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
76static void	cubic_ack_received(struct cc_var *ccv, ccsignal_t type);
77static void	cubic_cb_destroy(struct cc_var *ccv);
78static int	cubic_cb_init(struct cc_var *ccv, void *ptr);
79static void	cubic_cong_signal(struct cc_var *ccv, ccsignal_t type);
80static void	cubic_conn_init(struct cc_var *ccv);
81static int	cubic_mod_init(void);
82static void	cubic_post_recovery(struct cc_var *ccv);
83static void	cubic_record_rtt(struct cc_var *ccv);
84static void	cubic_ssthresh_update(struct cc_var *ccv, uint32_t maxseg);
85static void	cubic_after_idle(struct cc_var *ccv);
86static size_t	cubic_data_sz(void);
87static void	cubic_newround(struct cc_var *ccv, uint32_t round_cnt);
88static void	cubic_rttsample(struct cc_var *ccv, uint32_t usec_rtt,
89       uint32_t rxtcnt, uint32_t fas);
90
91struct 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
106static void
107cubic_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->ccvc.tcp;
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
157static void
158cubic_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	u_int incr = CCV(ccv, t_maxseg);
172	uint16_t abc_val;
173
174	cubicd->flags |= CUBICFLAG_IN_SLOWSTART;
175	if (ccv->flags & CCF_USE_LOCAL_ABC)
176		abc_val = ccv->labc;
177	else
178		abc_val = V_tcp_abc_l_var;
179	if ((ccv->flags & CCF_HYSTART_ALLOWED) &&
180	    (cubicd->flags & CUBICFLAG_HYSTART_ENABLED) &&
181	    ((cubicd->flags & CUBICFLAG_HYSTART_IN_CSS) == 0)) {
182		/*
183		 * Hystart is allowed and still enabled and we are not yet
184		 * in CSS. Lets check to see if we can make a decision on
185		 * if we need to go into CSS.
186		 */
187		if ((cubicd->css_rttsample_count >= hystart_n_rttsamples) &&
188		    (cubicd->css_current_round_minrtt != 0xffffffff) &&
189		    (cubicd->css_lastround_minrtt != 0xffffffff)) {
190			uint32_t rtt_thresh;
191
192			/* Clamp (minrtt_thresh, lastround/8, maxrtt_thresh) */
193			rtt_thresh = (cubicd->css_lastround_minrtt >> 3);
194			if (rtt_thresh < hystart_minrtt_thresh)
195				rtt_thresh = hystart_minrtt_thresh;
196			if (rtt_thresh > hystart_maxrtt_thresh)
197				rtt_thresh = hystart_maxrtt_thresh;
198			cubic_log_hystart_event(ccv, cubicd, 1, rtt_thresh);
199
200			if (cubicd->css_current_round_minrtt >= (cubicd->css_lastround_minrtt + rtt_thresh)) {
201				/* Enter CSS */
202				cubicd->flags |= CUBICFLAG_HYSTART_IN_CSS;
203				cubicd->css_fas_at_css_entry = cubicd->css_lowrtt_fas;
204				/*
205				 * The draft (v4) calls for us to set baseline to css_current_round_min
206				 * but that can cause an oscillation. We probably shoudl be using
207				 * css_lastround_minrtt, but the authors insist that will cause
208				 * issues on exiting early. We will leave the draft version for now
209				 * but I suspect this is incorrect.
210				 */
211				cubicd->css_baseline_minrtt = cubicd->css_current_round_minrtt;
212				cubicd->css_entered_at_round = cubicd->css_current_round;
213				cubic_log_hystart_event(ccv, cubicd, 2, rtt_thresh);
214			}
215		}
216	}
217	if (CCV(ccv, snd_nxt) == CCV(ccv, snd_max))
218		incr = min(ccv->bytes_this_ack,
219			   ccv->nsegs * abc_val *
220			   CCV(ccv, t_maxseg));
221	else
222		incr = min(ccv->bytes_this_ack, CCV(ccv, t_maxseg));
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
235static void
236cubic_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
242	cubic_data = ccv->cc_data;
243	cubic_record_rtt(ccv);
244
245	/*
246	 * For a regular ACK and we're not in cong/fast recovery and
247	 * we're cwnd limited, always recalculate cwnd.
