Loading...
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/proto_memory.h>
76#include <net/inet_common.h>
77#include <linux/ipsec.h>
78#include <linux/unaligned.h>
79#include <linux/errqueue.h>
80#include <trace/events/tcp.h>
81#include <linux/jump_label_ratelimit.h>
82#include <net/busy_poll.h>
83#include <net/mptcp.h>
84
85int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
86
87#define FLAG_DATA 0x01 /* Incoming frame contained data. */
88#define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
89#define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
90#define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
91#define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
92#define FLAG_DATA_SACKED 0x20 /* New SACK. */
93#define FLAG_ECE 0x40 /* ECE in this ACK */
94#define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */
95#define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
96#define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
97#define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
98#define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
99#define FLAG_SET_XMIT_TIMER 0x1000 /* Set TLP or RTO timer */
100#define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
101#define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
102#define FLAG_NO_CHALLENGE_ACK 0x8000 /* do not call tcp_send_challenge_ack() */
103#define FLAG_ACK_MAYBE_DELAYED 0x10000 /* Likely a delayed ACK */
104#define FLAG_DSACK_TLP 0x20000 /* DSACK for tail loss probe */
105
106#define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
107#define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
108#define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE|FLAG_DSACKING_ACK)
109#define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
110
111#define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
112#define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
113
114#define REXMIT_NONE 0 /* no loss recovery to do */
115#define REXMIT_LOST 1 /* retransmit packets marked lost */
116#define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */
117
118#if IS_ENABLED(CONFIG_TLS_DEVICE)
119static DEFINE_STATIC_KEY_DEFERRED_FALSE(clean_acked_data_enabled, HZ);
120
121void clean_acked_data_enable(struct inet_connection_sock *icsk,
122 void (*cad)(struct sock *sk, u32 ack_seq))
123{
124 icsk->icsk_clean_acked = cad;
125 static_branch_deferred_inc(&clean_acked_data_enabled);
126}
127EXPORT_SYMBOL_GPL(clean_acked_data_enable);
128
129void clean_acked_data_disable(struct inet_connection_sock *icsk)
130{
131 static_branch_slow_dec_deferred(&clean_acked_data_enabled);
132 icsk->icsk_clean_acked = NULL;
133}
134EXPORT_SYMBOL_GPL(clean_acked_data_disable);
135
136void clean_acked_data_flush(void)
137{
138 static_key_deferred_flush(&clean_acked_data_enabled);
139}
140EXPORT_SYMBOL_GPL(clean_acked_data_flush);
141#endif
142
143#ifdef CONFIG_CGROUP_BPF
144static void bpf_skops_parse_hdr(struct sock *sk, struct sk_buff *skb)
145{
146 bool unknown_opt = tcp_sk(sk)->rx_opt.saw_unknown &&
147 BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk),
148 BPF_SOCK_OPS_PARSE_UNKNOWN_HDR_OPT_CB_FLAG);
149 bool parse_all_opt = BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk),
150 BPF_SOCK_OPS_PARSE_ALL_HDR_OPT_CB_FLAG);
151 struct bpf_sock_ops_kern sock_ops;
152
153 if (likely(!unknown_opt && !parse_all_opt))
154 return;
155
156 /* The skb will be handled in the
157 * bpf_skops_established() or
158 * bpf_skops_write_hdr_opt().
159 */
160 switch (sk->sk_state) {
161 case TCP_SYN_RECV:
162 case TCP_SYN_SENT:
163 case TCP_LISTEN:
164 return;
165 }
166
167 sock_owned_by_me(sk);
168
169 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
170 sock_ops.op = BPF_SOCK_OPS_PARSE_HDR_OPT_CB;
171 sock_ops.is_fullsock = 1;
172 sock_ops.sk = sk;
173 bpf_skops_init_skb(&sock_ops, skb, tcp_hdrlen(skb));
174
175 BPF_CGROUP_RUN_PROG_SOCK_OPS(&sock_ops);
176}
177
178static void bpf_skops_established(struct sock *sk, int bpf_op,
179 struct sk_buff *skb)
180{
181 struct bpf_sock_ops_kern sock_ops;
182
183 sock_owned_by_me(sk);
184
185 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
186 sock_ops.op = bpf_op;
187 sock_ops.is_fullsock = 1;
188 sock_ops.sk = sk;
189 /* sk with TCP_REPAIR_ON does not have skb in tcp_finish_connect */
190 if (skb)
191 bpf_skops_init_skb(&sock_ops, skb, tcp_hdrlen(skb));
192
193 BPF_CGROUP_RUN_PROG_SOCK_OPS(&sock_ops);
194}
195#else
196static void bpf_skops_parse_hdr(struct sock *sk, struct sk_buff *skb)
197{
198}
199
200static void bpf_skops_established(struct sock *sk, int bpf_op,
201 struct sk_buff *skb)
202{
203}
204#endif
205
206static __cold void tcp_gro_dev_warn(const struct sock *sk, const struct sk_buff *skb,
207 unsigned int len)
208{
209 struct net_device *dev;
210
211 rcu_read_lock();
212 dev = dev_get_by_index_rcu(sock_net(sk), skb->skb_iif);
213 if (!dev || len >= READ_ONCE(dev->mtu))
214 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n",
215 dev ? dev->name : "Unknown driver");
216 rcu_read_unlock();
217}
218
219/* Adapt the MSS value used to make delayed ack decision to the
220 * real world.
221 */
222static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
223{
224 struct inet_connection_sock *icsk = inet_csk(sk);
225 const unsigned int lss = icsk->icsk_ack.last_seg_size;
226 unsigned int len;
227
228 icsk->icsk_ack.last_seg_size = 0;
229
230 /* skb->len may jitter because of SACKs, even if peer
231 * sends good full-sized frames.
232 */
233 len = skb_shinfo(skb)->gso_size ? : skb->len;
234 if (len >= icsk->icsk_ack.rcv_mss) {
235 /* Note: divides are still a bit expensive.
236 * For the moment, only adjust scaling_ratio
237 * when we update icsk_ack.rcv_mss.
238 */
239 if (unlikely(len != icsk->icsk_ack.rcv_mss)) {
240 u64 val = (u64)skb->len << TCP_RMEM_TO_WIN_SCALE;
241 u8 old_ratio = tcp_sk(sk)->scaling_ratio;
242
243 do_div(val, skb->truesize);
244 tcp_sk(sk)->scaling_ratio = val ? val : 1;
245
246 if (old_ratio != tcp_sk(sk)->scaling_ratio) {
247 struct tcp_sock *tp = tcp_sk(sk);
248
249 val = tcp_win_from_space(sk, sk->sk_rcvbuf);
250 tcp_set_window_clamp(sk, val);
251
252 if (tp->window_clamp < tp->rcvq_space.space)
253 tp->rcvq_space.space = tp->window_clamp;
254 }
255 }
256 icsk->icsk_ack.rcv_mss = min_t(unsigned int, len,
257 tcp_sk(sk)->advmss);
258 /* Account for possibly-removed options */
259 DO_ONCE_LITE_IF(len > icsk->icsk_ack.rcv_mss + MAX_TCP_OPTION_SPACE,
260 tcp_gro_dev_warn, sk, skb, len);
261 /* If the skb has a len of exactly 1*MSS and has the PSH bit
262 * set then it is likely the end of an application write. So
263 * more data may not be arriving soon, and yet the data sender
264 * may be waiting for an ACK if cwnd-bound or using TX zero
265 * copy. So we set ICSK_ACK_PUSHED here so that
266 * tcp_cleanup_rbuf() will send an ACK immediately if the app
267 * reads all of the data and is not ping-pong. If len > MSS
268 * then this logic does not matter (and does not hurt) because
269 * tcp_cleanup_rbuf() will always ACK immediately if the app
270 * reads data and there is more than an MSS of unACKed data.
271 */
272 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_PSH)
273 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
274 } else {
275 /* Otherwise, we make more careful check taking into account,
276 * that SACKs block is variable.
277 *
278 * "len" is invariant segment length, including TCP header.
279 */
280 len += skb->data - skb_transport_header(skb);
281 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
282 /* If PSH is not set, packet should be
283 * full sized, provided peer TCP is not badly broken.
284 * This observation (if it is correct 8)) allows
285 * to handle super-low mtu links fairly.
286 */
287 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
288 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
289 /* Subtract also invariant (if peer is RFC compliant),
290 * tcp header plus fixed timestamp option length.
291 * Resulting "len" is MSS free of SACK jitter.
292 */
293 len -= tcp_sk(sk)->tcp_header_len;
294 icsk->icsk_ack.last_seg_size = len;
295 if (len == lss) {
296 icsk->icsk_ack.rcv_mss = len;
297 return;
298 }
299 }
300 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
301 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
302 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
303 }
304}
305
306static void tcp_incr_quickack(struct sock *sk, unsigned int max_quickacks)
307{
308 struct inet_connection_sock *icsk = inet_csk(sk);
309 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
310
311 if (quickacks == 0)
312 quickacks = 2;
313 quickacks = min(quickacks, max_quickacks);
314 if (quickacks > icsk->icsk_ack.quick)
315 icsk->icsk_ack.quick = quickacks;
316}
317
318static void tcp_enter_quickack_mode(struct sock *sk, unsigned int max_quickacks)
319{
320 struct inet_connection_sock *icsk = inet_csk(sk);
321
322 tcp_incr_quickack(sk, max_quickacks);
323 inet_csk_exit_pingpong_mode(sk);
324 icsk->icsk_ack.ato = TCP_ATO_MIN;
325}
326
327/* Send ACKs quickly, if "quick" count is not exhausted
328 * and the session is not interactive.
329 */
330
331static bool tcp_in_quickack_mode(struct sock *sk)
332{
333 const struct inet_connection_sock *icsk = inet_csk(sk);
334 const struct dst_entry *dst = __sk_dst_get(sk);
335
336 return (dst && dst_metric(dst, RTAX_QUICKACK)) ||
337 (icsk->icsk_ack.quick && !inet_csk_in_pingpong_mode(sk));
338}
339
340static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
341{
342 if (tp->ecn_flags & TCP_ECN_OK)
343 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
344}
345
346static void tcp_ecn_accept_cwr(struct sock *sk, const struct sk_buff *skb)
347{
348 if (tcp_hdr(skb)->cwr) {
349 tcp_sk(sk)->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
350
351 /* If the sender is telling us it has entered CWR, then its
352 * cwnd may be very low (even just 1 packet), so we should ACK
353 * immediately.
354 */
355 if (TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq)
356 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
357 }
358}
359
360static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
361{
362 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
363}
364
365static void __tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
366{
367 struct tcp_sock *tp = tcp_sk(sk);
368
369 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
370 case INET_ECN_NOT_ECT:
371 /* Funny extension: if ECT is not set on a segment,
372 * and we already seen ECT on a previous segment,
373 * it is probably a retransmit.
374 */
375 if (tp->ecn_flags & TCP_ECN_SEEN)
376 tcp_enter_quickack_mode(sk, 2);
377 break;
378 case INET_ECN_CE:
379 if (tcp_ca_needs_ecn(sk))
380 tcp_ca_event(sk, CA_EVENT_ECN_IS_CE);
381
382 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
383 /* Better not delay acks, sender can have a very low cwnd */
384 tcp_enter_quickack_mode(sk, 2);
385 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
386 }
387 tp->ecn_flags |= TCP_ECN_SEEN;
388 break;
389 default:
390 if (tcp_ca_needs_ecn(sk))
391 tcp_ca_event(sk, CA_EVENT_ECN_NO_CE);
392 tp->ecn_flags |= TCP_ECN_SEEN;
393 break;
394 }
395}
396
397static void tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
398{
399 if (tcp_sk(sk)->ecn_flags & TCP_ECN_OK)
400 __tcp_ecn_check_ce(sk, skb);
401}
402
403static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
404{
405 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
406 tp->ecn_flags &= ~TCP_ECN_OK;
407}
408
409static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
410{
411 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
412 tp->ecn_flags &= ~TCP_ECN_OK;
413}
414
415static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
416{
417 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
418 return true;
419 return false;
420}
421
422/* Buffer size and advertised window tuning.
423 *
424 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
425 */
426
427static void tcp_sndbuf_expand(struct sock *sk)
428{
429 const struct tcp_sock *tp = tcp_sk(sk);
430 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
431 int sndmem, per_mss;
432 u32 nr_segs;
433
434 /* Worst case is non GSO/TSO : each frame consumes one skb
435 * and skb->head is kmalloced using power of two area of memory
436 */
437 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
438 MAX_TCP_HEADER +
439 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
440
441 per_mss = roundup_pow_of_two(per_mss) +
442 SKB_DATA_ALIGN(sizeof(struct sk_buff));
443
444 nr_segs = max_t(u32, TCP_INIT_CWND, tcp_snd_cwnd(tp));
445 nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
446
447 /* Fast Recovery (RFC 5681 3.2) :
448 * Cubic needs 1.7 factor, rounded to 2 to include
449 * extra cushion (application might react slowly to EPOLLOUT)
450 */
451 sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2;
452 sndmem *= nr_segs * per_mss;
453
454 if (sk->sk_sndbuf < sndmem)
455 WRITE_ONCE(sk->sk_sndbuf,
456 min(sndmem, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_wmem[2])));
457}
458
459/* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
460 *
461 * All tcp_full_space() is split to two parts: "network" buffer, allocated
462 * forward and advertised in receiver window (tp->rcv_wnd) and
463 * "application buffer", required to isolate scheduling/application
464 * latencies from network.
465 * window_clamp is maximal advertised window. It can be less than
466 * tcp_full_space(), in this case tcp_full_space() - window_clamp
467 * is reserved for "application" buffer. The less window_clamp is
468 * the smoother our behaviour from viewpoint of network, but the lower
469 * throughput and the higher sensitivity of the connection to losses. 8)
470 *
471 * rcv_ssthresh is more strict window_clamp used at "slow start"
472 * phase to predict further behaviour of this connection.
473 * It is used for two goals:
474 * - to enforce header prediction at sender, even when application
475 * requires some significant "application buffer". It is check #1.
476 * - to prevent pruning of receive queue because of misprediction
477 * of receiver window. Check #2.
478 *
479 * The scheme does not work when sender sends good segments opening
480 * window and then starts to feed us spaghetti. But it should work
481 * in common situations. Otherwise, we have to rely on queue collapsing.
482 */
483
484/* Slow part of check#2. */
485static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb,
486 unsigned int skbtruesize)
487{
488 const struct tcp_sock *tp = tcp_sk(sk);
489 /* Optimize this! */
490 int truesize = tcp_win_from_space(sk, skbtruesize) >> 1;
491 int window = tcp_win_from_space(sk, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2])) >> 1;
492
493 while (tp->rcv_ssthresh <= window) {
494 if (truesize <= skb->len)
495 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
496
497 truesize >>= 1;
498 window >>= 1;
499 }
500 return 0;
501}
502
503/* Even if skb appears to have a bad len/truesize ratio, TCP coalescing
504 * can play nice with us, as sk_buff and skb->head might be either
505 * freed or shared with up to MAX_SKB_FRAGS segments.
506 * Only give a boost to drivers using page frag(s) to hold the frame(s),
507 * and if no payload was pulled in skb->head before reaching us.
508 */
509static u32 truesize_adjust(bool adjust, const struct sk_buff *skb)
510{
511 u32 truesize = skb->truesize;
512
513 if (adjust && !skb_headlen(skb)) {
514 truesize -= SKB_TRUESIZE(skb_end_offset(skb));
515 /* paranoid check, some drivers might be buggy */
516 if (unlikely((int)truesize < (int)skb->len))
517 truesize = skb->truesize;
518 }
519 return truesize;
520}
521
522static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb,
523 bool adjust)
524{
525 struct tcp_sock *tp = tcp_sk(sk);
526 int room;
527
528 room = min_t(int, tp->window_clamp, tcp_space(sk)) - tp->rcv_ssthresh;
529
530 if (room <= 0)
531 return;
532
533 /* Check #1 */
534 if (!tcp_under_memory_pressure(sk)) {
535 unsigned int truesize = truesize_adjust(adjust, skb);
536 int incr;
537
538 /* Check #2. Increase window, if skb with such overhead
539 * will fit to rcvbuf in future.
540 */
541 if (tcp_win_from_space(sk, truesize) <= skb->len)
542 incr = 2 * tp->advmss;
543 else
544 incr = __tcp_grow_window(sk, skb, truesize);
545
546 if (incr) {
547 incr = max_t(int, incr, 2 * skb->len);
548 tp->rcv_ssthresh += min(room, incr);
549 inet_csk(sk)->icsk_ack.quick |= 1;
550 }
551 } else {
552 /* Under pressure:
553 * Adjust rcv_ssthresh according to reserved mem
554 */
555 tcp_adjust_rcv_ssthresh(sk);
556 }
557}
558
559/* 3. Try to fixup all. It is made immediately after connection enters
560 * established state.
561 */
562static void tcp_init_buffer_space(struct sock *sk)
563{
564 int tcp_app_win = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_app_win);
565 struct tcp_sock *tp = tcp_sk(sk);
566 int maxwin;
567
568 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
569 tcp_sndbuf_expand(sk);
570
571 tcp_mstamp_refresh(tp);
572 tp->rcvq_space.time = tp->tcp_mstamp;
573 tp->rcvq_space.seq = tp->copied_seq;
574
575 maxwin = tcp_full_space(sk);
576
577 if (tp->window_clamp >= maxwin) {
578 WRITE_ONCE(tp->window_clamp, maxwin);
579
580 if (tcp_app_win && maxwin > 4 * tp->advmss)
581 WRITE_ONCE(tp->window_clamp,
582 max(maxwin - (maxwin >> tcp_app_win),
583 4 * tp->advmss));
584 }
585
586 /* Force reservation of one segment. */
587 if (tcp_app_win &&
588 tp->window_clamp > 2 * tp->advmss &&
589 tp->window_clamp + tp->advmss > maxwin)
590 WRITE_ONCE(tp->window_clamp,
591 max(2 * tp->advmss, maxwin - tp->advmss));
592
593 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
594 tp->snd_cwnd_stamp = tcp_jiffies32;
595 tp->rcvq_space.space = min3(tp->rcv_ssthresh, tp->rcv_wnd,
596 (u32)TCP_INIT_CWND * tp->advmss);
597}
598
599/* 4. Recalculate window clamp after socket hit its memory bounds. */
600static void tcp_clamp_window(struct sock *sk)
601{
602 struct tcp_sock *tp = tcp_sk(sk);
603 struct inet_connection_sock *icsk = inet_csk(sk);
604 struct net *net = sock_net(sk);
605 int rmem2;
606
607 icsk->icsk_ack.quick = 0;
608 rmem2 = READ_ONCE(net->ipv4.sysctl_tcp_rmem[2]);
609
610 if (sk->sk_rcvbuf < rmem2 &&
611 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
612 !tcp_under_memory_pressure(sk) &&
613 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
614 WRITE_ONCE(sk->sk_rcvbuf,
615 min(atomic_read(&sk->sk_rmem_alloc), rmem2));
616 }
617 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
618 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
619}
620
621/* Initialize RCV_MSS value.
622 * RCV_MSS is an our guess about MSS used by the peer.
623 * We haven't any direct information about the MSS.
624 * It's better to underestimate the RCV_MSS rather than overestimate.
625 * Overestimations make us ACKing less frequently than needed.
626 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
627 */
628void tcp_initialize_rcv_mss(struct sock *sk)
629{
630 const struct tcp_sock *tp = tcp_sk(sk);
631 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
632
633 hint = min(hint, tp->rcv_wnd / 2);
634 hint = min(hint, TCP_MSS_DEFAULT);
635 hint = max(hint, TCP_MIN_MSS);
636
637 inet_csk(sk)->icsk_ack.rcv_mss = hint;
638}
639EXPORT_SYMBOL(tcp_initialize_rcv_mss);
640
641/* Receiver "autotuning" code.
642 *
643 * The algorithm for RTT estimation w/o timestamps is based on
644 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
645 * <https://public.lanl.gov/radiant/pubs.html#DRS>
646 *
647 * More detail on this code can be found at
648 * <http://staff.psc.edu/jheffner/>,
649 * though this reference is out of date. A new paper
650 * is pending.
651 */
652static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
653{
654 u32 new_sample = tp->rcv_rtt_est.rtt_us;
655 long m = sample;
656
657 if (new_sample != 0) {
658 /* If we sample in larger samples in the non-timestamp
659 * case, we could grossly overestimate the RTT especially
660 * with chatty applications or bulk transfer apps which
661 * are stalled on filesystem I/O.
662 *
663 * Also, since we are only going for a minimum in the
664 * non-timestamp case, we do not smooth things out
665 * else with timestamps disabled convergence takes too
666 * long.
667 */
668 if (!win_dep) {
669 m -= (new_sample >> 3);
670 new_sample += m;
671 } else {
672 m <<= 3;
673 if (m < new_sample)
674 new_sample = m;
675 }
676 } else {
677 /* No previous measure. */
678 new_sample = m << 3;
679 }
680
681 tp->rcv_rtt_est.rtt_us = new_sample;
682}
683
684static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
685{
686 u32 delta_us;
687
688 if (tp->rcv_rtt_est.time == 0)
689 goto new_measure;
690 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
691 return;
692 delta_us = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcv_rtt_est.time);
693 if (!delta_us)
694 delta_us = 1;
695 tcp_rcv_rtt_update(tp, delta_us, 1);
696
697new_measure:
698 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
699 tp->rcv_rtt_est.time = tp->tcp_mstamp;
700}
701
702static s32 tcp_rtt_tsopt_us(const struct tcp_sock *tp)
703{
704 u32 delta, delta_us;
705
706 delta = tcp_time_stamp_ts(tp) - tp->rx_opt.rcv_tsecr;
707 if (tp->tcp_usec_ts)
708 return delta;
709
710 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) {
711 if (!delta)
712 delta = 1;
713 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
714 return delta_us;
715 }
716 return -1;
717}
718
719static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
720 const struct sk_buff *skb)
721{
722 struct tcp_sock *tp = tcp_sk(sk);
723
724 if (tp->rx_opt.rcv_tsecr == tp->rcv_rtt_last_tsecr)
725 return;
726 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
727
728 if (TCP_SKB_CB(skb)->end_seq -
729 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss) {
730 s32 delta = tcp_rtt_tsopt_us(tp);
731
732 if (delta >= 0)
733 tcp_rcv_rtt_update(tp, delta, 0);
734 }
735}
736
737/*
738 * This function should be called every time data is copied to user space.
739 * It calculates the appropriate TCP receive buffer space.
740 */
741void tcp_rcv_space_adjust(struct sock *sk)
742{
743 struct tcp_sock *tp = tcp_sk(sk);
744 u32 copied;
745 int time;
746
747 trace_tcp_rcv_space_adjust(sk);
748
749 tcp_mstamp_refresh(tp);
750 time = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcvq_space.time);
751 if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0)
752 return;
753
754 /* Number of bytes copied to user in last RTT */
755 copied = tp->copied_seq - tp->rcvq_space.seq;
756 if (copied <= tp->rcvq_space.space)
757 goto new_measure;
758
759 /* A bit of theory :
760 * copied = bytes received in previous RTT, our base window
761 * To cope with packet losses, we need a 2x factor
762 * To cope with slow start, and sender growing its cwin by 100 %
763 * every RTT, we need a 4x factor, because the ACK we are sending
764 * now is for the next RTT, not the current one :
765 * <prev RTT . ><current RTT .. ><next RTT .... >
766 */
767
768 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf) &&
769 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
770 u64 rcvwin, grow;
771 int rcvbuf;
772
773 /* minimal window to cope with packet losses, assuming
774 * steady state. Add some cushion because of small variations.
775 */
776 rcvwin = ((u64)copied << 1) + 16 * tp->advmss;
777
778 /* Accommodate for sender rate increase (eg. slow start) */
779 grow = rcvwin * (copied - tp->rcvq_space.space);
780 do_div(grow, tp->rcvq_space.space);
781 rcvwin += (grow << 1);
782
783 rcvbuf = min_t(u64, tcp_space_from_win(sk, rcvwin),
784 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
785 if (rcvbuf > sk->sk_rcvbuf) {
786 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
787
788 /* Make the window clamp follow along. */
789 WRITE_ONCE(tp->window_clamp,
790 tcp_win_from_space(sk, rcvbuf));
791 }
792 }
793 tp->rcvq_space.space = copied;
794
795new_measure:
796 tp->rcvq_space.seq = tp->copied_seq;
797 tp->rcvq_space.time = tp->tcp_mstamp;
798}
799
800static void tcp_save_lrcv_flowlabel(struct sock *sk, const struct sk_buff *skb)
801{
802#if IS_ENABLED(CONFIG_IPV6)
803 struct inet_connection_sock *icsk = inet_csk(sk);
804
805 if (skb->protocol == htons(ETH_P_IPV6))
806 icsk->icsk_ack.lrcv_flowlabel = ntohl(ip6_flowlabel(ipv6_hdr(skb)));
807#endif
808}
809
810/* There is something which you must keep in mind when you analyze the
811 * behavior of the tp->ato delayed ack timeout interval. When a
812 * connection starts up, we want to ack as quickly as possible. The
813 * problem is that "good" TCP's do slow start at the beginning of data
814 * transmission. The means that until we send the first few ACK's the
815 * sender will sit on his end and only queue most of his data, because
816 * he can only send snd_cwnd unacked packets at any given time. For
817 * each ACK we send, he increments snd_cwnd and transmits more of his
818 * queue. -DaveM
819 */
820static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
821{
822 struct tcp_sock *tp = tcp_sk(sk);
823 struct inet_connection_sock *icsk = inet_csk(sk);
824 u32 now;
825
826 inet_csk_schedule_ack(sk);
827
828 tcp_measure_rcv_mss(sk, skb);
829
830 tcp_rcv_rtt_measure(tp);
831
832 now = tcp_jiffies32;
833
834 if (!icsk->icsk_ack.ato) {
835 /* The _first_ data packet received, initialize
836 * delayed ACK engine.
837 */
838 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
839 icsk->icsk_ack.ato = TCP_ATO_MIN;
840 } else {
841 int m = now - icsk->icsk_ack.lrcvtime;
842
843 if (m <= TCP_ATO_MIN / 2) {
844 /* The fastest case is the first. */
845 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
846 } else if (m < icsk->icsk_ack.ato) {
847 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
848 if (icsk->icsk_ack.ato > icsk->icsk_rto)
849 icsk->icsk_ack.ato = icsk->icsk_rto;
850 } else if (m > icsk->icsk_rto) {
851 /* Too long gap. Apparently sender failed to
852 * restart window, so that we send ACKs quickly.
853 */
854 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
855 }
856 }
857 icsk->icsk_ack.lrcvtime = now;
858 tcp_save_lrcv_flowlabel(sk, skb);
859
860 tcp_ecn_check_ce(sk, skb);
861
862 if (skb->len >= 128)
863 tcp_grow_window(sk, skb, true);
864}
865
866/* Called to compute a smoothed rtt estimate. The data fed to this
867 * routine either comes from timestamps, or from segments that were
868 * known _not_ to have been retransmitted [see Karn/Partridge
869 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
870 * piece by Van Jacobson.
871 * NOTE: the next three routines used to be one big routine.
872 * To save cycles in the RFC 1323 implementation it was better to break
873 * it up into three procedures. -- erics
874 */
875static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
876{
877 struct tcp_sock *tp = tcp_sk(sk);
878 long m = mrtt_us; /* RTT */
879 u32 srtt = tp->srtt_us;
880
881 /* The following amusing code comes from Jacobson's
882 * article in SIGCOMM '88. Note that rtt and mdev
883 * are scaled versions of rtt and mean deviation.
884 * This is designed to be as fast as possible
885 * m stands for "measurement".
886 *
887 * On a 1990 paper the rto value is changed to:
888 * RTO = rtt + 4 * mdev
889 *
890 * Funny. This algorithm seems to be very broken.
891 * These formulae increase RTO, when it should be decreased, increase
892 * too slowly, when it should be increased quickly, decrease too quickly
893 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
894 * does not matter how to _calculate_ it. Seems, it was trap
895 * that VJ failed to avoid. 8)
896 */
897 if (srtt != 0) {
898 m -= (srtt >> 3); /* m is now error in rtt est */
899 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
900 if (m < 0) {
901 m = -m; /* m is now abs(error) */
902 m -= (tp->mdev_us >> 2); /* similar update on mdev */
903 /* This is similar to one of Eifel findings.
904 * Eifel blocks mdev updates when rtt decreases.
905 * This solution is a bit different: we use finer gain
906 * for mdev in this case (alpha*beta).
907 * Like Eifel it also prevents growth of rto,
908 * but also it limits too fast rto decreases,
909 * happening in pure Eifel.
910 */
911 if (m > 0)
912 m >>= 3;
913 } else {
914 m -= (tp->mdev_us >> 2); /* similar update on mdev */
915 }
916 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
917 if (tp->mdev_us > tp->mdev_max_us) {
918 tp->mdev_max_us = tp->mdev_us;
919 if (tp->mdev_max_us > tp->rttvar_us)
920 tp->rttvar_us = tp->mdev_max_us;
921 }
922 if (after(tp->snd_una, tp->rtt_seq)) {
923 if (tp->mdev_max_us < tp->rttvar_us)
924 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
925 tp->rtt_seq = tp->snd_nxt;
926 tp->mdev_max_us = tcp_rto_min_us(sk);
927
928 tcp_bpf_rtt(sk, mrtt_us, srtt);
929 }
930 } else {
931 /* no previous measure. */
932 srtt = m << 3; /* take the measured time to be rtt */
933 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
934 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
935 tp->mdev_max_us = tp->rttvar_us;
936 tp->rtt_seq = tp->snd_nxt;
937
938 tcp_bpf_rtt(sk, mrtt_us, srtt);
939 }
940 tp->srtt_us = max(1U, srtt);
941}
942
943static void tcp_update_pacing_rate(struct sock *sk)
944{
945 const struct tcp_sock *tp = tcp_sk(sk);
946 u64 rate;
947
948 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
949 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
950
951 /* current rate is (cwnd * mss) / srtt
952 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
953 * In Congestion Avoidance phase, set it to 120 % the current rate.
954 *
955 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
956 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
957 * end of slow start and should slow down.
958 */
959 if (tcp_snd_cwnd(tp) < tp->snd_ssthresh / 2)
960 rate *= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_pacing_ss_ratio);
961 else
962 rate *= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_pacing_ca_ratio);
963
964 rate *= max(tcp_snd_cwnd(tp), tp->packets_out);
965
966 if (likely(tp->srtt_us))
967 do_div(rate, tp->srtt_us);
968
969 /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate
970 * without any lock. We want to make sure compiler wont store
971 * intermediate values in this location.
972 */
973 WRITE_ONCE(sk->sk_pacing_rate,
974 min_t(u64, rate, READ_ONCE(sk->sk_max_pacing_rate)));
975}
976
977/* Calculate rto without backoff. This is the second half of Van Jacobson's
978 * routine referred to above.
979 */
980static void tcp_set_rto(struct sock *sk)
981{
982 const struct tcp_sock *tp = tcp_sk(sk);
983 /* Old crap is replaced with new one. 8)
984 *
985 * More seriously:
986 * 1. If rtt variance happened to be less 50msec, it is hallucination.
987 * It cannot be less due to utterly erratic ACK generation made
988 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
989 * to do with delayed acks, because at cwnd>2 true delack timeout
990 * is invisible. Actually, Linux-2.4 also generates erratic
991 * ACKs in some circumstances.
992 */
993 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
994
995 /* 2. Fixups made earlier cannot be right.
996 * If we do not estimate RTO correctly without them,
997 * all the algo is pure shit and should be replaced
998 * with correct one. It is exactly, which we pretend to do.
999 */
1000
1001 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
1002 * guarantees that rto is higher.
1003 */
1004 tcp_bound_rto(sk);
1005}
1006
1007__u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
1008{
1009 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
1010
1011 if (!cwnd)
1012 cwnd = TCP_INIT_CWND;
1013 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
1014}
1015
1016struct tcp_sacktag_state {
1017 /* Timestamps for earliest and latest never-retransmitted segment
1018 * that was SACKed. RTO needs the earliest RTT to stay conservative,
1019 * but congestion control should still get an accurate delay signal.
1020 */
1021 u64 first_sackt;
1022 u64 last_sackt;
1023 u32 reord;
1024 u32 sack_delivered;
1025 int flag;
1026 unsigned int mss_now;
1027 struct rate_sample *rate;
1028};
1029
1030/* Take a notice that peer is sending D-SACKs. Skip update of data delivery
1031 * and spurious retransmission information if this DSACK is unlikely caused by
1032 * sender's action:
1033 * - DSACKed sequence range is larger than maximum receiver's window.
1034 * - Total no. of DSACKed segments exceed the total no. of retransmitted segs.
1035 */
1036static u32 tcp_dsack_seen(struct tcp_sock *tp, u32 start_seq,
1037 u32 end_seq, struct tcp_sacktag_state *state)
1038{
1039 u32 seq_len, dup_segs = 1;
1040
1041 if (!before(start_seq, end_seq))
1042 return 0;
1043
1044 seq_len = end_seq - start_seq;
1045 /* Dubious DSACK: DSACKed range greater than maximum advertised rwnd */
1046 if (seq_len > tp->max_window)
1047 return 0;
1048 if (seq_len > tp->mss_cache)
1049 dup_segs = DIV_ROUND_UP(seq_len, tp->mss_cache);
1050 else if (tp->tlp_high_seq && tp->tlp_high_seq == end_seq)
1051 state->flag |= FLAG_DSACK_TLP;
1052
1053 tp->dsack_dups += dup_segs;
1054 /* Skip the DSACK if dup segs weren't retransmitted by sender */
1055 if (tp->dsack_dups > tp->total_retrans)
1056 return 0;
1057
1058 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
1059 /* We increase the RACK ordering window in rounds where we receive
1060 * DSACKs that may have been due to reordering causing RACK to trigger
1061 * a spurious fast recovery. Thus RACK ignores DSACKs that happen
1062 * without having seen reordering, or that match TLP probes (TLP
1063 * is timer-driven, not triggered by RACK).
1064 */
1065 if (tp->reord_seen && !(state->flag & FLAG_DSACK_TLP))
1066 tp->rack.dsack_seen = 1;
1067
1068 state->flag |= FLAG_DSACKING_ACK;
1069 /* A spurious retransmission is delivered */
1070 state->sack_delivered += dup_segs;
1071
1072 return dup_segs;
1073}
1074
1075/* It's reordering when higher sequence was delivered (i.e. sacked) before
1076 * some lower never-retransmitted sequence ("low_seq"). The maximum reordering
1077 * distance is approximated in full-mss packet distance ("reordering").
1078 */
1079static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq,
1080 const int ts)
1081{
1082 struct tcp_sock *tp = tcp_sk(sk);
1083 const u32 mss = tp->mss_cache;
1084 u32 fack, metric;
1085
1086 fack = tcp_highest_sack_seq(tp);
1087 if (!before(low_seq, fack))
1088 return;
1089
1090 metric = fack - low_seq;
1091 if ((metric > tp->reordering * mss) && mss) {
1092#if FASTRETRANS_DEBUG > 1
1093 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
1094 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
1095 tp->reordering,
1096 0,
1097 tp->sacked_out,
1098 tp->undo_marker ? tp->undo_retrans : 0);
1099#endif
1100 tp->reordering = min_t(u32, (metric + mss - 1) / mss,
1101 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering));
1102 }
1103
1104 /* This exciting event is worth to be remembered. 8) */
1105 tp->reord_seen++;
1106 NET_INC_STATS(sock_net(sk),
1107 ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER);
1108}
1109
1110 /* This must be called before lost_out or retrans_out are updated
1111 * on a new loss, because we want to know if all skbs previously
1112 * known to be lost have already been retransmitted, indicating
1113 * that this newly lost skb is our next skb to retransmit.
1114 */
1115static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
1116{
1117 if ((!tp->retransmit_skb_hint && tp->retrans_out >= tp->lost_out) ||
1118 (tp->retransmit_skb_hint &&
1119 before(TCP_SKB_CB(skb)->seq,
1120 TCP_SKB_CB(tp->retransmit_skb_hint)->seq)))
1121 tp->retransmit_skb_hint = skb;
1122}
1123
1124/* Sum the number of packets on the wire we have marked as lost, and
1125 * notify the congestion control module that the given skb was marked lost.
1126 */
1127static void tcp_notify_skb_loss_event(struct tcp_sock *tp, const struct sk_buff *skb)
1128{
1129 tp->lost += tcp_skb_pcount(skb);
1130}
1131
1132void tcp_mark_skb_lost(struct sock *sk, struct sk_buff *skb)
1133{
1134 __u8 sacked = TCP_SKB_CB(skb)->sacked;
1135 struct tcp_sock *tp = tcp_sk(sk);
1136
1137 if (sacked & TCPCB_SACKED_ACKED)
1138 return;
1139
1140 tcp_verify_retransmit_hint(tp, skb);
1141 if (sacked & TCPCB_LOST) {
1142 if (sacked & TCPCB_SACKED_RETRANS) {
1143 /* Account for retransmits that are lost again */
1144 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1145 tp->retrans_out -= tcp_skb_pcount(skb);
1146 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT,
1147 tcp_skb_pcount(skb));
1148 tcp_notify_skb_loss_event(tp, skb);
1149 }
1150 } else {
1151 tp->lost_out += tcp_skb_pcount(skb);
1152 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1153 tcp_notify_skb_loss_event(tp, skb);
1154 }
1155}
1156
1157/* Updates the delivered and delivered_ce counts */
1158static void tcp_count_delivered(struct tcp_sock *tp, u32 delivered,
1159 bool ece_ack)
1160{
1161 tp->delivered += delivered;
1162 if (ece_ack)
1163 tp->delivered_ce += delivered;
1164}
1165
1166/* This procedure tags the retransmission queue when SACKs arrive.
1167 *
1168 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1169 * Packets in queue with these bits set are counted in variables
1170 * sacked_out, retrans_out and lost_out, correspondingly.
1171 *
1172 * Valid combinations are:
1173 * Tag InFlight Description
1174 * 0 1 - orig segment is in flight.
1175 * S 0 - nothing flies, orig reached receiver.
1176 * L 0 - nothing flies, orig lost by net.
1177 * R 2 - both orig and retransmit are in flight.
1178 * L|R 1 - orig is lost, retransmit is in flight.
1179 * S|R 1 - orig reached receiver, retrans is still in flight.
1180 * (L|S|R is logically valid, it could occur when L|R is sacked,
1181 * but it is equivalent to plain S and code short-circuits it to S.
1182 * L|S is logically invalid, it would mean -1 packet in flight 8))
1183 *
1184 * These 6 states form finite state machine, controlled by the following events:
1185 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1186 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1187 * 3. Loss detection event of two flavors:
1188 * A. Scoreboard estimator decided the packet is lost.
1189 * A'. Reno "three dupacks" marks head of queue lost.
1190 * B. SACK arrives sacking SND.NXT at the moment, when the
1191 * segment was retransmitted.
1192 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1193 *
1194 * It is pleasant to note, that state diagram turns out to be commutative,
1195 * so that we are allowed not to be bothered by order of our actions,
1196 * when multiple events arrive simultaneously. (see the function below).
1197 *
1198 * Reordering detection.
1199 * --------------------
1200 * Reordering metric is maximal distance, which a packet can be displaced
1201 * in packet stream. With SACKs we can estimate it:
1202 *
1203 * 1. SACK fills old hole and the corresponding segment was not
1204 * ever retransmitted -> reordering. Alas, we cannot use it
1205 * when segment was retransmitted.
1206 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1207 * for retransmitted and already SACKed segment -> reordering..
1208 * Both of these heuristics are not used in Loss state, when we cannot
1209 * account for retransmits accurately.
1210 *
1211 * SACK block validation.
1212 * ----------------------
1213 *
1214 * SACK block range validation checks that the received SACK block fits to
1215 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1216 * Note that SND.UNA is not included to the range though being valid because
1217 * it means that the receiver is rather inconsistent with itself reporting
1218 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1219 * perfectly valid, however, in light of RFC2018 which explicitly states
1220 * that "SACK block MUST reflect the newest segment. Even if the newest
1221 * segment is going to be discarded ...", not that it looks very clever
1222 * in case of head skb. Due to potentional receiver driven attacks, we
1223 * choose to avoid immediate execution of a walk in write queue due to
1224 * reneging and defer head skb's loss recovery to standard loss recovery
1225 * procedure that will eventually trigger (nothing forbids us doing this).
1226 *
1227 * Implements also blockage to start_seq wrap-around. Problem lies in the
1228 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1229 * there's no guarantee that it will be before snd_nxt (n). The problem
1230 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1231 * wrap (s_w):
1232 *
1233 * <- outs wnd -> <- wrapzone ->
1234 * u e n u_w e_w s n_w
1235 * | | | | | | |
1236 * |<------------+------+----- TCP seqno space --------------+---------->|
1237 * ...-- <2^31 ->| |<--------...
1238 * ...---- >2^31 ------>| |<--------...
1239 *
1240 * Current code wouldn't be vulnerable but it's better still to discard such
1241 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1242 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1243 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1244 * equal to the ideal case (infinite seqno space without wrap caused issues).
1245 *
1246 * With D-SACK the lower bound is extended to cover sequence space below
1247 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1248 * again, D-SACK block must not to go across snd_una (for the same reason as
1249 * for the normal SACK blocks, explained above). But there all simplicity
1250 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1251 * fully below undo_marker they do not affect behavior in anyway and can
1252 * therefore be safely ignored. In rare cases (which are more or less
1253 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1254 * fragmentation and packet reordering past skb's retransmission. To consider
1255 * them correctly, the acceptable range must be extended even more though
1256 * the exact amount is rather hard to quantify. However, tp->max_window can
1257 * be used as an exaggerated estimate.
1258 */
1259static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1260 u32 start_seq, u32 end_seq)
1261{
1262 /* Too far in future, or reversed (interpretation is ambiguous) */
1263 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1264 return false;
1265
1266 /* Nasty start_seq wrap-around check (see comments above) */
1267 if (!before(start_seq, tp->snd_nxt))
1268 return false;
1269
1270 /* In outstanding window? ...This is valid exit for D-SACKs too.
1271 * start_seq == snd_una is non-sensical (see comments above)
1272 */
1273 if (after(start_seq, tp->snd_una))
1274 return true;
1275
1276 if (!is_dsack || !tp->undo_marker)
1277 return false;
1278
1279 /* ...Then it's D-SACK, and must reside below snd_una completely */
1280 if (after(end_seq, tp->snd_una))
1281 return false;
1282
1283 if (!before(start_seq, tp->undo_marker))
1284 return true;
1285
1286 /* Too old */
1287 if (!after(end_seq, tp->undo_marker))
1288 return false;
1289
1290 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1291 * start_seq < undo_marker and end_seq >= undo_marker.
1292 */
1293 return !before(start_seq, end_seq - tp->max_window);
1294}
1295
1296static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1297 struct tcp_sack_block_wire *sp, int num_sacks,
1298 u32 prior_snd_una, struct tcp_sacktag_state *state)
1299{
1300 struct tcp_sock *tp = tcp_sk(sk);
1301 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1302 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1303 u32 dup_segs;
1304
1305 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1306 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1307 } else if (num_sacks > 1) {
1308 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1309 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1310
1311 if (after(end_seq_0, end_seq_1) || before(start_seq_0, start_seq_1))
1312 return false;
1313 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKOFORECV);
1314 } else {
1315 return false;
1316 }
1317
1318 dup_segs = tcp_dsack_seen(tp, start_seq_0, end_seq_0, state);
1319 if (!dup_segs) { /* Skip dubious DSACK */
1320 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKIGNOREDDUBIOUS);
1321 return false;
1322 }
1323
1324 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECVSEGS, dup_segs);
1325
1326 /* D-SACK for already forgotten data... Do dumb counting. */
1327 if (tp->undo_marker && tp->undo_retrans > 0 &&
1328 !after(end_seq_0, prior_snd_una) &&
1329 after(end_seq_0, tp->undo_marker))
1330 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - dup_segs);
1331
1332 return true;
1333}
1334
1335/* Check if skb is fully within the SACK block. In presence of GSO skbs,
1336 * the incoming SACK may not exactly match but we can find smaller MSS
1337 * aligned portion of it that matches. Therefore we might need to fragment
1338 * which may fail and creates some hassle (caller must handle error case
1339 * returns).
1340 *
1341 * FIXME: this could be merged to shift decision code
1342 */
1343static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1344 u32 start_seq, u32 end_seq)
1345{
1346 int err;
1347 bool in_sack;
1348 unsigned int pkt_len;
1349 unsigned int mss;
1350
1351 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1352 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1353
1354 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1355 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1356 mss = tcp_skb_mss(skb);
1357 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1358
1359 if (!in_sack) {
1360 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1361 if (pkt_len < mss)
1362 pkt_len = mss;
1363 } else {
1364 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1365 if (pkt_len < mss)
1366 return -EINVAL;
1367 }
1368
1369 /* Round if necessary so that SACKs cover only full MSSes
1370 * and/or the remaining small portion (if present)
1371 */
1372 if (pkt_len > mss) {
1373 unsigned int new_len = (pkt_len / mss) * mss;
1374 if (!in_sack && new_len < pkt_len)
1375 new_len += mss;
1376 pkt_len = new_len;
1377 }
1378
1379 if (pkt_len >= skb->len && !in_sack)
1380 return 0;
1381
1382 err = tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
1383 pkt_len, mss, GFP_ATOMIC);
1384 if (err < 0)
1385 return err;
1386 }
1387
1388 return in_sack;
1389}
1390
1391/* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1392static u8 tcp_sacktag_one(struct sock *sk,
1393 struct tcp_sacktag_state *state, u8 sacked,
1394 u32 start_seq, u32 end_seq,
1395 int dup_sack, int pcount,
1396 u64 xmit_time)
1397{
1398 struct tcp_sock *tp = tcp_sk(sk);
1399
1400 /* Account D-SACK for retransmitted packet. */
1401 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1402 if (tp->undo_marker && tp->undo_retrans > 0 &&
1403 after(end_seq, tp->undo_marker))
1404 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - pcount);
1405 if ((sacked & TCPCB_SACKED_ACKED) &&
1406 before(start_seq, state->reord))
1407 state->reord = start_seq;
1408 }
1409
1410 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1411 if (!after(end_seq, tp->snd_una))
1412 return sacked;
1413
1414 if (!(sacked & TCPCB_SACKED_ACKED)) {
1415 tcp_rack_advance(tp, sacked, end_seq, xmit_time);
1416
1417 if (sacked & TCPCB_SACKED_RETRANS) {
1418 /* If the segment is not tagged as lost,
1419 * we do not clear RETRANS, believing
1420 * that retransmission is still in flight.
1421 */
1422 if (sacked & TCPCB_LOST) {
1423 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1424 tp->lost_out -= pcount;
1425 tp->retrans_out -= pcount;
1426 }
1427 } else {
1428 if (!(sacked & TCPCB_RETRANS)) {
1429 /* New sack for not retransmitted frame,
1430 * which was in hole. It is reordering.
1431 */
1432 if (before(start_seq,
1433 tcp_highest_sack_seq(tp)) &&
1434 before(start_seq, state->reord))
1435 state->reord = start_seq;
1436
1437 if (!after(end_seq, tp->high_seq))
1438 state->flag |= FLAG_ORIG_SACK_ACKED;
1439 if (state->first_sackt == 0)
1440 state->first_sackt = xmit_time;
1441 state->last_sackt = xmit_time;
1442 }
1443
1444 if (sacked & TCPCB_LOST) {
1445 sacked &= ~TCPCB_LOST;
1446 tp->lost_out -= pcount;
1447 }
1448 }
1449
1450 sacked |= TCPCB_SACKED_ACKED;
1451 state->flag |= FLAG_DATA_SACKED;
1452 tp->sacked_out += pcount;
1453 /* Out-of-order packets delivered */
1454 state->sack_delivered += pcount;
1455
1456 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1457 if (tp->lost_skb_hint &&
1458 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1459 tp->lost_cnt_hint += pcount;
1460 }
1461
1462 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1463 * frames and clear it. undo_retrans is decreased above, L|R frames
1464 * are accounted above as well.
1465 */
1466 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1467 sacked &= ~TCPCB_SACKED_RETRANS;
1468 tp->retrans_out -= pcount;
1469 }
1470
1471 return sacked;
1472}
1473
1474/* Shift newly-SACKed bytes from this skb to the immediately previous
1475 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1476 */
1477static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev,
1478 struct sk_buff *skb,
1479 struct tcp_sacktag_state *state,
1480 unsigned int pcount, int shifted, int mss,
1481 bool dup_sack)
1482{
1483 struct tcp_sock *tp = tcp_sk(sk);
1484 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1485 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1486
1487 BUG_ON(!pcount);
1488
1489 /* Adjust counters and hints for the newly sacked sequence
1490 * range but discard the return value since prev is already
1491 * marked. We must tag the range first because the seq
1492 * advancement below implicitly advances
1493 * tcp_highest_sack_seq() when skb is highest_sack.
1494 */
1495 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1496 start_seq, end_seq, dup_sack, pcount,
1497 tcp_skb_timestamp_us(skb));
1498 tcp_rate_skb_delivered(sk, skb, state->rate);
1499
1500 if (skb == tp->lost_skb_hint)
1501 tp->lost_cnt_hint += pcount;
1502
1503 TCP_SKB_CB(prev)->end_seq += shifted;
1504 TCP_SKB_CB(skb)->seq += shifted;
1505
1506 tcp_skb_pcount_add(prev, pcount);
1507 WARN_ON_ONCE(tcp_skb_pcount(skb) < pcount);
1508 tcp_skb_pcount_add(skb, -pcount);
1509
1510 /* When we're adding to gso_segs == 1, gso_size will be zero,
1511 * in theory this shouldn't be necessary but as long as DSACK
1512 * code can come after this skb later on it's better to keep
1513 * setting gso_size to something.
1514 */
1515 if (!TCP_SKB_CB(prev)->tcp_gso_size)
1516 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1517
1518 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1519 if (tcp_skb_pcount(skb) <= 1)
1520 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1521
1522 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1523 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1524
1525 if (skb->len > 0) {
1526 BUG_ON(!tcp_skb_pcount(skb));
1527 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1528 return false;
1529 }
1530
1531 /* Whole SKB was eaten :-) */
1532
1533 if (skb == tp->retransmit_skb_hint)
1534 tp->retransmit_skb_hint = prev;
1535 if (skb == tp->lost_skb_hint) {
1536 tp->lost_skb_hint = prev;
1537 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1538 }
1539
1540 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1541 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor;
1542 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1543 TCP_SKB_CB(prev)->end_seq++;
1544
1545 if (skb == tcp_highest_sack(sk))
1546 tcp_advance_highest_sack(sk, skb);
1547
1548 tcp_skb_collapse_tstamp(prev, skb);
1549 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp))
1550 TCP_SKB_CB(prev)->tx.delivered_mstamp = 0;
1551
1552 tcp_rtx_queue_unlink_and_free(skb, sk);
1553
1554 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED);
1555
1556 return true;
1557}
1558
1559/* I wish gso_size would have a bit more sane initialization than
1560 * something-or-zero which complicates things
1561 */
1562static int tcp_skb_seglen(const struct sk_buff *skb)
1563{
1564 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1565}
1566
1567/* Shifting pages past head area doesn't work */
1568static int skb_can_shift(const struct sk_buff *skb)
1569{
1570 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1571}
1572
1573int tcp_skb_shift(struct sk_buff *to, struct sk_buff *from,
1574 int pcount, int shiftlen)
1575{
1576 /* TCP min gso_size is 8 bytes (TCP_MIN_GSO_SIZE)
1577 * Since TCP_SKB_CB(skb)->tcp_gso_segs is 16 bits, we need
1578 * to make sure not storing more than 65535 * 8 bytes per skb,
1579 * even if current MSS is bigger.
1580 */
1581 if (unlikely(to->len + shiftlen >= 65535 * TCP_MIN_GSO_SIZE))
1582 return 0;
1583 if (unlikely(tcp_skb_pcount(to) + pcount > 65535))
1584 return 0;
1585 return skb_shift(to, from, shiftlen);
1586}
1587
1588/* Try collapsing SACK blocks spanning across multiple skbs to a single
1589 * skb.
1590 */
1591static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1592 struct tcp_sacktag_state *state,
1593 u32 start_seq, u32 end_seq,
1594 bool dup_sack)
1595{
1596 struct tcp_sock *tp = tcp_sk(sk);
1597 struct sk_buff *prev;
1598 int mss;
1599 int pcount = 0;
1600 int len;
1601 int in_sack;
1602
1603 /* Normally R but no L won't result in plain S */
1604 if (!dup_sack &&
1605 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1606 goto fallback;
1607 if (!skb_can_shift(skb))
1608 goto fallback;
1609 /* This frame is about to be dropped (was ACKed). */
1610 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1611 goto fallback;
1612
1613 /* Can only happen with delayed DSACK + discard craziness */
1614 prev = skb_rb_prev(skb);
1615 if (!prev)
1616 goto fallback;
1617
1618 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1619 goto fallback;
1620
1621 if (!tcp_skb_can_collapse(prev, skb))
1622 goto fallback;
1623
1624 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1625 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1626
1627 if (in_sack) {
1628 len = skb->len;
1629 pcount = tcp_skb_pcount(skb);
1630 mss = tcp_skb_seglen(skb);
1631
1632 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1633 * drop this restriction as unnecessary
1634 */
1635 if (mss != tcp_skb_seglen(prev))
1636 goto fallback;
1637 } else {
1638 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1639 goto noop;
1640 /* CHECKME: This is non-MSS split case only?, this will
1641 * cause skipped skbs due to advancing loop btw, original
1642 * has that feature too
1643 */
1644 if (tcp_skb_pcount(skb) <= 1)
1645 goto noop;
1646
1647 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1648 if (!in_sack) {
1649 /* TODO: head merge to next could be attempted here
1650 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1651 * though it might not be worth of the additional hassle
1652 *
1653 * ...we can probably just fallback to what was done
1654 * previously. We could try merging non-SACKed ones
1655 * as well but it probably isn't going to buy off
1656 * because later SACKs might again split them, and
1657 * it would make skb timestamp tracking considerably
1658 * harder problem.
1659 */
1660 goto fallback;
1661 }
1662
1663 len = end_seq - TCP_SKB_CB(skb)->seq;
1664 BUG_ON(len < 0);
1665 BUG_ON(len > skb->len);
1666
1667 /* MSS boundaries should be honoured or else pcount will
1668 * severely break even though it makes things bit trickier.
1669 * Optimize common case to avoid most of the divides
1670 */
1671 mss = tcp_skb_mss(skb);
1672
1673 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1674 * drop this restriction as unnecessary
1675 */
1676 if (mss != tcp_skb_seglen(prev))
1677 goto fallback;
1678
1679 if (len == mss) {
1680 pcount = 1;
1681 } else if (len < mss) {
1682 goto noop;
1683 } else {
1684 pcount = len / mss;
1685 len = pcount * mss;
1686 }
1687 }
1688
1689 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1690 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1691 goto fallback;
1692
1693 if (!tcp_skb_shift(prev, skb, pcount, len))
1694 goto fallback;
1695 if (!tcp_shifted_skb(sk, prev, skb, state, pcount, len, mss, dup_sack))
1696 goto out;
1697
1698 /* Hole filled allows collapsing with the next as well, this is very
1699 * useful when hole on every nth skb pattern happens
1700 */
1701 skb = skb_rb_next(prev);
1702 if (!skb)
1703 goto out;
1704
1705 if (!skb_can_shift(skb) ||
1706 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1707 (mss != tcp_skb_seglen(skb)))
1708 goto out;
1709
1710 if (!tcp_skb_can_collapse(prev, skb))
1711 goto out;
1712 len = skb->len;
1713 pcount = tcp_skb_pcount(skb);
1714 if (tcp_skb_shift(prev, skb, pcount, len))
1715 tcp_shifted_skb(sk, prev, skb, state, pcount,
1716 len, mss, 0);
1717
1718out:
1719 return prev;
1720
1721noop:
1722 return skb;
1723
1724fallback:
1725 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1726 return NULL;
1727}
1728
1729static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1730 struct tcp_sack_block *next_dup,
1731 struct tcp_sacktag_state *state,
1732 u32 start_seq, u32 end_seq,
1733 bool dup_sack_in)
1734{
1735 struct tcp_sock *tp = tcp_sk(sk);
1736 struct sk_buff *tmp;
1737
1738 skb_rbtree_walk_from(skb) {
1739 int in_sack = 0;
1740 bool dup_sack = dup_sack_in;
1741
1742 /* queue is in-order => we can short-circuit the walk early */
1743 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1744 break;
1745
1746 if (next_dup &&
1747 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1748 in_sack = tcp_match_skb_to_sack(sk, skb,
1749 next_dup->start_seq,
1750 next_dup->end_seq);
1751 if (in_sack > 0)
1752 dup_sack = true;
1753 }
1754
1755 /* skb reference here is a bit tricky to get right, since
1756 * shifting can eat and free both this skb and the next,
1757 * so not even _safe variant of the loop is enough.
