1 /*
2 * Copyright (c) 2013-2020 Apple Inc. All rights reserved.
3 *
4 * @APPLE_OSREFERENCE_LICENSE_HEADER_START@
5 *
6 * This file contains Original Code and/or Modifications of Original Code
7 * as defined in and that are subject to the Apple Public Source License
8 * Version 2.0 (the 'License'). You may not use this file except in
9 * compliance with the License. The rights granted to you under the License
10 * may not be used to create, or enable the creation or redistribution of,
11 * unlawful or unlicensed copies of an Apple operating system, or to
12 * circumvent, violate, or enable the circumvention or violation of, any
13 * terms of an Apple operating system software license agreement.
14 *
15 * Please obtain a copy of the License at
16 * http://www.opensource.apple.com/apsl/ and read it before using this file.
17 *
18 * The Original Code and all software distributed under the License are
19 * distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER
20 * EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES,
21 * INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF MERCHANTABILITY,
22 * FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT.
23 * Please see the License for the specific language governing rights and
24 * limitations under the License.
25 *
26 * @APPLE_OSREFERENCE_LICENSE_HEADER_END@
27 */
28 #include <sys/param.h>
29 #include <sys/systm.h>
30 #include <sys/kernel.h>
31 #include <sys/protosw.h>
32 #include <sys/socketvar.h>
33 #include <sys/syslog.h>
34
35 #include <net/route.h>
36 #include <netinet/in.h>
37 #include <netinet/in_systm.h>
38 #include <netinet/ip.h>
39
40 #include <netinet/ip6.h>
41
42 #include <netinet/ip_var.h>
43 #include <netinet/tcp.h>
44 #include <netinet/tcp_timer.h>
45 #include <netinet/tcp_var.h>
46 #include <netinet/tcp_fsm.h>
47 #include <netinet/tcp_var.h>
48 #include <netinet/tcp_cc.h>
49 #include <netinet/tcpip.h>
50 #include <netinet/tcp_seq.h>
51 #include <kern/task.h>
52 #include <libkern/OSAtomic.h>
53
54 static int tcp_cubic_init(struct tcpcb *tp);
55 static int tcp_cubic_cleanup(struct tcpcb *tp);
56 static void tcp_cubic_cwnd_init_or_reset(struct tcpcb *tp);
57 static void tcp_cubic_congestion_avd(struct tcpcb *tp, struct tcphdr *th);
58 static void tcp_cubic_ack_rcvd(struct tcpcb *tp, struct tcphdr *th);
59 static void tcp_cubic_pre_fr(struct tcpcb *tp);
60 static void tcp_cubic_post_fr(struct tcpcb *tp, struct tcphdr *th);
61 static void tcp_cubic_after_timeout(struct tcpcb *tp);
62 static int tcp_cubic_delay_ack(struct tcpcb *tp, struct tcphdr *th);
63 static void tcp_cubic_switch_cc(struct tcpcb *tp);
64 static uint32_t tcp_cubic_update(struct tcpcb *tp, uint32_t rtt);
65 static inline void tcp_cubic_clear_state(struct tcpcb *tp);
66
67
68 extern float cbrtf(float x);
69
70 struct tcp_cc_algo tcp_cc_cubic = {
71 .name = "cubic",
72 .init = tcp_cubic_init,
73 .cleanup = tcp_cubic_cleanup,
74 .cwnd_init = tcp_cubic_cwnd_init_or_reset,
75 .congestion_avd = tcp_cubic_congestion_avd,
76 .ack_rcvd = tcp_cubic_ack_rcvd,
77 .pre_fr = tcp_cubic_pre_fr,
78 .post_fr = tcp_cubic_post_fr,
79 .after_idle = tcp_cubic_cwnd_init_or_reset,
80 .after_timeout = tcp_cubic_after_timeout,
