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>
73 #include <net/inet_common.h>
74 #include <linux/ipsec.h>
75 #include <asm/unaligned.h>
76 #include <linux/errqueue.h>
78 int sysctl_tcp_timestamps __read_mostly = 1;
79 int sysctl_tcp_window_scaling __read_mostly = 1;
80 int sysctl_tcp_sack __read_mostly = 1;
81 int sysctl_tcp_fack __read_mostly = 1;
82 int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
83 EXPORT_SYMBOL(sysctl_tcp_reordering);
84 int sysctl_tcp_dsack __read_mostly = 1;
85 int sysctl_tcp_app_win __read_mostly = 31;
86 int sysctl_tcp_adv_win_scale __read_mostly = 1;
87 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
89 /* rfc5961 challenge ack rate limiting */
90 int sysctl_tcp_challenge_ack_limit = 100;
92 int sysctl_tcp_stdurg __read_mostly;
93 int sysctl_tcp_rfc1337 __read_mostly;
94 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
95 int sysctl_tcp_frto __read_mostly = 2;
97 int sysctl_tcp_thin_dupack __read_mostly;
99 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
100 int sysctl_tcp_early_retrans __read_mostly = 3;
102 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
103 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
104 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
105 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
106 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
107 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
108 #define FLAG_ECE 0x40 /* ECE in this ACK */
109 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
110 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
111 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
112 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
113 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
114 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
116 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
117 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
118 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
119 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
121 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
122 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
124 /* Adapt the MSS value used to make delayed ack decision to the
127 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
129 struct inet_connection_sock *icsk = inet_csk(sk);
130 const unsigned int lss = icsk->icsk_ack.last_seg_size;
133 icsk->icsk_ack.last_seg_size = 0;
135 /* skb->len may jitter because of SACKs, even if peer
136 * sends good full-sized frames.
138 len = skb_shinfo(skb)->gso_size ? : skb->len;
139 if (len >= icsk->icsk_ack.rcv_mss) {
140 icsk->icsk_ack.rcv_mss = len;
142 /* Otherwise, we make more careful check taking into account,
143 * that SACKs block is variable.
145 * "len" is invariant segment length, including TCP header.
147 len += skb->data - skb_transport_header(skb);
148 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
149 /* If PSH is not set, packet should be
150 * full sized, provided peer TCP is not badly broken.
151 * This observation (if it is correct 8)) allows
152 * to handle super-low mtu links fairly.
154 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
155 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
156 /* Subtract also invariant (if peer is RFC compliant),
157 * tcp header plus fixed timestamp option length.
158 * Resulting "len" is MSS free of SACK jitter.
160 len -= tcp_sk(sk)->tcp_header_len;
161 icsk->icsk_ack.last_seg_size = len;
163 icsk->icsk_ack.rcv_mss = len;
167 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
168 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
169 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
173 static void tcp_incr_quickack(struct sock *sk)
175 struct inet_connection_sock *icsk = inet_csk(sk);
176 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
180 if (quickacks > icsk->icsk_ack.quick)
181 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
184 static void tcp_enter_quickack_mode(struct sock *sk)
186 struct inet_connection_sock *icsk = inet_csk(sk);
187 tcp_incr_quickack(sk);
188 icsk->icsk_ack.pingpong = 0;
189 icsk->icsk_ack.ato = TCP_ATO_MIN;
192 /* Send ACKs quickly, if "quick" count is not exhausted
193 * and the session is not interactive.
196 static inline bool tcp_in_quickack_mode(const struct sock *sk)
198 const struct inet_connection_sock *icsk = inet_csk(sk);
200 return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
203 static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
205 if (tp->ecn_flags & TCP_ECN_OK)
206 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
209 static void tcp_ecn_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
211 if (tcp_hdr(skb)->cwr)
212 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
215 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
217 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
220 static void __tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
222 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
223 case INET_ECN_NOT_ECT:
224 /* Funny extension: if ECT is not set on a segment,
225 * and we already seen ECT on a previous segment,
226 * it is probably a retransmit.
228 if (tp->ecn_flags & TCP_ECN_SEEN)
229 tcp_enter_quickack_mode((struct sock *)tp);
232 if (tcp_ca_needs_ecn((struct sock *)tp))
233 tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_IS_CE);
235 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
236 /* Better not delay acks, sender can have a very low cwnd */
237 tcp_enter_quickack_mode((struct sock *)tp);
238 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
240 tp->ecn_flags |= TCP_ECN_SEEN;
243 if (tcp_ca_needs_ecn((struct sock *)tp))
244 tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_NO_CE);
245 tp->ecn_flags |= TCP_ECN_SEEN;
250 static void tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
252 if (tp->ecn_flags & TCP_ECN_OK)
253 __tcp_ecn_check_ce(tp, skb);
256 static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
258 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
259 tp->ecn_flags &= ~TCP_ECN_OK;
262 static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
264 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
265 tp->ecn_flags &= ~TCP_ECN_OK;
268 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
270 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
275 /* Buffer size and advertised window tuning.
277 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
280 static void tcp_sndbuf_expand(struct sock *sk)
282 const struct tcp_sock *tp = tcp_sk(sk);
286 /* Worst case is non GSO/TSO : each frame consumes one skb
287 * and skb->head is kmalloced using power of two area of memory
289 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
291 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
293 per_mss = roundup_pow_of_two(per_mss) +
294 SKB_DATA_ALIGN(sizeof(struct sk_buff));
296 nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd);
297 nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
299 /* Fast Recovery (RFC 5681 3.2) :
300 * Cubic needs 1.7 factor, rounded to 2 to include
301 * extra cushion (application might react slowly to POLLOUT)
303 sndmem = 2 * nr_segs * per_mss;
305 if (sk->sk_sndbuf < sndmem)
306 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
309 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
311 * All tcp_full_space() is split to two parts: "network" buffer, allocated
312 * forward and advertised in receiver window (tp->rcv_wnd) and
313 * "application buffer", required to isolate scheduling/application
314 * latencies from network.
315 * window_clamp is maximal advertised window. It can be less than
316 * tcp_full_space(), in this case tcp_full_space() - window_clamp
317 * is reserved for "application" buffer. The less window_clamp is
318 * the smoother our behaviour from viewpoint of network, but the lower
319 * throughput and the higher sensitivity of the connection to losses. 8)
321 * rcv_ssthresh is more strict window_clamp used at "slow start"
322 * phase to predict further behaviour of this connection.
323 * It is used for two goals:
324 * - to enforce header prediction at sender, even when application
325 * requires some significant "application buffer". It is check #1.
326 * - to prevent pruning of receive queue because of misprediction
327 * of receiver window. Check #2.
329 * The scheme does not work when sender sends good segments opening
330 * window and then starts to feed us spaghetti. But it should work
331 * in common situations. Otherwise, we have to rely on queue collapsing.
334 /* Slow part of check#2. */
335 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
337 struct tcp_sock *tp = tcp_sk(sk);
339 int truesize = tcp_win_from_space(skb->truesize) >> 1;
340 int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
342 while (tp->rcv_ssthresh <= window) {
343 if (truesize <= skb->len)
344 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
352 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
354 struct tcp_sock *tp = tcp_sk(sk);
357 if (tp->rcv_ssthresh < tp->window_clamp &&
358 (int)tp->rcv_ssthresh < tcp_space(sk) &&
359 !sk_under_memory_pressure(sk)) {
362 /* Check #2. Increase window, if skb with such overhead
363 * will fit to rcvbuf in future.
365 if (tcp_win_from_space(skb->truesize) <= skb->len)
366 incr = 2 * tp->advmss;
368 incr = __tcp_grow_window(sk, skb);
371 incr = max_t(int, incr, 2 * skb->len);
372 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
374 inet_csk(sk)->icsk_ack.quick |= 1;
379 /* 3. Tuning rcvbuf, when connection enters established state. */
380 static void tcp_fixup_rcvbuf(struct sock *sk)
382 u32 mss = tcp_sk(sk)->advmss;
385 rcvmem = 2 * SKB_TRUESIZE(mss + MAX_TCP_HEADER) *
386 tcp_default_init_rwnd(mss);
388 /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
389 * Allow enough cushion so that sender is not limited by our window
391 if (sysctl_tcp_moderate_rcvbuf)
394 if (sk->sk_rcvbuf < rcvmem)
395 sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]);
398 /* 4. Try to fixup all. It is made immediately after connection enters
401 void tcp_init_buffer_space(struct sock *sk)
403 struct tcp_sock *tp = tcp_sk(sk);
406 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
407 tcp_fixup_rcvbuf(sk);
408 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
409 tcp_sndbuf_expand(sk);
411 tp->rcvq_space.space = tp->rcv_wnd;
412 tp->rcvq_space.time = tcp_time_stamp;
413 tp->rcvq_space.seq = tp->copied_seq;
415 maxwin = tcp_full_space(sk);
417 if (tp->window_clamp >= maxwin) {
418 tp->window_clamp = maxwin;
420 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
421 tp->window_clamp = max(maxwin -
422 (maxwin >> sysctl_tcp_app_win),
426 /* Force reservation of one segment. */
427 if (sysctl_tcp_app_win &&
428 tp->window_clamp > 2 * tp->advmss &&
429 tp->window_clamp + tp->advmss > maxwin)
430 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
432 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
433 tp->snd_cwnd_stamp = tcp_time_stamp;
436 /* 5. Recalculate window clamp after socket hit its memory bounds. */
437 static void tcp_clamp_window(struct sock *sk)
439 struct tcp_sock *tp = tcp_sk(sk);
440 struct inet_connection_sock *icsk = inet_csk(sk);
442 icsk->icsk_ack.quick = 0;
444 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
445 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
446 !sk_under_memory_pressure(sk) &&
447 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
448 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
451 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
452 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
455 /* Initialize RCV_MSS value.
456 * RCV_MSS is an our guess about MSS used by the peer.
457 * We haven't any direct information about the MSS.
458 * It's better to underestimate the RCV_MSS rather than overestimate.
459 * Overestimations make us ACKing less frequently than needed.
460 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
462 void tcp_initialize_rcv_mss(struct sock *sk)
464 const struct tcp_sock *tp = tcp_sk(sk);
465 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
467 hint = min(hint, tp->rcv_wnd / 2);
468 hint = min(hint, TCP_MSS_DEFAULT);
469 hint = max(hint, TCP_MIN_MSS);
471 inet_csk(sk)->icsk_ack.rcv_mss = hint;
473 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
475 /* Receiver "autotuning" code.
477 * The algorithm for RTT estimation w/o timestamps is based on
478 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
479 * <http://public.lanl.gov/radiant/pubs.html#DRS>
481 * More detail on this code can be found at
482 * <http://staff.psc.edu/jheffner/>,
483 * though this reference is out of date. A new paper
486 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
488 u32 new_sample = tp->rcv_rtt_est.rtt;
494 if (new_sample != 0) {
495 /* If we sample in larger samples in the non-timestamp
496 * case, we could grossly overestimate the RTT especially
497 * with chatty applications or bulk transfer apps which
498 * are stalled on filesystem I/O.
500 * Also, since we are only going for a minimum in the
501 * non-timestamp case, we do not smooth things out
502 * else with timestamps disabled convergence takes too
506 m -= (new_sample >> 3);
514 /* No previous measure. */
518 if (tp->rcv_rtt_est.rtt != new_sample)
519 tp->rcv_rtt_est.rtt = new_sample;
522 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
524 if (tp->rcv_rtt_est.time == 0)
526 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
528 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rcv_rtt_est.time, 1);
531 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
532 tp->rcv_rtt_est.time = tcp_time_stamp;
535 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
536 const struct sk_buff *skb)
538 struct tcp_sock *tp = tcp_sk(sk);
539 if (tp->rx_opt.rcv_tsecr &&
540 (TCP_SKB_CB(skb)->end_seq -
541 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
542 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
546 * This function should be called every time data is copied to user space.
547 * It calculates the appropriate TCP receive buffer space.
549 void tcp_rcv_space_adjust(struct sock *sk)
551 struct tcp_sock *tp = tcp_sk(sk);
555 time = tcp_time_stamp - tp->rcvq_space.time;
556 if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
559 /* Number of bytes copied to user in last RTT */
560 copied = tp->copied_seq - tp->rcvq_space.seq;
561 if (copied <= tp->rcvq_space.space)
565 * copied = bytes received in previous RTT, our base window
566 * To cope with packet losses, we need a 2x factor
567 * To cope with slow start, and sender growing its cwin by 100 %
568 * every RTT, we need a 4x factor, because the ACK we are sending
569 * now is for the next RTT, not the current one :
570 * <prev RTT . ><current RTT .. ><next RTT .... >
573 if (sysctl_tcp_moderate_rcvbuf &&
574 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
575 int rcvwin, rcvmem, rcvbuf;
577 /* minimal window to cope with packet losses, assuming
578 * steady state. Add some cushion because of small variations.
580 rcvwin = (copied << 1) + 16 * tp->advmss;
582 /* If rate increased by 25%,
583 * assume slow start, rcvwin = 3 * copied
584 * If rate increased by 50%,
585 * assume sender can use 2x growth, rcvwin = 4 * copied
588 tp->rcvq_space.space + (tp->rcvq_space.space >> 2)) {
590 tp->rcvq_space.space + (tp->rcvq_space.space >> 1))
593 rcvwin += (rcvwin >> 1);
596 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
597 while (tcp_win_from_space(rcvmem) < tp->advmss)
600 rcvbuf = min(rcvwin / tp->advmss * rcvmem, sysctl_tcp_rmem[2]);
601 if (rcvbuf > sk->sk_rcvbuf) {
602 sk->sk_rcvbuf = rcvbuf;
604 /* Make the window clamp follow along. */
605 tp->window_clamp = rcvwin;
608 tp->rcvq_space.space = copied;
611 tp->rcvq_space.seq = tp->copied_seq;
612 tp->rcvq_space.time = tcp_time_stamp;
615 /* There is something which you must keep in mind when you analyze the
616 * behavior of the tp->ato delayed ack timeout interval. When a
617 * connection starts up, we want to ack as quickly as possible. The
618 * problem is that "good" TCP's do slow start at the beginning of data
619 * transmission. The means that until we send the first few ACK's the
620 * sender will sit on his end and only queue most of his data, because
621 * he can only send snd_cwnd unacked packets at any given time. For
622 * each ACK we send, he increments snd_cwnd and transmits more of his
625 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
627 struct tcp_sock *tp = tcp_sk(sk);
628 struct inet_connection_sock *icsk = inet_csk(sk);
631 inet_csk_schedule_ack(sk);
633 tcp_measure_rcv_mss(sk, skb);
635 tcp_rcv_rtt_measure(tp);
637 now = tcp_time_stamp;
639 if (!icsk->icsk_ack.ato) {
640 /* The _first_ data packet received, initialize
641 * delayed ACK engine.
643 tcp_incr_quickack(sk);
644 icsk->icsk_ack.ato = TCP_ATO_MIN;
646 int m = now - icsk->icsk_ack.lrcvtime;
648 if (m <= TCP_ATO_MIN / 2) {
649 /* The fastest case is the first. */
650 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
651 } else if (m < icsk->icsk_ack.ato) {
652 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
653 if (icsk->icsk_ack.ato > icsk->icsk_rto)
654 icsk->icsk_ack.ato = icsk->icsk_rto;
655 } else if (m > icsk->icsk_rto) {
656 /* Too long gap. Apparently sender failed to
657 * restart window, so that we send ACKs quickly.
659 tcp_incr_quickack(sk);
663 icsk->icsk_ack.lrcvtime = now;
665 tcp_ecn_check_ce(tp, skb);
668 tcp_grow_window(sk, skb);
671 /* Called to compute a smoothed rtt estimate. The data fed to this
672 * routine either comes from timestamps, or from segments that were
673 * known _not_ to have been retransmitted [see Karn/Partridge
674 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
675 * piece by Van Jacobson.
676 * NOTE: the next three routines used to be one big routine.
677 * To save cycles in the RFC 1323 implementation it was better to break
678 * it up into three procedures. -- erics
680 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
682 struct tcp_sock *tp = tcp_sk(sk);
683 long m = mrtt_us; /* RTT */
684 u32 srtt = tp->srtt_us;
686 /* The following amusing code comes from Jacobson's
687 * article in SIGCOMM '88. Note that rtt and mdev
688 * are scaled versions of rtt and mean deviation.
689 * This is designed to be as fast as possible
690 * m stands for "measurement".
692 * On a 1990 paper the rto value is changed to:
693 * RTO = rtt + 4 * mdev
695 * Funny. This algorithm seems to be very broken.
696 * These formulae increase RTO, when it should be decreased, increase
697 * too slowly, when it should be increased quickly, decrease too quickly
698 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
699 * does not matter how to _calculate_ it. Seems, it was trap
700 * that VJ failed to avoid. 8)
703 m -= (srtt >> 3); /* m is now error in rtt est */
704 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
706 m = -m; /* m is now abs(error) */
707 m -= (tp->mdev_us >> 2); /* similar update on mdev */
708 /* This is similar to one of Eifel findings.
709 * Eifel blocks mdev updates when rtt decreases.
710 * This solution is a bit different: we use finer gain
711 * for mdev in this case (alpha*beta).
712 * Like Eifel it also prevents growth of rto,
713 * but also it limits too fast rto decreases,
714 * happening in pure Eifel.
719 m -= (tp->mdev_us >> 2); /* similar update on mdev */
721 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
722 if (tp->mdev_us > tp->mdev_max_us) {
723 tp->mdev_max_us = tp->mdev_us;
724 if (tp->mdev_max_us > tp->rttvar_us)
725 tp->rttvar_us = tp->mdev_max_us;
727 if (after(tp->snd_una, tp->rtt_seq)) {
728 if (tp->mdev_max_us < tp->rttvar_us)
729 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
730 tp->rtt_seq = tp->snd_nxt;
731 tp->mdev_max_us = tcp_rto_min_us(sk);
734 /* no previous measure. */
735 srtt = m << 3; /* take the measured time to be rtt */
736 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
737 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
738 tp->mdev_max_us = tp->rttvar_us;
739 tp->rtt_seq = tp->snd_nxt;
741 tp->srtt_us = max(1U, srtt);
744 /* Set the sk_pacing_rate to allow proper sizing of TSO packets.
