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 inet_csk_exit_pingpong_mode(sk);
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 && !inet_csk_in_pingpong_mode(sk));
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. Try to fixup all. It is made immediately after connection enters
430 * established state.
431 */
432void tcp_init_buffer_space(struct sock *sk)
433{
434 int tcp_app_win = sock_net(sk)->ipv4.sysctl_tcp_app_win;
435 struct tcp_sock *tp = tcp_sk(sk);
436 int maxwin;
437
438 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
439 tcp_sndbuf_expand(sk);
440
441 tp->rcvq_space.space = min_t(u32, tp->rcv_wnd, TCP_INIT_CWND * tp->advmss);
442 tcp_mstamp_refresh(tp);
443 tp->rcvq_space.time = tp->tcp_mstamp;
444 tp->rcvq_space.seq = tp->copied_seq;
445
446 maxwin = tcp_full_space(sk);
447
448 if (tp->window_clamp >= maxwin) {
449 tp->window_clamp = maxwin;
450
451 if (tcp_app_win && maxwin > 4 * tp->advmss)
452 tp->window_clamp = max(maxwin -
453 (maxwin >> tcp_app_win),
454 4 * tp->advmss);
455 }
456
457 /* Force reservation of one segment. */
458 if (tcp_app_win &&
459 tp->window_clamp > 2 * tp->advmss &&
460 tp->window_clamp + tp->advmss > maxwin)
461 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
462
463 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
464 tp->snd_cwnd_stamp = tcp_jiffies32;
465}
466
467/* 4. Recalculate window clamp after socket hit its memory bounds. */
468static void tcp_clamp_window(struct sock *sk)
469{
470 struct tcp_sock *tp = tcp_sk(sk);
471 struct inet_connection_sock *icsk = inet_csk(sk);
472 struct net *net = sock_net(sk);
473
474 icsk->icsk_ack.quick = 0;
475
476 if (sk->sk_rcvbuf < net->ipv4.sysctl_tcp_rmem[2] &&
477 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
478 !tcp_under_memory_pressure(sk) &&
479 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
480 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
481 net->ipv4.sysctl_tcp_rmem[2]);
482 }
483 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
484 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
485}
486
487/* Initialize RCV_MSS value.
488 * RCV_MSS is an our guess about MSS used by the peer.
489 * We haven't any direct information about the MSS.
490 * It's better to underestimate the RCV_MSS rather than overestimate.
491 * Overestimations make us ACKing less frequently than needed.
492 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
493 */
494void tcp_initialize_rcv_mss(struct sock *sk)
495{
496 const struct tcp_sock *tp = tcp_sk(sk);
497 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
498
499 hint = min(hint, tp->rcv_wnd / 2);
500 hint = min(hint, TCP_MSS_DEFAULT);
501 hint = max(hint, TCP_MIN_MSS);
502
503 inet_csk(sk)->icsk_ack.rcv_mss = hint;
504}
505EXPORT_SYMBOL(tcp_initialize_rcv_mss);
506
507/* Receiver "autotuning" code.
508 *
509 * The algorithm for RTT estimation w/o timestamps is based on
510 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
511 * <http://public.lanl.gov/radiant/pubs.html#DRS>
512 *
513 * More detail on this code can be found at
514 * <http://staff.psc.edu/jheffner/>,
515 * though this reference is out of date. A new paper
516 * is pending.
517 */
518static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
519{
520 u32 new_sample = tp->rcv_rtt_est.rtt_us;
521 long m = sample;
522
523 if (new_sample != 0) {
524 /* If we sample in larger samples in the non-timestamp
525 * case, we could grossly overestimate the RTT especially
526 * with chatty applications or bulk transfer apps which
527 * are stalled on filesystem I/O.
528 *
529 * Also, since we are only going for a minimum in the
530 * non-timestamp case, we do not smooth things out
531 * else with timestamps disabled convergence takes too
532 * long.
533 */
534 if (!win_dep) {
535 m -= (new_sample >> 3);
536 new_sample += m;
537 } else {
538 m <<= 3;
539 if (m < new_sample)
540 new_sample = m;
541 }
542 } else {
543 /* No previous measure. */
544 new_sample = m << 3;
545 }
546
547 tp->rcv_rtt_est.rtt_us = new_sample;
548}
549
550static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
551{
552 u32 delta_us;
553
554 if (tp->rcv_rtt_est.time == 0)
555 goto new_measure;
556 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
557 return;
558 delta_us = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcv_rtt_est.time);
559 if (!delta_us)
560 delta_us = 1;
561 tcp_rcv_rtt_update(tp, delta_us, 1);
562
563new_measure:
564 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
565 tp->rcv_rtt_est.time = tp->tcp_mstamp;
566}
567
568static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
569 const struct sk_buff *skb)
570{
571 struct tcp_sock *tp = tcp_sk(sk);
572
573 if (tp->rx_opt.rcv_tsecr == tp->rcv_rtt_last_tsecr)
574 return;
575 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
576
577 if (TCP_SKB_CB(skb)->end_seq -
578 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss) {
579 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
580 u32 delta_us;
581
582 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) {
583 if (!delta)
584 delta = 1;
585 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
586 tcp_rcv_rtt_update(tp, delta_us, 0);
587 }
588 }
589}
590
591/*
592 * This function should be called every time data is copied to user space.
593 * It calculates the appropriate TCP receive buffer space.
594 */
595void tcp_rcv_space_adjust(struct sock *sk)
596{
597 struct tcp_sock *tp = tcp_sk(sk);
598 u32 copied;
599 int time;
600
601 trace_tcp_rcv_space_adjust(sk);
602
603 tcp_mstamp_refresh(tp);
604 time = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcvq_space.time);
605 if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0)
606 return;
607
608 /* Number of bytes copied to user in last RTT */
609 copied = tp->copied_seq - tp->rcvq_space.seq;
610 if (copied <= tp->rcvq_space.space)
611 goto new_measure;
612
613 /* A bit of theory :
614 * copied = bytes received in previous RTT, our base window
615 * To cope with packet losses, we need a 2x factor
616 * To cope with slow start, and sender growing its cwin by 100 %
617 * every RTT, we need a 4x factor, because the ACK we are sending
618 * now is for the next RTT, not the current one :
619 * <prev RTT . ><current RTT .. ><next RTT .... >
620 */
621
622 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf &&
623 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
624 int rcvmem, rcvbuf;
625 u64 rcvwin, grow;
626
627 /* minimal window to cope with packet losses, assuming
628 * steady state. Add some cushion because of small variations.
629 */
630 rcvwin = ((u64)copied << 1) + 16 * tp->advmss;
631
632 /* Accommodate for sender rate increase (eg. slow start) */
633 grow = rcvwin * (copied - tp->rcvq_space.space);
634 do_div(grow, tp->rcvq_space.space);
635 rcvwin += (grow << 1);
636
637 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
638 while (tcp_win_from_space(sk, rcvmem) < tp->advmss)
639 rcvmem += 128;
640
641 do_div(rcvwin, tp->advmss);
642 rcvbuf = min_t(u64, rcvwin * rcvmem,
643 sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
644 if (rcvbuf > sk->sk_rcvbuf) {
645 sk->sk_rcvbuf = rcvbuf;
646
647 /* Make the window clamp follow along. */
648 tp->window_clamp = tcp_win_from_space(sk, rcvbuf);
649 }
650 }
651 tp->rcvq_space.space = copied;
652
653new_measure:
654 tp->rcvq_space.seq = tp->copied_seq;
655 tp->rcvq_space.time = tp->tcp_mstamp;
656}
657
658/* There is something which you must keep in mind when you analyze the
659 * behavior of the tp->ato delayed ack timeout interval. When a
660 * connection starts up, we want to ack as quickly as possible. The
661 * problem is that "good" TCP's do slow start at the beginning of data
662 * transmission. The means that until we send the first few ACK's the
663 * sender will sit on his end and only queue most of his data, because
664 * he can only send snd_cwnd unacked packets at any given time. For
665 * each ACK we send, he increments snd_cwnd and transmits more of his
666 * queue. -DaveM
667 */
668static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
669{
670 struct tcp_sock *tp = tcp_sk(sk);
671 struct inet_connection_sock *icsk = inet_csk(sk);
672 u32 now;
673
674 inet_csk_schedule_ack(sk);
675
676 tcp_measure_rcv_mss(sk, skb);
677
678 tcp_rcv_rtt_measure(tp);
679
680 now = tcp_jiffies32;
681
682 if (!icsk->icsk_ack.ato) {
683 /* The _first_ data packet received, initialize
684 * delayed ACK engine.
685 */
686 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
687 icsk->icsk_ack.ato = TCP_ATO_MIN;
688 } else {
689 int m = now - icsk->icsk_ack.lrcvtime;
690
691 if (m <= TCP_ATO_MIN / 2) {
692 /* The fastest case is the first. */
693 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
694 } else if (m < icsk->icsk_ack.ato) {
695 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
696 if (icsk->icsk_ack.ato > icsk->icsk_rto)
697 icsk->icsk_ack.ato = icsk->icsk_rto;
698 } else if (m > icsk->icsk_rto) {
699 /* Too long gap. Apparently sender failed to
700 * restart window, so that we send ACKs quickly.
701 */
702 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
703 sk_mem_reclaim(sk);
704 }
705 }
706 icsk->icsk_ack.lrcvtime = now;
707
708 tcp_ecn_check_ce(sk, skb);
709
710 if (skb->len >= 128)
711 tcp_grow_window(sk, skb);
712}
713
714/* Called to compute a smoothed rtt estimate. The data fed to this
715 * routine either comes from timestamps, or from segments that were
716 * known _not_ to have been retransmitted [see Karn/Partridge
717 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
718 * piece by Van Jacobson.
