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 = 100;
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;
99 int sysctl_tcp_thin_dupack __read_mostly;
101 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
102 int sysctl_tcp_early_retrans __read_mostly = 3;
103 int sysctl_tcp_invalid_ratelimit __read_mostly = HZ/2;
105 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
106 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
107 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
108 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
109 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
110 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
111 #define FLAG_ECE 0x40 /* ECE in this ACK */
112 #define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */
113 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
114 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
115 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
116 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
117 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
118 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
120 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
121 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
122 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
123 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
125 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
126 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
128 /* Adapt the MSS value used to make delayed ack decision to the
131 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
133 struct inet_connection_sock *icsk = inet_csk(sk);
134 const unsigned int lss = icsk->icsk_ack.last_seg_size;
137 icsk->icsk_ack.last_seg_size = 0;
139 /* skb->len may jitter because of SACKs, even if peer
140 * sends good full-sized frames.
142 len = skb_shinfo(skb)->gso_size ? : skb->len;
143 if (len >= icsk->icsk_ack.rcv_mss) {
144 icsk->icsk_ack.rcv_mss = len;
146 /* Otherwise, we make more careful check taking into account,
147 * that SACKs block is variable.
149 * "len" is invariant segment length, including TCP header.
151 len += skb->data - skb_transport_header(skb);
152 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
153 /* If PSH is not set, packet should be
154 * full sized, provided peer TCP is not badly broken.
155 * This observation (if it is correct 8)) allows
156 * to handle super-low mtu links fairly.
158 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
159 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
160 /* Subtract also invariant (if peer is RFC compliant),
161 * tcp header plus fixed timestamp option length.
162 * Resulting "len" is MSS free of SACK jitter.
164 len -= tcp_sk(sk)->tcp_header_len;
165 icsk->icsk_ack.last_seg_size = len;
167 icsk->icsk_ack.rcv_mss = len;
171 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
172 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
173 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
177 static void tcp_incr_quickack(struct sock *sk)
179 struct inet_connection_sock *icsk = inet_csk(sk);
180 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
184 if (quickacks > icsk->icsk_ack.quick)
185 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
188 static void tcp_enter_quickack_mode(struct sock *sk)
190 struct inet_connection_sock *icsk = inet_csk(sk);
191 tcp_incr_quickack(sk);
192 icsk->icsk_ack.pingpong = 0;
193 icsk->icsk_ack.ato = TCP_ATO_MIN;
196 /* Send ACKs quickly, if "quick" count is not exhausted
197 * and the session is not interactive.
200 static inline bool tcp_in_quickack_mode(const struct sock *sk)
202 const struct inet_connection_sock *icsk = inet_csk(sk);
204 return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
207 static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
209 if (tp->ecn_flags & TCP_ECN_OK)
210 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
213 static void tcp_ecn_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
215 if (tcp_hdr(skb)->cwr)
216 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
219 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
221 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
224 static void __tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
226 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
227 case INET_ECN_NOT_ECT:
228 /* Funny extension: if ECT is not set on a segment,
229 * and we already seen ECT on a previous segment,
230 * it is probably a retransmit.
232 if (tp->ecn_flags & TCP_ECN_SEEN)
233 tcp_enter_quickack_mode((struct sock *)tp);
236 if (tcp_ca_needs_ecn((struct sock *)tp))
237 tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_IS_CE);
239 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
240 /* Better not delay acks, sender can have a very low cwnd */
241 tcp_enter_quickack_mode((struct sock *)tp);
242 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
244 tp->ecn_flags |= TCP_ECN_SEEN;
247 if (tcp_ca_needs_ecn((struct sock *)tp))
248 tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_NO_CE);
249 tp->ecn_flags |= TCP_ECN_SEEN;
254 static void tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
256 if (tp->ecn_flags & TCP_ECN_OK)
257 __tcp_ecn_check_ce(tp, skb);
260 static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
262 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
263 tp->ecn_flags &= ~TCP_ECN_OK;
266 static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
268 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
269 tp->ecn_flags &= ~TCP_ECN_OK;
272 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
274 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
279 /* Buffer size and advertised window tuning.
281 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
284 static void tcp_sndbuf_expand(struct sock *sk)
286 const struct tcp_sock *tp = tcp_sk(sk);
290 /* Worst case is non GSO/TSO : each frame consumes one skb
291 * and skb->head is kmalloced using power of two area of memory
293 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
295 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
297 per_mss = roundup_pow_of_two(per_mss) +
298 SKB_DATA_ALIGN(sizeof(struct sk_buff));
300 nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd);
301 nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
303 /* Fast Recovery (RFC 5681 3.2) :
304 * Cubic needs 1.7 factor, rounded to 2 to include
305 * extra cushion (application might react slowly to POLLOUT)
307 sndmem = 2 * nr_segs * per_mss;
309 if (sk->sk_sndbuf < sndmem)
310 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
313 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
315 * All tcp_full_space() is split to two parts: "network" buffer, allocated
316 * forward and advertised in receiver window (tp->rcv_wnd) and
317 * "application buffer", required to isolate scheduling/application
318 * latencies from network.
319 * window_clamp is maximal advertised window. It can be less than
320 * tcp_full_space(), in this case tcp_full_space() - window_clamp
321 * is reserved for "application" buffer. The less window_clamp is
322 * the smoother our behaviour from viewpoint of network, but the lower
323 * throughput and the higher sensitivity of the connection to losses. 8)
325 * rcv_ssthresh is more strict window_clamp used at "slow start"
326 * phase to predict further behaviour of this connection.
327 * It is used for two goals:
328 * - to enforce header prediction at sender, even when application
329 * requires some significant "application buffer". It is check #1.
330 * - to prevent pruning of receive queue because of misprediction
331 * of receiver window. Check #2.
333 * The scheme does not work when sender sends good segments opening
334 * window and then starts to feed us spaghetti. But it should work
335 * in common situations. Otherwise, we have to rely on queue collapsing.
338 /* Slow part of check#2. */
339 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
341 struct tcp_sock *tp = tcp_sk(sk);
343 int truesize = tcp_win_from_space(skb->truesize) >> 1;
344 int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
346 while (tp->rcv_ssthresh <= window) {
347 if (truesize <= skb->len)
348 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
356 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
358 struct tcp_sock *tp = tcp_sk(sk);
361 if (tp->rcv_ssthresh < tp->window_clamp &&
362 (int)tp->rcv_ssthresh < tcp_space(sk) &&
363 !tcp_under_memory_pressure(sk)) {
366 /* Check #2. Increase window, if skb with such overhead
367 * will fit to rcvbuf in future.
369 if (tcp_win_from_space(skb->truesize) <= skb->len)
370 incr = 2 * tp->advmss;
372 incr = __tcp_grow_window(sk, skb);
375 incr = max_t(int, incr, 2 * skb->len);
376 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
378 inet_csk(sk)->icsk_ack.quick |= 1;
383 /* 3. Tuning rcvbuf, when connection enters established state. */
384 static void tcp_fixup_rcvbuf(struct sock *sk)
386 u32 mss = tcp_sk(sk)->advmss;
389 rcvmem = 2 * SKB_TRUESIZE(mss + MAX_TCP_HEADER) *
390 tcp_default_init_rwnd(mss);
392 /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
393 * Allow enough cushion so that sender is not limited by our window
395 if (sysctl_tcp_moderate_rcvbuf)
398 if (sk->sk_rcvbuf < rcvmem)
399 sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]);
402 /* 4. Try to fixup all. It is made immediately after connection enters
405 void tcp_init_buffer_space(struct sock *sk)
407 struct tcp_sock *tp = tcp_sk(sk);
410 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
411 tcp_fixup_rcvbuf(sk);
412 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
413 tcp_sndbuf_expand(sk);
415 tp->rcvq_space.space = tp->rcv_wnd;
416 tp->rcvq_space.time = tcp_time_stamp;
417 tp->rcvq_space.seq = tp->copied_seq;
419 maxwin = tcp_full_space(sk);
421 if (tp->window_clamp >= maxwin) {
422 tp->window_clamp = maxwin;
424 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
425 tp->window_clamp = max(maxwin -
426 (maxwin >> sysctl_tcp_app_win),
430 /* Force reservation of one segment. */
431 if (sysctl_tcp_app_win &&
432 tp->window_clamp > 2 * tp->advmss &&
433 tp->window_clamp + tp->advmss > maxwin)
434 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
436 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
437 tp->snd_cwnd_stamp = tcp_time_stamp;
440 /* 5. Recalculate window clamp after socket hit its memory bounds. */
441 static void tcp_clamp_window(struct sock *sk)
443 struct tcp_sock *tp = tcp_sk(sk);
444 struct inet_connection_sock *icsk = inet_csk(sk);
446 icsk->icsk_ack.quick = 0;
448 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
449 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
450 !tcp_under_memory_pressure(sk) &&
451 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
452 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
455 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
456 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
459 /* Initialize RCV_MSS value.
460 * RCV_MSS is an our guess about MSS used by the peer.
461 * We haven't any direct information about the MSS.
462 * It's better to underestimate the RCV_MSS rather than overestimate.
463 * Overestimations make us ACKing less frequently than needed.
464 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
466 void tcp_initialize_rcv_mss(struct sock *sk)
468 const struct tcp_sock *tp = tcp_sk(sk);
469 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
471 hint = min(hint, tp->rcv_wnd / 2);
472 hint = min(hint, TCP_MSS_DEFAULT);
473 hint = max(hint, TCP_MIN_MSS);
475 inet_csk(sk)->icsk_ack.rcv_mss = hint;
477 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
479 /* Receiver "autotuning" code.
481 * The algorithm for RTT estimation w/o timestamps is based on
482 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
483 * <http://public.lanl.gov/radiant/pubs.html#DRS>
485 * More detail on this code can be found at
486 * <http://staff.psc.edu/jheffner/>,
487 * though this reference is out of date. A new paper
490 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
492 u32 new_sample = tp->rcv_rtt_est.rtt;
498 if (new_sample != 0) {
499 /* If we sample in larger samples in the non-timestamp
500 * case, we could grossly overestimate the RTT especially
501 * with chatty applications or bulk transfer apps which
502 * are stalled on filesystem I/O.
504 * Also, since we are only going for a minimum in the
505 * non-timestamp case, we do not smooth things out
506 * else with timestamps disabled convergence takes too
510 m -= (new_sample >> 3);
518 /* No previous measure. */
522 if (tp->rcv_rtt_est.rtt != new_sample)
523 tp->rcv_rtt_est.rtt = new_sample;
526 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
528 if (tp->rcv_rtt_est.time == 0)
530 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
532 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rcv_rtt_est.time, 1);
535 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
536 tp->rcv_rtt_est.time = tcp_time_stamp;
539 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
540 const struct sk_buff *skb)
542 struct tcp_sock *tp = tcp_sk(sk);
543 if (tp->rx_opt.rcv_tsecr &&
544 (TCP_SKB_CB(skb)->end_seq -
545 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
546 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
550 * This function should be called every time data is copied to user space.
551 * It calculates the appropriate TCP receive buffer space.
553 void tcp_rcv_space_adjust(struct sock *sk)
555 struct tcp_sock *tp = tcp_sk(sk);
559 time = tcp_time_stamp - tp->rcvq_space.time;
560 if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
563 /* Number of bytes copied to user in last RTT */
564 copied = tp->copied_seq - tp->rcvq_space.seq;
565 if (copied <= tp->rcvq_space.space)
569 * copied = bytes received in previous RTT, our base window
570 * To cope with packet losses, we need a 2x factor
571 * To cope with slow start, and sender growing its cwin by 100 %
572 * every RTT, we need a 4x factor, because the ACK we are sending
573 * now is for the next RTT, not the current one :
574 * <prev RTT . ><current RTT .. ><next RTT .... >
577 if (sysctl_tcp_moderate_rcvbuf &&
578 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
579 int rcvwin, rcvmem, rcvbuf;
581 /* minimal window to cope with packet losses, assuming
582 * steady state. Add some cushion because of small variations.
584 rcvwin = (copied << 1) + 16 * tp->advmss;
586 /* If rate increased by 25%,
587 * assume slow start, rcvwin = 3 * copied
588 * If rate increased by 50%,
589 * assume sender can use 2x growth, rcvwin = 4 * copied
592 tp->rcvq_space.space + (tp->rcvq_space.space >> 2)) {
594 tp->rcvq_space.space + (tp->rcvq_space.space >> 1))
597 rcvwin += (rcvwin >> 1);
600 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
601 while (tcp_win_from_space(rcvmem) < tp->advmss)
604 rcvbuf = min(rcvwin / tp->advmss * rcvmem, sysctl_tcp_rmem[2]);
605 if (rcvbuf > sk->sk_rcvbuf) {
606 sk->sk_rcvbuf = rcvbuf;
608 /* Make the window clamp follow along. */
609 tp->window_clamp = rcvwin;
612 tp->rcvq_space.space = copied;
615 tp->rcvq_space.seq = tp->copied_seq;
616 tp->rcvq_space.time = tcp_time_stamp;
619 /* There is something which you must keep in mind when you analyze the
620 * behavior of the tp->ato delayed ack timeout interval. When a
621 * connection starts up, we want to ack as quickly as possible. The
622 * problem is that "good" TCP's do slow start at the beginning of data
623 * transmission. The means that until we send the first few ACK's the
624 * sender will sit on his end and only queue most of his data, because
625 * he can only send snd_cwnd unacked packets at any given time. For
626 * each ACK we send, he increments snd_cwnd and transmits more of his
629 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
631 struct tcp_sock *tp = tcp_sk(sk);
632 struct inet_connection_sock *icsk = inet_csk(sk);
635 inet_csk_schedule_ack(sk);
637 tcp_measure_rcv_mss(sk, skb);
639 tcp_rcv_rtt_measure(tp);
641 now = tcp_time_stamp;
643 if (!icsk->icsk_ack.ato) {
644 /* The _first_ data packet received, initialize
645 * delayed ACK engine.
647 tcp_incr_quickack(sk);
648 icsk->icsk_ack.ato = TCP_ATO_MIN;
650 int m = now - icsk->icsk_ack.lrcvtime;
652 if (m <= TCP_ATO_MIN / 2) {
653 /* The fastest case is the first. */
654 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
655 } else if (m < icsk->icsk_ack.ato) {
656 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
657 if (icsk->icsk_ack.ato > icsk->icsk_rto)
658 icsk->icsk_ack.ato = icsk->icsk_rto;
659 } else if (m > icsk->icsk_rto) {
660 /* Too long gap. Apparently sender failed to
661 * restart window, so that we send ACKs quickly.
663 tcp_incr_quickack(sk);
667 icsk->icsk_ack.lrcvtime = now;
669 tcp_ecn_check_ce(tp, skb);
672 tcp_grow_window(sk, skb);
675 /* Called to compute a smoothed rtt estimate. The data fed to this
676 * routine either comes from timestamps, or from segments that were
677 * known _not_ to have been retransmitted [see Karn/Partridge
678 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
679 * piece by Van Jacobson.
680 * NOTE: the next three routines used to be one big routine.
681 * To save cycles in the RFC 1323 implementation it was better to break
682 * it up into three procedures. -- erics
684 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
686 struct tcp_sock *tp = tcp_sk(sk);
687 long m = mrtt_us; /* RTT */
688 u32 srtt = tp->srtt_us;
690 /* The following amusing code comes from Jacobson's
691 * article in SIGCOMM '88. Note that rtt and mdev
692 * are scaled versions of rtt and mean deviation.
693 * This is designed to be as fast as possible
694 * m stands for "measurement".
696 * On a 1990 paper the rto value is changed to:
697 * RTO = rtt + 4 * mdev
699 * Funny. This algorithm seems to be very broken.
700 * These formulae increase RTO, when it should be decreased, increase
701 * too slowly, when it should be increased quickly, decrease too quickly
702 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
703 * does not matter how to _calculate_ it. Seems, it was trap
704 * that VJ failed to avoid. 8)
707 m -= (srtt >> 3); /* m is now error in rtt est */
708 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
710 m = -m; /* m is now abs(error) */
711 m -= (tp->mdev_us >> 2); /* similar update on mdev */
712 /* This is similar to one of Eifel findings.
713 * Eifel blocks mdev updates when rtt decreases.
714 * This solution is a bit different: we use finer gain
715 * for mdev in this case (alpha*beta).
716 * Like Eifel it also prevents growth of rto,
717 * but also it limits too fast rto decreases,
718 * happening in pure Eifel.
723 m -= (tp->mdev_us >> 2); /* similar update on mdev */
725 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
726 if (tp->mdev_us > tp->mdev_max_us) {
727 tp->mdev_max_us = tp->mdev_us;
728 if (tp->mdev_max_us > tp->rttvar_us)
729 tp->rttvar_us = tp->mdev_max_us;
731 if (after(tp->snd_una, tp->rtt_seq)) {
732 if (tp->mdev_max_us < tp->rttvar_us)
733 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
734 tp->rtt_seq = tp->snd_nxt;
735 tp->mdev_max_us = tcp_rto_min_us(sk);
738 /* no previous measure. */
739 srtt = m << 3; /* take the measured time to be rtt */
740 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
741 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
742 tp->mdev_max_us = tp->rttvar_us;
743 tp->rtt_seq = tp->snd_nxt;
745 tp->srtt_us = max(1U, srtt);
748 /* Set the sk_pacing_rate to allow proper sizing of TSO packets.
749 * Note: TCP stack does not yet implement pacing.
750 * FQ packet scheduler can be used to implement cheap but effective
751 * TCP pacing, to smooth the burst on large writes when packets
752 * in flight is significantly lower than cwnd (or rwin)
754 static void tcp_update_pacing_rate(struct sock *sk)
756 const struct tcp_sock *tp = tcp_sk(sk);
759 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
760 rate = (u64)tp->mss_cache * 2 * (USEC_PER_SEC << 3);
762 rate *= max(tp->snd_cwnd, tp->packets_out);
764 if (likely(tp->srtt_us))
765 do_div(rate, tp->srtt_us);
767 /* ACCESS_ONCE() is needed because sch_fq fetches sk_pacing_rate
768 * without any lock. We want to make sure compiler wont store
769 * intermediate values in this location.
771 ACCESS_ONCE(sk->sk_pacing_rate) = min_t(u64, rate,
772 sk->sk_max_pacing_rate);
775 /* Calculate rto without backoff. This is the second half of Van Jacobson's
776 * routine referred to above.
778 static void tcp_set_rto(struct sock *sk)
780 const struct tcp_sock *tp = tcp_sk(sk);
781 /* Old crap is replaced with new one. 8)
784 * 1. If rtt variance happened to be less 50msec, it is hallucination.
785 * It cannot be less due to utterly erratic ACK generation made
786 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
787 * to do with delayed acks, because at cwnd>2 true delack timeout
788 * is invisible. Actually, Linux-2.4 also generates erratic
789 * ACKs in some circumstances.
791 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
793 /* 2. Fixups made earlier cannot be right.
794 * If we do not estimate RTO correctly without them,
795 * all the algo is pure shit and should be replaced
796 * with correct one. It is exactly, which we pretend to do.
799 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
800 * guarantees that rto is higher.
