2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
23 * Pedro Roque : Fast Retransmit/Recovery.
25 * Retransmit queue handled by TCP.
26 * Better retransmit timer handling.
27 * New congestion avoidance.
31 * Eric : Fast Retransmit.
32 * Randy Scott : MSS option defines.
33 * Eric Schenk : Fixes to slow start algorithm.
34 * Eric Schenk : Yet another double ACK bug.
35 * Eric Schenk : Delayed ACK bug fixes.
36 * Eric Schenk : Floyd style fast retrans war avoidance.
37 * David S. Miller : Don't allow zero congestion window.
38 * Eric Schenk : Fix retransmitter so that it sends
39 * next packet on ack of previous packet.
40 * Andi Kleen : Moved open_request checking here
41 * and process RSTs for open_requests.
42 * Andi Kleen : Better prune_queue, and other fixes.
43 * Andrey Savochkin: Fix RTT measurements in the presence of
45 * Andrey Savochkin: Check sequence numbers correctly when
46 * removing SACKs due to in sequence incoming
48 * Andi Kleen: Make sure we never ack data there is not
49 * enough room for. Also make this condition
50 * a fatal error if it might still happen.
51 * Andi Kleen: Add tcp_measure_rcv_mss to make
52 * connections with MSS<min(MTU,ann. MSS)
53 * work without delayed acks.
54 * Andi Kleen: Process packets with PSH set in the
56 * J Hadi Salim: ECN support
59 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
60 * engine. Lots of bugs are found.
61 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
64 #define pr_fmt(fmt) "TCP: " fmt
67 #include <linux/slab.h>
68 #include <linux/module.h>
69 #include <linux/sysctl.h>
70 #include <linux/kernel.h>
71 #include <linux/prefetch.h>
74 #include <net/inet_common.h>
75 #include <linux/ipsec.h>
76 #include <asm/unaligned.h>
77 #include <linux/errqueue.h>
79 int sysctl_tcp_timestamps __read_mostly = 1;
80 int sysctl_tcp_window_scaling __read_mostly = 1;
81 int sysctl_tcp_sack __read_mostly = 1;
82 int sysctl_tcp_fack __read_mostly = 1;
83 int sysctl_tcp_max_reordering __read_mostly = 300;
84 int sysctl_tcp_dsack __read_mostly = 1;
85 int sysctl_tcp_app_win __read_mostly = 31;
86 int sysctl_tcp_adv_win_scale __read_mostly = 1;
87 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
89 /* rfc5961 challenge ack rate limiting */
90 int sysctl_tcp_challenge_ack_limit = 1000;
92 int sysctl_tcp_stdurg __read_mostly;
93 int sysctl_tcp_rfc1337 __read_mostly;
94 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
95 int sysctl_tcp_frto __read_mostly = 2;
96 int sysctl_tcp_min_rtt_wlen __read_mostly = 300;
98 int sysctl_tcp_thin_dupack __read_mostly;
100 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
101 int sysctl_tcp_early_retrans __read_mostly = 3;
102 int sysctl_tcp_invalid_ratelimit __read_mostly = HZ/2;
104 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
105 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
106 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
107 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
108 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
109 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
110 #define FLAG_ECE 0x40 /* ECE in this ACK */
111 #define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */
112 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
113 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
114 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
115 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
116 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
117 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
118 #define FLAG_NO_CHALLENGE_ACK 0x8000 /* do not call tcp_send_challenge_ack() */
120 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
121 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
122 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
123 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
125 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
126 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
128 #define REXMIT_NONE 0 /* no loss recovery to do */
129 #define REXMIT_LOST 1 /* retransmit packets marked lost */
130 #define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */
132 static void tcp_gro_dev_warn(struct sock *sk, const struct sk_buff *skb,
135 static bool __once __read_mostly;
138 struct net_device *dev;
143 dev = dev_get_by_index_rcu(sock_net(sk), skb->skb_iif);
144 if (!dev || len >= dev->mtu)
145 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n",
146 dev ? dev->name : "Unknown driver");
151 /* Adapt the MSS value used to make delayed ack decision to the
154 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
156 struct inet_connection_sock *icsk = inet_csk(sk);
157 const unsigned int lss = icsk->icsk_ack.last_seg_size;
160 icsk->icsk_ack.last_seg_size = 0;
162 /* skb->len may jitter because of SACKs, even if peer
163 * sends good full-sized frames.
165 len = skb_shinfo(skb)->gso_size ? : skb->len;
166 if (len >= icsk->icsk_ack.rcv_mss) {
167 icsk->icsk_ack.rcv_mss = min_t(unsigned int, len,
169 /* Account for possibly-removed options */
170 if (unlikely(len > icsk->icsk_ack.rcv_mss +
171 MAX_TCP_OPTION_SPACE))
172 tcp_gro_dev_warn(sk, skb, len);
174 /* Otherwise, we make more careful check taking into account,
175 * that SACKs block is variable.
177 * "len" is invariant segment length, including TCP header.
179 len += skb->data - skb_transport_header(skb);
180 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
181 /* If PSH is not set, packet should be
182 * full sized, provided peer TCP is not badly broken.
183 * This observation (if it is correct 8)) allows
184 * to handle super-low mtu links fairly.
186 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
187 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
188 /* Subtract also invariant (if peer is RFC compliant),
189 * tcp header plus fixed timestamp option length.
190 * Resulting "len" is MSS free of SACK jitter.
192 len -= tcp_sk(sk)->tcp_header_len;
193 icsk->icsk_ack.last_seg_size = len;
195 icsk->icsk_ack.rcv_mss = len;
199 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
200 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
201 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
205 static void tcp_incr_quickack(struct sock *sk, unsigned int max_quickacks)
207 struct inet_connection_sock *icsk = inet_csk(sk);
208 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
212 quickacks = min(quickacks, max_quickacks);
213 if (quickacks > icsk->icsk_ack.quick)
214 icsk->icsk_ack.quick = quickacks;
217 void tcp_enter_quickack_mode(struct sock *sk, unsigned int max_quickacks)
219 struct inet_connection_sock *icsk = inet_csk(sk);
221 tcp_incr_quickack(sk, max_quickacks);
222 icsk->icsk_ack.pingpong = 0;
223 icsk->icsk_ack.ato = TCP_ATO_MIN;
225 EXPORT_SYMBOL(tcp_enter_quickack_mode);
227 /* Send ACKs quickly, if "quick" count is not exhausted
228 * and the session is not interactive.
231 static bool tcp_in_quickack_mode(struct sock *sk)
233 const struct inet_connection_sock *icsk = inet_csk(sk);
234 const struct dst_entry *dst = __sk_dst_get(sk);
236 return (dst && dst_metric(dst, RTAX_QUICKACK)) ||
237 (icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong);
240 static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
242 if (tp->ecn_flags & TCP_ECN_OK)
243 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
246 static void tcp_ecn_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
248 if (tcp_hdr(skb)->cwr)
249 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
252 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
254 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
257 static void __tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
259 struct tcp_sock *tp = tcp_sk(sk);
261 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
262 case INET_ECN_NOT_ECT:
263 /* Funny extension: if ECT is not set on a segment,
264 * and we already seen ECT on a previous segment,
265 * it is probably a retransmit.
267 if (tp->ecn_flags & TCP_ECN_SEEN)
268 tcp_enter_quickack_mode(sk, 2);
271 if (tcp_ca_needs_ecn(sk))
272 tcp_ca_event(sk, CA_EVENT_ECN_IS_CE);
274 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
275 /* Better not delay acks, sender can have a very low cwnd */
276 tcp_enter_quickack_mode(sk, 2);
277 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
279 tp->ecn_flags |= TCP_ECN_SEEN;
282 if (tcp_ca_needs_ecn(sk))
283 tcp_ca_event(sk, CA_EVENT_ECN_NO_CE);
284 tp->ecn_flags |= TCP_ECN_SEEN;
289 static void tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
291 if (tcp_sk(sk)->ecn_flags & TCP_ECN_OK)
292 __tcp_ecn_check_ce(sk, skb);
295 static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
297 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
298 tp->ecn_flags &= ~TCP_ECN_OK;
301 static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
303 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
304 tp->ecn_flags &= ~TCP_ECN_OK;
307 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
309 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
314 /* Buffer size and advertised window tuning.
316 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
319 static void tcp_sndbuf_expand(struct sock *sk)
321 const struct tcp_sock *tp = tcp_sk(sk);
322 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
326 /* Worst case is non GSO/TSO : each frame consumes one skb
327 * and skb->head is kmalloced using power of two area of memory
329 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
331 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
333 per_mss = roundup_pow_of_two(per_mss) +
334 SKB_DATA_ALIGN(sizeof(struct sk_buff));
336 nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd);
337 nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
339 /* Fast Recovery (RFC 5681 3.2) :
340 * Cubic needs 1.7 factor, rounded to 2 to include
341 * extra cushion (application might react slowly to POLLOUT)
343 sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2;
344 sndmem *= nr_segs * per_mss;
346 if (sk->sk_sndbuf < sndmem)
347 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
350 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
352 * All tcp_full_space() is split to two parts: "network" buffer, allocated
353 * forward and advertised in receiver window (tp->rcv_wnd) and
354 * "application buffer", required to isolate scheduling/application
355 * latencies from network.
356 * window_clamp is maximal advertised window. It can be less than
357 * tcp_full_space(), in this case tcp_full_space() - window_clamp
358 * is reserved for "application" buffer. The less window_clamp is
359 * the smoother our behaviour from viewpoint of network, but the lower
360 * throughput and the higher sensitivity of the connection to losses. 8)
362 * rcv_ssthresh is more strict window_clamp used at "slow start"
363 * phase to predict further behaviour of this connection.
364 * It is used for two goals:
365 * - to enforce header prediction at sender, even when application
366 * requires some significant "application buffer". It is check #1.
367 * - to prevent pruning of receive queue because of misprediction
368 * of receiver window. Check #2.
370 * The scheme does not work when sender sends good segments opening
371 * window and then starts to feed us spaghetti. But it should work
372 * in common situations. Otherwise, we have to rely on queue collapsing.
375 /* Slow part of check#2. */
376 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
378 struct tcp_sock *tp = tcp_sk(sk);
380 int truesize = tcp_win_from_space(skb->truesize) >> 1;
381 int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
383 while (tp->rcv_ssthresh <= window) {
384 if (truesize <= skb->len)
385 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
393 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
395 struct tcp_sock *tp = tcp_sk(sk);
398 room = min_t(int, tp->window_clamp, tcp_space(sk)) - tp->rcv_ssthresh;
401 if (room > 0 && !tcp_under_memory_pressure(sk)) {
404 /* Check #2. Increase window, if skb with such overhead
405 * will fit to rcvbuf in future.
407 if (tcp_win_from_space(skb->truesize) <= skb->len)
408 incr = 2 * tp->advmss;
410 incr = __tcp_grow_window(sk, skb);
413 incr = max_t(int, incr, 2 * skb->len);
414 tp->rcv_ssthresh += min(room, incr);
415 inet_csk(sk)->icsk_ack.quick |= 1;
420 /* 3. Tuning rcvbuf, when connection enters established state. */
421 static void tcp_fixup_rcvbuf(struct sock *sk)
423 u32 mss = tcp_sk(sk)->advmss;
426 rcvmem = 2 * SKB_TRUESIZE(mss + MAX_TCP_HEADER) *
427 tcp_default_init_rwnd(mss);
429 /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
430 * Allow enough cushion so that sender is not limited by our window
432 if (sysctl_tcp_moderate_rcvbuf)
435 if (sk->sk_rcvbuf < rcvmem)
436 sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]);
439 /* 4. Try to fixup all. It is made immediately after connection enters
442 void tcp_init_buffer_space(struct sock *sk)
444 struct tcp_sock *tp = tcp_sk(sk);
447 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
448 tcp_fixup_rcvbuf(sk);
449 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
450 tcp_sndbuf_expand(sk);
452 tp->rcvq_space.space = tp->rcv_wnd;
453 tp->rcvq_space.time = tcp_time_stamp;
454 tp->rcvq_space.seq = tp->copied_seq;
456 maxwin = tcp_full_space(sk);
458 if (tp->window_clamp >= maxwin) {
459 tp->window_clamp = maxwin;
461 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
462 tp->window_clamp = max(maxwin -
463 (maxwin >> sysctl_tcp_app_win),
467 /* Force reservation of one segment. */
468 if (sysctl_tcp_app_win &&
469 tp->window_clamp > 2 * tp->advmss &&
470 tp->window_clamp + tp->advmss > maxwin)
471 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
473 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
474 tp->snd_cwnd_stamp = tcp_time_stamp;
477 /* 5. Recalculate window clamp after socket hit its memory bounds. */
478 static void tcp_clamp_window(struct sock *sk)
480 struct tcp_sock *tp = tcp_sk(sk);
481 struct inet_connection_sock *icsk = inet_csk(sk);
483 icsk->icsk_ack.quick = 0;
485 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
486 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
487 !tcp_under_memory_pressure(sk) &&
488 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
489 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
492 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
493 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
496 /* Initialize RCV_MSS value.
497 * RCV_MSS is an our guess about MSS used by the peer.
498 * We haven't any direct information about the MSS.
499 * It's better to underestimate the RCV_MSS rather than overestimate.
500 * Overestimations make us ACKing less frequently than needed.
501 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
503 void tcp_initialize_rcv_mss(struct sock *sk)
505 const struct tcp_sock *tp = tcp_sk(sk);
506 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
508 hint = min(hint, tp->rcv_wnd / 2);
509 hint = min(hint, TCP_MSS_DEFAULT);
510 hint = max(hint, TCP_MIN_MSS);
512 inet_csk(sk)->icsk_ack.rcv_mss = hint;
514 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
516 /* Receiver "autotuning" code.
518 * The algorithm for RTT estimation w/o timestamps is based on
519 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
520 * <http://public.lanl.gov/radiant/pubs.html#DRS>
522 * More detail on this code can be found at
523 * <http://staff.psc.edu/jheffner/>,
524 * though this reference is out of date. A new paper
527 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
529 u32 new_sample = tp->rcv_rtt_est.rtt;
535 if (new_sample != 0) {
536 /* If we sample in larger samples in the non-timestamp
537 * case, we could grossly overestimate the RTT especially
538 * with chatty applications or bulk transfer apps which
539 * are stalled on filesystem I/O.
541 * Also, since we are only going for a minimum in the
542 * non-timestamp case, we do not smooth things out
543 * else with timestamps disabled convergence takes too
547 m -= (new_sample >> 3);
555 /* No previous measure. */
559 if (tp->rcv_rtt_est.rtt != new_sample)
560 tp->rcv_rtt_est.rtt = new_sample;
563 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
565 if (tp->rcv_rtt_est.time == 0)
567 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
569 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rcv_rtt_est.time, 1);
572 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
573 tp->rcv_rtt_est.time = tcp_time_stamp;
576 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
577 const struct sk_buff *skb)
579 struct tcp_sock *tp = tcp_sk(sk);
580 if (tp->rx_opt.rcv_tsecr &&
581 (TCP_SKB_CB(skb)->end_seq -
582 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
583 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
587 * This function should be called every time data is copied to user space.
588 * It calculates the appropriate TCP receive buffer space.
590 void tcp_rcv_space_adjust(struct sock *sk)
592 struct tcp_sock *tp = tcp_sk(sk);
596 time = tcp_time_stamp - tp->rcvq_space.time;
597 if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
600 /* Number of bytes copied to user in last RTT */
601 copied = tp->copied_seq - tp->rcvq_space.seq;
602 if (copied <= tp->rcvq_space.space)
606 * copied = bytes received in previous RTT, our base window
607 * To cope with packet losses, we need a 2x factor
608 * To cope with slow start, and sender growing its cwin by 100 %
609 * every RTT, we need a 4x factor, because the ACK we are sending
610 * now is for the next RTT, not the current one :
611 * <prev RTT . ><current RTT .. ><next RTT .... >
614 if (sysctl_tcp_moderate_rcvbuf &&
615 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
619 /* minimal window to cope with packet losses, assuming
620 * steady state. Add some cushion because of small variations.
622 rcvwin = ((u64)copied << 1) + 16 * tp->advmss;
624 /* If rate increased by 25%,
625 * assume slow start, rcvwin = 3 * copied
626 * If rate increased by 50%,
627 * assume sender can use 2x growth, rcvwin = 4 * copied
630 tp->rcvq_space.space + (tp->rcvq_space.space >> 2)) {
632 tp->rcvq_space.space + (tp->rcvq_space.space >> 1))
635 rcvwin += (rcvwin >> 1);
638 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
639 while (tcp_win_from_space(rcvmem) < tp->advmss)
642 do_div(rcvwin, tp->advmss);
643 rcvbuf = min_t(u64, rcvwin * rcvmem, sysctl_tcp_rmem[2]);
644 if (rcvbuf > sk->sk_rcvbuf) {
645 sk->sk_rcvbuf = rcvbuf;
647 /* Make the window clamp follow along. */
648 tp->window_clamp = tcp_win_from_space(rcvbuf);
651 tp->rcvq_space.space = copied;
654 tp->rcvq_space.seq = tp->copied_seq;
655 tp->rcvq_space.time = tcp_time_stamp;
658 /* There is something which you must keep in mind when you analyze the
659 * behavior of the tp->ato delayed ack timeout interval. When a
660 * connection starts up, we want to ack as quickly as possible. The
661 * problem is that "good" TCP's do slow start at the beginning of data
662 * transmission. The means that until we send the first few ACK's the
663 * sender will sit on his end and only queue most of his data, because
664 * he can only send snd_cwnd unacked packets at any given time. For
665 * each ACK we send, he increments snd_cwnd and transmits more of his
668 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
670 struct tcp_sock *tp = tcp_sk(sk);
671 struct inet_connection_sock *icsk = inet_csk(sk);
674 inet_csk_schedule_ack(sk);
676 tcp_measure_rcv_mss(sk, skb);
678 tcp_rcv_rtt_measure(tp);
680 now = tcp_time_stamp;
682 if (!icsk->icsk_ack.ato) {
683 /* The _first_ data packet received, initialize
684 * delayed ACK engine.
686 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
687 icsk->icsk_ack.ato = TCP_ATO_MIN;
689 int m = now - icsk->icsk_ack.lrcvtime;
691 if (m <= TCP_ATO_MIN / 2) {
692 /* The fastest case is the first. */
693 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
694 } else if (m < icsk->icsk_ack.ato) {
695 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
696 if (icsk->icsk_ack.ato > icsk->icsk_rto)
697 icsk->icsk_ack.ato = icsk->icsk_rto;
698 } else if (m > icsk->icsk_rto) {
699 /* Too long gap. Apparently sender failed to
700 * restart window, so that we send ACKs quickly.
702 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
706 icsk->icsk_ack.lrcvtime = now;
708 tcp_ecn_check_ce(sk, skb);
711 tcp_grow_window(sk, skb);
714 /* Called to compute a smoothed rtt estimate. The data fed to this
715 * routine either comes from timestamps, or from segments that were
716 * known _not_ to have been retransmitted [see Karn/Partridge
717 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
718 * piece by Van Jacobson.
719 * NOTE: the next three routines used to be one big routine.
720 * To save cycles in the RFC 1323 implementation it was better to break
721 * it up into three procedures. -- erics
723 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
725 struct tcp_sock *tp = tcp_sk(sk);
726 long m = mrtt_us; /* RTT */
727 u32 srtt = tp->srtt_us;
729 /* The following amusing code comes from Jacobson's
730 * article in SIGCOMM '88. Note that rtt and mdev
731 * are scaled versions of rtt and mean deviation.
732 * This is designed to be as fast as possible
733 * m stands for "measurement".
735 * On a 1990 paper the rto value is changed to:
736 * RTO = rtt + 4 * mdev
738 * Funny. This algorithm seems to be very broken.
739 * These formulae increase RTO, when it should be decreased, increase
740 * too slowly, when it should be increased quickly, decrease too quickly
741 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
742 * does not matter how to _calculate_ it. Seems, it was trap
743 * that VJ failed to avoid. 8)
746 m -= (srtt >> 3); /* m is now error in rtt est */
747 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
749 m = -m; /* m is now abs(error) */
750 m -= (tp->mdev_us >> 2); /* similar update on mdev */
751 /* This is similar to one of Eifel findings.
752 * Eifel blocks mdev updates when rtt decreases.
753 * This solution is a bit different: we use finer gain
754 * for mdev in this case (alpha*beta).
755 * Like Eifel it also prevents growth of rto,
756 * but also it limits too fast rto decreases,
757 * happening in pure Eifel.
762 m -= (tp->mdev_us >> 2); /* similar update on mdev */
764 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
765 if (tp->mdev_us > tp->mdev_max_us) {
766 tp->mdev_max_us = tp->mdev_us;
767 if (tp->mdev_max_us > tp->rttvar_us)
768 tp->rttvar_us = tp->mdev_max_us;
770 if (after(tp->snd_una, tp->rtt_seq)) {
771 if (tp->mdev_max_us < tp->rttvar_us)
772 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
773 tp->rtt_seq = tp->snd_nxt;
774 tp->mdev_max_us = tcp_rto_min_us(sk);
777 /* no previous measure. */
778 srtt = m << 3; /* take the measured time to be rtt */
779 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
780 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
781 tp->mdev_max_us = tp->rttvar_us;
782 tp->rtt_seq = tp->snd_nxt;
784 tp->srtt_us = max(1U, srtt);
787 /* Set the sk_pacing_rate to allow proper sizing of TSO packets.
