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_fack __read_mostly;
82 int sysctl_tcp_max_reordering __read_mostly = 300;
83 int sysctl_tcp_dsack __read_mostly = 1;
84 int sysctl_tcp_app_win __read_mostly = 31;
85 int sysctl_tcp_adv_win_scale __read_mostly = 1;
86 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
88 /* rfc5961 challenge ack rate limiting */
89 int sysctl_tcp_challenge_ack_limit = 1000;
91 int sysctl_tcp_stdurg __read_mostly;
92 int sysctl_tcp_rfc1337 __read_mostly;
93 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
94 int sysctl_tcp_frto __read_mostly = 2;
95 int sysctl_tcp_min_rtt_wlen __read_mostly = 300;
96 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
97 int sysctl_tcp_early_retrans __read_mostly = 3;
98 int sysctl_tcp_invalid_ratelimit __read_mostly = HZ/2;
100 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
101 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
102 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
103 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
104 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
105 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
106 #define FLAG_ECE 0x40 /* ECE in this ACK */
107 #define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */
108 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
109 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
110 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
111 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
112 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
113 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
114 #define FLAG_NO_CHALLENGE_ACK 0x8000 /* do not call tcp_send_challenge_ack() */
116 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
117 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
118 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
119 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
121 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
122 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
124 #define REXMIT_NONE 0 /* no loss recovery to do */
125 #define REXMIT_LOST 1 /* retransmit packets marked lost */
126 #define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */
128 static void tcp_gro_dev_warn(struct sock *sk, const struct sk_buff *skb,
131 static bool __once __read_mostly;
134 struct net_device *dev;
139 dev = dev_get_by_index_rcu(sock_net(sk), skb->skb_iif);
140 if (!dev || len >= dev->mtu)
141 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n",
142 dev ? dev->name : "Unknown driver");
147 /* Adapt the MSS value used to make delayed ack decision to the
150 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
152 struct inet_connection_sock *icsk = inet_csk(sk);
153 const unsigned int lss = icsk->icsk_ack.last_seg_size;
156 icsk->icsk_ack.last_seg_size = 0;
158 /* skb->len may jitter because of SACKs, even if peer
159 * sends good full-sized frames.
161 len = skb_shinfo(skb)->gso_size ? : skb->len;
162 if (len >= icsk->icsk_ack.rcv_mss) {
163 icsk->icsk_ack.rcv_mss = min_t(unsigned int, len,
165 /* Account for possibly-removed options */
166 if (unlikely(len > icsk->icsk_ack.rcv_mss +
167 MAX_TCP_OPTION_SPACE))
168 tcp_gro_dev_warn(sk, skb, len);
170 /* Otherwise, we make more careful check taking into account,
171 * that SACKs block is variable.
173 * "len" is invariant segment length, including TCP header.
175 len += skb->data - skb_transport_header(skb);
176 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
177 /* If PSH is not set, packet should be
178 * full sized, provided peer TCP is not badly broken.
179 * This observation (if it is correct 8)) allows
180 * to handle super-low mtu links fairly.
182 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
183 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
184 /* Subtract also invariant (if peer is RFC compliant),
185 * tcp header plus fixed timestamp option length.
186 * Resulting "len" is MSS free of SACK jitter.
188 len -= tcp_sk(sk)->tcp_header_len;
189 icsk->icsk_ack.last_seg_size = len;
191 icsk->icsk_ack.rcv_mss = len;
195 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
196 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
197 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
201 static void tcp_incr_quickack(struct sock *sk)
203 struct inet_connection_sock *icsk = inet_csk(sk);
204 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
208 if (quickacks > icsk->icsk_ack.quick)
209 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
212 static void tcp_enter_quickack_mode(struct sock *sk)
214 struct inet_connection_sock *icsk = inet_csk(sk);
215 tcp_incr_quickack(sk);
216 icsk->icsk_ack.pingpong = 0;
217 icsk->icsk_ack.ato = TCP_ATO_MIN;
220 /* Send ACKs quickly, if "quick" count is not exhausted
221 * and the session is not interactive.
224 static bool tcp_in_quickack_mode(struct sock *sk)
226 const struct inet_connection_sock *icsk = inet_csk(sk);
227 const struct dst_entry *dst = __sk_dst_get(sk);
229 return (dst && dst_metric(dst, RTAX_QUICKACK)) ||
230 (icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong);
233 static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
235 if (tp->ecn_flags & TCP_ECN_OK)
236 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
239 static void tcp_ecn_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
241 if (tcp_hdr(skb)->cwr)
242 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
245 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
247 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
250 static void __tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
252 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
253 case INET_ECN_NOT_ECT:
254 /* Funny extension: if ECT is not set on a segment,
255 * and we already seen ECT on a previous segment,
256 * it is probably a retransmit.
258 if (tp->ecn_flags & TCP_ECN_SEEN)
259 tcp_enter_quickack_mode((struct sock *)tp);
262 if (tcp_ca_needs_ecn((struct sock *)tp))
263 tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_IS_CE);
265 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
266 /* Better not delay acks, sender can have a very low cwnd */
267 tcp_enter_quickack_mode((struct sock *)tp);
268 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
270 tp->ecn_flags |= TCP_ECN_SEEN;
273 if (tcp_ca_needs_ecn((struct sock *)tp))
274 tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_NO_CE);
275 tp->ecn_flags |= TCP_ECN_SEEN;
280 static void tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
282 if (tp->ecn_flags & TCP_ECN_OK)
283 __tcp_ecn_check_ce(tp, skb);
286 static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
288 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
289 tp->ecn_flags &= ~TCP_ECN_OK;
292 static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
294 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
295 tp->ecn_flags &= ~TCP_ECN_OK;
298 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
300 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
305 /* Buffer size and advertised window tuning.
307 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
310 static void tcp_sndbuf_expand(struct sock *sk)
312 const struct tcp_sock *tp = tcp_sk(sk);
313 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
317 /* Worst case is non GSO/TSO : each frame consumes one skb
318 * and skb->head is kmalloced using power of two area of memory
320 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
322 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
324 per_mss = roundup_pow_of_two(per_mss) +
325 SKB_DATA_ALIGN(sizeof(struct sk_buff));
327 nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd);
328 nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
330 /* Fast Recovery (RFC 5681 3.2) :
331 * Cubic needs 1.7 factor, rounded to 2 to include
332 * extra cushion (application might react slowly to POLLOUT)
334 sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2;
335 sndmem *= nr_segs * per_mss;
337 if (sk->sk_sndbuf < sndmem)
338 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
341 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
343 * All tcp_full_space() is split to two parts: "network" buffer, allocated
344 * forward and advertised in receiver window (tp->rcv_wnd) and
345 * "application buffer", required to isolate scheduling/application
346 * latencies from network.
347 * window_clamp is maximal advertised window. It can be less than
348 * tcp_full_space(), in this case tcp_full_space() - window_clamp
349 * is reserved for "application" buffer. The less window_clamp is
350 * the smoother our behaviour from viewpoint of network, but the lower
351 * throughput and the higher sensitivity of the connection to losses. 8)
353 * rcv_ssthresh is more strict window_clamp used at "slow start"
354 * phase to predict further behaviour of this connection.
355 * It is used for two goals:
356 * - to enforce header prediction at sender, even when application
357 * requires some significant "application buffer". It is check #1.
358 * - to prevent pruning of receive queue because of misprediction
359 * of receiver window. Check #2.
361 * The scheme does not work when sender sends good segments opening
362 * window and then starts to feed us spaghetti. But it should work
363 * in common situations. Otherwise, we have to rely on queue collapsing.
366 /* Slow part of check#2. */
367 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
369 struct tcp_sock *tp = tcp_sk(sk);
371 int truesize = tcp_win_from_space(skb->truesize) >> 1;
372 int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
374 while (tp->rcv_ssthresh <= window) {
375 if (truesize <= skb->len)
376 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
384 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
386 struct tcp_sock *tp = tcp_sk(sk);
389 if (tp->rcv_ssthresh < tp->window_clamp &&
390 (int)tp->rcv_ssthresh < tcp_space(sk) &&
391 !tcp_under_memory_pressure(sk)) {
394 /* Check #2. Increase window, if skb with such overhead
395 * will fit to rcvbuf in future.
397 if (tcp_win_from_space(skb->truesize) <= skb->len)
398 incr = 2 * tp->advmss;
400 incr = __tcp_grow_window(sk, skb);
403 incr = max_t(int, incr, 2 * skb->len);
404 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
406 inet_csk(sk)->icsk_ack.quick |= 1;
411 /* 3. Tuning rcvbuf, when connection enters established state. */
412 static void tcp_fixup_rcvbuf(struct sock *sk)
414 u32 mss = tcp_sk(sk)->advmss;
417 rcvmem = 2 * SKB_TRUESIZE(mss + MAX_TCP_HEADER) *
418 tcp_default_init_rwnd(mss);
420 /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
421 * Allow enough cushion so that sender is not limited by our window
423 if (sysctl_tcp_moderate_rcvbuf)
426 if (sk->sk_rcvbuf < rcvmem)
427 sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]);
430 /* 4. Try to fixup all. It is made immediately after connection enters
433 void tcp_init_buffer_space(struct sock *sk)
435 struct tcp_sock *tp = tcp_sk(sk);
438 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
439 tcp_fixup_rcvbuf(sk);
440 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
441 tcp_sndbuf_expand(sk);
443 tp->rcvq_space.space = tp->rcv_wnd;
444 tcp_mstamp_refresh(tp);
445 tp->rcvq_space.time = tp->tcp_mstamp;
446 tp->rcvq_space.seq = tp->copied_seq;
448 maxwin = tcp_full_space(sk);
450 if (tp->window_clamp >= maxwin) {
451 tp->window_clamp = maxwin;
453 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
454 tp->window_clamp = max(maxwin -
455 (maxwin >> sysctl_tcp_app_win),
459 /* Force reservation of one segment. */
460 if (sysctl_tcp_app_win &&
461 tp->window_clamp > 2 * tp->advmss &&
462 tp->window_clamp + tp->advmss > maxwin)
463 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
465 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
466 tp->snd_cwnd_stamp = tcp_jiffies32;
469 /* 5. Recalculate window clamp after socket hit its memory bounds. */
470 static void tcp_clamp_window(struct sock *sk)
472 struct tcp_sock *tp = tcp_sk(sk);
473 struct inet_connection_sock *icsk = inet_csk(sk);
475 icsk->icsk_ack.quick = 0;
477 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
478 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
479 !tcp_under_memory_pressure(sk) &&
480 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
481 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
484 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
485 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
488 /* Initialize RCV_MSS value.
489 * RCV_MSS is an our guess about MSS used by the peer.
490 * We haven't any direct information about the MSS.
491 * It's better to underestimate the RCV_MSS rather than overestimate.
492 * Overestimations make us ACKing less frequently than needed.
493 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
495 void tcp_initialize_rcv_mss(struct sock *sk)
497 const struct tcp_sock *tp = tcp_sk(sk);
498 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
500 hint = min(hint, tp->rcv_wnd / 2);
501 hint = min(hint, TCP_MSS_DEFAULT);
502 hint = max(hint, TCP_MIN_MSS);
504 inet_csk(sk)->icsk_ack.rcv_mss = hint;
506 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
508 /* Receiver "autotuning" code.
510 * The algorithm for RTT estimation w/o timestamps is based on
511 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
512 * <http://public.lanl.gov/radiant/pubs.html#DRS>
514 * More detail on this code can be found at
515 * <http://staff.psc.edu/jheffner/>,
516 * though this reference is out of date. A new paper
519 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
521 u32 new_sample = tp->rcv_rtt_est.rtt_us;
527 if (new_sample != 0) {
528 /* If we sample in larger samples in the non-timestamp
529 * case, we could grossly overestimate the RTT especially
530 * with chatty applications or bulk transfer apps which
531 * are stalled on filesystem I/O.
533 * Also, since we are only going for a minimum in the
534 * non-timestamp case, we do not smooth things out
535 * else with timestamps disabled convergence takes too
539 m -= (new_sample >> 3);
547 /* No previous measure. */
551 tp->rcv_rtt_est.rtt_us = new_sample;
554 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
558 if (tp->rcv_rtt_est.time == 0)
560 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
562 delta_us = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcv_rtt_est.time);
563 tcp_rcv_rtt_update(tp, delta_us, 1);
566 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
567 tp->rcv_rtt_est.time = tp->tcp_mstamp;
570 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
571 const struct sk_buff *skb)
573 struct tcp_sock *tp = tcp_sk(sk);
575 if (tp->rx_opt.rcv_tsecr &&
576 (TCP_SKB_CB(skb)->end_seq -
577 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss)) {
578 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
579 u32 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
581 tcp_rcv_rtt_update(tp, delta_us, 0);
586 * This function should be called every time data is copied to user space.
587 * It calculates the appropriate TCP receive buffer space.
589 void tcp_rcv_space_adjust(struct sock *sk)
591 struct tcp_sock *tp = tcp_sk(sk);
595 time = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcvq_space.time);
596 if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0)
599 /* Number of bytes copied to user in last RTT */
600 copied = tp->copied_seq - tp->rcvq_space.seq;
601 if (copied <= tp->rcvq_space.space)
605 * copied = bytes received in previous RTT, our base window
606 * To cope with packet losses, we need a 2x factor
607 * To cope with slow start, and sender growing its cwin by 100 %
608 * every RTT, we need a 4x factor, because the ACK we are sending
609 * now is for the next RTT, not the current one :
610 * <prev RTT . ><current RTT .. ><next RTT .... >
613 if (sysctl_tcp_moderate_rcvbuf &&
614 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
615 int rcvwin, rcvmem, rcvbuf;
617 /* minimal window to cope with packet losses, assuming
618 * steady state. Add some cushion because of small variations.
620 rcvwin = (copied << 1) + 16 * tp->advmss;
622 /* If rate increased by 25%,
623 * assume slow start, rcvwin = 3 * copied
624 * If rate increased by 50%,
625 * assume sender can use 2x growth, rcvwin = 4 * copied
628 tp->rcvq_space.space + (tp->rcvq_space.space >> 2)) {
630 tp->rcvq_space.space + (tp->rcvq_space.space >> 1))
633 rcvwin += (rcvwin >> 1);
636 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
637 while (tcp_win_from_space(rcvmem) < tp->advmss)
640 rcvbuf = min(rcvwin / tp->advmss * rcvmem, sysctl_tcp_rmem[2]);
641 if (rcvbuf > sk->sk_rcvbuf) {
642 sk->sk_rcvbuf = rcvbuf;
644 /* Make the window clamp follow along. */
645 tp->window_clamp = rcvwin;
648 tp->rcvq_space.space = copied;
651 tp->rcvq_space.seq = tp->copied_seq;
652 tp->rcvq_space.time = tp->tcp_mstamp;
655 /* There is something which you must keep in mind when you analyze the
656 * behavior of the tp->ato delayed ack timeout interval. When a
657 * connection starts up, we want to ack as quickly as possible. The
658 * problem is that "good" TCP's do slow start at the beginning of data
659 * transmission. The means that until we send the first few ACK's the
660 * sender will sit on his end and only queue most of his data, because
661 * he can only send snd_cwnd unacked packets at any given time. For
662 * each ACK we send, he increments snd_cwnd and transmits more of his
665 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
667 struct tcp_sock *tp = tcp_sk(sk);
668 struct inet_connection_sock *icsk = inet_csk(sk);
671 inet_csk_schedule_ack(sk);
673 tcp_measure_rcv_mss(sk, skb);
675 tcp_rcv_rtt_measure(tp);
679 if (!icsk->icsk_ack.ato) {
680 /* The _first_ data packet received, initialize
681 * delayed ACK engine.
683 tcp_incr_quickack(sk);
684 icsk->icsk_ack.ato = TCP_ATO_MIN;
686 int m = now - icsk->icsk_ack.lrcvtime;
688 if (m <= TCP_ATO_MIN / 2) {
689 /* The fastest case is the first. */
690 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
691 } else if (m < icsk->icsk_ack.ato) {
692 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
693 if (icsk->icsk_ack.ato > icsk->icsk_rto)
694 icsk->icsk_ack.ato = icsk->icsk_rto;
695 } else if (m > icsk->icsk_rto) {
696 /* Too long gap. Apparently sender failed to
697 * restart window, so that we send ACKs quickly.
699 tcp_incr_quickack(sk);
703 icsk->icsk_ack.lrcvtime = now;
705 tcp_ecn_check_ce(tp, skb);
708 tcp_grow_window(sk, skb);
711 /* Called to compute a smoothed rtt estimate. The data fed to this
712 * routine either comes from timestamps, or from segments that were
713 * known _not_ to have been retransmitted [see Karn/Partridge
714 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
715 * piece by Van Jacobson.
716 * NOTE: the next three routines used to be one big routine.
717 * To save cycles in the RFC 1323 implementation it was better to break
718 * it up into three procedures. -- erics
720 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
722 struct tcp_sock *tp = tcp_sk(sk);
723 long m = mrtt_us; /* RTT */
724 u32 srtt = tp->srtt_us;
726 /* The following amusing code comes from Jacobson's
727 * article in SIGCOMM '88. Note that rtt and mdev
728 * are scaled versions of rtt and mean deviation.
729 * This is designed to be as fast as possible
730 * m stands for "measurement".
732 * On a 1990 paper the rto value is changed to:
733 * RTO = rtt + 4 * mdev
735 * Funny. This algorithm seems to be very broken.
736 * These formulae increase RTO, when it should be decreased, increase
737 * too slowly, when it should be increased quickly, decrease too quickly
738 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
739 * does not matter how to _calculate_ it. Seems, it was trap
740 * that VJ failed to avoid. 8)
743 m -= (srtt >> 3); /* m is now error in rtt est */
744 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
746 m = -m; /* m is now abs(error) */
747 m -= (tp->mdev_us >> 2); /* similar update on mdev */
748 /* This is similar to one of Eifel findings.
749 * Eifel blocks mdev updates when rtt decreases.
750 * This solution is a bit different: we use finer gain
751 * for mdev in this case (alpha*beta).
752 * Like Eifel it also prevents growth of rto,
753 * but also it limits too fast rto decreases,
754 * happening in pure Eifel.
759 m -= (tp->mdev_us >> 2); /* similar update on mdev */
761 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
762 if (tp->mdev_us > tp->mdev_max_us) {
763 tp->mdev_max_us = tp->mdev_us;
764 if (tp->mdev_max_us > tp->rttvar_us)
765 tp->rttvar_us = tp->mdev_max_us;
767 if (after(tp->snd_una, tp->rtt_seq)) {
768 if (tp->mdev_max_us < tp->rttvar_us)
769 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
770 tp->rtt_seq = tp->snd_nxt;
771 tp->mdev_max_us = tcp_rto_min_us(sk);
774 /* no previous measure. */
775 srtt = m << 3; /* take the measured time to be rtt */
776 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
777 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
778 tp->mdev_max_us = tp->rttvar_us;
779 tp->rtt_seq = tp->snd_nxt;
781 tp->srtt_us = max(1U, srtt);
784 /* Set the sk_pacing_rate to allow proper sizing of TSO packets.
785 * Note: TCP stack does not yet implement pacing.
786 * FQ packet scheduler can be used to implement cheap but effective
787 * TCP pacing, to smooth the burst on large writes when packets
788 * in flight is significantly lower than cwnd (or rwin)
790 int sysctl_tcp_pacing_ss_ratio __read_mostly = 200;
791 int sysctl_tcp_pacing_ca_ratio __read_mostly = 120;
793 static void tcp_update_pacing_rate(struct sock *sk)
795 const struct tcp_sock *tp = tcp_sk(sk);
798 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
799 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
801 /* current rate is (cwnd * mss) / srtt
802 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
803 * In Congestion Avoidance phase, set it to 120 % the current rate.
805 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
806 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
807 * end of slow start and should slow down.
809 if (tp->snd_cwnd < tp->snd_ssthresh / 2)
810 rate *= sysctl_tcp_pacing_ss_ratio;
812 rate *= sysctl_tcp_pacing_ca_ratio;
814 rate *= max(tp->snd_cwnd, tp->packets_out);
816 if (likely(tp->srtt_us))
817 do_div(rate, tp->srtt_us);
819 /* ACCESS_ONCE() is needed because sch_fq fetches sk_pacing_rate
820 * without any lock. We want to make sure compiler wont store
821 * intermediate values in this location.
