2 * net/dccp/ccids/ccid2.c
4 * Copyright (c) 2005, 2006 Andrea Bittau <a.bittau@cs.ucl.ac.uk>
6 * Changes to meet Linux coding standards, and DCCP infrastructure fixes.
8 * Copyright (c) 2006 Arnaldo Carvalho de Melo <acme@conectiva.com.br>
10 * This program is free software; you can redistribute it and/or modify
11 * it under the terms of the GNU General Public License as published by
12 * the Free Software Foundation; either version 2 of the License, or
13 * (at your option) any later version.
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
18 * GNU General Public License for more details.
20 * You should have received a copy of the GNU General Public License
21 * along with this program; if not, write to the Free Software
22 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
26 * This implementation should follow RFC 4341
34 #ifdef CONFIG_IP_DCCP_CCID2_DEBUG
35 static int ccid2_debug;
36 #define ccid2_pr_debug(format, a...) DCCP_PR_DEBUG(ccid2_debug, format, ##a)
38 #define ccid2_pr_debug(format, a...)
41 static int ccid2_hc_tx_alloc_seq(struct ccid2_hc_tx_sock *hctx)
43 struct ccid2_seq *seqp;
46 /* check if we have space to preserve the pointer to the buffer */
47 if (hctx->seqbufc >= sizeof(hctx->seqbuf) / sizeof(struct ccid2_seq *))
50 /* allocate buffer and initialize linked list */
51 seqp = kmalloc(CCID2_SEQBUF_LEN * sizeof(struct ccid2_seq), gfp_any());
55 for (i = 0; i < (CCID2_SEQBUF_LEN - 1); i++) {
56 seqp[i].ccid2s_next = &seqp[i + 1];
57 seqp[i + 1].ccid2s_prev = &seqp[i];
59 seqp[CCID2_SEQBUF_LEN - 1].ccid2s_next = seqp;
60 seqp->ccid2s_prev = &seqp[CCID2_SEQBUF_LEN - 1];
62 /* This is the first allocation. Initiate the head and tail. */
63 if (hctx->seqbufc == 0)
64 hctx->seqh = hctx->seqt = seqp;
66 /* link the existing list with the one we just created */
67 hctx->seqh->ccid2s_next = seqp;
68 seqp->ccid2s_prev = hctx->seqh;
70 hctx->seqt->ccid2s_prev = &seqp[CCID2_SEQBUF_LEN - 1];
71 seqp[CCID2_SEQBUF_LEN - 1].ccid2s_next = hctx->seqt;
74 /* store the original pointer to the buffer so we can free it */
75 hctx->seqbuf[hctx->seqbufc] = seqp;
81 static int ccid2_hc_tx_send_packet(struct sock *sk, struct sk_buff *skb)
83 if (ccid2_cwnd_network_limited(ccid2_hc_tx_sk(sk)))
84 return CCID_PACKET_WILL_DEQUEUE_LATER;
85 return CCID_PACKET_SEND_AT_ONCE;
88 static void ccid2_change_l_ack_ratio(struct sock *sk, u32 val)
90 struct dccp_sock *dp = dccp_sk(sk);
91 u32 max_ratio = DIV_ROUND_UP(ccid2_hc_tx_sk(sk)->cwnd, 2);
94 * Ensure that Ack Ratio does not exceed ceil(cwnd/2), which is (2) from
95 * RFC 4341, 6.1.2. We ignore the statement that Ack Ratio 2 is always
96 * acceptable since this causes starvation/deadlock whenever cwnd < 2.
97 * The same problem arises when Ack Ratio is 0 (ie. Ack Ratio disabled).
