1// SPDX-License-Identifier: GPL-2.0
2/*
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
6 *
7 * Implementation of the Transmission Control Protocol(TCP).
8 *
9 * Authors: Ross Biro
10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11 * Mark Evans, <evansmp@uhura.aston.ac.uk>
12 * Corey Minyard <wf-rch!minyard@relay.EU.net>
13 * Florian La Roche, <flla@stud.uni-sb.de>
14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15 * Linus Torvalds, <torvalds@cs.helsinki.fi>
16 * Alan Cox, <gw4pts@gw4pts.ampr.org>
17 * Matthew Dillon, <dillon@apollo.west.oic.com>
18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19 * Jorge Cwik, <jorge@laser.satlink.net>
20 */
21
22/*
23 * Changes:
24 * Pedro Roque : Fast Retransmit/Recovery.
25 * Two receive queues.
26 * Retransmit queue handled by TCP.
27 * Better retransmit timer handling.
28 * New congestion avoidance.
29 * Header prediction.
30 * Variable renaming.
31 *
32 * Eric : Fast Retransmit.
33 * Randy Scott : MSS option defines.
34 * Eric Schenk : Fixes to slow start algorithm.
35 * Eric Schenk : Yet another double ACK bug.
36 * Eric Schenk : Delayed ACK bug fixes.
37 * Eric Schenk : Floyd style fast retrans war avoidance.
38 * David S. Miller : Don't allow zero congestion window.
39 * Eric Schenk : Fix retransmitter so that it sends
40 * next packet on ack of previous packet.
41 * Andi Kleen : Moved open_request checking here
42 * and process RSTs for open_requests.
43 * Andi Kleen : Better prune_queue, and other fixes.
44 * Andrey Savochkin: Fix RTT measurements in the presence of
45 * timestamps.
46 * Andrey Savochkin: Check sequence numbers correctly when
47 * removing SACKs due to in sequence incoming
48 * data segments.
49 * Andi Kleen: Make sure we never ack data there is not
50 * enough room for. Also make this condition
51 * a fatal error if it might still happen.
52 * Andi Kleen: Add tcp_measure_rcv_mss to make
53 * connections with MSS<min(MTU,ann. MSS)
54 * work without delayed acks.
55 * Andi Kleen: Process packets with PSH set in the
56 * fast path.
57 * J Hadi Salim: ECN support
58 * Andrei Gurtov,
59 * Pasi Sarolahti,
60 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
61 * engine. Lots of bugs are found.
62 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
63 */
64
65#define pr_fmt(fmt) "TCP: " fmt
66
67#include <linux/mm.h>
68#include <linux/slab.h>
69#include <linux/module.h>
70#include <linux/sysctl.h>
71#include <linux/kernel.h>
72#include <linux/prefetch.h>
73#include <net/dst.h>
74#include <net/tcp.h>
75#include <net/inet_common.h>
76#include <linux/ipsec.h>
77#include <asm/unaligned.h>
78#include <linux/errqueue.h>
79#include <trace/events/tcp.h>
80#include <linux/static_key.h>
81#include <net/busy_poll.h>
82
83int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
84
85#define FLAG_DATA 0x01 /* Incoming frame contained data. */
86#define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
87#define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
88#define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
89#define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
90#define FLAG_DATA_SACKED 0x20 /* New SACK. */
91#define FLAG_ECE 0x40 /* ECE in this ACK */
92#define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */
93#define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
94#define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
95#define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
96#define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
97#define FLAG_SET_XMIT_TIMER 0x1000 /* Set TLP or RTO timer */
98#define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
99#define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
100#define FLAG_NO_CHALLENGE_ACK 0x8000 /* do not call tcp_send_challenge_ack() */
101#define FLAG_ACK_MAYBE_DELAYED 0x10000 /* Likely a delayed ACK */
102
103#define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
104#define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
105#define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE|FLAG_DSACKING_ACK)
106#define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
107
108#define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
109#define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
110
111#define REXMIT_NONE 0 /* no loss recovery to do */
112#define REXMIT_LOST 1 /* retransmit packets marked lost */
113#define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */
114
115#if IS_ENABLED(CONFIG_TLS_DEVICE)
116static DEFINE_STATIC_KEY_FALSE(clean_acked_data_enabled);
117
118void clean_acked_data_enable(struct inet_connection_sock *icsk,
119 void (*cad)(struct sock *sk, u32 ack_seq))
120{
121 icsk->icsk_clean_acked = cad;
122 static_branch_inc(&clean_acked_data_enabled);
123}
124EXPORT_SYMBOL_GPL(clean_acked_data_enable);
125
126void clean_acked_data_disable(struct inet_connection_sock *icsk)
127{
128 static_branch_dec(&clean_acked_data_enabled);
129 icsk->icsk_clean_acked = NULL;
130}
131EXPORT_SYMBOL_GPL(clean_acked_data_disable);
132#endif
133
134static void tcp_gro_dev_warn(struct sock *sk, const struct sk_buff *skb,
135 unsigned int len)
136{
137 static bool __once __read_mostly;
138
139 if (!__once) {
140 struct net_device *dev;
141
142 __once = true;
143
144 rcu_read_lock();
145 dev = dev_get_by_index_rcu(sock_net(sk), skb->skb_iif);
146 if (!dev || len >= dev->mtu)
147 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n",
148 dev ? dev->name : "Unknown driver");
149 rcu_read_unlock();
150 }
151}
152
153/* Adapt the MSS value used to make delayed ack decision to the
154 * real world.
155 */
156static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
157{
158 struct inet_connection_sock *icsk = inet_csk(sk);
159 const unsigned int lss = icsk->icsk_ack.last_seg_size;
160 unsigned int len;
161
162 icsk->icsk_ack.last_seg_size = 0;
163
164 /* skb->len may jitter because of SACKs, even if peer
165 * sends good full-sized frames.
166 */
167 len = skb_shinfo(skb)->gso_size ? : skb->len;
168 if (len >= icsk->icsk_ack.rcv_mss) {
169 icsk->icsk_ack.rcv_mss = min_t(unsigned int, len,
170 tcp_sk(sk)->advmss);
171 /* Account for possibly-removed options */
172 if (unlikely(len > icsk->icsk_ack.rcv_mss +
173 MAX_TCP_OPTION_SPACE))
174 tcp_gro_dev_warn(sk, skb, len);
175 } else {
176 /* Otherwise, we make more careful check taking into account,
177 * that SACKs block is variable.
178 *
179 * "len" is invariant segment length, including TCP header.
180 */
181 len += skb->data - skb_transport_header(skb);
182 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
183 /* If PSH is not set, packet should be
184 * full sized, provided peer TCP is not badly broken.
185 * This observation (if it is correct 8)) allows
186 * to handle super-low mtu links fairly.
187 */
188 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
189 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
190 /* Subtract also invariant (if peer is RFC compliant),
191 * tcp header plus fixed timestamp option length.
192 * Resulting "len" is MSS free of SACK jitter.
193 */
194 len -= tcp_sk(sk)->tcp_header_len;
195 icsk->icsk_ack.last_seg_size = len;
196 if (len == lss) {
197 icsk->icsk_ack.rcv_mss = len;
198 return;
199 }
200 }
201 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
202 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
203 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
204 }
205}
206
207static void tcp_incr_quickack(struct sock *sk, unsigned int max_quickacks)
208{
209 struct inet_connection_sock *icsk = inet_csk(sk);
210 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
211
212 if (quickacks == 0)
213 quickacks = 2;
214 quickacks = min(quickacks, max_quickacks);
215 if (quickacks > icsk->icsk_ack.quick)
216 icsk->icsk_ack.quick = quickacks;
217}
218
219void tcp_enter_quickack_mode(struct sock *sk, unsigned int max_quickacks)
220{
221 struct inet_connection_sock *icsk = inet_csk(sk);
222
223 tcp_incr_quickack(sk, max_quickacks);
224 icsk->icsk_ack.pingpong = 0;
225 icsk->icsk_ack.ato = TCP_ATO_MIN;
226}
227EXPORT_SYMBOL(tcp_enter_quickack_mode);
228
229/* Send ACKs quickly, if "quick" count is not exhausted
230 * and the session is not interactive.
231 */
232
233static bool tcp_in_quickack_mode(struct sock *sk)
234{
235 const struct inet_connection_sock *icsk = inet_csk(sk);
236 const struct dst_entry *dst = __sk_dst_get(sk);
237
238 return (dst && dst_metric(dst, RTAX_QUICKACK)) ||
239 (icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong);
240}
241
242static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
243{
244 if (tp->ecn_flags & TCP_ECN_OK)
245 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
246}
247
248static void tcp_ecn_accept_cwr(struct sock *sk, const struct sk_buff *skb)
249{
250 if (tcp_hdr(skb)->cwr) {
251 tcp_sk(sk)->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
252
253 /* If the sender is telling us it has entered CWR, then its
254 * cwnd may be very low (even just 1 packet), so we should ACK
255 * immediately.
256 */
257 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
258 }
259}
260
261static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
262{
263 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
264}
265
266static void __tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
267{
268 struct tcp_sock *tp = tcp_sk(sk);
269
270 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
271 case INET_ECN_NOT_ECT:
272 /* Funny extension: if ECT is not set on a segment,
273 * and we already seen ECT on a previous segment,
274 * it is probably a retransmit.
275 */
276 if (tp->ecn_flags & TCP_ECN_SEEN)
277 tcp_enter_quickack_mode(sk, 2);
278 break;
279 case INET_ECN_CE:
280 if (tcp_ca_needs_ecn(sk))
281 tcp_ca_event(sk, CA_EVENT_ECN_IS_CE);
282
283 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
284 /* Better not delay acks, sender can have a very low cwnd */
285 tcp_enter_quickack_mode(sk, 2);
286 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
287 }
288 tp->ecn_flags |= TCP_ECN_SEEN;
289 break;
290 default:
291 if (tcp_ca_needs_ecn(sk))
292 tcp_ca_event(sk, CA_EVENT_ECN_NO_CE);
293 tp->ecn_flags |= TCP_ECN_SEEN;
294 break;
295 }
296}
297
298static void tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
299{
300 if (tcp_sk(sk)->ecn_flags & TCP_ECN_OK)
301 __tcp_ecn_check_ce(sk, skb);
302}
303
304static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
305{
306 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
307 tp->ecn_flags &= ~TCP_ECN_OK;
308}
309
310static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
311{
312 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
313 tp->ecn_flags &= ~TCP_ECN_OK;
314}
315
316static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
317{
318 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
319 return true;
320 return false;
321}
322
323/* Buffer size and advertised window tuning.
324 *
325 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
326 */
327
328static void tcp_sndbuf_expand(struct sock *sk)
329{
330 const struct tcp_sock *tp = tcp_sk(sk);
331 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
332 int sndmem, per_mss;
333 u32 nr_segs;
334
335 /* Worst case is non GSO/TSO : each frame consumes one skb
336 * and skb->head is kmalloced using power of two area of memory
337 */
338 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
339 MAX_TCP_HEADER +
340 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
341
342 per_mss = roundup_pow_of_two(per_mss) +
343 SKB_DATA_ALIGN(sizeof(struct sk_buff));
344
345 nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd);
346 nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
347
348 /* Fast Recovery (RFC 5681 3.2) :
349 * Cubic needs 1.7 factor, rounded to 2 to include
350 * extra cushion (application might react slowly to EPOLLOUT)
351 */
352 sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2;
353 sndmem *= nr_segs * per_mss;
354
355 if (sk->sk_sndbuf < sndmem)
356 sk->sk_sndbuf = min(sndmem, sock_net(sk)->ipv4.sysctl_tcp_wmem[2]);
357}
358
359/* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
360 *
361 * All tcp_full_space() is split to two parts: "network" buffer, allocated
362 * forward and advertised in receiver window (tp->rcv_wnd) and
363 * "application buffer", required to isolate scheduling/application
364 * latencies from network.
365 * window_clamp is maximal advertised window. It can be less than
366 * tcp_full_space(), in this case tcp_full_space() - window_clamp
367 * is reserved for "application" buffer. The less window_clamp is
368 * the smoother our behaviour from viewpoint of network, but the lower
369 * throughput and the higher sensitivity of the connection to losses. 8)
370 *
371 * rcv_ssthresh is more strict window_clamp used at "slow start"
372 * phase to predict further behaviour of this connection.
373 * It is used for two goals:
374 * - to enforce header prediction at sender, even when application
375 * requires some significant "application buffer". It is check #1.
376 * - to prevent pruning of receive queue because of misprediction
377 * of receiver window. Check #2.
378 *
379 * The scheme does not work when sender sends good segments opening
380 * window and then starts to feed us spaghetti. But it should work
381 * in common situations. Otherwise, we have to rely on queue collapsing.
382 */
383
384/* Slow part of check#2. */
385static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
386{
387 struct tcp_sock *tp = tcp_sk(sk);
388 /* Optimize this! */
389 int truesize = tcp_win_from_space(sk, skb->truesize) >> 1;
390 int window = tcp_win_from_space(sk, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]) >> 1;
391
392 while (tp->rcv_ssthresh <= window) {
393 if (truesize <= skb->len)
394 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
395
396 truesize >>= 1;
397 window >>= 1;
398 }
399 return 0;
400}
401
402static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
403{
404 struct tcp_sock *tp = tcp_sk(sk);
405
406 /* Check #1 */
407 if (tp->rcv_ssthresh < tp->window_clamp &&
408 (int)tp->rcv_ssthresh < tcp_space(sk) &&
409 !tcp_under_memory_pressure(sk)) {
410 int incr;
411
412 /* Check #2. Increase window, if skb with such overhead
413 * will fit to rcvbuf in future.