1758 */
1759 if (in_sack <= 0) {
1760 tmp = tcp_shift_skb_data(sk, skb, state,
1761 start_seq, end_seq, dup_sack);
1762 if (tmp) {
1763 if (tmp != skb) {
1764 skb = tmp;
1765 continue;
1766 }
1767
1768 in_sack = 0;
1769 } else {
1770 in_sack = tcp_match_skb_to_sack(sk, skb,
1771 start_seq,
1772 end_seq);
1773 }
1774 }
1775
1776 if (unlikely(in_sack < 0))
1777 break;
1778
1779 if (in_sack) {
1780 TCP_SKB_CB(skb)->sacked =
1781 tcp_sacktag_one(sk,
1782 state,
1783 TCP_SKB_CB(skb)->sacked,
1784 TCP_SKB_CB(skb)->seq,
1785 TCP_SKB_CB(skb)->end_seq,
1786 dup_sack,
1787 tcp_skb_pcount(skb),
1788 tcp_skb_timestamp_us(skb));
1789 tcp_rate_skb_delivered(sk, skb, state->rate);
1790 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1791 list_del_init(&skb->tcp_tsorted_anchor);
1792
1793 if (!before(TCP_SKB_CB(skb)->seq,
1794 tcp_highest_sack_seq(tp)))
1795 tcp_advance_highest_sack(sk, skb);
1796 }
1797 }
1798 return skb;
1799}
1800
1801static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk, u32 seq)
1802{
1803 struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node;
1804 struct sk_buff *skb;
1805
1806 while (*p) {
1807 parent = *p;
1808 skb = rb_to_skb(parent);
1809 if (before(seq, TCP_SKB_CB(skb)->seq)) {
1810 p = &parent->rb_left;
1811 continue;
1812 }
1813 if (!before(seq, TCP_SKB_CB(skb)->end_seq)) {
1814 p = &parent->rb_right;
1815 continue;
1816 }
1817 return skb;
1818 }
1819 return NULL;
1820}
1821
1822static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1823 u32 skip_to_seq)
1824{
1825 if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq))
1826 return skb;
1827
1828 return tcp_sacktag_bsearch(sk, skip_to_seq);
1829}
1830
1831static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1832 struct sock *sk,
1833 struct tcp_sack_block *next_dup,
1834 struct tcp_sacktag_state *state,
1835 u32 skip_to_seq)
1836{
1837 if (!next_dup)
1838 return skb;
1839
1840 if (before(next_dup->start_seq, skip_to_seq)) {
1841 skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq);
1842 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1843 next_dup->start_seq, next_dup->end_seq,
1844 1);
1845 }
1846
1847 return skb;
1848}
1849
1850static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1851{
1852 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1853}
1854
1855static int
1856tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1857 u32 prior_snd_una, struct tcp_sacktag_state *state)
1858{
1859 struct tcp_sock *tp = tcp_sk(sk);
1860 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1861 TCP_SKB_CB(ack_skb)->sacked);
1862 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1863 struct tcp_sack_block sp[TCP_NUM_SACKS];
1864 struct tcp_sack_block *cache;
1865 struct sk_buff *skb;
1866 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1867 int used_sacks;
1868 bool found_dup_sack = false;
1869 int i, j;
1870 int first_sack_index;
1871
1872 state->flag = 0;
1873 state->reord = tp->snd_nxt;
1874
1875 if (!tp->sacked_out)
1876 tcp_highest_sack_reset(sk);
1877
1878 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1879 num_sacks, prior_snd_una, state);
1880
1881 /* Eliminate too old ACKs, but take into
1882 * account more or less fresh ones, they can
1883 * contain valid SACK info.
1884 */
1885 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1886 return 0;
1887
1888 if (!tp->packets_out)
1889 goto out;
1890
1891 used_sacks = 0;
1892 first_sack_index = 0;
1893 for (i = 0; i < num_sacks; i++) {
1894 bool dup_sack = !i && found_dup_sack;
1895
1896 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1897 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1898
1899 if (!tcp_is_sackblock_valid(tp, dup_sack,
1900 sp[used_sacks].start_seq,
1901 sp[used_sacks].end_seq)) {
1902 int mib_idx;
1903
1904 if (dup_sack) {
1905 if (!tp->undo_marker)
1906 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1907 else
1908 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1909 } else {
1910 /* Don't count olds caused by ACK reordering */
1911 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1912 !after(sp[used_sacks].end_seq, tp->snd_una))
1913 continue;
1914 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1915 }
1916
1917 NET_INC_STATS(sock_net(sk), mib_idx);
1918 if (i == 0)
1919 first_sack_index = -1;
1920 continue;
1921 }
1922
1923 /* Ignore very old stuff early */
1924 if (!after(sp[used_sacks].end_seq, prior_snd_una)) {
1925 if (i == 0)
1926 first_sack_index = -1;
1927 continue;
1928 }
1929
1930 used_sacks++;
1931 }
1932
1933 /* order SACK blocks to allow in order walk of the retrans queue */
1934 for (i = used_sacks - 1; i > 0; i--) {
1935 for (j = 0; j < i; j++) {
1936 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1937 swap(sp[j], sp[j + 1]);
1938
1939 /* Track where the first SACK block goes to */
1940 if (j == first_sack_index)
1941 first_sack_index = j + 1;
1942 }
1943 }
1944 }
1945
1946 state->mss_now = tcp_current_mss(sk);
1947 skb = NULL;
1948 i = 0;
1949
1950 if (!tp->sacked_out) {
1951 /* It's already past, so skip checking against it */
1952 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1953 } else {
1954 cache = tp->recv_sack_cache;
1955 /* Skip empty blocks in at head of the cache */
1956 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1957 !cache->end_seq)
1958 cache++;
1959 }
1960
1961 while (i < used_sacks) {
1962 u32 start_seq = sp[i].start_seq;
1963 u32 end_seq = sp[i].end_seq;
1964 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1965 struct tcp_sack_block *next_dup = NULL;
1966
1967 if (found_dup_sack && ((i + 1) == first_sack_index))
1968 next_dup = &sp[i + 1];
1969
1970 /* Skip too early cached blocks */
1971 while (tcp_sack_cache_ok(tp, cache) &&
1972 !before(start_seq, cache->end_seq))
1973 cache++;
1974
1975 /* Can skip some work by looking recv_sack_cache? */
1976 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1977 after(end_seq, cache->start_seq)) {
1978
1979 /* Head todo? */
1980 if (before(start_seq, cache->start_seq)) {
1981 skb = tcp_sacktag_skip(skb, sk, start_seq);
1982 skb = tcp_sacktag_walk(skb, sk, next_dup,
1983 state,
1984 start_seq,
1985 cache->start_seq,
1986 dup_sack);
1987 }
1988
1989 /* Rest of the block already fully processed? */
1990 if (!after(end_seq, cache->end_seq))
1991 goto advance_sp;
1992
1993 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1994 state,
1995 cache->end_seq);
1996
1997 /* ...tail remains todo... */
1998 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1999 /* ...but better entrypoint exists! */
2000 skb = tcp_highest_sack(sk);
2001 if (!skb)
2002 break;
2003 cache++;
2004 goto walk;
2005 }
2006
2007 skb = tcp_sacktag_skip(skb, sk, cache->end_seq);
2008 /* Check overlap against next cached too (past this one already) */
2009 cache++;
2010 continue;
2011 }
2012
2013 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
2014 skb = tcp_highest_sack(sk);
2015 if (!skb)
2016 break;
2017 }
2018 skb = tcp_sacktag_skip(skb, sk, start_seq);
2019
2020walk:
2021 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
2022 start_seq, end_seq, dup_sack);
2023
2024advance_sp:
2025 i++;
2026 }
2027
2028 /* Clear the head of the cache sack blocks so we can skip it next time */
2029 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
2030 tp->recv_sack_cache[i].start_seq = 0;
2031 tp->recv_sack_cache[i].end_seq = 0;
2032 }
2033 for (j = 0; j < used_sacks; j++)
2034 tp->recv_sack_cache[i++] = sp[j];
2035
2036 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss || tp->undo_marker)
2037 tcp_check_sack_reordering(sk, state->reord, 0);
2038
2039 tcp_verify_left_out(tp);
2040out:
2041
2042#if FASTRETRANS_DEBUG > 0
2043 WARN_ON((int)tp->sacked_out < 0);
2044 WARN_ON((int)tp->lost_out < 0);
2045 WARN_ON((int)tp->retrans_out < 0);
2046 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
2047#endif
2048 return state->flag;
2049}
2050
2051/* Limits sacked_out so that sum with lost_out isn't ever larger than
2052 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
2053 */
2054static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
2055{
2056 u32 holes;
2057
2058 holes = max(tp->lost_out, 1U);
2059 holes = min(holes, tp->packets_out);
2060
2061 if ((tp->sacked_out + holes) > tp->packets_out) {
2062 tp->sacked_out = tp->packets_out - holes;
2063 return true;
2064 }
2065 return false;
2066}
2067
2068/* If we receive more dupacks than we expected counting segments
2069 * in assumption of absent reordering, interpret this as reordering.
2070 * The only another reason could be bug in receiver TCP.
2071 */
2072static void tcp_check_reno_reordering(struct sock *sk, const int addend)
2073{
2074 struct tcp_sock *tp = tcp_sk(sk);
2075
2076 if (!tcp_limit_reno_sacked(tp))
2077 return;
2078
2079 tp->reordering = min_t(u32, tp->packets_out + addend,
2080 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering));
2081 tp->reord_seen++;
2082 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER);
2083}
2084
2085/* Emulate SACKs for SACKless connection: account for a new dupack. */
2086
2087static void tcp_add_reno_sack(struct sock *sk, int num_dupack, bool ece_ack)
2088{
2089 if (num_dupack) {
2090 struct tcp_sock *tp = tcp_sk(sk);
2091 u32 prior_sacked = tp->sacked_out;
2092 s32 delivered;
2093
2094 tp->sacked_out += num_dupack;
2095 tcp_check_reno_reordering(sk, 0);
2096 delivered = tp->sacked_out - prior_sacked;
2097 if (delivered > 0)
2098 tcp_count_delivered(tp, delivered, ece_ack);
2099 tcp_verify_left_out(tp);
2100 }
2101}
2102
2103/* Account for ACK, ACKing some data in Reno Recovery phase. */
2104
2105static void tcp_remove_reno_sacks(struct sock *sk, int acked, bool ece_ack)
2106{
2107 struct tcp_sock *tp = tcp_sk(sk);
2108
2109 if (acked > 0) {
2110 /* One ACK acked hole. The rest eat duplicate ACKs. */
2111 tcp_count_delivered(tp, max_t(int, acked - tp->sacked_out, 1),
2112 ece_ack);
2113 if (acked - 1 >= tp->sacked_out)
2114 tp->sacked_out = 0;
2115 else
2116 tp->sacked_out -= acked - 1;
2117 }
2118 tcp_check_reno_reordering(sk, acked);
2119 tcp_verify_left_out(tp);
2120}
2121
2122static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
2123{
2124 tp->sacked_out = 0;
2125}
2126
2127void tcp_clear_retrans(struct tcp_sock *tp)
2128{
2129 tp->retrans_out = 0;
2130 tp->lost_out = 0;
2131 tp->undo_marker = 0;
2132 tp->undo_retrans = -1;
2133 tp->sacked_out = 0;
2134 tp->rto_stamp = 0;
2135 tp->total_rto = 0;
2136 tp->total_rto_recoveries = 0;
2137 tp->total_rto_time = 0;
2138}
2139
2140static inline void tcp_init_undo(struct tcp_sock *tp)
2141{
2142 tp->undo_marker = tp->snd_una;
2143
2144 /* Retransmission still in flight may cause DSACKs later. */
2145 /* First, account for regular retransmits in flight: */
2146 tp->undo_retrans = tp->retrans_out;
2147 /* Next, account for TLP retransmits in flight: */
2148 if (tp->tlp_high_seq && tp->tlp_retrans)
2149 tp->undo_retrans++;
2150 /* Finally, avoid 0, because undo_retrans==0 means "can undo now": */
2151 if (!tp->undo_retrans)
2152 tp->undo_retrans = -1;
2153}
2154
2155static bool tcp_is_rack(const struct sock *sk)
2156{
2157 return READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_recovery) &
2158 TCP_RACK_LOSS_DETECTION;
2159}
2160
2161/* If we detect SACK reneging, forget all SACK information
2162 * and reset tags completely, otherwise preserve SACKs. If receiver
2163 * dropped its ofo queue, we will know this due to reneging detection.
2164 */
2165static void tcp_timeout_mark_lost(struct sock *sk)
2166{
2167 struct tcp_sock *tp = tcp_sk(sk);
2168 struct sk_buff *skb, *head;
2169 bool is_reneg; /* is receiver reneging on SACKs? */
2170
2171 head = tcp_rtx_queue_head(sk);
2172 is_reneg = head && (TCP_SKB_CB(head)->sacked & TCPCB_SACKED_ACKED);
2173 if (is_reneg) {
2174 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
2175 tp->sacked_out = 0;
2176 /* Mark SACK reneging until we recover from this loss event. */
2177 tp->is_sack_reneg = 1;
2178 } else if (tcp_is_reno(tp)) {
2179 tcp_reset_reno_sack(tp);
2180 }
2181
2182 skb = head;
2183 skb_rbtree_walk_from(skb) {
2184 if (is_reneg)
2185 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
2186 else if (tcp_is_rack(sk) && skb != head &&
2187 tcp_rack_skb_timeout(tp, skb, 0) > 0)
2188 continue; /* Don't mark recently sent ones lost yet */
2189 tcp_mark_skb_lost(sk, skb);
2190 }
2191 tcp_verify_left_out(tp);
2192 tcp_clear_all_retrans_hints(tp);
2193}
2194
2195/* Enter Loss state. */
2196void tcp_enter_loss(struct sock *sk)
2197{
2198 const struct inet_connection_sock *icsk = inet_csk(sk);
2199 struct tcp_sock *tp = tcp_sk(sk);
2200 struct net *net = sock_net(sk);
2201 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
2202 u8 reordering;
2203
2204 tcp_timeout_mark_lost(sk);
2205
2206 /* Reduce ssthresh if it has not yet been made inside this window. */
2207 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
2208 !after(tp->high_seq, tp->snd_una) ||
2209 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2210 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2211 tp->prior_cwnd = tcp_snd_cwnd(tp);
2212 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2213 tcp_ca_event(sk, CA_EVENT_LOSS);
2214 tcp_init_undo(tp);
2215 }
2216 tcp_snd_cwnd_set(tp, tcp_packets_in_flight(tp) + 1);
2217 tp->snd_cwnd_cnt = 0;
2218 tp->snd_cwnd_stamp = tcp_jiffies32;
2219
2220 /* Timeout in disordered state after receiving substantial DUPACKs
2221 * suggests that the degree of reordering is over-estimated.
2222 */
2223 reordering = READ_ONCE(net->ipv4.sysctl_tcp_reordering);
2224 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
2225 tp->sacked_out >= reordering)
2226 tp->reordering = min_t(unsigned int, tp->reordering,
2227 reordering);
2228
2229 tcp_set_ca_state(sk, TCP_CA_Loss);
2230 tp->high_seq = tp->snd_nxt;
2231 tp->tlp_high_seq = 0;
2232 tcp_ecn_queue_cwr(tp);
2233
2234 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
2235 * loss recovery is underway except recurring timeout(s) on
2236 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
2237 */
2238 tp->frto = READ_ONCE(net->ipv4.sysctl_tcp_frto) &&
2239 (new_recovery || icsk->icsk_retransmits) &&
2240 !inet_csk(sk)->icsk_mtup.probe_size;
2241}
2242
2243/* If ACK arrived pointing to a remembered SACK, it means that our
2244 * remembered SACKs do not reflect real state of receiver i.e.
2245 * receiver _host_ is heavily congested (or buggy).
2246 *
2247 * To avoid big spurious retransmission bursts due to transient SACK
2248 * scoreboard oddities that look like reneging, we give the receiver a
2249 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2250 * restore sanity to the SACK scoreboard. If the apparent reneging
2251 * persists until this RTO then we'll clear the SACK scoreboard.
2252 */
2253static bool tcp_check_sack_reneging(struct sock *sk, int *ack_flag)
2254{
2255 if (*ack_flag & FLAG_SACK_RENEGING &&
2256 *ack_flag & FLAG_SND_UNA_ADVANCED) {
2257 struct tcp_sock *tp = tcp_sk(sk);
2258 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2259 msecs_to_jiffies(10));
2260
2261 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2262 delay, TCP_RTO_MAX);
2263 *ack_flag &= ~FLAG_SET_XMIT_TIMER;
2264 return true;
2265 }
2266 return false;
2267}
2268
2269/* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2270 * counter when SACK is enabled (without SACK, sacked_out is used for
2271 * that purpose).
2272 *
2273 * With reordering, holes may still be in flight, so RFC3517 recovery
2274 * uses pure sacked_out (total number of SACKed segments) even though
2275 * it violates the RFC that uses duplicate ACKs, often these are equal
2276 * but when e.g. out-of-window ACKs or packet duplication occurs,
2277 * they differ. Since neither occurs due to loss, TCP should really
2278 * ignore them.
2279 */
2280static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2281{
2282 return tp->sacked_out + 1;
2283}
2284
2285/* Linux NewReno/SACK/ECN state machine.
2286 * --------------------------------------
2287 *
2288 * "Open" Normal state, no dubious events, fast path.
2289 * "Disorder" In all the respects it is "Open",
2290 * but requires a bit more attention. It is entered when
2291 * we see some SACKs or dupacks. It is split of "Open"
2292 * mainly to move some processing from fast path to slow one.
2293 * "CWR" CWND was reduced due to some Congestion Notification event.
2294 * It can be ECN, ICMP source quench, local device congestion.
2295 * "Recovery" CWND was reduced, we are fast-retransmitting.
2296 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2297 *
2298 * tcp_fastretrans_alert() is entered:
2299 * - each incoming ACK, if state is not "Open"
2300 * - when arrived ACK is unusual, namely:
2301 * * SACK
2302 * * Duplicate ACK.
2303 * * ECN ECE.
2304 *
2305 * Counting packets in flight is pretty simple.
2306 *
2307 * in_flight = packets_out - left_out + retrans_out
2308 *
2309 * packets_out is SND.NXT-SND.UNA counted in packets.
2310 *
2311 * retrans_out is number of retransmitted segments.
2312 *
2313 * left_out is number of segments left network, but not ACKed yet.
2314 *
2315 * left_out = sacked_out + lost_out
2316 *
2317 * sacked_out: Packets, which arrived to receiver out of order
2318 * and hence not ACKed. With SACKs this number is simply
2319 * amount of SACKed data. Even without SACKs
2320 * it is easy to give pretty reliable estimate of this number,
2321 * counting duplicate ACKs.
2322 *
2323 * lost_out: Packets lost by network. TCP has no explicit
2324 * "loss notification" feedback from network (for now).
2325 * It means that this number can be only _guessed_.
2326 * Actually, it is the heuristics to predict lossage that
2327 * distinguishes different algorithms.
2328 *
2329 * F.e. after RTO, when all the queue is considered as lost,
2330 * lost_out = packets_out and in_flight = retrans_out.
2331 *
2332 * Essentially, we have now a few algorithms detecting
2333 * lost packets.
2334 *
2335 * If the receiver supports SACK:
2336 *
2337 * RFC6675/3517: It is the conventional algorithm. A packet is
2338 * considered lost if the number of higher sequence packets
2339 * SACKed is greater than or equal the DUPACK thoreshold
2340 * (reordering). This is implemented in tcp_mark_head_lost and
2341 * tcp_update_scoreboard.
2342 *
2343 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm
2344 * (2017-) that checks timing instead of counting DUPACKs.
2345 * Essentially a packet is considered lost if it's not S/ACKed
2346 * after RTT + reordering_window, where both metrics are
2347 * dynamically measured and adjusted. This is implemented in
2348 * tcp_rack_mark_lost.
2349 *
2350 * If the receiver does not support SACK:
2351 *
2352 * NewReno (RFC6582): in Recovery we assume that one segment
2353 * is lost (classic Reno). While we are in Recovery and
2354 * a partial ACK arrives, we assume that one more packet
2355 * is lost (NewReno). This heuristics are the same in NewReno
2356 * and SACK.
2357 *
2358 * Really tricky (and requiring careful tuning) part of algorithm
2359 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2360 * The first determines the moment _when_ we should reduce CWND and,
2361 * hence, slow down forward transmission. In fact, it determines the moment
2362 * when we decide that hole is caused by loss, rather than by a reorder.
2363 *
2364 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2365 * holes, caused by lost packets.
2366 *
2367 * And the most logically complicated part of algorithm is undo
2368 * heuristics. We detect false retransmits due to both too early
2369 * fast retransmit (reordering) and underestimated RTO, analyzing
2370 * timestamps and D-SACKs. When we detect that some segments were
2371 * retransmitted by mistake and CWND reduction was wrong, we undo
2372 * window reduction and abort recovery phase. This logic is hidden
2373 * inside several functions named tcp_try_undo_<something>.
2374 */
2375
2376/* This function decides, when we should leave Disordered state
2377 * and enter Recovery phase, reducing congestion window.
2378 *
2379 * Main question: may we further continue forward transmission
2380 * with the same cwnd?
2381 */
2382static bool tcp_time_to_recover(struct sock *sk, int flag)
2383{
2384 struct tcp_sock *tp = tcp_sk(sk);
2385
2386 /* Trick#1: The loss is proven. */
2387 if (tp->lost_out)
2388 return true;
2389
2390 /* Not-A-Trick#2 : Classic rule... */
2391 if (!tcp_is_rack(sk) && tcp_dupack_heuristics(tp) > tp->reordering)
2392 return true;
2393
2394 return false;
2395}
2396
2397/* Detect loss in event "A" above by marking head of queue up as lost.
2398 * For RFC3517 SACK, a segment is considered lost if it
2399 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2400 * the maximum SACKed segments to pass before reaching this limit.
2401 */
2402static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2403{
2404 struct tcp_sock *tp = tcp_sk(sk);
2405 struct sk_buff *skb;
2406 int cnt;
2407 /* Use SACK to deduce losses of new sequences sent during recovery */
2408 const u32 loss_high = tp->snd_nxt;
2409
2410 WARN_ON(packets > tp->packets_out);
2411 skb = tp->lost_skb_hint;
2412 if (skb) {
2413 /* Head already handled? */
2414 if (mark_head && after(TCP_SKB_CB(skb)->seq, tp->snd_una))
2415 return;
2416 cnt = tp->lost_cnt_hint;
2417 } else {
2418 skb = tcp_rtx_queue_head(sk);
2419 cnt = 0;
2420 }
2421
2422 skb_rbtree_walk_from(skb) {
2423 /* TODO: do this better */
2424 /* this is not the most efficient way to do this... */
2425 tp->lost_skb_hint = skb;
2426 tp->lost_cnt_hint = cnt;
2427
2428 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2429 break;
2430
2431 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2432 cnt += tcp_skb_pcount(skb);
2433
2434 if (cnt > packets)
2435 break;
2436
2437 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_LOST))
2438 tcp_mark_skb_lost(sk, skb);
2439
2440 if (mark_head)
2441 break;
2442 }
2443 tcp_verify_left_out(tp);
2444}
2445
2446/* Account newly detected lost packet(s) */
2447
2448static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2449{
2450 struct tcp_sock *tp = tcp_sk(sk);
2451
2452 if (tcp_is_sack(tp)) {
2453 int sacked_upto = tp->sacked_out - tp->reordering;
2454 if (sacked_upto >= 0)
2455 tcp_mark_head_lost(sk, sacked_upto, 0);
2456 else if (fast_rexmit)
2457 tcp_mark_head_lost(sk, 1, 1);
2458 }
2459}
2460
2461static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2462{
2463 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2464 before(tp->rx_opt.rcv_tsecr, when);
2465}
2466
2467/* skb is spurious retransmitted if the returned timestamp echo
2468 * reply is prior to the skb transmission time
2469 */
2470static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2471 const struct sk_buff *skb)
2472{
2473 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2474 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp_ts(tp->tcp_usec_ts, skb));
2475}
2476
2477/* Nothing was retransmitted or returned timestamp is less
2478 * than timestamp of the first retransmission.
2479 */
2480static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2481{
2482 const struct sock *sk = (const struct sock *)tp;
2483
2484 if (tp->retrans_stamp &&
2485 tcp_tsopt_ecr_before(tp, tp->retrans_stamp))
2486 return true; /* got echoed TS before first retransmission */
2487
2488 /* Check if nothing was retransmitted (retrans_stamp==0), which may
2489 * happen in fast recovery due to TSQ. But we ignore zero retrans_stamp
2490 * in TCP_SYN_SENT, since when we set FLAG_SYN_ACKED we also clear
2491 * retrans_stamp even if we had retransmitted the SYN.
2492 */
2493 if (!tp->retrans_stamp && /* no record of a retransmit/SYN? */
2494 sk->sk_state != TCP_SYN_SENT) /* not the FLAG_SYN_ACKED case? */
2495 return true; /* nothing was retransmitted */
2496
2497 return false;
2498}
2499
2500/* Undo procedures. */
2501
2502/* We can clear retrans_stamp when there are no retransmissions in the
2503 * window. It would seem that it is trivially available for us in
2504 * tp->retrans_out, however, that kind of assumptions doesn't consider
2505 * what will happen if errors occur when sending retransmission for the
2506 * second time. ...It could the that such segment has only
2507 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2508 * the head skb is enough except for some reneging corner cases that
2509 * are not worth the effort.
2510 *
2511 * Main reason for all this complexity is the fact that connection dying
2512 * time now depends on the validity of the retrans_stamp, in particular,
2513 * that successive retransmissions of a segment must not advance
2514 * retrans_stamp under any conditions.
2515 */
2516static bool tcp_any_retrans_done(const struct sock *sk)
2517{
2518 const struct tcp_sock *tp = tcp_sk(sk);
2519 struct sk_buff *skb;
2520
2521 if (tp->retrans_out)
2522 return true;
2523
2524 skb = tcp_rtx_queue_head(sk);
2525 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2526 return true;
2527
2528 return false;
2529}
2530
2531/* If loss recovery is finished and there are no retransmits out in the
2532 * network, then we clear retrans_stamp so that upon the next loss recovery
2533 * retransmits_timed_out() and timestamp-undo are using the correct value.
2534 */
2535static void tcp_retrans_stamp_cleanup(struct sock *sk)
2536{
2537 if (!tcp_any_retrans_done(sk))
2538 tcp_sk(sk)->retrans_stamp = 0;
2539}
2540
2541static void DBGUNDO(struct sock *sk, const char *msg)
2542{
2543#if FASTRETRANS_DEBUG > 1
2544 struct tcp_sock *tp = tcp_sk(sk);
2545 struct inet_sock *inet = inet_sk(sk);
2546
2547 if (sk->sk_family == AF_INET) {
2548 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2549 msg,
2550 &inet->inet_daddr, ntohs(inet->inet_dport),
2551 tcp_snd_cwnd(tp), tcp_left_out(tp),
2552 tp->snd_ssthresh, tp->prior_ssthresh,
2553 tp->packets_out);
2554 }
2555#if IS_ENABLED(CONFIG_IPV6)
2556 else if (sk->sk_family == AF_INET6) {
2557 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2558 msg,
2559 &sk->sk_v6_daddr, ntohs(inet->inet_dport),
2560 tcp_snd_cwnd(tp), tcp_left_out(tp),
2561 tp->snd_ssthresh, tp->prior_ssthresh,
2562 tp->packets_out);
2563 }
2564#endif
2565#endif
2566}
2567
2568static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2569{
2570 struct tcp_sock *tp = tcp_sk(sk);
2571
2572 if (unmark_loss) {
2573 struct sk_buff *skb;
2574
2575 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2576 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2577 }
2578 tp->lost_out = 0;
2579 tcp_clear_all_retrans_hints(tp);
2580 }
2581
2582 if (tp->prior_ssthresh) {
2583 const struct inet_connection_sock *icsk = inet_csk(sk);
2584
2585 tcp_snd_cwnd_set(tp, icsk->icsk_ca_ops->undo_cwnd(sk));
2586
2587 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2588 tp->snd_ssthresh = tp->prior_ssthresh;
2589 tcp_ecn_withdraw_cwr(tp);
2590 }
2591 }
2592 tp->snd_cwnd_stamp = tcp_jiffies32;
2593 tp->undo_marker = 0;
2594 tp->rack.advanced = 1; /* Force RACK to re-exam losses */
2595}
2596
2597static inline bool tcp_may_undo(const struct tcp_sock *tp)
2598{
2599 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2600}
2601
2602static bool tcp_is_non_sack_preventing_reopen(struct sock *sk)
2603{
2604 struct tcp_sock *tp = tcp_sk(sk);
2605
2606 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2607 /* Hold old state until something *above* high_seq
2608 * is ACKed. For Reno it is MUST to prevent false
2609 * fast retransmits (RFC2582). SACK TCP is safe. */
2610 if (!tcp_any_retrans_done(sk))
2611 tp->retrans_stamp = 0;
2612 return true;
2613 }
2614 return false;
2615}
2616
2617/* People celebrate: "We love our President!" */
2618static bool tcp_try_undo_recovery(struct sock *sk)
2619{
2620 struct tcp_sock *tp = tcp_sk(sk);
2621
2622 if (tcp_may_undo(tp)) {
2623 int mib_idx;
2624
2625 /* Happy end! We did not retransmit anything
2626 * or our original transmission succeeded.
2627 */
2628 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2629 tcp_undo_cwnd_reduction(sk, false);
2630 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2631 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2632 else
2633 mib_idx = LINUX_MIB_TCPFULLUNDO;
2634
2635 NET_INC_STATS(sock_net(sk), mib_idx);
2636 } else if (tp->rack.reo_wnd_persist) {
2637 tp->rack.reo_wnd_persist--;
2638 }
2639 if (tcp_is_non_sack_preventing_reopen(sk))
2640 return true;
2641 tcp_set_ca_state(sk, TCP_CA_Open);
2642 tp->is_sack_reneg = 0;
2643 return false;
2644}
2645
2646/* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2647static bool tcp_try_undo_dsack(struct sock *sk)
2648{
2649 struct tcp_sock *tp = tcp_sk(sk);
2650
2651 if (tp->undo_marker && !tp->undo_retrans) {
2652 tp->rack.reo_wnd_persist = min(TCP_RACK_RECOVERY_THRESH,
2653 tp->rack.reo_wnd_persist + 1);
2654 DBGUNDO(sk, "D-SACK");
2655 tcp_undo_cwnd_reduction(sk, false);
2656 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2657 return true;
2658 }
2659 return false;
2660}
2661
2662/* Undo during loss recovery after partial ACK or using F-RTO. */
2663static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2664{
2665 struct tcp_sock *tp = tcp_sk(sk);
2666
2667 if (frto_undo || tcp_may_undo(tp)) {
2668 tcp_undo_cwnd_reduction(sk, true);
2669
2670 DBGUNDO(sk, "partial loss");
2671 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2672 if (frto_undo)
2673 NET_INC_STATS(sock_net(sk),
2674 LINUX_MIB_TCPSPURIOUSRTOS);
2675 inet_csk(sk)->icsk_retransmits = 0;
2676 if (tcp_is_non_sack_preventing_reopen(sk))
2677 return true;
2678 if (frto_undo || tcp_is_sack(tp)) {
2679 tcp_set_ca_state(sk, TCP_CA_Open);
2680 tp->is_sack_reneg = 0;
2681 }
2682 return true;
2683 }
2684 return false;
2685}
2686
2687/* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2688 * It computes the number of packets to send (sndcnt) based on packets newly
2689 * delivered:
2690 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2691 * cwnd reductions across a full RTT.
2692 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2693 * But when SND_UNA is acked without further losses,
2694 * slow starts cwnd up to ssthresh to speed up the recovery.
2695 */
2696static void tcp_init_cwnd_reduction(struct sock *sk)
2697{
2698 struct tcp_sock *tp = tcp_sk(sk);
2699
2700 tp->high_seq = tp->snd_nxt;
2701 tp->tlp_high_seq = 0;
2702 tp->snd_cwnd_cnt = 0;
2703 tp->prior_cwnd = tcp_snd_cwnd(tp);
2704 tp->prr_delivered = 0;
2705 tp->prr_out = 0;
2706 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2707 tcp_ecn_queue_cwr(tp);
2708}
2709
2710void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int newly_lost, int flag)
2711{
2712 struct tcp_sock *tp = tcp_sk(sk);
2713 int sndcnt = 0;
2714 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2715
2716 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2717 return;
2718
2719 tp->prr_delivered += newly_acked_sacked;
2720 if (delta < 0) {
2721 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2722 tp->prior_cwnd - 1;
2723 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2724 } else {
2725 sndcnt = max_t(int, tp->prr_delivered - tp->prr_out,
2726 newly_acked_sacked);
2727 if (flag & FLAG_SND_UNA_ADVANCED && !newly_lost)
2728 sndcnt++;
2729 sndcnt = min(delta, sndcnt);
2730 }
2731 /* Force a fast retransmit upon entering fast recovery */
2732 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
2733 tcp_snd_cwnd_set(tp, tcp_packets_in_flight(tp) + sndcnt);
2734}
2735
2736static inline void tcp_end_cwnd_reduction(struct sock *sk)
2737{
2738 struct tcp_sock *tp = tcp_sk(sk);
2739
2740 if (inet_csk(sk)->icsk_ca_ops->cong_control)
2741 return;
2742
2743 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2744 if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH &&
2745 (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) {
2746 tcp_snd_cwnd_set(tp, tp->snd_ssthresh);
2747 tp->snd_cwnd_stamp = tcp_jiffies32;
2748 }
2749 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2750}
2751
2752/* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2753void tcp_enter_cwr(struct sock *sk)
2754{
2755 struct tcp_sock *tp = tcp_sk(sk);
2756
2757 tp->prior_ssthresh = 0;
2758 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2759 tp->undo_marker = 0;
2760 tcp_init_cwnd_reduction(sk);
2761 tcp_set_ca_state(sk, TCP_CA_CWR);
2762 }
2763}
2764EXPORT_SYMBOL(tcp_enter_cwr);
2765
2766static void tcp_try_keep_open(struct sock *sk)
2767{
2768 struct tcp_sock *tp = tcp_sk(sk);
2769 int state = TCP_CA_Open;
2770
2771 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2772 state = TCP_CA_Disorder;
2773
2774 if (inet_csk(sk)->icsk_ca_state != state) {
2775 tcp_set_ca_state(sk, state);
2776 tp->high_seq = tp->snd_nxt;
2777 }
2778}
2779
2780static void tcp_try_to_open(struct sock *sk, int flag)
2781{
2782 struct tcp_sock *tp = tcp_sk(sk);
2783
2784 tcp_verify_left_out(tp);
2785
2786 if (!tcp_any_retrans_done(sk))
2787 tp->retrans_stamp = 0;
2788
2789 if (flag & FLAG_ECE)
2790 tcp_enter_cwr(sk);
2791
2792 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2793 tcp_try_keep_open(sk);
2794 }
2795}
2796
2797static void tcp_mtup_probe_failed(struct sock *sk)
2798{
2799 struct inet_connection_sock *icsk = inet_csk(sk);
2800
2801 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2802 icsk->icsk_mtup.probe_size = 0;
2803 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
2804}
2805
2806static void tcp_mtup_probe_success(struct sock *sk)
2807{
2808 struct tcp_sock *tp = tcp_sk(sk);
2809 struct inet_connection_sock *icsk = inet_csk(sk);
2810 u64 val;
2811
2812 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2813
2814 val = (u64)tcp_snd_cwnd(tp) * tcp_mss_to_mtu(sk, tp->mss_cache);
2815 do_div(val, icsk->icsk_mtup.probe_size);
2816 DEBUG_NET_WARN_ON_ONCE((u32)val != val);
2817 tcp_snd_cwnd_set(tp, max_t(u32, 1U, val));
2818
2819 tp->snd_cwnd_cnt = 0;
2820 tp->snd_cwnd_stamp = tcp_jiffies32;
2821 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2822
2823 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2824 icsk->icsk_mtup.probe_size = 0;
2825 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2826 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
2827}
2828
2829/* Sometimes we deduce that packets have been dropped due to reasons other than
2830 * congestion, like path MTU reductions or failed client TFO attempts. In these
2831 * cases we call this function to retransmit as many packets as cwnd allows,
2832 * without reducing cwnd. Given that retransmits will set retrans_stamp to a
2833 * non-zero value (and may do so in a later calling context due to TSQ), we
2834 * also enter CA_Loss so that we track when all retransmitted packets are ACKed
2835 * and clear retrans_stamp when that happens (to ensure later recurring RTOs
2836 * are using the correct retrans_stamp and don't declare ETIMEDOUT
2837 * prematurely).
2838 */
2839static void tcp_non_congestion_loss_retransmit(struct sock *sk)
2840{
2841 const struct inet_connection_sock *icsk = inet_csk(sk);
2842 struct tcp_sock *tp = tcp_sk(sk);
2843
2844 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2845 tp->high_seq = tp->snd_nxt;
2846 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2847 tp->prior_ssthresh = 0;
2848 tp->undo_marker = 0;
2849 tcp_set_ca_state(sk, TCP_CA_Loss);
2850 }
2851 tcp_xmit_retransmit_queue(sk);
2852}
2853
2854/* Do a simple retransmit without using the backoff mechanisms in
2855 * tcp_timer. This is used for path mtu discovery.
2856 * The socket is already locked here.
2857 */
2858void tcp_simple_retransmit(struct sock *sk)
2859{
2860 struct tcp_sock *tp = tcp_sk(sk);
2861 struct sk_buff *skb;
2862 int mss;
2863
2864 /* A fastopen SYN request is stored as two separate packets within
2865 * the retransmit queue, this is done by tcp_send_syn_data().
2866 * As a result simply checking the MSS of the frames in the queue
2867 * will not work for the SYN packet.
2868 *
2869 * Us being here is an indication of a path MTU issue so we can
2870 * assume that the fastopen SYN was lost and just mark all the
2871 * frames in the retransmit queue as lost. We will use an MSS of
2872 * -1 to mark all frames as lost, otherwise compute the current MSS.
2873 */
2874 if (tp->syn_data && sk->sk_state == TCP_SYN_SENT)
2875 mss = -1;
2876 else
2877 mss = tcp_current_mss(sk);
2878
2879 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2880 if (tcp_skb_seglen(skb) > mss)
2881 tcp_mark_skb_lost(sk, skb);
2882 }
2883
2884 tcp_clear_retrans_hints_partial(tp);
2885
2886 if (!tp->lost_out)
2887 return;
2888
2889 if (tcp_is_reno(tp))
2890 tcp_limit_reno_sacked(tp);
2891
2892 tcp_verify_left_out(tp);
2893
2894 /* Don't muck with the congestion window here.
2895 * Reason is that we do not increase amount of _data_
2896 * in network, but units changed and effective
2897 * cwnd/ssthresh really reduced now.
2898 */
2899 tcp_non_congestion_loss_retransmit(sk);
2900}
2901EXPORT_SYMBOL(tcp_simple_retransmit);
2902
2903void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2904{
2905 struct tcp_sock *tp = tcp_sk(sk);
2906 int mib_idx;
2907
2908 /* Start the clock with our fast retransmit, for undo and ETIMEDOUT. */
2909 tcp_retrans_stamp_cleanup(sk);
2910
2911 if (tcp_is_reno(tp))
2912 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2913 else
2914 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2915
2916 NET_INC_STATS(sock_net(sk), mib_idx);
2917
2918 tp->prior_ssthresh = 0;
2919 tcp_init_undo(tp);
2920
2921 if (!tcp_in_cwnd_reduction(sk)) {
2922 if (!ece_ack)
2923 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2924 tcp_init_cwnd_reduction(sk);
2925 }
2926 tcp_set_ca_state(sk, TCP_CA_Recovery);
2927}
2928
2929static void tcp_update_rto_time(struct tcp_sock *tp)
2930{
2931 if (tp->rto_stamp) {
2932 tp->total_rto_time += tcp_time_stamp_ms(tp) - tp->rto_stamp;
2933 tp->rto_stamp = 0;
2934 }
2935}
2936
2937/* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2938 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2939 */
2940static void tcp_process_loss(struct sock *sk, int flag, int num_dupack,
2941 int *rexmit)
2942{
2943 struct tcp_sock *tp = tcp_sk(sk);
2944 bool recovered = !before(tp->snd_una, tp->high_seq);
2945
2946 if ((flag & FLAG_SND_UNA_ADVANCED || rcu_access_pointer(tp->fastopen_rsk)) &&
2947 tcp_try_undo_loss(sk, false))
2948 return;
2949
2950 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2951 /* Step 3.b. A timeout is spurious if not all data are
2952 * lost, i.e., never-retransmitted data are (s)acked.
2953 */
2954 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2955 tcp_try_undo_loss(sk, true))
2956 return;
2957
2958 if (after(tp->snd_nxt, tp->high_seq)) {
2959 if (flag & FLAG_DATA_SACKED || num_dupack)
2960 tp->frto = 0; /* Step 3.a. loss was real */
2961 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2962 tp->high_seq = tp->snd_nxt;
2963 /* Step 2.b. Try send new data (but deferred until cwnd
2964 * is updated in tcp_ack()). Otherwise fall back to
2965 * the conventional recovery.
2966 */
2967 if (!tcp_write_queue_empty(sk) &&
2968 after(tcp_wnd_end(tp), tp->snd_nxt)) {
2969 *rexmit = REXMIT_NEW;
2970 return;
2971 }
2972 tp->frto = 0;
2973 }
2974 }
2975
2976 if (recovered) {
2977 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2978 tcp_try_undo_recovery(sk);
2979 return;
2980 }
2981 if (tcp_is_reno(tp)) {
2982 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2983 * delivered. Lower inflight to clock out (re)transmissions.
2984 */
2985 if (after(tp->snd_nxt, tp->high_seq) && num_dupack)
2986 tcp_add_reno_sack(sk, num_dupack, flag & FLAG_ECE);
2987 else if (flag & FLAG_SND_UNA_ADVANCED)
2988 tcp_reset_reno_sack(tp);
2989 }
2990 *rexmit = REXMIT_LOST;
2991}
2992
2993static bool tcp_force_fast_retransmit(struct sock *sk)
2994{
2995 struct tcp_sock *tp = tcp_sk(sk);
2996
2997 return after(tcp_highest_sack_seq(tp),
2998 tp->snd_una + tp->reordering * tp->mss_cache);
2999}
3000
3001/* Undo during fast recovery after partial ACK. */
3002static bool tcp_try_undo_partial(struct sock *sk, u32 prior_snd_una,
3003 bool *do_lost)
3004{
3005 struct tcp_sock *tp = tcp_sk(sk);
3006
3007 if (tp->undo_marker && tcp_packet_delayed(tp)) {
3008 /* Plain luck! Hole if filled with delayed
3009 * packet, rather than with a retransmit. Check reordering.
3010 */
3011 tcp_check_sack_reordering(sk, prior_snd_una, 1);
3012
3013 /* We are getting evidence that the reordering degree is higher
3014 * than we realized. If there are no retransmits out then we
3015 * can undo. Otherwise we clock out new packets but do not
3016 * mark more packets lost or retransmit more.
3017 */
3018 if (tp->retrans_out)
3019 return true;
3020
3021 if (!tcp_any_retrans_done(sk))
3022 tp->retrans_stamp = 0;
3023
3024 DBGUNDO(sk, "partial recovery");
3025 tcp_undo_cwnd_reduction(sk, true);
3026 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
3027 tcp_try_keep_open(sk);
3028 } else {
3029 /* Partial ACK arrived. Force fast retransmit. */
3030 *do_lost = tcp_force_fast_retransmit(sk);
3031 }
3032 return false;
3033}
3034
3035static void tcp_identify_packet_loss(struct sock *sk, int *ack_flag)
3036{
3037 struct tcp_sock *tp = tcp_sk(sk);
3038
3039 if (tcp_rtx_queue_empty(sk))
3040 return;
3041
3042 if (unlikely(tcp_is_reno(tp))) {
3043 tcp_newreno_mark_lost(sk, *ack_flag & FLAG_SND_UNA_ADVANCED);
3044 } else if (tcp_is_rack(sk)) {
3045 u32 prior_retrans = tp->retrans_out;
3046
3047 if (tcp_rack_mark_lost(sk))
3048 *ack_flag &= ~FLAG_SET_XMIT_TIMER;
3049 if (prior_retrans > tp->retrans_out)
3050 *ack_flag |= FLAG_LOST_RETRANS;
3051 }
3052}
3053
3054/* Process an event, which can update packets-in-flight not trivially.
3055 * Main goal of this function is to calculate new estimate for left_out,
3056 * taking into account both packets sitting in receiver's buffer and
3057 * packets lost by network.
3058 *
3059 * Besides that it updates the congestion state when packet loss or ECN
3060 * is detected. But it does not reduce the cwnd, it is done by the
3061 * congestion control later.
3062 *
3063 * It does _not_ decide what to send, it is made in function
3064 * tcp_xmit_retransmit_queue().
3065 */
3066static void tcp_fastretrans_alert(struct sock *sk, const u32 prior_snd_una,
3067 int num_dupack, int *ack_flag, int *rexmit)
3068{
3069 struct inet_connection_sock *icsk = inet_csk(sk);
3070 struct tcp_sock *tp = tcp_sk(sk);
3071 int fast_rexmit = 0, flag = *ack_flag;
3072 bool ece_ack = flag & FLAG_ECE;
3073 bool do_lost = num_dupack || ((flag & FLAG_DATA_SACKED) &&
3074 tcp_force_fast_retransmit(sk));
3075
3076 if (!tp->packets_out && tp->sacked_out)
3077 tp->sacked_out = 0;
3078
3079 /* Now state machine starts.
3080 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
3081 if (ece_ack)
3082 tp->prior_ssthresh = 0;
3083
3084 /* B. In all the states check for reneging SACKs. */
3085 if (tcp_check_sack_reneging(sk, ack_flag))
3086 return;
3087
3088 /* C. Check consistency of the current state. */
3089 tcp_verify_left_out(tp);
3090
3091 /* D. Check state exit conditions. State can be terminated
3092 * when high_seq is ACKed. */
3093 if (icsk->icsk_ca_state == TCP_CA_Open) {
3094 WARN_ON(tp->retrans_out != 0 && !tp->syn_data);
3095 tp->retrans_stamp = 0;
3096 } else if (!before(tp->snd_una, tp->high_seq)) {
3097 switch (icsk->icsk_ca_state) {
3098 case TCP_CA_CWR:
3099 /* CWR is to be held something *above* high_seq
3100 * is ACKed for CWR bit to reach receiver. */
3101 if (tp->snd_una != tp->high_seq) {
3102 tcp_end_cwnd_reduction(sk);
3103 tcp_set_ca_state(sk, TCP_CA_Open);
3104 }
3105 break;
3106
3107 case TCP_CA_Recovery:
3108 if (tcp_is_reno(tp))
3109 tcp_reset_reno_sack(tp);
3110 if (tcp_try_undo_recovery(sk))
3111 return;
3112 tcp_end_cwnd_reduction(sk);
3113 break;
3114 }
3115 }
3116
3117 /* E. Process state. */
3118 switch (icsk->icsk_ca_state) {
3119 case TCP_CA_Recovery:
3120 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
3121 if (tcp_is_reno(tp))
3122 tcp_add_reno_sack(sk, num_dupack, ece_ack);
3123 } else if (tcp_try_undo_partial(sk, prior_snd_una, &do_lost))
3124 return;
3125
3126 if (tcp_try_undo_dsack(sk))
3127 tcp_try_to_open(sk, flag);
3128
3129 tcp_identify_packet_loss(sk, ack_flag);
3130 if (icsk->icsk_ca_state != TCP_CA_Recovery) {
3131 if (!tcp_time_to_recover(sk, flag))
3132 return;
3133 /* Undo reverts the recovery state. If loss is evident,
3134 * starts a new recovery (e.g. reordering then loss);
3135 */
3136 tcp_enter_recovery(sk, ece_ack);
3137 }
3138 break;
3139 case TCP_CA_Loss:
3140 tcp_process_loss(sk, flag, num_dupack, rexmit);
3141 if (icsk->icsk_ca_state != TCP_CA_Loss)
3142 tcp_update_rto_time(tp);
3143 tcp_identify_packet_loss(sk, ack_flag);
3144 if (!(icsk->icsk_ca_state == TCP_CA_Open ||
3145 (*ack_flag & FLAG_LOST_RETRANS)))
3146 return;
3147 /* Change state if cwnd is undone or retransmits are lost */
3148 fallthrough;
3149 default:
3150 if (tcp_is_reno(tp)) {
3151 if (flag & FLAG_SND_UNA_ADVANCED)
3152 tcp_reset_reno_sack(tp);
3153 tcp_add_reno_sack(sk, num_dupack, ece_ack);
3154 }
3155
3156 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
3157 tcp_try_undo_dsack(sk);
3158
3159 tcp_identify_packet_loss(sk, ack_flag);
3160 if (!tcp_time_to_recover(sk, flag)) {
3161 tcp_try_to_open(sk, flag);
3162 return;
3163 }
3164
3165 /* MTU probe failure: don't reduce cwnd */
3166 if (icsk->icsk_ca_state < TCP_CA_CWR &&
3167 icsk->icsk_mtup.probe_size &&
3168 tp->snd_una == tp->mtu_probe.probe_seq_start) {
3169 tcp_mtup_probe_failed(sk);
3170 /* Restores the reduction we did in tcp_mtup_probe() */
3171 tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) + 1);
3172 tcp_simple_retransmit(sk);
3173 return;
3174 }
3175
3176 /* Otherwise enter Recovery state */
3177 tcp_enter_recovery(sk, ece_ack);
3178 fast_rexmit = 1;
3179 }
3180
3181 if (!tcp_is_rack(sk) && do_lost)
3182 tcp_update_scoreboard(sk, fast_rexmit);
3183 *rexmit = REXMIT_LOST;
3184}
3185
3186static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us, const int flag)
3187{
3188 u32 wlen = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_rtt_wlen) * HZ;
3189 struct tcp_sock *tp = tcp_sk(sk);
3190
3191 if ((flag & FLAG_ACK_MAYBE_DELAYED) && rtt_us > tcp_min_rtt(tp)) {
3192 /* If the remote keeps returning delayed ACKs, eventually
3193 * the min filter would pick it up and overestimate the
3194 * prop. delay when it expires. Skip suspected delayed ACKs.
3195 */
3196 return;
3197 }
3198 minmax_running_min(&tp->rtt_min, wlen, tcp_jiffies32,
3199 rtt_us ? : jiffies_to_usecs(1));
3200}
3201
3202static bool tcp_ack_update_rtt(struct sock *sk, const int flag,
3203 long seq_rtt_us, long sack_rtt_us,
3204 long ca_rtt_us, struct rate_sample *rs)
3205{
3206 const struct tcp_sock *tp = tcp_sk(sk);
3207
3208 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
3209 * broken middle-boxes or peers may corrupt TS-ECR fields. But
3210 * Karn's algorithm forbids taking RTT if some retransmitted data
3211 * is acked (RFC6298).
3212 */
3213 if (seq_rtt_us < 0)
3214 seq_rtt_us = sack_rtt_us;
3215
3216 /* RTTM Rule: A TSecr value received in a segment is used to
3217 * update the averaged RTT measurement only if the segment
3218 * acknowledges some new data, i.e., only if it advances the
3219 * left edge of the send window.
3220 * See draft-ietf-tcplw-high-performance-00, section 3.3.
3221 */
3222 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp &&
3223 tp->rx_opt.rcv_tsecr && flag & FLAG_ACKED)
3224 seq_rtt_us = ca_rtt_us = tcp_rtt_tsopt_us(tp);
3225
3226 rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */
3227 if (seq_rtt_us < 0)
3228 return false;
3229
3230 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
3231 * always taken together with ACK, SACK, or TS-opts. Any negative
3232 * values will be skipped with the seq_rtt_us < 0 check above.
3233 */
3234 tcp_update_rtt_min(sk, ca_rtt_us, flag);
3235 tcp_rtt_estimator(sk, seq_rtt_us);
3236 tcp_set_rto(sk);
3237
3238 /* RFC6298: only reset backoff on valid RTT measurement. */
3239 inet_csk(sk)->icsk_backoff = 0;
3240 return true;
3241}
3242
3243/* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
3244void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
3245{
3246 struct rate_sample rs;
3247 long rtt_us = -1L;
3248
3249 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack)
3250 rtt_us = tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req)->snt_synack);
3251
3252 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us, &rs);
3253}
3254
3255
3256static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
3257{
3258 const struct inet_connection_sock *icsk = inet_csk(sk);
3259
3260 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
3261 tcp_sk(sk)->snd_cwnd_stamp = tcp_jiffies32;
3262}
3263
3264/* Restart timer after forward progress on connection.
3265 * RFC2988 recommends to restart timer to now+rto.
3266 */
3267void tcp_rearm_rto(struct sock *sk)
3268{
3269 const struct inet_connection_sock *icsk = inet_csk(sk);
3270 struct tcp_sock *tp = tcp_sk(sk);
3271
3272 /* If the retrans timer is currently being used by Fast Open
3273 * for SYN-ACK retrans purpose, stay put.
3274 */
3275 if (rcu_access_pointer(tp->fastopen_rsk))
3276 return;
3277
3278 if (!tp->packets_out) {
3279 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3280 } else {
3281 u32 rto = inet_csk(sk)->icsk_rto;
3282 /* Offset the time elapsed after installing regular RTO */
3283 if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT ||
3284 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
3285 s64 delta_us = tcp_rto_delta_us(sk);
3286 /* delta_us may not be positive if the socket is locked
3287 * when the retrans timer fires and is rescheduled.
3288 */
3289 rto = usecs_to_jiffies(max_t(int, delta_us, 1));
3290 }
3291 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
3292 TCP_RTO_MAX);
3293 }
3294}
3295
3296/* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */
3297static void tcp_set_xmit_timer(struct sock *sk)
3298{
3299 if (!tcp_schedule_loss_probe(sk, true))
3300 tcp_rearm_rto(sk);
3301}
3302
3303/* If we get here, the whole TSO packet has not been acked. */
3304static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3305{
3306 struct tcp_sock *tp = tcp_sk(sk);
3307 u32 packets_acked;
3308
3309 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3310
3311 packets_acked = tcp_skb_pcount(skb);
3312 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3313 return 0;
3314 packets_acked -= tcp_skb_pcount(skb);
3315
3316 if (packets_acked) {
3317 BUG_ON(tcp_skb_pcount(skb) == 0);
3318 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3319 }
3320
3321 return packets_acked;
3322}
3323
3324static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3325 const struct sk_buff *ack_skb, u32 prior_snd_una)
3326{
3327 const struct skb_shared_info *shinfo;
3328
3329 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3330 if (likely(!TCP_SKB_CB(skb)->txstamp_ack))
3331 return;
3332
3333 shinfo = skb_shinfo(skb);
3334 if (!before(shinfo->tskey, prior_snd_una) &&
3335 before(shinfo->tskey, tcp_sk(sk)->snd_una)) {
3336 tcp_skb_tsorted_save(skb) {
3337 __skb_tstamp_tx(skb, ack_skb, NULL, sk, SCM_TSTAMP_ACK);
3338 } tcp_skb_tsorted_restore(skb);
3339 }
3340}
3341
3342/* Remove acknowledged frames from the retransmission queue. If our packet
3343 * is before the ack sequence we can discard it as it's confirmed to have
3344 * arrived at the other end.
