 			 * We have in flight what we are allowed by cwnd (if
11856 			 * it was rwnd blocking it would have hit above out
11857 			 * >= tp->snd_wnd).
11858 			 */
11859 			return (0);
11860 		}
11861 		len = sendwin - flight;
11862 		if ((len + ctf_outstanding(tp)) > tp->snd_wnd) {
11863 			/* We would send too much (beyond the rwnd) */
11864 			len = tp->snd_wnd - ctf_outstanding(tp);
11865 		}
11866 		if ((len + sb_offset) > avail) {
11867 			/*
11868 			 * We don't have that much in the SB, how much is
11869 			 * there?
11870 			 */
11871 			len = avail - sb_offset;
11872 		}
11873 	}
11874 	return (len);
11875 }
11876 
11877 static int
rack_output(struct tcpcb * tp)11878 rack_output(struct tcpcb *tp)
11879 {
11880 	struct socket *so;
11881 	uint32_t recwin;
11882 	uint32_t sb_offset;
11883 	int32_t len, flags, error = 0;
11884 	struct mbuf *m;
11885 	struct mbuf *mb;
11886 	uint32_t if_hw_tsomaxsegcount = 0;
11887 	uint32_t if_hw_tsomaxsegsize;
11888 	int32_t segsiz, minseg;
11889 	long tot_len_this_send = 0;
11890 	struct ip *ip = NULL;
11891 #ifdef TCPDEBUG
11892 	struct ipovly *ipov = NULL;
11893 #endif
11894 	struct udphdr *udp = NULL;
11895 	struct tcp_rack *rack;
11896 	struct tcphdr *th;
11897 	uint8_t pass = 0;
11898 	uint8_t mark = 0;
11899 	uint8_t wanted_cookie = 0;
11900 	u_char opt[TCP_MAXOLEN];
11901 	unsigned ipoptlen, optlen, hdrlen, ulen=0;
11902 	uint32_t rack_seq;
11903 
11904 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
11905 	unsigned ipsec_optlen = 0;
11906 
11907 #endif
11908 	int32_t idle, sendalot;
11909 	int32_t sub_from_prr = 0;
11910 	volatile int32_t sack_rxmit;
11911 	struct rack_sendmap *rsm = NULL;
11912 	int32_t tso, mtu;
11913 	struct tcpopt to;
11914 	int32_t slot = 0;
11915 	int32_t sup_rack = 0;
11916 	uint32_t cts, us_cts, delayed, early;
11917 	uint8_t hpts_calling, new_data_tlp = 0, doing_tlp = 0;
11918 	uint32_t cwnd_to_use;
11919 	int32_t do_a_prefetch;
11920 	int32_t prefetch_rsm = 0;
11921 	int32_t orig_len;
11922 	struct timeval tv;
11923 	int32_t prefetch_so_done = 0;
11924 	struct tcp_log_buffer *lgb = NULL;
11925 	struct inpcb *inp;
11926 	struct sockbuf *sb;
11927 #ifdef INET6
11928 	struct ip6_hdr *ip6 = NULL;
11929 	int32_t isipv6;
11930 #endif
11931 	uint8_t filled_all = 0;
11932 	bool hw_tls = false;
11933 
11934 	/* setup and take the cache hits here */
11935 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11936 	inp = rack->rc_inp;
11937 	so = inp->inp_socket;
11938 	sb = &so->so_snd;
11939 	kern_prefetch(sb, &do_a_prefetch);
11940 	do_a_prefetch = 1;
11941 	hpts_calling = inp->inp_hpts_calls;
11942 	hw_tls = (so->so_snd.sb_flags & SB_TLS_IFNET) != 0;
11943 
11944 	NET_EPOCH_ASSERT();
11945 	INP_WLOCK_ASSERT(inp);
11946 #ifdef TCP_OFFLOAD
11947 	if (tp->t_flags & TF_TOE)
11948 		return (tcp_offload_output(tp));
11949 #endif
11950 	/*
11951 	 * For TFO connections in SYN_RECEIVED, only allow the initial
11952 	 * SYN|ACK and those sent by the retransmit timer.
11953 	 */
11954 	if (IS_FASTOPEN(tp->t_flags) &&
11955 	    (tp->t_state == TCPS_SYN_RECEIVED) &&
11956 	    SEQ_GT(tp->snd_max, tp->snd_una) &&    /* initial SYN|ACK sent */
11957 	    (rack->r_ctl.rc_resend == NULL))         /* not a retransmit */
11958 		return (0);
11959 #ifdef INET6
11960 	if (rack->r_state) {
11961 		/* Use the cache line loaded if possible */
11962 		isipv6 = rack->r_is_v6;
11963 	} else {
11964 		isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
11965 	}
11966 #endif
11967 	early = 0;
11968 	us_cts = tcp_get_usecs(&tv);
11969 	cts = tcp_tv_to_mssectick(&tv);
11970 	if (((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) &&
11971 	    inp->inp_in_hpts) {
11972 		/*
11973 		 * We are on the hpts for some timer but not hptsi output.
11974 		 * Remove from the hpts unconditionally.
11975 		 */
11976 		rack_timer_cancel(tp, rack, cts, __LINE__);
11977 	}
11978 	/* Are we pacing and late? */
11979 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
11980 	    TSTMP_GEQ(us_cts, rack->r_ctl.rc_last_output_to)) {
11981 		/* We are delayed */
11982 		delayed = us_cts - rack->r_ctl.rc_last_output_to;
11983 	} else {
11984 		delayed = 0;
11985 	}
11986 	/* Do the timers, which may override the pacer  */
11987 	if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
11988 		if (rack_process_timers(tp, rack, cts, hpts_calling)) {
11989 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_ATIMER], 1);
11990 			return (0);
11991 		}
11992 	}
11993 	if ((rack->r_timer_override) ||
11994 	    (delayed) ||
11995 	    (tp->t_state < TCPS_ESTABLISHED)) {
11996 		if (tp->t_inpcb->inp_in_hpts)
11997 			tcp_hpts_remove(tp->t_inpcb, HPTS_REMOVE_OUTPUT);
11998 	} else if (tp->t_inpcb->inp_in_hpts) {
11999 		/*
12000 		 * On the hpts you can't pass even if ACKNOW is on, we will
12001 		 * when the hpts fires.
12002 		 */
12003 		counter_u64_add(rack_out_size[TCP_MSS_ACCT_INPACE], 1);
12004 		return (0);
12005 	}
12006 	inp->inp_hpts_calls = 0;
12007 	/* Finish out both pacing early and late accounting */
12008 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
12009 	    TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
12010 		early = rack->r_ctl.rc_last_output_to - us_cts;
12011 	} else
12012 		early = 0;
12013 	if (delayed) {
12014 		rack->r_ctl.rc_agg_delayed += delayed;
12015 		rack->r_late = 1;
12016 	} else if (early) {
12017 		rack->r_ctl.rc_agg_early += early;
12018 		rack->r_early = 1;
12019 	}
12020 	/* Now that early/late accounting is done turn off the flag */
12021 	rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
12022 	rack->r_wanted_output = 0;
12023 	rack->r_timer_override = 0;
12024 	/*
12025 	 * For TFO connections in SYN_SENT or SYN_RECEIVED,
12026 	 * only allow the initial SYN or SYN|ACK and those sent
12027 	 * by the retransmit timer.
12028 	 */
12029 	if (IS_FASTOPEN(tp->t_flags) &&
12030 	    ((tp->t_state == TCPS_SYN_RECEIVED) ||
12031 	     (tp->t_state == TCPS_SYN_SENT)) &&
12032 	    SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
12033 	    (tp->t_rxtshift == 0)) {              /* not a retransmit */
12034 		cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
12035 		goto just_return_nolock;
12036 	}
12037 	/*
12038 	 * Determine length of data that should be transmitted, and flags
12039 	 * that will be used. If there is some data or critical controls
12040 	 * (SYN, RST) to send, then transmit; otherwise, investigate
12041 	 * further.
12042 	 */
12043 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
12044 	if (tp->t_idle_reduce) {
12045 		if (idle && ((ticks - tp->t_rcvtime) >= tp->t_rxtcur))
12046 			rack_cc_after_idle(rack, tp);
12047 	}
12048 	tp->t_flags &= ~TF_LASTIDLE;
12049 	if (idle) {
12050 		if (tp->t_flags & TF_MORETOCOME) {
12051 			tp->t_flags |= TF_LASTIDLE;
12052 			idle = 0;
12053 		}
12054 	}
12055 	if ((tp->snd_una == tp->snd_max) &&
12056 	    rack->r_ctl.rc_went_idle_time &&
12057 	    TSTMP_GT(us_cts, rack->r_ctl.rc_went_idle_time)) {
12058 		idle = us_cts - rack->r_ctl.rc_went_idle_time;
12059 		if (idle > rack_min_probertt_hold) {
12060 			/* Count as a probe rtt */
12061 			if (rack->in_probe_rtt == 0) {
12062 				rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
12063 				rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
12064 				rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
12065 				rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
12066 			} else {
12067 				rack_exit_probertt(rack, us_cts);
12068 			}
12069 		}
12070 		idle = 0;
12071 	}
12072 again:
12073 	/*
12074 	 * If we've recently taken a timeout, snd_max will be greater than
12075 	 * snd_nxt.  There may be SACK information that allows us to avoid
12076 	 * resending already delivered data.  Adjust snd_nxt accordingly.
12077 	 */
12078 	sendalot = 0;
12079 	us_cts = tcp_get_usecs(&tv);
12080 	cts = tcp_tv_to_mssectick(&tv);
12081 	tso = 0;
12082 	mtu = 0;
12083 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
12084 	minseg = segsiz;
12085 	sb_offset = tp->snd_max - tp->snd_una;
12086 	cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
12087 #ifdef NETFLIX_SHARED_CWND
12088 	if ((tp->t_flags2 & TF2_TCP_SCWND_ALLOWED) &&
12089 	    rack->rack_enable_scwnd) {
12090 		/* We are doing cwnd sharing */
12091 		if (rack->rc_gp_filled &&
12092 		    (rack->rack_attempted_scwnd == 0) &&
12093 		    (rack->r_ctl.rc_scw == NULL) &&
12094 		    tp->t_lib) {
12095 			/* The pcbid is in, lets make an attempt */
12096 			counter_u64_add(rack_try_scwnd, 1);
12097 			rack->rack_attempted_scwnd = 1;
12098 			rack->r_ctl.rc_scw = tcp_shared_cwnd_alloc(tp,
12099 								   &rack->r_ctl.rc_scw_index,
12100 								   segsiz);
12101 		}
12102 		if (rack->r_ctl.rc_scw &&
12103 		    (rack->rack_scwnd_is_idle == 1) &&
12104 		    (rack->rc_in_persist == 0) &&
12105 		    sbavail(sb)) {
12106 			/* we are no longer out of data */
12107 			tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
12108 			rack->rack_scwnd_is_idle = 0;
12109 		}
12110 		if (rack->r_ctl.rc_scw) {
12111 			/* First lets update and get the cwnd */
12112 			rack->r_ctl.cwnd_to_use = cwnd_to_use = tcp_shared_cwnd_update(rack->r_ctl.rc_scw,
12113 								    rack->r_ctl.rc_scw_index,
12114 								    tp->snd_cwnd, tp->snd_wnd, segsiz);
12115 		}
12116 	}
12117 #endif
12118 	flags = tcp_outflags[tp->t_state];
12119 	while (rack->rc_free_cnt < rack_free_cache) {
12120 		rsm = rack_alloc(rack);
12121 		if (rsm == NULL) {
12122 			if (inp->inp_hpts_calls)
12123 				/* Retry in a ms */
12124 				slot = (1 * HPTS_USEC_IN_MSEC);
12125 			goto just_return_nolock;
12126 		}
12127 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_tnext);
12128 		rack->rc_free_cnt++;
12129 		rsm = NULL;
12130 	}
12131 	if (inp->inp_hpts_calls)
12132 		inp->inp_hpts_calls = 0;
12133 	sack_rxmit = 0;
12134 	len = 0;
12135 	rsm = NULL;
12136 	if (flags & TH_RST) {
12137 		SOCKBUF_LOCK(sb);
12138 		goto send;
12139 	}
12140 	if (rack->r_ctl.rc_resend) {
12141 		/* Retransmit timer */
12142 		rsm = rack->r_ctl.rc_resend;
12143 		rack->r_ctl.rc_resend = NULL;
12144 		rsm->r_flags &= ~RACK_TLP;
12145 		len = rsm->r_end - rsm->r_start;
12146 		sack_rxmit = 1;
12147 		sendalot = 0;
12148 		KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
12149 			("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
12150 			 __func__, __LINE__,
12151 			 rsm->r_start, tp->snd_una, tp, rack, rsm));
12152 		sb_offset = rsm->r_start - tp->snd_una;
12153 		if (len >= segsiz)
12154 			len = segsiz;
12155 	} else if ((rack->rc_in_persist == 0) &&
12156 		   ((rsm = tcp_rack_output(tp, rack, cts)) != NULL)) {
12157 		/* We have a retransmit that takes precedence */
12158 		rsm->r_flags &= ~RACK_TLP;
12159 		if ((!IN_RECOVERY(tp->t_flags)) &&
12160 		    ((tp->t_flags & (TF_WASFRECOVERY | TF_WASCRECOVERY)) == 0)) {
12161 			/* Enter recovery if not induced by a time-out */
12162 			rack->r_ctl.rc_rsm_start = rsm->r_start;
12163 			rack->r_ctl.rc_cwnd_at = tp->snd_cwnd;
12164 			rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh;
12165 			rack_cong_signal(tp, NULL, CC_NDUPACK);
12166 			/*
12167 			 * When we enter recovery we need to assure we send
12168 			 * one packet.
12169 			 */
12170 			if (rack->rack_no_prr == 0) {
12171 				rack->r_ctl.rc_prr_sndcnt = segsiz;
12172 				rack_log_to_prr(rack, 13, 0);
12173 			}
12174 		}
12175 #ifdef INVARIANTS
12176 		if (SEQ_LT(rsm->r_start, tp->snd_una)) {
12177 			panic("Huh, tp:%p rack:%p rsm:%p start:%u < snd_una:%u\n",
12178 			      tp, rack, rsm, rsm->r_start, tp->snd_una);
12179 		}
12180 #endif
12181 		len = rsm->r_end - rsm->r_start;
12182 		KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
12183 			("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
12184 			 __func__, __LINE__,
12185 			 rsm->r_start, tp->snd_una, tp, rack, rsm));
12186 		sb_offset = rsm->r_start - tp->snd_una;
12187 		/* Can we send it within the PRR boundary? */
12188 		if (rack->rack_no_prr == 0) {
12189 			if ((rack->use_rack_rr == 0) && (len > rack->r_ctl.rc_prr_sndcnt)) {
12190 				/* It does not fit */
12191 				if ((ctf_flight_size(tp, rack->r_ctl.rc_sacked) > len) &&
12192 				    (rack->r_ctl.rc_prr_sndcnt < segsiz)) {
12193 					/*
12194 					 * prr is less than a segment, we
12195 					 * have more acks due in besides
12196 					 * what we need to resend. Lets not send
12197 					 * to avoid sending small pieces of
12198 					 * what we need to retransmit.
12199 					 */
12200 					len = 0;
12201 					goto just_return_nolock;
12202 				}
12203 				len = rack->r_ctl.rc_prr_sndcnt;
12204 			}
12205 		}
12206 		sendalot = 0;
12207 		if (len >= segsiz)
12208 			len = segsiz;
12209 		if (len > 0) {
12210 			sub_from_prr = 1;
12211 			sack_rxmit = 1;
12212 			KMOD_TCPSTAT_INC(tcps_sack_rexmits);
12213 			KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes,
12214 			    min(len, segsiz));
12215 			counter_u64_add(rack_rtm_prr_retran, 1);
12216 		}
12217 	} else 	if (rack->r_ctl.rc_tlpsend) {
12218 		/* Tail loss probe */
12219 		long cwin;
12220 		long tlen;
12221 
12222 		doing_tlp = 1;
12223 		/*
12224 		 * Check if we can do a TLP with a RACK'd packet
12225 		 * this can happen if we are not doing the rack
12226 		 * cheat and we skipped to a TLP and it
12227 		 * went off.
