Loading...
1/*
2 * Copyright (c) 2005, 2006 Andrea Bittau <a.bittau@cs.ucl.ac.uk>
3 *
4 * Changes to meet Linux coding standards, and DCCP infrastructure fixes.
5 *
6 * Copyright (c) 2006 Arnaldo Carvalho de Melo <acme@conectiva.com.br>
7 *
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License as published by
10 * the Free Software Foundation; either version 2 of the License, or
11 * (at your option) any later version.
12 *
13 * This program is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 * GNU General Public License for more details.
17 *
18 * You should have received a copy of the GNU General Public License
19 * along with this program; if not, write to the Free Software
20 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
21 */
22
23/*
24 * This implementation should follow RFC 4341
25 */
26#include <linux/slab.h>
27#include "../feat.h"
28#include "ccid2.h"
29
30
31#ifdef CONFIG_IP_DCCP_CCID2_DEBUG
32static int ccid2_debug;
33#define ccid2_pr_debug(format, a...) DCCP_PR_DEBUG(ccid2_debug, format, ##a)
34#else
35#define ccid2_pr_debug(format, a...)
36#endif
37
38static int ccid2_hc_tx_alloc_seq(struct ccid2_hc_tx_sock *hc)
39{
40 struct ccid2_seq *seqp;
41 int i;
42
43 /* check if we have space to preserve the pointer to the buffer */
44 if (hc->tx_seqbufc >= (sizeof(hc->tx_seqbuf) /
45 sizeof(struct ccid2_seq *)))
46 return -ENOMEM;
47
48 /* allocate buffer and initialize linked list */
49 seqp = kmalloc(CCID2_SEQBUF_LEN * sizeof(struct ccid2_seq), gfp_any());
50 if (seqp == NULL)
51 return -ENOMEM;
52
53 for (i = 0; i < (CCID2_SEQBUF_LEN - 1); i++) {
54 seqp[i].ccid2s_next = &seqp[i + 1];
55 seqp[i + 1].ccid2s_prev = &seqp[i];
56 }
57 seqp[CCID2_SEQBUF_LEN - 1].ccid2s_next = seqp;
58 seqp->ccid2s_prev = &seqp[CCID2_SEQBUF_LEN - 1];
59
60 /* This is the first allocation. Initiate the head and tail. */
61 if (hc->tx_seqbufc == 0)
62 hc->tx_seqh = hc->tx_seqt = seqp;
63 else {
64 /* link the existing list with the one we just created */
65 hc->tx_seqh->ccid2s_next = seqp;
66 seqp->ccid2s_prev = hc->tx_seqh;
67
68 hc->tx_seqt->ccid2s_prev = &seqp[CCID2_SEQBUF_LEN - 1];
69 seqp[CCID2_SEQBUF_LEN - 1].ccid2s_next = hc->tx_seqt;
70 }
71
72 /* store the original pointer to the buffer so we can free it */
73 hc->tx_seqbuf[hc->tx_seqbufc] = seqp;
74 hc->tx_seqbufc++;
75
76 return 0;
77}
78
79static int ccid2_hc_tx_send_packet(struct sock *sk, struct sk_buff *skb)
80{
81 if (ccid2_cwnd_network_limited(ccid2_hc_tx_sk(sk)))
82 return CCID_PACKET_WILL_DEQUEUE_LATER;
83 return CCID_PACKET_SEND_AT_ONCE;
84}
85
86static void ccid2_change_l_ack_ratio(struct sock *sk, u32 val)
87{
88 struct dccp_sock *dp = dccp_sk(sk);
89 u32 max_ratio = DIV_ROUND_UP(ccid2_hc_tx_sk(sk)->tx_cwnd, 2);
90
91 /*
92 * Ensure that Ack Ratio does not exceed ceil(cwnd/2), which is (2) from
93 * RFC 4341, 6.1.2. We ignore the statement that Ack Ratio 2 is always
94 * acceptable since this causes starvation/deadlock whenever cwnd < 2.
95 * The same problem arises when Ack Ratio is 0 (ie. Ack Ratio disabled).
96 */
97 if (val == 0 || val > max_ratio) {
98 DCCP_WARN("Limiting Ack Ratio (%u) to %u\n", val, max_ratio);
99 val = max_ratio;
100 }
101 if (val > DCCPF_ACK_RATIO_MAX)
102 val = DCCPF_ACK_RATIO_MAX;
103
104 if (val == dp->dccps_l_ack_ratio)
105 return;
106
107 ccid2_pr_debug("changing local ack ratio to %u\n", val);
108 dp->dccps_l_ack_ratio = val;
109}
110
111static void ccid2_hc_tx_rto_expire(unsigned long data)
112{
113 struct sock *sk = (struct sock *)data;
114 struct ccid2_hc_tx_sock *hc = ccid2_hc_tx_sk(sk);
115 const bool sender_was_blocked = ccid2_cwnd_network_limited(hc);
116
117 bh_lock_sock(sk);
118 if (sock_owned_by_user(sk)) {
119 sk_reset_timer(sk, &hc->tx_rtotimer, jiffies + HZ / 5);
120 goto out;
121 }
122
123 ccid2_pr_debug("RTO_EXPIRE\n");
124
125 /* back-off timer */
126 hc->tx_rto <<= 1;
127 if (hc->tx_rto > DCCP_RTO_MAX)
128 hc->tx_rto = DCCP_RTO_MAX;
129
130 /* adjust pipe, cwnd etc */
131 hc->tx_ssthresh = hc->tx_cwnd / 2;
132 if (hc->tx_ssthresh < 2)
133 hc->tx_ssthresh = 2;
134 hc->tx_cwnd = 1;
135 hc->tx_pipe = 0;
136
137 /* clear state about stuff we sent */
138 hc->tx_seqt = hc->tx_seqh;
139 hc->tx_packets_acked = 0;
140
141 /* clear ack ratio state. */
142 hc->tx_rpseq = 0;
143 hc->tx_rpdupack = -1;
144 ccid2_change_l_ack_ratio(sk, 1);
145
146 /* if we were blocked before, we may now send cwnd=1 packet */
147 if (sender_was_blocked)
148 tasklet_schedule(&dccp_sk(sk)->dccps_xmitlet);
149 /* restart backed-off timer */
150 sk_reset_timer(sk, &hc->tx_rtotimer, jiffies + hc->tx_rto);
151out:
152 bh_unlock_sock(sk);
153 sock_put(sk);
154}
155
156/*
157 * Congestion window validation (RFC 2861).
158 */
159static int ccid2_do_cwv = 1;
160module_param(ccid2_do_cwv, bool, 0644);
161MODULE_PARM_DESC(ccid2_do_cwv, "Perform RFC2861 Congestion Window Validation");
162
163/**
164 * ccid2_update_used_window - Track how much of cwnd is actually used
165 * This is done in addition to CWV. The sender needs to have an idea of how many
166 * packets may be in flight, to set the local Sequence Window value accordingly
167 * (RFC 4340, 7.5.2). The CWV mechanism is exploited to keep track of the
168 * maximum-used window. We use an EWMA low-pass filter to filter out noise.
