1 /*-
2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3 *
4 * Copyright (c) 2009-2013
5 * Swinburne University of Technology, Melbourne, Australia
6 * All rights reserved.
7 *
8 * This software was developed at the Centre for Advanced Internet
9 * Architectures, Swinburne University of Technology, by David Hayes, made
10 * possible in part by a gift from The Cisco University Research Program Fund,
11 * a corporate advised fund of Silicon Valley Community Foundation. Development
12 * and testing were further assisted by a grant from the FreeBSD Foundation.
13 *
14 * Redistribution and use in source and binary forms, with or without
15 * modification, are permitted provided that the following conditions
16 * are met:
17 * 1. Redistributions of source code must retain the above copyright
18 * notice, this list of conditions and the following disclaimer.
19 * 2. Redistributions in binary form must reproduce the above copyright
20 * notice, this list of conditions and the following disclaimer in the
21 * documentation and/or other materials provided with the distribution.
22 *
23 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
24 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
25 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
26 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
27 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
28 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
29 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
30 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
31 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
32 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33 * SUCH DAMAGE.
34 */
35
36 /*
37 * CAIA Delay-Gradient (CDG) congestion control algorithm
38 *
39 * An implemention of the delay-gradient congestion control algorithm proposed
40 * in the following paper:
41 *
42 * D. A. Hayes and G. Armitage, "Revisiting TCP Congestion Control using Delay
43 * Gradients", in IFIP Networking, Valencia, Spain, 9-13 May 2011.
44 *
45 * Developed as part of the NewTCP research project at Swinburne University of
46 * Technology's Centre for Advanced Internet Architectures, Melbourne,
47 * Australia. More details are available at:
48 * http://caia.swin.edu.au/urp/newtcp/
49 */
50
51 #include <sys/cdefs.h>
52 __FBSDID("$FreeBSD$");
53
54 #include <sys/param.h>
55 #include <sys/hhook.h>
56 #include <sys/kernel.h>
57 #include <sys/khelp.h>
58 #include <sys/limits.h>
59 #include <sys/lock.h>
60 #include <sys/malloc.h>
61 #include <sys/module.h>
62 #include <sys/queue.h>
63 #include <sys/socket.h>
64 #include <sys/socketvar.h>
65 #include <sys/sysctl.h>
66 #include <sys/systm.h>
67
68 #include <net/vnet.h>
69
70 #include <netinet/tcp.h>
71 #include <netinet/tcp_seq.h>
72 #include <netinet/tcp_timer.h>
73 #include <netinet/tcp_var.h>
74 #include <netinet/cc/cc.h>
75 #include <netinet/cc/cc_module.h>
76
77 #include <netinet/khelp/h_ertt.h>
78
79 #include <vm/uma.h>
80
81 #define CDG_VERSION "0.1"
82
83 /* Private delay-gradient induced congestion control signal. */
84 #define CC_CDG_DELAY 0x01000000
85
86 /* NewReno window deflation factor on loss (as a percentage). */
87 #define RENO_BETA 50
88
89 /* Queue states. */
90 #define CDG_Q_EMPTY 1
91 #define CDG_Q_RISING 2
92 #define CDG_Q_FALLING 3
93 #define CDG_Q_FULL 4
94 #define CDG_Q_UNKNOWN 9999
95
96 /* Number of bit shifts used in probexp lookup table. */
97 #define EXP_PREC 15
98
99 /* Largest gradient represented in probexp lookup table. */
100 #define MAXGRAD 5
101
102 /*
103 * Delay Precision Enhance - number of bit shifts used for qtrend related
104 * integer arithmetic precision.
