2 * Copyright (c) 2009-2010
3 * Swinburne University of Technology, Melbourne, Australia
4 * Copyright (c) 2010-2011 The FreeBSD Foundation
7 * This software was developed at the Centre for Advanced Internet
8 * Architectures, Swinburne University of Technology, by David Hayes and
9 * Lawrence Stewart, made possible in part by a grant from the Cisco University
10 * Research Program Fund at Community Foundation Silicon Valley.
12 * Portions of this software were developed at the Centre for Advanced Internet
13 * Architectures, Swinburne University of Technology, Melbourne, Australia by
14 * David Hayes under sponsorship from the FreeBSD Foundation.
16 * Redistribution and use in source and binary forms, with or without
17 * modification, are permitted provided that the following conditions
19 * 1. Redistributions of source code must retain the above copyright
20 * notice, this list of conditions and the following disclaimer.
21 * 2. Redistributions in binary form must reproduce the above copyright
22 * notice, this list of conditions and the following disclaimer in the
23 * documentation and/or other materials provided with the distribution.
25 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
26 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
27 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
28 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
29 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
30 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
31 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
32 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
33 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
34 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
39 * An implementation of the CAIA-Hamilton delay based congestion control
40 * algorithm, based on "Improved coexistence and loss tolerance for delay based
41 * TCP congestion control" by D. A. Hayes and G. Armitage., in 35th Annual IEEE
42 * Conference on Local Computer Networks (LCN 2010), Denver, Colorado, USA,
45 * Originally released as part of the NewTCP research project at Swinburne
46 * University of Technology's Centre for Advanced Internet Architectures,
47 * Melbourne, Australia, which was made possible in part by a grant from the
48 * Cisco University Research Program Fund at Community Foundation Silicon
49 * Valley. More details are available at:
50 * http://caia.swin.edu.au/urp/newtcp/
53 #include <sys/cdefs.h>
54 __FBSDID("$FreeBSD$");
56 #include <sys/param.h>
57 #include <sys/kernel.h>
58 #include <sys/khelp.h>
59 #include <sys/limits.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>
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>
77 #include <netinet/khelp/h_ertt.h>
79 #define CAST_PTR_INT(X) (*((int*)(X)))
82 * Private signal type for rate based congestion signal.
83 * See <netinet/cc.h> for appropriate bit-range to use for private signals.
85 #define CC_CHD_DELAY 0x02000000
87 /* Largest possible number returned by random(). */
88 #define RANDOM_MAX INT_MAX
90 static void chd_ack_received(struct cc_var *ccv, uint16_t ack_type);
91 static void chd_cb_destroy(struct cc_var *ccv);
92 static int chd_cb_init(struct cc_var *ccv);
93 static void chd_cong_signal(struct cc_var *ccv, uint32_t signal_type);
94 static void chd_conn_init(struct cc_var *ccv);
95 static int chd_mod_init(void);
99 * Shadow window - keeps track of what the NewReno congestion window
100 * would have been if delay-based cwnd backoffs had not been made. This
101 * functionality aids coexistence with loss-based TCP flows which may be
102 * sharing links along the path.
104 unsigned long shadow_w;
106 * Loss-based TCP compatibility flag - When set, it turns on the shadow
107 * window functionality.
