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. Skip update of data delivery
889 * and spurious retransmission information if this DSACK is unlikely caused by
891 * - DSACKed sequence range is larger than maximum receiver's window.
892 * - Total no. of DSACKed segments exceed the total no. of retransmitted segs.
894 static u32 tcp_dsack_seen(struct tcp_sock *tp, u32 start_seq,
895 u32 end_seq, struct tcp_sacktag_state *state)
897 u32 seq_len, dup_segs = 1;
899 if (!before(start_seq, end_seq))
902 seq_len = end_seq - start_seq;
903 /* Dubious DSACK: DSACKed range greater than maximum advertised rwnd */
904 if (seq_len > tp->max_window)
906 if (seq_len > tp->mss_cache)
907 dup_segs = DIV_ROUND_UP(seq_len, tp->mss_cache);
909 tp->dsack_dups += dup_segs;
910 /* Skip the DSACK if dup segs weren't retransmitted by sender */
911 if (tp->dsack_dups > tp->total_retrans)
914 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
915 tp->rack.dsack_seen = 1;
917 state->flag |= FLAG_DSACKING_ACK;
918 /* A spurious retransmission is delivered */
919 state->sack_delivered += dup_segs;
924 /* It's reordering when higher sequence was delivered (i.e. sacked) before
925 * some lower never-retransmitted sequence ("low_seq"). The maximum reordering
926 * distance is approximated in full-mss packet distance ("reordering").
928 static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq,
931 struct tcp_sock *tp = tcp_sk(sk);
932 const u32 mss = tp->mss_cache;
935 fack = tcp_highest_sack_seq(tp);
936 if (!before(low_seq, fack))
939 metric = fack - low_seq;
940 if ((metric > tp->reordering * mss) && mss) {
941 #if FASTRETRANS_DEBUG > 1
942 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
943 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
947 tp->undo_marker ? tp->undo_retrans : 0);
949 tp->reordering = min_t(u32, (metric + mss - 1) / mss,
950 sock_net(sk)->ipv4.sysctl_tcp_max_reordering);
953 /* This exciting event is worth to be remembered. 8) */
955 NET_INC_STATS(sock_net(sk),
956 ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER);
959 /* This must be called before lost_out is incremented */
960 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
962 if ((!tp->retransmit_skb_hint && tp->retrans_out >= tp->lost_out) ||
963 (tp->retransmit_skb_hint &&
964 before(TCP_SKB_CB(skb)->seq,
965 TCP_SKB_CB(tp->retransmit_skb_hint)->seq)))
966 tp->retransmit_skb_hint = skb;
969 /* Sum the number of packets on the wire we have marked as lost.
970 * There are two cases we care about here:
971 * a) Packet hasn't been marked lost (nor retransmitted),
972 * and this is the first loss.
973 * b) Packet has been marked both lost and retransmitted,
974 * and this means we think it was lost again.
976 static void tcp_sum_lost(struct tcp_sock *tp, struct sk_buff *skb)
978 __u8 sacked = TCP_SKB_CB(skb)->sacked;
980 if (!(sacked & TCPCB_LOST) ||
981 ((sacked & TCPCB_LOST) && (sacked & TCPCB_SACKED_RETRANS)))
982 tp->lost += tcp_skb_pcount(skb);
985 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
987 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
988 tcp_verify_retransmit_hint(tp, skb);
990 tp->lost_out += tcp_skb_pcount(skb);
991 tcp_sum_lost(tp, skb);
992 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
996 void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, struct sk_buff *skb)
998 tcp_verify_retransmit_hint(tp, skb);
1000 tcp_sum_lost(tp, skb);
1001 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1002 tp->lost_out += tcp_skb_pcount(skb);
1003 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1007 /* Updates the delivered and delivered_ce counts */
1008 static void tcp_count_delivered(struct tcp_sock *tp, u32 delivered,
1011 tp->delivered += delivered;
1013 tp->delivered_ce += delivered;
1016 /* This procedure tags the retransmission queue when SACKs arrive.
1018 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1019 * Packets in queue with these bits set are counted in variables
1020 * sacked_out, retrans_out and lost_out, correspondingly.
1022 * Valid combinations are:
1023 * Tag InFlight Description
1024 * 0 1 - orig segment is in flight.
1025 * S 0 - nothing flies, orig reached receiver.
1026 * L 0 - nothing flies, orig lost by net.
1027 * R 2 - both orig and retransmit are in flight.
1028 * L|R 1 - orig is lost, retransmit is in flight.
1029 * S|R 1 - orig reached receiver, retrans is still in flight.
1030 * (L|S|R is logically valid, it could occur when L|R is sacked,
1031 * but it is equivalent to plain S and code short-curcuits it to S.
1032 * L|S is logically invalid, it would mean -1 packet in flight 8))
1034 * These 6 states form finite state machine, controlled by the following events:
1035 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1036 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1037 * 3. Loss detection event of two flavors:
1038 * A. Scoreboard estimator decided the packet is lost.
1039 * A'. Reno "three dupacks" marks head of queue lost.
1040 * B. SACK arrives sacking SND.NXT at the moment, when the
1041 * segment was retransmitted.
1042 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1044 * It is pleasant to note, that state diagram turns out to be commutative,
1045 * so that we are allowed not to be bothered by order of our actions,
1046 * when multiple events arrive simultaneously. (see the function below).
1048 * Reordering detection.
1049 * --------------------
1050 * Reordering metric is maximal distance, which a packet can be displaced
1051 * in packet stream. With SACKs we can estimate it:
1053 * 1. SACK fills old hole and the corresponding segment was not
1054 * ever retransmitted -> reordering. Alas, we cannot use it
1055 * when segment was retransmitted.
1056 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1057 * for retransmitted and already SACKed segment -> reordering..
1058 * Both of these heuristics are not used in Loss state, when we cannot
1059 * account for retransmits accurately.
1061 * SACK block validation.
1062 * ----------------------
1064 * SACK block range validation checks that the received SACK block fits to
1065 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1066 * Note that SND.UNA is not included to the range though being valid because
1067 * it means that the receiver is rather inconsistent with itself reporting
1068 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1069 * perfectly valid, however, in light of RFC2018 which explicitly states
1070 * that "SACK block MUST reflect the newest segment. Even if the newest
1071 * segment is going to be discarded ...", not that it looks very clever
1072 * in case of head skb. Due to potentional receiver driven attacks, we
1073 * choose to avoid immediate execution of a walk in write queue due to
1074 * reneging and defer head skb's loss recovery to standard loss recovery
1075 * procedure that will eventually trigger (nothing forbids us doing this).
1077 * Implements also blockage to start_seq wrap-around. Problem lies in the
1078 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1079 * there's no guarantee that it will be before snd_nxt (n). The problem
1080 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1083 * <- outs wnd -> <- wrapzone ->
1084 * u e n u_w e_w s n_w
1086 * |<------------+------+----- TCP seqno space --------------+---------->|
1087 * ...-- <2^31 ->| |<--------...
1088 * ...---- >2^31 ------>| |<--------...
1090 * Current code wouldn't be vulnerable but it's better still to discard such
1091 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1092 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1093 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1094 * equal to the ideal case (infinite seqno space without wrap caused issues).
1096 * With D-SACK the lower bound is extended to cover sequence space below
1097 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1098 * again, D-SACK block must not to go across snd_una (for the same reason as
1099 * for the normal SACK blocks, explained above). But there all simplicity
1100 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1101 * fully below undo_marker they do not affect behavior in anyway and can
1102 * therefore be safely ignored. In rare cases (which are more or less
1103 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1104 * fragmentation and packet reordering past skb's retransmission. To consider
1105 * them correctly, the acceptable range must be extended even more though
1106 * the exact amount is rather hard to quantify. However, tp->max_window can
1107 * be used as an exaggerated estimate.
1109 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1110 u32 start_seq, u32 end_seq)
1112 /* Too far in future, or reversed (interpretation is ambiguous) */
1113 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1116 /* Nasty start_seq wrap-around check (see comments above) */
1117 if (!before(start_seq, tp->snd_nxt))
1120 /* In outstanding window? ...This is valid exit for D-SACKs too.
1121 * start_seq == snd_una is non-sensical (see comments above)
1123 if (after(start_seq, tp->snd_una))
1126 if (!is_dsack || !tp->undo_marker)
1129 /* ...Then it's D-SACK, and must reside below snd_una completely */
1130 if (after(end_seq, tp->snd_una))
1133 if (!before(start_seq, tp->undo_marker))
1137 if (!after(end_seq, tp->undo_marker))
1140 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1141 * start_seq < undo_marker and end_seq >= undo_marker.
1143 return !before(start_seq, end_seq - tp->max_window);
1146 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1147 struct tcp_sack_block_wire *sp, int num_sacks,
1148 u32 prior_snd_una, struct tcp_sacktag_state *state)
1150 struct tcp_sock *tp = tcp_sk(sk);
1151 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1152 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1155 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1156 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1157 } else if (num_sacks > 1) {
1158 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1159 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1161 if (after(end_seq_0, end_seq_1) || before(start_seq_0, start_seq_1))
1163 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKOFORECV);
1168 dup_segs = tcp_dsack_seen(tp, start_seq_0, end_seq_0, state);
1169 if (!dup_segs) { /* Skip dubious DSACK */
1170 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKIGNOREDDUBIOUS);
1174 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECVSEGS, dup_segs);
1176 /* D-SACK for already forgotten data... Do dumb counting. */
1177 if (tp->undo_marker && tp->undo_retrans > 0 &&
1178 !after(end_seq_0, prior_snd_una) &&
1179 after(end_seq_0, tp->undo_marker))
1180 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - dup_segs);
1185 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1186 * the incoming SACK may not exactly match but we can find smaller MSS
1187 * aligned portion of it that matches. Therefore we might need to fragment
1188 * which may fail and creates some hassle (caller must handle error case
1191 * FIXME: this could be merged to shift decision code
1193 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1194 u32 start_seq, u32 end_seq)
1198 unsigned int pkt_len;
1201 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1202 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1204 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1205 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1206 mss = tcp_skb_mss(skb);
1207 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1210 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1214 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1219 /* Round if necessary so that SACKs cover only full MSSes
1220 * and/or the remaining small portion (if present)
1222 if (pkt_len > mss) {
1223 unsigned int new_len = (pkt_len / mss) * mss;
1224 if (!in_sack && new_len < pkt_len)
1229 if (pkt_len >= skb->len && !in_sack)
1232 err = tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
1233 pkt_len, mss, GFP_ATOMIC);
1241 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1242 static u8 tcp_sacktag_one(struct sock *sk,
1243 struct tcp_sacktag_state *state, u8 sacked,
1244 u32 start_seq, u32 end_seq,
1245 int dup_sack, int pcount,
1248 struct tcp_sock *tp = tcp_sk(sk);
1250 /* Account D-SACK for retransmitted packet. */
1251 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1252 if (tp->undo_marker && tp->undo_retrans > 0 &&
1253 after(end_seq, tp->undo_marker))
1255 if ((sacked & TCPCB_SACKED_ACKED) &&
1256 before(start_seq, state->reord))
1257 state->reord = start_seq;
1260 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1261 if (!after(end_seq, tp->snd_una))
1264 if (!(sacked & TCPCB_SACKED_ACKED)) {
1265 tcp_rack_advance(tp, sacked, end_seq, xmit_time);
1267 if (sacked & TCPCB_SACKED_RETRANS) {
1268 /* If the segment is not tagged as lost,
1269 * we do not clear RETRANS, believing
1270 * that retransmission is still in flight.
1272 if (sacked & TCPCB_LOST) {
1273 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1274 tp->lost_out -= pcount;
1275 tp->retrans_out -= pcount;
1278 if (!(sacked & TCPCB_RETRANS)) {
1279 /* New sack for not retransmitted frame,
1280 * which was in hole. It is reordering.
1282 if (before(start_seq,
1283 tcp_highest_sack_seq(tp)) &&
1284 before(start_seq, state->reord))
1285 state->reord = start_seq;
1287 if (!after(end_seq, tp->high_seq))
1288 state->flag |= FLAG_ORIG_SACK_ACKED;
1289 if (state->first_sackt == 0)
1290 state->first_sackt = xmit_time;
1291 state->last_sackt = xmit_time;
1294 if (sacked & TCPCB_LOST) {
1295 sacked &= ~TCPCB_LOST;
1296 tp->lost_out -= pcount;
1300 sacked |= TCPCB_SACKED_ACKED;
1301 state->flag |= FLAG_DATA_SACKED;
1302 tp->sacked_out += pcount;
1303 /* Out-of-order packets delivered */
1304 state->sack_delivered += pcount;
1306 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1307 if (tp->lost_skb_hint &&
1308 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1309 tp->lost_cnt_hint += pcount;
1312 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1313 * frames and clear it. undo_retrans is decreased above, L|R frames
1314 * are accounted above as well.
1316 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1317 sacked &= ~TCPCB_SACKED_RETRANS;
1318 tp->retrans_out -= pcount;
1324 /* Shift newly-SACKed bytes from this skb to the immediately previous
1325 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1327 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev,
1328 struct sk_buff *skb,
1329 struct tcp_sacktag_state *state,
1330 unsigned int pcount, int shifted, int mss,
1333 struct tcp_sock *tp = tcp_sk(sk);
1334 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1335 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1339 /* Adjust counters and hints for the newly sacked sequence
1340 * range but discard the return value since prev is already
1341 * marked. We must tag the range first because the seq
1342 * advancement below implicitly advances
1343 * tcp_highest_sack_seq() when skb is highest_sack.
1345 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1346 start_seq, end_seq, dup_sack, pcount,
1347 tcp_skb_timestamp_us(skb));
1348 tcp_rate_skb_delivered(sk, skb, state->rate);
1350 if (skb == tp->lost_skb_hint)
1351 tp->lost_cnt_hint += pcount;
1353 TCP_SKB_CB(prev)->end_seq += shifted;
1354 TCP_SKB_CB(skb)->seq += shifted;
1356 tcp_skb_pcount_add(prev, pcount);
1357 WARN_ON_ONCE(tcp_skb_pcount(skb) < pcount);
1358 tcp_skb_pcount_add(skb, -pcount);
1360 /* When we're adding to gso_segs == 1, gso_size will be zero,
1361 * in theory this shouldn't be necessary but as long as DSACK
1362 * code can come after this skb later on it's better to keep
1363 * setting gso_size to something.
1365 if (!TCP_SKB_CB(prev)->tcp_gso_size)
1366 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1368 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1369 if (tcp_skb_pcount(skb) <= 1)
1370 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1372 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1373 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1376 BUG_ON(!tcp_skb_pcount(skb));
1377 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1381 /* Whole SKB was eaten :-) */
1383 if (skb == tp->retransmit_skb_hint)
1384 tp->retransmit_skb_hint = prev;
1385 if (skb == tp->lost_skb_hint) {
1386 tp->lost_skb_hint = prev;
1387 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1390 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1391 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor;
1392 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1393 TCP_SKB_CB(prev)->end_seq++;
1395 if (skb == tcp_highest_sack(sk))
1396 tcp_advance_highest_sack(sk, skb);
1398 tcp_skb_collapse_tstamp(prev, skb);
1399 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp))
1400 TCP_SKB_CB(prev)->tx.delivered_mstamp = 0;
1402 tcp_rtx_queue_unlink_and_free(skb, sk);
1404 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED);
1409 /* I wish gso_size would have a bit more sane initialization than
1410 * something-or-zero which complicates things
1412 static int tcp_skb_seglen(const struct sk_buff *skb)
1414 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1417 /* Shifting pages past head area doesn't work */
1418 static int skb_can_shift(const struct sk_buff *skb)
1420 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1423 int tcp_skb_shift(struct sk_buff *to, struct sk_buff *from,
1424 int pcount, int shiftlen)
1426 /* TCP min gso_size is 8 bytes (TCP_MIN_GSO_SIZE)
1427 * Since TCP_SKB_CB(skb)->tcp_gso_segs is 16 bits, we need
1428 * to make sure not storing more than 65535 * 8 bytes per skb,
1429 * even if current MSS is bigger.
1431 if (unlikely(to->len + shiftlen >= 65535 * TCP_MIN_GSO_SIZE))
1433 if (unlikely(tcp_skb_pcount(to) + pcount > 65535))
1435 return skb_shift(to, from, shiftlen);
1438 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1441 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1442 struct tcp_sacktag_state *state,
1443 u32 start_seq, u32 end_seq,
1446 struct tcp_sock *tp = tcp_sk(sk);
1447 struct sk_buff *prev;
1453 /* Normally R but no L won't result in plain S */
1455 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1457 if (!skb_can_shift(skb))
1459 /* This frame is about to be dropped (was ACKed). */
1460 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1463 /* Can only happen with delayed DSACK + discard craziness */
1464 prev = skb_rb_prev(skb);
1468 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1471 if (!tcp_skb_can_collapse(prev, skb))
1474 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1475 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1479 pcount = tcp_skb_pcount(skb);
1480 mss = tcp_skb_seglen(skb);
1482 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1483 * drop this restriction as unnecessary
1485 if (mss != tcp_skb_seglen(prev))
1488 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1490 /* CHECKME: This is non-MSS split case only?, this will
1491 * cause skipped skbs due to advancing loop btw, original
1492 * has that feature too
1494 if (tcp_skb_pcount(skb) <= 1)
1497 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1499 /* TODO: head merge to next could be attempted here
1500 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1501 * though it might not be worth of the additional hassle
1503 * ...we can probably just fallback to what was done
1504 * previously. We could try merging non-SACKed ones
1505 * as well but it probably isn't going to buy off
1506 * because later SACKs might again split them, and
1507 * it would make skb timestamp tracking considerably
1513 len = end_seq - TCP_SKB_CB(skb)->seq;
1515 BUG_ON(len > skb->len);
1517 /* MSS boundaries should be honoured or else pcount will
1518 * severely break even though it makes things bit trickier.
1519 * Optimize common case to avoid most of the divides
1521 mss = tcp_skb_mss(skb);
1523 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1524 * drop this restriction as unnecessary
1526 if (mss != tcp_skb_seglen(prev))
1531 } else if (len < mss) {
1539 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1540 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1543 if (!tcp_skb_shift(prev, skb, pcount, len))
1545 if (!tcp_shifted_skb(sk, prev, skb, state, pcount, len, mss, dup_sack))
1548 /* Hole filled allows collapsing with the next as well, this is very
1549 * useful when hole on every nth skb pattern happens
1551 skb = skb_rb_next(prev);
1555 if (!skb_can_shift(skb) ||
1556 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1557 (mss != tcp_skb_seglen(skb)))
1561 pcount = tcp_skb_pcount(skb);
1562 if (tcp_skb_shift(prev, skb, pcount, len))
1563 tcp_shifted_skb(sk, prev, skb, state, pcount,
1573 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1577 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1578 struct tcp_sack_block *next_dup,
1579 struct tcp_sacktag_state *state,
1580 u32 start_seq, u32 end_seq,
1583 struct tcp_sock *tp = tcp_sk(sk);
1584 struct sk_buff *tmp;
1586 skb_rbtree_walk_from(skb) {
1588 bool dup_sack = dup_sack_in;
1590 /* queue is in-order => we can short-circuit the walk early */
1591 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1595 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1596 in_sack = tcp_match_skb_to_sack(sk, skb,
1597 next_dup->start_seq,
1603 /* skb reference here is a bit tricky to get right, since
1604 * shifting can eat and free both this skb and the next,
1605 * so not even _safe variant of the loop is enough.