99 if (val == 0 || val > max_ratio) {
100 DCCP_WARN("Limiting Ack Ratio (%u) to %u\n", val, max_ratio);
103 if (val > DCCPF_ACK_RATIO_MAX)
104 val = DCCPF_ACK_RATIO_MAX;
106 if (val == dp->dccps_l_ack_ratio)
109 ccid2_pr_debug("changing local ack ratio to %u\n", val);
110 dp->dccps_l_ack_ratio = val;
113 static void ccid2_start_rto_timer(struct sock *sk);
115 static void ccid2_hc_tx_rto_expire(unsigned long data)
117 struct sock *sk = (struct sock *)data;
118 struct ccid2_hc_tx_sock *hctx = ccid2_hc_tx_sk(sk);
119 const bool sender_was_blocked = ccid2_cwnd_network_limited(hctx);
122 if (sock_owned_by_user(sk)) {
123 sk_reset_timer(sk, &hctx->rtotimer, jiffies + HZ / 5);
127 ccid2_pr_debug("RTO_EXPIRE\n");
131 if (hctx->rto > DCCP_RTO_MAX)
132 hctx->rto = DCCP_RTO_MAX;
134 /* adjust pipe, cwnd etc */
135 hctx->ssthresh = hctx->cwnd / 2;
136 if (hctx->ssthresh < 2)
141 /* clear state about stuff we sent */
142 hctx->seqt = hctx->seqh;
143 hctx->packets_acked = 0;
145 /* clear ack ratio state. */
148 ccid2_change_l_ack_ratio(sk, 1);
150 /* if we were blocked before, we may now send cwnd=1 packet */
151 if (sender_was_blocked)
152 tasklet_schedule(&dccp_sk(sk)->dccps_xmitlet);
153 ccid2_start_rto_timer(sk);
159 static void ccid2_start_rto_timer(struct sock *sk)
161 struct ccid2_hc_tx_sock *hctx = ccid2_hc_tx_sk(sk);
163 ccid2_pr_debug("setting RTO timeout=%ld\n", hctx->rto);
165 BUG_ON(timer_pending(&hctx->rtotimer));
166 sk_reset_timer(sk, &hctx->rtotimer,
167 jiffies + hctx->rto);
170 static void ccid2_hc_tx_packet_sent(struct sock *sk, unsigned int len)
172 struct dccp_sock *dp = dccp_sk(sk);
173 struct ccid2_hc_tx_sock *hctx = ccid2_hc_tx_sk(sk);
174 struct ccid2_seq *next;
178 hctx->seqh->ccid2s_seq = dp->dccps_gss;
179 hctx->seqh->ccid2s_acked = 0;
180 hctx->seqh->ccid2s_sent = jiffies;
182 next = hctx->seqh->ccid2s_next;
183 /* check if we need to alloc more space */
184 if (next == hctx->seqt) {
185 if (ccid2_hc_tx_alloc_seq(hctx)) {
186 DCCP_CRIT("packet history - out of memory!");
187 /* FIXME: find a more graceful way to bail out */
190 next = hctx->seqh->ccid2s_next;
191 BUG_ON(next == hctx->seqt);
195 ccid2_pr_debug("cwnd=%d pipe=%d\n", hctx->cwnd, hctx->pipe);
198 * FIXME: The code below is broken and the variables have been removed
199 * from the socket struct. The `ackloss' variable was always set to 0,
200 * and with arsent there are several problems:
201 * (i) it doesn't just count the number of Acks, but all sent packets;
202 * (ii) it is expressed in # of packets, not # of windows, so the
203 * comparison below uses the wrong formula: Appendix A of RFC 4341
204 * comes up with the number K = cwnd / (R^2 - R) of consecutive windows
205 * of data with no lost or marked Ack packets. If arsent were the # of
206 * consecutive Acks received without loss, then Ack Ratio needs to be
207 * decreased by 1 when
208 * arsent >= K * cwnd / R = cwnd^2 / (R^3 - R^2)
209 * where cwnd / R is the number of Acks received per window of data
210 * (cf. RFC 4341, App. A). The problems are that
211 * - arsent counts other packets as well;
212 * - the comparison uses a formula different from RFC 4341;
213 * - computing a cubic/quadratic equation each time is too complicated.
214 * Hence a different algorithm is needed.