788 * Note: TCP stack does not yet implement pacing.
789 * FQ packet scheduler can be used to implement cheap but effective
790 * TCP pacing, to smooth the burst on large writes when packets
791 * in flight is significantly lower than cwnd (or rwin)
793 int sysctl_tcp_pacing_ss_ratio __read_mostly = 200;
794 int sysctl_tcp_pacing_ca_ratio __read_mostly = 120;
796 static void tcp_update_pacing_rate(struct sock *sk)
798 const struct tcp_sock *tp = tcp_sk(sk);
801 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
802 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
804 /* current rate is (cwnd * mss) / srtt
805 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
806 * In Congestion Avoidance phase, set it to 120 % the current rate.
808 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
809 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
810 * end of slow start and should slow down.
812 if (tp->snd_cwnd < tp->snd_ssthresh / 2)
813 rate *= sysctl_tcp_pacing_ss_ratio;
815 rate *= sysctl_tcp_pacing_ca_ratio;
817 rate *= max(tp->snd_cwnd, tp->packets_out);
819 if (likely(tp->srtt_us))
820 do_div(rate, tp->srtt_us);
822 /* ACCESS_ONCE() is needed because sch_fq fetches sk_pacing_rate
823 * without any lock. We want to make sure compiler wont store
824 * intermediate values in this location.
826 ACCESS_ONCE(sk->sk_pacing_rate) = min_t(u64, rate,
827 sk->sk_max_pacing_rate);
830 /* Calculate rto without backoff. This is the second half of Van Jacobson's
831 * routine referred to above.
833 static void tcp_set_rto(struct sock *sk)
835 const struct tcp_sock *tp = tcp_sk(sk);
836 /* Old crap is replaced with new one. 8)
839 * 1. If rtt variance happened to be less 50msec, it is hallucination.
840 * It cannot be less due to utterly erratic ACK generation made
841 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
842 * to do with delayed acks, because at cwnd>2 true delack timeout
843 * is invisible. Actually, Linux-2.4 also generates erratic
844 * ACKs in some circumstances.
846 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
848 /* 2. Fixups made earlier cannot be right.
849 * If we do not estimate RTO correctly without them,
850 * all the algo is pure shit and should be replaced
851 * with correct one. It is exactly, which we pretend to do.
854 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
855 * guarantees that rto is higher.
860 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
862 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
865 cwnd = TCP_INIT_CWND;
866 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
870 * Packet counting of FACK is based on in-order assumptions, therefore TCP
871 * disables it when reordering is detected
873 void tcp_disable_fack(struct tcp_sock *tp)
875 /* RFC3517 uses different metric in lost marker => reset on change */
877 tp->lost_skb_hint = NULL;
878 tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED;
881 /* Take a notice that peer is sending D-SACKs */
882 static void tcp_dsack_seen(struct tcp_sock *tp)
884 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
887 static void tcp_update_reordering(struct sock *sk, const int metric,
890 struct tcp_sock *tp = tcp_sk(sk);
891 if (metric > tp->reordering) {
894 tp->reordering = min(sysctl_tcp_max_reordering, metric);
896 /* This exciting event is worth to be remembered. 8) */
898 mib_idx = LINUX_MIB_TCPTSREORDER;
899 else if (tcp_is_reno(tp))
900 mib_idx = LINUX_MIB_TCPRENOREORDER;
901 else if (tcp_is_fack(tp))
902 mib_idx = LINUX_MIB_TCPFACKREORDER;
904 mib_idx = LINUX_MIB_TCPSACKREORDER;
906 NET_INC_STATS(sock_net(sk), mib_idx);
907 #if FASTRETRANS_DEBUG > 1
908 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
909 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
913 tp->undo_marker ? tp->undo_retrans : 0);
915 tcp_disable_fack(tp);
919 tcp_disable_early_retrans(tp);
923 /* This must be called before lost_out is incremented */
924 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
926 if ((!tp->retransmit_skb_hint && tp->retrans_out >= tp->lost_out) ||
927 (tp->retransmit_skb_hint &&
928 before(TCP_SKB_CB(skb)->seq,
929 TCP_SKB_CB(tp->retransmit_skb_hint)->seq)))
930 tp->retransmit_skb_hint = skb;
933 after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
934 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
937 /* Sum the number of packets on the wire we have marked as lost.
938 * There are two cases we care about here:
939 * a) Packet hasn't been marked lost (nor retransmitted),
940 * and this is the first loss.
941 * b) Packet has been marked both lost and retransmitted,
942 * and this means we think it was lost again.
944 static void tcp_sum_lost(struct tcp_sock *tp, struct sk_buff *skb)
946 __u8 sacked = TCP_SKB_CB(skb)->sacked;
948 if (!(sacked & TCPCB_LOST) ||
949 ((sacked & TCPCB_LOST) && (sacked & TCPCB_SACKED_RETRANS)))
950 tp->lost += tcp_skb_pcount(skb);
953 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
955 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
956 tcp_verify_retransmit_hint(tp, skb);
958 tp->lost_out += tcp_skb_pcount(skb);
959 tcp_sum_lost(tp, skb);
960 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
964 void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, struct sk_buff *skb)
966 tcp_verify_retransmit_hint(tp, skb);
968 tcp_sum_lost(tp, skb);
969 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
970 tp->lost_out += tcp_skb_pcount(skb);
971 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
975 /* This procedure tags the retransmission queue when SACKs arrive.
977 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
978 * Packets in queue with these bits set are counted in variables
979 * sacked_out, retrans_out and lost_out, correspondingly.
981 * Valid combinations are:
982 * Tag InFlight Description
983 * 0 1 - orig segment is in flight.
984 * S 0 - nothing flies, orig reached receiver.
985 * L 0 - nothing flies, orig lost by net.
986 * R 2 - both orig and retransmit are in flight.
987 * L|R 1 - orig is lost, retransmit is in flight.
988 * S|R 1 - orig reached receiver, retrans is still in flight.
989 * (L|S|R is logically valid, it could occur when L|R is sacked,
990 * but it is equivalent to plain S and code short-curcuits it to S.
991 * L|S is logically invalid, it would mean -1 packet in flight 8))
993 * These 6 states form finite state machine, controlled by the following events:
994 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
995 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
996 * 3. Loss detection event of two flavors:
997 * A. Scoreboard estimator decided the packet is lost.
998 * A'. Reno "three dupacks" marks head of queue lost.
999 * A''. Its FACK modification, head until snd.fack is lost.
1000 * B. SACK arrives sacking SND.NXT at the moment, when the
1001 * segment was retransmitted.
1002 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1004 * It is pleasant to note, that state diagram turns out to be commutative,
1005 * so that we are allowed not to be bothered by order of our actions,
1006 * when multiple events arrive simultaneously. (see the function below).
1008 * Reordering detection.
1009 * --------------------
1010 * Reordering metric is maximal distance, which a packet can be displaced
1011 * in packet stream. With SACKs we can estimate it:
1013 * 1. SACK fills old hole and the corresponding segment was not
1014 * ever retransmitted -> reordering. Alas, we cannot use it
1015 * when segment was retransmitted.
1016 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1017 * for retransmitted and already SACKed segment -> reordering..
1018 * Both of these heuristics are not used in Loss state, when we cannot
1019 * account for retransmits accurately.
1021 * SACK block validation.
1022 * ----------------------
1024 * SACK block range validation checks that the received SACK block fits to
1025 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1026 * Note that SND.UNA is not included to the range though being valid because
1027 * it means that the receiver is rather inconsistent with itself reporting
1028 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1029 * perfectly valid, however, in light of RFC2018 which explicitly states
1030 * that "SACK block MUST reflect the newest segment. Even if the newest
1031 * segment is going to be discarded ...", not that it looks very clever
1032 * in case of head skb. Due to potentional receiver driven attacks, we
1033 * choose to avoid immediate execution of a walk in write queue due to
1034 * reneging and defer head skb's loss recovery to standard loss recovery
1035 * procedure that will eventually trigger (nothing forbids us doing this).
1037 * Implements also blockage to start_seq wrap-around. Problem lies in the
1038 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1039 * there's no guarantee that it will be before snd_nxt (n). The problem
1040 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1043 * <- outs wnd -> <- wrapzone ->
1044 * u e n u_w e_w s n_w
1046 * |<------------+------+----- TCP seqno space --------------+---------->|
1047 * ...-- <2^31 ->| |<--------...
1048 * ...---- >2^31 ------>| |<--------...
1050 * Current code wouldn't be vulnerable but it's better still to discard such
1051 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1052 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1053 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1054 * equal to the ideal case (infinite seqno space without wrap caused issues).
1056 * With D-SACK the lower bound is extended to cover sequence space below
1057 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1058 * again, D-SACK block must not to go across snd_una (for the same reason as
1059 * for the normal SACK blocks, explained above). But there all simplicity
1060 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1061 * fully below undo_marker they do not affect behavior in anyway and can
1062 * therefore be safely ignored. In rare cases (which are more or less
1063 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1064 * fragmentation and packet reordering past skb's retransmission. To consider
1065 * them correctly, the acceptable range must be extended even more though
1066 * the exact amount is rather hard to quantify. However, tp->max_window can
1067 * be used as an exaggerated estimate.
1069 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1070 u32 start_seq, u32 end_seq)
1072 /* Too far in future, or reversed (interpretation is ambiguous) */
1073 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1076 /* Nasty start_seq wrap-around check (see comments above) */
1077 if (!before(start_seq, tp->snd_nxt))
1080 /* In outstanding window? ...This is valid exit for D-SACKs too.
1081 * start_seq == snd_una is non-sensical (see comments above)
1083 if (after(start_seq, tp->snd_una))
1086 if (!is_dsack || !tp->undo_marker)
1089 /* ...Then it's D-SACK, and must reside below snd_una completely */
1090 if (after(end_seq, tp->snd_una))
1093 if (!before(start_seq, tp->undo_marker))
1097 if (!after(end_seq, tp->undo_marker))
1100 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1101 * start_seq < undo_marker and end_seq >= undo_marker.
1103 return !before(start_seq, end_seq - tp->max_window);
1106 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1107 struct tcp_sack_block_wire *sp, int num_sacks,
1110 struct tcp_sock *tp = tcp_sk(sk);
1111 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1112 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1113 bool dup_sack = false;
1115 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1118 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1119 } else if (num_sacks > 1) {
1120 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1121 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1123 if (!after(end_seq_0, end_seq_1) &&
1124 !before(start_seq_0, start_seq_1)) {
1127 NET_INC_STATS(sock_net(sk),
1128 LINUX_MIB_TCPDSACKOFORECV);
1132 /* D-SACK for already forgotten data... Do dumb counting. */
1133 if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 &&
1134 !after(end_seq_0, prior_snd_una) &&
1135 after(end_seq_0, tp->undo_marker))
1141 struct tcp_sacktag_state {
1144 /* Timestamps for earliest and latest never-retransmitted segment
1145 * that was SACKed. RTO needs the earliest RTT to stay conservative,
1146 * but congestion control should still get an accurate delay signal.
1148 struct skb_mstamp first_sackt;
1149 struct skb_mstamp last_sackt;
1150 struct rate_sample *rate;
1154 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1155 * the incoming SACK may not exactly match but we can find smaller MSS
1156 * aligned portion of it that matches. Therefore we might need to fragment
1157 * which may fail and creates some hassle (caller must handle error case
1160 * FIXME: this could be merged to shift decision code
1162 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1163 u32 start_seq, u32 end_seq)
1167 unsigned int pkt_len;
1170 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1171 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1173 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1174 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1175 mss = tcp_skb_mss(skb);
1176 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1179 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1183 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1188 /* Round if necessary so that SACKs cover only full MSSes
1189 * and/or the remaining small portion (if present)
1191 if (pkt_len > mss) {
1192 unsigned int new_len = (pkt_len / mss) * mss;
1193 if (!in_sack && new_len < pkt_len)
1198 if (pkt_len >= skb->len && !in_sack)
1201 err = tcp_fragment(sk, skb, pkt_len, mss, GFP_ATOMIC);
1209 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1210 static u8 tcp_sacktag_one(struct sock *sk,
1211 struct tcp_sacktag_state *state, u8 sacked,
1212 u32 start_seq, u32 end_seq,
1213 int dup_sack, int pcount,
1214 const struct skb_mstamp *xmit_time)
1216 struct tcp_sock *tp = tcp_sk(sk);
1217 int fack_count = state->fack_count;
1219 /* Account D-SACK for retransmitted packet. */
1220 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1221 if (tp->undo_marker && tp->undo_retrans > 0 &&
1222 after(end_seq, tp->undo_marker))
1223 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - pcount);
1224 if (sacked & TCPCB_SACKED_ACKED)
1225 state->reord = min(fack_count, state->reord);
1228 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1229 if (!after(end_seq, tp->snd_una))
1232 if (!(sacked & TCPCB_SACKED_ACKED)) {
1233 tcp_rack_advance(tp, xmit_time, sacked);
1235 if (sacked & TCPCB_SACKED_RETRANS) {
1236 /* If the segment is not tagged as lost,
1237 * we do not clear RETRANS, believing
1238 * that retransmission is still in flight.
1240 if (sacked & TCPCB_LOST) {
1241 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1242 tp->lost_out -= pcount;
1243 tp->retrans_out -= pcount;
1246 if (!(sacked & TCPCB_RETRANS)) {
1247 /* New sack for not retransmitted frame,
1248 * which was in hole. It is reordering.
1250 if (before(start_seq,
1251 tcp_highest_sack_seq(tp)))
1252 state->reord = min(fack_count,
1254 if (!after(end_seq, tp->high_seq))
1255 state->flag |= FLAG_ORIG_SACK_ACKED;
1256 if (state->first_sackt.v64 == 0)
1257 state->first_sackt = *xmit_time;
1258 state->last_sackt = *xmit_time;
1261 if (sacked & TCPCB_LOST) {
1262 sacked &= ~TCPCB_LOST;
1263 tp->lost_out -= pcount;
1267 sacked |= TCPCB_SACKED_ACKED;
1268 state->flag |= FLAG_DATA_SACKED;
1269 tp->sacked_out += pcount;
1270 tp->delivered += pcount; /* Out-of-order packets delivered */
1272 fack_count += pcount;
1274 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1275 if (!tcp_is_fack(tp) && tp->lost_skb_hint &&
1276 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1277 tp->lost_cnt_hint += pcount;
1279 if (fack_count > tp->fackets_out)
1280 tp->fackets_out = fack_count;
1283 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1284 * frames and clear it. undo_retrans is decreased above, L|R frames
1285 * are accounted above as well.
1287 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1288 sacked &= ~TCPCB_SACKED_RETRANS;
1289 tp->retrans_out -= pcount;
1295 /* Shift newly-SACKed bytes from this skb to the immediately previous
1296 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1298 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1299 struct tcp_sacktag_state *state,
1300 unsigned int pcount, int shifted, int mss,
1303 struct tcp_sock *tp = tcp_sk(sk);
1304 struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1305 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1306 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1310 /* Adjust counters and hints for the newly sacked sequence
1311 * range but discard the return value since prev is already
1312 * marked. We must tag the range first because the seq
1313 * advancement below implicitly advances
1314 * tcp_highest_sack_seq() when skb is highest_sack.
1316 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1317 start_seq, end_seq, dup_sack, pcount,
1319 tcp_rate_skb_delivered(sk, skb, state->rate);
1321 if (skb == tp->lost_skb_hint)
1322 tp->lost_cnt_hint += pcount;
1324 TCP_SKB_CB(prev)->end_seq += shifted;
1325 TCP_SKB_CB(skb)->seq += shifted;
1327 tcp_skb_pcount_add(prev, pcount);
1328 WARN_ON_ONCE(tcp_skb_pcount(skb) < pcount);
1329 tcp_skb_pcount_add(skb, -pcount);
1331 /* When we're adding to gso_segs == 1, gso_size will be zero,
1332 * in theory this shouldn't be necessary but as long as DSACK
1333 * code can come after this skb later on it's better to keep
1334 * setting gso_size to something.
1336 if (!TCP_SKB_CB(prev)->tcp_gso_size)
1337 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1339 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1340 if (tcp_skb_pcount(skb) <= 1)
1341 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1343 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1344 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1347 BUG_ON(!tcp_skb_pcount(skb));
1348 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1352 /* Whole SKB was eaten :-) */
1354 if (skb == tp->retransmit_skb_hint)
1355 tp->retransmit_skb_hint = prev;
1356 if (skb == tp->lost_skb_hint) {
1357 tp->lost_skb_hint = prev;
1358 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1361 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1362 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor;
1363 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1364 TCP_SKB_CB(prev)->end_seq++;
1366 if (skb == tcp_highest_sack(sk))
1367 tcp_advance_highest_sack(sk, skb);
1369 tcp_skb_collapse_tstamp(prev, skb);
1370 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp.v64))
1371 TCP_SKB_CB(prev)->tx.delivered_mstamp.v64 = 0;
1373 tcp_unlink_write_queue(skb, sk);
1374 sk_wmem_free_skb(sk, skb);
1376 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED);
1381 /* I wish gso_size would have a bit more sane initialization than
1382 * something-or-zero which complicates things
1384 static int tcp_skb_seglen(const struct sk_buff *skb)
1386 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1389 /* Shifting pages past head area doesn't work */
1390 static int skb_can_shift(const struct sk_buff *skb)
1392 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1395 int tcp_skb_shift(struct sk_buff *to, struct sk_buff *from,
1396 int pcount, int shiftlen)
1398 /* TCP min gso_size is 8 bytes (TCP_MIN_GSO_SIZE)
1399 * Since TCP_SKB_CB(skb)->tcp_gso_segs is 16 bits, we need
1400 * to make sure not storing more than 65535 * 8 bytes per skb,
1401 * even if current MSS is bigger.
1403 if (unlikely(to->len + shiftlen >= 65535 * TCP_MIN_GSO_SIZE))
1405 if (unlikely(tcp_skb_pcount(to) + pcount > 65535))
1407 return skb_shift(to, from, shiftlen);
1410 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1413 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1414 struct tcp_sacktag_state *state,
1415 u32 start_seq, u32 end_seq,
1418 struct tcp_sock *tp = tcp_sk(sk);
1419 struct sk_buff *prev;
1426 if (!sk_can_gso(sk))
1429 /* Normally R but no L won't result in plain S */
1431 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1433 if (!skb_can_shift(skb))
1435 /* This frame is about to be dropped (was ACKed). */
1436 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1439 /* Can only happen with delayed DSACK + discard craziness */
1440 if (unlikely(skb == tcp_write_queue_head(sk)))
1442 prev = tcp_write_queue_prev(sk, skb);
1444 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1447 if (!tcp_skb_can_collapse_to(prev))
1450 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1451 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1455 pcount = tcp_skb_pcount(skb);
1456 mss = tcp_skb_seglen(skb);
1458 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1459 * drop this restriction as unnecessary
1461 if (mss != tcp_skb_seglen(prev))
1464 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1466 /* CHECKME: This is non-MSS split case only?, this will
1467 * cause skipped skbs due to advancing loop btw, original
1468 * has that feature too
1470 if (tcp_skb_pcount(skb) <= 1)
1473 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1475 /* TODO: head merge to next could be attempted here
1476 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1477 * though it might not be worth of the additional hassle
1479 * ...we can probably just fallback to what was done
1480 * previously. We could try merging non-SACKed ones
1481 * as well but it probably isn't going to buy off
1482 * because later SACKs might again split them, and
1483 * it would make skb timestamp tracking considerably
1489 len = end_seq - TCP_SKB_CB(skb)->seq;
1491 BUG_ON(len > skb->len);
1493 /* MSS boundaries should be honoured or else pcount will
1494 * severely break even though it makes things bit trickier.
1495 * Optimize common case to avoid most of the divides
1497 mss = tcp_skb_mss(skb);
1499 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1500 * drop this restriction as unnecessary
1502 if (mss != tcp_skb_seglen(prev))
1507 } else if (len < mss) {
1515 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1516 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1519 if (!tcp_skb_shift(prev, skb, pcount, len))
1521 if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1524 /* Hole filled allows collapsing with the next as well, this is very
1525 * useful when hole on every nth skb pattern happens
1527 if (prev == tcp_write_queue_tail(sk))
1529 skb = tcp_write_queue_next(sk, prev);
1531 if (!skb_can_shift(skb) ||
1532 (skb == tcp_send_head(sk)) ||
1533 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1534 (mss != tcp_skb_seglen(skb)))
1538 next_pcount = tcp_skb_pcount(skb);
1539 if (tcp_skb_shift(prev, skb, next_pcount, len)) {
1540 pcount += next_pcount;
1541 tcp_shifted_skb(sk, skb, state, next_pcount, len, mss, 0);
1544 state->fack_count += pcount;
1551 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1555 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1556 struct tcp_sack_block *next_dup,
1557 struct tcp_sacktag_state *state,
1558 u32 start_seq, u32 end_seq,
1561 struct tcp_sock *tp = tcp_sk(sk);
1562 struct sk_buff *tmp;
1564 tcp_for_write_queue_from(skb, sk) {
1566 bool dup_sack = dup_sack_in;
1568 if (skb == tcp_send_head(sk))
1571 /* queue is in-order => we can short-circuit the walk early */
1572 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1576 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1577 in_sack = tcp_match_skb_to_sack(sk, skb,
1578 next_dup->start_seq,
1584 /* skb reference here is a bit tricky to get right, since
1585 * shifting can eat and free both this skb and the next,
1586 * so not even _safe variant of the loop is enough.
