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 <asm/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
85 int 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)
119 static DEFINE_STATIC_KEY_DEFERRED_FALSE(clean_acked_data_enabled, HZ);
120
121 void 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 }
127 EXPORT_SYMBOL_GPL(clean_acked_data_enable);
128
129 void 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 }
134 EXPORT_SYMBOL_GPL(clean_acked_data_disable);
135
136 void clean_acked_data_flush(void)
137 {
138 static_key_deferred_flush(&clean_acked_data_enabled);
139 }
140 EXPORT_SYMBOL_GPL(clean_acked_data_flush);
141 #endif
142
143 #ifdef CONFIG_CGROUP_BPF
144 static 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
178 static 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
196 static void bpf_skops_parse_hdr(struct sock *sk, struct sk_buff *skb)
197 {
198 }
199
200 static void bpf_skops_established(struct sock *sk, int bpf_op,
201 struct sk_buff *skb)
202 {
203 }
204 #endif
205
206 static __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 */
222 static 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 WRITE_ONCE(tcp_sk(sk)->window_clamp,
248 tcp_win_from_space(sk, sk->sk_rcvbuf));
249 }
250 icsk->icsk_ack.rcv_mss = min_t(unsigned int, len,
251 tcp_sk(sk)->advmss);
252 /* Account for possibly-removed options */
253 DO_ONCE_LITE_IF(len > icsk->icsk_ack.rcv_mss + MAX_TCP_OPTION_SPACE,
254 tcp_gro_dev_warn, sk, skb, len);
255 /* If the skb has a len of exactly 1*MSS and has the PSH bit
256 * set then it is likely the end of an application write. So
257 * more data may not be arriving soon, and yet the data sender
258 * may be waiting for an ACK if cwnd-bound or using TX zero
259 * copy. So we set ICSK_ACK_PUSHED here so that
260 * tcp_cleanup_rbuf() will send an ACK immediately if the app
261 * reads all of the data and is not ping-pong. If len > MSS
262 * then this logic does not matter (and does not hurt) because
263 * tcp_cleanup_rbuf() will always ACK immediately if the app
264 * reads data and there is more than an MSS of unACKed data.
265 */
266 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_PSH)
267 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
268 } else {
269 /* Otherwise, we make more careful check taking into account,
270 * that SACKs block is variable.
271 *
272 * "len" is invariant segment length, including TCP header.
273 */
274 len += skb->data - skb_transport_header(skb);
275 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
276 /* If PSH is not set, packet should be
277 * full sized, provided peer TCP is not badly broken.
278 * This observation (if it is correct 8)) allows
279 * to handle super-low mtu links fairly.
280 */
281 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
282 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
283 /* Subtract also invariant (if peer is RFC compliant),
284 * tcp header plus fixed timestamp option length.
285 * Resulting "len" is MSS free of SACK jitter.
286 */
287 len -= tcp_sk(sk)->tcp_header_len;
288 icsk->icsk_ack.last_seg_size = len;
289 if (len == lss) {
290 icsk->icsk_ack.rcv_mss = len;
291 return;
292 }
293 }
294 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
295 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
296 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
297 }
298 }
299
300 static void tcp_incr_quickack(struct sock *sk, unsigned int max_quickacks)
301 {
302 struct inet_connection_sock *icsk = inet_csk(sk);
303 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
304
305 if (quickacks == 0)
306 quickacks = 2;
307 quickacks = min(quickacks, max_quickacks);
308 if (quickacks > icsk->icsk_ack.quick)
309 icsk->icsk_ack.quick = quickacks;
310 }
311
312 static void tcp_enter_quickack_mode(struct sock *sk, unsigned int max_quickacks)
313 {
314 struct inet_connection_sock *icsk = inet_csk(sk);
315
316 tcp_incr_quickack(sk, max_quickacks);
317 inet_csk_exit_pingpong_mode(sk);
318 icsk->icsk_ack.ato = TCP_ATO_MIN;
319 }
320
321 /* Send ACKs quickly, if "quick" count is not exhausted
322 * and the session is not interactive.
323 */
324
325 static bool tcp_in_quickack_mode(struct sock *sk)
326 {
327 const struct inet_connection_sock *icsk = inet_csk(sk);
328 const struct dst_entry *dst = __sk_dst_get(sk);
329
330 return (dst && dst_metric(dst, RTAX_QUICKACK)) ||
331 (icsk->icsk_ack.quick && !inet_csk_in_pingpong_mode(sk));
332 }
333
334 static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
335 {
336 if (tp->ecn_flags & TCP_ECN_OK)
337 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
338 }
339
340 static void tcp_ecn_accept_cwr(struct sock *sk, const struct sk_buff *skb)
341 {
342 if (tcp_hdr(skb)->cwr) {
343 tcp_sk(sk)->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
344
345 /* If the sender is telling us it has entered CWR, then its
346 * cwnd may be very low (even just 1 packet), so we should ACK
347 * immediately.
348 */
349 if (TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq)
350 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
351 }
352 }
353
354 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
355 {
356 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
357 }
358
359 static void __tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
360 {
361 struct tcp_sock *tp = tcp_sk(sk);
362
363 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
364 case INET_ECN_NOT_ECT:
365 /* Funny extension: if ECT is not set on a segment,
366 * and we already seen ECT on a previous segment,
367 * it is probably a retransmit.
368 */
369 if (tp->ecn_flags & TCP_ECN_SEEN)
370 tcp_enter_quickack_mode(sk, 2);
371 break;
372 case INET_ECN_CE:
373 if (tcp_ca_needs_ecn(sk))
374 tcp_ca_event(sk, CA_EVENT_ECN_IS_CE);
375
376 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
377 /* Better not delay acks, sender can have a very low cwnd */
378 tcp_enter_quickack_mode(sk, 2);
379 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
380 }
381 tp->ecn_flags |= TCP_ECN_SEEN;
382 break;
383 default:
384 if (tcp_ca_needs_ecn(sk))
385 tcp_ca_event(sk, CA_EVENT_ECN_NO_CE);
386 tp->ecn_flags |= TCP_ECN_SEEN;
387 break;
388 }
389 }
390
391 static void tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
392 {
393 if (tcp_sk(sk)->ecn_flags & TCP_ECN_OK)
394 __tcp_ecn_check_ce(sk, skb);
395 }
396
397 static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
398 {
399 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
400 tp->ecn_flags &= ~TCP_ECN_OK;
401 }
402
403 static void tcp_ecn_rcv_syn(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
409 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
410 {
411 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
412 return true;
413 return false;
414 }
415
416 /* Buffer size and advertised window tuning.
417 *
418 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
419 */
420
421 static void tcp_sndbuf_expand(struct sock *sk)
422 {
423 const struct tcp_sock *tp = tcp_sk(sk);
424 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
425 int sndmem, per_mss;
426 u32 nr_segs;
427
428 /* Worst case is non GSO/TSO : each frame consumes one skb
429 * and skb->head is kmalloced using power of two area of memory
430 */
431 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
432 MAX_TCP_HEADER +
433 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
434
435 per_mss = roundup_pow_of_two(per_mss) +
436 SKB_DATA_ALIGN(sizeof(struct sk_buff));
437
438 nr_segs = max_t(u32, TCP_INIT_CWND, tcp_snd_cwnd(tp));
439 nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
440
441 /* Fast Recovery (RFC 5681 3.2) :
442 * Cubic needs 1.7 factor, rounded to 2 to include
443 * extra cushion (application might react slowly to EPOLLOUT)
444 */
445 sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2;
446 sndmem *= nr_segs * per_mss;
447
448 if (sk->sk_sndbuf < sndmem)
449 WRITE_ONCE(sk->sk_sndbuf,
450 min(sndmem, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_wmem[2])));
451 }
452
453 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
454 *
455 * All tcp_full_space() is split to two parts: "network" buffer, allocated
456 * forward and advertised in receiver window (tp->rcv_wnd) and
457 * "application buffer", required to isolate scheduling/application
458 * latencies from network.
459 * window_clamp is maximal advertised window. It can be less than
460 * tcp_full_space(), in this case tcp_full_space() - window_clamp
461 * is reserved for "application" buffer. The less window_clamp is
462 * the smoother our behaviour from viewpoint of network, but the lower
463 * throughput and the higher sensitivity of the connection to losses. 8)
464 *
465 * rcv_ssthresh is more strict window_clamp used at "slow start"
466 * phase to predict further behaviour of this connection.
467 * It is used for two goals:
468 * - to enforce header prediction at sender, even when application
469 * requires some significant "application buffer". It is check #1.
470 * - to prevent pruning of receive queue because of misprediction
471 * of receiver window. Check #2.
472 *
473 * The scheme does not work when sender sends good segments opening
474 * window and then starts to feed us spaghetti. But it should work
475 * in common situations. Otherwise, we have to rely on queue collapsing.
476 */
477
478 /* Slow part of check#2. */
479 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb,
480 unsigned int skbtruesize)
481 {
482 const struct tcp_sock *tp = tcp_sk(sk);
483 /* Optimize this! */
484 int truesize = tcp_win_from_space(sk, skbtruesize) >> 1;
485 int window = tcp_win_from_space(sk, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2])) >> 1;
486
487 while (tp->rcv_ssthresh <= window) {
488 if (truesize <= skb->len)
489 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
490
491 truesize >>= 1;
492 window >>= 1;
493 }
494 return 0;
495 }
496
497 /* Even if skb appears to have a bad len/truesize ratio, TCP coalescing
498 * can play nice with us, as sk_buff and skb->head might be either
499 * freed or shared with up to MAX_SKB_FRAGS segments.
500 * Only give a boost to drivers using page frag(s) to hold the frame(s),
501 * and if no payload was pulled in skb->head before reaching us.
502 */
503 static u32 truesize_adjust(bool adjust, const struct sk_buff *skb)
504 {
505 u32 truesize = skb->truesize;
506
507 if (adjust && !skb_headlen(skb)) {
508 truesize -= SKB_TRUESIZE(skb_end_offset(skb));
509 /* paranoid check, some drivers might be buggy */
510 if (unlikely((int)truesize < (int)skb->len))
511 truesize = skb->truesize;
512 }
513 return truesize;
514 }
515
516 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb,
517 bool adjust)
518 {
519 struct tcp_sock *tp = tcp_sk(sk);
520 int room;
521
522 room = min_t(int, tp->window_clamp, tcp_space(sk)) - tp->rcv_ssthresh;
523
524 if (room <= 0)
525 return;
526
527 /* Check #1 */
528 if (!tcp_under_memory_pressure(sk)) {
529 unsigned int truesize = truesize_adjust(adjust, skb);
530 int incr;
531
532 /* Check #2. Increase window, if skb with such overhead
533 * will fit to rcvbuf in future.
534 */
535 if (tcp_win_from_space(sk, truesize) <= skb->len)
536 incr = 2 * tp->advmss;
537 else
538 incr = __tcp_grow_window(sk, skb, truesize);
539
540 if (incr) {
541 incr = max_t(int, incr, 2 * skb->len);
542 tp->rcv_ssthresh += min(room, incr);
543 inet_csk(sk)->icsk_ack.quick |= 1;
544 }
545 } else {
546 /* Under pressure:
547 * Adjust rcv_ssthresh according to reserved mem
548 */
549 tcp_adjust_rcv_ssthresh(sk);
550 }
551 }
552
553 /* 3. Try to fixup all. It is made immediately after connection enters
554 * established state.
555 */
556 static void tcp_init_buffer_space(struct sock *sk)
557 {
558 int tcp_app_win = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_app_win);
559 struct tcp_sock *tp = tcp_sk(sk);
560 int maxwin;
561
562 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
563 tcp_sndbuf_expand(sk);
564
565 tcp_mstamp_refresh(tp);
566 tp->rcvq_space.time = tp->tcp_mstamp;
567 tp->rcvq_space.seq = tp->copied_seq;
568
569 maxwin = tcp_full_space(sk);
570
571 if (tp->window_clamp >= maxwin) {
572 WRITE_ONCE(tp->window_clamp, maxwin);
573
574 if (tcp_app_win && maxwin > 4 * tp->advmss)
575 WRITE_ONCE(tp->window_clamp,
576 max(maxwin - (maxwin >> tcp_app_win),
577 4 * tp->advmss));
578 }
579
580 /* Force reservation of one segment. */
581 if (tcp_app_win &&
582 tp->window_clamp > 2 * tp->advmss &&
583 tp->window_clamp + tp->advmss > maxwin)
584 WRITE_ONCE(tp->window_clamp,
585 max(2 * tp->advmss, maxwin - tp->advmss));
586
587 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
588 tp->snd_cwnd_stamp = tcp_jiffies32;
589 tp->rcvq_space.space = min3(tp->rcv_ssthresh, tp->rcv_wnd,
590 (u32)TCP_INIT_CWND * tp->advmss);
591 }
592
593 /* 4. Recalculate window clamp after socket hit its memory bounds. */
594 static void tcp_clamp_window(struct sock *sk)
595 {
596 struct tcp_sock *tp = tcp_sk(sk);
597 struct inet_connection_sock *icsk = inet_csk(sk);
598 struct net *net = sock_net(sk);
599 int rmem2;
600
601 icsk->icsk_ack.quick = 0;
602 rmem2 = READ_ONCE(net->ipv4.sysctl_tcp_rmem[2]);
603
604 if (sk->sk_rcvbuf < rmem2 &&
605 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
606 !tcp_under_memory_pressure(sk) &&
607 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
608 WRITE_ONCE(sk->sk_rcvbuf,
609 min(atomic_read(&sk->sk_rmem_alloc), rmem2));
610 }
611 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
612 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
613 }
614
615 /* Initialize RCV_MSS value.
616 * RCV_MSS is an our guess about MSS used by the peer.
617 * We haven't any direct information about the MSS.
618 * It's better to underestimate the RCV_MSS rather than overestimate.
619 * Overestimations make us ACKing less frequently than needed.
620 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
621 */
622 void tcp_initialize_rcv_mss(struct sock *sk)
623 {
624 const struct tcp_sock *tp = tcp_sk(sk);
625 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
626
627 hint = min(hint, tp->rcv_wnd / 2);
628 hint = min(hint, TCP_MSS_DEFAULT);
629 hint = max(hint, TCP_MIN_MSS);
630
631 inet_csk(sk)->icsk_ack.rcv_mss = hint;
632 }
633 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
634
635 /* Receiver "autotuning" code.
636 *
637 * The algorithm for RTT estimation w/o timestamps is based on
638 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
639 * <https://public.lanl.gov/radiant/pubs.html#DRS>
640 *
641 * More detail on this code can be found at
642 * <http://staff.psc.edu/jheffner/>,
643 * though this reference is out of date. A new paper
644 * is pending.
645 */
646 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
647 {
648 u32 new_sample = tp->rcv_rtt_est.rtt_us;
649 long m = sample;
650
651 if (new_sample != 0) {
652 /* If we sample in larger samples in the non-timestamp
653 * case, we could grossly overestimate the RTT especially
654 * with chatty applications or bulk transfer apps which
655 * are stalled on filesystem I/O.
656 *
657 * Also, since we are only going for a minimum in the
658 * non-timestamp case, we do not smooth things out
659 * else with timestamps disabled convergence takes too
660 * long.
661 */
662 if (!win_dep) {
663 m -= (new_sample >> 3);
664 new_sample += m;
665 } else {
666 m <<= 3;
667 if (m < new_sample)
668 new_sample = m;
669 }
670 } else {
671 /* No previous measure. */
672 new_sample = m << 3;
673 }
674
675 tp->rcv_rtt_est.rtt_us = new_sample;
676 }
677
678 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
679 {
680 u32 delta_us;
681
682 if (tp->rcv_rtt_est.time == 0)
683 goto new_measure;
684 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
685 return;
686 delta_us = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcv_rtt_est.time);
687 if (!delta_us)
688 delta_us = 1;
689 tcp_rcv_rtt_update(tp, delta_us, 1);
690
691 new_measure:
692 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
693 tp->rcv_rtt_est.time = tp->tcp_mstamp;
694 }
695
696 static s32 tcp_rtt_tsopt_us(const struct tcp_sock *tp)
697 {
698 u32 delta, delta_us;
699
700 delta = tcp_time_stamp_ts(tp) - tp->rx_opt.rcv_tsecr;
701 if (tp->tcp_usec_ts)
702 return delta;
703
704 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) {
705 if (!delta)
706 delta = 1;
707 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
708 return delta_us;
709 }
710 return -1;
711 }
712
713 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
714 const struct sk_buff *skb)
715 {
716 struct tcp_sock *tp = tcp_sk(sk);
717
718 if (tp->rx_opt.rcv_tsecr == tp->rcv_rtt_last_tsecr)
719 return;
720 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
721
722 if (TCP_SKB_CB(skb)->end_seq -
723 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss) {
724 s32 delta = tcp_rtt_tsopt_us(tp);
725
726 if (delta >= 0)
727 tcp_rcv_rtt_update(tp, delta, 0);
728 }
729 }
730
731 /*
732 * This function should be called every time data is copied to user space.
733 * It calculates the appropriate TCP receive buffer space.
734 */
735 void tcp_rcv_space_adjust(struct sock *sk)
736 {
737 struct tcp_sock *tp = tcp_sk(sk);
738 u32 copied;
739 int time;
740
741 trace_tcp_rcv_space_adjust(sk);
742
743 tcp_mstamp_refresh(tp);
744 time = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcvq_space.time);
745 if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0)
746 return;
747
748 /* Number of bytes copied to user in last RTT */
749 copied = tp->copied_seq - tp->rcvq_space.seq;
750 if (copied <= tp->rcvq_space.space)
751 goto new_measure;
752
753 /* A bit of theory :
754 * copied = bytes received in previous RTT, our base window
755 * To cope with packet losses, we need a 2x factor
756 * To cope with slow start, and sender growing its cwin by 100 %
757 * every RTT, we need a 4x factor, because the ACK we are sending
758 * now is for the next RTT, not the current one :
759 * <prev RTT . ><current RTT .. ><next RTT .... >
760 */
761
762 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf) &&
763 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
764 u64 rcvwin, grow;
765 int rcvbuf;
766
767 /* minimal window to cope with packet losses, assuming
768 * steady state. Add some cushion because of small variations.
769 */
770 rcvwin = ((u64)copied << 1) + 16 * tp->advmss;
771
772 /* Accommodate for sender rate increase (eg. slow start) */
773 grow = rcvwin * (copied - tp->rcvq_space.space);
774 do_div(grow, tp->rcvq_space.space);
775 rcvwin += (grow << 1);
776
777 rcvbuf = min_t(u64, tcp_space_from_win(sk, rcvwin),
778 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
779 if (rcvbuf > sk->sk_rcvbuf) {
780 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
781
782 /* Make the window clamp follow along. */
783 WRITE_ONCE(tp->window_clamp,
784 tcp_win_from_space(sk, rcvbuf));
785 }
786 }
787 tp->rcvq_space.space = copied;
788
789 new_measure:
790 tp->rcvq_space.seq = tp->copied_seq;
791 tp->rcvq_space.time = tp->tcp_mstamp;
792 }
793
794 static void tcp_save_lrcv_flowlabel(struct sock *sk, const struct sk_buff *skb)
795 {
796 #if IS_ENABLED(CONFIG_IPV6)
797 struct inet_connection_sock *icsk = inet_csk(sk);
798
799 if (skb->protocol == htons(ETH_P_IPV6))
800 icsk->icsk_ack.lrcv_flowlabel = ntohl(ip6_flowlabel(ipv6_hdr(skb)));
801 #endif
802 }
803
804 /* There is something which you must keep in mind when you analyze the
805 * behavior of the tp->ato delayed ack timeout interval. When a
806 * connection starts up, we want to ack as quickly as possible. The
807 * problem is that "good" TCP's do slow start at the beginning of data
808 * transmission. The means that until we send the first few ACK's the
809 * sender will sit on his end and only queue most of his data, because
810 * he can only send snd_cwnd unacked packets at any given time. For
811 * each ACK we send, he increments snd_cwnd and transmits more of his
812 * queue. -DaveM
813 */
814 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
815 {
816 struct tcp_sock *tp = tcp_sk(sk);
817 struct inet_connection_sock *icsk = inet_csk(sk);
818 u32 now;
819
820 inet_csk_schedule_ack(sk);
821
822 tcp_measure_rcv_mss(sk, skb);
823
824 tcp_rcv_rtt_measure(tp);
825
826 now = tcp_jiffies32;
827
828 if (!icsk->icsk_ack.ato) {
829 /* The _first_ data packet received, initialize
830 * delayed ACK engine.
831 */
832 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
833 icsk->icsk_ack.ato = TCP_ATO_MIN;
834 } else {
835 int m = now - icsk->icsk_ack.lrcvtime;
836
837 if (m <= TCP_ATO_MIN / 2) {
838 /* The fastest case is the first. */
839 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
840 } else if (m < icsk->icsk_ack.ato) {
841 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
842 if (icsk->icsk_ack.ato > icsk->icsk_rto)
843 icsk->icsk_ack.ato = icsk->icsk_rto;
844 } else if (m > icsk->icsk_rto) {
845 /* Too long gap. Apparently sender failed to
846 * restart window, so that we send ACKs quickly.
847 */
848 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
849 }
850 }
851 icsk->icsk_ack.lrcvtime = now;
852 tcp_save_lrcv_flowlabel(sk, skb);
853
854 tcp_ecn_check_ce(sk, skb);
855
856 if (skb->len >= 128)
857 tcp_grow_window(sk, skb, true);
858 }
859
860 /* Called to compute a smoothed rtt estimate. The data fed to this
861 * routine either comes from timestamps, or from segments that were
862 * known _not_ to have been retransmitted [see Karn/Partridge
863 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
864 * piece by Van Jacobson.
865 * NOTE: the next three routines used to be one big routine.
866 * To save cycles in the RFC 1323 implementation it was better to break
867 * it up into three procedures. -- erics
868 */
869 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
870 {
871 struct tcp_sock *tp = tcp_sk(sk);
872 long m = mrtt_us; /* RTT */
873 u32 srtt = tp->srtt_us;
874
875 /* The following amusing code comes from Jacobson's
876 * article in SIGCOMM '88. Note that rtt and mdev
877 * are scaled versions of rtt and mean deviation.
878 * This is designed to be as fast as possible
879 * m stands for "measurement".
880 *
881 * On a 1990 paper the rto value is changed to:
882 * RTO = rtt + 4 * mdev
883 *
884 * Funny. This algorithm seems to be very broken.
885 * These formulae increase RTO, when it should be decreased, increase
886 * too slowly, when it should be increased quickly, decrease too quickly
887 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
888 * does not matter how to _calculate_ it. Seems, it was trap
889 * that VJ failed to avoid. 8)
890 */
891 if (srtt != 0) {
892 m -= (srtt >> 3); /* m is now error in rtt est */
893 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
894 if (m < 0) {
895 m = -m; /* m is now abs(error) */
896 m -= (tp->mdev_us >> 2); /* similar update on mdev */
897 /* This is similar to one of Eifel findings.
898 * Eifel blocks mdev updates when rtt decreases.
899 * This solution is a bit different: we use finer gain
900 * for mdev in this case (alpha*beta).
901 * Like Eifel it also prevents growth of rto,
902 * but also it limits too fast rto decreases,
903 * happening in pure Eifel.
904 */
905 if (m > 0)
906 m >>= 3;
907 } else {
908 m -= (tp->mdev_us >> 2); /* similar update on mdev */
909 }
910 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
911 if (tp->mdev_us > tp->mdev_max_us) {
912 tp->mdev_max_us = tp->mdev_us;
913 if (tp->mdev_max_us > tp->rttvar_us)
914 tp->rttvar_us = tp->mdev_max_us;
915 }
916 if (after(tp->snd_una, tp->rtt_seq)) {
917 if (tp->mdev_max_us < tp->rttvar_us)
918 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
919 tp->rtt_seq = tp->snd_nxt;
920 tp->mdev_max_us = tcp_rto_min_us(sk);
921
922 tcp_bpf_rtt(sk, mrtt_us, srtt);
923 }
924 } else {
925 /* no previous measure. */
926 srtt = m << 3; /* take the measured time to be rtt */
927 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
928 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
929 tp->mdev_max_us = tp->rttvar_us;
930 tp->rtt_seq = tp->snd_nxt;
931
932 tcp_bpf_rtt(sk, mrtt_us, srtt);
933 }
934 tp->srtt_us = max(1U, srtt);
935 }
936
937 static void tcp_update_pacing_rate(struct sock *sk)
938 {
939 const struct tcp_sock *tp = tcp_sk(sk);
940 u64 rate;
941
942 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
943 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
944
945 /* current rate is (cwnd * mss) / srtt
946 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
947 * In Congestion Avoidance phase, set it to 120 % the current rate.
948 *
949 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
950 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
951 * end of slow start and should slow down.
952 */
953 if (tcp_snd_cwnd(tp) < tp->snd_ssthresh / 2)
954 rate *= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_pacing_ss_ratio);
955 else
956 rate *= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_pacing_ca_ratio);
957
958 rate *= max(tcp_snd_cwnd(tp), tp->packets_out);
959
960 if (likely(tp->srtt_us))
961 do_div(rate, tp->srtt_us);
962
963 /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate
964 * without any lock. We want to make sure compiler wont store
965 * intermediate values in this location.
966 */
967 WRITE_ONCE(sk->sk_pacing_rate,
968 min_t(u64, rate, READ_ONCE(sk->sk_max_pacing_rate)));
969 }
970
971 /* Calculate rto without backoff. This is the second half of Van Jacobson's
972 * routine referred to above.
973 */
974 static void tcp_set_rto(struct sock *sk)
975 {
976 const struct tcp_sock *tp = tcp_sk(sk);
977 /* Old crap is replaced with new one. 8)
978 *
979 * More seriously:
980 * 1. If rtt variance happened to be less 50msec, it is hallucination.
981 * It cannot be less due to utterly erratic ACK generation made
982 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
983 * to do with delayed acks, because at cwnd>2 true delack timeout
984 * is invisible. Actually, Linux-2.4 also generates erratic
985 * ACKs in some circumstances.
986 */
987 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
988
989 /* 2. Fixups made earlier cannot be right.
990 * If we do not estimate RTO correctly without them,
991 * all the algo is pure shit and should be replaced
992 * with correct one. It is exactly, which we pretend to do.
993 */
994
995 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
996 * guarantees that rto is higher.
997 */
998 tcp_bound_rto(sk);
999 }
1000
1001 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
1002 {
1003 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
1004
1005 if (!cwnd)
1006 cwnd = TCP_INIT_CWND;
1007 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
1008 }
1009
1010 struct tcp_sacktag_state {
1011 /* Timestamps for earliest and latest never-retransmitted segment
1012 * that was SACKed. RTO needs the earliest RTT to stay conservative,
1013 * but congestion control should still get an accurate delay signal.
1014 */
1015 u64 first_sackt;
1016 u64 last_sackt;
1017 u32 reord;
1018 u32 sack_delivered;
1019 int flag;
1020 unsigned int mss_now;
1021 struct rate_sample *rate;
1022 };
1023
1024 /* Take a notice that peer is sending D-SACKs. Skip update of data delivery
1025 * and spurious retransmission information if this DSACK is unlikely caused by
1026 * sender's action:
1027 * - DSACKed sequence range is larger than maximum receiver's window.
1028 * - Total no. of DSACKed segments exceed the total no. of retransmitted segs.
1029 */
1030 static u32 tcp_dsack_seen(struct tcp_sock *tp, u32 start_seq,
1031 u32 end_seq, struct tcp_sacktag_state *state)
1032 {
1033 u32 seq_len, dup_segs = 1;
1034
1035 if (!before(start_seq, end_seq))
1036 return 0;
1037
1038 seq_len = end_seq - start_seq;
1039 /* Dubious DSACK: DSACKed range greater than maximum advertised rwnd */
1040 if (seq_len > tp->max_window)
1041 return 0;
1042 if (seq_len > tp->mss_cache)
1043 dup_segs = DIV_ROUND_UP(seq_len, tp->mss_cache);
1044 else if (tp->tlp_high_seq && tp->tlp_high_seq == end_seq)
1045 state->flag |= FLAG_DSACK_TLP;
1046
1047 tp->dsack_dups += dup_segs;
1048 /* Skip the DSACK if dup segs weren't retransmitted by sender */
1049 if (tp->dsack_dups > tp->total_retrans)
1050 return 0;
1051
1052 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
1053 /* We increase the RACK ordering window in rounds where we receive
1054 * DSACKs that may have been due to reordering causing RACK to trigger
1055 * a spurious fast recovery. Thus RACK ignores DSACKs that happen
1056 * without having seen reordering, or that match TLP probes (TLP
1057 * is timer-driven, not triggered by RACK).
1058 */
1059 if (tp->reord_seen && !(state->flag & FLAG_DSACK_TLP))
1060 tp->rack.dsack_seen = 1;
1061
1062 state->flag |= FLAG_DSACKING_ACK;
1063 /* A spurious retransmission is delivered */
1064 state->sack_delivered += dup_segs;
1065
1066 return dup_segs;
1067 }
1068
1069 /* It's reordering when higher sequence was delivered (i.e. sacked) before
1070 * some lower never-retransmitted sequence ("low_seq"). The maximum reordering
1071 * distance is approximated in full-mss packet distance ("reordering").
1072 */
1073 static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq,
1074 const int ts)
1075 {
1076 struct tcp_sock *tp = tcp_sk(sk);
1077 const u32 mss = tp->mss_cache;
1078 u32 fack, metric;
1079
1080 fack = tcp_highest_sack_seq(tp);
1081 if (!before(low_seq, fack))
1082 return;
1083
1084 metric = fack - low_seq;
1085 if ((metric > tp->reordering * mss) && mss) {
1086 #if FASTRETRANS_DEBUG > 1
1087 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
1088 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
1089 tp->reordering,
1090 0,
1091 tp->sacked_out,
1092 tp->undo_marker ? tp->undo_retrans : 0);
1093 #endif
1094 tp->reordering = min_t(u32, (metric + mss - 1) / mss,
1095 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering));
1096 }
1097
1098 /* This exciting event is worth to be remembered. 8) */
1099 tp->reord_seen++;
1100 NET_INC_STATS(sock_net(sk),
1101 ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER);
1102 }
1103
1104 /* This must be called before lost_out or retrans_out are updated
1105 * on a new loss, because we want to know if all skbs previously
1106 * known to be lost have already been retransmitted, indicating
1107 * that this newly lost skb is our next skb to retransmit.
1108 */
1109 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
1110 {
1111 if ((!tp->retransmit_skb_hint && tp->retrans_out >= tp->lost_out) ||
1112 (tp->retransmit_skb_hint &&
1113 before(TCP_SKB_CB(skb)->seq,
1114 TCP_SKB_CB(tp->retransmit_skb_hint)->seq)))
1115 tp->retransmit_skb_hint = skb;
1116 }
1117
1118 /* Sum the number of packets on the wire we have marked as lost, and
1119 * notify the congestion control module that the given skb was marked lost.