414 */
415 if (tcp_win_from_space(sk, skb->truesize) <= skb->len)
416 incr = 2 * tp->advmss;
417 else
418 incr = __tcp_grow_window(sk, skb);
419
420 if (incr) {
421 incr = max_t(int, incr, 2 * skb->len);
422 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
423 tp->window_clamp);
424 inet_csk(sk)->icsk_ack.quick |= 1;
425 }
426 }
427}
428
429/* 3. Tuning rcvbuf, when connection enters established state. */
430static void tcp_fixup_rcvbuf(struct sock *sk)
431{
432 u32 mss = tcp_sk(sk)->advmss;
433 int rcvmem;
434
435 rcvmem = 2 * SKB_TRUESIZE(mss + MAX_TCP_HEADER) *
436 tcp_default_init_rwnd(mss);
437
438 /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
439 * Allow enough cushion so that sender is not limited by our window
440 */
441 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf)
442 rcvmem <<= 2;
443
444 if (sk->sk_rcvbuf < rcvmem)
445 sk->sk_rcvbuf = min(rcvmem, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
446}
447
448/* 4. Try to fixup all. It is made immediately after connection enters
449 * established state.
450 */
451void tcp_init_buffer_space(struct sock *sk)
452{
453 int tcp_app_win = sock_net(sk)->ipv4.sysctl_tcp_app_win;
454 struct tcp_sock *tp = tcp_sk(sk);
455 int maxwin;
456
457 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
458 tcp_fixup_rcvbuf(sk);
459 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
460 tcp_sndbuf_expand(sk);
461
462 tp->rcvq_space.space = tp->rcv_wnd;
463 tcp_mstamp_refresh(tp);
464 tp->rcvq_space.time = tp->tcp_mstamp;
465 tp->rcvq_space.seq = tp->copied_seq;
466
467 maxwin = tcp_full_space(sk);
468
469 if (tp->window_clamp >= maxwin) {
470 tp->window_clamp = maxwin;
471
472 if (tcp_app_win && maxwin > 4 * tp->advmss)
473 tp->window_clamp = max(maxwin -
474 (maxwin >> tcp_app_win),
475 4 * tp->advmss);
476 }
477
478 /* Force reservation of one segment. */
479 if (tcp_app_win &&
480 tp->window_clamp > 2 * tp->advmss &&
481 tp->window_clamp + tp->advmss > maxwin)
482 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
483
484 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
485 tp->snd_cwnd_stamp = tcp_jiffies32;
486}
487
488/* 5. Recalculate window clamp after socket hit its memory bounds. */
489static void tcp_clamp_window(struct sock *sk)
490{
491 struct tcp_sock *tp = tcp_sk(sk);
492 struct inet_connection_sock *icsk = inet_csk(sk);
493 struct net *net = sock_net(sk);
494
495 icsk->icsk_ack.quick = 0;
496
497 if (sk->sk_rcvbuf < net->ipv4.sysctl_tcp_rmem[2] &&
498 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
499 !tcp_under_memory_pressure(sk) &&
500 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
501 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
502 net->ipv4.sysctl_tcp_rmem[2]);
503 }
504 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
505 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
506}
507
508/* Initialize RCV_MSS value.
509 * RCV_MSS is an our guess about MSS used by the peer.
510 * We haven't any direct information about the MSS.
511 * It's better to underestimate the RCV_MSS rather than overestimate.
512 * Overestimations make us ACKing less frequently than needed.
513 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
514 */
515void tcp_initialize_rcv_mss(struct sock *sk)
516{
517 const struct tcp_sock *tp = tcp_sk(sk);
518 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
519
520 hint = min(hint, tp->rcv_wnd / 2);
521 hint = min(hint, TCP_MSS_DEFAULT);
522 hint = max(hint, TCP_MIN_MSS);
523
524 inet_csk(sk)->icsk_ack.rcv_mss = hint;
525}
526EXPORT_SYMBOL(tcp_initialize_rcv_mss);
527
528/* Receiver "autotuning" code.
529 *
530 * The algorithm for RTT estimation w/o timestamps is based on
531 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
532 * <http://public.lanl.gov/radiant/pubs.html#DRS>
533 *
534 * More detail on this code can be found at
535 * <http://staff.psc.edu/jheffner/>,
536 * though this reference is out of date. A new paper
537 * is pending.
538 */
539static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
540{
541 u32 new_sample = tp->rcv_rtt_est.rtt_us;
542 long m = sample;
543
544 if (new_sample != 0) {
545 /* If we sample in larger samples in the non-timestamp
546 * case, we could grossly overestimate the RTT especially
547 * with chatty applications or bulk transfer apps which
548 * are stalled on filesystem I/O.
549 *
550 * Also, since we are only going for a minimum in the
551 * non-timestamp case, we do not smooth things out
552 * else with timestamps disabled convergence takes too
553 * long.
554 */
555 if (!win_dep) {
556 m -= (new_sample >> 3);
557 new_sample += m;
558 } else {
559 m <<= 3;
560 if (m < new_sample)
561 new_sample = m;
562 }
563 } else {
564 /* No previous measure. */
565 new_sample = m << 3;
566 }
567
568 tp->rcv_rtt_est.rtt_us = new_sample;
569}
570
571static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
572{
573 u32 delta_us;
574
575 if (tp->rcv_rtt_est.time == 0)
576 goto new_measure;
577 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
578 return;
579 delta_us = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcv_rtt_est.time);
580 if (!delta_us)
581 delta_us = 1;
582 tcp_rcv_rtt_update(tp, delta_us, 1);
583
584new_measure:
585 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
586 tp->rcv_rtt_est.time = tp->tcp_mstamp;
587}
588
589static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
590 const struct sk_buff *skb)
591{
592 struct tcp_sock *tp = tcp_sk(sk);
593
594 if (tp->rx_opt.rcv_tsecr == tp->rcv_rtt_last_tsecr)
595 return;
596 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
597
598 if (TCP_SKB_CB(skb)->end_seq -
599 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss) {
600 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
601 u32 delta_us;
602
603 if (!delta)
604 delta = 1;
605 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
606 tcp_rcv_rtt_update(tp, delta_us, 0);
607 }
608}
609
610/*
611 * This function should be called every time data is copied to user space.
612 * It calculates the appropriate TCP receive buffer space.
613 */
614void tcp_rcv_space_adjust(struct sock *sk)
615{
616 struct tcp_sock *tp = tcp_sk(sk);
617 u32 copied;
618 int time;
619
620 trace_tcp_rcv_space_adjust(sk);
621
622 tcp_mstamp_refresh(tp);
623 time = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcvq_space.time);
624 if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0)
625 return;
626
627 /* Number of bytes copied to user in last RTT */
628 copied = tp->copied_seq - tp->rcvq_space.seq;
629 if (copied <= tp->rcvq_space.space)
630 goto new_measure;
631
632 /* A bit of theory :
633 * copied = bytes received in previous RTT, our base window
634 * To cope with packet losses, we need a 2x factor
635 * To cope with slow start, and sender growing its cwin by 100 %
636 * every RTT, we need a 4x factor, because the ACK we are sending
637 * now is for the next RTT, not the current one :
638 * <prev RTT . ><current RTT .. ><next RTT .... >
639 */
640
641 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf &&
642 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
643 int rcvmem, rcvbuf;
644 u64 rcvwin, grow;
645
646 /* minimal window to cope with packet losses, assuming
647 * steady state. Add some cushion because of small variations.
648 */
649 rcvwin = ((u64)copied << 1) + 16 * tp->advmss;
650
651 /* Accommodate for sender rate increase (eg. slow start) */
652 grow = rcvwin * (copied - tp->rcvq_space.space);
653 do_div(grow, tp->rcvq_space.space);
654 rcvwin += (grow << 1);
655
656 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
657 while (tcp_win_from_space(sk, rcvmem) < tp->advmss)
658 rcvmem += 128;
659
660 do_div(rcvwin, tp->advmss);
661 rcvbuf = min_t(u64, rcvwin * rcvmem,
662 sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
663 if (rcvbuf > sk->sk_rcvbuf) {
664 sk->sk_rcvbuf = rcvbuf;
665
666 /* Make the window clamp follow along. */
667 tp->window_clamp = tcp_win_from_space(sk, rcvbuf);
668 }
669 }
670 tp->rcvq_space.space = copied;
671
672new_measure:
673 tp->rcvq_space.seq = tp->copied_seq;
674 tp->rcvq_space.time = tp->tcp_mstamp;
675}
676
677/* There is something which you must keep in mind when you analyze the
678 * behavior of the tp->ato delayed ack timeout interval. When a
679 * connection starts up, we want to ack as quickly as possible. The
680 * problem is that "good" TCP's do slow start at the beginning of data
681 * transmission. The means that until we send the first few ACK's the
682 * sender will sit on his end and only queue most of his data, because
683 * he can only send snd_cwnd unacked packets at any given time. For
684 * each ACK we send, he increments snd_cwnd and transmits more of his
685 * queue. -DaveM
686 */
687static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
688{
689 struct tcp_sock *tp = tcp_sk(sk);
690 struct inet_connection_sock *icsk = inet_csk(sk);
691 u32 now;
692
693 inet_csk_schedule_ack(sk);
694
695 tcp_measure_rcv_mss(sk, skb);
696
697 tcp_rcv_rtt_measure(tp);
698
699 now = tcp_jiffies32;
700
701 if (!icsk->icsk_ack.ato) {
702 /* The _first_ data packet received, initialize
703 * delayed ACK engine.
704 */
705 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
706 icsk->icsk_ack.ato = TCP_ATO_MIN;
707 } else {
708 int m = now - icsk->icsk_ack.lrcvtime;
709
710 if (m <= TCP_ATO_MIN / 2) {
711 /* The fastest case is the first. */
712 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
713 } else if (m < icsk->icsk_ack.ato) {
714 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
715 if (icsk->icsk_ack.ato > icsk->icsk_rto)
716 icsk->icsk_ack.ato = icsk->icsk_rto;
717 } else if (m > icsk->icsk_rto) {
718 /* Too long gap. Apparently sender failed to
719 * restart window, so that we send ACKs quickly.
720 */
721 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
722 sk_mem_reclaim(sk);
723 }
724 }
725 icsk->icsk_ack.lrcvtime = now;
726
727 tcp_ecn_check_ce(sk, skb);
728
729 if (skb->len >= 128)
730 tcp_grow_window(sk, skb);
731}
732
733/* Called to compute a smoothed rtt estimate. The data fed to this
734 * routine either comes from timestamps, or from segments that were
735 * known _not_ to have been retransmitted [see Karn/Partridge
736 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
737 * piece by Van Jacobson.
738 * NOTE: the next three routines used to be one big routine.
739 * To save cycles in the RFC 1323 implementation it was better to break
740 * it up into three procedures. -- erics
741 */
742static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
743{
744 struct tcp_sock *tp = tcp_sk(sk);
745 long m = mrtt_us; /* RTT */
746 u32 srtt = tp->srtt_us;
747
748 /* The following amusing code comes from Jacobson's
749 * article in SIGCOMM '88. Note that rtt and mdev
750 * are scaled versions of rtt and mean deviation.
751 * This is designed to be as fast as possible
752 * m stands for "measurement".
753 *
754 * On a 1990 paper the rto value is changed to:
755 * RTO = rtt + 4 * mdev
756 *
757 * Funny. This algorithm seems to be very broken.
758 * These formulae increase RTO, when it should be decreased, increase
759 * too slowly, when it should be increased quickly, decrease too quickly
760 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
761 * does not matter how to _calculate_ it. Seems, it was trap
762 * that VJ failed to avoid. 8)
763 */
764 if (srtt != 0) {
765 m -= (srtt >> 3); /* m is now error in rtt est */
766 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
767 if (m < 0) {
768 m = -m; /* m is now abs(error) */
769 m -= (tp->mdev_us >> 2); /* similar update on mdev */
770 /* This is similar to one of Eifel findings.
771 * Eifel blocks mdev updates when rtt decreases.
772 * This solution is a bit different: we use finer gain
773 * for mdev in this case (alpha*beta).
774 * Like Eifel it also prevents growth of rto,
775 * but also it limits too fast rto decreases,
776 * happening in pure Eifel.
777 */
778 if (m > 0)
779 m >>= 3;
780 } else {
781 m -= (tp->mdev_us >> 2); /* similar update on mdev */
782 }
783 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
784 if (tp->mdev_us > tp->mdev_max_us) {
785 tp->mdev_max_us = tp->mdev_us;
786 if (tp->mdev_max_us > tp->rttvar_us)
787 tp->rttvar_us = tp->mdev_max_us;
788 }
789 if (after(tp->snd_una, tp->rtt_seq)) {
790 if (tp->mdev_max_us < tp->rttvar_us)
791 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
792 tp->rtt_seq = tp->snd_nxt;
793 tp->mdev_max_us = tcp_rto_min_us(sk);
794 }
795 } else {
796 /* no previous measure. */
797 srtt = m << 3; /* take the measured time to be rtt */
798 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
799 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
800 tp->mdev_max_us = tp->rttvar_us;
801 tp->rtt_seq = tp->snd_nxt;
802 }
803 tp->srtt_us = max(1U, srtt);
804}
805
806static void tcp_update_pacing_rate(struct sock *sk)
807{
808 const struct tcp_sock *tp = tcp_sk(sk);
809 u64 rate;
810
811 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
812 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
813
814 /* current rate is (cwnd * mss) / srtt
815 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
816 * In Congestion Avoidance phase, set it to 120 % the current rate.