1589 tmp = tcp_shift_skb_data(sk, skb, state,
1590 start_seq, end_seq, dup_sack);
1599 in_sack = tcp_match_skb_to_sack(sk, skb,
1605 if (unlikely(in_sack < 0))
1609 TCP_SKB_CB(skb)->sacked =
1612 TCP_SKB_CB(skb)->sacked,
1613 TCP_SKB_CB(skb)->seq,
1614 TCP_SKB_CB(skb)->end_seq,
1616 tcp_skb_pcount(skb),
1618 tcp_rate_skb_delivered(sk, skb, state->rate);
1620 if (!before(TCP_SKB_CB(skb)->seq,
1621 tcp_highest_sack_seq(tp)))
1622 tcp_advance_highest_sack(sk, skb);
1625 state->fack_count += tcp_skb_pcount(skb);
1630 /* Avoid all extra work that is being done by sacktag while walking in
1633 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1634 struct tcp_sacktag_state *state,
1637 tcp_for_write_queue_from(skb, sk) {
1638 if (skb == tcp_send_head(sk))
1641 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1644 state->fack_count += tcp_skb_pcount(skb);
1649 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1651 struct tcp_sack_block *next_dup,
1652 struct tcp_sacktag_state *state,
1658 if (before(next_dup->start_seq, skip_to_seq)) {
1659 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1660 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1661 next_dup->start_seq, next_dup->end_seq,
1668 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1670 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1674 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1675 u32 prior_snd_una, struct tcp_sacktag_state *state)
1677 struct tcp_sock *tp = tcp_sk(sk);
1678 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1679 TCP_SKB_CB(ack_skb)->sacked);
1680 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1681 struct tcp_sack_block sp[TCP_NUM_SACKS];
1682 struct tcp_sack_block *cache;
1683 struct sk_buff *skb;
1684 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1686 bool found_dup_sack = false;
1688 int first_sack_index;
1691 state->reord = tp->packets_out;
1693 if (!tp->sacked_out) {
1694 if (WARN_ON(tp->fackets_out))
1695 tp->fackets_out = 0;
1696 tcp_highest_sack_reset(sk);
1699 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1700 num_sacks, prior_snd_una);
1701 if (found_dup_sack) {
1702 state->flag |= FLAG_DSACKING_ACK;
1703 tp->delivered++; /* A spurious retransmission is delivered */
1706 /* Eliminate too old ACKs, but take into
1707 * account more or less fresh ones, they can
1708 * contain valid SACK info.
1710 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1713 if (!tp->packets_out)
1717 first_sack_index = 0;
1718 for (i = 0; i < num_sacks; i++) {
1719 bool dup_sack = !i && found_dup_sack;
1721 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1722 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1724 if (!tcp_is_sackblock_valid(tp, dup_sack,
1725 sp[used_sacks].start_seq,
1726 sp[used_sacks].end_seq)) {
1730 if (!tp->undo_marker)
1731 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1733 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1735 /* Don't count olds caused by ACK reordering */
1736 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1737 !after(sp[used_sacks].end_seq, tp->snd_una))
1739 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1742 NET_INC_STATS(sock_net(sk), mib_idx);
1744 first_sack_index = -1;
1748 /* Ignore very old stuff early */
1749 if (!after(sp[used_sacks].end_seq, prior_snd_una)) {
1751 first_sack_index = -1;
1758 /* order SACK blocks to allow in order walk of the retrans queue */
1759 for (i = used_sacks - 1; i > 0; i--) {
1760 for (j = 0; j < i; j++) {
1761 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1762 swap(sp[j], sp[j + 1]);
1764 /* Track where the first SACK block goes to */
1765 if (j == first_sack_index)
1766 first_sack_index = j + 1;
1771 skb = tcp_write_queue_head(sk);
1772 state->fack_count = 0;
1775 if (!tp->sacked_out) {
1776 /* It's already past, so skip checking against it */
1777 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1779 cache = tp->recv_sack_cache;
1780 /* Skip empty blocks in at head of the cache */
1781 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1786 while (i < used_sacks) {
1787 u32 start_seq = sp[i].start_seq;
1788 u32 end_seq = sp[i].end_seq;
1789 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1790 struct tcp_sack_block *next_dup = NULL;
1792 if (found_dup_sack && ((i + 1) == first_sack_index))
1793 next_dup = &sp[i + 1];
1795 /* Skip too early cached blocks */
1796 while (tcp_sack_cache_ok(tp, cache) &&
1797 !before(start_seq, cache->end_seq))
1800 /* Can skip some work by looking recv_sack_cache? */
1801 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1802 after(end_seq, cache->start_seq)) {
1805 if (before(start_seq, cache->start_seq)) {
1806 skb = tcp_sacktag_skip(skb, sk, state,
1808 skb = tcp_sacktag_walk(skb, sk, next_dup,
1815 /* Rest of the block already fully processed? */
1816 if (!after(end_seq, cache->end_seq))
1819 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1823 /* ...tail remains todo... */
1824 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1825 /* ...but better entrypoint exists! */
1826 skb = tcp_highest_sack(sk);
1829 state->fack_count = tp->fackets_out;
1834 skb = tcp_sacktag_skip(skb, sk, state, cache->end_seq);
1835 /* Check overlap against next cached too (past this one already) */
1840 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1841 skb = tcp_highest_sack(sk);
1844 state->fack_count = tp->fackets_out;
1846 skb = tcp_sacktag_skip(skb, sk, state, start_seq);
1849 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
1850 start_seq, end_seq, dup_sack);
1856 /* Clear the head of the cache sack blocks so we can skip it next time */
1857 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1858 tp->recv_sack_cache[i].start_seq = 0;
1859 tp->recv_sack_cache[i].end_seq = 0;
1861 for (j = 0; j < used_sacks; j++)
1862 tp->recv_sack_cache[i++] = sp[j];
1864 if ((state->reord < tp->fackets_out) &&
1865 ((inet_csk(sk)->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker))
1866 tcp_update_reordering(sk, tp->fackets_out - state->reord, 0);
1868 tcp_verify_left_out(tp);
1871 #if FASTRETRANS_DEBUG > 0
1872 WARN_ON((int)tp->sacked_out < 0);
1873 WARN_ON((int)tp->lost_out < 0);
1874 WARN_ON((int)tp->retrans_out < 0);
1875 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1880 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1881 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1883 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1887 holes = max(tp->lost_out, 1U);
1888 holes = min(holes, tp->packets_out);
1890 if ((tp->sacked_out + holes) > tp->packets_out) {
1891 tp->sacked_out = tp->packets_out - holes;
1897 /* If we receive more dupacks than we expected counting segments
1898 * in assumption of absent reordering, interpret this as reordering.
1899 * The only another reason could be bug in receiver TCP.
1901 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1903 struct tcp_sock *tp = tcp_sk(sk);
1904 if (tcp_limit_reno_sacked(tp))
1905 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1908 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1910 static void tcp_add_reno_sack(struct sock *sk)
1912 struct tcp_sock *tp = tcp_sk(sk);
1913 u32 prior_sacked = tp->sacked_out;
1916 tcp_check_reno_reordering(sk, 0);
1917 if (tp->sacked_out > prior_sacked)
1918 tp->delivered++; /* Some out-of-order packet is delivered */
1919 tcp_verify_left_out(tp);
1922 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1924 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1926 struct tcp_sock *tp = tcp_sk(sk);
1929 /* One ACK acked hole. The rest eat duplicate ACKs. */
1930 tp->delivered += max_t(int, acked - tp->sacked_out, 1);
1931 if (acked - 1 >= tp->sacked_out)
1934 tp->sacked_out -= acked - 1;
1936 tcp_check_reno_reordering(sk, acked);
1937 tcp_verify_left_out(tp);
1940 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1945 void tcp_clear_retrans(struct tcp_sock *tp)
1947 tp->retrans_out = 0;
1949 tp->undo_marker = 0;
1950 tp->undo_retrans = -1;
1951 tp->fackets_out = 0;
1955 static inline void tcp_init_undo(struct tcp_sock *tp)
1957 tp->undo_marker = tp->snd_una;
1958 /* Retransmission still in flight may cause DSACKs later. */
1959 tp->undo_retrans = tp->retrans_out ? : -1;
1962 /* Enter Loss state. If we detect SACK reneging, forget all SACK information
1963 * and reset tags completely, otherwise preserve SACKs. If receiver
1964 * dropped its ofo queue, we will know this due to reneging detection.
1966 void tcp_enter_loss(struct sock *sk)
1968 const struct inet_connection_sock *icsk = inet_csk(sk);
1969 struct tcp_sock *tp = tcp_sk(sk);
1970 struct net *net = sock_net(sk);
1971 struct sk_buff *skb;
1972 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
1973 bool is_reneg; /* is receiver reneging on SACKs? */
1976 /* Reduce ssthresh if it has not yet been made inside this window. */
1977 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1978 !after(tp->high_seq, tp->snd_una) ||
1979 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1980 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1981 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1982 tcp_ca_event(sk, CA_EVENT_LOSS);
1986 tp->snd_cwnd_cnt = 0;
1987 tp->snd_cwnd_stamp = tcp_time_stamp;
1989 tp->retrans_out = 0;
1992 if (tcp_is_reno(tp))
1993 tcp_reset_reno_sack(tp);
1995 skb = tcp_write_queue_head(sk);
1996 is_reneg = skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED);
1998 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
2000 tp->fackets_out = 0;
2001 /* Mark SACK reneging until we recover from this loss event. */
2002 tp->is_sack_reneg = 1;
2004 tcp_clear_all_retrans_hints(tp);
2006 tcp_for_write_queue(skb, sk) {
2007 if (skb == tcp_send_head(sk))
2010 mark_lost = (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2013 tcp_sum_lost(tp, skb);
2014 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
2016 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
2017 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2018 tp->lost_out += tcp_skb_pcount(skb);
2019 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
2022 tcp_verify_left_out(tp);
2024 /* Timeout in disordered state after receiving substantial DUPACKs
2025 * suggests that the degree of reordering is over-estimated.
2027 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
2028 tp->sacked_out >= net->ipv4.sysctl_tcp_reordering)
2029 tp->reordering = min_t(unsigned int, tp->reordering,
2030 net->ipv4.sysctl_tcp_reordering);
2031 tcp_set_ca_state(sk, TCP_CA_Loss);
2032 tp->high_seq = tp->snd_nxt;
2033 tcp_ecn_queue_cwr(tp);
2035 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
2036 * loss recovery is underway except recurring timeout(s) on
2037 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
2039 tp->frto = sysctl_tcp_frto &&
2040 (new_recovery || icsk->icsk_retransmits) &&
2041 !inet_csk(sk)->icsk_mtup.probe_size;
2044 /* If ACK arrived pointing to a remembered SACK, it means that our
2045 * remembered SACKs do not reflect real state of receiver i.e.
2046 * receiver _host_ is heavily congested (or buggy).
2048 * To avoid big spurious retransmission bursts due to transient SACK
2049 * scoreboard oddities that look like reneging, we give the receiver a
2050 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2051 * restore sanity to the SACK scoreboard. If the apparent reneging
2052 * persists until this RTO then we'll clear the SACK scoreboard.
2054 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
2056 if (flag & FLAG_SACK_RENEGING) {
2057 struct tcp_sock *tp = tcp_sk(sk);
2058 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2059 msecs_to_jiffies(10));
2061 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2062 delay, TCP_RTO_MAX);
2068 static inline int tcp_fackets_out(const struct tcp_sock *tp)
2070 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
2073 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2074 * counter when SACK is enabled (without SACK, sacked_out is used for
2077 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2078 * segments up to the highest received SACK block so far and holes in
2081 * With reordering, holes may still be in flight, so RFC3517 recovery
2082 * uses pure sacked_out (total number of SACKed segments) even though
2083 * it violates the RFC that uses duplicate ACKs, often these are equal
2084 * but when e.g. out-of-window ACKs or packet duplication occurs,
2085 * they differ. Since neither occurs due to loss, TCP should really
2088 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2090 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2093 static bool tcp_pause_early_retransmit(struct sock *sk, int flag)
2095 struct tcp_sock *tp = tcp_sk(sk);
2096 unsigned long delay;
2098 /* Delay early retransmit and entering fast recovery for
2099 * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
2100 * available, or RTO is scheduled to fire first.
2102 if (sysctl_tcp_early_retrans < 2 || sysctl_tcp_early_retrans > 3 ||
2103 (flag & FLAG_ECE) || !tp->srtt_us)
2106 delay = max(usecs_to_jiffies(tp->srtt_us >> 5),
2107 msecs_to_jiffies(2));
2109 if (!time_after(inet_csk(sk)->icsk_timeout, (jiffies + delay)))
2112 inet_csk_reset_xmit_timer(sk, ICSK_TIME_EARLY_RETRANS, delay,
2117 /* Linux NewReno/SACK/FACK/ECN state machine.
2118 * --------------------------------------
2120 * "Open" Normal state, no dubious events, fast path.
2121 * "Disorder" In all the respects it is "Open",
2122 * but requires a bit more attention. It is entered when
2123 * we see some SACKs or dupacks. It is split of "Open"
2124 * mainly to move some processing from fast path to slow one.
2125 * "CWR" CWND was reduced due to some Congestion Notification event.
2126 * It can be ECN, ICMP source quench, local device congestion.
2127 * "Recovery" CWND was reduced, we are fast-retransmitting.
2128 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2130 * tcp_fastretrans_alert() is entered:
2131 * - each incoming ACK, if state is not "Open"
2132 * - when arrived ACK is unusual, namely:
2137 * Counting packets in flight is pretty simple.
2139 * in_flight = packets_out - left_out + retrans_out
2141 * packets_out is SND.NXT-SND.UNA counted in packets.
2143 * retrans_out is number of retransmitted segments.
2145 * left_out is number of segments left network, but not ACKed yet.
2147 * left_out = sacked_out + lost_out
2149 * sacked_out: Packets, which arrived to receiver out of order
2150 * and hence not ACKed. With SACKs this number is simply
2151 * amount of SACKed data. Even without SACKs
2152 * it is easy to give pretty reliable estimate of this number,
2153 * counting duplicate ACKs.
2155 * lost_out: Packets lost by network. TCP has no explicit
2156 * "loss notification" feedback from network (for now).
2157 * It means that this number can be only _guessed_.
2158 * Actually, it is the heuristics to predict lossage that
2159 * distinguishes different algorithms.
2161 * F.e. after RTO, when all the queue is considered as lost,
2162 * lost_out = packets_out and in_flight = retrans_out.
2164 * Essentially, we have now two algorithms counting
2167 * FACK: It is the simplest heuristics. As soon as we decided
2168 * that something is lost, we decide that _all_ not SACKed
2169 * packets until the most forward SACK are lost. I.e.
2170 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2171 * It is absolutely correct estimate, if network does not reorder
2172 * packets. And it loses any connection to reality when reordering
2173 * takes place. We use FACK by default until reordering
2174 * is suspected on the path to this destination.
2176 * NewReno: when Recovery is entered, we assume that one segment
2177 * is lost (classic Reno). While we are in Recovery and
2178 * a partial ACK arrives, we assume that one more packet
2179 * is lost (NewReno). This heuristics are the same in NewReno
2182 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2183 * deflation etc. CWND is real congestion window, never inflated, changes
2184 * only according to classic VJ rules.
2186 * Really tricky (and requiring careful tuning) part of algorithm
2187 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2188 * The first determines the moment _when_ we should reduce CWND and,
2189 * hence, slow down forward transmission. In fact, it determines the moment
2190 * when we decide that hole is caused by loss, rather than by a reorder.
2192 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2193 * holes, caused by lost packets.
2195 * And the most logically complicated part of algorithm is undo
2196 * heuristics. We detect false retransmits due to both too early
2197 * fast retransmit (reordering) and underestimated RTO, analyzing
2198 * timestamps and D-SACKs. When we detect that some segments were
2199 * retransmitted by mistake and CWND reduction was wrong, we undo
2200 * window reduction and abort recovery phase. This logic is hidden
2201 * inside several functions named tcp_try_undo_<something>.
2204 /* This function decides, when we should leave Disordered state
2205 * and enter Recovery phase, reducing congestion window.
2207 * Main question: may we further continue forward transmission
2208 * with the same cwnd?
2210 static bool tcp_time_to_recover(struct sock *sk, int flag)
2212 struct tcp_sock *tp = tcp_sk(sk);
2214 int tcp_reordering = sock_net(sk)->ipv4.sysctl_tcp_reordering;
2216 /* Trick#1: The loss is proven. */
2220 /* Not-A-Trick#2 : Classic rule... */
2221 if (tcp_dupack_heuristics(tp) > tp->reordering)
2224 /* Trick#4: It is still not OK... But will it be useful to delay
2227 packets_out = tp->packets_out;
2228 if (packets_out <= tp->reordering &&
2229 tp->sacked_out >= max_t(__u32, packets_out/2, tcp_reordering) &&
2230 !tcp_may_send_now(sk)) {
2231 /* We have nothing to send. This connection is limited
2232 * either by receiver window or by application.
2237 /* If a thin stream is detected, retransmit after first
2238 * received dupack. Employ only if SACK is supported in order
2239 * to avoid possible corner-case series of spurious retransmissions
2240 * Use only if there are no unsent data.
2242 if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2243 tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2244 tcp_is_sack(tp) && !tcp_send_head(sk))
2247 /* Trick#6: TCP early retransmit, per RFC5827. To avoid spurious
2248 * retransmissions due to small network reorderings, we implement
2249 * Mitigation A.3 in the RFC and delay the retransmission for a short
2250 * interval if appropriate.
2252 if (tp->do_early_retrans && !tp->retrans_out && tp->sacked_out &&
2253 (tp->packets_out >= (tp->sacked_out + 1) && tp->packets_out < 4) &&
2254 !tcp_may_send_now(sk))
2255 return !tcp_pause_early_retransmit(sk, flag);
2260 /* Detect loss in event "A" above by marking head of queue up as lost.
2261 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2262 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2263 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2264 * the maximum SACKed segments to pass before reaching this limit.
2266 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2268 struct tcp_sock *tp = tcp_sk(sk);
2269 struct sk_buff *skb;
2270 int cnt, oldcnt, lost;
2272 /* Use SACK to deduce losses of new sequences sent during recovery */
2273 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq;
2275 WARN_ON(packets > tp->packets_out);
2276 if (tp->lost_skb_hint) {
2277 skb = tp->lost_skb_hint;
2278 cnt = tp->lost_cnt_hint;
2279 /* Head already handled? */
2280 if (mark_head && skb != tcp_write_queue_head(sk))
2283 skb = tcp_write_queue_head(sk);
2287 tcp_for_write_queue_from(skb, sk) {
2288 if (skb == tcp_send_head(sk))
2290 /* TODO: do this better */
2291 /* this is not the most efficient way to do this... */
2292 tp->lost_skb_hint = skb;
2293 tp->lost_cnt_hint = cnt;
2295 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2299 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2300 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2301 cnt += tcp_skb_pcount(skb);
2303 if (cnt > packets) {
2304 if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2305 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2306 (oldcnt >= packets))
2309 mss = tcp_skb_mss(skb);
2310 /* If needed, chop off the prefix to mark as lost. */
2311 lost = (packets - oldcnt) * mss;
2312 if (lost < skb->len &&
2313 tcp_fragment(sk, skb, lost, mss, GFP_ATOMIC) < 0)
2318 tcp_skb_mark_lost(tp, skb);
2323 tcp_verify_left_out(tp);
2326 /* Account newly detected lost packet(s) */
2328 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2330 struct tcp_sock *tp = tcp_sk(sk);
2332 if (tcp_is_reno(tp)) {
2333 tcp_mark_head_lost(sk, 1, 1);
2334 } else if (tcp_is_fack(tp)) {
2335 int lost = tp->fackets_out - tp->reordering;
2338 tcp_mark_head_lost(sk, lost, 0);
2340 int sacked_upto = tp->sacked_out - tp->reordering;
2341 if (sacked_upto >= 0)
2342 tcp_mark_head_lost(sk, sacked_upto, 0);
2343 else if (fast_rexmit)
2344 tcp_mark_head_lost(sk, 1, 1);
2348 static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2350 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2351 before(tp->rx_opt.rcv_tsecr, when);
2354 /* skb is spurious retransmitted if the returned timestamp echo
2355 * reply is prior to the skb transmission time
2357 static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2358 const struct sk_buff *skb)
2360 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2361 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp(skb));
2364 /* Nothing was retransmitted or returned timestamp is less
2365 * than timestamp of the first retransmission.
2367 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2369 return !tp->retrans_stamp ||
2370 tcp_tsopt_ecr_before(tp, tp->retrans_stamp);
2373 /* Undo procedures. */
2375 /* We can clear retrans_stamp when there are no retransmissions in the
2376 * window. It would seem that it is trivially available for us in
2377 * tp->retrans_out, however, that kind of assumptions doesn't consider
2378 * what will happen if errors occur when sending retransmission for the
2379 * second time. ...It could the that such segment has only
2380 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2381 * the head skb is enough except for some reneging corner cases that
2382 * are not worth the effort.
2384 * Main reason for all this complexity is the fact that connection dying
2385 * time now depends on the validity of the retrans_stamp, in particular,
2386 * that successive retransmissions of a segment must not advance
2387 * retrans_stamp under any conditions.
