1 // SPDX-License-Identifier: GPL-2.0
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
7 * Implementation of the Transmission Control Protocol(TCP).
10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11 * Mark Evans, <evansmp@uhura.aston.ac.uk>
12 * Corey Minyard <wf-rch!minyard@relay.EU.net>
13 * Florian La Roche, <flla@stud.uni-sb.de>
14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15 * Linus Torvalds, <torvalds@cs.helsinki.fi>
16 * Alan Cox, <gw4pts@gw4pts.ampr.org>
17 * Matthew Dillon, <dillon@apollo.west.oic.com>
18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19 * Jorge Cwik, <jorge@laser.satlink.net>
24 * Pedro Roque : Fast Retransmit/Recovery.
26 * Retransmit queue handled by TCP.
27 * Better retransmit timer handling.
28 * New congestion avoidance.
32 * Eric : Fast Retransmit.
33 * Randy Scott : MSS option defines.
34 * Eric Schenk : Fixes to slow start algorithm.
35 * Eric Schenk : Yet another double ACK bug.
36 * Eric Schenk : Delayed ACK bug fixes.
37 * Eric Schenk : Floyd style fast retrans war avoidance.
38 * David S. Miller : Don't allow zero congestion window.
39 * Eric Schenk : Fix retransmitter so that it sends
40 * next packet on ack of previous packet.
41 * Andi Kleen : Moved open_request checking here
42 * and process RSTs for open_requests.
43 * Andi Kleen : Better prune_queue, and other fixes.
44 * Andrey Savochkin: Fix RTT measurements in the presence of
46 * Andrey Savochkin: Check sequence numbers correctly when
47 * removing SACKs due to in sequence incoming
49 * Andi Kleen: Make sure we never ack data there is not
50 * enough room for. Also make this condition
51 * a fatal error if it might still happen.
52 * Andi Kleen: Add tcp_measure_rcv_mss to make
53 * connections with MSS<min(MTU,ann. MSS)
54 * work without delayed acks.
55 * Andi Kleen: Process packets with PSH set in the
57 * J Hadi Salim: ECN support
60 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
61 * engine. Lots of bugs are found.
62 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
65 #define pr_fmt(fmt) "TCP: " fmt
68 #include <linux/slab.h>
69 #include <linux/module.h>
70 #include <linux/sysctl.h>
71 #include <linux/kernel.h>
72 #include <linux/prefetch.h>
75 #include <net/inet_common.h>
76 #include <linux/ipsec.h>
77 #include <asm/unaligned.h>
78 #include <linux/errqueue.h>
79 #include <trace/events/tcp.h>
80 #include <linux/jump_label_ratelimit.h>
81 #include <net/busy_poll.h>
82 #include <net/mptcp.h>
84 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
86 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
87 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
88 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
89 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
90 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
91 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
92 #define FLAG_ECE 0x40 /* ECE in this ACK */
93 #define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */
94 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
95 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
96 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
97 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
98 #define FLAG_SET_XMIT_TIMER 0x1000 /* Set TLP or RTO timer */
99 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
100 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
101 #define FLAG_NO_CHALLENGE_ACK 0x8000 /* do not call tcp_send_challenge_ack() */
102 #define FLAG_ACK_MAYBE_DELAYED 0x10000 /* Likely a delayed ACK */
104 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
105 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
106 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE|FLAG_DSACKING_ACK)
107 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
109 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
110 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
112 #define REXMIT_NONE 0 /* no loss recovery to do */
113 #define REXMIT_LOST 1 /* retransmit packets marked lost */
114 #define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */
116 #if IS_ENABLED(CONFIG_TLS_DEVICE)
117 static DEFINE_STATIC_KEY_DEFERRED_FALSE(clean_acked_data_enabled, HZ);
119 void clean_acked_data_enable(struct inet_connection_sock *icsk,
120 void (*cad)(struct sock *sk, u32 ack_seq))
122 icsk->icsk_clean_acked = cad;
123 static_branch_deferred_inc(&clean_acked_data_enabled);
125 EXPORT_SYMBOL_GPL(clean_acked_data_enable);
127 void clean_acked_data_disable(struct inet_connection_sock *icsk)
129 static_branch_slow_dec_deferred(&clean_acked_data_enabled);
130 icsk->icsk_clean_acked = NULL;
132 EXPORT_SYMBOL_GPL(clean_acked_data_disable);
134 void clean_acked_data_flush(void)
136 static_key_deferred_flush(&clean_acked_data_enabled);
138 EXPORT_SYMBOL_GPL(clean_acked_data_flush);
141 static void tcp_gro_dev_warn(struct sock *sk, const struct sk_buff *skb,
144 static bool __once __read_mostly;
147 struct net_device *dev;
152 dev = dev_get_by_index_rcu(sock_net(sk), skb->skb_iif);
153 if (!dev || len >= dev->mtu)
154 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n",
155 dev ? dev->name : "Unknown driver");
160 /* Adapt the MSS value used to make delayed ack decision to the
163 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
165 struct inet_connection_sock *icsk = inet_csk(sk);
166 const unsigned int lss = icsk->icsk_ack.last_seg_size;
169 icsk->icsk_ack.last_seg_size = 0;
171 /* skb->len may jitter because of SACKs, even if peer
172 * sends good full-sized frames.
174 len = skb_shinfo(skb)->gso_size ? : skb->len;
175 if (len >= icsk->icsk_ack.rcv_mss) {
176 icsk->icsk_ack.rcv_mss = min_t(unsigned int, len,
178 /* Account for possibly-removed options */
179 if (unlikely(len > icsk->icsk_ack.rcv_mss +
180 MAX_TCP_OPTION_SPACE))
181 tcp_gro_dev_warn(sk, skb, len);
183 /* Otherwise, we make more careful check taking into account,
184 * that SACKs block is variable.
186 * "len" is invariant segment length, including TCP header.
188 len += skb->data - skb_transport_header(skb);
189 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
190 /* If PSH is not set, packet should be
191 * full sized, provided peer TCP is not badly broken.
192 * This observation (if it is correct 8)) allows
193 * to handle super-low mtu links fairly.
195 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
196 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
197 /* Subtract also invariant (if peer is RFC compliant),
198 * tcp header plus fixed timestamp option length.
199 * Resulting "len" is MSS free of SACK jitter.
201 len -= tcp_sk(sk)->tcp_header_len;
202 icsk->icsk_ack.last_seg_size = len;
204 icsk->icsk_ack.rcv_mss = len;
208 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
209 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
210 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
214 static void tcp_incr_quickack(struct sock *sk, unsigned int max_quickacks)
216 struct inet_connection_sock *icsk = inet_csk(sk);
217 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
221 quickacks = min(quickacks, max_quickacks);
222 if (quickacks > icsk->icsk_ack.quick)
223 icsk->icsk_ack.quick = quickacks;
226 void tcp_enter_quickack_mode(struct sock *sk, unsigned int max_quickacks)
228 struct inet_connection_sock *icsk = inet_csk(sk);
230 tcp_incr_quickack(sk, max_quickacks);
231 inet_csk_exit_pingpong_mode(sk);
232 icsk->icsk_ack.ato = TCP_ATO_MIN;
234 EXPORT_SYMBOL(tcp_enter_quickack_mode);
236 /* Send ACKs quickly, if "quick" count is not exhausted
237 * and the session is not interactive.
240 static bool tcp_in_quickack_mode(struct sock *sk)
242 const struct inet_connection_sock *icsk = inet_csk(sk);
243 const struct dst_entry *dst = __sk_dst_get(sk);
245 return (dst && dst_metric(dst, RTAX_QUICKACK)) ||
246 (icsk->icsk_ack.quick && !inet_csk_in_pingpong_mode(sk));
249 static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
251 if (tp->ecn_flags & TCP_ECN_OK)
252 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
255 static void tcp_ecn_accept_cwr(struct sock *sk, const struct sk_buff *skb)
257 if (tcp_hdr(skb)->cwr) {
258 tcp_sk(sk)->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
260 /* If the sender is telling us it has entered CWR, then its
261 * cwnd may be very low (even just 1 packet), so we should ACK
264 if (TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq)
265 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
269 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
271 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
274 static void __tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
276 struct tcp_sock *tp = tcp_sk(sk);
278 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
279 case INET_ECN_NOT_ECT:
280 /* Funny extension: if ECT is not set on a segment,
281 * and we already seen ECT on a previous segment,
282 * it is probably a retransmit.
284 if (tp->ecn_flags & TCP_ECN_SEEN)
285 tcp_enter_quickack_mode(sk, 2);
288 if (tcp_ca_needs_ecn(sk))
289 tcp_ca_event(sk, CA_EVENT_ECN_IS_CE);
291 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
292 /* Better not delay acks, sender can have a very low cwnd */
293 tcp_enter_quickack_mode(sk, 2);
294 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
296 tp->ecn_flags |= TCP_ECN_SEEN;
299 if (tcp_ca_needs_ecn(sk))
300 tcp_ca_event(sk, CA_EVENT_ECN_NO_CE);
301 tp->ecn_flags |= TCP_ECN_SEEN;
306 static void tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
308 if (tcp_sk(sk)->ecn_flags & TCP_ECN_OK)
309 __tcp_ecn_check_ce(sk, skb);
312 static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
314 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
315 tp->ecn_flags &= ~TCP_ECN_OK;
318 static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
320 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
321 tp->ecn_flags &= ~TCP_ECN_OK;
324 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
326 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
331 /* Buffer size and advertised window tuning.
333 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
336 static void tcp_sndbuf_expand(struct sock *sk)
338 const struct tcp_sock *tp = tcp_sk(sk);
339 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
343 /* Worst case is non GSO/TSO : each frame consumes one skb
344 * and skb->head is kmalloced using power of two area of memory
346 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
348 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
350 per_mss = roundup_pow_of_two(per_mss) +
351 SKB_DATA_ALIGN(sizeof(struct sk_buff));
353 nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd);
354 nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
356 /* Fast Recovery (RFC 5681 3.2) :
357 * Cubic needs 1.7 factor, rounded to 2 to include
358 * extra cushion (application might react slowly to EPOLLOUT)
360 sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2;
361 sndmem *= nr_segs * per_mss;
363 if (sk->sk_sndbuf < sndmem)
364 WRITE_ONCE(sk->sk_sndbuf,
365 min(sndmem, sock_net(sk)->ipv4.sysctl_tcp_wmem[2]));
368 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
370 * All tcp_full_space() is split to two parts: "network" buffer, allocated
371 * forward and advertised in receiver window (tp->rcv_wnd) and
372 * "application buffer", required to isolate scheduling/application
373 * latencies from network.
374 * window_clamp is maximal advertised window. It can be less than
375 * tcp_full_space(), in this case tcp_full_space() - window_clamp
376 * is reserved for "application" buffer. The less window_clamp is
377 * the smoother our behaviour from viewpoint of network, but the lower
378 * throughput and the higher sensitivity of the connection to losses. 8)
380 * rcv_ssthresh is more strict window_clamp used at "slow start"
381 * phase to predict further behaviour of this connection.
382 * It is used for two goals:
383 * - to enforce header prediction at sender, even when application
384 * requires some significant "application buffer". It is check #1.
385 * - to prevent pruning of receive queue because of misprediction
386 * of receiver window. Check #2.
388 * The scheme does not work when sender sends good segments opening
389 * window and then starts to feed us spaghetti. But it should work
390 * in common situations. Otherwise, we have to rely on queue collapsing.
393 /* Slow part of check#2. */
394 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
396 struct tcp_sock *tp = tcp_sk(sk);
398 int truesize = tcp_win_from_space(sk, skb->truesize) >> 1;
399 int window = tcp_win_from_space(sk, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]) >> 1;
401 while (tp->rcv_ssthresh <= window) {
402 if (truesize <= skb->len)
403 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
411 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
413 struct tcp_sock *tp = tcp_sk(sk);
416 room = min_t(int, tp->window_clamp, tcp_space(sk)) - tp->rcv_ssthresh;
419 if (room > 0 && !tcp_under_memory_pressure(sk)) {
422 /* Check #2. Increase window, if skb with such overhead
423 * will fit to rcvbuf in future.
425 if (tcp_win_from_space(sk, skb->truesize) <= skb->len)
426 incr = 2 * tp->advmss;
428 incr = __tcp_grow_window(sk, skb);
431 incr = max_t(int, incr, 2 * skb->len);
432 tp->rcv_ssthresh += min(room, incr);
433 inet_csk(sk)->icsk_ack.quick |= 1;
438 /* 3. Try to fixup all. It is made immediately after connection enters
441 static void tcp_init_buffer_space(struct sock *sk)
443 int tcp_app_win = sock_net(sk)->ipv4.sysctl_tcp_app_win;
444 struct tcp_sock *tp = tcp_sk(sk);
447 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
448 tcp_sndbuf_expand(sk);
450 tp->rcvq_space.space = min_t(u32, tp->rcv_wnd, TCP_INIT_CWND * tp->advmss);
451 tcp_mstamp_refresh(tp);
452 tp->rcvq_space.time = tp->tcp_mstamp;
453 tp->rcvq_space.seq = tp->copied_seq;
455 maxwin = tcp_full_space(sk);
457 if (tp->window_clamp >= maxwin) {
458 tp->window_clamp = maxwin;
460 if (tcp_app_win && maxwin > 4 * tp->advmss)
461 tp->window_clamp = max(maxwin -
462 (maxwin >> tcp_app_win),
466 /* Force reservation of one segment. */
468 tp->window_clamp > 2 * tp->advmss &&
469 tp->window_clamp + tp->advmss > maxwin)
470 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
472 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
473 tp->snd_cwnd_stamp = tcp_jiffies32;
476 /* 4. Recalculate window clamp after socket hit its memory bounds. */
477 static void tcp_clamp_window(struct sock *sk)
479 struct tcp_sock *tp = tcp_sk(sk);
480 struct inet_connection_sock *icsk = inet_csk(sk);
481 struct net *net = sock_net(sk);
483 icsk->icsk_ack.quick = 0;
485 if (sk->sk_rcvbuf < net->ipv4.sysctl_tcp_rmem[2] &&
486 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
487 !tcp_under_memory_pressure(sk) &&
488 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
489 WRITE_ONCE(sk->sk_rcvbuf,
490 min(atomic_read(&sk->sk_rmem_alloc),
491 net->ipv4.sysctl_tcp_rmem[2]));
493 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
494 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
497 /* Initialize RCV_MSS value.
498 * RCV_MSS is an our guess about MSS used by the peer.
499 * We haven't any direct information about the MSS.
500 * It's better to underestimate the RCV_MSS rather than overestimate.
501 * Overestimations make us ACKing less frequently than needed.
502 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
504 void tcp_initialize_rcv_mss(struct sock *sk)
506 const struct tcp_sock *tp = tcp_sk(sk);
507 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
509 hint = min(hint, tp->rcv_wnd / 2);
510 hint = min(hint, TCP_MSS_DEFAULT);
511 hint = max(hint, TCP_MIN_MSS);
513 inet_csk(sk)->icsk_ack.rcv_mss = hint;
515 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
517 /* Receiver "autotuning" code.
519 * The algorithm for RTT estimation w/o timestamps is based on
520 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
521 * <https://public.lanl.gov/radiant/pubs.html#DRS>
523 * More detail on this code can be found at
524 * <http://staff.psc.edu/jheffner/>,
525 * though this reference is out of date. A new paper
528 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
530 u32 new_sample = tp->rcv_rtt_est.rtt_us;
533 if (new_sample != 0) {
534 /* If we sample in larger samples in the non-timestamp
535 * case, we could grossly overestimate the RTT especially
536 * with chatty applications or bulk transfer apps which
537 * are stalled on filesystem I/O.
539 * Also, since we are only going for a minimum in the
540 * non-timestamp case, we do not smooth things out
541 * else with timestamps disabled convergence takes too
545 m -= (new_sample >> 3);
553 /* No previous measure. */
557 tp->rcv_rtt_est.rtt_us = new_sample;
560 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
564 if (tp->rcv_rtt_est.time == 0)
566 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
568 delta_us = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcv_rtt_est.time);
571 tcp_rcv_rtt_update(tp, delta_us, 1);
574 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
575 tp->rcv_rtt_est.time = tp->tcp_mstamp;
578 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
579 const struct sk_buff *skb)
581 struct tcp_sock *tp = tcp_sk(sk);
583 if (tp->rx_opt.rcv_tsecr == tp->rcv_rtt_last_tsecr)
585 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
587 if (TCP_SKB_CB(skb)->end_seq -
588 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss) {
589 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
592 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) {
595 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
596 tcp_rcv_rtt_update(tp, delta_us, 0);
602 * This function should be called every time data is copied to user space.
603 * It calculates the appropriate TCP receive buffer space.
605 void tcp_rcv_space_adjust(struct sock *sk)
607 struct tcp_sock *tp = tcp_sk(sk);
611 trace_tcp_rcv_space_adjust(sk);
613 tcp_mstamp_refresh(tp);
614 time = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcvq_space.time);
615 if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0)
618 /* Number of bytes copied to user in last RTT */
619 copied = tp->copied_seq - tp->rcvq_space.seq;
620 if (copied <= tp->rcvq_space.space)
624 * copied = bytes received in previous RTT, our base window
625 * To cope with packet losses, we need a 2x factor
626 * To cope with slow start, and sender growing its cwin by 100 %
627 * every RTT, we need a 4x factor, because the ACK we are sending
628 * now is for the next RTT, not the current one :
629 * <prev RTT . ><current RTT .. ><next RTT .... >
632 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf &&
633 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
637 /* minimal window to cope with packet losses, assuming
638 * steady state. Add some cushion because of small variations.
640 rcvwin = ((u64)copied << 1) + 16 * tp->advmss;
642 /* Accommodate for sender rate increase (eg. slow start) */
643 grow = rcvwin * (copied - tp->rcvq_space.space);
644 do_div(grow, tp->rcvq_space.space);
645 rcvwin += (grow << 1);
647 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
648 while (tcp_win_from_space(sk, rcvmem) < tp->advmss)
651 do_div(rcvwin, tp->advmss);
652 rcvbuf = min_t(u64, rcvwin * rcvmem,
653 sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
654 if (rcvbuf > sk->sk_rcvbuf) {
655 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
657 /* Make the window clamp follow along. */
658 tp->window_clamp = tcp_win_from_space(sk, rcvbuf);
661 tp->rcvq_space.space = copied;
664 tp->rcvq_space.seq = tp->copied_seq;
665 tp->rcvq_space.time = tp->tcp_mstamp;
668 /* There is something which you must keep in mind when you analyze the
669 * behavior of the tp->ato delayed ack timeout interval. When a
670 * connection starts up, we want to ack as quickly as possible. The
671 * problem is that "good" TCP's do slow start at the beginning of data
672 * transmission. The means that until we send the first few ACK's the
673 * sender will sit on his end and only queue most of his data, because
674 * he can only send snd_cwnd unacked packets at any given time. For
675 * each ACK we send, he increments snd_cwnd and transmits more of his
678 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
680 struct tcp_sock *tp = tcp_sk(sk);
681 struct inet_connection_sock *icsk = inet_csk(sk);
684 inet_csk_schedule_ack(sk);
686 tcp_measure_rcv_mss(sk, skb);
688 tcp_rcv_rtt_measure(tp);
692 if (!icsk->icsk_ack.ato) {
693 /* The _first_ data packet received, initialize
694 * delayed ACK engine.
696 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
697 icsk->icsk_ack.ato = TCP_ATO_MIN;
699 int m = now - icsk->icsk_ack.lrcvtime;
701 if (m <= TCP_ATO_MIN / 2) {
702 /* The fastest case is the first. */
703 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
704 } else if (m < icsk->icsk_ack.ato) {
705 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
706 if (icsk->icsk_ack.ato > icsk->icsk_rto)
707 icsk->icsk_ack.ato = icsk->icsk_rto;
708 } else if (m > icsk->icsk_rto) {
709 /* Too long gap. Apparently sender failed to
710 * restart window, so that we send ACKs quickly.
712 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
716 icsk->icsk_ack.lrcvtime = now;
718 tcp_ecn_check_ce(sk, skb);
721 tcp_grow_window(sk, skb);
724 /* Called to compute a smoothed rtt estimate. The data fed to this
725 * routine either comes from timestamps, or from segments that were
726 * known _not_ to have been retransmitted [see Karn/Partridge
727 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
728 * piece by Van Jacobson.
729 * NOTE: the next three routines used to be one big routine.
730 * To save cycles in the RFC 1323 implementation it was better to break
731 * it up into three procedures. -- erics
733 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
735 struct tcp_sock *tp = tcp_sk(sk);
736 long m = mrtt_us; /* RTT */
737 u32 srtt = tp->srtt_us;
739 /* The following amusing code comes from Jacobson's
740 * article in SIGCOMM '88. Note that rtt and mdev
741 * are scaled versions of rtt and mean deviation.
742 * This is designed to be as fast as possible
743 * m stands for "measurement".
745 * On a 1990 paper the rto value is changed to:
746 * RTO = rtt + 4 * mdev
748 * Funny. This algorithm seems to be very broken.
749 * These formulae increase RTO, when it should be decreased, increase
750 * too slowly, when it should be increased quickly, decrease too quickly
751 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
752 * does not matter how to _calculate_ it. Seems, it was trap
753 * that VJ failed to avoid. 8)
756 m -= (srtt >> 3); /* m is now error in rtt est */
757 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
759 m = -m; /* m is now abs(error) */
760 m -= (tp->mdev_us >> 2); /* similar update on mdev */
761 /* This is similar to one of Eifel findings.
762 * Eifel blocks mdev updates when rtt decreases.
763 * This solution is a bit different: we use finer gain
764 * for mdev in this case (alpha*beta).
765 * Like Eifel it also prevents growth of rto,
766 * but also it limits too fast rto decreases,
767 * happening in pure Eifel.
772 m -= (tp->mdev_us >> 2); /* similar update on mdev */
774 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
775 if (tp->mdev_us > tp->mdev_max_us) {
776 tp->mdev_max_us = tp->mdev_us;
777 if (tp->mdev_max_us > tp->rttvar_us)
778 tp->rttvar_us = tp->mdev_max_us;
780 if (after(tp->snd_una, tp->rtt_seq)) {
781 if (tp->mdev_max_us < tp->rttvar_us)
782 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
783 tp->rtt_seq = tp->snd_nxt;
784 tp->mdev_max_us = tcp_rto_min_us(sk);
789 /* no previous measure. */
790 srtt = m << 3; /* take the measured time to be rtt */
791 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
792 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
793 tp->mdev_max_us = tp->rttvar_us;
794 tp->rtt_seq = tp->snd_nxt;
798 tp->srtt_us = max(1U, srtt);
801 static void tcp_update_pacing_rate(struct sock *sk)
803 const struct tcp_sock *tp = tcp_sk(sk);
806 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
807 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
809 /* current rate is (cwnd * mss) / srtt
810 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
811 * In Congestion Avoidance phase, set it to 120 % the current rate.