719 * NOTE: the next three routines used to be one big routine.
720 * To save cycles in the RFC 1323 implementation it was better to break
721 * it up into three procedures. -- erics
722 */
723static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
724{
725 struct tcp_sock *tp = tcp_sk(sk);
726 long m = mrtt_us; /* RTT */
727 u32 srtt = tp->srtt_us;
728
729 /* The following amusing code comes from Jacobson's
730 * article in SIGCOMM '88. Note that rtt and mdev
731 * are scaled versions of rtt and mean deviation.
732 * This is designed to be as fast as possible
733 * m stands for "measurement".
734 *
735 * On a 1990 paper the rto value is changed to:
736 * RTO = rtt + 4 * mdev
737 *
738 * Funny. This algorithm seems to be very broken.
739 * These formulae increase RTO, when it should be decreased, increase
740 * too slowly, when it should be increased quickly, decrease too quickly
741 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
742 * does not matter how to _calculate_ it. Seems, it was trap
743 * that VJ failed to avoid. 8)
744 */
745 if (srtt != 0) {
746 m -= (srtt >> 3); /* m is now error in rtt est */
747 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
748 if (m < 0) {
749 m = -m; /* m is now abs(error) */
750 m -= (tp->mdev_us >> 2); /* similar update on mdev */
751 /* This is similar to one of Eifel findings.
752 * Eifel blocks mdev updates when rtt decreases.
753 * This solution is a bit different: we use finer gain
754 * for mdev in this case (alpha*beta).
755 * Like Eifel it also prevents growth of rto,
756 * but also it limits too fast rto decreases,
757 * happening in pure Eifel.
758 */
759 if (m > 0)
760 m >>= 3;
761 } else {
762 m -= (tp->mdev_us >> 2); /* similar update on mdev */
763 }
764 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
765 if (tp->mdev_us > tp->mdev_max_us) {
766 tp->mdev_max_us = tp->mdev_us;
767 if (tp->mdev_max_us > tp->rttvar_us)
768 tp->rttvar_us = tp->mdev_max_us;
769 }
770 if (after(tp->snd_una, tp->rtt_seq)) {
771 if (tp->mdev_max_us < tp->rttvar_us)
772 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
773 tp->rtt_seq = tp->snd_nxt;
774 tp->mdev_max_us = tcp_rto_min_us(sk);
775 }
776 } else {
777 /* no previous measure. */
778 srtt = m << 3; /* take the measured time to be rtt */
779 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
780 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
781 tp->mdev_max_us = tp->rttvar_us;
782 tp->rtt_seq = tp->snd_nxt;
783 }
784 tp->srtt_us = max(1U, srtt);
785}
786
787static void tcp_update_pacing_rate(struct sock *sk)
788{
789 const struct tcp_sock *tp = tcp_sk(sk);
790 u64 rate;
791
792 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
793 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
794
795 /* current rate is (cwnd * mss) / srtt
796 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
797 * In Congestion Avoidance phase, set it to 120 % the current rate.
798 *
799 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
800 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
801 * end of slow start and should slow down.
802 */
803 if (tp->snd_cwnd < tp->snd_ssthresh / 2)
804 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ss_ratio;
805 else
806 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ca_ratio;
807
808 rate *= max(tp->snd_cwnd, tp->packets_out);
809
810 if (likely(tp->srtt_us))
811 do_div(rate, tp->srtt_us);
812
813 /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate
814 * without any lock. We want to make sure compiler wont store
815 * intermediate values in this location.
816 */
817 WRITE_ONCE(sk->sk_pacing_rate, min_t(u64, rate,
818 sk->sk_max_pacing_rate));
819}
820
821/* Calculate rto without backoff. This is the second half of Van Jacobson's
822 * routine referred to above.
823 */
824static void tcp_set_rto(struct sock *sk)
825{
826 const struct tcp_sock *tp = tcp_sk(sk);
827 /* Old crap is replaced with new one. 8)
828 *
829 * More seriously:
830 * 1. If rtt variance happened to be less 50msec, it is hallucination.
831 * It cannot be less due to utterly erratic ACK generation made
832 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
833 * to do with delayed acks, because at cwnd>2 true delack timeout
834 * is invisible. Actually, Linux-2.4 also generates erratic
835 * ACKs in some circumstances.
836 */
837 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
838
839 /* 2. Fixups made earlier cannot be right.
840 * If we do not estimate RTO correctly without them,
841 * all the algo is pure shit and should be replaced
842 * with correct one. It is exactly, which we pretend to do.
843 */
844
845 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
846 * guarantees that rto is higher.
847 */
848 tcp_bound_rto(sk);
849}
850
851__u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
852{
853 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
854
855 if (!cwnd)
856 cwnd = TCP_INIT_CWND;
857 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
858}
859
860/* Take a notice that peer is sending D-SACKs */
861static void tcp_dsack_seen(struct tcp_sock *tp)
862{
863 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
864 tp->rack.dsack_seen = 1;
865 tp->dsack_dups++;
866}
867
868/* It's reordering when higher sequence was delivered (i.e. sacked) before
869 * some lower never-retransmitted sequence ("low_seq"). The maximum reordering
870 * distance is approximated in full-mss packet distance ("reordering").
871 */
872static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq,
873 const int ts)
874{
875 struct tcp_sock *tp = tcp_sk(sk);
876 const u32 mss = tp->mss_cache;
877 u32 fack, metric;
878
879 fack = tcp_highest_sack_seq(tp);
880 if (!before(low_seq, fack))
881 return;
882
883 metric = fack - low_seq;
884 if ((metric > tp->reordering * mss) && mss) {
885#if FASTRETRANS_DEBUG > 1
886 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
887 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
888 tp->reordering,
889 0,
890 tp->sacked_out,
891 tp->undo_marker ? tp->undo_retrans : 0);
892#endif
893 tp->reordering = min_t(u32, (metric + mss - 1) / mss,
894 sock_net(sk)->ipv4.sysctl_tcp_max_reordering);
895 }
896
897 /* This exciting event is worth to be remembered. 8) */
898 tp->reord_seen++;
899 NET_INC_STATS(sock_net(sk),
900 ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER);
901}
902
903/* This must be called before lost_out is incremented */
904static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
905{
906 if (!tp->retransmit_skb_hint ||
907 before(TCP_SKB_CB(skb)->seq,
908 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
909 tp->retransmit_skb_hint = skb;
910}
911
912/* Sum the number of packets on the wire we have marked as lost.
913 * There are two cases we care about here:
914 * a) Packet hasn't been marked lost (nor retransmitted),
915 * and this is the first loss.
916 * b) Packet has been marked both lost and retransmitted,
917 * and this means we think it was lost again.
918 */
919static void tcp_sum_lost(struct tcp_sock *tp, struct sk_buff *skb)
920{
921 __u8 sacked = TCP_SKB_CB(skb)->sacked;
922
923 if (!(sacked & TCPCB_LOST) ||
924 ((sacked & TCPCB_LOST) && (sacked & TCPCB_SACKED_RETRANS)))
925 tp->lost += tcp_skb_pcount(skb);
926}
927
928static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
929{
930 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
931 tcp_verify_retransmit_hint(tp, skb);
932
933 tp->lost_out += tcp_skb_pcount(skb);
934 tcp_sum_lost(tp, skb);
935 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
936 }
937}
938
939void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, struct sk_buff *skb)
940{
941 tcp_verify_retransmit_hint(tp, skb);
942
943 tcp_sum_lost(tp, skb);
944 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
945 tp->lost_out += tcp_skb_pcount(skb);
946 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
947 }
948}
949
950/* This procedure tags the retransmission queue when SACKs arrive.
951 *
952 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
953 * Packets in queue with these bits set are counted in variables
954 * sacked_out, retrans_out and lost_out, correspondingly.
955 *
956 * Valid combinations are:
957 * Tag InFlight Description
958 * 0 1 - orig segment is in flight.
959 * S 0 - nothing flies, orig reached receiver.
960 * L 0 - nothing flies, orig lost by net.
961 * R 2 - both orig and retransmit are in flight.
962 * L|R 1 - orig is lost, retransmit is in flight.
963 * S|R 1 - orig reached receiver, retrans is still in flight.
964 * (L|S|R is logically valid, it could occur when L|R is sacked,
965 * but it is equivalent to plain S and code short-curcuits it to S.
966 * L|S is logically invalid, it would mean -1 packet in flight 8))
967 *
968 * These 6 states form finite state machine, controlled by the following events:
969 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
970 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
971 * 3. Loss detection event of two flavors:
972 * A. Scoreboard estimator decided the packet is lost.
973 * A'. Reno "three dupacks" marks head of queue lost.
974 * B. SACK arrives sacking SND.NXT at the moment, when the
975 * segment was retransmitted.
976 * 4. D-SACK added new rule: D-SACK changes any tag to S.
977 *
978 * It is pleasant to note, that state diagram turns out to be commutative,
979 * so that we are allowed not to be bothered by order of our actions,
980 * when multiple events arrive simultaneously. (see the function below).
981 *
982 * Reordering detection.
983 * --------------------
984 * Reordering metric is maximal distance, which a packet can be displaced
985 * in packet stream. With SACKs we can estimate it:
986 *
987 * 1. SACK fills old hole and the corresponding segment was not
988 * ever retransmitted -> reordering. Alas, we cannot use it
989 * when segment was retransmitted.
990 * 2. The last flaw is solved with D-SACK. D-SACK arrives
991 * for retransmitted and already SACKed segment -> reordering..
992 * Both of these heuristics are not used in Loss state, when we cannot
993 * account for retransmits accurately.
994 *
995 * SACK block validation.