81 .delay_ack = tcp_cubic_delay_ack,
82 .switch_to = tcp_cubic_switch_cc
83 };
84
85 static float tcp_cubic_backoff = 0.2f; /* multiplicative decrease factor */
86 static float tcp_cubic_coeff = 0.4f;
87 static float tcp_cubic_fast_convergence_factor = 0.875f;
88
89 static float tcp_cubic_beta = 0.8f;
90
91 SYSCTL_SKMEM_TCP_INT(OID_AUTO, cubic_tcp_friendliness, CTLFLAG_RW | CTLFLAG_LOCKED,
92 static int, tcp_cubic_tcp_friendliness, 0, "Enable TCP friendliness");
93
94 SYSCTL_SKMEM_TCP_INT(OID_AUTO, cubic_fast_convergence, CTLFLAG_RW | CTLFLAG_LOCKED,
95 static int, tcp_cubic_fast_convergence, 0, "Enable fast convergence");
96
97 SYSCTL_SKMEM_TCP_INT(OID_AUTO, cubic_use_minrtt, CTLFLAG_RW | CTLFLAG_LOCKED,
98 static int, tcp_cubic_use_minrtt, 0, "use a min of 5 sec rtt");
99
100 SYSCTL_SKMEM_TCP_INT(OID_AUTO, cubic_minor_fixes, CTLFLAG_RW | CTLFLAG_LOCKED,
101 int, tcp_cubic_minor_fixes, 1, "Minor fixes to TCP Cubic");
102
103 SYSCTL_SKMEM_TCP_INT(OID_AUTO, cubic_rfc_compliant, CTLFLAG_RW | CTLFLAG_LOCKED,
104 int, tcp_cubic_rfc_compliant, 1, "RFC Compliance for TCP Cubic");
105
106 static int
tcp_cubic_init(struct tcpcb * tp)107 tcp_cubic_init(struct tcpcb *tp)
108 {
109 OSIncrementAtomic((volatile SInt32 *)&tcp_cc_cubic.num_sockets);
110
111 if (tcp_cubic_rfc_compliant) {
112 tcp_cubic_backoff = 0.3f; /* multiplicative decrease factor */
113 tcp_cubic_fast_convergence_factor = 0.85f;
114 tcp_cubic_beta = 0.7f;
115 } else {
116 tcp_cubic_backoff = 0.2f; /* multiplicative decrease factor */
117 tcp_cubic_fast_convergence_factor = 0.875f;
118 tcp_cubic_beta = 0.8f;
119 }
120
121 VERIFY(tp->t_ccstate != NULL);
122 tcp_cubic_clear_state(tp);
123 return 0;
124 }
125
126 static int
tcp_cubic_cleanup(struct tcpcb * tp)127 tcp_cubic_cleanup(struct tcpcb *tp)
128 {
129 #pragma unused(tp)
130 OSDecrementAtomic((volatile SInt32 *)&tcp_cc_cubic.num_sockets);
131 return 0;
132 }
133
134 /*
135 * Initialize the congestion window at the beginning of a connection or
136 * after idle time
137 */
138 static void
tcp_cubic_cwnd_init_or_reset(struct tcpcb * tp)139 tcp_cubic_cwnd_init_or_reset(struct tcpcb *tp)
140 {
141 VERIFY(tp->t_ccstate != NULL);
142
143 tcp_cubic_clear_state(tp);
144 tcp_cc_cwnd_init_or_reset(tp);
145 tp->t_pipeack = 0;
146 tcp_clear_pipeack_state(tp);
147
148 /* Start counting bytes for RFC 3465 again */
149 tp->t_bytes_acked = 0;
150
151 /*
152 * slow start threshold could get initialized to a lower value
153 * when there is a cached value in the route metrics. In this case,
154 * the connection can enter congestion avoidance without any packet
155 * loss and Cubic will enter steady-state too early. It is better
156 * to always probe to find the initial slow-start threshold.
157 */
158 if (tp->t_inpcb->inp_stat->txbytes <= tcp_initial_cwnd(tp) &&
159 tp->snd_ssthresh < (TCP_MAXWIN << TCP_MAX_WINSHIFT)) {
160 tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT;
161 }
162
163 /* Initialize cubic last max to be same as ssthresh */
164 tp->t_ccstate->cub_last_max = tp->snd_ssthresh;
165 }
166
167 /*
168 * Compute the target congestion window for the next RTT according to
169 * cubic equation when an ack is received.
