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_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
84 int sysctl_tcp_max_reordering __read_mostly = 300;
85 EXPORT_SYMBOL(sysctl_tcp_reordering);
86 int sysctl_tcp_dsack __read_mostly = 1;
87 int sysctl_tcp_app_win __read_mostly = 31;
88 int sysctl_tcp_adv_win_scale __read_mostly = 1;
89 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
91 /* rfc5961 challenge ack rate limiting */
92 int sysctl_tcp_challenge_ack_limit = 1000;
94 int sysctl_tcp_stdurg __read_mostly;
95 int sysctl_tcp_rfc1337 __read_mostly;
96 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
97 int sysctl_tcp_frto __read_mostly = 2;
98 int sysctl_tcp_min_rtt_wlen __read_mostly = 300;
100 int sysctl_tcp_thin_dupack __read_mostly;
102 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
103 int sysctl_tcp_early_retrans __read_mostly = 3;
104 int sysctl_tcp_invalid_ratelimit __read_mostly = HZ/2;
105 int sysctl_tcp_default_init_rwnd __read_mostly = TCP_INIT_CWND * 2;
107 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
108 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
109 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
110 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
111 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
112 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
113 #define FLAG_ECE 0x40 /* ECE in this ACK */
114 #define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */
115 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
116 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
117 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
118 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
119 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
120 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
122 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
123 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
124 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
125 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
127 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
128 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
130 /* Adapt the MSS value used to make delayed ack decision to the
133 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
135 struct inet_connection_sock *icsk = inet_csk(sk);
136 const unsigned int lss = icsk->icsk_ack.last_seg_size;
139 icsk->icsk_ack.last_seg_size = 0;
141 /* skb->len may jitter because of SACKs, even if peer
142 * sends good full-sized frames.
144 len = skb_shinfo(skb)->gso_size ? : skb->len;
145 if (len >= icsk->icsk_ack.rcv_mss) {
146 icsk->icsk_ack.rcv_mss = len;
148 /* Otherwise, we make more careful check taking into account,
149 * that SACKs block is variable.
151 * "len" is invariant segment length, including TCP header.
153 len += skb->data - skb_transport_header(skb);
154 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
155 /* If PSH is not set, packet should be
156 * full sized, provided peer TCP is not badly broken.
157 * This observation (if it is correct 8)) allows
158 * to handle super-low mtu links fairly.
160 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
161 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
162 /* Subtract also invariant (if peer is RFC compliant),
163 * tcp header plus fixed timestamp option length.
164 * Resulting "len" is MSS free of SACK jitter.
166 len -= tcp_sk(sk)->tcp_header_len;
167 icsk->icsk_ack.last_seg_size = len;
169 icsk->icsk_ack.rcv_mss = len;
173 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
174 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
175 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
179 static void tcp_incr_quickack(struct sock *sk)
181 struct inet_connection_sock *icsk = inet_csk(sk);
182 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
186 if (quickacks > icsk->icsk_ack.quick)
187 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
190 static void tcp_enter_quickack_mode(struct sock *sk)
192 struct inet_connection_sock *icsk = inet_csk(sk);
193 tcp_incr_quickack(sk);
194 icsk->icsk_ack.pingpong = 0;
195 icsk->icsk_ack.ato = TCP_ATO_MIN;
198 /* Send ACKs quickly, if "quick" count is not exhausted
199 * and the session is not interactive.
202 static bool tcp_in_quickack_mode(struct sock *sk)
204 const struct inet_connection_sock *icsk = inet_csk(sk);
205 const struct dst_entry *dst = __sk_dst_get(sk);
207 return (dst && dst_metric(dst, RTAX_QUICKACK)) ||
208 (icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong);
211 static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
213 if (tp->ecn_flags & TCP_ECN_OK)
214 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
217 static void tcp_ecn_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
219 if (tcp_hdr(skb)->cwr)
220 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
223 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
225 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
228 static void __tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
230 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
231 case INET_ECN_NOT_ECT:
232 /* Funny extension: if ECT is not set on a segment,
233 * and we already seen ECT on a previous segment,
234 * it is probably a retransmit.
236 if (tp->ecn_flags & TCP_ECN_SEEN)
237 tcp_enter_quickack_mode((struct sock *)tp);
240 if (tcp_ca_needs_ecn((struct sock *)tp))
241 tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_IS_CE);
243 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
244 /* Better not delay acks, sender can have a very low cwnd */
245 tcp_enter_quickack_mode((struct sock *)tp);
246 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
248 tp->ecn_flags |= TCP_ECN_SEEN;
251 if (tcp_ca_needs_ecn((struct sock *)tp))
252 tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_NO_CE);
253 tp->ecn_flags |= TCP_ECN_SEEN;
258 static void tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
260 if (tp->ecn_flags & TCP_ECN_OK)
261 __tcp_ecn_check_ce(tp, skb);
264 static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
266 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
267 tp->ecn_flags &= ~TCP_ECN_OK;
270 static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
272 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
273 tp->ecn_flags &= ~TCP_ECN_OK;
276 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
278 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
283 /* Buffer size and advertised window tuning.
285 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
288 static void tcp_sndbuf_expand(struct sock *sk)
290 const struct tcp_sock *tp = tcp_sk(sk);
294 /* Worst case is non GSO/TSO : each frame consumes one skb
295 * and skb->head is kmalloced using power of two area of memory
297 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
299 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
301 per_mss = roundup_pow_of_two(per_mss) +
302 SKB_DATA_ALIGN(sizeof(struct sk_buff));
304 nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd);
305 nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
307 /* Fast Recovery (RFC 5681 3.2) :
308 * Cubic needs 1.7 factor, rounded to 2 to include
309 * extra cushion (application might react slowly to POLLOUT)
311 sndmem = 2 * nr_segs * per_mss;
313 if (sk->sk_sndbuf < sndmem)
314 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
317 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
319 * All tcp_full_space() is split to two parts: "network" buffer, allocated
320 * forward and advertised in receiver window (tp->rcv_wnd) and
321 * "application buffer", required to isolate scheduling/application
322 * latencies from network.
323 * window_clamp is maximal advertised window. It can be less than
324 * tcp_full_space(), in this case tcp_full_space() - window_clamp
325 * is reserved for "application" buffer. The less window_clamp is
326 * the smoother our behaviour from viewpoint of network, but the lower
327 * throughput and the higher sensitivity of the connection to losses. 8)
329 * rcv_ssthresh is more strict window_clamp used at "slow start"
330 * phase to predict further behaviour of this connection.
331 * It is used for two goals:
332 * - to enforce header prediction at sender, even when application
333 * requires some significant "application buffer". It is check #1.
334 * - to prevent pruning of receive queue because of misprediction
335 * of receiver window. Check #2.
337 * The scheme does not work when sender sends good segments opening
338 * window and then starts to feed us spaghetti. But it should work
339 * in common situations. Otherwise, we have to rely on queue collapsing.
342 /* Slow part of check#2. */
343 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
345 struct tcp_sock *tp = tcp_sk(sk);
347 int truesize = tcp_win_from_space(skb->truesize) >> 1;
348 int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
350 while (tp->rcv_ssthresh <= window) {
351 if (truesize <= skb->len)
352 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
360 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
362 struct tcp_sock *tp = tcp_sk(sk);
365 if (tp->rcv_ssthresh < tp->window_clamp &&
366 (int)tp->rcv_ssthresh < tcp_space(sk) &&
367 !tcp_under_memory_pressure(sk)) {
370 /* Check #2. Increase window, if skb with such overhead
371 * will fit to rcvbuf in future.
373 if (tcp_win_from_space(skb->truesize) <= skb->len)
374 incr = 2 * tp->advmss;
376 incr = __tcp_grow_window(sk, skb);
379 incr = max_t(int, incr, 2 * skb->len);
380 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
382 inet_csk(sk)->icsk_ack.quick |= 1;
387 /* 3. Tuning rcvbuf, when connection enters established state. */
388 static void tcp_fixup_rcvbuf(struct sock *sk)
390 u32 mss = tcp_sk(sk)->advmss;
393 rcvmem = 2 * SKB_TRUESIZE(mss + MAX_TCP_HEADER) *
394 tcp_default_init_rwnd(mss);
396 /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
397 * Allow enough cushion so that sender is not limited by our window
399 if (sysctl_tcp_moderate_rcvbuf)
402 if (sk->sk_rcvbuf < rcvmem)
403 sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]);
406 /* 4. Try to fixup all. It is made immediately after connection enters
409 void tcp_init_buffer_space(struct sock *sk)
411 struct tcp_sock *tp = tcp_sk(sk);
414 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
415 tcp_fixup_rcvbuf(sk);
416 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
417 tcp_sndbuf_expand(sk);
419 tp->rcvq_space.space = tp->rcv_wnd;
420 tp->rcvq_space.time = tcp_time_stamp;
421 tp->rcvq_space.seq = tp->copied_seq;
423 maxwin = tcp_full_space(sk);
425 if (tp->window_clamp >= maxwin) {
426 tp->window_clamp = maxwin;
428 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
429 tp->window_clamp = max(maxwin -
430 (maxwin >> sysctl_tcp_app_win),
434 /* Force reservation of one segment. */
435 if (sysctl_tcp_app_win &&
436 tp->window_clamp > 2 * tp->advmss &&
437 tp->window_clamp + tp->advmss > maxwin)
438 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
440 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
441 tp->snd_cwnd_stamp = tcp_time_stamp;
444 /* 5. Recalculate window clamp after socket hit its memory bounds. */
445 static void tcp_clamp_window(struct sock *sk)
447 struct tcp_sock *tp = tcp_sk(sk);
448 struct inet_connection_sock *icsk = inet_csk(sk);
450 icsk->icsk_ack.quick = 0;
452 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
453 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
454 !tcp_under_memory_pressure(sk) &&
455 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
456 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
459 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
460 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
463 /* Initialize RCV_MSS value.
464 * RCV_MSS is an our guess about MSS used by the peer.
465 * We haven't any direct information about the MSS.
466 * It's better to underestimate the RCV_MSS rather than overestimate.
467 * Overestimations make us ACKing less frequently than needed.
468 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
470 void tcp_initialize_rcv_mss(struct sock *sk)
472 const struct tcp_sock *tp = tcp_sk(sk);
473 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
475 hint = min(hint, tp->rcv_wnd / 2);
476 hint = min(hint, TCP_MSS_DEFAULT);
477 hint = max(hint, TCP_MIN_MSS);
479 inet_csk(sk)->icsk_ack.rcv_mss = hint;
481 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
483 /* Receiver "autotuning" code.
485 * The algorithm for RTT estimation w/o timestamps is based on
486 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
487 * <http://public.lanl.gov/radiant/pubs.html#DRS>
489 * More detail on this code can be found at
490 * <http://staff.psc.edu/jheffner/>,
491 * though this reference is out of date. A new paper
494 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
496 u32 new_sample = tp->rcv_rtt_est.rtt;
502 if (new_sample != 0) {
503 /* If we sample in larger samples in the non-timestamp
504 * case, we could grossly overestimate the RTT especially
505 * with chatty applications or bulk transfer apps which
506 * are stalled on filesystem I/O.
508 * Also, since we are only going for a minimum in the
509 * non-timestamp case, we do not smooth things out
510 * else with timestamps disabled convergence takes too
514 m -= (new_sample >> 3);
522 /* No previous measure. */
526 if (tp->rcv_rtt_est.rtt != new_sample)
527 tp->rcv_rtt_est.rtt = new_sample;
530 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
532 if (tp->rcv_rtt_est.time == 0)
534 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
536 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rcv_rtt_est.time, 1);
539 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
540 tp->rcv_rtt_est.time = tcp_time_stamp;
543 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
544 const struct sk_buff *skb)
546 struct tcp_sock *tp = tcp_sk(sk);
547 if (tp->rx_opt.rcv_tsecr &&
548 (TCP_SKB_CB(skb)->end_seq -
549 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
550 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
554 * This function should be called every time data is copied to user space.
555 * It calculates the appropriate TCP receive buffer space.
557 void tcp_rcv_space_adjust(struct sock *sk)
559 struct tcp_sock *tp = tcp_sk(sk);
563 time = tcp_time_stamp - tp->rcvq_space.time;
564 if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
567 /* Number of bytes copied to user in last RTT */
568 copied = tp->copied_seq - tp->rcvq_space.seq;
569 if (copied <= tp->rcvq_space.space)
573 * copied = bytes received in previous RTT, our base window
574 * To cope with packet losses, we need a 2x factor
575 * To cope with slow start, and sender growing its cwin by 100 %
576 * every RTT, we need a 4x factor, because the ACK we are sending
577 * now is for the next RTT, not the current one :
578 * <prev RTT . ><current RTT .. ><next RTT .... >
581 if (sysctl_tcp_moderate_rcvbuf &&
582 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
583 int rcvwin, rcvmem, rcvbuf;
585 /* minimal window to cope with packet losses, assuming
586 * steady state. Add some cushion because of small variations.
588 rcvwin = (copied << 1) + 16 * tp->advmss;
590 /* If rate increased by 25%,
591 * assume slow start, rcvwin = 3 * copied
592 * If rate increased by 50%,
593 * assume sender can use 2x growth, rcvwin = 4 * copied
596 tp->rcvq_space.space + (tp->rcvq_space.space >> 2)) {
598 tp->rcvq_space.space + (tp->rcvq_space.space >> 1))
601 rcvwin += (rcvwin >> 1);
604 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
605 while (tcp_win_from_space(rcvmem) < tp->advmss)
608 rcvbuf = min(rcvwin / tp->advmss * rcvmem, sysctl_tcp_rmem[2]);
609 if (rcvbuf > sk->sk_rcvbuf) {
610 sk->sk_rcvbuf = rcvbuf;
612 /* Make the window clamp follow along. */
613 tp->window_clamp = rcvwin;
616 tp->rcvq_space.space = copied;
619 tp->rcvq_space.seq = tp->copied_seq;
620 tp->rcvq_space.time = tcp_time_stamp;
623 /* There is something which you must keep in mind when you analyze the
624 * behavior of the tp->ato delayed ack timeout interval. When a
625 * connection starts up, we want to ack as quickly as possible. The
626 * problem is that "good" TCP's do slow start at the beginning of data
627 * transmission. The means that until we send the first few ACK's the
628 * sender will sit on his end and only queue most of his data, because
629 * he can only send snd_cwnd unacked packets at any given time. For
630 * each ACK we send, he increments snd_cwnd and transmits more of his
633 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
635 struct tcp_sock *tp = tcp_sk(sk);
636 struct inet_connection_sock *icsk = inet_csk(sk);
639 inet_csk_schedule_ack(sk);
641 tcp_measure_rcv_mss(sk, skb);
643 tcp_rcv_rtt_measure(tp);
645 now = tcp_time_stamp;
647 if (!icsk->icsk_ack.ato) {
648 /* The _first_ data packet received, initialize
649 * delayed ACK engine.
651 tcp_incr_quickack(sk);
652 icsk->icsk_ack.ato = TCP_ATO_MIN;
654 int m = now - icsk->icsk_ack.lrcvtime;
656 if (m <= TCP_ATO_MIN / 2) {
657 /* The fastest case is the first. */
658 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
659 } else if (m < icsk->icsk_ack.ato) {
660 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
661 if (icsk->icsk_ack.ato > icsk->icsk_rto)
662 icsk->icsk_ack.ato = icsk->icsk_rto;
663 } else if (m > icsk->icsk_rto) {
664 /* Too long gap. Apparently sender failed to
665 * restart window, so that we send ACKs quickly.
667 tcp_incr_quickack(sk);
671 icsk->icsk_ack.lrcvtime = now;
673 tcp_ecn_check_ce(tp, skb);
676 tcp_grow_window(sk, skb);
679 /* Called to compute a smoothed rtt estimate. The data fed to this
680 * routine either comes from timestamps, or from segments that were
681 * known _not_ to have been retransmitted [see Karn/Partridge
682 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
683 * piece by Van Jacobson.
684 * NOTE: the next three routines used to be one big routine.
685 * To save cycles in the RFC 1323 implementation it was better to break
686 * it up into three procedures. -- erics
688 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
690 struct tcp_sock *tp = tcp_sk(sk);
691 long m = mrtt_us; /* RTT */
692 u32 srtt = tp->srtt_us;
694 /* The following amusing code comes from Jacobson's
695 * article in SIGCOMM '88. Note that rtt and mdev
696 * are scaled versions of rtt and mean deviation.
697 * This is designed to be as fast as possible
698 * m stands for "measurement".
700 * On a 1990 paper the rto value is changed to:
701 * RTO = rtt + 4 * mdev
703 * Funny. This algorithm seems to be very broken.
704 * These formulae increase RTO, when it should be decreased, increase
705 * too slowly, when it should be increased quickly, decrease too quickly
706 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
707 * does not matter how to _calculate_ it. Seems, it was trap
708 * that VJ failed to avoid. 8)
711 m -= (srtt >> 3); /* m is now error in rtt est */
712 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
714 m = -m; /* m is now abs(error) */
715 m -= (tp->mdev_us >> 2); /* similar update on mdev */
716 /* This is similar to one of Eifel findings.
717 * Eifel blocks mdev updates when rtt decreases.
718 * This solution is a bit different: we use finer gain
719 * for mdev in this case (alpha*beta).
720 * Like Eifel it also prevents growth of rto,
721 * but also it limits too fast rto decreases,
722 * happening in pure Eifel.
727 m -= (tp->mdev_us >> 2); /* similar update on mdev */
729 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
730 if (tp->mdev_us > tp->mdev_max_us) {
731 tp->mdev_max_us = tp->mdev_us;
732 if (tp->mdev_max_us > tp->rttvar_us)
733 tp->rttvar_us = tp->mdev_max_us;
735 if (after(tp->snd_una, tp->rtt_seq)) {
736 if (tp->mdev_max_us < tp->rttvar_us)
737 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
738 tp->rtt_seq = tp->snd_nxt;
739 tp->mdev_max_us = tcp_rto_min_us(sk);
742 /* no previous measure. */
743 srtt = m << 3; /* take the measured time to be rtt */
744 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
745 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
746 tp->mdev_max_us = tp->rttvar_us;
747 tp->rtt_seq = tp->snd_nxt;
749 tp->srtt_us = max(1U, srtt);
752 /* Set the sk_pacing_rate to allow proper sizing of TSO packets.
753 * Note: TCP stack does not yet implement pacing.
754 * FQ packet scheduler can be used to implement cheap but effective
755 * TCP pacing, to smooth the burst on large writes when packets
756 * in flight is significantly lower than cwnd (or rwin)
758 int sysctl_tcp_pacing_ss_ratio __read_mostly = 200;
759 int sysctl_tcp_pacing_ca_ratio __read_mostly = 120;
761 static void tcp_update_pacing_rate(struct sock *sk)
763 const struct tcp_sock *tp = tcp_sk(sk);
766 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
767 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
769 /* current rate is (cwnd * mss) / srtt
770 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
771 * In Congestion Avoidance phase, set it to 120 % the current rate.
773 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
774 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
775 * end of slow start and should slow down.
777 if (tp->snd_cwnd < tp->snd_ssthresh / 2)
778 rate *= sysctl_tcp_pacing_ss_ratio;
780 rate *= sysctl_tcp_pacing_ca_ratio;
782 rate *= max(tp->snd_cwnd, tp->packets_out);
784 if (likely(tp->srtt_us))
785 do_div(rate, tp->srtt_us);
787 /* ACCESS_ONCE() is needed because sch_fq fetches sk_pacing_rate
788 * without any lock. We want to make sure compiler wont store
789 * intermediate values in this location.
791 ACCESS_ONCE(sk->sk_pacing_rate) = min_t(u64, rate,
792 sk->sk_max_pacing_rate);
795 /* Calculate rto without backoff. This is the second half of Van Jacobson's
796 * routine referred to above.