1608 tmp = tcp_shift_skb_data(sk, skb, state,
1609 start_seq, end_seq, dup_sack);
1618 in_sack = tcp_match_skb_to_sack(sk, skb,
1624 if (unlikely(in_sack < 0))
1628 TCP_SKB_CB(skb)->sacked =
1631 TCP_SKB_CB(skb)->sacked,
1632 TCP_SKB_CB(skb)->seq,
1633 TCP_SKB_CB(skb)->end_seq,
1635 tcp_skb_pcount(skb),
1636 tcp_skb_timestamp_us(skb));
1637 tcp_rate_skb_delivered(sk, skb, state->rate);
1638 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1639 list_del_init(&skb->tcp_tsorted_anchor);
1641 if (!before(TCP_SKB_CB(skb)->seq,
1642 tcp_highest_sack_seq(tp)))
1643 tcp_advance_highest_sack(sk, skb);
1649 static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk, u32 seq)
1651 struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node;
1652 struct sk_buff *skb;
1656 skb = rb_to_skb(parent);
1657 if (before(seq, TCP_SKB_CB(skb)->seq)) {
1658 p = &parent->rb_left;
1661 if (!before(seq, TCP_SKB_CB(skb)->end_seq)) {
1662 p = &parent->rb_right;
1670 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1673 if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq))
1676 return tcp_sacktag_bsearch(sk, skip_to_seq);
1679 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1681 struct tcp_sack_block *next_dup,
1682 struct tcp_sacktag_state *state,
1688 if (before(next_dup->start_seq, skip_to_seq)) {
1689 skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq);
1690 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1691 next_dup->start_seq, next_dup->end_seq,
1698 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1700 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1704 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1705 u32 prior_snd_una, struct tcp_sacktag_state *state)
1707 struct tcp_sock *tp = tcp_sk(sk);
1708 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1709 TCP_SKB_CB(ack_skb)->sacked);
1710 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1711 struct tcp_sack_block sp[TCP_NUM_SACKS];
1712 struct tcp_sack_block *cache;
1713 struct sk_buff *skb;
1714 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1716 bool found_dup_sack = false;
1718 int first_sack_index;
1721 state->reord = tp->snd_nxt;
1723 if (!tp->sacked_out)
1724 tcp_highest_sack_reset(sk);
1726 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1727 num_sacks, prior_snd_una, state);
1729 /* Eliminate too old ACKs, but take into
1730 * account more or less fresh ones, they can
1731 * contain valid SACK info.
1733 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1736 if (!tp->packets_out)
1740 first_sack_index = 0;
1741 for (i = 0; i < num_sacks; i++) {
1742 bool dup_sack = !i && found_dup_sack;
1744 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1745 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1747 if (!tcp_is_sackblock_valid(tp, dup_sack,
1748 sp[used_sacks].start_seq,
1749 sp[used_sacks].end_seq)) {
1753 if (!tp->undo_marker)
1754 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1756 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1758 /* Don't count olds caused by ACK reordering */
1759 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1760 !after(sp[used_sacks].end_seq, tp->snd_una))
1762 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1765 NET_INC_STATS(sock_net(sk), mib_idx);
1767 first_sack_index = -1;
1771 /* Ignore very old stuff early */
1772 if (!after(sp[used_sacks].end_seq, prior_snd_una)) {
1774 first_sack_index = -1;
1781 /* order SACK blocks to allow in order walk of the retrans queue */
1782 for (i = used_sacks - 1; i > 0; i--) {
1783 for (j = 0; j < i; j++) {
1784 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1785 swap(sp[j], sp[j + 1]);
1787 /* Track where the first SACK block goes to */
1788 if (j == first_sack_index)
1789 first_sack_index = j + 1;
1794 state->mss_now = tcp_current_mss(sk);
1798 if (!tp->sacked_out) {
1799 /* It's already past, so skip checking against it */
1800 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1802 cache = tp->recv_sack_cache;
1803 /* Skip empty blocks in at head of the cache */
1804 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1809 while (i < used_sacks) {
1810 u32 start_seq = sp[i].start_seq;
1811 u32 end_seq = sp[i].end_seq;
1812 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1813 struct tcp_sack_block *next_dup = NULL;
1815 if (found_dup_sack && ((i + 1) == first_sack_index))
1816 next_dup = &sp[i + 1];
1818 /* Skip too early cached blocks */
1819 while (tcp_sack_cache_ok(tp, cache) &&
1820 !before(start_seq, cache->end_seq))
1823 /* Can skip some work by looking recv_sack_cache? */
1824 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1825 after(end_seq, cache->start_seq)) {
1828 if (before(start_seq, cache->start_seq)) {
1829 skb = tcp_sacktag_skip(skb, sk, start_seq);
1830 skb = tcp_sacktag_walk(skb, sk, next_dup,
1837 /* Rest of the block already fully processed? */
1838 if (!after(end_seq, cache->end_seq))
1841 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1845 /* ...tail remains todo... */
1846 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1847 /* ...but better entrypoint exists! */
1848 skb = tcp_highest_sack(sk);
1855 skb = tcp_sacktag_skip(skb, sk, cache->end_seq);
1856 /* Check overlap against next cached too (past this one already) */
1861 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1862 skb = tcp_highest_sack(sk);
1866 skb = tcp_sacktag_skip(skb, sk, start_seq);
1869 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
1870 start_seq, end_seq, dup_sack);
1876 /* Clear the head of the cache sack blocks so we can skip it next time */
1877 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1878 tp->recv_sack_cache[i].start_seq = 0;
1879 tp->recv_sack_cache[i].end_seq = 0;
1881 for (j = 0; j < used_sacks; j++)
1882 tp->recv_sack_cache[i++] = sp[j];
1884 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss || tp->undo_marker)
1885 tcp_check_sack_reordering(sk, state->reord, 0);
1887 tcp_verify_left_out(tp);
1890 #if FASTRETRANS_DEBUG > 0
1891 WARN_ON((int)tp->sacked_out < 0);
1892 WARN_ON((int)tp->lost_out < 0);
1893 WARN_ON((int)tp->retrans_out < 0);
1894 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1899 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1900 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1902 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1906 holes = max(tp->lost_out, 1U);
1907 holes = min(holes, tp->packets_out);
1909 if ((tp->sacked_out + holes) > tp->packets_out) {
1910 tp->sacked_out = tp->packets_out - holes;
1916 /* If we receive more dupacks than we expected counting segments
1917 * in assumption of absent reordering, interpret this as reordering.
1918 * The only another reason could be bug in receiver TCP.
1920 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1922 struct tcp_sock *tp = tcp_sk(sk);
1924 if (!tcp_limit_reno_sacked(tp))
1927 tp->reordering = min_t(u32, tp->packets_out + addend,
1928 sock_net(sk)->ipv4.sysctl_tcp_max_reordering);
1930 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER);
1933 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1935 static void tcp_add_reno_sack(struct sock *sk, int num_dupack, bool ece_ack)
1938 struct tcp_sock *tp = tcp_sk(sk);
1939 u32 prior_sacked = tp->sacked_out;
1942 tp->sacked_out += num_dupack;
1943 tcp_check_reno_reordering(sk, 0);
1944 delivered = tp->sacked_out - prior_sacked;
1946 tcp_count_delivered(tp, delivered, ece_ack);
1947 tcp_verify_left_out(tp);
1951 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1953 static void tcp_remove_reno_sacks(struct sock *sk, int acked, bool ece_ack)
1955 struct tcp_sock *tp = tcp_sk(sk);
1958 /* One ACK acked hole. The rest eat duplicate ACKs. */
1959 tcp_count_delivered(tp, max_t(int, acked - tp->sacked_out, 1),
1961 if (acked - 1 >= tp->sacked_out)
1964 tp->sacked_out -= acked - 1;
1966 tcp_check_reno_reordering(sk, acked);
1967 tcp_verify_left_out(tp);
1970 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1975 void tcp_clear_retrans(struct tcp_sock *tp)
1977 tp->retrans_out = 0;
1979 tp->undo_marker = 0;
1980 tp->undo_retrans = -1;
1984 static inline void tcp_init_undo(struct tcp_sock *tp)
1986 tp->undo_marker = tp->snd_una;
1987 /* Retransmission still in flight may cause DSACKs later. */
1988 tp->undo_retrans = tp->retrans_out ? : -1;
1991 static bool tcp_is_rack(const struct sock *sk)
1993 return sock_net(sk)->ipv4.sysctl_tcp_recovery & TCP_RACK_LOSS_DETECTION;
1996 /* If we detect SACK reneging, forget all SACK information
1997 * and reset tags completely, otherwise preserve SACKs. If receiver
1998 * dropped its ofo queue, we will know this due to reneging detection.
2000 static void tcp_timeout_mark_lost(struct sock *sk)
2002 struct tcp_sock *tp = tcp_sk(sk);
2003 struct sk_buff *skb, *head;
2004 bool is_reneg; /* is receiver reneging on SACKs? */
2006 head = tcp_rtx_queue_head(sk);
2007 is_reneg = head && (TCP_SKB_CB(head)->sacked & TCPCB_SACKED_ACKED);
2009 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
2011 /* Mark SACK reneging until we recover from this loss event. */
2012 tp->is_sack_reneg = 1;
2013 } else if (tcp_is_reno(tp)) {
2014 tcp_reset_reno_sack(tp);
2018 skb_rbtree_walk_from(skb) {
2020 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
2021 else if (tcp_is_rack(sk) && skb != head &&
2022 tcp_rack_skb_timeout(tp, skb, 0) > 0)
2023 continue; /* Don't mark recently sent ones lost yet */
2024 tcp_mark_skb_lost(sk, skb);
2026 tcp_verify_left_out(tp);
2027 tcp_clear_all_retrans_hints(tp);
2030 /* Enter Loss state. */
2031 void tcp_enter_loss(struct sock *sk)
2033 const struct inet_connection_sock *icsk = inet_csk(sk);
2034 struct tcp_sock *tp = tcp_sk(sk);
2035 struct net *net = sock_net(sk);
2036 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
2038 tcp_timeout_mark_lost(sk);
2040 /* Reduce ssthresh if it has not yet been made inside this window. */
2041 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
2042 !after(tp->high_seq, tp->snd_una) ||
2043 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2044 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2045 tp->prior_cwnd = tp->snd_cwnd;
2046 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2047 tcp_ca_event(sk, CA_EVENT_LOSS);
2050 tp->snd_cwnd = tcp_packets_in_flight(tp) + 1;
2051 tp->snd_cwnd_cnt = 0;
2052 tp->snd_cwnd_stamp = tcp_jiffies32;
2054 /* Timeout in disordered state after receiving substantial DUPACKs
2055 * suggests that the degree of reordering is over-estimated.
2057 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
2058 tp->sacked_out >= net->ipv4.sysctl_tcp_reordering)
2059 tp->reordering = min_t(unsigned int, tp->reordering,
2060 net->ipv4.sysctl_tcp_reordering);
2061 tcp_set_ca_state(sk, TCP_CA_Loss);
2062 tp->high_seq = tp->snd_nxt;
2063 tcp_ecn_queue_cwr(tp);
2065 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
2066 * loss recovery is underway except recurring timeout(s) on
2067 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
2069 tp->frto = net->ipv4.sysctl_tcp_frto &&
2070 (new_recovery || icsk->icsk_retransmits) &&
2071 !inet_csk(sk)->icsk_mtup.probe_size;
2074 /* If ACK arrived pointing to a remembered SACK, it means that our
2075 * remembered SACKs do not reflect real state of receiver i.e.
2076 * receiver _host_ is heavily congested (or buggy).
2078 * To avoid big spurious retransmission bursts due to transient SACK
2079 * scoreboard oddities that look like reneging, we give the receiver a
2080 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2081 * restore sanity to the SACK scoreboard. If the apparent reneging
2082 * persists until this RTO then we'll clear the SACK scoreboard.
2084 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
2086 if (flag & FLAG_SACK_RENEGING) {
2087 struct tcp_sock *tp = tcp_sk(sk);
2088 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2089 msecs_to_jiffies(10));
2091 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2092 delay, TCP_RTO_MAX);
2098 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2099 * counter when SACK is enabled (without SACK, sacked_out is used for
2102 * With reordering, holes may still be in flight, so RFC3517 recovery
2103 * uses pure sacked_out (total number of SACKed segments) even though
2104 * it violates the RFC that uses duplicate ACKs, often these are equal
2105 * but when e.g. out-of-window ACKs or packet duplication occurs,
2106 * they differ. Since neither occurs due to loss, TCP should really
2109 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2111 return tp->sacked_out + 1;
2114 /* Linux NewReno/SACK/ECN state machine.
2115 * --------------------------------------
2117 * "Open" Normal state, no dubious events, fast path.
2118 * "Disorder" In all the respects it is "Open",
2119 * but requires a bit more attention. It is entered when
2120 * we see some SACKs or dupacks. It is split of "Open"
2121 * mainly to move some processing from fast path to slow one.
2122 * "CWR" CWND was reduced due to some Congestion Notification event.
2123 * It can be ECN, ICMP source quench, local device congestion.
2124 * "Recovery" CWND was reduced, we are fast-retransmitting.
2125 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2127 * tcp_fastretrans_alert() is entered:
2128 * - each incoming ACK, if state is not "Open"
2129 * - when arrived ACK is unusual, namely:
2134 * Counting packets in flight is pretty simple.
2136 * in_flight = packets_out - left_out + retrans_out
2138 * packets_out is SND.NXT-SND.UNA counted in packets.
2140 * retrans_out is number of retransmitted segments.
2142 * left_out is number of segments left network, but not ACKed yet.
2144 * left_out = sacked_out + lost_out
2146 * sacked_out: Packets, which arrived to receiver out of order
2147 * and hence not ACKed. With SACKs this number is simply
2148 * amount of SACKed data. Even without SACKs
2149 * it is easy to give pretty reliable estimate of this number,
2150 * counting duplicate ACKs.
2152 * lost_out: Packets lost by network. TCP has no explicit
2153 * "loss notification" feedback from network (for now).
2154 * It means that this number can be only _guessed_.
2155 * Actually, it is the heuristics to predict lossage that
2156 * distinguishes different algorithms.
2158 * F.e. after RTO, when all the queue is considered as lost,
2159 * lost_out = packets_out and in_flight = retrans_out.
2161 * Essentially, we have now a few algorithms detecting
2164 * If the receiver supports SACK:
2166 * RFC6675/3517: It is the conventional algorithm. A packet is
2167 * considered lost if the number of higher sequence packets
2168 * SACKed is greater than or equal the DUPACK thoreshold
2169 * (reordering). This is implemented in tcp_mark_head_lost and
2170 * tcp_update_scoreboard.
2172 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm
2173 * (2017-) that checks timing instead of counting DUPACKs.
2174 * Essentially a packet is considered lost if it's not S/ACKed
2175 * after RTT + reordering_window, where both metrics are
2176 * dynamically measured and adjusted. This is implemented in
2177 * tcp_rack_mark_lost.
2179 * If the receiver does not support SACK:
2181 * NewReno (RFC6582): in Recovery we assume that one segment
2182 * is lost (classic Reno). While we are in Recovery and
2183 * a partial ACK arrives, we assume that one more packet
2184 * is lost (NewReno). This heuristics are the same in NewReno
2187 * Really tricky (and requiring careful tuning) part of algorithm
2188 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2189 * The first determines the moment _when_ we should reduce CWND and,
2190 * hence, slow down forward transmission. In fact, it determines the moment
2191 * when we decide that hole is caused by loss, rather than by a reorder.
2193 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2194 * holes, caused by lost packets.
2196 * And the most logically complicated part of algorithm is undo
2197 * heuristics. We detect false retransmits due to both too early
2198 * fast retransmit (reordering) and underestimated RTO, analyzing
2199 * timestamps and D-SACKs. When we detect that some segments were
2200 * retransmitted by mistake and CWND reduction was wrong, we undo
2201 * window reduction and abort recovery phase. This logic is hidden
2202 * inside several functions named tcp_try_undo_<something>.
2205 /* This function decides, when we should leave Disordered state
2206 * and enter Recovery phase, reducing congestion window.
2208 * Main question: may we further continue forward transmission
2209 * with the same cwnd?
2211 static bool tcp_time_to_recover(struct sock *sk, int flag)
2213 struct tcp_sock *tp = tcp_sk(sk);
2215 /* Trick#1: The loss is proven. */
2219 /* Not-A-Trick#2 : Classic rule... */
2220 if (!tcp_is_rack(sk) && tcp_dupack_heuristics(tp) > tp->reordering)
2226 /* Detect loss in event "A" above by marking head of queue up as lost.
2227 * For RFC3517 SACK, a segment is considered lost if it
2228 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2229 * the maximum SACKed segments to pass before reaching this limit.
2231 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2233 struct tcp_sock *tp = tcp_sk(sk);
2234 struct sk_buff *skb;
2236 /* Use SACK to deduce losses of new sequences sent during recovery */
2237 const u32 loss_high = tp->snd_nxt;
2239 WARN_ON(packets > tp->packets_out);
2240 skb = tp->lost_skb_hint;
2242 /* Head already handled? */
2243 if (mark_head && after(TCP_SKB_CB(skb)->seq, tp->snd_una))
2245 cnt = tp->lost_cnt_hint;
2247 skb = tcp_rtx_queue_head(sk);
2251 skb_rbtree_walk_from(skb) {
2252 /* TODO: do this better */
2253 /* this is not the most efficient way to do this... */
2254 tp->lost_skb_hint = skb;
2255 tp->lost_cnt_hint = cnt;
2257 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2260 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2261 cnt += tcp_skb_pcount(skb);
2266 tcp_skb_mark_lost(tp, skb);
2271 tcp_verify_left_out(tp);
2274 /* Account newly detected lost packet(s) */
2276 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2278 struct tcp_sock *tp = tcp_sk(sk);
2280 if (tcp_is_sack(tp)) {
2281 int sacked_upto = tp->sacked_out - tp->reordering;
2282 if (sacked_upto >= 0)
2283 tcp_mark_head_lost(sk, sacked_upto, 0);
2284 else if (fast_rexmit)
2285 tcp_mark_head_lost(sk, 1, 1);
2289 static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2291 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2292 before(tp->rx_opt.rcv_tsecr, when);
2295 /* skb is spurious retransmitted if the returned timestamp echo
2296 * reply is prior to the skb transmission time
2298 static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2299 const struct sk_buff *skb)
2301 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2302 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp(skb));
2305 /* Nothing was retransmitted or returned timestamp is less
2306 * than timestamp of the first retransmission.
2308 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2310 return tp->retrans_stamp &&
2311 tcp_tsopt_ecr_before(tp, tp->retrans_stamp);
2314 /* Undo procedures. */
2316 /* We can clear retrans_stamp when there are no retransmissions in the
2317 * window. It would seem that it is trivially available for us in
2318 * tp->retrans_out, however, that kind of assumptions doesn't consider
2319 * what will happen if errors occur when sending retransmission for the
2320 * second time. ...It could the that such segment has only
2321 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2322 * the head skb is enough except for some reneging corner cases that
2323 * are not worth the effort.
2325 * Main reason for all this complexity is the fact that connection dying
2326 * time now depends on the validity of the retrans_stamp, in particular,
2327 * that successive retransmissions of a segment must not advance
2328 * retrans_stamp under any conditions.