217 /* Ack Ratio. Need to maintain a concept of how many windows we sent */
219 /* We had an ack loss in this window... */
221 if (hctx->arsent >= hctx->cwnd) {
226 /* No acks lost up to now... */
227 /* decrease ack ratio if enough packets were sent */
228 if (dp->dccps_l_ack_ratio > 1) {
229 /* XXX don't calculate denominator each time */
230 int denom = dp->dccps_l_ack_ratio * dp->dccps_l_ack_ratio -
231 dp->dccps_l_ack_ratio;
233 denom = hctx->cwnd * hctx->cwnd / denom;
235 if (hctx->arsent >= denom) {
236 ccid2_change_l_ack_ratio(sk, dp->dccps_l_ack_ratio - 1);
240 /* we can't increase ack ratio further [1] */
241 hctx->arsent = 0; /* or maybe set it to cwnd*/
246 /* setup RTO timer */
247 if (!timer_pending(&hctx->rtotimer))
248 ccid2_start_rto_timer(sk);
250 #ifdef CONFIG_IP_DCCP_CCID2_DEBUG
252 struct ccid2_seq *seqp = hctx->seqt;
254 while (seqp != hctx->seqh) {
255 ccid2_pr_debug("out seq=%llu acked=%d time=%lu\n",
256 (unsigned long long)seqp->ccid2s_seq,
257 seqp->ccid2s_acked, seqp->ccid2s_sent);
258 seqp = seqp->ccid2s_next;
261 ccid2_pr_debug("=========\n");
265 static void ccid2_hc_tx_kill_rto_timer(struct sock *sk)
267 struct ccid2_hc_tx_sock *hctx = ccid2_hc_tx_sk(sk);
269 sk_stop_timer(sk, &hctx->rtotimer);
270 ccid2_pr_debug("deleted RTO timer\n");
274 * ccid2_rtt_estimator - Sample RTT and compute RTO using RFC2988 algorithm
275 * This code is almost identical with TCP's tcp_rtt_estimator(), since
276 * - it has a higher sampling frequency (recommended by RFC 1323),
277 * - the RTO does not collapse into RTT due to RTTVAR going towards zero,
278 * - it is simple (cf. more complex proposals such as Eifel timer or research
279 * which suggests that the gain should be set according to window size),
280 * - in tests it was found to work well with CCID2 [gerrit].
282 static void ccid2_rtt_estimator(struct sock *sk, const long mrtt)
284 struct ccid2_hc_tx_sock *hctx = ccid2_hc_tx_sk(sk);
287 if (hctx->srtt == 0) {
288 /* First measurement m */
292 hctx->mdev_max = max(TCP_RTO_MIN, hctx->mdev);
293 hctx->rttvar = hctx->mdev_max;
294 hctx->rtt_seq = dccp_sk(sk)->dccps_gss;
296 /* Update scaled SRTT as SRTT += 1/8 * (m - SRTT) */
297 m -= (hctx->srtt >> 3);
300 /* Similarly, update scaled mdev with regard to |m| */
303 m -= (hctx->mdev >> 2);
305 * This neutralises RTO increase when RTT < SRTT - mdev
306 * (see P. Sarolahti, A. Kuznetsov,"Congestion Control
307 * in Linux TCP", USENIX 2002, pp. 49-62).
312 m -= (hctx->mdev >> 2);
316 if (hctx->mdev > hctx->mdev_max) {
317 hctx->mdev_max = hctx->mdev;
318 if (hctx->mdev_max > hctx->rttvar)
319 hctx->rttvar = hctx->mdev_max;
323 * Decay RTTVAR at most once per flight, exploiting that
324 * 1) pipe <= cwnd <= Sequence_Window = W (RFC 4340, 7.5.2)
325 * 2) AWL = GSS-W+1 <= GAR <= GSS (RFC 4340, 7.5.1)
326 * GAR is a useful bound for FlightSize = pipe, AWL is probably
327 * too low as it over-estimates pipe.
329 if (after48(dccp_sk(sk)->dccps_gar, hctx->rtt_seq)) {
330 if (hctx->mdev_max < hctx->rttvar)
331 hctx->rttvar -= (hctx->rttvar -
332 hctx->mdev_max) >> 2;
333 hctx->rtt_seq = dccp_sk(sk)->dccps_gss;
334 hctx->mdev_max = TCP_RTO_MIN;
339 * Set RTO from SRTT and RTTVAR
340 * Clock granularity is ignored since the minimum error for RTTVAR is
341 * clamped to 50msec (corresponding to HZ=20). This leads to a minimum
342 * RTO of 200msec. This agrees with TCP and RFC 4341, 5.: "Because DCCP
343 * does not retransmit data, DCCP does not require TCP's recommended
344 * minimum timeout of one second".