2389 static bool tcp_any_retrans_done(const struct sock *sk)
2391 const struct tcp_sock *tp = tcp_sk(sk);
2392 struct sk_buff *skb;
2394 if (tp->retrans_out)
2397 skb = tcp_write_queue_head(sk);
2398 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2404 #if FASTRETRANS_DEBUG > 1
2405 static void DBGUNDO(struct sock *sk, const char *msg)
2407 struct tcp_sock *tp = tcp_sk(sk);
2408 struct inet_sock *inet = inet_sk(sk);
2410 if (sk->sk_family == AF_INET) {
2411 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2413 &inet->inet_daddr, ntohs(inet->inet_dport),
2414 tp->snd_cwnd, tcp_left_out(tp),
2415 tp->snd_ssthresh, tp->prior_ssthresh,
2418 #if IS_ENABLED(CONFIG_IPV6)
2419 else if (sk->sk_family == AF_INET6) {
2420 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2422 &sk->sk_v6_daddr, ntohs(inet->inet_dport),
2423 tp->snd_cwnd, tcp_left_out(tp),
2424 tp->snd_ssthresh, tp->prior_ssthresh,
2430 #define DBGUNDO(x...) do { } while (0)
2433 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2435 struct tcp_sock *tp = tcp_sk(sk);
2438 struct sk_buff *skb;
2440 tcp_for_write_queue(skb, sk) {
2441 if (skb == tcp_send_head(sk))
2443 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2446 tcp_clear_all_retrans_hints(tp);
2449 if (tp->prior_ssthresh) {
2450 const struct inet_connection_sock *icsk = inet_csk(sk);
2452 if (icsk->icsk_ca_ops->undo_cwnd)
2453 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2455 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2457 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2458 tp->snd_ssthresh = tp->prior_ssthresh;
2459 tcp_ecn_withdraw_cwr(tp);
2462 tp->snd_cwnd_stamp = tcp_time_stamp;
2463 tp->undo_marker = 0;
2466 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2468 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2471 static bool tcp_is_non_sack_preventing_reopen(struct sock *sk)
2473 struct tcp_sock *tp = tcp_sk(sk);
2475 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2476 /* Hold old state until something *above* high_seq
2477 * is ACKed. For Reno it is MUST to prevent false
2478 * fast retransmits (RFC2582). SACK TCP is safe. */
2479 if (!tcp_any_retrans_done(sk))
2480 tp->retrans_stamp = 0;
2486 /* People celebrate: "We love our President!" */
2487 static bool tcp_try_undo_recovery(struct sock *sk)
2489 struct tcp_sock *tp = tcp_sk(sk);
2491 if (tcp_may_undo(tp)) {
2494 /* Happy end! We did not retransmit anything
2495 * or our original transmission succeeded.
2497 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2498 tcp_undo_cwnd_reduction(sk, false);
2499 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2500 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2502 mib_idx = LINUX_MIB_TCPFULLUNDO;
2504 NET_INC_STATS(sock_net(sk), mib_idx);
2506 if (tcp_is_non_sack_preventing_reopen(sk))
2508 tcp_set_ca_state(sk, TCP_CA_Open);
2509 tp->is_sack_reneg = 0;
2513 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2514 static bool tcp_try_undo_dsack(struct sock *sk)
2516 struct tcp_sock *tp = tcp_sk(sk);
2518 if (tp->undo_marker && !tp->undo_retrans) {
2519 DBGUNDO(sk, "D-SACK");
2520 tcp_undo_cwnd_reduction(sk, false);
2521 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2527 /* Undo during loss recovery after partial ACK or using F-RTO. */
2528 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2530 struct tcp_sock *tp = tcp_sk(sk);
2532 if (frto_undo || tcp_may_undo(tp)) {
2533 tcp_undo_cwnd_reduction(sk, true);
2535 DBGUNDO(sk, "partial loss");
2536 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2538 NET_INC_STATS(sock_net(sk),
2539 LINUX_MIB_TCPSPURIOUSRTOS);
2540 inet_csk(sk)->icsk_retransmits = 0;
2541 if (tcp_is_non_sack_preventing_reopen(sk))
2543 if (frto_undo || tcp_is_sack(tp)) {
2544 tcp_set_ca_state(sk, TCP_CA_Open);
2545 tp->is_sack_reneg = 0;
2552 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2553 * It computes the number of packets to send (sndcnt) based on packets newly
2555 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2556 * cwnd reductions across a full RTT.
2557 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2558 * But when the retransmits are acked without further losses, PRR
2559 * slow starts cwnd up to ssthresh to speed up the recovery.
2561 static void tcp_init_cwnd_reduction(struct sock *sk)
2563 struct tcp_sock *tp = tcp_sk(sk);
2565 tp->high_seq = tp->snd_nxt;
2566 tp->tlp_high_seq = 0;
2567 tp->snd_cwnd_cnt = 0;
2568 tp->prior_cwnd = tp->snd_cwnd;
2569 tp->prr_delivered = 0;
2571 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2572 tcp_ecn_queue_cwr(tp);
2575 static void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked,
2578 struct tcp_sock *tp = tcp_sk(sk);
2580 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2582 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2585 tp->prr_delivered += newly_acked_sacked;
2587 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2589 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2590 } else if ((flag & FLAG_RETRANS_DATA_ACKED) &&
2591 !(flag & FLAG_LOST_RETRANS)) {
2592 sndcnt = min_t(int, delta,
2593 max_t(int, tp->prr_delivered - tp->prr_out,
2594 newly_acked_sacked) + 1);
2596 sndcnt = min(delta, newly_acked_sacked);
2598 /* Force a fast retransmit upon entering fast recovery */
2599 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
2600 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2603 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2605 struct tcp_sock *tp = tcp_sk(sk);
2607 if (inet_csk(sk)->icsk_ca_ops->cong_control)
2610 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2611 if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH &&
2612 (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) {
2613 tp->snd_cwnd = tp->snd_ssthresh;
2614 tp->snd_cwnd_stamp = tcp_time_stamp;
2616 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2619 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2620 void tcp_enter_cwr(struct sock *sk)
2622 struct tcp_sock *tp = tcp_sk(sk);
2624 tp->prior_ssthresh = 0;
2625 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2626 tp->undo_marker = 0;
2627 tcp_init_cwnd_reduction(sk);
2628 tcp_set_ca_state(sk, TCP_CA_CWR);
2631 EXPORT_SYMBOL(tcp_enter_cwr);
2633 static void tcp_try_keep_open(struct sock *sk)
2635 struct tcp_sock *tp = tcp_sk(sk);
2636 int state = TCP_CA_Open;
2638 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2639 state = TCP_CA_Disorder;
2641 if (inet_csk(sk)->icsk_ca_state != state) {
2642 tcp_set_ca_state(sk, state);
2643 tp->high_seq = tp->snd_nxt;
2647 static void tcp_try_to_open(struct sock *sk, int flag)
2649 struct tcp_sock *tp = tcp_sk(sk);
2651 tcp_verify_left_out(tp);
2653 if (!tcp_any_retrans_done(sk))
2654 tp->retrans_stamp = 0;
2656 if (flag & FLAG_ECE)
2659 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2660 tcp_try_keep_open(sk);
2664 static void tcp_mtup_probe_failed(struct sock *sk)
2666 struct inet_connection_sock *icsk = inet_csk(sk);
2668 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2669 icsk->icsk_mtup.probe_size = 0;
2670 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
2673 static void tcp_mtup_probe_success(struct sock *sk)
2675 struct tcp_sock *tp = tcp_sk(sk);
2676 struct inet_connection_sock *icsk = inet_csk(sk);
2679 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2681 val = (u64)tp->snd_cwnd * tcp_mss_to_mtu(sk, tp->mss_cache);
2682 do_div(val, icsk->icsk_mtup.probe_size);
2683 WARN_ON_ONCE((u32)val != val);
2684 tp->snd_cwnd = max_t(u32, 1U, val);
2686 tp->snd_cwnd_cnt = 0;
2687 tp->snd_cwnd_stamp = tcp_time_stamp;
2688 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2690 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2691 icsk->icsk_mtup.probe_size = 0;
2692 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2693 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
2696 /* Do a simple retransmit without using the backoff mechanisms in
2697 * tcp_timer. This is used for path mtu discovery.
2698 * The socket is already locked here.
2700 void tcp_simple_retransmit(struct sock *sk)
2702 const struct inet_connection_sock *icsk = inet_csk(sk);
2703 struct tcp_sock *tp = tcp_sk(sk);
2704 struct sk_buff *skb;
2705 unsigned int mss = tcp_current_mss(sk);
2706 u32 prior_lost = tp->lost_out;
2708 tcp_for_write_queue(skb, sk) {
2709 if (skb == tcp_send_head(sk))
2711 if (tcp_skb_seglen(skb) > mss &&
2712 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2713 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2714 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2715 tp->retrans_out -= tcp_skb_pcount(skb);
2717 tcp_skb_mark_lost_uncond_verify(tp, skb);
2721 tcp_clear_retrans_hints_partial(tp);
2723 if (prior_lost == tp->lost_out)
2726 if (tcp_is_reno(tp))
2727 tcp_limit_reno_sacked(tp);
2729 tcp_verify_left_out(tp);
2731 /* Don't muck with the congestion window here.
2732 * Reason is that we do not increase amount of _data_
2733 * in network, but units changed and effective
2734 * cwnd/ssthresh really reduced now.
2736 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2737 tp->high_seq = tp->snd_nxt;
2738 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2739 tp->prior_ssthresh = 0;
2740 tp->undo_marker = 0;
2741 tcp_set_ca_state(sk, TCP_CA_Loss);
2743 tcp_xmit_retransmit_queue(sk);
2745 EXPORT_SYMBOL(tcp_simple_retransmit);
2747 static void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2749 struct tcp_sock *tp = tcp_sk(sk);
2752 if (tcp_is_reno(tp))
2753 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2755 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2757 NET_INC_STATS(sock_net(sk), mib_idx);
2759 tp->prior_ssthresh = 0;
2762 if (!tcp_in_cwnd_reduction(sk)) {
2764 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2765 tcp_init_cwnd_reduction(sk);
2767 tcp_set_ca_state(sk, TCP_CA_Recovery);
2770 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2771 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2773 static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack,
2776 struct tcp_sock *tp = tcp_sk(sk);
2777 bool recovered = !before(tp->snd_una, tp->high_seq);
2779 if ((flag & FLAG_SND_UNA_ADVANCED) &&
2780 tcp_try_undo_loss(sk, false))
2783 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2784 /* Step 3.b. A timeout is spurious if not all data are
2785 * lost, i.e., never-retransmitted data are (s)acked.
2787 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2788 tcp_try_undo_loss(sk, true))
2791 if (after(tp->snd_nxt, tp->high_seq)) {
2792 if (flag & FLAG_DATA_SACKED || is_dupack)
2793 tp->frto = 0; /* Step 3.a. loss was real */
2794 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2795 tp->high_seq = tp->snd_nxt;
2796 /* Step 2.b. Try send new data (but deferred until cwnd
2797 * is updated in tcp_ack()). Otherwise fall back to
2798 * the conventional recovery.
2800 if (tcp_send_head(sk) &&
2801 after(tcp_wnd_end(tp), tp->snd_nxt)) {
2802 *rexmit = REXMIT_NEW;
2810 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2811 tcp_try_undo_recovery(sk);
2814 if (tcp_is_reno(tp)) {
2815 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2816 * delivered. Lower inflight to clock out (re)tranmissions.
2818 if (after(tp->snd_nxt, tp->high_seq) && is_dupack)
2819 tcp_add_reno_sack(sk);
2820 else if (flag & FLAG_SND_UNA_ADVANCED)
2821 tcp_reset_reno_sack(tp);
2823 *rexmit = REXMIT_LOST;
2826 /* Undo during fast recovery after partial ACK. */
2827 static bool tcp_try_undo_partial(struct sock *sk, const int acked)
2829 struct tcp_sock *tp = tcp_sk(sk);
2831 if (tp->undo_marker && tcp_packet_delayed(tp)) {
2832 /* Plain luck! Hole if filled with delayed
2833 * packet, rather than with a retransmit.
2835 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2837 /* We are getting evidence that the reordering degree is higher
2838 * than we realized. If there are no retransmits out then we
2839 * can undo. Otherwise we clock out new packets but do not
2840 * mark more packets lost or retransmit more.
2842 if (tp->retrans_out)
2845 if (!tcp_any_retrans_done(sk))
2846 tp->retrans_stamp = 0;
2848 DBGUNDO(sk, "partial recovery");
2849 tcp_undo_cwnd_reduction(sk, true);
2850 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2851 tcp_try_keep_open(sk);
2857 /* Process an event, which can update packets-in-flight not trivially.
2858 * Main goal of this function is to calculate new estimate for left_out,
2859 * taking into account both packets sitting in receiver's buffer and
2860 * packets lost by network.
2862 * Besides that it updates the congestion state when packet loss or ECN
2863 * is detected. But it does not reduce the cwnd, it is done by the
2864 * congestion control later.
2866 * It does _not_ decide what to send, it is made in function
2867 * tcp_xmit_retransmit_queue().
2869 static void tcp_fastretrans_alert(struct sock *sk, const int acked,
2870 bool is_dupack, int *ack_flag, int *rexmit)
2872 struct inet_connection_sock *icsk = inet_csk(sk);
2873 struct tcp_sock *tp = tcp_sk(sk);
2874 int fast_rexmit = 0, flag = *ack_flag;
2875 bool do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2876 (tcp_fackets_out(tp) > tp->reordering));
2878 if (!tp->packets_out && tp->sacked_out)
2880 if (!tp->sacked_out && tp->fackets_out)
2881 tp->fackets_out = 0;
2883 /* Now state machine starts.
2884 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2885 if (flag & FLAG_ECE)
2886 tp->prior_ssthresh = 0;
2888 /* B. In all the states check for reneging SACKs. */
2889 if (tcp_check_sack_reneging(sk, flag))
2892 /* C. Check consistency of the current state. */
2893 tcp_verify_left_out(tp);
2895 /* D. Check state exit conditions. State can be terminated
2896 * when high_seq is ACKed. */
2897 if (icsk->icsk_ca_state == TCP_CA_Open) {
2898 WARN_ON(tp->retrans_out != 0);
2899 tp->retrans_stamp = 0;
2900 } else if (!before(tp->snd_una, tp->high_seq)) {
2901 switch (icsk->icsk_ca_state) {
2903 /* CWR is to be held something *above* high_seq
2904 * is ACKed for CWR bit to reach receiver. */
2905 if (tp->snd_una != tp->high_seq) {
2906 tcp_end_cwnd_reduction(sk);
2907 tcp_set_ca_state(sk, TCP_CA_Open);
2911 case TCP_CA_Recovery:
2912 if (tcp_is_reno(tp))
2913 tcp_reset_reno_sack(tp);
2914 if (tcp_try_undo_recovery(sk))
2916 tcp_end_cwnd_reduction(sk);
2921 /* Use RACK to detect loss */
2922 if (sysctl_tcp_recovery & TCP_RACK_LOST_RETRANS &&
2923 tcp_rack_mark_lost(sk)) {
2924 flag |= FLAG_LOST_RETRANS;
2925 *ack_flag |= FLAG_LOST_RETRANS;
2928 /* E. Process state. */
2929 switch (icsk->icsk_ca_state) {
2930 case TCP_CA_Recovery:
2931 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2932 if (tcp_is_reno(tp) && is_dupack)
2933 tcp_add_reno_sack(sk);
2935 if (tcp_try_undo_partial(sk, acked))
2937 /* Partial ACK arrived. Force fast retransmit. */
2938 do_lost = tcp_is_reno(tp) ||
2939 tcp_fackets_out(tp) > tp->reordering;
2941 if (tcp_try_undo_dsack(sk)) {
2942 tcp_try_keep_open(sk);
2947 tcp_process_loss(sk, flag, is_dupack, rexmit);
2948 if (icsk->icsk_ca_state != TCP_CA_Open &&
2949 !(flag & FLAG_LOST_RETRANS))
2951 /* Change state if cwnd is undone or retransmits are lost */
2953 if (tcp_is_reno(tp)) {
2954 if (flag & FLAG_SND_UNA_ADVANCED)
2955 tcp_reset_reno_sack(tp);
2957 tcp_add_reno_sack(sk);
2960 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2961 tcp_try_undo_dsack(sk);
2963 if (!tcp_time_to_recover(sk, flag)) {
2964 tcp_try_to_open(sk, flag);
2968 /* MTU probe failure: don't reduce cwnd */
2969 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2970 icsk->icsk_mtup.probe_size &&
2971 tp->snd_una == tp->mtu_probe.probe_seq_start) {
2972 tcp_mtup_probe_failed(sk);
2973 /* Restores the reduction we did in tcp_mtup_probe() */
2975 tcp_simple_retransmit(sk);
2979 /* Otherwise enter Recovery state */
2980 tcp_enter_recovery(sk, (flag & FLAG_ECE));
2985 tcp_update_scoreboard(sk, fast_rexmit);
2986 *rexmit = REXMIT_LOST;
2989 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us)
2991 struct tcp_sock *tp = tcp_sk(sk);
2992 u32 wlen = sysctl_tcp_min_rtt_wlen * HZ;
2994 minmax_running_min(&tp->rtt_min, wlen, tcp_time_stamp,
2995 rtt_us ? : jiffies_to_usecs(1));
2998 static inline bool tcp_ack_update_rtt(struct sock *sk, const int flag,
2999 long seq_rtt_us, long sack_rtt_us,
3002 const struct tcp_sock *tp = tcp_sk(sk);
3004 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
3005 * broken middle-boxes or peers may corrupt TS-ECR fields. But
3006 * Karn's algorithm forbids taking RTT if some retransmitted data
3007 * is acked (RFC6298).
3010 seq_rtt_us = sack_rtt_us;
3012 /* RTTM Rule: A TSecr value received in a segment is used to
3013 * update the averaged RTT measurement only if the segment
3014 * acknowledges some new data, i.e., only if it advances the
3015 * left edge of the send window.
3016 * See draft-ietf-tcplw-high-performance-00, section 3.3.
3018 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
3020 seq_rtt_us = ca_rtt_us = jiffies_to_usecs(tcp_time_stamp -
3021 tp->rx_opt.rcv_tsecr);
3025 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
3026 * always taken together with ACK, SACK, or TS-opts. Any negative
3027 * values will be skipped with the seq_rtt_us < 0 check above.
3029 tcp_update_rtt_min(sk, ca_rtt_us);
3030 tcp_rtt_estimator(sk, seq_rtt_us);
3033 /* RFC6298: only reset backoff on valid RTT measurement. */
3034 inet_csk(sk)->icsk_backoff = 0;
3038 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
3039 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
3043 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack.v64) {
3044 struct skb_mstamp now;
3046 skb_mstamp_get(&now);
3047 rtt_us = skb_mstamp_us_delta(&now, &tcp_rsk(req)->snt_synack);
3050 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us);
3054 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
3056 const struct inet_connection_sock *icsk = inet_csk(sk);
3058 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
3059 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
3062 /* Restart timer after forward progress on connection.
3063 * RFC2988 recommends to restart timer to now+rto.
3065 void tcp_rearm_rto(struct sock *sk)
3067 const struct inet_connection_sock *icsk = inet_csk(sk);
3068 struct tcp_sock *tp = tcp_sk(sk);
3070 /* If the retrans timer is currently being used by Fast Open
3071 * for SYN-ACK retrans purpose, stay put.
3073 if (tp->fastopen_rsk)
3076 if (!tp->packets_out) {
3077 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3079 u32 rto = inet_csk(sk)->icsk_rto;
3080 /* Offset the time elapsed after installing regular RTO */
3081 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
3082 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
3083 struct sk_buff *skb = tcp_write_queue_head(sk);
3084 const u32 rto_time_stamp =
3085 tcp_skb_timestamp(skb) + rto;
3086 s32 delta = (s32)(rto_time_stamp - tcp_time_stamp);
3087 /* delta may not be positive if the socket is locked
3088 * when the retrans timer fires and is rescheduled.
3090 rto = max(delta, 1);
3092 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
3097 /* This function is called when the delayed ER timer fires. TCP enters
3098 * fast recovery and performs fast-retransmit.
3100 void tcp_resume_early_retransmit(struct sock *sk)
3102 struct tcp_sock *tp = tcp_sk(sk);
3106 /* Stop if ER is disabled after the delayed ER timer is scheduled */
3107 if (!tp->do_early_retrans)
3110 tcp_enter_recovery(sk, false);
3111 tcp_update_scoreboard(sk, 1);
3112 tcp_xmit_retransmit_queue(sk);
3115 /* If we get here, the whole TSO packet has not been acked. */
3116 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3118 struct tcp_sock *tp = tcp_sk(sk);
3121 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3123 packets_acked = tcp_skb_pcount(skb);
3124 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3126 packets_acked -= tcp_skb_pcount(skb);
3128 if (packets_acked) {
3129 BUG_ON(tcp_skb_pcount(skb) == 0);
3130 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3133 return packets_acked;
3136 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3139 const struct skb_shared_info *shinfo;
3141 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3142 if (likely(!TCP_SKB_CB(skb)->txstamp_ack))
3145 shinfo = skb_shinfo(skb);
3146 if (!before(shinfo->tskey, prior_snd_una) &&
3147 before(shinfo->tskey, tcp_sk(sk)->snd_una))
3148 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK);
3151 /* Remove acknowledged frames from the retransmission queue. If our packet
3152 * is before the ack sequence we can discard it as it's confirmed to have
3153 * arrived at the other end.