110 /* The maximum round trip time seen within a measured rtt period. */
112 /* The previous qdly that caused cwnd to backoff. */
113 int prev_backoff_qdly;
118 static VNET_DEFINE(uint32_t, chd_qmin) = 5;
119 static VNET_DEFINE(uint32_t, chd_pmax) = 50;
120 static VNET_DEFINE(uint32_t, chd_loss_fair) = 1;
121 static VNET_DEFINE(uint32_t, chd_use_max) = 1;
122 static VNET_DEFINE(uint32_t, chd_qthresh) = 20;
123 #define V_chd_qthresh VNET(chd_qthresh)
124 #define V_chd_qmin VNET(chd_qmin)
125 #define V_chd_pmax VNET(chd_pmax)
126 #define V_chd_loss_fair VNET(chd_loss_fair)
127 #define V_chd_use_max VNET(chd_use_max)
129 static MALLOC_DEFINE(M_CHD, "chd data",
130 "Per connection data required for the CHD congestion control algorithm");
132 struct cc_algo chd_cc_algo = {
134 .ack_received = chd_ack_received,
135 .cb_destroy = chd_cb_destroy,
136 .cb_init = chd_cb_init,
137 .cong_signal = chd_cong_signal,
138 .conn_init = chd_conn_init,
139 .mod_init = chd_mod_init
143 chd_window_decrease(struct cc_var *ccv)
147 win = min(CCV(ccv, snd_wnd), CCV(ccv, snd_cwnd)) / CCV(ccv, t_maxseg);
148 win -= max((win / 2), 1);
149 CCV(ccv, snd_ssthresh) = max(win, 2) * CCV(ccv, t_maxseg);
153 * Probabilistic backoff function. Returns 1 if we should backoff or 0
154 * otherwise. The calculation of p is similar to the calculation of p in cc_hd.
157 should_backoff(int qdly, int maxqdly, struct chd *chd_data)
159 unsigned long p, rand;
163 if (qdly < V_chd_qthresh) {
164 chd_data->loss_compete = 0;
165 p = (((RANDOM_MAX / 100) * V_chd_pmax) /
166 (V_chd_qthresh - V_chd_qmin)) *
169 if (qdly > V_chd_qthresh) {
170 p = (((RANDOM_MAX / 100) * V_chd_pmax) /
171 (maxqdly - V_chd_qthresh)) *
173 if (V_chd_loss_fair && rand < p)
174 chd_data->loss_compete = 1;
176 p = (RANDOM_MAX / 100) * V_chd_pmax;
177 chd_data->loss_compete = 0;
185 chd_window_increase(struct cc_var *ccv, int new_measurement)
187 struct chd *chd_data;
190 chd_data = ccv->cc_data;
193 if (CCV(ccv, snd_cwnd) <= CCV(ccv, snd_ssthresh)) {
194 /* Adapted from NewReno slow start. */
195 if (V_tcp_do_rfc3465) {
196 /* In slow-start with ABC enabled. */
197 if (CCV(ccv, snd_nxt) == CCV(ccv, snd_max)) {
198 /* Not due to RTO. */
199 incr = min(ccv->bytes_this_ack,
200 V_tcp_abc_l_var * CCV(ccv, t_maxseg));
203 incr = min(ccv->bytes_this_ack,
207 incr = CCV(ccv, t_maxseg);
209 } else { /* Congestion avoidance. */
210 if (V_tcp_do_rfc3465) {
211 if (ccv->flags & CCF_ABC_SENTAWND) {
212 ccv->flags &= ~CCF_ABC_SENTAWND;
213 incr = CCV(ccv, t_maxseg);
215 } else if (new_measurement)
216 incr = CCV(ccv, t_maxseg);
219 if (chd_data->shadow_w > 0) {
220 /* Track NewReno window. */
221 chd_data->shadow_w = min(chd_data->shadow_w + incr,
222 TCP_MAXWIN << CCV(ccv, snd_scale));
225 CCV(ccv,snd_cwnd) = min(CCV(ccv, snd_cwnd) + incr,
226 TCP_MAXWIN << CCV(ccv, snd_scale));
230 * All ACK signals are used for timing measurements to determine delay-based
231 * congestion. However, window increases are only performed when
232 * ack_type == CC_ACK.