346 hctx->rto = (hctx->srtt >> 3) + hctx->rttvar;
348 if (hctx->rto > DCCP_RTO_MAX)
349 hctx->rto = DCCP_RTO_MAX;
352 static void ccid2_new_ack(struct sock *sk, struct ccid2_seq *seqp,
353 unsigned int *maxincr)
355 struct ccid2_hc_tx_sock *hctx = ccid2_hc_tx_sk(sk);
357 if (hctx->cwnd < hctx->ssthresh) {
358 if (*maxincr > 0 && ++hctx->packets_acked == 2) {
361 hctx->packets_acked = 0;
363 } else if (++hctx->packets_acked >= hctx->cwnd) {
365 hctx->packets_acked = 0;
368 * FIXME: RTT is sampled several times per acknowledgment (for each
369 * entry in the Ack Vector), instead of once per Ack (as in TCP SACK).
370 * This causes the RTT to be over-estimated, since the older entries
371 * in the Ack Vector have earlier sending times.
372 * The cleanest solution is to not use the ccid2s_sent field at all
373 * and instead use DCCP timestamps - need to be resolved at some time.
375 ccid2_rtt_estimator(sk, jiffies - seqp->ccid2s_sent);
378 static void ccid2_congestion_event(struct sock *sk, struct ccid2_seq *seqp)
380 struct ccid2_hc_tx_sock *hctx = ccid2_hc_tx_sk(sk);
382 if (time_before(seqp->ccid2s_sent, hctx->last_cong)) {
383 ccid2_pr_debug("Multiple losses in an RTT---treating as one\n");
387 hctx->last_cong = jiffies;
389 hctx->cwnd = hctx->cwnd / 2 ? : 1U;
390 hctx->ssthresh = max(hctx->cwnd, 2U);
392 /* Avoid spurious timeouts resulting from Ack Ratio > cwnd */
393 if (dccp_sk(sk)->dccps_l_ack_ratio > hctx->cwnd)
394 ccid2_change_l_ack_ratio(sk, hctx->cwnd);
397 static int ccid2_hc_tx_parse_options(struct sock *sk, u8 packet_type,
398 u8 option, u8 *optval, u8 optlen)
400 struct ccid2_hc_tx_sock *hctx = ccid2_hc_tx_sk(sk);
403 case DCCPO_ACK_VECTOR_0:
404 case DCCPO_ACK_VECTOR_1:
405 return dccp_ackvec_parsed_add(&hctx->av_chunks, optval, optlen,
406 option - DCCPO_ACK_VECTOR_0);
411 static void ccid2_hc_tx_packet_recv(struct sock *sk, struct sk_buff *skb)
413 struct dccp_sock *dp = dccp_sk(sk);
414 struct ccid2_hc_tx_sock *hctx = ccid2_hc_tx_sk(sk);
415 const bool sender_was_blocked = ccid2_cwnd_network_limited(hctx);
416 struct dccp_ackvec_parsed *avp;
418 struct ccid2_seq *seqp;
420 unsigned int maxincr = 0;
422 /* check reverse path congestion */
423 seqno = DCCP_SKB_CB(skb)->dccpd_seq;
425 /* XXX this whole "algorithm" is broken. Need to fix it to keep track
426 * of the seqnos of the dupacks so that rpseq and rpdupack are correct
429 /* need to bootstrap */
430 if (hctx->rpdupack == -1) {
434 /* check if packet is consecutive */
435 if (dccp_delta_seqno(hctx->rpseq, seqno) == 1)
437 /* it's a later packet */
438 else if (after48(seqno, hctx->rpseq)) {
441 /* check if we got enough dupacks */
442 if (hctx->rpdupack >= NUMDUPACK) {
443 hctx->rpdupack = -1; /* XXX lame */
446 ccid2_change_l_ack_ratio(sk, 2 * dp->dccps_l_ack_ratio);
451 /* check forward path congestion */
452 if (dccp_packet_without_ack(skb))
455 /* still didn't send out new data packets */
456 if (hctx->seqh == hctx->seqt)
459 ackno = DCCP_SKB_CB(skb)->dccpd_ack_seq;
460 if (after48(ackno, hctx->high_ack))
461 hctx->high_ack = ackno;
464 while (before48(seqp->ccid2s_seq, ackno)) {
465 seqp = seqp->ccid2s_next;
466 if (seqp == hctx->seqh) {
467 seqp = hctx->seqh->ccid2s_prev;
473 * In slow-start, cwnd can increase up to a maximum of Ack Ratio/2
474 * packets per acknowledgement. Rounding up avoids that cwnd is not
475 * advanced when Ack Ratio is 1 and gives a slight edge otherwise.
