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_hc_tx_rto_expire(unsigned long data)
115 struct sock *sk = (struct sock *)data;
116 struct ccid2_hc_tx_sock *hctx = ccid2_hc_tx_sk(sk);
117 const bool sender_was_blocked = ccid2_cwnd_network_limited(hctx);
120 if (sock_owned_by_user(sk)) {
121 sk_reset_timer(sk, &hctx->rtotimer, jiffies + HZ / 5);
125 ccid2_pr_debug("RTO_EXPIRE\n");
129 if (hctx->rto > DCCP_RTO_MAX)
130 hctx->rto = DCCP_RTO_MAX;
132 /* adjust pipe, cwnd etc */
133 hctx->ssthresh = hctx->cwnd / 2;
134 if (hctx->ssthresh < 2)
139 /* clear state about stuff we sent */
140 hctx->seqt = hctx->seqh;
141 hctx->packets_acked = 0;
143 /* clear ack ratio state. */
146 ccid2_change_l_ack_ratio(sk, 1);
148 /* if we were blocked before, we may now send cwnd=1 packet */
149 if (sender_was_blocked)
150 tasklet_schedule(&dccp_sk(sk)->dccps_xmitlet);
151 /* restart backed-off timer */
152 sk_reset_timer(sk, &hctx->rtotimer, jiffies + hctx->rto);
158 static void ccid2_hc_tx_packet_sent(struct sock *sk, unsigned int len)
160 struct dccp_sock *dp = dccp_sk(sk);
161 struct ccid2_hc_tx_sock *hctx = ccid2_hc_tx_sk(sk);
162 struct ccid2_seq *next;
166 hctx->seqh->ccid2s_seq = dp->dccps_gss;
167 hctx->seqh->ccid2s_acked = 0;
168 hctx->seqh->ccid2s_sent = jiffies;
170 next = hctx->seqh->ccid2s_next;
171 /* check if we need to alloc more space */
172 if (next == hctx->seqt) {
173 if (ccid2_hc_tx_alloc_seq(hctx)) {
174 DCCP_CRIT("packet history - out of memory!");
175 /* FIXME: find a more graceful way to bail out */
178 next = hctx->seqh->ccid2s_next;
179 BUG_ON(next == hctx->seqt);
183 ccid2_pr_debug("cwnd=%d pipe=%d\n", hctx->cwnd, hctx->pipe);
186 * FIXME: The code below is broken and the variables have been removed
187 * from the socket struct. The `ackloss' variable was always set to 0,
188 * and with arsent there are several problems:
189 * (i) it doesn't just count the number of Acks, but all sent packets;
190 * (ii) it is expressed in # of packets, not # of windows, so the
191 * comparison below uses the wrong formula: Appendix A of RFC 4341
192 * comes up with the number K = cwnd / (R^2 - R) of consecutive windows
193 * of data with no lost or marked Ack packets. If arsent were the # of
194 * consecutive Acks received without loss, then Ack Ratio needs to be
195 * decreased by 1 when
196 * arsent >= K * cwnd / R = cwnd^2 / (R^3 - R^2)
197 * where cwnd / R is the number of Acks received per window of data
198 * (cf. RFC 4341, App. A). The problems are that
199 * - arsent counts other packets as well;
200 * - the comparison uses a formula different from RFC 4341;
201 * - computing a cubic/quadratic equation each time is too complicated.
202 * Hence a different algorithm is needed.