3345 */
3346static int tcp_clean_rtx_queue(struct sock *sk, const struct sk_buff *ack_skb,
3347 u32 prior_fack, u32 prior_snd_una,
3348 struct tcp_sacktag_state *sack, bool ece_ack)
3349{
3350 const struct inet_connection_sock *icsk = inet_csk(sk);
3351 u64 first_ackt, last_ackt;
3352 struct tcp_sock *tp = tcp_sk(sk);
3353 u32 prior_sacked = tp->sacked_out;
3354 u32 reord = tp->snd_nxt; /* lowest acked un-retx un-sacked seq */
3355 struct sk_buff *skb, *next;
3356 bool fully_acked = true;
3357 long sack_rtt_us = -1L;
3358 long seq_rtt_us = -1L;
3359 long ca_rtt_us = -1L;
3360 u32 pkts_acked = 0;
3361 bool rtt_update;
3362 int flag = 0;
3363
3364 first_ackt = 0;
3365
3366 for (skb = skb_rb_first(&sk->tcp_rtx_queue); skb; skb = next) {
3367 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3368 const u32 start_seq = scb->seq;
3369 u8 sacked = scb->sacked;
3370 u32 acked_pcount;
3371
3372 /* Determine how many packets and what bytes were acked, tso and else */
3373 if (after(scb->end_seq, tp->snd_una)) {
3374 if (tcp_skb_pcount(skb) == 1 ||
3375 !after(tp->snd_una, scb->seq))
3376 break;
3377
3378 acked_pcount = tcp_tso_acked(sk, skb);
3379 if (!acked_pcount)
3380 break;
3381 fully_acked = false;
3382 } else {
3383 acked_pcount = tcp_skb_pcount(skb);
3384 }
3385
3386 if (unlikely(sacked & TCPCB_RETRANS)) {
3387 if (sacked & TCPCB_SACKED_RETRANS)
3388 tp->retrans_out -= acked_pcount;
3389 flag |= FLAG_RETRANS_DATA_ACKED;
3390 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3391 last_ackt = tcp_skb_timestamp_us(skb);
3392 WARN_ON_ONCE(last_ackt == 0);
3393 if (!first_ackt)
3394 first_ackt = last_ackt;
3395
3396 if (before(start_seq, reord))
3397 reord = start_seq;
3398 if (!after(scb->end_seq, tp->high_seq))
3399 flag |= FLAG_ORIG_SACK_ACKED;
3400 }
3401
3402 if (sacked & TCPCB_SACKED_ACKED) {
3403 tp->sacked_out -= acked_pcount;
3404 } else if (tcp_is_sack(tp)) {
3405 tcp_count_delivered(tp, acked_pcount, ece_ack);
3406 if (!tcp_skb_spurious_retrans(tp, skb))
3407 tcp_rack_advance(tp, sacked, scb->end_seq,
3408 tcp_skb_timestamp_us(skb));
3409 }
3410 if (sacked & TCPCB_LOST)
3411 tp->lost_out -= acked_pcount;
3412
3413 tp->packets_out -= acked_pcount;
3414 pkts_acked += acked_pcount;
3415 tcp_rate_skb_delivered(sk, skb, sack->rate);
3416
3417 /* Initial outgoing SYN's get put onto the write_queue
3418 * just like anything else we transmit. It is not
3419 * true data, and if we misinform our callers that
3420 * this ACK acks real data, we will erroneously exit
3421 * connection startup slow start one packet too
3422 * quickly. This is severely frowned upon behavior.
3423 */
3424 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3425 flag |= FLAG_DATA_ACKED;
3426 } else {
3427 flag |= FLAG_SYN_ACKED;
3428 tp->retrans_stamp = 0;
3429 }
3430
3431 if (!fully_acked)
3432 break;
3433
3434 tcp_ack_tstamp(sk, skb, ack_skb, prior_snd_una);
3435
3436 next = skb_rb_next(skb);
3437 if (unlikely(skb == tp->retransmit_skb_hint))
3438 tp->retransmit_skb_hint = NULL;
3439 if (unlikely(skb == tp->lost_skb_hint))
3440 tp->lost_skb_hint = NULL;
3441 tcp_highest_sack_replace(sk, skb, next);
3442 tcp_rtx_queue_unlink_and_free(skb, sk);
3443 }
3444
3445 if (!skb)
3446 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3447
3448 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3449 tp->snd_up = tp->snd_una;
3450
3451 if (skb) {
3452 tcp_ack_tstamp(sk, skb, ack_skb, prior_snd_una);
3453 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
3454 flag |= FLAG_SACK_RENEGING;
3455 }
3456
3457 if (likely(first_ackt) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
3458 seq_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, first_ackt);
3459 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, last_ackt);
3460
3461 if (pkts_acked == 1 && fully_acked && !prior_sacked &&
3462 (tp->snd_una - prior_snd_una) < tp->mss_cache &&
3463 sack->rate->prior_delivered + 1 == tp->delivered &&
3464 !(flag & (FLAG_CA_ALERT | FLAG_SYN_ACKED))) {
3465 /* Conservatively mark a delayed ACK. It's typically
3466 * from a lone runt packet over the round trip to
3467 * a receiver w/o out-of-order or CE events.
3468 */
3469 flag |= FLAG_ACK_MAYBE_DELAYED;
3470 }
3471 }
3472 if (sack->first_sackt) {
3473 sack_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->first_sackt);
3474 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->last_sackt);
3475 }
3476 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
3477 ca_rtt_us, sack->rate);
3478
3479 if (flag & FLAG_ACKED) {
3480 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3481 if (unlikely(icsk->icsk_mtup.probe_size &&
3482 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3483 tcp_mtup_probe_success(sk);
3484 }
3485
3486 if (tcp_is_reno(tp)) {
3487 tcp_remove_reno_sacks(sk, pkts_acked, ece_ack);
3488
3489 /* If any of the cumulatively ACKed segments was
3490 * retransmitted, non-SACK case cannot confirm that
3491 * progress was due to original transmission due to
3492 * lack of TCPCB_SACKED_ACKED bits even if some of
3493 * the packets may have been never retransmitted.
3494 */
3495 if (flag & FLAG_RETRANS_DATA_ACKED)
3496 flag &= ~FLAG_ORIG_SACK_ACKED;
3497 } else {
3498 int delta;
3499
3500 /* Non-retransmitted hole got filled? That's reordering */
3501 if (before(reord, prior_fack))
3502 tcp_check_sack_reordering(sk, reord, 0);
3503
3504 delta = prior_sacked - tp->sacked_out;
3505 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3506 }
3507 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3508 sack_rtt_us > tcp_stamp_us_delta(tp->tcp_mstamp,
3509 tcp_skb_timestamp_us(skb))) {
3510 /* Do not re-arm RTO if the sack RTT is measured from data sent
3511 * after when the head was last (re)transmitted. Otherwise the
3512 * timeout may continue to extend in loss recovery.
3513 */
3514 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3515 }
3516
3517 if (icsk->icsk_ca_ops->pkts_acked) {
3518 struct ack_sample sample = { .pkts_acked = pkts_acked,
3519 .rtt_us = sack->rate->rtt_us };
3520
3521 sample.in_flight = tp->mss_cache *
3522 (tp->delivered - sack->rate->prior_delivered);
3523 icsk->icsk_ca_ops->pkts_acked(sk, &sample);
3524 }
3525
3526#if FASTRETRANS_DEBUG > 0
3527 WARN_ON((int)tp->sacked_out < 0);
3528 WARN_ON((int)tp->lost_out < 0);
3529 WARN_ON((int)tp->retrans_out < 0);
3530 if (!tp->packets_out && tcp_is_sack(tp)) {
3531 icsk = inet_csk(sk);
3532 if (tp->lost_out) {
3533 pr_debug("Leak l=%u %d\n",
3534 tp->lost_out, icsk->icsk_ca_state);
3535 tp->lost_out = 0;
3536 }
3537 if (tp->sacked_out) {
3538 pr_debug("Leak s=%u %d\n",
3539 tp->sacked_out, icsk->icsk_ca_state);
3540 tp->sacked_out = 0;
3541 }
3542 if (tp->retrans_out) {
3543 pr_debug("Leak r=%u %d\n",
3544 tp->retrans_out, icsk->icsk_ca_state);
3545 tp->retrans_out = 0;
3546 }
3547 }
3548#endif
3549 return flag;
3550}
3551
3552static void tcp_ack_probe(struct sock *sk)
3553{
3554 struct inet_connection_sock *icsk = inet_csk(sk);
3555 struct sk_buff *head = tcp_send_head(sk);
3556 const struct tcp_sock *tp = tcp_sk(sk);
3557
3558 /* Was it a usable window open? */
3559 if (!head)
3560 return;
3561 if (!after(TCP_SKB_CB(head)->end_seq, tcp_wnd_end(tp))) {
3562 icsk->icsk_backoff = 0;
3563 icsk->icsk_probes_tstamp = 0;
3564 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3565 /* Socket must be waked up by subsequent tcp_data_snd_check().
3566 * This function is not for random using!
3567 */
3568 } else {
3569 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
3570
3571 when = tcp_clamp_probe0_to_user_timeout(sk, when);
3572 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, when, TCP_RTO_MAX);
3573 }
3574}
3575
3576static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3577{
3578 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3579 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3580}
3581
3582/* Decide wheather to run the increase function of congestion control. */
3583static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3584{
3585 /* If reordering is high then always grow cwnd whenever data is
3586 * delivered regardless of its ordering. Otherwise stay conservative
3587 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3588 * new SACK or ECE mark may first advance cwnd here and later reduce
3589 * cwnd in tcp_fastretrans_alert() based on more states.
3590 */
3591 if (tcp_sk(sk)->reordering >
3592 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_reordering))
3593 return flag & FLAG_FORWARD_PROGRESS;
3594
3595 return flag & FLAG_DATA_ACKED;
3596}
3597
3598/* The "ultimate" congestion control function that aims to replace the rigid
3599 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3600 * It's called toward the end of processing an ACK with precise rate
3601 * information. All transmission or retransmission are delayed afterwards.
3602 */
3603static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
3604 int flag, const struct rate_sample *rs)
3605{
3606 const struct inet_connection_sock *icsk = inet_csk(sk);
3607
3608 if (icsk->icsk_ca_ops->cong_control) {
3609 icsk->icsk_ca_ops->cong_control(sk, ack, flag, rs);
3610 return;
3611 }
3612
3613 if (tcp_in_cwnd_reduction(sk)) {
3614 /* Reduce cwnd if state mandates */
3615 tcp_cwnd_reduction(sk, acked_sacked, rs->losses, flag);
3616 } else if (tcp_may_raise_cwnd(sk, flag)) {
3617 /* Advance cwnd if state allows */
3618 tcp_cong_avoid(sk, ack, acked_sacked);
3619 }
3620 tcp_update_pacing_rate(sk);
3621}
3622
3623/* Check that window update is acceptable.
3624 * The function assumes that snd_una<=ack<=snd_next.
3625 */
3626static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3627 const u32 ack, const u32 ack_seq,
3628 const u32 nwin)
3629{
3630 return after(ack, tp->snd_una) ||
3631 after(ack_seq, tp->snd_wl1) ||
3632 (ack_seq == tp->snd_wl1 && (nwin > tp->snd_wnd || !nwin));
3633}
3634
3635static void tcp_snd_sne_update(struct tcp_sock *tp, u32 ack)
3636{
3637#ifdef CONFIG_TCP_AO
3638 struct tcp_ao_info *ao;
3639
3640 if (!static_branch_unlikely(&tcp_ao_needed.key))
3641 return;
3642
3643 ao = rcu_dereference_protected(tp->ao_info,
3644 lockdep_sock_is_held((struct sock *)tp));
3645 if (ao && ack < tp->snd_una) {
3646 ao->snd_sne++;
3647 trace_tcp_ao_snd_sne_update((struct sock *)tp, ao->snd_sne);
3648 }
3649#endif
3650}
3651
3652/* If we update tp->snd_una, also update tp->bytes_acked */
3653static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3654{
3655 u32 delta = ack - tp->snd_una;
3656
3657 sock_owned_by_me((struct sock *)tp);
3658 tp->bytes_acked += delta;
3659 tcp_snd_sne_update(tp, ack);
3660 tp->snd_una = ack;
3661}
3662
3663static void tcp_rcv_sne_update(struct tcp_sock *tp, u32 seq)
3664{
3665#ifdef CONFIG_TCP_AO
3666 struct tcp_ao_info *ao;
3667
3668 if (!static_branch_unlikely(&tcp_ao_needed.key))
3669 return;
3670
3671 ao = rcu_dereference_protected(tp->ao_info,
3672 lockdep_sock_is_held((struct sock *)tp));
3673 if (ao && seq < tp->rcv_nxt) {
3674 ao->rcv_sne++;
3675 trace_tcp_ao_rcv_sne_update((struct sock *)tp, ao->rcv_sne);
3676 }
3677#endif
3678}
3679
3680/* If we update tp->rcv_nxt, also update tp->bytes_received */
3681static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3682{
3683 u32 delta = seq - tp->rcv_nxt;
3684
3685 sock_owned_by_me((struct sock *)tp);
3686 tp->bytes_received += delta;
3687 tcp_rcv_sne_update(tp, seq);
3688 WRITE_ONCE(tp->rcv_nxt, seq);
3689}
3690
3691/* Update our send window.
3692 *
3693 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3694 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3695 */
3696static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3697 u32 ack_seq)
3698{
3699 struct tcp_sock *tp = tcp_sk(sk);
3700 int flag = 0;
3701 u32 nwin = ntohs(tcp_hdr(skb)->window);
3702
3703 if (likely(!tcp_hdr(skb)->syn))
3704 nwin <<= tp->rx_opt.snd_wscale;
3705
3706 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3707 flag |= FLAG_WIN_UPDATE;
3708 tcp_update_wl(tp, ack_seq);
3709
3710 if (tp->snd_wnd != nwin) {
3711 tp->snd_wnd = nwin;
3712
3713 /* Note, it is the only place, where
3714 * fast path is recovered for sending TCP.
3715 */
3716 tp->pred_flags = 0;
3717 tcp_fast_path_check(sk);
3718
3719 if (!tcp_write_queue_empty(sk))
3720 tcp_slow_start_after_idle_check(sk);
3721
3722 if (nwin > tp->max_window) {
3723 tp->max_window = nwin;
3724 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3725 }
3726 }
3727 }
3728
3729 tcp_snd_una_update(tp, ack);
3730
3731 return flag;
3732}
3733
3734static bool __tcp_oow_rate_limited(struct net *net, int mib_idx,
3735 u32 *last_oow_ack_time)
3736{
3737 /* Paired with the WRITE_ONCE() in this function. */
3738 u32 val = READ_ONCE(*last_oow_ack_time);
3739
3740 if (val) {
3741 s32 elapsed = (s32)(tcp_jiffies32 - val);
3742
3743 if (0 <= elapsed &&
3744 elapsed < READ_ONCE(net->ipv4.sysctl_tcp_invalid_ratelimit)) {
3745 NET_INC_STATS(net, mib_idx);
3746 return true; /* rate-limited: don't send yet! */
3747 }
3748 }
3749
3750 /* Paired with the prior READ_ONCE() and with itself,
3751 * as we might be lockless.
3752 */
3753 WRITE_ONCE(*last_oow_ack_time, tcp_jiffies32);
3754
3755 return false; /* not rate-limited: go ahead, send dupack now! */
3756}
3757
3758/* Return true if we're currently rate-limiting out-of-window ACKs and
3759 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3760 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3761 * attacks that send repeated SYNs or ACKs for the same connection. To
3762 * do this, we do not send a duplicate SYNACK or ACK if the remote
3763 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3764 */
3765bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3766 int mib_idx, u32 *last_oow_ack_time)
3767{
3768 /* Data packets without SYNs are not likely part of an ACK loop. */
3769 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3770 !tcp_hdr(skb)->syn)
3771 return false;
3772
3773 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time);
3774}
3775
3776/* RFC 5961 7 [ACK Throttling] */
3777static void tcp_send_challenge_ack(struct sock *sk)
3778{
3779 struct tcp_sock *tp = tcp_sk(sk);
3780 struct net *net = sock_net(sk);
3781 u32 count, now, ack_limit;
3782
3783 /* First check our per-socket dupack rate limit. */
3784 if (__tcp_oow_rate_limited(net,
3785 LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3786 &tp->last_oow_ack_time))
3787 return;
3788
3789 ack_limit = READ_ONCE(net->ipv4.sysctl_tcp_challenge_ack_limit);
3790 if (ack_limit == INT_MAX)
3791 goto send_ack;
3792
3793 /* Then check host-wide RFC 5961 rate limit. */
3794 now = jiffies / HZ;
3795 if (now != READ_ONCE(net->ipv4.tcp_challenge_timestamp)) {
3796 u32 half = (ack_limit + 1) >> 1;
3797
3798 WRITE_ONCE(net->ipv4.tcp_challenge_timestamp, now);
3799 WRITE_ONCE(net->ipv4.tcp_challenge_count,
3800 get_random_u32_inclusive(half, ack_limit + half - 1));
3801 }
3802 count = READ_ONCE(net->ipv4.tcp_challenge_count);
3803 if (count > 0) {
3804 WRITE_ONCE(net->ipv4.tcp_challenge_count, count - 1);
3805send_ack:
3806 NET_INC_STATS(net, LINUX_MIB_TCPCHALLENGEACK);
3807 tcp_send_ack(sk);
3808 }
3809}
3810
3811static void tcp_store_ts_recent(struct tcp_sock *tp)
3812{
3813 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3814 tp->rx_opt.ts_recent_stamp = ktime_get_seconds();
3815}
3816
3817static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3818{
3819 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3820 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3821 * extra check below makes sure this can only happen
3822 * for pure ACK frames. -DaveM
3823 *
3824 * Not only, also it occurs for expired timestamps.
3825 */
3826
3827 if (tcp_paws_check(&tp->rx_opt, 0))
3828 tcp_store_ts_recent(tp);
3829 }
3830}
3831
3832/* This routine deals with acks during a TLP episode and ends an episode by
3833 * resetting tlp_high_seq. Ref: TLP algorithm in draft-ietf-tcpm-rack
3834 */
3835static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3836{
3837 struct tcp_sock *tp = tcp_sk(sk);
3838
3839 if (before(ack, tp->tlp_high_seq))
3840 return;
3841
3842 if (!tp->tlp_retrans) {
3843 /* TLP of new data has been acknowledged */
3844 tp->tlp_high_seq = 0;
3845 } else if (flag & FLAG_DSACK_TLP) {
3846 /* This DSACK means original and TLP probe arrived; no loss */
3847 tp->tlp_high_seq = 0;
3848 } else if (after(ack, tp->tlp_high_seq)) {
3849 /* ACK advances: there was a loss, so reduce cwnd. Reset
3850 * tlp_high_seq in tcp_init_cwnd_reduction()
3851 */
3852 tcp_init_cwnd_reduction(sk);
3853 tcp_set_ca_state(sk, TCP_CA_CWR);
3854 tcp_end_cwnd_reduction(sk);
3855 tcp_try_keep_open(sk);
3856 NET_INC_STATS(sock_net(sk),
3857 LINUX_MIB_TCPLOSSPROBERECOVERY);
3858 } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3859 FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3860 /* Pure dupack: original and TLP probe arrived; no loss */
3861 tp->tlp_high_seq = 0;
3862 }
3863}
3864
3865static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3866{
3867 const struct inet_connection_sock *icsk = inet_csk(sk);
3868
3869 if (icsk->icsk_ca_ops->in_ack_event)
3870 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3871}
3872
3873/* Congestion control has updated the cwnd already. So if we're in
3874 * loss recovery then now we do any new sends (for FRTO) or
3875 * retransmits (for CA_Loss or CA_recovery) that make sense.
3876 */
3877static void tcp_xmit_recovery(struct sock *sk, int rexmit)
3878{
3879 struct tcp_sock *tp = tcp_sk(sk);
3880
3881 if (rexmit == REXMIT_NONE || sk->sk_state == TCP_SYN_SENT)
3882 return;
3883
3884 if (unlikely(rexmit == REXMIT_NEW)) {
3885 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
3886 TCP_NAGLE_OFF);
3887 if (after(tp->snd_nxt, tp->high_seq))
3888 return;
3889 tp->frto = 0;
3890 }
3891 tcp_xmit_retransmit_queue(sk);
3892}
3893
3894/* Returns the number of packets newly acked or sacked by the current ACK */
3895static u32 tcp_newly_delivered(struct sock *sk, u32 prior_delivered, int flag)
3896{
3897 const struct net *net = sock_net(sk);
3898 struct tcp_sock *tp = tcp_sk(sk);
3899 u32 delivered;
3900
3901 delivered = tp->delivered - prior_delivered;
3902 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVERED, delivered);
3903 if (flag & FLAG_ECE)
3904 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVEREDCE, delivered);
3905
3906 return delivered;
3907}
3908
3909/* This routine deals with incoming acks, but not outgoing ones. */
3910static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3911{
3912 struct inet_connection_sock *icsk = inet_csk(sk);
3913 struct tcp_sock *tp = tcp_sk(sk);
3914 struct tcp_sacktag_state sack_state;
3915 struct rate_sample rs = { .prior_delivered = 0 };
3916 u32 prior_snd_una = tp->snd_una;
3917 bool is_sack_reneg = tp->is_sack_reneg;
3918 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3919 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3920 int num_dupack = 0;
3921 int prior_packets = tp->packets_out;
3922 u32 delivered = tp->delivered;
3923 u32 lost = tp->lost;
3924 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
3925 u32 prior_fack;
3926
3927 sack_state.first_sackt = 0;
3928 sack_state.rate = &rs;
3929 sack_state.sack_delivered = 0;
3930
3931 /* We very likely will need to access rtx queue. */
3932 prefetch(sk->tcp_rtx_queue.rb_node);
3933
3934 /* If the ack is older than previous acks
3935 * then we can probably ignore it.
3936 */
3937 if (before(ack, prior_snd_una)) {
3938 u32 max_window;
3939
3940 /* do not accept ACK for bytes we never sent. */
3941 max_window = min_t(u64, tp->max_window, tp->bytes_acked);
3942 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3943 if (before(ack, prior_snd_una - max_window)) {
3944 if (!(flag & FLAG_NO_CHALLENGE_ACK))
3945 tcp_send_challenge_ack(sk);
3946 return -SKB_DROP_REASON_TCP_TOO_OLD_ACK;
3947 }
3948 goto old_ack;
3949 }
3950
3951 /* If the ack includes data we haven't sent yet, discard
3952 * this segment (RFC793 Section 3.9).
3953 */
3954 if (after(ack, tp->snd_nxt))
3955 return -SKB_DROP_REASON_TCP_ACK_UNSENT_DATA;
3956
3957 if (after(ack, prior_snd_una)) {
3958 flag |= FLAG_SND_UNA_ADVANCED;
3959 icsk->icsk_retransmits = 0;
3960
3961#if IS_ENABLED(CONFIG_TLS_DEVICE)
3962 if (static_branch_unlikely(&clean_acked_data_enabled.key))
3963 if (icsk->icsk_clean_acked)
3964 icsk->icsk_clean_acked(sk, ack);
3965#endif
3966 }
3967
3968 prior_fack = tcp_is_sack(tp) ? tcp_highest_sack_seq(tp) : tp->snd_una;
3969 rs.prior_in_flight = tcp_packets_in_flight(tp);
3970
3971 /* ts_recent update must be made after we are sure that the packet
3972 * is in window.
3973 */
3974 if (flag & FLAG_UPDATE_TS_RECENT)
3975 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3976
3977 if ((flag & (FLAG_SLOWPATH | FLAG_SND_UNA_ADVANCED)) ==
3978 FLAG_SND_UNA_ADVANCED) {
3979 /* Window is constant, pure forward advance.
3980 * No more checks are required.
3981 * Note, we use the fact that SND.UNA>=SND.WL2.
3982 */
3983 tcp_update_wl(tp, ack_seq);
3984 tcp_snd_una_update(tp, ack);
3985 flag |= FLAG_WIN_UPDATE;
3986
3987 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3988
3989 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS);
3990 } else {
3991 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3992
3993 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3994 flag |= FLAG_DATA;
3995 else
3996 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3997
3998 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3999
4000 if (TCP_SKB_CB(skb)->sacked)
4001 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
4002 &sack_state);
4003
4004 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
4005 flag |= FLAG_ECE;
4006 ack_ev_flags |= CA_ACK_ECE;
4007 }
4008
4009 if (sack_state.sack_delivered)
4010 tcp_count_delivered(tp, sack_state.sack_delivered,
4011 flag & FLAG_ECE);
4012
4013 if (flag & FLAG_WIN_UPDATE)
4014 ack_ev_flags |= CA_ACK_WIN_UPDATE;
4015
4016 tcp_in_ack_event(sk, ack_ev_flags);
4017 }
4018
4019 /* This is a deviation from RFC3168 since it states that:
4020 * "When the TCP data sender is ready to set the CWR bit after reducing
4021 * the congestion window, it SHOULD set the CWR bit only on the first
4022 * new data packet that it transmits."
4023 * We accept CWR on pure ACKs to be more robust
4024 * with widely-deployed TCP implementations that do this.
4025 */
4026 tcp_ecn_accept_cwr(sk, skb);
4027
4028 /* We passed data and got it acked, remove any soft error
4029 * log. Something worked...
4030 */
4031 WRITE_ONCE(sk->sk_err_soft, 0);
4032 icsk->icsk_probes_out = 0;
4033 tp->rcv_tstamp = tcp_jiffies32;
4034 if (!prior_packets)
4035 goto no_queue;
4036
4037 /* See if we can take anything off of the retransmit queue. */
4038 flag |= tcp_clean_rtx_queue(sk, skb, prior_fack, prior_snd_una,
4039 &sack_state, flag & FLAG_ECE);
4040
4041 tcp_rack_update_reo_wnd(sk, &rs);
4042
4043 if (tp->tlp_high_seq)
4044 tcp_process_tlp_ack(sk, ack, flag);
4045
4046 if (tcp_ack_is_dubious(sk, flag)) {
4047 if (!(flag & (FLAG_SND_UNA_ADVANCED |
4048 FLAG_NOT_DUP | FLAG_DSACKING_ACK))) {
4049 num_dupack = 1;
4050 /* Consider if pure acks were aggregated in tcp_add_backlog() */
4051 if (!(flag & FLAG_DATA))
4052 num_dupack = max_t(u16, 1, skb_shinfo(skb)->gso_segs);
4053 }
4054 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
4055 &rexmit);
4056 }
4057
4058 /* If needed, reset TLP/RTO timer when RACK doesn't set. */
4059 if (flag & FLAG_SET_XMIT_TIMER)
4060 tcp_set_xmit_timer(sk);
4061
4062 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
4063 sk_dst_confirm(sk);
4064
4065 delivered = tcp_newly_delivered(sk, delivered, flag);
4066 lost = tp->lost - lost; /* freshly marked lost */
4067 rs.is_ack_delayed = !!(flag & FLAG_ACK_MAYBE_DELAYED);
4068 tcp_rate_gen(sk, delivered, lost, is_sack_reneg, sack_state.rate);
4069 tcp_cong_control(sk, ack, delivered, flag, sack_state.rate);
4070 tcp_xmit_recovery(sk, rexmit);
4071 return 1;
4072
4073no_queue:
4074 /* If data was DSACKed, see if we can undo a cwnd reduction. */
4075 if (flag & FLAG_DSACKING_ACK) {
4076 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
4077 &rexmit);
4078 tcp_newly_delivered(sk, delivered, flag);
4079 }
4080 /* If this ack opens up a zero window, clear backoff. It was
4081 * being used to time the probes, and is probably far higher than
4082 * it needs to be for normal retransmission.
4083 */
4084 tcp_ack_probe(sk);
4085
4086 if (tp->tlp_high_seq)
4087 tcp_process_tlp_ack(sk, ack, flag);
4088 return 1;
4089
4090old_ack:
4091 /* If data was SACKed, tag it and see if we should send more data.
4092 * If data was DSACKed, see if we can undo a cwnd reduction.
4093 */
4094 if (TCP_SKB_CB(skb)->sacked) {
4095 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
4096 &sack_state);
4097 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
4098 &rexmit);
4099 tcp_newly_delivered(sk, delivered, flag);
4100 tcp_xmit_recovery(sk, rexmit);
4101 }
4102
4103 return 0;
4104}
4105
4106static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
4107 bool syn, struct tcp_fastopen_cookie *foc,
4108 bool exp_opt)
4109{
4110 /* Valid only in SYN or SYN-ACK with an even length. */
4111 if (!foc || !syn || len < 0 || (len & 1))
4112 return;
4113
4114 if (len >= TCP_FASTOPEN_COOKIE_MIN &&
4115 len <= TCP_FASTOPEN_COOKIE_MAX)
4116 memcpy(foc->val, cookie, len);
4117 else if (len != 0)
4118 len = -1;
4119 foc->len = len;
4120 foc->exp = exp_opt;
4121}
4122
4123static bool smc_parse_options(const struct tcphdr *th,
4124 struct tcp_options_received *opt_rx,
4125 const unsigned char *ptr,
4126 int opsize)
4127{
4128#if IS_ENABLED(CONFIG_SMC)
4129 if (static_branch_unlikely(&tcp_have_smc)) {
4130 if (th->syn && !(opsize & 1) &&
4131 opsize >= TCPOLEN_EXP_SMC_BASE &&
4132 get_unaligned_be32(ptr) == TCPOPT_SMC_MAGIC) {
4133 opt_rx->smc_ok = 1;
4134 return true;
4135 }
4136 }
4137#endif
4138 return false;
4139}
4140
4141/* Try to parse the MSS option from the TCP header. Return 0 on failure, clamped
4142 * value on success.
4143 */
4144u16 tcp_parse_mss_option(const struct tcphdr *th, u16 user_mss)
4145{
4146 const unsigned char *ptr = (const unsigned char *)(th + 1);
4147 int length = (th->doff * 4) - sizeof(struct tcphdr);
4148 u16 mss = 0;
4149
4150 while (length > 0) {
4151 int opcode = *ptr++;
4152 int opsize;
4153
4154 switch (opcode) {
4155 case TCPOPT_EOL:
4156 return mss;
4157 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
4158 length--;
4159 continue;
4160 default:
4161 if (length < 2)
4162 return mss;
4163 opsize = *ptr++;
4164 if (opsize < 2) /* "silly options" */
4165 return mss;
4166 if (opsize > length)
4167 return mss; /* fail on partial options */
4168 if (opcode == TCPOPT_MSS && opsize == TCPOLEN_MSS) {
4169 u16 in_mss = get_unaligned_be16(ptr);
4170
4171 if (in_mss) {
4172 if (user_mss && user_mss < in_mss)
4173 in_mss = user_mss;
4174 mss = in_mss;
4175 }
4176 }
4177 ptr += opsize - 2;
4178 length -= opsize;
4179 }
4180 }
4181 return mss;
4182}
4183EXPORT_SYMBOL_GPL(tcp_parse_mss_option);
4184
4185/* Look for tcp options. Normally only called on SYN and SYNACK packets.
4186 * But, this can also be called on packets in the established flow when
4187 * the fast version below fails.
4188 */
4189void tcp_parse_options(const struct net *net,
4190 const struct sk_buff *skb,
4191 struct tcp_options_received *opt_rx, int estab,
4192 struct tcp_fastopen_cookie *foc)
4193{
4194 const unsigned char *ptr;
4195 const struct tcphdr *th = tcp_hdr(skb);
4196 int length = (th->doff * 4) - sizeof(struct tcphdr);
4197
4198 ptr = (const unsigned char *)(th + 1);
4199 opt_rx->saw_tstamp = 0;
4200 opt_rx->saw_unknown = 0;
4201
4202 while (length > 0) {
4203 int opcode = *ptr++;
4204 int opsize;
4205
4206 switch (opcode) {
4207 case TCPOPT_EOL:
4208 return;
4209 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
4210 length--;
4211 continue;
4212 default:
4213 if (length < 2)
4214 return;
4215 opsize = *ptr++;
4216 if (opsize < 2) /* "silly options" */
4217 return;
4218 if (opsize > length)
4219 return; /* don't parse partial options */
4220 switch (opcode) {
4221 case TCPOPT_MSS:
4222 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
4223 u16 in_mss = get_unaligned_be16(ptr);
4224 if (in_mss) {
4225 if (opt_rx->user_mss &&
4226 opt_rx->user_mss < in_mss)
4227 in_mss = opt_rx->user_mss;
4228 opt_rx->mss_clamp = in_mss;
4229 }
4230 }
4231 break;
4232 case TCPOPT_WINDOW:
4233 if (opsize == TCPOLEN_WINDOW && th->syn &&
4234 !estab && READ_ONCE(net->ipv4.sysctl_tcp_window_scaling)) {
4235 __u8 snd_wscale = *(__u8 *)ptr;
4236 opt_rx->wscale_ok = 1;
4237 if (snd_wscale > TCP_MAX_WSCALE) {
4238 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n",
4239 __func__,
4240 snd_wscale,
4241 TCP_MAX_WSCALE);
4242 snd_wscale = TCP_MAX_WSCALE;
4243 }
4244 opt_rx->snd_wscale = snd_wscale;
4245 }
4246 break;
4247 case TCPOPT_TIMESTAMP:
4248 if ((opsize == TCPOLEN_TIMESTAMP) &&
4249 ((estab && opt_rx->tstamp_ok) ||
4250 (!estab && READ_ONCE(net->ipv4.sysctl_tcp_timestamps)))) {
4251 opt_rx->saw_tstamp = 1;
4252 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
4253 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
4254 }
4255 break;
4256 case TCPOPT_SACK_PERM:
4257 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
4258 !estab && READ_ONCE(net->ipv4.sysctl_tcp_sack)) {
4259 opt_rx->sack_ok = TCP_SACK_SEEN;
4260 tcp_sack_reset(opt_rx);
4261 }
4262 break;
4263
4264 case TCPOPT_SACK:
4265 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
4266 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
4267 opt_rx->sack_ok) {
4268 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
4269 }
4270 break;
4271#ifdef CONFIG_TCP_MD5SIG
4272 case TCPOPT_MD5SIG:
4273 /* The MD5 Hash has already been
4274 * checked (see tcp_v{4,6}_rcv()).
4275 */
4276 break;
4277#endif
4278#ifdef CONFIG_TCP_AO
4279 case TCPOPT_AO:
4280 /* TCP AO has already been checked
4281 * (see tcp_inbound_ao_hash()).
4282 */
4283 break;
4284#endif
4285 case TCPOPT_FASTOPEN:
4286 tcp_parse_fastopen_option(
4287 opsize - TCPOLEN_FASTOPEN_BASE,
4288 ptr, th->syn, foc, false);
4289 break;
4290
4291 case TCPOPT_EXP:
4292 /* Fast Open option shares code 254 using a
4293 * 16 bits magic number.
4294 */
4295 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
4296 get_unaligned_be16(ptr) ==
4297 TCPOPT_FASTOPEN_MAGIC) {
4298 tcp_parse_fastopen_option(opsize -
4299 TCPOLEN_EXP_FASTOPEN_BASE,
4300 ptr + 2, th->syn, foc, true);
4301 break;
4302 }
4303
4304 if (smc_parse_options(th, opt_rx, ptr, opsize))
4305 break;
4306
4307 opt_rx->saw_unknown = 1;
4308 break;
4309
4310 default:
4311 opt_rx->saw_unknown = 1;
4312 }
4313 ptr += opsize-2;
4314 length -= opsize;
4315 }
4316 }
4317}
4318EXPORT_SYMBOL(tcp_parse_options);
4319
4320static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
4321{
4322 const __be32 *ptr = (const __be32 *)(th + 1);
4323
4324 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
4325 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
4326 tp->rx_opt.saw_tstamp = 1;
4327 ++ptr;
4328 tp->rx_opt.rcv_tsval = ntohl(*ptr);
4329 ++ptr;
4330 if (*ptr)
4331 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
4332 else
4333 tp->rx_opt.rcv_tsecr = 0;
4334 return true;
4335 }
4336 return false;
4337}
4338
4339/* Fast parse options. This hopes to only see timestamps.
4340 * If it is wrong it falls back on tcp_parse_options().
4341 */
4342static bool tcp_fast_parse_options(const struct net *net,
4343 const struct sk_buff *skb,
4344 const struct tcphdr *th, struct tcp_sock *tp)
4345{
4346 /* In the spirit of fast parsing, compare doff directly to constant
4347 * values. Because equality is used, short doff can be ignored here.
4348 */
4349 if (th->doff == (sizeof(*th) / 4)) {
4350 tp->rx_opt.saw_tstamp = 0;
4351 return false;
4352 } else if (tp->rx_opt.tstamp_ok &&
4353 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
4354 if (tcp_parse_aligned_timestamp(tp, th))
4355 return true;
4356 }
4357
4358 tcp_parse_options(net, skb, &tp->rx_opt, 1, NULL);
4359 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
4360 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
4361
4362 return true;
4363}
4364
4365#if defined(CONFIG_TCP_MD5SIG) || defined(CONFIG_TCP_AO)
4366/*
4367 * Parse Signature options
4368 */
4369int tcp_do_parse_auth_options(const struct tcphdr *th,
4370 const u8 **md5_hash, const u8 **ao_hash)
4371{
4372 int length = (th->doff << 2) - sizeof(*th);
4373 const u8 *ptr = (const u8 *)(th + 1);
4374 unsigned int minlen = TCPOLEN_MD5SIG;
4375
4376 if (IS_ENABLED(CONFIG_TCP_AO))
4377 minlen = sizeof(struct tcp_ao_hdr) + 1;
4378
4379 *md5_hash = NULL;
4380 *ao_hash = NULL;
4381
4382 /* If not enough data remaining, we can short cut */
4383 while (length >= minlen) {
4384 int opcode = *ptr++;
4385 int opsize;
4386
4387 switch (opcode) {
4388 case TCPOPT_EOL:
4389 return 0;
4390 case TCPOPT_NOP:
4391 length--;
4392 continue;
4393 default:
4394 opsize = *ptr++;
4395 if (opsize < 2 || opsize > length)
4396 return -EINVAL;
4397 if (opcode == TCPOPT_MD5SIG) {
4398 if (opsize != TCPOLEN_MD5SIG)
4399 return -EINVAL;
4400 if (unlikely(*md5_hash || *ao_hash))
4401 return -EEXIST;
4402 *md5_hash = ptr;
4403 } else if (opcode == TCPOPT_AO) {
4404 if (opsize <= sizeof(struct tcp_ao_hdr))
4405 return -EINVAL;
4406 if (unlikely(*md5_hash || *ao_hash))
4407 return -EEXIST;
4408 *ao_hash = ptr;
4409 }
4410 }
4411 ptr += opsize - 2;
4412 length -= opsize;
4413 }
4414 return 0;
4415}
4416EXPORT_SYMBOL(tcp_do_parse_auth_options);
4417#endif
4418
4419/* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
4420 *
4421 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
4422 * it can pass through stack. So, the following predicate verifies that
4423 * this segment is not used for anything but congestion avoidance or
4424 * fast retransmit. Moreover, we even are able to eliminate most of such
4425 * second order effects, if we apply some small "replay" window (~RTO)
4426 * to timestamp space.
4427 *
4428 * All these measures still do not guarantee that we reject wrapped ACKs
4429 * on networks with high bandwidth, when sequence space is recycled fastly,
4430 * but it guarantees that such events will be very rare and do not affect
4431 * connection seriously. This doesn't look nice, but alas, PAWS is really
4432 * buggy extension.
4433 *
4434 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
4435 * states that events when retransmit arrives after original data are rare.
4436 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
4437 * the biggest problem on large power networks even with minor reordering.
4438 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
4439 * up to bandwidth of 18Gigabit/sec. 8) ]
4440 */
4441
4442/* Estimates max number of increments of remote peer TSval in
4443 * a replay window (based on our current RTO estimation).
4444 */
4445static u32 tcp_tsval_replay(const struct sock *sk)
4446{
4447 /* If we use usec TS resolution,
4448 * then expect the remote peer to use the same resolution.
4449 */
4450 if (tcp_sk(sk)->tcp_usec_ts)
4451 return inet_csk(sk)->icsk_rto * (USEC_PER_SEC / HZ);
4452
4453 /* RFC 7323 recommends a TSval clock between 1ms and 1sec.
4454 * We know that some OS (including old linux) can use 1200 Hz.
4455 */
4456 return inet_csk(sk)->icsk_rto * 1200 / HZ;
4457}
4458
4459static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
4460{
4461 const struct tcp_sock *tp = tcp_sk(sk);
4462 const struct tcphdr *th = tcp_hdr(skb);
4463 u32 seq = TCP_SKB_CB(skb)->seq;
4464 u32 ack = TCP_SKB_CB(skb)->ack_seq;
4465
4466 return /* 1. Pure ACK with correct sequence number. */
4467 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
4468
4469 /* 2. ... and duplicate ACK. */
4470 ack == tp->snd_una &&
4471
4472 /* 3. ... and does not update window. */
4473 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
4474
4475 /* 4. ... and sits in replay window. */
4476 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <=
4477 tcp_tsval_replay(sk);
4478}
4479
4480static inline bool tcp_paws_discard(const struct sock *sk,
4481 const struct sk_buff *skb)
4482{
4483 const struct tcp_sock *tp = tcp_sk(sk);
4484
4485 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4486 !tcp_disordered_ack(sk, skb);
4487}
4488
4489/* Check segment sequence number for validity.
4490 *
4491 * Segment controls are considered valid, if the segment
4492 * fits to the window after truncation to the window. Acceptability
4493 * of data (and SYN, FIN, of course) is checked separately.
4494 * See tcp_data_queue(), for example.
4495 *
4496 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4497 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4498 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4499 * (borrowed from freebsd)
4500 */
4501
4502static enum skb_drop_reason tcp_sequence(const struct tcp_sock *tp,
4503 u32 seq, u32 end_seq)
4504{
4505 if (before(end_seq, tp->rcv_wup))
4506 return SKB_DROP_REASON_TCP_OLD_SEQUENCE;
4507
4508 if (after(seq, tp->rcv_nxt + tcp_receive_window(tp)))
4509 return SKB_DROP_REASON_TCP_INVALID_SEQUENCE;
4510
4511 return SKB_NOT_DROPPED_YET;
4512}
4513
4514
4515void tcp_done_with_error(struct sock *sk, int err)
4516{
4517 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4518 WRITE_ONCE(sk->sk_err, err);
4519 smp_wmb();
4520
4521 tcp_write_queue_purge(sk);
4522 tcp_done(sk);
4523
4524 if (!sock_flag(sk, SOCK_DEAD))
4525 sk_error_report(sk);
4526}
4527EXPORT_SYMBOL(tcp_done_with_error);
4528
4529/* When we get a reset we do this. */
4530void tcp_reset(struct sock *sk, struct sk_buff *skb)
4531{
4532 int err;
4533
4534 trace_tcp_receive_reset(sk);
4535
4536 /* mptcp can't tell us to ignore reset pkts,
4537 * so just ignore the return value of mptcp_incoming_options().
4538 */
4539 if (sk_is_mptcp(sk))
4540 mptcp_incoming_options(sk, skb);
4541
4542 /* We want the right error as BSD sees it (and indeed as we do). */
4543 switch (sk->sk_state) {
4544 case TCP_SYN_SENT:
4545 err = ECONNREFUSED;
4546 break;
4547 case TCP_CLOSE_WAIT:
4548 err = EPIPE;
4549 break;
4550 case TCP_CLOSE:
4551 return;
4552 default:
4553 err = ECONNRESET;
4554 }
4555 tcp_done_with_error(sk, err);
4556}
4557
4558/*
4559 * Process the FIN bit. This now behaves as it is supposed to work
4560 * and the FIN takes effect when it is validly part of sequence
4561 * space. Not before when we get holes.
4562 *
4563 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4564 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4565 * TIME-WAIT)
4566 *
4567 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4568 * close and we go into CLOSING (and later onto TIME-WAIT)
4569 *
4570 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4571 */
4572void tcp_fin(struct sock *sk)
4573{
4574 struct tcp_sock *tp = tcp_sk(sk);
4575
4576 inet_csk_schedule_ack(sk);
4577
4578 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | RCV_SHUTDOWN);
4579 sock_set_flag(sk, SOCK_DONE);
4580
4581 switch (sk->sk_state) {
4582 case TCP_SYN_RECV:
4583 case TCP_ESTABLISHED:
4584 /* Move to CLOSE_WAIT */
4585 tcp_set_state(sk, TCP_CLOSE_WAIT);
4586 inet_csk_enter_pingpong_mode(sk);
4587 break;
4588
4589 case TCP_CLOSE_WAIT:
4590 case TCP_CLOSING:
4591 /* Received a retransmission of the FIN, do
4592 * nothing.
4593 */
4594 break;
4595 case TCP_LAST_ACK:
4596 /* RFC793: Remain in the LAST-ACK state. */
4597 break;
4598
4599 case TCP_FIN_WAIT1:
4600 /* This case occurs when a simultaneous close
4601 * happens, we must ack the received FIN and
4602 * enter the CLOSING state.
4603 */
4604 tcp_send_ack(sk);
4605 tcp_set_state(sk, TCP_CLOSING);
4606 break;
4607 case TCP_FIN_WAIT2:
4608 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4609 tcp_send_ack(sk);
4610 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4611 break;
4612 default:
4613 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4614 * cases we should never reach this piece of code.
4615 */
4616 pr_err("%s: Impossible, sk->sk_state=%d\n",
4617 __func__, sk->sk_state);
4618 break;
4619 }
4620
4621 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4622 * Probably, we should reset in this case. For now drop them.
4623 */
4624 skb_rbtree_purge(&tp->out_of_order_queue);
4625 if (tcp_is_sack(tp))
4626 tcp_sack_reset(&tp->rx_opt);
4627
4628 if (!sock_flag(sk, SOCK_DEAD)) {
4629 sk->sk_state_change(sk);
4630
4631 /* Do not send POLL_HUP for half duplex close. */
4632 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4633 sk->sk_state == TCP_CLOSE)
4634 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4635 else
4636 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4637 }
4638}
4639
4640static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4641 u32 end_seq)
4642{
4643 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4644 if (before(seq, sp->start_seq))
4645 sp->start_seq = seq;
4646 if (after(end_seq, sp->end_seq))
4647 sp->end_seq = end_seq;
4648 return true;
4649 }
4650 return false;
4651}
4652
4653static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4654{
4655 struct tcp_sock *tp = tcp_sk(sk);
4656
4657 if (tcp_is_sack(tp) && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_dsack)) {
4658 int mib_idx;
4659
4660 if (before(seq, tp->rcv_nxt))
4661 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4662 else
4663 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4664
4665 NET_INC_STATS(sock_net(sk), mib_idx);
4666
4667 tp->rx_opt.dsack = 1;
4668 tp->duplicate_sack[0].start_seq = seq;
4669 tp->duplicate_sack[0].end_seq = end_seq;
4670 }
4671}
4672
4673static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4674{
4675 struct tcp_sock *tp = tcp_sk(sk);
4676
4677 if (!tp->rx_opt.dsack)
4678 tcp_dsack_set(sk, seq, end_seq);
4679 else
4680 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4681}
4682
4683static void tcp_rcv_spurious_retrans(struct sock *sk, const struct sk_buff *skb)
4684{
4685 /* When the ACK path fails or drops most ACKs, the sender would
4686 * timeout and spuriously retransmit the same segment repeatedly.
4687 * If it seems our ACKs are not reaching the other side,
4688 * based on receiving a duplicate data segment with new flowlabel
4689 * (suggesting the sender suffered an RTO), and we are not already
4690 * repathing due to our own RTO, then rehash the socket to repath our
4691 * packets.
4692 */
4693#if IS_ENABLED(CONFIG_IPV6)
4694 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss &&
4695 skb->protocol == htons(ETH_P_IPV6) &&
4696 (tcp_sk(sk)->inet_conn.icsk_ack.lrcv_flowlabel !=
4697 ntohl(ip6_flowlabel(ipv6_hdr(skb)))) &&
4698 sk_rethink_txhash(sk))
4699 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDUPLICATEDATAREHASH);
4700
4701 /* Save last flowlabel after a spurious retrans. */
4702 tcp_save_lrcv_flowlabel(sk, skb);
4703#endif
4704}
4705
4706static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4707{
4708 struct tcp_sock *tp = tcp_sk(sk);
4709
4710 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4711 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4712 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4713 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
4714
4715 if (tcp_is_sack(tp) && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_dsack)) {
4716 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4717
4718 tcp_rcv_spurious_retrans(sk, skb);
4719 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4720 end_seq = tp->rcv_nxt;
4721 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4722 }
4723 }
4724
4725 tcp_send_ack(sk);
4726}
4727
4728/* These routines update the SACK block as out-of-order packets arrive or
4729 * in-order packets close up the sequence space.
4730 */
4731static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4732{
4733 int this_sack;
4734 struct tcp_sack_block *sp = &tp->selective_acks[0];
4735 struct tcp_sack_block *swalk = sp + 1;
4736
4737 /* See if the recent change to the first SACK eats into
4738 * or hits the sequence space of other SACK blocks, if so coalesce.
4739 */
4740 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4741 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4742 int i;
4743
4744 /* Zap SWALK, by moving every further SACK up by one slot.
4745 * Decrease num_sacks.
4746 */
4747 tp->rx_opt.num_sacks--;
4748 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4749 sp[i] = sp[i + 1];
4750 continue;
4751 }
4752 this_sack++;
4753 swalk++;
4754 }
4755}
4756
4757void tcp_sack_compress_send_ack(struct sock *sk)
4758{
4759 struct tcp_sock *tp = tcp_sk(sk);
4760
4761 if (!tp->compressed_ack)
4762 return;
4763
4764 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
4765 __sock_put(sk);
4766
4767 /* Since we have to send one ack finally,
4768 * substract one from tp->compressed_ack to keep
4769 * LINUX_MIB_TCPACKCOMPRESSED accurate.
4770 */
4771 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
4772 tp->compressed_ack - 1);
4773
4774 tp->compressed_ack = 0;
4775 tcp_send_ack(sk);
4776}
4777
4778/* Reasonable amount of sack blocks included in TCP SACK option
4779 * The max is 4, but this becomes 3 if TCP timestamps are there.
4780 * Given that SACK packets might be lost, be conservative and use 2.
4781 */
4782#define TCP_SACK_BLOCKS_EXPECTED 2
4783
4784static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4785{
4786 struct tcp_sock *tp = tcp_sk(sk);
4787 struct tcp_sack_block *sp = &tp->selective_acks[0];
4788 int cur_sacks = tp->rx_opt.num_sacks;
4789 int this_sack;
4790
4791 if (!cur_sacks)
4792 goto new_sack;
4793
4794 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4795 if (tcp_sack_extend(sp, seq, end_seq)) {
4796 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED)
4797 tcp_sack_compress_send_ack(sk);
4798 /* Rotate this_sack to the first one. */
4799 for (; this_sack > 0; this_sack--, sp--)
4800 swap(*sp, *(sp - 1));
4801 if (cur_sacks > 1)
4802 tcp_sack_maybe_coalesce(tp);
4803 return;
4804 }
4805 }
4806
4807 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED)
4808 tcp_sack_compress_send_ack(sk);
4809
4810 /* Could not find an adjacent existing SACK, build a new one,
4811 * put it at the front, and shift everyone else down. We
4812 * always know there is at least one SACK present already here.
4813 *
4814 * If the sack array is full, forget about the last one.
4815 */
4816 if (this_sack >= TCP_NUM_SACKS) {
4817 this_sack--;
4818 tp->rx_opt.num_sacks--;
4819 sp--;
4820 }
4821 for (; this_sack > 0; this_sack--, sp--)
4822 *sp = *(sp - 1);
4823
4824new_sack:
4825 /* Build the new head SACK, and we're done. */
4826 sp->start_seq = seq;
4827 sp->end_seq = end_seq;
4828 tp->rx_opt.num_sacks++;
4829}
4830
4831/* RCV.NXT advances, some SACKs should be eaten. */
4832
4833static void tcp_sack_remove(struct tcp_sock *tp)
4834{
4835 struct tcp_sack_block *sp = &tp->selective_acks[0];
4836 int num_sacks = tp->rx_opt.num_sacks;
4837 int this_sack;
4838
4839 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4840 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4841 tp->rx_opt.num_sacks = 0;
4842 return;
4843 }
4844
4845 for (this_sack = 0; this_sack < num_sacks;) {
4846 /* Check if the start of the sack is covered by RCV.NXT. */
4847 if (!before(tp->rcv_nxt, sp->start_seq)) {
4848 int i;
4849
4850 /* RCV.NXT must cover all the block! */
4851 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4852
4853 /* Zap this SACK, by moving forward any other SACKS. */
4854 for (i = this_sack+1; i < num_sacks; i++)
4855 tp->selective_acks[i-1] = tp->selective_acks[i];
4856 num_sacks--;
4857 continue;
4858 }
4859 this_sack++;
4860 sp++;
4861 }
4862 tp->rx_opt.num_sacks = num_sacks;
4863}
4864
4865/**
4866 * tcp_try_coalesce - try to merge skb to prior one
4867 * @sk: socket
4868 * @to: prior buffer
4869 * @from: buffer to add in queue
4870 * @fragstolen: pointer to boolean
4871 *
4872 * Before queueing skb @from after @to, try to merge them
4873 * to reduce overall memory use and queue lengths, if cost is small.
4874 * Packets in ofo or receive queues can stay a long time.
4875 * Better try to coalesce them right now to avoid future collapses.
4876 * Returns true if caller should free @from instead of queueing it
4877 */
4878static bool tcp_try_coalesce(struct sock *sk,
4879 struct sk_buff *to,
4880 struct sk_buff *from,
4881 bool *fragstolen)
4882{
4883 int delta;
4884
4885 *fragstolen = false;
4886
4887 /* Its possible this segment overlaps with prior segment in queue */
4888 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4889 return false;
4890
4891 if (!tcp_skb_can_collapse_rx(to, from))
4892 return false;
4893
4894 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4895 return false;
4896
4897 atomic_add(delta, &sk->sk_rmem_alloc);
4898 sk_mem_charge(sk, delta);
4899 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4900 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4901 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4902 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4903
4904 if (TCP_SKB_CB(from)->has_rxtstamp) {
4905 TCP_SKB_CB(to)->has_rxtstamp = true;
4906 to->tstamp = from->tstamp;
4907 skb_hwtstamps(to)->hwtstamp = skb_hwtstamps(from)->hwtstamp;
4908 }
4909
4910 return true;
4911}
4912
4913static bool tcp_ooo_try_coalesce(struct sock *sk,
4914 struct sk_buff *to,
4915 struct sk_buff *from,
4916 bool *fragstolen)
4917{
4918 bool res = tcp_try_coalesce(sk, to, from, fragstolen);
4919
4920 /* In case tcp_drop_reason() is called later, update to->gso_segs */
4921 if (res) {
4922 u32 gso_segs = max_t(u16, 1, skb_shinfo(to)->gso_segs) +
4923 max_t(u16, 1, skb_shinfo(from)->gso_segs);
4924
4925 skb_shinfo(to)->gso_segs = min_t(u32, gso_segs, 0xFFFF);
4926 }
4927 return res;
4928}
4929
4930noinline_for_tracing static void
4931tcp_drop_reason(struct sock *sk, struct sk_buff *skb, enum skb_drop_reason reason)
4932{
4933 sk_drops_add(sk, skb);
4934 sk_skb_reason_drop(sk, skb, reason);
4935}
4936
4937/* This one checks to see if we can put data from the
4938 * out_of_order queue into the receive_queue.
4939 */
4940static void tcp_ofo_queue(struct sock *sk)
4941{
4942 struct tcp_sock *tp = tcp_sk(sk);
4943 __u32 dsack_high = tp->rcv_nxt;
4944 bool fin, fragstolen, eaten;
4945 struct sk_buff *skb, *tail;
4946 struct rb_node *p;
4947
4948 p = rb_first(&tp->out_of_order_queue);
4949 while (p) {
4950 skb = rb_to_skb(p);
4951 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4952 break;
4953
4954 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4955 __u32 dsack = dsack_high;
4956 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4957 dsack_high = TCP_SKB_CB(skb)->end_seq;
4958 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4959 }
4960 p = rb_next(p);
4961 rb_erase(&skb->rbnode, &tp->out_of_order_queue);
4962
4963 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) {
4964 tcp_drop_reason(sk, skb, SKB_DROP_REASON_TCP_OFO_DROP);
4965 continue;
4966 }
4967
4968 tail = skb_peek_tail(&sk->sk_receive_queue);
4969 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4970 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4971 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
4972 if (!eaten)
4973 tcp_add_receive_queue(sk, skb);
4974 else
4975 kfree_skb_partial(skb, fragstolen);
4976
4977 if (unlikely(fin)) {
4978 tcp_fin(sk);
4979 /* tcp_fin() purges tp->out_of_order_queue,
4980 * so we must end this loop right now.
