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// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * Copyright (c) 2005, 2006 Andrea Bittau <a.bittau@cs.ucl.ac.uk>
4 *
5 * Changes to meet Linux coding standards, and DCCP infrastructure fixes.
6 *
7 * Copyright (c) 2006 Arnaldo Carvalho de Melo <acme@conectiva.com.br>
8 */
9
10/*
11 * This implementation should follow RFC 4341
12 */
13#include <linux/slab.h>
14#include "../feat.h"
15#include "ccid2.h"
16
17
18#ifdef CONFIG_IP_DCCP_CCID2_DEBUG
19static bool ccid2_debug;
20#define ccid2_pr_debug(format, a...) DCCP_PR_DEBUG(ccid2_debug, format, ##a)
21#else
22#define ccid2_pr_debug(format, a...)
23#endif
24
25static int ccid2_hc_tx_alloc_seq(struct ccid2_hc_tx_sock *hc)
26{
27 struct ccid2_seq *seqp;
28 int i;
29
30 /* check if we have space to preserve the pointer to the buffer */
31 if (hc->tx_seqbufc >= (sizeof(hc->tx_seqbuf) /
32 sizeof(struct ccid2_seq *)))
33 return -ENOMEM;
34
35 /* allocate buffer and initialize linked list */
36 seqp = kmalloc_array(CCID2_SEQBUF_LEN, sizeof(struct ccid2_seq),
37 gfp_any());
38 if (seqp == NULL)
39 return -ENOMEM;
40
41 for (i = 0; i < (CCID2_SEQBUF_LEN - 1); i++) {
42 seqp[i].ccid2s_next = &seqp[i + 1];
43 seqp[i + 1].ccid2s_prev = &seqp[i];
44 }
45 seqp[CCID2_SEQBUF_LEN - 1].ccid2s_next = seqp;
46 seqp->ccid2s_prev = &seqp[CCID2_SEQBUF_LEN - 1];
47
48 /* This is the first allocation. Initiate the head and tail. */
49 if (hc->tx_seqbufc == 0)
50 hc->tx_seqh = hc->tx_seqt = seqp;
51 else {
52 /* link the existing list with the one we just created */
53 hc->tx_seqh->ccid2s_next = seqp;
54 seqp->ccid2s_prev = hc->tx_seqh;
55
56 hc->tx_seqt->ccid2s_prev = &seqp[CCID2_SEQBUF_LEN - 1];
57 seqp[CCID2_SEQBUF_LEN - 1].ccid2s_next = hc->tx_seqt;
58 }
59
60 /* store the original pointer to the buffer so we can free it */
61 hc->tx_seqbuf[hc->tx_seqbufc] = seqp;
62 hc->tx_seqbufc++;
63
64 return 0;
65}
66
67static int ccid2_hc_tx_send_packet(struct sock *sk, struct sk_buff *skb)
68{
69 if (ccid2_cwnd_network_limited(ccid2_hc_tx_sk(sk)))
70 return CCID_PACKET_WILL_DEQUEUE_LATER;
71 return CCID_PACKET_SEND_AT_ONCE;
72}
73
74static void ccid2_change_l_ack_ratio(struct sock *sk, u32 val)
75{
76 u32 max_ratio = DIV_ROUND_UP(ccid2_hc_tx_sk(sk)->tx_cwnd, 2);
77
78 /*
79 * Ensure that Ack Ratio does not exceed ceil(cwnd/2), which is (2) from
80 * RFC 4341, 6.1.2. We ignore the statement that Ack Ratio 2 is always
81 * acceptable since this causes starvation/deadlock whenever cwnd < 2.
82 * The same problem arises when Ack Ratio is 0 (ie. Ack Ratio disabled).
83 */
84 if (val == 0 || val > max_ratio) {
85 DCCP_WARN("Limiting Ack Ratio (%u) to %u\n", val, max_ratio);
86 val = max_ratio;
87 }
88 dccp_feat_signal_nn_change(sk, DCCPF_ACK_RATIO,
89 min_t(u32, val, DCCPF_ACK_RATIO_MAX));
90}
91
92static void ccid2_check_l_ack_ratio(struct sock *sk)
93{
94 struct ccid2_hc_tx_sock *hc = ccid2_hc_tx_sk(sk);
95
96 /*
97 * After a loss, idle period, application limited period, or RTO we
98 * need to check that the ack ratio is still less than the congestion
99 * window. Otherwise, we will send an entire congestion window of
100 * packets and got no response because we haven't sent ack ratio
101 * packets yet.