1120 */
1121 static void tcp_notify_skb_loss_event(struct tcp_sock *tp, const struct sk_buff *skb)
1122 {
1123 tp->lost += tcp_skb_pcount(skb);
1124 }
1125
1126 void tcp_mark_skb_lost(struct sock *sk, struct sk_buff *skb)
1127 {
1128 __u8 sacked = TCP_SKB_CB(skb)->sacked;
1129 struct tcp_sock *tp = tcp_sk(sk);
1130
1131 if (sacked & TCPCB_SACKED_ACKED)
1132 return;
1133
1134 tcp_verify_retransmit_hint(tp, skb);
1135 if (sacked & TCPCB_LOST) {
1136 if (sacked & TCPCB_SACKED_RETRANS) {
1137 /* Account for retransmits that are lost again */
1138 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1139 tp->retrans_out -= tcp_skb_pcount(skb);
1140 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT,
1141 tcp_skb_pcount(skb));
1142 tcp_notify_skb_loss_event(tp, skb);
1143 }
1144 } else {
1145 tp->lost_out += tcp_skb_pcount(skb);
1146 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1147 tcp_notify_skb_loss_event(tp, skb);
1148 }
1149 }
1150
1151 /* Updates the delivered and delivered_ce counts */
1152 static void tcp_count_delivered(struct tcp_sock *tp, u32 delivered,
1153 bool ece_ack)
1154 {
1155 tp->delivered += delivered;
1156 if (ece_ack)
1157 tp->delivered_ce += delivered;
1158 }
1159
1160 /* This procedure tags the retransmission queue when SACKs arrive.
1161 *
1162 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1163 * Packets in queue with these bits set are counted in variables
1164 * sacked_out, retrans_out and lost_out, correspondingly.
1165 *
1166 * Valid combinations are:
1167 * Tag InFlight Description
1168 * 0 1 - orig segment is in flight.
1169 * S 0 - nothing flies, orig reached receiver.
1170 * L 0 - nothing flies, orig lost by net.
1171 * R 2 - both orig and retransmit are in flight.
1172 * L|R 1 - orig is lost, retransmit is in flight.
1173 * S|R 1 - orig reached receiver, retrans is still in flight.
1174 * (L|S|R is logically valid, it could occur when L|R is sacked,
1175 * but it is equivalent to plain S and code short-circuits it to S.
1176 * L|S is logically invalid, it would mean -1 packet in flight 8))
1177 *
1178 * These 6 states form finite state machine, controlled by the following events:
1179 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1180 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1181 * 3. Loss detection event of two flavors:
1182 * A. Scoreboard estimator decided the packet is lost.
1183 * A'. Reno "three dupacks" marks head of queue lost.
1184 * B. SACK arrives sacking SND.NXT at the moment, when the
1185 * segment was retransmitted.
1186 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1187 *
1188 * It is pleasant to note, that state diagram turns out to be commutative,
1189 * so that we are allowed not to be bothered by order of our actions,
1190 * when multiple events arrive simultaneously. (see the function below).
1191 *
1192 * Reordering detection.
1193 * --------------------
1194 * Reordering metric is maximal distance, which a packet can be displaced
1195 * in packet stream. With SACKs we can estimate it:
1196 *
1197 * 1. SACK fills old hole and the corresponding segment was not
1198 * ever retransmitted -> reordering. Alas, we cannot use it
1199 * when segment was retransmitted.
1200 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1201 * for retransmitted and already SACKed segment -> reordering..
1202 * Both of these heuristics are not used in Loss state, when we cannot
1203 * account for retransmits accurately.
1204 *
1205 * SACK block validation.
1206 * ----------------------
1207 *
1208 * SACK block range validation checks that the received SACK block fits to
1209 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1210 * Note that SND.UNA is not included to the range though being valid because
1211 * it means that the receiver is rather inconsistent with itself reporting
1212 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1213 * perfectly valid, however, in light of RFC2018 which explicitly states
1214 * that "SACK block MUST reflect the newest segment. Even if the newest
1215 * segment is going to be discarded ...", not that it looks very clever
1216 * in case of head skb. Due to potentional receiver driven attacks, we
1217 * choose to avoid immediate execution of a walk in write queue due to
1218 * reneging and defer head skb's loss recovery to standard loss recovery
1219 * procedure that will eventually trigger (nothing forbids us doing this).
1220 *
1221 * Implements also blockage to start_seq wrap-around. Problem lies in the
1222 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1223 * there's no guarantee that it will be before snd_nxt (n). The problem
1224 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1225 * wrap (s_w):
1226 *
1227 * <- outs wnd -> <- wrapzone ->
1228 * u e n u_w e_w s n_w
1229 * | | | | | | |
1230 * |<------------+------+----- TCP seqno space --------------+---------->|
1231 * ...-- <2^31 ->| |<--------...
1232 * ...---- >2^31 ------>| |<--------...
1233 *
1234 * Current code wouldn't be vulnerable but it's better still to discard such
1235 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1236 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1237 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1238 * equal to the ideal case (infinite seqno space without wrap caused issues).
1239 *
1240 * With D-SACK the lower bound is extended to cover sequence space below
1241 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1242 * again, D-SACK block must not to go across snd_una (for the same reason as
1243 * for the normal SACK blocks, explained above). But there all simplicity
1244 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1245 * fully below undo_marker they do not affect behavior in anyway and can
1246 * therefore be safely ignored. In rare cases (which are more or less
1247 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1248 * fragmentation and packet reordering past skb's retransmission. To consider
1249 * them correctly, the acceptable range must be extended even more though
1250 * the exact amount is rather hard to quantify. However, tp->max_window can
1251 * be used as an exaggerated estimate.
1252 */
1253 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1254 u32 start_seq, u32 end_seq)
1255 {
1256 /* Too far in future, or reversed (interpretation is ambiguous) */
1257 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1258 return false;
1259
1260 /* Nasty start_seq wrap-around check (see comments above) */
1261 if (!before(start_seq, tp->snd_nxt))
1262 return false;
1263
1264 /* In outstanding window? ...This is valid exit for D-SACKs too.
1265 * start_seq == snd_una is non-sensical (see comments above)
1266 */
1267 if (after(start_seq, tp->snd_una))
1268 return true;
1269
1270 if (!is_dsack || !tp->undo_marker)
1271 return false;
1272
1273 /* ...Then it's D-SACK, and must reside below snd_una completely */
1274 if (after(end_seq, tp->snd_una))
1275 return false;
1276
1277 if (!before(start_seq, tp->undo_marker))
1278 return true;
1279
1280 /* Too old */
1281 if (!after(end_seq, tp->undo_marker))
1282 return false;
1283
1284 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1285 * start_seq < undo_marker and end_seq >= undo_marker.
1286 */
1287 return !before(start_seq, end_seq - tp->max_window);
1288 }
1289
1290 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1291 struct tcp_sack_block_wire *sp, int num_sacks,
1292 u32 prior_snd_una, struct tcp_sacktag_state *state)
1293 {
1294 struct tcp_sock *tp = tcp_sk(sk);
1295 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1296 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1297 u32 dup_segs;
1298
1299 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1300 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1301 } else if (num_sacks > 1) {
1302 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1303 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1304
1305 if (after(end_seq_0, end_seq_1) || before(start_seq_0, start_seq_1))
1306 return false;
1307 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKOFORECV);
1308 } else {
1309 return false;
1310 }
1311
1312 dup_segs = tcp_dsack_seen(tp, start_seq_0, end_seq_0, state);
1313 if (!dup_segs) { /* Skip dubious DSACK */
1314 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKIGNOREDDUBIOUS);
1315 return false;
1316 }
1317
1318 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECVSEGS, dup_segs);
1319
1320 /* D-SACK for already forgotten data... Do dumb counting. */
1321 if (tp->undo_marker && tp->undo_retrans > 0 &&
1322 !after(end_seq_0, prior_snd_una) &&
1323 after(end_seq_0, tp->undo_marker))
1324 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - dup_segs);
1325
1326 return true;
1327 }
1328
1329 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1330 * the incoming SACK may not exactly match but we can find smaller MSS
1331 * aligned portion of it that matches. Therefore we might need to fragment
1332 * which may fail and creates some hassle (caller must handle error case
1333 * returns).
1334 *
1335 * FIXME: this could be merged to shift decision code
1336 */
1337 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1338 u32 start_seq, u32 end_seq)
1339 {
1340 int err;
1341 bool in_sack;
1342 unsigned int pkt_len;
1343 unsigned int mss;
1344
1345 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1346 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1347
1348 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1349 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1350 mss = tcp_skb_mss(skb);
1351 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1352
1353 if (!in_sack) {
1354 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1355 if (pkt_len < mss)
1356 pkt_len = mss;
1357 } else {
1358 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1359 if (pkt_len < mss)
1360 return -EINVAL;
1361 }
1362
1363 /* Round if necessary so that SACKs cover only full MSSes
1364 * and/or the remaining small portion (if present)
1365 */
1366 if (pkt_len > mss) {
1367 unsigned int new_len = (pkt_len / mss) * mss;
1368 if (!in_sack && new_len < pkt_len)
1369 new_len += mss;
1370 pkt_len = new_len;
1371 }
1372
1373 if (pkt_len >= skb->len && !in_sack)
1374 return 0;
1375
1376 err = tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
1377 pkt_len, mss, GFP_ATOMIC);
1378 if (err < 0)
1379 return err;
1380 }
1381
1382 return in_sack;
1383 }
1384
1385 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1386 static u8 tcp_sacktag_one(struct sock *sk,
1387 struct tcp_sacktag_state *state, u8 sacked,
1388 u32 start_seq, u32 end_seq,
1389 int dup_sack, int pcount,
1390 u64 xmit_time)
1391 {
1392 struct tcp_sock *tp = tcp_sk(sk);
1393
1394 /* Account D-SACK for retransmitted packet. */
1395 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1396 if (tp->undo_marker && tp->undo_retrans > 0 &&
1397 after(end_seq, tp->undo_marker))
1398 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - pcount);
1399 if ((sacked & TCPCB_SACKED_ACKED) &&
1400 before(start_seq, state->reord))
1401 state->reord = start_seq;
1402 }
1403
1404 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1405 if (!after(end_seq, tp->snd_una))
1406 return sacked;
1407
1408 if (!(sacked & TCPCB_SACKED_ACKED)) {
1409 tcp_rack_advance(tp, sacked, end_seq, xmit_time);
1410
1411 if (sacked & TCPCB_SACKED_RETRANS) {
1412 /* If the segment is not tagged as lost,
1413 * we do not clear RETRANS, believing
1414 * that retransmission is still in flight.
1415 */
1416 if (sacked & TCPCB_LOST) {
1417 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1418 tp->lost_out -= pcount;
1419 tp->retrans_out -= pcount;
1420 }
1421 } else {
1422 if (!(sacked & TCPCB_RETRANS)) {
1423 /* New sack for not retransmitted frame,
1424 * which was in hole. It is reordering.
1425 */
1426 if (before(start_seq,
1427 tcp_highest_sack_seq(tp)) &&
1428 before(start_seq, state->reord))
1429 state->reord = start_seq;
1430
1431 if (!after(end_seq, tp->high_seq))
1432 state->flag |= FLAG_ORIG_SACK_ACKED;
1433 if (state->first_sackt == 0)
1434 state->first_sackt = xmit_time;
1435 state->last_sackt = xmit_time;
1436 }
1437
1438 if (sacked & TCPCB_LOST) {
1439 sacked &= ~TCPCB_LOST;
1440 tp->lost_out -= pcount;
1441 }
1442 }
1443
1444 sacked |= TCPCB_SACKED_ACKED;
1445 state->flag |= FLAG_DATA_SACKED;
1446 tp->sacked_out += pcount;
1447 /* Out-of-order packets delivered */
1448 state->sack_delivered += pcount;
1449
1450 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1451 if (tp->lost_skb_hint &&
1452 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1453 tp->lost_cnt_hint += pcount;
1454 }
1455
1456 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1457 * frames and clear it. undo_retrans is decreased above, L|R frames
1458 * are accounted above as well.
1459 */
1460 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1461 sacked &= ~TCPCB_SACKED_RETRANS;
1462 tp->retrans_out -= pcount;
1463 }
1464
1465 return sacked;
1466 }
1467
1468 /* Shift newly-SACKed bytes from this skb to the immediately previous
1469 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1470 */
1471 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev,
1472 struct sk_buff *skb,
1473 struct tcp_sacktag_state *state,
1474 unsigned int pcount, int shifted, int mss,
1475 bool dup_sack)
1476 {
1477 struct tcp_sock *tp = tcp_sk(sk);
1478 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1479 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1480
1481 BUG_ON(!pcount);
1482
1483 /* Adjust counters and hints for the newly sacked sequence
1484 * range but discard the return value since prev is already
1485 * marked. We must tag the range first because the seq
1486 * advancement below implicitly advances
1487 * tcp_highest_sack_seq() when skb is highest_sack.
1488 */
1489 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1490 start_seq, end_seq, dup_sack, pcount,
1491 tcp_skb_timestamp_us(skb));
1492 tcp_rate_skb_delivered(sk, skb, state->rate);
1493
1494 if (skb == tp->lost_skb_hint)
1495 tp->lost_cnt_hint += pcount;
1496
1497 TCP_SKB_CB(prev)->end_seq += shifted;
1498 TCP_SKB_CB(skb)->seq += shifted;
1499
1500 tcp_skb_pcount_add(prev, pcount);
1501 WARN_ON_ONCE(tcp_skb_pcount(skb) < pcount);
1502 tcp_skb_pcount_add(skb, -pcount);
1503
1504 /* When we're adding to gso_segs == 1, gso_size will be zero,
1505 * in theory this shouldn't be necessary but as long as DSACK
1506 * code can come after this skb later on it's better to keep
1507 * setting gso_size to something.
1508 */
1509 if (!TCP_SKB_CB(prev)->tcp_gso_size)
1510 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1511
1512 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1513 if (tcp_skb_pcount(skb) <= 1)
1514 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1515
1516 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1517 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1518
1519 if (skb->len > 0) {
1520 BUG_ON(!tcp_skb_pcount(skb));
1521 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1522 return false;
1523 }
1524
1525 /* Whole SKB was eaten :-) */
1526
1527 if (skb == tp->retransmit_skb_hint)
1528 tp->retransmit_skb_hint = prev;
1529 if (skb == tp->lost_skb_hint) {
1530 tp->lost_skb_hint = prev;
1531 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1532 }
1533
1534 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1535 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor;
1536 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1537 TCP_SKB_CB(prev)->end_seq++;
1538
1539 if (skb == tcp_highest_sack(sk))
1540 tcp_advance_highest_sack(sk, skb);
1541
1542 tcp_skb_collapse_tstamp(prev, skb);
1543 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp))
1544 TCP_SKB_CB(prev)->tx.delivered_mstamp = 0;
1545
1546 tcp_rtx_queue_unlink_and_free(skb, sk);
1547
1548 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED);
1549
1550 return true;
1551 }
1552
1553 /* I wish gso_size would have a bit more sane initialization than
1554 * something-or-zero which complicates things
1555 */
1556 static int tcp_skb_seglen(const struct sk_buff *skb)
1557 {
1558 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1559 }
1560
1561 /* Shifting pages past head area doesn't work */
1562 static int skb_can_shift(const struct sk_buff *skb)
1563 {
1564 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1565 }
1566
1567 int tcp_skb_shift(struct sk_buff *to, struct sk_buff *from,
1568 int pcount, int shiftlen)
1569 {
1570 /* TCP min gso_size is 8 bytes (TCP_MIN_GSO_SIZE)
1571 * Since TCP_SKB_CB(skb)->tcp_gso_segs is 16 bits, we need
1572 * to make sure not storing more than 65535 * 8 bytes per skb,
1573 * even if current MSS is bigger.
1574 */
1575 if (unlikely(to->len + shiftlen >= 65535 * TCP_MIN_GSO_SIZE))
1576 return 0;
1577 if (unlikely(tcp_skb_pcount(to) + pcount > 65535))
1578 return 0;
1579 return skb_shift(to, from, shiftlen);
1580 }
1581
1582 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1583 * skb.
1584 */
1585 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1586 struct tcp_sacktag_state *state,
1587 u32 start_seq, u32 end_seq,
1588 bool dup_sack)
1589 {
1590 struct tcp_sock *tp = tcp_sk(sk);
1591 struct sk_buff *prev;
1592 int mss;
1593 int pcount = 0;
1594 int len;
1595 int in_sack;
1596
1597 /* Normally R but no L won't result in plain S */
1598 if (!dup_sack &&
1599 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1600 goto fallback;
1601 if (!skb_can_shift(skb))
1602 goto fallback;
1603 /* This frame is about to be dropped (was ACKed). */
1604 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1605 goto fallback;
1606
1607 /* Can only happen with delayed DSACK + discard craziness */
1608 prev = skb_rb_prev(skb);
1609 if (!prev)
1610 goto fallback;
1611
1612 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1613 goto fallback;
1614
1615 if (!tcp_skb_can_collapse(prev, skb))
1616 goto fallback;
1617
1618 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1619 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1620
1621 if (in_sack) {
1622 len = skb->len;
1623 pcount = tcp_skb_pcount(skb);
1624 mss = tcp_skb_seglen(skb);
1625
1626 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1627 * drop this restriction as unnecessary
1628 */
1629 if (mss != tcp_skb_seglen(prev))
1630 goto fallback;
1631 } else {
1632 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1633 goto noop;
1634 /* CHECKME: This is non-MSS split case only?, this will
1635 * cause skipped skbs due to advancing loop btw, original
1636 * has that feature too
1637 */
1638 if (tcp_skb_pcount(skb) <= 1)
1639 goto noop;
1640
1641 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1642 if (!in_sack) {
1643 /* TODO: head merge to next could be attempted here
1644 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1645 * though it might not be worth of the additional hassle
1646 *
1647 * ...we can probably just fallback to what was done
1648 * previously. We could try merging non-SACKed ones
1649 * as well but it probably isn't going to buy off
1650 * because later SACKs might again split them, and
1651 * it would make skb timestamp tracking considerably
1652 * harder problem.
1653 */
1654 goto fallback;
1655 }
1656
1657 len = end_seq - TCP_SKB_CB(skb)->seq;
1658 BUG_ON(len < 0);
1659 BUG_ON(len > skb->len);
1660
1661 /* MSS boundaries should be honoured or else pcount will
1662 * severely break even though it makes things bit trickier.
1663 * Optimize common case to avoid most of the divides
1664 */
1665 mss = tcp_skb_mss(skb);
1666
1667 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1668 * drop this restriction as unnecessary
1669 */
1670 if (mss != tcp_skb_seglen(prev))
1671 goto fallback;
1672
1673 if (len == mss) {
1674 pcount = 1;
1675 } else if (len < mss) {
1676 goto noop;
1677 } else {
1678 pcount = len / mss;
1679 len = pcount * mss;
1680 }
1681 }
1682
1683 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1684 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1685 goto fallback;
1686
1687 if (!tcp_skb_shift(prev, skb, pcount, len))
1688 goto fallback;
1689 if (!tcp_shifted_skb(sk, prev, skb, state, pcount, len, mss, dup_sack))
1690 goto out;
1691
1692 /* Hole filled allows collapsing with the next as well, this is very
1693 * useful when hole on every nth skb pattern happens
1694 */
1695 skb = skb_rb_next(prev);
1696 if (!skb)
1697 goto out;
1698
1699 if (!skb_can_shift(skb) ||
1700 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1701 (mss != tcp_skb_seglen(skb)))
1702 goto out;
1703
1704 if (!tcp_skb_can_collapse(prev, skb))
1705 goto out;
1706 len = skb->len;
1707 pcount = tcp_skb_pcount(skb);
1708 if (tcp_skb_shift(prev, skb, pcount, len))
1709 tcp_shifted_skb(sk, prev, skb, state, pcount,
1710 len, mss, 0);
1711
1712 out:
1713 return prev;
1714
1715 noop:
1716 return skb;
1717
1718 fallback:
1719 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1720 return NULL;
1721 }
1722
1723 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1724 struct tcp_sack_block *next_dup,
1725 struct tcp_sacktag_state *state,
1726 u32 start_seq, u32 end_seq,
1727 bool dup_sack_in)
1728 {
1729 struct tcp_sock *tp = tcp_sk(sk);
1730 struct sk_buff *tmp;
1731
1732 skb_rbtree_walk_from(skb) {
1733 int in_sack = 0;
1734 bool dup_sack = dup_sack_in;
1735
1736 /* queue is in-order => we can short-circuit the walk early */
1737 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1738 break;
1739
1740 if (next_dup &&
1741 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1742 in_sack = tcp_match_skb_to_sack(sk, skb,
1743 next_dup->start_seq,
1744 next_dup->end_seq);
1745 if (in_sack > 0)
1746 dup_sack = true;
1747 }
1748
1749 /* skb reference here is a bit tricky to get right, since
1750 * shifting can eat and free both this skb and the next,
1751 * so not even _safe variant of the loop is enough.
1752 */
1753 if (in_sack <= 0) {
1754 tmp = tcp_shift_skb_data(sk, skb, state,
1755 start_seq, end_seq, dup_sack);
1756 if (tmp) {
1757 if (tmp != skb) {
1758 skb = tmp;
1759 continue;
1760 }
1761
1762 in_sack = 0;
1763 } else {
1764 in_sack = tcp_match_skb_to_sack(sk, skb,
1765 start_seq,
1766 end_seq);
1767 }
1768 }
1769
1770 if (unlikely(in_sack < 0))
1771 break;
1772
1773 if (in_sack) {
1774 TCP_SKB_CB(skb)->sacked =
1775 tcp_sacktag_one(sk,
1776 state,
1777 TCP_SKB_CB(skb)->sacked,
1778 TCP_SKB_CB(skb)->seq,
1779 TCP_SKB_CB(skb)->end_seq,
1780 dup_sack,
1781 tcp_skb_pcount(skb),
1782 tcp_skb_timestamp_us(skb));
1783 tcp_rate_skb_delivered(sk, skb, state->rate);
1784 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1785 list_del_init(&skb->tcp_tsorted_anchor);
1786
1787 if (!before(TCP_SKB_CB(skb)->seq,
1788 tcp_highest_sack_seq(tp)))
1789 tcp_advance_highest_sack(sk, skb);
1790 }
1791 }
1792 return skb;
1793 }
1794
1795 static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk, u32 seq)
1796 {
1797 struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node;
1798 struct sk_buff *skb;
1799
1800 while (*p) {
1801 parent = *p;
1802 skb = rb_to_skb(parent);
1803 if (before(seq, TCP_SKB_CB(skb)->seq)) {
1804 p = &parent->rb_left;
1805 continue;
1806 }
1807 if (!before(seq, TCP_SKB_CB(skb)->end_seq)) {
1808 p = &parent->rb_right;
1809 continue;
1810 }
1811 return skb;
1812 }
1813 return NULL;
1814 }
1815
1816 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1817 u32 skip_to_seq)
1818 {
1819 if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq))
1820 return skb;
1821
1822 return tcp_sacktag_bsearch(sk, skip_to_seq);
1823 }
1824
1825 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1826 struct sock *sk,
1827 struct tcp_sack_block *next_dup,
1828 struct tcp_sacktag_state *state,
1829 u32 skip_to_seq)
1830 {
1831 if (!next_dup)
1832 return skb;
1833
1834 if (before(next_dup->start_seq, skip_to_seq)) {
1835 skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq);
1836 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1837 next_dup->start_seq, next_dup->end_seq,
1838 1);
1839 }
1840
1841 return skb;
1842 }
1843
1844 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1845 {
1846 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1847 }
1848
1849 static int
1850 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1851 u32 prior_snd_una, struct tcp_sacktag_state *state)
1852 {
1853 struct tcp_sock *tp = tcp_sk(sk);
1854 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1855 TCP_SKB_CB(ack_skb)->sacked);
1856 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1857 struct tcp_sack_block sp[TCP_NUM_SACKS];
1858 struct tcp_sack_block *cache;
1859 struct sk_buff *skb;
1860 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1861 int used_sacks;
1862 bool found_dup_sack = false;
1863 int i, j;
1864 int first_sack_index;
1865
1866 state->flag = 0;
1867 state->reord = tp->snd_nxt;
1868
1869 if (!tp->sacked_out)
1870 tcp_highest_sack_reset(sk);
1871
1872 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1873 num_sacks, prior_snd_una, state);
1874
1875 /* Eliminate too old ACKs, but take into
1876 * account more or less fresh ones, they can
1877 * contain valid SACK info.
1878 */
1879 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1880 return 0;
1881
1882 if (!tp->packets_out)
1883 goto out;
1884
1885 used_sacks = 0;
1886 first_sack_index = 0;
1887 for (i = 0; i < num_sacks; i++) {
1888 bool dup_sack = !i && found_dup_sack;
1889
1890 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1891 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1892
1893 if (!tcp_is_sackblock_valid(tp, dup_sack,
1894 sp[used_sacks].start_seq,
1895 sp[used_sacks].end_seq)) {
1896 int mib_idx;
1897
1898 if (dup_sack) {
1899 if (!tp->undo_marker)
1900 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1901 else
1902 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1903 } else {
1904 /* Don't count olds caused by ACK reordering */
1905 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1906 !after(sp[used_sacks].end_seq, tp->snd_una))
1907 continue;
1908 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1909 }
1910
1911 NET_INC_STATS(sock_net(sk), mib_idx);
1912 if (i == 0)
1913 first_sack_index = -1;
1914 continue;
1915 }
1916
1917 /* Ignore very old stuff early */
1918 if (!after(sp[used_sacks].end_seq, prior_snd_una)) {
1919 if (i == 0)
1920 first_sack_index = -1;
1921 continue;
1922 }
1923
1924 used_sacks++;
1925 }
1926
1927 /* order SACK blocks to allow in order walk of the retrans queue */
1928 for (i = used_sacks - 1; i > 0; i--) {
1929 for (j = 0; j < i; j++) {
1930 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1931 swap(sp[j], sp[j + 1]);
1932
1933 /* Track where the first SACK block goes to */
1934 if (j == first_sack_index)
1935 first_sack_index = j + 1;
1936 }
1937 }
1938 }
1939
1940 state->mss_now = tcp_current_mss(sk);
1941 skb = NULL;
1942 i = 0;
1943
1944 if (!tp->sacked_out) {
1945 /* It's already past, so skip checking against it */
1946 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1947 } else {
1948 cache = tp->recv_sack_cache;
1949 /* Skip empty blocks in at head of the cache */
1950 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1951 !cache->end_seq)
1952 cache++;
1953 }
1954
1955 while (i < used_sacks) {
1956 u32 start_seq = sp[i].start_seq;
1957 u32 end_seq = sp[i].end_seq;
1958 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1959 struct tcp_sack_block *next_dup = NULL;
1960
1961 if (found_dup_sack && ((i + 1) == first_sack_index))
1962 next_dup = &sp[i + 1];
1963
1964 /* Skip too early cached blocks */
1965 while (tcp_sack_cache_ok(tp, cache) &&
1966 !before(start_seq, cache->end_seq))
1967 cache++;
1968
1969 /* Can skip some work by looking recv_sack_cache? */
1970 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1971 after(end_seq, cache->start_seq)) {
1972
1973 /* Head todo? */
1974 if (before(start_seq, cache->start_seq)) {
1975 skb = tcp_sacktag_skip(skb, sk, start_seq);
1976 skb = tcp_sacktag_walk(skb, sk, next_dup,
1977 state,
1978 start_seq,
1979 cache->start_seq,
1980 dup_sack);
1981 }
1982
1983 /* Rest of the block already fully processed? */
1984 if (!after(end_seq, cache->end_seq))
1985 goto advance_sp;
1986
1987 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1988 state,
1989 cache->end_seq);
1990
1991 /* ...tail remains todo... */
1992 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1993 /* ...but better entrypoint exists! */
1994 skb = tcp_highest_sack(sk);
1995 if (!skb)
1996 break;
1997 cache++;
1998 goto walk;
1999 }
2000
2001 skb = tcp_sacktag_skip(skb, sk, cache->end_seq);
2002 /* Check overlap against next cached too (past this one already) */
2003 cache++;
2004 continue;
2005 }
2006
2007 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
2008 skb = tcp_highest_sack(sk);
2009 if (!skb)
2010 break;
2011 }
2012 skb = tcp_sacktag_skip(skb, sk, start_seq);
2013
2014 walk:
2015 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
2016 start_seq, end_seq, dup_sack);
2017
2018 advance_sp:
2019 i++;
2020 }
2021
2022 /* Clear the head of the cache sack blocks so we can skip it next time */
2023 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
2024 tp->recv_sack_cache[i].start_seq = 0;
2025 tp->recv_sack_cache[i].end_seq = 0;
2026 }
2027 for (j = 0; j < used_sacks; j++)
2028 tp->recv_sack_cache[i++] = sp[j];
2029
2030 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss || tp->undo_marker)
2031 tcp_check_sack_reordering(sk, state->reord, 0);
2032
2033 tcp_verify_left_out(tp);
2034 out:
2035
2036 #if FASTRETRANS_DEBUG > 0
2037 WARN_ON((int)tp->sacked_out < 0);
2038 WARN_ON((int)tp->lost_out < 0);
2039 WARN_ON((int)tp->retrans_out < 0);
2040 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
2041 #endif
2042 return state->flag;
2043 }
2044
2045 /* Limits sacked_out so that sum with lost_out isn't ever larger than
2046 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
2047 */
2048 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
2049 {
2050 u32 holes;
2051
2052 holes = max(tp->lost_out, 1U);
2053 holes = min(holes, tp->packets_out);
2054
2055 if ((tp->sacked_out + holes) > tp->packets_out) {
2056 tp->sacked_out = tp->packets_out - holes;
2057 return true;
2058 }
2059 return false;
2060 }
2061
2062 /* If we receive more dupacks than we expected counting segments
2063 * in assumption of absent reordering, interpret this as reordering.
2064 * The only another reason could be bug in receiver TCP.