105 */
106 #define D_P_E 7
107
108 struct qdiff_sample {
109 long qdiff;
110 STAILQ_ENTRY(qdiff_sample) qdiff_lnk;
111 };
112
113 struct cdg {
114 long max_qtrend;
115 long min_qtrend;
116 STAILQ_HEAD(minrtts_head, qdiff_sample) qdiffmin_q;
117 STAILQ_HEAD(maxrtts_head, qdiff_sample) qdiffmax_q;
118 long window_incr;
119 /* rttcount for window increase when in congestion avoidance */
120 long rtt_count;
121 /* maximum measured rtt within an rtt period */
122 int maxrtt_in_rtt;
123 /* maximum measured rtt within prev rtt period */
124 int maxrtt_in_prevrtt;
125 /* minimum measured rtt within an rtt period */
126 int minrtt_in_rtt;
127 /* minimum measured rtt within prev rtt period */
128 int minrtt_in_prevrtt;
129 /* consecutive congestion episode counter */
130 uint32_t consec_cong_cnt;
131 /* when tracking a new reno type loss window */
132 uint32_t shadow_w;
133 /* maximum number of samples in the moving average queue */
134 int sample_q_size;
135 /* number of samples in the moving average queue */
136 int num_samples;
137 /* estimate of the queue state of the path */
138 int queue_state;
139 };
140
141 /*
142 * Lookup table for:
143 * (1 - exp(-x)) << EXP_PREC, where x = [0,MAXGRAD] in 2^-7 increments
144 *
145 * Note: probexp[0] is set to 10 (not 0) as a safety for very low increase
146 * gradients.
147 */
148 static const int probexp[641] = {
149 10,255,508,759,1008,1255,1501,1744,1985,2225,2463,2698,2932,3165,3395,3624,
150 3850,4075,4299,4520,4740,4958,5175,5389,5602,5814,6024,6232,6438,6643,6846,
151 7048,7248,7447,7644,7839,8033,8226,8417,8606,8794,8981,9166,9350,9532,9713,
152 9892,10070,10247,10422,10596,10769,10940,11110,11278,11445,11611,11776,11939,
153 12101,12262,12422,12580,12737,12893,13048,13201,13354,13505,13655,13803,13951,
154 14097,14243,14387,14530,14672,14813,14952,15091,15229,15365,15500,15635,15768,
155 15900,16032,16162,16291,16419,16547,16673,16798,16922,17046,17168,17289,17410,
156 17529,17648,17766,17882,17998,18113,18227,18340,18453,18564,18675,18784,18893,
157 19001,19108,19215,19320,19425,19529,19632,19734,19835,19936,20036,20135,20233,
158 20331,20427,20523,20619,20713,20807,20900,20993,21084,21175,21265,21355,21444,
159 21532,21619,21706,21792,21878,21962,22046,22130,22213,22295,22376,22457,22537,
160 22617,22696,22774,22852,22929,23006,23082,23157,23232,23306,23380,23453,23525,
161 23597,23669,23739,23810,23879,23949,24017,24085,24153,24220,24286,24352,24418,
162 24483,24547,24611,24675,24738,24800,24862,24924,24985,25045,25106,25165,25224,
163 25283,25341,25399,25456,25513,25570,25626,25681,25737,25791,25846,25899,25953,
164 26006,26059,26111,26163,26214,26265,26316,26366,26416,26465,26514,26563,26611,
165 26659,26707,26754,26801,26847,26893,26939,26984,27029,27074,27118,27162,27206,
166 27249,27292,27335,27377,27419,27460,27502,27543,27583,27624,27664,27703,27743,
167 27782,27821,27859,27897,27935,27973,28010,28047,28084,28121,28157,28193,28228,
168 28263,28299,28333,28368,28402,28436,28470,28503,28536,28569,28602,28634,28667,
169 28699,28730,28762,28793,28824,28854,28885,28915,28945,28975,29004,29034,29063,
170 29092,29120,29149,29177,29205,29232,29260,29287,29314,29341,29368,29394,29421,
171 29447,29472,29498,29524,29549,29574,29599,29623,29648,29672,29696,29720,29744,
