1 /*-
2 * Copyright (c) 2009-2013
3 * Swinburne University of Technology, Melbourne, Australia
4 * All rights reserved.
5 *
6 * This software was developed at the Centre for Advanced Internet
7 * Architectures, Swinburne University of Technology, by David Hayes, made
8 * possible in part by a gift from The Cisco University Research Program Fund,
9 * a corporate advised fund of Silicon Valley Community Foundation. Development
10 * and testing were further assisted by a grant from the FreeBSD Foundation.
11 *
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
14 * are met:
15 * 1. Redistributions of source code must retain the above copyright
16 * notice, this list of conditions and the following disclaimer.
17 * 2. Redistributions in binary form must reproduce the above copyright
18 * notice, this list of conditions and the following disclaimer in the
19 * documentation and/or other materials provided with the distribution.
20 *
21 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
22 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
25 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31 * SUCH DAMAGE.
32 */
33
34 /*
35 * CAIA Delay-Gradient (CDG) congestion control algorithm
36 *
37 * An implemention of the delay-gradient congestion control algorithm proposed
38 * in the following paper:
39 *
40 * D. A. Hayes and G. Armitage, "Revisiting TCP Congestion Control using Delay
41 * Gradients", in IFIP Networking, Valencia, Spain, 9-13 May 2011.
42 *
43 * Developed as part of the NewTCP research project at Swinburne University of
44 * Technology's Centre for Advanced Internet Architectures, Melbourne,
45 * Australia. More details are available at:
46 * http://caia.swin.edu.au/urp/newtcp/
47 */
48
49 #include <sys/cdefs.h>
50 __FBSDID("$FreeBSD: stable/9/sys/netinet/cc/cc_cdg.c 271653 2014-09-16 00:04:38Z lstewart $");
51
52 #include <sys/param.h>
53 #include <sys/hhook.h>
54 #include <sys/kernel.h>
55 #include <sys/khelp.h>
56 #include <sys/limits.h>
57 #include <sys/lock.h>
58 #include <sys/malloc.h>
59 #include <sys/module.h>
60 #include <sys/queue.h>
61 #include <sys/socket.h>
62 #include <sys/socketvar.h>
63 #include <sys/sysctl.h>
64 #include <sys/systm.h>
65
66 #include <net/if.h>
67 #include <net/vnet.h>
68
69 #include <netinet/cc.h>
70 #include <netinet/tcp_seq.h>
71 #include <netinet/tcp_timer.h>
72 #include <netinet/tcp_var.h>
73
74 #include <netinet/cc/cc_module.h>
75
76 #include <netinet/khelp/h_ertt.h>
77
78 #include <vm/uma.h>
79
80 #define CDG_VERSION "0.1"
81
82 #define CAST_PTR_INT(X) (*((int*)(X)))
83
84 #ifndef VIMAGE
85 #define vnet_sysctl_handle_uint(oidp, arg1, arg2, req) \
86 sysctl_handle_int(oidp, arg1, arg2, req)
87 #endif
88
89 /* Private delay-gradient induced congestion control signal. */
90 #define CC_CDG_DELAY 0x01000000
91
92 /* NewReno window deflation factor on loss (as a percentage). */
93 #define RENO_BETA 50
94
95 /* Queue states. */
96 #define CDG_Q_EMPTY 1
97 #define CDG_Q_RISING 2
98 #define CDG_Q_FALLING 3
99 #define CDG_Q_FULL 4
100 #define CDG_Q_UNKNOWN 9999
101
102 /* Number of bit shifts used in probexp lookup table. */
103 #define EXP_PREC 15
104
105 /* Largest gradient represented in probexp lookup table. */
106 #define MAXGRAD 5
107
108 /*
109 * Delay Precision Enhance - number of bit shifts used for qtrend related
110 * integer arithmetic precision.