12228 		 */
12229 		rsm = rack->r_ctl.rc_tlpsend;
12230 		rsm->r_flags |= RACK_TLP;
12231 		rack->r_ctl.rc_tlpsend = NULL;
12232 		sack_rxmit = 1;
12233 		tlen = rsm->r_end - rsm->r_start;
12234 		if (tlen > segsiz)
12235 			tlen = segsiz;
12236 		KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
12237 			("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
12238 			 __func__, __LINE__,
12239 			 rsm->r_start, tp->snd_una, tp, rack, rsm));
12240 		sb_offset = rsm->r_start - tp->snd_una;
12241 		cwin = min(tp->snd_wnd, tlen);
12242 		len = cwin;
12243 	}
12244 	/*
12245 	 * Enforce a connection sendmap count limit if set
12246 	 * as long as we are not retransmiting.
12247 	 */
12248 	if ((rsm == NULL) &&
12249 	    (rack->do_detection == 0) &&
12250 	    (V_tcp_map_entries_limit > 0) &&
12251 	    (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
12252 		counter_u64_add(rack_to_alloc_limited, 1);
12253 		if (!rack->alloc_limit_reported) {
12254 			rack->alloc_limit_reported = 1;
12255 			counter_u64_add(rack_alloc_limited_conns, 1);
12256 		}
12257 		goto just_return_nolock;
12258 	}
12259 	if (rsm && (rsm->r_flags & RACK_HAS_FIN)) {
12260 		/* we are retransmitting the fin */
12261 		len--;
12262 		if (len) {
12263 			/*
12264 			 * When retransmitting data do *not* include the
12265 			 * FIN. This could happen from a TLP probe.
12266 			 */
12267 			flags &= ~TH_FIN;
12268 		}
12269 	}
12270 #ifdef INVARIANTS
12271 	/* For debugging */
12272 	rack->r_ctl.rc_rsm_at_retran = rsm;
12273 #endif
12274 	/*
12275 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
12276 	 * state flags.
12277 	 */
12278 	if (tp->t_flags & TF_NEEDFIN)
12279 		flags |= TH_FIN;
12280 	if (tp->t_flags & TF_NEEDSYN)
12281 		flags |= TH_SYN;
12282 	if ((sack_rxmit == 0) && (prefetch_rsm == 0)) {
12283 		void *end_rsm;
12284 		end_rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
12285 		if (end_rsm)
12286 			kern_prefetch(end_rsm, &prefetch_rsm);
12287 		prefetch_rsm = 1;
12288 	}
12289 	SOCKBUF_LOCK(sb);
12290 	/*
12291 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
12292 	 * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a
12293 	 * negative length.  This can also occur when TCP opens up its
12294 	 * congestion window while receiving additional duplicate acks after
12295 	 * fast-retransmit because TCP will reset snd_nxt to snd_max after
12296 	 * the fast-retransmit.
12297 	 *
12298 	 * In the normal retransmit-FIN-only case, however, snd_nxt will be
12299 	 * set to snd_una, the sb_offset will be 0, and the length may wind
12300 	 * up 0.
12301 	 *
12302 	 * If sack_rxmit is true we are retransmitting from the scoreboard
12303 	 * in which case len is already set.
12304 	 */
12305 	if ((sack_rxmit == 0) &&
12306 	    (TCPS_HAVEESTABLISHED(tp->t_state) || IS_FASTOPEN(tp->t_flags))) {
12307 		uint32_t avail;
12308 
12309 		avail = sbavail(sb);
12310 		if (SEQ_GT(tp->snd_nxt, tp->snd_una) && avail)
12311 			sb_offset = tp->snd_nxt - tp->snd_una;
12312 		else
12313 			sb_offset = 0;
12314 		if ((IN_FASTRECOVERY(tp->t_flags) == 0) || rack->rack_no_prr) {
12315 			if (rack->r_ctl.rc_tlp_new_data) {
12316 				/* TLP is forcing out new data */
12317 				if (rack->r_ctl.rc_tlp_new_data > (uint32_t) (avail - sb_offset)) {
12318 					rack->r_ctl.rc_tlp_new_data = (uint32_t) (avail - sb_offset);
12319 				}
12320 				if (rack->r_ctl.rc_tlp_new_data > tp->snd_wnd)
12321 					len = tp->snd_wnd;
12322 				else
12323 					len = rack->r_ctl.rc_tlp_new_data;
12324 				rack->r_ctl.rc_tlp_new_data = 0;
12325 				new_data_tlp = doing_tlp = 1;
12326 			}  else
12327 				len = rack_what_can_we_send(tp, rack, cwnd_to_use, avail, sb_offset);
12328 			if (IN_FASTRECOVERY(tp->t_flags) && (len > segsiz)) {
12329 				/*
12330 				 * For prr=off, we need to send only 1 MSS
12331 				 * at a time. We do this because another sack could
12332 				 * be arriving that causes us to send retransmits and
12333 				 * we don't want to be on a long pace due to a larger send
12334 				 * that keeps us from sending out the retransmit.
12335 				 */
12336 				len = segsiz;
12337 			}
12338 		} else {
12339 			uint32_t outstanding;
12340 
12341 			/*
12342 			 * We are inside of a SACK recovery episode and are
12343 			 * sending new data, having retransmitted all the
12344 			 * data possible so far in the scoreboard.
12345 			 */
12346 			outstanding = tp->snd_max - tp->snd_una;
12347 			if ((rack->r_ctl.rc_prr_sndcnt + outstanding) > tp->snd_wnd) {
12348 				if (tp->snd_wnd > outstanding) {
12349 					len = tp->snd_wnd - outstanding;
12350 					/* Check to see if we have the data */
12351 					if ((sb_offset + len) > avail) {
12352 						/* It does not all fit */
12353 						if (avail > sb_offset)
12354 							len = avail - sb_offset;
12355 						else
12356 							len = 0;
12357 					}
12358 				} else
12359 					len = 0;
12360 			} else if (avail > sb_offset)
12361 				len = avail - sb_offset;
12362 			else
12363 				len = 0;
12364 			if (len > 0) {
12365 				if (len > rack->r_ctl.rc_prr_sndcnt)
12366 					len = rack->r_ctl.rc_prr_sndcnt;
12367 				if (len > 0) {
12368 					sub_from_prr = 1;
12369 					counter_u64_add(rack_rtm_prr_newdata, 1);
12370 				}
12371 			}
12372 			if (len > segsiz) {
12373 				/*
12374 				 * We should never send more than a MSS when
12375 				 * retransmitting or sending new data in prr
12376 				 * mode unless the override flag is on. Most
12377 				 * likely the PRR algorithm is not going to
12378 				 * let us send a lot as well :-)
12379 				 */
12380 				if (rack->r_ctl.rc_prr_sendalot == 0)
12381 					len = segsiz;
12382 			} else if (len < segsiz) {
12383 				/*
12384 				 * Do we send any? The idea here is if the
12385 				 * send empty's the socket buffer we want to
12386 				 * do it. However if not then lets just wait
12387 				 * for our prr_sndcnt to get bigger.
12388 				 */
12389 				long leftinsb;
12390 
12391 				leftinsb = sbavail(sb) - sb_offset;
12392 				if (leftinsb > len) {
12393 					/* This send does not empty the sb */
12394 					len = 0;
12395 				}
12396 			}
12397 		}
12398 	} else if (!TCPS_HAVEESTABLISHED(tp->t_state)) {
12399 		/*
12400 		 * If you have not established
12401 		 * and are not doing FAST OPEN
12402 		 * no data please.
12403 		 */
12404 		if ((sack_rxmit == 0) &&
12405 		    (!IS_FASTOPEN(tp->t_flags))){
12406 			len = 0;
12407 			sb_offset = 0;
12408 		}
12409 	}
12410 	if (prefetch_so_done == 0) {
12411 		kern_prefetch(so, &prefetch_so_done);
12412 		prefetch_so_done = 1;
12413 	}
12414 	/*
12415 	 * Lop off SYN bit if it has already been sent.  However, if this is
12416 	 * SYN-SENT state and if segment contains data and if we don't know
12417 	 * that foreign host supports TAO, suppress sending segment.
12418 	 */
12419 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una) &&
12420 	    ((sack_rxmit == 0) && (tp->t_rxtshift == 0))) {
12421 		/*
12422 		 * When sending additional segments following a TFO SYN|ACK,
12423 		 * do not include the SYN bit.
12424 		 */
12425 		if (IS_FASTOPEN(tp->t_flags) &&
12426 		    (tp->t_state == TCPS_SYN_RECEIVED))
12427 			flags &= ~TH_SYN;
12428 	}
12429 	/*
12430 	 * Be careful not to send data and/or FIN on SYN segments. This
12431 	 * measure is needed to prevent interoperability problems with not
12432 	 * fully conformant TCP implementations.
12433 	 */
12434 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
12435 		len = 0;
12436 		flags &= ~TH_FIN;
12437 	}
12438 	/*
12439 	 * On TFO sockets, ensure no data is sent in the following cases:
12440 	 *
12441 	 *  - When retransmitting SYN|ACK on a passively-created socket
12442 	 *
12443 	 *  - When retransmitting SYN on an actively created socket
12444 	 *
12445 	 *  - When sending a zero-length cookie (cookie request) on an
12446 	 *    actively created socket
12447 	 *
12448 	 *  - When the socket is in the CLOSED state (RST is being sent)
12449 	 */
12450 	if (IS_FASTOPEN(tp->t_flags) &&
12451 	    (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
12452 	     ((tp->t_state == TCPS_SYN_SENT) &&
12453 	      (tp->t_tfo_client_cookie_len == 0)) ||
12454 	     (flags & TH_RST))) {
12455 		sack_rxmit = 0;
12456 		len = 0;
12457 	}
12458 	/* Without fast-open there should never be data sent on a SYN */
12459 	if ((flags & TH_SYN) && (!IS_FASTOPEN(tp->t_flags))) {
12460 		tp->snd_nxt = tp->iss;
12461 		len = 0;
12462 	}
12463 	orig_len = len;
12464 	if (len <= 0) {
12465 		/*
12466 		 * If FIN has been sent but not acked, but we haven't been
12467 		 * called to retransmit, len will be < 0.  Otherwise, window
12468 		 * shrank after we sent into it.  If window shrank to 0,
12469 		 * cancel pending retransmit, pull snd_nxt back to (closed)
12470 		 * window, and set the persist timer if it isn't already
12471 		 * going.  If the window didn't close completely, just wait
12472 		 * for an ACK.
12473 		 *
12474 		 * We also do a general check here to ensure that we will
12475 		 * set the persist timer when we have data to send, but a
12476 		 * 0-byte window. This makes sure the persist timer is set
12477 		 * even if the packet hits one of the "goto send" lines
12478 		 * below.
12479 		 */
12480 		len = 0;
12481 		if ((tp->snd_wnd == 0) &&
12482 		    (TCPS_HAVEESTABLISHED(tp->t_state)) &&
12483 		    (tp->snd_una == tp->snd_max) &&
12484 		    (sb_offset < (int)sbavail(sb))) {
12485 			tp->snd_nxt = tp->snd_una;
12486 			rack_enter_persist(tp, rack, cts);
12487 		}
12488 	} else if ((rsm == NULL) &&
12489 		   ((doing_tlp == 0) || (new_data_tlp == 1)) &&
12490 		   (len < rack->r_ctl.rc_pace_max_segs)) {
12491 		/*
12492 		 * We are not sending a maximum sized segment for
12493 		 * some reason. Should we not send anything (think
12494 		 * sws or persists)?
12495 		 */
12496 		if ((tp->snd_wnd < min(max(segsiz, (rack->r_ctl.rc_high_rwnd/2)), minseg)) &&
12497 		    (TCPS_HAVEESTABLISHED(tp->t_state)) &&
12498 		    (len < minseg) &&
12499 		    (len < (int)(sbavail(sb) - sb_offset))) {
12500 			/*
12501 			 * Here the rwnd is less than
12502 			 * the minimum pacing size, this is not a retransmit,
12503 			 * we are established and
12504 			 * the send is not the last in the socket buffer
12505 			 * we send nothing, and we may enter persists
12506 			 * if nothing is outstanding.
12507 			 */
12508 			len = 0;
12509 			if (tp->snd_max == tp->snd_una) {
12510 				/*
12511 				 * Nothing out we can
12512 				 * go into persists.
12513 				 */
12514 				rack_enter_persist(tp, rack, cts);
12515 				tp->snd_nxt = tp->snd_una;
12516 			}
12517 		} else if ((cwnd_to_use >= max(minseg, (segsiz * 4))) &&
12518 			   (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
12519 			   (len < (int)(sbavail(sb) - sb_offset)) &&
12520 			   (len < minseg)) {
12521 			/*
12522 			 * Here we are not retransmitting, and
12523 			 * the cwnd is not so small that we could
12524 			 * not send at least a min size (rxt timer
12525 			 * not having gone off), We have 2 segments or
12526 			 * more already in flight, its not the tail end
12527 			 * of the socket buffer  and the cwnd is blocking
12528 			 * us from sending out a minimum pacing segment size.
12529 			 * Lets not send anything.
12530 			 */
12531 			len = 0;
12532 		} else if (((tp->snd_wnd - ctf_outstanding(tp)) <
12533 			    min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
12534 			   (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
12535 			   (len < (int)(sbavail(sb) - sb_offset)) &&
12536 			   (TCPS_HAVEESTABLISHED(tp->t_state))) {
12537 			/*
12538 			 * Here we have a send window but we have
12539 			 * filled it up and we can't send another pacing segment.
12540 			 * We also have in flight more than 2 segments
12541 			 * and we are not completing the sb i.e. we allow
12542 			 * the last bytes of the sb to go out even if
12543 			 * its not a full pacing segment.
12544 			 */
12545 			len = 0;
12546 		}
12547 	}
12548 	/* len will be >= 0 after this point. */
12549 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
12550 	tcp_sndbuf_autoscale(tp, so, min(tp->snd_wnd, cwnd_to_use));
12551 	/*
12552 	 * Decide if we can use TCP Segmentation Offloading (if supported by
12553 	 * hardware).
12554 	 *
12555 	 * TSO may only be used if we are in a pure bulk sending state.  The
12556 	 * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP
12557 	 * options prevent using TSO.  With TSO the TCP header is the same
12558 	 * (except for the sequence number) for all generated packets.  This
12559 	 * makes it impossible to transmit any options which vary per
12560 	 * generated segment or packet.
12561 	 *
12562 	 * IPv4 handling has a clear separation of ip options and ip header
12563 	 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
12564 	 * the right thing below to provide length of just ip options and thus
12565 	 * checking for ipoptlen is enough to decide if ip options are present.
12566 	 */
12567 
12568 #ifdef INET6
12569 	if (isipv6)
12570 		ipoptlen = ip6_optlen(tp->t_inpcb);
12571 	else
12572 #endif
12573 		if (tp->t_inpcb->inp_options)
12574 			ipoptlen = tp->t_inpcb->inp_options->m_len -
12575 				offsetof(struct ipoption, ipopt_list);
12576 		else
12577 			ipoptlen = 0;
12578 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
12579 	/*
12580 	 * Pre-calculate here as we save another lookup into the darknesses
12581 	 * of IPsec that way and can actually decide if TSO is ok.
