1 /*-
2 * SPDX-License-Identifier: BSD-2-Clause
3 *
4 * Copyright (c) 2009-2010
5 * Swinburne University of Technology, Melbourne, Australia
6 * Copyright (c) 2010 Lawrence Stewart <[email protected]>
7 * Copyright (c) 2010-2011 The FreeBSD Foundation
8 * All rights reserved.
9 *
10 * This software was developed at the Centre for Advanced Internet
11 * Architectures, Swinburne University of Technology, by David Hayes, made
12 * possible in part by a grant from the Cisco University Research Program Fund
13 * at Community Foundation Silicon Valley.
14 *
15 * Portions of this software were developed at the Centre for Advanced
16 * Internet Architectures, Swinburne University of Technology, Melbourne,
17 * Australia by David Hayes under sponsorship from the FreeBSD Foundation.
18 *
19 * Redistribution and use in source and binary forms, with or without
20 * modification, are permitted provided that the following conditions
21 * are met:
22 * 1. Redistributions of source code must retain the above copyright
23 * notice, this list of conditions and the following disclaimer.
24 * 2. Redistributions in binary form must reproduce the above copyright
25 * notice, this list of conditions and the following disclaimer in the
26 * documentation and/or other materials provided with the distribution.
27 *
28 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
29 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
30 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
31 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
32 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
33 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
34 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
35 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
36 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
37 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
38 * SUCH DAMAGE.
39 */
40
41 #include <sys/cdefs.h>
42 #include <sys/param.h>
43 #include <sys/kernel.h>
44 #include <sys/mbuf.h>
45 #include <sys/module.h>
46 #include <sys/hhook.h>
47 #include <sys/khelp.h>
48 #include <sys/module_khelp.h>
49 #include <sys/socket.h>
50 #include <sys/sockopt.h>
51
52 #include <net/vnet.h>
53
54 #include <netinet/in.h>
55 #include <netinet/in_pcb.h>
56 #include <netinet/tcp_seq.h>
57 #include <netinet/tcp_var.h>
58
59 #include <netinet/khelp/h_ertt.h>
60
61 #include <vm/uma.h>
62
63 uma_zone_t txseginfo_zone;
64
65 /* Smoothing factor for delayed ack guess. */
66 #define DLYACK_SMOOTH 5
67
68 /* Max number of time stamp errors allowed in a session. */
69 #define MAX_TS_ERR 10
70
71 static int ertt_packet_measurement_hook(int hhook_type, int hhook_id,
72 void *udata, void *ctx_data, void *hdata, struct osd *hosd);
73 static int ertt_add_tx_segment_info_hook(int hhook_type, int hhook_id,
74 void *udata, void *ctx_data, void *hdata, struct osd *hosd);
75 static int ertt_mod_init(void);
76 static int ertt_mod_destroy(void);
77 static int ertt_uma_ctor(void *mem, int size, void *arg, int flags);
78 static void ertt_uma_dtor(void *mem, int size, void *arg);
79
80 /*
81 * Contains information about the sent segment for comparison with the
82 * corresponding ack.
83 */
84 struct txseginfo {
85 /* Segment length. */
86 uint32_t len;
87 /* Segment sequence number. */
88 tcp_seq seq;
89 /* Time stamp indicating when the packet was sent. */
90 uint32_t tx_ts;
91 /* Last received receiver ts (if the TCP option is used). */
92 uint32_t rx_ts;
93 uint32_t flags;
94 TAILQ_ENTRY (txseginfo) txsegi_lnk;
95 };
96
97 /* Flags for struct txseginfo. */
98 #define TXSI_TSO 0x01 /* TSO was used for this entry. */
99 #define TXSI_RTT_MEASURE_START 0x02 /* Start a per RTT measurement. */
100 #define TXSI_RX_MEASURE_END 0x04 /* Measure the rx rate until this txsi. */
101
102 struct helper ertt_helper = {
103 .mod_init = ertt_mod_init,
104 .mod_destroy = ertt_mod_destroy,
105 .h_flags = HELPER_NEEDS_OSD,
106 .h_classes = HELPER_CLASS_TCP
107 };
108
109 /* Define the helper hook info required by ERTT. */
110 struct hookinfo ertt_hooks[] = {
111 {
112 .hook_type = HHOOK_TYPE_TCP,
113 .hook_id = HHOOK_TCP_EST_IN,
114 .hook_udata = NULL,
115 .hook_func = &ertt_packet_measurement_hook
116 },
117 {
118 .hook_type = HHOOK_TYPE_TCP,
119 .hook_id = HHOOK_TCP_EST_OUT,
120 .hook_udata = NULL,
121 .hook_func = &ertt_add_tx_segment_info_hook
122 }
123 };
124
125 /* Flags to indicate how marked_packet_rtt should handle this txsi. */
126 #define MULTI_ACK 0x01 /* More than this txsi is acked. */
127 #define OLD_TXSI 0x02 /* TXSI is old according to timestamps. */
128 #define CORRECT_ACK 0X04 /* Acks this TXSI. */
129 #define FORCED_MEASUREMENT 0X08 /* Force an RTT measurement. */
130
131 /*
132 * This fuction measures the RTT of a particular segment/ack pair, or the next
133 * closest if this will yield an inaccurate result due to delayed acking or
134 * other issues.