169 */
170static void ccid2_update_used_window(struct ccid2_hc_tx_sock *hc, u32 new_wnd)
171{
172 hc->tx_expected_wnd = (3 * hc->tx_expected_wnd + new_wnd) / 4;
173}
174
175/* This borrows the code of tcp_cwnd_application_limited() */
176static void ccid2_cwnd_application_limited(struct sock *sk, const u32 now)
177{
178 struct ccid2_hc_tx_sock *hc = ccid2_hc_tx_sk(sk);
179 /* don't reduce cwnd below the initial window (IW) */
180 u32 init_win = rfc3390_bytes_to_packets(dccp_sk(sk)->dccps_mss_cache),
181 win_used = max(hc->tx_cwnd_used, init_win);
182
183 if (win_used < hc->tx_cwnd) {
184 hc->tx_ssthresh = max(hc->tx_ssthresh,
185 (hc->tx_cwnd >> 1) + (hc->tx_cwnd >> 2));
186 hc->tx_cwnd = (hc->tx_cwnd + win_used) >> 1;
187 }
188 hc->tx_cwnd_used = 0;
189 hc->tx_cwnd_stamp = now;
190}
191
192/* This borrows the code of tcp_cwnd_restart() */
193static void ccid2_cwnd_restart(struct sock *sk, const u32 now)
194{
195 struct ccid2_hc_tx_sock *hc = ccid2_hc_tx_sk(sk);
196 u32 cwnd = hc->tx_cwnd, restart_cwnd,
197 iwnd = rfc3390_bytes_to_packets(dccp_sk(sk)->dccps_mss_cache);
198
199 hc->tx_ssthresh = max(hc->tx_ssthresh, (cwnd >> 1) + (cwnd >> 2));
200
201 /* don't reduce cwnd below the initial window (IW) */
202 restart_cwnd = min(cwnd, iwnd);
203 cwnd >>= (now - hc->tx_lsndtime) / hc->tx_rto;
204 hc->tx_cwnd = max(cwnd, restart_cwnd);
205
206 hc->tx_cwnd_stamp = now;
207 hc->tx_cwnd_used = 0;
208}
209
210static void ccid2_hc_tx_packet_sent(struct sock *sk, unsigned int len)
211{
212 struct dccp_sock *dp = dccp_sk(sk);
213 struct ccid2_hc_tx_sock *hc = ccid2_hc_tx_sk(sk);
214 const u32 now = ccid2_time_stamp;
215 struct ccid2_seq *next;
216
217 /* slow-start after idle periods (RFC 2581, RFC 2861) */
218 if (ccid2_do_cwv && !hc->tx_pipe &&
219 (s32)(now - hc->tx_lsndtime) >= hc->tx_rto)
220 ccid2_cwnd_restart(sk, now);
221
222 hc->tx_lsndtime = now;
223 hc->tx_pipe += 1;
224
225 /* see whether cwnd was fully used (RFC 2861), update expected window */
226 if (ccid2_cwnd_network_limited(hc)) {
227 ccid2_update_used_window(hc, hc->tx_cwnd);
228 hc->tx_cwnd_used = 0;
229 hc->tx_cwnd_stamp = now;
230 } else {
231 if (hc->tx_pipe > hc->tx_cwnd_used)
232 hc->tx_cwnd_used = hc->tx_pipe;
233
234 ccid2_update_used_window(hc, hc->tx_cwnd_used);
235
236 if (ccid2_do_cwv && (s32)(now - hc->tx_cwnd_stamp) >= hc->tx_rto)
237 ccid2_cwnd_application_limited(sk, now);
238 }
239
240 hc->tx_seqh->ccid2s_seq = dp->dccps_gss;
241 hc->tx_seqh->ccid2s_acked = 0;
242 hc->tx_seqh->ccid2s_sent = now;
243
244 next = hc->tx_seqh->ccid2s_next;
245 /* check if we need to alloc more space */
246 if (next == hc->tx_seqt) {
247 if (ccid2_hc_tx_alloc_seq(hc)) {
248 DCCP_CRIT("packet history - out of memory!");
249 /* FIXME: find a more graceful way to bail out */
250 return;
251 }
252 next = hc->tx_seqh->ccid2s_next;
253 BUG_ON(next == hc->tx_seqt);
254 }
255 hc->tx_seqh = next;
256
257 ccid2_pr_debug("cwnd=%d pipe=%d\n", hc->tx_cwnd, hc->tx_pipe);
258
259 /*
260 * FIXME: The code below is broken and the variables have been removed
261 * from the socket struct. The `ackloss' variable was always set to 0,
262 * and with arsent there are several problems:
263 * (i) it doesn't just count the number of Acks, but all sent packets;
264 * (ii) it is expressed in # of packets, not # of windows, so the
265 * comparison below uses the wrong formula: Appendix A of RFC 4341
266 * comes up with the number K = cwnd / (R^2 - R) of consecutive windows
267 * of data with no lost or marked Ack packets. If arsent were the # of
268 * consecutive Acks received without loss, then Ack Ratio needs to be
269 * decreased by 1 when
270 * arsent >= K * cwnd / R = cwnd^2 / (R^3 - R^2)
271 * where cwnd / R is the number of Acks received per window of data
272 * (cf. RFC 4341, App. A). The problems are that
273 * - arsent counts other packets as well;
274 * - the comparison uses a formula different from RFC 4341;
275 * - computing a cubic/quadratic equation each time is too complicated.
276 * Hence a different algorithm is needed.
277 */
278#if 0
279 /* Ack Ratio. Need to maintain a concept of how many windows we sent */
280 hc->tx_arsent++;
281 /* We had an ack loss in this window... */
282 if (hc->tx_ackloss) {
283 if (hc->tx_arsent >= hc->tx_cwnd) {
284 hc->tx_arsent = 0;
285 hc->tx_ackloss = 0;
286 }
287 } else {
288 /* No acks lost up to now... */
289 /* decrease ack ratio if enough packets were sent */
290 if (dp->dccps_l_ack_ratio > 1) {
291 /* XXX don't calculate denominator each time */
292 int denom = dp->dccps_l_ack_ratio * dp->dccps_l_ack_ratio -
293 dp->dccps_l_ack_ratio;
294
295 denom = hc->tx_cwnd * hc->tx_cwnd / denom;
296
297 if (hc->tx_arsent >= denom) {
298 ccid2_change_l_ack_ratio(sk, dp->dccps_l_ack_ratio - 1);
299 hc->tx_arsent = 0;
300 }
301 } else {
302 /* we can't increase ack ratio further [1] */
303 hc->tx_arsent = 0; /* or maybe set it to cwnd*/
304 }
305 }
306#endif
307
308 sk_reset_timer(sk, &hc->tx_rtotimer, jiffies + hc->tx_rto);
309
310#ifdef CONFIG_IP_DCCP_CCID2_DEBUG
311 do {
312 struct ccid2_seq *seqp = hc->tx_seqt;
313
314 while (seqp != hc->tx_seqh) {
315 ccid2_pr_debug("out seq=%llu acked=%d time=%u\n",
316 (unsigned long long)seqp->ccid2s_seq,
317 seqp->ccid2s_acked, seqp->ccid2s_sent);
318 seqp = seqp->ccid2s_next;
319 }
320 } while (0);
321 ccid2_pr_debug("=========\n");
322#endif
323}
324
325/**
326 * ccid2_rtt_estimator - Sample RTT and compute RTO using RFC2988 algorithm
327 * This code is almost identical with TCP's tcp_rtt_estimator(), since
328 * - it has a higher sampling frequency (recommended by RFC 1323),
329 * - the RTO does not collapse into RTT due to RTTVAR going towards zero,
330 * - it is simple (cf. more complex proposals such as Eifel timer or research
331 * which suggests that the gain should be set according to window size),
332 * - in tests it was found to work well with CCID2 [gerrit].
333 */
334static void ccid2_rtt_estimator(struct sock *sk, const long mrtt)
335{
336 struct ccid2_hc_tx_sock *hc = ccid2_hc_tx_sk(sk);
337 long m = mrtt ? : 1;
338
339 if (hc->tx_srtt == 0) {
340 /* First measurement m */
341 hc->tx_srtt = m << 3;
342 hc->tx_mdev = m << 1;
343
344 hc->tx_mdev_max = max(hc->tx_mdev, tcp_rto_min(sk));
345 hc->tx_rttvar = hc->tx_mdev_max;
346
347 hc->tx_rtt_seq = dccp_sk(sk)->dccps_gss;
348 } else {
349 /* Update scaled SRTT as SRTT += 1/8 * (m - SRTT) */
350 m -= (hc->tx_srtt >> 3);
351 hc->tx_srtt += m;
352
353 /* Similarly, update scaled mdev with regard to |m| */
354 if (m < 0) {
355 m = -m;
356 m -= (hc->tx_mdev >> 2);
357 /*
358 * This neutralises RTO increase when RTT < SRTT - mdev
359 * (see P. Sarolahti, A. Kuznetsov,"Congestion Control
360 * in Linux TCP", USENIX 2002, pp. 49-62).
361 */
362 if (m > 0)
363 m >>= 3;
364 } else {
365 m -= (hc->tx_mdev >> 2);
366 }
367 hc->tx_mdev += m;
368
369 if (hc->tx_mdev > hc->tx_mdev_max) {
370 hc->tx_mdev_max = hc->tx_mdev;
371 if (hc->tx_mdev_max > hc->tx_rttvar)
372 hc->tx_rttvar = hc->tx_mdev_max;
373 }
374
375 /*
376 * Decay RTTVAR at most once per flight, exploiting that
377 * 1) pipe <= cwnd <= Sequence_Window = W (RFC 4340, 7.5.2)
378 * 2) AWL = GSS-W+1 <= GAR <= GSS (RFC 4340, 7.5.1)
379 * GAR is a useful bound for FlightSize = pipe.