172 29767,29791,29814,29837,29860,29882,29905,29927,29949,29971,29993,30014,30036,
173 30057,30078,30099,30120,30141,30161,30181,30201,30221,30241,30261,30280,30300,
174 30319,30338,30357,30376,30394,30413,30431,30449,30467,30485,30503,30521,30538,
175 30555,30573,30590,30607,30624,30640,30657,30673,30690,30706,30722,30738,30753,
176 30769,30785,30800,30815,30831,30846,30861,30876,30890,30905,30919,30934,30948,
177 30962,30976,30990,31004,31018,31031,31045,31058,31072,31085,31098,31111,31124,
178 31137,31149,31162,31174,31187,31199,31211,31223,31235,31247,31259,31271,31283,
179 31294,31306,31317,31328,31339,31351,31362,31373,31383,31394,31405,31416,31426,
180 31436,31447,31457,31467,31477,31487,31497,31507,31517,31527,31537,31546,31556,
181 31565,31574,31584,31593,31602,31611,31620,31629,31638,31647,31655,31664,31673,
182 31681,31690,31698,31706,31715,31723,31731,31739,31747,31755,31763,31771,31778,
183 31786,31794,31801,31809,31816,31824,31831,31838,31846,31853,31860,31867,31874,
184 31881,31888,31895,31902,31908,31915,31922,31928,31935,31941,31948,31954,31960,
185 31967,31973,31979,31985,31991,31997,32003,32009,32015,32021,32027,32033,32038,
186 32044,32050,32055,32061,32066,32072,32077,32083,32088,32093,32098,32104,32109,
187 32114,32119,32124,32129,32134,32139,32144,32149,32154,32158,32163,32168,32173,
188 32177,32182,32186,32191,32195,32200,32204,32209,32213,32217,32222,32226,32230,
189 32234,32238,32242,32247,32251,32255,32259,32263,32267,32270,32274,32278,32282,
190 32286,32290,32293,32297,32301,32304,32308,32311,32315,32318,32322,32325,32329,
191 32332,32336,32339,32342,32346,32349,32352,32356,32359,32362,32365,32368,32371,
192 32374,32377,32381,32384,32387,32389,32392,32395,32398,32401,32404,32407,32410,
193 32412,32415,32418,32421,32423,32426,32429,32431,32434,32437,32439,32442,32444,
194 32447,32449,32452,32454,32457,32459,32461,32464,32466,32469,32471,32473,32476,
195 32478,32480,32482,32485,32487,32489,32491,32493,32495,32497,32500,32502,32504,
196 32506,32508,32510,32512,32514,32516,32518,32520,32522,32524,32526,32527,32529,
197 32531,32533,32535,32537,32538,32540,32542,32544,32545,32547};
198
199 static uma_zone_t qdiffsample_zone;
200
201 static MALLOC_DEFINE(M_CDG, "cdg data",
202 "Per connection data required for the CDG congestion control algorithm");
203
204 static int ertt_id;
205
206 VNET_DEFINE_STATIC(uint32_t, cdg_alpha_inc);
207 VNET_DEFINE_STATIC(uint32_t, cdg_beta_delay);
208 VNET_DEFINE_STATIC(uint32_t, cdg_beta_loss);
209 VNET_DEFINE_STATIC(uint32_t, cdg_smoothing_factor);
210 VNET_DEFINE_STATIC(uint32_t, cdg_exp_backoff_scale);
211 VNET_DEFINE_STATIC(uint32_t, cdg_consec_cong);
212 VNET_DEFINE_STATIC(uint32_t, cdg_hold_backoff);
213 #define V_cdg_alpha_inc VNET(cdg_alpha_inc)
214 #define V_cdg_beta_delay VNET(cdg_beta_delay)
215 #define V_cdg_beta_loss VNET(cdg_beta_loss)
216 #define V_cdg_smoothing_factor VNET(cdg_smoothing_factor)
217 #define V_cdg_exp_backoff_scale VNET(cdg_exp_backoff_scale)
218 #define V_cdg_consec_cong VNET(cdg_consec_cong)
219 #define V_cdg_hold_backoff VNET(cdg_hold_backoff)
220
221 /* Function prototypes. */
222 static int cdg_mod_init(void);
223 static int cdg_mod_destroy(void);
224 static void cdg_conn_init(struct cc_var *ccv);
