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/static_key.h>
82 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
84 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
85 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
86 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
87 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
88 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
89 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
90 #define FLAG_ECE 0x40 /* ECE in this ACK */
91 #define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */
92 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
93 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
94 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
95 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
96 #define FLAG_SET_XMIT_TIMER 0x1000 /* Set TLP or RTO timer */
97 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
98 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
99 #define FLAG_NO_CHALLENGE_ACK 0x8000 /* do not call tcp_send_challenge_ack() */
100 #define FLAG_ACK_MAYBE_DELAYED 0x10000 /* Likely a delayed ACK */
102 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
103 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
104 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE|FLAG_DSACKING_ACK)
105 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
107 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
108 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
110 #define REXMIT_NONE 0 /* no loss recovery to do */
111 #define REXMIT_LOST 1 /* retransmit packets marked lost */
112 #define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */
114 #if IS_ENABLED(CONFIG_TLS_DEVICE)
115 static DEFINE_STATIC_KEY_FALSE(clean_acked_data_enabled);
117 void clean_acked_data_enable(struct inet_connection_sock *icsk,
118 void (*cad)(struct sock *sk, u32 ack_seq))
120 icsk->icsk_clean_acked = cad;
121 static_branch_inc(&clean_acked_data_enabled);
123 EXPORT_SYMBOL_GPL(clean_acked_data_enable);
125 void clean_acked_data_disable(struct inet_connection_sock *icsk)
127 static_branch_dec(&clean_acked_data_enabled);
128 icsk->icsk_clean_acked = NULL;
130 EXPORT_SYMBOL_GPL(clean_acked_data_disable);
133 static void tcp_gro_dev_warn(struct sock *sk, const struct sk_buff *skb,
136 static bool __once __read_mostly;
139 struct net_device *dev;
144 dev = dev_get_by_index_rcu(sock_net(sk), skb->skb_iif);
145 if (!dev || len >= dev->mtu)
146 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n",
147 dev ? dev->name : "Unknown driver");
152 /* Adapt the MSS value used to make delayed ack decision to the
155 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
157 struct inet_connection_sock *icsk = inet_csk(sk);
158 const unsigned int lss = icsk->icsk_ack.last_seg_size;
161 icsk->icsk_ack.last_seg_size = 0;
163 /* skb->len may jitter because of SACKs, even if peer
164 * sends good full-sized frames.
166 len = skb_shinfo(skb)->gso_size ? : skb->len;
167 if (len >= icsk->icsk_ack.rcv_mss) {
168 icsk->icsk_ack.rcv_mss = min_t(unsigned int, len,
170 /* Account for possibly-removed options */
171 if (unlikely(len > icsk->icsk_ack.rcv_mss +
172 MAX_TCP_OPTION_SPACE))
173 tcp_gro_dev_warn(sk, skb, len);
175 /* Otherwise, we make more careful check taking into account,
176 * that SACKs block is variable.
178 * "len" is invariant segment length, including TCP header.
180 len += skb->data - skb_transport_header(skb);
181 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
182 /* If PSH is not set, packet should be
183 * full sized, provided peer TCP is not badly broken.
184 * This observation (if it is correct 8)) allows
185 * to handle super-low mtu links fairly.
187 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
188 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
189 /* Subtract also invariant (if peer is RFC compliant),
190 * tcp header plus fixed timestamp option length.
191 * Resulting "len" is MSS free of SACK jitter.
193 len -= tcp_sk(sk)->tcp_header_len;
194 icsk->icsk_ack.last_seg_size = len;
196 icsk->icsk_ack.rcv_mss = len;
200 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
201 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
202 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
206 static void tcp_incr_quickack(struct sock *sk, unsigned int max_quickacks)
208 struct inet_connection_sock *icsk = inet_csk(sk);
209 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
213 quickacks = min(quickacks, max_quickacks);
214 if (quickacks > icsk->icsk_ack.quick)
215 icsk->icsk_ack.quick = quickacks;
218 void tcp_enter_quickack_mode(struct sock *sk, unsigned int max_quickacks)
220 struct inet_connection_sock *icsk = inet_csk(sk);
222 tcp_incr_quickack(sk, max_quickacks);
223 icsk->icsk_ack.pingpong = 0;
224 icsk->icsk_ack.ato = TCP_ATO_MIN;
226 EXPORT_SYMBOL(tcp_enter_quickack_mode);
228 /* Send ACKs quickly, if "quick" count is not exhausted
229 * and the session is not interactive.
232 static bool tcp_in_quickack_mode(struct sock *sk)
234 const struct inet_connection_sock *icsk = inet_csk(sk);
235 const struct dst_entry *dst = __sk_dst_get(sk);
237 return (dst && dst_metric(dst, RTAX_QUICKACK)) ||
238 (icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong);
241 static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
243 if (tp->ecn_flags & TCP_ECN_OK)
244 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
247 static void tcp_ecn_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
249 if (tcp_hdr(skb)->cwr)
250 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
253 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
255 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
258 static void __tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
260 struct tcp_sock *tp = tcp_sk(sk);
262 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
263 case INET_ECN_NOT_ECT:
264 /* Funny extension: if ECT is not set on a segment,
265 * and we already seen ECT on a previous segment,
266 * it is probably a retransmit.
268 if (tp->ecn_flags & TCP_ECN_SEEN)
269 tcp_enter_quickack_mode(sk, 2);
272 if (tcp_ca_needs_ecn(sk))
273 tcp_ca_event(sk, CA_EVENT_ECN_IS_CE);
275 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
276 /* Better not delay acks, sender can have a very low cwnd */
277 tcp_enter_quickack_mode(sk, 2);
278 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
280 tp->ecn_flags |= TCP_ECN_SEEN;
283 if (tcp_ca_needs_ecn(sk))
284 tcp_ca_event(sk, CA_EVENT_ECN_NO_CE);
285 tp->ecn_flags |= TCP_ECN_SEEN;
290 static void tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
292 if (tcp_sk(sk)->ecn_flags & TCP_ECN_OK)
293 __tcp_ecn_check_ce(sk, skb);
296 static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
298 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
299 tp->ecn_flags &= ~TCP_ECN_OK;
302 static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
304 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
305 tp->ecn_flags &= ~TCP_ECN_OK;
308 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
310 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
315 /* Buffer size and advertised window tuning.
317 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
320 static void tcp_sndbuf_expand(struct sock *sk)
322 const struct tcp_sock *tp = tcp_sk(sk);
323 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
327 /* Worst case is non GSO/TSO : each frame consumes one skb
328 * and skb->head is kmalloced using power of two area of memory
330 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
332 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
334 per_mss = roundup_pow_of_two(per_mss) +
335 SKB_DATA_ALIGN(sizeof(struct sk_buff));
337 nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd);
338 nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
340 /* Fast Recovery (RFC 5681 3.2) :
341 * Cubic needs 1.7 factor, rounded to 2 to include
342 * extra cushion (application might react slowly to EPOLLOUT)
344 sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2;
345 sndmem *= nr_segs * per_mss;
347 if (sk->sk_sndbuf < sndmem)
348 sk->sk_sndbuf = min(sndmem, sock_net(sk)->ipv4.sysctl_tcp_wmem[2]);
351 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
353 * All tcp_full_space() is split to two parts: "network" buffer, allocated
354 * forward and advertised in receiver window (tp->rcv_wnd) and
355 * "application buffer", required to isolate scheduling/application
356 * latencies from network.
357 * window_clamp is maximal advertised window. It can be less than
358 * tcp_full_space(), in this case tcp_full_space() - window_clamp
359 * is reserved for "application" buffer. The less window_clamp is
360 * the smoother our behaviour from viewpoint of network, but the lower
361 * throughput and the higher sensitivity of the connection to losses. 8)
363 * rcv_ssthresh is more strict window_clamp used at "slow start"
364 * phase to predict further behaviour of this connection.
365 * It is used for two goals:
366 * - to enforce header prediction at sender, even when application
367 * requires some significant "application buffer". It is check #1.
368 * - to prevent pruning of receive queue because of misprediction
369 * of receiver window. Check #2.
371 * The scheme does not work when sender sends good segments opening
372 * window and then starts to feed us spaghetti. But it should work
373 * in common situations. Otherwise, we have to rely on queue collapsing.
376 /* Slow part of check#2. */
377 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
379 struct tcp_sock *tp = tcp_sk(sk);
381 int truesize = tcp_win_from_space(sk, skb->truesize) >> 1;
382 int window = tcp_win_from_space(sk, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]) >> 1;
384 while (tp->rcv_ssthresh <= window) {
385 if (truesize <= skb->len)
386 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
394 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
396 struct tcp_sock *tp = tcp_sk(sk);
399 if (tp->rcv_ssthresh < tp->window_clamp &&
400 (int)tp->rcv_ssthresh < tcp_space(sk) &&
401 !tcp_under_memory_pressure(sk)) {
404 /* Check #2. Increase window, if skb with such overhead
405 * will fit to rcvbuf in future.
407 if (tcp_win_from_space(sk, skb->truesize) <= skb->len)
408 incr = 2 * tp->advmss;
410 incr = __tcp_grow_window(sk, skb);
413 incr = max_t(int, incr, 2 * skb->len);
414 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
416 inet_csk(sk)->icsk_ack.quick |= 1;
421 /* 3. Tuning rcvbuf, when connection enters established state. */
422 static void tcp_fixup_rcvbuf(struct sock *sk)
424 u32 mss = tcp_sk(sk)->advmss;
427 rcvmem = 2 * SKB_TRUESIZE(mss + MAX_TCP_HEADER) *
428 tcp_default_init_rwnd(mss);
430 /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
431 * Allow enough cushion so that sender is not limited by our window
433 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf)
436 if (sk->sk_rcvbuf < rcvmem)
437 sk->sk_rcvbuf = min(rcvmem, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
440 /* 4. Try to fixup all. It is made immediately after connection enters
443 void tcp_init_buffer_space(struct sock *sk)
445 int tcp_app_win = sock_net(sk)->ipv4.sysctl_tcp_app_win;
446 struct tcp_sock *tp = tcp_sk(sk);
449 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
450 tcp_fixup_rcvbuf(sk);
451 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
452 tcp_sndbuf_expand(sk);
454 tp->rcvq_space.space = tp->rcv_wnd;
455 tcp_mstamp_refresh(tp);
456 tp->rcvq_space.time = tp->tcp_mstamp;
457 tp->rcvq_space.seq = tp->copied_seq;
459 maxwin = tcp_full_space(sk);
461 if (tp->window_clamp >= maxwin) {
462 tp->window_clamp = maxwin;
464 if (tcp_app_win && maxwin > 4 * tp->advmss)
465 tp->window_clamp = max(maxwin -
466 (maxwin >> tcp_app_win),
470 /* Force reservation of one segment. */
472 tp->window_clamp > 2 * tp->advmss &&
473 tp->window_clamp + tp->advmss > maxwin)
474 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
476 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
477 tp->snd_cwnd_stamp = tcp_jiffies32;
480 /* 5. Recalculate window clamp after socket hit its memory bounds. */
481 static void tcp_clamp_window(struct sock *sk)
483 struct tcp_sock *tp = tcp_sk(sk);
484 struct inet_connection_sock *icsk = inet_csk(sk);
485 struct net *net = sock_net(sk);
487 icsk->icsk_ack.quick = 0;
489 if (sk->sk_rcvbuf < net->ipv4.sysctl_tcp_rmem[2] &&
490 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
491 !tcp_under_memory_pressure(sk) &&
492 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
493 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
494 net->ipv4.sysctl_tcp_rmem[2]);
496 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
497 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
500 /* Initialize RCV_MSS value.
501 * RCV_MSS is an our guess about MSS used by the peer.
502 * We haven't any direct information about the MSS.
503 * It's better to underestimate the RCV_MSS rather than overestimate.
504 * Overestimations make us ACKing less frequently than needed.
505 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
507 void tcp_initialize_rcv_mss(struct sock *sk)
509 const struct tcp_sock *tp = tcp_sk(sk);
510 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
512 hint = min(hint, tp->rcv_wnd / 2);
513 hint = min(hint, TCP_MSS_DEFAULT);
514 hint = max(hint, TCP_MIN_MSS);
516 inet_csk(sk)->icsk_ack.rcv_mss = hint;
518 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
520 /* Receiver "autotuning" code.
522 * The algorithm for RTT estimation w/o timestamps is based on
523 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
524 * <http://public.lanl.gov/radiant/pubs.html#DRS>
526 * More detail on this code can be found at
527 * <http://staff.psc.edu/jheffner/>,
528 * though this reference is out of date. A new paper
531 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
533 u32 new_sample = tp->rcv_rtt_est.rtt_us;
536 if (new_sample != 0) {
537 /* If we sample in larger samples in the non-timestamp
538 * case, we could grossly overestimate the RTT especially
539 * with chatty applications or bulk transfer apps which
540 * are stalled on filesystem I/O.
542 * Also, since we are only going for a minimum in the
543 * non-timestamp case, we do not smooth things out
544 * else with timestamps disabled convergence takes too
548 m -= (new_sample >> 3);
556 /* No previous measure. */
560 tp->rcv_rtt_est.rtt_us = new_sample;
563 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
567 if (tp->rcv_rtt_est.time == 0)
569 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
571 delta_us = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcv_rtt_est.time);
574 tcp_rcv_rtt_update(tp, delta_us, 1);
577 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
578 tp->rcv_rtt_est.time = tp->tcp_mstamp;
581 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
582 const struct sk_buff *skb)
584 struct tcp_sock *tp = tcp_sk(sk);
586 if (tp->rx_opt.rcv_tsecr &&
587 (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;
594 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
595 tcp_rcv_rtt_update(tp, delta_us, 0);
600 * This function should be called every time data is copied to user space.
601 * It calculates the appropriate TCP receive buffer space.
603 void tcp_rcv_space_adjust(struct sock *sk)
605 struct tcp_sock *tp = tcp_sk(sk);
609 trace_tcp_rcv_space_adjust(sk);
611 tcp_mstamp_refresh(tp);
612 time = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcvq_space.time);
613 if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0)
616 /* Number of bytes copied to user in last RTT */
617 copied = tp->copied_seq - tp->rcvq_space.seq;
618 if (copied <= tp->rcvq_space.space)
622 * copied = bytes received in previous RTT, our base window
623 * To cope with packet losses, we need a 2x factor
624 * To cope with slow start, and sender growing its cwin by 100 %
625 * every RTT, we need a 4x factor, because the ACK we are sending
626 * now is for the next RTT, not the current one :
627 * <prev RTT . ><current RTT .. ><next RTT .... >
630 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf &&
631 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
635 /* minimal window to cope with packet losses, assuming
636 * steady state. Add some cushion because of small variations.
638 rcvwin = ((u64)copied << 1) + 16 * tp->advmss;
640 /* Accommodate for sender rate increase (eg. slow start) */
641 grow = rcvwin * (copied - tp->rcvq_space.space);
642 do_div(grow, tp->rcvq_space.space);
643 rcvwin += (grow << 1);
645 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
646 while (tcp_win_from_space(sk, rcvmem) < tp->advmss)
649 do_div(rcvwin, tp->advmss);
650 rcvbuf = min_t(u64, rcvwin * rcvmem,
651 sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
652 if (rcvbuf > sk->sk_rcvbuf) {
653 sk->sk_rcvbuf = rcvbuf;
655 /* Make the window clamp follow along. */
656 tp->window_clamp = tcp_win_from_space(sk, rcvbuf);
659 tp->rcvq_space.space = copied;
662 tp->rcvq_space.seq = tp->copied_seq;
663 tp->rcvq_space.time = tp->tcp_mstamp;
666 /* There is something which you must keep in mind when you analyze the
667 * behavior of the tp->ato delayed ack timeout interval. When a
668 * connection starts up, we want to ack as quickly as possible. The
669 * problem is that "good" TCP's do slow start at the beginning of data
670 * transmission. The means that until we send the first few ACK's the
671 * sender will sit on his end and only queue most of his data, because
672 * he can only send snd_cwnd unacked packets at any given time. For
673 * each ACK we send, he increments snd_cwnd and transmits more of his
676 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
678 struct tcp_sock *tp = tcp_sk(sk);
679 struct inet_connection_sock *icsk = inet_csk(sk);
682 inet_csk_schedule_ack(sk);
684 tcp_measure_rcv_mss(sk, skb);
686 tcp_rcv_rtt_measure(tp);
690 if (!icsk->icsk_ack.ato) {
691 /* The _first_ data packet received, initialize
692 * delayed ACK engine.
694 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
695 icsk->icsk_ack.ato = TCP_ATO_MIN;
697 int m = now - icsk->icsk_ack.lrcvtime;
699 if (m <= TCP_ATO_MIN / 2) {
700 /* The fastest case is the first. */
701 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
702 } else if (m < icsk->icsk_ack.ato) {
703 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
704 if (icsk->icsk_ack.ato > icsk->icsk_rto)
705 icsk->icsk_ack.ato = icsk->icsk_rto;
706 } else if (m > icsk->icsk_rto) {
707 /* Too long gap. Apparently sender failed to
708 * restart window, so that we send ACKs quickly.
710 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
714 icsk->icsk_ack.lrcvtime = now;
716 tcp_ecn_check_ce(sk, skb);
719 tcp_grow_window(sk, skb);
722 /* Called to compute a smoothed rtt estimate. The data fed to this
723 * routine either comes from timestamps, or from segments that were
724 * known _not_ to have been retransmitted [see Karn/Partridge
725 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
726 * piece by Van Jacobson.
727 * NOTE: the next three routines used to be one big routine.
728 * To save cycles in the RFC 1323 implementation it was better to break
729 * it up into three procedures. -- erics
731 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
733 struct tcp_sock *tp = tcp_sk(sk);
734 long m = mrtt_us; /* RTT */
735 u32 srtt = tp->srtt_us;
737 /* The following amusing code comes from Jacobson's
738 * article in SIGCOMM '88. Note that rtt and mdev
739 * are scaled versions of rtt and mean deviation.
740 * This is designed to be as fast as possible
741 * m stands for "measurement".