380 * AWL is probably too low here, as it over-estimates pipe.
381 */
382 if (after48(dccp_sk(sk)->dccps_gar, hc->tx_rtt_seq)) {
383 if (hc->tx_mdev_max < hc->tx_rttvar)
384 hc->tx_rttvar -= (hc->tx_rttvar -
385 hc->tx_mdev_max) >> 2;
386 hc->tx_rtt_seq = dccp_sk(sk)->dccps_gss;
387 hc->tx_mdev_max = tcp_rto_min(sk);
388 }
389 }
390
391 /*
392 * Set RTO from SRTT and RTTVAR
393 * As in TCP, 4 * RTTVAR >= TCP_RTO_MIN, giving a minimum RTO of 200 ms.
394 * This agrees with RFC 4341, 5:
395 * "Because DCCP does not retransmit data, DCCP does not require
396 * TCP's recommended minimum timeout of one second".
397 */
398 hc->tx_rto = (hc->tx_srtt >> 3) + hc->tx_rttvar;
399
400 if (hc->tx_rto > DCCP_RTO_MAX)
401 hc->tx_rto = DCCP_RTO_MAX;
402}
403
404static void ccid2_new_ack(struct sock *sk, struct ccid2_seq *seqp,
405 unsigned int *maxincr)
406{
407 struct ccid2_hc_tx_sock *hc = ccid2_hc_tx_sk(sk);
408
409 if (hc->tx_cwnd < hc->tx_ssthresh) {
410 if (*maxincr > 0 && ++hc->tx_packets_acked == 2) {
411 hc->tx_cwnd += 1;
412 *maxincr -= 1;
413 hc->tx_packets_acked = 0;
414 }
415 } else if (++hc->tx_packets_acked >= hc->tx_cwnd) {
416 hc->tx_cwnd += 1;
417 hc->tx_packets_acked = 0;
418 }
419 /*
420 * FIXME: RTT is sampled several times per acknowledgment (for each
421 * entry in the Ack Vector), instead of once per Ack (as in TCP SACK).
422 * This causes the RTT to be over-estimated, since the older entries
423 * in the Ack Vector have earlier sending times.
424 * The cleanest solution is to not use the ccid2s_sent field at all
425 * and instead use DCCP timestamps: requires changes in other places.
426 */
427 ccid2_rtt_estimator(sk, ccid2_time_stamp - seqp->ccid2s_sent);
428}
429
430static void ccid2_congestion_event(struct sock *sk, struct ccid2_seq *seqp)
431{
432 struct ccid2_hc_tx_sock *hc = ccid2_hc_tx_sk(sk);
433
434 if ((s32)(seqp->ccid2s_sent - hc->tx_last_cong) < 0) {
435 ccid2_pr_debug("Multiple losses in an RTT---treating as one\n");
436 return;
437 }
438
439 hc->tx_last_cong = ccid2_time_stamp;
440
441 hc->tx_cwnd = hc->tx_cwnd / 2 ? : 1U;
442 hc->tx_ssthresh = max(hc->tx_cwnd, 2U);
443
444 /* Avoid spurious timeouts resulting from Ack Ratio > cwnd */
445 if (dccp_sk(sk)->dccps_l_ack_ratio > hc->tx_cwnd)
446 ccid2_change_l_ack_ratio(sk, hc->tx_cwnd);
447}
448
449static int ccid2_hc_tx_parse_options(struct sock *sk, u8 packet_type,
450 u8 option, u8 *optval, u8 optlen)
451{
452 struct ccid2_hc_tx_sock *hc = ccid2_hc_tx_sk(sk);
453
454 switch (option) {
455 case DCCPO_ACK_VECTOR_0:
456 case DCCPO_ACK_VECTOR_1:
457 return dccp_ackvec_parsed_add(&hc->tx_av_chunks, optval, optlen,
458 option - DCCPO_ACK_VECTOR_0);
459 }
460 return 0;
461}
462
463static void ccid2_hc_tx_packet_recv(struct sock *sk, struct sk_buff *skb)
464{
465 struct dccp_sock *dp = dccp_sk(sk);
466 struct ccid2_hc_tx_sock *hc = ccid2_hc_tx_sk(sk);
467 const bool sender_was_blocked = ccid2_cwnd_network_limited(hc);
468 struct dccp_ackvec_parsed *avp;
469 u64 ackno, seqno;
470 struct ccid2_seq *seqp;
471 int done = 0;
472 unsigned int maxincr = 0;
473
474 /* check reverse path congestion */
475 seqno = DCCP_SKB_CB(skb)->dccpd_seq;
476
477 /* XXX this whole "algorithm" is broken. Need to fix it to keep track
478 * of the seqnos of the dupacks so that rpseq and rpdupack are correct
479 * -sorbo.
480 */
481 /* need to bootstrap */
482 if (hc->tx_rpdupack == -1) {
483 hc->tx_rpdupack = 0;
484 hc->tx_rpseq = seqno;
485 } else {
486 /* check if packet is consecutive */
487 if (dccp_delta_seqno(hc->tx_rpseq, seqno) == 1)
488 hc->tx_rpseq = seqno;
489 /* it's a later packet */
490 else if (after48(seqno, hc->tx_rpseq)) {
491 hc->tx_rpdupack++;
492
493 /* check if we got enough dupacks */
494 if (hc->tx_rpdupack >= NUMDUPACK) {
495 hc->tx_rpdupack = -1; /* XXX lame */
496 hc->tx_rpseq = 0;
497
498 ccid2_change_l_ack_ratio(sk, 2 * dp->dccps_l_ack_ratio);
499 }
500 }
501 }
502
503 /* check forward path congestion */
504 if (dccp_packet_without_ack(skb))
505 return;
506
507 /* still didn't send out new data packets */
508 if (hc->tx_seqh == hc->tx_seqt)
509 goto done;
510
511 ackno = DCCP_SKB_CB(skb)->dccpd_ack_seq;
512 if (after48(ackno, hc->tx_high_ack))
513 hc->tx_high_ack = ackno;
514
515 seqp = hc->tx_seqt;
516 while (before48(seqp->ccid2s_seq, ackno)) {
517 seqp = seqp->ccid2s_next;
518 if (seqp == hc->tx_seqh) {
519 seqp = hc->tx_seqh->ccid2s_prev;
520 break;
521 }
522 }
523
524 /*
525 * In slow-start, cwnd can increase up to a maximum of Ack Ratio/2
526 * packets per acknowledgement. Rounding up avoids that cwnd is not
527 * advanced when Ack Ratio is 1 and gives a slight edge otherwise.
528 */
529 if (hc->tx_cwnd < hc->tx_ssthresh)
530 maxincr = DIV_ROUND_UP(dp->dccps_l_ack_ratio, 2);
531
532 /* go through all ack vectors */
533 list_for_each_entry(avp, &hc->tx_av_chunks, node) {
534 /* go through this ack vector */
535 for (; avp->len--; avp->vec++) {
536 u64 ackno_end_rl = SUB48(ackno,
537 dccp_ackvec_runlen(avp->vec));
538
539 ccid2_pr_debug("ackvec %llu |%u,%u|\n",
540 (unsigned long long)ackno,
541 dccp_ackvec_state(avp->vec) >> 6,
542 dccp_ackvec_runlen(avp->vec));
543 /* if the seqno we are analyzing is larger than the
544 * current ackno, then move towards the tail of our
545 * seqnos.