2065 */
2066 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
2067 {
2068 struct tcp_sock *tp = tcp_sk(sk);
2069
2070 if (!tcp_limit_reno_sacked(tp))
2071 return;
2072
2073 tp->reordering = min_t(u32, tp->packets_out + addend,
2074 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering));
2075 tp->reord_seen++;
2076 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER);
2077 }
2078
2079 /* Emulate SACKs for SACKless connection: account for a new dupack. */
2080
2081 static void tcp_add_reno_sack(struct sock *sk, int num_dupack, bool ece_ack)
2082 {
2083 if (num_dupack) {
2084 struct tcp_sock *tp = tcp_sk(sk);
2085 u32 prior_sacked = tp->sacked_out;
2086 s32 delivered;
2087
2088 tp->sacked_out += num_dupack;
2089 tcp_check_reno_reordering(sk, 0);
2090 delivered = tp->sacked_out - prior_sacked;
2091 if (delivered > 0)
2092 tcp_count_delivered(tp, delivered, ece_ack);
2093 tcp_verify_left_out(tp);
2094 }
2095 }
2096
2097 /* Account for ACK, ACKing some data in Reno Recovery phase. */
2098
2099 static void tcp_remove_reno_sacks(struct sock *sk, int acked, bool ece_ack)
2100 {
2101 struct tcp_sock *tp = tcp_sk(sk);
2102
2103 if (acked > 0) {
2104 /* One ACK acked hole. The rest eat duplicate ACKs. */
2105 tcp_count_delivered(tp, max_t(int, acked - tp->sacked_out, 1),
2106 ece_ack);
2107 if (acked - 1 >= tp->sacked_out)
2108 tp->sacked_out = 0;
2109 else
2110 tp->sacked_out -= acked - 1;
2111 }
2112 tcp_check_reno_reordering(sk, acked);
2113 tcp_verify_left_out(tp);
2114 }
2115
2116 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
2117 {
2118 tp->sacked_out = 0;
2119 }
2120
2121 void tcp_clear_retrans(struct tcp_sock *tp)
2122 {
2123 tp->retrans_out = 0;
2124 tp->lost_out = 0;
2125 tp->undo_marker = 0;
2126 tp->undo_retrans = -1;
2127 tp->sacked_out = 0;
2128 tp->rto_stamp = 0;
2129 tp->total_rto = 0;
2130 tp->total_rto_recoveries = 0;
2131 tp->total_rto_time = 0;
2132 }
2133
2134 static inline void tcp_init_undo(struct tcp_sock *tp)
2135 {
2136 tp->undo_marker = tp->snd_una;
2137
2138 /* Retransmission still in flight may cause DSACKs later. */
2139 /* First, account for regular retransmits in flight: */
2140 tp->undo_retrans = tp->retrans_out;
2141 /* Next, account for TLP retransmits in flight: */
2142 if (tp->tlp_high_seq && tp->tlp_retrans)
2143 tp->undo_retrans++;
2144 /* Finally, avoid 0, because undo_retrans==0 means "can undo now": */
2145 if (!tp->undo_retrans)
2146 tp->undo_retrans = -1;
2147 }
2148
2149 static bool tcp_is_rack(const struct sock *sk)
2150 {
2151 return READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_recovery) &
2152 TCP_RACK_LOSS_DETECTION;
2153 }
2154
2155 /* If we detect SACK reneging, forget all SACK information
2156 * and reset tags completely, otherwise preserve SACKs. If receiver
2157 * dropped its ofo queue, we will know this due to reneging detection.
2158 */
2159 static void tcp_timeout_mark_lost(struct sock *sk)
2160 {
2161 struct tcp_sock *tp = tcp_sk(sk);
2162 struct sk_buff *skb, *head;
2163 bool is_reneg; /* is receiver reneging on SACKs? */
2164
2165 head = tcp_rtx_queue_head(sk);
2166 is_reneg = head && (TCP_SKB_CB(head)->sacked & TCPCB_SACKED_ACKED);
2167 if (is_reneg) {
2168 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
2169 tp->sacked_out = 0;
2170 /* Mark SACK reneging until we recover from this loss event. */
2171 tp->is_sack_reneg = 1;
2172 } else if (tcp_is_reno(tp)) {
2173 tcp_reset_reno_sack(tp);
2174 }
2175
2176 skb = head;
2177 skb_rbtree_walk_from(skb) {
2178 if (is_reneg)
2179 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
2180 else if (tcp_is_rack(sk) && skb != head &&
2181 tcp_rack_skb_timeout(tp, skb, 0) > 0)
2182 continue; /* Don't mark recently sent ones lost yet */
2183 tcp_mark_skb_lost(sk, skb);
2184 }
2185 tcp_verify_left_out(tp);
2186 tcp_clear_all_retrans_hints(tp);
2187 }
2188
2189 /* Enter Loss state. */
2190 void tcp_enter_loss(struct sock *sk)
2191 {
2192 const struct inet_connection_sock *icsk = inet_csk(sk);
2193 struct tcp_sock *tp = tcp_sk(sk);
2194 struct net *net = sock_net(sk);
2195 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
2196 u8 reordering;
2197
2198 tcp_timeout_mark_lost(sk);
2199
2200 /* Reduce ssthresh if it has not yet been made inside this window. */
2201 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
2202 !after(tp->high_seq, tp->snd_una) ||
2203 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2204 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2205 tp->prior_cwnd = tcp_snd_cwnd(tp);
2206 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2207 tcp_ca_event(sk, CA_EVENT_LOSS);
2208 tcp_init_undo(tp);
2209 }
2210 tcp_snd_cwnd_set(tp, tcp_packets_in_flight(tp) + 1);
2211 tp->snd_cwnd_cnt = 0;
2212 tp->snd_cwnd_stamp = tcp_jiffies32;
2213
2214 /* Timeout in disordered state after receiving substantial DUPACKs
2215 * suggests that the degree of reordering is over-estimated.
2216 */
2217 reordering = READ_ONCE(net->ipv4.sysctl_tcp_reordering);
2218 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
2219 tp->sacked_out >= reordering)
2220 tp->reordering = min_t(unsigned int, tp->reordering,
2221 reordering);
2222
2223 tcp_set_ca_state(sk, TCP_CA_Loss);
2224 tp->high_seq = tp->snd_nxt;
2225 tp->tlp_high_seq = 0;
2226 tcp_ecn_queue_cwr(tp);
2227
2228 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
2229 * loss recovery is underway except recurring timeout(s) on
2230 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
2231 */
2232 tp->frto = READ_ONCE(net->ipv4.sysctl_tcp_frto) &&
2233 (new_recovery || icsk->icsk_retransmits) &&
2234 !inet_csk(sk)->icsk_mtup.probe_size;
2235 }
2236
2237 /* If ACK arrived pointing to a remembered SACK, it means that our
2238 * remembered SACKs do not reflect real state of receiver i.e.
2239 * receiver _host_ is heavily congested (or buggy).
2240 *
2241 * To avoid big spurious retransmission bursts due to transient SACK
2242 * scoreboard oddities that look like reneging, we give the receiver a
2243 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2244 * restore sanity to the SACK scoreboard. If the apparent reneging
2245 * persists until this RTO then we'll clear the SACK scoreboard.
2246 */
2247 static bool tcp_check_sack_reneging(struct sock *sk, int *ack_flag)
2248 {
2249 if (*ack_flag & FLAG_SACK_RENEGING &&
2250 *ack_flag & FLAG_SND_UNA_ADVANCED) {
2251 struct tcp_sock *tp = tcp_sk(sk);
2252 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2253 msecs_to_jiffies(10));
2254
2255 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2256 delay, TCP_RTO_MAX);
2257 *ack_flag &= ~FLAG_SET_XMIT_TIMER;
2258 return true;
2259 }
2260 return false;
2261 }
2262
2263 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2264 * counter when SACK is enabled (without SACK, sacked_out is used for
2265 * that purpose).
2266 *
2267 * With reordering, holes may still be in flight, so RFC3517 recovery
2268 * uses pure sacked_out (total number of SACKed segments) even though
2269 * it violates the RFC that uses duplicate ACKs, often these are equal
2270 * but when e.g. out-of-window ACKs or packet duplication occurs,
2271 * they differ. Since neither occurs due to loss, TCP should really
2272 * ignore them.
2273 */
2274 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2275 {
2276 return tp->sacked_out + 1;
2277 }
2278
2279 /* Linux NewReno/SACK/ECN state machine.
2280 * --------------------------------------
2281 *
2282 * "Open" Normal state, no dubious events, fast path.
2283 * "Disorder" In all the respects it is "Open",
2284 * but requires a bit more attention. It is entered when
2285 * we see some SACKs or dupacks. It is split of "Open"
2286 * mainly to move some processing from fast path to slow one.
2287 * "CWR" CWND was reduced due to some Congestion Notification event.
2288 * It can be ECN, ICMP source quench, local device congestion.
2289 * "Recovery" CWND was reduced, we are fast-retransmitting.
2290 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2291 *
2292 * tcp_fastretrans_alert() is entered:
2293 * - each incoming ACK, if state is not "Open"
2294 * - when arrived ACK is unusual, namely:
2295 * * SACK
2296 * * Duplicate ACK.
2297 * * ECN ECE.
2298 *
2299 * Counting packets in flight is pretty simple.
2300 *
2301 * in_flight = packets_out - left_out + retrans_out
2302 *
2303 * packets_out is SND.NXT-SND.UNA counted in packets.
2304 *
2305 * retrans_out is number of retransmitted segments.
2306 *
2307 * left_out is number of segments left network, but not ACKed yet.
2308 *
2309 * left_out = sacked_out + lost_out
2310 *
2311 * sacked_out: Packets, which arrived to receiver out of order
2312 * and hence not ACKed. With SACKs this number is simply
2313 * amount of SACKed data. Even without SACKs
2314 * it is easy to give pretty reliable estimate of this number,
2315 * counting duplicate ACKs.
2316 *
2317 * lost_out: Packets lost by network. TCP has no explicit
2318 * "loss notification" feedback from network (for now).
2319 * It means that this number can be only _guessed_.
2320 * Actually, it is the heuristics to predict lossage that
2321 * distinguishes different algorithms.
2322 *
2323 * F.e. after RTO, when all the queue is considered as lost,
2324 * lost_out = packets_out and in_flight = retrans_out.
2325 *
2326 * Essentially, we have now a few algorithms detecting
2327 * lost packets.
2328 *
2329 * If the receiver supports SACK:
2330 *
2331 * RFC6675/3517: It is the conventional algorithm. A packet is
2332 * considered lost if the number of higher sequence packets
2333 * SACKed is greater than or equal the DUPACK thoreshold
2334 * (reordering). This is implemented in tcp_mark_head_lost and
2335 * tcp_update_scoreboard.
2336 *
2337 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm
2338 * (2017-) that checks timing instead of counting DUPACKs.
2339 * Essentially a packet is considered lost if it's not S/ACKed
2340 * after RTT + reordering_window, where both metrics are
2341 * dynamically measured and adjusted. This is implemented in
2342 * tcp_rack_mark_lost.
2343 *
2344 * If the receiver does not support SACK:
2345 *
2346 * NewReno (RFC6582): in Recovery we assume that one segment
2347 * is lost (classic Reno). While we are in Recovery and
2348 * a partial ACK arrives, we assume that one more packet
2349 * is lost (NewReno). This heuristics are the same in NewReno
2350 * and SACK.
2351 *
2352 * Really tricky (and requiring careful tuning) part of algorithm
2353 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2354 * The first determines the moment _when_ we should reduce CWND and,
2355 * hence, slow down forward transmission. In fact, it determines the moment
2356 * when we decide that hole is caused by loss, rather than by a reorder.
2357 *
2358 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2359 * holes, caused by lost packets.
2360 *
2361 * And the most logically complicated part of algorithm is undo
2362 * heuristics. We detect false retransmits due to both too early
2363 * fast retransmit (reordering) and underestimated RTO, analyzing
2364 * timestamps and D-SACKs. When we detect that some segments were
2365 * retransmitted by mistake and CWND reduction was wrong, we undo
2366 * window reduction and abort recovery phase. This logic is hidden
2367 * inside several functions named tcp_try_undo_<something>.
2368 */
2369
2370 /* This function decides, when we should leave Disordered state
2371 * and enter Recovery phase, reducing congestion window.
2372 *
2373 * Main question: may we further continue forward transmission
2374 * with the same cwnd?
2375 */
2376 static bool tcp_time_to_recover(struct sock *sk, int flag)
2377 {
2378 struct tcp_sock *tp = tcp_sk(sk);
2379
2380 /* Trick#1: The loss is proven. */
2381 if (tp->lost_out)
2382 return true;
2383
2384 /* Not-A-Trick#2 : Classic rule... */
2385 if (!tcp_is_rack(sk) && tcp_dupack_heuristics(tp) > tp->reordering)
2386 return true;
2387
2388 return false;
2389 }
2390
2391 /* Detect loss in event "A" above by marking head of queue up as lost.
2392 * For RFC3517 SACK, a segment is considered lost if it
2393 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2394 * the maximum SACKed segments to pass before reaching this limit.
2395 */
2396 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2397 {
2398 struct tcp_sock *tp = tcp_sk(sk);
2399 struct sk_buff *skb;
2400 int cnt;
2401 /* Use SACK to deduce losses of new sequences sent during recovery */
2402 const u32 loss_high = tp->snd_nxt;
2403
2404 WARN_ON(packets > tp->packets_out);
2405 skb = tp->lost_skb_hint;
2406 if (skb) {
2407 /* Head already handled? */
2408 if (mark_head && after(TCP_SKB_CB(skb)->seq, tp->snd_una))
2409 return;
2410 cnt = tp->lost_cnt_hint;
2411 } else {
2412 skb = tcp_rtx_queue_head(sk);
2413 cnt = 0;
2414 }
2415
2416 skb_rbtree_walk_from(skb) {
2417 /* TODO: do this better */
2418 /* this is not the most efficient way to do this... */
2419 tp->lost_skb_hint = skb;
2420 tp->lost_cnt_hint = cnt;
2421
2422 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2423 break;
2424
2425 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2426 cnt += tcp_skb_pcount(skb);
2427
2428 if (cnt > packets)
2429 break;
2430
2431 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_LOST))
2432 tcp_mark_skb_lost(sk, skb);
2433
2434 if (mark_head)
2435 break;
2436 }
2437 tcp_verify_left_out(tp);
2438 }
2439
2440 /* Account newly detected lost packet(s) */
2441
2442 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2443 {
2444 struct tcp_sock *tp = tcp_sk(sk);
2445
2446 if (tcp_is_sack(tp)) {
2447 int sacked_upto = tp->sacked_out - tp->reordering;
2448 if (sacked_upto >= 0)
2449 tcp_mark_head_lost(sk, sacked_upto, 0);
2450 else if (fast_rexmit)
2451 tcp_mark_head_lost(sk, 1, 1);
2452 }
2453 }
2454
2455 static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2456 {
2457 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2458 before(tp->rx_opt.rcv_tsecr, when);
2459 }
2460
2461 /* skb is spurious retransmitted if the returned timestamp echo
2462 * reply is prior to the skb transmission time
2463 */
2464 static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2465 const struct sk_buff *skb)
2466 {
2467 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2468 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp_ts(tp->tcp_usec_ts, skb));
2469 }
2470
2471 /* Nothing was retransmitted or returned timestamp is less
2472 * than timestamp of the first retransmission.
2473 */
2474 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2475 {
2476 const struct sock *sk = (const struct sock *)tp;
2477
2478 if (tp->retrans_stamp &&
2479 tcp_tsopt_ecr_before(tp, tp->retrans_stamp))
2480 return true; /* got echoed TS before first retransmission */
2481
2482 /* Check if nothing was retransmitted (retrans_stamp==0), which may
2483 * happen in fast recovery due to TSQ. But we ignore zero retrans_stamp
2484 * in TCP_SYN_SENT, since when we set FLAG_SYN_ACKED we also clear
2485 * retrans_stamp even if we had retransmitted the SYN.
2486 */
2487 if (!tp->retrans_stamp && /* no record of a retransmit/SYN? */
2488 sk->sk_state != TCP_SYN_SENT) /* not the FLAG_SYN_ACKED case? */
2489 return true; /* nothing was retransmitted */
2490
2491 return false;
2492 }
2493
2494 /* Undo procedures. */
2495
2496 /* We can clear retrans_stamp when there are no retransmissions in the
2497 * window. It would seem that it is trivially available for us in
2498 * tp->retrans_out, however, that kind of assumptions doesn't consider
2499 * what will happen if errors occur when sending retransmission for the
2500 * second time. ...It could the that such segment has only
2501 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2502 * the head skb is enough except for some reneging corner cases that
2503 * are not worth the effort.
2504 *
2505 * Main reason for all this complexity is the fact that connection dying
2506 * time now depends on the validity of the retrans_stamp, in particular,
2507 * that successive retransmissions of a segment must not advance
2508 * retrans_stamp under any conditions.
2509 */
2510 static bool tcp_any_retrans_done(const struct sock *sk)
2511 {
2512 const struct tcp_sock *tp = tcp_sk(sk);
2513 struct sk_buff *skb;
2514
2515 if (tp->retrans_out)
2516 return true;
2517
2518 skb = tcp_rtx_queue_head(sk);
2519 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2520 return true;
2521
2522 return false;
2523 }
2524
2525 /* If loss recovery is finished and there are no retransmits out in the
2526 * network, then we clear retrans_stamp so that upon the next loss recovery
2527 * retransmits_timed_out() and timestamp-undo are using the correct value.
2528 */
2529 static void tcp_retrans_stamp_cleanup(struct sock *sk)
2530 {
2531 if (!tcp_any_retrans_done(sk))
2532 tcp_sk(sk)->retrans_stamp = 0;
2533 }
2534
2535 static void DBGUNDO(struct sock *sk, const char *msg)
2536 {
2537 #if FASTRETRANS_DEBUG > 1
2538 struct tcp_sock *tp = tcp_sk(sk);
2539 struct inet_sock *inet = inet_sk(sk);
2540
2541 if (sk->sk_family == AF_INET) {
2542 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2543 msg,
2544 &inet->inet_daddr, ntohs(inet->inet_dport),
2545 tcp_snd_cwnd(tp), tcp_left_out(tp),
2546 tp->snd_ssthresh, tp->prior_ssthresh,
2547 tp->packets_out);
2548 }
2549 #if IS_ENABLED(CONFIG_IPV6)
2550 else if (sk->sk_family == AF_INET6) {
2551 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2552 msg,
2553 &sk->sk_v6_daddr, ntohs(inet->inet_dport),
2554 tcp_snd_cwnd(tp), tcp_left_out(tp),
2555 tp->snd_ssthresh, tp->prior_ssthresh,
2556 tp->packets_out);
2557 }
2558 #endif
2559 #endif
2560 }
2561
2562 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2563 {
2564 struct tcp_sock *tp = tcp_sk(sk);
2565
2566 if (unmark_loss) {
2567 struct sk_buff *skb;
2568
2569 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2570 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2571 }
2572 tp->lost_out = 0;
2573 tcp_clear_all_retrans_hints(tp);
2574 }
2575
2576 if (tp->prior_ssthresh) {
2577 const struct inet_connection_sock *icsk = inet_csk(sk);
2578
2579 tcp_snd_cwnd_set(tp, icsk->icsk_ca_ops->undo_cwnd(sk));
2580
2581 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2582 tp->snd_ssthresh = tp->prior_ssthresh;
2583 tcp_ecn_withdraw_cwr(tp);
2584 }
2585 }
2586 tp->snd_cwnd_stamp = tcp_jiffies32;
2587 tp->undo_marker = 0;
2588 tp->rack.advanced = 1; /* Force RACK to re-exam losses */
2589 }
2590
2591 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2592 {
2593 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2594 }
2595
2596 static bool tcp_is_non_sack_preventing_reopen(struct sock *sk)
2597 {
2598 struct tcp_sock *tp = tcp_sk(sk);
2599
2600 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2601 /* Hold old state until something *above* high_seq
2602 * is ACKed. For Reno it is MUST to prevent false
2603 * fast retransmits (RFC2582). SACK TCP is safe. */
2604 if (!tcp_any_retrans_done(sk))
2605 tp->retrans_stamp = 0;
2606 return true;
2607 }
2608 return false;
2609 }
2610
2611 /* People celebrate: "We love our President!" */
2612 static bool tcp_try_undo_recovery(struct sock *sk)
2613 {
2614 struct tcp_sock *tp = tcp_sk(sk);
2615
2616 if (tcp_may_undo(tp)) {
2617 int mib_idx;
2618
2619 /* Happy end! We did not retransmit anything
2620 * or our original transmission succeeded.
2621 */
2622 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2623 tcp_undo_cwnd_reduction(sk, false);
2624 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2625 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2626 else
2627 mib_idx = LINUX_MIB_TCPFULLUNDO;
2628
2629 NET_INC_STATS(sock_net(sk), mib_idx);
2630 } else if (tp->rack.reo_wnd_persist) {
2631 tp->rack.reo_wnd_persist--;
2632 }
2633 if (tcp_is_non_sack_preventing_reopen(sk))
2634 return true;
2635 tcp_set_ca_state(sk, TCP_CA_Open);
2636 tp->is_sack_reneg = 0;
2637 return false;
2638 }
2639
2640 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2641 static bool tcp_try_undo_dsack(struct sock *sk)
2642 {
2643 struct tcp_sock *tp = tcp_sk(sk);
2644
2645 if (tp->undo_marker && !tp->undo_retrans) {
2646 tp->rack.reo_wnd_persist = min(TCP_RACK_RECOVERY_THRESH,
2647 tp->rack.reo_wnd_persist + 1);
2648 DBGUNDO(sk, "D-SACK");
2649 tcp_undo_cwnd_reduction(sk, false);
2650 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2651 return true;
2652 }
2653 return false;
2654 }
2655
2656 /* Undo during loss recovery after partial ACK or using F-RTO. */
2657 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2658 {
2659 struct tcp_sock *tp = tcp_sk(sk);
2660
2661 if (frto_undo || tcp_may_undo(tp)) {
2662 tcp_undo_cwnd_reduction(sk, true);
2663
2664 DBGUNDO(sk, "partial loss");
2665 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2666 if (frto_undo)
2667 NET_INC_STATS(sock_net(sk),
2668 LINUX_MIB_TCPSPURIOUSRTOS);
2669 inet_csk(sk)->icsk_retransmits = 0;
2670 if (tcp_is_non_sack_preventing_reopen(sk))
2671 return true;
2672 if (frto_undo || tcp_is_sack(tp)) {
2673 tcp_set_ca_state(sk, TCP_CA_Open);
2674 tp->is_sack_reneg = 0;
2675 }
2676 return true;
2677 }
2678 return false;
2679 }
2680
2681 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2682 * It computes the number of packets to send (sndcnt) based on packets newly
2683 * delivered:
2684 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2685 * cwnd reductions across a full RTT.
2686 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2687 * But when SND_UNA is acked without further losses,
2688 * slow starts cwnd up to ssthresh to speed up the recovery.
2689 */
2690 static void tcp_init_cwnd_reduction(struct sock *sk)
2691 {
2692 struct tcp_sock *tp = tcp_sk(sk);
2693
2694 tp->high_seq = tp->snd_nxt;
2695 tp->tlp_high_seq = 0;
2696 tp->snd_cwnd_cnt = 0;
2697 tp->prior_cwnd = tcp_snd_cwnd(tp);
2698 tp->prr_delivered = 0;
2699 tp->prr_out = 0;
2700 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2701 tcp_ecn_queue_cwr(tp);
2702 }
2703
2704 void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int newly_lost, int flag)
2705 {
2706 struct tcp_sock *tp = tcp_sk(sk);
2707 int sndcnt = 0;
2708 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2709
2710 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2711 return;
2712
2713 tp->prr_delivered += newly_acked_sacked;
2714 if (delta < 0) {
2715 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2716 tp->prior_cwnd - 1;
2717 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2718 } else {
2719 sndcnt = max_t(int, tp->prr_delivered - tp->prr_out,
2720 newly_acked_sacked);
2721 if (flag & FLAG_SND_UNA_ADVANCED && !newly_lost)
2722 sndcnt++;
2723 sndcnt = min(delta, sndcnt);
2724 }
2725 /* Force a fast retransmit upon entering fast recovery */
2726 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
2727 tcp_snd_cwnd_set(tp, tcp_packets_in_flight(tp) + sndcnt);
2728 }
2729
2730 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2731 {
2732 struct tcp_sock *tp = tcp_sk(sk);
2733
2734 if (inet_csk(sk)->icsk_ca_ops->cong_control)
2735 return;
2736
2737 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2738 if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH &&
2739 (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) {
2740 tcp_snd_cwnd_set(tp, tp->snd_ssthresh);
2741 tp->snd_cwnd_stamp = tcp_jiffies32;
2742 }
2743 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2744 }
2745
2746 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2747 void tcp_enter_cwr(struct sock *sk)
2748 {
2749 struct tcp_sock *tp = tcp_sk(sk);
2750
2751 tp->prior_ssthresh = 0;
2752 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2753 tp->undo_marker = 0;
2754 tcp_init_cwnd_reduction(sk);
2755 tcp_set_ca_state(sk, TCP_CA_CWR);
2756 }
2757 }
2758 EXPORT_SYMBOL(tcp_enter_cwr);
2759
2760 static void tcp_try_keep_open(struct sock *sk)
2761 {
2762 struct tcp_sock *tp = tcp_sk(sk);
2763 int state = TCP_CA_Open;
2764
2765 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2766 state = TCP_CA_Disorder;
2767
2768 if (inet_csk(sk)->icsk_ca_state != state) {
2769 tcp_set_ca_state(sk, state);
2770 tp->high_seq = tp->snd_nxt;
2771 }
2772 }
2773
2774 static void tcp_try_to_open(struct sock *sk, int flag)
2775 {
2776 struct tcp_sock *tp = tcp_sk(sk);
2777
2778 tcp_verify_left_out(tp);
2779
2780 if (!tcp_any_retrans_done(sk))
2781 tp->retrans_stamp = 0;
2782
2783 if (flag & FLAG_ECE)
2784 tcp_enter_cwr(sk);
2785
2786 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2787 tcp_try_keep_open(sk);
2788 }
2789 }
2790
2791 static void tcp_mtup_probe_failed(struct sock *sk)
2792 {
2793 struct inet_connection_sock *icsk = inet_csk(sk);
2794
2795 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2796 icsk->icsk_mtup.probe_size = 0;
2797 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
2798 }
2799
2800 static void tcp_mtup_probe_success(struct sock *sk)
2801 {
2802 struct tcp_sock *tp = tcp_sk(sk);
2803 struct inet_connection_sock *icsk = inet_csk(sk);
2804 u64 val;
2805
2806 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2807
2808 val = (u64)tcp_snd_cwnd(tp) * tcp_mss_to_mtu(sk, tp->mss_cache);
2809 do_div(val, icsk->icsk_mtup.probe_size);
2810 DEBUG_NET_WARN_ON_ONCE((u32)val != val);
2811 tcp_snd_cwnd_set(tp, max_t(u32, 1U, val));
2812
2813 tp->snd_cwnd_cnt = 0;
2814 tp->snd_cwnd_stamp = tcp_jiffies32;
2815 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2816
2817 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2818 icsk->icsk_mtup.probe_size = 0;
2819 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2820 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
2821 }
2822
2823 /* Sometimes we deduce that packets have been dropped due to reasons other than
2824 * congestion, like path MTU reductions or failed client TFO attempts. In these
2825 * cases we call this function to retransmit as many packets as cwnd allows,
2826 * without reducing cwnd. Given that retransmits will set retrans_stamp to a
2827 * non-zero value (and may do so in a later calling context due to TSQ), we
2828 * also enter CA_Loss so that we track when all retransmitted packets are ACKed
2829 * and clear retrans_stamp when that happens (to ensure later recurring RTOs
2830 * are using the correct retrans_stamp and don't declare ETIMEDOUT
2831 * prematurely).
2832 */
2833 static void tcp_non_congestion_loss_retransmit(struct sock *sk)
2834 {
2835 const struct inet_connection_sock *icsk = inet_csk(sk);
2836 struct tcp_sock *tp = tcp_sk(sk);
2837
2838 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2839 tp->high_seq = tp->snd_nxt;
2840 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2841 tp->prior_ssthresh = 0;
2842 tp->undo_marker = 0;
2843 tcp_set_ca_state(sk, TCP_CA_Loss);
2844 }
2845 tcp_xmit_retransmit_queue(sk);
2846 }
2847
2848 /* Do a simple retransmit without using the backoff mechanisms in
2849 * tcp_timer. This is used for path mtu discovery.
2850 * The socket is already locked here.
2851 */
2852 void tcp_simple_retransmit(struct sock *sk)
2853 {
2854 struct tcp_sock *tp = tcp_sk(sk);
2855 struct sk_buff *skb;
2856 int mss;
2857
2858 /* A fastopen SYN request is stored as two separate packets within
2859 * the retransmit queue, this is done by tcp_send_syn_data().
2860 * As a result simply checking the MSS of the frames in the queue
2861 * will not work for the SYN packet.
2862 *
2863 * Us being here is an indication of a path MTU issue so we can
2864 * assume that the fastopen SYN was lost and just mark all the
2865 * frames in the retransmit queue as lost. We will use an MSS of
2866 * -1 to mark all frames as lost, otherwise compute the current MSS.
2867 */
2868 if (tp->syn_data && sk->sk_state == TCP_SYN_SENT)
2869 mss = -1;
2870 else
2871 mss = tcp_current_mss(sk);
2872
2873 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2874 if (tcp_skb_seglen(skb) > mss)
2875 tcp_mark_skb_lost(sk, skb);
2876 }
2877
2878 tcp_clear_retrans_hints_partial(tp);
2879
2880 if (!tp->lost_out)
2881 return;
2882
2883 if (tcp_is_reno(tp))
2884 tcp_limit_reno_sacked(tp);
2885
2886 tcp_verify_left_out(tp);
2887
2888 /* Don't muck with the congestion window here.
2889 * Reason is that we do not increase amount of _data_
2890 * in network, but units changed and effective
2891 * cwnd/ssthresh really reduced now.
2892 */
2893 tcp_non_congestion_loss_retransmit(sk);
2894 }
2895 EXPORT_SYMBOL(tcp_simple_retransmit);
2896
2897 void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2898 {
2899 struct tcp_sock *tp = tcp_sk(sk);
2900 int mib_idx;
2901
2902 /* Start the clock with our fast retransmit, for undo and ETIMEDOUT. */
2903 tcp_retrans_stamp_cleanup(sk);
2904
2905 if (tcp_is_reno(tp))
2906 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2907 else
2908 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2909
2910 NET_INC_STATS(sock_net(sk), mib_idx);
2911
2912 tp->prior_ssthresh = 0;
2913 tcp_init_undo(tp);
2914
2915 if (!tcp_in_cwnd_reduction(sk)) {
2916 if (!ece_ack)
2917 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2918 tcp_init_cwnd_reduction(sk);
2919 }
2920 tcp_set_ca_state(sk, TCP_CA_Recovery);
2921 }
2922
2923 static void tcp_update_rto_time(struct tcp_sock *tp)
2924 {
2925 if (tp->rto_stamp) {
2926 tp->total_rto_time += tcp_time_stamp_ms(tp) - tp->rto_stamp;
2927 tp->rto_stamp = 0;
2928 }
2929 }
2930
2931 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2932 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2933 */
2934 static void tcp_process_loss(struct sock *sk, int flag, int num_dupack,
2935 int *rexmit)
2936 {
2937 struct tcp_sock *tp = tcp_sk(sk);
2938 bool recovered = !before(tp->snd_una, tp->high_seq);
2939
2940 if ((flag & FLAG_SND_UNA_ADVANCED || rcu_access_pointer(tp->fastopen_rsk)) &&
2941 tcp_try_undo_loss(sk, false))
2942 return;
2943
2944 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2945 /* Step 3.b. A timeout is spurious if not all data are
2946 * lost, i.e., never-retransmitted data are (s)acked.