745 * Note: TCP stack does not yet implement pacing.
746 * FQ packet scheduler can be used to implement cheap but effective
747 * TCP pacing, to smooth the burst on large writes when packets
748 * in flight is significantly lower than cwnd (or rwin)
750 static void tcp_update_pacing_rate(struct sock *sk)
752 const struct tcp_sock *tp = tcp_sk(sk);
755 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
756 rate = (u64)tp->mss_cache * 2 * (USEC_PER_SEC << 3);
758 rate *= max(tp->snd_cwnd, tp->packets_out);
760 if (likely(tp->srtt_us))
761 do_div(rate, tp->srtt_us);
763 /* ACCESS_ONCE() is needed because sch_fq fetches sk_pacing_rate
764 * without any lock. We want to make sure compiler wont store
765 * intermediate values in this location.
767 ACCESS_ONCE(sk->sk_pacing_rate) = min_t(u64, rate,
768 sk->sk_max_pacing_rate);
771 /* Calculate rto without backoff. This is the second half of Van Jacobson's
772 * routine referred to above.
774 static void tcp_set_rto(struct sock *sk)
776 const struct tcp_sock *tp = tcp_sk(sk);
777 /* Old crap is replaced with new one. 8)
780 * 1. If rtt variance happened to be less 50msec, it is hallucination.
781 * It cannot be less due to utterly erratic ACK generation made
782 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
783 * to do with delayed acks, because at cwnd>2 true delack timeout
784 * is invisible. Actually, Linux-2.4 also generates erratic
785 * ACKs in some circumstances.
787 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
789 /* 2. Fixups made earlier cannot be right.
790 * If we do not estimate RTO correctly without them,
791 * all the algo is pure shit and should be replaced
792 * with correct one. It is exactly, which we pretend to do.
795 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
796 * guarantees that rto is higher.
801 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
803 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
806 cwnd = TCP_INIT_CWND;
807 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
811 * Packet counting of FACK is based on in-order assumptions, therefore TCP
812 * disables it when reordering is detected
814 void tcp_disable_fack(struct tcp_sock *tp)
816 /* RFC3517 uses different metric in lost marker => reset on change */
818 tp->lost_skb_hint = NULL;
819 tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED;
822 /* Take a notice that peer is sending D-SACKs */
823 static void tcp_dsack_seen(struct tcp_sock *tp)
825 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
828 static void tcp_update_reordering(struct sock *sk, const int metric,
831 struct tcp_sock *tp = tcp_sk(sk);
832 if (metric > tp->reordering) {
835 tp->reordering = min(TCP_MAX_REORDERING, metric);
837 /* This exciting event is worth to be remembered. 8) */
839 mib_idx = LINUX_MIB_TCPTSREORDER;
840 else if (tcp_is_reno(tp))
841 mib_idx = LINUX_MIB_TCPRENOREORDER;
842 else if (tcp_is_fack(tp))
843 mib_idx = LINUX_MIB_TCPFACKREORDER;
845 mib_idx = LINUX_MIB_TCPSACKREORDER;
847 NET_INC_STATS_BH(sock_net(sk), mib_idx);
848 #if FASTRETRANS_DEBUG > 1
849 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
850 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
854 tp->undo_marker ? tp->undo_retrans : 0);
856 tcp_disable_fack(tp);
860 tcp_disable_early_retrans(tp);
863 /* This must be called before lost_out is incremented */
864 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
866 if ((tp->retransmit_skb_hint == NULL) ||
867 before(TCP_SKB_CB(skb)->seq,
868 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
869 tp->retransmit_skb_hint = skb;
872 after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
873 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
876 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
878 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
879 tcp_verify_retransmit_hint(tp, skb);
881 tp->lost_out += tcp_skb_pcount(skb);
882 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
886 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp,
889 tcp_verify_retransmit_hint(tp, skb);
891 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
892 tp->lost_out += tcp_skb_pcount(skb);
893 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
897 /* This procedure tags the retransmission queue when SACKs arrive.
899 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
900 * Packets in queue with these bits set are counted in variables
901 * sacked_out, retrans_out and lost_out, correspondingly.
903 * Valid combinations are:
904 * Tag InFlight Description
905 * 0 1 - orig segment is in flight.
906 * S 0 - nothing flies, orig reached receiver.
907 * L 0 - nothing flies, orig lost by net.
908 * R 2 - both orig and retransmit are in flight.
909 * L|R 1 - orig is lost, retransmit is in flight.
910 * S|R 1 - orig reached receiver, retrans is still in flight.
911 * (L|S|R is logically valid, it could occur when L|R is sacked,
912 * but it is equivalent to plain S and code short-curcuits it to S.
913 * L|S is logically invalid, it would mean -1 packet in flight 8))
915 * These 6 states form finite state machine, controlled by the following events:
916 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
917 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
918 * 3. Loss detection event of two flavors:
919 * A. Scoreboard estimator decided the packet is lost.
920 * A'. Reno "three dupacks" marks head of queue lost.
921 * A''. Its FACK modification, head until snd.fack is lost.
922 * B. SACK arrives sacking SND.NXT at the moment, when the
923 * segment was retransmitted.
924 * 4. D-SACK added new rule: D-SACK changes any tag to S.
926 * It is pleasant to note, that state diagram turns out to be commutative,
927 * so that we are allowed not to be bothered by order of our actions,
928 * when multiple events arrive simultaneously. (see the function below).
930 * Reordering detection.
931 * --------------------
932 * Reordering metric is maximal distance, which a packet can be displaced
933 * in packet stream. With SACKs we can estimate it:
935 * 1. SACK fills old hole and the corresponding segment was not
936 * ever retransmitted -> reordering. Alas, we cannot use it
937 * when segment was retransmitted.
938 * 2. The last flaw is solved with D-SACK. D-SACK arrives
939 * for retransmitted and already SACKed segment -> reordering..
940 * Both of these heuristics are not used in Loss state, when we cannot
941 * account for retransmits accurately.
943 * SACK block validation.
944 * ----------------------
946 * SACK block range validation checks that the received SACK block fits to
947 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
948 * Note that SND.UNA is not included to the range though being valid because
949 * it means that the receiver is rather inconsistent with itself reporting
950 * SACK reneging when it should advance SND.UNA. Such SACK block this is
951 * perfectly valid, however, in light of RFC2018 which explicitly states
952 * that "SACK block MUST reflect the newest segment. Even if the newest
953 * segment is going to be discarded ...", not that it looks very clever
954 * in case of head skb. Due to potentional receiver driven attacks, we
955 * choose to avoid immediate execution of a walk in write queue due to
956 * reneging and defer head skb's loss recovery to standard loss recovery
957 * procedure that will eventually trigger (nothing forbids us doing this).
959 * Implements also blockage to start_seq wrap-around. Problem lies in the
960 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
961 * there's no guarantee that it will be before snd_nxt (n). The problem
962 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
965 * <- outs wnd -> <- wrapzone ->
966 * u e n u_w e_w s n_w
968 * |<------------+------+----- TCP seqno space --------------+---------->|
969 * ...-- <2^31 ->| |<--------...
970 * ...---- >2^31 ------>| |<--------...
972 * Current code wouldn't be vulnerable but it's better still to discard such
973 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
974 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
975 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
976 * equal to the ideal case (infinite seqno space without wrap caused issues).
978 * With D-SACK the lower bound is extended to cover sequence space below
979 * SND.UNA down to undo_marker, which is the last point of interest. Yet
980 * again, D-SACK block must not to go across snd_una (for the same reason as
981 * for the normal SACK blocks, explained above). But there all simplicity
982 * ends, TCP might receive valid D-SACKs below that. As long as they reside
983 * fully below undo_marker they do not affect behavior in anyway and can
984 * therefore be safely ignored. In rare cases (which are more or less
985 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
986 * fragmentation and packet reordering past skb's retransmission. To consider
987 * them correctly, the acceptable range must be extended even more though
988 * the exact amount is rather hard to quantify. However, tp->max_window can
989 * be used as an exaggerated estimate.
991 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
992 u32 start_seq, u32 end_seq)
994 /* Too far in future, or reversed (interpretation is ambiguous) */
995 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
998 /* Nasty start_seq wrap-around check (see comments above) */
999 if (!before(start_seq, tp->snd_nxt))
1002 /* In outstanding window? ...This is valid exit for D-SACKs too.
1003 * start_seq == snd_una is non-sensical (see comments above)
1005 if (after(start_seq, tp->snd_una))
1008 if (!is_dsack || !tp->undo_marker)
1011 /* ...Then it's D-SACK, and must reside below snd_una completely */
1012 if (after(end_seq, tp->snd_una))
1015 if (!before(start_seq, tp->undo_marker))
1019 if (!after(end_seq, tp->undo_marker))
1022 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1023 * start_seq < undo_marker and end_seq >= undo_marker.
1025 return !before(start_seq, end_seq - tp->max_window);
1028 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1029 * Event "B". Later note: FACK people cheated me again 8), we have to account
1030 * for reordering! Ugly, but should help.
1032 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1033 * less than what is now known to be received by the other end (derived from
1034 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1035 * retransmitted skbs to avoid some costly processing per ACKs.
1037 static void tcp_mark_lost_retrans(struct sock *sk)
1039 const struct inet_connection_sock *icsk = inet_csk(sk);
1040 struct tcp_sock *tp = tcp_sk(sk);
1041 struct sk_buff *skb;
1043 u32 new_low_seq = tp->snd_nxt;
1044 u32 received_upto = tcp_highest_sack_seq(tp);
1046 if (!tcp_is_fack(tp) || !tp->retrans_out ||
1047 !after(received_upto, tp->lost_retrans_low) ||
1048 icsk->icsk_ca_state != TCP_CA_Recovery)
1051 tcp_for_write_queue(skb, sk) {
1052 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1054 if (skb == tcp_send_head(sk))
1056 if (cnt == tp->retrans_out)
1058 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1061 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1064 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1065 * constraint here (see above) but figuring out that at
1066 * least tp->reordering SACK blocks reside between ack_seq
1067 * and received_upto is not easy task to do cheaply with
1068 * the available datastructures.
1070 * Whether FACK should check here for tp->reordering segs
1071 * in-between one could argue for either way (it would be
1072 * rather simple to implement as we could count fack_count
1073 * during the walk and do tp->fackets_out - fack_count).
1075 if (after(received_upto, ack_seq)) {
1076 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1077 tp->retrans_out -= tcp_skb_pcount(skb);
1079 tcp_skb_mark_lost_uncond_verify(tp, skb);
1080 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
1082 if (before(ack_seq, new_low_seq))
1083 new_low_seq = ack_seq;
1084 cnt += tcp_skb_pcount(skb);
1088 if (tp->retrans_out)
1089 tp->lost_retrans_low = new_low_seq;
1092 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1093 struct tcp_sack_block_wire *sp, int num_sacks,
1096 struct tcp_sock *tp = tcp_sk(sk);
1097 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1098 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1099 bool dup_sack = false;
1101 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1104 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1105 } else if (num_sacks > 1) {
1106 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1107 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1109 if (!after(end_seq_0, end_seq_1) &&
1110 !before(start_seq_0, start_seq_1)) {
1113 NET_INC_STATS_BH(sock_net(sk),
1114 LINUX_MIB_TCPDSACKOFORECV);
1118 /* D-SACK for already forgotten data... Do dumb counting. */
1119 if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 &&
1120 !after(end_seq_0, prior_snd_una) &&
1121 after(end_seq_0, tp->undo_marker))
1127 struct tcp_sacktag_state {
1130 long rtt_us; /* RTT measured by SACKing never-retransmitted data */
1134 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1135 * the incoming SACK may not exactly match but we can find smaller MSS
1136 * aligned portion of it that matches. Therefore we might need to fragment
1137 * which may fail and creates some hassle (caller must handle error case
1140 * FIXME: this could be merged to shift decision code
1142 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1143 u32 start_seq, u32 end_seq)
1147 unsigned int pkt_len;
1150 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1151 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1153 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1154 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1155 mss = tcp_skb_mss(skb);
1156 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1159 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1163 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1168 /* Round if necessary so that SACKs cover only full MSSes
1169 * and/or the remaining small portion (if present)
1171 if (pkt_len > mss) {
1172 unsigned int new_len = (pkt_len / mss) * mss;
1173 if (!in_sack && new_len < pkt_len) {
1175 if (new_len >= skb->len)
1180 err = tcp_fragment(sk, skb, pkt_len, mss, GFP_ATOMIC);
1188 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1189 static u8 tcp_sacktag_one(struct sock *sk,
1190 struct tcp_sacktag_state *state, u8 sacked,
1191 u32 start_seq, u32 end_seq,
1192 int dup_sack, int pcount,
1193 const struct skb_mstamp *xmit_time)
1195 struct tcp_sock *tp = tcp_sk(sk);
1196 int fack_count = state->fack_count;
1198 /* Account D-SACK for retransmitted packet. */
1199 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1200 if (tp->undo_marker && tp->undo_retrans > 0 &&
1201 after(end_seq, tp->undo_marker))
1203 if (sacked & TCPCB_SACKED_ACKED)
1204 state->reord = min(fack_count, state->reord);
1207 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1208 if (!after(end_seq, tp->snd_una))
1211 if (!(sacked & TCPCB_SACKED_ACKED)) {
1212 if (sacked & TCPCB_SACKED_RETRANS) {
1213 /* If the segment is not tagged as lost,
1214 * we do not clear RETRANS, believing
1215 * that retransmission is still in flight.
1217 if (sacked & TCPCB_LOST) {
1218 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1219 tp->lost_out -= pcount;
1220 tp->retrans_out -= pcount;
1223 if (!(sacked & TCPCB_RETRANS)) {
1224 /* New sack for not retransmitted frame,
1225 * which was in hole. It is reordering.
1227 if (before(start_seq,
1228 tcp_highest_sack_seq(tp)))
1229 state->reord = min(fack_count,
1231 if (!after(end_seq, tp->high_seq))
1232 state->flag |= FLAG_ORIG_SACK_ACKED;
1233 /* Pick the earliest sequence sacked for RTT */
1234 if (state->rtt_us < 0) {
1235 struct skb_mstamp now;
1237 skb_mstamp_get(&now);
1238 state->rtt_us = skb_mstamp_us_delta(&now,
1243 if (sacked & TCPCB_LOST) {
1244 sacked &= ~TCPCB_LOST;
1245 tp->lost_out -= pcount;
1249 sacked |= TCPCB_SACKED_ACKED;
1250 state->flag |= FLAG_DATA_SACKED;
1251 tp->sacked_out += pcount;
1253 fack_count += pcount;
1255 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1256 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
1257 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1258 tp->lost_cnt_hint += pcount;
1260 if (fack_count > tp->fackets_out)
1261 tp->fackets_out = fack_count;
1264 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1265 * frames and clear it. undo_retrans is decreased above, L|R frames
1266 * are accounted above as well.
1268 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1269 sacked &= ~TCPCB_SACKED_RETRANS;
1270 tp->retrans_out -= pcount;
1276 /* Shift newly-SACKed bytes from this skb to the immediately previous
1277 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1279 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1280 struct tcp_sacktag_state *state,
1281 unsigned int pcount, int shifted, int mss,
1284 struct tcp_sock *tp = tcp_sk(sk);
1285 struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1286 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1287 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1291 /* Adjust counters and hints for the newly sacked sequence
1292 * range but discard the return value since prev is already
1293 * marked. We must tag the range first because the seq
1294 * advancement below implicitly advances
1295 * tcp_highest_sack_seq() when skb is highest_sack.
1297 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1298 start_seq, end_seq, dup_sack, pcount,
1301 if (skb == tp->lost_skb_hint)
1302 tp->lost_cnt_hint += pcount;
1304 TCP_SKB_CB(prev)->end_seq += shifted;
1305 TCP_SKB_CB(skb)->seq += shifted;
1307 tcp_skb_pcount_add(prev, pcount);
1308 BUG_ON(tcp_skb_pcount(skb) < pcount);
1309 tcp_skb_pcount_add(skb, -pcount);
1311 /* When we're adding to gso_segs == 1, gso_size will be zero,
1312 * in theory this shouldn't be necessary but as long as DSACK
1313 * code can come after this skb later on it's better to keep
1314 * setting gso_size to something.
1316 if (!skb_shinfo(prev)->gso_size) {
1317 skb_shinfo(prev)->gso_size = mss;
1318 skb_shinfo(prev)->gso_type = sk->sk_gso_type;
1321 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1322 if (tcp_skb_pcount(skb) <= 1) {
1323 skb_shinfo(skb)->gso_size = 0;
1324 skb_shinfo(skb)->gso_type = 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))
1535 if ((next_dup != NULL) &&
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,
1614 if (next_dup == NULL)
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, long *sack_rtt_us)
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 tcp_sacktag_state state;
1643 struct sk_buff *skb;
1644 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1646 bool found_dup_sack = false;
1648 int first_sack_index;
1651 state.reord = tp->packets_out;
1654 if (!tp->sacked_out) {
1655 if (WARN_ON(tp->fackets_out))
1656 tp->fackets_out = 0;
1657 tcp_highest_sack_reset(sk);
1660 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1661 num_sacks, prior_snd_una);
1663 state.flag |= FLAG_DSACKING_ACK;
1665 /* Eliminate too old ACKs, but take into
1666 * account more or less fresh ones, they can
1667 * contain valid SACK info.