823 ACCESS_ONCE(sk->sk_pacing_rate) = min_t(u64, rate,
824 sk->sk_max_pacing_rate);
827 /* Calculate rto without backoff. This is the second half of Van Jacobson's
828 * routine referred to above.
830 static void tcp_set_rto(struct sock *sk)
832 const struct tcp_sock *tp = tcp_sk(sk);
833 /* Old crap is replaced with new one. 8)
836 * 1. If rtt variance happened to be less 50msec, it is hallucination.
837 * It cannot be less due to utterly erratic ACK generation made
838 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
839 * to do with delayed acks, because at cwnd>2 true delack timeout
840 * is invisible. Actually, Linux-2.4 also generates erratic
841 * ACKs in some circumstances.
843 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
845 /* 2. Fixups made earlier cannot be right.
846 * If we do not estimate RTO correctly without them,
847 * all the algo is pure shit and should be replaced
848 * with correct one. It is exactly, which we pretend to do.
851 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
852 * guarantees that rto is higher.
857 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
859 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
862 cwnd = TCP_INIT_CWND;
863 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
867 * Packet counting of FACK is based on in-order assumptions, therefore TCP
868 * disables it when reordering is detected
870 void tcp_disable_fack(struct tcp_sock *tp)
872 /* RFC3517 uses different metric in lost marker => reset on change */
874 tp->lost_skb_hint = NULL;
875 tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED;
878 /* Take a notice that peer is sending D-SACKs */
879 static void tcp_dsack_seen(struct tcp_sock *tp)
881 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
884 static void tcp_update_reordering(struct sock *sk, const int metric,
887 struct tcp_sock *tp = tcp_sk(sk);
890 if (WARN_ON_ONCE(metric < 0))
893 if (metric > tp->reordering) {
894 tp->reordering = min(sysctl_tcp_max_reordering, metric);
896 #if FASTRETRANS_DEBUG > 1
897 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
898 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
902 tp->undo_marker ? tp->undo_retrans : 0);
904 tcp_disable_fack(tp);
909 /* This exciting event is worth to be remembered. 8) */
911 mib_idx = LINUX_MIB_TCPTSREORDER;
912 else if (tcp_is_reno(tp))
913 mib_idx = LINUX_MIB_TCPRENOREORDER;
914 else if (tcp_is_fack(tp))
915 mib_idx = LINUX_MIB_TCPFACKREORDER;
917 mib_idx = LINUX_MIB_TCPSACKREORDER;
919 NET_INC_STATS(sock_net(sk), mib_idx);
922 /* This must be called before lost_out is incremented */
923 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
925 if (!tp->retransmit_skb_hint ||
926 before(TCP_SKB_CB(skb)->seq,
927 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
928 tp->retransmit_skb_hint = skb;
931 /* Sum the number of packets on the wire we have marked as lost.
932 * There are two cases we care about here:
933 * a) Packet hasn't been marked lost (nor retransmitted),
934 * and this is the first loss.
935 * b) Packet has been marked both lost and retransmitted,
936 * and this means we think it was lost again.
938 static void tcp_sum_lost(struct tcp_sock *tp, struct sk_buff *skb)
940 __u8 sacked = TCP_SKB_CB(skb)->sacked;
942 if (!(sacked & TCPCB_LOST) ||
943 ((sacked & TCPCB_LOST) && (sacked & TCPCB_SACKED_RETRANS)))
944 tp->lost += tcp_skb_pcount(skb);
947 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
949 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
950 tcp_verify_retransmit_hint(tp, skb);
952 tp->lost_out += tcp_skb_pcount(skb);
953 tcp_sum_lost(tp, skb);
954 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
958 void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, struct sk_buff *skb)
960 tcp_verify_retransmit_hint(tp, skb);
962 tcp_sum_lost(tp, skb);
963 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
964 tp->lost_out += tcp_skb_pcount(skb);
965 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
969 /* This procedure tags the retransmission queue when SACKs arrive.
971 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
972 * Packets in queue with these bits set are counted in variables
973 * sacked_out, retrans_out and lost_out, correspondingly.
975 * Valid combinations are:
976 * Tag InFlight Description
977 * 0 1 - orig segment is in flight.
978 * S 0 - nothing flies, orig reached receiver.
979 * L 0 - nothing flies, orig lost by net.
980 * R 2 - both orig and retransmit are in flight.
981 * L|R 1 - orig is lost, retransmit is in flight.
982 * S|R 1 - orig reached receiver, retrans is still in flight.
983 * (L|S|R is logically valid, it could occur when L|R is sacked,
984 * but it is equivalent to plain S and code short-curcuits it to S.
985 * L|S is logically invalid, it would mean -1 packet in flight 8))
987 * These 6 states form finite state machine, controlled by the following events:
988 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
989 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
990 * 3. Loss detection event of two flavors:
991 * A. Scoreboard estimator decided the packet is lost.
992 * A'. Reno "three dupacks" marks head of queue lost.
993 * A''. Its FACK modification, head until snd.fack is lost.
994 * B. SACK arrives sacking SND.NXT at the moment, when the
995 * segment was retransmitted.
996 * 4. D-SACK added new rule: D-SACK changes any tag to S.
998 * It is pleasant to note, that state diagram turns out to be commutative,
999 * so that we are allowed not to be bothered by order of our actions,
1000 * when multiple events arrive simultaneously. (see the function below).
1002 * Reordering detection.
1003 * --------------------
1004 * Reordering metric is maximal distance, which a packet can be displaced
1005 * in packet stream. With SACKs we can estimate it:
1007 * 1. SACK fills old hole and the corresponding segment was not
1008 * ever retransmitted -> reordering. Alas, we cannot use it
1009 * when segment was retransmitted.
1010 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1011 * for retransmitted and already SACKed segment -> reordering..
1012 * Both of these heuristics are not used in Loss state, when we cannot
1013 * account for retransmits accurately.
1015 * SACK block validation.
1016 * ----------------------
1018 * SACK block range validation checks that the received SACK block fits to
1019 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1020 * Note that SND.UNA is not included to the range though being valid because
1021 * it means that the receiver is rather inconsistent with itself reporting
1022 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1023 * perfectly valid, however, in light of RFC2018 which explicitly states
1024 * that "SACK block MUST reflect the newest segment. Even if the newest
1025 * segment is going to be discarded ...", not that it looks very clever
1026 * in case of head skb. Due to potentional receiver driven attacks, we
1027 * choose to avoid immediate execution of a walk in write queue due to
1028 * reneging and defer head skb's loss recovery to standard loss recovery
1029 * procedure that will eventually trigger (nothing forbids us doing this).
1031 * Implements also blockage to start_seq wrap-around. Problem lies in the
1032 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1033 * there's no guarantee that it will be before snd_nxt (n). The problem
1034 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1037 * <- outs wnd -> <- wrapzone ->
1038 * u e n u_w e_w s n_w
1040 * |<------------+------+----- TCP seqno space --------------+---------->|
1041 * ...-- <2^31 ->| |<--------...
1042 * ...---- >2^31 ------>| |<--------...
1044 * Current code wouldn't be vulnerable but it's better still to discard such
1045 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1046 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1047 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1048 * equal to the ideal case (infinite seqno space without wrap caused issues).
1050 * With D-SACK the lower bound is extended to cover sequence space below
1051 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1052 * again, D-SACK block must not to go across snd_una (for the same reason as
1053 * for the normal SACK blocks, explained above). But there all simplicity
1054 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1055 * fully below undo_marker they do not affect behavior in anyway and can
1056 * therefore be safely ignored. In rare cases (which are more or less
1057 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1058 * fragmentation and packet reordering past skb's retransmission. To consider
1059 * them correctly, the acceptable range must be extended even more though
1060 * the exact amount is rather hard to quantify. However, tp->max_window can
1061 * be used as an exaggerated estimate.
1063 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1064 u32 start_seq, u32 end_seq)
1066 /* Too far in future, or reversed (interpretation is ambiguous) */
1067 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1070 /* Nasty start_seq wrap-around check (see comments above) */
1071 if (!before(start_seq, tp->snd_nxt))
1074 /* In outstanding window? ...This is valid exit for D-SACKs too.
1075 * start_seq == snd_una is non-sensical (see comments above)
1077 if (after(start_seq, tp->snd_una))
1080 if (!is_dsack || !tp->undo_marker)
1083 /* ...Then it's D-SACK, and must reside below snd_una completely */
1084 if (after(end_seq, tp->snd_una))
1087 if (!before(start_seq, tp->undo_marker))
1091 if (!after(end_seq, tp->undo_marker))
1094 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1095 * start_seq < undo_marker and end_seq >= undo_marker.
1097 return !before(start_seq, end_seq - tp->max_window);
1100 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1101 struct tcp_sack_block_wire *sp, int num_sacks,
1104 struct tcp_sock *tp = tcp_sk(sk);
1105 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1106 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1107 bool dup_sack = false;
1109 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1112 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1113 } else if (num_sacks > 1) {
1114 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1115 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1117 if (!after(end_seq_0, end_seq_1) &&
1118 !before(start_seq_0, start_seq_1)) {
1121 NET_INC_STATS(sock_net(sk),
1122 LINUX_MIB_TCPDSACKOFORECV);
1126 /* D-SACK for already forgotten data... Do dumb counting. */
1127 if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 &&
1128 !after(end_seq_0, prior_snd_una) &&
1129 after(end_seq_0, tp->undo_marker))
1135 struct tcp_sacktag_state {
1138 /* Timestamps for earliest and latest never-retransmitted segment
1139 * that was SACKed. RTO needs the earliest RTT to stay conservative,
1140 * but congestion control should still get an accurate delay signal.
1144 struct rate_sample *rate;
1148 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1149 * the incoming SACK may not exactly match but we can find smaller MSS
1150 * aligned portion of it that matches. Therefore we might need to fragment
1151 * which may fail and creates some hassle (caller must handle error case
1154 * FIXME: this could be merged to shift decision code
1156 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1157 u32 start_seq, u32 end_seq)
1161 unsigned int pkt_len;
1164 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1165 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1167 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1168 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1169 mss = tcp_skb_mss(skb);
1170 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1173 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1177 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1182 /* Round if necessary so that SACKs cover only full MSSes
1183 * and/or the remaining small portion (if present)
1185 if (pkt_len > mss) {
1186 unsigned int new_len = (pkt_len / mss) * mss;
1187 if (!in_sack && new_len < pkt_len)
1192 if (pkt_len >= skb->len && !in_sack)
1195 err = tcp_fragment(sk, skb, pkt_len, mss, GFP_ATOMIC);
1203 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1204 static u8 tcp_sacktag_one(struct sock *sk,
1205 struct tcp_sacktag_state *state, u8 sacked,
1206 u32 start_seq, u32 end_seq,
1207 int dup_sack, int pcount,
1210 struct tcp_sock *tp = tcp_sk(sk);
1211 int fack_count = state->fack_count;
1213 /* Account D-SACK for retransmitted packet. */
1214 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1215 if (tp->undo_marker && tp->undo_retrans > 0 &&
1216 after(end_seq, tp->undo_marker))
1218 if (sacked & TCPCB_SACKED_ACKED)
1219 state->reord = min(fack_count, state->reord);
1222 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1223 if (!after(end_seq, tp->snd_una))
1226 if (!(sacked & TCPCB_SACKED_ACKED)) {
1227 tcp_rack_advance(tp, sacked, end_seq, xmit_time);
1229 if (sacked & TCPCB_SACKED_RETRANS) {
1230 /* If the segment is not tagged as lost,
1231 * we do not clear RETRANS, believing
1232 * that retransmission is still in flight.
1234 if (sacked & TCPCB_LOST) {
1235 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1236 tp->lost_out -= pcount;
1237 tp->retrans_out -= pcount;
1240 if (!(sacked & TCPCB_RETRANS)) {
1241 /* New sack for not retransmitted frame,
1242 * which was in hole. It is reordering.
1244 if (before(start_seq,
1245 tcp_highest_sack_seq(tp)))
1246 state->reord = min(fack_count,
1248 if (!after(end_seq, tp->high_seq))
1249 state->flag |= FLAG_ORIG_SACK_ACKED;
1250 if (state->first_sackt == 0)
1251 state->first_sackt = xmit_time;
1252 state->last_sackt = xmit_time;
1255 if (sacked & TCPCB_LOST) {
1256 sacked &= ~TCPCB_LOST;
1257 tp->lost_out -= pcount;
1261 sacked |= TCPCB_SACKED_ACKED;
1262 state->flag |= FLAG_DATA_SACKED;
1263 tp->sacked_out += pcount;
1264 tp->delivered += pcount; /* Out-of-order packets delivered */
1266 fack_count += pcount;
1268 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1269 if (!tcp_is_fack(tp) && tp->lost_skb_hint &&
1270 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1271 tp->lost_cnt_hint += pcount;
1273 if (fack_count > tp->fackets_out)
1274 tp->fackets_out = fack_count;
1277 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1278 * frames and clear it. undo_retrans is decreased above, L|R frames
1279 * are accounted above as well.
1281 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1282 sacked &= ~TCPCB_SACKED_RETRANS;
1283 tp->retrans_out -= pcount;
1289 /* Shift newly-SACKed bytes from this skb to the immediately previous
1290 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1292 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1293 struct tcp_sacktag_state *state,
1294 unsigned int pcount, int shifted, int mss,
1297 struct tcp_sock *tp = tcp_sk(sk);
1298 struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1299 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1300 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1304 /* Adjust counters and hints for the newly sacked sequence
1305 * range but discard the return value since prev is already
1306 * marked. We must tag the range first because the seq
1307 * advancement below implicitly advances
1308 * tcp_highest_sack_seq() when skb is highest_sack.
1310 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1311 start_seq, end_seq, dup_sack, pcount,
1313 tcp_rate_skb_delivered(sk, skb, state->rate);
1315 if (skb == tp->lost_skb_hint)
1316 tp->lost_cnt_hint += pcount;
1318 TCP_SKB_CB(prev)->end_seq += shifted;
1319 TCP_SKB_CB(skb)->seq += shifted;
1321 tcp_skb_pcount_add(prev, pcount);
1322 BUG_ON(tcp_skb_pcount(skb) < pcount);
1323 tcp_skb_pcount_add(skb, -pcount);
1325 /* When we're adding to gso_segs == 1, gso_size will be zero,
1326 * in theory this shouldn't be necessary but as long as DSACK
1327 * code can come after this skb later on it's better to keep
1328 * setting gso_size to something.
1330 if (!TCP_SKB_CB(prev)->tcp_gso_size)
1331 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1333 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1334 if (tcp_skb_pcount(skb) <= 1)
1335 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1337 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1338 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1341 BUG_ON(!tcp_skb_pcount(skb));
1342 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1346 /* Whole SKB was eaten :-) */
1348 if (skb == tp->retransmit_skb_hint)
1349 tp->retransmit_skb_hint = prev;
1350 if (skb == tp->lost_skb_hint) {
1351 tp->lost_skb_hint = prev;
1352 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1355 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1356 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor;
1357 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1358 TCP_SKB_CB(prev)->end_seq++;
1360 if (skb == tcp_highest_sack(sk))
1361 tcp_advance_highest_sack(sk, skb);
1363 tcp_skb_collapse_tstamp(prev, skb);
1364 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp))
1365 TCP_SKB_CB(prev)->tx.delivered_mstamp = 0;
1367 tcp_unlink_write_queue(skb, sk);
1368 sk_wmem_free_skb(sk, skb);
1370 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED);
1375 /* I wish gso_size would have a bit more sane initialization than
1376 * something-or-zero which complicates things
1378 static int tcp_skb_seglen(const struct sk_buff *skb)
1380 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1383 /* Shifting pages past head area doesn't work */
1384 static int skb_can_shift(const struct sk_buff *skb)
1386 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1389 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1392 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1393 struct tcp_sacktag_state *state,
1394 u32 start_seq, u32 end_seq,
1397 struct tcp_sock *tp = tcp_sk(sk);
1398 struct sk_buff *prev;
1404 if (!sk_can_gso(sk))
1407 /* Normally R but no L won't result in plain S */
1409 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1411 if (!skb_can_shift(skb))
1413 /* This frame is about to be dropped (was ACKed). */
1414 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1417 /* Can only happen with delayed DSACK + discard craziness */
1418 if (unlikely(skb == tcp_write_queue_head(sk)))
1420 prev = tcp_write_queue_prev(sk, skb);
1422 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1425 if (!tcp_skb_can_collapse_to(prev))
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(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),
1596 tcp_rate_skb_delivered(sk, skb, state->rate);
1598 if (!before(TCP_SKB_CB(skb)->seq,
1599 tcp_highest_sack_seq(tp)))
1600 tcp_advance_highest_sack(sk, skb);
1603 state->fack_count += tcp_skb_pcount(skb);
1608 /* Avoid all extra work that is being done by sacktag while walking in
1611 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1612 struct tcp_sacktag_state *state,
1615 tcp_for_write_queue_from(skb, sk) {
1616 if (skb == tcp_send_head(sk))
1619 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1622 state->fack_count += tcp_skb_pcount(skb);
1627 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1629 struct tcp_sack_block *next_dup,
1630 struct tcp_sacktag_state *state,
1636 if (before(next_dup->start_seq, skip_to_seq)) {
1637 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1638 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1639 next_dup->start_seq, next_dup->end_seq,
1646 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1648 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1652 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1653 u32 prior_snd_una, struct tcp_sacktag_state *state)
1655 struct tcp_sock *tp = tcp_sk(sk);
1656 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1657 TCP_SKB_CB(ack_skb)->sacked);
1658 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1659 struct tcp_sack_block sp[TCP_NUM_SACKS];
1660 struct tcp_sack_block *cache;
1661 struct sk_buff *skb;
1662 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1664 bool found_dup_sack = false;
1666 int first_sack_index;
1669 state->reord = tp->packets_out;
1671 if (!tp->sacked_out) {
1672 if (WARN_ON(tp->fackets_out))
1673 tp->fackets_out = 0;
1674 tcp_highest_sack_reset(sk);
1677 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1678 num_sacks, prior_snd_una);
1679 if (found_dup_sack) {
1680 state->flag |= FLAG_DSACKING_ACK;
1681 tp->delivered++; /* A spurious retransmission is delivered */
1684 /* Eliminate too old ACKs, but take into
1685 * account more or less fresh ones, they can
1686 * contain valid SACK info.
