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/*
2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
5 *
6 * Implementation of the Transmission Control Protocol(TCP).
7 *
8 * Authors: Ross Biro
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
19 */
20
21/*
22 * Changes:
23 * Pedro Roque : Fast Retransmit/Recovery.
24 * Two receive queues.
25 * Retransmit queue handled by TCP.
26 * Better retransmit timer handling.
27 * New congestion avoidance.
28 * Header prediction.
29 * Variable renaming.
30 *
31 * Eric : Fast Retransmit.
32 * Randy Scott : MSS option defines.
33 * Eric Schenk : Fixes to slow start algorithm.
34 * Eric Schenk : Yet another double ACK bug.
35 * Eric Schenk : Delayed ACK bug fixes.
36 * Eric Schenk : Floyd style fast retrans war avoidance.
37 * David S. Miller : Don't allow zero congestion window.
38 * Eric Schenk : Fix retransmitter so that it sends
39 * next packet on ack of previous packet.
40 * Andi Kleen : Moved open_request checking here
41 * and process RSTs for open_requests.
42 * Andi Kleen : Better prune_queue, and other fixes.
43 * Andrey Savochkin: Fix RTT measurements in the presence of
44 * timestamps.
45 * Andrey Savochkin: Check sequence numbers correctly when
46 * removing SACKs due to in sequence incoming
47 * data segments.
48 * Andi Kleen: Make sure we never ack data there is not
49 * enough room for. Also make this condition
50 * a fatal error if it might still happen.
51 * Andi Kleen: Add tcp_measure_rcv_mss to make
52 * connections with MSS<min(MTU,ann. MSS)
53 * work without delayed acks.
54 * Andi Kleen: Process packets with PSH set in the
55 * fast path.
56 * J Hadi Salim: ECN support
57 * Andrei Gurtov,
58 * Pasi Sarolahti,
59 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
60 * engine. Lots of bugs are found.
61 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
62 */
63
64#include <linux/mm.h>
65#include <linux/slab.h>
66#include <linux/module.h>
67#include <linux/sysctl.h>
68#include <linux/kernel.h>
69#include <net/dst.h>
70#include <net/tcp.h>
71#include <net/inet_common.h>
72#include <linux/ipsec.h>
73#include <asm/unaligned.h>
74#include <net/netdma.h>
75
76int sysctl_tcp_timestamps __read_mostly = 1;
77int sysctl_tcp_window_scaling __read_mostly = 1;
78int sysctl_tcp_sack __read_mostly = 1;
79int sysctl_tcp_fack __read_mostly = 1;
80int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
81EXPORT_SYMBOL(sysctl_tcp_reordering);
82int sysctl_tcp_ecn __read_mostly = 2;
83EXPORT_SYMBOL(sysctl_tcp_ecn);
84int sysctl_tcp_dsack __read_mostly = 1;
85int sysctl_tcp_app_win __read_mostly = 31;
86int sysctl_tcp_adv_win_scale __read_mostly = 2;
87EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
88
89int sysctl_tcp_stdurg __read_mostly;
90int sysctl_tcp_rfc1337 __read_mostly;
91int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
92int sysctl_tcp_frto __read_mostly = 2;
93int sysctl_tcp_frto_response __read_mostly;
94int sysctl_tcp_nometrics_save __read_mostly;
95
96int sysctl_tcp_thin_dupack __read_mostly;
97
98int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
99int sysctl_tcp_abc __read_mostly;
100
101#define FLAG_DATA 0x01 /* Incoming frame contained data. */
102#define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
103#define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
104#define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
105#define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
106#define FLAG_DATA_SACKED 0x20 /* New SACK. */
107#define FLAG_ECE 0x40 /* ECE in this ACK */
108#define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */
109#define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
110#define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */
111#define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
112#define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
113#define FLAG_NONHEAD_RETRANS_ACKED 0x1000 /* Non-head rexmitted data was ACKed */
114#define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
115
116#define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
117#define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
118#define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
119#define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
120#define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED)
121
122#define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
123#define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
124
125/* Adapt the MSS value used to make delayed ack decision to the
126 * real world.
127 */
128static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
129{
130 struct inet_connection_sock *icsk = inet_csk(sk);
131 const unsigned int lss = icsk->icsk_ack.last_seg_size;
132 unsigned int len;
133
134 icsk->icsk_ack.last_seg_size = 0;
135
136 /* skb->len may jitter because of SACKs, even if peer
137 * sends good full-sized frames.
138 */
139 len = skb_shinfo(skb)->gso_size ? : skb->len;
140 if (len >= icsk->icsk_ack.rcv_mss) {
141 icsk->icsk_ack.rcv_mss = len;
142 } else {
143 /* Otherwise, we make more careful check taking into account,
144 * that SACKs block is variable.
145 *
146 * "len" is invariant segment length, including TCP header.
147 */
148 len += skb->data - skb_transport_header(skb);
149 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
150 /* If PSH is not set, packet should be
151 * full sized, provided peer TCP is not badly broken.
152 * This observation (if it is correct 8)) allows
153 * to handle super-low mtu links fairly.
154 */
155 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
156 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
157 /* Subtract also invariant (if peer is RFC compliant),
158 * tcp header plus fixed timestamp option length.
159 * Resulting "len" is MSS free of SACK jitter.
160 */
161 len -= tcp_sk(sk)->tcp_header_len;
162 icsk->icsk_ack.last_seg_size = len;
163 if (len == lss) {
164 icsk->icsk_ack.rcv_mss = len;
165 return;
166 }
167 }
168 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
169 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
170 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
171 }
172}
173
174static void tcp_incr_quickack(struct sock *sk)
175{
176 struct inet_connection_sock *icsk = inet_csk(sk);
177 unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
178
179 if (quickacks == 0)
180 quickacks = 2;
181 if (quickacks > icsk->icsk_ack.quick)
182 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
183}
184
185static void tcp_enter_quickack_mode(struct sock *sk)
186{
187 struct inet_connection_sock *icsk = inet_csk(sk);
188 tcp_incr_quickack(sk);
189 icsk->icsk_ack.pingpong = 0;
190 icsk->icsk_ack.ato = TCP_ATO_MIN;
191}
192
193/* Send ACKs quickly, if "quick" count is not exhausted
194 * and the session is not interactive.
195 */
196
197static inline int tcp_in_quickack_mode(const struct sock *sk)
198{
199 const struct inet_connection_sock *icsk = inet_csk(sk);
200 return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
201}
202
203static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp)
204{
205 if (tp->ecn_flags & TCP_ECN_OK)
206 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
207}
208
209static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, struct sk_buff *skb)
210{
211 if (tcp_hdr(skb)->cwr)
212 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
213}
214
215static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp)
216{
217 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
218}
219
220static inline void TCP_ECN_check_ce(struct tcp_sock *tp, struct sk_buff *skb)
221{
222 if (tp->ecn_flags & TCP_ECN_OK) {
223 if (INET_ECN_is_ce(TCP_SKB_CB(skb)->flags))
224 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
225 /* Funny extension: if ECT is not set on a segment,
226 * it is surely retransmit. It is not in ECN RFC,
227 * but Linux follows this rule. */
228 else if (INET_ECN_is_not_ect((TCP_SKB_CB(skb)->flags)))
229 tcp_enter_quickack_mode((struct sock *)tp);
230 }
231}
232
233static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, struct tcphdr *th)
234{
235 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
236 tp->ecn_flags &= ~TCP_ECN_OK;
237}
238
239static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, struct tcphdr *th)
240{
241 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
242 tp->ecn_flags &= ~TCP_ECN_OK;
243}
244
245static inline int TCP_ECN_rcv_ecn_echo(struct tcp_sock *tp, struct tcphdr *th)
246{
247 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
248 return 1;
249 return 0;
250}
251
252/* Buffer size and advertised window tuning.
253 *
254 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
255 */
256
257static void tcp_fixup_sndbuf(struct sock *sk)
258{
259 int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 +
260 sizeof(struct sk_buff);
261
262 if (sk->sk_sndbuf < 3 * sndmem) {
263 sk->sk_sndbuf = 3 * sndmem;
264 if (sk->sk_sndbuf > sysctl_tcp_wmem[2])
265 sk->sk_sndbuf = sysctl_tcp_wmem[2];
266 }
267}
268
269/* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
270 *
271 * All tcp_full_space() is split to two parts: "network" buffer, allocated
272 * forward and advertised in receiver window (tp->rcv_wnd) and
273 * "application buffer", required to isolate scheduling/application
274 * latencies from network.
275 * window_clamp is maximal advertised window. It can be less than
276 * tcp_full_space(), in this case tcp_full_space() - window_clamp
277 * is reserved for "application" buffer. The less window_clamp is
278 * the smoother our behaviour from viewpoint of network, but the lower
279 * throughput and the higher sensitivity of the connection to losses. 8)
280 *
281 * rcv_ssthresh is more strict window_clamp used at "slow start"
282 * phase to predict further behaviour of this connection.
283 * It is used for two goals:
284 * - to enforce header prediction at sender, even when application
285 * requires some significant "application buffer". It is check #1.
286 * - to prevent pruning of receive queue because of misprediction
287 * of receiver window. Check #2.
288 *
289 * The scheme does not work when sender sends good segments opening
290 * window and then starts to feed us spaghetti. But it should work
291 * in common situations. Otherwise, we have to rely on queue collapsing.
292 */
293
294/* Slow part of check#2. */
295static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
296{
297 struct tcp_sock *tp = tcp_sk(sk);
298 /* Optimize this! */
299 int truesize = tcp_win_from_space(skb->truesize) >> 1;
300 int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
301
302 while (tp->rcv_ssthresh <= window) {
303 if (truesize <= skb->len)
304 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
305
306 truesize >>= 1;
307 window >>= 1;
308 }
309 return 0;
310}
311
312static void tcp_grow_window(struct sock *sk, struct sk_buff *skb)
313{
314 struct tcp_sock *tp = tcp_sk(sk);
315
316 /* Check #1 */
317 if (tp->rcv_ssthresh < tp->window_clamp &&
318 (int)tp->rcv_ssthresh < tcp_space(sk) &&
319 !tcp_memory_pressure) {
320 int incr;
321
322 /* Check #2. Increase window, if skb with such overhead
323 * will fit to rcvbuf in future.
324 */
325 if (tcp_win_from_space(skb->truesize) <= skb->len)
326 incr = 2 * tp->advmss;
327 else
328 incr = __tcp_grow_window(sk, skb);
329
330 if (incr) {
331 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
332 tp->window_clamp);
333 inet_csk(sk)->icsk_ack.quick |= 1;
334 }
335 }
336}
337
338/* 3. Tuning rcvbuf, when connection enters established state. */
339
340static void tcp_fixup_rcvbuf(struct sock *sk)
341{
342 struct tcp_sock *tp = tcp_sk(sk);
343 int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
344
345 /* Try to select rcvbuf so that 4 mss-sized segments
346 * will fit to window and corresponding skbs will fit to our rcvbuf.
347 * (was 3; 4 is minimum to allow fast retransmit to work.)
348 */
349 while (tcp_win_from_space(rcvmem) < tp->advmss)
350 rcvmem += 128;
351 if (sk->sk_rcvbuf < 4 * rcvmem)
352 sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]);
353}
354
355/* 4. Try to fixup all. It is made immediately after connection enters
356 * established state.
357 */
358static void tcp_init_buffer_space(struct sock *sk)
359{
360 struct tcp_sock *tp = tcp_sk(sk);
361 int maxwin;
362
363 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
364 tcp_fixup_rcvbuf(sk);
365 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
366 tcp_fixup_sndbuf(sk);
367
368 tp->rcvq_space.space = tp->rcv_wnd;
369
370 maxwin = tcp_full_space(sk);
371
372 if (tp->window_clamp >= maxwin) {
373 tp->window_clamp = maxwin;
374
375 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
376 tp->window_clamp = max(maxwin -
377 (maxwin >> sysctl_tcp_app_win),
378 4 * tp->advmss);
379 }
380
381 /* Force reservation of one segment. */
382 if (sysctl_tcp_app_win &&
383 tp->window_clamp > 2 * tp->advmss &&
384 tp->window_clamp + tp->advmss > maxwin)
385 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
386
387 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
388 tp->snd_cwnd_stamp = tcp_time_stamp;
389}
390
391/* 5. Recalculate window clamp after socket hit its memory bounds. */
392static void tcp_clamp_window(struct sock *sk)
393{
394 struct tcp_sock *tp = tcp_sk(sk);
395 struct inet_connection_sock *icsk = inet_csk(sk);
396
397 icsk->icsk_ack.quick = 0;
398
399 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
400 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
401 !tcp_memory_pressure &&
402 atomic_long_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) {
403 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
404 sysctl_tcp_rmem[2]);
405 }
406 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
407 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
408}
409
410/* Initialize RCV_MSS value.
411 * RCV_MSS is an our guess about MSS used by the peer.
412 * We haven't any direct information about the MSS.
413 * It's better to underestimate the RCV_MSS rather than overestimate.
414 * Overestimations make us ACKing less frequently than needed.
415 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
416 */
417void tcp_initialize_rcv_mss(struct sock *sk)
418{
419 struct tcp_sock *tp = tcp_sk(sk);
420 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
421
422 hint = min(hint, tp->rcv_wnd / 2);
423 hint = min(hint, TCP_MSS_DEFAULT);
424 hint = max(hint, TCP_MIN_MSS);
425
426 inet_csk(sk)->icsk_ack.rcv_mss = hint;
427}
428EXPORT_SYMBOL(tcp_initialize_rcv_mss);
429
430/* Receiver "autotuning" code.
431 *
432 * The algorithm for RTT estimation w/o timestamps is based on
433 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
434 * <http://public.lanl.gov/radiant/pubs.html#DRS>
435 *
436 * More detail on this code can be found at
437 * <http://staff.psc.edu/jheffner/>,
438 * though this reference is out of date. A new paper
439 * is pending.
440 */
441static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
442{
443 u32 new_sample = tp->rcv_rtt_est.rtt;
444 long m = sample;
445
446 if (m == 0)
447 m = 1;
448
449 if (new_sample != 0) {
450 /* If we sample in larger samples in the non-timestamp
451 * case, we could grossly overestimate the RTT especially
452 * with chatty applications or bulk transfer apps which
453 * are stalled on filesystem I/O.
454 *
455 * Also, since we are only going for a minimum in the
456 * non-timestamp case, we do not smooth things out
457 * else with timestamps disabled convergence takes too
458 * long.
459 */
460 if (!win_dep) {
461 m -= (new_sample >> 3);
462 new_sample += m;
463 } else if (m < new_sample)
464 new_sample = m << 3;
465 } else {
466 /* No previous measure. */
467 new_sample = m << 3;
468 }
469
470 if (tp->rcv_rtt_est.rtt != new_sample)
471 tp->rcv_rtt_est.rtt = new_sample;
472}
473
474static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
475{
476 if (tp->rcv_rtt_est.time == 0)
477 goto new_measure;
478 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
479 return;
480 tcp_rcv_rtt_update(tp, jiffies - tp->rcv_rtt_est.time, 1);
481
482new_measure:
483 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
484 tp->rcv_rtt_est.time = tcp_time_stamp;
485}
486
487static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
488 const struct sk_buff *skb)
489{
490 struct tcp_sock *tp = tcp_sk(sk);
491 if (tp->rx_opt.rcv_tsecr &&
492 (TCP_SKB_CB(skb)->end_seq -
493 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
494 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
495}
496
497/*
498 * This function should be called every time data is copied to user space.
499 * It calculates the appropriate TCP receive buffer space.
500 */
501void tcp_rcv_space_adjust(struct sock *sk)
502{
503 struct tcp_sock *tp = tcp_sk(sk);
504 int time;
505 int space;
506
507 if (tp->rcvq_space.time == 0)
508 goto new_measure;
509
510 time = tcp_time_stamp - tp->rcvq_space.time;
511 if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
512 return;
513
514 space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
515
516 space = max(tp->rcvq_space.space, space);
517
518 if (tp->rcvq_space.space != space) {
519 int rcvmem;
520
521 tp->rcvq_space.space = space;
522
523 if (sysctl_tcp_moderate_rcvbuf &&
524 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
525 int new_clamp = space;
526
527 /* Receive space grows, normalize in order to
528 * take into account packet headers and sk_buff
529 * structure overhead.
530 */
531 space /= tp->advmss;
532 if (!space)
533 space = 1;
534 rcvmem = (tp->advmss + MAX_TCP_HEADER +
535 16 + sizeof(struct sk_buff));
536 while (tcp_win_from_space(rcvmem) < tp->advmss)
537 rcvmem += 128;
538 space *= rcvmem;
539 space = min(space, sysctl_tcp_rmem[2]);
540 if (space > sk->sk_rcvbuf) {
541 sk->sk_rcvbuf = space;
542
543 /* Make the window clamp follow along. */
544 tp->window_clamp = new_clamp;
545 }
546 }
547 }
548
549new_measure:
550 tp->rcvq_space.seq = tp->copied_seq;
551 tp->rcvq_space.time = tcp_time_stamp;
552}
553
554/* There is something which you must keep in mind when you analyze the
555 * behavior of the tp->ato delayed ack timeout interval. When a
556 * connection starts up, we want to ack as quickly as possible. The
557 * problem is that "good" TCP's do slow start at the beginning of data
558 * transmission. The means that until we send the first few ACK's the
559 * sender will sit on his end and only queue most of his data, because
560 * he can only send snd_cwnd unacked packets at any given time. For
561 * each ACK we send, he increments snd_cwnd and transmits more of his
562 * queue. -DaveM
563 */
564static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
565{
566 struct tcp_sock *tp = tcp_sk(sk);
567 struct inet_connection_sock *icsk = inet_csk(sk);
568 u32 now;
569
570 inet_csk_schedule_ack(sk);
571
572 tcp_measure_rcv_mss(sk, skb);
573
574 tcp_rcv_rtt_measure(tp);
575
576 now = tcp_time_stamp;
577
578 if (!icsk->icsk_ack.ato) {
579 /* The _first_ data packet received, initialize
580 * delayed ACK engine.
581 */
582 tcp_incr_quickack(sk);
583 icsk->icsk_ack.ato = TCP_ATO_MIN;
584 } else {
585 int m = now - icsk->icsk_ack.lrcvtime;
586
587 if (m <= TCP_ATO_MIN / 2) {
588 /* The fastest case is the first. */
589 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
590 } else if (m < icsk->icsk_ack.ato) {
591 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
592 if (icsk->icsk_ack.ato > icsk->icsk_rto)
593 icsk->icsk_ack.ato = icsk->icsk_rto;
594 } else if (m > icsk->icsk_rto) {
595 /* Too long gap. Apparently sender failed to
596 * restart window, so that we send ACKs quickly.
597 */
598 tcp_incr_quickack(sk);
599 sk_mem_reclaim(sk);
600 }
601 }
602 icsk->icsk_ack.lrcvtime = now;
603
604 TCP_ECN_check_ce(tp, skb);
605
606 if (skb->len >= 128)
607 tcp_grow_window(sk, skb);
608}
609
610/* Called to compute a smoothed rtt estimate. The data fed to this
611 * routine either comes from timestamps, or from segments that were
612 * known _not_ to have been retransmitted [see Karn/Partridge
613 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
614 * piece by Van Jacobson.
615 * NOTE: the next three routines used to be one big routine.
616 * To save cycles in the RFC 1323 implementation it was better to break
617 * it up into three procedures. -- erics
618 */
619static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
620{
621 struct tcp_sock *tp = tcp_sk(sk);
622 long m = mrtt; /* RTT */
623
624 /* The following amusing code comes from Jacobson's
625 * article in SIGCOMM '88. Note that rtt and mdev
626 * are scaled versions of rtt and mean deviation.
627 * This is designed to be as fast as possible
628 * m stands for "measurement".
629 *
630 * On a 1990 paper the rto value is changed to:
631 * RTO = rtt + 4 * mdev
632 *
633 * Funny. This algorithm seems to be very broken.
634 * These formulae increase RTO, when it should be decreased, increase
635 * too slowly, when it should be increased quickly, decrease too quickly
636 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
637 * does not matter how to _calculate_ it. Seems, it was trap
638 * that VJ failed to avoid. 8)
639 */
640 if (m == 0)
641 m = 1;
642 if (tp->srtt != 0) {
643 m -= (tp->srtt >> 3); /* m is now error in rtt est */
644 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */
645 if (m < 0) {
646 m = -m; /* m is now abs(error) */
647 m -= (tp->mdev >> 2); /* similar update on mdev */
648 /* This is similar to one of Eifel findings.
649 * Eifel blocks mdev updates when rtt decreases.
650 * This solution is a bit different: we use finer gain
651 * for mdev in this case (alpha*beta).
652 * Like Eifel it also prevents growth of rto,
653 * but also it limits too fast rto decreases,
654 * happening in pure Eifel.
655 */
656 if (m > 0)
657 m >>= 3;
658 } else {
659 m -= (tp->mdev >> 2); /* similar update on mdev */
660 }
661 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */
662 if (tp->mdev > tp->mdev_max) {
663 tp->mdev_max = tp->mdev;
664 if (tp->mdev_max > tp->rttvar)
665 tp->rttvar = tp->mdev_max;
666 }
667 if (after(tp->snd_una, tp->rtt_seq)) {
668 if (tp->mdev_max < tp->rttvar)
669 tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2;
670 tp->rtt_seq = tp->snd_nxt;
671 tp->mdev_max = tcp_rto_min(sk);
672 }
673 } else {
674 /* no previous measure. */
675 tp->srtt = m << 3; /* take the measured time to be rtt */
676 tp->mdev = m << 1; /* make sure rto = 3*rtt */
677 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
678 tp->rtt_seq = tp->snd_nxt;
679 }
680}
681
682/* Calculate rto without backoff. This is the second half of Van Jacobson's
683 * routine referred to above.
684 */
685static inline void tcp_set_rto(struct sock *sk)
686{
687 const struct tcp_sock *tp = tcp_sk(sk);
688 /* Old crap is replaced with new one. 8)
689 *
690 * More seriously:
691 * 1. If rtt variance happened to be less 50msec, it is hallucination.
692 * It cannot be less due to utterly erratic ACK generation made
693 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
694 * to do with delayed acks, because at cwnd>2 true delack timeout
695 * is invisible. Actually, Linux-2.4 also generates erratic
696 * ACKs in some circumstances.
697 */
698 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
699
700 /* 2. Fixups made earlier cannot be right.
701 * If we do not estimate RTO correctly without them,
702 * all the algo is pure shit and should be replaced
703 * with correct one. It is exactly, which we pretend to do.
704 */
705
706 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
707 * guarantees that rto is higher.
708 */
709 tcp_bound_rto(sk);
710}
711
712/* Save metrics learned by this TCP session.
713 This function is called only, when TCP finishes successfully
714 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
715 */
716void tcp_update_metrics(struct sock *sk)
717{
718 struct tcp_sock *tp = tcp_sk(sk);
719 struct dst_entry *dst = __sk_dst_get(sk);
720
721 if (sysctl_tcp_nometrics_save)
722 return;
723
724 dst_confirm(dst);
725
726 if (dst && (dst->flags & DST_HOST)) {
727 const struct inet_connection_sock *icsk = inet_csk(sk);
728 int m;
729 unsigned long rtt;
730
731 if (icsk->icsk_backoff || !tp->srtt) {
732 /* This session failed to estimate rtt. Why?
733 * Probably, no packets returned in time.
734 * Reset our results.
735 */
736 if (!(dst_metric_locked(dst, RTAX_RTT)))
737 dst_metric_set(dst, RTAX_RTT, 0);
738 return;
739 }
740
741 rtt = dst_metric_rtt(dst, RTAX_RTT);
742 m = rtt - tp->srtt;
743
744 /* If newly calculated rtt larger than stored one,
745 * store new one. Otherwise, use EWMA. Remember,
746 * rtt overestimation is always better than underestimation.