477 if (hctx->cwnd < hctx->ssthresh)
478 maxincr = DIV_ROUND_UP(dp->dccps_l_ack_ratio, 2);
480 /* go through all ack vectors */
481 list_for_each_entry(avp, &hctx->av_chunks, node) {
482 /* go through this ack vector */
483 for (; avp->len--; avp->vec++) {
484 u64 ackno_end_rl = SUB48(ackno,
485 dccp_ackvec_runlen(avp->vec));
487 ccid2_pr_debug("ackvec %llu |%u,%u|\n",
488 (unsigned long long)ackno,
489 dccp_ackvec_state(avp->vec) >> 6,
490 dccp_ackvec_runlen(avp->vec));
491 /* if the seqno we are analyzing is larger than the
492 * current ackno, then move towards the tail of our
495 while (after48(seqp->ccid2s_seq, ackno)) {
496 if (seqp == hctx->seqt) {
500 seqp = seqp->ccid2s_prev;
505 /* check all seqnos in the range of the vector
508 while (between48(seqp->ccid2s_seq,ackno_end_rl,ackno)) {
509 const u8 state = dccp_ackvec_state(avp->vec);
511 /* new packet received or marked */
512 if (state != DCCPAV_NOT_RECEIVED &&
513 !seqp->ccid2s_acked) {
514 if (state == DCCPAV_ECN_MARKED)
515 ccid2_congestion_event(sk,
518 ccid2_new_ack(sk, seqp,
521 seqp->ccid2s_acked = 1;
522 ccid2_pr_debug("Got ack for %llu\n",
523 (unsigned long long)seqp->ccid2s_seq);
526 if (seqp == hctx->seqt) {
530 seqp = seqp->ccid2s_prev;
535 ackno = SUB48(ackno_end_rl, 1);
541 /* The state about what is acked should be correct now
542 * Check for NUMDUPACK
545 while (before48(seqp->ccid2s_seq, hctx->high_ack)) {
546 seqp = seqp->ccid2s_next;
547 if (seqp == hctx->seqh) {
548 seqp = hctx->seqh->ccid2s_prev;
554 if (seqp->ccid2s_acked) {
556 if (done == NUMDUPACK)
559 if (seqp == hctx->seqt)
561 seqp = seqp->ccid2s_prev;
564 /* If there are at least 3 acknowledgements, anything unacknowledged
565 * below the last sequence number is considered lost
567 if (done == NUMDUPACK) {
568 struct ccid2_seq *last_acked = seqp;
570 /* check for lost packets */
572 if (!seqp->ccid2s_acked) {
573 ccid2_pr_debug("Packet lost: %llu\n",
574 (unsigned long long)seqp->ccid2s_seq);
575 /* XXX need to traverse from tail -> head in
576 * order to detect multiple congestion events in
579 ccid2_congestion_event(sk, seqp);
582 if (seqp == hctx->seqt)
584 seqp = seqp->ccid2s_prev;
587 hctx->seqt = last_acked;
590 /* trim acked packets in tail */
591 while (hctx->seqt != hctx->seqh) {
592 if (!hctx->seqt->ccid2s_acked)
595 hctx->seqt = hctx->seqt->ccid2s_next;
598 /* restart RTO timer if not all outstanding data has been acked */
600 sk_stop_timer(sk, &hctx->rtotimer);
602 sk_reset_timer(sk, &hctx->rtotimer, jiffies + hctx->rto);
604 /* check if incoming Acks allow pending packets to be sent */
605 if (sender_was_blocked && !ccid2_cwnd_network_limited(hctx))
606 tasklet_schedule(&dccp_sk(sk)->dccps_xmitlet);
607 dccp_ackvec_parsed_cleanup(&hctx->av_chunks);
610 static int ccid2_hc_tx_init(struct ccid *ccid, struct sock *sk)
612 struct ccid2_hc_tx_sock *hctx = ccid_priv(ccid);
613 struct dccp_sock *dp = dccp_sk(sk);
616 /* RFC 4341, 5: initialise ssthresh to arbitrarily high (max) value */
617 hctx->ssthresh = ~0U;
620 * RFC 4341, 5: "The cwnd parameter is initialized to at most four
621 * packets for new connections, following the rules from [RFC3390]".