2947 */
2948 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2949 tcp_try_undo_loss(sk, true))
2950 return;
2951
2952 if (after(tp->snd_nxt, tp->high_seq)) {
2953 if (flag & FLAG_DATA_SACKED || num_dupack)
2954 tp->frto = 0; /* Step 3.a. loss was real */
2955 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2956 tp->high_seq = tp->snd_nxt;
2957 /* Step 2.b. Try send new data (but deferred until cwnd
2958 * is updated in tcp_ack()). Otherwise fall back to
2959 * the conventional recovery.
2960 */
2961 if (!tcp_write_queue_empty(sk) &&
2962 after(tcp_wnd_end(tp), tp->snd_nxt)) {
2963 *rexmit = REXMIT_NEW;
2964 return;
2965 }
2966 tp->frto = 0;
2967 }
2968 }
2969
2970 if (recovered) {
2971 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2972 tcp_try_undo_recovery(sk);
2973 return;
2974 }
2975 if (tcp_is_reno(tp)) {
2976 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2977 * delivered. Lower inflight to clock out (re)transmissions.
2978 */
2979 if (after(tp->snd_nxt, tp->high_seq) && num_dupack)
2980 tcp_add_reno_sack(sk, num_dupack, flag & FLAG_ECE);
2981 else if (flag & FLAG_SND_UNA_ADVANCED)
2982 tcp_reset_reno_sack(tp);
2983 }
2984 *rexmit = REXMIT_LOST;
2985 }
2986
2987 static bool tcp_force_fast_retransmit(struct sock *sk)
2988 {
2989 struct tcp_sock *tp = tcp_sk(sk);
2990
2991 return after(tcp_highest_sack_seq(tp),
2992 tp->snd_una + tp->reordering * tp->mss_cache);
2993 }
2994
2995 /* Undo during fast recovery after partial ACK. */
2996 static bool tcp_try_undo_partial(struct sock *sk, u32 prior_snd_una,
2997 bool *do_lost)
2998 {
2999 struct tcp_sock *tp = tcp_sk(sk);
3000
3001 if (tp->undo_marker && tcp_packet_delayed(tp)) {
3002 /* Plain luck! Hole if filled with delayed
3003 * packet, rather than with a retransmit. Check reordering.
3004 */
3005 tcp_check_sack_reordering(sk, prior_snd_una, 1);
3006
3007 /* We are getting evidence that the reordering degree is higher
3008 * than we realized. If there are no retransmits out then we
3009 * can undo. Otherwise we clock out new packets but do not
3010 * mark more packets lost or retransmit more.
3011 */
3012 if (tp->retrans_out)
3013 return true;
3014
3015 if (!tcp_any_retrans_done(sk))
3016 tp->retrans_stamp = 0;
3017
3018 DBGUNDO(sk, "partial recovery");
3019 tcp_undo_cwnd_reduction(sk, true);
3020 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
3021 tcp_try_keep_open(sk);
3022 } else {
3023 /* Partial ACK arrived. Force fast retransmit. */
3024 *do_lost = tcp_force_fast_retransmit(sk);
3025 }
3026 return false;
3027 }
3028
3029 static void tcp_identify_packet_loss(struct sock *sk, int *ack_flag)
3030 {
3031 struct tcp_sock *tp = tcp_sk(sk);
3032
3033 if (tcp_rtx_queue_empty(sk))
3034 return;
3035
3036 if (unlikely(tcp_is_reno(tp))) {
3037 tcp_newreno_mark_lost(sk, *ack_flag & FLAG_SND_UNA_ADVANCED);
3038 } else if (tcp_is_rack(sk)) {
3039 u32 prior_retrans = tp->retrans_out;
3040
3041 if (tcp_rack_mark_lost(sk))
3042 *ack_flag &= ~FLAG_SET_XMIT_TIMER;
3043 if (prior_retrans > tp->retrans_out)
3044 *ack_flag |= FLAG_LOST_RETRANS;
3045 }
3046 }
3047
3048 /* Process an event, which can update packets-in-flight not trivially.
3049 * Main goal of this function is to calculate new estimate for left_out,
3050 * taking into account both packets sitting in receiver's buffer and
3051 * packets lost by network.
3052 *
3053 * Besides that it updates the congestion state when packet loss or ECN
3054 * is detected. But it does not reduce the cwnd, it is done by the
3055 * congestion control later.
3056 *
3057 * It does _not_ decide what to send, it is made in function
3058 * tcp_xmit_retransmit_queue().
3059 */
3060 static void tcp_fastretrans_alert(struct sock *sk, const u32 prior_snd_una,
3061 int num_dupack, int *ack_flag, int *rexmit)
3062 {
3063 struct inet_connection_sock *icsk = inet_csk(sk);
3064 struct tcp_sock *tp = tcp_sk(sk);
3065 int fast_rexmit = 0, flag = *ack_flag;
3066 bool ece_ack = flag & FLAG_ECE;
3067 bool do_lost = num_dupack || ((flag & FLAG_DATA_SACKED) &&
3068 tcp_force_fast_retransmit(sk));
3069
3070 if (!tp->packets_out && tp->sacked_out)
3071 tp->sacked_out = 0;
3072
3073 /* Now state machine starts.
3074 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
3075 if (ece_ack)
3076 tp->prior_ssthresh = 0;
3077
3078 /* B. In all the states check for reneging SACKs. */
3079 if (tcp_check_sack_reneging(sk, ack_flag))
3080 return;
3081
3082 /* C. Check consistency of the current state. */
3083 tcp_verify_left_out(tp);
3084
3085 /* D. Check state exit conditions. State can be terminated
3086 * when high_seq is ACKed. */
3087 if (icsk->icsk_ca_state == TCP_CA_Open) {
3088 WARN_ON(tp->retrans_out != 0 && !tp->syn_data);
3089 tp->retrans_stamp = 0;
3090 } else if (!before(tp->snd_una, tp->high_seq)) {
3091 switch (icsk->icsk_ca_state) {
3092 case TCP_CA_CWR:
3093 /* CWR is to be held something *above* high_seq
3094 * is ACKed for CWR bit to reach receiver. */
3095 if (tp->snd_una != tp->high_seq) {
3096 tcp_end_cwnd_reduction(sk);
3097 tcp_set_ca_state(sk, TCP_CA_Open);
3098 }
3099 break;
3100
3101 case TCP_CA_Recovery:
3102 if (tcp_is_reno(tp))
3103 tcp_reset_reno_sack(tp);
3104 if (tcp_try_undo_recovery(sk))
3105 return;
3106 tcp_end_cwnd_reduction(sk);
3107 break;
3108 }
3109 }
3110
3111 /* E. Process state. */
3112 switch (icsk->icsk_ca_state) {
3113 case TCP_CA_Recovery:
3114 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
3115 if (tcp_is_reno(tp))
3116 tcp_add_reno_sack(sk, num_dupack, ece_ack);
3117 } else if (tcp_try_undo_partial(sk, prior_snd_una, &do_lost))
3118 return;
3119
3120 if (tcp_try_undo_dsack(sk))
3121 tcp_try_to_open(sk, flag);
3122
3123 tcp_identify_packet_loss(sk, ack_flag);
3124 if (icsk->icsk_ca_state != TCP_CA_Recovery) {
3125 if (!tcp_time_to_recover(sk, flag))
3126 return;
3127 /* Undo reverts the recovery state. If loss is evident,
3128 * starts a new recovery (e.g. reordering then loss);
3129 */
3130 tcp_enter_recovery(sk, ece_ack);
3131 }
3132 break;
3133 case TCP_CA_Loss:
3134 tcp_process_loss(sk, flag, num_dupack, rexmit);
3135 if (icsk->icsk_ca_state != TCP_CA_Loss)
3136 tcp_update_rto_time(tp);
3137 tcp_identify_packet_loss(sk, ack_flag);
3138 if (!(icsk->icsk_ca_state == TCP_CA_Open ||
3139 (*ack_flag & FLAG_LOST_RETRANS)))
3140 return;
3141 /* Change state if cwnd is undone or retransmits are lost */
3142 fallthrough;
3143 default:
3144 if (tcp_is_reno(tp)) {
3145 if (flag & FLAG_SND_UNA_ADVANCED)
3146 tcp_reset_reno_sack(tp);
3147 tcp_add_reno_sack(sk, num_dupack, ece_ack);
3148 }
3149
3150 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
3151 tcp_try_undo_dsack(sk);
3152
3153 tcp_identify_packet_loss(sk, ack_flag);
3154 if (!tcp_time_to_recover(sk, flag)) {
3155 tcp_try_to_open(sk, flag);
3156 return;
3157 }
3158
3159 /* MTU probe failure: don't reduce cwnd */
3160 if (icsk->icsk_ca_state < TCP_CA_CWR &&
3161 icsk->icsk_mtup.probe_size &&
3162 tp->snd_una == tp->mtu_probe.probe_seq_start) {
3163 tcp_mtup_probe_failed(sk);
3164 /* Restores the reduction we did in tcp_mtup_probe() */
3165 tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) + 1);
3166 tcp_simple_retransmit(sk);
3167 return;
3168 }
3169
3170 /* Otherwise enter Recovery state */
3171 tcp_enter_recovery(sk, ece_ack);
3172 fast_rexmit = 1;
3173 }
3174
3175 if (!tcp_is_rack(sk) && do_lost)
3176 tcp_update_scoreboard(sk, fast_rexmit);
3177 *rexmit = REXMIT_LOST;
3178 }
3179
3180 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us, const int flag)
3181 {
3182 u32 wlen = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_rtt_wlen) * HZ;
3183 struct tcp_sock *tp = tcp_sk(sk);
3184
3185 if ((flag & FLAG_ACK_MAYBE_DELAYED) && rtt_us > tcp_min_rtt(tp)) {
3186 /* If the remote keeps returning delayed ACKs, eventually
3187 * the min filter would pick it up and overestimate the
3188 * prop. delay when it expires. Skip suspected delayed ACKs.
3189 */
3190 return;
3191 }
3192 minmax_running_min(&tp->rtt_min, wlen, tcp_jiffies32,
3193 rtt_us ? : jiffies_to_usecs(1));
3194 }
3195
3196 static bool tcp_ack_update_rtt(struct sock *sk, const int flag,
3197 long seq_rtt_us, long sack_rtt_us,
3198 long ca_rtt_us, struct rate_sample *rs)
3199 {
3200 const struct tcp_sock *tp = tcp_sk(sk);
3201
3202 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
3203 * broken middle-boxes or peers may corrupt TS-ECR fields. But
3204 * Karn's algorithm forbids taking RTT if some retransmitted data
3205 * is acked (RFC6298).
3206 */
3207 if (seq_rtt_us < 0)
3208 seq_rtt_us = sack_rtt_us;
3209
3210 /* RTTM Rule: A TSecr value received in a segment is used to
3211 * update the averaged RTT measurement only if the segment
3212 * acknowledges some new data, i.e., only if it advances the
3213 * left edge of the send window.
3214 * See draft-ietf-tcplw-high-performance-00, section 3.3.
3215 */
3216 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp &&
3217 tp->rx_opt.rcv_tsecr && flag & FLAG_ACKED)
3218 seq_rtt_us = ca_rtt_us = tcp_rtt_tsopt_us(tp);
3219
3220 rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */
3221 if (seq_rtt_us < 0)
3222 return false;
3223
3224 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
3225 * always taken together with ACK, SACK, or TS-opts. Any negative
3226 * values will be skipped with the seq_rtt_us < 0 check above.
3227 */
3228 tcp_update_rtt_min(sk, ca_rtt_us, flag);
3229 tcp_rtt_estimator(sk, seq_rtt_us);
3230 tcp_set_rto(sk);
3231
3232 /* RFC6298: only reset backoff on valid RTT measurement. */
3233 inet_csk(sk)->icsk_backoff = 0;
3234 return true;
3235 }
3236
3237 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
3238 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
3239 {
3240 struct rate_sample rs;
3241 long rtt_us = -1L;
3242
3243 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack)
3244 rtt_us = tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req)->snt_synack);
3245
3246 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us, &rs);
3247 }
3248
3249
3250 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
3251 {
3252 const struct inet_connection_sock *icsk = inet_csk(sk);
3253
3254 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
3255 tcp_sk(sk)->snd_cwnd_stamp = tcp_jiffies32;
3256 }
3257
3258 /* Restart timer after forward progress on connection.
3259 * RFC2988 recommends to restart timer to now+rto.
3260 */
3261 void tcp_rearm_rto(struct sock *sk)
3262 {
3263 const struct inet_connection_sock *icsk = inet_csk(sk);
3264 struct tcp_sock *tp = tcp_sk(sk);
3265
3266 /* If the retrans timer is currently being used by Fast Open
3267 * for SYN-ACK retrans purpose, stay put.
3268 */
3269 if (rcu_access_pointer(tp->fastopen_rsk))
3270 return;
3271
3272 if (!tp->packets_out) {
3273 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3274 } else {
3275 u32 rto = inet_csk(sk)->icsk_rto;
3276 /* Offset the time elapsed after installing regular RTO */
3277 if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT ||
3278 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
3279 s64 delta_us = tcp_rto_delta_us(sk);
3280 /* delta_us may not be positive if the socket is locked
3281 * when the retrans timer fires and is rescheduled.
3282 */
3283 rto = usecs_to_jiffies(max_t(int, delta_us, 1));
3284 }
3285 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
3286 TCP_RTO_MAX);
3287 }
3288 }
3289
3290 /* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */
3291 static void tcp_set_xmit_timer(struct sock *sk)
3292 {
3293 if (!tcp_schedule_loss_probe(sk, true))
3294 tcp_rearm_rto(sk);
3295 }
3296
3297 /* If we get here, the whole TSO packet has not been acked. */
3298 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3299 {
3300 struct tcp_sock *tp = tcp_sk(sk);
3301 u32 packets_acked;
3302
3303 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3304
3305 packets_acked = tcp_skb_pcount(skb);
3306 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3307 return 0;
3308 packets_acked -= tcp_skb_pcount(skb);
3309
3310 if (packets_acked) {
3311 BUG_ON(tcp_skb_pcount(skb) == 0);
3312 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3313 }
3314
3315 return packets_acked;
3316 }
3317
3318 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3319 const struct sk_buff *ack_skb, u32 prior_snd_una)
3320 {
3321 const struct skb_shared_info *shinfo;
3322
3323 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3324 if (likely(!TCP_SKB_CB(skb)->txstamp_ack))
3325 return;
3326
3327 shinfo = skb_shinfo(skb);
3328 if (!before(shinfo->tskey, prior_snd_una) &&
3329 before(shinfo->tskey, tcp_sk(sk)->snd_una)) {
3330 tcp_skb_tsorted_save(skb) {
3331 __skb_tstamp_tx(skb, ack_skb, NULL, sk, SCM_TSTAMP_ACK);
3332 } tcp_skb_tsorted_restore(skb);
3333 }
3334 }
3335
3336 /* Remove acknowledged frames from the retransmission queue. If our packet
3337 * is before the ack sequence we can discard it as it's confirmed to have
3338 * arrived at the other end.
3339 */
3340 static int tcp_clean_rtx_queue(struct sock *sk, const struct sk_buff *ack_skb,
3341 u32 prior_fack, u32 prior_snd_una,
3342 struct tcp_sacktag_state *sack, bool ece_ack)
3343 {
3344 const struct inet_connection_sock *icsk = inet_csk(sk);
3345 u64 first_ackt, last_ackt;
3346 struct tcp_sock *tp = tcp_sk(sk);
3347 u32 prior_sacked = tp->sacked_out;
3348 u32 reord = tp->snd_nxt; /* lowest acked un-retx un-sacked seq */
3349 struct sk_buff *skb, *next;
3350 bool fully_acked = true;
3351 long sack_rtt_us = -1L;
3352 long seq_rtt_us = -1L;
3353 long ca_rtt_us = -1L;
3354 u32 pkts_acked = 0;
3355 bool rtt_update;
3356 int flag = 0;
3357
3358 first_ackt = 0;
3359
3360 for (skb = skb_rb_first(&sk->tcp_rtx_queue); skb; skb = next) {
3361 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3362 const u32 start_seq = scb->seq;
3363 u8 sacked = scb->sacked;
3364 u32 acked_pcount;
3365
3366 /* Determine how many packets and what bytes were acked, tso and else */
3367 if (after(scb->end_seq, tp->snd_una)) {
3368 if (tcp_skb_pcount(skb) == 1 ||
3369 !after(tp->snd_una, scb->seq))
3370 break;
3371
3372 acked_pcount = tcp_tso_acked(sk, skb);
3373 if (!acked_pcount)
3374 break;
3375 fully_acked = false;
3376 } else {
3377 acked_pcount = tcp_skb_pcount(skb);
3378 }
3379
3380 if (unlikely(sacked & TCPCB_RETRANS)) {
3381 if (sacked & TCPCB_SACKED_RETRANS)
3382 tp->retrans_out -= acked_pcount;
3383 flag |= FLAG_RETRANS_DATA_ACKED;
3384 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3385 last_ackt = tcp_skb_timestamp_us(skb);
3386 WARN_ON_ONCE(last_ackt == 0);
3387 if (!first_ackt)
3388 first_ackt = last_ackt;
3389
3390 if (before(start_seq, reord))
3391 reord = start_seq;
3392 if (!after(scb->end_seq, tp->high_seq))
3393 flag |= FLAG_ORIG_SACK_ACKED;
3394 }
3395
3396 if (sacked & TCPCB_SACKED_ACKED) {
3397 tp->sacked_out -= acked_pcount;
3398 } else if (tcp_is_sack(tp)) {
3399 tcp_count_delivered(tp, acked_pcount, ece_ack);
3400 if (!tcp_skb_spurious_retrans(tp, skb))
3401 tcp_rack_advance(tp, sacked, scb->end_seq,
3402 tcp_skb_timestamp_us(skb));
3403 }
3404 if (sacked & TCPCB_LOST)
3405 tp->lost_out -= acked_pcount;
3406
3407 tp->packets_out -= acked_pcount;
3408 pkts_acked += acked_pcount;
3409 tcp_rate_skb_delivered(sk, skb, sack->rate);
3410
3411 /* Initial outgoing SYN's get put onto the write_queue
3412 * just like anything else we transmit. It is not
3413 * true data, and if we misinform our callers that
3414 * this ACK acks real data, we will erroneously exit
3415 * connection startup slow start one packet too
3416 * quickly. This is severely frowned upon behavior.
3417 */
3418 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3419 flag |= FLAG_DATA_ACKED;
3420 } else {
3421 flag |= FLAG_SYN_ACKED;
3422 tp->retrans_stamp = 0;
3423 }
3424
3425 if (!fully_acked)
3426 break;
3427
3428 tcp_ack_tstamp(sk, skb, ack_skb, prior_snd_una);
3429
3430 next = skb_rb_next(skb);
3431 if (unlikely(skb == tp->retransmit_skb_hint))
3432 tp->retransmit_skb_hint = NULL;
3433 if (unlikely(skb == tp->lost_skb_hint))
3434 tp->lost_skb_hint = NULL;
3435 tcp_highest_sack_replace(sk, skb, next);
3436 tcp_rtx_queue_unlink_and_free(skb, sk);
3437 }
3438
3439 if (!skb)
3440 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3441
3442 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3443 tp->snd_up = tp->snd_una;
3444
3445 if (skb) {
3446 tcp_ack_tstamp(sk, skb, ack_skb, prior_snd_una);
3447 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
3448 flag |= FLAG_SACK_RENEGING;
3449 }
3450
3451 if (likely(first_ackt) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
3452 seq_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, first_ackt);
3453 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, last_ackt);
3454
3455 if (pkts_acked == 1 && fully_acked && !prior_sacked &&
3456 (tp->snd_una - prior_snd_una) < tp->mss_cache &&
3457 sack->rate->prior_delivered + 1 == tp->delivered &&
3458 !(flag & (FLAG_CA_ALERT | FLAG_SYN_ACKED))) {
3459 /* Conservatively mark a delayed ACK. It's typically
3460 * from a lone runt packet over the round trip to
3461 * a receiver w/o out-of-order or CE events.
3462 */
3463 flag |= FLAG_ACK_MAYBE_DELAYED;
3464 }
3465 }
3466 if (sack->first_sackt) {
3467 sack_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->first_sackt);
3468 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->last_sackt);
3469 }
3470 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
3471 ca_rtt_us, sack->rate);
3472
3473 if (flag & FLAG_ACKED) {
3474 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3475 if (unlikely(icsk->icsk_mtup.probe_size &&
3476 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3477 tcp_mtup_probe_success(sk);
3478 }
3479
3480 if (tcp_is_reno(tp)) {
3481 tcp_remove_reno_sacks(sk, pkts_acked, ece_ack);
3482
3483 /* If any of the cumulatively ACKed segments was
3484 * retransmitted, non-SACK case cannot confirm that
3485 * progress was due to original transmission due to
3486 * lack of TCPCB_SACKED_ACKED bits even if some of
3487 * the packets may have been never retransmitted.
3488 */
3489 if (flag & FLAG_RETRANS_DATA_ACKED)
3490 flag &= ~FLAG_ORIG_SACK_ACKED;
3491 } else {
3492 int delta;
3493
3494 /* Non-retransmitted hole got filled? That's reordering */
3495 if (before(reord, prior_fack))
3496 tcp_check_sack_reordering(sk, reord, 0);
3497
3498 delta = prior_sacked - tp->sacked_out;
3499 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3500 }
3501 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3502 sack_rtt_us > tcp_stamp_us_delta(tp->tcp_mstamp,
3503 tcp_skb_timestamp_us(skb))) {
3504 /* Do not re-arm RTO if the sack RTT is measured from data sent
3505 * after when the head was last (re)transmitted. Otherwise the
3506 * timeout may continue to extend in loss recovery.
3507 */
3508 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3509 }
3510
3511 if (icsk->icsk_ca_ops->pkts_acked) {
3512 struct ack_sample sample = { .pkts_acked = pkts_acked,
3513 .rtt_us = sack->rate->rtt_us };
3514
3515 sample.in_flight = tp->mss_cache *
3516 (tp->delivered - sack->rate->prior_delivered);
3517 icsk->icsk_ca_ops->pkts_acked(sk, &sample);
3518 }
3519
3520 #if FASTRETRANS_DEBUG > 0
3521 WARN_ON((int)tp->sacked_out < 0);
3522 WARN_ON((int)tp->lost_out < 0);
3523 WARN_ON((int)tp->retrans_out < 0);
3524 if (!tp->packets_out && tcp_is_sack(tp)) {
3525 icsk = inet_csk(sk);
3526 if (tp->lost_out) {
3527 pr_debug("Leak l=%u %d\n",
3528 tp->lost_out, icsk->icsk_ca_state);
3529 tp->lost_out = 0;
3530 }
3531 if (tp->sacked_out) {
3532 pr_debug("Leak s=%u %d\n",
3533 tp->sacked_out, icsk->icsk_ca_state);
3534 tp->sacked_out = 0;
3535 }
3536 if (tp->retrans_out) {
3537 pr_debug("Leak r=%u %d\n",
3538 tp->retrans_out, icsk->icsk_ca_state);
3539 tp->retrans_out = 0;
3540 }
3541 }
3542 #endif
3543 return flag;
3544 }
3545
3546 static void tcp_ack_probe(struct sock *sk)
3547 {
3548 struct inet_connection_sock *icsk = inet_csk(sk);
3549 struct sk_buff *head = tcp_send_head(sk);
3550 const struct tcp_sock *tp = tcp_sk(sk);
3551
3552 /* Was it a usable window open? */
3553 if (!head)
3554 return;
3555 if (!after(TCP_SKB_CB(head)->end_seq, tcp_wnd_end(tp))) {
3556 icsk->icsk_backoff = 0;
3557 icsk->icsk_probes_tstamp = 0;
3558 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3559 /* Socket must be waked up by subsequent tcp_data_snd_check().
3560 * This function is not for random using!
3561 */
3562 } else {
3563 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
3564
3565 when = tcp_clamp_probe0_to_user_timeout(sk, when);
3566 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, when, TCP_RTO_MAX);
3567 }
3568 }
3569
3570 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3571 {
3572 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3573 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3574 }
3575
3576 /* Decide wheather to run the increase function of congestion control. */
3577 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3578 {
3579 /* If reordering is high then always grow cwnd whenever data is
3580 * delivered regardless of its ordering. Otherwise stay conservative
3581 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3582 * new SACK or ECE mark may first advance cwnd here and later reduce
3583 * cwnd in tcp_fastretrans_alert() based on more states.
3584 */
3585 if (tcp_sk(sk)->reordering >
3586 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_reordering))
3587 return flag & FLAG_FORWARD_PROGRESS;
3588
3589 return flag & FLAG_DATA_ACKED;
3590 }
3591
3592 /* The "ultimate" congestion control function that aims to replace the rigid
3593 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3594 * It's called toward the end of processing an ACK with precise rate
3595 * information. All transmission or retransmission are delayed afterwards.
3596 */
3597 static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
3598 int flag, const struct rate_sample *rs)
3599 {
3600 const struct inet_connection_sock *icsk = inet_csk(sk);
3601
3602 if (icsk->icsk_ca_ops->cong_control) {
3603 icsk->icsk_ca_ops->cong_control(sk, ack, flag, rs);
3604 return;
3605 }
3606
3607 if (tcp_in_cwnd_reduction(sk)) {
3608 /* Reduce cwnd if state mandates */
3609 tcp_cwnd_reduction(sk, acked_sacked, rs->losses, flag);
3610 } else if (tcp_may_raise_cwnd(sk, flag)) {
3611 /* Advance cwnd if state allows */
3612 tcp_cong_avoid(sk, ack, acked_sacked);
3613 }
3614 tcp_update_pacing_rate(sk);
3615 }
3616
3617 /* Check that window update is acceptable.
3618 * The function assumes that snd_una<=ack<=snd_next.
3619 */
3620 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3621 const u32 ack, const u32 ack_seq,
3622 const u32 nwin)
3623 {
3624 return after(ack, tp->snd_una) ||
3625 after(ack_seq, tp->snd_wl1) ||
3626 (ack_seq == tp->snd_wl1 && (nwin > tp->snd_wnd || !nwin));
3627 }
3628
3629 static void tcp_snd_sne_update(struct tcp_sock *tp, u32 ack)
3630 {
3631 #ifdef CONFIG_TCP_AO
3632 struct tcp_ao_info *ao;
3633
3634 if (!static_branch_unlikely(&tcp_ao_needed.key))
3635 return;
3636
3637 ao = rcu_dereference_protected(tp->ao_info,
3638 lockdep_sock_is_held((struct sock *)tp));
3639 if (ao && ack < tp->snd_una) {
3640 ao->snd_sne++;
3641 trace_tcp_ao_snd_sne_update((struct sock *)tp, ao->snd_sne);
3642 }
3643 #endif
3644 }
3645
3646 /* If we update tp->snd_una, also update tp->bytes_acked */
3647 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3648 {
3649 u32 delta = ack - tp->snd_una;
3650
3651 sock_owned_by_me((struct sock *)tp);
3652 tp->bytes_acked += delta;
3653 tcp_snd_sne_update(tp, ack);
3654 tp->snd_una = ack;
3655 }
3656
3657 static void tcp_rcv_sne_update(struct tcp_sock *tp, u32 seq)
3658 {
3659 #ifdef CONFIG_TCP_AO
3660 struct tcp_ao_info *ao;
3661
3662 if (!static_branch_unlikely(&tcp_ao_needed.key))
3663 return;
3664
3665 ao = rcu_dereference_protected(tp->ao_info,
3666 lockdep_sock_is_held((struct sock *)tp));
3667 if (ao && seq < tp->rcv_nxt) {
3668 ao->rcv_sne++;
3669 trace_tcp_ao_rcv_sne_update((struct sock *)tp, ao->rcv_sne);
3670 }
3671 #endif
3672 }
3673
3674 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3675 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3676 {
3677 u32 delta = seq - tp->rcv_nxt;
3678
3679 sock_owned_by_me((struct sock *)tp);
3680 tp->bytes_received += delta;
3681 tcp_rcv_sne_update(tp, seq);
3682 WRITE_ONCE(tp->rcv_nxt, seq);
3683 }
3684
3685 /* Update our send window.
3686 *
3687 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3688 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3689 */
3690 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3691 u32 ack_seq)
3692 {
3693 struct tcp_sock *tp = tcp_sk(sk);
3694 int flag = 0;
3695 u32 nwin = ntohs(tcp_hdr(skb)->window);
3696
3697 if (likely(!tcp_hdr(skb)->syn))
3698 nwin <<= tp->rx_opt.snd_wscale;
3699
3700 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3701 flag |= FLAG_WIN_UPDATE;
3702 tcp_update_wl(tp, ack_seq);
3703
3704 if (tp->snd_wnd != nwin) {
3705 tp->snd_wnd = nwin;
3706
3707 /* Note, it is the only place, where
3708 * fast path is recovered for sending TCP.
3709 */
3710 tp->pred_flags = 0;
3711 tcp_fast_path_check(sk);
3712
3713 if (!tcp_write_queue_empty(sk))
3714 tcp_slow_start_after_idle_check(sk);
3715
3716 if (nwin > tp->max_window) {
3717 tp->max_window = nwin;
3718 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3719 }
3720 }
3721 }
3722
3723 tcp_snd_una_update(tp, ack);
3724
3725 return flag;
3726 }
3727
3728 static bool __tcp_oow_rate_limited(struct net *net, int mib_idx,
3729 u32 *last_oow_ack_time)
3730 {
3731 /* Paired with the WRITE_ONCE() in this function. */
3732 u32 val = READ_ONCE(*last_oow_ack_time);
3733
3734 if (val) {
3735 s32 elapsed = (s32)(tcp_jiffies32 - val);
3736
3737 if (0 <= elapsed &&
3738 elapsed < READ_ONCE(net->ipv4.sysctl_tcp_invalid_ratelimit)) {
3739 NET_INC_STATS(net, mib_idx);
3740 return true; /* rate-limited: don't send yet! */
3741 }
3742 }
3743
3744 /* Paired with the prior READ_ONCE() and with itself,
3745 * as we might be lockless.
