1 /*-
2 * SPDX-License-Identifier: BSD-2-Clause
3 *
4 * Copyright (c) 2008-2010 Lawrence Stewart <lstewart@freebsd.org>
5 * Copyright (c) 2010 The FreeBSD Foundation
6 * All rights reserved.
7 *
8 * This software was developed by Lawrence Stewart while studying at the Centre
9 * for Advanced Internet Architectures, Swinburne University of Technology, made
10 * possible in part by a grant from the Cisco University Research Program Fund
11 * at Community Foundation Silicon Valley.
12 *
13 * Portions of this software were developed at the Centre for Advanced
14 * Internet Architectures, Swinburne University of Technology, Melbourne,
15 * Australia by David Hayes under sponsorship from the FreeBSD Foundation.
16 *
17 * Redistribution and use in source and binary forms, with or without
18 * modification, are permitted provided that the following conditions
19 * are met:
20 * 1. Redistributions of source code must retain the above copyright
21 * notice, this list of conditions and the following disclaimer.
22 * 2. Redistributions in binary form must reproduce the above copyright
23 * notice, this list of conditions and the following disclaimer in the
24 * documentation and/or other materials provided with the distribution.
25 *
26 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
27 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
28 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
29 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
30 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
31 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
32 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
33 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
34 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
35 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36 * SUCH DAMAGE.
37 */
38
39 /*
40 * An implementation of the CUBIC congestion control algorithm for FreeBSD,
41 * based on the Internet Draft "draft-rhee-tcpm-cubic-02" by Rhee, Xu and Ha.
42 * Originally released as part of the NewTCP research project at Swinburne
43 * University of Technology's Centre for Advanced Internet Architectures,
44 * Melbourne, Australia, which was made possible in part by a grant from the
45 * Cisco University Research Program Fund at Community Foundation Silicon
46 * Valley. More details are available at:
47 * http://caia.swin.edu.au/urp/newtcp/
48 */
49
50 #include <sys/param.h>
51 #include <sys/kernel.h>
52 #include <sys/limits.h>
53 #include <sys/malloc.h>
54 #include <sys/module.h>
55 #include <sys/socket.h>
56 #include <sys/socketvar.h>
57 #include <sys/sysctl.h>
58 #include <sys/systm.h>
59
60 #include <net/vnet.h>
61
62 #include <net/route.h>
63 #include <net/route/nhop.h>
64
65 #include <netinet/in_pcb.h>
66 #include <netinet/tcp.h>
67 #include <netinet/tcp_seq.h>
68 #include <netinet/tcp_timer.h>
69 #include <netinet/tcp_var.h>
70 #include <netinet/tcp_log_buf.h>
71 #include <netinet/tcp_hpts.h>
72 #include <netinet/cc/cc.h>
73 #include <netinet/cc/cc_cubic.h>
74 #include <netinet/cc/cc_module.h>
75
76 static void cubic_ack_received(struct cc_var *ccv, ccsignal_t type);
77 static void cubic_cb_destroy(struct cc_var *ccv);
78 static int cubic_cb_init(struct cc_var *ccv, void *ptr);
79 static void cubic_cong_signal(struct cc_var *ccv, ccsignal_t type);
80 static void cubic_conn_init(struct cc_var *ccv);
81 static int cubic_mod_init(void);
82 static void cubic_post_recovery(struct cc_var *ccv);
83 static void cubic_record_rtt(struct cc_var *ccv);
84 static void cubic_ssthresh_update(struct cc_var *ccv, uint32_t maxseg);
85 static void cubic_after_idle(struct cc_var *ccv);
86 static size_t cubic_data_sz(void);
87 static void cubic_newround(struct cc_var *ccv, uint32_t round_cnt);
88 static void cubic_rttsample(struct cc_var *ccv, uint32_t usec_rtt,
89 uint32_t rxtcnt, uint32_t fas);
90
91 struct cc_algo cubic_cc_algo = {
92 .name = "cubic",
93 .ack_received = cubic_ack_received,
94 .cb_destroy = cubic_cb_destroy,
95 .cb_init = cubic_cb_init,
96 .cong_signal = cubic_cong_signal,
97 .conn_init = cubic_conn_init,
98 .mod_init = cubic_mod_init,
99 .post_recovery = cubic_post_recovery,
100 .after_idle = cubic_after_idle,
101 .cc_data_sz = cubic_data_sz,
102 .rttsample = cubic_rttsample,
103 .newround = cubic_newround
104 };
105
106 static void
cubic_log_hystart_event(struct cc_var * ccv,struct cubic * cubicd,uint8_t mod,uint32_t flex1)107 cubic_log_hystart_event(struct cc_var *ccv, struct cubic *cubicd, uint8_t mod, uint32_t flex1)
108 {
109 /*
110 * Types of logs (mod value)
111 * 1 - rtt_thresh in flex1, checking to see if RTT is to great.