1688 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1691 if (!tp->packets_out)
1695 first_sack_index = 0;
1696 for (i = 0; i < num_sacks; i++) {
1697 bool dup_sack = !i && found_dup_sack;
1699 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1700 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1702 if (!tcp_is_sackblock_valid(tp, dup_sack,
1703 sp[used_sacks].start_seq,
1704 sp[used_sacks].end_seq)) {
1708 if (!tp->undo_marker)
1709 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1711 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1713 /* Don't count olds caused by ACK reordering */
1714 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1715 !after(sp[used_sacks].end_seq, tp->snd_una))
1717 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1720 NET_INC_STATS(sock_net(sk), mib_idx);
1722 first_sack_index = -1;
1726 /* Ignore very old stuff early */
1727 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1733 /* order SACK blocks to allow in order walk of the retrans queue */
1734 for (i = used_sacks - 1; i > 0; i--) {
1735 for (j = 0; j < i; j++) {
1736 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1737 swap(sp[j], sp[j + 1]);
1739 /* Track where the first SACK block goes to */
1740 if (j == first_sack_index)
1741 first_sack_index = j + 1;
1746 skb = tcp_write_queue_head(sk);
1747 state->fack_count = 0;
1750 if (!tp->sacked_out) {
1751 /* It's already past, so skip checking against it */
1752 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1754 cache = tp->recv_sack_cache;
1755 /* Skip empty blocks in at head of the cache */
1756 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1761 while (i < used_sacks) {
1762 u32 start_seq = sp[i].start_seq;
1763 u32 end_seq = sp[i].end_seq;
1764 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1765 struct tcp_sack_block *next_dup = NULL;
1767 if (found_dup_sack && ((i + 1) == first_sack_index))
1768 next_dup = &sp[i + 1];
1770 /* Skip too early cached blocks */
1771 while (tcp_sack_cache_ok(tp, cache) &&
1772 !before(start_seq, cache->end_seq))
1775 /* Can skip some work by looking recv_sack_cache? */
1776 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1777 after(end_seq, cache->start_seq)) {
1780 if (before(start_seq, cache->start_seq)) {
1781 skb = tcp_sacktag_skip(skb, sk, state,
1783 skb = tcp_sacktag_walk(skb, sk, next_dup,
1790 /* Rest of the block already fully processed? */
1791 if (!after(end_seq, cache->end_seq))
1794 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1798 /* ...tail remains todo... */
1799 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1800 /* ...but better entrypoint exists! */
1801 skb = tcp_highest_sack(sk);
1804 state->fack_count = tp->fackets_out;
1809 skb = tcp_sacktag_skip(skb, sk, state, cache->end_seq);
1810 /* Check overlap against next cached too (past this one already) */
1815 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1816 skb = tcp_highest_sack(sk);
1819 state->fack_count = tp->fackets_out;
1821 skb = tcp_sacktag_skip(skb, sk, state, start_seq);
1824 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
1825 start_seq, end_seq, dup_sack);
1831 /* Clear the head of the cache sack blocks so we can skip it next time */
1832 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1833 tp->recv_sack_cache[i].start_seq = 0;
1834 tp->recv_sack_cache[i].end_seq = 0;
1836 for (j = 0; j < used_sacks; j++)
1837 tp->recv_sack_cache[i++] = sp[j];
1839 if ((state->reord < tp->fackets_out) &&
1840 ((inet_csk(sk)->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker))
1841 tcp_update_reordering(sk, tp->fackets_out - state->reord, 0);
1843 tcp_verify_left_out(tp);
1846 #if FASTRETRANS_DEBUG > 0
1847 WARN_ON((int)tp->sacked_out < 0);
1848 WARN_ON((int)tp->lost_out < 0);
1849 WARN_ON((int)tp->retrans_out < 0);
1850 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1855 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1856 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1858 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1862 holes = max(tp->lost_out, 1U);
1863 holes = min(holes, tp->packets_out);
1865 if ((tp->sacked_out + holes) > tp->packets_out) {
1866 tp->sacked_out = tp->packets_out - holes;
1872 /* If we receive more dupacks than we expected counting segments
1873 * in assumption of absent reordering, interpret this as reordering.
1874 * The only another reason could be bug in receiver TCP.
1876 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1878 struct tcp_sock *tp = tcp_sk(sk);
1879 if (tcp_limit_reno_sacked(tp))
1880 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1883 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1885 static void tcp_add_reno_sack(struct sock *sk)
1887 struct tcp_sock *tp = tcp_sk(sk);
1888 u32 prior_sacked = tp->sacked_out;
1891 tcp_check_reno_reordering(sk, 0);
1892 if (tp->sacked_out > prior_sacked)
1893 tp->delivered++; /* Some out-of-order packet is delivered */
1894 tcp_verify_left_out(tp);
1897 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1899 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1901 struct tcp_sock *tp = tcp_sk(sk);
1904 /* One ACK acked hole. The rest eat duplicate ACKs. */
1905 tp->delivered += max_t(int, acked - tp->sacked_out, 1);
1906 if (acked - 1 >= tp->sacked_out)
1909 tp->sacked_out -= acked - 1;
1911 tcp_check_reno_reordering(sk, acked);
1912 tcp_verify_left_out(tp);
1915 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1920 void tcp_clear_retrans(struct tcp_sock *tp)
1922 tp->retrans_out = 0;
1924 tp->undo_marker = 0;
1925 tp->undo_retrans = -1;
1926 tp->fackets_out = 0;
1930 static inline void tcp_init_undo(struct tcp_sock *tp)
1932 tp->undo_marker = tp->snd_una;
1933 /* Retransmission still in flight may cause DSACKs later. */
1934 tp->undo_retrans = tp->retrans_out ? : -1;
1937 /* Enter Loss state. If we detect SACK reneging, forget all SACK information
1938 * and reset tags completely, otherwise preserve SACKs. If receiver
1939 * dropped its ofo queue, we will know this due to reneging detection.
1941 void tcp_enter_loss(struct sock *sk)
1943 const struct inet_connection_sock *icsk = inet_csk(sk);
1944 struct tcp_sock *tp = tcp_sk(sk);
1945 struct net *net = sock_net(sk);
1946 struct sk_buff *skb;
1947 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
1948 bool is_reneg; /* is receiver reneging on SACKs? */
1951 /* Reduce ssthresh if it has not yet been made inside this window. */
1952 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1953 !after(tp->high_seq, tp->snd_una) ||
1954 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1955 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1956 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1957 tcp_ca_event(sk, CA_EVENT_LOSS);
1961 tp->snd_cwnd_cnt = 0;
1962 tp->snd_cwnd_stamp = tcp_jiffies32;
1964 tp->retrans_out = 0;
1967 if (tcp_is_reno(tp))
1968 tcp_reset_reno_sack(tp);
1970 skb = tcp_write_queue_head(sk);
1971 is_reneg = skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED);
1973 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
1975 tp->fackets_out = 0;
1977 tcp_clear_all_retrans_hints(tp);
1979 tcp_for_write_queue(skb, sk) {
1980 if (skb == tcp_send_head(sk))
1983 mark_lost = (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
1986 tcp_sum_lost(tp, skb);
1987 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1989 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1990 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1991 tp->lost_out += tcp_skb_pcount(skb);
1994 tcp_verify_left_out(tp);
1996 /* Timeout in disordered state after receiving substantial DUPACKs
1997 * suggests that the degree of reordering is over-estimated.
1999 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
2000 tp->sacked_out >= net->ipv4.sysctl_tcp_reordering)
2001 tp->reordering = min_t(unsigned int, tp->reordering,
2002 net->ipv4.sysctl_tcp_reordering);
2003 tcp_set_ca_state(sk, TCP_CA_Loss);
2004 tp->high_seq = tp->snd_nxt;
2005 tcp_ecn_queue_cwr(tp);
2007 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
2008 * loss recovery is underway except recurring timeout(s) on
2009 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
2011 * In theory F-RTO can be used repeatedly during loss recovery.
2012 * In practice this interacts badly with broken middle-boxes that
2013 * falsely raise the receive window, which results in repeated
2014 * timeouts and stop-and-go behavior.
2016 tp->frto = sysctl_tcp_frto &&
2017 (new_recovery || icsk->icsk_retransmits) &&
2018 !inet_csk(sk)->icsk_mtup.probe_size;
2021 /* If ACK arrived pointing to a remembered SACK, it means that our
2022 * remembered SACKs do not reflect real state of receiver i.e.
2023 * receiver _host_ is heavily congested (or buggy).
2025 * To avoid big spurious retransmission bursts due to transient SACK
2026 * scoreboard oddities that look like reneging, we give the receiver a
2027 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2028 * restore sanity to the SACK scoreboard. If the apparent reneging
2029 * persists until this RTO then we'll clear the SACK scoreboard.
2031 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
2033 if (flag & FLAG_SACK_RENEGING) {
2034 struct tcp_sock *tp = tcp_sk(sk);
2035 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2036 msecs_to_jiffies(10));
2038 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2039 delay, TCP_RTO_MAX);
2045 static inline int tcp_fackets_out(const struct tcp_sock *tp)
2047 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
2050 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2051 * counter when SACK is enabled (without SACK, sacked_out is used for
2054 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2055 * segments up to the highest received SACK block so far and holes in
2058 * With reordering, holes may still be in flight, so RFC3517 recovery
2059 * uses pure sacked_out (total number of SACKed segments) even though
2060 * it violates the RFC that uses duplicate ACKs, often these are equal
2061 * but when e.g. out-of-window ACKs or packet duplication occurs,
2062 * they differ. Since neither occurs due to loss, TCP should really
2065 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2067 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2070 /* Linux NewReno/SACK/FACK/ECN state machine.
2071 * --------------------------------------
2073 * "Open" Normal state, no dubious events, fast path.
2074 * "Disorder" In all the respects it is "Open",
2075 * but requires a bit more attention. It is entered when
2076 * we see some SACKs or dupacks. It is split of "Open"
2077 * mainly to move some processing from fast path to slow one.
2078 * "CWR" CWND was reduced due to some Congestion Notification event.
2079 * It can be ECN, ICMP source quench, local device congestion.
2080 * "Recovery" CWND was reduced, we are fast-retransmitting.
2081 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2083 * tcp_fastretrans_alert() is entered:
2084 * - each incoming ACK, if state is not "Open"
2085 * - when arrived ACK is unusual, namely:
2090 * Counting packets in flight is pretty simple.
2092 * in_flight = packets_out - left_out + retrans_out
2094 * packets_out is SND.NXT-SND.UNA counted in packets.
2096 * retrans_out is number of retransmitted segments.
2098 * left_out is number of segments left network, but not ACKed yet.
2100 * left_out = sacked_out + lost_out
2102 * sacked_out: Packets, which arrived to receiver out of order
2103 * and hence not ACKed. With SACKs this number is simply
2104 * amount of SACKed data. Even without SACKs
2105 * it is easy to give pretty reliable estimate of this number,
2106 * counting duplicate ACKs.
2108 * lost_out: Packets lost by network. TCP has no explicit
2109 * "loss notification" feedback from network (for now).
2110 * It means that this number can be only _guessed_.
2111 * Actually, it is the heuristics to predict lossage that
2112 * distinguishes different algorithms.
2114 * F.e. after RTO, when all the queue is considered as lost,
2115 * lost_out = packets_out and in_flight = retrans_out.
2117 * Essentially, we have now a few algorithms detecting
2120 * If the receiver supports SACK:
2122 * RFC6675/3517: It is the conventional algorithm. A packet is
2123 * considered lost if the number of higher sequence packets
2124 * SACKed is greater than or equal the DUPACK thoreshold
2125 * (reordering). This is implemented in tcp_mark_head_lost and
2126 * tcp_update_scoreboard.
2128 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm
2129 * (2017-) that checks timing instead of counting DUPACKs.
2130 * Essentially a packet is considered lost if it's not S/ACKed
2131 * after RTT + reordering_window, where both metrics are
2132 * dynamically measured and adjusted. This is implemented in
2133 * tcp_rack_mark_lost.
2135 * FACK (Disabled by default. Subsumbed by RACK):
2136 * It is the simplest heuristics. As soon as we decided
2137 * that something is lost, we decide that _all_ not SACKed
2138 * packets until the most forward SACK are lost. I.e.
2139 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2140 * It is absolutely correct estimate, if network does not reorder
2141 * packets. And it loses any connection to reality when reordering
2142 * takes place. We use FACK by default until reordering
2143 * is suspected on the path to this destination.
2145 * If the receiver does not support SACK:
2147 * NewReno (RFC6582): in Recovery we assume that one segment
2148 * is lost (classic Reno). While we are in Recovery and
2149 * a partial ACK arrives, we assume that one more packet
2150 * is lost (NewReno). This heuristics are the same in NewReno
2153 * Really tricky (and requiring careful tuning) part of algorithm
2154 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2155 * The first determines the moment _when_ we should reduce CWND and,
2156 * hence, slow down forward transmission. In fact, it determines the moment
2157 * when we decide that hole is caused by loss, rather than by a reorder.
2159 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2160 * holes, caused by lost packets.
2162 * And the most logically complicated part of algorithm is undo
2163 * heuristics. We detect false retransmits due to both too early
2164 * fast retransmit (reordering) and underestimated RTO, analyzing
2165 * timestamps and D-SACKs. When we detect that some segments were
2166 * retransmitted by mistake and CWND reduction was wrong, we undo
2167 * window reduction and abort recovery phase. This logic is hidden
2168 * inside several functions named tcp_try_undo_<something>.
2171 /* This function decides, when we should leave Disordered state
2172 * and enter Recovery phase, reducing congestion window.
2174 * Main question: may we further continue forward transmission
2175 * with the same cwnd?
2177 static bool tcp_time_to_recover(struct sock *sk, int flag)
2179 struct tcp_sock *tp = tcp_sk(sk);
2181 /* Trick#1: The loss is proven. */
2185 /* Not-A-Trick#2 : Classic rule... */
2186 if (tcp_dupack_heuristics(tp) > tp->reordering)
2192 /* Detect loss in event "A" above by marking head of queue up as lost.
2193 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2194 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2195 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2196 * the maximum SACKed segments to pass before reaching this limit.
2198 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2200 struct tcp_sock *tp = tcp_sk(sk);
2201 struct sk_buff *skb;
2202 int cnt, oldcnt, lost;
2204 /* Use SACK to deduce losses of new sequences sent during recovery */
2205 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq;
2207 WARN_ON(packets > tp->packets_out);
2208 if (tp->lost_skb_hint) {
2209 skb = tp->lost_skb_hint;
2210 cnt = tp->lost_cnt_hint;
2211 /* Head already handled? */
2212 if (mark_head && skb != tcp_write_queue_head(sk))
2215 skb = tcp_write_queue_head(sk);
2219 tcp_for_write_queue_from(skb, sk) {
2220 if (skb == tcp_send_head(sk))
2222 /* TODO: do this better */
2223 /* this is not the most efficient way to do this... */
2224 tp->lost_skb_hint = skb;
2225 tp->lost_cnt_hint = cnt;
2227 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2231 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2232 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2233 cnt += tcp_skb_pcount(skb);
2235 if (cnt > packets) {
2236 if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2237 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2238 (oldcnt >= packets))
2241 mss = tcp_skb_mss(skb);
2242 /* If needed, chop off the prefix to mark as lost. */
2243 lost = (packets - oldcnt) * mss;
2244 if (lost < skb->len &&
2245 tcp_fragment(sk, skb, lost, mss, GFP_ATOMIC) < 0)
2250 tcp_skb_mark_lost(tp, skb);
2255 tcp_verify_left_out(tp);
2258 /* Account newly detected lost packet(s) */
2260 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2262 struct tcp_sock *tp = tcp_sk(sk);
2264 if (tcp_is_reno(tp)) {
2265 tcp_mark_head_lost(sk, 1, 1);
2266 } else if (tcp_is_fack(tp)) {
2267 int lost = tp->fackets_out - tp->reordering;
2270 tcp_mark_head_lost(sk, lost, 0);
2272 int sacked_upto = tp->sacked_out - tp->reordering;
2273 if (sacked_upto >= 0)
2274 tcp_mark_head_lost(sk, sacked_upto, 0);
2275 else if (fast_rexmit)
2276 tcp_mark_head_lost(sk, 1, 1);
2280 static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2282 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2283 before(tp->rx_opt.rcv_tsecr, when);
2286 /* skb is spurious retransmitted if the returned timestamp echo
2287 * reply is prior to the skb transmission time
2289 static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2290 const struct sk_buff *skb)
2292 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2293 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp(skb));
2296 /* Nothing was retransmitted or returned timestamp is less
2297 * than timestamp of the first retransmission.
2299 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2301 return !tp->retrans_stamp ||
2302 tcp_tsopt_ecr_before(tp, tp->retrans_stamp);
2305 /* Undo procedures. */
2307 /* We can clear retrans_stamp when there are no retransmissions in the
2308 * window. It would seem that it is trivially available for us in
2309 * tp->retrans_out, however, that kind of assumptions doesn't consider
2310 * what will happen if errors occur when sending retransmission for the
2311 * second time. ...It could the that such segment has only
2312 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2313 * the head skb is enough except for some reneging corner cases that
2314 * are not worth the effort.
2316 * Main reason for all this complexity is the fact that connection dying
2317 * time now depends on the validity of the retrans_stamp, in particular,
2318 * that successive retransmissions of a segment must not advance
2319 * retrans_stamp under any conditions.
2321 static bool tcp_any_retrans_done(const struct sock *sk)
2323 const struct tcp_sock *tp = tcp_sk(sk);
2324 struct sk_buff *skb;
2326 if (tp->retrans_out)
2329 skb = tcp_write_queue_head(sk);
2330 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2336 #if FASTRETRANS_DEBUG > 1
2337 static void DBGUNDO(struct sock *sk, const char *msg)
2339 struct tcp_sock *tp = tcp_sk(sk);
2340 struct inet_sock *inet = inet_sk(sk);
2342 if (sk->sk_family == AF_INET) {
2343 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2345 &inet->inet_daddr, ntohs(inet->inet_dport),
2346 tp->snd_cwnd, tcp_left_out(tp),
2347 tp->snd_ssthresh, tp->prior_ssthresh,
2350 #if IS_ENABLED(CONFIG_IPV6)
2351 else if (sk->sk_family == AF_INET6) {
2352 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2354 &sk->sk_v6_daddr, ntohs(inet->inet_dport),
2355 tp->snd_cwnd, tcp_left_out(tp),
2356 tp->snd_ssthresh, tp->prior_ssthresh,
2362 #define DBGUNDO(x...) do { } while (0)
2365 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2367 struct tcp_sock *tp = tcp_sk(sk);
2370 struct sk_buff *skb;
2372 tcp_for_write_queue(skb, sk) {
2373 if (skb == tcp_send_head(sk))
2375 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2378 tcp_clear_all_retrans_hints(tp);
2381 if (tp->prior_ssthresh) {
2382 const struct inet_connection_sock *icsk = inet_csk(sk);
2384 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2386 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2387 tp->snd_ssthresh = tp->prior_ssthresh;
2388 tcp_ecn_withdraw_cwr(tp);
2391 tp->snd_cwnd_stamp = tcp_jiffies32;
2392 tp->undo_marker = 0;
2395 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2397 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2400 /* People celebrate: "We love our President!" */
2401 static bool tcp_try_undo_recovery(struct sock *sk)
2403 struct tcp_sock *tp = tcp_sk(sk);
2405 if (tcp_may_undo(tp)) {
2408 /* Happy end! We did not retransmit anything
2409 * or our original transmission succeeded.
2411 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2412 tcp_undo_cwnd_reduction(sk, false);
2413 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2414 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2416 mib_idx = LINUX_MIB_TCPFULLUNDO;
2418 NET_INC_STATS(sock_net(sk), mib_idx);
2420 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2421 /* Hold old state until something *above* high_seq
2422 * is ACKed. For Reno it is MUST to prevent false
2423 * fast retransmits (RFC2582). SACK TCP is safe. */
2424 if (!tcp_any_retrans_done(sk))
2425 tp->retrans_stamp = 0;
2428 tcp_set_ca_state(sk, TCP_CA_Open);
2432 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2433 static bool tcp_try_undo_dsack(struct sock *sk)
2435 struct tcp_sock *tp = tcp_sk(sk);
2437 if (tp->undo_marker && !tp->undo_retrans) {
2438 DBGUNDO(sk, "D-SACK");
2439 tcp_undo_cwnd_reduction(sk, false);
2440 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2446 /* Undo during loss recovery after partial ACK or using F-RTO. */
2447 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2449 struct tcp_sock *tp = tcp_sk(sk);
2451 if (frto_undo || tcp_may_undo(tp)) {
2452 tcp_undo_cwnd_reduction(sk, true);
2454 DBGUNDO(sk, "partial loss");
2455 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2457 NET_INC_STATS(sock_net(sk),
2458 LINUX_MIB_TCPSPURIOUSRTOS);
2459 inet_csk(sk)->icsk_retransmits = 0;
2460 if (frto_undo || tcp_is_sack(tp))
2461 tcp_set_ca_state(sk, TCP_CA_Open);
2467 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2468 * It computes the number of packets to send (sndcnt) based on packets newly
2470 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2471 * cwnd reductions across a full RTT.
2472 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2473 * But when the retransmits are acked without further losses, PRR
2474 * slow starts cwnd up to ssthresh to speed up the recovery.