170 *
171 * W(t) = C(t-K)^3 + W(last_max)
172 */
173 static uint32_t
tcp_cubic_update(struct tcpcb * tp,u_int32_t rtt)174 tcp_cubic_update(struct tcpcb *tp, u_int32_t rtt)
175 {
176 float K, var;
177 u_int32_t elapsed_time, win;
178
179 win = min(tp->snd_cwnd, tp->snd_wnd);
180 if (tp->t_ccstate->cub_last_max == 0) {
181 tp->t_ccstate->cub_last_max = tp->snd_ssthresh;
182 }
183
184 if (tp->t_ccstate->cub_epoch_start == 0) {
185 /*
186 * This is the beginning of a new epoch, initialize some of
187 * the variables that we need to use for computing the
188 * congestion window later.
189 */
190 tp->t_ccstate->cub_epoch_start = tcp_now;
191 if (tp->t_ccstate->cub_epoch_start == 0) {
192 tp->t_ccstate->cub_epoch_start = 1;
193 }
194 if (win < tp->t_ccstate->cub_last_max) {
195 /*
196 * Compute cubic epoch period, this is the time
197 * period that the window will take to increase to
198 * last_max again after backoff due to loss.
199 */
200 if (tcp_cubic_minor_fixes) {
201 K = ((float)tp->t_ccstate->cub_last_max - win) / tp->t_maxseg / tcp_cubic_coeff;
202 } else {
203 K = (tp->t_ccstate->cub_last_max - win) / tp->t_maxseg / tcp_cubic_coeff;
204 }
205 K = cbrtf(K);
206 tp->t_ccstate->cub_epoch_period = K * TCP_RETRANSHZ;
207 /* Origin point */
208 tp->t_ccstate->cub_origin_point = tp->t_ccstate->cub_last_max;
209 } else {
210 tp->t_ccstate->cub_epoch_period = 0;
211 tp->t_ccstate->cub_origin_point = win;
212 }
213 }
214
215 VERIFY(tp->t_ccstate->cub_origin_point > 0);
216 /*
217 * Compute the target window for the next RTT using smoothed RTT
218 * as an estimate for next RTT.
219 */
220 elapsed_time = timer_diff(tcp_now, 0, tp->t_ccstate->cub_epoch_start, 0);
221
222 if (tcp_cubic_use_minrtt) {
223 elapsed_time += max(tcp_cubic_use_minrtt, rtt);
224 } else {
225 elapsed_time += rtt;
226 }
227 var = (elapsed_time - tp->t_ccstate->cub_epoch_period) / TCP_RETRANSHZ;
228 var = var * var * var * (tcp_cubic_coeff * tp->t_maxseg);
229
230 return (u_int32_t)(tp->t_ccstate->cub_origin_point + var);
231 }
232
233 /*
234 * Standard TCP utilizes bandwidth well in low RTT and low BDP connections
235 * even when there is some packet loss. Enabling TCP mode will help Cubic
236 * to achieve this kind of utilization.
237 *
238 * But if there is a bottleneck link in the path with a fixed size queue
239 * and fixed bandwidth, TCP Cubic will help to reduce packet loss at this
240 * link because of the steady-state behavior. Using average and mean
241 * absolute deviation of W(lastmax), we try to detect if the congestion
242 * window is close to the bottleneck bandwidth. In that case, disabling
243 * TCP mode will help to minimize packet loss at this link.
244 *
245 * Disable TCP mode if the W(lastmax) (the window where previous packet
246 * loss happened) is within a small range from the average last max
247 * calculated.
248 */
249 #define TCP_CUBIC_ENABLE_TCPMODE(_tp_) \
250 ((!soissrcrealtime((_tp_)->t_inpcb->inp_socket) && \
251 (_tp_)->t_ccstate->cub_mean_dev > (tp->t_maxseg << 1)) ? 1 : 0)
252
253 /*
254 * Compute the window growth if standard TCP (AIMD) was used with
255 * a backoff of 0.5 and additive increase of 1 packet per RTT.