12582 	 */
12583 #ifdef INET6
12584 	if (isipv6 && IPSEC_ENABLED(ipv6))
12585 		ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb);
12586 #ifdef INET
12587 	else
12588 #endif
12589 #endif				/* INET6 */
12590 #ifdef INET
12591 		if (IPSEC_ENABLED(ipv4))
12592 			ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb);
12593 #endif				/* INET */
12594 #endif
12595 
12596 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
12597 	ipoptlen += ipsec_optlen;
12598 #endif
12599 	if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > segsiz &&
12600 	    (tp->t_port == 0) &&
12601 	    ((tp->t_flags & TF_SIGNATURE) == 0) &&
12602 	    tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
12603 	    ipoptlen == 0)
12604 		tso = 1;
12605 	{
12606 		uint32_t outstanding;
12607 
12608 		outstanding = tp->snd_max - tp->snd_una;
12609 		if (tp->t_flags & TF_SENTFIN) {
12610 			/*
12611 			 * If we sent a fin, snd_max is 1 higher than
12612 			 * snd_una
12613 			 */
12614 			outstanding--;
12615 		}
12616 		if (sack_rxmit) {
12617 			if ((rsm->r_flags & RACK_HAS_FIN) == 0)
12618 				flags &= ~TH_FIN;
12619 		} else {
12620 			if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
12621 				   sbused(sb)))
12622 				flags &= ~TH_FIN;
12623 		}
12624 	}
12625 	recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
12626 	    (long)TCP_MAXWIN << tp->rcv_scale);
12627 
12628 	/*
12629 	 * Sender silly window avoidance.   We transmit under the following
12630 	 * conditions when len is non-zero:
12631 	 *
12632 	 * - We have a full segment (or more with TSO) - This is the last
12633 	 * buffer in a write()/send() and we are either idle or running
12634 	 * NODELAY - we've timed out (e.g. persist timer) - we have more
12635 	 * then 1/2 the maximum send window's worth of data (receiver may be
12636 	 * limited the window size) - we need to retransmit
12637 	 */
12638 	if (len) {
12639 		if (len >= segsiz) {
12640 			goto send;
12641 		}
12642 		/*
12643 		 * NOTE! on localhost connections an 'ack' from the remote
12644 		 * end may occur synchronously with the output and cause us
12645 		 * to flush a buffer queued with moretocome.  XXX
12646 		 *
12647 		 */
12648 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
12649 		    (idle || (tp->t_flags & TF_NODELAY)) &&
12650 		    ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
12651 		    (tp->t_flags & TF_NOPUSH) == 0) {
12652 			pass = 2;
12653 			goto send;
12654 		}
12655 		if ((tp->snd_una == tp->snd_max) && len) {	/* Nothing outstanding */
12656 			pass = 22;
12657 			goto send;
12658 		}
12659 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) {
12660 			pass = 4;
12661 			goto send;
12662 		}
12663 		if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {	/* retransmit case */
12664 			pass = 5;
12665 			goto send;
12666 		}
12667 		if (sack_rxmit) {
12668 			pass = 6;
12669 			goto send;
12670 		}
12671 		if (((tp->snd_wnd - ctf_outstanding(tp)) < segsiz) &&
12672 		    (ctf_outstanding(tp) < (segsiz * 2))) {
12673 			/*
12674 			 * We have less than two MSS outstanding (delayed ack)
12675 			 * and our rwnd will not let us send a full sized
12676 			 * MSS. Lets go ahead and let this small segment
12677 			 * out because we want to try to have at least two
12678 			 * packets inflight to not be caught by delayed ack.
12679 			 */
12680 			pass = 12;
12681 			goto send;
12682 		}
12683 	}
12684 	/*
12685 	 * Sending of standalone window updates.
12686 	 *
12687 	 * Window updates are important when we close our window due to a
12688 	 * full socket buffer and are opening it again after the application
12689 	 * reads data from it.  Once the window has opened again and the
12690 	 * remote end starts to send again the ACK clock takes over and
12691 	 * provides the most current window information.
12692 	 *
12693 	 * We must avoid the silly window syndrome whereas every read from
12694 	 * the receive buffer, no matter how small, causes a window update
12695 	 * to be sent.  We also should avoid sending a flurry of window
12696 	 * updates when the socket buffer had queued a lot of data and the
12697 	 * application is doing small reads.
12698 	 *
12699 	 * Prevent a flurry of pointless window updates by only sending an
12700 	 * update when we can increase the advertized window by more than
12701 	 * 1/4th of the socket buffer capacity.  When the buffer is getting
12702 	 * full or is very small be more aggressive and send an update
12703 	 * whenever we can increase by two mss sized segments. In all other
12704 	 * situations the ACK's to new incoming data will carry further
12705 	 * window increases.
12706 	 *
12707 	 * Don't send an independent window update if a delayed ACK is
12708 	 * pending (it will get piggy-backed on it) or the remote side
12709 	 * already has done a half-close and won't send more data.  Skip
12710 	 * this if the connection is in T/TCP half-open state.
12711 	 */
12712 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
12713 	    !(tp->t_flags & TF_DELACK) &&
12714 	    !TCPS_HAVERCVDFIN(tp->t_state)) {
12715 		/*
12716 		 * "adv" is the amount we could increase the window, taking
12717 		 * into account that we are limited by TCP_MAXWIN <<
12718 		 * tp->rcv_scale.
12719 		 */
12720 		int32_t adv;
12721 		int oldwin;
12722 
12723 		adv = recwin;
12724 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
12725 			oldwin = (tp->rcv_adv - tp->rcv_nxt);
12726 			if (adv > oldwin)
12727 			    adv -= oldwin;
12728 			else {
12729 				/* We can't increase the window */
12730 				adv = 0;
12731 			}
12732 		} else
12733 			oldwin = 0;
12734 
12735 		/*
12736 		 * If the new window size ends up being the same as or less
12737 		 * than the old size when it is scaled, then don't force
12738 		 * a window update.
12739 		 */
12740 		if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
12741 			goto dontupdate;
12742 
12743 		if (adv >= (int32_t)(2 * segsiz) &&
12744 		    (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
12745 		     recwin <= (int32_t)(so->so_rcv.sb_hiwat / 8) ||
12746 		     so->so_rcv.sb_hiwat <= 8 * segsiz)) {
12747 			pass = 7;
12748 			goto send;
12749 		}
12750 		if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) {
12751 			pass = 23;
12752 			goto send;
12753 		}
12754 	}
12755 dontupdate:
12756 
12757 	/*
12758 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
12759 	 * is also a catch-all for the retransmit timer timeout case.
12760 	 */
12761 	if (tp->t_flags & TF_ACKNOW) {
12762 		pass = 8;
12763 		goto send;
12764 	}
12765 	if (((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) {
12766 		pass = 9;
12767 		goto send;
12768 	}
12769 	/*
12770 	 * If our state indicates that FIN should be sent and we have not
12771 	 * yet done so, then we need to send.
12772 	 */
12773 	if ((flags & TH_FIN) &&
12774 	    (tp->snd_nxt == tp->snd_una)) {
12775 		pass = 11;
12776 		goto send;
12777 	}
12778 	/*
12779 	 * No reason to send a segment, just return.
12780 	 */
12781 just_return:
12782 	SOCKBUF_UNLOCK(sb);
12783 just_return_nolock:
12784 	{
12785 		int app_limited = CTF_JR_SENT_DATA;
12786 
12787 		if (tot_len_this_send > 0) {
12788 			/* Make sure snd_nxt is up to max */
12789 			if (SEQ_GT(tp->snd_max, tp->snd_nxt))
12790 				tp->snd_nxt = tp->snd_max;
12791 			slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, NULL, segsiz);
12792 		} else {
12793 			int end_window = 0;
12794 			uint32_t seq = tp->gput_ack;
12795 
12796 			rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
12797 			if (rsm) {
12798 				/*
12799 				 * Mark the last sent that we just-returned (hinting
12800 				 * that delayed ack may play a role in any rtt measurement).
12801 				 */
12802 				rsm->r_just_ret = 1;
12803 			}
12804 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_JUSTRET], 1);
12805 			rack->r_ctl.rc_agg_delayed = 0;
12806 			rack->r_early = 0;
12807 			rack->r_late = 0;
12808 			rack->r_ctl.rc_agg_early = 0;
12809 			if ((ctf_outstanding(tp) +
12810 			     min(max(segsiz, (rack->r_ctl.rc_high_rwnd/2)),
12811 				 minseg)) >= tp->snd_wnd) {
12812 				/* We are limited by the rwnd */
12813 				app_limited = CTF_JR_RWND_LIMITED;
12814 			} else if (ctf_outstanding(tp) >= sbavail(sb)) {
12815 				/* We are limited by whats available -- app limited */
12816 				app_limited = CTF_JR_APP_LIMITED;
12817 			} else if ((idle == 0) &&
12818 				   ((tp->t_flags & TF_NODELAY) == 0) &&
12819 				   ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
12820 				   (len < segsiz)) {
12821 				/*
12822 				 * No delay is not on and the
12823 				 * user is sending less than 1MSS. This
12824 				 * brings out SWS avoidance so we
12825 				 * don't send. Another app-limited case.
12826 				 */
12827 				app_limited = CTF_JR_APP_LIMITED;
12828 			} else if (tp->t_flags & TF_NOPUSH) {
12829 				/*
12830 				 * The user has requested no push of
12831 				 * the last segment and we are
12832 				 * at the last segment. Another app
12833 				 * limited case.
12834 				 */
12835 				app_limited = CTF_JR_APP_LIMITED;
12836 			} else if ((ctf_outstanding(tp) + minseg) > cwnd_to_use) {
12837 				/* Its the cwnd */
12838 				app_limited = CTF_JR_CWND_LIMITED;
12839 			} else if (rack->rc_in_persist == 1) {
12840 				/* We are in persists */
12841 				app_limited = CTF_JR_PERSISTS;
12842 			} else if (IN_RECOVERY(tp->t_flags) &&
12843 				   (rack->rack_no_prr == 0) &&
12844 				   (rack->r_ctl.rc_prr_sndcnt < segsiz)) {
12845 				app_limited = CTF_JR_PRR;
12846 			} else {
12847 				/* Now why here are we not sending? */
12848 #ifdef NOW
12849 #ifdef INVARIANTS
12850 				panic("rack:%p hit JR_ASSESSING case cwnd_to_use:%u?", rack, cwnd_to_use);
12851 #endif
12852 #endif
12853 				app_limited = CTF_JR_ASSESSING;
12854 			}
12855 			/*
12856 			 * App limited in some fashion, for our pacing GP
12857 			 * measurements we don't want any gap (even cwnd).
12858 			 * Close  down the measurement window.
12859 			 */
12860 			if (rack_cwnd_block_ends_measure &&
12861 			    ((app_limited == CTF_JR_CWND_LIMITED) ||
12862 			     (app_limited == CTF_JR_PRR))) {
12863 				/*
12864 				 * The reason we are not sending is
12865 				 * the cwnd (or prr). We have been configured
12866 				 * to end the measurement window in
12867 				 * this case.
12868 				 */
12869 				end_window = 1;
12870 			} else if (app_limited == CTF_JR_PERSISTS) {
12871 				/*
12872 				 * We never end the measurement window
12873 				 * in persists, though in theory we
12874 				 * should be only entering after everything
12875 				 * is acknowledged (so we will probably
12876 				 * never come here).
12877 				 */
12878 				end_window = 0;
12879 			} else if (rack_rwnd_block_ends_measure &&
12880 				   (app_limited == CTF_JR_RWND_LIMITED)) {
12881 				/*
12882 				 * We are rwnd limited and have been
12883 				 * configured to end the measurement
12884 				 * window in this case.
12885 				 */
12886 				end_window = 1;
12887 			} else if (app_limited == CTF_JR_APP_LIMITED) {
12888 				/*
12889 				 * A true application limited period, we have
12890 				 * ran out of data.
12891 				 */
12892 				end_window = 1;
12893 			} else if (app_limited == CTF_JR_ASSESSING) {
12894 				/*
12895 				 * In the assessing case we hit the end of
12896 				 * the if/else and had no known reason
12897 				 * This will panic us under invariants..
12898 				 *
12899 				 * If we get this out in logs we need to
12900 				 * investagate which reason we missed.
12901 				 */
12902 				end_window = 1;
12903 			}
12904 			if (end_window) {
12905 				uint8_t log = 0;
12906 
12907 				if ((tp->t_flags & TF_GPUTINPROG) &&
12908 				    SEQ_GT(tp->gput_ack, tp->snd_max)) {
12909 					/* Mark the last packet has app limited */
12910 					tp->gput_ack = tp->snd_max;
12911 					log = 1;
12912 				}
12913 				rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
12914 				if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
12915 					if (rack->r_ctl.rc_app_limited_cnt == 0)
12916 						rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
12917 					else {
12918 						/*
12919 						 * Go out to the end app limited and mark
12920 						 * this new one as next and move the end_appl up
12921 						 * to this guy.
12922 						 */
12923 						if (rack->r_ctl.rc_end_appl)
12924 							rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
12925 						rack->r_ctl.rc_end_appl = rsm;
12926 					}
12927 					rsm->r_flags |= RACK_APP_LIMITED;
12928 					rack->r_ctl.rc_app_limited_cnt++;
12929 				}
12930 				if (log)
12931 					rack_log_pacing_delay_calc(rack,
12932 								   rack->r_ctl.rc_app_limited_cnt, seq,
12933 								   tp->gput_ack, 0, 0, 4, __LINE__, NULL);
12934 			}
12935 		}
12936 		if (slot) {
12937 			/* set the rack tcb into the slot N */
12938 			counter_u64_add(rack_paced_segments, 1);
12939 		} else if (tot_len_this_send) {
12940 			counter_u64_add(rack_unpaced_segments, 1);
12941 		}
12942 		/* Check if we need to go into persists or not */
12943 		if ((rack->rc_in_persist == 0) &&
12944 		    (tp->snd_max == tp->snd_una) &&
12945 		    TCPS_HAVEESTABLISHED(tp->t_state) &&
12946 		    sbavail(sb) &&
12947 		    (sbavail(sb) > tp->snd_wnd) &&
12948 		    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg))) {
12949 			/* Yes lets make sure to move to persist before timer-start */
12950 			rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
12951 		}
12952 		rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, sup_rack);
12953 		rack_log_type_just_return(rack, cts, tot_len_this_send, slot, hpts_calling, app_limited, cwnd_to_use);
12954 	}
12955 #ifdef NETFLIX_SHARED_CWND
12956 	if ((sbavail(sb) == 0) &&
12957 	    rack->r_ctl.rc_scw) {
12958 		tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
12959 		rack->rack_scwnd_is_idle = 1;
12960 	}
12961 #endif
12962 	return (0);
12963 
12964 send:
12965 	if ((flags & TH_FIN) &&
12966 	    sbavail(sb)) {
12967 		/*
12968 		 * We do not transmit a FIN
12969 		 * with data outstanding. We
12970 		 * need to make it so all data
12971 		 * is acked first.
12972 		 */
12973 		flags &= ~TH_FIN;
12974 	}
12975 	/* Enforce stack imposed max seg size if we have one */
12976 	if (rack->r_ctl.rc_pace_max_segs &&
12977 	    (len > rack->r_ctl.rc_pace_max_segs)) {
12978 		mark = 1;
12979 		len = rack->r_ctl.rc_pace_max_segs;
12980 	}
12981 	SOCKBUF_LOCK_ASSERT(sb);
12982 	if (len > 0) {
12983 		if (len >= segsiz)
12984 			tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
12985 		else
12986 			tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
12987 	}
12988 	/*
12989 	 * Before ESTABLISHED, force sending of initial options unless TCP
12990 	 * set not to do any options. NOTE: we assume that the IP/TCP header
12991 	 * plus TCP options always fit in a single mbuf, leaving room for a
12992 	 * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr)
12993 	 * + optlen <= MCLBYTES
12994 	 */
12995 	optlen = 0;
12996 #ifdef INET6
12997 	if (isipv6)
12998 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
12999 	else
13000 #endif
13001 		hdrlen = sizeof(struct tcpiphdr);
13002 
13003 	/*
13004 	 * Compute options for segment. We only have to care about SYN and
13005 	 * established connection segments.  Options for SYN-ACK segments
13006 	 * are handled in TCP syncache.
13007 	 */
13008 	to.to_flags = 0;
13009 	if ((tp->t_flags & TF_NOOPT) == 0) {
13010 		/* Maximum segment size. */
13011 		if (flags & TH_SYN) {
13012 			tp->snd_nxt = tp->iss;
13013 			to.to_mss = tcp_mssopt(&inp->inp_inc);
13014 #ifdef NETFLIX_TCPOUDP
13015 			if (tp->t_port)
13016 				to.to_mss -= V_tcp_udp_tunneling_overhead;
13017 #endif
13018 			to.to_flags |= TOF_MSS;
13019 
13020 			/*
13021 			 * On SYN or SYN|ACK transmits on TFO connections,
13022 			 * only include the TFO option if it is not a
13023 			 * retransmit, as the presence of the TFO option may
13024 			 * have caused the original SYN or SYN|ACK to have
13025 			 * been dropped by a middlebox.