2330 static bool tcp_any_retrans_done(const struct sock *sk)
2332 const struct tcp_sock *tp = tcp_sk(sk);
2333 struct sk_buff *skb;
2335 if (tp->retrans_out)
2338 skb = tcp_rtx_queue_head(sk);
2339 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2345 static void DBGUNDO(struct sock *sk, const char *msg)
2347 #if FASTRETRANS_DEBUG > 1
2348 struct tcp_sock *tp = tcp_sk(sk);
2349 struct inet_sock *inet = inet_sk(sk);
2351 if (sk->sk_family == AF_INET) {
2352 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2354 &inet->inet_daddr, ntohs(inet->inet_dport),
2355 tp->snd_cwnd, tcp_left_out(tp),
2356 tp->snd_ssthresh, tp->prior_ssthresh,
2359 #if IS_ENABLED(CONFIG_IPV6)
2360 else if (sk->sk_family == AF_INET6) {
2361 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2363 &sk->sk_v6_daddr, ntohs(inet->inet_dport),
2364 tp->snd_cwnd, tcp_left_out(tp),
2365 tp->snd_ssthresh, tp->prior_ssthresh,
2372 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2374 struct tcp_sock *tp = tcp_sk(sk);
2377 struct sk_buff *skb;
2379 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2380 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2383 tcp_clear_all_retrans_hints(tp);
2386 if (tp->prior_ssthresh) {
2387 const struct inet_connection_sock *icsk = inet_csk(sk);
2389 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2391 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2392 tp->snd_ssthresh = tp->prior_ssthresh;
2393 tcp_ecn_withdraw_cwr(tp);
2396 tp->snd_cwnd_stamp = tcp_jiffies32;
2397 tp->undo_marker = 0;
2398 tp->rack.advanced = 1; /* Force RACK to re-exam losses */
2401 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2403 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2406 /* People celebrate: "We love our President!" */
2407 static bool tcp_try_undo_recovery(struct sock *sk)
2409 struct tcp_sock *tp = tcp_sk(sk);
2411 if (tcp_may_undo(tp)) {
2414 /* Happy end! We did not retransmit anything
2415 * or our original transmission succeeded.
2417 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2418 tcp_undo_cwnd_reduction(sk, false);
2419 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2420 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2422 mib_idx = LINUX_MIB_TCPFULLUNDO;
2424 NET_INC_STATS(sock_net(sk), mib_idx);
2425 } else if (tp->rack.reo_wnd_persist) {
2426 tp->rack.reo_wnd_persist--;
2428 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2429 /* Hold old state until something *above* high_seq
2430 * is ACKed. For Reno it is MUST to prevent false
2431 * fast retransmits (RFC2582). SACK TCP is safe. */
2432 if (!tcp_any_retrans_done(sk))
2433 tp->retrans_stamp = 0;
2436 tcp_set_ca_state(sk, TCP_CA_Open);
2437 tp->is_sack_reneg = 0;
2441 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2442 static bool tcp_try_undo_dsack(struct sock *sk)
2444 struct tcp_sock *tp = tcp_sk(sk);
2446 if (tp->undo_marker && !tp->undo_retrans) {
2447 tp->rack.reo_wnd_persist = min(TCP_RACK_RECOVERY_THRESH,
2448 tp->rack.reo_wnd_persist + 1);
2449 DBGUNDO(sk, "D-SACK");
2450 tcp_undo_cwnd_reduction(sk, false);
2451 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2457 /* Undo during loss recovery after partial ACK or using F-RTO. */
2458 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2460 struct tcp_sock *tp = tcp_sk(sk);
2462 if (frto_undo || tcp_may_undo(tp)) {
2463 tcp_undo_cwnd_reduction(sk, true);
2465 DBGUNDO(sk, "partial loss");
2466 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2468 NET_INC_STATS(sock_net(sk),
2469 LINUX_MIB_TCPSPURIOUSRTOS);
2470 inet_csk(sk)->icsk_retransmits = 0;
2471 if (frto_undo || tcp_is_sack(tp)) {
2472 tcp_set_ca_state(sk, TCP_CA_Open);
2473 tp->is_sack_reneg = 0;
2480 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2481 * It computes the number of packets to send (sndcnt) based on packets newly
2483 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2484 * cwnd reductions across a full RTT.
2485 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2486 * But when the retransmits are acked without further losses, PRR
2487 * slow starts cwnd up to ssthresh to speed up the recovery.
2489 static void tcp_init_cwnd_reduction(struct sock *sk)
2491 struct tcp_sock *tp = tcp_sk(sk);
2493 tp->high_seq = tp->snd_nxt;
2494 tp->tlp_high_seq = 0;
2495 tp->snd_cwnd_cnt = 0;
2496 tp->prior_cwnd = tp->snd_cwnd;
2497 tp->prr_delivered = 0;
2499 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2500 tcp_ecn_queue_cwr(tp);
2503 void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int flag)
2505 struct tcp_sock *tp = tcp_sk(sk);
2507 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2509 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2512 tp->prr_delivered += newly_acked_sacked;
2514 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2516 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2517 } else if ((flag & (FLAG_RETRANS_DATA_ACKED | FLAG_LOST_RETRANS)) ==
2518 FLAG_RETRANS_DATA_ACKED) {
2519 sndcnt = min_t(int, delta,
2520 max_t(int, tp->prr_delivered - tp->prr_out,
2521 newly_acked_sacked) + 1);
2523 sndcnt = min(delta, newly_acked_sacked);
2525 /* Force a fast retransmit upon entering fast recovery */
2526 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
2527 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2530 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2532 struct tcp_sock *tp = tcp_sk(sk);
2534 if (inet_csk(sk)->icsk_ca_ops->cong_control)
2537 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2538 if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH &&
2539 (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) {
2540 tp->snd_cwnd = tp->snd_ssthresh;
2541 tp->snd_cwnd_stamp = tcp_jiffies32;
2543 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2546 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2547 void tcp_enter_cwr(struct sock *sk)
2549 struct tcp_sock *tp = tcp_sk(sk);
2551 tp->prior_ssthresh = 0;
2552 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2553 tp->undo_marker = 0;
2554 tcp_init_cwnd_reduction(sk);
2555 tcp_set_ca_state(sk, TCP_CA_CWR);
2558 EXPORT_SYMBOL(tcp_enter_cwr);
2560 static void tcp_try_keep_open(struct sock *sk)
2562 struct tcp_sock *tp = tcp_sk(sk);
2563 int state = TCP_CA_Open;
2565 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2566 state = TCP_CA_Disorder;
2568 if (inet_csk(sk)->icsk_ca_state != state) {
2569 tcp_set_ca_state(sk, state);
2570 tp->high_seq = tp->snd_nxt;
2574 static void tcp_try_to_open(struct sock *sk, int flag)
2576 struct tcp_sock *tp = tcp_sk(sk);
2578 tcp_verify_left_out(tp);
2580 if (!tcp_any_retrans_done(sk))
2581 tp->retrans_stamp = 0;
2583 if (flag & FLAG_ECE)
2586 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2587 tcp_try_keep_open(sk);
2591 static void tcp_mtup_probe_failed(struct sock *sk)
2593 struct inet_connection_sock *icsk = inet_csk(sk);
2595 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2596 icsk->icsk_mtup.probe_size = 0;
2597 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
2600 static void tcp_mtup_probe_success(struct sock *sk)
2602 struct tcp_sock *tp = tcp_sk(sk);
2603 struct inet_connection_sock *icsk = inet_csk(sk);
2605 /* FIXME: breaks with very large cwnd */
2606 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2607 tp->snd_cwnd = tp->snd_cwnd *
2608 tcp_mss_to_mtu(sk, tp->mss_cache) /
2609 icsk->icsk_mtup.probe_size;
2610 tp->snd_cwnd_cnt = 0;
2611 tp->snd_cwnd_stamp = tcp_jiffies32;
2612 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2614 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2615 icsk->icsk_mtup.probe_size = 0;
2616 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2617 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
2620 /* Do a simple retransmit without using the backoff mechanisms in
2621 * tcp_timer. This is used for path mtu discovery.
2622 * The socket is already locked here.
2624 void tcp_simple_retransmit(struct sock *sk)
2626 const struct inet_connection_sock *icsk = inet_csk(sk);
2627 struct tcp_sock *tp = tcp_sk(sk);
2628 struct sk_buff *skb;
2629 unsigned int mss = tcp_current_mss(sk);
2631 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2632 if (tcp_skb_seglen(skb) > mss &&
2633 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2634 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2635 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2636 tp->retrans_out -= tcp_skb_pcount(skb);
2638 tcp_skb_mark_lost_uncond_verify(tp, skb);
2642 tcp_clear_retrans_hints_partial(tp);
2647 if (tcp_is_reno(tp))
2648 tcp_limit_reno_sacked(tp);
2650 tcp_verify_left_out(tp);
2652 /* Don't muck with the congestion window here.
2653 * Reason is that we do not increase amount of _data_
2654 * in network, but units changed and effective
2655 * cwnd/ssthresh really reduced now.
2657 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2658 tp->high_seq = tp->snd_nxt;
2659 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2660 tp->prior_ssthresh = 0;
2661 tp->undo_marker = 0;
2662 tcp_set_ca_state(sk, TCP_CA_Loss);
2664 tcp_xmit_retransmit_queue(sk);
2666 EXPORT_SYMBOL(tcp_simple_retransmit);
2668 void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2670 struct tcp_sock *tp = tcp_sk(sk);
2673 if (tcp_is_reno(tp))
2674 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2676 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2678 NET_INC_STATS(sock_net(sk), mib_idx);
2680 tp->prior_ssthresh = 0;
2683 if (!tcp_in_cwnd_reduction(sk)) {
2685 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2686 tcp_init_cwnd_reduction(sk);
2688 tcp_set_ca_state(sk, TCP_CA_Recovery);
2691 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2692 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2694 static void tcp_process_loss(struct sock *sk, int flag, int num_dupack,
2697 struct tcp_sock *tp = tcp_sk(sk);
2698 bool recovered = !before(tp->snd_una, tp->high_seq);
2700 if ((flag & FLAG_SND_UNA_ADVANCED || rcu_access_pointer(tp->fastopen_rsk)) &&
2701 tcp_try_undo_loss(sk, false))
2704 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2705 /* Step 3.b. A timeout is spurious if not all data are
2706 * lost, i.e., never-retransmitted data are (s)acked.
2708 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2709 tcp_try_undo_loss(sk, true))
2712 if (after(tp->snd_nxt, tp->high_seq)) {
2713 if (flag & FLAG_DATA_SACKED || num_dupack)
2714 tp->frto = 0; /* Step 3.a. loss was real */
2715 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2716 tp->high_seq = tp->snd_nxt;
2717 /* Step 2.b. Try send new data (but deferred until cwnd
2718 * is updated in tcp_ack()). Otherwise fall back to
2719 * the conventional recovery.
2721 if (!tcp_write_queue_empty(sk) &&
2722 after(tcp_wnd_end(tp), tp->snd_nxt)) {
2723 *rexmit = REXMIT_NEW;
2731 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2732 tcp_try_undo_recovery(sk);
2735 if (tcp_is_reno(tp)) {
2736 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2737 * delivered. Lower inflight to clock out (re)tranmissions.
2739 if (after(tp->snd_nxt, tp->high_seq) && num_dupack)
2740 tcp_add_reno_sack(sk, num_dupack, flag & FLAG_ECE);
2741 else if (flag & FLAG_SND_UNA_ADVANCED)
2742 tcp_reset_reno_sack(tp);
2744 *rexmit = REXMIT_LOST;
2747 /* Undo during fast recovery after partial ACK. */
2748 static bool tcp_try_undo_partial(struct sock *sk, u32 prior_snd_una)
2750 struct tcp_sock *tp = tcp_sk(sk);
2752 if (tp->undo_marker && tcp_packet_delayed(tp)) {
2753 /* Plain luck! Hole if filled with delayed
2754 * packet, rather than with a retransmit. Check reordering.
2756 tcp_check_sack_reordering(sk, prior_snd_una, 1);
2758 /* We are getting evidence that the reordering degree is higher
2759 * than we realized. If there are no retransmits out then we
2760 * can undo. Otherwise we clock out new packets but do not
2761 * mark more packets lost or retransmit more.
2763 if (tp->retrans_out)
2766 if (!tcp_any_retrans_done(sk))
2767 tp->retrans_stamp = 0;
2769 DBGUNDO(sk, "partial recovery");
2770 tcp_undo_cwnd_reduction(sk, true);
2771 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2772 tcp_try_keep_open(sk);
2778 static void tcp_identify_packet_loss(struct sock *sk, int *ack_flag)
2780 struct tcp_sock *tp = tcp_sk(sk);
2782 if (tcp_rtx_queue_empty(sk))
2785 if (unlikely(tcp_is_reno(tp))) {
2786 tcp_newreno_mark_lost(sk, *ack_flag & FLAG_SND_UNA_ADVANCED);
2787 } else if (tcp_is_rack(sk)) {
2788 u32 prior_retrans = tp->retrans_out;
2790 tcp_rack_mark_lost(sk);
2791 if (prior_retrans > tp->retrans_out)
2792 *ack_flag |= FLAG_LOST_RETRANS;
2796 static bool tcp_force_fast_retransmit(struct sock *sk)
2798 struct tcp_sock *tp = tcp_sk(sk);
2800 return after(tcp_highest_sack_seq(tp),
2801 tp->snd_una + tp->reordering * tp->mss_cache);
2804 /* Process an event, which can update packets-in-flight not trivially.
2805 * Main goal of this function is to calculate new estimate for left_out,
2806 * taking into account both packets sitting in receiver's buffer and
2807 * packets lost by network.
2809 * Besides that it updates the congestion state when packet loss or ECN
2810 * is detected. But it does not reduce the cwnd, it is done by the
2811 * congestion control later.
2813 * It does _not_ decide what to send, it is made in function
2814 * tcp_xmit_retransmit_queue().
2816 static void tcp_fastretrans_alert(struct sock *sk, const u32 prior_snd_una,
2817 int num_dupack, int *ack_flag, int *rexmit)
2819 struct inet_connection_sock *icsk = inet_csk(sk);
2820 struct tcp_sock *tp = tcp_sk(sk);
2821 int fast_rexmit = 0, flag = *ack_flag;
2822 bool ece_ack = flag & FLAG_ECE;
2823 bool do_lost = num_dupack || ((flag & FLAG_DATA_SACKED) &&
2824 tcp_force_fast_retransmit(sk));
2826 if (!tp->packets_out && tp->sacked_out)
2829 /* Now state machine starts.
2830 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2832 tp->prior_ssthresh = 0;
2834 /* B. In all the states check for reneging SACKs. */
2835 if (tcp_check_sack_reneging(sk, flag))
2838 /* C. Check consistency of the current state. */
2839 tcp_verify_left_out(tp);
2841 /* D. Check state exit conditions. State can be terminated
2842 * when high_seq is ACKed. */
2843 if (icsk->icsk_ca_state == TCP_CA_Open) {
2844 WARN_ON(tp->retrans_out != 0);
2845 tp->retrans_stamp = 0;
2846 } else if (!before(tp->snd_una, tp->high_seq)) {
2847 switch (icsk->icsk_ca_state) {
2849 /* CWR is to be held something *above* high_seq
2850 * is ACKed for CWR bit to reach receiver. */
2851 if (tp->snd_una != tp->high_seq) {
2852 tcp_end_cwnd_reduction(sk);
2853 tcp_set_ca_state(sk, TCP_CA_Open);
2857 case TCP_CA_Recovery:
2858 if (tcp_is_reno(tp))
2859 tcp_reset_reno_sack(tp);
2860 if (tcp_try_undo_recovery(sk))
2862 tcp_end_cwnd_reduction(sk);
2867 /* E. Process state. */
2868 switch (icsk->icsk_ca_state) {
2869 case TCP_CA_Recovery:
2870 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2871 if (tcp_is_reno(tp))
2872 tcp_add_reno_sack(sk, num_dupack, ece_ack);
2874 if (tcp_try_undo_partial(sk, prior_snd_una))
2876 /* Partial ACK arrived. Force fast retransmit. */
2877 do_lost = tcp_force_fast_retransmit(sk);
2879 if (tcp_try_undo_dsack(sk)) {
2880 tcp_try_keep_open(sk);
2883 tcp_identify_packet_loss(sk, ack_flag);
2886 tcp_process_loss(sk, flag, num_dupack, rexmit);
2887 tcp_identify_packet_loss(sk, ack_flag);
2888 if (!(icsk->icsk_ca_state == TCP_CA_Open ||
2889 (*ack_flag & FLAG_LOST_RETRANS)))
2891 /* Change state if cwnd is undone or retransmits are lost */
2894 if (tcp_is_reno(tp)) {
2895 if (flag & FLAG_SND_UNA_ADVANCED)
2896 tcp_reset_reno_sack(tp);
2897 tcp_add_reno_sack(sk, num_dupack, ece_ack);
2900 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2901 tcp_try_undo_dsack(sk);
2903 tcp_identify_packet_loss(sk, ack_flag);
2904 if (!tcp_time_to_recover(sk, flag)) {
2905 tcp_try_to_open(sk, flag);
2909 /* MTU probe failure: don't reduce cwnd */
2910 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2911 icsk->icsk_mtup.probe_size &&
2912 tp->snd_una == tp->mtu_probe.probe_seq_start) {
2913 tcp_mtup_probe_failed(sk);
2914 /* Restores the reduction we did in tcp_mtup_probe() */
2916 tcp_simple_retransmit(sk);
2920 /* Otherwise enter Recovery state */
2921 tcp_enter_recovery(sk, ece_ack);
2925 if (!tcp_is_rack(sk) && do_lost)
2926 tcp_update_scoreboard(sk, fast_rexmit);
2927 *rexmit = REXMIT_LOST;
2930 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us, const int flag)
2932 u32 wlen = sock_net(sk)->ipv4.sysctl_tcp_min_rtt_wlen * HZ;
2933 struct tcp_sock *tp = tcp_sk(sk);
2935 if ((flag & FLAG_ACK_MAYBE_DELAYED) && rtt_us > tcp_min_rtt(tp)) {
2936 /* If the remote keeps returning delayed ACKs, eventually
2937 * the min filter would pick it up and overestimate the
2938 * prop. delay when it expires. Skip suspected delayed ACKs.
2942 minmax_running_min(&tp->rtt_min, wlen, tcp_jiffies32,
2943 rtt_us ? : jiffies_to_usecs(1));
2946 static bool tcp_ack_update_rtt(struct sock *sk, const int flag,
2947 long seq_rtt_us, long sack_rtt_us,
2948 long ca_rtt_us, struct rate_sample *rs)
2950 const struct tcp_sock *tp = tcp_sk(sk);
2952 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2953 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2954 * Karn's algorithm forbids taking RTT if some retransmitted data
2955 * is acked (RFC6298).
2958 seq_rtt_us = sack_rtt_us;
2960 /* RTTM Rule: A TSecr value received in a segment is used to
2961 * update the averaged RTT measurement only if the segment
2962 * acknowledges some new data, i.e., only if it advances the
2963 * left edge of the send window.
2964 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2966 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2967 flag & FLAG_ACKED) {
2968 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
2970 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) {
2973 seq_rtt_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
2974 ca_rtt_us = seq_rtt_us;
2977 rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */
2981 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
2982 * always taken together with ACK, SACK, or TS-opts. Any negative
2983 * values will be skipped with the seq_rtt_us < 0 check above.
2985 tcp_update_rtt_min(sk, ca_rtt_us, flag);
2986 tcp_rtt_estimator(sk, seq_rtt_us);
2989 /* RFC6298: only reset backoff on valid RTT measurement. */
2990 inet_csk(sk)->icsk_backoff = 0;
2994 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2995 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
2997 struct rate_sample rs;
3000 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack)
3001 rtt_us = tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req)->snt_synack);
3003 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us, &rs);
3007 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
3009 const struct inet_connection_sock *icsk = inet_csk(sk);
3011 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
3012 tcp_sk(sk)->snd_cwnd_stamp = tcp_jiffies32;
3015 /* Restart timer after forward progress on connection.
3016 * RFC2988 recommends to restart timer to now+rto.
3018 void tcp_rearm_rto(struct sock *sk)
3020 const struct inet_connection_sock *icsk = inet_csk(sk);
3021 struct tcp_sock *tp = tcp_sk(sk);
3023 /* If the retrans timer is currently being used by Fast Open
3024 * for SYN-ACK retrans purpose, stay put.