256 *
257 * TCP window at time t can be calculated using the following equation
258 * with tcp_beta_cubic
259 *
260 * W(t) <- Wmax * tcp_beta_cubic + 3 * ((1 - tcp_beta_cubic)/(1 + tcp_beta_cubic)) * t/RTT
261 *
262 */
263 static uint32_t
tcp_cubic_tcpwin(struct tcpcb * tp,struct tcphdr * th)264 tcp_cubic_tcpwin(struct tcpcb *tp, struct tcphdr *th)
265 {
266 if (tp->t_ccstate->cub_tcp_win == 0) {
267 /* Start of the epoch, we set the tcp_win to whatever Cubic decided
268 * at the beginning of the epoch.
269 */
270 tp->t_ccstate->cub_tcp_win = min(tp->snd_cwnd, tp->snd_wnd);
271 if (tcp_cubic_minor_fixes) {
272 tp->t_ccstate->cub_tcp_bytes_acked = BYTES_ACKED(th, tp);
273 } else {
274 tp->t_ccstate->cub_tcp_bytes_acked = 0;
275 }
276 } else {
277 tp->t_ccstate->cub_tcp_bytes_acked += BYTES_ACKED(th, tp);
278
279 if (tcp_cubic_minor_fixes) {
280 /*
281 * Increase by ai_factor * MSS, once per RTT. Counting bytes_acked
282 * against the snd_cwnd represents exactly one RTT at full rate.
283 */
284 while (tp->t_ccstate->cub_tcp_bytes_acked >= tp->snd_cwnd) {
285 /* Enough bytes have been ACK'd for TCP to do AIMD*/
286 tp->t_ccstate->cub_tcp_bytes_acked -= tp->snd_cwnd;
287
288 if (tp->snd_cwnd >= tp->t_ccstate->cub_last_max || !tcp_cubic_rfc_compliant) {
289 tp->t_ccstate->cub_tcp_win += tp->t_maxseg;
290 } else {
291 /* Increase-rate from Section 4.2, RFC 8312 */
292 float ai_factor = (float)3 * (1 - tcp_cubic_beta) / (1 + tcp_cubic_beta);
293
294 tp->t_ccstate->cub_tcp_win += (uint32_t)(tp->t_maxseg * ai_factor);
295 }
296 }
297 } else {
298 if (tp->t_ccstate->cub_tcp_bytes_acked >= tp->t_ccstate->cub_tcp_win) {
299 tp->t_ccstate->cub_tcp_bytes_acked -= tp->t_ccstate->cub_tcp_win;
300 tp->t_ccstate->cub_tcp_win += tp->t_maxseg;
301 }
302 }
303 }
304 return tp->t_ccstate->cub_tcp_win;
305 }
306
307 /*
308 * Handle an in-sequence ack during congestion avoidance phase.
309 */
310 static void
tcp_cubic_congestion_avd(struct tcpcb * tp,struct tcphdr * th)311 tcp_cubic_congestion_avd(struct tcpcb *tp, struct tcphdr *th)
312 {
313 u_int32_t cubic_target_win, tcp_win, rtt;
314 u_int64_t incr_win = UINT32_MAX;
315
316 /* Do not increase congestion window in non-validated phase */
317 if (tcp_cc_is_cwnd_nonvalidated(tp) != 0) {
318 return;
319 }
320
321 tp->t_bytes_acked += BYTES_ACKED(th, tp);
322
323 rtt = get_base_rtt(tp);
324 /*
325 * First compute cubic window. If cubic variables are not
326 * initialized (after coming out of recovery), this call will
327 * initialize them.
328 */
329 cubic_target_win = tcp_cubic_update(tp, rtt);
330
331 /* Compute TCP window if a multiplicative decrease of 0.2 is used */
332 tcp_win = tcp_cubic_tcpwin(tp, th);
333
334 if (tp->snd_cwnd < tcp_win && tcp_cubic_minor_fixes == 0 && TCP_CUBIC_ENABLE_TCPMODE(tp)) {
335 /* this connection is in TCP-friendly region */
336 if (tp->t_bytes_acked >= tp->snd_cwnd) {
337 tp->t_bytes_acked -= tp->snd_cwnd;
338 tp->snd_cwnd = min(tcp_win, TCP_MAXWIN << tp->snd_scale);
339 }
340 } else {
341 if (cubic_target_win > tp->snd_cwnd) {
342 /*
343 * The target win is computed for the next RTT.