248	 */
249	if (type == CC_ACK && !IN_RECOVERY(CCV(ccv, t_flags)) &&
250	    (ccv->flags & CCF_CWND_LIMITED)) {
251		 /* Use the logic in NewReno ack_received() for slow start. */
252		if (CCV(ccv, snd_cwnd) <= CCV(ccv, snd_ssthresh) ||
253		    cubic_data->min_rtt_usecs == TCPTV_SRTTBASE) {
254			cubic_does_slow_start(ccv, cubic_data);
255		} else {
256			if (cubic_data->flags & CUBICFLAG_HYSTART_IN_CSS) {
257				/*
258				 * We have slipped into CA with
259				 * CSS active. Deactivate all.
260				 */
261				/* Turn off the CSS flag */
262				cubic_data->flags &= ~CUBICFLAG_HYSTART_IN_CSS;
263				/* Disable use of CSS in the future except long idle  */
264				cubic_data->flags &= ~CUBICFLAG_HYSTART_ENABLED;
265				cubic_log_hystart_event(ccv, cubic_data, 11, CCV(ccv, snd_ssthresh));
266			}
267			if ((cubic_data->flags & CUBICFLAG_RTO_EVENT) &&
268			    (cubic_data->flags & CUBICFLAG_IN_SLOWSTART)) {
269				/* RFC8312 Section 4.7 */
270				cubic_data->flags &= ~(CUBICFLAG_RTO_EVENT |
271						       CUBICFLAG_IN_SLOWSTART);
272				cubic_data->W_max = CCV(ccv, snd_cwnd);
273				cubic_data->t_epoch = ticks;
274				cubic_data->K = 0;
275			} else if (cubic_data->flags & (CUBICFLAG_IN_SLOWSTART |
276						 CUBICFLAG_IN_APPLIMIT)) {
277				cubic_data->flags &= ~(CUBICFLAG_IN_SLOWSTART |
278						       CUBICFLAG_IN_APPLIMIT);
279				cubic_data->t_epoch = ticks;
280				cubic_data->K = cubic_k(cubic_data->W_max /
281							CCV(ccv, t_maxseg));
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			 * The mean RTT is used to best reflect the equations in
293			 * the I-D. Using min_rtt in the tf_cwnd calculation
294			 * causes W_est to grow much faster than it should if the
295			 * RTT is dominated by network buffering rather than
296			 * propagation delay.
297			 */
298			W_est = tf_cwnd(usecs_since_epoch, cubic_data->mean_rtt_usecs,
299				       cubic_data->W_max, CCV(ccv, t_maxseg));
300
301			W_cubic = cubic_cwnd(usecs_since_epoch +
302					     cubic_data->mean_rtt_usecs,
303					     cubic_data->W_max,
304					     CCV(ccv, t_maxseg),
305					     cubic_data->K);
306
307			ccv->flags &= ~CCF_ABC_SENTAWND;
308
309			if (W_cubic < W_est) {
310				/*
311				 * TCP-friendly region, follow tf
312				 * cwnd growth.
313				 */
314				if (CCV(ccv, snd_cwnd) < W_est)
315					CCV(ccv, snd_cwnd) = ulmin(W_est, INT_MAX);
316			} else if (CCV(ccv, snd_cwnd) < W_cubic) {
317				/*
318				 * Concave or convex region, follow CUBIC
319				 * cwnd growth.
320				 * Only update snd_cwnd, if it doesn't shrink.
321				 */
322				CCV(ccv, snd_cwnd) = ulmin(W_cubic, INT_MAX);
323			}
324
325			/*
326			 * If we're not in slow start and we're probing for a
327			 * new cwnd limit at the start of a connection
328			 * (happens when hostcache has a relevant entry),
329			 * keep updating our current estimate of the
330			 * W_max.
331			 */
332			if (((cubic_data->flags & CUBICFLAG_CONG_EVENT) == 0) &&
333			    cubic_data->W_max < CCV(ccv, snd_cwnd)) {
334				cubic_data->W_max = CCV(ccv, snd_cwnd);
335				cubic_data->K = cubic_k(cubic_data->W_max /
336				    CCV(ccv, t_maxseg));
337			}
338		}
339	} else if (type == CC_ACK && !IN_RECOVERY(CCV(ccv, t_flags)) &&
340	    !(ccv->flags & CCF_CWND_LIMITED)) {
341		cubic_data->flags |= CUBICFLAG_IN_APPLIMIT;
342	}
343}
344
345/*
346 * This is a CUBIC specific implementation of after_idle.