743 * On a 1990 paper the rto value is changed to:
744 * RTO = rtt + 4 * mdev
746 * Funny. This algorithm seems to be very broken.
747 * These formulae increase RTO, when it should be decreased, increase
748 * too slowly, when it should be increased quickly, decrease too quickly
749 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
750 * does not matter how to _calculate_ it. Seems, it was trap
751 * that VJ failed to avoid. 8)
754 m -= (srtt >> 3); /* m is now error in rtt est */
755 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
757 m = -m; /* m is now abs(error) */
758 m -= (tp->mdev_us >> 2); /* similar update on mdev */
759 /* This is similar to one of Eifel findings.
760 * Eifel blocks mdev updates when rtt decreases.
761 * This solution is a bit different: we use finer gain
762 * for mdev in this case (alpha*beta).
763 * Like Eifel it also prevents growth of rto,
764 * but also it limits too fast rto decreases,
765 * happening in pure Eifel.
770 m -= (tp->mdev_us >> 2); /* similar update on mdev */
772 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
773 if (tp->mdev_us > tp->mdev_max_us) {
774 tp->mdev_max_us = tp->mdev_us;
775 if (tp->mdev_max_us > tp->rttvar_us)
776 tp->rttvar_us = tp->mdev_max_us;
778 if (after(tp->snd_una, tp->rtt_seq)) {
779 if (tp->mdev_max_us < tp->rttvar_us)
780 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
781 tp->rtt_seq = tp->snd_nxt;
782 tp->mdev_max_us = tcp_rto_min_us(sk);
785 /* no previous measure. */
786 srtt = m << 3; /* take the measured time to be rtt */
787 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
788 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
789 tp->mdev_max_us = tp->rttvar_us;
790 tp->rtt_seq = tp->snd_nxt;
792 tp->srtt_us = max(1U, srtt);
795 static void tcp_update_pacing_rate(struct sock *sk)
797 const struct tcp_sock *tp = tcp_sk(sk);
800 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
801 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
803 /* current rate is (cwnd * mss) / srtt
804 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
805 * In Congestion Avoidance phase, set it to 120 % the current rate.
807 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
808 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
809 * end of slow start and should slow down.
811 if (tp->snd_cwnd < tp->snd_ssthresh / 2)
812 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ss_ratio;
814 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ca_ratio;
816 rate *= max(tp->snd_cwnd, tp->packets_out);
818 if (likely(tp->srtt_us))
819 do_div(rate, tp->srtt_us);
821 /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate
822 * without any lock. We want to make sure compiler wont store
823 * intermediate values in this location.
825 WRITE_ONCE(sk->sk_pacing_rate, min_t(u64, rate,
826 sk->sk_max_pacing_rate));
829 /* Calculate rto without backoff. This is the second half of Van Jacobson's
830 * routine referred to above.
832 static void tcp_set_rto(struct sock *sk)
834 const struct tcp_sock *tp = tcp_sk(sk);
835 /* Old crap is replaced with new one. 8)
838 * 1. If rtt variance happened to be less 50msec, it is hallucination.
839 * It cannot be less due to utterly erratic ACK generation made
840 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
841 * to do with delayed acks, because at cwnd>2 true delack timeout
842 * is invisible. Actually, Linux-2.4 also generates erratic
843 * ACKs in some circumstances.
845 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
847 /* 2. Fixups made earlier cannot be right.
848 * If we do not estimate RTO correctly without them,
849 * all the algo is pure shit and should be replaced
850 * with correct one. It is exactly, which we pretend to do.
853 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
854 * guarantees that rto is higher.
859 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
861 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
864 cwnd = TCP_INIT_CWND;
865 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
868 /* Take a notice that peer is sending D-SACKs */
869 static void tcp_dsack_seen(struct tcp_sock *tp)
871 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
872 tp->rack.dsack_seen = 1;
875 /* It's reordering when higher sequence was delivered (i.e. sacked) before
876 * some lower never-retransmitted sequence ("low_seq"). The maximum reordering
877 * distance is approximated in full-mss packet distance ("reordering").
879 static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq,
882 struct tcp_sock *tp = tcp_sk(sk);
883 const u32 mss = tp->mss_cache;
886 fack = tcp_highest_sack_seq(tp);
887 if (!before(low_seq, fack))
890 metric = fack - low_seq;
891 if ((metric > tp->reordering * mss) && mss) {
892 #if FASTRETRANS_DEBUG > 1
893 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
894 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
898 tp->undo_marker ? tp->undo_retrans : 0);
900 tp->reordering = min_t(u32, (metric + mss - 1) / mss,
901 sock_net(sk)->ipv4.sysctl_tcp_max_reordering);
905 /* This exciting event is worth to be remembered. 8) */
906 NET_INC_STATS(sock_net(sk),
907 ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER);
910 /* This must be called before lost_out is incremented */
911 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
913 if (!tp->retransmit_skb_hint ||
914 before(TCP_SKB_CB(skb)->seq,
915 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
916 tp->retransmit_skb_hint = skb;
919 /* Sum the number of packets on the wire we have marked as lost.
920 * There are two cases we care about here:
921 * a) Packet hasn't been marked lost (nor retransmitted),
922 * and this is the first loss.
923 * b) Packet has been marked both lost and retransmitted,
924 * and this means we think it was lost again.
926 static void tcp_sum_lost(struct tcp_sock *tp, struct sk_buff *skb)
928 __u8 sacked = TCP_SKB_CB(skb)->sacked;
930 if (!(sacked & TCPCB_LOST) ||
931 ((sacked & TCPCB_LOST) && (sacked & TCPCB_SACKED_RETRANS)))
932 tp->lost += tcp_skb_pcount(skb);
935 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
937 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
938 tcp_verify_retransmit_hint(tp, skb);
940 tp->lost_out += tcp_skb_pcount(skb);
941 tcp_sum_lost(tp, skb);
942 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
946 void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, struct sk_buff *skb)
948 tcp_verify_retransmit_hint(tp, skb);
950 tcp_sum_lost(tp, skb);
951 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
952 tp->lost_out += tcp_skb_pcount(skb);
953 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
957 /* This procedure tags the retransmission queue when SACKs arrive.
959 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
960 * Packets in queue with these bits set are counted in variables
961 * sacked_out, retrans_out and lost_out, correspondingly.
963 * Valid combinations are:
964 * Tag InFlight Description
965 * 0 1 - orig segment is in flight.
966 * S 0 - nothing flies, orig reached receiver.
967 * L 0 - nothing flies, orig lost by net.
968 * R 2 - both orig and retransmit are in flight.
969 * L|R 1 - orig is lost, retransmit is in flight.
970 * S|R 1 - orig reached receiver, retrans is still in flight.
971 * (L|S|R is logically valid, it could occur when L|R is sacked,
972 * but it is equivalent to plain S and code short-curcuits it to S.
973 * L|S is logically invalid, it would mean -1 packet in flight 8))
975 * These 6 states form finite state machine, controlled by the following events:
976 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
977 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
978 * 3. Loss detection event of two flavors:
979 * A. Scoreboard estimator decided the packet is lost.
980 * A'. Reno "three dupacks" marks head of queue lost.
981 * B. SACK arrives sacking SND.NXT at the moment, when the
982 * segment was retransmitted.
983 * 4. D-SACK added new rule: D-SACK changes any tag to S.
985 * It is pleasant to note, that state diagram turns out to be commutative,
986 * so that we are allowed not to be bothered by order of our actions,
987 * when multiple events arrive simultaneously. (see the function below).
989 * Reordering detection.
990 * --------------------
991 * Reordering metric is maximal distance, which a packet can be displaced
992 * in packet stream. With SACKs we can estimate it:
994 * 1. SACK fills old hole and the corresponding segment was not
995 * ever retransmitted -> reordering. Alas, we cannot use it
996 * when segment was retransmitted.
997 * 2. The last flaw is solved with D-SACK. D-SACK arrives
998 * for retransmitted and already SACKed segment -> reordering..
999 * Both of these heuristics are not used in Loss state, when we cannot
1000 * account for retransmits accurately.
1002 * SACK block validation.
1003 * ----------------------
1005 * SACK block range validation checks that the received SACK block fits to
1006 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1007 * Note that SND.UNA is not included to the range though being valid because
1008 * it means that the receiver is rather inconsistent with itself reporting
1009 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1010 * perfectly valid, however, in light of RFC2018 which explicitly states
1011 * that "SACK block MUST reflect the newest segment. Even if the newest
1012 * segment is going to be discarded ...", not that it looks very clever
1013 * in case of head skb. Due to potentional receiver driven attacks, we
1014 * choose to avoid immediate execution of a walk in write queue due to
1015 * reneging and defer head skb's loss recovery to standard loss recovery
1016 * procedure that will eventually trigger (nothing forbids us doing this).
1018 * Implements also blockage to start_seq wrap-around. Problem lies in the
1019 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1020 * there's no guarantee that it will be before snd_nxt (n). The problem
1021 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1024 * <- outs wnd -> <- wrapzone ->
1025 * u e n u_w e_w s n_w
1027 * |<------------+------+----- TCP seqno space --------------+---------->|
1028 * ...-- <2^31 ->| |<--------...
1029 * ...---- >2^31 ------>| |<--------...
1031 * Current code wouldn't be vulnerable but it's better still to discard such
1032 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1033 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1034 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1035 * equal to the ideal case (infinite seqno space without wrap caused issues).
1037 * With D-SACK the lower bound is extended to cover sequence space below
1038 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1039 * again, D-SACK block must not to go across snd_una (for the same reason as
1040 * for the normal SACK blocks, explained above). But there all simplicity
1041 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1042 * fully below undo_marker they do not affect behavior in anyway and can
1043 * therefore be safely ignored. In rare cases (which are more or less
1044 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1045 * fragmentation and packet reordering past skb's retransmission. To consider
1046 * them correctly, the acceptable range must be extended even more though
1047 * the exact amount is rather hard to quantify. However, tp->max_window can
1048 * be used as an exaggerated estimate.
1050 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1051 u32 start_seq, u32 end_seq)
1053 /* Too far in future, or reversed (interpretation is ambiguous) */
1054 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1057 /* Nasty start_seq wrap-around check (see comments above) */
1058 if (!before(start_seq, tp->snd_nxt))
1061 /* In outstanding window? ...This is valid exit for D-SACKs too.
1062 * start_seq == snd_una is non-sensical (see comments above)
1064 if (after(start_seq, tp->snd_una))
1067 if (!is_dsack || !tp->undo_marker)
1070 /* ...Then it's D-SACK, and must reside below snd_una completely */
1071 if (after(end_seq, tp->snd_una))
1074 if (!before(start_seq, tp->undo_marker))
1078 if (!after(end_seq, tp->undo_marker))
1081 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1082 * start_seq < undo_marker and end_seq >= undo_marker.
1084 return !before(start_seq, end_seq - tp->max_window);
1087 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1088 struct tcp_sack_block_wire *sp, int num_sacks,
1091 struct tcp_sock *tp = tcp_sk(sk);
1092 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1093 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1094 bool dup_sack = false;
1096 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1099 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1100 } else if (num_sacks > 1) {
1101 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1102 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1104 if (!after(end_seq_0, end_seq_1) &&
1105 !before(start_seq_0, start_seq_1)) {
1108 NET_INC_STATS(sock_net(sk),
1109 LINUX_MIB_TCPDSACKOFORECV);
1113 /* D-SACK for already forgotten data... Do dumb counting. */
1114 if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 &&
1115 !after(end_seq_0, prior_snd_una) &&
1116 after(end_seq_0, tp->undo_marker))
1122 struct tcp_sacktag_state {
1124 /* Timestamps for earliest and latest never-retransmitted segment
1125 * that was SACKed. RTO needs the earliest RTT to stay conservative,
1126 * but congestion control should still get an accurate delay signal.
1130 struct rate_sample *rate;
1132 unsigned int mss_now;
1135 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1136 * the incoming SACK may not exactly match but we can find smaller MSS
1137 * aligned portion of it that matches. Therefore we might need to fragment
1138 * which may fail and creates some hassle (caller must handle error case
1141 * FIXME: this could be merged to shift decision code
1143 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1144 u32 start_seq, u32 end_seq)
1148 unsigned int pkt_len;
1151 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1152 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1154 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1155 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1156 mss = tcp_skb_mss(skb);
1157 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1160 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1164 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1169 /* Round if necessary so that SACKs cover only full MSSes
1170 * and/or the remaining small portion (if present)
1172 if (pkt_len > mss) {
1173 unsigned int new_len = (pkt_len / mss) * mss;
1174 if (!in_sack && new_len < pkt_len)
1179 if (pkt_len >= skb->len && !in_sack)
1182 err = tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
1183 pkt_len, mss, GFP_ATOMIC);
1191 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1192 static u8 tcp_sacktag_one(struct sock *sk,
1193 struct tcp_sacktag_state *state, u8 sacked,
1194 u32 start_seq, u32 end_seq,
1195 int dup_sack, int pcount,
1198 struct tcp_sock *tp = tcp_sk(sk);
1200 /* Account D-SACK for retransmitted packet. */
1201 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1202 if (tp->undo_marker && tp->undo_retrans > 0 &&
1203 after(end_seq, tp->undo_marker))
1205 if ((sacked & TCPCB_SACKED_ACKED) &&
1206 before(start_seq, state->reord))
1207 state->reord = start_seq;
1210 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1211 if (!after(end_seq, tp->snd_una))
1214 if (!(sacked & TCPCB_SACKED_ACKED)) {
1215 tcp_rack_advance(tp, sacked, end_seq, xmit_time);
1217 if (sacked & TCPCB_SACKED_RETRANS) {
1218 /* If the segment is not tagged as lost,
1219 * we do not clear RETRANS, believing
1220 * that retransmission is still in flight.
1222 if (sacked & TCPCB_LOST) {
1223 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1224 tp->lost_out -= pcount;
1225 tp->retrans_out -= pcount;
1228 if (!(sacked & TCPCB_RETRANS)) {
1229 /* New sack for not retransmitted frame,
1230 * which was in hole. It is reordering.
1232 if (before(start_seq,
1233 tcp_highest_sack_seq(tp)) &&
1234 before(start_seq, state->reord))
1235 state->reord = start_seq;
1237 if (!after(end_seq, tp->high_seq))
1238 state->flag |= FLAG_ORIG_SACK_ACKED;
1239 if (state->first_sackt == 0)
1240 state->first_sackt = xmit_time;
1241 state->last_sackt = xmit_time;
1244 if (sacked & TCPCB_LOST) {
1245 sacked &= ~TCPCB_LOST;
1246 tp->lost_out -= pcount;
1250 sacked |= TCPCB_SACKED_ACKED;
1251 state->flag |= FLAG_DATA_SACKED;
1252 tp->sacked_out += pcount;
1253 tp->delivered += pcount; /* Out-of-order packets delivered */
1255 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1256 if (tp->lost_skb_hint &&
1257 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1258 tp->lost_cnt_hint += pcount;
1261 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1262 * frames and clear it. undo_retrans is decreased above, L|R frames
1263 * are accounted above as well.
1265 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1266 sacked &= ~TCPCB_SACKED_RETRANS;
1267 tp->retrans_out -= pcount;
1273 /* Shift newly-SACKed bytes from this skb to the immediately previous
1274 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1276 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev,
1277 struct sk_buff *skb,
1278 struct tcp_sacktag_state *state,
1279 unsigned int pcount, int shifted, int mss,
1282 struct tcp_sock *tp = tcp_sk(sk);
1283 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1284 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1288 /* Adjust counters and hints for the newly sacked sequence
1289 * range but discard the return value since prev is already
1290 * marked. We must tag the range first because the seq
1291 * advancement below implicitly advances
1292 * tcp_highest_sack_seq() when skb is highest_sack.
1294 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1295 start_seq, end_seq, dup_sack, pcount,
1297 tcp_rate_skb_delivered(sk, skb, state->rate);
1299 if (skb == tp->lost_skb_hint)
1300 tp->lost_cnt_hint += pcount;
1302 TCP_SKB_CB(prev)->end_seq += shifted;
1303 TCP_SKB_CB(skb)->seq += shifted;
1305 tcp_skb_pcount_add(prev, pcount);
1306 BUG_ON(tcp_skb_pcount(skb) < pcount);
1307 tcp_skb_pcount_add(skb, -pcount);
1309 /* When we're adding to gso_segs == 1, gso_size will be zero,
1310 * in theory this shouldn't be necessary but as long as DSACK
1311 * code can come after this skb later on it's better to keep
1312 * setting gso_size to something.
1314 if (!TCP_SKB_CB(prev)->tcp_gso_size)
1315 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1317 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1318 if (tcp_skb_pcount(skb) <= 1)
1319 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1321 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1322 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1325 BUG_ON(!tcp_skb_pcount(skb));
1326 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1330 /* Whole SKB was eaten :-) */
1332 if (skb == tp->retransmit_skb_hint)
1333 tp->retransmit_skb_hint = prev;
1334 if (skb == tp->lost_skb_hint) {
1335 tp->lost_skb_hint = prev;
1336 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1339 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1340 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor;
1341 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1342 TCP_SKB_CB(prev)->end_seq++;
1344 if (skb == tcp_highest_sack(sk))
1345 tcp_advance_highest_sack(sk, skb);
1347 tcp_skb_collapse_tstamp(prev, skb);
1348 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp))
1349 TCP_SKB_CB(prev)->tx.delivered_mstamp = 0;
1351 tcp_rtx_queue_unlink_and_free(skb, sk);
1353 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED);
1358 /* I wish gso_size would have a bit more sane initialization than
1359 * something-or-zero which complicates things
1361 static int tcp_skb_seglen(const struct sk_buff *skb)
1363 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1366 /* Shifting pages past head area doesn't work */
1367 static int skb_can_shift(const struct sk_buff *skb)
1369 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1372 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1375 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1376 struct tcp_sacktag_state *state,
1377 u32 start_seq, u32 end_seq,
1380 struct tcp_sock *tp = tcp_sk(sk);
1381 struct sk_buff *prev;
1387 /* Normally R but no L won't result in plain S */
1389 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1391 if (!skb_can_shift(skb))
1393 /* This frame is about to be dropped (was ACKed). */
1394 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1397 /* Can only happen with delayed DSACK + discard craziness */
1398 prev = skb_rb_prev(skb);
1402 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1405 if (!tcp_skb_can_collapse_to(prev))
1408 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1409 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1413 pcount = tcp_skb_pcount(skb);
1414 mss = tcp_skb_seglen(skb);
1416 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1417 * drop this restriction as unnecessary
1419 if (mss != tcp_skb_seglen(prev))
1422 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1424 /* CHECKME: This is non-MSS split case only?, this will
1425 * cause skipped skbs due to advancing loop btw, original
1426 * has that feature too
1428 if (tcp_skb_pcount(skb) <= 1)
1431 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1433 /* TODO: head merge to next could be attempted here
1434 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1435 * though it might not be worth of the additional hassle
1437 * ...we can probably just fallback to what was done
1438 * previously. We could try merging non-SACKed ones
1439 * as well but it probably isn't going to buy off
1440 * because later SACKs might again split them, and
1441 * it would make skb timestamp tracking considerably
1447 len = end_seq - TCP_SKB_CB(skb)->seq;
1449 BUG_ON(len > skb->len);
1451 /* MSS boundaries should be honoured or else pcount will
1452 * severely break even though it makes things bit trickier.