344 * To reach this value, cwnd will have to be updated
345 * one segment at a time. Compute how many bytes
346 * need to be acknowledged before we can increase
347 * the cwnd by one segment.
348 */
349 incr_win = (uint64_t)tp->snd_cwnd * tp->t_maxseg;
350 incr_win /= (cubic_target_win - tp->snd_cwnd);
351 if (!tcp_cubic_minor_fixes) {
352 if (incr_win > 0 &&
353 tp->t_bytes_acked >= incr_win) {
354 tp->t_bytes_acked -= incr_win;
355 tp->snd_cwnd =
356 min((tp->snd_cwnd + tp->t_maxseg),
357 TCP_MAXWIN << tp->snd_scale);
358 }
359 }
360 }
361 }
362
363 if (tcp_cubic_minor_fixes) {
364 tcp_win = tcp_round_to(tcp_win, tp->t_maxseg);
365
366 if (tp->snd_cwnd < tcp_win) {
367 uint64_t tcp_incr_win;
368
369 tcp_incr_win = (uint64_t)tp->snd_cwnd * tp->t_maxseg;
370 tcp_incr_win /= (tcp_win - tp->snd_cwnd);
371
372 if (tcp_incr_win < incr_win) {
373 /* this connection is in TCP-friendly region */
374 incr_win = tcp_incr_win;
375 }
376 }
377
378 if (incr_win > 0 && tp->t_bytes_acked >= incr_win) {
379 tp->t_bytes_acked -= incr_win;
380 tp->snd_cwnd = min(tp->snd_cwnd + tp->t_maxseg, TCP_MAXWIN << tp->snd_scale);
381 }
382 }
383 }
384
385 static void
tcp_cubic_ack_rcvd(struct tcpcb * tp,struct tcphdr * th)386 tcp_cubic_ack_rcvd(struct tcpcb *tp, struct tcphdr *th)
387 {
388 /* Do not increase the congestion window in non-validated phase */
389 if (tcp_cc_is_cwnd_nonvalidated(tp) != 0) {
390 return;
391 }
392
393 if (tp->snd_cwnd >= tp->snd_ssthresh) {
394 /* Congestion avoidance phase */
395 tcp_cubic_congestion_avd(tp, th);
396 } else {
397 /*
398 * Use 2*SMSS as limit on increment as suggested
399 * by RFC 3465 section 2.3
400 */
401 uint32_t acked, abc_lim, incr;
402
403 acked = BYTES_ACKED(th, tp);
404 if (tcp_cubic_minor_fixes) {
405 /*
406 * Maximum burst-size is limited to the initial congestion-window.
407 * We know that the network can survive this kind of burst.
408 */
409 abc_lim = tcp_initial_cwnd(tp);
410 } else {
411 abc_lim = (tp->snd_nxt == tp->snd_max) ? 2 * tp->t_maxseg : tp->t_maxseg;
412 }
413 incr = min(acked, abc_lim);
414
415 tp->snd_cwnd += incr;
416 tp->snd_cwnd = min(tp->snd_cwnd, TCP_MAXWIN << tp->snd_scale);
417 }
418 }
419
420 static void
tcp_cubic_pre_fr(struct tcpcb * tp)421 tcp_cubic_pre_fr(struct tcpcb *tp)
422 {
423 u_int32_t win, avg;
424 int32_t dev;
425 tp->t_ccstate->cub_epoch_start = 0;
426 tp->t_ccstate->cub_tcp_win = 0;
427 tp->t_ccstate->cub_tcp_bytes_acked = 0;
428
429 win = min(tp->snd_cwnd, tp->snd_wnd);
430 if (tp->t_flagsext & TF_CWND_NONVALIDATED) {
431 tp->t_lossflightsize = tp->snd_max - tp->snd_una;
432 if (tcp_flow_control_response) {
433 win = max(tp->t_pipeack, tp->t_lossflightsize);
434 } else {
435 win = (max(tp->t_pipeack, tp->t_lossflightsize)) >> 1;
436 }
437 } else {
438 tp->t_lossflightsize = 0;
439 }
440 /*
441 * Note the congestion window at which packet loss occurred as
442 * cub_last_max.