747 */
748 if (!(dst_metric_locked(dst, RTAX_RTT))) {
749 if (m <= 0)
750 set_dst_metric_rtt(dst, RTAX_RTT, tp->srtt);
751 else
752 set_dst_metric_rtt(dst, RTAX_RTT, rtt - (m >> 3));
753 }
754
755 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
756 unsigned long var;
757 if (m < 0)
758 m = -m;
759
760 /* Scale deviation to rttvar fixed point */
761 m >>= 1;
762 if (m < tp->mdev)
763 m = tp->mdev;
764
765 var = dst_metric_rtt(dst, RTAX_RTTVAR);
766 if (m >= var)
767 var = m;
768 else
769 var -= (var - m) >> 2;
770
771 set_dst_metric_rtt(dst, RTAX_RTTVAR, var);
772 }
773
774 if (tcp_in_initial_slowstart(tp)) {
775 /* Slow start still did not finish. */
776 if (dst_metric(dst, RTAX_SSTHRESH) &&
777 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
778 (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
779 dst_metric_set(dst, RTAX_SSTHRESH, tp->snd_cwnd >> 1);
780 if (!dst_metric_locked(dst, RTAX_CWND) &&
781 tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
782 dst_metric_set(dst, RTAX_CWND, tp->snd_cwnd);
783 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
784 icsk->icsk_ca_state == TCP_CA_Open) {
785 /* Cong. avoidance phase, cwnd is reliable. */
786 if (!dst_metric_locked(dst, RTAX_SSTHRESH))
787 dst_metric_set(dst, RTAX_SSTHRESH,
788 max(tp->snd_cwnd >> 1, tp->snd_ssthresh));
789 if (!dst_metric_locked(dst, RTAX_CWND))
790 dst_metric_set(dst, RTAX_CWND,
791 (dst_metric(dst, RTAX_CWND) +
792 tp->snd_cwnd) >> 1);
793 } else {
794 /* Else slow start did not finish, cwnd is non-sense,
795 ssthresh may be also invalid.
796 */
797 if (!dst_metric_locked(dst, RTAX_CWND))
798 dst_metric_set(dst, RTAX_CWND,
799 (dst_metric(dst, RTAX_CWND) +
800 tp->snd_ssthresh) >> 1);
801 if (dst_metric(dst, RTAX_SSTHRESH) &&
802 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
803 tp->snd_ssthresh > dst_metric(dst, RTAX_SSTHRESH))
804 dst_metric_set(dst, RTAX_SSTHRESH, tp->snd_ssthresh);
805 }
806
807 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
808 if (dst_metric(dst, RTAX_REORDERING) < tp->reordering &&
809 tp->reordering != sysctl_tcp_reordering)
810 dst_metric_set(dst, RTAX_REORDERING, tp->reordering);
811 }
812 }
813}
814
815__u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst)
816{
817 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
818
819 if (!cwnd)
820 cwnd = TCP_INIT_CWND;
821 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
822}
823
824/* Set slow start threshold and cwnd not falling to slow start */
825void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
826{
827 struct tcp_sock *tp = tcp_sk(sk);
828 const struct inet_connection_sock *icsk = inet_csk(sk);
829
830 tp->prior_ssthresh = 0;
831 tp->bytes_acked = 0;
832 if (icsk->icsk_ca_state < TCP_CA_CWR) {
833 tp->undo_marker = 0;
834 if (set_ssthresh)
835 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
836 tp->snd_cwnd = min(tp->snd_cwnd,
837 tcp_packets_in_flight(tp) + 1U);
838 tp->snd_cwnd_cnt = 0;
839 tp->high_seq = tp->snd_nxt;
840 tp->snd_cwnd_stamp = tcp_time_stamp;
841 TCP_ECN_queue_cwr(tp);
842
843 tcp_set_ca_state(sk, TCP_CA_CWR);
844 }
845}
846
847/*
848 * Packet counting of FACK is based on in-order assumptions, therefore TCP
849 * disables it when reordering is detected
850 */
851static void tcp_disable_fack(struct tcp_sock *tp)
852{
853 /* RFC3517 uses different metric in lost marker => reset on change */
854 if (tcp_is_fack(tp))
855 tp->lost_skb_hint = NULL;
856 tp->rx_opt.sack_ok &= ~2;
857}
858
859/* Take a notice that peer is sending D-SACKs */
860static void tcp_dsack_seen(struct tcp_sock *tp)
861{
862 tp->rx_opt.sack_ok |= 4;
863}
864
865/* Initialize metrics on socket. */
866
867static void tcp_init_metrics(struct sock *sk)
868{
869 struct tcp_sock *tp = tcp_sk(sk);
870 struct dst_entry *dst = __sk_dst_get(sk);
871
872 if (dst == NULL)
873 goto reset;
874
875 dst_confirm(dst);
876
877 if (dst_metric_locked(dst, RTAX_CWND))
878 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
879 if (dst_metric(dst, RTAX_SSTHRESH)) {
880 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
881 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
882 tp->snd_ssthresh = tp->snd_cwnd_clamp;
883 } else {
884 /* ssthresh may have been reduced unnecessarily during.
885 * 3WHS. Restore it back to its initial default.
886 */
887 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
888 }
889 if (dst_metric(dst, RTAX_REORDERING) &&
890 tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
891 tcp_disable_fack(tp);
892 tp->reordering = dst_metric(dst, RTAX_REORDERING);
893 }
894
895 if (dst_metric(dst, RTAX_RTT) == 0 || tp->srtt == 0)
896 goto reset;
897
898 /* Initial rtt is determined from SYN,SYN-ACK.
899 * The segment is small and rtt may appear much
900 * less than real one. Use per-dst memory
901 * to make it more realistic.
902 *
903 * A bit of theory. RTT is time passed after "normal" sized packet
904 * is sent until it is ACKed. In normal circumstances sending small
905 * packets force peer to delay ACKs and calculation is correct too.
906 * The algorithm is adaptive and, provided we follow specs, it
907 * NEVER underestimate RTT. BUT! If peer tries to make some clever
908 * tricks sort of "quick acks" for time long enough to decrease RTT
909 * to low value, and then abruptly stops to do it and starts to delay
910 * ACKs, wait for troubles.
911 */
912 if (dst_metric_rtt(dst, RTAX_RTT) > tp->srtt) {
913 tp->srtt = dst_metric_rtt(dst, RTAX_RTT);
914 tp->rtt_seq = tp->snd_nxt;
915 }
916 if (dst_metric_rtt(dst, RTAX_RTTVAR) > tp->mdev) {
917 tp->mdev = dst_metric_rtt(dst, RTAX_RTTVAR);
918 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
919 }
920 tcp_set_rto(sk);
921reset:
922 if (tp->srtt == 0) {
923 /* RFC2988bis: We've failed to get a valid RTT sample from
924 * 3WHS. This is most likely due to retransmission,
925 * including spurious one. Reset the RTO back to 3secs
926 * from the more aggressive 1sec to avoid more spurious
927 * retransmission.
928 */
929 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_FALLBACK;
930 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK;
931 }
932 /* Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
933 * retransmitted. In light of RFC2988bis' more aggressive 1sec
934 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
935 * retransmission has occurred.
936 */
937 if (tp->total_retrans > 1)
938 tp->snd_cwnd = 1;
939 else
940 tp->snd_cwnd = tcp_init_cwnd(tp, dst);
941 tp->snd_cwnd_stamp = tcp_time_stamp;
942}
943
944static void tcp_update_reordering(struct sock *sk, const int metric,
945 const int ts)
946{
947 struct tcp_sock *tp = tcp_sk(sk);
948 if (metric > tp->reordering) {
949 int mib_idx;
950
951 tp->reordering = min(TCP_MAX_REORDERING, metric);
952
953 /* This exciting event is worth to be remembered. 8) */
954 if (ts)
955 mib_idx = LINUX_MIB_TCPTSREORDER;
956 else if (tcp_is_reno(tp))
957 mib_idx = LINUX_MIB_TCPRENOREORDER;
958 else if (tcp_is_fack(tp))
959 mib_idx = LINUX_MIB_TCPFACKREORDER;
960 else
961 mib_idx = LINUX_MIB_TCPSACKREORDER;
962
963 NET_INC_STATS_BH(sock_net(sk), mib_idx);
964#if FASTRETRANS_DEBUG > 1
965 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
966 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
967 tp->reordering,
968 tp->fackets_out,
969 tp->sacked_out,
970 tp->undo_marker ? tp->undo_retrans : 0);
971#endif
972 tcp_disable_fack(tp);
973 }
974}
975
976/* This must be called before lost_out is incremented */
977static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
978{
979 if ((tp->retransmit_skb_hint == NULL) ||
980 before(TCP_SKB_CB(skb)->seq,
981 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
982 tp->retransmit_skb_hint = skb;
983
984 if (!tp->lost_out ||
985 after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
986 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
987}
988
989static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
990{
991 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
992 tcp_verify_retransmit_hint(tp, skb);
993
994 tp->lost_out += tcp_skb_pcount(skb);
995 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
996 }
997}
998
999static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp,
1000 struct sk_buff *skb)
1001{
1002 tcp_verify_retransmit_hint(tp, skb);
1003
1004 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1005 tp->lost_out += tcp_skb_pcount(skb);
1006 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1007 }
1008}
1009
1010/* This procedure tags the retransmission queue when SACKs arrive.
1011 *
1012 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1013 * Packets in queue with these bits set are counted in variables
1014 * sacked_out, retrans_out and lost_out, correspondingly.
1015 *
1016 * Valid combinations are:
1017 * Tag InFlight Description
1018 * 0 1 - orig segment is in flight.
1019 * S 0 - nothing flies, orig reached receiver.
1020 * L 0 - nothing flies, orig lost by net.
1021 * R 2 - both orig and retransmit are in flight.
1022 * L|R 1 - orig is lost, retransmit is in flight.
1023 * S|R 1 - orig reached receiver, retrans is still in flight.
1024 * (L|S|R is logically valid, it could occur when L|R is sacked,
1025 * but it is equivalent to plain S and code short-curcuits it to S.
1026 * L|S is logically invalid, it would mean -1 packet in flight 8))
1027 *
1028 * These 6 states form finite state machine, controlled by the following events:
1029 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1030 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1031 * 3. Loss detection event of one of three flavors:
1032 * A. Scoreboard estimator decided the packet is lost.
1033 * A'. Reno "three dupacks" marks head of queue lost.
1034 * A''. Its FACK modfication, head until snd.fack is lost.
1035 * B. SACK arrives sacking data transmitted after never retransmitted
1036 * hole was sent out.
1037 * C. SACK arrives sacking SND.NXT at the moment, when the
1038 * segment was retransmitted.
1039 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1040 *
1041 * It is pleasant to note, that state diagram turns out to be commutative,
1042 * so that we are allowed not to be bothered by order of our actions,
1043 * when multiple events arrive simultaneously. (see the function below).
1044 *
1045 * Reordering detection.
1046 * --------------------
1047 * Reordering metric is maximal distance, which a packet can be displaced
1048 * in packet stream. With SACKs we can estimate it:
1049 *
1050 * 1. SACK fills old hole and the corresponding segment was not
1051 * ever retransmitted -> reordering. Alas, we cannot use it
1052 * when segment was retransmitted.
1053 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1054 * for retransmitted and already SACKed segment -> reordering..
1055 * Both of these heuristics are not used in Loss state, when we cannot
1056 * account for retransmits accurately.
1057 *
1058 * SACK block validation.
1059 * ----------------------
1060 *
1061 * SACK block range validation checks that the received SACK block fits to
1062 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1063 * Note that SND.UNA is not included to the range though being valid because
1064 * it means that the receiver is rather inconsistent with itself reporting
1065 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1066 * perfectly valid, however, in light of RFC2018 which explicitly states
1067 * that "SACK block MUST reflect the newest segment. Even if the newest
1068 * segment is going to be discarded ...", not that it looks very clever
1069 * in case of head skb. Due to potentional receiver driven attacks, we
1070 * choose to avoid immediate execution of a walk in write queue due to
1071 * reneging and defer head skb's loss recovery to standard loss recovery
1072 * procedure that will eventually trigger (nothing forbids us doing this).
1073 *
1074 * Implements also blockage to start_seq wrap-around. Problem lies in the
1075 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1076 * there's no guarantee that it will be before snd_nxt (n). The problem
1077 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1078 * wrap (s_w):
1079 *
1080 * <- outs wnd -> <- wrapzone ->
1081 * u e n u_w e_w s n_w
1082 * | | | | | | |
1083 * |<------------+------+----- TCP seqno space --------------+---------->|
1084 * ...-- <2^31 ->| |<--------...
1085 * ...---- >2^31 ------>| |<--------...
1086 *
1087 * Current code wouldn't be vulnerable but it's better still to discard such
1088 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1089 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1090 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1091 * equal to the ideal case (infinite seqno space without wrap caused issues).
1092 *
1093 * With D-SACK the lower bound is extended to cover sequence space below
1094 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1095 * again, D-SACK block must not to go across snd_una (for the same reason as
1096 * for the normal SACK blocks, explained above). But there all simplicity
1097 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1098 * fully below undo_marker they do not affect behavior in anyway and can
1099 * therefore be safely ignored. In rare cases (which are more or less
1100 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1101 * fragmentation and packet reordering past skb's retransmission. To consider
1102 * them correctly, the acceptable range must be extended even more though
1103 * the exact amount is rather hard to quantify. However, tp->max_window can
1104 * be used as an exaggerated estimate.
1105 */
1106static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack,
1107 u32 start_seq, u32 end_seq)
1108{
1109 /* Too far in future, or reversed (interpretation is ambiguous) */
1110 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1111 return 0;
1112
1113 /* Nasty start_seq wrap-around check (see comments above) */
1114 if (!before(start_seq, tp->snd_nxt))
1115 return 0;
1116
1117 /* In outstanding window? ...This is valid exit for D-SACKs too.
1118 * start_seq == snd_una is non-sensical (see comments above)
1119 */
1120 if (after(start_seq, tp->snd_una))
1121 return 1;
1122
1123 if (!is_dsack || !tp->undo_marker)
1124 return 0;
1125
1126 /* ...Then it's D-SACK, and must reside below snd_una completely */
1127 if (after(end_seq, tp->snd_una))
1128 return 0;
1129
1130 if (!before(start_seq, tp->undo_marker))
1131 return 1;
1132
1133 /* Too old */
1134 if (!after(end_seq, tp->undo_marker))
1135 return 0;
1136
1137 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1138 * start_seq < undo_marker and end_seq >= undo_marker.
1139 */
1140 return !before(start_seq, end_seq - tp->max_window);
1141}
1142
1143/* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1144 * Event "C". Later note: FACK people cheated me again 8), we have to account
1145 * for reordering! Ugly, but should help.
1146 *
1147 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1148 * less than what is now known to be received by the other end (derived from
1149 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1150 * retransmitted skbs to avoid some costly processing per ACKs.
1151 */
1152static void tcp_mark_lost_retrans(struct sock *sk)
1153{
1154 const struct inet_connection_sock *icsk = inet_csk(sk);
1155 struct tcp_sock *tp = tcp_sk(sk);
1156 struct sk_buff *skb;
1157 int cnt = 0;
1158 u32 new_low_seq = tp->snd_nxt;
1159 u32 received_upto = tcp_highest_sack_seq(tp);
1160
1161 if (!tcp_is_fack(tp) || !tp->retrans_out ||
1162 !after(received_upto, tp->lost_retrans_low) ||
1163 icsk->icsk_ca_state != TCP_CA_Recovery)
1164 return;
1165
1166 tcp_for_write_queue(skb, sk) {
1167 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1168
1169 if (skb == tcp_send_head(sk))
1170 break;
1171 if (cnt == tp->retrans_out)
1172 break;
1173 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1174 continue;
1175
1176 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1177 continue;
1178
1179 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1180 * constraint here (see above) but figuring out that at
1181 * least tp->reordering SACK blocks reside between ack_seq
1182 * and received_upto is not easy task to do cheaply with
1183 * the available datastructures.
1184 *
1185 * Whether FACK should check here for tp->reordering segs
1186 * in-between one could argue for either way (it would be
1187 * rather simple to implement as we could count fack_count
1188 * during the walk and do tp->fackets_out - fack_count).
1189 */
1190 if (after(received_upto, ack_seq)) {
1191 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1192 tp->retrans_out -= tcp_skb_pcount(skb);
1193
1194 tcp_skb_mark_lost_uncond_verify(tp, skb);
1195 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
1196 } else {
1197 if (before(ack_seq, new_low_seq))
1198 new_low_seq = ack_seq;
1199 cnt += tcp_skb_pcount(skb);
1200 }
1201 }
1202
1203 if (tp->retrans_out)
1204 tp->lost_retrans_low = new_low_seq;
1205}
1206
1207static int tcp_check_dsack(struct sock *sk, struct sk_buff *ack_skb,
1208 struct tcp_sack_block_wire *sp, int num_sacks,
1209 u32 prior_snd_una)
1210{
1211 struct tcp_sock *tp = tcp_sk(sk);
1212 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1213 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1214 int dup_sack = 0;
1215
1216 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1217 dup_sack = 1;
1218 tcp_dsack_seen(tp);
1219 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1220 } else if (num_sacks > 1) {
1221 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1222 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1223
1224 if (!after(end_seq_0, end_seq_1) &&
1225 !before(start_seq_0, start_seq_1)) {
1226 dup_sack = 1;
1227 tcp_dsack_seen(tp);
1228 NET_INC_STATS_BH(sock_net(sk),
1229 LINUX_MIB_TCPDSACKOFORECV);
1230 }
1231 }
1232
1233 /* D-SACK for already forgotten data... Do dumb counting. */
1234 if (dup_sack && tp->undo_marker && tp->undo_retrans &&
1235 !after(end_seq_0, prior_snd_una) &&
1236 after(end_seq_0, tp->undo_marker))
1237 tp->undo_retrans--;
1238
1239 return dup_sack;
1240}
1241
1242struct tcp_sacktag_state {
1243 int reord;
1244 int fack_count;
1245 int flag;
1246};
1247
1248/* Check if skb is fully within the SACK block. In presence of GSO skbs,
1249 * the incoming SACK may not exactly match but we can find smaller MSS
1250 * aligned portion of it that matches. Therefore we might need to fragment
1251 * which may fail and creates some hassle (caller must handle error case
1252 * returns).
1253 *
1254 * FIXME: this could be merged to shift decision code
1255 */
1256static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1257 u32 start_seq, u32 end_seq)
1258{
1259 int in_sack, err;
1260 unsigned int pkt_len;
1261 unsigned int mss;
1262
1263 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1264 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1265
1266 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1267 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1268 mss = tcp_skb_mss(skb);
1269 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1270
1271 if (!in_sack) {
1272 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1273 if (pkt_len < mss)
1274 pkt_len = mss;
1275 } else {
1276 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1277 if (pkt_len < mss)
1278 return -EINVAL;
1279 }
1280
1281 /* Round if necessary so that SACKs cover only full MSSes
1282 * and/or the remaining small portion (if present)
1283 */
1284 if (pkt_len > mss) {
1285 unsigned int new_len = (pkt_len / mss) * mss;
1286 if (!in_sack && new_len < pkt_len) {
1287 new_len += mss;
1288 if (new_len > skb->len)
1289 return 0;
1290 }
1291 pkt_len = new_len;
1292 }
1293 err = tcp_fragment(sk, skb, pkt_len, mss);
1294 if (err < 0)
1295 return err;
1296 }
1297
1298 return in_sack;
1299}
1300
1301static u8 tcp_sacktag_one(struct sk_buff *skb, struct sock *sk,
1302 struct tcp_sacktag_state *state,
1303 int dup_sack, int pcount)
1304{
1305 struct tcp_sock *tp = tcp_sk(sk);
1306 u8 sacked = TCP_SKB_CB(skb)->sacked;
1307 int fack_count = state->fack_count;
1308
1309 /* Account D-SACK for retransmitted packet. */
1310 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1311 if (tp->undo_marker && tp->undo_retrans &&
1312 after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1313 tp->undo_retrans--;
1314 if (sacked & TCPCB_SACKED_ACKED)
1315 state->reord = min(fack_count, state->reord);
1316 }
1317
1318 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1319 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1320 return sacked;
1321
1322 if (!(sacked & TCPCB_SACKED_ACKED)) {
1323 if (sacked & TCPCB_SACKED_RETRANS) {
1324 /* If the segment is not tagged as lost,
1325 * we do not clear RETRANS, believing
1326 * that retransmission is still in flight.
1327 */
1328 if (sacked & TCPCB_LOST) {
1329 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1330 tp->lost_out -= pcount;
1331 tp->retrans_out -= pcount;
1332 }
1333 } else {
1334 if (!(sacked & TCPCB_RETRANS)) {
1335 /* New sack for not retransmitted frame,
1336 * which was in hole. It is reordering.
1337 */
1338 if (before(TCP_SKB_CB(skb)->seq,
1339 tcp_highest_sack_seq(tp)))
1340 state->reord = min(fack_count,
1341 state->reord);
1342
1343 /* SACK enhanced F-RTO (RFC4138; Appendix B) */
1344 if (!after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark))
1345 state->flag |= FLAG_ONLY_ORIG_SACKED;
1346 }
1347
1348 if (sacked & TCPCB_LOST) {
1349 sacked &= ~TCPCB_LOST;
1350 tp->lost_out -= pcount;
1351 }
1352 }
1353
1354 sacked |= TCPCB_SACKED_ACKED;
1355 state->flag |= FLAG_DATA_SACKED;
1356 tp->sacked_out += pcount;
1357
1358 fack_count += pcount;
1359
1360 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1361 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
1362 before(TCP_SKB_CB(skb)->seq,
1363 TCP_SKB_CB(tp->lost_skb_hint)->seq))
1364 tp->lost_cnt_hint += pcount;
1365
1366 if (fack_count > tp->fackets_out)
1367 tp->fackets_out = fack_count;
1368 }
1369
1370 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1371 * frames and clear it. undo_retrans is decreased above, L|R frames
1372 * are accounted above as well.
1373 */
1374 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1375 sacked &= ~TCPCB_SACKED_RETRANS;
1376 tp->retrans_out -= pcount;
1377 }
1378
1379 return sacked;
1380}
1381
1382static int tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1383 struct tcp_sacktag_state *state,
1384 unsigned int pcount, int shifted, int mss,
1385 int dup_sack)
1386{
1387 struct tcp_sock *tp = tcp_sk(sk);
1388 struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1389
1390 BUG_ON(!pcount);
1391
1392 if (skb == tp->lost_skb_hint)
1393 tp->lost_cnt_hint += pcount;
1394
1395 TCP_SKB_CB(prev)->end_seq += shifted;
1396 TCP_SKB_CB(skb)->seq += shifted;
1397
1398 skb_shinfo(prev)->gso_segs += pcount;
1399 BUG_ON(skb_shinfo(skb)->gso_segs < pcount);
1400 skb_shinfo(skb)->gso_segs -= pcount;
1401
1402 /* When we're adding to gso_segs == 1, gso_size will be zero,
1403 * in theory this shouldn't be necessary but as long as DSACK
1404 * code can come after this skb later on it's better to keep
1405 * setting gso_size to something.
1406 */
1407 if (!skb_shinfo(prev)->gso_size) {
1408 skb_shinfo(prev)->gso_size = mss;
1409 skb_shinfo(prev)->gso_type = sk->sk_gso_type;
1410 }
1411
1412 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1413 if (skb_shinfo(skb)->gso_segs <= 1) {
1414 skb_shinfo(skb)->gso_size = 0;
1415 skb_shinfo(skb)->gso_type = 0;
1416 }
1417
1418 /* We discard results */
1419 tcp_sacktag_one(skb, sk, state, dup_sack, pcount);
1420
1421 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1422 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1423
1424 if (skb->len > 0) {
1425 BUG_ON(!tcp_skb_pcount(skb));
1426 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1427 return 0;
1428 }
1429
1430 /* Whole SKB was eaten :-) */
1431
1432 if (skb == tp->retransmit_skb_hint)
1433 tp->retransmit_skb_hint = prev;
1434 if (skb == tp->scoreboard_skb_hint)
1435 tp->scoreboard_skb_hint = prev;
1436 if (skb == tp->lost_skb_hint) {
1437 tp->lost_skb_hint = prev;
1438 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1439 }
1440
1441 TCP_SKB_CB(skb)->flags |= TCP_SKB_CB(prev)->flags;
1442 if (skb == tcp_highest_sack(sk))
1443 tcp_advance_highest_sack(sk, skb);
1444
1445 tcp_unlink_write_queue(skb, sk);
1446 sk_wmem_free_skb(sk, skb);
1447
1448 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1449
1450 return 1;
1451}
1452
1453/* I wish gso_size would have a bit more sane initialization than
1454 * something-or-zero which complicates things
1455 */
1456static int tcp_skb_seglen(struct sk_buff *skb)
1457{
1458 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1459}
1460
1461/* Shifting pages past head area doesn't work */
1462static int skb_can_shift(struct sk_buff *skb)
1463{
1464 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1465}
1466
1467/* Try collapsing SACK blocks spanning across multiple skbs to a single
1468 * skb.