111 */
112 #define D_P_E 7
113
114 struct qdiff_sample {
115 long qdiff;
116 STAILQ_ENTRY(qdiff_sample) qdiff_lnk;
117 };
118
119 struct cdg {
120 long max_qtrend;
121 long min_qtrend;
122 STAILQ_HEAD(minrtts_head, qdiff_sample) qdiffmin_q;
123 STAILQ_HEAD(maxrtts_head, qdiff_sample) qdiffmax_q;
124 long window_incr;
125 /* rttcount for window increase when in congestion avoidance */
126 long rtt_count;
127 /* maximum measured rtt within an rtt period */
128 int maxrtt_in_rtt;
129 /* maximum measured rtt within prev rtt period */
130 int maxrtt_in_prevrtt;
131 /* minimum measured rtt within an rtt period */
132 int minrtt_in_rtt;
133 /* minimum measured rtt within prev rtt period */
134 int minrtt_in_prevrtt;
135 /* consecutive congestion episode counter */
136 uint32_t consec_cong_cnt;
137 /* when tracking a new reno type loss window */
138 uint32_t shadow_w;
139 /* maximum number of samples in the moving average queue */
140 int sample_q_size;
141 /* number of samples in the moving average queue */
142 int num_samples;
143 /* estimate of the queue state of the path */
144 int queue_state;
145 };
146
147 /*
148 * Lookup table for:
149 * (1 - exp(-x)) << EXP_PREC, where x = [0,MAXGRAD] in 2^-7 increments
150 *
151 * Note: probexp[0] is set to 10 (not 0) as a safety for very low increase
152 * gradients.
153 */
154 static const int probexp[641] = {
155 10,255,508,759,1008,1255,1501,1744,1985,2225,2463,2698,2932,3165,3395,3624,
156 3850,4075,4299,4520,4740,4958,5175,5389,5602,5814,6024,6232,6438,6643,6846,
157 7048,7248,7447,7644,7839,8033,8226,8417,8606,8794,8981,9166,9350,9532,9713,
158 9892,10070,10247,10422,10596,10769,10940,11110,11278,11445,11611,11776,11939,
159 12101,12262,12422,12580,12737,12893,13048,13201,13354,13505,13655,13803,13951,
160 14097,14243,14387,14530,14672,14813,14952,15091,15229,15365,15500,15635,15768,
161 15900,16032,16162,16291,16419,16547,16673,16798,16922,17046,17168,17289,17410,
162 17529,17648,17766,17882,17998,18113,18227,18340,18453,18564,18675,18784,18893,
163 19001,19108,19215,19320,19425,19529,19632,19734,19835,19936,20036,20135,20233,
164 20331,20427,20523,20619,20713,20807,20900,20993,21084,21175,21265,21355,21444,
165 21532,21619,21706,21792,21878,21962,22046,22130,22213,22295,22376,22457,22537,
166 22617,22696,22774,22852,22929,23006,23082,23157,23232,23306,23380,23453,23525,
167 23597,23669,23739,23810,23879,23949,24017,24085,24153,24220,24286,24352,24418,
168 24483,24547,24611,24675,24738,24800,24862,24924,24985,25045,25106,25165,25224,
169 25283,25341,25399,25456,25513,25570,25626,25681,25737,25791,25846,25899,25953,
170 26006,26059,26111,26163,26214,26265,26316,26366,26416,26465,26514,26563,26611,
171 26659,26707,26754,26801,26847,26893,26939,26984,27029,27074,27118,27162,27206,
172 27249,27292,27335,27377,27419,27460,27502,27543,27583,27624,27664,27703,27743,
173 27782,27821,27859,27897,27935,27973,28010,28047,28084,28121,28157,28193,28228,
174 28263,28299,28333,28368,28402,28436,28470,28503,28536,28569,28602,28634,28667,
175 28699,28730,28762,28793,28824,28854,28885,28915,28945,28975,29004,29034,29063,
176 29092,29120,29149,29177,29205,29232,29260,29287,29314,29341,29368,29394,29421,