443 *
444 * If the congestion window is less than the last max window when
445 * loss occurred, it indicates that capacity available in the
446 * network has gone down. This can happen if a new flow has started
447 * and it is capturing some of the bandwidth. To reach convergence
448 * quickly, backoff a little more.
449 */
450 if (win < tp->t_ccstate->cub_last_max && tcp_cubic_minor_fixes) {
451 tp->t_ccstate->cub_last_max = (uint32_t)((float)win * tcp_cubic_fast_convergence_factor);
452 } else {
453 tp->t_ccstate->cub_last_max = win;
454 }
455
456 if (tp->t_ccstate->cub_last_max == 0) {
457 /*
458 * If last_max is zero because snd_wnd is zero or for
459 * any other reason, initialize it to the amount of data
460 * in flight
461 */
462 tp->t_ccstate->cub_last_max = tp->snd_max - tp->snd_una;
463 }
464
465 /*
466 * Compute average and mean absolute deviation of the
467 * window at which packet loss occurred.
468 */
469 if (tp->t_ccstate->cub_avg_lastmax == 0) {
470 tp->t_ccstate->cub_avg_lastmax = tp->t_ccstate->cub_last_max;
471 } else {
472 /*
473 * Average is computed by taking 63 parts of
474 * history and one part of the most recent value
475 */
476 avg = tp->t_ccstate->cub_avg_lastmax;
477 avg = (avg << 6) - avg;
478 tp->t_ccstate->cub_avg_lastmax =
479 (avg + tp->t_ccstate->cub_last_max) >> 6;
480 }
481
482 /* caluclate deviation from average */
483 dev = tp->t_ccstate->cub_avg_lastmax - tp->t_ccstate->cub_last_max;
484
485 /* Take the absolute value */
486 if (dev < 0) {
487 dev = -dev;
488 }
489
490 if (tp->t_ccstate->cub_mean_dev == 0) {
491 tp->t_ccstate->cub_mean_dev = dev;
492 } else {
493 dev = dev + ((tp->t_ccstate->cub_mean_dev << 4)
494 - tp->t_ccstate->cub_mean_dev);
495 tp->t_ccstate->cub_mean_dev = dev >> 4;
496 }
497
498 /* Backoff congestion window by tcp_cubic_backoff factor */
499 win = (u_int32_t)(win - (win * tcp_cubic_backoff));
500 win = tcp_round_to(win, tp->t_maxseg);
501 if (win < 2 * tp->t_maxseg) {
502 win = 2 * tp->t_maxseg;
503 }
504 tp->snd_ssthresh = win;
505 tcp_cc_resize_sndbuf(tp);
506 }
507
508 static void
tcp_cubic_post_fr(struct tcpcb * tp,struct tcphdr * th)509 tcp_cubic_post_fr(struct tcpcb *tp, struct tcphdr *th)
510 {
511 uint32_t flight_size = 0;
512 uint32_t ack;
513
514 if (th != NULL) {
515 ack = th->th_ack;
516 } else {
517 ack = tp->snd_una;
518 }
519
520 if (SEQ_LEQ(ack, tp->snd_max) && (!tcp_cubic_minor_fixes || tcp_flow_control_response)) {
521 flight_size = tp->snd_max - ack;
522 } else if (tcp_cubic_minor_fixes) {
523 /*
524 * Cubic Minor Fixes: snd_max - th_ack is a very very bad estimate
525 * of the flight size. Either the app is sending at full speed and
526 * flight_size *is* snd_sshtresh, or the app is not sending at full
527 * speed and congestion-window validation would have kicked in earlier.
528 *
529 * Except that for the latter, snd_ssthresh is way too high.
530 * When we exit recovery we will burst a lot of data out...
531 *
532 * So, tcp_flow_control_response brings us back to the old behavior.
533 * Too many feature-flags...
534 */
535 flight_size = tp->snd_ssthresh;
536 }
537
538 /*
539 * Cubic Minor Fixes: t_lossflightsize is always 0, because of
540 * EXIT_FASTRECOVERY. This here is basically dead code...