13026 			 */
13027 			if (IS_FASTOPEN(tp->t_flags) &&
13028 			    (tp->t_rxtshift == 0)) {
13029 				if (tp->t_state == TCPS_SYN_RECEIVED) {
13030 					to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
13031 					to.to_tfo_cookie =
13032 						(u_int8_t *)&tp->t_tfo_cookie.server;
13033 					to.to_flags |= TOF_FASTOPEN;
13034 					wanted_cookie = 1;
13035 				} else if (tp->t_state == TCPS_SYN_SENT) {
13036 					to.to_tfo_len =
13037 						tp->t_tfo_client_cookie_len;
13038 					to.to_tfo_cookie =
13039 						tp->t_tfo_cookie.client;
13040 					to.to_flags |= TOF_FASTOPEN;
13041 					wanted_cookie = 1;
13042 					/*
13043 					 * If we wind up having more data to
13044 					 * send with the SYN than can fit in
13045 					 * one segment, don't send any more
13046 					 * until the SYN|ACK comes back from
13047 					 * the other end.
13048 					 */
13049 					sendalot = 0;
13050 				}
13051 			}
13052 		}
13053 		/* Window scaling. */
13054 		if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
13055 			to.to_wscale = tp->request_r_scale;
13056 			to.to_flags |= TOF_SCALE;
13057 		}
13058 		/* Timestamps. */
13059 		if ((tp->t_flags & TF_RCVD_TSTMP) ||
13060 		    ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
13061 			to.to_tsval = cts + tp->ts_offset;
13062 			to.to_tsecr = tp->ts_recent;
13063 			to.to_flags |= TOF_TS;
13064 		}
13065 		/* Set receive buffer autosizing timestamp. */
13066 		if (tp->rfbuf_ts == 0 &&
13067 		    (so->so_rcv.sb_flags & SB_AUTOSIZE))
13068 			tp->rfbuf_ts = tcp_ts_getticks();
13069 		/* Selective ACK's. */
13070 		if (flags & TH_SYN)
13071 			to.to_flags |= TOF_SACKPERM;
13072 		else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
13073 			 tp->rcv_numsacks > 0) {
13074 			to.to_flags |= TOF_SACK;
13075 			to.to_nsacks = tp->rcv_numsacks;
13076 			to.to_sacks = (u_char *)tp->sackblks;
13077 		}
13078 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
13079 		/* TCP-MD5 (RFC2385). */
13080 		if (tp->t_flags & TF_SIGNATURE)
13081 			to.to_flags |= TOF_SIGNATURE;
13082 #endif				/* TCP_SIGNATURE */
13083 
13084 		/* Processing the options. */
13085 		hdrlen += optlen = tcp_addoptions(&to, opt);
13086 		/*
13087 		 * If we wanted a TFO option to be added, but it was unable
13088 		 * to fit, ensure no data is sent.
13089 		 */
13090 		if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
13091 		    !(to.to_flags & TOF_FASTOPEN))
13092 			len = 0;
13093 	}
13094 #ifdef NETFLIX_TCPOUDP
13095 	if (tp->t_port) {
13096 		if (V_tcp_udp_tunneling_port == 0) {
13097 			/* The port was removed?? */
13098 			SOCKBUF_UNLOCK(&so->so_snd);
13099 			return (EHOSTUNREACH);
13100 		}
13101 		hdrlen += sizeof(struct udphdr);
13102 	}
13103 #endif
13104 #ifdef INET6
13105 	if (isipv6)
13106 		ipoptlen = ip6_optlen(tp->t_inpcb);
13107 	else
13108 #endif
13109 		if (tp->t_inpcb->inp_options)
13110 			ipoptlen = tp->t_inpcb->inp_options->m_len -
13111 				offsetof(struct ipoption, ipopt_list);
13112 		else
13113 			ipoptlen = 0;
13114 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
13115 	ipoptlen += ipsec_optlen;
13116 #endif
13117 
13118 	/*
13119 	 * Adjust data length if insertion of options will bump the packet
13120 	 * length beyond the t_maxseg length. Clear the FIN bit because we
13121 	 * cut off the tail of the segment.
13122 	 */
13123 	if (len + optlen + ipoptlen > tp->t_maxseg) {
13124 		if (tso) {
13125 			uint32_t if_hw_tsomax;
13126 			uint32_t moff;
13127 			int32_t max_len;
13128 
13129 			/* extract TSO information */
13130 			if_hw_tsomax = tp->t_tsomax;
13131 			if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
13132 			if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
13133 			KASSERT(ipoptlen == 0,
13134 				("%s: TSO can't do IP options", __func__));
13135 
13136 			/*
13137 			 * Check if we should limit by maximum payload
13138 			 * length:
13139 			 */
13140 			if (if_hw_tsomax != 0) {
13141 				/* compute maximum TSO length */
13142 				max_len = (if_hw_tsomax - hdrlen -
13143 					   max_linkhdr);
13144 				if (max_len <= 0) {
13145 					len = 0;
13146 				} else if (len > max_len) {
13147 					sendalot = 1;
13148 					len = max_len;
13149 					mark = 2;
13150 				}
13151 			}
13152 			/*
13153 			 * Prevent the last segment from being fractional
13154 			 * unless the send sockbuf can be emptied:
13155 			 */
13156 			max_len = (tp->t_maxseg - optlen);
13157 			if ((sb_offset + len) < sbavail(sb)) {
13158 				moff = len % (u_int)max_len;
13159 				if (moff != 0) {
13160 					mark = 3;
13161 					len -= moff;
13162 				}
13163 			}
13164                         /*
13165 			 * In case there are too many small fragments don't
13166 			 * use TSO:
13167 			 */
13168 			if (len <= segsiz) {
13169 				mark = 4;
13170 				tso = 0;
13171 			}
13172 			/*
13173 			 * Send the FIN in a separate segment after the bulk
13174 			 * sending is done. We don't trust the TSO
13175 			 * implementations to clear the FIN flag on all but
13176 			 * the last segment.
13177 			 */
13178 			if (tp->t_flags & TF_NEEDFIN) {
13179 				sendalot = 4;
13180 			}
13181 		} else {
13182 			mark = 5;
13183 			if (optlen + ipoptlen >= tp->t_maxseg) {
13184 				/*
13185 				 * Since we don't have enough space to put
13186 				 * the IP header chain and the TCP header in
13187 				 * one packet as required by RFC 7112, don't
13188 				 * send it. Also ensure that at least one
13189 				 * byte of the payload can be put into the
13190 				 * TCP segment.
13191 				 */
13192 				SOCKBUF_UNLOCK(&so->so_snd);
13193 				error = EMSGSIZE;
13194 				sack_rxmit = 0;
13195 				goto out;
13196 			}
13197 			len = tp->t_maxseg - optlen - ipoptlen;
13198 			sendalot = 5;
13199 		}
13200 	} else {
13201 		tso = 0;
13202 		mark = 6;
13203 	}
13204 	KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
13205 		("%s: len > IP_MAXPACKET", __func__));
13206 #ifdef DIAGNOSTIC
13207 #ifdef INET6
13208 	if (max_linkhdr + hdrlen > MCLBYTES)
13209 #else
13210 		if (max_linkhdr + hdrlen > MHLEN)
13211 #endif
13212 			panic("tcphdr too big");
13213 #endif
13214 
13215 	/*
13216 	 * This KASSERT is here to catch edge cases at a well defined place.
13217 	 * Before, those had triggered (random) panic conditions further
13218 	 * down.
13219 	 */
13220 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
13221 	if ((len == 0) &&
13222 	    (flags & TH_FIN) &&
13223 	    (sbused(sb))) {
13224 		/*
13225 		 * We have outstanding data, don't send a fin by itself!.
13226 		 */
13227 		goto just_return;
13228 	}
13229 	/*
13230 	 * Grab a header mbuf, attaching a copy of data to be transmitted,
13231 	 * and initialize the header from the template for sends on this
13232 	 * connection.
13233 	 */
13234 	if (len) {
13235 		uint32_t max_val;
13236 		uint32_t moff;
13237 
13238 		if (rack->r_ctl.rc_pace_max_segs)
13239 			max_val = rack->r_ctl.rc_pace_max_segs;
13240 		else if (rack->rc_user_set_max_segs)
13241 			max_val = rack->rc_user_set_max_segs * segsiz;
13242 		else
13243 			max_val = len;
13244 		/*
13245 		 * We allow a limit on sending with hptsi.
13246 		 */
13247 		if (len > max_val) {
13248 			mark = 7;
13249 			len = max_val;
13250 		}
13251 #ifdef INET6
13252 		if (MHLEN < hdrlen + max_linkhdr)
13253 			m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
13254 		else
13255 #endif
13256 			m = m_gethdr(M_NOWAIT, MT_DATA);
13257 
13258 		if (m == NULL) {
13259 			SOCKBUF_UNLOCK(sb);
13260 			error = ENOBUFS;
13261 			sack_rxmit = 0;
13262 			goto out;
13263 		}
13264 		m->m_data += max_linkhdr;
13265 		m->m_len = hdrlen;
13266 
13267 		/*
13268 		 * Start the m_copy functions from the closest mbuf to the
13269 		 * sb_offset in the socket buffer chain.
13270 		 */
13271 		mb = sbsndptr_noadv(sb, sb_offset, &moff);
13272 		if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
13273 			m_copydata(mb, moff, (int)len,
13274 				   mtod(m, caddr_t)+hdrlen);
13275 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
13276 				sbsndptr_adv(sb, mb, len);
13277 			m->m_len += len;
13278 		} else {
13279 			struct sockbuf *msb;
13280 
13281 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
13282 				msb = NULL;
13283 			else
13284 				msb = sb;
13285 			m->m_next = tcp_m_copym(
13286 				mb, moff, &len,
13287 				if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, msb,
13288 				((rsm == NULL) ? hw_tls : 0)
13289 #ifdef NETFLIX_COPY_ARGS
13290 				, &filled_all
13291 #endif
13292 				);
13293 			if (len <= (tp->t_maxseg - optlen)) {
13294 				/*
13295 				 * Must have ran out of mbufs for the copy
13296 				 * shorten it to no longer need tso. Lets
13297 				 * not put on sendalot since we are low on
13298 				 * mbufs.
13299 				 */
13300 				tso = 0;
13301 			}
13302 			if (m->m_next == NULL) {
13303 				SOCKBUF_UNLOCK(sb);
13304 				(void)m_free(m);
13305 				error = ENOBUFS;
13306 				sack_rxmit = 0;
13307 				goto out;
13308 			}
13309 		}
13310 		if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
13311 			if (rsm && (rsm->r_flags & RACK_TLP)) {
13312 				/*
13313 				 * TLP should not count in retran count, but
13314 				 * in its own bin
13315 				 */
13316 				counter_u64_add(rack_tlp_retran, 1);
13317 				counter_u64_add(rack_tlp_retran_bytes, len);
13318 			} else {
13319 				tp->t_sndrexmitpack++;
13320 				KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
13321 				KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
13322 			}
13323 #ifdef STATS
13324 			stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
13325 						 len);
13326 #endif
13327 		} else {
13328 			KMOD_TCPSTAT_INC(tcps_sndpack);
13329 			KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
13330 #ifdef STATS
13331 			stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
13332 						 len);
13333 #endif
13334 		}
13335 		/*
13336 		 * If we're sending everything we've got, set PUSH. (This
13337 		 * will keep happy those implementations which only give
13338 		 * data to the user when a buffer fills or a PUSH comes in.)
13339 		 */
13340 		if (sb_offset + len == sbused(sb) &&
13341 		    sbused(sb) &&
13342 		    !(flags & TH_SYN))
13343 			flags |= TH_PUSH;
13344 
13345 		SOCKBUF_UNLOCK(sb);
13346 	} else {
13347 		SOCKBUF_UNLOCK(sb);
13348 		if (tp->t_flags & TF_ACKNOW)
13349 			KMOD_TCPSTAT_INC(tcps_sndacks);
13350 		else if (flags & (TH_SYN | TH_FIN | TH_RST))
13351 			KMOD_TCPSTAT_INC(tcps_sndctrl);
13352 		else
13353 			KMOD_TCPSTAT_INC(tcps_sndwinup);
13354 
13355 		m = m_gethdr(M_NOWAIT, MT_DATA);
13356 		if (m == NULL) {
13357 			error = ENOBUFS;
13358 			sack_rxmit = 0;
13359 			goto out;
13360 		}
13361 #ifdef INET6
13362 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
13363 		    MHLEN >= hdrlen) {
13364 			M_ALIGN(m, hdrlen);
13365 		} else
13366 #endif
13367 			m->m_data += max_linkhdr;
13368 		m->m_len = hdrlen;
13369 	}
13370 	SOCKBUF_UNLOCK_ASSERT(sb);
13371 	m->m_pkthdr.rcvif = (struct ifnet *)0;
13372 #ifdef MAC
13373 	mac_inpcb_create_mbuf(inp, m);
13374 #endif
13375 #ifdef INET6
13376 	if (isipv6) {
13377 		ip6 = mtod(m, struct ip6_hdr *);
13378 #ifdef NETFLIX_TCPOUDP
13379 		if (tp->t_port) {
13380 			udp = (struct udphdr *)((caddr_t)ip6 + ipoptlen + sizeof(struct ip6_hdr));
13381 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
13382 			udp->uh_dport = tp->t_port;
13383 			ulen = hdrlen + len - sizeof(struct ip6_hdr);
13384 			udp->uh_ulen = htons(ulen);
13385 			th = (struct tcphdr *)(udp + 1);
13386 		} else
13387 #endif
13388 			th = (struct tcphdr *)(ip6 + 1);
13389 		tcpip_fillheaders(inp,
13390 #ifdef NETFLIX_TCPOUDP
13391 				  tp->t_port,
13392 #endif
13393 				  ip6, th);
13394 	} else
13395 #endif				/* INET6 */
13396 	{
13397 		ip = mtod(m, struct ip *);
13398 #ifdef TCPDEBUG
13399 		ipov = (struct ipovly *)ip;
13400 #endif
13401 #ifdef NETFLIX_TCPOUDP
13402 		if (tp->t_port) {
13403 			udp = (struct udphdr *)((caddr_t)ip + ipoptlen + sizeof(struct ip));
13404 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
13405 			udp->uh_dport = tp->t_port;
13406 			ulen = hdrlen + len - sizeof(struct ip);
13407 			udp->uh_ulen = htons(ulen);
13408 			th = (struct tcphdr *)(udp + 1);
13409 		} else
13410 #endif
13411 			th = (struct tcphdr *)(ip + 1);
13412 		tcpip_fillheaders(inp,
13413 #ifdef NETFLIX_TCPOUDP
13414 				  tp->t_port,
13415 #endif
13416 				  ip, th);
13417 	}
13418 	/*
13419 	 * Fill in fields, remembering maximum advertised window for use in
13420 	 * delaying messages about window sizes. If resending a FIN, be sure
13421 	 * not to use a new sequence number.
13422 	 */
13423 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
13424 	    tp->snd_nxt == tp->snd_max)
13425 		tp->snd_nxt--;
13426 	/*
13427 	 * If we are starting a connection, send ECN setup SYN packet. If we
13428 	 * are on a retransmit, we may resend those bits a number of times
13429 	 * as per RFC 3168.
13430 	 */
13431 	if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn == 1) {
13432 		if (tp->t_rxtshift >= 1) {
13433 			if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
13434 				flags |= TH_ECE | TH_CWR;
13435 		} else
13436 			flags |= TH_ECE | TH_CWR;
13437 	}
13438 	/* Handle parallel SYN for ECN */
13439 	if ((tp->t_state == TCPS_SYN_RECEIVED) &&
13440 	    (tp->t_flags2 & TF2_ECN_SND_ECE)) {
13441 		flags |= TH_ECE;
13442 		tp->t_flags2 &= ~TF2_ECN_SND_ECE;
13443 	}
13444 	if (tp->t_state == TCPS_ESTABLISHED &&
13445 	    (tp->t_flags2 & TF2_ECN_PERMIT)) {
13446 		/*
13447 		 * If the peer has ECN, mark data packets with ECN capable
13448 		 * transmission (ECT). Ignore pure ack packets,
13449 		 * retransmissions.