2476 static void tcp_init_cwnd_reduction(struct sock *sk)
2478 struct tcp_sock *tp = tcp_sk(sk);
2480 tp->high_seq = tp->snd_nxt;
2481 tp->tlp_high_seq = 0;
2482 tp->snd_cwnd_cnt = 0;
2483 tp->prior_cwnd = tp->snd_cwnd;
2484 tp->prr_delivered = 0;
2486 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2487 tcp_ecn_queue_cwr(tp);
2490 void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int flag)
2492 struct tcp_sock *tp = tcp_sk(sk);
2494 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2496 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2499 tp->prr_delivered += newly_acked_sacked;
2501 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2503 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2504 } else if ((flag & FLAG_RETRANS_DATA_ACKED) &&
2505 !(flag & FLAG_LOST_RETRANS)) {
2506 sndcnt = min_t(int, delta,
2507 max_t(int, tp->prr_delivered - tp->prr_out,
2508 newly_acked_sacked) + 1);
2510 sndcnt = min(delta, newly_acked_sacked);
2512 /* Force a fast retransmit upon entering fast recovery */
2513 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
2514 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2517 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2519 struct tcp_sock *tp = tcp_sk(sk);
2521 if (inet_csk(sk)->icsk_ca_ops->cong_control)
2524 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2525 if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR ||
2526 (tp->undo_marker && tp->snd_ssthresh < TCP_INFINITE_SSTHRESH)) {
2527 tp->snd_cwnd = tp->snd_ssthresh;
2528 tp->snd_cwnd_stamp = tcp_jiffies32;
2530 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2533 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2534 void tcp_enter_cwr(struct sock *sk)
2536 struct tcp_sock *tp = tcp_sk(sk);
2538 tp->prior_ssthresh = 0;
2539 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2540 tp->undo_marker = 0;
2541 tcp_init_cwnd_reduction(sk);
2542 tcp_set_ca_state(sk, TCP_CA_CWR);
2545 EXPORT_SYMBOL(tcp_enter_cwr);
2547 static void tcp_try_keep_open(struct sock *sk)
2549 struct tcp_sock *tp = tcp_sk(sk);
2550 int state = TCP_CA_Open;
2552 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2553 state = TCP_CA_Disorder;
2555 if (inet_csk(sk)->icsk_ca_state != state) {
2556 tcp_set_ca_state(sk, state);
2557 tp->high_seq = tp->snd_nxt;
2561 static void tcp_try_to_open(struct sock *sk, int flag)
2563 struct tcp_sock *tp = tcp_sk(sk);
2565 tcp_verify_left_out(tp);
2567 if (!tcp_any_retrans_done(sk))
2568 tp->retrans_stamp = 0;
2570 if (flag & FLAG_ECE)
2573 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2574 tcp_try_keep_open(sk);
2578 static void tcp_mtup_probe_failed(struct sock *sk)
2580 struct inet_connection_sock *icsk = inet_csk(sk);
2582 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2583 icsk->icsk_mtup.probe_size = 0;
2584 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
2587 static void tcp_mtup_probe_success(struct sock *sk)
2589 struct tcp_sock *tp = tcp_sk(sk);
2590 struct inet_connection_sock *icsk = inet_csk(sk);
2592 /* FIXME: breaks with very large cwnd */
2593 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2594 tp->snd_cwnd = tp->snd_cwnd *
2595 tcp_mss_to_mtu(sk, tp->mss_cache) /
2596 icsk->icsk_mtup.probe_size;
2597 tp->snd_cwnd_cnt = 0;
2598 tp->snd_cwnd_stamp = tcp_jiffies32;
2599 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2601 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2602 icsk->icsk_mtup.probe_size = 0;
2603 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2604 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
2607 /* Do a simple retransmit without using the backoff mechanisms in
2608 * tcp_timer. This is used for path mtu discovery.
2609 * The socket is already locked here.
2611 void tcp_simple_retransmit(struct sock *sk)
2613 const struct inet_connection_sock *icsk = inet_csk(sk);
2614 struct tcp_sock *tp = tcp_sk(sk);
2615 struct sk_buff *skb;
2616 unsigned int mss = tcp_current_mss(sk);
2617 u32 prior_lost = tp->lost_out;
2619 tcp_for_write_queue(skb, sk) {
2620 if (skb == tcp_send_head(sk))
2622 if (tcp_skb_seglen(skb) > mss &&
2623 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2624 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2625 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2626 tp->retrans_out -= tcp_skb_pcount(skb);
2628 tcp_skb_mark_lost_uncond_verify(tp, skb);
2632 tcp_clear_retrans_hints_partial(tp);
2634 if (prior_lost == tp->lost_out)
2637 if (tcp_is_reno(tp))
2638 tcp_limit_reno_sacked(tp);
2640 tcp_verify_left_out(tp);
2642 /* Don't muck with the congestion window here.
2643 * Reason is that we do not increase amount of _data_
2644 * in network, but units changed and effective
2645 * cwnd/ssthresh really reduced now.
2647 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2648 tp->high_seq = tp->snd_nxt;
2649 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2650 tp->prior_ssthresh = 0;
2651 tp->undo_marker = 0;
2652 tcp_set_ca_state(sk, TCP_CA_Loss);
2654 tcp_xmit_retransmit_queue(sk);
2656 EXPORT_SYMBOL(tcp_simple_retransmit);
2658 void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2660 struct tcp_sock *tp = tcp_sk(sk);
2663 if (tcp_is_reno(tp))
2664 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2666 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2668 NET_INC_STATS(sock_net(sk), mib_idx);
2670 tp->prior_ssthresh = 0;
2673 if (!tcp_in_cwnd_reduction(sk)) {
2675 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2676 tcp_init_cwnd_reduction(sk);
2678 tcp_set_ca_state(sk, TCP_CA_Recovery);
2681 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2682 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2684 static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack,
2687 struct tcp_sock *tp = tcp_sk(sk);
2688 bool recovered = !before(tp->snd_una, tp->high_seq);
2690 if ((flag & FLAG_SND_UNA_ADVANCED) &&
2691 tcp_try_undo_loss(sk, false))
2694 /* The ACK (s)acks some never-retransmitted data meaning not all
2695 * the data packets before the timeout were lost. Therefore we
2696 * undo the congestion window and state. This is essentially
2697 * the operation in F-RTO (RFC5682 section 3.1 step 3.b). Since
2698 * a retransmitted skb is permantly marked, we can apply such an
2699 * operation even if F-RTO was not used.
2701 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2702 tcp_try_undo_loss(sk, tp->undo_marker))
2705 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2706 if (after(tp->snd_nxt, tp->high_seq)) {
2707 if (flag & FLAG_DATA_SACKED || is_dupack)
2708 tp->frto = 0; /* Step 3.a. loss was real */
2709 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2710 tp->high_seq = tp->snd_nxt;
2711 /* Step 2.b. Try send new data (but deferred until cwnd
2712 * is updated in tcp_ack()). Otherwise fall back to
2713 * the conventional recovery.
2715 if (tcp_send_head(sk) &&
2716 after(tcp_wnd_end(tp), tp->snd_nxt)) {
2717 *rexmit = REXMIT_NEW;
2725 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2726 tcp_try_undo_recovery(sk);
2729 if (tcp_is_reno(tp)) {
2730 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2731 * delivered. Lower inflight to clock out (re)tranmissions.
2733 if (after(tp->snd_nxt, tp->high_seq) && is_dupack)
2734 tcp_add_reno_sack(sk);
2735 else if (flag & FLAG_SND_UNA_ADVANCED)
2736 tcp_reset_reno_sack(tp);
2738 *rexmit = REXMIT_LOST;
2741 /* Undo during fast recovery after partial ACK. */
2742 static bool tcp_try_undo_partial(struct sock *sk, const int acked)
2744 struct tcp_sock *tp = tcp_sk(sk);
2746 if (tp->undo_marker && tcp_packet_delayed(tp)) {
2747 /* Plain luck! Hole if filled with delayed
2748 * packet, rather than with a retransmit.
2750 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2752 /* We are getting evidence that the reordering degree is higher
2753 * than we realized. If there are no retransmits out then we
2754 * can undo. Otherwise we clock out new packets but do not
2755 * mark more packets lost or retransmit more.
2757 if (tp->retrans_out)
2760 if (!tcp_any_retrans_done(sk))
2761 tp->retrans_stamp = 0;
2763 DBGUNDO(sk, "partial recovery");
2764 tcp_undo_cwnd_reduction(sk, true);
2765 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2766 tcp_try_keep_open(sk);
2772 static void tcp_rack_identify_loss(struct sock *sk, int *ack_flag)
2774 struct tcp_sock *tp = tcp_sk(sk);
2776 /* Use RACK to detect loss */
2777 if (sysctl_tcp_recovery & TCP_RACK_LOSS_DETECTION) {
2778 u32 prior_retrans = tp->retrans_out;
2780 tcp_rack_mark_lost(sk);
2781 if (prior_retrans > tp->retrans_out)
2782 *ack_flag |= FLAG_LOST_RETRANS;
2786 /* Process an event, which can update packets-in-flight not trivially.
2787 * Main goal of this function is to calculate new estimate for left_out,
2788 * taking into account both packets sitting in receiver's buffer and
2789 * packets lost by network.
2791 * Besides that it updates the congestion state when packet loss or ECN
2792 * is detected. But it does not reduce the cwnd, it is done by the
2793 * congestion control later.
2795 * It does _not_ decide what to send, it is made in function
2796 * tcp_xmit_retransmit_queue().
2798 static void tcp_fastretrans_alert(struct sock *sk, const int acked,
2799 bool is_dupack, int *ack_flag, int *rexmit)
2801 struct inet_connection_sock *icsk = inet_csk(sk);
2802 struct tcp_sock *tp = tcp_sk(sk);
2803 int fast_rexmit = 0, flag = *ack_flag;
2804 bool do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2805 (tcp_fackets_out(tp) > tp->reordering));
2807 if (WARN_ON(!tp->packets_out && tp->sacked_out))
2809 if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2810 tp->fackets_out = 0;
2812 /* Now state machine starts.
2813 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2814 if (flag & FLAG_ECE)
2815 tp->prior_ssthresh = 0;
2817 /* B. In all the states check for reneging SACKs. */
2818 if (tcp_check_sack_reneging(sk, flag))
2821 /* C. Check consistency of the current state. */
2822 tcp_verify_left_out(tp);
2824 /* D. Check state exit conditions. State can be terminated
2825 * when high_seq is ACKed. */
2826 if (icsk->icsk_ca_state == TCP_CA_Open) {
2827 WARN_ON(tp->retrans_out != 0);
2828 tp->retrans_stamp = 0;
2829 } else if (!before(tp->snd_una, tp->high_seq)) {
2830 switch (icsk->icsk_ca_state) {
2832 /* CWR is to be held something *above* high_seq
2833 * is ACKed for CWR bit to reach receiver. */
2834 if (tp->snd_una != tp->high_seq) {
2835 tcp_end_cwnd_reduction(sk);
2836 tcp_set_ca_state(sk, TCP_CA_Open);
2840 case TCP_CA_Recovery:
2841 if (tcp_is_reno(tp))
2842 tcp_reset_reno_sack(tp);
2843 if (tcp_try_undo_recovery(sk))
2845 tcp_end_cwnd_reduction(sk);
2850 /* E. Process state. */
2851 switch (icsk->icsk_ca_state) {
2852 case TCP_CA_Recovery:
2853 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2854 if (tcp_is_reno(tp) && is_dupack)
2855 tcp_add_reno_sack(sk);
2857 if (tcp_try_undo_partial(sk, acked))
2859 /* Partial ACK arrived. Force fast retransmit. */
2860 do_lost = tcp_is_reno(tp) ||
2861 tcp_fackets_out(tp) > tp->reordering;
2863 if (tcp_try_undo_dsack(sk)) {
2864 tcp_try_keep_open(sk);
2867 tcp_rack_identify_loss(sk, ack_flag);
2870 tcp_process_loss(sk, flag, is_dupack, rexmit);
2871 tcp_rack_identify_loss(sk, ack_flag);
2872 if (!(icsk->icsk_ca_state == TCP_CA_Open ||
2873 (*ack_flag & FLAG_LOST_RETRANS)))
2875 /* Change state if cwnd is undone or retransmits are lost */
2877 if (tcp_is_reno(tp)) {
2878 if (flag & FLAG_SND_UNA_ADVANCED)
2879 tcp_reset_reno_sack(tp);
2881 tcp_add_reno_sack(sk);
2884 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2885 tcp_try_undo_dsack(sk);
2887 tcp_rack_identify_loss(sk, ack_flag);
2888 if (!tcp_time_to_recover(sk, flag)) {
2889 tcp_try_to_open(sk, flag);
2893 /* MTU probe failure: don't reduce cwnd */
2894 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2895 icsk->icsk_mtup.probe_size &&
2896 tp->snd_una == tp->mtu_probe.probe_seq_start) {
2897 tcp_mtup_probe_failed(sk);
2898 /* Restores the reduction we did in tcp_mtup_probe() */
2900 tcp_simple_retransmit(sk);
2904 /* Otherwise enter Recovery state */
2905 tcp_enter_recovery(sk, (flag & FLAG_ECE));
2910 tcp_update_scoreboard(sk, fast_rexmit);
2911 *rexmit = REXMIT_LOST;
2914 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us)
2916 struct tcp_sock *tp = tcp_sk(sk);
2917 u32 wlen = sysctl_tcp_min_rtt_wlen * HZ;
2919 minmax_running_min(&tp->rtt_min, wlen, tcp_jiffies32,
2920 rtt_us ? : jiffies_to_usecs(1));
2923 static bool tcp_ack_update_rtt(struct sock *sk, const int flag,
2924 long seq_rtt_us, long sack_rtt_us,
2925 long ca_rtt_us, struct rate_sample *rs)
2927 const struct tcp_sock *tp = tcp_sk(sk);
2929 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2930 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2931 * Karn's algorithm forbids taking RTT if some retransmitted data
2932 * is acked (RFC6298).
2935 seq_rtt_us = sack_rtt_us;
2937 /* RTTM Rule: A TSecr value received in a segment is used to
2938 * update the averaged RTT measurement only if the segment
2939 * acknowledges some new data, i.e., only if it advances the
2940 * left edge of the send window.
2941 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2943 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2944 flag & FLAG_ACKED) {
2945 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
2946 u32 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
2948 seq_rtt_us = ca_rtt_us = delta_us;
2950 rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */
2954 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
2955 * always taken together with ACK, SACK, or TS-opts. Any negative
2956 * values will be skipped with the seq_rtt_us < 0 check above.
2958 tcp_update_rtt_min(sk, ca_rtt_us);
2959 tcp_rtt_estimator(sk, seq_rtt_us);
2962 /* RFC6298: only reset backoff on valid RTT measurement. */
2963 inet_csk(sk)->icsk_backoff = 0;
2967 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2968 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
2970 struct rate_sample rs;
2973 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack)
2974 rtt_us = tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req)->snt_synack);
2976 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us, &rs);
2980 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
2982 const struct inet_connection_sock *icsk = inet_csk(sk);
2984 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
2985 tcp_sk(sk)->snd_cwnd_stamp = tcp_jiffies32;
2988 /* Restart timer after forward progress on connection.
2989 * RFC2988 recommends to restart timer to now+rto.
2991 void tcp_rearm_rto(struct sock *sk)
2993 const struct inet_connection_sock *icsk = inet_csk(sk);
2994 struct tcp_sock *tp = tcp_sk(sk);
2996 /* If the retrans timer is currently being used by Fast Open
2997 * for SYN-ACK retrans purpose, stay put.
2999 if (tp->fastopen_rsk)
3002 if (!tp->packets_out) {
3003 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3005 u32 rto = inet_csk(sk)->icsk_rto;
3006 /* Offset the time elapsed after installing regular RTO */
3007 if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT ||
3008 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
3009 struct sk_buff *skb = tcp_write_queue_head(sk);
3010 u64 rto_time_stamp = skb->skb_mstamp +
3011 jiffies_to_usecs(rto);
3012 s64 delta_us = rto_time_stamp - tp->tcp_mstamp;
3013 /* delta_us may not be positive if the socket is locked
3014 * when the retrans timer fires and is rescheduled.
3017 rto = usecs_to_jiffies(delta_us);
3019 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
3024 /* If we get here, the whole TSO packet has not been acked. */
3025 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3027 struct tcp_sock *tp = tcp_sk(sk);
3030 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3032 packets_acked = tcp_skb_pcount(skb);
3033 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3035 packets_acked -= tcp_skb_pcount(skb);
3037 if (packets_acked) {
3038 BUG_ON(tcp_skb_pcount(skb) == 0);
3039 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3042 return packets_acked;
3045 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3048 const struct skb_shared_info *shinfo;
3050 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3051 if (likely(!TCP_SKB_CB(skb)->txstamp_ack))
3054 shinfo = skb_shinfo(skb);
3055 if (!before(shinfo->tskey, prior_snd_una) &&
3056 before(shinfo->tskey, tcp_sk(sk)->snd_una))
3057 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK);
3060 /* Remove acknowledged frames from the retransmission queue. If our packet
3061 * is before the ack sequence we can discard it as it's confirmed to have
3062 * arrived at the other end.
3064 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3065 u32 prior_snd_una, int *acked,
3066 struct tcp_sacktag_state *sack)
3068 const struct inet_connection_sock *icsk = inet_csk(sk);
3069 u64 first_ackt, last_ackt;
3070 struct tcp_sock *tp = tcp_sk(sk);
3071 u32 prior_sacked = tp->sacked_out;
3072 u32 reord = tp->packets_out;
3073 bool fully_acked = true;
3074 long sack_rtt_us = -1L;
3075 long seq_rtt_us = -1L;
3076 long ca_rtt_us = -1L;
3077 struct sk_buff *skb;
3079 u32 last_in_flight = 0;
3085 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3086 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3087 u8 sacked = scb->sacked;
3090 tcp_ack_tstamp(sk, skb, prior_snd_una);
3092 /* Determine how many packets and what bytes were acked, tso and else */
3093 if (after(scb->end_seq, tp->snd_una)) {
3094 if (tcp_skb_pcount(skb) == 1 ||
3095 !after(tp->snd_una, scb->seq))
3098 acked_pcount = tcp_tso_acked(sk, skb);
3101 fully_acked = false;
3103 /* Speedup tcp_unlink_write_queue() and next loop */
3104 prefetchw(skb->next);
3105 acked_pcount = tcp_skb_pcount(skb);
3108 if (unlikely(sacked & TCPCB_RETRANS)) {
3109 if (sacked & TCPCB_SACKED_RETRANS)
3110 tp->retrans_out -= acked_pcount;
3111 flag |= FLAG_RETRANS_DATA_ACKED;
3112 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3113 last_ackt = skb->skb_mstamp;
3114 WARN_ON_ONCE(last_ackt == 0);
3116 first_ackt = last_ackt;
3118 last_in_flight = TCP_SKB_CB(skb)->tx.in_flight;
3119 reord = min(pkts_acked, reord);
3120 if (!after(scb->end_seq, tp->high_seq))
3121 flag |= FLAG_ORIG_SACK_ACKED;
3124 if (sacked & TCPCB_SACKED_ACKED) {
3125 tp->sacked_out -= acked_pcount;
3126 } else if (tcp_is_sack(tp)) {
3127 tp->delivered += acked_pcount;
3128 if (!tcp_skb_spurious_retrans(tp, skb))
3129 tcp_rack_advance(tp, sacked, scb->end_seq,
3132 if (sacked & TCPCB_LOST)
3133 tp->lost_out -= acked_pcount;
3135 tp->packets_out -= acked_pcount;
3136 pkts_acked += acked_pcount;
3137 tcp_rate_skb_delivered(sk, skb, sack->rate);
3139 /* Initial outgoing SYN's get put onto the write_queue
3140 * just like anything else we transmit. It is not
3141 * true data, and if we misinform our callers that
3142 * this ACK acks real data, we will erroneously exit
3143 * connection startup slow start one packet too
3144 * quickly. This is severely frowned upon behavior.