3026 if (rcu_access_pointer(tp->fastopen_rsk))
3029 if (!tp->packets_out) {
3030 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3032 u32 rto = inet_csk(sk)->icsk_rto;
3033 /* Offset the time elapsed after installing regular RTO */
3034 if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT ||
3035 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
3036 s64 delta_us = tcp_rto_delta_us(sk);
3037 /* delta_us may not be positive if the socket is locked
3038 * when the retrans timer fires and is rescheduled.
3040 rto = usecs_to_jiffies(max_t(int, delta_us, 1));
3042 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
3047 /* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */
3048 static void tcp_set_xmit_timer(struct sock *sk)
3050 if (!tcp_schedule_loss_probe(sk, true))
3054 /* If we get here, the whole TSO packet has not been acked. */
3055 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3057 struct tcp_sock *tp = tcp_sk(sk);
3060 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3062 packets_acked = tcp_skb_pcount(skb);
3063 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3065 packets_acked -= tcp_skb_pcount(skb);
3067 if (packets_acked) {
3068 BUG_ON(tcp_skb_pcount(skb) == 0);
3069 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3072 return packets_acked;
3075 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3078 const struct skb_shared_info *shinfo;
3080 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3081 if (likely(!TCP_SKB_CB(skb)->txstamp_ack))
3084 shinfo = skb_shinfo(skb);
3085 if (!before(shinfo->tskey, prior_snd_una) &&
3086 before(shinfo->tskey, tcp_sk(sk)->snd_una)) {
3087 tcp_skb_tsorted_save(skb) {
3088 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK);
3089 } tcp_skb_tsorted_restore(skb);
3093 /* Remove acknowledged frames from the retransmission queue. If our packet
3094 * is before the ack sequence we can discard it as it's confirmed to have
3095 * arrived at the other end.
3097 static int tcp_clean_rtx_queue(struct sock *sk, u32 prior_fack,
3099 struct tcp_sacktag_state *sack, bool ece_ack)
3101 const struct inet_connection_sock *icsk = inet_csk(sk);
3102 u64 first_ackt, last_ackt;
3103 struct tcp_sock *tp = tcp_sk(sk);
3104 u32 prior_sacked = tp->sacked_out;
3105 u32 reord = tp->snd_nxt; /* lowest acked un-retx un-sacked seq */
3106 struct sk_buff *skb, *next;
3107 bool fully_acked = true;
3108 long sack_rtt_us = -1L;
3109 long seq_rtt_us = -1L;
3110 long ca_rtt_us = -1L;
3112 u32 last_in_flight = 0;
3118 for (skb = skb_rb_first(&sk->tcp_rtx_queue); skb; skb = next) {
3119 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3120 const u32 start_seq = scb->seq;
3121 u8 sacked = scb->sacked;
3124 /* Determine how many packets and what bytes were acked, tso and else */
3125 if (after(scb->end_seq, tp->snd_una)) {
3126 if (tcp_skb_pcount(skb) == 1 ||
3127 !after(tp->snd_una, scb->seq))
3130 acked_pcount = tcp_tso_acked(sk, skb);
3133 fully_acked = false;
3135 acked_pcount = tcp_skb_pcount(skb);
3138 if (unlikely(sacked & TCPCB_RETRANS)) {
3139 if (sacked & TCPCB_SACKED_RETRANS)
3140 tp->retrans_out -= acked_pcount;
3141 flag |= FLAG_RETRANS_DATA_ACKED;
3142 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3143 last_ackt = tcp_skb_timestamp_us(skb);
3144 WARN_ON_ONCE(last_ackt == 0);
3146 first_ackt = last_ackt;
3148 last_in_flight = TCP_SKB_CB(skb)->tx.in_flight;
3149 if (before(start_seq, reord))
3151 if (!after(scb->end_seq, tp->high_seq))
3152 flag |= FLAG_ORIG_SACK_ACKED;
3155 if (sacked & TCPCB_SACKED_ACKED) {
3156 tp->sacked_out -= acked_pcount;
3157 } else if (tcp_is_sack(tp)) {
3158 tcp_count_delivered(tp, acked_pcount, ece_ack);
3159 if (!tcp_skb_spurious_retrans(tp, skb))
3160 tcp_rack_advance(tp, sacked, scb->end_seq,
3161 tcp_skb_timestamp_us(skb));
3163 if (sacked & TCPCB_LOST)
3164 tp->lost_out -= acked_pcount;
3166 tp->packets_out -= acked_pcount;
3167 pkts_acked += acked_pcount;
3168 tcp_rate_skb_delivered(sk, skb, sack->rate);
3170 /* Initial outgoing SYN's get put onto the write_queue
3171 * just like anything else we transmit. It is not
3172 * true data, and if we misinform our callers that
3173 * this ACK acks real data, we will erroneously exit
3174 * connection startup slow start one packet too
3175 * quickly. This is severely frowned upon behavior.
3177 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3178 flag |= FLAG_DATA_ACKED;
3180 flag |= FLAG_SYN_ACKED;
3181 tp->retrans_stamp = 0;
3187 tcp_ack_tstamp(sk, skb, prior_snd_una);
3189 next = skb_rb_next(skb);
3190 if (unlikely(skb == tp->retransmit_skb_hint))
3191 tp->retransmit_skb_hint = NULL;
3192 if (unlikely(skb == tp->lost_skb_hint))
3193 tp->lost_skb_hint = NULL;
3194 tcp_highest_sack_replace(sk, skb, next);
3195 tcp_rtx_queue_unlink_and_free(skb, sk);
3199 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3201 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3202 tp->snd_up = tp->snd_una;
3205 tcp_ack_tstamp(sk, skb, prior_snd_una);
3206 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
3207 flag |= FLAG_SACK_RENEGING;
3210 if (likely(first_ackt) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
3211 seq_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, first_ackt);
3212 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, last_ackt);
3214 if (pkts_acked == 1 && last_in_flight < tp->mss_cache &&
3215 last_in_flight && !prior_sacked && fully_acked &&
3216 sack->rate->prior_delivered + 1 == tp->delivered &&
3217 !(flag & (FLAG_CA_ALERT | FLAG_SYN_ACKED))) {
3218 /* Conservatively mark a delayed ACK. It's typically
3219 * from a lone runt packet over the round trip to
3220 * a receiver w/o out-of-order or CE events.
3222 flag |= FLAG_ACK_MAYBE_DELAYED;
3225 if (sack->first_sackt) {
3226 sack_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->first_sackt);
3227 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->last_sackt);
3229 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
3230 ca_rtt_us, sack->rate);
3232 if (flag & FLAG_ACKED) {
3233 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3234 if (unlikely(icsk->icsk_mtup.probe_size &&
3235 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3236 tcp_mtup_probe_success(sk);
3239 if (tcp_is_reno(tp)) {
3240 tcp_remove_reno_sacks(sk, pkts_acked, ece_ack);
3242 /* If any of the cumulatively ACKed segments was
3243 * retransmitted, non-SACK case cannot confirm that
3244 * progress was due to original transmission due to
3245 * lack of TCPCB_SACKED_ACKED bits even if some of
3246 * the packets may have been never retransmitted.
3248 if (flag & FLAG_RETRANS_DATA_ACKED)
3249 flag &= ~FLAG_ORIG_SACK_ACKED;
3253 /* Non-retransmitted hole got filled? That's reordering */
3254 if (before(reord, prior_fack))
3255 tcp_check_sack_reordering(sk, reord, 0);
3257 delta = prior_sacked - tp->sacked_out;
3258 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3260 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3261 sack_rtt_us > tcp_stamp_us_delta(tp->tcp_mstamp,
3262 tcp_skb_timestamp_us(skb))) {
3263 /* Do not re-arm RTO if the sack RTT is measured from data sent
3264 * after when the head was last (re)transmitted. Otherwise the
3265 * timeout may continue to extend in loss recovery.
3267 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3270 if (icsk->icsk_ca_ops->pkts_acked) {
3271 struct ack_sample sample = { .pkts_acked = pkts_acked,
3272 .rtt_us = sack->rate->rtt_us,
3273 .in_flight = last_in_flight };
3275 icsk->icsk_ca_ops->pkts_acked(sk, &sample);
3278 #if FASTRETRANS_DEBUG > 0
3279 WARN_ON((int)tp->sacked_out < 0);
3280 WARN_ON((int)tp->lost_out < 0);
3281 WARN_ON((int)tp->retrans_out < 0);
3282 if (!tp->packets_out && tcp_is_sack(tp)) {
3283 icsk = inet_csk(sk);
3285 pr_debug("Leak l=%u %d\n",
3286 tp->lost_out, icsk->icsk_ca_state);
3289 if (tp->sacked_out) {
3290 pr_debug("Leak s=%u %d\n",
3291 tp->sacked_out, icsk->icsk_ca_state);
3294 if (tp->retrans_out) {
3295 pr_debug("Leak r=%u %d\n",
3296 tp->retrans_out, icsk->icsk_ca_state);
3297 tp->retrans_out = 0;
3304 static void tcp_ack_probe(struct sock *sk)
3306 struct inet_connection_sock *icsk = inet_csk(sk);
3307 struct sk_buff *head = tcp_send_head(sk);
3308 const struct tcp_sock *tp = tcp_sk(sk);
3310 /* Was it a usable window open? */
3313 if (!after(TCP_SKB_CB(head)->end_seq, tcp_wnd_end(tp))) {
3314 icsk->icsk_backoff = 0;
3315 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3316 /* Socket must be waked up by subsequent tcp_data_snd_check().
3317 * This function is not for random using!
3320 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
3322 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3327 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3329 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3330 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3333 /* Decide wheather to run the increase function of congestion control. */
3334 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3336 /* If reordering is high then always grow cwnd whenever data is
3337 * delivered regardless of its ordering. Otherwise stay conservative
3338 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3339 * new SACK or ECE mark may first advance cwnd here and later reduce
3340 * cwnd in tcp_fastretrans_alert() based on more states.
3342 if (tcp_sk(sk)->reordering > sock_net(sk)->ipv4.sysctl_tcp_reordering)
3343 return flag & FLAG_FORWARD_PROGRESS;
3345 return flag & FLAG_DATA_ACKED;
3348 /* The "ultimate" congestion control function that aims to replace the rigid
3349 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3350 * It's called toward the end of processing an ACK with precise rate
3351 * information. All transmission or retransmission are delayed afterwards.
3353 static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
3354 int flag, const struct rate_sample *rs)
3356 const struct inet_connection_sock *icsk = inet_csk(sk);
3358 if (icsk->icsk_ca_ops->cong_control) {
3359 icsk->icsk_ca_ops->cong_control(sk, rs);
3363 if (tcp_in_cwnd_reduction(sk)) {
3364 /* Reduce cwnd if state mandates */
3365 tcp_cwnd_reduction(sk, acked_sacked, flag);
3366 } else if (tcp_may_raise_cwnd(sk, flag)) {
3367 /* Advance cwnd if state allows */
3368 tcp_cong_avoid(sk, ack, acked_sacked);
3370 tcp_update_pacing_rate(sk);
3373 /* Check that window update is acceptable.
3374 * The function assumes that snd_una<=ack<=snd_next.
3376 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3377 const u32 ack, const u32 ack_seq,
3380 return after(ack, tp->snd_una) ||
3381 after(ack_seq, tp->snd_wl1) ||
3382 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3385 /* If we update tp->snd_una, also update tp->bytes_acked */
3386 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3388 u32 delta = ack - tp->snd_una;
3390 sock_owned_by_me((struct sock *)tp);
3391 tp->bytes_acked += delta;
3395 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3396 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3398 u32 delta = seq - tp->rcv_nxt;
3400 sock_owned_by_me((struct sock *)tp);
3401 tp->bytes_received += delta;
3402 WRITE_ONCE(tp->rcv_nxt, seq);
3405 /* Update our send window.
3407 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3408 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3410 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3413 struct tcp_sock *tp = tcp_sk(sk);
3415 u32 nwin = ntohs(tcp_hdr(skb)->window);
3417 if (likely(!tcp_hdr(skb)->syn))
3418 nwin <<= tp->rx_opt.snd_wscale;
3420 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3421 flag |= FLAG_WIN_UPDATE;
3422 tcp_update_wl(tp, ack_seq);
3424 if (tp->snd_wnd != nwin) {
3427 /* Note, it is the only place, where
3428 * fast path is recovered for sending TCP.
3431 tcp_fast_path_check(sk);
3433 if (!tcp_write_queue_empty(sk))
3434 tcp_slow_start_after_idle_check(sk);
3436 if (nwin > tp->max_window) {
3437 tp->max_window = nwin;
3438 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3443 tcp_snd_una_update(tp, ack);
3448 static bool __tcp_oow_rate_limited(struct net *net, int mib_idx,
3449 u32 *last_oow_ack_time)
3451 if (*last_oow_ack_time) {
3452 s32 elapsed = (s32)(tcp_jiffies32 - *last_oow_ack_time);
3454 if (0 <= elapsed && elapsed < net->ipv4.sysctl_tcp_invalid_ratelimit) {
3455 NET_INC_STATS(net, mib_idx);
3456 return true; /* rate-limited: don't send yet! */
3460 *last_oow_ack_time = tcp_jiffies32;
3462 return false; /* not rate-limited: go ahead, send dupack now! */
3465 /* Return true if we're currently rate-limiting out-of-window ACKs and
3466 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3467 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3468 * attacks that send repeated SYNs or ACKs for the same connection. To
3469 * do this, we do not send a duplicate SYNACK or ACK if the remote
3470 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3472 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3473 int mib_idx, u32 *last_oow_ack_time)
3475 /* Data packets without SYNs are not likely part of an ACK loop. */
3476 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3480 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time);
3483 /* RFC 5961 7 [ACK Throttling] */
3484 static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb)
3486 /* unprotected vars, we dont care of overwrites */
3487 static u32 challenge_timestamp;
3488 static unsigned int challenge_count;
3489 struct tcp_sock *tp = tcp_sk(sk);
3490 struct net *net = sock_net(sk);
3493 /* First check our per-socket dupack rate limit. */
3494 if (__tcp_oow_rate_limited(net,
3495 LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3496 &tp->last_oow_ack_time))
3499 /* Then check host-wide RFC 5961 rate limit. */
3501 if (now != challenge_timestamp) {
3502 u32 ack_limit = net->ipv4.sysctl_tcp_challenge_ack_limit;
3503 u32 half = (ack_limit + 1) >> 1;
3505 challenge_timestamp = now;
3506 WRITE_ONCE(challenge_count, half + prandom_u32_max(ack_limit));
3508 count = READ_ONCE(challenge_count);
3510 WRITE_ONCE(challenge_count, count - 1);
3511 NET_INC_STATS(net, LINUX_MIB_TCPCHALLENGEACK);
3516 static void tcp_store_ts_recent(struct tcp_sock *tp)
3518 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3519 tp->rx_opt.ts_recent_stamp = ktime_get_seconds();
3522 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3524 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3525 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3526 * extra check below makes sure this can only happen
3527 * for pure ACK frames. -DaveM
3529 * Not only, also it occurs for expired timestamps.
3532 if (tcp_paws_check(&tp->rx_opt, 0))
3533 tcp_store_ts_recent(tp);
3537 /* This routine deals with acks during a TLP episode and ends an episode by
3538 * resetting tlp_high_seq. Ref: TLP algorithm in draft-ietf-tcpm-rack
3540 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3542 struct tcp_sock *tp = tcp_sk(sk);
3544 if (before(ack, tp->tlp_high_seq))
3547 if (!tp->tlp_retrans) {
3548 /* TLP of new data has been acknowledged */
3549 tp->tlp_high_seq = 0;
3550 } else if (flag & FLAG_DSACKING_ACK) {
3551 /* This DSACK means original and TLP probe arrived; no loss */
3552 tp->tlp_high_seq = 0;
3553 } else if (after(ack, tp->tlp_high_seq)) {
3554 /* ACK advances: there was a loss, so reduce cwnd. Reset
3555 * tlp_high_seq in tcp_init_cwnd_reduction()
3557 tcp_init_cwnd_reduction(sk);
3558 tcp_set_ca_state(sk, TCP_CA_CWR);
3559 tcp_end_cwnd_reduction(sk);
3560 tcp_try_keep_open(sk);
3561 NET_INC_STATS(sock_net(sk),
3562 LINUX_MIB_TCPLOSSPROBERECOVERY);
3563 } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3564 FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3565 /* Pure dupack: original and TLP probe arrived; no loss */
3566 tp->tlp_high_seq = 0;
3570 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3572 const struct inet_connection_sock *icsk = inet_csk(sk);
3574 if (icsk->icsk_ca_ops->in_ack_event)
3575 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3578 /* Congestion control has updated the cwnd already. So if we're in
3579 * loss recovery then now we do any new sends (for FRTO) or
3580 * retransmits (for CA_Loss or CA_recovery) that make sense.
3582 static void tcp_xmit_recovery(struct sock *sk, int rexmit)
3584 struct tcp_sock *tp = tcp_sk(sk);
3586 if (rexmit == REXMIT_NONE || sk->sk_state == TCP_SYN_SENT)
3589 if (unlikely(rexmit == REXMIT_NEW)) {
3590 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
3592 if (after(tp->snd_nxt, tp->high_seq))
3596 tcp_xmit_retransmit_queue(sk);
3599 /* Returns the number of packets newly acked or sacked by the current ACK */
3600 static u32 tcp_newly_delivered(struct sock *sk, u32 prior_delivered, int flag)
3602 const struct net *net = sock_net(sk);
3603 struct tcp_sock *tp = tcp_sk(sk);
3606 delivered = tp->delivered - prior_delivered;
3607 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVERED, delivered);
3608 if (flag & FLAG_ECE)
3609 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVEREDCE, delivered);
3614 /* This routine deals with incoming acks, but not outgoing ones. */
3615 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3617 struct inet_connection_sock *icsk = inet_csk(sk);
3618 struct tcp_sock *tp = tcp_sk(sk);
3619 struct tcp_sacktag_state sack_state;
3620 struct rate_sample rs = { .prior_delivered = 0 };
3621 u32 prior_snd_una = tp->snd_una;
3622 bool is_sack_reneg = tp->is_sack_reneg;
3623 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3624 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3626 int prior_packets = tp->packets_out;
3627 u32 delivered = tp->delivered;
3628 u32 lost = tp->lost;
3629 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
3632 sack_state.first_sackt = 0;
3633 sack_state.rate = &rs;
3634 sack_state.sack_delivered = 0;
3636 /* We very likely will need to access rtx queue. */
3637 prefetch(sk->tcp_rtx_queue.rb_node);
3639 /* If the ack is older than previous acks
3640 * then we can probably ignore it.
3642 if (before(ack, prior_snd_una)) {
3643 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3644 if (before(ack, prior_snd_una - tp->max_window)) {
3645 if (!(flag & FLAG_NO_CHALLENGE_ACK))
3646 tcp_send_challenge_ack(sk, skb);
3652 /* If the ack includes data we haven't sent yet, discard
3653 * this segment (RFC793 Section 3.9).
3655 if (after(ack, tp->snd_nxt))
3658 if (after(ack, prior_snd_una)) {
3659 flag |= FLAG_SND_UNA_ADVANCED;
3660 icsk->icsk_retransmits = 0;
3662 #if IS_ENABLED(CONFIG_TLS_DEVICE)
3663 if (static_branch_unlikely(&clean_acked_data_enabled.key))
3664 if (icsk->icsk_clean_acked)
3665 icsk->icsk_clean_acked(sk, ack);
3669 prior_fack = tcp_is_sack(tp) ? tcp_highest_sack_seq(tp) : tp->snd_una;
3670 rs.prior_in_flight = tcp_packets_in_flight(tp);
3672 /* ts_recent update must be made after we are sure that the packet
3675 if (flag & FLAG_UPDATE_TS_RECENT)
3676 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3678 if ((flag & (FLAG_SLOWPATH | FLAG_SND_UNA_ADVANCED)) ==
3679 FLAG_SND_UNA_ADVANCED) {
3680 /* Window is constant, pure forward advance.