546 */
547 while (after48(seqp->ccid2s_seq, ackno)) {
548 if (seqp == hc->tx_seqt) {
549 done = 1;
550 break;
551 }
552 seqp = seqp->ccid2s_prev;
553 }
554 if (done)
555 break;
556
557 /* check all seqnos in the range of the vector
558 * run length
559 */
560 while (between48(seqp->ccid2s_seq,ackno_end_rl,ackno)) {
561 const u8 state = dccp_ackvec_state(avp->vec);
562
563 /* new packet received or marked */
564 if (state != DCCPAV_NOT_RECEIVED &&
565 !seqp->ccid2s_acked) {
566 if (state == DCCPAV_ECN_MARKED)
567 ccid2_congestion_event(sk,
568 seqp);
569 else
570 ccid2_new_ack(sk, seqp,
571 &maxincr);
572
573 seqp->ccid2s_acked = 1;
574 ccid2_pr_debug("Got ack for %llu\n",
575 (unsigned long long)seqp->ccid2s_seq);
576 hc->tx_pipe--;
577 }
578 if (seqp == hc->tx_seqt) {
579 done = 1;
580 break;
581 }
582 seqp = seqp->ccid2s_prev;
583 }
584 if (done)
585 break;
586
587 ackno = SUB48(ackno_end_rl, 1);
588 }
589 if (done)
590 break;
591 }
592
593 /* The state about what is acked should be correct now
594 * Check for NUMDUPACK
595 */
596 seqp = hc->tx_seqt;
597 while (before48(seqp->ccid2s_seq, hc->tx_high_ack)) {
598 seqp = seqp->ccid2s_next;
599 if (seqp == hc->tx_seqh) {
600 seqp = hc->tx_seqh->ccid2s_prev;
601 break;
602 }
603 }
604 done = 0;
605 while (1) {
606 if (seqp->ccid2s_acked) {
607 done++;
608 if (done == NUMDUPACK)
609 break;
610 }
611 if (seqp == hc->tx_seqt)
612 break;
613 seqp = seqp->ccid2s_prev;
614 }
615
616 /* If there are at least 3 acknowledgements, anything unacknowledged
617 * below the last sequence number is considered lost
618 */
619 if (done == NUMDUPACK) {
620 struct ccid2_seq *last_acked = seqp;
621
622 /* check for lost packets */
623 while (1) {
624 if (!seqp->ccid2s_acked) {
625 ccid2_pr_debug("Packet lost: %llu\n",
626 (unsigned long long)seqp->ccid2s_seq);
627 /* XXX need to traverse from tail -> head in
628 * order to detect multiple congestion events in
629 * one ack vector.
630 */
631 ccid2_congestion_event(sk, seqp);
632 hc->tx_pipe--;
633 }
634 if (seqp == hc->tx_seqt)
635 break;
636 seqp = seqp->ccid2s_prev;
637 }
638
639 hc->tx_seqt = last_acked;
640 }
641
642 /* trim acked packets in tail */
643 while (hc->tx_seqt != hc->tx_seqh) {
644 if (!hc->tx_seqt->ccid2s_acked)
645 break;
646
647 hc->tx_seqt = hc->tx_seqt->ccid2s_next;
648 }
649
650 /* restart RTO timer if not all outstanding data has been acked */
651 if (hc->tx_pipe == 0)
652 sk_stop_timer(sk, &hc->tx_rtotimer);
653 else
654 sk_reset_timer(sk, &hc->tx_rtotimer, jiffies + hc->tx_rto);
655done:
656 /* check if incoming Acks allow pending packets to be sent */
657 if (sender_was_blocked && !ccid2_cwnd_network_limited(hc))
658 tasklet_schedule(&dccp_sk(sk)->dccps_xmitlet);
659 dccp_ackvec_parsed_cleanup(&hc->tx_av_chunks);
660}
661
662static int ccid2_hc_tx_init(struct ccid *ccid, struct sock *sk)
663{
664 struct ccid2_hc_tx_sock *hc = ccid_priv(ccid);
665 struct dccp_sock *dp = dccp_sk(sk);
666 u32 max_ratio;
667
668 /* RFC 4341, 5: initialise ssthresh to arbitrarily high (max) value */
669 hc->tx_ssthresh = ~0U;
670
671 /* Use larger initial windows (RFC 4341, section 5). */
672 hc->tx_cwnd = rfc3390_bytes_to_packets(dp->dccps_mss_cache);
673 hc->tx_expected_wnd = hc->tx_cwnd;
674
675 /* Make sure that Ack Ratio is enabled and within bounds. */
676 max_ratio = DIV_ROUND_UP(hc->tx_cwnd, 2);
677 if (dp->dccps_l_ack_ratio == 0 || dp->dccps_l_ack_ratio > max_ratio)
678 dp->dccps_l_ack_ratio = max_ratio;
679
680 /* XXX init ~ to window size... */
681 if (ccid2_hc_tx_alloc_seq(hc))
682 return -ENOMEM;
683
684 hc->tx_rto = DCCP_TIMEOUT_INIT;
685 hc->tx_rpdupack = -1;
686 hc->tx_last_cong = hc->tx_lsndtime = hc->tx_cwnd_stamp = ccid2_time_stamp;
687 hc->tx_cwnd_used = 0;
688 setup_timer(&hc->tx_rtotimer, ccid2_hc_tx_rto_expire,
689 (unsigned long)sk);
690 INIT_LIST_HEAD(&hc->tx_av_chunks);
691 return 0;
692}
693
694static void ccid2_hc_tx_exit(struct sock *sk)
695{
696 struct ccid2_hc_tx_sock *hc = ccid2_hc_tx_sk(sk);
697 int i;
698
699 sk_stop_timer(sk, &hc->tx_rtotimer);
700
701 for (i = 0; i < hc->tx_seqbufc; i++)
702 kfree(hc->tx_seqbuf[i]);
703 hc->tx_seqbufc = 0;
704}
705
706static void ccid2_hc_rx_packet_recv(struct sock *sk, struct sk_buff *skb)
707{
708 struct ccid2_hc_rx_sock *hc = ccid2_hc_rx_sk(sk);
709
710 if (!dccp_data_packet(skb))
711 return;
712
713 if (++hc->rx_num_data_pkts >= dccp_sk(sk)->dccps_r_ack_ratio) {
714 dccp_send_ack(sk);
715 hc->rx_num_data_pkts = 0;
716 }
717}
718
719struct ccid_operations ccid2_ops = {
720 .ccid_id = DCCPC_CCID2,
721 .ccid_name = "TCP-like",
722 .ccid_hc_tx_obj_size = sizeof(struct ccid2_hc_tx_sock),
723 .ccid_hc_tx_init = ccid2_hc_tx_init,
724 .ccid_hc_tx_exit = ccid2_hc_tx_exit,
725 .ccid_hc_tx_send_packet = ccid2_hc_tx_send_packet,
726 .ccid_hc_tx_packet_sent = ccid2_hc_tx_packet_sent,
727 .ccid_hc_tx_parse_options = ccid2_hc_tx_parse_options,
728 .ccid_hc_tx_packet_recv = ccid2_hc_tx_packet_recv,
729 .ccid_hc_rx_obj_size = sizeof(struct ccid2_hc_rx_sock),
730 .ccid_hc_rx_packet_recv = ccid2_hc_rx_packet_recv,
731};
732
733#ifdef CONFIG_IP_DCCP_CCID2_DEBUG
734module_param(ccid2_debug, bool, 0644);
735MODULE_PARM_DESC(ccid2_debug, "Enable CCID-2 debug messages");
736#endif
1/*
2 * Copyright (c) 2005, 2006 Andrea Bittau <a.bittau@cs.ucl.ac.uk>
3 *
4 * Changes to meet Linux coding standards, and DCCP infrastructure fixes.
5 *
6 * Copyright (c) 2006 Arnaldo Carvalho de Melo <acme@conectiva.com.br>
7 *
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License as published by
10 * the Free Software Foundation; either version 2 of the License, or
11 * (at your option) any later version.
12 *
13 * This program is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 * GNU General Public License for more details.
17 *
18 * You should have received a copy of the GNU General Public License
19 * along with this program; if not, write to the Free Software
20 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
21 */
22
23/*
24 * This implementation should follow RFC 4341
25 */
26#include <linux/slab.h>
27#include "../feat.h"
28#include "ccid2.h"
29
30
31#ifdef CONFIG_IP_DCCP_CCID2_DEBUG
32static bool ccid2_debug;
33#define ccid2_pr_debug(format, a...) DCCP_PR_DEBUG(ccid2_debug, format, ##a)
34#else
35#define ccid2_pr_debug(format, a...)