112 * 2 - rtt is too great, rtt_thresh in flex1.
113 * 3 - CSS is active incr in flex1
114 * 4 - A new round is beginning flex1 is round count
115 * 5 - A new RTT measurement flex1 is the new measurement.
116 * 6 - We enter CA ssthresh is also in flex1.
117 * 7 - Socket option to change hystart executed opt.val in flex1.
118 * 8 - Back out of CSS into SS, flex1 is the css_baseline_minrtt
119 * 9 - We enter CA, via an ECN mark.
120 * 10 - We enter CA, via a loss.
121 * 11 - We have slipped out of SS into CA via cwnd growth.
122 * 12 - After idle has re-enabled hystart++
123 */
124 struct tcpcb *tp;
125
126 if (hystart_bblogs == 0)
127 return;
128 tp = ccv->tp;
129 if (tcp_bblogging_on(tp)) {
130 union tcp_log_stackspecific log;
131 struct timeval tv;
132
133 memset(&log, 0, sizeof(log));
134 log.u_bbr.flex1 = flex1;
135 log.u_bbr.flex2 = cubicd->css_current_round_minrtt;
136 log.u_bbr.flex3 = cubicd->css_lastround_minrtt;
137 log.u_bbr.flex4 = cubicd->css_rttsample_count;
138 log.u_bbr.flex5 = cubicd->css_entered_at_round;
139 log.u_bbr.flex6 = cubicd->css_baseline_minrtt;
140 /* We only need bottom 16 bits of flags */
141 log.u_bbr.flex7 = cubicd->flags & 0x0000ffff;
142 log.u_bbr.flex8 = mod;
143 log.u_bbr.epoch = cubicd->css_current_round;
144 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
145 log.u_bbr.lt_epoch = cubicd->css_fas_at_css_entry;
146 log.u_bbr.pkts_out = cubicd->css_last_fas;
147 log.u_bbr.delivered = cubicd->css_lowrtt_fas;
148 log.u_bbr.pkt_epoch = ccv->flags;
149 TCP_LOG_EVENTP(tp, NULL,
150 &tptosocket(tp)->so_rcv,
151 &tptosocket(tp)->so_snd,
152 TCP_HYSTART, 0,
153 0, &log, false, &tv);
154 }
155 }
156
157 static void
cubic_does_slow_start(struct cc_var * ccv,struct cubic * cubicd)158 cubic_does_slow_start(struct cc_var *ccv, struct cubic *cubicd)
159 {
160 /*
161 * In slow-start with ABC enabled and no RTO in sight?
162 * (Must not use abc_l_var > 1 if slow starting after
163 * an RTO. On RTO, snd_nxt = snd_una, so the
164 * snd_nxt == snd_max check is sufficient to
165 * handle this).
166 *
167 * XXXLAS: Find a way to signal SS after RTO that
168 * doesn't rely on tcpcb vars.
169 */
170 u_int cw = CCV(ccv, snd_cwnd);
171 uint32_t mss = tcp_fixed_maxseg(ccv->tp);
172 u_int incr = mss;
173 uint16_t abc_val;
174
175 cubicd->flags |= CUBICFLAG_IN_SLOWSTART;
176 if (ccv->flags & CCF_USE_LOCAL_ABC)
177 abc_val = ccv->labc;
178 else
179 abc_val = V_tcp_abc_l_var;
180 if ((ccv->flags & CCF_HYSTART_ALLOWED) &&
181 (cubicd->flags & CUBICFLAG_HYSTART_ENABLED) &&
182 ((cubicd->flags & CUBICFLAG_HYSTART_IN_CSS) == 0)) {
183 /*
184 * Hystart is allowed and still enabled and we are not yet
185 * in CSS. Lets check to see if we can make a decision on
186 * if we need to go into CSS.