177 29447,29472,29498,29524,29549,29574,29599,29623,29648,29672,29696,29720,29744,
178 29767,29791,29814,29837,29860,29882,29905,29927,29949,29971,29993,30014,30036,
179 30057,30078,30099,30120,30141,30161,30181,30201,30221,30241,30261,30280,30300,
180 30319,30338,30357,30376,30394,30413,30431,30449,30467,30485,30503,30521,30538,
181 30555,30573,30590,30607,30624,30640,30657,30673,30690,30706,30722,30738,30753,
182 30769,30785,30800,30815,30831,30846,30861,30876,30890,30905,30919,30934,30948,
183 30962,30976,30990,31004,31018,31031,31045,31058,31072,31085,31098,31111,31124,
184 31137,31149,31162,31174,31187,31199,31211,31223,31235,31247,31259,31271,31283,
185 31294,31306,31317,31328,31339,31351,31362,31373,31383,31394,31405,31416,31426,
186 31436,31447,31457,31467,31477,31487,31497,31507,31517,31527,31537,31546,31556,
187 31565,31574,31584,31593,31602,31611,31620,31629,31638,31647,31655,31664,31673,
188 31681,31690,31698,31706,31715,31723,31731,31739,31747,31755,31763,31771,31778,
189 31786,31794,31801,31809,31816,31824,31831,31838,31846,31853,31860,31867,31874,
190 31881,31888,31895,31902,31908,31915,31922,31928,31935,31941,31948,31954,31960,
191 31967,31973,31979,31985,31991,31997,32003,32009,32015,32021,32027,32033,32038,
192 32044,32050,32055,32061,32066,32072,32077,32083,32088,32093,32098,32104,32109,
193 32114,32119,32124,32129,32134,32139,32144,32149,32154,32158,32163,32168,32173,
194 32177,32182,32186,32191,32195,32200,32204,32209,32213,32217,32222,32226,32230,
195 32234,32238,32242,32247,32251,32255,32259,32263,32267,32270,32274,32278,32282,
196 32286,32290,32293,32297,32301,32304,32308,32311,32315,32318,32322,32325,32329,
197 32332,32336,32339,32342,32346,32349,32352,32356,32359,32362,32365,32368,32371,
198 32374,32377,32381,32384,32387,32389,32392,32395,32398,32401,32404,32407,32410,
199 32412,32415,32418,32421,32423,32426,32429,32431,32434,32437,32439,32442,32444,
200 32447,32449,32452,32454,32457,32459,32461,32464,32466,32469,32471,32473,32476,
201 32478,32480,32482,32485,32487,32489,32491,32493,32495,32497,32500,32502,32504,
202 32506,32508,32510,32512,32514,32516,32518,32520,32522,32524,32526,32527,32529,
203 32531,32533,32535,32537,32538,32540,32542,32544,32545,32547};
204
205 static uma_zone_t qdiffsample_zone;
206
207 static MALLOC_DEFINE(M_CDG, "cdg data",
208 "Per connection data required for the CDG congestion control algorithm");
209
210 static int ertt_id;
211
212 static VNET_DEFINE(uint32_t, cdg_alpha_inc);
213 static VNET_DEFINE(uint32_t, cdg_beta_delay);
214 static VNET_DEFINE(uint32_t, cdg_beta_loss);
215 static VNET_DEFINE(uint32_t, cdg_smoothing_factor);
216 static VNET_DEFINE(uint32_t, cdg_exp_backoff_scale);
217 static VNET_DEFINE(uint32_t, cdg_consec_cong);
218 static VNET_DEFINE(uint32_t, cdg_hold_backoff);
219 #define V_cdg_alpha_inc VNET(cdg_alpha_inc)
220 #define V_cdg_beta_delay VNET(cdg_beta_delay)
221 #define V_cdg_beta_loss VNET(cdg_beta_loss)
222 #define V_cdg_smoothing_factor VNET(cdg_smoothing_factor)
223 #define V_cdg_exp_backoff_scale VNET(cdg_exp_backoff_scale)
224 #define V_cdg_consec_cong VNET(cdg_consec_cong)
225 #define V_cdg_hold_backoff VNET(cdg_hold_backoff)
226
227 /* Function prototypes. */