3746 */
3747 WRITE_ONCE(*last_oow_ack_time, tcp_jiffies32);
3748
3749 return false; /* not rate-limited: go ahead, send dupack now! */
3750 }
3751
3752 /* Return true if we're currently rate-limiting out-of-window ACKs and
3753 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3754 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3755 * attacks that send repeated SYNs or ACKs for the same connection. To
3756 * do this, we do not send a duplicate SYNACK or ACK if the remote
3757 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3758 */
3759 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3760 int mib_idx, u32 *last_oow_ack_time)
3761 {
3762 /* Data packets without SYNs are not likely part of an ACK loop. */
3763 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3764 !tcp_hdr(skb)->syn)
3765 return false;
3766
3767 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time);
3768 }
3769
3770 /* RFC 5961 7 [ACK Throttling] */
3771 static void tcp_send_challenge_ack(struct sock *sk)
3772 {
3773 struct tcp_sock *tp = tcp_sk(sk);
3774 struct net *net = sock_net(sk);
3775 u32 count, now, ack_limit;
3776
3777 /* First check our per-socket dupack rate limit. */
3778 if (__tcp_oow_rate_limited(net,
3779 LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3780 &tp->last_oow_ack_time))
3781 return;
3782
3783 ack_limit = READ_ONCE(net->ipv4.sysctl_tcp_challenge_ack_limit);
3784 if (ack_limit == INT_MAX)
3785 goto send_ack;
3786
3787 /* Then check host-wide RFC 5961 rate limit. */
3788 now = jiffies / HZ;
3789 if (now != READ_ONCE(net->ipv4.tcp_challenge_timestamp)) {
3790 u32 half = (ack_limit + 1) >> 1;
3791
3792 WRITE_ONCE(net->ipv4.tcp_challenge_timestamp, now);
3793 WRITE_ONCE(net->ipv4.tcp_challenge_count,
3794 get_random_u32_inclusive(half, ack_limit + half - 1));
3795 }
3796 count = READ_ONCE(net->ipv4.tcp_challenge_count);
3797 if (count > 0) {
3798 WRITE_ONCE(net->ipv4.tcp_challenge_count, count - 1);
3799 send_ack:
3800 NET_INC_STATS(net, LINUX_MIB_TCPCHALLENGEACK);
3801 tcp_send_ack(sk);
3802 }
3803 }
3804
3805 static void tcp_store_ts_recent(struct tcp_sock *tp)
3806 {
3807 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3808 tp->rx_opt.ts_recent_stamp = ktime_get_seconds();
3809 }
3810
3811 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3812 {
3813 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3814 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3815 * extra check below makes sure this can only happen
3816 * for pure ACK frames. -DaveM
3817 *
3818 * Not only, also it occurs for expired timestamps.
3819 */
3820
3821 if (tcp_paws_check(&tp->rx_opt, 0))
3822 tcp_store_ts_recent(tp);
3823 }
3824 }
3825
3826 /* This routine deals with acks during a TLP episode and ends an episode by
3827 * resetting tlp_high_seq. Ref: TLP algorithm in draft-ietf-tcpm-rack
3828 */
3829 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3830 {
3831 struct tcp_sock *tp = tcp_sk(sk);
3832
3833 if (before(ack, tp->tlp_high_seq))
3834 return;
3835
3836 if (!tp->tlp_retrans) {
3837 /* TLP of new data has been acknowledged */
3838 tp->tlp_high_seq = 0;
3839 } else if (flag & FLAG_DSACK_TLP) {
3840 /* This DSACK means original and TLP probe arrived; no loss */
3841 tp->tlp_high_seq = 0;
3842 } else if (after(ack, tp->tlp_high_seq)) {
3843 /* ACK advances: there was a loss, so reduce cwnd. Reset
3844 * tlp_high_seq in tcp_init_cwnd_reduction()
3845 */
3846 tcp_init_cwnd_reduction(sk);
3847 tcp_set_ca_state(sk, TCP_CA_CWR);
3848 tcp_end_cwnd_reduction(sk);
3849 tcp_try_keep_open(sk);
3850 NET_INC_STATS(sock_net(sk),
3851 LINUX_MIB_TCPLOSSPROBERECOVERY);
3852 } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3853 FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3854 /* Pure dupack: original and TLP probe arrived; no loss */
3855 tp->tlp_high_seq = 0;
3856 }
3857 }
3858
3859 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3860 {
3861 const struct inet_connection_sock *icsk = inet_csk(sk);
3862
3863 if (icsk->icsk_ca_ops->in_ack_event)
3864 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3865 }
3866
3867 /* Congestion control has updated the cwnd already. So if we're in
3868 * loss recovery then now we do any new sends (for FRTO) or
3869 * retransmits (for CA_Loss or CA_recovery) that make sense.
3870 */
3871 static void tcp_xmit_recovery(struct sock *sk, int rexmit)
3872 {
3873 struct tcp_sock *tp = tcp_sk(sk);
3874
3875 if (rexmit == REXMIT_NONE || sk->sk_state == TCP_SYN_SENT)
3876 return;
3877
3878 if (unlikely(rexmit == REXMIT_NEW)) {
3879 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
3880 TCP_NAGLE_OFF);
3881 if (after(tp->snd_nxt, tp->high_seq))
3882 return;
3883 tp->frto = 0;
3884 }
3885 tcp_xmit_retransmit_queue(sk);
3886 }
3887
3888 /* Returns the number of packets newly acked or sacked by the current ACK */
3889 static u32 tcp_newly_delivered(struct sock *sk, u32 prior_delivered, int flag)
3890 {
3891 const struct net *net = sock_net(sk);
3892 struct tcp_sock *tp = tcp_sk(sk);
3893 u32 delivered;
3894
3895 delivered = tp->delivered - prior_delivered;
3896 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVERED, delivered);
3897 if (flag & FLAG_ECE)
3898 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVEREDCE, delivered);
3899
3900 return delivered;
3901 }
3902
3903 /* This routine deals with incoming acks, but not outgoing ones. */
3904 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3905 {
3906 struct inet_connection_sock *icsk = inet_csk(sk);
3907 struct tcp_sock *tp = tcp_sk(sk);
3908 struct tcp_sacktag_state sack_state;
3909 struct rate_sample rs = { .prior_delivered = 0 };
3910 u32 prior_snd_una = tp->snd_una;
3911 bool is_sack_reneg = tp->is_sack_reneg;
3912 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3913 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3914 int num_dupack = 0;
3915 int prior_packets = tp->packets_out;
3916 u32 delivered = tp->delivered;
3917 u32 lost = tp->lost;
3918 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
3919 u32 prior_fack;
3920
3921 sack_state.first_sackt = 0;
3922 sack_state.rate = &rs;
3923 sack_state.sack_delivered = 0;
3924
3925 /* We very likely will need to access rtx queue. */
3926 prefetch(sk->tcp_rtx_queue.rb_node);
3927
3928 /* If the ack is older than previous acks
3929 * then we can probably ignore it.
3930 */
3931 if (before(ack, prior_snd_una)) {
3932 u32 max_window;
3933
3934 /* do not accept ACK for bytes we never sent. */
3935 max_window = min_t(u64, tp->max_window, tp->bytes_acked);
3936 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3937 if (before(ack, prior_snd_una - max_window)) {
3938 if (!(flag & FLAG_NO_CHALLENGE_ACK))
3939 tcp_send_challenge_ack(sk);
3940 return -SKB_DROP_REASON_TCP_TOO_OLD_ACK;
3941 }
3942 goto old_ack;
3943 }
3944
3945 /* If the ack includes data we haven't sent yet, discard
3946 * this segment (RFC793 Section 3.9).
3947 */
3948 if (after(ack, tp->snd_nxt))
3949 return -SKB_DROP_REASON_TCP_ACK_UNSENT_DATA;
3950
3951 if (after(ack, prior_snd_una)) {
3952 flag |= FLAG_SND_UNA_ADVANCED;
3953 icsk->icsk_retransmits = 0;
3954
3955 #if IS_ENABLED(CONFIG_TLS_DEVICE)
3956 if (static_branch_unlikely(&clean_acked_data_enabled.key))
3957 if (icsk->icsk_clean_acked)
3958 icsk->icsk_clean_acked(sk, ack);
3959 #endif
3960 }
3961
3962 prior_fack = tcp_is_sack(tp) ? tcp_highest_sack_seq(tp) : tp->snd_una;
3963 rs.prior_in_flight = tcp_packets_in_flight(tp);
3964
3965 /* ts_recent update must be made after we are sure that the packet
3966 * is in window.
3967 */
3968 if (flag & FLAG_UPDATE_TS_RECENT)
3969 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3970
3971 if ((flag & (FLAG_SLOWPATH | FLAG_SND_UNA_ADVANCED)) ==
3972 FLAG_SND_UNA_ADVANCED) {
3973 /* Window is constant, pure forward advance.
3974 * No more checks are required.
3975 * Note, we use the fact that SND.UNA>=SND.WL2.
3976 */
3977 tcp_update_wl(tp, ack_seq);
3978 tcp_snd_una_update(tp, ack);
3979 flag |= FLAG_WIN_UPDATE;
3980
3981 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3982
3983 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS);
3984 } else {
3985 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3986
3987 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3988 flag |= FLAG_DATA;
3989 else
3990 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3991
3992 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3993
3994 if (TCP_SKB_CB(skb)->sacked)
3995 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3996 &sack_state);
3997
3998 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3999 flag |= FLAG_ECE;
4000 ack_ev_flags |= CA_ACK_ECE;
4001 }
4002
4003 if (sack_state.sack_delivered)
4004 tcp_count_delivered(tp, sack_state.sack_delivered,
4005 flag & FLAG_ECE);
4006
4007 if (flag & FLAG_WIN_UPDATE)
4008 ack_ev_flags |= CA_ACK_WIN_UPDATE;
4009
4010 tcp_in_ack_event(sk, ack_ev_flags);
4011 }
4012
4013 /* This is a deviation from RFC3168 since it states that:
4014 * "When the TCP data sender is ready to set the CWR bit after reducing
4015 * the congestion window, it SHOULD set the CWR bit only on the first
4016 * new data packet that it transmits."
4017 * We accept CWR on pure ACKs to be more robust
4018 * with widely-deployed TCP implementations that do this.
4019 */
4020 tcp_ecn_accept_cwr(sk, skb);
4021
4022 /* We passed data and got it acked, remove any soft error
4023 * log. Something worked...
4024 */
4025 WRITE_ONCE(sk->sk_err_soft, 0);
4026 icsk->icsk_probes_out = 0;
4027 tp->rcv_tstamp = tcp_jiffies32;
4028 if (!prior_packets)
4029 goto no_queue;
4030
4031 /* See if we can take anything off of the retransmit queue. */
4032 flag |= tcp_clean_rtx_queue(sk, skb, prior_fack, prior_snd_una,
4033 &sack_state, flag & FLAG_ECE);
4034
4035 tcp_rack_update_reo_wnd(sk, &rs);
4036
4037 if (tp->tlp_high_seq)
4038 tcp_process_tlp_ack(sk, ack, flag);
4039
4040 if (tcp_ack_is_dubious(sk, flag)) {
4041 if (!(flag & (FLAG_SND_UNA_ADVANCED |
4042 FLAG_NOT_DUP | FLAG_DSACKING_ACK))) {
4043 num_dupack = 1;
4044 /* Consider if pure acks were aggregated in tcp_add_backlog() */
4045 if (!(flag & FLAG_DATA))
4046 num_dupack = max_t(u16, 1, skb_shinfo(skb)->gso_segs);
4047 }
4048 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
4049 &rexmit);
4050 }
4051
4052 /* If needed, reset TLP/RTO timer when RACK doesn't set. */
4053 if (flag & FLAG_SET_XMIT_TIMER)
4054 tcp_set_xmit_timer(sk);
4055
4056 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
4057 sk_dst_confirm(sk);
4058
4059 delivered = tcp_newly_delivered(sk, delivered, flag);
4060 lost = tp->lost - lost; /* freshly marked lost */
4061 rs.is_ack_delayed = !!(flag & FLAG_ACK_MAYBE_DELAYED);
4062 tcp_rate_gen(sk, delivered, lost, is_sack_reneg, sack_state.rate);
4063 tcp_cong_control(sk, ack, delivered, flag, sack_state.rate);
4064 tcp_xmit_recovery(sk, rexmit);
4065 return 1;
4066
4067 no_queue:
4068 /* If data was DSACKed, see if we can undo a cwnd reduction. */
4069 if (flag & FLAG_DSACKING_ACK) {
4070 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
4071 &rexmit);
4072 tcp_newly_delivered(sk, delivered, flag);
4073 }
4074 /* If this ack opens up a zero window, clear backoff. It was
4075 * being used to time the probes, and is probably far higher than
4076 * it needs to be for normal retransmission.
4077 */
4078 tcp_ack_probe(sk);
4079
4080 if (tp->tlp_high_seq)
4081 tcp_process_tlp_ack(sk, ack, flag);
4082 return 1;
4083
4084 old_ack:
4085 /* If data was SACKed, tag it and see if we should send more data.
4086 * If data was DSACKed, see if we can undo a cwnd reduction.
4087 */
4088 if (TCP_SKB_CB(skb)->sacked) {
4089 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
4090 &sack_state);
4091 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
4092 &rexmit);
4093 tcp_newly_delivered(sk, delivered, flag);
4094 tcp_xmit_recovery(sk, rexmit);
4095 }
4096
4097 return 0;
4098 }
4099
4100 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
4101 bool syn, struct tcp_fastopen_cookie *foc,
4102 bool exp_opt)
4103 {
4104 /* Valid only in SYN or SYN-ACK with an even length. */
4105 if (!foc || !syn || len < 0 || (len & 1))
4106 return;
4107
4108 if (len >= TCP_FASTOPEN_COOKIE_MIN &&
4109 len <= TCP_FASTOPEN_COOKIE_MAX)
4110 memcpy(foc->val, cookie, len);
4111 else if (len != 0)
4112 len = -1;
4113 foc->len = len;
4114 foc->exp = exp_opt;
4115 }
4116
4117 static bool smc_parse_options(const struct tcphdr *th,
4118 struct tcp_options_received *opt_rx,
4119 const unsigned char *ptr,
4120 int opsize)
4121 {
4122 #if IS_ENABLED(CONFIG_SMC)
4123 if (static_branch_unlikely(&tcp_have_smc)) {
4124 if (th->syn && !(opsize & 1) &&
4125 opsize >= TCPOLEN_EXP_SMC_BASE &&
4126 get_unaligned_be32(ptr) == TCPOPT_SMC_MAGIC) {
4127 opt_rx->smc_ok = 1;
4128 return true;
4129 }
4130 }
4131 #endif
4132 return false;
4133 }
4134
4135 /* Try to parse the MSS option from the TCP header. Return 0 on failure, clamped
4136 * value on success.
4137 */
4138 u16 tcp_parse_mss_option(const struct tcphdr *th, u16 user_mss)
4139 {
4140 const unsigned char *ptr = (const unsigned char *)(th + 1);
4141 int length = (th->doff * 4) - sizeof(struct tcphdr);
4142 u16 mss = 0;
4143
4144 while (length > 0) {
4145 int opcode = *ptr++;
4146 int opsize;
4147
4148 switch (opcode) {
4149 case TCPOPT_EOL:
4150 return mss;
4151 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
4152 length--;
4153 continue;
4154 default:
4155 if (length < 2)
4156 return mss;
4157 opsize = *ptr++;
4158 if (opsize < 2) /* "silly options" */
4159 return mss;
4160 if (opsize > length)
4161 return mss; /* fail on partial options */
4162 if (opcode == TCPOPT_MSS && opsize == TCPOLEN_MSS) {
4163 u16 in_mss = get_unaligned_be16(ptr);
4164
4165 if (in_mss) {
4166 if (user_mss && user_mss < in_mss)
4167 in_mss = user_mss;
4168 mss = in_mss;
4169 }
4170 }
4171 ptr += opsize - 2;
4172 length -= opsize;
4173 }
4174 }
4175 return mss;
4176 }
4177 EXPORT_SYMBOL_GPL(tcp_parse_mss_option);
4178
4179 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
4180 * But, this can also be called on packets in the established flow when
4181 * the fast version below fails.
4182 */
4183 void tcp_parse_options(const struct net *net,
4184 const struct sk_buff *skb,
4185 struct tcp_options_received *opt_rx, int estab,
4186 struct tcp_fastopen_cookie *foc)
4187 {
4188 const unsigned char *ptr;
4189 const struct tcphdr *th = tcp_hdr(skb);
4190 int length = (th->doff * 4) - sizeof(struct tcphdr);
4191
4192 ptr = (const unsigned char *)(th + 1);
4193 opt_rx->saw_tstamp = 0;
4194 opt_rx->saw_unknown = 0;
4195
4196 while (length > 0) {
4197 int opcode = *ptr++;
4198 int opsize;
4199
4200 switch (opcode) {
4201 case TCPOPT_EOL:
4202 return;
4203 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
4204 length--;
4205 continue;
4206 default:
4207 if (length < 2)
4208 return;
4209 opsize = *ptr++;
4210 if (opsize < 2) /* "silly options" */
4211 return;
4212 if (opsize > length)
4213 return; /* don't parse partial options */
4214 switch (opcode) {
4215 case TCPOPT_MSS:
4216 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
4217 u16 in_mss = get_unaligned_be16(ptr);
4218 if (in_mss) {
4219 if (opt_rx->user_mss &&
4220 opt_rx->user_mss < in_mss)
4221 in_mss = opt_rx->user_mss;
4222 opt_rx->mss_clamp = in_mss;
4223 }
4224 }
4225 break;
4226 case TCPOPT_WINDOW:
4227 if (opsize == TCPOLEN_WINDOW && th->syn &&
4228 !estab && READ_ONCE(net->ipv4.sysctl_tcp_window_scaling)) {
4229 __u8 snd_wscale = *(__u8 *)ptr;
4230 opt_rx->wscale_ok = 1;
4231 if (snd_wscale > TCP_MAX_WSCALE) {
4232 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n",
4233 __func__,
4234 snd_wscale,
4235 TCP_MAX_WSCALE);
4236 snd_wscale = TCP_MAX_WSCALE;
4237 }
4238 opt_rx->snd_wscale = snd_wscale;
4239 }
4240 break;
4241 case TCPOPT_TIMESTAMP:
4242 if ((opsize == TCPOLEN_TIMESTAMP) &&
4243 ((estab && opt_rx->tstamp_ok) ||
4244 (!estab && READ_ONCE(net->ipv4.sysctl_tcp_timestamps)))) {
4245 opt_rx->saw_tstamp = 1;
4246 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
4247 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
4248 }
4249 break;
4250 case TCPOPT_SACK_PERM:
4251 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
4252 !estab && READ_ONCE(net->ipv4.sysctl_tcp_sack)) {
4253 opt_rx->sack_ok = TCP_SACK_SEEN;
4254 tcp_sack_reset(opt_rx);
4255 }
4256 break;
4257
4258 case TCPOPT_SACK:
4259 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
4260 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
4261 opt_rx->sack_ok) {
4262 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
4263 }
4264 break;
4265 #ifdef CONFIG_TCP_MD5SIG
4266 case TCPOPT_MD5SIG:
4267 /* The MD5 Hash has already been
4268 * checked (see tcp_v{4,6}_rcv()).
4269 */
4270 break;
4271 #endif
4272 #ifdef CONFIG_TCP_AO
4273 case TCPOPT_AO:
4274 /* TCP AO has already been checked
4275 * (see tcp_inbound_ao_hash()).
4276 */
4277 break;
4278 #endif
4279 case TCPOPT_FASTOPEN:
4280 tcp_parse_fastopen_option(
4281 opsize - TCPOLEN_FASTOPEN_BASE,
4282 ptr, th->syn, foc, false);
4283 break;
4284
4285 case TCPOPT_EXP:
4286 /* Fast Open option shares code 254 using a
4287 * 16 bits magic number.
4288 */
4289 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
4290 get_unaligned_be16(ptr) ==
4291 TCPOPT_FASTOPEN_MAGIC) {
4292 tcp_parse_fastopen_option(opsize -
4293 TCPOLEN_EXP_FASTOPEN_BASE,
4294 ptr + 2, th->syn, foc, true);
4295 break;
4296 }
4297
4298 if (smc_parse_options(th, opt_rx, ptr, opsize))
4299 break;
4300
4301 opt_rx->saw_unknown = 1;
4302 break;
4303
4304 default:
4305 opt_rx->saw_unknown = 1;
4306 }
4307 ptr += opsize-2;
4308 length -= opsize;
4309 }
4310 }
4311 }
4312 EXPORT_SYMBOL(tcp_parse_options);
4313
4314 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
4315 {
4316 const __be32 *ptr = (const __be32 *)(th + 1);
4317
4318 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
4319 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
4320 tp->rx_opt.saw_tstamp = 1;
4321 ++ptr;
4322 tp->rx_opt.rcv_tsval = ntohl(*ptr);
4323 ++ptr;
4324 if (*ptr)
4325 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
4326 else
4327 tp->rx_opt.rcv_tsecr = 0;
4328 return true;
4329 }
4330 return false;
4331 }
4332
4333 /* Fast parse options. This hopes to only see timestamps.
4334 * If it is wrong it falls back on tcp_parse_options().
4335 */
4336 static bool tcp_fast_parse_options(const struct net *net,
4337 const struct sk_buff *skb,
4338 const struct tcphdr *th, struct tcp_sock *tp)
4339 {
4340 /* In the spirit of fast parsing, compare doff directly to constant
4341 * values. Because equality is used, short doff can be ignored here.
4342 */
4343 if (th->doff == (sizeof(*th) / 4)) {
4344 tp->rx_opt.saw_tstamp = 0;
4345 return false;
4346 } else if (tp->rx_opt.tstamp_ok &&
4347 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
4348 if (tcp_parse_aligned_timestamp(tp, th))
4349 return true;
4350 }
4351
4352 tcp_parse_options(net, skb, &tp->rx_opt, 1, NULL);
4353 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
4354 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
4355
4356 return true;
4357 }
4358
4359 #if defined(CONFIG_TCP_MD5SIG) || defined(CONFIG_TCP_AO)
4360 /*
4361 * Parse Signature options
4362 */
4363 int tcp_do_parse_auth_options(const struct tcphdr *th,
4364 const u8 **md5_hash, const u8 **ao_hash)
4365 {
4366 int length = (th->doff << 2) - sizeof(*th);
4367 const u8 *ptr = (const u8 *)(th + 1);
4368 unsigned int minlen = TCPOLEN_MD5SIG;
4369
4370 if (IS_ENABLED(CONFIG_TCP_AO))
4371 minlen = sizeof(struct tcp_ao_hdr) + 1;
4372
4373 *md5_hash = NULL;
4374 *ao_hash = NULL;
4375
4376 /* If not enough data remaining, we can short cut */
4377 while (length >= minlen) {
4378 int opcode = *ptr++;
4379 int opsize;
4380
4381 switch (opcode) {
4382 case TCPOPT_EOL:
4383 return 0;
4384 case TCPOPT_NOP:
4385 length--;
4386 continue;
4387 default:
4388 opsize = *ptr++;
4389 if (opsize < 2 || opsize > length)
4390 return -EINVAL;
4391 if (opcode == TCPOPT_MD5SIG) {
4392 if (opsize != TCPOLEN_MD5SIG)
4393 return -EINVAL;
4394 if (unlikely(*md5_hash || *ao_hash))
4395 return -EEXIST;
4396 *md5_hash = ptr;
4397 } else if (opcode == TCPOPT_AO) {
4398 if (opsize <= sizeof(struct tcp_ao_hdr))
4399 return -EINVAL;
4400 if (unlikely(*md5_hash || *ao_hash))
4401 return -EEXIST;
4402 *ao_hash = ptr;
4403 }
4404 }
4405 ptr += opsize - 2;
4406 length -= opsize;
4407 }
4408 return 0;
4409 }
4410 EXPORT_SYMBOL(tcp_do_parse_auth_options);
4411 #endif
4412
4413 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
4414 *
4415 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
4416 * it can pass through stack. So, the following predicate verifies that
4417 * this segment is not used for anything but congestion avoidance or
4418 * fast retransmit. Moreover, we even are able to eliminate most of such
4419 * second order effects, if we apply some small "replay" window (~RTO)
4420 * to timestamp space.
4421 *
4422 * All these measures still do not guarantee that we reject wrapped ACKs
4423 * on networks with high bandwidth, when sequence space is recycled fastly,
4424 * but it guarantees that such events will be very rare and do not affect
4425 * connection seriously. This doesn't look nice, but alas, PAWS is really
4426 * buggy extension.
4427 *
4428 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
4429 * states that events when retransmit arrives after original data are rare.
4430 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
4431 * the biggest problem on large power networks even with minor reordering.
4432 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
4433 * up to bandwidth of 18Gigabit/sec. 8) ]
4434 */
4435
4436 /* Estimates max number of increments of remote peer TSval in
4437 * a replay window (based on our current RTO estimation).
4438 */
4439 static u32 tcp_tsval_replay(const struct sock *sk)
4440 {
4441 /* If we use usec TS resolution,
4442 * then expect the remote peer to use the same resolution.
4443 */
4444 if (tcp_sk(sk)->tcp_usec_ts)
4445 return inet_csk(sk)->icsk_rto * (USEC_PER_SEC / HZ);
4446
4447 /* RFC 7323 recommends a TSval clock between 1ms and 1sec.
4448 * We know that some OS (including old linux) can use 1200 Hz.
4449 */
4450 return inet_csk(sk)->icsk_rto * 1200 / HZ;
4451 }
4452
4453 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
4454 {
4455 const struct tcp_sock *tp = tcp_sk(sk);
4456 const struct tcphdr *th = tcp_hdr(skb);
4457 u32 seq = TCP_SKB_CB(skb)->seq;
4458 u32 ack = TCP_SKB_CB(skb)->ack_seq;
4459
4460 return /* 1. Pure ACK with correct sequence number. */
4461 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
4462
4463 /* 2. ... and duplicate ACK. */
4464 ack == tp->snd_una &&
4465
4466 /* 3. ... and does not update window. */
4467 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
4468
4469 /* 4. ... and sits in replay window. */
4470 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <=
4471 tcp_tsval_replay(sk);
4472 }
4473
4474 static inline bool tcp_paws_discard(const struct sock *sk,
4475 const struct sk_buff *skb)
4476 {
4477 const struct tcp_sock *tp = tcp_sk(sk);
4478
4479 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4480 !tcp_disordered_ack(sk, skb);
4481 }
4482
4483 /* Check segment sequence number for validity.
4484 *
4485 * Segment controls are considered valid, if the segment
4486 * fits to the window after truncation to the window. Acceptability
4487 * of data (and SYN, FIN, of course) is checked separately.
4488 * See tcp_data_queue(), for example.
4489 *
4490 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4491 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4492 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4493 * (borrowed from freebsd)
4494 */
4495
4496 static enum skb_drop_reason tcp_sequence(const struct tcp_sock *tp,
4497 u32 seq, u32 end_seq)
4498 {
4499 if (before(end_seq, tp->rcv_wup))
4500 return SKB_DROP_REASON_TCP_OLD_SEQUENCE;
4501
4502 if (after(seq, tp->rcv_nxt + tcp_receive_window(tp)))
4503 return SKB_DROP_REASON_TCP_INVALID_SEQUENCE;
4504
4505 return SKB_NOT_DROPPED_YET;
4506 }
4507
4508
4509 void tcp_done_with_error(struct sock *sk, int err)
4510 {
4511 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4512 WRITE_ONCE(sk->sk_err, err);
4513 smp_wmb();
4514
4515 tcp_write_queue_purge(sk);
4516 tcp_done(sk);
4517
4518 if (!sock_flag(sk, SOCK_DEAD))
4519 sk_error_report(sk);
4520 }
4521 EXPORT_SYMBOL(tcp_done_with_error);
4522
4523 /* When we get a reset we do this. */
4524 void tcp_reset(struct sock *sk, struct sk_buff *skb)
4525 {
4526 int err;
4527
4528 trace_tcp_receive_reset(sk);
4529
4530 /* mptcp can't tell us to ignore reset pkts,
4531 * so just ignore the return value of mptcp_incoming_options().
4532 */
4533 if (sk_is_mptcp(sk))
4534 mptcp_incoming_options(sk, skb);
4535
4536 /* We want the right error as BSD sees it (and indeed as we do). */
4537 switch (sk->sk_state) {
4538 case TCP_SYN_SENT:
4539 err = ECONNREFUSED;
4540 break;
4541 case TCP_CLOSE_WAIT:
4542 err = EPIPE;
4543 break;
4544 case TCP_CLOSE:
4545 return;
4546 default:
4547 err = ECONNRESET;
4548 }
4549 tcp_done_with_error(sk, err);
4550 }
4551
4552 /*
4553 * Process the FIN bit. This now behaves as it is supposed to work
4554 * and the FIN takes effect when it is validly part of sequence
4555 * space. Not before when we get holes.
4556 *
4557 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4558 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4559 * TIME-WAIT)
4560 *
4561 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4562 * close and we go into CLOSING (and later onto TIME-WAIT)
4563 *
4564 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4565 */
4566 void tcp_fin(struct sock *sk)
4567 {
4568 struct tcp_sock *tp = tcp_sk(sk);
4569
4570 inet_csk_schedule_ack(sk);
4571
4572 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | RCV_SHUTDOWN);
4573 sock_set_flag(sk, SOCK_DONE);
4574
4575 switch (sk->sk_state) {
4576 case TCP_SYN_RECV:
4577 case TCP_ESTABLISHED:
4578 /* Move to CLOSE_WAIT */
4579 tcp_set_state(sk, TCP_CLOSE_WAIT);
4580 inet_csk_enter_pingpong_mode(sk);
4581 break;
4582
4583 case TCP_CLOSE_WAIT:
4584 case TCP_CLOSING:
4585 /* Received a retransmission of the FIN, do
4586 * nothing.
4587 */
4588 break;
4589 case TCP_LAST_ACK:
4590 /* RFC793: Remain in the LAST-ACK state. */
4591 break;
4592
4593 case TCP_FIN_WAIT1:
4594 /* This case occurs when a simultaneous close
4595 * happens, we must ack the received FIN and
4596 * enter the CLOSING state.
4597 */
4598 tcp_send_ack(sk);
4599 tcp_set_state(sk, TCP_CLOSING);
4600 break;
4601 case TCP_FIN_WAIT2:
4602 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4603 tcp_send_ack(sk);
4604 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4605 break;
4606 default:
4607 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4608 * cases we should never reach this piece of code.
4609 */
4610 pr_err("%s: Impossible, sk->sk_state=%d\n",
4611 __func__, sk->sk_state);
4612 break;
4613 }
4614
4615 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4616 * Probably, we should reset in this case. For now drop them.
4617 */
4618 skb_rbtree_purge(&tp->out_of_order_queue);
4619 if (tcp_is_sack(tp))
4620 tcp_sack_reset(&tp->rx_opt);
4621
4622 if (!sock_flag(sk, SOCK_DEAD)) {
4623 sk->sk_state_change(sk);
4624
4625 /* Do not send POLL_HUP for half duplex close. */
4626 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4627 sk->sk_state == TCP_CLOSE)
4628 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4629 else
4630 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4631 }
4632 }
4633
4634 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4635 u32 end_seq)
4636 {
4637 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4638 if (before(seq, sp->start_seq))
4639 sp->start_seq = seq;
4640 if (after(end_seq, sp->end_seq))
4641 sp->end_seq = end_seq;
4642 return true;
4643 }
4644 return false;
4645 }
4646
4647 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4648 {
4649 struct tcp_sock *tp = tcp_sk(sk);
4650
4651 if (tcp_is_sack(tp) && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_dsack)) {
4652 int mib_idx;
4653
4654 if (before(seq, tp->rcv_nxt))
4655 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4656 else
4657 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4658
4659 NET_INC_STATS(sock_net(sk), mib_idx);
4660
4661 tp->rx_opt.dsack = 1;
4662 tp->duplicate_sack[0].start_seq = seq;
4663 tp->duplicate_sack[0].end_seq = end_seq;
4664 }
4665 }
4666
4667 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4668 {
4669 struct tcp_sock *tp = tcp_sk(sk);
4670
4671 if (!tp->rx_opt.dsack)
4672 tcp_dsack_set(sk, seq, end_seq);
4673 else
4674 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4675 }
4676
4677 static void tcp_rcv_spurious_retrans(struct sock *sk, const struct sk_buff *skb)
4678 {
4679 /* When the ACK path fails or drops most ACKs, the sender would
4680 * timeout and spuriously retransmit the same segment repeatedly.