813 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
814 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
815 * end of slow start and should slow down.
817 if (tp->snd_cwnd < tp->snd_ssthresh / 2)
818 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ss_ratio;
820 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ca_ratio;
822 rate *= max(tp->snd_cwnd, tp->packets_out);
824 if (likely(tp->srtt_us))
825 do_div(rate, tp->srtt_us);
827 /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate
828 * without any lock. We want to make sure compiler wont store
829 * intermediate values in this location.
831 WRITE_ONCE(sk->sk_pacing_rate, min_t(u64, rate,
832 sk->sk_max_pacing_rate));
835 /* Calculate rto without backoff. This is the second half of Van Jacobson's
836 * routine referred to above.
838 static void tcp_set_rto(struct sock *sk)
840 const struct tcp_sock *tp = tcp_sk(sk);
841 /* Old crap is replaced with new one. 8)
844 * 1. If rtt variance happened to be less 50msec, it is hallucination.
845 * It cannot be less due to utterly erratic ACK generation made
846 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
847 * to do with delayed acks, because at cwnd>2 true delack timeout
848 * is invisible. Actually, Linux-2.4 also generates erratic
849 * ACKs in some circumstances.
851 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
853 /* 2. Fixups made earlier cannot be right.
854 * If we do not estimate RTO correctly without them,
855 * all the algo is pure shit and should be replaced
856 * with correct one. It is exactly, which we pretend to do.
859 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
860 * guarantees that rto is higher.
865 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
867 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
870 cwnd = TCP_INIT_CWND;
871 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
874 struct tcp_sacktag_state {
875 /* Timestamps for earliest and latest never-retransmitted segment
876 * that was SACKed. RTO needs the earliest RTT to stay conservative,
877 * but congestion control should still get an accurate delay signal.
884 unsigned int mss_now;
885 struct rate_sample *rate;
888 /* Take a notice that peer is sending D-SACKs */
889 static u32 tcp_dsack_seen(struct tcp_sock *tp, u32 start_seq,
890 u32 end_seq, struct tcp_sacktag_state *state)
892 u32 seq_len, dup_segs = 1;
894 if (before(start_seq, end_seq)) {
895 seq_len = end_seq - start_seq;
896 if (seq_len > tp->mss_cache)
897 dup_segs = DIV_ROUND_UP(seq_len, tp->mss_cache);
900 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
901 tp->rack.dsack_seen = 1;
902 tp->dsack_dups += dup_segs;
904 state->flag |= FLAG_DSACKING_ACK;
905 /* A spurious retransmission is delivered */
906 state->sack_delivered += dup_segs;
911 /* It's reordering when higher sequence was delivered (i.e. sacked) before
912 * some lower never-retransmitted sequence ("low_seq"). The maximum reordering
913 * distance is approximated in full-mss packet distance ("reordering").
915 static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq,
918 struct tcp_sock *tp = tcp_sk(sk);
919 const u32 mss = tp->mss_cache;
922 fack = tcp_highest_sack_seq(tp);
923 if (!before(low_seq, fack))
926 metric = fack - low_seq;
927 if ((metric > tp->reordering * mss) && mss) {
928 #if FASTRETRANS_DEBUG > 1
929 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
930 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
934 tp->undo_marker ? tp->undo_retrans : 0);
936 tp->reordering = min_t(u32, (metric + mss - 1) / mss,
937 sock_net(sk)->ipv4.sysctl_tcp_max_reordering);
940 /* This exciting event is worth to be remembered. 8) */
942 NET_INC_STATS(sock_net(sk),
943 ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER);
946 /* This must be called before lost_out is incremented */
947 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
949 if ((!tp->retransmit_skb_hint && tp->retrans_out >= tp->lost_out) ||
950 (tp->retransmit_skb_hint &&
951 before(TCP_SKB_CB(skb)->seq,
952 TCP_SKB_CB(tp->retransmit_skb_hint)->seq)))
953 tp->retransmit_skb_hint = skb;
956 /* Sum the number of packets on the wire we have marked as lost.
957 * There are two cases we care about here:
958 * a) Packet hasn't been marked lost (nor retransmitted),
959 * and this is the first loss.
960 * b) Packet has been marked both lost and retransmitted,
961 * and this means we think it was lost again.
963 static void tcp_sum_lost(struct tcp_sock *tp, struct sk_buff *skb)
965 __u8 sacked = TCP_SKB_CB(skb)->sacked;
967 if (!(sacked & TCPCB_LOST) ||
968 ((sacked & TCPCB_LOST) && (sacked & TCPCB_SACKED_RETRANS)))
969 tp->lost += tcp_skb_pcount(skb);
972 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
974 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
975 tcp_verify_retransmit_hint(tp, skb);
977 tp->lost_out += tcp_skb_pcount(skb);
978 tcp_sum_lost(tp, skb);
979 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
983 void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, struct sk_buff *skb)
985 tcp_verify_retransmit_hint(tp, skb);
987 tcp_sum_lost(tp, skb);
988 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
989 tp->lost_out += tcp_skb_pcount(skb);
990 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
994 /* Updates the delivered and delivered_ce counts */
995 static void tcp_count_delivered(struct tcp_sock *tp, u32 delivered,
998 tp->delivered += delivered;
1000 tp->delivered_ce += delivered;
1003 /* This procedure tags the retransmission queue when SACKs arrive.
1005 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1006 * Packets in queue with these bits set are counted in variables
1007 * sacked_out, retrans_out and lost_out, correspondingly.
1009 * Valid combinations are:
1010 * Tag InFlight Description
1011 * 0 1 - orig segment is in flight.
1012 * S 0 - nothing flies, orig reached receiver.
1013 * L 0 - nothing flies, orig lost by net.
1014 * R 2 - both orig and retransmit are in flight.
1015 * L|R 1 - orig is lost, retransmit is in flight.
1016 * S|R 1 - orig reached receiver, retrans is still in flight.
1017 * (L|S|R is logically valid, it could occur when L|R is sacked,
1018 * but it is equivalent to plain S and code short-curcuits it to S.
1019 * L|S is logically invalid, it would mean -1 packet in flight 8))
1021 * These 6 states form finite state machine, controlled by the following events:
1022 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1023 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1024 * 3. Loss detection event of two flavors:
1025 * A. Scoreboard estimator decided the packet is lost.
1026 * A'. Reno "three dupacks" marks head of queue lost.
1027 * B. SACK arrives sacking SND.NXT at the moment, when the
1028 * segment was retransmitted.
1029 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1031 * It is pleasant to note, that state diagram turns out to be commutative,
1032 * so that we are allowed not to be bothered by order of our actions,
1033 * when multiple events arrive simultaneously. (see the function below).
1035 * Reordering detection.
1036 * --------------------
1037 * Reordering metric is maximal distance, which a packet can be displaced
1038 * in packet stream. With SACKs we can estimate it:
1040 * 1. SACK fills old hole and the corresponding segment was not
1041 * ever retransmitted -> reordering. Alas, we cannot use it
1042 * when segment was retransmitted.
1043 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1044 * for retransmitted and already SACKed segment -> reordering..
1045 * Both of these heuristics are not used in Loss state, when we cannot
1046 * account for retransmits accurately.
1048 * SACK block validation.
1049 * ----------------------
1051 * SACK block range validation checks that the received SACK block fits to
1052 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1053 * Note that SND.UNA is not included to the range though being valid because
1054 * it means that the receiver is rather inconsistent with itself reporting
1055 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1056 * perfectly valid, however, in light of RFC2018 which explicitly states
1057 * that "SACK block MUST reflect the newest segment. Even if the newest
1058 * segment is going to be discarded ...", not that it looks very clever
1059 * in case of head skb. Due to potentional receiver driven attacks, we
1060 * choose to avoid immediate execution of a walk in write queue due to
1061 * reneging and defer head skb's loss recovery to standard loss recovery
1062 * procedure that will eventually trigger (nothing forbids us doing this).
1064 * Implements also blockage to start_seq wrap-around. Problem lies in the
1065 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1066 * there's no guarantee that it will be before snd_nxt (n). The problem
1067 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1070 * <- outs wnd -> <- wrapzone ->
1071 * u e n u_w e_w s n_w
1073 * |<------------+------+----- TCP seqno space --------------+---------->|
1074 * ...-- <2^31 ->| |<--------...
1075 * ...---- >2^31 ------>| |<--------...
1077 * Current code wouldn't be vulnerable but it's better still to discard such
1078 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1079 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1080 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1081 * equal to the ideal case (infinite seqno space without wrap caused issues).
1083 * With D-SACK the lower bound is extended to cover sequence space below
1084 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1085 * again, D-SACK block must not to go across snd_una (for the same reason as
1086 * for the normal SACK blocks, explained above). But there all simplicity
1087 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1088 * fully below undo_marker they do not affect behavior in anyway and can
1089 * therefore be safely ignored. In rare cases (which are more or less
1090 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1091 * fragmentation and packet reordering past skb's retransmission. To consider
1092 * them correctly, the acceptable range must be extended even more though
1093 * the exact amount is rather hard to quantify. However, tp->max_window can
1094 * be used as an exaggerated estimate.
1096 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1097 u32 start_seq, u32 end_seq)
1099 /* Too far in future, or reversed (interpretation is ambiguous) */
1100 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1103 /* Nasty start_seq wrap-around check (see comments above) */
1104 if (!before(start_seq, tp->snd_nxt))
1107 /* In outstanding window? ...This is valid exit for D-SACKs too.
1108 * start_seq == snd_una is non-sensical (see comments above)
1110 if (after(start_seq, tp->snd_una))
1113 if (!is_dsack || !tp->undo_marker)
1116 /* ...Then it's D-SACK, and must reside below snd_una completely */
1117 if (after(end_seq, tp->snd_una))
1120 if (!before(start_seq, tp->undo_marker))
1124 if (!after(end_seq, tp->undo_marker))
1127 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1128 * start_seq < undo_marker and end_seq >= undo_marker.
1130 return !before(start_seq, end_seq - tp->max_window);
1133 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1134 struct tcp_sack_block_wire *sp, int num_sacks,
1135 u32 prior_snd_una, struct tcp_sacktag_state *state)
1137 struct tcp_sock *tp = tcp_sk(sk);
1138 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1139 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1142 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1143 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1144 } else if (num_sacks > 1) {
1145 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1146 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1148 if (after(end_seq_0, end_seq_1) || before(start_seq_0, start_seq_1))
1150 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKOFORECV);
1155 dup_segs = tcp_dsack_seen(tp, start_seq_0, end_seq_0, state);
1156 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECVSEGS, dup_segs);
1158 /* D-SACK for already forgotten data... Do dumb counting. */
1159 if (tp->undo_marker && tp->undo_retrans > 0 &&
1160 !after(end_seq_0, prior_snd_una) &&
1161 after(end_seq_0, tp->undo_marker))
1162 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - dup_segs);
1167 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1168 * the incoming SACK may not exactly match but we can find smaller MSS
1169 * aligned portion of it that matches. Therefore we might need to fragment
1170 * which may fail and creates some hassle (caller must handle error case
1173 * FIXME: this could be merged to shift decision code
1175 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1176 u32 start_seq, u32 end_seq)
1180 unsigned int pkt_len;
1183 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1184 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1186 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1187 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1188 mss = tcp_skb_mss(skb);
1189 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1192 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1196 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1201 /* Round if necessary so that SACKs cover only full MSSes
1202 * and/or the remaining small portion (if present)
1204 if (pkt_len > mss) {
1205 unsigned int new_len = (pkt_len / mss) * mss;
1206 if (!in_sack && new_len < pkt_len)
1211 if (pkt_len >= skb->len && !in_sack)
1214 err = tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
1215 pkt_len, mss, GFP_ATOMIC);
1223 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1224 static u8 tcp_sacktag_one(struct sock *sk,
1225 struct tcp_sacktag_state *state, u8 sacked,
1226 u32 start_seq, u32 end_seq,
1227 int dup_sack, int pcount,
1230 struct tcp_sock *tp = tcp_sk(sk);
1232 /* Account D-SACK for retransmitted packet. */
1233 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1234 if (tp->undo_marker && tp->undo_retrans > 0 &&
1235 after(end_seq, tp->undo_marker))
1237 if ((sacked & TCPCB_SACKED_ACKED) &&
1238 before(start_seq, state->reord))
1239 state->reord = start_seq;
1242 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1243 if (!after(end_seq, tp->snd_una))
1246 if (!(sacked & TCPCB_SACKED_ACKED)) {
1247 tcp_rack_advance(tp, sacked, end_seq, xmit_time);
1249 if (sacked & TCPCB_SACKED_RETRANS) {
1250 /* If the segment is not tagged as lost,
1251 * we do not clear RETRANS, believing
1252 * that retransmission is still in flight.
1254 if (sacked & TCPCB_LOST) {
1255 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1256 tp->lost_out -= pcount;
1257 tp->retrans_out -= pcount;
1260 if (!(sacked & TCPCB_RETRANS)) {
1261 /* New sack for not retransmitted frame,
1262 * which was in hole. It is reordering.
1264 if (before(start_seq,
1265 tcp_highest_sack_seq(tp)) &&
1266 before(start_seq, state->reord))
1267 state->reord = start_seq;
1269 if (!after(end_seq, tp->high_seq))
1270 state->flag |= FLAG_ORIG_SACK_ACKED;
1271 if (state->first_sackt == 0)
1272 state->first_sackt = xmit_time;
1273 state->last_sackt = xmit_time;
1276 if (sacked & TCPCB_LOST) {
1277 sacked &= ~TCPCB_LOST;
1278 tp->lost_out -= pcount;
1282 sacked |= TCPCB_SACKED_ACKED;
1283 state->flag |= FLAG_DATA_SACKED;
1284 tp->sacked_out += pcount;
1285 /* Out-of-order packets delivered */
1286 state->sack_delivered += pcount;
1288 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1289 if (tp->lost_skb_hint &&
1290 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1291 tp->lost_cnt_hint += pcount;
1294 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1295 * frames and clear it. undo_retrans is decreased above, L|R frames
1296 * are accounted above as well.
1298 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1299 sacked &= ~TCPCB_SACKED_RETRANS;
1300 tp->retrans_out -= pcount;
1306 /* Shift newly-SACKed bytes from this skb to the immediately previous
1307 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1309 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev,
1310 struct sk_buff *skb,
1311 struct tcp_sacktag_state *state,
1312 unsigned int pcount, int shifted, int mss,
1315 struct tcp_sock *tp = tcp_sk(sk);
1316 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1317 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1321 /* Adjust counters and hints for the newly sacked sequence
1322 * range but discard the return value since prev is already
1323 * marked. We must tag the range first because the seq
1324 * advancement below implicitly advances
1325 * tcp_highest_sack_seq() when skb is highest_sack.
1327 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1328 start_seq, end_seq, dup_sack, pcount,
1329 tcp_skb_timestamp_us(skb));
1330 tcp_rate_skb_delivered(sk, skb, state->rate);
1332 if (skb == tp->lost_skb_hint)
1333 tp->lost_cnt_hint += pcount;
1335 TCP_SKB_CB(prev)->end_seq += shifted;
1336 TCP_SKB_CB(skb)->seq += shifted;
1338 tcp_skb_pcount_add(prev, pcount);
1339 WARN_ON_ONCE(tcp_skb_pcount(skb) < pcount);
1340 tcp_skb_pcount_add(skb, -pcount);
1342 /* When we're adding to gso_segs == 1, gso_size will be zero,
1343 * in theory this shouldn't be necessary but as long as DSACK
1344 * code can come after this skb later on it's better to keep
1345 * setting gso_size to something.
1347 if (!TCP_SKB_CB(prev)->tcp_gso_size)
1348 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1350 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1351 if (tcp_skb_pcount(skb) <= 1)
1352 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1354 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1355 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1358 BUG_ON(!tcp_skb_pcount(skb));
1359 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1363 /* Whole SKB was eaten :-) */
1365 if (skb == tp->retransmit_skb_hint)
1366 tp->retransmit_skb_hint = prev;
1367 if (skb == tp->lost_skb_hint) {
1368 tp->lost_skb_hint = prev;
1369 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1372 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1373 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor;
1374 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1375 TCP_SKB_CB(prev)->end_seq++;
1377 if (skb == tcp_highest_sack(sk))
1378 tcp_advance_highest_sack(sk, skb);
1380 tcp_skb_collapse_tstamp(prev, skb);
1381 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp))
1382 TCP_SKB_CB(prev)->tx.delivered_mstamp = 0;
1384 tcp_rtx_queue_unlink_and_free(skb, sk);
1386 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED);
1391 /* I wish gso_size would have a bit more sane initialization than
1392 * something-or-zero which complicates things
1394 static int tcp_skb_seglen(const struct sk_buff *skb)
1396 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1399 /* Shifting pages past head area doesn't work */
1400 static int skb_can_shift(const struct sk_buff *skb)
1402 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1405 int tcp_skb_shift(struct sk_buff *to, struct sk_buff *from,
1406 int pcount, int shiftlen)
1408 /* TCP min gso_size is 8 bytes (TCP_MIN_GSO_SIZE)
1409 * Since TCP_SKB_CB(skb)->tcp_gso_segs is 16 bits, we need
1410 * to make sure not storing more than 65535 * 8 bytes per skb,
1411 * even if current MSS is bigger.
1413 if (unlikely(to->len + shiftlen >= 65535 * TCP_MIN_GSO_SIZE))
1415 if (unlikely(tcp_skb_pcount(to) + pcount > 65535))
1417 return skb_shift(to, from, shiftlen);
1420 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1423 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1424 struct tcp_sacktag_state *state,
1425 u32 start_seq, u32 end_seq,
1428 struct tcp_sock *tp = tcp_sk(sk);
1429 struct sk_buff *prev;
1435 /* Normally R but no L won't result in plain S */
1437 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1439 if (!skb_can_shift(skb))
1441 /* This frame is about to be dropped (was ACKed). */
1442 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1445 /* Can only happen with delayed DSACK + discard craziness */
1446 prev = skb_rb_prev(skb);
1450 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1453 if (!tcp_skb_can_collapse(prev, skb))
1456 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1457 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1461 pcount = tcp_skb_pcount(skb);
1462 mss = tcp_skb_seglen(skb);
1464 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1465 * drop this restriction as unnecessary
1467 if (mss != tcp_skb_seglen(prev))
1470 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1472 /* CHECKME: This is non-MSS split case only?, this will
1473 * cause skipped skbs due to advancing loop btw, original
1474 * has that feature too
1476 if (tcp_skb_pcount(skb) <= 1)
1479 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1481 /* TODO: head merge to next could be attempted here
1482 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1483 * though it might not be worth of the additional hassle
1485 * ...we can probably just fallback to what was done
1486 * previously. We could try merging non-SACKed ones
1487 * as well but it probably isn't going to buy off
1488 * because later SACKs might again split them, and
1489 * it would make skb timestamp tracking considerably
1495 len = end_seq - TCP_SKB_CB(skb)->seq;
1497 BUG_ON(len > skb->len);
1499 /* MSS boundaries should be honoured or else pcount will
1500 * severely break even though it makes things bit trickier.
1501 * Optimize common case to avoid most of the divides
1503 mss = tcp_skb_mss(skb);
1505 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1506 * drop this restriction as unnecessary
1508 if (mss != tcp_skb_seglen(prev))
1513 } else if (len < mss) {
1521 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1522 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1525 if (!tcp_skb_shift(prev, skb, pcount, len))
1527 if (!tcp_shifted_skb(sk, prev, skb, state, pcount, len, mss, dup_sack))
1530 /* Hole filled allows collapsing with the next as well, this is very
1531 * useful when hole on every nth skb pattern happens
1533 skb = skb_rb_next(prev);
1537 if (!skb_can_shift(skb) ||
1538 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1539 (mss != tcp_skb_seglen(skb)))
1543 pcount = tcp_skb_pcount(skb);
1544 if (tcp_skb_shift(prev, skb, pcount, len))
1545 tcp_shifted_skb(sk, prev, skb, state, pcount,
1555 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1559 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1560 struct tcp_sack_block *next_dup,
1561 struct tcp_sacktag_state *state,
1562 u32 start_seq, u32 end_seq,
1565 struct tcp_sock *tp = tcp_sk(sk);
1566 struct sk_buff *tmp;
1568 skb_rbtree_walk_from(skb) {
1570 bool dup_sack = dup_sack_in;
1572 /* queue is in-order => we can short-circuit the walk early */
1573 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1577 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1578 in_sack = tcp_match_skb_to_sack(sk, skb,
1579 next_dup->start_seq,
1585 /* skb reference here is a bit tricky to get right, since
1586 * shifting can eat and free both this skb and the next,
1587 * so not even _safe variant of the loop is enough.
1590 tmp = tcp_shift_skb_data(sk, skb, state,
1591 start_seq, end_seq, dup_sack);
1600 in_sack = tcp_match_skb_to_sack(sk, skb,
1606 if (unlikely(in_sack < 0))
1610 TCP_SKB_CB(skb)->sacked =
1613 TCP_SKB_CB(skb)->sacked,
1614 TCP_SKB_CB(skb)->seq,
1615 TCP_SKB_CB(skb)->end_seq,
1617 tcp_skb_pcount(skb),
1618 tcp_skb_timestamp_us(skb));
1619 tcp_rate_skb_delivered(sk, skb, state->rate);
1620 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1621 list_del_init(&skb->tcp_tsorted_anchor);
1623 if (!before(TCP_SKB_CB(skb)->seq,
1624 tcp_highest_sack_seq(tp)))
1625 tcp_advance_highest_sack(sk, skb);
1631 static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk, u32 seq)
1633 struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node;
1634 struct sk_buff *skb;
1638 skb = rb_to_skb(parent);
1639 if (before(seq, TCP_SKB_CB(skb)->seq)) {
1640 p = &parent->rb_left;
1643 if (!before(seq, TCP_SKB_CB(skb)->end_seq)) {
1644 p = &parent->rb_right;
1652 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1655 if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq))
1658 return tcp_sacktag_bsearch(sk, skip_to_seq);
1661 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1663 struct tcp_sack_block *next_dup,
1664 struct tcp_sacktag_state *state,
1670 if (before(next_dup->start_seq, skip_to_seq)) {
1671 skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq);
1672 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1673 next_dup->start_seq, next_dup->end_seq,
1680 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1682 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1686 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1687 u32 prior_snd_una, struct tcp_sacktag_state *state)
1689 struct tcp_sock *tp = tcp_sk(sk);
1690 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1691 TCP_SKB_CB(ack_skb)->sacked);
1692 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1693 struct tcp_sack_block sp[TCP_NUM_SACKS];
1694 struct tcp_sack_block *cache;
1695 struct sk_buff *skb;
1696 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1698 bool found_dup_sack = false;
1700 int first_sack_index;
1703 state->reord = tp->snd_nxt;
1705 if (!tp->sacked_out)
1706 tcp_highest_sack_reset(sk);
1708 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1709 num_sacks, prior_snd_una, state);
1711 /* Eliminate too old ACKs, but take into
1712 * account more or less fresh ones, they can
1713 * contain valid SACK info.
