xref: /freebsd/sys/netinet/cc/cc_cdg.c (revision f39bffc62c1395bde25d152c7f68fdf7cbaab414)
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 #define	CAST_PTR_INT(X) (*((int*)(X)))
84 
85 /* Private delay-gradient induced congestion control signal. */
86 #define	CC_CDG_DELAY 0x01000000
87 
88 /* NewReno window deflation factor on loss (as a percentage). */
89 #define	RENO_BETA 50
90 
91 /* Queue states. */
92 #define	CDG_Q_EMPTY	1
93 #define	CDG_Q_RISING	2
94 #define	CDG_Q_FALLING	3
95 #define	CDG_Q_FULL	4
96 #define	CDG_Q_UNKNOWN	9999
97 
98 /* Number of bit shifts used in probexp lookup table. */
99 #define	EXP_PREC 15
100 
101 /* Largest gradient represented in probexp lookup table. */
102 #define	MAXGRAD 5
103 
104 /*
105  * Delay Precision Enhance - number of bit shifts used for qtrend related
106  * integer arithmetic precision.
107  */
108 #define	D_P_E 7
109 
110 struct qdiff_sample {
111 	long qdiff;
112 	STAILQ_ENTRY(qdiff_sample) qdiff_lnk;
113 };
114 
115 struct cdg {
116 	long max_qtrend;
117 	long min_qtrend;
118 	STAILQ_HEAD(minrtts_head, qdiff_sample) qdiffmin_q;
119 	STAILQ_HEAD(maxrtts_head, qdiff_sample) qdiffmax_q;
120 	long window_incr;
121 	/* rttcount for window increase when in congestion avoidance */
122 	long rtt_count;
123 	/* maximum measured rtt within an rtt period */
124 	int maxrtt_in_rtt;
125 	/* maximum measured rtt within prev rtt period */
126 	int maxrtt_in_prevrtt;
127 	/* minimum measured rtt within an rtt period */
128 	int minrtt_in_rtt;
129 	/* minimum measured rtt within prev rtt period */
130 	int minrtt_in_prevrtt;
131 	/* consecutive congestion episode counter */
132 	uint32_t consec_cong_cnt;
133 	/* when tracking a new reno type loss window */
134 	uint32_t shadow_w;
135 	/* maximum number of samples in the moving average queue */
136 	int sample_q_size;
137 	/* number of samples in the moving average queue */
138 	int num_samples;
139 	/* estimate of the queue state of the path */
140 	int queue_state;
141 };
142 
143 /*
144  * Lookup table for:
145  *   (1 - exp(-x)) << EXP_PREC, where x = [0,MAXGRAD] in 2^-7 increments
146  *
147  * Note: probexp[0] is set to 10 (not 0) as a safety for very low increase
148  * gradients.
149  */
150 static const int probexp[641] = {
151    10,255,508,759,1008,1255,1501,1744,1985,2225,2463,2698,2932,3165,3395,3624,
152    3850,4075,4299,4520,4740,4958,5175,5389,5602,5814,6024,6232,6438,6643,6846,
153    7048,7248,7447,7644,7839,8033,8226,8417,8606,8794,8981,9166,9350,9532,9713,
154    9892,10070,10247,10422,10596,10769,10940,11110,11278,11445,11611,11776,11939,
155    12101,12262,12422,12580,12737,12893,13048,13201,13354,13505,13655,13803,13951,
156    14097,14243,14387,14530,14672,14813,14952,15091,15229,15365,15500,15635,15768,
157    15900,16032,16162,16291,16419,16547,16673,16798,16922,17046,17168,17289,17410,
158    17529,17648,17766,17882,17998,18113,18227,18340,18453,18564,18675,18784,18893,