541 */
542 if (SACK_ENABLED(tp) && tp->t_lossflightsize > 0 && !tcp_cubic_minor_fixes) {
543 u_int32_t total_rxt_size = 0, ncwnd;
544 /*
545 * When SACK is enabled, the number of retransmitted bytes
546 * can be counted more accurately.
547 */
548 total_rxt_size = tcp_rxtseg_total_size(tp);
549 ncwnd = max(tp->t_pipeack, tp->t_lossflightsize);
550 if (total_rxt_size <= ncwnd) {
551 ncwnd = ncwnd - total_rxt_size;
552 }
553
554 /*
555 * To avoid sending a large burst at the end of recovery
556 * set a max limit on ncwnd
557 */
558 ncwnd = min(ncwnd, (tp->t_maxseg << 6));
559 ncwnd = ncwnd >> 1;
560 flight_size = max(ncwnd, flight_size);
561 }
562 /*
563 * Complete ack. The current window was inflated for fast recovery.
564 * It has to be deflated post recovery.
565 *
566 * Window inflation should have left us with approx snd_ssthresh
567 * outstanding data. If the flight size is zero or one segment,
568 * make congestion window to be at least as big as 2 segments to
569 * avoid delayed acknowledgements. This is according to RFC 6582.
570 */
571 if (flight_size < tp->snd_ssthresh) {
572 tp->snd_cwnd = max(flight_size, tp->t_maxseg)
573 + tp->t_maxseg;
574 } else {
575 tp->snd_cwnd = tp->snd_ssthresh;
576 }
577 tp->t_ccstate->cub_tcp_win = 0;
578 tp->t_ccstate->cub_tcp_bytes_acked = 0;
579 }
580
581 static void
tcp_cubic_after_timeout(struct tcpcb * tp)582 tcp_cubic_after_timeout(struct tcpcb *tp)
583 {
584 VERIFY(tp->t_ccstate != NULL);
585
586 /*
587 * Avoid adjusting congestion window due to SYN retransmissions.
588 * If more than one byte (SYN) is outstanding then it is still
589 * needed to adjust the window.
590 */
591 if (tp->t_state < TCPS_ESTABLISHED &&
592 ((int)(tp->snd_max - tp->snd_una) <= 1)) {
593 return;
594 }
595
596 if (!IN_FASTRECOVERY(tp)) {
597 tcp_cubic_clear_state(tp);
598 tcp_cubic_pre_fr(tp);
599 }
600
601 /*
602 * Close the congestion window down to one segment as a retransmit
603 * timeout might indicate severe congestion.
604 */
605 tp->snd_cwnd = tp->t_maxseg;
606 }
607
608 static int
tcp_cubic_delay_ack(struct tcpcb * tp,struct tcphdr * th)609 tcp_cubic_delay_ack(struct tcpcb *tp, struct tcphdr *th)
610 {
611 return tcp_cc_delay_ack(tp, th);
612 }
613
614 /*
615 * When switching from a different CC it is better for Cubic to start
616 * fresh. The state required for Cubic calculation might be stale and it
617 * might not represent the current state of the network. If it starts as
618 * a new connection it will probe and learn the existing network conditions.
619 */
620 static void
tcp_cubic_switch_cc(struct tcpcb * tp)621 tcp_cubic_switch_cc(struct tcpcb *tp)
622 {
623 tcp_cubic_cwnd_init_or_reset(tp);
624
625 OSIncrementAtomic((volatile SInt32 *)&tcp_cc_cubic.num_sockets);
626 }
627
628 static inline void
tcp_cubic_clear_state(struct tcpcb * tp)629 tcp_cubic_clear_state(struct tcpcb *tp)
630 {
631 tp->t_ccstate->cub_last_max = 0;
632 tp->t_ccstate->cub_epoch_start = 0;
633 tp->t_ccstate->cub_origin_point = 0;
634 tp->t_ccstate->cub_tcp_win = 0;
635 tp->t_ccstate->cub_tcp_bytes_acked = 0;
636 tp->t_ccstate->cub_epoch_period = 0;
637 }
638