102 * If the ack ratio does need to be reduced, we reduce it to half of
103 * the congestion window (or 1 if that's zero) instead of to the
104 * congestion window. This prevents problems if one ack is lost.
105 */
106 if (dccp_feat_nn_get(sk, DCCPF_ACK_RATIO) > hc->tx_cwnd)
107 ccid2_change_l_ack_ratio(sk, hc->tx_cwnd/2 ? : 1U);
108}
109
110static void ccid2_change_l_seq_window(struct sock *sk, u64 val)
111{
112 dccp_feat_signal_nn_change(sk, DCCPF_SEQUENCE_WINDOW,
113 clamp_val(val, DCCPF_SEQ_WMIN,
114 DCCPF_SEQ_WMAX));
115}
116
117static void dccp_tasklet_schedule(struct sock *sk)
118{
119 struct tasklet_struct *t = &dccp_sk(sk)->dccps_xmitlet;
120
121 if (!test_and_set_bit(TASKLET_STATE_SCHED, &t->state)) {
122 sock_hold(sk);
123 __tasklet_schedule(t);
124 }
125}
126
127static void ccid2_hc_tx_rto_expire(struct timer_list *t)
128{
129 struct ccid2_hc_tx_sock *hc = from_timer(hc, t, tx_rtotimer);
130 struct sock *sk = hc->sk;
131 const bool sender_was_blocked = ccid2_cwnd_network_limited(hc);
132
133 bh_lock_sock(sk);
134 if (sock_owned_by_user(sk)) {
135 sk_reset_timer(sk, &hc->tx_rtotimer, jiffies + HZ / 5);
136 goto out;
137 }
138
139 ccid2_pr_debug("RTO_EXPIRE\n");
140
141 if (sk->sk_state == DCCP_CLOSED)
142 goto out;
143
144 /* back-off timer */
145 hc->tx_rto <<= 1;
146 if (hc->tx_rto > DCCP_RTO_MAX)
147 hc->tx_rto = DCCP_RTO_MAX;
148
149 /* adjust pipe, cwnd etc */
150 hc->tx_ssthresh = hc->tx_cwnd / 2;
151 if (hc->tx_ssthresh < 2)
152 hc->tx_ssthresh = 2;
153 hc->tx_cwnd = 1;
154 hc->tx_pipe = 0;
155
156 /* clear state about stuff we sent */
157 hc->tx_seqt = hc->tx_seqh;
158 hc->tx_packets_acked = 0;
159
160 /* clear ack ratio state. */
161 hc->tx_rpseq = 0;
162 hc->tx_rpdupack = -1;
163 ccid2_change_l_ack_ratio(sk, 1);
164
165 /* if we were blocked before, we may now send cwnd=1 packet */
166 if (sender_was_blocked)
167 dccp_tasklet_schedule(sk);
168 /* restart backed-off timer */
169 sk_reset_timer(sk, &hc->tx_rtotimer, jiffies + hc->tx_rto);
170out:
171 bh_unlock_sock(sk);
172 sock_put(sk);
173}
174
175/*
176 * Congestion window validation (RFC 2861).
177 */
178static bool ccid2_do_cwv = true;
179module_param(ccid2_do_cwv, bool, 0644);
180MODULE_PARM_DESC(ccid2_do_cwv, "Perform RFC2861 Congestion Window Validation");
181
182/**
183 * ccid2_update_used_window - Track how much of cwnd is actually used
184 * @hc: socket to update window
185 * @new_wnd: new window values to add into the filter
186 *
187 * This is done in addition to CWV. The sender needs to have an idea of how many
188 * packets may be in flight, to set the local Sequence Window value accordingly
189 * (RFC 4340, 7.5.2). The CWV mechanism is exploited to keep track of the
190 * maximum-used window. We use an EWMA low-pass filter to filter out noise.