159    19001,19108,19215,19320,19425,19529,19632,19734,19835,19936,20036,20135,20233,
160    20331,20427,20523,20619,20713,20807,20900,20993,21084,21175,21265,21355,21444,
161    21532,21619,21706,21792,21878,21962,22046,22130,22213,22295,22376,22457,22537,
162    22617,22696,22774,22852,22929,23006,23082,23157,23232,23306,23380,23453,23525,
163    23597,23669,23739,23810,23879,23949,24017,24085,24153,24220,24286,24352,24418,
164    24483,24547,24611,24675,24738,24800,24862,24924,24985,25045,25106,25165,25224,
165    25283,25341,25399,25456,25513,25570,25626,25681,25737,25791,25846,25899,25953,
166    26006,26059,26111,26163,26214,26265,26316,26366,26416,26465,26514,26563,26611,
167    26659,26707,26754,26801,26847,26893,26939,26984,27029,27074,27118,27162,27206,
168    27249,27292,27335,27377,27419,27460,27502,27543,27583,27624,27664,27703,27743,
169    27782,27821,27859,27897,27935,27973,28010,28047,28084,28121,28157,28193,28228,
170    28263,28299,28333,28368,28402,28436,28470,28503,28536,28569,28602,28634,28667,
171    28699,28730,28762,28793,28824,28854,28885,28915,28945,28975,29004,29034,29063,
172    29092,29120,29149,29177,29205,29232,29260,29287,29314,29341,29368,29394,29421,
173    29447,29472,29498,29524,29549,29574,29599,29623,29648,29672,29696,29720,29744,
174    29767,29791,29814,29837,29860,29882,29905,29927,29949,29971,29993,30014,30036,
175    30057,30078,30099,30120,30141,30161,30181,30201,30221,30241,30261,30280,30300,
176    30319,30338,30357,30376,30394,30413,30431,30449,30467,30485,30503,30521,30538,
177    30555,30573,30590,30607,30624,30640,30657,30673,30690,30706,30722,30738,30753,
178    30769,30785,30800,30815,30831,30846,30861,30876,30890,30905,30919,30934,30948,
179    30962,30976,30990,31004,31018,31031,31045,31058,31072,31085,31098,31111,31124,
180    31137,31149,31162,31174,31187,31199,31211,31223,31235,31247,31259,31271,31283,
181    31294,31306,31317,31328,31339,31351,31362,31373,31383,31394,31405,31416,31426,
182    31436,31447,31457,31467,31477,31487,31497,31507,31517,31527,31537,31546,31556,
183    31565,31574,31584,31593,31602,31611,31620,31629,31638,31647,31655,31664,31673,
184    31681,31690,31698,31706,31715,31723,31731,31739,31747,31755,31763,31771,31778,
185    31786,31794,31801,31809,31816,31824,31831,31838,31846,31853,31860,31867,31874,
186    31881,31888,31895,31902,31908,31915,31922,31928,31935,31941,31948,31954,31960,
187    31967,31973,31979,31985,31991,31997,32003,32009,32015,32021,32027,32033,32038,
188    32044,32050,32055,32061,32066,32072,32077,32083,32088,32093,32098,32104,32109,
189    32114,32119,32124,32129,32134,32139,32144,32149,32154,32158,32163,32168,32173,
190    32177,32182,32186,32191,32195,32200,32204,32209,32213,32217,32222,32226,32230,
191    32234,32238,32242,32247,32251,32255,32259,32263,32267,32270,32274,32278,32282,
192    32286,32290,32293,32297,32301,32304,32308,32311,32315,32318,32322,32325,32329,