1469 */
1470static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1471 struct tcp_sacktag_state *state,
1472 u32 start_seq, u32 end_seq,
1473 int dup_sack)
1474{
1475 struct tcp_sock *tp = tcp_sk(sk);
1476 struct sk_buff *prev;
1477 int mss;
1478 int pcount = 0;
1479 int len;
1480 int in_sack;
1481
1482 if (!sk_can_gso(sk))
1483 goto fallback;
1484
1485 /* Normally R but no L won't result in plain S */
1486 if (!dup_sack &&
1487 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1488 goto fallback;
1489 if (!skb_can_shift(skb))
1490 goto fallback;
1491 /* This frame is about to be dropped (was ACKed). */
1492 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1493 goto fallback;
1494
1495 /* Can only happen with delayed DSACK + discard craziness */
1496 if (unlikely(skb == tcp_write_queue_head(sk)))
1497 goto fallback;
1498 prev = tcp_write_queue_prev(sk, skb);
1499
1500 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1501 goto fallback;
1502
1503 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1504 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1505
1506 if (in_sack) {
1507 len = skb->len;
1508 pcount = tcp_skb_pcount(skb);
1509 mss = tcp_skb_seglen(skb);
1510
1511 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1512 * drop this restriction as unnecessary
1513 */
1514 if (mss != tcp_skb_seglen(prev))
1515 goto fallback;
1516 } else {
1517 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1518 goto noop;
1519 /* CHECKME: This is non-MSS split case only?, this will
1520 * cause skipped skbs due to advancing loop btw, original
1521 * has that feature too
1522 */
1523 if (tcp_skb_pcount(skb) <= 1)
1524 goto noop;
1525
1526 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1527 if (!in_sack) {
1528 /* TODO: head merge to next could be attempted here
1529 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1530 * though it might not be worth of the additional hassle
1531 *
1532 * ...we can probably just fallback to what was done
1533 * previously. We could try merging non-SACKed ones
1534 * as well but it probably isn't going to buy off
1535 * because later SACKs might again split them, and
1536 * it would make skb timestamp tracking considerably
1537 * harder problem.
1538 */
1539 goto fallback;
1540 }
1541
1542 len = end_seq - TCP_SKB_CB(skb)->seq;
1543 BUG_ON(len < 0);
1544 BUG_ON(len > skb->len);
1545
1546 /* MSS boundaries should be honoured or else pcount will
1547 * severely break even though it makes things bit trickier.
1548 * Optimize common case to avoid most of the divides
1549 */
1550 mss = tcp_skb_mss(skb);
1551
1552 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1553 * drop this restriction as unnecessary
1554 */
1555 if (mss != tcp_skb_seglen(prev))
1556 goto fallback;
1557
1558 if (len == mss) {
1559 pcount = 1;
1560 } else if (len < mss) {
1561 goto noop;
1562 } else {
1563 pcount = len / mss;
1564 len = pcount * mss;
1565 }
1566 }
1567
1568 if (!skb_shift(prev, skb, len))
1569 goto fallback;
1570 if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1571 goto out;
1572
1573 /* Hole filled allows collapsing with the next as well, this is very
1574 * useful when hole on every nth skb pattern happens
1575 */
1576 if (prev == tcp_write_queue_tail(sk))
1577 goto out;
1578 skb = tcp_write_queue_next(sk, prev);
1579
1580 if (!skb_can_shift(skb) ||
1581 (skb == tcp_send_head(sk)) ||
1582 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1583 (mss != tcp_skb_seglen(skb)))
1584 goto out;
1585
1586 len = skb->len;
1587 if (skb_shift(prev, skb, len)) {
1588 pcount += tcp_skb_pcount(skb);
1589 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1590 }
1591
1592out:
1593 state->fack_count += pcount;
1594 return prev;
1595
1596noop:
1597 return skb;
1598
1599fallback:
1600 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1601 return NULL;
1602}
1603
1604static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1605 struct tcp_sack_block *next_dup,
1606 struct tcp_sacktag_state *state,
1607 u32 start_seq, u32 end_seq,
1608 int dup_sack_in)
1609{
1610 struct tcp_sock *tp = tcp_sk(sk);
1611 struct sk_buff *tmp;
1612
1613 tcp_for_write_queue_from(skb, sk) {
1614 int in_sack = 0;
1615 int dup_sack = dup_sack_in;
1616
1617 if (skb == tcp_send_head(sk))
1618 break;
1619
1620 /* queue is in-order => we can short-circuit the walk early */
1621 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1622 break;
1623
1624 if ((next_dup != NULL) &&
1625 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1626 in_sack = tcp_match_skb_to_sack(sk, skb,
1627 next_dup->start_seq,
1628 next_dup->end_seq);
1629 if (in_sack > 0)
1630 dup_sack = 1;
1631 }
1632
1633 /* skb reference here is a bit tricky to get right, since
1634 * shifting can eat and free both this skb and the next,
1635 * so not even _safe variant of the loop is enough.
1636 */
1637 if (in_sack <= 0) {
1638 tmp = tcp_shift_skb_data(sk, skb, state,
1639 start_seq, end_seq, dup_sack);
1640 if (tmp != NULL) {
1641 if (tmp != skb) {
1642 skb = tmp;
1643 continue;
1644 }
1645
1646 in_sack = 0;
1647 } else {
1648 in_sack = tcp_match_skb_to_sack(sk, skb,
1649 start_seq,
1650 end_seq);
1651 }
1652 }
1653
1654 if (unlikely(in_sack < 0))
1655 break;
1656
1657 if (in_sack) {
1658 TCP_SKB_CB(skb)->sacked = tcp_sacktag_one(skb, sk,
1659 state,
1660 dup_sack,
1661 tcp_skb_pcount(skb));
1662
1663 if (!before(TCP_SKB_CB(skb)->seq,
1664 tcp_highest_sack_seq(tp)))
1665 tcp_advance_highest_sack(sk, skb);
1666 }
1667
1668 state->fack_count += tcp_skb_pcount(skb);
1669 }
1670 return skb;
1671}
1672
1673/* Avoid all extra work that is being done by sacktag while walking in
1674 * a normal way
1675 */
1676static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1677 struct tcp_sacktag_state *state,
1678 u32 skip_to_seq)
1679{
1680 tcp_for_write_queue_from(skb, sk) {
1681 if (skb == tcp_send_head(sk))
1682 break;
1683
1684 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1685 break;
1686
1687 state->fack_count += tcp_skb_pcount(skb);
1688 }
1689 return skb;
1690}
1691
1692static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1693 struct sock *sk,
1694 struct tcp_sack_block *next_dup,
1695 struct tcp_sacktag_state *state,
1696 u32 skip_to_seq)
1697{
1698 if (next_dup == NULL)
1699 return skb;
1700
1701 if (before(next_dup->start_seq, skip_to_seq)) {
1702 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1703 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1704 next_dup->start_seq, next_dup->end_seq,
1705 1);
1706 }
1707
1708 return skb;
1709}
1710
1711static int tcp_sack_cache_ok(struct tcp_sock *tp, struct tcp_sack_block *cache)
1712{
1713 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1714}
1715
1716static int
1717tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb,
1718 u32 prior_snd_una)
1719{
1720 const struct inet_connection_sock *icsk = inet_csk(sk);
1721 struct tcp_sock *tp = tcp_sk(sk);
1722 unsigned char *ptr = (skb_transport_header(ack_skb) +
1723 TCP_SKB_CB(ack_skb)->sacked);
1724 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1725 struct tcp_sack_block sp[TCP_NUM_SACKS];
1726 struct tcp_sack_block *cache;
1727 struct tcp_sacktag_state state;
1728 struct sk_buff *skb;
1729 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1730 int used_sacks;
1731 int found_dup_sack = 0;
1732 int i, j;
1733 int first_sack_index;
1734
1735 state.flag = 0;
1736 state.reord = tp->packets_out;
1737
1738 if (!tp->sacked_out) {
1739 if (WARN_ON(tp->fackets_out))
1740 tp->fackets_out = 0;
1741 tcp_highest_sack_reset(sk);
1742 }
1743
1744 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1745 num_sacks, prior_snd_una);
1746 if (found_dup_sack)
1747 state.flag |= FLAG_DSACKING_ACK;
1748
1749 /* Eliminate too old ACKs, but take into
1750 * account more or less fresh ones, they can
1751 * contain valid SACK info.
1752 */
1753 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1754 return 0;
1755
1756 if (!tp->packets_out)
1757 goto out;
1758
1759 used_sacks = 0;
1760 first_sack_index = 0;
1761 for (i = 0; i < num_sacks; i++) {
1762 int dup_sack = !i && found_dup_sack;
1763
1764 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1765 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1766
1767 if (!tcp_is_sackblock_valid(tp, dup_sack,
1768 sp[used_sacks].start_seq,
1769 sp[used_sacks].end_seq)) {
1770 int mib_idx;
1771
1772 if (dup_sack) {
1773 if (!tp->undo_marker)
1774 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1775 else
1776 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1777 } else {
1778 /* Don't count olds caused by ACK reordering */
1779 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1780 !after(sp[used_sacks].end_seq, tp->snd_una))
1781 continue;
1782 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1783 }
1784
1785 NET_INC_STATS_BH(sock_net(sk), mib_idx);
1786 if (i == 0)
1787 first_sack_index = -1;
1788 continue;
1789 }
1790
1791 /* Ignore very old stuff early */
1792 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1793 continue;
1794
1795 used_sacks++;
1796 }
1797
1798 /* order SACK blocks to allow in order walk of the retrans queue */
1799 for (i = used_sacks - 1; i > 0; i--) {
1800 for (j = 0; j < i; j++) {
1801 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1802 swap(sp[j], sp[j + 1]);
1803
1804 /* Track where the first SACK block goes to */
1805 if (j == first_sack_index)
1806 first_sack_index = j + 1;
1807 }
1808 }
1809 }
1810
1811 skb = tcp_write_queue_head(sk);
1812 state.fack_count = 0;
1813 i = 0;
1814
1815 if (!tp->sacked_out) {
1816 /* It's already past, so skip checking against it */
1817 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1818 } else {
1819 cache = tp->recv_sack_cache;
1820 /* Skip empty blocks in at head of the cache */
1821 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1822 !cache->end_seq)
1823 cache++;
1824 }
1825
1826 while (i < used_sacks) {
1827 u32 start_seq = sp[i].start_seq;
1828 u32 end_seq = sp[i].end_seq;
1829 int dup_sack = (found_dup_sack && (i == first_sack_index));
1830 struct tcp_sack_block *next_dup = NULL;
1831
1832 if (found_dup_sack && ((i + 1) == first_sack_index))
1833 next_dup = &sp[i + 1];
1834
1835 /* Event "B" in the comment above. */
1836 if (after(end_seq, tp->high_seq))
1837 state.flag |= FLAG_DATA_LOST;
1838
1839 /* Skip too early cached blocks */
1840 while (tcp_sack_cache_ok(tp, cache) &&
1841 !before(start_seq, cache->end_seq))
1842 cache++;
1843
1844 /* Can skip some work by looking recv_sack_cache? */
1845 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1846 after(end_seq, cache->start_seq)) {
1847
1848 /* Head todo? */
1849 if (before(start_seq, cache->start_seq)) {
1850 skb = tcp_sacktag_skip(skb, sk, &state,
1851 start_seq);
1852 skb = tcp_sacktag_walk(skb, sk, next_dup,
1853 &state,
1854 start_seq,
1855 cache->start_seq,
1856 dup_sack);
1857 }
1858
1859 /* Rest of the block already fully processed? */
1860 if (!after(end_seq, cache->end_seq))
1861 goto advance_sp;
1862
1863 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1864 &state,
1865 cache->end_seq);
1866
1867 /* ...tail remains todo... */
1868 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1869 /* ...but better entrypoint exists! */
1870 skb = tcp_highest_sack(sk);
1871 if (skb == NULL)
1872 break;
1873 state.fack_count = tp->fackets_out;
1874 cache++;
1875 goto walk;
1876 }
1877
1878 skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq);
1879 /* Check overlap against next cached too (past this one already) */
1880 cache++;
1881 continue;
1882 }
1883
1884 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1885 skb = tcp_highest_sack(sk);
1886 if (skb == NULL)
1887 break;
1888 state.fack_count = tp->fackets_out;
1889 }
1890 skb = tcp_sacktag_skip(skb, sk, &state, start_seq);
1891
1892walk:
1893 skb = tcp_sacktag_walk(skb, sk, next_dup, &state,
1894 start_seq, end_seq, dup_sack);
1895
1896advance_sp:
1897 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1898 * due to in-order walk
1899 */
1900 if (after(end_seq, tp->frto_highmark))
1901 state.flag &= ~FLAG_ONLY_ORIG_SACKED;
1902
1903 i++;
1904 }
1905
1906 /* Clear the head of the cache sack blocks so we can skip it next time */
1907 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1908 tp->recv_sack_cache[i].start_seq = 0;
1909 tp->recv_sack_cache[i].end_seq = 0;
1910 }
1911 for (j = 0; j < used_sacks; j++)
1912 tp->recv_sack_cache[i++] = sp[j];
1913
1914 tcp_mark_lost_retrans(sk);
1915
1916 tcp_verify_left_out(tp);
1917
1918 if ((state.reord < tp->fackets_out) &&
1919 ((icsk->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker) &&
1920 (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
1921 tcp_update_reordering(sk, tp->fackets_out - state.reord, 0);
1922
1923out:
1924
1925#if FASTRETRANS_DEBUG > 0
1926 WARN_ON((int)tp->sacked_out < 0);
1927 WARN_ON((int)tp->lost_out < 0);
1928 WARN_ON((int)tp->retrans_out < 0);
1929 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1930#endif
1931 return state.flag;
1932}
1933
1934/* Limits sacked_out so that sum with lost_out isn't ever larger than
1935 * packets_out. Returns zero if sacked_out adjustement wasn't necessary.
1936 */
1937static int tcp_limit_reno_sacked(struct tcp_sock *tp)
1938{
1939 u32 holes;
1940
1941 holes = max(tp->lost_out, 1U);
1942 holes = min(holes, tp->packets_out);
1943
1944 if ((tp->sacked_out + holes) > tp->packets_out) {
1945 tp->sacked_out = tp->packets_out - holes;
1946 return 1;
1947 }
1948 return 0;
1949}
1950
1951/* If we receive more dupacks than we expected counting segments
1952 * in assumption of absent reordering, interpret this as reordering.
1953 * The only another reason could be bug in receiver TCP.
1954 */
1955static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1956{
1957 struct tcp_sock *tp = tcp_sk(sk);
1958 if (tcp_limit_reno_sacked(tp))
1959 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1960}
1961
1962/* Emulate SACKs for SACKless connection: account for a new dupack. */
1963
1964static void tcp_add_reno_sack(struct sock *sk)
1965{
1966 struct tcp_sock *tp = tcp_sk(sk);
1967 tp->sacked_out++;
1968 tcp_check_reno_reordering(sk, 0);
1969 tcp_verify_left_out(tp);
1970}
1971
1972/* Account for ACK, ACKing some data in Reno Recovery phase. */
1973
1974static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1975{
1976 struct tcp_sock *tp = tcp_sk(sk);
1977
1978 if (acked > 0) {
1979 /* One ACK acked hole. The rest eat duplicate ACKs. */
1980 if (acked - 1 >= tp->sacked_out)
1981 tp->sacked_out = 0;
1982 else
1983 tp->sacked_out -= acked - 1;
1984 }
1985 tcp_check_reno_reordering(sk, acked);
1986 tcp_verify_left_out(tp);
1987}
1988
1989static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1990{
1991 tp->sacked_out = 0;
1992}
1993
1994static int tcp_is_sackfrto(const struct tcp_sock *tp)
1995{
1996 return (sysctl_tcp_frto == 0x2) && !tcp_is_reno(tp);
1997}
1998
1999/* F-RTO can only be used if TCP has never retransmitted anything other than
2000 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
2001 */
2002int tcp_use_frto(struct sock *sk)
2003{
2004 const struct tcp_sock *tp = tcp_sk(sk);
2005 const struct inet_connection_sock *icsk = inet_csk(sk);
2006 struct sk_buff *skb;
2007
2008 if (!sysctl_tcp_frto)
2009 return 0;
2010
2011 /* MTU probe and F-RTO won't really play nicely along currently */
2012 if (icsk->icsk_mtup.probe_size)
2013 return 0;
2014
2015 if (tcp_is_sackfrto(tp))
2016 return 1;
2017
2018 /* Avoid expensive walking of rexmit queue if possible */
2019 if (tp->retrans_out > 1)
2020 return 0;
2021
2022 skb = tcp_write_queue_head(sk);
2023 if (tcp_skb_is_last(sk, skb))
2024 return 1;
2025 skb = tcp_write_queue_next(sk, skb); /* Skips head */
2026 tcp_for_write_queue_from(skb, sk) {
2027 if (skb == tcp_send_head(sk))
2028 break;
2029 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2030 return 0;
2031 /* Short-circuit when first non-SACKed skb has been checked */
2032 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2033 break;
2034 }
2035 return 1;
2036}
2037
2038/* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
2039 * recovery a bit and use heuristics in tcp_process_frto() to detect if
2040 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
2041 * keep retrans_out counting accurate (with SACK F-RTO, other than head
2042 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
2043 * bits are handled if the Loss state is really to be entered (in
2044 * tcp_enter_frto_loss).
2045 *
2046 * Do like tcp_enter_loss() would; when RTO expires the second time it
2047 * does:
2048 * "Reduce ssthresh if it has not yet been made inside this window."
2049 */
2050void tcp_enter_frto(struct sock *sk)
2051{
2052 const struct inet_connection_sock *icsk = inet_csk(sk);
2053 struct tcp_sock *tp = tcp_sk(sk);
2054 struct sk_buff *skb;
2055
2056 if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
2057 tp->snd_una == tp->high_seq ||
2058 ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
2059 !icsk->icsk_retransmits)) {
2060 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2061 /* Our state is too optimistic in ssthresh() call because cwnd
2062 * is not reduced until tcp_enter_frto_loss() when previous F-RTO
2063 * recovery has not yet completed. Pattern would be this: RTO,
2064 * Cumulative ACK, RTO (2xRTO for the same segment does not end
2065 * up here twice).
2066 * RFC4138 should be more specific on what to do, even though
2067 * RTO is quite unlikely to occur after the first Cumulative ACK
2068 * due to back-off and complexity of triggering events ...
2069 */
2070 if (tp->frto_counter) {
2071 u32 stored_cwnd;
2072 stored_cwnd = tp->snd_cwnd;
2073 tp->snd_cwnd = 2;
2074 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2075 tp->snd_cwnd = stored_cwnd;
2076 } else {
2077 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2078 }
2079 /* ... in theory, cong.control module could do "any tricks" in
2080 * ssthresh(), which means that ca_state, lost bits and lost_out
2081 * counter would have to be faked before the call occurs. We
2082 * consider that too expensive, unlikely and hacky, so modules
2083 * using these in ssthresh() must deal these incompatibility
2084 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
2085 */
2086 tcp_ca_event(sk, CA_EVENT_FRTO);
2087 }
2088
2089 tp->undo_marker = tp->snd_una;
2090 tp->undo_retrans = 0;
2091
2092 skb = tcp_write_queue_head(sk);
2093 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2094 tp->undo_marker = 0;
2095 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2096 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2097 tp->retrans_out -= tcp_skb_pcount(skb);
2098 }
2099 tcp_verify_left_out(tp);
2100
2101 /* Too bad if TCP was application limited */
2102 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2103
2104 /* Earlier loss recovery underway (see RFC4138; Appendix B).
2105 * The last condition is necessary at least in tp->frto_counter case.
2106 */
2107 if (tcp_is_sackfrto(tp) && (tp->frto_counter ||
2108 ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
2109 after(tp->high_seq, tp->snd_una)) {
2110 tp->frto_highmark = tp->high_seq;
2111 } else {
2112 tp->frto_highmark = tp->snd_nxt;
2113 }
2114 tcp_set_ca_state(sk, TCP_CA_Disorder);
2115 tp->high_seq = tp->snd_nxt;
2116 tp->frto_counter = 1;
2117}
2118
2119/* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
2120 * which indicates that we should follow the traditional RTO recovery,
2121 * i.e. mark everything lost and do go-back-N retransmission.
2122 */
2123static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
2124{
2125 struct tcp_sock *tp = tcp_sk(sk);
2126 struct sk_buff *skb;
2127
2128 tp->lost_out = 0;
2129 tp->retrans_out = 0;
2130 if (tcp_is_reno(tp))
2131 tcp_reset_reno_sack(tp);
2132
2133 tcp_for_write_queue(skb, sk) {
2134 if (skb == tcp_send_head(sk))
2135 break;
2136
2137 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2138 /*
2139 * Count the retransmission made on RTO correctly (only when
2140 * waiting for the first ACK and did not get it)...
2141 */
2142 if ((tp->frto_counter == 1) && !(flag & FLAG_DATA_ACKED)) {
2143 /* For some reason this R-bit might get cleared? */
2144 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
2145 tp->retrans_out += tcp_skb_pcount(skb);
2146 /* ...enter this if branch just for the first segment */
2147 flag |= FLAG_DATA_ACKED;
2148 } else {
2149 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2150 tp->undo_marker = 0;
2151 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2152 }
2153
2154 /* Marking forward transmissions that were made after RTO lost
2155 * can cause unnecessary retransmissions in some scenarios,
2156 * SACK blocks will mitigate that in some but not in all cases.
2157 * We used to not mark them but it was causing break-ups with
2158 * receivers that do only in-order receival.
2159 *
2160 * TODO: we could detect presence of such receiver and select
2161 * different behavior per flow.
2162 */
2163 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2164 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2165 tp->lost_out += tcp_skb_pcount(skb);
2166 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
2167 }
2168 }
2169 tcp_verify_left_out(tp);
2170
2171 tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
2172 tp->snd_cwnd_cnt = 0;
2173 tp->snd_cwnd_stamp = tcp_time_stamp;
2174 tp->frto_counter = 0;
2175 tp->bytes_acked = 0;
2176
2177 tp->reordering = min_t(unsigned int, tp->reordering,
2178 sysctl_tcp_reordering);
2179 tcp_set_ca_state(sk, TCP_CA_Loss);
2180 tp->high_seq = tp->snd_nxt;
2181 TCP_ECN_queue_cwr(tp);
2182
2183 tcp_clear_all_retrans_hints(tp);
2184}
2185
2186static void tcp_clear_retrans_partial(struct tcp_sock *tp)
2187{
2188 tp->retrans_out = 0;
2189 tp->lost_out = 0;
2190
2191 tp->undo_marker = 0;
2192 tp->undo_retrans = 0;
2193}
2194
2195void tcp_clear_retrans(struct tcp_sock *tp)
2196{
2197 tcp_clear_retrans_partial(tp);
2198
2199 tp->fackets_out = 0;
2200 tp->sacked_out = 0;
2201}
2202
2203/* Enter Loss state. If "how" is not zero, forget all SACK information
2204 * and reset tags completely, otherwise preserve SACKs. If receiver
2205 * dropped its ofo queue, we will know this due to reneging detection.