225 static int cdg_cb_init(struct cc_var *ccv);
226 static void cdg_cb_destroy(struct cc_var *ccv);
227 static void cdg_cong_signal(struct cc_var *ccv, uint32_t signal_type);
228 static void cdg_ack_received(struct cc_var *ccv, uint16_t ack_type);
229
230 struct cc_algo cdg_cc_algo = {
231 .name = "cdg",
232 .mod_init = cdg_mod_init,
233 .ack_received = cdg_ack_received,
234 .cb_destroy = cdg_cb_destroy,
235 .cb_init = cdg_cb_init,
236 .conn_init = cdg_conn_init,
237 .cong_signal = cdg_cong_signal,
238 .mod_destroy = cdg_mod_destroy
239 };
240
241 /* Vnet created and being initialised. */
242 static void
cdg_init_vnet(const void * unused __unused)243 cdg_init_vnet(const void *unused __unused)
244 {
245
246 V_cdg_alpha_inc = 0;
247 V_cdg_beta_delay = 70;
248 V_cdg_beta_loss = 50;
249 V_cdg_smoothing_factor = 8;
250 V_cdg_exp_backoff_scale = 3;
251 V_cdg_consec_cong = 5;
252 V_cdg_hold_backoff = 5;
253 }
254
255 static int
cdg_mod_init(void)256 cdg_mod_init(void)
257 {
258 VNET_ITERATOR_DECL(v);
259
260 ertt_id = khelp_get_id("ertt");
261 if (ertt_id <= 0)
262 return (EINVAL);
263
264 qdiffsample_zone = uma_zcreate("cdg_qdiffsample",
265 sizeof(struct qdiff_sample), NULL, NULL, NULL, NULL, 0, 0);
266
267 VNET_LIST_RLOCK();
268 VNET_FOREACH(v) {
269 CURVNET_SET(v);
270 cdg_init_vnet(NULL);
271 CURVNET_RESTORE();
272 }
273 VNET_LIST_RUNLOCK();
274
275 cdg_cc_algo.post_recovery = newreno_cc_algo.post_recovery;
276 cdg_cc_algo.after_idle = newreno_cc_algo.after_idle;
277
278 return (0);
279 }
280
281 static int
cdg_mod_destroy(void)282 cdg_mod_destroy(void)
283 {
284
285 uma_zdestroy(qdiffsample_zone);
286 return (0);
287 }
288
289 static int
cdg_cb_init(struct cc_var * ccv)290 cdg_cb_init(struct cc_var *ccv)
291 {
292 struct cdg *cdg_data;
293
294 cdg_data = malloc(sizeof(struct cdg), M_CDG, M_NOWAIT);
295 if (cdg_data == NULL)
296 return (ENOMEM);
297
298 cdg_data->shadow_w = 0;
299 cdg_data->max_qtrend = 0;
300 cdg_data->min_qtrend = 0;
301 cdg_data->queue_state = CDG_Q_UNKNOWN;
302 cdg_data->maxrtt_in_rtt = 0;
303 cdg_data->maxrtt_in_prevrtt = 0;
304 cdg_data->minrtt_in_rtt = INT_MAX;
305 cdg_data->minrtt_in_prevrtt = 0;
306 cdg_data->window_incr = 0;
307 cdg_data->rtt_count = 0;
308 cdg_data->consec_cong_cnt = 0;
309 cdg_data->sample_q_size = V_cdg_smoothing_factor;
310 cdg_data->num_samples = 0;
311 STAILQ_INIT(&cdg_data->qdiffmin_q);
312 STAILQ_INIT(&cdg_data->qdiffmax_q);
313
314 ccv->cc_data = cdg_data;
315
316 return (0);
317 }
318
319 static void
cdg_conn_init(struct cc_var * ccv)320 cdg_conn_init(struct cc_var *ccv)
321 {
322 struct cdg *cdg_data = ccv->cc_data;
323
324 /*
325 * Initialise the shadow_cwnd in case we are competing with loss based
326 * flows from the start
327 */
328 cdg_data->shadow_w = CCV(ccv, snd_cwnd);
329 }
330
331 static void
cdg_cb_destroy(struct cc_var * ccv)332 cdg_cb_destroy(struct cc_var *ccv)
333 {
334 struct cdg *cdg_data;
335 struct qdiff_sample *qds, *qds_n;
336
337 cdg_data = ccv->cc_data;
338
339 qds = STAILQ_FIRST(&cdg_data->qdiffmin_q);
340 while (qds != NULL) {
341 qds_n = STAILQ_NEXT(qds, qdiff_lnk);
342 uma_zfree(qdiffsample_zone,qds);
343 qds = qds_n;
344 }
345
346 qds = STAILQ_FIRST(&cdg_data->qdiffmax_q);
347 while (qds != NULL) {