3146 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3147 flag |= FLAG_DATA_ACKED;
3149 flag |= FLAG_SYN_ACKED;
3150 tp->retrans_stamp = 0;
3156 tcp_unlink_write_queue(skb, sk);
3157 sk_wmem_free_skb(sk, skb);
3158 if (unlikely(skb == tp->retransmit_skb_hint))
3159 tp->retransmit_skb_hint = NULL;
3160 if (unlikely(skb == tp->lost_skb_hint))
3161 tp->lost_skb_hint = NULL;
3165 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3167 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3168 tp->snd_up = tp->snd_una;
3170 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3171 flag |= FLAG_SACK_RENEGING;
3173 if (likely(first_ackt) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
3174 seq_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, first_ackt);
3175 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, last_ackt);
3177 if (sack->first_sackt) {
3178 sack_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->first_sackt);
3179 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->last_sackt);
3181 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
3182 ca_rtt_us, sack->rate);
3184 if (flag & FLAG_ACKED) {
3186 if (unlikely(icsk->icsk_mtup.probe_size &&
3187 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3188 tcp_mtup_probe_success(sk);
3191 if (tcp_is_reno(tp)) {
3192 tcp_remove_reno_sacks(sk, pkts_acked);
3196 /* Non-retransmitted hole got filled? That's reordering */
3197 if (reord < prior_fackets && reord <= tp->fackets_out)
3198 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3200 delta = tcp_is_fack(tp) ? pkts_acked :
3201 prior_sacked - tp->sacked_out;
3202 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3205 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3207 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3208 sack_rtt_us > tcp_stamp_us_delta(tp->tcp_mstamp, skb->skb_mstamp)) {
3209 /* Do not re-arm RTO if the sack RTT is measured from data sent
3210 * after when the head was last (re)transmitted. Otherwise the
3211 * timeout may continue to extend in loss recovery.
3216 if (icsk->icsk_ca_ops->pkts_acked) {
3217 struct ack_sample sample = { .pkts_acked = pkts_acked,
3218 .rtt_us = sack->rate->rtt_us,
3219 .in_flight = last_in_flight };
3221 icsk->icsk_ca_ops->pkts_acked(sk, &sample);
3224 #if FASTRETRANS_DEBUG > 0
3225 WARN_ON((int)tp->sacked_out < 0);
3226 WARN_ON((int)tp->lost_out < 0);
3227 WARN_ON((int)tp->retrans_out < 0);
3228 if (!tp->packets_out && tcp_is_sack(tp)) {
3229 icsk = inet_csk(sk);
3231 pr_debug("Leak l=%u %d\n",
3232 tp->lost_out, icsk->icsk_ca_state);
3235 if (tp->sacked_out) {
3236 pr_debug("Leak s=%u %d\n",
3237 tp->sacked_out, icsk->icsk_ca_state);
3240 if (tp->retrans_out) {
3241 pr_debug("Leak r=%u %d\n",
3242 tp->retrans_out, icsk->icsk_ca_state);
3243 tp->retrans_out = 0;
3247 *acked = pkts_acked;
3251 static void tcp_ack_probe(struct sock *sk)
3253 const struct tcp_sock *tp = tcp_sk(sk);
3254 struct inet_connection_sock *icsk = inet_csk(sk);
3256 /* Was it a usable window open? */
3258 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3259 icsk->icsk_backoff = 0;
3260 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3261 /* Socket must be waked up by subsequent tcp_data_snd_check().
3262 * This function is not for random using!
3265 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
3267 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3272 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3274 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3275 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3278 /* Decide wheather to run the increase function of congestion control. */
3279 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3281 /* If reordering is high then always grow cwnd whenever data is
3282 * delivered regardless of its ordering. Otherwise stay conservative
3283 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3284 * new SACK or ECE mark may first advance cwnd here and later reduce
3285 * cwnd in tcp_fastretrans_alert() based on more states.
3287 if (tcp_sk(sk)->reordering > sock_net(sk)->ipv4.sysctl_tcp_reordering)
3288 return flag & FLAG_FORWARD_PROGRESS;
3290 return flag & FLAG_DATA_ACKED;
3293 /* The "ultimate" congestion control function that aims to replace the rigid
3294 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3295 * It's called toward the end of processing an ACK with precise rate
3296 * information. All transmission or retransmission are delayed afterwards.
3298 static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
3299 int flag, const struct rate_sample *rs)
3301 const struct inet_connection_sock *icsk = inet_csk(sk);
3303 if (icsk->icsk_ca_ops->cong_control) {
3304 icsk->icsk_ca_ops->cong_control(sk, rs);
3308 if (tcp_in_cwnd_reduction(sk)) {
3309 /* Reduce cwnd if state mandates */
3310 tcp_cwnd_reduction(sk, acked_sacked, flag);
3311 } else if (tcp_may_raise_cwnd(sk, flag)) {
3312 /* Advance cwnd if state allows */
3313 tcp_cong_avoid(sk, ack, acked_sacked);
3315 tcp_update_pacing_rate(sk);
3318 /* Check that window update is acceptable.
3319 * The function assumes that snd_una<=ack<=snd_next.
3321 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3322 const u32 ack, const u32 ack_seq,
3325 return after(ack, tp->snd_una) ||
3326 after(ack_seq, tp->snd_wl1) ||
3327 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3330 /* If we update tp->snd_una, also update tp->bytes_acked */
3331 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3333 u32 delta = ack - tp->snd_una;
3335 sock_owned_by_me((struct sock *)tp);
3336 tp->bytes_acked += delta;
3340 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3341 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3343 u32 delta = seq - tp->rcv_nxt;
3345 sock_owned_by_me((struct sock *)tp);
3346 tp->bytes_received += delta;
3350 /* Update our send window.
3352 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3353 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3355 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3358 struct tcp_sock *tp = tcp_sk(sk);
3360 u32 nwin = ntohs(tcp_hdr(skb)->window);
3362 if (likely(!tcp_hdr(skb)->syn))
3363 nwin <<= tp->rx_opt.snd_wscale;
3365 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3366 flag |= FLAG_WIN_UPDATE;
3367 tcp_update_wl(tp, ack_seq);
3369 if (tp->snd_wnd != nwin) {
3372 /* Note, it is the only place, where
3373 * fast path is recovered for sending TCP.
3376 tcp_fast_path_check(sk);
3378 if (tcp_send_head(sk))
3379 tcp_slow_start_after_idle_check(sk);
3381 if (nwin > tp->max_window) {
3382 tp->max_window = nwin;
3383 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3388 tcp_snd_una_update(tp, ack);
3393 static bool __tcp_oow_rate_limited(struct net *net, int mib_idx,
3394 u32 *last_oow_ack_time)
3396 if (*last_oow_ack_time) {
3397 s32 elapsed = (s32)(tcp_jiffies32 - *last_oow_ack_time);
3399 if (0 <= elapsed && elapsed < sysctl_tcp_invalid_ratelimit) {
3400 NET_INC_STATS(net, mib_idx);
3401 return true; /* rate-limited: don't send yet! */
3405 *last_oow_ack_time = tcp_jiffies32;
3407 return false; /* not rate-limited: go ahead, send dupack now! */
3410 /* Return true if we're currently rate-limiting out-of-window ACKs and
3411 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3412 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3413 * attacks that send repeated SYNs or ACKs for the same connection. To
3414 * do this, we do not send a duplicate SYNACK or ACK if the remote
3415 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3417 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3418 int mib_idx, u32 *last_oow_ack_time)
3420 /* Data packets without SYNs are not likely part of an ACK loop. */
3421 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3425 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time);
3428 /* RFC 5961 7 [ACK Throttling] */
3429 static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb)
3431 /* unprotected vars, we dont care of overwrites */
3432 static u32 challenge_timestamp;
3433 static unsigned int challenge_count;
3434 struct tcp_sock *tp = tcp_sk(sk);
3437 /* First check our per-socket dupack rate limit. */
3438 if (__tcp_oow_rate_limited(sock_net(sk),
3439 LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3440 &tp->last_oow_ack_time))
3443 /* Then check host-wide RFC 5961 rate limit. */
3445 if (now != challenge_timestamp) {
3446 u32 half = (sysctl_tcp_challenge_ack_limit + 1) >> 1;
3448 challenge_timestamp = now;
3449 WRITE_ONCE(challenge_count, half +
3450 prandom_u32_max(sysctl_tcp_challenge_ack_limit));
3452 count = READ_ONCE(challenge_count);
3454 WRITE_ONCE(challenge_count, count - 1);
3455 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPCHALLENGEACK);
3460 static void tcp_store_ts_recent(struct tcp_sock *tp)
3462 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3463 tp->rx_opt.ts_recent_stamp = get_seconds();
3466 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3468 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3469 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3470 * extra check below makes sure this can only happen
3471 * for pure ACK frames. -DaveM
3473 * Not only, also it occurs for expired timestamps.
3476 if (tcp_paws_check(&tp->rx_opt, 0))
3477 tcp_store_ts_recent(tp);
3481 /* This routine deals with acks during a TLP episode.
3482 * We mark the end of a TLP episode on receiving TLP dupack or when
3483 * ack is after tlp_high_seq.
3484 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3486 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3488 struct tcp_sock *tp = tcp_sk(sk);
3490 if (before(ack, tp->tlp_high_seq))
3493 if (flag & FLAG_DSACKING_ACK) {
3494 /* This DSACK means original and TLP probe arrived; no loss */
3495 tp->tlp_high_seq = 0;
3496 } else if (after(ack, tp->tlp_high_seq)) {
3497 /* ACK advances: there was a loss, so reduce cwnd. Reset
3498 * tlp_high_seq in tcp_init_cwnd_reduction()
3500 tcp_init_cwnd_reduction(sk);
3501 tcp_set_ca_state(sk, TCP_CA_CWR);
3502 tcp_end_cwnd_reduction(sk);
3503 tcp_try_keep_open(sk);
3504 NET_INC_STATS(sock_net(sk),
3505 LINUX_MIB_TCPLOSSPROBERECOVERY);
3506 } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3507 FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3508 /* Pure dupack: original and TLP probe arrived; no loss */
3509 tp->tlp_high_seq = 0;
3513 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3515 const struct inet_connection_sock *icsk = inet_csk(sk);
3517 if (icsk->icsk_ca_ops->in_ack_event)
3518 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3521 /* Congestion control has updated the cwnd already. So if we're in
3522 * loss recovery then now we do any new sends (for FRTO) or
3523 * retransmits (for CA_Loss or CA_recovery) that make sense.
3525 static void tcp_xmit_recovery(struct sock *sk, int rexmit)
3527 struct tcp_sock *tp = tcp_sk(sk);
3529 if (rexmit == REXMIT_NONE)
3532 if (unlikely(rexmit == 2)) {
3533 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
3535 if (after(tp->snd_nxt, tp->high_seq))
3539 tcp_xmit_retransmit_queue(sk);
3542 /* This routine deals with incoming acks, but not outgoing ones. */
3543 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3545 struct inet_connection_sock *icsk = inet_csk(sk);
3546 struct tcp_sock *tp = tcp_sk(sk);
3547 struct tcp_sacktag_state sack_state;
3548 struct rate_sample rs = { .prior_delivered = 0 };
3549 u32 prior_snd_una = tp->snd_una;
3550 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3551 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3552 bool is_dupack = false;
3554 int prior_packets = tp->packets_out;
3555 u32 delivered = tp->delivered;
3556 u32 lost = tp->lost;
3557 int acked = 0; /* Number of packets newly acked */
3558 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
3560 sack_state.first_sackt = 0;
3561 sack_state.rate = &rs;
3563 /* We very likely will need to access write queue head. */
3564 prefetchw(sk->sk_write_queue.next);
3566 /* If the ack is older than previous acks
3567 * then we can probably ignore it.
3569 if (before(ack, prior_snd_una)) {
3570 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3571 if (before(ack, prior_snd_una - tp->max_window)) {
3572 if (!(flag & FLAG_NO_CHALLENGE_ACK))
3573 tcp_send_challenge_ack(sk, skb);
3579 /* If the ack includes data we haven't sent yet, discard
3580 * this segment (RFC793 Section 3.9).
3582 if (after(ack, tp->snd_nxt))
3585 if (icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
3588 if (after(ack, prior_snd_una)) {
3589 flag |= FLAG_SND_UNA_ADVANCED;
3590 icsk->icsk_retransmits = 0;
3593 prior_fackets = tp->fackets_out;
3594 rs.prior_in_flight = tcp_packets_in_flight(tp);
3596 /* ts_recent update must be made after we are sure that the packet
3599 if (flag & FLAG_UPDATE_TS_RECENT)
3600 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3602 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3603 /* Window is constant, pure forward advance.
3604 * No more checks are required.
3605 * Note, we use the fact that SND.UNA>=SND.WL2.
3607 tcp_update_wl(tp, ack_seq);
3608 tcp_snd_una_update(tp, ack);
3609 flag |= FLAG_WIN_UPDATE;
3611 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3613 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS);
3615 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3617 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3620 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3622 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3624 if (TCP_SKB_CB(skb)->sacked)
3625 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3628 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3630 ack_ev_flags |= CA_ACK_ECE;
3633 if (flag & FLAG_WIN_UPDATE)
3634 ack_ev_flags |= CA_ACK_WIN_UPDATE;
3636 tcp_in_ack_event(sk, ack_ev_flags);
3639 /* We passed data and got it acked, remove any soft error
3640 * log. Something worked...
3642 sk->sk_err_soft = 0;
3643 icsk->icsk_probes_out = 0;
3644 tp->rcv_tstamp = tcp_jiffies32;
3648 /* See if we can take anything off of the retransmit queue. */
3649 flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una, &acked,
3652 if (tcp_ack_is_dubious(sk, flag)) {
3653 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3654 tcp_fastretrans_alert(sk, acked, is_dupack, &flag, &rexmit);
3656 if (tp->tlp_high_seq)
3657 tcp_process_tlp_ack(sk, ack, flag);
3659 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3662 if (icsk->icsk_pending == ICSK_TIME_RETRANS)
3663 tcp_schedule_loss_probe(sk);
3664 delivered = tp->delivered - delivered; /* freshly ACKed or SACKed */
3665 lost = tp->lost - lost; /* freshly marked lost */
3666 tcp_rate_gen(sk, delivered, lost, sack_state.rate);
3667 tcp_cong_control(sk, ack, delivered, flag, sack_state.rate);
3668 tcp_xmit_recovery(sk, rexmit);
3672 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3673 if (flag & FLAG_DSACKING_ACK)
3674 tcp_fastretrans_alert(sk, acked, is_dupack, &flag, &rexmit);
3675 /* If this ack opens up a zero window, clear backoff. It was
3676 * being used to time the probes, and is probably far higher than
3677 * it needs to be for normal retransmission.
3679 if (tcp_send_head(sk))
3682 if (tp->tlp_high_seq)
3683 tcp_process_tlp_ack(sk, ack, flag);
3687 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3691 /* If data was SACKed, tag it and see if we should send more data.
3692 * If data was DSACKed, see if we can undo a cwnd reduction.
3694 if (TCP_SKB_CB(skb)->sacked) {
3695 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3697 tcp_fastretrans_alert(sk, acked, is_dupack, &flag, &rexmit);
3698 tcp_xmit_recovery(sk, rexmit);
3701 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3705 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3706 bool syn, struct tcp_fastopen_cookie *foc,
3709 /* Valid only in SYN or SYN-ACK with an even length. */
3710 if (!foc || !syn || len < 0 || (len & 1))
3713 if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3714 len <= TCP_FASTOPEN_COOKIE_MAX)
3715 memcpy(foc->val, cookie, len);
3722 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3723 * But, this can also be called on packets in the established flow when
3724 * the fast version below fails.
3726 void tcp_parse_options(const struct net *net,
3727 const struct sk_buff *skb,
3728 struct tcp_options_received *opt_rx, int estab,
3729 struct tcp_fastopen_cookie *foc)
3731 const unsigned char *ptr;
3732 const struct tcphdr *th = tcp_hdr(skb);
3733 int length = (th->doff * 4) - sizeof(struct tcphdr);
3735 ptr = (const unsigned char *)(th + 1);
3736 opt_rx->saw_tstamp = 0;
3738 while (length > 0) {
3739 int opcode = *ptr++;
3745 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3750 if (opsize < 2) /* "silly options" */
3752 if (opsize > length)
3753 return; /* don't parse partial options */
3756 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3757 u16 in_mss = get_unaligned_be16(ptr);
3759 if (opt_rx->user_mss &&
3760 opt_rx->user_mss < in_mss)
3761 in_mss = opt_rx->user_mss;
3762 opt_rx->mss_clamp = in_mss;
3767 if (opsize == TCPOLEN_WINDOW && th->syn &&
3768 !estab && sysctl_tcp_window_scaling) {
3769 __u8 snd_wscale = *(__u8 *)ptr;
3770 opt_rx->wscale_ok = 1;
3771 if (snd_wscale > TCP_MAX_WSCALE) {
3772 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n",
3776 snd_wscale = TCP_MAX_WSCALE;
3778 opt_rx->snd_wscale = snd_wscale;
3781 case TCPOPT_TIMESTAMP:
3782 if ((opsize == TCPOLEN_TIMESTAMP) &&
3783 ((estab && opt_rx->tstamp_ok) ||
3784 (!estab && sysctl_tcp_timestamps))) {
3785 opt_rx->saw_tstamp = 1;
3786 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3787 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3790 case TCPOPT_SACK_PERM:
3791 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3792 !estab && net->ipv4.sysctl_tcp_sack) {
3793 opt_rx->sack_ok = TCP_SACK_SEEN;
3794 tcp_sack_reset(opt_rx);
3799 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3800 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3802 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3805 #ifdef CONFIG_TCP_MD5SIG
3808 * The MD5 Hash has already been
3809 * checked (see tcp_v{4,6}_do_rcv()).
3813 case TCPOPT_FASTOPEN:
3814 tcp_parse_fastopen_option(
3815 opsize - TCPOLEN_FASTOPEN_BASE,
3816 ptr, th->syn, foc, false);
3820 /* Fast Open option shares code 254 using a
3821 * 16 bits magic number.
3823 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
3824 get_unaligned_be16(ptr) ==
3825 TCPOPT_FASTOPEN_MAGIC)
3826 tcp_parse_fastopen_option(opsize -
3827 TCPOLEN_EXP_FASTOPEN_BASE,
3828 ptr + 2, th->syn, foc, true);
3837 EXPORT_SYMBOL(tcp_parse_options);
3839 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3841 const __be32 *ptr = (const __be32 *)(th + 1);
3843 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3844 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3845 tp->rx_opt.saw_tstamp = 1;
3847 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3850 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3852 tp->rx_opt.rcv_tsecr = 0;
3858 /* Fast parse options. This hopes to only see timestamps.
3859 * If it is wrong it falls back on tcp_parse_options().
3861 static bool tcp_fast_parse_options(const struct net *net,
3862 const struct sk_buff *skb,
3863 const struct tcphdr *th, struct tcp_sock *tp)
3865 /* In the spirit of fast parsing, compare doff directly to constant
3866 * values. Because equality is used, short doff can be ignored here.
3868 if (th->doff == (sizeof(*th) / 4)) {
3869 tp->rx_opt.saw_tstamp = 0;
3871 } else if (tp->rx_opt.tstamp_ok &&
3872 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3873 if (tcp_parse_aligned_timestamp(tp, th))
3877 tcp_parse_options(net, skb, &tp->rx_opt, 1, NULL);
3878 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3879 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3884 #ifdef CONFIG_TCP_MD5SIG
3886 * Parse MD5 Signature option
3888 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3890 int length = (th->doff << 2) - sizeof(*th);
3891 const u8 *ptr = (const u8 *)(th + 1);
3893 /* If the TCP option is too short, we can short cut */
3894 if (length < TCPOLEN_MD5SIG)
3897 while (length > 0) {
3898 int opcode = *ptr++;
3909 if (opsize < 2 || opsize > length)
3911 if (opcode == TCPOPT_MD5SIG)
3912 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3919 EXPORT_SYMBOL(tcp_parse_md5sig_option);
3922 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3924 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3925 * it can pass through stack. So, the following predicate verifies that
3926 * this segment is not used for anything but congestion avoidance or
3927 * fast retransmit. Moreover, we even are able to eliminate most of such
3928 * second order effects, if we apply some small "replay" window (~RTO)
3929 * to timestamp space.
3931 * All these measures still do not guarantee that we reject wrapped ACKs
3932 * on networks with high bandwidth, when sequence space is recycled fastly,
3933 * but it guarantees that such events will be very rare and do not affect
3934 * connection seriously. This doesn't look nice, but alas, PAWS is really
3937 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3938 * states that events when retransmit arrives after original data are rare.