191 */
192static void ccid2_update_used_window(struct ccid2_hc_tx_sock *hc, u32 new_wnd)
193{
194 hc->tx_expected_wnd = (3 * hc->tx_expected_wnd + new_wnd) / 4;
195}
196
197/* This borrows the code of tcp_cwnd_application_limited() */
198static void ccid2_cwnd_application_limited(struct sock *sk, const u32 now)
199{
200 struct ccid2_hc_tx_sock *hc = ccid2_hc_tx_sk(sk);
201 /* don't reduce cwnd below the initial window (IW) */
202 u32 init_win = rfc3390_bytes_to_packets(dccp_sk(sk)->dccps_mss_cache),
203 win_used = max(hc->tx_cwnd_used, init_win);
204
205 if (win_used < hc->tx_cwnd) {
206 hc->tx_ssthresh = max(hc->tx_ssthresh,
207 (hc->tx_cwnd >> 1) + (hc->tx_cwnd >> 2));
208 hc->tx_cwnd = (hc->tx_cwnd + win_used) >> 1;
209 }
210 hc->tx_cwnd_used = 0;
211 hc->tx_cwnd_stamp = now;
212
213 ccid2_check_l_ack_ratio(sk);
214}
215
216/* This borrows the code of tcp_cwnd_restart() */
217static void ccid2_cwnd_restart(struct sock *sk, const u32 now)
218{
219 struct ccid2_hc_tx_sock *hc = ccid2_hc_tx_sk(sk);
220 u32 cwnd = hc->tx_cwnd, restart_cwnd,
221 iwnd = rfc3390_bytes_to_packets(dccp_sk(sk)->dccps_mss_cache);
222 s32 delta = now - hc->tx_lsndtime;
223
224 hc->tx_ssthresh = max(hc->tx_ssthresh, (cwnd >> 1) + (cwnd >> 2));
225
226 /* don't reduce cwnd below the initial window (IW) */
227 restart_cwnd = min(cwnd, iwnd);
228
229 while ((delta -= hc->tx_rto) >= 0 && cwnd > restart_cwnd)
230 cwnd >>= 1;
231 hc->tx_cwnd = max(cwnd, restart_cwnd);
232 hc->tx_cwnd_stamp = now;
233 hc->tx_cwnd_used = 0;
234
235 ccid2_check_l_ack_ratio(sk);
236}
237
238static void ccid2_hc_tx_packet_sent(struct sock *sk, unsigned int len)
239{
240 struct dccp_sock *dp = dccp_sk(sk);
241 struct ccid2_hc_tx_sock *hc = ccid2_hc_tx_sk(sk);
242 const u32 now = ccid2_jiffies32;
243 struct ccid2_seq *next;
244
245 /* slow-start after idle periods (RFC 2581, RFC 2861) */
246 if (ccid2_do_cwv && !hc->tx_pipe &&
247 (s32)(now - hc->tx_lsndtime) >= hc->tx_rto)
248 ccid2_cwnd_restart(sk, now);
249
250 hc->tx_lsndtime = now;
251 hc->tx_pipe += 1;
252
253 /* see whether cwnd was fully used (RFC 2861), update expected window */
254 if (ccid2_cwnd_network_limited(hc)) {
255 ccid2_update_used_window(hc, hc->tx_cwnd);
256 hc->tx_cwnd_used = 0;
257 hc->tx_cwnd_stamp = now;
258 } else {
259 if (hc->tx_pipe > hc->tx_cwnd_used)
260 hc->tx_cwnd_used = hc->tx_pipe;
261
262 ccid2_update_used_window(hc, hc->tx_cwnd_used);
263
264 if (ccid2_do_cwv && (s32)(now - hc->tx_cwnd_stamp) >= hc->tx_rto)
265 ccid2_cwnd_application_limited(sk, now);
266 }
267
268 hc->tx_seqh->ccid2s_seq = dp->dccps_gss;
269 hc->tx_seqh->ccid2s_acked = 0;
270 hc->tx_seqh->ccid2s_sent = now;
271
272 next = hc->tx_seqh->ccid2s_next;
273 /* check if we need to alloc more space */
274 if (next == hc->tx_seqt) {
275 if (ccid2_hc_tx_alloc_seq(hc)) {
276 DCCP_CRIT("packet history - out of memory!");
277 /* FIXME: find a more graceful way to bail out */
278 return;
279 }
280 next = hc->tx_seqh->ccid2s_next;
281 BUG_ON(next == hc->tx_seqt);
282 }
283 hc->tx_seqh = next;
284
285 ccid2_pr_debug("cwnd=%d pipe=%d\n", hc->tx_cwnd, hc->tx_pipe);
286
287 /*
288 * FIXME: The code below is broken and the variables have been removed
289 * from the socket struct. The `ackloss' variable was always set to 0,
290 * and with arsent there are several problems:
291 * (i) it doesn't just count the number of Acks, but all sent packets;
292 * (ii) it is expressed in # of packets, not # of windows, so the
293 * comparison below uses the wrong formula: Appendix A of RFC 4341
294 * comes up with the number K = cwnd / (R^2 - R) of consecutive windows
295 * of data with no lost or marked Ack packets. If arsent were the # of
296 * consecutive Acks received without loss, then Ack Ratio needs to be
297 * decreased by 1 when
298 * arsent >= K * cwnd / R = cwnd^2 / (R^3 - R^2)
299 * where cwnd / R is the number of Acks received per window of data
300 * (cf. RFC 4341, App. A). The problems are that
301 * - arsent counts other packets as well;
302 * - the comparison uses a formula different from RFC 4341;
303 * - computing a cubic/quadratic equation each time is too complicated.