205 /* Ack Ratio. Need to maintain a concept of how many windows we sent */
207 /* We had an ack loss in this window... */
209 if (hctx->arsent >= hctx->cwnd) {
214 /* No acks lost up to now... */
215 /* decrease ack ratio if enough packets were sent */
216 if (dp->dccps_l_ack_ratio > 1) {
217 /* XXX don't calculate denominator each time */
218 int denom = dp->dccps_l_ack_ratio * dp->dccps_l_ack_ratio -
219 dp->dccps_l_ack_ratio;
221 denom = hctx->cwnd * hctx->cwnd / denom;
223 if (hctx->arsent >= denom) {
224 ccid2_change_l_ack_ratio(sk, dp->dccps_l_ack_ratio - 1);
228 /* we can't increase ack ratio further [1] */
229 hctx->arsent = 0; /* or maybe set it to cwnd*/
234 /* setup RTO timer */
235 if (!timer_pending(&hctx->rtotimer))
236 sk_reset_timer(sk, &hctx->rtotimer, jiffies + hctx->rto);
238 #ifdef CONFIG_IP_DCCP_CCID2_DEBUG
240 struct ccid2_seq *seqp = hctx->seqt;
242 while (seqp != hctx->seqh) {
243 ccid2_pr_debug("out seq=%llu acked=%d time=%lu\n",
244 (unsigned long long)seqp->ccid2s_seq,
245 seqp->ccid2s_acked, seqp->ccid2s_sent);
246 seqp = seqp->ccid2s_next;
249 ccid2_pr_debug("=========\n");
254 * ccid2_rtt_estimator - Sample RTT and compute RTO using RFC2988 algorithm
255 * This code is almost identical with TCP's tcp_rtt_estimator(), since
256 * - it has a higher sampling frequency (recommended by RFC 1323),
257 * - the RTO does not collapse into RTT due to RTTVAR going towards zero,
258 * - it is simple (cf. more complex proposals such as Eifel timer or research
259 * which suggests that the gain should be set according to window size),
260 * - in tests it was found to work well with CCID2 [gerrit].
262 static void ccid2_rtt_estimator(struct sock *sk, const long mrtt)
264 struct ccid2_hc_tx_sock *hctx = ccid2_hc_tx_sk(sk);
267 if (hctx->srtt == 0) {
268 /* First measurement m */
272 hctx->mdev_max = max(TCP_RTO_MIN, hctx->mdev);
273 hctx->rttvar = hctx->mdev_max;
274 hctx->rtt_seq = dccp_sk(sk)->dccps_gss;
276 /* Update scaled SRTT as SRTT += 1/8 * (m - SRTT) */
277 m -= (hctx->srtt >> 3);
280 /* Similarly, update scaled mdev with regard to |m| */
283 m -= (hctx->mdev >> 2);
285 * This neutralises RTO increase when RTT < SRTT - mdev
286 * (see P. Sarolahti, A. Kuznetsov,"Congestion Control
287 * in Linux TCP", USENIX 2002, pp. 49-62).
292 m -= (hctx->mdev >> 2);
296 if (hctx->mdev > hctx->mdev_max) {
297 hctx->mdev_max = hctx->mdev;
298 if (hctx->mdev_max > hctx->rttvar)
299 hctx->rttvar = hctx->mdev_max;
303 * Decay RTTVAR at most once per flight, exploiting that
304 * 1) pipe <= cwnd <= Sequence_Window = W (RFC 4340, 7.5.2)
305 * 2) AWL = GSS-W+1 <= GAR <= GSS (RFC 4340, 7.5.1)
306 * GAR is a useful bound for FlightSize = pipe, AWL is probably
307 * too low as it over-estimates pipe.
309 if (after48(dccp_sk(sk)->dccps_gar, hctx->rtt_seq)) {
310 if (hctx->mdev_max < hctx->rttvar)
311 hctx->rttvar -= (hctx->rttvar -
312 hctx->mdev_max) >> 2;
313 hctx->rtt_seq = dccp_sk(sk)->dccps_gss;
314 hctx->mdev_max = TCP_RTO_MIN;
319 * Set RTO from SRTT and RTTVAR
320 * Clock granularity is ignored since the minimum error for RTTVAR is
321 * clamped to 50msec (corresponding to HZ=20). This leads to a minimum
322 * RTO of 200msec. This agrees with TCP and RFC 4341, 5.: "Because DCCP
323 * does not retransmit data, DCCP does not require TCP's recommended
324 * minimum timeout of one second".