1669 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1672 if (!tp->packets_out)
1676 first_sack_index = 0;
1677 for (i = 0; i < num_sacks; i++) {
1678 bool dup_sack = !i && found_dup_sack;
1680 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1681 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1683 if (!tcp_is_sackblock_valid(tp, dup_sack,
1684 sp[used_sacks].start_seq,
1685 sp[used_sacks].end_seq)) {
1689 if (!tp->undo_marker)
1690 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1692 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1694 /* Don't count olds caused by ACK reordering */
1695 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1696 !after(sp[used_sacks].end_seq, tp->snd_una))
1698 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1701 NET_INC_STATS_BH(sock_net(sk), mib_idx);
1703 first_sack_index = -1;
1707 /* Ignore very old stuff early */
1708 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1714 /* order SACK blocks to allow in order walk of the retrans queue */
1715 for (i = used_sacks - 1; i > 0; i--) {
1716 for (j = 0; j < i; j++) {
1717 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1718 swap(sp[j], sp[j + 1]);
1720 /* Track where the first SACK block goes to */
1721 if (j == first_sack_index)
1722 first_sack_index = j + 1;
1727 skb = tcp_write_queue_head(sk);
1728 state.fack_count = 0;
1731 if (!tp->sacked_out) {
1732 /* It's already past, so skip checking against it */
1733 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1735 cache = tp->recv_sack_cache;
1736 /* Skip empty blocks in at head of the cache */
1737 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1742 while (i < used_sacks) {
1743 u32 start_seq = sp[i].start_seq;
1744 u32 end_seq = sp[i].end_seq;
1745 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1746 struct tcp_sack_block *next_dup = NULL;
1748 if (found_dup_sack && ((i + 1) == first_sack_index))
1749 next_dup = &sp[i + 1];
1751 /* Skip too early cached blocks */
1752 while (tcp_sack_cache_ok(tp, cache) &&
1753 !before(start_seq, cache->end_seq))
1756 /* Can skip some work by looking recv_sack_cache? */
1757 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1758 after(end_seq, cache->start_seq)) {
1761 if (before(start_seq, cache->start_seq)) {
1762 skb = tcp_sacktag_skip(skb, sk, &state,
1764 skb = tcp_sacktag_walk(skb, sk, next_dup,
1771 /* Rest of the block already fully processed? */
1772 if (!after(end_seq, cache->end_seq))
1775 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1779 /* ...tail remains todo... */
1780 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1781 /* ...but better entrypoint exists! */
1782 skb = tcp_highest_sack(sk);
1785 state.fack_count = tp->fackets_out;
1790 skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq);
1791 /* Check overlap against next cached too (past this one already) */
1796 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1797 skb = tcp_highest_sack(sk);
1800 state.fack_count = tp->fackets_out;
1802 skb = tcp_sacktag_skip(skb, sk, &state, start_seq);
1805 skb = tcp_sacktag_walk(skb, sk, next_dup, &state,
1806 start_seq, end_seq, dup_sack);
1812 /* Clear the head of the cache sack blocks so we can skip it next time */
1813 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1814 tp->recv_sack_cache[i].start_seq = 0;
1815 tp->recv_sack_cache[i].end_seq = 0;
1817 for (j = 0; j < used_sacks; j++)
1818 tp->recv_sack_cache[i++] = sp[j];
1820 tcp_mark_lost_retrans(sk);
1822 tcp_verify_left_out(tp);
1824 if ((state.reord < tp->fackets_out) &&
1825 ((inet_csk(sk)->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker))
1826 tcp_update_reordering(sk, tp->fackets_out - state.reord, 0);
1830 #if FASTRETRANS_DEBUG > 0
1831 WARN_ON((int)tp->sacked_out < 0);
1832 WARN_ON((int)tp->lost_out < 0);
1833 WARN_ON((int)tp->retrans_out < 0);
1834 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1836 *sack_rtt_us = state.rtt_us;
1840 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1841 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1843 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1847 holes = max(tp->lost_out, 1U);
1848 holes = min(holes, tp->packets_out);
1850 if ((tp->sacked_out + holes) > tp->packets_out) {
1851 tp->sacked_out = tp->packets_out - holes;
1857 /* If we receive more dupacks than we expected counting segments
1858 * in assumption of absent reordering, interpret this as reordering.
1859 * The only another reason could be bug in receiver TCP.
1861 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1863 struct tcp_sock *tp = tcp_sk(sk);
1864 if (tcp_limit_reno_sacked(tp))
1865 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1868 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1870 static void tcp_add_reno_sack(struct sock *sk)
1872 struct tcp_sock *tp = tcp_sk(sk);
1874 tcp_check_reno_reordering(sk, 0);
1875 tcp_verify_left_out(tp);
1878 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1880 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1882 struct tcp_sock *tp = tcp_sk(sk);
1885 /* One ACK acked hole. The rest eat duplicate ACKs. */
1886 if (acked - 1 >= tp->sacked_out)
1889 tp->sacked_out -= acked - 1;
1891 tcp_check_reno_reordering(sk, acked);
1892 tcp_verify_left_out(tp);
1895 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1900 void tcp_clear_retrans(struct tcp_sock *tp)
1902 tp->retrans_out = 0;
1904 tp->undo_marker = 0;
1905 tp->undo_retrans = -1;
1906 tp->fackets_out = 0;
1910 static inline void tcp_init_undo(struct tcp_sock *tp)
1912 tp->undo_marker = tp->snd_una;
1913 /* Retransmission still in flight may cause DSACKs later. */
1914 tp->undo_retrans = tp->retrans_out ? : -1;
1917 /* Enter Loss state. If we detect SACK reneging, forget all SACK information
1918 * and reset tags completely, otherwise preserve SACKs. If receiver
1919 * dropped its ofo queue, we will know this due to reneging detection.
1921 void tcp_enter_loss(struct sock *sk)
1923 const struct inet_connection_sock *icsk = inet_csk(sk);
1924 struct tcp_sock *tp = tcp_sk(sk);
1925 struct sk_buff *skb;
1926 bool new_recovery = false;
1927 bool is_reneg; /* is receiver reneging on SACKs? */
1929 /* Reduce ssthresh if it has not yet been made inside this window. */
1930 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1931 !after(tp->high_seq, tp->snd_una) ||
1932 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1933 new_recovery = true;
1934 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1935 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1936 tcp_ca_event(sk, CA_EVENT_LOSS);
1940 tp->snd_cwnd_cnt = 0;
1941 tp->snd_cwnd_stamp = tcp_time_stamp;
1943 tp->retrans_out = 0;
1946 if (tcp_is_reno(tp))
1947 tcp_reset_reno_sack(tp);
1949 skb = tcp_write_queue_head(sk);
1950 is_reneg = skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED);
1952 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
1954 tp->fackets_out = 0;
1956 tcp_clear_all_retrans_hints(tp);
1958 tcp_for_write_queue(skb, sk) {
1959 if (skb == tcp_send_head(sk))
1962 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1963 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || is_reneg) {
1964 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1965 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1966 tp->lost_out += tcp_skb_pcount(skb);
1967 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
1970 tcp_verify_left_out(tp);
1972 /* Timeout in disordered state after receiving substantial DUPACKs
1973 * suggests that the degree of reordering is over-estimated.
1975 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
1976 tp->sacked_out >= sysctl_tcp_reordering)
1977 tp->reordering = min_t(unsigned int, tp->reordering,
1978 sysctl_tcp_reordering);
1979 tcp_set_ca_state(sk, TCP_CA_Loss);
1980 tp->high_seq = tp->snd_nxt;
1981 tcp_ecn_queue_cwr(tp);
1983 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1984 * loss recovery is underway except recurring timeout(s) on
1985 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1987 tp->frto = sysctl_tcp_frto &&
1988 (new_recovery || icsk->icsk_retransmits) &&
1989 !inet_csk(sk)->icsk_mtup.probe_size;
1992 /* If ACK arrived pointing to a remembered SACK, it means that our
1993 * remembered SACKs do not reflect real state of receiver i.e.
1994 * receiver _host_ is heavily congested (or buggy).
1996 * To avoid big spurious retransmission bursts due to transient SACK
1997 * scoreboard oddities that look like reneging, we give the receiver a
1998 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
1999 * restore sanity to the SACK scoreboard. If the apparent reneging
2000 * persists until this RTO then we'll clear the SACK scoreboard.
2002 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
2004 if (flag & FLAG_SACK_RENEGING) {
2005 struct tcp_sock *tp = tcp_sk(sk);
2006 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2007 msecs_to_jiffies(10));
2009 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2010 delay, TCP_RTO_MAX);
2016 static inline int tcp_fackets_out(const struct tcp_sock *tp)
2018 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
2021 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2022 * counter when SACK is enabled (without SACK, sacked_out is used for
2025 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2026 * segments up to the highest received SACK block so far and holes in
2029 * With reordering, holes may still be in flight, so RFC3517 recovery
2030 * uses pure sacked_out (total number of SACKed segments) even though
2031 * it violates the RFC that uses duplicate ACKs, often these are equal
2032 * but when e.g. out-of-window ACKs or packet duplication occurs,
2033 * they differ. Since neither occurs due to loss, TCP should really
2036 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2038 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2041 static bool tcp_pause_early_retransmit(struct sock *sk, int flag)
2043 struct tcp_sock *tp = tcp_sk(sk);
2044 unsigned long delay;
2046 /* Delay early retransmit and entering fast recovery for
2047 * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
2048 * available, or RTO is scheduled to fire first.
2050 if (sysctl_tcp_early_retrans < 2 || sysctl_tcp_early_retrans > 3 ||
2051 (flag & FLAG_ECE) || !tp->srtt_us)
2054 delay = max(usecs_to_jiffies(tp->srtt_us >> 5),
2055 msecs_to_jiffies(2));
2057 if (!time_after(inet_csk(sk)->icsk_timeout, (jiffies + delay)))
2060 inet_csk_reset_xmit_timer(sk, ICSK_TIME_EARLY_RETRANS, delay,
2065 /* Linux NewReno/SACK/FACK/ECN state machine.
2066 * --------------------------------------
2068 * "Open" Normal state, no dubious events, fast path.
2069 * "Disorder" In all the respects it is "Open",
2070 * but requires a bit more attention. It is entered when
2071 * we see some SACKs or dupacks. It is split of "Open"
2072 * mainly to move some processing from fast path to slow one.
2073 * "CWR" CWND was reduced due to some Congestion Notification event.
2074 * It can be ECN, ICMP source quench, local device congestion.
2075 * "Recovery" CWND was reduced, we are fast-retransmitting.
2076 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2078 * tcp_fastretrans_alert() is entered:
2079 * - each incoming ACK, if state is not "Open"
2080 * - when arrived ACK is unusual, namely:
2085 * Counting packets in flight is pretty simple.
2087 * in_flight = packets_out - left_out + retrans_out
2089 * packets_out is SND.NXT-SND.UNA counted in packets.
2091 * retrans_out is number of retransmitted segments.
2093 * left_out is number of segments left network, but not ACKed yet.
2095 * left_out = sacked_out + lost_out
2097 * sacked_out: Packets, which arrived to receiver out of order
2098 * and hence not ACKed. With SACKs this number is simply
2099 * amount of SACKed data. Even without SACKs
2100 * it is easy to give pretty reliable estimate of this number,
2101 * counting duplicate ACKs.
2103 * lost_out: Packets lost by network. TCP has no explicit
2104 * "loss notification" feedback from network (for now).
2105 * It means that this number can be only _guessed_.
2106 * Actually, it is the heuristics to predict lossage that
2107 * distinguishes different algorithms.
2109 * F.e. after RTO, when all the queue is considered as lost,
2110 * lost_out = packets_out and in_flight = retrans_out.
2112 * Essentially, we have now two algorithms counting
2115 * FACK: It is the simplest heuristics. As soon as we decided
2116 * that something is lost, we decide that _all_ not SACKed
2117 * packets until the most forward SACK are lost. I.e.
2118 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2119 * It is absolutely correct estimate, if network does not reorder
2120 * packets. And it loses any connection to reality when reordering
2121 * takes place. We use FACK by default until reordering
2122 * is suspected on the path to this destination.
2124 * NewReno: when Recovery is entered, we assume that one segment
2125 * is lost (classic Reno). While we are in Recovery and
2126 * a partial ACK arrives, we assume that one more packet
2127 * is lost (NewReno). This heuristics are the same in NewReno
2130 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2131 * deflation etc. CWND is real congestion window, never inflated, changes
2132 * only according to classic VJ rules.
2134 * Really tricky (and requiring careful tuning) part of algorithm
2135 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2136 * The first determines the moment _when_ we should reduce CWND and,
2137 * hence, slow down forward transmission. In fact, it determines the moment
2138 * when we decide that hole is caused by loss, rather than by a reorder.
2140 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2141 * holes, caused by lost packets.
2143 * And the most logically complicated part of algorithm is undo
2144 * heuristics. We detect false retransmits due to both too early
2145 * fast retransmit (reordering) and underestimated RTO, analyzing
2146 * timestamps and D-SACKs. When we detect that some segments were
2147 * retransmitted by mistake and CWND reduction was wrong, we undo
2148 * window reduction and abort recovery phase. This logic is hidden
2149 * inside several functions named tcp_try_undo_<something>.
2152 /* This function decides, when we should leave Disordered state
2153 * and enter Recovery phase, reducing congestion window.
2155 * Main question: may we further continue forward transmission
2156 * with the same cwnd?
2158 static bool tcp_time_to_recover(struct sock *sk, int flag)
2160 struct tcp_sock *tp = tcp_sk(sk);
2163 /* Trick#1: The loss is proven. */
2167 /* Not-A-Trick#2 : Classic rule... */
2168 if (tcp_dupack_heuristics(tp) > tp->reordering)
2171 /* Trick#4: It is still not OK... But will it be useful to delay
2174 packets_out = tp->packets_out;
2175 if (packets_out <= tp->reordering &&
2176 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2177 !tcp_may_send_now(sk)) {
2178 /* We have nothing to send. This connection is limited
2179 * either by receiver window or by application.
2184 /* If a thin stream is detected, retransmit after first
2185 * received dupack. Employ only if SACK is supported in order
2186 * to avoid possible corner-case series of spurious retransmissions
2187 * Use only if there are no unsent data.
2189 if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2190 tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2191 tcp_is_sack(tp) && !tcp_send_head(sk))
2194 /* Trick#6: TCP early retransmit, per RFC5827. To avoid spurious
2195 * retransmissions due to small network reorderings, we implement
2196 * Mitigation A.3 in the RFC and delay the retransmission for a short
2197 * interval if appropriate.
2199 if (tp->do_early_retrans && !tp->retrans_out && tp->sacked_out &&
2200 (tp->packets_out >= (tp->sacked_out + 1) && tp->packets_out < 4) &&
2201 !tcp_may_send_now(sk))
2202 return !tcp_pause_early_retransmit(sk, flag);
2207 /* Detect loss in event "A" above by marking head of queue up as lost.
2208 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2209 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2210 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2211 * the maximum SACKed segments to pass before reaching this limit.
2213 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2215 struct tcp_sock *tp = tcp_sk(sk);
2216 struct sk_buff *skb;
2220 /* Use SACK to deduce losses of new sequences sent during recovery */
2221 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq;
2223 WARN_ON(packets > tp->packets_out);
2224 if (tp->lost_skb_hint) {
2225 skb = tp->lost_skb_hint;
2226 cnt = tp->lost_cnt_hint;
2227 /* Head already handled? */
2228 if (mark_head && skb != tcp_write_queue_head(sk))
2231 skb = tcp_write_queue_head(sk);
2235 tcp_for_write_queue_from(skb, sk) {
2236 if (skb == tcp_send_head(sk))
2238 /* TODO: do this better */
2239 /* this is not the most efficient way to do this... */
2240 tp->lost_skb_hint = skb;
2241 tp->lost_cnt_hint = cnt;
2243 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2247 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2248 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2249 cnt += tcp_skb_pcount(skb);
2251 if (cnt > packets) {
2252 if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2253 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2254 (oldcnt >= packets))
2257 mss = skb_shinfo(skb)->gso_size;
2258 err = tcp_fragment(sk, skb, (packets - oldcnt) * mss,
2265 tcp_skb_mark_lost(tp, skb);
2270 tcp_verify_left_out(tp);
2273 /* Account newly detected lost packet(s) */
2275 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2277 struct tcp_sock *tp = tcp_sk(sk);
2279 if (tcp_is_reno(tp)) {
2280 tcp_mark_head_lost(sk, 1, 1);
2281 } else if (tcp_is_fack(tp)) {
2282 int lost = tp->fackets_out - tp->reordering;
2285 tcp_mark_head_lost(sk, lost, 0);
2287 int sacked_upto = tp->sacked_out - tp->reordering;
2288 if (sacked_upto >= 0)
2289 tcp_mark_head_lost(sk, sacked_upto, 0);
2290 else if (fast_rexmit)
2291 tcp_mark_head_lost(sk, 1, 1);
2295 /* CWND moderation, preventing bursts due to too big ACKs
2296 * in dubious situations.
2298 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2300 tp->snd_cwnd = min(tp->snd_cwnd,
2301 tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2302 tp->snd_cwnd_stamp = tcp_time_stamp;
2305 /* Nothing was retransmitted or returned timestamp is less
2306 * than timestamp of the first retransmission.
2308 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2310 return !tp->retrans_stamp ||
2311 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2312 before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
2315 /* Undo procedures. */
2317 #if FASTRETRANS_DEBUG > 1
2318 static void DBGUNDO(struct sock *sk, const char *msg)
2320 struct tcp_sock *tp = tcp_sk(sk);
2321 struct inet_sock *inet = inet_sk(sk);
2323 if (sk->sk_family == AF_INET) {
2324 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2326 &inet->inet_daddr, ntohs(inet->inet_dport),
2327 tp->snd_cwnd, tcp_left_out(tp),
2328 tp->snd_ssthresh, tp->prior_ssthresh,
2331 #if IS_ENABLED(CONFIG_IPV6)
2332 else if (sk->sk_family == AF_INET6) {
2333 struct ipv6_pinfo *np = inet6_sk(sk);
2334 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2336 &np->daddr, ntohs(inet->inet_dport),
2337 tp->snd_cwnd, tcp_left_out(tp),
2338 tp->snd_ssthresh, tp->prior_ssthresh,
2344 #define DBGUNDO(x...) do { } while (0)
2347 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2349 struct tcp_sock *tp = tcp_sk(sk);
2352 struct sk_buff *skb;
2354 tcp_for_write_queue(skb, sk) {
2355 if (skb == tcp_send_head(sk))
2357 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2360 tcp_clear_all_retrans_hints(tp);
2363 if (tp->prior_ssthresh) {
2364 const struct inet_connection_sock *icsk = inet_csk(sk);
2366 if (icsk->icsk_ca_ops->undo_cwnd)
2367 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2369 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2371 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2372 tp->snd_ssthresh = tp->prior_ssthresh;
2373 tcp_ecn_withdraw_cwr(tp);
2376 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2378 tp->snd_cwnd_stamp = tcp_time_stamp;
2379 tp->undo_marker = 0;
2382 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2384 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2387 /* People celebrate: "We love our President!" */
2388 static bool tcp_try_undo_recovery(struct sock *sk)
2390 struct tcp_sock *tp = tcp_sk(sk);
2392 if (tcp_may_undo(tp)) {
2395 /* Happy end! We did not retransmit anything
2396 * or our original transmission succeeded.