798 static void tcp_set_rto(struct sock *sk)
800 const struct tcp_sock *tp = tcp_sk(sk);
801 /* Old crap is replaced with new one. 8)
804 * 1. If rtt variance happened to be less 50msec, it is hallucination.
805 * It cannot be less due to utterly erratic ACK generation made
806 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
807 * to do with delayed acks, because at cwnd>2 true delack timeout
808 * is invisible. Actually, Linux-2.4 also generates erratic
809 * ACKs in some circumstances.
811 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
813 /* 2. Fixups made earlier cannot be right.
814 * If we do not estimate RTO correctly without them,
815 * all the algo is pure shit and should be replaced
816 * with correct one. It is exactly, which we pretend to do.
819 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
820 * guarantees that rto is higher.
825 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
827 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
830 cwnd = TCP_INIT_CWND;
831 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
835 * Packet counting of FACK is based on in-order assumptions, therefore TCP
836 * disables it when reordering is detected
838 void tcp_disable_fack(struct tcp_sock *tp)
840 /* RFC3517 uses different metric in lost marker => reset on change */
842 tp->lost_skb_hint = NULL;
843 tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED;
846 /* Take a notice that peer is sending D-SACKs */
847 static void tcp_dsack_seen(struct tcp_sock *tp)
849 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
852 static void tcp_update_reordering(struct sock *sk, const int metric,
855 struct tcp_sock *tp = tcp_sk(sk);
856 if (metric > tp->reordering) {
859 tp->reordering = min(sysctl_tcp_max_reordering, metric);
861 /* This exciting event is worth to be remembered. 8) */
863 mib_idx = LINUX_MIB_TCPTSREORDER;
864 else if (tcp_is_reno(tp))
865 mib_idx = LINUX_MIB_TCPRENOREORDER;
866 else if (tcp_is_fack(tp))
867 mib_idx = LINUX_MIB_TCPFACKREORDER;
869 mib_idx = LINUX_MIB_TCPSACKREORDER;
871 NET_INC_STATS_BH(sock_net(sk), mib_idx);
872 #if FASTRETRANS_DEBUG > 1
873 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
874 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
878 tp->undo_marker ? tp->undo_retrans : 0);
880 tcp_disable_fack(tp);
884 tcp_disable_early_retrans(tp);
888 /* This must be called before lost_out is incremented */
889 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
891 if (!tp->retransmit_skb_hint ||
892 before(TCP_SKB_CB(skb)->seq,
893 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
894 tp->retransmit_skb_hint = skb;
897 after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
898 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
901 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
903 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
904 tcp_verify_retransmit_hint(tp, skb);
906 tp->lost_out += tcp_skb_pcount(skb);
907 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
911 void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, struct sk_buff *skb)
913 tcp_verify_retransmit_hint(tp, skb);
915 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
916 tp->lost_out += tcp_skb_pcount(skb);
917 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
921 /* This procedure tags the retransmission queue when SACKs arrive.
923 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
924 * Packets in queue with these bits set are counted in variables
925 * sacked_out, retrans_out and lost_out, correspondingly.
927 * Valid combinations are:
928 * Tag InFlight Description
929 * 0 1 - orig segment is in flight.
930 * S 0 - nothing flies, orig reached receiver.
931 * L 0 - nothing flies, orig lost by net.
932 * R 2 - both orig and retransmit are in flight.
933 * L|R 1 - orig is lost, retransmit is in flight.
934 * S|R 1 - orig reached receiver, retrans is still in flight.
935 * (L|S|R is logically valid, it could occur when L|R is sacked,
936 * but it is equivalent to plain S and code short-curcuits it to S.
937 * L|S is logically invalid, it would mean -1 packet in flight 8))
939 * These 6 states form finite state machine, controlled by the following events:
940 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
941 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
942 * 3. Loss detection event of two flavors:
943 * A. Scoreboard estimator decided the packet is lost.
944 * A'. Reno "three dupacks" marks head of queue lost.
945 * A''. Its FACK modification, head until snd.fack is lost.
946 * B. SACK arrives sacking SND.NXT at the moment, when the
947 * segment was retransmitted.
948 * 4. D-SACK added new rule: D-SACK changes any tag to S.
950 * It is pleasant to note, that state diagram turns out to be commutative,
951 * so that we are allowed not to be bothered by order of our actions,
952 * when multiple events arrive simultaneously. (see the function below).
954 * Reordering detection.
955 * --------------------
956 * Reordering metric is maximal distance, which a packet can be displaced
957 * in packet stream. With SACKs we can estimate it:
959 * 1. SACK fills old hole and the corresponding segment was not
960 * ever retransmitted -> reordering. Alas, we cannot use it
961 * when segment was retransmitted.
962 * 2. The last flaw is solved with D-SACK. D-SACK arrives
963 * for retransmitted and already SACKed segment -> reordering..
964 * Both of these heuristics are not used in Loss state, when we cannot
965 * account for retransmits accurately.
967 * SACK block validation.
968 * ----------------------
970 * SACK block range validation checks that the received SACK block fits to
971 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
972 * Note that SND.UNA is not included to the range though being valid because
973 * it means that the receiver is rather inconsistent with itself reporting
974 * SACK reneging when it should advance SND.UNA. Such SACK block this is
975 * perfectly valid, however, in light of RFC2018 which explicitly states
976 * that "SACK block MUST reflect the newest segment. Even if the newest
977 * segment is going to be discarded ...", not that it looks very clever
978 * in case of head skb. Due to potentional receiver driven attacks, we
979 * choose to avoid immediate execution of a walk in write queue due to
980 * reneging and defer head skb's loss recovery to standard loss recovery
981 * procedure that will eventually trigger (nothing forbids us doing this).
983 * Implements also blockage to start_seq wrap-around. Problem lies in the
984 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
985 * there's no guarantee that it will be before snd_nxt (n). The problem
986 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
989 * <- outs wnd -> <- wrapzone ->
990 * u e n u_w e_w s n_w
992 * |<------------+------+----- TCP seqno space --------------+---------->|
993 * ...-- <2^31 ->| |<--------...
994 * ...---- >2^31 ------>| |<--------...
996 * Current code wouldn't be vulnerable but it's better still to discard such
997 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
998 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
999 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1000 * equal to the ideal case (infinite seqno space without wrap caused issues).
1002 * With D-SACK the lower bound is extended to cover sequence space below
1003 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1004 * again, D-SACK block must not to go across snd_una (for the same reason as
1005 * for the normal SACK blocks, explained above). But there all simplicity
1006 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1007 * fully below undo_marker they do not affect behavior in anyway and can
1008 * therefore be safely ignored. In rare cases (which are more or less
1009 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1010 * fragmentation and packet reordering past skb's retransmission. To consider
1011 * them correctly, the acceptable range must be extended even more though
1012 * the exact amount is rather hard to quantify. However, tp->max_window can
1013 * be used as an exaggerated estimate.
1015 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1016 u32 start_seq, u32 end_seq)
1018 /* Too far in future, or reversed (interpretation is ambiguous) */
1019 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1022 /* Nasty start_seq wrap-around check (see comments above) */
1023 if (!before(start_seq, tp->snd_nxt))
1026 /* In outstanding window? ...This is valid exit for D-SACKs too.
1027 * start_seq == snd_una is non-sensical (see comments above)
1029 if (after(start_seq, tp->snd_una))
1032 if (!is_dsack || !tp->undo_marker)
1035 /* ...Then it's D-SACK, and must reside below snd_una completely */
1036 if (after(end_seq, tp->snd_una))
1039 if (!before(start_seq, tp->undo_marker))
1043 if (!after(end_seq, tp->undo_marker))
1046 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1047 * start_seq < undo_marker and end_seq >= undo_marker.
1049 return !before(start_seq, end_seq - tp->max_window);
1052 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1053 struct tcp_sack_block_wire *sp, int num_sacks,
1056 struct tcp_sock *tp = tcp_sk(sk);
1057 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1058 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1059 bool dup_sack = false;
1061 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1064 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1065 } else if (num_sacks > 1) {
1066 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1067 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1069 if (!after(end_seq_0, end_seq_1) &&
1070 !before(start_seq_0, start_seq_1)) {
1073 NET_INC_STATS_BH(sock_net(sk),
1074 LINUX_MIB_TCPDSACKOFORECV);
1078 /* D-SACK for already forgotten data... Do dumb counting. */
1079 if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 &&
1080 !after(end_seq_0, prior_snd_una) &&
1081 after(end_seq_0, tp->undo_marker))
1087 struct tcp_sacktag_state {
1090 /* Timestamps for earliest and latest never-retransmitted segment
1091 * that was SACKed. RTO needs the earliest RTT to stay conservative,
1092 * but congestion control should still get an accurate delay signal.
1094 struct skb_mstamp first_sackt;
1095 struct skb_mstamp last_sackt;
1099 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1100 * the incoming SACK may not exactly match but we can find smaller MSS
1101 * aligned portion of it that matches. Therefore we might need to fragment
1102 * which may fail and creates some hassle (caller must handle error case
1105 * FIXME: this could be merged to shift decision code
1107 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1108 u32 start_seq, u32 end_seq)
1112 unsigned int pkt_len;
1115 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1116 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1118 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1119 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1120 mss = tcp_skb_mss(skb);
1121 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1124 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1128 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1133 /* Round if necessary so that SACKs cover only full MSSes
1134 * and/or the remaining small portion (if present)
1136 if (pkt_len > mss) {
1137 unsigned int new_len = (pkt_len / mss) * mss;
1138 if (!in_sack && new_len < pkt_len)
1143 if (pkt_len >= skb->len && !in_sack)
1146 err = tcp_fragment(sk, skb, pkt_len, mss, GFP_ATOMIC);
1154 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1155 static u8 tcp_sacktag_one(struct sock *sk,
1156 struct tcp_sacktag_state *state, u8 sacked,
1157 u32 start_seq, u32 end_seq,
1158 int dup_sack, int pcount,
1159 const struct skb_mstamp *xmit_time)
1161 struct tcp_sock *tp = tcp_sk(sk);
1162 int fack_count = state->fack_count;
1164 /* Account D-SACK for retransmitted packet. */
1165 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1166 if (tp->undo_marker && tp->undo_retrans > 0 &&
1167 after(end_seq, tp->undo_marker))
1169 if (sacked & TCPCB_SACKED_ACKED)
1170 state->reord = min(fack_count, state->reord);
1173 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1174 if (!after(end_seq, tp->snd_una))
1177 if (!(sacked & TCPCB_SACKED_ACKED)) {
1178 tcp_rack_advance(tp, xmit_time, sacked);
1180 if (sacked & TCPCB_SACKED_RETRANS) {
1181 /* If the segment is not tagged as lost,
1182 * we do not clear RETRANS, believing
1183 * that retransmission is still in flight.
1185 if (sacked & TCPCB_LOST) {
1186 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1187 tp->lost_out -= pcount;
1188 tp->retrans_out -= pcount;
1191 if (!(sacked & TCPCB_RETRANS)) {
1192 /* New sack for not retransmitted frame,
1193 * which was in hole. It is reordering.
1195 if (before(start_seq,
1196 tcp_highest_sack_seq(tp)))
1197 state->reord = min(fack_count,
1199 if (!after(end_seq, tp->high_seq))
1200 state->flag |= FLAG_ORIG_SACK_ACKED;
1201 if (state->first_sackt.v64 == 0)
1202 state->first_sackt = *xmit_time;
1203 state->last_sackt = *xmit_time;
1206 if (sacked & TCPCB_LOST) {
1207 sacked &= ~TCPCB_LOST;
1208 tp->lost_out -= pcount;
1212 sacked |= TCPCB_SACKED_ACKED;
1213 state->flag |= FLAG_DATA_SACKED;
1214 tp->sacked_out += pcount;
1216 fack_count += pcount;
1218 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1219 if (!tcp_is_fack(tp) && tp->lost_skb_hint &&
1220 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1221 tp->lost_cnt_hint += pcount;
1223 if (fack_count > tp->fackets_out)
1224 tp->fackets_out = fack_count;
1227 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1228 * frames and clear it. undo_retrans is decreased above, L|R frames
1229 * are accounted above as well.
1231 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1232 sacked &= ~TCPCB_SACKED_RETRANS;
1233 tp->retrans_out -= pcount;
1239 /* Shift newly-SACKed bytes from this skb to the immediately previous
1240 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1242 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1243 struct tcp_sacktag_state *state,
1244 unsigned int pcount, int shifted, int mss,
1247 struct tcp_sock *tp = tcp_sk(sk);
1248 struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1249 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1250 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1254 /* Adjust counters and hints for the newly sacked sequence
1255 * range but discard the return value since prev is already
1256 * marked. We must tag the range first because the seq
1257 * advancement below implicitly advances
1258 * tcp_highest_sack_seq() when skb is highest_sack.
1260 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1261 start_seq, end_seq, dup_sack, pcount,
1264 if (skb == tp->lost_skb_hint)
1265 tp->lost_cnt_hint += pcount;
1267 TCP_SKB_CB(prev)->end_seq += shifted;
1268 TCP_SKB_CB(skb)->seq += shifted;
1270 tcp_skb_pcount_add(prev, pcount);
1271 BUG_ON(tcp_skb_pcount(skb) < pcount);
1272 tcp_skb_pcount_add(skb, -pcount);
1274 /* When we're adding to gso_segs == 1, gso_size will be zero,
1275 * in theory this shouldn't be necessary but as long as DSACK
1276 * code can come after this skb later on it's better to keep
1277 * setting gso_size to something.
1279 if (!TCP_SKB_CB(prev)->tcp_gso_size)
1280 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1282 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1283 if (tcp_skb_pcount(skb) <= 1)
1284 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1286 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1287 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1290 BUG_ON(!tcp_skb_pcount(skb));
1291 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1295 /* Whole SKB was eaten :-) */
1297 if (skb == tp->retransmit_skb_hint)
1298 tp->retransmit_skb_hint = prev;
1299 if (skb == tp->lost_skb_hint) {
1300 tp->lost_skb_hint = prev;
1301 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1304 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1305 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1306 TCP_SKB_CB(prev)->end_seq++;
1308 if (skb == tcp_highest_sack(sk))
1309 tcp_advance_highest_sack(sk, skb);
1311 tcp_unlink_write_queue(skb, sk);
1312 sk_wmem_free_skb(sk, skb);
1314 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1319 /* I wish gso_size would have a bit more sane initialization than
1320 * something-or-zero which complicates things
1322 static int tcp_skb_seglen(const struct sk_buff *skb)
1324 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1327 /* Shifting pages past head area doesn't work */
1328 static int skb_can_shift(const struct sk_buff *skb)
1330 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1333 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1336 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1337 struct tcp_sacktag_state *state,
1338 u32 start_seq, u32 end_seq,
1341 struct tcp_sock *tp = tcp_sk(sk);
1342 struct sk_buff *prev;
1348 if (!sk_can_gso(sk))
1351 /* Normally R but no L won't result in plain S */
1353 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1355 if (!skb_can_shift(skb))
1357 /* This frame is about to be dropped (was ACKed). */
1358 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1361 /* Can only happen with delayed DSACK + discard craziness */
1362 if (unlikely(skb == tcp_write_queue_head(sk)))
1364 prev = tcp_write_queue_prev(sk, skb);
1366 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1369 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1370 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1374 pcount = tcp_skb_pcount(skb);
1375 mss = tcp_skb_seglen(skb);
1377 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1378 * drop this restriction as unnecessary
1380 if (mss != tcp_skb_seglen(prev))
1383 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1385 /* CHECKME: This is non-MSS split case only?, this will
1386 * cause skipped skbs due to advancing loop btw, original
1387 * has that feature too
1389 if (tcp_skb_pcount(skb) <= 1)
1392 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1394 /* TODO: head merge to next could be attempted here
1395 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1396 * though it might not be worth of the additional hassle
1398 * ...we can probably just fallback to what was done
1399 * previously. We could try merging non-SACKed ones
1400 * as well but it probably isn't going to buy off
1401 * because later SACKs might again split them, and
1402 * it would make skb timestamp tracking considerably
1408 len = end_seq - TCP_SKB_CB(skb)->seq;
1410 BUG_ON(len > skb->len);
1412 /* MSS boundaries should be honoured or else pcount will
1413 * severely break even though it makes things bit trickier.
1414 * Optimize common case to avoid most of the divides
1416 mss = tcp_skb_mss(skb);
1418 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1419 * drop this restriction as unnecessary
1421 if (mss != tcp_skb_seglen(prev))
1426 } else if (len < mss) {
1434 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1435 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1438 if (!skb_shift(prev, skb, len))
1440 if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1443 /* Hole filled allows collapsing with the next as well, this is very
1444 * useful when hole on every nth skb pattern happens
1446 if (prev == tcp_write_queue_tail(sk))
1448 skb = tcp_write_queue_next(sk, prev);
1450 if (!skb_can_shift(skb) ||
1451 (skb == tcp_send_head(sk)) ||
1452 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1453 (mss != tcp_skb_seglen(skb)))
1457 if (skb_shift(prev, skb, len)) {
1458 pcount += tcp_skb_pcount(skb);
1459 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1463 state->fack_count += pcount;
1470 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1474 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1475 struct tcp_sack_block *next_dup,
1476 struct tcp_sacktag_state *state,
1477 u32 start_seq, u32 end_seq,
1480 struct tcp_sock *tp = tcp_sk(sk);
1481 struct sk_buff *tmp;
1483 tcp_for_write_queue_from(skb, sk) {
1485 bool dup_sack = dup_sack_in;
1487 if (skb == tcp_send_head(sk))
1490 /* queue is in-order => we can short-circuit the walk early */
1491 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1495 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1496 in_sack = tcp_match_skb_to_sack(sk, skb,
1497 next_dup->start_seq,
1503 /* skb reference here is a bit tricky to get right, since
1504 * shifting can eat and free both this skb and the next,
1505 * so not even _safe variant of the loop is enough.
1508 tmp = tcp_shift_skb_data(sk, skb, state,
1509 start_seq, end_seq, dup_sack);
1518 in_sack = tcp_match_skb_to_sack(sk, skb,
1524 if (unlikely(in_sack < 0))
1528 TCP_SKB_CB(skb)->sacked =
1531 TCP_SKB_CB(skb)->sacked,
1532 TCP_SKB_CB(skb)->seq,
1533 TCP_SKB_CB(skb)->end_seq,
1535 tcp_skb_pcount(skb),
1538 if (!before(TCP_SKB_CB(skb)->seq,
1539 tcp_highest_sack_seq(tp)))
1540 tcp_advance_highest_sack(sk, skb);
1543 state->fack_count += tcp_skb_pcount(skb);
1548 /* Avoid all extra work that is being done by sacktag while walking in
1551 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1552 struct tcp_sacktag_state *state,
1555 tcp_for_write_queue_from(skb, sk) {
1556 if (skb == tcp_send_head(sk))
1559 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1562 state->fack_count += tcp_skb_pcount(skb);
1567 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1569 struct tcp_sack_block *next_dup,
1570 struct tcp_sacktag_state *state,
1576 if (before(next_dup->start_seq, skip_to_seq)) {
1577 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1578 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1579 next_dup->start_seq, next_dup->end_seq,
1586 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1588 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1592 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1593 u32 prior_snd_una, struct tcp_sacktag_state *state)
1595 struct tcp_sock *tp = tcp_sk(sk);
1596 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1597 TCP_SKB_CB(ack_skb)->sacked);
1598 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1599 struct tcp_sack_block sp[TCP_NUM_SACKS];
1600 struct tcp_sack_block *cache;
1601 struct sk_buff *skb;
1602 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1604 bool found_dup_sack = false;
1606 int first_sack_index;
1609 state->reord = tp->packets_out;
1611 if (!tp->sacked_out) {
1612 if (WARN_ON(tp->fackets_out))
1613 tp->fackets_out = 0;
1614 tcp_highest_sack_reset(sk);
1617 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1618 num_sacks, prior_snd_una);
1620 state->flag |= FLAG_DSACKING_ACK;
1622 /* Eliminate too old ACKs, but take into
1623 * account more or less fresh ones, they can
1624 * contain valid SACK info.