1453 * Optimize common case to avoid most of the divides
1455 mss = tcp_skb_mss(skb);
1457 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1458 * drop this restriction as unnecessary
1460 if (mss != tcp_skb_seglen(prev))
1465 } else if (len < mss) {
1473 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1474 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1477 if (!skb_shift(prev, skb, len))
1479 if (!tcp_shifted_skb(sk, prev, skb, state, pcount, len, mss, dup_sack))
1482 /* Hole filled allows collapsing with the next as well, this is very
1483 * useful when hole on every nth skb pattern happens
1485 skb = skb_rb_next(prev);
1489 if (!skb_can_shift(skb) ||
1490 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1491 (mss != tcp_skb_seglen(skb)))
1495 if (skb_shift(prev, skb, len)) {
1496 pcount += tcp_skb_pcount(skb);
1497 tcp_shifted_skb(sk, prev, skb, state, tcp_skb_pcount(skb),
1508 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1512 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1513 struct tcp_sack_block *next_dup,
1514 struct tcp_sacktag_state *state,
1515 u32 start_seq, u32 end_seq,
1518 struct tcp_sock *tp = tcp_sk(sk);
1519 struct sk_buff *tmp;
1521 skb_rbtree_walk_from(skb) {
1523 bool dup_sack = dup_sack_in;
1525 /* queue is in-order => we can short-circuit the walk early */
1526 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1530 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1531 in_sack = tcp_match_skb_to_sack(sk, skb,
1532 next_dup->start_seq,
1538 /* skb reference here is a bit tricky to get right, since
1539 * shifting can eat and free both this skb and the next,
1540 * so not even _safe variant of the loop is enough.
1543 tmp = tcp_shift_skb_data(sk, skb, state,
1544 start_seq, end_seq, dup_sack);
1553 in_sack = tcp_match_skb_to_sack(sk, skb,
1559 if (unlikely(in_sack < 0))
1563 TCP_SKB_CB(skb)->sacked =
1566 TCP_SKB_CB(skb)->sacked,
1567 TCP_SKB_CB(skb)->seq,
1568 TCP_SKB_CB(skb)->end_seq,
1570 tcp_skb_pcount(skb),
1572 tcp_rate_skb_delivered(sk, skb, state->rate);
1573 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1574 list_del_init(&skb->tcp_tsorted_anchor);
1576 if (!before(TCP_SKB_CB(skb)->seq,
1577 tcp_highest_sack_seq(tp)))
1578 tcp_advance_highest_sack(sk, skb);
1584 static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk,
1585 struct tcp_sacktag_state *state,
1588 struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node;
1589 struct sk_buff *skb;
1593 skb = rb_to_skb(parent);
1594 if (before(seq, TCP_SKB_CB(skb)->seq)) {
1595 p = &parent->rb_left;
1598 if (!before(seq, TCP_SKB_CB(skb)->end_seq)) {
1599 p = &parent->rb_right;
1607 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1608 struct tcp_sacktag_state *state,
1611 if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq))
1614 return tcp_sacktag_bsearch(sk, state, skip_to_seq);
1617 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1619 struct tcp_sack_block *next_dup,
1620 struct tcp_sacktag_state *state,
1626 if (before(next_dup->start_seq, skip_to_seq)) {
1627 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1628 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1629 next_dup->start_seq, next_dup->end_seq,
1636 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1638 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1642 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1643 u32 prior_snd_una, struct tcp_sacktag_state *state)
1645 struct tcp_sock *tp = tcp_sk(sk);
1646 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1647 TCP_SKB_CB(ack_skb)->sacked);
1648 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1649 struct tcp_sack_block sp[TCP_NUM_SACKS];
1650 struct tcp_sack_block *cache;
1651 struct sk_buff *skb;
1652 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1654 bool found_dup_sack = false;
1656 int first_sack_index;
1659 state->reord = tp->snd_nxt;
1661 if (!tp->sacked_out)
1662 tcp_highest_sack_reset(sk);
1664 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1665 num_sacks, prior_snd_una);
1666 if (found_dup_sack) {
1667 state->flag |= FLAG_DSACKING_ACK;
1668 tp->delivered++; /* A spurious retransmission is delivered */
1671 /* Eliminate too old ACKs, but take into
1672 * account more or less fresh ones, they can
1673 * contain valid SACK info.
1675 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1678 if (!tp->packets_out)
1682 first_sack_index = 0;
1683 for (i = 0; i < num_sacks; i++) {
1684 bool dup_sack = !i && found_dup_sack;
1686 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1687 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1689 if (!tcp_is_sackblock_valid(tp, dup_sack,
1690 sp[used_sacks].start_seq,
1691 sp[used_sacks].end_seq)) {
1695 if (!tp->undo_marker)
1696 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1698 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1700 /* Don't count olds caused by ACK reordering */
1701 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1702 !after(sp[used_sacks].end_seq, tp->snd_una))
1704 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1707 NET_INC_STATS(sock_net(sk), mib_idx);
1709 first_sack_index = -1;
1713 /* Ignore very old stuff early */
1714 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1720 /* order SACK blocks to allow in order walk of the retrans queue */
1721 for (i = used_sacks - 1; i > 0; i--) {
1722 for (j = 0; j < i; j++) {
1723 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1724 swap(sp[j], sp[j + 1]);
1726 /* Track where the first SACK block goes to */
1727 if (j == first_sack_index)
1728 first_sack_index = j + 1;
1733 state->mss_now = tcp_current_mss(sk);
1737 if (!tp->sacked_out) {
1738 /* It's already past, so skip checking against it */
1739 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1741 cache = tp->recv_sack_cache;
1742 /* Skip empty blocks in at head of the cache */
1743 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1748 while (i < used_sacks) {
1749 u32 start_seq = sp[i].start_seq;
1750 u32 end_seq = sp[i].end_seq;
1751 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1752 struct tcp_sack_block *next_dup = NULL;
1754 if (found_dup_sack && ((i + 1) == first_sack_index))
1755 next_dup = &sp[i + 1];
1757 /* Skip too early cached blocks */
1758 while (tcp_sack_cache_ok(tp, cache) &&
1759 !before(start_seq, cache->end_seq))
1762 /* Can skip some work by looking recv_sack_cache? */
1763 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1764 after(end_seq, cache->start_seq)) {
1767 if (before(start_seq, cache->start_seq)) {
1768 skb = tcp_sacktag_skip(skb, sk, state,
1770 skb = tcp_sacktag_walk(skb, sk, next_dup,
1777 /* Rest of the block already fully processed? */
1778 if (!after(end_seq, cache->end_seq))
1781 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1785 /* ...tail remains todo... */
1786 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1787 /* ...but better entrypoint exists! */
1788 skb = tcp_highest_sack(sk);
1795 skb = tcp_sacktag_skip(skb, sk, state, cache->end_seq);
1796 /* Check overlap against next cached too (past this one already) */
1801 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1802 skb = tcp_highest_sack(sk);
1806 skb = tcp_sacktag_skip(skb, sk, state, start_seq);
1809 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
1810 start_seq, end_seq, dup_sack);
1816 /* Clear the head of the cache sack blocks so we can skip it next time */
1817 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1818 tp->recv_sack_cache[i].start_seq = 0;
1819 tp->recv_sack_cache[i].end_seq = 0;
1821 for (j = 0; j < used_sacks; j++)
1822 tp->recv_sack_cache[i++] = sp[j];
1824 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss || tp->undo_marker)
1825 tcp_check_sack_reordering(sk, state->reord, 0);
1827 tcp_verify_left_out(tp);
1830 #if FASTRETRANS_DEBUG > 0
1831 WARN_ON((int)tp->sacked_out < 0);
1832 WARN_ON((int)tp->lost_out < 0);
1833 WARN_ON((int)tp->retrans_out < 0);
1834 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1839 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1840 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1842 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1846 holes = max(tp->lost_out, 1U);
1847 holes = min(holes, tp->packets_out);
1849 if ((tp->sacked_out + holes) > tp->packets_out) {
1850 tp->sacked_out = tp->packets_out - holes;
1856 /* If we receive more dupacks than we expected counting segments
1857 * in assumption of absent reordering, interpret this as reordering.
1858 * The only another reason could be bug in receiver TCP.
1860 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1862 struct tcp_sock *tp = tcp_sk(sk);
1864 if (!tcp_limit_reno_sacked(tp))
1867 tp->reordering = min_t(u32, tp->packets_out + addend,
1868 sock_net(sk)->ipv4.sysctl_tcp_max_reordering);
1869 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER);
1872 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1874 static void tcp_add_reno_sack(struct sock *sk)
1876 struct tcp_sock *tp = tcp_sk(sk);
1877 u32 prior_sacked = tp->sacked_out;
1880 tcp_check_reno_reordering(sk, 0);
1881 if (tp->sacked_out > prior_sacked)
1882 tp->delivered++; /* Some out-of-order packet is delivered */
1883 tcp_verify_left_out(tp);
1886 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1888 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1890 struct tcp_sock *tp = tcp_sk(sk);
1893 /* One ACK acked hole. The rest eat duplicate ACKs. */
1894 tp->delivered += max_t(int, acked - tp->sacked_out, 1);
1895 if (acked - 1 >= tp->sacked_out)
1898 tp->sacked_out -= acked - 1;
1900 tcp_check_reno_reordering(sk, acked);
1901 tcp_verify_left_out(tp);
1904 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1909 void tcp_clear_retrans(struct tcp_sock *tp)
1911 tp->retrans_out = 0;
1913 tp->undo_marker = 0;
1914 tp->undo_retrans = -1;
1918 static inline void tcp_init_undo(struct tcp_sock *tp)
1920 tp->undo_marker = tp->snd_una;
1921 /* Retransmission still in flight may cause DSACKs later. */
1922 tp->undo_retrans = tp->retrans_out ? : -1;
1925 static bool tcp_is_rack(const struct sock *sk)
1927 return sock_net(sk)->ipv4.sysctl_tcp_recovery & TCP_RACK_LOSS_DETECTION;
1930 /* If we detect SACK reneging, forget all SACK information
1931 * and reset tags completely, otherwise preserve SACKs. If receiver
1932 * dropped its ofo queue, we will know this due to reneging detection.
1934 static void tcp_timeout_mark_lost(struct sock *sk)
1936 struct tcp_sock *tp = tcp_sk(sk);
1937 struct sk_buff *skb, *head;
1938 bool is_reneg; /* is receiver reneging on SACKs? */
1940 head = tcp_rtx_queue_head(sk);
1941 is_reneg = head && (TCP_SKB_CB(head)->sacked & TCPCB_SACKED_ACKED);
1943 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
1945 /* Mark SACK reneging until we recover from this loss event. */
1946 tp->is_sack_reneg = 1;
1947 } else if (tcp_is_reno(tp)) {
1948 tcp_reset_reno_sack(tp);
1952 skb_rbtree_walk_from(skb) {
1954 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1955 else if (tcp_is_rack(sk) && skb != head &&
1956 tcp_rack_skb_timeout(tp, skb, 0) > 0)
1957 continue; /* Don't mark recently sent ones lost yet */
1958 tcp_mark_skb_lost(sk, skb);
1960 tcp_verify_left_out(tp);
1961 tcp_clear_all_retrans_hints(tp);
1964 /* Enter Loss state. */
1965 void tcp_enter_loss(struct sock *sk)
1967 const struct inet_connection_sock *icsk = inet_csk(sk);
1968 struct tcp_sock *tp = tcp_sk(sk);
1969 struct net *net = sock_net(sk);
1970 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
1972 tcp_timeout_mark_lost(sk);
1974 /* Reduce ssthresh if it has not yet been made inside this window. */
1975 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1976 !after(tp->high_seq, tp->snd_una) ||
1977 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1978 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1979 tp->prior_cwnd = tp->snd_cwnd;
1980 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1981 tcp_ca_event(sk, CA_EVENT_LOSS);
1984 tp->snd_cwnd = tcp_packets_in_flight(tp) + 1;
1985 tp->snd_cwnd_cnt = 0;
1986 tp->snd_cwnd_stamp = tcp_jiffies32;
1988 /* Timeout in disordered state after receiving substantial DUPACKs
1989 * suggests that the degree of reordering is over-estimated.
1991 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
1992 tp->sacked_out >= net->ipv4.sysctl_tcp_reordering)
1993 tp->reordering = min_t(unsigned int, tp->reordering,
1994 net->ipv4.sysctl_tcp_reordering);
1995 tcp_set_ca_state(sk, TCP_CA_Loss);
1996 tp->high_seq = tp->snd_nxt;
1997 tcp_ecn_queue_cwr(tp);
1999 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
2000 * loss recovery is underway except recurring timeout(s) on
2001 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
2003 tp->frto = net->ipv4.sysctl_tcp_frto &&
2004 (new_recovery || icsk->icsk_retransmits) &&
2005 !inet_csk(sk)->icsk_mtup.probe_size;
2008 /* If ACK arrived pointing to a remembered SACK, it means that our
2009 * remembered SACKs do not reflect real state of receiver i.e.
2010 * receiver _host_ is heavily congested (or buggy).
2012 * To avoid big spurious retransmission bursts due to transient SACK
2013 * scoreboard oddities that look like reneging, we give the receiver a
2014 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2015 * restore sanity to the SACK scoreboard. If the apparent reneging
2016 * persists until this RTO then we'll clear the SACK scoreboard.
2018 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
2020 if (flag & FLAG_SACK_RENEGING) {
2021 struct tcp_sock *tp = tcp_sk(sk);
2022 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2023 msecs_to_jiffies(10));
2025 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2026 delay, TCP_RTO_MAX);
2032 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2033 * counter when SACK is enabled (without SACK, sacked_out is used for
2036 * With reordering, holes may still be in flight, so RFC3517 recovery
2037 * uses pure sacked_out (total number of SACKed segments) even though
2038 * it violates the RFC that uses duplicate ACKs, often these are equal
2039 * but when e.g. out-of-window ACKs or packet duplication occurs,
2040 * they differ. Since neither occurs due to loss, TCP should really
2043 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2045 return tp->sacked_out + 1;
2048 /* Linux NewReno/SACK/ECN state machine.
2049 * --------------------------------------
2051 * "Open" Normal state, no dubious events, fast path.
2052 * "Disorder" In all the respects it is "Open",
2053 * but requires a bit more attention. It is entered when
2054 * we see some SACKs or dupacks. It is split of "Open"
2055 * mainly to move some processing from fast path to slow one.
2056 * "CWR" CWND was reduced due to some Congestion Notification event.
2057 * It can be ECN, ICMP source quench, local device congestion.
2058 * "Recovery" CWND was reduced, we are fast-retransmitting.
2059 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2061 * tcp_fastretrans_alert() is entered:
2062 * - each incoming ACK, if state is not "Open"
2063 * - when arrived ACK is unusual, namely:
2068 * Counting packets in flight is pretty simple.
2070 * in_flight = packets_out - left_out + retrans_out
2072 * packets_out is SND.NXT-SND.UNA counted in packets.
2074 * retrans_out is number of retransmitted segments.
2076 * left_out is number of segments left network, but not ACKed yet.
2078 * left_out = sacked_out + lost_out
2080 * sacked_out: Packets, which arrived to receiver out of order
2081 * and hence not ACKed. With SACKs this number is simply
2082 * amount of SACKed data. Even without SACKs
2083 * it is easy to give pretty reliable estimate of this number,
2084 * counting duplicate ACKs.
2086 * lost_out: Packets lost by network. TCP has no explicit
2087 * "loss notification" feedback from network (for now).
2088 * It means that this number can be only _guessed_.
2089 * Actually, it is the heuristics to predict lossage that
2090 * distinguishes different algorithms.
2092 * F.e. after RTO, when all the queue is considered as lost,
2093 * lost_out = packets_out and in_flight = retrans_out.
2095 * Essentially, we have now a few algorithms detecting
2098 * If the receiver supports SACK:
2100 * RFC6675/3517: It is the conventional algorithm. A packet is
2101 * considered lost if the number of higher sequence packets
2102 * SACKed is greater than or equal the DUPACK thoreshold
2103 * (reordering). This is implemented in tcp_mark_head_lost and
2104 * tcp_update_scoreboard.
2106 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm
2107 * (2017-) that checks timing instead of counting DUPACKs.
2108 * Essentially a packet is considered lost if it's not S/ACKed
2109 * after RTT + reordering_window, where both metrics are
2110 * dynamically measured and adjusted. This is implemented in
2111 * tcp_rack_mark_lost.
2113 * If the receiver does not support SACK:
2115 * NewReno (RFC6582): in Recovery we assume that one segment
2116 * is lost (classic Reno). While we are in Recovery and
2117 * a partial ACK arrives, we assume that one more packet
2118 * is lost (NewReno). This heuristics are the same in NewReno
2121 * Really tricky (and requiring careful tuning) part of algorithm
2122 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2123 * The first determines the moment _when_ we should reduce CWND and,
2124 * hence, slow down forward transmission. In fact, it determines the moment
2125 * when we decide that hole is caused by loss, rather than by a reorder.