348 qds_n = STAILQ_NEXT(qds, qdiff_lnk);
349 uma_zfree(qdiffsample_zone,qds);
350 qds = qds_n;
351 }
352
353 free(ccv->cc_data, M_CDG);
354 }
355
356 static int
cdg_beta_handler(SYSCTL_HANDLER_ARGS)357 cdg_beta_handler(SYSCTL_HANDLER_ARGS)
358 {
359 int error;
360 uint32_t new;
361
362 new = *(uint32_t *)arg1;
363 error = sysctl_handle_int(oidp, &new, 0, req);
364 if (error == 0 && req->newptr != NULL) {
365 if (new == 0 || new > 100)
366 error = EINVAL;
367 else
368 *(uint32_t *)arg1 = new;
369 }
370
371 return (error);
372 }
373
374 static int
cdg_exp_backoff_scale_handler(SYSCTL_HANDLER_ARGS)375 cdg_exp_backoff_scale_handler(SYSCTL_HANDLER_ARGS)
376 {
377 int error;
378 uint32_t new;
379
380 new = *(uint32_t *)arg1;
381 error = sysctl_handle_int(oidp, &new, 0, req);
382 if (error == 0 && req->newptr != NULL) {
383 if (new < 1)
384 error = EINVAL;
385 else
386 *(uint32_t *)arg1 = new;
387 }
388
389 return (error);
390 }
391
392 static inline uint32_t
cdg_window_decrease(struct cc_var * ccv,unsigned long owin,unsigned int beta)393 cdg_window_decrease(struct cc_var *ccv, unsigned long owin, unsigned int beta)
394 {
395
396 return ((ulmin(CCV(ccv, snd_wnd), owin) * beta) / 100);
397 }
398
399 /*
400 * Window increase function
401 * This window increase function is independent of the initial window size
402 * to ensure small window flows are not discriminated against (i.e. fairness).
403 * It increases at 1pkt/rtt like Reno for alpha_inc rtts, and then 2pkts/rtt for
404 * the next alpha_inc rtts, etc.
405 */
406 static void
cdg_window_increase(struct cc_var * ccv,int new_measurement)407 cdg_window_increase(struct cc_var *ccv, int new_measurement)
408 {
409 struct cdg *cdg_data;
410 int incr, s_w_incr;
411
412 cdg_data = ccv->cc_data;
413 incr = s_w_incr = 0;
414
415 if (CCV(ccv, snd_cwnd) <= CCV(ccv, snd_ssthresh)) {
416 /* Slow start. */
417 incr = CCV(ccv, t_maxseg);
418 s_w_incr = incr;
419 cdg_data->window_incr = cdg_data->rtt_count = 0;
420 } else {
421 /* Congestion avoidance. */
422 if (new_measurement) {
423 s_w_incr = CCV(ccv, t_maxseg);
424 if (V_cdg_alpha_inc == 0) {
425 incr = CCV(ccv, t_maxseg);
426 } else {
427 if (++cdg_data->rtt_count >= V_cdg_alpha_inc) {
428 cdg_data->window_incr++;
429 cdg_data->rtt_count = 0;
430 }
431 incr = CCV(ccv, t_maxseg) *
432 cdg_data->window_incr;
433 }
434 }
435 }
436
437 if (cdg_data->shadow_w > 0)
438 cdg_data->shadow_w = ulmin(cdg_data->shadow_w + s_w_incr,
439 TCP_MAXWIN << CCV(ccv, snd_scale));
440
441 CCV(ccv, snd_cwnd) = ulmin(CCV(ccv, snd_cwnd) + incr,
442 TCP_MAXWIN << CCV(ccv, snd_scale));
443 }
444
445 static void
cdg_cong_signal(struct cc_var * ccv,uint32_t signal_type)446 cdg_cong_signal(struct cc_var *ccv, uint32_t signal_type)
447 {
448 struct cdg *cdg_data = ccv->cc_data;
449
450 switch(signal_type) {
451 case CC_CDG_DELAY:
452 CCV(ccv, snd_ssthresh) = cdg_window_decrease(ccv,
453 CCV(ccv, snd_cwnd), V_cdg_beta_delay);
454 CCV(ccv, snd_cwnd) = CCV(ccv, snd_ssthresh);
455 CCV(ccv, snd_recover) = CCV(ccv, snd_max);
456 cdg_data->window_incr = cdg_data->rtt_count = 0;
457 ENTER_CONGRECOVERY(CCV(ccv, t_flags));
458 break;
459 case CC_NDUPACK:
460 /*
461 * If already responding to congestion OR we have guessed no
462 * queue in the path is full.