1626 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1629 if (!tp->packets_out)
1633 first_sack_index = 0;
1634 for (i = 0; i < num_sacks; i++) {
1635 bool dup_sack = !i && found_dup_sack;
1637 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1638 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1640 if (!tcp_is_sackblock_valid(tp, dup_sack,
1641 sp[used_sacks].start_seq,
1642 sp[used_sacks].end_seq)) {
1646 if (!tp->undo_marker)
1647 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1649 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1651 /* Don't count olds caused by ACK reordering */
1652 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1653 !after(sp[used_sacks].end_seq, tp->snd_una))
1655 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1658 NET_INC_STATS_BH(sock_net(sk), mib_idx);
1660 first_sack_index = -1;
1664 /* Ignore very old stuff early */
1665 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1671 /* order SACK blocks to allow in order walk of the retrans queue */
1672 for (i = used_sacks - 1; i > 0; i--) {
1673 for (j = 0; j < i; j++) {
1674 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1675 swap(sp[j], sp[j + 1]);
1677 /* Track where the first SACK block goes to */
1678 if (j == first_sack_index)
1679 first_sack_index = j + 1;
1684 skb = tcp_write_queue_head(sk);
1685 state->fack_count = 0;
1688 if (!tp->sacked_out) {
1689 /* It's already past, so skip checking against it */
1690 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1692 cache = tp->recv_sack_cache;
1693 /* Skip empty blocks in at head of the cache */
1694 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1699 while (i < used_sacks) {
1700 u32 start_seq = sp[i].start_seq;
1701 u32 end_seq = sp[i].end_seq;
1702 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1703 struct tcp_sack_block *next_dup = NULL;
1705 if (found_dup_sack && ((i + 1) == first_sack_index))
1706 next_dup = &sp[i + 1];
1708 /* Skip too early cached blocks */
1709 while (tcp_sack_cache_ok(tp, cache) &&
1710 !before(start_seq, cache->end_seq))
1713 /* Can skip some work by looking recv_sack_cache? */
1714 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1715 after(end_seq, cache->start_seq)) {
1718 if (before(start_seq, cache->start_seq)) {
1719 skb = tcp_sacktag_skip(skb, sk, state,
1721 skb = tcp_sacktag_walk(skb, sk, next_dup,
1728 /* Rest of the block already fully processed? */
1729 if (!after(end_seq, cache->end_seq))
1732 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1736 /* ...tail remains todo... */
1737 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1738 /* ...but better entrypoint exists! */
1739 skb = tcp_highest_sack(sk);
1742 state->fack_count = tp->fackets_out;
1747 skb = tcp_sacktag_skip(skb, sk, state, cache->end_seq);
1748 /* Check overlap against next cached too (past this one already) */
1753 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1754 skb = tcp_highest_sack(sk);
1757 state->fack_count = tp->fackets_out;
1759 skb = tcp_sacktag_skip(skb, sk, state, start_seq);
1762 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
1763 start_seq, end_seq, dup_sack);
1769 /* Clear the head of the cache sack blocks so we can skip it next time */
1770 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1771 tp->recv_sack_cache[i].start_seq = 0;
1772 tp->recv_sack_cache[i].end_seq = 0;
1774 for (j = 0; j < used_sacks; j++)
1775 tp->recv_sack_cache[i++] = sp[j];
1777 if ((state->reord < tp->fackets_out) &&
1778 ((inet_csk(sk)->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker))
1779 tcp_update_reordering(sk, tp->fackets_out - state->reord, 0);
1781 tcp_verify_left_out(tp);
1784 #if FASTRETRANS_DEBUG > 0
1785 WARN_ON((int)tp->sacked_out < 0);
1786 WARN_ON((int)tp->lost_out < 0);
1787 WARN_ON((int)tp->retrans_out < 0);
1788 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1793 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1794 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1796 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1800 holes = max(tp->lost_out, 1U);
1801 holes = min(holes, tp->packets_out);
1803 if ((tp->sacked_out + holes) > tp->packets_out) {
1804 tp->sacked_out = tp->packets_out - holes;
1810 /* If we receive more dupacks than we expected counting segments
1811 * in assumption of absent reordering, interpret this as reordering.
1812 * The only another reason could be bug in receiver TCP.
1814 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1816 struct tcp_sock *tp = tcp_sk(sk);
1817 if (tcp_limit_reno_sacked(tp))
1818 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1821 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1823 static void tcp_add_reno_sack(struct sock *sk)
1825 struct tcp_sock *tp = tcp_sk(sk);
1827 tcp_check_reno_reordering(sk, 0);
1828 tcp_verify_left_out(tp);
1831 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1833 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1835 struct tcp_sock *tp = tcp_sk(sk);
1838 /* One ACK acked hole. The rest eat duplicate ACKs. */
1839 if (acked - 1 >= tp->sacked_out)
1842 tp->sacked_out -= acked - 1;
1844 tcp_check_reno_reordering(sk, acked);
1845 tcp_verify_left_out(tp);
1848 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1853 void tcp_clear_retrans(struct tcp_sock *tp)
1855 tp->retrans_out = 0;
1857 tp->undo_marker = 0;
1858 tp->undo_retrans = -1;
1859 tp->fackets_out = 0;
1863 static inline void tcp_init_undo(struct tcp_sock *tp)
1865 tp->undo_marker = tp->snd_una;
1866 /* Retransmission still in flight may cause DSACKs later. */
1867 tp->undo_retrans = tp->retrans_out ? : -1;
1870 /* Enter Loss state. If we detect SACK reneging, forget all SACK information
1871 * and reset tags completely, otherwise preserve SACKs. If receiver
1872 * dropped its ofo queue, we will know this due to reneging detection.
1874 void tcp_enter_loss(struct sock *sk)
1876 const struct inet_connection_sock *icsk = inet_csk(sk);
1877 struct tcp_sock *tp = tcp_sk(sk);
1878 struct sk_buff *skb;
1879 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
1880 bool is_reneg; /* is receiver reneging on SACKs? */
1882 /* Reduce ssthresh if it has not yet been made inside this window. */
1883 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1884 !after(tp->high_seq, tp->snd_una) ||
1885 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1886 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1887 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1888 tcp_ca_event(sk, CA_EVENT_LOSS);
1892 tp->snd_cwnd_cnt = 0;
1893 tp->snd_cwnd_stamp = tcp_time_stamp;
1895 tp->retrans_out = 0;
1898 if (tcp_is_reno(tp))
1899 tcp_reset_reno_sack(tp);
1901 skb = tcp_write_queue_head(sk);
1902 is_reneg = skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED);
1904 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
1906 tp->fackets_out = 0;
1908 tcp_clear_all_retrans_hints(tp);
1910 tcp_for_write_queue(skb, sk) {
1911 if (skb == tcp_send_head(sk))
1914 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1915 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || is_reneg) {
1916 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1917 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1918 tp->lost_out += tcp_skb_pcount(skb);
1919 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
1922 tcp_verify_left_out(tp);
1924 /* Timeout in disordered state after receiving substantial DUPACKs
1925 * suggests that the degree of reordering is over-estimated.
1927 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
1928 tp->sacked_out >= sysctl_tcp_reordering)
1929 tp->reordering = min_t(unsigned int, tp->reordering,
1930 sysctl_tcp_reordering);
1931 tcp_set_ca_state(sk, TCP_CA_Loss);
1932 tp->high_seq = tp->snd_nxt;
1933 tcp_ecn_queue_cwr(tp);
1935 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1936 * loss recovery is underway except recurring timeout(s) on
1937 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1939 tp->frto = sysctl_tcp_frto &&
1940 (new_recovery || icsk->icsk_retransmits) &&
1941 !inet_csk(sk)->icsk_mtup.probe_size;
1944 /* If ACK arrived pointing to a remembered SACK, it means that our
1945 * remembered SACKs do not reflect real state of receiver i.e.
1946 * receiver _host_ is heavily congested (or buggy).
1948 * To avoid big spurious retransmission bursts due to transient SACK
1949 * scoreboard oddities that look like reneging, we give the receiver a
1950 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
1951 * restore sanity to the SACK scoreboard. If the apparent reneging
1952 * persists until this RTO then we'll clear the SACK scoreboard.
1954 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
1956 if (flag & FLAG_SACK_RENEGING) {
1957 struct tcp_sock *tp = tcp_sk(sk);
1958 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
1959 msecs_to_jiffies(10));
1961 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1962 delay, TCP_RTO_MAX);
1968 static inline int tcp_fackets_out(const struct tcp_sock *tp)
1970 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
1973 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
1974 * counter when SACK is enabled (without SACK, sacked_out is used for
1977 * Instead, with FACK TCP uses fackets_out that includes both SACKed
1978 * segments up to the highest received SACK block so far and holes in
1981 * With reordering, holes may still be in flight, so RFC3517 recovery
1982 * uses pure sacked_out (total number of SACKed segments) even though
1983 * it violates the RFC that uses duplicate ACKs, often these are equal
1984 * but when e.g. out-of-window ACKs or packet duplication occurs,
1985 * they differ. Since neither occurs due to loss, TCP should really
1988 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
1990 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
1993 static bool tcp_pause_early_retransmit(struct sock *sk, int flag)
1995 struct tcp_sock *tp = tcp_sk(sk);
1996 unsigned long delay;
1998 /* Delay early retransmit and entering fast recovery for
1999 * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
2000 * available, or RTO is scheduled to fire first.
2002 if (sysctl_tcp_early_retrans < 2 || sysctl_tcp_early_retrans > 3 ||
2003 (flag & FLAG_ECE) || !tp->srtt_us)
2006 delay = max(usecs_to_jiffies(tp->srtt_us >> 5),
2007 msecs_to_jiffies(2));
2009 if (!time_after(inet_csk(sk)->icsk_timeout, (jiffies + delay)))
2012 inet_csk_reset_xmit_timer(sk, ICSK_TIME_EARLY_RETRANS, delay,
2017 /* Linux NewReno/SACK/FACK/ECN state machine.
2018 * --------------------------------------
2020 * "Open" Normal state, no dubious events, fast path.
2021 * "Disorder" In all the respects it is "Open",
2022 * but requires a bit more attention. It is entered when
2023 * we see some SACKs or dupacks. It is split of "Open"
2024 * mainly to move some processing from fast path to slow one.
2025 * "CWR" CWND was reduced due to some Congestion Notification event.
2026 * It can be ECN, ICMP source quench, local device congestion.
2027 * "Recovery" CWND was reduced, we are fast-retransmitting.
2028 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2030 * tcp_fastretrans_alert() is entered:
2031 * - each incoming ACK, if state is not "Open"
2032 * - when arrived ACK is unusual, namely:
2037 * Counting packets in flight is pretty simple.
2039 * in_flight = packets_out - left_out + retrans_out
2041 * packets_out is SND.NXT-SND.UNA counted in packets.
2043 * retrans_out is number of retransmitted segments.
2045 * left_out is number of segments left network, but not ACKed yet.
2047 * left_out = sacked_out + lost_out
2049 * sacked_out: Packets, which arrived to receiver out of order
2050 * and hence not ACKed. With SACKs this number is simply
2051 * amount of SACKed data. Even without SACKs
2052 * it is easy to give pretty reliable estimate of this number,
2053 * counting duplicate ACKs.
2055 * lost_out: Packets lost by network. TCP has no explicit
2056 * "loss notification" feedback from network (for now).
2057 * It means that this number can be only _guessed_.
2058 * Actually, it is the heuristics to predict lossage that
2059 * distinguishes different algorithms.
2061 * F.e. after RTO, when all the queue is considered as lost,
2062 * lost_out = packets_out and in_flight = retrans_out.
2064 * Essentially, we have now two algorithms counting
2067 * FACK: It is the simplest heuristics. As soon as we decided
2068 * that something is lost, we decide that _all_ not SACKed
2069 * packets until the most forward SACK are lost. I.e.
2070 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2071 * It is absolutely correct estimate, if network does not reorder
2072 * packets. And it loses any connection to reality when reordering
2073 * takes place. We use FACK by default until reordering
2074 * is suspected on the path to this destination.
2076 * NewReno: when Recovery is entered, we assume that one segment
2077 * is lost (classic Reno). While we are in Recovery and
2078 * a partial ACK arrives, we assume that one more packet
2079 * is lost (NewReno). This heuristics are the same in NewReno
2082 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2083 * deflation etc. CWND is real congestion window, never inflated, changes
2084 * only according to classic VJ rules.
2086 * Really tricky (and requiring careful tuning) part of algorithm
2087 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2088 * The first determines the moment _when_ we should reduce CWND and,
2089 * hence, slow down forward transmission. In fact, it determines the moment
2090 * when we decide that hole is caused by loss, rather than by a reorder.
2092 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2093 * holes, caused by lost packets.
2095 * And the most logically complicated part of algorithm is undo
2096 * heuristics. We detect false retransmits due to both too early
2097 * fast retransmit (reordering) and underestimated RTO, analyzing
2098 * timestamps and D-SACKs. When we detect that some segments were
2099 * retransmitted by mistake and CWND reduction was wrong, we undo
2100 * window reduction and abort recovery phase. This logic is hidden
2101 * inside several functions named tcp_try_undo_<something>.
2104 /* This function decides, when we should leave Disordered state
2105 * and enter Recovery phase, reducing congestion window.
2107 * Main question: may we further continue forward transmission
2108 * with the same cwnd?
2110 static bool tcp_time_to_recover(struct sock *sk, int flag)
2112 struct tcp_sock *tp = tcp_sk(sk);
2115 /* Trick#1: The loss is proven. */
2119 /* Not-A-Trick#2 : Classic rule... */
2120 if (tcp_dupack_heuristics(tp) > tp->reordering)
2123 /* Trick#4: It is still not OK... But will it be useful to delay
2126 packets_out = tp->packets_out;
2127 if (packets_out <= tp->reordering &&
2128 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2129 !tcp_may_send_now(sk)) {
2130 /* We have nothing to send. This connection is limited
2131 * either by receiver window or by application.
2136 /* If a thin stream is detected, retransmit after first
2137 * received dupack. Employ only if SACK is supported in order
2138 * to avoid possible corner-case series of spurious retransmissions
2139 * Use only if there are no unsent data.
2141 if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2142 tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2143 tcp_is_sack(tp) && !tcp_send_head(sk))
2146 /* Trick#6: TCP early retransmit, per RFC5827. To avoid spurious
2147 * retransmissions due to small network reorderings, we implement
2148 * Mitigation A.3 in the RFC and delay the retransmission for a short
2149 * interval if appropriate.
2151 if (tp->do_early_retrans && !tp->retrans_out && tp->sacked_out &&
2152 (tp->packets_out >= (tp->sacked_out + 1) && tp->packets_out < 4) &&
2153 !tcp_may_send_now(sk))
2154 return !tcp_pause_early_retransmit(sk, flag);
2159 /* Detect loss in event "A" above by marking head of queue up as lost.
2160 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2161 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2162 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2163 * the maximum SACKed segments to pass before reaching this limit.
2165 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2167 struct tcp_sock *tp = tcp_sk(sk);
2168 struct sk_buff *skb;
2169 int cnt, oldcnt, lost;
2171 /* Use SACK to deduce losses of new sequences sent during recovery */
2172 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq;
2174 WARN_ON(packets > tp->packets_out);
2175 if (tp->lost_skb_hint) {
2176 skb = tp->lost_skb_hint;
2177 cnt = tp->lost_cnt_hint;
2178 /* Head already handled? */
2179 if (mark_head && skb != tcp_write_queue_head(sk))
2182 skb = tcp_write_queue_head(sk);
2186 tcp_for_write_queue_from(skb, sk) {
2187 if (skb == tcp_send_head(sk))
2189 /* TODO: do this better */
2190 /* this is not the most efficient way to do this... */
2191 tp->lost_skb_hint = skb;
2192 tp->lost_cnt_hint = cnt;
2194 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2198 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2199 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2200 cnt += tcp_skb_pcount(skb);
2202 if (cnt > packets) {
2203 if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2204 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2205 (oldcnt >= packets))
2208 mss = tcp_skb_mss(skb);
2209 /* If needed, chop off the prefix to mark as lost. */
2210 lost = (packets - oldcnt) * mss;
2211 if (lost < skb->len &&
2212 tcp_fragment(sk, skb, lost, mss, GFP_ATOMIC) < 0)
2217 tcp_skb_mark_lost(tp, skb);
2222 tcp_verify_left_out(tp);
2225 /* Account newly detected lost packet(s) */
2227 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2229 struct tcp_sock *tp = tcp_sk(sk);
2231 if (tcp_is_reno(tp)) {
2232 tcp_mark_head_lost(sk, 1, 1);
2233 } else if (tcp_is_fack(tp)) {
2234 int lost = tp->fackets_out - tp->reordering;
2237 tcp_mark_head_lost(sk, lost, 0);
2239 int sacked_upto = tp->sacked_out - tp->reordering;
2240 if (sacked_upto >= 0)
2241 tcp_mark_head_lost(sk, sacked_upto, 0);
2242 else if (fast_rexmit)
2243 tcp_mark_head_lost(sk, 1, 1);
2247 /* CWND moderation, preventing bursts due to too big ACKs
2248 * in dubious situations.
2250 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2252 tp->snd_cwnd = min(tp->snd_cwnd,
2253 tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2254 tp->snd_cwnd_stamp = tcp_time_stamp;
2257 static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2259 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2260 before(tp->rx_opt.rcv_tsecr, when);
2263 /* skb is spurious retransmitted if the returned timestamp echo
2264 * reply is prior to the skb transmission time
2266 static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2267 const struct sk_buff *skb)
2269 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2270 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp(skb));
2273 /* Nothing was retransmitted or returned timestamp is less
2274 * than timestamp of the first retransmission.
2276 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2278 return !tp->retrans_stamp ||
2279 tcp_tsopt_ecr_before(tp, tp->retrans_stamp);
2282 /* Undo procedures. */
2284 /* We can clear retrans_stamp when there are no retransmissions in the
2285 * window. It would seem that it is trivially available for us in
2286 * tp->retrans_out, however, that kind of assumptions doesn't consider
2287 * what will happen if errors occur when sending retransmission for the
2288 * second time. ...It could the that such segment has only
2289 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2290 * the head skb is enough except for some reneging corner cases that
2291 * are not worth the effort.
2293 * Main reason for all this complexity is the fact that connection dying
2294 * time now depends on the validity of the retrans_stamp, in particular,
2295 * that successive retransmissions of a segment must not advance
2296 * retrans_stamp under any conditions.
2298 static bool tcp_any_retrans_done(const struct sock *sk)
2300 const struct tcp_sock *tp = tcp_sk(sk);
2301 struct sk_buff *skb;
2303 if (tp->retrans_out)
2306 skb = tcp_write_queue_head(sk);
2307 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2313 #if FASTRETRANS_DEBUG > 1
2314 static void DBGUNDO(struct sock *sk, const char *msg)
2316 struct tcp_sock *tp = tcp_sk(sk);
2317 struct inet_sock *inet = inet_sk(sk);
2319 if (sk->sk_family == AF_INET) {
2320 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2322 &inet->inet_daddr, ntohs(inet->inet_dport),
2323 tp->snd_cwnd, tcp_left_out(tp),
2324 tp->snd_ssthresh, tp->prior_ssthresh,
2327 #if IS_ENABLED(CONFIG_IPV6)
2328 else if (sk->sk_family == AF_INET6) {
2329 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2331 &sk->sk_v6_daddr, ntohs(inet->inet_dport),
2332 tp->snd_cwnd, tcp_left_out(tp),
2333 tp->snd_ssthresh, tp->prior_ssthresh,
2339 #define DBGUNDO(x...) do { } while (0)
2342 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2344 struct tcp_sock *tp = tcp_sk(sk);
2347 struct sk_buff *skb;
2349 tcp_for_write_queue(skb, sk) {
2350 if (skb == tcp_send_head(sk))
2352 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2355 tcp_clear_all_retrans_hints(tp);
2358 if (tp->prior_ssthresh) {
2359 const struct inet_connection_sock *icsk = inet_csk(sk);
2361 if (icsk->icsk_ca_ops->undo_cwnd)
2362 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2364 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2366 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2367 tp->snd_ssthresh = tp->prior_ssthresh;
2368 tcp_ecn_withdraw_cwr(tp);
2371 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2373 tp->snd_cwnd_stamp = tcp_time_stamp;
2374 tp->undo_marker = 0;
2377 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2379 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2382 /* People celebrate: "We love our President!" */
2383 static bool tcp_try_undo_recovery(struct sock *sk)
2385 struct tcp_sock *tp = tcp_sk(sk);
2387 if (tcp_may_undo(tp)) {
2390 /* Happy end! We did not retransmit anything
2391 * or our original transmission succeeded.