2127 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2128 * holes, caused by lost packets.
2130 * And the most logically complicated part of algorithm is undo
2131 * heuristics. We detect false retransmits due to both too early
2132 * fast retransmit (reordering) and underestimated RTO, analyzing
2133 * timestamps and D-SACKs. When we detect that some segments were
2134 * retransmitted by mistake and CWND reduction was wrong, we undo
2135 * window reduction and abort recovery phase. This logic is hidden
2136 * inside several functions named tcp_try_undo_<something>.
2139 /* This function decides, when we should leave Disordered state
2140 * and enter Recovery phase, reducing congestion window.
2142 * Main question: may we further continue forward transmission
2143 * with the same cwnd?
2145 static bool tcp_time_to_recover(struct sock *sk, int flag)
2147 struct tcp_sock *tp = tcp_sk(sk);
2149 /* Trick#1: The loss is proven. */
2153 /* Not-A-Trick#2 : Classic rule... */
2154 if (!tcp_is_rack(sk) && tcp_dupack_heuristics(tp) > tp->reordering)
2160 /* Detect loss in event "A" above by marking head of queue up as lost.
2161 * For non-SACK(Reno) senders, the first "packets" number of segments
2162 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2163 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2164 * the maximum SACKed segments to pass before reaching this limit.
2166 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2168 struct tcp_sock *tp = tcp_sk(sk);
2169 struct sk_buff *skb;
2170 int cnt, oldcnt, lost;
2172 /* Use SACK to deduce losses of new sequences sent during recovery */
2173 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq;
2175 WARN_ON(packets > tp->packets_out);
2176 skb = tp->lost_skb_hint;
2178 /* Head already handled? */
2179 if (mark_head && after(TCP_SKB_CB(skb)->seq, tp->snd_una))
2181 cnt = tp->lost_cnt_hint;
2183 skb = tcp_rtx_queue_head(sk);
2187 skb_rbtree_walk_from(skb) {
2188 /* TODO: do this better */
2189 /* this is not the most efficient way to do this... */
2190 tp->lost_skb_hint = skb;
2191 tp->lost_cnt_hint = cnt;
2193 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2197 if (tcp_is_reno(tp) ||
2198 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2199 cnt += tcp_skb_pcount(skb);
2201 if (cnt > packets) {
2202 if (tcp_is_sack(tp) ||
2203 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2204 (oldcnt >= packets))
2207 mss = tcp_skb_mss(skb);
2208 /* If needed, chop off the prefix to mark as lost. */
2209 lost = (packets - oldcnt) * mss;
2210 if (lost < skb->len &&
2211 tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2212 lost, mss, GFP_ATOMIC) < 0)
2217 tcp_skb_mark_lost(tp, skb);
2222 tcp_verify_left_out(tp);
2225 /* Account newly detected lost packet(s) */
2227 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2229 struct tcp_sock *tp = tcp_sk(sk);
2231 if (tcp_is_sack(tp)) {
2232 int sacked_upto = tp->sacked_out - tp->reordering;
2233 if (sacked_upto >= 0)
2234 tcp_mark_head_lost(sk, sacked_upto, 0);
2235 else if (fast_rexmit)
2236 tcp_mark_head_lost(sk, 1, 1);
2240 static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2242 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2243 before(tp->rx_opt.rcv_tsecr, when);
2246 /* skb is spurious retransmitted if the returned timestamp echo
2247 * reply is prior to the skb transmission time
2249 static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2250 const struct sk_buff *skb)
2252 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2253 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp(skb));
2256 /* Nothing was retransmitted or returned timestamp is less
2257 * than timestamp of the first retransmission.
2259 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2261 return !tp->retrans_stamp ||
2262 tcp_tsopt_ecr_before(tp, tp->retrans_stamp);
2265 /* Undo procedures. */
2267 /* We can clear retrans_stamp when there are no retransmissions in the
2268 * window. It would seem that it is trivially available for us in
2269 * tp->retrans_out, however, that kind of assumptions doesn't consider
2270 * what will happen if errors occur when sending retransmission for the
2271 * second time. ...It could the that such segment has only
2272 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2273 * the head skb is enough except for some reneging corner cases that
2274 * are not worth the effort.
2276 * Main reason for all this complexity is the fact that connection dying
2277 * time now depends on the validity of the retrans_stamp, in particular,
2278 * that successive retransmissions of a segment must not advance
2279 * retrans_stamp under any conditions.
2281 static bool tcp_any_retrans_done(const struct sock *sk)
2283 const struct tcp_sock *tp = tcp_sk(sk);
2284 struct sk_buff *skb;
2286 if (tp->retrans_out)
2289 skb = tcp_rtx_queue_head(sk);
2290 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2296 static void DBGUNDO(struct sock *sk, const char *msg)
2298 #if FASTRETRANS_DEBUG > 1
2299 struct tcp_sock *tp = tcp_sk(sk);
2300 struct inet_sock *inet = inet_sk(sk);
2302 if (sk->sk_family == AF_INET) {
2303 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2305 &inet->inet_daddr, ntohs(inet->inet_dport),
2306 tp->snd_cwnd, tcp_left_out(tp),
2307 tp->snd_ssthresh, tp->prior_ssthresh,
2310 #if IS_ENABLED(CONFIG_IPV6)
2311 else if (sk->sk_family == AF_INET6) {
2312 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2314 &sk->sk_v6_daddr, ntohs(inet->inet_dport),
2315 tp->snd_cwnd, tcp_left_out(tp),
2316 tp->snd_ssthresh, tp->prior_ssthresh,
2323 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2325 struct tcp_sock *tp = tcp_sk(sk);
2328 struct sk_buff *skb;
2330 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2331 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2334 tcp_clear_all_retrans_hints(tp);
2337 if (tp->prior_ssthresh) {
2338 const struct inet_connection_sock *icsk = inet_csk(sk);
2340 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2342 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2343 tp->snd_ssthresh = tp->prior_ssthresh;
2344 tcp_ecn_withdraw_cwr(tp);
2347 tp->snd_cwnd_stamp = tcp_jiffies32;
2348 tp->undo_marker = 0;
2349 tp->rack.advanced = 1; /* Force RACK to re-exam losses */
2352 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2354 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2357 /* People celebrate: "We love our President!" */
2358 static bool tcp_try_undo_recovery(struct sock *sk)
2360 struct tcp_sock *tp = tcp_sk(sk);
2362 if (tcp_may_undo(tp)) {
2365 /* Happy end! We did not retransmit anything
2366 * or our original transmission succeeded.
2368 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2369 tcp_undo_cwnd_reduction(sk, false);
2370 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2371 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2373 mib_idx = LINUX_MIB_TCPFULLUNDO;
2375 NET_INC_STATS(sock_net(sk), mib_idx);
2376 } else if (tp->rack.reo_wnd_persist) {
2377 tp->rack.reo_wnd_persist--;
2379 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2380 /* Hold old state until something *above* high_seq
2381 * is ACKed. For Reno it is MUST to prevent false
2382 * fast retransmits (RFC2582). SACK TCP is safe. */
2383 if (!tcp_any_retrans_done(sk))
2384 tp->retrans_stamp = 0;
2387 tcp_set_ca_state(sk, TCP_CA_Open);
2388 tp->is_sack_reneg = 0;
2392 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2393 static bool tcp_try_undo_dsack(struct sock *sk)
2395 struct tcp_sock *tp = tcp_sk(sk);
2397 if (tp->undo_marker && !tp->undo_retrans) {
2398 tp->rack.reo_wnd_persist = min(TCP_RACK_RECOVERY_THRESH,
2399 tp->rack.reo_wnd_persist + 1);
2400 DBGUNDO(sk, "D-SACK");
2401 tcp_undo_cwnd_reduction(sk, false);
2402 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2408 /* Undo during loss recovery after partial ACK or using F-RTO. */
2409 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2411 struct tcp_sock *tp = tcp_sk(sk);
2413 if (frto_undo || tcp_may_undo(tp)) {
2414 tcp_undo_cwnd_reduction(sk, true);
2416 DBGUNDO(sk, "partial loss");
2417 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2419 NET_INC_STATS(sock_net(sk),
2420 LINUX_MIB_TCPSPURIOUSRTOS);
2421 inet_csk(sk)->icsk_retransmits = 0;
2422 if (frto_undo || tcp_is_sack(tp)) {
2423 tcp_set_ca_state(sk, TCP_CA_Open);
2424 tp->is_sack_reneg = 0;
2431 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2432 * It computes the number of packets to send (sndcnt) based on packets newly
2434 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2435 * cwnd reductions across a full RTT.
2436 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2437 * But when the retransmits are acked without further losses, PRR
2438 * slow starts cwnd up to ssthresh to speed up the recovery.
2440 static void tcp_init_cwnd_reduction(struct sock *sk)
2442 struct tcp_sock *tp = tcp_sk(sk);
2444 tp->high_seq = tp->snd_nxt;
2445 tp->tlp_high_seq = 0;
2446 tp->snd_cwnd_cnt = 0;
2447 tp->prior_cwnd = tp->snd_cwnd;
2448 tp->prr_delivered = 0;
2450 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2451 tcp_ecn_queue_cwr(tp);
2454 void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int flag)
2456 struct tcp_sock *tp = tcp_sk(sk);
2458 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2460 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2463 tp->prr_delivered += newly_acked_sacked;
2465 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2467 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2468 } else if ((flag & FLAG_RETRANS_DATA_ACKED) &&
2469 !(flag & FLAG_LOST_RETRANS)) {
2470 sndcnt = min_t(int, delta,
2471 max_t(int, tp->prr_delivered - tp->prr_out,
2472 newly_acked_sacked) + 1);
2474 sndcnt = min(delta, newly_acked_sacked);
2476 /* Force a fast retransmit upon entering fast recovery */
2477 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
2478 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2481 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2483 struct tcp_sock *tp = tcp_sk(sk);
2485 if (inet_csk(sk)->icsk_ca_ops->cong_control)
2488 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2489 if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH &&
2490 (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) {
2491 tp->snd_cwnd = tp->snd_ssthresh;
2492 tp->snd_cwnd_stamp = tcp_jiffies32;
2494 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2497 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2498 void tcp_enter_cwr(struct sock *sk)
2500 struct tcp_sock *tp = tcp_sk(sk);
2502 tp->prior_ssthresh = 0;
2503 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2504 tp->undo_marker = 0;
2505 tcp_init_cwnd_reduction(sk);
2506 tcp_set_ca_state(sk, TCP_CA_CWR);
2509 EXPORT_SYMBOL(tcp_enter_cwr);
2511 static void tcp_try_keep_open(struct sock *sk)
2513 struct tcp_sock *tp = tcp_sk(sk);
2514 int state = TCP_CA_Open;
2516 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2517 state = TCP_CA_Disorder;
2519 if (inet_csk(sk)->icsk_ca_state != state) {
2520 tcp_set_ca_state(sk, state);
2521 tp->high_seq = tp->snd_nxt;
2525 static void tcp_try_to_open(struct sock *sk, int flag)
2527 struct tcp_sock *tp = tcp_sk(sk);
2529 tcp_verify_left_out(tp);
2531 if (!tcp_any_retrans_done(sk))
2532 tp->retrans_stamp = 0;
2534 if (flag & FLAG_ECE)
2537 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2538 tcp_try_keep_open(sk);
2542 static void tcp_mtup_probe_failed(struct sock *sk)
2544 struct inet_connection_sock *icsk = inet_csk(sk);
2546 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2547 icsk->icsk_mtup.probe_size = 0;
2548 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
2551 static void tcp_mtup_probe_success(struct sock *sk)
2553 struct tcp_sock *tp = tcp_sk(sk);
2554 struct inet_connection_sock *icsk = inet_csk(sk);
2556 /* FIXME: breaks with very large cwnd */
2557 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2558 tp->snd_cwnd = tp->snd_cwnd *
2559 tcp_mss_to_mtu(sk, tp->mss_cache) /
2560 icsk->icsk_mtup.probe_size;
2561 tp->snd_cwnd_cnt = 0;
2562 tp->snd_cwnd_stamp = tcp_jiffies32;
2563 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2565 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2566 icsk->icsk_mtup.probe_size = 0;
2567 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2568 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
2571 /* Do a simple retransmit without using the backoff mechanisms in
2572 * tcp_timer. This is used for path mtu discovery.
2573 * The socket is already locked here.
2575 void tcp_simple_retransmit(struct sock *sk)
2577 const struct inet_connection_sock *icsk = inet_csk(sk);
2578 struct tcp_sock *tp = tcp_sk(sk);
2579 struct sk_buff *skb;
2580 unsigned int mss = tcp_current_mss(sk);
2582 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2583 if (tcp_skb_seglen(skb) > mss &&
2584 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2585 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2586 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2587 tp->retrans_out -= tcp_skb_pcount(skb);
2589 tcp_skb_mark_lost_uncond_verify(tp, skb);
2593 tcp_clear_retrans_hints_partial(tp);
2598 if (tcp_is_reno(tp))
2599 tcp_limit_reno_sacked(tp);
2601 tcp_verify_left_out(tp);
2603 /* Don't muck with the congestion window here.
2604 * Reason is that we do not increase amount of _data_
2605 * in network, but units changed and effective
2606 * cwnd/ssthresh really reduced now.
2608 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2609 tp->high_seq = tp->snd_nxt;
2610 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2611 tp->prior_ssthresh = 0;
2612 tp->undo_marker = 0;
2613 tcp_set_ca_state(sk, TCP_CA_Loss);
2615 tcp_xmit_retransmit_queue(sk);
2617 EXPORT_SYMBOL(tcp_simple_retransmit);
2619 void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2621 struct tcp_sock *tp = tcp_sk(sk);
2624 if (tcp_is_reno(tp))
2625 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2627 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2629 NET_INC_STATS(sock_net(sk), mib_idx);
2631 tp->prior_ssthresh = 0;
2634 if (!tcp_in_cwnd_reduction(sk)) {
2636 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2637 tcp_init_cwnd_reduction(sk);
2639 tcp_set_ca_state(sk, TCP_CA_Recovery);
2642 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2643 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2645 static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack,
2648 struct tcp_sock *tp = tcp_sk(sk);
2649 bool recovered = !before(tp->snd_una, tp->high_seq);
2651 if ((flag & FLAG_SND_UNA_ADVANCED) &&
2652 tcp_try_undo_loss(sk, false))
2655 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2656 /* Step 3.b. A timeout is spurious if not all data are
2657 * lost, i.e., never-retransmitted data are (s)acked.
2659 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2660 tcp_try_undo_loss(sk, true))
2663 if (after(tp->snd_nxt, tp->high_seq)) {
2664 if (flag & FLAG_DATA_SACKED || is_dupack)
2665 tp->frto = 0; /* Step 3.a. loss was real */
2666 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2667 tp->high_seq = tp->snd_nxt;
2668 /* Step 2.b. Try send new data (but deferred until cwnd
2669 * is updated in tcp_ack()). Otherwise fall back to
2670 * the conventional recovery.
2672 if (!tcp_write_queue_empty(sk) &&
2673 after(tcp_wnd_end(tp), tp->snd_nxt)) {
2674 *rexmit = REXMIT_NEW;
2682 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2683 tcp_try_undo_recovery(sk);
2686 if (tcp_is_reno(tp)) {
2687 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2688 * delivered. Lower inflight to clock out (re)tranmissions.
2690 if (after(tp->snd_nxt, tp->high_seq) && is_dupack)
2691 tcp_add_reno_sack(sk);
2692 else if (flag & FLAG_SND_UNA_ADVANCED)
2693 tcp_reset_reno_sack(tp);
2695 *rexmit = REXMIT_LOST;
2698 /* Undo during fast recovery after partial ACK. */
2699 static bool tcp_try_undo_partial(struct sock *sk, u32 prior_snd_una)
2701 struct tcp_sock *tp = tcp_sk(sk);
2703 if (tp->undo_marker && tcp_packet_delayed(tp)) {
2704 /* Plain luck! Hole if filled with delayed
2705 * packet, rather than with a retransmit. Check reordering.
2707 tcp_check_sack_reordering(sk, prior_snd_una, 1);
2709 /* We are getting evidence that the reordering degree is higher
2710 * than we realized. If there are no retransmits out then we
2711 * can undo. Otherwise we clock out new packets but do not
2712 * mark more packets lost or retransmit more.
2714 if (tp->retrans_out)
2717 if (!tcp_any_retrans_done(sk))
2718 tp->retrans_stamp = 0;
2720 DBGUNDO(sk, "partial recovery");
2721 tcp_undo_cwnd_reduction(sk, true);
2722 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2723 tcp_try_keep_open(sk);
2729 static void tcp_identify_packet_loss(struct sock *sk, int *ack_flag)
2731 struct tcp_sock *tp = tcp_sk(sk);
2733 if (tcp_rtx_queue_empty(sk))
2736 if (unlikely(tcp_is_reno(tp))) {
2737 tcp_newreno_mark_lost(sk, *ack_flag & FLAG_SND_UNA_ADVANCED);
2738 } else if (tcp_is_rack(sk)) {
2739 u32 prior_retrans = tp->retrans_out;
2741 tcp_rack_mark_lost(sk);
2742 if (prior_retrans > tp->retrans_out)
2743 *ack_flag |= FLAG_LOST_RETRANS;
2747 static bool tcp_force_fast_retransmit(struct sock *sk)
2749 struct tcp_sock *tp = tcp_sk(sk);
2751 return after(tcp_highest_sack_seq(tp),
2752 tp->snd_una + tp->reordering * tp->mss_cache);
2755 /* Process an event, which can update packets-in-flight not trivially.
2756 * Main goal of this function is to calculate new estimate for left_out,
2757 * taking into account both packets sitting in receiver's buffer and
2758 * packets lost by network.