817 *
818 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
819 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
820 * end of slow start and should slow down.
821 */
822 if (tp->snd_cwnd < tp->snd_ssthresh / 2)
823 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ss_ratio;
824 else
825 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ca_ratio;
826
827 rate *= max(tp->snd_cwnd, tp->packets_out);
828
829 if (likely(tp->srtt_us))
830 do_div(rate, tp->srtt_us);
831
832 /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate
833 * without any lock. We want to make sure compiler wont store
834 * intermediate values in this location.
835 */
836 WRITE_ONCE(sk->sk_pacing_rate, min_t(u64, rate,
837 sk->sk_max_pacing_rate));
838}
839
840/* Calculate rto without backoff. This is the second half of Van Jacobson's
841 * routine referred to above.
842 */
843static void tcp_set_rto(struct sock *sk)
844{
845 const struct tcp_sock *tp = tcp_sk(sk);
846 /* Old crap is replaced with new one. 8)
847 *
848 * More seriously:
849 * 1. If rtt variance happened to be less 50msec, it is hallucination.
850 * It cannot be less due to utterly erratic ACK generation made
851 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
852 * to do with delayed acks, because at cwnd>2 true delack timeout
853 * is invisible. Actually, Linux-2.4 also generates erratic
854 * ACKs in some circumstances.
855 */
856 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
857
858 /* 2. Fixups made earlier cannot be right.
859 * If we do not estimate RTO correctly without them,
860 * all the algo is pure shit and should be replaced
861 * with correct one. It is exactly, which we pretend to do.
862 */
863
864 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
865 * guarantees that rto is higher.
866 */
867 tcp_bound_rto(sk);
868}
869
870__u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
871{
872 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
873
874 if (!cwnd)
875 cwnd = TCP_INIT_CWND;
876 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
877}
878
879/* Take a notice that peer is sending D-SACKs */
880static void tcp_dsack_seen(struct tcp_sock *tp)
881{
882 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
883 tp->rack.dsack_seen = 1;
884 tp->dsack_dups++;
885}
886
887/* It's reordering when higher sequence was delivered (i.e. sacked) before
888 * some lower never-retransmitted sequence ("low_seq"). The maximum reordering
889 * distance is approximated in full-mss packet distance ("reordering").
890 */
891static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq,
892 const int ts)
893{
894 struct tcp_sock *tp = tcp_sk(sk);
895 const u32 mss = tp->mss_cache;
896 u32 fack, metric;
897
898 fack = tcp_highest_sack_seq(tp);
899 if (!before(low_seq, fack))
900 return;
901
902 metric = fack - low_seq;
903 if ((metric > tp->reordering * mss) && mss) {
904#if FASTRETRANS_DEBUG > 1
905 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
906 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
907 tp->reordering,
908 0,
909 tp->sacked_out,
910 tp->undo_marker ? tp->undo_retrans : 0);
911#endif
912 tp->reordering = min_t(u32, (metric + mss - 1) / mss,
913 sock_net(sk)->ipv4.sysctl_tcp_max_reordering);
914 }
915
916 /* This exciting event is worth to be remembered. 8) */
917 tp->reord_seen++;
918 NET_INC_STATS(sock_net(sk),
919 ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER);
920}
921
922/* This must be called before lost_out is incremented */
923static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
924{
925 if (!tp->retransmit_skb_hint ||
926 before(TCP_SKB_CB(skb)->seq,
927 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
928 tp->retransmit_skb_hint = skb;
929}
930
931/* Sum the number of packets on the wire we have marked as lost.
932 * There are two cases we care about here:
933 * a) Packet hasn't been marked lost (nor retransmitted),
934 * and this is the first loss.
935 * b) Packet has been marked both lost and retransmitted,
936 * and this means we think it was lost again.
937 */
938static void tcp_sum_lost(struct tcp_sock *tp, struct sk_buff *skb)
939{
940 __u8 sacked = TCP_SKB_CB(skb)->sacked;
941
942 if (!(sacked & TCPCB_LOST) ||
943 ((sacked & TCPCB_LOST) && (sacked & TCPCB_SACKED_RETRANS)))
944 tp->lost += tcp_skb_pcount(skb);
945}
946
947static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
948{
949 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
950 tcp_verify_retransmit_hint(tp, skb);
951
952 tp->lost_out += tcp_skb_pcount(skb);
953 tcp_sum_lost(tp, skb);
954 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
955 }
956}
957
958void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, struct sk_buff *skb)
959{
960 tcp_verify_retransmit_hint(tp, skb);
961
962 tcp_sum_lost(tp, skb);
963 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
964 tp->lost_out += tcp_skb_pcount(skb);
965 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
966 }
967}
968
969/* This procedure tags the retransmission queue when SACKs arrive.
970 *
971 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
972 * Packets in queue with these bits set are counted in variables
973 * sacked_out, retrans_out and lost_out, correspondingly.
974 *
975 * Valid combinations are:
976 * Tag InFlight Description
977 * 0 1 - orig segment is in flight.
978 * S 0 - nothing flies, orig reached receiver.
979 * L 0 - nothing flies, orig lost by net.
980 * R 2 - both orig and retransmit are in flight.
981 * L|R 1 - orig is lost, retransmit is in flight.
982 * S|R 1 - orig reached receiver, retrans is still in flight.
983 * (L|S|R is logically valid, it could occur when L|R is sacked,
984 * but it is equivalent to plain S and code short-curcuits it to S.
985 * L|S is logically invalid, it would mean -1 packet in flight 8))
986 *
987 * These 6 states form finite state machine, controlled by the following events:
988 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
989 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
990 * 3. Loss detection event of two flavors:
991 * A. Scoreboard estimator decided the packet is lost.
992 * A'. Reno "three dupacks" marks head of queue lost.
993 * B. SACK arrives sacking SND.NXT at the moment, when the
994 * segment was retransmitted.
995 * 4. D-SACK added new rule: D-SACK changes any tag to S.
996 *
997 * It is pleasant to note, that state diagram turns out to be commutative,
998 * so that we are allowed not to be bothered by order of our actions,
999 * when multiple events arrive simultaneously. (see the function below).
1000 *
1001 * Reordering detection.
1002 * --------------------
1003 * Reordering metric is maximal distance, which a packet can be displaced
1004 * in packet stream. With SACKs we can estimate it:
1005 *
1006 * 1. SACK fills old hole and the corresponding segment was not
1007 * ever retransmitted -> reordering. Alas, we cannot use it
1008 * when segment was retransmitted.
1009 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1010 * for retransmitted and already SACKed segment -> reordering..
1011 * Both of these heuristics are not used in Loss state, when we cannot
1012 * account for retransmits accurately.
1013 *
1014 * SACK block validation.
1015 * ----------------------
1016 *
1017 * SACK block range validation checks that the received SACK block fits to
1018 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1019 * Note that SND.UNA is not included to the range though being valid because
1020 * it means that the receiver is rather inconsistent with itself reporting
1021 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1022 * perfectly valid, however, in light of RFC2018 which explicitly states
1023 * that "SACK block MUST reflect the newest segment. Even if the newest
1024 * segment is going to be discarded ...", not that it looks very clever
1025 * in case of head skb. Due to potentional receiver driven attacks, we
1026 * choose to avoid immediate execution of a walk in write queue due to
1027 * reneging and defer head skb's loss recovery to standard loss recovery
1028 * procedure that will eventually trigger (nothing forbids us doing this).
1029 *
1030 * Implements also blockage to start_seq wrap-around. Problem lies in the
1031 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1032 * there's no guarantee that it will be before snd_nxt (n). The problem
1033 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1034 * wrap (s_w):
1035 *
1036 * <- outs wnd -> <- wrapzone ->
1037 * u e n u_w e_w s n_w
1038 * | | | | | | |
1039 * |<------------+------+----- TCP seqno space --------------+---------->|
1040 * ...-- <2^31 ->| |<--------...
1041 * ...---- >2^31 ------>| |<--------...
1042 *
1043 * Current code wouldn't be vulnerable but it's better still to discard such
1044 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1045 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1046 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1047 * equal to the ideal case (infinite seqno space without wrap caused issues).
1048 *
1049 * With D-SACK the lower bound is extended to cover sequence space below
1050 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1051 * again, D-SACK block must not to go across snd_una (for the same reason as
1052 * for the normal SACK blocks, explained above). But there all simplicity
1053 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1054 * fully below undo_marker they do not affect behavior in anyway and can
1055 * therefore be safely ignored. In rare cases (which are more or less
1056 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1057 * fragmentation and packet reordering past skb's retransmission. To consider
1058 * them correctly, the acceptable range must be extended even more though
1059 * the exact amount is rather hard to quantify. However, tp->max_window can
1060 * be used as an exaggerated estimate.
1061 */
1062static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1063 u32 start_seq, u32 end_seq)
1064{
1065 /* Too far in future, or reversed (interpretation is ambiguous) */
1066 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1067 return false;
1068
1069 /* Nasty start_seq wrap-around check (see comments above) */
1070 if (!before(start_seq, tp->snd_nxt))
1071 return false;
1072
1073 /* In outstanding window? ...This is valid exit for D-SACKs too.
1074 * start_seq == snd_una is non-sensical (see comments above)
1075 */
1076 if (after(start_seq, tp->snd_una))
1077 return true;
1078
1079 if (!is_dsack || !tp->undo_marker)
1080 return false;
1081
1082 /* ...Then it's D-SACK, and must reside below snd_una completely */
1083 if (after(end_seq, tp->snd_una))
1084 return false;
1085
1086 if (!before(start_seq, tp->undo_marker))
1087 return true;
1088
1089 /* Too old */
1090 if (!after(end_seq, tp->undo_marker))
1091 return false;
1092
1093 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1094 * start_seq < undo_marker and end_seq >= undo_marker.
1095 */
1096 return !before(start_seq, end_seq - tp->max_window);
1097}
1098
1099static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1100 struct tcp_sack_block_wire *sp, int num_sacks,
1101 u32 prior_snd_una)
1102{
1103 struct tcp_sock *tp = tcp_sk(sk);
1104 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1105 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1106 bool dup_sack = false;
1107
1108 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1109 dup_sack = true;
1110 tcp_dsack_seen(tp);
1111 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1112 } else if (num_sacks > 1) {
1113 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1114 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1115
1116 if (!after(end_seq_0, end_seq_1) &&
1117 !before(start_seq_0, start_seq_1)) {
1118 dup_sack = true;
1119 tcp_dsack_seen(tp);
1120 NET_INC_STATS(sock_net(sk),
1121 LINUX_MIB_TCPDSACKOFORECV);
1122 }
1123 }
1124
1125 /* D-SACK for already forgotten data... Do dumb counting. */
1126 if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 &&
1127 !after(end_seq_0, prior_snd_una) &&
1128 after(end_seq_0, tp->undo_marker))
1129 tp->undo_retrans--;
1130
1131 return dup_sack;
1132}
1133
1134struct tcp_sacktag_state {
1135 u32 reord;
1136 /* Timestamps for earliest and latest never-retransmitted segment
1137 * that was SACKed. RTO needs the earliest RTT to stay conservative,
1138 * but congestion control should still get an accurate delay signal.
1139 */
1140 u64 first_sackt;
1141 u64 last_sackt;
1142 struct rate_sample *rate;
1143 int flag;
1144 unsigned int mss_now;
1145};
1146
1147/* Check if skb is fully within the SACK block. In presence of GSO skbs,
1148 * the incoming SACK may not exactly match but we can find smaller MSS
1149 * aligned portion of it that matches. Therefore we might need to fragment
1150 * which may fail and creates some hassle (caller must handle error case
1151 * returns).