4681 * If it seems our ACKs are not reaching the other side,
4682 * based on receiving a duplicate data segment with new flowlabel
4683 * (suggesting the sender suffered an RTO), and we are not already
4684 * repathing due to our own RTO, then rehash the socket to repath our
4685 * packets.
4686 */
4687 #if IS_ENABLED(CONFIG_IPV6)
4688 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss &&
4689 skb->protocol == htons(ETH_P_IPV6) &&
4690 (tcp_sk(sk)->inet_conn.icsk_ack.lrcv_flowlabel !=
4691 ntohl(ip6_flowlabel(ipv6_hdr(skb)))) &&
4692 sk_rethink_txhash(sk))
4693 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDUPLICATEDATAREHASH);
4694
4695 /* Save last flowlabel after a spurious retrans. */
4696 tcp_save_lrcv_flowlabel(sk, skb);
4697 #endif
4698 }
4699
4700 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4701 {
4702 struct tcp_sock *tp = tcp_sk(sk);
4703
4704 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4705 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4706 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4707 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
4708
4709 if (tcp_is_sack(tp) && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_dsack)) {
4710 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4711
4712 tcp_rcv_spurious_retrans(sk, skb);
4713 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4714 end_seq = tp->rcv_nxt;
4715 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4716 }
4717 }
4718
4719 tcp_send_ack(sk);
4720 }
4721
4722 /* These routines update the SACK block as out-of-order packets arrive or
4723 * in-order packets close up the sequence space.
4724 */
4725 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4726 {
4727 int this_sack;
4728 struct tcp_sack_block *sp = &tp->selective_acks[0];
4729 struct tcp_sack_block *swalk = sp + 1;
4730
4731 /* See if the recent change to the first SACK eats into
4732 * or hits the sequence space of other SACK blocks, if so coalesce.
4733 */
4734 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4735 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4736 int i;
4737
4738 /* Zap SWALK, by moving every further SACK up by one slot.
4739 * Decrease num_sacks.
4740 */
4741 tp->rx_opt.num_sacks--;
4742 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4743 sp[i] = sp[i + 1];
4744 continue;
4745 }
4746 this_sack++;
4747 swalk++;
4748 }
4749 }
4750
4751 void tcp_sack_compress_send_ack(struct sock *sk)
4752 {
4753 struct tcp_sock *tp = tcp_sk(sk);
4754
4755 if (!tp->compressed_ack)
4756 return;
4757
4758 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
4759 __sock_put(sk);
4760
4761 /* Since we have to send one ack finally,
4762 * substract one from tp->compressed_ack to keep
4763 * LINUX_MIB_TCPACKCOMPRESSED accurate.
4764 */
4765 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
4766 tp->compressed_ack - 1);
4767
4768 tp->compressed_ack = 0;
4769 tcp_send_ack(sk);
4770 }
4771
4772 /* Reasonable amount of sack blocks included in TCP SACK option
4773 * The max is 4, but this becomes 3 if TCP timestamps are there.
4774 * Given that SACK packets might be lost, be conservative and use 2.
4775 */
4776 #define TCP_SACK_BLOCKS_EXPECTED 2
4777
4778 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4779 {
4780 struct tcp_sock *tp = tcp_sk(sk);
4781 struct tcp_sack_block *sp = &tp->selective_acks[0];
4782 int cur_sacks = tp->rx_opt.num_sacks;
4783 int this_sack;
4784
4785 if (!cur_sacks)
4786 goto new_sack;
4787
4788 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4789 if (tcp_sack_extend(sp, seq, end_seq)) {
4790 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED)
4791 tcp_sack_compress_send_ack(sk);
4792 /* Rotate this_sack to the first one. */
4793 for (; this_sack > 0; this_sack--, sp--)
4794 swap(*sp, *(sp - 1));
4795 if (cur_sacks > 1)
4796 tcp_sack_maybe_coalesce(tp);
4797 return;
4798 }
4799 }
4800
4801 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED)
4802 tcp_sack_compress_send_ack(sk);
4803
4804 /* Could not find an adjacent existing SACK, build a new one,
4805 * put it at the front, and shift everyone else down. We
4806 * always know there is at least one SACK present already here.
4807 *
4808 * If the sack array is full, forget about the last one.
4809 */
4810 if (this_sack >= TCP_NUM_SACKS) {
4811 this_sack--;
4812 tp->rx_opt.num_sacks--;
4813 sp--;
4814 }
4815 for (; this_sack > 0; this_sack--, sp--)
4816 *sp = *(sp - 1);
4817
4818 new_sack:
4819 /* Build the new head SACK, and we're done. */
4820 sp->start_seq = seq;
4821 sp->end_seq = end_seq;
4822 tp->rx_opt.num_sacks++;
4823 }
4824
4825 /* RCV.NXT advances, some SACKs should be eaten. */
4826
4827 static void tcp_sack_remove(struct tcp_sock *tp)
4828 {
4829 struct tcp_sack_block *sp = &tp->selective_acks[0];
4830 int num_sacks = tp->rx_opt.num_sacks;
4831 int this_sack;
4832
4833 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4834 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4835 tp->rx_opt.num_sacks = 0;
4836 return;
4837 }
4838
4839 for (this_sack = 0; this_sack < num_sacks;) {
4840 /* Check if the start of the sack is covered by RCV.NXT. */
4841 if (!before(tp->rcv_nxt, sp->start_seq)) {
4842 int i;
4843
4844 /* RCV.NXT must cover all the block! */
4845 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4846
4847 /* Zap this SACK, by moving forward any other SACKS. */
4848 for (i = this_sack+1; i < num_sacks; i++)
4849 tp->selective_acks[i-1] = tp->selective_acks[i];
4850 num_sacks--;
4851 continue;
4852 }
4853 this_sack++;
4854 sp++;
4855 }
4856 tp->rx_opt.num_sacks = num_sacks;
4857 }
4858
4859 /**
4860 * tcp_try_coalesce - try to merge skb to prior one
4861 * @sk: socket
4862 * @to: prior buffer
4863 * @from: buffer to add in queue
4864 * @fragstolen: pointer to boolean
4865 *
4866 * Before queueing skb @from after @to, try to merge them
4867 * to reduce overall memory use and queue lengths, if cost is small.
4868 * Packets in ofo or receive queues can stay a long time.
4869 * Better try to coalesce them right now to avoid future collapses.
4870 * Returns true if caller should free @from instead of queueing it
4871 */
4872 static bool tcp_try_coalesce(struct sock *sk,
4873 struct sk_buff *to,
4874 struct sk_buff *from,
4875 bool *fragstolen)
4876 {
4877 int delta;
4878
4879 *fragstolen = false;
4880
4881 /* Its possible this segment overlaps with prior segment in queue */
4882 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4883 return false;
4884
4885 if (!tcp_skb_can_collapse_rx(to, from))
4886 return false;
4887
4888 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4889 return false;
4890
4891 atomic_add(delta, &sk->sk_rmem_alloc);
4892 sk_mem_charge(sk, delta);
4893 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4894 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4895 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4896 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4897
4898 if (TCP_SKB_CB(from)->has_rxtstamp) {
4899 TCP_SKB_CB(to)->has_rxtstamp = true;
4900 to->tstamp = from->tstamp;
4901 skb_hwtstamps(to)->hwtstamp = skb_hwtstamps(from)->hwtstamp;
4902 }
4903
4904 return true;
4905 }
4906
4907 static bool tcp_ooo_try_coalesce(struct sock *sk,
4908 struct sk_buff *to,
4909 struct sk_buff *from,
4910 bool *fragstolen)
4911 {
4912 bool res = tcp_try_coalesce(sk, to, from, fragstolen);
4913
4914 /* In case tcp_drop_reason() is called later, update to->gso_segs */
4915 if (res) {
4916 u32 gso_segs = max_t(u16, 1, skb_shinfo(to)->gso_segs) +
4917 max_t(u16, 1, skb_shinfo(from)->gso_segs);
4918
4919 skb_shinfo(to)->gso_segs = min_t(u32, gso_segs, 0xFFFF);
4920 }
4921 return res;
4922 }
4923
4924 static void tcp_drop_reason(struct sock *sk, struct sk_buff *skb,
4925 enum skb_drop_reason reason)
4926 {
4927 sk_drops_add(sk, skb);
4928 sk_skb_reason_drop(sk, skb, reason);
4929 }
4930
4931 /* This one checks to see if we can put data from the
4932 * out_of_order queue into the receive_queue.
4933 */
4934 static void tcp_ofo_queue(struct sock *sk)
4935 {
4936 struct tcp_sock *tp = tcp_sk(sk);
4937 __u32 dsack_high = tp->rcv_nxt;
4938 bool fin, fragstolen, eaten;
4939 struct sk_buff *skb, *tail;
4940 struct rb_node *p;
4941
4942 p = rb_first(&tp->out_of_order_queue);
4943 while (p) {
4944 skb = rb_to_skb(p);
4945 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4946 break;
4947
4948 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4949 __u32 dsack = dsack_high;
4950 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4951 dsack_high = TCP_SKB_CB(skb)->end_seq;
4952 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4953 }
4954 p = rb_next(p);
4955 rb_erase(&skb->rbnode, &tp->out_of_order_queue);
4956
4957 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) {
4958 tcp_drop_reason(sk, skb, SKB_DROP_REASON_TCP_OFO_DROP);
4959 continue;
4960 }
4961
4962 tail = skb_peek_tail(&sk->sk_receive_queue);
4963 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4964 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4965 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
4966 if (!eaten)
4967 __skb_queue_tail(&sk->sk_receive_queue, skb);
4968 else
4969 kfree_skb_partial(skb, fragstolen);
4970
4971 if (unlikely(fin)) {
4972 tcp_fin(sk);
4973 /* tcp_fin() purges tp->out_of_order_queue,
4974 * so we must end this loop right now.
4975 */
4976 break;
4977 }
4978 }
4979 }
4980
4981 static bool tcp_prune_ofo_queue(struct sock *sk, const struct sk_buff *in_skb);
4982 static int tcp_prune_queue(struct sock *sk, const struct sk_buff *in_skb);
4983
4984 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4985 unsigned int size)
4986 {
4987 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4988 !sk_rmem_schedule(sk, skb, size)) {
4989
4990 if (tcp_prune_queue(sk, skb) < 0)
4991 return -1;
4992
4993 while (!sk_rmem_schedule(sk, skb, size)) {
4994 if (!tcp_prune_ofo_queue(sk, skb))
4995 return -1;
4996 }
4997 }
4998 return 0;
4999 }
5000
5001 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
5002 {
5003 struct tcp_sock *tp = tcp_sk(sk);
5004 struct rb_node **p, *parent;
5005 struct sk_buff *skb1;
5006 u32 seq, end_seq;
5007 bool fragstolen;
5008
5009 tcp_save_lrcv_flowlabel(sk, skb);
5010 tcp_ecn_check_ce(sk, skb);
5011
5012 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
5013 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP);
5014 sk->sk_data_ready(sk);
5015 tcp_drop_reason(sk, skb, SKB_DROP_REASON_PROTO_MEM);
5016 return;
5017 }
5018
5019 /* Disable header prediction. */
5020 tp->pred_flags = 0;
5021 inet_csk_schedule_ack(sk);
5022
5023 tp->rcv_ooopack += max_t(u16, 1, skb_shinfo(skb)->gso_segs);
5024 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
5025 seq = TCP_SKB_CB(skb)->seq;
5026 end_seq = TCP_SKB_CB(skb)->end_seq;
5027
5028 p = &tp->out_of_order_queue.rb_node;
5029 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5030 /* Initial out of order segment, build 1 SACK. */
5031 if (tcp_is_sack(tp)) {
5032 tp->rx_opt.num_sacks = 1;
5033 tp->selective_acks[0].start_seq = seq;
5034 tp->selective_acks[0].end_seq = end_seq;
5035 }
5036 rb_link_node(&skb->rbnode, NULL, p);
5037 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
5038 tp->ooo_last_skb = skb;
5039 goto end;
5040 }
5041
5042 /* In the typical case, we are adding an skb to the end of the list.
5043 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
5044 */
5045 if (tcp_ooo_try_coalesce(sk, tp->ooo_last_skb,
5046 skb, &fragstolen)) {
5047 coalesce_done:
5048 /* For non sack flows, do not grow window to force DUPACK
5049 * and trigger fast retransmit.
5050 */
5051 if (tcp_is_sack(tp))
5052 tcp_grow_window(sk, skb, true);
5053 kfree_skb_partial(skb, fragstolen);
5054 skb = NULL;
5055 goto add_sack;
5056 }
5057 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
5058 if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) {
5059 parent = &tp->ooo_last_skb->rbnode;
5060 p = &parent->rb_right;
5061 goto insert;
5062 }
5063
5064 /* Find place to insert this segment. Handle overlaps on the way. */
5065 parent = NULL;
5066 while (*p) {
5067 parent = *p;
5068 skb1 = rb_to_skb(parent);
5069 if (before(seq, TCP_SKB_CB(skb1)->seq)) {
5070 p = &parent->rb_left;
5071 continue;
5072 }
5073 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) {
5074 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
5075 /* All the bits are present. Drop. */
5076 NET_INC_STATS(sock_net(sk),
5077 LINUX_MIB_TCPOFOMERGE);
5078 tcp_drop_reason(sk, skb,
5079 SKB_DROP_REASON_TCP_OFOMERGE);
5080 skb = NULL;
5081 tcp_dsack_set(sk, seq, end_seq);
5082 goto add_sack;
5083 }
5084 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
5085 /* Partial overlap. */
5086 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq);
5087 } else {
5088 /* skb's seq == skb1's seq and skb covers skb1.
5089 * Replace skb1 with skb.
5090 */
5091 rb_replace_node(&skb1->rbnode, &skb->rbnode,
5092 &tp->out_of_order_queue);
5093 tcp_dsack_extend(sk,
5094 TCP_SKB_CB(skb1)->seq,
5095 TCP_SKB_CB(skb1)->end_seq);
5096 NET_INC_STATS(sock_net(sk),
5097 LINUX_MIB_TCPOFOMERGE);
5098 tcp_drop_reason(sk, skb1,
5099 SKB_DROP_REASON_TCP_OFOMERGE);
5100 goto merge_right;
5101 }
5102 } else if (tcp_ooo_try_coalesce(sk, skb1,
5103 skb, &fragstolen)) {
5104 goto coalesce_done;
5105 }
5106 p = &parent->rb_right;
5107 }
5108 insert:
5109 /* Insert segment into RB tree. */
5110 rb_link_node(&skb->rbnode, parent, p);
5111 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
5112
5113 merge_right:
5114 /* Remove other segments covered by skb. */
5115 while ((skb1 = skb_rb_next(skb)) != NULL) {
5116 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
5117 break;
5118 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
5119 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
5120 end_seq);
5121 break;
5122 }
5123 rb_erase(&skb1->rbnode, &tp->out_of_order_queue);
5124 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
5125 TCP_SKB_CB(skb1)->end_seq);
5126 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
5127 tcp_drop_reason(sk, skb1, SKB_DROP_REASON_TCP_OFOMERGE);
5128 }
5129 /* If there is no skb after us, we are the last_skb ! */
5130 if (!skb1)
5131 tp->ooo_last_skb = skb;
5132
5133 add_sack:
5134 if (tcp_is_sack(tp))
5135 tcp_sack_new_ofo_skb(sk, seq, end_seq);
5136 end:
5137 if (skb) {
5138 /* For non sack flows, do not grow window to force DUPACK
5139 * and trigger fast retransmit.
5140 */
5141 if (tcp_is_sack(tp))
5142 tcp_grow_window(sk, skb, false);
5143 skb_condense(skb);
5144 skb_set_owner_r(skb, sk);
5145 }
5146 }
5147
5148 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb,
5149 bool *fragstolen)
5150 {
5151 int eaten;
5152 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
5153
5154 eaten = (tail &&
5155 tcp_try_coalesce(sk, tail,
5156 skb, fragstolen)) ? 1 : 0;
5157 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
5158 if (!eaten) {
5159 __skb_queue_tail(&sk->sk_receive_queue, skb);
5160 skb_set_owner_r(skb, sk);
5161 }
5162 return eaten;
5163 }
5164
5165 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
5166 {
5167 struct sk_buff *skb;
5168 int err = -ENOMEM;
5169 int data_len = 0;
5170 bool fragstolen;
5171
5172 if (size == 0)
5173 return 0;
5174
5175 if (size > PAGE_SIZE) {
5176 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
5177
5178 data_len = npages << PAGE_SHIFT;
5179 size = data_len + (size & ~PAGE_MASK);
5180 }
5181 skb = alloc_skb_with_frags(size - data_len, data_len,
5182 PAGE_ALLOC_COSTLY_ORDER,
5183 &err, sk->sk_allocation);
5184 if (!skb)
5185 goto err;
5186
5187 skb_put(skb, size - data_len);
5188 skb->data_len = data_len;
5189 skb->len = size;
5190
5191 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
5192 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
5193 goto err_free;
5194 }
5195
5196 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
5197 if (err)
5198 goto err_free;
5199
5200 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
5201 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
5202 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
5203
5204 if (tcp_queue_rcv(sk, skb, &fragstolen)) {
5205 WARN_ON_ONCE(fragstolen); /* should not happen */
5206 __kfree_skb(skb);
5207 }
5208 return size;
5209
5210 err_free:
5211 kfree_skb(skb);
5212 err:
5213 return err;
5214
5215 }
5216
5217 void tcp_data_ready(struct sock *sk)
5218 {
5219 if (tcp_epollin_ready(sk, sk->sk_rcvlowat) || sock_flag(sk, SOCK_DONE))
5220 sk->sk_data_ready(sk);
5221 }
5222
5223 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
5224 {
5225 struct tcp_sock *tp = tcp_sk(sk);
5226 enum skb_drop_reason reason;
5227 bool fragstolen;
5228 int eaten;
5229
5230 /* If a subflow has been reset, the packet should not continue
5231 * to be processed, drop the packet.
5232 */
5233 if (sk_is_mptcp(sk) && !mptcp_incoming_options(sk, skb)) {
5234 __kfree_skb(skb);
5235 return;
5236 }
5237
5238 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
5239 __kfree_skb(skb);
5240 return;
5241 }
5242 skb_dst_drop(skb);
5243 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
5244
5245 reason = SKB_DROP_REASON_NOT_SPECIFIED;
5246 tp->rx_opt.dsack = 0;
5247
5248 /* Queue data for delivery to the user.
5249 * Packets in sequence go to the receive queue.
5250 * Out of sequence packets to the out_of_order_queue.
5251 */
5252 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
5253 if (tcp_receive_window(tp) == 0) {
5254 /* Some stacks are known to send bare FIN packets
5255 * in a loop even if we send RWIN 0 in our ACK.
5256 * Accepting this FIN does not hurt memory pressure
5257 * because the FIN flag will simply be merged to the
5258 * receive queue tail skb in most cases.
5259 */
5260 if (!skb->len &&
5261 (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
5262 goto queue_and_out;
5263
5264 reason = SKB_DROP_REASON_TCP_ZEROWINDOW;
5265 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
5266 goto out_of_window;
5267 }
5268
5269 /* Ok. In sequence. In window. */
5270 queue_and_out:
5271 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
5272 /* TODO: maybe ratelimit these WIN 0 ACK ? */
5273 inet_csk(sk)->icsk_ack.pending |=
5274 (ICSK_ACK_NOMEM | ICSK_ACK_NOW);
5275 inet_csk_schedule_ack(sk);
5276 sk->sk_data_ready(sk);
5277
5278 if (skb_queue_len(&sk->sk_receive_queue) && skb->len) {
5279 reason = SKB_DROP_REASON_PROTO_MEM;
5280 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
5281 goto drop;
5282 }
5283 sk_forced_mem_schedule(sk, skb->truesize);
5284 }
5285
5286 eaten = tcp_queue_rcv(sk, skb, &fragstolen);
5287 if (skb->len)
5288 tcp_event_data_recv(sk, skb);
5289 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
5290 tcp_fin(sk);
5291
5292 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5293 tcp_ofo_queue(sk);
5294
5295 /* RFC5681. 4.2. SHOULD send immediate ACK, when
5296 * gap in queue is filled.
5297 */
5298 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
5299 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
5300 }
5301
5302 if (tp->rx_opt.num_sacks)
5303 tcp_sack_remove(tp);
5304
5305 tcp_fast_path_check(sk);
5306
5307 if (eaten > 0)
5308 kfree_skb_partial(skb, fragstolen);
5309 if (!sock_flag(sk, SOCK_DEAD))
5310 tcp_data_ready(sk);
5311 return;
5312 }
5313
5314 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
5315 tcp_rcv_spurious_retrans(sk, skb);
5316 /* A retransmit, 2nd most common case. Force an immediate ack. */
5317 reason = SKB_DROP_REASON_TCP_OLD_DATA;
5318 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
5319 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
5320
5321 out_of_window:
5322 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
5323 inet_csk_schedule_ack(sk);
5324 drop:
5325 tcp_drop_reason(sk, skb, reason);
5326 return;
5327 }
5328
5329 /* Out of window. F.e. zero window probe. */
5330 if (!before(TCP_SKB_CB(skb)->seq,
5331 tp->rcv_nxt + tcp_receive_window(tp))) {
5332 reason = SKB_DROP_REASON_TCP_OVERWINDOW;
5333 goto out_of_window;
5334 }
5335
5336 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
5337 /* Partial packet, seq < rcv_next < end_seq */
5338 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
5339
5340 /* If window is closed, drop tail of packet. But after
5341 * remembering D-SACK for its head made in previous line.
5342 */
5343 if (!tcp_receive_window(tp)) {
5344 reason = SKB_DROP_REASON_TCP_ZEROWINDOW;
5345 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
5346 goto out_of_window;
5347 }
5348 goto queue_and_out;
5349 }
5350
5351 tcp_data_queue_ofo(sk, skb);
5352 }
5353
5354 static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list)
5355 {
5356 if (list)
5357 return !skb_queue_is_last(list, skb) ? skb->next : NULL;
5358
5359 return skb_rb_next(skb);
5360 }
5361
5362 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
5363 struct sk_buff_head *list,
5364 struct rb_root *root)
5365 {
5366 struct sk_buff *next = tcp_skb_next(skb, list);
5367
5368 if (list)
5369 __skb_unlink(skb, list);
5370 else
5371 rb_erase(&skb->rbnode, root);
5372
5373 __kfree_skb(skb);
5374 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
5375
5376 return next;
5377 }
5378
5379 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
5380 void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb)
5381 {
5382 struct rb_node **p = &root->rb_node;
5383 struct rb_node *parent = NULL;
5384 struct sk_buff *skb1;
5385
5386 while (*p) {
5387 parent = *p;
5388 skb1 = rb_to_skb(parent);
5389 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq))
5390 p = &parent->rb_left;
5391 else
5392 p = &parent->rb_right;
5393 }
5394 rb_link_node(&skb->rbnode, parent, p);
5395 rb_insert_color(&skb->rbnode, root);
5396 }
5397
5398 /* Collapse contiguous sequence of skbs head..tail with
5399 * sequence numbers start..end.
5400 *
5401 * If tail is NULL, this means until the end of the queue.
5402 *
5403 * Segments with FIN/SYN are not collapsed (only because this
5404 * simplifies code)
5405 */
5406 static void
5407 tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root,
5408 struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end)
5409 {
5410 struct sk_buff *skb = head, *n;
5411 struct sk_buff_head tmp;
5412 bool end_of_skbs;
5413
5414 /* First, check that queue is collapsible and find
5415 * the point where collapsing can be useful.
5416 */
5417 restart:
5418 for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) {
5419 n = tcp_skb_next(skb, list);
5420
5421 /* No new bits? It is possible on ofo queue. */
5422 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
5423 skb = tcp_collapse_one(sk, skb, list, root);
5424 if (!skb)
5425 break;
5426 goto restart;
5427 }
5428
5429 /* The first skb to collapse is:
5430 * - not SYN/FIN and
5431 * - bloated or contains data before "start" or
5432 * overlaps to the next one and mptcp allow collapsing.
5433 */
5434 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
5435 (tcp_win_from_space(sk, skb->truesize) > skb->len ||
5436 before(TCP_SKB_CB(skb)->seq, start))) {
5437 end_of_skbs = false;
5438 break;
5439 }
5440
5441 if (n && n != tail && tcp_skb_can_collapse_rx(skb, n) &&
5442 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) {
5443 end_of_skbs = false;
5444 break;
5445 }
5446
5447 /* Decided to skip this, advance start seq. */
5448 start = TCP_SKB_CB(skb)->end_seq;
5449 }
5450 if (end_of_skbs ||
5451 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
5452 return;
5453
5454 __skb_queue_head_init(&tmp);
5455
5456 while (before(start, end)) {
5457 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
5458 struct sk_buff *nskb;
5459
5460 nskb = alloc_skb(copy, GFP_ATOMIC);
5461 if (!nskb)
5462 break;
5463
5464 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
5465 skb_copy_decrypted(nskb, skb);
5466 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
5467 if (list)
5468 __skb_queue_before(list, skb, nskb);
5469 else
5470 __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */
5471 skb_set_owner_r(nskb, sk);
5472 mptcp_skb_ext_move(nskb, skb);
5473
5474 /* Copy data, releasing collapsed skbs. */
5475 while (copy > 0) {
5476 int offset = start - TCP_SKB_CB(skb)->seq;
5477 int size = TCP_SKB_CB(skb)->end_seq - start;
5478
5479 BUG_ON(offset < 0);
5480 if (size > 0) {
5481 size = min(copy, size);
5482 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
5483 BUG();
5484 TCP_SKB_CB(nskb)->end_seq += size;
5485 copy -= size;
5486 start += size;
5487 }
5488 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
5489 skb = tcp_collapse_one(sk, skb, list, root);
5490 if (!skb ||
5491 skb == tail ||
5492 !tcp_skb_can_collapse_rx(nskb, skb) ||
5493 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
5494 goto end;
5495 }
5496 }
5497 }
5498 end:
5499 skb_queue_walk_safe(&tmp, skb, n)
5500 tcp_rbtree_insert(root, skb);
5501 }
5502
5503 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
5504 * and tcp_collapse() them until all the queue is collapsed.
5505 */
5506 static void tcp_collapse_ofo_queue(struct sock *sk)
5507 {
5508 struct tcp_sock *tp = tcp_sk(sk);
5509 u32 range_truesize, sum_tiny = 0;
5510 struct sk_buff *skb, *head;
5511 u32 start, end;
5512
5513 skb = skb_rb_first(&tp->out_of_order_queue);
5514 new_range:
5515 if (!skb) {
5516 tp->ooo_last_skb = skb_rb_last(&tp->out_of_order_queue);
5517 return;
5518 }
5519 start = TCP_SKB_CB(skb)->seq;
5520 end = TCP_SKB_CB(skb)->end_seq;
5521 range_truesize = skb->truesize;
5522
5523 for (head = skb;;) {
5524 skb = skb_rb_next(skb);
5525
5526 /* Range is terminated when we see a gap or when
5527 * we are at the queue end.
5528 */
5529 if (!skb ||
5530 after(TCP_SKB_CB(skb)->seq, end) ||
5531 before(TCP_SKB_CB(skb)->end_seq, start)) {
5532 /* Do not attempt collapsing tiny skbs */
5533 if (range_truesize != head->truesize ||
5534 end - start >= SKB_WITH_OVERHEAD(PAGE_SIZE)) {
5535 tcp_collapse(sk, NULL, &tp->out_of_order_queue,
5536 head, skb, start, end);
5537 } else {
5538 sum_tiny += range_truesize;
5539 if (sum_tiny > sk->sk_rcvbuf >> 3)
5540 return;
5541 }
5542 goto new_range;
5543 }
5544
5545 range_truesize += skb->truesize;
5546 if (unlikely(before(TCP_SKB_CB(skb)->seq, start)))
5547 start = TCP_SKB_CB(skb)->seq;
5548 if (after(TCP_SKB_CB(skb)->end_seq, end))
5549 end = TCP_SKB_CB(skb)->end_seq;
5550 }
5551 }
5552
5553 /*
5554 * Clean the out-of-order queue to make room.
5555 * We drop high sequences packets to :
5556 * 1) Let a chance for holes to be filled.
5557 * This means we do not drop packets from ooo queue if their sequence
5558 * is before incoming packet sequence.
5559 * 2) not add too big latencies if thousands of packets sit there.
5560 * (But if application shrinks SO_RCVBUF, we could still end up
5561 * freeing whole queue here)
5562 * 3) Drop at least 12.5 % of sk_rcvbuf to avoid malicious attacks.
5563 *
5564 * Return true if queue has shrunk.
5565 */
5566 static bool tcp_prune_ofo_queue(struct sock *sk, const struct sk_buff *in_skb)
5567 {
5568 struct tcp_sock *tp = tcp_sk(sk);
5569 struct rb_node *node, *prev;
5570 bool pruned = false;
5571 int goal;
5572
5573 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
5574 return false;
5575
5576 goal = sk->sk_rcvbuf >> 3;
5577 node = &tp->ooo_last_skb->rbnode;
5578
5579 do {
5580 struct sk_buff *skb = rb_to_skb(node);
5581
5582 /* If incoming skb would land last in ofo queue, stop pruning. */
5583 if (after(TCP_SKB_CB(in_skb)->seq, TCP_SKB_CB(skb)->seq))
5584 break;
5585 pruned = true;
5586 prev = rb_prev(node);
5587 rb_erase(node, &tp->out_of_order_queue);
5588 goal -= skb->truesize;
5589 tcp_drop_reason(sk, skb, SKB_DROP_REASON_TCP_OFO_QUEUE_PRUNE);
5590 tp->ooo_last_skb = rb_to_skb(prev);
5591 if (!prev || goal <= 0) {
5592 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf &&
5593 !tcp_under_memory_pressure(sk))
5594 break;
5595 goal = sk->sk_rcvbuf >> 3;
5596 }
5597 node = prev;
5598 } while (node);
5599
5600 if (pruned) {
5601 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED);
5602 /* Reset SACK state. A conforming SACK implementation will
5603 * do the same at a timeout based retransmit. When a connection
5604 * is in a sad state like this, we care only about integrity
5605 * of the connection not performance.