996 * ----------------------
997 *
998 * SACK block range validation checks that the received SACK block fits to
999 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1000 * Note that SND.UNA is not included to the range though being valid because
1001 * it means that the receiver is rather inconsistent with itself reporting
1002 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1003 * perfectly valid, however, in light of RFC2018 which explicitly states
1004 * that "SACK block MUST reflect the newest segment. Even if the newest
1005 * segment is going to be discarded ...", not that it looks very clever
1006 * in case of head skb. Due to potentional receiver driven attacks, we
1007 * choose to avoid immediate execution of a walk in write queue due to
1008 * reneging and defer head skb's loss recovery to standard loss recovery
1009 * procedure that will eventually trigger (nothing forbids us doing this).
1010 *
1011 * Implements also blockage to start_seq wrap-around. Problem lies in the
1012 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1013 * there's no guarantee that it will be before snd_nxt (n). The problem
1014 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1015 * wrap (s_w):
1016 *
1017 * <- outs wnd -> <- wrapzone ->
1018 * u e n u_w e_w s n_w
1019 * | | | | | | |
1020 * |<------------+------+----- TCP seqno space --------------+---------->|
1021 * ...-- <2^31 ->| |<--------...
1022 * ...---- >2^31 ------>| |<--------...
1023 *
1024 * Current code wouldn't be vulnerable but it's better still to discard such
1025 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1026 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1027 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1028 * equal to the ideal case (infinite seqno space without wrap caused issues).
1029 *
1030 * With D-SACK the lower bound is extended to cover sequence space below
1031 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1032 * again, D-SACK block must not to go across snd_una (for the same reason as
1033 * for the normal SACK blocks, explained above). But there all simplicity
1034 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1035 * fully below undo_marker they do not affect behavior in anyway and can
1036 * therefore be safely ignored. In rare cases (which are more or less
1037 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1038 * fragmentation and packet reordering past skb's retransmission. To consider
1039 * them correctly, the acceptable range must be extended even more though
1040 * the exact amount is rather hard to quantify. However, tp->max_window can
1041 * be used as an exaggerated estimate.
1042 */
1043static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1044 u32 start_seq, u32 end_seq)
1045{
1046 /* Too far in future, or reversed (interpretation is ambiguous) */
1047 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1048 return false;
1049
1050 /* Nasty start_seq wrap-around check (see comments above) */
1051 if (!before(start_seq, tp->snd_nxt))
1052 return false;
1053
1054 /* In outstanding window? ...This is valid exit for D-SACKs too.
1055 * start_seq == snd_una is non-sensical (see comments above)
1056 */
1057 if (after(start_seq, tp->snd_una))
1058 return true;
1059
1060 if (!is_dsack || !tp->undo_marker)
1061 return false;
1062
1063 /* ...Then it's D-SACK, and must reside below snd_una completely */
1064 if (after(end_seq, tp->snd_una))
1065 return false;
1066
1067 if (!before(start_seq, tp->undo_marker))
1068 return true;
1069
1070 /* Too old */
1071 if (!after(end_seq, tp->undo_marker))
1072 return false;
1073
1074 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1075 * start_seq < undo_marker and end_seq >= undo_marker.
1076 */
1077 return !before(start_seq, end_seq - tp->max_window);
1078}
1079
1080static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1081 struct tcp_sack_block_wire *sp, int num_sacks,
1082 u32 prior_snd_una)
1083{
1084 struct tcp_sock *tp = tcp_sk(sk);
1085 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1086 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1087 bool dup_sack = false;
1088
1089 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1090 dup_sack = true;
1091 tcp_dsack_seen(tp);
1092 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1093 } else if (num_sacks > 1) {
1094 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1095 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1096
1097 if (!after(end_seq_0, end_seq_1) &&
1098 !before(start_seq_0, start_seq_1)) {
1099 dup_sack = true;
1100 tcp_dsack_seen(tp);
1101 NET_INC_STATS(sock_net(sk),
1102 LINUX_MIB_TCPDSACKOFORECV);
1103 }
1104 }
1105
1106 /* D-SACK for already forgotten data... Do dumb counting. */
1107 if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 &&
1108 !after(end_seq_0, prior_snd_una) &&
1109 after(end_seq_0, tp->undo_marker))
1110 tp->undo_retrans--;
1111
1112 return dup_sack;
1113}
1114
1115struct tcp_sacktag_state {
1116 u32 reord;
1117 /* Timestamps for earliest and latest never-retransmitted segment
1118 * that was SACKed. RTO needs the earliest RTT to stay conservative,
1119 * but congestion control should still get an accurate delay signal.
1120 */
1121 u64 first_sackt;
1122 u64 last_sackt;
1123 struct rate_sample *rate;
1124 int flag;
1125 unsigned int mss_now;
1126};
1127
1128/* Check if skb is fully within the SACK block. In presence of GSO skbs,
1129 * the incoming SACK may not exactly match but we can find smaller MSS
1130 * aligned portion of it that matches. Therefore we might need to fragment
1131 * which may fail and creates some hassle (caller must handle error case
1132 * returns).
1133 *
1134 * FIXME: this could be merged to shift decision code
1135 */
1136static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1137 u32 start_seq, u32 end_seq)
1138{
1139 int err;
1140 bool in_sack;
1141 unsigned int pkt_len;
1142 unsigned int mss;
1143
1144 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1145 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1146
1147 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1148 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1149 mss = tcp_skb_mss(skb);
1150 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1151
1152 if (!in_sack) {
1153 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1154 if (pkt_len < mss)
1155 pkt_len = mss;
1156 } else {
1157 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1158 if (pkt_len < mss)
1159 return -EINVAL;
1160 }
1161
1162 /* Round if necessary so that SACKs cover only full MSSes
1163 * and/or the remaining small portion (if present)
1164 */
1165 if (pkt_len > mss) {
1166 unsigned int new_len = (pkt_len / mss) * mss;
1167 if (!in_sack && new_len < pkt_len)
1168 new_len += mss;
1169 pkt_len = new_len;
1170 }
1171
1172 if (pkt_len >= skb->len && !in_sack)
1173 return 0;
1174
1175 err = tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
1176 pkt_len, mss, GFP_ATOMIC);
1177 if (err < 0)
1178 return err;
1179 }
1180
1181 return in_sack;
1182}
1183
1184/* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1185static u8 tcp_sacktag_one(struct sock *sk,
1186 struct tcp_sacktag_state *state, u8 sacked,
1187 u32 start_seq, u32 end_seq,
1188 int dup_sack, int pcount,
1189 u64 xmit_time)
1190{
1191 struct tcp_sock *tp = tcp_sk(sk);
1192
1193 /* Account D-SACK for retransmitted packet. */
1194 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1195 if (tp->undo_marker && tp->undo_retrans > 0 &&
1196 after(end_seq, tp->undo_marker))
1197 tp->undo_retrans--;
1198 if ((sacked & TCPCB_SACKED_ACKED) &&
1199 before(start_seq, state->reord))
1200 state->reord = start_seq;
1201 }
1202
1203 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1204 if (!after(end_seq, tp->snd_una))
1205 return sacked;
1206
1207 if (!(sacked & TCPCB_SACKED_ACKED)) {
1208 tcp_rack_advance(tp, sacked, end_seq, xmit_time);
1209
1210 if (sacked & TCPCB_SACKED_RETRANS) {
1211 /* If the segment is not tagged as lost,
1212 * we do not clear RETRANS, believing
1213 * that retransmission is still in flight.
1214 */
1215 if (sacked & TCPCB_LOST) {
1216 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1217 tp->lost_out -= pcount;
1218 tp->retrans_out -= pcount;
1219 }
1220 } else {
1221 if (!(sacked & TCPCB_RETRANS)) {
1222 /* New sack for not retransmitted frame,
1223 * which was in hole. It is reordering.
1224 */
1225 if (before(start_seq,
1226 tcp_highest_sack_seq(tp)) &&
1227 before(start_seq, state->reord))
1228 state->reord = start_seq;
1229
1230 if (!after(end_seq, tp->high_seq))
1231 state->flag |= FLAG_ORIG_SACK_ACKED;
1232 if (state->first_sackt == 0)
1233 state->first_sackt = xmit_time;
1234 state->last_sackt = xmit_time;
1235 }
1236
1237 if (sacked & TCPCB_LOST) {
1238 sacked &= ~TCPCB_LOST;
1239 tp->lost_out -= pcount;
1240 }
1241 }
1242
1243 sacked |= TCPCB_SACKED_ACKED;
1244 state->flag |= FLAG_DATA_SACKED;
1245 tp->sacked_out += pcount;
1246 tp->delivered += pcount; /* Out-of-order packets delivered */
1247
1248 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1249 if (tp->lost_skb_hint &&
1250 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1251 tp->lost_cnt_hint += pcount;
1252 }
1253
1254 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1255 * frames and clear it. undo_retrans is decreased above, L|R frames
1256 * are accounted above as well.
1257 */
1258 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1259 sacked &= ~TCPCB_SACKED_RETRANS;
1260 tp->retrans_out -= pcount;
1261 }
1262
1263 return sacked;
1264}
1265
1266/* Shift newly-SACKed bytes from this skb to the immediately previous
1267 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1268 */
1269static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev,
1270 struct sk_buff *skb,
1271 struct tcp_sacktag_state *state,
1272 unsigned int pcount, int shifted, int mss,
1273 bool dup_sack)
1274{
1275 struct tcp_sock *tp = tcp_sk(sk);
1276 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1277 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1278
1279 BUG_ON(!pcount);
1280
1281 /* Adjust counters and hints for the newly sacked sequence
1282 * range but discard the return value since prev is already
1283 * marked. We must tag the range first because the seq
1284 * advancement below implicitly advances
1285 * tcp_highest_sack_seq() when skb is highest_sack.