2760 * Besides that it updates the congestion state when packet loss or ECN
2761 * is detected. But it does not reduce the cwnd, it is done by the
2762 * congestion control later.
2764 * It does _not_ decide what to send, it is made in function
2765 * tcp_xmit_retransmit_queue().
2767 static void tcp_fastretrans_alert(struct sock *sk, const u32 prior_snd_una,
2768 bool is_dupack, int *ack_flag, int *rexmit)
2770 struct inet_connection_sock *icsk = inet_csk(sk);
2771 struct tcp_sock *tp = tcp_sk(sk);
2772 int fast_rexmit = 0, flag = *ack_flag;
2773 bool do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2774 tcp_force_fast_retransmit(sk));
2776 if (!tp->packets_out && tp->sacked_out)
2779 /* Now state machine starts.
2780 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2781 if (flag & FLAG_ECE)
2782 tp->prior_ssthresh = 0;
2784 /* B. In all the states check for reneging SACKs. */
2785 if (tcp_check_sack_reneging(sk, flag))
2788 /* C. Check consistency of the current state. */
2789 tcp_verify_left_out(tp);
2791 /* D. Check state exit conditions. State can be terminated
2792 * when high_seq is ACKed. */
2793 if (icsk->icsk_ca_state == TCP_CA_Open) {
2794 WARN_ON(tp->retrans_out != 0);
2795 tp->retrans_stamp = 0;
2796 } else if (!before(tp->snd_una, tp->high_seq)) {
2797 switch (icsk->icsk_ca_state) {
2799 /* CWR is to be held something *above* high_seq
2800 * is ACKed for CWR bit to reach receiver. */
2801 if (tp->snd_una != tp->high_seq) {
2802 tcp_end_cwnd_reduction(sk);
2803 tcp_set_ca_state(sk, TCP_CA_Open);
2807 case TCP_CA_Recovery:
2808 if (tcp_is_reno(tp))
2809 tcp_reset_reno_sack(tp);
2810 if (tcp_try_undo_recovery(sk))
2812 tcp_end_cwnd_reduction(sk);
2817 /* E. Process state. */
2818 switch (icsk->icsk_ca_state) {
2819 case TCP_CA_Recovery:
2820 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2821 if (tcp_is_reno(tp) && is_dupack)
2822 tcp_add_reno_sack(sk);
2824 if (tcp_try_undo_partial(sk, prior_snd_una))
2826 /* Partial ACK arrived. Force fast retransmit. */
2827 do_lost = tcp_is_reno(tp) ||
2828 tcp_force_fast_retransmit(sk);
2830 if (tcp_try_undo_dsack(sk)) {
2831 tcp_try_keep_open(sk);
2834 tcp_identify_packet_loss(sk, ack_flag);
2837 tcp_process_loss(sk, flag, is_dupack, rexmit);
2838 tcp_identify_packet_loss(sk, ack_flag);
2839 if (!(icsk->icsk_ca_state == TCP_CA_Open ||
2840 (*ack_flag & FLAG_LOST_RETRANS)))
2842 /* Change state if cwnd is undone or retransmits are lost */
2845 if (tcp_is_reno(tp)) {
2846 if (flag & FLAG_SND_UNA_ADVANCED)
2847 tcp_reset_reno_sack(tp);
2849 tcp_add_reno_sack(sk);
2852 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2853 tcp_try_undo_dsack(sk);
2855 tcp_identify_packet_loss(sk, ack_flag);
2856 if (!tcp_time_to_recover(sk, flag)) {
2857 tcp_try_to_open(sk, flag);
2861 /* MTU probe failure: don't reduce cwnd */
2862 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2863 icsk->icsk_mtup.probe_size &&
2864 tp->snd_una == tp->mtu_probe.probe_seq_start) {
2865 tcp_mtup_probe_failed(sk);
2866 /* Restores the reduction we did in tcp_mtup_probe() */
2868 tcp_simple_retransmit(sk);
2872 /* Otherwise enter Recovery state */
2873 tcp_enter_recovery(sk, (flag & FLAG_ECE));
2877 if (!tcp_is_rack(sk) && do_lost)
2878 tcp_update_scoreboard(sk, fast_rexmit);
2879 *rexmit = REXMIT_LOST;
2882 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us, const int flag)
2884 u32 wlen = sock_net(sk)->ipv4.sysctl_tcp_min_rtt_wlen * HZ;
2885 struct tcp_sock *tp = tcp_sk(sk);
2887 if ((flag & FLAG_ACK_MAYBE_DELAYED) && rtt_us > tcp_min_rtt(tp)) {
2888 /* If the remote keeps returning delayed ACKs, eventually
2889 * the min filter would pick it up and overestimate the
2890 * prop. delay when it expires. Skip suspected delayed ACKs.
2894 minmax_running_min(&tp->rtt_min, wlen, tcp_jiffies32,
2895 rtt_us ? : jiffies_to_usecs(1));
2898 static bool tcp_ack_update_rtt(struct sock *sk, const int flag,
2899 long seq_rtt_us, long sack_rtt_us,
2900 long ca_rtt_us, struct rate_sample *rs)
2902 const struct tcp_sock *tp = tcp_sk(sk);
2904 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2905 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2906 * Karn's algorithm forbids taking RTT if some retransmitted data
2907 * is acked (RFC6298).
2910 seq_rtt_us = sack_rtt_us;
2912 /* RTTM Rule: A TSecr value received in a segment is used to
2913 * update the averaged RTT measurement only if the segment
2914 * acknowledges some new data, i.e., only if it advances the
2915 * left edge of the send window.
2916 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2918 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2919 flag & FLAG_ACKED) {
2920 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
2921 u32 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
2923 seq_rtt_us = ca_rtt_us = delta_us;
2925 rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */
2929 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
2930 * always taken together with ACK, SACK, or TS-opts. Any negative
2931 * values will be skipped with the seq_rtt_us < 0 check above.
2933 tcp_update_rtt_min(sk, ca_rtt_us, flag);
2934 tcp_rtt_estimator(sk, seq_rtt_us);
2937 /* RFC6298: only reset backoff on valid RTT measurement. */
2938 inet_csk(sk)->icsk_backoff = 0;
2942 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2943 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
2945 struct rate_sample rs;
2948 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack)
2949 rtt_us = tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req)->snt_synack);
2951 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us, &rs);
2955 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
2957 const struct inet_connection_sock *icsk = inet_csk(sk);
2959 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
2960 tcp_sk(sk)->snd_cwnd_stamp = tcp_jiffies32;
2963 /* Restart timer after forward progress on connection.
2964 * RFC2988 recommends to restart timer to now+rto.
2966 void tcp_rearm_rto(struct sock *sk)
2968 const struct inet_connection_sock *icsk = inet_csk(sk);
2969 struct tcp_sock *tp = tcp_sk(sk);
2971 /* If the retrans timer is currently being used by Fast Open
2972 * for SYN-ACK retrans purpose, stay put.
2974 if (tp->fastopen_rsk)
2977 if (!tp->packets_out) {
2978 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2980 u32 rto = inet_csk(sk)->icsk_rto;
2981 /* Offset the time elapsed after installing regular RTO */
2982 if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT ||
2983 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
2984 s64 delta_us = tcp_rto_delta_us(sk);
2985 /* delta_us may not be positive if the socket is locked
2986 * when the retrans timer fires and is rescheduled.
2988 rto = usecs_to_jiffies(max_t(int, delta_us, 1));
2990 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
2995 /* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */
2996 static void tcp_set_xmit_timer(struct sock *sk)
2998 if (!tcp_schedule_loss_probe(sk, true))
3002 /* If we get here, the whole TSO packet has not been acked. */
3003 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3005 struct tcp_sock *tp = tcp_sk(sk);
3008 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3010 packets_acked = tcp_skb_pcount(skb);
3011 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3013 packets_acked -= tcp_skb_pcount(skb);
3015 if (packets_acked) {
3016 BUG_ON(tcp_skb_pcount(skb) == 0);
3017 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3020 return packets_acked;
3023 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3026 const struct skb_shared_info *shinfo;
3028 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3029 if (likely(!TCP_SKB_CB(skb)->txstamp_ack))
3032 shinfo = skb_shinfo(skb);
3033 if (!before(shinfo->tskey, prior_snd_una) &&
3034 before(shinfo->tskey, tcp_sk(sk)->snd_una)) {
3035 tcp_skb_tsorted_save(skb) {
3036 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK);
3037 } tcp_skb_tsorted_restore(skb);
3041 /* Remove acknowledged frames from the retransmission queue. If our packet
3042 * is before the ack sequence we can discard it as it's confirmed to have
3043 * arrived at the other end.
3045 static int tcp_clean_rtx_queue(struct sock *sk, u32 prior_fack,
3047 struct tcp_sacktag_state *sack)
3049 const struct inet_connection_sock *icsk = inet_csk(sk);
3050 u64 first_ackt, last_ackt;
3051 struct tcp_sock *tp = tcp_sk(sk);
3052 u32 prior_sacked = tp->sacked_out;
3053 u32 reord = tp->snd_nxt; /* lowest acked un-retx un-sacked seq */
3054 struct sk_buff *skb, *next;
3055 bool fully_acked = true;
3056 long sack_rtt_us = -1L;
3057 long seq_rtt_us = -1L;
3058 long ca_rtt_us = -1L;
3060 u32 last_in_flight = 0;
3066 for (skb = skb_rb_first(&sk->tcp_rtx_queue); skb; skb = next) {
3067 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3068 const u32 start_seq = scb->seq;
3069 u8 sacked = scb->sacked;
3072 tcp_ack_tstamp(sk, skb, prior_snd_una);
3074 /* Determine how many packets and what bytes were acked, tso and else */
3075 if (after(scb->end_seq, tp->snd_una)) {
3076 if (tcp_skb_pcount(skb) == 1 ||
3077 !after(tp->snd_una, scb->seq))
3080 acked_pcount = tcp_tso_acked(sk, skb);
3083 fully_acked = false;
3085 acked_pcount = tcp_skb_pcount(skb);
3088 if (unlikely(sacked & TCPCB_RETRANS)) {
3089 if (sacked & TCPCB_SACKED_RETRANS)
3090 tp->retrans_out -= acked_pcount;
3091 flag |= FLAG_RETRANS_DATA_ACKED;
3092 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3093 last_ackt = skb->skb_mstamp;
3094 WARN_ON_ONCE(last_ackt == 0);
3096 first_ackt = last_ackt;
3098 last_in_flight = TCP_SKB_CB(skb)->tx.in_flight;
3099 if (before(start_seq, reord))
3101 if (!after(scb->end_seq, tp->high_seq))
3102 flag |= FLAG_ORIG_SACK_ACKED;
3105 if (sacked & TCPCB_SACKED_ACKED) {
3106 tp->sacked_out -= acked_pcount;
3107 } else if (tcp_is_sack(tp)) {
3108 tp->delivered += acked_pcount;
3109 if (!tcp_skb_spurious_retrans(tp, skb))
3110 tcp_rack_advance(tp, sacked, scb->end_seq,
3113 if (sacked & TCPCB_LOST)
3114 tp->lost_out -= acked_pcount;
3116 tp->packets_out -= acked_pcount;
3117 pkts_acked += acked_pcount;
3118 tcp_rate_skb_delivered(sk, skb, sack->rate);
3120 /* Initial outgoing SYN's get put onto the write_queue
3121 * just like anything else we transmit. It is not
3122 * true data, and if we misinform our callers that
3123 * this ACK acks real data, we will erroneously exit
3124 * connection startup slow start one packet too
3125 * quickly. This is severely frowned upon behavior.
3127 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3128 flag |= FLAG_DATA_ACKED;
3130 flag |= FLAG_SYN_ACKED;
3131 tp->retrans_stamp = 0;
3137 next = skb_rb_next(skb);
3138 if (unlikely(skb == tp->retransmit_skb_hint))
3139 tp->retransmit_skb_hint = NULL;
3140 if (unlikely(skb == tp->lost_skb_hint))
3141 tp->lost_skb_hint = NULL;
3142 tcp_rtx_queue_unlink_and_free(skb, sk);
3146 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3148 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3149 tp->snd_up = tp->snd_una;
3151 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3152 flag |= FLAG_SACK_RENEGING;
3154 if (likely(first_ackt) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
3155 seq_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, first_ackt);
3156 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, last_ackt);
3158 if (pkts_acked == 1 && last_in_flight < tp->mss_cache &&
3159 last_in_flight && !prior_sacked && fully_acked &&
3160 sack->rate->prior_delivered + 1 == tp->delivered &&
3161 !(flag & (FLAG_CA_ALERT | FLAG_SYN_ACKED))) {
3162 /* Conservatively mark a delayed ACK. It's typically
3163 * from a lone runt packet over the round trip to
3164 * a receiver w/o out-of-order or CE events.
3166 flag |= FLAG_ACK_MAYBE_DELAYED;
3169 if (sack->first_sackt) {
3170 sack_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->first_sackt);
3171 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->last_sackt);
3173 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
3174 ca_rtt_us, sack->rate);
3176 if (flag & FLAG_ACKED) {
3177 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3178 if (unlikely(icsk->icsk_mtup.probe_size &&
3179 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3180 tcp_mtup_probe_success(sk);
3183 if (tcp_is_reno(tp)) {
3184 tcp_remove_reno_sacks(sk, pkts_acked);
3186 /* If any of the cumulatively ACKed segments was
3187 * retransmitted, non-SACK case cannot confirm that
3188 * progress was due to original transmission due to
3189 * lack of TCPCB_SACKED_ACKED bits even if some of
3190 * the packets may have been never retransmitted.
3192 if (flag & FLAG_RETRANS_DATA_ACKED)
3193 flag &= ~FLAG_ORIG_SACK_ACKED;
3197 /* Non-retransmitted hole got filled? That's reordering */
3198 if (before(reord, prior_fack))
3199 tcp_check_sack_reordering(sk, reord, 0);
3201 delta = prior_sacked - tp->sacked_out;
3202 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3204 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3205 sack_rtt_us > tcp_stamp_us_delta(tp->tcp_mstamp, skb->skb_mstamp)) {
3206 /* Do not re-arm RTO if the sack RTT is measured from data sent
3207 * after when the head was last (re)transmitted. Otherwise the
3208 * timeout may continue to extend in loss recovery.
3210 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3213 if (icsk->icsk_ca_ops->pkts_acked) {
3214 struct ack_sample sample = { .pkts_acked = pkts_acked,
3215 .rtt_us = sack->rate->rtt_us,
3216 .in_flight = last_in_flight };
3218 icsk->icsk_ca_ops->pkts_acked(sk, &sample);
3221 #if FASTRETRANS_DEBUG > 0
3222 WARN_ON((int)tp->sacked_out < 0);
3223 WARN_ON((int)tp->lost_out < 0);
3224 WARN_ON((int)tp->retrans_out < 0);
3225 if (!tp->packets_out && tcp_is_sack(tp)) {
3226 icsk = inet_csk(sk);
3228 pr_debug("Leak l=%u %d\n",
3229 tp->lost_out, icsk->icsk_ca_state);
3232 if (tp->sacked_out) {
3233 pr_debug("Leak s=%u %d\n",
3234 tp->sacked_out, icsk->icsk_ca_state);
3237 if (tp->retrans_out) {
3238 pr_debug("Leak r=%u %d\n",
3239 tp->retrans_out, icsk->icsk_ca_state);
3240 tp->retrans_out = 0;
3247 static void tcp_ack_probe(struct sock *sk)
3249 struct inet_connection_sock *icsk = inet_csk(sk);
3250 struct sk_buff *head = tcp_send_head(sk);
3251 const struct tcp_sock *tp = tcp_sk(sk);
3253 /* Was it a usable window open? */
3256 if (!after(TCP_SKB_CB(head)->end_seq, tcp_wnd_end(tp))) {
3257 icsk->icsk_backoff = 0;
3258 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3259 /* Socket must be waked up by subsequent tcp_data_snd_check().
3260 * This function is not for random using!
3263 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
3265 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3270 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3272 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3273 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3276 /* Decide wheather to run the increase function of congestion control. */
3277 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3279 /* If reordering is high then always grow cwnd whenever data is
3280 * delivered regardless of its ordering. Otherwise stay conservative
3281 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3282 * new SACK or ECE mark may first advance cwnd here and later reduce
3283 * cwnd in tcp_fastretrans_alert() based on more states.
3285 if (tcp_sk(sk)->reordering > sock_net(sk)->ipv4.sysctl_tcp_reordering)
3286 return flag & FLAG_FORWARD_PROGRESS;
3288 return flag & FLAG_DATA_ACKED;
3291 /* The "ultimate" congestion control function that aims to replace the rigid
3292 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3293 * It's called toward the end of processing an ACK with precise rate
3294 * information. All transmission or retransmission are delayed afterwards.
3296 static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
3297 int flag, const struct rate_sample *rs)
3299 const struct inet_connection_sock *icsk = inet_csk(sk);
3301 if (icsk->icsk_ca_ops->cong_control) {
3302 icsk->icsk_ca_ops->cong_control(sk, rs);
3306 if (tcp_in_cwnd_reduction(sk)) {
3307 /* Reduce cwnd if state mandates */
3308 tcp_cwnd_reduction(sk, acked_sacked, flag);
3309 } else if (tcp_may_raise_cwnd(sk, flag)) {
3310 /* Advance cwnd if state allows */
3311 tcp_cong_avoid(sk, ack, acked_sacked);
3313 tcp_update_pacing_rate(sk);
3316 /* Check that window update is acceptable.
3317 * The function assumes that snd_una<=ack<=snd_next.
3319 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3320 const u32 ack, const u32 ack_seq,
3323 return after(ack, tp->snd_una) ||
3324 after(ack_seq, tp->snd_wl1) ||
3325 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3328 /* If we update tp->snd_una, also update tp->bytes_acked */
3329 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3331 u32 delta = ack - tp->snd_una;
3333 sock_owned_by_me((struct sock *)tp);
3334 tp->bytes_acked += delta;
3338 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3339 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3341 u32 delta = seq - tp->rcv_nxt;
3343 sock_owned_by_me((struct sock *)tp);
3344 tp->bytes_received += delta;
3348 /* Update our send window.