36#endif
37
38static int ccid2_hc_tx_alloc_seq(struct ccid2_hc_tx_sock *hc)
39{
40 struct ccid2_seq *seqp;
41 int i;
42
43 /* check if we have space to preserve the pointer to the buffer */
44 if (hc->tx_seqbufc >= (sizeof(hc->tx_seqbuf) /
45 sizeof(struct ccid2_seq *)))
46 return -ENOMEM;
47
48 /* allocate buffer and initialize linked list */
49 seqp = kmalloc(CCID2_SEQBUF_LEN * sizeof(struct ccid2_seq), gfp_any());
50 if (seqp == NULL)
51 return -ENOMEM;
52
53 for (i = 0; i < (CCID2_SEQBUF_LEN - 1); i++) {
54 seqp[i].ccid2s_next = &seqp[i + 1];
55 seqp[i + 1].ccid2s_prev = &seqp[i];
56 }
57 seqp[CCID2_SEQBUF_LEN - 1].ccid2s_next = seqp;
58 seqp->ccid2s_prev = &seqp[CCID2_SEQBUF_LEN - 1];
59
60 /* This is the first allocation. Initiate the head and tail. */
61 if (hc->tx_seqbufc == 0)
62 hc->tx_seqh = hc->tx_seqt = seqp;
63 else {
64 /* link the existing list with the one we just created */
65 hc->tx_seqh->ccid2s_next = seqp;
66 seqp->ccid2s_prev = hc->tx_seqh;
67
68 hc->tx_seqt->ccid2s_prev = &seqp[CCID2_SEQBUF_LEN - 1];
69 seqp[CCID2_SEQBUF_LEN - 1].ccid2s_next = hc->tx_seqt;
70 }
71
72 /* store the original pointer to the buffer so we can free it */
73 hc->tx_seqbuf[hc->tx_seqbufc] = seqp;
74 hc->tx_seqbufc++;
75
76 return 0;
77}
78
79static int ccid2_hc_tx_send_packet(struct sock *sk, struct sk_buff *skb)
80{
81 if (ccid2_cwnd_network_limited(ccid2_hc_tx_sk(sk)))
82 return CCID_PACKET_WILL_DEQUEUE_LATER;
83 return CCID_PACKET_SEND_AT_ONCE;
84}
85
86static void ccid2_change_l_ack_ratio(struct sock *sk, u32 val)
87{
88 u32 max_ratio = DIV_ROUND_UP(ccid2_hc_tx_sk(sk)->tx_cwnd, 2);
89
90 /*
91 * Ensure that Ack Ratio does not exceed ceil(cwnd/2), which is (2) from
92 * RFC 4341, 6.1.2. We ignore the statement that Ack Ratio 2 is always
93 * acceptable since this causes starvation/deadlock whenever cwnd < 2.
94 * The same problem arises when Ack Ratio is 0 (ie. Ack Ratio disabled).
95 */
96 if (val == 0 || val > max_ratio) {
97 DCCP_WARN("Limiting Ack Ratio (%u) to %u\n", val, max_ratio);
98 val = max_ratio;
99 }
100 dccp_feat_signal_nn_change(sk, DCCPF_ACK_RATIO,
101 min_t(u32, val, DCCPF_ACK_RATIO_MAX));
102}
103
104static void ccid2_check_l_ack_ratio(struct sock *sk)
105{
106 struct ccid2_hc_tx_sock *hc = ccid2_hc_tx_sk(sk);
107
108 /*
109 * After a loss, idle period, application limited period, or RTO we
110 * need to check that the ack ratio is still less than the congestion
111 * window. Otherwise, we will send an entire congestion window of
112 * packets and got no response because we haven't sent ack ratio
113 * packets yet.
114 * If the ack ratio does need to be reduced, we reduce it to half of
115 * the congestion window (or 1 if that's zero) instead of to the
116 * congestion window. This prevents problems if one ack is lost.
117 */
118 if (dccp_feat_nn_get(sk, DCCPF_ACK_RATIO) > hc->tx_cwnd)
119 ccid2_change_l_ack_ratio(sk, hc->tx_cwnd/2 ? : 1U);
120}
121
122static void ccid2_change_l_seq_window(struct sock *sk, u64 val)
123{
124 dccp_feat_signal_nn_change(sk, DCCPF_SEQUENCE_WINDOW,
125 clamp_val(val, DCCPF_SEQ_WMIN,
126 DCCPF_SEQ_WMAX));
127}
128
129static void dccp_tasklet_schedule(struct sock *sk)
130{
131 struct tasklet_struct *t = &dccp_sk(sk)->dccps_xmitlet;
132
133 if (!test_and_set_bit(TASKLET_STATE_SCHED, &t->state)) {
134 sock_hold(sk);
135 __tasklet_schedule(t);
136 }
137}
138
139static void ccid2_hc_tx_rto_expire(struct timer_list *t)
140{
141 struct ccid2_hc_tx_sock *hc = from_timer(hc, t, tx_rtotimer);
142 struct sock *sk = hc->sk;
143 const bool sender_was_blocked = ccid2_cwnd_network_limited(hc);
144
145 bh_lock_sock(sk);
146 if (sock_owned_by_user(sk)) {
147 sk_reset_timer(sk, &hc->tx_rtotimer, jiffies + HZ / 5);
148 goto out;
149 }
150
151 ccid2_pr_debug("RTO_EXPIRE\n");
152
153 if (sk->sk_state == DCCP_CLOSED)
154 goto out;
155
156 /* back-off timer */
157 hc->tx_rto <<= 1;
158 if (hc->tx_rto > DCCP_RTO_MAX)
159 hc->tx_rto = DCCP_RTO_MAX;
160
161 /* adjust pipe, cwnd etc */
162 hc->tx_ssthresh = hc->tx_cwnd / 2;
163 if (hc->tx_ssthresh < 2)
164 hc->tx_ssthresh = 2;
165 hc->tx_cwnd = 1;
166 hc->tx_pipe = 0;
167
168 /* clear state about stuff we sent */
169 hc->tx_seqt = hc->tx_seqh;
170 hc->tx_packets_acked = 0;
171
172 /* clear ack ratio state. */
173 hc->tx_rpseq = 0;
174 hc->tx_rpdupack = -1;
175 ccid2_change_l_ack_ratio(sk, 1);
176
177 /* if we were blocked before, we may now send cwnd=1 packet */
178 if (sender_was_blocked)
179 dccp_tasklet_schedule(sk);
180 /* restart backed-off timer */
181 sk_reset_timer(sk, &hc->tx_rtotimer, jiffies + hc->tx_rto);
182out:
183 bh_unlock_sock(sk);
184 sock_put(sk);
185}
186
187/*
188 * Congestion window validation (RFC 2861).
189 */
190static bool ccid2_do_cwv = true;
191module_param(ccid2_do_cwv, bool, 0644);
192MODULE_PARM_DESC(ccid2_do_cwv, "Perform RFC2861 Congestion Window Validation");
193
194/**
195 * ccid2_update_used_window - Track how much of cwnd is actually used
196 * This is done in addition to CWV. The sender needs to have an idea of how many
197 * packets may be in flight, to set the local Sequence Window value accordingly
198 * (RFC 4340, 7.5.2). The CWV mechanism is exploited to keep track of the
199 * maximum-used window. We use an EWMA low-pass filter to filter out noise.