2206 */
2207void tcp_enter_loss(struct sock *sk, int how)
2208{
2209 const struct inet_connection_sock *icsk = inet_csk(sk);
2210 struct tcp_sock *tp = tcp_sk(sk);
2211 struct sk_buff *skb;
2212
2213 /* Reduce ssthresh if it has not yet been made inside this window. */
2214 if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
2215 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2216 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2217 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2218 tcp_ca_event(sk, CA_EVENT_LOSS);
2219 }
2220 tp->snd_cwnd = 1;
2221 tp->snd_cwnd_cnt = 0;
2222 tp->snd_cwnd_stamp = tcp_time_stamp;
2223
2224 tp->bytes_acked = 0;
2225 tcp_clear_retrans_partial(tp);
2226
2227 if (tcp_is_reno(tp))
2228 tcp_reset_reno_sack(tp);
2229
2230 if (!how) {
2231 /* Push undo marker, if it was plain RTO and nothing
2232 * was retransmitted. */
2233 tp->undo_marker = tp->snd_una;
2234 } else {
2235 tp->sacked_out = 0;
2236 tp->fackets_out = 0;
2237 }
2238 tcp_clear_all_retrans_hints(tp);
2239
2240 tcp_for_write_queue(skb, sk) {
2241 if (skb == tcp_send_head(sk))
2242 break;
2243
2244 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2245 tp->undo_marker = 0;
2246 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
2247 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
2248 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
2249 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2250 tp->lost_out += tcp_skb_pcount(skb);
2251 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
2252 }
2253 }
2254 tcp_verify_left_out(tp);
2255
2256 tp->reordering = min_t(unsigned int, tp->reordering,
2257 sysctl_tcp_reordering);
2258 tcp_set_ca_state(sk, TCP_CA_Loss);
2259 tp->high_seq = tp->snd_nxt;
2260 TCP_ECN_queue_cwr(tp);
2261 /* Abort F-RTO algorithm if one is in progress */
2262 tp->frto_counter = 0;
2263}
2264
2265/* If ACK arrived pointing to a remembered SACK, it means that our
2266 * remembered SACKs do not reflect real state of receiver i.e.
2267 * receiver _host_ is heavily congested (or buggy).
2268 *
2269 * Do processing similar to RTO timeout.
2270 */
2271static int tcp_check_sack_reneging(struct sock *sk, int flag)
2272{
2273 if (flag & FLAG_SACK_RENEGING) {
2274 struct inet_connection_sock *icsk = inet_csk(sk);
2275 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
2276
2277 tcp_enter_loss(sk, 1);
2278 icsk->icsk_retransmits++;
2279 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
2280 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2281 icsk->icsk_rto, TCP_RTO_MAX);
2282 return 1;
2283 }
2284 return 0;
2285}
2286
2287static inline int tcp_fackets_out(struct tcp_sock *tp)
2288{
2289 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
2290}
2291
2292/* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2293 * counter when SACK is enabled (without SACK, sacked_out is used for
2294 * that purpose).
2295 *
2296 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2297 * segments up to the highest received SACK block so far and holes in
2298 * between them.
2299 *
2300 * With reordering, holes may still be in flight, so RFC3517 recovery
2301 * uses pure sacked_out (total number of SACKed segments) even though
2302 * it violates the RFC that uses duplicate ACKs, often these are equal
2303 * but when e.g. out-of-window ACKs or packet duplication occurs,
2304 * they differ. Since neither occurs due to loss, TCP should really
2305 * ignore them.
2306 */
2307static inline int tcp_dupack_heuristics(struct tcp_sock *tp)
2308{
2309 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2310}
2311
2312static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb)
2313{
2314 return tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto;
2315}
2316
2317static inline int tcp_head_timedout(struct sock *sk)
2318{
2319 struct tcp_sock *tp = tcp_sk(sk);
2320
2321 return tp->packets_out &&
2322 tcp_skb_timedout(sk, tcp_write_queue_head(sk));
2323}
2324
2325/* Linux NewReno/SACK/FACK/ECN state machine.
2326 * --------------------------------------
2327 *
2328 * "Open" Normal state, no dubious events, fast path.
2329 * "Disorder" In all the respects it is "Open",
2330 * but requires a bit more attention. It is entered when
2331 * we see some SACKs or dupacks. It is split of "Open"
2332 * mainly to move some processing from fast path to slow one.
2333 * "CWR" CWND was reduced due to some Congestion Notification event.
2334 * It can be ECN, ICMP source quench, local device congestion.
2335 * "Recovery" CWND was reduced, we are fast-retransmitting.
2336 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2337 *
2338 * tcp_fastretrans_alert() is entered:
2339 * - each incoming ACK, if state is not "Open"
2340 * - when arrived ACK is unusual, namely:
2341 * * SACK
2342 * * Duplicate ACK.
2343 * * ECN ECE.
2344 *
2345 * Counting packets in flight is pretty simple.
2346 *
2347 * in_flight = packets_out - left_out + retrans_out
2348 *
2349 * packets_out is SND.NXT-SND.UNA counted in packets.
2350 *
2351 * retrans_out is number of retransmitted segments.
2352 *
2353 * left_out is number of segments left network, but not ACKed yet.
2354 *
2355 * left_out = sacked_out + lost_out
2356 *
2357 * sacked_out: Packets, which arrived to receiver out of order
2358 * and hence not ACKed. With SACKs this number is simply
2359 * amount of SACKed data. Even without SACKs
2360 * it is easy to give pretty reliable estimate of this number,
2361 * counting duplicate ACKs.
2362 *
2363 * lost_out: Packets lost by network. TCP has no explicit
2364 * "loss notification" feedback from network (for now).
2365 * It means that this number can be only _guessed_.
2366 * Actually, it is the heuristics to predict lossage that
2367 * distinguishes different algorithms.
2368 *
2369 * F.e. after RTO, when all the queue is considered as lost,
2370 * lost_out = packets_out and in_flight = retrans_out.
2371 *
2372 * Essentially, we have now two algorithms counting
2373 * lost packets.
2374 *
2375 * FACK: It is the simplest heuristics. As soon as we decided
2376 * that something is lost, we decide that _all_ not SACKed
2377 * packets until the most forward SACK are lost. I.e.
2378 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2379 * It is absolutely correct estimate, if network does not reorder
2380 * packets. And it loses any connection to reality when reordering
2381 * takes place. We use FACK by default until reordering
2382 * is suspected on the path to this destination.
2383 *
2384 * NewReno: when Recovery is entered, we assume that one segment
2385 * is lost (classic Reno). While we are in Recovery and
2386 * a partial ACK arrives, we assume that one more packet
2387 * is lost (NewReno). This heuristics are the same in NewReno
2388 * and SACK.
2389 *
2390 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2391 * deflation etc. CWND is real congestion window, never inflated, changes
2392 * only according to classic VJ rules.
2393 *
2394 * Really tricky (and requiring careful tuning) part of algorithm
2395 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2396 * The first determines the moment _when_ we should reduce CWND and,
2397 * hence, slow down forward transmission. In fact, it determines the moment
2398 * when we decide that hole is caused by loss, rather than by a reorder.
2399 *
2400 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2401 * holes, caused by lost packets.
2402 *
2403 * And the most logically complicated part of algorithm is undo
2404 * heuristics. We detect false retransmits due to both too early
2405 * fast retransmit (reordering) and underestimated RTO, analyzing
2406 * timestamps and D-SACKs. When we detect that some segments were
2407 * retransmitted by mistake and CWND reduction was wrong, we undo
2408 * window reduction and abort recovery phase. This logic is hidden
2409 * inside several functions named tcp_try_undo_<something>.
2410 */
2411
2412/* This function decides, when we should leave Disordered state
2413 * and enter Recovery phase, reducing congestion window.
2414 *
2415 * Main question: may we further continue forward transmission
2416 * with the same cwnd?
2417 */
2418static int tcp_time_to_recover(struct sock *sk)
2419{
2420 struct tcp_sock *tp = tcp_sk(sk);
2421 __u32 packets_out;
2422
2423 /* Do not perform any recovery during F-RTO algorithm */
2424 if (tp->frto_counter)
2425 return 0;
2426
2427 /* Trick#1: The loss is proven. */
2428 if (tp->lost_out)
2429 return 1;
2430
2431 /* Not-A-Trick#2 : Classic rule... */
2432 if (tcp_dupack_heuristics(tp) > tp->reordering)
2433 return 1;
2434
2435 /* Trick#3 : when we use RFC2988 timer restart, fast
2436 * retransmit can be triggered by timeout of queue head.
2437 */
2438 if (tcp_is_fack(tp) && tcp_head_timedout(sk))
2439 return 1;
2440
2441 /* Trick#4: It is still not OK... But will it be useful to delay
2442 * recovery more?
2443 */
2444 packets_out = tp->packets_out;
2445 if (packets_out <= tp->reordering &&
2446 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2447 !tcp_may_send_now(sk)) {
2448 /* We have nothing to send. This connection is limited
2449 * either by receiver window or by application.
2450 */
2451 return 1;
2452 }
2453
2454 /* If a thin stream is detected, retransmit after first
2455 * received dupack. Employ only if SACK is supported in order
2456 * to avoid possible corner-case series of spurious retransmissions
2457 * Use only if there are no unsent data.
2458 */
2459 if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2460 tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2461 tcp_is_sack(tp) && !tcp_send_head(sk))
2462 return 1;
2463
2464 return 0;
2465}
2466
2467/* New heuristics: it is possible only after we switched to restart timer
2468 * each time when something is ACKed. Hence, we can detect timed out packets
2469 * during fast retransmit without falling to slow start.
2470 *
2471 * Usefulness of this as is very questionable, since we should know which of
2472 * the segments is the next to timeout which is relatively expensive to find
2473 * in general case unless we add some data structure just for that. The
2474 * current approach certainly won't find the right one too often and when it
2475 * finally does find _something_ it usually marks large part of the window
2476 * right away (because a retransmission with a larger timestamp blocks the
2477 * loop from advancing). -ij
2478 */
2479static void tcp_timeout_skbs(struct sock *sk)
2480{
2481 struct tcp_sock *tp = tcp_sk(sk);
2482 struct sk_buff *skb;
2483
2484 if (!tcp_is_fack(tp) || !tcp_head_timedout(sk))
2485 return;
2486
2487 skb = tp->scoreboard_skb_hint;
2488 if (tp->scoreboard_skb_hint == NULL)
2489 skb = tcp_write_queue_head(sk);
2490
2491 tcp_for_write_queue_from(skb, sk) {
2492 if (skb == tcp_send_head(sk))
2493 break;
2494 if (!tcp_skb_timedout(sk, skb))
2495 break;
2496
2497 tcp_skb_mark_lost(tp, skb);
2498 }
2499
2500 tp->scoreboard_skb_hint = skb;
2501
2502 tcp_verify_left_out(tp);
2503}
2504
2505/* Mark head of queue up as lost. With RFC3517 SACK, the packets is
2506 * is against sacked "cnt", otherwise it's against facked "cnt"
2507 */
2508static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2509{
2510 struct tcp_sock *tp = tcp_sk(sk);
2511 struct sk_buff *skb;
2512 int cnt, oldcnt;
2513 int err;
2514 unsigned int mss;
2515
2516 WARN_ON(packets > tp->packets_out);
2517 if (tp->lost_skb_hint) {
2518 skb = tp->lost_skb_hint;
2519 cnt = tp->lost_cnt_hint;
2520 /* Head already handled? */
2521 if (mark_head && skb != tcp_write_queue_head(sk))
2522 return;
2523 } else {
2524 skb = tcp_write_queue_head(sk);
2525 cnt = 0;
2526 }
2527
2528 tcp_for_write_queue_from(skb, sk) {
2529 if (skb == tcp_send_head(sk))
2530 break;
2531 /* TODO: do this better */
2532 /* this is not the most efficient way to do this... */
2533 tp->lost_skb_hint = skb;
2534 tp->lost_cnt_hint = cnt;
2535
2536 if (after(TCP_SKB_CB(skb)->end_seq, tp->high_seq))
2537 break;
2538
2539 oldcnt = cnt;
2540 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2541 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2542 cnt += tcp_skb_pcount(skb);
2543
2544 if (cnt > packets) {
2545 if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2546 (oldcnt >= packets))
2547 break;
2548
2549 mss = skb_shinfo(skb)->gso_size;
2550 err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss);
2551 if (err < 0)
2552 break;
2553 cnt = packets;
2554 }
2555
2556 tcp_skb_mark_lost(tp, skb);
2557
2558 if (mark_head)
2559 break;
2560 }
2561 tcp_verify_left_out(tp);
2562}
2563
2564/* Account newly detected lost packet(s) */
2565
2566static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2567{
2568 struct tcp_sock *tp = tcp_sk(sk);
2569
2570 if (tcp_is_reno(tp)) {
2571 tcp_mark_head_lost(sk, 1, 1);
2572 } else if (tcp_is_fack(tp)) {
2573 int lost = tp->fackets_out - tp->reordering;
2574 if (lost <= 0)
2575 lost = 1;
2576 tcp_mark_head_lost(sk, lost, 0);
2577 } else {
2578 int sacked_upto = tp->sacked_out - tp->reordering;
2579 if (sacked_upto >= 0)
2580 tcp_mark_head_lost(sk, sacked_upto, 0);
2581 else if (fast_rexmit)
2582 tcp_mark_head_lost(sk, 1, 1);
2583 }
2584
2585 tcp_timeout_skbs(sk);
2586}
2587
2588/* CWND moderation, preventing bursts due to too big ACKs
2589 * in dubious situations.
2590 */
2591static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2592{
2593 tp->snd_cwnd = min(tp->snd_cwnd,
2594 tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2595 tp->snd_cwnd_stamp = tcp_time_stamp;
2596}
2597
2598/* Lower bound on congestion window is slow start threshold
2599 * unless congestion avoidance choice decides to overide it.
2600 */
2601static inline u32 tcp_cwnd_min(const struct sock *sk)
2602{
2603 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
2604
2605 return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
2606}
2607
2608/* Decrease cwnd each second ack. */
2609static void tcp_cwnd_down(struct sock *sk, int flag)
2610{
2611 struct tcp_sock *tp = tcp_sk(sk);
2612 int decr = tp->snd_cwnd_cnt + 1;
2613
2614 if ((flag & (FLAG_ANY_PROGRESS | FLAG_DSACKING_ACK)) ||
2615 (tcp_is_reno(tp) && !(flag & FLAG_NOT_DUP))) {
2616 tp->snd_cwnd_cnt = decr & 1;
2617 decr >>= 1;
2618
2619 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
2620 tp->snd_cwnd -= decr;
2621
2622 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2623 tp->snd_cwnd_stamp = tcp_time_stamp;
2624 }
2625}
2626
2627/* Nothing was retransmitted or returned timestamp is less
2628 * than timestamp of the first retransmission.
2629 */
2630static inline int tcp_packet_delayed(struct tcp_sock *tp)
2631{
2632 return !tp->retrans_stamp ||
2633 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2634 before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
2635}
2636
2637/* Undo procedures. */
2638
2639#if FASTRETRANS_DEBUG > 1
2640static void DBGUNDO(struct sock *sk, const char *msg)
2641{
2642 struct tcp_sock *tp = tcp_sk(sk);
2643 struct inet_sock *inet = inet_sk(sk);
2644
2645 if (sk->sk_family == AF_INET) {
2646 printk(KERN_DEBUG "Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2647 msg,
2648 &inet->inet_daddr, ntohs(inet->inet_dport),
2649 tp->snd_cwnd, tcp_left_out(tp),
2650 tp->snd_ssthresh, tp->prior_ssthresh,
2651 tp->packets_out);
2652 }
2653#if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
2654 else if (sk->sk_family == AF_INET6) {
2655 struct ipv6_pinfo *np = inet6_sk(sk);
2656 printk(KERN_DEBUG "Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2657 msg,
2658 &np->daddr, ntohs(inet->inet_dport),
2659 tp->snd_cwnd, tcp_left_out(tp),
2660 tp->snd_ssthresh, tp->prior_ssthresh,
2661 tp->packets_out);
2662 }
2663#endif
2664}
2665#else
2666#define DBGUNDO(x...) do { } while (0)
2667#endif
2668
2669static void tcp_undo_cwr(struct sock *sk, const bool undo_ssthresh)
2670{
2671 struct tcp_sock *tp = tcp_sk(sk);
2672
2673 if (tp->prior_ssthresh) {
2674 const struct inet_connection_sock *icsk = inet_csk(sk);
2675
2676 if (icsk->icsk_ca_ops->undo_cwnd)
2677 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2678 else
2679 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2680
2681 if (undo_ssthresh && tp->prior_ssthresh > tp->snd_ssthresh) {
2682 tp->snd_ssthresh = tp->prior_ssthresh;
2683 TCP_ECN_withdraw_cwr(tp);
2684 }
2685 } else {
2686 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2687 }
2688 tp->snd_cwnd_stamp = tcp_time_stamp;
2689}
2690
2691static inline int tcp_may_undo(struct tcp_sock *tp)
2692{
2693 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2694}
2695
2696/* People celebrate: "We love our President!" */
2697static int tcp_try_undo_recovery(struct sock *sk)
2698{
2699 struct tcp_sock *tp = tcp_sk(sk);
2700
2701 if (tcp_may_undo(tp)) {
2702 int mib_idx;
2703
2704 /* Happy end! We did not retransmit anything
2705 * or our original transmission succeeded.
2706 */
2707 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2708 tcp_undo_cwr(sk, true);
2709 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2710 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2711 else
2712 mib_idx = LINUX_MIB_TCPFULLUNDO;
2713
2714 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2715 tp->undo_marker = 0;
2716 }
2717 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2718 /* Hold old state until something *above* high_seq
2719 * is ACKed. For Reno it is MUST to prevent false
2720 * fast retransmits (RFC2582). SACK TCP is safe. */
2721 tcp_moderate_cwnd(tp);
2722 return 1;
2723 }
2724 tcp_set_ca_state(sk, TCP_CA_Open);
2725 return 0;
2726}
2727
2728/* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2729static void tcp_try_undo_dsack(struct sock *sk)
2730{
2731 struct tcp_sock *tp = tcp_sk(sk);
2732
2733 if (tp->undo_marker && !tp->undo_retrans) {
2734 DBGUNDO(sk, "D-SACK");
2735 tcp_undo_cwr(sk, true);
2736 tp->undo_marker = 0;
2737 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2738 }
2739}
2740
2741/* We can clear retrans_stamp when there are no retransmissions in the
2742 * window. It would seem that it is trivially available for us in
2743 * tp->retrans_out, however, that kind of assumptions doesn't consider
2744 * what will happen if errors occur when sending retransmission for the
2745 * second time. ...It could the that such segment has only
2746 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2747 * the head skb is enough except for some reneging corner cases that
2748 * are not worth the effort.
2749 *
2750 * Main reason for all this complexity is the fact that connection dying
2751 * time now depends on the validity of the retrans_stamp, in particular,
2752 * that successive retransmissions of a segment must not advance
2753 * retrans_stamp under any conditions.
2754 */
2755static int tcp_any_retrans_done(struct sock *sk)
2756{
2757 struct tcp_sock *tp = tcp_sk(sk);
2758 struct sk_buff *skb;
2759
2760 if (tp->retrans_out)
2761 return 1;
2762
2763 skb = tcp_write_queue_head(sk);
2764 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2765 return 1;
2766
2767 return 0;
2768}
2769
2770/* Undo during fast recovery after partial ACK. */
2771
2772static int tcp_try_undo_partial(struct sock *sk, int acked)
2773{
2774 struct tcp_sock *tp = tcp_sk(sk);
2775 /* Partial ACK arrived. Force Hoe's retransmit. */
2776 int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering);
2777
2778 if (tcp_may_undo(tp)) {
2779 /* Plain luck! Hole if filled with delayed
2780 * packet, rather than with a retransmit.
2781 */
2782 if (!tcp_any_retrans_done(sk))
2783 tp->retrans_stamp = 0;
2784
2785 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2786
2787 DBGUNDO(sk, "Hoe");
2788 tcp_undo_cwr(sk, false);
2789 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2790
2791 /* So... Do not make Hoe's retransmit yet.
2792 * If the first packet was delayed, the rest
2793 * ones are most probably delayed as well.
2794 */
2795 failed = 0;
2796 }
2797 return failed;
2798}
2799
2800/* Undo during loss recovery after partial ACK. */
2801static int tcp_try_undo_loss(struct sock *sk)
2802{
2803 struct tcp_sock *tp = tcp_sk(sk);
2804
2805 if (tcp_may_undo(tp)) {
2806 struct sk_buff *skb;
2807 tcp_for_write_queue(skb, sk) {
2808 if (skb == tcp_send_head(sk))
2809 break;
2810 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2811 }
2812
2813 tcp_clear_all_retrans_hints(tp);
2814
2815 DBGUNDO(sk, "partial loss");
2816 tp->lost_out = 0;
2817 tcp_undo_cwr(sk, true);
2818 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2819 inet_csk(sk)->icsk_retransmits = 0;
2820 tp->undo_marker = 0;
2821 if (tcp_is_sack(tp))
2822 tcp_set_ca_state(sk, TCP_CA_Open);
2823 return 1;
2824 }
2825 return 0;
2826}
2827
2828static inline void tcp_complete_cwr(struct sock *sk)
2829{
2830 struct tcp_sock *tp = tcp_sk(sk);
2831 /* Do not moderate cwnd if it's already undone in cwr or recovery */
2832 if (tp->undo_marker && tp->snd_cwnd > tp->snd_ssthresh) {
2833 tp->snd_cwnd = tp->snd_ssthresh;
2834 tp->snd_cwnd_stamp = tcp_time_stamp;
2835 }
2836 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2837}
2838
2839static void tcp_try_keep_open(struct sock *sk)
2840{
2841 struct tcp_sock *tp = tcp_sk(sk);
2842 int state = TCP_CA_Open;
2843
2844 if (tcp_left_out(tp) || tcp_any_retrans_done(sk) || tp->undo_marker)
2845 state = TCP_CA_Disorder;
2846
2847 if (inet_csk(sk)->icsk_ca_state != state) {
2848 tcp_set_ca_state(sk, state);
2849 tp->high_seq = tp->snd_nxt;
2850 }
2851}
2852
2853static void tcp_try_to_open(struct sock *sk, int flag)
2854{
2855 struct tcp_sock *tp = tcp_sk(sk);
2856
2857 tcp_verify_left_out(tp);
2858
2859 if (!tp->frto_counter && !tcp_any_retrans_done(sk))
2860 tp->retrans_stamp = 0;
2861
2862 if (flag & FLAG_ECE)
2863 tcp_enter_cwr(sk, 1);
2864
2865 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2866 tcp_try_keep_open(sk);
2867 tcp_moderate_cwnd(tp);
2868 } else {
2869 tcp_cwnd_down(sk, flag);
2870 }
2871}
2872
2873static void tcp_mtup_probe_failed(struct sock *sk)
2874{
2875 struct inet_connection_sock *icsk = inet_csk(sk);
2876
2877 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2878 icsk->icsk_mtup.probe_size = 0;
2879}
2880
2881static void tcp_mtup_probe_success(struct sock *sk)
2882{
2883 struct tcp_sock *tp = tcp_sk(sk);
2884 struct inet_connection_sock *icsk = inet_csk(sk);
2885
2886 /* FIXME: breaks with very large cwnd */
2887 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2888 tp->snd_cwnd = tp->snd_cwnd *
2889 tcp_mss_to_mtu(sk, tp->mss_cache) /
2890 icsk->icsk_mtup.probe_size;
2891 tp->snd_cwnd_cnt = 0;
2892 tp->snd_cwnd_stamp = tcp_time_stamp;
2893 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2894
2895 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2896 icsk->icsk_mtup.probe_size = 0;
2897 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2898}
2899
2900/* Do a simple retransmit without using the backoff mechanisms in
2901 * tcp_timer. This is used for path mtu discovery.
2902 * The socket is already locked here.