1715 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1718 if (!tp->packets_out)
1722 first_sack_index = 0;
1723 for (i = 0; i < num_sacks; i++) {
1724 bool dup_sack = !i && found_dup_sack;
1726 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1727 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1729 if (!tcp_is_sackblock_valid(tp, dup_sack,
1730 sp[used_sacks].start_seq,
1731 sp[used_sacks].end_seq)) {
1735 if (!tp->undo_marker)
1736 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1738 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1740 /* Don't count olds caused by ACK reordering */
1741 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1742 !after(sp[used_sacks].end_seq, tp->snd_una))
1744 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1747 NET_INC_STATS(sock_net(sk), mib_idx);
1749 first_sack_index = -1;
1753 /* Ignore very old stuff early */
1754 if (!after(sp[used_sacks].end_seq, prior_snd_una)) {
1756 first_sack_index = -1;
1763 /* order SACK blocks to allow in order walk of the retrans queue */
1764 for (i = used_sacks - 1; i > 0; i--) {
1765 for (j = 0; j < i; j++) {
1766 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1767 swap(sp[j], sp[j + 1]);
1769 /* Track where the first SACK block goes to */
1770 if (j == first_sack_index)
1771 first_sack_index = j + 1;
1776 state->mss_now = tcp_current_mss(sk);
1780 if (!tp->sacked_out) {
1781 /* It's already past, so skip checking against it */
1782 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1784 cache = tp->recv_sack_cache;
1785 /* Skip empty blocks in at head of the cache */
1786 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1791 while (i < used_sacks) {
1792 u32 start_seq = sp[i].start_seq;
1793 u32 end_seq = sp[i].end_seq;
1794 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1795 struct tcp_sack_block *next_dup = NULL;
1797 if (found_dup_sack && ((i + 1) == first_sack_index))
1798 next_dup = &sp[i + 1];
1800 /* Skip too early cached blocks */
1801 while (tcp_sack_cache_ok(tp, cache) &&
1802 !before(start_seq, cache->end_seq))
1805 /* Can skip some work by looking recv_sack_cache? */
1806 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1807 after(end_seq, cache->start_seq)) {
1810 if (before(start_seq, cache->start_seq)) {
1811 skb = tcp_sacktag_skip(skb, sk, start_seq);
1812 skb = tcp_sacktag_walk(skb, sk, next_dup,
1819 /* Rest of the block already fully processed? */
1820 if (!after(end_seq, cache->end_seq))
1823 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1827 /* ...tail remains todo... */
1828 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1829 /* ...but better entrypoint exists! */
1830 skb = tcp_highest_sack(sk);
1837 skb = tcp_sacktag_skip(skb, sk, cache->end_seq);
1838 /* Check overlap against next cached too (past this one already) */
1843 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1844 skb = tcp_highest_sack(sk);
1848 skb = tcp_sacktag_skip(skb, sk, start_seq);
1851 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
1852 start_seq, end_seq, dup_sack);
1858 /* Clear the head of the cache sack blocks so we can skip it next time */
1859 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1860 tp->recv_sack_cache[i].start_seq = 0;
1861 tp->recv_sack_cache[i].end_seq = 0;
1863 for (j = 0; j < used_sacks; j++)
1864 tp->recv_sack_cache[i++] = sp[j];
1866 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss || tp->undo_marker)
1867 tcp_check_sack_reordering(sk, state->reord, 0);
1869 tcp_verify_left_out(tp);
1872 #if FASTRETRANS_DEBUG > 0
1873 WARN_ON((int)tp->sacked_out < 0);
1874 WARN_ON((int)tp->lost_out < 0);
1875 WARN_ON((int)tp->retrans_out < 0);
1876 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1881 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1882 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1884 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1888 holes = max(tp->lost_out, 1U);
1889 holes = min(holes, tp->packets_out);
1891 if ((tp->sacked_out + holes) > tp->packets_out) {
1892 tp->sacked_out = tp->packets_out - holes;
1898 /* If we receive more dupacks than we expected counting segments
1899 * in assumption of absent reordering, interpret this as reordering.
1900 * The only another reason could be bug in receiver TCP.
1902 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1904 struct tcp_sock *tp = tcp_sk(sk);
1906 if (!tcp_limit_reno_sacked(tp))
1909 tp->reordering = min_t(u32, tp->packets_out + addend,
1910 sock_net(sk)->ipv4.sysctl_tcp_max_reordering);
1912 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER);
1915 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1917 static void tcp_add_reno_sack(struct sock *sk, int num_dupack, bool ece_ack)
1920 struct tcp_sock *tp = tcp_sk(sk);
1921 u32 prior_sacked = tp->sacked_out;
1924 tp->sacked_out += num_dupack;
1925 tcp_check_reno_reordering(sk, 0);
1926 delivered = tp->sacked_out - prior_sacked;
1928 tcp_count_delivered(tp, delivered, ece_ack);
1929 tcp_verify_left_out(tp);
1933 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1935 static void tcp_remove_reno_sacks(struct sock *sk, int acked, bool ece_ack)
1937 struct tcp_sock *tp = tcp_sk(sk);
1940 /* One ACK acked hole. The rest eat duplicate ACKs. */
1941 tcp_count_delivered(tp, max_t(int, acked - tp->sacked_out, 1),
1943 if (acked - 1 >= tp->sacked_out)
1946 tp->sacked_out -= acked - 1;
1948 tcp_check_reno_reordering(sk, acked);
1949 tcp_verify_left_out(tp);
1952 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1957 void tcp_clear_retrans(struct tcp_sock *tp)
1959 tp->retrans_out = 0;
1961 tp->undo_marker = 0;
1962 tp->undo_retrans = -1;
1966 static inline void tcp_init_undo(struct tcp_sock *tp)
1968 tp->undo_marker = tp->snd_una;
1969 /* Retransmission still in flight may cause DSACKs later. */
1970 tp->undo_retrans = tp->retrans_out ? : -1;
1973 static bool tcp_is_rack(const struct sock *sk)
1975 return sock_net(sk)->ipv4.sysctl_tcp_recovery & TCP_RACK_LOSS_DETECTION;
1978 /* If we detect SACK reneging, forget all SACK information
1979 * and reset tags completely, otherwise preserve SACKs. If receiver
1980 * dropped its ofo queue, we will know this due to reneging detection.
1982 static void tcp_timeout_mark_lost(struct sock *sk)
1984 struct tcp_sock *tp = tcp_sk(sk);
1985 struct sk_buff *skb, *head;
1986 bool is_reneg; /* is receiver reneging on SACKs? */
1988 head = tcp_rtx_queue_head(sk);
1989 is_reneg = head && (TCP_SKB_CB(head)->sacked & TCPCB_SACKED_ACKED);
1991 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
1993 /* Mark SACK reneging until we recover from this loss event. */
1994 tp->is_sack_reneg = 1;
1995 } else if (tcp_is_reno(tp)) {
1996 tcp_reset_reno_sack(tp);
2000 skb_rbtree_walk_from(skb) {
2002 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
2003 else if (tcp_is_rack(sk) && skb != head &&
2004 tcp_rack_skb_timeout(tp, skb, 0) > 0)
2005 continue; /* Don't mark recently sent ones lost yet */
2006 tcp_mark_skb_lost(sk, skb);
2008 tcp_verify_left_out(tp);
2009 tcp_clear_all_retrans_hints(tp);
2012 /* Enter Loss state. */
2013 void tcp_enter_loss(struct sock *sk)
2015 const struct inet_connection_sock *icsk = inet_csk(sk);
2016 struct tcp_sock *tp = tcp_sk(sk);
2017 struct net *net = sock_net(sk);
2018 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
2020 tcp_timeout_mark_lost(sk);
2022 /* Reduce ssthresh if it has not yet been made inside this window. */
2023 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
2024 !after(tp->high_seq, tp->snd_una) ||
2025 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2026 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2027 tp->prior_cwnd = tp->snd_cwnd;
2028 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2029 tcp_ca_event(sk, CA_EVENT_LOSS);
2032 tp->snd_cwnd = tcp_packets_in_flight(tp) + 1;
2033 tp->snd_cwnd_cnt = 0;
2034 tp->snd_cwnd_stamp = tcp_jiffies32;
2036 /* Timeout in disordered state after receiving substantial DUPACKs
2037 * suggests that the degree of reordering is over-estimated.
2039 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
2040 tp->sacked_out >= net->ipv4.sysctl_tcp_reordering)
2041 tp->reordering = min_t(unsigned int, tp->reordering,
2042 net->ipv4.sysctl_tcp_reordering);
2043 tcp_set_ca_state(sk, TCP_CA_Loss);
2044 tp->high_seq = tp->snd_nxt;
2045 tcp_ecn_queue_cwr(tp);
2047 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
2048 * loss recovery is underway except recurring timeout(s) on
2049 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
2051 tp->frto = net->ipv4.sysctl_tcp_frto &&
2052 (new_recovery || icsk->icsk_retransmits) &&
2053 !inet_csk(sk)->icsk_mtup.probe_size;
2056 /* If ACK arrived pointing to a remembered SACK, it means that our
2057 * remembered SACKs do not reflect real state of receiver i.e.
2058 * receiver _host_ is heavily congested (or buggy).
2060 * To avoid big spurious retransmission bursts due to transient SACK
2061 * scoreboard oddities that look like reneging, we give the receiver a
2062 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2063 * restore sanity to the SACK scoreboard. If the apparent reneging
2064 * persists until this RTO then we'll clear the SACK scoreboard.
2066 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
2068 if (flag & FLAG_SACK_RENEGING) {
2069 struct tcp_sock *tp = tcp_sk(sk);
2070 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2071 msecs_to_jiffies(10));
2073 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2074 delay, TCP_RTO_MAX);
2080 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2081 * counter when SACK is enabled (without SACK, sacked_out is used for
2084 * With reordering, holes may still be in flight, so RFC3517 recovery
2085 * uses pure sacked_out (total number of SACKed segments) even though
2086 * it violates the RFC that uses duplicate ACKs, often these are equal
2087 * but when e.g. out-of-window ACKs or packet duplication occurs,
2088 * they differ. Since neither occurs due to loss, TCP should really
2091 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2093 return tp->sacked_out + 1;
2096 /* Linux NewReno/SACK/ECN state machine.
2097 * --------------------------------------
2099 * "Open" Normal state, no dubious events, fast path.
2100 * "Disorder" In all the respects it is "Open",
2101 * but requires a bit more attention. It is entered when
2102 * we see some SACKs or dupacks. It is split of "Open"
2103 * mainly to move some processing from fast path to slow one.
2104 * "CWR" CWND was reduced due to some Congestion Notification event.
2105 * It can be ECN, ICMP source quench, local device congestion.
2106 * "Recovery" CWND was reduced, we are fast-retransmitting.
2107 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2109 * tcp_fastretrans_alert() is entered:
2110 * - each incoming ACK, if state is not "Open"
2111 * - when arrived ACK is unusual, namely:
2116 * Counting packets in flight is pretty simple.
2118 * in_flight = packets_out - left_out + retrans_out
2120 * packets_out is SND.NXT-SND.UNA counted in packets.
2122 * retrans_out is number of retransmitted segments.
2124 * left_out is number of segments left network, but not ACKed yet.
2126 * left_out = sacked_out + lost_out
2128 * sacked_out: Packets, which arrived to receiver out of order
2129 * and hence not ACKed. With SACKs this number is simply
2130 * amount of SACKed data. Even without SACKs
2131 * it is easy to give pretty reliable estimate of this number,
2132 * counting duplicate ACKs.
2134 * lost_out: Packets lost by network. TCP has no explicit
2135 * "loss notification" feedback from network (for now).
2136 * It means that this number can be only _guessed_.
2137 * Actually, it is the heuristics to predict lossage that
2138 * distinguishes different algorithms.
2140 * F.e. after RTO, when all the queue is considered as lost,
2141 * lost_out = packets_out and in_flight = retrans_out.
2143 * Essentially, we have now a few algorithms detecting
2146 * If the receiver supports SACK:
2148 * RFC6675/3517: It is the conventional algorithm. A packet is
2149 * considered lost if the number of higher sequence packets
2150 * SACKed is greater than or equal the DUPACK thoreshold
2151 * (reordering). This is implemented in tcp_mark_head_lost and
2152 * tcp_update_scoreboard.
2154 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm
2155 * (2017-) that checks timing instead of counting DUPACKs.
2156 * Essentially a packet is considered lost if it's not S/ACKed
2157 * after RTT + reordering_window, where both metrics are
2158 * dynamically measured and adjusted. This is implemented in
2159 * tcp_rack_mark_lost.
2161 * If the receiver does not support SACK:
2163 * NewReno (RFC6582): in Recovery we assume that one segment
2164 * is lost (classic Reno). While we are in Recovery and
2165 * a partial ACK arrives, we assume that one more packet
2166 * is lost (NewReno). This heuristics are the same in NewReno
2169 * Really tricky (and requiring careful tuning) part of algorithm
2170 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2171 * The first determines the moment _when_ we should reduce CWND and,
2172 * hence, slow down forward transmission. In fact, it determines the moment
2173 * when we decide that hole is caused by loss, rather than by a reorder.
2175 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2176 * holes, caused by lost packets.
2178 * And the most logically complicated part of algorithm is undo
2179 * heuristics. We detect false retransmits due to both too early
2180 * fast retransmit (reordering) and underestimated RTO, analyzing
2181 * timestamps and D-SACKs. When we detect that some segments were
2182 * retransmitted by mistake and CWND reduction was wrong, we undo
2183 * window reduction and abort recovery phase. This logic is hidden
2184 * inside several functions named tcp_try_undo_<something>.
2187 /* This function decides, when we should leave Disordered state
2188 * and enter Recovery phase, reducing congestion window.
2190 * Main question: may we further continue forward transmission
2191 * with the same cwnd?
2193 static bool tcp_time_to_recover(struct sock *sk, int flag)
2195 struct tcp_sock *tp = tcp_sk(sk);
2197 /* Trick#1: The loss is proven. */
2201 /* Not-A-Trick#2 : Classic rule... */
2202 if (!tcp_is_rack(sk) && tcp_dupack_heuristics(tp) > tp->reordering)
2208 /* Detect loss in event "A" above by marking head of queue up as lost.
2209 * For RFC3517 SACK, a segment is considered lost if it
2210 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2211 * the maximum SACKed segments to pass before reaching this limit.
2213 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2215 struct tcp_sock *tp = tcp_sk(sk);
2216 struct sk_buff *skb;
2218 /* Use SACK to deduce losses of new sequences sent during recovery */
2219 const u32 loss_high = tp->snd_nxt;
2221 WARN_ON(packets > tp->packets_out);
2222 skb = tp->lost_skb_hint;
2224 /* Head already handled? */
2225 if (mark_head && after(TCP_SKB_CB(skb)->seq, tp->snd_una))
2227 cnt = tp->lost_cnt_hint;
2229 skb = tcp_rtx_queue_head(sk);
2233 skb_rbtree_walk_from(skb) {
2234 /* TODO: do this better */
2235 /* this is not the most efficient way to do this... */
2236 tp->lost_skb_hint = skb;
2237 tp->lost_cnt_hint = cnt;
2239 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2242 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2243 cnt += tcp_skb_pcount(skb);
2248 tcp_skb_mark_lost(tp, skb);
2253 tcp_verify_left_out(tp);
2256 /* Account newly detected lost packet(s) */
2258 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2260 struct tcp_sock *tp = tcp_sk(sk);
2262 if (tcp_is_sack(tp)) {
2263 int sacked_upto = tp->sacked_out - tp->reordering;
2264 if (sacked_upto >= 0)
2265 tcp_mark_head_lost(sk, sacked_upto, 0);
2266 else if (fast_rexmit)
2267 tcp_mark_head_lost(sk, 1, 1);
2271 static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2273 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2274 before(tp->rx_opt.rcv_tsecr, when);
2277 /* skb is spurious retransmitted if the returned timestamp echo
2278 * reply is prior to the skb transmission time
2280 static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2281 const struct sk_buff *skb)
2283 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2284 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp(skb));
2287 /* Nothing was retransmitted or returned timestamp is less
2288 * than timestamp of the first retransmission.
2290 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2292 return tp->retrans_stamp &&
2293 tcp_tsopt_ecr_before(tp, tp->retrans_stamp);
2296 /* Undo procedures. */
2298 /* We can clear retrans_stamp when there are no retransmissions in the
2299 * window. It would seem that it is trivially available for us in
2300 * tp->retrans_out, however, that kind of assumptions doesn't consider
2301 * what will happen if errors occur when sending retransmission for the
2302 * second time. ...It could the that such segment has only
2303 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2304 * the head skb is enough except for some reneging corner cases that
2305 * are not worth the effort.
2307 * Main reason for all this complexity is the fact that connection dying
2308 * time now depends on the validity of the retrans_stamp, in particular,
2309 * that successive retransmissions of a segment must not advance
2310 * retrans_stamp under any conditions.
2312 static bool tcp_any_retrans_done(const struct sock *sk)
2314 const struct tcp_sock *tp = tcp_sk(sk);
2315 struct sk_buff *skb;
2317 if (tp->retrans_out)
2320 skb = tcp_rtx_queue_head(sk);
2321 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2327 static void DBGUNDO(struct sock *sk, const char *msg)
2329 #if FASTRETRANS_DEBUG > 1
2330 struct tcp_sock *tp = tcp_sk(sk);
2331 struct inet_sock *inet = inet_sk(sk);
2333 if (sk->sk_family == AF_INET) {
2334 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2336 &inet->inet_daddr, ntohs(inet->inet_dport),
2337 tp->snd_cwnd, tcp_left_out(tp),
2338 tp->snd_ssthresh, tp->prior_ssthresh,
2341 #if IS_ENABLED(CONFIG_IPV6)
2342 else if (sk->sk_family == AF_INET6) {
2343 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2345 &sk->sk_v6_daddr, ntohs(inet->inet_dport),
2346 tp->snd_cwnd, tcp_left_out(tp),
2347 tp->snd_ssthresh, tp->prior_ssthresh,
2354 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2356 struct tcp_sock *tp = tcp_sk(sk);
2359 struct sk_buff *skb;
2361 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2362 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2365 tcp_clear_all_retrans_hints(tp);
2368 if (tp->prior_ssthresh) {
2369 const struct inet_connection_sock *icsk = inet_csk(sk);
2371 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2373 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2374 tp->snd_ssthresh = tp->prior_ssthresh;
2375 tcp_ecn_withdraw_cwr(tp);
2378 tp->snd_cwnd_stamp = tcp_jiffies32;
2379 tp->undo_marker = 0;
2380 tp->rack.advanced = 1; /* Force RACK to re-exam losses */
2383 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2385 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2388 /* People celebrate: "We love our President!" */
2389 static bool tcp_try_undo_recovery(struct sock *sk)
2391 struct tcp_sock *tp = tcp_sk(sk);
2393 if (tcp_may_undo(tp)) {
2396 /* Happy end! We did not retransmit anything
2397 * or our original transmission succeeded.
2399 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2400 tcp_undo_cwnd_reduction(sk, false);
2401 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2402 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2404 mib_idx = LINUX_MIB_TCPFULLUNDO;
2406 NET_INC_STATS(sock_net(sk), mib_idx);
2407 } else if (tp->rack.reo_wnd_persist) {
2408 tp->rack.reo_wnd_persist--;
2410 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2411 /* Hold old state until something *above* high_seq
2412 * is ACKed. For Reno it is MUST to prevent false
2413 * fast retransmits (RFC2582). SACK TCP is safe. */
2414 if (!tcp_any_retrans_done(sk))
2415 tp->retrans_stamp = 0;
2418 tcp_set_ca_state(sk, TCP_CA_Open);
2419 tp->is_sack_reneg = 0;
2423 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2424 static bool tcp_try_undo_dsack(struct sock *sk)
2426 struct tcp_sock *tp = tcp_sk(sk);
2428 if (tp->undo_marker && !tp->undo_retrans) {
2429 tp->rack.reo_wnd_persist = min(TCP_RACK_RECOVERY_THRESH,
2430 tp->rack.reo_wnd_persist + 1);
2431 DBGUNDO(sk, "D-SACK");
2432 tcp_undo_cwnd_reduction(sk, false);
2433 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2439 /* Undo during loss recovery after partial ACK or using F-RTO. */
2440 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2442 struct tcp_sock *tp = tcp_sk(sk);
2444 if (frto_undo || tcp_may_undo(tp)) {
2445 tcp_undo_cwnd_reduction(sk, true);
2447 DBGUNDO(sk, "partial loss");
2448 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2450 NET_INC_STATS(sock_net(sk),
2451 LINUX_MIB_TCPSPURIOUSRTOS);
2452 inet_csk(sk)->icsk_retransmits = 0;
2453 if (frto_undo || tcp_is_sack(tp)) {
2454 tcp_set_ca_state(sk, TCP_CA_Open);
2455 tp->is_sack_reneg = 0;
2462 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2463 * It computes the number of packets to send (sndcnt) based on packets newly
2465 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2466 * cwnd reductions across a full RTT.
2467 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2468 * But when the retransmits are acked without further losses, PRR
2469 * slow starts cwnd up to ssthresh to speed up the recovery.