805 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
807 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
810 cwnd = TCP_INIT_CWND;
811 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
815 * Packet counting of FACK is based on in-order assumptions, therefore TCP
816 * disables it when reordering is detected
818 void tcp_disable_fack(struct tcp_sock *tp)
820 /* RFC3517 uses different metric in lost marker => reset on change */
822 tp->lost_skb_hint = NULL;
823 tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED;
826 /* Take a notice that peer is sending D-SACKs */
827 static void tcp_dsack_seen(struct tcp_sock *tp)
829 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
832 static void tcp_update_reordering(struct sock *sk, const int metric,
835 struct tcp_sock *tp = tcp_sk(sk);
836 if (metric > tp->reordering) {
839 tp->reordering = min(sysctl_tcp_max_reordering, metric);
841 /* This exciting event is worth to be remembered. 8) */
843 mib_idx = LINUX_MIB_TCPTSREORDER;
844 else if (tcp_is_reno(tp))
845 mib_idx = LINUX_MIB_TCPRENOREORDER;
846 else if (tcp_is_fack(tp))
847 mib_idx = LINUX_MIB_TCPFACKREORDER;
849 mib_idx = LINUX_MIB_TCPSACKREORDER;
851 NET_INC_STATS_BH(sock_net(sk), mib_idx);
852 #if FASTRETRANS_DEBUG > 1
853 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
854 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
858 tp->undo_marker ? tp->undo_retrans : 0);
860 tcp_disable_fack(tp);
864 tcp_disable_early_retrans(tp);
867 /* This must be called before lost_out is incremented */
868 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
870 if (!tp->retransmit_skb_hint ||
871 before(TCP_SKB_CB(skb)->seq,
872 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
873 tp->retransmit_skb_hint = skb;
876 after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
877 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
880 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
882 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
883 tcp_verify_retransmit_hint(tp, skb);
885 tp->lost_out += tcp_skb_pcount(skb);
886 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
890 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp,
893 tcp_verify_retransmit_hint(tp, skb);
895 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
896 tp->lost_out += tcp_skb_pcount(skb);
897 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
901 /* This procedure tags the retransmission queue when SACKs arrive.
903 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
904 * Packets in queue with these bits set are counted in variables
905 * sacked_out, retrans_out and lost_out, correspondingly.
907 * Valid combinations are:
908 * Tag InFlight Description
909 * 0 1 - orig segment is in flight.
910 * S 0 - nothing flies, orig reached receiver.
911 * L 0 - nothing flies, orig lost by net.
912 * R 2 - both orig and retransmit are in flight.
913 * L|R 1 - orig is lost, retransmit is in flight.
914 * S|R 1 - orig reached receiver, retrans is still in flight.
915 * (L|S|R is logically valid, it could occur when L|R is sacked,
916 * but it is equivalent to plain S and code short-curcuits it to S.
917 * L|S is logically invalid, it would mean -1 packet in flight 8))
919 * These 6 states form finite state machine, controlled by the following events:
920 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
921 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
922 * 3. Loss detection event of two flavors:
923 * A. Scoreboard estimator decided the packet is lost.
924 * A'. Reno "three dupacks" marks head of queue lost.
925 * A''. Its FACK modification, head until snd.fack is lost.
926 * B. SACK arrives sacking SND.NXT at the moment, when the
927 * segment was retransmitted.
928 * 4. D-SACK added new rule: D-SACK changes any tag to S.
930 * It is pleasant to note, that state diagram turns out to be commutative,
931 * so that we are allowed not to be bothered by order of our actions,
932 * when multiple events arrive simultaneously. (see the function below).
934 * Reordering detection.
935 * --------------------
936 * Reordering metric is maximal distance, which a packet can be displaced
937 * in packet stream. With SACKs we can estimate it:
939 * 1. SACK fills old hole and the corresponding segment was not
940 * ever retransmitted -> reordering. Alas, we cannot use it
941 * when segment was retransmitted.
942 * 2. The last flaw is solved with D-SACK. D-SACK arrives
943 * for retransmitted and already SACKed segment -> reordering..
944 * Both of these heuristics are not used in Loss state, when we cannot
945 * account for retransmits accurately.
947 * SACK block validation.
948 * ----------------------
950 * SACK block range validation checks that the received SACK block fits to
951 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
952 * Note that SND.UNA is not included to the range though being valid because
953 * it means that the receiver is rather inconsistent with itself reporting
954 * SACK reneging when it should advance SND.UNA. Such SACK block this is
955 * perfectly valid, however, in light of RFC2018 which explicitly states
956 * that "SACK block MUST reflect the newest segment. Even if the newest
957 * segment is going to be discarded ...", not that it looks very clever
958 * in case of head skb. Due to potentional receiver driven attacks, we
959 * choose to avoid immediate execution of a walk in write queue due to
960 * reneging and defer head skb's loss recovery to standard loss recovery
961 * procedure that will eventually trigger (nothing forbids us doing this).
963 * Implements also blockage to start_seq wrap-around. Problem lies in the
964 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
965 * there's no guarantee that it will be before snd_nxt (n). The problem
966 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
969 * <- outs wnd -> <- wrapzone ->
970 * u e n u_w e_w s n_w
972 * |<------------+------+----- TCP seqno space --------------+---------->|
973 * ...-- <2^31 ->| |<--------...
974 * ...---- >2^31 ------>| |<--------...
976 * Current code wouldn't be vulnerable but it's better still to discard such
977 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
978 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
979 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
980 * equal to the ideal case (infinite seqno space without wrap caused issues).
982 * With D-SACK the lower bound is extended to cover sequence space below
983 * SND.UNA down to undo_marker, which is the last point of interest. Yet
984 * again, D-SACK block must not to go across snd_una (for the same reason as
985 * for the normal SACK blocks, explained above). But there all simplicity
986 * ends, TCP might receive valid D-SACKs below that. As long as they reside
987 * fully below undo_marker they do not affect behavior in anyway and can
988 * therefore be safely ignored. In rare cases (which are more or less
989 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
990 * fragmentation and packet reordering past skb's retransmission. To consider
991 * them correctly, the acceptable range must be extended even more though
992 * the exact amount is rather hard to quantify. However, tp->max_window can
993 * be used as an exaggerated estimate.
995 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
996 u32 start_seq, u32 end_seq)
998 /* Too far in future, or reversed (interpretation is ambiguous) */
999 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1002 /* Nasty start_seq wrap-around check (see comments above) */
1003 if (!before(start_seq, tp->snd_nxt))
1006 /* In outstanding window? ...This is valid exit for D-SACKs too.
1007 * start_seq == snd_una is non-sensical (see comments above)
1009 if (after(start_seq, tp->snd_una))
1012 if (!is_dsack || !tp->undo_marker)
1015 /* ...Then it's D-SACK, and must reside below snd_una completely */
1016 if (after(end_seq, tp->snd_una))
1019 if (!before(start_seq, tp->undo_marker))
1023 if (!after(end_seq, tp->undo_marker))
1026 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1027 * start_seq < undo_marker and end_seq >= undo_marker.
1029 return !before(start_seq, end_seq - tp->max_window);
1032 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1033 * Event "B". Later note: FACK people cheated me again 8), we have to account
1034 * for reordering! Ugly, but should help.
1036 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1037 * less than what is now known to be received by the other end (derived from
1038 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1039 * retransmitted skbs to avoid some costly processing per ACKs.
1041 static void tcp_mark_lost_retrans(struct sock *sk, int *flag)
1043 const struct inet_connection_sock *icsk = inet_csk(sk);
1044 struct tcp_sock *tp = tcp_sk(sk);
1045 struct sk_buff *skb;
1047 u32 new_low_seq = tp->snd_nxt;
1048 u32 received_upto = tcp_highest_sack_seq(tp);
1050 if (!tcp_is_fack(tp) || !tp->retrans_out ||
1051 !after(received_upto, tp->lost_retrans_low) ||
1052 icsk->icsk_ca_state != TCP_CA_Recovery)
1055 tcp_for_write_queue(skb, sk) {
1056 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1058 if (skb == tcp_send_head(sk))
1060 if (cnt == tp->retrans_out)
1062 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1065 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1068 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1069 * constraint here (see above) but figuring out that at
1070 * least tp->reordering SACK blocks reside between ack_seq
1071 * and received_upto is not easy task to do cheaply with
1072 * the available datastructures.
1074 * Whether FACK should check here for tp->reordering segs
1075 * in-between one could argue for either way (it would be
1076 * rather simple to implement as we could count fack_count
1077 * during the walk and do tp->fackets_out - fack_count).
1079 if (after(received_upto, ack_seq)) {
1080 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1081 tp->retrans_out -= tcp_skb_pcount(skb);
1082 *flag |= FLAG_LOST_RETRANS;
1083 tcp_skb_mark_lost_uncond_verify(tp, skb);
1084 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
1086 if (before(ack_seq, new_low_seq))
1087 new_low_seq = ack_seq;
1088 cnt += tcp_skb_pcount(skb);
1092 if (tp->retrans_out)
1093 tp->lost_retrans_low = new_low_seq;
1096 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1097 struct tcp_sack_block_wire *sp, int num_sacks,
1100 struct tcp_sock *tp = tcp_sk(sk);
1101 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1102 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1103 bool dup_sack = false;
1105 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1108 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1109 } else if (num_sacks > 1) {
1110 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1111 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1113 if (!after(end_seq_0, end_seq_1) &&
1114 !before(start_seq_0, start_seq_1)) {
1117 NET_INC_STATS_BH(sock_net(sk),
1118 LINUX_MIB_TCPDSACKOFORECV);
1122 /* D-SACK for already forgotten data... Do dumb counting. */
1123 if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 &&
1124 !after(end_seq_0, prior_snd_una) &&
1125 after(end_seq_0, tp->undo_marker))
1131 struct tcp_sacktag_state {
1134 /* Timestamps for earliest and latest never-retransmitted segment
1135 * that was SACKed. RTO needs the earliest RTT to stay conservative,
1136 * but congestion control should still get an accurate delay signal.
1138 struct skb_mstamp first_sackt;
1139 struct skb_mstamp last_sackt;
1143 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1144 * the incoming SACK may not exactly match but we can find smaller MSS
1145 * aligned portion of it that matches. Therefore we might need to fragment
1146 * which may fail and creates some hassle (caller must handle error case
1149 * FIXME: this could be merged to shift decision code
1151 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1152 u32 start_seq, u32 end_seq)
1156 unsigned int pkt_len;
1159 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1160 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1162 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1163 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1164 mss = tcp_skb_mss(skb);
1165 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1168 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1172 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1177 /* Round if necessary so that SACKs cover only full MSSes
1178 * and/or the remaining small portion (if present)
1180 if (pkt_len > mss) {
1181 unsigned int new_len = (pkt_len / mss) * mss;
1182 if (!in_sack && new_len < pkt_len) {
1184 if (new_len >= skb->len)
1189 err = tcp_fragment(sk, skb, pkt_len, mss, GFP_ATOMIC);
1197 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1198 static u8 tcp_sacktag_one(struct sock *sk,
1199 struct tcp_sacktag_state *state, u8 sacked,
1200 u32 start_seq, u32 end_seq,
1201 int dup_sack, int pcount,
1202 const struct skb_mstamp *xmit_time)
1204 struct tcp_sock *tp = tcp_sk(sk);
1205 int fack_count = state->fack_count;
1207 /* Account D-SACK for retransmitted packet. */
1208 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1209 if (tp->undo_marker && tp->undo_retrans > 0 &&
1210 after(end_seq, tp->undo_marker))
1212 if (sacked & TCPCB_SACKED_ACKED)
1213 state->reord = min(fack_count, state->reord);
1216 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1217 if (!after(end_seq, tp->snd_una))
1220 if (!(sacked & TCPCB_SACKED_ACKED)) {
1221 if (sacked & TCPCB_SACKED_RETRANS) {
1222 /* If the segment is not tagged as lost,
1223 * we do not clear RETRANS, believing
1224 * that retransmission is still in flight.
1226 if (sacked & TCPCB_LOST) {
1227 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1228 tp->lost_out -= pcount;
1229 tp->retrans_out -= pcount;
1232 if (!(sacked & TCPCB_RETRANS)) {
1233 /* New sack for not retransmitted frame,
1234 * which was in hole. It is reordering.
1236 if (before(start_seq,
1237 tcp_highest_sack_seq(tp)))
1238 state->reord = min(fack_count,
1240 if (!after(end_seq, tp->high_seq))
1241 state->flag |= FLAG_ORIG_SACK_ACKED;
1242 if (state->first_sackt.v64 == 0)
1243 state->first_sackt = *xmit_time;
1244 state->last_sackt = *xmit_time;
1247 if (sacked & TCPCB_LOST) {
1248 sacked &= ~TCPCB_LOST;
1249 tp->lost_out -= pcount;
1253 sacked |= TCPCB_SACKED_ACKED;
1254 state->flag |= FLAG_DATA_SACKED;
1255 tp->sacked_out += pcount;
1257 fack_count += pcount;
1259 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1260 if (!tcp_is_fack(tp) && tp->lost_skb_hint &&
1261 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1262 tp->lost_cnt_hint += pcount;
1264 if (fack_count > tp->fackets_out)
1265 tp->fackets_out = fack_count;
1268 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1269 * frames and clear it. undo_retrans is decreased above, L|R frames
1270 * are accounted above as well.
1272 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1273 sacked &= ~TCPCB_SACKED_RETRANS;
1274 tp->retrans_out -= pcount;
1280 /* Shift newly-SACKed bytes from this skb to the immediately previous
1281 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1283 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1284 struct tcp_sacktag_state *state,
1285 unsigned int pcount, int shifted, int mss,
1288 struct tcp_sock *tp = tcp_sk(sk);
1289 struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1290 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1291 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1295 /* Adjust counters and hints for the newly sacked sequence
1296 * range but discard the return value since prev is already
1297 * marked. We must tag the range first because the seq
1298 * advancement below implicitly advances
1299 * tcp_highest_sack_seq() when skb is highest_sack.
1301 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1302 start_seq, end_seq, dup_sack, pcount,
1305 if (skb == tp->lost_skb_hint)
1306 tp->lost_cnt_hint += pcount;
1308 TCP_SKB_CB(prev)->end_seq += shifted;
1309 TCP_SKB_CB(skb)->seq += shifted;
1311 tcp_skb_pcount_add(prev, pcount);
1312 BUG_ON(tcp_skb_pcount(skb) < pcount);
1313 tcp_skb_pcount_add(skb, -pcount);
1315 /* When we're adding to gso_segs == 1, gso_size will be zero,
1316 * in theory this shouldn't be necessary but as long as DSACK
1317 * code can come after this skb later on it's better to keep
1318 * setting gso_size to something.
1320 if (!TCP_SKB_CB(prev)->tcp_gso_size)
1321 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1323 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1324 if (tcp_skb_pcount(skb) <= 1)
1325 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1327 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1328 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1331 BUG_ON(!tcp_skb_pcount(skb));
1332 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1336 /* Whole SKB was eaten :-) */
1338 if (skb == tp->retransmit_skb_hint)
1339 tp->retransmit_skb_hint = prev;
1340 if (skb == tp->lost_skb_hint) {
1341 tp->lost_skb_hint = prev;
1342 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1345 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1346 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1347 TCP_SKB_CB(prev)->end_seq++;
1349 if (skb == tcp_highest_sack(sk))
1350 tcp_advance_highest_sack(sk, skb);
1352 tcp_unlink_write_queue(skb, sk);
1353 sk_wmem_free_skb(sk, skb);
1355 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1360 /* I wish gso_size would have a bit more sane initialization than
1361 * something-or-zero which complicates things
1363 static int tcp_skb_seglen(const struct sk_buff *skb)
1365 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1368 /* Shifting pages past head area doesn't work */
1369 static int skb_can_shift(const struct sk_buff *skb)
1371 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1374 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1377 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1378 struct tcp_sacktag_state *state,
1379 u32 start_seq, u32 end_seq,
1382 struct tcp_sock *tp = tcp_sk(sk);
1383 struct sk_buff *prev;
1389 if (!sk_can_gso(sk))
1392 /* Normally R but no L won't result in plain S */
1394 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1396 if (!skb_can_shift(skb))
1398 /* This frame is about to be dropped (was ACKed). */
1399 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1402 /* Can only happen with delayed DSACK + discard craziness */
1403 if (unlikely(skb == tcp_write_queue_head(sk)))
1405 prev = tcp_write_queue_prev(sk, skb);
1407 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1410 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1411 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1415 pcount = tcp_skb_pcount(skb);
1416 mss = tcp_skb_seglen(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))
1424 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1426 /* CHECKME: This is non-MSS split case only?, this will
1427 * cause skipped skbs due to advancing loop btw, original
1428 * has that feature too
1430 if (tcp_skb_pcount(skb) <= 1)
1433 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1435 /* TODO: head merge to next could be attempted here
1436 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1437 * though it might not be worth of the additional hassle
1439 * ...we can probably just fallback to what was done
1440 * previously. We could try merging non-SACKed ones
1441 * as well but it probably isn't going to buy off
1442 * because later SACKs might again split them, and
1443 * it would make skb timestamp tracking considerably
1449 len = end_seq - TCP_SKB_CB(skb)->seq;
1451 BUG_ON(len > skb->len);
1453 /* MSS boundaries should be honoured or else pcount will
1454 * severely break even though it makes things bit trickier.
1455 * Optimize common case to avoid most of the divides
1457 mss = tcp_skb_mss(skb);
1459 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1460 * drop this restriction as unnecessary
1462 if (mss != tcp_skb_seglen(prev))
1467 } else if (len < mss) {
1475 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1476 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1479 if (!skb_shift(prev, skb, len))
1481 if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1484 /* Hole filled allows collapsing with the next as well, this is very
1485 * useful when hole on every nth skb pattern happens
1487 if (prev == tcp_write_queue_tail(sk))
1489 skb = tcp_write_queue_next(sk, prev);
1491 if (!skb_can_shift(skb) ||
1492 (skb == tcp_send_head(sk)) ||
1493 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1494 (mss != tcp_skb_seglen(skb)))
1498 if (skb_shift(prev, skb, len)) {
1499 pcount += tcp_skb_pcount(skb);
1500 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1504 state->fack_count += pcount;
1511 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1515 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1516 struct tcp_sack_block *next_dup,
1517 struct tcp_sacktag_state *state,
1518 u32 start_seq, u32 end_seq,
1521 struct tcp_sock *tp = tcp_sk(sk);
1522 struct sk_buff *tmp;
1524 tcp_for_write_queue_from(skb, sk) {
1526 bool dup_sack = dup_sack_in;
1528 if (skb == tcp_send_head(sk))
1531 /* queue is in-order => we can short-circuit the walk early */
1532 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1536 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1537 in_sack = tcp_match_skb_to_sack(sk, skb,
1538 next_dup->start_seq,
1544 /* skb reference here is a bit tricky to get right, since
1545 * shifting can eat and free both this skb and the next,
1546 * so not even _safe variant of the loop is enough.