13450 		 */
13451 		if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
13452 		    (sack_rxmit == 0)) {
13453 #ifdef INET6
13454 			if (isipv6)
13455 				ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
13456 			else
13457 #endif
13458 				ip->ip_tos |= IPTOS_ECN_ECT0;
13459 			KMOD_TCPSTAT_INC(tcps_ecn_ect0);
13460 			/*
13461 			 * Reply with proper ECN notifications.
13462 			 * Only set CWR on new data segments.
13463 			 */
13464 			if (tp->t_flags2 & TF2_ECN_SND_CWR) {
13465 				flags |= TH_CWR;
13466 				tp->t_flags2 &= ~TF2_ECN_SND_CWR;
13467 			}
13468 		}
13469 		if (tp->t_flags2 & TF2_ECN_SND_ECE)
13470 			flags |= TH_ECE;
13471 	}
13472 	/*
13473 	 * If we are doing retransmissions, then snd_nxt will not reflect
13474 	 * the first unsent octet.  For ACK only packets, we do not want the
13475 	 * sequence number of the retransmitted packet, we want the sequence
13476 	 * number of the next unsent octet.  So, if there is no data (and no
13477 	 * SYN or FIN), use snd_max instead of snd_nxt when filling in
13478 	 * ti_seq.  But if we are in persist state, snd_max might reflect
13479 	 * one byte beyond the right edge of the window, so use snd_nxt in
13480 	 * that case, since we know we aren't doing a retransmission.
13481 	 * (retransmit and persist are mutually exclusive...)
13482 	 */
13483 	if (sack_rxmit == 0) {
13484 		if (len || (flags & (TH_SYN | TH_FIN)) ||
13485 		    rack->rc_in_persist) {
13486 			th->th_seq = htonl(tp->snd_nxt);
13487 			rack_seq = tp->snd_nxt;
13488 		} else if (flags & TH_RST) {
13489 			/*
13490 			 * For a Reset send the last cum ack in sequence
13491 			 * (this like any other choice may still generate a
13492 			 * challenge ack, if a ack-update packet is in
13493 			 * flight).
13494 			 */
13495 			th->th_seq = htonl(tp->snd_una);
13496 			rack_seq = tp->snd_una;
13497 		} else {
13498 			th->th_seq = htonl(tp->snd_max);
13499 			rack_seq = tp->snd_max;
13500 		}
13501 	} else {
13502 		th->th_seq = htonl(rsm->r_start);
13503 		rack_seq = rsm->r_start;
13504 	}
13505 	th->th_ack = htonl(tp->rcv_nxt);
13506 	if (optlen) {
13507 		bcopy(opt, th + 1, optlen);
13508 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
13509 	}
13510 	th->th_flags = flags;
13511 	/*
13512 	 * Calculate receive window.  Don't shrink window, but avoid silly
13513 	 * window syndrome.
13514 	 * If a RST segment is sent, advertise a window of zero.
13515 	 */
13516 	if (flags & TH_RST) {
13517 		recwin = 0;
13518 	} else {
13519 		if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
13520 		    recwin < (long)segsiz)
13521 			recwin = 0;
13522 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
13523 		    recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
13524 			recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
13525 	}
13526 
13527 	/*
13528 	 * According to RFC1323 the window field in a SYN (i.e., a <SYN> or
13529 	 * <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK> case is
13530 	 * handled in syncache.
13531 	 */
13532 	if (flags & TH_SYN)
13533 		th->th_win = htons((u_short)
13534 				   (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
13535 	else {
13536 		/* Avoid shrinking window with window scaling. */
13537 		recwin = roundup2(recwin, 1 << tp->rcv_scale);
13538 		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
13539 	}
13540 	/*
13541 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0
13542 	 * window.  This may cause the remote transmitter to stall.  This
13543 	 * flag tells soreceive() to disable delayed acknowledgements when
13544 	 * draining the buffer.  This can occur if the receiver is
13545 	 * attempting to read more data than can be buffered prior to
13546 	 * transmitting on the connection.
13547 	 */
13548 	if (th->th_win == 0) {
13549 		tp->t_sndzerowin++;
13550 		tp->t_flags |= TF_RXWIN0SENT;
13551 	} else
13552 		tp->t_flags &= ~TF_RXWIN0SENT;
13553 	tp->snd_up = tp->snd_una;	/* drag it along, its deprecated  */
13554 
13555 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
13556 	if (to.to_flags & TOF_SIGNATURE) {
13557 		/*
13558 		 * Calculate MD5 signature and put it into the place
13559 		 * determined before.
13560 		 * NOTE: since TCP options buffer doesn't point into
13561 		 * mbuf's data, calculate offset and use it.
13562 		 */
13563 		if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
13564 						       (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
13565 			/*
13566 			 * Do not send segment if the calculation of MD5
13567 			 * digest has failed.
13568 			 */
13569 			goto out;
13570 		}
13571 	}
13572 #endif
13573 
13574 	/*
13575 	 * Put TCP length in extended header, and then checksum extended
13576 	 * header and data.
13577 	 */
13578 	m->m_pkthdr.len = hdrlen + len;	/* in6_cksum() need this */
13579 #ifdef INET6
13580 	if (isipv6) {
13581 		/*
13582 		 * ip6_plen is not need to be filled now, and will be filled
13583 		 * in ip6_output.
13584 		 */
13585 		if (tp->t_port) {
13586 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
13587 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
13588 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
13589 			th->th_sum = htons(0);
13590 			UDPSTAT_INC(udps_opackets);
13591 		} else {
13592 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
13593 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
13594 			th->th_sum = in6_cksum_pseudo(ip6,
13595 						      sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
13596 						      0);
13597 		}
13598 	}
13599 #endif
13600 #if defined(INET6) && defined(INET)
13601 	else
13602 #endif
13603 #ifdef INET
13604 	{
13605 		if (tp->t_port) {
13606 			m->m_pkthdr.csum_flags = CSUM_UDP;
13607 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
13608 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
13609 						ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
13610 			th->th_sum = htons(0);
13611 			UDPSTAT_INC(udps_opackets);
13612 		} else {
13613 			m->m_pkthdr.csum_flags = CSUM_TCP;
13614 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
13615 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
13616 					       ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
13617 									IPPROTO_TCP + len + optlen));
13618 		}
13619 		/* IP version must be set here for ipv4/ipv6 checking later */
13620 		KASSERT(ip->ip_v == IPVERSION,
13621 			("%s: IP version incorrect: %d", __func__, ip->ip_v));
13622 	}
13623 #endif
13624 	/*
13625 	 * Enable TSO and specify the size of the segments. The TCP pseudo
13626 	 * header checksum is always provided. XXX: Fixme: This is currently
13627 	 * not the case for IPv6.
13628 	 */
13629 	if (tso) {
13630 		KASSERT(len > tp->t_maxseg - optlen,
13631 			("%s: len <= tso_segsz", __func__));
13632 		m->m_pkthdr.csum_flags |= CSUM_TSO;
13633 		m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
13634 	}
13635 	KASSERT(len + hdrlen == m_length(m, NULL),
13636 		("%s: mbuf chain different than expected: %d + %u != %u",
13637 		 __func__, len, hdrlen, m_length(m, NULL)));
13638 
13639 #ifdef TCP_HHOOK
13640 	/* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
13641 	hhook_run_tcp_est_out(tp, th, &to, len, tso);
13642 #endif
13643 #ifdef TCPDEBUG
13644 	/*
13645 	 * Trace.
13646 	 */
13647 	if (so->so_options & SO_DEBUG) {
13648 		u_short save = 0;
13649 
13650 #ifdef INET6
13651 		if (!isipv6)
13652 #endif
13653 		{
13654 			save = ipov->ih_len;
13655 			ipov->ih_len = htons(m->m_pkthdr.len	/* - hdrlen +
13656 								 * (th->th_off << 2) */ );
13657 		}
13658 		tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
13659 #ifdef INET6
13660 		if (!isipv6)
13661 #endif
13662 			ipov->ih_len = save;
13663 	}
13664 #endif				/* TCPDEBUG */
13665 
13666 	/* We're getting ready to send; log now. */
13667 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
13668 		union tcp_log_stackspecific log;
13669 		struct timeval tv;
13670 
13671 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
13672 		log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
13673 		log.u_bbr.ininput = rack->rc_inp->inp_in_input;
13674 		if (rack->rack_no_prr)
13675 			log.u_bbr.flex1 = 0;
13676 		else
13677 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
13678 		log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
13679 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
13680 		log.u_bbr.flex4 = orig_len;
13681 		if (filled_all)
13682 			log.u_bbr.flex5 = 0x80000000;
13683 		else
13684 			log.u_bbr.flex5 = 0;
13685 		/* Save off the early/late values */
13686 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
13687 		log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
13688 		log.u_bbr.bw_inuse = rack_get_bw(rack);
13689 		if (rsm || sack_rxmit) {
13690 			if (doing_tlp)
13691 				log.u_bbr.flex8 = 2;
13692 			else
13693 				log.u_bbr.flex8 = 1;
13694 		} else {
13695 			log.u_bbr.flex8 = 0;
13696 		}
13697 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
13698 		log.u_bbr.flex7 = mark;
13699 		log.u_bbr.pkts_out = tp->t_maxseg;
13700 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
13701 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
13702 		log.u_bbr.lt_epoch = cwnd_to_use;
13703 		log.u_bbr.delivered = sendalot;
13704 		lgb = tcp_log_event_(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
13705 				     len, &log, false, NULL, NULL, 0, &tv);
13706 	} else
13707 		lgb = NULL;
13708 
13709 	/*
13710 	 * Fill in IP length and desired time to live and send to IP level.
13711 	 * There should be a better way to handle ttl and tos; we could keep
13712 	 * them in the template, but need a way to checksum without them.
13713 	 */
13714 	/*
13715 	 * m->m_pkthdr.len should have been set before cksum calcuration,
13716 	 * because in6_cksum() need it.
13717 	 */
13718 #ifdef INET6
13719 	if (isipv6) {
13720 		/*
13721 		 * we separately set hoplimit for every segment, since the
13722 		 * user might want to change the value via setsockopt. Also,
13723 		 * desired default hop limit might be changed via Neighbor
13724 		 * Discovery.
13725 		 */
13726 		ip6->ip6_hlim = in6_selecthlim(inp, NULL);
13727 
13728 		/*
13729 		 * Set the packet size here for the benefit of DTrace
13730 		 * probes. ip6_output() will set it properly; it's supposed
13731 		 * to include the option header lengths as well.
13732 		 */
13733 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
13734 
13735 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
13736 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
13737 		else
13738 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
13739 
13740 		if (tp->t_state == TCPS_SYN_SENT)
13741 			TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
13742 
13743 		TCP_PROBE5(send, NULL, tp, ip6, tp, th);
13744 		/* TODO: IPv6 IP6TOS_ECT bit on */
13745 		error = ip6_output(m, inp->in6p_outputopts,
13746 				   &inp->inp_route6,
13747 				   ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0),
13748 				   NULL, NULL, inp);
13749 
13750 		if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL)
13751 			mtu = inp->inp_route6.ro_nh->nh_mtu;
13752 	}
13753 #endif				/* INET6 */
13754 #if defined(INET) && defined(INET6)
13755 	else
13756 #endif
13757 #ifdef INET
13758 	{
13759 		ip->ip_len = htons(m->m_pkthdr.len);
13760 #ifdef INET6
13761 		if (inp->inp_vflag & INP_IPV6PROTO)
13762 			ip->ip_ttl = in6_selecthlim(inp, NULL);
13763 #endif				/* INET6 */
13764 		/*
13765 		 * If we do path MTU discovery, then we set DF on every
13766 		 * packet. This might not be the best thing to do according
13767 		 * to RFC3390 Section 2. However the tcp hostcache migitates
13768 		 * the problem so it affects only the first tcp connection
13769 		 * with a host.
13770 		 *
13771 		 * NB: Don't set DF on small MTU/MSS to have a safe
13772 		 * fallback.
13773 		 */
13774 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
13775 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
13776 			if (tp->t_port == 0 || len < V_tcp_minmss) {
13777 				ip->ip_off |= htons(IP_DF);
13778 			}
13779 		} else {
13780 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
13781 		}
13782 
13783 		if (tp->t_state == TCPS_SYN_SENT)
13784 			TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
13785 
13786 		TCP_PROBE5(send, NULL, tp, ip, tp, th);
13787 
13788 		error = ip_output(m, inp->inp_options, &inp->inp_route,
13789 				  ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0,
13790 				  inp);
13791 		if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL)
13792 			mtu = inp->inp_route.ro_nh->nh_mtu;
13793 	}
13794 #endif				/* INET */
13795 
13796 out:
13797 	if (lgb) {
13798 		lgb->tlb_errno = error;
13799 		lgb = NULL;
13800 	}
13801 	/*
13802 	 * In transmit state, time the transmission and arrange for the
13803 	 * retransmit.  In persist state, just set snd_max.
13804 	 */
13805 	if (error == 0) {
13806 		rack->forced_ack = 0;	/* If we send something zap the FA flag */
13807 		if (rsm && (doing_tlp == 0)) {
13808 			/* Set we retransmitted */
13809 			rack->rc_gp_saw_rec = 1;
13810 		} else {
13811 			if (cwnd_to_use > tp->snd_ssthresh) {
13812 				/* Set we sent in CA */
13813 				rack->rc_gp_saw_ca = 1;
13814 			} else {
13815 				/* Set we sent in SS */
13816 				rack->rc_gp_saw_ss = 1;
13817 			}
13818 		}
13819 		if (TCPS_HAVEESTABLISHED(tp->t_state) &&
13820 		    (tp->t_flags & TF_SACK_PERMIT) &&
13821 		    tp->rcv_numsacks > 0)
13822 			tcp_clean_dsack_blocks(tp);
13823 		tot_len_this_send += len;
13824 		if (len == 0)
13825 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_SNDACK], 1);
13826 		else if (len == 1) {
13827 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_PERSIST], 1);
13828 		} else if (len > 1) {
13829 			int idx;
13830 
13831 			idx = (len / segsiz) + 3;
13832 			if (idx >= TCP_MSS_ACCT_ATIMER)
13833 				counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
13834 			else
13835 				counter_u64_add(rack_out_size[idx], 1);
13836 		}
13837 	}
13838 	if (rack->rack_no_prr == 0) {
13839 		if (sub_from_prr && (error == 0)) {
13840 			if (rack->r_ctl.rc_prr_sndcnt >= len)
13841 				rack->r_ctl.rc_prr_sndcnt -= len;
13842 			else
13843 				rack->r_ctl.rc_prr_sndcnt = 0;
13844 		}
13845  	}
13846 	sub_from_prr = 0;
13847 	rack_log_output(tp, &to, len, rack_seq, (uint8_t) flags, error, cts,
13848 			pass, rsm, us_cts);
13849 	if ((error == 0) &&
13850 	    (len > 0) &&
13851 	    (tp->snd_una == tp->snd_max))
13852 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
13853 	/* Now are we in persists? */
13854 	if (rack->rc_in_persist == 0) {
13855 		tcp_seq startseq = tp->snd_nxt;
13856 
13857 		/* Track our lost count */
13858 		if (rsm && (doing_tlp == 0))
13859 			rack->r_ctl.rc_loss_count += rsm->r_end - rsm->r_start;
13860 		/*
13861 		 * Advance snd_nxt over sequence space of this segment.
13862 		 */
13863 		if (error)
13864 			/* We don't log or do anything with errors */
13865 			goto nomore;
13866 		if (doing_tlp == 0) {
13867 			if (rsm == NULL) {
13868 				/*
13869 				 * Not a retransmission of some
13870 				 * sort, new data is going out so
13871 				 * clear our TLP count and flag.