347 *   - Reset cwnd by calling New Reno implementation of after_idle.
348 *   - Reset t_epoch.
349 */
350static void
351cubic_after_idle(struct cc_var *ccv)
352{
353	struct cubic *cubic_data;
354
355	cubic_data = ccv->cc_data;
356
357	cubic_data->W_max = ulmax(cubic_data->W_max, CCV(ccv, snd_cwnd));
358	cubic_data->K = cubic_k(cubic_data->W_max / CCV(ccv, t_maxseg));
359	if ((cubic_data->flags & CUBICFLAG_HYSTART_ENABLED) == 0) {
360		/*
361		 * Re-enable hystart if we have been idle.
362		 */
363		cubic_data->flags &= ~CUBICFLAG_HYSTART_IN_CSS;
364		cubic_data->flags |= CUBICFLAG_HYSTART_ENABLED;
365		cubic_log_hystart_event(ccv, cubic_data, 12, CCV(ccv, snd_ssthresh));
366	}
367	newreno_cc_after_idle(ccv);
368	cubic_data->t_epoch = ticks;
369}
370
371static void
372cubic_cb_destroy(struct cc_var *ccv)
373{
374	free(ccv->cc_data, M_CC_MEM);
375}
376
377static size_t
378cubic_data_sz(void)
379{
380	return (sizeof(struct cubic));
381}
382
383static int
384cubic_cb_init(struct cc_var *ccv, void *ptr)
385{
386	struct cubic *cubic_data;
387
388	INP_WLOCK_ASSERT(tptoinpcb(ccv->ccvc.tcp));
389	if (ptr == NULL) {
390		cubic_data = malloc(sizeof(struct cubic), M_CC_MEM, M_NOWAIT|M_ZERO);
391		if (cubic_data == NULL)
392			return (ENOMEM);
393	} else
394		cubic_data = ptr;
395
396	/* Init some key variables with sensible defaults. */
397	cubic_data->t_epoch = ticks;
398	cubic_data->min_rtt_usecs = TCPTV_SRTTBASE;
399	cubic_data->mean_rtt_usecs = 1;
400
401	ccv->cc_data = cubic_data;
402	cubic_data->flags = CUBICFLAG_HYSTART_ENABLED;
403	/* At init set both to infinity */
404	cubic_data->css_lastround_minrtt = 0xffffffff;
405	cubic_data->css_current_round_minrtt = 0xffffffff;
406	cubic_data->css_current_round = 0;
407	cubic_data->css_baseline_minrtt = 0xffffffff;
408	cubic_data->css_rttsample_count = 0;
409	cubic_data->css_entered_at_round = 0;
410	cubic_data->css_fas_at_css_entry = 0;
411	cubic_data->css_lowrtt_fas = 0;
412	cubic_data->css_last_fas = 0;
413
414	return (0);
415}
416
417/*
418 * Perform any necessary tasks before we enter congestion recovery.
419 */
420static void
421cubic_cong_signal(struct cc_var *ccv, ccsignal_t type)
422{
423	struct cubic *cubic_data;
424	uint32_t mss, pipe;
425
426	cubic_data = ccv->cc_data;
427	mss = tcp_fixed_maxseg(ccv->ccvc.tcp);
428
429	switch (type) {
430	case CC_NDUPACK:
431		if (cubic_data->flags & CUBICFLAG_HYSTART_ENABLED) {
432			/* Make sure the flags are all off we had a loss */
433			cubic_data->flags &= ~CUBICFLAG_HYSTART_ENABLED;
434			cubic_data->flags &= ~CUBICFLAG_HYSTART_IN_CSS;
435			cubic_log_hystart_event(ccv, cubic_data, 10, CCV(ccv, snd_ssthresh));
436		}
437		if (!IN_FASTRECOVERY(CCV(ccv, t_flags))) {
438			if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) {
439				cubic_ssthresh_update(ccv, mss);
440				cubic_data->flags |= CUBICFLAG_CONG_EVENT;
441				cubic_data->t_epoch = ticks;
442				cubic_data->K = cubic_k(cubic_data->W_max / mss);
443			}
444			ENTER_RECOVERY(CCV(ccv, t_flags));
445		}
446		break;
447
448	case CC_ECN:
449		if (cubic_data->flags & CUBICFLAG_HYSTART_ENABLED) {
450			/* Make sure the flags are all off we had a loss */
451			cubic_data->flags &= ~CUBICFLAG_HYSTART_ENABLED;
452			cubic_data->flags &= ~CUBICFLAG_HYSTART_IN_CSS;
453			cubic_log_hystart_event(ccv, cubic_data, 9, CCV(ccv, snd_ssthresh));
454		}
455		if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) {
456			cubic_ssthresh_update(ccv, mss);
457			cubic_data->flags |= CUBICFLAG_CONG_EVENT;
458			cubic_data->t_epoch = ticks;
459			cubic_data->K = cubic_k(cubic_data->W_max / mss);
460			CCV(ccv, snd_cwnd) = CCV(ccv, snd_ssthresh);
461			ENTER_CONGRECOVERY(CCV(ccv, t_flags));
462		}
463		break;
464
465	case CC_RTO:
466		/* RFC8312 Section 4.7 */
467		if (CCV(ccv, t_rxtshift) == 1) {
468			/*
469			 * Remember the state only for the first RTO event. This
470			 * will help us restore the state to the values seen
471			 * at the most recent congestion avoidance stage before
472			 * the current RTO event.