2398 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2399 tcp_undo_cwnd_reduction(sk, false);
2400 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2401 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2403 mib_idx = LINUX_MIB_TCPFULLUNDO;
2405 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2407 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2408 /* Hold old state until something *above* high_seq
2409 * is ACKed. For Reno it is MUST to prevent false
2410 * fast retransmits (RFC2582). SACK TCP is safe. */
2411 tcp_moderate_cwnd(tp);
2414 tcp_set_ca_state(sk, TCP_CA_Open);
2418 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2419 static bool tcp_try_undo_dsack(struct sock *sk)
2421 struct tcp_sock *tp = tcp_sk(sk);
2423 if (tp->undo_marker && !tp->undo_retrans) {
2424 DBGUNDO(sk, "D-SACK");
2425 tcp_undo_cwnd_reduction(sk, false);
2426 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2432 /* We can clear retrans_stamp when there are no retransmissions in the
2433 * window. It would seem that it is trivially available for us in
2434 * tp->retrans_out, however, that kind of assumptions doesn't consider
2435 * what will happen if errors occur when sending retransmission for the
2436 * second time. ...It could the that such segment has only
2437 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2438 * the head skb is enough except for some reneging corner cases that
2439 * are not worth the effort.
2441 * Main reason for all this complexity is the fact that connection dying
2442 * time now depends on the validity of the retrans_stamp, in particular,
2443 * that successive retransmissions of a segment must not advance
2444 * retrans_stamp under any conditions.
2446 static bool tcp_any_retrans_done(const struct sock *sk)
2448 const struct tcp_sock *tp = tcp_sk(sk);
2449 struct sk_buff *skb;
2451 if (tp->retrans_out)
2454 skb = tcp_write_queue_head(sk);
2455 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2461 /* Undo during loss recovery after partial ACK or using F-RTO. */
2462 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2464 struct tcp_sock *tp = tcp_sk(sk);
2466 if (frto_undo || tcp_may_undo(tp)) {
2467 tcp_undo_cwnd_reduction(sk, true);
2469 DBGUNDO(sk, "partial loss");
2470 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2472 NET_INC_STATS_BH(sock_net(sk),
2473 LINUX_MIB_TCPSPURIOUSRTOS);
2474 inet_csk(sk)->icsk_retransmits = 0;
2475 if (frto_undo || tcp_is_sack(tp))
2476 tcp_set_ca_state(sk, TCP_CA_Open);
2482 /* The cwnd reduction in CWR and Recovery use the PRR algorithm
2483 * https://datatracker.ietf.org/doc/draft-ietf-tcpm-proportional-rate-reduction/
2484 * It computes the number of packets to send (sndcnt) based on packets newly
2486 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2487 * cwnd reductions across a full RTT.
2488 * 2) If packets in flight is lower than ssthresh (such as due to excess
2489 * losses and/or application stalls), do not perform any further cwnd
2490 * reductions, but instead slow start up to ssthresh.
2492 static void tcp_init_cwnd_reduction(struct sock *sk)
2494 struct tcp_sock *tp = tcp_sk(sk);
2496 tp->high_seq = tp->snd_nxt;
2497 tp->tlp_high_seq = 0;
2498 tp->snd_cwnd_cnt = 0;
2499 tp->prior_cwnd = tp->snd_cwnd;
2500 tp->prr_delivered = 0;
2502 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2503 tcp_ecn_queue_cwr(tp);
2506 static void tcp_cwnd_reduction(struct sock *sk, const int prior_unsacked,
2509 struct tcp_sock *tp = tcp_sk(sk);
2511 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2512 int newly_acked_sacked = prior_unsacked -
2513 (tp->packets_out - tp->sacked_out);
2515 tp->prr_delivered += newly_acked_sacked;
2516 if (tcp_packets_in_flight(tp) > tp->snd_ssthresh) {
2517 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2519 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2521 sndcnt = min_t(int, delta,
2522 max_t(int, tp->prr_delivered - tp->prr_out,
2523 newly_acked_sacked) + 1);
2526 sndcnt = max(sndcnt, (fast_rexmit ? 1 : 0));
2527 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2530 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2532 struct tcp_sock *tp = tcp_sk(sk);
2534 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2535 if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR ||
2536 (tp->undo_marker && tp->snd_ssthresh < TCP_INFINITE_SSTHRESH)) {
2537 tp->snd_cwnd = tp->snd_ssthresh;
2538 tp->snd_cwnd_stamp = tcp_time_stamp;
2540 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2543 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2544 void tcp_enter_cwr(struct sock *sk)
2546 struct tcp_sock *tp = tcp_sk(sk);
2548 tp->prior_ssthresh = 0;
2549 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2550 tp->undo_marker = 0;
2551 tcp_init_cwnd_reduction(sk);
2552 tcp_set_ca_state(sk, TCP_CA_CWR);
2556 static void tcp_try_keep_open(struct sock *sk)
2558 struct tcp_sock *tp = tcp_sk(sk);
2559 int state = TCP_CA_Open;
2561 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2562 state = TCP_CA_Disorder;
2564 if (inet_csk(sk)->icsk_ca_state != state) {
2565 tcp_set_ca_state(sk, state);
2566 tp->high_seq = tp->snd_nxt;
2570 static void tcp_try_to_open(struct sock *sk, int flag, const int prior_unsacked)
2572 struct tcp_sock *tp = tcp_sk(sk);
2574 tcp_verify_left_out(tp);
2576 if (!tcp_any_retrans_done(sk))
2577 tp->retrans_stamp = 0;
2579 if (flag & FLAG_ECE)
2582 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2583 tcp_try_keep_open(sk);
2585 tcp_cwnd_reduction(sk, prior_unsacked, 0);
2589 static void tcp_mtup_probe_failed(struct sock *sk)
2591 struct inet_connection_sock *icsk = inet_csk(sk);
2593 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2594 icsk->icsk_mtup.probe_size = 0;
2597 static void tcp_mtup_probe_success(struct sock *sk)
2599 struct tcp_sock *tp = tcp_sk(sk);
2600 struct inet_connection_sock *icsk = inet_csk(sk);
2602 /* FIXME: breaks with very large cwnd */
2603 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2604 tp->snd_cwnd = tp->snd_cwnd *
2605 tcp_mss_to_mtu(sk, tp->mss_cache) /
2606 icsk->icsk_mtup.probe_size;
2607 tp->snd_cwnd_cnt = 0;
2608 tp->snd_cwnd_stamp = tcp_time_stamp;
2609 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2611 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2612 icsk->icsk_mtup.probe_size = 0;
2613 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2616 /* Do a simple retransmit without using the backoff mechanisms in
2617 * tcp_timer. This is used for path mtu discovery.
2618 * The socket is already locked here.
2620 void tcp_simple_retransmit(struct sock *sk)
2622 const struct inet_connection_sock *icsk = inet_csk(sk);
2623 struct tcp_sock *tp = tcp_sk(sk);
2624 struct sk_buff *skb;
2625 unsigned int mss = tcp_current_mss(sk);
2626 u32 prior_lost = tp->lost_out;
2628 tcp_for_write_queue(skb, sk) {
2629 if (skb == tcp_send_head(sk))
2631 if (tcp_skb_seglen(skb) > mss &&
2632 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2633 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2634 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2635 tp->retrans_out -= tcp_skb_pcount(skb);
2637 tcp_skb_mark_lost_uncond_verify(tp, skb);
2641 tcp_clear_retrans_hints_partial(tp);
2643 if (prior_lost == tp->lost_out)
2646 if (tcp_is_reno(tp))
2647 tcp_limit_reno_sacked(tp);
2649 tcp_verify_left_out(tp);
2651 /* Don't muck with the congestion window here.
2652 * Reason is that we do not increase amount of _data_
2653 * in network, but units changed and effective
2654 * cwnd/ssthresh really reduced now.
2656 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2657 tp->high_seq = tp->snd_nxt;
2658 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2659 tp->prior_ssthresh = 0;
2660 tp->undo_marker = 0;
2661 tcp_set_ca_state(sk, TCP_CA_Loss);
2663 tcp_xmit_retransmit_queue(sk);
2665 EXPORT_SYMBOL(tcp_simple_retransmit);
2667 static void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2669 struct tcp_sock *tp = tcp_sk(sk);
2672 if (tcp_is_reno(tp))
2673 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2675 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2677 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2679 tp->prior_ssthresh = 0;
2682 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2684 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2685 tcp_init_cwnd_reduction(sk);
2687 tcp_set_ca_state(sk, TCP_CA_Recovery);
2690 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2691 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2693 static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack)
2695 struct tcp_sock *tp = tcp_sk(sk);
2696 bool recovered = !before(tp->snd_una, tp->high_seq);
2698 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2699 /* Step 3.b. A timeout is spurious if not all data are
2700 * lost, i.e., never-retransmitted data are (s)acked.
2702 if (tcp_try_undo_loss(sk, flag & FLAG_ORIG_SACK_ACKED))
2705 if (after(tp->snd_nxt, tp->high_seq) &&
2706 (flag & FLAG_DATA_SACKED || is_dupack)) {
2707 tp->frto = 0; /* Loss was real: 2nd part of step 3.a */
2708 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2709 tp->high_seq = tp->snd_nxt;
2710 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
2712 if (after(tp->snd_nxt, tp->high_seq))
2713 return; /* Step 2.b */
2719 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2720 tcp_try_undo_recovery(sk);
2723 if (tcp_is_reno(tp)) {
2724 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2725 * delivered. Lower inflight to clock out (re)tranmissions.
2727 if (after(tp->snd_nxt, tp->high_seq) && is_dupack)
2728 tcp_add_reno_sack(sk);
2729 else if (flag & FLAG_SND_UNA_ADVANCED)
2730 tcp_reset_reno_sack(tp);
2732 if (tcp_try_undo_loss(sk, false))
2734 tcp_xmit_retransmit_queue(sk);
2737 /* Undo during fast recovery after partial ACK. */
2738 static bool tcp_try_undo_partial(struct sock *sk, const int acked,
2739 const int prior_unsacked)
2741 struct tcp_sock *tp = tcp_sk(sk);
2743 if (tp->undo_marker && tcp_packet_delayed(tp)) {
2744 /* Plain luck! Hole if filled with delayed
2745 * packet, rather than with a retransmit.
2747 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2749 /* We are getting evidence that the reordering degree is higher
2750 * than we realized. If there are no retransmits out then we
2751 * can undo. Otherwise we clock out new packets but do not
2752 * mark more packets lost or retransmit more.
2754 if (tp->retrans_out) {
2755 tcp_cwnd_reduction(sk, prior_unsacked, 0);
2759 if (!tcp_any_retrans_done(sk))
2760 tp->retrans_stamp = 0;
2762 DBGUNDO(sk, "partial recovery");
2763 tcp_undo_cwnd_reduction(sk, true);
2764 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2765 tcp_try_keep_open(sk);
2771 /* Process an event, which can update packets-in-flight not trivially.
2772 * Main goal of this function is to calculate new estimate for left_out,
2773 * taking into account both packets sitting in receiver's buffer and
2774 * packets lost by network.
2776 * Besides that it does CWND reduction, when packet loss is detected
2777 * and changes state of machine.
2779 * It does _not_ decide what to send, it is made in function
2780 * tcp_xmit_retransmit_queue().
2782 static void tcp_fastretrans_alert(struct sock *sk, const int acked,
2783 const int prior_unsacked,
2784 bool is_dupack, int flag)
2786 struct inet_connection_sock *icsk = inet_csk(sk);
2787 struct tcp_sock *tp = tcp_sk(sk);
2788 bool do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2789 (tcp_fackets_out(tp) > tp->reordering));
2790 int fast_rexmit = 0;
2792 if (WARN_ON(!tp->packets_out && tp->sacked_out))
2794 if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2795 tp->fackets_out = 0;
2797 /* Now state machine starts.
2798 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2799 if (flag & FLAG_ECE)
2800 tp->prior_ssthresh = 0;
2802 /* B. In all the states check for reneging SACKs. */
2803 if (tcp_check_sack_reneging(sk, flag))
2806 /* C. Check consistency of the current state. */
2807 tcp_verify_left_out(tp);
2809 /* D. Check state exit conditions. State can be terminated
2810 * when high_seq is ACKed. */
2811 if (icsk->icsk_ca_state == TCP_CA_Open) {
2812 WARN_ON(tp->retrans_out != 0);
2813 tp->retrans_stamp = 0;
2814 } else if (!before(tp->snd_una, tp->high_seq)) {
2815 switch (icsk->icsk_ca_state) {
2817 /* CWR is to be held something *above* high_seq
2818 * is ACKed for CWR bit to reach receiver. */
2819 if (tp->snd_una != tp->high_seq) {
2820 tcp_end_cwnd_reduction(sk);
2821 tcp_set_ca_state(sk, TCP_CA_Open);
2825 case TCP_CA_Recovery:
2826 if (tcp_is_reno(tp))
2827 tcp_reset_reno_sack(tp);
2828 if (tcp_try_undo_recovery(sk))
2830 tcp_end_cwnd_reduction(sk);
2835 /* E. Process state. */
2836 switch (icsk->icsk_ca_state) {
2837 case TCP_CA_Recovery:
2838 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2839 if (tcp_is_reno(tp) && is_dupack)
2840 tcp_add_reno_sack(sk);
2842 if (tcp_try_undo_partial(sk, acked, prior_unsacked))
2844 /* Partial ACK arrived. Force fast retransmit. */
2845 do_lost = tcp_is_reno(tp) ||
2846 tcp_fackets_out(tp) > tp->reordering;
2848 if (tcp_try_undo_dsack(sk)) {
2849 tcp_try_keep_open(sk);
2854 tcp_process_loss(sk, flag, is_dupack);
2855 if (icsk->icsk_ca_state != TCP_CA_Open)
2857 /* Fall through to processing in Open state. */
2859 if (tcp_is_reno(tp)) {
2860 if (flag & FLAG_SND_UNA_ADVANCED)
2861 tcp_reset_reno_sack(tp);
2863 tcp_add_reno_sack(sk);
2866 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2867 tcp_try_undo_dsack(sk);
2869 if (!tcp_time_to_recover(sk, flag)) {
2870 tcp_try_to_open(sk, flag, prior_unsacked);
2874 /* MTU probe failure: don't reduce cwnd */
2875 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2876 icsk->icsk_mtup.probe_size &&
2877 tp->snd_una == tp->mtu_probe.probe_seq_start) {
2878 tcp_mtup_probe_failed(sk);
2879 /* Restores the reduction we did in tcp_mtup_probe() */
2881 tcp_simple_retransmit(sk);
2885 /* Otherwise enter Recovery state */
2886 tcp_enter_recovery(sk, (flag & FLAG_ECE));
2891 tcp_update_scoreboard(sk, fast_rexmit);
2892 tcp_cwnd_reduction(sk, prior_unsacked, fast_rexmit);
2893 tcp_xmit_retransmit_queue(sk);
2896 static inline bool tcp_ack_update_rtt(struct sock *sk, const int flag,
2897 long seq_rtt_us, long sack_rtt_us)
2899 const struct tcp_sock *tp = tcp_sk(sk);
2901 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2902 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2903 * Karn's algorithm forbids taking RTT if some retransmitted data
2904 * is acked (RFC6298).
2906 if (flag & FLAG_RETRANS_DATA_ACKED)
2910 seq_rtt_us = sack_rtt_us;
2912 /* RTTM Rule: A TSecr value received in a segment is used to
2913 * update the averaged RTT measurement only if the segment
2914 * acknowledges some new data, i.e., only if it advances the
2915 * left edge of the send window.
2916 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2918 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2920 seq_rtt_us = jiffies_to_usecs(tcp_time_stamp - tp->rx_opt.rcv_tsecr);
2925 tcp_rtt_estimator(sk, seq_rtt_us);
2928 /* RFC6298: only reset backoff on valid RTT measurement. */
2929 inet_csk(sk)->icsk_backoff = 0;
2933 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2934 static void tcp_synack_rtt_meas(struct sock *sk, const u32 synack_stamp)
2936 struct tcp_sock *tp = tcp_sk(sk);
2937 long seq_rtt_us = -1L;
2939 if (synack_stamp && !tp->total_retrans)
2940 seq_rtt_us = jiffies_to_usecs(tcp_time_stamp - synack_stamp);
2942 /* If the ACK acks both the SYNACK and the (Fast Open'd) data packets
2943 * sent in SYN_RECV, SYNACK RTT is the smooth RTT computed in tcp_ack()
2946 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, seq_rtt_us, -1L);
2949 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
2951 const struct inet_connection_sock *icsk = inet_csk(sk);
2953 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
2954 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
2957 /* Restart timer after forward progress on connection.
2958 * RFC2988 recommends to restart timer to now+rto.
2960 void tcp_rearm_rto(struct sock *sk)
2962 const struct inet_connection_sock *icsk = inet_csk(sk);
2963 struct tcp_sock *tp = tcp_sk(sk);
2965 /* If the retrans timer is currently being used by Fast Open
2966 * for SYN-ACK retrans purpose, stay put.
2968 if (tp->fastopen_rsk)
2971 if (!tp->packets_out) {
2972 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2974 u32 rto = inet_csk(sk)->icsk_rto;
2975 /* Offset the time elapsed after installing regular RTO */
2976 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
2977 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
2978 struct sk_buff *skb = tcp_write_queue_head(sk);
2979 const u32 rto_time_stamp =
2980 tcp_skb_timestamp(skb) + rto;
2981 s32 delta = (s32)(rto_time_stamp - tcp_time_stamp);
2982 /* delta may not be positive if the socket is locked
2983 * when the retrans timer fires and is rescheduled.