228 static int cdg_mod_init(void);
229 static int cdg_mod_destroy(void);
230 static void cdg_conn_init(struct cc_var *ccv);
231 static int cdg_cb_init(struct cc_var *ccv);
232 static void cdg_cb_destroy(struct cc_var *ccv);
233 static void cdg_cong_signal(struct cc_var *ccv, uint32_t signal_type);
234 static void cdg_ack_received(struct cc_var *ccv, uint16_t ack_type);
235
236 struct cc_algo cdg_cc_algo = {
237 .name = "cdg",
238 .mod_init = cdg_mod_init,
239 .ack_received = cdg_ack_received,
240 .cb_destroy = cdg_cb_destroy,
241 .cb_init = cdg_cb_init,
242 .conn_init = cdg_conn_init,
243 .cong_signal = cdg_cong_signal,
244 .mod_destroy = cdg_mod_destroy
245 };
246
247 /* Vnet created and being initialised. */
248 static void
cdg_init_vnet(const void * unused __unused)249 cdg_init_vnet(const void *unused __unused)
250 {
251
252 V_cdg_alpha_inc = 0;
253 V_cdg_beta_delay = 70;
254 V_cdg_beta_loss = 50;
255 V_cdg_smoothing_factor = 8;
256 V_cdg_exp_backoff_scale = 3;
257 V_cdg_consec_cong = 5;
258 V_cdg_hold_backoff = 5;
259 }
260
261 static int
cdg_mod_init(void)262 cdg_mod_init(void)
263 {
264 VNET_ITERATOR_DECL(v);
265
266 ertt_id = khelp_get_id("ertt");
267 if (ertt_id <= 0)
268 return (EINVAL);
269
270 qdiffsample_zone = uma_zcreate("cdg_qdiffsample",
271 sizeof(struct qdiff_sample), NULL, NULL, NULL, NULL, 0, 0);
272
273 VNET_LIST_RLOCK();
274 VNET_FOREACH(v) {
275 CURVNET_SET(v);
276 cdg_init_vnet(NULL);
277 CURVNET_RESTORE();
278 }
279 VNET_LIST_RUNLOCK();
280
281 cdg_cc_algo.post_recovery = newreno_cc_algo.post_recovery;
282 cdg_cc_algo.after_idle = newreno_cc_algo.after_idle;
283
284 return (0);
285 }
286
287 static int
cdg_mod_destroy(void)288 cdg_mod_destroy(void)
289 {
290
291 uma_zdestroy(qdiffsample_zone);
292 return (0);
293 }
294
295 static int
cdg_cb_init(struct cc_var * ccv)296 cdg_cb_init(struct cc_var *ccv)
297 {
298 struct cdg *cdg_data;
299
300 cdg_data = malloc(sizeof(struct cdg), M_CDG, M_NOWAIT);
301 if (cdg_data == NULL)
302 return (ENOMEM);
303
304 cdg_data->shadow_w = 0;
305 cdg_data->max_qtrend = 0;
306 cdg_data->min_qtrend = 0;
307 cdg_data->queue_state = CDG_Q_UNKNOWN;
308 cdg_data->maxrtt_in_rtt = 0;
309 cdg_data->maxrtt_in_prevrtt = 0;
310 cdg_data->minrtt_in_rtt = INT_MAX;
311 cdg_data->minrtt_in_prevrtt = 0;
312 cdg_data->window_incr = 0;
313 cdg_data->rtt_count = 0;
314 cdg_data->consec_cong_cnt = 0;
315 cdg_data->sample_q_size = V_cdg_smoothing_factor;
316 cdg_data->num_samples = 0;
317 STAILQ_INIT(&cdg_data->qdiffmin_q);
318 STAILQ_INIT(&cdg_data->qdiffmax_q);
319
320 ccv->cc_data = cdg_data;
321
322 return (0);
323 }
324
325 static void
cdg_conn_init(struct cc_var * ccv)326 cdg_conn_init(struct cc_var *ccv)
327 {
328 struct cdg *cdg_data = ccv->cc_data;
329
330 /*
331 * Initialise the shadow_cwnd in case we are competing with loss based
332 * flows from the start
333 */
334 cdg_data->shadow_w = CCV(ccv, snd_cwnd);
335 }
336
337 static void
cdg_cb_destroy(struct cc_var * ccv)338 cdg_cb_destroy(struct cc_var *ccv)
339 {
340 struct cdg *cdg_data;
341 struct qdiff_sample *qds, *qds_n;
342
343 cdg_data = ccv->cc_data;
344
345 qds = STAILQ_FIRST(&cdg_data->qdiffmin_q);
346 while (qds != NULL) {
347 qds_n = STAILQ_NEXT(qds, qdiff_lnk);