3681 * No more checks are required.
3682 * Note, we use the fact that SND.UNA>=SND.WL2.
3684 tcp_update_wl(tp, ack_seq);
3685 tcp_snd_una_update(tp, ack);
3686 flag |= FLAG_WIN_UPDATE;
3688 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3690 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS);
3692 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3694 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3697 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3699 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3701 if (TCP_SKB_CB(skb)->sacked)
3702 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3705 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3707 ack_ev_flags |= CA_ACK_ECE;
3710 if (sack_state.sack_delivered)
3711 tcp_count_delivered(tp, sack_state.sack_delivered,
3714 if (flag & FLAG_WIN_UPDATE)
3715 ack_ev_flags |= CA_ACK_WIN_UPDATE;
3717 tcp_in_ack_event(sk, ack_ev_flags);
3720 /* This is a deviation from RFC3168 since it states that:
3721 * "When the TCP data sender is ready to set the CWR bit after reducing
3722 * the congestion window, it SHOULD set the CWR bit only on the first
3723 * new data packet that it transmits."
3724 * We accept CWR on pure ACKs to be more robust
3725 * with widely-deployed TCP implementations that do this.
3727 tcp_ecn_accept_cwr(sk, skb);
3729 /* We passed data and got it acked, remove any soft error
3730 * log. Something worked...
3732 sk->sk_err_soft = 0;
3733 icsk->icsk_probes_out = 0;
3734 tp->rcv_tstamp = tcp_jiffies32;
3738 /* See if we can take anything off of the retransmit queue. */
3739 flag |= tcp_clean_rtx_queue(sk, prior_fack, prior_snd_una, &sack_state,
3742 tcp_rack_update_reo_wnd(sk, &rs);
3744 if (tp->tlp_high_seq)
3745 tcp_process_tlp_ack(sk, ack, flag);
3746 /* If needed, reset TLP/RTO timer; RACK may later override this. */
3747 if (flag & FLAG_SET_XMIT_TIMER)
3748 tcp_set_xmit_timer(sk);
3750 if (tcp_ack_is_dubious(sk, flag)) {
3751 if (!(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP))) {
3753 /* Consider if pure acks were aggregated in tcp_add_backlog() */
3754 if (!(flag & FLAG_DATA))
3755 num_dupack = max_t(u16, 1, skb_shinfo(skb)->gso_segs);
3757 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3761 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3764 delivered = tcp_newly_delivered(sk, delivered, flag);
3765 lost = tp->lost - lost; /* freshly marked lost */
3766 rs.is_ack_delayed = !!(flag & FLAG_ACK_MAYBE_DELAYED);
3767 tcp_rate_gen(sk, delivered, lost, is_sack_reneg, sack_state.rate);
3768 tcp_cong_control(sk, ack, delivered, flag, sack_state.rate);
3769 tcp_xmit_recovery(sk, rexmit);
3773 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3774 if (flag & FLAG_DSACKING_ACK) {
3775 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3777 tcp_newly_delivered(sk, delivered, flag);
3779 /* If this ack opens up a zero window, clear backoff. It was
3780 * being used to time the probes, and is probably far higher than
3781 * it needs to be for normal retransmission.
3785 if (tp->tlp_high_seq)
3786 tcp_process_tlp_ack(sk, ack, flag);
3790 /* If data was SACKed, tag it and see if we should send more data.
3791 * If data was DSACKed, see if we can undo a cwnd reduction.
3793 if (TCP_SKB_CB(skb)->sacked) {
3794 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3796 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3798 tcp_newly_delivered(sk, delivered, flag);
3799 tcp_xmit_recovery(sk, rexmit);
3805 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3806 bool syn, struct tcp_fastopen_cookie *foc,
3809 /* Valid only in SYN or SYN-ACK with an even length. */
3810 if (!foc || !syn || len < 0 || (len & 1))
3813 if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3814 len <= TCP_FASTOPEN_COOKIE_MAX)
3815 memcpy(foc->val, cookie, len);
3822 static void smc_parse_options(const struct tcphdr *th,
3823 struct tcp_options_received *opt_rx,
3824 const unsigned char *ptr,
3827 #if IS_ENABLED(CONFIG_SMC)
3828 if (static_branch_unlikely(&tcp_have_smc)) {
3829 if (th->syn && !(opsize & 1) &&
3830 opsize >= TCPOLEN_EXP_SMC_BASE &&
3831 get_unaligned_be32(ptr) == TCPOPT_SMC_MAGIC)
3837 /* Try to parse the MSS option from the TCP header. Return 0 on failure, clamped
3840 static u16 tcp_parse_mss_option(const struct tcphdr *th, u16 user_mss)
3842 const unsigned char *ptr = (const unsigned char *)(th + 1);
3843 int length = (th->doff * 4) - sizeof(struct tcphdr);
3846 while (length > 0) {
3847 int opcode = *ptr++;
3853 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3860 if (opsize < 2) /* "silly options" */
3862 if (opsize > length)
3863 return mss; /* fail on partial options */
3864 if (opcode == TCPOPT_MSS && opsize == TCPOLEN_MSS) {
3865 u16 in_mss = get_unaligned_be16(ptr);
3868 if (user_mss && user_mss < in_mss)
3880 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3881 * But, this can also be called on packets in the established flow when
3882 * the fast version below fails.
3884 void tcp_parse_options(const struct net *net,
3885 const struct sk_buff *skb,
3886 struct tcp_options_received *opt_rx, int estab,
3887 struct tcp_fastopen_cookie *foc)
3889 const unsigned char *ptr;
3890 const struct tcphdr *th = tcp_hdr(skb);
3891 int length = (th->doff * 4) - sizeof(struct tcphdr);
3893 ptr = (const unsigned char *)(th + 1);
3894 opt_rx->saw_tstamp = 0;
3896 while (length > 0) {
3897 int opcode = *ptr++;
3903 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3910 if (opsize < 2) /* "silly options" */
3912 if (opsize > length)
3913 return; /* don't parse partial options */
3916 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3917 u16 in_mss = get_unaligned_be16(ptr);
3919 if (opt_rx->user_mss &&
3920 opt_rx->user_mss < in_mss)
3921 in_mss = opt_rx->user_mss;
3922 opt_rx->mss_clamp = in_mss;
3927 if (opsize == TCPOLEN_WINDOW && th->syn &&
3928 !estab && net->ipv4.sysctl_tcp_window_scaling) {
3929 __u8 snd_wscale = *(__u8 *)ptr;
3930 opt_rx->wscale_ok = 1;
3931 if (snd_wscale > TCP_MAX_WSCALE) {
3932 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n",
3936 snd_wscale = TCP_MAX_WSCALE;
3938 opt_rx->snd_wscale = snd_wscale;
3941 case TCPOPT_TIMESTAMP:
3942 if ((opsize == TCPOLEN_TIMESTAMP) &&
3943 ((estab && opt_rx->tstamp_ok) ||
3944 (!estab && net->ipv4.sysctl_tcp_timestamps))) {
3945 opt_rx->saw_tstamp = 1;
3946 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3947 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3950 case TCPOPT_SACK_PERM:
3951 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3952 !estab && net->ipv4.sysctl_tcp_sack) {
3953 opt_rx->sack_ok = TCP_SACK_SEEN;
3954 tcp_sack_reset(opt_rx);
3959 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3960 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3962 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3965 #ifdef CONFIG_TCP_MD5SIG
3968 * The MD5 Hash has already been
3969 * checked (see tcp_v{4,6}_do_rcv()).
3973 case TCPOPT_FASTOPEN:
3974 tcp_parse_fastopen_option(
3975 opsize - TCPOLEN_FASTOPEN_BASE,
3976 ptr, th->syn, foc, false);
3980 /* Fast Open option shares code 254 using a
3981 * 16 bits magic number.
3983 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
3984 get_unaligned_be16(ptr) ==
3985 TCPOPT_FASTOPEN_MAGIC)
3986 tcp_parse_fastopen_option(opsize -
3987 TCPOLEN_EXP_FASTOPEN_BASE,
3988 ptr + 2, th->syn, foc, true);
3990 smc_parse_options(th, opt_rx, ptr,
4000 EXPORT_SYMBOL(tcp_parse_options);
4002 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
4004 const __be32 *ptr = (const __be32 *)(th + 1);
4006 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
4007 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
4008 tp->rx_opt.saw_tstamp = 1;
4010 tp->rx_opt.rcv_tsval = ntohl(*ptr);
4013 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
4015 tp->rx_opt.rcv_tsecr = 0;
4021 /* Fast parse options. This hopes to only see timestamps.
4022 * If it is wrong it falls back on tcp_parse_options().
4024 static bool tcp_fast_parse_options(const struct net *net,
4025 const struct sk_buff *skb,
4026 const struct tcphdr *th, struct tcp_sock *tp)
4028 /* In the spirit of fast parsing, compare doff directly to constant
4029 * values. Because equality is used, short doff can be ignored here.
4031 if (th->doff == (sizeof(*th) / 4)) {
4032 tp->rx_opt.saw_tstamp = 0;
4034 } else if (tp->rx_opt.tstamp_ok &&
4035 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
4036 if (tcp_parse_aligned_timestamp(tp, th))
4040 tcp_parse_options(net, skb, &tp->rx_opt, 1, NULL);
4041 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
4042 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
4047 #ifdef CONFIG_TCP_MD5SIG
4049 * Parse MD5 Signature option
4051 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
4053 int length = (th->doff << 2) - sizeof(*th);
4054 const u8 *ptr = (const u8 *)(th + 1);
4056 /* If not enough data remaining, we can short cut */
4057 while (length >= TCPOLEN_MD5SIG) {
4058 int opcode = *ptr++;
4069 if (opsize < 2 || opsize > length)
4071 if (opcode == TCPOPT_MD5SIG)
4072 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
4079 EXPORT_SYMBOL(tcp_parse_md5sig_option);
4082 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
4084 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
4085 * it can pass through stack. So, the following predicate verifies that
4086 * this segment is not used for anything but congestion avoidance or
4087 * fast retransmit. Moreover, we even are able to eliminate most of such
4088 * second order effects, if we apply some small "replay" window (~RTO)
4089 * to timestamp space.
4091 * All these measures still do not guarantee that we reject wrapped ACKs
4092 * on networks with high bandwidth, when sequence space is recycled fastly,
4093 * but it guarantees that such events will be very rare and do not affect
4094 * connection seriously. This doesn't look nice, but alas, PAWS is really
4097 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
4098 * states that events when retransmit arrives after original data are rare.
4099 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
4100 * the biggest problem on large power networks even with minor reordering.
4101 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
4102 * up to bandwidth of 18Gigabit/sec. 8) ]
4105 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
4107 const struct tcp_sock *tp = tcp_sk(sk);
4108 const struct tcphdr *th = tcp_hdr(skb);
4109 u32 seq = TCP_SKB_CB(skb)->seq;
4110 u32 ack = TCP_SKB_CB(skb)->ack_seq;
4112 return (/* 1. Pure ACK with correct sequence number. */
4113 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
4115 /* 2. ... and duplicate ACK. */
4116 ack == tp->snd_una &&
4118 /* 3. ... and does not update window. */
4119 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
4121 /* 4. ... and sits in replay window. */
4122 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
4125 static inline bool tcp_paws_discard(const struct sock *sk,
4126 const struct sk_buff *skb)
4128 const struct tcp_sock *tp = tcp_sk(sk);
4130 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4131 !tcp_disordered_ack(sk, skb);
4134 /* Check segment sequence number for validity.
4136 * Segment controls are considered valid, if the segment
4137 * fits to the window after truncation to the window. Acceptability
4138 * of data (and SYN, FIN, of course) is checked separately.
4139 * See tcp_data_queue(), for example.
4141 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4142 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4143 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4144 * (borrowed from freebsd)
4147 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
4149 return !before(end_seq, tp->rcv_wup) &&
4150 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
4153 /* When we get a reset we do this. */
4154 void tcp_reset(struct sock *sk)
4156 trace_tcp_receive_reset(sk);
4158 /* We want the right error as BSD sees it (and indeed as we do). */
4159 switch (sk->sk_state) {
4161 sk->sk_err = ECONNREFUSED;
4163 case TCP_CLOSE_WAIT:
4169 sk->sk_err = ECONNRESET;
4171 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4174 tcp_write_queue_purge(sk);
4177 if (!sock_flag(sk, SOCK_DEAD))
4178 sk->sk_error_report(sk);
4182 * Process the FIN bit. This now behaves as it is supposed to work
4183 * and the FIN takes effect when it is validly part of sequence
4184 * space. Not before when we get holes.
4186 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4187 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4190 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4191 * close and we go into CLOSING (and later onto TIME-WAIT)
4193 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4195 void tcp_fin(struct sock *sk)
4197 struct tcp_sock *tp = tcp_sk(sk);
4199 inet_csk_schedule_ack(sk);
4201 sk->sk_shutdown |= RCV_SHUTDOWN;
4202 sock_set_flag(sk, SOCK_DONE);
4204 switch (sk->sk_state) {
4206 case TCP_ESTABLISHED:
4207 /* Move to CLOSE_WAIT */
4208 tcp_set_state(sk, TCP_CLOSE_WAIT);
4209 inet_csk_enter_pingpong_mode(sk);
4212 case TCP_CLOSE_WAIT:
4214 /* Received a retransmission of the FIN, do
4219 /* RFC793: Remain in the LAST-ACK state. */
4223 /* This case occurs when a simultaneous close
4224 * happens, we must ack the received FIN and
4225 * enter the CLOSING state.
4228 tcp_set_state(sk, TCP_CLOSING);
4231 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4233 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4236 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4237 * cases we should never reach this piece of code.
4239 pr_err("%s: Impossible, sk->sk_state=%d\n",
4240 __func__, sk->sk_state);
4244 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4245 * Probably, we should reset in this case. For now drop them.
4247 skb_rbtree_purge(&tp->out_of_order_queue);
4248 if (tcp_is_sack(tp))
4249 tcp_sack_reset(&tp->rx_opt);
4252 if (!sock_flag(sk, SOCK_DEAD)) {
4253 sk->sk_state_change(sk);
4255 /* Do not send POLL_HUP for half duplex close. */
4256 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4257 sk->sk_state == TCP_CLOSE)
4258 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4260 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4264 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4267 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4268 if (before(seq, sp->start_seq))
4269 sp->start_seq = seq;
4270 if (after(end_seq, sp->end_seq))
4271 sp->end_seq = end_seq;
4277 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4279 struct tcp_sock *tp = tcp_sk(sk);
4281 if (tcp_is_sack(tp) && sock_net(sk)->ipv4.sysctl_tcp_dsack) {
4284 if (before(seq, tp->rcv_nxt))
4285 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4287 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4289 NET_INC_STATS(sock_net(sk), mib_idx);
4291 tp->rx_opt.dsack = 1;
4292 tp->duplicate_sack[0].start_seq = seq;
4293 tp->duplicate_sack[0].end_seq = end_seq;
4297 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4299 struct tcp_sock *tp = tcp_sk(sk);
4301 if (!tp->rx_opt.dsack)
4302 tcp_dsack_set(sk, seq, end_seq);
4304 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4307 static void tcp_rcv_spurious_retrans(struct sock *sk, const struct sk_buff *skb)
4309 /* When the ACK path fails or drops most ACKs, the sender would
4310 * timeout and spuriously retransmit the same segment repeatedly.
4311 * The receiver remembers and reflects via DSACKs. Leverage the
4312 * DSACK state and change the txhash to re-route speculatively.
4314 if (TCP_SKB_CB(skb)->seq == tcp_sk(sk)->duplicate_sack[0].start_seq) {
4315 sk_rethink_txhash(sk);
4316 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDUPLICATEDATAREHASH);
4320 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4322 struct tcp_sock *tp = tcp_sk(sk);
4324 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4325 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4326 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4327 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
4329 if (tcp_is_sack(tp) && sock_net(sk)->ipv4.sysctl_tcp_dsack) {
4330 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4332 tcp_rcv_spurious_retrans(sk, skb);
4333 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4334 end_seq = tp->rcv_nxt;
4335 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4342 /* These routines update the SACK block as out-of-order packets arrive or
4343 * in-order packets close up the sequence space.
4345 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4348 struct tcp_sack_block *sp = &tp->selective_acks[0];
4349 struct tcp_sack_block *swalk = sp + 1;
4351 /* See if the recent change to the first SACK eats into
4352 * or hits the sequence space of other SACK blocks, if so coalesce.
4354 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4355 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4358 /* Zap SWALK, by moving every further SACK up by one slot.
4359 * Decrease num_sacks.
4361 tp->rx_opt.num_sacks--;
4362 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4366 this_sack++, swalk++;
4370 static void tcp_sack_compress_send_ack(struct sock *sk)
4372 struct tcp_sock *tp = tcp_sk(sk);
4374 if (!tp->compressed_ack)
4377 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
4380 /* Since we have to send one ack finally,
4381 * substract one from tp->compressed_ack to keep
4382 * LINUX_MIB_TCPACKCOMPRESSED accurate.
4384 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
4385 tp->compressed_ack - 1);
4387 tp->compressed_ack = 0;
4391 /* Reasonable amount of sack blocks included in TCP SACK option
4392 * The max is 4, but this becomes 3 if TCP timestamps are there.
4393 * Given that SACK packets might be lost, be conservative and use 2.
4395 #define TCP_SACK_BLOCKS_EXPECTED 2
4397 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4399 struct tcp_sock *tp = tcp_sk(sk);
4400 struct tcp_sack_block *sp = &tp->selective_acks[0];
4401 int cur_sacks = tp->rx_opt.num_sacks;
4407 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4408 if (tcp_sack_extend(sp, seq, end_seq)) {
4409 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED)
4410 tcp_sack_compress_send_ack(sk);
4411 /* Rotate this_sack to the first one. */
4412 for (; this_sack > 0; this_sack--, sp--)
4413 swap(*sp, *(sp - 1));
4415 tcp_sack_maybe_coalesce(tp);
4420 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED)
4421 tcp_sack_compress_send_ack(sk);
4423 /* Could not find an adjacent existing SACK, build a new one,
4424 * put it at the front, and shift everyone else down. We
4425 * always know there is at least one SACK present already here.
4427 * If the sack array is full, forget about the last one.
4429 if (this_sack >= TCP_NUM_SACKS) {
4431 tp->rx_opt.num_sacks--;
4434 for (; this_sack > 0; this_sack--, sp--)
4438 /* Build the new head SACK, and we're done. */
4439 sp->start_seq = seq;
4440 sp->end_seq = end_seq;
4441 tp->rx_opt.num_sacks++;
4444 /* RCV.NXT advances, some SACKs should be eaten. */
4446 static void tcp_sack_remove(struct tcp_sock *tp)
4448 struct tcp_sack_block *sp = &tp->selective_acks[0];
4449 int num_sacks = tp->rx_opt.num_sacks;
4452 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4453 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4454 tp->rx_opt.num_sacks = 0;
4458 for (this_sack = 0; this_sack < num_sacks;) {
4459 /* Check if the start of the sack is covered by RCV.NXT. */
4460 if (!before(tp->rcv_nxt, sp->start_seq)) {
4463 /* RCV.NXT must cover all the block! */
4464 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4466 /* Zap this SACK, by moving forward any other SACKS. */
4467 for (i = this_sack+1; i < num_sacks; i++)
4468 tp->selective_acks[i-1] = tp->selective_acks[i];
4475 tp->rx_opt.num_sacks = num_sacks;
4479 * tcp_try_coalesce - try to merge skb to prior one
4482 * @from: buffer to add in queue
4483 * @fragstolen: pointer to boolean
4485 * Before queueing skb @from after @to, try to merge them
4486 * to reduce overall memory use and queue lengths, if cost is small.