1549 tmp = tcp_shift_skb_data(sk, skb, state,
1550 start_seq, end_seq, dup_sack);
1559 in_sack = tcp_match_skb_to_sack(sk, skb,
1565 if (unlikely(in_sack < 0))
1569 TCP_SKB_CB(skb)->sacked =
1572 TCP_SKB_CB(skb)->sacked,
1573 TCP_SKB_CB(skb)->seq,
1574 TCP_SKB_CB(skb)->end_seq,
1576 tcp_skb_pcount(skb),
1579 if (!before(TCP_SKB_CB(skb)->seq,
1580 tcp_highest_sack_seq(tp)))
1581 tcp_advance_highest_sack(sk, skb);
1584 state->fack_count += tcp_skb_pcount(skb);
1589 /* Avoid all extra work that is being done by sacktag while walking in
1592 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1593 struct tcp_sacktag_state *state,
1596 tcp_for_write_queue_from(skb, sk) {
1597 if (skb == tcp_send_head(sk))
1600 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1603 state->fack_count += tcp_skb_pcount(skb);
1608 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1610 struct tcp_sack_block *next_dup,
1611 struct tcp_sacktag_state *state,
1617 if (before(next_dup->start_seq, skip_to_seq)) {
1618 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1619 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1620 next_dup->start_seq, next_dup->end_seq,
1627 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1629 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1633 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1634 u32 prior_snd_una, struct tcp_sacktag_state *state)
1636 struct tcp_sock *tp = tcp_sk(sk);
1637 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1638 TCP_SKB_CB(ack_skb)->sacked);
1639 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1640 struct tcp_sack_block sp[TCP_NUM_SACKS];
1641 struct tcp_sack_block *cache;
1642 struct sk_buff *skb;
1643 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1645 bool found_dup_sack = false;
1647 int first_sack_index;
1650 state->reord = tp->packets_out;
1652 if (!tp->sacked_out) {
1653 if (WARN_ON(tp->fackets_out))
1654 tp->fackets_out = 0;
1655 tcp_highest_sack_reset(sk);
1658 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1659 num_sacks, prior_snd_una);
1661 state->flag |= FLAG_DSACKING_ACK;
1663 /* Eliminate too old ACKs, but take into
1664 * account more or less fresh ones, they can
1665 * contain valid SACK info.
1667 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1670 if (!tp->packets_out)
1674 first_sack_index = 0;
1675 for (i = 0; i < num_sacks; i++) {
1676 bool dup_sack = !i && found_dup_sack;
1678 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1679 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1681 if (!tcp_is_sackblock_valid(tp, dup_sack,
1682 sp[used_sacks].start_seq,
1683 sp[used_sacks].end_seq)) {
1687 if (!tp->undo_marker)
1688 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1690 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1692 /* Don't count olds caused by ACK reordering */
1693 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1694 !after(sp[used_sacks].end_seq, tp->snd_una))
1696 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1699 NET_INC_STATS_BH(sock_net(sk), mib_idx);
1701 first_sack_index = -1;
1705 /* Ignore very old stuff early */
1706 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1712 /* order SACK blocks to allow in order walk of the retrans queue */
1713 for (i = used_sacks - 1; i > 0; i--) {
1714 for (j = 0; j < i; j++) {
1715 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1716 swap(sp[j], sp[j + 1]);
1718 /* Track where the first SACK block goes to */
1719 if (j == first_sack_index)
1720 first_sack_index = j + 1;
1725 skb = tcp_write_queue_head(sk);
1726 state->fack_count = 0;
1729 if (!tp->sacked_out) {
1730 /* It's already past, so skip checking against it */
1731 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1733 cache = tp->recv_sack_cache;
1734 /* Skip empty blocks in at head of the cache */
1735 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1740 while (i < used_sacks) {
1741 u32 start_seq = sp[i].start_seq;
1742 u32 end_seq = sp[i].end_seq;
1743 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1744 struct tcp_sack_block *next_dup = NULL;
1746 if (found_dup_sack && ((i + 1) == first_sack_index))
1747 next_dup = &sp[i + 1];
1749 /* Skip too early cached blocks */
1750 while (tcp_sack_cache_ok(tp, cache) &&
1751 !before(start_seq, cache->end_seq))
1754 /* Can skip some work by looking recv_sack_cache? */
1755 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1756 after(end_seq, cache->start_seq)) {
1759 if (before(start_seq, cache->start_seq)) {
1760 skb = tcp_sacktag_skip(skb, sk, state,
1762 skb = tcp_sacktag_walk(skb, sk, next_dup,
1769 /* Rest of the block already fully processed? */
1770 if (!after(end_seq, cache->end_seq))
1773 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1777 /* ...tail remains todo... */
1778 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1779 /* ...but better entrypoint exists! */
1780 skb = tcp_highest_sack(sk);
1783 state->fack_count = tp->fackets_out;
1788 skb = tcp_sacktag_skip(skb, sk, state, cache->end_seq);
1789 /* Check overlap against next cached too (past this one already) */
1794 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1795 skb = tcp_highest_sack(sk);
1798 state->fack_count = tp->fackets_out;
1800 skb = tcp_sacktag_skip(skb, sk, state, start_seq);
1803 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
1804 start_seq, end_seq, dup_sack);
1810 /* Clear the head of the cache sack blocks so we can skip it next time */
1811 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1812 tp->recv_sack_cache[i].start_seq = 0;
1813 tp->recv_sack_cache[i].end_seq = 0;
1815 for (j = 0; j < used_sacks; j++)
1816 tp->recv_sack_cache[i++] = sp[j];
1818 if ((state->reord < tp->fackets_out) &&
1819 ((inet_csk(sk)->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker))
1820 tcp_update_reordering(sk, tp->fackets_out - state->reord, 0);
1822 tcp_mark_lost_retrans(sk, &state->flag);
1823 tcp_verify_left_out(tp);
1826 #if FASTRETRANS_DEBUG > 0
1827 WARN_ON((int)tp->sacked_out < 0);
1828 WARN_ON((int)tp->lost_out < 0);
1829 WARN_ON((int)tp->retrans_out < 0);
1830 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1835 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1836 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1838 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1842 holes = max(tp->lost_out, 1U);
1843 holes = min(holes, tp->packets_out);
1845 if ((tp->sacked_out + holes) > tp->packets_out) {
1846 tp->sacked_out = tp->packets_out - holes;
1852 /* If we receive more dupacks than we expected counting segments
1853 * in assumption of absent reordering, interpret this as reordering.
1854 * The only another reason could be bug in receiver TCP.
1856 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1858 struct tcp_sock *tp = tcp_sk(sk);
1859 if (tcp_limit_reno_sacked(tp))
1860 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1863 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1865 static void tcp_add_reno_sack(struct sock *sk)
1867 struct tcp_sock *tp = tcp_sk(sk);
1869 tcp_check_reno_reordering(sk, 0);
1870 tcp_verify_left_out(tp);
1873 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1875 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1877 struct tcp_sock *tp = tcp_sk(sk);
1880 /* One ACK acked hole. The rest eat duplicate ACKs. */
1881 if (acked - 1 >= tp->sacked_out)
1884 tp->sacked_out -= acked - 1;
1886 tcp_check_reno_reordering(sk, acked);
1887 tcp_verify_left_out(tp);
1890 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1895 void tcp_clear_retrans(struct tcp_sock *tp)
1897 tp->retrans_out = 0;
1899 tp->undo_marker = 0;
1900 tp->undo_retrans = -1;
1901 tp->fackets_out = 0;
1905 static inline void tcp_init_undo(struct tcp_sock *tp)
1907 tp->undo_marker = tp->snd_una;
1908 /* Retransmission still in flight may cause DSACKs later. */
1909 tp->undo_retrans = tp->retrans_out ? : -1;
1912 /* Enter Loss state. If we detect SACK reneging, forget all SACK information
1913 * and reset tags completely, otherwise preserve SACKs. If receiver
1914 * dropped its ofo queue, we will know this due to reneging detection.
1916 void tcp_enter_loss(struct sock *sk)
1918 const struct inet_connection_sock *icsk = inet_csk(sk);
1919 struct tcp_sock *tp = tcp_sk(sk);
1920 struct sk_buff *skb;
1921 bool new_recovery = false;
1922 bool is_reneg; /* is receiver reneging on SACKs? */
1924 /* Reduce ssthresh if it has not yet been made inside this window. */
1925 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1926 !after(tp->high_seq, tp->snd_una) ||
1927 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1928 new_recovery = true;
1929 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1930 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1931 tcp_ca_event(sk, CA_EVENT_LOSS);
1935 tp->snd_cwnd_cnt = 0;
1936 tp->snd_cwnd_stamp = tcp_time_stamp;
1938 tp->retrans_out = 0;
1941 if (tcp_is_reno(tp))
1942 tcp_reset_reno_sack(tp);
1944 skb = tcp_write_queue_head(sk);
1945 is_reneg = skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED);
1947 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
1949 tp->fackets_out = 0;
1951 tcp_clear_all_retrans_hints(tp);
1953 tcp_for_write_queue(skb, sk) {
1954 if (skb == tcp_send_head(sk))
1957 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1958 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || is_reneg) {
1959 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1960 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1961 tp->lost_out += tcp_skb_pcount(skb);
1962 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
1965 tcp_verify_left_out(tp);
1967 /* Timeout in disordered state after receiving substantial DUPACKs
1968 * suggests that the degree of reordering is over-estimated.
1970 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
1971 tp->sacked_out >= sysctl_tcp_reordering)
1972 tp->reordering = min_t(unsigned int, tp->reordering,
1973 sysctl_tcp_reordering);
1974 tcp_set_ca_state(sk, TCP_CA_Loss);
1975 tp->high_seq = tp->snd_nxt;
1976 tcp_ecn_queue_cwr(tp);
1978 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1979 * loss recovery is underway except recurring timeout(s) on
1980 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1982 tp->frto = sysctl_tcp_frto &&
1983 (new_recovery || icsk->icsk_retransmits) &&
1984 !inet_csk(sk)->icsk_mtup.probe_size;
1987 /* If ACK arrived pointing to a remembered SACK, it means that our
1988 * remembered SACKs do not reflect real state of receiver i.e.
1989 * receiver _host_ is heavily congested (or buggy).
1991 * To avoid big spurious retransmission bursts due to transient SACK
1992 * scoreboard oddities that look like reneging, we give the receiver a
1993 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
1994 * restore sanity to the SACK scoreboard. If the apparent reneging
1995 * persists until this RTO then we'll clear the SACK scoreboard.
1997 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
1999 if (flag & FLAG_SACK_RENEGING) {
2000 struct tcp_sock *tp = tcp_sk(sk);
2001 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2002 msecs_to_jiffies(10));
2004 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2005 delay, TCP_RTO_MAX);
2011 static inline int tcp_fackets_out(const struct tcp_sock *tp)
2013 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
2016 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2017 * counter when SACK is enabled (without SACK, sacked_out is used for
2020 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2021 * segments up to the highest received SACK block so far and holes in
2024 * With reordering, holes may still be in flight, so RFC3517 recovery
2025 * uses pure sacked_out (total number of SACKed segments) even though
2026 * it violates the RFC that uses duplicate ACKs, often these are equal
2027 * but when e.g. out-of-window ACKs or packet duplication occurs,
2028 * they differ. Since neither occurs due to loss, TCP should really
2031 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2033 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2036 static bool tcp_pause_early_retransmit(struct sock *sk, int flag)
2038 struct tcp_sock *tp = tcp_sk(sk);
2039 unsigned long delay;
2041 /* Delay early retransmit and entering fast recovery for
2042 * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
2043 * available, or RTO is scheduled to fire first.
2045 if (sysctl_tcp_early_retrans < 2 || sysctl_tcp_early_retrans > 3 ||
2046 (flag & FLAG_ECE) || !tp->srtt_us)
2049 delay = max(usecs_to_jiffies(tp->srtt_us >> 5),
2050 msecs_to_jiffies(2));
2052 if (!time_after(inet_csk(sk)->icsk_timeout, (jiffies + delay)))
2055 inet_csk_reset_xmit_timer(sk, ICSK_TIME_EARLY_RETRANS, delay,
2060 /* Linux NewReno/SACK/FACK/ECN state machine.
2061 * --------------------------------------
2063 * "Open" Normal state, no dubious events, fast path.
2064 * "Disorder" In all the respects it is "Open",
2065 * but requires a bit more attention. It is entered when
2066 * we see some SACKs or dupacks. It is split of "Open"
2067 * mainly to move some processing from fast path to slow one.
2068 * "CWR" CWND was reduced due to some Congestion Notification event.
2069 * It can be ECN, ICMP source quench, local device congestion.
2070 * "Recovery" CWND was reduced, we are fast-retransmitting.
2071 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2073 * tcp_fastretrans_alert() is entered:
2074 * - each incoming ACK, if state is not "Open"
2075 * - when arrived ACK is unusual, namely:
2080 * Counting packets in flight is pretty simple.
2082 * in_flight = packets_out - left_out + retrans_out
2084 * packets_out is SND.NXT-SND.UNA counted in packets.
2086 * retrans_out is number of retransmitted segments.
2088 * left_out is number of segments left network, but not ACKed yet.
2090 * left_out = sacked_out + lost_out
2092 * sacked_out: Packets, which arrived to receiver out of order
2093 * and hence not ACKed. With SACKs this number is simply
2094 * amount of SACKed data. Even without SACKs
2095 * it is easy to give pretty reliable estimate of this number,
2096 * counting duplicate ACKs.
2098 * lost_out: Packets lost by network. TCP has no explicit
2099 * "loss notification" feedback from network (for now).
2100 * It means that this number can be only _guessed_.
2101 * Actually, it is the heuristics to predict lossage that
2102 * distinguishes different algorithms.
2104 * F.e. after RTO, when all the queue is considered as lost,
2105 * lost_out = packets_out and in_flight = retrans_out.
2107 * Essentially, we have now two algorithms counting
2110 * FACK: It is the simplest heuristics. As soon as we decided
2111 * that something is lost, we decide that _all_ not SACKed
2112 * packets until the most forward SACK are lost. I.e.
2113 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2114 * It is absolutely correct estimate, if network does not reorder
2115 * packets. And it loses any connection to reality when reordering
2116 * takes place. We use FACK by default until reordering
2117 * is suspected on the path to this destination.
2119 * NewReno: when Recovery is entered, we assume that one segment
2120 * is lost (classic Reno). While we are in Recovery and
2121 * a partial ACK arrives, we assume that one more packet
2122 * is lost (NewReno). This heuristics are the same in NewReno
2125 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2126 * deflation etc. CWND is real congestion window, never inflated, changes
2127 * only according to classic VJ rules.
2129 * Really tricky (and requiring careful tuning) part of algorithm
2130 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2131 * The first determines the moment _when_ we should reduce CWND and,
2132 * hence, slow down forward transmission. In fact, it determines the moment
2133 * when we decide that hole is caused by loss, rather than by a reorder.
2135 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2136 * holes, caused by lost packets.
2138 * And the most logically complicated part of algorithm is undo
2139 * heuristics. We detect false retransmits due to both too early
2140 * fast retransmit (reordering) and underestimated RTO, analyzing
2141 * timestamps and D-SACKs. When we detect that some segments were
2142 * retransmitted by mistake and CWND reduction was wrong, we undo
2143 * window reduction and abort recovery phase. This logic is hidden
2144 * inside several functions named tcp_try_undo_<something>.
2147 /* This function decides, when we should leave Disordered state
2148 * and enter Recovery phase, reducing congestion window.
2150 * Main question: may we further continue forward transmission
2151 * with the same cwnd?
2153 static bool tcp_time_to_recover(struct sock *sk, int flag)
2155 struct tcp_sock *tp = tcp_sk(sk);
2158 /* Trick#1: The loss is proven. */
2162 /* Not-A-Trick#2 : Classic rule... */
2163 if (tcp_dupack_heuristics(tp) > tp->reordering)
2166 /* Trick#4: It is still not OK... But will it be useful to delay
2169 packets_out = tp->packets_out;
2170 if (packets_out <= tp->reordering &&
2171 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2172 !tcp_may_send_now(sk)) {
2173 /* We have nothing to send. This connection is limited
2174 * either by receiver window or by application.
2179 /* If a thin stream is detected, retransmit after first
2180 * received dupack. Employ only if SACK is supported in order
2181 * to avoid possible corner-case series of spurious retransmissions
2182 * Use only if there are no unsent data.
2184 if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2185 tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2186 tcp_is_sack(tp) && !tcp_send_head(sk))
2189 /* Trick#6: TCP early retransmit, per RFC5827. To avoid spurious
2190 * retransmissions due to small network reorderings, we implement
2191 * Mitigation A.3 in the RFC and delay the retransmission for a short
2192 * interval if appropriate.
2194 if (tp->do_early_retrans && !tp->retrans_out && tp->sacked_out &&
2195 (tp->packets_out >= (tp->sacked_out + 1) && tp->packets_out < 4) &&
2196 !tcp_may_send_now(sk))
2197 return !tcp_pause_early_retransmit(sk, flag);
2202 /* Detect loss in event "A" above by marking head of queue up as lost.
2203 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2204 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2205 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2206 * the maximum SACKed segments to pass before reaching this limit.
2208 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2210 struct tcp_sock *tp = tcp_sk(sk);
2211 struct sk_buff *skb;
2215 /* Use SACK to deduce losses of new sequences sent during recovery */
2216 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq;
2218 WARN_ON(packets > tp->packets_out);
2219 if (tp->lost_skb_hint) {
2220 skb = tp->lost_skb_hint;
2221 cnt = tp->lost_cnt_hint;
2222 /* Head already handled? */
2223 if (mark_head && skb != tcp_write_queue_head(sk))
2226 skb = tcp_write_queue_head(sk);
2230 tcp_for_write_queue_from(skb, sk) {
2231 if (skb == tcp_send_head(sk))
2233 /* TODO: do this better */
2234 /* this is not the most efficient way to do this... */
2235 tp->lost_skb_hint = skb;
2236 tp->lost_cnt_hint = cnt;
2238 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2242 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2243 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2244 cnt += tcp_skb_pcount(skb);
2246 if (cnt > packets) {
2247 if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2248 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2249 (oldcnt >= packets))
2252 mss = tcp_skb_mss(skb);
2253 err = tcp_fragment(sk, skb, (packets - oldcnt) * mss,
2260 tcp_skb_mark_lost(tp, skb);
2265 tcp_verify_left_out(tp);
2268 /* Account newly detected lost packet(s) */
2270 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2272 struct tcp_sock *tp = tcp_sk(sk);
2274 if (tcp_is_reno(tp)) {
2275 tcp_mark_head_lost(sk, 1, 1);
2276 } else if (tcp_is_fack(tp)) {
2277 int lost = tp->fackets_out - tp->reordering;
2280 tcp_mark_head_lost(sk, lost, 0);
2282 int sacked_upto = tp->sacked_out - tp->reordering;
2283 if (sacked_upto >= 0)
2284 tcp_mark_head_lost(sk, sacked_upto, 0);
2285 else if (fast_rexmit)
2286 tcp_mark_head_lost(sk, 1, 1);
2290 /* CWND moderation, preventing bursts due to too big ACKs
2291 * in dubious situations.
2293 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2295 tp->snd_cwnd = min(tp->snd_cwnd,
2296 tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2297 tp->snd_cwnd_stamp = tcp_time_stamp;
2300 /* Nothing was retransmitted or returned timestamp is less
2301 * than timestamp of the first retransmission.
2303 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2305 return !tp->retrans_stamp ||
2306 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2307 before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
2310 /* Undo procedures. */
2312 /* We can clear retrans_stamp when there are no retransmissions in the
2313 * window. It would seem that it is trivially available for us in
2314 * tp->retrans_out, however, that kind of assumptions doesn't consider
2315 * what will happen if errors occur when sending retransmission for the
2316 * second time. ...It could the that such segment has only
2317 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2318 * the head skb is enough except for some reneging corner cases that
2319 * are not worth the effort.
2321 * Main reason for all this complexity is the fact that connection dying
2322 * time now depends on the validity of the retrans_stamp, in particular,
2323 * that successive retransmissions of a segment must not advance
2324 * retrans_stamp under any conditions.