235 chd_ack_received(struct cc_var *ccv, uint16_t ack_type)
237 struct chd *chd_data;
239 int backoff, new_measurement, qdly, rtt;
241 e_t = khelp_get_osd(CCV(ccv, osd), ertt_id);
242 chd_data = ccv->cc_data;
243 new_measurement = e_t->flags & ERTT_NEW_MEASUREMENT;
246 chd_data->maxrtt_in_rtt = imax(e_t->rtt, chd_data->maxrtt_in_rtt);
248 if (new_measurement) {
250 * There is a new per RTT measurement, so check to see if there
251 * is delay based congestion.
253 rtt = V_chd_use_max ? chd_data->maxrtt_in_rtt : e_t->rtt;
254 chd_data->maxrtt_in_rtt = 0;
256 if (rtt && e_t->minrtt && !IN_RECOVERY(CCV(ccv, t_flags))) {
257 qdly = rtt - e_t->minrtt;
258 if (qdly > V_chd_qmin) {
260 * Probabilistic delay based congestion
263 backoff = should_backoff(qdly,
264 e_t->maxrtt - e_t->minrtt, chd_data);
266 chd_data->loss_compete = 0;
268 /* Reset per RTT measurement flag to start a new measurement. */
269 e_t->flags &= ~ERTT_NEW_MEASUREMENT;
274 * Update shadow_w before delay based backoff.
276 if (chd_data->loss_compete ||
277 qdly > chd_data->prev_backoff_qdly) {
279 * Delay is higher than when we backed off previously,
280 * so it is possible that this flow is competing with
283 chd_data->shadow_w = max(CCV(ccv, snd_cwnd),
287 * Reset shadow_w, as it is probable that this flow is
288 * not competing with loss based flows at the moment.
290 chd_data->shadow_w = 0;
293 chd_data->prev_backoff_qdly = qdly;
295 * Send delay-based congestion signal to the congestion signal
298 chd_cong_signal(ccv, CC_CHD_DELAY);
300 } else if (ack_type == CC_ACK)
301 chd_window_increase(ccv, new_measurement);
305 chd_cb_destroy(struct cc_var *ccv)
308 if (ccv->cc_data != NULL)
309 free(ccv->cc_data, M_CHD);
313 chd_cb_init(struct cc_var *ccv)
315 struct chd *chd_data;
317 chd_data = malloc(sizeof(struct chd), M_CHD, M_NOWAIT);
318 if (chd_data == NULL)
321 chd_data->shadow_w = 0;
322 ccv->cc_data = chd_data;
328 chd_cong_signal(struct cc_var *ccv, uint32_t signal_type)
331 struct chd *chd_data;
334 e_t = khelp_get_osd(CCV(ccv, osd), ertt_id);
335 chd_data = ccv->cc_data;
336 qdly = imax(e_t->rtt, chd_data->maxrtt_in_rtt) - e_t->minrtt;
338 switch(signal_type) {
340 chd_window_decrease(ccv); /* Set new ssthresh. */
341 CCV(ccv, snd_cwnd) = CCV(ccv, snd_ssthresh);
342 CCV(ccv, snd_recover) = CCV(ccv, snd_max);
343 ENTER_CONGRECOVERY(CCV(ccv, t_flags));
346 case CC_NDUPACK: /* Packet loss. */
348 * Only react to loss as a congestion signal if qdly >
349 * V_chd_qthresh. If qdly is less than qthresh, presume that
350 * this is a non congestion related loss. If qdly is greater
351 * than qthresh, assume that we are competing with loss based
352 * tcp flows and restore window from any unnecessary backoffs,
353 * before the decrease.
355 if (!IN_RECOVERY(CCV(ccv, t_flags)) && qdly > V_chd_qthresh) {
356 if (chd_data->loss_compete) {
357 CCV(ccv, snd_cwnd) = max(CCV(ccv, snd_cwnd),
360 chd_window_decrease(ccv);
363 * This loss isn't congestion related, or already
364 * recovering from congestion.