2393 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2394 tcp_undo_cwnd_reduction(sk, false);
2395 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2396 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2398 mib_idx = LINUX_MIB_TCPFULLUNDO;
2400 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2402 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2403 /* Hold old state until something *above* high_seq
2404 * is ACKed. For Reno it is MUST to prevent false
2405 * fast retransmits (RFC2582). SACK TCP is safe. */
2406 tcp_moderate_cwnd(tp);
2407 if (!tcp_any_retrans_done(sk))
2408 tp->retrans_stamp = 0;
2411 tcp_set_ca_state(sk, TCP_CA_Open);
2415 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2416 static bool tcp_try_undo_dsack(struct sock *sk)
2418 struct tcp_sock *tp = tcp_sk(sk);
2420 if (tp->undo_marker && !tp->undo_retrans) {
2421 DBGUNDO(sk, "D-SACK");
2422 tcp_undo_cwnd_reduction(sk, false);
2423 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2429 /* Undo during loss recovery after partial ACK or using F-RTO. */
2430 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2432 struct tcp_sock *tp = tcp_sk(sk);
2434 if (frto_undo || tcp_may_undo(tp)) {
2435 tcp_undo_cwnd_reduction(sk, true);
2437 DBGUNDO(sk, "partial loss");
2438 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2440 NET_INC_STATS_BH(sock_net(sk),
2441 LINUX_MIB_TCPSPURIOUSRTOS);
2442 inet_csk(sk)->icsk_retransmits = 0;
2443 if (frto_undo || tcp_is_sack(tp))
2444 tcp_set_ca_state(sk, TCP_CA_Open);
2450 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2451 * It computes the number of packets to send (sndcnt) based on packets newly
2453 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2454 * cwnd reductions across a full RTT.
2455 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2456 * But when the retransmits are acked without further losses, PRR
2457 * slow starts cwnd up to ssthresh to speed up the recovery.
2459 static void tcp_init_cwnd_reduction(struct sock *sk)
2461 struct tcp_sock *tp = tcp_sk(sk);
2463 tp->high_seq = tp->snd_nxt;
2464 tp->tlp_high_seq = 0;
2465 tp->snd_cwnd_cnt = 0;
2466 tp->prior_cwnd = tp->snd_cwnd;
2467 tp->prr_delivered = 0;
2469 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2470 tcp_ecn_queue_cwr(tp);
2473 static void tcp_cwnd_reduction(struct sock *sk, const int prior_unsacked,
2474 int fast_rexmit, int flag)
2476 struct tcp_sock *tp = tcp_sk(sk);
2478 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2479 int newly_acked_sacked = prior_unsacked -
2480 (tp->packets_out - tp->sacked_out);
2482 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2485 tp->prr_delivered += newly_acked_sacked;
2487 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2489 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2490 } else if ((flag & FLAG_RETRANS_DATA_ACKED) &&
2491 !(flag & FLAG_LOST_RETRANS)) {
2492 sndcnt = min_t(int, delta,
2493 max_t(int, tp->prr_delivered - tp->prr_out,
2494 newly_acked_sacked) + 1);
2496 sndcnt = min(delta, newly_acked_sacked);
2498 sndcnt = max(sndcnt, (fast_rexmit ? 1 : 0));
2499 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2502 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2504 struct tcp_sock *tp = tcp_sk(sk);
2506 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2507 if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR ||
2508 (tp->undo_marker && tp->snd_ssthresh < TCP_INFINITE_SSTHRESH)) {
2509 tp->snd_cwnd = tp->snd_ssthresh;
2510 tp->snd_cwnd_stamp = tcp_time_stamp;
2512 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2515 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2516 void tcp_enter_cwr(struct sock *sk)
2518 struct tcp_sock *tp = tcp_sk(sk);
2520 tp->prior_ssthresh = 0;
2521 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2522 tp->undo_marker = 0;
2523 tcp_init_cwnd_reduction(sk);
2524 tcp_set_ca_state(sk, TCP_CA_CWR);
2527 EXPORT_SYMBOL(tcp_enter_cwr);
2529 static void tcp_try_keep_open(struct sock *sk)
2531 struct tcp_sock *tp = tcp_sk(sk);
2532 int state = TCP_CA_Open;
2534 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2535 state = TCP_CA_Disorder;
2537 if (inet_csk(sk)->icsk_ca_state != state) {
2538 tcp_set_ca_state(sk, state);
2539 tp->high_seq = tp->snd_nxt;
2543 static void tcp_try_to_open(struct sock *sk, int flag, const int prior_unsacked)
2545 struct tcp_sock *tp = tcp_sk(sk);
2547 tcp_verify_left_out(tp);
2549 if (!tcp_any_retrans_done(sk))
2550 tp->retrans_stamp = 0;
2552 if (flag & FLAG_ECE)
2555 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2556 tcp_try_keep_open(sk);
2558 tcp_cwnd_reduction(sk, prior_unsacked, 0, flag);
2562 static void tcp_mtup_probe_failed(struct sock *sk)
2564 struct inet_connection_sock *icsk = inet_csk(sk);
2566 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2567 icsk->icsk_mtup.probe_size = 0;
2568 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
2571 static void tcp_mtup_probe_success(struct sock *sk)
2573 struct tcp_sock *tp = tcp_sk(sk);
2574 struct inet_connection_sock *icsk = inet_csk(sk);
2576 /* FIXME: breaks with very large cwnd */
2577 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2578 tp->snd_cwnd = tp->snd_cwnd *
2579 tcp_mss_to_mtu(sk, tp->mss_cache) /
2580 icsk->icsk_mtup.probe_size;
2581 tp->snd_cwnd_cnt = 0;
2582 tp->snd_cwnd_stamp = tcp_time_stamp;
2583 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2585 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2586 icsk->icsk_mtup.probe_size = 0;
2587 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2588 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
2591 /* Do a simple retransmit without using the backoff mechanisms in
2592 * tcp_timer. This is used for path mtu discovery.
2593 * The socket is already locked here.
2595 void tcp_simple_retransmit(struct sock *sk)
2597 const struct inet_connection_sock *icsk = inet_csk(sk);
2598 struct tcp_sock *tp = tcp_sk(sk);
2599 struct sk_buff *skb;
2600 unsigned int mss = tcp_current_mss(sk);
2601 u32 prior_lost = tp->lost_out;
2603 tcp_for_write_queue(skb, sk) {
2604 if (skb == tcp_send_head(sk))
2606 if (tcp_skb_seglen(skb) > mss &&
2607 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2608 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2609 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2610 tp->retrans_out -= tcp_skb_pcount(skb);
2612 tcp_skb_mark_lost_uncond_verify(tp, skb);
2616 tcp_clear_retrans_hints_partial(tp);
2618 if (prior_lost == tp->lost_out)
2621 if (tcp_is_reno(tp))
2622 tcp_limit_reno_sacked(tp);
2624 tcp_verify_left_out(tp);
2626 /* Don't muck with the congestion window here.
2627 * Reason is that we do not increase amount of _data_
2628 * in network, but units changed and effective
2629 * cwnd/ssthresh really reduced now.
2631 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2632 tp->high_seq = tp->snd_nxt;
2633 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2634 tp->prior_ssthresh = 0;
2635 tp->undo_marker = 0;
2636 tcp_set_ca_state(sk, TCP_CA_Loss);
2638 tcp_xmit_retransmit_queue(sk);
2640 EXPORT_SYMBOL(tcp_simple_retransmit);
2642 static void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2644 struct tcp_sock *tp = tcp_sk(sk);
2647 if (tcp_is_reno(tp))
2648 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2650 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2652 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2654 tp->prior_ssthresh = 0;
2657 if (!tcp_in_cwnd_reduction(sk)) {
2659 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2660 tcp_init_cwnd_reduction(sk);
2662 tcp_set_ca_state(sk, TCP_CA_Recovery);
2665 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2666 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2668 static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack)
2670 struct tcp_sock *tp = tcp_sk(sk);
2671 bool recovered = !before(tp->snd_una, tp->high_seq);
2673 if ((flag & FLAG_SND_UNA_ADVANCED) &&
2674 tcp_try_undo_loss(sk, false))
2677 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2678 /* Step 3.b. A timeout is spurious if not all data are
2679 * lost, i.e., never-retransmitted data are (s)acked.
2681 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2682 tcp_try_undo_loss(sk, true))
2685 if (after(tp->snd_nxt, tp->high_seq)) {
2686 if (flag & FLAG_DATA_SACKED || is_dupack)
2687 tp->frto = 0; /* Step 3.a. loss was real */
2688 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2689 tp->high_seq = tp->snd_nxt;
2690 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
2692 if (after(tp->snd_nxt, tp->high_seq))
2693 return; /* Step 2.b */
2699 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2700 tcp_try_undo_recovery(sk);
2703 if (tcp_is_reno(tp)) {
2704 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2705 * delivered. Lower inflight to clock out (re)tranmissions.
2707 if (after(tp->snd_nxt, tp->high_seq) && is_dupack)
2708 tcp_add_reno_sack(sk);
2709 else if (flag & FLAG_SND_UNA_ADVANCED)
2710 tcp_reset_reno_sack(tp);
2712 tcp_xmit_retransmit_queue(sk);
2715 /* Undo during fast recovery after partial ACK. */
2716 static bool tcp_try_undo_partial(struct sock *sk, const int acked,
2717 const int prior_unsacked, int flag)
2719 struct tcp_sock *tp = tcp_sk(sk);
2721 if (tp->undo_marker && tcp_packet_delayed(tp)) {
2722 /* Plain luck! Hole if filled with delayed
2723 * packet, rather than with a retransmit.
2725 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2727 /* We are getting evidence that the reordering degree is higher
2728 * than we realized. If there are no retransmits out then we
2729 * can undo. Otherwise we clock out new packets but do not
2730 * mark more packets lost or retransmit more.
2732 if (tp->retrans_out) {
2733 tcp_cwnd_reduction(sk, prior_unsacked, 0, flag);
2737 if (!tcp_any_retrans_done(sk))
2738 tp->retrans_stamp = 0;
2740 DBGUNDO(sk, "partial recovery");
2741 tcp_undo_cwnd_reduction(sk, true);
2742 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2743 tcp_try_keep_open(sk);
2749 /* Process an event, which can update packets-in-flight not trivially.
2750 * Main goal of this function is to calculate new estimate for left_out,
2751 * taking into account both packets sitting in receiver's buffer and
2752 * packets lost by network.
2754 * Besides that it does CWND reduction, when packet loss is detected
2755 * and changes state of machine.
2757 * It does _not_ decide what to send, it is made in function
2758 * tcp_xmit_retransmit_queue().
2760 static void tcp_fastretrans_alert(struct sock *sk, const int acked,
2761 const int prior_unsacked,
2762 bool is_dupack, int flag)
2764 struct inet_connection_sock *icsk = inet_csk(sk);
2765 struct tcp_sock *tp = tcp_sk(sk);
2766 bool do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2767 (tcp_fackets_out(tp) > tp->reordering));
2768 int fast_rexmit = 0;
2770 if (WARN_ON(!tp->packets_out && tp->sacked_out))
2772 if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2773 tp->fackets_out = 0;
2775 /* Now state machine starts.
2776 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2777 if (flag & FLAG_ECE)
2778 tp->prior_ssthresh = 0;
2780 /* B. In all the states check for reneging SACKs. */
2781 if (tcp_check_sack_reneging(sk, flag))
2784 /* C. Check consistency of the current state. */
2785 tcp_verify_left_out(tp);
2787 /* D. Check state exit conditions. State can be terminated
2788 * when high_seq is ACKed. */
2789 if (icsk->icsk_ca_state == TCP_CA_Open) {
2790 WARN_ON(tp->retrans_out != 0);
2791 tp->retrans_stamp = 0;
2792 } else if (!before(tp->snd_una, tp->high_seq)) {
2793 switch (icsk->icsk_ca_state) {
2795 /* CWR is to be held something *above* high_seq
2796 * is ACKed for CWR bit to reach receiver. */
2797 if (tp->snd_una != tp->high_seq) {
2798 tcp_end_cwnd_reduction(sk);
2799 tcp_set_ca_state(sk, TCP_CA_Open);
2803 case TCP_CA_Recovery:
2804 if (tcp_is_reno(tp))
2805 tcp_reset_reno_sack(tp);
2806 if (tcp_try_undo_recovery(sk))
2808 tcp_end_cwnd_reduction(sk);
2813 /* Use RACK to detect loss */
2814 if (sysctl_tcp_recovery & TCP_RACK_LOST_RETRANS &&
2815 tcp_rack_mark_lost(sk))
2816 flag |= FLAG_LOST_RETRANS;
2818 /* E. Process state. */
2819 switch (icsk->icsk_ca_state) {
2820 case TCP_CA_Recovery:
2821 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2822 if (tcp_is_reno(tp) && is_dupack)
2823 tcp_add_reno_sack(sk);
2825 if (tcp_try_undo_partial(sk, acked, prior_unsacked, flag))
2827 /* Partial ACK arrived. Force fast retransmit. */
2828 do_lost = tcp_is_reno(tp) ||
2829 tcp_fackets_out(tp) > tp->reordering;
2831 if (tcp_try_undo_dsack(sk)) {
2832 tcp_try_keep_open(sk);
2837 tcp_process_loss(sk, flag, is_dupack);
2838 if (icsk->icsk_ca_state != TCP_CA_Open &&
2839 !(flag & FLAG_LOST_RETRANS))
2841 /* Change state if cwnd is undone or retransmits are lost */
2843 if (tcp_is_reno(tp)) {
2844 if (flag & FLAG_SND_UNA_ADVANCED)
2845 tcp_reset_reno_sack(tp);
2847 tcp_add_reno_sack(sk);
2850 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2851 tcp_try_undo_dsack(sk);
2853 if (!tcp_time_to_recover(sk, flag)) {
2854 tcp_try_to_open(sk, flag, prior_unsacked);
2858 /* MTU probe failure: don't reduce cwnd */
2859 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2860 icsk->icsk_mtup.probe_size &&
2861 tp->snd_una == tp->mtu_probe.probe_seq_start) {
2862 tcp_mtup_probe_failed(sk);
2863 /* Restores the reduction we did in tcp_mtup_probe() */
2865 tcp_simple_retransmit(sk);
2869 /* Otherwise enter Recovery state */
2870 tcp_enter_recovery(sk, (flag & FLAG_ECE));
2875 tcp_update_scoreboard(sk, fast_rexmit);
2876 tcp_cwnd_reduction(sk, prior_unsacked, fast_rexmit, flag);
2877 tcp_xmit_retransmit_queue(sk);
2880 /* Kathleen Nichols' algorithm for tracking the minimum value of
2881 * a data stream over some fixed time interval. (E.g., the minimum
2882 * RTT over the past five minutes.) It uses constant space and constant
2883 * time per update yet almost always delivers the same minimum as an
2884 * implementation that has to keep all the data in the window.
2886 * The algorithm keeps track of the best, 2nd best & 3rd best min
2887 * values, maintaining an invariant that the measurement time of the
2888 * n'th best >= n-1'th best. It also makes sure that the three values
2889 * are widely separated in the time window since that bounds the worse
2890 * case error when that data is monotonically increasing over the window.
2892 * Upon getting a new min, we can forget everything earlier because it
2893 * has no value - the new min is <= everything else in the window by
2894 * definition and it's the most recent. So we restart fresh on every new min
2895 * and overwrites 2nd & 3rd choices. The same property holds for 2nd & 3rd
2898 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us)
2900 const u32 now = tcp_time_stamp, wlen = sysctl_tcp_min_rtt_wlen * HZ;
2901 struct rtt_meas *m = tcp_sk(sk)->rtt_min;
2902 struct rtt_meas rttm = { .rtt = (rtt_us ? : 1), .ts = now };
2905 /* Check if the new measurement updates the 1st, 2nd, or 3rd choices */
2906 if (unlikely(rttm.rtt <= m[0].rtt))
2907 m[0] = m[1] = m[2] = rttm;
2908 else if (rttm.rtt <= m[1].rtt)
2910 else if (rttm.rtt <= m[2].rtt)
2913 elapsed = now - m[0].ts;
2914 if (unlikely(elapsed > wlen)) {
2915 /* Passed entire window without a new min so make 2nd choice
2916 * the new min & 3rd choice the new 2nd. So forth and so on.
2921 if (now - m[0].ts > wlen) {
2924 if (now - m[0].ts > wlen)
2927 } else if (m[1].ts == m[0].ts && elapsed > wlen / 4) {
2928 /* Passed a quarter of the window without a new min so
2929 * take 2nd choice from the 2nd quarter of the window.
2932 } else if (m[2].ts == m[1].ts && elapsed > wlen / 2) {
2933 /* Passed half the window without a new min so take the 3rd
2934 * choice from the last half of the window.
2940 static inline bool tcp_ack_update_rtt(struct sock *sk, const int flag,
2941 long seq_rtt_us, long sack_rtt_us,
2944 const struct tcp_sock *tp = tcp_sk(sk);
2946 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2947 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2948 * Karn's algorithm forbids taking RTT if some retransmitted data
2949 * is acked (RFC6298).
2952 seq_rtt_us = sack_rtt_us;
2954 /* RTTM Rule: A TSecr value received in a segment is used to
2955 * update the averaged RTT measurement only if the segment
2956 * acknowledges some new data, i.e., only if it advances the
2957 * left edge of the send window.
2958 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2960 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2962 seq_rtt_us = ca_rtt_us = jiffies_to_usecs(tcp_time_stamp -
2963 tp->rx_opt.rcv_tsecr);
2967 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
2968 * always taken together with ACK, SACK, or TS-opts. Any negative
2969 * values will be skipped with the seq_rtt_us < 0 check above.
2971 tcp_update_rtt_min(sk, ca_rtt_us);
2972 tcp_rtt_estimator(sk, seq_rtt_us);
2975 /* RFC6298: only reset backoff on valid RTT measurement. */
2976 inet_csk(sk)->icsk_backoff = 0;
2980 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2981 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
2985 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack.v64) {
2986 struct skb_mstamp now;
2988 skb_mstamp_get(&now);
2989 rtt_us = skb_mstamp_us_delta(&now, &tcp_rsk(req)->snt_synack);
2992 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us);
2996 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
2998 const struct inet_connection_sock *icsk = inet_csk(sk);
3000 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
3001 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
3004 /* Restart timer after forward progress on connection.
3005 * RFC2988 recommends to restart timer to now+rto.
3007 void tcp_rearm_rto(struct sock *sk)
3009 const struct inet_connection_sock *icsk = inet_csk(sk);
3010 struct tcp_sock *tp = tcp_sk(sk);
3012 /* If the retrans timer is currently being used by Fast Open
3013 * for SYN-ACK retrans purpose, stay put.