3350 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3351 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3353 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3356 struct tcp_sock *tp = tcp_sk(sk);
3358 u32 nwin = ntohs(tcp_hdr(skb)->window);
3360 if (likely(!tcp_hdr(skb)->syn))
3361 nwin <<= tp->rx_opt.snd_wscale;
3363 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3364 flag |= FLAG_WIN_UPDATE;
3365 tcp_update_wl(tp, ack_seq);
3367 if (tp->snd_wnd != nwin) {
3370 /* Note, it is the only place, where
3371 * fast path is recovered for sending TCP.
3374 tcp_fast_path_check(sk);
3376 if (!tcp_write_queue_empty(sk))
3377 tcp_slow_start_after_idle_check(sk);
3379 if (nwin > tp->max_window) {
3380 tp->max_window = nwin;
3381 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3386 tcp_snd_una_update(tp, ack);
3391 static bool __tcp_oow_rate_limited(struct net *net, int mib_idx,
3392 u32 *last_oow_ack_time)
3394 if (*last_oow_ack_time) {
3395 s32 elapsed = (s32)(tcp_jiffies32 - *last_oow_ack_time);
3397 if (0 <= elapsed && elapsed < net->ipv4.sysctl_tcp_invalid_ratelimit) {
3398 NET_INC_STATS(net, mib_idx);
3399 return true; /* rate-limited: don't send yet! */
3403 *last_oow_ack_time = tcp_jiffies32;
3405 return false; /* not rate-limited: go ahead, send dupack now! */
3408 /* Return true if we're currently rate-limiting out-of-window ACKs and
3409 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3410 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3411 * attacks that send repeated SYNs or ACKs for the same connection. To
3412 * do this, we do not send a duplicate SYNACK or ACK if the remote
3413 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3415 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3416 int mib_idx, u32 *last_oow_ack_time)
3418 /* Data packets without SYNs are not likely part of an ACK loop. */
3419 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3423 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time);
3426 /* RFC 5961 7 [ACK Throttling] */
3427 static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb)
3429 /* unprotected vars, we dont care of overwrites */
3430 static u32 challenge_timestamp;
3431 static unsigned int challenge_count;
3432 struct tcp_sock *tp = tcp_sk(sk);
3433 struct net *net = sock_net(sk);
3436 /* First check our per-socket dupack rate limit. */
3437 if (__tcp_oow_rate_limited(net,
3438 LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3439 &tp->last_oow_ack_time))
3442 /* Then check host-wide RFC 5961 rate limit. */
3444 if (now != challenge_timestamp) {
3445 u32 ack_limit = net->ipv4.sysctl_tcp_challenge_ack_limit;
3446 u32 half = (ack_limit + 1) >> 1;
3448 challenge_timestamp = now;
3449 WRITE_ONCE(challenge_count, half + prandom_u32_max(ack_limit));
3451 count = READ_ONCE(challenge_count);
3453 WRITE_ONCE(challenge_count, count - 1);
3454 NET_INC_STATS(net, LINUX_MIB_TCPCHALLENGEACK);
3459 static void tcp_store_ts_recent(struct tcp_sock *tp)
3461 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3462 tp->rx_opt.ts_recent_stamp = get_seconds();
3465 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3467 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3468 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3469 * extra check below makes sure this can only happen
3470 * for pure ACK frames. -DaveM
3472 * Not only, also it occurs for expired timestamps.
3475 if (tcp_paws_check(&tp->rx_opt, 0))
3476 tcp_store_ts_recent(tp);
3480 /* This routine deals with acks during a TLP episode.
3481 * We mark the end of a TLP episode on receiving TLP dupack or when
3482 * ack is after tlp_high_seq.
3483 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3485 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3487 struct tcp_sock *tp = tcp_sk(sk);
3489 if (before(ack, tp->tlp_high_seq))
3492 if (flag & FLAG_DSACKING_ACK) {
3493 /* This DSACK means original and TLP probe arrived; no loss */
3494 tp->tlp_high_seq = 0;
3495 } else if (after(ack, tp->tlp_high_seq)) {
3496 /* ACK advances: there was a loss, so reduce cwnd. Reset
3497 * tlp_high_seq in tcp_init_cwnd_reduction()
3499 tcp_init_cwnd_reduction(sk);
3500 tcp_set_ca_state(sk, TCP_CA_CWR);
3501 tcp_end_cwnd_reduction(sk);
3502 tcp_try_keep_open(sk);
3503 NET_INC_STATS(sock_net(sk),
3504 LINUX_MIB_TCPLOSSPROBERECOVERY);
3505 } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3506 FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3507 /* Pure dupack: original and TLP probe arrived; no loss */
3508 tp->tlp_high_seq = 0;
3512 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3514 const struct inet_connection_sock *icsk = inet_csk(sk);
3516 if (icsk->icsk_ca_ops->in_ack_event)
3517 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3520 /* Congestion control has updated the cwnd already. So if we're in
3521 * loss recovery then now we do any new sends (for FRTO) or
3522 * retransmits (for CA_Loss or CA_recovery) that make sense.
3524 static void tcp_xmit_recovery(struct sock *sk, int rexmit)
3526 struct tcp_sock *tp = tcp_sk(sk);
3528 if (rexmit == REXMIT_NONE)
3531 if (unlikely(rexmit == 2)) {
3532 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
3534 if (after(tp->snd_nxt, tp->high_seq))
3538 tcp_xmit_retransmit_queue(sk);
3541 /* Returns the number of packets newly acked or sacked by the current ACK */
3542 static u32 tcp_newly_delivered(struct sock *sk, u32 prior_delivered, int flag)
3544 const struct net *net = sock_net(sk);
3545 struct tcp_sock *tp = tcp_sk(sk);
3548 delivered = tp->delivered - prior_delivered;
3549 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVERED, delivered);
3550 if (flag & FLAG_ECE) {
3551 tp->delivered_ce += delivered;
3552 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVEREDCE, delivered);
3557 /* This routine deals with incoming acks, but not outgoing ones. */
3558 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3560 struct inet_connection_sock *icsk = inet_csk(sk);
3561 struct tcp_sock *tp = tcp_sk(sk);
3562 struct tcp_sacktag_state sack_state;
3563 struct rate_sample rs = { .prior_delivered = 0 };
3564 u32 prior_snd_una = tp->snd_una;
3565 bool is_sack_reneg = tp->is_sack_reneg;
3566 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3567 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3568 bool is_dupack = false;
3569 int prior_packets = tp->packets_out;
3570 u32 delivered = tp->delivered;
3571 u32 lost = tp->lost;
3572 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
3575 sack_state.first_sackt = 0;
3576 sack_state.rate = &rs;
3578 /* We very likely will need to access rtx queue. */
3579 prefetch(sk->tcp_rtx_queue.rb_node);
3581 /* If the ack is older than previous acks
3582 * then we can probably ignore it.
3584 if (before(ack, prior_snd_una)) {
3585 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3586 if (before(ack, prior_snd_una - tp->max_window)) {
3587 if (!(flag & FLAG_NO_CHALLENGE_ACK))
3588 tcp_send_challenge_ack(sk, skb);
3594 /* If the ack includes data we haven't sent yet, discard
3595 * this segment (RFC793 Section 3.9).
3597 if (after(ack, tp->snd_nxt))
3600 if (after(ack, prior_snd_una)) {
3601 flag |= FLAG_SND_UNA_ADVANCED;
3602 icsk->icsk_retransmits = 0;
3604 #if IS_ENABLED(CONFIG_TLS_DEVICE)
3605 if (static_branch_unlikely(&clean_acked_data_enabled))
3606 if (icsk->icsk_clean_acked)
3607 icsk->icsk_clean_acked(sk, ack);
3611 prior_fack = tcp_is_sack(tp) ? tcp_highest_sack_seq(tp) : tp->snd_una;
3612 rs.prior_in_flight = tcp_packets_in_flight(tp);
3614 /* ts_recent update must be made after we are sure that the packet
3617 if (flag & FLAG_UPDATE_TS_RECENT)
3618 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3620 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3621 /* Window is constant, pure forward advance.
3622 * No more checks are required.
3623 * Note, we use the fact that SND.UNA>=SND.WL2.
3625 tcp_update_wl(tp, ack_seq);
3626 tcp_snd_una_update(tp, ack);
3627 flag |= FLAG_WIN_UPDATE;
3629 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3631 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS);
3633 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3635 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3638 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3640 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3642 if (TCP_SKB_CB(skb)->sacked)
3643 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3646 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3648 ack_ev_flags |= CA_ACK_ECE;
3651 if (flag & FLAG_WIN_UPDATE)
3652 ack_ev_flags |= CA_ACK_WIN_UPDATE;
3654 tcp_in_ack_event(sk, ack_ev_flags);
3657 /* We passed data and got it acked, remove any soft error
3658 * log. Something worked...
3660 sk->sk_err_soft = 0;
3661 icsk->icsk_probes_out = 0;
3662 tp->rcv_tstamp = tcp_jiffies32;
3666 /* See if we can take anything off of the retransmit queue. */
3667 flag |= tcp_clean_rtx_queue(sk, prior_fack, prior_snd_una, &sack_state);
3669 tcp_rack_update_reo_wnd(sk, &rs);
3671 if (tp->tlp_high_seq)
3672 tcp_process_tlp_ack(sk, ack, flag);
3673 /* If needed, reset TLP/RTO timer; RACK may later override this. */
3674 if (flag & FLAG_SET_XMIT_TIMER)
3675 tcp_set_xmit_timer(sk);
3677 if (tcp_ack_is_dubious(sk, flag)) {
3678 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3679 tcp_fastretrans_alert(sk, prior_snd_una, is_dupack, &flag,
3683 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3686 delivered = tcp_newly_delivered(sk, delivered, flag);
3687 lost = tp->lost - lost; /* freshly marked lost */
3688 rs.is_ack_delayed = !!(flag & FLAG_ACK_MAYBE_DELAYED);
3689 tcp_rate_gen(sk, delivered, lost, is_sack_reneg, sack_state.rate);
3690 tcp_cong_control(sk, ack, delivered, flag, sack_state.rate);
3691 tcp_xmit_recovery(sk, rexmit);
3695 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3696 if (flag & FLAG_DSACKING_ACK) {
3697 tcp_fastretrans_alert(sk, prior_snd_una, is_dupack, &flag,
3699 tcp_newly_delivered(sk, delivered, flag);
3701 /* If this ack opens up a zero window, clear backoff. It was
3702 * being used to time the probes, and is probably far higher than
3703 * it needs to be for normal retransmission.
3707 if (tp->tlp_high_seq)
3708 tcp_process_tlp_ack(sk, ack, flag);
3712 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3716 /* If data was SACKed, tag it and see if we should send more data.
3717 * If data was DSACKed, see if we can undo a cwnd reduction.
3719 if (TCP_SKB_CB(skb)->sacked) {
3720 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3722 tcp_fastretrans_alert(sk, prior_snd_una, is_dupack, &flag,
3724 tcp_newly_delivered(sk, delivered, flag);
3725 tcp_xmit_recovery(sk, rexmit);
3728 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3732 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3733 bool syn, struct tcp_fastopen_cookie *foc,
3736 /* Valid only in SYN or SYN-ACK with an even length. */
3737 if (!foc || !syn || len < 0 || (len & 1))
3740 if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3741 len <= TCP_FASTOPEN_COOKIE_MAX)
3742 memcpy(foc->val, cookie, len);
3749 static void smc_parse_options(const struct tcphdr *th,
3750 struct tcp_options_received *opt_rx,
3751 const unsigned char *ptr,
3754 #if IS_ENABLED(CONFIG_SMC)
3755 if (static_branch_unlikely(&tcp_have_smc)) {
3756 if (th->syn && !(opsize & 1) &&
3757 opsize >= TCPOLEN_EXP_SMC_BASE &&
3758 get_unaligned_be32(ptr) == TCPOPT_SMC_MAGIC)
3764 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3765 * But, this can also be called on packets in the established flow when
3766 * the fast version below fails.
3768 void tcp_parse_options(const struct net *net,
3769 const struct sk_buff *skb,
3770 struct tcp_options_received *opt_rx, int estab,
3771 struct tcp_fastopen_cookie *foc)
3773 const unsigned char *ptr;
3774 const struct tcphdr *th = tcp_hdr(skb);
3775 int length = (th->doff * 4) - sizeof(struct tcphdr);
3777 ptr = (const unsigned char *)(th + 1);
3778 opt_rx->saw_tstamp = 0;
3780 while (length > 0) {
3781 int opcode = *ptr++;
3787 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3792 if (opsize < 2) /* "silly options" */
3794 if (opsize > length)
3795 return; /* don't parse partial options */
3798 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3799 u16 in_mss = get_unaligned_be16(ptr);
3801 if (opt_rx->user_mss &&
3802 opt_rx->user_mss < in_mss)
3803 in_mss = opt_rx->user_mss;
3804 opt_rx->mss_clamp = in_mss;
3809 if (opsize == TCPOLEN_WINDOW && th->syn &&
3810 !estab && net->ipv4.sysctl_tcp_window_scaling) {
3811 __u8 snd_wscale = *(__u8 *)ptr;
3812 opt_rx->wscale_ok = 1;
3813 if (snd_wscale > TCP_MAX_WSCALE) {
3814 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n",
3818 snd_wscale = TCP_MAX_WSCALE;
3820 opt_rx->snd_wscale = snd_wscale;
3823 case TCPOPT_TIMESTAMP:
3824 if ((opsize == TCPOLEN_TIMESTAMP) &&
3825 ((estab && opt_rx->tstamp_ok) ||
3826 (!estab && net->ipv4.sysctl_tcp_timestamps))) {
3827 opt_rx->saw_tstamp = 1;
3828 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3829 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3832 case TCPOPT_SACK_PERM:
3833 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3834 !estab && net->ipv4.sysctl_tcp_sack) {
3835 opt_rx->sack_ok = TCP_SACK_SEEN;
3836 tcp_sack_reset(opt_rx);
3841 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3842 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3844 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3847 #ifdef CONFIG_TCP_MD5SIG
3850 * The MD5 Hash has already been
3851 * checked (see tcp_v{4,6}_do_rcv()).
3855 case TCPOPT_FASTOPEN:
3856 tcp_parse_fastopen_option(
3857 opsize - TCPOLEN_FASTOPEN_BASE,
3858 ptr, th->syn, foc, false);
3862 /* Fast Open option shares code 254 using a
3863 * 16 bits magic number.
3865 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
3866 get_unaligned_be16(ptr) ==
3867 TCPOPT_FASTOPEN_MAGIC)
3868 tcp_parse_fastopen_option(opsize -
3869 TCPOLEN_EXP_FASTOPEN_BASE,
3870 ptr + 2, th->syn, foc, true);
3872 smc_parse_options(th, opt_rx, ptr,
3882 EXPORT_SYMBOL(tcp_parse_options);
3884 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3886 const __be32 *ptr = (const __be32 *)(th + 1);
3888 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3889 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3890 tp->rx_opt.saw_tstamp = 1;
3892 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3895 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3897 tp->rx_opt.rcv_tsecr = 0;
3903 /* Fast parse options. This hopes to only see timestamps.
3904 * If it is wrong it falls back on tcp_parse_options().
3906 static bool tcp_fast_parse_options(const struct net *net,
3907 const struct sk_buff *skb,
3908 const struct tcphdr *th, struct tcp_sock *tp)
3910 /* In the spirit of fast parsing, compare doff directly to constant
3911 * values. Because equality is used, short doff can be ignored here.
3913 if (th->doff == (sizeof(*th) / 4)) {
3914 tp->rx_opt.saw_tstamp = 0;
3916 } else if (tp->rx_opt.tstamp_ok &&
3917 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3918 if (tcp_parse_aligned_timestamp(tp, th))
3922 tcp_parse_options(net, skb, &tp->rx_opt, 1, NULL);
3923 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3924 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3929 #ifdef CONFIG_TCP_MD5SIG
3931 * Parse MD5 Signature option
3933 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3935 int length = (th->doff << 2) - sizeof(*th);
3936 const u8 *ptr = (const u8 *)(th + 1);
3938 /* If not enough data remaining, we can short cut */
3939 while (length >= TCPOLEN_MD5SIG) {
3940 int opcode = *ptr++;
3951 if (opsize < 2 || opsize > length)
3953 if (opcode == TCPOPT_MD5SIG)
3954 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3961 EXPORT_SYMBOL(tcp_parse_md5sig_option);
3964 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3966 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3967 * it can pass through stack. So, the following predicate verifies that
3968 * this segment is not used for anything but congestion avoidance or
3969 * fast retransmit. Moreover, we even are able to eliminate most of such
3970 * second order effects, if we apply some small "replay" window (~RTO)
3971 * to timestamp space.
3973 * All these measures still do not guarantee that we reject wrapped ACKs
3974 * on networks with high bandwidth, when sequence space is recycled fastly,
3975 * but it guarantees that such events will be very rare and do not affect
3976 * connection seriously. This doesn't look nice, but alas, PAWS is really
3979 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3980 * states that events when retransmit arrives after original data are rare.
3981 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3982 * the biggest problem on large power networks even with minor reordering.
3983 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3984 * up to bandwidth of 18Gigabit/sec. 8) ]
3987 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3989 const struct tcp_sock *tp = tcp_sk(sk);
3990 const struct tcphdr *th = tcp_hdr(skb);
3991 u32 seq = TCP_SKB_CB(skb)->seq;
3992 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3994 return (/* 1. Pure ACK with correct sequence number. */
3995 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3997 /* 2. ... and duplicate ACK. */
3998 ack == tp->snd_una &&
4000 /* 3. ... and does not update window. */
4001 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
4003 /* 4. ... and sits in replay window. */
4004 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
4007 static inline bool tcp_paws_discard(const struct sock *sk,
4008 const struct sk_buff *skb)
4010 const struct tcp_sock *tp = tcp_sk(sk);
4012 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4013 !tcp_disordered_ack(sk, skb);
4016 /* Check segment sequence number for validity.