326 hctx->rto = (hctx->srtt >> 3) + hctx->rttvar;
328 if (hctx->rto > DCCP_RTO_MAX)
329 hctx->rto = DCCP_RTO_MAX;
332 static void ccid2_new_ack(struct sock *sk, struct ccid2_seq *seqp,
333 unsigned int *maxincr)
335 struct ccid2_hc_tx_sock *hctx = ccid2_hc_tx_sk(sk);
337 if (hctx->cwnd < hctx->ssthresh) {
338 if (*maxincr > 0 && ++hctx->packets_acked == 2) {
341 hctx->packets_acked = 0;
343 } else if (++hctx->packets_acked >= hctx->cwnd) {
345 hctx->packets_acked = 0;
348 * FIXME: RTT is sampled several times per acknowledgment (for each
349 * entry in the Ack Vector), instead of once per Ack (as in TCP SACK).
350 * This causes the RTT to be over-estimated, since the older entries
351 * in the Ack Vector have earlier sending times.
352 * The cleanest solution is to not use the ccid2s_sent field at all
353 * and instead use DCCP timestamps - need to be resolved at some time.
355 ccid2_rtt_estimator(sk, jiffies - seqp->ccid2s_sent);
358 static void ccid2_congestion_event(struct sock *sk, struct ccid2_seq *seqp)
360 struct ccid2_hc_tx_sock *hctx = ccid2_hc_tx_sk(sk);
362 if (time_before(seqp->ccid2s_sent, hctx->last_cong)) {
363 ccid2_pr_debug("Multiple losses in an RTT---treating as one\n");
367 hctx->last_cong = jiffies;
369 hctx->cwnd = hctx->cwnd / 2 ? : 1U;
370 hctx->ssthresh = max(hctx->cwnd, 2U);
372 /* Avoid spurious timeouts resulting from Ack Ratio > cwnd */
373 if (dccp_sk(sk)->dccps_l_ack_ratio > hctx->cwnd)
374 ccid2_change_l_ack_ratio(sk, hctx->cwnd);
377 static int ccid2_hc_tx_parse_options(struct sock *sk, u8 packet_type,
378 u8 option, u8 *optval, u8 optlen)
380 struct ccid2_hc_tx_sock *hctx = ccid2_hc_tx_sk(sk);
383 case DCCPO_ACK_VECTOR_0:
384 case DCCPO_ACK_VECTOR_1:
385 return dccp_ackvec_parsed_add(&hctx->av_chunks, optval, optlen,
386 option - DCCPO_ACK_VECTOR_0);
391 static void ccid2_hc_tx_packet_recv(struct sock *sk, struct sk_buff *skb)
393 struct dccp_sock *dp = dccp_sk(sk);
394 struct ccid2_hc_tx_sock *hctx = ccid2_hc_tx_sk(sk);
395 const bool sender_was_blocked = ccid2_cwnd_network_limited(hctx);
396 struct dccp_ackvec_parsed *avp;
398 struct ccid2_seq *seqp;
400 unsigned int maxincr = 0;
402 /* check reverse path congestion */
403 seqno = DCCP_SKB_CB(skb)->dccpd_seq;
405 /* XXX this whole "algorithm" is broken. Need to fix it to keep track
406 * of the seqnos of the dupacks so that rpseq and rpdupack are correct
409 /* need to bootstrap */
410 if (hctx->rpdupack == -1) {
414 /* check if packet is consecutive */
415 if (dccp_delta_seqno(hctx->rpseq, seqno) == 1)
417 /* it's a later packet */
418 else if (after48(seqno, hctx->rpseq)) {
421 /* check if we got enough dupacks */
422 if (hctx->rpdupack >= NUMDUPACK) {
423 hctx->rpdupack = -1; /* XXX lame */
426 ccid2_change_l_ack_ratio(sk, 2 * dp->dccps_l_ack_ratio);
431 /* check forward path congestion */
432 if (dccp_packet_without_ack(skb))
435 /* still didn't send out new data packets */
436 if (hctx->seqh == hctx->seqt)
439 ackno = DCCP_SKB_CB(skb)->dccpd_ack_seq;
440 if (after48(ackno, hctx->high_ack))
441 hctx->high_ack = ackno;
444 while (before48(seqp->ccid2s_seq, ackno)) {
445 seqp = seqp->ccid2s_next;
446 if (seqp == hctx->seqh) {
447 seqp = hctx->seqh->ccid2s_prev;
453 * In slow-start, cwnd can increase up to a maximum of Ack Ratio/2
454 * packets per acknowledgement. Rounding up avoids that cwnd is not
455 * advanced when Ack Ratio is 1 and gives a slight edge otherwise.