4487 * Packets in ofo or receive queues can stay a long time.
4488 * Better try to coalesce them right now to avoid future collapses.
4489 * Returns true if caller should free @from instead of queueing it
4491 static bool tcp_try_coalesce(struct sock *sk,
4493 struct sk_buff *from,
4498 *fragstolen = false;
4500 /* Its possible this segment overlaps with prior segment in queue */
4501 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4504 if (!mptcp_skb_can_collapse(to, from))
4507 #ifdef CONFIG_TLS_DEVICE
4508 if (from->decrypted != to->decrypted)
4512 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4515 atomic_add(delta, &sk->sk_rmem_alloc);
4516 sk_mem_charge(sk, delta);
4517 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4518 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4519 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4520 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4522 if (TCP_SKB_CB(from)->has_rxtstamp) {
4523 TCP_SKB_CB(to)->has_rxtstamp = true;
4524 to->tstamp = from->tstamp;
4525 skb_hwtstamps(to)->hwtstamp = skb_hwtstamps(from)->hwtstamp;
4531 static bool tcp_ooo_try_coalesce(struct sock *sk,
4533 struct sk_buff *from,
4536 bool res = tcp_try_coalesce(sk, to, from, fragstolen);
4538 /* In case tcp_drop() is called later, update to->gso_segs */
4540 u32 gso_segs = max_t(u16, 1, skb_shinfo(to)->gso_segs) +
4541 max_t(u16, 1, skb_shinfo(from)->gso_segs);
4543 skb_shinfo(to)->gso_segs = min_t(u32, gso_segs, 0xFFFF);
4548 static void tcp_drop(struct sock *sk, struct sk_buff *skb)
4550 sk_drops_add(sk, skb);
4554 /* This one checks to see if we can put data from the
4555 * out_of_order queue into the receive_queue.
4557 static void tcp_ofo_queue(struct sock *sk)
4559 struct tcp_sock *tp = tcp_sk(sk);
4560 __u32 dsack_high = tp->rcv_nxt;
4561 bool fin, fragstolen, eaten;
4562 struct sk_buff *skb, *tail;
4565 p = rb_first(&tp->out_of_order_queue);
4568 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4571 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4572 __u32 dsack = dsack_high;
4573 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4574 dsack_high = TCP_SKB_CB(skb)->end_seq;
4575 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4578 rb_erase(&skb->rbnode, &tp->out_of_order_queue);
4580 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) {
4585 tail = skb_peek_tail(&sk->sk_receive_queue);
4586 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4587 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4588 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
4590 __skb_queue_tail(&sk->sk_receive_queue, skb);
4592 kfree_skb_partial(skb, fragstolen);
4594 if (unlikely(fin)) {
4596 /* tcp_fin() purges tp->out_of_order_queue,
4597 * so we must end this loop right now.
4604 static bool tcp_prune_ofo_queue(struct sock *sk);
4605 static int tcp_prune_queue(struct sock *sk);
4607 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4610 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4611 !sk_rmem_schedule(sk, skb, size)) {
4613 if (tcp_prune_queue(sk) < 0)
4616 while (!sk_rmem_schedule(sk, skb, size)) {
4617 if (!tcp_prune_ofo_queue(sk))
4624 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4626 struct tcp_sock *tp = tcp_sk(sk);
4627 struct rb_node **p, *parent;
4628 struct sk_buff *skb1;
4632 tcp_ecn_check_ce(sk, skb);
4634 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4635 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP);
4636 sk->sk_data_ready(sk);
4641 /* Disable header prediction. */
4643 inet_csk_schedule_ack(sk);
4645 tp->rcv_ooopack += max_t(u16, 1, skb_shinfo(skb)->gso_segs);
4646 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4647 seq = TCP_SKB_CB(skb)->seq;
4648 end_seq = TCP_SKB_CB(skb)->end_seq;
4650 p = &tp->out_of_order_queue.rb_node;
4651 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4652 /* Initial out of order segment, build 1 SACK. */
4653 if (tcp_is_sack(tp)) {
4654 tp->rx_opt.num_sacks = 1;
4655 tp->selective_acks[0].start_seq = seq;
4656 tp->selective_acks[0].end_seq = end_seq;
4658 rb_link_node(&skb->rbnode, NULL, p);
4659 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4660 tp->ooo_last_skb = skb;
4664 /* In the typical case, we are adding an skb to the end of the list.
4665 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4667 if (tcp_ooo_try_coalesce(sk, tp->ooo_last_skb,
4668 skb, &fragstolen)) {
4670 /* For non sack flows, do not grow window to force DUPACK
4671 * and trigger fast retransmit.
4673 if (tcp_is_sack(tp))
4674 tcp_grow_window(sk, skb);
4675 kfree_skb_partial(skb, fragstolen);
4679 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4680 if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) {
4681 parent = &tp->ooo_last_skb->rbnode;
4682 p = &parent->rb_right;
4686 /* Find place to insert this segment. Handle overlaps on the way. */
4690 skb1 = rb_to_skb(parent);
4691 if (before(seq, TCP_SKB_CB(skb1)->seq)) {
4692 p = &parent->rb_left;
4695 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4696 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4697 /* All the bits are present. Drop. */
4698 NET_INC_STATS(sock_net(sk),
4699 LINUX_MIB_TCPOFOMERGE);
4702 tcp_dsack_set(sk, seq, end_seq);
4705 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4706 /* Partial overlap. */
4707 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq);
4709 /* skb's seq == skb1's seq and skb covers skb1.
4710 * Replace skb1 with skb.
4712 rb_replace_node(&skb1->rbnode, &skb->rbnode,
4713 &tp->out_of_order_queue);
4714 tcp_dsack_extend(sk,
4715 TCP_SKB_CB(skb1)->seq,
4716 TCP_SKB_CB(skb1)->end_seq);
4717 NET_INC_STATS(sock_net(sk),
4718 LINUX_MIB_TCPOFOMERGE);
4722 } else if (tcp_ooo_try_coalesce(sk, skb1,
4723 skb, &fragstolen)) {
4726 p = &parent->rb_right;
4729 /* Insert segment into RB tree. */
4730 rb_link_node(&skb->rbnode, parent, p);
4731 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4734 /* Remove other segments covered by skb. */
4735 while ((skb1 = skb_rb_next(skb)) != NULL) {
4736 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4738 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4739 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4743 rb_erase(&skb1->rbnode, &tp->out_of_order_queue);
4744 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4745 TCP_SKB_CB(skb1)->end_seq);
4746 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4749 /* If there is no skb after us, we are the last_skb ! */
4751 tp->ooo_last_skb = skb;
4754 if (tcp_is_sack(tp))
4755 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4758 /* For non sack flows, do not grow window to force DUPACK
4759 * and trigger fast retransmit.
4761 if (tcp_is_sack(tp))
4762 tcp_grow_window(sk, skb);
4764 skb_set_owner_r(skb, sk);
4768 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb,
4772 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4775 tcp_try_coalesce(sk, tail,
4776 skb, fragstolen)) ? 1 : 0;
4777 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
4779 __skb_queue_tail(&sk->sk_receive_queue, skb);
4780 skb_set_owner_r(skb, sk);
4785 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4787 struct sk_buff *skb;
4795 if (size > PAGE_SIZE) {
4796 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
4798 data_len = npages << PAGE_SHIFT;
4799 size = data_len + (size & ~PAGE_MASK);
4801 skb = alloc_skb_with_frags(size - data_len, data_len,
4802 PAGE_ALLOC_COSTLY_ORDER,
4803 &err, sk->sk_allocation);
4807 skb_put(skb, size - data_len);
4808 skb->data_len = data_len;
4811 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
4812 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
4816 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
4820 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4821 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4822 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4824 if (tcp_queue_rcv(sk, skb, &fragstolen)) {
4825 WARN_ON_ONCE(fragstolen); /* should not happen */
4837 void tcp_data_ready(struct sock *sk)
4839 const struct tcp_sock *tp = tcp_sk(sk);
4840 int avail = tp->rcv_nxt - tp->copied_seq;
4842 if (avail < sk->sk_rcvlowat && !tcp_rmem_pressure(sk) &&
4843 !sock_flag(sk, SOCK_DONE))
4846 sk->sk_data_ready(sk);
4849 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4851 struct tcp_sock *tp = tcp_sk(sk);
4855 if (sk_is_mptcp(sk))
4856 mptcp_incoming_options(sk, skb, &tp->rx_opt);
4858 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
4863 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
4865 tp->rx_opt.dsack = 0;
4867 /* Queue data for delivery to the user.
4868 * Packets in sequence go to the receive queue.
4869 * Out of sequence packets to the out_of_order_queue.
4871 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4872 if (tcp_receive_window(tp) == 0) {
4873 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
4877 /* Ok. In sequence. In window. */
4879 if (skb_queue_len(&sk->sk_receive_queue) == 0)
4880 sk_forced_mem_schedule(sk, skb->truesize);
4881 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
4882 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
4883 sk->sk_data_ready(sk);
4887 eaten = tcp_queue_rcv(sk, skb, &fragstolen);
4889 tcp_event_data_recv(sk, skb);
4890 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4893 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4896 /* RFC5681. 4.2. SHOULD send immediate ACK, when
4897 * gap in queue is filled.
4899 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
4900 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
4903 if (tp->rx_opt.num_sacks)
4904 tcp_sack_remove(tp);
4906 tcp_fast_path_check(sk);
4909 kfree_skb_partial(skb, fragstolen);
4910 if (!sock_flag(sk, SOCK_DEAD))
4915 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4916 tcp_rcv_spurious_retrans(sk, skb);
4917 /* A retransmit, 2nd most common case. Force an immediate ack. */
4918 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4919 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4922 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
4923 inet_csk_schedule_ack(sk);
4929 /* Out of window. F.e. zero window probe. */
4930 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4933 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4934 /* Partial packet, seq < rcv_next < end_seq */
4935 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4937 /* If window is closed, drop tail of packet. But after
4938 * remembering D-SACK for its head made in previous line.
4940 if (!tcp_receive_window(tp)) {
4941 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
4947 tcp_data_queue_ofo(sk, skb);
4950 static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list)
4953 return !skb_queue_is_last(list, skb) ? skb->next : NULL;
4955 return skb_rb_next(skb);
4958 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4959 struct sk_buff_head *list,
4960 struct rb_root *root)
4962 struct sk_buff *next = tcp_skb_next(skb, list);
4965 __skb_unlink(skb, list);
4967 rb_erase(&skb->rbnode, root);
4970 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4975 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
4976 void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb)
4978 struct rb_node **p = &root->rb_node;
4979 struct rb_node *parent = NULL;
4980 struct sk_buff *skb1;
4984 skb1 = rb_to_skb(parent);
4985 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq))
4986 p = &parent->rb_left;
4988 p = &parent->rb_right;
4990 rb_link_node(&skb->rbnode, parent, p);
4991 rb_insert_color(&skb->rbnode, root);
4994 /* Collapse contiguous sequence of skbs head..tail with
4995 * sequence numbers start..end.
4997 * If tail is NULL, this means until the end of the queue.
4999 * Segments with FIN/SYN are not collapsed (only because this
5003 tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root,
5004 struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end)
5006 struct sk_buff *skb = head, *n;
5007 struct sk_buff_head tmp;
5010 /* First, check that queue is collapsible and find
5011 * the point where collapsing can be useful.
5014 for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) {
5015 n = tcp_skb_next(skb, list);
5017 /* No new bits? It is possible on ofo queue. */
5018 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
5019 skb = tcp_collapse_one(sk, skb, list, root);
5025 /* The first skb to collapse is:
5027 * - bloated or contains data before "start" or
5028 * overlaps to the next one and mptcp allow collapsing.
5030 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
5031 (tcp_win_from_space(sk, skb->truesize) > skb->len ||
5032 before(TCP_SKB_CB(skb)->seq, start))) {
5033 end_of_skbs = false;
5037 if (n && n != tail && mptcp_skb_can_collapse(skb, n) &&
5038 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) {
5039 end_of_skbs = false;
5043 /* Decided to skip this, advance start seq. */
5044 start = TCP_SKB_CB(skb)->end_seq;
5047 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
5050 __skb_queue_head_init(&tmp);
5052 while (before(start, end)) {
5053 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
5054 struct sk_buff *nskb;
5056 nskb = alloc_skb(copy, GFP_ATOMIC);
5060 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
5061 #ifdef CONFIG_TLS_DEVICE
5062 nskb->decrypted = skb->decrypted;
5064 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
5066 __skb_queue_before(list, skb, nskb);
5068 __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */
5069 skb_set_owner_r(nskb, sk);
5070 mptcp_skb_ext_move(nskb, skb);
5072 /* Copy data, releasing collapsed skbs. */
5074 int offset = start - TCP_SKB_CB(skb)->seq;
5075 int size = TCP_SKB_CB(skb)->end_seq - start;
5079 size = min(copy, size);
5080 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
5082 TCP_SKB_CB(nskb)->end_seq += size;
5086 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
5087 skb = tcp_collapse_one(sk, skb, list, root);
5090 !mptcp_skb_can_collapse(nskb, skb) ||
5091 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
5093 #ifdef CONFIG_TLS_DEVICE
5094 if (skb->decrypted != nskb->decrypted)
5101 skb_queue_walk_safe(&tmp, skb, n)
5102 tcp_rbtree_insert(root, skb);
5105 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
5106 * and tcp_collapse() them until all the queue is collapsed.
5108 static void tcp_collapse_ofo_queue(struct sock *sk)
5110 struct tcp_sock *tp = tcp_sk(sk);
5111 u32 range_truesize, sum_tiny = 0;
5112 struct sk_buff *skb, *head;
5115 skb = skb_rb_first(&tp->out_of_order_queue);
5118 tp->ooo_last_skb = skb_rb_last(&tp->out_of_order_queue);
5121 start = TCP_SKB_CB(skb)->seq;
5122 end = TCP_SKB_CB(skb)->end_seq;
5123 range_truesize = skb->truesize;
5125 for (head = skb;;) {
5126 skb = skb_rb_next(skb);
5128 /* Range is terminated when we see a gap or when
5129 * we are at the queue end.
5132 after(TCP_SKB_CB(skb)->seq, end) ||
5133 before(TCP_SKB_CB(skb)->end_seq, start)) {
5134 /* Do not attempt collapsing tiny skbs */
5135 if (range_truesize != head->truesize ||
5136 end - start >= SKB_WITH_OVERHEAD(SK_MEM_QUANTUM)) {
5137 tcp_collapse(sk, NULL, &tp->out_of_order_queue,
5138 head, skb, start, end);
5140 sum_tiny += range_truesize;
5141 if (sum_tiny > sk->sk_rcvbuf >> 3)
5147 range_truesize += skb->truesize;
5148 if (unlikely(before(TCP_SKB_CB(skb)->seq, start)))
5149 start = TCP_SKB_CB(skb)->seq;
5150 if (after(TCP_SKB_CB(skb)->end_seq, end))
5151 end = TCP_SKB_CB(skb)->end_seq;
5156 * Clean the out-of-order queue to make room.
5157 * We drop high sequences packets to :
5158 * 1) Let a chance for holes to be filled.
5159 * 2) not add too big latencies if thousands of packets sit there.
5160 * (But if application shrinks SO_RCVBUF, we could still end up
5161 * freeing whole queue here)
5162 * 3) Drop at least 12.5 % of sk_rcvbuf to avoid malicious attacks.
5164 * Return true if queue has shrunk.
5166 static bool tcp_prune_ofo_queue(struct sock *sk)
5168 struct tcp_sock *tp = tcp_sk(sk);
5169 struct rb_node *node, *prev;
5172 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
5175 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED);
5176 goal = sk->sk_rcvbuf >> 3;
5177 node = &tp->ooo_last_skb->rbnode;
5179 prev = rb_prev(node);
5180 rb_erase(node, &tp->out_of_order_queue);
5181 goal -= rb_to_skb(node)->truesize;
5182 tcp_drop(sk, rb_to_skb(node));
5183 if (!prev || goal <= 0) {
5185 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf &&
5186 !tcp_under_memory_pressure(sk))
5188 goal = sk->sk_rcvbuf >> 3;
5192 tp->ooo_last_skb = rb_to_skb(prev);
5194 /* Reset SACK state. A conforming SACK implementation will
5195 * do the same at a timeout based retransmit. When a connection
5196 * is in a sad state like this, we care only about integrity
5197 * of the connection not performance.
5199 if (tp->rx_opt.sack_ok)
5200 tcp_sack_reset(&tp->rx_opt);
5204 /* Reduce allocated memory if we can, trying to get
5205 * the socket within its memory limits again.
5207 * Return less than zero if we should start dropping frames
5208 * until the socket owning process reads some of the data
5209 * to stabilize the situation.
5211 static int tcp_prune_queue(struct sock *sk)
5213 struct tcp_sock *tp = tcp_sk(sk);
5215 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED);
5217 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
5218 tcp_clamp_window(sk);
5219 else if (tcp_under_memory_pressure(sk))
5220 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
5222 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5225 tcp_collapse_ofo_queue(sk);
5226 if (!skb_queue_empty(&sk->sk_receive_queue))
5227 tcp_collapse(sk, &sk->sk_receive_queue, NULL,
5228 skb_peek(&sk->sk_receive_queue),
5230 tp->copied_seq, tp->rcv_nxt);
5233 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5236 /* Collapsing did not help, destructive actions follow.
5237 * This must not ever occur. */
5239 tcp_prune_ofo_queue(sk);
5241 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5244 /* If we are really being abused, tell the caller to silently
5245 * drop receive data on the floor. It will get retransmitted
5246 * and hopefully then we'll have sufficient space.
5248 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED);
5250 /* Massive buffer overcommit. */
5255 static bool tcp_should_expand_sndbuf(const struct sock *sk)
5257 const struct tcp_sock *tp = tcp_sk(sk);
5259 /* If the user specified a specific send buffer setting, do
5262 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
5265 /* If we are under global TCP memory pressure, do not expand. */
5266 if (tcp_under_memory_pressure(sk))
5269 /* If we are under soft global TCP memory pressure, do not expand. */
5270 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
5273 /* If we filled the congestion window, do not expand. */
5274 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
5280 /* When incoming ACK allowed to free some skb from write_queue,
5281 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
5282 * on the exit from tcp input handler.
5284 * PROBLEM: sndbuf expansion does not work well with largesend.
5286 static void tcp_new_space(struct sock *sk)
5288 struct tcp_sock *tp = tcp_sk(sk);
5290 if (tcp_should_expand_sndbuf(sk)) {
5291 tcp_sndbuf_expand(sk);
5292 tp->snd_cwnd_stamp = tcp_jiffies32;
5295 sk->sk_write_space(sk);
5298 static void tcp_check_space(struct sock *sk)
5300 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
5301 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
5302 /* pairs with tcp_poll() */
5304 if (sk->sk_socket &&
5305 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
5307 if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5308 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
5313 static inline void tcp_data_snd_check(struct sock *sk)
5315 tcp_push_pending_frames(sk);
5316 tcp_check_space(sk);
5320 * Check if sending an ack is needed.
5322 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5324 struct tcp_sock *tp = tcp_sk(sk);
5325 unsigned long rtt, delay;
5327 /* More than one full frame received... */
5328 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5329 /* ... and right edge of window advances far enough.
5330 * (tcp_recvmsg() will send ACK otherwise).
5331 * If application uses SO_RCVLOWAT, we want send ack now if
5332 * we have not received enough bytes to satisfy the condition.