304 * Hence a different algorithm is needed.
305 */
306#if 0
307 /* Ack Ratio. Need to maintain a concept of how many windows we sent */
308 hc->tx_arsent++;
309 /* We had an ack loss in this window... */
310 if (hc->tx_ackloss) {
311 if (hc->tx_arsent >= hc->tx_cwnd) {
312 hc->tx_arsent = 0;
313 hc->tx_ackloss = 0;
314 }
315 } else {
316 /* No acks lost up to now... */
317 /* decrease ack ratio if enough packets were sent */
318 if (dp->dccps_l_ack_ratio > 1) {
319 /* XXX don't calculate denominator each time */
320 int denom = dp->dccps_l_ack_ratio * dp->dccps_l_ack_ratio -
321 dp->dccps_l_ack_ratio;
322
323 denom = hc->tx_cwnd * hc->tx_cwnd / denom;
324
325 if (hc->tx_arsent >= denom) {
326 ccid2_change_l_ack_ratio(sk, dp->dccps_l_ack_ratio - 1);
327 hc->tx_arsent = 0;
328 }
329 } else {
330 /* we can't increase ack ratio further [1] */
331 hc->tx_arsent = 0; /* or maybe set it to cwnd*/
332 }
333 }
334#endif
335
336 sk_reset_timer(sk, &hc->tx_rtotimer, jiffies + hc->tx_rto);
337
338#ifdef CONFIG_IP_DCCP_CCID2_DEBUG
339 do {
340 struct ccid2_seq *seqp = hc->tx_seqt;
341
342 while (seqp != hc->tx_seqh) {
343 ccid2_pr_debug("out seq=%llu acked=%d time=%u\n",
344 (unsigned long long)seqp->ccid2s_seq,
345 seqp->ccid2s_acked, seqp->ccid2s_sent);
346 seqp = seqp->ccid2s_next;
347 }
348 } while (0);
349 ccid2_pr_debug("=========\n");
350#endif
351}
352
353/**
354 * ccid2_rtt_estimator - Sample RTT and compute RTO using RFC2988 algorithm
355 * @sk: socket to perform estimator on
356 * @mrtt: measured RTT
357 *
358 * This code is almost identical with TCP's tcp_rtt_estimator(), since
359 * - it has a higher sampling frequency (recommended by RFC 1323),
360 * - the RTO does not collapse into RTT due to RTTVAR going towards zero,
361 * - it is simple (cf. more complex proposals such as Eifel timer or research
362 * which suggests that the gain should be set according to window size),
363 * - in tests it was found to work well with CCID2 [gerrit].
364 */
365static void ccid2_rtt_estimator(struct sock *sk, const long mrtt)
366{
367 struct ccid2_hc_tx_sock *hc = ccid2_hc_tx_sk(sk);
368 long m = mrtt ? : 1;
369
370 if (hc->tx_srtt == 0) {
371 /* First measurement m */
372 hc->tx_srtt = m << 3;
373 hc->tx_mdev = m << 1;
374
375 hc->tx_mdev_max = max(hc->tx_mdev, tcp_rto_min(sk));
376 hc->tx_rttvar = hc->tx_mdev_max;
377
378 hc->tx_rtt_seq = dccp_sk(sk)->dccps_gss;
379 } else {
380 /* Update scaled SRTT as SRTT += 1/8 * (m - SRTT) */
381 m -= (hc->tx_srtt >> 3);
382 hc->tx_srtt += m;
383
384 /* Similarly, update scaled mdev with regard to |m| */
385 if (m < 0) {
386 m = -m;
387 m -= (hc->tx_mdev >> 2);
388 /*
389 * This neutralises RTO increase when RTT < SRTT - mdev
390 * (see P. Sarolahti, A. Kuznetsov,"Congestion Control
391 * in Linux TCP", USENIX 2002, pp. 49-62).