2471 static void tcp_init_cwnd_reduction(struct sock *sk)
2473 struct tcp_sock *tp = tcp_sk(sk);
2475 tp->high_seq = tp->snd_nxt;
2476 tp->tlp_high_seq = 0;
2477 tp->snd_cwnd_cnt = 0;
2478 tp->prior_cwnd = tp->snd_cwnd;
2479 tp->prr_delivered = 0;
2481 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2482 tcp_ecn_queue_cwr(tp);
2485 void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int flag)
2487 struct tcp_sock *tp = tcp_sk(sk);
2489 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2491 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2494 tp->prr_delivered += newly_acked_sacked;
2496 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2498 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2499 } else if ((flag & (FLAG_RETRANS_DATA_ACKED | FLAG_LOST_RETRANS)) ==
2500 FLAG_RETRANS_DATA_ACKED) {
2501 sndcnt = min_t(int, delta,
2502 max_t(int, tp->prr_delivered - tp->prr_out,
2503 newly_acked_sacked) + 1);
2505 sndcnt = min(delta, newly_acked_sacked);
2507 /* Force a fast retransmit upon entering fast recovery */
2508 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
2509 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2512 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2514 struct tcp_sock *tp = tcp_sk(sk);
2516 if (inet_csk(sk)->icsk_ca_ops->cong_control)
2519 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2520 if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH &&
2521 (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) {
2522 tp->snd_cwnd = tp->snd_ssthresh;
2523 tp->snd_cwnd_stamp = tcp_jiffies32;
2525 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2528 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2529 void tcp_enter_cwr(struct sock *sk)
2531 struct tcp_sock *tp = tcp_sk(sk);
2533 tp->prior_ssthresh = 0;
2534 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2535 tp->undo_marker = 0;
2536 tcp_init_cwnd_reduction(sk);
2537 tcp_set_ca_state(sk, TCP_CA_CWR);
2540 EXPORT_SYMBOL(tcp_enter_cwr);
2542 static void tcp_try_keep_open(struct sock *sk)
2544 struct tcp_sock *tp = tcp_sk(sk);
2545 int state = TCP_CA_Open;
2547 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2548 state = TCP_CA_Disorder;
2550 if (inet_csk(sk)->icsk_ca_state != state) {
2551 tcp_set_ca_state(sk, state);
2552 tp->high_seq = tp->snd_nxt;
2556 static void tcp_try_to_open(struct sock *sk, int flag)
2558 struct tcp_sock *tp = tcp_sk(sk);
2560 tcp_verify_left_out(tp);
2562 if (!tcp_any_retrans_done(sk))
2563 tp->retrans_stamp = 0;
2565 if (flag & FLAG_ECE)
2568 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2569 tcp_try_keep_open(sk);
2573 static void tcp_mtup_probe_failed(struct sock *sk)
2575 struct inet_connection_sock *icsk = inet_csk(sk);
2577 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2578 icsk->icsk_mtup.probe_size = 0;
2579 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
2582 static void tcp_mtup_probe_success(struct sock *sk)
2584 struct tcp_sock *tp = tcp_sk(sk);
2585 struct inet_connection_sock *icsk = inet_csk(sk);
2587 /* FIXME: breaks with very large cwnd */
2588 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2589 tp->snd_cwnd = tp->snd_cwnd *
2590 tcp_mss_to_mtu(sk, tp->mss_cache) /
2591 icsk->icsk_mtup.probe_size;
2592 tp->snd_cwnd_cnt = 0;
2593 tp->snd_cwnd_stamp = tcp_jiffies32;
2594 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2596 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2597 icsk->icsk_mtup.probe_size = 0;
2598 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2599 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
2602 /* Do a simple retransmit without using the backoff mechanisms in
2603 * tcp_timer. This is used for path mtu discovery.
2604 * The socket is already locked here.
2606 void tcp_simple_retransmit(struct sock *sk)
2608 const struct inet_connection_sock *icsk = inet_csk(sk);
2609 struct tcp_sock *tp = tcp_sk(sk);
2610 struct sk_buff *skb;
2611 unsigned int mss = tcp_current_mss(sk);
2613 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2614 if (tcp_skb_seglen(skb) > mss &&
2615 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2616 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2617 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2618 tp->retrans_out -= tcp_skb_pcount(skb);
2620 tcp_skb_mark_lost_uncond_verify(tp, skb);
2624 tcp_clear_retrans_hints_partial(tp);
2629 if (tcp_is_reno(tp))
2630 tcp_limit_reno_sacked(tp);
2632 tcp_verify_left_out(tp);
2634 /* Don't muck with the congestion window here.
2635 * Reason is that we do not increase amount of _data_
2636 * in network, but units changed and effective
2637 * cwnd/ssthresh really reduced now.
2639 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2640 tp->high_seq = tp->snd_nxt;
2641 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2642 tp->prior_ssthresh = 0;
2643 tp->undo_marker = 0;
2644 tcp_set_ca_state(sk, TCP_CA_Loss);
2646 tcp_xmit_retransmit_queue(sk);
2648 EXPORT_SYMBOL(tcp_simple_retransmit);
2650 void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2652 struct tcp_sock *tp = tcp_sk(sk);
2655 if (tcp_is_reno(tp))
2656 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2658 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2660 NET_INC_STATS(sock_net(sk), mib_idx);
2662 tp->prior_ssthresh = 0;
2665 if (!tcp_in_cwnd_reduction(sk)) {
2667 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2668 tcp_init_cwnd_reduction(sk);
2670 tcp_set_ca_state(sk, TCP_CA_Recovery);
2673 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2674 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2676 static void tcp_process_loss(struct sock *sk, int flag, int num_dupack,
2679 struct tcp_sock *tp = tcp_sk(sk);
2680 bool recovered = !before(tp->snd_una, tp->high_seq);
2682 if ((flag & FLAG_SND_UNA_ADVANCED || rcu_access_pointer(tp->fastopen_rsk)) &&
2683 tcp_try_undo_loss(sk, false))
2686 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2687 /* Step 3.b. A timeout is spurious if not all data are
2688 * lost, i.e., never-retransmitted data are (s)acked.
2690 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2691 tcp_try_undo_loss(sk, true))
2694 if (after(tp->snd_nxt, tp->high_seq)) {
2695 if (flag & FLAG_DATA_SACKED || num_dupack)
2696 tp->frto = 0; /* Step 3.a. loss was real */
2697 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2698 tp->high_seq = tp->snd_nxt;
2699 /* Step 2.b. Try send new data (but deferred until cwnd
2700 * is updated in tcp_ack()). Otherwise fall back to
2701 * the conventional recovery.
2703 if (!tcp_write_queue_empty(sk) &&
2704 after(tcp_wnd_end(tp), tp->snd_nxt)) {
2705 *rexmit = REXMIT_NEW;
2713 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2714 tcp_try_undo_recovery(sk);
2717 if (tcp_is_reno(tp)) {
2718 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2719 * delivered. Lower inflight to clock out (re)tranmissions.
2721 if (after(tp->snd_nxt, tp->high_seq) && num_dupack)
2722 tcp_add_reno_sack(sk, num_dupack, flag & FLAG_ECE);
2723 else if (flag & FLAG_SND_UNA_ADVANCED)
2724 tcp_reset_reno_sack(tp);
2726 *rexmit = REXMIT_LOST;
2729 /* Undo during fast recovery after partial ACK. */
2730 static bool tcp_try_undo_partial(struct sock *sk, u32 prior_snd_una)
2732 struct tcp_sock *tp = tcp_sk(sk);
2734 if (tp->undo_marker && tcp_packet_delayed(tp)) {
2735 /* Plain luck! Hole if filled with delayed
2736 * packet, rather than with a retransmit. Check reordering.
2738 tcp_check_sack_reordering(sk, prior_snd_una, 1);
2740 /* We are getting evidence that the reordering degree is higher
2741 * than we realized. If there are no retransmits out then we
2742 * can undo. Otherwise we clock out new packets but do not
2743 * mark more packets lost or retransmit more.
2745 if (tp->retrans_out)
2748 if (!tcp_any_retrans_done(sk))
2749 tp->retrans_stamp = 0;
2751 DBGUNDO(sk, "partial recovery");
2752 tcp_undo_cwnd_reduction(sk, true);
2753 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2754 tcp_try_keep_open(sk);
2760 static void tcp_identify_packet_loss(struct sock *sk, int *ack_flag)
2762 struct tcp_sock *tp = tcp_sk(sk);
2764 if (tcp_rtx_queue_empty(sk))
2767 if (unlikely(tcp_is_reno(tp))) {
2768 tcp_newreno_mark_lost(sk, *ack_flag & FLAG_SND_UNA_ADVANCED);
2769 } else if (tcp_is_rack(sk)) {
2770 u32 prior_retrans = tp->retrans_out;
2772 tcp_rack_mark_lost(sk);
2773 if (prior_retrans > tp->retrans_out)
2774 *ack_flag |= FLAG_LOST_RETRANS;
2778 static bool tcp_force_fast_retransmit(struct sock *sk)
2780 struct tcp_sock *tp = tcp_sk(sk);
2782 return after(tcp_highest_sack_seq(tp),
2783 tp->snd_una + tp->reordering * tp->mss_cache);
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 u32 prior_snd_una,
2799 int num_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 ece_ack = flag & FLAG_ECE;
2805 bool do_lost = num_dupack || ((flag & FLAG_DATA_SACKED) &&
2806 tcp_force_fast_retransmit(sk));
2808 if (!tp->packets_out && tp->sacked_out)
2811 /* Now state machine starts.
2812 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2814 tp->prior_ssthresh = 0;
2816 /* B. In all the states check for reneging SACKs. */
2817 if (tcp_check_sack_reneging(sk, flag))
2820 /* C. Check consistency of the current state. */
2821 tcp_verify_left_out(tp);
2823 /* D. Check state exit conditions. State can be terminated
2824 * when high_seq is ACKed. */
2825 if (icsk->icsk_ca_state == TCP_CA_Open) {
2826 WARN_ON(tp->retrans_out != 0);
2827 tp->retrans_stamp = 0;
2828 } else if (!before(tp->snd_una, tp->high_seq)) {
2829 switch (icsk->icsk_ca_state) {
2831 /* CWR is to be held something *above* high_seq
2832 * is ACKed for CWR bit to reach receiver. */
2833 if (tp->snd_una != tp->high_seq) {
2834 tcp_end_cwnd_reduction(sk);
2835 tcp_set_ca_state(sk, TCP_CA_Open);
2839 case TCP_CA_Recovery:
2840 if (tcp_is_reno(tp))
2841 tcp_reset_reno_sack(tp);
2842 if (tcp_try_undo_recovery(sk))
2844 tcp_end_cwnd_reduction(sk);
2849 /* E. Process state. */
2850 switch (icsk->icsk_ca_state) {
2851 case TCP_CA_Recovery:
2852 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2853 if (tcp_is_reno(tp))
2854 tcp_add_reno_sack(sk, num_dupack, ece_ack);
2856 if (tcp_try_undo_partial(sk, prior_snd_una))
2858 /* Partial ACK arrived. Force fast retransmit. */
2859 do_lost = tcp_force_fast_retransmit(sk);
2861 if (tcp_try_undo_dsack(sk)) {
2862 tcp_try_keep_open(sk);
2865 tcp_identify_packet_loss(sk, ack_flag);
2868 tcp_process_loss(sk, flag, num_dupack, rexmit);
2869 tcp_identify_packet_loss(sk, ack_flag);
2870 if (!(icsk->icsk_ca_state == TCP_CA_Open ||
2871 (*ack_flag & FLAG_LOST_RETRANS)))
2873 /* Change state if cwnd is undone or retransmits are lost */
2876 if (tcp_is_reno(tp)) {
2877 if (flag & FLAG_SND_UNA_ADVANCED)
2878 tcp_reset_reno_sack(tp);
2879 tcp_add_reno_sack(sk, num_dupack, ece_ack);
2882 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2883 tcp_try_undo_dsack(sk);
2885 tcp_identify_packet_loss(sk, ack_flag);
2886 if (!tcp_time_to_recover(sk, flag)) {
2887 tcp_try_to_open(sk, flag);
2891 /* MTU probe failure: don't reduce cwnd */
2892 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2893 icsk->icsk_mtup.probe_size &&
2894 tp->snd_una == tp->mtu_probe.probe_seq_start) {
2895 tcp_mtup_probe_failed(sk);
2896 /* Restores the reduction we did in tcp_mtup_probe() */
2898 tcp_simple_retransmit(sk);
2902 /* Otherwise enter Recovery state */
2903 tcp_enter_recovery(sk, ece_ack);
2907 if (!tcp_is_rack(sk) && do_lost)
2908 tcp_update_scoreboard(sk, fast_rexmit);
2909 *rexmit = REXMIT_LOST;
2912 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us, const int flag)
2914 u32 wlen = sock_net(sk)->ipv4.sysctl_tcp_min_rtt_wlen * HZ;
2915 struct tcp_sock *tp = tcp_sk(sk);
2917 if ((flag & FLAG_ACK_MAYBE_DELAYED) && rtt_us > tcp_min_rtt(tp)) {
2918 /* If the remote keeps returning delayed ACKs, eventually
2919 * the min filter would pick it up and overestimate the
2920 * prop. delay when it expires. Skip suspected delayed ACKs.
2924 minmax_running_min(&tp->rtt_min, wlen, tcp_jiffies32,
2925 rtt_us ? : jiffies_to_usecs(1));
2928 static bool tcp_ack_update_rtt(struct sock *sk, const int flag,
2929 long seq_rtt_us, long sack_rtt_us,
2930 long ca_rtt_us, struct rate_sample *rs)
2932 const struct tcp_sock *tp = tcp_sk(sk);
2934 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2935 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2936 * Karn's algorithm forbids taking RTT if some retransmitted data
2937 * is acked (RFC6298).
2940 seq_rtt_us = sack_rtt_us;
2942 /* RTTM Rule: A TSecr value received in a segment is used to
2943 * update the averaged RTT measurement only if the segment
2944 * acknowledges some new data, i.e., only if it advances the
2945 * left edge of the send window.
2946 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2948 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2949 flag & FLAG_ACKED) {
2950 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
2952 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) {
2953 seq_rtt_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
2954 ca_rtt_us = seq_rtt_us;
2957 rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */
2961 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
2962 * always taken together with ACK, SACK, or TS-opts. Any negative
2963 * values will be skipped with the seq_rtt_us < 0 check above.
2965 tcp_update_rtt_min(sk, ca_rtt_us, flag);
2966 tcp_rtt_estimator(sk, seq_rtt_us);
2969 /* RFC6298: only reset backoff on valid RTT measurement. */
2970 inet_csk(sk)->icsk_backoff = 0;
2974 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2975 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
2977 struct rate_sample rs;
2980 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack)
2981 rtt_us = tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req)->snt_synack);
2983 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us, &rs);
2987 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
2989 const struct inet_connection_sock *icsk = inet_csk(sk);
2991 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
2992 tcp_sk(sk)->snd_cwnd_stamp = tcp_jiffies32;
2995 /* Restart timer after forward progress on connection.
2996 * RFC2988 recommends to restart timer to now+rto.
2998 void tcp_rearm_rto(struct sock *sk)
3000 const struct inet_connection_sock *icsk = inet_csk(sk);
3001 struct tcp_sock *tp = tcp_sk(sk);
3003 /* If the retrans timer is currently being used by Fast Open
3004 * for SYN-ACK retrans purpose, stay put.
3006 if (rcu_access_pointer(tp->fastopen_rsk))
3009 if (!tp->packets_out) {
3010 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3012 u32 rto = inet_csk(sk)->icsk_rto;
3013 /* Offset the time elapsed after installing regular RTO */
3014 if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT ||
3015 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
3016 s64 delta_us = tcp_rto_delta_us(sk);
3017 /* delta_us may not be positive if the socket is locked
3018 * when the retrans timer fires and is rescheduled.
3020 rto = usecs_to_jiffies(max_t(int, delta_us, 1));
3022 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
3027 /* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */
3028 static void tcp_set_xmit_timer(struct sock *sk)
3030 if (!tcp_schedule_loss_probe(sk, true))
3034 /* If we get here, the whole TSO packet has not been acked. */
3035 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3037 struct tcp_sock *tp = tcp_sk(sk);
3040 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3042 packets_acked = tcp_skb_pcount(skb);
3043 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3045 packets_acked -= tcp_skb_pcount(skb);
3047 if (packets_acked) {
3048 BUG_ON(tcp_skb_pcount(skb) == 0);
3049 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3052 return packets_acked;
3055 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3058 const struct skb_shared_info *shinfo;
3060 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3061 if (likely(!TCP_SKB_CB(skb)->txstamp_ack))
3064 shinfo = skb_shinfo(skb);
3065 if (!before(shinfo->tskey, prior_snd_una) &&
3066 before(shinfo->tskey, tcp_sk(sk)->snd_una)) {
3067 tcp_skb_tsorted_save(skb) {
3068 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK);
3069 } tcp_skb_tsorted_restore(skb);
3073 /* Remove acknowledged frames from the retransmission queue. If our packet
3074 * is before the ack sequence we can discard it as it's confirmed to have
3075 * arrived at the other end.
3077 static int tcp_clean_rtx_queue(struct sock *sk, u32 prior_fack,
3079 struct tcp_sacktag_state *sack, bool ece_ack)
3081 const struct inet_connection_sock *icsk = inet_csk(sk);
3082 u64 first_ackt, last_ackt;
3083 struct tcp_sock *tp = tcp_sk(sk);
3084 u32 prior_sacked = tp->sacked_out;
3085 u32 reord = tp->snd_nxt; /* lowest acked un-retx un-sacked seq */
3086 struct sk_buff *skb, *next;
3087 bool fully_acked = true;
3088 long sack_rtt_us = -1L;
3089 long seq_rtt_us = -1L;
3090 long ca_rtt_us = -1L;
3092 u32 last_in_flight = 0;
3098 for (skb = skb_rb_first(&sk->tcp_rtx_queue); skb; skb = next) {
3099 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3100 const u32 start_seq = scb->seq;
3101 u8 sacked = scb->sacked;
3104 /* Determine how many packets and what bytes were acked, tso and else */
3105 if (after(scb->end_seq, tp->snd_una)) {
3106 if (tcp_skb_pcount(skb) == 1 ||
3107 !after(tp->snd_una, scb->seq))
3110 acked_pcount = tcp_tso_acked(sk, skb);
3113 fully_acked = false;
3115 acked_pcount = tcp_skb_pcount(skb);
3118 if (unlikely(sacked & TCPCB_RETRANS)) {
3119 if (sacked & TCPCB_SACKED_RETRANS)
3120 tp->retrans_out -= acked_pcount;
3121 flag |= FLAG_RETRANS_DATA_ACKED;
3122 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3123 last_ackt = tcp_skb_timestamp_us(skb);
3124 WARN_ON_ONCE(last_ackt == 0);
3126 first_ackt = last_ackt;
3128 last_in_flight = TCP_SKB_CB(skb)->tx.in_flight;
3129 if (before(start_seq, reord))
3131 if (!after(scb->end_seq, tp->high_seq))
3132 flag |= FLAG_ORIG_SACK_ACKED;
3135 if (sacked & TCPCB_SACKED_ACKED) {
3136 tp->sacked_out -= acked_pcount;
3137 } else if (tcp_is_sack(tp)) {
3138 tcp_count_delivered(tp, acked_pcount, ece_ack);
3139 if (!tcp_skb_spurious_retrans(tp, skb))
3140 tcp_rack_advance(tp, sacked, scb->end_seq,
3141 tcp_skb_timestamp_us(skb));
3143 if (sacked & TCPCB_LOST)
3144 tp->lost_out -= acked_pcount;
3146 tp->packets_out -= acked_pcount;
3147 pkts_acked += acked_pcount;
3148 tcp_rate_skb_delivered(sk, skb, sack->rate);
3150 /* Initial outgoing SYN's get put onto the write_queue
3151 * just like anything else we transmit. It is not
3152 * true data, and if we misinform our callers that
3153 * this ACK acks real data, we will erroneously exit
3154 * connection startup slow start one packet too
3155 * quickly. This is severely frowned upon behavior.
3157 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3158 flag |= FLAG_DATA_ACKED;
3160 flag |= FLAG_SYN_ACKED;
3161 tp->retrans_stamp = 0;
3167 tcp_ack_tstamp(sk, skb, prior_snd_una);
3169 next = skb_rb_next(skb);
3170 if (unlikely(skb == tp->retransmit_skb_hint))
3171 tp->retransmit_skb_hint = NULL;
3172 if (unlikely(skb == tp->lost_skb_hint))
3173 tp->lost_skb_hint = NULL;
3174 tcp_highest_sack_replace(sk, skb, next);
3175 tcp_rtx_queue_unlink_and_free(skb, sk);
3179 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3181 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3182 tp->snd_up = tp->snd_una;
3185 tcp_ack_tstamp(sk, skb, prior_snd_una);
3186 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
3187 flag |= FLAG_SACK_RENEGING;
3190 if (likely(first_ackt) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
3191 seq_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, first_ackt);
3192 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, last_ackt);
3194 if (pkts_acked == 1 && last_in_flight < tp->mss_cache &&
3195 last_in_flight && !prior_sacked && fully_acked &&
3196 sack->rate->prior_delivered + 1 == tp->delivered &&
3197 !(flag & (FLAG_CA_ALERT | FLAG_SYN_ACKED))) {
3198 /* Conservatively mark a delayed ACK. It's typically
3199 * from a lone runt packet over the round trip to
3200 * a receiver w/o out-of-order or CE events.
3202 flag |= FLAG_ACK_MAYBE_DELAYED;
3205 if (sack->first_sackt) {
3206 sack_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->first_sackt);
3207 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->last_sackt);
3209 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
3210 ca_rtt_us, sack->rate);
3212 if (flag & FLAG_ACKED) {
3213 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3214 if (unlikely(icsk->icsk_mtup.probe_size &&
3215 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3216 tcp_mtup_probe_success(sk);
3219 if (tcp_is_reno(tp)) {
3220 tcp_remove_reno_sacks(sk, pkts_acked, ece_ack);
3222 /* If any of the cumulatively ACKed segments was
3223 * retransmitted, non-SACK case cannot confirm that
3224 * progress was due to original transmission due to
3225 * lack of TCPCB_SACKED_ACKED bits even if some of
3226 * the packets may have been never retransmitted.
3228 if (flag & FLAG_RETRANS_DATA_ACKED)
3229 flag &= ~FLAG_ORIG_SACK_ACKED;
3233 /* Non-retransmitted hole got filled? That's reordering */
3234 if (before(reord, prior_fack))
3235 tcp_check_sack_reordering(sk, reord, 0);
3237 delta = prior_sacked - tp->sacked_out;
3238 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3240 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3241 sack_rtt_us > tcp_stamp_us_delta(tp->tcp_mstamp,
3242 tcp_skb_timestamp_us(skb))) {
3243 /* Do not re-arm RTO if the sack RTT is measured from data sent
3244 * after when the head was last (re)transmitted. Otherwise the
3245 * timeout may continue to extend in loss recovery.