193    32332,32336,32339,32342,32346,32349,32352,32356,32359,32362,32365,32368,32371,
194    32374,32377,32381,32384,32387,32389,32392,32395,32398,32401,32404,32407,32410,
195    32412,32415,32418,32421,32423,32426,32429,32431,32434,32437,32439,32442,32444,
196    32447,32449,32452,32454,32457,32459,32461,32464,32466,32469,32471,32473,32476,
197    32478,32480,32482,32485,32487,32489,32491,32493,32495,32497,32500,32502,32504,
198    32506,32508,32510,32512,32514,32516,32518,32520,32522,32524,32526,32527,32529,
199    32531,32533,32535,32537,32538,32540,32542,32544,32545,32547};
200 
201 static uma_zone_t qdiffsample_zone;
202 
203 static MALLOC_DEFINE(M_CDG, "cdg data",
204   "Per connection data required for the CDG congestion control algorithm");
205 
206 static int ertt_id;
207 
208 static VNET_DEFINE(uint32_t, cdg_alpha_inc);
209 static VNET_DEFINE(uint32_t, cdg_beta_delay);
210 static VNET_DEFINE(uint32_t, cdg_beta_loss);
211 static VNET_DEFINE(uint32_t, cdg_smoothing_factor);
212 static VNET_DEFINE(uint32_t, cdg_exp_backoff_scale);
213 static VNET_DEFINE(uint32_t, cdg_consec_cong);
214 static VNET_DEFINE(uint32_t, cdg_hold_backoff);
215 #define	V_cdg_alpha_inc		VNET(cdg_alpha_inc)
216 #define	V_cdg_beta_delay	VNET(cdg_beta_delay)
217 #define	V_cdg_beta_loss		VNET(cdg_beta_loss)
218 #define	V_cdg_smoothing_factor	VNET(cdg_smoothing_factor)
219 #define	V_cdg_exp_backoff_scale	VNET(cdg_exp_backoff_scale)
220 #define	V_cdg_consec_cong	VNET(cdg_consec_cong)
221 #define	V_cdg_hold_backoff	VNET(cdg_hold_backoff)
222 
223 /* Function prototypes. */
224 static int cdg_mod_init(void);
225 static int cdg_mod_destroy(void);
226 static void cdg_conn_init(struct cc_var *ccv);
227 static int cdg_cb_init(struct cc_var *ccv);
228 static void cdg_cb_destroy(struct cc_var *ccv);
229 static void cdg_cong_signal(struct cc_var *ccv, uint32_t signal_type);
230 static void cdg_ack_received(struct cc_var *ccv, uint16_t ack_type);
231 
232 struct cc_algo cdg_cc_algo = {
233 	.name = "cdg",
234 	.mod_init = cdg_mod_init,
235 	.ack_received = cdg_ack_received,
236 	.cb_destroy = cdg_cb_destroy,
237 	.cb_init = cdg_cb_init,
238 	.conn_init = cdg_conn_init,
239 	.cong_signal = cdg_cong_signal,
240 	.mod_destroy = cdg_mod_destroy
241 };
242 
243 /* Vnet created and being initialised. */
244 static void
245 cdg_init_vnet(const void *unused __unused)
246 {
247 
248 	V_cdg_alpha_inc = 0;
249 	V_cdg_beta_delay = 70;
250 	V_cdg_beta_loss = 50;
251 	V_cdg_smoothing_factor = 8;
252 	V_cdg_exp_backoff_scale = 3;
253 	V_cdg_consec_cong = 5;
254 	V_cdg_hold_backoff = 5;
255 }
256 
257 static int
258 cdg_mod_init(void)
259 {
260 	VNET_ITERATOR_DECL(v);
261 
262 	ertt_id = khelp_get_id("ertt");
263 	if (ertt_id <= 0)
264 		return (EINVAL);
265 
266 	qdiffsample_zone = uma_zcreate("cdg_qdiffsample",
267 	    sizeof(struct qdiff_sample), NULL, NULL, NULL, NULL, 0, 0);
268 
269 	VNET_LIST_RLOCK();
270 	VNET_FOREACH(v) {
271 		CURVNET_SET(v);
272 		cdg_init_vnet(NULL);