2326 static bool tcp_any_retrans_done(const struct sock *sk)
2328 const struct tcp_sock *tp = tcp_sk(sk);
2329 struct sk_buff *skb;
2331 if (tp->retrans_out)
2334 skb = tcp_write_queue_head(sk);
2335 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2341 #if FASTRETRANS_DEBUG > 1
2342 static void DBGUNDO(struct sock *sk, const char *msg)
2344 struct tcp_sock *tp = tcp_sk(sk);
2345 struct inet_sock *inet = inet_sk(sk);
2347 if (sk->sk_family == AF_INET) {
2348 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2350 &inet->inet_daddr, ntohs(inet->inet_dport),
2351 tp->snd_cwnd, tcp_left_out(tp),
2352 tp->snd_ssthresh, tp->prior_ssthresh,
2355 #if IS_ENABLED(CONFIG_IPV6)
2356 else if (sk->sk_family == AF_INET6) {
2357 struct ipv6_pinfo *np = inet6_sk(sk);
2358 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2360 &np->daddr, ntohs(inet->inet_dport),
2361 tp->snd_cwnd, tcp_left_out(tp),
2362 tp->snd_ssthresh, tp->prior_ssthresh,
2368 #define DBGUNDO(x...) do { } while (0)
2371 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2373 struct tcp_sock *tp = tcp_sk(sk);
2376 struct sk_buff *skb;
2378 tcp_for_write_queue(skb, sk) {
2379 if (skb == tcp_send_head(sk))
2381 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2384 tcp_clear_all_retrans_hints(tp);
2387 if (tp->prior_ssthresh) {
2388 const struct inet_connection_sock *icsk = inet_csk(sk);
2390 if (icsk->icsk_ca_ops->undo_cwnd)
2391 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2393 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2395 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2396 tp->snd_ssthresh = tp->prior_ssthresh;
2397 tcp_ecn_withdraw_cwr(tp);
2400 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2402 tp->snd_cwnd_stamp = tcp_time_stamp;
2403 tp->undo_marker = 0;
2406 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2408 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2411 /* People celebrate: "We love our President!" */
2412 static bool tcp_try_undo_recovery(struct sock *sk)
2414 struct tcp_sock *tp = tcp_sk(sk);
2416 if (tcp_may_undo(tp)) {
2419 /* Happy end! We did not retransmit anything
2420 * or our original transmission succeeded.
2422 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2423 tcp_undo_cwnd_reduction(sk, false);
2424 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2425 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2427 mib_idx = LINUX_MIB_TCPFULLUNDO;
2429 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2431 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2432 /* Hold old state until something *above* high_seq
2433 * is ACKed. For Reno it is MUST to prevent false
2434 * fast retransmits (RFC2582). SACK TCP is safe. */
2435 tcp_moderate_cwnd(tp);
2436 if (!tcp_any_retrans_done(sk))
2437 tp->retrans_stamp = 0;
2440 tcp_set_ca_state(sk, TCP_CA_Open);
2444 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2445 static bool tcp_try_undo_dsack(struct sock *sk)
2447 struct tcp_sock *tp = tcp_sk(sk);
2449 if (tp->undo_marker && !tp->undo_retrans) {
2450 DBGUNDO(sk, "D-SACK");
2451 tcp_undo_cwnd_reduction(sk, false);
2452 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2458 /* Undo during loss recovery after partial ACK or using F-RTO. */
2459 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2461 struct tcp_sock *tp = tcp_sk(sk);
2463 if (frto_undo || tcp_may_undo(tp)) {
2464 tcp_undo_cwnd_reduction(sk, true);
2466 DBGUNDO(sk, "partial loss");
2467 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2469 NET_INC_STATS_BH(sock_net(sk),
2470 LINUX_MIB_TCPSPURIOUSRTOS);
2471 inet_csk(sk)->icsk_retransmits = 0;
2472 if (frto_undo || tcp_is_sack(tp))
2473 tcp_set_ca_state(sk, TCP_CA_Open);
2479 /* The cwnd reduction in CWR and Recovery use the PRR algorithm
2480 * https://datatracker.ietf.org/doc/draft-ietf-tcpm-proportional-rate-reduction/
2481 * It computes the number of packets to send (sndcnt) based on packets newly
2483 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2484 * cwnd reductions across a full RTT.
2485 * 2) If packets in flight is lower than ssthresh (such as due to excess
2486 * losses and/or application stalls), do not perform any further cwnd
2487 * reductions, but instead slow start up to ssthresh.
2489 static void tcp_init_cwnd_reduction(struct sock *sk)
2491 struct tcp_sock *tp = tcp_sk(sk);
2493 tp->high_seq = tp->snd_nxt;
2494 tp->tlp_high_seq = 0;
2495 tp->snd_cwnd_cnt = 0;
2496 tp->prior_cwnd = tp->snd_cwnd;
2497 tp->prr_delivered = 0;
2499 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2500 tcp_ecn_queue_cwr(tp);
2503 static void tcp_cwnd_reduction(struct sock *sk, const int prior_unsacked,
2506 struct tcp_sock *tp = tcp_sk(sk);
2508 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2509 int newly_acked_sacked = prior_unsacked -
2510 (tp->packets_out - tp->sacked_out);
2512 tp->prr_delivered += newly_acked_sacked;
2513 if (tcp_packets_in_flight(tp) > tp->snd_ssthresh) {
2514 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2516 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2518 sndcnt = min_t(int, delta,
2519 max_t(int, tp->prr_delivered - tp->prr_out,
2520 newly_acked_sacked) + 1);
2523 sndcnt = max(sndcnt, (fast_rexmit ? 1 : 0));
2524 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2527 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2529 struct tcp_sock *tp = tcp_sk(sk);
2531 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2532 if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR ||
2533 (tp->undo_marker && tp->snd_ssthresh < TCP_INFINITE_SSTHRESH)) {
2534 tp->snd_cwnd = tp->snd_ssthresh;
2535 tp->snd_cwnd_stamp = tcp_time_stamp;
2537 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2540 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2541 void tcp_enter_cwr(struct sock *sk)
2543 struct tcp_sock *tp = tcp_sk(sk);
2545 tp->prior_ssthresh = 0;
2546 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2547 tp->undo_marker = 0;
2548 tcp_init_cwnd_reduction(sk);
2549 tcp_set_ca_state(sk, TCP_CA_CWR);
2552 EXPORT_SYMBOL(tcp_enter_cwr);
2554 static void tcp_try_keep_open(struct sock *sk)
2556 struct tcp_sock *tp = tcp_sk(sk);
2557 int state = TCP_CA_Open;
2559 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2560 state = TCP_CA_Disorder;
2562 if (inet_csk(sk)->icsk_ca_state != state) {
2563 tcp_set_ca_state(sk, state);
2564 tp->high_seq = tp->snd_nxt;
2568 static void tcp_try_to_open(struct sock *sk, int flag, const int prior_unsacked)
2570 struct tcp_sock *tp = tcp_sk(sk);
2572 tcp_verify_left_out(tp);
2574 if (!tcp_any_retrans_done(sk))
2575 tp->retrans_stamp = 0;
2577 if (flag & FLAG_ECE)
2580 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2581 tcp_try_keep_open(sk);
2583 tcp_cwnd_reduction(sk, prior_unsacked, 0);
2587 static void tcp_mtup_probe_failed(struct sock *sk)
2589 struct inet_connection_sock *icsk = inet_csk(sk);
2591 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2592 icsk->icsk_mtup.probe_size = 0;
2595 static void tcp_mtup_probe_success(struct sock *sk)
2597 struct tcp_sock *tp = tcp_sk(sk);
2598 struct inet_connection_sock *icsk = inet_csk(sk);
2600 /* FIXME: breaks with very large cwnd */
2601 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2602 tp->snd_cwnd = tp->snd_cwnd *
2603 tcp_mss_to_mtu(sk, tp->mss_cache) /
2604 icsk->icsk_mtup.probe_size;
2605 tp->snd_cwnd_cnt = 0;
2606 tp->snd_cwnd_stamp = tcp_time_stamp;
2607 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2609 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2610 icsk->icsk_mtup.probe_size = 0;
2611 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2614 /* Do a simple retransmit without using the backoff mechanisms in
2615 * tcp_timer. This is used for path mtu discovery.
2616 * The socket is already locked here.
2618 void tcp_simple_retransmit(struct sock *sk)
2620 const struct inet_connection_sock *icsk = inet_csk(sk);
2621 struct tcp_sock *tp = tcp_sk(sk);
2622 struct sk_buff *skb;
2623 unsigned int mss = tcp_current_mss(sk);
2624 u32 prior_lost = tp->lost_out;
2626 tcp_for_write_queue(skb, sk) {
2627 if (skb == tcp_send_head(sk))
2629 if (tcp_skb_seglen(skb) > mss &&
2630 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2631 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2632 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2633 tp->retrans_out -= tcp_skb_pcount(skb);
2635 tcp_skb_mark_lost_uncond_verify(tp, skb);
2639 tcp_clear_retrans_hints_partial(tp);
2641 if (prior_lost == tp->lost_out)
2644 if (tcp_is_reno(tp))
2645 tcp_limit_reno_sacked(tp);
2647 tcp_verify_left_out(tp);
2649 /* Don't muck with the congestion window here.
2650 * Reason is that we do not increase amount of _data_
2651 * in network, but units changed and effective
2652 * cwnd/ssthresh really reduced now.
2654 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2655 tp->high_seq = tp->snd_nxt;
2656 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2657 tp->prior_ssthresh = 0;
2658 tp->undo_marker = 0;
2659 tcp_set_ca_state(sk, TCP_CA_Loss);
2661 tcp_xmit_retransmit_queue(sk);
2663 EXPORT_SYMBOL(tcp_simple_retransmit);
2665 static void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2667 struct tcp_sock *tp = tcp_sk(sk);
2670 if (tcp_is_reno(tp))
2671 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2673 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2675 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2677 tp->prior_ssthresh = 0;
2680 if (!tcp_in_cwnd_reduction(sk)) {
2682 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2683 tcp_init_cwnd_reduction(sk);
2685 tcp_set_ca_state(sk, TCP_CA_Recovery);
2688 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2689 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2691 static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack)
2693 struct tcp_sock *tp = tcp_sk(sk);
2694 bool recovered = !before(tp->snd_una, tp->high_seq);
2696 if ((flag & FLAG_SND_UNA_ADVANCED) &&
2697 tcp_try_undo_loss(sk, false))
2700 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2701 /* Step 3.b. A timeout is spurious if not all data are
2702 * lost, i.e., never-retransmitted data are (s)acked.
2704 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2705 tcp_try_undo_loss(sk, true))
2708 if (after(tp->snd_nxt, tp->high_seq)) {
2709 if (flag & FLAG_DATA_SACKED || is_dupack)
2710 tp->frto = 0; /* Step 3.a. loss was real */
2711 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2712 tp->high_seq = tp->snd_nxt;
2713 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
2715 if (after(tp->snd_nxt, tp->high_seq))
2716 return; /* Step 2.b */
2722 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2723 tcp_try_undo_recovery(sk);
2726 if (tcp_is_reno(tp)) {
2727 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2728 * delivered. Lower inflight to clock out (re)tranmissions.
2730 if (after(tp->snd_nxt, tp->high_seq) && is_dupack)
2731 tcp_add_reno_sack(sk);
2732 else if (flag & FLAG_SND_UNA_ADVANCED)
2733 tcp_reset_reno_sack(tp);
2735 tcp_xmit_retransmit_queue(sk);
2738 /* Undo during fast recovery after partial ACK. */
2739 static bool tcp_try_undo_partial(struct sock *sk, const int acked,
2740 const int prior_unsacked)
2742 struct tcp_sock *tp = tcp_sk(sk);
2744 if (tp->undo_marker && tcp_packet_delayed(tp)) {
2745 /* Plain luck! Hole if filled with delayed
2746 * packet, rather than with a retransmit.
2748 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2750 /* We are getting evidence that the reordering degree is higher
2751 * than we realized. If there are no retransmits out then we
2752 * can undo. Otherwise we clock out new packets but do not
2753 * mark more packets lost or retransmit more.
2755 if (tp->retrans_out) {
2756 tcp_cwnd_reduction(sk, prior_unsacked, 0);
2760 if (!tcp_any_retrans_done(sk))
2761 tp->retrans_stamp = 0;
2763 DBGUNDO(sk, "partial recovery");
2764 tcp_undo_cwnd_reduction(sk, true);
2765 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2766 tcp_try_keep_open(sk);
2772 /* Process an event, which can update packets-in-flight not trivially.
2773 * Main goal of this function is to calculate new estimate for left_out,
2774 * taking into account both packets sitting in receiver's buffer and
2775 * packets lost by network.
2777 * Besides that it does CWND reduction, when packet loss is detected
2778 * and changes state of machine.
2780 * It does _not_ decide what to send, it is made in function
2781 * tcp_xmit_retransmit_queue().
2783 static void tcp_fastretrans_alert(struct sock *sk, const int acked,
2784 const int prior_unsacked,
2785 bool is_dupack, int flag)
2787 struct inet_connection_sock *icsk = inet_csk(sk);
2788 struct tcp_sock *tp = tcp_sk(sk);
2789 bool do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2790 (tcp_fackets_out(tp) > tp->reordering));
2791 int fast_rexmit = 0;
2793 if (WARN_ON(!tp->packets_out && tp->sacked_out))
2795 if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2796 tp->fackets_out = 0;
2798 /* Now state machine starts.
2799 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2800 if (flag & FLAG_ECE)
2801 tp->prior_ssthresh = 0;
2803 /* B. In all the states check for reneging SACKs. */
2804 if (tcp_check_sack_reneging(sk, flag))
2807 /* C. Check consistency of the current state. */
2808 tcp_verify_left_out(tp);
2810 /* D. Check state exit conditions. State can be terminated
2811 * when high_seq is ACKed. */
2812 if (icsk->icsk_ca_state == TCP_CA_Open) {
2813 WARN_ON(tp->retrans_out != 0);
2814 tp->retrans_stamp = 0;
2815 } else if (!before(tp->snd_una, tp->high_seq)) {
2816 switch (icsk->icsk_ca_state) {
2818 /* CWR is to be held something *above* high_seq
2819 * is ACKed for CWR bit to reach receiver. */
2820 if (tp->snd_una != tp->high_seq) {
2821 tcp_end_cwnd_reduction(sk);
2822 tcp_set_ca_state(sk, TCP_CA_Open);
2826 case TCP_CA_Recovery:
2827 if (tcp_is_reno(tp))
2828 tcp_reset_reno_sack(tp);
2829 if (tcp_try_undo_recovery(sk))
2831 tcp_end_cwnd_reduction(sk);
2836 /* E. Process state. */
2837 switch (icsk->icsk_ca_state) {
2838 case TCP_CA_Recovery:
2839 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2840 if (tcp_is_reno(tp) && is_dupack)
2841 tcp_add_reno_sack(sk);
2843 if (tcp_try_undo_partial(sk, acked, prior_unsacked))
2845 /* Partial ACK arrived. Force fast retransmit. */
2846 do_lost = tcp_is_reno(tp) ||
2847 tcp_fackets_out(tp) > tp->reordering;
2849 if (tcp_try_undo_dsack(sk)) {
2850 tcp_try_keep_open(sk);
2855 tcp_process_loss(sk, flag, is_dupack);
2856 if (icsk->icsk_ca_state != TCP_CA_Open &&
2857 !(flag & FLAG_LOST_RETRANS))
2859 /* Change state if cwnd is undone or retransmits are lost */
2861 if (tcp_is_reno(tp)) {
2862 if (flag & FLAG_SND_UNA_ADVANCED)
2863 tcp_reset_reno_sack(tp);
2865 tcp_add_reno_sack(sk);
2868 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2869 tcp_try_undo_dsack(sk);
2871 if (!tcp_time_to_recover(sk, flag)) {
2872 tcp_try_to_open(sk, flag, prior_unsacked);
2876 /* MTU probe failure: don't reduce cwnd */
2877 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2878 icsk->icsk_mtup.probe_size &&
2879 tp->snd_una == tp->mtu_probe.probe_seq_start) {
2880 tcp_mtup_probe_failed(sk);
2881 /* Restores the reduction we did in tcp_mtup_probe() */
2883 tcp_simple_retransmit(sk);
2887 /* Otherwise enter Recovery state */
2888 tcp_enter_recovery(sk, (flag & FLAG_ECE));
2893 tcp_update_scoreboard(sk, fast_rexmit);
2894 tcp_cwnd_reduction(sk, prior_unsacked, fast_rexmit);
2895 tcp_xmit_retransmit_queue(sk);
2898 static inline bool tcp_ack_update_rtt(struct sock *sk, const int flag,
2899 long seq_rtt_us, long sack_rtt_us)
2901 const struct tcp_sock *tp = tcp_sk(sk);
2903 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2904 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2905 * Karn's algorithm forbids taking RTT if some retransmitted data
2906 * is acked (RFC6298).
2908 if (flag & FLAG_RETRANS_DATA_ACKED)
2912 seq_rtt_us = sack_rtt_us;
2914 /* RTTM Rule: A TSecr value received in a segment is used to
2915 * update the averaged RTT measurement only if the segment
2916 * acknowledges some new data, i.e., only if it advances the
2917 * left edge of the send window.
2918 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2920 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2922 seq_rtt_us = jiffies_to_usecs(tcp_time_stamp - tp->rx_opt.rcv_tsecr);
2927 tcp_rtt_estimator(sk, seq_rtt_us);
2930 /* RFC6298: only reset backoff on valid RTT measurement. */
2931 inet_csk(sk)->icsk_backoff = 0;
2935 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2936 static void tcp_synack_rtt_meas(struct sock *sk, const u32 synack_stamp)
2938 struct tcp_sock *tp = tcp_sk(sk);
2939 long seq_rtt_us = -1L;
2941 if (synack_stamp && !tp->total_retrans)
2942 seq_rtt_us = jiffies_to_usecs(tcp_time_stamp - synack_stamp);
2944 /* If the ACK acks both the SYNACK and the (Fast Open'd) data packets
2945 * sent in SYN_RECV, SYNACK RTT is the smooth RTT computed in tcp_ack()
2948 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, seq_rtt_us, -1L);
2951 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
2953 const struct inet_connection_sock *icsk = inet_csk(sk);
2955 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
2956 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
2959 /* Restart timer after forward progress on connection.
2960 * RFC2988 recommends to restart timer to now+rto.
2962 void tcp_rearm_rto(struct sock *sk)
2964 const struct inet_connection_sock *icsk = inet_csk(sk);
2965 struct tcp_sock *tp = tcp_sk(sk);
2967 /* If the retrans timer is currently being used by Fast Open
2968 * for SYN-ACK retrans purpose, stay put.
2970 if (tp->fastopen_rsk)
2973 if (!tp->packets_out) {
2974 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2976 u32 rto = inet_csk(sk)->icsk_rto;
2977 /* Offset the time elapsed after installing regular RTO */
2978 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
2979 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
2980 struct sk_buff *skb = tcp_write_queue_head(sk);
2981 const u32 rto_time_stamp =
2982 tcp_skb_timestamp(skb) + rto;
2983 s32 delta = (s32)(rto_time_stamp - tcp_time_stamp);
2984 /* delta may not be positive if the socket is locked
2985 * when the retrans timer fires and is rescheduled.
2990 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
2995 /* This function is called when the delayed ER timer fires. TCP enters
2996 * fast recovery and performs fast-retransmit.