473			 */
474			cubic_data->undo_t_epoch = cubic_data->t_epoch;
475			cubic_data->undo_cwnd_epoch = cubic_data->cwnd_epoch;
476			cubic_data->undo_W_est = cubic_data->W_est;
477			cubic_data->undo_cwnd_prior = cubic_data->cwnd_prior;
478			cubic_data->undo_W_max = cubic_data->W_max;
479			cubic_data->undo_K = cubic_data->K;
480			if (V_tcp_do_newsack) {
481				pipe = tcp_compute_pipe(ccv->ccvc.tcp);
482			} else {
483				pipe = CCV(ccv, snd_max) -
484					CCV(ccv, snd_fack) +
485					CCV(ccv, sackhint.sack_bytes_rexmit);
486			}
487			CCV(ccv, snd_ssthresh) = max(2,
488				(((uint64_t)min(CCV(ccv, snd_wnd), pipe) *
489				CUBIC_BETA) >> CUBIC_SHIFT) / mss) * mss;
490		}
491		cubic_data->flags |= CUBICFLAG_CONG_EVENT | CUBICFLAG_RTO_EVENT;
492		cubic_data->undo_W_max = cubic_data->W_max;
493		cubic_data->num_cong_events++;
494		CCV(ccv, snd_cwnd) = mss;
495		break;
496
497	case CC_RTO_ERR:
498		cubic_data->flags &= ~(CUBICFLAG_CONG_EVENT | CUBICFLAG_RTO_EVENT);
499		cubic_data->num_cong_events--;
500		cubic_data->K = cubic_data->undo_K;
501		cubic_data->cwnd_prior = cubic_data->undo_cwnd_prior;
502		cubic_data->W_max = cubic_data->undo_W_max;
503		cubic_data->W_est = cubic_data->undo_W_est;
504		cubic_data->cwnd_epoch = cubic_data->undo_cwnd_epoch;
505		cubic_data->t_epoch = cubic_data->undo_t_epoch;
506		break;
507	default:
508		break;
509	}
510}
511
512static void
513cubic_conn_init(struct cc_var *ccv)
514{
515	struct cubic *cubic_data;
516
517	cubic_data = ccv->cc_data;
518
519	/*
520	 * Ensure we have a sane initial value for W_max recorded. Without
521	 * this here bad things happen when entries from the TCP hostcache
522	 * get used.
523	 */
524	cubic_data->W_max = CCV(ccv, snd_cwnd);
525}
526
527static int
528cubic_mod_init(void)
529{
530	return (0);
531}
532
533/*
534 * Perform any necessary tasks before we exit congestion recovery.
535 */
536static void
537cubic_post_recovery(struct cc_var *ccv)
538{
539	struct cubic *cubic_data;
540	int pipe;
541
542	cubic_data = ccv->cc_data;
543	pipe = 0;
544
545	if (IN_FASTRECOVERY(CCV(ccv, t_flags))) {
546		/*
547		 * If inflight data is less than ssthresh, set cwnd
548		 * conservatively to avoid a burst of data, as suggested in
549		 * the NewReno RFC. Otherwise, use the CUBIC method.