13872 				 */
13873 				rack->rc_tlp_in_progress = 0;
13874 				rack->r_ctl.rc_tlp_cnt_out = 0;
13875 			}
13876 		} else {
13877 			/*
13878 			 * We have just sent a TLP, mark that it is true
13879 			 * and make sure our in progress is set so we
13880 			 * continue to check the count.
13881 			 */
13882 			rack->rc_tlp_in_progress = 1;
13883 			rack->r_ctl.rc_tlp_cnt_out++;
13884 		}
13885 		if (flags & (TH_SYN | TH_FIN)) {
13886 			if (flags & TH_SYN)
13887 				tp->snd_nxt++;
13888 			if (flags & TH_FIN) {
13889 				tp->snd_nxt++;
13890 				tp->t_flags |= TF_SENTFIN;
13891 			}
13892 		}
13893 		/* In the ENOBUFS case we do *not* update snd_max */
13894 		if (sack_rxmit)
13895 			goto nomore;
13896 
13897 		tp->snd_nxt += len;
13898 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
13899 			if (tp->snd_una == tp->snd_max) {
13900 				/*
13901 				 * Update the time we just added data since
13902 				 * none was outstanding.
13903 				 */
13904 				rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
13905 				tp->t_acktime = ticks;
13906 			}
13907 			tp->snd_max = tp->snd_nxt;
13908 			/*
13909 			 * Time this transmission if not a retransmission and
13910 			 * not currently timing anything.
13911 			 * This is only relevant in case of switching back to
13912 			 * the base stack.
13913 			 */
13914 			if (tp->t_rtttime == 0) {
13915 				tp->t_rtttime = ticks;
13916 				tp->t_rtseq = startseq;
13917 				KMOD_TCPSTAT_INC(tcps_segstimed);
13918 			}
13919 			if (len &&
13920 			    ((tp->t_flags & TF_GPUTINPROG) == 0))
13921 				rack_start_gp_measurement(tp, rack, startseq, sb_offset);
13922 		}
13923 	} else {
13924 		/*
13925 		 * Persist case, update snd_max but since we are in persist
13926 		 * mode (no window) we do not update snd_nxt.
13927 		 */
13928 		int32_t xlen = len;
13929 
13930 		if (error)
13931 			goto nomore;
13932 
13933 		if (flags & TH_SYN)
13934 			++xlen;
13935 		if (flags & TH_FIN) {
13936 			++xlen;
13937 			tp->t_flags |= TF_SENTFIN;
13938 		}
13939 		/* In the ENOBUFS case we do *not* update snd_max */
13940 		if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max)) {
13941 			if (tp->snd_una == tp->snd_max) {
13942 				/*
13943 				 * Update the time we just added data since
13944 				 * none was outstanding.
13945 				 */
13946 				rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
13947 				tp->t_acktime = ticks;
13948 			}
13949 			tp->snd_max = tp->snd_nxt + len;
13950 		}
13951 	}
13952 nomore:
13953 	if (error) {
13954 		rack->r_ctl.rc_agg_delayed = 0;
13955 		rack->r_early = 0;
13956 		rack->r_late = 0;
13957 		rack->r_ctl.rc_agg_early = 0;
13958 		SOCKBUF_UNLOCK_ASSERT(sb);	/* Check gotos. */
13959 		/*
13960 		 * Failures do not advance the seq counter above. For the
13961 		 * case of ENOBUFS we will fall out and retry in 1ms with
13962 		 * the hpts. Everything else will just have to retransmit
13963 		 * with the timer.
13964 		 *
13965 		 * In any case, we do not want to loop around for another
13966 		 * send without a good reason.
13967 		 */
13968 		sendalot = 0;
13969 		switch (error) {
13970 		case EPERM:
13971 			tp->t_softerror = error;
13972 			return (error);
13973 		case ENOBUFS:
13974 			if (slot == 0) {
13975 				/*
13976 				 * Pace us right away to retry in a some
13977 				 * time
13978 				 */
13979 				slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
13980 				if (rack->rc_enobuf < 126)
13981 					rack->rc_enobuf++;
13982 				if (slot > ((rack->rc_rack_rtt / 2) * HPTS_USEC_IN_MSEC)) {
13983 					slot = (rack->rc_rack_rtt / 2) * HPTS_USEC_IN_MSEC;
13984 				}
13985 				if (slot < (10 * HPTS_USEC_IN_MSEC))
13986 					slot = 10 * HPTS_USEC_IN_MSEC;
13987 			}
13988 			counter_u64_add(rack_saw_enobuf, 1);
13989 			error = 0;
13990 			goto enobufs;
13991 		case EMSGSIZE:
13992 			/*
13993 			 * For some reason the interface we used initially
13994 			 * to send segments changed to another or lowered
13995 			 * its MTU. If TSO was active we either got an
13996 			 * interface without TSO capabilits or TSO was
13997 			 * turned off. If we obtained mtu from ip_output()
13998 			 * then update it and try again.
13999 			 */
14000 			if (tso)
14001 				tp->t_flags &= ~TF_TSO;
14002 			if (mtu != 0) {
14003 				tcp_mss_update(tp, -1, mtu, NULL, NULL);
14004 				goto again;
14005 			}
14006 			slot = 10 * HPTS_USEC_IN_MSEC;
14007 			rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
14008 			return (error);
14009 		case ENETUNREACH:
14010 			counter_u64_add(rack_saw_enetunreach, 1);
14011 		case EHOSTDOWN:
14012 		case EHOSTUNREACH:
14013 		case ENETDOWN:
14014 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
14015 				tp->t_softerror = error;
14016 			}
14017 			/* FALLTHROUGH */
14018 		default:
14019 			slot = 10 * HPTS_USEC_IN_MSEC;
14020 			rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
14021 			return (error);
14022 		}
14023 	} else {
14024 		rack->rc_enobuf = 0;
14025 	}
14026 	KMOD_TCPSTAT_INC(tcps_sndtotal);
14027 
14028 	/*
14029 	 * Data sent (as far as we can tell). If this advertises a larger
14030 	 * window than any other segment, then remember the size of the
14031 	 * advertised window. Any pending ACK has now been sent.
14032 	 */
14033 	if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
14034 		tp->rcv_adv = tp->rcv_nxt + recwin;
14035 	tp->last_ack_sent = tp->rcv_nxt;
14036 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
14037 enobufs:
14038 	/* Assure when we leave that snd_nxt will point to top */
14039 	if (SEQ_GT(tp->snd_max, tp->snd_nxt))
14040 		tp->snd_nxt = tp->snd_max;
14041 	if (sendalot) {
14042 		/* Do we need to turn off sendalot? */
14043 		if (rack->r_ctl.rc_pace_max_segs &&
14044 		    (tot_len_this_send >= rack->r_ctl.rc_pace_max_segs)) {
14045 			/* We hit our max. */
14046 			sendalot = 0;
14047 		} else if ((rack->rc_user_set_max_segs) &&
14048 			   (tot_len_this_send >= (rack->rc_user_set_max_segs * segsiz))) {
14049 			/* We hit the user defined max */
14050 			sendalot = 0;
14051 		}
14052 	}
14053 	if ((error == 0) && (flags & TH_FIN))
14054 		tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_FIN);
14055 	if (flags & TH_RST) {
14056 		/*
14057 		 * We don't send again after sending a RST.
14058 		 */
14059 		slot = 0;
14060 		sendalot = 0;
14061 		if (error == 0)
14062 			tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
14063 	} else if ((slot == 0) && (sendalot == 0) && tot_len_this_send) {
14064 		/*
14065 		 * Get our pacing rate, if an error
14066 		 * occured in sending (ENOBUF) we would
14067 		 * hit the else if with slot preset. Other
14068 		 * errors return.
14069 		 */
14070 		slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, rsm, segsiz);
14071 	}
14072 	if (rsm &&
14073 	    rack->use_rack_rr) {
14074 		/* Its a retransmit and we use the rack cheat? */
14075 		if ((slot == 0) ||
14076 		    (rack->rc_always_pace == 0) ||
14077 		    (rack->r_rr_config == 1)) {
14078 			/*
14079 			 * We have no pacing set or we
14080 			 * are using old-style rack or
14081 			 * we are overriden to use the old 1ms pacing.
14082 			 */
14083 			slot = rack->r_ctl.rc_min_to * HPTS_USEC_IN_MSEC;
14084 		}
14085 	}
14086 	if (slot) {
14087 		/* set the rack tcb into the slot N */
14088 		counter_u64_add(rack_paced_segments, 1);
14089 	} else if (sendalot) {
14090 		if (len)
14091 			counter_u64_add(rack_unpaced_segments, 1);
14092 		sack_rxmit = 0;
14093 		goto again;
14094 	} else if (len) {
14095 		counter_u64_add(rack_unpaced_segments, 1);
14096 	}
14097 	rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, 0);
14098 	return (error);
14099 }
14100 
14101 static void
rack_update_seg(struct tcp_rack * rack)14102 rack_update_seg(struct tcp_rack *rack)
14103 {
14104 	uint32_t orig_val;
14105 
14106 	orig_val = rack->r_ctl.rc_pace_max_segs;
14107 	rack_set_pace_segments(rack->rc_tp, rack, __LINE__);
14108 	if (orig_val != rack->r_ctl.rc_pace_max_segs)
14109 		rack_log_pacing_delay_calc(rack, 0, 0, orig_val, 0, 0, 15, __LINE__, NULL);
14110 }
14111 
14112 /*
14113  * rack_ctloutput() must drop the inpcb lock before performing copyin on
14114  * socket option arguments.  When it re-acquires the lock after the copy, it
14115  * has to revalidate that the connection is still valid for the socket
14116  * option.
14117  */
14118 static int
rack_set_sockopt(struct socket * so,struct sockopt * sopt,struct inpcb * inp,struct tcpcb * tp,struct tcp_rack * rack)14119 rack_set_sockopt(struct socket *so, struct sockopt *sopt,
14120     struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack)
14121 {
14122 	struct epoch_tracker et;
14123 	uint64_t val;
14124 	int32_t error = 0, optval;
14125 	uint16_t ca, ss;
14126 
14127 	switch (sopt->sopt_name) {
14128 	case TCP_RACK_PROP_RATE:		/*  URL:prop_rate */
14129 	case TCP_RACK_PROP	:		/*  URL:prop */
14130 	case TCP_RACK_TLP_REDUCE:		/*  URL:tlp_reduce */
14131 	case TCP_RACK_EARLY_RECOV:		/*  URL:early_recov */
14132 	case TCP_RACK_PACE_REDUCE:		/*  Not used */
14133         /*  Pacing related ones */
14134 	case TCP_RACK_PACE_ALWAYS:		/*  URL:pace_always */
14135 	case TCP_BBR_RACK_INIT_RATE:		/*  URL:irate */
14136 	case TCP_BBR_IWINTSO:			/*  URL:tso_iwin */
14137 	case TCP_RACK_PACE_MAX_SEG:		/*  URL:pace_max_seg */
14138 	case TCP_RACK_FORCE_MSEG:		/*  URL:force_max_seg */
14139 	case TCP_RACK_PACE_RATE_CA:		/*  URL:pr_ca */
14140 	case TCP_RACK_PACE_RATE_SS:		/*  URL:pr_ss*/
14141 	case TCP_RACK_PACE_RATE_REC:		/*  URL:pr_rec */
14142 	case TCP_RACK_GP_INCREASE_CA:		/*  URL:gp_inc_ca */
14143 	case TCP_RACK_GP_INCREASE_SS:		/*  URL:gp_inc_ss */
14144 	case TCP_RACK_GP_INCREASE_REC:		/*  URL:gp_inc_rec */
14145 	case TCP_RACK_RR_CONF:			/*  URL:rrr_conf */
14146 	case TCP_BBR_HDWR_PACE:			/*  URL:hdwrpace */
14147        /* End pacing related */
14148 	case TCP_DELACK:
14149 	case TCP_RACK_PRR_SENDALOT:		/*  URL:prr_sendalot */
14150 	case TCP_RACK_MIN_TO:			/*  URL:min_to */
14151 	case TCP_RACK_EARLY_SEG:		/*  URL:early_seg */
14152 	case TCP_RACK_REORD_THRESH:		/*  URL:reord_thresh */
14153 	case TCP_RACK_REORD_FADE:		/*  URL:reord_fade */
14154 	case TCP_RACK_TLP_THRESH:		/*  URL:tlp_thresh */
14155 	case TCP_RACK_PKT_DELAY:		/*  URL:pkt_delay */
14156 	case TCP_RACK_TLP_USE:			/*  URL:tlp_use */
14157 	case TCP_RACK_TLP_INC_VAR:		/*  URL:tlp_inc_var */
14158 	case TCP_RACK_IDLE_REDUCE_HIGH:		/*  URL:idle_reduce_high */
14159 	case TCP_BBR_RACK_RTT_USE:		/*  URL:rttuse */
14160 	case TCP_BBR_USE_RACK_RR:		/*  URL:rackrr */
14161 	case TCP_RACK_DO_DETECTION:		/*  URL:detect */
14162 	case TCP_NO_PRR:			/*  URL:noprr */
14163 	case TCP_TIMELY_DYN_ADJ:		/*  URL:dynamic */
14164 	case TCP_DATA_AFTER_CLOSE:
14165 	case TCP_RACK_NONRXT_CFG_RATE:		/*  URL:nonrxtcr */
14166 	case TCP_SHARED_CWND_ENABLE:		/*  URL:scwnd */
14167 	case TCP_RACK_MBUF_QUEUE:		/*  URL:mqueue */
14168 	case TCP_RACK_NO_PUSH_AT_MAX:		/*  URL:npush */
14169 	case TCP_RACK_PACE_TO_FILL:		/*  URL:fillcw */
14170 	case TCP_SHARED_CWND_TIME_LIMIT:	/*  URL:lscwnd */
14171 	case TCP_RACK_PROFILE:			/*  URL:profile */
14172 		break;
14173 	default:
14174 		return (tcp_default_ctloutput(so, sopt, inp, tp));
14175 		break;
14176 	}
14177 	INP_WUNLOCK(inp);
14178 	error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval));
14179 	if (error)
14180 		return (error);
14181 	INP_WLOCK(inp);
14182 	if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
14183 		INP_WUNLOCK(inp);
14184 		return (ECONNRESET);
14185 	}
14186 	tp = intotcpcb(inp);
14187 	rack = (struct tcp_rack *)tp->t_fb_ptr;
14188 	switch (sopt->sopt_name) {
14189 	case TCP_RACK_PROFILE:
14190 		RACK_OPTS_INC(tcp_profile);
14191 		if (optval == 1) {
14192 			/* pace_always=1 */
14193 			rack->rc_always_pace = 1;
14194 			tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
14195 			/* scwnd=1 */
14196 			rack->rack_enable_scwnd = 1;
14197 			/* dynamic=100 */
14198 			rack->rc_gp_dyn_mul = 1;
14199 			rack->r_ctl.rack_per_of_gp_ca = 100;
14200 			/* rrr_conf=3 */
14201 			rack->r_rr_config = 3;
14202 			/* npush=2 */
14203 			rack->r_ctl.rc_no_push_at_mrtt = 2;
14204 			/* fillcw=1 */
14205 			rack->rc_pace_to_cwnd = 1;
14206 			rack->rc_pace_fill_if_rttin_range = 0;
14207 			rack->rtt_limit_mul = 0;
14208 			/* noprr=1 */
14209 			rack->rack_no_prr = 1;
14210 			/* lscwnd=1 */
14211 			rack->r_limit_scw = 1;
14212 		} else if (optval == 2) {
14213 			/* pace_always=1 */
14214 			rack->rc_always_pace = 1;
14215 			tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
14216 			/* scwnd=1 */
14217 			rack->rack_enable_scwnd = 1;
14218 			/* dynamic=100 */
14219 			rack->rc_gp_dyn_mul = 1;
14220 			rack->r_ctl.rack_per_of_gp_ca = 100;
14221 			/* rrr_conf=3 */
14222 			rack->r_rr_config = 3;
14223 			/* npush=2 */
14224 			rack->r_ctl.rc_no_push_at_mrtt = 2;
14225 			/* fillcw=1 */
14226 			rack->rc_pace_to_cwnd = 1;
14227 			rack->rc_pace_fill_if_rttin_range = 0;
14228 			rack->rtt_limit_mul = 0;
14229 			/* noprr=1 */
14230 			rack->rack_no_prr = 1;
14231 			/* lscwnd=0 */
14232 			rack->r_limit_scw = 0;
14233 		}
14234 		break;
14235 	case TCP_SHARED_CWND_TIME_LIMIT:
14236 		RACK_OPTS_INC(tcp_lscwnd);
14237 		if (optval)
14238 			rack->r_limit_scw = 1;
14239 		else
14240 			rack->r_limit_scw = 0;
14241 		break;
14242  	case TCP_RACK_PACE_TO_FILL:
14243 		RACK_OPTS_INC(tcp_fillcw);
14244 		if (optval == 0)
14245 			rack->rc_pace_to_cwnd = 0;
14246 		else
14247 			rack->rc_pace_to_cwnd = 1;
14248 		if ((optval >= rack_gp_rtt_maxmul) &&
14249 		    rack_gp_rtt_maxmul &&
14250 		    (optval < 0xf)) {
14251 			rack->rc_pace_fill_if_rttin_range = 1;
14252 			rack->rtt_limit_mul = optval;
14253 		} else {
14254 			rack->rc_pace_fill_if_rttin_range = 0;
14255 			rack->rtt_limit_mul = 0;
14256 		}
14257 		break;
14258 	case TCP_RACK_NO_PUSH_AT_MAX:
14259 		RACK_OPTS_INC(tcp_npush);
14260 		if (optval == 0)
14261 			rack->r_ctl.rc_no_push_at_mrtt = 0;
14262 		else if (optval < 0xff)
14263 			rack->r_ctl.rc_no_push_at_mrtt = optval;
14264 		else
14265 			error = EINVAL;
14266 		break;
14267 	case TCP_SHARED_CWND_ENABLE:
14268 		RACK_OPTS_INC(tcp_rack_scwnd);
14269 		if (optval == 0)
14270 			rack->rack_enable_scwnd = 0;
14271 		else
14272 			rack->rack_enable_scwnd = 1;
14273 		break;
14274 	case TCP_RACK_MBUF_QUEUE:
14275 		/* Now do we use the LRO mbuf-queue feature */
14276 		RACK_OPTS_INC(tcp_rack_mbufq);
14277 		if (optval)
14278 			rack->r_mbuf_queue = 1;
14279 		else
14280 			rack->r_mbuf_queue = 0;
14281 		if  (rack->r_mbuf_queue || rack->rc_always_pace)
14282 			tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
14283 		else
14284 			tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
14285 		break;
14286 	case TCP_RACK_NONRXT_CFG_RATE:
14287 		RACK_OPTS_INC(tcp_rack_cfg_rate);
14288 		if (optval == 0)
14289 			rack->rack_rec_nonrxt_use_cr = 0;
14290 		else
14291 			rack->rack_rec_nonrxt_use_cr = 1;
14292 		break;
14293 	case TCP_NO_PRR:
14294 		RACK_OPTS_INC(tcp_rack_noprr);
14295 		if (optval == 0)
14296 			rack->rack_no_prr = 0;
14297 		else
14298 			rack->rack_no_prr = 1;
14299 		break;
14300 	case TCP_TIMELY_DYN_ADJ:
14301 		RACK_OPTS_INC(tcp_timely_dyn);
14302 		if (optval == 0)
14303 			rack->rc_gp_dyn_mul = 0;
14304 		else {
14305 			rack->rc_gp_dyn_mul = 1;
14306 			if (optval >= 100) {
14307 				/*
14308 				 * If the user sets something 100 or more
14309 				 * its the gp_ca value.