3155 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3156 u32 prior_snd_una, int *acked,
3157 struct tcp_sacktag_state *sack,
3158 struct skb_mstamp *now)
3160 const struct inet_connection_sock *icsk = inet_csk(sk);
3161 struct skb_mstamp first_ackt, last_ackt;
3162 struct tcp_sock *tp = tcp_sk(sk);
3163 u32 prior_sacked = tp->sacked_out;
3164 u32 reord = tp->packets_out;
3165 bool fully_acked = true;
3166 long sack_rtt_us = -1L;
3167 long seq_rtt_us = -1L;
3168 long ca_rtt_us = -1L;
3169 struct sk_buff *skb;
3171 u32 last_in_flight = 0;
3177 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3178 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3179 u8 sacked = scb->sacked;
3182 tcp_ack_tstamp(sk, skb, prior_snd_una);
3184 /* Determine how many packets and what bytes were acked, tso and else */
3185 if (after(scb->end_seq, tp->snd_una)) {
3186 if (tcp_skb_pcount(skb) == 1 ||
3187 !after(tp->snd_una, scb->seq))
3190 acked_pcount = tcp_tso_acked(sk, skb);
3193 fully_acked = false;
3195 /* Speedup tcp_unlink_write_queue() and next loop */
3196 prefetchw(skb->next);
3197 acked_pcount = tcp_skb_pcount(skb);
3200 if (unlikely(sacked & TCPCB_RETRANS)) {
3201 if (sacked & TCPCB_SACKED_RETRANS)
3202 tp->retrans_out -= acked_pcount;
3203 flag |= FLAG_RETRANS_DATA_ACKED;
3204 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3205 last_ackt = skb->skb_mstamp;
3206 WARN_ON_ONCE(last_ackt.v64 == 0);
3207 if (!first_ackt.v64)
3208 first_ackt = last_ackt;
3210 last_in_flight = TCP_SKB_CB(skb)->tx.in_flight;
3211 reord = min(pkts_acked, reord);
3212 if (!after(scb->end_seq, tp->high_seq))
3213 flag |= FLAG_ORIG_SACK_ACKED;
3216 if (sacked & TCPCB_SACKED_ACKED) {
3217 tp->sacked_out -= acked_pcount;
3218 } else if (tcp_is_sack(tp)) {
3219 tp->delivered += acked_pcount;
3220 if (!tcp_skb_spurious_retrans(tp, skb))
3221 tcp_rack_advance(tp, &skb->skb_mstamp, sacked);
3223 if (sacked & TCPCB_LOST)
3224 tp->lost_out -= acked_pcount;
3226 tp->packets_out -= acked_pcount;
3227 pkts_acked += acked_pcount;
3228 tcp_rate_skb_delivered(sk, skb, sack->rate);
3230 /* Initial outgoing SYN's get put onto the write_queue
3231 * just like anything else we transmit. It is not
3232 * true data, and if we misinform our callers that
3233 * this ACK acks real data, we will erroneously exit
3234 * connection startup slow start one packet too
3235 * quickly. This is severely frowned upon behavior.
3237 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3238 flag |= FLAG_DATA_ACKED;
3240 flag |= FLAG_SYN_ACKED;
3241 tp->retrans_stamp = 0;
3247 tcp_unlink_write_queue(skb, sk);
3248 sk_wmem_free_skb(sk, skb);
3249 if (unlikely(skb == tp->retransmit_skb_hint))
3250 tp->retransmit_skb_hint = NULL;
3251 if (unlikely(skb == tp->lost_skb_hint))
3252 tp->lost_skb_hint = NULL;
3255 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3256 tp->snd_up = tp->snd_una;
3258 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3259 flag |= FLAG_SACK_RENEGING;
3261 if (likely(first_ackt.v64) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
3262 seq_rtt_us = skb_mstamp_us_delta(now, &first_ackt);
3263 ca_rtt_us = skb_mstamp_us_delta(now, &last_ackt);
3265 if (sack->first_sackt.v64) {
3266 sack_rtt_us = skb_mstamp_us_delta(now, &sack->first_sackt);
3267 ca_rtt_us = skb_mstamp_us_delta(now, &sack->last_sackt);
3269 sack->rate->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet, or -1 */
3270 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
3273 if (flag & FLAG_ACKED) {
3275 if (unlikely(icsk->icsk_mtup.probe_size &&
3276 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3277 tcp_mtup_probe_success(sk);
3280 if (tcp_is_reno(tp)) {
3281 tcp_remove_reno_sacks(sk, pkts_acked);
3283 /* If any of the cumulatively ACKed segments was
3284 * retransmitted, non-SACK case cannot confirm that
3285 * progress was due to original transmission due to
3286 * lack of TCPCB_SACKED_ACKED bits even if some of
3287 * the packets may have been never retransmitted.
3289 if (flag & FLAG_RETRANS_DATA_ACKED)
3290 flag &= ~FLAG_ORIG_SACK_ACKED;
3294 /* Non-retransmitted hole got filled? That's reordering */
3295 if (reord < prior_fackets && reord <= tp->fackets_out)
3296 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3298 delta = tcp_is_fack(tp) ? pkts_acked :
3299 prior_sacked - tp->sacked_out;
3300 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3303 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3305 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3306 sack_rtt_us > skb_mstamp_us_delta(now, &skb->skb_mstamp)) {
3307 /* Do not re-arm RTO if the sack RTT is measured from data sent
3308 * after when the head was last (re)transmitted. Otherwise the
3309 * timeout may continue to extend in loss recovery.
3314 if (icsk->icsk_ca_ops->pkts_acked) {
3315 struct ack_sample sample = { .pkts_acked = pkts_acked,
3316 .rtt_us = ca_rtt_us,
3317 .in_flight = last_in_flight };
3319 icsk->icsk_ca_ops->pkts_acked(sk, &sample);
3322 #if FASTRETRANS_DEBUG > 0
3323 WARN_ON((int)tp->sacked_out < 0);
3324 WARN_ON((int)tp->lost_out < 0);
3325 WARN_ON((int)tp->retrans_out < 0);
3326 if (!tp->packets_out && tcp_is_sack(tp)) {
3327 icsk = inet_csk(sk);
3329 pr_debug("Leak l=%u %d\n",
3330 tp->lost_out, icsk->icsk_ca_state);
3333 if (tp->sacked_out) {
3334 pr_debug("Leak s=%u %d\n",
3335 tp->sacked_out, icsk->icsk_ca_state);
3338 if (tp->retrans_out) {
3339 pr_debug("Leak r=%u %d\n",
3340 tp->retrans_out, icsk->icsk_ca_state);
3341 tp->retrans_out = 0;
3345 *acked = pkts_acked;
3349 static void tcp_ack_probe(struct sock *sk)
3351 const struct tcp_sock *tp = tcp_sk(sk);
3352 struct inet_connection_sock *icsk = inet_csk(sk);
3354 /* Was it a usable window open? */
3356 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3357 icsk->icsk_backoff = 0;
3358 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3359 /* Socket must be waked up by subsequent tcp_data_snd_check().
3360 * This function is not for random using!
3363 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
3365 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3370 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3372 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3373 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3376 /* Decide wheather to run the increase function of congestion control. */
3377 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3379 /* If reordering is high then always grow cwnd whenever data is
3380 * delivered regardless of its ordering. Otherwise stay conservative
3381 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3382 * new SACK or ECE mark may first advance cwnd here and later reduce
3383 * cwnd in tcp_fastretrans_alert() based on more states.
3385 if (tcp_sk(sk)->reordering > sock_net(sk)->ipv4.sysctl_tcp_reordering)
3386 return flag & FLAG_FORWARD_PROGRESS;
3388 return flag & FLAG_DATA_ACKED;
3391 /* The "ultimate" congestion control function that aims to replace the rigid
3392 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3393 * It's called toward the end of processing an ACK with precise rate
3394 * information. All transmission or retransmission are delayed afterwards.
3396 static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
3397 int flag, const struct rate_sample *rs)
3399 const struct inet_connection_sock *icsk = inet_csk(sk);
3401 if (icsk->icsk_ca_ops->cong_control) {
3402 icsk->icsk_ca_ops->cong_control(sk, rs);
3406 if (tcp_in_cwnd_reduction(sk)) {
3407 /* Reduce cwnd if state mandates */
3408 tcp_cwnd_reduction(sk, acked_sacked, flag);
3409 } else if (tcp_may_raise_cwnd(sk, flag)) {
3410 /* Advance cwnd if state allows */
3411 tcp_cong_avoid(sk, ack, acked_sacked);
3413 tcp_update_pacing_rate(sk);
3416 /* Check that window update is acceptable.
3417 * The function assumes that snd_una<=ack<=snd_next.
3419 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3420 const u32 ack, const u32 ack_seq,
3423 return after(ack, tp->snd_una) ||
3424 after(ack_seq, tp->snd_wl1) ||
3425 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3428 /* If we update tp->snd_una, also update tp->bytes_acked */
3429 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3431 u32 delta = ack - tp->snd_una;
3433 sock_owned_by_me((struct sock *)tp);
3434 u64_stats_update_begin_raw(&tp->syncp);
3435 tp->bytes_acked += delta;
3436 u64_stats_update_end_raw(&tp->syncp);
3440 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3441 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3443 u32 delta = seq - tp->rcv_nxt;
3445 sock_owned_by_me((struct sock *)tp);
3446 u64_stats_update_begin_raw(&tp->syncp);
3447 tp->bytes_received += delta;
3448 u64_stats_update_end_raw(&tp->syncp);
3452 /* Update our send window.
3454 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3455 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3457 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3460 struct tcp_sock *tp = tcp_sk(sk);
3462 u32 nwin = ntohs(tcp_hdr(skb)->window);
3464 if (likely(!tcp_hdr(skb)->syn))
3465 nwin <<= tp->rx_opt.snd_wscale;
3467 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3468 flag |= FLAG_WIN_UPDATE;
3469 tcp_update_wl(tp, ack_seq);
3471 if (tp->snd_wnd != nwin) {
3474 /* Note, it is the only place, where
3475 * fast path is recovered for sending TCP.
3478 tcp_fast_path_check(sk);
3480 if (tcp_send_head(sk))
3481 tcp_slow_start_after_idle_check(sk);
3483 if (nwin > tp->max_window) {
3484 tp->max_window = nwin;
3485 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3490 tcp_snd_una_update(tp, ack);
3495 static bool __tcp_oow_rate_limited(struct net *net, int mib_idx,
3496 u32 *last_oow_ack_time)
3498 if (*last_oow_ack_time) {
3499 s32 elapsed = (s32)(tcp_time_stamp - *last_oow_ack_time);
3501 if (0 <= elapsed && elapsed < sysctl_tcp_invalid_ratelimit) {
3502 NET_INC_STATS(net, mib_idx);
3503 return true; /* rate-limited: don't send yet! */
3507 *last_oow_ack_time = tcp_time_stamp;
3509 return false; /* not rate-limited: go ahead, send dupack now! */
3512 /* Return true if we're currently rate-limiting out-of-window ACKs and
3513 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3514 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3515 * attacks that send repeated SYNs or ACKs for the same connection. To
3516 * do this, we do not send a duplicate SYNACK or ACK if the remote
3517 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3519 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3520 int mib_idx, u32 *last_oow_ack_time)
3522 /* Data packets without SYNs are not likely part of an ACK loop. */
3523 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3527 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time);
3530 /* RFC 5961 7 [ACK Throttling] */
3531 static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb)
3533 /* unprotected vars, we dont care of overwrites */
3534 static u32 challenge_timestamp;
3535 static unsigned int challenge_count;
3536 struct tcp_sock *tp = tcp_sk(sk);
3539 /* First check our per-socket dupack rate limit. */
3540 if (__tcp_oow_rate_limited(sock_net(sk),
3541 LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3542 &tp->last_oow_ack_time))
3545 /* Then check host-wide RFC 5961 rate limit. */
3547 if (now != challenge_timestamp) {
3548 u32 half = (sysctl_tcp_challenge_ack_limit + 1) >> 1;
3550 challenge_timestamp = now;
3551 WRITE_ONCE(challenge_count, half +
3552 prandom_u32_max(sysctl_tcp_challenge_ack_limit));
3554 count = READ_ONCE(challenge_count);
3556 WRITE_ONCE(challenge_count, count - 1);
3557 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPCHALLENGEACK);
3562 static void tcp_store_ts_recent(struct tcp_sock *tp)
3564 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3565 tp->rx_opt.ts_recent_stamp = get_seconds();
3568 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3570 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3571 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3572 * extra check below makes sure this can only happen
3573 * for pure ACK frames. -DaveM
3575 * Not only, also it occurs for expired timestamps.
3578 if (tcp_paws_check(&tp->rx_opt, 0))
3579 tcp_store_ts_recent(tp);
3583 /* This routine deals with acks during a TLP episode and ends an episode by
3584 * resetting tlp_high_seq. Ref: TLP algorithm in draft-ietf-tcpm-rack
3586 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3588 struct tcp_sock *tp = tcp_sk(sk);
3590 if (before(ack, tp->tlp_high_seq))
3593 if (!tp->tlp_retrans) {
3594 /* TLP of new data has been acknowledged */
3595 tp->tlp_high_seq = 0;
3596 } else if (flag & FLAG_DSACKING_ACK) {
3597 /* This DSACK means original and TLP probe arrived; no loss */
3598 tp->tlp_high_seq = 0;
3599 } else if (after(ack, tp->tlp_high_seq)) {
3600 /* ACK advances: there was a loss, so reduce cwnd. Reset
3601 * tlp_high_seq in tcp_init_cwnd_reduction()
3603 tcp_init_cwnd_reduction(sk);
3604 tcp_set_ca_state(sk, TCP_CA_CWR);
3605 tcp_end_cwnd_reduction(sk);
3606 tcp_try_keep_open(sk);
3607 NET_INC_STATS(sock_net(sk),
3608 LINUX_MIB_TCPLOSSPROBERECOVERY);
3609 } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3610 FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3611 /* Pure dupack: original and TLP probe arrived; no loss */
3612 tp->tlp_high_seq = 0;
3616 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3618 const struct inet_connection_sock *icsk = inet_csk(sk);
3620 if (icsk->icsk_ca_ops->in_ack_event)
3621 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3624 /* Congestion control has updated the cwnd already. So if we're in
3625 * loss recovery then now we do any new sends (for FRTO) or
3626 * retransmits (for CA_Loss or CA_recovery) that make sense.
3628 static void tcp_xmit_recovery(struct sock *sk, int rexmit)
3630 struct tcp_sock *tp = tcp_sk(sk);
3632 if (rexmit == REXMIT_NONE)
3635 if (unlikely(rexmit == 2)) {
3636 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
3638 if (after(tp->snd_nxt, tp->high_seq))
3642 tcp_xmit_retransmit_queue(sk);
3645 /* This routine deals with incoming acks, but not outgoing ones. */
3646 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3648 struct inet_connection_sock *icsk = inet_csk(sk);
3649 struct tcp_sock *tp = tcp_sk(sk);
3650 struct tcp_sacktag_state sack_state;
3651 struct rate_sample rs = { .prior_delivered = 0 };
3652 u32 prior_snd_una = tp->snd_una;
3653 bool is_sack_reneg = tp->is_sack_reneg;
3654 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3655 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3656 bool is_dupack = false;
3658 int prior_packets = tp->packets_out;
3659 u32 delivered = tp->delivered;
3660 u32 lost = tp->lost;
3661 int acked = 0; /* Number of packets newly acked */
3662 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
3663 struct skb_mstamp now;
3665 sack_state.first_sackt.v64 = 0;
3666 sack_state.rate = &rs;
3668 /* We very likely will need to access write queue head. */
3669 prefetchw(sk->sk_write_queue.next);
3671 /* If the ack is older than previous acks
3672 * then we can probably ignore it.
3674 if (before(ack, prior_snd_una)) {
3675 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3676 if (before(ack, prior_snd_una - tp->max_window)) {
3677 if (!(flag & FLAG_NO_CHALLENGE_ACK))
3678 tcp_send_challenge_ack(sk, skb);
3684 /* If the ack includes data we haven't sent yet, discard
3685 * this segment (RFC793 Section 3.9).
3687 if (after(ack, tp->snd_nxt))
3690 skb_mstamp_get(&now);
3692 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
3693 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
3696 if (after(ack, prior_snd_una)) {
3697 flag |= FLAG_SND_UNA_ADVANCED;
3698 icsk->icsk_retransmits = 0;
3701 prior_fackets = tp->fackets_out;
3702 rs.prior_in_flight = tcp_packets_in_flight(tp);
3704 /* ts_recent update must be made after we are sure that the packet
3707 if (flag & FLAG_UPDATE_TS_RECENT)
3708 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3710 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3711 /* Window is constant, pure forward advance.
3712 * No more checks are required.
3713 * Note, we use the fact that SND.UNA>=SND.WL2.
3715 tcp_update_wl(tp, ack_seq);
3716 tcp_snd_una_update(tp, ack);
3717 flag |= FLAG_WIN_UPDATE;
3719 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3721 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS);
3723 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3725 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3728 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3730 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3732 if (TCP_SKB_CB(skb)->sacked)
3733 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3736 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3738 ack_ev_flags |= CA_ACK_ECE;
3741 if (flag & FLAG_WIN_UPDATE)
3742 ack_ev_flags |= CA_ACK_WIN_UPDATE;
3744 tcp_in_ack_event(sk, ack_ev_flags);
3747 /* We passed data and got it acked, remove any soft error
3748 * log. Something worked...
3750 sk->sk_err_soft = 0;
3751 icsk->icsk_probes_out = 0;
3752 tp->rcv_tstamp = tcp_time_stamp;
3756 /* See if we can take anything off of the retransmit queue. */
3757 flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una, &acked,
3760 if (tcp_ack_is_dubious(sk, flag)) {
3761 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3762 tcp_fastretrans_alert(sk, acked, is_dupack, &flag, &rexmit);
3764 if (tp->tlp_high_seq)
3765 tcp_process_tlp_ack(sk, ack, flag);
3767 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) {
3768 struct dst_entry *dst = __sk_dst_get(sk);
3773 if (icsk->icsk_pending == ICSK_TIME_RETRANS)
3774 tcp_schedule_loss_probe(sk);
3775 delivered = tp->delivered - delivered; /* freshly ACKed or SACKed */
3776 lost = tp->lost - lost; /* freshly marked lost */
3777 tcp_rate_gen(sk, delivered, lost, is_sack_reneg, &now, &rs);
3778 tcp_cong_control(sk, ack, delivered, flag, &rs);
3779 tcp_xmit_recovery(sk, rexmit);
3783 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3784 if (flag & FLAG_DSACKING_ACK)
3785 tcp_fastretrans_alert(sk, acked, is_dupack, &flag, &rexmit);
3786 /* If this ack opens up a zero window, clear backoff. It was
3787 * being used to time the probes, and is probably far higher than
3788 * it needs to be for normal retransmission.
3790 if (tcp_send_head(sk))
3793 if (tp->tlp_high_seq)
3794 tcp_process_tlp_ack(sk, ack, flag);
3798 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3802 /* If data was SACKed, tag it and see if we should send more data.
3803 * If data was DSACKed, see if we can undo a cwnd reduction.
3805 if (TCP_SKB_CB(skb)->sacked) {
3806 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3808 tcp_fastretrans_alert(sk, acked, is_dupack, &flag, &rexmit);
3809 tcp_xmit_recovery(sk, rexmit);
3812 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3816 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3817 bool syn, struct tcp_fastopen_cookie *foc,
3820 /* Valid only in SYN or SYN-ACK with an even length. */
3821 if (!foc || !syn || len < 0 || (len & 1))
3824 if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3825 len <= TCP_FASTOPEN_COOKIE_MAX)
3826 memcpy(foc->val, cookie, len);
3833 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3834 * But, this can also be called on packets in the established flow when
3835 * the fast version below fails.
3837 void tcp_parse_options(const struct sk_buff *skb,
3838 struct tcp_options_received *opt_rx, int estab,
3839 struct tcp_fastopen_cookie *foc)
3841 const unsigned char *ptr;
3842 const struct tcphdr *th = tcp_hdr(skb);
3843 int length = (th->doff * 4) - sizeof(struct tcphdr);
3845 ptr = (const unsigned char *)(th + 1);
3846 opt_rx->saw_tstamp = 0;
3848 while (length > 0) {
3849 int opcode = *ptr++;
3855 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3860 if (opsize < 2) /* "silly options" */
3862 if (opsize > length)
3863 return; /* don't parse partial options */
3866 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3867 u16 in_mss = get_unaligned_be16(ptr);
3869 if (opt_rx->user_mss &&
3870 opt_rx->user_mss < in_mss)
3871 in_mss = opt_rx->user_mss;
3872 opt_rx->mss_clamp = in_mss;
3877 if (opsize == TCPOLEN_WINDOW && th->syn &&
3878 !estab && sysctl_tcp_window_scaling) {
3879 __u8 snd_wscale = *(__u8 *)ptr;
3880 opt_rx->wscale_ok = 1;
3881 if (snd_wscale > 14) {
3882 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3887 opt_rx->snd_wscale = snd_wscale;
3890 case TCPOPT_TIMESTAMP:
3891 if ((opsize == TCPOLEN_TIMESTAMP) &&
3892 ((estab && opt_rx->tstamp_ok) ||
3893 (!estab && sysctl_tcp_timestamps))) {
3894 opt_rx->saw_tstamp = 1;
3895 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3896 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3899 case TCPOPT_SACK_PERM:
3900 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3901 !estab && sysctl_tcp_sack) {
3902 opt_rx->sack_ok = TCP_SACK_SEEN;
3903 tcp_sack_reset(opt_rx);
3908 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3909 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3911 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3914 #ifdef CONFIG_TCP_MD5SIG
3917 * The MD5 Hash has already been
3918 * checked (see tcp_v{4,6}_do_rcv()).