3015 if (tp->fastopen_rsk)
3018 if (!tp->packets_out) {
3019 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3021 u32 rto = inet_csk(sk)->icsk_rto;
3022 /* Offset the time elapsed after installing regular RTO */
3023 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
3024 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
3025 struct sk_buff *skb = tcp_write_queue_head(sk);
3026 const u32 rto_time_stamp =
3027 tcp_skb_timestamp(skb) + rto;
3028 s32 delta = (s32)(rto_time_stamp - tcp_time_stamp);
3029 /* delta may not be positive if the socket is locked
3030 * when the retrans timer fires and is rescheduled.
3035 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
3040 /* This function is called when the delayed ER timer fires. TCP enters
3041 * fast recovery and performs fast-retransmit.
3043 void tcp_resume_early_retransmit(struct sock *sk)
3045 struct tcp_sock *tp = tcp_sk(sk);
3049 /* Stop if ER is disabled after the delayed ER timer is scheduled */
3050 if (!tp->do_early_retrans)
3053 tcp_enter_recovery(sk, false);
3054 tcp_update_scoreboard(sk, 1);
3055 tcp_xmit_retransmit_queue(sk);
3058 /* If we get here, the whole TSO packet has not been acked. */
3059 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3061 struct tcp_sock *tp = tcp_sk(sk);
3064 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3066 packets_acked = tcp_skb_pcount(skb);
3067 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3069 packets_acked -= tcp_skb_pcount(skb);
3071 if (packets_acked) {
3072 BUG_ON(tcp_skb_pcount(skb) == 0);
3073 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3076 return packets_acked;
3079 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3082 const struct skb_shared_info *shinfo;
3084 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3085 if (likely(!(sk->sk_tsflags & SOF_TIMESTAMPING_TX_ACK)))
3088 shinfo = skb_shinfo(skb);
3089 if ((shinfo->tx_flags & SKBTX_ACK_TSTAMP) &&
3090 between(shinfo->tskey, prior_snd_una, tcp_sk(sk)->snd_una - 1))
3091 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK);
3094 /* Remove acknowledged frames from the retransmission queue. If our packet
3095 * is before the ack sequence we can discard it as it's confirmed to have
3096 * arrived at the other end.
3098 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3100 struct tcp_sacktag_state *sack)
3102 const struct inet_connection_sock *icsk = inet_csk(sk);
3103 struct skb_mstamp first_ackt, last_ackt, now;
3104 struct tcp_sock *tp = tcp_sk(sk);
3105 u32 prior_sacked = tp->sacked_out;
3106 u32 reord = tp->packets_out;
3107 bool fully_acked = true;
3108 long sack_rtt_us = -1L;
3109 long seq_rtt_us = -1L;
3110 long ca_rtt_us = -1L;
3111 struct sk_buff *skb;
3118 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3119 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3120 u8 sacked = scb->sacked;
3123 tcp_ack_tstamp(sk, skb, prior_snd_una);
3125 /* Determine how many packets and what bytes were acked, tso and else */
3126 if (after(scb->end_seq, tp->snd_una)) {
3127 if (tcp_skb_pcount(skb) == 1 ||
3128 !after(tp->snd_una, scb->seq))
3131 acked_pcount = tcp_tso_acked(sk, skb);
3135 fully_acked = false;
3137 /* Speedup tcp_unlink_write_queue() and next loop */
3138 prefetchw(skb->next);
3139 acked_pcount = tcp_skb_pcount(skb);
3142 if (unlikely(sacked & TCPCB_RETRANS)) {
3143 if (sacked & TCPCB_SACKED_RETRANS)
3144 tp->retrans_out -= acked_pcount;
3145 flag |= FLAG_RETRANS_DATA_ACKED;
3146 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3147 last_ackt = skb->skb_mstamp;
3148 WARN_ON_ONCE(last_ackt.v64 == 0);
3149 if (!first_ackt.v64)
3150 first_ackt = last_ackt;
3152 reord = min(pkts_acked, reord);
3153 if (!after(scb->end_seq, tp->high_seq))
3154 flag |= FLAG_ORIG_SACK_ACKED;
3157 if (sacked & TCPCB_SACKED_ACKED)
3158 tp->sacked_out -= acked_pcount;
3159 else if (tcp_is_sack(tp) && !tcp_skb_spurious_retrans(tp, skb))
3160 tcp_rack_advance(tp, &skb->skb_mstamp, sacked);
3161 if (sacked & TCPCB_LOST)
3162 tp->lost_out -= acked_pcount;
3164 tp->packets_out -= acked_pcount;
3165 pkts_acked += acked_pcount;
3167 /* Initial outgoing SYN's get put onto the write_queue
3168 * just like anything else we transmit. It is not
3169 * true data, and if we misinform our callers that
3170 * this ACK acks real data, we will erroneously exit
3171 * connection startup slow start one packet too
3172 * quickly. This is severely frowned upon behavior.
3174 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3175 flag |= FLAG_DATA_ACKED;
3177 flag |= FLAG_SYN_ACKED;
3178 tp->retrans_stamp = 0;
3184 tcp_unlink_write_queue(skb, sk);
3185 sk_wmem_free_skb(sk, skb);
3186 if (unlikely(skb == tp->retransmit_skb_hint))
3187 tp->retransmit_skb_hint = NULL;
3188 if (unlikely(skb == tp->lost_skb_hint))
3189 tp->lost_skb_hint = NULL;
3192 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3193 tp->snd_up = tp->snd_una;
3195 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3196 flag |= FLAG_SACK_RENEGING;
3198 skb_mstamp_get(&now);
3199 if (likely(first_ackt.v64) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
3200 seq_rtt_us = skb_mstamp_us_delta(&now, &first_ackt);
3201 ca_rtt_us = skb_mstamp_us_delta(&now, &last_ackt);
3203 if (sack->first_sackt.v64) {
3204 sack_rtt_us = skb_mstamp_us_delta(&now, &sack->first_sackt);
3205 ca_rtt_us = skb_mstamp_us_delta(&now, &sack->last_sackt);
3208 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
3211 if (flag & FLAG_ACKED) {
3213 if (unlikely(icsk->icsk_mtup.probe_size &&
3214 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3215 tcp_mtup_probe_success(sk);
3218 if (tcp_is_reno(tp)) {
3219 tcp_remove_reno_sacks(sk, pkts_acked);
3223 /* Non-retransmitted hole got filled? That's reordering */
3224 if (reord < prior_fackets && reord <= tp->fackets_out)
3225 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3227 delta = tcp_is_fack(tp) ? pkts_acked :
3228 prior_sacked - tp->sacked_out;
3229 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3232 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3234 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3235 sack_rtt_us > skb_mstamp_us_delta(&now, &skb->skb_mstamp)) {
3236 /* Do not re-arm RTO if the sack RTT is measured from data sent
3237 * after when the head was last (re)transmitted. Otherwise the
3238 * timeout may continue to extend in loss recovery.
3243 if (icsk->icsk_ca_ops->pkts_acked)
3244 icsk->icsk_ca_ops->pkts_acked(sk, pkts_acked, ca_rtt_us);
3246 #if FASTRETRANS_DEBUG > 0
3247 WARN_ON((int)tp->sacked_out < 0);
3248 WARN_ON((int)tp->lost_out < 0);
3249 WARN_ON((int)tp->retrans_out < 0);
3250 if (!tp->packets_out && tcp_is_sack(tp)) {
3251 icsk = inet_csk(sk);
3253 pr_debug("Leak l=%u %d\n",
3254 tp->lost_out, icsk->icsk_ca_state);
3257 if (tp->sacked_out) {
3258 pr_debug("Leak s=%u %d\n",
3259 tp->sacked_out, icsk->icsk_ca_state);
3262 if (tp->retrans_out) {
3263 pr_debug("Leak r=%u %d\n",
3264 tp->retrans_out, icsk->icsk_ca_state);
3265 tp->retrans_out = 0;
3272 static void tcp_ack_probe(struct sock *sk)
3274 const struct tcp_sock *tp = tcp_sk(sk);
3275 struct inet_connection_sock *icsk = inet_csk(sk);
3277 /* Was it a usable window open? */
3279 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3280 icsk->icsk_backoff = 0;
3281 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3282 /* Socket must be waked up by subsequent tcp_data_snd_check().
3283 * This function is not for random using!
3286 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
3288 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3293 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3295 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3296 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3299 /* Decide wheather to run the increase function of congestion control. */
3300 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3302 if (tcp_in_cwnd_reduction(sk))
3305 /* If reordering is high then always grow cwnd whenever data is
3306 * delivered regardless of its ordering. Otherwise stay conservative
3307 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3308 * new SACK or ECE mark may first advance cwnd here and later reduce
3309 * cwnd in tcp_fastretrans_alert() based on more states.
3311 if (tcp_sk(sk)->reordering > sysctl_tcp_reordering)
3312 return flag & FLAG_FORWARD_PROGRESS;
3314 return flag & FLAG_DATA_ACKED;
3317 /* Check that window update is acceptable.
3318 * The function assumes that snd_una<=ack<=snd_next.
3320 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3321 const u32 ack, const u32 ack_seq,
3324 return after(ack, tp->snd_una) ||
3325 after(ack_seq, tp->snd_wl1) ||
3326 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3329 /* If we update tp->snd_una, also update tp->bytes_acked */
3330 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3332 u32 delta = ack - tp->snd_una;
3334 u64_stats_update_begin(&tp->syncp);
3335 tp->bytes_acked += delta;
3336 u64_stats_update_end(&tp->syncp);
3340 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3341 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3343 u32 delta = seq - tp->rcv_nxt;
3345 u64_stats_update_begin(&tp->syncp);
3346 tp->bytes_received += delta;
3347 u64_stats_update_end(&tp->syncp);
3351 /* Update our send window.
3353 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3354 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3356 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3359 struct tcp_sock *tp = tcp_sk(sk);
3361 u32 nwin = ntohs(tcp_hdr(skb)->window);
3363 if (likely(!tcp_hdr(skb)->syn))
3364 nwin <<= tp->rx_opt.snd_wscale;
3366 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3367 flag |= FLAG_WIN_UPDATE;
3368 tcp_update_wl(tp, ack_seq);
3370 if (tp->snd_wnd != nwin) {
3373 /* Note, it is the only place, where
3374 * fast path is recovered for sending TCP.
3377 tcp_fast_path_check(sk);
3379 if (tcp_send_head(sk))
3380 tcp_slow_start_after_idle_check(sk);
3382 if (nwin > tp->max_window) {
3383 tp->max_window = nwin;
3384 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3389 tcp_snd_una_update(tp, ack);
3394 static bool __tcp_oow_rate_limited(struct net *net, int mib_idx,
3395 u32 *last_oow_ack_time)
3397 if (*last_oow_ack_time) {
3398 s32 elapsed = (s32)(tcp_time_stamp - *last_oow_ack_time);
3400 if (0 <= elapsed && elapsed < sysctl_tcp_invalid_ratelimit) {
3401 NET_INC_STATS_BH(net, mib_idx);
3402 return true; /* rate-limited: don't send yet! */
3406 *last_oow_ack_time = tcp_time_stamp;
3408 return false; /* not rate-limited: go ahead, send dupack now! */
3411 /* Return true if we're currently rate-limiting out-of-window ACKs and
3412 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3413 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3414 * attacks that send repeated SYNs or ACKs for the same connection. To
3415 * do this, we do not send a duplicate SYNACK or ACK if the remote
3416 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3418 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3419 int mib_idx, u32 *last_oow_ack_time)
3421 /* Data packets without SYNs are not likely part of an ACK loop. */
3422 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3426 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time);
3429 /* RFC 5961 7 [ACK Throttling] */
3430 static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb)
3432 /* unprotected vars, we dont care of overwrites */
3433 static u32 challenge_timestamp;
3434 static unsigned int challenge_count;
3435 struct tcp_sock *tp = tcp_sk(sk);
3438 /* First check our per-socket dupack rate limit. */
3439 if (__tcp_oow_rate_limited(sock_net(sk),
3440 LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3441 &tp->last_oow_ack_time))
3444 /* Then check host-wide RFC 5961 rate limit. */
3446 if (now != challenge_timestamp) {
3447 u32 half = (sysctl_tcp_challenge_ack_limit + 1) >> 1;
3449 challenge_timestamp = now;
3450 WRITE_ONCE(challenge_count, half +
3451 prandom_u32_max(sysctl_tcp_challenge_ack_limit));
3453 count = READ_ONCE(challenge_count);
3455 WRITE_ONCE(challenge_count, count - 1);
3456 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPCHALLENGEACK);
3461 static void tcp_store_ts_recent(struct tcp_sock *tp)
3463 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3464 tp->rx_opt.ts_recent_stamp = get_seconds();
3467 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3469 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3470 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3471 * extra check below makes sure this can only happen
3472 * for pure ACK frames. -DaveM
3474 * Not only, also it occurs for expired timestamps.
3477 if (tcp_paws_check(&tp->rx_opt, 0))
3478 tcp_store_ts_recent(tp);
3482 /* This routine deals with acks during a TLP episode.
3483 * We mark the end of a TLP episode on receiving TLP dupack or when
3484 * ack is after tlp_high_seq.
3485 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3487 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3489 struct tcp_sock *tp = tcp_sk(sk);
3491 if (before(ack, tp->tlp_high_seq))
3494 if (flag & FLAG_DSACKING_ACK) {
3495 /* This DSACK means original and TLP probe arrived; no loss */
3496 tp->tlp_high_seq = 0;
3497 } else if (after(ack, tp->tlp_high_seq)) {
3498 /* ACK advances: there was a loss, so reduce cwnd. Reset
3499 * tlp_high_seq in tcp_init_cwnd_reduction()
3501 tcp_init_cwnd_reduction(sk);
3502 tcp_set_ca_state(sk, TCP_CA_CWR);
3503 tcp_end_cwnd_reduction(sk);
3504 tcp_try_keep_open(sk);
3505 NET_INC_STATS_BH(sock_net(sk),
3506 LINUX_MIB_TCPLOSSPROBERECOVERY);
3507 } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3508 FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3509 /* Pure dupack: original and TLP probe arrived; no loss */
3510 tp->tlp_high_seq = 0;
3514 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3516 const struct inet_connection_sock *icsk = inet_csk(sk);
3518 if (icsk->icsk_ca_ops->in_ack_event)
3519 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3522 /* This routine deals with incoming acks, but not outgoing ones. */
3523 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3525 struct inet_connection_sock *icsk = inet_csk(sk);
3526 struct tcp_sock *tp = tcp_sk(sk);
3527 struct tcp_sacktag_state sack_state;
3528 u32 prior_snd_una = tp->snd_una;
3529 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3530 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3531 bool is_dupack = false;
3533 int prior_packets = tp->packets_out;
3534 const int prior_unsacked = tp->packets_out - tp->sacked_out;
3535 int acked = 0; /* Number of packets newly acked */
3537 sack_state.first_sackt.v64 = 0;
3539 /* We very likely will need to access write queue head. */
3540 prefetchw(sk->sk_write_queue.next);
3542 /* If the ack is older than previous acks
3543 * then we can probably ignore it.
3545 if (before(ack, prior_snd_una)) {
3546 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3547 if (before(ack, prior_snd_una - tp->max_window)) {
3548 tcp_send_challenge_ack(sk, skb);
3554 /* If the ack includes data we haven't sent yet, discard
3555 * this segment (RFC793 Section 3.9).
3557 if (after(ack, tp->snd_nxt))
3560 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
3561 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
3564 if (after(ack, prior_snd_una)) {
3565 flag |= FLAG_SND_UNA_ADVANCED;
3566 icsk->icsk_retransmits = 0;
3569 prior_fackets = tp->fackets_out;
3571 /* ts_recent update must be made after we are sure that the packet
3574 if (flag & FLAG_UPDATE_TS_RECENT)
3575 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3577 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3578 /* Window is constant, pure forward advance.
3579 * No more checks are required.
3580 * Note, we use the fact that SND.UNA>=SND.WL2.
3582 tcp_update_wl(tp, ack_seq);
3583 tcp_snd_una_update(tp, ack);
3584 flag |= FLAG_WIN_UPDATE;
3586 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3588 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3590 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3592 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3595 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3597 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3599 if (TCP_SKB_CB(skb)->sacked)
3600 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3603 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3605 ack_ev_flags |= CA_ACK_ECE;
3608 if (flag & FLAG_WIN_UPDATE)
3609 ack_ev_flags |= CA_ACK_WIN_UPDATE;
3611 tcp_in_ack_event(sk, ack_ev_flags);
3614 /* We passed data and got it acked, remove any soft error
3615 * log. Something worked...
3617 sk->sk_err_soft = 0;
3618 icsk->icsk_probes_out = 0;
3619 tp->rcv_tstamp = tcp_time_stamp;
3623 /* See if we can take anything off of the retransmit queue. */
3624 acked = tp->packets_out;
3625 flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una,
3627 acked -= tp->packets_out;
3629 if (tcp_ack_is_dubious(sk, flag)) {
3630 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3631 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3634 if (tp->tlp_high_seq)
3635 tcp_process_tlp_ack(sk, ack, flag);
3637 /* Advance cwnd if state allows */
3638 if (tcp_may_raise_cwnd(sk, flag))
3639 tcp_cong_avoid(sk, ack, acked);
3641 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) {
3642 struct dst_entry *dst = __sk_dst_get(sk);
3647 if (icsk->icsk_pending == ICSK_TIME_RETRANS)
3648 tcp_schedule_loss_probe(sk);
3649 tcp_update_pacing_rate(sk);
3653 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3654 if (flag & FLAG_DSACKING_ACK)
3655 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3657 /* If this ack opens up a zero window, clear backoff. It was
3658 * being used to time the probes, and is probably far higher than
3659 * it needs to be for normal retransmission.
3661 if (tcp_send_head(sk))
3664 if (tp->tlp_high_seq)
3665 tcp_process_tlp_ack(sk, ack, flag);
3669 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3673 /* If data was SACKed, tag it and see if we should send more data.
3674 * If data was DSACKed, see if we can undo a cwnd reduction.
3676 if (TCP_SKB_CB(skb)->sacked) {
3677 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3679 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3683 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3687 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3688 bool syn, struct tcp_fastopen_cookie *foc,
3691 /* Valid only in SYN or SYN-ACK with an even length. */
3692 if (!foc || !syn || len < 0 || (len & 1))
3695 if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3696 len <= TCP_FASTOPEN_COOKIE_MAX)
3697 memcpy(foc->val, cookie, len);
3704 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3705 * But, this can also be called on packets in the established flow when
3706 * the fast version below fails.
3708 void tcp_parse_options(const struct sk_buff *skb,
3709 struct tcp_options_received *opt_rx, int estab,
3710 struct tcp_fastopen_cookie *foc)
3712 const unsigned char *ptr;
3713 const struct tcphdr *th = tcp_hdr(skb);
3714 int length = (th->doff * 4) - sizeof(struct tcphdr);
3716 ptr = (const unsigned char *)(th + 1);
3717 opt_rx->saw_tstamp = 0;
3719 while (length > 0) {
3720 int opcode = *ptr++;
3726 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3731 if (opsize < 2) /* "silly options" */
3733 if (opsize > length)
3734 return; /* don't parse partial options */
3737 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3738 u16 in_mss = get_unaligned_be16(ptr);
3740 if (opt_rx->user_mss &&
3741 opt_rx->user_mss < in_mss)
3742 in_mss = opt_rx->user_mss;
3743 opt_rx->mss_clamp = in_mss;
3748 if (opsize == TCPOLEN_WINDOW && th->syn &&
3749 !estab && sysctl_tcp_window_scaling) {
3750 __u8 snd_wscale = *(__u8 *)ptr;
3751 opt_rx->wscale_ok = 1;
3752 if (snd_wscale > 14) {
3753 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3758 opt_rx->snd_wscale = snd_wscale;
3761 case TCPOPT_TIMESTAMP:
3762 if ((opsize == TCPOLEN_TIMESTAMP) &&
3763 ((estab && opt_rx->tstamp_ok) ||
3764 (!estab && sysctl_tcp_timestamps))) {
3765 opt_rx->saw_tstamp = 1;
3766 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3767 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3770 case TCPOPT_SACK_PERM:
3771 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3772 !estab && sysctl_tcp_sack) {
3773 opt_rx->sack_ok = TCP_SACK_SEEN;
3774 tcp_sack_reset(opt_rx);
3779 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3780 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3782 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3785 #ifdef CONFIG_TCP_MD5SIG
3788 * The MD5 Hash has already been
3789 * checked (see tcp_v{4,6}_do_rcv()).
