2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
23 * Pedro Roque : Fast Retransmit/Recovery.
25 * Retransmit queue handled by TCP.
26 * Better retransmit timer handling.
27 * New congestion avoidance.
31 * Eric : Fast Retransmit.
32 * Randy Scott : MSS option defines.
33 * Eric Schenk : Fixes to slow start algorithm.
34 * Eric Schenk : Yet another double ACK bug.
35 * Eric Schenk : Delayed ACK bug fixes.
36 * Eric Schenk : Floyd style fast retrans war avoidance.
37 * David S. Miller : Don't allow zero congestion window.
38 * Eric Schenk : Fix retransmitter so that it sends
39 * next packet on ack of previous packet.
40 * Andi Kleen : Moved open_request checking here
41 * and process RSTs for open_requests.
42 * Andi Kleen : Better prune_queue, and other fixes.
43 * Andrey Savochkin: Fix RTT measurements in the presence of
45 * Andrey Savochkin: Check sequence numbers correctly when
46 * removing SACKs due to in sequence incoming
48 * Andi Kleen: Make sure we never ack data there is not
49 * enough room for. Also make this condition
50 * a fatal error if it might still happen.
51 * Andi Kleen: Add tcp_measure_rcv_mss to make
52 * connections with MSS<min(MTU,ann. MSS)
53 * work without delayed acks.
54 * Andi Kleen: Process packets with PSH set in the
56 * J Hadi Salim: ECN support
59 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
60 * engine. Lots of bugs are found.
61 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
64 #define pr_fmt(fmt) "TCP: " fmt
67 #include <linux/slab.h>
68 #include <linux/module.h>
69 #include <linux/sysctl.h>
70 #include <linux/kernel.h>
71 #include <linux/prefetch.h>
74 #include <net/inet_common.h>
75 #include <linux/ipsec.h>
76 #include <asm/unaligned.h>
77 #include <linux/errqueue.h>
79 int sysctl_tcp_timestamps __read_mostly = 1;
80 int sysctl_tcp_window_scaling __read_mostly = 1;
81 int sysctl_tcp_sack __read_mostly = 1;
82 int sysctl_tcp_fack __read_mostly = 1;
83 int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
84 int sysctl_tcp_max_reordering __read_mostly = 300;
85 EXPORT_SYMBOL(sysctl_tcp_reordering);
86 int sysctl_tcp_dsack __read_mostly = 1;
87 int sysctl_tcp_app_win __read_mostly = 31;
88 int sysctl_tcp_adv_win_scale __read_mostly = 1;
89 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
91 /* rfc5961 challenge ack rate limiting */
92 int sysctl_tcp_challenge_ack_limit = 100;
94 int sysctl_tcp_stdurg __read_mostly;
95 int sysctl_tcp_rfc1337 __read_mostly;
96 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
97 int sysctl_tcp_frto __read_mostly = 2;
99 int sysctl_tcp_thin_dupack __read_mostly;
101 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
102 int sysctl_tcp_early_retrans __read_mostly = 3;
103 int sysctl_tcp_invalid_ratelimit __read_mostly = HZ/2;
105 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
106 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
107 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
108 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
109 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
110 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
111 #define FLAG_ECE 0x40 /* ECE in this ACK */
112 #define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */
113 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
114 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
115 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
116 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
117 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
118 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
120 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
121 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
122 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
123 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
125 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
126 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
128 /* Adapt the MSS value used to make delayed ack decision to the
131 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
133 struct inet_connection_sock *icsk = inet_csk(sk);
134 const unsigned int lss = icsk->icsk_ack.last_seg_size;
137 icsk->icsk_ack.last_seg_size = 0;
139 /* skb->len may jitter because of SACKs, even if peer
140 * sends good full-sized frames.
142 len = skb_shinfo(skb)->gso_size ? : skb->len;
143 if (len >= icsk->icsk_ack.rcv_mss) {
144 icsk->icsk_ack.rcv_mss = len;
146 /* Otherwise, we make more careful check taking into account,
147 * that SACKs block is variable.
149 * "len" is invariant segment length, including TCP header.
151 len += skb->data - skb_transport_header(skb);
152 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
153 /* If PSH is not set, packet should be
154 * full sized, provided peer TCP is not badly broken.
155 * This observation (if it is correct 8)) allows
156 * to handle super-low mtu links fairly.
158 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
159 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
160 /* Subtract also invariant (if peer is RFC compliant),
161 * tcp header plus fixed timestamp option length.
162 * Resulting "len" is MSS free of SACK jitter.
164 len -= tcp_sk(sk)->tcp_header_len;
165 icsk->icsk_ack.last_seg_size = len;
167 icsk->icsk_ack.rcv_mss = len;
171 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
172 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
173 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
177 static void tcp_incr_quickack(struct sock *sk)
179 struct inet_connection_sock *icsk = inet_csk(sk);
180 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
184 if (quickacks > icsk->icsk_ack.quick)
185 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
188 static void tcp_enter_quickack_mode(struct sock *sk)
190 struct inet_connection_sock *icsk = inet_csk(sk);
191 tcp_incr_quickack(sk);
192 icsk->icsk_ack.pingpong = 0;
193 icsk->icsk_ack.ato = TCP_ATO_MIN;
196 /* Send ACKs quickly, if "quick" count is not exhausted
197 * and the session is not interactive.
200 static bool tcp_in_quickack_mode(struct sock *sk)
202 const struct inet_connection_sock *icsk = inet_csk(sk);
203 const struct dst_entry *dst = __sk_dst_get(sk);
205 return (dst && dst_metric(dst, RTAX_QUICKACK)) ||
206 (icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong);
209 static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
211 if (tp->ecn_flags & TCP_ECN_OK)
212 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
215 static void tcp_ecn_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
217 if (tcp_hdr(skb)->cwr)
218 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
221 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
223 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
226 static void __tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
228 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
229 case INET_ECN_NOT_ECT:
230 /* Funny extension: if ECT is not set on a segment,
231 * and we already seen ECT on a previous segment,
232 * it is probably a retransmit.
234 if (tp->ecn_flags & TCP_ECN_SEEN)
235 tcp_enter_quickack_mode((struct sock *)tp);
238 if (tcp_ca_needs_ecn((struct sock *)tp))
239 tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_IS_CE);
241 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
242 /* Better not delay acks, sender can have a very low cwnd */
243 tcp_enter_quickack_mode((struct sock *)tp);
244 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
246 tp->ecn_flags |= TCP_ECN_SEEN;
249 if (tcp_ca_needs_ecn((struct sock *)tp))
250 tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_NO_CE);
251 tp->ecn_flags |= TCP_ECN_SEEN;
256 static void tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
258 if (tp->ecn_flags & TCP_ECN_OK)
259 __tcp_ecn_check_ce(tp, skb);
262 static void tcp_ecn_rcv_synack(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 void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
270 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
271 tp->ecn_flags &= ~TCP_ECN_OK;
274 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
276 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
281 /* Buffer size and advertised window tuning.
283 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
286 static void tcp_sndbuf_expand(struct sock *sk)
288 const struct tcp_sock *tp = tcp_sk(sk);
292 /* Worst case is non GSO/TSO : each frame consumes one skb
293 * and skb->head is kmalloced using power of two area of memory
295 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
297 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
299 per_mss = roundup_pow_of_two(per_mss) +
300 SKB_DATA_ALIGN(sizeof(struct sk_buff));
302 nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd);
303 nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
305 /* Fast Recovery (RFC 5681 3.2) :
306 * Cubic needs 1.7 factor, rounded to 2 to include
307 * extra cushion (application might react slowly to POLLOUT)
309 sndmem = 2 * nr_segs * per_mss;
311 if (sk->sk_sndbuf < sndmem)
312 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
315 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
317 * All tcp_full_space() is split to two parts: "network" buffer, allocated
318 * forward and advertised in receiver window (tp->rcv_wnd) and
319 * "application buffer", required to isolate scheduling/application
320 * latencies from network.
321 * window_clamp is maximal advertised window. It can be less than
322 * tcp_full_space(), in this case tcp_full_space() - window_clamp
323 * is reserved for "application" buffer. The less window_clamp is
324 * the smoother our behaviour from viewpoint of network, but the lower
325 * throughput and the higher sensitivity of the connection to losses. 8)
327 * rcv_ssthresh is more strict window_clamp used at "slow start"
328 * phase to predict further behaviour of this connection.
329 * It is used for two goals:
330 * - to enforce header prediction at sender, even when application
331 * requires some significant "application buffer". It is check #1.
332 * - to prevent pruning of receive queue because of misprediction
333 * of receiver window. Check #2.
335 * The scheme does not work when sender sends good segments opening
336 * window and then starts to feed us spaghetti. But it should work
337 * in common situations. Otherwise, we have to rely on queue collapsing.
340 /* Slow part of check#2. */
341 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
343 struct tcp_sock *tp = tcp_sk(sk);
345 int truesize = tcp_win_from_space(skb->truesize) >> 1;
346 int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
348 while (tp->rcv_ssthresh <= window) {
349 if (truesize <= skb->len)
350 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
358 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
360 struct tcp_sock *tp = tcp_sk(sk);
363 if (tp->rcv_ssthresh < tp->window_clamp &&
364 (int)tp->rcv_ssthresh < tcp_space(sk) &&
365 !tcp_under_memory_pressure(sk)) {
368 /* Check #2. Increase window, if skb with such overhead
369 * will fit to rcvbuf in future.
371 if (tcp_win_from_space(skb->truesize) <= skb->len)
372 incr = 2 * tp->advmss;
374 incr = __tcp_grow_window(sk, skb);
377 incr = max_t(int, incr, 2 * skb->len);
378 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
380 inet_csk(sk)->icsk_ack.quick |= 1;
385 /* 3. Tuning rcvbuf, when connection enters established state. */
386 static void tcp_fixup_rcvbuf(struct sock *sk)
388 u32 mss = tcp_sk(sk)->advmss;
391 rcvmem = 2 * SKB_TRUESIZE(mss + MAX_TCP_HEADER) *
392 tcp_default_init_rwnd(mss);
394 /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
395 * Allow enough cushion so that sender is not limited by our window
397 if (sysctl_tcp_moderate_rcvbuf)
400 if (sk->sk_rcvbuf < rcvmem)
401 sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]);
404 /* 4. Try to fixup all. It is made immediately after connection enters
407 void tcp_init_buffer_space(struct sock *sk)
409 struct tcp_sock *tp = tcp_sk(sk);
412 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
413 tcp_fixup_rcvbuf(sk);
414 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
415 tcp_sndbuf_expand(sk);
417 tp->rcvq_space.space = tp->rcv_wnd;
418 tp->rcvq_space.time = tcp_time_stamp;
419 tp->rcvq_space.seq = tp->copied_seq;
421 maxwin = tcp_full_space(sk);
423 if (tp->window_clamp >= maxwin) {
424 tp->window_clamp = maxwin;
426 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
427 tp->window_clamp = max(maxwin -
428 (maxwin >> sysctl_tcp_app_win),
432 /* Force reservation of one segment. */
433 if (sysctl_tcp_app_win &&
434 tp->window_clamp > 2 * tp->advmss &&
435 tp->window_clamp + tp->advmss > maxwin)
436 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
438 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
439 tp->snd_cwnd_stamp = tcp_time_stamp;
442 /* 5. Recalculate window clamp after socket hit its memory bounds. */
443 static void tcp_clamp_window(struct sock *sk)
445 struct tcp_sock *tp = tcp_sk(sk);
446 struct inet_connection_sock *icsk = inet_csk(sk);
448 icsk->icsk_ack.quick = 0;
450 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
451 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
452 !tcp_under_memory_pressure(sk) &&
453 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
454 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
457 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
458 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
461 /* Initialize RCV_MSS value.
462 * RCV_MSS is an our guess about MSS used by the peer.
463 * We haven't any direct information about the MSS.
464 * It's better to underestimate the RCV_MSS rather than overestimate.
465 * Overestimations make us ACKing less frequently than needed.
466 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
468 void tcp_initialize_rcv_mss(struct sock *sk)
470 const struct tcp_sock *tp = tcp_sk(sk);
471 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
473 hint = min(hint, tp->rcv_wnd / 2);
474 hint = min(hint, TCP_MSS_DEFAULT);
475 hint = max(hint, TCP_MIN_MSS);
477 inet_csk(sk)->icsk_ack.rcv_mss = hint;
479 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
481 /* Receiver "autotuning" code.
483 * The algorithm for RTT estimation w/o timestamps is based on
484 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
485 * <http://public.lanl.gov/radiant/pubs.html#DRS>
487 * More detail on this code can be found at
488 * <http://staff.psc.edu/jheffner/>,
489 * though this reference is out of date. A new paper
492 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
494 u32 new_sample = tp->rcv_rtt_est.rtt;
500 if (new_sample != 0) {
501 /* If we sample in larger samples in the non-timestamp
502 * case, we could grossly overestimate the RTT especially
503 * with chatty applications or bulk transfer apps which
504 * are stalled on filesystem I/O.
506 * Also, since we are only going for a minimum in the
507 * non-timestamp case, we do not smooth things out
508 * else with timestamps disabled convergence takes too
512 m -= (new_sample >> 3);
520 /* No previous measure. */
524 if (tp->rcv_rtt_est.rtt != new_sample)
525 tp->rcv_rtt_est.rtt = new_sample;
528 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
530 if (tp->rcv_rtt_est.time == 0)
532 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
534 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rcv_rtt_est.time, 1);
537 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
538 tp->rcv_rtt_est.time = tcp_time_stamp;
541 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
542 const struct sk_buff *skb)
544 struct tcp_sock *tp = tcp_sk(sk);
545 if (tp->rx_opt.rcv_tsecr &&
546 (TCP_SKB_CB(skb)->end_seq -
547 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
548 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
552 * This function should be called every time data is copied to user space.
553 * It calculates the appropriate TCP receive buffer space.
555 void tcp_rcv_space_adjust(struct sock *sk)
557 struct tcp_sock *tp = tcp_sk(sk);
561 time = tcp_time_stamp - tp->rcvq_space.time;
562 if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
565 /* Number of bytes copied to user in last RTT */
566 copied = tp->copied_seq - tp->rcvq_space.seq;
567 if (copied <= tp->rcvq_space.space)
571 * copied = bytes received in previous RTT, our base window
572 * To cope with packet losses, we need a 2x factor
573 * To cope with slow start, and sender growing its cwin by 100 %
574 * every RTT, we need a 4x factor, because the ACK we are sending
575 * now is for the next RTT, not the current one :
576 * <prev RTT . ><current RTT .. ><next RTT .... >
579 if (sysctl_tcp_moderate_rcvbuf &&
580 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
581 int rcvwin, rcvmem, rcvbuf;
583 /* minimal window to cope with packet losses, assuming
584 * steady state. Add some cushion because of small variations.
586 rcvwin = (copied << 1) + 16 * tp->advmss;
588 /* If rate increased by 25%,
589 * assume slow start, rcvwin = 3 * copied
590 * If rate increased by 50%,
591 * assume sender can use 2x growth, rcvwin = 4 * copied
594 tp->rcvq_space.space + (tp->rcvq_space.space >> 2)) {
596 tp->rcvq_space.space + (tp->rcvq_space.space >> 1))
599 rcvwin += (rcvwin >> 1);
602 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
603 while (tcp_win_from_space(rcvmem) < tp->advmss)
606 rcvbuf = min(rcvwin / tp->advmss * rcvmem, sysctl_tcp_rmem[2]);
607 if (rcvbuf > sk->sk_rcvbuf) {
608 sk->sk_rcvbuf = rcvbuf;
610 /* Make the window clamp follow along. */
611 tp->window_clamp = rcvwin;
614 tp->rcvq_space.space = copied;
617 tp->rcvq_space.seq = tp->copied_seq;
618 tp->rcvq_space.time = tcp_time_stamp;
621 /* There is something which you must keep in mind when you analyze the
622 * behavior of the tp->ato delayed ack timeout interval. When a
623 * connection starts up, we want to ack as quickly as possible. The
624 * problem is that "good" TCP's do slow start at the beginning of data
625 * transmission. The means that until we send the first few ACK's the
626 * sender will sit on his end and only queue most of his data, because
627 * he can only send snd_cwnd unacked packets at any given time. For
628 * each ACK we send, he increments snd_cwnd and transmits more of his
631 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
633 struct tcp_sock *tp = tcp_sk(sk);
634 struct inet_connection_sock *icsk = inet_csk(sk);
637 inet_csk_schedule_ack(sk);
639 tcp_measure_rcv_mss(sk, skb);
641 tcp_rcv_rtt_measure(tp);
643 now = tcp_time_stamp;
645 if (!icsk->icsk_ack.ato) {
646 /* The _first_ data packet received, initialize
647 * delayed ACK engine.
649 tcp_incr_quickack(sk);
650 icsk->icsk_ack.ato = TCP_ATO_MIN;
652 int m = now - icsk->icsk_ack.lrcvtime;
654 if (m <= TCP_ATO_MIN / 2) {
655 /* The fastest case is the first. */
656 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
657 } else if (m < icsk->icsk_ack.ato) {
658 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
659 if (icsk->icsk_ack.ato > icsk->icsk_rto)
660 icsk->icsk_ack.ato = icsk->icsk_rto;
661 } else if (m > icsk->icsk_rto) {
662 /* Too long gap. Apparently sender failed to
663 * restart window, so that we send ACKs quickly.
665 tcp_incr_quickack(sk);
669 icsk->icsk_ack.lrcvtime = now;
671 tcp_ecn_check_ce(tp, skb);
674 tcp_grow_window(sk, skb);
677 /* Called to compute a smoothed rtt estimate. The data fed to this
678 * routine either comes from timestamps, or from segments that were
679 * known _not_ to have been retransmitted [see Karn/Partridge
680 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
681 * piece by Van Jacobson.
682 * NOTE: the next three routines used to be one big routine.
683 * To save cycles in the RFC 1323 implementation it was better to break
684 * it up into three procedures. -- erics
686 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
688 struct tcp_sock *tp = tcp_sk(sk);
689 long m = mrtt_us; /* RTT */
690 u32 srtt = tp->srtt_us;
692 /* The following amusing code comes from Jacobson's
693 * article in SIGCOMM '88. Note that rtt and mdev
694 * are scaled versions of rtt and mean deviation.
695 * This is designed to be as fast as possible
696 * m stands for "measurement".
698 * On a 1990 paper the rto value is changed to:
699 * RTO = rtt + 4 * mdev
701 * Funny. This algorithm seems to be very broken.
702 * These formulae increase RTO, when it should be decreased, increase
703 * too slowly, when it should be increased quickly, decrease too quickly
704 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
705 * does not matter how to _calculate_ it. Seems, it was trap
706 * that VJ failed to avoid. 8)
709 m -= (srtt >> 3); /* m is now error in rtt est */
710 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
712 m = -m; /* m is now abs(error) */
713 m -= (tp->mdev_us >> 2); /* similar update on mdev */
714 /* This is similar to one of Eifel findings.
715 * Eifel blocks mdev updates when rtt decreases.
716 * This solution is a bit different: we use finer gain
717 * for mdev in this case (alpha*beta).
718 * Like Eifel it also prevents growth of rto,
719 * but also it limits too fast rto decreases,
720 * happening in pure Eifel.
725 m -= (tp->mdev_us >> 2); /* similar update on mdev */
727 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
728 if (tp->mdev_us > tp->mdev_max_us) {
729 tp->mdev_max_us = tp->mdev_us;
730 if (tp->mdev_max_us > tp->rttvar_us)
731 tp->rttvar_us = tp->mdev_max_us;
733 if (after(tp->snd_una, tp->rtt_seq)) {
734 if (tp->mdev_max_us < tp->rttvar_us)
735 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
736 tp->rtt_seq = tp->snd_nxt;
737 tp->mdev_max_us = tcp_rto_min_us(sk);
740 /* no previous measure. */
741 srtt = m << 3; /* take the measured time to be rtt */
742 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
743 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
744 tp->mdev_max_us = tp->rttvar_us;
745 tp->rtt_seq = tp->snd_nxt;
747 tp->srtt_us = max(1U, srtt);
750 /* Set the sk_pacing_rate to allow proper sizing of TSO packets.