457 if (hctx->cwnd < hctx->ssthresh)
458 maxincr = DIV_ROUND_UP(dp->dccps_l_ack_ratio, 2);
460 /* go through all ack vectors */
461 list_for_each_entry(avp, &hctx->av_chunks, node) {
462 /* go through this ack vector */
463 for (; avp->len--; avp->vec++) {
464 u64 ackno_end_rl = SUB48(ackno,
465 dccp_ackvec_runlen(avp->vec));
467 ccid2_pr_debug("ackvec %llu |%u,%u|\n",
468 (unsigned long long)ackno,
469 dccp_ackvec_state(avp->vec) >> 6,
470 dccp_ackvec_runlen(avp->vec));
471 /* if the seqno we are analyzing is larger than the
472 * current ackno, then move towards the tail of our
475 while (after48(seqp->ccid2s_seq, ackno)) {
476 if (seqp == hctx->seqt) {
480 seqp = seqp->ccid2s_prev;
485 /* check all seqnos in the range of the vector
488 while (between48(seqp->ccid2s_seq,ackno_end_rl,ackno)) {
489 const u8 state = dccp_ackvec_state(avp->vec);
491 /* new packet received or marked */
492 if (state != DCCPAV_NOT_RECEIVED &&
493 !seqp->ccid2s_acked) {
494 if (state == DCCPAV_ECN_MARKED)
495 ccid2_congestion_event(sk,
498 ccid2_new_ack(sk, seqp,
501 seqp->ccid2s_acked = 1;
502 ccid2_pr_debug("Got ack for %llu\n",
503 (unsigned long long)seqp->ccid2s_seq);
506 if (seqp == hctx->seqt) {
510 seqp = seqp->ccid2s_prev;
515 ackno = SUB48(ackno_end_rl, 1);
521 /* The state about what is acked should be correct now
522 * Check for NUMDUPACK
525 while (before48(seqp->ccid2s_seq, hctx->high_ack)) {
526 seqp = seqp->ccid2s_next;
527 if (seqp == hctx->seqh) {
528 seqp = hctx->seqh->ccid2s_prev;
534 if (seqp->ccid2s_acked) {
536 if (done == NUMDUPACK)
539 if (seqp == hctx->seqt)
541 seqp = seqp->ccid2s_prev;
544 /* If there are at least 3 acknowledgements, anything unacknowledged
545 * below the last sequence number is considered lost
547 if (done == NUMDUPACK) {
548 struct ccid2_seq *last_acked = seqp;
550 /* check for lost packets */
552 if (!seqp->ccid2s_acked) {
553 ccid2_pr_debug("Packet lost: %llu\n",
554 (unsigned long long)seqp->ccid2s_seq);
555 /* XXX need to traverse from tail -> head in
556 * order to detect multiple congestion events in
559 ccid2_congestion_event(sk, seqp);
562 if (seqp == hctx->seqt)
564 seqp = seqp->ccid2s_prev;
567 hctx->seqt = last_acked;
570 /* trim acked packets in tail */
571 while (hctx->seqt != hctx->seqh) {
572 if (!hctx->seqt->ccid2s_acked)
575 hctx->seqt = hctx->seqt->ccid2s_next;
578 /* restart RTO timer if not all outstanding data has been acked */
580 sk_stop_timer(sk, &hctx->rtotimer);
582 sk_reset_timer(sk, &hctx->rtotimer, jiffies + hctx->rto);
584 /* check if incoming Acks allow pending packets to be sent */
585 if (sender_was_blocked && !ccid2_cwnd_network_limited(hctx))