135 */
136 static void inline
marked_packet_rtt(struct txseginfo * txsi,struct ertt * e_t,struct tcpcb * tp,uint32_t * pmeasurenext,int * pmeasurenext_len,int * prtt_bytes_adjust,int mflag)137 marked_packet_rtt(struct txseginfo *txsi, struct ertt *e_t, struct tcpcb *tp,
138 uint32_t *pmeasurenext, int *pmeasurenext_len, int *prtt_bytes_adjust,
139 int mflag)
140 {
141
142 /*
143 * If we can't measure this one properly due to delayed acking adjust
144 * byte counters and flag to measure next txsi. Note that since the
145 * marked packet's transmitted bytes are measured we need to subtract the
146 * transmitted bytes. Then pretend the next txsi was marked.
147 */
148 if (mflag & (MULTI_ACK|OLD_TXSI)) {
149 *pmeasurenext = txsi->tx_ts;
150 *pmeasurenext_len = txsi->len;
151 *prtt_bytes_adjust += *pmeasurenext_len;
152 } else {
153 if (mflag & FORCED_MEASUREMENT) {
154 e_t->markedpkt_rtt = tcp_ts_getticks() -
155 *pmeasurenext + 1;
156 e_t->bytes_tx_in_marked_rtt = e_t->bytes_tx_in_rtt +
157 *pmeasurenext_len - *prtt_bytes_adjust;
158 } else {
159 e_t->markedpkt_rtt = tcp_ts_getticks() -
160 txsi->tx_ts + 1;
161 e_t->bytes_tx_in_marked_rtt = e_t->bytes_tx_in_rtt -
162 *prtt_bytes_adjust;
163 }
164 e_t->marked_snd_cwnd = tp->snd_cwnd;
165
166 /*
167 * Reset the ERTT_MEASUREMENT_IN_PROGRESS flag to indicate to
168 * add_tx_segment_info that a new measurement should be started.
169 */
170 e_t->flags &= ~ERTT_MEASUREMENT_IN_PROGRESS;
171 /*
172 * Set ERTT_NEW_MEASUREMENT to tell the congestion control
173 * algorithm that a new marked RTT measurement has has been made
174 * and is available for use.
175 */
176 e_t->flags |= ERTT_NEW_MEASUREMENT;
177
178 if (tp->t_flags & TF_TSO) {
179 /* Temporarily disable TSO to aid a new measurement. */
180 tp->t_flags &= ~TF_TSO;
181 /* Keep track that we've disabled it. */
182 e_t->flags |= ERTT_TSO_DISABLED;
183 }
184 }
185 }
186
187 /*
188 * Ertt_packet_measurements uses a small amount of state kept on each packet
189 * sent to match incoming acknowledgements. This enables more accurate and
190 * secure round trip time measurements. The resulting measurement is used for
191 * congestion control algorithms which require a more accurate time.