187 */
188 if ((cubicd->css_rttsample_count >= hystart_n_rttsamples) &&
189 (cubicd->css_current_round_minrtt != 0xffffffff) &&
190 (cubicd->css_lastround_minrtt != 0xffffffff)) {
191 uint32_t rtt_thresh;
192
193 /* Clamp (minrtt_thresh, lastround/8, maxrtt_thresh) */
194 rtt_thresh = (cubicd->css_lastround_minrtt >> 3);
195 if (rtt_thresh < hystart_minrtt_thresh)
196 rtt_thresh = hystart_minrtt_thresh;
197 if (rtt_thresh > hystart_maxrtt_thresh)
198 rtt_thresh = hystart_maxrtt_thresh;
199 cubic_log_hystart_event(ccv, cubicd, 1, rtt_thresh);
200
201 if (cubicd->css_current_round_minrtt >= (cubicd->css_lastround_minrtt + rtt_thresh)) {
202 /* Enter CSS */
203 cubicd->flags |= CUBICFLAG_HYSTART_IN_CSS;
204 cubicd->css_fas_at_css_entry = cubicd->css_lowrtt_fas;
205 /*
206 * The draft (v4) calls for us to set baseline to css_current_round_min
207 * but that can cause an oscillation. We probably shoudl be using
208 * css_lastround_minrtt, but the authors insist that will cause
209 * issues on exiting early. We will leave the draft version for now
210 * but I suspect this is incorrect.
211 */
212 cubicd->css_baseline_minrtt = cubicd->css_current_round_minrtt;
213 cubicd->css_entered_at_round = cubicd->css_current_round;
214 cubic_log_hystart_event(ccv, cubicd, 2, rtt_thresh);
215 }
216 }
217 }
218 if (CCV(ccv, snd_nxt) == CCV(ccv, snd_max))
219 incr = min(ccv->bytes_this_ack,
220 ccv->nsegs * abc_val * mss);
221 else
222 incr = min(ccv->bytes_this_ack, mss);
223
224 /* Only if Hystart is enabled will the flag get set */
225 if (cubicd->flags & CUBICFLAG_HYSTART_IN_CSS) {
226 incr /= hystart_css_growth_div;
227 cubic_log_hystart_event(ccv, cubicd, 3, incr);
228 }
229 /* ABC is on by default, so incr equals 0 frequently. */
230 if (incr > 0)
231 CCV(ccv, snd_cwnd) = min((cw + incr),
232 TCP_MAXWIN << CCV(ccv, snd_scale));
233 }
234
235 static void
cubic_ack_received(struct cc_var * ccv,ccsignal_t type)236 cubic_ack_received(struct cc_var *ccv, ccsignal_t type)
237 {
238 struct cubic *cubic_data;
239 unsigned long W_est, W_cubic;
240 int usecs_since_epoch;
241 uint32_t mss = tcp_fixed_maxseg(ccv->tp);
242
243 cubic_data = ccv->cc_data;
244 cubic_record_rtt(ccv);
245
246 /*
247 * For a regular ACK and we're not in cong/fast recovery and
248 * we're cwnd limited, always recalculate cwnd.
249 */
250 if (type == CC_ACK && !IN_RECOVERY(CCV(ccv, t_flags)) &&
251 (ccv->flags & CCF_CWND_LIMITED)) {
252 /* Use the logic in NewReno ack_received() for slow start. */
253 if (CCV(ccv, snd_cwnd) <= CCV(ccv, snd_ssthresh) ||
254 cubic_data->min_rtt_usecs == TCPTV_SRTTBASE) {
255 cubic_does_slow_start(ccv, cubic_data);
256 } else {
257 if (cubic_data->flags & CUBICFLAG_HYSTART_IN_CSS) {
258 /*
259 * We have slipped into CA with
260 * CSS active. Deactivate all.