3939 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3940 * the biggest problem on large power networks even with minor reordering.
3941 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3942 * up to bandwidth of 18Gigabit/sec. 8) ]
3945 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3947 const struct tcp_sock *tp = tcp_sk(sk);
3948 const struct tcphdr *th = tcp_hdr(skb);
3949 u32 seq = TCP_SKB_CB(skb)->seq;
3950 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3952 return (/* 1. Pure ACK with correct sequence number. */
3953 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3955 /* 2. ... and duplicate ACK. */
3956 ack == tp->snd_una &&
3958 /* 3. ... and does not update window. */
3959 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3961 /* 4. ... and sits in replay window. */
3962 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3965 static inline bool tcp_paws_discard(const struct sock *sk,
3966 const struct sk_buff *skb)
3968 const struct tcp_sock *tp = tcp_sk(sk);
3970 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
3971 !tcp_disordered_ack(sk, skb);
3974 /* Check segment sequence number for validity.
3976 * Segment controls are considered valid, if the segment
3977 * fits to the window after truncation to the window. Acceptability
3978 * of data (and SYN, FIN, of course) is checked separately.
3979 * See tcp_data_queue(), for example.
3981 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3982 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3983 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3984 * (borrowed from freebsd)
3987 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
3989 return !before(end_seq, tp->rcv_wup) &&
3990 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3993 /* When we get a reset we do this. */
3994 void tcp_reset(struct sock *sk)
3996 /* We want the right error as BSD sees it (and indeed as we do). */
3997 switch (sk->sk_state) {
3999 sk->sk_err = ECONNREFUSED;
4001 case TCP_CLOSE_WAIT:
4007 sk->sk_err = ECONNRESET;
4009 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4014 if (!sock_flag(sk, SOCK_DEAD))
4015 sk->sk_error_report(sk);
4019 * Process the FIN bit. This now behaves as it is supposed to work
4020 * and the FIN takes effect when it is validly part of sequence
4021 * space. Not before when we get holes.
4023 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4024 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4027 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4028 * close and we go into CLOSING (and later onto TIME-WAIT)
4030 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4032 void tcp_fin(struct sock *sk)
4034 struct tcp_sock *tp = tcp_sk(sk);
4036 inet_csk_schedule_ack(sk);
4038 sk->sk_shutdown |= RCV_SHUTDOWN;
4039 sock_set_flag(sk, SOCK_DONE);
4041 switch (sk->sk_state) {
4043 case TCP_ESTABLISHED:
4044 /* Move to CLOSE_WAIT */
4045 tcp_set_state(sk, TCP_CLOSE_WAIT);
4046 inet_csk(sk)->icsk_ack.pingpong = 1;
4049 case TCP_CLOSE_WAIT:
4051 /* Received a retransmission of the FIN, do
4056 /* RFC793: Remain in the LAST-ACK state. */
4060 /* This case occurs when a simultaneous close
4061 * happens, we must ack the received FIN and
4062 * enter the CLOSING state.
4065 tcp_set_state(sk, TCP_CLOSING);
4068 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4070 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4073 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4074 * cases we should never reach this piece of code.
4076 pr_err("%s: Impossible, sk->sk_state=%d\n",
4077 __func__, sk->sk_state);
4081 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4082 * Probably, we should reset in this case. For now drop them.
4084 skb_rbtree_purge(&tp->out_of_order_queue);
4085 if (tcp_is_sack(tp))
4086 tcp_sack_reset(&tp->rx_opt);
4089 if (!sock_flag(sk, SOCK_DEAD)) {
4090 sk->sk_state_change(sk);
4092 /* Do not send POLL_HUP for half duplex close. */
4093 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4094 sk->sk_state == TCP_CLOSE)
4095 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4097 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4101 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4104 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4105 if (before(seq, sp->start_seq))
4106 sp->start_seq = seq;
4107 if (after(end_seq, sp->end_seq))
4108 sp->end_seq = end_seq;
4114 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4116 struct tcp_sock *tp = tcp_sk(sk);
4118 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4121 if (before(seq, tp->rcv_nxt))
4122 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4124 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4126 NET_INC_STATS(sock_net(sk), mib_idx);
4128 tp->rx_opt.dsack = 1;
4129 tp->duplicate_sack[0].start_seq = seq;
4130 tp->duplicate_sack[0].end_seq = end_seq;
4134 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4136 struct tcp_sock *tp = tcp_sk(sk);
4138 if (!tp->rx_opt.dsack)
4139 tcp_dsack_set(sk, seq, end_seq);
4141 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4144 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4146 struct tcp_sock *tp = tcp_sk(sk);
4148 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4149 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4150 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4151 tcp_enter_quickack_mode(sk);
4153 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4154 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4156 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4157 end_seq = tp->rcv_nxt;
4158 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4165 /* These routines update the SACK block as out-of-order packets arrive or
4166 * in-order packets close up the sequence space.
4168 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4171 struct tcp_sack_block *sp = &tp->selective_acks[0];
4172 struct tcp_sack_block *swalk = sp + 1;
4174 /* See if the recent change to the first SACK eats into
4175 * or hits the sequence space of other SACK blocks, if so coalesce.
4177 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4178 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4181 /* Zap SWALK, by moving every further SACK up by one slot.
4182 * Decrease num_sacks.
4184 tp->rx_opt.num_sacks--;
4185 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4189 this_sack++, swalk++;
4193 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4195 struct tcp_sock *tp = tcp_sk(sk);
4196 struct tcp_sack_block *sp = &tp->selective_acks[0];
4197 int cur_sacks = tp->rx_opt.num_sacks;
4203 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4204 if (tcp_sack_extend(sp, seq, end_seq)) {
4205 /* Rotate this_sack to the first one. */
4206 for (; this_sack > 0; this_sack--, sp--)
4207 swap(*sp, *(sp - 1));
4209 tcp_sack_maybe_coalesce(tp);
4214 /* Could not find an adjacent existing SACK, build a new one,
4215 * put it at the front, and shift everyone else down. We
4216 * always know there is at least one SACK present already here.
4218 * If the sack array is full, forget about the last one.
4220 if (this_sack >= TCP_NUM_SACKS) {
4222 tp->rx_opt.num_sacks--;
4225 for (; this_sack > 0; this_sack--, sp--)
4229 /* Build the new head SACK, and we're done. */
4230 sp->start_seq = seq;
4231 sp->end_seq = end_seq;
4232 tp->rx_opt.num_sacks++;
4235 /* RCV.NXT advances, some SACKs should be eaten. */
4237 static void tcp_sack_remove(struct tcp_sock *tp)
4239 struct tcp_sack_block *sp = &tp->selective_acks[0];
4240 int num_sacks = tp->rx_opt.num_sacks;
4243 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4244 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4245 tp->rx_opt.num_sacks = 0;
4249 for (this_sack = 0; this_sack < num_sacks;) {
4250 /* Check if the start of the sack is covered by RCV.NXT. */
4251 if (!before(tp->rcv_nxt, sp->start_seq)) {
4254 /* RCV.NXT must cover all the block! */
4255 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4257 /* Zap this SACK, by moving forward any other SACKS. */
4258 for (i = this_sack+1; i < num_sacks; i++)
4259 tp->selective_acks[i-1] = tp->selective_acks[i];
4266 tp->rx_opt.num_sacks = num_sacks;
4270 * tcp_try_coalesce - try to merge skb to prior one
4273 * @from: buffer to add in queue
4274 * @fragstolen: pointer to boolean
4276 * Before queueing skb @from after @to, try to merge them
4277 * to reduce overall memory use and queue lengths, if cost is small.
4278 * Packets in ofo or receive queues can stay a long time.
4279 * Better try to coalesce them right now to avoid future collapses.
4280 * Returns true if caller should free @from instead of queueing it
4282 static bool tcp_try_coalesce(struct sock *sk,
4284 struct sk_buff *from,
4289 *fragstolen = false;
4291 /* Its possible this segment overlaps with prior segment in queue */
4292 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4295 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4298 atomic_add(delta, &sk->sk_rmem_alloc);
4299 sk_mem_charge(sk, delta);
4300 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4301 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4302 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4303 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4307 static void tcp_drop(struct sock *sk, struct sk_buff *skb)
4309 sk_drops_add(sk, skb);
4313 /* This one checks to see if we can put data from the
4314 * out_of_order queue into the receive_queue.
4316 static void tcp_ofo_queue(struct sock *sk)
4318 struct tcp_sock *tp = tcp_sk(sk);
4319 __u32 dsack_high = tp->rcv_nxt;
4320 bool fin, fragstolen, eaten;
4321 struct sk_buff *skb, *tail;
4324 p = rb_first(&tp->out_of_order_queue);
4326 skb = rb_entry(p, struct sk_buff, rbnode);
4327 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4330 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4331 __u32 dsack = dsack_high;
4332 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4333 dsack_high = TCP_SKB_CB(skb)->end_seq;
4334 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4337 rb_erase(&skb->rbnode, &tp->out_of_order_queue);
4339 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) {
4340 SOCK_DEBUG(sk, "ofo packet was already received\n");
4344 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4345 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4346 TCP_SKB_CB(skb)->end_seq);
4348 tail = skb_peek_tail(&sk->sk_receive_queue);
4349 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4350 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4351 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
4353 __skb_queue_tail(&sk->sk_receive_queue, skb);
4355 kfree_skb_partial(skb, fragstolen);
4357 if (unlikely(fin)) {
4359 /* tcp_fin() purges tp->out_of_order_queue,
4360 * so we must end this loop right now.
4367 static bool tcp_prune_ofo_queue(struct sock *sk);
4368 static int tcp_prune_queue(struct sock *sk);
4370 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4373 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4374 !sk_rmem_schedule(sk, skb, size)) {
4376 if (tcp_prune_queue(sk) < 0)
4379 while (!sk_rmem_schedule(sk, skb, size)) {
4380 if (!tcp_prune_ofo_queue(sk))
4387 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4389 struct tcp_sock *tp = tcp_sk(sk);
4390 struct rb_node **p, *q, *parent;
4391 struct sk_buff *skb1;
4395 tcp_ecn_check_ce(tp, skb);
4397 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4398 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP);
4403 /* Disable header prediction. */
4405 inet_csk_schedule_ack(sk);
4407 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4408 seq = TCP_SKB_CB(skb)->seq;
4409 end_seq = TCP_SKB_CB(skb)->end_seq;
4410 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4411 tp->rcv_nxt, seq, end_seq);
4413 p = &tp->out_of_order_queue.rb_node;
4414 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4415 /* Initial out of order segment, build 1 SACK. */
4416 if (tcp_is_sack(tp)) {
4417 tp->rx_opt.num_sacks = 1;
4418 tp->selective_acks[0].start_seq = seq;
4419 tp->selective_acks[0].end_seq = end_seq;
4421 rb_link_node(&skb->rbnode, NULL, p);
4422 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4423 tp->ooo_last_skb = skb;
4427 /* In the typical case, we are adding an skb to the end of the list.
4428 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4430 if (tcp_try_coalesce(sk, tp->ooo_last_skb, skb, &fragstolen)) {
4432 tcp_grow_window(sk, skb);
4433 kfree_skb_partial(skb, fragstolen);
4437 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4438 if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) {
4439 parent = &tp->ooo_last_skb->rbnode;
4440 p = &parent->rb_right;
4444 /* Find place to insert this segment. Handle overlaps on the way. */
4448 skb1 = rb_entry(parent, struct sk_buff, rbnode);
4449 if (before(seq, TCP_SKB_CB(skb1)->seq)) {
4450 p = &parent->rb_left;
4453 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4454 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4455 /* All the bits are present. Drop. */
4456 NET_INC_STATS(sock_net(sk),
4457 LINUX_MIB_TCPOFOMERGE);
4460 tcp_dsack_set(sk, seq, end_seq);
4463 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4464 /* Partial overlap. */
4465 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq);
4467 /* skb's seq == skb1's seq and skb covers skb1.
4468 * Replace skb1 with skb.
4470 rb_replace_node(&skb1->rbnode, &skb->rbnode,
4471 &tp->out_of_order_queue);
4472 tcp_dsack_extend(sk,
4473 TCP_SKB_CB(skb1)->seq,
4474 TCP_SKB_CB(skb1)->end_seq);
4475 NET_INC_STATS(sock_net(sk),
4476 LINUX_MIB_TCPOFOMERGE);
4480 } else if (tcp_try_coalesce(sk, skb1, skb, &fragstolen)) {
4483 p = &parent->rb_right;
4486 /* Insert segment into RB tree. */
4487 rb_link_node(&skb->rbnode, parent, p);
4488 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4491 /* Remove other segments covered by skb. */
4492 while ((q = rb_next(&skb->rbnode)) != NULL) {
4493 skb1 = rb_entry(q, struct sk_buff, rbnode);
4495 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4497 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4498 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4502 rb_erase(&skb1->rbnode, &tp->out_of_order_queue);
4503 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4504 TCP_SKB_CB(skb1)->end_seq);
4505 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4508 /* If there is no skb after us, we are the last_skb ! */
4510 tp->ooo_last_skb = skb;
4513 if (tcp_is_sack(tp))
4514 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4517 tcp_grow_window(sk, skb);
4519 skb_set_owner_r(skb, sk);
4523 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
4527 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4529 __skb_pull(skb, hdrlen);
4531 tcp_try_coalesce(sk, tail, skb, fragstolen)) ? 1 : 0;
4532 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
4534 __skb_queue_tail(&sk->sk_receive_queue, skb);
4535 skb_set_owner_r(skb, sk);
4540 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4542 struct sk_buff *skb;
4550 if (size > PAGE_SIZE) {
4551 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
4553 data_len = npages << PAGE_SHIFT;
4554 size = data_len + (size & ~PAGE_MASK);
4556 skb = alloc_skb_with_frags(size - data_len, data_len,
4557 PAGE_ALLOC_COSTLY_ORDER,
4558 &err, sk->sk_allocation);
4562 skb_put(skb, size - data_len);
4563 skb->data_len = data_len;
4566 if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4569 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
4573 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4574 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4575 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4577 if (tcp_queue_rcv(sk, skb, 0, &fragstolen)) {
4578 WARN_ON_ONCE(fragstolen); /* should not happen */
4590 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4592 struct tcp_sock *tp = tcp_sk(sk);
4593 bool fragstolen = false;
4596 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
4601 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
4603 tcp_ecn_accept_cwr(tp, skb);
4605 tp->rx_opt.dsack = 0;
4607 /* Queue data for delivery to the user.
4608 * Packets in sequence go to the receive queue.
4609 * Out of sequence packets to the out_of_order_queue.
4611 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4612 if (tcp_receive_window(tp) == 0)
4615 /* Ok. In sequence. In window. */
4616 if (tp->ucopy.task == current &&
4617 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4618 sock_owned_by_user(sk) && !tp->urg_data) {
4619 int chunk = min_t(unsigned int, skb->len,
4622 __set_current_state(TASK_RUNNING);
4624 if (!skb_copy_datagram_msg(skb, 0, tp->ucopy.msg, chunk)) {
4625 tp->ucopy.len -= chunk;
4626 tp->copied_seq += chunk;
4627 eaten = (chunk == skb->len);
4628 tcp_rcv_space_adjust(sk);
4635 if (skb_queue_len(&sk->sk_receive_queue) == 0)
4636 sk_forced_mem_schedule(sk, skb->truesize);
4637 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4640 eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
4642 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4644 tcp_event_data_recv(sk, skb);
4645 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4648 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4651 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4652 * gap in queue is filled.
4654 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
4655 inet_csk(sk)->icsk_ack.pingpong = 0;
4658 if (tp->rx_opt.num_sacks)
4659 tcp_sack_remove(tp);
4661 tcp_fast_path_check(sk);
4664 kfree_skb_partial(skb, fragstolen);
4665 if (!sock_flag(sk, SOCK_DEAD))
4666 sk->sk_data_ready(sk);
4670 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4671 /* A retransmit, 2nd most common case. Force an immediate ack. */
4672 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4673 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4676 tcp_enter_quickack_mode(sk);
4677 inet_csk_schedule_ack(sk);
4683 /* Out of window. F.e. zero window probe. */
4684 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4687 tcp_enter_quickack_mode(sk);
4689 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4690 /* Partial packet, seq < rcv_next < end_seq */
4691 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4692 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4693 TCP_SKB_CB(skb)->end_seq);
4695 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4697 /* If window is closed, drop tail of packet. But after
4698 * remembering D-SACK for its head made in previous line.
4700 if (!tcp_receive_window(tp))
4705 tcp_data_queue_ofo(sk, skb);
4708 static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list)
4711 return !skb_queue_is_last(list, skb) ? skb->next : NULL;
4713 return rb_entry_safe(rb_next(&skb->rbnode), struct sk_buff, rbnode);
4716 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4717 struct sk_buff_head *list,
4718 struct rb_root *root)
4720 struct sk_buff *next = tcp_skb_next(skb, list);
4723 __skb_unlink(skb, list);
4725 rb_erase(&skb->rbnode, root);
4728 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4733 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
4734 static void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb)
4736 struct rb_node **p = &root->rb_node;
4737 struct rb_node *parent = NULL;
4738 struct sk_buff *skb1;
4742 skb1 = rb_entry(parent, struct sk_buff, rbnode);
4743 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq))
4744 p = &parent->rb_left;
4746 p = &parent->rb_right;
4748 rb_link_node(&skb->rbnode, parent, p);
4749 rb_insert_color(&skb->rbnode, root);
4752 /* Collapse contiguous sequence of skbs head..tail with
4753 * sequence numbers start..end.
4755 * If tail is NULL, this means until the end of the queue.
4757 * Segments with FIN/SYN are not collapsed (only because this
4761 tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root,
4762 struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end)
4764 struct sk_buff *skb = head, *n;
4765 struct sk_buff_head tmp;
4768 /* First, check that queue is collapsible and find
4769 * the point where collapsing can be useful.
4772 for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) {
4773 n = tcp_skb_next(skb, list);
4775 /* No new bits? It is possible on ofo queue. */
4776 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4777 skb = tcp_collapse_one(sk, skb, list, root);
4783 /* The first skb to collapse is:
4785 * - bloated or contains data before "start" or
4786 * overlaps to the next one.
4788 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
4789 (tcp_win_from_space(skb->truesize) > skb->len ||
4790 before(TCP_SKB_CB(skb)->seq, start))) {
4791 end_of_skbs = false;
4795 if (n && n != tail &&
4796 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) {
4797 end_of_skbs = false;
4801 /* Decided to skip this, advance start seq. */
4802 start = TCP_SKB_CB(skb)->end_seq;
4805 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4808 __skb_queue_head_init(&tmp);
4810 while (before(start, end)) {
4811 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
4812 struct sk_buff *nskb;
4814 nskb = alloc_skb(copy, GFP_ATOMIC);
4818 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4819 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4821 __skb_queue_before(list, skb, nskb);
4823 __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */
4824 skb_set_owner_r(nskb, sk);
4826 /* Copy data, releasing collapsed skbs. */
4828 int offset = start - TCP_SKB_CB(skb)->seq;
4829 int size = TCP_SKB_CB(skb)->end_seq - start;
4833 size = min(copy, size);
4834 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4836 TCP_SKB_CB(nskb)->end_seq += size;
4840 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4841 skb = tcp_collapse_one(sk, skb, list, root);
4844 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4850 skb_queue_walk_safe(&tmp, skb, n)
4851 tcp_rbtree_insert(root, skb);
4854 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4855 * and tcp_collapse() them until all the queue is collapsed.
4857 static void tcp_collapse_ofo_queue(struct sock *sk)
4859 struct tcp_sock *tp = tcp_sk(sk);
4860 struct sk_buff *skb, *head;
4864 p = rb_first(&tp->out_of_order_queue);
4865 skb = rb_entry_safe(p, struct sk_buff, rbnode);
4868 p = rb_last(&tp->out_of_order_queue);
4869 /* Note: This is possible p is NULL here. We do not
4870 * use rb_entry_safe(), as ooo_last_skb is valid only
4871 * if rbtree is not empty.