5606 */
5607 if (tp->rx_opt.sack_ok)
5608 tcp_sack_reset(&tp->rx_opt);
5609 }
5610 return pruned;
5611 }
5612
5613 /* Reduce allocated memory if we can, trying to get
5614 * the socket within its memory limits again.
5615 *
5616 * Return less than zero if we should start dropping frames
5617 * until the socket owning process reads some of the data
5618 * to stabilize the situation.
5619 */
5620 static int tcp_prune_queue(struct sock *sk, const struct sk_buff *in_skb)
5621 {
5622 struct tcp_sock *tp = tcp_sk(sk);
5623
5624 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED);
5625
5626 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
5627 tcp_clamp_window(sk);
5628 else if (tcp_under_memory_pressure(sk))
5629 tcp_adjust_rcv_ssthresh(sk);
5630
5631 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5632 return 0;
5633
5634 tcp_collapse_ofo_queue(sk);
5635 if (!skb_queue_empty(&sk->sk_receive_queue))
5636 tcp_collapse(sk, &sk->sk_receive_queue, NULL,
5637 skb_peek(&sk->sk_receive_queue),
5638 NULL,
5639 tp->copied_seq, tp->rcv_nxt);
5640
5641 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5642 return 0;
5643
5644 /* Collapsing did not help, destructive actions follow.
5645 * This must not ever occur. */
5646
5647 tcp_prune_ofo_queue(sk, in_skb);
5648
5649 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5650 return 0;
5651
5652 /* If we are really being abused, tell the caller to silently
5653 * drop receive data on the floor. It will get retransmitted
5654 * and hopefully then we'll have sufficient space.
5655 */
5656 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED);
5657
5658 /* Massive buffer overcommit. */
5659 tp->pred_flags = 0;
5660 return -1;
5661 }
5662
5663 static bool tcp_should_expand_sndbuf(struct sock *sk)
5664 {
5665 const struct tcp_sock *tp = tcp_sk(sk);
5666
5667 /* If the user specified a specific send buffer setting, do
5668 * not modify it.
5669 */
5670 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
5671 return false;
5672
5673 /* If we are under global TCP memory pressure, do not expand. */
5674 if (tcp_under_memory_pressure(sk)) {
5675 int unused_mem = sk_unused_reserved_mem(sk);
5676
5677 /* Adjust sndbuf according to reserved mem. But make sure
5678 * it never goes below SOCK_MIN_SNDBUF.
5679 * See sk_stream_moderate_sndbuf() for more details.
5680 */
5681 if (unused_mem > SOCK_MIN_SNDBUF)
5682 WRITE_ONCE(sk->sk_sndbuf, unused_mem);
5683
5684 return false;
5685 }
5686
5687 /* If we are under soft global TCP memory pressure, do not expand. */
5688 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
5689 return false;
5690
5691 /* If we filled the congestion window, do not expand. */
5692 if (tcp_packets_in_flight(tp) >= tcp_snd_cwnd(tp))
5693 return false;
5694
5695 return true;
5696 }
5697
5698 static void tcp_new_space(struct sock *sk)
5699 {
5700 struct tcp_sock *tp = tcp_sk(sk);
5701
5702 if (tcp_should_expand_sndbuf(sk)) {
5703 tcp_sndbuf_expand(sk);
5704 tp->snd_cwnd_stamp = tcp_jiffies32;
5705 }
5706
5707 INDIRECT_CALL_1(sk->sk_write_space, sk_stream_write_space, sk);
5708 }
5709
5710 /* Caller made space either from:
5711 * 1) Freeing skbs in rtx queues (after tp->snd_una has advanced)
5712 * 2) Sent skbs from output queue (and thus advancing tp->snd_nxt)
5713 *
5714 * We might be able to generate EPOLLOUT to the application if:
5715 * 1) Space consumed in output/rtx queues is below sk->sk_sndbuf/2
5716 * 2) notsent amount (tp->write_seq - tp->snd_nxt) became
5717 * small enough that tcp_stream_memory_free() decides it
5718 * is time to generate EPOLLOUT.
5719 */
5720 void tcp_check_space(struct sock *sk)
5721 {
5722 /* pairs with tcp_poll() */
5723 smp_mb();
5724 if (sk->sk_socket &&
5725 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
5726 tcp_new_space(sk);
5727 if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5728 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
5729 }
5730 }
5731
5732 static inline void tcp_data_snd_check(struct sock *sk)
5733 {
5734 tcp_push_pending_frames(sk);
5735 tcp_check_space(sk);
5736 }
5737
5738 /*
5739 * Check if sending an ack is needed.
5740 */
5741 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5742 {
5743 struct tcp_sock *tp = tcp_sk(sk);
5744 unsigned long rtt, delay;
5745
5746 /* More than one full frame received... */
5747 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5748 /* ... and right edge of window advances far enough.
5749 * (tcp_recvmsg() will send ACK otherwise).
5750 * If application uses SO_RCVLOWAT, we want send ack now if
5751 * we have not received enough bytes to satisfy the condition.
5752 */
5753 (tp->rcv_nxt - tp->copied_seq < sk->sk_rcvlowat ||
5754 __tcp_select_window(sk) >= tp->rcv_wnd)) ||
5755 /* We ACK each frame or... */
5756 tcp_in_quickack_mode(sk) ||
5757 /* Protocol state mandates a one-time immediate ACK */
5758 inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOW) {
5759 /* If we are running from __release_sock() in user context,
5760 * Defer the ack until tcp_release_cb().
5761 */
5762 if (sock_owned_by_user_nocheck(sk) &&
5763 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_backlog_ack_defer)) {
5764 set_bit(TCP_ACK_DEFERRED, &sk->sk_tsq_flags);
5765 return;
5766 }
5767 send_now:
5768 tcp_send_ack(sk);
5769 return;
5770 }
5771
5772 if (!ofo_possible || RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5773 tcp_send_delayed_ack(sk);
5774 return;
5775 }
5776
5777 if (!tcp_is_sack(tp) ||
5778 tp->compressed_ack >= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_nr))
5779 goto send_now;
5780
5781 if (tp->compressed_ack_rcv_nxt != tp->rcv_nxt) {
5782 tp->compressed_ack_rcv_nxt = tp->rcv_nxt;
5783 tp->dup_ack_counter = 0;
5784 }
5785 if (tp->dup_ack_counter < TCP_FASTRETRANS_THRESH) {
5786 tp->dup_ack_counter++;
5787 goto send_now;
5788 }
5789 tp->compressed_ack++;
5790 if (hrtimer_is_queued(&tp->compressed_ack_timer))
5791 return;
5792
5793 /* compress ack timer : 5 % of rtt, but no more than tcp_comp_sack_delay_ns */
5794
5795 rtt = tp->rcv_rtt_est.rtt_us;
5796 if (tp->srtt_us && tp->srtt_us < rtt)
5797 rtt = tp->srtt_us;
5798
5799 delay = min_t(unsigned long,
5800 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_delay_ns),
5801 rtt * (NSEC_PER_USEC >> 3)/20);
5802 sock_hold(sk);
5803 hrtimer_start_range_ns(&tp->compressed_ack_timer, ns_to_ktime(delay),
5804 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_slack_ns),
5805 HRTIMER_MODE_REL_PINNED_SOFT);
5806 }
5807
5808 static inline void tcp_ack_snd_check(struct sock *sk)
5809 {
5810 if (!inet_csk_ack_scheduled(sk)) {
5811 /* We sent a data segment already. */
5812 return;
5813 }
5814 __tcp_ack_snd_check(sk, 1);
5815 }
5816
5817 /*
5818 * This routine is only called when we have urgent data
5819 * signaled. Its the 'slow' part of tcp_urg. It could be
5820 * moved inline now as tcp_urg is only called from one
5821 * place. We handle URGent data wrong. We have to - as
5822 * BSD still doesn't use the correction from RFC961.
5823 * For 1003.1g we should support a new option TCP_STDURG to permit
5824 * either form (or just set the sysctl tcp_stdurg).
5825 */
5826
5827 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5828 {
5829 struct tcp_sock *tp = tcp_sk(sk);
5830 u32 ptr = ntohs(th->urg_ptr);
5831
5832 if (ptr && !READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_stdurg))
5833 ptr--;
5834 ptr += ntohl(th->seq);
5835
5836 /* Ignore urgent data that we've already seen and read. */
5837 if (after(tp->copied_seq, ptr))
5838 return;
5839
5840 /* Do not replay urg ptr.
5841 *
5842 * NOTE: interesting situation not covered by specs.
5843 * Misbehaving sender may send urg ptr, pointing to segment,
5844 * which we already have in ofo queue. We are not able to fetch
5845 * such data and will stay in TCP_URG_NOTYET until will be eaten
5846 * by recvmsg(). Seems, we are not obliged to handle such wicked
5847 * situations. But it is worth to think about possibility of some
5848 * DoSes using some hypothetical application level deadlock.
5849 */
5850 if (before(ptr, tp->rcv_nxt))
5851 return;
5852
5853 /* Do we already have a newer (or duplicate) urgent pointer? */
5854 if (tp->urg_data && !after(ptr, tp->urg_seq))
5855 return;
5856
5857 /* Tell the world about our new urgent pointer. */
5858 sk_send_sigurg(sk);
5859
5860 /* We may be adding urgent data when the last byte read was
5861 * urgent. To do this requires some care. We cannot just ignore
5862 * tp->copied_seq since we would read the last urgent byte again
5863 * as data, nor can we alter copied_seq until this data arrives
5864 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5865 *
5866 * NOTE. Double Dutch. Rendering to plain English: author of comment
5867 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5868 * and expect that both A and B disappear from stream. This is _wrong_.
5869 * Though this happens in BSD with high probability, this is occasional.
5870 * Any application relying on this is buggy. Note also, that fix "works"
5871 * only in this artificial test. Insert some normal data between A and B and we will
5872 * decline of BSD again. Verdict: it is better to remove to trap
5873 * buggy users.
5874 */
5875 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5876 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5877 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5878 tp->copied_seq++;
5879 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5880 __skb_unlink(skb, &sk->sk_receive_queue);
5881 __kfree_skb(skb);
5882 }
5883 }
5884
5885 WRITE_ONCE(tp->urg_data, TCP_URG_NOTYET);
5886 WRITE_ONCE(tp->urg_seq, ptr);
5887
5888 /* Disable header prediction. */
5889 tp->pred_flags = 0;
5890 }
5891
5892 /* This is the 'fast' part of urgent handling. */
5893 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5894 {
5895 struct tcp_sock *tp = tcp_sk(sk);
5896
5897 /* Check if we get a new urgent pointer - normally not. */
5898 if (unlikely(th->urg))
5899 tcp_check_urg(sk, th);
5900
5901 /* Do we wait for any urgent data? - normally not... */
5902 if (unlikely(tp->urg_data == TCP_URG_NOTYET)) {
5903 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5904 th->syn;
5905
5906 /* Is the urgent pointer pointing into this packet? */
5907 if (ptr < skb->len) {
5908 u8 tmp;
5909 if (skb_copy_bits(skb, ptr, &tmp, 1))
5910 BUG();
5911 WRITE_ONCE(tp->urg_data, TCP_URG_VALID | tmp);
5912 if (!sock_flag(sk, SOCK_DEAD))
5913 sk->sk_data_ready(sk);
5914 }
5915 }
5916 }
5917
5918 /* Accept RST for rcv_nxt - 1 after a FIN.
5919 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a
5920 * FIN is sent followed by a RST packet. The RST is sent with the same
5921 * sequence number as the FIN, and thus according to RFC 5961 a challenge
5922 * ACK should be sent. However, Mac OSX rate limits replies to challenge
5923 * ACKs on the closed socket. In addition middleboxes can drop either the
5924 * challenge ACK or a subsequent RST.
5925 */
5926 static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb)
5927 {
5928 const struct tcp_sock *tp = tcp_sk(sk);
5929
5930 return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) &&
5931 (1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK |
5932 TCPF_CLOSING));
5933 }
5934
5935 /* Does PAWS and seqno based validation of an incoming segment, flags will
5936 * play significant role here.
5937 */
5938 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5939 const struct tcphdr *th, int syn_inerr)
5940 {
5941 struct tcp_sock *tp = tcp_sk(sk);
5942 SKB_DR(reason);
5943
5944 /* RFC1323: H1. Apply PAWS check first. */
5945 if (tcp_fast_parse_options(sock_net(sk), skb, th, tp) &&
5946 tp->rx_opt.saw_tstamp &&
5947 tcp_paws_discard(sk, skb)) {
5948 if (!th->rst) {
5949 if (unlikely(th->syn))
5950 goto syn_challenge;
5951 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5952 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5953 LINUX_MIB_TCPACKSKIPPEDPAWS,
5954 &tp->last_oow_ack_time))
5955 tcp_send_dupack(sk, skb);
5956 SKB_DR_SET(reason, TCP_RFC7323_PAWS);
5957 goto discard;
5958 }
5959 /* Reset is accepted even if it did not pass PAWS. */
5960 }
5961
5962 /* Step 1: check sequence number */
5963 reason = tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
5964 if (reason) {
5965 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5966 * (RST) segments are validated by checking their SEQ-fields."
5967 * And page 69: "If an incoming segment is not acceptable,
5968 * an acknowledgment should be sent in reply (unless the RST
5969 * bit is set, if so drop the segment and return)".
5970 */
5971 if (!th->rst) {
5972 if (th->syn)
5973 goto syn_challenge;
5974 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5975 LINUX_MIB_TCPACKSKIPPEDSEQ,
5976 &tp->last_oow_ack_time))
5977 tcp_send_dupack(sk, skb);
5978 } else if (tcp_reset_check(sk, skb)) {
5979 goto reset;
5980 }
5981 goto discard;
5982 }
5983
5984 /* Step 2: check RST bit */
5985 if (th->rst) {
5986 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a
5987 * FIN and SACK too if available):
5988 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or
5989 * the right-most SACK block,
5990 * then
5991 * RESET the connection
5992 * else
5993 * Send a challenge ACK
5994 */
5995 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt ||
5996 tcp_reset_check(sk, skb))
5997 goto reset;
5998
5999 if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) {
6000 struct tcp_sack_block *sp = &tp->selective_acks[0];
6001 int max_sack = sp[0].end_seq;
6002 int this_sack;
6003
6004 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;
6005 ++this_sack) {
6006 max_sack = after(sp[this_sack].end_seq,
6007 max_sack) ?
6008 sp[this_sack].end_seq : max_sack;
6009 }
6010
6011 if (TCP_SKB_CB(skb)->seq == max_sack)
6012 goto reset;
6013 }
6014
6015 /* Disable TFO if RST is out-of-order
6016 * and no data has been received
6017 * for current active TFO socket
6018 */
6019 if (tp->syn_fastopen && !tp->data_segs_in &&
6020 sk->sk_state == TCP_ESTABLISHED)
6021 tcp_fastopen_active_disable(sk);
6022 tcp_send_challenge_ack(sk);
6023 SKB_DR_SET(reason, TCP_RESET);
6024 goto discard;
6025 }
6026
6027 /* step 3: check security and precedence [ignored] */
6028
6029 /* step 4: Check for a SYN
6030 * RFC 5961 4.2 : Send a challenge ack
6031 */
6032 if (th->syn) {
6033 if (sk->sk_state == TCP_SYN_RECV && sk->sk_socket && th->ack &&
6034 TCP_SKB_CB(skb)->seq + 1 == TCP_SKB_CB(skb)->end_seq &&
6035 TCP_SKB_CB(skb)->seq + 1 == tp->rcv_nxt &&
6036 TCP_SKB_CB(skb)->ack_seq == tp->snd_nxt)
6037 goto pass;
6038 syn_challenge:
6039 if (syn_inerr)
6040 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
6041 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
6042 tcp_send_challenge_ack(sk);
6043 SKB_DR_SET(reason, TCP_INVALID_SYN);
6044 goto discard;
6045 }
6046
6047 pass:
6048 bpf_skops_parse_hdr(sk, skb);
6049
6050 return true;
6051
6052 discard:
6053 tcp_drop_reason(sk, skb, reason);
6054 return false;
6055
6056 reset:
6057 tcp_reset(sk, skb);
6058 __kfree_skb(skb);
6059 return false;
6060 }
6061
6062 /*
6063 * TCP receive function for the ESTABLISHED state.
6064 *
6065 * It is split into a fast path and a slow path. The fast path is
6066 * disabled when:
6067 * - A zero window was announced from us - zero window probing
6068 * is only handled properly in the slow path.
6069 * - Out of order segments arrived.
6070 * - Urgent data is expected.
6071 * - There is no buffer space left
6072 * - Unexpected TCP flags/window values/header lengths are received
6073 * (detected by checking the TCP header against pred_flags)
6074 * - Data is sent in both directions. Fast path only supports pure senders
6075 * or pure receivers (this means either the sequence number or the ack
6076 * value must stay constant)
6077 * - Unexpected TCP option.
6078 *
6079 * When these conditions are not satisfied it drops into a standard
6080 * receive procedure patterned after RFC793 to handle all cases.
6081 * The first three cases are guaranteed by proper pred_flags setting,
6082 * the rest is checked inline. Fast processing is turned on in
6083 * tcp_data_queue when everything is OK.
6084 */
6085 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb)
6086 {
6087 enum skb_drop_reason reason = SKB_DROP_REASON_NOT_SPECIFIED;
6088 const struct tcphdr *th = (const struct tcphdr *)skb->data;
6089 struct tcp_sock *tp = tcp_sk(sk);
6090 unsigned int len = skb->len;
6091
6092 /* TCP congestion window tracking */
6093 trace_tcp_probe(sk, skb);
6094
6095 tcp_mstamp_refresh(tp);
6096 if (unlikely(!rcu_access_pointer(sk->sk_rx_dst)))
6097 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
6098 /*
6099 * Header prediction.
6100 * The code loosely follows the one in the famous
6101 * "30 instruction TCP receive" Van Jacobson mail.
6102 *
6103 * Van's trick is to deposit buffers into socket queue
6104 * on a device interrupt, to call tcp_recv function
6105 * on the receive process context and checksum and copy
6106 * the buffer to user space. smart...
6107 *
6108 * Our current scheme is not silly either but we take the
6109 * extra cost of the net_bh soft interrupt processing...
6110 * We do checksum and copy also but from device to kernel.
6111 */
6112
6113 tp->rx_opt.saw_tstamp = 0;
6114
6115 /* pred_flags is 0xS?10 << 16 + snd_wnd
6116 * if header_prediction is to be made
6117 * 'S' will always be tp->tcp_header_len >> 2
6118 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
6119 * turn it off (when there are holes in the receive
6120 * space for instance)
6121 * PSH flag is ignored.
6122 */
6123
6124 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
6125 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
6126 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
6127 int tcp_header_len = tp->tcp_header_len;
6128
6129 /* Timestamp header prediction: tcp_header_len
6130 * is automatically equal to th->doff*4 due to pred_flags
6131 * match.
6132 */
6133
6134 /* Check timestamp */
6135 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
6136 /* No? Slow path! */
6137 if (!tcp_parse_aligned_timestamp(tp, th))
6138 goto slow_path;
6139
6140 /* If PAWS failed, check it more carefully in slow path */
6141 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
6142 goto slow_path;
6143
6144 /* DO NOT update ts_recent here, if checksum fails
6145 * and timestamp was corrupted part, it will result
6146 * in a hung connection since we will drop all
6147 * future packets due to the PAWS test.
6148 */
6149 }
6150
6151 if (len <= tcp_header_len) {
6152 /* Bulk data transfer: sender */
6153 if (len == tcp_header_len) {
6154 /* Predicted packet is in window by definition.
6155 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
6156 * Hence, check seq<=rcv_wup reduces to:
6157 */
6158 if (tcp_header_len ==
6159 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
6160 tp->rcv_nxt == tp->rcv_wup)
6161 tcp_store_ts_recent(tp);
6162
6163 /* We know that such packets are checksummed
6164 * on entry.
6165 */
6166 tcp_ack(sk, skb, 0);
6167 __kfree_skb(skb);
6168 tcp_data_snd_check(sk);
6169 /* When receiving pure ack in fast path, update
6170 * last ts ecr directly instead of calling
6171 * tcp_rcv_rtt_measure_ts()
6172 */
6173 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
6174 return;
6175 } else { /* Header too small */
6176 reason = SKB_DROP_REASON_PKT_TOO_SMALL;
6177 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
6178 goto discard;
6179 }
6180 } else {
6181 int eaten = 0;
6182 bool fragstolen = false;
6183
6184 if (tcp_checksum_complete(skb))
6185 goto csum_error;
6186
6187 if ((int)skb->truesize > sk->sk_forward_alloc)
6188 goto step5;
6189
6190 /* Predicted packet is in window by definition.
6191 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
6192 * Hence, check seq<=rcv_wup reduces to:
6193 */
6194 if (tcp_header_len ==
6195 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
6196 tp->rcv_nxt == tp->rcv_wup)
6197 tcp_store_ts_recent(tp);
6198
6199 tcp_rcv_rtt_measure_ts(sk, skb);
6200
6201 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS);
6202
6203 /* Bulk data transfer: receiver */
6204 skb_dst_drop(skb);
6205 __skb_pull(skb, tcp_header_len);
6206 eaten = tcp_queue_rcv(sk, skb, &fragstolen);
6207
6208 tcp_event_data_recv(sk, skb);
6209
6210 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
6211 /* Well, only one small jumplet in fast path... */
6212 tcp_ack(sk, skb, FLAG_DATA);
6213 tcp_data_snd_check(sk);
6214 if (!inet_csk_ack_scheduled(sk))
6215 goto no_ack;
6216 } else {
6217 tcp_update_wl(tp, TCP_SKB_CB(skb)->seq);
6218 }
6219
6220 __tcp_ack_snd_check(sk, 0);
6221 no_ack:
6222 if (eaten)
6223 kfree_skb_partial(skb, fragstolen);
6224 tcp_data_ready(sk);
6225 return;
6226 }
6227 }
6228
6229 slow_path:
6230 if (len < (th->doff << 2) || tcp_checksum_complete(skb))
6231 goto csum_error;
6232
6233 if (!th->ack && !th->rst && !th->syn) {
6234 reason = SKB_DROP_REASON_TCP_FLAGS;
6235 goto discard;
6236 }
6237
6238 /*
6239 * Standard slow path.
6240 */
6241
6242 if (!tcp_validate_incoming(sk, skb, th, 1))
6243 return;
6244
6245 step5:
6246 reason = tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT);
6247 if ((int)reason < 0) {
6248 reason = -reason;
6249 goto discard;
6250 }
6251 tcp_rcv_rtt_measure_ts(sk, skb);
6252
6253 /* Process urgent data. */
6254 tcp_urg(sk, skb, th);
6255
6256 /* step 7: process the segment text */
6257 tcp_data_queue(sk, skb);
6258
6259 tcp_data_snd_check(sk);
6260 tcp_ack_snd_check(sk);
6261 return;
6262
6263 csum_error:
6264 reason = SKB_DROP_REASON_TCP_CSUM;
6265 trace_tcp_bad_csum(skb);
6266 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
6267 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
6268
6269 discard:
6270 tcp_drop_reason(sk, skb, reason);
6271 }
6272 EXPORT_SYMBOL(tcp_rcv_established);
6273
6274 void tcp_init_transfer(struct sock *sk, int bpf_op, struct sk_buff *skb)
6275 {
6276 struct inet_connection_sock *icsk = inet_csk(sk);
6277 struct tcp_sock *tp = tcp_sk(sk);
6278
6279 tcp_mtup_init(sk);
6280 icsk->icsk_af_ops->rebuild_header(sk);
6281 tcp_init_metrics(sk);
6282
6283 /* Initialize the congestion window to start the transfer.
6284 * Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
6285 * retransmitted. In light of RFC6298 more aggressive 1sec
6286 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
6287 * retransmission has occurred.
6288 */
6289 if (tp->total_retrans > 1 && tp->undo_marker)
6290 tcp_snd_cwnd_set(tp, 1);
6291 else
6292 tcp_snd_cwnd_set(tp, tcp_init_cwnd(tp, __sk_dst_get(sk)));
6293 tp->snd_cwnd_stamp = tcp_jiffies32;
6294
6295 bpf_skops_established(sk, bpf_op, skb);
6296 /* Initialize congestion control unless BPF initialized it already: */
6297 if (!icsk->icsk_ca_initialized)
6298 tcp_init_congestion_control(sk);
6299 tcp_init_buffer_space(sk);
6300 }
6301
6302 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
6303 {
6304 struct tcp_sock *tp = tcp_sk(sk);
6305 struct inet_connection_sock *icsk = inet_csk(sk);
6306
6307 tcp_ao_finish_connect(sk, skb);
6308 tcp_set_state(sk, TCP_ESTABLISHED);
6309 icsk->icsk_ack.lrcvtime = tcp_jiffies32;
6310
6311 if (skb) {
6312 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
6313 security_inet_conn_established(sk, skb);
6314 sk_mark_napi_id(sk, skb);
6315 }
6316
6317 tcp_init_transfer(sk, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB, skb);
6318
6319 /* Prevent spurious tcp_cwnd_restart() on first data
6320 * packet.
6321 */
6322 tp->lsndtime = tcp_jiffies32;
6323
6324 if (sock_flag(sk, SOCK_KEEPOPEN))
6325 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
6326
6327 if (!tp->rx_opt.snd_wscale)
6328 __tcp_fast_path_on(tp, tp->snd_wnd);
6329 else
6330 tp->pred_flags = 0;
6331 }
6332
6333 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
6334 struct tcp_fastopen_cookie *cookie)
6335 {
6336 struct tcp_sock *tp = tcp_sk(sk);
6337 struct sk_buff *data = tp->syn_data ? tcp_rtx_queue_head(sk) : NULL;
6338 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
6339 bool syn_drop = false;
6340
6341 if (mss == tp->rx_opt.user_mss) {
6342 struct tcp_options_received opt;
6343
6344 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
6345 tcp_clear_options(&opt);
6346 opt.user_mss = opt.mss_clamp = 0;
6347 tcp_parse_options(sock_net(sk), synack, &opt, 0, NULL);
6348 mss = opt.mss_clamp;
6349 }
6350
6351 if (!tp->syn_fastopen) {
6352 /* Ignore an unsolicited cookie */
6353 cookie->len = -1;
6354 } else if (tp->total_retrans) {
6355 /* SYN timed out and the SYN-ACK neither has a cookie nor
6356 * acknowledges data. Presumably the remote received only
6357 * the retransmitted (regular) SYNs: either the original
6358 * SYN-data or the corresponding SYN-ACK was dropped.
6359 */
6360 syn_drop = (cookie->len < 0 && data);
6361 } else if (cookie->len < 0 && !tp->syn_data) {
6362 /* We requested a cookie but didn't get it. If we did not use
6363 * the (old) exp opt format then try so next time (try_exp=1).
6364 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
6365 */
6366 try_exp = tp->syn_fastopen_exp ? 2 : 1;
6367 }
6368
6369 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
6370
6371 if (data) { /* Retransmit unacked data in SYN */
6372 if (tp->total_retrans)
6373 tp->fastopen_client_fail = TFO_SYN_RETRANSMITTED;
6374 else
6375 tp->fastopen_client_fail = TFO_DATA_NOT_ACKED;
6376 skb_rbtree_walk_from(data)
6377 tcp_mark_skb_lost(sk, data);
6378 tcp_non_congestion_loss_retransmit(sk);
6379 NET_INC_STATS(sock_net(sk),
6380 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
6381 return true;
6382 }
6383 tp->syn_data_acked = tp->syn_data;
6384 if (tp->syn_data_acked) {
6385 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
6386 /* SYN-data is counted as two separate packets in tcp_ack() */
6387 if (tp->delivered > 1)
6388 --tp->delivered;
6389 }
6390
6391 tcp_fastopen_add_skb(sk, synack);
6392
6393 return false;
6394 }
6395
6396 static void smc_check_reset_syn(struct tcp_sock *tp)
6397 {
6398 #if IS_ENABLED(CONFIG_SMC)
6399 if (static_branch_unlikely(&tcp_have_smc)) {
6400 if (tp->syn_smc && !tp->rx_opt.smc_ok)
6401 tp->syn_smc = 0;
6402 }
6403 #endif
6404 }
6405
6406 static void tcp_try_undo_spurious_syn(struct sock *sk)
6407 {
6408 struct tcp_sock *tp = tcp_sk(sk);
6409 u32 syn_stamp;
6410
6411 /* undo_marker is set when SYN or SYNACK times out. The timeout is
6412 * spurious if the ACK's timestamp option echo value matches the
6413 * original SYN timestamp.
6414 */
6415 syn_stamp = tp->retrans_stamp;
6416 if (tp->undo_marker && syn_stamp && tp->rx_opt.saw_tstamp &&
6417 syn_stamp == tp->rx_opt.rcv_tsecr)
6418 tp->undo_marker = 0;
6419 }
6420
6421 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
6422 const struct tcphdr *th)
6423 {
6424 struct inet_connection_sock *icsk = inet_csk(sk);
6425 struct tcp_sock *tp = tcp_sk(sk);
6426 struct tcp_fastopen_cookie foc = { .len = -1 };
6427 int saved_clamp = tp->rx_opt.mss_clamp;
6428 bool fastopen_fail;
6429 SKB_DR(reason);
6430
6431 tcp_parse_options(sock_net(sk), skb, &tp->rx_opt, 0, &foc);
6432 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
6433 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
6434
6435 if (th->ack) {
6436 /* rfc793:
6437 * "If the state is SYN-SENT then
6438 * first check the ACK bit
6439 * If the ACK bit is set
6440 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
6441 * a reset (unless the RST bit is set, if so drop
6442 * the segment and return)"
6443 */
6444 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
6445 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
6446 /* Previous FIN/ACK or RST/ACK might be ignored. */
6447 if (icsk->icsk_retransmits == 0)
6448 inet_csk_reset_xmit_timer(sk,
6449 ICSK_TIME_RETRANS,
6450 TCP_TIMEOUT_MIN, TCP_RTO_MAX);
6451 SKB_DR_SET(reason, TCP_INVALID_ACK_SEQUENCE);
6452 goto reset_and_undo;
6453 }
6454
6455 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
6456 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
6457 tcp_time_stamp_ts(tp))) {
6458 NET_INC_STATS(sock_net(sk),
6459 LINUX_MIB_PAWSACTIVEREJECTED);
6460 SKB_DR_SET(reason, TCP_RFC7323_PAWS);
6461 goto reset_and_undo;
6462 }
6463
6464 /* Now ACK is acceptable.
6465 *
6466 * "If the RST bit is set
6467 * If the ACK was acceptable then signal the user "error:
6468 * connection reset", drop the segment, enter CLOSED state,
6469 * delete TCB, and return."
6470 */
6471
6472 if (th->rst) {
6473 tcp_reset(sk, skb);
6474 consume:
6475 __kfree_skb(skb);
6476 return 0;
6477 }
6478
6479 /* rfc793:
6480 * "fifth, if neither of the SYN or RST bits is set then
6481 * drop the segment and return."
6482 *
6483 * See note below!