3793 case TCPOPT_FASTOPEN:
3794 tcp_parse_fastopen_option(
3795 opsize - TCPOLEN_FASTOPEN_BASE,
3796 ptr, th->syn, foc, false);
3800 /* Fast Open option shares code 254 using a
3801 * 16 bits magic number.
3803 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
3804 get_unaligned_be16(ptr) ==
3805 TCPOPT_FASTOPEN_MAGIC)
3806 tcp_parse_fastopen_option(opsize -
3807 TCPOLEN_EXP_FASTOPEN_BASE,
3808 ptr + 2, th->syn, foc, true);
3817 EXPORT_SYMBOL(tcp_parse_options);
3819 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3821 const __be32 *ptr = (const __be32 *)(th + 1);
3823 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3824 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3825 tp->rx_opt.saw_tstamp = 1;
3827 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3830 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3832 tp->rx_opt.rcv_tsecr = 0;
3838 /* Fast parse options. This hopes to only see timestamps.
3839 * If it is wrong it falls back on tcp_parse_options().
3841 static bool tcp_fast_parse_options(const struct sk_buff *skb,
3842 const struct tcphdr *th, struct tcp_sock *tp)
3844 /* In the spirit of fast parsing, compare doff directly to constant
3845 * values. Because equality is used, short doff can be ignored here.
3847 if (th->doff == (sizeof(*th) / 4)) {
3848 tp->rx_opt.saw_tstamp = 0;
3850 } else if (tp->rx_opt.tstamp_ok &&
3851 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3852 if (tcp_parse_aligned_timestamp(tp, th))
3856 tcp_parse_options(skb, &tp->rx_opt, 1, NULL);
3857 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3858 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3863 #ifdef CONFIG_TCP_MD5SIG
3865 * Parse MD5 Signature option
3867 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3869 int length = (th->doff << 2) - sizeof(*th);
3870 const u8 *ptr = (const u8 *)(th + 1);
3872 /* If the TCP option is too short, we can short cut */
3873 if (length < TCPOLEN_MD5SIG)
3876 while (length > 0) {
3877 int opcode = *ptr++;
3888 if (opsize < 2 || opsize > length)
3890 if (opcode == TCPOPT_MD5SIG)
3891 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3898 EXPORT_SYMBOL(tcp_parse_md5sig_option);
3901 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3903 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3904 * it can pass through stack. So, the following predicate verifies that
3905 * this segment is not used for anything but congestion avoidance or
3906 * fast retransmit. Moreover, we even are able to eliminate most of such
3907 * second order effects, if we apply some small "replay" window (~RTO)
3908 * to timestamp space.
3910 * All these measures still do not guarantee that we reject wrapped ACKs
3911 * on networks with high bandwidth, when sequence space is recycled fastly,
3912 * but it guarantees that such events will be very rare and do not affect
3913 * connection seriously. This doesn't look nice, but alas, PAWS is really
3916 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3917 * states that events when retransmit arrives after original data are rare.
3918 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3919 * the biggest problem on large power networks even with minor reordering.
3920 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3921 * up to bandwidth of 18Gigabit/sec. 8) ]
3924 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3926 const struct tcp_sock *tp = tcp_sk(sk);
3927 const struct tcphdr *th = tcp_hdr(skb);
3928 u32 seq = TCP_SKB_CB(skb)->seq;
3929 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3931 return (/* 1. Pure ACK with correct sequence number. */
3932 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3934 /* 2. ... and duplicate ACK. */
3935 ack == tp->snd_una &&
3937 /* 3. ... and does not update window. */
3938 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3940 /* 4. ... and sits in replay window. */
3941 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3944 static inline bool tcp_paws_discard(const struct sock *sk,
3945 const struct sk_buff *skb)
3947 const struct tcp_sock *tp = tcp_sk(sk);
3949 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
3950 !tcp_disordered_ack(sk, skb);
3953 /* Check segment sequence number for validity.
3955 * Segment controls are considered valid, if the segment
3956 * fits to the window after truncation to the window. Acceptability
3957 * of data (and SYN, FIN, of course) is checked separately.
3958 * See tcp_data_queue(), for example.
3960 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3961 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3962 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3963 * (borrowed from freebsd)
3966 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
3968 return !before(end_seq, tp->rcv_wup) &&
3969 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3972 /* When we get a reset we do this. */
3973 void tcp_reset(struct sock *sk)
3975 /* We want the right error as BSD sees it (and indeed as we do). */
3976 switch (sk->sk_state) {
3978 sk->sk_err = ECONNREFUSED;
3980 case TCP_CLOSE_WAIT:
3986 sk->sk_err = ECONNRESET;
3988 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3991 if (!sock_flag(sk, SOCK_DEAD))
3992 sk->sk_error_report(sk);
3998 * Process the FIN bit. This now behaves as it is supposed to work
3999 * and the FIN takes effect when it is validly part of sequence
4000 * space. Not before when we get holes.
4002 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4003 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4006 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4007 * close and we go into CLOSING (and later onto TIME-WAIT)
4009 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4011 static void tcp_fin(struct sock *sk)
4013 struct tcp_sock *tp = tcp_sk(sk);
4015 inet_csk_schedule_ack(sk);
4017 sk->sk_shutdown |= RCV_SHUTDOWN;
4018 sock_set_flag(sk, SOCK_DONE);
4020 switch (sk->sk_state) {
4022 case TCP_ESTABLISHED:
4023 /* Move to CLOSE_WAIT */
4024 tcp_set_state(sk, TCP_CLOSE_WAIT);
4025 inet_csk(sk)->icsk_ack.pingpong = 1;
4028 case TCP_CLOSE_WAIT:
4030 /* Received a retransmission of the FIN, do
4035 /* RFC793: Remain in the LAST-ACK state. */
4039 /* This case occurs when a simultaneous close
4040 * happens, we must ack the received FIN and
4041 * enter the CLOSING state.
4044 tcp_set_state(sk, TCP_CLOSING);
4047 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4049 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4052 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4053 * cases we should never reach this piece of code.
4055 pr_err("%s: Impossible, sk->sk_state=%d\n",
4056 __func__, sk->sk_state);
4060 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4061 * Probably, we should reset in this case. For now drop them.
4063 __skb_queue_purge(&tp->out_of_order_queue);
4064 if (tcp_is_sack(tp))
4065 tcp_sack_reset(&tp->rx_opt);
4068 if (!sock_flag(sk, SOCK_DEAD)) {
4069 sk->sk_state_change(sk);
4071 /* Do not send POLL_HUP for half duplex close. */
4072 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4073 sk->sk_state == TCP_CLOSE)
4074 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4076 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4080 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4083 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4084 if (before(seq, sp->start_seq))
4085 sp->start_seq = seq;
4086 if (after(end_seq, sp->end_seq))
4087 sp->end_seq = end_seq;
4093 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4095 struct tcp_sock *tp = tcp_sk(sk);
4097 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4100 if (before(seq, tp->rcv_nxt))
4101 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4103 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4105 NET_INC_STATS_BH(sock_net(sk), mib_idx);
4107 tp->rx_opt.dsack = 1;
4108 tp->duplicate_sack[0].start_seq = seq;
4109 tp->duplicate_sack[0].end_seq = end_seq;
4113 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4115 struct tcp_sock *tp = tcp_sk(sk);
4117 if (!tp->rx_opt.dsack)
4118 tcp_dsack_set(sk, seq, end_seq);
4120 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4123 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4125 struct tcp_sock *tp = tcp_sk(sk);
4127 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4128 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4129 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4130 tcp_enter_quickack_mode(sk);
4132 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4133 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4135 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4136 end_seq = tp->rcv_nxt;
4137 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4144 /* These routines update the SACK block as out-of-order packets arrive or
4145 * in-order packets close up the sequence space.
4147 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4150 struct tcp_sack_block *sp = &tp->selective_acks[0];
4151 struct tcp_sack_block *swalk = sp + 1;
4153 /* See if the recent change to the first SACK eats into
4154 * or hits the sequence space of other SACK blocks, if so coalesce.
4156 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4157 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4160 /* Zap SWALK, by moving every further SACK up by one slot.
4161 * Decrease num_sacks.
4163 tp->rx_opt.num_sacks--;
4164 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4168 this_sack++, swalk++;
4172 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4174 struct tcp_sock *tp = tcp_sk(sk);
4175 struct tcp_sack_block *sp = &tp->selective_acks[0];
4176 int cur_sacks = tp->rx_opt.num_sacks;
4182 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4183 if (tcp_sack_extend(sp, seq, end_seq)) {
4184 /* Rotate this_sack to the first one. */
4185 for (; this_sack > 0; this_sack--, sp--)
4186 swap(*sp, *(sp - 1));
4188 tcp_sack_maybe_coalesce(tp);
4193 /* Could not find an adjacent existing SACK, build a new one,
4194 * put it at the front, and shift everyone else down. We
4195 * always know there is at least one SACK present already here.
4197 * If the sack array is full, forget about the last one.
4199 if (this_sack >= TCP_NUM_SACKS) {
4201 tp->rx_opt.num_sacks--;
4204 for (; this_sack > 0; this_sack--, sp--)
4208 /* Build the new head SACK, and we're done. */
4209 sp->start_seq = seq;
4210 sp->end_seq = end_seq;
4211 tp->rx_opt.num_sacks++;
4214 /* RCV.NXT advances, some SACKs should be eaten. */
4216 static void tcp_sack_remove(struct tcp_sock *tp)
4218 struct tcp_sack_block *sp = &tp->selective_acks[0];
4219 int num_sacks = tp->rx_opt.num_sacks;
4222 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4223 if (skb_queue_empty(&tp->out_of_order_queue)) {
4224 tp->rx_opt.num_sacks = 0;
4228 for (this_sack = 0; this_sack < num_sacks;) {
4229 /* Check if the start of the sack is covered by RCV.NXT. */
4230 if (!before(tp->rcv_nxt, sp->start_seq)) {
4233 /* RCV.NXT must cover all the block! */
4234 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4236 /* Zap this SACK, by moving forward any other SACKS. */
4237 for (i = this_sack+1; i < num_sacks; i++)
4238 tp->selective_acks[i-1] = tp->selective_acks[i];
4245 tp->rx_opt.num_sacks = num_sacks;
4249 * tcp_try_coalesce - try to merge skb to prior one
4252 * @from: buffer to add in queue
4253 * @fragstolen: pointer to boolean
4255 * Before queueing skb @from after @to, try to merge them
4256 * to reduce overall memory use and queue lengths, if cost is small.
4257 * Packets in ofo or receive queues can stay a long time.
4258 * Better try to coalesce them right now to avoid future collapses.
4259 * Returns true if caller should free @from instead of queueing it
4261 static bool tcp_try_coalesce(struct sock *sk,
4263 struct sk_buff *from,
4268 *fragstolen = false;
4270 /* Its possible this segment overlaps with prior segment in queue */
4271 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4274 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4277 atomic_add(delta, &sk->sk_rmem_alloc);
4278 sk_mem_charge(sk, delta);
4279 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4280 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4281 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4282 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4286 /* This one checks to see if we can put data from the
4287 * out_of_order queue into the receive_queue.
4289 static void tcp_ofo_queue(struct sock *sk)
4291 struct tcp_sock *tp = tcp_sk(sk);
4292 __u32 dsack_high = tp->rcv_nxt;
4293 struct sk_buff *skb, *tail;
4294 bool fragstolen, eaten;
4296 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4297 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4300 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4301 __u32 dsack = dsack_high;
4302 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4303 dsack_high = TCP_SKB_CB(skb)->end_seq;
4304 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4307 __skb_unlink(skb, &tp->out_of_order_queue);
4308 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4309 SOCK_DEBUG(sk, "ofo packet was already received\n");
4313 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4314 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4315 TCP_SKB_CB(skb)->end_seq);
4317 tail = skb_peek_tail(&sk->sk_receive_queue);
4318 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4319 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4321 __skb_queue_tail(&sk->sk_receive_queue, skb);
4322 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4325 kfree_skb_partial(skb, fragstolen);
4329 static bool tcp_prune_ofo_queue(struct sock *sk);
4330 static int tcp_prune_queue(struct sock *sk);
4332 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4335 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4336 !sk_rmem_schedule(sk, skb, size)) {
4338 if (tcp_prune_queue(sk) < 0)
4341 if (!sk_rmem_schedule(sk, skb, size)) {
4342 if (!tcp_prune_ofo_queue(sk))
4345 if (!sk_rmem_schedule(sk, skb, size))
4352 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4354 struct tcp_sock *tp = tcp_sk(sk);
4355 struct sk_buff *skb1;
4358 tcp_ecn_check_ce(tp, skb);
4360 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4361 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFODROP);
4366 /* Disable header prediction. */
4368 inet_csk_schedule_ack(sk);
4370 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4371 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4372 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4374 skb1 = skb_peek_tail(&tp->out_of_order_queue);
4376 /* Initial out of order segment, build 1 SACK. */
4377 if (tcp_is_sack(tp)) {
4378 tp->rx_opt.num_sacks = 1;
4379 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4380 tp->selective_acks[0].end_seq =
4381 TCP_SKB_CB(skb)->end_seq;
4383 __skb_queue_head(&tp->out_of_order_queue, skb);
4387 seq = TCP_SKB_CB(skb)->seq;
4388 end_seq = TCP_SKB_CB(skb)->end_seq;
4390 if (seq == TCP_SKB_CB(skb1)->end_seq) {
4393 if (!tcp_try_coalesce(sk, skb1, skb, &fragstolen)) {
4394 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4396 tcp_grow_window(sk, skb);
4397 kfree_skb_partial(skb, fragstolen);
4401 if (!tp->rx_opt.num_sacks ||
4402 tp->selective_acks[0].end_seq != seq)
4405 /* Common case: data arrive in order after hole. */
4406 tp->selective_acks[0].end_seq = end_seq;
4410 /* Find place to insert this segment. */
4412 if (!after(TCP_SKB_CB(skb1)->seq, seq))
4414 if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
4418 skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
4421 /* Do skb overlap to previous one? */
4422 if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4423 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4424 /* All the bits are present. Drop. */
4425 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4428 tcp_dsack_set(sk, seq, end_seq);
4431 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4432 /* Partial overlap. */
4433 tcp_dsack_set(sk, seq,
4434 TCP_SKB_CB(skb1)->end_seq);
4436 if (skb_queue_is_first(&tp->out_of_order_queue,
4440 skb1 = skb_queue_prev(
4441 &tp->out_of_order_queue,
4446 __skb_queue_head(&tp->out_of_order_queue, skb);
4448 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4450 /* And clean segments covered by new one as whole. */
4451 while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
4452 skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
4454 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4456 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4457 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4461 __skb_unlink(skb1, &tp->out_of_order_queue);
4462 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4463 TCP_SKB_CB(skb1)->end_seq);
4464 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4469 if (tcp_is_sack(tp))
4470 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4473 tcp_grow_window(sk, skb);
4474 skb_set_owner_r(skb, sk);
4478 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
4482 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4484 __skb_pull(skb, hdrlen);
4486 tcp_try_coalesce(sk, tail, skb, fragstolen)) ? 1 : 0;
4487 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
4489 __skb_queue_tail(&sk->sk_receive_queue, skb);
4490 skb_set_owner_r(skb, sk);
4495 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4497 struct sk_buff *skb;
4505 if (size > PAGE_SIZE) {
4506 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
4508 data_len = npages << PAGE_SHIFT;
4509 size = data_len + (size & ~PAGE_MASK);
4511 skb = alloc_skb_with_frags(size - data_len, data_len,
4512 PAGE_ALLOC_COSTLY_ORDER,
4513 &err, sk->sk_allocation);
4517 skb_put(skb, size - data_len);
4518 skb->data_len = data_len;
4521 if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4524 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
4528 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4529 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4530 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4532 if (tcp_queue_rcv(sk, skb, 0, &fragstolen)) {
4533 WARN_ON_ONCE(fragstolen); /* should not happen */
4545 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4547 struct tcp_sock *tp = tcp_sk(sk);
4549 bool fragstolen = false;
4551 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4555 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
4557 tcp_ecn_accept_cwr(tp, skb);
4559 tp->rx_opt.dsack = 0;
4561 /* Queue data for delivery to the user.
4562 * Packets in sequence go to the receive queue.
4563 * Out of sequence packets to the out_of_order_queue.
4565 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4566 if (tcp_receive_window(tp) == 0)
4569 /* Ok. In sequence. In window. */
4570 if (tp->ucopy.task == current &&
4571 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4572 sock_owned_by_user(sk) && !tp->urg_data) {
4573 int chunk = min_t(unsigned int, skb->len,
4576 __set_current_state(TASK_RUNNING);
4579 if (!skb_copy_datagram_msg(skb, 0, tp->ucopy.msg, chunk)) {
4580 tp->ucopy.len -= chunk;
4581 tp->copied_seq += chunk;
4582 eaten = (chunk == skb->len);
4583 tcp_rcv_space_adjust(sk);
4591 if (skb_queue_len(&sk->sk_receive_queue) == 0)
4592 sk_forced_mem_schedule(sk, skb->truesize);
4593 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4596 eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
4598 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4600 tcp_event_data_recv(sk, skb);
4601 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4604 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4607 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4608 * gap in queue is filled.
4610 if (skb_queue_empty(&tp->out_of_order_queue))
4611 inet_csk(sk)->icsk_ack.pingpong = 0;
4614 if (tp->rx_opt.num_sacks)
4615 tcp_sack_remove(tp);
4617 tcp_fast_path_check(sk);
4620 kfree_skb_partial(skb, fragstolen);
4621 if (!sock_flag(sk, SOCK_DEAD))
4622 sk->sk_data_ready(sk);
4626 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4627 /* A retransmit, 2nd most common case. Force an immediate ack. */
4628 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4629 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4632 tcp_enter_quickack_mode(sk);
4633 inet_csk_schedule_ack(sk);
4639 /* Out of window. F.e. zero window probe. */
4640 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4643 tcp_enter_quickack_mode(sk);
4645 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4646 /* Partial packet, seq < rcv_next < end_seq */
4647 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4648 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4649 TCP_SKB_CB(skb)->end_seq);
4651 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4653 /* If window is closed, drop tail of packet. But after
4654 * remembering D-SACK for its head made in previous line.
4656 if (!tcp_receive_window(tp))
4661 tcp_data_queue_ofo(sk, skb);
4664 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4665 struct sk_buff_head *list)
4667 struct sk_buff *next = NULL;
4669 if (!skb_queue_is_last(list, skb))
4670 next = skb_queue_next(list, skb);
4672 __skb_unlink(skb, list);
4674 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4679 /* Collapse contiguous sequence of skbs head..tail with
4680 * sequence numbers start..end.
4682 * If tail is NULL, this means until the end of the list.
4684 * Segments with FIN/SYN are not collapsed (only because this
4688 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4689 struct sk_buff *head, struct sk_buff *tail,
4692 struct sk_buff *skb, *n;
4695 /* First, check that queue is collapsible and find
4696 * the point where collapsing can be useful. */
4700 skb_queue_walk_from_safe(list, skb, n) {
4703 /* No new bits? It is possible on ofo queue. */
4704 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4705 skb = tcp_collapse_one(sk, skb, list);
4711 /* The first skb to collapse is:
4713 * - bloated or contains data before "start" or
4714 * overlaps to the next one.