4981 */
4982 break;
4983 }
4984 }
4985}
4986
4987static bool tcp_prune_ofo_queue(struct sock *sk, const struct sk_buff *in_skb);
4988static int tcp_prune_queue(struct sock *sk, const struct sk_buff *in_skb);
4989
4990static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4991 unsigned int size)
4992{
4993 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4994 !sk_rmem_schedule(sk, skb, size)) {
4995
4996 if (tcp_prune_queue(sk, skb) < 0)
4997 return -1;
4998
4999 while (!sk_rmem_schedule(sk, skb, size)) {
5000 if (!tcp_prune_ofo_queue(sk, skb))
5001 return -1;
5002 }
5003 }
5004 return 0;
5005}
5006
5007static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
5008{
5009 struct tcp_sock *tp = tcp_sk(sk);
5010 struct rb_node **p, *parent;
5011 struct sk_buff *skb1;
5012 u32 seq, end_seq;
5013 bool fragstolen;
5014
5015 tcp_save_lrcv_flowlabel(sk, skb);
5016 tcp_ecn_check_ce(sk, skb);
5017
5018 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
5019 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP);
5020 sk->sk_data_ready(sk);
5021 tcp_drop_reason(sk, skb, SKB_DROP_REASON_PROTO_MEM);
5022 return;
5023 }
5024
5025 /* Disable header prediction. */
5026 tp->pred_flags = 0;
5027 inet_csk_schedule_ack(sk);
5028
5029 tp->rcv_ooopack += max_t(u16, 1, skb_shinfo(skb)->gso_segs);
5030 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
5031 seq = TCP_SKB_CB(skb)->seq;
5032 end_seq = TCP_SKB_CB(skb)->end_seq;
5033
5034 p = &tp->out_of_order_queue.rb_node;
5035 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5036 /* Initial out of order segment, build 1 SACK. */
5037 if (tcp_is_sack(tp)) {
5038 tp->rx_opt.num_sacks = 1;
5039 tp->selective_acks[0].start_seq = seq;
5040 tp->selective_acks[0].end_seq = end_seq;
5041 }
5042 rb_link_node(&skb->rbnode, NULL, p);
5043 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
5044 tp->ooo_last_skb = skb;
5045 goto end;
5046 }
5047
5048 /* In the typical case, we are adding an skb to the end of the list.
5049 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
5050 */
5051 if (tcp_ooo_try_coalesce(sk, tp->ooo_last_skb,
5052 skb, &fragstolen)) {
5053coalesce_done:
5054 /* For non sack flows, do not grow window to force DUPACK
5055 * and trigger fast retransmit.
5056 */
5057 if (tcp_is_sack(tp))
5058 tcp_grow_window(sk, skb, true);
5059 kfree_skb_partial(skb, fragstolen);
5060 skb = NULL;
5061 goto add_sack;
5062 }
5063 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
5064 if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) {
5065 parent = &tp->ooo_last_skb->rbnode;
5066 p = &parent->rb_right;
5067 goto insert;
5068 }
5069
5070 /* Find place to insert this segment. Handle overlaps on the way. */
5071 parent = NULL;
5072 while (*p) {
5073 parent = *p;
5074 skb1 = rb_to_skb(parent);
5075 if (before(seq, TCP_SKB_CB(skb1)->seq)) {
5076 p = &parent->rb_left;
5077 continue;
5078 }
5079 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) {
5080 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
5081 /* All the bits are present. Drop. */
5082 NET_INC_STATS(sock_net(sk),
5083 LINUX_MIB_TCPOFOMERGE);
5084 tcp_drop_reason(sk, skb,
5085 SKB_DROP_REASON_TCP_OFOMERGE);
5086 skb = NULL;
5087 tcp_dsack_set(sk, seq, end_seq);
5088 goto add_sack;
5089 }
5090 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
5091 /* Partial overlap. */
5092 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq);
5093 } else {
5094 /* skb's seq == skb1's seq and skb covers skb1.
5095 * Replace skb1 with skb.
5096 */
5097 rb_replace_node(&skb1->rbnode, &skb->rbnode,
5098 &tp->out_of_order_queue);
5099 tcp_dsack_extend(sk,
5100 TCP_SKB_CB(skb1)->seq,
5101 TCP_SKB_CB(skb1)->end_seq);
5102 NET_INC_STATS(sock_net(sk),
5103 LINUX_MIB_TCPOFOMERGE);
5104 tcp_drop_reason(sk, skb1,
5105 SKB_DROP_REASON_TCP_OFOMERGE);
5106 goto merge_right;
5107 }
5108 } else if (tcp_ooo_try_coalesce(sk, skb1,
5109 skb, &fragstolen)) {
5110 goto coalesce_done;
5111 }
5112 p = &parent->rb_right;
5113 }
5114insert:
5115 /* Insert segment into RB tree. */
5116 rb_link_node(&skb->rbnode, parent, p);
5117 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
5118
5119merge_right:
5120 /* Remove other segments covered by skb. */
5121 while ((skb1 = skb_rb_next(skb)) != NULL) {
5122 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
5123 break;
5124 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
5125 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
5126 end_seq);
5127 break;
5128 }
5129 rb_erase(&skb1->rbnode, &tp->out_of_order_queue);
5130 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
5131 TCP_SKB_CB(skb1)->end_seq);
5132 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
5133 tcp_drop_reason(sk, skb1, SKB_DROP_REASON_TCP_OFOMERGE);
5134 }
5135 /* If there is no skb after us, we are the last_skb ! */
5136 if (!skb1)
5137 tp->ooo_last_skb = skb;
5138
5139add_sack:
5140 if (tcp_is_sack(tp))
5141 tcp_sack_new_ofo_skb(sk, seq, end_seq);
5142end:
5143 if (skb) {
5144 /* For non sack flows, do not grow window to force DUPACK
5145 * and trigger fast retransmit.
5146 */
5147 if (tcp_is_sack(tp))
5148 tcp_grow_window(sk, skb, false);
5149 skb_condense(skb);
5150 skb_set_owner_r(skb, sk);
5151 }
5152}
5153
5154static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb,
5155 bool *fragstolen)
5156{
5157 int eaten;
5158 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
5159
5160 eaten = (tail &&
5161 tcp_try_coalesce(sk, tail,
5162 skb, fragstolen)) ? 1 : 0;
5163 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
5164 if (!eaten) {
5165 tcp_add_receive_queue(sk, skb);
5166 skb_set_owner_r(skb, sk);
5167 }
5168 return eaten;
5169}
5170
5171int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
5172{
5173 struct sk_buff *skb;
5174 int err = -ENOMEM;
5175 int data_len = 0;
5176 bool fragstolen;
5177
5178 if (size == 0)
5179 return 0;
5180
5181 if (size > PAGE_SIZE) {
5182 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
5183
5184 data_len = npages << PAGE_SHIFT;
5185 size = data_len + (size & ~PAGE_MASK);
5186 }
5187 skb = alloc_skb_with_frags(size - data_len, data_len,
5188 PAGE_ALLOC_COSTLY_ORDER,
5189 &err, sk->sk_allocation);
5190 if (!skb)
5191 goto err;
5192
5193 skb_put(skb, size - data_len);
5194 skb->data_len = data_len;
5195 skb->len = size;
5196
5197 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
5198 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
5199 goto err_free;
5200 }
5201
5202 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
5203 if (err)
5204 goto err_free;
5205
5206 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
5207 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
5208 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
5209
5210 if (tcp_queue_rcv(sk, skb, &fragstolen)) {
5211 WARN_ON_ONCE(fragstolen); /* should not happen */
5212 __kfree_skb(skb);
5213 }
5214 return size;
5215
5216err_free:
5217 kfree_skb(skb);
5218err:
5219 return err;
5220
5221}
5222
5223void tcp_data_ready(struct sock *sk)
5224{
5225 if (tcp_epollin_ready(sk, sk->sk_rcvlowat) || sock_flag(sk, SOCK_DONE))
5226 sk->sk_data_ready(sk);
5227}
5228
5229static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
5230{
5231 struct tcp_sock *tp = tcp_sk(sk);
5232 enum skb_drop_reason reason;
5233 bool fragstolen;
5234 int eaten;
5235
5236 /* If a subflow has been reset, the packet should not continue
5237 * to be processed, drop the packet.
5238 */
5239 if (sk_is_mptcp(sk) && !mptcp_incoming_options(sk, skb)) {
5240 __kfree_skb(skb);
5241 return;
5242 }
5243
5244 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
5245 __kfree_skb(skb);
5246 return;
5247 }
5248 tcp_cleanup_skb(skb);
5249 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
5250
5251 reason = SKB_DROP_REASON_NOT_SPECIFIED;
5252 tp->rx_opt.dsack = 0;
5253
5254 /* Queue data for delivery to the user.
5255 * Packets in sequence go to the receive queue.
5256 * Out of sequence packets to the out_of_order_queue.
5257 */
5258 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
5259 if (tcp_receive_window(tp) == 0) {
5260 /* Some stacks are known to send bare FIN packets
5261 * in a loop even if we send RWIN 0 in our ACK.
5262 * Accepting this FIN does not hurt memory pressure
5263 * because the FIN flag will simply be merged to the
5264 * receive queue tail skb in most cases.
5265 */
5266 if (!skb->len &&
5267 (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
5268 goto queue_and_out;
5269
5270 reason = SKB_DROP_REASON_TCP_ZEROWINDOW;
5271 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
5272 goto out_of_window;
5273 }
5274
5275 /* Ok. In sequence. In window. */
5276queue_and_out:
5277 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
5278 /* TODO: maybe ratelimit these WIN 0 ACK ? */
5279 inet_csk(sk)->icsk_ack.pending |=
5280 (ICSK_ACK_NOMEM | ICSK_ACK_NOW);
5281 inet_csk_schedule_ack(sk);
5282 sk->sk_data_ready(sk);
5283
5284 if (skb_queue_len(&sk->sk_receive_queue) && skb->len) {
5285 reason = SKB_DROP_REASON_PROTO_MEM;
5286 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
5287 goto drop;
5288 }
5289 sk_forced_mem_schedule(sk, skb->truesize);
5290 }
5291
5292 eaten = tcp_queue_rcv(sk, skb, &fragstolen);
5293 if (skb->len)
5294 tcp_event_data_recv(sk, skb);
5295 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
5296 tcp_fin(sk);
5297
5298 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5299 tcp_ofo_queue(sk);
5300
5301 /* RFC5681. 4.2. SHOULD send immediate ACK, when
5302 * gap in queue is filled.
5303 */
5304 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
5305 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
5306 }
5307
5308 if (tp->rx_opt.num_sacks)
5309 tcp_sack_remove(tp);
5310
5311 tcp_fast_path_check(sk);
5312
5313 if (eaten > 0)
5314 kfree_skb_partial(skb, fragstolen);
5315 if (!sock_flag(sk, SOCK_DEAD))
5316 tcp_data_ready(sk);
5317 return;
5318 }
5319
5320 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
5321 tcp_rcv_spurious_retrans(sk, skb);
5322 /* A retransmit, 2nd most common case. Force an immediate ack. */
5323 reason = SKB_DROP_REASON_TCP_OLD_DATA;
5324 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
5325 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
5326
5327out_of_window:
5328 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
5329 inet_csk_schedule_ack(sk);
5330drop:
5331 tcp_drop_reason(sk, skb, reason);
5332 return;
5333 }
5334
5335 /* Out of window. F.e. zero window probe. */
5336 if (!before(TCP_SKB_CB(skb)->seq,
5337 tp->rcv_nxt + tcp_receive_window(tp))) {
5338 reason = SKB_DROP_REASON_TCP_OVERWINDOW;
5339 goto out_of_window;
5340 }
5341
5342 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
5343 /* Partial packet, seq < rcv_next < end_seq */
5344 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
5345
5346 /* If window is closed, drop tail of packet. But after
5347 * remembering D-SACK for its head made in previous line.
5348 */
5349 if (!tcp_receive_window(tp)) {
5350 reason = SKB_DROP_REASON_TCP_ZEROWINDOW;
5351 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
5352 goto out_of_window;
5353 }
5354 goto queue_and_out;
5355 }
5356
5357 tcp_data_queue_ofo(sk, skb);
5358}
5359
5360static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list)
5361{
5362 if (list)
5363 return !skb_queue_is_last(list, skb) ? skb->next : NULL;
5364
5365 return skb_rb_next(skb);
5366}
5367
5368static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
5369 struct sk_buff_head *list,
5370 struct rb_root *root)
5371{
5372 struct sk_buff *next = tcp_skb_next(skb, list);
5373
5374 if (list)
5375 __skb_unlink(skb, list);
5376 else
5377 rb_erase(&skb->rbnode, root);
5378
5379 __kfree_skb(skb);
5380 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
5381
5382 return next;
5383}
5384
5385/* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
5386void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb)
5387{
5388 struct rb_node **p = &root->rb_node;
5389 struct rb_node *parent = NULL;
5390 struct sk_buff *skb1;
5391
5392 while (*p) {
5393 parent = *p;
5394 skb1 = rb_to_skb(parent);
5395 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq))
5396 p = &parent->rb_left;
5397 else
5398 p = &parent->rb_right;
5399 }
5400 rb_link_node(&skb->rbnode, parent, p);
5401 rb_insert_color(&skb->rbnode, root);
5402}
5403
5404/* Collapse contiguous sequence of skbs head..tail with
5405 * sequence numbers start..end.
5406 *
5407 * If tail is NULL, this means until the end of the queue.
5408 *
5409 * Segments with FIN/SYN are not collapsed (only because this
5410 * simplifies code)
5411 */
5412static void
5413tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root,
5414 struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end)
5415{
5416 struct sk_buff *skb = head, *n;
5417 struct sk_buff_head tmp;
5418 bool end_of_skbs;
5419
5420 /* First, check that queue is collapsible and find
5421 * the point where collapsing can be useful.
5422 */
5423restart:
5424 for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) {
5425 n = tcp_skb_next(skb, list);
5426
5427 if (!skb_frags_readable(skb))
5428 goto skip_this;
5429
5430 /* No new bits? It is possible on ofo queue. */
5431 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
5432 skb = tcp_collapse_one(sk, skb, list, root);
5433 if (!skb)
5434 break;
5435 goto restart;
5436 }
5437
5438 /* The first skb to collapse is:
5439 * - not SYN/FIN and
5440 * - bloated or contains data before "start" or
5441 * overlaps to the next one and mptcp allow collapsing.
5442 */
5443 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
5444 (tcp_win_from_space(sk, skb->truesize) > skb->len ||
5445 before(TCP_SKB_CB(skb)->seq, start))) {
5446 end_of_skbs = false;
5447 break;
5448 }
5449
5450 if (n && n != tail && skb_frags_readable(n) &&
5451 tcp_skb_can_collapse_rx(skb, n) &&
5452 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) {
5453 end_of_skbs = false;
5454 break;
5455 }
5456
5457skip_this:
5458 /* Decided to skip this, advance start seq. */
5459 start = TCP_SKB_CB(skb)->end_seq;
5460 }
5461 if (end_of_skbs ||
5462 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) ||
5463 !skb_frags_readable(skb))
5464 return;
5465
5466 __skb_queue_head_init(&tmp);
5467
5468 while (before(start, end)) {
5469 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
5470 struct sk_buff *nskb;
5471
5472 nskb = alloc_skb(copy, GFP_ATOMIC);
5473 if (!nskb)
5474 break;
5475
5476 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
5477 skb_copy_decrypted(nskb, skb);
5478 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
5479 if (list)
5480 __skb_queue_before(list, skb, nskb);
5481 else
5482 __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */
5483 skb_set_owner_r(nskb, sk);
5484 mptcp_skb_ext_move(nskb, skb);
5485
5486 /* Copy data, releasing collapsed skbs. */
5487 while (copy > 0) {
5488 int offset = start - TCP_SKB_CB(skb)->seq;
5489 int size = TCP_SKB_CB(skb)->end_seq - start;
5490
5491 BUG_ON(offset < 0);
5492 if (size > 0) {
5493 size = min(copy, size);
5494 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
5495 BUG();
5496 TCP_SKB_CB(nskb)->end_seq += size;
5497 copy -= size;
5498 start += size;
5499 }
5500 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
5501 skb = tcp_collapse_one(sk, skb, list, root);
5502 if (!skb ||
5503 skb == tail ||
5504 !tcp_skb_can_collapse_rx(nskb, skb) ||
5505 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) ||
5506 !skb_frags_readable(skb))
5507 goto end;
5508 }
5509 }
5510 }
5511end:
5512 skb_queue_walk_safe(&tmp, skb, n)
5513 tcp_rbtree_insert(root, skb);
5514}
5515
5516/* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
5517 * and tcp_collapse() them until all the queue is collapsed.
5518 */
5519static void tcp_collapse_ofo_queue(struct sock *sk)
5520{
5521 struct tcp_sock *tp = tcp_sk(sk);
5522 u32 range_truesize, sum_tiny = 0;
5523 struct sk_buff *skb, *head;
5524 u32 start, end;
5525
5526 skb = skb_rb_first(&tp->out_of_order_queue);
5527new_range:
5528 if (!skb) {
5529 tp->ooo_last_skb = skb_rb_last(&tp->out_of_order_queue);
5530 return;
5531 }
5532 start = TCP_SKB_CB(skb)->seq;
5533 end = TCP_SKB_CB(skb)->end_seq;
5534 range_truesize = skb->truesize;
5535
5536 for (head = skb;;) {
5537 skb = skb_rb_next(skb);
5538
5539 /* Range is terminated when we see a gap or when
5540 * we are at the queue end.
5541 */
5542 if (!skb ||
5543 after(TCP_SKB_CB(skb)->seq, end) ||
5544 before(TCP_SKB_CB(skb)->end_seq, start)) {
5545 /* Do not attempt collapsing tiny skbs */
5546 if (range_truesize != head->truesize ||
5547 end - start >= SKB_WITH_OVERHEAD(PAGE_SIZE)) {
5548 tcp_collapse(sk, NULL, &tp->out_of_order_queue,
5549 head, skb, start, end);
5550 } else {
5551 sum_tiny += range_truesize;
5552 if (sum_tiny > sk->sk_rcvbuf >> 3)
5553 return;
5554 }
5555 goto new_range;
5556 }
5557
5558 range_truesize += skb->truesize;
5559 if (unlikely(before(TCP_SKB_CB(skb)->seq, start)))
5560 start = TCP_SKB_CB(skb)->seq;
5561 if (after(TCP_SKB_CB(skb)->end_seq, end))
5562 end = TCP_SKB_CB(skb)->end_seq;
5563 }
5564}
5565
5566/*
5567 * Clean the out-of-order queue to make room.
5568 * We drop high sequences packets to :
5569 * 1) Let a chance for holes to be filled.
5570 * This means we do not drop packets from ooo queue if their sequence
5571 * is before incoming packet sequence.
5572 * 2) not add too big latencies if thousands of packets sit there.
5573 * (But if application shrinks SO_RCVBUF, we could still end up
5574 * freeing whole queue here)
5575 * 3) Drop at least 12.5 % of sk_rcvbuf to avoid malicious attacks.
5576 *
5577 * Return true if queue has shrunk.
5578 */
5579static bool tcp_prune_ofo_queue(struct sock *sk, const struct sk_buff *in_skb)
5580{
5581 struct tcp_sock *tp = tcp_sk(sk);
5582 struct rb_node *node, *prev;
5583 bool pruned = false;
5584 int goal;
5585
5586 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
5587 return false;
5588
5589 goal = sk->sk_rcvbuf >> 3;
5590 node = &tp->ooo_last_skb->rbnode;
5591
5592 do {
5593 struct sk_buff *skb = rb_to_skb(node);
5594
5595 /* If incoming skb would land last in ofo queue, stop pruning. */
5596 if (after(TCP_SKB_CB(in_skb)->seq, TCP_SKB_CB(skb)->seq))
5597 break;
5598 pruned = true;
5599 prev = rb_prev(node);
5600 rb_erase(node, &tp->out_of_order_queue);
5601 goal -= skb->truesize;
5602 tcp_drop_reason(sk, skb, SKB_DROP_REASON_TCP_OFO_QUEUE_PRUNE);
5603 tp->ooo_last_skb = rb_to_skb(prev);
5604 if (!prev || goal <= 0) {
5605 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf &&
5606 !tcp_under_memory_pressure(sk))
5607 break;
5608 goal = sk->sk_rcvbuf >> 3;
5609 }
5610 node = prev;
5611 } while (node);
5612
5613 if (pruned) {
5614 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED);
5615 /* Reset SACK state. A conforming SACK implementation will
5616 * do the same at a timeout based retransmit. When a connection
5617 * is in a sad state like this, we care only about integrity
5618 * of the connection not performance.
5619 */
5620 if (tp->rx_opt.sack_ok)
5621 tcp_sack_reset(&tp->rx_opt);
5622 }
5623 return pruned;
5624}
5625
5626/* Reduce allocated memory if we can, trying to get
5627 * the socket within its memory limits again.
5628 *
5629 * Return less than zero if we should start dropping frames
5630 * until the socket owning process reads some of the data
5631 * to stabilize the situation.
5632 */
5633static int tcp_prune_queue(struct sock *sk, const struct sk_buff *in_skb)
5634{
5635 struct tcp_sock *tp = tcp_sk(sk);
5636
5637 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED);
5638
5639 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
5640 tcp_clamp_window(sk);
5641 else if (tcp_under_memory_pressure(sk))
5642 tcp_adjust_rcv_ssthresh(sk);
5643
5644 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5645 return 0;
5646
5647 tcp_collapse_ofo_queue(sk);
5648 if (!skb_queue_empty(&sk->sk_receive_queue))
5649 tcp_collapse(sk, &sk->sk_receive_queue, NULL,
5650 skb_peek(&sk->sk_receive_queue),
5651 NULL,
5652 tp->copied_seq, tp->rcv_nxt);
5653
5654 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5655 return 0;
5656
5657 /* Collapsing did not help, destructive actions follow.
5658 * This must not ever occur. */
5659
5660 tcp_prune_ofo_queue(sk, in_skb);
5661
5662 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5663 return 0;
5664
5665 /* If we are really being abused, tell the caller to silently
5666 * drop receive data on the floor. It will get retransmitted
5667 * and hopefully then we'll have sufficient space.
5668 */
5669 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED);
5670
5671 /* Massive buffer overcommit. */
5672 tp->pred_flags = 0;
5673 return -1;
5674}
5675
5676static bool tcp_should_expand_sndbuf(struct sock *sk)
5677{
5678 const struct tcp_sock *tp = tcp_sk(sk);
5679
5680 /* If the user specified a specific send buffer setting, do
5681 * not modify it.
5682 */
5683 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
5684 return false;
5685
5686 /* If we are under global TCP memory pressure, do not expand. */
5687 if (tcp_under_memory_pressure(sk)) {
5688 int unused_mem = sk_unused_reserved_mem(sk);
5689
5690 /* Adjust sndbuf according to reserved mem. But make sure
5691 * it never goes below SOCK_MIN_SNDBUF.
5692 * See sk_stream_moderate_sndbuf() for more details.
5693 */
5694 if (unused_mem > SOCK_MIN_SNDBUF)
5695 WRITE_ONCE(sk->sk_sndbuf, unused_mem);
5696
5697 return false;
5698 }
5699
5700 /* If we are under soft global TCP memory pressure, do not expand. */
5701 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
5702 return false;
5703
5704 /* If we filled the congestion window, do not expand. */
5705 if (tcp_packets_in_flight(tp) >= tcp_snd_cwnd(tp))
5706 return false;
5707
5708 return true;
5709}
5710
5711static void tcp_new_space(struct sock *sk)
5712{
5713 struct tcp_sock *tp = tcp_sk(sk);
5714
5715 if (tcp_should_expand_sndbuf(sk)) {
5716 tcp_sndbuf_expand(sk);
5717 tp->snd_cwnd_stamp = tcp_jiffies32;
5718 }
5719
5720 INDIRECT_CALL_1(sk->sk_write_space, sk_stream_write_space, sk);
5721}
5722
5723/* Caller made space either from:
5724 * 1) Freeing skbs in rtx queues (after tp->snd_una has advanced)
5725 * 2) Sent skbs from output queue (and thus advancing tp->snd_nxt)
5726 *
5727 * We might be able to generate EPOLLOUT to the application if:
5728 * 1) Space consumed in output/rtx queues is below sk->sk_sndbuf/2
5729 * 2) notsent amount (tp->write_seq - tp->snd_nxt) became
5730 * small enough that tcp_stream_memory_free() decides it
5731 * is time to generate EPOLLOUT.
5732 */
5733void tcp_check_space(struct sock *sk)
5734{
5735 /* pairs with tcp_poll() */
5736 smp_mb();
5737 if (sk->sk_socket &&
5738 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
5739 tcp_new_space(sk);
5740 if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5741 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
5742 }
5743}
5744
5745static inline void tcp_data_snd_check(struct sock *sk)
5746{
5747 tcp_push_pending_frames(sk);
5748 tcp_check_space(sk);
5749}
5750
5751/*
5752 * Check if sending an ack is needed.
5753 */
5754static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5755{
5756 struct tcp_sock *tp = tcp_sk(sk);
5757 unsigned long rtt, delay;
5758
5759 /* More than one full frame received... */
5760 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5761 /* ... and right edge of window advances far enough.
5762 * (tcp_recvmsg() will send ACK otherwise).
5763 * If application uses SO_RCVLOWAT, we want send ack now if
5764 * we have not received enough bytes to satisfy the condition.
5765 */
5766 (tp->rcv_nxt - tp->copied_seq < sk->sk_rcvlowat ||
5767 __tcp_select_window(sk) >= tp->rcv_wnd)) ||
5768 /* We ACK each frame or... */
5769 tcp_in_quickack_mode(sk) ||
5770 /* Protocol state mandates a one-time immediate ACK */
5771 inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOW) {
5772 /* If we are running from __release_sock() in user context,
5773 * Defer the ack until tcp_release_cb().
5774 */
5775 if (sock_owned_by_user_nocheck(sk) &&
5776 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_backlog_ack_defer)) {
5777 set_bit(TCP_ACK_DEFERRED, &sk->sk_tsq_flags);
5778 return;
5779 }
5780send_now:
5781 tcp_send_ack(sk);
5782 return;
5783 }
5784
5785 if (!ofo_possible || RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5786 tcp_send_delayed_ack(sk);
5787 return;
5788 }
5789
5790 if (!tcp_is_sack(tp) ||
5791 tp->compressed_ack >= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_nr))
5792 goto send_now;
5793
5794 if (tp->compressed_ack_rcv_nxt != tp->rcv_nxt) {
5795 tp->compressed_ack_rcv_nxt = tp->rcv_nxt;
5796 tp->dup_ack_counter = 0;
5797 }
5798 if (tp->dup_ack_counter < TCP_FASTRETRANS_THRESH) {
5799 tp->dup_ack_counter++;
5800 goto send_now;
5801 }
5802 tp->compressed_ack++;
5803 if (hrtimer_is_queued(&tp->compressed_ack_timer))
5804 return;
5805
5806 /* compress ack timer : 5 % of rtt, but no more than tcp_comp_sack_delay_ns */
5807
5808 rtt = tp->rcv_rtt_est.rtt_us;
5809 if (tp->srtt_us && tp->srtt_us < rtt)
5810 rtt = tp->srtt_us;
5811
5812 delay = min_t(unsigned long,
5813 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_delay_ns),
5814 rtt * (NSEC_PER_USEC >> 3)/20);
5815 sock_hold(sk);
5816 hrtimer_start_range_ns(&tp->compressed_ack_timer, ns_to_ktime(delay),
5817 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_slack_ns),
5818 HRTIMER_MODE_REL_PINNED_SOFT);
5819}
5820
5821static inline void tcp_ack_snd_check(struct sock *sk)
5822{
5823 if (!inet_csk_ack_scheduled(sk)) {
5824 /* We sent a data segment already. */
5825 return;
5826 }
5827 __tcp_ack_snd_check(sk, 1);
5828}
5829
5830/*
5831 * This routine is only called when we have urgent data
5832 * signaled. Its the 'slow' part of tcp_urg. It could be
5833 * moved inline now as tcp_urg is only called from one
5834 * place. We handle URGent data wrong. We have to - as
5835 * BSD still doesn't use the correction from RFC961.
5836 * For 1003.1g we should support a new option TCP_STDURG to permit
5837 * either form (or just set the sysctl tcp_stdurg).
5838 */
5839
5840static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5841{
5842 struct tcp_sock *tp = tcp_sk(sk);
5843 u32 ptr = ntohs(th->urg_ptr);
5844
5845 if (ptr && !READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_stdurg))
5846 ptr--;
5847 ptr += ntohl(th->seq);
5848
5849 /* Ignore urgent data that we've already seen and read. */
5850 if (after(tp->copied_seq, ptr))
5851 return;
5852
5853 /* Do not replay urg ptr.
5854 *
5855 * NOTE: interesting situation not covered by specs.
5856 * Misbehaving sender may send urg ptr, pointing to segment,
5857 * which we already have in ofo queue. We are not able to fetch
5858 * such data and will stay in TCP_URG_NOTYET until will be eaten
5859 * by recvmsg(). Seems, we are not obliged to handle such wicked
5860 * situations. But it is worth to think about possibility of some
5861 * DoSes using some hypothetical application level deadlock.
5862 */
5863 if (before(ptr, tp->rcv_nxt))
5864 return;
5865
5866 /* Do we already have a newer (or duplicate) urgent pointer? */
5867 if (tp->urg_data && !after(ptr, tp->urg_seq))
5868 return;
5869
5870 /* Tell the world about our new urgent pointer. */
5871 sk_send_sigurg(sk);
5872
5873 /* We may be adding urgent data when the last byte read was
5874 * urgent. To do this requires some care. We cannot just ignore
5875 * tp->copied_seq since we would read the last urgent byte again
5876 * as data, nor can we alter copied_seq until this data arrives
5877 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5878 *
5879 * NOTE. Double Dutch. Rendering to plain English: author of comment
5880 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5881 * and expect that both A and B disappear from stream. This is _wrong_.
5882 * Though this happens in BSD with high probability, this is occasional.
5883 * Any application relying on this is buggy. Note also, that fix "works"
5884 * only in this artificial test. Insert some normal data between A and B and we will
5885 * decline of BSD again. Verdict: it is better to remove to trap
5886 * buggy users.
5887 */
5888 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5889 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5890 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5891 tp->copied_seq++;
5892 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5893 __skb_unlink(skb, &sk->sk_receive_queue);
5894 __kfree_skb(skb);
5895 }
5896 }
5897
5898 WRITE_ONCE(tp->urg_data, TCP_URG_NOTYET);
5899 WRITE_ONCE(tp->urg_seq, ptr);
5900
5901 /* Disable header prediction. */
5902 tp->pred_flags = 0;
5903}
5904
5905/* This is the 'fast' part of urgent handling. */
5906static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5907{
5908 struct tcp_sock *tp = tcp_sk(sk);
5909
5910 /* Check if we get a new urgent pointer - normally not. */
5911 if (unlikely(th->urg))
5912 tcp_check_urg(sk, th);
5913
5914 /* Do we wait for any urgent data? - normally not... */
5915 if (unlikely(tp->urg_data == TCP_URG_NOTYET)) {
5916 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5917 th->syn;
5918
5919 /* Is the urgent pointer pointing into this packet? */
5920 if (ptr < skb->len) {
5921 u8 tmp;
5922 if (skb_copy_bits(skb, ptr, &tmp, 1))
5923 BUG();
5924 WRITE_ONCE(tp->urg_data, TCP_URG_VALID | tmp);
5925 if (!sock_flag(sk, SOCK_DEAD))
5926 sk->sk_data_ready(sk);
5927 }
5928 }
5929}
5930
5931/* Accept RST for rcv_nxt - 1 after a FIN.
5932 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a
5933 * FIN is sent followed by a RST packet. The RST is sent with the same
5934 * sequence number as the FIN, and thus according to RFC 5961 a challenge
5935 * ACK should be sent. However, Mac OSX rate limits replies to challenge
5936 * ACKs on the closed socket. In addition middleboxes can drop either the
5937 * challenge ACK or a subsequent RST.
5938 */
5939static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb)
5940{
5941 const struct tcp_sock *tp = tcp_sk(sk);
5942
5943 return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) &&
5944 (1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK |
5945 TCPF_CLOSING));
5946}
5947
5948/* Does PAWS and seqno based validation of an incoming segment, flags will
5949 * play significant role here.
5950 */
5951static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5952 const struct tcphdr *th, int syn_inerr)
5953{
5954 struct tcp_sock *tp = tcp_sk(sk);
5955 SKB_DR(reason);
5956
5957 /* RFC1323: H1. Apply PAWS check first. */
5958 if (tcp_fast_parse_options(sock_net(sk), skb, th, tp) &&
5959 tp->rx_opt.saw_tstamp &&
5960 tcp_paws_discard(sk, skb)) {
5961 if (!th->rst) {
5962 if (unlikely(th->syn))
5963 goto syn_challenge;
5964 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5965 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5966 LINUX_MIB_TCPACKSKIPPEDPAWS,
5967 &tp->last_oow_ack_time))
5968 tcp_send_dupack(sk, skb);
5969 SKB_DR_SET(reason, TCP_RFC7323_PAWS);
5970 goto discard;
5971 }
5972 /* Reset is accepted even if it did not pass PAWS. */
5973 }
5974
5975 /* Step 1: check sequence number */
5976 reason = tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
5977 if (reason) {
5978 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5979 * (RST) segments are validated by checking their SEQ-fields."
5980 * And page 69: "If an incoming segment is not acceptable,
5981 * an acknowledgment should be sent in reply (unless the RST
5982 * bit is set, if so drop the segment and return)".
5983 */
5984 if (!th->rst) {
5985 if (th->syn)
5986 goto syn_challenge;
5987 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5988 LINUX_MIB_TCPACKSKIPPEDSEQ,
5989 &tp->last_oow_ack_time))
5990 tcp_send_dupack(sk, skb);
5991 } else if (tcp_reset_check(sk, skb)) {
5992 goto reset;
5993 }
5994 goto discard;
5995 }
5996
5997 /* Step 2: check RST bit */
5998 if (th->rst) {
5999 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a
6000 * FIN and SACK too if available):
6001 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or
6002 * the right-most SACK block,
6003 * then
6004 * RESET the connection
6005 * else
6006 * Send a challenge ACK
6007 */
6008 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt ||
6009 tcp_reset_check(sk, skb))
6010 goto reset;
6011
6012 if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) {
6013 struct tcp_sack_block *sp = &tp->selective_acks[0];
6014 int max_sack = sp[0].end_seq;
6015 int this_sack;
6016
6017 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;
6018 ++this_sack) {
6019 max_sack = after(sp[this_sack].end_seq,
6020 max_sack) ?
6021 sp[this_sack].end_seq : max_sack;
6022 }
6023
6024 if (TCP_SKB_CB(skb)->seq == max_sack)
6025 goto reset;
6026 }
6027
6028 /* Disable TFO if RST is out-of-order
6029 * and no data has been received
6030 * for current active TFO socket
6031 */
6032 if (tp->syn_fastopen && !tp->data_segs_in &&
6033 sk->sk_state == TCP_ESTABLISHED)
6034 tcp_fastopen_active_disable(sk);
6035 tcp_send_challenge_ack(sk);
6036 SKB_DR_SET(reason, TCP_RESET);
6037 goto discard;
6038 }
6039
6040 /* step 3: check security and precedence [ignored] */
6041
6042 /* step 4: Check for a SYN
6043 * RFC 5961 4.2 : Send a challenge ack
6044 */
6045 if (th->syn) {
6046 if (sk->sk_state == TCP_SYN_RECV && sk->sk_socket && th->ack &&
6047 TCP_SKB_CB(skb)->seq + 1 == TCP_SKB_CB(skb)->end_seq &&
6048 TCP_SKB_CB(skb)->seq + 1 == tp->rcv_nxt &&
6049 TCP_SKB_CB(skb)->ack_seq == tp->snd_nxt)
6050 goto pass;
6051syn_challenge:
6052 if (syn_inerr)
6053 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
6054 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
6055 tcp_send_challenge_ack(sk);
6056 SKB_DR_SET(reason, TCP_INVALID_SYN);
6057 goto discard;
6058 }
6059
6060pass:
6061 bpf_skops_parse_hdr(sk, skb);
6062
6063 return true;
6064
6065discard:
6066 tcp_drop_reason(sk, skb, reason);
6067 return false;
6068
6069reset:
6070 tcp_reset(sk, skb);
6071 __kfree_skb(skb);
6072 return false;
6073}
6074
6075/*
6076 * TCP receive function for the ESTABLISHED state.
6077 *
6078 * It is split into a fast path and a slow path. The fast path is
6079 * disabled when:
6080 * - A zero window was announced from us - zero window probing
6081 * is only handled properly in the slow path.
6082 * - Out of order segments arrived.
6083 * - Urgent data is expected.
6084 * - There is no buffer space left
6085 * - Unexpected TCP flags/window values/header lengths are received
6086 * (detected by checking the TCP header against pred_flags)
6087 * - Data is sent in both directions. Fast path only supports pure senders
6088 * or pure receivers (this means either the sequence number or the ack
6089 * value must stay constant)
6090 * - Unexpected TCP option.
6091 *
6092 * When these conditions are not satisfied it drops into a standard
6093 * receive procedure patterned after RFC793 to handle all cases.
6094 * The first three cases are guaranteed by proper pred_flags setting,
6095 * the rest is checked inline. Fast processing is turned on in
6096 * tcp_data_queue when everything is OK.
6097 */
6098void tcp_rcv_established(struct sock *sk, struct sk_buff *skb)
6099{
6100 enum skb_drop_reason reason = SKB_DROP_REASON_NOT_SPECIFIED;
6101 const struct tcphdr *th = (const struct tcphdr *)skb->data;
6102 struct tcp_sock *tp = tcp_sk(sk);
6103 unsigned int len = skb->len;
6104
6105 /* TCP congestion window tracking */
6106 trace_tcp_probe(sk, skb);
6107
6108 tcp_mstamp_refresh(tp);
6109 if (unlikely(!rcu_access_pointer(sk->sk_rx_dst)))
6110 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
6111 /*
6112 * Header prediction.
6113 * The code loosely follows the one in the famous
6114 * "30 instruction TCP receive" Van Jacobson mail.
6115 *
6116 * Van's trick is to deposit buffers into socket queue
6117 * on a device interrupt, to call tcp_recv function
6118 * on the receive process context and checksum and copy
6119 * the buffer to user space. smart...
6120 *
6121 * Our current scheme is not silly either but we take the
6122 * extra cost of the net_bh soft interrupt processing...
6123 * We do checksum and copy also but from device to kernel.
6124 */
6125
6126 tp->rx_opt.saw_tstamp = 0;
6127
6128 /* pred_flags is 0xS?10 << 16 + snd_wnd
6129 * if header_prediction is to be made
6130 * 'S' will always be tp->tcp_header_len >> 2
6131 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
6132 * turn it off (when there are holes in the receive
6133 * space for instance)
6134 * PSH flag is ignored.
6135 */
6136
6137 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
6138 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
6139 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
6140 int tcp_header_len = tp->tcp_header_len;
6141
6142 /* Timestamp header prediction: tcp_header_len
6143 * is automatically equal to th->doff*4 due to pred_flags
6144 * match.
6145 */
6146
6147 /* Check timestamp */
6148 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
6149 /* No? Slow path! */
6150 if (!tcp_parse_aligned_timestamp(tp, th))
6151 goto slow_path;
6152
6153 /* If PAWS failed, check it more carefully in slow path */
6154 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
6155 goto slow_path;
6156
6157 /* DO NOT update ts_recent here, if checksum fails
6158 * and timestamp was corrupted part, it will result
6159 * in a hung connection since we will drop all
6160 * future packets due to the PAWS test.
6161 */
6162 }
6163
6164 if (len <= tcp_header_len) {
6165 /* Bulk data transfer: sender */
6166 if (len == tcp_header_len) {
6167 /* Predicted packet is in window by definition.
6168 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
6169 * Hence, check seq<=rcv_wup reduces to:
6170 */
6171 if (tcp_header_len ==
6172 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
6173 tp->rcv_nxt == tp->rcv_wup)
6174 tcp_store_ts_recent(tp);
6175
6176 /* We know that such packets are checksummed
6177 * on entry.
6178 */
6179 tcp_ack(sk, skb, 0);
6180 __kfree_skb(skb);
6181 tcp_data_snd_check(sk);
6182 /* When receiving pure ack in fast path, update
6183 * last ts ecr directly instead of calling
6184 * tcp_rcv_rtt_measure_ts()
6185 */
6186 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
6187 return;
6188 } else { /* Header too small */
6189 reason = SKB_DROP_REASON_PKT_TOO_SMALL;
6190 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
6191 goto discard;
6192 }
6193 } else {
6194 int eaten = 0;
6195 bool fragstolen = false;
6196
6197 if (tcp_checksum_complete(skb))
6198 goto csum_error;
6199
6200 if ((int)skb->truesize > sk->sk_forward_alloc)
6201 goto step5;
6202
6203 /* Predicted packet is in window by definition.
6204 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
6205 * Hence, check seq<=rcv_wup reduces to:
6206 */
6207 if (tcp_header_len ==
6208 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
6209 tp->rcv_nxt == tp->rcv_wup)
6210 tcp_store_ts_recent(tp);
6211
6212 tcp_rcv_rtt_measure_ts(sk, skb);
6213
6214 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS);
6215
6216 /* Bulk data transfer: receiver */
6217 tcp_cleanup_skb(skb);
6218 __skb_pull(skb, tcp_header_len);
6219 eaten = tcp_queue_rcv(sk, skb, &fragstolen);
6220
6221 tcp_event_data_recv(sk, skb);
6222
6223 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
6224 /* Well, only one small jumplet in fast path... */
6225 tcp_ack(sk, skb, FLAG_DATA);
6226 tcp_data_snd_check(sk);
6227 if (!inet_csk_ack_scheduled(sk))
6228 goto no_ack;
6229 } else {
6230 tcp_update_wl(tp, TCP_SKB_CB(skb)->seq);
6231 }
6232
6233 __tcp_ack_snd_check(sk, 0);
6234no_ack:
6235 if (eaten)
6236 kfree_skb_partial(skb, fragstolen);
6237 tcp_data_ready(sk);
6238 return;
6239 }
6240 }
6241
6242slow_path:
6243 if (len < (th->doff << 2) || tcp_checksum_complete(skb))
6244 goto csum_error;
6245
6246 if (!th->ack && !th->rst && !th->syn) {
6247 reason = SKB_DROP_REASON_TCP_FLAGS;
6248 goto discard;
6249 }
6250
6251 /*
6252 * Standard slow path.
6253 */
6254
6255 if (!tcp_validate_incoming(sk, skb, th, 1))
6256 return;
6257
6258step5:
6259 reason = tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT);
6260 if ((int)reason < 0) {
6261 reason = -reason;
6262 goto discard;
6263 }
6264 tcp_rcv_rtt_measure_ts(sk, skb);
6265
6266 /* Process urgent data. */
6267 tcp_urg(sk, skb, th);
6268
6269 /* step 7: process the segment text */
6270 tcp_data_queue(sk, skb);
6271
6272 tcp_data_snd_check(sk);
6273 tcp_ack_snd_check(sk);
6274 return;
6275
6276csum_error:
6277 reason = SKB_DROP_REASON_TCP_CSUM;
6278 trace_tcp_bad_csum(skb);
6279 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
6280 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
6281
6282discard:
6283 tcp_drop_reason(sk, skb, reason);
6284}
6285EXPORT_SYMBOL(tcp_rcv_established);
6286
6287void tcp_init_transfer(struct sock *sk, int bpf_op, struct sk_buff *skb)
6288{
6289 struct inet_connection_sock *icsk = inet_csk(sk);
6290 struct tcp_sock *tp = tcp_sk(sk);
6291
6292 tcp_mtup_init(sk);
6293 icsk->icsk_af_ops->rebuild_header(sk);
6294 tcp_init_metrics(sk);
6295
6296 /* Initialize the congestion window to start the transfer.
6297 * Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
6298 * retransmitted. In light of RFC6298 more aggressive 1sec
6299 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
6300 * retransmission has occurred.
6301 */
6302 if (tp->total_retrans > 1 && tp->undo_marker)
6303 tcp_snd_cwnd_set(tp, 1);
6304 else
6305 tcp_snd_cwnd_set(tp, tcp_init_cwnd(tp, __sk_dst_get(sk)));
6306 tp->snd_cwnd_stamp = tcp_jiffies32;
6307
6308 bpf_skops_established(sk, bpf_op, skb);
6309 /* Initialize congestion control unless BPF initialized it already: */
6310 if (!icsk->icsk_ca_initialized)
6311 tcp_init_congestion_control(sk);
6312 tcp_init_buffer_space(sk);
6313}
6314
6315void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
6316{
6317 struct tcp_sock *tp = tcp_sk(sk);
6318 struct inet_connection_sock *icsk = inet_csk(sk);
6319
6320 tcp_ao_finish_connect(sk, skb);
6321 tcp_set_state(sk, TCP_ESTABLISHED);
6322 icsk->icsk_ack.lrcvtime = tcp_jiffies32;
6323
6324 if (skb) {
6325 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
6326 security_inet_conn_established(sk, skb);
6327 sk_mark_napi_id(sk, skb);
6328 }
6329
6330 tcp_init_transfer(sk, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB, skb);
6331
6332 /* Prevent spurious tcp_cwnd_restart() on first data
6333 * packet.
6334 */
6335 tp->lsndtime = tcp_jiffies32;
6336
6337 if (sock_flag(sk, SOCK_KEEPOPEN))
6338 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
6339
6340 if (!tp->rx_opt.snd_wscale)
6341 __tcp_fast_path_on(tp, tp->snd_wnd);
6342 else
6343 tp->pred_flags = 0;
6344}
6345
6346static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
6347 struct tcp_fastopen_cookie *cookie)
6348{
6349 struct tcp_sock *tp = tcp_sk(sk);
6350 struct sk_buff *data = tp->syn_data ? tcp_rtx_queue_head(sk) : NULL;
6351 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
6352 bool syn_drop = false;
6353
6354 if (mss == tp->rx_opt.user_mss) {
6355 struct tcp_options_received opt;
6356
6357 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
6358 tcp_clear_options(&opt);
6359 opt.user_mss = opt.mss_clamp = 0;
6360 tcp_parse_options(sock_net(sk), synack, &opt, 0, NULL);
6361 mss = opt.mss_clamp;
6362 }
6363
6364 if (!tp->syn_fastopen) {
6365 /* Ignore an unsolicited cookie */
6366 cookie->len = -1;
6367 } else if (tp->total_retrans) {
6368 /* SYN timed out and the SYN-ACK neither has a cookie nor
6369 * acknowledges data. Presumably the remote received only
6370 * the retransmitted (regular) SYNs: either the original
6371 * SYN-data or the corresponding SYN-ACK was dropped.
6372 */
6373 syn_drop = (cookie->len < 0 && data);
6374 } else if (cookie->len < 0 && !tp->syn_data) {
6375 /* We requested a cookie but didn't get it. If we did not use
6376 * the (old) exp opt format then try so next time (try_exp=1).
6377 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
6378 */
6379 try_exp = tp->syn_fastopen_exp ? 2 : 1;
6380 }
6381
6382 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
6383
6384 if (data) { /* Retransmit unacked data in SYN */
6385 if (tp->total_retrans)
6386 tp->fastopen_client_fail = TFO_SYN_RETRANSMITTED;
6387 else
6388 tp->fastopen_client_fail = TFO_DATA_NOT_ACKED;
6389 skb_rbtree_walk_from(data)
6390 tcp_mark_skb_lost(sk, data);
6391 tcp_non_congestion_loss_retransmit(sk);
6392 NET_INC_STATS(sock_net(sk),
6393 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
6394 return true;
6395 }
6396 tp->syn_data_acked = tp->syn_data;
6397 if (tp->syn_data_acked) {
6398 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
6399 /* SYN-data is counted as two separate packets in tcp_ack() */
6400 if (tp->delivered > 1)
6401 --tp->delivered;
6402 }
6403
6404 tcp_fastopen_add_skb(sk, synack);
6405
6406 return false;
6407}
6408
6409static void smc_check_reset_syn(struct tcp_sock *tp)
6410{
6411#if IS_ENABLED(CONFIG_SMC)
6412 if (static_branch_unlikely(&tcp_have_smc)) {
6413 if (tp->syn_smc && !tp->rx_opt.smc_ok)
6414 tp->syn_smc = 0;
6415 }
6416#endif
6417}
6418
6419static void tcp_try_undo_spurious_syn(struct sock *sk)
6420{
6421 struct tcp_sock *tp = tcp_sk(sk);
6422 u32 syn_stamp;
6423
6424 /* undo_marker is set when SYN or SYNACK times out. The timeout is
6425 * spurious if the ACK's timestamp option echo value matches the
6426 * original SYN timestamp.
6427 */
6428 syn_stamp = tp->retrans_stamp;
6429 if (tp->undo_marker && syn_stamp && tp->rx_opt.saw_tstamp &&
6430 syn_stamp == tp->rx_opt.rcv_tsecr)
6431 tp->undo_marker = 0;
6432}
6433
6434static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
6435 const struct tcphdr *th)
6436{
6437 struct inet_connection_sock *icsk = inet_csk(sk);
6438 struct tcp_sock *tp = tcp_sk(sk);
6439 struct tcp_fastopen_cookie foc = { .len = -1 };
6440 int saved_clamp = tp->rx_opt.mss_clamp;
6441 bool fastopen_fail;
6442 SKB_DR(reason);
6443
6444 tcp_parse_options(sock_net(sk), skb, &tp->rx_opt, 0, &foc);
6445 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
6446 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
6447
6448 if (th->ack) {
6449 /* rfc793:
6450 * "If the state is SYN-SENT then
6451 * first check the ACK bit
6452 * If the ACK bit is set
6453 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
6454 * a reset (unless the RST bit is set, if so drop
6455 * the segment and return)"
6456 */
6457 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
6458 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
6459 /* Previous FIN/ACK or RST/ACK might be ignored. */
6460 if (icsk->icsk_retransmits == 0)
6461 inet_csk_reset_xmit_timer(sk,
6462 ICSK_TIME_RETRANS,
6463 TCP_TIMEOUT_MIN, TCP_RTO_MAX);
6464 SKB_DR_SET(reason, TCP_INVALID_ACK_SEQUENCE);
6465 goto reset_and_undo;
6466 }
6467
6468 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
6469 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
6470 tcp_time_stamp_ts(tp))) {
6471 NET_INC_STATS(sock_net(sk),
6472 LINUX_MIB_PAWSACTIVEREJECTED);
6473 SKB_DR_SET(reason, TCP_RFC7323_PAWS);
6474 goto reset_and_undo;
6475 }
6476
6477 /* Now ACK is acceptable.
6478 *
6479 * "If the RST bit is set
6480 * If the ACK was acceptable then signal the user "error:
6481 * connection reset", drop the segment, enter CLOSED state,
6482 * delete TCB, and return."
6483 */
6484
6485 if (th->rst) {
6486 tcp_reset(sk, skb);
6487consume:
6488 __kfree_skb(skb);
6489 return 0;
6490 }
6491
6492 /* rfc793:
6493 * "fifth, if neither of the SYN or RST bits is set then
6494 * drop the segment and return."
6495 *
6496 * See note below!
6497 * --ANK(990513)
6498 */
6499 if (!th->syn) {
6500 SKB_DR_SET(reason, TCP_FLAGS);
6501 goto discard_and_undo;
6502 }
6503 /* rfc793:
6504 * "If the SYN bit is on ...
6505 * are acceptable then ...
6506 * (our SYN has been ACKed), change the connection
6507 * state to ESTABLISHED..."
6508 */
6509
6510 tcp_ecn_rcv_synack(tp, th);
6511
6512 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6513 tcp_try_undo_spurious_syn(sk);
6514 tcp_ack(sk, skb, FLAG_SLOWPATH);
6515
6516 /* Ok.. it's good. Set up sequence numbers and
6517 * move to established.
6518 */
6519 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1);
6520 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
6521
6522 /* RFC1323: The window in SYN & SYN/ACK segments is
6523 * never scaled.
6524 */
6525 tp->snd_wnd = ntohs(th->window);
6526
6527 if (!tp->rx_opt.wscale_ok) {
6528 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
6529 WRITE_ONCE(tp->window_clamp,
6530 min(tp->window_clamp, 65535U));
6531 }
6532
6533 if (tp->rx_opt.saw_tstamp) {
6534 tp->rx_opt.tstamp_ok = 1;
6535 tp->tcp_header_len =
6536 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
6537 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6538 tcp_store_ts_recent(tp);
6539 } else {
6540 tp->tcp_header_len = sizeof(struct tcphdr);
6541 }
6542
6543 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
6544 tcp_initialize_rcv_mss(sk);
6545
6546 /* Remember, tcp_poll() does not lock socket!
6547 * Change state from SYN-SENT only after copied_seq
6548 * is initialized. */
6549 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6550
6551 smc_check_reset_syn(tp);
6552
6553 smp_mb();
6554
6555 tcp_finish_connect(sk, skb);
6556
6557 fastopen_fail = (tp->syn_fastopen || tp->syn_data) &&
6558 tcp_rcv_fastopen_synack(sk, skb, &foc);
6559
6560 if (!sock_flag(sk, SOCK_DEAD)) {
6561 sk->sk_state_change(sk);
6562 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6563 }
6564 if (fastopen_fail)
6565 return -1;
6566 if (sk->sk_write_pending ||
6567 READ_ONCE(icsk->icsk_accept_queue.rskq_defer_accept) ||
6568 inet_csk_in_pingpong_mode(sk)) {
6569 /* Save one ACK. Data will be ready after
6570 * several ticks, if write_pending is set.
6571 *
6572 * It may be deleted, but with this feature tcpdumps
6573 * look so _wonderfully_ clever, that I was not able
6574 * to stand against the temptation 8) --ANK
6575 */
6576 inet_csk_schedule_ack(sk);
6577 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
6578 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
6579 TCP_DELACK_MAX, TCP_RTO_MAX);
6580 goto consume;
6581 }
6582 tcp_send_ack(sk);
6583 return -1;
6584 }
6585
6586 /* No ACK in the segment */
6587
6588 if (th->rst) {
6589 /* rfc793:
6590 * "If the RST bit is set
6591 *
6592 * Otherwise (no ACK) drop the segment and return."
6593 */
6594 SKB_DR_SET(reason, TCP_RESET);
6595 goto discard_and_undo;
6596 }
6597
6598 /* PAWS check. */
6599 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
6600 tcp_paws_reject(&tp->rx_opt, 0)) {
6601 SKB_DR_SET(reason, TCP_RFC7323_PAWS);
6602 goto discard_and_undo;
6603 }
6604 if (th->syn) {
6605 /* We see SYN without ACK. It is attempt of
6606 * simultaneous connect with crossed SYNs.
6607 * Particularly, it can be connect to self.
6608 */
6609#ifdef CONFIG_TCP_AO
6610 struct tcp_ao_info *ao;
6611
6612 ao = rcu_dereference_protected(tp->ao_info,
6613 lockdep_sock_is_held(sk));
6614 if (ao) {
6615 WRITE_ONCE(ao->risn, th->seq);
6616 ao->rcv_sne = 0;
6617 }
6618#endif
6619 tcp_set_state(sk, TCP_SYN_RECV);
6620
6621 if (tp->rx_opt.saw_tstamp) {
6622 tp->rx_opt.tstamp_ok = 1;
6623 tcp_store_ts_recent(tp);
6624 tp->tcp_header_len =
6625 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
6626 } else {
6627 tp->tcp_header_len = sizeof(struct tcphdr);
6628 }
6629
6630 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1);
6631 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6632 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
6633
6634 /* RFC1323: The window in SYN & SYN/ACK segments is
6635 * never scaled.
6636 */
6637 tp->snd_wnd = ntohs(th->window);
6638 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
6639 tp->max_window = tp->snd_wnd;
6640
6641 tcp_ecn_rcv_syn(tp, th);
6642
6643 tcp_mtup_init(sk);
6644 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
6645 tcp_initialize_rcv_mss(sk);
6646
6647 tcp_send_synack(sk);
6648#if 0
6649 /* Note, we could accept data and URG from this segment.
6650 * There are no obstacles to make this (except that we must
6651 * either change tcp_recvmsg() to prevent it from returning data
6652 * before 3WHS completes per RFC793, or employ TCP Fast Open).
6653 *
6654 * However, if we ignore data in ACKless segments sometimes,
6655 * we have no reasons to accept it sometimes.
6656 * Also, seems the code doing it in step6 of tcp_rcv_state_process
6657 * is not flawless. So, discard packet for sanity.
6658 * Uncomment this return to process the data.
6659 */
6660 return -1;
6661#else
6662 goto consume;
6663#endif
6664 }
6665 /* "fifth, if neither of the SYN or RST bits is set then
6666 * drop the segment and return."
6667 */
6668
6669discard_and_undo:
6670 tcp_clear_options(&tp->rx_opt);
6671 tp->rx_opt.mss_clamp = saved_clamp;
6672 tcp_drop_reason(sk, skb, reason);
6673 return 0;
6674
6675reset_and_undo:
6676 tcp_clear_options(&tp->rx_opt);
6677 tp->rx_opt.mss_clamp = saved_clamp;
6678 /* we can reuse/return @reason to its caller to handle the exception */
6679 return reason;
6680}
6681
6682static void tcp_rcv_synrecv_state_fastopen(struct sock *sk)
6683{
6684 struct tcp_sock *tp = tcp_sk(sk);
6685 struct request_sock *req;
6686
6687 /* If we are still handling the SYNACK RTO, see if timestamp ECR allows
6688 * undo. If peer SACKs triggered fast recovery, we can't undo here.