6484 * --ANK(990513)
6485 */
6486 if (!th->syn) {
6487 SKB_DR_SET(reason, TCP_FLAGS);
6488 goto discard_and_undo;
6489 }
6490 /* rfc793:
6491 * "If the SYN bit is on ...
6492 * are acceptable then ...
6493 * (our SYN has been ACKed), change the connection
6494 * state to ESTABLISHED..."
6495 */
6496
6497 tcp_ecn_rcv_synack(tp, th);
6498
6499 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6500 tcp_try_undo_spurious_syn(sk);
6501 tcp_ack(sk, skb, FLAG_SLOWPATH);
6502
6503 /* Ok.. it's good. Set up sequence numbers and
6504 * move to established.
6505 */
6506 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1);
6507 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
6508
6509 /* RFC1323: The window in SYN & SYN/ACK segments is
6510 * never scaled.
6511 */
6512 tp->snd_wnd = ntohs(th->window);
6513
6514 if (!tp->rx_opt.wscale_ok) {
6515 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
6516 WRITE_ONCE(tp->window_clamp,
6517 min(tp->window_clamp, 65535U));
6518 }
6519
6520 if (tp->rx_opt.saw_tstamp) {
6521 tp->rx_opt.tstamp_ok = 1;
6522 tp->tcp_header_len =
6523 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
6524 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6525 tcp_store_ts_recent(tp);
6526 } else {
6527 tp->tcp_header_len = sizeof(struct tcphdr);
6528 }
6529
6530 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
6531 tcp_initialize_rcv_mss(sk);
6532
6533 /* Remember, tcp_poll() does not lock socket!
6534 * Change state from SYN-SENT only after copied_seq
6535 * is initialized. */
6536 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6537
6538 smc_check_reset_syn(tp);
6539
6540 smp_mb();
6541
6542 tcp_finish_connect(sk, skb);
6543
6544 fastopen_fail = (tp->syn_fastopen || tp->syn_data) &&
6545 tcp_rcv_fastopen_synack(sk, skb, &foc);
6546
6547 if (!sock_flag(sk, SOCK_DEAD)) {
6548 sk->sk_state_change(sk);
6549 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6550 }
6551 if (fastopen_fail)
6552 return -1;
6553 if (sk->sk_write_pending ||
6554 READ_ONCE(icsk->icsk_accept_queue.rskq_defer_accept) ||
6555 inet_csk_in_pingpong_mode(sk)) {
6556 /* Save one ACK. Data will be ready after
6557 * several ticks, if write_pending is set.
6558 *
6559 * It may be deleted, but with this feature tcpdumps
6560 * look so _wonderfully_ clever, that I was not able
6561 * to stand against the temptation 8) --ANK
6562 */
6563 inet_csk_schedule_ack(sk);
6564 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
6565 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
6566 TCP_DELACK_MAX, TCP_RTO_MAX);
6567 goto consume;
6568 }
6569 tcp_send_ack(sk);
6570 return -1;
6571 }
6572
6573 /* No ACK in the segment */
6574
6575 if (th->rst) {
6576 /* rfc793:
6577 * "If the RST bit is set
6578 *
6579 * Otherwise (no ACK) drop the segment and return."
6580 */
6581 SKB_DR_SET(reason, TCP_RESET);
6582 goto discard_and_undo;
6583 }
6584
6585 /* PAWS check. */
6586 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
6587 tcp_paws_reject(&tp->rx_opt, 0)) {
6588 SKB_DR_SET(reason, TCP_RFC7323_PAWS);
6589 goto discard_and_undo;
6590 }
6591 if (th->syn) {
6592 /* We see SYN without ACK. It is attempt of
6593 * simultaneous connect with crossed SYNs.
6594 * Particularly, it can be connect to self.
6595 */
6596 #ifdef CONFIG_TCP_AO
6597 struct tcp_ao_info *ao;
6598
6599 ao = rcu_dereference_protected(tp->ao_info,
6600 lockdep_sock_is_held(sk));
6601 if (ao) {
6602 WRITE_ONCE(ao->risn, th->seq);
6603 ao->rcv_sne = 0;
6604 }
6605 #endif
6606 tcp_set_state(sk, TCP_SYN_RECV);
6607
6608 if (tp->rx_opt.saw_tstamp) {
6609 tp->rx_opt.tstamp_ok = 1;
6610 tcp_store_ts_recent(tp);
6611 tp->tcp_header_len =
6612 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
6613 } else {
6614 tp->tcp_header_len = sizeof(struct tcphdr);
6615 }
6616
6617 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1);
6618 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6619 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
6620
6621 /* RFC1323: The window in SYN & SYN/ACK segments is
6622 * never scaled.
6623 */
6624 tp->snd_wnd = ntohs(th->window);
6625 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
6626 tp->max_window = tp->snd_wnd;
6627
6628 tcp_ecn_rcv_syn(tp, th);
6629
6630 tcp_mtup_init(sk);
6631 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
6632 tcp_initialize_rcv_mss(sk);
6633
6634 tcp_send_synack(sk);
6635 #if 0
6636 /* Note, we could accept data and URG from this segment.
6637 * There are no obstacles to make this (except that we must
6638 * either change tcp_recvmsg() to prevent it from returning data
6639 * before 3WHS completes per RFC793, or employ TCP Fast Open).
6640 *
6641 * However, if we ignore data in ACKless segments sometimes,
6642 * we have no reasons to accept it sometimes.
6643 * Also, seems the code doing it in step6 of tcp_rcv_state_process
6644 * is not flawless. So, discard packet for sanity.
6645 * Uncomment this return to process the data.
6646 */
6647 return -1;
6648 #else
6649 goto consume;
6650 #endif
6651 }
6652 /* "fifth, if neither of the SYN or RST bits is set then
6653 * drop the segment and return."
6654 */
6655
6656 discard_and_undo:
6657 tcp_clear_options(&tp->rx_opt);
6658 tp->rx_opt.mss_clamp = saved_clamp;
6659 tcp_drop_reason(sk, skb, reason);
6660 return 0;
6661
6662 reset_and_undo:
6663 tcp_clear_options(&tp->rx_opt);
6664 tp->rx_opt.mss_clamp = saved_clamp;
6665 /* we can reuse/return @reason to its caller to handle the exception */
6666 return reason;
6667 }
6668
6669 static void tcp_rcv_synrecv_state_fastopen(struct sock *sk)
6670 {
6671 struct tcp_sock *tp = tcp_sk(sk);
6672 struct request_sock *req;
6673
6674 /* If we are still handling the SYNACK RTO, see if timestamp ECR allows
6675 * undo. If peer SACKs triggered fast recovery, we can't undo here.
6676 */
6677 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss && !tp->packets_out)
6678 tcp_try_undo_recovery(sk);
6679
6680 tcp_update_rto_time(tp);
6681 inet_csk(sk)->icsk_retransmits = 0;
6682 /* In tcp_fastopen_synack_timer() on the first SYNACK RTO we set
6683 * retrans_stamp but don't enter CA_Loss, so in case that happened we
6684 * need to zero retrans_stamp here to prevent spurious
6685 * retransmits_timed_out(). However, if the ACK of our SYNACK caused us
6686 * to enter CA_Recovery then we need to leave retrans_stamp as it was
6687 * set entering CA_Recovery, for correct retransmits_timed_out() and
6688 * undo behavior.
6689 */
6690 tcp_retrans_stamp_cleanup(sk);
6691
6692 /* Once we leave TCP_SYN_RECV or TCP_FIN_WAIT_1,
6693 * we no longer need req so release it.
6694 */
6695 req = rcu_dereference_protected(tp->fastopen_rsk,
6696 lockdep_sock_is_held(sk));
6697 reqsk_fastopen_remove(sk, req, false);
6698
6699 /* Re-arm the timer because data may have been sent out.
6700 * This is similar to the regular data transmission case
6701 * when new data has just been ack'ed.
6702 *
6703 * (TFO) - we could try to be more aggressive and
6704 * retransmitting any data sooner based on when they
6705 * are sent out.
6706 */
6707 tcp_rearm_rto(sk);
6708 }
6709
6710 /*
6711 * This function implements the receiving procedure of RFC 793 for
6712 * all states except ESTABLISHED and TIME_WAIT.
6713 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
6714 * address independent.
6715 */
6716
6717 enum skb_drop_reason
6718 tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
6719 {
6720 struct tcp_sock *tp = tcp_sk(sk);
6721 struct inet_connection_sock *icsk = inet_csk(sk);
6722 const struct tcphdr *th = tcp_hdr(skb);
6723 struct request_sock *req;
6724 int queued = 0;
6725 SKB_DR(reason);
6726
6727 switch (sk->sk_state) {
6728 case TCP_CLOSE:
6729 SKB_DR_SET(reason, TCP_CLOSE);
6730 goto discard;
6731
6732 case TCP_LISTEN:
6733 if (th->ack)
6734 return SKB_DROP_REASON_TCP_FLAGS;
6735
6736 if (th->rst) {
6737 SKB_DR_SET(reason, TCP_RESET);
6738 goto discard;
6739 }
6740 if (th->syn) {
6741 if (th->fin) {
6742 SKB_DR_SET(reason, TCP_FLAGS);
6743 goto discard;
6744 }
6745 /* It is possible that we process SYN packets from backlog,
6746 * so we need to make sure to disable BH and RCU right there.
6747 */
6748 rcu_read_lock();
6749 local_bh_disable();
6750 icsk->icsk_af_ops->conn_request(sk, skb);
6751 local_bh_enable();
6752 rcu_read_unlock();
6753
6754 consume_skb(skb);
6755 return 0;
6756 }
6757 SKB_DR_SET(reason, TCP_FLAGS);
6758 goto discard;
6759
6760 case TCP_SYN_SENT:
6761 tp->rx_opt.saw_tstamp = 0;
6762 tcp_mstamp_refresh(tp);
6763 queued = tcp_rcv_synsent_state_process(sk, skb, th);
6764 if (queued >= 0)
6765 return queued;
6766
6767 /* Do step6 onward by hand. */
6768 tcp_urg(sk, skb, th);
6769 __kfree_skb(skb);
6770 tcp_data_snd_check(sk);
6771 return 0;
6772 }
6773
6774 tcp_mstamp_refresh(tp);
6775 tp->rx_opt.saw_tstamp = 0;
6776 req = rcu_dereference_protected(tp->fastopen_rsk,
6777 lockdep_sock_is_held(sk));
6778 if (req) {
6779 bool req_stolen;
6780
6781 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
6782 sk->sk_state != TCP_FIN_WAIT1);
6783
6784 if (!tcp_check_req(sk, skb, req, true, &req_stolen)) {
6785 SKB_DR_SET(reason, TCP_FASTOPEN);
6786 goto discard;
6787 }
6788 }
6789
6790 if (!th->ack && !th->rst && !th->syn) {
6791 SKB_DR_SET(reason, TCP_FLAGS);
6792 goto discard;
6793 }
6794 if (!tcp_validate_incoming(sk, skb, th, 0))
6795 return 0;
6796
6797 /* step 5: check the ACK field */
6798 reason = tcp_ack(sk, skb, FLAG_SLOWPATH |
6799 FLAG_UPDATE_TS_RECENT |
6800 FLAG_NO_CHALLENGE_ACK);
6801
6802 if ((int)reason <= 0) {
6803 if (sk->sk_state == TCP_SYN_RECV) {
6804 /* send one RST */
6805 if (!reason)
6806 return SKB_DROP_REASON_TCP_OLD_ACK;
6807 return -reason;
6808 }
6809 /* accept old ack during closing */
6810 if ((int)reason < 0) {
6811 tcp_send_challenge_ack(sk);
6812 reason = -reason;
6813 goto discard;
6814 }
6815 }
6816 SKB_DR_SET(reason, NOT_SPECIFIED);
6817 switch (sk->sk_state) {
6818 case TCP_SYN_RECV:
6819 tp->delivered++; /* SYN-ACK delivery isn't tracked in tcp_ack */
6820 if (!tp->srtt_us)
6821 tcp_synack_rtt_meas(sk, req);
6822
6823 if (req) {
6824 tcp_rcv_synrecv_state_fastopen(sk);
6825 } else {
6826 tcp_try_undo_spurious_syn(sk);
6827 tp->retrans_stamp = 0;
6828 tcp_init_transfer(sk, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB,
6829 skb);
6830 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6831 }
6832 tcp_ao_established(sk);
6833 smp_mb();
6834 tcp_set_state(sk, TCP_ESTABLISHED);
6835 sk->sk_state_change(sk);
6836
6837 /* Note, that this wakeup is only for marginal crossed SYN case.
6838 * Passively open sockets are not waked up, because
6839 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
6840 */
6841 if (sk->sk_socket)
6842 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6843
6844 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
6845 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
6846 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6847
6848 if (tp->rx_opt.tstamp_ok)
6849 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6850
6851 if (!inet_csk(sk)->icsk_ca_ops->cong_control)
6852 tcp_update_pacing_rate(sk);
6853
6854 /* Prevent spurious tcp_cwnd_restart() on first data packet */
6855 tp->lsndtime = tcp_jiffies32;
6856
6857 tcp_initialize_rcv_mss(sk);
6858 tcp_fast_path_on(tp);
6859 if (sk->sk_shutdown & SEND_SHUTDOWN)
6860 tcp_shutdown(sk, SEND_SHUTDOWN);
6861 break;
6862
6863 case TCP_FIN_WAIT1: {
6864 int tmo;
6865
6866 if (req)
6867 tcp_rcv_synrecv_state_fastopen(sk);
6868
6869 if (tp->snd_una != tp->write_seq)
6870 break;
6871
6872 tcp_set_state(sk, TCP_FIN_WAIT2);
6873 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | SEND_SHUTDOWN);
6874
6875 sk_dst_confirm(sk);
6876
6877 if (!sock_flag(sk, SOCK_DEAD)) {
6878 /* Wake up lingering close() */
6879 sk->sk_state_change(sk);
6880 break;
6881 }
6882
6883 if (READ_ONCE(tp->linger2) < 0) {
6884 tcp_done(sk);
6885 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6886 return SKB_DROP_REASON_TCP_ABORT_ON_DATA;
6887 }
6888 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6889 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6890 /* Receive out of order FIN after close() */
6891 if (tp->syn_fastopen && th->fin)
6892 tcp_fastopen_active_disable(sk);
6893 tcp_done(sk);
6894 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6895 return SKB_DROP_REASON_TCP_ABORT_ON_DATA;
6896 }
6897
6898 tmo = tcp_fin_time(sk);
6899 if (tmo > TCP_TIMEWAIT_LEN) {
6900 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
6901 } else if (th->fin || sock_owned_by_user(sk)) {
6902 /* Bad case. We could lose such FIN otherwise.
6903 * It is not a big problem, but it looks confusing
6904 * and not so rare event. We still can lose it now,
6905 * if it spins in bh_lock_sock(), but it is really
6906 * marginal case.
6907 */
6908 inet_csk_reset_keepalive_timer(sk, tmo);
6909 } else {
6910 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
6911 goto consume;
6912 }
6913 break;
6914 }
6915
6916 case TCP_CLOSING:
6917 if (tp->snd_una == tp->write_seq) {
6918 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6919 goto consume;
6920 }
6921 break;
6922
6923 case TCP_LAST_ACK:
6924 if (tp->snd_una == tp->write_seq) {
6925 tcp_update_metrics(sk);
6926 tcp_done(sk);
6927 goto consume;
6928 }
6929 break;
6930 }
6931
6932 /* step 6: check the URG bit */
6933 tcp_urg(sk, skb, th);
6934
6935 /* step 7: process the segment text */
6936 switch (sk->sk_state) {
6937 case TCP_CLOSE_WAIT:
6938 case TCP_CLOSING:
6939 case TCP_LAST_ACK:
6940 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
6941 /* If a subflow has been reset, the packet should not
6942 * continue to be processed, drop the packet.
6943 */
6944 if (sk_is_mptcp(sk) && !mptcp_incoming_options(sk, skb))
6945 goto discard;
6946 break;
6947 }
6948 fallthrough;
6949 case TCP_FIN_WAIT1:
6950 case TCP_FIN_WAIT2:
6951 /* RFC 793 says to queue data in these states,
6952 * RFC 1122 says we MUST send a reset.
6953 * BSD 4.4 also does reset.
6954 */
6955 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6956 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6957 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6958 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6959 tcp_reset(sk, skb);
6960 return SKB_DROP_REASON_TCP_ABORT_ON_DATA;
6961 }
6962 }
6963 fallthrough;
6964 case TCP_ESTABLISHED:
6965 tcp_data_queue(sk, skb);
6966 queued = 1;
6967 break;
6968 }
6969
6970 /* tcp_data could move socket to TIME-WAIT */
6971 if (sk->sk_state != TCP_CLOSE) {
6972 tcp_data_snd_check(sk);
6973 tcp_ack_snd_check(sk);
6974 }
6975
6976 if (!queued) {
6977 discard:
6978 tcp_drop_reason(sk, skb, reason);
6979 }
6980 return 0;
6981
6982 consume:
6983 __kfree_skb(skb);
6984 return 0;
6985 }
6986 EXPORT_SYMBOL(tcp_rcv_state_process);
6987
6988 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
6989 {
6990 struct inet_request_sock *ireq = inet_rsk(req);
6991
6992 if (family == AF_INET)
6993 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6994 &ireq->ir_rmt_addr, port);
6995 #if IS_ENABLED(CONFIG_IPV6)
6996 else if (family == AF_INET6)
6997 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6998 &ireq->ir_v6_rmt_addr, port);
6999 #endif
7000 }
7001
7002 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
7003 *
7004 * If we receive a SYN packet with these bits set, it means a
7005 * network is playing bad games with TOS bits. In order to
7006 * avoid possible false congestion notifications, we disable
7007 * TCP ECN negotiation.
7008 *
7009 * Exception: tcp_ca wants ECN. This is required for DCTCP
7010 * congestion control: Linux DCTCP asserts ECT on all packets,
7011 * including SYN, which is most optimal solution; however,
7012 * others, such as FreeBSD do not.
7013 *
7014 * Exception: At least one of the reserved bits of the TCP header (th->res1) is
7015 * set, indicating the use of a future TCP extension (such as AccECN). See
7016 * RFC8311 ยง4.3 which updates RFC3168 to allow the development of such
7017 * extensions.
7018 */
7019 static void tcp_ecn_create_request(struct request_sock *req,
7020 const struct sk_buff *skb,
7021 const struct sock *listen_sk,
7022 const struct dst_entry *dst)
7023 {
7024 const struct tcphdr *th = tcp_hdr(skb);
7025 const struct net *net = sock_net(listen_sk);
7026 bool th_ecn = th->ece && th->cwr;
7027 bool ect, ecn_ok;
7028 u32 ecn_ok_dst;
7029
7030 if (!th_ecn)
7031 return;
7032
7033 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
7034 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK);
7035 ecn_ok = READ_ONCE(net->ipv4.sysctl_tcp_ecn) || ecn_ok_dst;
7036
7037 if (((!ect || th->res1) && ecn_ok) || tcp_ca_needs_ecn(listen_sk) ||
7038 (ecn_ok_dst & DST_FEATURE_ECN_CA) ||
7039 tcp_bpf_ca_needs_ecn((struct sock *)req))
7040 inet_rsk(req)->ecn_ok = 1;
7041 }
7042
7043 static void tcp_openreq_init(struct request_sock *req,
7044 const struct tcp_options_received *rx_opt,
7045 struct sk_buff *skb, const struct sock *sk)
7046 {
7047 struct inet_request_sock *ireq = inet_rsk(req);
7048
7049 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */
7050 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
7051 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
7052 tcp_rsk(req)->snt_synack = 0;
7053 tcp_rsk(req)->last_oow_ack_time = 0;
7054 req->mss = rx_opt->mss_clamp;
7055 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
7056 ireq->tstamp_ok = rx_opt->tstamp_ok;
7057 ireq->sack_ok = rx_opt->sack_ok;
7058 ireq->snd_wscale = rx_opt->snd_wscale;
7059 ireq->wscale_ok = rx_opt->wscale_ok;
7060 ireq->acked = 0;
7061 ireq->ecn_ok = 0;
7062 ireq->ir_rmt_port = tcp_hdr(skb)->source;
7063 ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
7064 ireq->ir_mark = inet_request_mark(sk, skb);
7065 #if IS_ENABLED(CONFIG_SMC)
7066 ireq->smc_ok = rx_opt->smc_ok && !(tcp_sk(sk)->smc_hs_congested &&
7067 tcp_sk(sk)->smc_hs_congested(sk));
7068 #endif
7069 }
7070
7071 /*
7072 * Return true if a syncookie should be sent
7073 */
7074 static bool tcp_syn_flood_action(struct sock *sk, const char *proto)
7075 {
7076 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
7077 const char *msg = "Dropping request";
7078 struct net *net = sock_net(sk);
7079 bool want_cookie = false;
7080 u8 syncookies;
7081
7082 syncookies = READ_ONCE(net->ipv4.sysctl_tcp_syncookies);
7083
7084 #ifdef CONFIG_SYN_COOKIES
7085 if (syncookies) {
7086 msg = "Sending cookies";
7087 want_cookie = true;
7088 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
7089 } else
7090 #endif
7091 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
7092
7093 if (!READ_ONCE(queue->synflood_warned) && syncookies != 2 &&
7094 xchg(&queue->synflood_warned, 1) == 0) {
7095 if (IS_ENABLED(CONFIG_IPV6) && sk->sk_family == AF_INET6) {
7096 net_info_ratelimited("%s: Possible SYN flooding on port [%pI6c]:%u. %s.\n",
7097 proto, inet6_rcv_saddr(sk),
7098 sk->sk_num, msg);
7099 } else {
7100 net_info_ratelimited("%s: Possible SYN flooding on port %pI4:%u. %s.\n",
7101 proto, &sk->sk_rcv_saddr,
7102 sk->sk_num, msg);
7103 }
7104 }
7105
7106 return want_cookie;
7107 }
7108
7109 static void tcp_reqsk_record_syn(const struct sock *sk,
7110 struct request_sock *req,
7111 const struct sk_buff *skb)
7112 {
7113 if (tcp_sk(sk)->save_syn) {
7114 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
7115 struct saved_syn *saved_syn;
7116 u32 mac_hdrlen;
7117 void *base;
7118
7119 if (tcp_sk(sk)->save_syn == 2) { /* Save full header. */
7120 base = skb_mac_header(skb);
7121 mac_hdrlen = skb_mac_header_len(skb);
7122 len += mac_hdrlen;
7123 } else {
7124 base = skb_network_header(skb);
7125 mac_hdrlen = 0;
7126 }
7127
7128 saved_syn = kmalloc(struct_size(saved_syn, data, len),
7129 GFP_ATOMIC);
7130 if (saved_syn) {
7131 saved_syn->mac_hdrlen = mac_hdrlen;
7132 saved_syn->network_hdrlen = skb_network_header_len(skb);
7133 saved_syn->tcp_hdrlen = tcp_hdrlen(skb);
7134 memcpy(saved_syn->data, base, len);
7135 req->saved_syn = saved_syn;
7136 }
7137 }
7138 }
7139
7140 /* If a SYN cookie is required and supported, returns a clamped MSS value to be
7141 * used for SYN cookie generation.
7142 */
7143 u16 tcp_get_syncookie_mss(struct request_sock_ops *rsk_ops,
7144 const struct tcp_request_sock_ops *af_ops,
7145 struct sock *sk, struct tcphdr *th)
7146 {
7147 struct tcp_sock *tp = tcp_sk(sk);
7148 u16 mss;
7149
7150 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_syncookies) != 2 &&
7151 !inet_csk_reqsk_queue_is_full(sk))
7152 return 0;
7153
7154 if (!tcp_syn_flood_action(sk, rsk_ops->slab_name))
7155 return 0;
7156
7157 if (sk_acceptq_is_full(sk)) {
7158 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
7159 return 0;
7160 }
7161
7162 mss = tcp_parse_mss_option(th, tp->rx_opt.user_mss);
7163 if (!mss)
7164 mss = af_ops->mss_clamp;
7165
7166 return mss;
7167 }
7168 EXPORT_SYMBOL_GPL(tcp_get_syncookie_mss);
7169
7170 int tcp_conn_request(struct request_sock_ops *rsk_ops,
7171 const struct tcp_request_sock_ops *af_ops,
7172 struct sock *sk, struct sk_buff *skb)
7173 {
7174 struct tcp_fastopen_cookie foc = { .len = -1 };
7175 struct tcp_options_received tmp_opt;
7176 struct tcp_sock *tp = tcp_sk(sk);
7177 struct net *net = sock_net(sk);
7178 struct sock *fastopen_sk = NULL;
7179 struct request_sock *req;
7180 bool want_cookie = false;
7181 struct dst_entry *dst;
7182 struct flowi fl;
7183 u8 syncookies;
7184 u32 isn;
7185
7186 #ifdef CONFIG_TCP_AO
7187 const struct tcp_ao_hdr *aoh;
7188 #endif
7189
7190 isn = __this_cpu_read(tcp_tw_isn);
7191 if (isn) {
7192 /* TW buckets are converted to open requests without
7193 * limitations, they conserve resources and peer is
7194 * evidently real one.
7195 */
7196 __this_cpu_write(tcp_tw_isn, 0);
7197 } else {
7198 syncookies = READ_ONCE(net->ipv4.sysctl_tcp_syncookies);
7199
7200 if (syncookies == 2 || inet_csk_reqsk_queue_is_full(sk)) {
7201 want_cookie = tcp_syn_flood_action(sk,
7202 rsk_ops->slab_name);
7203 if (!want_cookie)
7204 goto drop;
7205 }
7206 }
7207
7208 if (sk_acceptq_is_full(sk)) {
7209 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
7210 goto drop;
7211 }
7212
7213 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie);
7214 if (!req)
7215 goto drop;
7216
7217 req->syncookie = want_cookie;
7218 tcp_rsk(req)->af_specific = af_ops;
7219 tcp_rsk(req)->ts_off = 0;
7220 tcp_rsk(req)->req_usec_ts = false;
7221 #if IS_ENABLED(CONFIG_MPTCP)
7222 tcp_rsk(req)->is_mptcp = 0;
7223 #endif
7224
7225 tcp_clear_options(&tmp_opt);
7226 tmp_opt.mss_clamp = af_ops->mss_clamp;
7227 tmp_opt.user_mss = tp->rx_opt.user_mss;
7228 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0,
7229 want_cookie ? NULL : &foc);
7230
7231 if (want_cookie && !tmp_opt.saw_tstamp)
7232 tcp_clear_options(&tmp_opt);
7233
7234 if (IS_ENABLED(CONFIG_SMC) && want_cookie)
7235 tmp_opt.smc_ok = 0;
7236
7237 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
7238 tcp_openreq_init(req, &tmp_opt, skb, sk);
7239 inet_rsk(req)->no_srccheck = inet_test_bit(TRANSPARENT, sk);
7240
7241 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
7242 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb);
7243
7244 dst = af_ops->route_req(sk, skb, &fl, req, isn);
7245 if (!dst)
7246 goto drop_and_free;
7247
7248 if (tmp_opt.tstamp_ok) {
7249 tcp_rsk(req)->req_usec_ts = dst_tcp_usec_ts(dst);
7250 tcp_rsk(req)->ts_off = af_ops->init_ts_off(net, skb);
7251 }
7252 if (!want_cookie && !isn) {
7253 int max_syn_backlog = READ_ONCE(net->ipv4.sysctl_max_syn_backlog);
7254
7255 /* Kill the following clause, if you dislike this way. */
7256 if (!syncookies &&
7257 (max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
7258 (max_syn_backlog >> 2)) &&
7259 !tcp_peer_is_proven(req, dst)) {
7260 /* Without syncookies last quarter of
7261 * backlog is filled with destinations,
7262 * proven to be alive.
7263 * It means that we continue to communicate
7264 * to destinations, already remembered
7265 * to the moment of synflood.
7266 */
7267 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
7268 rsk_ops->family);
7269 goto drop_and_release;
7270 }
7271
7272 isn = af_ops->init_seq(skb);
7273 }
7274
7275 tcp_ecn_create_request(req, skb, sk, dst);
7276
7277 if (want_cookie) {
7278 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
7279 if (!tmp_opt.tstamp_ok)
7280 inet_rsk(req)->ecn_ok = 0;
7281 }
7282
7283 #ifdef CONFIG_TCP_AO
7284 if (tcp_parse_auth_options(tcp_hdr(skb), NULL, &aoh))
7285 goto drop_and_release; /* Invalid TCP options */
7286 if (aoh) {
7287 tcp_rsk(req)->used_tcp_ao = true;
7288 tcp_rsk(req)->ao_rcv_next = aoh->keyid;
7289 tcp_rsk(req)->ao_keyid = aoh->rnext_keyid;
7290
7291 } else {
7292 tcp_rsk(req)->used_tcp_ao = false;
7293 }
7294 #endif
7295 tcp_rsk(req)->snt_isn = isn;
7296 tcp_rsk(req)->txhash = net_tx_rndhash();
7297 tcp_rsk(req)->syn_tos = TCP_SKB_CB(skb)->ip_dsfield;
7298 tcp_openreq_init_rwin(req, sk, dst);
7299 sk_rx_queue_set(req_to_sk(req), skb);
7300 if (!want_cookie) {
7301 tcp_reqsk_record_syn(sk, req, skb);
7302 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst);
7303 }
7304 if (fastopen_sk) {
7305 af_ops->send_synack(fastopen_sk, dst, &fl, req,
7306 &foc, TCP_SYNACK_FASTOPEN, skb);
7307 /* Add the child socket directly into the accept queue */
7308 if (!inet_csk_reqsk_queue_add(sk, req, fastopen_sk)) {
7309 reqsk_fastopen_remove(fastopen_sk, req, false);
7310 bh_unlock_sock(fastopen_sk);
7311 sock_put(fastopen_sk);
7312 goto drop_and_free;
7313 }
7314 sk->sk_data_ready(sk);
7315 bh_unlock_sock(fastopen_sk);
7316 sock_put(fastopen_sk);
7317 } else {
7318 tcp_rsk(req)->tfo_listener = false;
7319 if (!want_cookie) {
7320 req->timeout = tcp_timeout_init((struct sock *)req);
7321 if (unlikely(!inet_csk_reqsk_queue_hash_add(sk, req,
7322 req->timeout))) {
7323 reqsk_free(req);
7324 return 0;
7325 }
7326
7327 }
7328 af_ops->send_synack(sk, dst, &fl, req, &foc,
7329 !want_cookie ? TCP_SYNACK_NORMAL :
7330 TCP_SYNACK_COOKIE,
7331 skb);
7332 if (want_cookie) {
7333 reqsk_free(req);
7334 return 0;
7335 }
7336 }
7337 reqsk_put(req);
7338 return 0;
7339
7340 drop_and_release:
7341 dst_release(dst);
7342 drop_and_free:
7343 __reqsk_free(req);
7344 drop:
7345 tcp_listendrop(sk);
7346 return 0;
7347 }
7348 EXPORT_SYMBOL(tcp_conn_request);
7349
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.