3922 case TCPOPT_FASTOPEN:
3923 tcp_parse_fastopen_option(
3924 opsize - TCPOLEN_FASTOPEN_BASE,
3925 ptr, th->syn, foc, false);
3929 /* Fast Open option shares code 254 using a
3930 * 16 bits magic number.
3932 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
3933 get_unaligned_be16(ptr) ==
3934 TCPOPT_FASTOPEN_MAGIC)
3935 tcp_parse_fastopen_option(opsize -
3936 TCPOLEN_EXP_FASTOPEN_BASE,
3937 ptr + 2, th->syn, foc, true);
3946 EXPORT_SYMBOL(tcp_parse_options);
3948 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3950 const __be32 *ptr = (const __be32 *)(th + 1);
3952 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3953 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3954 tp->rx_opt.saw_tstamp = 1;
3956 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3959 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3961 tp->rx_opt.rcv_tsecr = 0;
3967 /* Fast parse options. This hopes to only see timestamps.
3968 * If it is wrong it falls back on tcp_parse_options().
3970 static bool tcp_fast_parse_options(const struct sk_buff *skb,
3971 const struct tcphdr *th, struct tcp_sock *tp)
3973 /* In the spirit of fast parsing, compare doff directly to constant
3974 * values. Because equality is used, short doff can be ignored here.
3976 if (th->doff == (sizeof(*th) / 4)) {
3977 tp->rx_opt.saw_tstamp = 0;
3979 } else if (tp->rx_opt.tstamp_ok &&
3980 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3981 if (tcp_parse_aligned_timestamp(tp, th))
3985 tcp_parse_options(skb, &tp->rx_opt, 1, NULL);
3986 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3987 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3992 #ifdef CONFIG_TCP_MD5SIG
3994 * Parse MD5 Signature option
3996 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3998 int length = (th->doff << 2) - sizeof(*th);
3999 const u8 *ptr = (const u8 *)(th + 1);
4001 /* If not enough data remaining, we can short cut */
4002 while (length >= TCPOLEN_MD5SIG) {
4003 int opcode = *ptr++;
4014 if (opsize < 2 || opsize > length)
4016 if (opcode == TCPOPT_MD5SIG)
4017 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
4024 EXPORT_SYMBOL(tcp_parse_md5sig_option);
4027 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
4029 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
4030 * it can pass through stack. So, the following predicate verifies that
4031 * this segment is not used for anything but congestion avoidance or
4032 * fast retransmit. Moreover, we even are able to eliminate most of such
4033 * second order effects, if we apply some small "replay" window (~RTO)
4034 * to timestamp space.
4036 * All these measures still do not guarantee that we reject wrapped ACKs
4037 * on networks with high bandwidth, when sequence space is recycled fastly,
4038 * but it guarantees that such events will be very rare and do not affect
4039 * connection seriously. This doesn't look nice, but alas, PAWS is really
4042 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
4043 * states that events when retransmit arrives after original data are rare.
4044 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
4045 * the biggest problem on large power networks even with minor reordering.
4046 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
4047 * up to bandwidth of 18Gigabit/sec. 8) ]
4050 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
4052 const struct tcp_sock *tp = tcp_sk(sk);
4053 const struct tcphdr *th = tcp_hdr(skb);
4054 u32 seq = TCP_SKB_CB(skb)->seq;
4055 u32 ack = TCP_SKB_CB(skb)->ack_seq;
4057 return (/* 1. Pure ACK with correct sequence number. */
4058 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
4060 /* 2. ... and duplicate ACK. */
4061 ack == tp->snd_una &&
4063 /* 3. ... and does not update window. */
4064 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
4066 /* 4. ... and sits in replay window. */
4067 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
4070 static inline bool tcp_paws_discard(const struct sock *sk,
4071 const struct sk_buff *skb)
4073 const struct tcp_sock *tp = tcp_sk(sk);
4075 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4076 !tcp_disordered_ack(sk, skb);
4079 /* Check segment sequence number for validity.
4081 * Segment controls are considered valid, if the segment
4082 * fits to the window after truncation to the window. Acceptability
4083 * of data (and SYN, FIN, of course) is checked separately.
4084 * See tcp_data_queue(), for example.
4086 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4087 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4088 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4089 * (borrowed from freebsd)
4092 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
4094 return !before(end_seq, tp->rcv_wup) &&
4095 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
4098 /* When we get a reset we do this. */
4099 void tcp_reset(struct sock *sk)
4101 /* We want the right error as BSD sees it (and indeed as we do). */
4102 switch (sk->sk_state) {
4104 sk->sk_err = ECONNREFUSED;
4106 case TCP_CLOSE_WAIT:
4112 sk->sk_err = ECONNRESET;
4114 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4117 if (!sock_flag(sk, SOCK_DEAD))
4118 sk->sk_error_report(sk);
4124 * Process the FIN bit. This now behaves as it is supposed to work
4125 * and the FIN takes effect when it is validly part of sequence
4126 * space. Not before when we get holes.
4128 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4129 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4132 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4133 * close and we go into CLOSING (and later onto TIME-WAIT)
4135 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4137 void tcp_fin(struct sock *sk)
4139 struct tcp_sock *tp = tcp_sk(sk);
4141 inet_csk_schedule_ack(sk);
4143 sk->sk_shutdown |= RCV_SHUTDOWN;
4144 sock_set_flag(sk, SOCK_DONE);
4146 switch (sk->sk_state) {
4148 case TCP_ESTABLISHED:
4149 /* Move to CLOSE_WAIT */
4150 tcp_set_state(sk, TCP_CLOSE_WAIT);
4151 inet_csk(sk)->icsk_ack.pingpong = 1;
4154 case TCP_CLOSE_WAIT:
4156 /* Received a retransmission of the FIN, do
4161 /* RFC793: Remain in the LAST-ACK state. */
4165 /* This case occurs when a simultaneous close
4166 * happens, we must ack the received FIN and
4167 * enter the CLOSING state.
4170 tcp_set_state(sk, TCP_CLOSING);
4173 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4175 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4178 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4179 * cases we should never reach this piece of code.
4181 pr_err("%s: Impossible, sk->sk_state=%d\n",
4182 __func__, sk->sk_state);
4186 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4187 * Probably, we should reset in this case. For now drop them.
4189 skb_rbtree_purge(&tp->out_of_order_queue);
4190 if (tcp_is_sack(tp))
4191 tcp_sack_reset(&tp->rx_opt);
4194 if (!sock_flag(sk, SOCK_DEAD)) {
4195 sk->sk_state_change(sk);
4197 /* Do not send POLL_HUP for half duplex close. */
4198 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4199 sk->sk_state == TCP_CLOSE)
4200 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4202 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4206 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4209 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4210 if (before(seq, sp->start_seq))
4211 sp->start_seq = seq;
4212 if (after(end_seq, sp->end_seq))
4213 sp->end_seq = end_seq;
4219 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4221 struct tcp_sock *tp = tcp_sk(sk);
4223 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4226 if (before(seq, tp->rcv_nxt))
4227 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4229 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4231 NET_INC_STATS(sock_net(sk), mib_idx);
4233 tp->rx_opt.dsack = 1;
4234 tp->duplicate_sack[0].start_seq = seq;
4235 tp->duplicate_sack[0].end_seq = end_seq;
4239 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4241 struct tcp_sock *tp = tcp_sk(sk);
4243 if (!tp->rx_opt.dsack)
4244 tcp_dsack_set(sk, seq, end_seq);
4246 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4249 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4251 struct tcp_sock *tp = tcp_sk(sk);
4253 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4254 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4255 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4256 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
4258 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4259 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4261 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4262 end_seq = tp->rcv_nxt;
4263 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4270 /* These routines update the SACK block as out-of-order packets arrive or
4271 * in-order packets close up the sequence space.
4273 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4276 struct tcp_sack_block *sp = &tp->selective_acks[0];
4277 struct tcp_sack_block *swalk = sp + 1;
4279 /* See if the recent change to the first SACK eats into
4280 * or hits the sequence space of other SACK blocks, if so coalesce.
4282 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4283 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4286 /* Zap SWALK, by moving every further SACK up by one slot.
4287 * Decrease num_sacks.
4289 tp->rx_opt.num_sacks--;
4290 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4294 this_sack++, swalk++;
4298 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4300 struct tcp_sock *tp = tcp_sk(sk);
4301 struct tcp_sack_block *sp = &tp->selective_acks[0];
4302 int cur_sacks = tp->rx_opt.num_sacks;
4308 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4309 if (tcp_sack_extend(sp, seq, end_seq)) {
4310 /* Rotate this_sack to the first one. */
4311 for (; this_sack > 0; this_sack--, sp--)
4312 swap(*sp, *(sp - 1));
4314 tcp_sack_maybe_coalesce(tp);
4319 /* Could not find an adjacent existing SACK, build a new one,
4320 * put it at the front, and shift everyone else down. We
4321 * always know there is at least one SACK present already here.
4323 * If the sack array is full, forget about the last one.
4325 if (this_sack >= TCP_NUM_SACKS) {
4327 tp->rx_opt.num_sacks--;
4330 for (; this_sack > 0; this_sack--, sp--)
4334 /* Build the new head SACK, and we're done. */
4335 sp->start_seq = seq;
4336 sp->end_seq = end_seq;
4337 tp->rx_opt.num_sacks++;
4340 /* RCV.NXT advances, some SACKs should be eaten. */
4342 static void tcp_sack_remove(struct tcp_sock *tp)
4344 struct tcp_sack_block *sp = &tp->selective_acks[0];
4345 int num_sacks = tp->rx_opt.num_sacks;
4348 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4349 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4350 tp->rx_opt.num_sacks = 0;
4354 for (this_sack = 0; this_sack < num_sacks;) {
4355 /* Check if the start of the sack is covered by RCV.NXT. */
4356 if (!before(tp->rcv_nxt, sp->start_seq)) {
4359 /* RCV.NXT must cover all the block! */
4360 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4362 /* Zap this SACK, by moving forward any other SACKS. */
4363 for (i = this_sack+1; i < num_sacks; i++)
4364 tp->selective_acks[i-1] = tp->selective_acks[i];
4371 tp->rx_opt.num_sacks = num_sacks;
4375 * tcp_try_coalesce - try to merge skb to prior one
4378 * @from: buffer to add in queue
4379 * @fragstolen: pointer to boolean
4381 * Before queueing skb @from after @to, try to merge them
4382 * to reduce overall memory use and queue lengths, if cost is small.
4383 * Packets in ofo or receive queues can stay a long time.
4384 * Better try to coalesce them right now to avoid future collapses.
4385 * Returns true if caller should free @from instead of queueing it
4387 static bool tcp_try_coalesce(struct sock *sk,
4389 struct sk_buff *from,
4394 *fragstolen = false;
4396 /* Its possible this segment overlaps with prior segment in queue */
4397 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4400 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4403 atomic_add(delta, &sk->sk_rmem_alloc);
4404 sk_mem_charge(sk, delta);
4405 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4406 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4407 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4408 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4412 static bool tcp_ooo_try_coalesce(struct sock *sk,
4414 struct sk_buff *from,
4417 bool res = tcp_try_coalesce(sk, to, from, fragstolen);
4419 /* In case tcp_drop() is called later, update to->gso_segs */
4421 u32 gso_segs = max_t(u16, 1, skb_shinfo(to)->gso_segs) +
4422 max_t(u16, 1, skb_shinfo(from)->gso_segs);
4424 skb_shinfo(to)->gso_segs = min_t(u32, gso_segs, 0xFFFF);
4429 static void tcp_drop(struct sock *sk, struct sk_buff *skb)
4431 sk_drops_add(sk, skb);
4435 /* This one checks to see if we can put data from the
4436 * out_of_order queue into the receive_queue.
4438 static void tcp_ofo_queue(struct sock *sk)
4440 struct tcp_sock *tp = tcp_sk(sk);
4441 __u32 dsack_high = tp->rcv_nxt;
4442 bool fin, fragstolen, eaten;
4443 struct sk_buff *skb, *tail;
4446 p = rb_first(&tp->out_of_order_queue);
4449 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4452 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4453 __u32 dsack = dsack_high;
4454 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4455 dsack_high = TCP_SKB_CB(skb)->end_seq;
4456 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4459 rb_erase(&skb->rbnode, &tp->out_of_order_queue);
4461 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) {
4462 SOCK_DEBUG(sk, "ofo packet was already received\n");
4466 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4467 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4468 TCP_SKB_CB(skb)->end_seq);
4470 tail = skb_peek_tail(&sk->sk_receive_queue);
4471 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4472 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4473 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
4475 __skb_queue_tail(&sk->sk_receive_queue, skb);
4477 kfree_skb_partial(skb, fragstolen);
4479 if (unlikely(fin)) {
4481 /* tcp_fin() purges tp->out_of_order_queue,
4482 * so we must end this loop right now.
4489 static bool tcp_prune_ofo_queue(struct sock *sk);
4490 static int tcp_prune_queue(struct sock *sk);
4492 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4495 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4496 !sk_rmem_schedule(sk, skb, size)) {
4498 if (tcp_prune_queue(sk) < 0)
4501 while (!sk_rmem_schedule(sk, skb, size)) {
4502 if (!tcp_prune_ofo_queue(sk))
4509 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4511 struct tcp_sock *tp = tcp_sk(sk);
4512 struct rb_node **p, *parent;
4513 struct sk_buff *skb1;
4517 tcp_ecn_check_ce(sk, skb);
4519 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4520 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP);
4525 /* Disable header prediction. */
4527 inet_csk_schedule_ack(sk);
4529 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4530 seq = TCP_SKB_CB(skb)->seq;
4531 end_seq = TCP_SKB_CB(skb)->end_seq;
4532 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4533 tp->rcv_nxt, seq, end_seq);
4535 p = &tp->out_of_order_queue.rb_node;
4536 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4537 /* Initial out of order segment, build 1 SACK. */
4538 if (tcp_is_sack(tp)) {
4539 tp->rx_opt.num_sacks = 1;
4540 tp->selective_acks[0].start_seq = seq;
4541 tp->selective_acks[0].end_seq = end_seq;
4543 rb_link_node(&skb->rbnode, NULL, p);
4544 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4545 tp->ooo_last_skb = skb;
4549 /* In the typical case, we are adding an skb to the end of the list.
4550 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4552 if (tcp_ooo_try_coalesce(sk, tp->ooo_last_skb,
4553 skb, &fragstolen)) {
4555 /* For non sack flows, do not grow window to force DUPACK
4556 * and trigger fast retransmit.
4558 if (tcp_is_sack(tp))
4559 tcp_grow_window(sk, skb);
4560 kfree_skb_partial(skb, fragstolen);
4564 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4565 if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) {
4566 parent = &tp->ooo_last_skb->rbnode;
4567 p = &parent->rb_right;
4571 /* Find place to insert this segment. Handle overlaps on the way. */
4575 skb1 = rb_to_skb(parent);
4576 if (before(seq, TCP_SKB_CB(skb1)->seq)) {
4577 p = &parent->rb_left;
4580 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4581 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4582 /* All the bits are present. Drop. */
4583 NET_INC_STATS(sock_net(sk),
4584 LINUX_MIB_TCPOFOMERGE);
4587 tcp_dsack_set(sk, seq, end_seq);
4590 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4591 /* Partial overlap. */
4592 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq);
4594 /* skb's seq == skb1's seq and skb covers skb1.
4595 * Replace skb1 with skb.
4597 rb_replace_node(&skb1->rbnode, &skb->rbnode,
4598 &tp->out_of_order_queue);
4599 tcp_dsack_extend(sk,
4600 TCP_SKB_CB(skb1)->seq,
4601 TCP_SKB_CB(skb1)->end_seq);
4602 NET_INC_STATS(sock_net(sk),
4603 LINUX_MIB_TCPOFOMERGE);
4607 } else if (tcp_ooo_try_coalesce(sk, skb1,
4608 skb, &fragstolen)) {
4611 p = &parent->rb_right;
4614 /* Insert segment into RB tree. */
4615 rb_link_node(&skb->rbnode, parent, p);
4616 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4619 /* Remove other segments covered by skb. */
4620 while ((skb1 = skb_rb_next(skb)) != NULL) {
4621 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4623 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4624 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4628 rb_erase(&skb1->rbnode, &tp->out_of_order_queue);
4629 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4630 TCP_SKB_CB(skb1)->end_seq);
4631 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4634 /* If there is no skb after us, we are the last_skb ! */
4636 tp->ooo_last_skb = skb;
4639 if (tcp_is_sack(tp))
4640 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4643 /* For non sack flows, do not grow window to force DUPACK
4644 * and trigger fast retransmit.
4646 if (tcp_is_sack(tp))
4647 tcp_grow_window(sk, skb);
4648 skb_set_owner_r(skb, sk);
4652 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
4656 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4658 __skb_pull(skb, hdrlen);
4660 tcp_try_coalesce(sk, tail, skb, fragstolen)) ? 1 : 0;
4661 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
4663 __skb_queue_tail(&sk->sk_receive_queue, skb);
4664 skb_set_owner_r(skb, sk);
4669 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4671 struct sk_buff *skb;
4679 if (size > PAGE_SIZE) {
4680 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
4682 data_len = npages << PAGE_SHIFT;
4683 size = data_len + (size & ~PAGE_MASK);
4685 skb = alloc_skb_with_frags(size - data_len, data_len,
4686 PAGE_ALLOC_COSTLY_ORDER,
4687 &err, sk->sk_allocation);
4691 skb_put(skb, size - data_len);
4692 skb->data_len = data_len;
4695 if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4698 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
4702 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4703 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4704 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4706 if (tcp_queue_rcv(sk, skb, 0, &fragstolen)) {
4707 WARN_ON_ONCE(fragstolen); /* should not happen */
4719 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4721 struct tcp_sock *tp = tcp_sk(sk);
4722 bool fragstolen = false;
4725 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
4730 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
4732 tcp_ecn_accept_cwr(tp, skb);
4734 tp->rx_opt.dsack = 0;
4736 /* Queue data for delivery to the user.
4737 * Packets in sequence go to the receive queue.
4738 * Out of sequence packets to the out_of_order_queue.
4740 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4741 if (tcp_receive_window(tp) == 0)
4744 /* Ok. In sequence. In window. */
4745 if (tp->ucopy.task == current &&
4746 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4747 sock_owned_by_user(sk) && !tp->urg_data) {
4748 int chunk = min_t(unsigned int, skb->len,
4751 __set_current_state(TASK_RUNNING);
4753 if (!skb_copy_datagram_msg(skb, 0, tp->ucopy.msg, chunk)) {
4754 tp->ucopy.len -= chunk;
4755 tp->copied_seq += chunk;
4756 eaten = (chunk == skb->len);
4757 tcp_rcv_space_adjust(sk);
4764 if (skb_queue_len(&sk->sk_receive_queue) == 0)
4765 sk_forced_mem_schedule(sk, skb->truesize);
4766 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4769 eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
4771 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4773 tcp_event_data_recv(sk, skb);
4774 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4777 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4780 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4781 * gap in queue is filled.
4783 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
4784 inet_csk(sk)->icsk_ack.pingpong = 0;
4787 if (tp->rx_opt.num_sacks)
4788 tcp_sack_remove(tp);
4790 tcp_fast_path_check(sk);
4793 kfree_skb_partial(skb, fragstolen);
4794 if (!sock_flag(sk, SOCK_DEAD))
4795 sk->sk_data_ready(sk);
4799 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4800 /* A retransmit, 2nd most common case. Force an immediate ack. */
4801 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4802 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4805 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
4806 inet_csk_schedule_ack(sk);
4812 /* Out of window. F.e. zero window probe. */
4813 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4816 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4817 /* Partial packet, seq < rcv_next < end_seq */
4818 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4819 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4820 TCP_SKB_CB(skb)->end_seq);
4822 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4824 /* If window is closed, drop tail of packet. But after
4825 * remembering D-SACK for its head made in previous line.
4827 if (!tcp_receive_window(tp))
4832 tcp_data_queue_ofo(sk, skb);
4835 static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list)
4838 return !skb_queue_is_last(list, skb) ? skb->next : NULL;
4840 return skb_rb_next(skb);
4843 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4844 struct sk_buff_head *list,
4845 struct rb_root *root)
4847 struct sk_buff *next = tcp_skb_next(skb, list);
4850 __skb_unlink(skb, list);
4852 rb_erase(&skb->rbnode, root);
4855 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4860 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
4861 static void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb)
4863 struct rb_node **p = &root->rb_node;
4864 struct rb_node *parent = NULL;
4865 struct sk_buff *skb1;
4869 skb1 = rb_to_skb(parent);
4870 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq))
4871 p = &parent->rb_left;
4873 p = &parent->rb_right;
4875 rb_link_node(&skb->rbnode, parent, p);
4876 rb_insert_color(&skb->rbnode, root);
4879 /* Collapse contiguous sequence of skbs head..tail with
4880 * sequence numbers start..end.
4882 * If tail is NULL, this means until the end of the queue.
4884 * Segments with FIN/SYN are not collapsed (only because this
4888 tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root,
4889 struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end)
4891 struct sk_buff *skb = head, *n;
4892 struct sk_buff_head tmp;
4895 /* First, check that queue is collapsible and find
4896 * the point where collapsing can be useful.