751 * Note: TCP stack does not yet implement pacing.
752 * FQ packet scheduler can be used to implement cheap but effective
753 * TCP pacing, to smooth the burst on large writes when packets
754 * in flight is significantly lower than cwnd (or rwin)
756 int sysctl_tcp_pacing_ss_ratio __read_mostly = 200;
757 int sysctl_tcp_pacing_ca_ratio __read_mostly = 120;
759 static void tcp_update_pacing_rate(struct sock *sk)
761 const struct tcp_sock *tp = tcp_sk(sk);
764 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
765 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
767 /* current rate is (cwnd * mss) / srtt
768 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
769 * In Congestion Avoidance phase, set it to 120 % the current rate.
771 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
772 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
773 * end of slow start and should slow down.
775 if (tp->snd_cwnd < tp->snd_ssthresh / 2)
776 rate *= sysctl_tcp_pacing_ss_ratio;
778 rate *= sysctl_tcp_pacing_ca_ratio;
780 rate *= max(tp->snd_cwnd, tp->packets_out);
782 if (likely(tp->srtt_us))
783 do_div(rate, tp->srtt_us);
785 /* ACCESS_ONCE() is needed because sch_fq fetches sk_pacing_rate
786 * without any lock. We want to make sure compiler wont store
787 * intermediate values in this location.
789 ACCESS_ONCE(sk->sk_pacing_rate) = min_t(u64, rate,
790 sk->sk_max_pacing_rate);
793 /* Calculate rto without backoff. This is the second half of Van Jacobson's
794 * routine referred to above.
796 static void tcp_set_rto(struct sock *sk)
798 const struct tcp_sock *tp = tcp_sk(sk);
799 /* Old crap is replaced with new one. 8)
802 * 1. If rtt variance happened to be less 50msec, it is hallucination.
803 * It cannot be less due to utterly erratic ACK generation made
804 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
805 * to do with delayed acks, because at cwnd>2 true delack timeout
806 * is invisible. Actually, Linux-2.4 also generates erratic
807 * ACKs in some circumstances.
809 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
811 /* 2. Fixups made earlier cannot be right.
812 * If we do not estimate RTO correctly without them,
813 * all the algo is pure shit and should be replaced
814 * with correct one. It is exactly, which we pretend to do.
817 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
818 * guarantees that rto is higher.
823 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
825 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
828 cwnd = TCP_INIT_CWND;
829 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
833 * Packet counting of FACK is based on in-order assumptions, therefore TCP
834 * disables it when reordering is detected
836 void tcp_disable_fack(struct tcp_sock *tp)
838 /* RFC3517 uses different metric in lost marker => reset on change */
840 tp->lost_skb_hint = NULL;
841 tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED;
844 /* Take a notice that peer is sending D-SACKs */
845 static void tcp_dsack_seen(struct tcp_sock *tp)
847 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
850 static void tcp_update_reordering(struct sock *sk, const int metric,
853 struct tcp_sock *tp = tcp_sk(sk);
854 if (metric > tp->reordering) {
857 tp->reordering = min(sysctl_tcp_max_reordering, metric);
859 /* This exciting event is worth to be remembered. 8) */
861 mib_idx = LINUX_MIB_TCPTSREORDER;
862 else if (tcp_is_reno(tp))
863 mib_idx = LINUX_MIB_TCPRENOREORDER;
864 else if (tcp_is_fack(tp))
865 mib_idx = LINUX_MIB_TCPFACKREORDER;
867 mib_idx = LINUX_MIB_TCPSACKREORDER;
869 NET_INC_STATS_BH(sock_net(sk), mib_idx);
870 #if FASTRETRANS_DEBUG > 1
871 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
872 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
876 tp->undo_marker ? tp->undo_retrans : 0);
878 tcp_disable_fack(tp);
882 tcp_disable_early_retrans(tp);
885 /* This must be called before lost_out is incremented */
886 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
888 if (!tp->retransmit_skb_hint ||
889 before(TCP_SKB_CB(skb)->seq,
890 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
891 tp->retransmit_skb_hint = skb;
894 after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
895 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
898 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
900 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
901 tcp_verify_retransmit_hint(tp, skb);
903 tp->lost_out += tcp_skb_pcount(skb);
904 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
908 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp,
911 tcp_verify_retransmit_hint(tp, skb);
913 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
914 tp->lost_out += tcp_skb_pcount(skb);
915 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
919 /* This procedure tags the retransmission queue when SACKs arrive.
921 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
922 * Packets in queue with these bits set are counted in variables
923 * sacked_out, retrans_out and lost_out, correspondingly.
925 * Valid combinations are:
926 * Tag InFlight Description
927 * 0 1 - orig segment is in flight.
928 * S 0 - nothing flies, orig reached receiver.
929 * L 0 - nothing flies, orig lost by net.
930 * R 2 - both orig and retransmit are in flight.
931 * L|R 1 - orig is lost, retransmit is in flight.
932 * S|R 1 - orig reached receiver, retrans is still in flight.
933 * (L|S|R is logically valid, it could occur when L|R is sacked,
934 * but it is equivalent to plain S and code short-curcuits it to S.
935 * L|S is logically invalid, it would mean -1 packet in flight 8))
937 * These 6 states form finite state machine, controlled by the following events:
938 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
939 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
940 * 3. Loss detection event of two flavors:
941 * A. Scoreboard estimator decided the packet is lost.
942 * A'. Reno "three dupacks" marks head of queue lost.
943 * A''. Its FACK modification, head until snd.fack is lost.
944 * B. SACK arrives sacking SND.NXT at the moment, when the
945 * segment was retransmitted.
946 * 4. D-SACK added new rule: D-SACK changes any tag to S.
948 * It is pleasant to note, that state diagram turns out to be commutative,
949 * so that we are allowed not to be bothered by order of our actions,
950 * when multiple events arrive simultaneously. (see the function below).
952 * Reordering detection.
953 * --------------------
954 * Reordering metric is maximal distance, which a packet can be displaced
955 * in packet stream. With SACKs we can estimate it:
957 * 1. SACK fills old hole and the corresponding segment was not
958 * ever retransmitted -> reordering. Alas, we cannot use it
959 * when segment was retransmitted.
960 * 2. The last flaw is solved with D-SACK. D-SACK arrives
961 * for retransmitted and already SACKed segment -> reordering..
962 * Both of these heuristics are not used in Loss state, when we cannot
963 * account for retransmits accurately.
965 * SACK block validation.
966 * ----------------------
968 * SACK block range validation checks that the received SACK block fits to
969 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
970 * Note that SND.UNA is not included to the range though being valid because
971 * it means that the receiver is rather inconsistent with itself reporting
972 * SACK reneging when it should advance SND.UNA. Such SACK block this is
973 * perfectly valid, however, in light of RFC2018 which explicitly states
974 * that "SACK block MUST reflect the newest segment. Even if the newest
975 * segment is going to be discarded ...", not that it looks very clever
976 * in case of head skb. Due to potentional receiver driven attacks, we
977 * choose to avoid immediate execution of a walk in write queue due to
978 * reneging and defer head skb's loss recovery to standard loss recovery
979 * procedure that will eventually trigger (nothing forbids us doing this).
981 * Implements also blockage to start_seq wrap-around. Problem lies in the
982 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
983 * there's no guarantee that it will be before snd_nxt (n). The problem
984 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
987 * <- outs wnd -> <- wrapzone ->
988 * u e n u_w e_w s n_w
990 * |<------------+------+----- TCP seqno space --------------+---------->|
991 * ...-- <2^31 ->| |<--------...
992 * ...---- >2^31 ------>| |<--------...
994 * Current code wouldn't be vulnerable but it's better still to discard such
995 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
996 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
997 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
998 * equal to the ideal case (infinite seqno space without wrap caused issues).
1000 * With D-SACK the lower bound is extended to cover sequence space below
1001 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1002 * again, D-SACK block must not to go across snd_una (for the same reason as
1003 * for the normal SACK blocks, explained above). But there all simplicity
1004 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1005 * fully below undo_marker they do not affect behavior in anyway and can
1006 * therefore be safely ignored. In rare cases (which are more or less
1007 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1008 * fragmentation and packet reordering past skb's retransmission. To consider
1009 * them correctly, the acceptable range must be extended even more though
1010 * the exact amount is rather hard to quantify. However, tp->max_window can
1011 * be used as an exaggerated estimate.
1013 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1014 u32 start_seq, u32 end_seq)
1016 /* Too far in future, or reversed (interpretation is ambiguous) */
1017 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1020 /* Nasty start_seq wrap-around check (see comments above) */
1021 if (!before(start_seq, tp->snd_nxt))
1024 /* In outstanding window? ...This is valid exit for D-SACKs too.
1025 * start_seq == snd_una is non-sensical (see comments above)
1027 if (after(start_seq, tp->snd_una))
1030 if (!is_dsack || !tp->undo_marker)
1033 /* ...Then it's D-SACK, and must reside below snd_una completely */
1034 if (after(end_seq, tp->snd_una))
1037 if (!before(start_seq, tp->undo_marker))
1041 if (!after(end_seq, tp->undo_marker))
1044 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1045 * start_seq < undo_marker and end_seq >= undo_marker.
1047 return !before(start_seq, end_seq - tp->max_window);
1050 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1051 * Event "B". Later note: FACK people cheated me again 8), we have to account
1052 * for reordering! Ugly, but should help.
1054 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1055 * less than what is now known to be received by the other end (derived from
1056 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1057 * retransmitted skbs to avoid some costly processing per ACKs.
1059 static void tcp_mark_lost_retrans(struct sock *sk, int *flag)
1061 const struct inet_connection_sock *icsk = inet_csk(sk);
1062 struct tcp_sock *tp = tcp_sk(sk);
1063 struct sk_buff *skb;
1065 u32 new_low_seq = tp->snd_nxt;
1066 u32 received_upto = tcp_highest_sack_seq(tp);
1068 if (!tcp_is_fack(tp) || !tp->retrans_out ||
1069 !after(received_upto, tp->lost_retrans_low) ||
1070 icsk->icsk_ca_state != TCP_CA_Recovery)
1073 tcp_for_write_queue(skb, sk) {
1074 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1076 if (skb == tcp_send_head(sk))
1078 if (cnt == tp->retrans_out)
1080 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1083 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1086 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1087 * constraint here (see above) but figuring out that at
1088 * least tp->reordering SACK blocks reside between ack_seq
1089 * and received_upto is not easy task to do cheaply with
1090 * the available datastructures.
1092 * Whether FACK should check here for tp->reordering segs
1093 * in-between one could argue for either way (it would be
1094 * rather simple to implement as we could count fack_count
1095 * during the walk and do tp->fackets_out - fack_count).
1097 if (after(received_upto, ack_seq)) {
1098 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1099 tp->retrans_out -= tcp_skb_pcount(skb);
1100 *flag |= FLAG_LOST_RETRANS;
1101 tcp_skb_mark_lost_uncond_verify(tp, skb);
1102 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
1104 if (before(ack_seq, new_low_seq))
1105 new_low_seq = ack_seq;
1106 cnt += tcp_skb_pcount(skb);
1110 if (tp->retrans_out)
1111 tp->lost_retrans_low = new_low_seq;
1114 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1115 struct tcp_sack_block_wire *sp, int num_sacks,
1118 struct tcp_sock *tp = tcp_sk(sk);
1119 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1120 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1121 bool dup_sack = false;
1123 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1126 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1127 } else if (num_sacks > 1) {
1128 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1129 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1131 if (!after(end_seq_0, end_seq_1) &&
1132 !before(start_seq_0, start_seq_1)) {
1135 NET_INC_STATS_BH(sock_net(sk),
1136 LINUX_MIB_TCPDSACKOFORECV);
1140 /* D-SACK for already forgotten data... Do dumb counting. */
1141 if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 &&
1142 !after(end_seq_0, prior_snd_una) &&
1143 after(end_seq_0, tp->undo_marker))
1149 struct tcp_sacktag_state {
1152 /* Timestamps for earliest and latest never-retransmitted segment
1153 * that was SACKed. RTO needs the earliest RTT to stay conservative,
1154 * but congestion control should still get an accurate delay signal.
1156 struct skb_mstamp first_sackt;
1157 struct skb_mstamp last_sackt;
1161 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1162 * the incoming SACK may not exactly match but we can find smaller MSS
1163 * aligned portion of it that matches. Therefore we might need to fragment
1164 * which may fail and creates some hassle (caller must handle error case
1167 * FIXME: this could be merged to shift decision code
1169 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1170 u32 start_seq, u32 end_seq)
1174 unsigned int pkt_len;
1177 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1178 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1180 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1181 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1182 mss = tcp_skb_mss(skb);
1183 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1186 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1190 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1195 /* Round if necessary so that SACKs cover only full MSSes
1196 * and/or the remaining small portion (if present)
1198 if (pkt_len > mss) {
1199 unsigned int new_len = (pkt_len / mss) * mss;
1200 if (!in_sack && new_len < pkt_len) {
1202 if (new_len >= skb->len)
1207 err = tcp_fragment(sk, skb, pkt_len, mss, GFP_ATOMIC);
1215 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1216 static u8 tcp_sacktag_one(struct sock *sk,
1217 struct tcp_sacktag_state *state, u8 sacked,
1218 u32 start_seq, u32 end_seq,
1219 int dup_sack, int pcount,
1220 const struct skb_mstamp *xmit_time)
1222 struct tcp_sock *tp = tcp_sk(sk);
1223 int fack_count = state->fack_count;
1225 /* Account D-SACK for retransmitted packet. */
1226 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1227 if (tp->undo_marker && tp->undo_retrans > 0 &&
1228 after(end_seq, tp->undo_marker))
1230 if (sacked & TCPCB_SACKED_ACKED)
1231 state->reord = min(fack_count, state->reord);
1234 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1235 if (!after(end_seq, tp->snd_una))
1238 if (!(sacked & TCPCB_SACKED_ACKED)) {
1239 if (sacked & TCPCB_SACKED_RETRANS) {
1240 /* If the segment is not tagged as lost,
1241 * we do not clear RETRANS, believing
1242 * that retransmission is still in flight.
1244 if (sacked & TCPCB_LOST) {
1245 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1246 tp->lost_out -= pcount;
1247 tp->retrans_out -= pcount;
1250 if (!(sacked & TCPCB_RETRANS)) {
1251 /* New sack for not retransmitted frame,
1252 * which was in hole. It is reordering.
1254 if (before(start_seq,
1255 tcp_highest_sack_seq(tp)))
1256 state->reord = min(fack_count,
1258 if (!after(end_seq, tp->high_seq))
1259 state->flag |= FLAG_ORIG_SACK_ACKED;
1260 if (state->first_sackt.v64 == 0)
1261 state->first_sackt = *xmit_time;
1262 state->last_sackt = *xmit_time;
1265 if (sacked & TCPCB_LOST) {
1266 sacked &= ~TCPCB_LOST;
1267 tp->lost_out -= pcount;
1271 sacked |= TCPCB_SACKED_ACKED;
1272 state->flag |= FLAG_DATA_SACKED;
1273 tp->sacked_out += pcount;
1275 fack_count += pcount;
1277 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1278 if (!tcp_is_fack(tp) && tp->lost_skb_hint &&
1279 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1280 tp->lost_cnt_hint += pcount;
1282 if (fack_count > tp->fackets_out)
1283 tp->fackets_out = fack_count;
1286 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1287 * frames and clear it. undo_retrans is decreased above, L|R frames
1288 * are accounted above as well.
1290 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1291 sacked &= ~TCPCB_SACKED_RETRANS;
1292 tp->retrans_out -= pcount;
1298 /* Shift newly-SACKed bytes from this skb to the immediately previous
1299 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1301 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1302 struct tcp_sacktag_state *state,
1303 unsigned int pcount, int shifted, int mss,
1306 struct tcp_sock *tp = tcp_sk(sk);
1307 struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1308 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1309 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1313 /* Adjust counters and hints for the newly sacked sequence
1314 * range but discard the return value since prev is already
1315 * marked. We must tag the range first because the seq
1316 * advancement below implicitly advances
1317 * tcp_highest_sack_seq() when skb is highest_sack.
1319 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1320 start_seq, end_seq, dup_sack, pcount,
1323 if (skb == tp->lost_skb_hint)
1324 tp->lost_cnt_hint += pcount;
1326 TCP_SKB_CB(prev)->end_seq += shifted;
1327 TCP_SKB_CB(skb)->seq += shifted;
1329 tcp_skb_pcount_add(prev, pcount);
1330 BUG_ON(tcp_skb_pcount(skb) < pcount);
1331 tcp_skb_pcount_add(skb, -pcount);
1333 /* When we're adding to gso_segs == 1, gso_size will be zero,
1334 * in theory this shouldn't be necessary but as long as DSACK
1335 * code can come after this skb later on it's better to keep
1336 * setting gso_size to something.
1338 if (!TCP_SKB_CB(prev)->tcp_gso_size)
1339 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1341 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1342 if (tcp_skb_pcount(skb) <= 1)
1343 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1345 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1346 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1349 BUG_ON(!tcp_skb_pcount(skb));
1350 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1354 /* Whole SKB was eaten :-) */
1356 if (skb == tp->retransmit_skb_hint)
1357 tp->retransmit_skb_hint = prev;
1358 if (skb == tp->lost_skb_hint) {
1359 tp->lost_skb_hint = prev;
1360 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1363 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1364 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1365 TCP_SKB_CB(prev)->end_seq++;
1367 if (skb == tcp_highest_sack(sk))
1368 tcp_advance_highest_sack(sk, skb);
1370 tcp_unlink_write_queue(skb, sk);
1371 sk_wmem_free_skb(sk, skb);
1373 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1378 /* I wish gso_size would have a bit more sane initialization than
1379 * something-or-zero which complicates things
1381 static int tcp_skb_seglen(const struct sk_buff *skb)
1383 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1386 /* Shifting pages past head area doesn't work */
1387 static int skb_can_shift(const struct sk_buff *skb)
1389 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1392 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1395 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1396 struct tcp_sacktag_state *state,
1397 u32 start_seq, u32 end_seq,
1400 struct tcp_sock *tp = tcp_sk(sk);
1401 struct sk_buff *prev;
1407 if (!sk_can_gso(sk))
1410 /* Normally R but no L won't result in plain S */
1412 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1414 if (!skb_can_shift(skb))
1416 /* This frame is about to be dropped (was ACKed). */
1417 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1420 /* Can only happen with delayed DSACK + discard craziness */
1421 if (unlikely(skb == tcp_write_queue_head(sk)))
1423 prev = tcp_write_queue_prev(sk, skb);
1425 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1428 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1429 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1433 pcount = tcp_skb_pcount(skb);
1434 mss = tcp_skb_seglen(skb);
1436 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1437 * drop this restriction as unnecessary
1439 if (mss != tcp_skb_seglen(prev))
1442 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1444 /* CHECKME: This is non-MSS split case only?, this will
1445 * cause skipped skbs due to advancing loop btw, original
1446 * has that feature too
1448 if (tcp_skb_pcount(skb) <= 1)
1451 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1453 /* TODO: head merge to next could be attempted here
1454 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1455 * though it might not be worth of the additional hassle
1457 * ...we can probably just fallback to what was done
1458 * previously. We could try merging non-SACKed ones
1459 * as well but it probably isn't going to buy off
1460 * because later SACKs might again split them, and
1461 * it would make skb timestamp tracking considerably
1467 len = end_seq - TCP_SKB_CB(skb)->seq;
1469 BUG_ON(len > skb->len);
1471 /* MSS boundaries should be honoured or else pcount will
1472 * severely break even though it makes things bit trickier.