2903 */
2904void tcp_simple_retransmit(struct sock *sk)
2905{
2906 const struct inet_connection_sock *icsk = inet_csk(sk);
2907 struct tcp_sock *tp = tcp_sk(sk);
2908 struct sk_buff *skb;
2909 unsigned int mss = tcp_current_mss(sk);
2910 u32 prior_lost = tp->lost_out;
2911
2912 tcp_for_write_queue(skb, sk) {
2913 if (skb == tcp_send_head(sk))
2914 break;
2915 if (tcp_skb_seglen(skb) > mss &&
2916 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2917 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2918 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2919 tp->retrans_out -= tcp_skb_pcount(skb);
2920 }
2921 tcp_skb_mark_lost_uncond_verify(tp, skb);
2922 }
2923 }
2924
2925 tcp_clear_retrans_hints_partial(tp);
2926
2927 if (prior_lost == tp->lost_out)
2928 return;
2929
2930 if (tcp_is_reno(tp))
2931 tcp_limit_reno_sacked(tp);
2932
2933 tcp_verify_left_out(tp);
2934
2935 /* Don't muck with the congestion window here.
2936 * Reason is that we do not increase amount of _data_
2937 * in network, but units changed and effective
2938 * cwnd/ssthresh really reduced now.
2939 */
2940 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2941 tp->high_seq = tp->snd_nxt;
2942 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2943 tp->prior_ssthresh = 0;
2944 tp->undo_marker = 0;
2945 tcp_set_ca_state(sk, TCP_CA_Loss);
2946 }
2947 tcp_xmit_retransmit_queue(sk);
2948}
2949EXPORT_SYMBOL(tcp_simple_retransmit);
2950
2951/* Process an event, which can update packets-in-flight not trivially.
2952 * Main goal of this function is to calculate new estimate for left_out,
2953 * taking into account both packets sitting in receiver's buffer and
2954 * packets lost by network.
2955 *
2956 * Besides that it does CWND reduction, when packet loss is detected
2957 * and changes state of machine.
2958 *
2959 * It does _not_ decide what to send, it is made in function
2960 * tcp_xmit_retransmit_queue().
2961 */
2962static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked, int flag)
2963{
2964 struct inet_connection_sock *icsk = inet_csk(sk);
2965 struct tcp_sock *tp = tcp_sk(sk);
2966 int is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
2967 int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2968 (tcp_fackets_out(tp) > tp->reordering));
2969 int fast_rexmit = 0, mib_idx;
2970
2971 if (WARN_ON(!tp->packets_out && tp->sacked_out))
2972 tp->sacked_out = 0;
2973 if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2974 tp->fackets_out = 0;
2975
2976 /* Now state machine starts.
2977 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2978 if (flag & FLAG_ECE)
2979 tp->prior_ssthresh = 0;
2980
2981 /* B. In all the states check for reneging SACKs. */
2982 if (tcp_check_sack_reneging(sk, flag))
2983 return;
2984
2985 /* C. Process data loss notification, provided it is valid. */
2986 if (tcp_is_fack(tp) && (flag & FLAG_DATA_LOST) &&
2987 before(tp->snd_una, tp->high_seq) &&
2988 icsk->icsk_ca_state != TCP_CA_Open &&
2989 tp->fackets_out > tp->reordering) {
2990 tcp_mark_head_lost(sk, tp->fackets_out - tp->reordering, 0);
2991 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSS);
2992 }
2993
2994 /* D. Check consistency of the current state. */
2995 tcp_verify_left_out(tp);
2996
2997 /* E. Check state exit conditions. State can be terminated
2998 * when high_seq is ACKed. */
2999 if (icsk->icsk_ca_state == TCP_CA_Open) {
3000 WARN_ON(tp->retrans_out != 0);
3001 tp->retrans_stamp = 0;
3002 } else if (!before(tp->snd_una, tp->high_seq)) {
3003 switch (icsk->icsk_ca_state) {
3004 case TCP_CA_Loss:
3005 icsk->icsk_retransmits = 0;
3006 if (tcp_try_undo_recovery(sk))
3007 return;
3008 break;
3009
3010 case TCP_CA_CWR:
3011 /* CWR is to be held something *above* high_seq
3012 * is ACKed for CWR bit to reach receiver. */
3013 if (tp->snd_una != tp->high_seq) {
3014 tcp_complete_cwr(sk);
3015 tcp_set_ca_state(sk, TCP_CA_Open);
3016 }
3017 break;
3018
3019 case TCP_CA_Disorder:
3020 tcp_try_undo_dsack(sk);
3021 if (!tp->undo_marker ||
3022 /* For SACK case do not Open to allow to undo
3023 * catching for all duplicate ACKs. */
3024 tcp_is_reno(tp) || tp->snd_una != tp->high_seq) {
3025 tp->undo_marker = 0;
3026 tcp_set_ca_state(sk, TCP_CA_Open);
3027 }
3028 break;
3029
3030 case TCP_CA_Recovery:
3031 if (tcp_is_reno(tp))
3032 tcp_reset_reno_sack(tp);
3033 if (tcp_try_undo_recovery(sk))
3034 return;
3035 tcp_complete_cwr(sk);
3036 break;
3037 }
3038 }
3039
3040 /* F. Process state. */
3041 switch (icsk->icsk_ca_state) {
3042 case TCP_CA_Recovery:
3043 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
3044 if (tcp_is_reno(tp) && is_dupack)
3045 tcp_add_reno_sack(sk);
3046 } else
3047 do_lost = tcp_try_undo_partial(sk, pkts_acked);
3048 break;
3049 case TCP_CA_Loss:
3050 if (flag & FLAG_DATA_ACKED)
3051 icsk->icsk_retransmits = 0;
3052 if (tcp_is_reno(tp) && flag & FLAG_SND_UNA_ADVANCED)
3053 tcp_reset_reno_sack(tp);
3054 if (!tcp_try_undo_loss(sk)) {
3055 tcp_moderate_cwnd(tp);
3056 tcp_xmit_retransmit_queue(sk);
3057 return;
3058 }
3059 if (icsk->icsk_ca_state != TCP_CA_Open)
3060 return;
3061 /* Loss is undone; fall through to processing in Open state. */
3062 default:
3063 if (tcp_is_reno(tp)) {
3064 if (flag & FLAG_SND_UNA_ADVANCED)
3065 tcp_reset_reno_sack(tp);
3066 if (is_dupack)
3067 tcp_add_reno_sack(sk);
3068 }
3069
3070 if (icsk->icsk_ca_state == TCP_CA_Disorder)
3071 tcp_try_undo_dsack(sk);
3072
3073 if (!tcp_time_to_recover(sk)) {
3074 tcp_try_to_open(sk, flag);
3075 return;
3076 }
3077
3078 /* MTU probe failure: don't reduce cwnd */
3079 if (icsk->icsk_ca_state < TCP_CA_CWR &&
3080 icsk->icsk_mtup.probe_size &&
3081 tp->snd_una == tp->mtu_probe.probe_seq_start) {
3082 tcp_mtup_probe_failed(sk);
3083 /* Restores the reduction we did in tcp_mtup_probe() */
3084 tp->snd_cwnd++;
3085 tcp_simple_retransmit(sk);
3086 return;
3087 }
3088
3089 /* Otherwise enter Recovery state */
3090
3091 if (tcp_is_reno(tp))
3092 mib_idx = LINUX_MIB_TCPRENORECOVERY;
3093 else
3094 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
3095
3096 NET_INC_STATS_BH(sock_net(sk), mib_idx);
3097
3098 tp->high_seq = tp->snd_nxt;
3099 tp->prior_ssthresh = 0;
3100 tp->undo_marker = tp->snd_una;
3101 tp->undo_retrans = tp->retrans_out;
3102
3103 if (icsk->icsk_ca_state < TCP_CA_CWR) {
3104 if (!(flag & FLAG_ECE))
3105 tp->prior_ssthresh = tcp_current_ssthresh(sk);
3106 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
3107 TCP_ECN_queue_cwr(tp);
3108 }
3109
3110 tp->bytes_acked = 0;
3111 tp->snd_cwnd_cnt = 0;
3112 tcp_set_ca_state(sk, TCP_CA_Recovery);
3113 fast_rexmit = 1;
3114 }
3115
3116 if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk)))
3117 tcp_update_scoreboard(sk, fast_rexmit);
3118 tcp_cwnd_down(sk, flag);
3119 tcp_xmit_retransmit_queue(sk);
3120}
3121
3122void tcp_valid_rtt_meas(struct sock *sk, u32 seq_rtt)
3123{
3124 tcp_rtt_estimator(sk, seq_rtt);
3125 tcp_set_rto(sk);
3126 inet_csk(sk)->icsk_backoff = 0;
3127}
3128EXPORT_SYMBOL(tcp_valid_rtt_meas);
3129
3130/* Read draft-ietf-tcplw-high-performance before mucking
3131 * with this code. (Supersedes RFC1323)
3132 */
3133static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
3134{
3135 /* RTTM Rule: A TSecr value received in a segment is used to
3136 * update the averaged RTT measurement only if the segment
3137 * acknowledges some new data, i.e., only if it advances the
3138 * left edge of the send window.
3139 *
3140 * See draft-ietf-tcplw-high-performance-00, section 3.3.
3141 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
3142 *
3143 * Changed: reset backoff as soon as we see the first valid sample.
3144 * If we do not, we get strongly overestimated rto. With timestamps
3145 * samples are accepted even from very old segments: f.e., when rtt=1
3146 * increases to 8, we retransmit 5 times and after 8 seconds delayed
3147 * answer arrives rto becomes 120 seconds! If at least one of segments
3148 * in window is lost... Voila. --ANK (010210)
3149 */
3150 struct tcp_sock *tp = tcp_sk(sk);
3151
3152 tcp_valid_rtt_meas(sk, tcp_time_stamp - tp->rx_opt.rcv_tsecr);
3153}
3154
3155static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
3156{
3157 /* We don't have a timestamp. Can only use
3158 * packets that are not retransmitted to determine
3159 * rtt estimates. Also, we must not reset the
3160 * backoff for rto until we get a non-retransmitted
3161 * packet. This allows us to deal with a situation
3162 * where the network delay has increased suddenly.
3163 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
3164 */
3165
3166 if (flag & FLAG_RETRANS_DATA_ACKED)
3167 return;
3168
3169 tcp_valid_rtt_meas(sk, seq_rtt);
3170}
3171
3172static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
3173 const s32 seq_rtt)
3174{
3175 const struct tcp_sock *tp = tcp_sk(sk);
3176 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
3177 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3178 tcp_ack_saw_tstamp(sk, flag);
3179 else if (seq_rtt >= 0)
3180 tcp_ack_no_tstamp(sk, seq_rtt, flag);
3181}
3182
3183static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
3184{
3185 const struct inet_connection_sock *icsk = inet_csk(sk);
3186 icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight);
3187 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
3188}
3189
3190/* Restart timer after forward progress on connection.
3191 * RFC2988 recommends to restart timer to now+rto.
3192 */
3193static void tcp_rearm_rto(struct sock *sk)
3194{
3195 struct tcp_sock *tp = tcp_sk(sk);
3196
3197 if (!tp->packets_out) {
3198 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3199 } else {
3200 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3201 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3202 }
3203}
3204
3205/* If we get here, the whole TSO packet has not been acked. */
3206static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3207{
3208 struct tcp_sock *tp = tcp_sk(sk);
3209 u32 packets_acked;
3210
3211 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3212
3213 packets_acked = tcp_skb_pcount(skb);
3214 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3215 return 0;
3216 packets_acked -= tcp_skb_pcount(skb);
3217
3218 if (packets_acked) {
3219 BUG_ON(tcp_skb_pcount(skb) == 0);
3220 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3221 }
3222
3223 return packets_acked;
3224}
3225
3226/* Remove acknowledged frames from the retransmission queue. If our packet
3227 * is before the ack sequence we can discard it as it's confirmed to have
3228 * arrived at the other end.
3229 */
3230static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3231 u32 prior_snd_una)
3232{
3233 struct tcp_sock *tp = tcp_sk(sk);
3234 const struct inet_connection_sock *icsk = inet_csk(sk);
3235 struct sk_buff *skb;
3236 u32 now = tcp_time_stamp;
3237 int fully_acked = 1;
3238 int flag = 0;
3239 u32 pkts_acked = 0;
3240 u32 reord = tp->packets_out;
3241 u32 prior_sacked = tp->sacked_out;
3242 s32 seq_rtt = -1;
3243 s32 ca_seq_rtt = -1;
3244 ktime_t last_ackt = net_invalid_timestamp();
3245
3246 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3247 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3248 u32 acked_pcount;
3249 u8 sacked = scb->sacked;
3250
3251 /* Determine how many packets and what bytes were acked, tso and else */
3252 if (after(scb->end_seq, tp->snd_una)) {
3253 if (tcp_skb_pcount(skb) == 1 ||
3254 !after(tp->snd_una, scb->seq))
3255 break;
3256
3257 acked_pcount = tcp_tso_acked(sk, skb);
3258 if (!acked_pcount)
3259 break;
3260
3261 fully_acked = 0;
3262 } else {
3263 acked_pcount = tcp_skb_pcount(skb);
3264 }
3265
3266 if (sacked & TCPCB_RETRANS) {
3267 if (sacked & TCPCB_SACKED_RETRANS)
3268 tp->retrans_out -= acked_pcount;
3269 flag |= FLAG_RETRANS_DATA_ACKED;
3270 ca_seq_rtt = -1;
3271 seq_rtt = -1;
3272 if ((flag & FLAG_DATA_ACKED) || (acked_pcount > 1))
3273 flag |= FLAG_NONHEAD_RETRANS_ACKED;
3274 } else {
3275 ca_seq_rtt = now - scb->when;
3276 last_ackt = skb->tstamp;
3277 if (seq_rtt < 0) {
3278 seq_rtt = ca_seq_rtt;
3279 }
3280 if (!(sacked & TCPCB_SACKED_ACKED))
3281 reord = min(pkts_acked, reord);
3282 }
3283
3284 if (sacked & TCPCB_SACKED_ACKED)
3285 tp->sacked_out -= acked_pcount;
3286 if (sacked & TCPCB_LOST)
3287 tp->lost_out -= acked_pcount;
3288
3289 tp->packets_out -= acked_pcount;
3290 pkts_acked += acked_pcount;
3291
3292 /* Initial outgoing SYN's get put onto the write_queue
3293 * just like anything else we transmit. It is not
3294 * true data, and if we misinform our callers that
3295 * this ACK acks real data, we will erroneously exit
3296 * connection startup slow start one packet too
3297 * quickly. This is severely frowned upon behavior.
3298 */
3299 if (!(scb->flags & TCPHDR_SYN)) {
3300 flag |= FLAG_DATA_ACKED;
3301 } else {
3302 flag |= FLAG_SYN_ACKED;
3303 tp->retrans_stamp = 0;
3304 }
3305
3306 if (!fully_acked)
3307 break;
3308
3309 tcp_unlink_write_queue(skb, sk);
3310 sk_wmem_free_skb(sk, skb);
3311 tp->scoreboard_skb_hint = NULL;
3312 if (skb == tp->retransmit_skb_hint)
3313 tp->retransmit_skb_hint = NULL;
3314 if (skb == tp->lost_skb_hint)
3315 tp->lost_skb_hint = NULL;
3316 }
3317
3318 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3319 tp->snd_up = tp->snd_una;
3320
3321 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3322 flag |= FLAG_SACK_RENEGING;
3323
3324 if (flag & FLAG_ACKED) {
3325 const struct tcp_congestion_ops *ca_ops
3326 = inet_csk(sk)->icsk_ca_ops;
3327
3328 if (unlikely(icsk->icsk_mtup.probe_size &&
3329 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3330 tcp_mtup_probe_success(sk);
3331 }
3332
3333 tcp_ack_update_rtt(sk, flag, seq_rtt);
3334 tcp_rearm_rto(sk);
3335
3336 if (tcp_is_reno(tp)) {
3337 tcp_remove_reno_sacks(sk, pkts_acked);
3338 } else {
3339 int delta;
3340
3341 /* Non-retransmitted hole got filled? That's reordering */
3342 if (reord < prior_fackets)
3343 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3344
3345 delta = tcp_is_fack(tp) ? pkts_acked :
3346 prior_sacked - tp->sacked_out;
3347 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3348 }
3349
3350 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3351
3352 if (ca_ops->pkts_acked) {
3353 s32 rtt_us = -1;
3354
3355 /* Is the ACK triggering packet unambiguous? */
3356 if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
3357 /* High resolution needed and available? */
3358 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
3359 !ktime_equal(last_ackt,
3360 net_invalid_timestamp()))
3361 rtt_us = ktime_us_delta(ktime_get_real(),
3362 last_ackt);
3363 else if (ca_seq_rtt >= 0)
3364 rtt_us = jiffies_to_usecs(ca_seq_rtt);
3365 }
3366
3367 ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
3368 }
3369 }
3370
3371#if FASTRETRANS_DEBUG > 0
3372 WARN_ON((int)tp->sacked_out < 0);
3373 WARN_ON((int)tp->lost_out < 0);
3374 WARN_ON((int)tp->retrans_out < 0);
3375 if (!tp->packets_out && tcp_is_sack(tp)) {
3376 icsk = inet_csk(sk);
3377 if (tp->lost_out) {
3378 printk(KERN_DEBUG "Leak l=%u %d\n",
3379 tp->lost_out, icsk->icsk_ca_state);
3380 tp->lost_out = 0;
3381 }
3382 if (tp->sacked_out) {
3383 printk(KERN_DEBUG "Leak s=%u %d\n",
3384 tp->sacked_out, icsk->icsk_ca_state);
3385 tp->sacked_out = 0;
3386 }
3387 if (tp->retrans_out) {
3388 printk(KERN_DEBUG "Leak r=%u %d\n",
3389 tp->retrans_out, icsk->icsk_ca_state);
3390 tp->retrans_out = 0;
3391 }
3392 }
3393#endif
3394 return flag;
3395}
3396
3397static void tcp_ack_probe(struct sock *sk)
3398{
3399 const struct tcp_sock *tp = tcp_sk(sk);
3400 struct inet_connection_sock *icsk = inet_csk(sk);
3401
3402 /* Was it a usable window open? */
3403
3404 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3405 icsk->icsk_backoff = 0;
3406 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3407 /* Socket must be waked up by subsequent tcp_data_snd_check().
3408 * This function is not for random using!
3409 */
3410 } else {
3411 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3412 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3413 TCP_RTO_MAX);
3414 }
3415}
3416
3417static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
3418{
3419 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3420 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3421}
3422
3423static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3424{
3425 const struct tcp_sock *tp = tcp_sk(sk);
3426 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
3427 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
3428}
3429
3430/* Check that window update is acceptable.
3431 * The function assumes that snd_una<=ack<=snd_next.
3432 */
3433static inline int tcp_may_update_window(const struct tcp_sock *tp,
3434 const u32 ack, const u32 ack_seq,
3435 const u32 nwin)
3436{
3437 return after(ack, tp->snd_una) ||
3438 after(ack_seq, tp->snd_wl1) ||
3439 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3440}
3441
3442/* Update our send window.
3443 *
3444 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3445 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3446 */
3447static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack,
3448 u32 ack_seq)
3449{
3450 struct tcp_sock *tp = tcp_sk(sk);
3451 int flag = 0;
3452 u32 nwin = ntohs(tcp_hdr(skb)->window);
3453
3454 if (likely(!tcp_hdr(skb)->syn))
3455 nwin <<= tp->rx_opt.snd_wscale;
3456
3457 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3458 flag |= FLAG_WIN_UPDATE;
3459 tcp_update_wl(tp, ack_seq);
3460
3461 if (tp->snd_wnd != nwin) {
3462 tp->snd_wnd = nwin;
3463
3464 /* Note, it is the only place, where
3465 * fast path is recovered for sending TCP.
3466 */
3467 tp->pred_flags = 0;
3468 tcp_fast_path_check(sk);
3469
3470 if (nwin > tp->max_window) {
3471 tp->max_window = nwin;
3472 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3473 }
3474 }
3475 }
3476
3477 tp->snd_una = ack;
3478
3479 return flag;
3480}
3481
3482/* A very conservative spurious RTO response algorithm: reduce cwnd and
3483 * continue in congestion avoidance.
3484 */
3485static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
3486{
3487 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
3488 tp->snd_cwnd_cnt = 0;
3489 tp->bytes_acked = 0;
3490 TCP_ECN_queue_cwr(tp);
3491 tcp_moderate_cwnd(tp);
3492}
3493
3494/* A conservative spurious RTO response algorithm: reduce cwnd using
3495 * rate halving and continue in congestion avoidance.
3496 */
3497static void tcp_ratehalving_spur_to_response(struct sock *sk)
3498{
3499 tcp_enter_cwr(sk, 0);
3500}
3501
3502static void tcp_undo_spur_to_response(struct sock *sk, int flag)
3503{
3504 if (flag & FLAG_ECE)
3505 tcp_ratehalving_spur_to_response(sk);
3506 else
3507 tcp_undo_cwr(sk, true);
3508}
3509
3510/* F-RTO spurious RTO detection algorithm (RFC4138)
3511 *
3512 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3513 * comments). State (ACK number) is kept in frto_counter. When ACK advances
3514 * window (but not to or beyond highest sequence sent before RTO):
3515 * On First ACK, send two new segments out.
3516 * On Second ACK, RTO was likely spurious. Do spurious response (response
3517 * algorithm is not part of the F-RTO detection algorithm
3518 * given in RFC4138 but can be selected separately).
3519 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
3520 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3521 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3522 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
3523 *
3524 * Rationale: if the RTO was spurious, new ACKs should arrive from the
3525 * original window even after we transmit two new data segments.
3526 *
3527 * SACK version:
3528 * on first step, wait until first cumulative ACK arrives, then move to
3529 * the second step. In second step, the next ACK decides.
3530 *
3531 * F-RTO is implemented (mainly) in four functions:
3532 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
3533 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3534 * called when tcp_use_frto() showed green light
3535 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3536 * - tcp_enter_frto_loss() is called if there is not enough evidence
3537 * to prove that the RTO is indeed spurious. It transfers the control
3538 * from F-RTO to the conventional RTO recovery
3539 */
3540static int tcp_process_frto(struct sock *sk, int flag)
3541{
3542 struct tcp_sock *tp = tcp_sk(sk);
3543
3544 tcp_verify_left_out(tp);
3545
3546 /* Duplicate the behavior from Loss state (fastretrans_alert) */
3547 if (flag & FLAG_DATA_ACKED)
3548 inet_csk(sk)->icsk_retransmits = 0;
3549
3550 if ((flag & FLAG_NONHEAD_RETRANS_ACKED) ||
3551 ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED)))
3552 tp->undo_marker = 0;
3553
3554 if (!before(tp->snd_una, tp->frto_highmark)) {
3555 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
3556 return 1;
3557 }
3558
3559 if (!tcp_is_sackfrto(tp)) {
3560 /* RFC4138 shortcoming in step 2; should also have case c):
3561 * ACK isn't duplicate nor advances window, e.g., opposite dir
3562 * data, winupdate
3563 */
3564 if (!(flag & FLAG_ANY_PROGRESS) && (flag & FLAG_NOT_DUP))
3565 return 1;
3566
3567 if (!(flag & FLAG_DATA_ACKED)) {
3568 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
3569 flag);
3570 return 1;
3571 }
3572 } else {
3573 if (!(flag & FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
3574 /* Prevent sending of new data. */
3575 tp->snd_cwnd = min(tp->snd_cwnd,
3576 tcp_packets_in_flight(tp));
3577 return 1;
3578 }
3579
3580 if ((tp->frto_counter >= 2) &&
3581 (!(flag & FLAG_FORWARD_PROGRESS) ||
3582 ((flag & FLAG_DATA_SACKED) &&
3583 !(flag & FLAG_ONLY_ORIG_SACKED)))) {
3584 /* RFC4138 shortcoming (see comment above) */
3585 if (!(flag & FLAG_FORWARD_PROGRESS) &&
3586 (flag & FLAG_NOT_DUP))
3587 return 1;
3588
3589 tcp_enter_frto_loss(sk, 3, flag);
3590 return 1;
3591 }
3592 }
3593
3594 if (tp->frto_counter == 1) {
3595 /* tcp_may_send_now needs to see updated state */
3596 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
3597 tp->frto_counter = 2;
3598
3599 if (!tcp_may_send_now(sk))
3600 tcp_enter_frto_loss(sk, 2, flag);
3601
3602 return 1;
3603 } else {
3604 switch (sysctl_tcp_frto_response) {
3605 case 2:
3606 tcp_undo_spur_to_response(sk, flag);
3607 break;
3608 case 1:
3609 tcp_conservative_spur_to_response(tp);
3610 break;
3611 default:
3612 tcp_ratehalving_spur_to_response(sk);
3613 break;
3614 }
3615 tp->frto_counter = 0;
3616 tp->undo_marker = 0;
3617 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSPURIOUSRTOS);
3618 }
3619 return 0;
3620}
3621
3622/* This routine deals with incoming acks, but not outgoing ones. */
3623static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
3624{
3625 struct inet_connection_sock *icsk = inet_csk(sk);
3626 struct tcp_sock *tp = tcp_sk(sk);
3627 u32 prior_snd_una = tp->snd_una;
3628 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3629 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3630 u32 prior_in_flight;
3631 u32 prior_fackets;
3632 int prior_packets;
3633 int frto_cwnd = 0;
3634
3635 /* If the ack is older than previous acks
3636 * then we can probably ignore it.