348 uma_zfree(qdiffsample_zone,qds);
349 qds = qds_n;
350 }
351
352 qds = STAILQ_FIRST(&cdg_data->qdiffmax_q);
353 while (qds != NULL) {
354 qds_n = STAILQ_NEXT(qds, qdiff_lnk);
355 uma_zfree(qdiffsample_zone,qds);
356 qds = qds_n;
357 }
358
359 free(ccv->cc_data, M_CDG);
360 }
361
362 static int
cdg_beta_handler(SYSCTL_HANDLER_ARGS)363 cdg_beta_handler(SYSCTL_HANDLER_ARGS)
364 {
365
366 if (req->newptr != NULL &&
367 (CAST_PTR_INT(req->newptr) == 0 || CAST_PTR_INT(req->newptr) > 100))
368 return (EINVAL);
369
370 return (vnet_sysctl_handle_uint(oidp, arg1, arg2, req));
371 }
372
373 static int
cdg_exp_backoff_scale_handler(SYSCTL_HANDLER_ARGS)374 cdg_exp_backoff_scale_handler(SYSCTL_HANDLER_ARGS)
375 {
376
377 if (req->newptr != NULL && CAST_PTR_INT(req->newptr) < 1)
378 return (EINVAL);
379
380 return (vnet_sysctl_handle_uint(oidp, arg1, arg2, req));
381 }
382
383 static inline unsigned long
cdg_window_decrease(struct cc_var * ccv,unsigned long owin,unsigned int beta)384 cdg_window_decrease(struct cc_var *ccv, unsigned long owin, unsigned int beta)
385 {
386
387 return ((ulmin(CCV(ccv, snd_wnd), owin) * beta) / 100);
388 }
389
390 /*
391 * Window increase function
392 * This window increase function is independent of the initial window size
393 * to ensure small window flows are not discriminated against (i.e. fairness).
394 * It increases at 1pkt/rtt like Reno for alpha_inc rtts, and then 2pkts/rtt for
395 * the next alpha_inc rtts, etc.
396 */
397 static void
cdg_window_increase(struct cc_var * ccv,int new_measurement)398 cdg_window_increase(struct cc_var *ccv, int new_measurement)
399 {
400 struct cdg *cdg_data;
401 int incr, s_w_incr;
402
403 cdg_data = ccv->cc_data;
404 incr = s_w_incr = 0;
405
406 if (CCV(ccv, snd_cwnd) <= CCV(ccv, snd_ssthresh)) {
407 /* Slow start. */
408 incr = CCV(ccv, t_maxseg);
409 s_w_incr = incr;
410 cdg_data->window_incr = cdg_data->rtt_count = 0;
411 } else {
412 /* Congestion avoidance. */
413 if (new_measurement) {
414 s_w_incr = CCV(ccv, t_maxseg);
415 if (V_cdg_alpha_inc == 0) {
416 incr = CCV(ccv, t_maxseg);
417 } else {
418 if (++cdg_data->rtt_count >= V_cdg_alpha_inc) {
419 cdg_data->window_incr++;
420 cdg_data->rtt_count = 0;
421 }
422 incr = CCV(ccv, t_maxseg) *
423 cdg_data->window_incr;
424 }
425 }
426 }
427
428 if (cdg_data->shadow_w > 0)
429 cdg_data->shadow_w = ulmin(cdg_data->shadow_w + s_w_incr,
430 TCP_MAXWIN << CCV(ccv, snd_scale));
431
432 CCV(ccv, snd_cwnd) = ulmin(CCV(ccv, snd_cwnd) + incr,
433 TCP_MAXWIN << CCV(ccv, snd_scale));
434 }
435
436 static void
cdg_cong_signal(struct cc_var * ccv,uint32_t signal_type)437 cdg_cong_signal(struct cc_var *ccv, uint32_t signal_type)
438 {
439 struct cdg *cdg_data = ccv->cc_data;
440
441 switch(signal_type) {
442 case CC_CDG_DELAY:
443 CCV(ccv, snd_ssthresh) = cdg_window_decrease(ccv,
444 CCV(ccv, snd_cwnd), V_cdg_beta_delay);
445 CCV(ccv, snd_cwnd) = CCV(ccv, snd_ssthresh);
446 CCV(ccv, snd_recover) = CCV(ccv, snd_max);
447 cdg_data->window_incr = cdg_data->rtt_count = 0;
448 ENTER_CONGRECOVERY(CCV(ccv, t_flags));
449 break;
450 case CC_NDUPACK:
451 /*
452 * If already responding to congestion OR we have guessed no
453 * queue in the path is full.