1152 *
1153 * FIXME: this could be merged to shift decision code
1154 */
1155static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1156 u32 start_seq, u32 end_seq)
1157{
1158 int err;
1159 bool in_sack;
1160 unsigned int pkt_len;
1161 unsigned int mss;
1162
1163 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1164 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1165
1166 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1167 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1168 mss = tcp_skb_mss(skb);
1169 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1170
1171 if (!in_sack) {
1172 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1173 if (pkt_len < mss)
1174 pkt_len = mss;
1175 } else {
1176 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1177 if (pkt_len < mss)
1178 return -EINVAL;
1179 }
1180
1181 /* Round if necessary so that SACKs cover only full MSSes
1182 * and/or the remaining small portion (if present)
1183 */
1184 if (pkt_len > mss) {
1185 unsigned int new_len = (pkt_len / mss) * mss;
1186 if (!in_sack && new_len < pkt_len)
1187 new_len += mss;
1188 pkt_len = new_len;
1189 }
1190
1191 if (pkt_len >= skb->len && !in_sack)
1192 return 0;
1193
1194 err = tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
1195 pkt_len, mss, GFP_ATOMIC);
1196 if (err < 0)
1197 return err;
1198 }
1199
1200 return in_sack;
1201}
1202
1203/* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1204static u8 tcp_sacktag_one(struct sock *sk,
1205 struct tcp_sacktag_state *state, u8 sacked,
1206 u32 start_seq, u32 end_seq,
1207 int dup_sack, int pcount,
1208 u64 xmit_time)
1209{
1210 struct tcp_sock *tp = tcp_sk(sk);
1211
1212 /* Account D-SACK for retransmitted packet. */
1213 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1214 if (tp->undo_marker && tp->undo_retrans > 0 &&
1215 after(end_seq, tp->undo_marker))
1216 tp->undo_retrans--;
1217 if ((sacked & TCPCB_SACKED_ACKED) &&
1218 before(start_seq, state->reord))
1219 state->reord = start_seq;
1220 }
1221
1222 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1223 if (!after(end_seq, tp->snd_una))
1224 return sacked;
1225
1226 if (!(sacked & TCPCB_SACKED_ACKED)) {
1227 tcp_rack_advance(tp, sacked, end_seq, xmit_time);
1228
1229 if (sacked & TCPCB_SACKED_RETRANS) {
1230 /* If the segment is not tagged as lost,
1231 * we do not clear RETRANS, believing
1232 * that retransmission is still in flight.
1233 */
1234 if (sacked & TCPCB_LOST) {
1235 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1236 tp->lost_out -= pcount;
1237 tp->retrans_out -= pcount;
1238 }
1239 } else {
1240 if (!(sacked & TCPCB_RETRANS)) {
1241 /* New sack for not retransmitted frame,
1242 * which was in hole. It is reordering.
1243 */
1244 if (before(start_seq,
1245 tcp_highest_sack_seq(tp)) &&
1246 before(start_seq, state->reord))
1247 state->reord = start_seq;
1248
1249 if (!after(end_seq, tp->high_seq))
1250 state->flag |= FLAG_ORIG_SACK_ACKED;
1251 if (state->first_sackt == 0)
1252 state->first_sackt = xmit_time;
1253 state->last_sackt = xmit_time;
1254 }
1255
1256 if (sacked & TCPCB_LOST) {
1257 sacked &= ~TCPCB_LOST;
1258 tp->lost_out -= pcount;
1259 }
1260 }
1261
1262 sacked |= TCPCB_SACKED_ACKED;
1263 state->flag |= FLAG_DATA_SACKED;
1264 tp->sacked_out += pcount;
1265 tp->delivered += pcount; /* Out-of-order packets delivered */
1266
1267 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1268 if (tp->lost_skb_hint &&
1269 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1270 tp->lost_cnt_hint += pcount;
1271 }
1272
1273 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1274 * frames and clear it. undo_retrans is decreased above, L|R frames
1275 * are accounted above as well.
1276 */
1277 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1278 sacked &= ~TCPCB_SACKED_RETRANS;
1279 tp->retrans_out -= pcount;
1280 }
1281
1282 return sacked;
1283}
1284
1285/* Shift newly-SACKed bytes from this skb to the immediately previous
1286 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1287 */
1288static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev,
1289 struct sk_buff *skb,
1290 struct tcp_sacktag_state *state,
1291 unsigned int pcount, int shifted, int mss,
1292 bool dup_sack)
1293{
1294 struct tcp_sock *tp = tcp_sk(sk);
1295 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1296 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1297
1298 BUG_ON(!pcount);
1299
1300 /* Adjust counters and hints for the newly sacked sequence
1301 * range but discard the return value since prev is already
1302 * marked. We must tag the range first because the seq
1303 * advancement below implicitly advances
1304 * tcp_highest_sack_seq() when skb is highest_sack.
1305 */
1306 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1307 start_seq, end_seq, dup_sack, pcount,
1308 skb->skb_mstamp);
1309 tcp_rate_skb_delivered(sk, skb, state->rate);
1310
1311 if (skb == tp->lost_skb_hint)
1312 tp->lost_cnt_hint += pcount;
1313
1314 TCP_SKB_CB(prev)->end_seq += shifted;
1315 TCP_SKB_CB(skb)->seq += shifted;
1316
1317 tcp_skb_pcount_add(prev, pcount);
1318 BUG_ON(tcp_skb_pcount(skb) < pcount);
1319 tcp_skb_pcount_add(skb, -pcount);
1320
1321 /* When we're adding to gso_segs == 1, gso_size will be zero,
1322 * in theory this shouldn't be necessary but as long as DSACK
1323 * code can come after this skb later on it's better to keep
1324 * setting gso_size to something.
1325 */
1326 if (!TCP_SKB_CB(prev)->tcp_gso_size)
1327 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1328
1329 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1330 if (tcp_skb_pcount(skb) <= 1)
1331 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1332
1333 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1334 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1335
1336 if (skb->len > 0) {
1337 BUG_ON(!tcp_skb_pcount(skb));
1338 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1339 return false;
1340 }
1341
1342 /* Whole SKB was eaten :-) */
1343
1344 if (skb == tp->retransmit_skb_hint)
1345 tp->retransmit_skb_hint = prev;
1346 if (skb == tp->lost_skb_hint) {
1347 tp->lost_skb_hint = prev;
1348 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1349 }
1350
1351 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1352 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor;
1353 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1354 TCP_SKB_CB(prev)->end_seq++;
1355
1356 if (skb == tcp_highest_sack(sk))
1357 tcp_advance_highest_sack(sk, skb);
1358
1359 tcp_skb_collapse_tstamp(prev, skb);
1360 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp))
1361 TCP_SKB_CB(prev)->tx.delivered_mstamp = 0;
1362
1363 tcp_rtx_queue_unlink_and_free(skb, sk);
1364
1365 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED);
1366
1367 return true;
1368}
1369
1370/* I wish gso_size would have a bit more sane initialization than
1371 * something-or-zero which complicates things
1372 */
1373static int tcp_skb_seglen(const struct sk_buff *skb)
1374{
1375 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1376}
1377
1378/* Shifting pages past head area doesn't work */
1379static int skb_can_shift(const struct sk_buff *skb)
1380{
1381 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1382}
1383
1384/* Try collapsing SACK blocks spanning across multiple skbs to a single
1385 * skb.
1386 */
1387static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1388 struct tcp_sacktag_state *state,
1389 u32 start_seq, u32 end_seq,
1390 bool dup_sack)
1391{
1392 struct tcp_sock *tp = tcp_sk(sk);
1393 struct sk_buff *prev;
1394 int mss;
1395 int pcount = 0;
1396 int len;
1397 int in_sack;
1398
1399 /* Normally R but no L won't result in plain S */
1400 if (!dup_sack &&
1401 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1402 goto fallback;
1403 if (!skb_can_shift(skb))
1404 goto fallback;
1405 /* This frame is about to be dropped (was ACKed). */
1406 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1407 goto fallback;
1408
1409 /* Can only happen with delayed DSACK + discard craziness */
1410 prev = skb_rb_prev(skb);
1411 if (!prev)
1412 goto fallback;
1413
1414 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1415 goto fallback;
1416
1417 if (!tcp_skb_can_collapse_to(prev))
1418 goto fallback;
1419
1420 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1421 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1422
1423 if (in_sack) {
1424 len = skb->len;
1425 pcount = tcp_skb_pcount(skb);
1426 mss = tcp_skb_seglen(skb);
1427
1428 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1429 * drop this restriction as unnecessary
1430 */
1431 if (mss != tcp_skb_seglen(prev))
1432 goto fallback;
1433 } else {
1434 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1435 goto noop;
1436 /* CHECKME: This is non-MSS split case only?, this will
1437 * cause skipped skbs due to advancing loop btw, original
1438 * has that feature too
1439 */
1440 if (tcp_skb_pcount(skb) <= 1)
1441 goto noop;
1442
1443 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1444 if (!in_sack) {
1445 /* TODO: head merge to next could be attempted here
1446 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1447 * though it might not be worth of the additional hassle
1448 *
1449 * ...we can probably just fallback to what was done
1450 * previously. We could try merging non-SACKed ones
1451 * as well but it probably isn't going to buy off
1452 * because later SACKs might again split them, and
1453 * it would make skb timestamp tracking considerably
1454 * harder problem.
1455 */
1456 goto fallback;
1457 }
1458
1459 len = end_seq - TCP_SKB_CB(skb)->seq;
1460 BUG_ON(len < 0);
1461 BUG_ON(len > skb->len);
1462
1463 /* MSS boundaries should be honoured or else pcount will
1464 * severely break even though it makes things bit trickier.
1465 * Optimize common case to avoid most of the divides
1466 */
1467 mss = tcp_skb_mss(skb);
1468
1469 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1470 * drop this restriction as unnecessary
1471 */
1472 if (mss != tcp_skb_seglen(prev))
1473 goto fallback;
1474
1475 if (len == mss) {
1476 pcount = 1;
1477 } else if (len < mss) {
1478 goto noop;
1479 } else {
1480 pcount = len / mss;
1481 len = pcount * mss;
1482 }
1483 }
1484
1485 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1486 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1487 goto fallback;
1488
1489 if (!skb_shift(prev, skb, len))
1490 goto fallback;
1491 if (!tcp_shifted_skb(sk, prev, skb, state, pcount, len, mss, dup_sack))
1492 goto out;
1493
1494 /* Hole filled allows collapsing with the next as well, this is very
1495 * useful when hole on every nth skb pattern happens
1496 */
1497 skb = skb_rb_next(prev);
1498 if (!skb)
1499 goto out;
1500
1501 if (!skb_can_shift(skb) ||
1502 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1503 (mss != tcp_skb_seglen(skb)))
1504 goto out;
1505
1506 len = skb->len;
1507 if (skb_shift(prev, skb, len)) {
1508 pcount += tcp_skb_pcount(skb);
1509 tcp_shifted_skb(sk, prev, skb, state, tcp_skb_pcount(skb),
1510 len, mss, 0);
1511 }
1512
1513out:
1514 return prev;
1515
1516noop:
1517 return skb;
1518
1519fallback:
1520 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1521 return NULL;
1522}
1523
1524static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1525 struct tcp_sack_block *next_dup,
1526 struct tcp_sacktag_state *state,
1527 u32 start_seq, u32 end_seq,
1528 bool dup_sack_in)
1529{
1530 struct tcp_sock *tp = tcp_sk(sk);
1531 struct sk_buff *tmp;
1532
1533 skb_rbtree_walk_from(skb) {
1534 int in_sack = 0;
1535 bool dup_sack = dup_sack_in;
1536
1537 /* queue is in-order => we can short-circuit the walk early */
1538 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1539 break;
1540
1541 if (next_dup &&
1542 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1543 in_sack = tcp_match_skb_to_sack(sk, skb,
1544 next_dup->start_seq,
1545 next_dup->end_seq);
1546 if (in_sack > 0)
1547 dup_sack = true;
1548 }
1549
1550 /* skb reference here is a bit tricky to get right, since
1551 * shifting can eat and free both this skb and the next,
1552 * so not even _safe variant of the loop is enough.
1553 */
1554 if (in_sack <= 0) {
1555 tmp = tcp_shift_skb_data(sk, skb, state,
1556 start_seq, end_seq, dup_sack);
1557 if (tmp) {
1558 if (tmp != skb) {
1559 skb = tmp;
1560 continue;
1561 }
1562
1563 in_sack = 0;
1564 } else {
1565 in_sack = tcp_match_skb_to_sack(sk, skb,
1566 start_seq,
1567 end_seq);
1568 }
1569 }
1570
1571 if (unlikely(in_sack < 0))
1572 break;
1573
1574 if (in_sack) {
1575 TCP_SKB_CB(skb)->sacked =
1576 tcp_sacktag_one(sk,
1577 state,
1578 TCP_SKB_CB(skb)->sacked,
1579 TCP_SKB_CB(skb)->seq,
1580 TCP_SKB_CB(skb)->end_seq,
1581 dup_sack,
1582 tcp_skb_pcount(skb),
1583 skb->skb_mstamp);
1584 tcp_rate_skb_delivered(sk, skb, state->rate);
1585 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1586 list_del_init(&skb->tcp_tsorted_anchor);
1587
1588 if (!before(TCP_SKB_CB(skb)->seq,
1589 tcp_highest_sack_seq(tp)))
1590 tcp_advance_highest_sack(sk, skb);
1591 }
1592 }
1593 return skb;
1594}
1595
1596static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk,
1597 struct tcp_sacktag_state *state,
1598 u32 seq)
1599{
1600 struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node;
1601 struct sk_buff *skb;
1602
1603 while (*p) {
1604 parent = *p;
1605 skb = rb_to_skb(parent);
1606 if (before(seq, TCP_SKB_CB(skb)->seq)) {
1607 p = &parent->rb_left;
1608 continue;
1609 }
1610 if (!before(seq, TCP_SKB_CB(skb)->end_seq)) {
1611 p = &parent->rb_right;
1612 continue;
1613 }
1614 return skb;
1615 }
1616 return NULL;
1617}
1618
1619static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1620 struct tcp_sacktag_state *state,
1621 u32 skip_to_seq)
1622{
1623 if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq))
1624 return skb;
1625
1626 return tcp_sacktag_bsearch(sk, state, skip_to_seq);
1627}
1628
1629static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1630 struct sock *sk,
1631 struct tcp_sack_block *next_dup,
1632 struct tcp_sacktag_state *state,
1633 u32 skip_to_seq)
1634{
1635 if (!next_dup)
1636 return skb;
1637
1638 if (before(next_dup->start_seq, skip_to_seq)) {
1639 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1640 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1641 next_dup->start_seq, next_dup->end_seq,
1642 1);
1643 }
1644
1645 return skb;
1646}
1647
1648static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1649{
1650 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1651}
1652
1653static int
1654tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1655 u32 prior_snd_una, struct tcp_sacktag_state *state)
1656{
1657 struct tcp_sock *tp = tcp_sk(sk);
1658 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1659 TCP_SKB_CB(ack_skb)->sacked);
1660 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1661 struct tcp_sack_block sp[TCP_NUM_SACKS];
1662 struct tcp_sack_block *cache;
1663 struct sk_buff *skb;
1664 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1665 int used_sacks;
1666 bool found_dup_sack = false;
1667 int i, j;
1668 int first_sack_index;
1669
1670 state->flag = 0;
1671 state->reord = tp->snd_nxt;
1672
1673 if (!tp->sacked_out)
1674 tcp_highest_sack_reset(sk);
1675
1676 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1677 num_sacks, prior_snd_una);
1678 if (found_dup_sack) {
1679 state->flag |= FLAG_DSACKING_ACK;
1680 tp->delivered++; /* A spurious retransmission is delivered */
1681 }
1682
1683 /* Eliminate too old ACKs, but take into
1684 * account more or less fresh ones, they can
1685 * contain valid SACK info.