2998 void tcp_resume_early_retransmit(struct sock *sk)
3000 struct tcp_sock *tp = tcp_sk(sk);
3004 /* Stop if ER is disabled after the delayed ER timer is scheduled */
3005 if (!tp->do_early_retrans)
3008 tcp_enter_recovery(sk, false);
3009 tcp_update_scoreboard(sk, 1);
3010 tcp_xmit_retransmit_queue(sk);
3013 /* If we get here, the whole TSO packet has not been acked. */
3014 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3016 struct tcp_sock *tp = tcp_sk(sk);
3019 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3021 packets_acked = tcp_skb_pcount(skb);
3022 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3024 packets_acked -= tcp_skb_pcount(skb);
3026 if (packets_acked) {
3027 BUG_ON(tcp_skb_pcount(skb) == 0);
3028 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3031 return packets_acked;
3034 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3037 const struct skb_shared_info *shinfo;
3039 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3040 if (likely(!(sk->sk_tsflags & SOF_TIMESTAMPING_TX_ACK)))
3043 shinfo = skb_shinfo(skb);
3044 if ((shinfo->tx_flags & SKBTX_ACK_TSTAMP) &&
3045 between(shinfo->tskey, prior_snd_una, tcp_sk(sk)->snd_una - 1))
3046 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK);
3049 /* Remove acknowledged frames from the retransmission queue. If our packet
3050 * is before the ack sequence we can discard it as it's confirmed to have
3051 * arrived at the other end.
3053 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3055 struct tcp_sacktag_state *sack)
3057 const struct inet_connection_sock *icsk = inet_csk(sk);
3058 struct skb_mstamp first_ackt, last_ackt, now;
3059 struct tcp_sock *tp = tcp_sk(sk);
3060 u32 prior_sacked = tp->sacked_out;
3061 u32 reord = tp->packets_out;
3062 bool fully_acked = true;
3063 long sack_rtt_us = -1L;
3064 long seq_rtt_us = -1L;
3065 long ca_rtt_us = -1L;
3066 struct sk_buff *skb;
3073 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3074 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3075 u8 sacked = scb->sacked;
3078 tcp_ack_tstamp(sk, skb, prior_snd_una);
3080 /* Determine how many packets and what bytes were acked, tso and else */
3081 if (after(scb->end_seq, tp->snd_una)) {
3082 if (tcp_skb_pcount(skb) == 1 ||
3083 !after(tp->snd_una, scb->seq))
3086 acked_pcount = tcp_tso_acked(sk, skb);
3090 fully_acked = false;
3092 /* Speedup tcp_unlink_write_queue() and next loop */
3093 prefetchw(skb->next);
3094 acked_pcount = tcp_skb_pcount(skb);
3097 if (unlikely(sacked & TCPCB_RETRANS)) {
3098 if (sacked & TCPCB_SACKED_RETRANS)
3099 tp->retrans_out -= acked_pcount;
3100 flag |= FLAG_RETRANS_DATA_ACKED;
3101 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3102 last_ackt = skb->skb_mstamp;
3103 WARN_ON_ONCE(last_ackt.v64 == 0);
3104 if (!first_ackt.v64)
3105 first_ackt = last_ackt;
3107 reord = min(pkts_acked, reord);
3108 if (!after(scb->end_seq, tp->high_seq))
3109 flag |= FLAG_ORIG_SACK_ACKED;
3112 if (sacked & TCPCB_SACKED_ACKED)
3113 tp->sacked_out -= acked_pcount;
3114 if (sacked & TCPCB_LOST)
3115 tp->lost_out -= acked_pcount;
3117 tp->packets_out -= acked_pcount;
3118 pkts_acked += acked_pcount;
3120 /* Initial outgoing SYN's get put onto the write_queue
3121 * just like anything else we transmit. It is not
3122 * true data, and if we misinform our callers that
3123 * this ACK acks real data, we will erroneously exit
3124 * connection startup slow start one packet too
3125 * quickly. This is severely frowned upon behavior.
3127 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3128 flag |= FLAG_DATA_ACKED;
3130 flag |= FLAG_SYN_ACKED;
3131 tp->retrans_stamp = 0;
3137 tcp_unlink_write_queue(skb, sk);
3138 sk_wmem_free_skb(sk, skb);
3139 if (unlikely(skb == tp->retransmit_skb_hint))
3140 tp->retransmit_skb_hint = NULL;
3141 if (unlikely(skb == tp->lost_skb_hint))
3142 tp->lost_skb_hint = NULL;
3145 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3146 tp->snd_up = tp->snd_una;
3148 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3149 flag |= FLAG_SACK_RENEGING;
3151 skb_mstamp_get(&now);
3152 if (likely(first_ackt.v64)) {
3153 seq_rtt_us = skb_mstamp_us_delta(&now, &first_ackt);
3154 ca_rtt_us = skb_mstamp_us_delta(&now, &last_ackt);
3156 if (sack->first_sackt.v64) {
3157 sack_rtt_us = skb_mstamp_us_delta(&now, &sack->first_sackt);
3158 ca_rtt_us = skb_mstamp_us_delta(&now, &sack->last_sackt);
3161 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us);
3163 if (flag & FLAG_ACKED) {
3165 if (unlikely(icsk->icsk_mtup.probe_size &&
3166 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3167 tcp_mtup_probe_success(sk);
3170 if (tcp_is_reno(tp)) {
3171 tcp_remove_reno_sacks(sk, pkts_acked);
3175 /* Non-retransmitted hole got filled? That's reordering */
3176 if (reord < prior_fackets)
3177 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3179 delta = tcp_is_fack(tp) ? pkts_acked :
3180 prior_sacked - tp->sacked_out;
3181 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3184 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3186 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3187 sack_rtt_us > skb_mstamp_us_delta(&now, &skb->skb_mstamp)) {
3188 /* Do not re-arm RTO if the sack RTT is measured from data sent
3189 * after when the head was last (re)transmitted. Otherwise the
3190 * timeout may continue to extend in loss recovery.
3195 if (icsk->icsk_ca_ops->pkts_acked)
3196 icsk->icsk_ca_ops->pkts_acked(sk, pkts_acked, ca_rtt_us);
3198 #if FASTRETRANS_DEBUG > 0
3199 WARN_ON((int)tp->sacked_out < 0);
3200 WARN_ON((int)tp->lost_out < 0);
3201 WARN_ON((int)tp->retrans_out < 0);
3202 if (!tp->packets_out && tcp_is_sack(tp)) {
3203 icsk = inet_csk(sk);
3205 pr_debug("Leak l=%u %d\n",
3206 tp->lost_out, icsk->icsk_ca_state);
3209 if (tp->sacked_out) {
3210 pr_debug("Leak s=%u %d\n",
3211 tp->sacked_out, icsk->icsk_ca_state);
3214 if (tp->retrans_out) {
3215 pr_debug("Leak r=%u %d\n",
3216 tp->retrans_out, icsk->icsk_ca_state);
3217 tp->retrans_out = 0;
3224 static void tcp_ack_probe(struct sock *sk)
3226 const struct tcp_sock *tp = tcp_sk(sk);
3227 struct inet_connection_sock *icsk = inet_csk(sk);
3229 /* Was it a usable window open? */
3231 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3232 icsk->icsk_backoff = 0;
3233 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3234 /* Socket must be waked up by subsequent tcp_data_snd_check().
3235 * This function is not for random using!
3238 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
3240 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3245 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3247 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3248 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3251 /* Decide wheather to run the increase function of congestion control. */
3252 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3254 if (tcp_in_cwnd_reduction(sk))
3257 /* If reordering is high then always grow cwnd whenever data is
3258 * delivered regardless of its ordering. Otherwise stay conservative
3259 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3260 * new SACK or ECE mark may first advance cwnd here and later reduce
3261 * cwnd in tcp_fastretrans_alert() based on more states.
3263 if (tcp_sk(sk)->reordering > sysctl_tcp_reordering)
3264 return flag & FLAG_FORWARD_PROGRESS;
3266 return flag & FLAG_DATA_ACKED;
3269 /* Check that window update is acceptable.
3270 * The function assumes that snd_una<=ack<=snd_next.
3272 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3273 const u32 ack, const u32 ack_seq,
3276 return after(ack, tp->snd_una) ||
3277 after(ack_seq, tp->snd_wl1) ||
3278 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3281 /* If we update tp->snd_una, also update tp->bytes_acked */
3282 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3284 u32 delta = ack - tp->snd_una;
3286 u64_stats_update_begin(&tp->syncp);
3287 tp->bytes_acked += delta;
3288 u64_stats_update_end(&tp->syncp);
3292 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3293 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3295 u32 delta = seq - tp->rcv_nxt;
3297 u64_stats_update_begin(&tp->syncp);
3298 tp->bytes_received += delta;
3299 u64_stats_update_end(&tp->syncp);
3303 /* Update our send window.
3305 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3306 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3308 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3311 struct tcp_sock *tp = tcp_sk(sk);
3313 u32 nwin = ntohs(tcp_hdr(skb)->window);
3315 if (likely(!tcp_hdr(skb)->syn))
3316 nwin <<= tp->rx_opt.snd_wscale;
3318 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3319 flag |= FLAG_WIN_UPDATE;
3320 tcp_update_wl(tp, ack_seq);
3322 if (tp->snd_wnd != nwin) {
3325 /* Note, it is the only place, where
3326 * fast path is recovered for sending TCP.
3329 tcp_fast_path_check(sk);
3331 if (nwin > tp->max_window) {
3332 tp->max_window = nwin;
3333 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3338 tcp_snd_una_update(tp, ack);
3343 /* Return true if we're currently rate-limiting out-of-window ACKs and
3344 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3345 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3346 * attacks that send repeated SYNs or ACKs for the same connection. To
3347 * do this, we do not send a duplicate SYNACK or ACK if the remote
3348 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3350 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3351 int mib_idx, u32 *last_oow_ack_time)
3353 /* Data packets without SYNs are not likely part of an ACK loop. */
3354 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3356 goto not_rate_limited;
3358 if (*last_oow_ack_time) {
3359 s32 elapsed = (s32)(tcp_time_stamp - *last_oow_ack_time);
3361 if (0 <= elapsed && elapsed < sysctl_tcp_invalid_ratelimit) {
3362 NET_INC_STATS_BH(net, mib_idx);
3363 return true; /* rate-limited: don't send yet! */
3367 *last_oow_ack_time = tcp_time_stamp;
3370 return false; /* not rate-limited: go ahead, send dupack now! */
3373 /* RFC 5961 7 [ACK Throttling] */
3374 static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb)
3376 /* unprotected vars, we dont care of overwrites */
3377 static u32 challenge_timestamp;
3378 static unsigned int challenge_count;
3379 struct tcp_sock *tp = tcp_sk(sk);
3382 /* First check our per-socket dupack rate limit. */
3383 if (tcp_oow_rate_limited(sock_net(sk), skb,
3384 LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3385 &tp->last_oow_ack_time))
3388 /* Then check the check host-wide RFC 5961 rate limit. */
3390 if (now != challenge_timestamp) {
3391 challenge_timestamp = now;
3392 challenge_count = 0;
3394 if (++challenge_count <= sysctl_tcp_challenge_ack_limit) {
3395 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPCHALLENGEACK);
3400 static void tcp_store_ts_recent(struct tcp_sock *tp)
3402 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3403 tp->rx_opt.ts_recent_stamp = get_seconds();
3406 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3408 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3409 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3410 * extra check below makes sure this can only happen
3411 * for pure ACK frames. -DaveM
3413 * Not only, also it occurs for expired timestamps.
3416 if (tcp_paws_check(&tp->rx_opt, 0))
3417 tcp_store_ts_recent(tp);
3421 /* This routine deals with acks during a TLP episode.
3422 * We mark the end of a TLP episode on receiving TLP dupack or when
3423 * ack is after tlp_high_seq.
3424 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3426 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3428 struct tcp_sock *tp = tcp_sk(sk);
3430 if (before(ack, tp->tlp_high_seq))
3433 if (flag & FLAG_DSACKING_ACK) {
3434 /* This DSACK means original and TLP probe arrived; no loss */
3435 tp->tlp_high_seq = 0;
3436 } else if (after(ack, tp->tlp_high_seq)) {
3437 /* ACK advances: there was a loss, so reduce cwnd. Reset
3438 * tlp_high_seq in tcp_init_cwnd_reduction()
3440 tcp_init_cwnd_reduction(sk);
3441 tcp_set_ca_state(sk, TCP_CA_CWR);
3442 tcp_end_cwnd_reduction(sk);
3443 tcp_try_keep_open(sk);
3444 NET_INC_STATS_BH(sock_net(sk),
3445 LINUX_MIB_TCPLOSSPROBERECOVERY);
3446 } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3447 FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3448 /* Pure dupack: original and TLP probe arrived; no loss */
3449 tp->tlp_high_seq = 0;
3453 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3455 const struct inet_connection_sock *icsk = inet_csk(sk);
3457 if (icsk->icsk_ca_ops->in_ack_event)
3458 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3461 /* This routine deals with incoming acks, but not outgoing ones. */
3462 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3464 struct inet_connection_sock *icsk = inet_csk(sk);
3465 struct tcp_sock *tp = tcp_sk(sk);
3466 struct tcp_sacktag_state sack_state;
3467 u32 prior_snd_una = tp->snd_una;
3468 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3469 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3470 bool is_dupack = false;
3472 int prior_packets = tp->packets_out;
3473 const int prior_unsacked = tp->packets_out - tp->sacked_out;
3474 int acked = 0; /* Number of packets newly acked */
3476 sack_state.first_sackt.v64 = 0;
3478 /* We very likely will need to access write queue head. */
3479 prefetchw(sk->sk_write_queue.next);
3481 /* If the ack is older than previous acks
3482 * then we can probably ignore it.
3484 if (before(ack, prior_snd_una)) {
3485 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3486 if (before(ack, prior_snd_una - tp->max_window)) {
3487 tcp_send_challenge_ack(sk, skb);
3493 /* If the ack includes data we haven't sent yet, discard
3494 * this segment (RFC793 Section 3.9).
3496 if (after(ack, tp->snd_nxt))
3499 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
3500 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
3503 if (after(ack, prior_snd_una)) {
3504 flag |= FLAG_SND_UNA_ADVANCED;
3505 icsk->icsk_retransmits = 0;
3508 prior_fackets = tp->fackets_out;
3510 /* ts_recent update must be made after we are sure that the packet
3513 if (flag & FLAG_UPDATE_TS_RECENT)
3514 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3516 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3517 /* Window is constant, pure forward advance.
3518 * No more checks are required.
3519 * Note, we use the fact that SND.UNA>=SND.WL2.
3521 tcp_update_wl(tp, ack_seq);
3522 tcp_snd_una_update(tp, ack);
3523 flag |= FLAG_WIN_UPDATE;
3525 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3527 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3529 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3531 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3534 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3536 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3538 if (TCP_SKB_CB(skb)->sacked)
3539 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3542 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3544 ack_ev_flags |= CA_ACK_ECE;
3547 if (flag & FLAG_WIN_UPDATE)
3548 ack_ev_flags |= CA_ACK_WIN_UPDATE;
3550 tcp_in_ack_event(sk, ack_ev_flags);
3553 /* We passed data and got it acked, remove any soft error
3554 * log. Something worked...
3556 sk->sk_err_soft = 0;
3557 icsk->icsk_probes_out = 0;
3558 tp->rcv_tstamp = tcp_time_stamp;
3562 /* See if we can take anything off of the retransmit queue. */
3563 acked = tp->packets_out;
3564 flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una,
3566 acked -= tp->packets_out;
3568 /* Advance cwnd if state allows */
3569 if (tcp_may_raise_cwnd(sk, flag))
3570 tcp_cong_avoid(sk, ack, acked);
3572 if (tcp_ack_is_dubious(sk, flag)) {
3573 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3574 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3577 if (tp->tlp_high_seq)
3578 tcp_process_tlp_ack(sk, ack, flag);
3580 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) {
3581 struct dst_entry *dst = __sk_dst_get(sk);
3586 if (icsk->icsk_pending == ICSK_TIME_RETRANS)
3587 tcp_schedule_loss_probe(sk);
3588 tcp_update_pacing_rate(sk);
3592 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3593 if (flag & FLAG_DSACKING_ACK)
3594 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3596 /* If this ack opens up a zero window, clear backoff. It was
3597 * being used to time the probes, and is probably far higher than
3598 * it needs to be for normal retransmission.
3600 if (tcp_send_head(sk))
3603 if (tp->tlp_high_seq)
3604 tcp_process_tlp_ack(sk, ack, flag);
3608 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3612 /* If data was SACKed, tag it and see if we should send more data.
3613 * If data was DSACKed, see if we can undo a cwnd reduction.
3615 if (TCP_SKB_CB(skb)->sacked) {
3616 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3618 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3622 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3626 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3627 bool syn, struct tcp_fastopen_cookie *foc,
3630 /* Valid only in SYN or SYN-ACK with an even length. */
3631 if (!foc || !syn || len < 0 || (len & 1))
3634 if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3635 len <= TCP_FASTOPEN_COOKIE_MAX)
3636 memcpy(foc->val, cookie, len);
3643 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3644 * But, this can also be called on packets in the established flow when
3645 * the fast version below fails.
3647 void tcp_parse_options(const struct sk_buff *skb,
3648 struct tcp_options_received *opt_rx, int estab,
3649 struct tcp_fastopen_cookie *foc)
3651 const unsigned char *ptr;
3652 const struct tcphdr *th = tcp_hdr(skb);
3653 int length = (th->doff * 4) - sizeof(struct tcphdr);
3655 ptr = (const unsigned char *)(th + 1);
3656 opt_rx->saw_tstamp = 0;
3658 while (length > 0) {
3659 int opcode = *ptr++;
3665 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3670 if (opsize < 2) /* "silly options" */
3672 if (opsize > length)
3673 return; /* don't parse partial options */
3676 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3677 u16 in_mss = get_unaligned_be16(ptr);
3679 if (opt_rx->user_mss &&
3680 opt_rx->user_mss < in_mss)
3681 in_mss = opt_rx->user_mss;
3682 opt_rx->mss_clamp = in_mss;
3687 if (opsize == TCPOLEN_WINDOW && th->syn &&
3688 !estab && sysctl_tcp_window_scaling) {
3689 __u8 snd_wscale = *(__u8 *)ptr;
3690 opt_rx->wscale_ok = 1;
3691 if (snd_wscale > 14) {
3692 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3697 opt_rx->snd_wscale = snd_wscale;
3700 case TCPOPT_TIMESTAMP:
3701 if ((opsize == TCPOLEN_TIMESTAMP) &&
3702 ((estab && opt_rx->tstamp_ok) ||
3703 (!estab && sysctl_tcp_timestamps))) {
3704 opt_rx->saw_tstamp = 1;
3705 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3706 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3709 case TCPOPT_SACK_PERM:
3710 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3711 !estab && sysctl_tcp_sack) {
3712 opt_rx->sack_ok = TCP_SACK_SEEN;
3713 tcp_sack_reset(opt_rx);
3718 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3719 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3721 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3724 #ifdef CONFIG_TCP_MD5SIG
3727 * The MD5 Hash has already been
3728 * checked (see tcp_v{4,6}_do_rcv()).
3732 case TCPOPT_FASTOPEN:
3733 tcp_parse_fastopen_option(
3734 opsize - TCPOLEN_FASTOPEN_BASE,
3735 ptr, th->syn, foc, false);
3739 /* Fast Open option shares code 254 using a
3740 * 16 bits magic number.