550		 *
551		 * XXXLAS: Find a way to do this without needing curack
552		 */
553		if (V_tcp_do_newsack)
554			pipe = tcp_compute_pipe(ccv->ccvc.tcp);
555		else
556			pipe = CCV(ccv, snd_max) - ccv->curack;
557
558		if (pipe < CCV(ccv, snd_ssthresh))
559			/*
560			 * Ensure that cwnd does not collapse to 1 MSS under
561			 * adverse conditions. Implements RFC6582
562			 */
563			CCV(ccv, snd_cwnd) = max(pipe, CCV(ccv, t_maxseg)) +
564			    CCV(ccv, t_maxseg);
565		else
566			/* Update cwnd based on beta and adjusted W_max. */
567			CCV(ccv, snd_cwnd) = max(((uint64_t)cubic_data->W_max *
568			    CUBIC_BETA) >> CUBIC_SHIFT,
569			    2 * CCV(ccv, t_maxseg));
570	}
571
572	/* Calculate the average RTT between congestion epochs. */
573	if (cubic_data->epoch_ack_count > 0 &&
574	    cubic_data->sum_rtt_usecs >= cubic_data->epoch_ack_count) {
575		cubic_data->mean_rtt_usecs = (int)(cubic_data->sum_rtt_usecs /
576		    cubic_data->epoch_ack_count);
577	}
578
579	cubic_data->epoch_ack_count = 0;
580	cubic_data->sum_rtt_usecs = 0;
581}
582
583/*
584 * Record the min RTT and sum samples for the epoch average RTT calculation.
585 */
586static void
587cubic_record_rtt(struct cc_var *ccv)
588{
589	struct cubic *cubic_data;
590	uint32_t t_srtt_usecs;
591
592	/* Ignore srtt until a min number of samples have been taken. */
593	if (CCV(ccv, t_rttupdated) >= CUBIC_MIN_RTT_SAMPLES) {
594		cubic_data = ccv->cc_data;
595		t_srtt_usecs = tcp_get_srtt(ccv->ccvc.tcp,
596					    TCP_TMR_GRANULARITY_USEC);
597		/*
598		 * Record the current SRTT as our minrtt if it's the smallest
599		 * we've seen or minrtt is currently equal to its initialised
600		 * value.
601		 *
602		 * XXXLAS: Should there be some hysteresis for minrtt?
603		 */
604		if ((t_srtt_usecs < cubic_data->min_rtt_usecs ||
605		    cubic_data->min_rtt_usecs == TCPTV_SRTTBASE)) {
606			/* A minimal rtt is a single unshifted tick of a ticks
607			 * timer. */
608			cubic_data->min_rtt_usecs = max(tick >> TCP_RTT_SHIFT,
609							t_srtt_usecs);
610
611			/*
612			 * If the connection is within its first congestion
613			 * epoch, ensure we prime mean_rtt_usecs with a
614			 * reasonable value until the epoch average RTT is
615			 * calculated in cubic_post_recovery().
616			 */
617			if (cubic_data->min_rtt_usecs >
618			    cubic_data->mean_rtt_usecs)
619				cubic_data->mean_rtt_usecs =
620				    cubic_data->min_rtt_usecs;
621		}
622
623		/* Sum samples for epoch average RTT calculation. */
624		cubic_data->sum_rtt_usecs += t_srtt_usecs;
625		cubic_data->epoch_ack_count++;
626	}
627}
628
629/*
630 * Update the ssthresh in the event of congestion.
631 */
632static void
633cubic_ssthresh_update(struct cc_var *ccv, uint32_t maxseg)
634{
635	struct cubic *cubic_data;
636	uint32_t ssthresh;
637	uint32_t cwnd;
638
639	cubic_data = ccv->cc_data;
640	cwnd = CCV(ccv, snd_cwnd);
641
642	/* Fast convergence heuristic. */
643	if (cwnd < cubic_data->W_max) {
644		cwnd = ((uint64_t)cwnd * CUBIC_FC_FACTOR) >> CUBIC_SHIFT;
645	}
646	cubic_data->undo_W_max = cubic_data->W_max;
647	cubic_data->W_max = cwnd;
648
649	/*
650	 * On the first congestion event, set ssthresh to cwnd * 0.5
651	 * and reduce W_max to cwnd * beta. This aligns the cubic concave
652	 * region appropriately. On subsequent congestion events, set
653	 * ssthresh to cwnd * beta.