4716 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
4717 (tcp_win_from_space(skb->truesize) > skb->len ||
4718 before(TCP_SKB_CB(skb)->seq, start))) {
4719 end_of_skbs = false;
4723 if (!skb_queue_is_last(list, skb)) {
4724 struct sk_buff *next = skb_queue_next(list, skb);
4726 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
4727 end_of_skbs = false;
4732 /* Decided to skip this, advance start seq. */
4733 start = TCP_SKB_CB(skb)->end_seq;
4736 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4739 while (before(start, end)) {
4740 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
4741 struct sk_buff *nskb;
4743 nskb = alloc_skb(copy, GFP_ATOMIC);
4747 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4748 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4749 __skb_queue_before(list, skb, nskb);
4750 skb_set_owner_r(nskb, sk);
4752 /* Copy data, releasing collapsed skbs. */
4754 int offset = start - TCP_SKB_CB(skb)->seq;
4755 int size = TCP_SKB_CB(skb)->end_seq - start;
4759 size = min(copy, size);
4760 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4762 TCP_SKB_CB(nskb)->end_seq += size;
4766 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4767 skb = tcp_collapse_one(sk, skb, list);
4770 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4777 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4778 * and tcp_collapse() them until all the queue is collapsed.
4780 static void tcp_collapse_ofo_queue(struct sock *sk)
4782 struct tcp_sock *tp = tcp_sk(sk);
4783 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4784 struct sk_buff *head;
4790 start = TCP_SKB_CB(skb)->seq;
4791 end = TCP_SKB_CB(skb)->end_seq;
4795 struct sk_buff *next = NULL;
4797 if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
4798 next = skb_queue_next(&tp->out_of_order_queue, skb);
4801 /* Segment is terminated when we see gap or when
4802 * we are at the end of all the queue. */
4804 after(TCP_SKB_CB(skb)->seq, end) ||
4805 before(TCP_SKB_CB(skb)->end_seq, start)) {
4806 tcp_collapse(sk, &tp->out_of_order_queue,
4807 head, skb, start, end);
4811 /* Start new segment */
4812 start = TCP_SKB_CB(skb)->seq;
4813 end = TCP_SKB_CB(skb)->end_seq;
4815 if (before(TCP_SKB_CB(skb)->seq, start))
4816 start = TCP_SKB_CB(skb)->seq;
4817 if (after(TCP_SKB_CB(skb)->end_seq, end))
4818 end = TCP_SKB_CB(skb)->end_seq;
4824 * Purge the out-of-order queue.
4825 * Return true if queue was pruned.
4827 static bool tcp_prune_ofo_queue(struct sock *sk)
4829 struct tcp_sock *tp = tcp_sk(sk);
4832 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4833 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
4834 __skb_queue_purge(&tp->out_of_order_queue);
4836 /* Reset SACK state. A conforming SACK implementation will
4837 * do the same at a timeout based retransmit. When a connection
4838 * is in a sad state like this, we care only about integrity
4839 * of the connection not performance.
4841 if (tp->rx_opt.sack_ok)
4842 tcp_sack_reset(&tp->rx_opt);
4849 /* Reduce allocated memory if we can, trying to get
4850 * the socket within its memory limits again.
4852 * Return less than zero if we should start dropping frames
4853 * until the socket owning process reads some of the data
4854 * to stabilize the situation.
4856 static int tcp_prune_queue(struct sock *sk)
4858 struct tcp_sock *tp = tcp_sk(sk);
4860 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4862 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
4864 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4865 tcp_clamp_window(sk);
4866 else if (tcp_under_memory_pressure(sk))
4867 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4869 tcp_collapse_ofo_queue(sk);
4870 if (!skb_queue_empty(&sk->sk_receive_queue))
4871 tcp_collapse(sk, &sk->sk_receive_queue,
4872 skb_peek(&sk->sk_receive_queue),
4874 tp->copied_seq, tp->rcv_nxt);
4877 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4880 /* Collapsing did not help, destructive actions follow.
4881 * This must not ever occur. */
4883 tcp_prune_ofo_queue(sk);
4885 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4888 /* If we are really being abused, tell the caller to silently
4889 * drop receive data on the floor. It will get retransmitted
4890 * and hopefully then we'll have sufficient space.
4892 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
4894 /* Massive buffer overcommit. */
4899 static bool tcp_should_expand_sndbuf(const struct sock *sk)
4901 const struct tcp_sock *tp = tcp_sk(sk);
4903 /* If the user specified a specific send buffer setting, do
4906 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4909 /* If we are under global TCP memory pressure, do not expand. */
4910 if (tcp_under_memory_pressure(sk))
4913 /* If we are under soft global TCP memory pressure, do not expand. */
4914 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
4917 /* If we filled the congestion window, do not expand. */
4918 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
4924 /* When incoming ACK allowed to free some skb from write_queue,
4925 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4926 * on the exit from tcp input handler.
4928 * PROBLEM: sndbuf expansion does not work well with largesend.
4930 static void tcp_new_space(struct sock *sk)
4932 struct tcp_sock *tp = tcp_sk(sk);
4934 if (tcp_should_expand_sndbuf(sk)) {
4935 tcp_sndbuf_expand(sk);
4936 tp->snd_cwnd_stamp = tcp_time_stamp;
4939 sk->sk_write_space(sk);
4942 static void tcp_check_space(struct sock *sk)
4944 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4945 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4946 /* pairs with tcp_poll() */
4947 smp_mb__after_atomic();
4948 if (sk->sk_socket &&
4949 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4954 static inline void tcp_data_snd_check(struct sock *sk)
4956 tcp_push_pending_frames(sk);
4957 tcp_check_space(sk);
4961 * Check if sending an ack is needed.
4963 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4965 struct tcp_sock *tp = tcp_sk(sk);
4967 /* More than one full frame received... */
4968 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
4969 /* ... and right edge of window advances far enough.
4970 * (tcp_recvmsg() will send ACK otherwise). Or...
4972 __tcp_select_window(sk) >= tp->rcv_wnd) ||
4973 /* We ACK each frame or... */
4974 tcp_in_quickack_mode(sk) ||
4975 /* We have out of order data. */
4976 (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
4977 /* Then ack it now */
4980 /* Else, send delayed ack. */
4981 tcp_send_delayed_ack(sk);
4985 static inline void tcp_ack_snd_check(struct sock *sk)
4987 if (!inet_csk_ack_scheduled(sk)) {
4988 /* We sent a data segment already. */
4991 __tcp_ack_snd_check(sk, 1);
4995 * This routine is only called when we have urgent data
4996 * signaled. Its the 'slow' part of tcp_urg. It could be
4997 * moved inline now as tcp_urg is only called from one
4998 * place. We handle URGent data wrong. We have to - as
4999 * BSD still doesn't use the correction from RFC961.
5000 * For 1003.1g we should support a new option TCP_STDURG to permit
5001 * either form (or just set the sysctl tcp_stdurg).
5004 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5006 struct tcp_sock *tp = tcp_sk(sk);
5007 u32 ptr = ntohs(th->urg_ptr);
5009 if (ptr && !sysctl_tcp_stdurg)
5011 ptr += ntohl(th->seq);
5013 /* Ignore urgent data that we've already seen and read. */
5014 if (after(tp->copied_seq, ptr))
5017 /* Do not replay urg ptr.
5019 * NOTE: interesting situation not covered by specs.
5020 * Misbehaving sender may send urg ptr, pointing to segment,
5021 * which we already have in ofo queue. We are not able to fetch
5022 * such data and will stay in TCP_URG_NOTYET until will be eaten
5023 * by recvmsg(). Seems, we are not obliged to handle such wicked
5024 * situations. But it is worth to think about possibility of some
5025 * DoSes using some hypothetical application level deadlock.
5027 if (before(ptr, tp->rcv_nxt))
5030 /* Do we already have a newer (or duplicate) urgent pointer? */
5031 if (tp->urg_data && !after(ptr, tp->urg_seq))
5034 /* Tell the world about our new urgent pointer. */
5037 /* We may be adding urgent data when the last byte read was
5038 * urgent. To do this requires some care. We cannot just ignore
5039 * tp->copied_seq since we would read the last urgent byte again
5040 * as data, nor can we alter copied_seq until this data arrives
5041 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5043 * NOTE. Double Dutch. Rendering to plain English: author of comment
5044 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5045 * and expect that both A and B disappear from stream. This is _wrong_.
5046 * Though this happens in BSD with high probability, this is occasional.
5047 * Any application relying on this is buggy. Note also, that fix "works"
5048 * only in this artificial test. Insert some normal data between A and B and we will
5049 * decline of BSD again. Verdict: it is better to remove to trap
5052 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5053 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5054 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5056 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5057 __skb_unlink(skb, &sk->sk_receive_queue);
5062 tp->urg_data = TCP_URG_NOTYET;
5065 /* Disable header prediction. */
5069 /* This is the 'fast' part of urgent handling. */
5070 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5072 struct tcp_sock *tp = tcp_sk(sk);
5074 /* Check if we get a new urgent pointer - normally not. */
5076 tcp_check_urg(sk, th);
5078 /* Do we wait for any urgent data? - normally not... */
5079 if (tp->urg_data == TCP_URG_NOTYET) {
5080 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5083 /* Is the urgent pointer pointing into this packet? */
5084 if (ptr < skb->len) {
5086 if (skb_copy_bits(skb, ptr, &tmp, 1))
5088 tp->urg_data = TCP_URG_VALID | tmp;
5089 if (!sock_flag(sk, SOCK_DEAD))
5090 sk->sk_data_ready(sk);
5095 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
5097 struct tcp_sock *tp = tcp_sk(sk);
5098 int chunk = skb->len - hlen;
5102 if (skb_csum_unnecessary(skb))
5103 err = skb_copy_datagram_msg(skb, hlen, tp->ucopy.msg, chunk);
5105 err = skb_copy_and_csum_datagram_msg(skb, hlen, tp->ucopy.msg);
5108 tp->ucopy.len -= chunk;
5109 tp->copied_seq += chunk;
5110 tcp_rcv_space_adjust(sk);
5117 static __sum16 __tcp_checksum_complete_user(struct sock *sk,
5118 struct sk_buff *skb)
5122 if (sock_owned_by_user(sk)) {
5124 result = __tcp_checksum_complete(skb);
5127 result = __tcp_checksum_complete(skb);
5132 static inline bool tcp_checksum_complete_user(struct sock *sk,
5133 struct sk_buff *skb)
5135 return !skb_csum_unnecessary(skb) &&
5136 __tcp_checksum_complete_user(sk, skb);
5139 /* Does PAWS and seqno based validation of an incoming segment, flags will
5140 * play significant role here.
5142 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5143 const struct tcphdr *th, int syn_inerr)
5145 struct tcp_sock *tp = tcp_sk(sk);
5147 /* RFC1323: H1. Apply PAWS check first. */
5148 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
5149 tcp_paws_discard(sk, skb)) {
5151 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5152 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5153 LINUX_MIB_TCPACKSKIPPEDPAWS,
5154 &tp->last_oow_ack_time))
5155 tcp_send_dupack(sk, skb);
5158 /* Reset is accepted even if it did not pass PAWS. */
5161 /* Step 1: check sequence number */
5162 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5163 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5164 * (RST) segments are validated by checking their SEQ-fields."
5165 * And page 69: "If an incoming segment is not acceptable,
5166 * an acknowledgment should be sent in reply (unless the RST
5167 * bit is set, if so drop the segment and return)".
5172 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5173 LINUX_MIB_TCPACKSKIPPEDSEQ,
5174 &tp->last_oow_ack_time))
5175 tcp_send_dupack(sk, skb);
5180 /* Step 2: check RST bit */
5183 * If sequence number exactly matches RCV.NXT, then
5184 * RESET the connection
5186 * Send a challenge ACK
5188 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt)
5191 tcp_send_challenge_ack(sk, skb);
5195 /* step 3: check security and precedence [ignored] */
5197 /* step 4: Check for a SYN
5198 * RFC 5961 4.2 : Send a challenge ack
5203 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5204 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5205 tcp_send_challenge_ack(sk, skb);
5217 * TCP receive function for the ESTABLISHED state.
5219 * It is split into a fast path and a slow path. The fast path is
5221 * - A zero window was announced from us - zero window probing
5222 * is only handled properly in the slow path.
5223 * - Out of order segments arrived.
5224 * - Urgent data is expected.
5225 * - There is no buffer space left
5226 * - Unexpected TCP flags/window values/header lengths are received
5227 * (detected by checking the TCP header against pred_flags)
5228 * - Data is sent in both directions. Fast path only supports pure senders
5229 * or pure receivers (this means either the sequence number or the ack
5230 * value must stay constant)
5231 * - Unexpected TCP option.
5233 * When these conditions are not satisfied it drops into a standard
5234 * receive procedure patterned after RFC793 to handle all cases.
5235 * The first three cases are guaranteed by proper pred_flags setting,
5236 * the rest is checked inline. Fast processing is turned on in
5237 * tcp_data_queue when everything is OK.
5239 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5240 const struct tcphdr *th, unsigned int len)
5242 struct tcp_sock *tp = tcp_sk(sk);
5244 if (unlikely(!sk->sk_rx_dst))
5245 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5247 * Header prediction.
5248 * The code loosely follows the one in the famous
5249 * "30 instruction TCP receive" Van Jacobson mail.
5251 * Van's trick is to deposit buffers into socket queue
5252 * on a device interrupt, to call tcp_recv function
5253 * on the receive process context and checksum and copy
5254 * the buffer to user space. smart...
5256 * Our current scheme is not silly either but we take the
5257 * extra cost of the net_bh soft interrupt processing...
5258 * We do checksum and copy also but from device to kernel.
5261 tp->rx_opt.saw_tstamp = 0;
5263 /* pred_flags is 0xS?10 << 16 + snd_wnd
5264 * if header_prediction is to be made
5265 * 'S' will always be tp->tcp_header_len >> 2
5266 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5267 * turn it off (when there are holes in the receive
5268 * space for instance)
5269 * PSH flag is ignored.
5272 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5273 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5274 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5275 int tcp_header_len = tp->tcp_header_len;
5277 /* Timestamp header prediction: tcp_header_len
5278 * is automatically equal to th->doff*4 due to pred_flags
5282 /* Check timestamp */
5283 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5284 /* No? Slow path! */
5285 if (!tcp_parse_aligned_timestamp(tp, th))
5288 /* If PAWS failed, check it more carefully in slow path */
5289 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5292 /* DO NOT update ts_recent here, if checksum fails
5293 * and timestamp was corrupted part, it will result
5294 * in a hung connection since we will drop all
5295 * future packets due to the PAWS test.
5299 if (len <= tcp_header_len) {
5300 /* Bulk data transfer: sender */
5301 if (len == tcp_header_len) {
5302 /* Predicted packet is in window by definition.
5303 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5304 * Hence, check seq<=rcv_wup reduces to:
5306 if (tcp_header_len ==
5307 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5308 tp->rcv_nxt == tp->rcv_wup)
5309 tcp_store_ts_recent(tp);
5311 /* We know that such packets are checksummed
5314 tcp_ack(sk, skb, 0);
5316 tcp_data_snd_check(sk);
5318 } else { /* Header too small */
5319 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5324 bool fragstolen = false;
5326 if (tp->ucopy.task == current &&
5327 tp->copied_seq == tp->rcv_nxt &&
5328 len - tcp_header_len <= tp->ucopy.len &&
5329 sock_owned_by_user(sk)) {
5330 __set_current_state(TASK_RUNNING);
5332 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) {
5333 /* Predicted packet is in window by definition.
5334 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5335 * Hence, check seq<=rcv_wup reduces to:
5337 if (tcp_header_len ==
5338 (sizeof(struct tcphdr) +
5339 TCPOLEN_TSTAMP_ALIGNED) &&
5340 tp->rcv_nxt == tp->rcv_wup)
5341 tcp_store_ts_recent(tp);
5343 tcp_rcv_rtt_measure_ts(sk, skb);
5345 __skb_pull(skb, tcp_header_len);
5346 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
5347 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
5352 if (tcp_checksum_complete_user(sk, skb))
5355 if ((int)skb->truesize > sk->sk_forward_alloc)
5358 /* Predicted packet is in window by definition.
5359 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5360 * Hence, check seq<=rcv_wup reduces to:
5362 if (tcp_header_len ==
5363 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5364 tp->rcv_nxt == tp->rcv_wup)
5365 tcp_store_ts_recent(tp);
5367 tcp_rcv_rtt_measure_ts(sk, skb);
5369 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
5371 /* Bulk data transfer: receiver */
5372 eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5376 tcp_event_data_recv(sk, skb);
5378 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5379 /* Well, only one small jumplet in fast path... */
5380 tcp_ack(sk, skb, FLAG_DATA);
5381 tcp_data_snd_check(sk);
5382 if (!inet_csk_ack_scheduled(sk))
5386 __tcp_ack_snd_check(sk, 0);
5389 kfree_skb_partial(skb, fragstolen);
5390 sk->sk_data_ready(sk);
5396 if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
5399 if (!th->ack && !th->rst && !th->syn)
5403 * Standard slow path.
5406 if (!tcp_validate_incoming(sk, skb, th, 1))
5410 if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5413 tcp_rcv_rtt_measure_ts(sk, skb);
5415 /* Process urgent data. */
5416 tcp_urg(sk, skb, th);
5418 /* step 7: process the segment text */
5419 tcp_data_queue(sk, skb);
5421 tcp_data_snd_check(sk);
5422 tcp_ack_snd_check(sk);
5426 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_CSUMERRORS);
5427 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5432 EXPORT_SYMBOL(tcp_rcv_established);
5434 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5436 struct tcp_sock *tp = tcp_sk(sk);
5437 struct inet_connection_sock *icsk = inet_csk(sk);
5439 tcp_set_state(sk, TCP_ESTABLISHED);
5440 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5443 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5444 security_inet_conn_established(sk, skb);
5447 /* Make sure socket is routed, for correct metrics. */
5448 icsk->icsk_af_ops->rebuild_header(sk);
5450 tcp_init_metrics(sk);
5452 tcp_init_congestion_control(sk);
5454 /* Prevent spurious tcp_cwnd_restart() on first data
5457 tp->lsndtime = tcp_time_stamp;
5459 tcp_init_buffer_space(sk);
5461 if (sock_flag(sk, SOCK_KEEPOPEN))
5462 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5464 if (!tp->rx_opt.snd_wscale)
5465 __tcp_fast_path_on(tp, tp->snd_wnd);
5469 if (!sock_flag(sk, SOCK_DEAD)) {
5470 sk->sk_state_change(sk);
5471 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5475 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5476 struct tcp_fastopen_cookie *cookie)
5478 struct tcp_sock *tp = tcp_sk(sk);
5479 struct sk_buff *data = tp->syn_data ? tcp_write_queue_head(sk) : NULL;
5480 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
5481 bool syn_drop = false;
5483 if (mss == tp->rx_opt.user_mss) {
5484 struct tcp_options_received opt;
5486 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5487 tcp_clear_options(&opt);
5488 opt.user_mss = opt.mss_clamp = 0;
5489 tcp_parse_options(synack, &opt, 0, NULL);
5490 mss = opt.mss_clamp;
5493 if (!tp->syn_fastopen) {
5494 /* Ignore an unsolicited cookie */
5496 } else if (tp->total_retrans) {
5497 /* SYN timed out and the SYN-ACK neither has a cookie nor
5498 * acknowledges data. Presumably the remote received only
5499 * the retransmitted (regular) SYNs: either the original
5500 * SYN-data or the corresponding SYN-ACK was dropped.
5502 syn_drop = (cookie->len < 0 && data);
5503 } else if (cookie->len < 0 && !tp->syn_data) {
5504 /* We requested a cookie but didn't get it. If we did not use
5505 * the (old) exp opt format then try so next time (try_exp=1).