6689 */
6690 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss && !tp->packets_out)
6691 tcp_try_undo_recovery(sk);
6692
6693 tcp_update_rto_time(tp);
6694 inet_csk(sk)->icsk_retransmits = 0;
6695 /* In tcp_fastopen_synack_timer() on the first SYNACK RTO we set
6696 * retrans_stamp but don't enter CA_Loss, so in case that happened we
6697 * need to zero retrans_stamp here to prevent spurious
6698 * retransmits_timed_out(). However, if the ACK of our SYNACK caused us
6699 * to enter CA_Recovery then we need to leave retrans_stamp as it was
6700 * set entering CA_Recovery, for correct retransmits_timed_out() and
6701 * undo behavior.
6702 */
6703 tcp_retrans_stamp_cleanup(sk);
6704
6705 /* Once we leave TCP_SYN_RECV or TCP_FIN_WAIT_1,
6706 * we no longer need req so release it.
6707 */
6708 req = rcu_dereference_protected(tp->fastopen_rsk,
6709 lockdep_sock_is_held(sk));
6710 reqsk_fastopen_remove(sk, req, false);
6711
6712 /* Re-arm the timer because data may have been sent out.
6713 * This is similar to the regular data transmission case
6714 * when new data has just been ack'ed.
6715 *
6716 * (TFO) - we could try to be more aggressive and
6717 * retransmitting any data sooner based on when they
6718 * are sent out.
6719 */
6720 tcp_rearm_rto(sk);
6721}
6722
6723/*
6724 * This function implements the receiving procedure of RFC 793 for
6725 * all states except ESTABLISHED and TIME_WAIT.
6726 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
6727 * address independent.
6728 */
6729
6730enum skb_drop_reason
6731tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
6732{
6733 struct tcp_sock *tp = tcp_sk(sk);
6734 struct inet_connection_sock *icsk = inet_csk(sk);
6735 const struct tcphdr *th = tcp_hdr(skb);
6736 struct request_sock *req;
6737 int queued = 0;
6738 SKB_DR(reason);
6739
6740 switch (sk->sk_state) {
6741 case TCP_CLOSE:
6742 SKB_DR_SET(reason, TCP_CLOSE);
6743 goto discard;
6744
6745 case TCP_LISTEN:
6746 if (th->ack)
6747 return SKB_DROP_REASON_TCP_FLAGS;
6748
6749 if (th->rst) {
6750 SKB_DR_SET(reason, TCP_RESET);
6751 goto discard;
6752 }
6753 if (th->syn) {
6754 if (th->fin) {
6755 SKB_DR_SET(reason, TCP_FLAGS);
6756 goto discard;
6757 }
6758 /* It is possible that we process SYN packets from backlog,
6759 * so we need to make sure to disable BH and RCU right there.
6760 */
6761 rcu_read_lock();
6762 local_bh_disable();
6763 icsk->icsk_af_ops->conn_request(sk, skb);
6764 local_bh_enable();
6765 rcu_read_unlock();
6766
6767 consume_skb(skb);
6768 return 0;
6769 }
6770 SKB_DR_SET(reason, TCP_FLAGS);
6771 goto discard;
6772
6773 case TCP_SYN_SENT:
6774 tp->rx_opt.saw_tstamp = 0;
6775 tcp_mstamp_refresh(tp);
6776 queued = tcp_rcv_synsent_state_process(sk, skb, th);
6777 if (queued >= 0)
6778 return queued;
6779
6780 /* Do step6 onward by hand. */
6781 tcp_urg(sk, skb, th);
6782 __kfree_skb(skb);
6783 tcp_data_snd_check(sk);
6784 return 0;
6785 }
6786
6787 tcp_mstamp_refresh(tp);
6788 tp->rx_opt.saw_tstamp = 0;
6789 req = rcu_dereference_protected(tp->fastopen_rsk,
6790 lockdep_sock_is_held(sk));
6791 if (req) {
6792 bool req_stolen;
6793
6794 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
6795 sk->sk_state != TCP_FIN_WAIT1);
6796
6797 if (!tcp_check_req(sk, skb, req, true, &req_stolen)) {
6798 SKB_DR_SET(reason, TCP_FASTOPEN);
6799 goto discard;
6800 }
6801 }
6802
6803 if (!th->ack && !th->rst && !th->syn) {
6804 SKB_DR_SET(reason, TCP_FLAGS);
6805 goto discard;
6806 }
6807 if (!tcp_validate_incoming(sk, skb, th, 0))
6808 return 0;
6809
6810 /* step 5: check the ACK field */
6811 reason = tcp_ack(sk, skb, FLAG_SLOWPATH |
6812 FLAG_UPDATE_TS_RECENT |
6813 FLAG_NO_CHALLENGE_ACK);
6814
6815 if ((int)reason <= 0) {
6816 if (sk->sk_state == TCP_SYN_RECV) {
6817 /* send one RST */
6818 if (!reason)
6819 return SKB_DROP_REASON_TCP_OLD_ACK;
6820 return -reason;
6821 }
6822 /* accept old ack during closing */
6823 if ((int)reason < 0) {
6824 tcp_send_challenge_ack(sk);
6825 reason = -reason;
6826 goto discard;
6827 }
6828 }
6829 SKB_DR_SET(reason, NOT_SPECIFIED);
6830 switch (sk->sk_state) {
6831 case TCP_SYN_RECV:
6832 tp->delivered++; /* SYN-ACK delivery isn't tracked in tcp_ack */
6833 if (!tp->srtt_us)
6834 tcp_synack_rtt_meas(sk, req);
6835
6836 if (req) {
6837 tcp_rcv_synrecv_state_fastopen(sk);
6838 } else {
6839 tcp_try_undo_spurious_syn(sk);
6840 tp->retrans_stamp = 0;
6841 tcp_init_transfer(sk, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB,
6842 skb);
6843 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6844 }
6845 tcp_ao_established(sk);
6846 smp_mb();
6847 tcp_set_state(sk, TCP_ESTABLISHED);
6848 sk->sk_state_change(sk);
6849
6850 /* Note, that this wakeup is only for marginal crossed SYN case.
6851 * Passively open sockets are not waked up, because
6852 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
6853 */
6854 if (sk->sk_socket)
6855 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6856
6857 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
6858 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
6859 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6860
6861 if (tp->rx_opt.tstamp_ok)
6862 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6863
6864 if (!inet_csk(sk)->icsk_ca_ops->cong_control)
6865 tcp_update_pacing_rate(sk);
6866
6867 /* Prevent spurious tcp_cwnd_restart() on first data packet */
6868 tp->lsndtime = tcp_jiffies32;
6869
6870 tcp_initialize_rcv_mss(sk);
6871 tcp_fast_path_on(tp);
6872 if (sk->sk_shutdown & SEND_SHUTDOWN)
6873 tcp_shutdown(sk, SEND_SHUTDOWN);
6874 break;
6875
6876 case TCP_FIN_WAIT1: {
6877 int tmo;
6878
6879 if (req)
6880 tcp_rcv_synrecv_state_fastopen(sk);
6881
6882 if (tp->snd_una != tp->write_seq)
6883 break;
6884
6885 tcp_set_state(sk, TCP_FIN_WAIT2);
6886 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | SEND_SHUTDOWN);
6887
6888 sk_dst_confirm(sk);
6889
6890 if (!sock_flag(sk, SOCK_DEAD)) {
6891 /* Wake up lingering close() */
6892 sk->sk_state_change(sk);
6893 break;
6894 }
6895
6896 if (READ_ONCE(tp->linger2) < 0) {
6897 tcp_done(sk);
6898 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6899 return SKB_DROP_REASON_TCP_ABORT_ON_DATA;
6900 }
6901 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6902 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6903 /* Receive out of order FIN after close() */
6904 if (tp->syn_fastopen && th->fin)
6905 tcp_fastopen_active_disable(sk);
6906 tcp_done(sk);
6907 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6908 return SKB_DROP_REASON_TCP_ABORT_ON_DATA;
6909 }
6910
6911 tmo = tcp_fin_time(sk);
6912 if (tmo > TCP_TIMEWAIT_LEN) {
6913 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
6914 } else if (th->fin || sock_owned_by_user(sk)) {
6915 /* Bad case. We could lose such FIN otherwise.
6916 * It is not a big problem, but it looks confusing
6917 * and not so rare event. We still can lose it now,
6918 * if it spins in bh_lock_sock(), but it is really
6919 * marginal case.
6920 */
6921 inet_csk_reset_keepalive_timer(sk, tmo);
6922 } else {
6923 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
6924 goto consume;
6925 }
6926 break;
6927 }
6928
6929 case TCP_CLOSING:
6930 if (tp->snd_una == tp->write_seq) {
6931 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6932 goto consume;
6933 }
6934 break;
6935
6936 case TCP_LAST_ACK:
6937 if (tp->snd_una == tp->write_seq) {
6938 tcp_update_metrics(sk);
6939 tcp_done(sk);
6940 goto consume;
6941 }
6942 break;
6943 }
6944
6945 /* step 6: check the URG bit */
6946 tcp_urg(sk, skb, th);
6947
6948 /* step 7: process the segment text */
6949 switch (sk->sk_state) {
6950 case TCP_CLOSE_WAIT:
6951 case TCP_CLOSING:
6952 case TCP_LAST_ACK:
6953 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
6954 /* If a subflow has been reset, the packet should not
6955 * continue to be processed, drop the packet.
6956 */
6957 if (sk_is_mptcp(sk) && !mptcp_incoming_options(sk, skb))
6958 goto discard;
6959 break;
6960 }
6961 fallthrough;
6962 case TCP_FIN_WAIT1:
6963 case TCP_FIN_WAIT2:
6964 /* RFC 793 says to queue data in these states,
6965 * RFC 1122 says we MUST send a reset.
6966 * BSD 4.4 also does reset.
6967 */
6968 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6969 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6970 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6971 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6972 tcp_reset(sk, skb);
6973 return SKB_DROP_REASON_TCP_ABORT_ON_DATA;
6974 }
6975 }
6976 fallthrough;
6977 case TCP_ESTABLISHED:
6978 tcp_data_queue(sk, skb);
6979 queued = 1;
6980 break;
6981 }
6982
6983 /* tcp_data could move socket to TIME-WAIT */
6984 if (sk->sk_state != TCP_CLOSE) {
6985 tcp_data_snd_check(sk);
6986 tcp_ack_snd_check(sk);
6987 }
6988
6989 if (!queued) {
6990discard:
6991 tcp_drop_reason(sk, skb, reason);
6992 }
6993 return 0;
6994
6995consume:
6996 __kfree_skb(skb);
6997 return 0;
6998}
6999EXPORT_SYMBOL(tcp_rcv_state_process);
7000
7001static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
7002{
7003 struct inet_request_sock *ireq = inet_rsk(req);
7004
7005 if (family == AF_INET)
7006 net_dbg_ratelimited("drop open request from %pI4/%u\n",
7007 &ireq->ir_rmt_addr, port);
7008#if IS_ENABLED(CONFIG_IPV6)
7009 else if (family == AF_INET6)
7010 net_dbg_ratelimited("drop open request from %pI6/%u\n",
7011 &ireq->ir_v6_rmt_addr, port);
7012#endif
7013}
7014
7015/* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
7016 *
7017 * If we receive a SYN packet with these bits set, it means a
7018 * network is playing bad games with TOS bits. In order to
7019 * avoid possible false congestion notifications, we disable
7020 * TCP ECN negotiation.
7021 *
7022 * Exception: tcp_ca wants ECN. This is required for DCTCP
7023 * congestion control: Linux DCTCP asserts ECT on all packets,
7024 * including SYN, which is most optimal solution; however,
7025 * others, such as FreeBSD do not.
7026 *
7027 * Exception: At least one of the reserved bits of the TCP header (th->res1) is
7028 * set, indicating the use of a future TCP extension (such as AccECN). See
7029 * RFC8311 §4.3 which updates RFC3168 to allow the development of such
7030 * extensions.
7031 */
7032static void tcp_ecn_create_request(struct request_sock *req,
7033 const struct sk_buff *skb,
7034 const struct sock *listen_sk,
7035 const struct dst_entry *dst)
7036{
7037 const struct tcphdr *th = tcp_hdr(skb);
7038 const struct net *net = sock_net(listen_sk);
7039 bool th_ecn = th->ece && th->cwr;
7040 bool ect, ecn_ok;
7041 u32 ecn_ok_dst;
7042
7043 if (!th_ecn)
7044 return;
7045
7046 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
7047 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK);
7048 ecn_ok = READ_ONCE(net->ipv4.sysctl_tcp_ecn) || ecn_ok_dst;
7049
7050 if (((!ect || th->res1) && ecn_ok) || tcp_ca_needs_ecn(listen_sk) ||
7051 (ecn_ok_dst & DST_FEATURE_ECN_CA) ||
7052 tcp_bpf_ca_needs_ecn((struct sock *)req))
7053 inet_rsk(req)->ecn_ok = 1;
7054}
7055
7056static void tcp_openreq_init(struct request_sock *req,
7057 const struct tcp_options_received *rx_opt,
7058 struct sk_buff *skb, const struct sock *sk)
7059{
7060 struct inet_request_sock *ireq = inet_rsk(req);
7061
7062 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */
7063 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
7064 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
7065 tcp_rsk(req)->snt_synack = 0;
7066 tcp_rsk(req)->last_oow_ack_time = 0;
7067 req->mss = rx_opt->mss_clamp;
7068 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
7069 ireq->tstamp_ok = rx_opt->tstamp_ok;
7070 ireq->sack_ok = rx_opt->sack_ok;
7071 ireq->snd_wscale = rx_opt->snd_wscale;
7072 ireq->wscale_ok = rx_opt->wscale_ok;
7073 ireq->acked = 0;
7074 ireq->ecn_ok = 0;
7075 ireq->ir_rmt_port = tcp_hdr(skb)->source;
7076 ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
7077 ireq->ir_mark = inet_request_mark(sk, skb);
7078#if IS_ENABLED(CONFIG_SMC)
7079 ireq->smc_ok = rx_opt->smc_ok && !(tcp_sk(sk)->smc_hs_congested &&
7080 tcp_sk(sk)->smc_hs_congested(sk));
7081#endif
7082}
7083
7084/*
7085 * Return true if a syncookie should be sent
7086 */
7087static bool tcp_syn_flood_action(struct sock *sk, const char *proto)
7088{
7089 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
7090 const char *msg = "Dropping request";
7091 struct net *net = sock_net(sk);
7092 bool want_cookie = false;
7093 u8 syncookies;
7094
7095 syncookies = READ_ONCE(net->ipv4.sysctl_tcp_syncookies);
7096
7097#ifdef CONFIG_SYN_COOKIES
7098 if (syncookies) {
7099 msg = "Sending cookies";
7100 want_cookie = true;
7101 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
7102 } else
7103#endif
7104 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
7105
7106 if (!READ_ONCE(queue->synflood_warned) && syncookies != 2 &&
7107 xchg(&queue->synflood_warned, 1) == 0) {
7108 if (IS_ENABLED(CONFIG_IPV6) && sk->sk_family == AF_INET6) {
7109 net_info_ratelimited("%s: Possible SYN flooding on port [%pI6c]:%u. %s.\n",
7110 proto, inet6_rcv_saddr(sk),
7111 sk->sk_num, msg);
7112 } else {
7113 net_info_ratelimited("%s: Possible SYN flooding on port %pI4:%u. %s.\n",
7114 proto, &sk->sk_rcv_saddr,
7115 sk->sk_num, msg);
7116 }
7117 }
7118
7119 return want_cookie;
7120}
7121
7122static void tcp_reqsk_record_syn(const struct sock *sk,
7123 struct request_sock *req,
7124 const struct sk_buff *skb)
7125{
7126 if (tcp_sk(sk)->save_syn) {
7127 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
7128 struct saved_syn *saved_syn;
7129 u32 mac_hdrlen;
7130 void *base;
7131
7132 if (tcp_sk(sk)->save_syn == 2) { /* Save full header. */
7133 base = skb_mac_header(skb);
7134 mac_hdrlen = skb_mac_header_len(skb);
7135 len += mac_hdrlen;
7136 } else {
7137 base = skb_network_header(skb);
7138 mac_hdrlen = 0;
7139 }
7140
7141 saved_syn = kmalloc(struct_size(saved_syn, data, len),
7142 GFP_ATOMIC);
7143 if (saved_syn) {
7144 saved_syn->mac_hdrlen = mac_hdrlen;
7145 saved_syn->network_hdrlen = skb_network_header_len(skb);
7146 saved_syn->tcp_hdrlen = tcp_hdrlen(skb);
7147 memcpy(saved_syn->data, base, len);
7148 req->saved_syn = saved_syn;
7149 }
7150 }
7151}
7152
7153/* If a SYN cookie is required and supported, returns a clamped MSS value to be
7154 * used for SYN cookie generation.
7155 */
7156u16 tcp_get_syncookie_mss(struct request_sock_ops *rsk_ops,
7157 const struct tcp_request_sock_ops *af_ops,
7158 struct sock *sk, struct tcphdr *th)
7159{
7160 struct tcp_sock *tp = tcp_sk(sk);
7161 u16 mss;
7162
7163 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_syncookies) != 2 &&
7164 !inet_csk_reqsk_queue_is_full(sk))
7165 return 0;
7166
7167 if (!tcp_syn_flood_action(sk, rsk_ops->slab_name))
7168 return 0;
7169
7170 if (sk_acceptq_is_full(sk)) {
7171 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
7172 return 0;
7173 }
7174
7175 mss = tcp_parse_mss_option(th, tp->rx_opt.user_mss);
7176 if (!mss)
7177 mss = af_ops->mss_clamp;
7178
7179 return mss;
7180}
7181EXPORT_SYMBOL_GPL(tcp_get_syncookie_mss);
7182
7183int tcp_conn_request(struct request_sock_ops *rsk_ops,
7184 const struct tcp_request_sock_ops *af_ops,
7185 struct sock *sk, struct sk_buff *skb)
7186{
7187 struct tcp_fastopen_cookie foc = { .len = -1 };
7188 struct tcp_options_received tmp_opt;
7189 struct tcp_sock *tp = tcp_sk(sk);
7190 struct net *net = sock_net(sk);
7191 struct sock *fastopen_sk = NULL;
7192 struct request_sock *req;
7193 bool want_cookie = false;
7194 struct dst_entry *dst;
7195 struct flowi fl;
7196 u8 syncookies;
7197 u32 isn;
7198
7199#ifdef CONFIG_TCP_AO
7200 const struct tcp_ao_hdr *aoh;
7201#endif
7202
7203 isn = __this_cpu_read(tcp_tw_isn);
7204 if (isn) {
7205 /* TW buckets are converted to open requests without
7206 * limitations, they conserve resources and peer is
7207 * evidently real one.
7208 */
7209 __this_cpu_write(tcp_tw_isn, 0);
7210 } else {
7211 syncookies = READ_ONCE(net->ipv4.sysctl_tcp_syncookies);
7212
7213 if (syncookies == 2 || inet_csk_reqsk_queue_is_full(sk)) {
7214 want_cookie = tcp_syn_flood_action(sk,
7215 rsk_ops->slab_name);
7216 if (!want_cookie)
7217 goto drop;
7218 }
7219 }
7220
7221 if (sk_acceptq_is_full(sk)) {
7222 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
7223 goto drop;
7224 }
7225
7226 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie);
7227 if (!req)
7228 goto drop;
7229
7230 req->syncookie = want_cookie;
7231 tcp_rsk(req)->af_specific = af_ops;
7232 tcp_rsk(req)->ts_off = 0;
7233 tcp_rsk(req)->req_usec_ts = false;
7234#if IS_ENABLED(CONFIG_MPTCP)
7235 tcp_rsk(req)->is_mptcp = 0;
7236#endif
7237
7238 tcp_clear_options(&tmp_opt);
7239 tmp_opt.mss_clamp = af_ops->mss_clamp;
7240 tmp_opt.user_mss = tp->rx_opt.user_mss;
7241 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0,
7242 want_cookie ? NULL : &foc);
7243
7244 if (want_cookie && !tmp_opt.saw_tstamp)
7245 tcp_clear_options(&tmp_opt);
7246
7247 if (IS_ENABLED(CONFIG_SMC) && want_cookie)
7248 tmp_opt.smc_ok = 0;
7249
7250 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
7251 tcp_openreq_init(req, &tmp_opt, skb, sk);
7252 inet_rsk(req)->no_srccheck = inet_test_bit(TRANSPARENT, sk);
7253
7254 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
7255 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb);
7256
7257 dst = af_ops->route_req(sk, skb, &fl, req, isn);
7258 if (!dst)
7259 goto drop_and_free;
7260
7261 if (tmp_opt.tstamp_ok) {
7262 tcp_rsk(req)->req_usec_ts = dst_tcp_usec_ts(dst);
7263 tcp_rsk(req)->ts_off = af_ops->init_ts_off(net, skb);
7264 }
7265 if (!want_cookie && !isn) {
7266 int max_syn_backlog = READ_ONCE(net->ipv4.sysctl_max_syn_backlog);
7267
7268 /* Kill the following clause, if you dislike this way. */
7269 if (!syncookies &&
7270 (max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
7271 (max_syn_backlog >> 2)) &&
7272 !tcp_peer_is_proven(req, dst)) {
7273 /* Without syncookies last quarter of
7274 * backlog is filled with destinations,
7275 * proven to be alive.
7276 * It means that we continue to communicate
7277 * to destinations, already remembered
7278 * to the moment of synflood.
7279 */
7280 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
7281 rsk_ops->family);
7282 goto drop_and_release;
7283 }
7284
7285 isn = af_ops->init_seq(skb);
7286 }
7287
7288 tcp_ecn_create_request(req, skb, sk, dst);
7289
7290 if (want_cookie) {
7291 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
7292 if (!tmp_opt.tstamp_ok)
7293 inet_rsk(req)->ecn_ok = 0;
7294 }
7295
7296#ifdef CONFIG_TCP_AO
7297 if (tcp_parse_auth_options(tcp_hdr(skb), NULL, &aoh))
7298 goto drop_and_release; /* Invalid TCP options */
7299 if (aoh) {
7300 tcp_rsk(req)->used_tcp_ao = true;
7301 tcp_rsk(req)->ao_rcv_next = aoh->keyid;
7302 tcp_rsk(req)->ao_keyid = aoh->rnext_keyid;
7303
7304 } else {
7305 tcp_rsk(req)->used_tcp_ao = false;
7306 }
7307#endif
7308 tcp_rsk(req)->snt_isn = isn;
7309 tcp_rsk(req)->txhash = net_tx_rndhash();
7310 tcp_rsk(req)->syn_tos = TCP_SKB_CB(skb)->ip_dsfield;
7311 tcp_openreq_init_rwin(req, sk, dst);
7312 sk_rx_queue_set(req_to_sk(req), skb);
7313 if (!want_cookie) {
7314 tcp_reqsk_record_syn(sk, req, skb);
7315 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst);
7316 }
7317 if (fastopen_sk) {
7318 af_ops->send_synack(fastopen_sk, dst, &fl, req,
7319 &foc, TCP_SYNACK_FASTOPEN, skb);
7320 /* Add the child socket directly into the accept queue */
7321 if (!inet_csk_reqsk_queue_add(sk, req, fastopen_sk)) {
7322 reqsk_fastopen_remove(fastopen_sk, req, false);
7323 bh_unlock_sock(fastopen_sk);
7324 sock_put(fastopen_sk);
7325 goto drop_and_free;
7326 }
7327 sk->sk_data_ready(sk);
7328 bh_unlock_sock(fastopen_sk);
7329 sock_put(fastopen_sk);
7330 } else {
7331 tcp_rsk(req)->tfo_listener = false;
7332 if (!want_cookie) {
7333 req->timeout = tcp_timeout_init((struct sock *)req);
7334 if (unlikely(!inet_csk_reqsk_queue_hash_add(sk, req,
7335 req->timeout))) {
7336 reqsk_free(req);
7337 dst_release(dst);
7338 return 0;
7339 }
7340
7341 }
7342 af_ops->send_synack(sk, dst, &fl, req, &foc,
7343 !want_cookie ? TCP_SYNACK_NORMAL :
7344 TCP_SYNACK_COOKIE,
7345 skb);
7346 if (want_cookie) {
7347 reqsk_free(req);
7348 return 0;
7349 }
7350 }
7351 reqsk_put(req);
7352 return 0;
7353
7354drop_and_release:
7355 dst_release(dst);
7356drop_and_free:
7357 __reqsk_free(req);
7358drop:
7359 tcp_listendrop(sk);
7360 return 0;
7361}
7362EXPORT_SYMBOL(tcp_conn_request);
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
82int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
83
84#define FLAG_DATA 0x01 /* Incoming frame contained data. */
85#define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
86#define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
87#define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
88#define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
89#define FLAG_DATA_SACKED 0x20 /* New SACK. */
90#define FLAG_ECE 0x40 /* ECE in this ACK */
91#define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */
92#define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
93#define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
94#define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
95#define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
96#define FLAG_SET_XMIT_TIMER 0x1000 /* Set TLP or RTO timer */
97#define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
98#define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
99#define FLAG_NO_CHALLENGE_ACK 0x8000 /* do not call tcp_send_challenge_ack() */
100#define FLAG_ACK_MAYBE_DELAYED 0x10000 /* Likely a delayed ACK */
101
102#define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
103#define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
104#define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE|FLAG_DSACKING_ACK)
105#define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
106
107#define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
108#define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
109
110#define REXMIT_NONE 0 /* no loss recovery to do */
111#define REXMIT_LOST 1 /* retransmit packets marked lost */
112#define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */
113
114static void tcp_gro_dev_warn(struct sock *sk, const struct sk_buff *skb,
115 unsigned int len)
116{
117 static bool __once __read_mostly;
118
119 if (!__once) {
120 struct net_device *dev;
121
122 __once = true;
123
124 rcu_read_lock();
125 dev = dev_get_by_index_rcu(sock_net(sk), skb->skb_iif);
126 if (!dev || len >= dev->mtu)
127 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n",
128 dev ? dev->name : "Unknown driver");
129 rcu_read_unlock();
130 }
131}
132
133/* Adapt the MSS value used to make delayed ack decision to the
134 * real world.
135 */
136static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
137{
138 struct inet_connection_sock *icsk = inet_csk(sk);
139 const unsigned int lss = icsk->icsk_ack.last_seg_size;
140 unsigned int len;
141
142 icsk->icsk_ack.last_seg_size = 0;
143
144 /* skb->len may jitter because of SACKs, even if peer
145 * sends good full-sized frames.
146 */
147 len = skb_shinfo(skb)->gso_size ? : skb->len;
148 if (len >= icsk->icsk_ack.rcv_mss) {
149 icsk->icsk_ack.rcv_mss = min_t(unsigned int, len,
150 tcp_sk(sk)->advmss);
151 /* Account for possibly-removed options */
152 if (unlikely(len > icsk->icsk_ack.rcv_mss +
153 MAX_TCP_OPTION_SPACE))
154 tcp_gro_dev_warn(sk, skb, len);
155 } else {
156 /* Otherwise, we make more careful check taking into account,
157 * that SACKs block is variable.
158 *
159 * "len" is invariant segment length, including TCP header.
160 */
161 len += skb->data - skb_transport_header(skb);
162 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
163 /* If PSH is not set, packet should be
164 * full sized, provided peer TCP is not badly broken.
165 * This observation (if it is correct 8)) allows
166 * to handle super-low mtu links fairly.
167 */
168 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
169 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
170 /* Subtract also invariant (if peer is RFC compliant),
171 * tcp header plus fixed timestamp option length.
172 * Resulting "len" is MSS free of SACK jitter.
173 */
174 len -= tcp_sk(sk)->tcp_header_len;
175 icsk->icsk_ack.last_seg_size = len;
176 if (len == lss) {
177 icsk->icsk_ack.rcv_mss = len;
178 return;
179 }
180 }
181 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
182 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
183 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
184 }
185}
186
187static void tcp_incr_quickack(struct sock *sk)
188{
189 struct inet_connection_sock *icsk = inet_csk(sk);
190 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
191
192 if (quickacks == 0)
193 quickacks = 2;
194 if (quickacks > icsk->icsk_ack.quick)
195 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
196}
197
198static void tcp_enter_quickack_mode(struct sock *sk)
199{
200 struct inet_connection_sock *icsk = inet_csk(sk);
201 tcp_incr_quickack(sk);
202 icsk->icsk_ack.pingpong = 0;
203 icsk->icsk_ack.ato = TCP_ATO_MIN;
204}
205
206/* Send ACKs quickly, if "quick" count is not exhausted
207 * and the session is not interactive.
208 */
209
210static bool tcp_in_quickack_mode(struct sock *sk)
211{
212 const struct inet_connection_sock *icsk = inet_csk(sk);
213 const struct dst_entry *dst = __sk_dst_get(sk);
214
215 return (dst && dst_metric(dst, RTAX_QUICKACK)) ||
216 (icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong);
217}
218
219static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
220{
221 if (tp->ecn_flags & TCP_ECN_OK)
222 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
223}
224
225static void tcp_ecn_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
226{
227 if (tcp_hdr(skb)->cwr)
228 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
229}
230
231static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
232{
233 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
234}
235
236static void __tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
237{
238 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
239 case INET_ECN_NOT_ECT:
240 /* Funny extension: if ECT is not set on a segment,
241 * and we already seen ECT on a previous segment,
242 * it is probably a retransmit.
243 */
244 if (tp->ecn_flags & TCP_ECN_SEEN)
245 tcp_enter_quickack_mode((struct sock *)tp);
246 break;
247 case INET_ECN_CE:
248 if (tcp_ca_needs_ecn((struct sock *)tp))
249 tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_IS_CE);
250
251 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
252 /* Better not delay acks, sender can have a very low cwnd */
253 tcp_enter_quickack_mode((struct sock *)tp);
254 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
255 }
256 tp->ecn_flags |= TCP_ECN_SEEN;
257 break;
258 default:
259 if (tcp_ca_needs_ecn((struct sock *)tp))
260 tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_NO_CE);
261 tp->ecn_flags |= TCP_ECN_SEEN;
262 break;
263 }
264}
265
266static void tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
267{
268 if (tp->ecn_flags & TCP_ECN_OK)
269 __tcp_ecn_check_ce(tp, skb);
270}
271
272static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
273{
274 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
275 tp->ecn_flags &= ~TCP_ECN_OK;
276}
277
278static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
279{
280 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
281 tp->ecn_flags &= ~TCP_ECN_OK;
282}
283
284static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
285{
286 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
287 return true;
288 return false;
289}
290
291/* Buffer size and advertised window tuning.
292 *
293 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
294 */
295
296static void tcp_sndbuf_expand(struct sock *sk)
297{
298 const struct tcp_sock *tp = tcp_sk(sk);
299 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
300 int sndmem, per_mss;
301 u32 nr_segs;
302
303 /* Worst case is non GSO/TSO : each frame consumes one skb
304 * and skb->head is kmalloced using power of two area of memory
305 */
306 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
307 MAX_TCP_HEADER +
308 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
309
310 per_mss = roundup_pow_of_two(per_mss) +
311 SKB_DATA_ALIGN(sizeof(struct sk_buff));
312
313 nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd);
314 nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
315
316 /* Fast Recovery (RFC 5681 3.2) :
317 * Cubic needs 1.7 factor, rounded to 2 to include
318 * extra cushion (application might react slowly to EPOLLOUT)
319 */
320 sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2;
321 sndmem *= nr_segs * per_mss;
322
323 if (sk->sk_sndbuf < sndmem)
324 sk->sk_sndbuf = min(sndmem, sock_net(sk)->ipv4.sysctl_tcp_wmem[2]);
325}
326
327/* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
328 *
329 * All tcp_full_space() is split to two parts: "network" buffer, allocated
330 * forward and advertised in receiver window (tp->rcv_wnd) and
331 * "application buffer", required to isolate scheduling/application
332 * latencies from network.
333 * window_clamp is maximal advertised window. It can be less than
334 * tcp_full_space(), in this case tcp_full_space() - window_clamp
335 * is reserved for "application" buffer. The less window_clamp is
336 * the smoother our behaviour from viewpoint of network, but the lower
337 * throughput and the higher sensitivity of the connection to losses. 8)
338 *
339 * rcv_ssthresh is more strict window_clamp used at "slow start"
340 * phase to predict further behaviour of this connection.
341 * It is used for two goals:
342 * - to enforce header prediction at sender, even when application
343 * requires some significant "application buffer". It is check #1.
344 * - to prevent pruning of receive queue because of misprediction
345 * of receiver window. Check #2.
346 *
347 * The scheme does not work when sender sends good segments opening
348 * window and then starts to feed us spaghetti. But it should work
349 * in common situations. Otherwise, we have to rely on queue collapsing.
350 */
351
352/* Slow part of check#2. */
353static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
354{
355 struct tcp_sock *tp = tcp_sk(sk);
356 /* Optimize this! */
357 int truesize = tcp_win_from_space(sk, skb->truesize) >> 1;
358 int window = tcp_win_from_space(sk, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]) >> 1;
359
360 while (tp->rcv_ssthresh <= window) {
361 if (truesize <= skb->len)
362 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
363
364 truesize >>= 1;
365 window >>= 1;
366 }
367 return 0;
368}
369
370static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
371{
372 struct tcp_sock *tp = tcp_sk(sk);
373
374 /* Check #1 */
375 if (tp->rcv_ssthresh < tp->window_clamp &&
376 (int)tp->rcv_ssthresh < tcp_space(sk) &&
377 !tcp_under_memory_pressure(sk)) {
378 int incr;
379
380 /* Check #2. Increase window, if skb with such overhead
381 * will fit to rcvbuf in future.
382 */
383 if (tcp_win_from_space(sk, skb->truesize) <= skb->len)
384 incr = 2 * tp->advmss;
385 else
386 incr = __tcp_grow_window(sk, skb);
387
388 if (incr) {
389 incr = max_t(int, incr, 2 * skb->len);
390 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
391 tp->window_clamp);
392 inet_csk(sk)->icsk_ack.quick |= 1;
393 }
394 }
395}
396
397/* 3. Tuning rcvbuf, when connection enters established state. */
398static void tcp_fixup_rcvbuf(struct sock *sk)
399{
400 u32 mss = tcp_sk(sk)->advmss;
401 int rcvmem;
402
403 rcvmem = 2 * SKB_TRUESIZE(mss + MAX_TCP_HEADER) *
404 tcp_default_init_rwnd(mss);
405
406 /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
407 * Allow enough cushion so that sender is not limited by our window
408 */
409 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf)
410 rcvmem <<= 2;
411
412 if (sk->sk_rcvbuf < rcvmem)
413 sk->sk_rcvbuf = min(rcvmem, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
414}
415
416/* 4. Try to fixup all. It is made immediately after connection enters
417 * established state.
418 */
419void tcp_init_buffer_space(struct sock *sk)
420{
421 int tcp_app_win = sock_net(sk)->ipv4.sysctl_tcp_app_win;
422 struct tcp_sock *tp = tcp_sk(sk);
423 int maxwin;
424
425 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
426 tcp_fixup_rcvbuf(sk);
427 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
428 tcp_sndbuf_expand(sk);
429
430 tp->rcvq_space.space = tp->rcv_wnd;
431 tcp_mstamp_refresh(tp);
432 tp->rcvq_space.time = tp->tcp_mstamp;
433 tp->rcvq_space.seq = tp->copied_seq;
434
435 maxwin = tcp_full_space(sk);
436
437 if (tp->window_clamp >= maxwin) {
438 tp->window_clamp = maxwin;
439
440 if (tcp_app_win && maxwin > 4 * tp->advmss)
441 tp->window_clamp = max(maxwin -
442 (maxwin >> tcp_app_win),
443 4 * tp->advmss);
444 }
445
446 /* Force reservation of one segment. */
447 if (tcp_app_win &&
448 tp->window_clamp > 2 * tp->advmss &&
449 tp->window_clamp + tp->advmss > maxwin)
450 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
451
452 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
453 tp->snd_cwnd_stamp = tcp_jiffies32;
454}
455
456/* 5. Recalculate window clamp after socket hit its memory bounds. */
457static void tcp_clamp_window(struct sock *sk)
458{
459 struct tcp_sock *tp = tcp_sk(sk);
460 struct inet_connection_sock *icsk = inet_csk(sk);
461 struct net *net = sock_net(sk);
462
463 icsk->icsk_ack.quick = 0;
464
465 if (sk->sk_rcvbuf < net->ipv4.sysctl_tcp_rmem[2] &&
466 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
467 !tcp_under_memory_pressure(sk) &&
468 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
469 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
470 net->ipv4.sysctl_tcp_rmem[2]);
471 }
472 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
473 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
474}
475
476/* Initialize RCV_MSS value.
477 * RCV_MSS is an our guess about MSS used by the peer.
478 * We haven't any direct information about the MSS.
479 * It's better to underestimate the RCV_MSS rather than overestimate.
480 * Overestimations make us ACKing less frequently than needed.
481 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
482 */
483void tcp_initialize_rcv_mss(struct sock *sk)
484{
485 const struct tcp_sock *tp = tcp_sk(sk);
486 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
487
488 hint = min(hint, tp->rcv_wnd / 2);
489 hint = min(hint, TCP_MSS_DEFAULT);
490 hint = max(hint, TCP_MIN_MSS);
491
492 inet_csk(sk)->icsk_ack.rcv_mss = hint;
493}
494EXPORT_SYMBOL(tcp_initialize_rcv_mss);
495
496/* Receiver "autotuning" code.
497 *
498 * The algorithm for RTT estimation w/o timestamps is based on
499 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
500 * <http://public.lanl.gov/radiant/pubs.html#DRS>
501 *
502 * More detail on this code can be found at
503 * <http://staff.psc.edu/jheffner/>,
504 * though this reference is out of date. A new paper
505 * is pending.
506 */
507static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
508{
509 u32 new_sample = tp->rcv_rtt_est.rtt_us;
510 long m = sample;
511
512 if (new_sample != 0) {
513 /* If we sample in larger samples in the non-timestamp
514 * case, we could grossly overestimate the RTT especially
515 * with chatty applications or bulk transfer apps which
516 * are stalled on filesystem I/O.
517 *
518 * Also, since we are only going for a minimum in the
519 * non-timestamp case, we do not smooth things out
520 * else with timestamps disabled convergence takes too
521 * long.
522 */
523 if (!win_dep) {
524 m -= (new_sample >> 3);
525 new_sample += m;
526 } else {
527 m <<= 3;
528 if (m < new_sample)
529 new_sample = m;
530 }
531 } else {
532 /* No previous measure. */
533 new_sample = m << 3;
534 }
535
536 tp->rcv_rtt_est.rtt_us = new_sample;
537}
538
539static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
540{
541 u32 delta_us;
542
543 if (tp->rcv_rtt_est.time == 0)
544 goto new_measure;
545 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
546 return;
547 delta_us = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcv_rtt_est.time);
548 if (!delta_us)
549 delta_us = 1;
550 tcp_rcv_rtt_update(tp, delta_us, 1);
551
552new_measure:
553 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
554 tp->rcv_rtt_est.time = tp->tcp_mstamp;
555}
556
557static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
558 const struct sk_buff *skb)
559{
560 struct tcp_sock *tp = tcp_sk(sk);
561
562 if (tp->rx_opt.rcv_tsecr &&
563 (TCP_SKB_CB(skb)->end_seq -
564 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss)) {
565 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
566 u32 delta_us;
567
568 if (!delta)
569 delta = 1;
570 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
571 tcp_rcv_rtt_update(tp, delta_us, 0);
572 }
573}
574
575/*
576 * This function should be called every time data is copied to user space.
577 * It calculates the appropriate TCP receive buffer space.
578 */
579void tcp_rcv_space_adjust(struct sock *sk)
580{
581 struct tcp_sock *tp = tcp_sk(sk);
582 u32 copied;
583 int time;
584
585 tcp_mstamp_refresh(tp);
586 time = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcvq_space.time);
587 if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0)
588 return;
589
590 /* Number of bytes copied to user in last RTT */
591 copied = tp->copied_seq - tp->rcvq_space.seq;
592 if (copied <= tp->rcvq_space.space)
593 goto new_measure;
594
595 /* A bit of theory :
596 * copied = bytes received in previous RTT, our base window
597 * To cope with packet losses, we need a 2x factor
598 * To cope with slow start, and sender growing its cwin by 100 %
599 * every RTT, we need a 4x factor, because the ACK we are sending
600 * now is for the next RTT, not the current one :
601 * <prev RTT . ><current RTT .. ><next RTT .... >
602 */
603
604 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf &&
605 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
606 int rcvmem, rcvbuf;
607 u64 rcvwin, grow;
608
609 /* minimal window to cope with packet losses, assuming
610 * steady state. Add some cushion because of small variations.
611 */
612 rcvwin = ((u64)copied << 1) + 16 * tp->advmss;
613
614 /* Accommodate for sender rate increase (eg. slow start) */
615 grow = rcvwin * (copied - tp->rcvq_space.space);
616 do_div(grow, tp->rcvq_space.space);
617 rcvwin += (grow << 1);
618
619 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
620 while (tcp_win_from_space(sk, rcvmem) < tp->advmss)
621 rcvmem += 128;
622
623 do_div(rcvwin, tp->advmss);
624 rcvbuf = min_t(u64, rcvwin * rcvmem,
625 sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
626 if (rcvbuf > sk->sk_rcvbuf) {
627 sk->sk_rcvbuf = rcvbuf;
628
629 /* Make the window clamp follow along. */
630 tp->window_clamp = tcp_win_from_space(sk, rcvbuf);
631 }
632 }
633 tp->rcvq_space.space = copied;
634
635new_measure:
636 tp->rcvq_space.seq = tp->copied_seq;
637 tp->rcvq_space.time = tp->tcp_mstamp;
638}
639
640/* There is something which you must keep in mind when you analyze the
641 * behavior of the tp->ato delayed ack timeout interval. When a
642 * connection starts up, we want to ack as quickly as possible. The
643 * problem is that "good" TCP's do slow start at the beginning of data
644 * transmission. The means that until we send the first few ACK's the
645 * sender will sit on his end and only queue most of his data, because
646 * he can only send snd_cwnd unacked packets at any given time. For
647 * each ACK we send, he increments snd_cwnd and transmits more of his
648 * queue. -DaveM
649 */
650static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
651{
652 struct tcp_sock *tp = tcp_sk(sk);
653 struct inet_connection_sock *icsk = inet_csk(sk);
654 u32 now;
655
656 inet_csk_schedule_ack(sk);
657
658 tcp_measure_rcv_mss(sk, skb);
659
660 tcp_rcv_rtt_measure(tp);
661
662 now = tcp_jiffies32;
663
664 if (!icsk->icsk_ack.ato) {
665 /* The _first_ data packet received, initialize
666 * delayed ACK engine.
667 */
668 tcp_incr_quickack(sk);
669 icsk->icsk_ack.ato = TCP_ATO_MIN;
670 } else {
671 int m = now - icsk->icsk_ack.lrcvtime;
672
673 if (m <= TCP_ATO_MIN / 2) {
674 /* The fastest case is the first. */
675 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
676 } else if (m < icsk->icsk_ack.ato) {
677 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
678 if (icsk->icsk_ack.ato > icsk->icsk_rto)
679 icsk->icsk_ack.ato = icsk->icsk_rto;
680 } else if (m > icsk->icsk_rto) {
681 /* Too long gap. Apparently sender failed to
682 * restart window, so that we send ACKs quickly.
683 */
684 tcp_incr_quickack(sk);
685 sk_mem_reclaim(sk);
686 }
687 }
688 icsk->icsk_ack.lrcvtime = now;
689
690 tcp_ecn_check_ce(tp, skb);
691
692 if (skb->len >= 128)
693 tcp_grow_window(sk, skb);
694}
695
696/* Called to compute a smoothed rtt estimate. The data fed to this
697 * routine either comes from timestamps, or from segments that were
698 * known _not_ to have been retransmitted [see Karn/Partridge
699 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
700 * piece by Van Jacobson.
701 * NOTE: the next three routines used to be one big routine.
702 * To save cycles in the RFC 1323 implementation it was better to break
703 * it up into three procedures. -- erics
704 */
705static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
706{
707 struct tcp_sock *tp = tcp_sk(sk);
708 long m = mrtt_us; /* RTT */
709 u32 srtt = tp->srtt_us;
710
711 /* The following amusing code comes from Jacobson's
712 * article in SIGCOMM '88. Note that rtt and mdev
713 * are scaled versions of rtt and mean deviation.
714 * This is designed to be as fast as possible
715 * m stands for "measurement".
716 *
717 * On a 1990 paper the rto value is changed to:
718 * RTO = rtt + 4 * mdev
719 *
720 * Funny. This algorithm seems to be very broken.
721 * These formulae increase RTO, when it should be decreased, increase
722 * too slowly, when it should be increased quickly, decrease too quickly
723 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
724 * does not matter how to _calculate_ it. Seems, it was trap
725 * that VJ failed to avoid. 8)
726 */
727 if (srtt != 0) {
728 m -= (srtt >> 3); /* m is now error in rtt est */
729 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
730 if (m < 0) {
731 m = -m; /* m is now abs(error) */
732 m -= (tp->mdev_us >> 2); /* similar update on mdev */
733 /* This is similar to one of Eifel findings.
734 * Eifel blocks mdev updates when rtt decreases.
735 * This solution is a bit different: we use finer gain
736 * for mdev in this case (alpha*beta).
737 * Like Eifel it also prevents growth of rto,
738 * but also it limits too fast rto decreases,
739 * happening in pure Eifel.
740 */
741 if (m > 0)
742 m >>= 3;
743 } else {
744 m -= (tp->mdev_us >> 2); /* similar update on mdev */
745 }
746 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
747 if (tp->mdev_us > tp->mdev_max_us) {
748 tp->mdev_max_us = tp->mdev_us;
749 if (tp->mdev_max_us > tp->rttvar_us)
750 tp->rttvar_us = tp->mdev_max_us;
751 }
752 if (after(tp->snd_una, tp->rtt_seq)) {
753 if (tp->mdev_max_us < tp->rttvar_us)
754 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
755 tp->rtt_seq = tp->snd_nxt;
756 tp->mdev_max_us = tcp_rto_min_us(sk);
757 }
758 } else {
759 /* no previous measure. */
760 srtt = m << 3; /* take the measured time to be rtt */
761 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
762 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
763 tp->mdev_max_us = tp->rttvar_us;
764 tp->rtt_seq = tp->snd_nxt;
765 }
766 tp->srtt_us = max(1U, srtt);
767}
768
769static void tcp_update_pacing_rate(struct sock *sk)
770{
771 const struct tcp_sock *tp = tcp_sk(sk);
772 u64 rate;
773
774 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
775 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
776
777 /* current rate is (cwnd * mss) / srtt
778 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
779 * In Congestion Avoidance phase, set it to 120 % the current rate.
780 *
781 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
782 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
783 * end of slow start and should slow down.
784 */
785 if (tp->snd_cwnd < tp->snd_ssthresh / 2)
786 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ss_ratio;
787 else
788 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ca_ratio;
789
790 rate *= max(tp->snd_cwnd, tp->packets_out);
791
792 if (likely(tp->srtt_us))
793 do_div(rate, tp->srtt_us);
794
795 /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate
796 * without any lock. We want to make sure compiler wont store
797 * intermediate values in this location.
798 */
799 WRITE_ONCE(sk->sk_pacing_rate, min_t(u64, rate,
800 sk->sk_max_pacing_rate));
801}
802
803/* Calculate rto without backoff. This is the second half of Van Jacobson's
804 * routine referred to above.
805 */
806static void tcp_set_rto(struct sock *sk)
807{
808 const struct tcp_sock *tp = tcp_sk(sk);
809 /* Old crap is replaced with new one. 8)
810 *
811 * More seriously:
812 * 1. If rtt variance happened to be less 50msec, it is hallucination.
813 * It cannot be less due to utterly erratic ACK generation made
814 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
815 * to do with delayed acks, because at cwnd>2 true delack timeout
816 * is invisible. Actually, Linux-2.4 also generates erratic
817 * ACKs in some circumstances.
818 */
819 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
820
821 /* 2. Fixups made earlier cannot be right.
822 * If we do not estimate RTO correctly without them,
823 * all the algo is pure shit and should be replaced
824 * with correct one. It is exactly, which we pretend to do.
825 */
826
827 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
828 * guarantees that rto is higher.
829 */
830 tcp_bound_rto(sk);
831}
832
833__u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
834{
835 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
836
837 if (!cwnd)
838 cwnd = TCP_INIT_CWND;
839 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
840}
841
842/* Take a notice that peer is sending D-SACKs */
843static void tcp_dsack_seen(struct tcp_sock *tp)
844{
845 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
846 tp->rack.dsack_seen = 1;
847}
848
849/* It's reordering when higher sequence was delivered (i.e. sacked) before
850 * some lower never-retransmitted sequence ("low_seq"). The maximum reordering
851 * distance is approximated in full-mss packet distance ("reordering").
852 */
853static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq,
854 const int ts)
855{
856 struct tcp_sock *tp = tcp_sk(sk);
857 const u32 mss = tp->mss_cache;
858 u32 fack, metric;
859
860 fack = tcp_highest_sack_seq(tp);
861 if (!before(low_seq, fack))
862 return;
863
864 metric = fack - low_seq;
865 if ((metric > tp->reordering * mss) && mss) {
866#if FASTRETRANS_DEBUG > 1
867 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
868 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
869 tp->reordering,
870 0,
871 tp->sacked_out,
872 tp->undo_marker ? tp->undo_retrans : 0);
873#endif
874 tp->reordering = min_t(u32, (metric + mss - 1) / mss,
875 sock_net(sk)->ipv4.sysctl_tcp_max_reordering);
876 }
877
878 tp->rack.reord = 1;
879 /* This exciting event is worth to be remembered. 8) */
880 NET_INC_STATS(sock_net(sk),
881 ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER);
882}
883
884/* This must be called before lost_out is incremented */
885static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
886{
887 if (!tp->retransmit_skb_hint ||
888 before(TCP_SKB_CB(skb)->seq,
889 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
890 tp->retransmit_skb_hint = skb;
891}
892
893/* Sum the number of packets on the wire we have marked as lost.
894 * There are two cases we care about here:
895 * a) Packet hasn't been marked lost (nor retransmitted),
896 * and this is the first loss.
897 * b) Packet has been marked both lost and retransmitted,
898 * and this means we think it was lost again.
899 */
900static void tcp_sum_lost(struct tcp_sock *tp, struct sk_buff *skb)
901{
902 __u8 sacked = TCP_SKB_CB(skb)->sacked;
903
904 if (!(sacked & TCPCB_LOST) ||
905 ((sacked & TCPCB_LOST) && (sacked & TCPCB_SACKED_RETRANS)))
906 tp->lost += tcp_skb_pcount(skb);
907}
908
909static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
910{
911 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
912 tcp_verify_retransmit_hint(tp, skb);
913
914 tp->lost_out += tcp_skb_pcount(skb);
915 tcp_sum_lost(tp, skb);
916 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
917 }
918}
919
920void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, struct sk_buff *skb)
921{
922 tcp_verify_retransmit_hint(tp, skb);
923
924 tcp_sum_lost(tp, skb);
925 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
926 tp->lost_out += tcp_skb_pcount(skb);
927 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
928 }
929}
930
931/* This procedure tags the retransmission queue when SACKs arrive.
932 *
933 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
934 * Packets in queue with these bits set are counted in variables
935 * sacked_out, retrans_out and lost_out, correspondingly.
936 *
937 * Valid combinations are:
938 * Tag InFlight Description
939 * 0 1 - orig segment is in flight.
940 * S 0 - nothing flies, orig reached receiver.
941 * L 0 - nothing flies, orig lost by net.
942 * R 2 - both orig and retransmit are in flight.
943 * L|R 1 - orig is lost, retransmit is in flight.
944 * S|R 1 - orig reached receiver, retrans is still in flight.
945 * (L|S|R is logically valid, it could occur when L|R is sacked,
946 * but it is equivalent to plain S and code short-curcuits it to S.
947 * L|S is logically invalid, it would mean -1 packet in flight 8))
948 *
949 * These 6 states form finite state machine, controlled by the following events:
950 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
951 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
952 * 3. Loss detection event of two flavors:
953 * A. Scoreboard estimator decided the packet is lost.
954 * A'. Reno "three dupacks" marks head of queue lost.
955 * B. SACK arrives sacking SND.NXT at the moment, when the
956 * segment was retransmitted.
957 * 4. D-SACK added new rule: D-SACK changes any tag to S.
958 *
959 * It is pleasant to note, that state diagram turns out to be commutative,
960 * so that we are allowed not to be bothered by order of our actions,
961 * when multiple events arrive simultaneously. (see the function below).
962 *
963 * Reordering detection.
964 * --------------------
965 * Reordering metric is maximal distance, which a packet can be displaced
966 * in packet stream. With SACKs we can estimate it:
967 *
968 * 1. SACK fills old hole and the corresponding segment was not
969 * ever retransmitted -> reordering. Alas, we cannot use it
970 * when segment was retransmitted.
971 * 2. The last flaw is solved with D-SACK. D-SACK arrives
972 * for retransmitted and already SACKed segment -> reordering..
973 * Both of these heuristics are not used in Loss state, when we cannot
974 * account for retransmits accurately.
975 *
976 * SACK block validation.
977 * ----------------------
978 *
979 * SACK block range validation checks that the received SACK block fits to
980 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
981 * Note that SND.UNA is not included to the range though being valid because
982 * it means that the receiver is rather inconsistent with itself reporting
983 * SACK reneging when it should advance SND.UNA. Such SACK block this is
984 * perfectly valid, however, in light of RFC2018 which explicitly states
985 * that "SACK block MUST reflect the newest segment. Even if the newest
986 * segment is going to be discarded ...", not that it looks very clever
987 * in case of head skb. Due to potentional receiver driven attacks, we
988 * choose to avoid immediate execution of a walk in write queue due to
989 * reneging and defer head skb's loss recovery to standard loss recovery
990 * procedure that will eventually trigger (nothing forbids us doing this).
991 *
992 * Implements also blockage to start_seq wrap-around. Problem lies in the
993 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
994 * there's no guarantee that it will be before snd_nxt (n). The problem
995 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
996 * wrap (s_w):
997 *
998 * <- outs wnd -> <- wrapzone ->
999 * u e n u_w e_w s n_w
1000 * | | | | | | |
1001 * |<------------+------+----- TCP seqno space --------------+---------->|
1002 * ...-- <2^31 ->| |<--------...
1003 * ...---- >2^31 ------>| |<--------...
1004 *
1005 * Current code wouldn't be vulnerable but it's better still to discard such
1006 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1007 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1008 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1009 * equal to the ideal case (infinite seqno space without wrap caused issues).
1010 *
1011 * With D-SACK the lower bound is extended to cover sequence space below
1012 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1013 * again, D-SACK block must not to go across snd_una (for the same reason as
1014 * for the normal SACK blocks, explained above). But there all simplicity
1015 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1016 * fully below undo_marker they do not affect behavior in anyway and can
1017 * therefore be safely ignored. In rare cases (which are more or less
1018 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1019 * fragmentation and packet reordering past skb's retransmission. To consider
1020 * them correctly, the acceptable range must be extended even more though
1021 * the exact amount is rather hard to quantify. However, tp->max_window can
1022 * be used as an exaggerated estimate.
1023 */
1024static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1025 u32 start_seq, u32 end_seq)
1026{
1027 /* Too far in future, or reversed (interpretation is ambiguous) */
1028 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1029 return false;
1030
1031 /* Nasty start_seq wrap-around check (see comments above) */
1032 if (!before(start_seq, tp->snd_nxt))
1033 return false;
1034
1035 /* In outstanding window? ...This is valid exit for D-SACKs too.
1036 * start_seq == snd_una is non-sensical (see comments above)
1037 */
1038 if (after(start_seq, tp->snd_una))
1039 return true;
1040
1041 if (!is_dsack || !tp->undo_marker)
1042 return false;
1043
1044 /* ...Then it's D-SACK, and must reside below snd_una completely */
1045 if (after(end_seq, tp->snd_una))
1046 return false;
1047
1048 if (!before(start_seq, tp->undo_marker))
1049 return true;
1050
1051 /* Too old */
1052 if (!after(end_seq, tp->undo_marker))
1053 return false;
1054
1055 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1056 * start_seq < undo_marker and end_seq >= undo_marker.