3742 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
3743 get_unaligned_be16(ptr) ==
3744 TCPOPT_FASTOPEN_MAGIC)
3745 tcp_parse_fastopen_option(opsize -
3746 TCPOLEN_EXP_FASTOPEN_BASE,
3747 ptr + 2, th->syn, foc, true);
3756 EXPORT_SYMBOL(tcp_parse_options);
3758 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3760 const __be32 *ptr = (const __be32 *)(th + 1);
3762 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3763 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3764 tp->rx_opt.saw_tstamp = 1;
3766 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3769 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3771 tp->rx_opt.rcv_tsecr = 0;
3777 /* Fast parse options. This hopes to only see timestamps.
3778 * If it is wrong it falls back on tcp_parse_options().
3780 static bool tcp_fast_parse_options(const struct sk_buff *skb,
3781 const struct tcphdr *th, struct tcp_sock *tp)
3783 /* In the spirit of fast parsing, compare doff directly to constant
3784 * values. Because equality is used, short doff can be ignored here.
3786 if (th->doff == (sizeof(*th) / 4)) {
3787 tp->rx_opt.saw_tstamp = 0;
3789 } else if (tp->rx_opt.tstamp_ok &&
3790 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3791 if (tcp_parse_aligned_timestamp(tp, th))
3795 tcp_parse_options(skb, &tp->rx_opt, 1, NULL);
3796 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3797 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3802 #ifdef CONFIG_TCP_MD5SIG
3804 * Parse MD5 Signature option
3806 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3808 int length = (th->doff << 2) - sizeof(*th);
3809 const u8 *ptr = (const u8 *)(th + 1);
3811 /* If the TCP option is too short, we can short cut */
3812 if (length < TCPOLEN_MD5SIG)
3815 while (length > 0) {
3816 int opcode = *ptr++;
3827 if (opsize < 2 || opsize > length)
3829 if (opcode == TCPOPT_MD5SIG)
3830 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3837 EXPORT_SYMBOL(tcp_parse_md5sig_option);
3840 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3842 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3843 * it can pass through stack. So, the following predicate verifies that
3844 * this segment is not used for anything but congestion avoidance or
3845 * fast retransmit. Moreover, we even are able to eliminate most of such
3846 * second order effects, if we apply some small "replay" window (~RTO)
3847 * to timestamp space.
3849 * All these measures still do not guarantee that we reject wrapped ACKs
3850 * on networks with high bandwidth, when sequence space is recycled fastly,
3851 * but it guarantees that such events will be very rare and do not affect
3852 * connection seriously. This doesn't look nice, but alas, PAWS is really
3855 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3856 * states that events when retransmit arrives after original data are rare.
3857 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3858 * the biggest problem on large power networks even with minor reordering.
3859 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3860 * up to bandwidth of 18Gigabit/sec. 8) ]
3863 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3865 const struct tcp_sock *tp = tcp_sk(sk);
3866 const struct tcphdr *th = tcp_hdr(skb);
3867 u32 seq = TCP_SKB_CB(skb)->seq;
3868 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3870 return (/* 1. Pure ACK with correct sequence number. */
3871 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3873 /* 2. ... and duplicate ACK. */
3874 ack == tp->snd_una &&
3876 /* 3. ... and does not update window. */
3877 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3879 /* 4. ... and sits in replay window. */
3880 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3883 static inline bool tcp_paws_discard(const struct sock *sk,
3884 const struct sk_buff *skb)
3886 const struct tcp_sock *tp = tcp_sk(sk);
3888 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
3889 !tcp_disordered_ack(sk, skb);
3892 /* Check segment sequence number for validity.
3894 * Segment controls are considered valid, if the segment
3895 * fits to the window after truncation to the window. Acceptability
3896 * of data (and SYN, FIN, of course) is checked separately.
3897 * See tcp_data_queue(), for example.
3899 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3900 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3901 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3902 * (borrowed from freebsd)
3905 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
3907 return !before(end_seq, tp->rcv_wup) &&
3908 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3911 /* When we get a reset we do this. */
3912 void tcp_reset(struct sock *sk)
3914 /* We want the right error as BSD sees it (and indeed as we do). */
3915 switch (sk->sk_state) {
3917 sk->sk_err = ECONNREFUSED;
3919 case TCP_CLOSE_WAIT:
3925 sk->sk_err = ECONNRESET;
3927 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3930 if (!sock_flag(sk, SOCK_DEAD))
3931 sk->sk_error_report(sk);
3937 * Process the FIN bit. This now behaves as it is supposed to work
3938 * and the FIN takes effect when it is validly part of sequence
3939 * space. Not before when we get holes.
3941 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3942 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3945 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3946 * close and we go into CLOSING (and later onto TIME-WAIT)
3948 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3950 static void tcp_fin(struct sock *sk)
3952 struct tcp_sock *tp = tcp_sk(sk);
3953 const struct dst_entry *dst;
3955 inet_csk_schedule_ack(sk);
3957 sk->sk_shutdown |= RCV_SHUTDOWN;
3958 sock_set_flag(sk, SOCK_DONE);
3960 switch (sk->sk_state) {
3962 case TCP_ESTABLISHED:
3963 /* Move to CLOSE_WAIT */
3964 tcp_set_state(sk, TCP_CLOSE_WAIT);
3965 dst = __sk_dst_get(sk);
3966 if (!dst || !dst_metric(dst, RTAX_QUICKACK))
3967 inet_csk(sk)->icsk_ack.pingpong = 1;
3970 case TCP_CLOSE_WAIT:
3972 /* Received a retransmission of the FIN, do
3977 /* RFC793: Remain in the LAST-ACK state. */
3981 /* This case occurs when a simultaneous close
3982 * happens, we must ack the received FIN and
3983 * enter the CLOSING state.
3986 tcp_set_state(sk, TCP_CLOSING);
3989 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3991 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
3994 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3995 * cases we should never reach this piece of code.
3997 pr_err("%s: Impossible, sk->sk_state=%d\n",
3998 __func__, sk->sk_state);
4002 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4003 * Probably, we should reset in this case. For now drop them.
4005 __skb_queue_purge(&tp->out_of_order_queue);
4006 if (tcp_is_sack(tp))
4007 tcp_sack_reset(&tp->rx_opt);
4010 if (!sock_flag(sk, SOCK_DEAD)) {
4011 sk->sk_state_change(sk);
4013 /* Do not send POLL_HUP for half duplex close. */
4014 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4015 sk->sk_state == TCP_CLOSE)
4016 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4018 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4022 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4025 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4026 if (before(seq, sp->start_seq))
4027 sp->start_seq = seq;
4028 if (after(end_seq, sp->end_seq))
4029 sp->end_seq = end_seq;
4035 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4037 struct tcp_sock *tp = tcp_sk(sk);
4039 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4042 if (before(seq, tp->rcv_nxt))
4043 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4045 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4047 NET_INC_STATS_BH(sock_net(sk), mib_idx);
4049 tp->rx_opt.dsack = 1;
4050 tp->duplicate_sack[0].start_seq = seq;
4051 tp->duplicate_sack[0].end_seq = end_seq;
4055 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4057 struct tcp_sock *tp = tcp_sk(sk);
4059 if (!tp->rx_opt.dsack)
4060 tcp_dsack_set(sk, seq, end_seq);
4062 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4065 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4067 struct tcp_sock *tp = tcp_sk(sk);
4069 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4070 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4071 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4072 tcp_enter_quickack_mode(sk);
4074 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4075 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4077 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4078 end_seq = tp->rcv_nxt;
4079 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4086 /* These routines update the SACK block as out-of-order packets arrive or
4087 * in-order packets close up the sequence space.
4089 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4092 struct tcp_sack_block *sp = &tp->selective_acks[0];
4093 struct tcp_sack_block *swalk = sp + 1;
4095 /* See if the recent change to the first SACK eats into
4096 * or hits the sequence space of other SACK blocks, if so coalesce.
4098 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4099 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4102 /* Zap SWALK, by moving every further SACK up by one slot.
4103 * Decrease num_sacks.
4105 tp->rx_opt.num_sacks--;
4106 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4110 this_sack++, swalk++;
4114 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4116 struct tcp_sock *tp = tcp_sk(sk);
4117 struct tcp_sack_block *sp = &tp->selective_acks[0];
4118 int cur_sacks = tp->rx_opt.num_sacks;
4124 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4125 if (tcp_sack_extend(sp, seq, end_seq)) {
4126 /* Rotate this_sack to the first one. */
4127 for (; this_sack > 0; this_sack--, sp--)
4128 swap(*sp, *(sp - 1));
4130 tcp_sack_maybe_coalesce(tp);
4135 /* Could not find an adjacent existing SACK, build a new one,
4136 * put it at the front, and shift everyone else down. We
4137 * always know there is at least one SACK present already here.
4139 * If the sack array is full, forget about the last one.
4141 if (this_sack >= TCP_NUM_SACKS) {
4143 tp->rx_opt.num_sacks--;
4146 for (; this_sack > 0; this_sack--, sp--)
4150 /* Build the new head SACK, and we're done. */
4151 sp->start_seq = seq;
4152 sp->end_seq = end_seq;
4153 tp->rx_opt.num_sacks++;
4156 /* RCV.NXT advances, some SACKs should be eaten. */
4158 static void tcp_sack_remove(struct tcp_sock *tp)
4160 struct tcp_sack_block *sp = &tp->selective_acks[0];
4161 int num_sacks = tp->rx_opt.num_sacks;
4164 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4165 if (skb_queue_empty(&tp->out_of_order_queue)) {
4166 tp->rx_opt.num_sacks = 0;
4170 for (this_sack = 0; this_sack < num_sacks;) {
4171 /* Check if the start of the sack is covered by RCV.NXT. */
4172 if (!before(tp->rcv_nxt, sp->start_seq)) {
4175 /* RCV.NXT must cover all the block! */
4176 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4178 /* Zap this SACK, by moving forward any other SACKS. */
4179 for (i = this_sack+1; i < num_sacks; i++)
4180 tp->selective_acks[i-1] = tp->selective_acks[i];
4187 tp->rx_opt.num_sacks = num_sacks;
4191 * tcp_try_coalesce - try to merge skb to prior one
4194 * @from: buffer to add in queue
4195 * @fragstolen: pointer to boolean
4197 * Before queueing skb @from after @to, try to merge them
4198 * to reduce overall memory use and queue lengths, if cost is small.
4199 * Packets in ofo or receive queues can stay a long time.
4200 * Better try to coalesce them right now to avoid future collapses.
4201 * Returns true if caller should free @from instead of queueing it
4203 static bool tcp_try_coalesce(struct sock *sk,
4205 struct sk_buff *from,
4210 *fragstolen = false;
4212 /* Its possible this segment overlaps with prior segment in queue */
4213 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4216 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4219 atomic_add(delta, &sk->sk_rmem_alloc);
4220 sk_mem_charge(sk, delta);
4221 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4222 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4223 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4224 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4228 /* This one checks to see if we can put data from the
4229 * out_of_order queue into the receive_queue.
4231 static void tcp_ofo_queue(struct sock *sk)
4233 struct tcp_sock *tp = tcp_sk(sk);
4234 __u32 dsack_high = tp->rcv_nxt;
4235 struct sk_buff *skb, *tail;
4236 bool fragstolen, eaten;
4238 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4239 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4242 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4243 __u32 dsack = dsack_high;
4244 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4245 dsack_high = TCP_SKB_CB(skb)->end_seq;
4246 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4249 __skb_unlink(skb, &tp->out_of_order_queue);
4250 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4251 SOCK_DEBUG(sk, "ofo packet was already received\n");
4255 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4256 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4257 TCP_SKB_CB(skb)->end_seq);
4259 tail = skb_peek_tail(&sk->sk_receive_queue);
4260 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4261 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4263 __skb_queue_tail(&sk->sk_receive_queue, skb);
4264 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4267 kfree_skb_partial(skb, fragstolen);
4271 static bool tcp_prune_ofo_queue(struct sock *sk);
4272 static int tcp_prune_queue(struct sock *sk);
4274 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4277 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4278 !sk_rmem_schedule(sk, skb, size)) {
4280 if (tcp_prune_queue(sk) < 0)
4283 if (!sk_rmem_schedule(sk, skb, size)) {
4284 if (!tcp_prune_ofo_queue(sk))
4287 if (!sk_rmem_schedule(sk, skb, size))
4294 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4296 struct tcp_sock *tp = tcp_sk(sk);
4297 struct sk_buff *skb1;
4300 tcp_ecn_check_ce(tp, skb);
4302 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4303 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFODROP);
4308 /* Disable header prediction. */
4310 inet_csk_schedule_ack(sk);
4312 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4313 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4314 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4316 skb1 = skb_peek_tail(&tp->out_of_order_queue);
4318 /* Initial out of order segment, build 1 SACK. */
4319 if (tcp_is_sack(tp)) {
4320 tp->rx_opt.num_sacks = 1;
4321 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4322 tp->selective_acks[0].end_seq =
4323 TCP_SKB_CB(skb)->end_seq;
4325 __skb_queue_head(&tp->out_of_order_queue, skb);
4329 seq = TCP_SKB_CB(skb)->seq;
4330 end_seq = TCP_SKB_CB(skb)->end_seq;
4332 if (seq == TCP_SKB_CB(skb1)->end_seq) {
4335 if (!tcp_try_coalesce(sk, skb1, skb, &fragstolen)) {
4336 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4338 tcp_grow_window(sk, skb);
4339 kfree_skb_partial(skb, fragstolen);
4343 if (!tp->rx_opt.num_sacks ||
4344 tp->selective_acks[0].end_seq != seq)
4347 /* Common case: data arrive in order after hole. */
4348 tp->selective_acks[0].end_seq = end_seq;
4352 /* Find place to insert this segment. */
4354 if (!after(TCP_SKB_CB(skb1)->seq, seq))
4356 if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
4360 skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
4363 /* Do skb overlap to previous one? */
4364 if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4365 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4366 /* All the bits are present. Drop. */
4367 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4370 tcp_dsack_set(sk, seq, end_seq);
4373 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4374 /* Partial overlap. */
4375 tcp_dsack_set(sk, seq,
4376 TCP_SKB_CB(skb1)->end_seq);
4378 if (skb_queue_is_first(&tp->out_of_order_queue,
4382 skb1 = skb_queue_prev(
4383 &tp->out_of_order_queue,
4388 __skb_queue_head(&tp->out_of_order_queue, skb);
4390 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4392 /* And clean segments covered by new one as whole. */
4393 while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
4394 skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
4396 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4398 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4399 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4403 __skb_unlink(skb1, &tp->out_of_order_queue);
4404 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4405 TCP_SKB_CB(skb1)->end_seq);
4406 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4411 if (tcp_is_sack(tp))
4412 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4415 tcp_grow_window(sk, skb);
4416 skb_set_owner_r(skb, sk);
4420 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
4424 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4426 __skb_pull(skb, hdrlen);
4428 tcp_try_coalesce(sk, tail, skb, fragstolen)) ? 1 : 0;
4429 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
4431 __skb_queue_tail(&sk->sk_receive_queue, skb);
4432 skb_set_owner_r(skb, sk);
4437 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4439 struct sk_buff *skb;
4445 skb = alloc_skb(size, sk->sk_allocation);
4449 if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4452 if (memcpy_from_msg(skb_put(skb, size), msg, size))
4455 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4456 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4457 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4459 if (tcp_queue_rcv(sk, skb, 0, &fragstolen)) {
4460 WARN_ON_ONCE(fragstolen); /* should not happen */
4471 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4473 struct tcp_sock *tp = tcp_sk(sk);
4475 bool fragstolen = false;
4477 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4481 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
4483 tcp_ecn_accept_cwr(tp, skb);
4485 tp->rx_opt.dsack = 0;
4487 /* Queue data for delivery to the user.
4488 * Packets in sequence go to the receive queue.
4489 * Out of sequence packets to the out_of_order_queue.
4491 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4492 if (tcp_receive_window(tp) == 0)
4495 /* Ok. In sequence. In window. */
4496 if (tp->ucopy.task == current &&
4497 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4498 sock_owned_by_user(sk) && !tp->urg_data) {
4499 int chunk = min_t(unsigned int, skb->len,
4502 __set_current_state(TASK_RUNNING);
4505 if (!skb_copy_datagram_msg(skb, 0, tp->ucopy.msg, chunk)) {
4506 tp->ucopy.len -= chunk;
4507 tp->copied_seq += chunk;
4508 eaten = (chunk == skb->len);
4509 tcp_rcv_space_adjust(sk);
4517 if (skb_queue_len(&sk->sk_receive_queue) == 0)
4518 sk_forced_mem_schedule(sk, skb->truesize);
4519 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4522 eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
4524 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4526 tcp_event_data_recv(sk, skb);
4527 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4530 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4533 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4534 * gap in queue is filled.
4536 if (skb_queue_empty(&tp->out_of_order_queue))
4537 inet_csk(sk)->icsk_ack.pingpong = 0;
4540 if (tp->rx_opt.num_sacks)
4541 tcp_sack_remove(tp);
4543 tcp_fast_path_check(sk);
4546 kfree_skb_partial(skb, fragstolen);
4547 if (!sock_flag(sk, SOCK_DEAD))
4548 sk->sk_data_ready(sk);
4552 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4553 /* A retransmit, 2nd most common case. Force an immediate ack. */
4554 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4555 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4558 tcp_enter_quickack_mode(sk);
4559 inet_csk_schedule_ack(sk);
4565 /* Out of window. F.e. zero window probe. */
4566 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4569 tcp_enter_quickack_mode(sk);
4571 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4572 /* Partial packet, seq < rcv_next < end_seq */
4573 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4574 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4575 TCP_SKB_CB(skb)->end_seq);
4577 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4579 /* If window is closed, drop tail of packet. But after
4580 * remembering D-SACK for its head made in previous line.
4582 if (!tcp_receive_window(tp))
4587 tcp_data_queue_ofo(sk, skb);
4590 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4591 struct sk_buff_head *list)
4593 struct sk_buff *next = NULL;
4595 if (!skb_queue_is_last(list, skb))
4596 next = skb_queue_next(list, skb);
4598 __skb_unlink(skb, list);
4600 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4605 /* Collapse contiguous sequence of skbs head..tail with
4606 * sequence numbers start..end.
4608 * If tail is NULL, this means until the end of the list.
4610 * Segments with FIN/SYN are not collapsed (only because this
4614 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4615 struct sk_buff *head, struct sk_buff *tail,
4618 struct sk_buff *skb, *n;
4621 /* First, check that queue is collapsible and find
4622 * the point where collapsing can be useful. */
4626 skb_queue_walk_from_safe(list, skb, n) {
4629 /* No new bits? It is possible on ofo queue. */
4630 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4631 skb = tcp_collapse_one(sk, skb, list);
4637 /* The first skb to collapse is:
4639 * - bloated or contains data before "start" or
4640 * overlaps to the next one.
4642 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
4643 (tcp_win_from_space(skb->truesize) > skb->len ||
4644 before(TCP_SKB_CB(skb)->seq, start))) {
4645 end_of_skbs = false;
4649 if (!skb_queue_is_last(list, skb)) {
4650 struct sk_buff *next = skb_queue_next(list, skb);
4652 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
4653 end_of_skbs = false;
4658 /* Decided to skip this, advance start seq. */
4659 start = TCP_SKB_CB(skb)->end_seq;
4662 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4665 while (before(start, end)) {
4666 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
4667 struct sk_buff *nskb;
4669 nskb = alloc_skb(copy, GFP_ATOMIC);
4673 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4674 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4675 __skb_queue_before(list, skb, nskb);
4676 skb_set_owner_r(nskb, sk);
4678 /* Copy data, releasing collapsed skbs. */
4680 int offset = start - TCP_SKB_CB(skb)->seq;
4681 int size = TCP_SKB_CB(skb)->end_seq - start;
4685 size = min(copy, size);
4686 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4688 TCP_SKB_CB(nskb)->end_seq += size;
4692 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4693 skb = tcp_collapse_one(sk, skb, list);
4696 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4703 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4704 * and tcp_collapse() them until all the queue is collapsed.