3637 */
3638 if (before(ack, prior_snd_una))
3639 goto old_ack;
3640
3641 /* If the ack includes data we haven't sent yet, discard
3642 * this segment (RFC793 Section 3.9).
3643 */
3644 if (after(ack, tp->snd_nxt))
3645 goto invalid_ack;
3646
3647 if (after(ack, prior_snd_una))
3648 flag |= FLAG_SND_UNA_ADVANCED;
3649
3650 if (sysctl_tcp_abc) {
3651 if (icsk->icsk_ca_state < TCP_CA_CWR)
3652 tp->bytes_acked += ack - prior_snd_una;
3653 else if (icsk->icsk_ca_state == TCP_CA_Loss)
3654 /* we assume just one segment left network */
3655 tp->bytes_acked += min(ack - prior_snd_una,
3656 tp->mss_cache);
3657 }
3658
3659 prior_fackets = tp->fackets_out;
3660 prior_in_flight = tcp_packets_in_flight(tp);
3661
3662 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3663 /* Window is constant, pure forward advance.
3664 * No more checks are required.
3665 * Note, we use the fact that SND.UNA>=SND.WL2.
3666 */
3667 tcp_update_wl(tp, ack_seq);
3668 tp->snd_una = ack;
3669 flag |= FLAG_WIN_UPDATE;
3670
3671 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
3672
3673 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3674 } else {
3675 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3676 flag |= FLAG_DATA;
3677 else
3678 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3679
3680 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3681
3682 if (TCP_SKB_CB(skb)->sacked)
3683 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3684
3685 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
3686 flag |= FLAG_ECE;
3687
3688 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
3689 }
3690
3691 /* We passed data and got it acked, remove any soft error
3692 * log. Something worked...
3693 */
3694 sk->sk_err_soft = 0;
3695 icsk->icsk_probes_out = 0;
3696 tp->rcv_tstamp = tcp_time_stamp;
3697 prior_packets = tp->packets_out;
3698 if (!prior_packets)
3699 goto no_queue;
3700
3701 /* See if we can take anything off of the retransmit queue. */
3702 flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una);
3703
3704 if (tp->frto_counter)
3705 frto_cwnd = tcp_process_frto(sk, flag);
3706 /* Guarantee sacktag reordering detection against wrap-arounds */
3707 if (before(tp->frto_highmark, tp->snd_una))
3708 tp->frto_highmark = 0;
3709
3710 if (tcp_ack_is_dubious(sk, flag)) {
3711 /* Advance CWND, if state allows this. */
3712 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
3713 tcp_may_raise_cwnd(sk, flag))
3714 tcp_cong_avoid(sk, ack, prior_in_flight);
3715 tcp_fastretrans_alert(sk, prior_packets - tp->packets_out,
3716 flag);
3717 } else {
3718 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
3719 tcp_cong_avoid(sk, ack, prior_in_flight);
3720 }
3721
3722 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3723 dst_confirm(__sk_dst_get(sk));
3724
3725 return 1;
3726
3727no_queue:
3728 /* If this ack opens up a zero window, clear backoff. It was
3729 * being used to time the probes, and is probably far higher than
3730 * it needs to be for normal retransmission.
3731 */
3732 if (tcp_send_head(sk))
3733 tcp_ack_probe(sk);
3734 return 1;
3735
3736invalid_ack:
3737 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3738 return -1;
3739
3740old_ack:
3741 if (TCP_SKB_CB(skb)->sacked) {
3742 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3743 if (icsk->icsk_ca_state == TCP_CA_Open)
3744 tcp_try_keep_open(sk);
3745 }
3746
3747 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3748 return 0;
3749}
3750
3751/* Look for tcp options. Normally only called on SYN and SYNACK packets.
3752 * But, this can also be called on packets in the established flow when
3753 * the fast version below fails.
3754 */
3755void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx,
3756 u8 **hvpp, int estab)
3757{
3758 unsigned char *ptr;
3759 struct tcphdr *th = tcp_hdr(skb);
3760 int length = (th->doff * 4) - sizeof(struct tcphdr);
3761
3762 ptr = (unsigned char *)(th + 1);
3763 opt_rx->saw_tstamp = 0;
3764
3765 while (length > 0) {
3766 int opcode = *ptr++;
3767 int opsize;
3768
3769 switch (opcode) {
3770 case TCPOPT_EOL:
3771 return;
3772 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3773 length--;
3774 continue;
3775 default:
3776 opsize = *ptr++;
3777 if (opsize < 2) /* "silly options" */
3778 return;
3779 if (opsize > length)
3780 return; /* don't parse partial options */
3781 switch (opcode) {
3782 case TCPOPT_MSS:
3783 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3784 u16 in_mss = get_unaligned_be16(ptr);
3785 if (in_mss) {
3786 if (opt_rx->user_mss &&
3787 opt_rx->user_mss < in_mss)
3788 in_mss = opt_rx->user_mss;
3789 opt_rx->mss_clamp = in_mss;
3790 }
3791 }
3792 break;
3793 case TCPOPT_WINDOW:
3794 if (opsize == TCPOLEN_WINDOW && th->syn &&
3795 !estab && sysctl_tcp_window_scaling) {
3796 __u8 snd_wscale = *(__u8 *)ptr;
3797 opt_rx->wscale_ok = 1;
3798 if (snd_wscale > 14) {
3799 if (net_ratelimit())
3800 printk(KERN_INFO "tcp_parse_options: Illegal window "
3801 "scaling value %d >14 received.\n",
3802 snd_wscale);
3803 snd_wscale = 14;
3804 }
3805 opt_rx->snd_wscale = snd_wscale;
3806 }
3807 break;
3808 case TCPOPT_TIMESTAMP:
3809 if ((opsize == TCPOLEN_TIMESTAMP) &&
3810 ((estab && opt_rx->tstamp_ok) ||
3811 (!estab && sysctl_tcp_timestamps))) {
3812 opt_rx->saw_tstamp = 1;
3813 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3814 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3815 }
3816 break;
3817 case TCPOPT_SACK_PERM:
3818 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3819 !estab && sysctl_tcp_sack) {
3820 opt_rx->sack_ok = 1;
3821 tcp_sack_reset(opt_rx);
3822 }
3823 break;
3824
3825 case TCPOPT_SACK:
3826 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3827 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3828 opt_rx->sack_ok) {
3829 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3830 }
3831 break;
3832#ifdef CONFIG_TCP_MD5SIG
3833 case TCPOPT_MD5SIG:
3834 /*
3835 * The MD5 Hash has already been
3836 * checked (see tcp_v{4,6}_do_rcv()).
3837 */
3838 break;
3839#endif
3840 case TCPOPT_COOKIE:
3841 /* This option is variable length.
3842 */
3843 switch (opsize) {
3844 case TCPOLEN_COOKIE_BASE:
3845 /* not yet implemented */
3846 break;
3847 case TCPOLEN_COOKIE_PAIR:
3848 /* not yet implemented */
3849 break;
3850 case TCPOLEN_COOKIE_MIN+0:
3851 case TCPOLEN_COOKIE_MIN+2:
3852 case TCPOLEN_COOKIE_MIN+4:
3853 case TCPOLEN_COOKIE_MIN+6:
3854 case TCPOLEN_COOKIE_MAX:
3855 /* 16-bit multiple */
3856 opt_rx->cookie_plus = opsize;
3857 *hvpp = ptr;
3858 break;
3859 default:
3860 /* ignore option */
3861 break;
3862 }
3863 break;
3864 }
3865
3866 ptr += opsize-2;
3867 length -= opsize;
3868 }
3869 }
3870}
3871EXPORT_SYMBOL(tcp_parse_options);
3872
3873static int tcp_parse_aligned_timestamp(struct tcp_sock *tp, struct tcphdr *th)
3874{
3875 __be32 *ptr = (__be32 *)(th + 1);
3876
3877 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3878 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3879 tp->rx_opt.saw_tstamp = 1;
3880 ++ptr;
3881 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3882 ++ptr;
3883 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3884 return 1;
3885 }
3886 return 0;
3887}
3888
3889/* Fast parse options. This hopes to only see timestamps.
3890 * If it is wrong it falls back on tcp_parse_options().
3891 */
3892static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
3893 struct tcp_sock *tp, u8 **hvpp)
3894{
3895 /* In the spirit of fast parsing, compare doff directly to constant
3896 * values. Because equality is used, short doff can be ignored here.
3897 */
3898 if (th->doff == (sizeof(*th) / 4)) {
3899 tp->rx_opt.saw_tstamp = 0;
3900 return 0;
3901 } else if (tp->rx_opt.tstamp_ok &&
3902 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3903 if (tcp_parse_aligned_timestamp(tp, th))
3904 return 1;
3905 }
3906 tcp_parse_options(skb, &tp->rx_opt, hvpp, 1);
3907 return 1;
3908}
3909
3910#ifdef CONFIG_TCP_MD5SIG
3911/*
3912 * Parse MD5 Signature option
3913 */
3914u8 *tcp_parse_md5sig_option(struct tcphdr *th)
3915{
3916 int length = (th->doff << 2) - sizeof (*th);
3917 u8 *ptr = (u8*)(th + 1);
3918
3919 /* If the TCP option is too short, we can short cut */
3920 if (length < TCPOLEN_MD5SIG)
3921 return NULL;
3922
3923 while (length > 0) {
3924 int opcode = *ptr++;
3925 int opsize;
3926
3927 switch(opcode) {
3928 case TCPOPT_EOL:
3929 return NULL;
3930 case TCPOPT_NOP:
3931 length--;
3932 continue;
3933 default:
3934 opsize = *ptr++;
3935 if (opsize < 2 || opsize > length)
3936 return NULL;
3937 if (opcode == TCPOPT_MD5SIG)
3938 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3939 }
3940 ptr += opsize - 2;
3941 length -= opsize;
3942 }
3943 return NULL;
3944}
3945EXPORT_SYMBOL(tcp_parse_md5sig_option);
3946#endif
3947
3948static inline void tcp_store_ts_recent(struct tcp_sock *tp)
3949{
3950 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3951 tp->rx_opt.ts_recent_stamp = get_seconds();
3952}
3953
3954static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3955{
3956 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3957 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3958 * extra check below makes sure this can only happen
3959 * for pure ACK frames. -DaveM
3960 *
3961 * Not only, also it occurs for expired timestamps.
3962 */
3963
3964 if (tcp_paws_check(&tp->rx_opt, 0))
3965 tcp_store_ts_recent(tp);
3966 }
3967}
3968
3969/* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3970 *
3971 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3972 * it can pass through stack. So, the following predicate verifies that
3973 * this segment is not used for anything but congestion avoidance or
3974 * fast retransmit. Moreover, we even are able to eliminate most of such
3975 * second order effects, if we apply some small "replay" window (~RTO)
3976 * to timestamp space.
3977 *
3978 * All these measures still do not guarantee that we reject wrapped ACKs
3979 * on networks with high bandwidth, when sequence space is recycled fastly,
3980 * but it guarantees that such events will be very rare and do not affect
3981 * connection seriously. This doesn't look nice, but alas, PAWS is really
3982 * buggy extension.
3983 *
3984 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3985 * states that events when retransmit arrives after original data are rare.
3986 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3987 * the biggest problem on large power networks even with minor reordering.
3988 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3989 * up to bandwidth of 18Gigabit/sec. 8) ]
3990 */
3991
3992static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3993{
3994 struct tcp_sock *tp = tcp_sk(sk);
3995 struct tcphdr *th = tcp_hdr(skb);
3996 u32 seq = TCP_SKB_CB(skb)->seq;
3997 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3998
3999 return (/* 1. Pure ACK with correct sequence number. */
4000 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
4001
4002 /* 2. ... and duplicate ACK. */
4003 ack == tp->snd_una &&
4004
4005 /* 3. ... and does not update window. */
4006 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
4007
4008 /* 4. ... and sits in replay window. */
4009 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
4010}
4011
4012static inline int tcp_paws_discard(const struct sock *sk,
4013 const struct sk_buff *skb)
4014{
4015 const struct tcp_sock *tp = tcp_sk(sk);
4016
4017 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4018 !tcp_disordered_ack(sk, skb);
4019}
4020
4021/* Check segment sequence number for validity.
4022 *
4023 * Segment controls are considered valid, if the segment
4024 * fits to the window after truncation to the window. Acceptability
4025 * of data (and SYN, FIN, of course) is checked separately.
4026 * See tcp_data_queue(), for example.
4027 *
4028 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4029 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4030 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4031 * (borrowed from freebsd)
4032 */
4033
4034static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
4035{
4036 return !before(end_seq, tp->rcv_wup) &&
4037 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
4038}
4039
4040/* When we get a reset we do this. */
4041static void tcp_reset(struct sock *sk)
4042{
4043 /* We want the right error as BSD sees it (and indeed as we do). */
4044 switch (sk->sk_state) {
4045 case TCP_SYN_SENT:
4046 sk->sk_err = ECONNREFUSED;
4047 break;
4048 case TCP_CLOSE_WAIT:
4049 sk->sk_err = EPIPE;
4050 break;
4051 case TCP_CLOSE:
4052 return;
4053 default:
4054 sk->sk_err = ECONNRESET;
4055 }
4056 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4057 smp_wmb();
4058
4059 if (!sock_flag(sk, SOCK_DEAD))
4060 sk->sk_error_report(sk);
4061
4062 tcp_done(sk);
4063}
4064
4065/*
4066 * Process the FIN bit. This now behaves as it is supposed to work
4067 * and the FIN takes effect when it is validly part of sequence
4068 * space. Not before when we get holes.
4069 *
4070 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4071 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4072 * TIME-WAIT)
4073 *
4074 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4075 * close and we go into CLOSING (and later onto TIME-WAIT)
4076 *
4077 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4078 */
4079static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
4080{
4081 struct tcp_sock *tp = tcp_sk(sk);
4082
4083 inet_csk_schedule_ack(sk);
4084
4085 sk->sk_shutdown |= RCV_SHUTDOWN;
4086 sock_set_flag(sk, SOCK_DONE);
4087
4088 switch (sk->sk_state) {
4089 case TCP_SYN_RECV:
4090 case TCP_ESTABLISHED:
4091 /* Move to CLOSE_WAIT */
4092 tcp_set_state(sk, TCP_CLOSE_WAIT);
4093 inet_csk(sk)->icsk_ack.pingpong = 1;
4094 break;
4095
4096 case TCP_CLOSE_WAIT:
4097 case TCP_CLOSING:
4098 /* Received a retransmission of the FIN, do
4099 * nothing.
4100 */
4101 break;
4102 case TCP_LAST_ACK:
4103 /* RFC793: Remain in the LAST-ACK state. */
4104 break;
4105
4106 case TCP_FIN_WAIT1:
4107 /* This case occurs when a simultaneous close
4108 * happens, we must ack the received FIN and
4109 * enter the CLOSING state.
4110 */
4111 tcp_send_ack(sk);
4112 tcp_set_state(sk, TCP_CLOSING);
4113 break;
4114 case TCP_FIN_WAIT2:
4115 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4116 tcp_send_ack(sk);
4117 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4118 break;
4119 default:
4120 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4121 * cases we should never reach this piece of code.
4122 */
4123 printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
4124 __func__, sk->sk_state);
4125 break;
4126 }
4127
4128 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4129 * Probably, we should reset in this case. For now drop them.
4130 */
4131 __skb_queue_purge(&tp->out_of_order_queue);
4132 if (tcp_is_sack(tp))
4133 tcp_sack_reset(&tp->rx_opt);
4134 sk_mem_reclaim(sk);
4135
4136 if (!sock_flag(sk, SOCK_DEAD)) {
4137 sk->sk_state_change(sk);
4138
4139 /* Do not send POLL_HUP for half duplex close. */
4140 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4141 sk->sk_state == TCP_CLOSE)
4142 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4143 else
4144 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4145 }
4146}
4147
4148static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4149 u32 end_seq)
4150{
4151 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4152 if (before(seq, sp->start_seq))
4153 sp->start_seq = seq;
4154 if (after(end_seq, sp->end_seq))
4155 sp->end_seq = end_seq;
4156 return 1;
4157 }
4158 return 0;
4159}
4160
4161static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4162{
4163 struct tcp_sock *tp = tcp_sk(sk);
4164
4165 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4166 int mib_idx;
4167
4168 if (before(seq, tp->rcv_nxt))
4169 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4170 else
4171 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4172
4173 NET_INC_STATS_BH(sock_net(sk), mib_idx);
4174
4175 tp->rx_opt.dsack = 1;
4176 tp->duplicate_sack[0].start_seq = seq;
4177 tp->duplicate_sack[0].end_seq = end_seq;
4178 }
4179}
4180
4181static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4182{
4183 struct tcp_sock *tp = tcp_sk(sk);
4184
4185 if (!tp->rx_opt.dsack)
4186 tcp_dsack_set(sk, seq, end_seq);
4187 else
4188 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4189}
4190
4191static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
4192{
4193 struct tcp_sock *tp = tcp_sk(sk);
4194
4195 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4196 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4197 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4198 tcp_enter_quickack_mode(sk);
4199
4200 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4201 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4202
4203 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4204 end_seq = tp->rcv_nxt;
4205 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4206 }
4207 }
4208
4209 tcp_send_ack(sk);
4210}
4211
4212/* These routines update the SACK block as out-of-order packets arrive or
4213 * in-order packets close up the sequence space.
4214 */
4215static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4216{
4217 int this_sack;
4218 struct tcp_sack_block *sp = &tp->selective_acks[0];
4219 struct tcp_sack_block *swalk = sp + 1;
4220
4221 /* See if the recent change to the first SACK eats into
4222 * or hits the sequence space of other SACK blocks, if so coalesce.
4223 */
4224 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4225 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4226 int i;
4227
4228 /* Zap SWALK, by moving every further SACK up by one slot.
4229 * Decrease num_sacks.
4230 */
4231 tp->rx_opt.num_sacks--;
4232 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4233 sp[i] = sp[i + 1];
4234 continue;
4235 }
4236 this_sack++, swalk++;
4237 }
4238}
4239
4240static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4241{
4242 struct tcp_sock *tp = tcp_sk(sk);
4243 struct tcp_sack_block *sp = &tp->selective_acks[0];
4244 int cur_sacks = tp->rx_opt.num_sacks;
4245 int this_sack;
4246
4247 if (!cur_sacks)
4248 goto new_sack;
4249
4250 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4251 if (tcp_sack_extend(sp, seq, end_seq)) {
4252 /* Rotate this_sack to the first one. */
4253 for (; this_sack > 0; this_sack--, sp--)
4254 swap(*sp, *(sp - 1));
4255 if (cur_sacks > 1)
4256 tcp_sack_maybe_coalesce(tp);
4257 return;
4258 }
4259 }
4260
4261 /* Could not find an adjacent existing SACK, build a new one,
4262 * put it at the front, and shift everyone else down. We
4263 * always know there is at least one SACK present already here.
4264 *
4265 * If the sack array is full, forget about the last one.
4266 */
4267 if (this_sack >= TCP_NUM_SACKS) {
4268 this_sack--;
4269 tp->rx_opt.num_sacks--;
4270 sp--;
4271 }
4272 for (; this_sack > 0; this_sack--, sp--)
4273 *sp = *(sp - 1);
4274
4275new_sack:
4276 /* Build the new head SACK, and we're done. */
4277 sp->start_seq = seq;
4278 sp->end_seq = end_seq;
4279 tp->rx_opt.num_sacks++;
4280}
4281
4282/* RCV.NXT advances, some SACKs should be eaten. */
4283
4284static void tcp_sack_remove(struct tcp_sock *tp)
4285{
4286 struct tcp_sack_block *sp = &tp->selective_acks[0];
4287 int num_sacks = tp->rx_opt.num_sacks;
4288 int this_sack;
4289
4290 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4291 if (skb_queue_empty(&tp->out_of_order_queue)) {
4292 tp->rx_opt.num_sacks = 0;
4293 return;
4294 }
4295
4296 for (this_sack = 0; this_sack < num_sacks;) {
4297 /* Check if the start of the sack is covered by RCV.NXT. */
4298 if (!before(tp->rcv_nxt, sp->start_seq)) {
4299 int i;
4300
4301 /* RCV.NXT must cover all the block! */
4302 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4303
4304 /* Zap this SACK, by moving forward any other SACKS. */
4305 for (i=this_sack+1; i < num_sacks; i++)
4306 tp->selective_acks[i-1] = tp->selective_acks[i];
4307 num_sacks--;
4308 continue;
4309 }
4310 this_sack++;
4311 sp++;
4312 }
4313 tp->rx_opt.num_sacks = num_sacks;
4314}
4315
4316/* This one checks to see if we can put data from the
4317 * out_of_order queue into the receive_queue.
4318 */
4319static void tcp_ofo_queue(struct sock *sk)
4320{
4321 struct tcp_sock *tp = tcp_sk(sk);
4322 __u32 dsack_high = tp->rcv_nxt;
4323 struct sk_buff *skb;
4324
4325 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4326 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4327 break;
4328
4329 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4330 __u32 dsack = dsack_high;
4331 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4332 dsack_high = TCP_SKB_CB(skb)->end_seq;
4333 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4334 }
4335
4336 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4337 SOCK_DEBUG(sk, "ofo packet was already received\n");
4338 __skb_unlink(skb, &tp->out_of_order_queue);
4339 __kfree_skb(skb);
4340 continue;
4341 }
4342 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4343 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4344 TCP_SKB_CB(skb)->end_seq);
4345
4346 __skb_unlink(skb, &tp->out_of_order_queue);
4347 __skb_queue_tail(&sk->sk_receive_queue, skb);
4348 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4349 if (tcp_hdr(skb)->fin)
4350 tcp_fin(skb, sk, tcp_hdr(skb));
4351 }
4352}
4353
4354static int tcp_prune_ofo_queue(struct sock *sk);
4355static int tcp_prune_queue(struct sock *sk);
4356
4357static inline int tcp_try_rmem_schedule(struct sock *sk, unsigned int size)
4358{
4359 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4360 !sk_rmem_schedule(sk, size)) {
4361
4362 if (tcp_prune_queue(sk) < 0)
4363 return -1;
4364
4365 if (!sk_rmem_schedule(sk, size)) {
4366 if (!tcp_prune_ofo_queue(sk))
4367 return -1;
4368
4369 if (!sk_rmem_schedule(sk, size))
4370 return -1;
4371 }
4372 }
4373 return 0;
4374}
4375
4376static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4377{
4378 struct tcphdr *th = tcp_hdr(skb);
4379 struct tcp_sock *tp = tcp_sk(sk);
4380 int eaten = -1;
4381
4382 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4383 goto drop;
4384
4385 skb_dst_drop(skb);
4386 __skb_pull(skb, th->doff * 4);
4387
4388 TCP_ECN_accept_cwr(tp, skb);
4389
4390 tp->rx_opt.dsack = 0;
4391
4392 /* Queue data for delivery to the user.