463 */
464 if (IN_CONGRECOVERY(CCV(ccv, t_flags)) ||
465 cdg_data->queue_state < CDG_Q_FULL) {
466 CCV(ccv, snd_ssthresh) = CCV(ccv, snd_cwnd);
467 CCV(ccv, snd_recover) = CCV(ccv, snd_max);
468 } else {
469 /*
470 * Loss is likely to be congestion related. We have
471 * inferred a queue full state, so have shadow window
472 * react to loss as NewReno would.
473 */
474 if (cdg_data->shadow_w > 0)
475 cdg_data->shadow_w = cdg_window_decrease(ccv,
476 cdg_data->shadow_w, RENO_BETA);
477
478 CCV(ccv, snd_ssthresh) = max(cdg_data->shadow_w,
479 cdg_window_decrease(ccv, CCV(ccv, snd_cwnd),
480 V_cdg_beta_loss));
481
482 cdg_data->window_incr = cdg_data->rtt_count = 0;
483 }
484 ENTER_RECOVERY(CCV(ccv, t_flags));
485 break;
486 default:
487 newreno_cc_algo.cong_signal(ccv, signal_type);
488 break;
489 }
490 }
491
492 /*
493 * Using a negative exponential probabilistic backoff so that sources with
494 * varying RTTs which share the same link will, on average, have the same
495 * probability of backoff over time.
496 *
497 * Prob_backoff = 1 - exp(-qtrend / V_cdg_exp_backoff_scale), where
498 * V_cdg_exp_backoff_scale is the average qtrend for the exponential backoff.
499 */
500 static inline int
prob_backoff(long qtrend)501 prob_backoff(long qtrend)
502 {
503 int backoff, idx, p;
504
505 backoff = (qtrend > ((MAXGRAD * V_cdg_exp_backoff_scale) << D_P_E));
506
507 if (!backoff) {
508 if (V_cdg_exp_backoff_scale > 1)
509 idx = (qtrend + V_cdg_exp_backoff_scale / 2) /
510 V_cdg_exp_backoff_scale;
511 else
512 idx = qtrend;
513
514 /* Backoff probability proportional to rate of queue growth. */
515 p = (INT_MAX / (1 << EXP_PREC)) * probexp[idx];
516 backoff = (random() < p);
517 }
518
519 return (backoff);
520 }
521
522 static inline void
calc_moving_average(struct cdg * cdg_data,long qdiff_max,long qdiff_min)523 calc_moving_average(struct cdg *cdg_data, long qdiff_max, long qdiff_min)
524 {
525 struct qdiff_sample *qds;
526
527 ++cdg_data->num_samples;
528 if (cdg_data->num_samples > cdg_data->sample_q_size) {
529 /* Minimum RTT. */
530 qds = STAILQ_FIRST(&cdg_data->qdiffmin_q);
531 cdg_data->min_qtrend = cdg_data->min_qtrend +
532 (qdiff_min - qds->qdiff) / cdg_data->sample_q_size;
533 STAILQ_REMOVE_HEAD(&cdg_data->qdiffmin_q, qdiff_lnk);
534 qds->qdiff = qdiff_min;
535 STAILQ_INSERT_TAIL(&cdg_data->qdiffmin_q, qds, qdiff_lnk);
536
537 /* Maximum RTT. */
538 qds = STAILQ_FIRST(&cdg_data->qdiffmax_q);
539 cdg_data->max_qtrend = cdg_data->max_qtrend +
540 (qdiff_max - qds->qdiff) / cdg_data->sample_q_size;
541 STAILQ_REMOVE_HEAD(&cdg_data->qdiffmax_q, qdiff_lnk);
542 qds->qdiff = qdiff_max;
543 STAILQ_INSERT_TAIL(&cdg_data->qdiffmax_q, qds, qdiff_lnk);
544 --cdg_data->num_samples;
545 } else {
546 qds = uma_zalloc(qdiffsample_zone, M_NOWAIT);
547 if (qds != NULL) {
548 cdg_data->min_qtrend = cdg_data->min_qtrend +
549 qdiff_min / cdg_data->sample_q_size;
550 qds->qdiff = qdiff_min;
551 STAILQ_INSERT_TAIL(&cdg_data->qdiffmin_q, qds,
552 qdiff_lnk);
553 }
554
555 qds = uma_zalloc(qdiffsample_zone, M_NOWAIT);
556 if (qds) {
557 cdg_data->max_qtrend = cdg_data->max_qtrend +
558 qdiff_max / cdg_data->sample_q_size;