392 */
393 if (m > 0)
394 m >>= 3;
395 } else {
396 m -= (hc->tx_mdev >> 2);
397 }
398 hc->tx_mdev += m;
399
400 if (hc->tx_mdev > hc->tx_mdev_max) {
401 hc->tx_mdev_max = hc->tx_mdev;
402 if (hc->tx_mdev_max > hc->tx_rttvar)
403 hc->tx_rttvar = hc->tx_mdev_max;
404 }
405
406 /*
407 * Decay RTTVAR at most once per flight, exploiting that
408 * 1) pipe <= cwnd <= Sequence_Window = W (RFC 4340, 7.5.2)
409 * 2) AWL = GSS-W+1 <= GAR <= GSS (RFC 4340, 7.5.1)
410 * GAR is a useful bound for FlightSize = pipe.
411 * AWL is probably too low here, as it over-estimates pipe.
412 */
413 if (after48(dccp_sk(sk)->dccps_gar, hc->tx_rtt_seq)) {
414 if (hc->tx_mdev_max < hc->tx_rttvar)
415 hc->tx_rttvar -= (hc->tx_rttvar -
416 hc->tx_mdev_max) >> 2;
417 hc->tx_rtt_seq = dccp_sk(sk)->dccps_gss;
418 hc->tx_mdev_max = tcp_rto_min(sk);
419 }
420 }
421
422 /*
423 * Set RTO from SRTT and RTTVAR
424 * As in TCP, 4 * RTTVAR >= TCP_RTO_MIN, giving a minimum RTO of 200 ms.
425 * This agrees with RFC 4341, 5:
426 * "Because DCCP does not retransmit data, DCCP does not require
427 * TCP's recommended minimum timeout of one second".
428 */
429 hc->tx_rto = (hc->tx_srtt >> 3) + hc->tx_rttvar;
430
431 if (hc->tx_rto > DCCP_RTO_MAX)
432 hc->tx_rto = DCCP_RTO_MAX;
433}
434
435static void ccid2_new_ack(struct sock *sk, struct ccid2_seq *seqp,
436 unsigned int *maxincr)
437{
438 struct ccid2_hc_tx_sock *hc = ccid2_hc_tx_sk(sk);
439 struct dccp_sock *dp = dccp_sk(sk);
440 int r_seq_used = hc->tx_cwnd / dp->dccps_l_ack_ratio;
441
442 if (hc->tx_cwnd < dp->dccps_l_seq_win &&
443 r_seq_used < dp->dccps_r_seq_win) {
444 if (hc->tx_cwnd < hc->tx_ssthresh) {
445 if (*maxincr > 0 && ++hc->tx_packets_acked >= 2) {
446 hc->tx_cwnd += 1;
447 *maxincr -= 1;
448 hc->tx_packets_acked = 0;
449 }
450 } else if (++hc->tx_packets_acked >= hc->tx_cwnd) {
451 hc->tx_cwnd += 1;
452 hc->tx_packets_acked = 0;
453 }
454 }
455
456 /*
457 * Adjust the local sequence window and the ack ratio to allow about
458 * 5 times the number of packets in the network (RFC 4340 7.5.2)
459 */
460 if (r_seq_used * CCID2_WIN_CHANGE_FACTOR >= dp->dccps_r_seq_win)
461 ccid2_change_l_ack_ratio(sk, dp->dccps_l_ack_ratio * 2);
462 else if (r_seq_used * CCID2_WIN_CHANGE_FACTOR < dp->dccps_r_seq_win/2)
463 ccid2_change_l_ack_ratio(sk, dp->dccps_l_ack_ratio / 2 ? : 1U);
464
465 if (hc->tx_cwnd * CCID2_WIN_CHANGE_FACTOR >= dp->dccps_l_seq_win)
466 ccid2_change_l_seq_window(sk, dp->dccps_l_seq_win * 2);
467 else if (hc->tx_cwnd * CCID2_WIN_CHANGE_FACTOR < dp->dccps_l_seq_win/2)
468 ccid2_change_l_seq_window(sk, dp->dccps_l_seq_win / 2);
469
470 /*
471 * FIXME: RTT is sampled several times per acknowledgment (for each
472 * entry in the Ack Vector), instead of once per Ack (as in TCP SACK).
473 * This causes the RTT to be over-estimated, since the older entries
474 * in the Ack Vector have earlier sending times.