273 		CURVNET_RESTORE();
274 	}
275 	VNET_LIST_RUNLOCK();
276 
277 	cdg_cc_algo.post_recovery = newreno_cc_algo.post_recovery;
278 	cdg_cc_algo.after_idle = newreno_cc_algo.after_idle;
279 
280 	return (0);
281 }
282 
283 static int
284 cdg_mod_destroy(void)
285 {
286 
287 	uma_zdestroy(qdiffsample_zone);
288 	return (0);
289 }
290 
291 static int
292 cdg_cb_init(struct cc_var *ccv)
293 {
294 	struct cdg *cdg_data;
295 
296 	cdg_data = malloc(sizeof(struct cdg), M_CDG, M_NOWAIT);
297 	if (cdg_data == NULL)
298 		return (ENOMEM);
299 
300 	cdg_data->shadow_w = 0;
301 	cdg_data->max_qtrend = 0;
302 	cdg_data->min_qtrend = 0;
303 	cdg_data->queue_state = CDG_Q_UNKNOWN;
304 	cdg_data->maxrtt_in_rtt = 0;
305 	cdg_data->maxrtt_in_prevrtt = 0;
306 	cdg_data->minrtt_in_rtt = INT_MAX;
307 	cdg_data->minrtt_in_prevrtt = 0;
308 	cdg_data->window_incr = 0;
309 	cdg_data->rtt_count = 0;
310 	cdg_data->consec_cong_cnt = 0;
311 	cdg_data->sample_q_size = V_cdg_smoothing_factor;
312 	cdg_data->num_samples = 0;
313 	STAILQ_INIT(&cdg_data->qdiffmin_q);
314 	STAILQ_INIT(&cdg_data->qdiffmax_q);
315 
316 	ccv->cc_data = cdg_data;
317 
318 	return (0);
319 }
320 
321 static void
322 cdg_conn_init(struct cc_var *ccv)
323 {
324 	struct cdg *cdg_data = ccv->cc_data;
325 
326 	/*
327 	 * Initialise the shadow_cwnd in case we are competing with loss based
328 	 * flows from the start
329 	 */
330 	cdg_data->shadow_w = CCV(ccv, snd_cwnd);
331 }
332 
333 static void
334 cdg_cb_destroy(struct cc_var *ccv)
335 {
336 	struct cdg *cdg_data;
337 	struct qdiff_sample *qds, *qds_n;
338 
339 	cdg_data = ccv->cc_data;
340 
341 	qds = STAILQ_FIRST(&cdg_data->qdiffmin_q);
342 	while (qds != NULL) {
343 		qds_n = STAILQ_NEXT(qds, qdiff_lnk);
344 		uma_zfree(qdiffsample_zone,qds);
345 		qds = qds_n;
346 	}
347 
348 	qds = STAILQ_FIRST(&cdg_data->qdiffmax_q);
349 	while (qds != NULL) {
350 		qds_n = STAILQ_NEXT(qds, qdiff_lnk);
351 		uma_zfree(qdiffsample_zone,qds);
352 		qds = qds_n;
353 	}
354 
355 	free(ccv->cc_data, M_CDG);
356 }
357 
358 static int
359 cdg_beta_handler(SYSCTL_HANDLER_ARGS)
360 {
361 
362 	if (req->newptr != NULL &&
363 	    (CAST_PTR_INT(req->newptr) == 0 || CAST_PTR_INT(req->newptr) > 100))
364 		return (EINVAL);
365 
366 	return (sysctl_handle_int(oidp, arg1, arg2, req));
367 }
368 
369 static int
370 cdg_exp_backoff_scale_handler(SYSCTL_HANDLER_ARGS)
371 {
372 
373 	if (req->newptr != NULL && CAST_PTR_INT(req->newptr) < 1)
374 		return (EINVAL);
375 
376 	return (sysctl_handle_int(oidp, arg1, arg2, req));
377 }
378 
379 static inline uint32_t
380 cdg_window_decrease(struct cc_var *ccv, unsigned long owin, unsigned int beta)
381 {
382 
383 	return ((ulmin(CCV(ccv, snd_wnd), owin) * beta) / 100);
384 }
385 
386 /*
387  * Window increase function
388  * This window increase function is independent of the initial window size
389  * to ensure small window flows are not discriminated against (i.e. fairness).