1286 */
1287 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1288 start_seq, end_seq, dup_sack, pcount,
1289 tcp_skb_timestamp_us(skb));
1290 tcp_rate_skb_delivered(sk, skb, state->rate);
1291
1292 if (skb == tp->lost_skb_hint)
1293 tp->lost_cnt_hint += pcount;
1294
1295 TCP_SKB_CB(prev)->end_seq += shifted;
1296 TCP_SKB_CB(skb)->seq += shifted;
1297
1298 tcp_skb_pcount_add(prev, pcount);
1299 BUG_ON(tcp_skb_pcount(skb) < pcount);
1300 tcp_skb_pcount_add(skb, -pcount);
1301
1302 /* When we're adding to gso_segs == 1, gso_size will be zero,
1303 * in theory this shouldn't be necessary but as long as DSACK
1304 * code can come after this skb later on it's better to keep
1305 * setting gso_size to something.
1306 */
1307 if (!TCP_SKB_CB(prev)->tcp_gso_size)
1308 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1309
1310 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1311 if (tcp_skb_pcount(skb) <= 1)
1312 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1313
1314 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1315 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1316
1317 if (skb->len > 0) {
1318 BUG_ON(!tcp_skb_pcount(skb));
1319 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1320 return false;
1321 }
1322
1323 /* Whole SKB was eaten :-) */
1324
1325 if (skb == tp->retransmit_skb_hint)
1326 tp->retransmit_skb_hint = prev;
1327 if (skb == tp->lost_skb_hint) {
1328 tp->lost_skb_hint = prev;
1329 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1330 }
1331
1332 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1333 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor;
1334 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1335 TCP_SKB_CB(prev)->end_seq++;
1336
1337 if (skb == tcp_highest_sack(sk))
1338 tcp_advance_highest_sack(sk, skb);
1339
1340 tcp_skb_collapse_tstamp(prev, skb);
1341 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp))
1342 TCP_SKB_CB(prev)->tx.delivered_mstamp = 0;
1343
1344 tcp_rtx_queue_unlink_and_free(skb, sk);
1345
1346 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED);
1347
1348 return true;
1349}
1350
1351/* I wish gso_size would have a bit more sane initialization than
1352 * something-or-zero which complicates things
1353 */
1354static int tcp_skb_seglen(const struct sk_buff *skb)
1355{
1356 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1357}
1358
1359/* Shifting pages past head area doesn't work */
1360static int skb_can_shift(const struct sk_buff *skb)
1361{
1362 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1363}
1364
1365/* Try collapsing SACK blocks spanning across multiple skbs to a single
1366 * skb.
1367 */
1368static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1369 struct tcp_sacktag_state *state,
1370 u32 start_seq, u32 end_seq,
1371 bool dup_sack)
1372{
1373 struct tcp_sock *tp = tcp_sk(sk);
1374 struct sk_buff *prev;
1375 int mss;
1376 int pcount = 0;
1377 int len;
1378 int in_sack;
1379
1380 /* Normally R but no L won't result in plain S */
1381 if (!dup_sack &&
1382 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1383 goto fallback;
1384 if (!skb_can_shift(skb))
1385 goto fallback;
1386 /* This frame is about to be dropped (was ACKed). */
1387 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1388 goto fallback;
1389
1390 /* Can only happen with delayed DSACK + discard craziness */
1391 prev = skb_rb_prev(skb);
1392 if (!prev)
1393 goto fallback;
1394
1395 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1396 goto fallback;
1397
1398 if (!tcp_skb_can_collapse_to(prev))
1399 goto fallback;
1400
1401 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1402 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1403
1404 if (in_sack) {
1405 len = skb->len;
1406 pcount = tcp_skb_pcount(skb);
1407 mss = tcp_skb_seglen(skb);
1408
1409 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1410 * drop this restriction as unnecessary
1411 */
1412 if (mss != tcp_skb_seglen(prev))
1413 goto fallback;
1414 } else {
1415 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1416 goto noop;
1417 /* CHECKME: This is non-MSS split case only?, this will
1418 * cause skipped skbs due to advancing loop btw, original
1419 * has that feature too
1420 */
1421 if (tcp_skb_pcount(skb) <= 1)
1422 goto noop;
1423
1424 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1425 if (!in_sack) {
1426 /* TODO: head merge to next could be attempted here
1427 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1428 * though it might not be worth of the additional hassle
1429 *
1430 * ...we can probably just fallback to what was done
1431 * previously. We could try merging non-SACKed ones
1432 * as well but it probably isn't going to buy off
1433 * because later SACKs might again split them, and
1434 * it would make skb timestamp tracking considerably
1435 * harder problem.
1436 */
1437 goto fallback;
1438 }
1439
1440 len = end_seq - TCP_SKB_CB(skb)->seq;
1441 BUG_ON(len < 0);
1442 BUG_ON(len > skb->len);
1443
1444 /* MSS boundaries should be honoured or else pcount will
1445 * severely break even though it makes things bit trickier.
1446 * Optimize common case to avoid most of the divides
1447 */
1448 mss = tcp_skb_mss(skb);
1449
1450 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1451 * drop this restriction as unnecessary
1452 */
1453 if (mss != tcp_skb_seglen(prev))
1454 goto fallback;
1455
1456 if (len == mss) {
1457 pcount = 1;
1458 } else if (len < mss) {
1459 goto noop;
1460 } else {
1461 pcount = len / mss;
1462 len = pcount * mss;
1463 }
1464 }
1465
1466 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1467 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1468 goto fallback;
1469
1470 if (!skb_shift(prev, skb, len))
1471 goto fallback;
1472 if (!tcp_shifted_skb(sk, prev, skb, state, pcount, len, mss, dup_sack))
1473 goto out;
1474
1475 /* Hole filled allows collapsing with the next as well, this is very
1476 * useful when hole on every nth skb pattern happens
1477 */
1478 skb = skb_rb_next(prev);
1479 if (!skb)
1480 goto out;
1481
1482 if (!skb_can_shift(skb) ||
1483 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1484 (mss != tcp_skb_seglen(skb)))
1485 goto out;
1486
1487 len = skb->len;
1488 if (skb_shift(prev, skb, len)) {
1489 pcount += tcp_skb_pcount(skb);
1490 tcp_shifted_skb(sk, prev, skb, state, tcp_skb_pcount(skb),
1491 len, mss, 0);
1492 }
1493
1494out:
1495 return prev;
1496
1497noop:
1498 return skb;
1499
1500fallback:
1501 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1502 return NULL;
1503}
1504
1505static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1506 struct tcp_sack_block *next_dup,
1507 struct tcp_sacktag_state *state,
1508 u32 start_seq, u32 end_seq,
1509 bool dup_sack_in)
1510{
1511 struct tcp_sock *tp = tcp_sk(sk);
1512 struct sk_buff *tmp;
1513
1514 skb_rbtree_walk_from(skb) {
1515 int in_sack = 0;
1516 bool dup_sack = dup_sack_in;
1517
1518 /* queue is in-order => we can short-circuit the walk early */
1519 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1520 break;
1521
1522 if (next_dup &&
1523 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1524 in_sack = tcp_match_skb_to_sack(sk, skb,
1525 next_dup->start_seq,
1526 next_dup->end_seq);
1527 if (in_sack > 0)
1528 dup_sack = true;
1529 }
1530
1531 /* skb reference here is a bit tricky to get right, since
1532 * shifting can eat and free both this skb and the next,
1533 * so not even _safe variant of the loop is enough.
1534 */
1535 if (in_sack <= 0) {
1536 tmp = tcp_shift_skb_data(sk, skb, state,
1537 start_seq, end_seq, dup_sack);
1538 if (tmp) {
1539 if (tmp != skb) {
1540 skb = tmp;
1541 continue;
1542 }
1543
1544 in_sack = 0;
1545 } else {
1546 in_sack = tcp_match_skb_to_sack(sk, skb,
1547 start_seq,
1548 end_seq);
1549 }
1550 }
1551
1552 if (unlikely(in_sack < 0))
1553 break;
1554
1555 if (in_sack) {
1556 TCP_SKB_CB(skb)->sacked =
1557 tcp_sacktag_one(sk,
1558 state,
1559 TCP_SKB_CB(skb)->sacked,
1560 TCP_SKB_CB(skb)->seq,
1561 TCP_SKB_CB(skb)->end_seq,
1562 dup_sack,
1563 tcp_skb_pcount(skb),
1564 tcp_skb_timestamp_us(skb));
1565 tcp_rate_skb_delivered(sk, skb, state->rate);
1566 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1567 list_del_init(&skb->tcp_tsorted_anchor);
1568
1569 if (!before(TCP_SKB_CB(skb)->seq,
1570 tcp_highest_sack_seq(tp)))
1571 tcp_advance_highest_sack(sk, skb);
1572 }
1573 }
1574 return skb;
1575}
1576
1577static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk, u32 seq)
1578{
1579 struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node;
1580 struct sk_buff *skb;
1581
1582 while (*p) {
1583 parent = *p;
1584 skb = rb_to_skb(parent);
1585 if (before(seq, TCP_SKB_CB(skb)->seq)) {
1586 p = &parent->rb_left;
1587 continue;
1588 }
1589 if (!before(seq, TCP_SKB_CB(skb)->end_seq)) {
1590 p = &parent->rb_right;
1591 continue;
1592 }
1593 return skb;
1594 }
1595 return NULL;
1596}
1597
1598static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1599 u32 skip_to_seq)
1600{
1601 if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq))
1602 return skb;
1603
1604 return tcp_sacktag_bsearch(sk, skip_to_seq);
1605}
1606
1607static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1608 struct sock *sk,
1609 struct tcp_sack_block *next_dup,
1610 struct tcp_sacktag_state *state,
1611 u32 skip_to_seq)
1612{
1613 if (!next_dup)
1614 return skb;
1615
1616 if (before(next_dup->start_seq, skip_to_seq)) {
1617 skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq);
1618 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1619 next_dup->start_seq, next_dup->end_seq,
1620 1);
1621 }
1622
1623 return skb;
1624}
1625
1626static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1627{
1628 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1629}
1630
1631static int
1632tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1633 u32 prior_snd_una, struct tcp_sacktag_state *state)
1634{
1635 struct tcp_sock *tp = tcp_sk(sk);
1636 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1637 TCP_SKB_CB(ack_skb)->sacked);
1638 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1639 struct tcp_sack_block sp[TCP_NUM_SACKS];
1640 struct tcp_sack_block *cache;
1641 struct sk_buff *skb;
1642 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1643 int used_sacks;
1644 bool found_dup_sack = false;
1645 int i, j;
1646 int first_sack_index;
1647
1648 state->flag = 0;
1649 state->reord = tp->snd_nxt;
1650
1651 if (!tp->sacked_out)
1652 tcp_highest_sack_reset(sk);
1653
1654 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1655 num_sacks, prior_snd_una);
1656 if (found_dup_sack) {
1657 state->flag |= FLAG_DSACKING_ACK;
1658 tp->delivered++; /* A spurious retransmission is delivered */
1659 }
1660
1661 /* Eliminate too old ACKs, but take into
1662 * account more or less fresh ones, they can
1663 * contain valid SACK info.