475 * The cleanest solution is to not use the ccid2s_sent field at all
476 * and instead use DCCP timestamps: requires changes in other places.
477 */
478 ccid2_rtt_estimator(sk, ccid2_jiffies32 - seqp->ccid2s_sent);
479}
480
481static void ccid2_congestion_event(struct sock *sk, struct ccid2_seq *seqp)
482{
483 struct ccid2_hc_tx_sock *hc = ccid2_hc_tx_sk(sk);
484
485 if ((s32)(seqp->ccid2s_sent - hc->tx_last_cong) < 0) {
486 ccid2_pr_debug("Multiple losses in an RTT---treating as one\n");
487 return;
488 }
489
490 hc->tx_last_cong = ccid2_jiffies32;
491
492 hc->tx_cwnd = hc->tx_cwnd / 2 ? : 1U;
493 hc->tx_ssthresh = max(hc->tx_cwnd, 2U);
494
495 ccid2_check_l_ack_ratio(sk);
496}
497
498static int ccid2_hc_tx_parse_options(struct sock *sk, u8 packet_type,
499 u8 option, u8 *optval, u8 optlen)
500{
501 struct ccid2_hc_tx_sock *hc = ccid2_hc_tx_sk(sk);
502
503 switch (option) {
504 case DCCPO_ACK_VECTOR_0:
505 case DCCPO_ACK_VECTOR_1:
506 return dccp_ackvec_parsed_add(&hc->tx_av_chunks, optval, optlen,
507 option - DCCPO_ACK_VECTOR_0);
508 }
509 return 0;
510}
511
512static void ccid2_hc_tx_packet_recv(struct sock *sk, struct sk_buff *skb)
513{
514 struct dccp_sock *dp = dccp_sk(sk);
515 struct ccid2_hc_tx_sock *hc = ccid2_hc_tx_sk(sk);
516 const bool sender_was_blocked = ccid2_cwnd_network_limited(hc);
517 struct dccp_ackvec_parsed *avp;
518 u64 ackno, seqno;
519 struct ccid2_seq *seqp;
520 int done = 0;
521 unsigned int maxincr = 0;
522
523 /* check reverse path congestion */
524 seqno = DCCP_SKB_CB(skb)->dccpd_seq;
525
526 /* XXX this whole "algorithm" is broken. Need to fix it to keep track
527 * of the seqnos of the dupacks so that rpseq and rpdupack are correct
528 * -sorbo.
529 */
530 /* need to bootstrap */
531 if (hc->tx_rpdupack == -1) {
532 hc->tx_rpdupack = 0;
533 hc->tx_rpseq = seqno;
534 } else {
535 /* check if packet is consecutive */
536 if (dccp_delta_seqno(hc->tx_rpseq, seqno) == 1)
537 hc->tx_rpseq = seqno;
538 /* it's a later packet */
539 else if (after48(seqno, hc->tx_rpseq)) {
540 hc->tx_rpdupack++;
541
542 /* check if we got enough dupacks */
543 if (hc->tx_rpdupack >= NUMDUPACK) {
544 hc->tx_rpdupack = -1; /* XXX lame */
545 hc->tx_rpseq = 0;
546#ifdef __CCID2_COPES_GRACEFULLY_WITH_ACK_CONGESTION_CONTROL__
547 /*
548 * FIXME: Ack Congestion Control is broken; in
549 * the current state instabilities occurred with
550 * Ack Ratios greater than 1; causing hang-ups
551 * and long RTO timeouts. This needs to be fixed
552 * before opening up dynamic changes. -- gerrit
553 */
554 ccid2_change_l_ack_ratio(sk, 2 * dp->dccps_l_ack_ratio);
555#endif
556 }
557 }
558 }
559
560 /* check forward path congestion */
561 if (dccp_packet_without_ack(skb))
562 return;
563
564 /* still didn't send out new data packets */
565 if (hc->tx_seqh == hc->tx_seqt)
566 goto done;
567
568 ackno = DCCP_SKB_CB(skb)->dccpd_ack_seq;
569 if (after48(ackno, hc->tx_high_ack))
570 hc->tx_high_ack = ackno;
571
572 seqp = hc->tx_seqt;
573 while (before48(seqp->ccid2s_seq, ackno)) {
574 seqp = seqp->ccid2s_next;
575 if (seqp == hc->tx_seqh) {
576 seqp = hc->tx_seqh->ccid2s_prev;
577 break;
578 }
579 }
580
581 /*
582 * In slow-start, cwnd can increase up to a maximum of Ack Ratio/2
583 * packets per acknowledgement. Rounding up avoids that cwnd is not
584 * advanced when Ack Ratio is 1 and gives a slight edge otherwise.