390  * It increases at 1pkt/rtt like Reno for alpha_inc rtts, and then 2pkts/rtt for
391  * the next alpha_inc rtts, etc.
392  */
393 static void
394 cdg_window_increase(struct cc_var *ccv, int new_measurement)
395 {
396 	struct cdg *cdg_data;
397 	int incr, s_w_incr;
398 
399 	cdg_data = ccv->cc_data;
400 	incr = s_w_incr = 0;
401 
402 	if (CCV(ccv, snd_cwnd) <= CCV(ccv, snd_ssthresh)) {
403 		/* Slow start. */
404 		incr = CCV(ccv, t_maxseg);
405 		s_w_incr = incr;
406 		cdg_data->window_incr = cdg_data->rtt_count = 0;
407 	} else {
408 		/* Congestion avoidance. */
409 		if (new_measurement) {
410 			s_w_incr = CCV(ccv, t_maxseg);
411 			if (V_cdg_alpha_inc == 0) {
412 				incr = CCV(ccv, t_maxseg);
413 			} else {
414 				if (++cdg_data->rtt_count >= V_cdg_alpha_inc) {
415 					cdg_data->window_incr++;
416 					cdg_data->rtt_count = 0;
417 				}
418 				incr = CCV(ccv, t_maxseg) *
419 				    cdg_data->window_incr;
420 			}
421 		}
422 	}
423 
424 	if (cdg_data->shadow_w > 0)
425 		cdg_data->shadow_w = ulmin(cdg_data->shadow_w + s_w_incr,
426 		    TCP_MAXWIN << CCV(ccv, snd_scale));
427 
428 	CCV(ccv, snd_cwnd) = ulmin(CCV(ccv, snd_cwnd) + incr,
429 	    TCP_MAXWIN << CCV(ccv, snd_scale));
430 }
431 
432 static void
433 cdg_cong_signal(struct cc_var *ccv, uint32_t signal_type)
434 {
435 	struct cdg *cdg_data = ccv->cc_data;
436 
437 	switch(signal_type) {
438 	case CC_CDG_DELAY:
439 		CCV(ccv, snd_ssthresh) = cdg_window_decrease(ccv,
440 		    CCV(ccv, snd_cwnd), V_cdg_beta_delay);
441 		CCV(ccv, snd_cwnd) = CCV(ccv, snd_ssthresh);
442 		CCV(ccv, snd_recover) = CCV(ccv, snd_max);
443 		cdg_data->window_incr = cdg_data->rtt_count = 0;
444 		ENTER_CONGRECOVERY(CCV(ccv, t_flags));
445 		break;
446 	case CC_NDUPACK:
447 		/*
448 		 * If already responding to congestion OR we have guessed no
449 		 * queue in the path is full.
450 		 */
451 		if (IN_CONGRECOVERY(CCV(ccv, t_flags)) ||
452 		    cdg_data->queue_state < CDG_Q_FULL) {
453 			CCV(ccv, snd_ssthresh) = CCV(ccv, snd_cwnd);
454 			CCV(ccv, snd_recover) = CCV(ccv, snd_max);
455 		} else {
456 			/*
457 			 * Loss is likely to be congestion related. We have
458 			 * inferred a queue full state, so have shadow window
459 			 * react to loss as NewReno would.
460 			 */
461 			if (cdg_data->shadow_w > 0)
462 				cdg_data->shadow_w = cdg_window_decrease(ccv,
463 				    cdg_data->shadow_w, RENO_BETA);
464 
465 			CCV(ccv, snd_ssthresh) = max(cdg_data->shadow_w,
466 			    cdg_window_decrease(ccv, CCV(ccv, snd_cwnd),
467 			    V_cdg_beta_loss));
468 
469 			cdg_data->window_incr = cdg_data->rtt_count = 0;
470 		}
471 		ENTER_RECOVERY(CCV(ccv, t_flags));
472 		break;
473 	default:
474 		newreno_cc_algo.cong_signal(ccv, signal_type);
475 		break;
476 	}
477 }
478 
479 /*
480  * Using a negative exponential probabilistic backoff so that sources with
481  * varying RTTs which share the same link will, on average, have the same
482  * probability of backoff over time.