1664 */
1665 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1666 return 0;
1667
1668 if (!tp->packets_out)
1669 goto out;
1670
1671 used_sacks = 0;
1672 first_sack_index = 0;
1673 for (i = 0; i < num_sacks; i++) {
1674 bool dup_sack = !i && found_dup_sack;
1675
1676 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1677 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1678
1679 if (!tcp_is_sackblock_valid(tp, dup_sack,
1680 sp[used_sacks].start_seq,
1681 sp[used_sacks].end_seq)) {
1682 int mib_idx;
1683
1684 if (dup_sack) {
1685 if (!tp->undo_marker)
1686 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1687 else
1688 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1689 } else {
1690 /* Don't count olds caused by ACK reordering */
1691 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1692 !after(sp[used_sacks].end_seq, tp->snd_una))
1693 continue;
1694 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1695 }
1696
1697 NET_INC_STATS(sock_net(sk), mib_idx);
1698 if (i == 0)
1699 first_sack_index = -1;
1700 continue;
1701 }
1702
1703 /* Ignore very old stuff early */
1704 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1705 continue;
1706
1707 used_sacks++;
1708 }
1709
1710 /* order SACK blocks to allow in order walk of the retrans queue */
1711 for (i = used_sacks - 1; i > 0; i--) {
1712 for (j = 0; j < i; j++) {
1713 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1714 swap(sp[j], sp[j + 1]);
1715
1716 /* Track where the first SACK block goes to */
1717 if (j == first_sack_index)
1718 first_sack_index = j + 1;
1719 }
1720 }
1721 }
1722
1723 state->mss_now = tcp_current_mss(sk);
1724 skb = NULL;
1725 i = 0;
1726
1727 if (!tp->sacked_out) {
1728 /* It's already past, so skip checking against it */
1729 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1730 } else {
1731 cache = tp->recv_sack_cache;
1732 /* Skip empty blocks in at head of the cache */
1733 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1734 !cache->end_seq)
1735 cache++;
1736 }
1737
1738 while (i < used_sacks) {
1739 u32 start_seq = sp[i].start_seq;
1740 u32 end_seq = sp[i].end_seq;
1741 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1742 struct tcp_sack_block *next_dup = NULL;
1743
1744 if (found_dup_sack && ((i + 1) == first_sack_index))
1745 next_dup = &sp[i + 1];
1746
1747 /* Skip too early cached blocks */
1748 while (tcp_sack_cache_ok(tp, cache) &&
1749 !before(start_seq, cache->end_seq))
1750 cache++;
1751
1752 /* Can skip some work by looking recv_sack_cache? */
1753 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1754 after(end_seq, cache->start_seq)) {
1755
1756 /* Head todo? */
1757 if (before(start_seq, cache->start_seq)) {
1758 skb = tcp_sacktag_skip(skb, sk, start_seq);
1759 skb = tcp_sacktag_walk(skb, sk, next_dup,
1760 state,
1761 start_seq,
1762 cache->start_seq,
1763 dup_sack);
1764 }
1765
1766 /* Rest of the block already fully processed? */
1767 if (!after(end_seq, cache->end_seq))
1768 goto advance_sp;
1769
1770 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1771 state,
1772 cache->end_seq);
1773
1774 /* ...tail remains todo... */
1775 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1776 /* ...but better entrypoint exists! */
1777 skb = tcp_highest_sack(sk);
1778 if (!skb)
1779 break;
1780 cache++;
1781 goto walk;
1782 }
1783
1784 skb = tcp_sacktag_skip(skb, sk, cache->end_seq);
1785 /* Check overlap against next cached too (past this one already) */
1786 cache++;
1787 continue;
1788 }
1789
1790 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1791 skb = tcp_highest_sack(sk);
1792 if (!skb)
1793 break;
1794 }
1795 skb = tcp_sacktag_skip(skb, sk, start_seq);
1796
1797walk:
1798 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
1799 start_seq, end_seq, dup_sack);
1800
1801advance_sp:
1802 i++;
1803 }
1804
1805 /* Clear the head of the cache sack blocks so we can skip it next time */
1806 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1807 tp->recv_sack_cache[i].start_seq = 0;
1808 tp->recv_sack_cache[i].end_seq = 0;
1809 }
1810 for (j = 0; j < used_sacks; j++)
1811 tp->recv_sack_cache[i++] = sp[j];
1812
1813 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss || tp->undo_marker)
1814 tcp_check_sack_reordering(sk, state->reord, 0);
1815
1816 tcp_verify_left_out(tp);
1817out:
1818
1819#if FASTRETRANS_DEBUG > 0
1820 WARN_ON((int)tp->sacked_out < 0);
1821 WARN_ON((int)tp->lost_out < 0);
1822 WARN_ON((int)tp->retrans_out < 0);
1823 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1824#endif
1825 return state->flag;
1826}
1827
1828/* Limits sacked_out so that sum with lost_out isn't ever larger than
1829 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1830 */
1831static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1832{
1833 u32 holes;
1834
1835 holes = max(tp->lost_out, 1U);
1836 holes = min(holes, tp->packets_out);
1837
1838 if ((tp->sacked_out + holes) > tp->packets_out) {
1839 tp->sacked_out = tp->packets_out - holes;
1840 return true;
1841 }
1842 return false;
1843}
1844
1845/* If we receive more dupacks than we expected counting segments
1846 * in assumption of absent reordering, interpret this as reordering.
1847 * The only another reason could be bug in receiver TCP.
1848 */
1849static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1850{
1851 struct tcp_sock *tp = tcp_sk(sk);
1852
1853 if (!tcp_limit_reno_sacked(tp))
1854 return;
1855
1856 tp->reordering = min_t(u32, tp->packets_out + addend,
1857 sock_net(sk)->ipv4.sysctl_tcp_max_reordering);
1858 tp->reord_seen++;
1859 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER);
1860}
1861
1862/* Emulate SACKs for SACKless connection: account for a new dupack. */
1863
1864static void tcp_add_reno_sack(struct sock *sk, int num_dupack)
1865{
1866 if (num_dupack) {
1867 struct tcp_sock *tp = tcp_sk(sk);
1868 u32 prior_sacked = tp->sacked_out;
1869 s32 delivered;
1870
1871 tp->sacked_out += num_dupack;
1872 tcp_check_reno_reordering(sk, 0);
1873 delivered = tp->sacked_out - prior_sacked;
1874 if (delivered > 0)
1875 tp->delivered += delivered;
1876 tcp_verify_left_out(tp);
1877 }
1878}
1879
1880/* Account for ACK, ACKing some data in Reno Recovery phase. */
1881
1882static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1883{
1884 struct tcp_sock *tp = tcp_sk(sk);
1885
1886 if (acked > 0) {
1887 /* One ACK acked hole. The rest eat duplicate ACKs. */
1888 tp->delivered += max_t(int, acked - tp->sacked_out, 1);
1889 if (acked - 1 >= tp->sacked_out)
1890 tp->sacked_out = 0;
1891 else
1892 tp->sacked_out -= acked - 1;
1893 }
1894 tcp_check_reno_reordering(sk, acked);
1895 tcp_verify_left_out(tp);
1896}
1897
1898static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1899{
1900 tp->sacked_out = 0;
1901}
1902
1903void tcp_clear_retrans(struct tcp_sock *tp)
1904{
1905 tp->retrans_out = 0;
1906 tp->lost_out = 0;
1907 tp->undo_marker = 0;
1908 tp->undo_retrans = -1;
1909 tp->sacked_out = 0;
1910}
1911
1912static inline void tcp_init_undo(struct tcp_sock *tp)
1913{
1914 tp->undo_marker = tp->snd_una;
1915 /* Retransmission still in flight may cause DSACKs later. */
1916 tp->undo_retrans = tp->retrans_out ? : -1;
1917}
1918
1919static bool tcp_is_rack(const struct sock *sk)
1920{
1921 return sock_net(sk)->ipv4.sysctl_tcp_recovery & TCP_RACK_LOSS_DETECTION;
1922}
1923
1924/* If we detect SACK reneging, forget all SACK information
1925 * and reset tags completely, otherwise preserve SACKs. If receiver
1926 * dropped its ofo queue, we will know this due to reneging detection.