4706 static void tcp_collapse_ofo_queue(struct sock *sk)
4708 struct tcp_sock *tp = tcp_sk(sk);
4709 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4710 struct sk_buff *head;
4716 start = TCP_SKB_CB(skb)->seq;
4717 end = TCP_SKB_CB(skb)->end_seq;
4721 struct sk_buff *next = NULL;
4723 if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
4724 next = skb_queue_next(&tp->out_of_order_queue, skb);
4727 /* Segment is terminated when we see gap or when
4728 * we are at the end of all the queue. */
4730 after(TCP_SKB_CB(skb)->seq, end) ||
4731 before(TCP_SKB_CB(skb)->end_seq, start)) {
4732 tcp_collapse(sk, &tp->out_of_order_queue,
4733 head, skb, start, end);
4737 /* Start new segment */
4738 start = TCP_SKB_CB(skb)->seq;
4739 end = TCP_SKB_CB(skb)->end_seq;
4741 if (before(TCP_SKB_CB(skb)->seq, start))
4742 start = TCP_SKB_CB(skb)->seq;
4743 if (after(TCP_SKB_CB(skb)->end_seq, end))
4744 end = TCP_SKB_CB(skb)->end_seq;
4750 * Purge the out-of-order queue.
4751 * Return true if queue was pruned.
4753 static bool tcp_prune_ofo_queue(struct sock *sk)
4755 struct tcp_sock *tp = tcp_sk(sk);
4758 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4759 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
4760 __skb_queue_purge(&tp->out_of_order_queue);
4762 /* Reset SACK state. A conforming SACK implementation will
4763 * do the same at a timeout based retransmit. When a connection
4764 * is in a sad state like this, we care only about integrity
4765 * of the connection not performance.
4767 if (tp->rx_opt.sack_ok)
4768 tcp_sack_reset(&tp->rx_opt);
4775 /* Reduce allocated memory if we can, trying to get
4776 * the socket within its memory limits again.
4778 * Return less than zero if we should start dropping frames
4779 * until the socket owning process reads some of the data
4780 * to stabilize the situation.
4782 static int tcp_prune_queue(struct sock *sk)
4784 struct tcp_sock *tp = tcp_sk(sk);
4786 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4788 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
4790 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4791 tcp_clamp_window(sk);
4792 else if (tcp_under_memory_pressure(sk))
4793 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4795 tcp_collapse_ofo_queue(sk);
4796 if (!skb_queue_empty(&sk->sk_receive_queue))
4797 tcp_collapse(sk, &sk->sk_receive_queue,
4798 skb_peek(&sk->sk_receive_queue),
4800 tp->copied_seq, tp->rcv_nxt);
4803 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4806 /* Collapsing did not help, destructive actions follow.
4807 * This must not ever occur. */
4809 tcp_prune_ofo_queue(sk);
4811 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4814 /* If we are really being abused, tell the caller to silently
4815 * drop receive data on the floor. It will get retransmitted
4816 * and hopefully then we'll have sufficient space.
4818 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
4820 /* Massive buffer overcommit. */
4825 static bool tcp_should_expand_sndbuf(const struct sock *sk)
4827 const struct tcp_sock *tp = tcp_sk(sk);
4829 /* If the user specified a specific send buffer setting, do
4832 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4835 /* If we are under global TCP memory pressure, do not expand. */
4836 if (tcp_under_memory_pressure(sk))
4839 /* If we are under soft global TCP memory pressure, do not expand. */
4840 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
4843 /* If we filled the congestion window, do not expand. */
4844 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
4850 /* When incoming ACK allowed to free some skb from write_queue,
4851 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4852 * on the exit from tcp input handler.
4854 * PROBLEM: sndbuf expansion does not work well with largesend.
4856 static void tcp_new_space(struct sock *sk)
4858 struct tcp_sock *tp = tcp_sk(sk);
4860 if (tcp_should_expand_sndbuf(sk)) {
4861 tcp_sndbuf_expand(sk);
4862 tp->snd_cwnd_stamp = tcp_time_stamp;
4865 sk->sk_write_space(sk);
4868 static void tcp_check_space(struct sock *sk)
4870 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4871 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4872 /* pairs with tcp_poll() */
4873 smp_mb__after_atomic();
4874 if (sk->sk_socket &&
4875 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4880 static inline void tcp_data_snd_check(struct sock *sk)
4882 tcp_push_pending_frames(sk);
4883 tcp_check_space(sk);
4887 * Check if sending an ack is needed.
4889 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4891 struct tcp_sock *tp = tcp_sk(sk);
4893 /* More than one full frame received... */
4894 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
4895 /* ... and right edge of window advances far enough.
4896 * (tcp_recvmsg() will send ACK otherwise). Or...
4898 __tcp_select_window(sk) >= tp->rcv_wnd) ||
4899 /* We ACK each frame or... */
4900 tcp_in_quickack_mode(sk) ||
4901 /* We have out of order data. */
4902 (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
4903 /* Then ack it now */
4906 /* Else, send delayed ack. */
4907 tcp_send_delayed_ack(sk);
4911 static inline void tcp_ack_snd_check(struct sock *sk)
4913 if (!inet_csk_ack_scheduled(sk)) {
4914 /* We sent a data segment already. */
4917 __tcp_ack_snd_check(sk, 1);
4921 * This routine is only called when we have urgent data
4922 * signaled. Its the 'slow' part of tcp_urg. It could be
4923 * moved inline now as tcp_urg is only called from one
4924 * place. We handle URGent data wrong. We have to - as
4925 * BSD still doesn't use the correction from RFC961.
4926 * For 1003.1g we should support a new option TCP_STDURG to permit
4927 * either form (or just set the sysctl tcp_stdurg).
4930 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
4932 struct tcp_sock *tp = tcp_sk(sk);
4933 u32 ptr = ntohs(th->urg_ptr);
4935 if (ptr && !sysctl_tcp_stdurg)
4937 ptr += ntohl(th->seq);
4939 /* Ignore urgent data that we've already seen and read. */
4940 if (after(tp->copied_seq, ptr))
4943 /* Do not replay urg ptr.
4945 * NOTE: interesting situation not covered by specs.
4946 * Misbehaving sender may send urg ptr, pointing to segment,
4947 * which we already have in ofo queue. We are not able to fetch
4948 * such data and will stay in TCP_URG_NOTYET until will be eaten
4949 * by recvmsg(). Seems, we are not obliged to handle such wicked
4950 * situations. But it is worth to think about possibility of some
4951 * DoSes using some hypothetical application level deadlock.
4953 if (before(ptr, tp->rcv_nxt))
4956 /* Do we already have a newer (or duplicate) urgent pointer? */
4957 if (tp->urg_data && !after(ptr, tp->urg_seq))
4960 /* Tell the world about our new urgent pointer. */
4963 /* We may be adding urgent data when the last byte read was
4964 * urgent. To do this requires some care. We cannot just ignore
4965 * tp->copied_seq since we would read the last urgent byte again
4966 * as data, nor can we alter copied_seq until this data arrives
4967 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4969 * NOTE. Double Dutch. Rendering to plain English: author of comment
4970 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4971 * and expect that both A and B disappear from stream. This is _wrong_.
4972 * Though this happens in BSD with high probability, this is occasional.
4973 * Any application relying on this is buggy. Note also, that fix "works"
4974 * only in this artificial test. Insert some normal data between A and B and we will
4975 * decline of BSD again. Verdict: it is better to remove to trap
4978 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
4979 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
4980 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4982 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
4983 __skb_unlink(skb, &sk->sk_receive_queue);
4988 tp->urg_data = TCP_URG_NOTYET;
4991 /* Disable header prediction. */
4995 /* This is the 'fast' part of urgent handling. */
4996 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
4998 struct tcp_sock *tp = tcp_sk(sk);
5000 /* Check if we get a new urgent pointer - normally not. */
5002 tcp_check_urg(sk, th);
5004 /* Do we wait for any urgent data? - normally not... */
5005 if (tp->urg_data == TCP_URG_NOTYET) {
5006 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5009 /* Is the urgent pointer pointing into this packet? */
5010 if (ptr < skb->len) {
5012 if (skb_copy_bits(skb, ptr, &tmp, 1))
5014 tp->urg_data = TCP_URG_VALID | tmp;
5015 if (!sock_flag(sk, SOCK_DEAD))
5016 sk->sk_data_ready(sk);
5021 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
5023 struct tcp_sock *tp = tcp_sk(sk);
5024 int chunk = skb->len - hlen;
5028 if (skb_csum_unnecessary(skb))
5029 err = skb_copy_datagram_msg(skb, hlen, tp->ucopy.msg, chunk);
5031 err = skb_copy_and_csum_datagram_msg(skb, hlen, tp->ucopy.msg);
5034 tp->ucopy.len -= chunk;
5035 tp->copied_seq += chunk;
5036 tcp_rcv_space_adjust(sk);
5043 static __sum16 __tcp_checksum_complete_user(struct sock *sk,
5044 struct sk_buff *skb)
5048 if (sock_owned_by_user(sk)) {
5050 result = __tcp_checksum_complete(skb);
5053 result = __tcp_checksum_complete(skb);
5058 static inline bool tcp_checksum_complete_user(struct sock *sk,
5059 struct sk_buff *skb)
5061 return !skb_csum_unnecessary(skb) &&
5062 __tcp_checksum_complete_user(sk, skb);
5065 /* Does PAWS and seqno based validation of an incoming segment, flags will
5066 * play significant role here.
5068 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5069 const struct tcphdr *th, int syn_inerr)
5071 struct tcp_sock *tp = tcp_sk(sk);
5073 /* RFC1323: H1. Apply PAWS check first. */
5074 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
5075 tcp_paws_discard(sk, skb)) {
5077 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5078 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5079 LINUX_MIB_TCPACKSKIPPEDPAWS,
5080 &tp->last_oow_ack_time))
5081 tcp_send_dupack(sk, skb);
5084 /* Reset is accepted even if it did not pass PAWS. */
5087 /* Step 1: check sequence number */
5088 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5089 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5090 * (RST) segments are validated by checking their SEQ-fields."
5091 * And page 69: "If an incoming segment is not acceptable,
5092 * an acknowledgment should be sent in reply (unless the RST
5093 * bit is set, if so drop the segment and return)".
5098 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5099 LINUX_MIB_TCPACKSKIPPEDSEQ,
5100 &tp->last_oow_ack_time))
5101 tcp_send_dupack(sk, skb);
5106 /* Step 2: check RST bit */
5109 * If sequence number exactly matches RCV.NXT, then
5110 * RESET the connection
5112 * Send a challenge ACK
5114 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt)
5117 tcp_send_challenge_ack(sk, skb);
5121 /* step 3: check security and precedence [ignored] */
5123 /* step 4: Check for a SYN
5124 * RFC 5961 4.2 : Send a challenge ack
5129 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5130 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5131 tcp_send_challenge_ack(sk, skb);
5143 * TCP receive function for the ESTABLISHED state.
5145 * It is split into a fast path and a slow path. The fast path is
5147 * - A zero window was announced from us - zero window probing
5148 * is only handled properly in the slow path.
5149 * - Out of order segments arrived.
5150 * - Urgent data is expected.
5151 * - There is no buffer space left
5152 * - Unexpected TCP flags/window values/header lengths are received
5153 * (detected by checking the TCP header against pred_flags)
5154 * - Data is sent in both directions. Fast path only supports pure senders
5155 * or pure receivers (this means either the sequence number or the ack
5156 * value must stay constant)
5157 * - Unexpected TCP option.
5159 * When these conditions are not satisfied it drops into a standard
5160 * receive procedure patterned after RFC793 to handle all cases.
5161 * The first three cases are guaranteed by proper pred_flags setting,
5162 * the rest is checked inline. Fast processing is turned on in
5163 * tcp_data_queue when everything is OK.
5165 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5166 const struct tcphdr *th, unsigned int len)
5168 struct tcp_sock *tp = tcp_sk(sk);
5170 if (unlikely(!sk->sk_rx_dst))
5171 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5173 * Header prediction.
5174 * The code loosely follows the one in the famous
5175 * "30 instruction TCP receive" Van Jacobson mail.
5177 * Van's trick is to deposit buffers into socket queue
5178 * on a device interrupt, to call tcp_recv function
5179 * on the receive process context and checksum and copy
5180 * the buffer to user space. smart...
5182 * Our current scheme is not silly either but we take the
5183 * extra cost of the net_bh soft interrupt processing...
5184 * We do checksum and copy also but from device to kernel.
5187 tp->rx_opt.saw_tstamp = 0;
5189 /* pred_flags is 0xS?10 << 16 + snd_wnd
5190 * if header_prediction is to be made
5191 * 'S' will always be tp->tcp_header_len >> 2
5192 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5193 * turn it off (when there are holes in the receive
5194 * space for instance)
5195 * PSH flag is ignored.
5198 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5199 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5200 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5201 int tcp_header_len = tp->tcp_header_len;
5203 /* Timestamp header prediction: tcp_header_len
5204 * is automatically equal to th->doff*4 due to pred_flags
5208 /* Check timestamp */
5209 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5210 /* No? Slow path! */
5211 if (!tcp_parse_aligned_timestamp(tp, th))
5214 /* If PAWS failed, check it more carefully in slow path */
5215 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5218 /* DO NOT update ts_recent here, if checksum fails
5219 * and timestamp was corrupted part, it will result
5220 * in a hung connection since we will drop all
5221 * future packets due to the PAWS test.
5225 if (len <= tcp_header_len) {
5226 /* Bulk data transfer: sender */
5227 if (len == tcp_header_len) {
5228 /* Predicted packet is in window by definition.
5229 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5230 * Hence, check seq<=rcv_wup reduces to:
5232 if (tcp_header_len ==
5233 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5234 tp->rcv_nxt == tp->rcv_wup)
5235 tcp_store_ts_recent(tp);
5237 /* We know that such packets are checksummed
5240 tcp_ack(sk, skb, 0);
5242 tcp_data_snd_check(sk);
5244 } else { /* Header too small */
5245 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5250 bool fragstolen = false;
5252 if (tp->ucopy.task == current &&
5253 tp->copied_seq == tp->rcv_nxt &&
5254 len - tcp_header_len <= tp->ucopy.len &&
5255 sock_owned_by_user(sk)) {
5256 __set_current_state(TASK_RUNNING);
5258 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) {
5259 /* Predicted packet is in window by definition.
5260 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5261 * Hence, check seq<=rcv_wup reduces to:
5263 if (tcp_header_len ==
5264 (sizeof(struct tcphdr) +
5265 TCPOLEN_TSTAMP_ALIGNED) &&
5266 tp->rcv_nxt == tp->rcv_wup)
5267 tcp_store_ts_recent(tp);
5269 tcp_rcv_rtt_measure_ts(sk, skb);
5271 __skb_pull(skb, tcp_header_len);
5272 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
5273 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
5278 if (tcp_checksum_complete_user(sk, skb))
5281 if ((int)skb->truesize > sk->sk_forward_alloc)
5284 /* Predicted packet is in window by definition.
5285 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5286 * Hence, check seq<=rcv_wup reduces to:
5288 if (tcp_header_len ==
5289 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5290 tp->rcv_nxt == tp->rcv_wup)
5291 tcp_store_ts_recent(tp);
5293 tcp_rcv_rtt_measure_ts(sk, skb);
5295 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
5297 /* Bulk data transfer: receiver */
5298 eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5302 tcp_event_data_recv(sk, skb);
5304 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5305 /* Well, only one small jumplet in fast path... */
5306 tcp_ack(sk, skb, FLAG_DATA);
5307 tcp_data_snd_check(sk);
5308 if (!inet_csk_ack_scheduled(sk))
5312 __tcp_ack_snd_check(sk, 0);
5315 kfree_skb_partial(skb, fragstolen);
5316 sk->sk_data_ready(sk);
5322 if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
5325 if (!th->ack && !th->rst && !th->syn)
5329 * Standard slow path.
5332 if (!tcp_validate_incoming(sk, skb, th, 1))
5336 if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5339 tcp_rcv_rtt_measure_ts(sk, skb);
5341 /* Process urgent data. */
5342 tcp_urg(sk, skb, th);
5344 /* step 7: process the segment text */
5345 tcp_data_queue(sk, skb);
5347 tcp_data_snd_check(sk);
5348 tcp_ack_snd_check(sk);
5352 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_CSUMERRORS);
5353 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5358 EXPORT_SYMBOL(tcp_rcv_established);
5360 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5362 struct tcp_sock *tp = tcp_sk(sk);
5363 struct inet_connection_sock *icsk = inet_csk(sk);
5365 tcp_set_state(sk, TCP_ESTABLISHED);
5368 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5369 security_inet_conn_established(sk, skb);
5372 /* Make sure socket is routed, for correct metrics. */
5373 icsk->icsk_af_ops->rebuild_header(sk);
5375 tcp_init_metrics(sk);
5377 tcp_init_congestion_control(sk);
5379 /* Prevent spurious tcp_cwnd_restart() on first data
5382 tp->lsndtime = tcp_time_stamp;
5384 tcp_init_buffer_space(sk);
5386 if (sock_flag(sk, SOCK_KEEPOPEN))
5387 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5389 if (!tp->rx_opt.snd_wscale)
5390 __tcp_fast_path_on(tp, tp->snd_wnd);
5394 if (!sock_flag(sk, SOCK_DEAD)) {
5395 sk->sk_state_change(sk);
5396 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5400 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5401 struct tcp_fastopen_cookie *cookie)
5403 struct tcp_sock *tp = tcp_sk(sk);
5404 struct sk_buff *data = tp->syn_data ? tcp_write_queue_head(sk) : NULL;
5405 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
5406 bool syn_drop = false;
5408 if (mss == tp->rx_opt.user_mss) {
5409 struct tcp_options_received opt;
5411 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5412 tcp_clear_options(&opt);
5413 opt.user_mss = opt.mss_clamp = 0;
5414 tcp_parse_options(synack, &opt, 0, NULL);
5415 mss = opt.mss_clamp;
5418 if (!tp->syn_fastopen) {
5419 /* Ignore an unsolicited cookie */
5421 } else if (tp->total_retrans) {
5422 /* SYN timed out and the SYN-ACK neither has a cookie nor
5423 * acknowledges data. Presumably the remote received only
5424 * the retransmitted (regular) SYNs: either the original
5425 * SYN-data or the corresponding SYN-ACK was dropped.
5427 syn_drop = (cookie->len < 0 && data);
5428 } else if (cookie->len < 0 && !tp->syn_data) {
5429 /* We requested a cookie but didn't get it. If we did not use
5430 * the (old) exp opt format then try so next time (try_exp=1).