4873 tp->ooo_last_skb = rb_entry(p, struct sk_buff, rbnode);
4876 start = TCP_SKB_CB(skb)->seq;
4877 end = TCP_SKB_CB(skb)->end_seq;
4879 for (head = skb;;) {
4880 skb = tcp_skb_next(skb, NULL);
4882 /* Range is terminated when we see a gap or when
4883 * we are at the queue end.
4886 after(TCP_SKB_CB(skb)->seq, end) ||
4887 before(TCP_SKB_CB(skb)->end_seq, start)) {
4888 tcp_collapse(sk, NULL, &tp->out_of_order_queue,
4889 head, skb, start, end);
4893 if (unlikely(before(TCP_SKB_CB(skb)->seq, start)))
4894 start = TCP_SKB_CB(skb)->seq;
4895 if (after(TCP_SKB_CB(skb)->end_seq, end))
4896 end = TCP_SKB_CB(skb)->end_seq;
4901 * Clean the out-of-order queue to make room.
4902 * We drop high sequences packets to :
4903 * 1) Let a chance for holes to be filled.
4904 * 2) not add too big latencies if thousands of packets sit there.
4905 * (But if application shrinks SO_RCVBUF, we could still end up
4906 * freeing whole queue here)
4908 * Return true if queue has shrunk.
4910 static bool tcp_prune_ofo_queue(struct sock *sk)
4912 struct tcp_sock *tp = tcp_sk(sk);
4913 struct rb_node *node, *prev;
4915 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
4918 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED);
4919 node = &tp->ooo_last_skb->rbnode;
4921 prev = rb_prev(node);
4922 rb_erase(node, &tp->out_of_order_queue);
4923 tcp_drop(sk, rb_entry(node, struct sk_buff, rbnode));
4925 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf &&
4926 !tcp_under_memory_pressure(sk))
4930 tp->ooo_last_skb = rb_entry(prev, struct sk_buff, rbnode);
4932 /* Reset SACK state. A conforming SACK implementation will
4933 * do the same at a timeout based retransmit. When a connection
4934 * is in a sad state like this, we care only about integrity
4935 * of the connection not performance.
4937 if (tp->rx_opt.sack_ok)
4938 tcp_sack_reset(&tp->rx_opt);
4942 /* Reduce allocated memory if we can, trying to get
4943 * the socket within its memory limits again.
4945 * Return less than zero if we should start dropping frames
4946 * until the socket owning process reads some of the data
4947 * to stabilize the situation.
4949 static int tcp_prune_queue(struct sock *sk)
4951 struct tcp_sock *tp = tcp_sk(sk);
4953 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4955 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED);
4957 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4958 tcp_clamp_window(sk);
4959 else if (tcp_under_memory_pressure(sk))
4960 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4962 tcp_collapse_ofo_queue(sk);
4963 if (!skb_queue_empty(&sk->sk_receive_queue))
4964 tcp_collapse(sk, &sk->sk_receive_queue, NULL,
4965 skb_peek(&sk->sk_receive_queue),
4967 tp->copied_seq, tp->rcv_nxt);
4970 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4973 /* Collapsing did not help, destructive actions follow.
4974 * This must not ever occur. */
4976 tcp_prune_ofo_queue(sk);
4978 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4981 /* If we are really being abused, tell the caller to silently
4982 * drop receive data on the floor. It will get retransmitted
4983 * and hopefully then we'll have sufficient space.
4985 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED);
4987 /* Massive buffer overcommit. */
4992 static bool tcp_should_expand_sndbuf(const struct sock *sk)
4994 const struct tcp_sock *tp = tcp_sk(sk);
4996 /* If the user specified a specific send buffer setting, do
4999 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
5002 /* If we are under global TCP memory pressure, do not expand. */
5003 if (tcp_under_memory_pressure(sk))
5006 /* If we are under soft global TCP memory pressure, do not expand. */
5007 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
5010 /* If we filled the congestion window, do not expand. */
5011 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
5017 /* When incoming ACK allowed to free some skb from write_queue,
5018 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
5019 * on the exit from tcp input handler.
5021 * PROBLEM: sndbuf expansion does not work well with largesend.
5023 static void tcp_new_space(struct sock *sk)
5025 struct tcp_sock *tp = tcp_sk(sk);
5027 if (tcp_should_expand_sndbuf(sk)) {
5028 tcp_sndbuf_expand(sk);
5029 tp->snd_cwnd_stamp = tcp_jiffies32;
5032 sk->sk_write_space(sk);
5035 static void tcp_check_space(struct sock *sk)
5037 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
5038 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
5039 /* pairs with tcp_poll() */
5041 if (sk->sk_socket &&
5042 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
5044 if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5045 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
5050 static inline void tcp_data_snd_check(struct sock *sk)
5052 tcp_push_pending_frames(sk);
5053 tcp_check_space(sk);
5057 * Check if sending an ack is needed.
5059 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5061 struct tcp_sock *tp = tcp_sk(sk);
5063 /* More than one full frame received... */
5064 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5065 /* ... and right edge of window advances far enough.
5066 * (tcp_recvmsg() will send ACK otherwise). Or...
5068 __tcp_select_window(sk) >= tp->rcv_wnd) ||
5069 /* We ACK each frame or... */
5070 tcp_in_quickack_mode(sk) ||
5071 /* We have out of order data. */
5072 (ofo_possible && !RB_EMPTY_ROOT(&tp->out_of_order_queue))) {
5073 /* Then ack it now */
5076 /* Else, send delayed ack. */
5077 tcp_send_delayed_ack(sk);
5081 static inline void tcp_ack_snd_check(struct sock *sk)
5083 if (!inet_csk_ack_scheduled(sk)) {
5084 /* We sent a data segment already. */
5087 __tcp_ack_snd_check(sk, 1);
5091 * This routine is only called when we have urgent data
5092 * signaled. Its the 'slow' part of tcp_urg. It could be
5093 * moved inline now as tcp_urg is only called from one
5094 * place. We handle URGent data wrong. We have to - as
5095 * BSD still doesn't use the correction from RFC961.
5096 * For 1003.1g we should support a new option TCP_STDURG to permit
5097 * either form (or just set the sysctl tcp_stdurg).
5100 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5102 struct tcp_sock *tp = tcp_sk(sk);
5103 u32 ptr = ntohs(th->urg_ptr);
5105 if (ptr && !sysctl_tcp_stdurg)
5107 ptr += ntohl(th->seq);
5109 /* Ignore urgent data that we've already seen and read. */
5110 if (after(tp->copied_seq, ptr))
5113 /* Do not replay urg ptr.
5115 * NOTE: interesting situation not covered by specs.
5116 * Misbehaving sender may send urg ptr, pointing to segment,
5117 * which we already have in ofo queue. We are not able to fetch
5118 * such data and will stay in TCP_URG_NOTYET until will be eaten
5119 * by recvmsg(). Seems, we are not obliged to handle such wicked
5120 * situations. But it is worth to think about possibility of some
5121 * DoSes using some hypothetical application level deadlock.
5123 if (before(ptr, tp->rcv_nxt))
5126 /* Do we already have a newer (or duplicate) urgent pointer? */
5127 if (tp->urg_data && !after(ptr, tp->urg_seq))
5130 /* Tell the world about our new urgent pointer. */
5133 /* We may be adding urgent data when the last byte read was
5134 * urgent. To do this requires some care. We cannot just ignore
5135 * tp->copied_seq since we would read the last urgent byte again
5136 * as data, nor can we alter copied_seq until this data arrives
5137 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5139 * NOTE. Double Dutch. Rendering to plain English: author of comment
5140 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5141 * and expect that both A and B disappear from stream. This is _wrong_.
5142 * Though this happens in BSD with high probability, this is occasional.
5143 * Any application relying on this is buggy. Note also, that fix "works"
5144 * only in this artificial test. Insert some normal data between A and B and we will
5145 * decline of BSD again. Verdict: it is better to remove to trap
5148 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5149 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5150 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5152 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5153 __skb_unlink(skb, &sk->sk_receive_queue);
5158 tp->urg_data = TCP_URG_NOTYET;
5161 /* Disable header prediction. */
5165 /* This is the 'fast' part of urgent handling. */
5166 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5168 struct tcp_sock *tp = tcp_sk(sk);
5170 /* Check if we get a new urgent pointer - normally not. */
5172 tcp_check_urg(sk, th);
5174 /* Do we wait for any urgent data? - normally not... */
5175 if (tp->urg_data == TCP_URG_NOTYET) {
5176 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5179 /* Is the urgent pointer pointing into this packet? */
5180 if (ptr < skb->len) {
5182 if (skb_copy_bits(skb, ptr, &tmp, 1))
5184 tp->urg_data = TCP_URG_VALID | tmp;
5185 if (!sock_flag(sk, SOCK_DEAD))
5186 sk->sk_data_ready(sk);
5191 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
5193 struct tcp_sock *tp = tcp_sk(sk);
5194 int chunk = skb->len - hlen;
5197 if (skb_csum_unnecessary(skb))
5198 err = skb_copy_datagram_msg(skb, hlen, tp->ucopy.msg, chunk);
5200 err = skb_copy_and_csum_datagram_msg(skb, hlen, tp->ucopy.msg);
5203 tp->ucopy.len -= chunk;
5204 tp->copied_seq += chunk;
5205 tcp_rcv_space_adjust(sk);
5211 /* Accept RST for rcv_nxt - 1 after a FIN.
5212 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a
5213 * FIN is sent followed by a RST packet. The RST is sent with the same
5214 * sequence number as the FIN, and thus according to RFC 5961 a challenge
5215 * ACK should be sent. However, Mac OSX rate limits replies to challenge
5216 * ACKs on the closed socket. In addition middleboxes can drop either the
5217 * challenge ACK or a subsequent RST.
5219 static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb)
5221 struct tcp_sock *tp = tcp_sk(sk);
5223 return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) &&
5224 (1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK |
5228 /* Does PAWS and seqno based validation of an incoming segment, flags will
5229 * play significant role here.
5231 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5232 const struct tcphdr *th, int syn_inerr)
5234 struct tcp_sock *tp = tcp_sk(sk);
5235 bool rst_seq_match = false;
5237 /* RFC1323: H1. Apply PAWS check first. */
5238 if (tcp_fast_parse_options(sock_net(sk), skb, th, tp) &&
5239 tp->rx_opt.saw_tstamp &&
5240 tcp_paws_discard(sk, skb)) {
5242 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5243 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5244 LINUX_MIB_TCPACKSKIPPEDPAWS,
5245 &tp->last_oow_ack_time))
5246 tcp_send_dupack(sk, skb);
5249 /* Reset is accepted even if it did not pass PAWS. */
5252 /* Step 1: check sequence number */
5253 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5254 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5255 * (RST) segments are validated by checking their SEQ-fields."
5256 * And page 69: "If an incoming segment is not acceptable,
5257 * an acknowledgment should be sent in reply (unless the RST
5258 * bit is set, if so drop the segment and return)".
5263 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5264 LINUX_MIB_TCPACKSKIPPEDSEQ,
5265 &tp->last_oow_ack_time))
5266 tcp_send_dupack(sk, skb);
5267 } else if (tcp_reset_check(sk, skb)) {
5273 /* Step 2: check RST bit */
5275 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a
5276 * FIN and SACK too if available):
5277 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or
5278 * the right-most SACK block,
5280 * RESET the connection
5282 * Send a challenge ACK
5284 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt ||
5285 tcp_reset_check(sk, skb)) {
5286 rst_seq_match = true;
5287 } else if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) {
5288 struct tcp_sack_block *sp = &tp->selective_acks[0];
5289 int max_sack = sp[0].end_seq;
5292 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;
5294 max_sack = after(sp[this_sack].end_seq,
5296 sp[this_sack].end_seq : max_sack;
5299 if (TCP_SKB_CB(skb)->seq == max_sack)
5300 rst_seq_match = true;
5306 /* Disable TFO if RST is out-of-order
5307 * and no data has been received
5308 * for current active TFO socket
5310 if (tp->syn_fastopen && !tp->data_segs_in &&
5311 sk->sk_state == TCP_ESTABLISHED)
5312 tcp_fastopen_active_disable(sk);
5313 tcp_send_challenge_ack(sk, skb);
5318 /* step 3: check security and precedence [ignored] */
5320 /* step 4: Check for a SYN
5321 * RFC 5961 4.2 : Send a challenge ack
5326 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5327 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5328 tcp_send_challenge_ack(sk, skb);
5340 * TCP receive function for the ESTABLISHED state.
5342 * It is split into a fast path and a slow path. The fast path is
5344 * - A zero window was announced from us - zero window probing
5345 * is only handled properly in the slow path.
5346 * - Out of order segments arrived.
5347 * - Urgent data is expected.
5348 * - There is no buffer space left
5349 * - Unexpected TCP flags/window values/header lengths are received
5350 * (detected by checking the TCP header against pred_flags)
5351 * - Data is sent in both directions. Fast path only supports pure senders
5352 * or pure receivers (this means either the sequence number or the ack
5353 * value must stay constant)
5354 * - Unexpected TCP option.
5356 * When these conditions are not satisfied it drops into a standard
5357 * receive procedure patterned after RFC793 to handle all cases.
5358 * The first three cases are guaranteed by proper pred_flags setting,
5359 * the rest is checked inline. Fast processing is turned on in
5360 * tcp_data_queue when everything is OK.
5362 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5363 const struct tcphdr *th, unsigned int len)
5365 struct tcp_sock *tp = tcp_sk(sk);
5367 tcp_mstamp_refresh(tp);
5368 if (unlikely(!sk->sk_rx_dst))
5369 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5371 * Header prediction.
5372 * The code loosely follows the one in the famous
5373 * "30 instruction TCP receive" Van Jacobson mail.
5375 * Van's trick is to deposit buffers into socket queue
5376 * on a device interrupt, to call tcp_recv function
5377 * on the receive process context and checksum and copy
5378 * the buffer to user space. smart...
5380 * Our current scheme is not silly either but we take the
5381 * extra cost of the net_bh soft interrupt processing...
5382 * We do checksum and copy also but from device to kernel.
5385 tp->rx_opt.saw_tstamp = 0;
5387 /* pred_flags is 0xS?10 << 16 + snd_wnd
5388 * if header_prediction is to be made
5389 * 'S' will always be tp->tcp_header_len >> 2
5390 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5391 * turn it off (when there are holes in the receive
5392 * space for instance)
5393 * PSH flag is ignored.
5396 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5397 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5398 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5399 int tcp_header_len = tp->tcp_header_len;
5401 /* Timestamp header prediction: tcp_header_len
5402 * is automatically equal to th->doff*4 due to pred_flags
5406 /* Check timestamp */
5407 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5408 /* No? Slow path! */
5409 if (!tcp_parse_aligned_timestamp(tp, th))
5412 /* If PAWS failed, check it more carefully in slow path */
5413 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5416 /* DO NOT update ts_recent here, if checksum fails
5417 * and timestamp was corrupted part, it will result
5418 * in a hung connection since we will drop all
5419 * future packets due to the PAWS test.
5423 if (len <= tcp_header_len) {
5424 /* Bulk data transfer: sender */
5425 if (len == tcp_header_len) {
5426 /* Predicted packet is in window by definition.
5427 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5428 * Hence, check seq<=rcv_wup reduces to:
5430 if (tcp_header_len ==
5431 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5432 tp->rcv_nxt == tp->rcv_wup)
5433 tcp_store_ts_recent(tp);
5435 /* We know that such packets are checksummed
5438 tcp_ack(sk, skb, 0);
5440 tcp_data_snd_check(sk);
5442 } else { /* Header too small */
5443 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5448 bool fragstolen = false;
5450 if (tp->ucopy.task == current &&
5451 tp->copied_seq == tp->rcv_nxt &&
5452 len - tcp_header_len <= tp->ucopy.len &&
5453 sock_owned_by_user(sk)) {
5454 __set_current_state(TASK_RUNNING);
5456 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) {
5457 /* Predicted packet is in window by definition.
5458 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5459 * Hence, check seq<=rcv_wup reduces to:
5461 if (tcp_header_len ==
5462 (sizeof(struct tcphdr) +
5463 TCPOLEN_TSTAMP_ALIGNED) &&
5464 tp->rcv_nxt == tp->rcv_wup)
5465 tcp_store_ts_recent(tp);
5467 tcp_rcv_rtt_measure_ts(sk, skb);
5469 __skb_pull(skb, tcp_header_len);
5470 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
5471 NET_INC_STATS(sock_net(sk),
5472 LINUX_MIB_TCPHPHITSTOUSER);
5477 if (tcp_checksum_complete(skb))
5480 if ((int)skb->truesize > sk->sk_forward_alloc)
5483 /* Predicted packet is in window by definition.
5484 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5485 * Hence, check seq<=rcv_wup reduces to:
5487 if (tcp_header_len ==
5488 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5489 tp->rcv_nxt == tp->rcv_wup)
5490 tcp_store_ts_recent(tp);
5492 tcp_rcv_rtt_measure_ts(sk, skb);
5494 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS);
5496 /* Bulk data transfer: receiver */
5497 eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5501 tcp_event_data_recv(sk, skb);
5503 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5504 /* Well, only one small jumplet in fast path... */
5505 tcp_ack(sk, skb, FLAG_DATA);
5506 tcp_data_snd_check(sk);
5507 if (!inet_csk_ack_scheduled(sk))
5511 __tcp_ack_snd_check(sk, 0);
5514 kfree_skb_partial(skb, fragstolen);
5515 sk->sk_data_ready(sk);
5521 if (len < (th->doff << 2) || tcp_checksum_complete(skb))
5524 if (!th->ack && !th->rst && !th->syn)
5528 * Standard slow path.
5531 if (!tcp_validate_incoming(sk, skb, th, 1))
5535 if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5538 tcp_rcv_rtt_measure_ts(sk, skb);
5540 /* Process urgent data. */
5541 tcp_urg(sk, skb, th);
5543 /* step 7: process the segment text */
5544 tcp_data_queue(sk, skb);
5546 tcp_data_snd_check(sk);
5547 tcp_ack_snd_check(sk);
5551 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
5552 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5557 EXPORT_SYMBOL(tcp_rcv_established);
5559 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5561 struct tcp_sock *tp = tcp_sk(sk);
5562 struct inet_connection_sock *icsk = inet_csk(sk);
5564 tcp_set_state(sk, TCP_ESTABLISHED);
5565 icsk->icsk_ack.lrcvtime = tcp_jiffies32;
5568 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5569 security_inet_conn_established(sk, skb);
5572 /* Make sure socket is routed, for correct metrics. */
5573 icsk->icsk_af_ops->rebuild_header(sk);
5575 tcp_init_metrics(sk);
5577 tcp_init_congestion_control(sk);
5579 /* Prevent spurious tcp_cwnd_restart() on first data
5582 tp->lsndtime = tcp_jiffies32;
5584 tcp_init_buffer_space(sk);
5586 if (sock_flag(sk, SOCK_KEEPOPEN))
5587 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5589 if (!tp->rx_opt.snd_wscale)
5590 __tcp_fast_path_on(tp, tp->snd_wnd);
5596 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5597 struct tcp_fastopen_cookie *cookie)
5599 struct tcp_sock *tp = tcp_sk(sk);
5600 struct sk_buff *data = tp->syn_data ? tcp_write_queue_head(sk) : NULL;
5601 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
5602 bool syn_drop = false;
5604 if (mss == tp->rx_opt.user_mss) {
5605 struct tcp_options_received opt;
5607 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5608 tcp_clear_options(&opt);
5609 opt.user_mss = opt.mss_clamp = 0;
5610 tcp_parse_options(sock_net(sk), synack, &opt, 0, NULL);
5611 mss = opt.mss_clamp;
5614 if (!tp->syn_fastopen) {
5615 /* Ignore an unsolicited cookie */
5617 } else if (tp->total_retrans) {
5618 /* SYN timed out and the SYN-ACK neither has a cookie nor
5619 * acknowledges data. Presumably the remote received only
5620 * the retransmitted (regular) SYNs: either the original
5621 * SYN-data or the corresponding SYN-ACK was dropped.