559 qds->qdiff = qdiff_max;
560 STAILQ_INSERT_TAIL(&cdg_data->qdiffmax_q, qds,
561 qdiff_lnk);
562 }
563 }
564 }
565
566 static void
cdg_ack_received(struct cc_var * ccv,uint16_t ack_type)567 cdg_ack_received(struct cc_var *ccv, uint16_t ack_type)
568 {
569 struct cdg *cdg_data;
570 struct ertt *e_t;
571 long qdiff_max, qdiff_min;
572 int congestion, new_measurement, slowstart;
573
574 cdg_data = ccv->cc_data;
575 e_t = (struct ertt *)khelp_get_osd(CCV(ccv, osd), ertt_id);
576 new_measurement = e_t->flags & ERTT_NEW_MEASUREMENT;
577 congestion = 0;
578 cdg_data->maxrtt_in_rtt = imax(e_t->rtt, cdg_data->maxrtt_in_rtt);
579 cdg_data->minrtt_in_rtt = imin(e_t->rtt, cdg_data->minrtt_in_rtt);
580
581 if (new_measurement) {
582 slowstart = (CCV(ccv, snd_cwnd) <= CCV(ccv, snd_ssthresh));
583 /*
584 * Update smoothed gradient measurements. Since we are only
585 * using one measurement per RTT, use max or min rtt_in_rtt.
586 * This is also less noisy than a sample RTT measurement. Max
587 * RTT measurements can have trouble due to OS issues.
588 */
589 if (cdg_data->maxrtt_in_prevrtt) {
590 qdiff_max = ((long)(cdg_data->maxrtt_in_rtt -
591 cdg_data->maxrtt_in_prevrtt) << D_P_E );
592 qdiff_min = ((long)(cdg_data->minrtt_in_rtt -
593 cdg_data->minrtt_in_prevrtt) << D_P_E );
594
595 if (cdg_data->sample_q_size == 0) {
596 cdg_data->max_qtrend = qdiff_max;
597 cdg_data->min_qtrend = qdiff_min;
598 } else
599 calc_moving_average(cdg_data, qdiff_max, qdiff_min);
600
601 /* Probabilistic backoff with respect to gradient. */
602 if (slowstart && qdiff_min > 0)
603 congestion = prob_backoff(qdiff_min);
604 else if (cdg_data->min_qtrend > 0)
605 congestion = prob_backoff(cdg_data->min_qtrend);
606 else if (slowstart && qdiff_max > 0)
607 congestion = prob_backoff(qdiff_max);
608 else if (cdg_data->max_qtrend > 0)
609 congestion = prob_backoff(cdg_data->max_qtrend);
610
611 /* Update estimate of queue state. */
612 if (cdg_data->min_qtrend > 0 &&
613 cdg_data->max_qtrend <= 0) {
614 cdg_data->queue_state = CDG_Q_FULL;
615 } else if (cdg_data->min_qtrend >= 0 &&
616 cdg_data->max_qtrend < 0) {
617 cdg_data->queue_state = CDG_Q_EMPTY;
618 cdg_data->shadow_w = 0;
619 } else if (cdg_data->min_qtrend > 0 &&
620 cdg_data->max_qtrend > 0) {
621 cdg_data->queue_state = CDG_Q_RISING;
622 } else if (cdg_data->min_qtrend < 0 &&
623 cdg_data->max_qtrend < 0) {
624 cdg_data->queue_state = CDG_Q_FALLING;
625 }
626
627 if (cdg_data->min_qtrend < 0 ||
628 cdg_data->max_qtrend < 0)
629 cdg_data->consec_cong_cnt = 0;
630 }
631
632 cdg_data->minrtt_in_prevrtt = cdg_data->minrtt_in_rtt;
633 cdg_data->minrtt_in_rtt = INT_MAX;
634 cdg_data->maxrtt_in_prevrtt = cdg_data->maxrtt_in_rtt;
635 cdg_data->maxrtt_in_rtt = 0;
636 e_t->flags &= ~ERTT_NEW_MEASUREMENT;
637 }
638
639 if (congestion) {
640 cdg_data->consec_cong_cnt++;
641 if (!IN_RECOVERY(CCV(ccv, t_flags))) {
642 if (cdg_data->consec_cong_cnt <= V_cdg_consec_cong)
643 cdg_cong_signal(ccv, CC_CDG_DELAY);
644 else
645 /*
646 * We have been backing off but the queue is not
647 * falling. Assume we are competing with
648 * loss-based flows and don't back off for the
649 * next V_cdg_hold_backoff RTT periods.