454 */
455 if (IN_CONGRECOVERY(CCV(ccv, t_flags)) ||
456 cdg_data->queue_state < CDG_Q_FULL) {
457 CCV(ccv, snd_ssthresh) = CCV(ccv, snd_cwnd);
458 CCV(ccv, snd_recover) = CCV(ccv, snd_max);
459 } else {
460 /*
461 * Loss is likely to be congestion related. We have
462 * inferred a queue full state, so have shadow window
463 * react to loss as NewReno would.
464 */
465 if (cdg_data->shadow_w > 0)
466 cdg_data->shadow_w = cdg_window_decrease(ccv,
467 cdg_data->shadow_w, RENO_BETA);
468
469 CCV(ccv, snd_ssthresh) = ulmax(cdg_data->shadow_w,
470 cdg_window_decrease(ccv, CCV(ccv, snd_cwnd),
471 V_cdg_beta_loss));
472
473 cdg_data->window_incr = cdg_data->rtt_count = 0;
474 }
475 ENTER_RECOVERY(CCV(ccv, t_flags));
476 break;
477 default:
478 newreno_cc_algo.cong_signal(ccv, signal_type);
479 break;
480 }
481 }
482
483 /*
484 * Using a negative exponential probabilistic backoff so that sources with
485 * varying RTTs which share the same link will, on average, have the same
486 * probability of backoff over time.
487 *
488 * Prob_backoff = 1 - exp(-qtrend / V_cdg_exp_backoff_scale), where
489 * V_cdg_exp_backoff_scale is the average qtrend for the exponential backoff.
490 */
491 static inline int
prob_backoff(long qtrend)492 prob_backoff(long qtrend)
493 {
494 int backoff, idx, p;
495
496 backoff = (qtrend > ((MAXGRAD * V_cdg_exp_backoff_scale) << D_P_E));
497
498 if (!backoff) {
499 if (V_cdg_exp_backoff_scale > 1)
500 idx = (qtrend + V_cdg_exp_backoff_scale / 2) /
501 V_cdg_exp_backoff_scale;
502 else
503 idx = qtrend;
504
505 /* Backoff probability proportional to rate of queue growth. */
506 p = (INT_MAX / (1 << EXP_PREC)) * probexp[idx];
507 backoff = (random() < p);
508 }
509
510 return (backoff);
511 }
512
513 static inline void
calc_moving_average(struct cdg * cdg_data,long qdiff_max,long qdiff_min)514 calc_moving_average(struct cdg *cdg_data, long qdiff_max, long qdiff_min)
515 {
516 struct qdiff_sample *qds;
517
518 ++cdg_data->num_samples;
519 if (cdg_data->num_samples > cdg_data->sample_q_size) {
520 /* Minimum RTT. */
521 qds = STAILQ_FIRST(&cdg_data->qdiffmin_q);
522 cdg_data->min_qtrend = cdg_data->min_qtrend +
523 (qdiff_min - qds->qdiff) / cdg_data->sample_q_size;
524 STAILQ_REMOVE_HEAD(&cdg_data->qdiffmin_q, qdiff_lnk);
525 qds->qdiff = qdiff_min;
526 STAILQ_INSERT_TAIL(&cdg_data->qdiffmin_q, qds, qdiff_lnk);
527
528 /* Maximum RTT. */
529 qds = STAILQ_FIRST(&cdg_data->qdiffmax_q);
530 cdg_data->max_qtrend = cdg_data->max_qtrend +
531 (qdiff_max - qds->qdiff) / cdg_data->sample_q_size;
532 STAILQ_REMOVE_HEAD(&cdg_data->qdiffmax_q, qdiff_lnk);
533 qds->qdiff = qdiff_max;
534 STAILQ_INSERT_TAIL(&cdg_data->qdiffmax_q, qds, qdiff_lnk);
535 --cdg_data->num_samples;
536 } else {
537 qds = uma_zalloc(qdiffsample_zone, M_NOWAIT);
538 if (qds != NULL) {
539 cdg_data->min_qtrend = cdg_data->min_qtrend +
540 qdiff_min / cdg_data->sample_q_size;
541 qds->qdiff = qdiff_min;
542 STAILQ_INSERT_TAIL(&cdg_data->qdiffmin_q, qds,
543 qdiff_lnk);
544 }
545
546 qds = uma_zalloc(qdiffsample_zone, M_NOWAIT);
547 if (qds) {
548 cdg_data->max_qtrend = cdg_data->max_qtrend +