5623 syn_drop = (cookie->len < 0 && data);
5624 } else if (cookie->len < 0 && !tp->syn_data) {
5625 /* We requested a cookie but didn't get it. If we did not use
5626 * the (old) exp opt format then try so next time (try_exp=1).
5627 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5629 try_exp = tp->syn_fastopen_exp ? 2 : 1;
5632 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
5634 if (data) { /* Retransmit unacked data in SYN */
5635 tcp_for_write_queue_from(data, sk) {
5636 if (data == tcp_send_head(sk) ||
5637 __tcp_retransmit_skb(sk, data, 1))
5641 NET_INC_STATS(sock_net(sk),
5642 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
5645 tp->syn_data_acked = tp->syn_data;
5646 if (tp->syn_data_acked)
5647 NET_INC_STATS(sock_net(sk),
5648 LINUX_MIB_TCPFASTOPENACTIVE);
5650 tcp_fastopen_add_skb(sk, synack);
5655 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5656 const struct tcphdr *th)
5658 struct inet_connection_sock *icsk = inet_csk(sk);
5659 struct tcp_sock *tp = tcp_sk(sk);
5660 struct tcp_fastopen_cookie foc = { .len = -1 };
5661 int saved_clamp = tp->rx_opt.mss_clamp;
5664 tcp_parse_options(sock_net(sk), skb, &tp->rx_opt, 0, &foc);
5665 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
5666 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
5670 * "If the state is SYN-SENT then
5671 * first check the ACK bit
5672 * If the ACK bit is set
5673 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5674 * a reset (unless the RST bit is set, if so drop
5675 * the segment and return)"
5677 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5678 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
5679 goto reset_and_undo;
5681 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5682 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5683 tcp_time_stamp(tp))) {
5684 NET_INC_STATS(sock_net(sk),
5685 LINUX_MIB_PAWSACTIVEREJECTED);
5686 goto reset_and_undo;
5689 /* Now ACK is acceptable.
5691 * "If the RST bit is set
5692 * If the ACK was acceptable then signal the user "error:
5693 * connection reset", drop the segment, enter CLOSED state,
5694 * delete TCB, and return."
5703 * "fifth, if neither of the SYN or RST bits is set then
5704 * drop the segment and return."
5710 goto discard_and_undo;
5713 * "If the SYN bit is on ...
5714 * are acceptable then ...
5715 * (our SYN has been ACKed), change the connection
5716 * state to ESTABLISHED..."
5719 tcp_ecn_rcv_synack(tp, th);
5721 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5722 tcp_ack(sk, skb, FLAG_SLOWPATH);
5724 /* Ok.. it's good. Set up sequence numbers and
5725 * move to established.
5727 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5728 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5730 /* RFC1323: The window in SYN & SYN/ACK segments is
5733 tp->snd_wnd = ntohs(th->window);
5735 if (!tp->rx_opt.wscale_ok) {
5736 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5737 tp->window_clamp = min(tp->window_clamp, 65535U);
5740 if (tp->rx_opt.saw_tstamp) {
5741 tp->rx_opt.tstamp_ok = 1;
5742 tp->tcp_header_len =
5743 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5744 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5745 tcp_store_ts_recent(tp);
5747 tp->tcp_header_len = sizeof(struct tcphdr);
5750 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5751 tcp_enable_fack(tp);
5754 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5755 tcp_initialize_rcv_mss(sk);
5757 /* Remember, tcp_poll() does not lock socket!
5758 * Change state from SYN-SENT only after copied_seq
5759 * is initialized. */
5760 tp->copied_seq = tp->rcv_nxt;
5764 tcp_finish_connect(sk, skb);
5766 fastopen_fail = (tp->syn_fastopen || tp->syn_data) &&
5767 tcp_rcv_fastopen_synack(sk, skb, &foc);
5769 if (!sock_flag(sk, SOCK_DEAD)) {
5770 sk->sk_state_change(sk);
5771 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5775 if (sk->sk_write_pending ||
5776 icsk->icsk_accept_queue.rskq_defer_accept ||
5777 icsk->icsk_ack.pingpong) {
5778 /* Save one ACK. Data will be ready after
5779 * several ticks, if write_pending is set.
5781 * It may be deleted, but with this feature tcpdumps
5782 * look so _wonderfully_ clever, that I was not able
5783 * to stand against the temptation 8) --ANK
5785 inet_csk_schedule_ack(sk);
5786 tcp_enter_quickack_mode(sk);
5787 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5788 TCP_DELACK_MAX, TCP_RTO_MAX);
5799 /* No ACK in the segment */
5803 * "If the RST bit is set
5805 * Otherwise (no ACK) drop the segment and return."
5808 goto discard_and_undo;
5812 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5813 tcp_paws_reject(&tp->rx_opt, 0))
5814 goto discard_and_undo;
5817 /* We see SYN without ACK. It is attempt of
5818 * simultaneous connect with crossed SYNs.
5819 * Particularly, it can be connect to self.
5821 tcp_set_state(sk, TCP_SYN_RECV);
5823 if (tp->rx_opt.saw_tstamp) {
5824 tp->rx_opt.tstamp_ok = 1;
5825 tcp_store_ts_recent(tp);
5826 tp->tcp_header_len =
5827 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5829 tp->tcp_header_len = sizeof(struct tcphdr);
5832 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5833 tp->copied_seq = tp->rcv_nxt;
5834 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5836 /* RFC1323: The window in SYN & SYN/ACK segments is
5839 tp->snd_wnd = ntohs(th->window);
5840 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5841 tp->max_window = tp->snd_wnd;
5843 tcp_ecn_rcv_syn(tp, th);
5846 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5847 tcp_initialize_rcv_mss(sk);
5849 tcp_send_synack(sk);
5851 /* Note, we could accept data and URG from this segment.
5852 * There are no obstacles to make this (except that we must
5853 * either change tcp_recvmsg() to prevent it from returning data
5854 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5856 * However, if we ignore data in ACKless segments sometimes,
5857 * we have no reasons to accept it sometimes.
5858 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5859 * is not flawless. So, discard packet for sanity.
5860 * Uncomment this return to process the data.
5867 /* "fifth, if neither of the SYN or RST bits is set then
5868 * drop the segment and return."
5872 tcp_clear_options(&tp->rx_opt);
5873 tp->rx_opt.mss_clamp = saved_clamp;
5877 tcp_clear_options(&tp->rx_opt);
5878 tp->rx_opt.mss_clamp = saved_clamp;
5883 * This function implements the receiving procedure of RFC 793 for
5884 * all states except ESTABLISHED and TIME_WAIT.
5885 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5886 * address independent.
5889 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
5891 struct tcp_sock *tp = tcp_sk(sk);
5892 struct inet_connection_sock *icsk = inet_csk(sk);
5893 const struct tcphdr *th = tcp_hdr(skb);
5894 struct request_sock *req;
5898 switch (sk->sk_state) {
5912 /* It is possible that we process SYN packets from backlog,
5913 * so we need to make sure to disable BH right there.
5916 acceptable = icsk->icsk_af_ops->conn_request(sk, skb) >= 0;
5927 tp->rx_opt.saw_tstamp = 0;
5928 tcp_mstamp_refresh(tp);
5929 queued = tcp_rcv_synsent_state_process(sk, skb, th);
5933 /* Do step6 onward by hand. */
5934 tcp_urg(sk, skb, th);
5936 tcp_data_snd_check(sk);
5940 tcp_mstamp_refresh(tp);
5941 tp->rx_opt.saw_tstamp = 0;
5942 req = tp->fastopen_rsk;
5944 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
5945 sk->sk_state != TCP_FIN_WAIT1);
5947 if (!tcp_check_req(sk, skb, req, true))
5951 if (!th->ack && !th->rst && !th->syn)
5954 if (!tcp_validate_incoming(sk, skb, th, 0))
5957 /* step 5: check the ACK field */
5958 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
5959 FLAG_UPDATE_TS_RECENT |
5960 FLAG_NO_CHALLENGE_ACK) > 0;
5963 if (sk->sk_state == TCP_SYN_RECV)
5964 return 1; /* send one RST */
5965 tcp_send_challenge_ack(sk, skb);
5968 switch (sk->sk_state) {
5971 tcp_synack_rtt_meas(sk, req);
5973 /* Once we leave TCP_SYN_RECV, we no longer need req
5977 inet_csk(sk)->icsk_retransmits = 0;
5978 reqsk_fastopen_remove(sk, req, false);
5980 /* Make sure socket is routed, for correct metrics. */
5981 icsk->icsk_af_ops->rebuild_header(sk);
5982 tcp_init_congestion_control(sk);
5985 tp->copied_seq = tp->rcv_nxt;
5986 tcp_init_buffer_space(sk);
5989 tcp_set_state(sk, TCP_ESTABLISHED);
5990 sk->sk_state_change(sk);
5992 /* Note, that this wakeup is only for marginal crossed SYN case.
5993 * Passively open sockets are not waked up, because
5994 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5997 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5999 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
6000 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
6001 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6003 if (tp->rx_opt.tstamp_ok)
6004 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6007 /* Re-arm the timer because data may have been sent out.
6008 * This is similar to the regular data transmission case
6009 * when new data has just been ack'ed.
6011 * (TFO) - we could try to be more aggressive and
6012 * retransmitting any data sooner based on when they
6017 tcp_init_metrics(sk);
6019 if (!inet_csk(sk)->icsk_ca_ops->cong_control)
6020 tcp_update_pacing_rate(sk);
6022 /* Prevent spurious tcp_cwnd_restart() on first data packet */
6023 tp->lsndtime = tcp_jiffies32;
6025 tcp_initialize_rcv_mss(sk);
6026 tcp_fast_path_on(tp);
6029 case TCP_FIN_WAIT1: {
6032 /* If we enter the TCP_FIN_WAIT1 state and we are a
6033 * Fast Open socket and this is the first acceptable
6034 * ACK we have received, this would have acknowledged
6035 * our SYNACK so stop the SYNACK timer.
6038 /* We no longer need the request sock. */
6039 reqsk_fastopen_remove(sk, req, false);
6042 if (tp->snd_una != tp->write_seq)
6045 tcp_set_state(sk, TCP_FIN_WAIT2);
6046 sk->sk_shutdown |= SEND_SHUTDOWN;
6050 if (!sock_flag(sk, SOCK_DEAD)) {
6051 /* Wake up lingering close() */
6052 sk->sk_state_change(sk);
6056 if (tp->linger2 < 0) {
6058 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6061 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6062 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6063 /* Receive out of order FIN after close() */
6064 if (tp->syn_fastopen && th->fin)
6065 tcp_fastopen_active_disable(sk);
6067 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6071 tmo = tcp_fin_time(sk);
6072 if (tmo > TCP_TIMEWAIT_LEN) {
6073 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
6074 } else if (th->fin || sock_owned_by_user(sk)) {
6075 /* Bad case. We could lose such FIN otherwise.
6076 * It is not a big problem, but it looks confusing
6077 * and not so rare event. We still can lose it now,
6078 * if it spins in bh_lock_sock(), but it is really
6081 inet_csk_reset_keepalive_timer(sk, tmo);
6083 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
6090 if (tp->snd_una == tp->write_seq) {
6091 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6097 if (tp->snd_una == tp->write_seq) {
6098 tcp_update_metrics(sk);
6105 /* step 6: check the URG bit */
6106 tcp_urg(sk, skb, th);
6108 /* step 7: process the segment text */
6109 switch (sk->sk_state) {
6110 case TCP_CLOSE_WAIT:
6113 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
6117 /* RFC 793 says to queue data in these states,
6118 * RFC 1122 says we MUST send a reset.
6119 * BSD 4.4 also does reset.
6121 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6122 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6123 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6124 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6130 case TCP_ESTABLISHED:
6131 tcp_data_queue(sk, skb);
6136 /* tcp_data could move socket to TIME-WAIT */
6137 if (sk->sk_state != TCP_CLOSE) {
6138 tcp_data_snd_check(sk);
6139 tcp_ack_snd_check(sk);
6148 EXPORT_SYMBOL(tcp_rcv_state_process);
6150 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
6152 struct inet_request_sock *ireq = inet_rsk(req);
6154 if (family == AF_INET)
6155 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6156 &ireq->ir_rmt_addr, port);
6157 #if IS_ENABLED(CONFIG_IPV6)
6158 else if (family == AF_INET6)
6159 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6160 &ireq->ir_v6_rmt_addr, port);
6164 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6166 * If we receive a SYN packet with these bits set, it means a
6167 * network is playing bad games with TOS bits. In order to
6168 * avoid possible false congestion notifications, we disable
6169 * TCP ECN negotiation.
6171 * Exception: tcp_ca wants ECN. This is required for DCTCP
6172 * congestion control: Linux DCTCP asserts ECT on all packets,
6173 * including SYN, which is most optimal solution; however,
6174 * others, such as FreeBSD do not.
6176 static void tcp_ecn_create_request(struct request_sock *req,
6177 const struct sk_buff *skb,
6178 const struct sock *listen_sk,
6179 const struct dst_entry *dst)
6181 const struct tcphdr *th = tcp_hdr(skb);
6182 const struct net *net = sock_net(listen_sk);
6183 bool th_ecn = th->ece && th->cwr;
6190 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
6191 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK);
6192 ecn_ok = net->ipv4.sysctl_tcp_ecn || ecn_ok_dst;
6194 if ((!ect && ecn_ok) || tcp_ca_needs_ecn(listen_sk) ||
6195 (ecn_ok_dst & DST_FEATURE_ECN_CA))
6196 inet_rsk(req)->ecn_ok = 1;
6199 static void tcp_openreq_init(struct request_sock *req,
6200 const struct tcp_options_received *rx_opt,
6201 struct sk_buff *skb, const struct sock *sk)
6203 struct inet_request_sock *ireq = inet_rsk(req);
6205 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */
6207 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6208 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
6209 tcp_rsk(req)->snt_synack = tcp_clock_us();
6210 tcp_rsk(req)->last_oow_ack_time = 0;
6211 req->mss = rx_opt->mss_clamp;
6212 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6213 ireq->tstamp_ok = rx_opt->tstamp_ok;
6214 ireq->sack_ok = rx_opt->sack_ok;
6215 ireq->snd_wscale = rx_opt->snd_wscale;
6216 ireq->wscale_ok = rx_opt->wscale_ok;
6219 ireq->ir_rmt_port = tcp_hdr(skb)->source;
6220 ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6221 ireq->ir_mark = inet_request_mark(sk, skb);
6224 struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
6225 struct sock *sk_listener,
6226 bool attach_listener)
6228 struct request_sock *req = reqsk_alloc(ops, sk_listener,
6232 struct inet_request_sock *ireq = inet_rsk(req);
6234 kmemcheck_annotate_bitfield(ireq, flags);
6236 #if IS_ENABLED(CONFIG_IPV6)
6237 ireq->pktopts = NULL;
6239 atomic64_set(&ireq->ir_cookie, 0);
6240 ireq->ireq_state = TCP_NEW_SYN_RECV;
6241 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
6242 ireq->ireq_family = sk_listener->sk_family;
6247 EXPORT_SYMBOL(inet_reqsk_alloc);
6250 * Return true if a syncookie should be sent
6252 static bool tcp_syn_flood_action(const struct sock *sk,
6253 const struct sk_buff *skb,
6256 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
6257 const char *msg = "Dropping request";
6258 bool want_cookie = false;
6259 struct net *net = sock_net(sk);
6261 #ifdef CONFIG_SYN_COOKIES
6262 if (net->ipv4.sysctl_tcp_syncookies) {
6263 msg = "Sending cookies";
6265 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
6268 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
6270 if (!queue->synflood_warned &&
6271 net->ipv4.sysctl_tcp_syncookies != 2 &&
6272 xchg(&queue->synflood_warned, 1) == 0)
6273 pr_info("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6274 proto, ntohs(tcp_hdr(skb)->dest), msg);
6279 static void tcp_reqsk_record_syn(const struct sock *sk,
6280 struct request_sock *req,
6281 const struct sk_buff *skb)
6283 if (tcp_sk(sk)->save_syn) {
6284 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
6287 copy = kmalloc(len + sizeof(u32), GFP_ATOMIC);
6290 memcpy(©[1], skb_network_header(skb), len);
6291 req->saved_syn = copy;
6296 int tcp_conn_request(struct request_sock_ops *rsk_ops,
6297 const struct tcp_request_sock_ops *af_ops,
6298 struct sock *sk, struct sk_buff *skb)
6300 struct tcp_fastopen_cookie foc = { .len = -1 };
6301 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
6302 struct tcp_options_received tmp_opt;
6303 struct tcp_sock *tp = tcp_sk(sk);
6304 struct net *net = sock_net(sk);
6305 struct sock *fastopen_sk = NULL;
6306 struct dst_entry *dst = NULL;
6307 struct request_sock *req;
6308 bool want_cookie = false;
6311 /* TW buckets are converted to open requests without
6312 * limitations, they conserve resources and peer is
6313 * evidently real one.
6315 if ((net->ipv4.sysctl_tcp_syncookies == 2 ||
6316 inet_csk_reqsk_queue_is_full(sk)) && !isn) {
6317 want_cookie = tcp_syn_flood_action(sk, skb, rsk_ops->slab_name);
6322 if (sk_acceptq_is_full(sk)) {
6323 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6327 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie);
6331 tcp_rsk(req)->af_specific = af_ops;
6332 tcp_rsk(req)->ts_off = 0;
6334 tcp_clear_options(&tmp_opt);
6335 tmp_opt.mss_clamp = af_ops->mss_clamp;
6336 tmp_opt.user_mss = tp->rx_opt.user_mss;
6337 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0,
6338 want_cookie ? NULL : &foc);
6340 if (want_cookie && !tmp_opt.saw_tstamp)
6341 tcp_clear_options(&tmp_opt);
6343 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
6344 tcp_openreq_init(req, &tmp_opt, skb, sk);
6345 inet_rsk(req)->no_srccheck = inet_sk(sk)->transparent;
6347 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6348 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb);
6350 af_ops->init_req(req, sk, skb);
6352 if (security_inet_conn_request(sk, skb, req))
6355 if (tmp_opt.tstamp_ok)
6356 tcp_rsk(req)->ts_off = af_ops->init_ts_off(skb);
6358 if (!want_cookie && !isn) {
6359 /* Kill the following clause, if you dislike this way. */
6360 if (!net->ipv4.sysctl_tcp_syncookies &&
6361 (net->ipv4.sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
6362 (net->ipv4.sysctl_max_syn_backlog >> 2)) &&
6363 !tcp_peer_is_proven(req, dst)) {
6364 /* Without syncookies last quarter of
6365 * backlog is filled with destinations,
6366 * proven to be alive.
6367 * It means that we continue to communicate
6368 * to destinations, already remembered
6369 * to the moment of synflood.
6371 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
6373 goto drop_and_release;
6376 isn = af_ops->init_seq(skb);
6379 dst = af_ops->route_req(sk, &fl, req);
6384 tcp_ecn_create_request(req, skb, sk, dst);
6387 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
6388 req->cookie_ts = tmp_opt.tstamp_ok;
6389 if (!tmp_opt.tstamp_ok)
6390 inet_rsk(req)->ecn_ok = 0;
6393 tcp_rsk(req)->snt_isn = isn;
6394 tcp_rsk(req)->txhash = net_tx_rndhash();
6395 tcp_openreq_init_rwin(req, sk, dst);
6397 tcp_reqsk_record_syn(sk, req, skb);
6398 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst);
6401 af_ops->send_synack(fastopen_sk, dst, &fl, req,
6402 &foc, TCP_SYNACK_FASTOPEN);
6403 /* Add the child socket directly into the accept queue */
6404 inet_csk_reqsk_queue_add(sk, req, fastopen_sk);
6405 sk->sk_data_ready(sk);
6406 bh_unlock_sock(fastopen_sk);
6407 sock_put(fastopen_sk);
6409 tcp_rsk(req)->tfo_listener = false;
6411 inet_csk_reqsk_queue_hash_add(sk, req, TCP_TIMEOUT_INIT);
6412 af_ops->send_synack(sk, dst, &fl, req, &foc,
6413 !want_cookie ? TCP_SYNACK_NORMAL :
6431 EXPORT_SYMBOL(tcp_conn_request);