3247 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3250 if (icsk->icsk_ca_ops->pkts_acked) {
3251 struct ack_sample sample = { .pkts_acked = pkts_acked,
3252 .rtt_us = sack->rate->rtt_us,
3253 .in_flight = last_in_flight };
3255 icsk->icsk_ca_ops->pkts_acked(sk, &sample);
3258 #if FASTRETRANS_DEBUG > 0
3259 WARN_ON((int)tp->sacked_out < 0);
3260 WARN_ON((int)tp->lost_out < 0);
3261 WARN_ON((int)tp->retrans_out < 0);
3262 if (!tp->packets_out && tcp_is_sack(tp)) {
3263 icsk = inet_csk(sk);
3265 pr_debug("Leak l=%u %d\n",
3266 tp->lost_out, icsk->icsk_ca_state);
3269 if (tp->sacked_out) {
3270 pr_debug("Leak s=%u %d\n",
3271 tp->sacked_out, icsk->icsk_ca_state);
3274 if (tp->retrans_out) {
3275 pr_debug("Leak r=%u %d\n",
3276 tp->retrans_out, icsk->icsk_ca_state);
3277 tp->retrans_out = 0;
3284 static void tcp_ack_probe(struct sock *sk)
3286 struct inet_connection_sock *icsk = inet_csk(sk);
3287 struct sk_buff *head = tcp_send_head(sk);
3288 const struct tcp_sock *tp = tcp_sk(sk);
3290 /* Was it a usable window open? */
3293 if (!after(TCP_SKB_CB(head)->end_seq, tcp_wnd_end(tp))) {
3294 icsk->icsk_backoff = 0;
3295 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3296 /* Socket must be waked up by subsequent tcp_data_snd_check().
3297 * This function is not for random using!
3300 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
3302 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3307 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3309 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3310 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3313 /* Decide wheather to run the increase function of congestion control. */
3314 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3316 /* If reordering is high then always grow cwnd whenever data is
3317 * delivered regardless of its ordering. Otherwise stay conservative
3318 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3319 * new SACK or ECE mark may first advance cwnd here and later reduce
3320 * cwnd in tcp_fastretrans_alert() based on more states.
3322 if (tcp_sk(sk)->reordering > sock_net(sk)->ipv4.sysctl_tcp_reordering)
3323 return flag & FLAG_FORWARD_PROGRESS;
3325 return flag & FLAG_DATA_ACKED;
3328 /* The "ultimate" congestion control function that aims to replace the rigid
3329 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3330 * It's called toward the end of processing an ACK with precise rate
3331 * information. All transmission or retransmission are delayed afterwards.
3333 static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
3334 int flag, const struct rate_sample *rs)
3336 const struct inet_connection_sock *icsk = inet_csk(sk);
3338 if (icsk->icsk_ca_ops->cong_control) {
3339 icsk->icsk_ca_ops->cong_control(sk, rs);
3343 if (tcp_in_cwnd_reduction(sk)) {
3344 /* Reduce cwnd if state mandates */
3345 tcp_cwnd_reduction(sk, acked_sacked, flag);
3346 } else if (tcp_may_raise_cwnd(sk, flag)) {
3347 /* Advance cwnd if state allows */
3348 tcp_cong_avoid(sk, ack, acked_sacked);
3350 tcp_update_pacing_rate(sk);
3353 /* Check that window update is acceptable.
3354 * The function assumes that snd_una<=ack<=snd_next.
3356 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3357 const u32 ack, const u32 ack_seq,
3360 return after(ack, tp->snd_una) ||
3361 after(ack_seq, tp->snd_wl1) ||
3362 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3365 /* If we update tp->snd_una, also update tp->bytes_acked */
3366 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3368 u32 delta = ack - tp->snd_una;
3370 sock_owned_by_me((struct sock *)tp);
3371 tp->bytes_acked += delta;
3375 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3376 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3378 u32 delta = seq - tp->rcv_nxt;
3380 sock_owned_by_me((struct sock *)tp);
3381 tp->bytes_received += delta;
3382 WRITE_ONCE(tp->rcv_nxt, seq);
3385 /* Update our send window.
3387 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3388 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3390 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3393 struct tcp_sock *tp = tcp_sk(sk);
3395 u32 nwin = ntohs(tcp_hdr(skb)->window);
3397 if (likely(!tcp_hdr(skb)->syn))
3398 nwin <<= tp->rx_opt.snd_wscale;
3400 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3401 flag |= FLAG_WIN_UPDATE;
3402 tcp_update_wl(tp, ack_seq);
3404 if (tp->snd_wnd != nwin) {
3407 /* Note, it is the only place, where
3408 * fast path is recovered for sending TCP.
3411 tcp_fast_path_check(sk);
3413 if (!tcp_write_queue_empty(sk))
3414 tcp_slow_start_after_idle_check(sk);
3416 if (nwin > tp->max_window) {
3417 tp->max_window = nwin;
3418 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3423 tcp_snd_una_update(tp, ack);
3428 static bool __tcp_oow_rate_limited(struct net *net, int mib_idx,
3429 u32 *last_oow_ack_time)
3431 if (*last_oow_ack_time) {
3432 s32 elapsed = (s32)(tcp_jiffies32 - *last_oow_ack_time);
3434 if (0 <= elapsed && elapsed < net->ipv4.sysctl_tcp_invalid_ratelimit) {
3435 NET_INC_STATS(net, mib_idx);
3436 return true; /* rate-limited: don't send yet! */
3440 *last_oow_ack_time = tcp_jiffies32;
3442 return false; /* not rate-limited: go ahead, send dupack now! */
3445 /* Return true if we're currently rate-limiting out-of-window ACKs and
3446 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3447 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3448 * attacks that send repeated SYNs or ACKs for the same connection. To
3449 * do this, we do not send a duplicate SYNACK or ACK if the remote
3450 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3452 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3453 int mib_idx, u32 *last_oow_ack_time)
3455 /* Data packets without SYNs are not likely part of an ACK loop. */
3456 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3460 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time);
3463 /* RFC 5961 7 [ACK Throttling] */
3464 static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb)
3466 /* unprotected vars, we dont care of overwrites */
3467 static u32 challenge_timestamp;
3468 static unsigned int challenge_count;
3469 struct tcp_sock *tp = tcp_sk(sk);
3470 struct net *net = sock_net(sk);
3473 /* First check our per-socket dupack rate limit. */
3474 if (__tcp_oow_rate_limited(net,
3475 LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3476 &tp->last_oow_ack_time))
3479 /* Then check host-wide RFC 5961 rate limit. */
3481 if (now != challenge_timestamp) {
3482 u32 ack_limit = net->ipv4.sysctl_tcp_challenge_ack_limit;
3483 u32 half = (ack_limit + 1) >> 1;
3485 challenge_timestamp = now;
3486 WRITE_ONCE(challenge_count, half + prandom_u32_max(ack_limit));
3488 count = READ_ONCE(challenge_count);
3490 WRITE_ONCE(challenge_count, count - 1);
3491 NET_INC_STATS(net, LINUX_MIB_TCPCHALLENGEACK);
3496 static void tcp_store_ts_recent(struct tcp_sock *tp)
3498 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3499 tp->rx_opt.ts_recent_stamp = ktime_get_seconds();
3502 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3504 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3505 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3506 * extra check below makes sure this can only happen
3507 * for pure ACK frames. -DaveM
3509 * Not only, also it occurs for expired timestamps.
3512 if (tcp_paws_check(&tp->rx_opt, 0))
3513 tcp_store_ts_recent(tp);
3517 /* This routine deals with acks during a TLP episode and ends an episode by
3518 * resetting tlp_high_seq. Ref: TLP algorithm in draft-ietf-tcpm-rack
3520 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3522 struct tcp_sock *tp = tcp_sk(sk);
3524 if (before(ack, tp->tlp_high_seq))
3527 if (!tp->tlp_retrans) {
3528 /* TLP of new data has been acknowledged */
3529 tp->tlp_high_seq = 0;
3530 } else if (flag & FLAG_DSACKING_ACK) {
3531 /* This DSACK means original and TLP probe arrived; no loss */
3532 tp->tlp_high_seq = 0;
3533 } else if (after(ack, tp->tlp_high_seq)) {
3534 /* ACK advances: there was a loss, so reduce cwnd. Reset
3535 * tlp_high_seq in tcp_init_cwnd_reduction()
3537 tcp_init_cwnd_reduction(sk);
3538 tcp_set_ca_state(sk, TCP_CA_CWR);
3539 tcp_end_cwnd_reduction(sk);
3540 tcp_try_keep_open(sk);
3541 NET_INC_STATS(sock_net(sk),
3542 LINUX_MIB_TCPLOSSPROBERECOVERY);
3543 } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3544 FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3545 /* Pure dupack: original and TLP probe arrived; no loss */
3546 tp->tlp_high_seq = 0;
3550 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3552 const struct inet_connection_sock *icsk = inet_csk(sk);
3554 if (icsk->icsk_ca_ops->in_ack_event)
3555 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3558 /* Congestion control has updated the cwnd already. So if we're in
3559 * loss recovery then now we do any new sends (for FRTO) or
3560 * retransmits (for CA_Loss or CA_recovery) that make sense.
3562 static void tcp_xmit_recovery(struct sock *sk, int rexmit)
3564 struct tcp_sock *tp = tcp_sk(sk);
3566 if (rexmit == REXMIT_NONE || sk->sk_state == TCP_SYN_SENT)
3569 if (unlikely(rexmit == REXMIT_NEW)) {
3570 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
3572 if (after(tp->snd_nxt, tp->high_seq))
3576 tcp_xmit_retransmit_queue(sk);
3579 /* Returns the number of packets newly acked or sacked by the current ACK */
3580 static u32 tcp_newly_delivered(struct sock *sk, u32 prior_delivered, int flag)
3582 const struct net *net = sock_net(sk);
3583 struct tcp_sock *tp = tcp_sk(sk);
3586 delivered = tp->delivered - prior_delivered;
3587 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVERED, delivered);
3588 if (flag & FLAG_ECE)
3589 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVEREDCE, delivered);
3594 /* This routine deals with incoming acks, but not outgoing ones. */
3595 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3597 struct inet_connection_sock *icsk = inet_csk(sk);
3598 struct tcp_sock *tp = tcp_sk(sk);
3599 struct tcp_sacktag_state sack_state;
3600 struct rate_sample rs = { .prior_delivered = 0 };
3601 u32 prior_snd_una = tp->snd_una;
3602 bool is_sack_reneg = tp->is_sack_reneg;
3603 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3604 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3606 int prior_packets = tp->packets_out;
3607 u32 delivered = tp->delivered;
3608 u32 lost = tp->lost;
3609 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
3612 sack_state.first_sackt = 0;
3613 sack_state.rate = &rs;
3614 sack_state.sack_delivered = 0;
3616 /* We very likely will need to access rtx queue. */
3617 prefetch(sk->tcp_rtx_queue.rb_node);
3619 /* If the ack is older than previous acks
3620 * then we can probably ignore it.
3622 if (before(ack, prior_snd_una)) {
3623 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3624 if (before(ack, prior_snd_una - tp->max_window)) {
3625 if (!(flag & FLAG_NO_CHALLENGE_ACK))
3626 tcp_send_challenge_ack(sk, skb);
3632 /* If the ack includes data we haven't sent yet, discard
3633 * this segment (RFC793 Section 3.9).
3635 if (after(ack, tp->snd_nxt))
3638 if (after(ack, prior_snd_una)) {
3639 flag |= FLAG_SND_UNA_ADVANCED;
3640 icsk->icsk_retransmits = 0;
3642 #if IS_ENABLED(CONFIG_TLS_DEVICE)
3643 if (static_branch_unlikely(&clean_acked_data_enabled.key))
3644 if (icsk->icsk_clean_acked)
3645 icsk->icsk_clean_acked(sk, ack);
3649 prior_fack = tcp_is_sack(tp) ? tcp_highest_sack_seq(tp) : tp->snd_una;
3650 rs.prior_in_flight = tcp_packets_in_flight(tp);
3652 /* ts_recent update must be made after we are sure that the packet
3655 if (flag & FLAG_UPDATE_TS_RECENT)
3656 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3658 if ((flag & (FLAG_SLOWPATH | FLAG_SND_UNA_ADVANCED)) ==
3659 FLAG_SND_UNA_ADVANCED) {
3660 /* Window is constant, pure forward advance.
3661 * No more checks are required.
3662 * Note, we use the fact that SND.UNA>=SND.WL2.
3664 tcp_update_wl(tp, ack_seq);
3665 tcp_snd_una_update(tp, ack);
3666 flag |= FLAG_WIN_UPDATE;
3668 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3670 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS);
3672 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3674 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3677 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3679 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3681 if (TCP_SKB_CB(skb)->sacked)
3682 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3685 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3687 ack_ev_flags |= CA_ACK_ECE;
3690 if (sack_state.sack_delivered)
3691 tcp_count_delivered(tp, sack_state.sack_delivered,
3694 if (flag & FLAG_WIN_UPDATE)
3695 ack_ev_flags |= CA_ACK_WIN_UPDATE;
3697 tcp_in_ack_event(sk, ack_ev_flags);
3700 /* This is a deviation from RFC3168 since it states that:
3701 * "When the TCP data sender is ready to set the CWR bit after reducing
3702 * the congestion window, it SHOULD set the CWR bit only on the first
3703 * new data packet that it transmits."
3704 * We accept CWR on pure ACKs to be more robust
3705 * with widely-deployed TCP implementations that do this.
3707 tcp_ecn_accept_cwr(sk, skb);
3709 /* We passed data and got it acked, remove any soft error
3710 * log. Something worked...
3712 sk->sk_err_soft = 0;
3713 icsk->icsk_probes_out = 0;
3714 tp->rcv_tstamp = tcp_jiffies32;
3718 /* See if we can take anything off of the retransmit queue. */
3719 flag |= tcp_clean_rtx_queue(sk, prior_fack, prior_snd_una, &sack_state,
3722 tcp_rack_update_reo_wnd(sk, &rs);
3724 if (tp->tlp_high_seq)
3725 tcp_process_tlp_ack(sk, ack, flag);
3726 /* If needed, reset TLP/RTO timer; RACK may later override this. */
3727 if (flag & FLAG_SET_XMIT_TIMER)
3728 tcp_set_xmit_timer(sk);
3730 if (tcp_ack_is_dubious(sk, flag)) {
3731 if (!(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP))) {
3733 /* Consider if pure acks were aggregated in tcp_add_backlog() */
3734 if (!(flag & FLAG_DATA))
3735 num_dupack = max_t(u16, 1, skb_shinfo(skb)->gso_segs);
3737 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3741 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3744 delivered = tcp_newly_delivered(sk, delivered, flag);
3745 lost = tp->lost - lost; /* freshly marked lost */
3746 rs.is_ack_delayed = !!(flag & FLAG_ACK_MAYBE_DELAYED);
3747 tcp_rate_gen(sk, delivered, lost, is_sack_reneg, sack_state.rate);
3748 tcp_cong_control(sk, ack, delivered, flag, sack_state.rate);
3749 tcp_xmit_recovery(sk, rexmit);
3753 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3754 if (flag & FLAG_DSACKING_ACK) {
3755 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3757 tcp_newly_delivered(sk, delivered, flag);
3759 /* If this ack opens up a zero window, clear backoff. It was
3760 * being used to time the probes, and is probably far higher than
3761 * it needs to be for normal retransmission.
3765 if (tp->tlp_high_seq)
3766 tcp_process_tlp_ack(sk, ack, flag);
3770 /* If data was SACKed, tag it and see if we should send more data.
3771 * If data was DSACKed, see if we can undo a cwnd reduction.
3773 if (TCP_SKB_CB(skb)->sacked) {
3774 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3776 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3778 tcp_newly_delivered(sk, delivered, flag);
3779 tcp_xmit_recovery(sk, rexmit);
3785 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3786 bool syn, struct tcp_fastopen_cookie *foc,
3789 /* Valid only in SYN or SYN-ACK with an even length. */
3790 if (!foc || !syn || len < 0 || (len & 1))
3793 if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3794 len <= TCP_FASTOPEN_COOKIE_MAX)
3795 memcpy(foc->val, cookie, len);
3802 static void smc_parse_options(const struct tcphdr *th,
3803 struct tcp_options_received *opt_rx,
3804 const unsigned char *ptr,
3807 #if IS_ENABLED(CONFIG_SMC)
3808 if (static_branch_unlikely(&tcp_have_smc)) {
3809 if (th->syn && !(opsize & 1) &&
3810 opsize >= TCPOLEN_EXP_SMC_BASE &&
3811 get_unaligned_be32(ptr) == TCPOPT_SMC_MAGIC)
3817 /* Try to parse the MSS option from the TCP header. Return 0 on failure, clamped
3820 static u16 tcp_parse_mss_option(const struct tcphdr *th, u16 user_mss)
3822 const unsigned char *ptr = (const unsigned char *)(th + 1);
3823 int length = (th->doff * 4) - sizeof(struct tcphdr);
3826 while (length > 0) {
3827 int opcode = *ptr++;
3833 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3840 if (opsize < 2) /* "silly options" */
3842 if (opsize > length)
3843 return mss; /* fail on partial options */
3844 if (opcode == TCPOPT_MSS && opsize == TCPOLEN_MSS) {
3845 u16 in_mss = get_unaligned_be16(ptr);
3848 if (user_mss && user_mss < in_mss)
3860 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3861 * But, this can also be called on packets in the established flow when
3862 * the fast version below fails.
3864 void tcp_parse_options(const struct net *net,
3865 const struct sk_buff *skb,
3866 struct tcp_options_received *opt_rx, int estab,
3867 struct tcp_fastopen_cookie *foc)
3869 const unsigned char *ptr;
3870 const struct tcphdr *th = tcp_hdr(skb);
3871 int length = (th->doff * 4) - sizeof(struct tcphdr);
3873 ptr = (const unsigned char *)(th + 1);
3874 opt_rx->saw_tstamp = 0;
3876 while (length > 0) {
3877 int opcode = *ptr++;
3883 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3890 if (opsize < 2) /* "silly options" */
3892 if (opsize > length)
3893 return; /* don't parse partial options */
3896 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3897 u16 in_mss = get_unaligned_be16(ptr);
3899 if (opt_rx->user_mss &&
3900 opt_rx->user_mss < in_mss)
3901 in_mss = opt_rx->user_mss;
3902 opt_rx->mss_clamp = in_mss;
3907 if (opsize == TCPOLEN_WINDOW && th->syn &&
3908 !estab && net->ipv4.sysctl_tcp_window_scaling) {
3909 __u8 snd_wscale = *(__u8 *)ptr;
3910 opt_rx->wscale_ok = 1;
3911 if (snd_wscale > TCP_MAX_WSCALE) {
3912 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n",
3916 snd_wscale = TCP_MAX_WSCALE;
3918 opt_rx->snd_wscale = snd_wscale;
3921 case TCPOPT_TIMESTAMP:
3922 if ((opsize == TCPOLEN_TIMESTAMP) &&
3923 ((estab && opt_rx->tstamp_ok) ||
3924 (!estab && net->ipv4.sysctl_tcp_timestamps))) {
3925 opt_rx->saw_tstamp = 1;
3926 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3927 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3930 case TCPOPT_SACK_PERM:
3931 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3932 !estab && net->ipv4.sysctl_tcp_sack) {
3933 opt_rx->sack_ok = TCP_SACK_SEEN;
3934 tcp_sack_reset(opt_rx);
3939 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3940 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3942 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3945 #ifdef CONFIG_TCP_MD5SIG
3948 * The MD5 Hash has already been
3949 * checked (see tcp_v{4,6}_do_rcv()).
3953 case TCPOPT_FASTOPEN:
3954 tcp_parse_fastopen_option(
3955 opsize - TCPOLEN_FASTOPEN_BASE,
3956 ptr, th->syn, foc, false);
3960 /* Fast Open option shares code 254 using a
3961 * 16 bits magic number.
3963 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
3964 get_unaligned_be16(ptr) ==
3965 TCPOPT_FASTOPEN_MAGIC)
3966 tcp_parse_fastopen_option(opsize -
3967 TCPOLEN_EXP_FASTOPEN_BASE,
3968 ptr + 2, th->syn, foc, true);
3970 smc_parse_options(th, opt_rx, ptr,
3980 EXPORT_SYMBOL(tcp_parse_options);
3982 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3984 const __be32 *ptr = (const __be32 *)(th + 1);
3986 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3987 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3988 tp->rx_opt.saw_tstamp = 1;
3990 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3993 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3995 tp->rx_opt.rcv_tsecr = 0;
4001 /* Fast parse options. This hopes to only see timestamps.
4002 * If it is wrong it falls back on tcp_parse_options().
4004 static bool tcp_fast_parse_options(const struct net *net,
4005 const struct sk_buff *skb,
4006 const struct tcphdr *th, struct tcp_sock *tp)
4008 /* In the spirit of fast parsing, compare doff directly to constant
4009 * values. Because equality is used, short doff can be ignored here.
4011 if (th->doff == (sizeof(*th) / 4)) {
4012 tp->rx_opt.saw_tstamp = 0;
4014 } else if (tp->rx_opt.tstamp_ok &&
4015 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
4016 if (tcp_parse_aligned_timestamp(tp, th))
4020 tcp_parse_options(net, skb, &tp->rx_opt, 1, NULL);
4021 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
4022 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
4027 #ifdef CONFIG_TCP_MD5SIG
4029 * Parse MD5 Signature option
4031 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
4033 int length = (th->doff << 2) - sizeof(*th);
4034 const u8 *ptr = (const u8 *)(th + 1);
4036 /* If not enough data remaining, we can short cut */
4037 while (length >= TCPOLEN_MD5SIG) {
4038 int opcode = *ptr++;
4049 if (opsize < 2 || opsize > length)
4051 if (opcode == TCPOPT_MD5SIG)
4052 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
4059 EXPORT_SYMBOL(tcp_parse_md5sig_option);
4062 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
4064 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
4065 * it can pass through stack. So, the following predicate verifies that
4066 * this segment is not used for anything but congestion avoidance or
4067 * fast retransmit. Moreover, we even are able to eliminate most of such
4068 * second order effects, if we apply some small "replay" window (~RTO)
4069 * to timestamp space.
4071 * All these measures still do not guarantee that we reject wrapped ACKs
4072 * on networks with high bandwidth, when sequence space is recycled fastly,
4073 * but it guarantees that such events will be very rare and do not affect
4074 * connection seriously. This doesn't look nice, but alas, PAWS is really
4077 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
4078 * states that events when retransmit arrives after original data are rare.
4079 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
4080 * the biggest problem on large power networks even with minor reordering.
4081 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
4082 * up to bandwidth of 18Gigabit/sec. 8) ]
4085 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
4087 const struct tcp_sock *tp = tcp_sk(sk);
4088 const struct tcphdr *th = tcp_hdr(skb);
4089 u32 seq = TCP_SKB_CB(skb)->seq;
4090 u32 ack = TCP_SKB_CB(skb)->ack_seq;
4092 return (/* 1. Pure ACK with correct sequence number. */
4093 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
4095 /* 2. ... and duplicate ACK. */
4096 ack == tp->snd_una &&
4098 /* 3. ... and does not update window. */
4099 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
4101 /* 4. ... and sits in replay window. */
4102 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
4105 static inline bool tcp_paws_discard(const struct sock *sk,
4106 const struct sk_buff *skb)
4108 const struct tcp_sock *tp = tcp_sk(sk);
4110 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4111 !tcp_disordered_ack(sk, skb);
4114 /* Check segment sequence number for validity.