4018 * Segment controls are considered valid, if the segment
4019 * fits to the window after truncation to the window. Acceptability
4020 * of data (and SYN, FIN, of course) is checked separately.
4021 * See tcp_data_queue(), for example.
4023 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4024 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4025 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4026 * (borrowed from freebsd)
4029 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
4031 return !before(end_seq, tp->rcv_wup) &&
4032 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
4035 /* When we get a reset we do this. */
4036 void tcp_reset(struct sock *sk)
4038 trace_tcp_receive_reset(sk);
4040 /* We want the right error as BSD sees it (and indeed as we do). */
4041 switch (sk->sk_state) {
4043 sk->sk_err = ECONNREFUSED;
4045 case TCP_CLOSE_WAIT:
4051 sk->sk_err = ECONNRESET;
4053 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4056 tcp_write_queue_purge(sk);
4059 if (!sock_flag(sk, SOCK_DEAD))
4060 sk->sk_error_report(sk);
4064 * Process the FIN bit. This now behaves as it is supposed to work
4065 * and the FIN takes effect when it is validly part of sequence
4066 * space. Not before when we get holes.
4068 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4069 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4072 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4073 * close and we go into CLOSING (and later onto TIME-WAIT)
4075 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4077 void tcp_fin(struct sock *sk)
4079 struct tcp_sock *tp = tcp_sk(sk);
4081 inet_csk_schedule_ack(sk);
4083 sk->sk_shutdown |= RCV_SHUTDOWN;
4084 sock_set_flag(sk, SOCK_DONE);
4086 switch (sk->sk_state) {
4088 case TCP_ESTABLISHED:
4089 /* Move to CLOSE_WAIT */
4090 tcp_set_state(sk, TCP_CLOSE_WAIT);
4091 inet_csk(sk)->icsk_ack.pingpong = 1;
4094 case TCP_CLOSE_WAIT:
4096 /* Received a retransmission of the FIN, do
4101 /* RFC793: Remain in the LAST-ACK state. */
4105 /* This case occurs when a simultaneous close
4106 * happens, we must ack the received FIN and
4107 * enter the CLOSING state.
4110 tcp_set_state(sk, TCP_CLOSING);
4113 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4115 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4118 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4119 * cases we should never reach this piece of code.
4121 pr_err("%s: Impossible, sk->sk_state=%d\n",
4122 __func__, sk->sk_state);
4126 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4127 * Probably, we should reset in this case. For now drop them.
4129 skb_rbtree_purge(&tp->out_of_order_queue);
4130 if (tcp_is_sack(tp))
4131 tcp_sack_reset(&tp->rx_opt);
4134 if (!sock_flag(sk, SOCK_DEAD)) {
4135 sk->sk_state_change(sk);
4137 /* Do not send POLL_HUP for half duplex close. */
4138 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4139 sk->sk_state == TCP_CLOSE)
4140 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4142 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4146 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4149 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4150 if (before(seq, sp->start_seq))
4151 sp->start_seq = seq;
4152 if (after(end_seq, sp->end_seq))
4153 sp->end_seq = end_seq;
4159 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4161 struct tcp_sock *tp = tcp_sk(sk);
4163 if (tcp_is_sack(tp) && sock_net(sk)->ipv4.sysctl_tcp_dsack) {
4166 if (before(seq, tp->rcv_nxt))
4167 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4169 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4171 NET_INC_STATS(sock_net(sk), mib_idx);
4173 tp->rx_opt.dsack = 1;
4174 tp->duplicate_sack[0].start_seq = seq;
4175 tp->duplicate_sack[0].end_seq = end_seq;
4179 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4181 struct tcp_sock *tp = tcp_sk(sk);
4183 if (!tp->rx_opt.dsack)
4184 tcp_dsack_set(sk, seq, end_seq);
4186 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4189 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4191 struct tcp_sock *tp = tcp_sk(sk);
4193 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4194 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4195 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4196 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
4198 if (tcp_is_sack(tp) && sock_net(sk)->ipv4.sysctl_tcp_dsack) {
4199 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4201 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4202 end_seq = tp->rcv_nxt;
4203 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4210 /* These routines update the SACK block as out-of-order packets arrive or
4211 * in-order packets close up the sequence space.
4213 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4216 struct tcp_sack_block *sp = &tp->selective_acks[0];
4217 struct tcp_sack_block *swalk = sp + 1;
4219 /* See if the recent change to the first SACK eats into
4220 * or hits the sequence space of other SACK blocks, if so coalesce.
4222 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4223 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4226 /* Zap SWALK, by moving every further SACK up by one slot.
4227 * Decrease num_sacks.
4229 tp->rx_opt.num_sacks--;
4230 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4234 this_sack++, swalk++;
4238 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4240 struct tcp_sock *tp = tcp_sk(sk);
4241 struct tcp_sack_block *sp = &tp->selective_acks[0];
4242 int cur_sacks = tp->rx_opt.num_sacks;
4248 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4249 if (tcp_sack_extend(sp, seq, end_seq)) {
4250 /* Rotate this_sack to the first one. */
4251 for (; this_sack > 0; this_sack--, sp--)
4252 swap(*sp, *(sp - 1));
4254 tcp_sack_maybe_coalesce(tp);
4259 /* Could not find an adjacent existing SACK, build a new one,
4260 * put it at the front, and shift everyone else down. We
4261 * always know there is at least one SACK present already here.
4263 * If the sack array is full, forget about the last one.
4265 if (this_sack >= TCP_NUM_SACKS) {
4266 if (tp->compressed_ack)
4269 tp->rx_opt.num_sacks--;
4272 for (; this_sack > 0; this_sack--, sp--)
4276 /* Build the new head SACK, and we're done. */
4277 sp->start_seq = seq;
4278 sp->end_seq = end_seq;
4279 tp->rx_opt.num_sacks++;
4282 /* RCV.NXT advances, some SACKs should be eaten. */
4284 static void tcp_sack_remove(struct tcp_sock *tp)
4286 struct tcp_sack_block *sp = &tp->selective_acks[0];
4287 int num_sacks = tp->rx_opt.num_sacks;
4290 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4291 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4292 tp->rx_opt.num_sacks = 0;
4296 for (this_sack = 0; this_sack < num_sacks;) {
4297 /* Check if the start of the sack is covered by RCV.NXT. */
4298 if (!before(tp->rcv_nxt, sp->start_seq)) {
4301 /* RCV.NXT must cover all the block! */
4302 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4304 /* Zap this SACK, by moving forward any other SACKS. */
4305 for (i = this_sack+1; i < num_sacks; i++)
4306 tp->selective_acks[i-1] = tp->selective_acks[i];
4313 tp->rx_opt.num_sacks = num_sacks;
4317 * tcp_try_coalesce - try to merge skb to prior one
4319 * @dest: destination queue
4321 * @from: buffer to add in queue
4322 * @fragstolen: pointer to boolean
4324 * Before queueing skb @from after @to, try to merge them
4325 * to reduce overall memory use and queue lengths, if cost is small.
4326 * Packets in ofo or receive queues can stay a long time.
4327 * Better try to coalesce them right now to avoid future collapses.
4328 * Returns true if caller should free @from instead of queueing it
4330 static bool tcp_try_coalesce(struct sock *sk,
4332 struct sk_buff *from,
4337 *fragstolen = false;
4339 /* Its possible this segment overlaps with prior segment in queue */
4340 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4343 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4346 atomic_add(delta, &sk->sk_rmem_alloc);
4347 sk_mem_charge(sk, delta);
4348 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4349 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4350 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4351 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4353 if (TCP_SKB_CB(from)->has_rxtstamp) {
4354 TCP_SKB_CB(to)->has_rxtstamp = true;
4355 to->tstamp = from->tstamp;
4361 static void tcp_drop(struct sock *sk, struct sk_buff *skb)
4363 sk_drops_add(sk, skb);
4367 /* This one checks to see if we can put data from the
4368 * out_of_order queue into the receive_queue.
4370 static void tcp_ofo_queue(struct sock *sk)
4372 struct tcp_sock *tp = tcp_sk(sk);
4373 __u32 dsack_high = tp->rcv_nxt;
4374 bool fin, fragstolen, eaten;
4375 struct sk_buff *skb, *tail;
4378 p = rb_first(&tp->out_of_order_queue);
4381 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4384 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4385 __u32 dsack = dsack_high;
4386 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4387 dsack_high = TCP_SKB_CB(skb)->end_seq;
4388 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4391 rb_erase(&skb->rbnode, &tp->out_of_order_queue);
4393 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) {
4394 SOCK_DEBUG(sk, "ofo packet was already received\n");
4398 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4399 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4400 TCP_SKB_CB(skb)->end_seq);
4402 tail = skb_peek_tail(&sk->sk_receive_queue);
4403 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4404 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4405 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
4407 __skb_queue_tail(&sk->sk_receive_queue, skb);
4409 kfree_skb_partial(skb, fragstolen);
4411 if (unlikely(fin)) {
4413 /* tcp_fin() purges tp->out_of_order_queue,
4414 * so we must end this loop right now.
4421 static bool tcp_prune_ofo_queue(struct sock *sk);
4422 static int tcp_prune_queue(struct sock *sk);
4424 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4427 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4428 !sk_rmem_schedule(sk, skb, size)) {
4430 if (tcp_prune_queue(sk) < 0)
4433 while (!sk_rmem_schedule(sk, skb, size)) {
4434 if (!tcp_prune_ofo_queue(sk))
4441 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4443 struct tcp_sock *tp = tcp_sk(sk);
4444 struct rb_node **p, *parent;
4445 struct sk_buff *skb1;
4449 tcp_ecn_check_ce(sk, skb);
4451 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4452 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP);
4457 /* Disable header prediction. */
4459 inet_csk_schedule_ack(sk);
4461 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4462 seq = TCP_SKB_CB(skb)->seq;
4463 end_seq = TCP_SKB_CB(skb)->end_seq;
4464 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4465 tp->rcv_nxt, seq, end_seq);
4467 p = &tp->out_of_order_queue.rb_node;
4468 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4469 /* Initial out of order segment, build 1 SACK. */
4470 if (tcp_is_sack(tp)) {
4471 tp->rx_opt.num_sacks = 1;
4472 tp->selective_acks[0].start_seq = seq;
4473 tp->selective_acks[0].end_seq = end_seq;
4475 rb_link_node(&skb->rbnode, NULL, p);
4476 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4477 tp->ooo_last_skb = skb;
4481 /* In the typical case, we are adding an skb to the end of the list.
4482 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4484 if (tcp_try_coalesce(sk, tp->ooo_last_skb,
4485 skb, &fragstolen)) {
4487 tcp_grow_window(sk, skb);
4488 kfree_skb_partial(skb, fragstolen);
4492 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4493 if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) {
4494 parent = &tp->ooo_last_skb->rbnode;
4495 p = &parent->rb_right;
4499 /* Find place to insert this segment. Handle overlaps on the way. */
4503 skb1 = rb_to_skb(parent);
4504 if (before(seq, TCP_SKB_CB(skb1)->seq)) {
4505 p = &parent->rb_left;
4508 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4509 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4510 /* All the bits are present. Drop. */
4511 NET_INC_STATS(sock_net(sk),
4512 LINUX_MIB_TCPOFOMERGE);
4515 tcp_dsack_set(sk, seq, end_seq);
4518 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4519 /* Partial overlap. */
4520 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq);
4522 /* skb's seq == skb1's seq and skb covers skb1.
4523 * Replace skb1 with skb.
4525 rb_replace_node(&skb1->rbnode, &skb->rbnode,
4526 &tp->out_of_order_queue);
4527 tcp_dsack_extend(sk,
4528 TCP_SKB_CB(skb1)->seq,
4529 TCP_SKB_CB(skb1)->end_seq);
4530 NET_INC_STATS(sock_net(sk),
4531 LINUX_MIB_TCPOFOMERGE);
4535 } else if (tcp_try_coalesce(sk, skb1,
4536 skb, &fragstolen)) {
4539 p = &parent->rb_right;
4542 /* Insert segment into RB tree. */
4543 rb_link_node(&skb->rbnode, parent, p);
4544 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4547 /* Remove other segments covered by skb. */
4548 while ((skb1 = skb_rb_next(skb)) != NULL) {
4549 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4551 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4552 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4556 rb_erase(&skb1->rbnode, &tp->out_of_order_queue);
4557 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4558 TCP_SKB_CB(skb1)->end_seq);
4559 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4562 /* If there is no skb after us, we are the last_skb ! */
4564 tp->ooo_last_skb = skb;
4567 if (tcp_is_sack(tp))
4568 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4571 tcp_grow_window(sk, skb);
4573 skb_set_owner_r(skb, sk);
4577 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
4581 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4583 __skb_pull(skb, hdrlen);
4585 tcp_try_coalesce(sk, tail,
4586 skb, fragstolen)) ? 1 : 0;
4587 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
4589 __skb_queue_tail(&sk->sk_receive_queue, skb);
4590 skb_set_owner_r(skb, sk);
4595 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4597 struct sk_buff *skb;
4605 if (size > PAGE_SIZE) {
4606 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
4608 data_len = npages << PAGE_SHIFT;
4609 size = data_len + (size & ~PAGE_MASK);
4611 skb = alloc_skb_with_frags(size - data_len, data_len,
4612 PAGE_ALLOC_COSTLY_ORDER,
4613 &err, sk->sk_allocation);
4617 skb_put(skb, size - data_len);
4618 skb->data_len = data_len;
4621 if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4624 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
4628 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4629 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4630 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4632 if (tcp_queue_rcv(sk, skb, 0, &fragstolen)) {
4633 WARN_ON_ONCE(fragstolen); /* should not happen */
4645 void tcp_data_ready(struct sock *sk)
4647 const struct tcp_sock *tp = tcp_sk(sk);
4648 int avail = tp->rcv_nxt - tp->copied_seq;
4650 if (avail < sk->sk_rcvlowat && !sock_flag(sk, SOCK_DONE))
4653 sk->sk_data_ready(sk);
4656 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4658 struct tcp_sock *tp = tcp_sk(sk);
4662 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
4667 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
4669 tcp_ecn_accept_cwr(tp, skb);
4671 tp->rx_opt.dsack = 0;
4673 /* Queue data for delivery to the user.
4674 * Packets in sequence go to the receive queue.
4675 * Out of sequence packets to the out_of_order_queue.
4677 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4678 if (tcp_receive_window(tp) == 0)
4681 /* Ok. In sequence. In window. */
4683 if (skb_queue_len(&sk->sk_receive_queue) == 0)
4684 sk_forced_mem_schedule(sk, skb->truesize);
4685 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4688 eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
4689 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4691 tcp_event_data_recv(sk, skb);
4692 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4695 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4698 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4699 * gap in queue is filled.
4701 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
4702 inet_csk(sk)->icsk_ack.pingpong = 0;
4705 if (tp->rx_opt.num_sacks)
4706 tcp_sack_remove(tp);
4708 tcp_fast_path_check(sk);
4711 kfree_skb_partial(skb, fragstolen);
4712 if (!sock_flag(sk, SOCK_DEAD))
4717 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4718 /* A retransmit, 2nd most common case. Force an immediate ack. */
4719 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4720 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4723 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
4724 inet_csk_schedule_ack(sk);
4730 /* Out of window. F.e. zero window probe. */
4731 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4734 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4735 /* Partial packet, seq < rcv_next < end_seq */
4736 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4737 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4738 TCP_SKB_CB(skb)->end_seq);
4740 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4742 /* If window is closed, drop tail of packet. But after
4743 * remembering D-SACK for its head made in previous line.
4745 if (!tcp_receive_window(tp))
4750 tcp_data_queue_ofo(sk, skb);
4753 static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list)
4756 return !skb_queue_is_last(list, skb) ? skb->next : NULL;
4758 return skb_rb_next(skb);
4761 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4762 struct sk_buff_head *list,
4763 struct rb_root *root)
4765 struct sk_buff *next = tcp_skb_next(skb, list);
4768 __skb_unlink(skb, list);
4770 rb_erase(&skb->rbnode, root);
4773 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4778 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
4779 void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb)
4781 struct rb_node **p = &root->rb_node;
4782 struct rb_node *parent = NULL;
4783 struct sk_buff *skb1;
4787 skb1 = rb_to_skb(parent);
4788 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq))
4789 p = &parent->rb_left;
4791 p = &parent->rb_right;
4793 rb_link_node(&skb->rbnode, parent, p);
4794 rb_insert_color(&skb->rbnode, root);
4797 /* Collapse contiguous sequence of skbs head..tail with
4798 * sequence numbers start..end.
4800 * If tail is NULL, this means until the end of the queue.
4802 * Segments with FIN/SYN are not collapsed (only because this
4806 tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root,
4807 struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end)
4809 struct sk_buff *skb = head, *n;
4810 struct sk_buff_head tmp;
4813 /* First, check that queue is collapsible and find
4814 * the point where collapsing can be useful.
4817 for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) {
4818 n = tcp_skb_next(skb, list);
4820 /* No new bits? It is possible on ofo queue. */
4821 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4822 skb = tcp_collapse_one(sk, skb, list, root);
4828 /* The first skb to collapse is:
4830 * - bloated or contains data before "start" or
4831 * overlaps to the next one.
4833 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
4834 (tcp_win_from_space(sk, skb->truesize) > skb->len ||
4835 before(TCP_SKB_CB(skb)->seq, start))) {
4836 end_of_skbs = false;
4840 if (n && n != tail &&
4841 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) {
4842 end_of_skbs = false;
4846 /* Decided to skip this, advance start seq. */
4847 start = TCP_SKB_CB(skb)->end_seq;
4850 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4853 __skb_queue_head_init(&tmp);
4855 while (before(start, end)) {
4856 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
4857 struct sk_buff *nskb;
4859 nskb = alloc_skb(copy, GFP_ATOMIC);
4863 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4864 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4866 __skb_queue_before(list, skb, nskb);
4868 __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */
4869 skb_set_owner_r(nskb, sk);
4871 /* Copy data, releasing collapsed skbs. */
4873 int offset = start - TCP_SKB_CB(skb)->seq;
4874 int size = TCP_SKB_CB(skb)->end_seq - start;
4878 size = min(copy, size);
4879 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4881 TCP_SKB_CB(nskb)->end_seq += size;
4885 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4886 skb = tcp_collapse_one(sk, skb, list, root);
4889 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4895 skb_queue_walk_safe(&tmp, skb, n)
4896 tcp_rbtree_insert(root, skb);
4899 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4900 * and tcp_collapse() them until all the queue is collapsed.