585 */
586 if (hc->tx_cwnd < hc->tx_ssthresh)
587 maxincr = DIV_ROUND_UP(dp->dccps_l_ack_ratio, 2);
588
589 /* go through all ack vectors */
590 list_for_each_entry(avp, &hc->tx_av_chunks, node) {
591 /* go through this ack vector */
592 for (; avp->len--; avp->vec++) {
593 u64 ackno_end_rl = SUB48(ackno,
594 dccp_ackvec_runlen(avp->vec));
595
596 ccid2_pr_debug("ackvec %llu |%u,%u|\n",
597 (unsigned long long)ackno,
598 dccp_ackvec_state(avp->vec) >> 6,
599 dccp_ackvec_runlen(avp->vec));
600 /* if the seqno we are analyzing is larger than the
601 * current ackno, then move towards the tail of our
602 * seqnos.
603 */
604 while (after48(seqp->ccid2s_seq, ackno)) {
605 if (seqp == hc->tx_seqt) {
606 done = 1;
607 break;
608 }
609 seqp = seqp->ccid2s_prev;
610 }
611 if (done)
612 break;
613
614 /* check all seqnos in the range of the vector
615 * run length
616 */
617 while (between48(seqp->ccid2s_seq,ackno_end_rl,ackno)) {
618 const u8 state = dccp_ackvec_state(avp->vec);
619
620 /* new packet received or marked */
621 if (state != DCCPAV_NOT_RECEIVED &&
622 !seqp->ccid2s_acked) {
623 if (state == DCCPAV_ECN_MARKED)
624 ccid2_congestion_event(sk,
625 seqp);
626 else
627 ccid2_new_ack(sk, seqp,
628 &maxincr);
629
630 seqp->ccid2s_acked = 1;
631 ccid2_pr_debug("Got ack for %llu\n",
632 (unsigned long long)seqp->ccid2s_seq);
633 hc->tx_pipe--;
634 }
635 if (seqp == hc->tx_seqt) {
636 done = 1;
637 break;
638 }
639 seqp = seqp->ccid2s_prev;
640 }
641 if (done)
642 break;
643
644 ackno = SUB48(ackno_end_rl, 1);
645 }
646 if (done)
647 break;
648 }
649
650 /* The state about what is acked should be correct now
651 * Check for NUMDUPACK
652 */
653 seqp = hc->tx_seqt;
654 while (before48(seqp->ccid2s_seq, hc->tx_high_ack)) {
655 seqp = seqp->ccid2s_next;
656 if (seqp == hc->tx_seqh) {
657 seqp = hc->tx_seqh->ccid2s_prev;
658 break;
659 }
660 }
661 done = 0;
662 while (1) {
663 if (seqp->ccid2s_acked) {
664 done++;
665 if (done == NUMDUPACK)
666 break;
667 }
668 if (seqp == hc->tx_seqt)
669 break;
670 seqp = seqp->ccid2s_prev;
671 }
672
673 /* If there are at least 3 acknowledgements, anything unacknowledged
674 * below the last sequence number is considered lost
675 */
676 if (done == NUMDUPACK) {
677 struct ccid2_seq *last_acked = seqp;
678
679 /* check for lost packets */
680 while (1) {
681 if (!seqp->ccid2s_acked) {
682 ccid2_pr_debug("Packet lost: %llu\n",
683 (unsigned long long)seqp->ccid2s_seq);
684 /* XXX need to traverse from tail -> head in
685 * order to detect multiple congestion events in
686 * one ack vector.