14310 				 */
14311 				rack->r_ctl.rack_per_of_gp_ca  = optval;
14312 			}
14313 		}
14314 		break;
14315 	case TCP_RACK_DO_DETECTION:
14316 		RACK_OPTS_INC(tcp_rack_do_detection);
14317 		if (optval == 0)
14318 			rack->do_detection = 0;
14319 		else
14320 			rack->do_detection = 1;
14321 		break;
14322 	case TCP_RACK_PROP_RATE:
14323 		if ((optval <= 0) || (optval >= 100)) {
14324 			error = EINVAL;
14325 			break;
14326 		}
14327 		RACK_OPTS_INC(tcp_rack_prop_rate);
14328 		rack->r_ctl.rc_prop_rate = optval;
14329 		break;
14330 	case TCP_RACK_TLP_USE:
14331 		if ((optval < TLP_USE_ID) || (optval > TLP_USE_TWO_TWO)) {
14332 			error = EINVAL;
14333 			break;
14334 		}
14335 		RACK_OPTS_INC(tcp_tlp_use);
14336 		rack->rack_tlp_threshold_use = optval;
14337 		break;
14338 	case TCP_RACK_PROP:
14339 		/* RACK proportional rate reduction (bool) */
14340 		RACK_OPTS_INC(tcp_rack_prop);
14341 		rack->r_ctl.rc_prop_reduce = optval;
14342 		break;
14343 	case TCP_RACK_TLP_REDUCE:
14344 		/* RACK TLP cwnd reduction (bool) */
14345 		RACK_OPTS_INC(tcp_rack_tlp_reduce);
14346 		rack->r_ctl.rc_tlp_cwnd_reduce = optval;
14347 		break;
14348 	case TCP_RACK_EARLY_RECOV:
14349 		/* Should recovery happen early (bool) */
14350 		RACK_OPTS_INC(tcp_rack_early_recov);
14351 		rack->r_ctl.rc_early_recovery = optval;
14352 		break;
14353 
14354         /*  Pacing related ones */
14355 	case TCP_RACK_PACE_ALWAYS:
14356 		/*
14357 		 * zero is old rack method, 1 is new
14358 		 * method using a pacing rate.
14359 		 */
14360 		RACK_OPTS_INC(tcp_rack_pace_always);
14361 		if (optval > 0)
14362 			rack->rc_always_pace = 1;
14363 		else
14364 			rack->rc_always_pace = 0;
14365 		if  (rack->r_mbuf_queue || rack->rc_always_pace)
14366 			tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
14367 		else
14368 			tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
14369 		/* A rate may be set irate or other, if so set seg size */
14370 		rack_update_seg(rack);
14371 		break;
14372 	case TCP_BBR_RACK_INIT_RATE:
14373 		RACK_OPTS_INC(tcp_initial_rate);
14374 		val = optval;
14375 		/* Change from kbits per second to bytes per second */
14376 		val *= 1000;
14377 		val /= 8;
14378 		rack->r_ctl.init_rate = val;
14379 		if (rack->rc_init_win != rack_default_init_window) {
14380 			uint32_t win, snt;
14381 
14382 			/*
14383 			 * Options don't always get applied
14384 			 * in the order you think. So in order
14385 			 * to assure we update a cwnd we need
14386 			 * to check and see if we are still
14387 			 * where we should raise the cwnd.
14388 			 */
14389 			win = rc_init_window(rack);
14390 			if (SEQ_GT(tp->snd_max, tp->iss))
14391 				snt = tp->snd_max - tp->iss;
14392 			else
14393 				snt = 0;
14394 			if ((snt < win) &&
14395 			    (tp->snd_cwnd < win))
14396 				tp->snd_cwnd = win;
14397 		}
14398 		if (rack->rc_always_pace)
14399 			rack_update_seg(rack);
14400 		break;
14401 	case TCP_BBR_IWINTSO:
14402 		RACK_OPTS_INC(tcp_initial_win);
14403 		if (optval && (optval <= 0xff)) {
14404 			uint32_t win, snt;
14405 
14406 			rack->rc_init_win = optval;
14407 			win = rc_init_window(rack);
14408 			if (SEQ_GT(tp->snd_max, tp->iss))
14409 				snt = tp->snd_max - tp->iss;
14410 			else
14411 				snt = 0;
14412 			if ((snt < win) &&
14413 			    (tp->t_srtt |
14414 #ifdef NETFLIX_PEAKRATE
14415 			     tp->t_maxpeakrate |
14416 #endif
14417 			     rack->r_ctl.init_rate)) {
14418 				/*
14419 				 * We are not past the initial window
14420 				 * and we have some bases for pacing,
14421 				 * so we need to possibly adjust up
14422 				 * the cwnd. Note even if we don't set
14423 				 * the cwnd, its still ok to raise the rc_init_win
14424 				 * which can be used coming out of idle when we
14425 				 * would have a rate.
14426 				 */
14427 				if (tp->snd_cwnd < win)
14428 					tp->snd_cwnd = win;
14429 			}
14430 			if (rack->rc_always_pace)
14431 				rack_update_seg(rack);
14432 		} else
14433 			error = EINVAL;
14434 		break;
14435 	case TCP_RACK_FORCE_MSEG:
14436 		RACK_OPTS_INC(tcp_rack_force_max_seg);
14437 		if (optval)
14438 			rack->rc_force_max_seg = 1;
14439 		else
14440 			rack->rc_force_max_seg = 0;
14441 		break;
14442 	case TCP_RACK_PACE_MAX_SEG:
14443 		/* Max segments size in a pace in bytes */
14444 		RACK_OPTS_INC(tcp_rack_max_seg);
14445 		rack->rc_user_set_max_segs = optval;
14446 		rack_set_pace_segments(tp, rack, __LINE__);
14447 		break;
14448 	case TCP_RACK_PACE_RATE_REC:
14449 		/* Set the fixed pacing rate in Bytes per second ca */
14450 		RACK_OPTS_INC(tcp_rack_pace_rate_rec);
14451 		rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
14452 		if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
14453 			rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
14454 		if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
14455 			rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
14456 		rack->use_fixed_rate = 1;
14457 		rack_log_pacing_delay_calc(rack,
14458 					   rack->r_ctl.rc_fixed_pacing_rate_ss,
14459 					   rack->r_ctl.rc_fixed_pacing_rate_ca,
14460 					   rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
14461 					   __LINE__, NULL);
14462 		break;
14463 
14464 	case TCP_RACK_PACE_RATE_SS:
14465 		/* Set the fixed pacing rate in Bytes per second ca */
14466 		RACK_OPTS_INC(tcp_rack_pace_rate_ss);
14467 		rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
14468 		if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
14469 			rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
14470 		if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
14471 			rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
14472 		rack->use_fixed_rate = 1;
14473 		rack_log_pacing_delay_calc(rack,
14474 					   rack->r_ctl.rc_fixed_pacing_rate_ss,
14475 					   rack->r_ctl.rc_fixed_pacing_rate_ca,
14476 					   rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
14477 					   __LINE__, NULL);
14478 		break;
14479 
14480 	case TCP_RACK_PACE_RATE_CA:
14481 		/* Set the fixed pacing rate in Bytes per second ca */
14482 		RACK_OPTS_INC(tcp_rack_pace_rate_ca);
14483 		rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
14484 		if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
14485 			rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
14486 		if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
14487 			rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
14488 		rack->use_fixed_rate = 1;
14489 		rack_log_pacing_delay_calc(rack,
14490 					   rack->r_ctl.rc_fixed_pacing_rate_ss,
14491 					   rack->r_ctl.rc_fixed_pacing_rate_ca,
14492 					   rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
14493 					   __LINE__, NULL);
14494 		break;
14495 	case TCP_RACK_GP_INCREASE_REC:
14496 		RACK_OPTS_INC(tcp_gp_inc_rec);
14497 		rack->r_ctl.rack_per_of_gp_rec = optval;
14498 		rack_log_pacing_delay_calc(rack,
14499 					   rack->r_ctl.rack_per_of_gp_ss,
14500 					   rack->r_ctl.rack_per_of_gp_ca,
14501 					   rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
14502 					   __LINE__, NULL);
14503 		break;
14504 	case TCP_RACK_GP_INCREASE_CA:
14505 		RACK_OPTS_INC(tcp_gp_inc_ca);
14506 		ca = optval;
14507 		if (ca < 100) {
14508 			/*
14509 			 * We don't allow any reduction
14510 			 * over the GP b/w.
14511 			 */
14512 			error = EINVAL;
14513 			break;
14514 		}
14515 		rack->r_ctl.rack_per_of_gp_ca = ca;
14516 		rack_log_pacing_delay_calc(rack,
14517 					   rack->r_ctl.rack_per_of_gp_ss,
14518 					   rack->r_ctl.rack_per_of_gp_ca,
14519 					   rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
14520 					   __LINE__, NULL);
14521 		break;
14522 	case TCP_RACK_GP_INCREASE_SS:
14523 		RACK_OPTS_INC(tcp_gp_inc_ss);
14524 		ss = optval;
14525 		if (ss < 100) {
14526 			/*
14527 			 * We don't allow any reduction
14528 			 * over the GP b/w.
14529 			 */
14530 			error = EINVAL;
14531 			break;
14532 		}
14533 		rack->r_ctl.rack_per_of_gp_ss = ss;
14534 		rack_log_pacing_delay_calc(rack,
14535 					   rack->r_ctl.rack_per_of_gp_ss,
14536 					   rack->r_ctl.rack_per_of_gp_ca,
14537 					   rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
14538 					   __LINE__, NULL);
14539 		break;
14540 	case TCP_RACK_RR_CONF:
14541 		RACK_OPTS_INC(tcp_rack_rrr_no_conf_rate);
14542 		if (optval && optval <= 3)
14543 			rack->r_rr_config = optval;
14544 		else
14545 			rack->r_rr_config = 0;
14546 		break;
14547 	case TCP_BBR_HDWR_PACE:
14548 		RACK_OPTS_INC(tcp_hdwr_pacing);
14549 		if (optval){
14550 			if (rack->rack_hdrw_pacing == 0) {
14551 				rack->rack_hdw_pace_ena = 1;
14552 				rack->rack_attempt_hdwr_pace = 0;
14553 			} else
14554 				error = EALREADY;
14555 		} else {
14556 			rack->rack_hdw_pace_ena = 0;
14557 #ifdef RATELIMIT
14558 			if (rack->rack_hdrw_pacing) {
14559 				rack->rack_hdrw_pacing = 0;
14560 				in_pcbdetach_txrtlmt(rack->rc_inp);
14561 			}
14562 #endif
14563 		}
14564 		break;
14565         /*  End Pacing related ones */
14566 	case TCP_RACK_PRR_SENDALOT:
14567 		/* Allow PRR to send more than one seg */
14568 		RACK_OPTS_INC(tcp_rack_prr_sendalot);
14569 		rack->r_ctl.rc_prr_sendalot = optval;
14570 		break;
14571 	case TCP_RACK_MIN_TO:
14572 		/* Minimum time between rack t-o's in ms */
14573 		RACK_OPTS_INC(tcp_rack_min_to);
14574 		rack->r_ctl.rc_min_to = optval;
14575 		break;
14576 	case TCP_RACK_EARLY_SEG:
14577 		/* If early recovery max segments */
14578 		RACK_OPTS_INC(tcp_rack_early_seg);
14579 		rack->r_ctl.rc_early_recovery_segs = optval;
14580 		break;
14581 	case TCP_RACK_REORD_THRESH:
14582 		/* RACK reorder threshold (shift amount) */
14583 		RACK_OPTS_INC(tcp_rack_reord_thresh);
14584 		if ((optval > 0) && (optval < 31))
14585 			rack->r_ctl.rc_reorder_shift = optval;
14586 		else
14587 			error = EINVAL;
14588 		break;
14589 	case TCP_RACK_REORD_FADE:
14590 		/* Does reordering fade after ms time */
14591 		RACK_OPTS_INC(tcp_rack_reord_fade);
14592 		rack->r_ctl.rc_reorder_fade = optval;
14593 		break;
14594 	case TCP_RACK_TLP_THRESH:
14595 		/* RACK TLP theshold i.e. srtt+(srtt/N) */
14596 		RACK_OPTS_INC(tcp_rack_tlp_thresh);
14597 		if (optval)
14598 			rack->r_ctl.rc_tlp_threshold = optval;
14599 		else
14600 			error = EINVAL;
14601 		break;
14602 	case TCP_BBR_USE_RACK_RR:
14603 		RACK_OPTS_INC(tcp_rack_rr);
14604 		if (optval)
14605 			rack->use_rack_rr = 1;
14606 		else
14607 			rack->use_rack_rr = 0;
14608 		break;
14609 	case TCP_RACK_PKT_DELAY:
14610 		/* RACK added ms i.e. rack-rtt + reord + N */
14611 		RACK_OPTS_INC(tcp_rack_pkt_delay);
14612 		rack->r_ctl.rc_pkt_delay = optval;
14613 		break;
14614 	case TCP_RACK_TLP_INC_VAR:
14615 		/* Does TLP include rtt variance in t-o */
14616 		error = EINVAL;
14617 		break;
14618 	case TCP_RACK_IDLE_REDUCE_HIGH:
14619 		error = EINVAL;
14620 		break;
14621 	case TCP_DELACK:
14622 		if (optval == 0)
14623 			tp->t_delayed_ack = 0;
14624 		else
14625 			tp->t_delayed_ack = 1;
14626 		if (tp->t_flags & TF_DELACK) {
14627 			tp->t_flags &= ~TF_DELACK;
14628 			tp->t_flags |= TF_ACKNOW;
14629 			NET_EPOCH_ENTER(et);
14630 			rack_output(tp);
14631 			NET_EPOCH_EXIT(et);
14632 		}
14633 		break;
14634 
14635 	case TCP_BBR_RACK_RTT_USE:
14636 		if ((optval != USE_RTT_HIGH) &&
14637 		    (optval != USE_RTT_LOW) &&
14638 		    (optval != USE_RTT_AVG))
14639 			error = EINVAL;
14640 		else
14641 			rack->r_ctl.rc_rate_sample_method = optval;
14642 		break;
14643 	case TCP_DATA_AFTER_CLOSE:
14644 		if (optval)
14645 			rack->rc_allow_data_af_clo = 1;
14646 		else
14647 			rack->rc_allow_data_af_clo = 0;
14648 		break;
14649 	case TCP_RACK_PACE_REDUCE:
14650 		/* sysctl only now */
14651 		error = EINVAL;
14652 		break;
14653 	default:
14654 		return (tcp_default_ctloutput(so, sopt, inp, tp));
14655 		break;
14656 	}
14657 #ifdef NETFLIX_STATS
14658 	tcp_log_socket_option(tp, sopt->sopt_name, optval, error);
14659 #endif
14660 	INP_WUNLOCK(inp);
14661 	return (error);
14662 }
14663 
14664 static int
rack_get_sockopt(struct socket * so,struct sockopt * sopt,struct inpcb * inp,struct tcpcb * tp,struct tcp_rack * rack)14665 rack_get_sockopt(struct socket *so, struct sockopt *sopt,
14666     struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack)
14667 {
14668 	int32_t error, optval;
14669 	uint64_t val;
14670 	/*
14671 	 * Because all our options are either boolean or an int, we can just
14672 	 * pull everything into optval and then unlock and copy. If we ever
14673 	 * add a option that is not a int, then this will have quite an
14674 	 * impact to this routine.