1686 */
1687 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1688 return 0;
1689
1690 if (!tp->packets_out)
1691 goto out;
1692
1693 used_sacks = 0;
1694 first_sack_index = 0;
1695 for (i = 0; i < num_sacks; i++) {
1696 bool dup_sack = !i && found_dup_sack;
1697
1698 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1699 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1700
1701 if (!tcp_is_sackblock_valid(tp, dup_sack,
1702 sp[used_sacks].start_seq,
1703 sp[used_sacks].end_seq)) {
1704 int mib_idx;
1705
1706 if (dup_sack) {
1707 if (!tp->undo_marker)
1708 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1709 else
1710 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1711 } else {
1712 /* Don't count olds caused by ACK reordering */
1713 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1714 !after(sp[used_sacks].end_seq, tp->snd_una))
1715 continue;
1716 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1717 }
1718
1719 NET_INC_STATS(sock_net(sk), mib_idx);
1720 if (i == 0)
1721 first_sack_index = -1;
1722 continue;
1723 }
1724
1725 /* Ignore very old stuff early */
1726 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1727 continue;
1728
1729 used_sacks++;
1730 }
1731
1732 /* order SACK blocks to allow in order walk of the retrans queue */
1733 for (i = used_sacks - 1; i > 0; i--) {
1734 for (j = 0; j < i; j++) {
1735 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1736 swap(sp[j], sp[j + 1]);
1737
1738 /* Track where the first SACK block goes to */
1739 if (j == first_sack_index)
1740 first_sack_index = j + 1;
1741 }
1742 }
1743 }
1744
1745 state->mss_now = tcp_current_mss(sk);
1746 skb = NULL;
1747 i = 0;
1748
1749 if (!tp->sacked_out) {
1750 /* It's already past, so skip checking against it */
1751 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1752 } else {
1753 cache = tp->recv_sack_cache;
1754 /* Skip empty blocks in at head of the cache */
1755 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1756 !cache->end_seq)
1757 cache++;
1758 }
1759
1760 while (i < used_sacks) {
1761 u32 start_seq = sp[i].start_seq;
1762 u32 end_seq = sp[i].end_seq;
1763 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1764 struct tcp_sack_block *next_dup = NULL;
1765
1766 if (found_dup_sack && ((i + 1) == first_sack_index))
1767 next_dup = &sp[i + 1];
1768
1769 /* Skip too early cached blocks */
1770 while (tcp_sack_cache_ok(tp, cache) &&
1771 !before(start_seq, cache->end_seq))
1772 cache++;
1773
1774 /* Can skip some work by looking recv_sack_cache? */
1775 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1776 after(end_seq, cache->start_seq)) {
1777
1778 /* Head todo? */
1779 if (before(start_seq, cache->start_seq)) {
1780 skb = tcp_sacktag_skip(skb, sk, state,
1781 start_seq);
1782 skb = tcp_sacktag_walk(skb, sk, next_dup,
1783 state,
1784 start_seq,
1785 cache->start_seq,
1786 dup_sack);
1787 }
1788
1789 /* Rest of the block already fully processed? */
1790 if (!after(end_seq, cache->end_seq))
1791 goto advance_sp;
1792
1793 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1794 state,
1795 cache->end_seq);
1796
1797 /* ...tail remains todo... */
1798 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1799 /* ...but better entrypoint exists! */
1800 skb = tcp_highest_sack(sk);
1801 if (!skb)
1802 break;
1803 cache++;
1804 goto walk;
1805 }
1806
1807 skb = tcp_sacktag_skip(skb, sk, state, cache->end_seq);
1808 /* Check overlap against next cached too (past this one already) */
1809 cache++;
1810 continue;
1811 }
1812
1813 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1814 skb = tcp_highest_sack(sk);
1815 if (!skb)
1816 break;
1817 }
1818 skb = tcp_sacktag_skip(skb, sk, state, start_seq);
1819
1820walk:
1821 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
1822 start_seq, end_seq, dup_sack);
1823
1824advance_sp:
1825 i++;
1826 }
1827
1828 /* Clear the head of the cache sack blocks so we can skip it next time */
1829 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1830 tp->recv_sack_cache[i].start_seq = 0;
1831 tp->recv_sack_cache[i].end_seq = 0;
1832 }
1833 for (j = 0; j < used_sacks; j++)
1834 tp->recv_sack_cache[i++] = sp[j];
1835
1836 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss || tp->undo_marker)
1837 tcp_check_sack_reordering(sk, state->reord, 0);
1838
1839 tcp_verify_left_out(tp);
1840out:
1841
1842#if FASTRETRANS_DEBUG > 0
1843 WARN_ON((int)tp->sacked_out < 0);
1844 WARN_ON((int)tp->lost_out < 0);
1845 WARN_ON((int)tp->retrans_out < 0);
1846 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1847#endif
1848 return state->flag;
1849}
1850
1851/* Limits sacked_out so that sum with lost_out isn't ever larger than
1852 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1853 */
1854static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1855{
1856 u32 holes;
1857
1858 holes = max(tp->lost_out, 1U);
1859 holes = min(holes, tp->packets_out);
1860
1861 if ((tp->sacked_out + holes) > tp->packets_out) {
1862 tp->sacked_out = tp->packets_out - holes;
1863 return true;
1864 }
1865 return false;
1866}
1867
1868/* If we receive more dupacks than we expected counting segments
1869 * in assumption of absent reordering, interpret this as reordering.
1870 * The only another reason could be bug in receiver TCP.
1871 */
1872static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1873{
1874 struct tcp_sock *tp = tcp_sk(sk);
1875
1876 if (!tcp_limit_reno_sacked(tp))
1877 return;
1878
1879 tp->reordering = min_t(u32, tp->packets_out + addend,
1880 sock_net(sk)->ipv4.sysctl_tcp_max_reordering);
1881 tp->reord_seen++;
1882 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER);
1883}
1884
1885/* Emulate SACKs for SACKless connection: account for a new dupack. */
1886
1887static void tcp_add_reno_sack(struct sock *sk)
1888{
1889 struct tcp_sock *tp = tcp_sk(sk);
1890 u32 prior_sacked = tp->sacked_out;
1891
1892 tp->sacked_out++;
1893 tcp_check_reno_reordering(sk, 0);
1894 if (tp->sacked_out > prior_sacked)
1895 tp->delivered++; /* Some out-of-order packet is delivered */
1896 tcp_verify_left_out(tp);
1897}
1898
1899/* Account for ACK, ACKing some data in Reno Recovery phase. */
1900
1901static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1902{
1903 struct tcp_sock *tp = tcp_sk(sk);
1904
1905 if (acked > 0) {
1906 /* One ACK acked hole. The rest eat duplicate ACKs. */
1907 tp->delivered += max_t(int, acked - tp->sacked_out, 1);
1908 if (acked - 1 >= tp->sacked_out)
1909 tp->sacked_out = 0;
1910 else
1911 tp->sacked_out -= acked - 1;
1912 }
1913 tcp_check_reno_reordering(sk, acked);
1914 tcp_verify_left_out(tp);
1915}
1916
1917static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1918{
1919 tp->sacked_out = 0;
1920}
1921
1922void tcp_clear_retrans(struct tcp_sock *tp)
1923{
1924 tp->retrans_out = 0;
1925 tp->lost_out = 0;
1926 tp->undo_marker = 0;
1927 tp->undo_retrans = -1;
1928 tp->sacked_out = 0;
1929}
1930
1931static inline void tcp_init_undo(struct tcp_sock *tp)
1932{
1933 tp->undo_marker = tp->snd_una;
1934 /* Retransmission still in flight may cause DSACKs later. */
1935 tp->undo_retrans = tp->retrans_out ? : -1;
1936}
1937
1938static bool tcp_is_rack(const struct sock *sk)
1939{
1940 return sock_net(sk)->ipv4.sysctl_tcp_recovery & TCP_RACK_LOSS_DETECTION;
1941}
1942
1943/* If we detect SACK reneging, forget all SACK information
1944 * and reset tags completely, otherwise preserve SACKs. If receiver
1945 * dropped its ofo queue, we will know this due to reneging detection.
1946 */
1947static void tcp_timeout_mark_lost(struct sock *sk)
1948{
1949 struct tcp_sock *tp = tcp_sk(sk);
1950 struct sk_buff *skb, *head;
1951 bool is_reneg; /* is receiver reneging on SACKs? */
1952
1953 head = tcp_rtx_queue_head(sk);
1954 is_reneg = head && (TCP_SKB_CB(head)->sacked & TCPCB_SACKED_ACKED);
1955 if (is_reneg) {
1956 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
1957 tp->sacked_out = 0;
1958 /* Mark SACK reneging until we recover from this loss event. */
1959 tp->is_sack_reneg = 1;
1960 } else if (tcp_is_reno(tp)) {
1961 tcp_reset_reno_sack(tp);
1962 }
1963
1964 skb = head;
1965 skb_rbtree_walk_from(skb) {
1966 if (is_reneg)
1967 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1968 else if (tcp_is_rack(sk) && skb != head &&
1969 tcp_rack_skb_timeout(tp, skb, 0) > 0)
1970 continue; /* Don't mark recently sent ones lost yet */
1971 tcp_mark_skb_lost(sk, skb);
1972 }
1973 tcp_verify_left_out(tp);
1974 tcp_clear_all_retrans_hints(tp);
1975}
1976
1977/* Enter Loss state. */
1978void tcp_enter_loss(struct sock *sk)
1979{
1980 const struct inet_connection_sock *icsk = inet_csk(sk);
1981 struct tcp_sock *tp = tcp_sk(sk);
1982 struct net *net = sock_net(sk);
1983 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
1984
1985 tcp_timeout_mark_lost(sk);
1986
1987 /* Reduce ssthresh if it has not yet been made inside this window. */
1988 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1989 !after(tp->high_seq, tp->snd_una) ||
1990 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1991 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1992 tp->prior_cwnd = tp->snd_cwnd;
1993 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1994 tcp_ca_event(sk, CA_EVENT_LOSS);
1995 tcp_init_undo(tp);
1996 }
1997 tp->snd_cwnd = tcp_packets_in_flight(tp) + 1;
1998 tp->snd_cwnd_cnt = 0;
1999 tp->snd_cwnd_stamp = tcp_jiffies32;
2000
2001 /* Timeout in disordered state after receiving substantial DUPACKs
2002 * suggests that the degree of reordering is over-estimated.
2003 */
2004 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
2005 tp->sacked_out >= net->ipv4.sysctl_tcp_reordering)
2006 tp->reordering = min_t(unsigned int, tp->reordering,
2007 net->ipv4.sysctl_tcp_reordering);
2008 tcp_set_ca_state(sk, TCP_CA_Loss);
2009 tp->high_seq = tp->snd_nxt;
2010 tcp_ecn_queue_cwr(tp);
2011
2012 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
2013 * loss recovery is underway except recurring timeout(s) on
2014 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
2015 */
2016 tp->frto = net->ipv4.sysctl_tcp_frto &&
2017 (new_recovery || icsk->icsk_retransmits) &&
2018 !inet_csk(sk)->icsk_mtup.probe_size;
2019}
2020
2021/* If ACK arrived pointing to a remembered SACK, it means that our
2022 * remembered SACKs do not reflect real state of receiver i.e.
2023 * receiver _host_ is heavily congested (or buggy).
2024 *
2025 * To avoid big spurious retransmission bursts due to transient SACK
2026 * scoreboard oddities that look like reneging, we give the receiver a
2027 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2028 * restore sanity to the SACK scoreboard. If the apparent reneging
2029 * persists until this RTO then we'll clear the SACK scoreboard.