2988 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
2993 /* This function is called when the delayed ER timer fires. TCP enters
2994 * fast recovery and performs fast-retransmit.
2996 void tcp_resume_early_retransmit(struct sock *sk)
2998 struct tcp_sock *tp = tcp_sk(sk);
3002 /* Stop if ER is disabled after the delayed ER timer is scheduled */
3003 if (!tp->do_early_retrans)
3006 tcp_enter_recovery(sk, false);
3007 tcp_update_scoreboard(sk, 1);
3008 tcp_xmit_retransmit_queue(sk);
3011 /* If we get here, the whole TSO packet has not been acked. */
3012 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3014 struct tcp_sock *tp = tcp_sk(sk);
3017 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3019 packets_acked = tcp_skb_pcount(skb);
3020 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3022 packets_acked -= tcp_skb_pcount(skb);
3024 if (packets_acked) {
3025 BUG_ON(tcp_skb_pcount(skb) == 0);
3026 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3029 return packets_acked;
3032 /* Remove acknowledged frames from the retransmission queue. If our packet
3033 * is before the ack sequence we can discard it as it's confirmed to have
3034 * arrived at the other end.
3036 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3037 u32 prior_snd_una, long sack_rtt_us)
3039 const struct inet_connection_sock *icsk = inet_csk(sk);
3040 struct skb_mstamp first_ackt, last_ackt, now;
3041 struct tcp_sock *tp = tcp_sk(sk);
3042 u32 prior_sacked = tp->sacked_out;
3043 u32 reord = tp->packets_out;
3044 bool fully_acked = true;
3045 long ca_seq_rtt_us = -1L;
3046 long seq_rtt_us = -1L;
3047 struct sk_buff *skb;
3054 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3055 struct skb_shared_info *shinfo = skb_shinfo(skb);
3056 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3057 u8 sacked = scb->sacked;
3060 if (unlikely(shinfo->tx_flags & SKBTX_ACK_TSTAMP) &&
3061 between(shinfo->tskey, prior_snd_una, tp->snd_una - 1))
3062 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK);
3064 /* Determine how many packets and what bytes were acked, tso and else */
3065 if (after(scb->end_seq, tp->snd_una)) {
3066 if (tcp_skb_pcount(skb) == 1 ||
3067 !after(tp->snd_una, scb->seq))
3070 acked_pcount = tcp_tso_acked(sk, skb);
3074 fully_acked = false;
3076 acked_pcount = tcp_skb_pcount(skb);
3079 if (sacked & TCPCB_RETRANS) {
3080 if (sacked & TCPCB_SACKED_RETRANS)
3081 tp->retrans_out -= acked_pcount;
3082 flag |= FLAG_RETRANS_DATA_ACKED;
3084 last_ackt = skb->skb_mstamp;
3085 WARN_ON_ONCE(last_ackt.v64 == 0);
3086 if (!first_ackt.v64)
3087 first_ackt = last_ackt;
3089 if (!(sacked & TCPCB_SACKED_ACKED))
3090 reord = min(pkts_acked, reord);
3091 if (!after(scb->end_seq, tp->high_seq))
3092 flag |= FLAG_ORIG_SACK_ACKED;
3095 if (sacked & TCPCB_SACKED_ACKED)
3096 tp->sacked_out -= acked_pcount;
3097 if (sacked & TCPCB_LOST)
3098 tp->lost_out -= acked_pcount;
3100 tp->packets_out -= acked_pcount;
3101 pkts_acked += acked_pcount;
3103 /* Initial outgoing SYN's get put onto the write_queue
3104 * just like anything else we transmit. It is not
3105 * true data, and if we misinform our callers that
3106 * this ACK acks real data, we will erroneously exit
3107 * connection startup slow start one packet too
3108 * quickly. This is severely frowned upon behavior.
3110 if (!(scb->tcp_flags & TCPHDR_SYN)) {
3111 flag |= FLAG_DATA_ACKED;
3113 flag |= FLAG_SYN_ACKED;
3114 tp->retrans_stamp = 0;
3120 tcp_unlink_write_queue(skb, sk);
3121 sk_wmem_free_skb(sk, skb);
3122 if (skb == tp->retransmit_skb_hint)
3123 tp->retransmit_skb_hint = NULL;
3124 if (skb == tp->lost_skb_hint)
3125 tp->lost_skb_hint = NULL;
3128 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3129 tp->snd_up = tp->snd_una;
3131 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3132 flag |= FLAG_SACK_RENEGING;
3134 skb_mstamp_get(&now);
3135 if (first_ackt.v64) {
3136 seq_rtt_us = skb_mstamp_us_delta(&now, &first_ackt);
3137 ca_seq_rtt_us = skb_mstamp_us_delta(&now, &last_ackt);
3140 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us);
3142 if (flag & FLAG_ACKED) {
3143 const struct tcp_congestion_ops *ca_ops
3144 = inet_csk(sk)->icsk_ca_ops;
3147 if (unlikely(icsk->icsk_mtup.probe_size &&
3148 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3149 tcp_mtup_probe_success(sk);
3152 if (tcp_is_reno(tp)) {
3153 tcp_remove_reno_sacks(sk, pkts_acked);
3157 /* Non-retransmitted hole got filled? That's reordering */
3158 if (reord < prior_fackets)
3159 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3161 delta = tcp_is_fack(tp) ? pkts_acked :
3162 prior_sacked - tp->sacked_out;
3163 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3166 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3168 if (ca_ops->pkts_acked)
3169 ca_ops->pkts_acked(sk, pkts_acked, ca_seq_rtt_us);
3171 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3172 sack_rtt_us > skb_mstamp_us_delta(&now, &skb->skb_mstamp)) {
3173 /* Do not re-arm RTO if the sack RTT is measured from data sent
3174 * after when the head was last (re)transmitted. Otherwise the
3175 * timeout may continue to extend in loss recovery.
3180 #if FASTRETRANS_DEBUG > 0
3181 WARN_ON((int)tp->sacked_out < 0);
3182 WARN_ON((int)tp->lost_out < 0);
3183 WARN_ON((int)tp->retrans_out < 0);
3184 if (!tp->packets_out && tcp_is_sack(tp)) {
3185 icsk = inet_csk(sk);
3187 pr_debug("Leak l=%u %d\n",
3188 tp->lost_out, icsk->icsk_ca_state);
3191 if (tp->sacked_out) {
3192 pr_debug("Leak s=%u %d\n",
3193 tp->sacked_out, icsk->icsk_ca_state);
3196 if (tp->retrans_out) {
3197 pr_debug("Leak r=%u %d\n",
3198 tp->retrans_out, icsk->icsk_ca_state);
3199 tp->retrans_out = 0;
3206 static void tcp_ack_probe(struct sock *sk)
3208 const struct tcp_sock *tp = tcp_sk(sk);
3209 struct inet_connection_sock *icsk = inet_csk(sk);
3211 /* Was it a usable window open? */
3213 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3214 icsk->icsk_backoff = 0;
3215 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3216 /* Socket must be waked up by subsequent tcp_data_snd_check().
3217 * This function is not for random using!
3220 unsigned long when = inet_csk_rto_backoff(icsk, TCP_RTO_MAX);
3222 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3227 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3229 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3230 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3233 /* Decide wheather to run the increase function of congestion control. */
3234 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3236 if (tcp_in_cwnd_reduction(sk))
3239 /* If reordering is high then always grow cwnd whenever data is
3240 * delivered regardless of its ordering. Otherwise stay conservative
3241 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3242 * new SACK or ECE mark may first advance cwnd here and later reduce
3243 * cwnd in tcp_fastretrans_alert() based on more states.
3245 if (tcp_sk(sk)->reordering > sysctl_tcp_reordering)
3246 return flag & FLAG_FORWARD_PROGRESS;
3248 return flag & FLAG_DATA_ACKED;
3251 /* Check that window update is acceptable.
3252 * The function assumes that snd_una<=ack<=snd_next.
3254 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3255 const u32 ack, const u32 ack_seq,
3258 return after(ack, tp->snd_una) ||
3259 after(ack_seq, tp->snd_wl1) ||
3260 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3263 /* Update our send window.
3265 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3266 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3268 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3271 struct tcp_sock *tp = tcp_sk(sk);
3273 u32 nwin = ntohs(tcp_hdr(skb)->window);
3275 if (likely(!tcp_hdr(skb)->syn))
3276 nwin <<= tp->rx_opt.snd_wscale;
3278 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3279 flag |= FLAG_WIN_UPDATE;
3280 tcp_update_wl(tp, ack_seq);
3282 if (tp->snd_wnd != nwin) {
3285 /* Note, it is the only place, where
3286 * fast path is recovered for sending TCP.
3289 tcp_fast_path_check(sk);
3291 if (nwin > tp->max_window) {
3292 tp->max_window = nwin;
3293 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3303 /* RFC 5961 7 [ACK Throttling] */
3304 static void tcp_send_challenge_ack(struct sock *sk)
3306 /* unprotected vars, we dont care of overwrites */
3307 static u32 challenge_timestamp;
3308 static unsigned int challenge_count;
3309 u32 now = jiffies / HZ;
3311 if (now != challenge_timestamp) {
3312 challenge_timestamp = now;
3313 challenge_count = 0;
3315 if (++challenge_count <= sysctl_tcp_challenge_ack_limit) {
3316 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPCHALLENGEACK);
3321 static void tcp_store_ts_recent(struct tcp_sock *tp)
3323 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3324 tp->rx_opt.ts_recent_stamp = get_seconds();
3327 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3329 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3330 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3331 * extra check below makes sure this can only happen
3332 * for pure ACK frames. -DaveM
3334 * Not only, also it occurs for expired timestamps.
3337 if (tcp_paws_check(&tp->rx_opt, 0))
3338 tcp_store_ts_recent(tp);
3342 /* This routine deals with acks during a TLP episode.
3343 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3345 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3347 struct tcp_sock *tp = tcp_sk(sk);
3348 bool is_tlp_dupack = (ack == tp->tlp_high_seq) &&
3349 !(flag & (FLAG_SND_UNA_ADVANCED |
3350 FLAG_NOT_DUP | FLAG_DATA_SACKED));
3352 /* Mark the end of TLP episode on receiving TLP dupack or when
3353 * ack is after tlp_high_seq.
3355 if (is_tlp_dupack) {
3356 tp->tlp_high_seq = 0;
3360 if (after(ack, tp->tlp_high_seq)) {
3361 tp->tlp_high_seq = 0;
3362 /* Don't reduce cwnd if DSACK arrives for TLP retrans. */
3363 if (!(flag & FLAG_DSACKING_ACK)) {
3364 tcp_init_cwnd_reduction(sk);
3365 tcp_set_ca_state(sk, TCP_CA_CWR);
3366 tcp_end_cwnd_reduction(sk);
3367 tcp_try_keep_open(sk);
3368 NET_INC_STATS_BH(sock_net(sk),
3369 LINUX_MIB_TCPLOSSPROBERECOVERY);
3374 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3376 const struct inet_connection_sock *icsk = inet_csk(sk);
3378 if (icsk->icsk_ca_ops->in_ack_event)
3379 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3382 /* This routine deals with incoming acks, but not outgoing ones. */
3383 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3385 struct inet_connection_sock *icsk = inet_csk(sk);
3386 struct tcp_sock *tp = tcp_sk(sk);
3387 u32 prior_snd_una = tp->snd_una;
3388 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3389 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3390 bool is_dupack = false;
3392 int prior_packets = tp->packets_out;
3393 const int prior_unsacked = tp->packets_out - tp->sacked_out;
3394 int acked = 0; /* Number of packets newly acked */
3395 long sack_rtt_us = -1L;
3397 /* If the ack is older than previous acks
3398 * then we can probably ignore it.
3400 if (before(ack, prior_snd_una)) {
3401 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3402 if (before(ack, prior_snd_una - tp->max_window)) {
3403 tcp_send_challenge_ack(sk);
3409 /* If the ack includes data we haven't sent yet, discard
3410 * this segment (RFC793 Section 3.9).
3412 if (after(ack, tp->snd_nxt))
3415 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
3416 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
3419 if (after(ack, prior_snd_una)) {
3420 flag |= FLAG_SND_UNA_ADVANCED;
3421 icsk->icsk_retransmits = 0;
3424 prior_fackets = tp->fackets_out;
3426 /* ts_recent update must be made after we are sure that the packet
3429 if (flag & FLAG_UPDATE_TS_RECENT)
3430 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3432 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3433 /* Window is constant, pure forward advance.
3434 * No more checks are required.
3435 * Note, we use the fact that SND.UNA>=SND.WL2.
3437 tcp_update_wl(tp, ack_seq);
3439 flag |= FLAG_WIN_UPDATE;
3441 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3443 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3445 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3447 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3450 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3452 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3454 if (TCP_SKB_CB(skb)->sacked)
3455 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3458 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3460 ack_ev_flags |= CA_ACK_ECE;
3463 if (flag & FLAG_WIN_UPDATE)
3464 ack_ev_flags |= CA_ACK_WIN_UPDATE;
3466 tcp_in_ack_event(sk, ack_ev_flags);
3469 /* We passed data and got it acked, remove any soft error
3470 * log. Something worked...
3472 sk->sk_err_soft = 0;
3473 icsk->icsk_probes_out = 0;
3474 tp->rcv_tstamp = tcp_time_stamp;
3478 /* See if we can take anything off of the retransmit queue. */
3479 acked = tp->packets_out;
3480 flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una,
3482 acked -= tp->packets_out;
3484 /* Advance cwnd if state allows */
3485 if (tcp_may_raise_cwnd(sk, flag))
3486 tcp_cong_avoid(sk, ack, acked);
3488 if (tcp_ack_is_dubious(sk, flag)) {
3489 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3490 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3493 if (tp->tlp_high_seq)
3494 tcp_process_tlp_ack(sk, ack, flag);
3496 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) {
3497 struct dst_entry *dst = __sk_dst_get(sk);
3502 if (icsk->icsk_pending == ICSK_TIME_RETRANS)
3503 tcp_schedule_loss_probe(sk);
3504 tcp_update_pacing_rate(sk);
3508 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3509 if (flag & FLAG_DSACKING_ACK)
3510 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3512 /* If this ack opens up a zero window, clear backoff. It was
3513 * being used to time the probes, and is probably far higher than
3514 * it needs to be for normal retransmission.
3516 if (tcp_send_head(sk))
3519 if (tp->tlp_high_seq)
3520 tcp_process_tlp_ack(sk, ack, flag);
3524 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3528 /* If data was SACKed, tag it and see if we should send more data.
3529 * If data was DSACKed, see if we can undo a cwnd reduction.
3531 if (TCP_SKB_CB(skb)->sacked) {
3532 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3534 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3538 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3542 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3543 * But, this can also be called on packets in the established flow when
3544 * the fast version below fails.
3546 void tcp_parse_options(const struct sk_buff *skb,
3547 struct tcp_options_received *opt_rx, int estab,
3548 struct tcp_fastopen_cookie *foc)
3550 const unsigned char *ptr;
3551 const struct tcphdr *th = tcp_hdr(skb);
3552 int length = (th->doff * 4) - sizeof(struct tcphdr);
3554 ptr = (const unsigned char *)(th + 1);
3555 opt_rx->saw_tstamp = 0;
3557 while (length > 0) {
3558 int opcode = *ptr++;
3564 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3569 if (opsize < 2) /* "silly options" */
3571 if (opsize > length)
3572 return; /* don't parse partial options */
3575 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3576 u16 in_mss = get_unaligned_be16(ptr);
3578 if (opt_rx->user_mss &&
3579 opt_rx->user_mss < in_mss)
3580 in_mss = opt_rx->user_mss;
3581 opt_rx->mss_clamp = in_mss;
3586 if (opsize == TCPOLEN_WINDOW && th->syn &&
3587 !estab && sysctl_tcp_window_scaling) {
3588 __u8 snd_wscale = *(__u8 *)ptr;
3589 opt_rx->wscale_ok = 1;
3590 if (snd_wscale > 14) {
3591 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3596 opt_rx->snd_wscale = snd_wscale;
3599 case TCPOPT_TIMESTAMP:
3600 if ((opsize == TCPOLEN_TIMESTAMP) &&
3601 ((estab && opt_rx->tstamp_ok) ||
3602 (!estab && sysctl_tcp_timestamps))) {
3603 opt_rx->saw_tstamp = 1;
3604 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3605 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3608 case TCPOPT_SACK_PERM:
3609 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3610 !estab && sysctl_tcp_sack) {
3611 opt_rx->sack_ok = TCP_SACK_SEEN;
3612 tcp_sack_reset(opt_rx);
3617 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3618 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3620 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3623 #ifdef CONFIG_TCP_MD5SIG
3626 * The MD5 Hash has already been
3627 * checked (see tcp_v{4,6}_do_rcv()).
3632 /* Fast Open option shares code 254 using a
3633 * 16 bits magic number. It's valid only in
3634 * SYN or SYN-ACK with an even size.
3636 if (opsize < TCPOLEN_EXP_FASTOPEN_BASE ||
3637 get_unaligned_be16(ptr) != TCPOPT_FASTOPEN_MAGIC ||
3638 foc == NULL || !th->syn || (opsize & 1))
3640 foc->len = opsize - TCPOLEN_EXP_FASTOPEN_BASE;
3641 if (foc->len >= TCP_FASTOPEN_COOKIE_MIN &&
3642 foc->len <= TCP_FASTOPEN_COOKIE_MAX)
3643 memcpy(foc->val, ptr + 2, foc->len);
3644 else if (foc->len != 0)
3654 EXPORT_SYMBOL(tcp_parse_options);
3656 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3658 const __be32 *ptr = (const __be32 *)(th + 1);
3660 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3661 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3662 tp->rx_opt.saw_tstamp = 1;
3664 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3667 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3669 tp->rx_opt.rcv_tsecr = 0;
3675 /* Fast parse options. This hopes to only see timestamps.
3676 * If it is wrong it falls back on tcp_parse_options().