1473 * Optimize common case to avoid most of the divides
1475 mss = tcp_skb_mss(skb);
1477 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1478 * drop this restriction as unnecessary
1480 if (mss != tcp_skb_seglen(prev))
1485 } else if (len < mss) {
1493 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1494 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1497 if (!skb_shift(prev, skb, len))
1499 if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1502 /* Hole filled allows collapsing with the next as well, this is very
1503 * useful when hole on every nth skb pattern happens
1505 if (prev == tcp_write_queue_tail(sk))
1507 skb = tcp_write_queue_next(sk, prev);
1509 if (!skb_can_shift(skb) ||
1510 (skb == tcp_send_head(sk)) ||
1511 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1512 (mss != tcp_skb_seglen(skb)))
1516 if (skb_shift(prev, skb, len)) {
1517 pcount += tcp_skb_pcount(skb);
1518 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1522 state->fack_count += pcount;
1529 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1533 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1534 struct tcp_sack_block *next_dup,
1535 struct tcp_sacktag_state *state,
1536 u32 start_seq, u32 end_seq,
1539 struct tcp_sock *tp = tcp_sk(sk);
1540 struct sk_buff *tmp;
1542 tcp_for_write_queue_from(skb, sk) {
1544 bool dup_sack = dup_sack_in;
1546 if (skb == tcp_send_head(sk))
1549 /* queue is in-order => we can short-circuit the walk early */
1550 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1554 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1555 in_sack = tcp_match_skb_to_sack(sk, skb,
1556 next_dup->start_seq,
1562 /* skb reference here is a bit tricky to get right, since
1563 * shifting can eat and free both this skb and the next,
1564 * so not even _safe variant of the loop is enough.
1567 tmp = tcp_shift_skb_data(sk, skb, state,
1568 start_seq, end_seq, dup_sack);
1577 in_sack = tcp_match_skb_to_sack(sk, skb,
1583 if (unlikely(in_sack < 0))
1587 TCP_SKB_CB(skb)->sacked =
1590 TCP_SKB_CB(skb)->sacked,
1591 TCP_SKB_CB(skb)->seq,
1592 TCP_SKB_CB(skb)->end_seq,
1594 tcp_skb_pcount(skb),
1597 if (!before(TCP_SKB_CB(skb)->seq,
1598 tcp_highest_sack_seq(tp)))
1599 tcp_advance_highest_sack(sk, skb);
1602 state->fack_count += tcp_skb_pcount(skb);
1607 /* Avoid all extra work that is being done by sacktag while walking in
1610 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1611 struct tcp_sacktag_state *state,
1614 tcp_for_write_queue_from(skb, sk) {
1615 if (skb == tcp_send_head(sk))
1618 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1621 state->fack_count += tcp_skb_pcount(skb);
1626 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1628 struct tcp_sack_block *next_dup,
1629 struct tcp_sacktag_state *state,
1635 if (before(next_dup->start_seq, skip_to_seq)) {
1636 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1637 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1638 next_dup->start_seq, next_dup->end_seq,
1645 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1647 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1651 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1652 u32 prior_snd_una, struct tcp_sacktag_state *state)
1654 struct tcp_sock *tp = tcp_sk(sk);
1655 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1656 TCP_SKB_CB(ack_skb)->sacked);
1657 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1658 struct tcp_sack_block sp[TCP_NUM_SACKS];
1659 struct tcp_sack_block *cache;
1660 struct sk_buff *skb;
1661 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1663 bool found_dup_sack = false;
1665 int first_sack_index;
1668 state->reord = tp->packets_out;
1670 if (!tp->sacked_out) {
1671 if (WARN_ON(tp->fackets_out))
1672 tp->fackets_out = 0;
1673 tcp_highest_sack_reset(sk);
1676 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1677 num_sacks, prior_snd_una);
1679 state->flag |= FLAG_DSACKING_ACK;
1681 /* Eliminate too old ACKs, but take into
1682 * account more or less fresh ones, they can
1683 * contain valid SACK info.
1685 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1688 if (!tp->packets_out)
1692 first_sack_index = 0;
1693 for (i = 0; i < num_sacks; i++) {
1694 bool dup_sack = !i && found_dup_sack;
1696 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1697 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1699 if (!tcp_is_sackblock_valid(tp, dup_sack,
1700 sp[used_sacks].start_seq,
1701 sp[used_sacks].end_seq)) {
1705 if (!tp->undo_marker)
1706 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1708 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1710 /* Don't count olds caused by ACK reordering */
1711 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1712 !after(sp[used_sacks].end_seq, tp->snd_una))
1714 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1717 NET_INC_STATS_BH(sock_net(sk), mib_idx);
1719 first_sack_index = -1;
1723 /* Ignore very old stuff early */
1724 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1730 /* order SACK blocks to allow in order walk of the retrans queue */
1731 for (i = used_sacks - 1; i > 0; i--) {
1732 for (j = 0; j < i; j++) {
1733 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1734 swap(sp[j], sp[j + 1]);
1736 /* Track where the first SACK block goes to */
1737 if (j == first_sack_index)
1738 first_sack_index = j + 1;
1743 skb = tcp_write_queue_head(sk);
1744 state->fack_count = 0;
1747 if (!tp->sacked_out) {
1748 /* It's already past, so skip checking against it */
1749 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1751 cache = tp->recv_sack_cache;
1752 /* Skip empty blocks in at head of the cache */
1753 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1758 while (i < used_sacks) {
1759 u32 start_seq = sp[i].start_seq;
1760 u32 end_seq = sp[i].end_seq;
1761 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1762 struct tcp_sack_block *next_dup = NULL;
1764 if (found_dup_sack && ((i + 1) == first_sack_index))
1765 next_dup = &sp[i + 1];
1767 /* Skip too early cached blocks */
1768 while (tcp_sack_cache_ok(tp, cache) &&
1769 !before(start_seq, cache->end_seq))
1772 /* Can skip some work by looking recv_sack_cache? */
1773 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1774 after(end_seq, cache->start_seq)) {
1777 if (before(start_seq, cache->start_seq)) {
1778 skb = tcp_sacktag_skip(skb, sk, state,
1780 skb = tcp_sacktag_walk(skb, sk, next_dup,
1787 /* Rest of the block already fully processed? */
1788 if (!after(end_seq, cache->end_seq))
1791 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1795 /* ...tail remains todo... */
1796 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1797 /* ...but better entrypoint exists! */
1798 skb = tcp_highest_sack(sk);
1801 state->fack_count = tp->fackets_out;
1806 skb = tcp_sacktag_skip(skb, sk, state, cache->end_seq);
1807 /* Check overlap against next cached too (past this one already) */
1812 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1813 skb = tcp_highest_sack(sk);
1816 state->fack_count = tp->fackets_out;
1818 skb = tcp_sacktag_skip(skb, sk, state, start_seq);
1821 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
1822 start_seq, end_seq, dup_sack);
1828 /* Clear the head of the cache sack blocks so we can skip it next time */
1829 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1830 tp->recv_sack_cache[i].start_seq = 0;
1831 tp->recv_sack_cache[i].end_seq = 0;
1833 for (j = 0; j < used_sacks; j++)
1834 tp->recv_sack_cache[i++] = sp[j];
1836 if ((state->reord < tp->fackets_out) &&
1837 ((inet_csk(sk)->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker))
1838 tcp_update_reordering(sk, tp->fackets_out - state->reord, 0);
1840 tcp_mark_lost_retrans(sk, &state->flag);
1841 tcp_verify_left_out(tp);
1844 #if FASTRETRANS_DEBUG > 0
1845 WARN_ON((int)tp->sacked_out < 0);
1846 WARN_ON((int)tp->lost_out < 0);
1847 WARN_ON((int)tp->retrans_out < 0);
1848 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1853 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1854 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1856 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1860 holes = max(tp->lost_out, 1U);
1861 holes = min(holes, tp->packets_out);
1863 if ((tp->sacked_out + holes) > tp->packets_out) {
1864 tp->sacked_out = tp->packets_out - holes;
1870 /* If we receive more dupacks than we expected counting segments
1871 * in assumption of absent reordering, interpret this as reordering.
1872 * The only another reason could be bug in receiver TCP.
1874 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1876 struct tcp_sock *tp = tcp_sk(sk);
1877 if (tcp_limit_reno_sacked(tp))
1878 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1881 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1883 static void tcp_add_reno_sack(struct sock *sk)
1885 struct tcp_sock *tp = tcp_sk(sk);
1887 tcp_check_reno_reordering(sk, 0);
1888 tcp_verify_left_out(tp);
1891 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1893 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1895 struct tcp_sock *tp = tcp_sk(sk);
1898 /* One ACK acked hole. The rest eat duplicate ACKs. */
1899 if (acked - 1 >= tp->sacked_out)
1902 tp->sacked_out -= acked - 1;
1904 tcp_check_reno_reordering(sk, acked);
1905 tcp_verify_left_out(tp);
1908 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1913 void tcp_clear_retrans(struct tcp_sock *tp)
1915 tp->retrans_out = 0;
1917 tp->undo_marker = 0;
1918 tp->undo_retrans = -1;
1919 tp->fackets_out = 0;
1923 static inline void tcp_init_undo(struct tcp_sock *tp)
1925 tp->undo_marker = tp->snd_una;
1926 /* Retransmission still in flight may cause DSACKs later. */
1927 tp->undo_retrans = tp->retrans_out ? : -1;
1930 /* Enter Loss state. If we detect SACK reneging, forget all SACK information
1931 * and reset tags completely, otherwise preserve SACKs. If receiver
1932 * dropped its ofo queue, we will know this due to reneging detection.
1934 void tcp_enter_loss(struct sock *sk)
1936 const struct inet_connection_sock *icsk = inet_csk(sk);
1937 struct tcp_sock *tp = tcp_sk(sk);
1938 struct sk_buff *skb;
1939 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
1940 bool is_reneg; /* is receiver reneging on SACKs? */
1942 /* Reduce ssthresh if it has not yet been made inside this window. */
1943 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1944 !after(tp->high_seq, tp->snd_una) ||
1945 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1946 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1947 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1948 tcp_ca_event(sk, CA_EVENT_LOSS);
1952 tp->snd_cwnd_cnt = 0;
1953 tp->snd_cwnd_stamp = tcp_time_stamp;
1955 tp->retrans_out = 0;
1958 if (tcp_is_reno(tp))
1959 tcp_reset_reno_sack(tp);
1961 skb = tcp_write_queue_head(sk);
1962 is_reneg = skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED);
1964 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
1966 tp->fackets_out = 0;
1968 tcp_clear_all_retrans_hints(tp);
1970 tcp_for_write_queue(skb, sk) {
1971 if (skb == tcp_send_head(sk))
1974 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1975 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || is_reneg) {
1976 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1977 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1978 tp->lost_out += tcp_skb_pcount(skb);
1979 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
1982 tcp_verify_left_out(tp);
1984 /* Timeout in disordered state after receiving substantial DUPACKs
1985 * suggests that the degree of reordering is over-estimated.
1987 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
1988 tp->sacked_out >= sysctl_tcp_reordering)
1989 tp->reordering = min_t(unsigned int, tp->reordering,
1990 sysctl_tcp_reordering);
1991 tcp_set_ca_state(sk, TCP_CA_Loss);
1992 tp->high_seq = tp->snd_nxt;
1993 tcp_ecn_queue_cwr(tp);
1995 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1996 * loss recovery is underway except recurring timeout(s) on
1997 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1999 tp->frto = sysctl_tcp_frto &&
2000 (new_recovery || icsk->icsk_retransmits) &&
2001 !inet_csk(sk)->icsk_mtup.probe_size;
2004 /* If ACK arrived pointing to a remembered SACK, it means that our
2005 * remembered SACKs do not reflect real state of receiver i.e.
2006 * receiver _host_ is heavily congested (or buggy).
2008 * To avoid big spurious retransmission bursts due to transient SACK
2009 * scoreboard oddities that look like reneging, we give the receiver a
2010 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2011 * restore sanity to the SACK scoreboard. If the apparent reneging
2012 * persists until this RTO then we'll clear the SACK scoreboard.
2014 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
2016 if (flag & FLAG_SACK_RENEGING) {
2017 struct tcp_sock *tp = tcp_sk(sk);
2018 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2019 msecs_to_jiffies(10));
2021 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2022 delay, TCP_RTO_MAX);
2028 static inline int tcp_fackets_out(const struct tcp_sock *tp)
2030 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
2033 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2034 * counter when SACK is enabled (without SACK, sacked_out is used for
2037 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2038 * segments up to the highest received SACK block so far and holes in
2041 * With reordering, holes may still be in flight, so RFC3517 recovery
2042 * uses pure sacked_out (total number of SACKed segments) even though
2043 * it violates the RFC that uses duplicate ACKs, often these are equal
2044 * but when e.g. out-of-window ACKs or packet duplication occurs,
2045 * they differ. Since neither occurs due to loss, TCP should really
2048 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2050 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2053 static bool tcp_pause_early_retransmit(struct sock *sk, int flag)
2055 struct tcp_sock *tp = tcp_sk(sk);
2056 unsigned long delay;
2058 /* Delay early retransmit and entering fast recovery for
2059 * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
2060 * available, or RTO is scheduled to fire first.
2062 if (sysctl_tcp_early_retrans < 2 || sysctl_tcp_early_retrans > 3 ||
2063 (flag & FLAG_ECE) || !tp->srtt_us)
2066 delay = max(usecs_to_jiffies(tp->srtt_us >> 5),
2067 msecs_to_jiffies(2));
2069 if (!time_after(inet_csk(sk)->icsk_timeout, (jiffies + delay)))
2072 inet_csk_reset_xmit_timer(sk, ICSK_TIME_EARLY_RETRANS, delay,
2077 /* Linux NewReno/SACK/FACK/ECN state machine.
2078 * --------------------------------------
2080 * "Open" Normal state, no dubious events, fast path.
2081 * "Disorder" In all the respects it is "Open",
2082 * but requires a bit more attention. It is entered when
2083 * we see some SACKs or dupacks. It is split of "Open"
2084 * mainly to move some processing from fast path to slow one.
2085 * "CWR" CWND was reduced due to some Congestion Notification event.
2086 * It can be ECN, ICMP source quench, local device congestion.
2087 * "Recovery" CWND was reduced, we are fast-retransmitting.
2088 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2090 * tcp_fastretrans_alert() is entered:
2091 * - each incoming ACK, if state is not "Open"
2092 * - when arrived ACK is unusual, namely:
2097 * Counting packets in flight is pretty simple.
2099 * in_flight = packets_out - left_out + retrans_out
2101 * packets_out is SND.NXT-SND.UNA counted in packets.
2103 * retrans_out is number of retransmitted segments.
2105 * left_out is number of segments left network, but not ACKed yet.
2107 * left_out = sacked_out + lost_out
2109 * sacked_out: Packets, which arrived to receiver out of order
2110 * and hence not ACKed. With SACKs this number is simply
2111 * amount of SACKed data. Even without SACKs
2112 * it is easy to give pretty reliable estimate of this number,
2113 * counting duplicate ACKs.
2115 * lost_out: Packets lost by network. TCP has no explicit
2116 * "loss notification" feedback from network (for now).
2117 * It means that this number can be only _guessed_.
2118 * Actually, it is the heuristics to predict lossage that
2119 * distinguishes different algorithms.
2121 * F.e. after RTO, when all the queue is considered as lost,
2122 * lost_out = packets_out and in_flight = retrans_out.
2124 * Essentially, we have now two algorithms counting
2127 * FACK: It is the simplest heuristics. As soon as we decided
2128 * that something is lost, we decide that _all_ not SACKed
2129 * packets until the most forward SACK are lost. I.e.
2130 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2131 * It is absolutely correct estimate, if network does not reorder
2132 * packets. And it loses any connection to reality when reordering
2133 * takes place. We use FACK by default until reordering
2134 * is suspected on the path to this destination.
2136 * NewReno: when Recovery is entered, we assume that one segment
2137 * is lost (classic Reno). While we are in Recovery and
2138 * a partial ACK arrives, we assume that one more packet
2139 * is lost (NewReno). This heuristics are the same in NewReno
2142 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2143 * deflation etc. CWND is real congestion window, never inflated, changes
2144 * only according to classic VJ rules.
2146 * Really tricky (and requiring careful tuning) part of algorithm
2147 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2148 * The first determines the moment _when_ we should reduce CWND and,
2149 * hence, slow down forward transmission. In fact, it determines the moment
2150 * when we decide that hole is caused by loss, rather than by a reorder.
2152 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2153 * holes, caused by lost packets.
2155 * And the most logically complicated part of algorithm is undo
2156 * heuristics. We detect false retransmits due to both too early
2157 * fast retransmit (reordering) and underestimated RTO, analyzing
2158 * timestamps and D-SACKs. When we detect that some segments were
2159 * retransmitted by mistake and CWND reduction was wrong, we undo
2160 * window reduction and abort recovery phase. This logic is hidden
2161 * inside several functions named tcp_try_undo_<something>.
2164 /* This function decides, when we should leave Disordered state
2165 * and enter Recovery phase, reducing congestion window.
2167 * Main question: may we further continue forward transmission
2168 * with the same cwnd?
2170 static bool tcp_time_to_recover(struct sock *sk, int flag)
2172 struct tcp_sock *tp = tcp_sk(sk);
2175 /* Trick#1: The loss is proven. */
2179 /* Not-A-Trick#2 : Classic rule... */
2180 if (tcp_dupack_heuristics(tp) > tp->reordering)
2183 /* Trick#4: It is still not OK... But will it be useful to delay
2186 packets_out = tp->packets_out;
2187 if (packets_out <= tp->reordering &&
2188 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2189 !tcp_may_send_now(sk)) {
2190 /* We have nothing to send. This connection is limited
2191 * either by receiver window or by application.
2196 /* If a thin stream is detected, retransmit after first
2197 * received dupack. Employ only if SACK is supported in order
2198 * to avoid possible corner-case series of spurious retransmissions
2199 * Use only if there are no unsent data.
2201 if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2202 tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2203 tcp_is_sack(tp) && !tcp_send_head(sk))
2206 /* Trick#6: TCP early retransmit, per RFC5827. To avoid spurious
2207 * retransmissions due to small network reorderings, we implement
2208 * Mitigation A.3 in the RFC and delay the retransmission for a short
2209 * interval if appropriate.
2211 if (tp->do_early_retrans && !tp->retrans_out && tp->sacked_out &&
2212 (tp->packets_out >= (tp->sacked_out + 1) && tp->packets_out < 4) &&
2213 !tcp_may_send_now(sk))
2214 return !tcp_pause_early_retransmit(sk, flag);
2219 /* Detect loss in event "A" above by marking head of queue up as lost.
2220 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2221 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2222 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2223 * the maximum SACKed segments to pass before reaching this limit.
2225 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2227 struct tcp_sock *tp = tcp_sk(sk);
2228 struct sk_buff *skb;
2232 /* Use SACK to deduce losses of new sequences sent during recovery */
2233 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq;
2235 WARN_ON(packets > tp->packets_out);
2236 if (tp->lost_skb_hint) {
2237 skb = tp->lost_skb_hint;
2238 cnt = tp->lost_cnt_hint;
2239 /* Head already handled? */
2240 if (mark_head && skb != tcp_write_queue_head(sk))
2243 skb = tcp_write_queue_head(sk);
2247 tcp_for_write_queue_from(skb, sk) {
2248 if (skb == tcp_send_head(sk))
2250 /* TODO: do this better */
2251 /* this is not the most efficient way to do this... */
2252 tp->lost_skb_hint = skb;
2253 tp->lost_cnt_hint = cnt;
2255 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2259 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2260 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2261 cnt += tcp_skb_pcount(skb);
2263 if (cnt > packets) {
2264 if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2265 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2266 (oldcnt >= packets))
2269 mss = tcp_skb_mss(skb);
2270 err = tcp_fragment(sk, skb, (packets - oldcnt) * mss,
2277 tcp_skb_mark_lost(tp, skb);
2282 tcp_verify_left_out(tp);
2285 /* Account newly detected lost packet(s) */
2287 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2289 struct tcp_sock *tp = tcp_sk(sk);
2291 if (tcp_is_reno(tp)) {
2292 tcp_mark_head_lost(sk, 1, 1);
2293 } else if (tcp_is_fack(tp)) {
2294 int lost = tp->fackets_out - tp->reordering;
2297 tcp_mark_head_lost(sk, lost, 0);
2299 int sacked_upto = tp->sacked_out - tp->reordering;
2300 if (sacked_upto >= 0)
2301 tcp_mark_head_lost(sk, sacked_upto, 0);
2302 else if (fast_rexmit)
2303 tcp_mark_head_lost(sk, 1, 1);
2307 /* CWND moderation, preventing bursts due to too big ACKs
2308 * in dubious situations.