2030 */
2031static bool tcp_check_sack_reneging(struct sock *sk, int flag)
2032{
2033 if (flag & FLAG_SACK_RENEGING) {
2034 struct tcp_sock *tp = tcp_sk(sk);
2035 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2036 msecs_to_jiffies(10));
2037
2038 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2039 delay, TCP_RTO_MAX);
2040 return true;
2041 }
2042 return false;
2043}
2044
2045/* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2046 * counter when SACK is enabled (without SACK, sacked_out is used for
2047 * that purpose).
2048 *
2049 * With reordering, holes may still be in flight, so RFC3517 recovery
2050 * uses pure sacked_out (total number of SACKed segments) even though
2051 * it violates the RFC that uses duplicate ACKs, often these are equal
2052 * but when e.g. out-of-window ACKs or packet duplication occurs,
2053 * they differ. Since neither occurs due to loss, TCP should really
2054 * ignore them.
2055 */
2056static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2057{
2058 return tp->sacked_out + 1;
2059}
2060
2061/* Linux NewReno/SACK/ECN state machine.
2062 * --------------------------------------
2063 *
2064 * "Open" Normal state, no dubious events, fast path.
2065 * "Disorder" In all the respects it is "Open",
2066 * but requires a bit more attention. It is entered when
2067 * we see some SACKs or dupacks. It is split of "Open"
2068 * mainly to move some processing from fast path to slow one.
2069 * "CWR" CWND was reduced due to some Congestion Notification event.
2070 * It can be ECN, ICMP source quench, local device congestion.
2071 * "Recovery" CWND was reduced, we are fast-retransmitting.
2072 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2073 *
2074 * tcp_fastretrans_alert() is entered:
2075 * - each incoming ACK, if state is not "Open"
2076 * - when arrived ACK is unusual, namely:
2077 * * SACK
2078 * * Duplicate ACK.
2079 * * ECN ECE.
2080 *
2081 * Counting packets in flight is pretty simple.
2082 *
2083 * in_flight = packets_out - left_out + retrans_out
2084 *
2085 * packets_out is SND.NXT-SND.UNA counted in packets.
2086 *
2087 * retrans_out is number of retransmitted segments.
2088 *
2089 * left_out is number of segments left network, but not ACKed yet.
2090 *
2091 * left_out = sacked_out + lost_out
2092 *
2093 * sacked_out: Packets, which arrived to receiver out of order
2094 * and hence not ACKed. With SACKs this number is simply
2095 * amount of SACKed data. Even without SACKs
2096 * it is easy to give pretty reliable estimate of this number,
2097 * counting duplicate ACKs.
2098 *
2099 * lost_out: Packets lost by network. TCP has no explicit
2100 * "loss notification" feedback from network (for now).
2101 * It means that this number can be only _guessed_.
2102 * Actually, it is the heuristics to predict lossage that
2103 * distinguishes different algorithms.
2104 *
2105 * F.e. after RTO, when all the queue is considered as lost,
2106 * lost_out = packets_out and in_flight = retrans_out.
2107 *
2108 * Essentially, we have now a few algorithms detecting
2109 * lost packets.
2110 *
2111 * If the receiver supports SACK:
2112 *
2113 * RFC6675/3517: It is the conventional algorithm. A packet is
2114 * considered lost if the number of higher sequence packets
2115 * SACKed is greater than or equal the DUPACK thoreshold
2116 * (reordering). This is implemented in tcp_mark_head_lost and
2117 * tcp_update_scoreboard.
2118 *
2119 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm
2120 * (2017-) that checks timing instead of counting DUPACKs.
2121 * Essentially a packet is considered lost if it's not S/ACKed
2122 * after RTT + reordering_window, where both metrics are
2123 * dynamically measured and adjusted. This is implemented in
2124 * tcp_rack_mark_lost.
2125 *
2126 * If the receiver does not support SACK:
2127 *
2128 * NewReno (RFC6582): in Recovery we assume that one segment
2129 * is lost (classic Reno). While we are in Recovery and
2130 * a partial ACK arrives, we assume that one more packet
2131 * is lost (NewReno). This heuristics are the same in NewReno
2132 * and SACK.
2133 *
2134 * Really tricky (and requiring careful tuning) part of algorithm
2135 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2136 * The first determines the moment _when_ we should reduce CWND and,
2137 * hence, slow down forward transmission. In fact, it determines the moment
2138 * when we decide that hole is caused by loss, rather than by a reorder.
2139 *
2140 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2141 * holes, caused by lost packets.
2142 *
2143 * And the most logically complicated part of algorithm is undo
2144 * heuristics. We detect false retransmits due to both too early
2145 * fast retransmit (reordering) and underestimated RTO, analyzing
2146 * timestamps and D-SACKs. When we detect that some segments were
2147 * retransmitted by mistake and CWND reduction was wrong, we undo
2148 * window reduction and abort recovery phase. This logic is hidden
2149 * inside several functions named tcp_try_undo_<something>.
2150 */
2151
2152/* This function decides, when we should leave Disordered state
2153 * and enter Recovery phase, reducing congestion window.
2154 *
2155 * Main question: may we further continue forward transmission
2156 * with the same cwnd?
2157 */
2158static bool tcp_time_to_recover(struct sock *sk, int flag)
2159{
2160 struct tcp_sock *tp = tcp_sk(sk);
2161
2162 /* Trick#1: The loss is proven. */
2163 if (tp->lost_out)
2164 return true;
2165
2166 /* Not-A-Trick#2 : Classic rule... */
2167 if (!tcp_is_rack(sk) && tcp_dupack_heuristics(tp) > tp->reordering)
2168 return true;
2169
2170 return false;
2171}
2172
2173/* Detect loss in event "A" above by marking head of queue up as lost.
2174 * For non-SACK(Reno) senders, the first "packets" number of segments
2175 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2176 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2177 * the maximum SACKed segments to pass before reaching this limit.
2178 */
2179static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2180{
2181 struct tcp_sock *tp = tcp_sk(sk);
2182 struct sk_buff *skb;
2183 int cnt, oldcnt, lost;
2184 unsigned int mss;
2185 /* Use SACK to deduce losses of new sequences sent during recovery */
2186 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq;
2187
2188 WARN_ON(packets > tp->packets_out);
2189 skb = tp->lost_skb_hint;
2190 if (skb) {
2191 /* Head already handled? */
2192 if (mark_head && after(TCP_SKB_CB(skb)->seq, tp->snd_una))
2193 return;
2194 cnt = tp->lost_cnt_hint;
2195 } else {
2196 skb = tcp_rtx_queue_head(sk);
2197 cnt = 0;
2198 }
2199
2200 skb_rbtree_walk_from(skb) {
2201 /* TODO: do this better */
2202 /* this is not the most efficient way to do this... */
2203 tp->lost_skb_hint = skb;
2204 tp->lost_cnt_hint = cnt;
2205
2206 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2207 break;
2208
2209 oldcnt = cnt;
2210 if (tcp_is_reno(tp) ||
2211 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2212 cnt += tcp_skb_pcount(skb);
2213
2214 if (cnt > packets) {
2215 if (tcp_is_sack(tp) ||
2216 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2217 (oldcnt >= packets))
2218 break;
2219
2220 mss = tcp_skb_mss(skb);
2221 /* If needed, chop off the prefix to mark as lost. */
2222 lost = (packets - oldcnt) * mss;
2223 if (lost < skb->len &&
2224 tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2225 lost, mss, GFP_ATOMIC) < 0)
2226 break;
2227 cnt = packets;
2228 }
2229
2230 tcp_skb_mark_lost(tp, skb);
2231
2232 if (mark_head)
2233 break;
2234 }
2235 tcp_verify_left_out(tp);
2236}
2237
2238/* Account newly detected lost packet(s) */
2239
2240static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2241{
2242 struct tcp_sock *tp = tcp_sk(sk);
2243
2244 if (tcp_is_sack(tp)) {
2245 int sacked_upto = tp->sacked_out - tp->reordering;
2246 if (sacked_upto >= 0)
2247 tcp_mark_head_lost(sk, sacked_upto, 0);
2248 else if (fast_rexmit)
2249 tcp_mark_head_lost(sk, 1, 1);
2250 }
2251}
2252
2253static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2254{
2255 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2256 before(tp->rx_opt.rcv_tsecr, when);
2257}
2258
2259/* skb is spurious retransmitted if the returned timestamp echo
2260 * reply is prior to the skb transmission time
2261 */
2262static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2263 const struct sk_buff *skb)
2264{
2265 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2266 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp(skb));
2267}
2268
2269/* Nothing was retransmitted or returned timestamp is less
2270 * than timestamp of the first retransmission.
2271 */
2272static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2273{
2274 return !tp->retrans_stamp ||
2275 tcp_tsopt_ecr_before(tp, tp->retrans_stamp);
2276}
2277
2278/* Undo procedures. */
2279
2280/* We can clear retrans_stamp when there are no retransmissions in the
2281 * window. It would seem that it is trivially available for us in
2282 * tp->retrans_out, however, that kind of assumptions doesn't consider
2283 * what will happen if errors occur when sending retransmission for the
2284 * second time. ...It could the that such segment has only
2285 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2286 * the head skb is enough except for some reneging corner cases that
2287 * are not worth the effort.
2288 *
2289 * Main reason for all this complexity is the fact that connection dying
2290 * time now depends on the validity of the retrans_stamp, in particular,
2291 * that successive retransmissions of a segment must not advance
2292 * retrans_stamp under any conditions.
2293 */
2294static bool tcp_any_retrans_done(const struct sock *sk)
2295{
2296 const struct tcp_sock *tp = tcp_sk(sk);
2297 struct sk_buff *skb;
2298
2299 if (tp->retrans_out)
2300 return true;
2301
2302 skb = tcp_rtx_queue_head(sk);
2303 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2304 return true;
2305
2306 return false;
2307}
2308
2309static void DBGUNDO(struct sock *sk, const char *msg)
2310{
2311#if FASTRETRANS_DEBUG > 1
2312 struct tcp_sock *tp = tcp_sk(sk);
2313 struct inet_sock *inet = inet_sk(sk);
2314
2315 if (sk->sk_family == AF_INET) {
2316 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2317 msg,
2318 &inet->inet_daddr, ntohs(inet->inet_dport),
2319 tp->snd_cwnd, tcp_left_out(tp),
2320 tp->snd_ssthresh, tp->prior_ssthresh,
2321 tp->packets_out);
2322 }
2323#if IS_ENABLED(CONFIG_IPV6)
2324 else if (sk->sk_family == AF_INET6) {
2325 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2326 msg,
2327 &sk->sk_v6_daddr, ntohs(inet->inet_dport),
2328 tp->snd_cwnd, tcp_left_out(tp),
2329 tp->snd_ssthresh, tp->prior_ssthresh,
2330 tp->packets_out);
2331 }
2332#endif
2333#endif
2334}
2335
2336static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2337{
2338 struct tcp_sock *tp = tcp_sk(sk);
2339
2340 if (unmark_loss) {
2341 struct sk_buff *skb;
2342
2343 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2344 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2345 }
2346 tp->lost_out = 0;
2347 tcp_clear_all_retrans_hints(tp);
2348 }
2349
2350 if (tp->prior_ssthresh) {
2351 const struct inet_connection_sock *icsk = inet_csk(sk);
2352
2353 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2354
2355 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2356 tp->snd_ssthresh = tp->prior_ssthresh;
2357 tcp_ecn_withdraw_cwr(tp);
2358 }
2359 }
2360 tp->snd_cwnd_stamp = tcp_jiffies32;
2361 tp->undo_marker = 0;
2362 tp->rack.advanced = 1; /* Force RACK to re-exam losses */
2363}
2364
2365static inline bool tcp_may_undo(const struct tcp_sock *tp)
2366{
2367 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2368}
2369
2370/* People celebrate: "We love our President!" */
2371static bool tcp_try_undo_recovery(struct sock *sk)
2372{
2373 struct tcp_sock *tp = tcp_sk(sk);
2374
2375 if (tcp_may_undo(tp)) {
2376 int mib_idx;
2377
2378 /* Happy end! We did not retransmit anything
2379 * or our original transmission succeeded.
2380 */
2381 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2382 tcp_undo_cwnd_reduction(sk, false);
2383 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2384 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2385 else
2386 mib_idx = LINUX_MIB_TCPFULLUNDO;
2387
2388 NET_INC_STATS(sock_net(sk), mib_idx);
2389 } else if (tp->rack.reo_wnd_persist) {
2390 tp->rack.reo_wnd_persist--;
2391 }
2392 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2393 /* Hold old state until something *above* high_seq
2394 * is ACKed. For Reno it is MUST to prevent false
2395 * fast retransmits (RFC2582). SACK TCP is safe. */
2396 if (!tcp_any_retrans_done(sk))
2397 tp->retrans_stamp = 0;
2398 return true;
2399 }
2400 tcp_set_ca_state(sk, TCP_CA_Open);
2401 tp->is_sack_reneg = 0;
2402 return false;
2403}
2404
2405/* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2406static bool tcp_try_undo_dsack(struct sock *sk)
2407{
2408 struct tcp_sock *tp = tcp_sk(sk);
2409
2410 if (tp->undo_marker && !tp->undo_retrans) {
2411 tp->rack.reo_wnd_persist = min(TCP_RACK_RECOVERY_THRESH,
2412 tp->rack.reo_wnd_persist + 1);
2413 DBGUNDO(sk, "D-SACK");
2414 tcp_undo_cwnd_reduction(sk, false);
2415 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2416 return true;
2417 }
2418 return false;
2419}
2420
2421/* Undo during loss recovery after partial ACK or using F-RTO. */
2422static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2423{
2424 struct tcp_sock *tp = tcp_sk(sk);
2425
2426 if (frto_undo || tcp_may_undo(tp)) {
2427 tcp_undo_cwnd_reduction(sk, true);
2428
2429 DBGUNDO(sk, "partial loss");
2430 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2431 if (frto_undo)
2432 NET_INC_STATS(sock_net(sk),
2433 LINUX_MIB_TCPSPURIOUSRTOS);
2434 inet_csk(sk)->icsk_retransmits = 0;
2435 if (frto_undo || tcp_is_sack(tp)) {
2436 tcp_set_ca_state(sk, TCP_CA_Open);
2437 tp->is_sack_reneg = 0;
2438 }
2439 return true;
2440 }
2441 return false;
2442}
2443
2444/* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2445 * It computes the number of packets to send (sndcnt) based on packets newly
2446 * delivered:
2447 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2448 * cwnd reductions across a full RTT.