14675 	 */
14676 	error = 0;
14677 	switch (sopt->sopt_name) {
14678 	case TCP_RACK_PROFILE:
14679 		/* You cannot retrieve a profile, its write only */
14680 		error = EINVAL;
14681 		break;
14682 	case TCP_RACK_PACE_TO_FILL:
14683 		optval = rack->rc_pace_to_cwnd;
14684 		break;
14685 	case TCP_RACK_NO_PUSH_AT_MAX:
14686 		optval = rack->r_ctl.rc_no_push_at_mrtt;
14687 		break;
14688 	case TCP_SHARED_CWND_ENABLE:
14689 		optval = rack->rack_enable_scwnd;
14690 		break;
14691 	case TCP_RACK_NONRXT_CFG_RATE:
14692 		optval = rack->rack_rec_nonrxt_use_cr;
14693 		break;
14694 	case TCP_NO_PRR:
14695 		optval = rack->rack_no_prr;
14696 		break;
14697 	case TCP_RACK_DO_DETECTION:
14698 		optval = rack->do_detection;
14699 		break;
14700 	case TCP_RACK_MBUF_QUEUE:
14701 		/* Now do we use the LRO mbuf-queue feature */
14702 		optval = rack->r_mbuf_queue;
14703 		break;
14704 	case TCP_TIMELY_DYN_ADJ:
14705 		optval = rack->rc_gp_dyn_mul;
14706 		break;
14707 	case TCP_BBR_IWINTSO:
14708 		optval = rack->rc_init_win;
14709 		break;
14710 	case TCP_RACK_PROP_RATE:
14711 		optval = rack->r_ctl.rc_prop_rate;
14712 		break;
14713 	case TCP_RACK_PROP:
14714 		/* RACK proportional rate reduction (bool) */
14715 		optval = rack->r_ctl.rc_prop_reduce;
14716 		break;
14717 	case TCP_RACK_TLP_REDUCE:
14718 		/* RACK TLP cwnd reduction (bool) */
14719 		optval = rack->r_ctl.rc_tlp_cwnd_reduce;
14720 		break;
14721 	case TCP_RACK_EARLY_RECOV:
14722 		/* Should recovery happen early (bool) */
14723 		optval = rack->r_ctl.rc_early_recovery;
14724 		break;
14725 	case TCP_RACK_PACE_REDUCE:
14726 		/* RACK Hptsi reduction factor (divisor) */
14727 		error = EINVAL;
14728 		break;
14729 	case TCP_BBR_RACK_INIT_RATE:
14730 		val = rack->r_ctl.init_rate;
14731 		/* convert to kbits per sec */
14732 		val *= 8;
14733 		val /= 1000;
14734 		optval = (uint32_t)val;
14735 		break;
14736 	case TCP_RACK_FORCE_MSEG:
14737 		optval = rack->rc_force_max_seg;
14738 		break;
14739 	case TCP_RACK_PACE_MAX_SEG:
14740 		/* Max segments in a pace */
14741 		optval = rack->rc_user_set_max_segs;
14742 		break;
14743 	case TCP_RACK_PACE_ALWAYS:
14744 		/* Use the always pace method */
14745 		optval = rack->rc_always_pace;
14746 		break;
14747 	case TCP_RACK_PRR_SENDALOT:
14748 		/* Allow PRR to send more than one seg */
14749 		optval = rack->r_ctl.rc_prr_sendalot;
14750 		break;
14751 	case TCP_RACK_MIN_TO:
14752 		/* Minimum time between rack t-o's in ms */
14753 		optval = rack->r_ctl.rc_min_to;
14754 		break;
14755 	case TCP_RACK_EARLY_SEG:
14756 		/* If early recovery max segments */
14757 		optval = rack->r_ctl.rc_early_recovery_segs;
14758 		break;
14759 	case TCP_RACK_REORD_THRESH:
14760 		/* RACK reorder threshold (shift amount) */
14761 		optval = rack->r_ctl.rc_reorder_shift;
14762 		break;
14763 	case TCP_RACK_REORD_FADE:
14764 		/* Does reordering fade after ms time */
14765 		optval = rack->r_ctl.rc_reorder_fade;
14766 		break;
14767 	case TCP_BBR_USE_RACK_RR:
14768 		/* Do we use the rack cheat for rxt */
14769 		optval = rack->use_rack_rr;
14770 		break;
14771 	case TCP_RACK_RR_CONF:
14772 		optval = rack->r_rr_config;
14773 		break;
14774 	case TCP_BBR_HDWR_PACE:
14775 		optval = rack->rack_hdw_pace_ena;
14776 		break;
14777 	case TCP_RACK_TLP_THRESH:
14778 		/* RACK TLP theshold i.e. srtt+(srtt/N) */
14779 		optval = rack->r_ctl.rc_tlp_threshold;
14780 		break;
14781 	case TCP_RACK_PKT_DELAY:
14782 		/* RACK added ms i.e. rack-rtt + reord + N */
14783 		optval = rack->r_ctl.rc_pkt_delay;
14784 		break;
14785 	case TCP_RACK_TLP_USE:
14786 		optval = rack->rack_tlp_threshold_use;
14787 		break;
14788 	case TCP_RACK_TLP_INC_VAR:
14789 		/* Does TLP include rtt variance in t-o */
14790 		error = EINVAL;
14791 		break;
14792 	case TCP_RACK_IDLE_REDUCE_HIGH:
14793 		error = EINVAL;
14794 		break;
14795 	case TCP_RACK_PACE_RATE_CA:
14796 		optval = rack->r_ctl.rc_fixed_pacing_rate_ca;
14797 		break;
14798 	case TCP_RACK_PACE_RATE_SS:
14799 		optval = rack->r_ctl.rc_fixed_pacing_rate_ss;
14800 		break;
14801 	case TCP_RACK_PACE_RATE_REC:
14802 		optval = rack->r_ctl.rc_fixed_pacing_rate_rec;
14803 		break;
14804 	case TCP_RACK_GP_INCREASE_SS:
14805 		optval = rack->r_ctl.rack_per_of_gp_ca;
14806 		break;
14807 	case TCP_RACK_GP_INCREASE_CA:
14808 		optval = rack->r_ctl.rack_per_of_gp_ss;
14809 		break;
14810 	case TCP_BBR_RACK_RTT_USE:
14811 		optval = rack->r_ctl.rc_rate_sample_method;
14812 		break;
14813 	case TCP_DELACK:
14814 		optval = tp->t_delayed_ack;
14815 		break;
14816 	case TCP_DATA_AFTER_CLOSE:
14817 		optval = rack->rc_allow_data_af_clo;
14818 		break;
14819 	case TCP_SHARED_CWND_TIME_LIMIT:
14820 		optval = rack->r_limit_scw;
14821 		break;
14822 	default:
14823 		return (tcp_default_ctloutput(so, sopt, inp, tp));
14824 		break;
14825 	}
14826 	INP_WUNLOCK(inp);
14827 	if (error == 0) {
14828 		error = sooptcopyout(sopt, &optval, sizeof optval);
14829 	}
14830 	return (error);
14831 }
14832 
14833 static int
rack_ctloutput(struct socket * so,struct sockopt * sopt,struct inpcb * inp,struct tcpcb * tp)14834 rack_ctloutput(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp)
14835 {
14836 	int32_t error = EINVAL;
14837 	struct tcp_rack *rack;
14838 
14839 	rack = (struct tcp_rack *)tp->t_fb_ptr;
14840 	if (rack == NULL) {
14841 		/* Huh? */
14842 		goto out;
14843 	}
14844 	if (sopt->sopt_dir == SOPT_SET) {
14845 		return (rack_set_sockopt(so, sopt, inp, tp, rack));
14846 	} else if (sopt->sopt_dir == SOPT_GET) {
14847 		return (rack_get_sockopt(so, sopt, inp, tp, rack));
14848 	}
14849 out:
14850 	INP_WUNLOCK(inp);
14851 	return (error);
14852 }
14853 
14854 static int
rack_pru_options(struct tcpcb * tp,int flags)14855 rack_pru_options(struct tcpcb *tp, int flags)
14856 {
14857 	if (flags & PRUS_OOB)
14858 		return (EOPNOTSUPP);
14859 	return (0);
14860 }
14861 
14862 static struct tcp_function_block __tcp_rack = {
14863 	.tfb_tcp_block_name = __XSTRING(STACKNAME),
14864 	.tfb_tcp_output = rack_output,
14865 	.tfb_do_queued_segments = ctf_do_queued_segments,
14866 	.tfb_do_segment_nounlock = rack_do_segment_nounlock,
14867 	.tfb_tcp_do_segment = rack_do_segment,
14868 	.tfb_tcp_ctloutput = rack_ctloutput,
14869 	.tfb_tcp_fb_init = rack_init,
14870 	.tfb_tcp_fb_fini = rack_fini,
14871 	.tfb_tcp_timer_stop_all = rack_stopall,
14872 	.tfb_tcp_timer_activate = rack_timer_activate,
14873 	.tfb_tcp_timer_active = rack_timer_active,
14874 	.tfb_tcp_timer_stop = rack_timer_stop,
14875 	.tfb_tcp_rexmit_tmr = rack_remxt_tmr,
14876 	.tfb_tcp_handoff_ok = rack_handoff_ok,
14877 	.tfb_pru_options = rack_pru_options,
14878 };
14879 
14880 static const char *rack_stack_names[] = {
14881 	__XSTRING(STACKNAME),
14882 #ifdef STACKALIAS
14883 	__XSTRING(STACKALIAS),
14884 #endif
14885 };
14886 
14887 static int
rack_ctor(void * mem,int32_t size,void * arg,int32_t how)14888 rack_ctor(void *mem, int32_t size, void *arg, int32_t how)
14889 {
14890 	memset(mem, 0, size);
14891 	return (0);
14892 }
14893 
14894 static void
rack_dtor(void * mem,int32_t size,void * arg)14895 rack_dtor(void *mem, int32_t size, void *arg)
14896 {
14897 
14898 }
14899 
14900 static bool rack_mod_inited = false;
14901 
14902 static int
tcp_addrack(module_t mod,int32_t type,void * data)14903 tcp_addrack(module_t mod, int32_t type, void *data)
14904 {
14905 	int32_t err = 0;
14906 	int num_stacks;
14907 
14908 	switch (type) {
14909 	case MOD_LOAD:
14910 		rack_zone = uma_zcreate(__XSTRING(MODNAME) "_map",
14911 		    sizeof(struct rack_sendmap),
14912 		    rack_ctor, rack_dtor, NULL, NULL, UMA_ALIGN_PTR, 0);
14913 
14914 		rack_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb",
14915 		    sizeof(struct tcp_rack),
14916 		    rack_ctor, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0);
14917 
14918 		sysctl_ctx_init(&rack_sysctl_ctx);
14919 		rack_sysctl_root = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
14920 		    SYSCTL_STATIC_CHILDREN(_net_inet_tcp),
14921 		    OID_AUTO,
14922 #ifdef STACKALIAS
14923 		    __XSTRING(STACKALIAS),
14924 #else
14925 		    __XSTRING(STACKNAME),
14926 #endif
14927 		    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
14928 		    "");
14929 		if (rack_sysctl_root == NULL) {
14930 			printf("Failed to add sysctl node\n");
14931 			err = EFAULT;
14932 			goto free_uma;
14933 		}
14934 		rack_init_sysctls();
14935 		num_stacks = nitems(rack_stack_names);
14936 		err = register_tcp_functions_as_names(&__tcp_rack, M_WAITOK,
14937 		    rack_stack_names, &num_stacks);
14938 		if (err) {
14939 			printf("Failed to register %s stack name for "
14940 			    "%s module\n", rack_stack_names[num_stacks],
14941 			    __XSTRING(MODNAME));
14942 			sysctl_ctx_free(&rack_sysctl_ctx);
14943 free_uma:
14944 			uma_zdestroy(rack_zone);
14945 			uma_zdestroy(rack_pcb_zone);
14946 			rack_counter_destroy();
14947 			printf("Failed to register rack module -- err:%d\n", err);
14948 			return (err);
14949 		}
14950 		tcp_lro_reg_mbufq();
14951 		rack_mod_inited = true;
14952 		break;
14953 	case MOD_QUIESCE:
14954 		err = deregister_tcp_functions(&__tcp_rack, true, false);
14955 		break;
14956 	case MOD_UNLOAD:
14957 		err = deregister_tcp_functions(&__tcp_rack, false, true);
14958 		if (err == EBUSY)
14959 			break;
14960 		if (rack_mod_inited) {
14961 			uma_zdestroy(rack_zone);
14962 			uma_zdestroy(rack_pcb_zone);
14963 			sysctl_ctx_free(&rack_sysctl_ctx);
14964 			rack_counter_destroy();
14965 			rack_mod_inited = false;
14966 		}
14967 		tcp_lro_dereg_mbufq();
14968 		err = 0;
14969 		break;
14970 	default:
14971 		return (EOPNOTSUPP);
14972 	}
14973 	return (err);
14974 }
14975 
14976 static moduledata_t tcp_rack = {
14977 	.name = __XSTRING(MODNAME),
14978 	.evhand = tcp_addrack,
14979 	.priv = 0
14980 };
14981 
14982 MODULE_VERSION(MODNAME, 1);
14983 DECLARE_MODULE(MODNAME, tcp_rack, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);
14984 MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1);
14985