4393 * Packets in sequence go to the receive queue.
4394 * Out of sequence packets to the out_of_order_queue.
4395 */
4396 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4397 if (tcp_receive_window(tp) == 0)
4398 goto out_of_window;
4399
4400 /* Ok. In sequence. In window. */
4401 if (tp->ucopy.task == current &&
4402 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4403 sock_owned_by_user(sk) && !tp->urg_data) {
4404 int chunk = min_t(unsigned int, skb->len,
4405 tp->ucopy.len);
4406
4407 __set_current_state(TASK_RUNNING);
4408
4409 local_bh_enable();
4410 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
4411 tp->ucopy.len -= chunk;
4412 tp->copied_seq += chunk;
4413 eaten = (chunk == skb->len);
4414 tcp_rcv_space_adjust(sk);
4415 }
4416 local_bh_disable();
4417 }
4418
4419 if (eaten <= 0) {
4420queue_and_out:
4421 if (eaten < 0 &&
4422 tcp_try_rmem_schedule(sk, skb->truesize))
4423 goto drop;
4424
4425 skb_set_owner_r(skb, sk);
4426 __skb_queue_tail(&sk->sk_receive_queue, skb);
4427 }
4428 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4429 if (skb->len)
4430 tcp_event_data_recv(sk, skb);
4431 if (th->fin)
4432 tcp_fin(skb, sk, th);
4433
4434 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4435 tcp_ofo_queue(sk);
4436
4437 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4438 * gap in queue is filled.
4439 */
4440 if (skb_queue_empty(&tp->out_of_order_queue))
4441 inet_csk(sk)->icsk_ack.pingpong = 0;
4442 }
4443
4444 if (tp->rx_opt.num_sacks)
4445 tcp_sack_remove(tp);
4446
4447 tcp_fast_path_check(sk);
4448
4449 if (eaten > 0)
4450 __kfree_skb(skb);
4451 else if (!sock_flag(sk, SOCK_DEAD))
4452 sk->sk_data_ready(sk, 0);
4453 return;
4454 }
4455
4456 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4457 /* A retransmit, 2nd most common case. Force an immediate ack. */
4458 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4459 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4460
4461out_of_window:
4462 tcp_enter_quickack_mode(sk);
4463 inet_csk_schedule_ack(sk);
4464drop:
4465 __kfree_skb(skb);
4466 return;
4467 }
4468
4469 /* Out of window. F.e. zero window probe. */
4470 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4471 goto out_of_window;
4472
4473 tcp_enter_quickack_mode(sk);
4474
4475 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4476 /* Partial packet, seq < rcv_next < end_seq */
4477 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4478 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4479 TCP_SKB_CB(skb)->end_seq);
4480
4481 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4482
4483 /* If window is closed, drop tail of packet. But after
4484 * remembering D-SACK for its head made in previous line.
4485 */
4486 if (!tcp_receive_window(tp))
4487 goto out_of_window;
4488 goto queue_and_out;
4489 }
4490
4491 TCP_ECN_check_ce(tp, skb);
4492
4493 if (tcp_try_rmem_schedule(sk, skb->truesize))
4494 goto drop;
4495
4496 /* Disable header prediction. */
4497 tp->pred_flags = 0;
4498 inet_csk_schedule_ack(sk);
4499
4500 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4501 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4502
4503 skb_set_owner_r(skb, sk);
4504
4505 if (!skb_peek(&tp->out_of_order_queue)) {
4506 /* Initial out of order segment, build 1 SACK. */
4507 if (tcp_is_sack(tp)) {
4508 tp->rx_opt.num_sacks = 1;
4509 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4510 tp->selective_acks[0].end_seq =
4511 TCP_SKB_CB(skb)->end_seq;
4512 }
4513 __skb_queue_head(&tp->out_of_order_queue, skb);
4514 } else {
4515 struct sk_buff *skb1 = skb_peek_tail(&tp->out_of_order_queue);
4516 u32 seq = TCP_SKB_CB(skb)->seq;
4517 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4518
4519 if (seq == TCP_SKB_CB(skb1)->end_seq) {
4520 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4521
4522 if (!tp->rx_opt.num_sacks ||
4523 tp->selective_acks[0].end_seq != seq)
4524 goto add_sack;
4525
4526 /* Common case: data arrive in order after hole. */
4527 tp->selective_acks[0].end_seq = end_seq;
4528 return;
4529 }
4530
4531 /* Find place to insert this segment. */
4532 while (1) {
4533 if (!after(TCP_SKB_CB(skb1)->seq, seq))
4534 break;
4535 if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
4536 skb1 = NULL;
4537 break;
4538 }
4539 skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
4540 }
4541
4542 /* Do skb overlap to previous one? */
4543 if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4544 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4545 /* All the bits are present. Drop. */
4546 __kfree_skb(skb);
4547 tcp_dsack_set(sk, seq, end_seq);
4548 goto add_sack;
4549 }
4550 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4551 /* Partial overlap. */
4552 tcp_dsack_set(sk, seq,
4553 TCP_SKB_CB(skb1)->end_seq);
4554 } else {
4555 if (skb_queue_is_first(&tp->out_of_order_queue,
4556 skb1))
4557 skb1 = NULL;
4558 else
4559 skb1 = skb_queue_prev(
4560 &tp->out_of_order_queue,
4561 skb1);
4562 }
4563 }
4564 if (!skb1)
4565 __skb_queue_head(&tp->out_of_order_queue, skb);
4566 else
4567 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4568
4569 /* And clean segments covered by new one as whole. */
4570 while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
4571 skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
4572
4573 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4574 break;
4575 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4576 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4577 end_seq);
4578 break;
4579 }
4580 __skb_unlink(skb1, &tp->out_of_order_queue);
4581 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4582 TCP_SKB_CB(skb1)->end_seq);
4583 __kfree_skb(skb1);
4584 }
4585
4586add_sack:
4587 if (tcp_is_sack(tp))
4588 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4589 }
4590}
4591
4592static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4593 struct sk_buff_head *list)
4594{
4595 struct sk_buff *next = NULL;
4596
4597 if (!skb_queue_is_last(list, skb))
4598 next = skb_queue_next(list, skb);
4599
4600 __skb_unlink(skb, list);
4601 __kfree_skb(skb);
4602 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4603
4604 return next;
4605}
4606
4607/* Collapse contiguous sequence of skbs head..tail with
4608 * sequence numbers start..end.
4609 *
4610 * If tail is NULL, this means until the end of the list.
4611 *
4612 * Segments with FIN/SYN are not collapsed (only because this
4613 * simplifies code)
4614 */
4615static void
4616tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4617 struct sk_buff *head, struct sk_buff *tail,
4618 u32 start, u32 end)
4619{
4620 struct sk_buff *skb, *n;
4621 bool end_of_skbs;
4622
4623 /* First, check that queue is collapsible and find
4624 * the point where collapsing can be useful. */
4625 skb = head;
4626restart:
4627 end_of_skbs = true;
4628 skb_queue_walk_from_safe(list, skb, n) {
4629 if (skb == tail)
4630 break;
4631 /* No new bits? It is possible on ofo queue. */
4632 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4633 skb = tcp_collapse_one(sk, skb, list);
4634 if (!skb)
4635 break;
4636 goto restart;
4637 }
4638
4639 /* The first skb to collapse is:
4640 * - not SYN/FIN and
4641 * - bloated or contains data before "start" or
4642 * overlaps to the next one.
4643 */
4644 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
4645 (tcp_win_from_space(skb->truesize) > skb->len ||
4646 before(TCP_SKB_CB(skb)->seq, start))) {
4647 end_of_skbs = false;
4648 break;
4649 }
4650
4651 if (!skb_queue_is_last(list, skb)) {
4652 struct sk_buff *next = skb_queue_next(list, skb);
4653 if (next != tail &&
4654 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
4655 end_of_skbs = false;
4656 break;
4657 }
4658 }
4659
4660 /* Decided to skip this, advance start seq. */
4661 start = TCP_SKB_CB(skb)->end_seq;
4662 }
4663 if (end_of_skbs || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
4664 return;
4665
4666 while (before(start, end)) {
4667 struct sk_buff *nskb;
4668 unsigned int header = skb_headroom(skb);
4669 int copy = SKB_MAX_ORDER(header, 0);
4670
4671 /* Too big header? This can happen with IPv6. */
4672 if (copy < 0)
4673 return;
4674 if (end - start < copy)
4675 copy = end - start;
4676 nskb = alloc_skb(copy + header, GFP_ATOMIC);
4677 if (!nskb)
4678 return;
4679
4680 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
4681 skb_set_network_header(nskb, (skb_network_header(skb) -
4682 skb->head));
4683 skb_set_transport_header(nskb, (skb_transport_header(skb) -
4684 skb->head));
4685 skb_reserve(nskb, header);
4686 memcpy(nskb->head, skb->head, header);
4687 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4688 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4689 __skb_queue_before(list, skb, nskb);
4690 skb_set_owner_r(nskb, sk);
4691
4692 /* Copy data, releasing collapsed skbs. */
4693 while (copy > 0) {
4694 int offset = start - TCP_SKB_CB(skb)->seq;
4695 int size = TCP_SKB_CB(skb)->end_seq - start;
4696
4697 BUG_ON(offset < 0);
4698 if (size > 0) {
4699 size = min(copy, size);
4700 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4701 BUG();
4702 TCP_SKB_CB(nskb)->end_seq += size;
4703 copy -= size;
4704 start += size;
4705 }
4706 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4707 skb = tcp_collapse_one(sk, skb, list);
4708 if (!skb ||
4709 skb == tail ||
4710 tcp_hdr(skb)->syn ||
4711 tcp_hdr(skb)->fin)
4712 return;
4713 }
4714 }
4715 }
4716}
4717
4718/* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4719 * and tcp_collapse() them until all the queue is collapsed.
4720 */
4721static void tcp_collapse_ofo_queue(struct sock *sk)
4722{
4723 struct tcp_sock *tp = tcp_sk(sk);
4724 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4725 struct sk_buff *head;
4726 u32 start, end;
4727
4728 if (skb == NULL)
4729 return;
4730
4731 start = TCP_SKB_CB(skb)->seq;
4732 end = TCP_SKB_CB(skb)->end_seq;
4733 head = skb;
4734
4735 for (;;) {
4736 struct sk_buff *next = NULL;
4737
4738 if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
4739 next = skb_queue_next(&tp->out_of_order_queue, skb);
4740 skb = next;
4741
4742 /* Segment is terminated when we see gap or when
4743 * we are at the end of all the queue. */
4744 if (!skb ||
4745 after(TCP_SKB_CB(skb)->seq, end) ||
4746 before(TCP_SKB_CB(skb)->end_seq, start)) {
4747 tcp_collapse(sk, &tp->out_of_order_queue,
4748 head, skb, start, end);
4749 head = skb;
4750 if (!skb)
4751 break;
4752 /* Start new segment */
4753 start = TCP_SKB_CB(skb)->seq;
4754 end = TCP_SKB_CB(skb)->end_seq;
4755 } else {
4756 if (before(TCP_SKB_CB(skb)->seq, start))
4757 start = TCP_SKB_CB(skb)->seq;
4758 if (after(TCP_SKB_CB(skb)->end_seq, end))
4759 end = TCP_SKB_CB(skb)->end_seq;
4760 }
4761 }
4762}
4763
4764/*
4765 * Purge the out-of-order queue.
4766 * Return true if queue was pruned.
4767 */
4768static int tcp_prune_ofo_queue(struct sock *sk)
4769{
4770 struct tcp_sock *tp = tcp_sk(sk);
4771 int res = 0;
4772
4773 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4774 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
4775 __skb_queue_purge(&tp->out_of_order_queue);
4776
4777 /* Reset SACK state. A conforming SACK implementation will
4778 * do the same at a timeout based retransmit. When a connection
4779 * is in a sad state like this, we care only about integrity
4780 * of the connection not performance.
4781 */
4782 if (tp->rx_opt.sack_ok)
4783 tcp_sack_reset(&tp->rx_opt);
4784 sk_mem_reclaim(sk);
4785 res = 1;
4786 }
4787 return res;
4788}
4789
4790/* Reduce allocated memory if we can, trying to get
4791 * the socket within its memory limits again.
4792 *
4793 * Return less than zero if we should start dropping frames
4794 * until the socket owning process reads some of the data
4795 * to stabilize the situation.
4796 */
4797static int tcp_prune_queue(struct sock *sk)
4798{
4799 struct tcp_sock *tp = tcp_sk(sk);
4800
4801 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4802
4803 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
4804
4805 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4806 tcp_clamp_window(sk);
4807 else if (tcp_memory_pressure)
4808 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4809
4810 tcp_collapse_ofo_queue(sk);
4811 if (!skb_queue_empty(&sk->sk_receive_queue))
4812 tcp_collapse(sk, &sk->sk_receive_queue,
4813 skb_peek(&sk->sk_receive_queue),
4814 NULL,
4815 tp->copied_seq, tp->rcv_nxt);
4816 sk_mem_reclaim(sk);
4817
4818 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4819 return 0;
4820
4821 /* Collapsing did not help, destructive actions follow.
4822 * This must not ever occur. */
4823
4824 tcp_prune_ofo_queue(sk);
4825
4826 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4827 return 0;
4828
4829 /* If we are really being abused, tell the caller to silently
4830 * drop receive data on the floor. It will get retransmitted
4831 * and hopefully then we'll have sufficient space.
4832 */
4833 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
4834
4835 /* Massive buffer overcommit. */
4836 tp->pred_flags = 0;
4837 return -1;
4838}
4839
4840/* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4841 * As additional protections, we do not touch cwnd in retransmission phases,
4842 * and if application hit its sndbuf limit recently.
4843 */
4844void tcp_cwnd_application_limited(struct sock *sk)
4845{
4846 struct tcp_sock *tp = tcp_sk(sk);
4847
4848 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
4849 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4850 /* Limited by application or receiver window. */
4851 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4852 u32 win_used = max(tp->snd_cwnd_used, init_win);
4853 if (win_used < tp->snd_cwnd) {
4854 tp->snd_ssthresh = tcp_current_ssthresh(sk);
4855 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4856 }
4857 tp->snd_cwnd_used = 0;
4858 }
4859 tp->snd_cwnd_stamp = tcp_time_stamp;
4860}
4861
4862static int tcp_should_expand_sndbuf(struct sock *sk)
4863{
4864 struct tcp_sock *tp = tcp_sk(sk);
4865
4866 /* If the user specified a specific send buffer setting, do
4867 * not modify it.
4868 */
4869 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4870 return 0;
4871
4872 /* If we are under global TCP memory pressure, do not expand. */
4873 if (tcp_memory_pressure)
4874 return 0;
4875
4876 /* If we are under soft global TCP memory pressure, do not expand. */
4877 if (atomic_long_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
4878 return 0;
4879
4880 /* If we filled the congestion window, do not expand. */
4881 if (tp->packets_out >= tp->snd_cwnd)
4882 return 0;
4883
4884 return 1;
4885}
4886
4887/* When incoming ACK allowed to free some skb from write_queue,
4888 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4889 * on the exit from tcp input handler.
4890 *
4891 * PROBLEM: sndbuf expansion does not work well with largesend.
4892 */
4893static void tcp_new_space(struct sock *sk)
4894{
4895 struct tcp_sock *tp = tcp_sk(sk);
4896
4897 if (tcp_should_expand_sndbuf(sk)) {
4898 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
4899 MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
4900 int demanded = max_t(unsigned int, tp->snd_cwnd,
4901 tp->reordering + 1);
4902 sndmem *= 2 * demanded;
4903 if (sndmem > sk->sk_sndbuf)
4904 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
4905 tp->snd_cwnd_stamp = tcp_time_stamp;
4906 }
4907
4908 sk->sk_write_space(sk);
4909}
4910
4911static void tcp_check_space(struct sock *sk)
4912{
4913 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4914 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4915 if (sk->sk_socket &&
4916 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4917 tcp_new_space(sk);
4918 }
4919}
4920
4921static inline void tcp_data_snd_check(struct sock *sk)
4922{
4923 tcp_push_pending_frames(sk);
4924 tcp_check_space(sk);
4925}
4926
4927/*
4928 * Check if sending an ack is needed.
4929 */
4930static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4931{
4932 struct tcp_sock *tp = tcp_sk(sk);
4933
4934 /* More than one full frame received... */
4935 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
4936 /* ... and right edge of window advances far enough.
4937 * (tcp_recvmsg() will send ACK otherwise). Or...
4938 */
4939 __tcp_select_window(sk) >= tp->rcv_wnd) ||
4940 /* We ACK each frame or... */
4941 tcp_in_quickack_mode(sk) ||
4942 /* We have out of order data. */
4943 (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
4944 /* Then ack it now */
4945 tcp_send_ack(sk);
4946 } else {
4947 /* Else, send delayed ack. */
4948 tcp_send_delayed_ack(sk);
4949 }
4950}
4951
4952static inline void tcp_ack_snd_check(struct sock *sk)
4953{
4954 if (!inet_csk_ack_scheduled(sk)) {
4955 /* We sent a data segment already. */
4956 return;
4957 }
4958 __tcp_ack_snd_check(sk, 1);
4959}
4960
4961/*
4962 * This routine is only called when we have urgent data
4963 * signaled. Its the 'slow' part of tcp_urg. It could be
4964 * moved inline now as tcp_urg is only called from one
4965 * place. We handle URGent data wrong. We have to - as
4966 * BSD still doesn't use the correction from RFC961.
4967 * For 1003.1g we should support a new option TCP_STDURG to permit
4968 * either form (or just set the sysctl tcp_stdurg).
4969 */
4970
4971static void tcp_check_urg(struct sock *sk, struct tcphdr *th)
4972{
4973 struct tcp_sock *tp = tcp_sk(sk);
4974 u32 ptr = ntohs(th->urg_ptr);
4975
4976 if (ptr && !sysctl_tcp_stdurg)
4977 ptr--;
4978 ptr += ntohl(th->seq);
4979
4980 /* Ignore urgent data that we've already seen and read. */
4981 if (after(tp->copied_seq, ptr))
4982 return;
4983
4984 /* Do not replay urg ptr.
4985 *
4986 * NOTE: interesting situation not covered by specs.
4987 * Misbehaving sender may send urg ptr, pointing to segment,
4988 * which we already have in ofo queue. We are not able to fetch
4989 * such data and will stay in TCP_URG_NOTYET until will be eaten
4990 * by recvmsg(). Seems, we are not obliged to handle such wicked
4991 * situations. But it is worth to think about possibility of some
4992 * DoSes using some hypothetical application level deadlock.
4993 */
4994 if (before(ptr, tp->rcv_nxt))
4995 return;
4996
4997 /* Do we already have a newer (or duplicate) urgent pointer? */
4998 if (tp->urg_data && !after(ptr, tp->urg_seq))
4999 return;
5000
5001 /* Tell the world about our new urgent pointer. */
5002 sk_send_sigurg(sk);
5003
5004 /* We may be adding urgent data when the last byte read was
5005 * urgent. To do this requires some care. We cannot just ignore
5006 * tp->copied_seq since we would read the last urgent byte again
5007 * as data, nor can we alter copied_seq until this data arrives
5008 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5009 *
5010 * NOTE. Double Dutch. Rendering to plain English: author of comment
5011 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5012 * and expect that both A and B disappear from stream. This is _wrong_.
5013 * Though this happens in BSD with high probability, this is occasional.
5014 * Any application relying on this is buggy. Note also, that fix "works"
5015 * only in this artificial test. Insert some normal data between A and B and we will
5016 * decline of BSD again. Verdict: it is better to remove to trap
5017 * buggy users.
5018 */
5019 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5020 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5021 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5022 tp->copied_seq++;
5023 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5024 __skb_unlink(skb, &sk->sk_receive_queue);
5025 __kfree_skb(skb);
5026 }
5027 }
5028
5029 tp->urg_data = TCP_URG_NOTYET;
5030 tp->urg_seq = ptr;
5031
5032 /* Disable header prediction. */
5033 tp->pred_flags = 0;
5034}
5035
5036/* This is the 'fast' part of urgent handling. */
5037static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
5038{
5039 struct tcp_sock *tp = tcp_sk(sk);
5040
5041 /* Check if we get a new urgent pointer - normally not. */
5042 if (th->urg)
5043 tcp_check_urg(sk, th);
5044
5045 /* Do we wait for any urgent data? - normally not... */
5046 if (tp->urg_data == TCP_URG_NOTYET) {
5047 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5048 th->syn;
5049
5050 /* Is the urgent pointer pointing into this packet? */
5051 if (ptr < skb->len) {
5052 u8 tmp;
5053 if (skb_copy_bits(skb, ptr, &tmp, 1))
5054 BUG();
5055 tp->urg_data = TCP_URG_VALID | tmp;
5056 if (!sock_flag(sk, SOCK_DEAD))
5057 sk->sk_data_ready(sk, 0);
5058 }
5059 }
5060}
5061
5062static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
5063{
5064 struct tcp_sock *tp = tcp_sk(sk);
5065 int chunk = skb->len - hlen;
5066 int err;
5067
5068 local_bh_enable();
5069 if (skb_csum_unnecessary(skb))
5070 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
5071 else
5072 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
5073 tp->ucopy.iov);
5074
5075 if (!err) {
5076 tp->ucopy.len -= chunk;
5077 tp->copied_seq += chunk;
5078 tcp_rcv_space_adjust(sk);
5079 }
5080
5081 local_bh_disable();
5082 return err;
5083}
5084
5085static __sum16 __tcp_checksum_complete_user(struct sock *sk,
5086 struct sk_buff *skb)
5087{
5088 __sum16 result;
5089
5090 if (sock_owned_by_user(sk)) {
5091 local_bh_enable();
5092 result = __tcp_checksum_complete(skb);
5093 local_bh_disable();
5094 } else {
5095 result = __tcp_checksum_complete(skb);
5096 }
5097 return result;
5098}
5099
5100static inline int tcp_checksum_complete_user(struct sock *sk,
5101 struct sk_buff *skb)
5102{
5103 return !skb_csum_unnecessary(skb) &&
5104 __tcp_checksum_complete_user(sk, skb);
5105}
5106
5107#ifdef CONFIG_NET_DMA
5108static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb,
5109 int hlen)
5110{
5111 struct tcp_sock *tp = tcp_sk(sk);
5112 int chunk = skb->len - hlen;
5113 int dma_cookie;
5114 int copied_early = 0;
5115
5116 if (tp->ucopy.wakeup)
5117 return 0;
5118
5119 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
5120 tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY);
5121
5122 if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
5123
5124 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
5125 skb, hlen,
5126 tp->ucopy.iov, chunk,
5127 tp->ucopy.pinned_list);
5128
5129 if (dma_cookie < 0)
5130 goto out;
5131
5132 tp->ucopy.dma_cookie = dma_cookie;
5133 copied_early = 1;
5134
5135 tp->ucopy.len -= chunk;
5136 tp->copied_seq += chunk;
5137 tcp_rcv_space_adjust(sk);
5138
5139 if ((tp->ucopy.len == 0) ||
5140 (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
5141 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
5142 tp->ucopy.wakeup = 1;
5143 sk->sk_data_ready(sk, 0);
5144 }
5145 } else if (chunk > 0) {
5146 tp->ucopy.wakeup = 1;
5147 sk->sk_data_ready(sk, 0);
5148 }
5149out:
5150 return copied_early;
5151}
5152#endif /* CONFIG_NET_DMA */
5153
5154/* Does PAWS and seqno based validation of an incoming segment, flags will
5155 * play significant role here.