192 * Ertt_packet_measurements is called via the helper hook in tcp_input.c
193 */
194 static int
ertt_packet_measurement_hook(int hhook_type,int hhook_id,void * udata,void * ctx_data,void * hdata,struct osd * hosd)195 ertt_packet_measurement_hook(int hhook_type, int hhook_id, void *udata,
196 void *ctx_data, void *hdata, struct osd *hosd)
197 {
198 struct ertt *e_t;
199 struct tcpcb *tp;
200 struct tcphdr *th;
201 struct tcpopt *to;
202 struct tcp_hhook_data *thdp;
203 struct txseginfo *txsi;
204 int acked, measurenext_len, multiack, new_sacked_bytes, rtt_bytes_adjust;
205 uint32_t measurenext, rts;
206 tcp_seq ack;
207
208 KASSERT(ctx_data != NULL, ("%s: ctx_data is NULL!", __func__));
209 KASSERT(hdata != NULL, ("%s: hdata is NULL!", __func__));
210
211 e_t = (struct ertt *)hdata;
212 thdp = ctx_data;
213 tp = thdp->tp;
214 th = thdp->th;
215 to = thdp->to;
216 new_sacked_bytes = (tp->sackhint.last_sack_ack != 0);
217 measurenext = measurenext_len = multiack = rts = rtt_bytes_adjust = 0;
218 acked = th->th_ack - tp->snd_una;
219
220 INP_WLOCK_ASSERT(tptoinpcb(tp));
221
222 /* Packet has provided new acknowledgements. */
223 if (acked > 0 || new_sacked_bytes) {
224 if (acked == 0 && new_sacked_bytes) {
225 /* Use last sacked data. */
226 ack = tp->sackhint.last_sack_ack;
227 } else
228 ack = th->th_ack;
229
230 txsi = TAILQ_FIRST(&e_t->txsegi_q);
231 while (txsi != NULL) {
232 rts = 0;
233
234 /* Acknowledgement is acking more than this txsi. */
235 if (SEQ_GT(ack, txsi->seq + txsi->len)) {
236 if (txsi->flags & TXSI_RTT_MEASURE_START ||
237 measurenext) {
238 marked_packet_rtt(txsi, e_t, tp,
239 &measurenext, &measurenext_len,
240 &rtt_bytes_adjust, MULTI_ACK);
241 }
242 TAILQ_REMOVE(&e_t->txsegi_q, txsi, txsegi_lnk);
243 uma_zfree(txseginfo_zone, txsi);
244 txsi = TAILQ_FIRST(&e_t->txsegi_q);
245 continue;
246 }
247
248 /*
249 * Guess if delayed acks are being used by the receiver.
250 *
251 * XXXDH: A simple heuristic that could be improved
252 */
253 if (!new_sacked_bytes) {
254 if (acked > tp->t_maxseg) {
255 e_t->dlyack_rx +=
256 (e_t->dlyack_rx < DLYACK_SMOOTH) ?
257 1 : 0;
258 multiack = 1;
259 } else if (acked > txsi->len) {
260 multiack = 1;
261 e_t->dlyack_rx +=
262 (e_t->dlyack_rx < DLYACK_SMOOTH) ?
263 1 : 0;
264 } else if (acked == tp->t_maxseg ||
265 acked == txsi->len) {
266 e_t->dlyack_rx -=
267 (e_t->dlyack_rx > 0) ? 1 : 0;
268 }
269 /* Otherwise leave dlyack_rx the way it was. */
270 }
271
272 /*
273 * Time stamps are only to help match the txsi with the
274 * received acknowledgements.
275 */
276 if (e_t->timestamp_errors < MAX_TS_ERR &&
277 (to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
278 /*
279 * Note: All packets sent with the offload will
280 * have the same time stamp. If we are sending
281 * on a fast interface and the t_maxseg is much
282 * smaller than one tick, this will be fine. The
283 * time stamp would be the same whether we were
284 * using tso or not. However, if the interface
285 * is slow, this will cause problems with the
286 * calculations. If the interface is slow, there
287 * is not reason to be using tso, and it should
288 * be turned off.
289 */
290 /*
291 * If there are too many time stamp errors, time
292 * stamps won't be trusted
293 */
294 rts = to->to_tsecr;
295 /* Before this packet. */
296 if (!e_t->dlyack_rx && TSTMP_LT(rts, txsi->tx_ts))
297 /* When delayed acking is used, the
298 * reflected time stamp is of the first
299 * packet and thus may be before
300 * txsi->tx_ts.
301 */
302 break;
303 if (TSTMP_GT(rts, txsi->tx_ts)) {
304 /*
305 * If reflected time stamp is later than
306 * tx_tsi, then this txsi is old.