4899 for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) {
4900 n = tcp_skb_next(skb, list);
4902 /* No new bits? It is possible on ofo queue. */
4903 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4904 skb = tcp_collapse_one(sk, skb, list, root);
4910 /* The first skb to collapse is:
4912 * - bloated or contains data before "start" or
4913 * overlaps to the next one.
4915 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
4916 (tcp_win_from_space(skb->truesize) > skb->len ||
4917 before(TCP_SKB_CB(skb)->seq, start))) {
4918 end_of_skbs = false;
4922 if (n && n != tail &&
4923 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) {
4924 end_of_skbs = false;
4928 /* Decided to skip this, advance start seq. */
4929 start = TCP_SKB_CB(skb)->end_seq;
4932 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4935 __skb_queue_head_init(&tmp);
4937 while (before(start, end)) {
4938 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
4939 struct sk_buff *nskb;
4941 nskb = alloc_skb(copy, GFP_ATOMIC);
4945 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4946 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4948 __skb_queue_before(list, skb, nskb);
4950 __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */
4951 skb_set_owner_r(nskb, sk);
4953 /* Copy data, releasing collapsed skbs. */
4955 int offset = start - TCP_SKB_CB(skb)->seq;
4956 int size = TCP_SKB_CB(skb)->end_seq - start;
4960 size = min(copy, size);
4961 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4963 TCP_SKB_CB(nskb)->end_seq += size;
4967 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4968 skb = tcp_collapse_one(sk, skb, list, root);
4971 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4977 skb_queue_walk_safe(&tmp, skb, n)
4978 tcp_rbtree_insert(root, skb);
4981 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4982 * and tcp_collapse() them until all the queue is collapsed.
4984 static void tcp_collapse_ofo_queue(struct sock *sk)
4986 struct tcp_sock *tp = tcp_sk(sk);
4987 u32 range_truesize, sum_tiny = 0;
4988 struct sk_buff *skb, *head;
4991 skb = skb_rb_first(&tp->out_of_order_queue);
4994 tp->ooo_last_skb = skb_rb_last(&tp->out_of_order_queue);
4997 start = TCP_SKB_CB(skb)->seq;
4998 end = TCP_SKB_CB(skb)->end_seq;
4999 range_truesize = skb->truesize;
5001 for (head = skb;;) {
5002 skb = skb_rb_next(skb);
5004 /* Range is terminated when we see a gap or when
5005 * we are at the queue end.
5008 after(TCP_SKB_CB(skb)->seq, end) ||
5009 before(TCP_SKB_CB(skb)->end_seq, start)) {
5010 /* Do not attempt collapsing tiny skbs */
5011 if (range_truesize != head->truesize ||
5012 end - start >= SKB_WITH_OVERHEAD(SK_MEM_QUANTUM)) {
5013 tcp_collapse(sk, NULL, &tp->out_of_order_queue,
5014 head, skb, start, end);
5016 sum_tiny += range_truesize;
5017 if (sum_tiny > sk->sk_rcvbuf >> 3)
5023 range_truesize += skb->truesize;
5024 if (unlikely(before(TCP_SKB_CB(skb)->seq, start)))
5025 start = TCP_SKB_CB(skb)->seq;
5026 if (after(TCP_SKB_CB(skb)->end_seq, end))
5027 end = TCP_SKB_CB(skb)->end_seq;
5032 * Clean the out-of-order queue to make room.
5033 * We drop high sequences packets to :
5034 * 1) Let a chance for holes to be filled.
5035 * 2) not add too big latencies if thousands of packets sit there.
5036 * (But if application shrinks SO_RCVBUF, we could still end up
5037 * freeing whole queue here)
5038 * 3) Drop at least 12.5 % of sk_rcvbuf to avoid malicious attacks.
5040 * Return true if queue has shrunk.
5042 static bool tcp_prune_ofo_queue(struct sock *sk)
5044 struct tcp_sock *tp = tcp_sk(sk);
5045 struct rb_node *node, *prev;
5048 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
5051 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED);
5052 goal = sk->sk_rcvbuf >> 3;
5053 node = &tp->ooo_last_skb->rbnode;
5055 prev = rb_prev(node);
5056 rb_erase(node, &tp->out_of_order_queue);
5057 goal -= rb_to_skb(node)->truesize;
5058 tcp_drop(sk, rb_to_skb(node));
5059 if (!prev || goal <= 0) {
5061 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf &&
5062 !tcp_under_memory_pressure(sk))
5064 goal = sk->sk_rcvbuf >> 3;
5068 tp->ooo_last_skb = rb_to_skb(prev);
5070 /* Reset SACK state. A conforming SACK implementation will
5071 * do the same at a timeout based retransmit. When a connection
5072 * is in a sad state like this, we care only about integrity
5073 * of the connection not performance.
5075 if (tp->rx_opt.sack_ok)
5076 tcp_sack_reset(&tp->rx_opt);
5080 /* Reduce allocated memory if we can, trying to get
5081 * the socket within its memory limits again.
5083 * Return less than zero if we should start dropping frames
5084 * until the socket owning process reads some of the data
5085 * to stabilize the situation.
5087 static int tcp_prune_queue(struct sock *sk)
5089 struct tcp_sock *tp = tcp_sk(sk);
5091 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
5093 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED);
5095 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
5096 tcp_clamp_window(sk);
5097 else if (tcp_under_memory_pressure(sk))
5098 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
5100 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5103 tcp_collapse_ofo_queue(sk);
5104 if (!skb_queue_empty(&sk->sk_receive_queue))
5105 tcp_collapse(sk, &sk->sk_receive_queue, NULL,
5106 skb_peek(&sk->sk_receive_queue),
5108 tp->copied_seq, tp->rcv_nxt);
5111 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5114 /* Collapsing did not help, destructive actions follow.
5115 * This must not ever occur. */
5117 tcp_prune_ofo_queue(sk);
5119 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5122 /* If we are really being abused, tell the caller to silently
5123 * drop receive data on the floor. It will get retransmitted
5124 * and hopefully then we'll have sufficient space.
5126 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED);
5128 /* Massive buffer overcommit. */
5133 static bool tcp_should_expand_sndbuf(const struct sock *sk)
5135 const struct tcp_sock *tp = tcp_sk(sk);
5137 /* If the user specified a specific send buffer setting, do
5140 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
5143 /* If we are under global TCP memory pressure, do not expand. */
5144 if (tcp_under_memory_pressure(sk))
5147 /* If we are under soft global TCP memory pressure, do not expand. */
5148 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
5151 /* If we filled the congestion window, do not expand. */
5152 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
5158 /* When incoming ACK allowed to free some skb from write_queue,
5159 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
5160 * on the exit from tcp input handler.
5162 * PROBLEM: sndbuf expansion does not work well with largesend.
5164 static void tcp_new_space(struct sock *sk)
5166 struct tcp_sock *tp = tcp_sk(sk);
5168 if (tcp_should_expand_sndbuf(sk)) {
5169 tcp_sndbuf_expand(sk);
5170 tp->snd_cwnd_stamp = tcp_time_stamp;
5173 sk->sk_write_space(sk);
5176 /* Caller made space either from:
5177 * 1) Freeing skbs in rtx queues (after tp->snd_una has advanced)
5178 * 2) Sent skbs from output queue (and thus advancing tp->snd_nxt)
5180 * We might be able to generate EPOLLOUT to the application if:
5181 * 1) Space consumed in output/rtx queues is below sk->sk_sndbuf/2
5182 * 2) notsent amount (tp->write_seq - tp->snd_nxt) became
5183 * small enough that tcp_stream_memory_free() decides it
5184 * is time to generate EPOLLOUT.
5186 void tcp_check_space(struct sock *sk)
5188 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
5189 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
5190 /* pairs with tcp_poll() */
5192 if (sk->sk_socket &&
5193 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5198 static inline void tcp_data_snd_check(struct sock *sk)
5200 tcp_push_pending_frames(sk);
5201 tcp_check_space(sk);
5205 * Check if sending an ack is needed.
5207 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5209 struct tcp_sock *tp = tcp_sk(sk);
5211 /* More than one full frame received... */
5212 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5213 /* ... and right edge of window advances far enough.
5214 * (tcp_recvmsg() will send ACK otherwise). Or...
5216 __tcp_select_window(sk) >= tp->rcv_wnd) ||
5217 /* We ACK each frame or... */
5218 tcp_in_quickack_mode(sk) ||
5219 /* We have out of order data. */
5220 (ofo_possible && !RB_EMPTY_ROOT(&tp->out_of_order_queue))) {
5221 /* Then ack it now */
5224 /* Else, send delayed ack. */
5225 tcp_send_delayed_ack(sk);
5229 static inline void tcp_ack_snd_check(struct sock *sk)
5231 if (!inet_csk_ack_scheduled(sk)) {
5232 /* We sent a data segment already. */
5235 __tcp_ack_snd_check(sk, 1);
5239 * This routine is only called when we have urgent data
5240 * signaled. Its the 'slow' part of tcp_urg. It could be
5241 * moved inline now as tcp_urg is only called from one
5242 * place. We handle URGent data wrong. We have to - as
5243 * BSD still doesn't use the correction from RFC961.
5244 * For 1003.1g we should support a new option TCP_STDURG to permit
5245 * either form (or just set the sysctl tcp_stdurg).
5248 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5250 struct tcp_sock *tp = tcp_sk(sk);
5251 u32 ptr = ntohs(th->urg_ptr);
5253 if (ptr && !sysctl_tcp_stdurg)
5255 ptr += ntohl(th->seq);
5257 /* Ignore urgent data that we've already seen and read. */
5258 if (after(tp->copied_seq, ptr))
5261 /* Do not replay urg ptr.
5263 * NOTE: interesting situation not covered by specs.
5264 * Misbehaving sender may send urg ptr, pointing to segment,
5265 * which we already have in ofo queue. We are not able to fetch
5266 * such data and will stay in TCP_URG_NOTYET until will be eaten
5267 * by recvmsg(). Seems, we are not obliged to handle such wicked
5268 * situations. But it is worth to think about possibility of some
5269 * DoSes using some hypothetical application level deadlock.
5271 if (before(ptr, tp->rcv_nxt))
5274 /* Do we already have a newer (or duplicate) urgent pointer? */
5275 if (tp->urg_data && !after(ptr, tp->urg_seq))
5278 /* Tell the world about our new urgent pointer. */
5281 /* We may be adding urgent data when the last byte read was
5282 * urgent. To do this requires some care. We cannot just ignore
5283 * tp->copied_seq since we would read the last urgent byte again
5284 * as data, nor can we alter copied_seq until this data arrives
5285 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5287 * NOTE. Double Dutch. Rendering to plain English: author of comment
5288 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5289 * and expect that both A and B disappear from stream. This is _wrong_.
5290 * Though this happens in BSD with high probability, this is occasional.
5291 * Any application relying on this is buggy. Note also, that fix "works"
5292 * only in this artificial test. Insert some normal data between A and B and we will
5293 * decline of BSD again. Verdict: it is better to remove to trap
5296 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5297 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5298 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5300 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5301 __skb_unlink(skb, &sk->sk_receive_queue);
5306 tp->urg_data = TCP_URG_NOTYET;
5309 /* Disable header prediction. */
5313 /* This is the 'fast' part of urgent handling. */
5314 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5316 struct tcp_sock *tp = tcp_sk(sk);
5318 /* Check if we get a new urgent pointer - normally not. */
5320 tcp_check_urg(sk, th);
5322 /* Do we wait for any urgent data? - normally not... */
5323 if (tp->urg_data == TCP_URG_NOTYET) {
5324 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5327 /* Is the urgent pointer pointing into this packet? */
5328 if (ptr < skb->len) {
5330 if (skb_copy_bits(skb, ptr, &tmp, 1))
5332 tp->urg_data = TCP_URG_VALID | tmp;
5333 if (!sock_flag(sk, SOCK_DEAD))
5334 sk->sk_data_ready(sk);
5339 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
5341 struct tcp_sock *tp = tcp_sk(sk);
5342 int chunk = skb->len - hlen;
5345 if (skb_csum_unnecessary(skb))
5346 err = skb_copy_datagram_msg(skb, hlen, tp->ucopy.msg, chunk);
5348 err = skb_copy_and_csum_datagram_msg(skb, hlen, tp->ucopy.msg);
5351 tp->ucopy.len -= chunk;
5352 tp->copied_seq += chunk;
5353 tcp_rcv_space_adjust(sk);
5359 /* Does PAWS and seqno based validation of an incoming segment, flags will
5360 * play significant role here.
5362 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5363 const struct tcphdr *th, int syn_inerr)
5365 struct tcp_sock *tp = tcp_sk(sk);
5366 bool rst_seq_match = false;
5368 /* RFC1323: H1. Apply PAWS check first. */
5369 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
5370 tcp_paws_discard(sk, skb)) {
5372 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5373 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5374 LINUX_MIB_TCPACKSKIPPEDPAWS,
5375 &tp->last_oow_ack_time))
5376 tcp_send_dupack(sk, skb);
5379 /* Reset is accepted even if it did not pass PAWS. */
5382 /* Step 1: check sequence number */
5383 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5384 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5385 * (RST) segments are validated by checking their SEQ-fields."
5386 * And page 69: "If an incoming segment is not acceptable,
5387 * an acknowledgment should be sent in reply (unless the RST
5388 * bit is set, if so drop the segment and return)".
5393 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5394 LINUX_MIB_TCPACKSKIPPEDSEQ,
5395 &tp->last_oow_ack_time))
5396 tcp_send_dupack(sk, skb);
5401 /* Step 2: check RST bit */
5403 /* RFC 5961 3.2 (extend to match against SACK too if available):
5404 * If seq num matches RCV.NXT or the right-most SACK block,
5406 * RESET the connection
5408 * Send a challenge ACK
5410 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
5411 rst_seq_match = true;
5412 } else if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) {
5413 struct tcp_sack_block *sp = &tp->selective_acks[0];
5414 int max_sack = sp[0].end_seq;
5417 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;
5419 max_sack = after(sp[this_sack].end_seq,
5421 sp[this_sack].end_seq : max_sack;
5424 if (TCP_SKB_CB(skb)->seq == max_sack)
5425 rst_seq_match = true;
5431 tcp_send_challenge_ack(sk, skb);
5435 /* step 3: check security and precedence [ignored] */
5437 /* step 4: Check for a SYN
5438 * RFC 5961 4.2 : Send a challenge ack
5443 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5444 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5445 tcp_send_challenge_ack(sk, skb);
5457 * TCP receive function for the ESTABLISHED state.
5459 * It is split into a fast path and a slow path. The fast path is
5461 * - A zero window was announced from us - zero window probing
5462 * is only handled properly in the slow path.
5463 * - Out of order segments arrived.
5464 * - Urgent data is expected.
5465 * - There is no buffer space left
5466 * - Unexpected TCP flags/window values/header lengths are received
5467 * (detected by checking the TCP header against pred_flags)
5468 * - Data is sent in both directions. Fast path only supports pure senders
5469 * or pure receivers (this means either the sequence number or the ack
5470 * value must stay constant)
5471 * - Unexpected TCP option.
5473 * When these conditions are not satisfied it drops into a standard
5474 * receive procedure patterned after RFC793 to handle all cases.
5475 * The first three cases are guaranteed by proper pred_flags setting,
5476 * the rest is checked inline. Fast processing is turned on in
5477 * tcp_data_queue when everything is OK.
5479 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5480 const struct tcphdr *th, unsigned int len)
5482 struct tcp_sock *tp = tcp_sk(sk);
5484 if (unlikely(!sk->sk_rx_dst))
5485 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5487 * Header prediction.
5488 * The code loosely follows the one in the famous
5489 * "30 instruction TCP receive" Van Jacobson mail.
5491 * Van's trick is to deposit buffers into socket queue
5492 * on a device interrupt, to call tcp_recv function
5493 * on the receive process context and checksum and copy
5494 * the buffer to user space. smart...
5496 * Our current scheme is not silly either but we take the
5497 * extra cost of the net_bh soft interrupt processing...
5498 * We do checksum and copy also but from device to kernel.
5501 tp->rx_opt.saw_tstamp = 0;
5503 /* pred_flags is 0xS?10 << 16 + snd_wnd
5504 * if header_prediction is to be made
5505 * 'S' will always be tp->tcp_header_len >> 2
5506 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5507 * turn it off (when there are holes in the receive
5508 * space for instance)
5509 * PSH flag is ignored.
5512 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5513 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5514 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5515 int tcp_header_len = tp->tcp_header_len;
5517 /* Timestamp header prediction: tcp_header_len
5518 * is automatically equal to th->doff*4 due to pred_flags
5522 /* Check timestamp */
5523 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5524 /* No? Slow path! */
5525 if (!tcp_parse_aligned_timestamp(tp, th))
5528 /* If PAWS failed, check it more carefully in slow path */
5529 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5532 /* DO NOT update ts_recent here, if checksum fails
5533 * and timestamp was corrupted part, it will result
5534 * in a hung connection since we will drop all
5535 * future packets due to the PAWS test.
5539 if (len <= tcp_header_len) {
5540 /* Bulk data transfer: sender */
5541 if (len == tcp_header_len) {
5542 /* Predicted packet is in window by definition.
5543 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5544 * Hence, check seq<=rcv_wup reduces to:
5546 if (tcp_header_len ==
5547 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5548 tp->rcv_nxt == tp->rcv_wup)
5549 tcp_store_ts_recent(tp);
5551 /* We know that such packets are checksummed
5554 tcp_ack(sk, skb, 0);
5556 tcp_data_snd_check(sk);
5558 } else { /* Header too small */
5559 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5564 bool fragstolen = false;
5566 if (tp->ucopy.task == current &&
5567 tp->copied_seq == tp->rcv_nxt &&
5568 len - tcp_header_len <= tp->ucopy.len &&
5569 sock_owned_by_user(sk)) {
5570 __set_current_state(TASK_RUNNING);
5572 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) {
5573 /* Predicted packet is in window by definition.
5574 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5575 * Hence, check seq<=rcv_wup reduces to:
5577 if (tcp_header_len ==
5578 (sizeof(struct tcphdr) +
5579 TCPOLEN_TSTAMP_ALIGNED) &&
5580 tp->rcv_nxt == tp->rcv_wup)
5581 tcp_store_ts_recent(tp);
5583 tcp_rcv_rtt_measure_ts(sk, skb);
5585 __skb_pull(skb, tcp_header_len);
5586 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
5587 NET_INC_STATS(sock_net(sk),
5588 LINUX_MIB_TCPHPHITSTOUSER);
5593 if (tcp_checksum_complete(skb))
5596 if ((int)skb->truesize > sk->sk_forward_alloc)
5599 /* Predicted packet is in window by definition.
5600 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5601 * Hence, check seq<=rcv_wup reduces to:
5603 if (tcp_header_len ==
5604 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5605 tp->rcv_nxt == tp->rcv_wup)
5606 tcp_store_ts_recent(tp);
5608 tcp_rcv_rtt_measure_ts(sk, skb);
5610 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS);
5612 /* Bulk data transfer: receiver */
5613 eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5617 tcp_event_data_recv(sk, skb);
5619 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5620 /* Well, only one small jumplet in fast path... */
5621 tcp_ack(sk, skb, FLAG_DATA);
5622 tcp_data_snd_check(sk);
5623 if (!inet_csk_ack_scheduled(sk))
5626 tcp_update_wl(tp, TCP_SKB_CB(skb)->seq);
5629 __tcp_ack_snd_check(sk, 0);
5632 kfree_skb_partial(skb, fragstolen);
5633 sk->sk_data_ready(sk);
5639 if (len < (th->doff << 2) || tcp_checksum_complete(skb))
5642 if (!th->ack && !th->rst && !th->syn)
5646 * Standard slow path.
5649 if (!tcp_validate_incoming(sk, skb, th, 1))
5653 if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5656 tcp_rcv_rtt_measure_ts(sk, skb);
5658 /* Process urgent data. */
5659 tcp_urg(sk, skb, th);
5661 /* step 7: process the segment text */
5662 tcp_data_queue(sk, skb);
5664 tcp_data_snd_check(sk);
5665 tcp_ack_snd_check(sk);
5669 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
5670 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5675 EXPORT_SYMBOL(tcp_rcv_established);
5677 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5679 struct tcp_sock *tp = tcp_sk(sk);
5680 struct inet_connection_sock *icsk = inet_csk(sk);
5682 tcp_set_state(sk, TCP_ESTABLISHED);
5683 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5686 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5687 security_inet_conn_established(sk, skb);
5690 /* Make sure socket is routed, for correct metrics. */
5691 icsk->icsk_af_ops->rebuild_header(sk);
5693 tcp_init_metrics(sk);
5695 tcp_init_congestion_control(sk);
5697 /* Prevent spurious tcp_cwnd_restart() on first data
5700 tp->lsndtime = tcp_time_stamp;
5702 tcp_init_buffer_space(sk);
5704 if (sock_flag(sk, SOCK_KEEPOPEN))
5705 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5707 if (!tp->rx_opt.snd_wscale)
5708 __tcp_fast_path_on(tp, tp->snd_wnd);
5714 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5715 struct tcp_fastopen_cookie *cookie)
5717 struct tcp_sock *tp = tcp_sk(sk);
5718 struct sk_buff *data = tp->syn_data ? tcp_write_queue_head(sk) : NULL;
5719 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
5720 bool syn_drop = false;
5722 if (mss == tp->rx_opt.user_mss) {
5723 struct tcp_options_received opt;
5725 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5726 tcp_clear_options(&opt);
5727 opt.user_mss = opt.mss_clamp = 0;
5728 tcp_parse_options(synack, &opt, 0, NULL);
5729 mss = opt.mss_clamp;
5732 if (!tp->syn_fastopen) {
5733 /* Ignore an unsolicited cookie */
5735 } else if (tp->total_retrans) {
5736 /* SYN timed out and the SYN-ACK neither has a cookie nor
5737 * acknowledges data. Presumably the remote received only
5738 * the retransmitted (regular) SYNs: either the original
5739 * SYN-data or the corresponding SYN-ACK was dropped.