5156 */
5157static int tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5158 struct tcphdr *th, int syn_inerr)
5159{
5160 u8 *hash_location;
5161 struct tcp_sock *tp = tcp_sk(sk);
5162
5163 /* RFC1323: H1. Apply PAWS check first. */
5164 if (tcp_fast_parse_options(skb, th, tp, &hash_location) &&
5165 tp->rx_opt.saw_tstamp &&
5166 tcp_paws_discard(sk, skb)) {
5167 if (!th->rst) {
5168 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5169 tcp_send_dupack(sk, skb);
5170 goto discard;
5171 }
5172 /* Reset is accepted even if it did not pass PAWS. */
5173 }
5174
5175 /* Step 1: check sequence number */
5176 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5177 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5178 * (RST) segments are validated by checking their SEQ-fields."
5179 * And page 69: "If an incoming segment is not acceptable,
5180 * an acknowledgment should be sent in reply (unless the RST
5181 * bit is set, if so drop the segment and return)".
5182 */
5183 if (!th->rst)
5184 tcp_send_dupack(sk, skb);
5185 goto discard;
5186 }
5187
5188 /* Step 2: check RST bit */
5189 if (th->rst) {
5190 tcp_reset(sk);
5191 goto discard;
5192 }
5193
5194 /* ts_recent update must be made after we are sure that the packet
5195 * is in window.
5196 */
5197 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
5198
5199 /* step 3: check security and precedence [ignored] */
5200
5201 /* step 4: Check for a SYN in window. */
5202 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
5203 if (syn_inerr)
5204 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5205 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONSYN);
5206 tcp_reset(sk);
5207 return -1;
5208 }
5209
5210 return 1;
5211
5212discard:
5213 __kfree_skb(skb);
5214 return 0;
5215}
5216
5217/*
5218 * TCP receive function for the ESTABLISHED state.
5219 *
5220 * It is split into a fast path and a slow path. The fast path is
5221 * disabled when:
5222 * - A zero window was announced from us - zero window probing
5223 * is only handled properly in the slow path.
5224 * - Out of order segments arrived.
5225 * - Urgent data is expected.
5226 * - There is no buffer space left
5227 * - Unexpected TCP flags/window values/header lengths are received
5228 * (detected by checking the TCP header against pred_flags)
5229 * - Data is sent in both directions. Fast path only supports pure senders
5230 * or pure receivers (this means either the sequence number or the ack
5231 * value must stay constant)
5232 * - Unexpected TCP option.
5233 *
5234 * When these conditions are not satisfied it drops into a standard
5235 * receive procedure patterned after RFC793 to handle all cases.
5236 * The first three cases are guaranteed by proper pred_flags setting,
5237 * the rest is checked inline. Fast processing is turned on in
5238 * tcp_data_queue when everything is OK.
5239 */
5240int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5241 struct tcphdr *th, unsigned len)
5242{
5243 struct tcp_sock *tp = tcp_sk(sk);
5244 int res;
5245
5246 /*
5247 * Header prediction.
5248 * The code loosely follows the one in the famous
5249 * "30 instruction TCP receive" Van Jacobson mail.
5250 *
5251 * Van's trick is to deposit buffers into socket queue
5252 * on a device interrupt, to call tcp_recv function
5253 * on the receive process context and checksum and copy
5254 * the buffer to user space. smart...
5255 *
5256 * Our current scheme is not silly either but we take the
5257 * extra cost of the net_bh soft interrupt processing...
5258 * We do checksum and copy also but from device to kernel.
5259 */
5260
5261 tp->rx_opt.saw_tstamp = 0;
5262
5263 /* pred_flags is 0xS?10 << 16 + snd_wnd
5264 * if header_prediction is to be made
5265 * 'S' will always be tp->tcp_header_len >> 2
5266 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5267 * turn it off (when there are holes in the receive
5268 * space for instance)
5269 * PSH flag is ignored.
5270 */
5271
5272 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5273 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5274 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5275 int tcp_header_len = tp->tcp_header_len;
5276
5277 /* Timestamp header prediction: tcp_header_len
5278 * is automatically equal to th->doff*4 due to pred_flags
5279 * match.
5280 */
5281
5282 /* Check timestamp */
5283 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5284 /* No? Slow path! */
5285 if (!tcp_parse_aligned_timestamp(tp, th))
5286 goto slow_path;
5287
5288 /* If PAWS failed, check it more carefully in slow path */
5289 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5290 goto slow_path;
5291
5292 /* DO NOT update ts_recent here, if checksum fails
5293 * and timestamp was corrupted part, it will result
5294 * in a hung connection since we will drop all
5295 * future packets due to the PAWS test.
5296 */
5297 }
5298
5299 if (len <= tcp_header_len) {
5300 /* Bulk data transfer: sender */
5301 if (len == tcp_header_len) {
5302 /* Predicted packet is in window by definition.
5303 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5304 * Hence, check seq<=rcv_wup reduces to:
5305 */
5306 if (tcp_header_len ==
5307 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5308 tp->rcv_nxt == tp->rcv_wup)
5309 tcp_store_ts_recent(tp);
5310
5311 /* We know that such packets are checksummed
5312 * on entry.
5313 */
5314 tcp_ack(sk, skb, 0);
5315 __kfree_skb(skb);
5316 tcp_data_snd_check(sk);
5317 return 0;
5318 } else { /* Header too small */
5319 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5320 goto discard;
5321 }
5322 } else {
5323 int eaten = 0;
5324 int copied_early = 0;
5325
5326 if (tp->copied_seq == tp->rcv_nxt &&
5327 len - tcp_header_len <= tp->ucopy.len) {
5328#ifdef CONFIG_NET_DMA
5329 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
5330 copied_early = 1;
5331 eaten = 1;
5332 }
5333#endif
5334 if (tp->ucopy.task == current &&
5335 sock_owned_by_user(sk) && !copied_early) {
5336 __set_current_state(TASK_RUNNING);
5337
5338 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
5339 eaten = 1;
5340 }
5341 if (eaten) {
5342 /* Predicted packet is in window by definition.
5343 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5344 * Hence, check seq<=rcv_wup reduces to:
5345 */
5346 if (tcp_header_len ==
5347 (sizeof(struct tcphdr) +
5348 TCPOLEN_TSTAMP_ALIGNED) &&
5349 tp->rcv_nxt == tp->rcv_wup)
5350 tcp_store_ts_recent(tp);
5351
5352 tcp_rcv_rtt_measure_ts(sk, skb);
5353
5354 __skb_pull(skb, tcp_header_len);
5355 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5356 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
5357 }
5358 if (copied_early)
5359 tcp_cleanup_rbuf(sk, skb->len);
5360 }
5361 if (!eaten) {
5362 if (tcp_checksum_complete_user(sk, skb))
5363 goto csum_error;
5364
5365 /* Predicted packet is in window by definition.
5366 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5367 * Hence, check seq<=rcv_wup reduces to:
5368 */
5369 if (tcp_header_len ==
5370 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5371 tp->rcv_nxt == tp->rcv_wup)
5372 tcp_store_ts_recent(tp);
5373
5374 tcp_rcv_rtt_measure_ts(sk, skb);
5375
5376 if ((int)skb->truesize > sk->sk_forward_alloc)
5377 goto step5;
5378
5379 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
5380
5381 /* Bulk data transfer: receiver */
5382 __skb_pull(skb, tcp_header_len);
5383 __skb_queue_tail(&sk->sk_receive_queue, skb);
5384 skb_set_owner_r(skb, sk);
5385 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5386 }
5387
5388 tcp_event_data_recv(sk, skb);
5389
5390 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5391 /* Well, only one small jumplet in fast path... */
5392 tcp_ack(sk, skb, FLAG_DATA);
5393 tcp_data_snd_check(sk);
5394 if (!inet_csk_ack_scheduled(sk))
5395 goto no_ack;
5396 }
5397
5398 if (!copied_early || tp->rcv_nxt != tp->rcv_wup)
5399 __tcp_ack_snd_check(sk, 0);
5400no_ack:
5401#ifdef CONFIG_NET_DMA
5402 if (copied_early)
5403 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
5404 else
5405#endif
5406 if (eaten)
5407 __kfree_skb(skb);
5408 else
5409 sk->sk_data_ready(sk, 0);
5410 return 0;
5411 }
5412 }
5413
5414slow_path:
5415 if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
5416 goto csum_error;
5417
5418 /*
5419 * Standard slow path.
5420 */
5421
5422 res = tcp_validate_incoming(sk, skb, th, 1);
5423 if (res <= 0)
5424 return -res;
5425
5426step5:
5427 if (th->ack && tcp_ack(sk, skb, FLAG_SLOWPATH) < 0)
5428 goto discard;
5429
5430 tcp_rcv_rtt_measure_ts(sk, skb);
5431
5432 /* Process urgent data. */
5433 tcp_urg(sk, skb, th);
5434
5435 /* step 7: process the segment text */
5436 tcp_data_queue(sk, skb);
5437
5438 tcp_data_snd_check(sk);
5439 tcp_ack_snd_check(sk);
5440 return 0;
5441
5442csum_error:
5443 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5444
5445discard:
5446 __kfree_skb(skb);
5447 return 0;
5448}
5449EXPORT_SYMBOL(tcp_rcv_established);
5450
5451static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5452 struct tcphdr *th, unsigned len)
5453{
5454 u8 *hash_location;
5455 struct inet_connection_sock *icsk = inet_csk(sk);
5456 struct tcp_sock *tp = tcp_sk(sk);
5457 struct tcp_cookie_values *cvp = tp->cookie_values;
5458 int saved_clamp = tp->rx_opt.mss_clamp;
5459
5460 tcp_parse_options(skb, &tp->rx_opt, &hash_location, 0);
5461
5462 if (th->ack) {
5463 /* rfc793:
5464 * "If the state is SYN-SENT then
5465 * first check the ACK bit
5466 * If the ACK bit is set
5467 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5468 * a reset (unless the RST bit is set, if so drop
5469 * the segment and return)"
5470 *
5471 * We do not send data with SYN, so that RFC-correct
5472 * test reduces to:
5473 */
5474 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
5475 goto reset_and_undo;
5476
5477 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5478 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5479 tcp_time_stamp)) {
5480 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
5481 goto reset_and_undo;
5482 }
5483
5484 /* Now ACK is acceptable.
5485 *
5486 * "If the RST bit is set
5487 * If the ACK was acceptable then signal the user "error:
5488 * connection reset", drop the segment, enter CLOSED state,
5489 * delete TCB, and return."
5490 */
5491
5492 if (th->rst) {
5493 tcp_reset(sk);
5494 goto discard;
5495 }
5496
5497 /* rfc793:
5498 * "fifth, if neither of the SYN or RST bits is set then
5499 * drop the segment and return."
5500 *
5501 * See note below!
5502 * --ANK(990513)
5503 */
5504 if (!th->syn)
5505 goto discard_and_undo;
5506
5507 /* rfc793:
5508 * "If the SYN bit is on ...
5509 * are acceptable then ...
5510 * (our SYN has been ACKed), change the connection
5511 * state to ESTABLISHED..."
5512 */
5513
5514 TCP_ECN_rcv_synack(tp, th);
5515
5516 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5517 tcp_ack(sk, skb, FLAG_SLOWPATH);
5518
5519 /* Ok.. it's good. Set up sequence numbers and
5520 * move to established.
5521 */
5522 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5523 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5524
5525 /* RFC1323: The window in SYN & SYN/ACK segments is
5526 * never scaled.
5527 */
5528 tp->snd_wnd = ntohs(th->window);
5529 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5530
5531 if (!tp->rx_opt.wscale_ok) {
5532 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5533 tp->window_clamp = min(tp->window_clamp, 65535U);
5534 }
5535
5536 if (tp->rx_opt.saw_tstamp) {
5537 tp->rx_opt.tstamp_ok = 1;
5538 tp->tcp_header_len =
5539 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5540 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5541 tcp_store_ts_recent(tp);
5542 } else {
5543 tp->tcp_header_len = sizeof(struct tcphdr);
5544 }
5545
5546 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5547 tcp_enable_fack(tp);
5548
5549 tcp_mtup_init(sk);
5550 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5551 tcp_initialize_rcv_mss(sk);
5552
5553 /* Remember, tcp_poll() does not lock socket!
5554 * Change state from SYN-SENT only after copied_seq
5555 * is initialized. */
5556 tp->copied_seq = tp->rcv_nxt;
5557
5558 if (cvp != NULL &&
5559 cvp->cookie_pair_size > 0 &&
5560 tp->rx_opt.cookie_plus > 0) {
5561 int cookie_size = tp->rx_opt.cookie_plus
5562 - TCPOLEN_COOKIE_BASE;
5563 int cookie_pair_size = cookie_size
5564 + cvp->cookie_desired;
5565
5566 /* A cookie extension option was sent and returned.
5567 * Note that each incoming SYNACK replaces the
5568 * Responder cookie. The initial exchange is most
5569 * fragile, as protection against spoofing relies
5570 * entirely upon the sequence and timestamp (above).
5571 * This replacement strategy allows the correct pair to
5572 * pass through, while any others will be filtered via
5573 * Responder verification later.
5574 */
5575 if (sizeof(cvp->cookie_pair) >= cookie_pair_size) {
5576 memcpy(&cvp->cookie_pair[cvp->cookie_desired],
5577 hash_location, cookie_size);
5578 cvp->cookie_pair_size = cookie_pair_size;
5579 }
5580 }
5581
5582 smp_mb();
5583 tcp_set_state(sk, TCP_ESTABLISHED);
5584
5585 security_inet_conn_established(sk, skb);
5586
5587 /* Make sure socket is routed, for correct metrics. */
5588 icsk->icsk_af_ops->rebuild_header(sk);
5589
5590 tcp_init_metrics(sk);
5591
5592 tcp_init_congestion_control(sk);
5593
5594 /* Prevent spurious tcp_cwnd_restart() on first data
5595 * packet.
5596 */
5597 tp->lsndtime = tcp_time_stamp;
5598
5599 tcp_init_buffer_space(sk);
5600
5601 if (sock_flag(sk, SOCK_KEEPOPEN))
5602 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5603
5604 if (!tp->rx_opt.snd_wscale)
5605 __tcp_fast_path_on(tp, tp->snd_wnd);
5606 else
5607 tp->pred_flags = 0;
5608
5609 if (!sock_flag(sk, SOCK_DEAD)) {
5610 sk->sk_state_change(sk);
5611 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5612 }
5613
5614 if (sk->sk_write_pending ||
5615 icsk->icsk_accept_queue.rskq_defer_accept ||
5616 icsk->icsk_ack.pingpong) {
5617 /* Save one ACK. Data will be ready after
5618 * several ticks, if write_pending is set.
5619 *
5620 * It may be deleted, but with this feature tcpdumps
5621 * look so _wonderfully_ clever, that I was not able
5622 * to stand against the temptation 8) --ANK
5623 */
5624 inet_csk_schedule_ack(sk);
5625 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5626 icsk->icsk_ack.ato = TCP_ATO_MIN;
5627 tcp_incr_quickack(sk);
5628 tcp_enter_quickack_mode(sk);
5629 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5630 TCP_DELACK_MAX, TCP_RTO_MAX);
5631
5632discard:
5633 __kfree_skb(skb);
5634 return 0;
5635 } else {
5636 tcp_send_ack(sk);
5637 }
5638 return -1;
5639 }
5640
5641 /* No ACK in the segment */
5642
5643 if (th->rst) {
5644 /* rfc793:
5645 * "If the RST bit is set
5646 *
5647 * Otherwise (no ACK) drop the segment and return."
5648 */
5649
5650 goto discard_and_undo;
5651 }
5652
5653 /* PAWS check. */
5654 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5655 tcp_paws_reject(&tp->rx_opt, 0))
5656 goto discard_and_undo;
5657
5658 if (th->syn) {
5659 /* We see SYN without ACK. It is attempt of
5660 * simultaneous connect with crossed SYNs.
5661 * Particularly, it can be connect to self.
5662 */
5663 tcp_set_state(sk, TCP_SYN_RECV);
5664
5665 if (tp->rx_opt.saw_tstamp) {
5666 tp->rx_opt.tstamp_ok = 1;
5667 tcp_store_ts_recent(tp);
5668 tp->tcp_header_len =
5669 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5670 } else {
5671 tp->tcp_header_len = sizeof(struct tcphdr);
5672 }
5673
5674 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5675 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5676
5677 /* RFC1323: The window in SYN & SYN/ACK segments is
5678 * never scaled.
5679 */
5680 tp->snd_wnd = ntohs(th->window);
5681 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5682 tp->max_window = tp->snd_wnd;
5683
5684 TCP_ECN_rcv_syn(tp, th);
5685
5686 tcp_mtup_init(sk);
5687 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5688 tcp_initialize_rcv_mss(sk);
5689
5690 tcp_send_synack(sk);
5691#if 0
5692 /* Note, we could accept data and URG from this segment.
5693 * There are no obstacles to make this.
5694 *
5695 * However, if we ignore data in ACKless segments sometimes,
5696 * we have no reasons to accept it sometimes.
5697 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5698 * is not flawless. So, discard packet for sanity.
5699 * Uncomment this return to process the data.
5700 */
5701 return -1;
5702#else
5703 goto discard;
5704#endif
5705 }
5706 /* "fifth, if neither of the SYN or RST bits is set then
5707 * drop the segment and return."
5708 */
5709
5710discard_and_undo:
5711 tcp_clear_options(&tp->rx_opt);
5712 tp->rx_opt.mss_clamp = saved_clamp;
5713 goto discard;
5714
5715reset_and_undo:
5716 tcp_clear_options(&tp->rx_opt);
5717 tp->rx_opt.mss_clamp = saved_clamp;
5718 return 1;
5719}
5720
5721/*
5722 * This function implements the receiving procedure of RFC 793 for
5723 * all states except ESTABLISHED and TIME_WAIT.
5724 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5725 * address independent.
5726 */
5727
5728int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5729 struct tcphdr *th, unsigned len)
5730{
5731 struct tcp_sock *tp = tcp_sk(sk);
5732 struct inet_connection_sock *icsk = inet_csk(sk);
5733 int queued = 0;
5734 int res;
5735
5736 tp->rx_opt.saw_tstamp = 0;
5737
5738 switch (sk->sk_state) {
5739 case TCP_CLOSE:
5740 goto discard;
5741
5742 case TCP_LISTEN:
5743 if (th->ack)
5744 return 1;
5745
5746 if (th->rst)
5747 goto discard;
5748
5749 if (th->syn) {
5750 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5751 return 1;
5752
5753 /* Now we have several options: In theory there is
5754 * nothing else in the frame. KA9Q has an option to
5755 * send data with the syn, BSD accepts data with the
5756 * syn up to the [to be] advertised window and
5757 * Solaris 2.1 gives you a protocol error. For now
5758 * we just ignore it, that fits the spec precisely
5759 * and avoids incompatibilities. It would be nice in
5760 * future to drop through and process the data.
5761 *
5762 * Now that TTCP is starting to be used we ought to
5763 * queue this data.
5764 * But, this leaves one open to an easy denial of
5765 * service attack, and SYN cookies can't defend
5766 * against this problem. So, we drop the data
5767 * in the interest of security over speed unless
5768 * it's still in use.
5769 */
5770 kfree_skb(skb);
5771 return 0;
5772 }
5773 goto discard;
5774
5775 case TCP_SYN_SENT:
5776 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5777 if (queued >= 0)
5778 return queued;
5779
5780 /* Do step6 onward by hand. */
5781 tcp_urg(sk, skb, th);
5782 __kfree_skb(skb);
5783 tcp_data_snd_check(sk);
5784 return 0;
5785 }
5786
5787 res = tcp_validate_incoming(sk, skb, th, 0);
5788 if (res <= 0)
5789 return -res;
5790
5791 /* step 5: check the ACK field */
5792 if (th->ack) {
5793 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH) > 0;
5794
5795 switch (sk->sk_state) {
5796 case TCP_SYN_RECV:
5797 if (acceptable) {
5798 tp->copied_seq = tp->rcv_nxt;
5799 smp_mb();
5800 tcp_set_state(sk, TCP_ESTABLISHED);
5801 sk->sk_state_change(sk);
5802
5803 /* Note, that this wakeup is only for marginal
5804 * crossed SYN case. Passively open sockets
5805 * are not waked up, because sk->sk_sleep ==
5806 * NULL and sk->sk_socket == NULL.
5807 */
5808 if (sk->sk_socket)
5809 sk_wake_async(sk,
5810 SOCK_WAKE_IO, POLL_OUT);
5811
5812 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5813 tp->snd_wnd = ntohs(th->window) <<
5814 tp->rx_opt.snd_wscale;
5815 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5816
5817 if (tp->rx_opt.tstamp_ok)
5818 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5819
5820 /* Make sure socket is routed, for
5821 * correct metrics.
5822 */
5823 icsk->icsk_af_ops->rebuild_header(sk);
5824
5825 tcp_init_metrics(sk);
5826
5827 tcp_init_congestion_control(sk);
5828
5829 /* Prevent spurious tcp_cwnd_restart() on
5830 * first data packet.
5831 */
5832 tp->lsndtime = tcp_time_stamp;
5833
5834 tcp_mtup_init(sk);
5835 tcp_initialize_rcv_mss(sk);
5836 tcp_init_buffer_space(sk);
5837 tcp_fast_path_on(tp);
5838 } else {
5839 return 1;
5840 }
5841 break;
5842
5843 case TCP_FIN_WAIT1:
5844 if (tp->snd_una == tp->write_seq) {
5845 tcp_set_state(sk, TCP_FIN_WAIT2);
5846 sk->sk_shutdown |= SEND_SHUTDOWN;
5847 dst_confirm(__sk_dst_get(sk));
5848
5849 if (!sock_flag(sk, SOCK_DEAD))
5850 /* Wake up lingering close() */
5851 sk->sk_state_change(sk);
5852 else {
5853 int tmo;
5854
5855 if (tp->linger2 < 0 ||
5856 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5857 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5858 tcp_done(sk);
5859 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5860 return 1;
5861 }
5862
5863 tmo = tcp_fin_time(sk);
5864 if (tmo > TCP_TIMEWAIT_LEN) {
5865 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5866 } else if (th->fin || sock_owned_by_user(sk)) {
5867 /* Bad case. We could lose such FIN otherwise.
5868 * It is not a big problem, but it looks confusing
5869 * and not so rare event. We still can lose it now,
5870 * if it spins in bh_lock_sock(), but it is really
5871 * marginal case.
5872 */
5873 inet_csk_reset_keepalive_timer(sk, tmo);
5874 } else {
5875 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5876 goto discard;
5877 }
5878 }
5879 }
5880 break;
5881
5882 case TCP_CLOSING:
5883 if (tp->snd_una == tp->write_seq) {
5884 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5885 goto discard;
5886 }
5887 break;
5888
5889 case TCP_LAST_ACK:
5890 if (tp->snd_una == tp->write_seq) {
5891 tcp_update_metrics(sk);
5892 tcp_done(sk);
5893 goto discard;
5894 }
5895 break;
5896 }
5897 } else
5898 goto discard;
5899
5900 /* step 6: check the URG bit */
5901 tcp_urg(sk, skb, th);
5902
5903 /* step 7: process the segment text */
5904 switch (sk->sk_state) {
5905 case TCP_CLOSE_WAIT:
5906 case TCP_CLOSING:
5907 case TCP_LAST_ACK:
5908 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5909 break;
5910 case TCP_FIN_WAIT1:
5911 case TCP_FIN_WAIT2:
5912 /* RFC 793 says to queue data in these states,
5913 * RFC 1122 says we MUST send a reset.
5914 * BSD 4.4 also does reset.
5915 */
5916 if (sk->sk_shutdown & RCV_SHUTDOWN) {
5917 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5918 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5919 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5920 tcp_reset(sk);
5921 return 1;
5922 }
5923 }
5924 /* Fall through */
5925 case TCP_ESTABLISHED:
5926 tcp_data_queue(sk, skb);
5927 queued = 1;
5928 break;
5929 }
5930
5931 /* tcp_data could move socket to TIME-WAIT */
5932 if (sk->sk_state != TCP_CLOSE) {
5933 tcp_data_snd_check(sk);
5934 tcp_ack_snd_check(sk);
5935 }
5936
5937 if (!queued) {
5938discard:
5939 __kfree_skb(skb);
5940 }
5941 return 0;
5942}
5943EXPORT_SYMBOL(tcp_rcv_state_process);