4116 * Segment controls are considered valid, if the segment
4117 * fits to the window after truncation to the window. Acceptability
4118 * of data (and SYN, FIN, of course) is checked separately.
4119 * See tcp_data_queue(), for example.
4121 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4122 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4123 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4124 * (borrowed from freebsd)
4127 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
4129 return !before(end_seq, tp->rcv_wup) &&
4130 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
4133 /* When we get a reset we do this. */
4134 void tcp_reset(struct sock *sk)
4136 trace_tcp_receive_reset(sk);
4138 /* We want the right error as BSD sees it (and indeed as we do). */
4139 switch (sk->sk_state) {
4141 sk->sk_err = ECONNREFUSED;
4143 case TCP_CLOSE_WAIT:
4149 sk->sk_err = ECONNRESET;
4151 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4154 tcp_write_queue_purge(sk);
4157 if (!sock_flag(sk, SOCK_DEAD))
4158 sk->sk_error_report(sk);
4162 * Process the FIN bit. This now behaves as it is supposed to work
4163 * and the FIN takes effect when it is validly part of sequence
4164 * space. Not before when we get holes.
4166 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4167 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4170 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4171 * close and we go into CLOSING (and later onto TIME-WAIT)
4173 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4175 void tcp_fin(struct sock *sk)
4177 struct tcp_sock *tp = tcp_sk(sk);
4179 inet_csk_schedule_ack(sk);
4181 sk->sk_shutdown |= RCV_SHUTDOWN;
4182 sock_set_flag(sk, SOCK_DONE);
4184 switch (sk->sk_state) {
4186 case TCP_ESTABLISHED:
4187 /* Move to CLOSE_WAIT */
4188 tcp_set_state(sk, TCP_CLOSE_WAIT);
4189 inet_csk_enter_pingpong_mode(sk);
4192 case TCP_CLOSE_WAIT:
4194 /* Received a retransmission of the FIN, do
4199 /* RFC793: Remain in the LAST-ACK state. */
4203 /* This case occurs when a simultaneous close
4204 * happens, we must ack the received FIN and
4205 * enter the CLOSING state.
4208 tcp_set_state(sk, TCP_CLOSING);
4211 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4213 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4216 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4217 * cases we should never reach this piece of code.
4219 pr_err("%s: Impossible, sk->sk_state=%d\n",
4220 __func__, sk->sk_state);
4224 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4225 * Probably, we should reset in this case. For now drop them.
4227 skb_rbtree_purge(&tp->out_of_order_queue);
4228 if (tcp_is_sack(tp))
4229 tcp_sack_reset(&tp->rx_opt);
4232 if (!sock_flag(sk, SOCK_DEAD)) {
4233 sk->sk_state_change(sk);
4235 /* Do not send POLL_HUP for half duplex close. */
4236 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4237 sk->sk_state == TCP_CLOSE)
4238 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4240 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4244 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4247 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4248 if (before(seq, sp->start_seq))
4249 sp->start_seq = seq;
4250 if (after(end_seq, sp->end_seq))
4251 sp->end_seq = end_seq;
4257 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4259 struct tcp_sock *tp = tcp_sk(sk);
4261 if (tcp_is_sack(tp) && sock_net(sk)->ipv4.sysctl_tcp_dsack) {
4264 if (before(seq, tp->rcv_nxt))
4265 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4267 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4269 NET_INC_STATS(sock_net(sk), mib_idx);
4271 tp->rx_opt.dsack = 1;
4272 tp->duplicate_sack[0].start_seq = seq;
4273 tp->duplicate_sack[0].end_seq = end_seq;
4277 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4279 struct tcp_sock *tp = tcp_sk(sk);
4281 if (!tp->rx_opt.dsack)
4282 tcp_dsack_set(sk, seq, end_seq);
4284 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4287 static void tcp_rcv_spurious_retrans(struct sock *sk, const struct sk_buff *skb)
4289 /* When the ACK path fails or drops most ACKs, the sender would
4290 * timeout and spuriously retransmit the same segment repeatedly.
4291 * The receiver remembers and reflects via DSACKs. Leverage the
4292 * DSACK state and change the txhash to re-route speculatively.
4294 if (TCP_SKB_CB(skb)->seq == tcp_sk(sk)->duplicate_sack[0].start_seq) {
4295 sk_rethink_txhash(sk);
4296 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDUPLICATEDATAREHASH);
4300 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4302 struct tcp_sock *tp = tcp_sk(sk);
4304 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4305 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4306 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4307 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
4309 if (tcp_is_sack(tp) && sock_net(sk)->ipv4.sysctl_tcp_dsack) {
4310 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4312 tcp_rcv_spurious_retrans(sk, skb);
4313 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4314 end_seq = tp->rcv_nxt;
4315 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4322 /* These routines update the SACK block as out-of-order packets arrive or
4323 * in-order packets close up the sequence space.
4325 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4328 struct tcp_sack_block *sp = &tp->selective_acks[0];
4329 struct tcp_sack_block *swalk = sp + 1;
4331 /* See if the recent change to the first SACK eats into
4332 * or hits the sequence space of other SACK blocks, if so coalesce.
4334 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4335 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4338 /* Zap SWALK, by moving every further SACK up by one slot.
4339 * Decrease num_sacks.
4341 tp->rx_opt.num_sacks--;
4342 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4346 this_sack++, swalk++;
4350 static void tcp_sack_compress_send_ack(struct sock *sk)
4352 struct tcp_sock *tp = tcp_sk(sk);
4354 if (!tp->compressed_ack)
4357 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
4360 /* Since we have to send one ack finally,
4361 * substract one from tp->compressed_ack to keep
4362 * LINUX_MIB_TCPACKCOMPRESSED accurate.
4364 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
4365 tp->compressed_ack - 1);
4367 tp->compressed_ack = 0;
4371 /* Reasonable amount of sack blocks included in TCP SACK option
4372 * The max is 4, but this becomes 3 if TCP timestamps are there.
4373 * Given that SACK packets might be lost, be conservative and use 2.
4375 #define TCP_SACK_BLOCKS_EXPECTED 2
4377 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4379 struct tcp_sock *tp = tcp_sk(sk);
4380 struct tcp_sack_block *sp = &tp->selective_acks[0];
4381 int cur_sacks = tp->rx_opt.num_sacks;
4387 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4388 if (tcp_sack_extend(sp, seq, end_seq)) {
4389 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED)
4390 tcp_sack_compress_send_ack(sk);
4391 /* Rotate this_sack to the first one. */
4392 for (; this_sack > 0; this_sack--, sp--)
4393 swap(*sp, *(sp - 1));
4395 tcp_sack_maybe_coalesce(tp);
4400 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED)
4401 tcp_sack_compress_send_ack(sk);
4403 /* Could not find an adjacent existing SACK, build a new one,
4404 * put it at the front, and shift everyone else down. We
4405 * always know there is at least one SACK present already here.
4407 * If the sack array is full, forget about the last one.
4409 if (this_sack >= TCP_NUM_SACKS) {
4411 tp->rx_opt.num_sacks--;
4414 for (; this_sack > 0; this_sack--, sp--)
4418 /* Build the new head SACK, and we're done. */
4419 sp->start_seq = seq;
4420 sp->end_seq = end_seq;
4421 tp->rx_opt.num_sacks++;
4424 /* RCV.NXT advances, some SACKs should be eaten. */
4426 static void tcp_sack_remove(struct tcp_sock *tp)
4428 struct tcp_sack_block *sp = &tp->selective_acks[0];
4429 int num_sacks = tp->rx_opt.num_sacks;
4432 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4433 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4434 tp->rx_opt.num_sacks = 0;
4438 for (this_sack = 0; this_sack < num_sacks;) {
4439 /* Check if the start of the sack is covered by RCV.NXT. */
4440 if (!before(tp->rcv_nxt, sp->start_seq)) {
4443 /* RCV.NXT must cover all the block! */
4444 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4446 /* Zap this SACK, by moving forward any other SACKS. */
4447 for (i = this_sack+1; i < num_sacks; i++)
4448 tp->selective_acks[i-1] = tp->selective_acks[i];
4455 tp->rx_opt.num_sacks = num_sacks;
4459 * tcp_try_coalesce - try to merge skb to prior one
4462 * @from: buffer to add in queue
4463 * @fragstolen: pointer to boolean
4465 * Before queueing skb @from after @to, try to merge them
4466 * to reduce overall memory use and queue lengths, if cost is small.
4467 * Packets in ofo or receive queues can stay a long time.
4468 * Better try to coalesce them right now to avoid future collapses.
4469 * Returns true if caller should free @from instead of queueing it
4471 static bool tcp_try_coalesce(struct sock *sk,
4473 struct sk_buff *from,
4478 *fragstolen = false;
4480 /* Its possible this segment overlaps with prior segment in queue */
4481 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4484 if (!mptcp_skb_can_collapse(to, from))
4487 #ifdef CONFIG_TLS_DEVICE
4488 if (from->decrypted != to->decrypted)
4492 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4495 atomic_add(delta, &sk->sk_rmem_alloc);
4496 sk_mem_charge(sk, delta);
4497 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4498 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4499 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4500 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4502 if (TCP_SKB_CB(from)->has_rxtstamp) {
4503 TCP_SKB_CB(to)->has_rxtstamp = true;
4504 to->tstamp = from->tstamp;
4505 skb_hwtstamps(to)->hwtstamp = skb_hwtstamps(from)->hwtstamp;
4511 static bool tcp_ooo_try_coalesce(struct sock *sk,
4513 struct sk_buff *from,
4516 bool res = tcp_try_coalesce(sk, to, from, fragstolen);
4518 /* In case tcp_drop() is called later, update to->gso_segs */
4520 u32 gso_segs = max_t(u16, 1, skb_shinfo(to)->gso_segs) +
4521 max_t(u16, 1, skb_shinfo(from)->gso_segs);
4523 skb_shinfo(to)->gso_segs = min_t(u32, gso_segs, 0xFFFF);
4528 static void tcp_drop(struct sock *sk, struct sk_buff *skb)
4530 sk_drops_add(sk, skb);
4534 /* This one checks to see if we can put data from the
4535 * out_of_order queue into the receive_queue.
4537 static void tcp_ofo_queue(struct sock *sk)
4539 struct tcp_sock *tp = tcp_sk(sk);
4540 __u32 dsack_high = tp->rcv_nxt;
4541 bool fin, fragstolen, eaten;
4542 struct sk_buff *skb, *tail;
4545 p = rb_first(&tp->out_of_order_queue);
4548 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4551 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4552 __u32 dsack = dsack_high;
4553 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4554 dsack_high = TCP_SKB_CB(skb)->end_seq;
4555 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4558 rb_erase(&skb->rbnode, &tp->out_of_order_queue);
4560 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) {
4565 tail = skb_peek_tail(&sk->sk_receive_queue);
4566 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4567 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4568 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
4570 __skb_queue_tail(&sk->sk_receive_queue, skb);
4572 kfree_skb_partial(skb, fragstolen);
4574 if (unlikely(fin)) {
4576 /* tcp_fin() purges tp->out_of_order_queue,
4577 * so we must end this loop right now.
4584 static bool tcp_prune_ofo_queue(struct sock *sk);
4585 static int tcp_prune_queue(struct sock *sk);
4587 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4590 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4591 !sk_rmem_schedule(sk, skb, size)) {
4593 if (tcp_prune_queue(sk) < 0)
4596 while (!sk_rmem_schedule(sk, skb, size)) {
4597 if (!tcp_prune_ofo_queue(sk))
4604 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4606 struct tcp_sock *tp = tcp_sk(sk);
4607 struct rb_node **p, *parent;
4608 struct sk_buff *skb1;
4612 tcp_ecn_check_ce(sk, skb);
4614 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4615 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP);
4616 sk->sk_data_ready(sk);
4621 /* Disable header prediction. */
4623 inet_csk_schedule_ack(sk);
4625 tp->rcv_ooopack += max_t(u16, 1, skb_shinfo(skb)->gso_segs);
4626 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4627 seq = TCP_SKB_CB(skb)->seq;
4628 end_seq = TCP_SKB_CB(skb)->end_seq;
4630 p = &tp->out_of_order_queue.rb_node;
4631 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4632 /* Initial out of order segment, build 1 SACK. */
4633 if (tcp_is_sack(tp)) {
4634 tp->rx_opt.num_sacks = 1;
4635 tp->selective_acks[0].start_seq = seq;
4636 tp->selective_acks[0].end_seq = end_seq;
4638 rb_link_node(&skb->rbnode, NULL, p);
4639 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4640 tp->ooo_last_skb = skb;
4644 /* In the typical case, we are adding an skb to the end of the list.
4645 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4647 if (tcp_ooo_try_coalesce(sk, tp->ooo_last_skb,
4648 skb, &fragstolen)) {
4650 /* For non sack flows, do not grow window to force DUPACK
4651 * and trigger fast retransmit.
4653 if (tcp_is_sack(tp))
4654 tcp_grow_window(sk, skb);
4655 kfree_skb_partial(skb, fragstolen);
4659 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4660 if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) {
4661 parent = &tp->ooo_last_skb->rbnode;
4662 p = &parent->rb_right;
4666 /* Find place to insert this segment. Handle overlaps on the way. */
4670 skb1 = rb_to_skb(parent);
4671 if (before(seq, TCP_SKB_CB(skb1)->seq)) {
4672 p = &parent->rb_left;
4675 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4676 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4677 /* All the bits are present. Drop. */
4678 NET_INC_STATS(sock_net(sk),
4679 LINUX_MIB_TCPOFOMERGE);
4682 tcp_dsack_set(sk, seq, end_seq);
4685 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4686 /* Partial overlap. */
4687 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq);
4689 /* skb's seq == skb1's seq and skb covers skb1.
4690 * Replace skb1 with skb.
4692 rb_replace_node(&skb1->rbnode, &skb->rbnode,
4693 &tp->out_of_order_queue);
4694 tcp_dsack_extend(sk,
4695 TCP_SKB_CB(skb1)->seq,
4696 TCP_SKB_CB(skb1)->end_seq);
4697 NET_INC_STATS(sock_net(sk),
4698 LINUX_MIB_TCPOFOMERGE);
4702 } else if (tcp_ooo_try_coalesce(sk, skb1,
4703 skb, &fragstolen)) {
4706 p = &parent->rb_right;
4709 /* Insert segment into RB tree. */
4710 rb_link_node(&skb->rbnode, parent, p);
4711 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4714 /* Remove other segments covered by skb. */
4715 while ((skb1 = skb_rb_next(skb)) != NULL) {
4716 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4718 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4719 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4723 rb_erase(&skb1->rbnode, &tp->out_of_order_queue);
4724 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4725 TCP_SKB_CB(skb1)->end_seq);
4726 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4729 /* If there is no skb after us, we are the last_skb ! */
4731 tp->ooo_last_skb = skb;
4734 if (tcp_is_sack(tp))
4735 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4738 /* For non sack flows, do not grow window to force DUPACK
4739 * and trigger fast retransmit.
4741 if (tcp_is_sack(tp))
4742 tcp_grow_window(sk, skb);
4744 skb_set_owner_r(skb, sk);
4748 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb,
4752 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4755 tcp_try_coalesce(sk, tail,
4756 skb, fragstolen)) ? 1 : 0;
4757 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
4759 __skb_queue_tail(&sk->sk_receive_queue, skb);
4760 skb_set_owner_r(skb, sk);
4765 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4767 struct sk_buff *skb;
4775 if (size > PAGE_SIZE) {
4776 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
4778 data_len = npages << PAGE_SHIFT;
4779 size = data_len + (size & ~PAGE_MASK);
4781 skb = alloc_skb_with_frags(size - data_len, data_len,
4782 PAGE_ALLOC_COSTLY_ORDER,
4783 &err, sk->sk_allocation);
4787 skb_put(skb, size - data_len);
4788 skb->data_len = data_len;
4791 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
4792 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
4796 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
4800 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4801 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4802 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4804 if (tcp_queue_rcv(sk, skb, &fragstolen)) {
4805 WARN_ON_ONCE(fragstolen); /* should not happen */
4817 void tcp_data_ready(struct sock *sk)
4819 const struct tcp_sock *tp = tcp_sk(sk);
4820 int avail = tp->rcv_nxt - tp->copied_seq;
4822 if (avail < sk->sk_rcvlowat && !tcp_rmem_pressure(sk) &&
4823 !sock_flag(sk, SOCK_DONE))
4826 sk->sk_data_ready(sk);
4829 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4831 struct tcp_sock *tp = tcp_sk(sk);
4835 if (sk_is_mptcp(sk))
4836 mptcp_incoming_options(sk, skb, &tp->rx_opt);
4838 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
4843 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
4845 tp->rx_opt.dsack = 0;
4847 /* Queue data for delivery to the user.
4848 * Packets in sequence go to the receive queue.
4849 * Out of sequence packets to the out_of_order_queue.
4851 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4852 if (tcp_receive_window(tp) == 0) {
4853 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
4857 /* Ok. In sequence. In window. */
4859 if (skb_queue_len(&sk->sk_receive_queue) == 0)
4860 sk_forced_mem_schedule(sk, skb->truesize);
4861 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
4862 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
4863 sk->sk_data_ready(sk);
4867 eaten = tcp_queue_rcv(sk, skb, &fragstolen);
4869 tcp_event_data_recv(sk, skb);
4870 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4873 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4876 /* RFC5681. 4.2. SHOULD send immediate ACK, when
4877 * gap in queue is filled.
4879 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
4880 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
4883 if (tp->rx_opt.num_sacks)
4884 tcp_sack_remove(tp);
4886 tcp_fast_path_check(sk);
4889 kfree_skb_partial(skb, fragstolen);
4890 if (!sock_flag(sk, SOCK_DEAD))
4895 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4896 tcp_rcv_spurious_retrans(sk, skb);
4897 /* A retransmit, 2nd most common case. Force an immediate ack. */
4898 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4899 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4902 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
4903 inet_csk_schedule_ack(sk);
4909 /* Out of window. F.e. zero window probe. */
4910 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4913 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4914 /* Partial packet, seq < rcv_next < end_seq */
4915 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4917 /* If window is closed, drop tail of packet. But after
4918 * remembering D-SACK for its head made in previous line.
4920 if (!tcp_receive_window(tp)) {
4921 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
4927 tcp_data_queue_ofo(sk, skb);
4930 static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list)
4933 return !skb_queue_is_last(list, skb) ? skb->next : NULL;
4935 return skb_rb_next(skb);
4938 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4939 struct sk_buff_head *list,
4940 struct rb_root *root)
4942 struct sk_buff *next = tcp_skb_next(skb, list);
4945 __skb_unlink(skb, list);
4947 rb_erase(&skb->rbnode, root);
4950 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4955 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
4956 void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb)
4958 struct rb_node **p = &root->rb_node;
4959 struct rb_node *parent = NULL;
4960 struct sk_buff *skb1;
4964 skb1 = rb_to_skb(parent);
4965 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq))
4966 p = &parent->rb_left;
4968 p = &parent->rb_right;
4970 rb_link_node(&skb->rbnode, parent, p);
4971 rb_insert_color(&skb->rbnode, root);
4974 /* Collapse contiguous sequence of skbs head..tail with
4975 * sequence numbers start..end.
4977 * If tail is NULL, this means until the end of the queue.
4979 * Segments with FIN/SYN are not collapsed (only because this
4983 tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root,
4984 struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end)
4986 struct sk_buff *skb = head, *n;
4987 struct sk_buff_head tmp;
4990 /* First, check that queue is collapsible and find
4991 * the point where collapsing can be useful.
4994 for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) {
4995 n = tcp_skb_next(skb, list);
4997 /* No new bits? It is possible on ofo queue. */
4998 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4999 skb = tcp_collapse_one(sk, skb, list, root);
5005 /* The first skb to collapse is:
5007 * - bloated or contains data before "start" or
5008 * overlaps to the next one and mptcp allow collapsing.
5010 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
5011 (tcp_win_from_space(sk, skb->truesize) > skb->len ||
5012 before(TCP_SKB_CB(skb)->seq, start))) {
5013 end_of_skbs = false;
5017 if (n && n != tail && mptcp_skb_can_collapse(skb, n) &&
5018 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) {
5019 end_of_skbs = false;
5023 /* Decided to skip this, advance start seq. */
5024 start = TCP_SKB_CB(skb)->end_seq;
5027 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
5030 __skb_queue_head_init(&tmp);
5032 while (before(start, end)) {
5033 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
5034 struct sk_buff *nskb;
5036 nskb = alloc_skb(copy, GFP_ATOMIC);
5040 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
5041 #ifdef CONFIG_TLS_DEVICE
5042 nskb->decrypted = skb->decrypted;
5044 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
5046 __skb_queue_before(list, skb, nskb);
5048 __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */
5049 skb_set_owner_r(nskb, sk);
5050 mptcp_skb_ext_move(nskb, skb);
5052 /* Copy data, releasing collapsed skbs. */
5054 int offset = start - TCP_SKB_CB(skb)->seq;
5055 int size = TCP_SKB_CB(skb)->end_seq - start;
5059 size = min(copy, size);
5060 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
5062 TCP_SKB_CB(nskb)->end_seq += size;
5066 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
5067 skb = tcp_collapse_one(sk, skb, list, root);
5070 !mptcp_skb_can_collapse(nskb, skb) ||
5071 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
5073 #ifdef CONFIG_TLS_DEVICE
5074 if (skb->decrypted != nskb->decrypted)
5081 skb_queue_walk_safe(&tmp, skb, n)
5082 tcp_rbtree_insert(root, skb);
5085 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
5086 * and tcp_collapse() them until all the queue is collapsed.
5088 static void tcp_collapse_ofo_queue(struct sock *sk)
5090 struct tcp_sock *tp = tcp_sk(sk);
5091 u32 range_truesize, sum_tiny = 0;
5092 struct sk_buff *skb, *head;
5095 skb = skb_rb_first(&tp->out_of_order_queue);
5098 tp->ooo_last_skb = skb_rb_last(&tp->out_of_order_queue);
5101 start = TCP_SKB_CB(skb)->seq;
5102 end = TCP_SKB_CB(skb)->end_seq;
5103 range_truesize = skb->truesize;
5105 for (head = skb;;) {
5106 skb = skb_rb_next(skb);
5108 /* Range is terminated when we see a gap or when
5109 * we are at the queue end.