2310 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2312 tp->snd_cwnd = min(tp->snd_cwnd,
2313 tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2314 tp->snd_cwnd_stamp = tcp_time_stamp;
2317 /* Nothing was retransmitted or returned timestamp is less
2318 * than timestamp of the first retransmission.
2320 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2322 return !tp->retrans_stamp ||
2323 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2324 before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
2327 /* Undo procedures. */
2329 /* We can clear retrans_stamp when there are no retransmissions in the
2330 * window. It would seem that it is trivially available for us in
2331 * tp->retrans_out, however, that kind of assumptions doesn't consider
2332 * what will happen if errors occur when sending retransmission for the
2333 * second time. ...It could the that such segment has only
2334 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2335 * the head skb is enough except for some reneging corner cases that
2336 * are not worth the effort.
2338 * Main reason for all this complexity is the fact that connection dying
2339 * time now depends on the validity of the retrans_stamp, in particular,
2340 * that successive retransmissions of a segment must not advance
2341 * retrans_stamp under any conditions.
2343 static bool tcp_any_retrans_done(const struct sock *sk)
2345 const struct tcp_sock *tp = tcp_sk(sk);
2346 struct sk_buff *skb;
2348 if (tp->retrans_out)
2351 skb = tcp_write_queue_head(sk);
2352 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2358 #if FASTRETRANS_DEBUG > 1
2359 static void DBGUNDO(struct sock *sk, const char *msg)
2361 struct tcp_sock *tp = tcp_sk(sk);
2362 struct inet_sock *inet = inet_sk(sk);
2364 if (sk->sk_family == AF_INET) {
2365 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2367 &inet->inet_daddr, ntohs(inet->inet_dport),
2368 tp->snd_cwnd, tcp_left_out(tp),
2369 tp->snd_ssthresh, tp->prior_ssthresh,
2372 #if IS_ENABLED(CONFIG_IPV6)
2373 else if (sk->sk_family == AF_INET6) {
2374 struct ipv6_pinfo *np = inet6_sk(sk);
2375 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2377 &np->daddr, ntohs(inet->inet_dport),
2378 tp->snd_cwnd, tcp_left_out(tp),
2379 tp->snd_ssthresh, tp->prior_ssthresh,
2385 #define DBGUNDO(x...) do { } while (0)
2388 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2390 struct tcp_sock *tp = tcp_sk(sk);
2393 struct sk_buff *skb;
2395 tcp_for_write_queue(skb, sk) {
2396 if (skb == tcp_send_head(sk))
2398 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2401 tcp_clear_all_retrans_hints(tp);
2404 if (tp->prior_ssthresh) {
2405 const struct inet_connection_sock *icsk = inet_csk(sk);
2407 if (icsk->icsk_ca_ops->undo_cwnd)
2408 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2410 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2412 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2413 tp->snd_ssthresh = tp->prior_ssthresh;
2414 tcp_ecn_withdraw_cwr(tp);
2417 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2419 tp->snd_cwnd_stamp = tcp_time_stamp;
2420 tp->undo_marker = 0;
2423 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2425 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2428 /* People celebrate: "We love our President!" */
2429 static bool tcp_try_undo_recovery(struct sock *sk)
2431 struct tcp_sock *tp = tcp_sk(sk);
2433 if (tcp_may_undo(tp)) {
2436 /* Happy end! We did not retransmit anything
2437 * or our original transmission succeeded.
2439 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2440 tcp_undo_cwnd_reduction(sk, false);
2441 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2442 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2444 mib_idx = LINUX_MIB_TCPFULLUNDO;
2446 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2448 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2449 /* Hold old state until something *above* high_seq
2450 * is ACKed. For Reno it is MUST to prevent false
2451 * fast retransmits (RFC2582). SACK TCP is safe. */
2452 tcp_moderate_cwnd(tp);
2453 if (!tcp_any_retrans_done(sk))
2454 tp->retrans_stamp = 0;
2457 tcp_set_ca_state(sk, TCP_CA_Open);
2461 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2462 static bool tcp_try_undo_dsack(struct sock *sk)
2464 struct tcp_sock *tp = tcp_sk(sk);
2466 if (tp->undo_marker && !tp->undo_retrans) {
2467 DBGUNDO(sk, "D-SACK");
2468 tcp_undo_cwnd_reduction(sk, false);
2469 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2475 /* Undo during loss recovery after partial ACK or using F-RTO. */
2476 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2478 struct tcp_sock *tp = tcp_sk(sk);
2480 if (frto_undo || tcp_may_undo(tp)) {
2481 tcp_undo_cwnd_reduction(sk, true);
2483 DBGUNDO(sk, "partial loss");
2484 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2486 NET_INC_STATS_BH(sock_net(sk),
2487 LINUX_MIB_TCPSPURIOUSRTOS);
2488 inet_csk(sk)->icsk_retransmits = 0;
2489 if (frto_undo || tcp_is_sack(tp))
2490 tcp_set_ca_state(sk, TCP_CA_Open);
2496 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2497 * It computes the number of packets to send (sndcnt) based on packets newly
2499 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2500 * cwnd reductions across a full RTT.
2501 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2502 * But when the retransmits are acked without further losses, PRR
2503 * slow starts cwnd up to ssthresh to speed up the recovery.
2505 static void tcp_init_cwnd_reduction(struct sock *sk)
2507 struct tcp_sock *tp = tcp_sk(sk);
2509 tp->high_seq = tp->snd_nxt;
2510 tp->tlp_high_seq = 0;
2511 tp->snd_cwnd_cnt = 0;
2512 tp->prior_cwnd = tp->snd_cwnd;
2513 tp->prr_delivered = 0;
2515 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2516 tcp_ecn_queue_cwr(tp);
2519 static void tcp_cwnd_reduction(struct sock *sk, const int prior_unsacked,
2520 int fast_rexmit, int flag)
2522 struct tcp_sock *tp = tcp_sk(sk);
2524 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2525 int newly_acked_sacked = prior_unsacked -
2526 (tp->packets_out - tp->sacked_out);
2528 tp->prr_delivered += newly_acked_sacked;
2530 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2532 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2533 } else if ((flag & FLAG_RETRANS_DATA_ACKED) &&
2534 !(flag & FLAG_LOST_RETRANS)) {
2535 sndcnt = min_t(int, delta,
2536 max_t(int, tp->prr_delivered - tp->prr_out,
2537 newly_acked_sacked) + 1);
2539 sndcnt = min(delta, newly_acked_sacked);
2541 sndcnt = max(sndcnt, (fast_rexmit ? 1 : 0));
2542 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2545 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2547 struct tcp_sock *tp = tcp_sk(sk);
2549 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2550 if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR ||
2551 (tp->undo_marker && tp->snd_ssthresh < TCP_INFINITE_SSTHRESH)) {
2552 tp->snd_cwnd = tp->snd_ssthresh;
2553 tp->snd_cwnd_stamp = tcp_time_stamp;
2555 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2558 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2559 void tcp_enter_cwr(struct sock *sk)
2561 struct tcp_sock *tp = tcp_sk(sk);
2563 tp->prior_ssthresh = 0;
2564 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2565 tp->undo_marker = 0;
2566 tcp_init_cwnd_reduction(sk);
2567 tcp_set_ca_state(sk, TCP_CA_CWR);
2570 EXPORT_SYMBOL(tcp_enter_cwr);
2572 static void tcp_try_keep_open(struct sock *sk)
2574 struct tcp_sock *tp = tcp_sk(sk);
2575 int state = TCP_CA_Open;
2577 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2578 state = TCP_CA_Disorder;
2580 if (inet_csk(sk)->icsk_ca_state != state) {
2581 tcp_set_ca_state(sk, state);
2582 tp->high_seq = tp->snd_nxt;
2586 static void tcp_try_to_open(struct sock *sk, int flag, const int prior_unsacked)
2588 struct tcp_sock *tp = tcp_sk(sk);
2590 tcp_verify_left_out(tp);
2592 if (!tcp_any_retrans_done(sk))
2593 tp->retrans_stamp = 0;
2595 if (flag & FLAG_ECE)
2598 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2599 tcp_try_keep_open(sk);
2601 tcp_cwnd_reduction(sk, prior_unsacked, 0, flag);
2605 static void tcp_mtup_probe_failed(struct sock *sk)
2607 struct inet_connection_sock *icsk = inet_csk(sk);
2609 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2610 icsk->icsk_mtup.probe_size = 0;
2611 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
2614 static void tcp_mtup_probe_success(struct sock *sk)
2616 struct tcp_sock *tp = tcp_sk(sk);
2617 struct inet_connection_sock *icsk = inet_csk(sk);
2619 /* FIXME: breaks with very large cwnd */
2620 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2621 tp->snd_cwnd = tp->snd_cwnd *
2622 tcp_mss_to_mtu(sk, tp->mss_cache) /
2623 icsk->icsk_mtup.probe_size;
2624 tp->snd_cwnd_cnt = 0;
2625 tp->snd_cwnd_stamp = tcp_time_stamp;
2626 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2628 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2629 icsk->icsk_mtup.probe_size = 0;
2630 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2631 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
2634 /* Do a simple retransmit without using the backoff mechanisms in
2635 * tcp_timer. This is used for path mtu discovery.
2636 * The socket is already locked here.
2638 void tcp_simple_retransmit(struct sock *sk)
2640 const struct inet_connection_sock *icsk = inet_csk(sk);
2641 struct tcp_sock *tp = tcp_sk(sk);
2642 struct sk_buff *skb;
2643 unsigned int mss = tcp_current_mss(sk);
2644 u32 prior_lost = tp->lost_out;
2646 tcp_for_write_queue(skb, sk) {
2647 if (skb == tcp_send_head(sk))
2649 if (tcp_skb_seglen(skb) > mss &&
2650 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2651 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2652 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2653 tp->retrans_out -= tcp_skb_pcount(skb);
2655 tcp_skb_mark_lost_uncond_verify(tp, skb);
2659 tcp_clear_retrans_hints_partial(tp);
2661 if (prior_lost == tp->lost_out)
2664 if (tcp_is_reno(tp))
2665 tcp_limit_reno_sacked(tp);
2667 tcp_verify_left_out(tp);
2669 /* Don't muck with the congestion window here.
2670 * Reason is that we do not increase amount of _data_
2671 * in network, but units changed and effective
2672 * cwnd/ssthresh really reduced now.
2674 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2675 tp->high_seq = tp->snd_nxt;
2676 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2677 tp->prior_ssthresh = 0;
2678 tp->undo_marker = 0;
2679 tcp_set_ca_state(sk, TCP_CA_Loss);
2681 tcp_xmit_retransmit_queue(sk);
2683 EXPORT_SYMBOL(tcp_simple_retransmit);
2685 static void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2687 struct tcp_sock *tp = tcp_sk(sk);
2690 if (tcp_is_reno(tp))
2691 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2693 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2695 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2697 tp->prior_ssthresh = 0;
2700 if (!tcp_in_cwnd_reduction(sk)) {
2702 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2703 tcp_init_cwnd_reduction(sk);
2705 tcp_set_ca_state(sk, TCP_CA_Recovery);
2708 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2709 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2711 static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack)
2713 struct tcp_sock *tp = tcp_sk(sk);
2714 bool recovered = !before(tp->snd_una, tp->high_seq);
2716 if ((flag & FLAG_SND_UNA_ADVANCED) &&
2717 tcp_try_undo_loss(sk, false))
2720 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2721 /* Step 3.b. A timeout is spurious if not all data are
2722 * lost, i.e., never-retransmitted data are (s)acked.
2724 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2725 tcp_try_undo_loss(sk, true))
2728 if (after(tp->snd_nxt, tp->high_seq)) {
2729 if (flag & FLAG_DATA_SACKED || is_dupack)
2730 tp->frto = 0; /* Step 3.a. loss was real */
2731 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2732 tp->high_seq = tp->snd_nxt;
2733 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
2735 if (after(tp->snd_nxt, tp->high_seq))
2736 return; /* Step 2.b */
2742 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2743 tcp_try_undo_recovery(sk);
2746 if (tcp_is_reno(tp)) {
2747 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2748 * delivered. Lower inflight to clock out (re)tranmissions.
2750 if (after(tp->snd_nxt, tp->high_seq) && is_dupack)
2751 tcp_add_reno_sack(sk);
2752 else if (flag & FLAG_SND_UNA_ADVANCED)
2753 tcp_reset_reno_sack(tp);
2755 tcp_xmit_retransmit_queue(sk);
2758 /* Undo during fast recovery after partial ACK. */
2759 static bool tcp_try_undo_partial(struct sock *sk, const int acked,
2760 const int prior_unsacked, int flag)
2762 struct tcp_sock *tp = tcp_sk(sk);
2764 if (tp->undo_marker && tcp_packet_delayed(tp)) {
2765 /* Plain luck! Hole if filled with delayed
2766 * packet, rather than with a retransmit.
2768 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2770 /* We are getting evidence that the reordering degree is higher
2771 * than we realized. If there are no retransmits out then we
2772 * can undo. Otherwise we clock out new packets but do not
2773 * mark more packets lost or retransmit more.
2775 if (tp->retrans_out) {
2776 tcp_cwnd_reduction(sk, prior_unsacked, 0, flag);
2780 if (!tcp_any_retrans_done(sk))
2781 tp->retrans_stamp = 0;
2783 DBGUNDO(sk, "partial recovery");
2784 tcp_undo_cwnd_reduction(sk, true);
2785 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2786 tcp_try_keep_open(sk);
2792 /* Process an event, which can update packets-in-flight not trivially.
2793 * Main goal of this function is to calculate new estimate for left_out,
2794 * taking into account both packets sitting in receiver's buffer and
2795 * packets lost by network.
2797 * Besides that it does CWND reduction, when packet loss is detected
2798 * and changes state of machine.
2800 * It does _not_ decide what to send, it is made in function
2801 * tcp_xmit_retransmit_queue().
2803 static void tcp_fastretrans_alert(struct sock *sk, const int acked,
2804 const int prior_unsacked,
2805 bool is_dupack, int flag)
2807 struct inet_connection_sock *icsk = inet_csk(sk);
2808 struct tcp_sock *tp = tcp_sk(sk);
2809 bool do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2810 (tcp_fackets_out(tp) > tp->reordering));
2811 int fast_rexmit = 0;
2813 if (WARN_ON(!tp->packets_out && tp->sacked_out))
2815 if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2816 tp->fackets_out = 0;
2818 /* Now state machine starts.
2819 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2820 if (flag & FLAG_ECE)
2821 tp->prior_ssthresh = 0;
2823 /* B. In all the states check for reneging SACKs. */
2824 if (tcp_check_sack_reneging(sk, flag))
2827 /* C. Check consistency of the current state. */
2828 tcp_verify_left_out(tp);
2830 /* D. Check state exit conditions. State can be terminated
2831 * when high_seq is ACKed. */
2832 if (icsk->icsk_ca_state == TCP_CA_Open) {
2833 WARN_ON(tp->retrans_out != 0);
2834 tp->retrans_stamp = 0;
2835 } else if (!before(tp->snd_una, tp->high_seq)) {
2836 switch (icsk->icsk_ca_state) {
2838 /* CWR is to be held something *above* high_seq
2839 * is ACKed for CWR bit to reach receiver. */
2840 if (tp->snd_una != tp->high_seq) {
2841 tcp_end_cwnd_reduction(sk);
2842 tcp_set_ca_state(sk, TCP_CA_Open);
2846 case TCP_CA_Recovery:
2847 if (tcp_is_reno(tp))
2848 tcp_reset_reno_sack(tp);
2849 if (tcp_try_undo_recovery(sk))
2851 tcp_end_cwnd_reduction(sk);
2856 /* E. Process state. */
2857 switch (icsk->icsk_ca_state) {
2858 case TCP_CA_Recovery:
2859 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2860 if (tcp_is_reno(tp) && is_dupack)
2861 tcp_add_reno_sack(sk);
2863 if (tcp_try_undo_partial(sk, acked, prior_unsacked, flag))
2865 /* Partial ACK arrived. Force fast retransmit. */
2866 do_lost = tcp_is_reno(tp) ||
2867 tcp_fackets_out(tp) > tp->reordering;
2869 if (tcp_try_undo_dsack(sk)) {
2870 tcp_try_keep_open(sk);
2875 tcp_process_loss(sk, flag, is_dupack);
2876 if (icsk->icsk_ca_state != TCP_CA_Open &&
2877 !(flag & FLAG_LOST_RETRANS))
2879 /* Change state if cwnd is undone or retransmits are lost */
2881 if (tcp_is_reno(tp)) {
2882 if (flag & FLAG_SND_UNA_ADVANCED)
2883 tcp_reset_reno_sack(tp);
2885 tcp_add_reno_sack(sk);
2888 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2889 tcp_try_undo_dsack(sk);
2891 if (!tcp_time_to_recover(sk, flag)) {
2892 tcp_try_to_open(sk, flag, prior_unsacked);
2896 /* MTU probe failure: don't reduce cwnd */
2897 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2898 icsk->icsk_mtup.probe_size &&
2899 tp->snd_una == tp->mtu_probe.probe_seq_start) {
2900 tcp_mtup_probe_failed(sk);
2901 /* Restores the reduction we did in tcp_mtup_probe() */
2903 tcp_simple_retransmit(sk);
2907 /* Otherwise enter Recovery state */
2908 tcp_enter_recovery(sk, (flag & FLAG_ECE));
2913 tcp_update_scoreboard(sk, fast_rexmit);
2914 tcp_cwnd_reduction(sk, prior_unsacked, fast_rexmit, flag);
2915 tcp_xmit_retransmit_queue(sk);
2918 static inline bool tcp_ack_update_rtt(struct sock *sk, const int flag,
2919 long seq_rtt_us, long sack_rtt_us)
2921 const struct tcp_sock *tp = tcp_sk(sk);
2923 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2924 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2925 * Karn's algorithm forbids taking RTT if some retransmitted data
2926 * is acked (RFC6298).
2928 if (flag & FLAG_RETRANS_DATA_ACKED)
2932 seq_rtt_us = sack_rtt_us;
2934 /* RTTM Rule: A TSecr value received in a segment is used to
2935 * update the averaged RTT measurement only if the segment
2936 * acknowledges some new data, i.e., only if it advances the
2937 * left edge of the send window.
2938 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2940 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2942 seq_rtt_us = jiffies_to_usecs(tcp_time_stamp - tp->rx_opt.rcv_tsecr);
2947 tcp_rtt_estimator(sk, seq_rtt_us);
2950 /* RFC6298: only reset backoff on valid RTT measurement. */
2951 inet_csk(sk)->icsk_backoff = 0;
2955 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2956 static void tcp_synack_rtt_meas(struct sock *sk, const u32 synack_stamp)
2958 struct tcp_sock *tp = tcp_sk(sk);
2959 long seq_rtt_us = -1L;
2961 if (synack_stamp && !tp->total_retrans)
2962 seq_rtt_us = jiffies_to_usecs(tcp_time_stamp - synack_stamp);
2964 /* If the ACK acks both the SYNACK and the (Fast Open'd) data packets
2965 * sent in SYN_RECV, SYNACK RTT is the smooth RTT computed in tcp_ack()
2968 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, seq_rtt_us, -1L);
2971 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
2973 const struct inet_connection_sock *icsk = inet_csk(sk);
2975 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
2976 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
2979 /* Restart timer after forward progress on connection.