307 */
308 if (txsi->flags & TXSI_RTT_MEASURE_START
309 || measurenext) {
310 marked_packet_rtt(txsi, e_t, tp,
311 &measurenext, &measurenext_len,
312 &rtt_bytes_adjust, OLD_TXSI);
313 }
314 TAILQ_REMOVE(&e_t->txsegi_q, txsi,
315 txsegi_lnk);
316 uma_zfree(txseginfo_zone, txsi);
317 txsi = TAILQ_FIRST(&e_t->txsegi_q);
318 continue;
319 }
320 if (rts == txsi->tx_ts &&
321 TSTMP_LT(to->to_tsval, txsi->rx_ts)) {
322 /*
323 * Segment received before sent!
324 * Something is wrong with the received
325 * timestamps so increment errors. If
326 * this keeps up we will ignore
327 * timestamps.
328 */
329 e_t->timestamp_errors++;
330 }
331 }
332 /*
333 * Acknowledging a sequence number before this txsi.
334 * If it is an old txsi that may have had the same seq
335 * numbers, it should have been removed if time stamps
336 * are being used.
337 */
338 if (SEQ_LEQ(ack, txsi->seq))
339 break; /* Before first packet in txsi. */
340
341 /*
342 * Only ack > txsi->seq and ack <= txsi->seq+txsi->len
343 * past this point.
344 *
345 * If delayed acks are being used, an acknowledgement
346 * for a single segment will have been delayed by the
347 * receiver and will yield an inaccurate measurement. In
348 * this case, we only make the measurement if more than
349 * one segment is being acknowledged or sack is
350 * currently being used.
351 */
352 if (!e_t->dlyack_rx || multiack || new_sacked_bytes) {
353 /* Make an accurate new measurement. */
354 e_t->rtt = tcp_ts_getticks() - txsi->tx_ts + 1;
355
356 if (e_t->rtt < e_t->minrtt || e_t->minrtt == 0)
357 e_t->minrtt = e_t->rtt;
358
359 if (e_t->rtt > e_t->maxrtt || e_t->maxrtt == 0)
360 e_t->maxrtt = e_t->rtt;
361 }
362
363 if (txsi->flags & TXSI_RTT_MEASURE_START || measurenext)
364 marked_packet_rtt(txsi, e_t, tp,
365 &measurenext, &measurenext_len,
366 &rtt_bytes_adjust, CORRECT_ACK);
367
368 if (txsi->flags & TXSI_TSO) {
369 if (txsi->len > acked) {
370 txsi->len -= acked;
371 /*
372 * This presumes ack for first bytes in
373 * txsi, this may not be true but it
374 * shouldn't cause problems for the
375 * timing.
376 *
377 * We remeasure RTT even though we only
378 * have a single txsi. The rationale
379 * behind this is that it is better to
380 * have a slightly inaccurate
381 * measurement than no additional
382 * measurement for the rest of the bulk
383 * transfer. Since TSO is only used on
384 * high speed interface cards, so the
385 * packets should be transmitted at line
386 * rate back to back with little
387 * difference in transmission times (in
388 * ticks).
389 */
390 txsi->seq += acked;
391 /*
392 * Reset txsi measure flag so we don't
393 * use it for another RTT measurement.
394 */
395 txsi->flags &= ~TXSI_RTT_MEASURE_START;
396 /*
397 * There is still more data to be acked
398 * from tso bulk transmission, so we
399 * won't remove it from the TAILQ yet.
400 */
401 break;
402 }
403 txsi->len = 0;
404 }
405
406 TAILQ_REMOVE(&e_t->txsegi_q, txsi, txsegi_lnk);
407 uma_zfree(txseginfo_zone, txsi);
408 break;
409 }
410
411 if (measurenext) {
412 /*
413 * We need to do a RTT measurement. It won't be the best
414 * if we do it here.
415 */
416 marked_packet_rtt(txsi, e_t, tp,
417 &measurenext, &measurenext_len,
418 &rtt_bytes_adjust, FORCED_MEASUREMENT);
419 }
420 }
421
422 return (0);
423 }
424
425 /*
426 * Add information about a transmitted segment to a list.