200 */
201static void ccid2_update_used_window(struct ccid2_hc_tx_sock *hc, u32 new_wnd)
202{
203 hc->tx_expected_wnd = (3 * hc->tx_expected_wnd + new_wnd) / 4;
204}
205
206/* This borrows the code of tcp_cwnd_application_limited() */
207static void ccid2_cwnd_application_limited(struct sock *sk, const u32 now)
208{
209 struct ccid2_hc_tx_sock *hc = ccid2_hc_tx_sk(sk);
210 /* don't reduce cwnd below the initial window (IW) */
211 u32 init_win = rfc3390_bytes_to_packets(dccp_sk(sk)->dccps_mss_cache),
212 win_used = max(hc->tx_cwnd_used, init_win);
213
214 if (win_used < hc->tx_cwnd) {
215 hc->tx_ssthresh = max(hc->tx_ssthresh,
216 (hc->tx_cwnd >> 1) + (hc->tx_cwnd >> 2));
217 hc->tx_cwnd = (hc->tx_cwnd + win_used) >> 1;
218 }
219 hc->tx_cwnd_used = 0;
220 hc->tx_cwnd_stamp = now;
221
222 ccid2_check_l_ack_ratio(sk);
223}
224
225/* This borrows the code of tcp_cwnd_restart() */
226static void ccid2_cwnd_restart(struct sock *sk, const u32 now)
227{
228 struct ccid2_hc_tx_sock *hc = ccid2_hc_tx_sk(sk);
229 u32 cwnd = hc->tx_cwnd, restart_cwnd,
230 iwnd = rfc3390_bytes_to_packets(dccp_sk(sk)->dccps_mss_cache);
231
232 hc->tx_ssthresh = max(hc->tx_ssthresh, (cwnd >> 1) + (cwnd >> 2));
233
234 /* don't reduce cwnd below the initial window (IW) */
235 restart_cwnd = min(cwnd, iwnd);
236 cwnd >>= (now - hc->tx_lsndtime) / hc->tx_rto;
237 hc->tx_cwnd = max(cwnd, restart_cwnd);
238
239 hc->tx_cwnd_stamp = now;
240 hc->tx_cwnd_used = 0;
241
242 ccid2_check_l_ack_ratio(sk);
243}
244
245static void ccid2_hc_tx_packet_sent(struct sock *sk, unsigned int len)
246{
247 struct dccp_sock *dp = dccp_sk(sk);
248 struct ccid2_hc_tx_sock *hc = ccid2_hc_tx_sk(sk);
249 const u32 now = ccid2_jiffies32;
250 struct ccid2_seq *next;
251
252 /* slow-start after idle periods (RFC 2581, RFC 2861) */
253 if (ccid2_do_cwv && !hc->tx_pipe &&
254 (s32)(now - hc->tx_lsndtime) >= hc->tx_rto)
255 ccid2_cwnd_restart(sk, now);
256
257 hc->tx_lsndtime = now;
258 hc->tx_pipe += 1;
259
260 /* see whether cwnd was fully used (RFC 2861), update expected window */
261 if (ccid2_cwnd_network_limited(hc)) {
262 ccid2_update_used_window(hc, hc->tx_cwnd);
263 hc->tx_cwnd_used = 0;
264 hc->tx_cwnd_stamp = now;
265 } else {
266 if (hc->tx_pipe > hc->tx_cwnd_used)
267 hc->tx_cwnd_used = hc->tx_pipe;
268
269 ccid2_update_used_window(hc, hc->tx_cwnd_used);
270
271 if (ccid2_do_cwv && (s32)(now - hc->tx_cwnd_stamp) >= hc->tx_rto)
272 ccid2_cwnd_application_limited(sk, now);
273 }
274
275 hc->tx_seqh->ccid2s_seq = dp->dccps_gss;
276 hc->tx_seqh->ccid2s_acked = 0;
277 hc->tx_seqh->ccid2s_sent = now;
278
279 next = hc->tx_seqh->ccid2s_next;
280 /* check if we need to alloc more space */
281 if (next == hc->tx_seqt) {
282 if (ccid2_hc_tx_alloc_seq(hc)) {
283 DCCP_CRIT("packet history - out of memory!");
284 /* FIXME: find a more graceful way to bail out */
285 return;
286 }
287 next = hc->tx_seqh->ccid2s_next;
288 BUG_ON(next == hc->tx_seqt);
289 }
290 hc->tx_seqh = next;
291
292 ccid2_pr_debug("cwnd=%d pipe=%d\n", hc->tx_cwnd, hc->tx_pipe);
293
294 /*
295 * FIXME: The code below is broken and the variables have been removed
296 * from the socket struct. The `ackloss' variable was always set to 0,
297 * and with arsent there are several problems:
298 * (i) it doesn't just count the number of Acks, but all sent packets;
299 * (ii) it is expressed in # of packets, not # of windows, so the
300 * comparison below uses the wrong formula: Appendix A of RFC 4341
301 * comes up with the number K = cwnd / (R^2 - R) of consecutive windows
302 * of data with no lost or marked Ack packets. If arsent were the # of
303 * consecutive Acks received without loss, then Ack Ratio needs to be
304 * decreased by 1 when
305 * arsent >= K * cwnd / R = cwnd^2 / (R^3 - R^2)
306 * where cwnd / R is the number of Acks received per window of data
307 * (cf. RFC 4341, App. A). The problems are that
308 * - arsent counts other packets as well;
309 * - the comparison uses a formula different from RFC 4341;
310 * - computing a cubic/quadratic equation each time is too complicated.
311 * Hence a different algorithm is needed.
312 */
313#if 0
314 /* Ack Ratio. Need to maintain a concept of how many windows we sent */
315 hc->tx_arsent++;
316 /* We had an ack loss in this window... */
317 if (hc->tx_ackloss) {
318 if (hc->tx_arsent >= hc->tx_cwnd) {
319 hc->tx_arsent = 0;
320 hc->tx_ackloss = 0;
321 }
322 } else {
323 /* No acks lost up to now... */
324 /* decrease ack ratio if enough packets were sent */
325 if (dp->dccps_l_ack_ratio > 1) {
326 /* XXX don't calculate denominator each time */
327 int denom = dp->dccps_l_ack_ratio * dp->dccps_l_ack_ratio -
328 dp->dccps_l_ack_ratio;
329
330 denom = hc->tx_cwnd * hc->tx_cwnd / denom;
331
332 if (hc->tx_arsent >= denom) {
333 ccid2_change_l_ack_ratio(sk, dp->dccps_l_ack_ratio - 1);
334 hc->tx_arsent = 0;
335 }
336 } else {
337 /* we can't increase ack ratio further [1] */
338 hc->tx_arsent = 0; /* or maybe set it to cwnd*/
339 }
340 }
341#endif
342
343 sk_reset_timer(sk, &hc->tx_rtotimer, jiffies + hc->tx_rto);
344
345#ifdef CONFIG_IP_DCCP_CCID2_DEBUG
346 do {
347 struct ccid2_seq *seqp = hc->tx_seqt;
348
349 while (seqp != hc->tx_seqh) {
350 ccid2_pr_debug("out seq=%llu acked=%d time=%u\n",
351 (unsigned long long)seqp->ccid2s_seq,
352 seqp->ccid2s_acked, seqp->ccid2s_sent);
353 seqp = seqp->ccid2s_next;
354 }
355 } while (0);
356 ccid2_pr_debug("=========\n");
357#endif
358}
359
360/**
361 * ccid2_rtt_estimator - Sample RTT and compute RTO using RFC2988 algorithm
362 * This code is almost identical with TCP's tcp_rtt_estimator(), since
363 * - it has a higher sampling frequency (recommended by RFC 1323),
364 * - the RTO does not collapse into RTT due to RTTVAR going towards zero,
365 * - it is simple (cf. more complex proposals such as Eifel timer or research
366 * which suggests that the gain should be set according to window size),
367 * - in tests it was found to work well with CCID2 [gerrit].
368 */
369static void ccid2_rtt_estimator(struct sock *sk, const long mrtt)
370{
371 struct ccid2_hc_tx_sock *hc = ccid2_hc_tx_sk(sk);
372 long m = mrtt ? : 1;
373
374 if (hc->tx_srtt == 0) {
375 /* First measurement m */
376 hc->tx_srtt = m << 3;
377 hc->tx_mdev = m << 1;
378
379 hc->tx_mdev_max = max(hc->tx_mdev, tcp_rto_min(sk));
380 hc->tx_rttvar = hc->tx_mdev_max;
381
382 hc->tx_rtt_seq = dccp_sk(sk)->dccps_gss;
383 } else {
384 /* Update scaled SRTT as SRTT += 1/8 * (m - SRTT) */
385 m -= (hc->tx_srtt >> 3);
386 hc->tx_srtt += m;
387
388 /* Similarly, update scaled mdev with regard to |m| */
389 if (m < 0) {
390 m = -m;
391 m -= (hc->tx_mdev >> 2);
392 /*
393 * This neutralises RTO increase when RTT < SRTT - mdev
394 * (see P. Sarolahti, A. Kuznetsov,"Congestion Control
395 * in Linux TCP", USENIX 2002, pp. 49-62).
396 */
397 if (m > 0)
398 m >>= 3;
399 } else {
400 m -= (hc->tx_mdev >> 2);
401 }
402 hc->tx_mdev += m;
403
404 if (hc->tx_mdev > hc->tx_mdev_max) {
405 hc->tx_mdev_max = hc->tx_mdev;
406 if (hc->tx_mdev_max > hc->tx_rttvar)
407 hc->tx_rttvar = hc->tx_mdev_max;
408 }
409
410 /*
411 * Decay RTTVAR at most once per flight, exploiting that
412 * 1) pipe <= cwnd <= Sequence_Window = W (RFC 4340, 7.5.2)
413 * 2) AWL = GSS-W+1 <= GAR <= GSS (RFC 4340, 7.5.1)
414 * GAR is a useful bound for FlightSize = pipe.