1057 */
1058 return !before(start_seq, end_seq - tp->max_window);
1059}
1060
1061static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1062 struct tcp_sack_block_wire *sp, int num_sacks,
1063 u32 prior_snd_una)
1064{
1065 struct tcp_sock *tp = tcp_sk(sk);
1066 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1067 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1068 bool dup_sack = false;
1069
1070 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1071 dup_sack = true;
1072 tcp_dsack_seen(tp);
1073 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1074 } else if (num_sacks > 1) {
1075 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1076 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1077
1078 if (!after(end_seq_0, end_seq_1) &&
1079 !before(start_seq_0, start_seq_1)) {
1080 dup_sack = true;
1081 tcp_dsack_seen(tp);
1082 NET_INC_STATS(sock_net(sk),
1083 LINUX_MIB_TCPDSACKOFORECV);
1084 }
1085 }
1086
1087 /* D-SACK for already forgotten data... Do dumb counting. */
1088 if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 &&
1089 !after(end_seq_0, prior_snd_una) &&
1090 after(end_seq_0, tp->undo_marker))
1091 tp->undo_retrans--;
1092
1093 return dup_sack;
1094}
1095
1096struct tcp_sacktag_state {
1097 u32 reord;
1098 /* Timestamps for earliest and latest never-retransmitted segment
1099 * that was SACKed. RTO needs the earliest RTT to stay conservative,
1100 * but congestion control should still get an accurate delay signal.
1101 */
1102 u64 first_sackt;
1103 u64 last_sackt;
1104 struct rate_sample *rate;
1105 int flag;
1106 unsigned int mss_now;
1107};
1108
1109/* Check if skb is fully within the SACK block. In presence of GSO skbs,
1110 * the incoming SACK may not exactly match but we can find smaller MSS
1111 * aligned portion of it that matches. Therefore we might need to fragment
1112 * which may fail and creates some hassle (caller must handle error case
1113 * returns).
1114 *
1115 * FIXME: this could be merged to shift decision code
1116 */
1117static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1118 u32 start_seq, u32 end_seq)
1119{
1120 int err;
1121 bool in_sack;
1122 unsigned int pkt_len;
1123 unsigned int mss;
1124
1125 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1126 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1127
1128 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1129 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1130 mss = tcp_skb_mss(skb);
1131 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1132
1133 if (!in_sack) {
1134 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1135 if (pkt_len < mss)
1136 pkt_len = mss;
1137 } else {
1138 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1139 if (pkt_len < mss)
1140 return -EINVAL;
1141 }
1142
1143 /* Round if necessary so that SACKs cover only full MSSes
1144 * and/or the remaining small portion (if present)
1145 */
1146 if (pkt_len > mss) {
1147 unsigned int new_len = (pkt_len / mss) * mss;
1148 if (!in_sack && new_len < pkt_len)
1149 new_len += mss;
1150 pkt_len = new_len;
1151 }
1152
1153 if (pkt_len >= skb->len && !in_sack)
1154 return 0;
1155
1156 err = tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
1157 pkt_len, mss, GFP_ATOMIC);
1158 if (err < 0)
1159 return err;
1160 }
1161
1162 return in_sack;
1163}
1164
1165/* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1166static u8 tcp_sacktag_one(struct sock *sk,
1167 struct tcp_sacktag_state *state, u8 sacked,
1168 u32 start_seq, u32 end_seq,
1169 int dup_sack, int pcount,
1170 u64 xmit_time)
1171{
1172 struct tcp_sock *tp = tcp_sk(sk);
1173
1174 /* Account D-SACK for retransmitted packet. */
1175 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1176 if (tp->undo_marker && tp->undo_retrans > 0 &&
1177 after(end_seq, tp->undo_marker))
1178 tp->undo_retrans--;
1179 if ((sacked & TCPCB_SACKED_ACKED) &&
1180 before(start_seq, state->reord))
1181 state->reord = start_seq;
1182 }
1183
1184 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1185 if (!after(end_seq, tp->snd_una))
1186 return sacked;
1187
1188 if (!(sacked & TCPCB_SACKED_ACKED)) {
1189 tcp_rack_advance(tp, sacked, end_seq, xmit_time);
1190
1191 if (sacked & TCPCB_SACKED_RETRANS) {
1192 /* If the segment is not tagged as lost,
1193 * we do not clear RETRANS, believing
1194 * that retransmission is still in flight.
1195 */
1196 if (sacked & TCPCB_LOST) {
1197 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1198 tp->lost_out -= pcount;
1199 tp->retrans_out -= pcount;
1200 }
1201 } else {
1202 if (!(sacked & TCPCB_RETRANS)) {
1203 /* New sack for not retransmitted frame,
1204 * which was in hole. It is reordering.
1205 */
1206 if (before(start_seq,
1207 tcp_highest_sack_seq(tp)) &&
1208 before(start_seq, state->reord))
1209 state->reord = start_seq;
1210
1211 if (!after(end_seq, tp->high_seq))
1212 state->flag |= FLAG_ORIG_SACK_ACKED;
1213 if (state->first_sackt == 0)
1214 state->first_sackt = xmit_time;
1215 state->last_sackt = xmit_time;
1216 }
1217
1218 if (sacked & TCPCB_LOST) {
1219 sacked &= ~TCPCB_LOST;
1220 tp->lost_out -= pcount;
1221 }
1222 }
1223
1224 sacked |= TCPCB_SACKED_ACKED;
1225 state->flag |= FLAG_DATA_SACKED;
1226 tp->sacked_out += pcount;
1227 tp->delivered += pcount; /* Out-of-order packets delivered */
1228
1229 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1230 if (tp->lost_skb_hint &&
1231 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1232 tp->lost_cnt_hint += pcount;
1233 }
1234
1235 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1236 * frames and clear it. undo_retrans is decreased above, L|R frames
1237 * are accounted above as well.
1238 */
1239 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1240 sacked &= ~TCPCB_SACKED_RETRANS;
1241 tp->retrans_out -= pcount;
1242 }
1243
1244 return sacked;
1245}
1246
1247/* Shift newly-SACKed bytes from this skb to the immediately previous
1248 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1249 */
1250static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev,
1251 struct sk_buff *skb,
1252 struct tcp_sacktag_state *state,
1253 unsigned int pcount, int shifted, int mss,
1254 bool dup_sack)
1255{
1256 struct tcp_sock *tp = tcp_sk(sk);
1257 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1258 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1259
1260 BUG_ON(!pcount);
1261
1262 /* Adjust counters and hints for the newly sacked sequence
1263 * range but discard the return value since prev is already
1264 * marked. We must tag the range first because the seq
1265 * advancement below implicitly advances
1266 * tcp_highest_sack_seq() when skb is highest_sack.
1267 */
1268 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1269 start_seq, end_seq, dup_sack, pcount,
1270 skb->skb_mstamp);
1271 tcp_rate_skb_delivered(sk, skb, state->rate);
1272
1273 if (skb == tp->lost_skb_hint)
1274 tp->lost_cnt_hint += pcount;
1275
1276 TCP_SKB_CB(prev)->end_seq += shifted;
1277 TCP_SKB_CB(skb)->seq += shifted;
1278
1279 tcp_skb_pcount_add(prev, pcount);
1280 BUG_ON(tcp_skb_pcount(skb) < pcount);
1281 tcp_skb_pcount_add(skb, -pcount);
1282
1283 /* When we're adding to gso_segs == 1, gso_size will be zero,
1284 * in theory this shouldn't be necessary but as long as DSACK
1285 * code can come after this skb later on it's better to keep
1286 * setting gso_size to something.
1287 */
1288 if (!TCP_SKB_CB(prev)->tcp_gso_size)
1289 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1290
1291 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1292 if (tcp_skb_pcount(skb) <= 1)
1293 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1294
1295 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1296 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1297
1298 if (skb->len > 0) {
1299 BUG_ON(!tcp_skb_pcount(skb));
1300 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1301 return false;
1302 }
1303
1304 /* Whole SKB was eaten :-) */
1305
1306 if (skb == tp->retransmit_skb_hint)
1307 tp->retransmit_skb_hint = prev;
1308 if (skb == tp->lost_skb_hint) {
1309 tp->lost_skb_hint = prev;
1310 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1311 }
1312
1313 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1314 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor;
1315 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1316 TCP_SKB_CB(prev)->end_seq++;
1317
1318 if (skb == tcp_highest_sack(sk))
1319 tcp_advance_highest_sack(sk, skb);
1320
1321 tcp_skb_collapse_tstamp(prev, skb);
1322 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp))
1323 TCP_SKB_CB(prev)->tx.delivered_mstamp = 0;
1324
1325 tcp_rtx_queue_unlink_and_free(skb, sk);
1326
1327 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED);
1328
1329 return true;
1330}
1331
1332/* I wish gso_size would have a bit more sane initialization than
1333 * something-or-zero which complicates things
1334 */
1335static int tcp_skb_seglen(const struct sk_buff *skb)
1336{
1337 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1338}
1339
1340/* Shifting pages past head area doesn't work */
1341static int skb_can_shift(const struct sk_buff *skb)
1342{
1343 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1344}
1345
1346/* Try collapsing SACK blocks spanning across multiple skbs to a single
1347 * skb.
1348 */
1349static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1350 struct tcp_sacktag_state *state,
1351 u32 start_seq, u32 end_seq,
1352 bool dup_sack)
1353{
1354 struct tcp_sock *tp = tcp_sk(sk);
1355 struct sk_buff *prev;
1356 int mss;
1357 int pcount = 0;
1358 int len;
1359 int in_sack;
1360
1361 /* Normally R but no L won't result in plain S */
1362 if (!dup_sack &&
1363 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1364 goto fallback;
1365 if (!skb_can_shift(skb))
1366 goto fallback;
1367 /* This frame is about to be dropped (was ACKed). */
1368 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1369 goto fallback;
1370
1371 /* Can only happen with delayed DSACK + discard craziness */
1372 prev = skb_rb_prev(skb);
1373 if (!prev)
1374 goto fallback;
1375
1376 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1377 goto fallback;
1378
1379 if (!tcp_skb_can_collapse_to(prev))
1380 goto fallback;
1381
1382 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1383 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1384
1385 if (in_sack) {
1386 len = skb->len;
1387 pcount = tcp_skb_pcount(skb);
1388 mss = tcp_skb_seglen(skb);
1389
1390 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1391 * drop this restriction as unnecessary
1392 */
1393 if (mss != tcp_skb_seglen(prev))
1394 goto fallback;
1395 } else {
1396 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1397 goto noop;
1398 /* CHECKME: This is non-MSS split case only?, this will
1399 * cause skipped skbs due to advancing loop btw, original
1400 * has that feature too
1401 */
1402 if (tcp_skb_pcount(skb) <= 1)
1403 goto noop;
1404
1405 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1406 if (!in_sack) {
1407 /* TODO: head merge to next could be attempted here
1408 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1409 * though it might not be worth of the additional hassle
1410 *
1411 * ...we can probably just fallback to what was done
1412 * previously. We could try merging non-SACKed ones
1413 * as well but it probably isn't going to buy off
1414 * because later SACKs might again split them, and
1415 * it would make skb timestamp tracking considerably
1416 * harder problem.
1417 */
1418 goto fallback;
1419 }
1420
1421 len = end_seq - TCP_SKB_CB(skb)->seq;
1422 BUG_ON(len < 0);
1423 BUG_ON(len > skb->len);
1424
1425 /* MSS boundaries should be honoured or else pcount will
1426 * severely break even though it makes things bit trickier.
1427 * Optimize common case to avoid most of the divides
1428 */
1429 mss = tcp_skb_mss(skb);
1430
1431 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1432 * drop this restriction as unnecessary
1433 */
1434 if (mss != tcp_skb_seglen(prev))
1435 goto fallback;
1436
1437 if (len == mss) {
1438 pcount = 1;
1439 } else if (len < mss) {
1440 goto noop;
1441 } else {
1442 pcount = len / mss;
1443 len = pcount * mss;
1444 }
1445 }
1446
1447 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1448 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1449 goto fallback;
1450
1451 if (!skb_shift(prev, skb, len))
1452 goto fallback;
1453 if (!tcp_shifted_skb(sk, prev, skb, state, pcount, len, mss, dup_sack))
1454 goto out;
1455
1456 /* Hole filled allows collapsing with the next as well, this is very
1457 * useful when hole on every nth skb pattern happens
1458 */
1459 skb = skb_rb_next(prev);
1460 if (!skb)
1461 goto out;
1462
1463 if (!skb_can_shift(skb) ||
1464 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1465 (mss != tcp_skb_seglen(skb)))
1466 goto out;
1467
1468 len = skb->len;
1469 if (skb_shift(prev, skb, len)) {
1470 pcount += tcp_skb_pcount(skb);
1471 tcp_shifted_skb(sk, prev, skb, state, tcp_skb_pcount(skb),
1472 len, mss, 0);
1473 }
1474
1475out:
1476 return prev;
1477
1478noop:
1479 return skb;
1480
1481fallback:
1482 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1483 return NULL;
1484}
1485
1486static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1487 struct tcp_sack_block *next_dup,
1488 struct tcp_sacktag_state *state,
1489 u32 start_seq, u32 end_seq,
1490 bool dup_sack_in)
1491{
1492 struct tcp_sock *tp = tcp_sk(sk);
1493 struct sk_buff *tmp;
1494
1495 skb_rbtree_walk_from(skb) {
1496 int in_sack = 0;
1497 bool dup_sack = dup_sack_in;
1498
1499 /* queue is in-order => we can short-circuit the walk early */
1500 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1501 break;
1502
1503 if (next_dup &&
1504 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1505 in_sack = tcp_match_skb_to_sack(sk, skb,
1506 next_dup->start_seq,
1507 next_dup->end_seq);
1508 if (in_sack > 0)
1509 dup_sack = true;
1510 }
1511
1512 /* skb reference here is a bit tricky to get right, since
1513 * shifting can eat and free both this skb and the next,
1514 * so not even _safe variant of the loop is enough.
1515 */
1516 if (in_sack <= 0) {
1517 tmp = tcp_shift_skb_data(sk, skb, state,
1518 start_seq, end_seq, dup_sack);
1519 if (tmp) {
1520 if (tmp != skb) {
1521 skb = tmp;
1522 continue;
1523 }
1524
1525 in_sack = 0;
1526 } else {
1527 in_sack = tcp_match_skb_to_sack(sk, skb,
1528 start_seq,
1529 end_seq);
1530 }
1531 }
1532
1533 if (unlikely(in_sack < 0))
1534 break;
1535
1536 if (in_sack) {
1537 TCP_SKB_CB(skb)->sacked =
1538 tcp_sacktag_one(sk,
1539 state,
1540 TCP_SKB_CB(skb)->sacked,
1541 TCP_SKB_CB(skb)->seq,
1542 TCP_SKB_CB(skb)->end_seq,
1543 dup_sack,
1544 tcp_skb_pcount(skb),
1545 skb->skb_mstamp);
1546 tcp_rate_skb_delivered(sk, skb, state->rate);
1547 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1548 list_del_init(&skb->tcp_tsorted_anchor);
1549
1550 if (!before(TCP_SKB_CB(skb)->seq,
1551 tcp_highest_sack_seq(tp)))
1552 tcp_advance_highest_sack(sk, skb);
1553 }
1554 }
1555 return skb;
1556}
1557
1558static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk,
1559 struct tcp_sacktag_state *state,
1560 u32 seq)
1561{
1562 struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node;
1563 struct sk_buff *skb;
1564
1565 while (*p) {
1566 parent = *p;
1567 skb = rb_to_skb(parent);
1568 if (before(seq, TCP_SKB_CB(skb)->seq)) {
1569 p = &parent->rb_left;
1570 continue;
1571 }
1572 if (!before(seq, TCP_SKB_CB(skb)->end_seq)) {
1573 p = &parent->rb_right;
1574 continue;
1575 }
1576 return skb;
1577 }
1578 return NULL;
1579}
1580
1581static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1582 struct tcp_sacktag_state *state,
1583 u32 skip_to_seq)
1584{
1585 if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq))
1586 return skb;
1587
1588 return tcp_sacktag_bsearch(sk, state, skip_to_seq);
1589}
1590
1591static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1592 struct sock *sk,
1593 struct tcp_sack_block *next_dup,
1594 struct tcp_sacktag_state *state,
1595 u32 skip_to_seq)
1596{
1597 if (!next_dup)
1598 return skb;
1599
1600 if (before(next_dup->start_seq, skip_to_seq)) {
1601 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1602 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1603 next_dup->start_seq, next_dup->end_seq,
1604 1);
1605 }
1606
1607 return skb;
1608}
1609
1610static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1611{
1612 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1613}
1614
1615static int
1616tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1617 u32 prior_snd_una, struct tcp_sacktag_state *state)
1618{
1619 struct tcp_sock *tp = tcp_sk(sk);
1620 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1621 TCP_SKB_CB(ack_skb)->sacked);
1622 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1623 struct tcp_sack_block sp[TCP_NUM_SACKS];
1624 struct tcp_sack_block *cache;
1625 struct sk_buff *skb;
1626 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1627 int used_sacks;
1628 bool found_dup_sack = false;
1629 int i, j;
1630 int first_sack_index;
1631
1632 state->flag = 0;
1633 state->reord = tp->snd_nxt;
1634
1635 if (!tp->sacked_out)
1636 tcp_highest_sack_reset(sk);
1637
1638 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1639 num_sacks, prior_snd_una);
1640 if (found_dup_sack) {
1641 state->flag |= FLAG_DSACKING_ACK;
1642 tp->delivered++; /* A spurious retransmission is delivered */
1643 }
1644
1645 /* Eliminate too old ACKs, but take into
1646 * account more or less fresh ones, they can
1647 * contain valid SACK info.
1648 */
1649 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1650 return 0;
1651
1652 if (!tp->packets_out)
1653 goto out;
1654
1655 used_sacks = 0;
1656 first_sack_index = 0;
1657 for (i = 0; i < num_sacks; i++) {
1658 bool dup_sack = !i && found_dup_sack;
1659
1660 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1661 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1662
1663 if (!tcp_is_sackblock_valid(tp, dup_sack,
1664 sp[used_sacks].start_seq,
1665 sp[used_sacks].end_seq)) {
1666 int mib_idx;
1667
1668 if (dup_sack) {
1669 if (!tp->undo_marker)
1670 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1671 else
1672 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1673 } else {
1674 /* Don't count olds caused by ACK reordering */
1675 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1676 !after(sp[used_sacks].end_seq, tp->snd_una))
1677 continue;
1678 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1679 }
1680
1681 NET_INC_STATS(sock_net(sk), mib_idx);
1682 if (i == 0)
1683 first_sack_index = -1;
1684 continue;
1685 }
1686
1687 /* Ignore very old stuff early */
1688 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1689 continue;
1690
1691 used_sacks++;
1692 }
1693
1694 /* order SACK blocks to allow in order walk of the retrans queue */
1695 for (i = used_sacks - 1; i > 0; i--) {
1696 for (j = 0; j < i; j++) {
1697 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1698 swap(sp[j], sp[j + 1]);
1699
1700 /* Track where the first SACK block goes to */
1701 if (j == first_sack_index)
1702 first_sack_index = j + 1;
1703 }
1704 }
1705 }
1706
1707 state->mss_now = tcp_current_mss(sk);
1708 skb = NULL;
1709 i = 0;
1710
1711 if (!tp->sacked_out) {
1712 /* It's already past, so skip checking against it */
1713 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1714 } else {
1715 cache = tp->recv_sack_cache;
1716 /* Skip empty blocks in at head of the cache */
1717 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1718 !cache->end_seq)
1719 cache++;
1720 }
1721
1722 while (i < used_sacks) {
1723 u32 start_seq = sp[i].start_seq;
1724 u32 end_seq = sp[i].end_seq;
1725 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1726 struct tcp_sack_block *next_dup = NULL;
1727
1728 if (found_dup_sack && ((i + 1) == first_sack_index))
1729 next_dup = &sp[i + 1];
1730
1731 /* Skip too early cached blocks */
1732 while (tcp_sack_cache_ok(tp, cache) &&
1733 !before(start_seq, cache->end_seq))
1734 cache++;
1735
1736 /* Can skip some work by looking recv_sack_cache? */
1737 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1738 after(end_seq, cache->start_seq)) {
1739
1740 /* Head todo? */
1741 if (before(start_seq, cache->start_seq)) {
1742 skb = tcp_sacktag_skip(skb, sk, state,
1743 start_seq);
1744 skb = tcp_sacktag_walk(skb, sk, next_dup,
1745 state,
1746 start_seq,
1747 cache->start_seq,
1748 dup_sack);
1749 }
1750
1751 /* Rest of the block already fully processed? */
1752 if (!after(end_seq, cache->end_seq))
1753 goto advance_sp;
1754
1755 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1756 state,
1757 cache->end_seq);
1758
1759 /* ...tail remains todo... */
1760 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1761 /* ...but better entrypoint exists! */
1762 skb = tcp_highest_sack(sk);
1763 if (!skb)
1764 break;
1765 cache++;
1766 goto walk;
1767 }
1768
1769 skb = tcp_sacktag_skip(skb, sk, state, cache->end_seq);
1770 /* Check overlap against next cached too (past this one already) */
1771 cache++;
1772 continue;
1773 }
1774
1775 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1776 skb = tcp_highest_sack(sk);
1777 if (!skb)
1778 break;
1779 }
1780 skb = tcp_sacktag_skip(skb, sk, state, start_seq);
1781
1782walk:
1783 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
1784 start_seq, end_seq, dup_sack);
1785
1786advance_sp:
1787 i++;
1788 }
1789
1790 /* Clear the head of the cache sack blocks so we can skip it next time */
1791 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1792 tp->recv_sack_cache[i].start_seq = 0;
1793 tp->recv_sack_cache[i].end_seq = 0;
1794 }
1795 for (j = 0; j < used_sacks; j++)
1796 tp->recv_sack_cache[i++] = sp[j];
1797
1798 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss || tp->undo_marker)
1799 tcp_check_sack_reordering(sk, state->reord, 0);
1800
1801 tcp_verify_left_out(tp);
1802out:
1803
1804#if FASTRETRANS_DEBUG > 0
1805 WARN_ON((int)tp->sacked_out < 0);
1806 WARN_ON((int)tp->lost_out < 0);
1807 WARN_ON((int)tp->retrans_out < 0);
1808 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1809#endif
1810 return state->flag;
1811}
1812
1813/* Limits sacked_out so that sum with lost_out isn't ever larger than
1814 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1815 */
1816static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1817{
1818 u32 holes;
1819
1820 holes = max(tp->lost_out, 1U);
1821 holes = min(holes, tp->packets_out);
1822
1823 if ((tp->sacked_out + holes) > tp->packets_out) {
1824 tp->sacked_out = tp->packets_out - holes;
1825 return true;
1826 }
1827 return false;
1828}
1829
1830/* If we receive more dupacks than we expected counting segments
1831 * in assumption of absent reordering, interpret this as reordering.
1832 * The only another reason could be bug in receiver TCP.
1833 */
1834static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1835{
1836 struct tcp_sock *tp = tcp_sk(sk);
1837
1838 if (!tcp_limit_reno_sacked(tp))
1839 return;
1840
1841 tp->reordering = min_t(u32, tp->packets_out + addend,
1842 sock_net(sk)->ipv4.sysctl_tcp_max_reordering);
1843 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER);
1844}
1845
1846/* Emulate SACKs for SACKless connection: account for a new dupack. */
1847
1848static void tcp_add_reno_sack(struct sock *sk)
1849{
1850 struct tcp_sock *tp = tcp_sk(sk);
1851 u32 prior_sacked = tp->sacked_out;
1852
1853 tp->sacked_out++;
1854 tcp_check_reno_reordering(sk, 0);
1855 if (tp->sacked_out > prior_sacked)
1856 tp->delivered++; /* Some out-of-order packet is delivered */
1857 tcp_verify_left_out(tp);
1858}
1859
1860/* Account for ACK, ACKing some data in Reno Recovery phase. */
1861
1862static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1863{
1864 struct tcp_sock *tp = tcp_sk(sk);
1865
1866 if (acked > 0) {
1867 /* One ACK acked hole. The rest eat duplicate ACKs. */
1868 tp->delivered += max_t(int, acked - tp->sacked_out, 1);
1869 if (acked - 1 >= tp->sacked_out)
1870 tp->sacked_out = 0;
1871 else
1872 tp->sacked_out -= acked - 1;
1873 }
1874 tcp_check_reno_reordering(sk, acked);
1875 tcp_verify_left_out(tp);
1876}
1877
1878static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1879{
1880 tp->sacked_out = 0;
1881}
1882
1883void tcp_clear_retrans(struct tcp_sock *tp)
1884{
1885 tp->retrans_out = 0;
1886 tp->lost_out = 0;
1887 tp->undo_marker = 0;
1888 tp->undo_retrans = -1;
1889 tp->sacked_out = 0;
1890}
1891
1892static inline void tcp_init_undo(struct tcp_sock *tp)
1893{
1894 tp->undo_marker = tp->snd_una;
1895 /* Retransmission still in flight may cause DSACKs later. */
1896 tp->undo_retrans = tp->retrans_out ? : -1;
1897}
1898
1899/* Enter Loss state. If we detect SACK reneging, forget all SACK information
1900 * and reset tags completely, otherwise preserve SACKs. If receiver
1901 * dropped its ofo queue, we will know this due to reneging detection.
1902 */
1903void tcp_enter_loss(struct sock *sk)
1904{
1905 const struct inet_connection_sock *icsk = inet_csk(sk);
1906 struct tcp_sock *tp = tcp_sk(sk);
1907 struct net *net = sock_net(sk);
1908 struct sk_buff *skb;
1909 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
1910 bool is_reneg; /* is receiver reneging on SACKs? */
1911 bool mark_lost;
1912
1913 /* Reduce ssthresh if it has not yet been made inside this window. */
1914 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1915 !after(tp->high_seq, tp->snd_una) ||
1916 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1917 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1918 tp->prior_cwnd = tp->snd_cwnd;
1919 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1920 tcp_ca_event(sk, CA_EVENT_LOSS);
1921 tcp_init_undo(tp);
1922 }
1923 tp->snd_cwnd = 1;
1924 tp->snd_cwnd_cnt = 0;
1925 tp->snd_cwnd_stamp = tcp_jiffies32;
1926
1927 tp->retrans_out = 0;
1928 tp->lost_out = 0;
1929
1930 if (tcp_is_reno(tp))
1931 tcp_reset_reno_sack(tp);
1932
1933 skb = tcp_rtx_queue_head(sk);
1934 is_reneg = skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED);
1935 if (is_reneg) {
1936 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
1937 tp->sacked_out = 0;
1938 /* Mark SACK reneging until we recover from this loss event. */
1939 tp->is_sack_reneg = 1;
1940 }
1941 tcp_clear_all_retrans_hints(tp);
1942
1943 skb_rbtree_walk_from(skb) {
1944 mark_lost = (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
1945 is_reneg);
1946 if (mark_lost)
1947 tcp_sum_lost(tp, skb);
1948 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1949 if (mark_lost) {
1950 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1951 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1952 tp->lost_out += tcp_skb_pcount(skb);
1953 }
1954 }
1955 tcp_verify_left_out(tp);
1956
1957 /* Timeout in disordered state after receiving substantial DUPACKs
1958 * suggests that the degree of reordering is over-estimated.
1959 */
1960 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
1961 tp->sacked_out >= net->ipv4.sysctl_tcp_reordering)
1962 tp->reordering = min_t(unsigned int, tp->reordering,
1963 net->ipv4.sysctl_tcp_reordering);
1964 tcp_set_ca_state(sk, TCP_CA_Loss);
1965 tp->high_seq = tp->snd_nxt;
1966 tcp_ecn_queue_cwr(tp);
1967
1968 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1969 * loss recovery is underway except recurring timeout(s) on
1970 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1971 */
1972 tp->frto = net->ipv4.sysctl_tcp_frto &&
1973 (new_recovery || icsk->icsk_retransmits) &&
1974 !inet_csk(sk)->icsk_mtup.probe_size;
1975}
1976
1977/* If ACK arrived pointing to a remembered SACK, it means that our
1978 * remembered SACKs do not reflect real state of receiver i.e.
1979 * receiver _host_ is heavily congested (or buggy).
1980 *
1981 * To avoid big spurious retransmission bursts due to transient SACK
1982 * scoreboard oddities that look like reneging, we give the receiver a
1983 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
1984 * restore sanity to the SACK scoreboard. If the apparent reneging
1985 * persists until this RTO then we'll clear the SACK scoreboard.
1986 */
1987static bool tcp_check_sack_reneging(struct sock *sk, int flag)
1988{
1989 if (flag & FLAG_SACK_RENEGING) {
1990 struct tcp_sock *tp = tcp_sk(sk);
1991 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
1992 msecs_to_jiffies(10));
1993
1994 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1995 delay, TCP_RTO_MAX);
1996 return true;
1997 }
1998 return false;
1999}
2000
2001/* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2002 * counter when SACK is enabled (without SACK, sacked_out is used for
2003 * that purpose).
2004 *
2005 * With reordering, holes may still be in flight, so RFC3517 recovery
2006 * uses pure sacked_out (total number of SACKed segments) even though
2007 * it violates the RFC that uses duplicate ACKs, often these are equal
2008 * but when e.g. out-of-window ACKs or packet duplication occurs,
2009 * they differ. Since neither occurs due to loss, TCP should really
2010 * ignore them.
2011 */
2012static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2013{
2014 return tp->sacked_out + 1;
2015}
2016
2017/* Linux NewReno/SACK/ECN state machine.
2018 * --------------------------------------
2019 *
2020 * "Open" Normal state, no dubious events, fast path.
2021 * "Disorder" In all the respects it is "Open",
2022 * but requires a bit more attention. It is entered when
2023 * we see some SACKs or dupacks. It is split of "Open"
2024 * mainly to move some processing from fast path to slow one.
2025 * "CWR" CWND was reduced due to some Congestion Notification event.
2026 * It can be ECN, ICMP source quench, local device congestion.
2027 * "Recovery" CWND was reduced, we are fast-retransmitting.
2028 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2029 *
2030 * tcp_fastretrans_alert() is entered:
2031 * - each incoming ACK, if state is not "Open"
2032 * - when arrived ACK is unusual, namely:
2033 * * SACK
2034 * * Duplicate ACK.
2035 * * ECN ECE.
2036 *
2037 * Counting packets in flight is pretty simple.
2038 *
2039 * in_flight = packets_out - left_out + retrans_out
2040 *
2041 * packets_out is SND.NXT-SND.UNA counted in packets.
2042 *
2043 * retrans_out is number of retransmitted segments.
2044 *
2045 * left_out is number of segments left network, but not ACKed yet.
2046 *
2047 * left_out = sacked_out + lost_out
2048 *
2049 * sacked_out: Packets, which arrived to receiver out of order
2050 * and hence not ACKed. With SACKs this number is simply
2051 * amount of SACKed data. Even without SACKs
2052 * it is easy to give pretty reliable estimate of this number,
2053 * counting duplicate ACKs.
2054 *
2055 * lost_out: Packets lost by network. TCP has no explicit
2056 * "loss notification" feedback from network (for now).
2057 * It means that this number can be only _guessed_.
2058 * Actually, it is the heuristics to predict lossage that
2059 * distinguishes different algorithms.
2060 *
2061 * F.e. after RTO, when all the queue is considered as lost,
2062 * lost_out = packets_out and in_flight = retrans_out.
2063 *
2064 * Essentially, we have now a few algorithms detecting
2065 * lost packets.
2066 *
2067 * If the receiver supports SACK:
2068 *
2069 * RFC6675/3517: It is the conventional algorithm. A packet is
2070 * considered lost if the number of higher sequence packets
2071 * SACKed is greater than or equal the DUPACK thoreshold
2072 * (reordering). This is implemented in tcp_mark_head_lost and
2073 * tcp_update_scoreboard.
2074 *
2075 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm
2076 * (2017-) that checks timing instead of counting DUPACKs.
2077 * Essentially a packet is considered lost if it's not S/ACKed
2078 * after RTT + reordering_window, where both metrics are
2079 * dynamically measured and adjusted. This is implemented in
2080 * tcp_rack_mark_lost.
2081 *
2082 * If the receiver does not support SACK:
2083 *
2084 * NewReno (RFC6582): in Recovery we assume that one segment
2085 * is lost (classic Reno). While we are in Recovery and
2086 * a partial ACK arrives, we assume that one more packet
2087 * is lost (NewReno). This heuristics are the same in NewReno
2088 * and SACK.
2089 *
2090 * Really tricky (and requiring careful tuning) part of algorithm
2091 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2092 * The first determines the moment _when_ we should reduce CWND and,
2093 * hence, slow down forward transmission. In fact, it determines the moment
2094 * when we decide that hole is caused by loss, rather than by a reorder.
2095 *
2096 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2097 * holes, caused by lost packets.
2098 *
2099 * And the most logically complicated part of algorithm is undo
2100 * heuristics. We detect false retransmits due to both too early
2101 * fast retransmit (reordering) and underestimated RTO, analyzing
2102 * timestamps and D-SACKs. When we detect that some segments were
2103 * retransmitted by mistake and CWND reduction was wrong, we undo
2104 * window reduction and abort recovery phase. This logic is hidden
2105 * inside several functions named tcp_try_undo_<something>.
2106 */
2107
2108/* This function decides, when we should leave Disordered state
2109 * and enter Recovery phase, reducing congestion window.
2110 *
2111 * Main question: may we further continue forward transmission
2112 * with the same cwnd?
2113 */
2114static bool tcp_time_to_recover(struct sock *sk, int flag)
2115{
2116 struct tcp_sock *tp = tcp_sk(sk);
2117
2118 /* Trick#1: The loss is proven. */
2119 if (tp->lost_out)
2120 return true;
2121
2122 /* Not-A-Trick#2 : Classic rule... */
2123 if (tcp_dupack_heuristics(tp) > tp->reordering)
2124 return true;
2125
2126 return false;
2127}
2128
2129/* Detect loss in event "A" above by marking head of queue up as lost.
2130 * For non-SACK(Reno) senders, the first "packets" number of segments
2131 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2132 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2133 * the maximum SACKed segments to pass before reaching this limit.
2134 */
2135static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2136{
2137 struct tcp_sock *tp = tcp_sk(sk);
2138 struct sk_buff *skb;
2139 int cnt, oldcnt, lost;
2140 unsigned int mss;
2141 /* Use SACK to deduce losses of new sequences sent during recovery */
2142 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq;
2143
2144 WARN_ON(packets > tp->packets_out);
2145 skb = tp->lost_skb_hint;
2146 if (skb) {
2147 /* Head already handled? */
2148 if (mark_head && after(TCP_SKB_CB(skb)->seq, tp->snd_una))
2149 return;
2150 cnt = tp->lost_cnt_hint;
2151 } else {
2152 skb = tcp_rtx_queue_head(sk);
2153 cnt = 0;
2154 }
2155
2156 skb_rbtree_walk_from(skb) {
2157 /* TODO: do this better */
2158 /* this is not the most efficient way to do this... */
2159 tp->lost_skb_hint = skb;
2160 tp->lost_cnt_hint = cnt;
2161
2162 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2163 break;
2164
2165 oldcnt = cnt;
2166 if (tcp_is_reno(tp) ||
2167 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2168 cnt += tcp_skb_pcount(skb);
2169
2170 if (cnt > packets) {
2171 if (tcp_is_sack(tp) ||
2172 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2173 (oldcnt >= packets))
2174 break;
2175
2176 mss = tcp_skb_mss(skb);
2177 /* If needed, chop off the prefix to mark as lost. */
2178 lost = (packets - oldcnt) * mss;
2179 if (lost < skb->len &&
2180 tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2181 lost, mss, GFP_ATOMIC) < 0)
2182 break;
2183 cnt = packets;
2184 }
2185
2186 tcp_skb_mark_lost(tp, skb);
2187
2188 if (mark_head)
2189 break;
2190 }
2191 tcp_verify_left_out(tp);
2192}
2193
2194/* Account newly detected lost packet(s) */
2195
2196static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2197{
2198 struct tcp_sock *tp = tcp_sk(sk);
2199
2200 if (tcp_is_reno(tp)) {
2201 tcp_mark_head_lost(sk, 1, 1);
2202 } else {
2203 int sacked_upto = tp->sacked_out - tp->reordering;
2204 if (sacked_upto >= 0)
2205 tcp_mark_head_lost(sk, sacked_upto, 0);
2206 else if (fast_rexmit)
2207 tcp_mark_head_lost(sk, 1, 1);
2208 }
2209}
2210
2211static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2212{
2213 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2214 before(tp->rx_opt.rcv_tsecr, when);
2215}
2216
2217/* skb is spurious retransmitted if the returned timestamp echo
2218 * reply is prior to the skb transmission time
2219 */
2220static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2221 const struct sk_buff *skb)
2222{
2223 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2224 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp(skb));
2225}
2226
2227/* Nothing was retransmitted or returned timestamp is less
2228 * than timestamp of the first retransmission.
2229 */
2230static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2231{
2232 return !tp->retrans_stamp ||
2233 tcp_tsopt_ecr_before(tp, tp->retrans_stamp);
2234}
2235
2236/* Undo procedures. */
2237
2238/* We can clear retrans_stamp when there are no retransmissions in the
2239 * window. It would seem that it is trivially available for us in
2240 * tp->retrans_out, however, that kind of assumptions doesn't consider
2241 * what will happen if errors occur when sending retransmission for the
2242 * second time. ...It could the that such segment has only
2243 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2244 * the head skb is enough except for some reneging corner cases that
2245 * are not worth the effort.
2246 *
2247 * Main reason for all this complexity is the fact that connection dying
2248 * time now depends on the validity of the retrans_stamp, in particular,
2249 * that successive retransmissions of a segment must not advance
2250 * retrans_stamp under any conditions.
2251 */
2252static bool tcp_any_retrans_done(const struct sock *sk)
2253{
2254 const struct tcp_sock *tp = tcp_sk(sk);
2255 struct sk_buff *skb;
2256
2257 if (tp->retrans_out)
2258 return true;
2259
2260 skb = tcp_rtx_queue_head(sk);
2261 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2262 return true;
2263
2264 return false;
2265}
2266
2267static void DBGUNDO(struct sock *sk, const char *msg)
2268{
2269#if FASTRETRANS_DEBUG > 1
2270 struct tcp_sock *tp = tcp_sk(sk);
2271 struct inet_sock *inet = inet_sk(sk);
2272
2273 if (sk->sk_family == AF_INET) {
2274 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2275 msg,
2276 &inet->inet_daddr, ntohs(inet->inet_dport),
2277 tp->snd_cwnd, tcp_left_out(tp),
2278 tp->snd_ssthresh, tp->prior_ssthresh,
2279 tp->packets_out);
2280 }
2281#if IS_ENABLED(CONFIG_IPV6)
2282 else if (sk->sk_family == AF_INET6) {
2283 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2284 msg,
2285 &sk->sk_v6_daddr, ntohs(inet->inet_dport),
2286 tp->snd_cwnd, tcp_left_out(tp),
2287 tp->snd_ssthresh, tp->prior_ssthresh,
2288 tp->packets_out);
2289 }
2290#endif
2291#endif
2292}
2293
2294static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2295{
2296 struct tcp_sock *tp = tcp_sk(sk);
2297
2298 if (unmark_loss) {
2299 struct sk_buff *skb;
2300
2301 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2302 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2303 }
2304 tp->lost_out = 0;
2305 tcp_clear_all_retrans_hints(tp);
2306 }
2307
2308 if (tp->prior_ssthresh) {
2309 const struct inet_connection_sock *icsk = inet_csk(sk);
2310
2311 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2312
2313 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2314 tp->snd_ssthresh = tp->prior_ssthresh;
2315 tcp_ecn_withdraw_cwr(tp);
2316 }
2317 }
2318 tp->snd_cwnd_stamp = tcp_jiffies32;
2319 tp->undo_marker = 0;
2320 tp->rack.advanced = 1; /* Force RACK to re-exam losses */
2321}
2322
2323static inline bool tcp_may_undo(const struct tcp_sock *tp)
2324{
2325 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2326}
2327
2328/* People celebrate: "We love our President!" */
2329static bool tcp_try_undo_recovery(struct sock *sk)
2330{
2331 struct tcp_sock *tp = tcp_sk(sk);
2332
2333 if (tcp_may_undo(tp)) {
2334 int mib_idx;
2335
2336 /* Happy end! We did not retransmit anything
2337 * or our original transmission succeeded.
2338 */
2339 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2340 tcp_undo_cwnd_reduction(sk, false);
2341 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2342 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2343 else
2344 mib_idx = LINUX_MIB_TCPFULLUNDO;
2345
2346 NET_INC_STATS(sock_net(sk), mib_idx);
2347 } else if (tp->rack.reo_wnd_persist) {
2348 tp->rack.reo_wnd_persist--;
2349 }
2350 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2351 /* Hold old state until something *above* high_seq
2352 * is ACKed. For Reno it is MUST to prevent false
2353 * fast retransmits (RFC2582). SACK TCP is safe. */
2354 if (!tcp_any_retrans_done(sk))
2355 tp->retrans_stamp = 0;
2356 return true;
2357 }
2358 tcp_set_ca_state(sk, TCP_CA_Open);
2359 tp->is_sack_reneg = 0;
2360 return false;
2361}
2362
2363/* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2364static bool tcp_try_undo_dsack(struct sock *sk)
2365{
2366 struct tcp_sock *tp = tcp_sk(sk);
2367
2368 if (tp->undo_marker && !tp->undo_retrans) {
2369 tp->rack.reo_wnd_persist = min(TCP_RACK_RECOVERY_THRESH,
2370 tp->rack.reo_wnd_persist + 1);
2371 DBGUNDO(sk, "D-SACK");
2372 tcp_undo_cwnd_reduction(sk, false);
2373 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2374 return true;
2375 }
2376 return false;
2377}
2378
2379/* Undo during loss recovery after partial ACK or using F-RTO. */
2380static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2381{
2382 struct tcp_sock *tp = tcp_sk(sk);
2383
2384 if (frto_undo || tcp_may_undo(tp)) {
2385 tcp_undo_cwnd_reduction(sk, true);
2386
2387 DBGUNDO(sk, "partial loss");
2388 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2389 if (frto_undo)
2390 NET_INC_STATS(sock_net(sk),
2391 LINUX_MIB_TCPSPURIOUSRTOS);
2392 inet_csk(sk)->icsk_retransmits = 0;
2393 if (frto_undo || tcp_is_sack(tp)) {
2394 tcp_set_ca_state(sk, TCP_CA_Open);
2395 tp->is_sack_reneg = 0;
2396 }
2397 return true;
2398 }
2399 return false;
2400}
2401
2402/* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2403 * It computes the number of packets to send (sndcnt) based on packets newly
2404 * delivered:
2405 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2406 * cwnd reductions across a full RTT.
2407 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2408 * But when the retransmits are acked without further losses, PRR
2409 * slow starts cwnd up to ssthresh to speed up the recovery.
2410 */
2411static void tcp_init_cwnd_reduction(struct sock *sk)
2412{
2413 struct tcp_sock *tp = tcp_sk(sk);
2414
2415 tp->high_seq = tp->snd_nxt;
2416 tp->tlp_high_seq = 0;
2417 tp->snd_cwnd_cnt = 0;
2418 tp->prior_cwnd = tp->snd_cwnd;
2419 tp->prr_delivered = 0;
2420 tp->prr_out = 0;
2421 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2422 tcp_ecn_queue_cwr(tp);
2423}
2424
2425void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int flag)
2426{
2427 struct tcp_sock *tp = tcp_sk(sk);
2428 int sndcnt = 0;
2429 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2430
2431 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2432 return;
2433
2434 tp->prr_delivered += newly_acked_sacked;
2435 if (delta < 0) {
2436 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2437 tp->prior_cwnd - 1;
2438 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2439 } else if ((flag & FLAG_RETRANS_DATA_ACKED) &&
2440 !(flag & FLAG_LOST_RETRANS)) {
2441 sndcnt = min_t(int, delta,
2442 max_t(int, tp->prr_delivered - tp->prr_out,
2443 newly_acked_sacked) + 1);
2444 } else {
2445 sndcnt = min(delta, newly_acked_sacked);
2446 }
2447 /* Force a fast retransmit upon entering fast recovery */
2448 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
2449 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2450}
2451
2452static inline void tcp_end_cwnd_reduction(struct sock *sk)
2453{
2454 struct tcp_sock *tp = tcp_sk(sk);
2455
2456 if (inet_csk(sk)->icsk_ca_ops->cong_control)
2457 return;
2458
2459 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2460 if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH &&
2461 (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) {
2462 tp->snd_cwnd = tp->snd_ssthresh;
2463 tp->snd_cwnd_stamp = tcp_jiffies32;
2464 }
2465 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2466}
2467
2468/* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2469void tcp_enter_cwr(struct sock *sk)
2470{
2471 struct tcp_sock *tp = tcp_sk(sk);
2472
2473 tp->prior_ssthresh = 0;
2474 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2475 tp->undo_marker = 0;
2476 tcp_init_cwnd_reduction(sk);
2477 tcp_set_ca_state(sk, TCP_CA_CWR);
2478 }
2479}
2480EXPORT_SYMBOL(tcp_enter_cwr);
2481
2482static void tcp_try_keep_open(struct sock *sk)
2483{
2484 struct tcp_sock *tp = tcp_sk(sk);
2485 int state = TCP_CA_Open;
2486
2487 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2488 state = TCP_CA_Disorder;
2489
2490 if (inet_csk(sk)->icsk_ca_state != state) {
2491 tcp_set_ca_state(sk, state);
2492 tp->high_seq = tp->snd_nxt;
2493 }
2494}
2495
2496static void tcp_try_to_open(struct sock *sk, int flag)
2497{
2498 struct tcp_sock *tp = tcp_sk(sk);
2499
2500 tcp_verify_left_out(tp);
2501
2502 if (!tcp_any_retrans_done(sk))
2503 tp->retrans_stamp = 0;
2504
2505 if (flag & FLAG_ECE)
2506 tcp_enter_cwr(sk);
2507
2508 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2509 tcp_try_keep_open(sk);
2510 }
2511}
2512
2513static void tcp_mtup_probe_failed(struct sock *sk)
2514{
2515 struct inet_connection_sock *icsk = inet_csk(sk);
2516
2517 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2518 icsk->icsk_mtup.probe_size = 0;
2519 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
2520}
2521
2522static void tcp_mtup_probe_success(struct sock *sk)
2523{
2524 struct tcp_sock *tp = tcp_sk(sk);
2525 struct inet_connection_sock *icsk = inet_csk(sk);
2526
2527 /* FIXME: breaks with very large cwnd */
2528 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2529 tp->snd_cwnd = tp->snd_cwnd *
2530 tcp_mss_to_mtu(sk, tp->mss_cache) /
2531 icsk->icsk_mtup.probe_size;
2532 tp->snd_cwnd_cnt = 0;
2533 tp->snd_cwnd_stamp = tcp_jiffies32;
2534 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2535
2536 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2537 icsk->icsk_mtup.probe_size = 0;
2538 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2539 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
2540}
2541
2542/* Do a simple retransmit without using the backoff mechanisms in
2543 * tcp_timer. This is used for path mtu discovery.
2544 * The socket is already locked here.
2545 */
2546void tcp_simple_retransmit(struct sock *sk)
2547{
2548 const struct inet_connection_sock *icsk = inet_csk(sk);
2549 struct tcp_sock *tp = tcp_sk(sk);
2550 struct sk_buff *skb;
2551 unsigned int mss = tcp_current_mss(sk);
2552
2553 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2554 if (tcp_skb_seglen(skb) > mss &&
2555 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2556 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2557 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2558 tp->retrans_out -= tcp_skb_pcount(skb);
2559 }
2560 tcp_skb_mark_lost_uncond_verify(tp, skb);
2561 }
2562 }
2563
2564 tcp_clear_retrans_hints_partial(tp);
2565
2566 if (!tp->lost_out)
2567 return;
2568
2569 if (tcp_is_reno(tp))
2570 tcp_limit_reno_sacked(tp);
2571
2572 tcp_verify_left_out(tp);
2573
2574 /* Don't muck with the congestion window here.
2575 * Reason is that we do not increase amount of _data_
2576 * in network, but units changed and effective
2577 * cwnd/ssthresh really reduced now.
2578 */
2579 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2580 tp->high_seq = tp->snd_nxt;
2581 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2582 tp->prior_ssthresh = 0;
2583 tp->undo_marker = 0;
2584 tcp_set_ca_state(sk, TCP_CA_Loss);
2585 }
2586 tcp_xmit_retransmit_queue(sk);
2587}
2588EXPORT_SYMBOL(tcp_simple_retransmit);
2589
2590void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2591{
2592 struct tcp_sock *tp = tcp_sk(sk);
2593 int mib_idx;
2594
2595 if (tcp_is_reno(tp))
2596 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2597 else
2598 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2599
2600 NET_INC_STATS(sock_net(sk), mib_idx);
2601
2602 tp->prior_ssthresh = 0;
2603 tcp_init_undo(tp);
2604
2605 if (!tcp_in_cwnd_reduction(sk)) {
2606 if (!ece_ack)
2607 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2608 tcp_init_cwnd_reduction(sk);
2609 }
2610 tcp_set_ca_state(sk, TCP_CA_Recovery);
2611}
2612
2613/* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2614 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2615 */
2616static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack,
2617 int *rexmit)
2618{
2619 struct tcp_sock *tp = tcp_sk(sk);
2620 bool recovered = !before(tp->snd_una, tp->high_seq);
2621
2622 if ((flag & FLAG_SND_UNA_ADVANCED) &&
2623 tcp_try_undo_loss(sk, false))
2624 return;
2625
2626 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2627 /* Step 3.b. A timeout is spurious if not all data are
2628 * lost, i.e., never-retransmitted data are (s)acked.
2629 */
2630 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2631 tcp_try_undo_loss(sk, true))
2632 return;
2633
2634 if (after(tp->snd_nxt, tp->high_seq)) {
2635 if (flag & FLAG_DATA_SACKED || is_dupack)
2636 tp->frto = 0; /* Step 3.a. loss was real */
2637 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2638 tp->high_seq = tp->snd_nxt;
2639 /* Step 2.b. Try send new data (but deferred until cwnd
2640 * is updated in tcp_ack()). Otherwise fall back to
2641 * the conventional recovery.
2642 */
2643 if (!tcp_write_queue_empty(sk) &&
2644 after(tcp_wnd_end(tp), tp->snd_nxt)) {
2645 *rexmit = REXMIT_NEW;
2646 return;
2647 }
2648 tp->frto = 0;
2649 }
2650 }
2651
2652 if (recovered) {
2653 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2654 tcp_try_undo_recovery(sk);
2655 return;
2656 }
2657 if (tcp_is_reno(tp)) {
2658 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2659 * delivered. Lower inflight to clock out (re)tranmissions.
2660 */
2661 if (after(tp->snd_nxt, tp->high_seq) && is_dupack)
2662 tcp_add_reno_sack(sk);
2663 else if (flag & FLAG_SND_UNA_ADVANCED)
2664 tcp_reset_reno_sack(tp);
2665 }
2666 *rexmit = REXMIT_LOST;
2667}
2668
2669/* Undo during fast recovery after partial ACK. */
2670static bool tcp_try_undo_partial(struct sock *sk, u32 prior_snd_una)
2671{
2672 struct tcp_sock *tp = tcp_sk(sk);
2673
2674 if (tp->undo_marker && tcp_packet_delayed(tp)) {
2675 /* Plain luck! Hole if filled with delayed
2676 * packet, rather than with a retransmit. Check reordering.
2677 */
2678 tcp_check_sack_reordering(sk, prior_snd_una, 1);
2679
2680 /* We are getting evidence that the reordering degree is higher
2681 * than we realized. If there are no retransmits out then we
2682 * can undo. Otherwise we clock out new packets but do not
2683 * mark more packets lost or retransmit more.
2684 */
2685 if (tp->retrans_out)
2686 return true;
2687
2688 if (!tcp_any_retrans_done(sk))
2689 tp->retrans_stamp = 0;
2690
2691 DBGUNDO(sk, "partial recovery");
2692 tcp_undo_cwnd_reduction(sk, true);
2693 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2694 tcp_try_keep_open(sk);
2695 return true;
2696 }
2697 return false;
2698}
2699
2700static void tcp_rack_identify_loss(struct sock *sk, int *ack_flag)
2701{
2702 struct tcp_sock *tp = tcp_sk(sk);
2703
2704 /* Use RACK to detect loss */
2705 if (sock_net(sk)->ipv4.sysctl_tcp_recovery & TCP_RACK_LOSS_DETECTION) {
2706 u32 prior_retrans = tp->retrans_out;
2707
2708 tcp_rack_mark_lost(sk);
2709 if (prior_retrans > tp->retrans_out)
2710 *ack_flag |= FLAG_LOST_RETRANS;
2711 }
2712}
2713
2714static bool tcp_force_fast_retransmit(struct sock *sk)
2715{
2716 struct tcp_sock *tp = tcp_sk(sk);
2717
2718 return after(tcp_highest_sack_seq(tp),
2719 tp->snd_una + tp->reordering * tp->mss_cache);
2720}
2721
2722/* Process an event, which can update packets-in-flight not trivially.
2723 * Main goal of this function is to calculate new estimate for left_out,
2724 * taking into account both packets sitting in receiver's buffer and
2725 * packets lost by network.
2726 *
2727 * Besides that it updates the congestion state when packet loss or ECN
2728 * is detected. But it does not reduce the cwnd, it is done by the
2729 * congestion control later.
2730 *
2731 * It does _not_ decide what to send, it is made in function
2732 * tcp_xmit_retransmit_queue().
2733 */
2734static void tcp_fastretrans_alert(struct sock *sk, const u32 prior_snd_una,
2735 bool is_dupack, int *ack_flag, int *rexmit)
2736{
2737 struct inet_connection_sock *icsk = inet_csk(sk);
2738 struct tcp_sock *tp = tcp_sk(sk);
2739 int fast_rexmit = 0, flag = *ack_flag;
2740 bool do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2741 tcp_force_fast_retransmit(sk));
2742
2743 if (!tp->packets_out && tp->sacked_out)
2744 tp->sacked_out = 0;
2745
2746 /* Now state machine starts.
2747 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2748 if (flag & FLAG_ECE)
2749 tp->prior_ssthresh = 0;
2750
2751 /* B. In all the states check for reneging SACKs. */
2752 if (tcp_check_sack_reneging(sk, flag))
2753 return;
2754
2755 /* C. Check consistency of the current state. */
2756 tcp_verify_left_out(tp);
2757
2758 /* D. Check state exit conditions. State can be terminated
2759 * when high_seq is ACKed. */
2760 if (icsk->icsk_ca_state == TCP_CA_Open) {
2761 WARN_ON(tp->retrans_out != 0);
2762 tp->retrans_stamp = 0;
2763 } else if (!before(tp->snd_una, tp->high_seq)) {
2764 switch (icsk->icsk_ca_state) {
2765 case TCP_CA_CWR:
2766 /* CWR is to be held something *above* high_seq
2767 * is ACKed for CWR bit to reach receiver. */
2768 if (tp->snd_una != tp->high_seq) {
2769 tcp_end_cwnd_reduction(sk);
2770 tcp_set_ca_state(sk, TCP_CA_Open);
2771 }
2772 break;
2773
2774 case TCP_CA_Recovery:
2775 if (tcp_is_reno(tp))
2776 tcp_reset_reno_sack(tp);
2777 if (tcp_try_undo_recovery(sk))
2778 return;
2779 tcp_end_cwnd_reduction(sk);
2780 break;
2781 }
2782 }
2783
2784 /* E. Process state. */
2785 switch (icsk->icsk_ca_state) {
2786 case TCP_CA_Recovery:
2787 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2788 if (tcp_is_reno(tp) && is_dupack)
2789 tcp_add_reno_sack(sk);
2790 } else {
2791 if (tcp_try_undo_partial(sk, prior_snd_una))
2792 return;
2793 /* Partial ACK arrived. Force fast retransmit. */
2794 do_lost = tcp_is_reno(tp) ||
2795 tcp_force_fast_retransmit(sk);
2796 }
2797 if (tcp_try_undo_dsack(sk)) {
2798 tcp_try_keep_open(sk);
2799 return;
2800 }
2801 tcp_rack_identify_loss(sk, ack_flag);
2802 break;
2803 case TCP_CA_Loss:
2804 tcp_process_loss(sk, flag, is_dupack, rexmit);
2805 tcp_rack_identify_loss(sk, ack_flag);
2806 if (!(icsk->icsk_ca_state == TCP_CA_Open ||
2807 (*ack_flag & FLAG_LOST_RETRANS)))
2808 return;
2809 /* Change state if cwnd is undone or retransmits are lost */
2810 /* fall through */
2811 default:
2812 if (tcp_is_reno(tp)) {
2813 if (flag & FLAG_SND_UNA_ADVANCED)
2814 tcp_reset_reno_sack(tp);
2815 if (is_dupack)
2816 tcp_add_reno_sack(sk);
2817 }
2818
2819 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2820 tcp_try_undo_dsack(sk);
2821
2822 tcp_rack_identify_loss(sk, ack_flag);
2823 if (!tcp_time_to_recover(sk, flag)) {
2824 tcp_try_to_open(sk, flag);
2825 return;
2826 }
2827
2828 /* MTU probe failure: don't reduce cwnd */
2829 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2830 icsk->icsk_mtup.probe_size &&
2831 tp->snd_una == tp->mtu_probe.probe_seq_start) {
2832 tcp_mtup_probe_failed(sk);
2833 /* Restores the reduction we did in tcp_mtup_probe() */
2834 tp->snd_cwnd++;
2835 tcp_simple_retransmit(sk);
2836 return;
2837 }
2838
2839 /* Otherwise enter Recovery state */
2840 tcp_enter_recovery(sk, (flag & FLAG_ECE));
2841 fast_rexmit = 1;
2842 }
2843
2844 if (do_lost)
2845 tcp_update_scoreboard(sk, fast_rexmit);
2846 *rexmit = REXMIT_LOST;
2847}
2848
2849static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us, const int flag)
2850{
2851 u32 wlen = sock_net(sk)->ipv4.sysctl_tcp_min_rtt_wlen * HZ;
2852 struct tcp_sock *tp = tcp_sk(sk);
2853
2854 if ((flag & FLAG_ACK_MAYBE_DELAYED) && rtt_us > tcp_min_rtt(tp)) {
2855 /* If the remote keeps returning delayed ACKs, eventually
2856 * the min filter would pick it up and overestimate the
2857 * prop. delay when it expires. Skip suspected delayed ACKs.