427 * This is called via the helper hook in tcp_output.c
428 */
429 static int
ertt_add_tx_segment_info_hook(int hhook_type,int hhook_id,void * udata,void * ctx_data,void * hdata,struct osd * hosd)430 ertt_add_tx_segment_info_hook(int hhook_type, int hhook_id, void *udata,
431 void *ctx_data, void *hdata, struct osd *hosd)
432 {
433 struct ertt *e_t;
434 struct tcpcb *tp;
435 struct tcphdr *th;
436 struct tcpopt *to;
437 struct tcp_hhook_data *thdp;
438 struct txseginfo *txsi;
439 uint32_t len;
440 int tso;
441
442 KASSERT(ctx_data != NULL, ("%s: ctx_data is NULL!", __func__));
443 KASSERT(hdata != NULL, ("%s: hdata is NULL!", __func__));
444
445 e_t = (struct ertt *)hdata;
446 thdp = ctx_data;
447 tp = thdp->tp;
448 th = thdp->th;
449 to = thdp->to;
450 len = thdp->len;
451 tso = thdp->tso;
452
453 INP_WLOCK_ASSERT(tptoinpcb(tp));
454
455 if (len > 0) {
456 txsi = uma_zalloc(txseginfo_zone, M_NOWAIT);
457 if (txsi != NULL) {
458 /* Construct txsi setting the necessary flags. */
459 txsi->flags = 0; /* Needs to be initialised. */
460 txsi->seq = ntohl(th->th_seq);
461 txsi->len = len;
462 if (tso)
463 txsi->flags |= TXSI_TSO;
464 else if (e_t->flags & ERTT_TSO_DISABLED) {
465 tp->t_flags |= TF_TSO;
466 e_t->flags &= ~ERTT_TSO_DISABLED;
467 }
468
469 if (e_t->flags & ERTT_MEASUREMENT_IN_PROGRESS) {
470 e_t->bytes_tx_in_rtt += len;
471 } else {
472 txsi->flags |= TXSI_RTT_MEASURE_START;
473 e_t->flags |= ERTT_MEASUREMENT_IN_PROGRESS;
474 e_t->bytes_tx_in_rtt = len;
475 }
476
477 if (((tp->t_flags & TF_NOOPT) == 0) &&
478 (to->to_flags & TOF_TS)) {
479 txsi->tx_ts = ntohl(to->to_tsval) -
480 tp->ts_offset;
481 txsi->rx_ts = ntohl(to->to_tsecr);
482 } else {
483 txsi->tx_ts = tcp_ts_getticks();
484 txsi->rx_ts = 0; /* No received time stamp. */
485 }
486 TAILQ_INSERT_TAIL(&e_t->txsegi_q, txsi, txsegi_lnk);
487 }
488 }
489
490 return (0);
491 }
492
493 static int
ertt_mod_init(void)494 ertt_mod_init(void)
495 {
496
497 txseginfo_zone = uma_zcreate("ertt_txseginfo", sizeof(struct txseginfo),
498 NULL, NULL, NULL, NULL, 0, 0);
499
500 return (0);
501 }
502
503 static int
ertt_mod_destroy(void)504 ertt_mod_destroy(void)
505 {
506
507 uma_zdestroy(txseginfo_zone);
508
509 return (0);
510 }
511
512 static int
ertt_uma_ctor(void * mem,int size,void * arg,int flags)513 ertt_uma_ctor(void *mem, int size, void *arg, int flags)
514 {
515 struct ertt *e_t;
516
517 e_t = mem;
518
519 TAILQ_INIT(&e_t->txsegi_q);
520 e_t->timestamp_errors = 0;
521 e_t->minrtt = 0;
522 e_t->maxrtt = 0;
523 e_t->rtt = 0;
524 e_t->flags = 0;
525 e_t->dlyack_rx = 0;
526 e_t->bytes_tx_in_rtt = 0;
527 e_t->markedpkt_rtt = 0;
528
529 return (0);
530 }
531
532 static void
ertt_uma_dtor(void * mem,int size,void * arg)533 ertt_uma_dtor(void *mem, int size, void *arg)
534 {
535 struct ertt *e_t;
536 struct txseginfo *n_txsi, *txsi;
537
538 e_t = mem;
539 txsi = TAILQ_FIRST(&e_t->txsegi_q);
540 while (txsi != NULL) {
541 n_txsi = TAILQ_NEXT(txsi, txsegi_lnk);
542 uma_zfree(txseginfo_zone, txsi);
543 txsi = n_txsi;
544 }
545 }
546
547 KHELP_DECLARE_MOD_UMA(ertt, &ertt_helper, ertt_hooks, 1, sizeof(struct ertt),
548 ertt_uma_ctor, ertt_uma_dtor);
549