261 */
262 /* Turn off the CSS flag */
263 cubic_data->flags &= ~CUBICFLAG_HYSTART_IN_CSS;
264 /* Disable use of CSS in the future except long idle */
265 cubic_data->flags &= ~CUBICFLAG_HYSTART_ENABLED;
266 cubic_log_hystart_event(ccv, cubic_data, 11, CCV(ccv, snd_ssthresh));
267 }
268 if ((cubic_data->flags & CUBICFLAG_RTO_EVENT) &&
269 (cubic_data->flags & CUBICFLAG_IN_SLOWSTART)) {
270 /* RFC8312 Section 4.7 */
271 cubic_data->flags &= ~(CUBICFLAG_RTO_EVENT |
272 CUBICFLAG_IN_SLOWSTART);
273 cubic_data->W_max = CCV(ccv, snd_cwnd);
274 cubic_data->t_epoch = ticks;
275 cubic_data->K = 0;
276 } else if (cubic_data->flags & (CUBICFLAG_IN_SLOWSTART |
277 CUBICFLAG_IN_APPLIMIT)) {
278 cubic_data->flags &= ~(CUBICFLAG_IN_SLOWSTART |
279 CUBICFLAG_IN_APPLIMIT);
280 cubic_data->t_epoch = ticks;
281 cubic_data->K = cubic_k(cubic_data->W_max / mss);
282 }
283 usecs_since_epoch = (ticks - cubic_data->t_epoch) * tick;
284 if (usecs_since_epoch < 0) {
285 /*
286 * dragging t_epoch along
287 */
288 usecs_since_epoch = INT_MAX;
289 cubic_data->t_epoch = ticks - INT_MAX;
290 }
291
292 W_est = tf_cwnd(ccv);
293
294 /*
295 * The mean RTT is used to best reflect the equations in
296 * the I-D.
297 */
298 W_cubic = cubic_cwnd(usecs_since_epoch +
299 cubic_data->mean_rtt_usecs,
300 cubic_data->W_max,
301 tcp_fixed_maxseg(ccv->tp),
302 cubic_data->K);
303
304 if (W_cubic < W_est) {
305 /*
306 * TCP-friendly region, follow tf
307 * cwnd growth.
308 */
309 CCV(ccv, snd_cwnd) = ulmin(W_est, INT_MAX);
310 cubic_data->flags |= CUBICFLAG_IN_TF;
311 } else if (CCV(ccv, snd_cwnd) < W_cubic) {
312 /*
313 * Concave or convex region, follow CUBIC
314 * cwnd growth.
315 * Only update snd_cwnd, if it doesn't shrink.
316 */
317 CCV(ccv, snd_cwnd) = ulmin(W_cubic, INT_MAX);
318 cubic_data->flags &= ~CUBICFLAG_IN_TF;
319 }
320
321 /*
322 * If we're not in slow start and we're probing for a
323 * new cwnd limit at the start of a connection
324 * (happens when hostcache has a relevant entry),
325 * keep updating our current estimate of the
326 * W_max.
327 */
328 if (((cubic_data->flags & CUBICFLAG_CONG_EVENT) == 0) &&
329 cubic_data->W_max < CCV(ccv, snd_cwnd)) {
330 cubic_data->W_max = CCV(ccv, snd_cwnd);
331 cubic_data->K = cubic_k(cubic_data->W_max /
332 tcp_fixed_maxseg(ccv->tp));
333 }
334 }
335 } else if (type == CC_ACK && !IN_RECOVERY(CCV(ccv, t_flags)) &&
336 !(ccv->flags & CCF_CWND_LIMITED)) {
337 cubic_data->flags |= CUBICFLAG_IN_APPLIMIT;
338 }
339 }
340
341 /*
342 * This is a CUBIC specific implementation of after_idle.
343 * - Reset cwnd by calling New Reno implementation of after_idle.
344 * - Reset t_epoch.
345 */
346 static void
cubic_after_idle(struct cc_var * ccv)347 cubic_after_idle(struct cc_var *ccv)
348 {
349 struct cubic *cubic_data;
350
351 cubic_data = ccv->cc_data;
352
353 cubic_data->W_max = ulmax(cubic_data->W_max, CCV(ccv, snd_cwnd));
354 cubic_data->K = cubic_k(cubic_data->W_max / tcp_fixed_maxseg(ccv->tp));
355 if ((cubic_data->flags & CUBICFLAG_HYSTART_ENABLED) == 0) {
356 /*
357 * Re-enable hystart if we have been idle.