5334 (tp->rcv_nxt - tp->copied_seq < sk->sk_rcvlowat ||
5335 __tcp_select_window(sk) >= tp->rcv_wnd)) ||
5336 /* We ACK each frame or... */
5337 tcp_in_quickack_mode(sk) ||
5338 /* Protocol state mandates a one-time immediate ACK */
5339 inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOW) {
5345 if (!ofo_possible || RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5346 tcp_send_delayed_ack(sk);
5350 if (!tcp_is_sack(tp) ||
5351 tp->compressed_ack >= sock_net(sk)->ipv4.sysctl_tcp_comp_sack_nr)
5354 if (tp->compressed_ack_rcv_nxt != tp->rcv_nxt) {
5355 tp->compressed_ack_rcv_nxt = tp->rcv_nxt;
5356 tp->dup_ack_counter = 0;
5358 if (tp->dup_ack_counter < TCP_FASTRETRANS_THRESH) {
5359 tp->dup_ack_counter++;
5362 tp->compressed_ack++;
5363 if (hrtimer_is_queued(&tp->compressed_ack_timer))
5366 /* compress ack timer : 5 % of rtt, but no more than tcp_comp_sack_delay_ns */
5368 rtt = tp->rcv_rtt_est.rtt_us;
5369 if (tp->srtt_us && tp->srtt_us < rtt)
5372 delay = min_t(unsigned long, sock_net(sk)->ipv4.sysctl_tcp_comp_sack_delay_ns,
5373 rtt * (NSEC_PER_USEC >> 3)/20);
5375 hrtimer_start_range_ns(&tp->compressed_ack_timer, ns_to_ktime(delay),
5376 sock_net(sk)->ipv4.sysctl_tcp_comp_sack_slack_ns,
5377 HRTIMER_MODE_REL_PINNED_SOFT);
5380 static inline void tcp_ack_snd_check(struct sock *sk)
5382 if (!inet_csk_ack_scheduled(sk)) {
5383 /* We sent a data segment already. */
5386 __tcp_ack_snd_check(sk, 1);
5390 * This routine is only called when we have urgent data
5391 * signaled. Its the 'slow' part of tcp_urg. It could be
5392 * moved inline now as tcp_urg is only called from one
5393 * place. We handle URGent data wrong. We have to - as
5394 * BSD still doesn't use the correction from RFC961.
5395 * For 1003.1g we should support a new option TCP_STDURG to permit
5396 * either form (or just set the sysctl tcp_stdurg).
5399 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5401 struct tcp_sock *tp = tcp_sk(sk);
5402 u32 ptr = ntohs(th->urg_ptr);
5404 if (ptr && !sock_net(sk)->ipv4.sysctl_tcp_stdurg)
5406 ptr += ntohl(th->seq);
5408 /* Ignore urgent data that we've already seen and read. */
5409 if (after(tp->copied_seq, ptr))
5412 /* Do not replay urg ptr.
5414 * NOTE: interesting situation not covered by specs.
5415 * Misbehaving sender may send urg ptr, pointing to segment,
5416 * which we already have in ofo queue. We are not able to fetch
5417 * such data and will stay in TCP_URG_NOTYET until will be eaten
5418 * by recvmsg(). Seems, we are not obliged to handle such wicked
5419 * situations. But it is worth to think about possibility of some
5420 * DoSes using some hypothetical application level deadlock.
5422 if (before(ptr, tp->rcv_nxt))
5425 /* Do we already have a newer (or duplicate) urgent pointer? */
5426 if (tp->urg_data && !after(ptr, tp->urg_seq))
5429 /* Tell the world about our new urgent pointer. */
5432 /* We may be adding urgent data when the last byte read was
5433 * urgent. To do this requires some care. We cannot just ignore
5434 * tp->copied_seq since we would read the last urgent byte again
5435 * as data, nor can we alter copied_seq until this data arrives
5436 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5438 * NOTE. Double Dutch. Rendering to plain English: author of comment
5439 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5440 * and expect that both A and B disappear from stream. This is _wrong_.
5441 * Though this happens in BSD with high probability, this is occasional.
5442 * Any application relying on this is buggy. Note also, that fix "works"
5443 * only in this artificial test. Insert some normal data between A and B and we will
5444 * decline of BSD again. Verdict: it is better to remove to trap
5447 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5448 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5449 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5451 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5452 __skb_unlink(skb, &sk->sk_receive_queue);
5457 tp->urg_data = TCP_URG_NOTYET;
5458 WRITE_ONCE(tp->urg_seq, ptr);
5460 /* Disable header prediction. */
5464 /* This is the 'fast' part of urgent handling. */
5465 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5467 struct tcp_sock *tp = tcp_sk(sk);
5469 /* Check if we get a new urgent pointer - normally not. */
5471 tcp_check_urg(sk, th);
5473 /* Do we wait for any urgent data? - normally not... */
5474 if (tp->urg_data == TCP_URG_NOTYET) {
5475 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5478 /* Is the urgent pointer pointing into this packet? */
5479 if (ptr < skb->len) {
5481 if (skb_copy_bits(skb, ptr, &tmp, 1))
5483 tp->urg_data = TCP_URG_VALID | tmp;
5484 if (!sock_flag(sk, SOCK_DEAD))
5485 sk->sk_data_ready(sk);
5490 /* Accept RST for rcv_nxt - 1 after a FIN.
5491 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a
5492 * FIN is sent followed by a RST packet. The RST is sent with the same
5493 * sequence number as the FIN, and thus according to RFC 5961 a challenge
5494 * ACK should be sent. However, Mac OSX rate limits replies to challenge
5495 * ACKs on the closed socket. In addition middleboxes can drop either the
5496 * challenge ACK or a subsequent RST.
5498 static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb)
5500 struct tcp_sock *tp = tcp_sk(sk);
5502 return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) &&
5503 (1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK |
5507 /* Does PAWS and seqno based validation of an incoming segment, flags will
5508 * play significant role here.
5510 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5511 const struct tcphdr *th, int syn_inerr)
5513 struct tcp_sock *tp = tcp_sk(sk);
5514 bool rst_seq_match = false;
5516 /* RFC1323: H1. Apply PAWS check first. */
5517 if (tcp_fast_parse_options(sock_net(sk), skb, th, tp) &&
5518 tp->rx_opt.saw_tstamp &&
5519 tcp_paws_discard(sk, skb)) {
5521 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5522 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5523 LINUX_MIB_TCPACKSKIPPEDPAWS,
5524 &tp->last_oow_ack_time))
5525 tcp_send_dupack(sk, skb);
5528 /* Reset is accepted even if it did not pass PAWS. */
5531 /* Step 1: check sequence number */
5532 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5533 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5534 * (RST) segments are validated by checking their SEQ-fields."
5535 * And page 69: "If an incoming segment is not acceptable,
5536 * an acknowledgment should be sent in reply (unless the RST
5537 * bit is set, if so drop the segment and return)".
5542 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5543 LINUX_MIB_TCPACKSKIPPEDSEQ,
5544 &tp->last_oow_ack_time))
5545 tcp_send_dupack(sk, skb);
5546 } else if (tcp_reset_check(sk, skb)) {
5552 /* Step 2: check RST bit */
5554 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a
5555 * FIN and SACK too if available):
5556 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or
5557 * the right-most SACK block,
5559 * RESET the connection
5561 * Send a challenge ACK
5563 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt ||
5564 tcp_reset_check(sk, skb)) {
5565 rst_seq_match = true;
5566 } else if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) {
5567 struct tcp_sack_block *sp = &tp->selective_acks[0];
5568 int max_sack = sp[0].end_seq;
5571 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;
5573 max_sack = after(sp[this_sack].end_seq,
5575 sp[this_sack].end_seq : max_sack;
5578 if (TCP_SKB_CB(skb)->seq == max_sack)
5579 rst_seq_match = true;
5585 /* Disable TFO if RST is out-of-order
5586 * and no data has been received
5587 * for current active TFO socket
5589 if (tp->syn_fastopen && !tp->data_segs_in &&
5590 sk->sk_state == TCP_ESTABLISHED)
5591 tcp_fastopen_active_disable(sk);
5592 tcp_send_challenge_ack(sk, skb);
5597 /* step 3: check security and precedence [ignored] */
5599 /* step 4: Check for a SYN
5600 * RFC 5961 4.2 : Send a challenge ack
5605 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5606 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5607 tcp_send_challenge_ack(sk, skb);
5619 * TCP receive function for the ESTABLISHED state.
5621 * It is split into a fast path and a slow path. The fast path is
5623 * - A zero window was announced from us - zero window probing
5624 * is only handled properly in the slow path.
5625 * - Out of order segments arrived.
5626 * - Urgent data is expected.
5627 * - There is no buffer space left
5628 * - Unexpected TCP flags/window values/header lengths are received
5629 * (detected by checking the TCP header against pred_flags)
5630 * - Data is sent in both directions. Fast path only supports pure senders
5631 * or pure receivers (this means either the sequence number or the ack
5632 * value must stay constant)
5633 * - Unexpected TCP option.
5635 * When these conditions are not satisfied it drops into a standard
5636 * receive procedure patterned after RFC793 to handle all cases.
5637 * The first three cases are guaranteed by proper pred_flags setting,
5638 * the rest is checked inline. Fast processing is turned on in
5639 * tcp_data_queue when everything is OK.
5641 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb)
5643 const struct tcphdr *th = (const struct tcphdr *)skb->data;
5644 struct tcp_sock *tp = tcp_sk(sk);
5645 unsigned int len = skb->len;
5647 /* TCP congestion window tracking */
5648 trace_tcp_probe(sk, skb);
5650 tcp_mstamp_refresh(tp);
5651 if (unlikely(!sk->sk_rx_dst))
5652 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5654 * Header prediction.
5655 * The code loosely follows the one in the famous
5656 * "30 instruction TCP receive" Van Jacobson mail.
5658 * Van's trick is to deposit buffers into socket queue
5659 * on a device interrupt, to call tcp_recv function
5660 * on the receive process context and checksum and copy
5661 * the buffer to user space. smart...
5663 * Our current scheme is not silly either but we take the
5664 * extra cost of the net_bh soft interrupt processing...
5665 * We do checksum and copy also but from device to kernel.
5668 tp->rx_opt.saw_tstamp = 0;
5670 /* pred_flags is 0xS?10 << 16 + snd_wnd
5671 * if header_prediction is to be made
5672 * 'S' will always be tp->tcp_header_len >> 2
5673 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5674 * turn it off (when there are holes in the receive
5675 * space for instance)
5676 * PSH flag is ignored.
5679 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5680 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5681 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5682 int tcp_header_len = tp->tcp_header_len;
5684 /* Timestamp header prediction: tcp_header_len
5685 * is automatically equal to th->doff*4 due to pred_flags
5689 /* Check timestamp */
5690 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5691 /* No? Slow path! */
5692 if (!tcp_parse_aligned_timestamp(tp, th))
5695 /* If PAWS failed, check it more carefully in slow path */
5696 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5699 /* DO NOT update ts_recent here, if checksum fails
5700 * and timestamp was corrupted part, it will result
5701 * in a hung connection since we will drop all
5702 * future packets due to the PAWS test.
5706 if (len <= tcp_header_len) {
5707 /* Bulk data transfer: sender */
5708 if (len == tcp_header_len) {
5709 /* Predicted packet is in window by definition.
5710 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5711 * Hence, check seq<=rcv_wup reduces to:
5713 if (tcp_header_len ==
5714 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5715 tp->rcv_nxt == tp->rcv_wup)
5716 tcp_store_ts_recent(tp);
5718 /* We know that such packets are checksummed
5721 tcp_ack(sk, skb, 0);
5723 tcp_data_snd_check(sk);
5724 /* When receiving pure ack in fast path, update
5725 * last ts ecr directly instead of calling
5726 * tcp_rcv_rtt_measure_ts()
5728 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
5730 } else { /* Header too small */
5731 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5736 bool fragstolen = false;
5738 if (tcp_checksum_complete(skb))
5741 if ((int)skb->truesize > sk->sk_forward_alloc)
5744 /* Predicted packet is in window by definition.
5745 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5746 * Hence, check seq<=rcv_wup reduces to:
5748 if (tcp_header_len ==
5749 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5750 tp->rcv_nxt == tp->rcv_wup)
5751 tcp_store_ts_recent(tp);
5753 tcp_rcv_rtt_measure_ts(sk, skb);
5755 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS);
5757 /* Bulk data transfer: receiver */
5758 __skb_pull(skb, tcp_header_len);
5759 eaten = tcp_queue_rcv(sk, skb, &fragstolen);
5761 tcp_event_data_recv(sk, skb);
5763 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5764 /* Well, only one small jumplet in fast path... */
5765 tcp_ack(sk, skb, FLAG_DATA);
5766 tcp_data_snd_check(sk);
5767 if (!inet_csk_ack_scheduled(sk))
5771 __tcp_ack_snd_check(sk, 0);
5774 kfree_skb_partial(skb, fragstolen);
5781 if (len < (th->doff << 2) || tcp_checksum_complete(skb))
5784 if (!th->ack && !th->rst && !th->syn)
5788 * Standard slow path.
5791 if (!tcp_validate_incoming(sk, skb, th, 1))
5795 if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5798 tcp_rcv_rtt_measure_ts(sk, skb);
5800 /* Process urgent data. */
5801 tcp_urg(sk, skb, th);
5803 /* step 7: process the segment text */
5804 tcp_data_queue(sk, skb);
5806 tcp_data_snd_check(sk);
5807 tcp_ack_snd_check(sk);
5811 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
5812 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5817 EXPORT_SYMBOL(tcp_rcv_established);
5819 void tcp_init_transfer(struct sock *sk, int bpf_op)
5821 struct inet_connection_sock *icsk = inet_csk(sk);
5822 struct tcp_sock *tp = tcp_sk(sk);
5825 icsk->icsk_af_ops->rebuild_header(sk);
5826 tcp_init_metrics(sk);
5828 /* Initialize the congestion window to start the transfer.
5829 * Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
5830 * retransmitted. In light of RFC6298 more aggressive 1sec
5831 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
5832 * retransmission has occurred.
5834 if (tp->total_retrans > 1 && tp->undo_marker)
5837 tp->snd_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
5838 tp->snd_cwnd_stamp = tcp_jiffies32;
5840 tcp_call_bpf(sk, bpf_op, 0, NULL);
5841 tcp_init_congestion_control(sk);
5842 tcp_init_buffer_space(sk);
5845 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5847 struct tcp_sock *tp = tcp_sk(sk);
5848 struct inet_connection_sock *icsk = inet_csk(sk);
5850 tcp_set_state(sk, TCP_ESTABLISHED);
5851 icsk->icsk_ack.lrcvtime = tcp_jiffies32;
5854 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5855 security_inet_conn_established(sk, skb);
5856 sk_mark_napi_id(sk, skb);
5859 tcp_init_transfer(sk, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB);
5861 /* Prevent spurious tcp_cwnd_restart() on first data
5864 tp->lsndtime = tcp_jiffies32;
5866 if (sock_flag(sk, SOCK_KEEPOPEN))
5867 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5869 if (!tp->rx_opt.snd_wscale)
5870 __tcp_fast_path_on(tp, tp->snd_wnd);
5875 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5876 struct tcp_fastopen_cookie *cookie)
5878 struct tcp_sock *tp = tcp_sk(sk);
5879 struct sk_buff *data = tp->syn_data ? tcp_rtx_queue_head(sk) : NULL;
5880 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
5881 bool syn_drop = false;
5883 if (mss == tp->rx_opt.user_mss) {
5884 struct tcp_options_received opt;
5886 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5887 tcp_clear_options(&opt);
5888 opt.user_mss = opt.mss_clamp = 0;
5889 tcp_parse_options(sock_net(sk), synack, &opt, 0, NULL);
5890 mss = opt.mss_clamp;
5893 if (!tp->syn_fastopen) {
5894 /* Ignore an unsolicited cookie */
5896 } else if (tp->total_retrans) {
5897 /* SYN timed out and the SYN-ACK neither has a cookie nor
5898 * acknowledges data. Presumably the remote received only
5899 * the retransmitted (regular) SYNs: either the original
5900 * SYN-data or the corresponding SYN-ACK was dropped.
5902 syn_drop = (cookie->len < 0 && data);
5903 } else if (cookie->len < 0 && !tp->syn_data) {
5904 /* We requested a cookie but didn't get it. If we did not use
5905 * the (old) exp opt format then try so next time (try_exp=1).
5906 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5908 try_exp = tp->syn_fastopen_exp ? 2 : 1;
5911 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
5913 if (data) { /* Retransmit unacked data in SYN */
5914 if (tp->total_retrans)
5915 tp->fastopen_client_fail = TFO_SYN_RETRANSMITTED;
5917 tp->fastopen_client_fail = TFO_DATA_NOT_ACKED;
5918 skb_rbtree_walk_from(data) {
5919 if (__tcp_retransmit_skb(sk, data, 1))
5923 NET_INC_STATS(sock_net(sk),
5924 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
5927 tp->syn_data_acked = tp->syn_data;
5928 if (tp->syn_data_acked) {
5929 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
5930 /* SYN-data is counted as two separate packets in tcp_ack() */
5931 if (tp->delivered > 1)
5935 tcp_fastopen_add_skb(sk, synack);
5940 static void smc_check_reset_syn(struct tcp_sock *tp)
5942 #if IS_ENABLED(CONFIG_SMC)
5943 if (static_branch_unlikely(&tcp_have_smc)) {
5944 if (tp->syn_smc && !tp->rx_opt.smc_ok)
5950 static void tcp_try_undo_spurious_syn(struct sock *sk)
5952 struct tcp_sock *tp = tcp_sk(sk);
5955 /* undo_marker is set when SYN or SYNACK times out. The timeout is
5956 * spurious if the ACK's timestamp option echo value matches the
5957 * original SYN timestamp.
5959 syn_stamp = tp->retrans_stamp;
5960 if (tp->undo_marker && syn_stamp && tp->rx_opt.saw_tstamp &&
5961 syn_stamp == tp->rx_opt.rcv_tsecr)
5962 tp->undo_marker = 0;
5965 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5966 const struct tcphdr *th)
5968 struct inet_connection_sock *icsk = inet_csk(sk);
5969 struct tcp_sock *tp = tcp_sk(sk);
5970 struct tcp_fastopen_cookie foc = { .len = -1 };
5971 int saved_clamp = tp->rx_opt.mss_clamp;
5974 tcp_parse_options(sock_net(sk), skb, &tp->rx_opt, 0, &foc);
5975 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
5976 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
5980 * "If the state is SYN-SENT then
5981 * first check the ACK bit
5982 * If the ACK bit is set
5983 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5984 * a reset (unless the RST bit is set, if so drop
5985 * the segment and return)"
5987 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5988 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5989 /* Previous FIN/ACK or RST/ACK might be ignored. */
5990 if (icsk->icsk_retransmits == 0)
5991 inet_csk_reset_xmit_timer(sk,
5993 TCP_TIMEOUT_MIN, TCP_RTO_MAX);
5994 goto reset_and_undo;
5997 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5998 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5999 tcp_time_stamp(tp))) {
6000 NET_INC_STATS(sock_net(sk),
6001 LINUX_MIB_PAWSACTIVEREJECTED);
6002 goto reset_and_undo;
6005 /* Now ACK is acceptable.
6007 * "If the RST bit is set
6008 * If the ACK was acceptable then signal the user "error:
6009 * connection reset", drop the segment, enter CLOSED state,
6010 * delete TCB, and return."
6019 * "fifth, if neither of the SYN or RST bits is set then
6020 * drop the segment and return."
6026 goto discard_and_undo;
6029 * "If the SYN bit is on ...
6030 * are acceptable then ...
6031 * (our SYN has been ACKed), change the connection
6032 * state to ESTABLISHED..."
6035 tcp_ecn_rcv_synack(tp, th);
6037 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6038 tcp_try_undo_spurious_syn(sk);
6039 tcp_ack(sk, skb, FLAG_SLOWPATH);
6041 /* Ok.. it's good. Set up sequence numbers and
6042 * move to established.