1927 */
1928static void tcp_timeout_mark_lost(struct sock *sk)
1929{
1930 struct tcp_sock *tp = tcp_sk(sk);
1931 struct sk_buff *skb, *head;
1932 bool is_reneg; /* is receiver reneging on SACKs? */
1933
1934 head = tcp_rtx_queue_head(sk);
1935 is_reneg = head && (TCP_SKB_CB(head)->sacked & TCPCB_SACKED_ACKED);
1936 if (is_reneg) {
1937 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
1938 tp->sacked_out = 0;
1939 /* Mark SACK reneging until we recover from this loss event. */
1940 tp->is_sack_reneg = 1;
1941 } else if (tcp_is_reno(tp)) {
1942 tcp_reset_reno_sack(tp);
1943 }
1944
1945 skb = head;
1946 skb_rbtree_walk_from(skb) {
1947 if (is_reneg)
1948 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1949 else if (tcp_is_rack(sk) && skb != head &&
1950 tcp_rack_skb_timeout(tp, skb, 0) > 0)
1951 continue; /* Don't mark recently sent ones lost yet */
1952 tcp_mark_skb_lost(sk, skb);
1953 }
1954 tcp_verify_left_out(tp);
1955 tcp_clear_all_retrans_hints(tp);
1956}
1957
1958/* Enter Loss state. */
1959void tcp_enter_loss(struct sock *sk)
1960{
1961 const struct inet_connection_sock *icsk = inet_csk(sk);
1962 struct tcp_sock *tp = tcp_sk(sk);
1963 struct net *net = sock_net(sk);
1964 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
1965
1966 tcp_timeout_mark_lost(sk);
1967
1968 /* Reduce ssthresh if it has not yet been made inside this window. */
1969 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1970 !after(tp->high_seq, tp->snd_una) ||
1971 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1972 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1973 tp->prior_cwnd = tp->snd_cwnd;
1974 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1975 tcp_ca_event(sk, CA_EVENT_LOSS);
1976 tcp_init_undo(tp);
1977 }
1978 tp->snd_cwnd = tcp_packets_in_flight(tp) + 1;
1979 tp->snd_cwnd_cnt = 0;
1980 tp->snd_cwnd_stamp = tcp_jiffies32;
1981
1982 /* Timeout in disordered state after receiving substantial DUPACKs
1983 * suggests that the degree of reordering is over-estimated.
1984 */
1985 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
1986 tp->sacked_out >= net->ipv4.sysctl_tcp_reordering)
1987 tp->reordering = min_t(unsigned int, tp->reordering,
1988 net->ipv4.sysctl_tcp_reordering);
1989 tcp_set_ca_state(sk, TCP_CA_Loss);
1990 tp->high_seq = tp->snd_nxt;
1991 tcp_ecn_queue_cwr(tp);
1992
1993 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1994 * loss recovery is underway except recurring timeout(s) on
1995 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1996 */
1997 tp->frto = net->ipv4.sysctl_tcp_frto &&
1998 (new_recovery || icsk->icsk_retransmits) &&
1999 !inet_csk(sk)->icsk_mtup.probe_size;
2000}
2001
2002/* If ACK arrived pointing to a remembered SACK, it means that our
2003 * remembered SACKs do not reflect real state of receiver i.e.
2004 * receiver _host_ is heavily congested (or buggy).
2005 *
2006 * To avoid big spurious retransmission bursts due to transient SACK
2007 * scoreboard oddities that look like reneging, we give the receiver a
2008 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2009 * restore sanity to the SACK scoreboard. If the apparent reneging
2010 * persists until this RTO then we'll clear the SACK scoreboard.
2011 */
2012static bool tcp_check_sack_reneging(struct sock *sk, int flag)
2013{
2014 if (flag & FLAG_SACK_RENEGING) {
2015 struct tcp_sock *tp = tcp_sk(sk);
2016 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2017 msecs_to_jiffies(10));
2018
2019 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2020 delay, TCP_RTO_MAX);
2021 return true;
2022 }
2023 return false;
2024}
2025
2026/* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2027 * counter when SACK is enabled (without SACK, sacked_out is used for
2028 * that purpose).
2029 *
2030 * With reordering, holes may still be in flight, so RFC3517 recovery
2031 * uses pure sacked_out (total number of SACKed segments) even though
2032 * it violates the RFC that uses duplicate ACKs, often these are equal
2033 * but when e.g. out-of-window ACKs or packet duplication occurs,
2034 * they differ. Since neither occurs due to loss, TCP should really
2035 * ignore them.
2036 */
2037static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2038{
2039 return tp->sacked_out + 1;
2040}
2041
2042/* Linux NewReno/SACK/ECN state machine.
2043 * --------------------------------------
2044 *
2045 * "Open" Normal state, no dubious events, fast path.
2046 * "Disorder" In all the respects it is "Open",
2047 * but requires a bit more attention. It is entered when
2048 * we see some SACKs or dupacks. It is split of "Open"
2049 * mainly to move some processing from fast path to slow one.
2050 * "CWR" CWND was reduced due to some Congestion Notification event.
2051 * It can be ECN, ICMP source quench, local device congestion.
2052 * "Recovery" CWND was reduced, we are fast-retransmitting.
2053 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2054 *
2055 * tcp_fastretrans_alert() is entered:
2056 * - each incoming ACK, if state is not "Open"
2057 * - when arrived ACK is unusual, namely:
2058 * * SACK
2059 * * Duplicate ACK.
2060 * * ECN ECE.
2061 *
2062 * Counting packets in flight is pretty simple.
2063 *
2064 * in_flight = packets_out - left_out + retrans_out
2065 *
2066 * packets_out is SND.NXT-SND.UNA counted in packets.
2067 *
2068 * retrans_out is number of retransmitted segments.
2069 *
2070 * left_out is number of segments left network, but not ACKed yet.
2071 *
2072 * left_out = sacked_out + lost_out
2073 *
2074 * sacked_out: Packets, which arrived to receiver out of order
2075 * and hence not ACKed. With SACKs this number is simply
2076 * amount of SACKed data. Even without SACKs
2077 * it is easy to give pretty reliable estimate of this number,
2078 * counting duplicate ACKs.
2079 *
2080 * lost_out: Packets lost by network. TCP has no explicit
2081 * "loss notification" feedback from network (for now).
2082 * It means that this number can be only _guessed_.
2083 * Actually, it is the heuristics to predict lossage that
2084 * distinguishes different algorithms.
2085 *
2086 * F.e. after RTO, when all the queue is considered as lost,
2087 * lost_out = packets_out and in_flight = retrans_out.
2088 *
2089 * Essentially, we have now a few algorithms detecting
2090 * lost packets.
2091 *
2092 * If the receiver supports SACK:
2093 *
2094 * RFC6675/3517: It is the conventional algorithm. A packet is
2095 * considered lost if the number of higher sequence packets
2096 * SACKed is greater than or equal the DUPACK thoreshold
2097 * (reordering). This is implemented in tcp_mark_head_lost and
2098 * tcp_update_scoreboard.
2099 *
2100 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm
2101 * (2017-) that checks timing instead of counting DUPACKs.
2102 * Essentially a packet is considered lost if it's not S/ACKed
2103 * after RTT + reordering_window, where both metrics are
2104 * dynamically measured and adjusted. This is implemented in
2105 * tcp_rack_mark_lost.
2106 *
2107 * If the receiver does not support SACK:
2108 *
2109 * NewReno (RFC6582): in Recovery we assume that one segment
2110 * is lost (classic Reno). While we are in Recovery and
2111 * a partial ACK arrives, we assume that one more packet
2112 * is lost (NewReno). This heuristics are the same in NewReno
2113 * and SACK.
2114 *
2115 * Really tricky (and requiring careful tuning) part of algorithm
2116 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2117 * The first determines the moment _when_ we should reduce CWND and,
2118 * hence, slow down forward transmission. In fact, it determines the moment
2119 * when we decide that hole is caused by loss, rather than by a reorder.
2120 *
2121 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2122 * holes, caused by lost packets.
2123 *
2124 * And the most logically complicated part of algorithm is undo
2125 * heuristics. We detect false retransmits due to both too early
2126 * fast retransmit (reordering) and underestimated RTO, analyzing
2127 * timestamps and D-SACKs. When we detect that some segments were
2128 * retransmitted by mistake and CWND reduction was wrong, we undo
2129 * window reduction and abort recovery phase. This logic is hidden
2130 * inside several functions named tcp_try_undo_<something>.
2131 */
2132
2133/* This function decides, when we should leave Disordered state
2134 * and enter Recovery phase, reducing congestion window.
2135 *
2136 * Main question: may we further continue forward transmission
2137 * with the same cwnd?
2138 */
2139static bool tcp_time_to_recover(struct sock *sk, int flag)
2140{
2141 struct tcp_sock *tp = tcp_sk(sk);
2142
2143 /* Trick#1: The loss is proven. */
2144 if (tp->lost_out)
2145 return true;
2146
2147 /* Not-A-Trick#2 : Classic rule... */
2148 if (!tcp_is_rack(sk) && tcp_dupack_heuristics(tp) > tp->reordering)
2149 return true;
2150
2151 return false;
2152}
2153
2154/* Detect loss in event "A" above by marking head of queue up as lost.
2155 * For non-SACK(Reno) senders, the first "packets" number of segments
2156 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2157 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2158 * the maximum SACKed segments to pass before reaching this limit.