358 */
359 cubic_data->flags &= ~CUBICFLAG_HYSTART_IN_CSS;
360 cubic_data->flags |= CUBICFLAG_HYSTART_ENABLED;
361 cubic_log_hystart_event(ccv, cubic_data, 12, CCV(ccv, snd_ssthresh));
362 }
363 newreno_cc_after_idle(ccv);
364 cubic_data->t_epoch = ticks;
365 }
366
367 static void
cubic_cb_destroy(struct cc_var * ccv)368 cubic_cb_destroy(struct cc_var *ccv)
369 {
370 free(ccv->cc_data, M_CC_MEM);
371 }
372
373 static size_t
cubic_data_sz(void)374 cubic_data_sz(void)
375 {
376 return (sizeof(struct cubic));
377 }
378
379 static int
cubic_cb_init(struct cc_var * ccv,void * ptr)380 cubic_cb_init(struct cc_var *ccv, void *ptr)
381 {
382 struct cubic *cubic_data;
383
384 INP_WLOCK_ASSERT(tptoinpcb(ccv->tp));
385 if (ptr == NULL) {
386 cubic_data = malloc(sizeof(struct cubic), M_CC_MEM, M_NOWAIT|M_ZERO);
387 if (cubic_data == NULL)
388 return (ENOMEM);
389 } else
390 cubic_data = ptr;
391
392 /* Init some key variables with sensible defaults. */
393 cubic_data->t_epoch = ticks;
394 cubic_data->min_rtt_usecs = TCPTV_SRTTBASE;
395 cubic_data->mean_rtt_usecs = 1;
396
397 ccv->cc_data = cubic_data;
398 cubic_data->flags = CUBICFLAG_HYSTART_ENABLED;
399 /* At init set both to infinity */
400 cubic_data->css_lastround_minrtt = 0xffffffff;
401 cubic_data->css_current_round_minrtt = 0xffffffff;
402 cubic_data->css_current_round = 0;
403 cubic_data->css_baseline_minrtt = 0xffffffff;
404 cubic_data->css_rttsample_count = 0;
405 cubic_data->css_entered_at_round = 0;
406 cubic_data->css_fas_at_css_entry = 0;
407 cubic_data->css_lowrtt_fas = 0;
408 cubic_data->css_last_fas = 0;
409
410 return (0);
411 }
412
413 /*
414 * Perform any necessary tasks before we enter congestion recovery.
415 */
416 static void
cubic_cong_signal(struct cc_var * ccv,ccsignal_t type)417 cubic_cong_signal(struct cc_var *ccv, ccsignal_t type)
418 {
419 struct cubic *cubic_data;
420 uint32_t mss, pipe;
421
422 cubic_data = ccv->cc_data;
423 mss = tcp_fixed_maxseg(ccv->tp);
424
425 switch (type) {
426 case CC_NDUPACK:
427 if (cubic_data->flags & CUBICFLAG_HYSTART_ENABLED) {
428 /* Make sure the flags are all off we had a loss */
429 cubic_data->flags &= ~CUBICFLAG_HYSTART_ENABLED;
430 cubic_data->flags &= ~CUBICFLAG_HYSTART_IN_CSS;
431 cubic_log_hystart_event(ccv, cubic_data, 10, CCV(ccv, snd_ssthresh));
432 }
433 if (!IN_FASTRECOVERY(CCV(ccv, t_flags))) {
434 if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) {
435 cubic_ssthresh_update(ccv, mss);
436 cubic_data->flags |= CUBICFLAG_CONG_EVENT;
437 cubic_data->t_epoch = ticks;
438 cubic_data->K = cubic_k(cubic_data->W_max / mss);
439 }
440 ENTER_RECOVERY(CCV(ccv, t_flags));
441 }
442 break;
443
444 case CC_ECN:
445 if (cubic_data->flags & CUBICFLAG_HYSTART_ENABLED) {
446 /* Make sure the flags are all off we had a loss */
447 cubic_data->flags &= ~CUBICFLAG_HYSTART_ENABLED;
448 cubic_data->flags &= ~CUBICFLAG_HYSTART_IN_CSS;
449 cubic_log_hystart_event(ccv, cubic_data, 9, CCV(ccv, snd_ssthresh));
450 }
451 if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) {
452 cubic_ssthresh_update(ccv, mss);
453 cubic_data->flags |= CUBICFLAG_CONG_EVENT;
454 cubic_data->t_epoch = ticks;
455 cubic_data->K = cubic_k(cubic_data->W_max / mss);
456 CCV(ccv, snd_cwnd) = CCV(ccv, snd_ssthresh);
457 ENTER_CONGRECOVERY(CCV(ccv, t_flags));
458 }
459 break;
460
461 case CC_RTO:
462 /* RFC8312 Section 4.7 */
463 if (CCV(ccv, t_rxtshift) == 1) {
464 /*
465 * Remember the state only for the first RTO event. This
466 * will help us restore the state to the values seen
467 * at the most recent congestion avoidance stage before
468 * the current RTO event.