2858 */
2859 return;
2860 }
2861 minmax_running_min(&tp->rtt_min, wlen, tcp_jiffies32,
2862 rtt_us ? : jiffies_to_usecs(1));
2863}
2864
2865static bool tcp_ack_update_rtt(struct sock *sk, const int flag,
2866 long seq_rtt_us, long sack_rtt_us,
2867 long ca_rtt_us, struct rate_sample *rs)
2868{
2869 const struct tcp_sock *tp = tcp_sk(sk);
2870
2871 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2872 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2873 * Karn's algorithm forbids taking RTT if some retransmitted data
2874 * is acked (RFC6298).
2875 */
2876 if (seq_rtt_us < 0)
2877 seq_rtt_us = sack_rtt_us;
2878
2879 /* RTTM Rule: A TSecr value received in a segment is used to
2880 * update the averaged RTT measurement only if the segment
2881 * acknowledges some new data, i.e., only if it advances the
2882 * left edge of the send window.
2883 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2884 */
2885 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2886 flag & FLAG_ACKED) {
2887 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
2888 u32 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
2889
2890 seq_rtt_us = ca_rtt_us = delta_us;
2891 }
2892 rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */
2893 if (seq_rtt_us < 0)
2894 return false;
2895
2896 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
2897 * always taken together with ACK, SACK, or TS-opts. Any negative
2898 * values will be skipped with the seq_rtt_us < 0 check above.
2899 */
2900 tcp_update_rtt_min(sk, ca_rtt_us, flag);
2901 tcp_rtt_estimator(sk, seq_rtt_us);
2902 tcp_set_rto(sk);
2903
2904 /* RFC6298: only reset backoff on valid RTT measurement. */
2905 inet_csk(sk)->icsk_backoff = 0;
2906 return true;
2907}
2908
2909/* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2910void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
2911{
2912 struct rate_sample rs;
2913 long rtt_us = -1L;
2914
2915 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack)
2916 rtt_us = tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req)->snt_synack);
2917
2918 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us, &rs);
2919}
2920
2921
2922static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
2923{
2924 const struct inet_connection_sock *icsk = inet_csk(sk);
2925
2926 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
2927 tcp_sk(sk)->snd_cwnd_stamp = tcp_jiffies32;
2928}
2929
2930/* Restart timer after forward progress on connection.
2931 * RFC2988 recommends to restart timer to now+rto.
2932 */
2933void tcp_rearm_rto(struct sock *sk)
2934{
2935 const struct inet_connection_sock *icsk = inet_csk(sk);
2936 struct tcp_sock *tp = tcp_sk(sk);
2937
2938 /* If the retrans timer is currently being used by Fast Open
2939 * for SYN-ACK retrans purpose, stay put.
2940 */
2941 if (tp->fastopen_rsk)
2942 return;
2943
2944 if (!tp->packets_out) {
2945 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2946 } else {
2947 u32 rto = inet_csk(sk)->icsk_rto;
2948 /* Offset the time elapsed after installing regular RTO */
2949 if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT ||
2950 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
2951 s64 delta_us = tcp_rto_delta_us(sk);
2952 /* delta_us may not be positive if the socket is locked
2953 * when the retrans timer fires and is rescheduled.
2954 */
2955 rto = usecs_to_jiffies(max_t(int, delta_us, 1));
2956 }
2957 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
2958 TCP_RTO_MAX);
2959 }
2960}
2961
2962/* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */
2963static void tcp_set_xmit_timer(struct sock *sk)
2964{
2965 if (!tcp_schedule_loss_probe(sk, true))
2966 tcp_rearm_rto(sk);
2967}
2968
2969/* If we get here, the whole TSO packet has not been acked. */
2970static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
2971{
2972 struct tcp_sock *tp = tcp_sk(sk);
2973 u32 packets_acked;
2974
2975 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
2976
2977 packets_acked = tcp_skb_pcount(skb);
2978 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2979 return 0;
2980 packets_acked -= tcp_skb_pcount(skb);
2981
2982 if (packets_acked) {
2983 BUG_ON(tcp_skb_pcount(skb) == 0);
2984 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
2985 }
2986
2987 return packets_acked;
2988}
2989
2990static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
2991 u32 prior_snd_una)
2992{
2993 const struct skb_shared_info *shinfo;
2994
2995 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
2996 if (likely(!TCP_SKB_CB(skb)->txstamp_ack))
2997 return;
2998
2999 shinfo = skb_shinfo(skb);
3000 if (!before(shinfo->tskey, prior_snd_una) &&
3001 before(shinfo->tskey, tcp_sk(sk)->snd_una)) {
3002 tcp_skb_tsorted_save(skb) {
3003 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK);
3004 } tcp_skb_tsorted_restore(skb);
3005 }
3006}
3007
3008/* Remove acknowledged frames from the retransmission queue. If our packet
3009 * is before the ack sequence we can discard it as it's confirmed to have
3010 * arrived at the other end.
3011 */
3012static int tcp_clean_rtx_queue(struct sock *sk, u32 prior_fack,
3013 u32 prior_snd_una,
3014 struct tcp_sacktag_state *sack)
3015{
3016 const struct inet_connection_sock *icsk = inet_csk(sk);
3017 u64 first_ackt, last_ackt;
3018 struct tcp_sock *tp = tcp_sk(sk);
3019 u32 prior_sacked = tp->sacked_out;
3020 u32 reord = tp->snd_nxt; /* lowest acked un-retx un-sacked seq */
3021 struct sk_buff *skb, *next;
3022 bool fully_acked = true;
3023 long sack_rtt_us = -1L;
3024 long seq_rtt_us = -1L;
3025 long ca_rtt_us = -1L;
3026 u32 pkts_acked = 0;
3027 u32 last_in_flight = 0;
3028 bool rtt_update;
3029 int flag = 0;
3030
3031 first_ackt = 0;
3032
3033 for (skb = skb_rb_first(&sk->tcp_rtx_queue); skb; skb = next) {
3034 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3035 const u32 start_seq = scb->seq;
3036 u8 sacked = scb->sacked;
3037 u32 acked_pcount;
3038
3039 tcp_ack_tstamp(sk, skb, prior_snd_una);
3040
3041 /* Determine how many packets and what bytes were acked, tso and else */
3042 if (after(scb->end_seq, tp->snd_una)) {
3043 if (tcp_skb_pcount(skb) == 1 ||
3044 !after(tp->snd_una, scb->seq))
3045 break;
3046
3047 acked_pcount = tcp_tso_acked(sk, skb);
3048 if (!acked_pcount)
3049 break;
3050 fully_acked = false;
3051 } else {
3052 acked_pcount = tcp_skb_pcount(skb);
3053 }
3054
3055 if (unlikely(sacked & TCPCB_RETRANS)) {
3056 if (sacked & TCPCB_SACKED_RETRANS)
3057 tp->retrans_out -= acked_pcount;
3058 flag |= FLAG_RETRANS_DATA_ACKED;
3059 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3060 last_ackt = skb->skb_mstamp;
3061 WARN_ON_ONCE(last_ackt == 0);
3062 if (!first_ackt)
3063 first_ackt = last_ackt;
3064
3065 last_in_flight = TCP_SKB_CB(skb)->tx.in_flight;
3066 if (before(start_seq, reord))
3067 reord = start_seq;
3068 if (!after(scb->end_seq, tp->high_seq))
3069 flag |= FLAG_ORIG_SACK_ACKED;
3070 }
3071
3072 if (sacked & TCPCB_SACKED_ACKED) {
3073 tp->sacked_out -= acked_pcount;
3074 } else if (tcp_is_sack(tp)) {
3075 tp->delivered += acked_pcount;
3076 if (!tcp_skb_spurious_retrans(tp, skb))
3077 tcp_rack_advance(tp, sacked, scb->end_seq,
3078 skb->skb_mstamp);
3079 }
3080 if (sacked & TCPCB_LOST)
3081 tp->lost_out -= acked_pcount;
3082
3083 tp->packets_out -= acked_pcount;
3084 pkts_acked += acked_pcount;
3085 tcp_rate_skb_delivered(sk, skb, sack->rate);
3086
3087 /* Initial outgoing SYN's get put onto the write_queue
3088 * just like anything else we transmit. It is not
3089 * true data, and if we misinform our callers that
3090 * this ACK acks real data, we will erroneously exit
3091 * connection startup slow start one packet too
3092 * quickly. This is severely frowned upon behavior.
3093 */
3094 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3095 flag |= FLAG_DATA_ACKED;
3096 } else {
3097 flag |= FLAG_SYN_ACKED;
3098 tp->retrans_stamp = 0;
3099 }
3100
3101 if (!fully_acked)
3102 break;
3103
3104 next = skb_rb_next(skb);
3105 if (unlikely(skb == tp->retransmit_skb_hint))
3106 tp->retransmit_skb_hint = NULL;
3107 if (unlikely(skb == tp->lost_skb_hint))
3108 tp->lost_skb_hint = NULL;
3109 tcp_rtx_queue_unlink_and_free(skb, sk);
3110 }
3111
3112 if (!skb)
3113 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3114
3115 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3116 tp->snd_up = tp->snd_una;
3117
3118 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3119 flag |= FLAG_SACK_RENEGING;
3120
3121 if (likely(first_ackt) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
3122 seq_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, first_ackt);
3123 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, last_ackt);
3124
3125 if (pkts_acked == 1 && last_in_flight < tp->mss_cache &&
3126 last_in_flight && !prior_sacked && fully_acked &&
3127 sack->rate->prior_delivered + 1 == tp->delivered &&
3128 !(flag & (FLAG_CA_ALERT | FLAG_SYN_ACKED))) {
3129 /* Conservatively mark a delayed ACK. It's typically
3130 * from a lone runt packet over the round trip to
3131 * a receiver w/o out-of-order or CE events.
3132 */
3133 flag |= FLAG_ACK_MAYBE_DELAYED;
3134 }
3135 }
3136 if (sack->first_sackt) {
3137 sack_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->first_sackt);
3138 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->last_sackt);
3139 }
3140 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
3141 ca_rtt_us, sack->rate);
3142
3143 if (flag & FLAG_ACKED) {
3144 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3145 if (unlikely(icsk->icsk_mtup.probe_size &&
3146 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3147 tcp_mtup_probe_success(sk);
3148 }
3149
3150 if (tcp_is_reno(tp)) {
3151 tcp_remove_reno_sacks(sk, pkts_acked);
3152 } else {
3153 int delta;
3154
3155 /* Non-retransmitted hole got filled? That's reordering */
3156 if (before(reord, prior_fack))
3157 tcp_check_sack_reordering(sk, reord, 0);
3158
3159 delta = prior_sacked - tp->sacked_out;
3160 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3161 }
3162 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3163 sack_rtt_us > tcp_stamp_us_delta(tp->tcp_mstamp, skb->skb_mstamp)) {
3164 /* Do not re-arm RTO if the sack RTT is measured from data sent
3165 * after when the head was last (re)transmitted. Otherwise the
3166 * timeout may continue to extend in loss recovery.
3167 */
3168 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3169 }
3170
3171 if (icsk->icsk_ca_ops->pkts_acked) {
3172 struct ack_sample sample = { .pkts_acked = pkts_acked,
3173 .rtt_us = sack->rate->rtt_us,
3174 .in_flight = last_in_flight };
3175
3176 icsk->icsk_ca_ops->pkts_acked(sk, &sample);
3177 }
3178
3179#if FASTRETRANS_DEBUG > 0
3180 WARN_ON((int)tp->sacked_out < 0);
3181 WARN_ON((int)tp->lost_out < 0);
3182 WARN_ON((int)tp->retrans_out < 0);
3183 if (!tp->packets_out && tcp_is_sack(tp)) {
3184 icsk = inet_csk(sk);
3185 if (tp->lost_out) {
3186 pr_debug("Leak l=%u %d\n",
3187 tp->lost_out, icsk->icsk_ca_state);
3188 tp->lost_out = 0;
3189 }
3190 if (tp->sacked_out) {
3191 pr_debug("Leak s=%u %d\n",
3192 tp->sacked_out, icsk->icsk_ca_state);
3193 tp->sacked_out = 0;
3194 }
3195 if (tp->retrans_out) {
3196 pr_debug("Leak r=%u %d\n",
3197 tp->retrans_out, icsk->icsk_ca_state);
3198 tp->retrans_out = 0;
3199 }
3200 }
3201#endif
3202 return flag;
3203}
3204
3205static void tcp_ack_probe(struct sock *sk)
3206{
3207 struct inet_connection_sock *icsk = inet_csk(sk);
3208 struct sk_buff *head = tcp_send_head(sk);
3209 const struct tcp_sock *tp = tcp_sk(sk);
3210
3211 /* Was it a usable window open? */
3212 if (!head)
3213 return;
3214 if (!after(TCP_SKB_CB(head)->end_seq, tcp_wnd_end(tp))) {
3215 icsk->icsk_backoff = 0;
3216 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3217 /* Socket must be waked up by subsequent tcp_data_snd_check().
3218 * This function is not for random using!
3219 */
3220 } else {
3221 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
3222
3223 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3224 when, TCP_RTO_MAX);
3225 }
3226}
3227
3228static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3229{
3230 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3231 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3232}
3233
3234/* Decide wheather to run the increase function of congestion control. */
3235static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3236{
3237 /* If reordering is high then always grow cwnd whenever data is
3238 * delivered regardless of its ordering. Otherwise stay conservative
3239 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3240 * new SACK or ECE mark may first advance cwnd here and later reduce
3241 * cwnd in tcp_fastretrans_alert() based on more states.
3242 */
3243 if (tcp_sk(sk)->reordering > sock_net(sk)->ipv4.sysctl_tcp_reordering)
3244 return flag & FLAG_FORWARD_PROGRESS;
3245
3246 return flag & FLAG_DATA_ACKED;
3247}
3248
3249/* The "ultimate" congestion control function that aims to replace the rigid
3250 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3251 * It's called toward the end of processing an ACK with precise rate
3252 * information. All transmission or retransmission are delayed afterwards.
3253 */
3254static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
3255 int flag, const struct rate_sample *rs)
3256{
3257 const struct inet_connection_sock *icsk = inet_csk(sk);
3258
3259 if (icsk->icsk_ca_ops->cong_control) {
3260 icsk->icsk_ca_ops->cong_control(sk, rs);
3261 return;
3262 }
3263
3264 if (tcp_in_cwnd_reduction(sk)) {
3265 /* Reduce cwnd if state mandates */
3266 tcp_cwnd_reduction(sk, acked_sacked, flag);
3267 } else if (tcp_may_raise_cwnd(sk, flag)) {
3268 /* Advance cwnd if state allows */
3269 tcp_cong_avoid(sk, ack, acked_sacked);
3270 }
3271 tcp_update_pacing_rate(sk);
3272}
3273
3274/* Check that window update is acceptable.
3275 * The function assumes that snd_una<=ack<=snd_next.
3276 */
3277static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3278 const u32 ack, const u32 ack_seq,
3279 const u32 nwin)
3280{
3281 return after(ack, tp->snd_una) ||
3282 after(ack_seq, tp->snd_wl1) ||
3283 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3284}
3285
3286/* If we update tp->snd_una, also update tp->bytes_acked */
3287static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3288{
3289 u32 delta = ack - tp->snd_una;
3290
3291 sock_owned_by_me((struct sock *)tp);
3292 tp->bytes_acked += delta;
3293 tp->snd_una = ack;
3294}
3295
3296/* If we update tp->rcv_nxt, also update tp->bytes_received */
3297static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3298{
3299 u32 delta = seq - tp->rcv_nxt;
3300
3301 sock_owned_by_me((struct sock *)tp);
3302 tp->bytes_received += delta;
3303 tp->rcv_nxt = seq;
3304}
3305
3306/* Update our send window.
3307 *
3308 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3309 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3310 */
3311static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3312 u32 ack_seq)
3313{
3314 struct tcp_sock *tp = tcp_sk(sk);
3315 int flag = 0;
3316 u32 nwin = ntohs(tcp_hdr(skb)->window);
3317
3318 if (likely(!tcp_hdr(skb)->syn))
3319 nwin <<= tp->rx_opt.snd_wscale;
3320
3321 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3322 flag |= FLAG_WIN_UPDATE;
3323 tcp_update_wl(tp, ack_seq);
3324
3325 if (tp->snd_wnd != nwin) {
3326 tp->snd_wnd = nwin;
3327
3328 /* Note, it is the only place, where
3329 * fast path is recovered for sending TCP.
3330 */
3331 tp->pred_flags = 0;
3332 tcp_fast_path_check(sk);
3333
3334 if (!tcp_write_queue_empty(sk))
3335 tcp_slow_start_after_idle_check(sk);
3336
3337 if (nwin > tp->max_window) {
3338 tp->max_window = nwin;
3339 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3340 }
3341 }
3342 }
3343
3344 tcp_snd_una_update(tp, ack);
3345
3346 return flag;
3347}
3348
3349static bool __tcp_oow_rate_limited(struct net *net, int mib_idx,
3350 u32 *last_oow_ack_time)
3351{
3352 if (*last_oow_ack_time) {
3353 s32 elapsed = (s32)(tcp_jiffies32 - *last_oow_ack_time);
3354
3355 if (0 <= elapsed && elapsed < net->ipv4.sysctl_tcp_invalid_ratelimit) {
3356 NET_INC_STATS(net, mib_idx);
3357 return true; /* rate-limited: don't send yet! */
3358 }
3359 }
3360
3361 *last_oow_ack_time = tcp_jiffies32;
3362
3363 return false; /* not rate-limited: go ahead, send dupack now! */
3364}
3365
3366/* Return true if we're currently rate-limiting out-of-window ACKs and
3367 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3368 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3369 * attacks that send repeated SYNs or ACKs for the same connection. To
3370 * do this, we do not send a duplicate SYNACK or ACK if the remote
3371 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3372 */
3373bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3374 int mib_idx, u32 *last_oow_ack_time)
3375{
3376 /* Data packets without SYNs are not likely part of an ACK loop. */
3377 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3378 !tcp_hdr(skb)->syn)
3379 return false;
3380
3381 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time);
3382}
3383
3384/* RFC 5961 7 [ACK Throttling] */
3385static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb)
3386{
3387 /* unprotected vars, we dont care of overwrites */
3388 static u32 challenge_timestamp;
3389 static unsigned int challenge_count;
3390 struct tcp_sock *tp = tcp_sk(sk);
3391 struct net *net = sock_net(sk);
3392 u32 count, now;
3393
3394 /* First check our per-socket dupack rate limit. */
3395 if (__tcp_oow_rate_limited(net,
3396 LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3397 &tp->last_oow_ack_time))
3398 return;
3399
3400 /* Then check host-wide RFC 5961 rate limit. */
3401 now = jiffies / HZ;
3402 if (now != challenge_timestamp) {
3403 u32 ack_limit = net->ipv4.sysctl_tcp_challenge_ack_limit;
3404 u32 half = (ack_limit + 1) >> 1;
3405
3406 challenge_timestamp = now;
3407 WRITE_ONCE(challenge_count, half + prandom_u32_max(ack_limit));
3408 }
3409 count = READ_ONCE(challenge_count);
3410 if (count > 0) {
3411 WRITE_ONCE(challenge_count, count - 1);
3412 NET_INC_STATS(net, LINUX_MIB_TCPCHALLENGEACK);
3413 tcp_send_ack(sk);
3414 }
3415}
3416
3417static void tcp_store_ts_recent(struct tcp_sock *tp)
3418{
3419 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3420 tp->rx_opt.ts_recent_stamp = get_seconds();
3421}
3422
3423static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3424{
3425 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3426 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3427 * extra check below makes sure this can only happen
3428 * for pure ACK frames. -DaveM
3429 *
3430 * Not only, also it occurs for expired timestamps.
3431 */
3432
3433 if (tcp_paws_check(&tp->rx_opt, 0))
3434 tcp_store_ts_recent(tp);
3435 }
3436}
3437
3438/* This routine deals with acks during a TLP episode.
3439 * We mark the end of a TLP episode on receiving TLP dupack or when
3440 * ack is after tlp_high_seq.
3441 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3442 */
3443static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3444{
3445 struct tcp_sock *tp = tcp_sk(sk);
3446
3447 if (before(ack, tp->tlp_high_seq))
3448 return;
3449
3450 if (flag & FLAG_DSACKING_ACK) {
3451 /* This DSACK means original and TLP probe arrived; no loss */
3452 tp->tlp_high_seq = 0;
3453 } else if (after(ack, tp->tlp_high_seq)) {
3454 /* ACK advances: there was a loss, so reduce cwnd. Reset
3455 * tlp_high_seq in tcp_init_cwnd_reduction()
3456 */
3457 tcp_init_cwnd_reduction(sk);
3458 tcp_set_ca_state(sk, TCP_CA_CWR);
3459 tcp_end_cwnd_reduction(sk);
3460 tcp_try_keep_open(sk);
3461 NET_INC_STATS(sock_net(sk),
3462 LINUX_MIB_TCPLOSSPROBERECOVERY);
3463 } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3464 FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3465 /* Pure dupack: original and TLP probe arrived; no loss */
3466 tp->tlp_high_seq = 0;
3467 }
3468}
3469
3470static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3471{
3472 const struct inet_connection_sock *icsk = inet_csk(sk);
3473
3474 if (icsk->icsk_ca_ops->in_ack_event)
3475 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3476}
3477
3478/* Congestion control has updated the cwnd already. So if we're in
3479 * loss recovery then now we do any new sends (for FRTO) or
3480 * retransmits (for CA_Loss or CA_recovery) that make sense.
3481 */
3482static void tcp_xmit_recovery(struct sock *sk, int rexmit)
3483{
3484 struct tcp_sock *tp = tcp_sk(sk);
3485
3486 if (rexmit == REXMIT_NONE)
3487 return;
3488
3489 if (unlikely(rexmit == 2)) {
3490 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
3491 TCP_NAGLE_OFF);
3492 if (after(tp->snd_nxt, tp->high_seq))
3493 return;
3494 tp->frto = 0;
3495 }
3496 tcp_xmit_retransmit_queue(sk);
3497}
3498
3499/* This routine deals with incoming acks, but not outgoing ones. */
3500static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3501{
3502 struct inet_connection_sock *icsk = inet_csk(sk);
3503 struct tcp_sock *tp = tcp_sk(sk);
3504 struct tcp_sacktag_state sack_state;
3505 struct rate_sample rs = { .prior_delivered = 0 };
3506 u32 prior_snd_una = tp->snd_una;
3507 bool is_sack_reneg = tp->is_sack_reneg;
3508 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3509 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3510 bool is_dupack = false;
3511 int prior_packets = tp->packets_out;
3512 u32 delivered = tp->delivered;
3513 u32 lost = tp->lost;
3514 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
3515 u32 prior_fack;
3516
3517 sack_state.first_sackt = 0;
3518 sack_state.rate = &rs;
3519
3520 /* We very likely will need to access rtx queue. */
3521 prefetch(sk->tcp_rtx_queue.rb_node);
3522
3523 /* If the ack is older than previous acks
3524 * then we can probably ignore it.
3525 */
3526 if (before(ack, prior_snd_una)) {
3527 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3528 if (before(ack, prior_snd_una - tp->max_window)) {
3529 if (!(flag & FLAG_NO_CHALLENGE_ACK))
3530 tcp_send_challenge_ack(sk, skb);
3531 return -1;
3532 }
3533 goto old_ack;
3534 }
3535
3536 /* If the ack includes data we haven't sent yet, discard
3537 * this segment (RFC793 Section 3.9).
3538 */
3539 if (after(ack, tp->snd_nxt))
3540 goto invalid_ack;
3541
3542 if (after(ack, prior_snd_una)) {
3543 flag |= FLAG_SND_UNA_ADVANCED;
3544 icsk->icsk_retransmits = 0;
3545 }
3546
3547 prior_fack = tcp_is_sack(tp) ? tcp_highest_sack_seq(tp) : tp->snd_una;
3548 rs.prior_in_flight = tcp_packets_in_flight(tp);
3549
3550 /* ts_recent update must be made after we are sure that the packet
3551 * is in window.
3552 */
3553 if (flag & FLAG_UPDATE_TS_RECENT)
3554 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3555
3556 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3557 /* Window is constant, pure forward advance.
3558 * No more checks are required.
3559 * Note, we use the fact that SND.UNA>=SND.WL2.
3560 */
3561 tcp_update_wl(tp, ack_seq);
3562 tcp_snd_una_update(tp, ack);
3563 flag |= FLAG_WIN_UPDATE;
3564
3565 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3566
3567 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS);
3568 } else {
3569 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3570
3571 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3572 flag |= FLAG_DATA;
3573 else
3574 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3575
3576 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3577
3578 if (TCP_SKB_CB(skb)->sacked)
3579 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3580 &sack_state);
3581
3582 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3583 flag |= FLAG_ECE;
3584 ack_ev_flags |= CA_ACK_ECE;
3585 }
3586
3587 if (flag & FLAG_WIN_UPDATE)
3588 ack_ev_flags |= CA_ACK_WIN_UPDATE;
3589
3590 tcp_in_ack_event(sk, ack_ev_flags);
3591 }
3592
3593 /* We passed data and got it acked, remove any soft error
3594 * log. Something worked...
3595 */
3596 sk->sk_err_soft = 0;
3597 icsk->icsk_probes_out = 0;
3598 tp->rcv_tstamp = tcp_jiffies32;
3599 if (!prior_packets)
3600 goto no_queue;
3601
3602 /* See if we can take anything off of the retransmit queue. */
3603 flag |= tcp_clean_rtx_queue(sk, prior_fack, prior_snd_una, &sack_state);
3604
3605 tcp_rack_update_reo_wnd(sk, &rs);
3606
3607 if (tp->tlp_high_seq)
3608 tcp_process_tlp_ack(sk, ack, flag);
3609 /* If needed, reset TLP/RTO timer; RACK may later override this. */
3610 if (flag & FLAG_SET_XMIT_TIMER)
3611 tcp_set_xmit_timer(sk);
3612
3613 if (tcp_ack_is_dubious(sk, flag)) {
3614 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3615 tcp_fastretrans_alert(sk, prior_snd_una, is_dupack, &flag,
3616 &rexmit);
3617 }
3618
3619 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3620 sk_dst_confirm(sk);
3621
3622 delivered = tp->delivered - delivered; /* freshly ACKed or SACKed */
3623 lost = tp->lost - lost; /* freshly marked lost */
3624 rs.is_ack_delayed = !!(flag & FLAG_ACK_MAYBE_DELAYED);
3625 tcp_rate_gen(sk, delivered, lost, is_sack_reneg, sack_state.rate);
3626 tcp_cong_control(sk, ack, delivered, flag, sack_state.rate);
3627 tcp_xmit_recovery(sk, rexmit);
3628 return 1;
3629
3630no_queue:
3631 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3632 if (flag & FLAG_DSACKING_ACK)
3633 tcp_fastretrans_alert(sk, prior_snd_una, is_dupack, &flag,
3634 &rexmit);
3635 /* If this ack opens up a zero window, clear backoff. It was
3636 * being used to time the probes, and is probably far higher than
3637 * it needs to be for normal retransmission.
3638 */
3639 tcp_ack_probe(sk);
3640
3641 if (tp->tlp_high_seq)
3642 tcp_process_tlp_ack(sk, ack, flag);
3643 return 1;
3644
3645invalid_ack:
3646 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3647 return -1;
3648
3649old_ack:
3650 /* If data was SACKed, tag it and see if we should send more data.
3651 * If data was DSACKed, see if we can undo a cwnd reduction.
3652 */
3653 if (TCP_SKB_CB(skb)->sacked) {
3654 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3655 &sack_state);
3656 tcp_fastretrans_alert(sk, prior_snd_una, is_dupack, &flag,
3657 &rexmit);
3658 tcp_xmit_recovery(sk, rexmit);
3659 }
3660
3661 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3662 return 0;
3663}
3664
3665static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3666 bool syn, struct tcp_fastopen_cookie *foc,
3667 bool exp_opt)
3668{
3669 /* Valid only in SYN or SYN-ACK with an even length. */
3670 if (!foc || !syn || len < 0 || (len & 1))
3671 return;
3672
3673 if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3674 len <= TCP_FASTOPEN_COOKIE_MAX)
3675 memcpy(foc->val, cookie, len);
3676 else if (len != 0)
3677 len = -1;
3678 foc->len = len;
3679 foc->exp = exp_opt;
3680}
3681
3682static void smc_parse_options(const struct tcphdr *th,
3683 struct tcp_options_received *opt_rx,
3684 const unsigned char *ptr,
3685 int opsize)
3686{
3687#if IS_ENABLED(CONFIG_SMC)
3688 if (static_branch_unlikely(&tcp_have_smc)) {
3689 if (th->syn && !(opsize & 1) &&
3690 opsize >= TCPOLEN_EXP_SMC_BASE &&
3691 get_unaligned_be32(ptr) == TCPOPT_SMC_MAGIC)
3692 opt_rx->smc_ok = 1;
3693 }
3694#endif
3695}
3696
3697/* Look for tcp options. Normally only called on SYN and SYNACK packets.
3698 * But, this can also be called on packets in the established flow when
3699 * the fast version below fails.
3700 */
3701void tcp_parse_options(const struct net *net,
3702 const struct sk_buff *skb,
3703 struct tcp_options_received *opt_rx, int estab,
3704 struct tcp_fastopen_cookie *foc)
3705{
3706 const unsigned char *ptr;
3707 const struct tcphdr *th = tcp_hdr(skb);
3708 int length = (th->doff * 4) - sizeof(struct tcphdr);
3709
3710 ptr = (const unsigned char *)(th + 1);
3711 opt_rx->saw_tstamp = 0;
3712
3713 while (length > 0) {
3714 int opcode = *ptr++;
3715 int opsize;
3716
3717 switch (opcode) {
3718 case TCPOPT_EOL:
3719 return;
3720 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3721 length--;
3722 continue;
3723 default:
3724 opsize = *ptr++;
3725 if (opsize < 2) /* "silly options" */
3726 return;
3727 if (opsize > length)
3728 return; /* don't parse partial options */
3729 switch (opcode) {
3730 case TCPOPT_MSS:
3731 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3732 u16 in_mss = get_unaligned_be16(ptr);
3733 if (in_mss) {
3734 if (opt_rx->user_mss &&
3735 opt_rx->user_mss < in_mss)
3736 in_mss = opt_rx->user_mss;
3737 opt_rx->mss_clamp = in_mss;
3738 }
3739 }
3740 break;
3741 case TCPOPT_WINDOW:
3742 if (opsize == TCPOLEN_WINDOW && th->syn &&
3743 !estab && net->ipv4.sysctl_tcp_window_scaling) {
3744 __u8 snd_wscale = *(__u8 *)ptr;
3745 opt_rx->wscale_ok = 1;
3746 if (snd_wscale > TCP_MAX_WSCALE) {
3747 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n",
3748 __func__,
3749 snd_wscale,
3750 TCP_MAX_WSCALE);
3751 snd_wscale = TCP_MAX_WSCALE;
3752 }
3753 opt_rx->snd_wscale = snd_wscale;
3754 }
3755 break;
3756 case TCPOPT_TIMESTAMP:
3757 if ((opsize == TCPOLEN_TIMESTAMP) &&
3758 ((estab && opt_rx->tstamp_ok) ||
3759 (!estab && net->ipv4.sysctl_tcp_timestamps))) {
3760 opt_rx->saw_tstamp = 1;
3761 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3762 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3763 }
3764 break;
3765 case TCPOPT_SACK_PERM:
3766 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3767 !estab && net->ipv4.sysctl_tcp_sack) {
3768 opt_rx->sack_ok = TCP_SACK_SEEN;
3769 tcp_sack_reset(opt_rx);
3770 }
3771 break;
3772
3773 case TCPOPT_SACK:
3774 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3775 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3776 opt_rx->sack_ok) {
3777 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3778 }
3779 break;
3780#ifdef CONFIG_TCP_MD5SIG
3781 case TCPOPT_MD5SIG:
3782 /*
3783 * The MD5 Hash has already been
3784 * checked (see tcp_v{4,6}_do_rcv()).
3785 */
3786 break;
3787#endif
3788 case TCPOPT_FASTOPEN:
3789 tcp_parse_fastopen_option(
3790 opsize - TCPOLEN_FASTOPEN_BASE,
3791 ptr, th->syn, foc, false);
3792 break;
3793
3794 case TCPOPT_EXP:
3795 /* Fast Open option shares code 254 using a
3796 * 16 bits magic number.
3797 */
3798 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
3799 get_unaligned_be16(ptr) ==
3800 TCPOPT_FASTOPEN_MAGIC)
3801 tcp_parse_fastopen_option(opsize -
3802 TCPOLEN_EXP_FASTOPEN_BASE,
3803 ptr + 2, th->syn, foc, true);
3804 else
3805 smc_parse_options(th, opt_rx, ptr,
3806 opsize);
3807 break;
3808
3809 }
3810 ptr += opsize-2;
3811 length -= opsize;
3812 }
3813 }
3814}
3815EXPORT_SYMBOL(tcp_parse_options);
3816
3817static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3818{
3819 const __be32 *ptr = (const __be32 *)(th + 1);
3820
3821 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3822 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3823 tp->rx_opt.saw_tstamp = 1;
3824 ++ptr;
3825 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3826 ++ptr;
3827 if (*ptr)
3828 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3829 else
3830 tp->rx_opt.rcv_tsecr = 0;
3831 return true;
3832 }
3833 return false;
3834}
3835
3836/* Fast parse options. This hopes to only see timestamps.
3837 * If it is wrong it falls back on tcp_parse_options().
3838 */
3839static bool tcp_fast_parse_options(const struct net *net,
3840 const struct sk_buff *skb,
3841 const struct tcphdr *th, struct tcp_sock *tp)
3842{
3843 /* In the spirit of fast parsing, compare doff directly to constant
3844 * values. Because equality is used, short doff can be ignored here.
3845 */
3846 if (th->doff == (sizeof(*th) / 4)) {
3847 tp->rx_opt.saw_tstamp = 0;
3848 return false;
3849 } else if (tp->rx_opt.tstamp_ok &&
3850 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3851 if (tcp_parse_aligned_timestamp(tp, th))
3852 return true;
3853 }
3854
3855 tcp_parse_options(net, skb, &tp->rx_opt, 1, NULL);
3856 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3857 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3858
3859 return true;
3860}
3861
3862#ifdef CONFIG_TCP_MD5SIG
3863/*
3864 * Parse MD5 Signature option
3865 */
3866const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3867{
3868 int length = (th->doff << 2) - sizeof(*th);
3869 const u8 *ptr = (const u8 *)(th + 1);
3870
3871 /* If not enough data remaining, we can short cut */
3872 while (length >= TCPOLEN_MD5SIG) {
3873 int opcode = *ptr++;
3874 int opsize;
3875
3876 switch (opcode) {
3877 case TCPOPT_EOL:
3878 return NULL;
3879 case TCPOPT_NOP:
3880 length--;
3881 continue;
3882 default:
3883 opsize = *ptr++;
3884 if (opsize < 2 || opsize > length)
3885 return NULL;
3886 if (opcode == TCPOPT_MD5SIG)
3887 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3888 }
3889 ptr += opsize - 2;
3890 length -= opsize;
3891 }
3892 return NULL;
3893}
3894EXPORT_SYMBOL(tcp_parse_md5sig_option);
3895#endif
3896
3897/* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3898 *
3899 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3900 * it can pass through stack. So, the following predicate verifies that
3901 * this segment is not used for anything but congestion avoidance or
3902 * fast retransmit. Moreover, we even are able to eliminate most of such
3903 * second order effects, if we apply some small "replay" window (~RTO)
3904 * to timestamp space.
3905 *
3906 * All these measures still do not guarantee that we reject wrapped ACKs
3907 * on networks with high bandwidth, when sequence space is recycled fastly,
3908 * but it guarantees that such events will be very rare and do not affect
3909 * connection seriously. This doesn't look nice, but alas, PAWS is really
3910 * buggy extension.
3911 *
3912 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3913 * states that events when retransmit arrives after original data are rare.
3914 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3915 * the biggest problem on large power networks even with minor reordering.
3916 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3917 * up to bandwidth of 18Gigabit/sec. 8) ]
3918 */
3919
3920static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3921{
3922 const struct tcp_sock *tp = tcp_sk(sk);
3923 const struct tcphdr *th = tcp_hdr(skb);
3924 u32 seq = TCP_SKB_CB(skb)->seq;
3925 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3926
3927 return (/* 1. Pure ACK with correct sequence number. */
3928 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3929
3930 /* 2. ... and duplicate ACK. */
3931 ack == tp->snd_una &&
3932
3933 /* 3. ... and does not update window. */
3934 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3935
3936 /* 4. ... and sits in replay window. */
3937 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3938}
3939
3940static inline bool tcp_paws_discard(const struct sock *sk,
3941 const struct sk_buff *skb)
3942{
3943 const struct tcp_sock *tp = tcp_sk(sk);
3944
3945 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
3946 !tcp_disordered_ack(sk, skb);
3947}
3948
3949/* Check segment sequence number for validity.
3950 *
3951 * Segment controls are considered valid, if the segment
3952 * fits to the window after truncation to the window. Acceptability
3953 * of data (and SYN, FIN, of course) is checked separately.
3954 * See tcp_data_queue(), for example.
3955 *
3956 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3957 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3958 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3959 * (borrowed from freebsd)
3960 */
3961
3962static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
3963{
3964 return !before(end_seq, tp->rcv_wup) &&
3965 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3966}
3967
3968/* When we get a reset we do this. */
3969void tcp_reset(struct sock *sk)
3970{
3971 trace_tcp_receive_reset(sk);
3972
3973 /* We want the right error as BSD sees it (and indeed as we do). */
3974 switch (sk->sk_state) {
3975 case TCP_SYN_SENT:
3976 sk->sk_err = ECONNREFUSED;
3977 break;
3978 case TCP_CLOSE_WAIT:
3979 sk->sk_err = EPIPE;
3980 break;
3981 case TCP_CLOSE:
3982 return;
3983 default:
3984 sk->sk_err = ECONNRESET;
3985 }
3986 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3987 smp_wmb();
3988
3989 tcp_write_queue_purge(sk);
3990 tcp_done(sk);
3991
3992 if (!sock_flag(sk, SOCK_DEAD))
3993 sk->sk_error_report(sk);
3994}
3995
3996/*
3997 * Process the FIN bit. This now behaves as it is supposed to work
3998 * and the FIN takes effect when it is validly part of sequence
3999 * space. Not before when we get holes.
4000 *
4001 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4002 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4003 * TIME-WAIT)
4004 *
4005 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4006 * close and we go into CLOSING (and later onto TIME-WAIT)
4007 *
4008 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4009 */
4010void tcp_fin(struct sock *sk)
4011{
4012 struct tcp_sock *tp = tcp_sk(sk);
4013
4014 inet_csk_schedule_ack(sk);
4015
4016 sk->sk_shutdown |= RCV_SHUTDOWN;
4017 sock_set_flag(sk, SOCK_DONE);
4018
4019 switch (sk->sk_state) {
4020 case TCP_SYN_RECV:
4021 case TCP_ESTABLISHED:
4022 /* Move to CLOSE_WAIT */
4023 tcp_set_state(sk, TCP_CLOSE_WAIT);
4024 inet_csk(sk)->icsk_ack.pingpong = 1;
4025 break;
4026
4027 case TCP_CLOSE_WAIT:
4028 case TCP_CLOSING:
4029 /* Received a retransmission of the FIN, do
4030 * nothing.
4031 */
4032 break;
4033 case TCP_LAST_ACK:
4034 /* RFC793: Remain in the LAST-ACK state. */
4035 break;
4036
4037 case TCP_FIN_WAIT1:
4038 /* This case occurs when a simultaneous close
4039 * happens, we must ack the received FIN and
4040 * enter the CLOSING state.
4041 */
4042 tcp_send_ack(sk);
4043 tcp_set_state(sk, TCP_CLOSING);
4044 break;
4045 case TCP_FIN_WAIT2:
4046 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4047 tcp_send_ack(sk);
4048 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4049 break;
4050 default:
4051 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4052 * cases we should never reach this piece of code.
4053 */
4054 pr_err("%s: Impossible, sk->sk_state=%d\n",
4055 __func__, sk->sk_state);
4056 break;
4057 }
4058
4059 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4060 * Probably, we should reset in this case. For now drop them.
4061 */
4062 skb_rbtree_purge(&tp->out_of_order_queue);
4063 if (tcp_is_sack(tp))
4064 tcp_sack_reset(&tp->rx_opt);
4065 sk_mem_reclaim(sk);
4066
4067 if (!sock_flag(sk, SOCK_DEAD)) {
4068 sk->sk_state_change(sk);
4069
4070 /* Do not send POLL_HUP for half duplex close. */
4071 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4072 sk->sk_state == TCP_CLOSE)
4073 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4074 else
4075 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4076 }
4077}
4078
4079static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4080 u32 end_seq)
4081{
4082 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4083 if (before(seq, sp->start_seq))
4084 sp->start_seq = seq;
4085 if (after(end_seq, sp->end_seq))
4086 sp->end_seq = end_seq;
4087 return true;
4088 }
4089 return false;
4090}
4091
4092static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4093{
4094 struct tcp_sock *tp = tcp_sk(sk);
4095
4096 if (tcp_is_sack(tp) && sock_net(sk)->ipv4.sysctl_tcp_dsack) {
4097 int mib_idx;
4098
4099 if (before(seq, tp->rcv_nxt))
4100 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4101 else
4102 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4103
4104 NET_INC_STATS(sock_net(sk), mib_idx);
4105
4106 tp->rx_opt.dsack = 1;
4107 tp->duplicate_sack[0].start_seq = seq;
4108 tp->duplicate_sack[0].end_seq = end_seq;
4109 }
4110}
4111
4112static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4113{
4114 struct tcp_sock *tp = tcp_sk(sk);
4115
4116 if (!tp->rx_opt.dsack)
4117 tcp_dsack_set(sk, seq, end_seq);
4118 else
4119 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4120}
4121
4122static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4123{
4124 struct tcp_sock *tp = tcp_sk(sk);
4125
4126 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4127 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4128 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4129 tcp_enter_quickack_mode(sk);
4130
4131 if (tcp_is_sack(tp) && sock_net(sk)->ipv4.sysctl_tcp_dsack) {
4132 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4133
4134 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4135 end_seq = tp->rcv_nxt;
4136 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4137 }
4138 }
4139
4140 tcp_send_ack(sk);
4141}
4142
4143/* These routines update the SACK block as out-of-order packets arrive or
4144 * in-order packets close up the sequence space.
4145 */
4146static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4147{
4148 int this_sack;
4149 struct tcp_sack_block *sp = &tp->selective_acks[0];
4150 struct tcp_sack_block *swalk = sp + 1;
4151
4152 /* See if the recent change to the first SACK eats into
4153 * or hits the sequence space of other SACK blocks, if so coalesce.
4154 */
4155 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4156 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4157 int i;
4158
4159 /* Zap SWALK, by moving every further SACK up by one slot.
4160 * Decrease num_sacks.
4161 */
4162 tp->rx_opt.num_sacks--;
4163 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4164 sp[i] = sp[i + 1];
4165 continue;
4166 }
4167 this_sack++, swalk++;
4168 }
4169}
4170
4171static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4172{
4173 struct tcp_sock *tp = tcp_sk(sk);
4174 struct tcp_sack_block *sp = &tp->selective_acks[0];
4175 int cur_sacks = tp->rx_opt.num_sacks;
4176 int this_sack;
4177
4178 if (!cur_sacks)
4179 goto new_sack;
4180
4181 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4182 if (tcp_sack_extend(sp, seq, end_seq)) {
4183 /* Rotate this_sack to the first one. */
4184 for (; this_sack > 0; this_sack--, sp--)
4185 swap(*sp, *(sp - 1));
4186 if (cur_sacks > 1)
4187 tcp_sack_maybe_coalesce(tp);
4188 return;
4189 }
4190 }
4191
4192 /* Could not find an adjacent existing SACK, build a new one,
4193 * put it at the front, and shift everyone else down. We
4194 * always know there is at least one SACK present already here.
4195 *
4196 * If the sack array is full, forget about the last one.
4197 */
4198 if (this_sack >= TCP_NUM_SACKS) {
4199 this_sack--;
4200 tp->rx_opt.num_sacks--;
4201 sp--;
4202 }
4203 for (; this_sack > 0; this_sack--, sp--)
4204 *sp = *(sp - 1);
4205
4206new_sack:
4207 /* Build the new head SACK, and we're done. */
4208 sp->start_seq = seq;
4209 sp->end_seq = end_seq;
4210 tp->rx_opt.num_sacks++;
4211}
4212
4213/* RCV.NXT advances, some SACKs should be eaten. */
4214
4215static void tcp_sack_remove(struct tcp_sock *tp)
4216{
4217 struct tcp_sack_block *sp = &tp->selective_acks[0];
4218 int num_sacks = tp->rx_opt.num_sacks;
4219 int this_sack;
4220
4221 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4222 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4223 tp->rx_opt.num_sacks = 0;
4224 return;
4225 }
4226
4227 for (this_sack = 0; this_sack < num_sacks;) {
4228 /* Check if the start of the sack is covered by RCV.NXT. */
4229 if (!before(tp->rcv_nxt, sp->start_seq)) {
4230 int i;
4231
4232 /* RCV.NXT must cover all the block! */
4233 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4234
4235 /* Zap this SACK, by moving forward any other SACKS. */
4236 for (i = this_sack+1; i < num_sacks; i++)
4237 tp->selective_acks[i-1] = tp->selective_acks[i];
4238 num_sacks--;
4239 continue;
4240 }
4241 this_sack++;
4242 sp++;
4243 }
4244 tp->rx_opt.num_sacks = num_sacks;
4245}
4246
4247/**
4248 * tcp_try_coalesce - try to merge skb to prior one
4249 * @sk: socket
4250 * @dest: destination queue
4251 * @to: prior buffer
4252 * @from: buffer to add in queue
4253 * @fragstolen: pointer to boolean
4254 *
4255 * Before queueing skb @from after @to, try to merge them
4256 * to reduce overall memory use and queue lengths, if cost is small.
4257 * Packets in ofo or receive queues can stay a long time.
4258 * Better try to coalesce them right now to avoid future collapses.
4259 * Returns true if caller should free @from instead of queueing it
4260 */
4261static bool tcp_try_coalesce(struct sock *sk,
4262 struct sk_buff *to,
4263 struct sk_buff *from,
4264 bool *fragstolen)
4265{
4266 int delta;
4267
4268 *fragstolen = false;
4269
4270 /* Its possible this segment overlaps with prior segment in queue */
4271 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4272 return false;
4273
4274 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4275 return false;
4276
4277 atomic_add(delta, &sk->sk_rmem_alloc);
4278 sk_mem_charge(sk, delta);
4279 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4280 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4281 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4282 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4283
4284 if (TCP_SKB_CB(from)->has_rxtstamp) {
4285 TCP_SKB_CB(to)->has_rxtstamp = true;
4286 to->tstamp = from->tstamp;
4287 }
4288
4289 return true;
4290}
4291
4292static void tcp_drop(struct sock *sk, struct sk_buff *skb)
4293{
4294 sk_drops_add(sk, skb);
4295 __kfree_skb(skb);
4296}
4297
4298/* This one checks to see if we can put data from the
4299 * out_of_order queue into the receive_queue.
4300 */
4301static void tcp_ofo_queue(struct sock *sk)
4302{
4303 struct tcp_sock *tp = tcp_sk(sk);
4304 __u32 dsack_high = tp->rcv_nxt;
4305 bool fin, fragstolen, eaten;
4306 struct sk_buff *skb, *tail;
4307 struct rb_node *p;
4308
4309 p = rb_first(&tp->out_of_order_queue);
4310 while (p) {
4311 skb = rb_to_skb(p);
4312 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4313 break;
4314
4315 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4316 __u32 dsack = dsack_high;
4317 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4318 dsack_high = TCP_SKB_CB(skb)->end_seq;
4319 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4320 }
4321 p = rb_next(p);
4322 rb_erase(&skb->rbnode, &tp->out_of_order_queue);
4323
4324 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) {
4325 SOCK_DEBUG(sk, "ofo packet was already received\n");
4326 tcp_drop(sk, skb);
4327 continue;
4328 }
4329 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4330 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4331 TCP_SKB_CB(skb)->end_seq);
4332
4333 tail = skb_peek_tail(&sk->sk_receive_queue);
4334 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4335 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4336 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
4337 if (!eaten)
4338 __skb_queue_tail(&sk->sk_receive_queue, skb);
4339 else
4340 kfree_skb_partial(skb, fragstolen);
4341
4342 if (unlikely(fin)) {
4343 tcp_fin(sk);
4344 /* tcp_fin() purges tp->out_of_order_queue,
4345 * so we must end this loop right now.
4346 */
4347 break;
4348 }
4349 }
4350}
4351
4352static bool tcp_prune_ofo_queue(struct sock *sk);
4353static int tcp_prune_queue(struct sock *sk);
4354
4355static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4356 unsigned int size)
4357{
4358 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4359 !sk_rmem_schedule(sk, skb, size)) {
4360
4361 if (tcp_prune_queue(sk) < 0)
4362 return -1;
4363
4364 while (!sk_rmem_schedule(sk, skb, size)) {
4365 if (!tcp_prune_ofo_queue(sk))
4366 return -1;
4367 }
4368 }
4369 return 0;
4370}
4371
4372static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4373{
4374 struct tcp_sock *tp = tcp_sk(sk);
4375 struct rb_node **p, *parent;
4376 struct sk_buff *skb1;
4377 u32 seq, end_seq;
4378 bool fragstolen;
4379
4380 tcp_ecn_check_ce(tp, skb);
4381
4382 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4383 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP);
4384 tcp_drop(sk, skb);
4385 return;
4386 }
4387
4388 /* Disable header prediction. */
4389 tp->pred_flags = 0;
4390 inet_csk_schedule_ack(sk);
4391
4392 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4393 seq = TCP_SKB_CB(skb)->seq;
4394 end_seq = TCP_SKB_CB(skb)->end_seq;
4395 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4396 tp->rcv_nxt, seq, end_seq);
4397
4398 p = &tp->out_of_order_queue.rb_node;
4399 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4400 /* Initial out of order segment, build 1 SACK. */
4401 if (tcp_is_sack(tp)) {
4402 tp->rx_opt.num_sacks = 1;
4403 tp->selective_acks[0].start_seq = seq;
4404 tp->selective_acks[0].end_seq = end_seq;
4405 }
4406 rb_link_node(&skb->rbnode, NULL, p);
4407 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4408 tp->ooo_last_skb = skb;
4409 goto end;
4410 }
4411
4412 /* In the typical case, we are adding an skb to the end of the list.
4413 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4414 */
4415 if (tcp_try_coalesce(sk, tp->ooo_last_skb,
4416 skb, &fragstolen)) {
4417coalesce_done:
4418 tcp_grow_window(sk, skb);
4419 kfree_skb_partial(skb, fragstolen);
4420 skb = NULL;
4421 goto add_sack;
4422 }
4423 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4424 if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) {
4425 parent = &tp->ooo_last_skb->rbnode;
4426 p = &parent->rb_right;
4427 goto insert;
4428 }
4429
4430 /* Find place to insert this segment. Handle overlaps on the way. */
4431 parent = NULL;
4432 while (*p) {
4433 parent = *p;
4434 skb1 = rb_to_skb(parent);
4435 if (before(seq, TCP_SKB_CB(skb1)->seq)) {
4436 p = &parent->rb_left;
4437 continue;
4438 }
4439 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4440 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4441 /* All the bits are present. Drop. */
4442 NET_INC_STATS(sock_net(sk),
4443 LINUX_MIB_TCPOFOMERGE);
4444 __kfree_skb(skb);
4445 skb = NULL;
4446 tcp_dsack_set(sk, seq, end_seq);
4447 goto add_sack;
4448 }
4449 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4450 /* Partial overlap. */
4451 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq);
4452 } else {
4453 /* skb's seq == skb1's seq and skb covers skb1.
4454 * Replace skb1 with skb.
4455 */
4456 rb_replace_node(&skb1->rbnode, &skb->rbnode,
4457 &tp->out_of_order_queue);
4458 tcp_dsack_extend(sk,
4459 TCP_SKB_CB(skb1)->seq,
4460 TCP_SKB_CB(skb1)->end_seq);
4461 NET_INC_STATS(sock_net(sk),
4462 LINUX_MIB_TCPOFOMERGE);
4463 __kfree_skb(skb1);
4464 goto merge_right;
4465 }
4466 } else if (tcp_try_coalesce(sk, skb1,
4467 skb, &fragstolen)) {
4468 goto coalesce_done;
4469 }
4470 p = &parent->rb_right;
4471 }
4472insert:
4473 /* Insert segment into RB tree. */
4474 rb_link_node(&skb->rbnode, parent, p);
4475 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4476
4477merge_right:
4478 /* Remove other segments covered by skb. */
4479 while ((skb1 = skb_rb_next(skb)) != NULL) {
4480 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4481 break;
4482 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4483 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4484 end_seq);
4485 break;
4486 }
4487 rb_erase(&skb1->rbnode, &tp->out_of_order_queue);
4488 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4489 TCP_SKB_CB(skb1)->end_seq);
4490 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4491 tcp_drop(sk, skb1);
4492 }
4493 /* If there is no skb after us, we are the last_skb ! */
4494 if (!skb1)
4495 tp->ooo_last_skb = skb;
4496
4497add_sack:
4498 if (tcp_is_sack(tp))
4499 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4500end:
4501 if (skb) {
4502 tcp_grow_window(sk, skb);
4503 skb_condense(skb);
4504 skb_set_owner_r(skb, sk);
4505 }
4506}
4507
4508static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
4509 bool *fragstolen)
4510{
4511 int eaten;
4512 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4513
4514 __skb_pull(skb, hdrlen);
4515 eaten = (tail &&
4516 tcp_try_coalesce(sk, tail,
4517 skb, fragstolen)) ? 1 : 0;
4518 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
4519 if (!eaten) {
4520 __skb_queue_tail(&sk->sk_receive_queue, skb);
4521 skb_set_owner_r(skb, sk);
4522 }
4523 return eaten;
4524}
4525
4526int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4527{
4528 struct sk_buff *skb;
4529 int err = -ENOMEM;
4530 int data_len = 0;
4531 bool fragstolen;
4532
4533 if (size == 0)
4534 return 0;
4535
4536 if (size > PAGE_SIZE) {
4537 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
4538
4539 data_len = npages << PAGE_SHIFT;
4540 size = data_len + (size & ~PAGE_MASK);
4541 }
4542 skb = alloc_skb_with_frags(size - data_len, data_len,
4543 PAGE_ALLOC_COSTLY_ORDER,
4544 &err, sk->sk_allocation);
4545 if (!skb)
4546 goto err;
4547
4548 skb_put(skb, size - data_len);
4549 skb->data_len = data_len;
4550 skb->len = size;
4551
4552 if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4553 goto err_free;
4554
4555 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
4556 if (err)
4557 goto err_free;
4558
4559 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4560 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4561 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4562
4563 if (tcp_queue_rcv(sk, skb, 0, &fragstolen)) {
4564 WARN_ON_ONCE(fragstolen); /* should not happen */
4565 __kfree_skb(skb);
4566 }
4567 return size;
4568
4569err_free:
4570 kfree_skb(skb);
4571err:
4572 return err;
4573
4574}
4575
4576static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4577{
4578 struct tcp_sock *tp = tcp_sk(sk);
4579 bool fragstolen;
4580 int eaten;
4581
4582 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
4583 __kfree_skb(skb);
4584 return;
4585 }
4586 skb_dst_drop(skb);
4587 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
4588
4589 tcp_ecn_accept_cwr(tp, skb);
4590
4591 tp->rx_opt.dsack = 0;
4592
4593 /* Queue data for delivery to the user.