5431 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5433 try_exp = tp->syn_fastopen_exp ? 2 : 1;
5436 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
5438 if (data) { /* Retransmit unacked data in SYN */
5439 tcp_for_write_queue_from(data, sk) {
5440 if (data == tcp_send_head(sk) ||
5441 __tcp_retransmit_skb(sk, data))
5445 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVEFAIL);
5448 tp->syn_data_acked = tp->syn_data;
5449 if (tp->syn_data_acked)
5450 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
5454 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5455 const struct tcphdr *th, unsigned int len)
5457 struct inet_connection_sock *icsk = inet_csk(sk);
5458 struct tcp_sock *tp = tcp_sk(sk);
5459 struct tcp_fastopen_cookie foc = { .len = -1 };
5460 int saved_clamp = tp->rx_opt.mss_clamp;
5462 tcp_parse_options(skb, &tp->rx_opt, 0, &foc);
5463 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
5464 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
5468 * "If the state is SYN-SENT then
5469 * first check the ACK bit
5470 * If the ACK bit is set
5471 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5472 * a reset (unless the RST bit is set, if so drop
5473 * the segment and return)"
5475 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5476 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
5477 goto reset_and_undo;
5479 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5480 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5482 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
5483 goto reset_and_undo;
5486 /* Now ACK is acceptable.
5488 * "If the RST bit is set
5489 * If the ACK was acceptable then signal the user "error:
5490 * connection reset", drop the segment, enter CLOSED state,
5491 * delete TCB, and return."
5500 * "fifth, if neither of the SYN or RST bits is set then
5501 * drop the segment and return."
5507 goto discard_and_undo;
5510 * "If the SYN bit is on ...
5511 * are acceptable then ...
5512 * (our SYN has been ACKed), change the connection
5513 * state to ESTABLISHED..."
5516 tcp_ecn_rcv_synack(tp, th);
5518 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5519 tcp_ack(sk, skb, FLAG_SLOWPATH);
5521 /* Ok.. it's good. Set up sequence numbers and
5522 * move to established.
5524 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5525 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5527 /* RFC1323: The window in SYN & SYN/ACK segments is
5530 tp->snd_wnd = ntohs(th->window);
5532 if (!tp->rx_opt.wscale_ok) {
5533 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5534 tp->window_clamp = min(tp->window_clamp, 65535U);
5537 if (tp->rx_opt.saw_tstamp) {
5538 tp->rx_opt.tstamp_ok = 1;
5539 tp->tcp_header_len =
5540 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5541 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5542 tcp_store_ts_recent(tp);
5544 tp->tcp_header_len = sizeof(struct tcphdr);
5547 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5548 tcp_enable_fack(tp);
5551 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5552 tcp_initialize_rcv_mss(sk);
5554 /* Remember, tcp_poll() does not lock socket!
5555 * Change state from SYN-SENT only after copied_seq
5556 * is initialized. */
5557 tp->copied_seq = tp->rcv_nxt;
5561 tcp_finish_connect(sk, skb);
5563 if ((tp->syn_fastopen || tp->syn_data) &&
5564 tcp_rcv_fastopen_synack(sk, skb, &foc))
5567 if (sk->sk_write_pending ||
5568 icsk->icsk_accept_queue.rskq_defer_accept ||
5569 icsk->icsk_ack.pingpong) {
5570 /* Save one ACK. Data will be ready after
5571 * several ticks, if write_pending is set.
5573 * It may be deleted, but with this feature tcpdumps
5574 * look so _wonderfully_ clever, that I was not able
5575 * to stand against the temptation 8) --ANK
5577 inet_csk_schedule_ack(sk);
5578 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5579 tcp_enter_quickack_mode(sk);
5580 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5581 TCP_DELACK_MAX, TCP_RTO_MAX);
5592 /* No ACK in the segment */
5596 * "If the RST bit is set
5598 * Otherwise (no ACK) drop the segment and return."
5601 goto discard_and_undo;
5605 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5606 tcp_paws_reject(&tp->rx_opt, 0))
5607 goto discard_and_undo;
5610 /* We see SYN without ACK. It is attempt of
5611 * simultaneous connect with crossed SYNs.
5612 * Particularly, it can be connect to self.
5614 tcp_set_state(sk, TCP_SYN_RECV);
5616 if (tp->rx_opt.saw_tstamp) {
5617 tp->rx_opt.tstamp_ok = 1;
5618 tcp_store_ts_recent(tp);
5619 tp->tcp_header_len =
5620 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5622 tp->tcp_header_len = sizeof(struct tcphdr);
5625 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5626 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5628 /* RFC1323: The window in SYN & SYN/ACK segments is
5631 tp->snd_wnd = ntohs(th->window);
5632 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5633 tp->max_window = tp->snd_wnd;
5635 tcp_ecn_rcv_syn(tp, th);
5638 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5639 tcp_initialize_rcv_mss(sk);
5641 tcp_send_synack(sk);
5643 /* Note, we could accept data and URG from this segment.
5644 * There are no obstacles to make this (except that we must
5645 * either change tcp_recvmsg() to prevent it from returning data
5646 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5648 * However, if we ignore data in ACKless segments sometimes,
5649 * we have no reasons to accept it sometimes.
5650 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5651 * is not flawless. So, discard packet for sanity.
5652 * Uncomment this return to process the data.
5659 /* "fifth, if neither of the SYN or RST bits is set then
5660 * drop the segment and return."
5664 tcp_clear_options(&tp->rx_opt);
5665 tp->rx_opt.mss_clamp = saved_clamp;
5669 tcp_clear_options(&tp->rx_opt);
5670 tp->rx_opt.mss_clamp = saved_clamp;
5675 * This function implements the receiving procedure of RFC 793 for
5676 * all states except ESTABLISHED and TIME_WAIT.
5677 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5678 * address independent.
5681 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5682 const struct tcphdr *th, unsigned int len)
5684 struct tcp_sock *tp = tcp_sk(sk);
5685 struct inet_connection_sock *icsk = inet_csk(sk);
5686 struct request_sock *req;
5691 tp->rx_opt.saw_tstamp = 0;
5693 switch (sk->sk_state) {
5707 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5710 /* Now we have several options: In theory there is
5711 * nothing else in the frame. KA9Q has an option to
5712 * send data with the syn, BSD accepts data with the
5713 * syn up to the [to be] advertised window and
5714 * Solaris 2.1 gives you a protocol error. For now
5715 * we just ignore it, that fits the spec precisely
5716 * and avoids incompatibilities. It would be nice in
5717 * future to drop through and process the data.
5719 * Now that TTCP is starting to be used we ought to
5721 * But, this leaves one open to an easy denial of
5722 * service attack, and SYN cookies can't defend
5723 * against this problem. So, we drop the data
5724 * in the interest of security over speed unless
5725 * it's still in use.
5733 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5737 /* Do step6 onward by hand. */
5738 tcp_urg(sk, skb, th);
5740 tcp_data_snd_check(sk);
5744 req = tp->fastopen_rsk;
5746 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
5747 sk->sk_state != TCP_FIN_WAIT1);
5749 if (!tcp_check_req(sk, skb, req, true))
5753 if (!th->ack && !th->rst && !th->syn)
5756 if (!tcp_validate_incoming(sk, skb, th, 0))
5759 /* step 5: check the ACK field */
5760 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
5761 FLAG_UPDATE_TS_RECENT) > 0;
5763 switch (sk->sk_state) {
5768 /* Once we leave TCP_SYN_RECV, we no longer need req
5772 synack_stamp = tcp_rsk(req)->snt_synack;
5773 tp->total_retrans = req->num_retrans;
5774 reqsk_fastopen_remove(sk, req, false);
5776 synack_stamp = tp->lsndtime;
5777 /* Make sure socket is routed, for correct metrics. */
5778 icsk->icsk_af_ops->rebuild_header(sk);
5779 tcp_init_congestion_control(sk);
5782 tp->copied_seq = tp->rcv_nxt;
5783 tcp_init_buffer_space(sk);
5786 tcp_set_state(sk, TCP_ESTABLISHED);
5787 sk->sk_state_change(sk);
5789 /* Note, that this wakeup is only for marginal crossed SYN case.
5790 * Passively open sockets are not waked up, because
5791 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5794 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5796 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5797 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
5798 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5799 tcp_synack_rtt_meas(sk, synack_stamp);
5801 if (tp->rx_opt.tstamp_ok)
5802 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5805 /* Re-arm the timer because data may have been sent out.
5806 * This is similar to the regular data transmission case
5807 * when new data has just been ack'ed.
5809 * (TFO) - we could try to be more aggressive and
5810 * retransmitting any data sooner based on when they
5815 tcp_init_metrics(sk);
5817 tcp_update_pacing_rate(sk);
5819 /* Prevent spurious tcp_cwnd_restart() on first data packet */
5820 tp->lsndtime = tcp_time_stamp;
5822 tcp_initialize_rcv_mss(sk);
5823 tcp_fast_path_on(tp);
5826 case TCP_FIN_WAIT1: {
5827 struct dst_entry *dst;
5830 /* If we enter the TCP_FIN_WAIT1 state and we are a
5831 * Fast Open socket and this is the first acceptable
5832 * ACK we have received, this would have acknowledged
5833 * our SYNACK so stop the SYNACK timer.
5836 /* Return RST if ack_seq is invalid.
5837 * Note that RFC793 only says to generate a
5838 * DUPACK for it but for TCP Fast Open it seems
5839 * better to treat this case like TCP_SYN_RECV
5844 /* We no longer need the request sock. */
5845 reqsk_fastopen_remove(sk, req, false);
5848 if (tp->snd_una != tp->write_seq)
5851 tcp_set_state(sk, TCP_FIN_WAIT2);
5852 sk->sk_shutdown |= SEND_SHUTDOWN;
5854 dst = __sk_dst_get(sk);
5858 if (!sock_flag(sk, SOCK_DEAD)) {
5859 /* Wake up lingering close() */
5860 sk->sk_state_change(sk);
5864 if (tp->linger2 < 0 ||
5865 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5866 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5868 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5872 tmo = tcp_fin_time(sk);
5873 if (tmo > TCP_TIMEWAIT_LEN) {
5874 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5875 } else if (th->fin || sock_owned_by_user(sk)) {
5876 /* Bad case. We could lose such FIN otherwise.
5877 * It is not a big problem, but it looks confusing
5878 * and not so rare event. We still can lose it now,
5879 * if it spins in bh_lock_sock(), but it is really
5882 inet_csk_reset_keepalive_timer(sk, tmo);
5884 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5891 if (tp->snd_una == tp->write_seq) {
5892 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5898 if (tp->snd_una == tp->write_seq) {
5899 tcp_update_metrics(sk);
5906 /* step 6: check the URG bit */
5907 tcp_urg(sk, skb, th);
5909 /* step 7: process the segment text */
5910 switch (sk->sk_state) {
5911 case TCP_CLOSE_WAIT:
5914 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5918 /* RFC 793 says to queue data in these states,
5919 * RFC 1122 says we MUST send a reset.
5920 * BSD 4.4 also does reset.
5922 if (sk->sk_shutdown & RCV_SHUTDOWN) {
5923 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5924 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5925 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5931 case TCP_ESTABLISHED:
5932 tcp_data_queue(sk, skb);
5937 /* tcp_data could move socket to TIME-WAIT */
5938 if (sk->sk_state != TCP_CLOSE) {
5939 tcp_data_snd_check(sk);
5940 tcp_ack_snd_check(sk);
5949 EXPORT_SYMBOL(tcp_rcv_state_process);
5951 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
5953 struct inet_request_sock *ireq = inet_rsk(req);
5955 if (family == AF_INET)
5956 net_dbg_ratelimited("drop open request from %pI4/%u\n",
5957 &ireq->ir_rmt_addr, port);
5958 #if IS_ENABLED(CONFIG_IPV6)
5959 else if (family == AF_INET6)
5960 net_dbg_ratelimited("drop open request from %pI6/%u\n",
5961 &ireq->ir_v6_rmt_addr, port);
5965 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
5967 * If we receive a SYN packet with these bits set, it means a
5968 * network is playing bad games with TOS bits. In order to
5969 * avoid possible false congestion notifications, we disable
5970 * TCP ECN negotiation.
5972 * Exception: tcp_ca wants ECN. This is required for DCTCP
5973 * congestion control: Linux DCTCP asserts ECT on all packets,
5974 * including SYN, which is most optimal solution; however,
5975 * others, such as FreeBSD do not.
5977 static void tcp_ecn_create_request(struct request_sock *req,
5978 const struct sk_buff *skb,
5979 const struct sock *listen_sk,
5980 const struct dst_entry *dst)
5982 const struct tcphdr *th = tcp_hdr(skb);
5983 const struct net *net = sock_net(listen_sk);
5984 bool th_ecn = th->ece && th->cwr;
5990 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
5991 ecn_ok = net->ipv4.sysctl_tcp_ecn || dst_feature(dst, RTAX_FEATURE_ECN);
5993 if ((!ect && ecn_ok) || tcp_ca_needs_ecn(listen_sk))
5994 inet_rsk(req)->ecn_ok = 1;
5997 static void tcp_openreq_init(struct request_sock *req,
5998 const struct tcp_options_received *rx_opt,
5999 struct sk_buff *skb, const struct sock *sk)
6001 struct inet_request_sock *ireq = inet_rsk(req);
6003 req->rcv_wnd = 0; /* So that tcp_send_synack() knows! */
6005 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6006 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
6007 tcp_rsk(req)->snt_synack = tcp_time_stamp;
6008 tcp_rsk(req)->last_oow_ack_time = 0;
6009 req->mss = rx_opt->mss_clamp;
6010 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6011 ireq->tstamp_ok = rx_opt->tstamp_ok;
6012 ireq->sack_ok = rx_opt->sack_ok;
6013 ireq->snd_wscale = rx_opt->snd_wscale;
6014 ireq->wscale_ok = rx_opt->wscale_ok;
6017 ireq->ir_rmt_port = tcp_hdr(skb)->source;
6018 ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6019 ireq->ir_mark = inet_request_mark(sk, skb);
6022 struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
6023 struct sock *sk_listener)
6025 struct request_sock *req = reqsk_alloc(ops, sk_listener);
6028 struct inet_request_sock *ireq = inet_rsk(req);
6030 kmemcheck_annotate_bitfield(ireq, flags);
6032 atomic64_set(&ireq->ir_cookie, 0);
6033 ireq->ireq_state = TCP_NEW_SYN_RECV;
6034 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
6035 ireq->ireq_family = sk_listener->sk_family;
6040 EXPORT_SYMBOL(inet_reqsk_alloc);
6043 * Return true if a syncookie should be sent
6045 static bool tcp_syn_flood_action(struct sock *sk,
6046 const struct sk_buff *skb,
6049 const char *msg = "Dropping request";
6050 bool want_cookie = false;
6051 struct listen_sock *lopt;
6053 #ifdef CONFIG_SYN_COOKIES
6054 if (sysctl_tcp_syncookies) {
6055 msg = "Sending cookies";
6057 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
6060 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
6062 lopt = inet_csk(sk)->icsk_accept_queue.listen_opt;
6063 if (!lopt->synflood_warned && sysctl_tcp_syncookies != 2) {
6064 lopt->synflood_warned = 1;
6065 pr_info("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6066 proto, ntohs(tcp_hdr(skb)->dest), msg);
6071 static void tcp_reqsk_record_syn(const struct sock *sk,
6072 struct request_sock *req,
6073 const struct sk_buff *skb)
6075 if (tcp_sk(sk)->save_syn) {
6076 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
6079 copy = kmalloc(len + sizeof(u32), GFP_ATOMIC);
6082 memcpy(©[1], skb_network_header(skb), len);
6083 req->saved_syn = copy;
6088 int tcp_conn_request(struct request_sock_ops *rsk_ops,
6089 const struct tcp_request_sock_ops *af_ops,
6090 struct sock *sk, struct sk_buff *skb)
6092 struct tcp_options_received tmp_opt;
6093 struct request_sock *req;
6094 struct tcp_sock *tp = tcp_sk(sk);
6095 struct dst_entry *dst = NULL;
6096 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
6097 bool want_cookie = false, fastopen;
6099 struct tcp_fastopen_cookie foc = { .len = -1 };
6103 /* TW buckets are converted to open requests without
6104 * limitations, they conserve resources and peer is
6105 * evidently real one.
6107 if ((sysctl_tcp_syncookies == 2 ||
6108 inet_csk_reqsk_queue_is_full(sk)) && !isn) {
6109 want_cookie = tcp_syn_flood_action(sk, skb, rsk_ops->slab_name);
6115 /* Accept backlog is full. If we have already queued enough
6116 * of warm entries in syn queue, drop request. It is better than
6117 * clogging syn queue with openreqs with exponentially increasing
6120 if (sk_acceptq_is_full(sk) && inet_csk_reqsk_queue_young(sk) > 1) {
6121 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6125 req = inet_reqsk_alloc(rsk_ops, sk);
6129 tcp_rsk(req)->af_specific = af_ops;
6131 tcp_clear_options(&tmp_opt);
6132 tmp_opt.mss_clamp = af_ops->mss_clamp;
6133 tmp_opt.user_mss = tp->rx_opt.user_mss;
6134 tcp_parse_options(skb, &tmp_opt, 0, want_cookie ? NULL : &foc);
6136 if (want_cookie && !tmp_opt.saw_tstamp)
6137 tcp_clear_options(&tmp_opt);
6139 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
6140 tcp_openreq_init(req, &tmp_opt, skb, sk);
6142 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6143 inet_rsk(req)->ir_iif = sk->sk_bound_dev_if;
6145 af_ops->init_req(req, sk, skb);
6147 if (security_inet_conn_request(sk, skb, req))
6150 if (!want_cookie && !isn) {
6151 /* VJ's idea. We save last timestamp seen
6152 * from the destination in peer table, when entering
6153 * state TIME-WAIT, and check against it before
6154 * accepting new connection request.
6156 * If "isn" is not zero, this request hit alive
6157 * timewait bucket, so that all the necessary checks
6158 * are made in the function processing timewait state.
6160 if (tcp_death_row.sysctl_tw_recycle) {
6163 dst = af_ops->route_req(sk, &fl, req, &strict);
6165 if (dst && strict &&
6166 !tcp_peer_is_proven(req, dst, true,
6167 tmp_opt.saw_tstamp)) {
6168 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSPASSIVEREJECTED);
6169 goto drop_and_release;
6172 /* Kill the following clause, if you dislike this way. */
6173 else if (!sysctl_tcp_syncookies &&
6174 (sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
6175 (sysctl_max_syn_backlog >> 2)) &&
6176 !tcp_peer_is_proven(req, dst, false,
6177 tmp_opt.saw_tstamp)) {
6178 /* Without syncookies last quarter of
6179 * backlog is filled with destinations,
6180 * proven to be alive.
6181 * It means that we continue to communicate
6182 * to destinations, already remembered
6183 * to the moment of synflood.
6185 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
6187 goto drop_and_release;
6190 isn = af_ops->init_seq(skb);
6193 dst = af_ops->route_req(sk, &fl, req, NULL);
6198 tcp_ecn_create_request(req, skb, sk, dst);
6201 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
6202 req->cookie_ts = tmp_opt.tstamp_ok;
6203 if (!tmp_opt.tstamp_ok)
6204 inet_rsk(req)->ecn_ok = 0;
6207 tcp_rsk(req)->snt_isn = isn;
6208 tcp_openreq_init_rwin(req, sk, dst);
6209 fastopen = !want_cookie &&
6210 tcp_try_fastopen(sk, skb, req, &foc, dst);
6211 err = af_ops->send_synack(sk, dst, &fl, req,
6212 skb_get_queue_mapping(skb), &foc);
6214 if (err || want_cookie)
6217 tcp_rsk(req)->tfo_listener = false;
6218 af_ops->queue_hash_add(sk, req, TCP_TIMEOUT_INIT);
6220 tcp_reqsk_record_syn(sk, req, skb);
6229 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_LISTENDROPS);
6232 EXPORT_SYMBOL(tcp_conn_request);