3678 static bool tcp_fast_parse_options(const struct sk_buff *skb,
3679 const struct tcphdr *th, struct tcp_sock *tp)
3681 /* In the spirit of fast parsing, compare doff directly to constant
3682 * values. Because equality is used, short doff can be ignored here.
3684 if (th->doff == (sizeof(*th) / 4)) {
3685 tp->rx_opt.saw_tstamp = 0;
3687 } else if (tp->rx_opt.tstamp_ok &&
3688 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3689 if (tcp_parse_aligned_timestamp(tp, th))
3693 tcp_parse_options(skb, &tp->rx_opt, 1, NULL);
3694 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3695 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3700 #ifdef CONFIG_TCP_MD5SIG
3702 * Parse MD5 Signature option
3704 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3706 int length = (th->doff << 2) - sizeof(*th);
3707 const u8 *ptr = (const u8 *)(th + 1);
3709 /* If the TCP option is too short, we can short cut */
3710 if (length < TCPOLEN_MD5SIG)
3713 while (length > 0) {
3714 int opcode = *ptr++;
3725 if (opsize < 2 || opsize > length)
3727 if (opcode == TCPOPT_MD5SIG)
3728 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3735 EXPORT_SYMBOL(tcp_parse_md5sig_option);
3738 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3740 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3741 * it can pass through stack. So, the following predicate verifies that
3742 * this segment is not used for anything but congestion avoidance or
3743 * fast retransmit. Moreover, we even are able to eliminate most of such
3744 * second order effects, if we apply some small "replay" window (~RTO)
3745 * to timestamp space.
3747 * All these measures still do not guarantee that we reject wrapped ACKs
3748 * on networks with high bandwidth, when sequence space is recycled fastly,
3749 * but it guarantees that such events will be very rare and do not affect
3750 * connection seriously. This doesn't look nice, but alas, PAWS is really
3753 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3754 * states that events when retransmit arrives after original data are rare.
3755 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3756 * the biggest problem on large power networks even with minor reordering.
3757 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3758 * up to bandwidth of 18Gigabit/sec. 8) ]
3761 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3763 const struct tcp_sock *tp = tcp_sk(sk);
3764 const struct tcphdr *th = tcp_hdr(skb);
3765 u32 seq = TCP_SKB_CB(skb)->seq;
3766 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3768 return (/* 1. Pure ACK with correct sequence number. */
3769 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3771 /* 2. ... and duplicate ACK. */
3772 ack == tp->snd_una &&
3774 /* 3. ... and does not update window. */
3775 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3777 /* 4. ... and sits in replay window. */
3778 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3781 static inline bool tcp_paws_discard(const struct sock *sk,
3782 const struct sk_buff *skb)
3784 const struct tcp_sock *tp = tcp_sk(sk);
3786 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
3787 !tcp_disordered_ack(sk, skb);
3790 /* Check segment sequence number for validity.
3792 * Segment controls are considered valid, if the segment
3793 * fits to the window after truncation to the window. Acceptability
3794 * of data (and SYN, FIN, of course) is checked separately.
3795 * See tcp_data_queue(), for example.
3797 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3798 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3799 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3800 * (borrowed from freebsd)
3803 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
3805 return !before(end_seq, tp->rcv_wup) &&
3806 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3809 /* When we get a reset we do this. */
3810 void tcp_reset(struct sock *sk)
3812 /* We want the right error as BSD sees it (and indeed as we do). */
3813 switch (sk->sk_state) {
3815 sk->sk_err = ECONNREFUSED;
3817 case TCP_CLOSE_WAIT:
3823 sk->sk_err = ECONNRESET;
3825 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3828 if (!sock_flag(sk, SOCK_DEAD))
3829 sk->sk_error_report(sk);
3835 * Process the FIN bit. This now behaves as it is supposed to work
3836 * and the FIN takes effect when it is validly part of sequence
3837 * space. Not before when we get holes.
3839 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3840 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3843 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3844 * close and we go into CLOSING (and later onto TIME-WAIT)
3846 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3848 static void tcp_fin(struct sock *sk)
3850 struct tcp_sock *tp = tcp_sk(sk);
3851 const struct dst_entry *dst;
3853 inet_csk_schedule_ack(sk);
3855 sk->sk_shutdown |= RCV_SHUTDOWN;
3856 sock_set_flag(sk, SOCK_DONE);
3858 switch (sk->sk_state) {
3860 case TCP_ESTABLISHED:
3861 /* Move to CLOSE_WAIT */
3862 tcp_set_state(sk, TCP_CLOSE_WAIT);
3863 dst = __sk_dst_get(sk);
3864 if (!dst || !dst_metric(dst, RTAX_QUICKACK))
3865 inet_csk(sk)->icsk_ack.pingpong = 1;
3868 case TCP_CLOSE_WAIT:
3870 /* Received a retransmission of the FIN, do
3875 /* RFC793: Remain in the LAST-ACK state. */
3879 /* This case occurs when a simultaneous close
3880 * happens, we must ack the received FIN and
3881 * enter the CLOSING state.
3884 tcp_set_state(sk, TCP_CLOSING);
3887 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3889 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
3892 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3893 * cases we should never reach this piece of code.
3895 pr_err("%s: Impossible, sk->sk_state=%d\n",
3896 __func__, sk->sk_state);
3900 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3901 * Probably, we should reset in this case. For now drop them.
3903 __skb_queue_purge(&tp->out_of_order_queue);
3904 if (tcp_is_sack(tp))
3905 tcp_sack_reset(&tp->rx_opt);
3908 if (!sock_flag(sk, SOCK_DEAD)) {
3909 sk->sk_state_change(sk);
3911 /* Do not send POLL_HUP for half duplex close. */
3912 if (sk->sk_shutdown == SHUTDOWN_MASK ||
3913 sk->sk_state == TCP_CLOSE)
3914 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
3916 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
3920 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
3923 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
3924 if (before(seq, sp->start_seq))
3925 sp->start_seq = seq;
3926 if (after(end_seq, sp->end_seq))
3927 sp->end_seq = end_seq;
3933 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
3935 struct tcp_sock *tp = tcp_sk(sk);
3937 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3940 if (before(seq, tp->rcv_nxt))
3941 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
3943 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
3945 NET_INC_STATS_BH(sock_net(sk), mib_idx);
3947 tp->rx_opt.dsack = 1;
3948 tp->duplicate_sack[0].start_seq = seq;
3949 tp->duplicate_sack[0].end_seq = end_seq;
3953 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
3955 struct tcp_sock *tp = tcp_sk(sk);
3957 if (!tp->rx_opt.dsack)
3958 tcp_dsack_set(sk, seq, end_seq);
3960 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
3963 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
3965 struct tcp_sock *tp = tcp_sk(sk);
3967 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
3968 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3969 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
3970 tcp_enter_quickack_mode(sk);
3972 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3973 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3975 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
3976 end_seq = tp->rcv_nxt;
3977 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
3984 /* These routines update the SACK block as out-of-order packets arrive or
3985 * in-order packets close up the sequence space.
3987 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
3990 struct tcp_sack_block *sp = &tp->selective_acks[0];
3991 struct tcp_sack_block *swalk = sp + 1;
3993 /* See if the recent change to the first SACK eats into
3994 * or hits the sequence space of other SACK blocks, if so coalesce.
3996 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
3997 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4000 /* Zap SWALK, by moving every further SACK up by one slot.
4001 * Decrease num_sacks.
4003 tp->rx_opt.num_sacks--;
4004 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4008 this_sack++, swalk++;
4012 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4014 struct tcp_sock *tp = tcp_sk(sk);
4015 struct tcp_sack_block *sp = &tp->selective_acks[0];
4016 int cur_sacks = tp->rx_opt.num_sacks;
4022 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4023 if (tcp_sack_extend(sp, seq, end_seq)) {
4024 /* Rotate this_sack to the first one. */
4025 for (; this_sack > 0; this_sack--, sp--)
4026 swap(*sp, *(sp - 1));
4028 tcp_sack_maybe_coalesce(tp);
4033 /* Could not find an adjacent existing SACK, build a new one,
4034 * put it at the front, and shift everyone else down. We
4035 * always know there is at least one SACK present already here.
4037 * If the sack array is full, forget about the last one.
4039 if (this_sack >= TCP_NUM_SACKS) {
4041 tp->rx_opt.num_sacks--;
4044 for (; this_sack > 0; this_sack--, sp--)
4048 /* Build the new head SACK, and we're done. */
4049 sp->start_seq = seq;
4050 sp->end_seq = end_seq;
4051 tp->rx_opt.num_sacks++;
4054 /* RCV.NXT advances, some SACKs should be eaten. */
4056 static void tcp_sack_remove(struct tcp_sock *tp)
4058 struct tcp_sack_block *sp = &tp->selective_acks[0];
4059 int num_sacks = tp->rx_opt.num_sacks;
4062 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4063 if (skb_queue_empty(&tp->out_of_order_queue)) {
4064 tp->rx_opt.num_sacks = 0;
4068 for (this_sack = 0; this_sack < num_sacks;) {
4069 /* Check if the start of the sack is covered by RCV.NXT. */
4070 if (!before(tp->rcv_nxt, sp->start_seq)) {
4073 /* RCV.NXT must cover all the block! */
4074 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4076 /* Zap this SACK, by moving forward any other SACKS. */
4077 for (i = this_sack+1; i < num_sacks; i++)
4078 tp->selective_acks[i-1] = tp->selective_acks[i];
4085 tp->rx_opt.num_sacks = num_sacks;
4089 * tcp_try_coalesce - try to merge skb to prior one
4092 * @from: buffer to add in queue
4093 * @fragstolen: pointer to boolean
4095 * Before queueing skb @from after @to, try to merge them
4096 * to reduce overall memory use and queue lengths, if cost is small.
4097 * Packets in ofo or receive queues can stay a long time.
4098 * Better try to coalesce them right now to avoid future collapses.
4099 * Returns true if caller should free @from instead of queueing it
4101 static bool tcp_try_coalesce(struct sock *sk,
4103 struct sk_buff *from,
4108 *fragstolen = false;
4110 /* Its possible this segment overlaps with prior segment in queue */
4111 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4114 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4117 atomic_add(delta, &sk->sk_rmem_alloc);
4118 sk_mem_charge(sk, delta);
4119 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4120 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4121 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4122 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4126 /* This one checks to see if we can put data from the
4127 * out_of_order queue into the receive_queue.
4129 static void tcp_ofo_queue(struct sock *sk)
4131 struct tcp_sock *tp = tcp_sk(sk);
4132 __u32 dsack_high = tp->rcv_nxt;
4133 struct sk_buff *skb, *tail;
4134 bool fragstolen, eaten;
4136 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4137 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4140 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4141 __u32 dsack = dsack_high;
4142 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4143 dsack_high = TCP_SKB_CB(skb)->end_seq;
4144 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4147 __skb_unlink(skb, &tp->out_of_order_queue);
4148 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4149 SOCK_DEBUG(sk, "ofo packet was already received\n");
4153 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4154 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4155 TCP_SKB_CB(skb)->end_seq);
4157 tail = skb_peek_tail(&sk->sk_receive_queue);
4158 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4159 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4161 __skb_queue_tail(&sk->sk_receive_queue, skb);
4162 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4165 kfree_skb_partial(skb, fragstolen);
4169 static bool tcp_prune_ofo_queue(struct sock *sk);
4170 static int tcp_prune_queue(struct sock *sk);
4172 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4175 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4176 !sk_rmem_schedule(sk, skb, size)) {
4178 if (tcp_prune_queue(sk) < 0)
4181 if (!sk_rmem_schedule(sk, skb, size)) {
4182 if (!tcp_prune_ofo_queue(sk))
4185 if (!sk_rmem_schedule(sk, skb, size))
4192 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4194 struct tcp_sock *tp = tcp_sk(sk);
4195 struct sk_buff *skb1;
4198 tcp_ecn_check_ce(tp, skb);
4200 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4201 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFODROP);
4206 /* Disable header prediction. */
4208 inet_csk_schedule_ack(sk);
4210 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4211 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4212 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4214 skb1 = skb_peek_tail(&tp->out_of_order_queue);
4216 /* Initial out of order segment, build 1 SACK. */
4217 if (tcp_is_sack(tp)) {
4218 tp->rx_opt.num_sacks = 1;
4219 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4220 tp->selective_acks[0].end_seq =
4221 TCP_SKB_CB(skb)->end_seq;
4223 __skb_queue_head(&tp->out_of_order_queue, skb);
4227 seq = TCP_SKB_CB(skb)->seq;
4228 end_seq = TCP_SKB_CB(skb)->end_seq;
4230 if (seq == TCP_SKB_CB(skb1)->end_seq) {
4233 if (!tcp_try_coalesce(sk, skb1, skb, &fragstolen)) {
4234 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4236 tcp_grow_window(sk, skb);
4237 kfree_skb_partial(skb, fragstolen);
4241 if (!tp->rx_opt.num_sacks ||
4242 tp->selective_acks[0].end_seq != seq)
4245 /* Common case: data arrive in order after hole. */
4246 tp->selective_acks[0].end_seq = end_seq;
4250 /* Find place to insert this segment. */
4252 if (!after(TCP_SKB_CB(skb1)->seq, seq))
4254 if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
4258 skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
4261 /* Do skb overlap to previous one? */
4262 if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4263 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4264 /* All the bits are present. Drop. */
4265 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4268 tcp_dsack_set(sk, seq, end_seq);
4271 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4272 /* Partial overlap. */
4273 tcp_dsack_set(sk, seq,
4274 TCP_SKB_CB(skb1)->end_seq);
4276 if (skb_queue_is_first(&tp->out_of_order_queue,
4280 skb1 = skb_queue_prev(
4281 &tp->out_of_order_queue,
4286 __skb_queue_head(&tp->out_of_order_queue, skb);
4288 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4290 /* And clean segments covered by new one as whole. */
4291 while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
4292 skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
4294 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4296 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4297 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4301 __skb_unlink(skb1, &tp->out_of_order_queue);
4302 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4303 TCP_SKB_CB(skb1)->end_seq);
4304 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4309 if (tcp_is_sack(tp))
4310 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4313 tcp_grow_window(sk, skb);
4314 skb_set_owner_r(skb, sk);
4318 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
4322 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4324 __skb_pull(skb, hdrlen);
4326 tcp_try_coalesce(sk, tail, skb, fragstolen)) ? 1 : 0;
4327 tcp_sk(sk)->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4329 __skb_queue_tail(&sk->sk_receive_queue, skb);
4330 skb_set_owner_r(skb, sk);
4335 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4337 struct sk_buff *skb;
4343 skb = alloc_skb(size, sk->sk_allocation);
4347 if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4350 if (memcpy_fromiovec(skb_put(skb, size), msg->msg_iov, size))
4353 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4354 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4355 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4357 if (tcp_queue_rcv(sk, skb, 0, &fragstolen)) {
4358 WARN_ON_ONCE(fragstolen); /* should not happen */
4369 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4371 struct tcp_sock *tp = tcp_sk(sk);
4373 bool fragstolen = false;
4375 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4379 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
4381 tcp_ecn_accept_cwr(tp, skb);
4383 tp->rx_opt.dsack = 0;
4385 /* Queue data for delivery to the user.
4386 * Packets in sequence go to the receive queue.
4387 * Out of sequence packets to the out_of_order_queue.
4389 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4390 if (tcp_receive_window(tp) == 0)
4393 /* Ok. In sequence. In window. */
4394 if (tp->ucopy.task == current &&
4395 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4396 sock_owned_by_user(sk) && !tp->urg_data) {
4397 int chunk = min_t(unsigned int, skb->len,
4400 __set_current_state(TASK_RUNNING);
4403 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
4404 tp->ucopy.len -= chunk;
4405 tp->copied_seq += chunk;
4406 eaten = (chunk == skb->len);
4407 tcp_rcv_space_adjust(sk);
4415 tcp_try_rmem_schedule(sk, skb, skb->truesize))
4418 eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
4420 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4422 tcp_event_data_recv(sk, skb);
4423 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4426 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4429 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4430 * gap in queue is filled.
4432 if (skb_queue_empty(&tp->out_of_order_queue))
4433 inet_csk(sk)->icsk_ack.pingpong = 0;
4436 if (tp->rx_opt.num_sacks)
4437 tcp_sack_remove(tp);
4439 tcp_fast_path_check(sk);
4442 kfree_skb_partial(skb, fragstolen);
4443 if (!sock_flag(sk, SOCK_DEAD))
4444 sk->sk_data_ready(sk);
4448 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4449 /* A retransmit, 2nd most common case. Force an immediate ack. */
4450 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4451 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4454 tcp_enter_quickack_mode(sk);
4455 inet_csk_schedule_ack(sk);
4461 /* Out of window. F.e. zero window probe. */
4462 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4465 tcp_enter_quickack_mode(sk);
4467 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4468 /* Partial packet, seq < rcv_next < end_seq */
4469 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4470 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4471 TCP_SKB_CB(skb)->end_seq);
4473 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4475 /* If window is closed, drop tail of packet. But after
4476 * remembering D-SACK for its head made in previous line.
4478 if (!tcp_receive_window(tp))
4483 tcp_data_queue_ofo(sk, skb);
4486 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4487 struct sk_buff_head *list)
4489 struct sk_buff *next = NULL;
4491 if (!skb_queue_is_last(list, skb))
4492 next = skb_queue_next(list, skb);
4494 __skb_unlink(skb, list);
4496 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4501 /* Collapse contiguous sequence of skbs head..tail with
4502 * sequence numbers start..end.
4504 * If tail is NULL, this means until the end of the list.
4506 * Segments with FIN/SYN are not collapsed (only because this
4510 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4511 struct sk_buff *head, struct sk_buff *tail,
4514 struct sk_buff *skb, *n;
4517 /* First, check that queue is collapsible and find
4518 * the point where collapsing can be useful. */
4522 skb_queue_walk_from_safe(list, skb, n) {
4525 /* No new bits? It is possible on ofo queue. */
4526 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4527 skb = tcp_collapse_one(sk, skb, list);
4533 /* The first skb to collapse is:
4535 * - bloated or contains data before "start" or
4536 * overlaps to the next one.