483  *
484  * Prob_backoff = 1 - exp(-qtrend / V_cdg_exp_backoff_scale), where
485  * V_cdg_exp_backoff_scale is the average qtrend for the exponential backoff.
486  */
487 static inline int
488 prob_backoff(long qtrend)
489 {
490 	int backoff, idx, p;
491 
492 	backoff = (qtrend > ((MAXGRAD * V_cdg_exp_backoff_scale) << D_P_E));
493 
494 	if (!backoff) {
495 		if (V_cdg_exp_backoff_scale > 1)
496 			idx = (qtrend + V_cdg_exp_backoff_scale / 2) /
497 			    V_cdg_exp_backoff_scale;
498 		else
499 			idx = qtrend;
500 
501 		/* Backoff probability proportional to rate of queue growth. */
502 		p = (INT_MAX / (1 << EXP_PREC)) * probexp[idx];
503 		backoff = (random() < p);
504 	}
505 
506 	return (backoff);
507 }
508 
509 static inline void
510 calc_moving_average(struct cdg *cdg_data, long qdiff_max, long qdiff_min)
511 {
512 	struct qdiff_sample *qds;
513 
514 	++cdg_data->num_samples;
515 	if (cdg_data->num_samples > cdg_data->sample_q_size) {
516 		/* Minimum RTT. */
517 		qds = STAILQ_FIRST(&cdg_data->qdiffmin_q);
518 		cdg_data->min_qtrend =  cdg_data->min_qtrend +
519 		    (qdiff_min - qds->qdiff) / cdg_data->sample_q_size;
520 		STAILQ_REMOVE_HEAD(&cdg_data->qdiffmin_q, qdiff_lnk);
521 		qds->qdiff = qdiff_min;
522 		STAILQ_INSERT_TAIL(&cdg_data->qdiffmin_q, qds, qdiff_lnk);
523 
524 		/* Maximum RTT. */
525 		qds = STAILQ_FIRST(&cdg_data->qdiffmax_q);
526 		cdg_data->max_qtrend =  cdg_data->max_qtrend +
527 		    (qdiff_max - qds->qdiff) / cdg_data->sample_q_size;
528 		STAILQ_REMOVE_HEAD(&cdg_data->qdiffmax_q, qdiff_lnk);
529 		qds->qdiff = qdiff_max;
530 		STAILQ_INSERT_TAIL(&cdg_data->qdiffmax_q, qds, qdiff_lnk);
531 		--cdg_data->num_samples;
532 	} else {
533 		qds = uma_zalloc(qdiffsample_zone, M_NOWAIT);
534 		if (qds != NULL) {
535 			cdg_data->min_qtrend = cdg_data->min_qtrend +
536 			    qdiff_min / cdg_data->sample_q_size;
537 			qds->qdiff = qdiff_min;
538 			STAILQ_INSERT_TAIL(&cdg_data->qdiffmin_q, qds,
539 			    qdiff_lnk);
540 		}
541 
542 		qds = uma_zalloc(qdiffsample_zone, M_NOWAIT);
543 		if (qds) {
544 			cdg_data->max_qtrend = cdg_data->max_qtrend +
545 			    qdiff_max / cdg_data->sample_q_size;
546 			qds->qdiff = qdiff_max;
547 			STAILQ_INSERT_TAIL(&cdg_data->qdiffmax_q, qds,
548 			    qdiff_lnk);
549 		}
550 	}
551 }
552 
553 static void
554 cdg_ack_received(struct cc_var *ccv, uint16_t ack_type)
555 {
556 	struct cdg *cdg_data;
557 	struct ertt *e_t;
558 	long qdiff_max, qdiff_min;
559 	int congestion, new_measurement, slowstart;
560 
561 	cdg_data = ccv->cc_data;
562 	e_t = (struct ertt *)khelp_get_osd(CCV(ccv, osd), ertt_id);
563 	new_measurement = e_t->flags & ERTT_NEW_MEASUREMENT;
564 	congestion = 0;
565 	cdg_data->maxrtt_in_rtt = imax(e_t->rtt, cdg_data->maxrtt_in_rtt);
566 	cdg_data->minrtt_in_rtt = imin(e_t->rtt, cdg_data->minrtt_in_rtt);
567 
568 	if (new_measurement) {
569 		slowstart = (CCV(ccv, snd_cwnd) <= CCV(ccv, snd_ssthresh));
570 		/*
571 		 * Update smoothed gradient measurements. Since we are only
572 		 * using one measurement per RTT, use max or min rtt_in_rtt.