4594 * Packets in sequence go to the receive queue.
4595 * Out of sequence packets to the out_of_order_queue.
4596 */
4597 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4598 if (tcp_receive_window(tp) == 0)
4599 goto out_of_window;
4600
4601 /* Ok. In sequence. In window. */
4602queue_and_out:
4603 if (skb_queue_len(&sk->sk_receive_queue) == 0)
4604 sk_forced_mem_schedule(sk, skb->truesize);
4605 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4606 goto drop;
4607
4608 eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
4609 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4610 if (skb->len)
4611 tcp_event_data_recv(sk, skb);
4612 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4613 tcp_fin(sk);
4614
4615 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4616 tcp_ofo_queue(sk);
4617
4618 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4619 * gap in queue is filled.
4620 */
4621 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
4622 inet_csk(sk)->icsk_ack.pingpong = 0;
4623 }
4624
4625 if (tp->rx_opt.num_sacks)
4626 tcp_sack_remove(tp);
4627
4628 tcp_fast_path_check(sk);
4629
4630 if (eaten > 0)
4631 kfree_skb_partial(skb, fragstolen);
4632 if (!sock_flag(sk, SOCK_DEAD))
4633 sk->sk_data_ready(sk);
4634 return;
4635 }
4636
4637 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4638 /* A retransmit, 2nd most common case. Force an immediate ack. */
4639 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4640 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4641
4642out_of_window:
4643 tcp_enter_quickack_mode(sk);
4644 inet_csk_schedule_ack(sk);
4645drop:
4646 tcp_drop(sk, skb);
4647 return;
4648 }
4649
4650 /* Out of window. F.e. zero window probe. */
4651 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4652 goto out_of_window;
4653
4654 tcp_enter_quickack_mode(sk);
4655
4656 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4657 /* Partial packet, seq < rcv_next < end_seq */
4658 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4659 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4660 TCP_SKB_CB(skb)->end_seq);
4661
4662 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4663
4664 /* If window is closed, drop tail of packet. But after
4665 * remembering D-SACK for its head made in previous line.
4666 */
4667 if (!tcp_receive_window(tp))
4668 goto out_of_window;
4669 goto queue_and_out;
4670 }
4671
4672 tcp_data_queue_ofo(sk, skb);
4673}
4674
4675static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list)
4676{
4677 if (list)
4678 return !skb_queue_is_last(list, skb) ? skb->next : NULL;
4679
4680 return skb_rb_next(skb);
4681}
4682
4683static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4684 struct sk_buff_head *list,
4685 struct rb_root *root)
4686{
4687 struct sk_buff *next = tcp_skb_next(skb, list);
4688
4689 if (list)
4690 __skb_unlink(skb, list);
4691 else
4692 rb_erase(&skb->rbnode, root);
4693
4694 __kfree_skb(skb);
4695 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4696
4697 return next;
4698}
4699
4700/* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
4701void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb)
4702{
4703 struct rb_node **p = &root->rb_node;
4704 struct rb_node *parent = NULL;
4705 struct sk_buff *skb1;
4706
4707 while (*p) {
4708 parent = *p;
4709 skb1 = rb_to_skb(parent);
4710 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq))
4711 p = &parent->rb_left;
4712 else
4713 p = &parent->rb_right;
4714 }
4715 rb_link_node(&skb->rbnode, parent, p);
4716 rb_insert_color(&skb->rbnode, root);
4717}
4718
4719/* Collapse contiguous sequence of skbs head..tail with
4720 * sequence numbers start..end.
4721 *
4722 * If tail is NULL, this means until the end of the queue.
4723 *
4724 * Segments with FIN/SYN are not collapsed (only because this
4725 * simplifies code)
4726 */
4727static void
4728tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root,
4729 struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end)
4730{
4731 struct sk_buff *skb = head, *n;
4732 struct sk_buff_head tmp;
4733 bool end_of_skbs;
4734
4735 /* First, check that queue is collapsible and find
4736 * the point where collapsing can be useful.
4737 */
4738restart:
4739 for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) {
4740 n = tcp_skb_next(skb, list);
4741
4742 /* No new bits? It is possible on ofo queue. */
4743 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4744 skb = tcp_collapse_one(sk, skb, list, root);
4745 if (!skb)
4746 break;
4747 goto restart;
4748 }
4749
4750 /* The first skb to collapse is:
4751 * - not SYN/FIN and
4752 * - bloated or contains data before "start" or
4753 * overlaps to the next one.
4754 */
4755 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
4756 (tcp_win_from_space(sk, skb->truesize) > skb->len ||
4757 before(TCP_SKB_CB(skb)->seq, start))) {
4758 end_of_skbs = false;
4759 break;
4760 }
4761
4762 if (n && n != tail &&
4763 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) {
4764 end_of_skbs = false;
4765 break;
4766 }
4767
4768 /* Decided to skip this, advance start seq. */
4769 start = TCP_SKB_CB(skb)->end_seq;
4770 }
4771 if (end_of_skbs ||
4772 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4773 return;
4774
4775 __skb_queue_head_init(&tmp);
4776
4777 while (before(start, end)) {
4778 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
4779 struct sk_buff *nskb;
4780
4781 nskb = alloc_skb(copy, GFP_ATOMIC);
4782 if (!nskb)
4783 break;
4784
4785 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4786 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4787 if (list)
4788 __skb_queue_before(list, skb, nskb);
4789 else
4790 __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */
4791 skb_set_owner_r(nskb, sk);
4792
4793 /* Copy data, releasing collapsed skbs. */
4794 while (copy > 0) {
4795 int offset = start - TCP_SKB_CB(skb)->seq;
4796 int size = TCP_SKB_CB(skb)->end_seq - start;
4797
4798 BUG_ON(offset < 0);
4799 if (size > 0) {
4800 size = min(copy, size);
4801 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4802 BUG();
4803 TCP_SKB_CB(nskb)->end_seq += size;
4804 copy -= size;
4805 start += size;
4806 }
4807 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4808 skb = tcp_collapse_one(sk, skb, list, root);
4809 if (!skb ||
4810 skb == tail ||
4811 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4812 goto end;
4813 }
4814 }
4815 }
4816end:
4817 skb_queue_walk_safe(&tmp, skb, n)
4818 tcp_rbtree_insert(root, skb);
4819}
4820
4821/* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4822 * and tcp_collapse() them until all the queue is collapsed.
4823 */
4824static void tcp_collapse_ofo_queue(struct sock *sk)
4825{
4826 struct tcp_sock *tp = tcp_sk(sk);
4827 struct sk_buff *skb, *head;
4828 u32 start, end;
4829
4830 skb = skb_rb_first(&tp->out_of_order_queue);
4831new_range:
4832 if (!skb) {
4833 tp->ooo_last_skb = skb_rb_last(&tp->out_of_order_queue);
4834 return;
4835 }
4836 start = TCP_SKB_CB(skb)->seq;
4837 end = TCP_SKB_CB(skb)->end_seq;
4838
4839 for (head = skb;;) {
4840 skb = skb_rb_next(skb);
4841
4842 /* Range is terminated when we see a gap or when
4843 * we are at the queue end.
4844 */
4845 if (!skb ||
4846 after(TCP_SKB_CB(skb)->seq, end) ||
4847 before(TCP_SKB_CB(skb)->end_seq, start)) {
4848 tcp_collapse(sk, NULL, &tp->out_of_order_queue,
4849 head, skb, start, end);
4850 goto new_range;
4851 }
4852
4853 if (unlikely(before(TCP_SKB_CB(skb)->seq, start)))
4854 start = TCP_SKB_CB(skb)->seq;
4855 if (after(TCP_SKB_CB(skb)->end_seq, end))
4856 end = TCP_SKB_CB(skb)->end_seq;
4857 }
4858}
4859
4860/*
4861 * Clean the out-of-order queue to make room.
4862 * We drop high sequences packets to :
4863 * 1) Let a chance for holes to be filled.
4864 * 2) not add too big latencies if thousands of packets sit there.
4865 * (But if application shrinks SO_RCVBUF, we could still end up
4866 * freeing whole queue here)
4867 *
4868 * Return true if queue has shrunk.
4869 */
4870static bool tcp_prune_ofo_queue(struct sock *sk)
4871{
4872 struct tcp_sock *tp = tcp_sk(sk);
4873 struct rb_node *node, *prev;
4874
4875 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
4876 return false;
4877
4878 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED);
4879 node = &tp->ooo_last_skb->rbnode;
4880 do {
4881 prev = rb_prev(node);
4882 rb_erase(node, &tp->out_of_order_queue);
4883 tcp_drop(sk, rb_to_skb(node));
4884 sk_mem_reclaim(sk);
4885 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf &&
4886 !tcp_under_memory_pressure(sk))
4887 break;
4888 node = prev;
4889 } while (node);
4890 tp->ooo_last_skb = rb_to_skb(prev);
4891
4892 /* Reset SACK state. A conforming SACK implementation will
4893 * do the same at a timeout based retransmit. When a connection
4894 * is in a sad state like this, we care only about integrity
4895 * of the connection not performance.
4896 */
4897 if (tp->rx_opt.sack_ok)
4898 tcp_sack_reset(&tp->rx_opt);
4899 return true;
4900}
4901
4902/* Reduce allocated memory if we can, trying to get
4903 * the socket within its memory limits again.
4904 *
4905 * Return less than zero if we should start dropping frames
4906 * until the socket owning process reads some of the data
4907 * to stabilize the situation.
4908 */
4909static int tcp_prune_queue(struct sock *sk)
4910{
4911 struct tcp_sock *tp = tcp_sk(sk);
4912
4913 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4914
4915 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED);
4916
4917 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4918 tcp_clamp_window(sk);
4919 else if (tcp_under_memory_pressure(sk))
4920 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4921
4922 tcp_collapse_ofo_queue(sk);
4923 if (!skb_queue_empty(&sk->sk_receive_queue))
4924 tcp_collapse(sk, &sk->sk_receive_queue, NULL,
4925 skb_peek(&sk->sk_receive_queue),
4926 NULL,
4927 tp->copied_seq, tp->rcv_nxt);
4928 sk_mem_reclaim(sk);
4929
4930 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4931 return 0;
4932
4933 /* Collapsing did not help, destructive actions follow.
4934 * This must not ever occur. */
4935
4936 tcp_prune_ofo_queue(sk);
4937
4938 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4939 return 0;
4940
4941 /* If we are really being abused, tell the caller to silently
4942 * drop receive data on the floor. It will get retransmitted
4943 * and hopefully then we'll have sufficient space.
4944 */
4945 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED);
4946
4947 /* Massive buffer overcommit. */
4948 tp->pred_flags = 0;
4949 return -1;
4950}
4951
4952static bool tcp_should_expand_sndbuf(const struct sock *sk)
4953{
4954 const struct tcp_sock *tp = tcp_sk(sk);
4955
4956 /* If the user specified a specific send buffer setting, do
4957 * not modify it.
4958 */
4959 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4960 return false;
4961
4962 /* If we are under global TCP memory pressure, do not expand. */
4963 if (tcp_under_memory_pressure(sk))
4964 return false;
4965
4966 /* If we are under soft global TCP memory pressure, do not expand. */
4967 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
4968 return false;
4969
4970 /* If we filled the congestion window, do not expand. */
4971 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
4972 return false;
4973
4974 return true;
4975}
4976
4977/* When incoming ACK allowed to free some skb from write_queue,
4978 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4979 * on the exit from tcp input handler.
4980 *
4981 * PROBLEM: sndbuf expansion does not work well with largesend.
4982 */
4983static void tcp_new_space(struct sock *sk)
4984{
4985 struct tcp_sock *tp = tcp_sk(sk);
4986
4987 if (tcp_should_expand_sndbuf(sk)) {
4988 tcp_sndbuf_expand(sk);
4989 tp->snd_cwnd_stamp = tcp_jiffies32;
4990 }
4991
4992 sk->sk_write_space(sk);
4993}
4994
4995static void tcp_check_space(struct sock *sk)
4996{
4997 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4998 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4999 /* pairs with tcp_poll() */
5000 smp_mb();
5001 if (sk->sk_socket &&
5002 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
5003 tcp_new_space(sk);
5004 if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5005 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
5006 }
5007 }
5008}
5009
5010static inline void tcp_data_snd_check(struct sock *sk)
5011{
5012 tcp_push_pending_frames(sk);
5013 tcp_check_space(sk);
5014}
5015
5016/*
5017 * Check if sending an ack is needed.
5018 */
5019static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5020{
5021 struct tcp_sock *tp = tcp_sk(sk);
5022
5023 /* More than one full frame received... */
5024 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5025 /* ... and right edge of window advances far enough.
5026 * (tcp_recvmsg() will send ACK otherwise). Or...
5027 */
5028 __tcp_select_window(sk) >= tp->rcv_wnd) ||
5029 /* We ACK each frame or... */
5030 tcp_in_quickack_mode(sk) ||
5031 /* We have out of order data. */
5032 (ofo_possible && !RB_EMPTY_ROOT(&tp->out_of_order_queue))) {
5033 /* Then ack it now */
5034 tcp_send_ack(sk);
5035 } else {
5036 /* Else, send delayed ack. */
5037 tcp_send_delayed_ack(sk);
5038 }
5039}
5040
5041static inline void tcp_ack_snd_check(struct sock *sk)
5042{
5043 if (!inet_csk_ack_scheduled(sk)) {
5044 /* We sent a data segment already. */
5045 return;
5046 }
5047 __tcp_ack_snd_check(sk, 1);
5048}
5049
5050/*
5051 * This routine is only called when we have urgent data
5052 * signaled. Its the 'slow' part of tcp_urg. It could be
5053 * moved inline now as tcp_urg is only called from one
5054 * place. We handle URGent data wrong. We have to - as
5055 * BSD still doesn't use the correction from RFC961.
5056 * For 1003.1g we should support a new option TCP_STDURG to permit
5057 * either form (or just set the sysctl tcp_stdurg).
5058 */
5059
5060static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5061{
5062 struct tcp_sock *tp = tcp_sk(sk);
5063 u32 ptr = ntohs(th->urg_ptr);
5064
5065 if (ptr && !sock_net(sk)->ipv4.sysctl_tcp_stdurg)
5066 ptr--;
5067 ptr += ntohl(th->seq);
5068
5069 /* Ignore urgent data that we've already seen and read. */
5070 if (after(tp->copied_seq, ptr))
5071 return;
5072
5073 /* Do not replay urg ptr.
5074 *
5075 * NOTE: interesting situation not covered by specs.
5076 * Misbehaving sender may send urg ptr, pointing to segment,
5077 * which we already have in ofo queue. We are not able to fetch
5078 * such data and will stay in TCP_URG_NOTYET until will be eaten
5079 * by recvmsg(). Seems, we are not obliged to handle such wicked
5080 * situations. But it is worth to think about possibility of some
5081 * DoSes using some hypothetical application level deadlock.
5082 */
5083 if (before(ptr, tp->rcv_nxt))
5084 return;
5085
5086 /* Do we already have a newer (or duplicate) urgent pointer? */
5087 if (tp->urg_data && !after(ptr, tp->urg_seq))
5088 return;
5089
5090 /* Tell the world about our new urgent pointer. */
5091 sk_send_sigurg(sk);
5092
5093 /* We may be adding urgent data when the last byte read was
5094 * urgent. To do this requires some care. We cannot just ignore
5095 * tp->copied_seq since we would read the last urgent byte again
5096 * as data, nor can we alter copied_seq until this data arrives
5097 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5098 *
5099 * NOTE. Double Dutch. Rendering to plain English: author of comment
5100 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5101 * and expect that both A and B disappear from stream. This is _wrong_.
5102 * Though this happens in BSD with high probability, this is occasional.
5103 * Any application relying on this is buggy. Note also, that fix "works"
5104 * only in this artificial test. Insert some normal data between A and B and we will
5105 * decline of BSD again. Verdict: it is better to remove to trap
5106 * buggy users.
5107 */
5108 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5109 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5110 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5111 tp->copied_seq++;
5112 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5113 __skb_unlink(skb, &sk->sk_receive_queue);
5114 __kfree_skb(skb);
5115 }
5116 }
5117
5118 tp->urg_data = TCP_URG_NOTYET;
5119 tp->urg_seq = ptr;
5120
5121 /* Disable header prediction. */
5122 tp->pred_flags = 0;
5123}
5124
5125/* This is the 'fast' part of urgent handling. */
5126static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5127{
5128 struct tcp_sock *tp = tcp_sk(sk);
5129
5130 /* Check if we get a new urgent pointer - normally not. */
5131 if (th->urg)
5132 tcp_check_urg(sk, th);
5133
5134 /* Do we wait for any urgent data? - normally not... */
5135 if (tp->urg_data == TCP_URG_NOTYET) {
5136 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5137 th->syn;
5138
5139 /* Is the urgent pointer pointing into this packet? */
5140 if (ptr < skb->len) {
5141 u8 tmp;
5142 if (skb_copy_bits(skb, ptr, &tmp, 1))
5143 BUG();
5144 tp->urg_data = TCP_URG_VALID | tmp;
5145 if (!sock_flag(sk, SOCK_DEAD))
5146 sk->sk_data_ready(sk);
5147 }
5148 }
5149}
5150
5151/* Accept RST for rcv_nxt - 1 after a FIN.
5152 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a
5153 * FIN is sent followed by a RST packet. The RST is sent with the same
5154 * sequence number as the FIN, and thus according to RFC 5961 a challenge
5155 * ACK should be sent. However, Mac OSX rate limits replies to challenge
5156 * ACKs on the closed socket. In addition middleboxes can drop either the
5157 * challenge ACK or a subsequent RST.
5158 */
5159static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb)
5160{
5161 struct tcp_sock *tp = tcp_sk(sk);
5162
5163 return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) &&
5164 (1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK |
5165 TCPF_CLOSING));
5166}
5167
5168/* Does PAWS and seqno based validation of an incoming segment, flags will
5169 * play significant role here.
5170 */
5171static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5172 const struct tcphdr *th, int syn_inerr)
5173{
5174 struct tcp_sock *tp = tcp_sk(sk);
5175 bool rst_seq_match = false;
5176
5177 /* RFC1323: H1. Apply PAWS check first. */
5178 if (tcp_fast_parse_options(sock_net(sk), skb, th, tp) &&
5179 tp->rx_opt.saw_tstamp &&
5180 tcp_paws_discard(sk, skb)) {
5181 if (!th->rst) {
5182 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5183 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5184 LINUX_MIB_TCPACKSKIPPEDPAWS,
5185 &tp->last_oow_ack_time))
5186 tcp_send_dupack(sk, skb);
5187 goto discard;
5188 }
5189 /* Reset is accepted even if it did not pass PAWS. */
5190 }
5191
5192 /* Step 1: check sequence number */
5193 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5194 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5195 * (RST) segments are validated by checking their SEQ-fields."
5196 * And page 69: "If an incoming segment is not acceptable,
5197 * an acknowledgment should be sent in reply (unless the RST
5198 * bit is set, if so drop the segment and return)".
5199 */
5200 if (!th->rst) {
5201 if (th->syn)
5202 goto syn_challenge;
5203 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5204 LINUX_MIB_TCPACKSKIPPEDSEQ,
5205 &tp->last_oow_ack_time))
5206 tcp_send_dupack(sk, skb);
5207 } else if (tcp_reset_check(sk, skb)) {
5208 tcp_reset(sk);
5209 }
5210 goto discard;
5211 }
5212
5213 /* Step 2: check RST bit */
5214 if (th->rst) {
5215 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a
5216 * FIN and SACK too if available):
5217 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or
5218 * the right-most SACK block,
5219 * then
5220 * RESET the connection
5221 * else
5222 * Send a challenge ACK
5223 */
5224 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt ||
5225 tcp_reset_check(sk, skb)) {
5226 rst_seq_match = true;
5227 } else if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) {
5228 struct tcp_sack_block *sp = &tp->selective_acks[0];
5229 int max_sack = sp[0].end_seq;
5230 int this_sack;
5231
5232 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;
5233 ++this_sack) {
5234 max_sack = after(sp[this_sack].end_seq,
5235 max_sack) ?
5236 sp[this_sack].end_seq : max_sack;
5237 }
5238
5239 if (TCP_SKB_CB(skb)->seq == max_sack)
5240 rst_seq_match = true;
5241 }
5242
5243 if (rst_seq_match)
5244 tcp_reset(sk);
5245 else {
5246 /* Disable TFO if RST is out-of-order
5247 * and no data has been received
5248 * for current active TFO socket
5249 */
5250 if (tp->syn_fastopen && !tp->data_segs_in &&
5251 sk->sk_state == TCP_ESTABLISHED)
5252 tcp_fastopen_active_disable(sk);
5253 tcp_send_challenge_ack(sk, skb);
5254 }
5255 goto discard;
5256 }
5257
5258 /* step 3: check security and precedence [ignored] */
5259
5260 /* step 4: Check for a SYN
5261 * RFC 5961 4.2 : Send a challenge ack
5262 */
5263 if (th->syn) {
5264syn_challenge:
5265 if (syn_inerr)
5266 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5267 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5268 tcp_send_challenge_ack(sk, skb);
5269 goto discard;
5270 }
5271
5272 return true;
5273
5274discard:
5275 tcp_drop(sk, skb);
5276 return false;
5277}
5278
5279/*
5280 * TCP receive function for the ESTABLISHED state.
5281 *
5282 * It is split into a fast path and a slow path. The fast path is
5283 * disabled when:
5284 * - A zero window was announced from us - zero window probing
5285 * is only handled properly in the slow path.
5286 * - Out of order segments arrived.
5287 * - Urgent data is expected.
5288 * - There is no buffer space left
5289 * - Unexpected TCP flags/window values/header lengths are received
5290 * (detected by checking the TCP header against pred_flags)
5291 * - Data is sent in both directions. Fast path only supports pure senders
5292 * or pure receivers (this means either the sequence number or the ack
5293 * value must stay constant)
5294 * - Unexpected TCP option.
5295 *
5296 * When these conditions are not satisfied it drops into a standard
5297 * receive procedure patterned after RFC793 to handle all cases.
5298 * The first three cases are guaranteed by proper pred_flags setting,
5299 * the rest is checked inline. Fast processing is turned on in
5300 * tcp_data_queue when everything is OK.
5301 */
5302void tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5303 const struct tcphdr *th)
5304{
5305 unsigned int len = skb->len;
5306 struct tcp_sock *tp = tcp_sk(sk);
5307
5308 /* TCP congestion window tracking */
5309 trace_tcp_probe(sk, skb);
5310
5311 tcp_mstamp_refresh(tp);
5312 if (unlikely(!sk->sk_rx_dst))
5313 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5314 /*
5315 * Header prediction.
5316 * The code loosely follows the one in the famous
5317 * "30 instruction TCP receive" Van Jacobson mail.
5318 *
5319 * Van's trick is to deposit buffers into socket queue
5320 * on a device interrupt, to call tcp_recv function
5321 * on the receive process context and checksum and copy
5322 * the buffer to user space. smart...
5323 *
5324 * Our current scheme is not silly either but we take the
5325 * extra cost of the net_bh soft interrupt processing...
5326 * We do checksum and copy also but from device to kernel.
5327 */
5328
5329 tp->rx_opt.saw_tstamp = 0;
5330
5331 /* pred_flags is 0xS?10 << 16 + snd_wnd
5332 * if header_prediction is to be made
5333 * 'S' will always be tp->tcp_header_len >> 2
5334 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5335 * turn it off (when there are holes in the receive
5336 * space for instance)
5337 * PSH flag is ignored.
5338 */
5339
5340 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5341 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5342 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5343 int tcp_header_len = tp->tcp_header_len;
5344
5345 /* Timestamp header prediction: tcp_header_len
5346 * is automatically equal to th->doff*4 due to pred_flags
5347 * match.
5348 */
5349
5350 /* Check timestamp */
5351 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5352 /* No? Slow path! */
5353 if (!tcp_parse_aligned_timestamp(tp, th))
5354 goto slow_path;
5355
5356 /* If PAWS failed, check it more carefully in slow path */
5357 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5358 goto slow_path;
5359
5360 /* DO NOT update ts_recent here, if checksum fails
5361 * and timestamp was corrupted part, it will result
5362 * in a hung connection since we will drop all
5363 * future packets due to the PAWS test.
5364 */
5365 }
5366
5367 if (len <= tcp_header_len) {
5368 /* Bulk data transfer: sender */
5369 if (len == tcp_header_len) {
5370 /* Predicted packet is in window by definition.
5371 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5372 * Hence, check seq<=rcv_wup reduces to:
5373 */
5374 if (tcp_header_len ==
5375 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5376 tp->rcv_nxt == tp->rcv_wup)
5377 tcp_store_ts_recent(tp);
5378
5379 /* We know that such packets are checksummed
5380 * on entry.
5381 */
5382 tcp_ack(sk, skb, 0);
5383 __kfree_skb(skb);
5384 tcp_data_snd_check(sk);
5385 return;
5386 } else { /* Header too small */
5387 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5388 goto discard;
5389 }
5390 } else {
5391 int eaten = 0;
5392 bool fragstolen = false;
5393
5394 if (tcp_checksum_complete(skb))
5395 goto csum_error;
5396
5397 if ((int)skb->truesize > sk->sk_forward_alloc)
5398 goto step5;
5399
5400 /* Predicted packet is in window by definition.
5401 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5402 * Hence, check seq<=rcv_wup reduces to:
5403 */
5404 if (tcp_header_len ==
5405 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5406 tp->rcv_nxt == tp->rcv_wup)
5407 tcp_store_ts_recent(tp);
5408
5409 tcp_rcv_rtt_measure_ts(sk, skb);
5410
5411 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS);
5412
5413 /* Bulk data transfer: receiver */
5414 eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5415 &fragstolen);
5416
5417 tcp_event_data_recv(sk, skb);
5418
5419 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5420 /* Well, only one small jumplet in fast path... */
5421 tcp_ack(sk, skb, FLAG_DATA);
5422 tcp_data_snd_check(sk);
5423 if (!inet_csk_ack_scheduled(sk))
5424 goto no_ack;
5425 }
5426
5427 __tcp_ack_snd_check(sk, 0);
5428no_ack:
5429 if (eaten)
5430 kfree_skb_partial(skb, fragstolen);
5431 sk->sk_data_ready(sk);
5432 return;
5433 }
5434 }
5435
5436slow_path:
5437 if (len < (th->doff << 2) || tcp_checksum_complete(skb))
5438 goto csum_error;
5439
5440 if (!th->ack && !th->rst && !th->syn)
5441 goto discard;
5442
5443 /*
5444 * Standard slow path.
5445 */
5446
5447 if (!tcp_validate_incoming(sk, skb, th, 1))
5448 return;
5449
5450step5:
5451 if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5452 goto discard;
5453
5454 tcp_rcv_rtt_measure_ts(sk, skb);
5455
5456 /* Process urgent data. */
5457 tcp_urg(sk, skb, th);
5458
5459 /* step 7: process the segment text */
5460 tcp_data_queue(sk, skb);
5461
5462 tcp_data_snd_check(sk);
5463 tcp_ack_snd_check(sk);
5464 return;
5465
5466csum_error:
5467 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
5468 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5469
5470discard:
5471 tcp_drop(sk, skb);
5472}
5473EXPORT_SYMBOL(tcp_rcv_established);
5474
5475void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5476{
5477 struct tcp_sock *tp = tcp_sk(sk);
5478 struct inet_connection_sock *icsk = inet_csk(sk);
5479
5480 tcp_set_state(sk, TCP_ESTABLISHED);
5481 icsk->icsk_ack.lrcvtime = tcp_jiffies32;
5482
5483 if (skb) {
5484 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5485 security_inet_conn_established(sk, skb);
5486 }
5487
5488 tcp_init_transfer(sk, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB);
5489
5490 /* Prevent spurious tcp_cwnd_restart() on first data
5491 * packet.
5492 */
5493 tp->lsndtime = tcp_jiffies32;
5494
5495 if (sock_flag(sk, SOCK_KEEPOPEN))
5496 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5497
5498 if (!tp->rx_opt.snd_wscale)
5499 __tcp_fast_path_on(tp, tp->snd_wnd);
5500 else
5501 tp->pred_flags = 0;
5502}
5503
5504static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5505 struct tcp_fastopen_cookie *cookie)
5506{
5507 struct tcp_sock *tp = tcp_sk(sk);
5508 struct sk_buff *data = tp->syn_data ? tcp_rtx_queue_head(sk) : NULL;
5509 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
5510 bool syn_drop = false;
5511
5512 if (mss == tp->rx_opt.user_mss) {
5513 struct tcp_options_received opt;
5514
5515 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5516 tcp_clear_options(&opt);
5517 opt.user_mss = opt.mss_clamp = 0;
5518 tcp_parse_options(sock_net(sk), synack, &opt, 0, NULL);
5519 mss = opt.mss_clamp;
5520 }
5521
5522 if (!tp->syn_fastopen) {
5523 /* Ignore an unsolicited cookie */
5524 cookie->len = -1;
5525 } else if (tp->total_retrans) {
5526 /* SYN timed out and the SYN-ACK neither has a cookie nor
5527 * acknowledges data. Presumably the remote received only
5528 * the retransmitted (regular) SYNs: either the original
5529 * SYN-data or the corresponding SYN-ACK was dropped.
5530 */
5531 syn_drop = (cookie->len < 0 && data);
5532 } else if (cookie->len < 0 && !tp->syn_data) {
5533 /* We requested a cookie but didn't get it. If we did not use
5534 * the (old) exp opt format then try so next time (try_exp=1).
5535 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5536 */
5537 try_exp = tp->syn_fastopen_exp ? 2 : 1;
5538 }
5539
5540 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
5541
5542 if (data) { /* Retransmit unacked data in SYN */
5543 skb_rbtree_walk_from(data) {
5544 if (__tcp_retransmit_skb(sk, data, 1))
5545 break;
5546 }
5547 tcp_rearm_rto(sk);
5548 NET_INC_STATS(sock_net(sk),
5549 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
5550 return true;
5551 }
5552 tp->syn_data_acked = tp->syn_data;
5553 if (tp->syn_data_acked)
5554 NET_INC_STATS(sock_net(sk),
5555 LINUX_MIB_TCPFASTOPENACTIVE);
5556
5557 tcp_fastopen_add_skb(sk, synack);
5558
5559 return false;
5560}
5561
5562static void smc_check_reset_syn(struct tcp_sock *tp)
5563{
5564#if IS_ENABLED(CONFIG_SMC)
5565 if (static_branch_unlikely(&tcp_have_smc)) {
5566 if (tp->syn_smc && !tp->rx_opt.smc_ok)
5567 tp->syn_smc = 0;
5568 }
5569#endif
5570}
5571
5572static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5573 const struct tcphdr *th)
5574{
5575 struct inet_connection_sock *icsk = inet_csk(sk);
5576 struct tcp_sock *tp = tcp_sk(sk);
5577 struct tcp_fastopen_cookie foc = { .len = -1 };
5578 int saved_clamp = tp->rx_opt.mss_clamp;
5579 bool fastopen_fail;
5580
5581 tcp_parse_options(sock_net(sk), skb, &tp->rx_opt, 0, &foc);
5582 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
5583 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
5584
5585 if (th->ack) {
5586 /* rfc793:
5587 * "If the state is SYN-SENT then
5588 * first check the ACK bit
5589 * If the ACK bit is set
5590 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5591 * a reset (unless the RST bit is set, if so drop
5592 * the segment and return)"
5593 */
5594 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5595 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
5596 goto reset_and_undo;
5597
5598 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5599 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5600 tcp_time_stamp(tp))) {
5601 NET_INC_STATS(sock_net(sk),
5602 LINUX_MIB_PAWSACTIVEREJECTED);
5603 goto reset_and_undo;
5604 }
5605
5606 /* Now ACK is acceptable.
5607 *
5608 * "If the RST bit is set
5609 * If the ACK was acceptable then signal the user "error:
5610 * connection reset", drop the segment, enter CLOSED state,
5611 * delete TCB, and return."
5612 */
5613
5614 if (th->rst) {
5615 tcp_reset(sk);
5616 goto discard;
5617 }
5618
5619 /* rfc793:
5620 * "fifth, if neither of the SYN or RST bits is set then
5621 * drop the segment and return."
5622 *
5623 * See note below!
5624 * --ANK(990513)
5625 */
5626 if (!th->syn)
5627 goto discard_and_undo;
5628
5629 /* rfc793:
5630 * "If the SYN bit is on ...
5631 * are acceptable then ...
5632 * (our SYN has been ACKed), change the connection
5633 * state to ESTABLISHED..."
5634 */
5635
5636 tcp_ecn_rcv_synack(tp, th);
5637
5638 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5639 tcp_ack(sk, skb, FLAG_SLOWPATH);
5640
5641 /* Ok.. it's good. Set up sequence numbers and
5642 * move to established.
5643 */
5644 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5645 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5646
5647 /* RFC1323: The window in SYN & SYN/ACK segments is
5648 * never scaled.
5649 */
5650 tp->snd_wnd = ntohs(th->window);
5651
5652 if (!tp->rx_opt.wscale_ok) {
5653 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5654 tp->window_clamp = min(tp->window_clamp, 65535U);
5655 }
5656
5657 if (tp->rx_opt.saw_tstamp) {
5658 tp->rx_opt.tstamp_ok = 1;
5659 tp->tcp_header_len =
5660 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5661 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5662 tcp_store_ts_recent(tp);
5663 } else {
5664 tp->tcp_header_len = sizeof(struct tcphdr);
5665 }
5666
5667 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5668 tcp_initialize_rcv_mss(sk);
5669
5670 /* Remember, tcp_poll() does not lock socket!
5671 * Change state from SYN-SENT only after copied_seq
5672 * is initialized. */
5673 tp->copied_seq = tp->rcv_nxt;
5674
5675 smc_check_reset_syn(tp);
5676
5677 smp_mb();
5678
5679 tcp_finish_connect(sk, skb);
5680
5681 fastopen_fail = (tp->syn_fastopen || tp->syn_data) &&
5682 tcp_rcv_fastopen_synack(sk, skb, &foc);
5683
5684 if (!sock_flag(sk, SOCK_DEAD)) {
5685 sk->sk_state_change(sk);
5686 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5687 }
5688 if (fastopen_fail)
5689 return -1;
5690 if (sk->sk_write_pending ||
5691 icsk->icsk_accept_queue.rskq_defer_accept ||
5692 icsk->icsk_ack.pingpong) {
5693 /* Save one ACK. Data will be ready after
5694 * several ticks, if write_pending is set.
5695 *
5696 * It may be deleted, but with this feature tcpdumps
5697 * look so _wonderfully_ clever, that I was not able
5698 * to stand against the temptation 8) --ANK
5699 */
5700 inet_csk_schedule_ack(sk);
5701 tcp_enter_quickack_mode(sk);
5702 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5703 TCP_DELACK_MAX, TCP_RTO_MAX);
5704
5705discard:
5706 tcp_drop(sk, skb);
5707 return 0;
5708 } else {
5709 tcp_send_ack(sk);
5710 }
5711 return -1;
5712 }
5713
5714 /* No ACK in the segment */
5715
5716 if (th->rst) {
5717 /* rfc793:
5718 * "If the RST bit is set
5719 *
5720 * Otherwise (no ACK) drop the segment and return."
5721 */
5722
5723 goto discard_and_undo;
5724 }
5725
5726 /* PAWS check. */
5727 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5728 tcp_paws_reject(&tp->rx_opt, 0))
5729 goto discard_and_undo;
5730
5731 if (th->syn) {
5732 /* We see SYN without ACK. It is attempt of
5733 * simultaneous connect with crossed SYNs.
5734 * Particularly, it can be connect to self.
5735 */
5736 tcp_set_state(sk, TCP_SYN_RECV);
5737
5738 if (tp->rx_opt.saw_tstamp) {
5739 tp->rx_opt.tstamp_ok = 1;
5740 tcp_store_ts_recent(tp);
5741 tp->tcp_header_len =
5742 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5743 } else {
5744 tp->tcp_header_len = sizeof(struct tcphdr);
5745 }
5746
5747 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5748 tp->copied_seq = tp->rcv_nxt;
5749 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5750
5751 /* RFC1323: The window in SYN & SYN/ACK segments is
5752 * never scaled.
5753 */
5754 tp->snd_wnd = ntohs(th->window);
5755 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5756 tp->max_window = tp->snd_wnd;
5757
5758 tcp_ecn_rcv_syn(tp, th);
5759
5760 tcp_mtup_init(sk);
5761 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5762 tcp_initialize_rcv_mss(sk);
5763
5764 tcp_send_synack(sk);
5765#if 0
5766 /* Note, we could accept data and URG from this segment.
5767 * There are no obstacles to make this (except that we must
5768 * either change tcp_recvmsg() to prevent it from returning data
5769 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5770 *
5771 * However, if we ignore data in ACKless segments sometimes,
5772 * we have no reasons to accept it sometimes.
5773 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5774 * is not flawless. So, discard packet for sanity.
5775 * Uncomment this return to process the data.
5776 */
5777 return -1;
5778#else
5779 goto discard;
5780#endif
5781 }
5782 /* "fifth, if neither of the SYN or RST bits is set then
5783 * drop the segment and return."
5784 */
5785
5786discard_and_undo:
5787 tcp_clear_options(&tp->rx_opt);
5788 tp->rx_opt.mss_clamp = saved_clamp;
5789 goto discard;
5790
5791reset_and_undo:
5792 tcp_clear_options(&tp->rx_opt);
5793 tp->rx_opt.mss_clamp = saved_clamp;
5794 return 1;
5795}
5796
5797/*
5798 * This function implements the receiving procedure of RFC 793 for
5799 * all states except ESTABLISHED and TIME_WAIT.
5800 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5801 * address independent.
5802 */
5803
5804int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
5805{
5806 struct tcp_sock *tp = tcp_sk(sk);
5807 struct inet_connection_sock *icsk = inet_csk(sk);
5808 const struct tcphdr *th = tcp_hdr(skb);
5809 struct request_sock *req;
5810 int queued = 0;
5811 bool acceptable;
5812
5813 switch (sk->sk_state) {
5814 case TCP_CLOSE:
5815 goto discard;
5816
5817 case TCP_LISTEN:
5818 if (th->ack)
5819 return 1;
5820
5821 if (th->rst)
5822 goto discard;
5823
5824 if (th->syn) {
5825 if (th->fin)
5826 goto discard;
5827 /* It is possible that we process SYN packets from backlog,
5828 * so we need to make sure to disable BH right there.
5829 */
5830 local_bh_disable();
5831 acceptable = icsk->icsk_af_ops->conn_request(sk, skb) >= 0;
5832 local_bh_enable();
5833
5834 if (!acceptable)
5835 return 1;
5836 consume_skb(skb);
5837 return 0;
5838 }
5839 goto discard;
5840
5841 case TCP_SYN_SENT:
5842 tp->rx_opt.saw_tstamp = 0;
5843 tcp_mstamp_refresh(tp);
5844 queued = tcp_rcv_synsent_state_process(sk, skb, th);
5845 if (queued >= 0)
5846 return queued;
5847
5848 /* Do step6 onward by hand. */
5849 tcp_urg(sk, skb, th);
5850 __kfree_skb(skb);
5851 tcp_data_snd_check(sk);
5852 return 0;
5853 }
5854
5855 tcp_mstamp_refresh(tp);
5856 tp->rx_opt.saw_tstamp = 0;
5857 req = tp->fastopen_rsk;
5858 if (req) {
5859 bool req_stolen;
5860
5861 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
5862 sk->sk_state != TCP_FIN_WAIT1);
5863
5864 if (!tcp_check_req(sk, skb, req, true, &req_stolen))
5865 goto discard;
5866 }
5867
5868 if (!th->ack && !th->rst && !th->syn)
5869 goto discard;
5870
5871 if (!tcp_validate_incoming(sk, skb, th, 0))
5872 return 0;
5873
5874 /* step 5: check the ACK field */
5875 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
5876 FLAG_UPDATE_TS_RECENT |
5877 FLAG_NO_CHALLENGE_ACK) > 0;
5878
5879 if (!acceptable) {
5880 if (sk->sk_state == TCP_SYN_RECV)
5881 return 1; /* send one RST */
5882 tcp_send_challenge_ack(sk, skb);
5883 goto discard;
5884 }
5885 switch (sk->sk_state) {
5886 case TCP_SYN_RECV:
5887 if (!tp->srtt_us)
5888 tcp_synack_rtt_meas(sk, req);
5889
5890 /* Once we leave TCP_SYN_RECV, we no longer need req
5891 * so release it.
5892 */
5893 if (req) {
5894 inet_csk(sk)->icsk_retransmits = 0;
5895 reqsk_fastopen_remove(sk, req, false);
5896 /* Re-arm the timer because data may have been sent out.
5897 * This is similar to the regular data transmission case
5898 * when new data has just been ack'ed.
5899 *
5900 * (TFO) - we could try to be more aggressive and
5901 * retransmitting any data sooner based on when they
5902 * are sent out.
5903 */
5904 tcp_rearm_rto(sk);
5905 } else {
5906 tcp_init_transfer(sk, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB);
5907 tp->copied_seq = tp->rcv_nxt;
5908 }
5909 smp_mb();
5910 tcp_set_state(sk, TCP_ESTABLISHED);
5911 sk->sk_state_change(sk);
5912
5913 /* Note, that this wakeup is only for marginal crossed SYN case.
5914 * Passively open sockets are not waked up, because
5915 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5916 */
5917 if (sk->sk_socket)
5918 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5919
5920 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5921 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
5922 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5923
5924 if (tp->rx_opt.tstamp_ok)
5925 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5926
5927 if (!inet_csk(sk)->icsk_ca_ops->cong_control)
5928 tcp_update_pacing_rate(sk);
5929
5930 /* Prevent spurious tcp_cwnd_restart() on first data packet */
5931 tp->lsndtime = tcp_jiffies32;
5932
5933 tcp_initialize_rcv_mss(sk);
5934 tcp_fast_path_on(tp);
5935 break;
5936
5937 case TCP_FIN_WAIT1: {
5938 int tmo;
5939
5940 /* If we enter the TCP_FIN_WAIT1 state and we are a
5941 * Fast Open socket and this is the first acceptable
5942 * ACK we have received, this would have acknowledged
5943 * our SYNACK so stop the SYNACK timer.
5944 */
5945 if (req) {
5946 /* We no longer need the request sock. */
5947 reqsk_fastopen_remove(sk, req, false);
5948 tcp_rearm_rto(sk);
5949 }
5950 if (tp->snd_una != tp->write_seq)
5951 break;
5952
5953 tcp_set_state(sk, TCP_FIN_WAIT2);
5954 sk->sk_shutdown |= SEND_SHUTDOWN;
5955
5956 sk_dst_confirm(sk);
5957
5958 if (!sock_flag(sk, SOCK_DEAD)) {
5959 /* Wake up lingering close() */
5960 sk->sk_state_change(sk);
5961 break;
5962 }
5963
5964 if (tp->linger2 < 0) {
5965 tcp_done(sk);
5966 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5967 return 1;
5968 }
5969 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5970 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5971 /* Receive out of order FIN after close() */
5972 if (tp->syn_fastopen && th->fin)
5973 tcp_fastopen_active_disable(sk);
5974 tcp_done(sk);
5975 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5976 return 1;
5977 }
5978
5979 tmo = tcp_fin_time(sk);
5980 if (tmo > TCP_TIMEWAIT_LEN) {
5981 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5982 } else if (th->fin || sock_owned_by_user(sk)) {
5983 /* Bad case. We could lose such FIN otherwise.
5984 * It is not a big problem, but it looks confusing
5985 * and not so rare event. We still can lose it now,
5986 * if it spins in bh_lock_sock(), but it is really
5987 * marginal case.
5988 */
5989 inet_csk_reset_keepalive_timer(sk, tmo);
5990 } else {
5991 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5992 goto discard;
5993 }
5994 break;
5995 }
5996
5997 case TCP_CLOSING:
5998 if (tp->snd_una == tp->write_seq) {
5999 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6000 goto discard;
6001 }
6002 break;
6003
6004 case TCP_LAST_ACK:
6005 if (tp->snd_una == tp->write_seq) {
6006 tcp_update_metrics(sk);
6007 tcp_done(sk);
6008 goto discard;
6009 }
6010 break;
6011 }
6012
6013 /* step 6: check the URG bit */
6014 tcp_urg(sk, skb, th);
6015
6016 /* step 7: process the segment text */
6017 switch (sk->sk_state) {
6018 case TCP_CLOSE_WAIT:
6019 case TCP_CLOSING:
6020 case TCP_LAST_ACK:
6021 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
6022 break;
6023 /* fall through */
6024 case TCP_FIN_WAIT1:
6025 case TCP_FIN_WAIT2:
6026 /* RFC 793 says to queue data in these states,
6027 * RFC 1122 says we MUST send a reset.
6028 * BSD 4.4 also does reset.
6029 */
6030 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6031 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6032 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6033 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6034 tcp_reset(sk);
6035 return 1;
6036 }
6037 }
6038 /* Fall through */
6039 case TCP_ESTABLISHED:
6040 tcp_data_queue(sk, skb);
6041 queued = 1;
6042 break;
6043 }
6044
6045 /* tcp_data could move socket to TIME-WAIT */
6046 if (sk->sk_state != TCP_CLOSE) {
6047 tcp_data_snd_check(sk);
6048 tcp_ack_snd_check(sk);
6049 }
6050
6051 if (!queued) {
6052discard:
6053 tcp_drop(sk, skb);
6054 }
6055 return 0;
6056}
6057EXPORT_SYMBOL(tcp_rcv_state_process);
6058
6059static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
6060{
6061 struct inet_request_sock *ireq = inet_rsk(req);
6062
6063 if (family == AF_INET)
6064 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6065 &ireq->ir_rmt_addr, port);
6066#if IS_ENABLED(CONFIG_IPV6)
6067 else if (family == AF_INET6)
6068 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6069 &ireq->ir_v6_rmt_addr, port);
6070#endif
6071}
6072
6073/* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6074 *
6075 * If we receive a SYN packet with these bits set, it means a
6076 * network is playing bad games with TOS bits. In order to
6077 * avoid possible false congestion notifications, we disable
6078 * TCP ECN negotiation.
6079 *
6080 * Exception: tcp_ca wants ECN. This is required for DCTCP
6081 * congestion control: Linux DCTCP asserts ECT on all packets,
6082 * including SYN, which is most optimal solution; however,
6083 * others, such as FreeBSD do not.
6084 */
6085static void tcp_ecn_create_request(struct request_sock *req,
6086 const struct sk_buff *skb,
6087 const struct sock *listen_sk,
6088 const struct dst_entry *dst)
6089{
6090 const struct tcphdr *th = tcp_hdr(skb);
6091 const struct net *net = sock_net(listen_sk);
6092 bool th_ecn = th->ece && th->cwr;
6093 bool ect, ecn_ok;
6094 u32 ecn_ok_dst;
6095
6096 if (!th_ecn)
6097 return;
6098
6099 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
6100 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK);
6101 ecn_ok = net->ipv4.sysctl_tcp_ecn || ecn_ok_dst;
6102
6103 if ((!ect && ecn_ok) || tcp_ca_needs_ecn(listen_sk) ||
6104 (ecn_ok_dst & DST_FEATURE_ECN_CA) ||
6105 tcp_bpf_ca_needs_ecn((struct sock *)req))
6106 inet_rsk(req)->ecn_ok = 1;
6107}
6108
6109static void tcp_openreq_init(struct request_sock *req,
6110 const struct tcp_options_received *rx_opt,
6111 struct sk_buff *skb, const struct sock *sk)
6112{
6113 struct inet_request_sock *ireq = inet_rsk(req);
6114
6115 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */
6116 req->cookie_ts = 0;
6117 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6118 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
6119 tcp_rsk(req)->snt_synack = tcp_clock_us();
6120 tcp_rsk(req)->last_oow_ack_time = 0;
6121 req->mss = rx_opt->mss_clamp;
6122 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6123 ireq->tstamp_ok = rx_opt->tstamp_ok;
6124 ireq->sack_ok = rx_opt->sack_ok;
6125 ireq->snd_wscale = rx_opt->snd_wscale;
6126 ireq->wscale_ok = rx_opt->wscale_ok;
6127 ireq->acked = 0;
6128 ireq->ecn_ok = 0;
6129 ireq->ir_rmt_port = tcp_hdr(skb)->source;
6130 ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6131 ireq->ir_mark = inet_request_mark(sk, skb);
6132#if IS_ENABLED(CONFIG_SMC)
6133 ireq->smc_ok = rx_opt->smc_ok;
6134#endif
6135}
6136
6137struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
6138 struct sock *sk_listener,
6139 bool attach_listener)
6140{
6141 struct request_sock *req = reqsk_alloc(ops, sk_listener,
6142 attach_listener);
6143
6144 if (req) {
6145 struct inet_request_sock *ireq = inet_rsk(req);
6146
6147 ireq->ireq_opt = NULL;
6148#if IS_ENABLED(CONFIG_IPV6)
6149 ireq->pktopts = NULL;
6150#endif
6151 atomic64_set(&ireq->ir_cookie, 0);
6152 ireq->ireq_state = TCP_NEW_SYN_RECV;
6153 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
6154 ireq->ireq_family = sk_listener->sk_family;
6155 }
6156
6157 return req;
6158}
6159EXPORT_SYMBOL(inet_reqsk_alloc);
6160
6161/*
6162 * Return true if a syncookie should be sent
6163 */
6164static bool tcp_syn_flood_action(const struct sock *sk,
6165 const struct sk_buff *skb,
6166 const char *proto)
6167{
6168 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
6169 const char *msg = "Dropping request";
6170 bool want_cookie = false;
6171 struct net *net = sock_net(sk);
6172
6173#ifdef CONFIG_SYN_COOKIES
6174 if (net->ipv4.sysctl_tcp_syncookies) {
6175 msg = "Sending cookies";
6176 want_cookie = true;
6177 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
6178 } else
6179#endif
6180 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
6181
6182 if (!queue->synflood_warned &&
6183 net->ipv4.sysctl_tcp_syncookies != 2 &&
6184 xchg(&queue->synflood_warned, 1) == 0)
6185 pr_info("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6186 proto, ntohs(tcp_hdr(skb)->dest), msg);
6187
6188 return want_cookie;
6189}
6190
6191static void tcp_reqsk_record_syn(const struct sock *sk,
6192 struct request_sock *req,
6193 const struct sk_buff *skb)
6194{
6195 if (tcp_sk(sk)->save_syn) {
6196 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
6197 u32 *copy;
6198
6199 copy = kmalloc(len + sizeof(u32), GFP_ATOMIC);
6200 if (copy) {
6201 copy[0] = len;
6202 memcpy(©[1], skb_network_header(skb), len);
6203 req->saved_syn = copy;
6204 }
6205 }
6206}
6207
6208int tcp_conn_request(struct request_sock_ops *rsk_ops,
6209 const struct tcp_request_sock_ops *af_ops,
6210 struct sock *sk, struct sk_buff *skb)
6211{
6212 struct tcp_fastopen_cookie foc = { .len = -1 };
6213 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
6214 struct tcp_options_received tmp_opt;
6215 struct tcp_sock *tp = tcp_sk(sk);
6216 struct net *net = sock_net(sk);
6217 struct sock *fastopen_sk = NULL;
6218 struct request_sock *req;
6219 bool want_cookie = false;
6220 struct dst_entry *dst;
6221 struct flowi fl;
6222
6223 /* TW buckets are converted to open requests without
6224 * limitations, they conserve resources and peer is
6225 * evidently real one.
6226 */
6227 if ((net->ipv4.sysctl_tcp_syncookies == 2 ||
6228 inet_csk_reqsk_queue_is_full(sk)) && !isn) {
6229 want_cookie = tcp_syn_flood_action(sk, skb, rsk_ops->slab_name);
6230 if (!want_cookie)
6231 goto drop;
6232 }
6233
6234 if (sk_acceptq_is_full(sk)) {
6235 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6236 goto drop;
6237 }
6238
6239 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie);
6240 if (!req)
6241 goto drop;
6242
6243 tcp_rsk(req)->af_specific = af_ops;
6244 tcp_rsk(req)->ts_off = 0;
6245
6246 tcp_clear_options(&tmp_opt);
6247 tmp_opt.mss_clamp = af_ops->mss_clamp;
6248 tmp_opt.user_mss = tp->rx_opt.user_mss;
6249 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0,
6250 want_cookie ? NULL : &foc);
6251
6252 if (want_cookie && !tmp_opt.saw_tstamp)
6253 tcp_clear_options(&tmp_opt);
6254
6255 if (IS_ENABLED(CONFIG_SMC) && want_cookie)
6256 tmp_opt.smc_ok = 0;
6257
6258 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
6259 tcp_openreq_init(req, &tmp_opt, skb, sk);
6260 inet_rsk(req)->no_srccheck = inet_sk(sk)->transparent;
6261
6262 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6263 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb);
6264
6265 af_ops->init_req(req, sk, skb);
6266
6267 if (security_inet_conn_request(sk, skb, req))
6268 goto drop_and_free;
6269
6270 if (tmp_opt.tstamp_ok)
6271 tcp_rsk(req)->ts_off = af_ops->init_ts_off(net, skb);
6272
6273 dst = af_ops->route_req(sk, &fl, req);
6274 if (!dst)
6275 goto drop_and_free;
6276
6277 if (!want_cookie && !isn) {
6278 /* Kill the following clause, if you dislike this way. */
6279 if (!net->ipv4.sysctl_tcp_syncookies &&
6280 (net->ipv4.sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
6281 (net->ipv4.sysctl_max_syn_backlog >> 2)) &&
6282 !tcp_peer_is_proven(req, dst)) {
6283 /* Without syncookies last quarter of
6284 * backlog is filled with destinations,
6285 * proven to be alive.
6286 * It means that we continue to communicate
6287 * to destinations, already remembered
6288 * to the moment of synflood.
6289 */
6290 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
6291 rsk_ops->family);
6292 goto drop_and_release;
6293 }
6294
6295 isn = af_ops->init_seq(skb);
6296 }
6297
6298 tcp_ecn_create_request(req, skb, sk, dst);
6299
6300 if (want_cookie) {
6301 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
6302 req->cookie_ts = tmp_opt.tstamp_ok;
6303 if (!tmp_opt.tstamp_ok)
6304 inet_rsk(req)->ecn_ok = 0;
6305 }
6306
6307 tcp_rsk(req)->snt_isn = isn;
6308 tcp_rsk(req)->txhash = net_tx_rndhash();
6309 tcp_openreq_init_rwin(req, sk, dst);
6310 if (!want_cookie) {
6311 tcp_reqsk_record_syn(sk, req, skb);
6312 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst);
6313 }
6314 if (fastopen_sk) {
6315 af_ops->send_synack(fastopen_sk, dst, &fl, req,
6316 &foc, TCP_SYNACK_FASTOPEN);
6317 /* Add the child socket directly into the accept queue */
6318 inet_csk_reqsk_queue_add(sk, req, fastopen_sk);
6319 sk->sk_data_ready(sk);
6320 bh_unlock_sock(fastopen_sk);
6321 sock_put(fastopen_sk);
6322 } else {
6323 tcp_rsk(req)->tfo_listener = false;
6324 if (!want_cookie)
6325 inet_csk_reqsk_queue_hash_add(sk, req,
6326 tcp_timeout_init((struct sock *)req));
6327 af_ops->send_synack(sk, dst, &fl, req, &foc,
6328 !want_cookie ? TCP_SYNACK_NORMAL :
6329 TCP_SYNACK_COOKIE);
6330 if (want_cookie) {
6331 reqsk_free(req);
6332 return 0;
6333 }
6334 }
6335 reqsk_put(req);
6336 return 0;
6337
6338drop_and_release:
6339 dst_release(dst);
6340drop_and_free:
6341 reqsk_free(req);
6342drop:
6343 tcp_listendrop(sk);
6344 return 0;
6345}
6346EXPORT_SYMBOL(tcp_conn_request);