2159 */
2160static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2161{
2162 struct tcp_sock *tp = tcp_sk(sk);
2163 struct sk_buff *skb;
2164 int cnt, oldcnt, lost;
2165 unsigned int mss;
2166 /* Use SACK to deduce losses of new sequences sent during recovery */
2167 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq;
2168
2169 WARN_ON(packets > tp->packets_out);
2170 skb = tp->lost_skb_hint;
2171 if (skb) {
2172 /* Head already handled? */
2173 if (mark_head && after(TCP_SKB_CB(skb)->seq, tp->snd_una))
2174 return;
2175 cnt = tp->lost_cnt_hint;
2176 } else {
2177 skb = tcp_rtx_queue_head(sk);
2178 cnt = 0;
2179 }
2180
2181 skb_rbtree_walk_from(skb) {
2182 /* TODO: do this better */
2183 /* this is not the most efficient way to do this... */
2184 tp->lost_skb_hint = skb;
2185 tp->lost_cnt_hint = cnt;
2186
2187 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2188 break;
2189
2190 oldcnt = cnt;
2191 if (tcp_is_reno(tp) ||
2192 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2193 cnt += tcp_skb_pcount(skb);
2194
2195 if (cnt > packets) {
2196 if (tcp_is_sack(tp) ||
2197 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2198 (oldcnt >= packets))
2199 break;
2200
2201 mss = tcp_skb_mss(skb);
2202 /* If needed, chop off the prefix to mark as lost. */
2203 lost = (packets - oldcnt) * mss;
2204 if (lost < skb->len &&
2205 tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2206 lost, mss, GFP_ATOMIC) < 0)
2207 break;
2208 cnt = packets;
2209 }
2210
2211 tcp_skb_mark_lost(tp, skb);
2212
2213 if (mark_head)
2214 break;
2215 }
2216 tcp_verify_left_out(tp);
2217}
2218
2219/* Account newly detected lost packet(s) */
2220
2221static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2222{
2223 struct tcp_sock *tp = tcp_sk(sk);
2224
2225 if (tcp_is_sack(tp)) {
2226 int sacked_upto = tp->sacked_out - tp->reordering;
2227 if (sacked_upto >= 0)
2228 tcp_mark_head_lost(sk, sacked_upto, 0);
2229 else if (fast_rexmit)
2230 tcp_mark_head_lost(sk, 1, 1);
2231 }
2232}
2233
2234static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2235{
2236 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2237 before(tp->rx_opt.rcv_tsecr, when);
2238}
2239
2240/* skb is spurious retransmitted if the returned timestamp echo
2241 * reply is prior to the skb transmission time
2242 */
2243static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2244 const struct sk_buff *skb)
2245{
2246 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2247 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp(skb));
2248}
2249
2250/* Nothing was retransmitted or returned timestamp is less
2251 * than timestamp of the first retransmission.
2252 */
2253static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2254{
2255 return !tp->retrans_stamp ||
2256 tcp_tsopt_ecr_before(tp, tp->retrans_stamp);
2257}
2258
2259/* Undo procedures. */
2260
2261/* We can clear retrans_stamp when there are no retransmissions in the
2262 * window. It would seem that it is trivially available for us in
2263 * tp->retrans_out, however, that kind of assumptions doesn't consider
2264 * what will happen if errors occur when sending retransmission for the
2265 * second time. ...It could the that such segment has only
2266 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2267 * the head skb is enough except for some reneging corner cases that
2268 * are not worth the effort.
2269 *
2270 * Main reason for all this complexity is the fact that connection dying
2271 * time now depends on the validity of the retrans_stamp, in particular,
2272 * that successive retransmissions of a segment must not advance
2273 * retrans_stamp under any conditions.
2274 */
2275static bool tcp_any_retrans_done(const struct sock *sk)
2276{
2277 const struct tcp_sock *tp = tcp_sk(sk);
2278 struct sk_buff *skb;
2279
2280 if (tp->retrans_out)
2281 return true;
2282
2283 skb = tcp_rtx_queue_head(sk);
2284 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2285 return true;
2286
2287 return false;
2288}
2289
2290static void DBGUNDO(struct sock *sk, const char *msg)
2291{
2292#if FASTRETRANS_DEBUG > 1
2293 struct tcp_sock *tp = tcp_sk(sk);
2294 struct inet_sock *inet = inet_sk(sk);
2295
2296 if (sk->sk_family == AF_INET) {
2297 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2298 msg,
2299 &inet->inet_daddr, ntohs(inet->inet_dport),
2300 tp->snd_cwnd, tcp_left_out(tp),
2301 tp->snd_ssthresh, tp->prior_ssthresh,
2302 tp->packets_out);
2303 }
2304#if IS_ENABLED(CONFIG_IPV6)
2305 else if (sk->sk_family == AF_INET6) {
2306 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2307 msg,
2308 &sk->sk_v6_daddr, ntohs(inet->inet_dport),
2309 tp->snd_cwnd, tcp_left_out(tp),
2310 tp->snd_ssthresh, tp->prior_ssthresh,
2311 tp->packets_out);
2312 }
2313#endif
2314#endif
2315}
2316
2317static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2318{
2319 struct tcp_sock *tp = tcp_sk(sk);
2320
2321 if (unmark_loss) {
2322 struct sk_buff *skb;
2323
2324 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2325 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2326 }
2327 tp->lost_out = 0;
2328 tcp_clear_all_retrans_hints(tp);
2329 }
2330
2331 if (tp->prior_ssthresh) {
2332 const struct inet_connection_sock *icsk = inet_csk(sk);
2333
2334 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2335
2336 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2337 tp->snd_ssthresh = tp->prior_ssthresh;
2338 tcp_ecn_withdraw_cwr(tp);
2339 }
2340 }
2341 tp->snd_cwnd_stamp = tcp_jiffies32;
2342 tp->undo_marker = 0;
2343 tp->rack.advanced = 1; /* Force RACK to re-exam losses */
2344}
2345
2346static inline bool tcp_may_undo(const struct tcp_sock *tp)
2347{
2348 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2349}
2350
2351/* People celebrate: "We love our President!" */
2352static bool tcp_try_undo_recovery(struct sock *sk)
2353{
2354 struct tcp_sock *tp = tcp_sk(sk);
2355
2356 if (tcp_may_undo(tp)) {
2357 int mib_idx;
2358
2359 /* Happy end! We did not retransmit anything
2360 * or our original transmission succeeded.
2361 */
2362 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2363 tcp_undo_cwnd_reduction(sk, false);
2364 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2365 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2366 else
2367 mib_idx = LINUX_MIB_TCPFULLUNDO;
2368
2369 NET_INC_STATS(sock_net(sk), mib_idx);
2370 } else if (tp->rack.reo_wnd_persist) {
2371 tp->rack.reo_wnd_persist--;
2372 }
2373 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2374 /* Hold old state until something *above* high_seq
2375 * is ACKed. For Reno it is MUST to prevent false
2376 * fast retransmits (RFC2582). SACK TCP is safe. */
2377 if (!tcp_any_retrans_done(sk))
2378 tp->retrans_stamp = 0;
2379 return true;
2380 }
2381 tcp_set_ca_state(sk, TCP_CA_Open);
2382 tp->is_sack_reneg = 0;
2383 return false;
2384}
2385
2386/* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2387static bool tcp_try_undo_dsack(struct sock *sk)
2388{
2389 struct tcp_sock *tp = tcp_sk(sk);
2390
2391 if (tp->undo_marker && !tp->undo_retrans) {
2392 tp->rack.reo_wnd_persist = min(TCP_RACK_RECOVERY_THRESH,
2393 tp->rack.reo_wnd_persist + 1);
2394 DBGUNDO(sk, "D-SACK");
2395 tcp_undo_cwnd_reduction(sk, false);
2396 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2397 return true;
2398 }
2399 return false;
2400}
2401
2402/* Undo during loss recovery after partial ACK or using F-RTO. */
2403static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2404{
2405 struct tcp_sock *tp = tcp_sk(sk);
2406
2407 if (frto_undo || tcp_may_undo(tp)) {
2408 tcp_undo_cwnd_reduction(sk, true);
2409
2410 DBGUNDO(sk, "partial loss");
2411 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2412 if (frto_undo)
2413 NET_INC_STATS(sock_net(sk),
2414 LINUX_MIB_TCPSPURIOUSRTOS);
2415 inet_csk(sk)->icsk_retransmits = 0;
2416 if (frto_undo || tcp_is_sack(tp)) {
2417 tcp_set_ca_state(sk, TCP_CA_Open);
2418 tp->is_sack_reneg = 0;
2419 }
2420 return true;
2421 }
2422 return false;
2423}
2424
2425/* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2426 * It computes the number of packets to send (sndcnt) based on packets newly
2427 * delivered:
2428 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2429 * cwnd reductions across a full RTT.
2430 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2431 * But when the retransmits are acked without further losses, PRR
2432 * slow starts cwnd up to ssthresh to speed up the recovery.