5112 after(TCP_SKB_CB(skb)->seq, end) ||
5113 before(TCP_SKB_CB(skb)->end_seq, start)) {
5114 /* Do not attempt collapsing tiny skbs */
5115 if (range_truesize != head->truesize ||
5116 end - start >= SKB_WITH_OVERHEAD(SK_MEM_QUANTUM)) {
5117 tcp_collapse(sk, NULL, &tp->out_of_order_queue,
5118 head, skb, start, end);
5120 sum_tiny += range_truesize;
5121 if (sum_tiny > sk->sk_rcvbuf >> 3)
5127 range_truesize += skb->truesize;
5128 if (unlikely(before(TCP_SKB_CB(skb)->seq, start)))
5129 start = TCP_SKB_CB(skb)->seq;
5130 if (after(TCP_SKB_CB(skb)->end_seq, end))
5131 end = TCP_SKB_CB(skb)->end_seq;
5136 * Clean the out-of-order queue to make room.
5137 * We drop high sequences packets to :
5138 * 1) Let a chance for holes to be filled.
5139 * 2) not add too big latencies if thousands of packets sit there.
5140 * (But if application shrinks SO_RCVBUF, we could still end up
5141 * freeing whole queue here)
5142 * 3) Drop at least 12.5 % of sk_rcvbuf to avoid malicious attacks.
5144 * Return true if queue has shrunk.
5146 static bool tcp_prune_ofo_queue(struct sock *sk)
5148 struct tcp_sock *tp = tcp_sk(sk);
5149 struct rb_node *node, *prev;
5152 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
5155 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED);
5156 goal = sk->sk_rcvbuf >> 3;
5157 node = &tp->ooo_last_skb->rbnode;
5159 prev = rb_prev(node);
5160 rb_erase(node, &tp->out_of_order_queue);
5161 goal -= rb_to_skb(node)->truesize;
5162 tcp_drop(sk, rb_to_skb(node));
5163 if (!prev || goal <= 0) {
5165 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf &&
5166 !tcp_under_memory_pressure(sk))
5168 goal = sk->sk_rcvbuf >> 3;
5172 tp->ooo_last_skb = rb_to_skb(prev);
5174 /* Reset SACK state. A conforming SACK implementation will
5175 * do the same at a timeout based retransmit. When a connection
5176 * is in a sad state like this, we care only about integrity
5177 * of the connection not performance.
5179 if (tp->rx_opt.sack_ok)
5180 tcp_sack_reset(&tp->rx_opt);
5184 /* Reduce allocated memory if we can, trying to get
5185 * the socket within its memory limits again.
5187 * Return less than zero if we should start dropping frames
5188 * until the socket owning process reads some of the data
5189 * to stabilize the situation.
5191 static int tcp_prune_queue(struct sock *sk)
5193 struct tcp_sock *tp = tcp_sk(sk);
5195 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED);
5197 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
5198 tcp_clamp_window(sk);
5199 else if (tcp_under_memory_pressure(sk))
5200 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
5202 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5205 tcp_collapse_ofo_queue(sk);
5206 if (!skb_queue_empty(&sk->sk_receive_queue))
5207 tcp_collapse(sk, &sk->sk_receive_queue, NULL,
5208 skb_peek(&sk->sk_receive_queue),
5210 tp->copied_seq, tp->rcv_nxt);
5213 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5216 /* Collapsing did not help, destructive actions follow.
5217 * This must not ever occur. */
5219 tcp_prune_ofo_queue(sk);
5221 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5224 /* If we are really being abused, tell the caller to silently
5225 * drop receive data on the floor. It will get retransmitted
5226 * and hopefully then we'll have sufficient space.
5228 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED);
5230 /* Massive buffer overcommit. */
5235 static bool tcp_should_expand_sndbuf(const struct sock *sk)
5237 const struct tcp_sock *tp = tcp_sk(sk);
5239 /* If the user specified a specific send buffer setting, do
5242 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
5245 /* If we are under global TCP memory pressure, do not expand. */
5246 if (tcp_under_memory_pressure(sk))
5249 /* If we are under soft global TCP memory pressure, do not expand. */
5250 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
5253 /* If we filled the congestion window, do not expand. */
5254 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
5260 /* When incoming ACK allowed to free some skb from write_queue,
5261 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
5262 * on the exit from tcp input handler.
5264 * PROBLEM: sndbuf expansion does not work well with largesend.
5266 static void tcp_new_space(struct sock *sk)
5268 struct tcp_sock *tp = tcp_sk(sk);
5270 if (tcp_should_expand_sndbuf(sk)) {
5271 tcp_sndbuf_expand(sk);
5272 tp->snd_cwnd_stamp = tcp_jiffies32;
5275 sk->sk_write_space(sk);
5278 static void tcp_check_space(struct sock *sk)
5280 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
5281 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
5282 /* pairs with tcp_poll() */
5284 if (sk->sk_socket &&
5285 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
5287 if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5288 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
5293 static inline void tcp_data_snd_check(struct sock *sk)
5295 tcp_push_pending_frames(sk);
5296 tcp_check_space(sk);
5300 * Check if sending an ack is needed.
5302 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5304 struct tcp_sock *tp = tcp_sk(sk);
5305 unsigned long rtt, delay;
5307 /* More than one full frame received... */
5308 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5309 /* ... and right edge of window advances far enough.
5310 * (tcp_recvmsg() will send ACK otherwise).
5311 * If application uses SO_RCVLOWAT, we want send ack now if
5312 * we have not received enough bytes to satisfy the condition.
5314 (tp->rcv_nxt - tp->copied_seq < sk->sk_rcvlowat ||
5315 __tcp_select_window(sk) >= tp->rcv_wnd)) ||
5316 /* We ACK each frame or... */
5317 tcp_in_quickack_mode(sk) ||
5318 /* Protocol state mandates a one-time immediate ACK */
5319 inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOW) {
5325 if (!ofo_possible || RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5326 tcp_send_delayed_ack(sk);
5330 if (!tcp_is_sack(tp) ||
5331 tp->compressed_ack >= sock_net(sk)->ipv4.sysctl_tcp_comp_sack_nr)
5334 if (tp->compressed_ack_rcv_nxt != tp->rcv_nxt) {
5335 tp->compressed_ack_rcv_nxt = tp->rcv_nxt;
5336 tp->dup_ack_counter = 0;
5338 if (tp->dup_ack_counter < TCP_FASTRETRANS_THRESH) {
5339 tp->dup_ack_counter++;
5342 tp->compressed_ack++;
5343 if (hrtimer_is_queued(&tp->compressed_ack_timer))
5346 /* compress ack timer : 5 % of rtt, but no more than tcp_comp_sack_delay_ns */
5348 rtt = tp->rcv_rtt_est.rtt_us;
5349 if (tp->srtt_us && tp->srtt_us < rtt)
5352 delay = min_t(unsigned long, sock_net(sk)->ipv4.sysctl_tcp_comp_sack_delay_ns,
5353 rtt * (NSEC_PER_USEC >> 3)/20);
5355 hrtimer_start_range_ns(&tp->compressed_ack_timer, ns_to_ktime(delay),
5356 sock_net(sk)->ipv4.sysctl_tcp_comp_sack_slack_ns,
5357 HRTIMER_MODE_REL_PINNED_SOFT);
5360 static inline void tcp_ack_snd_check(struct sock *sk)
5362 if (!inet_csk_ack_scheduled(sk)) {
5363 /* We sent a data segment already. */
5366 __tcp_ack_snd_check(sk, 1);
5370 * This routine is only called when we have urgent data
5371 * signaled. Its the 'slow' part of tcp_urg. It could be
5372 * moved inline now as tcp_urg is only called from one
5373 * place. We handle URGent data wrong. We have to - as
5374 * BSD still doesn't use the correction from RFC961.
5375 * For 1003.1g we should support a new option TCP_STDURG to permit
5376 * either form (or just set the sysctl tcp_stdurg).
5379 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5381 struct tcp_sock *tp = tcp_sk(sk);
5382 u32 ptr = ntohs(th->urg_ptr);
5384 if (ptr && !sock_net(sk)->ipv4.sysctl_tcp_stdurg)
5386 ptr += ntohl(th->seq);
5388 /* Ignore urgent data that we've already seen and read. */
5389 if (after(tp->copied_seq, ptr))
5392 /* Do not replay urg ptr.
5394 * NOTE: interesting situation not covered by specs.
5395 * Misbehaving sender may send urg ptr, pointing to segment,
5396 * which we already have in ofo queue. We are not able to fetch
5397 * such data and will stay in TCP_URG_NOTYET until will be eaten
5398 * by recvmsg(). Seems, we are not obliged to handle such wicked
5399 * situations. But it is worth to think about possibility of some
5400 * DoSes using some hypothetical application level deadlock.
5402 if (before(ptr, tp->rcv_nxt))
5405 /* Do we already have a newer (or duplicate) urgent pointer? */
5406 if (tp->urg_data && !after(ptr, tp->urg_seq))
5409 /* Tell the world about our new urgent pointer. */
5412 /* We may be adding urgent data when the last byte read was
5413 * urgent. To do this requires some care. We cannot just ignore
5414 * tp->copied_seq since we would read the last urgent byte again
5415 * as data, nor can we alter copied_seq until this data arrives
5416 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5418 * NOTE. Double Dutch. Rendering to plain English: author of comment
5419 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5420 * and expect that both A and B disappear from stream. This is _wrong_.
5421 * Though this happens in BSD with high probability, this is occasional.
5422 * Any application relying on this is buggy. Note also, that fix "works"
5423 * only in this artificial test. Insert some normal data between A and B and we will
5424 * decline of BSD again. Verdict: it is better to remove to trap
5427 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5428 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5429 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5431 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5432 __skb_unlink(skb, &sk->sk_receive_queue);
5437 tp->urg_data = TCP_URG_NOTYET;
5438 WRITE_ONCE(tp->urg_seq, ptr);
5440 /* Disable header prediction. */
5444 /* This is the 'fast' part of urgent handling. */
5445 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5447 struct tcp_sock *tp = tcp_sk(sk);
5449 /* Check if we get a new urgent pointer - normally not. */
5451 tcp_check_urg(sk, th);
5453 /* Do we wait for any urgent data? - normally not... */
5454 if (tp->urg_data == TCP_URG_NOTYET) {
5455 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5458 /* Is the urgent pointer pointing into this packet? */
5459 if (ptr < skb->len) {
5461 if (skb_copy_bits(skb, ptr, &tmp, 1))
5463 tp->urg_data = TCP_URG_VALID | tmp;
5464 if (!sock_flag(sk, SOCK_DEAD))
5465 sk->sk_data_ready(sk);
5470 /* Accept RST for rcv_nxt - 1 after a FIN.
5471 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a
5472 * FIN is sent followed by a RST packet. The RST is sent with the same
5473 * sequence number as the FIN, and thus according to RFC 5961 a challenge
5474 * ACK should be sent. However, Mac OSX rate limits replies to challenge
5475 * ACKs on the closed socket. In addition middleboxes can drop either the
5476 * challenge ACK or a subsequent RST.
5478 static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb)
5480 struct tcp_sock *tp = tcp_sk(sk);
5482 return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) &&
5483 (1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK |
5487 /* Does PAWS and seqno based validation of an incoming segment, flags will
5488 * play significant role here.
5490 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5491 const struct tcphdr *th, int syn_inerr)
5493 struct tcp_sock *tp = tcp_sk(sk);
5494 bool rst_seq_match = false;
5496 /* RFC1323: H1. Apply PAWS check first. */
5497 if (tcp_fast_parse_options(sock_net(sk), skb, th, tp) &&
5498 tp->rx_opt.saw_tstamp &&
5499 tcp_paws_discard(sk, skb)) {
5501 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5502 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5503 LINUX_MIB_TCPACKSKIPPEDPAWS,
5504 &tp->last_oow_ack_time))
5505 tcp_send_dupack(sk, skb);
5508 /* Reset is accepted even if it did not pass PAWS. */
5511 /* Step 1: check sequence number */
5512 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5513 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5514 * (RST) segments are validated by checking their SEQ-fields."
5515 * And page 69: "If an incoming segment is not acceptable,
5516 * an acknowledgment should be sent in reply (unless the RST
5517 * bit is set, if so drop the segment and return)".
5522 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5523 LINUX_MIB_TCPACKSKIPPEDSEQ,
5524 &tp->last_oow_ack_time))
5525 tcp_send_dupack(sk, skb);
5526 } else if (tcp_reset_check(sk, skb)) {
5532 /* Step 2: check RST bit */
5534 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a
5535 * FIN and SACK too if available):
5536 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or
5537 * the right-most SACK block,
5539 * RESET the connection
5541 * Send a challenge ACK
5543 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt ||
5544 tcp_reset_check(sk, skb)) {
5545 rst_seq_match = true;
5546 } else if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) {
5547 struct tcp_sack_block *sp = &tp->selective_acks[0];
5548 int max_sack = sp[0].end_seq;
5551 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;
5553 max_sack = after(sp[this_sack].end_seq,
5555 sp[this_sack].end_seq : max_sack;
5558 if (TCP_SKB_CB(skb)->seq == max_sack)
5559 rst_seq_match = true;
5565 /* Disable TFO if RST is out-of-order
5566 * and no data has been received
5567 * for current active TFO socket
5569 if (tp->syn_fastopen && !tp->data_segs_in &&
5570 sk->sk_state == TCP_ESTABLISHED)
5571 tcp_fastopen_active_disable(sk);
5572 tcp_send_challenge_ack(sk, skb);
5577 /* step 3: check security and precedence [ignored] */
5579 /* step 4: Check for a SYN
5580 * RFC 5961 4.2 : Send a challenge ack
5585 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5586 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5587 tcp_send_challenge_ack(sk, skb);
5599 * TCP receive function for the ESTABLISHED state.
5601 * It is split into a fast path and a slow path. The fast path is
5603 * - A zero window was announced from us - zero window probing
5604 * is only handled properly in the slow path.
5605 * - Out of order segments arrived.
5606 * - Urgent data is expected.
5607 * - There is no buffer space left
5608 * - Unexpected TCP flags/window values/header lengths are received
5609 * (detected by checking the TCP header against pred_flags)
5610 * - Data is sent in both directions. Fast path only supports pure senders
5611 * or pure receivers (this means either the sequence number or the ack
5612 * value must stay constant)
5613 * - Unexpected TCP option.
5615 * When these conditions are not satisfied it drops into a standard
5616 * receive procedure patterned after RFC793 to handle all cases.
5617 * The first three cases are guaranteed by proper pred_flags setting,
5618 * the rest is checked inline. Fast processing is turned on in
5619 * tcp_data_queue when everything is OK.
5621 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb)
5623 const struct tcphdr *th = (const struct tcphdr *)skb->data;
5624 struct tcp_sock *tp = tcp_sk(sk);
5625 unsigned int len = skb->len;
5627 /* TCP congestion window tracking */
5628 trace_tcp_probe(sk, skb);
5630 tcp_mstamp_refresh(tp);
5631 if (unlikely(!sk->sk_rx_dst))
5632 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5634 * Header prediction.
5635 * The code loosely follows the one in the famous
5636 * "30 instruction TCP receive" Van Jacobson mail.
5638 * Van's trick is to deposit buffers into socket queue
5639 * on a device interrupt, to call tcp_recv function
5640 * on the receive process context and checksum and copy
5641 * the buffer to user space. smart...
5643 * Our current scheme is not silly either but we take the
5644 * extra cost of the net_bh soft interrupt processing...
5645 * We do checksum and copy also but from device to kernel.
5648 tp->rx_opt.saw_tstamp = 0;
5650 /* pred_flags is 0xS?10 << 16 + snd_wnd
5651 * if header_prediction is to be made
5652 * 'S' will always be tp->tcp_header_len >> 2
5653 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5654 * turn it off (when there are holes in the receive
5655 * space for instance)
5656 * PSH flag is ignored.
5659 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5660 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5661 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5662 int tcp_header_len = tp->tcp_header_len;
5664 /* Timestamp header prediction: tcp_header_len
5665 * is automatically equal to th->doff*4 due to pred_flags
5669 /* Check timestamp */
5670 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5671 /* No? Slow path! */
5672 if (!tcp_parse_aligned_timestamp(tp, th))
5675 /* If PAWS failed, check it more carefully in slow path */
5676 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5679 /* DO NOT update ts_recent here, if checksum fails
5680 * and timestamp was corrupted part, it will result
5681 * in a hung connection since we will drop all
5682 * future packets due to the PAWS test.
5686 if (len <= tcp_header_len) {
5687 /* Bulk data transfer: sender */
5688 if (len == tcp_header_len) {
5689 /* Predicted packet is in window by definition.
5690 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5691 * Hence, check seq<=rcv_wup reduces to:
5693 if (tcp_header_len ==
5694 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5695 tp->rcv_nxt == tp->rcv_wup)
5696 tcp_store_ts_recent(tp);
5698 /* We know that such packets are checksummed
5701 tcp_ack(sk, skb, 0);
5703 tcp_data_snd_check(sk);
5704 /* When receiving pure ack in fast path, update
5705 * last ts ecr directly instead of calling
5706 * tcp_rcv_rtt_measure_ts()
5708 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
5710 } else { /* Header too small */
5711 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5716 bool fragstolen = false;
5718 if (tcp_checksum_complete(skb))
5721 if ((int)skb->truesize > sk->sk_forward_alloc)
5724 /* Predicted packet is in window by definition.
5725 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5726 * Hence, check seq<=rcv_wup reduces to:
5728 if (tcp_header_len ==
5729 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5730 tp->rcv_nxt == tp->rcv_wup)
5731 tcp_store_ts_recent(tp);
5733 tcp_rcv_rtt_measure_ts(sk, skb);
5735 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS);
5737 /* Bulk data transfer: receiver */
5738 __skb_pull(skb, tcp_header_len);
5739 eaten = tcp_queue_rcv(sk, skb, &fragstolen);
5741 tcp_event_data_recv(sk, skb);
5743 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5744 /* Well, only one small jumplet in fast path... */
5745 tcp_ack(sk, skb, FLAG_DATA);
5746 tcp_data_snd_check(sk);
5747 if (!inet_csk_ack_scheduled(sk))
5751 __tcp_ack_snd_check(sk, 0);
5754 kfree_skb_partial(skb, fragstolen);
5761 if (len < (th->doff << 2) || tcp_checksum_complete(skb))
5764 if (!th->ack && !th->rst && !th->syn)
5768 * Standard slow path.
5771 if (!tcp_validate_incoming(sk, skb, th, 1))
5775 if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5778 tcp_rcv_rtt_measure_ts(sk, skb);
5780 /* Process urgent data. */
5781 tcp_urg(sk, skb, th);
5783 /* step 7: process the segment text */
5784 tcp_data_queue(sk, skb);
5786 tcp_data_snd_check(sk);
5787 tcp_ack_snd_check(sk);
5791 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
5792 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5797 EXPORT_SYMBOL(tcp_rcv_established);
5799 void tcp_init_transfer(struct sock *sk, int bpf_op)
5801 struct inet_connection_sock *icsk = inet_csk(sk);
5802 struct tcp_sock *tp = tcp_sk(sk);
5805 icsk->icsk_af_ops->rebuild_header(sk);
5806 tcp_init_metrics(sk);
5808 /* Initialize the congestion window to start the transfer.
5809 * Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
5810 * retransmitted. In light of RFC6298 more aggressive 1sec
5811 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
5812 * retransmission has occurred.
5814 if (tp->total_retrans > 1 && tp->undo_marker)
5817 tp->snd_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
5818 tp->snd_cwnd_stamp = tcp_jiffies32;
5820 tcp_call_bpf(sk, bpf_op, 0, NULL);
5821 tcp_init_congestion_control(sk);
5822 tcp_init_buffer_space(sk);
5825 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5827 struct tcp_sock *tp = tcp_sk(sk);
5828 struct inet_connection_sock *icsk = inet_csk(sk);
5830 tcp_set_state(sk, TCP_ESTABLISHED);
5831 icsk->icsk_ack.lrcvtime = tcp_jiffies32;
5834 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5835 security_inet_conn_established(sk, skb);
5836 sk_mark_napi_id(sk, skb);
5839 tcp_init_transfer(sk, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB);
5841 /* Prevent spurious tcp_cwnd_restart() on first data
5844 tp->lsndtime = tcp_jiffies32;
5846 if (sock_flag(sk, SOCK_KEEPOPEN))
5847 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5849 if (!tp->rx_opt.snd_wscale)
5850 __tcp_fast_path_on(tp, tp->snd_wnd);
5855 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5856 struct tcp_fastopen_cookie *cookie)
5858 struct tcp_sock *tp = tcp_sk(sk);
5859 struct sk_buff *data = tp->syn_data ? tcp_rtx_queue_head(sk) : NULL;
5860 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
5861 bool syn_drop = false;
5863 if (mss == tp->rx_opt.user_mss) {
5864 struct tcp_options_received opt;
5866 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5867 tcp_clear_options(&opt);
5868 opt.user_mss = opt.mss_clamp = 0;
5869 tcp_parse_options(sock_net(sk), synack, &opt, 0, NULL);
5870 mss = opt.mss_clamp;
5873 if (!tp->syn_fastopen) {
5874 /* Ignore an unsolicited cookie */
5876 } else if (tp->total_retrans) {
5877 /* SYN timed out and the SYN-ACK neither has a cookie nor
5878 * acknowledges data. Presumably the remote received only
5879 * the retransmitted (regular) SYNs: either the original
5880 * SYN-data or the corresponding SYN-ACK was dropped.
5882 syn_drop = (cookie->len < 0 && data);
5883 } else if (cookie->len < 0 && !tp->syn_data) {
5884 /* We requested a cookie but didn't get it. If we did not use
5885 * the (old) exp opt format then try so next time (try_exp=1).
5886 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5888 try_exp = tp->syn_fastopen_exp ? 2 : 1;
5891 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
5893 if (data) { /* Retransmit unacked data in SYN */
5894 if (tp->total_retrans)
5895 tp->fastopen_client_fail = TFO_SYN_RETRANSMITTED;
5897 tp->fastopen_client_fail = TFO_DATA_NOT_ACKED;
5898 skb_rbtree_walk_from(data) {
5899 if (__tcp_retransmit_skb(sk, data, 1))
5903 NET_INC_STATS(sock_net(sk),
5904 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
5907 tp->syn_data_acked = tp->syn_data;
5908 if (tp->syn_data_acked) {
5909 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
5910 /* SYN-data is counted as two separate packets in tcp_ack() */
5911 if (tp->delivered > 1)
5915 tcp_fastopen_add_skb(sk, synack);
5920 static void smc_check_reset_syn(struct tcp_sock *tp)
5922 #if IS_ENABLED(CONFIG_SMC)
5923 if (static_branch_unlikely(&tcp_have_smc)) {
5924 if (tp->syn_smc && !tp->rx_opt.smc_ok)
5930 static void tcp_try_undo_spurious_syn(struct sock *sk)
5932 struct tcp_sock *tp = tcp_sk(sk);
5935 /* undo_marker is set when SYN or SYNACK times out. The timeout is
5936 * spurious if the ACK's timestamp option echo value matches the
5937 * original SYN timestamp.