2980 * RFC2988 recommends to restart timer to now+rto.
2982 void tcp_rearm_rto(struct sock *sk)
2984 const struct inet_connection_sock *icsk = inet_csk(sk);
2985 struct tcp_sock *tp = tcp_sk(sk);
2987 /* If the retrans timer is currently being used by Fast Open
2988 * for SYN-ACK retrans purpose, stay put.
2990 if (tp->fastopen_rsk)
2993 if (!tp->packets_out) {
2994 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2996 u32 rto = inet_csk(sk)->icsk_rto;
2997 /* Offset the time elapsed after installing regular RTO */
2998 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
2999 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
3000 struct sk_buff *skb = tcp_write_queue_head(sk);
3001 const u32 rto_time_stamp =
3002 tcp_skb_timestamp(skb) + rto;
3003 s32 delta = (s32)(rto_time_stamp - tcp_time_stamp);
3004 /* delta may not be positive if the socket is locked
3005 * when the retrans timer fires and is rescheduled.
3010 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
3015 /* This function is called when the delayed ER timer fires. TCP enters
3016 * fast recovery and performs fast-retransmit.
3018 void tcp_resume_early_retransmit(struct sock *sk)
3020 struct tcp_sock *tp = tcp_sk(sk);
3024 /* Stop if ER is disabled after the delayed ER timer is scheduled */
3025 if (!tp->do_early_retrans)
3028 tcp_enter_recovery(sk, false);
3029 tcp_update_scoreboard(sk, 1);
3030 tcp_xmit_retransmit_queue(sk);
3033 /* If we get here, the whole TSO packet has not been acked. */
3034 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3036 struct tcp_sock *tp = tcp_sk(sk);
3039 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3041 packets_acked = tcp_skb_pcount(skb);
3042 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3044 packets_acked -= tcp_skb_pcount(skb);
3046 if (packets_acked) {
3047 BUG_ON(tcp_skb_pcount(skb) == 0);
3048 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3051 return packets_acked;
3054 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3057 const struct skb_shared_info *shinfo;
3059 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3060 if (likely(!(sk->sk_tsflags & SOF_TIMESTAMPING_TX_ACK)))
3063 shinfo = skb_shinfo(skb);
3064 if ((shinfo->tx_flags & SKBTX_ACK_TSTAMP) &&
3065 between(shinfo->tskey, prior_snd_una, tcp_sk(sk)->snd_una - 1))
3066 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK);
3069 /* Remove acknowledged frames from the retransmission queue. If our packet
3070 * is before the ack sequence we can discard it as it's confirmed to have
3071 * arrived at the other end.
3073 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3075 struct tcp_sacktag_state *sack)
3077 const struct inet_connection_sock *icsk = inet_csk(sk);
3078 struct skb_mstamp first_ackt, last_ackt, now;
3079 struct tcp_sock *tp = tcp_sk(sk);
3080 u32 prior_sacked = tp->sacked_out;
3081 u32 reord = tp->packets_out;
3082 bool fully_acked = true;
3083 long sack_rtt_us = -1L;
3084 long seq_rtt_us = -1L;
3085 long ca_rtt_us = -1L;
3086 struct sk_buff *skb;
3093 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3094 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3095 u8 sacked = scb->sacked;
3098 tcp_ack_tstamp(sk, skb, prior_snd_una);
3100 /* Determine how many packets and what bytes were acked, tso and else */
3101 if (after(scb->end_seq, tp->snd_una)) {
3102 if (tcp_skb_pcount(skb) == 1 ||
3103 !after(tp->snd_una, scb->seq))
3106 acked_pcount = tcp_tso_acked(sk, skb);
3110 fully_acked = false;
3112 /* Speedup tcp_unlink_write_queue() and next loop */
3113 prefetchw(skb->next);
3114 acked_pcount = tcp_skb_pcount(skb);
3117 if (unlikely(sacked & TCPCB_RETRANS)) {
3118 if (sacked & TCPCB_SACKED_RETRANS)
3119 tp->retrans_out -= acked_pcount;
3120 flag |= FLAG_RETRANS_DATA_ACKED;
3121 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3122 last_ackt = skb->skb_mstamp;
3123 WARN_ON_ONCE(last_ackt.v64 == 0);
3124 if (!first_ackt.v64)
3125 first_ackt = last_ackt;
3127 reord = min(pkts_acked, reord);
3128 if (!after(scb->end_seq, tp->high_seq))
3129 flag |= FLAG_ORIG_SACK_ACKED;
3132 if (sacked & TCPCB_SACKED_ACKED)
3133 tp->sacked_out -= acked_pcount;
3134 if (sacked & TCPCB_LOST)
3135 tp->lost_out -= acked_pcount;
3137 tp->packets_out -= acked_pcount;
3138 pkts_acked += acked_pcount;
3140 /* Initial outgoing SYN's get put onto the write_queue
3141 * just like anything else we transmit. It is not
3142 * true data, and if we misinform our callers that
3143 * this ACK acks real data, we will erroneously exit
3144 * connection startup slow start one packet too
3145 * quickly. This is severely frowned upon behavior.
3147 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3148 flag |= FLAG_DATA_ACKED;
3150 flag |= FLAG_SYN_ACKED;
3151 tp->retrans_stamp = 0;
3157 tcp_unlink_write_queue(skb, sk);
3158 sk_wmem_free_skb(sk, skb);
3159 if (unlikely(skb == tp->retransmit_skb_hint))
3160 tp->retransmit_skb_hint = NULL;
3161 if (unlikely(skb == tp->lost_skb_hint))
3162 tp->lost_skb_hint = NULL;
3165 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3166 tp->snd_up = tp->snd_una;
3168 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3169 flag |= FLAG_SACK_RENEGING;
3171 skb_mstamp_get(&now);
3172 if (likely(first_ackt.v64)) {
3173 seq_rtt_us = skb_mstamp_us_delta(&now, &first_ackt);
3174 ca_rtt_us = skb_mstamp_us_delta(&now, &last_ackt);
3176 if (sack->first_sackt.v64) {
3177 sack_rtt_us = skb_mstamp_us_delta(&now, &sack->first_sackt);
3178 ca_rtt_us = skb_mstamp_us_delta(&now, &sack->last_sackt);
3181 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us);
3183 if (flag & FLAG_ACKED) {
3185 if (unlikely(icsk->icsk_mtup.probe_size &&
3186 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3187 tcp_mtup_probe_success(sk);
3190 if (tcp_is_reno(tp)) {
3191 tcp_remove_reno_sacks(sk, pkts_acked);
3195 /* Non-retransmitted hole got filled? That's reordering */
3196 if (reord < prior_fackets)
3197 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3199 delta = tcp_is_fack(tp) ? pkts_acked :
3200 prior_sacked - tp->sacked_out;
3201 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3204 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3206 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3207 sack_rtt_us > skb_mstamp_us_delta(&now, &skb->skb_mstamp)) {
3208 /* Do not re-arm RTO if the sack RTT is measured from data sent
3209 * after when the head was last (re)transmitted. Otherwise the
3210 * timeout may continue to extend in loss recovery.
3215 if (icsk->icsk_ca_ops->pkts_acked)
3216 icsk->icsk_ca_ops->pkts_acked(sk, pkts_acked, ca_rtt_us);
3218 #if FASTRETRANS_DEBUG > 0
3219 WARN_ON((int)tp->sacked_out < 0);
3220 WARN_ON((int)tp->lost_out < 0);
3221 WARN_ON((int)tp->retrans_out < 0);
3222 if (!tp->packets_out && tcp_is_sack(tp)) {
3223 icsk = inet_csk(sk);
3225 pr_debug("Leak l=%u %d\n",
3226 tp->lost_out, icsk->icsk_ca_state);
3229 if (tp->sacked_out) {
3230 pr_debug("Leak s=%u %d\n",
3231 tp->sacked_out, icsk->icsk_ca_state);
3234 if (tp->retrans_out) {
3235 pr_debug("Leak r=%u %d\n",
3236 tp->retrans_out, icsk->icsk_ca_state);
3237 tp->retrans_out = 0;
3244 static void tcp_ack_probe(struct sock *sk)
3246 const struct tcp_sock *tp = tcp_sk(sk);
3247 struct inet_connection_sock *icsk = inet_csk(sk);
3249 /* Was it a usable window open? */
3251 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3252 icsk->icsk_backoff = 0;
3253 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3254 /* Socket must be waked up by subsequent tcp_data_snd_check().
3255 * This function is not for random using!
3258 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
3260 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3265 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3267 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3268 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3271 /* Decide wheather to run the increase function of congestion control. */
3272 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3274 if (tcp_in_cwnd_reduction(sk))
3277 /* If reordering is high then always grow cwnd whenever data is
3278 * delivered regardless of its ordering. Otherwise stay conservative
3279 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3280 * new SACK or ECE mark may first advance cwnd here and later reduce
3281 * cwnd in tcp_fastretrans_alert() based on more states.
3283 if (tcp_sk(sk)->reordering > sysctl_tcp_reordering)
3284 return flag & FLAG_FORWARD_PROGRESS;
3286 return flag & FLAG_DATA_ACKED;
3289 /* Check that window update is acceptable.
3290 * The function assumes that snd_una<=ack<=snd_next.
3292 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3293 const u32 ack, const u32 ack_seq,
3296 return after(ack, tp->snd_una) ||
3297 after(ack_seq, tp->snd_wl1) ||
3298 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3301 /* If we update tp->snd_una, also update tp->bytes_acked */
3302 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3304 u32 delta = ack - tp->snd_una;
3306 u64_stats_update_begin(&tp->syncp);
3307 tp->bytes_acked += delta;
3308 u64_stats_update_end(&tp->syncp);
3312 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3313 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3315 u32 delta = seq - tp->rcv_nxt;
3317 u64_stats_update_begin(&tp->syncp);
3318 tp->bytes_received += delta;
3319 u64_stats_update_end(&tp->syncp);
3323 /* Update our send window.
3325 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3326 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3328 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3331 struct tcp_sock *tp = tcp_sk(sk);
3333 u32 nwin = ntohs(tcp_hdr(skb)->window);
3335 if (likely(!tcp_hdr(skb)->syn))
3336 nwin <<= tp->rx_opt.snd_wscale;
3338 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3339 flag |= FLAG_WIN_UPDATE;
3340 tcp_update_wl(tp, ack_seq);
3342 if (tp->snd_wnd != nwin) {
3345 /* Note, it is the only place, where
3346 * fast path is recovered for sending TCP.
3349 tcp_fast_path_check(sk);
3351 if (tcp_send_head(sk))
3352 tcp_slow_start_after_idle_check(sk);
3354 if (nwin > tp->max_window) {
3355 tp->max_window = nwin;
3356 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3361 tcp_snd_una_update(tp, ack);
3366 /* Return true if we're currently rate-limiting out-of-window ACKs and
3367 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3368 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3369 * attacks that send repeated SYNs or ACKs for the same connection. To
3370 * do this, we do not send a duplicate SYNACK or ACK if the remote
3371 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3373 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3374 int mib_idx, u32 *last_oow_ack_time)
3376 /* Data packets without SYNs are not likely part of an ACK loop. */
3377 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3379 goto not_rate_limited;
3381 if (*last_oow_ack_time) {
3382 s32 elapsed = (s32)(tcp_time_stamp - *last_oow_ack_time);
3384 if (0 <= elapsed && elapsed < sysctl_tcp_invalid_ratelimit) {
3385 NET_INC_STATS_BH(net, mib_idx);
3386 return true; /* rate-limited: don't send yet! */
3390 *last_oow_ack_time = tcp_time_stamp;
3393 return false; /* not rate-limited: go ahead, send dupack now! */
3396 /* RFC 5961 7 [ACK Throttling] */
3397 static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb)
3399 /* unprotected vars, we dont care of overwrites */
3400 static u32 challenge_timestamp;
3401 static unsigned int challenge_count;
3402 struct tcp_sock *tp = tcp_sk(sk);
3405 /* First check our per-socket dupack rate limit. */
3406 if (tcp_oow_rate_limited(sock_net(sk), skb,
3407 LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3408 &tp->last_oow_ack_time))
3411 /* Then check the check host-wide RFC 5961 rate limit. */
3413 if (now != challenge_timestamp) {
3414 challenge_timestamp = now;
3415 challenge_count = 0;
3417 if (++challenge_count <= sysctl_tcp_challenge_ack_limit) {
3418 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPCHALLENGEACK);
3423 static void tcp_store_ts_recent(struct tcp_sock *tp)
3425 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3426 tp->rx_opt.ts_recent_stamp = get_seconds();
3429 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3431 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3432 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3433 * extra check below makes sure this can only happen
3434 * for pure ACK frames. -DaveM
3436 * Not only, also it occurs for expired timestamps.
3439 if (tcp_paws_check(&tp->rx_opt, 0))
3440 tcp_store_ts_recent(tp);
3444 /* This routine deals with acks during a TLP episode.
3445 * We mark the end of a TLP episode on receiving TLP dupack or when
3446 * ack is after tlp_high_seq.
3447 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3449 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3451 struct tcp_sock *tp = tcp_sk(sk);
3453 if (before(ack, tp->tlp_high_seq))
3456 if (flag & FLAG_DSACKING_ACK) {
3457 /* This DSACK means original and TLP probe arrived; no loss */
3458 tp->tlp_high_seq = 0;
3459 } else if (after(ack, tp->tlp_high_seq)) {
3460 /* ACK advances: there was a loss, so reduce cwnd. Reset
3461 * tlp_high_seq in tcp_init_cwnd_reduction()
3463 tcp_init_cwnd_reduction(sk);
3464 tcp_set_ca_state(sk, TCP_CA_CWR);
3465 tcp_end_cwnd_reduction(sk);
3466 tcp_try_keep_open(sk);
3467 NET_INC_STATS_BH(sock_net(sk),
3468 LINUX_MIB_TCPLOSSPROBERECOVERY);
3469 } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3470 FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3471 /* Pure dupack: original and TLP probe arrived; no loss */
3472 tp->tlp_high_seq = 0;
3476 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3478 const struct inet_connection_sock *icsk = inet_csk(sk);
3480 if (icsk->icsk_ca_ops->in_ack_event)
3481 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3484 /* This routine deals with incoming acks, but not outgoing ones. */
3485 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3487 struct inet_connection_sock *icsk = inet_csk(sk);
3488 struct tcp_sock *tp = tcp_sk(sk);
3489 struct tcp_sacktag_state sack_state;
3490 u32 prior_snd_una = tp->snd_una;
3491 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3492 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3493 bool is_dupack = false;
3495 int prior_packets = tp->packets_out;
3496 const int prior_unsacked = tp->packets_out - tp->sacked_out;
3497 int acked = 0; /* Number of packets newly acked */
3499 sack_state.first_sackt.v64 = 0;
3501 /* We very likely will need to access write queue head. */
3502 prefetchw(sk->sk_write_queue.next);
3504 /* If the ack is older than previous acks
3505 * then we can probably ignore it.
3507 if (before(ack, prior_snd_una)) {
3508 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3509 if (before(ack, prior_snd_una - tp->max_window)) {
3510 tcp_send_challenge_ack(sk, skb);
3516 /* If the ack includes data we haven't sent yet, discard
3517 * this segment (RFC793 Section 3.9).
3519 if (after(ack, tp->snd_nxt))
3522 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
3523 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
3526 if (after(ack, prior_snd_una)) {
3527 flag |= FLAG_SND_UNA_ADVANCED;
3528 icsk->icsk_retransmits = 0;
3531 prior_fackets = tp->fackets_out;
3533 /* ts_recent update must be made after we are sure that the packet
3536 if (flag & FLAG_UPDATE_TS_RECENT)
3537 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3539 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3540 /* Window is constant, pure forward advance.
3541 * No more checks are required.
3542 * Note, we use the fact that SND.UNA>=SND.WL2.
3544 tcp_update_wl(tp, ack_seq);
3545 tcp_snd_una_update(tp, ack);
3546 flag |= FLAG_WIN_UPDATE;
3548 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3550 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3552 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3554 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3557 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3559 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3561 if (TCP_SKB_CB(skb)->sacked)
3562 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3565 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3567 ack_ev_flags |= CA_ACK_ECE;
3570 if (flag & FLAG_WIN_UPDATE)
3571 ack_ev_flags |= CA_ACK_WIN_UPDATE;
3573 tcp_in_ack_event(sk, ack_ev_flags);
3576 /* We passed data and got it acked, remove any soft error
3577 * log. Something worked...
3579 sk->sk_err_soft = 0;
3580 icsk->icsk_probes_out = 0;
3581 tp->rcv_tstamp = tcp_time_stamp;
3585 /* See if we can take anything off of the retransmit queue. */
3586 acked = tp->packets_out;
3587 flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una,
3589 acked -= tp->packets_out;
3591 if (tcp_ack_is_dubious(sk, flag)) {
3592 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3593 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3596 if (tp->tlp_high_seq)
3597 tcp_process_tlp_ack(sk, ack, flag);
3599 /* Advance cwnd if state allows */
3600 if (tcp_may_raise_cwnd(sk, flag))
3601 tcp_cong_avoid(sk, ack, acked);
3603 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) {
3604 struct dst_entry *dst = __sk_dst_get(sk);
3609 if (icsk->icsk_pending == ICSK_TIME_RETRANS)
3610 tcp_schedule_loss_probe(sk);
3611 tcp_update_pacing_rate(sk);
3615 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3616 if (flag & FLAG_DSACKING_ACK)
3617 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3619 /* If this ack opens up a zero window, clear backoff. It was
3620 * being used to time the probes, and is probably far higher than
3621 * it needs to be for normal retransmission.
3623 if (tcp_send_head(sk))
3626 if (tp->tlp_high_seq)
3627 tcp_process_tlp_ack(sk, ack, flag);
3631 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3635 /* If data was SACKed, tag it and see if we should send more data.
3636 * If data was DSACKed, see if we can undo a cwnd reduction.
3638 if (TCP_SKB_CB(skb)->sacked) {
3639 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3641 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3645 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3649 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3650 bool syn, struct tcp_fastopen_cookie *foc,
3653 /* Valid only in SYN or SYN-ACK with an even length. */
3654 if (!foc || !syn || len < 0 || (len & 1))
3657 if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3658 len <= TCP_FASTOPEN_COOKIE_MAX)
3659 memcpy(foc->val, cookie, len);
3666 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3667 * But, this can also be called on packets in the established flow when
3668 * the fast version below fails.
3670 void tcp_parse_options(const struct sk_buff *skb,
3671 struct tcp_options_received *opt_rx, int estab,
3672 struct tcp_fastopen_cookie *foc)
3674 const unsigned char *ptr;
3675 const struct tcphdr *th = tcp_hdr(skb);
3676 int length = (th->doff * 4) - sizeof(struct tcphdr);
3678 ptr = (const unsigned char *)(th + 1);
3679 opt_rx->saw_tstamp = 0;
3681 while (length > 0) {
3682 int opcode = *ptr++;
3688 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3693 if (opsize < 2) /* "silly options" */
3695 if (opsize > length)
3696 return; /* don't parse partial options */
3699 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3700 u16 in_mss = get_unaligned_be16(ptr);
3702 if (opt_rx->user_mss &&
3703 opt_rx->user_mss < in_mss)
3704 in_mss = opt_rx->user_mss;
3705 opt_rx->mss_clamp = in_mss;
3710 if (opsize == TCPOLEN_WINDOW && th->syn &&
3711 !estab && sysctl_tcp_window_scaling) {
3712 __u8 snd_wscale = *(__u8 *)ptr;
3713 opt_rx->wscale_ok = 1;
3714 if (snd_wscale > 14) {
3715 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3720 opt_rx->snd_wscale = snd_wscale;
3723 case TCPOPT_TIMESTAMP:
3724 if ((opsize == TCPOLEN_TIMESTAMP) &&
3725 ((estab && opt_rx->tstamp_ok) ||
3726 (!estab && sysctl_tcp_timestamps))) {
3727 opt_rx->saw_tstamp = 1;
3728 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3729 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3732 case TCPOPT_SACK_PERM:
3733 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3734 !estab && sysctl_tcp_sack) {
3735 opt_rx->sack_ok = TCP_SACK_SEEN;
3736 tcp_sack_reset(opt_rx);
3741 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3742 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3744 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3747 #ifdef CONFIG_TCP_MD5SIG
3750 * The MD5 Hash has already been
3751 * checked (see tcp_v{4,6}_do_rcv()).