6044 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1);
6045 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
6047 /* RFC1323: The window in SYN & SYN/ACK segments is
6050 tp->snd_wnd = ntohs(th->window);
6052 if (!tp->rx_opt.wscale_ok) {
6053 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
6054 tp->window_clamp = min(tp->window_clamp, 65535U);
6057 if (tp->rx_opt.saw_tstamp) {
6058 tp->rx_opt.tstamp_ok = 1;
6059 tp->tcp_header_len =
6060 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
6061 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6062 tcp_store_ts_recent(tp);
6064 tp->tcp_header_len = sizeof(struct tcphdr);
6067 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
6068 tcp_initialize_rcv_mss(sk);
6070 /* Remember, tcp_poll() does not lock socket!
6071 * Change state from SYN-SENT only after copied_seq
6072 * is initialized. */
6073 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6075 smc_check_reset_syn(tp);
6079 tcp_finish_connect(sk, skb);
6081 fastopen_fail = (tp->syn_fastopen || tp->syn_data) &&
6082 tcp_rcv_fastopen_synack(sk, skb, &foc);
6084 if (!sock_flag(sk, SOCK_DEAD)) {
6085 sk->sk_state_change(sk);
6086 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6090 if (sk->sk_write_pending ||
6091 icsk->icsk_accept_queue.rskq_defer_accept ||
6092 inet_csk_in_pingpong_mode(sk)) {
6093 /* Save one ACK. Data will be ready after
6094 * several ticks, if write_pending is set.
6096 * It may be deleted, but with this feature tcpdumps
6097 * look so _wonderfully_ clever, that I was not able
6098 * to stand against the temptation 8) --ANK
6100 inet_csk_schedule_ack(sk);
6101 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
6102 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
6103 TCP_DELACK_MAX, TCP_RTO_MAX);
6114 /* No ACK in the segment */
6118 * "If the RST bit is set
6120 * Otherwise (no ACK) drop the segment and return."
6123 goto discard_and_undo;
6127 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
6128 tcp_paws_reject(&tp->rx_opt, 0))
6129 goto discard_and_undo;
6132 /* We see SYN without ACK. It is attempt of
6133 * simultaneous connect with crossed SYNs.
6134 * Particularly, it can be connect to self.
6136 tcp_set_state(sk, TCP_SYN_RECV);
6138 if (tp->rx_opt.saw_tstamp) {
6139 tp->rx_opt.tstamp_ok = 1;
6140 tcp_store_ts_recent(tp);
6141 tp->tcp_header_len =
6142 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
6144 tp->tcp_header_len = sizeof(struct tcphdr);
6147 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1);
6148 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6149 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
6151 /* RFC1323: The window in SYN & SYN/ACK segments is
6154 tp->snd_wnd = ntohs(th->window);
6155 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
6156 tp->max_window = tp->snd_wnd;
6158 tcp_ecn_rcv_syn(tp, th);
6161 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
6162 tcp_initialize_rcv_mss(sk);
6164 tcp_send_synack(sk);
6166 /* Note, we could accept data and URG from this segment.
6167 * There are no obstacles to make this (except that we must
6168 * either change tcp_recvmsg() to prevent it from returning data
6169 * before 3WHS completes per RFC793, or employ TCP Fast Open).
6171 * However, if we ignore data in ACKless segments sometimes,
6172 * we have no reasons to accept it sometimes.
6173 * Also, seems the code doing it in step6 of tcp_rcv_state_process
6174 * is not flawless. So, discard packet for sanity.
6175 * Uncomment this return to process the data.
6182 /* "fifth, if neither of the SYN or RST bits is set then
6183 * drop the segment and return."
6187 tcp_clear_options(&tp->rx_opt);
6188 tp->rx_opt.mss_clamp = saved_clamp;
6192 tcp_clear_options(&tp->rx_opt);
6193 tp->rx_opt.mss_clamp = saved_clamp;
6197 static void tcp_rcv_synrecv_state_fastopen(struct sock *sk)
6199 struct request_sock *req;
6201 /* If we are still handling the SYNACK RTO, see if timestamp ECR allows
6202 * undo. If peer SACKs triggered fast recovery, we can't undo here.
6204 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
6205 tcp_try_undo_loss(sk, false);
6207 /* Reset rtx states to prevent spurious retransmits_timed_out() */
6208 tcp_sk(sk)->retrans_stamp = 0;
6209 inet_csk(sk)->icsk_retransmits = 0;
6211 /* Once we leave TCP_SYN_RECV or TCP_FIN_WAIT_1,
6212 * we no longer need req so release it.
6214 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk,
6215 lockdep_sock_is_held(sk));
6216 reqsk_fastopen_remove(sk, req, false);
6218 /* Re-arm the timer because data may have been sent out.
6219 * This is similar to the regular data transmission case
6220 * when new data has just been ack'ed.
6222 * (TFO) - we could try to be more aggressive and
6223 * retransmitting any data sooner based on when they
6230 * This function implements the receiving procedure of RFC 793 for
6231 * all states except ESTABLISHED and TIME_WAIT.
6232 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
6233 * address independent.
6236 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
6238 struct tcp_sock *tp = tcp_sk(sk);
6239 struct inet_connection_sock *icsk = inet_csk(sk);
6240 const struct tcphdr *th = tcp_hdr(skb);
6241 struct request_sock *req;
6245 switch (sk->sk_state) {
6259 /* It is possible that we process SYN packets from backlog,
6260 * so we need to make sure to disable BH and RCU right there.
6264 acceptable = icsk->icsk_af_ops->conn_request(sk, skb) >= 0;
6276 tp->rx_opt.saw_tstamp = 0;
6277 tcp_mstamp_refresh(tp);
6278 queued = tcp_rcv_synsent_state_process(sk, skb, th);
6282 /* Do step6 onward by hand. */
6283 tcp_urg(sk, skb, th);
6285 tcp_data_snd_check(sk);
6289 tcp_mstamp_refresh(tp);
6290 tp->rx_opt.saw_tstamp = 0;
6291 req = rcu_dereference_protected(tp->fastopen_rsk,
6292 lockdep_sock_is_held(sk));
6296 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
6297 sk->sk_state != TCP_FIN_WAIT1);
6299 if (!tcp_check_req(sk, skb, req, true, &req_stolen))
6303 if (!th->ack && !th->rst && !th->syn)
6306 if (!tcp_validate_incoming(sk, skb, th, 0))
6309 /* step 5: check the ACK field */
6310 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
6311 FLAG_UPDATE_TS_RECENT |
6312 FLAG_NO_CHALLENGE_ACK) > 0;
6315 if (sk->sk_state == TCP_SYN_RECV)
6316 return 1; /* send one RST */
6317 tcp_send_challenge_ack(sk, skb);
6320 switch (sk->sk_state) {
6322 tp->delivered++; /* SYN-ACK delivery isn't tracked in tcp_ack */
6324 tcp_synack_rtt_meas(sk, req);
6327 tcp_rcv_synrecv_state_fastopen(sk);
6329 tcp_try_undo_spurious_syn(sk);
6330 tp->retrans_stamp = 0;
6331 tcp_init_transfer(sk, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB);
6332 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6335 tcp_set_state(sk, TCP_ESTABLISHED);
6336 sk->sk_state_change(sk);
6338 /* Note, that this wakeup is only for marginal crossed SYN case.
6339 * Passively open sockets are not waked up, because
6340 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
6343 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6345 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
6346 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
6347 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6349 if (tp->rx_opt.tstamp_ok)
6350 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6352 if (!inet_csk(sk)->icsk_ca_ops->cong_control)
6353 tcp_update_pacing_rate(sk);
6355 /* Prevent spurious tcp_cwnd_restart() on first data packet */
6356 tp->lsndtime = tcp_jiffies32;
6358 tcp_initialize_rcv_mss(sk);
6359 tcp_fast_path_on(tp);
6362 case TCP_FIN_WAIT1: {
6366 tcp_rcv_synrecv_state_fastopen(sk);
6368 if (tp->snd_una != tp->write_seq)
6371 tcp_set_state(sk, TCP_FIN_WAIT2);
6372 sk->sk_shutdown |= SEND_SHUTDOWN;
6376 if (!sock_flag(sk, SOCK_DEAD)) {
6377 /* Wake up lingering close() */
6378 sk->sk_state_change(sk);
6382 if (tp->linger2 < 0) {
6384 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6387 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6388 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6389 /* Receive out of order FIN after close() */
6390 if (tp->syn_fastopen && th->fin)
6391 tcp_fastopen_active_disable(sk);
6393 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6397 tmo = tcp_fin_time(sk);
6398 if (tmo > TCP_TIMEWAIT_LEN) {
6399 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
6400 } else if (th->fin || sock_owned_by_user(sk)) {
6401 /* Bad case. We could lose such FIN otherwise.
6402 * It is not a big problem, but it looks confusing
6403 * and not so rare event. We still can lose it now,
6404 * if it spins in bh_lock_sock(), but it is really
6407 inet_csk_reset_keepalive_timer(sk, tmo);
6409 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
6416 if (tp->snd_una == tp->write_seq) {
6417 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6423 if (tp->snd_una == tp->write_seq) {
6424 tcp_update_metrics(sk);
6431 /* step 6: check the URG bit */
6432 tcp_urg(sk, skb, th);
6434 /* step 7: process the segment text */
6435 switch (sk->sk_state) {
6436 case TCP_CLOSE_WAIT:
6439 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
6440 if (sk_is_mptcp(sk))
6441 mptcp_incoming_options(sk, skb, &tp->rx_opt);
6447 /* RFC 793 says to queue data in these states,
6448 * RFC 1122 says we MUST send a reset.
6449 * BSD 4.4 also does reset.
6451 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6452 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6453 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6454 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6460 case TCP_ESTABLISHED:
6461 tcp_data_queue(sk, skb);
6466 /* tcp_data could move socket to TIME-WAIT */
6467 if (sk->sk_state != TCP_CLOSE) {
6468 tcp_data_snd_check(sk);
6469 tcp_ack_snd_check(sk);
6478 EXPORT_SYMBOL(tcp_rcv_state_process);
6480 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
6482 struct inet_request_sock *ireq = inet_rsk(req);
6484 if (family == AF_INET)
6485 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6486 &ireq->ir_rmt_addr, port);
6487 #if IS_ENABLED(CONFIG_IPV6)
6488 else if (family == AF_INET6)
6489 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6490 &ireq->ir_v6_rmt_addr, port);
6494 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6496 * If we receive a SYN packet with these bits set, it means a
6497 * network is playing bad games with TOS bits. In order to
6498 * avoid possible false congestion notifications, we disable
6499 * TCP ECN negotiation.
6501 * Exception: tcp_ca wants ECN. This is required for DCTCP
6502 * congestion control: Linux DCTCP asserts ECT on all packets,
6503 * including SYN, which is most optimal solution; however,
6504 * others, such as FreeBSD do not.
6506 * Exception: At least one of the reserved bits of the TCP header (th->res1) is
6507 * set, indicating the use of a future TCP extension (such as AccECN). See
6508 * RFC8311 §4.3 which updates RFC3168 to allow the development of such
6511 static void tcp_ecn_create_request(struct request_sock *req,
6512 const struct sk_buff *skb,
6513 const struct sock *listen_sk,
6514 const struct dst_entry *dst)
6516 const struct tcphdr *th = tcp_hdr(skb);
6517 const struct net *net = sock_net(listen_sk);
6518 bool th_ecn = th->ece && th->cwr;
6525 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
6526 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK);
6527 ecn_ok = net->ipv4.sysctl_tcp_ecn || ecn_ok_dst;
6529 if (((!ect || th->res1) && ecn_ok) || tcp_ca_needs_ecn(listen_sk) ||
6530 (ecn_ok_dst & DST_FEATURE_ECN_CA) ||
6531 tcp_bpf_ca_needs_ecn((struct sock *)req))
6532 inet_rsk(req)->ecn_ok = 1;
6535 static void tcp_openreq_init(struct request_sock *req,
6536 const struct tcp_options_received *rx_opt,
6537 struct sk_buff *skb, const struct sock *sk)
6539 struct inet_request_sock *ireq = inet_rsk(req);
6541 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */
6542 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6543 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
6544 tcp_rsk(req)->snt_synack = 0;
6545 tcp_rsk(req)->last_oow_ack_time = 0;
6546 req->mss = rx_opt->mss_clamp;
6547 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6548 ireq->tstamp_ok = rx_opt->tstamp_ok;
6549 ireq->sack_ok = rx_opt->sack_ok;
6550 ireq->snd_wscale = rx_opt->snd_wscale;
6551 ireq->wscale_ok = rx_opt->wscale_ok;
6554 ireq->ir_rmt_port = tcp_hdr(skb)->source;
6555 ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6556 ireq->ir_mark = inet_request_mark(sk, skb);
6557 #if IS_ENABLED(CONFIG_SMC)
6558 ireq->smc_ok = rx_opt->smc_ok;
6562 struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
6563 struct sock *sk_listener,
6564 bool attach_listener)
6566 struct request_sock *req = reqsk_alloc(ops, sk_listener,
6570 struct inet_request_sock *ireq = inet_rsk(req);
6572 ireq->ireq_opt = NULL;
6573 #if IS_ENABLED(CONFIG_IPV6)
6574 ireq->pktopts = NULL;
6576 atomic64_set(&ireq->ir_cookie, 0);
6577 ireq->ireq_state = TCP_NEW_SYN_RECV;
6578 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
6579 ireq->ireq_family = sk_listener->sk_family;
6584 EXPORT_SYMBOL(inet_reqsk_alloc);
6587 * Return true if a syncookie should be sent
6589 static bool tcp_syn_flood_action(const struct sock *sk, const char *proto)
6591 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
6592 const char *msg = "Dropping request";
6593 bool want_cookie = false;
6594 struct net *net = sock_net(sk);
6596 #ifdef CONFIG_SYN_COOKIES
6597 if (net->ipv4.sysctl_tcp_syncookies) {
6598 msg = "Sending cookies";
6600 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
6603 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
6605 if (!queue->synflood_warned &&
6606 net->ipv4.sysctl_tcp_syncookies != 2 &&
6607 xchg(&queue->synflood_warned, 1) == 0)
6608 net_info_ratelimited("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6609 proto, sk->sk_num, msg);
6614 static void tcp_reqsk_record_syn(const struct sock *sk,
6615 struct request_sock *req,
6616 const struct sk_buff *skb)
6618 if (tcp_sk(sk)->save_syn) {
6619 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
6622 copy = kmalloc(len + sizeof(u32), GFP_ATOMIC);
6625 memcpy(©[1], skb_network_header(skb), len);
6626 req->saved_syn = copy;
6631 /* If a SYN cookie is required and supported, returns a clamped MSS value to be
6632 * used for SYN cookie generation.
6634 u16 tcp_get_syncookie_mss(struct request_sock_ops *rsk_ops,
6635 const struct tcp_request_sock_ops *af_ops,
6636 struct sock *sk, struct tcphdr *th)
6638 struct tcp_sock *tp = tcp_sk(sk);
6641 if (sock_net(sk)->ipv4.sysctl_tcp_syncookies != 2 &&
6642 !inet_csk_reqsk_queue_is_full(sk))
6645 if (!tcp_syn_flood_action(sk, rsk_ops->slab_name))
6648 if (sk_acceptq_is_full(sk)) {
6649 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6653 mss = tcp_parse_mss_option(th, tp->rx_opt.user_mss);
6655 mss = af_ops->mss_clamp;
6659 EXPORT_SYMBOL_GPL(tcp_get_syncookie_mss);
6661 int tcp_conn_request(struct request_sock_ops *rsk_ops,
6662 const struct tcp_request_sock_ops *af_ops,
6663 struct sock *sk, struct sk_buff *skb)
6665 struct tcp_fastopen_cookie foc = { .len = -1 };
6666 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
6667 struct tcp_options_received tmp_opt;
6668 struct tcp_sock *tp = tcp_sk(sk);
6669 struct net *net = sock_net(sk);
6670 struct sock *fastopen_sk = NULL;
6671 struct request_sock *req;
6672 bool want_cookie = false;
6673 struct dst_entry *dst;
6676 /* TW buckets are converted to open requests without
6677 * limitations, they conserve resources and peer is
6678 * evidently real one.
6680 if ((net->ipv4.sysctl_tcp_syncookies == 2 ||
6681 inet_csk_reqsk_queue_is_full(sk)) && !isn) {
6682 want_cookie = tcp_syn_flood_action(sk, rsk_ops->slab_name);
6687 if (sk_acceptq_is_full(sk)) {
6688 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6692 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie);
6696 req->syncookie = want_cookie;
6697 tcp_rsk(req)->af_specific = af_ops;
6698 tcp_rsk(req)->ts_off = 0;
6699 #if IS_ENABLED(CONFIG_MPTCP)
6700 tcp_rsk(req)->is_mptcp = 0;
6703 tcp_clear_options(&tmp_opt);
6704 tmp_opt.mss_clamp = af_ops->mss_clamp;
6705 tmp_opt.user_mss = tp->rx_opt.user_mss;
6706 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0,
6707 want_cookie ? NULL : &foc);
6709 if (want_cookie && !tmp_opt.saw_tstamp)
6710 tcp_clear_options(&tmp_opt);
6712 if (IS_ENABLED(CONFIG_SMC) && want_cookie)
6715 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
6716 tcp_openreq_init(req, &tmp_opt, skb, sk);
6717 inet_rsk(req)->no_srccheck = inet_sk(sk)->transparent;
6719 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6720 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb);
6722 af_ops->init_req(req, sk, skb);
6724 if (security_inet_conn_request(sk, skb, req))
6727 if (tmp_opt.tstamp_ok)
6728 tcp_rsk(req)->ts_off = af_ops->init_ts_off(net, skb);
6730 dst = af_ops->route_req(sk, &fl, req);
6734 if (!want_cookie && !isn) {
6735 /* Kill the following clause, if you dislike this way. */
6736 if (!net->ipv4.sysctl_tcp_syncookies &&
6737 (net->ipv4.sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
6738 (net->ipv4.sysctl_max_syn_backlog >> 2)) &&
6739 !tcp_peer_is_proven(req, dst)) {
6740 /* Without syncookies last quarter of
6741 * backlog is filled with destinations,
6742 * proven to be alive.
6743 * It means that we continue to communicate
6744 * to destinations, already remembered
6745 * to the moment of synflood.
6747 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
6749 goto drop_and_release;
6752 isn = af_ops->init_seq(skb);
6755 tcp_ecn_create_request(req, skb, sk, dst);
6758 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
6759 if (!tmp_opt.tstamp_ok)
6760 inet_rsk(req)->ecn_ok = 0;
6763 tcp_rsk(req)->snt_isn = isn;
6764 tcp_rsk(req)->txhash = net_tx_rndhash();
6765 tcp_openreq_init_rwin(req, sk, dst);
6766 sk_rx_queue_set(req_to_sk(req), skb);
6768 tcp_reqsk_record_syn(sk, req, skb);
6769 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst);
6772 af_ops->send_synack(fastopen_sk, dst, &fl, req,
6773 &foc, TCP_SYNACK_FASTOPEN);
6774 /* Add the child socket directly into the accept queue */
6775 if (!inet_csk_reqsk_queue_add(sk, req, fastopen_sk)) {
6776 reqsk_fastopen_remove(fastopen_sk, req, false);
6777 bh_unlock_sock(fastopen_sk);
6778 sock_put(fastopen_sk);
6781 sk->sk_data_ready(sk);
6782 bh_unlock_sock(fastopen_sk);
6783 sock_put(fastopen_sk);
6785 tcp_rsk(req)->tfo_listener = false;
6787 inet_csk_reqsk_queue_hash_add(sk, req,
6788 tcp_timeout_init((struct sock *)req));
6789 af_ops->send_synack(sk, dst, &fl, req, &foc,
6790 !want_cookie ? TCP_SYNACK_NORMAL :
6808 EXPORT_SYMBOL(tcp_conn_request);