5939 syn_stamp = tp->retrans_stamp;
5940 if (tp->undo_marker && syn_stamp && tp->rx_opt.saw_tstamp &&
5941 syn_stamp == tp->rx_opt.rcv_tsecr)
5942 tp->undo_marker = 0;
5945 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5946 const struct tcphdr *th)
5948 struct inet_connection_sock *icsk = inet_csk(sk);
5949 struct tcp_sock *tp = tcp_sk(sk);
5950 struct tcp_fastopen_cookie foc = { .len = -1 };
5951 int saved_clamp = tp->rx_opt.mss_clamp;
5954 tcp_parse_options(sock_net(sk), skb, &tp->rx_opt, 0, &foc);
5955 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
5956 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
5960 * "If the state is SYN-SENT then
5961 * first check the ACK bit
5962 * If the ACK bit is set
5963 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5964 * a reset (unless the RST bit is set, if so drop
5965 * the segment and return)"
5967 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5968 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5969 /* Previous FIN/ACK or RST/ACK might be ignored. */
5970 if (icsk->icsk_retransmits == 0)
5971 inet_csk_reset_xmit_timer(sk,
5973 TCP_TIMEOUT_MIN, TCP_RTO_MAX);
5974 goto reset_and_undo;
5977 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5978 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5979 tcp_time_stamp(tp))) {
5980 NET_INC_STATS(sock_net(sk),
5981 LINUX_MIB_PAWSACTIVEREJECTED);
5982 goto reset_and_undo;
5985 /* Now ACK is acceptable.
5987 * "If the RST bit is set
5988 * If the ACK was acceptable then signal the user "error:
5989 * connection reset", drop the segment, enter CLOSED state,
5990 * delete TCB, and return."
5999 * "fifth, if neither of the SYN or RST bits is set then
6000 * drop the segment and return."
6006 goto discard_and_undo;
6009 * "If the SYN bit is on ...
6010 * are acceptable then ...
6011 * (our SYN has been ACKed), change the connection
6012 * state to ESTABLISHED..."
6015 tcp_ecn_rcv_synack(tp, th);
6017 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6018 tcp_try_undo_spurious_syn(sk);
6019 tcp_ack(sk, skb, FLAG_SLOWPATH);
6021 /* Ok.. it's good. Set up sequence numbers and
6022 * move to established.
6024 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1);
6025 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
6027 /* RFC1323: The window in SYN & SYN/ACK segments is
6030 tp->snd_wnd = ntohs(th->window);
6032 if (!tp->rx_opt.wscale_ok) {
6033 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
6034 tp->window_clamp = min(tp->window_clamp, 65535U);
6037 if (tp->rx_opt.saw_tstamp) {
6038 tp->rx_opt.tstamp_ok = 1;
6039 tp->tcp_header_len =
6040 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
6041 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6042 tcp_store_ts_recent(tp);
6044 tp->tcp_header_len = sizeof(struct tcphdr);
6047 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
6048 tcp_initialize_rcv_mss(sk);
6050 /* Remember, tcp_poll() does not lock socket!
6051 * Change state from SYN-SENT only after copied_seq
6052 * is initialized. */
6053 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6055 smc_check_reset_syn(tp);
6059 tcp_finish_connect(sk, skb);
6061 fastopen_fail = (tp->syn_fastopen || tp->syn_data) &&
6062 tcp_rcv_fastopen_synack(sk, skb, &foc);
6064 if (!sock_flag(sk, SOCK_DEAD)) {
6065 sk->sk_state_change(sk);
6066 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6070 if (sk->sk_write_pending ||
6071 icsk->icsk_accept_queue.rskq_defer_accept ||
6072 inet_csk_in_pingpong_mode(sk)) {
6073 /* Save one ACK. Data will be ready after
6074 * several ticks, if write_pending is set.
6076 * It may be deleted, but with this feature tcpdumps
6077 * look so _wonderfully_ clever, that I was not able
6078 * to stand against the temptation 8) --ANK
6080 inet_csk_schedule_ack(sk);
6081 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
6082 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
6083 TCP_DELACK_MAX, TCP_RTO_MAX);
6094 /* No ACK in the segment */
6098 * "If the RST bit is set
6100 * Otherwise (no ACK) drop the segment and return."
6103 goto discard_and_undo;
6107 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
6108 tcp_paws_reject(&tp->rx_opt, 0))
6109 goto discard_and_undo;
6112 /* We see SYN without ACK. It is attempt of
6113 * simultaneous connect with crossed SYNs.
6114 * Particularly, it can be connect to self.
6116 tcp_set_state(sk, TCP_SYN_RECV);
6118 if (tp->rx_opt.saw_tstamp) {
6119 tp->rx_opt.tstamp_ok = 1;
6120 tcp_store_ts_recent(tp);
6121 tp->tcp_header_len =
6122 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
6124 tp->tcp_header_len = sizeof(struct tcphdr);
6127 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1);
6128 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6129 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
6131 /* RFC1323: The window in SYN & SYN/ACK segments is
6134 tp->snd_wnd = ntohs(th->window);
6135 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
6136 tp->max_window = tp->snd_wnd;
6138 tcp_ecn_rcv_syn(tp, th);
6141 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
6142 tcp_initialize_rcv_mss(sk);
6144 tcp_send_synack(sk);
6146 /* Note, we could accept data and URG from this segment.
6147 * There are no obstacles to make this (except that we must
6148 * either change tcp_recvmsg() to prevent it from returning data
6149 * before 3WHS completes per RFC793, or employ TCP Fast Open).
6151 * However, if we ignore data in ACKless segments sometimes,
6152 * we have no reasons to accept it sometimes.
6153 * Also, seems the code doing it in step6 of tcp_rcv_state_process
6154 * is not flawless. So, discard packet for sanity.
6155 * Uncomment this return to process the data.
6162 /* "fifth, if neither of the SYN or RST bits is set then
6163 * drop the segment and return."
6167 tcp_clear_options(&tp->rx_opt);
6168 tp->rx_opt.mss_clamp = saved_clamp;
6172 tcp_clear_options(&tp->rx_opt);
6173 tp->rx_opt.mss_clamp = saved_clamp;
6177 static void tcp_rcv_synrecv_state_fastopen(struct sock *sk)
6179 struct request_sock *req;
6181 /* If we are still handling the SYNACK RTO, see if timestamp ECR allows
6182 * undo. If peer SACKs triggered fast recovery, we can't undo here.
6184 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
6185 tcp_try_undo_loss(sk, false);
6187 /* Reset rtx states to prevent spurious retransmits_timed_out() */
6188 tcp_sk(sk)->retrans_stamp = 0;
6189 inet_csk(sk)->icsk_retransmits = 0;
6191 /* Once we leave TCP_SYN_RECV or TCP_FIN_WAIT_1,
6192 * we no longer need req so release it.
6194 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk,
6195 lockdep_sock_is_held(sk));
6196 reqsk_fastopen_remove(sk, req, false);
6198 /* Re-arm the timer because data may have been sent out.
6199 * This is similar to the regular data transmission case
6200 * when new data has just been ack'ed.
6202 * (TFO) - we could try to be more aggressive and
6203 * retransmitting any data sooner based on when they
6210 * This function implements the receiving procedure of RFC 793 for
6211 * all states except ESTABLISHED and TIME_WAIT.
6212 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
6213 * address independent.
6216 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
6218 struct tcp_sock *tp = tcp_sk(sk);
6219 struct inet_connection_sock *icsk = inet_csk(sk);
6220 const struct tcphdr *th = tcp_hdr(skb);
6221 struct request_sock *req;
6225 switch (sk->sk_state) {
6239 /* It is possible that we process SYN packets from backlog,
6240 * so we need to make sure to disable BH and RCU right there.
6244 acceptable = icsk->icsk_af_ops->conn_request(sk, skb) >= 0;
6256 tp->rx_opt.saw_tstamp = 0;
6257 tcp_mstamp_refresh(tp);
6258 queued = tcp_rcv_synsent_state_process(sk, skb, th);
6262 /* Do step6 onward by hand. */
6263 tcp_urg(sk, skb, th);
6265 tcp_data_snd_check(sk);
6269 tcp_mstamp_refresh(tp);
6270 tp->rx_opt.saw_tstamp = 0;
6271 req = rcu_dereference_protected(tp->fastopen_rsk,
6272 lockdep_sock_is_held(sk));
6276 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
6277 sk->sk_state != TCP_FIN_WAIT1);
6279 if (!tcp_check_req(sk, skb, req, true, &req_stolen))
6283 if (!th->ack && !th->rst && !th->syn)
6286 if (!tcp_validate_incoming(sk, skb, th, 0))
6289 /* step 5: check the ACK field */
6290 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
6291 FLAG_UPDATE_TS_RECENT |
6292 FLAG_NO_CHALLENGE_ACK) > 0;
6295 if (sk->sk_state == TCP_SYN_RECV)
6296 return 1; /* send one RST */
6297 tcp_send_challenge_ack(sk, skb);
6300 switch (sk->sk_state) {
6302 tp->delivered++; /* SYN-ACK delivery isn't tracked in tcp_ack */
6304 tcp_synack_rtt_meas(sk, req);
6307 tcp_rcv_synrecv_state_fastopen(sk);
6309 tcp_try_undo_spurious_syn(sk);
6310 tp->retrans_stamp = 0;
6311 tcp_init_transfer(sk, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB);
6312 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6315 tcp_set_state(sk, TCP_ESTABLISHED);
6316 sk->sk_state_change(sk);
6318 /* Note, that this wakeup is only for marginal crossed SYN case.
6319 * Passively open sockets are not waked up, because
6320 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
6323 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6325 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
6326 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
6327 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6329 if (tp->rx_opt.tstamp_ok)
6330 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6332 if (!inet_csk(sk)->icsk_ca_ops->cong_control)
6333 tcp_update_pacing_rate(sk);
6335 /* Prevent spurious tcp_cwnd_restart() on first data packet */
6336 tp->lsndtime = tcp_jiffies32;
6338 tcp_initialize_rcv_mss(sk);
6339 tcp_fast_path_on(tp);
6342 case TCP_FIN_WAIT1: {
6346 tcp_rcv_synrecv_state_fastopen(sk);
6348 if (tp->snd_una != tp->write_seq)
6351 tcp_set_state(sk, TCP_FIN_WAIT2);
6352 sk->sk_shutdown |= SEND_SHUTDOWN;
6356 if (!sock_flag(sk, SOCK_DEAD)) {
6357 /* Wake up lingering close() */
6358 sk->sk_state_change(sk);
6362 if (tp->linger2 < 0) {
6364 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6367 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6368 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6369 /* Receive out of order FIN after close() */
6370 if (tp->syn_fastopen && th->fin)
6371 tcp_fastopen_active_disable(sk);
6373 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6377 tmo = tcp_fin_time(sk);
6378 if (tmo > TCP_TIMEWAIT_LEN) {
6379 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
6380 } else if (th->fin || sock_owned_by_user(sk)) {
6381 /* Bad case. We could lose such FIN otherwise.
6382 * It is not a big problem, but it looks confusing
6383 * and not so rare event. We still can lose it now,
6384 * if it spins in bh_lock_sock(), but it is really
6387 inet_csk_reset_keepalive_timer(sk, tmo);
6389 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
6396 if (tp->snd_una == tp->write_seq) {
6397 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6403 if (tp->snd_una == tp->write_seq) {
6404 tcp_update_metrics(sk);
6411 /* step 6: check the URG bit */
6412 tcp_urg(sk, skb, th);
6414 /* step 7: process the segment text */
6415 switch (sk->sk_state) {
6416 case TCP_CLOSE_WAIT:
6419 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
6420 if (sk_is_mptcp(sk))
6421 mptcp_incoming_options(sk, skb, &tp->rx_opt);
6427 /* RFC 793 says to queue data in these states,
6428 * RFC 1122 says we MUST send a reset.
6429 * BSD 4.4 also does reset.
6431 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6432 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6433 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6434 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6440 case TCP_ESTABLISHED:
6441 tcp_data_queue(sk, skb);
6446 /* tcp_data could move socket to TIME-WAIT */
6447 if (sk->sk_state != TCP_CLOSE) {
6448 tcp_data_snd_check(sk);
6449 tcp_ack_snd_check(sk);
6458 EXPORT_SYMBOL(tcp_rcv_state_process);
6460 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
6462 struct inet_request_sock *ireq = inet_rsk(req);
6464 if (family == AF_INET)
6465 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6466 &ireq->ir_rmt_addr, port);
6467 #if IS_ENABLED(CONFIG_IPV6)
6468 else if (family == AF_INET6)
6469 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6470 &ireq->ir_v6_rmt_addr, port);
6474 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6476 * If we receive a SYN packet with these bits set, it means a
6477 * network is playing bad games with TOS bits. In order to
6478 * avoid possible false congestion notifications, we disable
6479 * TCP ECN negotiation.
6481 * Exception: tcp_ca wants ECN. This is required for DCTCP
6482 * congestion control: Linux DCTCP asserts ECT on all packets,
6483 * including SYN, which is most optimal solution; however,
6484 * others, such as FreeBSD do not.
6486 * Exception: At least one of the reserved bits of the TCP header (th->res1) is
6487 * set, indicating the use of a future TCP extension (such as AccECN). See
6488 * RFC8311 §4.3 which updates RFC3168 to allow the development of such
6491 static void tcp_ecn_create_request(struct request_sock *req,
6492 const struct sk_buff *skb,
6493 const struct sock *listen_sk,
6494 const struct dst_entry *dst)
6496 const struct tcphdr *th = tcp_hdr(skb);
6497 const struct net *net = sock_net(listen_sk);
6498 bool th_ecn = th->ece && th->cwr;
6505 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
6506 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK);
6507 ecn_ok = net->ipv4.sysctl_tcp_ecn || ecn_ok_dst;
6509 if (((!ect || th->res1) && ecn_ok) || tcp_ca_needs_ecn(listen_sk) ||
6510 (ecn_ok_dst & DST_FEATURE_ECN_CA) ||
6511 tcp_bpf_ca_needs_ecn((struct sock *)req))
6512 inet_rsk(req)->ecn_ok = 1;
6515 static void tcp_openreq_init(struct request_sock *req,
6516 const struct tcp_options_received *rx_opt,
6517 struct sk_buff *skb, const struct sock *sk)
6519 struct inet_request_sock *ireq = inet_rsk(req);
6521 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */
6522 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6523 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
6524 tcp_rsk(req)->snt_synack = 0;
6525 tcp_rsk(req)->last_oow_ack_time = 0;
6526 req->mss = rx_opt->mss_clamp;
6527 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6528 ireq->tstamp_ok = rx_opt->tstamp_ok;
6529 ireq->sack_ok = rx_opt->sack_ok;
6530 ireq->snd_wscale = rx_opt->snd_wscale;
6531 ireq->wscale_ok = rx_opt->wscale_ok;
6534 ireq->ir_rmt_port = tcp_hdr(skb)->source;
6535 ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6536 ireq->ir_mark = inet_request_mark(sk, skb);
6537 #if IS_ENABLED(CONFIG_SMC)
6538 ireq->smc_ok = rx_opt->smc_ok;
6542 struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
6543 struct sock *sk_listener,
6544 bool attach_listener)
6546 struct request_sock *req = reqsk_alloc(ops, sk_listener,
6550 struct inet_request_sock *ireq = inet_rsk(req);
6552 ireq->ireq_opt = NULL;
6553 #if IS_ENABLED(CONFIG_IPV6)
6554 ireq->pktopts = NULL;
6556 atomic64_set(&ireq->ir_cookie, 0);
6557 ireq->ireq_state = TCP_NEW_SYN_RECV;
6558 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
6559 ireq->ireq_family = sk_listener->sk_family;
6564 EXPORT_SYMBOL(inet_reqsk_alloc);
6567 * Return true if a syncookie should be sent
6569 static bool tcp_syn_flood_action(const struct sock *sk, const char *proto)
6571 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
6572 const char *msg = "Dropping request";
6573 bool want_cookie = false;
6574 struct net *net = sock_net(sk);
6576 #ifdef CONFIG_SYN_COOKIES
6577 if (net->ipv4.sysctl_tcp_syncookies) {
6578 msg = "Sending cookies";
6580 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
6583 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
6585 if (!queue->synflood_warned &&
6586 net->ipv4.sysctl_tcp_syncookies != 2 &&
6587 xchg(&queue->synflood_warned, 1) == 0)
6588 net_info_ratelimited("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6589 proto, sk->sk_num, msg);
6594 static void tcp_reqsk_record_syn(const struct sock *sk,
6595 struct request_sock *req,
6596 const struct sk_buff *skb)
6598 if (tcp_sk(sk)->save_syn) {
6599 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
6602 copy = kmalloc(len + sizeof(u32), GFP_ATOMIC);
6605 memcpy(©[1], skb_network_header(skb), len);
6606 req->saved_syn = copy;
6611 /* If a SYN cookie is required and supported, returns a clamped MSS value to be
6612 * used for SYN cookie generation.
6614 u16 tcp_get_syncookie_mss(struct request_sock_ops *rsk_ops,
6615 const struct tcp_request_sock_ops *af_ops,
6616 struct sock *sk, struct tcphdr *th)
6618 struct tcp_sock *tp = tcp_sk(sk);
6621 if (sock_net(sk)->ipv4.sysctl_tcp_syncookies != 2 &&
6622 !inet_csk_reqsk_queue_is_full(sk))
6625 if (!tcp_syn_flood_action(sk, rsk_ops->slab_name))
6628 if (sk_acceptq_is_full(sk)) {
6629 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6633 mss = tcp_parse_mss_option(th, tp->rx_opt.user_mss);
6635 mss = af_ops->mss_clamp;
6639 EXPORT_SYMBOL_GPL(tcp_get_syncookie_mss);
6641 int tcp_conn_request(struct request_sock_ops *rsk_ops,
6642 const struct tcp_request_sock_ops *af_ops,
6643 struct sock *sk, struct sk_buff *skb)
6645 struct tcp_fastopen_cookie foc = { .len = -1 };
6646 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
6647 struct tcp_options_received tmp_opt;
6648 struct tcp_sock *tp = tcp_sk(sk);
6649 struct net *net = sock_net(sk);
6650 struct sock *fastopen_sk = NULL;
6651 struct request_sock *req;
6652 bool want_cookie = false;
6653 struct dst_entry *dst;
6656 /* TW buckets are converted to open requests without
6657 * limitations, they conserve resources and peer is
6658 * evidently real one.
6660 if ((net->ipv4.sysctl_tcp_syncookies == 2 ||
6661 inet_csk_reqsk_queue_is_full(sk)) && !isn) {
6662 want_cookie = tcp_syn_flood_action(sk, rsk_ops->slab_name);
6667 if (sk_acceptq_is_full(sk)) {
6668 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6672 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie);
6676 req->syncookie = want_cookie;
6677 tcp_rsk(req)->af_specific = af_ops;
6678 tcp_rsk(req)->ts_off = 0;
6679 #if IS_ENABLED(CONFIG_MPTCP)
6680 tcp_rsk(req)->is_mptcp = 0;
6683 tcp_clear_options(&tmp_opt);
6684 tmp_opt.mss_clamp = af_ops->mss_clamp;
6685 tmp_opt.user_mss = tp->rx_opt.user_mss;
6686 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0,
6687 want_cookie ? NULL : &foc);
6689 if (want_cookie && !tmp_opt.saw_tstamp)
6690 tcp_clear_options(&tmp_opt);
6692 if (IS_ENABLED(CONFIG_SMC) && want_cookie)
6695 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
6696 tcp_openreq_init(req, &tmp_opt, skb, sk);
6697 inet_rsk(req)->no_srccheck = inet_sk(sk)->transparent;
6699 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6700 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb);
6702 af_ops->init_req(req, sk, skb);
6704 if (IS_ENABLED(CONFIG_MPTCP) && want_cookie)
6705 tcp_rsk(req)->is_mptcp = 0;
6707 if (security_inet_conn_request(sk, skb, req))
6710 if (tmp_opt.tstamp_ok)
6711 tcp_rsk(req)->ts_off = af_ops->init_ts_off(net, skb);
6713 dst = af_ops->route_req(sk, &fl, req);
6717 if (!want_cookie && !isn) {
6718 /* Kill the following clause, if you dislike this way. */
6719 if (!net->ipv4.sysctl_tcp_syncookies &&
6720 (net->ipv4.sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
6721 (net->ipv4.sysctl_max_syn_backlog >> 2)) &&
6722 !tcp_peer_is_proven(req, dst)) {
6723 /* Without syncookies last quarter of
6724 * backlog is filled with destinations,
6725 * proven to be alive.
6726 * It means that we continue to communicate
6727 * to destinations, already remembered
6728 * to the moment of synflood.
6730 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
6732 goto drop_and_release;
6735 isn = af_ops->init_seq(skb);
6738 tcp_ecn_create_request(req, skb, sk, dst);
6741 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
6742 if (!tmp_opt.tstamp_ok)
6743 inet_rsk(req)->ecn_ok = 0;
6746 tcp_rsk(req)->snt_isn = isn;
6747 tcp_rsk(req)->txhash = net_tx_rndhash();
6748 tcp_openreq_init_rwin(req, sk, dst);
6749 sk_rx_queue_set(req_to_sk(req), skb);
6751 tcp_reqsk_record_syn(sk, req, skb);
6752 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst);
6755 af_ops->send_synack(fastopen_sk, dst, &fl, req,
6756 &foc, TCP_SYNACK_FASTOPEN);
6757 /* Add the child socket directly into the accept queue */
6758 if (!inet_csk_reqsk_queue_add(sk, req, fastopen_sk)) {
6759 reqsk_fastopen_remove(fastopen_sk, req, false);
6760 bh_unlock_sock(fastopen_sk);
6761 sock_put(fastopen_sk);
6764 sk->sk_data_ready(sk);
6765 bh_unlock_sock(fastopen_sk);
6766 sock_put(fastopen_sk);
6768 tcp_rsk(req)->tfo_listener = false;
6770 inet_csk_reqsk_queue_hash_add(sk, req,
6771 tcp_timeout_init((struct sock *)req));
6772 af_ops->send_synack(sk, dst, &fl, req, &foc,
6773 !want_cookie ? TCP_SYNACK_NORMAL :
6791 EXPORT_SYMBOL(tcp_conn_request);