3755 case TCPOPT_FASTOPEN:
3756 tcp_parse_fastopen_option(
3757 opsize - TCPOLEN_FASTOPEN_BASE,
3758 ptr, th->syn, foc, false);
3762 /* Fast Open option shares code 254 using a
3763 * 16 bits magic number.
3765 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
3766 get_unaligned_be16(ptr) ==
3767 TCPOPT_FASTOPEN_MAGIC)
3768 tcp_parse_fastopen_option(opsize -
3769 TCPOLEN_EXP_FASTOPEN_BASE,
3770 ptr + 2, th->syn, foc, true);
3779 EXPORT_SYMBOL(tcp_parse_options);
3781 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3783 const __be32 *ptr = (const __be32 *)(th + 1);
3785 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3786 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3787 tp->rx_opt.saw_tstamp = 1;
3789 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3792 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3794 tp->rx_opt.rcv_tsecr = 0;
3800 /* Fast parse options. This hopes to only see timestamps.
3801 * If it is wrong it falls back on tcp_parse_options().
3803 static bool tcp_fast_parse_options(const struct sk_buff *skb,
3804 const struct tcphdr *th, struct tcp_sock *tp)
3806 /* In the spirit of fast parsing, compare doff directly to constant
3807 * values. Because equality is used, short doff can be ignored here.
3809 if (th->doff == (sizeof(*th) / 4)) {
3810 tp->rx_opt.saw_tstamp = 0;
3812 } else if (tp->rx_opt.tstamp_ok &&
3813 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3814 if (tcp_parse_aligned_timestamp(tp, th))
3818 tcp_parse_options(skb, &tp->rx_opt, 1, NULL);
3819 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3820 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3825 #ifdef CONFIG_TCP_MD5SIG
3827 * Parse MD5 Signature option
3829 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3831 int length = (th->doff << 2) - sizeof(*th);
3832 const u8 *ptr = (const u8 *)(th + 1);
3834 /* If the TCP option is too short, we can short cut */
3835 if (length < TCPOLEN_MD5SIG)
3838 while (length > 0) {
3839 int opcode = *ptr++;
3850 if (opsize < 2 || opsize > length)
3852 if (opcode == TCPOPT_MD5SIG)
3853 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3860 EXPORT_SYMBOL(tcp_parse_md5sig_option);
3863 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3865 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3866 * it can pass through stack. So, the following predicate verifies that
3867 * this segment is not used for anything but congestion avoidance or
3868 * fast retransmit. Moreover, we even are able to eliminate most of such
3869 * second order effects, if we apply some small "replay" window (~RTO)
3870 * to timestamp space.
3872 * All these measures still do not guarantee that we reject wrapped ACKs
3873 * on networks with high bandwidth, when sequence space is recycled fastly,
3874 * but it guarantees that such events will be very rare and do not affect
3875 * connection seriously. This doesn't look nice, but alas, PAWS is really
3878 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3879 * states that events when retransmit arrives after original data are rare.
3880 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3881 * the biggest problem on large power networks even with minor reordering.
3882 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3883 * up to bandwidth of 18Gigabit/sec. 8) ]
3886 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3888 const struct tcp_sock *tp = tcp_sk(sk);
3889 const struct tcphdr *th = tcp_hdr(skb);
3890 u32 seq = TCP_SKB_CB(skb)->seq;
3891 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3893 return (/* 1. Pure ACK with correct sequence number. */
3894 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3896 /* 2. ... and duplicate ACK. */
3897 ack == tp->snd_una &&
3899 /* 3. ... and does not update window. */
3900 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3902 /* 4. ... and sits in replay window. */
3903 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3906 static inline bool tcp_paws_discard(const struct sock *sk,
3907 const struct sk_buff *skb)
3909 const struct tcp_sock *tp = tcp_sk(sk);
3911 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
3912 !tcp_disordered_ack(sk, skb);
3915 /* Check segment sequence number for validity.
3917 * Segment controls are considered valid, if the segment
3918 * fits to the window after truncation to the window. Acceptability
3919 * of data (and SYN, FIN, of course) is checked separately.
3920 * See tcp_data_queue(), for example.
3922 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3923 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3924 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3925 * (borrowed from freebsd)
3928 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
3930 return !before(end_seq, tp->rcv_wup) &&
3931 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3934 /* When we get a reset we do this. */
3935 void tcp_reset(struct sock *sk)
3937 /* We want the right error as BSD sees it (and indeed as we do). */
3938 switch (sk->sk_state) {
3940 sk->sk_err = ECONNREFUSED;
3942 case TCP_CLOSE_WAIT:
3948 sk->sk_err = ECONNRESET;
3950 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3953 if (!sock_flag(sk, SOCK_DEAD))
3954 sk->sk_error_report(sk);
3960 * Process the FIN bit. This now behaves as it is supposed to work
3961 * and the FIN takes effect when it is validly part of sequence
3962 * space. Not before when we get holes.
3964 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3965 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3968 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3969 * close and we go into CLOSING (and later onto TIME-WAIT)
3971 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3973 static void tcp_fin(struct sock *sk)
3975 struct tcp_sock *tp = tcp_sk(sk);
3977 inet_csk_schedule_ack(sk);
3979 sk->sk_shutdown |= RCV_SHUTDOWN;
3980 sock_set_flag(sk, SOCK_DONE);
3982 switch (sk->sk_state) {
3984 case TCP_ESTABLISHED:
3985 /* Move to CLOSE_WAIT */
3986 tcp_set_state(sk, TCP_CLOSE_WAIT);
3987 inet_csk(sk)->icsk_ack.pingpong = 1;
3990 case TCP_CLOSE_WAIT:
3992 /* Received a retransmission of the FIN, do
3997 /* RFC793: Remain in the LAST-ACK state. */
4001 /* This case occurs when a simultaneous close
4002 * happens, we must ack the received FIN and
4003 * enter the CLOSING state.
4006 tcp_set_state(sk, TCP_CLOSING);
4009 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4011 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4014 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4015 * cases we should never reach this piece of code.
4017 pr_err("%s: Impossible, sk->sk_state=%d\n",
4018 __func__, sk->sk_state);
4022 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4023 * Probably, we should reset in this case. For now drop them.
4025 __skb_queue_purge(&tp->out_of_order_queue);
4026 if (tcp_is_sack(tp))
4027 tcp_sack_reset(&tp->rx_opt);
4030 if (!sock_flag(sk, SOCK_DEAD)) {
4031 sk->sk_state_change(sk);
4033 /* Do not send POLL_HUP for half duplex close. */
4034 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4035 sk->sk_state == TCP_CLOSE)
4036 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4038 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4042 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4045 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4046 if (before(seq, sp->start_seq))
4047 sp->start_seq = seq;
4048 if (after(end_seq, sp->end_seq))
4049 sp->end_seq = end_seq;
4055 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4057 struct tcp_sock *tp = tcp_sk(sk);
4059 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4062 if (before(seq, tp->rcv_nxt))
4063 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4065 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4067 NET_INC_STATS_BH(sock_net(sk), mib_idx);
4069 tp->rx_opt.dsack = 1;
4070 tp->duplicate_sack[0].start_seq = seq;
4071 tp->duplicate_sack[0].end_seq = end_seq;
4075 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4077 struct tcp_sock *tp = tcp_sk(sk);
4079 if (!tp->rx_opt.dsack)
4080 tcp_dsack_set(sk, seq, end_seq);
4082 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4085 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4087 struct tcp_sock *tp = tcp_sk(sk);
4089 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4090 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4091 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4092 tcp_enter_quickack_mode(sk);
4094 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4095 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4097 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4098 end_seq = tp->rcv_nxt;
4099 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4106 /* These routines update the SACK block as out-of-order packets arrive or
4107 * in-order packets close up the sequence space.
4109 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4112 struct tcp_sack_block *sp = &tp->selective_acks[0];
4113 struct tcp_sack_block *swalk = sp + 1;
4115 /* See if the recent change to the first SACK eats into
4116 * or hits the sequence space of other SACK blocks, if so coalesce.
4118 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4119 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4122 /* Zap SWALK, by moving every further SACK up by one slot.
4123 * Decrease num_sacks.
4125 tp->rx_opt.num_sacks--;
4126 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4130 this_sack++, swalk++;
4134 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4136 struct tcp_sock *tp = tcp_sk(sk);
4137 struct tcp_sack_block *sp = &tp->selective_acks[0];
4138 int cur_sacks = tp->rx_opt.num_sacks;
4144 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4145 if (tcp_sack_extend(sp, seq, end_seq)) {
4146 /* Rotate this_sack to the first one. */
4147 for (; this_sack > 0; this_sack--, sp--)
4148 swap(*sp, *(sp - 1));
4150 tcp_sack_maybe_coalesce(tp);
4155 /* Could not find an adjacent existing SACK, build a new one,
4156 * put it at the front, and shift everyone else down. We
4157 * always know there is at least one SACK present already here.
4159 * If the sack array is full, forget about the last one.
4161 if (this_sack >= TCP_NUM_SACKS) {
4163 tp->rx_opt.num_sacks--;
4166 for (; this_sack > 0; this_sack--, sp--)
4170 /* Build the new head SACK, and we're done. */
4171 sp->start_seq = seq;
4172 sp->end_seq = end_seq;
4173 tp->rx_opt.num_sacks++;
4176 /* RCV.NXT advances, some SACKs should be eaten. */
4178 static void tcp_sack_remove(struct tcp_sock *tp)
4180 struct tcp_sack_block *sp = &tp->selective_acks[0];
4181 int num_sacks = tp->rx_opt.num_sacks;
4184 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4185 if (skb_queue_empty(&tp->out_of_order_queue)) {
4186 tp->rx_opt.num_sacks = 0;
4190 for (this_sack = 0; this_sack < num_sacks;) {
4191 /* Check if the start of the sack is covered by RCV.NXT. */
4192 if (!before(tp->rcv_nxt, sp->start_seq)) {
4195 /* RCV.NXT must cover all the block! */
4196 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4198 /* Zap this SACK, by moving forward any other SACKS. */
4199 for (i = this_sack+1; i < num_sacks; i++)
4200 tp->selective_acks[i-1] = tp->selective_acks[i];
4207 tp->rx_opt.num_sacks = num_sacks;
4211 * tcp_try_coalesce - try to merge skb to prior one
4214 * @from: buffer to add in queue
4215 * @fragstolen: pointer to boolean
4217 * Before queueing skb @from after @to, try to merge them
4218 * to reduce overall memory use and queue lengths, if cost is small.
4219 * Packets in ofo or receive queues can stay a long time.
4220 * Better try to coalesce them right now to avoid future collapses.
4221 * Returns true if caller should free @from instead of queueing it
4223 static bool tcp_try_coalesce(struct sock *sk,
4225 struct sk_buff *from,
4230 *fragstolen = false;
4232 /* Its possible this segment overlaps with prior segment in queue */
4233 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4236 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4239 atomic_add(delta, &sk->sk_rmem_alloc);
4240 sk_mem_charge(sk, delta);
4241 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4242 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4243 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4244 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4248 /* This one checks to see if we can put data from the
4249 * out_of_order queue into the receive_queue.
4251 static void tcp_ofo_queue(struct sock *sk)
4253 struct tcp_sock *tp = tcp_sk(sk);
4254 __u32 dsack_high = tp->rcv_nxt;
4255 struct sk_buff *skb, *tail;
4256 bool fragstolen, eaten;
4258 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4259 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4262 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4263 __u32 dsack = dsack_high;
4264 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4265 dsack_high = TCP_SKB_CB(skb)->end_seq;
4266 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4269 __skb_unlink(skb, &tp->out_of_order_queue);
4270 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4271 SOCK_DEBUG(sk, "ofo packet was already received\n");
4275 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4276 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4277 TCP_SKB_CB(skb)->end_seq);
4279 tail = skb_peek_tail(&sk->sk_receive_queue);
4280 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4281 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4283 __skb_queue_tail(&sk->sk_receive_queue, skb);
4284 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4287 kfree_skb_partial(skb, fragstolen);
4291 static bool tcp_prune_ofo_queue(struct sock *sk);
4292 static int tcp_prune_queue(struct sock *sk);
4294 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4297 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4298 !sk_rmem_schedule(sk, skb, size)) {
4300 if (tcp_prune_queue(sk) < 0)
4303 if (!sk_rmem_schedule(sk, skb, size)) {
4304 if (!tcp_prune_ofo_queue(sk))
4307 if (!sk_rmem_schedule(sk, skb, size))
4314 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4316 struct tcp_sock *tp = tcp_sk(sk);
4317 struct sk_buff *skb1;
4320 tcp_ecn_check_ce(tp, skb);
4322 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4323 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFODROP);
4328 /* Disable header prediction. */
4330 inet_csk_schedule_ack(sk);
4332 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4333 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4334 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4336 skb1 = skb_peek_tail(&tp->out_of_order_queue);
4338 /* Initial out of order segment, build 1 SACK. */
4339 if (tcp_is_sack(tp)) {
4340 tp->rx_opt.num_sacks = 1;
4341 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4342 tp->selective_acks[0].end_seq =
4343 TCP_SKB_CB(skb)->end_seq;
4345 __skb_queue_head(&tp->out_of_order_queue, skb);
4349 seq = TCP_SKB_CB(skb)->seq;
4350 end_seq = TCP_SKB_CB(skb)->end_seq;
4352 if (seq == TCP_SKB_CB(skb1)->end_seq) {
4355 if (!tcp_try_coalesce(sk, skb1, skb, &fragstolen)) {
4356 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4358 tcp_grow_window(sk, skb);
4359 kfree_skb_partial(skb, fragstolen);
4363 if (!tp->rx_opt.num_sacks ||
4364 tp->selective_acks[0].end_seq != seq)
4367 /* Common case: data arrive in order after hole. */
4368 tp->selective_acks[0].end_seq = end_seq;
4372 /* Find place to insert this segment. */
4374 if (!after(TCP_SKB_CB(skb1)->seq, seq))
4376 if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
4380 skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
4383 /* Do skb overlap to previous one? */
4384 if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4385 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4386 /* All the bits are present. Drop. */
4387 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4390 tcp_dsack_set(sk, seq, end_seq);
4393 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4394 /* Partial overlap. */
4395 tcp_dsack_set(sk, seq,
4396 TCP_SKB_CB(skb1)->end_seq);
4398 if (skb_queue_is_first(&tp->out_of_order_queue,
4402 skb1 = skb_queue_prev(
4403 &tp->out_of_order_queue,
4408 __skb_queue_head(&tp->out_of_order_queue, skb);
4410 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4412 /* And clean segments covered by new one as whole. */
4413 while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
4414 skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
4416 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4418 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4419 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4423 __skb_unlink(skb1, &tp->out_of_order_queue);
4424 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4425 TCP_SKB_CB(skb1)->end_seq);
4426 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4431 if (tcp_is_sack(tp))
4432 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4435 tcp_grow_window(sk, skb);
4436 skb_set_owner_r(skb, sk);
4440 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
4444 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4446 __skb_pull(skb, hdrlen);
4448 tcp_try_coalesce(sk, tail, skb, fragstolen)) ? 1 : 0;
4449 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
4451 __skb_queue_tail(&sk->sk_receive_queue, skb);
4452 skb_set_owner_r(skb, sk);
4457 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4459 struct sk_buff *skb;
4465 skb = alloc_skb(size, sk->sk_allocation);
4469 if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4472 if (memcpy_from_msg(skb_put(skb, size), msg, size))
4475 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4476 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4477 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4479 if (tcp_queue_rcv(sk, skb, 0, &fragstolen)) {
4480 WARN_ON_ONCE(fragstolen); /* should not happen */
4491 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4493 struct tcp_sock *tp = tcp_sk(sk);
4495 bool fragstolen = false;
4497 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4501 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
4503 tcp_ecn_accept_cwr(tp, skb);
4505 tp->rx_opt.dsack = 0;
4507 /* Queue data for delivery to the user.
4508 * Packets in sequence go to the receive queue.
4509 * Out of sequence packets to the out_of_order_queue.
4511 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4512 if (tcp_receive_window(tp) == 0)
4515 /* Ok. In sequence. In window. */
4516 if (tp->ucopy.task == current &&
4517 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4518 sock_owned_by_user(sk) && !tp->urg_data) {
4519 int chunk = min_t(unsigned int, skb->len,
4522 __set_current_state(TASK_RUNNING);
4525 if (!skb_copy_datagram_msg(skb, 0, tp->ucopy.msg, chunk)) {
4526 tp->ucopy.len -= chunk;
4527 tp->copied_seq += chunk;
4528 eaten = (chunk == skb->len);
4529 tcp_rcv_space_adjust(sk);
4537 if (skb_queue_len(&sk->sk_receive_queue) == 0)
4538 sk_forced_mem_schedule(sk, skb->truesize);
4539 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4542 eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
4544 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4546 tcp_event_data_recv(sk, skb);
4547 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4550 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4553 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4554 * gap in queue is filled.
4556 if (skb_queue_empty(&tp->out_of_order_queue))
4557 inet_csk(sk)->icsk_ack.pingpong = 0;
4560 if (tp->rx_opt.num_sacks)
4561 tcp_sack_remove(tp);
4563 tcp_fast_path_check(sk);
4566 kfree_skb_partial(skb, fragstolen);
4567 if (!sock_flag(sk, SOCK_DEAD))
4568 sk->sk_data_ready(sk);
4572 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4573 /* A retransmit, 2nd most common case. Force an immediate ack. */
4574 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4575 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4578 tcp_enter_quickack_mode(sk);
4579 inet_csk_schedule_ack(sk);
4585 /* Out of window. F.e. zero window probe. */
4586 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4589 tcp_enter_quickack_mode(sk);
4591 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4592 /* Partial packet, seq < rcv_next < end_seq */
4593 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4594 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4595 TCP_SKB_CB(skb)->end_seq);
4597 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4599 /* If window is closed, drop tail of packet. But after
4600 * remembering D-SACK for its head made in previous line.
4602 if (!tcp_receive_window(tp))
4607 tcp_data_queue_ofo(sk, skb);
4610 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4611 struct sk_buff_head *list)
4613 struct sk_buff *next = NULL;
4615 if (!skb_queue_is_last(list, skb))
4616 next = skb_queue_next(list, skb);
4618 __skb_unlink(skb, list);
4620 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4625 /* Collapse contiguous sequence of skbs head..tail with
4626 * sequence numbers start..end.
4628 * If tail is NULL, this means until the end of the list.
4630 * Segments with FIN/SYN are not collapsed (only because this
4634 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4635 struct sk_buff *head, struct sk_buff *tail,
4638 struct sk_buff *skb, *n;
4641 /* First, check that queue is collapsible and find
4642 * the point where collapsing can be useful. */
4646 skb_queue_walk_from_safe(list, skb, n) {
4649 /* No new bits? It is possible on ofo queue. */
4650 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4651 skb = tcp_collapse_one(sk, skb, list);
4657 /* The first skb to collapse is:
4659 * - bloated or contains data before "start" or
4660 * overlaps to the next one.