2449 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2450 * But when the retransmits are acked without further losses, PRR
2451 * slow starts cwnd up to ssthresh to speed up the recovery.
2452 */
2453static void tcp_init_cwnd_reduction(struct sock *sk)
2454{
2455 struct tcp_sock *tp = tcp_sk(sk);
2456
2457 tp->high_seq = tp->snd_nxt;
2458 tp->tlp_high_seq = 0;
2459 tp->snd_cwnd_cnt = 0;
2460 tp->prior_cwnd = tp->snd_cwnd;
2461 tp->prr_delivered = 0;
2462 tp->prr_out = 0;
2463 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2464 tcp_ecn_queue_cwr(tp);
2465}
2466
2467void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int flag)
2468{
2469 struct tcp_sock *tp = tcp_sk(sk);
2470 int sndcnt = 0;
2471 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2472
2473 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2474 return;
2475
2476 tp->prr_delivered += newly_acked_sacked;
2477 if (delta < 0) {
2478 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2479 tp->prior_cwnd - 1;
2480 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2481 } else if ((flag & FLAG_RETRANS_DATA_ACKED) &&
2482 !(flag & FLAG_LOST_RETRANS)) {
2483 sndcnt = min_t(int, delta,
2484 max_t(int, tp->prr_delivered - tp->prr_out,
2485 newly_acked_sacked) + 1);
2486 } else {
2487 sndcnt = min(delta, newly_acked_sacked);
2488 }
2489 /* Force a fast retransmit upon entering fast recovery */
2490 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
2491 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2492}
2493
2494static inline void tcp_end_cwnd_reduction(struct sock *sk)
2495{
2496 struct tcp_sock *tp = tcp_sk(sk);
2497
2498 if (inet_csk(sk)->icsk_ca_ops->cong_control)
2499 return;
2500
2501 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2502 if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH &&
2503 (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) {
2504 tp->snd_cwnd = tp->snd_ssthresh;
2505 tp->snd_cwnd_stamp = tcp_jiffies32;
2506 }
2507 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2508}
2509
2510/* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2511void tcp_enter_cwr(struct sock *sk)
2512{
2513 struct tcp_sock *tp = tcp_sk(sk);
2514
2515 tp->prior_ssthresh = 0;
2516 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2517 tp->undo_marker = 0;
2518 tcp_init_cwnd_reduction(sk);
2519 tcp_set_ca_state(sk, TCP_CA_CWR);
2520 }
2521}
2522EXPORT_SYMBOL(tcp_enter_cwr);
2523
2524static void tcp_try_keep_open(struct sock *sk)
2525{
2526 struct tcp_sock *tp = tcp_sk(sk);
2527 int state = TCP_CA_Open;
2528
2529 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2530 state = TCP_CA_Disorder;
2531
2532 if (inet_csk(sk)->icsk_ca_state != state) {
2533 tcp_set_ca_state(sk, state);
2534 tp->high_seq = tp->snd_nxt;
2535 }
2536}
2537
2538static void tcp_try_to_open(struct sock *sk, int flag)
2539{
2540 struct tcp_sock *tp = tcp_sk(sk);
2541
2542 tcp_verify_left_out(tp);
2543
2544 if (!tcp_any_retrans_done(sk))
2545 tp->retrans_stamp = 0;
2546
2547 if (flag & FLAG_ECE)
2548 tcp_enter_cwr(sk);
2549
2550 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2551 tcp_try_keep_open(sk);
2552 }
2553}
2554
2555static void tcp_mtup_probe_failed(struct sock *sk)
2556{
2557 struct inet_connection_sock *icsk = inet_csk(sk);
2558
2559 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2560 icsk->icsk_mtup.probe_size = 0;
2561 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
2562}
2563
2564static void tcp_mtup_probe_success(struct sock *sk)
2565{
2566 struct tcp_sock *tp = tcp_sk(sk);
2567 struct inet_connection_sock *icsk = inet_csk(sk);
2568
2569 /* FIXME: breaks with very large cwnd */
2570 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2571 tp->snd_cwnd = tp->snd_cwnd *
2572 tcp_mss_to_mtu(sk, tp->mss_cache) /
2573 icsk->icsk_mtup.probe_size;
2574 tp->snd_cwnd_cnt = 0;
2575 tp->snd_cwnd_stamp = tcp_jiffies32;
2576 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2577
2578 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2579 icsk->icsk_mtup.probe_size = 0;
2580 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2581 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
2582}
2583
2584/* Do a simple retransmit without using the backoff mechanisms in
2585 * tcp_timer. This is used for path mtu discovery.
2586 * The socket is already locked here.
2587 */
2588void tcp_simple_retransmit(struct sock *sk)
2589{
2590 const struct inet_connection_sock *icsk = inet_csk(sk);
2591 struct tcp_sock *tp = tcp_sk(sk);
2592 struct sk_buff *skb;
2593 unsigned int mss = tcp_current_mss(sk);
2594
2595 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2596 if (tcp_skb_seglen(skb) > mss &&
2597 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2598 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2599 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2600 tp->retrans_out -= tcp_skb_pcount(skb);
2601 }
2602 tcp_skb_mark_lost_uncond_verify(tp, skb);
2603 }
2604 }
2605
2606 tcp_clear_retrans_hints_partial(tp);
2607
2608 if (!tp->lost_out)
2609 return;
2610
2611 if (tcp_is_reno(tp))
2612 tcp_limit_reno_sacked(tp);
2613
2614 tcp_verify_left_out(tp);
2615
2616 /* Don't muck with the congestion window here.
2617 * Reason is that we do not increase amount of _data_
2618 * in network, but units changed and effective
2619 * cwnd/ssthresh really reduced now.
2620 */
2621 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2622 tp->high_seq = tp->snd_nxt;
2623 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2624 tp->prior_ssthresh = 0;
2625 tp->undo_marker = 0;
2626 tcp_set_ca_state(sk, TCP_CA_Loss);
2627 }
2628 tcp_xmit_retransmit_queue(sk);
2629}
2630EXPORT_SYMBOL(tcp_simple_retransmit);
2631
2632void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2633{
2634 struct tcp_sock *tp = tcp_sk(sk);
2635 int mib_idx;
2636
2637 if (tcp_is_reno(tp))
2638 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2639 else
2640 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2641
2642 NET_INC_STATS(sock_net(sk), mib_idx);
2643
2644 tp->prior_ssthresh = 0;
2645 tcp_init_undo(tp);
2646
2647 if (!tcp_in_cwnd_reduction(sk)) {
2648 if (!ece_ack)
2649 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2650 tcp_init_cwnd_reduction(sk);
2651 }
2652 tcp_set_ca_state(sk, TCP_CA_Recovery);
2653}
2654
2655/* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2656 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2657 */
2658static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack,
2659 int *rexmit)
2660{
2661 struct tcp_sock *tp = tcp_sk(sk);
2662 bool recovered = !before(tp->snd_una, tp->high_seq);
2663
2664 if ((flag & FLAG_SND_UNA_ADVANCED) &&
2665 tcp_try_undo_loss(sk, false))
2666 return;
2667
2668 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2669 /* Step 3.b. A timeout is spurious if not all data are
2670 * lost, i.e., never-retransmitted data are (s)acked.
2671 */
2672 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2673 tcp_try_undo_loss(sk, true))
2674 return;
2675
2676 if (after(tp->snd_nxt, tp->high_seq)) {
2677 if (flag & FLAG_DATA_SACKED || is_dupack)
2678 tp->frto = 0; /* Step 3.a. loss was real */
2679 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2680 tp->high_seq = tp->snd_nxt;
2681 /* Step 2.b. Try send new data (but deferred until cwnd
2682 * is updated in tcp_ack()). Otherwise fall back to
2683 * the conventional recovery.
2684 */
2685 if (!tcp_write_queue_empty(sk) &&
2686 after(tcp_wnd_end(tp), tp->snd_nxt)) {
2687 *rexmit = REXMIT_NEW;
2688 return;
2689 }
2690 tp->frto = 0;
2691 }
2692 }
2693
2694 if (recovered) {
2695 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2696 tcp_try_undo_recovery(sk);
2697 return;
2698 }
2699 if (tcp_is_reno(tp)) {
2700 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2701 * delivered. Lower inflight to clock out (re)tranmissions.
2702 */
2703 if (after(tp->snd_nxt, tp->high_seq) && is_dupack)
2704 tcp_add_reno_sack(sk);
2705 else if (flag & FLAG_SND_UNA_ADVANCED)
2706 tcp_reset_reno_sack(tp);
2707 }
2708 *rexmit = REXMIT_LOST;
2709}
2710
2711/* Undo during fast recovery after partial ACK. */
2712static bool tcp_try_undo_partial(struct sock *sk, u32 prior_snd_una)
2713{
2714 struct tcp_sock *tp = tcp_sk(sk);
2715
2716 if (tp->undo_marker && tcp_packet_delayed(tp)) {
2717 /* Plain luck! Hole if filled with delayed
2718 * packet, rather than with a retransmit. Check reordering.
2719 */
2720 tcp_check_sack_reordering(sk, prior_snd_una, 1);
2721
2722 /* We are getting evidence that the reordering degree is higher
2723 * than we realized. If there are no retransmits out then we
2724 * can undo. Otherwise we clock out new packets but do not
2725 * mark more packets lost or retransmit more.
2726 */
2727 if (tp->retrans_out)
2728 return true;
2729
2730 if (!tcp_any_retrans_done(sk))
2731 tp->retrans_stamp = 0;
2732
2733 DBGUNDO(sk, "partial recovery");
2734 tcp_undo_cwnd_reduction(sk, true);
2735 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2736 tcp_try_keep_open(sk);
2737 return true;
2738 }
2739 return false;
2740}
2741
2742static void tcp_identify_packet_loss(struct sock *sk, int *ack_flag)
2743{
2744 struct tcp_sock *tp = tcp_sk(sk);
2745
2746 if (tcp_rtx_queue_empty(sk))
2747 return;
2748
2749 if (unlikely(tcp_is_reno(tp))) {
2750 tcp_newreno_mark_lost(sk, *ack_flag & FLAG_SND_UNA_ADVANCED);
2751 } else if (tcp_is_rack(sk)) {
2752 u32 prior_retrans = tp->retrans_out;
2753
2754 tcp_rack_mark_lost(sk);
2755 if (prior_retrans > tp->retrans_out)
2756 *ack_flag |= FLAG_LOST_RETRANS;
2757 }
2758}
2759
2760static bool tcp_force_fast_retransmit(struct sock *sk)
2761{
2762 struct tcp_sock *tp = tcp_sk(sk);
2763
2764 return after(tcp_highest_sack_seq(tp),
2765 tp->snd_una + tp->reordering * tp->mss_cache);
2766}
2767
2768/* Process an event, which can update packets-in-flight not trivially.
2769 * Main goal of this function is to calculate new estimate for left_out,
2770 * taking into account both packets sitting in receiver's buffer and
2771 * packets lost by network.
2772 *
2773 * Besides that it updates the congestion state when packet loss or ECN
2774 * is detected. But it does not reduce the cwnd, it is done by the
2775 * congestion control later.
2776 *
2777 * It does _not_ decide what to send, it is made in function
2778 * tcp_xmit_retransmit_queue().
2779 */
2780static void tcp_fastretrans_alert(struct sock *sk, const u32 prior_snd_una,
2781 bool is_dupack, int *ack_flag, int *rexmit)
2782{
2783 struct inet_connection_sock *icsk = inet_csk(sk);
2784 struct tcp_sock *tp = tcp_sk(sk);
2785 int fast_rexmit = 0, flag = *ack_flag;
2786 bool do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2787 tcp_force_fast_retransmit(sk));
2788
2789 if (!tp->packets_out && tp->sacked_out)
2790 tp->sacked_out = 0;
2791
2792 /* Now state machine starts.