4538 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
4539 (tcp_win_from_space(skb->truesize) > skb->len ||
4540 before(TCP_SKB_CB(skb)->seq, start))) {
4541 end_of_skbs = false;
4545 if (!skb_queue_is_last(list, skb)) {
4546 struct sk_buff *next = skb_queue_next(list, skb);
4548 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
4549 end_of_skbs = false;
4554 /* Decided to skip this, advance start seq. */
4555 start = TCP_SKB_CB(skb)->end_seq;
4558 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4561 while (before(start, end)) {
4562 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
4563 struct sk_buff *nskb;
4565 nskb = alloc_skb(copy, GFP_ATOMIC);
4569 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4570 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4571 __skb_queue_before(list, skb, nskb);
4572 skb_set_owner_r(nskb, sk);
4574 /* Copy data, releasing collapsed skbs. */
4576 int offset = start - TCP_SKB_CB(skb)->seq;
4577 int size = TCP_SKB_CB(skb)->end_seq - start;
4581 size = min(copy, size);
4582 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4584 TCP_SKB_CB(nskb)->end_seq += size;
4588 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4589 skb = tcp_collapse_one(sk, skb, list);
4592 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4599 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4600 * and tcp_collapse() them until all the queue is collapsed.
4602 static void tcp_collapse_ofo_queue(struct sock *sk)
4604 struct tcp_sock *tp = tcp_sk(sk);
4605 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4606 struct sk_buff *head;
4612 start = TCP_SKB_CB(skb)->seq;
4613 end = TCP_SKB_CB(skb)->end_seq;
4617 struct sk_buff *next = NULL;
4619 if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
4620 next = skb_queue_next(&tp->out_of_order_queue, skb);
4623 /* Segment is terminated when we see gap or when
4624 * we are at the end of all the queue. */
4626 after(TCP_SKB_CB(skb)->seq, end) ||
4627 before(TCP_SKB_CB(skb)->end_seq, start)) {
4628 tcp_collapse(sk, &tp->out_of_order_queue,
4629 head, skb, start, end);
4633 /* Start new segment */
4634 start = TCP_SKB_CB(skb)->seq;
4635 end = TCP_SKB_CB(skb)->end_seq;
4637 if (before(TCP_SKB_CB(skb)->seq, start))
4638 start = TCP_SKB_CB(skb)->seq;
4639 if (after(TCP_SKB_CB(skb)->end_seq, end))
4640 end = TCP_SKB_CB(skb)->end_seq;
4646 * Purge the out-of-order queue.
4647 * Return true if queue was pruned.
4649 static bool tcp_prune_ofo_queue(struct sock *sk)
4651 struct tcp_sock *tp = tcp_sk(sk);
4654 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4655 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
4656 __skb_queue_purge(&tp->out_of_order_queue);
4658 /* Reset SACK state. A conforming SACK implementation will
4659 * do the same at a timeout based retransmit. When a connection
4660 * is in a sad state like this, we care only about integrity
4661 * of the connection not performance.
4663 if (tp->rx_opt.sack_ok)
4664 tcp_sack_reset(&tp->rx_opt);
4671 /* Reduce allocated memory if we can, trying to get
4672 * the socket within its memory limits again.
4674 * Return less than zero if we should start dropping frames
4675 * until the socket owning process reads some of the data
4676 * to stabilize the situation.
4678 static int tcp_prune_queue(struct sock *sk)
4680 struct tcp_sock *tp = tcp_sk(sk);
4682 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4684 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
4686 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4687 tcp_clamp_window(sk);
4688 else if (sk_under_memory_pressure(sk))
4689 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4691 tcp_collapse_ofo_queue(sk);
4692 if (!skb_queue_empty(&sk->sk_receive_queue))
4693 tcp_collapse(sk, &sk->sk_receive_queue,
4694 skb_peek(&sk->sk_receive_queue),
4696 tp->copied_seq, tp->rcv_nxt);
4699 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4702 /* Collapsing did not help, destructive actions follow.
4703 * This must not ever occur. */
4705 tcp_prune_ofo_queue(sk);
4707 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4710 /* If we are really being abused, tell the caller to silently
4711 * drop receive data on the floor. It will get retransmitted
4712 * and hopefully then we'll have sufficient space.
4714 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
4716 /* Massive buffer overcommit. */
4721 static bool tcp_should_expand_sndbuf(const struct sock *sk)
4723 const struct tcp_sock *tp = tcp_sk(sk);
4725 /* If the user specified a specific send buffer setting, do
4728 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4731 /* If we are under global TCP memory pressure, do not expand. */
4732 if (sk_under_memory_pressure(sk))
4735 /* If we are under soft global TCP memory pressure, do not expand. */
4736 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
4739 /* If we filled the congestion window, do not expand. */
4740 if (tp->packets_out >= tp->snd_cwnd)
4746 /* When incoming ACK allowed to free some skb from write_queue,
4747 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4748 * on the exit from tcp input handler.
4750 * PROBLEM: sndbuf expansion does not work well with largesend.
4752 static void tcp_new_space(struct sock *sk)
4754 struct tcp_sock *tp = tcp_sk(sk);
4756 if (tcp_should_expand_sndbuf(sk)) {
4757 tcp_sndbuf_expand(sk);
4758 tp->snd_cwnd_stamp = tcp_time_stamp;
4761 sk->sk_write_space(sk);
4764 static void tcp_check_space(struct sock *sk)
4766 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4767 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4768 if (sk->sk_socket &&
4769 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4774 static inline void tcp_data_snd_check(struct sock *sk)
4776 tcp_push_pending_frames(sk);
4777 tcp_check_space(sk);
4781 * Check if sending an ack is needed.
4783 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4785 struct tcp_sock *tp = tcp_sk(sk);
4787 /* More than one full frame received... */
4788 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
4789 /* ... and right edge of window advances far enough.
4790 * (tcp_recvmsg() will send ACK otherwise). Or...
4792 __tcp_select_window(sk) >= tp->rcv_wnd) ||
4793 /* We ACK each frame or... */
4794 tcp_in_quickack_mode(sk) ||
4795 /* We have out of order data. */
4796 (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
4797 /* Then ack it now */
4800 /* Else, send delayed ack. */
4801 tcp_send_delayed_ack(sk);
4805 static inline void tcp_ack_snd_check(struct sock *sk)
4807 if (!inet_csk_ack_scheduled(sk)) {
4808 /* We sent a data segment already. */
4811 __tcp_ack_snd_check(sk, 1);
4815 * This routine is only called when we have urgent data
4816 * signaled. Its the 'slow' part of tcp_urg. It could be
4817 * moved inline now as tcp_urg is only called from one
4818 * place. We handle URGent data wrong. We have to - as
4819 * BSD still doesn't use the correction from RFC961.
4820 * For 1003.1g we should support a new option TCP_STDURG to permit
4821 * either form (or just set the sysctl tcp_stdurg).
4824 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
4826 struct tcp_sock *tp = tcp_sk(sk);
4827 u32 ptr = ntohs(th->urg_ptr);
4829 if (ptr && !sysctl_tcp_stdurg)
4831 ptr += ntohl(th->seq);
4833 /* Ignore urgent data that we've already seen and read. */
4834 if (after(tp->copied_seq, ptr))
4837 /* Do not replay urg ptr.
4839 * NOTE: interesting situation not covered by specs.
4840 * Misbehaving sender may send urg ptr, pointing to segment,
4841 * which we already have in ofo queue. We are not able to fetch
4842 * such data and will stay in TCP_URG_NOTYET until will be eaten
4843 * by recvmsg(). Seems, we are not obliged to handle such wicked
4844 * situations. But it is worth to think about possibility of some
4845 * DoSes using some hypothetical application level deadlock.
4847 if (before(ptr, tp->rcv_nxt))
4850 /* Do we already have a newer (or duplicate) urgent pointer? */
4851 if (tp->urg_data && !after(ptr, tp->urg_seq))
4854 /* Tell the world about our new urgent pointer. */
4857 /* We may be adding urgent data when the last byte read was
4858 * urgent. To do this requires some care. We cannot just ignore
4859 * tp->copied_seq since we would read the last urgent byte again
4860 * as data, nor can we alter copied_seq until this data arrives
4861 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4863 * NOTE. Double Dutch. Rendering to plain English: author of comment
4864 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4865 * and expect that both A and B disappear from stream. This is _wrong_.
4866 * Though this happens in BSD with high probability, this is occasional.
4867 * Any application relying on this is buggy. Note also, that fix "works"
4868 * only in this artificial test. Insert some normal data between A and B and we will
4869 * decline of BSD again. Verdict: it is better to remove to trap
4872 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
4873 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
4874 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4876 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
4877 __skb_unlink(skb, &sk->sk_receive_queue);
4882 tp->urg_data = TCP_URG_NOTYET;
4885 /* Disable header prediction. */
4889 /* This is the 'fast' part of urgent handling. */
4890 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
4892 struct tcp_sock *tp = tcp_sk(sk);
4894 /* Check if we get a new urgent pointer - normally not. */
4896 tcp_check_urg(sk, th);
4898 /* Do we wait for any urgent data? - normally not... */
4899 if (tp->urg_data == TCP_URG_NOTYET) {
4900 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
4903 /* Is the urgent pointer pointing into this packet? */
4904 if (ptr < skb->len) {
4906 if (skb_copy_bits(skb, ptr, &tmp, 1))
4908 tp->urg_data = TCP_URG_VALID | tmp;
4909 if (!sock_flag(sk, SOCK_DEAD))
4910 sk->sk_data_ready(sk);
4915 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4917 struct tcp_sock *tp = tcp_sk(sk);
4918 int chunk = skb->len - hlen;
4922 if (skb_csum_unnecessary(skb))
4923 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
4925 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
4929 tp->ucopy.len -= chunk;
4930 tp->copied_seq += chunk;
4931 tcp_rcv_space_adjust(sk);
4938 static __sum16 __tcp_checksum_complete_user(struct sock *sk,
4939 struct sk_buff *skb)
4943 if (sock_owned_by_user(sk)) {
4945 result = __tcp_checksum_complete(skb);
4948 result = __tcp_checksum_complete(skb);
4953 static inline bool tcp_checksum_complete_user(struct sock *sk,
4954 struct sk_buff *skb)
4956 return !skb_csum_unnecessary(skb) &&
4957 __tcp_checksum_complete_user(sk, skb);
4960 /* Does PAWS and seqno based validation of an incoming segment, flags will
4961 * play significant role here.
4963 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
4964 const struct tcphdr *th, int syn_inerr)
4966 struct tcp_sock *tp = tcp_sk(sk);
4968 /* RFC1323: H1. Apply PAWS check first. */
4969 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4970 tcp_paws_discard(sk, skb)) {
4972 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
4973 tcp_send_dupack(sk, skb);
4976 /* Reset is accepted even if it did not pass PAWS. */
4979 /* Step 1: check sequence number */
4980 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4981 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4982 * (RST) segments are validated by checking their SEQ-fields."
4983 * And page 69: "If an incoming segment is not acceptable,
4984 * an acknowledgment should be sent in reply (unless the RST
4985 * bit is set, if so drop the segment and return)".
4990 tcp_send_dupack(sk, skb);
4995 /* Step 2: check RST bit */
4998 * If sequence number exactly matches RCV.NXT, then
4999 * RESET the connection
5001 * Send a challenge ACK
5003 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt)
5006 tcp_send_challenge_ack(sk);
5010 /* step 3: check security and precedence [ignored] */
5012 /* step 4: Check for a SYN
5013 * RFC 5691 4.2 : Send a challenge ack
5018 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5019 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5020 tcp_send_challenge_ack(sk);
5032 * TCP receive function for the ESTABLISHED state.
5034 * It is split into a fast path and a slow path. The fast path is
5036 * - A zero window was announced from us - zero window probing
5037 * is only handled properly in the slow path.
5038 * - Out of order segments arrived.
5039 * - Urgent data is expected.
5040 * - There is no buffer space left
5041 * - Unexpected TCP flags/window values/header lengths are received
5042 * (detected by checking the TCP header against pred_flags)
5043 * - Data is sent in both directions. Fast path only supports pure senders
5044 * or pure receivers (this means either the sequence number or the ack
5045 * value must stay constant)
5046 * - Unexpected TCP option.
5048 * When these conditions are not satisfied it drops into a standard
5049 * receive procedure patterned after RFC793 to handle all cases.
5050 * The first three cases are guaranteed by proper pred_flags setting,
5051 * the rest is checked inline. Fast processing is turned on in
5052 * tcp_data_queue when everything is OK.
5054 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5055 const struct tcphdr *th, unsigned int len)
5057 struct tcp_sock *tp = tcp_sk(sk);
5059 if (unlikely(sk->sk_rx_dst == NULL))
5060 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5062 * Header prediction.
5063 * The code loosely follows the one in the famous
5064 * "30 instruction TCP receive" Van Jacobson mail.
5066 * Van's trick is to deposit buffers into socket queue
5067 * on a device interrupt, to call tcp_recv function
5068 * on the receive process context and checksum and copy
5069 * the buffer to user space. smart...
5071 * Our current scheme is not silly either but we take the
5072 * extra cost of the net_bh soft interrupt processing...
5073 * We do checksum and copy also but from device to kernel.
5076 tp->rx_opt.saw_tstamp = 0;
5078 /* pred_flags is 0xS?10 << 16 + snd_wnd
5079 * if header_prediction is to be made
5080 * 'S' will always be tp->tcp_header_len >> 2
5081 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5082 * turn it off (when there are holes in the receive
5083 * space for instance)
5084 * PSH flag is ignored.
5087 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5088 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5089 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5090 int tcp_header_len = tp->tcp_header_len;
5092 /* Timestamp header prediction: tcp_header_len
5093 * is automatically equal to th->doff*4 due to pred_flags
5097 /* Check timestamp */
5098 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5099 /* No? Slow path! */
5100 if (!tcp_parse_aligned_timestamp(tp, th))
5103 /* If PAWS failed, check it more carefully in slow path */
5104 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5107 /* DO NOT update ts_recent here, if checksum fails
5108 * and timestamp was corrupted part, it will result
5109 * in a hung connection since we will drop all
5110 * future packets due to the PAWS test.
5114 if (len <= tcp_header_len) {
5115 /* Bulk data transfer: sender */
5116 if (len == tcp_header_len) {
5117 /* Predicted packet is in window by definition.
5118 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5119 * Hence, check seq<=rcv_wup reduces to:
5121 if (tcp_header_len ==
5122 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5123 tp->rcv_nxt == tp->rcv_wup)
5124 tcp_store_ts_recent(tp);
5126 /* We know that such packets are checksummed
5129 tcp_ack(sk, skb, 0);
5131 tcp_data_snd_check(sk);
5133 } else { /* Header too small */
5134 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5139 bool fragstolen = false;
5141 if (tp->ucopy.task == current &&
5142 tp->copied_seq == tp->rcv_nxt &&
5143 len - tcp_header_len <= tp->ucopy.len &&
5144 sock_owned_by_user(sk)) {
5145 __set_current_state(TASK_RUNNING);
5147 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) {
5148 /* Predicted packet is in window by definition.
5149 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5150 * Hence, check seq<=rcv_wup reduces to:
5152 if (tcp_header_len ==
5153 (sizeof(struct tcphdr) +
5154 TCPOLEN_TSTAMP_ALIGNED) &&
5155 tp->rcv_nxt == tp->rcv_wup)
5156 tcp_store_ts_recent(tp);
5158 tcp_rcv_rtt_measure_ts(sk, skb);
5160 __skb_pull(skb, tcp_header_len);
5161 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5162 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
5167 if (tcp_checksum_complete_user(sk, skb))
5170 if ((int)skb->truesize > sk->sk_forward_alloc)
5173 /* Predicted packet is in window by definition.
5174 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5175 * Hence, check seq<=rcv_wup reduces to:
5177 if (tcp_header_len ==
5178 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5179 tp->rcv_nxt == tp->rcv_wup)
5180 tcp_store_ts_recent(tp);
5182 tcp_rcv_rtt_measure_ts(sk, skb);
5184 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
5186 /* Bulk data transfer: receiver */
5187 eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5191 tcp_event_data_recv(sk, skb);
5193 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5194 /* Well, only one small jumplet in fast path... */
5195 tcp_ack(sk, skb, FLAG_DATA);
5196 tcp_data_snd_check(sk);
5197 if (!inet_csk_ack_scheduled(sk))
5201 __tcp_ack_snd_check(sk, 0);
5204 kfree_skb_partial(skb, fragstolen);
5205 sk->sk_data_ready(sk);
5211 if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
5214 if (!th->ack && !th->rst)
5218 * Standard slow path.
5221 if (!tcp_validate_incoming(sk, skb, th, 1))
5225 if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5228 tcp_rcv_rtt_measure_ts(sk, skb);
5230 /* Process urgent data. */
5231 tcp_urg(sk, skb, th);
5233 /* step 7: process the segment text */
5234 tcp_data_queue(sk, skb);
5236 tcp_data_snd_check(sk);
5237 tcp_ack_snd_check(sk);
5241 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_CSUMERRORS);
5242 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5247 EXPORT_SYMBOL(tcp_rcv_established);
5249 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5251 struct tcp_sock *tp = tcp_sk(sk);
5252 struct inet_connection_sock *icsk = inet_csk(sk);
5254 tcp_set_state(sk, TCP_ESTABLISHED);
5257 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5258 security_inet_conn_established(sk, skb);
5261 /* Make sure socket is routed, for correct metrics. */
5262 icsk->icsk_af_ops->rebuild_header(sk);
5264 tcp_init_metrics(sk);
5266 tcp_init_congestion_control(sk);
5268 /* Prevent spurious tcp_cwnd_restart() on first data
5271 tp->lsndtime = tcp_time_stamp;
5273 tcp_init_buffer_space(sk);
5275 if (sock_flag(sk, SOCK_KEEPOPEN))
5276 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5278 if (!tp->rx_opt.snd_wscale)
5279 __tcp_fast_path_on(tp, tp->snd_wnd);
5283 if (!sock_flag(sk, SOCK_DEAD)) {
5284 sk->sk_state_change(sk);
5285 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5289 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5290 struct tcp_fastopen_cookie *cookie)
5292 struct tcp_sock *tp = tcp_sk(sk);
5293 struct sk_buff *data = tp->syn_data ? tcp_write_queue_head(sk) : NULL;
5294 u16 mss = tp->rx_opt.mss_clamp;
5297 if (mss == tp->rx_opt.user_mss) {
5298 struct tcp_options_received opt;
5300 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5301 tcp_clear_options(&opt);
5302 opt.user_mss = opt.mss_clamp = 0;
5303 tcp_parse_options(synack, &opt, 0, NULL);
5304 mss = opt.mss_clamp;
5307 if (!tp->syn_fastopen) /* Ignore an unsolicited cookie */
5310 /* The SYN-ACK neither has cookie nor acknowledges the data. Presumably
5311 * the remote receives only the retransmitted (regular) SYNs: either
5312 * the original SYN-data or the corresponding SYN-ACK is lost.