4662 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
4663 (tcp_win_from_space(skb->truesize) > skb->len ||
4664 before(TCP_SKB_CB(skb)->seq, start))) {
4665 end_of_skbs = false;
4669 if (!skb_queue_is_last(list, skb)) {
4670 struct sk_buff *next = skb_queue_next(list, skb);
4672 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
4673 end_of_skbs = false;
4678 /* Decided to skip this, advance start seq. */
4679 start = TCP_SKB_CB(skb)->end_seq;
4682 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4685 while (before(start, end)) {
4686 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
4687 struct sk_buff *nskb;
4689 nskb = alloc_skb(copy, GFP_ATOMIC);
4693 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4694 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4695 __skb_queue_before(list, skb, nskb);
4696 skb_set_owner_r(nskb, sk);
4698 /* Copy data, releasing collapsed skbs. */
4700 int offset = start - TCP_SKB_CB(skb)->seq;
4701 int size = TCP_SKB_CB(skb)->end_seq - start;
4705 size = min(copy, size);
4706 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4708 TCP_SKB_CB(nskb)->end_seq += size;
4712 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4713 skb = tcp_collapse_one(sk, skb, list);
4716 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4723 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4724 * and tcp_collapse() them until all the queue is collapsed.
4726 static void tcp_collapse_ofo_queue(struct sock *sk)
4728 struct tcp_sock *tp = tcp_sk(sk);
4729 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4730 struct sk_buff *head;
4736 start = TCP_SKB_CB(skb)->seq;
4737 end = TCP_SKB_CB(skb)->end_seq;
4741 struct sk_buff *next = NULL;
4743 if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
4744 next = skb_queue_next(&tp->out_of_order_queue, skb);
4747 /* Segment is terminated when we see gap or when
4748 * we are at the end of all the queue. */
4750 after(TCP_SKB_CB(skb)->seq, end) ||
4751 before(TCP_SKB_CB(skb)->end_seq, start)) {
4752 tcp_collapse(sk, &tp->out_of_order_queue,
4753 head, skb, start, end);
4757 /* Start new segment */
4758 start = TCP_SKB_CB(skb)->seq;
4759 end = TCP_SKB_CB(skb)->end_seq;
4761 if (before(TCP_SKB_CB(skb)->seq, start))
4762 start = TCP_SKB_CB(skb)->seq;
4763 if (after(TCP_SKB_CB(skb)->end_seq, end))
4764 end = TCP_SKB_CB(skb)->end_seq;
4770 * Purge the out-of-order queue.
4771 * Return true if queue was pruned.
4773 static bool tcp_prune_ofo_queue(struct sock *sk)
4775 struct tcp_sock *tp = tcp_sk(sk);
4778 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4779 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
4780 __skb_queue_purge(&tp->out_of_order_queue);
4782 /* Reset SACK state. A conforming SACK implementation will
4783 * do the same at a timeout based retransmit. When a connection
4784 * is in a sad state like this, we care only about integrity
4785 * of the connection not performance.
4787 if (tp->rx_opt.sack_ok)
4788 tcp_sack_reset(&tp->rx_opt);
4795 /* Reduce allocated memory if we can, trying to get
4796 * the socket within its memory limits again.
4798 * Return less than zero if we should start dropping frames
4799 * until the socket owning process reads some of the data
4800 * to stabilize the situation.
4802 static int tcp_prune_queue(struct sock *sk)
4804 struct tcp_sock *tp = tcp_sk(sk);
4806 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4808 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
4810 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4811 tcp_clamp_window(sk);
4812 else if (tcp_under_memory_pressure(sk))
4813 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4815 tcp_collapse_ofo_queue(sk);
4816 if (!skb_queue_empty(&sk->sk_receive_queue))
4817 tcp_collapse(sk, &sk->sk_receive_queue,
4818 skb_peek(&sk->sk_receive_queue),
4820 tp->copied_seq, tp->rcv_nxt);
4823 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4826 /* Collapsing did not help, destructive actions follow.
4827 * This must not ever occur. */
4829 tcp_prune_ofo_queue(sk);
4831 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4834 /* If we are really being abused, tell the caller to silently
4835 * drop receive data on the floor. It will get retransmitted
4836 * and hopefully then we'll have sufficient space.
4838 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
4840 /* Massive buffer overcommit. */
4845 static bool tcp_should_expand_sndbuf(const struct sock *sk)
4847 const struct tcp_sock *tp = tcp_sk(sk);
4849 /* If the user specified a specific send buffer setting, do
4852 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4855 /* If we are under global TCP memory pressure, do not expand. */
4856 if (tcp_under_memory_pressure(sk))
4859 /* If we are under soft global TCP memory pressure, do not expand. */
4860 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
4863 /* If we filled the congestion window, do not expand. */
4864 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
4870 /* When incoming ACK allowed to free some skb from write_queue,
4871 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4872 * on the exit from tcp input handler.
4874 * PROBLEM: sndbuf expansion does not work well with largesend.
4876 static void tcp_new_space(struct sock *sk)
4878 struct tcp_sock *tp = tcp_sk(sk);
4880 if (tcp_should_expand_sndbuf(sk)) {
4881 tcp_sndbuf_expand(sk);
4882 tp->snd_cwnd_stamp = tcp_time_stamp;
4885 sk->sk_write_space(sk);
4888 static void tcp_check_space(struct sock *sk)
4890 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4891 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4892 /* pairs with tcp_poll() */
4893 smp_mb__after_atomic();
4894 if (sk->sk_socket &&
4895 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4900 static inline void tcp_data_snd_check(struct sock *sk)
4902 tcp_push_pending_frames(sk);
4903 tcp_check_space(sk);
4907 * Check if sending an ack is needed.
4909 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4911 struct tcp_sock *tp = tcp_sk(sk);
4913 /* More than one full frame received... */
4914 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
4915 /* ... and right edge of window advances far enough.
4916 * (tcp_recvmsg() will send ACK otherwise). Or...
4918 __tcp_select_window(sk) >= tp->rcv_wnd) ||
4919 /* We ACK each frame or... */
4920 tcp_in_quickack_mode(sk) ||
4921 /* We have out of order data. */
4922 (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
4923 /* Then ack it now */
4926 /* Else, send delayed ack. */
4927 tcp_send_delayed_ack(sk);
4931 static inline void tcp_ack_snd_check(struct sock *sk)
4933 if (!inet_csk_ack_scheduled(sk)) {
4934 /* We sent a data segment already. */
4937 __tcp_ack_snd_check(sk, 1);
4941 * This routine is only called when we have urgent data
4942 * signaled. Its the 'slow' part of tcp_urg. It could be
4943 * moved inline now as tcp_urg is only called from one
4944 * place. We handle URGent data wrong. We have to - as
4945 * BSD still doesn't use the correction from RFC961.
4946 * For 1003.1g we should support a new option TCP_STDURG to permit
4947 * either form (or just set the sysctl tcp_stdurg).
4950 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
4952 struct tcp_sock *tp = tcp_sk(sk);
4953 u32 ptr = ntohs(th->urg_ptr);
4955 if (ptr && !sysctl_tcp_stdurg)
4957 ptr += ntohl(th->seq);
4959 /* Ignore urgent data that we've already seen and read. */
4960 if (after(tp->copied_seq, ptr))
4963 /* Do not replay urg ptr.
4965 * NOTE: interesting situation not covered by specs.
4966 * Misbehaving sender may send urg ptr, pointing to segment,
4967 * which we already have in ofo queue. We are not able to fetch
4968 * such data and will stay in TCP_URG_NOTYET until will be eaten
4969 * by recvmsg(). Seems, we are not obliged to handle such wicked
4970 * situations. But it is worth to think about possibility of some
4971 * DoSes using some hypothetical application level deadlock.
4973 if (before(ptr, tp->rcv_nxt))
4976 /* Do we already have a newer (or duplicate) urgent pointer? */
4977 if (tp->urg_data && !after(ptr, tp->urg_seq))
4980 /* Tell the world about our new urgent pointer. */
4983 /* We may be adding urgent data when the last byte read was
4984 * urgent. To do this requires some care. We cannot just ignore
4985 * tp->copied_seq since we would read the last urgent byte again
4986 * as data, nor can we alter copied_seq until this data arrives
4987 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4989 * NOTE. Double Dutch. Rendering to plain English: author of comment
4990 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4991 * and expect that both A and B disappear from stream. This is _wrong_.
4992 * Though this happens in BSD with high probability, this is occasional.
4993 * Any application relying on this is buggy. Note also, that fix "works"
4994 * only in this artificial test. Insert some normal data between A and B and we will
4995 * decline of BSD again. Verdict: it is better to remove to trap
4998 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
4999 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5000 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5002 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5003 __skb_unlink(skb, &sk->sk_receive_queue);
5008 tp->urg_data = TCP_URG_NOTYET;
5011 /* Disable header prediction. */
5015 /* This is the 'fast' part of urgent handling. */
5016 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5018 struct tcp_sock *tp = tcp_sk(sk);
5020 /* Check if we get a new urgent pointer - normally not. */
5022 tcp_check_urg(sk, th);
5024 /* Do we wait for any urgent data? - normally not... */
5025 if (tp->urg_data == TCP_URG_NOTYET) {
5026 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5029 /* Is the urgent pointer pointing into this packet? */
5030 if (ptr < skb->len) {
5032 if (skb_copy_bits(skb, ptr, &tmp, 1))
5034 tp->urg_data = TCP_URG_VALID | tmp;
5035 if (!sock_flag(sk, SOCK_DEAD))
5036 sk->sk_data_ready(sk);
5041 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
5043 struct tcp_sock *tp = tcp_sk(sk);
5044 int chunk = skb->len - hlen;
5048 if (skb_csum_unnecessary(skb))
5049 err = skb_copy_datagram_msg(skb, hlen, tp->ucopy.msg, chunk);
5051 err = skb_copy_and_csum_datagram_msg(skb, hlen, tp->ucopy.msg);
5054 tp->ucopy.len -= chunk;
5055 tp->copied_seq += chunk;
5056 tcp_rcv_space_adjust(sk);
5063 static __sum16 __tcp_checksum_complete_user(struct sock *sk,
5064 struct sk_buff *skb)
5068 if (sock_owned_by_user(sk)) {
5070 result = __tcp_checksum_complete(skb);
5073 result = __tcp_checksum_complete(skb);
5078 static inline bool tcp_checksum_complete_user(struct sock *sk,
5079 struct sk_buff *skb)
5081 return !skb_csum_unnecessary(skb) &&
5082 __tcp_checksum_complete_user(sk, skb);
5085 /* Does PAWS and seqno based validation of an incoming segment, flags will
5086 * play significant role here.
5088 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5089 const struct tcphdr *th, int syn_inerr)
5091 struct tcp_sock *tp = tcp_sk(sk);
5093 /* RFC1323: H1. Apply PAWS check first. */
5094 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
5095 tcp_paws_discard(sk, skb)) {
5097 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5098 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5099 LINUX_MIB_TCPACKSKIPPEDPAWS,
5100 &tp->last_oow_ack_time))
5101 tcp_send_dupack(sk, skb);
5104 /* Reset is accepted even if it did not pass PAWS. */
5107 /* Step 1: check sequence number */
5108 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5109 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5110 * (RST) segments are validated by checking their SEQ-fields."
5111 * And page 69: "If an incoming segment is not acceptable,
5112 * an acknowledgment should be sent in reply (unless the RST
5113 * bit is set, if so drop the segment and return)".
5118 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5119 LINUX_MIB_TCPACKSKIPPEDSEQ,
5120 &tp->last_oow_ack_time))
5121 tcp_send_dupack(sk, skb);
5126 /* Step 2: check RST bit */
5129 * If sequence number exactly matches RCV.NXT, then
5130 * RESET the connection
5132 * Send a challenge ACK
5134 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt)
5137 tcp_send_challenge_ack(sk, skb);
5141 /* step 3: check security and precedence [ignored] */
5143 /* step 4: Check for a SYN
5144 * RFC 5961 4.2 : Send a challenge ack
5149 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5150 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5151 tcp_send_challenge_ack(sk, skb);
5163 * TCP receive function for the ESTABLISHED state.
5165 * It is split into a fast path and a slow path. The fast path is
5167 * - A zero window was announced from us - zero window probing
5168 * is only handled properly in the slow path.
5169 * - Out of order segments arrived.
5170 * - Urgent data is expected.
5171 * - There is no buffer space left
5172 * - Unexpected TCP flags/window values/header lengths are received
5173 * (detected by checking the TCP header against pred_flags)
5174 * - Data is sent in both directions. Fast path only supports pure senders
5175 * or pure receivers (this means either the sequence number or the ack
5176 * value must stay constant)
5177 * - Unexpected TCP option.
5179 * When these conditions are not satisfied it drops into a standard
5180 * receive procedure patterned after RFC793 to handle all cases.
5181 * The first three cases are guaranteed by proper pred_flags setting,
5182 * the rest is checked inline. Fast processing is turned on in
5183 * tcp_data_queue when everything is OK.
5185 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5186 const struct tcphdr *th, unsigned int len)
5188 struct tcp_sock *tp = tcp_sk(sk);
5190 if (unlikely(!sk->sk_rx_dst))
5191 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5193 * Header prediction.
5194 * The code loosely follows the one in the famous
5195 * "30 instruction TCP receive" Van Jacobson mail.
5197 * Van's trick is to deposit buffers into socket queue
5198 * on a device interrupt, to call tcp_recv function
5199 * on the receive process context and checksum and copy
5200 * the buffer to user space. smart...
5202 * Our current scheme is not silly either but we take the
5203 * extra cost of the net_bh soft interrupt processing...
5204 * We do checksum and copy also but from device to kernel.
5207 tp->rx_opt.saw_tstamp = 0;
5209 /* pred_flags is 0xS?10 << 16 + snd_wnd
5210 * if header_prediction is to be made
5211 * 'S' will always be tp->tcp_header_len >> 2
5212 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5213 * turn it off (when there are holes in the receive
5214 * space for instance)
5215 * PSH flag is ignored.
5218 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5219 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5220 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5221 int tcp_header_len = tp->tcp_header_len;
5223 /* Timestamp header prediction: tcp_header_len
5224 * is automatically equal to th->doff*4 due to pred_flags
5228 /* Check timestamp */
5229 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5230 /* No? Slow path! */
5231 if (!tcp_parse_aligned_timestamp(tp, th))
5234 /* If PAWS failed, check it more carefully in slow path */
5235 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5238 /* DO NOT update ts_recent here, if checksum fails
5239 * and timestamp was corrupted part, it will result
5240 * in a hung connection since we will drop all
5241 * future packets due to the PAWS test.
5245 if (len <= tcp_header_len) {
5246 /* Bulk data transfer: sender */
5247 if (len == tcp_header_len) {
5248 /* Predicted packet is in window by definition.
5249 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5250 * Hence, check seq<=rcv_wup reduces to:
5252 if (tcp_header_len ==
5253 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5254 tp->rcv_nxt == tp->rcv_wup)
5255 tcp_store_ts_recent(tp);
5257 /* We know that such packets are checksummed
5260 tcp_ack(sk, skb, 0);
5262 tcp_data_snd_check(sk);
5264 } else { /* Header too small */
5265 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5270 bool fragstolen = false;
5272 if (tp->ucopy.task == current &&
5273 tp->copied_seq == tp->rcv_nxt &&
5274 len - tcp_header_len <= tp->ucopy.len &&
5275 sock_owned_by_user(sk)) {
5276 __set_current_state(TASK_RUNNING);
5278 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) {
5279 /* Predicted packet is in window by definition.
5280 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5281 * Hence, check seq<=rcv_wup reduces to:
5283 if (tcp_header_len ==
5284 (sizeof(struct tcphdr) +
5285 TCPOLEN_TSTAMP_ALIGNED) &&
5286 tp->rcv_nxt == tp->rcv_wup)
5287 tcp_store_ts_recent(tp);
5289 tcp_rcv_rtt_measure_ts(sk, skb);
5291 __skb_pull(skb, tcp_header_len);
5292 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
5293 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
5298 if (tcp_checksum_complete_user(sk, skb))
5301 if ((int)skb->truesize > sk->sk_forward_alloc)
5304 /* Predicted packet is in window by definition.
5305 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5306 * Hence, check seq<=rcv_wup reduces to:
5308 if (tcp_header_len ==
5309 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5310 tp->rcv_nxt == tp->rcv_wup)
5311 tcp_store_ts_recent(tp);
5313 tcp_rcv_rtt_measure_ts(sk, skb);
5315 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
5317 /* Bulk data transfer: receiver */
5318 eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5322 tcp_event_data_recv(sk, skb);
5324 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5325 /* Well, only one small jumplet in fast path... */
5326 tcp_ack(sk, skb, FLAG_DATA);
5327 tcp_data_snd_check(sk);
5328 if (!inet_csk_ack_scheduled(sk))
5332 __tcp_ack_snd_check(sk, 0);
5335 kfree_skb_partial(skb, fragstolen);
5336 sk->sk_data_ready(sk);
5342 if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
5345 if (!th->ack && !th->rst && !th->syn)
5349 * Standard slow path.
5352 if (!tcp_validate_incoming(sk, skb, th, 1))
5356 if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5359 tcp_rcv_rtt_measure_ts(sk, skb);
5361 /* Process urgent data. */
5362 tcp_urg(sk, skb, th);
5364 /* step 7: process the segment text */
5365 tcp_data_queue(sk, skb);
5367 tcp_data_snd_check(sk);
5368 tcp_ack_snd_check(sk);
5372 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_CSUMERRORS);
5373 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5378 EXPORT_SYMBOL(tcp_rcv_established);
5380 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5382 struct tcp_sock *tp = tcp_sk(sk);
5383 struct inet_connection_sock *icsk = inet_csk(sk);
5385 tcp_set_state(sk, TCP_ESTABLISHED);
5388 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5389 security_inet_conn_established(sk, skb);
5392 /* Make sure socket is routed, for correct metrics. */
5393 icsk->icsk_af_ops->rebuild_header(sk);
5395 tcp_init_metrics(sk);
5397 tcp_init_congestion_control(sk);
5399 /* Prevent spurious tcp_cwnd_restart() on first data
5402 tp->lsndtime = tcp_time_stamp;
5404 tcp_init_buffer_space(sk);
5406 if (sock_flag(sk, SOCK_KEEPOPEN))
5407 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5409 if (!tp->rx_opt.snd_wscale)
5410 __tcp_fast_path_on(tp, tp->snd_wnd);
5414 if (!sock_flag(sk, SOCK_DEAD)) {
5415 sk->sk_state_change(sk);
5416 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5420 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5421 struct tcp_fastopen_cookie *cookie)
5423 struct tcp_sock *tp = tcp_sk(sk);
5424 struct sk_buff *data = tp->syn_data ? tcp_write_queue_head(sk) : NULL;
5425 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
5426 bool syn_drop = false;
5428 if (mss == tp->rx_opt.user_mss) {
5429 struct tcp_options_received opt;
5431 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5432 tcp_clear_options(&opt);
5433 opt.user_mss = opt.mss_clamp = 0;
5434 tcp_parse_options(synack, &opt, 0, NULL);
5435 mss = opt.mss_clamp;
5438 if (!tp->syn_fastopen) {
5439 /* Ignore an unsolicited cookie */
5441 } else if (tp->total_retrans) {
5442 /* SYN timed out and the SYN-ACK neither has a cookie nor
5443 * acknowledges data. Presumably the remote received only
5444 * the retransmitted (regular) SYNs: either the original
5445 * SYN-data or the corresponding SYN-ACK was dropped.
5447 syn_drop = (cookie->len < 0 && data);
5448 } else if (cookie->len < 0 && !tp->syn_data) {
5449 /* We requested a cookie but didn't get it. If we did not use
5450 * the (old) exp opt format then try so next time (try_exp=1).