622 * We need to convert the bytes of RFC3390 into the packets of RFC 4341.
624 hctx->cwnd = clamp(4380U / dp->dccps_mss_cache, 2U, 4U);
626 /* Make sure that Ack Ratio is enabled and within bounds. */
627 max_ratio = DIV_ROUND_UP(hctx->cwnd, 2);
628 if (dp->dccps_l_ack_ratio == 0 || dp->dccps_l_ack_ratio > max_ratio)
629 dp->dccps_l_ack_ratio = max_ratio;
631 /* XXX init ~ to window size... */
632 if (ccid2_hc_tx_alloc_seq(hctx))
635 hctx->rto = DCCP_TIMEOUT_INIT;
637 hctx->last_cong = jiffies;
638 setup_timer(&hctx->rtotimer, ccid2_hc_tx_rto_expire, (unsigned long)sk);
639 INIT_LIST_HEAD(&hctx->av_chunks);
643 static void ccid2_hc_tx_exit(struct sock *sk)
645 struct ccid2_hc_tx_sock *hctx = ccid2_hc_tx_sk(sk);
648 ccid2_hc_tx_kill_rto_timer(sk);
650 for (i = 0; i < hctx->seqbufc; i++)
651 kfree(hctx->seqbuf[i]);
655 static void ccid2_hc_rx_packet_recv(struct sock *sk, struct sk_buff *skb)
657 const struct dccp_sock *dp = dccp_sk(sk);
658 struct ccid2_hc_rx_sock *hcrx = ccid2_hc_rx_sk(sk);
660 switch (DCCP_SKB_CB(skb)->dccpd_type) {
662 case DCCP_PKT_DATAACK:
664 if (hcrx->data >= dp->dccps_r_ack_ratio) {
672 static struct ccid_operations ccid2 = {
673 .ccid_id = DCCPC_CCID2,
674 .ccid_name = "TCP-like",
675 .ccid_owner = THIS_MODULE,
676 .ccid_hc_tx_obj_size = sizeof(struct ccid2_hc_tx_sock),
677 .ccid_hc_tx_init = ccid2_hc_tx_init,
678 .ccid_hc_tx_exit = ccid2_hc_tx_exit,
679 .ccid_hc_tx_send_packet = ccid2_hc_tx_send_packet,
680 .ccid_hc_tx_packet_sent = ccid2_hc_tx_packet_sent,
681 .ccid_hc_tx_parse_options = ccid2_hc_tx_parse_options,
682 .ccid_hc_tx_packet_recv = ccid2_hc_tx_packet_recv,
683 .ccid_hc_rx_obj_size = sizeof(struct ccid2_hc_rx_sock),
684 .ccid_hc_rx_packet_recv = ccid2_hc_rx_packet_recv,
687 #ifdef CONFIG_IP_DCCP_CCID2_DEBUG
688 module_param(ccid2_debug, bool, 0644);
689 MODULE_PARM_DESC(ccid2_debug, "Enable debug messages");
692 static __init int ccid2_module_init(void)
694 return ccid_register(&ccid2);
696 module_init(ccid2_module_init);
698 static __exit void ccid2_module_exit(void)
700 ccid_unregister(&ccid2);
702 module_exit(ccid2_module_exit);
704 MODULE_AUTHOR("Andrea Bittau <a.bittau@cs.ucl.ac.uk>");
705 MODULE_DESCRIPTION("DCCP TCP-Like (CCID2) CCID");
706 MODULE_LICENSE("GPL");
707 MODULE_ALIAS("net-dccp-ccid-2");