366 CCV(ccv, snd_ssthresh) = CCV(ccv, snd_cwnd);
367 CCV(ccv, snd_recover) = CCV(ccv, snd_max);
370 if (chd_data->shadow_w > 0) {
371 chd_data->shadow_w = max(chd_data->shadow_w /
372 CCV(ccv, t_maxseg) / 2, 2) * CCV(ccv, t_maxseg);
374 ENTER_FASTRECOVERY(CCV(ccv, t_flags));
378 newreno_cc_algo.cong_signal(ccv, signal_type);
383 chd_conn_init(struct cc_var *ccv)
385 struct chd *chd_data;
387 chd_data = ccv->cc_data;
388 chd_data->prev_backoff_qdly = 0;
389 chd_data->maxrtt_in_rtt = 0;
390 chd_data->loss_compete = 0;
392 * Initialise the shadow_cwnd to be equal to snd_cwnd in case we are
393 * competing with loss based flows from the start.
395 chd_data->shadow_w = CCV(ccv, snd_cwnd);
402 ertt_id = khelp_get_id("ertt");
404 printf("%s: h_ertt module not found\n", __func__);
408 chd_cc_algo.after_idle = newreno_cc_algo.after_idle;
409 chd_cc_algo.post_recovery = newreno_cc_algo.post_recovery;
415 chd_loss_fair_handler(SYSCTL_HANDLER_ARGS)
420 new = V_chd_loss_fair;
421 error = sysctl_handle_int(oidp, &new, 0, req);
422 if (error == 0 && req->newptr != NULL) {
423 if (CAST_PTR_INT(req->newptr) > 1)
426 V_chd_loss_fair = new;
433 chd_pmax_handler(SYSCTL_HANDLER_ARGS)
439 error = sysctl_handle_int(oidp, &new, 0, req);
440 if (error == 0 && req->newptr != NULL) {
441 if (CAST_PTR_INT(req->newptr) == 0 ||
442 CAST_PTR_INT(req->newptr) > 100)
452 chd_qthresh_handler(SYSCTL_HANDLER_ARGS)
458 error = sysctl_handle_int(oidp, &new, 0, req);
459 if (error == 0 && req->newptr != NULL) {
460 if (CAST_PTR_INT(req->newptr) <= V_chd_qmin)
469 SYSCTL_DECL(_net_inet_tcp_cc_chd);
470 SYSCTL_NODE(_net_inet_tcp_cc, OID_AUTO, chd, CTLFLAG_RW, NULL,
471 "CAIA Hamilton delay-based congestion control related settings");
473 SYSCTL_PROC(_net_inet_tcp_cc_chd, OID_AUTO, loss_fair,
474 CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW,
475 &VNET_NAME(chd_loss_fair), 1, &chd_loss_fair_handler,
476 "IU", "Flag to enable shadow window functionality.");
478 SYSCTL_PROC(_net_inet_tcp_cc_chd, OID_AUTO, pmax,
479 CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW,
480 &VNET_NAME(chd_pmax), 5, &chd_pmax_handler,
481 "IU", "Per RTT maximum backoff probability as a percentage");
483 SYSCTL_PROC(_net_inet_tcp_cc_chd, OID_AUTO, queue_threshold,
484 CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW,
485 &VNET_NAME(chd_qthresh), 20, &chd_qthresh_handler,
486 "IU", "Queueing congestion threshold in ticks");
488 SYSCTL_UINT(_net_inet_tcp_cc_chd, OID_AUTO, queue_min,
489 CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(chd_qmin), 5,
490 "Minimum queueing delay threshold in ticks");
492 SYSCTL_UINT(_net_inet_tcp_cc_chd, OID_AUTO, use_max,
493 CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(chd_use_max), 1,
494 "Use the maximum RTT seen within the measurement period (RTT) "
495 "as the basic delay measurement for the algorithm.");
497 DECLARE_CC_MODULE(chd, &chd_cc_algo);
498 MODULE_DEPEND(chd, ertt, 1, 1, 1);