5451 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5453 try_exp = tp->syn_fastopen_exp ? 2 : 1;
5456 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
5458 if (data) { /* Retransmit unacked data in SYN */
5459 tcp_for_write_queue_from(data, sk) {
5460 if (data == tcp_send_head(sk) ||
5461 __tcp_retransmit_skb(sk, data))
5465 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVEFAIL);
5468 tp->syn_data_acked = tp->syn_data;
5469 if (tp->syn_data_acked)
5470 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
5474 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5475 const struct tcphdr *th, unsigned int len)
5477 struct inet_connection_sock *icsk = inet_csk(sk);
5478 struct tcp_sock *tp = tcp_sk(sk);
5479 struct tcp_fastopen_cookie foc = { .len = -1 };
5480 int saved_clamp = tp->rx_opt.mss_clamp;
5482 tcp_parse_options(skb, &tp->rx_opt, 0, &foc);
5483 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
5484 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
5488 * "If the state is SYN-SENT then
5489 * first check the ACK bit
5490 * If the ACK bit is set
5491 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5492 * a reset (unless the RST bit is set, if so drop
5493 * the segment and return)"
5495 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5496 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
5497 goto reset_and_undo;
5499 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5500 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5502 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
5503 goto reset_and_undo;
5506 /* Now ACK is acceptable.
5508 * "If the RST bit is set
5509 * If the ACK was acceptable then signal the user "error:
5510 * connection reset", drop the segment, enter CLOSED state,
5511 * delete TCB, and return."
5520 * "fifth, if neither of the SYN or RST bits is set then
5521 * drop the segment and return."
5527 goto discard_and_undo;
5530 * "If the SYN bit is on ...
5531 * are acceptable then ...
5532 * (our SYN has been ACKed), change the connection
5533 * state to ESTABLISHED..."
5536 tcp_ecn_rcv_synack(tp, th);
5538 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5539 tcp_ack(sk, skb, FLAG_SLOWPATH);
5541 /* Ok.. it's good. Set up sequence numbers and
5542 * move to established.
5544 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5545 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5547 /* RFC1323: The window in SYN & SYN/ACK segments is
5550 tp->snd_wnd = ntohs(th->window);
5552 if (!tp->rx_opt.wscale_ok) {
5553 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5554 tp->window_clamp = min(tp->window_clamp, 65535U);
5557 if (tp->rx_opt.saw_tstamp) {
5558 tp->rx_opt.tstamp_ok = 1;
5559 tp->tcp_header_len =
5560 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5561 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5562 tcp_store_ts_recent(tp);
5564 tp->tcp_header_len = sizeof(struct tcphdr);
5567 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5568 tcp_enable_fack(tp);
5571 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5572 tcp_initialize_rcv_mss(sk);
5574 /* Remember, tcp_poll() does not lock socket!
5575 * Change state from SYN-SENT only after copied_seq
5576 * is initialized. */
5577 tp->copied_seq = tp->rcv_nxt;
5581 tcp_finish_connect(sk, skb);
5583 if ((tp->syn_fastopen || tp->syn_data) &&
5584 tcp_rcv_fastopen_synack(sk, skb, &foc))
5587 if (sk->sk_write_pending ||
5588 icsk->icsk_accept_queue.rskq_defer_accept ||
5589 icsk->icsk_ack.pingpong) {
5590 /* Save one ACK. Data will be ready after
5591 * several ticks, if write_pending is set.
5593 * It may be deleted, but with this feature tcpdumps
5594 * look so _wonderfully_ clever, that I was not able
5595 * to stand against the temptation 8) --ANK
5597 inet_csk_schedule_ack(sk);
5598 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5599 tcp_enter_quickack_mode(sk);
5600 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5601 TCP_DELACK_MAX, TCP_RTO_MAX);
5612 /* No ACK in the segment */
5616 * "If the RST bit is set
5618 * Otherwise (no ACK) drop the segment and return."
5621 goto discard_and_undo;
5625 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5626 tcp_paws_reject(&tp->rx_opt, 0))
5627 goto discard_and_undo;
5630 /* We see SYN without ACK. It is attempt of
5631 * simultaneous connect with crossed SYNs.
5632 * Particularly, it can be connect to self.
5634 tcp_set_state(sk, TCP_SYN_RECV);
5636 if (tp->rx_opt.saw_tstamp) {
5637 tp->rx_opt.tstamp_ok = 1;
5638 tcp_store_ts_recent(tp);
5639 tp->tcp_header_len =
5640 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5642 tp->tcp_header_len = sizeof(struct tcphdr);
5645 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5646 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5648 /* RFC1323: The window in SYN & SYN/ACK segments is
5651 tp->snd_wnd = ntohs(th->window);
5652 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5653 tp->max_window = tp->snd_wnd;
5655 tcp_ecn_rcv_syn(tp, th);
5658 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5659 tcp_initialize_rcv_mss(sk);
5661 tcp_send_synack(sk);
5663 /* Note, we could accept data and URG from this segment.
5664 * There are no obstacles to make this (except that we must
5665 * either change tcp_recvmsg() to prevent it from returning data
5666 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5668 * However, if we ignore data in ACKless segments sometimes,
5669 * we have no reasons to accept it sometimes.
5670 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5671 * is not flawless. So, discard packet for sanity.
5672 * Uncomment this return to process the data.
5679 /* "fifth, if neither of the SYN or RST bits is set then
5680 * drop the segment and return."
5684 tcp_clear_options(&tp->rx_opt);
5685 tp->rx_opt.mss_clamp = saved_clamp;
5689 tcp_clear_options(&tp->rx_opt);
5690 tp->rx_opt.mss_clamp = saved_clamp;
5695 * This function implements the receiving procedure of RFC 793 for
5696 * all states except ESTABLISHED and TIME_WAIT.
5697 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5698 * address independent.
5701 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5702 const struct tcphdr *th, unsigned int len)
5704 struct tcp_sock *tp = tcp_sk(sk);
5705 struct inet_connection_sock *icsk = inet_csk(sk);
5706 struct request_sock *req;
5711 tp->rx_opt.saw_tstamp = 0;
5713 switch (sk->sk_state) {
5727 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5730 /* Now we have several options: In theory there is
5731 * nothing else in the frame. KA9Q has an option to
5732 * send data with the syn, BSD accepts data with the
5733 * syn up to the [to be] advertised window and
5734 * Solaris 2.1 gives you a protocol error. For now
5735 * we just ignore it, that fits the spec precisely
5736 * and avoids incompatibilities. It would be nice in
5737 * future to drop through and process the data.
5739 * Now that TTCP is starting to be used we ought to
5741 * But, this leaves one open to an easy denial of
5742 * service attack, and SYN cookies can't defend
5743 * against this problem. So, we drop the data
5744 * in the interest of security over speed unless
5745 * it's still in use.
5753 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5757 /* Do step6 onward by hand. */
5758 tcp_urg(sk, skb, th);
5760 tcp_data_snd_check(sk);
5764 req = tp->fastopen_rsk;
5766 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
5767 sk->sk_state != TCP_FIN_WAIT1);
5769 if (!tcp_check_req(sk, skb, req, true))
5773 if (!th->ack && !th->rst && !th->syn)
5776 if (!tcp_validate_incoming(sk, skb, th, 0))
5779 /* step 5: check the ACK field */
5780 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
5781 FLAG_UPDATE_TS_RECENT) > 0;
5783 switch (sk->sk_state) {
5788 /* Once we leave TCP_SYN_RECV, we no longer need req
5792 synack_stamp = tcp_rsk(req)->snt_synack;
5793 tp->total_retrans = req->num_retrans;
5794 reqsk_fastopen_remove(sk, req, false);
5796 synack_stamp = tp->lsndtime;
5797 /* Make sure socket is routed, for correct metrics. */
5798 icsk->icsk_af_ops->rebuild_header(sk);
5799 tcp_init_congestion_control(sk);
5802 tp->copied_seq = tp->rcv_nxt;
5803 tcp_init_buffer_space(sk);
5806 tcp_set_state(sk, TCP_ESTABLISHED);
5807 sk->sk_state_change(sk);
5809 /* Note, that this wakeup is only for marginal crossed SYN case.
5810 * Passively open sockets are not waked up, because
5811 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5814 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5816 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5817 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
5818 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5819 tcp_synack_rtt_meas(sk, synack_stamp);
5821 if (tp->rx_opt.tstamp_ok)
5822 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5825 /* Re-arm the timer because data may have been sent out.
5826 * This is similar to the regular data transmission case
5827 * when new data has just been ack'ed.
5829 * (TFO) - we could try to be more aggressive and
5830 * retransmitting any data sooner based on when they
5835 tcp_init_metrics(sk);
5837 tcp_update_pacing_rate(sk);
5839 /* Prevent spurious tcp_cwnd_restart() on first data packet */
5840 tp->lsndtime = tcp_time_stamp;
5842 tcp_initialize_rcv_mss(sk);
5843 tcp_fast_path_on(tp);
5846 case TCP_FIN_WAIT1: {
5847 struct dst_entry *dst;
5850 /* If we enter the TCP_FIN_WAIT1 state and we are a
5851 * Fast Open socket and this is the first acceptable
5852 * ACK we have received, this would have acknowledged
5853 * our SYNACK so stop the SYNACK timer.
5856 /* Return RST if ack_seq is invalid.
5857 * Note that RFC793 only says to generate a
5858 * DUPACK for it but for TCP Fast Open it seems
5859 * better to treat this case like TCP_SYN_RECV
5864 /* We no longer need the request sock. */
5865 reqsk_fastopen_remove(sk, req, false);
5868 if (tp->snd_una != tp->write_seq)
5871 tcp_set_state(sk, TCP_FIN_WAIT2);
5872 sk->sk_shutdown |= SEND_SHUTDOWN;
5874 dst = __sk_dst_get(sk);
5878 if (!sock_flag(sk, SOCK_DEAD)) {
5879 /* Wake up lingering close() */
5880 sk->sk_state_change(sk);
5884 if (tp->linger2 < 0 ||
5885 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5886 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5888 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5892 tmo = tcp_fin_time(sk);
5893 if (tmo > TCP_TIMEWAIT_LEN) {
5894 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5895 } else if (th->fin || sock_owned_by_user(sk)) {
5896 /* Bad case. We could lose such FIN otherwise.
5897 * It is not a big problem, but it looks confusing
5898 * and not so rare event. We still can lose it now,
5899 * if it spins in bh_lock_sock(), but it is really
5902 inet_csk_reset_keepalive_timer(sk, tmo);
5904 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5911 if (tp->snd_una == tp->write_seq) {
5912 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5918 if (tp->snd_una == tp->write_seq) {
5919 tcp_update_metrics(sk);
5926 /* step 6: check the URG bit */
5927 tcp_urg(sk, skb, th);
5929 /* step 7: process the segment text */
5930 switch (sk->sk_state) {
5931 case TCP_CLOSE_WAIT:
5934 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5938 /* RFC 793 says to queue data in these states,
5939 * RFC 1122 says we MUST send a reset.
5940 * BSD 4.4 also does reset.
5942 if (sk->sk_shutdown & RCV_SHUTDOWN) {
5943 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5944 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5945 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5951 case TCP_ESTABLISHED:
5952 tcp_data_queue(sk, skb);
5957 /* tcp_data could move socket to TIME-WAIT */
5958 if (sk->sk_state != TCP_CLOSE) {
5959 tcp_data_snd_check(sk);
5960 tcp_ack_snd_check(sk);
5969 EXPORT_SYMBOL(tcp_rcv_state_process);
5971 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
5973 struct inet_request_sock *ireq = inet_rsk(req);
5975 if (family == AF_INET)
5976 net_dbg_ratelimited("drop open request from %pI4/%u\n",
5977 &ireq->ir_rmt_addr, port);
5978 #if IS_ENABLED(CONFIG_IPV6)
5979 else if (family == AF_INET6)
5980 net_dbg_ratelimited("drop open request from %pI6/%u\n",
5981 &ireq->ir_v6_rmt_addr, port);
5985 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
5987 * If we receive a SYN packet with these bits set, it means a
5988 * network is playing bad games with TOS bits. In order to
5989 * avoid possible false congestion notifications, we disable
5990 * TCP ECN negotiation.
5992 * Exception: tcp_ca wants ECN. This is required for DCTCP
5993 * congestion control: Linux DCTCP asserts ECT on all packets,
5994 * including SYN, which is most optimal solution; however,
5995 * others, such as FreeBSD do not.
5997 static void tcp_ecn_create_request(struct request_sock *req,
5998 const struct sk_buff *skb,
5999 const struct sock *listen_sk,
6000 const struct dst_entry *dst)
6002 const struct tcphdr *th = tcp_hdr(skb);
6003 const struct net *net = sock_net(listen_sk);
6004 bool th_ecn = th->ece && th->cwr;
6010 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
6011 ecn_ok = net->ipv4.sysctl_tcp_ecn || dst_feature(dst, RTAX_FEATURE_ECN);
6013 if ((!ect && ecn_ok) || tcp_ca_needs_ecn(listen_sk))
6014 inet_rsk(req)->ecn_ok = 1;
6017 static void tcp_openreq_init(struct request_sock *req,
6018 const struct tcp_options_received *rx_opt,
6019 struct sk_buff *skb, const struct sock *sk)
6021 struct inet_request_sock *ireq = inet_rsk(req);
6023 req->rcv_wnd = 0; /* So that tcp_send_synack() knows! */
6025 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6026 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
6027 tcp_rsk(req)->snt_synack = tcp_time_stamp;
6028 tcp_rsk(req)->last_oow_ack_time = 0;
6029 req->mss = rx_opt->mss_clamp;
6030 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6031 ireq->tstamp_ok = rx_opt->tstamp_ok;
6032 ireq->sack_ok = rx_opt->sack_ok;
6033 ireq->snd_wscale = rx_opt->snd_wscale;
6034 ireq->wscale_ok = rx_opt->wscale_ok;
6037 ireq->ir_rmt_port = tcp_hdr(skb)->source;
6038 ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6039 ireq->ir_mark = inet_request_mark(sk, skb);
6042 struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
6043 struct sock *sk_listener)
6045 struct request_sock *req = reqsk_alloc(ops, sk_listener);
6048 struct inet_request_sock *ireq = inet_rsk(req);
6050 kmemcheck_annotate_bitfield(ireq, flags);
6052 atomic64_set(&ireq->ir_cookie, 0);
6053 ireq->ireq_state = TCP_NEW_SYN_RECV;
6054 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
6055 ireq->ireq_family = sk_listener->sk_family;
6060 EXPORT_SYMBOL(inet_reqsk_alloc);
6063 * Return true if a syncookie should be sent
6065 static bool tcp_syn_flood_action(struct sock *sk,
6066 const struct sk_buff *skb,
6069 const char *msg = "Dropping request";
6070 bool want_cookie = false;
6071 struct listen_sock *lopt;
6073 #ifdef CONFIG_SYN_COOKIES
6074 if (sysctl_tcp_syncookies) {
6075 msg = "Sending cookies";
6077 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
6080 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
6082 lopt = inet_csk(sk)->icsk_accept_queue.listen_opt;
6083 if (!lopt->synflood_warned && sysctl_tcp_syncookies != 2) {
6084 lopt->synflood_warned = 1;
6085 pr_info("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6086 proto, ntohs(tcp_hdr(skb)->dest), msg);
6091 static void tcp_reqsk_record_syn(const struct sock *sk,
6092 struct request_sock *req,
6093 const struct sk_buff *skb)
6095 if (tcp_sk(sk)->save_syn) {
6096 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
6099 copy = kmalloc(len + sizeof(u32), GFP_ATOMIC);
6102 memcpy(©[1], skb_network_header(skb), len);
6103 req->saved_syn = copy;
6108 int tcp_conn_request(struct request_sock_ops *rsk_ops,
6109 const struct tcp_request_sock_ops *af_ops,
6110 struct sock *sk, struct sk_buff *skb)
6112 struct tcp_options_received tmp_opt;
6113 struct request_sock *req;
6114 struct tcp_sock *tp = tcp_sk(sk);
6115 struct dst_entry *dst = NULL;
6116 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
6117 bool want_cookie = false, fastopen;
6119 struct tcp_fastopen_cookie foc = { .len = -1 };
6123 /* TW buckets are converted to open requests without
6124 * limitations, they conserve resources and peer is
6125 * evidently real one.
6127 if ((sysctl_tcp_syncookies == 2 ||
6128 inet_csk_reqsk_queue_is_full(sk)) && !isn) {
6129 want_cookie = tcp_syn_flood_action(sk, skb, rsk_ops->slab_name);
6135 /* Accept backlog is full. If we have already queued enough
6136 * of warm entries in syn queue, drop request. It is better than
6137 * clogging syn queue with openreqs with exponentially increasing
6140 if (sk_acceptq_is_full(sk) && inet_csk_reqsk_queue_young(sk) > 1) {
6141 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6145 req = inet_reqsk_alloc(rsk_ops, sk);
6149 tcp_rsk(req)->af_specific = af_ops;
6151 tcp_clear_options(&tmp_opt);
6152 tmp_opt.mss_clamp = af_ops->mss_clamp;
6153 tmp_opt.user_mss = tp->rx_opt.user_mss;
6154 tcp_parse_options(skb, &tmp_opt, 0, want_cookie ? NULL : &foc);
6156 if (want_cookie && !tmp_opt.saw_tstamp)
6157 tcp_clear_options(&tmp_opt);
6159 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
6160 tcp_openreq_init(req, &tmp_opt, skb, sk);
6162 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6163 inet_rsk(req)->ir_iif = sk->sk_bound_dev_if;
6165 af_ops->init_req(req, sk, skb);
6167 if (security_inet_conn_request(sk, skb, req))
6170 if (!want_cookie && !isn) {
6171 /* VJ's idea. We save last timestamp seen
6172 * from the destination in peer table, when entering
6173 * state TIME-WAIT, and check against it before
6174 * accepting new connection request.
6176 * If "isn" is not zero, this request hit alive
6177 * timewait bucket, so that all the necessary checks
6178 * are made in the function processing timewait state.
6180 if (tcp_death_row.sysctl_tw_recycle) {
6183 dst = af_ops->route_req(sk, &fl, req, &strict);
6185 if (dst && strict &&
6186 !tcp_peer_is_proven(req, dst, true,
6187 tmp_opt.saw_tstamp)) {
6188 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSPASSIVEREJECTED);
6189 goto drop_and_release;
6192 /* Kill the following clause, if you dislike this way. */
6193 else if (!sysctl_tcp_syncookies &&
6194 (sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
6195 (sysctl_max_syn_backlog >> 2)) &&
6196 !tcp_peer_is_proven(req, dst, false,
6197 tmp_opt.saw_tstamp)) {
6198 /* Without syncookies last quarter of
6199 * backlog is filled with destinations,
6200 * proven to be alive.
6201 * It means that we continue to communicate
6202 * to destinations, already remembered
6203 * to the moment of synflood.
6205 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
6207 goto drop_and_release;
6210 isn = af_ops->init_seq(skb);
6213 dst = af_ops->route_req(sk, &fl, req, NULL);
6218 tcp_ecn_create_request(req, skb, sk, dst);
6221 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
6222 req->cookie_ts = tmp_opt.tstamp_ok;
6223 if (!tmp_opt.tstamp_ok)
6224 inet_rsk(req)->ecn_ok = 0;
6227 tcp_rsk(req)->snt_isn = isn;
6228 tcp_openreq_init_rwin(req, sk, dst);
6229 fastopen = !want_cookie &&
6230 tcp_try_fastopen(sk, skb, req, &foc, dst);
6231 err = af_ops->send_synack(sk, dst, &fl, req,
6232 skb_get_queue_mapping(skb), &foc);
6234 if (err || want_cookie)
6237 tcp_rsk(req)->tfo_listener = false;
6238 af_ops->queue_hash_add(sk, req, TCP_TIMEOUT_INIT);
6240 tcp_reqsk_record_syn(sk, req, skb);
6249 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_LISTENDROPS);
6252 EXPORT_SYMBOL(tcp_conn_request);