654	 */
655	if ((cubic_data->flags & CUBICFLAG_CONG_EVENT) == 0) {
656		ssthresh = cwnd >> 1;
657		cubic_data->W_max = ((uint64_t)cwnd *
658		    CUBIC_BETA) >> CUBIC_SHIFT;
659	} else {
660		ssthresh = ((uint64_t)cwnd *
661		    CUBIC_BETA) >> CUBIC_SHIFT;
662	}
663	CCV(ccv, snd_ssthresh) = max(ssthresh, 2 * maxseg);
664}
665
666static void
667cubic_rttsample(struct cc_var *ccv, uint32_t usec_rtt, uint32_t rxtcnt, uint32_t fas)
668{
669	struct cubic *cubicd;
670
671	cubicd = ccv->cc_data;
672	if (rxtcnt > 1) {
673		/*
674		 * Only look at RTT's that are non-ambiguous.
675		 */
676		return;
677	}
678	cubicd->css_rttsample_count++;
679	cubicd->css_last_fas = fas;
680	if (cubicd->css_current_round_minrtt > usec_rtt) {
681		cubicd->css_current_round_minrtt = usec_rtt;
682		cubicd->css_lowrtt_fas = cubicd->css_last_fas;
683	}
684	if ((cubicd->css_rttsample_count >= hystart_n_rttsamples) &&
685	    (cubicd->css_current_round_minrtt != 0xffffffff) &&
686	    (cubicd->css_current_round_minrtt < cubicd->css_baseline_minrtt) &&
687	    (cubicd->css_lastround_minrtt != 0xffffffff)) {
688		/*
689		 * We were in CSS and the RTT is now less, we
690		 * entered CSS erroneously.
691		 */
692		cubicd->flags &= ~CUBICFLAG_HYSTART_IN_CSS;
693		cubic_log_hystart_event(ccv, cubicd, 8, cubicd->css_baseline_minrtt);
694		cubicd->css_baseline_minrtt = 0xffffffff;
695	}
696	if (cubicd->flags & CUBICFLAG_HYSTART_ENABLED)
697		cubic_log_hystart_event(ccv, cubicd, 5, usec_rtt);
698}
699
700static void
701cubic_newround(struct cc_var *ccv, uint32_t round_cnt)
702{
703	struct cubic *cubicd;
704
705	cubicd = ccv->cc_data;
706	/* We have entered a new round */
707	cubicd->css_lastround_minrtt = cubicd->css_current_round_minrtt;
708	cubicd->css_current_round_minrtt = 0xffffffff;
709	cubicd->css_rttsample_count = 0;
710	cubicd->css_current_round = round_cnt;
711	if ((cubicd->flags & CUBICFLAG_HYSTART_IN_CSS) &&
712	    ((round_cnt - cubicd->css_entered_at_round) >= hystart_css_rounds)) {
713		/* Enter CA */
714		if (ccv->flags & CCF_HYSTART_CAN_SH_CWND) {
715			/*
716			 * We engage more than snd_ssthresh, engage
717			 * the brakes!! Though we will stay in SS to
718			 * creep back up again, so lets leave CSS active
719			 * and give us hystart_css_rounds more rounds.
720			 */
721			if (ccv->flags & CCF_HYSTART_CONS_SSTH) {
722				CCV(ccv, snd_ssthresh) = ((cubicd->css_lowrtt_fas + cubicd->css_fas_at_css_entry) / 2);
723			} else {
724				CCV(ccv, snd_ssthresh) = cubicd->css_lowrtt_fas;
725			}
726			CCV(ccv, snd_cwnd) = cubicd->css_fas_at_css_entry;
727			cubicd->css_entered_at_round = round_cnt;
728		} else {
729			CCV(ccv, snd_ssthresh) = CCV(ccv, snd_cwnd);
730			/* Turn off the CSS flag */
731			cubicd->flags &= ~CUBICFLAG_HYSTART_IN_CSS;
732			/* Disable use of CSS in the future except long idle  */
733			cubicd->flags &= ~CUBICFLAG_HYSTART_ENABLED;
734		}
735		cubic_log_hystart_event(ccv, cubicd, 6, CCV(ccv, snd_ssthresh));
736	}
737	if (cubicd->flags & CUBICFLAG_HYSTART_ENABLED)
738		cubic_log_hystart_event(ccv, cubicd, 4, round_cnt);
739}
740
741DECLARE_CC_MODULE(cubic, &cubic_cc_algo);
742MODULE_VERSION(cubic, 2);
743