469 */
470 cubic_data->undo_t_epoch = cubic_data->t_epoch;
471 cubic_data->undo_cwnd_epoch = cubic_data->cwnd_epoch;
472 cubic_data->undo_W_est = cubic_data->W_est;
473 cubic_data->undo_cwnd_prior = cubic_data->cwnd_prior;
474 cubic_data->undo_W_max = cubic_data->W_max;
475 cubic_data->undo_K = cubic_data->K;
476 if (V_tcp_do_newsack) {
477 pipe = tcp_compute_pipe(ccv->tp);
478 } else {
479 pipe = CCV(ccv, snd_max) -
480 CCV(ccv, snd_fack) +
481 CCV(ccv, sackhint.sack_bytes_rexmit);
482 }
483 CCV(ccv, snd_ssthresh) = max(2,
484 (((uint64_t)min(CCV(ccv, snd_wnd), pipe) *
485 CUBIC_BETA) >> CUBIC_SHIFT) / mss) * mss;
486 }
487 cubic_data->flags |= CUBICFLAG_CONG_EVENT | CUBICFLAG_RTO_EVENT;
488 cubic_data->undo_W_max = cubic_data->W_max;
489 CCV(ccv, snd_cwnd) = mss;
490 break;
491
492 case CC_RTO_ERR:
493 cubic_data->flags &= ~(CUBICFLAG_CONG_EVENT | CUBICFLAG_RTO_EVENT);
494 cubic_data->K = cubic_data->undo_K;
495 cubic_data->cwnd_prior = cubic_data->undo_cwnd_prior;
496 cubic_data->W_max = cubic_data->undo_W_max;
497 cubic_data->W_est = cubic_data->undo_W_est;
498 cubic_data->cwnd_epoch = cubic_data->undo_cwnd_epoch;
499 cubic_data->t_epoch = cubic_data->undo_t_epoch;
500 break;
501 default:
502 break;
503 }
504 }
505
506 static void
cubic_conn_init(struct cc_var * ccv)507 cubic_conn_init(struct cc_var *ccv)
508 {
509 struct cubic *cubic_data;
510
511 cubic_data = ccv->cc_data;
512
513 /*
514 * Ensure we have a sane initial value for W_max recorded. Without
515 * this here bad things happen when entries from the TCP hostcache
516 * get used.
517 */
518 cubic_data->W_max = CCV(ccv, snd_cwnd);
519 }
520
521 static int
cubic_mod_init(void)522 cubic_mod_init(void)
523 {
524 return (0);
525 }
526
527 /*
528 * Perform any necessary tasks before we exit congestion recovery.
529 */
530 static void
cubic_post_recovery(struct cc_var * ccv)531 cubic_post_recovery(struct cc_var *ccv)
532 {
533 struct cubic *cubic_data;
534 int pipe;
535 uint32_t mss = tcp_fixed_maxseg(ccv->tp);
536
537 cubic_data = ccv->cc_data;
538 pipe = 0;
539
540 if (IN_FASTRECOVERY(CCV(ccv, t_flags))) {
541 /*
542 * If inflight data is less than ssthresh, set cwnd
543 * conservatively to avoid a burst of data, as suggested in
544 * the NewReno RFC. Otherwise, use the CUBIC method.