650 */
651 if (cdg_data->consec_cong_cnt >=
652 V_cdg_consec_cong + V_cdg_hold_backoff)
653 cdg_data->consec_cong_cnt = 0;
654
655 /* Won't see effect until 2nd RTT. */
656 cdg_data->maxrtt_in_prevrtt = 0;
657 /*
658 * Resync shadow window in case we are competing with a
659 * loss based flow
660 */
661 cdg_data->shadow_w = ulmax(CCV(ccv, snd_cwnd),
662 cdg_data->shadow_w);
663 }
664 } else if (ack_type == CC_ACK)
665 cdg_window_increase(ccv, new_measurement);
666 }
667
668 /* When a vnet is created and being initialised, init the per-stack CDG vars. */
669 VNET_SYSINIT(cdg_init_vnet, SI_SUB_PROTO_BEGIN, SI_ORDER_FIRST,
670 cdg_init_vnet, NULL);
671
672 SYSCTL_DECL(_net_inet_tcp_cc_cdg);
673 SYSCTL_NODE(_net_inet_tcp_cc, OID_AUTO, cdg, CTLFLAG_RW | CTLFLAG_MPSAFE, NULL,
674 "CAIA delay-gradient congestion control related settings");
675
676 SYSCTL_STRING(_net_inet_tcp_cc_cdg, OID_AUTO, version,
677 CTLFLAG_RD, CDG_VERSION, sizeof(CDG_VERSION) - 1,
678 "Current algorithm/implementation version number");
679
680 SYSCTL_UINT(_net_inet_tcp_cc_cdg, OID_AUTO, alpha_inc,
681 CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(cdg_alpha_inc), 0,
682 "Increment the window increase factor alpha by 1 MSS segment every "
683 "alpha_inc RTTs during congestion avoidance mode.");
684
685 SYSCTL_PROC(_net_inet_tcp_cc_cdg, OID_AUTO, beta_delay,
686 CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
687 &VNET_NAME(cdg_beta_delay), 70, &cdg_beta_handler, "IU",
688 "Delay-based window decrease factor as a percentage "
689 "(on delay-based backoff, w = w * beta_delay / 100)");
690
691 SYSCTL_PROC(_net_inet_tcp_cc_cdg, OID_AUTO, beta_loss,
692 CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
693 &VNET_NAME(cdg_beta_loss), 50, &cdg_beta_handler, "IU",
694 "Loss-based window decrease factor as a percentage "
695 "(on loss-based backoff, w = w * beta_loss / 100)");
696
697 SYSCTL_PROC(_net_inet_tcp_cc_cdg, OID_AUTO, exp_backoff_scale,
698 CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
699 &VNET_NAME(cdg_exp_backoff_scale), 2, &cdg_exp_backoff_scale_handler, "IU",
700 "Scaling parameter for the probabilistic exponential backoff");
701
702 SYSCTL_UINT(_net_inet_tcp_cc_cdg, OID_AUTO, smoothing_factor,
703 CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(cdg_smoothing_factor), 8,
704 "Number of samples used for moving average smoothing (0 = no smoothing)");
705
706 SYSCTL_UINT(_net_inet_tcp_cc_cdg, OID_AUTO, loss_compete_consec_cong,
707 CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(cdg_consec_cong), 5,
708 "Number of consecutive delay-gradient based congestion episodes which will "
709 "trigger loss based CC compatibility");
710
711 SYSCTL_UINT(_net_inet_tcp_cc_cdg, OID_AUTO, loss_compete_hold_backoff,
712 CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(cdg_hold_backoff), 5,
713 "Number of consecutive delay-gradient based congestion episodes to hold "
714 "the window backoff for loss based CC compatibility");
715
716 DECLARE_CC_MODULE(cdg, &cdg_cc_algo);
717 MODULE_VERSION(cdg, 1);
718 MODULE_DEPEND(cdg, ertt, 1, 1, 1);
719