5314 syn_drop = (cookie->len <= 0 && data && tp->total_retrans);
5316 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop);
5318 if (data) { /* Retransmit unacked data in SYN */
5319 tcp_for_write_queue_from(data, sk) {
5320 if (data == tcp_send_head(sk) ||
5321 __tcp_retransmit_skb(sk, data))
5325 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVEFAIL);
5328 tp->syn_data_acked = tp->syn_data;
5329 if (tp->syn_data_acked)
5330 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
5334 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5335 const struct tcphdr *th, unsigned int len)
5337 struct inet_connection_sock *icsk = inet_csk(sk);
5338 struct tcp_sock *tp = tcp_sk(sk);
5339 struct tcp_fastopen_cookie foc = { .len = -1 };
5340 int saved_clamp = tp->rx_opt.mss_clamp;
5342 tcp_parse_options(skb, &tp->rx_opt, 0, &foc);
5343 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
5344 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
5348 * "If the state is SYN-SENT then
5349 * first check the ACK bit
5350 * If the ACK bit is set
5351 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5352 * a reset (unless the RST bit is set, if so drop
5353 * the segment and return)"
5355 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5356 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
5357 goto reset_and_undo;
5359 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5360 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5362 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
5363 goto reset_and_undo;
5366 /* Now ACK is acceptable.
5368 * "If the RST bit is set
5369 * If the ACK was acceptable then signal the user "error:
5370 * connection reset", drop the segment, enter CLOSED state,
5371 * delete TCB, and return."
5380 * "fifth, if neither of the SYN or RST bits is set then
5381 * drop the segment and return."
5387 goto discard_and_undo;
5390 * "If the SYN bit is on ...
5391 * are acceptable then ...
5392 * (our SYN has been ACKed), change the connection
5393 * state to ESTABLISHED..."
5396 tcp_ecn_rcv_synack(tp, th);
5398 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5399 tcp_ack(sk, skb, FLAG_SLOWPATH);
5401 /* Ok.. it's good. Set up sequence numbers and
5402 * move to established.
5404 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5405 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5407 /* RFC1323: The window in SYN & SYN/ACK segments is
5410 tp->snd_wnd = ntohs(th->window);
5412 if (!tp->rx_opt.wscale_ok) {
5413 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5414 tp->window_clamp = min(tp->window_clamp, 65535U);
5417 if (tp->rx_opt.saw_tstamp) {
5418 tp->rx_opt.tstamp_ok = 1;
5419 tp->tcp_header_len =
5420 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5421 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5422 tcp_store_ts_recent(tp);
5424 tp->tcp_header_len = sizeof(struct tcphdr);
5427 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5428 tcp_enable_fack(tp);
5431 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5432 tcp_initialize_rcv_mss(sk);
5434 /* Remember, tcp_poll() does not lock socket!
5435 * Change state from SYN-SENT only after copied_seq
5436 * is initialized. */
5437 tp->copied_seq = tp->rcv_nxt;
5441 tcp_finish_connect(sk, skb);
5443 if ((tp->syn_fastopen || tp->syn_data) &&
5444 tcp_rcv_fastopen_synack(sk, skb, &foc))
5447 if (sk->sk_write_pending ||
5448 icsk->icsk_accept_queue.rskq_defer_accept ||
5449 icsk->icsk_ack.pingpong) {
5450 /* Save one ACK. Data will be ready after
5451 * several ticks, if write_pending is set.
5453 * It may be deleted, but with this feature tcpdumps
5454 * look so _wonderfully_ clever, that I was not able
5455 * to stand against the temptation 8) --ANK
5457 inet_csk_schedule_ack(sk);
5458 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5459 tcp_enter_quickack_mode(sk);
5460 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5461 TCP_DELACK_MAX, TCP_RTO_MAX);
5472 /* No ACK in the segment */
5476 * "If the RST bit is set
5478 * Otherwise (no ACK) drop the segment and return."
5481 goto discard_and_undo;
5485 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5486 tcp_paws_reject(&tp->rx_opt, 0))
5487 goto discard_and_undo;
5490 /* We see SYN without ACK. It is attempt of
5491 * simultaneous connect with crossed SYNs.
5492 * Particularly, it can be connect to self.
5494 tcp_set_state(sk, TCP_SYN_RECV);
5496 if (tp->rx_opt.saw_tstamp) {
5497 tp->rx_opt.tstamp_ok = 1;
5498 tcp_store_ts_recent(tp);
5499 tp->tcp_header_len =
5500 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5502 tp->tcp_header_len = sizeof(struct tcphdr);
5505 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5506 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5508 /* RFC1323: The window in SYN & SYN/ACK segments is
5511 tp->snd_wnd = ntohs(th->window);
5512 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5513 tp->max_window = tp->snd_wnd;
5515 tcp_ecn_rcv_syn(tp, th);
5518 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5519 tcp_initialize_rcv_mss(sk);
5521 tcp_send_synack(sk);
5523 /* Note, we could accept data and URG from this segment.
5524 * There are no obstacles to make this (except that we must
5525 * either change tcp_recvmsg() to prevent it from returning data
5526 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5528 * However, if we ignore data in ACKless segments sometimes,
5529 * we have no reasons to accept it sometimes.
5530 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5531 * is not flawless. So, discard packet for sanity.
5532 * Uncomment this return to process the data.
5539 /* "fifth, if neither of the SYN or RST bits is set then
5540 * drop the segment and return."
5544 tcp_clear_options(&tp->rx_opt);
5545 tp->rx_opt.mss_clamp = saved_clamp;
5549 tcp_clear_options(&tp->rx_opt);
5550 tp->rx_opt.mss_clamp = saved_clamp;
5555 * This function implements the receiving procedure of RFC 793 for
5556 * all states except ESTABLISHED and TIME_WAIT.
5557 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5558 * address independent.
5561 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5562 const struct tcphdr *th, unsigned int len)
5564 struct tcp_sock *tp = tcp_sk(sk);
5565 struct inet_connection_sock *icsk = inet_csk(sk);
5566 struct request_sock *req;
5571 tp->rx_opt.saw_tstamp = 0;
5573 switch (sk->sk_state) {
5587 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5590 /* Now we have several options: In theory there is
5591 * nothing else in the frame. KA9Q has an option to
5592 * send data with the syn, BSD accepts data with the
5593 * syn up to the [to be] advertised window and
5594 * Solaris 2.1 gives you a protocol error. For now
5595 * we just ignore it, that fits the spec precisely
5596 * and avoids incompatibilities. It would be nice in
5597 * future to drop through and process the data.
5599 * Now that TTCP is starting to be used we ought to
5601 * But, this leaves one open to an easy denial of
5602 * service attack, and SYN cookies can't defend
5603 * against this problem. So, we drop the data
5604 * in the interest of security over speed unless
5605 * it's still in use.
5613 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5617 /* Do step6 onward by hand. */
5618 tcp_urg(sk, skb, th);
5620 tcp_data_snd_check(sk);
5624 req = tp->fastopen_rsk;
5626 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
5627 sk->sk_state != TCP_FIN_WAIT1);
5629 if (tcp_check_req(sk, skb, req, NULL, true) == NULL)
5633 if (!th->ack && !th->rst)
5636 if (!tcp_validate_incoming(sk, skb, th, 0))
5639 /* step 5: check the ACK field */
5640 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
5641 FLAG_UPDATE_TS_RECENT) > 0;
5643 switch (sk->sk_state) {
5648 /* Once we leave TCP_SYN_RECV, we no longer need req
5652 synack_stamp = tcp_rsk(req)->snt_synack;
5653 tp->total_retrans = req->num_retrans;
5654 reqsk_fastopen_remove(sk, req, false);
5656 synack_stamp = tp->lsndtime;
5657 /* Make sure socket is routed, for correct metrics. */
5658 icsk->icsk_af_ops->rebuild_header(sk);
5659 tcp_init_congestion_control(sk);
5662 tp->copied_seq = tp->rcv_nxt;
5663 tcp_init_buffer_space(sk);
5666 tcp_set_state(sk, TCP_ESTABLISHED);
5667 sk->sk_state_change(sk);
5669 /* Note, that this wakeup is only for marginal crossed SYN case.
5670 * Passively open sockets are not waked up, because
5671 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5674 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5676 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5677 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
5678 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5679 tcp_synack_rtt_meas(sk, synack_stamp);
5681 if (tp->rx_opt.tstamp_ok)
5682 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5685 /* Re-arm the timer because data may have been sent out.
5686 * This is similar to the regular data transmission case
5687 * when new data has just been ack'ed.
5689 * (TFO) - we could try to be more aggressive and
5690 * retransmitting any data sooner based on when they
5695 tcp_init_metrics(sk);
5697 tcp_update_pacing_rate(sk);
5699 /* Prevent spurious tcp_cwnd_restart() on first data packet */
5700 tp->lsndtime = tcp_time_stamp;
5702 tcp_initialize_rcv_mss(sk);
5703 tcp_fast_path_on(tp);
5706 case TCP_FIN_WAIT1: {
5707 struct dst_entry *dst;
5710 /* If we enter the TCP_FIN_WAIT1 state and we are a
5711 * Fast Open socket and this is the first acceptable
5712 * ACK we have received, this would have acknowledged
5713 * our SYNACK so stop the SYNACK timer.
5716 /* Return RST if ack_seq is invalid.
5717 * Note that RFC793 only says to generate a
5718 * DUPACK for it but for TCP Fast Open it seems
5719 * better to treat this case like TCP_SYN_RECV
5724 /* We no longer need the request sock. */
5725 reqsk_fastopen_remove(sk, req, false);
5728 if (tp->snd_una != tp->write_seq)
5731 tcp_set_state(sk, TCP_FIN_WAIT2);
5732 sk->sk_shutdown |= SEND_SHUTDOWN;
5734 dst = __sk_dst_get(sk);
5738 if (!sock_flag(sk, SOCK_DEAD)) {
5739 /* Wake up lingering close() */
5740 sk->sk_state_change(sk);
5744 if (tp->linger2 < 0 ||
5745 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5746 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5748 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5752 tmo = tcp_fin_time(sk);
5753 if (tmo > TCP_TIMEWAIT_LEN) {
5754 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5755 } else if (th->fin || sock_owned_by_user(sk)) {
5756 /* Bad case. We could lose such FIN otherwise.
5757 * It is not a big problem, but it looks confusing
5758 * and not so rare event. We still can lose it now,
5759 * if it spins in bh_lock_sock(), but it is really
5762 inet_csk_reset_keepalive_timer(sk, tmo);
5764 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5771 if (tp->snd_una == tp->write_seq) {
5772 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5778 if (tp->snd_una == tp->write_seq) {
5779 tcp_update_metrics(sk);
5786 /* step 6: check the URG bit */
5787 tcp_urg(sk, skb, th);
5789 /* step 7: process the segment text */
5790 switch (sk->sk_state) {
5791 case TCP_CLOSE_WAIT:
5794 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5798 /* RFC 793 says to queue data in these states,
5799 * RFC 1122 says we MUST send a reset.
5800 * BSD 4.4 also does reset.
5802 if (sk->sk_shutdown & RCV_SHUTDOWN) {
5803 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5804 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5805 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5811 case TCP_ESTABLISHED:
5812 tcp_data_queue(sk, skb);
5817 /* tcp_data could move socket to TIME-WAIT */
5818 if (sk->sk_state != TCP_CLOSE) {
5819 tcp_data_snd_check(sk);
5820 tcp_ack_snd_check(sk);
5829 EXPORT_SYMBOL(tcp_rcv_state_process);
5831 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
5833 struct inet_request_sock *ireq = inet_rsk(req);
5835 if (family == AF_INET)
5836 LIMIT_NETDEBUG(KERN_DEBUG pr_fmt("drop open request from %pI4/%u\n"),
5837 &ireq->ir_rmt_addr, port);
5838 #if IS_ENABLED(CONFIG_IPV6)
5839 else if (family == AF_INET6)
5840 LIMIT_NETDEBUG(KERN_DEBUG pr_fmt("drop open request from %pI6/%u\n"),
5841 &ireq->ir_v6_rmt_addr, port);
5845 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
5847 * If we receive a SYN packet with these bits set, it means a
5848 * network is playing bad games with TOS bits. In order to
5849 * avoid possible false congestion notifications, we disable
5850 * TCP ECN negociation.
5852 * Exception: tcp_ca wants ECN. This is required for DCTCP
5853 * congestion control; it requires setting ECT on all packets,
5854 * including SYN. We inverse the test in this case: If our
5855 * local socket wants ECN, but peer only set ece/cwr (but not
5856 * ECT in IP header) its probably a non-DCTCP aware sender.
5858 static void tcp_ecn_create_request(struct request_sock *req,
5859 const struct sk_buff *skb,
5860 const struct sock *listen_sk)
5862 const struct tcphdr *th = tcp_hdr(skb);
5863 const struct net *net = sock_net(listen_sk);
5864 bool th_ecn = th->ece && th->cwr;
5870 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
5871 need_ecn = tcp_ca_needs_ecn(listen_sk);
5873 if (!ect && !need_ecn && net->ipv4.sysctl_tcp_ecn)
5874 inet_rsk(req)->ecn_ok = 1;
5875 else if (ect && need_ecn)
5876 inet_rsk(req)->ecn_ok = 1;
5879 int tcp_conn_request(struct request_sock_ops *rsk_ops,
5880 const struct tcp_request_sock_ops *af_ops,
5881 struct sock *sk, struct sk_buff *skb)
5883 struct tcp_options_received tmp_opt;
5884 struct request_sock *req;
5885 struct tcp_sock *tp = tcp_sk(sk);
5886 struct dst_entry *dst = NULL;
5887 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
5888 bool want_cookie = false, fastopen;
5890 struct tcp_fastopen_cookie foc = { .len = -1 };
5894 /* TW buckets are converted to open requests without
5895 * limitations, they conserve resources and peer is
5896 * evidently real one.
5898 if ((sysctl_tcp_syncookies == 2 ||
5899 inet_csk_reqsk_queue_is_full(sk)) && !isn) {
5900 want_cookie = tcp_syn_flood_action(sk, skb, rsk_ops->slab_name);
5906 /* Accept backlog is full. If we have already queued enough
5907 * of warm entries in syn queue, drop request. It is better than
5908 * clogging syn queue with openreqs with exponentially increasing
5911 if (sk_acceptq_is_full(sk) && inet_csk_reqsk_queue_young(sk) > 1) {
5912 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
5916 req = inet_reqsk_alloc(rsk_ops);
5920 tcp_rsk(req)->af_specific = af_ops;
5922 tcp_clear_options(&tmp_opt);
5923 tmp_opt.mss_clamp = af_ops->mss_clamp;
5924 tmp_opt.user_mss = tp->rx_opt.user_mss;
5925 tcp_parse_options(skb, &tmp_opt, 0, want_cookie ? NULL : &foc);
5927 if (want_cookie && !tmp_opt.saw_tstamp)
5928 tcp_clear_options(&tmp_opt);
5930 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
5931 tcp_openreq_init(req, &tmp_opt, skb, sk);
5933 af_ops->init_req(req, sk, skb);
5935 if (security_inet_conn_request(sk, skb, req))
5938 if (!want_cookie || tmp_opt.tstamp_ok)
5939 tcp_ecn_create_request(req, skb, sk);
5942 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
5943 req->cookie_ts = tmp_opt.tstamp_ok;
5945 /* VJ's idea. We save last timestamp seen
5946 * from the destination in peer table, when entering
5947 * state TIME-WAIT, and check against it before
5948 * accepting new connection request.
5950 * If "isn" is not zero, this request hit alive
5951 * timewait bucket, so that all the necessary checks
5952 * are made in the function processing timewait state.
5954 if (tcp_death_row.sysctl_tw_recycle) {
5957 dst = af_ops->route_req(sk, &fl, req, &strict);
5959 if (dst && strict &&
5960 !tcp_peer_is_proven(req, dst, true,
5961 tmp_opt.saw_tstamp)) {
5962 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSPASSIVEREJECTED);
5963 goto drop_and_release;
5966 /* Kill the following clause, if you dislike this way. */
5967 else if (!sysctl_tcp_syncookies &&
5968 (sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
5969 (sysctl_max_syn_backlog >> 2)) &&
5970 !tcp_peer_is_proven(req, dst, false,
5971 tmp_opt.saw_tstamp)) {
5972 /* Without syncookies last quarter of
5973 * backlog is filled with destinations,
5974 * proven to be alive.
5975 * It means that we continue to communicate
5976 * to destinations, already remembered
5977 * to the moment of synflood.
5979 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
5981 goto drop_and_release;
5984 isn = af_ops->init_seq(skb);
5987 dst = af_ops->route_req(sk, &fl, req, NULL);
5992 tcp_rsk(req)->snt_isn = isn;
5993 tcp_openreq_init_rwin(req, sk, dst);
5994 fastopen = !want_cookie &&
5995 tcp_try_fastopen(sk, skb, req, &foc, dst);
5996 err = af_ops->send_synack(sk, dst, &fl, req,
5997 skb_get_queue_mapping(skb), &foc);
5999 if (err || want_cookie)
6002 tcp_rsk(req)->listener = NULL;
6003 af_ops->queue_hash_add(sk, req, TCP_TIMEOUT_INIT);
6013 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_LISTENDROPS);
6016 EXPORT_SYMBOL(tcp_conn_request);