2793 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2794 if (flag & FLAG_ECE)
2795 tp->prior_ssthresh = 0;
2796
2797 /* B. In all the states check for reneging SACKs. */
2798 if (tcp_check_sack_reneging(sk, flag))
2799 return;
2800
2801 /* C. Check consistency of the current state. */
2802 tcp_verify_left_out(tp);
2803
2804 /* D. Check state exit conditions. State can be terminated
2805 * when high_seq is ACKed. */
2806 if (icsk->icsk_ca_state == TCP_CA_Open) {
2807 WARN_ON(tp->retrans_out != 0);
2808 tp->retrans_stamp = 0;
2809 } else if (!before(tp->snd_una, tp->high_seq)) {
2810 switch (icsk->icsk_ca_state) {
2811 case TCP_CA_CWR:
2812 /* CWR is to be held something *above* high_seq
2813 * is ACKed for CWR bit to reach receiver. */
2814 if (tp->snd_una != tp->high_seq) {
2815 tcp_end_cwnd_reduction(sk);
2816 tcp_set_ca_state(sk, TCP_CA_Open);
2817 }
2818 break;
2819
2820 case TCP_CA_Recovery:
2821 if (tcp_is_reno(tp))
2822 tcp_reset_reno_sack(tp);
2823 if (tcp_try_undo_recovery(sk))
2824 return;
2825 tcp_end_cwnd_reduction(sk);
2826 break;
2827 }
2828 }
2829
2830 /* E. Process state. */
2831 switch (icsk->icsk_ca_state) {
2832 case TCP_CA_Recovery:
2833 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2834 if (tcp_is_reno(tp) && is_dupack)
2835 tcp_add_reno_sack(sk);
2836 } else {
2837 if (tcp_try_undo_partial(sk, prior_snd_una))
2838 return;
2839 /* Partial ACK arrived. Force fast retransmit. */
2840 do_lost = tcp_is_reno(tp) ||
2841 tcp_force_fast_retransmit(sk);
2842 }
2843 if (tcp_try_undo_dsack(sk)) {
2844 tcp_try_keep_open(sk);
2845 return;
2846 }
2847 tcp_identify_packet_loss(sk, ack_flag);
2848 break;
2849 case TCP_CA_Loss:
2850 tcp_process_loss(sk, flag, is_dupack, rexmit);
2851 tcp_identify_packet_loss(sk, ack_flag);
2852 if (!(icsk->icsk_ca_state == TCP_CA_Open ||
2853 (*ack_flag & FLAG_LOST_RETRANS)))
2854 return;
2855 /* Change state if cwnd is undone or retransmits are lost */
2856 /* fall through */
2857 default:
2858 if (tcp_is_reno(tp)) {
2859 if (flag & FLAG_SND_UNA_ADVANCED)
2860 tcp_reset_reno_sack(tp);
2861 if (is_dupack)
2862 tcp_add_reno_sack(sk);
2863 }
2864
2865 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2866 tcp_try_undo_dsack(sk);
2867
2868 tcp_identify_packet_loss(sk, ack_flag);
2869 if (!tcp_time_to_recover(sk, flag)) {
2870 tcp_try_to_open(sk, flag);
2871 return;
2872 }
2873
2874 /* MTU probe failure: don't reduce cwnd */
2875 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2876 icsk->icsk_mtup.probe_size &&
2877 tp->snd_una == tp->mtu_probe.probe_seq_start) {
2878 tcp_mtup_probe_failed(sk);
2879 /* Restores the reduction we did in tcp_mtup_probe() */
2880 tp->snd_cwnd++;
2881 tcp_simple_retransmit(sk);
2882 return;
2883 }
2884
2885 /* Otherwise enter Recovery state */
2886 tcp_enter_recovery(sk, (flag & FLAG_ECE));
2887 fast_rexmit = 1;
2888 }
2889
2890 if (!tcp_is_rack(sk) && do_lost)
2891 tcp_update_scoreboard(sk, fast_rexmit);
2892 *rexmit = REXMIT_LOST;
2893}
2894
2895static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us, const int flag)
2896{
2897 u32 wlen = sock_net(sk)->ipv4.sysctl_tcp_min_rtt_wlen * HZ;
2898 struct tcp_sock *tp = tcp_sk(sk);
2899
2900 if ((flag & FLAG_ACK_MAYBE_DELAYED) && rtt_us > tcp_min_rtt(tp)) {
2901 /* If the remote keeps returning delayed ACKs, eventually
2902 * the min filter would pick it up and overestimate the
2903 * prop. delay when it expires. Skip suspected delayed ACKs.
2904 */
2905 return;
2906 }
2907 minmax_running_min(&tp->rtt_min, wlen, tcp_jiffies32,
2908 rtt_us ? : jiffies_to_usecs(1));
2909}
2910
2911static bool tcp_ack_update_rtt(struct sock *sk, const int flag,
2912 long seq_rtt_us, long sack_rtt_us,
2913 long ca_rtt_us, struct rate_sample *rs)
2914{
2915 const struct tcp_sock *tp = tcp_sk(sk);
2916
2917 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2918 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2919 * Karn's algorithm forbids taking RTT if some retransmitted data
2920 * is acked (RFC6298).
2921 */
2922 if (seq_rtt_us < 0)
2923 seq_rtt_us = sack_rtt_us;
2924
2925 /* RTTM Rule: A TSecr value received in a segment is used to
2926 * update the averaged RTT measurement only if the segment
2927 * acknowledges some new data, i.e., only if it advances the
2928 * left edge of the send window.
2929 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2930 */
2931 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2932 flag & FLAG_ACKED) {
2933 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
2934 u32 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
2935
2936 seq_rtt_us = ca_rtt_us = delta_us;
2937 }
2938 rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */
2939 if (seq_rtt_us < 0)
2940 return false;
2941
2942 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
2943 * always taken together with ACK, SACK, or TS-opts. Any negative
2944 * values will be skipped with the seq_rtt_us < 0 check above.
2945 */
2946 tcp_update_rtt_min(sk, ca_rtt_us, flag);
2947 tcp_rtt_estimator(sk, seq_rtt_us);
2948 tcp_set_rto(sk);
2949
2950 /* RFC6298: only reset backoff on valid RTT measurement. */
2951 inet_csk(sk)->icsk_backoff = 0;
2952 return true;
2953}
2954
2955/* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2956void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
2957{
2958 struct rate_sample rs;
2959 long rtt_us = -1L;
2960
2961 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack)
2962 rtt_us = tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req)->snt_synack);
2963
2964 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us, &rs);
2965}
2966
2967
2968static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
2969{
2970 const struct inet_connection_sock *icsk = inet_csk(sk);
2971
2972 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
2973 tcp_sk(sk)->snd_cwnd_stamp = tcp_jiffies32;
2974}
2975
2976/* Restart timer after forward progress on connection.
2977 * RFC2988 recommends to restart timer to now+rto.
2978 */
2979void tcp_rearm_rto(struct sock *sk)
2980{
2981 const struct inet_connection_sock *icsk = inet_csk(sk);
2982 struct tcp_sock *tp = tcp_sk(sk);
2983
2984 /* If the retrans timer is currently being used by Fast Open
2985 * for SYN-ACK retrans purpose, stay put.
2986 */
2987 if (tp->fastopen_rsk)
2988 return;
2989
2990 if (!tp->packets_out) {
2991 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2992 } else {
2993 u32 rto = inet_csk(sk)->icsk_rto;
2994 /* Offset the time elapsed after installing regular RTO */
2995 if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT ||
2996 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
2997 s64 delta_us = tcp_rto_delta_us(sk);
2998 /* delta_us may not be positive if the socket is locked
2999 * when the retrans timer fires and is rescheduled.
3000 */
3001 rto = usecs_to_jiffies(max_t(int, delta_us, 1));
3002 }
3003 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
3004 TCP_RTO_MAX);
3005 }
3006}
3007
3008/* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */
3009static void tcp_set_xmit_timer(struct sock *sk)
3010{
3011 if (!tcp_schedule_loss_probe(sk, true))
3012 tcp_rearm_rto(sk);
3013}
3014
3015/* If we get here, the whole TSO packet has not been acked. */
3016static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3017{
3018 struct tcp_sock *tp = tcp_sk(sk);
3019 u32 packets_acked;
3020
3021 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3022
3023 packets_acked = tcp_skb_pcount(skb);
3024 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3025 return 0;
3026 packets_acked -= tcp_skb_pcount(skb);
3027
3028 if (packets_acked) {
3029 BUG_ON(tcp_skb_pcount(skb) == 0);
3030 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3031 }
3032
3033 return packets_acked;
3034}
3035
3036static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3037 u32 prior_snd_una)
3038{
3039 const struct skb_shared_info *shinfo;
3040
3041 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3042 if (likely(!TCP_SKB_CB(skb)->txstamp_ack))
3043 return;
3044
3045 shinfo = skb_shinfo(skb);
3046 if (!before(shinfo->tskey, prior_snd_una) &&
3047 before(shinfo->tskey, tcp_sk(sk)->snd_una)) {
3048 tcp_skb_tsorted_save(skb) {
3049 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK);
3050 } tcp_skb_tsorted_restore(skb);
3051 }
3052}
3053
3054/* Remove acknowledged frames from the retransmission queue. If our packet
3055 * is before the ack sequence we can discard it as it's confirmed to have
3056 * arrived at the other end.
3057 */
3058static int tcp_clean_rtx_queue(struct sock *sk, u32 prior_fack,
3059 u32 prior_snd_una,
3060 struct tcp_sacktag_state *sack)
3061{
3062 const struct inet_connection_sock *icsk = inet_csk(sk);
3063 u64 first_ackt, last_ackt;
3064 struct tcp_sock *tp = tcp_sk(sk);
3065 u32 prior_sacked = tp->sacked_out;
3066 u32 reord = tp->snd_nxt; /* lowest acked un-retx un-sacked seq */
3067 struct sk_buff *skb, *next;
3068 bool fully_acked = true;
3069 long sack_rtt_us = -1L;
3070 long seq_rtt_us = -1L;
3071 long ca_rtt_us = -1L;
3072 u32 pkts_acked = 0;
3073 u32 last_in_flight = 0;
3074 bool rtt_update;
3075 int flag = 0;
3076
3077 first_ackt = 0;
3078
3079 for (skb = skb_rb_first(&sk->tcp_rtx_queue); skb; skb = next) {
3080 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3081 const u32 start_seq = scb->seq;
3082 u8 sacked = scb->sacked;
3083 u32 acked_pcount;
3084
3085 tcp_ack_tstamp(sk, skb, prior_snd_una);
3086
3087 /* Determine how many packets and what bytes were acked, tso and else */
3088 if (after(scb->end_seq, tp->snd_una)) {
3089 if (tcp_skb_pcount(skb) == 1 ||
3090 !after(tp->snd_una, scb->seq))
3091 break;
3092
3093 acked_pcount = tcp_tso_acked(sk, skb);
3094 if (!acked_pcount)
3095 break;
3096 fully_acked = false;
3097 } else {
3098 acked_pcount = tcp_skb_pcount(skb);
3099 }
3100
3101 if (unlikely(sacked & TCPCB_RETRANS)) {
3102 if (sacked & TCPCB_SACKED_RETRANS)
3103 tp->retrans_out -= acked_pcount;
3104 flag |= FLAG_RETRANS_DATA_ACKED;
3105 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3106 last_ackt = skb->skb_mstamp;
3107 WARN_ON_ONCE(last_ackt == 0);
3108 if (!first_ackt)
3109 first_ackt = last_ackt;
3110
3111 last_in_flight = TCP_SKB_CB(skb)->tx.in_flight;
3112 if (before(start_seq, reord))
3113 reord = start_seq;
3114 if (!after(scb->end_seq, tp->high_seq))
3115 flag |= FLAG_ORIG_SACK_ACKED;
3116 }
3117
3118 if (sacked & TCPCB_SACKED_ACKED) {
3119 tp->sacked_out -= acked_pcount;
3120 } else if (tcp_is_sack(tp)) {
3121 tp->delivered += acked_pcount;
3122 if (!tcp_skb_spurious_retrans(tp, skb))
3123 tcp_rack_advance(tp, sacked, scb->end_seq,
3124 skb->skb_mstamp);
3125 }
3126 if (sacked & TCPCB_LOST)
3127 tp->lost_out -= acked_pcount;
3128
3129 tp->packets_out -= acked_pcount;
3130 pkts_acked += acked_pcount;
3131 tcp_rate_skb_delivered(sk, skb, sack->rate);
3132
3133 /* Initial outgoing SYN's get put onto the write_queue
3134 * just like anything else we transmit. It is not
3135 * true data, and if we misinform our callers that
3136 * this ACK acks real data, we will erroneously exit
3137 * connection startup slow start one packet too
3138 * quickly. This is severely frowned upon behavior.
3139 */
3140 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3141 flag |= FLAG_DATA_ACKED;
3142 } else {
3143 flag |= FLAG_SYN_ACKED;
3144 tp->retrans_stamp = 0;
3145 }
3146
3147 if (!fully_acked)
3148 break;
3149
3150 next = skb_rb_next(skb);
3151 if (unlikely(skb == tp->retransmit_skb_hint))
3152 tp->retransmit_skb_hint = NULL;
3153 if (unlikely(skb == tp->lost_skb_hint))
3154 tp->lost_skb_hint = NULL;
3155 tcp_rtx_queue_unlink_and_free(skb, sk);
3156 }
3157
3158 if (!skb)
3159 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3160
3161 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3162 tp->snd_up = tp->snd_una;
3163
3164 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3165 flag |= FLAG_SACK_RENEGING;
3166
3167 if (likely(first_ackt) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
3168 seq_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, first_ackt);
3169 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, last_ackt);
3170
3171 if (pkts_acked == 1 && last_in_flight < tp->mss_cache &&
3172 last_in_flight && !