573 		 * This is also less noisy than a sample RTT measurement. Max
574 		 * RTT measurements can have trouble due to OS issues.
575 		 */
576 		if (cdg_data->maxrtt_in_prevrtt) {
577 			qdiff_max = ((long)(cdg_data->maxrtt_in_rtt -
578 			    cdg_data->maxrtt_in_prevrtt) << D_P_E );
579 			qdiff_min = ((long)(cdg_data->minrtt_in_rtt -
580 			    cdg_data->minrtt_in_prevrtt) << D_P_E );
581 
582 			calc_moving_average(cdg_data, qdiff_max, qdiff_min);
583 
584 			/* Probabilistic backoff with respect to gradient. */
585 			if (slowstart && qdiff_min > 0)
586 				congestion = prob_backoff(qdiff_min);
587 			else if (cdg_data->min_qtrend > 0)
588 				congestion = prob_backoff(cdg_data->min_qtrend);
589 			else if (slowstart && qdiff_max > 0)
590 				congestion = prob_backoff(qdiff_max);
591 			else if (cdg_data->max_qtrend > 0)
592 				congestion = prob_backoff(cdg_data->max_qtrend);
593 
594 			/* Update estimate of queue state. */
595 			if (cdg_data->min_qtrend > 0 &&
596 			    cdg_data->max_qtrend <= 0) {
597 				cdg_data->queue_state = CDG_Q_FULL;
598 			} else if (cdg_data->min_qtrend >= 0 &&
599 			    cdg_data->max_qtrend < 0) {
600 				cdg_data->queue_state = CDG_Q_EMPTY;
601 				cdg_data->shadow_w = 0;
602 			} else if (cdg_data->min_qtrend > 0 &&
603 			    cdg_data->max_qtrend > 0) {
604 				cdg_data->queue_state = CDG_Q_RISING;
605 			} else if (cdg_data->min_qtrend < 0 &&
606 			    cdg_data->max_qtrend < 0) {
607 				cdg_data->queue_state = CDG_Q_FALLING;
608 			}
609 
610 			if (cdg_data->min_qtrend < 0 ||
611 			    cdg_data->max_qtrend < 0)
612 				cdg_data->consec_cong_cnt = 0;
613 		}
614 
615 		cdg_data->minrtt_in_prevrtt = cdg_data->minrtt_in_rtt;
616 		cdg_data->minrtt_in_rtt = INT_MAX;
617 		cdg_data->maxrtt_in_prevrtt = cdg_data->maxrtt_in_rtt;
618 		cdg_data->maxrtt_in_rtt = 0;
619 		e_t->flags &= ~ERTT_NEW_MEASUREMENT;
620 	}
621 
622 	if (congestion) {
623 		cdg_data->consec_cong_cnt++;
624 		if (!IN_RECOVERY(CCV(ccv, t_flags))) {
625 			if (cdg_data->consec_cong_cnt <= V_cdg_consec_cong)
626 				cdg_cong_signal(ccv, CC_CDG_DELAY);
627 			else
628 				/*
629 				 * We have been backing off but the queue is not
630 				 * falling. Assume we are competing with
631 				 * loss-based flows and don't back off for the
632 				 * next V_cdg_hold_backoff RTT periods.