415 * AWL is probably too low here, as it over-estimates pipe.
416 */
417 if (after48(dccp_sk(sk)->dccps_gar, hc->tx_rtt_seq)) {
418 if (hc->tx_mdev_max < hc->tx_rttvar)
419 hc->tx_rttvar -= (hc->tx_rttvar -
420 hc->tx_mdev_max) >> 2;
421 hc->tx_rtt_seq = dccp_sk(sk)->dccps_gss;
422 hc->tx_mdev_max = tcp_rto_min(sk);
423 }
424 }
425
426 /*
427 * Set RTO from SRTT and RTTVAR
428 * As in TCP, 4 * RTTVAR >= TCP_RTO_MIN, giving a minimum RTO of 200 ms.
429 * This agrees with RFC 4341, 5:
430 * "Because DCCP does not retransmit data, DCCP does not require
431 * TCP's recommended minimum timeout of one second".
432 */
433 hc->tx_rto = (hc->tx_srtt >> 3) + hc->tx_rttvar;
434
435 if (hc->tx_rto > DCCP_RTO_MAX)
436 hc->tx_rto = DCCP_RTO_MAX;
437}
438
439static void ccid2_new_ack(struct sock *sk, struct ccid2_seq *seqp,
440 unsigned int *maxincr)
441{
442 struct ccid2_hc_tx_sock *hc = ccid2_hc_tx_sk(sk);
443 struct dccp_sock *dp = dccp_sk(sk);
444 int r_seq_used = hc->tx_cwnd / dp->dccps_l_ack_ratio;
445
446 if (hc->tx_cwnd < dp->dccps_l_seq_win &&
447 r_seq_used < dp->dccps_r_seq_win) {
448 if (hc->tx_cwnd < hc->tx_ssthresh) {
449 if (*maxincr > 0 && ++hc->tx_packets_acked >= 2) {
450 hc->tx_cwnd += 1;
451 *maxincr -= 1;
452 hc->tx_packets_acked = 0;
453 }
454 } else if (++hc->tx_packets_acked >= hc->tx_cwnd) {
455 hc->tx_cwnd += 1;
456 hc->tx_packets_acked = 0;
457 }
458 }
459
460 /*
461 * Adjust the local sequence window and the ack ratio to allow about
462 * 5 times the number of packets in the network (RFC 4340 7.5.2)
463 */
464 if (r_seq_used * CCID2_WIN_CHANGE_FACTOR >= dp->dccps_r_seq_win)
465 ccid2_change_l_ack_ratio(sk, dp->dccps_l_ack_ratio * 2);
466 else if (r_seq_used * CCID2_WIN_CHANGE_FACTOR < dp->dccps_r_seq_win/2)
467 ccid2_change_l_ack_ratio(sk, dp->dccps_l_ack_ratio / 2 ? : 1U);
468
469 if (hc->tx_cwnd * CCID2_WIN_CHANGE_FACTOR >= dp->dccps_l_seq_win)
470 ccid2_change_l_seq_window(sk, dp->dccps_l_seq_win * 2);
471 else if (hc->tx_cwnd * CCID2_WIN_CHANGE_FACTOR < dp->dccps_l_seq_win/2)
472 ccid2_change_l_seq_window(sk, dp->dccps_l_seq_win / 2);
473
474 /*
475 * FIXME: RTT is sampled several times per acknowledgment (for each
476 * entry in the Ack Vector), instead of once per Ack (as in TCP SACK).
477 * This causes the RTT to be over-estimated, since the older entries
478 * in the Ack Vector have earlier sending times.
479 * The cleanest solution is to not use the ccid2s_sent field at all
480 * and instead use DCCP timestamps: requires changes in other places.
481 */
482 ccid2_rtt_estimator(sk, ccid2_jiffies32 - seqp->ccid2s_sent);
483}
484
485static void ccid2_congestion_event(struct sock *sk, struct ccid2_seq *seqp)
486{
487 struct ccid2_hc_tx_sock *hc = ccid2_hc_tx_sk(sk);
488
489 if ((s32)(seqp->ccid2s_sent - hc->tx_last_cong) < 0) {
490 ccid2_pr_debug("Multiple losses in an RTT---treating as one\n");
491 return;
492 }
493
494 hc->tx_last_cong = ccid2_jiffies32;
495
496 hc->tx_cwnd = hc->tx_cwnd / 2 ? : 1U;
497 hc->tx_ssthresh = max(hc->tx_cwnd, 2U);
498
499 ccid2_check_l_ack_ratio(sk);
500}
501
502static int ccid2_hc_tx_parse_options(struct sock *sk, u8 packet_type,
503 u8 option, u8 *optval, u8 optlen)
504{
505 struct ccid2_hc_tx_sock *hc = ccid2_hc_tx_sk(sk);
506
507 switch (option) {
508 case DCCPO_ACK_VECTOR_0:
509 case DCCPO_ACK_VECTOR_1:
510 return dccp_ackvec_parsed_add(&hc->tx_av_chunks, optval, optlen,
511 option - DCCPO_ACK_VECTOR_0);
512 }
513 return 0;
514}
515
516static void ccid2_hc_tx_packet_recv(struct sock *sk, struct sk_buff *skb)
517{
518 struct dccp_sock *dp = dccp_sk(sk);
519 struct ccid2_hc_tx_sock *hc = ccid2_hc_tx_sk(sk);
520 const bool sender_was_blocked = ccid2_cwnd_network_limited(hc);
521 struct dccp_ackvec_parsed *avp;
522 u64 ackno, seqno;
523 struct ccid2_seq *seqp;
524 int done = 0;
525 unsigned int maxincr = 0;
526
527 /* check reverse path congestion */
528 seqno = DCCP_SKB_CB(skb)->dccpd_seq;
529
530 /* XXX this whole "algorithm" is broken. Need to fix it to keep track
531 * of the seqnos of the dupacks so that rpseq and rpdupack are correct
532 * -sorbo.
533 */
534 /* need to bootstrap */
535 if (hc->tx_rpdupack == -1) {
536 hc->tx_rpdupack = 0;
537 hc->tx_rpseq = seqno;
538 } else {
539 /* check if packet is consecutive */
540 if (dccp_delta_seqno(hc->tx_rpseq, seqno) == 1)
541 hc->tx_rpseq = seqno;
542 /* it's a later packet */
543 else if (after48(seqno, hc->tx_rpseq)) {
544 hc->tx_rpdupack++;
545
546 /* check if we got enough dupacks */
547 if (hc->tx_rpdupack >= NUMDUPACK) {
548 hc->tx_rpdupack = -1; /* XXX lame */
549 hc->tx_rpseq = 0;
550#ifdef __CCID2_COPES_GRACEFULLY_WITH_ACK_CONGESTION_CONTROL__
551 /*
552 * FIXME: Ack Congestion Control is broken; in
553 * the current state instabilities occurred with
554 * Ack Ratios greater than 1; causing hang-ups
555 * and long RTO timeouts. This needs to be fixed
556 * before opening up dynamic changes. -- gerrit
557 */
558 ccid2_change_l_ack_ratio(sk, 2 * dp->dccps_l_ack_ratio);
559#endif
560 }
561 }
562 }
563
564 /* check forward path congestion */
565 if (dccp_packet_without_ack(skb))
566 return;
567
568 /* still didn't send out new data packets */
569 if (hc->tx_seqh == hc->tx_seqt)
570 goto done;
571
572 ackno = DCCP_SKB_CB(skb)->dccpd_ack_seq;
573 if (after48(ackno, hc->tx_high_ack))
574 hc->tx_high_ack = ackno;
575
576 seqp = hc->tx_seqt;
577 while (before48(seqp->ccid2s_seq, ackno)) {
578 seqp = seqp->ccid2s_next;
579 if (seqp == hc->tx_seqh) {
580 seqp = hc->tx_seqh->ccid2s_prev;
581 break;
582 }
583 }
584
585 /*
586 * In slow-start, cwnd can increase up to a maximum of Ack Ratio/2
587 * packets per acknowledgement. Rounding up avoids that cwnd is not
588 * advanced when Ack Ratio is 1 and gives a slight edge otherwise.