586 tasklet_schedule(&dccp_sk(sk)->dccps_xmitlet);
587 dccp_ackvec_parsed_cleanup(&hctx->av_chunks);
590 static int ccid2_hc_tx_init(struct ccid *ccid, struct sock *sk)
592 struct ccid2_hc_tx_sock *hctx = ccid_priv(ccid);
593 struct dccp_sock *dp = dccp_sk(sk);
596 /* RFC 4341, 5: initialise ssthresh to arbitrarily high (max) value */
597 hctx->ssthresh = ~0U;
599 /* Use larger initial windows (RFC 3390, rfc2581bis) */
600 hctx->cwnd = rfc3390_bytes_to_packets(dp->dccps_mss_cache);
602 /* Make sure that Ack Ratio is enabled and within bounds. */
603 max_ratio = DIV_ROUND_UP(hctx->cwnd, 2);
604 if (dp->dccps_l_ack_ratio == 0 || dp->dccps_l_ack_ratio > max_ratio)
605 dp->dccps_l_ack_ratio = max_ratio;
607 /* XXX init ~ to window size... */
608 if (ccid2_hc_tx_alloc_seq(hctx))
611 hctx->rto = DCCP_TIMEOUT_INIT;
613 hctx->last_cong = jiffies;
614 setup_timer(&hctx->rtotimer, ccid2_hc_tx_rto_expire, (unsigned long)sk);
615 INIT_LIST_HEAD(&hctx->av_chunks);
619 static void ccid2_hc_tx_exit(struct sock *sk)
621 struct ccid2_hc_tx_sock *hctx = ccid2_hc_tx_sk(sk);
624 sk_stop_timer(sk, &hctx->rtotimer);
626 for (i = 0; i < hctx->seqbufc; i++)
627 kfree(hctx->seqbuf[i]);
631 static void ccid2_hc_rx_packet_recv(struct sock *sk, struct sk_buff *skb)
633 const struct dccp_sock *dp = dccp_sk(sk);
634 struct ccid2_hc_rx_sock *hcrx = ccid2_hc_rx_sk(sk);
636 switch (DCCP_SKB_CB(skb)->dccpd_type) {
638 case DCCP_PKT_DATAACK:
640 if (hcrx->data >= dp->dccps_r_ack_ratio) {
648 static struct ccid_operations ccid2 = {
649 .ccid_id = DCCPC_CCID2,
650 .ccid_name = "TCP-like",
651 .ccid_owner = THIS_MODULE,
652 .ccid_hc_tx_obj_size = sizeof(struct ccid2_hc_tx_sock),
653 .ccid_hc_tx_init = ccid2_hc_tx_init,
654 .ccid_hc_tx_exit = ccid2_hc_tx_exit,
655 .ccid_hc_tx_send_packet = ccid2_hc_tx_send_packet,
656 .ccid_hc_tx_packet_sent = ccid2_hc_tx_packet_sent,
657 .ccid_hc_tx_parse_options = ccid2_hc_tx_parse_options,
658 .ccid_hc_tx_packet_recv = ccid2_hc_tx_packet_recv,
659 .ccid_hc_rx_obj_size = sizeof(struct ccid2_hc_rx_sock),
660 .ccid_hc_rx_packet_recv = ccid2_hc_rx_packet_recv,
663 #ifdef CONFIG_IP_DCCP_CCID2_DEBUG
664 module_param(ccid2_debug, bool, 0644);
665 MODULE_PARM_DESC(ccid2_debug, "Enable debug messages");
668 static __init int ccid2_module_init(void)
670 return ccid_register(&ccid2);
672 module_init(ccid2_module_init);
674 static __exit void ccid2_module_exit(void)
676 ccid_unregister(&ccid2);
678 module_exit(ccid2_module_exit);
680 MODULE_AUTHOR("Andrea Bittau <a.bittau@cs.ucl.ac.uk>");
681 MODULE_DESCRIPTION("DCCP TCP-Like (CCID2) CCID");
682 MODULE_LICENSE("GPL");
683 MODULE_ALIAS("net-dccp-ccid-2");