633 				 */
634 				if (cdg_data->consec_cong_cnt >=
635 				    V_cdg_consec_cong + V_cdg_hold_backoff)
636 					cdg_data->consec_cong_cnt = 0;
637 
638 			/* Won't see effect until 2nd RTT. */
639 			cdg_data->maxrtt_in_prevrtt = 0;
640 			/*
641 			 * Resync shadow window in case we are competing with a
642 			 * loss based flow
643 			 */
644 			cdg_data->shadow_w = ulmax(CCV(ccv, snd_cwnd),
645 			    cdg_data->shadow_w);
646 		}
647 	} else if (ack_type == CC_ACK)
648 		cdg_window_increase(ccv, new_measurement);
649 }
650 
651 /* When a vnet is created and being initialised, init the per-stack CDG vars. */
652 VNET_SYSINIT(cdg_init_vnet, SI_SUB_PROTO_BEGIN, SI_ORDER_FIRST,
653     cdg_init_vnet, NULL);
654 
655 SYSCTL_DECL(_net_inet_tcp_cc_cdg);
656 SYSCTL_NODE(_net_inet_tcp_cc, OID_AUTO, cdg, CTLFLAG_RW, NULL,
657     "CAIA delay-gradient congestion control related settings");
658 
659 SYSCTL_STRING(_net_inet_tcp_cc_cdg, OID_AUTO, version,
660     CTLFLAG_RD, CDG_VERSION, sizeof(CDG_VERSION) - 1,
661     "Current algorithm/implementation version number");
662 
663 SYSCTL_UINT(_net_inet_tcp_cc_cdg, OID_AUTO, alpha_inc,
664     CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(cdg_alpha_inc), 0,
665     "Increment the window increase factor alpha by 1 MSS segment every "
666     "alpha_inc RTTs during congestion avoidance mode.");
667 
668 SYSCTL_PROC(_net_inet_tcp_cc_cdg, OID_AUTO, beta_delay,
669     CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW, &VNET_NAME(cdg_beta_delay), 70,
670     &cdg_beta_handler, "IU",
671     "Delay-based window decrease factor as a percentage "
672     "(on delay-based backoff, w = w * beta_delay / 100)");
673 
674 SYSCTL_PROC(_net_inet_tcp_cc_cdg, OID_AUTO, beta_loss,
675     CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW, &VNET_NAME(cdg_beta_loss), 50,
676     &cdg_beta_handler, "IU",
677     "Loss-based window decrease factor as a percentage "
678     "(on loss-based backoff, w = w * beta_loss / 100)");
679 
680 SYSCTL_PROC(_net_inet_tcp_cc_cdg, OID_AUTO, exp_backoff_scale,
681     CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW,
682     &VNET_NAME(cdg_exp_backoff_scale), 2, &cdg_exp_backoff_scale_handler, "IU",
683     "Scaling parameter for the probabilistic exponential backoff");
684 
685 SYSCTL_UINT(_net_inet_tcp_cc_cdg,  OID_AUTO, smoothing_factor,
686     CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(cdg_smoothing_factor), 8,
687     "Number of samples used for moving average smoothing (0 = no smoothing)");
688 
689 SYSCTL_UINT(_net_inet_tcp_cc_cdg, OID_AUTO, loss_compete_consec_cong,
690     CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(cdg_consec_cong), 5,
691     "Number of consecutive delay-gradient based congestion episodes which will "
692     "trigger loss based CC compatibility");
693 
694 SYSCTL_UINT(_net_inet_tcp_cc_cdg, OID_AUTO, loss_compete_hold_backoff,
695     CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(cdg_hold_backoff), 5,
696     "Number of consecutive delay-gradient based congestion episodes to hold "
697     "the window backoff for loss based CC compatibility");
698 
699 DECLARE_CC_MODULE(cdg, &cdg_cc_algo);
700 
701 MODULE_DEPEND(cdg, ertt, 1, 1, 1);
702