687 */
688 ccid2_congestion_event(sk, seqp);
689 hc->tx_pipe--;
690 }
691 if (seqp == hc->tx_seqt)
692 break;
693 seqp = seqp->ccid2s_prev;
694 }
695
696 hc->tx_seqt = last_acked;
697 }
698
699 /* trim acked packets in tail */
700 while (hc->tx_seqt != hc->tx_seqh) {
701 if (!hc->tx_seqt->ccid2s_acked)
702 break;
703
704 hc->tx_seqt = hc->tx_seqt->ccid2s_next;
705 }
706
707 /* restart RTO timer if not all outstanding data has been acked */
708 if (hc->tx_pipe == 0)
709 sk_stop_timer(sk, &hc->tx_rtotimer);
710 else
711 sk_reset_timer(sk, &hc->tx_rtotimer, jiffies + hc->tx_rto);
712done:
713 /* check if incoming Acks allow pending packets to be sent */
714 if (sender_was_blocked && !ccid2_cwnd_network_limited(hc))
715 dccp_tasklet_schedule(sk);
716 dccp_ackvec_parsed_cleanup(&hc->tx_av_chunks);
717}
718
719static int ccid2_hc_tx_init(struct ccid *ccid, struct sock *sk)
720{
721 struct ccid2_hc_tx_sock *hc = ccid_priv(ccid);
722 struct dccp_sock *dp = dccp_sk(sk);
723 u32 max_ratio;
724
725 /* RFC 4341, 5: initialise ssthresh to arbitrarily high (max) value */
726 hc->tx_ssthresh = ~0U;
727
728 /* Use larger initial windows (RFC 4341, section 5). */
729 hc->tx_cwnd = rfc3390_bytes_to_packets(dp->dccps_mss_cache);
730 hc->tx_expected_wnd = hc->tx_cwnd;
731
732 /* Make sure that Ack Ratio is enabled and within bounds. */
733 max_ratio = DIV_ROUND_UP(hc->tx_cwnd, 2);
734 if (dp->dccps_l_ack_ratio == 0 || dp->dccps_l_ack_ratio > max_ratio)
735 dp->dccps_l_ack_ratio = max_ratio;
736
737 /* XXX init ~ to window size... */
738 if (ccid2_hc_tx_alloc_seq(hc))
739 return -ENOMEM;
740
741 hc->tx_rto = DCCP_TIMEOUT_INIT;
742 hc->tx_rpdupack = -1;
743 hc->tx_last_cong = hc->tx_lsndtime = hc->tx_cwnd_stamp = ccid2_jiffies32;
744 hc->tx_cwnd_used = 0;
745 hc->sk = sk;
746 timer_setup(&hc->tx_rtotimer, ccid2_hc_tx_rto_expire, 0);
747 INIT_LIST_HEAD(&hc->tx_av_chunks);
748 return 0;
749}
750
751static void ccid2_hc_tx_exit(struct sock *sk)
752{
753 struct ccid2_hc_tx_sock *hc = ccid2_hc_tx_sk(sk);
754 int i;
755
756 sk_stop_timer(sk, &hc->tx_rtotimer);
757
758 for (i = 0; i < hc->tx_seqbufc; i++)
759 kfree(hc->tx_seqbuf[i]);
760 hc->tx_seqbufc = 0;
761 dccp_ackvec_parsed_cleanup(&hc->tx_av_chunks);
762}
763
764static void ccid2_hc_rx_packet_recv(struct sock *sk, struct sk_buff *skb)
765{
766 struct ccid2_hc_rx_sock *hc = ccid2_hc_rx_sk(sk);
767
768 if (!dccp_data_packet(skb))
769 return;
770
771 if (++hc->rx_num_data_pkts >= dccp_sk(sk)->dccps_r_ack_ratio) {
772 dccp_send_ack(sk);
773 hc->rx_num_data_pkts = 0;
774 }
775}
776
777struct ccid_operations ccid2_ops = {
778 .ccid_id = DCCPC_CCID2,
779 .ccid_name = "TCP-like",
780 .ccid_hc_tx_obj_size = sizeof(struct ccid2_hc_tx_sock),
781 .ccid_hc_tx_init = ccid2_hc_tx_init,
782 .ccid_hc_tx_exit = ccid2_hc_tx_exit,
783 .ccid_hc_tx_send_packet = ccid2_hc_tx_send_packet,
784 .ccid_hc_tx_packet_sent = ccid2_hc_tx_packet_sent,
785 .ccid_hc_tx_parse_options = ccid2_hc_tx_parse_options,
786 .ccid_hc_tx_packet_recv = ccid2_hc_tx_packet_recv,
787 .ccid_hc_rx_obj_size = sizeof(struct ccid2_hc_rx_sock),
788 .ccid_hc_rx_packet_recv = ccid2_hc_rx_packet_recv,
789};
790
791#ifdef CONFIG_IP_DCCP_CCID2_DEBUG
792module_param(ccid2_debug, bool, 0644);
793MODULE_PARM_DESC(ccid2_debug, "Enable CCID-2 debug messages");
794#endif