2433 */
2434static void tcp_init_cwnd_reduction(struct sock *sk)
2435{
2436 struct tcp_sock *tp = tcp_sk(sk);
2437
2438 tp->high_seq = tp->snd_nxt;
2439 tp->tlp_high_seq = 0;
2440 tp->snd_cwnd_cnt = 0;
2441 tp->prior_cwnd = tp->snd_cwnd;
2442 tp->prr_delivered = 0;
2443 tp->prr_out = 0;
2444 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2445 tcp_ecn_queue_cwr(tp);
2446}
2447
2448void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int flag)
2449{
2450 struct tcp_sock *tp = tcp_sk(sk);
2451 int sndcnt = 0;
2452 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2453
2454 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2455 return;
2456
2457 tp->prr_delivered += newly_acked_sacked;
2458 if (delta < 0) {
2459 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2460 tp->prior_cwnd - 1;
2461 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2462 } else if ((flag & (FLAG_RETRANS_DATA_ACKED | FLAG_LOST_RETRANS)) ==
2463 FLAG_RETRANS_DATA_ACKED) {
2464 sndcnt = min_t(int, delta,
2465 max_t(int, tp->prr_delivered - tp->prr_out,
2466 newly_acked_sacked) + 1);
2467 } else {
2468 sndcnt = min(delta, newly_acked_sacked);
2469 }
2470 /* Force a fast retransmit upon entering fast recovery */
2471 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
2472 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2473}
2474
2475static inline void tcp_end_cwnd_reduction(struct sock *sk)
2476{
2477 struct tcp_sock *tp = tcp_sk(sk);
2478
2479 if (inet_csk(sk)->icsk_ca_ops->cong_control)
2480 return;
2481
2482 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2483 if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH &&
2484 (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) {
2485 tp->snd_cwnd = tp->snd_ssthresh;
2486 tp->snd_cwnd_stamp = tcp_jiffies32;
2487 }
2488 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2489}
2490
2491/* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2492void tcp_enter_cwr(struct sock *sk)
2493{
2494 struct tcp_sock *tp = tcp_sk(sk);
2495
2496 tp->prior_ssthresh = 0;
2497 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2498 tp->undo_marker = 0;
2499 tcp_init_cwnd_reduction(sk);
2500 tcp_set_ca_state(sk, TCP_CA_CWR);
2501 }
2502}
2503EXPORT_SYMBOL(tcp_enter_cwr);
2504
2505static void tcp_try_keep_open(struct sock *sk)
2506{
2507 struct tcp_sock *tp = tcp_sk(sk);
2508 int state = TCP_CA_Open;
2509
2510 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2511 state = TCP_CA_Disorder;
2512
2513 if (inet_csk(sk)->icsk_ca_state != state) {
2514 tcp_set_ca_state(sk, state);
2515 tp->high_seq = tp->snd_nxt;
2516 }
2517}
2518
2519static void tcp_try_to_open(struct sock *sk, int flag)
2520{
2521 struct tcp_sock *tp = tcp_sk(sk);
2522
2523 tcp_verify_left_out(tp);
2524
2525 if (!tcp_any_retrans_done(sk))
2526 tp->retrans_stamp = 0;
2527
2528 if (flag & FLAG_ECE)
2529 tcp_enter_cwr(sk);
2530
2531 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2532 tcp_try_keep_open(sk);
2533 }
2534}
2535
2536static void tcp_mtup_probe_failed(struct sock *sk)
2537{
2538 struct inet_connection_sock *icsk = inet_csk(sk);
2539
2540 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2541 icsk->icsk_mtup.probe_size = 0;
2542 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
2543}
2544
2545static void tcp_mtup_probe_success(struct sock *sk)
2546{
2547 struct tcp_sock *tp = tcp_sk(sk);
2548 struct inet_connection_sock *icsk = inet_csk(sk);
2549
2550 /* FIXME: breaks with very large cwnd */
2551 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2552 tp->snd_cwnd = tp->snd_cwnd *
2553 tcp_mss_to_mtu(sk, tp->mss_cache) /
2554 icsk->icsk_mtup.probe_size;
2555 tp->snd_cwnd_cnt = 0;
2556 tp->snd_cwnd_stamp = tcp_jiffies32;
2557 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2558
2559 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2560 icsk->icsk_mtup.probe_size = 0;
2561 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2562 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
2563}
2564
2565/* Do a simple retransmit without using the backoff mechanisms in
2566 * tcp_timer. This is used for path mtu discovery.
2567 * The socket is already locked here.
2568 */
2569void tcp_simple_retransmit(struct sock *sk)
2570{
2571 const struct inet_connection_sock *icsk = inet_csk(sk);
2572 struct tcp_sock *tp = tcp_sk(sk);
2573 struct sk_buff *skb;
2574 unsigned int mss = tcp_current_mss(sk);
2575
2576 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2577 if (tcp_skb_seglen(skb) > mss &&
2578 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2579 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2580 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2581 tp->retrans_out -= tcp_skb_pcount(skb);
2582 }
2583 tcp_skb_mark_lost_uncond_verify(tp, skb);
2584 }
2585 }
2586
2587 tcp_clear_retrans_hints_partial(tp);
2588
2589 if (!tp->lost_out)
2590 return;
2591
2592 if (tcp_is_reno(tp))
2593 tcp_limit_reno_sacked(tp);
2594
2595 tcp_verify_left_out(tp);
2596
2597 /* Don't muck with the congestion window here.
2598 * Reason is that we do not increase amount of _data_
2599 * in network, but units changed and effective
2600 * cwnd/ssthresh really reduced now.
2601 */
2602 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2603 tp->high_seq = tp->snd_nxt;
2604 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2605 tp->prior_ssthresh = 0;
2606 tp->undo_marker = 0;
2607 tcp_set_ca_state(sk, TCP_CA_Loss);
2608 }
2609 tcp_xmit_retransmit_queue(sk);
2610}
2611EXPORT_SYMBOL(tcp_simple_retransmit);
2612
2613void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2614{
2615 struct tcp_sock *tp = tcp_sk(sk);
2616 int mib_idx;
2617
2618 if (tcp_is_reno(tp))
2619 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2620 else
2621 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2622
2623 NET_INC_STATS(sock_net(sk), mib_idx);
2624
2625 tp->prior_ssthresh = 0;
2626 tcp_init_undo(tp);
2627
2628 if (!tcp_in_cwnd_reduction(sk)) {
2629 if (!ece_ack)
2630 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2631 tcp_init_cwnd_reduction(sk);
2632 }
2633 tcp_set_ca_state(sk, TCP_CA_Recovery);
2634}
2635
2636/* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2637 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2638 */
2639static void tcp_process_loss(struct sock *sk, int flag, int num_dupack,
2640 int *rexmit)
2641{
2642 struct tcp_sock *tp = tcp_sk(sk);
2643 bool recovered = !before(tp->snd_una, tp->high_seq);
2644
2645 if ((flag & FLAG_SND_UNA_ADVANCED) &&
2646 tcp_try_undo_loss(sk, false))
2647 return;
2648
2649 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2650 /* Step 3.b. A timeout is spurious if not all data are
2651 * lost, i.e., never-retransmitted data are (s)acked.
2652 */
2653 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2654 tcp_try_undo_loss(sk, true))
2655 return;
2656
2657 if (after(tp->snd_nxt, tp->high_seq)) {
2658 if (flag & FLAG_DATA_SACKED || num_dupack)
2659 tp->frto = 0; /* Step 3.a. loss was real */
2660 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2661 tp->high_seq = tp->snd_nxt;
2662 /* Step 2.b. Try send new data (but deferred until cwnd
2663 * is updated in tcp_ack()). Otherwise fall back to
2664 * the conventional recovery.
2665 */
2666 if (!tcp_write_queue_empty(sk) &&
2667 after(tcp_wnd_end(tp), tp->snd_nxt)) {
2668 *rexmit = REXMIT_NEW;
2669 return;
2670 }
2671 tp->frto = 0;
2672 }
2673 }
2674
2675 if (recovered) {
2676 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2677 tcp_try_undo_recovery(sk);
2678 return;
2679 }
2680 if (tcp_is_reno(tp)) {
2681 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2682 * delivered. Lower inflight to clock out (re)tranmissions.
2683 */
2684 if (after(tp->snd_nxt, tp->high_seq) && num_dupack)
2685 tcp_add_reno_sack(sk, num_dupack);
2686 else if (flag & FLAG_SND_UNA_ADVANCED)
2687 tcp_reset_reno_sack(tp);
2688 }
2689 *rexmit = REXMIT_LOST;
2690}
2691
2692/* Undo during fast recovery after partial ACK. */
2693static bool tcp_try_undo_partial(struct sock *sk, u32 prior_snd_una)
2694{
2695 struct tcp_sock *tp = tcp_sk(sk);
2696
2697 if (tp->undo_marker && tcp_packet_delayed(tp)) {
2698 /* Plain luck! Hole if filled with delayed
2699 * packet, rather than with a retransmit. Check reordering.
2700 */
2701 tcp_check_sack_reordering(sk, prior_snd_una, 1);
2702
2703 /* We are getting evidence that the reordering degree is higher
2704 * than we realized. If there are no retransmits out then we
2705 * can undo. Otherwise we clock out new packets but do not
2706 * mark more packets lost or retransmit more.
2707 */
2708 if (tp->retrans_out)
2709 return true;
2710
2711 if (!tcp_any_retrans_done(sk))
2712 tp->retrans_stamp = 0;
2713
2714 DBGUNDO(sk, "partial recovery");
2715 tcp_undo_cwnd_reduction(sk, true);
2716 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2717 tcp_try_keep_open(sk);
2718 return true;
2719 }
2720 return false;
2721}
2722
2723static void tcp_identify_packet_loss(struct sock *sk, int *ack_flag)
2724{
2725 struct tcp_sock *tp = tcp_sk(sk);
2726
2727 if (tcp_rtx_queue_empty(sk))
2728 return;
2729
2730 if (unlikely(tcp_is_reno(tp))) {
2731 tcp_newreno_mark_lost(sk, *ack_flag & FLAG_SND_UNA_ADVANCED);
2732 } else if (tcp_is_rack(sk)) {
2733 u32 prior_retrans = tp->retrans_out;
2734
2735 tcp_rack_mark_lost(sk);
2736 if (prior_retrans > tp->retrans_out)
2737 *ack_flag |= FLAG_LOST_RETRANS;
2738 }
2739}
2740
2741static bool tcp_force_fast_retransmit(struct sock *sk)
2742{
2743 struct tcp_sock *tp = tcp_sk(sk);
2744
2745 return after(tcp_highest_sack_seq(tp),
2746 tp->snd_una + tp->reordering * tp->mss_cache);
2747}
2748
2749/* Process an event, which can update packets-in-flight not trivially.
2750 * Main goal of this function is to calculate new estimate for left_out,
2751 * taking into account both packets sitting in receiver's buffer and
2752 * packets lost by network.
2753 *
2754 * Besides that it updates the congestion state when packet loss or ECN
2755 * is detected. But it does not reduce the cwnd, it is done by the
2756 * congestion control later.