549 qdiff_max / cdg_data->sample_q_size;
550 qds->qdiff = qdiff_max;
551 STAILQ_INSERT_TAIL(&cdg_data->qdiffmax_q, qds,
552 qdiff_lnk);
553 }
554 }
555 }
556
557 static void
cdg_ack_received(struct cc_var * ccv,uint16_t ack_type)558 cdg_ack_received(struct cc_var *ccv, uint16_t ack_type)
559 {
560 struct cdg *cdg_data;
561 struct ertt *e_t;
562 long qdiff_max, qdiff_min;
563 int congestion, new_measurement, slowstart;
564
565 cdg_data = ccv->cc_data;
566 e_t = (struct ertt *)khelp_get_osd(CCV(ccv, osd), ertt_id);
567 new_measurement = e_t->flags & ERTT_NEW_MEASUREMENT;
568 congestion = 0;
569 cdg_data->maxrtt_in_rtt = imax(e_t->rtt, cdg_data->maxrtt_in_rtt);
570 cdg_data->minrtt_in_rtt = imin(e_t->rtt, cdg_data->minrtt_in_rtt);
571
572 if (new_measurement) {
573 slowstart = (CCV(ccv, snd_cwnd) <= CCV(ccv, snd_ssthresh));
574 /*
575 * Update smoothed gradient measurements. Since we are only
576 * using one measurement per RTT, use max or min rtt_in_rtt.
577 * This is also less noisy than a sample RTT measurement. Max
578 * RTT measurements can have trouble due to OS issues.
579 */
580 if (cdg_data->maxrtt_in_prevrtt) {
581 qdiff_max = ((long)(cdg_data->maxrtt_in_rtt -
582 cdg_data->maxrtt_in_prevrtt) << D_P_E );
583 qdiff_min = ((long)(cdg_data->minrtt_in_rtt -
584 cdg_data->minrtt_in_prevrtt) << D_P_E );
585
586 calc_moving_average(cdg_data, qdiff_max, qdiff_min);
587
588 /* Probabilistic backoff with respect to gradient. */
589 if (slowstart && qdiff_min > 0)
590 congestion = prob_backoff(qdiff_min);
591 else if (cdg_data->min_qtrend > 0)
592 congestion = prob_backoff(cdg_data->min_qtrend);
593 else if (slowstart && qdiff_max > 0)
594 congestion = prob_backoff(qdiff_max);
595 else if (cdg_data->max_qtrend > 0)
596 congestion = prob_backoff(cdg_data->max_qtrend);
597
598 /* Update estimate of queue state. */
599 if (cdg_data->min_qtrend > 0 &&
600 cdg_data->max_qtrend <= 0) {
601 cdg_data->queue_state = CDG_Q_FULL;
602 } else if (cdg_data->min_qtrend >= 0 &&
603 cdg_data->max_qtrend < 0) {
604 cdg_data->queue_state = CDG_Q_EMPTY;
605 cdg_data->shadow_w = 0;
606 } else if (cdg_data->min_qtrend > 0 &&
607 cdg_data->max_qtrend > 0) {
608 cdg_data->queue_state = CDG_Q_RISING;
609 } else if (cdg_data->min_qtrend < 0 &&
610 cdg_data->max_qtrend < 0) {
611 cdg_data->queue_state = CDG_Q_FALLING;
612 }
613
614 if (cdg_data->min_qtrend < 0 ||
615 cdg_data->max_qtrend < 0)
616 cdg_data->consec_cong_cnt = 0;
617 }
618
619 cdg_data->minrtt_in_prevrtt = cdg_data->minrtt_in_rtt;
620 cdg_data->minrtt_in_rtt = INT_MAX;
621 cdg_data->maxrtt_in_prevrtt = cdg_data->maxrtt_in_rtt;
622 cdg_data->maxrtt_in_rtt = 0;
623 e_t->flags &= ~ERTT_NEW_MEASUREMENT;
624 }
625
626 if (congestion) {
627 cdg_data->consec_cong_cnt++;
628 if (!IN_RECOVERY(CCV(ccv, t_flags))) {
629 if (cdg_data->consec_cong_cnt <= V_cdg_consec_cong)
630 cdg_cong_signal(ccv, CC_CDG_DELAY);
631 else
632 /*
633 * We have been backing off but the queue is not
634 * falling. Assume we are competing with
635 * loss-based flows and don't back off for the
636 * next V_cdg_hold_backoff RTT periods.