589 */
590 if (hc->tx_cwnd < hc->tx_ssthresh)
591 maxincr = DIV_ROUND_UP(dp->dccps_l_ack_ratio, 2);
592
593 /* go through all ack vectors */
594 list_for_each_entry(avp, &hc->tx_av_chunks, node) {
595 /* go through this ack vector */
596 for (; avp->len--; avp->vec++) {
597 u64 ackno_end_rl = SUB48(ackno,
598 dccp_ackvec_runlen(avp->vec));
599
600 ccid2_pr_debug("ackvec %llu |%u,%u|\n",
601 (unsigned long long)ackno,
602 dccp_ackvec_state(avp->vec) >> 6,
603 dccp_ackvec_runlen(avp->vec));
604 /* if the seqno we are analyzing is larger than the
605 * current ackno, then move towards the tail of our
606 * seqnos.
607 */
608 while (after48(seqp->ccid2s_seq, ackno)) {
609 if (seqp == hc->tx_seqt) {
610 done = 1;
611 break;
612 }
613 seqp = seqp->ccid2s_prev;
614 }
615 if (done)
616 break;
617
618 /* check all seqnos in the range of the vector
619 * run length
620 */
621 while (between48(seqp->ccid2s_seq,ackno_end_rl,ackno)) {
622 const u8 state = dccp_ackvec_state(avp->vec);
623
624 /* new packet received or marked */
625 if (state != DCCPAV_NOT_RECEIVED &&
626 !seqp->ccid2s_acked) {
627 if (state == DCCPAV_ECN_MARKED)
628 ccid2_congestion_event(sk,
629 seqp);
630 else
631 ccid2_new_ack(sk, seqp,
632 &maxincr);
633
634 seqp->ccid2s_acked = 1;
635 ccid2_pr_debug("Got ack for %llu\n",
636 (unsigned long long)seqp->ccid2s_seq);
637 hc->tx_pipe--;
638 }
639 if (seqp == hc->tx_seqt) {
640 done = 1;
641 break;
642 }
643 seqp = seqp->ccid2s_prev;
644 }
645 if (done)
646 break;
647
648 ackno = SUB48(ackno_end_rl, 1);
649 }
650 if (done)
651 break;
652 }
653
654 /* The state about what is acked should be correct now
655 * Check for NUMDUPACK
656 */
657 seqp = hc->tx_seqt;
658 while (before48(seqp->ccid2s_seq, hc->tx_high_ack)) {
659 seqp = seqp->ccid2s_next;
660 if (seqp == hc->tx_seqh) {
661 seqp = hc->tx_seqh->ccid2s_prev;
662 break;
663 }
664 }
665 done = 0;
666 while (1) {
667 if (seqp->ccid2s_acked) {
668 done++;
669 if (done == NUMDUPACK)
670 break;
671 }
672 if (seqp == hc->tx_seqt)
673 break;
674 seqp = seqp->ccid2s_prev;
675 }
676
677 /* If there are at least 3 acknowledgements, anything unacknowledged
678 * below the last sequence number is considered lost
679 */
680 if (done == NUMDUPACK) {
681 struct ccid2_seq *last_acked = seqp;
682
683 /* check for lost packets */
684 while (1) {
685 if (!seqp->ccid2s_acked) {
686 ccid2_pr_debug("Packet lost: %llu\n",
687 (unsigned long long)seqp->ccid2s_seq);
688 /* XXX need to traverse from tail -> head in
689 * order to detect multiple congestion events in
690 * one ack vector.
691 */
692 ccid2_congestion_event(sk, seqp);
693 hc->tx_pipe--;
694 }
695 if (seqp == hc->tx_seqt)
696 break;
697 seqp = seqp->ccid2s_prev;
698 }
699
700 hc->tx_seqt = last_acked;
701 }
702
703 /* trim acked packets in tail */
704 while (hc->tx_seqt != hc->tx_seqh) {
705 if (!hc->tx_seqt->ccid2s_acked)
706 break;
707
708 hc->tx_seqt = hc->tx_seqt->ccid2s_next;
709 }
710
711 /* restart RTO timer if not all outstanding data has been acked */
712 if (hc->tx_pipe == 0)
713 sk_stop_timer(sk, &hc->tx_rtotimer);
714 else
715 sk_reset_timer(sk, &hc->tx_rtotimer, jiffies + hc->tx_rto);
716done:
717 /* check if incoming Acks allow pending packets to be sent */
718 if (sender_was_blocked && !ccid2_cwnd_network_limited(hc))
719 dccp_tasklet_schedule(sk);
720 dccp_ackvec_parsed_cleanup(&hc->tx_av_chunks);
721}
722
723static int ccid2_hc_tx_init(struct ccid *ccid, struct sock *sk)
724{
725 struct ccid2_hc_tx_sock *hc = ccid_priv(ccid);
726 struct dccp_sock *dp = dccp_sk(sk);
727 u32 max_ratio;
728
729 /* RFC 4341, 5: initialise ssthresh to arbitrarily high (max) value */
730 hc->tx_ssthresh = ~0U;
731
732 /* Use larger initial windows (RFC 4341, section 5). */
733 hc->tx_cwnd = rfc3390_bytes_to_packets(dp->dccps_mss_cache);
734 hc->tx_expected_wnd = hc->tx_cwnd;
735
736 /* Make sure that Ack Ratio is enabled and within bounds. */
737 max_ratio = DIV_ROUND_UP(hc->tx_cwnd, 2);
738 if (dp->dccps_l_ack_ratio == 0 || dp->dccps_l_ack_ratio > max_ratio)
739 dp->dccps_l_ack_ratio = max_ratio;
740
741 /* XXX init ~ to window size... */
742 if (ccid2_hc_tx_alloc_seq(hc))
743 return -ENOMEM;
744
745 hc->tx_rto = DCCP_TIMEOUT_INIT;
746 hc->tx_rpdupack = -1;
747 hc->tx_last_cong = hc->tx_lsndtime = hc->tx_cwnd_stamp = ccid2_jiffies32;
748 hc->tx_cwnd_used = 0;
749 hc->sk = sk;
750 timer_setup(&hc->tx_rtotimer, ccid2_hc_tx_rto_expire, 0);
751 INIT_LIST_HEAD(&hc->tx_av_chunks);
752 return 0;
753}
754
755static void ccid2_hc_tx_exit(struct sock *sk)
756{
757 struct ccid2_hc_tx_sock *hc = ccid2_hc_tx_sk(sk);
758 int i;
759
760 sk_stop_timer(sk, &hc->tx_rtotimer);
761
762 for (i = 0; i < hc->tx_seqbufc; i++)
763 kfree(hc->tx_seqbuf[i]);
764 hc->tx_seqbufc = 0;
765 dccp_ackvec_parsed_cleanup(&hc->tx_av_chunks);
766}
767
768static void ccid2_hc_rx_packet_recv(struct sock *sk, struct sk_buff *skb)
769{
770 struct ccid2_hc_rx_sock *hc = ccid2_hc_rx_sk(sk);
771
772 if (!dccp_data_packet(skb))
773 return;
774
775 if (++hc->rx_num_data_pkts >= dccp_sk(sk)->dccps_r_ack_ratio) {
776 dccp_send_ack(sk);
777 hc->rx_num_data_pkts = 0;
778 }
779}
780
781struct ccid_operations ccid2_ops = {
782 .ccid_id = DCCPC_CCID2,
783 .ccid_name = "TCP-like",
784 .ccid_hc_tx_obj_size = sizeof(struct ccid2_hc_tx_sock),
785 .ccid_hc_tx_init = ccid2_hc_tx_init,
786 .ccid_hc_tx_exit = ccid2_hc_tx_exit,
787 .ccid_hc_tx_send_packet = ccid2_hc_tx_send_packet,
788 .ccid_hc_tx_packet_sent = ccid2_hc_tx_packet_sent,
789 .ccid_hc_tx_parse_options = ccid2_hc_tx_parse_options,
790 .ccid_hc_tx_packet_recv = ccid2_hc_tx_packet_recv,
791 .ccid_hc_rx_obj_size = sizeof(struct ccid2_hc_rx_sock),
792 .ccid_hc_rx_packet_recv = ccid2_hc_rx_packet_recv,
793};
794
795#ifdef CONFIG_IP_DCCP_CCID2_DEBUG
796module_param(ccid2_debug, bool, 0644);
797MODULE_PARM_DESC(ccid2_debug, "Enable CCID-2 debug messages");
798#endif