545 *
546 * XXXLAS: Find a way to do this without needing curack
547 */
548 if (V_tcp_do_newsack)
549 pipe = tcp_compute_pipe(ccv->tp);
550 else
551 pipe = CCV(ccv, snd_max) - ccv->curack;
552
553 if (pipe < CCV(ccv, snd_ssthresh))
554 /*
555 * Ensure that cwnd does not collapse to 1 MSS under
556 * adverse conditions. Implements RFC6582
557 */
558 CCV(ccv, snd_cwnd) = max(pipe, mss) + mss;
559 else
560 /* Update cwnd based on beta and adjusted W_max. */
561 CCV(ccv, snd_cwnd) = max(((uint64_t)cubic_data->W_max *
562 CUBIC_BETA) >> CUBIC_SHIFT,
563 2 * mss);
564 }
565
566 /* Calculate the average RTT between congestion epochs. */
567 if (cubic_data->epoch_ack_count > 0 &&
568 cubic_data->sum_rtt_usecs >= cubic_data->epoch_ack_count) {
569 cubic_data->mean_rtt_usecs = (int)(cubic_data->sum_rtt_usecs /
570 cubic_data->epoch_ack_count);
571 }
572
573 cubic_data->epoch_ack_count = 0;
574 cubic_data->sum_rtt_usecs = 0;
575 }
576
577 /*
578 * Record the min RTT and sum samples for the epoch average RTT calculation.
579 */
580 static void
cubic_record_rtt(struct cc_var * ccv)581 cubic_record_rtt(struct cc_var *ccv)
582 {
583 struct cubic *cubic_data;
584 uint32_t t_srtt_usecs;
585
586 /* Ignore srtt until a min number of samples have been taken. */
587 if (CCV(ccv, t_rttupdated) >= CUBIC_MIN_RTT_SAMPLES) {
588 cubic_data = ccv->cc_data;
589 t_srtt_usecs = tcp_get_srtt(ccv->tp,
590 TCP_TMR_GRANULARITY_USEC);
591 /*
592 * Record the current SRTT as our minrtt if it's the smallest
593 * we've seen or minrtt is currently equal to its initialised
594 * value.
595 *
596 * XXXLAS: Should there be some hysteresis for minrtt?
597 */
598 if ((t_srtt_usecs < cubic_data->min_rtt_usecs ||
599 cubic_data->min_rtt_usecs == TCPTV_SRTTBASE)) {
600 /* A minimal rtt is a single unshifted tick of a ticks
601 * timer. */
602 cubic_data->min_rtt_usecs = max(tick >> TCP_RTT_SHIFT,
603 t_srtt_usecs);
604
605 /*
606 * If the connection is within its first congestion
607 * epoch, ensure we prime mean_rtt_usecs with a
608 * reasonable value until the epoch average RTT is
609 * calculated in cubic_post_recovery().
610 */
611 if (cubic_data->min_rtt_usecs >
612 cubic_data->mean_rtt_usecs)
613 cubic_data->mean_rtt_usecs =
614 cubic_data->min_rtt_usecs;
615 }
616
617 /* Sum samples for epoch average RTT calculation. */
618 cubic_data->sum_rtt_usecs += t_srtt_usecs;
619 cubic_data->epoch_ack_count++;
620 }
621 }
622
623 /*
624 * Update the ssthresh in the event of congestion.
625 */
626 static void
cubic_ssthresh_update(struct cc_var * ccv,uint32_t maxseg)627 cubic_ssthresh_update(struct cc_var *ccv, uint32_t maxseg)
628 {
629 struct cubic *cubic_data;
630 uint32_t ssthresh;
631 uint32_t cwnd;
632
633 cubic_data = ccv->cc_data;
634 cwnd = CCV(ccv, snd_cwnd);
635
636 /* Fast convergence heuristic. */
637 if (cwnd < cubic_data->W_max) {
638 cwnd = ((uint64_t)cwnd * CUBIC_FC_FACTOR) >> CUBIC_SHIFT;
639 }
640 cubic_data->undo_W_max = cubic_data->W_max;
641 cubic_data->W_max = cwnd;
642
643 if (cubic_data->flags & CUBICFLAG_IN_TF) {
644 /* If in the TCP friendly region, follow what newreno does */
645 ssthresh = newreno_cc_cwnd_on_multiplicative_decrease(ccv, maxseg);
646
647 } else if ((cubic_data->flags & CUBICFLAG_CONG_EVENT) == 0) {
648 /*
649 * On the first congestion event, set ssthresh to cwnd * 0.5
650 * and reduce W_max to cwnd * beta. This aligns the cubic
651 * concave region appropriately.
652 */
653 ssthresh = cwnd >> 1;
654 cubic_data->W_max = ((uint64_t)cwnd * CUBIC_BETA) >> CUBIC_SHIFT;
655 } else {
656 /*
657 * On subsequent congestion events, set ssthresh to cwnd * beta.