637 */
638 if (cdg_data->consec_cong_cnt >=
639 V_cdg_consec_cong + V_cdg_hold_backoff)
640 cdg_data->consec_cong_cnt = 0;
641
642 /* Won't see effect until 2nd RTT. */
643 cdg_data->maxrtt_in_prevrtt = 0;
644 /*
645 * Resync shadow window in case we are competing with a
646 * loss based flow
647 */
648 cdg_data->shadow_w = ulmax(CCV(ccv, snd_cwnd),
649 cdg_data->shadow_w);
650 }
651 } else if (ack_type == CC_ACK)
652 cdg_window_increase(ccv, new_measurement);
653 }
654
655 /* When a vnet is created and being initialised, init the per-stack CDG vars. */
656 VNET_SYSINIT(cdg_init_vnet, SI_SUB_PROTO_BEGIN, SI_ORDER_FIRST,
657 cdg_init_vnet, NULL);
658
659 SYSCTL_DECL(_net_inet_tcp_cc_cdg);
660 SYSCTL_NODE(_net_inet_tcp_cc, OID_AUTO, cdg, CTLFLAG_RW, NULL,
661 "CAIA delay-gradient congestion control related settings");
662
663 SYSCTL_STRING(_net_inet_tcp_cc_cdg, OID_AUTO, version,
664 CTLFLAG_RD, CDG_VERSION, sizeof(CDG_VERSION) - 1,
665 "Current algorithm/implementation version number");
666
667 SYSCTL_VNET_UINT(_net_inet_tcp_cc_cdg, OID_AUTO, alpha_inc,
668 CTLFLAG_RW, &VNET_NAME(cdg_alpha_inc), 0,
669 "Increment the window increase factor alpha by 1 MSS segment every "
670 "alpha_inc RTTs during congestion avoidance mode.");
671
672 SYSCTL_VNET_PROC(_net_inet_tcp_cc_cdg, OID_AUTO, beta_delay,
673 CTLTYPE_UINT|CTLFLAG_RW, &VNET_NAME(cdg_beta_delay), 70,
674 &cdg_beta_handler, "IU",
675 "Delay-based window decrease factor as a percentage "
676 "(on delay-based backoff, w = w * beta_delay / 100)");
677
678 SYSCTL_VNET_PROC(_net_inet_tcp_cc_cdg, OID_AUTO, beta_loss,
679 CTLTYPE_UINT|CTLFLAG_RW, &VNET_NAME(cdg_beta_loss), 50,
680 &cdg_beta_handler, "IU",
681 "Loss-based window decrease factor as a percentage "
682 "(on loss-based backoff, w = w * beta_loss / 100)");
683
684 SYSCTL_VNET_PROC(_net_inet_tcp_cc_cdg, OID_AUTO, exp_backoff_scale,
685 CTLTYPE_UINT|CTLFLAG_RW, &VNET_NAME(cdg_exp_backoff_scale), 2,
686 &cdg_exp_backoff_scale_handler, "IU",
687 "Scaling parameter for the probabilistic exponential backoff");
688
689 SYSCTL_VNET_UINT(_net_inet_tcp_cc_cdg, OID_AUTO, smoothing_factor,
690 CTLFLAG_RW, &VNET_NAME(cdg_smoothing_factor), 8,
691 "Number of samples used for moving average smoothing (0 = no smoothing)");
692
693 SYSCTL_VNET_UINT(_net_inet_tcp_cc_cdg, OID_AUTO, loss_compete_consec_cong,
694 CTLFLAG_RW, &VNET_NAME(cdg_consec_cong), 5,
695 "Number of consecutive delay-gradient based congestion episodes which will "
696 "trigger loss based CC compatibility");
697
698 SYSCTL_VNET_UINT(_net_inet_tcp_cc_cdg, OID_AUTO, loss_compete_hold_backoff,
699 CTLFLAG_RW, &VNET_NAME(cdg_hold_backoff), 5,
700 "Number of consecutive delay-gradient based congestion episodes to hold "
701 "the window backoff for loss based CC compatibility");
702
703 DECLARE_CC_MODULE(cdg, &cdg_cc_algo);
704
705 MODULE_DEPEND(cdg, ertt, 1, 1, 1);
706