5506 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5508 try_exp = tp->syn_fastopen_exp ? 2 : 1;
5511 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
5513 if (data) { /* Retransmit unacked data in SYN */
5514 tcp_for_write_queue_from(data, sk) {
5515 if (data == tcp_send_head(sk) ||
5516 __tcp_retransmit_skb(sk, data))
5520 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVEFAIL);
5523 tp->syn_data_acked = tp->syn_data;
5524 if (tp->syn_data_acked)
5525 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
5529 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5530 const struct tcphdr *th)
5532 struct inet_connection_sock *icsk = inet_csk(sk);
5533 struct tcp_sock *tp = tcp_sk(sk);
5534 struct tcp_fastopen_cookie foc = { .len = -1 };
5535 int saved_clamp = tp->rx_opt.mss_clamp;
5537 tcp_parse_options(skb, &tp->rx_opt, 0, &foc);
5538 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
5539 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
5543 * "If the state is SYN-SENT then
5544 * first check the ACK bit
5545 * If the ACK bit is set
5546 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5547 * a reset (unless the RST bit is set, if so drop
5548 * the segment and return)"
5550 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5551 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
5552 goto reset_and_undo;
5554 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5555 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5557 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
5558 goto reset_and_undo;
5561 /* Now ACK is acceptable.
5563 * "If the RST bit is set
5564 * If the ACK was acceptable then signal the user "error:
5565 * connection reset", drop the segment, enter CLOSED state,
5566 * delete TCB, and return."
5575 * "fifth, if neither of the SYN or RST bits is set then
5576 * drop the segment and return."
5582 goto discard_and_undo;
5585 * "If the SYN bit is on ...
5586 * are acceptable then ...
5587 * (our SYN has been ACKed), change the connection
5588 * state to ESTABLISHED..."
5591 tcp_ecn_rcv_synack(tp, th);
5593 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5594 tcp_ack(sk, skb, FLAG_SLOWPATH);
5596 /* Ok.. it's good. Set up sequence numbers and
5597 * move to established.
5599 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5600 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5602 /* RFC1323: The window in SYN & SYN/ACK segments is
5605 tp->snd_wnd = ntohs(th->window);
5607 if (!tp->rx_opt.wscale_ok) {
5608 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5609 tp->window_clamp = min(tp->window_clamp, 65535U);
5612 if (tp->rx_opt.saw_tstamp) {
5613 tp->rx_opt.tstamp_ok = 1;
5614 tp->tcp_header_len =
5615 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5616 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5617 tcp_store_ts_recent(tp);
5619 tp->tcp_header_len = sizeof(struct tcphdr);
5622 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5623 tcp_enable_fack(tp);
5626 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5627 tcp_initialize_rcv_mss(sk);
5629 /* Remember, tcp_poll() does not lock socket!
5630 * Change state from SYN-SENT only after copied_seq
5631 * is initialized. */
5632 tp->copied_seq = tp->rcv_nxt;
5636 tcp_finish_connect(sk, skb);
5638 if ((tp->syn_fastopen || tp->syn_data) &&
5639 tcp_rcv_fastopen_synack(sk, skb, &foc))
5642 if (sk->sk_write_pending ||
5643 icsk->icsk_accept_queue.rskq_defer_accept ||
5644 icsk->icsk_ack.pingpong) {
5645 /* Save one ACK. Data will be ready after
5646 * several ticks, if write_pending is set.
5648 * It may be deleted, but with this feature tcpdumps
5649 * look so _wonderfully_ clever, that I was not able
5650 * to stand against the temptation 8) --ANK
5652 inet_csk_schedule_ack(sk);
5653 tcp_enter_quickack_mode(sk);
5654 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5655 TCP_DELACK_MAX, TCP_RTO_MAX);
5666 /* No ACK in the segment */
5670 * "If the RST bit is set
5672 * Otherwise (no ACK) drop the segment and return."
5675 goto discard_and_undo;
5679 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5680 tcp_paws_reject(&tp->rx_opt, 0))
5681 goto discard_and_undo;
5684 /* We see SYN without ACK. It is attempt of
5685 * simultaneous connect with crossed SYNs.
5686 * Particularly, it can be connect to self.
5688 tcp_set_state(sk, TCP_SYN_RECV);
5690 if (tp->rx_opt.saw_tstamp) {
5691 tp->rx_opt.tstamp_ok = 1;
5692 tcp_store_ts_recent(tp);
5693 tp->tcp_header_len =
5694 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5696 tp->tcp_header_len = sizeof(struct tcphdr);
5699 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5700 tp->copied_seq = tp->rcv_nxt;
5701 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5703 /* RFC1323: The window in SYN & SYN/ACK segments is
5706 tp->snd_wnd = ntohs(th->window);
5707 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5708 tp->max_window = tp->snd_wnd;
5710 tcp_ecn_rcv_syn(tp, th);
5713 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5714 tcp_initialize_rcv_mss(sk);
5716 tcp_send_synack(sk);
5718 /* Note, we could accept data and URG from this segment.
5719 * There are no obstacles to make this (except that we must
5720 * either change tcp_recvmsg() to prevent it from returning data
5721 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5723 * However, if we ignore data in ACKless segments sometimes,
5724 * we have no reasons to accept it sometimes.
5725 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5726 * is not flawless. So, discard packet for sanity.
5727 * Uncomment this return to process the data.
5734 /* "fifth, if neither of the SYN or RST bits is set then
5735 * drop the segment and return."
5739 tcp_clear_options(&tp->rx_opt);
5740 tp->rx_opt.mss_clamp = saved_clamp;
5744 tcp_clear_options(&tp->rx_opt);
5745 tp->rx_opt.mss_clamp = saved_clamp;
5750 * This function implements the receiving procedure of RFC 793 for
5751 * all states except ESTABLISHED and TIME_WAIT.
5752 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5753 * address independent.
5756 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
5758 struct tcp_sock *tp = tcp_sk(sk);
5759 struct inet_connection_sock *icsk = inet_csk(sk);
5760 const struct tcphdr *th = tcp_hdr(skb);
5761 struct request_sock *req;
5765 tp->rx_opt.saw_tstamp = 0;
5767 switch (sk->sk_state) {
5781 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5784 /* Now we have several options: In theory there is
5785 * nothing else in the frame. KA9Q has an option to
5786 * send data with the syn, BSD accepts data with the
5787 * syn up to the [to be] advertised window and
5788 * Solaris 2.1 gives you a protocol error. For now
5789 * we just ignore it, that fits the spec precisely
5790 * and avoids incompatibilities. It would be nice in
5791 * future to drop through and process the data.
5793 * Now that TTCP is starting to be used we ought to
5795 * But, this leaves one open to an easy denial of
5796 * service attack, and SYN cookies can't defend
5797 * against this problem. So, we drop the data
5798 * in the interest of security over speed unless
5799 * it's still in use.
5807 queued = tcp_rcv_synsent_state_process(sk, skb, th);
5811 /* Do step6 onward by hand. */
5812 tcp_urg(sk, skb, th);
5814 tcp_data_snd_check(sk);
5818 req = tp->fastopen_rsk;
5820 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
5821 sk->sk_state != TCP_FIN_WAIT1);
5823 if (!tcp_check_req(sk, skb, req, true))
5827 if (!th->ack && !th->rst && !th->syn)
5830 if (!tcp_validate_incoming(sk, skb, th, 0))
5833 /* step 5: check the ACK field */
5834 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
5835 FLAG_UPDATE_TS_RECENT) > 0;
5837 switch (sk->sk_state) {
5843 tcp_synack_rtt_meas(sk, req);
5845 /* Once we leave TCP_SYN_RECV, we no longer need req
5849 tp->total_retrans = req->num_retrans;
5850 reqsk_fastopen_remove(sk, req, false);
5852 /* Make sure socket is routed, for correct metrics. */
5853 icsk->icsk_af_ops->rebuild_header(sk);
5854 tcp_init_congestion_control(sk);
5857 tp->copied_seq = tp->rcv_nxt;
5858 tcp_init_buffer_space(sk);
5861 tcp_set_state(sk, TCP_ESTABLISHED);
5862 sk->sk_state_change(sk);
5864 /* Note, that this wakeup is only for marginal crossed SYN case.
5865 * Passively open sockets are not waked up, because
5866 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5869 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5871 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5872 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
5873 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5875 if (tp->rx_opt.tstamp_ok)
5876 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5879 /* Re-arm the timer because data may have been sent out.
5880 * This is similar to the regular data transmission case
5881 * when new data has just been ack'ed.
5883 * (TFO) - we could try to be more aggressive and
5884 * retransmitting any data sooner based on when they
5889 tcp_init_metrics(sk);
5891 tcp_update_pacing_rate(sk);
5893 /* Prevent spurious tcp_cwnd_restart() on first data packet */
5894 tp->lsndtime = tcp_time_stamp;
5896 tcp_initialize_rcv_mss(sk);
5897 tcp_fast_path_on(tp);
5900 case TCP_FIN_WAIT1: {
5901 struct dst_entry *dst;
5904 /* If we enter the TCP_FIN_WAIT1 state and we are a
5905 * Fast Open socket and this is the first acceptable
5906 * ACK we have received, this would have acknowledged
5907 * our SYNACK so stop the SYNACK timer.
5910 /* Return RST if ack_seq is invalid.
5911 * Note that RFC793 only says to generate a
5912 * DUPACK for it but for TCP Fast Open it seems
5913 * better to treat this case like TCP_SYN_RECV
5918 /* We no longer need the request sock. */
5919 reqsk_fastopen_remove(sk, req, false);
5922 if (tp->snd_una != tp->write_seq)
5925 tcp_set_state(sk, TCP_FIN_WAIT2);
5926 sk->sk_shutdown |= SEND_SHUTDOWN;
5928 dst = __sk_dst_get(sk);
5932 if (!sock_flag(sk, SOCK_DEAD)) {
5933 /* Wake up lingering close() */
5934 sk->sk_state_change(sk);
5938 if (tp->linger2 < 0 ||
5939 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5940 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5942 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5946 tmo = tcp_fin_time(sk);
5947 if (tmo > TCP_TIMEWAIT_LEN) {
5948 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5949 } else if (th->fin || sock_owned_by_user(sk)) {
5950 /* Bad case. We could lose such FIN otherwise.
5951 * It is not a big problem, but it looks confusing
5952 * and not so rare event. We still can lose it now,
5953 * if it spins in bh_lock_sock(), but it is really
5956 inet_csk_reset_keepalive_timer(sk, tmo);
5958 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5965 if (tp->snd_una == tp->write_seq) {
5966 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5972 if (tp->snd_una == tp->write_seq) {
5973 tcp_update_metrics(sk);
5980 /* step 6: check the URG bit */
5981 tcp_urg(sk, skb, th);
5983 /* step 7: process the segment text */
5984 switch (sk->sk_state) {
5985 case TCP_CLOSE_WAIT:
5988 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5992 /* RFC 793 says to queue data in these states,
5993 * RFC 1122 says we MUST send a reset.
5994 * BSD 4.4 also does reset.
5996 if (sk->sk_shutdown & RCV_SHUTDOWN) {
5997 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5998 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5999 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6005 case TCP_ESTABLISHED:
6006 tcp_data_queue(sk, skb);
6011 /* tcp_data could move socket to TIME-WAIT */
6012 if (sk->sk_state != TCP_CLOSE) {
6013 tcp_data_snd_check(sk);
6014 tcp_ack_snd_check(sk);
6023 EXPORT_SYMBOL(tcp_rcv_state_process);
6025 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
6027 struct inet_request_sock *ireq = inet_rsk(req);
6029 if (family == AF_INET)
6030 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6031 &ireq->ir_rmt_addr, port);
6032 #if IS_ENABLED(CONFIG_IPV6)
6033 else if (family == AF_INET6)
6034 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6035 &ireq->ir_v6_rmt_addr, port);
6039 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6041 * If we receive a SYN packet with these bits set, it means a
6042 * network is playing bad games with TOS bits. In order to
6043 * avoid possible false congestion notifications, we disable
6044 * TCP ECN negotiation.
6046 * Exception: tcp_ca wants ECN. This is required for DCTCP
6047 * congestion control: Linux DCTCP asserts ECT on all packets,
6048 * including SYN, which is most optimal solution; however,
6049 * others, such as FreeBSD do not.
6051 static void tcp_ecn_create_request(struct request_sock *req,
6052 const struct sk_buff *skb,
6053 const struct sock *listen_sk,
6054 const struct dst_entry *dst)
6056 const struct tcphdr *th = tcp_hdr(skb);
6057 const struct net *net = sock_net(listen_sk);
6058 bool th_ecn = th->ece && th->cwr;
6065 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
6066 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK);
6067 ecn_ok = net->ipv4.sysctl_tcp_ecn || ecn_ok_dst;
6069 if ((!ect && ecn_ok) || tcp_ca_needs_ecn(listen_sk) ||
6070 (ecn_ok_dst & DST_FEATURE_ECN_CA))
6071 inet_rsk(req)->ecn_ok = 1;
6074 static void tcp_openreq_init(struct request_sock *req,
6075 const struct tcp_options_received *rx_opt,
6076 struct sk_buff *skb, const struct sock *sk)
6078 struct inet_request_sock *ireq = inet_rsk(req);
6080 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */
6082 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6083 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
6084 skb_mstamp_get(&tcp_rsk(req)->snt_synack);
6085 tcp_rsk(req)->last_oow_ack_time = 0;
6086 req->mss = rx_opt->mss_clamp;
6087 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6088 ireq->tstamp_ok = rx_opt->tstamp_ok;
6089 ireq->sack_ok = rx_opt->sack_ok;
6090 ireq->snd_wscale = rx_opt->snd_wscale;
6091 ireq->wscale_ok = rx_opt->wscale_ok;
6094 ireq->ir_rmt_port = tcp_hdr(skb)->source;
6095 ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6096 ireq->ir_mark = inet_request_mark(sk, skb);
6099 struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
6100 struct sock *sk_listener,
6101 bool attach_listener)
6103 struct request_sock *req = reqsk_alloc(ops, sk_listener,
6107 struct inet_request_sock *ireq = inet_rsk(req);
6109 kmemcheck_annotate_bitfield(ireq, flags);
6111 atomic64_set(&ireq->ir_cookie, 0);
6112 ireq->ireq_state = TCP_NEW_SYN_RECV;
6113 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
6114 ireq->ireq_family = sk_listener->sk_family;
6119 EXPORT_SYMBOL(inet_reqsk_alloc);
6122 * Return true if a syncookie should be sent
6124 static bool tcp_syn_flood_action(const struct sock *sk,
6125 const struct sk_buff *skb,
6128 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
6129 const char *msg = "Dropping request";
6130 bool want_cookie = false;
6132 #ifdef CONFIG_SYN_COOKIES
6133 if (sysctl_tcp_syncookies) {
6134 msg = "Sending cookies";
6136 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
6139 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
6141 if (!queue->synflood_warned &&
6142 sysctl_tcp_syncookies != 2 &&
6143 xchg(&queue->synflood_warned, 1) == 0)
6144 pr_info("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6145 proto, ntohs(tcp_hdr(skb)->dest), msg);
6150 static void tcp_reqsk_record_syn(const struct sock *sk,
6151 struct request_sock *req,
6152 const struct sk_buff *skb)
6154 if (tcp_sk(sk)->save_syn) {
6155 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
6158 copy = kmalloc(len + sizeof(u32), GFP_ATOMIC);
6161 memcpy(©[1], skb_network_header(skb), len);
6162 req->saved_syn = copy;
6167 int tcp_conn_request(struct request_sock_ops *rsk_ops,
6168 const struct tcp_request_sock_ops *af_ops,
6169 struct sock *sk, struct sk_buff *skb)
6171 struct tcp_fastopen_cookie foc = { .len = -1 };
6172 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
6173 struct tcp_options_received tmp_opt;
6174 struct tcp_sock *tp = tcp_sk(sk);
6175 struct sock *fastopen_sk = NULL;
6176 struct dst_entry *dst = NULL;
6177 struct request_sock *req;
6178 bool want_cookie = false;
6181 /* TW buckets are converted to open requests without
6182 * limitations, they conserve resources and peer is
6183 * evidently real one.
6185 if ((sysctl_tcp_syncookies == 2 ||
6186 inet_csk_reqsk_queue_is_full(sk)) && !isn) {
6187 want_cookie = tcp_syn_flood_action(sk, skb, rsk_ops->slab_name);
6193 /* Accept backlog is full. If we have already queued enough
6194 * of warm entries in syn queue, drop request. It is better than
6195 * clogging syn queue with openreqs with exponentially increasing
6198 if (sk_acceptq_is_full(sk) && inet_csk_reqsk_queue_young(sk) > 1) {
6199 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6203 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie);
6207 tcp_rsk(req)->af_specific = af_ops;
6209 tcp_clear_options(&tmp_opt);
6210 tmp_opt.mss_clamp = af_ops->mss_clamp;
6211 tmp_opt.user_mss = tp->rx_opt.user_mss;
6212 tcp_parse_options(skb, &tmp_opt, 0, want_cookie ? NULL : &foc);
6214 if (want_cookie && !tmp_opt.saw_tstamp)
6215 tcp_clear_options(&tmp_opt);
6217 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
6218 tcp_openreq_init(req, &tmp_opt, skb, sk);
6220 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6221 inet_rsk(req)->ir_iif = sk->sk_bound_dev_if;
6223 af_ops->init_req(req, sk, skb);
6225 if (security_inet_conn_request(sk, skb, req))
6228 if (!want_cookie && !isn) {
6229 /* VJ's idea. We save last timestamp seen
6230 * from the destination in peer table, when entering
6231 * state TIME-WAIT, and check against it before
6232 * accepting new connection request.
6234 * If "isn" is not zero, this request hit alive
6235 * timewait bucket, so that all the necessary checks
6236 * are made in the function processing timewait state.
6238 if (tcp_death_row.sysctl_tw_recycle) {
6241 dst = af_ops->route_req(sk, &fl, req, &strict);
6243 if (dst && strict &&
6244 !tcp_peer_is_proven(req, dst, true,
6245 tmp_opt.saw_tstamp)) {
6246 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSPASSIVEREJECTED);
6247 goto drop_and_release;
6250 /* Kill the following clause, if you dislike this way. */
6251 else if (!sysctl_tcp_syncookies &&
6252 (sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
6253 (sysctl_max_syn_backlog >> 2)) &&
6254 !tcp_peer_is_proven(req, dst, false,
6255 tmp_opt.saw_tstamp)) {
6256 /* Without syncookies last quarter of
6257 * backlog is filled with destinations,
6258 * proven to be alive.
6259 * It means that we continue to communicate
6260 * to destinations, already remembered
6261 * to the moment of synflood.
6263 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
6265 goto drop_and_release;
6268 isn = af_ops->init_seq(skb);
6271 dst = af_ops->route_req(sk, &fl, req, NULL);
6276 tcp_ecn_create_request(req, skb, sk, dst);
6279 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
6280 req->cookie_ts = tmp_opt.tstamp_ok;
6281 if (!tmp_opt.tstamp_ok)
6282 inet_rsk(req)->ecn_ok = 0;
6285 tcp_rsk(req)->snt_isn = isn;
6286 tcp_rsk(req)->txhash = net_tx_rndhash();
6287 tcp_openreq_init_rwin(req, sk, dst);
6289 tcp_reqsk_record_syn(sk, req, skb);
6290 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst);
6293 af_ops->send_synack(fastopen_sk, dst, &fl, req,
6295 /* Add the child socket directly into the accept queue */
6296 inet_csk_reqsk_queue_add(sk, req, fastopen_sk);
6297 sk->sk_data_ready(sk);
6298 bh_unlock_sock(fastopen_sk);
6299 sock_put(fastopen_sk);
6301 tcp_rsk(req)->tfo_listener = false;
6303 inet_csk_reqsk_queue_hash_add(sk, req, TCP_TIMEOUT_INIT);
6304 af_ops->send_synack(sk, dst, &fl, req,
6305 &foc, !want_cookie);
6317 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_LISTENDROPS);
6320 EXPORT_SYMBOL(tcp_conn_request);