658 */
659 ssthresh = ((uint64_t)cwnd * CUBIC_BETA) >> CUBIC_SHIFT;
660 }
661 CCV(ccv, snd_ssthresh) = max(ssthresh, 2 * maxseg);
662 }
663
664 static void
cubic_rttsample(struct cc_var * ccv,uint32_t usec_rtt,uint32_t rxtcnt,uint32_t fas)665 cubic_rttsample(struct cc_var *ccv, uint32_t usec_rtt, uint32_t rxtcnt, uint32_t fas)
666 {
667 struct cubic *cubicd;
668
669 cubicd = ccv->cc_data;
670 if (rxtcnt > 1) {
671 /*
672 * Only look at RTT's that are non-ambiguous.
673 */
674 return;
675 }
676 cubicd->css_rttsample_count++;
677 cubicd->css_last_fas = fas;
678 if (cubicd->css_current_round_minrtt > usec_rtt) {
679 cubicd->css_current_round_minrtt = usec_rtt;
680 cubicd->css_lowrtt_fas = cubicd->css_last_fas;
681 }
682 if ((cubicd->css_rttsample_count >= hystart_n_rttsamples) &&
683 (cubicd->css_current_round_minrtt != 0xffffffff) &&
684 (cubicd->css_current_round_minrtt < cubicd->css_baseline_minrtt) &&
685 (cubicd->css_lastround_minrtt != 0xffffffff)) {
686 /*
687 * We were in CSS and the RTT is now less, we
688 * entered CSS erroneously.
689 */
690 cubicd->flags &= ~CUBICFLAG_HYSTART_IN_CSS;
691 cubic_log_hystart_event(ccv, cubicd, 8, cubicd->css_baseline_minrtt);
692 cubicd->css_baseline_minrtt = 0xffffffff;
693 }
694 if (cubicd->flags & CUBICFLAG_HYSTART_ENABLED)
695 cubic_log_hystart_event(ccv, cubicd, 5, usec_rtt);
696 }
697
698 static void
cubic_newround(struct cc_var * ccv,uint32_t round_cnt)699 cubic_newround(struct cc_var *ccv, uint32_t round_cnt)
700 {
701 struct cubic *cubicd;
702
703 cubicd = ccv->cc_data;
704 /* We have entered a new round */
705 cubicd->css_lastround_minrtt = cubicd->css_current_round_minrtt;
706 cubicd->css_current_round_minrtt = 0xffffffff;
707 cubicd->css_rttsample_count = 0;
708 cubicd->css_current_round = round_cnt;
709 if ((cubicd->flags & CUBICFLAG_HYSTART_IN_CSS) &&
710 ((round_cnt - cubicd->css_entered_at_round) >= hystart_css_rounds)) {
711 /* Enter CA */
712 if (ccv->flags & CCF_HYSTART_CAN_SH_CWND) {
713 /*
714 * We engage more than snd_ssthresh, engage
715 * the brakes!! Though we will stay in SS to
716 * creep back up again, so lets leave CSS active
717 * and give us hystart_css_rounds more rounds.
718 */
719 if (ccv->flags & CCF_HYSTART_CONS_SSTH) {
720 CCV(ccv, snd_ssthresh) = ((cubicd->css_lowrtt_fas + cubicd->css_fas_at_css_entry) / 2);
721 } else {
722 CCV(ccv, snd_ssthresh) = cubicd->css_lowrtt_fas;
723 }
724 CCV(ccv, snd_cwnd) = cubicd->css_fas_at_css_entry;
725 cubicd->css_entered_at_round = round_cnt;
726 } else {
727 CCV(ccv, snd_ssthresh) = CCV(ccv, snd_cwnd);
728 /* Turn off the CSS flag */
729 cubicd->flags &= ~CUBICFLAG_HYSTART_IN_CSS;
730 /* Disable use of CSS in the future except long idle */
731 cubicd->flags &= ~CUBICFLAG_HYSTART_ENABLED;
732 }
733 cubic_log_hystart_event(ccv, cubicd, 6, CCV(ccv, snd_ssthresh));
734 }
735 if (cubicd->flags & CUBICFLAG_HYSTART_ENABLED)
736 cubic_log_hystart_event(ccv, cubicd, 4, round_cnt);
737 }
738
739 DECLARE_CC_MODULE(cubic, &cubic_cc_algo);
740 MODULE_VERSION(cubic, 2);
741