5741 syn_drop = (cookie->len < 0 && data);
5742 } else if (cookie->len < 0 && !tp->syn_data) {
5743 /* We requested a cookie but didn't get it. If we did not use
5744 * the (old) exp opt format then try so next time (try_exp=1).
5745 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5747 try_exp = tp->syn_fastopen_exp ? 2 : 1;
5750 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
5752 if (data) { /* Retransmit unacked data in SYN */
5753 tcp_for_write_queue_from(data, sk) {
5754 if (data == tcp_send_head(sk) ||
5755 __tcp_retransmit_skb(sk, data, 1))
5759 NET_INC_STATS(sock_net(sk),
5760 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
5763 tp->syn_data_acked = tp->syn_data;
5764 if (tp->syn_data_acked)
5765 NET_INC_STATS(sock_net(sk),
5766 LINUX_MIB_TCPFASTOPENACTIVE);
5768 tcp_fastopen_add_skb(sk, synack);
5773 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5774 const struct tcphdr *th)
5776 struct inet_connection_sock *icsk = inet_csk(sk);
5777 struct tcp_sock *tp = tcp_sk(sk);
5778 struct tcp_fastopen_cookie foc = { .len = -1 };
5779 int saved_clamp = tp->rx_opt.mss_clamp;
5782 tcp_parse_options(skb, &tp->rx_opt, 0, &foc);
5783 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
5784 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
5788 * "If the state is SYN-SENT then
5789 * first check the ACK bit
5790 * If the ACK bit is set
5791 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5792 * a reset (unless the RST bit is set, if so drop
5793 * the segment and return)"
5795 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5796 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
5797 goto reset_and_undo;
5799 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5800 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5802 NET_INC_STATS(sock_net(sk),
5803 LINUX_MIB_PAWSACTIVEREJECTED);
5804 goto reset_and_undo;
5807 /* Now ACK is acceptable.
5809 * "If the RST bit is set
5810 * If the ACK was acceptable then signal the user "error:
5811 * connection reset", drop the segment, enter CLOSED state,
5812 * delete TCB, and return."
5821 * "fifth, if neither of the SYN or RST bits is set then
5822 * drop the segment and return."
5828 goto discard_and_undo;
5831 * "If the SYN bit is on ...
5832 * are acceptable then ...
5833 * (our SYN has been ACKed), change the connection
5834 * state to ESTABLISHED..."
5837 tcp_ecn_rcv_synack(tp, th);
5839 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5840 tcp_ack(sk, skb, FLAG_SLOWPATH);
5842 /* Ok.. it's good. Set up sequence numbers and
5843 * move to established.
5845 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5846 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5848 /* RFC1323: The window in SYN & SYN/ACK segments is
5851 tp->snd_wnd = ntohs(th->window);
5853 if (!tp->rx_opt.wscale_ok) {
5854 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5855 tp->window_clamp = min(tp->window_clamp, 65535U);
5858 if (tp->rx_opt.saw_tstamp) {
5859 tp->rx_opt.tstamp_ok = 1;
5860 tp->tcp_header_len =
5861 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5862 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5863 tcp_store_ts_recent(tp);
5865 tp->tcp_header_len = sizeof(struct tcphdr);
5868 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5869 tcp_enable_fack(tp);
5872 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5873 tcp_initialize_rcv_mss(sk);
5875 /* Remember, tcp_poll() does not lock socket!
5876 * Change state from SYN-SENT only after copied_seq
5877 * is initialized. */
5878 tp->copied_seq = tp->rcv_nxt;
5882 tcp_finish_connect(sk, skb);
5884 fastopen_fail = (tp->syn_fastopen || tp->syn_data) &&
5885 tcp_rcv_fastopen_synack(sk, skb, &foc);
5887 if (!sock_flag(sk, SOCK_DEAD)) {
5888 sk->sk_state_change(sk);
5889 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5893 if (sk->sk_write_pending ||
5894 icsk->icsk_accept_queue.rskq_defer_accept ||
5895 icsk->icsk_ack.pingpong) {
5896 /* Save one ACK. Data will be ready after
5897 * several ticks, if write_pending is set.
5899 * It may be deleted, but with this feature tcpdumps
5900 * look so _wonderfully_ clever, that I was not able
5901 * to stand against the temptation 8) --ANK
5903 inet_csk_schedule_ack(sk);
5904 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
5905 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5906 TCP_DELACK_MAX, TCP_RTO_MAX);
5917 /* No ACK in the segment */
5921 * "If the RST bit is set
5923 * Otherwise (no ACK) drop the segment and return."
5926 goto discard_and_undo;
5930 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5931 tcp_paws_reject(&tp->rx_opt, 0))
5932 goto discard_and_undo;
5935 /* We see SYN without ACK. It is attempt of
5936 * simultaneous connect with crossed SYNs.
5937 * Particularly, it can be connect to self.
5939 tcp_set_state(sk, TCP_SYN_RECV);
5941 if (tp->rx_opt.saw_tstamp) {
5942 tp->rx_opt.tstamp_ok = 1;
5943 tcp_store_ts_recent(tp);
5944 tp->tcp_header_len =
5945 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5947 tp->tcp_header_len = sizeof(struct tcphdr);
5950 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5951 tp->copied_seq = tp->rcv_nxt;
5952 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5954 /* RFC1323: The window in SYN & SYN/ACK segments is
5957 tp->snd_wnd = ntohs(th->window);
5958 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5959 tp->max_window = tp->snd_wnd;
5961 tcp_ecn_rcv_syn(tp, th);
5964 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5965 tcp_initialize_rcv_mss(sk);
5967 tcp_send_synack(sk);
5969 /* Note, we could accept data and URG from this segment.
5970 * There are no obstacles to make this (except that we must
5971 * either change tcp_recvmsg() to prevent it from returning data
5972 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5974 * However, if we ignore data in ACKless segments sometimes,
5975 * we have no reasons to accept it sometimes.
5976 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5977 * is not flawless. So, discard packet for sanity.
5978 * Uncomment this return to process the data.
5985 /* "fifth, if neither of the SYN or RST bits is set then
5986 * drop the segment and return."
5990 tcp_clear_options(&tp->rx_opt);
5991 tp->rx_opt.mss_clamp = saved_clamp;
5995 tcp_clear_options(&tp->rx_opt);
5996 tp->rx_opt.mss_clamp = saved_clamp;
6001 * This function implements the receiving procedure of RFC 793 for
6002 * all states except ESTABLISHED and TIME_WAIT.
6003 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
6004 * address independent.
6007 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
6009 struct tcp_sock *tp = tcp_sk(sk);
6010 struct inet_connection_sock *icsk = inet_csk(sk);
6011 const struct tcphdr *th = tcp_hdr(skb);
6012 struct request_sock *req;
6016 switch (sk->sk_state) {
6030 /* It is possible that we process SYN packets from backlog,
6031 * so we need to make sure to disable BH and RCU right there.
6035 acceptable = icsk->icsk_af_ops->conn_request(sk, skb) >= 0;
6047 tp->rx_opt.saw_tstamp = 0;
6048 queued = tcp_rcv_synsent_state_process(sk, skb, th);
6052 /* Do step6 onward by hand. */
6053 tcp_urg(sk, skb, th);
6055 tcp_data_snd_check(sk);
6059 tp->rx_opt.saw_tstamp = 0;
6060 req = tp->fastopen_rsk;
6062 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
6063 sk->sk_state != TCP_FIN_WAIT1);
6065 if (!tcp_check_req(sk, skb, req, true))
6069 if (!th->ack && !th->rst && !th->syn)
6072 if (!tcp_validate_incoming(sk, skb, th, 0))
6075 /* step 5: check the ACK field */
6076 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
6077 FLAG_UPDATE_TS_RECENT |
6078 FLAG_NO_CHALLENGE_ACK) > 0;
6081 if (sk->sk_state == TCP_SYN_RECV)
6082 return 1; /* send one RST */
6083 tcp_send_challenge_ack(sk, skb);
6086 switch (sk->sk_state) {
6089 tcp_synack_rtt_meas(sk, req);
6091 /* Once we leave TCP_SYN_RECV, we no longer need req
6095 inet_csk(sk)->icsk_retransmits = 0;
6096 reqsk_fastopen_remove(sk, req, false);
6098 /* Make sure socket is routed, for correct metrics. */
6099 icsk->icsk_af_ops->rebuild_header(sk);
6100 tcp_init_congestion_control(sk);
6103 tp->copied_seq = tp->rcv_nxt;
6104 tcp_init_buffer_space(sk);
6107 tcp_set_state(sk, TCP_ESTABLISHED);
6108 sk->sk_state_change(sk);
6110 /* Note, that this wakeup is only for marginal crossed SYN case.
6111 * Passively open sockets are not waked up, because
6112 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
6115 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6117 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
6118 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
6119 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6121 if (tp->rx_opt.tstamp_ok)
6122 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6125 /* Re-arm the timer because data may have been sent out.
6126 * This is similar to the regular data transmission case
6127 * when new data has just been ack'ed.
6129 * (TFO) - we could try to be more aggressive and
6130 * retransmitting any data sooner based on when they
6135 tcp_init_metrics(sk);
6137 if (!inet_csk(sk)->icsk_ca_ops->cong_control)
6138 tcp_update_pacing_rate(sk);
6140 /* Prevent spurious tcp_cwnd_restart() on first data packet */
6141 tp->lsndtime = tcp_time_stamp;
6143 tcp_initialize_rcv_mss(sk);
6144 tcp_fast_path_on(tp);
6147 case TCP_FIN_WAIT1: {
6148 struct dst_entry *dst;
6151 /* If we enter the TCP_FIN_WAIT1 state and we are a
6152 * Fast Open socket and this is the first acceptable
6153 * ACK we have received, this would have acknowledged
6154 * our SYNACK so stop the SYNACK timer.
6157 /* We no longer need the request sock. */
6158 reqsk_fastopen_remove(sk, req, false);
6161 if (tp->snd_una != tp->write_seq)
6164 tcp_set_state(sk, TCP_FIN_WAIT2);
6165 sk->sk_shutdown |= SEND_SHUTDOWN;
6167 dst = __sk_dst_get(sk);
6171 if (!sock_flag(sk, SOCK_DEAD)) {
6172 /* Wake up lingering close() */
6173 sk->sk_state_change(sk);
6177 if (tp->linger2 < 0 ||
6178 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6179 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
6181 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6185 tmo = tcp_fin_time(sk);
6186 if (tmo > TCP_TIMEWAIT_LEN) {
6187 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
6188 } else if (th->fin || sock_owned_by_user(sk)) {
6189 /* Bad case. We could lose such FIN otherwise.
6190 * It is not a big problem, but it looks confusing
6191 * and not so rare event. We still can lose it now,
6192 * if it spins in bh_lock_sock(), but it is really
6195 inet_csk_reset_keepalive_timer(sk, tmo);
6197 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
6204 if (tp->snd_una == tp->write_seq) {
6205 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6211 if (tp->snd_una == tp->write_seq) {
6212 tcp_update_metrics(sk);
6219 /* step 6: check the URG bit */
6220 tcp_urg(sk, skb, th);
6222 /* step 7: process the segment text */
6223 switch (sk->sk_state) {
6224 case TCP_CLOSE_WAIT:
6227 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
6231 /* RFC 793 says to queue data in these states,
6232 * RFC 1122 says we MUST send a reset.
6233 * BSD 4.4 also does reset.
6235 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6236 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6237 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6238 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6244 case TCP_ESTABLISHED:
6245 tcp_data_queue(sk, skb);
6250 /* tcp_data could move socket to TIME-WAIT */
6251 if (sk->sk_state != TCP_CLOSE) {
6252 tcp_data_snd_check(sk);
6253 tcp_ack_snd_check(sk);
6262 EXPORT_SYMBOL(tcp_rcv_state_process);
6264 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
6266 struct inet_request_sock *ireq = inet_rsk(req);
6268 if (family == AF_INET)
6269 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6270 &ireq->ir_rmt_addr, port);
6271 #if IS_ENABLED(CONFIG_IPV6)
6272 else if (family == AF_INET6)
6273 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6274 &ireq->ir_v6_rmt_addr, port);
6278 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6280 * If we receive a SYN packet with these bits set, it means a
6281 * network is playing bad games with TOS bits. In order to
6282 * avoid possible false congestion notifications, we disable
6283 * TCP ECN negotiation.
6285 * Exception: tcp_ca wants ECN. This is required for DCTCP
6286 * congestion control: Linux DCTCP asserts ECT on all packets,
6287 * including SYN, which is most optimal solution; however,
6288 * others, such as FreeBSD do not.
6290 static void tcp_ecn_create_request(struct request_sock *req,
6291 const struct sk_buff *skb,
6292 const struct sock *listen_sk,
6293 const struct dst_entry *dst)
6295 const struct tcphdr *th = tcp_hdr(skb);
6296 const struct net *net = sock_net(listen_sk);
6297 bool th_ecn = th->ece && th->cwr;
6304 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
6305 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK);
6306 ecn_ok = net->ipv4.sysctl_tcp_ecn || ecn_ok_dst;
6308 if ((!ect && ecn_ok) || tcp_ca_needs_ecn(listen_sk) ||
6309 (ecn_ok_dst & DST_FEATURE_ECN_CA))
6310 inet_rsk(req)->ecn_ok = 1;
6313 static void tcp_openreq_init(struct request_sock *req,
6314 const struct tcp_options_received *rx_opt,
6315 struct sk_buff *skb, const struct sock *sk)
6317 struct inet_request_sock *ireq = inet_rsk(req);
6319 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */
6321 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6322 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
6323 skb_mstamp_get(&tcp_rsk(req)->snt_synack);
6324 tcp_rsk(req)->last_oow_ack_time = 0;
6325 req->mss = rx_opt->mss_clamp;
6326 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6327 ireq->tstamp_ok = rx_opt->tstamp_ok;
6328 ireq->sack_ok = rx_opt->sack_ok;
6329 ireq->snd_wscale = rx_opt->snd_wscale;
6330 ireq->wscale_ok = rx_opt->wscale_ok;
6333 ireq->ir_rmt_port = tcp_hdr(skb)->source;
6334 ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6335 ireq->ir_mark = inet_request_mark(sk, skb);
6338 struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
6339 struct sock *sk_listener,
6340 bool attach_listener)
6342 struct request_sock *req = reqsk_alloc(ops, sk_listener,
6346 struct inet_request_sock *ireq = inet_rsk(req);
6348 kmemcheck_annotate_bitfield(ireq, flags);
6349 ireq->ireq_opt = NULL;
6350 #if IS_ENABLED(CONFIG_IPV6)
6351 ireq->pktopts = NULL;
6353 atomic64_set(&ireq->ir_cookie, 0);
6354 ireq->ireq_state = TCP_NEW_SYN_RECV;
6355 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
6356 ireq->ireq_family = sk_listener->sk_family;
6361 EXPORT_SYMBOL(inet_reqsk_alloc);
6364 * Return true if a syncookie should be sent
6366 static bool tcp_syn_flood_action(const struct sock *sk,
6367 const struct sk_buff *skb,
6370 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
6371 const char *msg = "Dropping request";
6372 bool want_cookie = false;
6373 struct net *net = sock_net(sk);
6375 #ifdef CONFIG_SYN_COOKIES
6376 if (net->ipv4.sysctl_tcp_syncookies) {
6377 msg = "Sending cookies";
6379 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
6382 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
6384 if (!queue->synflood_warned &&
6385 net->ipv4.sysctl_tcp_syncookies != 2 &&
6386 xchg(&queue->synflood_warned, 1) == 0)
6387 pr_info("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6388 proto, ntohs(tcp_hdr(skb)->dest), msg);
6393 static void tcp_reqsk_record_syn(const struct sock *sk,
6394 struct request_sock *req,
6395 const struct sk_buff *skb)
6397 if (tcp_sk(sk)->save_syn) {
6398 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
6401 copy = kmalloc(len + sizeof(u32), GFP_ATOMIC);
6404 memcpy(©[1], skb_network_header(skb), len);
6405 req->saved_syn = copy;
6410 int tcp_conn_request(struct request_sock_ops *rsk_ops,
6411 const struct tcp_request_sock_ops *af_ops,
6412 struct sock *sk, struct sk_buff *skb)
6414 struct tcp_fastopen_cookie foc = { .len = -1 };
6415 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
6416 struct tcp_options_received tmp_opt;
6417 struct tcp_sock *tp = tcp_sk(sk);
6418 struct net *net = sock_net(sk);
6419 struct sock *fastopen_sk = NULL;
6420 struct dst_entry *dst = NULL;
6421 struct request_sock *req;
6422 bool want_cookie = false;
6425 /* TW buckets are converted to open requests without
6426 * limitations, they conserve resources and peer is
6427 * evidently real one.
6429 if ((net->ipv4.sysctl_tcp_syncookies == 2 ||
6430 inet_csk_reqsk_queue_is_full(sk)) && !isn) {
6431 want_cookie = tcp_syn_flood_action(sk, skb, rsk_ops->slab_name);
6436 if (sk_acceptq_is_full(sk)) {
6437 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6441 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie);
6445 tcp_rsk(req)->af_specific = af_ops;
6447 tcp_clear_options(&tmp_opt);
6448 tmp_opt.mss_clamp = af_ops->mss_clamp;
6449 tmp_opt.user_mss = tp->rx_opt.user_mss;
6450 tcp_parse_options(skb, &tmp_opt, 0, want_cookie ? NULL : &foc);
6452 if (want_cookie && !tmp_opt.saw_tstamp)
6453 tcp_clear_options(&tmp_opt);
6455 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
6456 tcp_openreq_init(req, &tmp_opt, skb, sk);
6457 inet_rsk(req)->no_srccheck = inet_sk(sk)->transparent;
6459 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6460 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb);
6462 af_ops->init_req(req, sk, skb);
6464 if (security_inet_conn_request(sk, skb, req))
6467 if (!want_cookie && !isn) {
6468 /* VJ's idea. We save last timestamp seen
6469 * from the destination in peer table, when entering
6470 * state TIME-WAIT, and check against it before
6471 * accepting new connection request.
6473 * If "isn" is not zero, this request hit alive
6474 * timewait bucket, so that all the necessary checks
6475 * are made in the function processing timewait state.
6477 if (tcp_death_row.sysctl_tw_recycle) {
6480 dst = af_ops->route_req(sk, &fl, req, &strict);
6482 if (dst && strict &&
6483 !tcp_peer_is_proven(req, dst, true,
6484 tmp_opt.saw_tstamp)) {
6485 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSPASSIVEREJECTED);
6486 goto drop_and_release;
6489 /* Kill the following clause, if you dislike this way. */
6490 else if (!net->ipv4.sysctl_tcp_syncookies &&
6491 (sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
6492 (sysctl_max_syn_backlog >> 2)) &&
6493 !tcp_peer_is_proven(req, dst, false,
6494 tmp_opt.saw_tstamp)) {
6495 /* Without syncookies last quarter of
6496 * backlog is filled with destinations,
6497 * proven to be alive.
6498 * It means that we continue to communicate
6499 * to destinations, already remembered
6500 * to the moment of synflood.
6502 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
6504 goto drop_and_release;
6507 isn = af_ops->init_seq(skb);
6510 dst = af_ops->route_req(sk, &fl, req, NULL);
6515 tcp_ecn_create_request(req, skb, sk, dst);
6518 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
6519 req->cookie_ts = tmp_opt.tstamp_ok;
6520 if (!tmp_opt.tstamp_ok)
6521 inet_rsk(req)->ecn_ok = 0;
6524 tcp_rsk(req)->snt_isn = isn;
6525 tcp_rsk(req)->txhash = net_tx_rndhash();
6526 tcp_openreq_init_rwin(req, sk, dst);
6528 tcp_reqsk_record_syn(sk, req, skb);
6529 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst);
6532 af_ops->send_synack(fastopen_sk, dst, &fl, req,
6533 &foc, TCP_SYNACK_FASTOPEN);
6534 /* Add the child socket directly into the accept queue */
6535 if (!inet_csk_reqsk_queue_add(sk, req, fastopen_sk)) {
6536 reqsk_fastopen_remove(fastopen_sk, req, false);
6537 bh_unlock_sock(fastopen_sk);
6538 sock_put(fastopen_sk);
6542 sk->sk_data_ready(sk);
6543 bh_unlock_sock(fastopen_sk);
6544 sock_put(fastopen_sk);
6546 tcp_rsk(req)->tfo_listener = false;
6548 inet_csk_reqsk_queue_hash_add(sk, req, TCP_TIMEOUT_INIT);
6549 af_ops->send_synack(sk, dst, &fl, req, &foc,
6550 !want_cookie ? TCP_SYNACK_NORMAL :
6568 EXPORT_SYMBOL(tcp_conn_request);