4902 static void tcp_collapse_ofo_queue(struct sock *sk)
4904 struct tcp_sock *tp = tcp_sk(sk);
4905 u32 range_truesize, sum_tiny = 0;
4906 struct sk_buff *skb, *head;
4909 skb = skb_rb_first(&tp->out_of_order_queue);
4912 tp->ooo_last_skb = skb_rb_last(&tp->out_of_order_queue);
4915 start = TCP_SKB_CB(skb)->seq;
4916 end = TCP_SKB_CB(skb)->end_seq;
4917 range_truesize = skb->truesize;
4919 for (head = skb;;) {
4920 skb = skb_rb_next(skb);
4922 /* Range is terminated when we see a gap or when
4923 * we are at the queue end.
4926 after(TCP_SKB_CB(skb)->seq, end) ||
4927 before(TCP_SKB_CB(skb)->end_seq, start)) {
4928 /* Do not attempt collapsing tiny skbs */
4929 if (range_truesize != head->truesize ||
4930 end - start >= SKB_WITH_OVERHEAD(SK_MEM_QUANTUM)) {
4931 tcp_collapse(sk, NULL, &tp->out_of_order_queue,
4932 head, skb, start, end);
4934 sum_tiny += range_truesize;
4935 if (sum_tiny > sk->sk_rcvbuf >> 3)
4941 range_truesize += skb->truesize;
4942 if (unlikely(before(TCP_SKB_CB(skb)->seq, start)))
4943 start = TCP_SKB_CB(skb)->seq;
4944 if (after(TCP_SKB_CB(skb)->end_seq, end))
4945 end = TCP_SKB_CB(skb)->end_seq;
4950 * Clean the out-of-order queue to make room.
4951 * We drop high sequences packets to :
4952 * 1) Let a chance for holes to be filled.
4953 * 2) not add too big latencies if thousands of packets sit there.
4954 * (But if application shrinks SO_RCVBUF, we could still end up
4955 * freeing whole queue here)
4956 * 3) Drop at least 12.5 % of sk_rcvbuf to avoid malicious attacks.
4958 * Return true if queue has shrunk.
4960 static bool tcp_prune_ofo_queue(struct sock *sk)
4962 struct tcp_sock *tp = tcp_sk(sk);
4963 struct rb_node *node, *prev;
4966 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
4969 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED);
4970 goal = sk->sk_rcvbuf >> 3;
4971 node = &tp->ooo_last_skb->rbnode;
4973 prev = rb_prev(node);
4974 rb_erase(node, &tp->out_of_order_queue);
4975 goal -= rb_to_skb(node)->truesize;
4976 tcp_drop(sk, rb_to_skb(node));
4977 if (!prev || goal <= 0) {
4979 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf &&
4980 !tcp_under_memory_pressure(sk))
4982 goal = sk->sk_rcvbuf >> 3;
4986 tp->ooo_last_skb = rb_to_skb(prev);
4988 /* Reset SACK state. A conforming SACK implementation will
4989 * do the same at a timeout based retransmit. When a connection
4990 * is in a sad state like this, we care only about integrity
4991 * of the connection not performance.
4993 if (tp->rx_opt.sack_ok)
4994 tcp_sack_reset(&tp->rx_opt);
4998 /* Reduce allocated memory if we can, trying to get
4999 * the socket within its memory limits again.
5001 * Return less than zero if we should start dropping frames
5002 * until the socket owning process reads some of the data
5003 * to stabilize the situation.
5005 static int tcp_prune_queue(struct sock *sk)
5007 struct tcp_sock *tp = tcp_sk(sk);
5009 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
5011 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED);
5013 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
5014 tcp_clamp_window(sk);
5015 else if (tcp_under_memory_pressure(sk))
5016 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
5018 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5021 tcp_collapse_ofo_queue(sk);
5022 if (!skb_queue_empty(&sk->sk_receive_queue))
5023 tcp_collapse(sk, &sk->sk_receive_queue, NULL,
5024 skb_peek(&sk->sk_receive_queue),
5026 tp->copied_seq, tp->rcv_nxt);
5029 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5032 /* Collapsing did not help, destructive actions follow.
5033 * This must not ever occur. */
5035 tcp_prune_ofo_queue(sk);
5037 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5040 /* If we are really being abused, tell the caller to silently
5041 * drop receive data on the floor. It will get retransmitted
5042 * and hopefully then we'll have sufficient space.
5044 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED);
5046 /* Massive buffer overcommit. */
5051 static bool tcp_should_expand_sndbuf(const struct sock *sk)
5053 const struct tcp_sock *tp = tcp_sk(sk);
5055 /* If the user specified a specific send buffer setting, do
5058 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
5061 /* If we are under global TCP memory pressure, do not expand. */
5062 if (tcp_under_memory_pressure(sk))
5065 /* If we are under soft global TCP memory pressure, do not expand. */
5066 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
5069 /* If we filled the congestion window, do not expand. */
5070 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
5076 /* When incoming ACK allowed to free some skb from write_queue,
5077 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
5078 * on the exit from tcp input handler.
5080 * PROBLEM: sndbuf expansion does not work well with largesend.
5082 static void tcp_new_space(struct sock *sk)
5084 struct tcp_sock *tp = tcp_sk(sk);
5086 if (tcp_should_expand_sndbuf(sk)) {
5087 tcp_sndbuf_expand(sk);
5088 tp->snd_cwnd_stamp = tcp_jiffies32;
5091 sk->sk_write_space(sk);
5094 static void tcp_check_space(struct sock *sk)
5096 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
5097 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
5098 /* pairs with tcp_poll() */
5100 if (sk->sk_socket &&
5101 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
5103 if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5104 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
5109 static inline void tcp_data_snd_check(struct sock *sk)
5111 tcp_push_pending_frames(sk);
5112 tcp_check_space(sk);
5116 * Check if sending an ack is needed.
5118 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5120 struct tcp_sock *tp = tcp_sk(sk);
5121 unsigned long rtt, delay;
5123 /* More than one full frame received... */
5124 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5125 /* ... and right edge of window advances far enough.
5126 * (tcp_recvmsg() will send ACK otherwise).
5127 * If application uses SO_RCVLOWAT, we want send ack now if
5128 * we have not received enough bytes to satisfy the condition.
5130 (tp->rcv_nxt - tp->copied_seq < sk->sk_rcvlowat ||
5131 __tcp_select_window(sk) >= tp->rcv_wnd)) ||
5132 /* We ACK each frame or... */
5133 tcp_in_quickack_mode(sk)) {
5139 if (!ofo_possible || RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5140 tcp_send_delayed_ack(sk);
5144 if (!tcp_is_sack(tp) ||
5145 tp->compressed_ack >= sock_net(sk)->ipv4.sysctl_tcp_comp_sack_nr)
5147 tp->compressed_ack++;
5149 if (hrtimer_is_queued(&tp->compressed_ack_timer))
5152 /* compress ack timer : 5 % of rtt, but no more than tcp_comp_sack_delay_ns */
5154 rtt = tp->rcv_rtt_est.rtt_us;
5155 if (tp->srtt_us && tp->srtt_us < rtt)
5158 delay = min_t(unsigned long, sock_net(sk)->ipv4.sysctl_tcp_comp_sack_delay_ns,
5159 rtt * (NSEC_PER_USEC >> 3)/20);
5161 hrtimer_start(&tp->compressed_ack_timer, ns_to_ktime(delay),
5162 HRTIMER_MODE_REL_PINNED_SOFT);
5165 static inline void tcp_ack_snd_check(struct sock *sk)
5167 if (!inet_csk_ack_scheduled(sk)) {
5168 /* We sent a data segment already. */
5171 __tcp_ack_snd_check(sk, 1);
5175 * This routine is only called when we have urgent data
5176 * signaled. Its the 'slow' part of tcp_urg. It could be
5177 * moved inline now as tcp_urg is only called from one
5178 * place. We handle URGent data wrong. We have to - as
5179 * BSD still doesn't use the correction from RFC961.
5180 * For 1003.1g we should support a new option TCP_STDURG to permit
5181 * either form (or just set the sysctl tcp_stdurg).
5184 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5186 struct tcp_sock *tp = tcp_sk(sk);
5187 u32 ptr = ntohs(th->urg_ptr);
5189 if (ptr && !sock_net(sk)->ipv4.sysctl_tcp_stdurg)
5191 ptr += ntohl(th->seq);
5193 /* Ignore urgent data that we've already seen and read. */
5194 if (after(tp->copied_seq, ptr))
5197 /* Do not replay urg ptr.
5199 * NOTE: interesting situation not covered by specs.
5200 * Misbehaving sender may send urg ptr, pointing to segment,
5201 * which we already have in ofo queue. We are not able to fetch
5202 * such data and will stay in TCP_URG_NOTYET until will be eaten
5203 * by recvmsg(). Seems, we are not obliged to handle such wicked
5204 * situations. But it is worth to think about possibility of some
5205 * DoSes using some hypothetical application level deadlock.
5207 if (before(ptr, tp->rcv_nxt))
5210 /* Do we already have a newer (or duplicate) urgent pointer? */
5211 if (tp->urg_data && !after(ptr, tp->urg_seq))
5214 /* Tell the world about our new urgent pointer. */
5217 /* We may be adding urgent data when the last byte read was
5218 * urgent. To do this requires some care. We cannot just ignore
5219 * tp->copied_seq since we would read the last urgent byte again
5220 * as data, nor can we alter copied_seq until this data arrives
5221 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5223 * NOTE. Double Dutch. Rendering to plain English: author of comment
5224 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5225 * and expect that both A and B disappear from stream. This is _wrong_.
5226 * Though this happens in BSD with high probability, this is occasional.
5227 * Any application relying on this is buggy. Note also, that fix "works"
5228 * only in this artificial test. Insert some normal data between A and B and we will
5229 * decline of BSD again. Verdict: it is better to remove to trap
5232 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5233 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5234 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5236 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5237 __skb_unlink(skb, &sk->sk_receive_queue);
5242 tp->urg_data = TCP_URG_NOTYET;
5245 /* Disable header prediction. */
5249 /* This is the 'fast' part of urgent handling. */
5250 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5252 struct tcp_sock *tp = tcp_sk(sk);
5254 /* Check if we get a new urgent pointer - normally not. */
5256 tcp_check_urg(sk, th);
5258 /* Do we wait for any urgent data? - normally not... */
5259 if (tp->urg_data == TCP_URG_NOTYET) {
5260 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5263 /* Is the urgent pointer pointing into this packet? */
5264 if (ptr < skb->len) {
5266 if (skb_copy_bits(skb, ptr, &tmp, 1))
5268 tp->urg_data = TCP_URG_VALID | tmp;
5269 if (!sock_flag(sk, SOCK_DEAD))
5270 sk->sk_data_ready(sk);
5275 /* Accept RST for rcv_nxt - 1 after a FIN.
5276 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a
5277 * FIN is sent followed by a RST packet. The RST is sent with the same
5278 * sequence number as the FIN, and thus according to RFC 5961 a challenge
5279 * ACK should be sent. However, Mac OSX rate limits replies to challenge
5280 * ACKs on the closed socket. In addition middleboxes can drop either the
5281 * challenge ACK or a subsequent RST.
5283 static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb)
5285 struct tcp_sock *tp = tcp_sk(sk);
5287 return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) &&
5288 (1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK |
5292 /* Does PAWS and seqno based validation of an incoming segment, flags will
5293 * play significant role here.
5295 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5296 const struct tcphdr *th, int syn_inerr)
5298 struct tcp_sock *tp = tcp_sk(sk);
5299 bool rst_seq_match = false;
5301 /* RFC1323: H1. Apply PAWS check first. */
5302 if (tcp_fast_parse_options(sock_net(sk), skb, th, tp) &&
5303 tp->rx_opt.saw_tstamp &&
5304 tcp_paws_discard(sk, skb)) {
5306 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5307 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5308 LINUX_MIB_TCPACKSKIPPEDPAWS,
5309 &tp->last_oow_ack_time))
5310 tcp_send_dupack(sk, skb);
5313 /* Reset is accepted even if it did not pass PAWS. */
5316 /* Step 1: check sequence number */
5317 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5318 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5319 * (RST) segments are validated by checking their SEQ-fields."
5320 * And page 69: "If an incoming segment is not acceptable,
5321 * an acknowledgment should be sent in reply (unless the RST
5322 * bit is set, if so drop the segment and return)".
5327 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5328 LINUX_MIB_TCPACKSKIPPEDSEQ,
5329 &tp->last_oow_ack_time))
5330 tcp_send_dupack(sk, skb);
5331 } else if (tcp_reset_check(sk, skb)) {
5337 /* Step 2: check RST bit */
5339 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a
5340 * FIN and SACK too if available):
5341 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or
5342 * the right-most SACK block,
5344 * RESET the connection
5346 * Send a challenge ACK
5348 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt ||
5349 tcp_reset_check(sk, skb)) {
5350 rst_seq_match = true;
5351 } else if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) {
5352 struct tcp_sack_block *sp = &tp->selective_acks[0];
5353 int max_sack = sp[0].end_seq;
5356 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;
5358 max_sack = after(sp[this_sack].end_seq,
5360 sp[this_sack].end_seq : max_sack;
5363 if (TCP_SKB_CB(skb)->seq == max_sack)
5364 rst_seq_match = true;
5370 /* Disable TFO if RST is out-of-order
5371 * and no data has been received
5372 * for current active TFO socket
5374 if (tp->syn_fastopen && !tp->data_segs_in &&
5375 sk->sk_state == TCP_ESTABLISHED)
5376 tcp_fastopen_active_disable(sk);
5377 tcp_send_challenge_ack(sk, skb);
5382 /* step 3: check security and precedence [ignored] */
5384 /* step 4: Check for a SYN
5385 * RFC 5961 4.2 : Send a challenge ack
5390 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5391 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5392 tcp_send_challenge_ack(sk, skb);
5404 * TCP receive function for the ESTABLISHED state.
5406 * It is split into a fast path and a slow path. The fast path is
5408 * - A zero window was announced from us - zero window probing
5409 * is only handled properly in the slow path.
5410 * - Out of order segments arrived.
5411 * - Urgent data is expected.
5412 * - There is no buffer space left
5413 * - Unexpected TCP flags/window values/header lengths are received
5414 * (detected by checking the TCP header against pred_flags)
5415 * - Data is sent in both directions. Fast path only supports pure senders
5416 * or pure receivers (this means either the sequence number or the ack
5417 * value must stay constant)
5418 * - Unexpected TCP option.
5420 * When these conditions are not satisfied it drops into a standard
5421 * receive procedure patterned after RFC793 to handle all cases.
5422 * The first three cases are guaranteed by proper pred_flags setting,
5423 * the rest is checked inline. Fast processing is turned on in
5424 * tcp_data_queue when everything is OK.
5426 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb)
5428 const struct tcphdr *th = (const struct tcphdr *)skb->data;
5429 struct tcp_sock *tp = tcp_sk(sk);
5430 unsigned int len = skb->len;
5432 /* TCP congestion window tracking */
5433 trace_tcp_probe(sk, skb);
5435 tcp_mstamp_refresh(tp);
5436 if (unlikely(!sk->sk_rx_dst))
5437 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5439 * Header prediction.
5440 * The code loosely follows the one in the famous
5441 * "30 instruction TCP receive" Van Jacobson mail.
5443 * Van's trick is to deposit buffers into socket queue
5444 * on a device interrupt, to call tcp_recv function
5445 * on the receive process context and checksum and copy
5446 * the buffer to user space. smart...
5448 * Our current scheme is not silly either but we take the
5449 * extra cost of the net_bh soft interrupt processing...
5450 * We do checksum and copy also but from device to kernel.
5453 tp->rx_opt.saw_tstamp = 0;
5455 /* pred_flags is 0xS?10 << 16 + snd_wnd
5456 * if header_prediction is to be made
5457 * 'S' will always be tp->tcp_header_len >> 2
5458 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5459 * turn it off (when there are holes in the receive
5460 * space for instance)
5461 * PSH flag is ignored.
5464 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5465 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5466 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5467 int tcp_header_len = tp->tcp_header_len;
5469 /* Timestamp header prediction: tcp_header_len
5470 * is automatically equal to th->doff*4 due to pred_flags
5474 /* Check timestamp */
5475 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5476 /* No? Slow path! */
5477 if (!tcp_parse_aligned_timestamp(tp, th))
5480 /* If PAWS failed, check it more carefully in slow path */
5481 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5484 /* DO NOT update ts_recent here, if checksum fails
5485 * and timestamp was corrupted part, it will result
5486 * in a hung connection since we will drop all
5487 * future packets due to the PAWS test.
5491 if (len <= tcp_header_len) {
5492 /* Bulk data transfer: sender */
5493 if (len == tcp_header_len) {
5494 /* Predicted packet is in window by definition.
5495 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5496 * Hence, check seq<=rcv_wup reduces to:
5498 if (tcp_header_len ==
5499 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5500 tp->rcv_nxt == tp->rcv_wup)
5501 tcp_store_ts_recent(tp);
5503 /* We know that such packets are checksummed
5506 tcp_ack(sk, skb, 0);
5508 tcp_data_snd_check(sk);
5510 } else { /* Header too small */
5511 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5516 bool fragstolen = false;
5518 if (tcp_checksum_complete(skb))
5521 if ((int)skb->truesize > sk->sk_forward_alloc)
5524 /* Predicted packet is in window by definition.
5525 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5526 * Hence, check seq<=rcv_wup reduces to:
5528 if (tcp_header_len ==
5529 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5530 tp->rcv_nxt == tp->rcv_wup)
5531 tcp_store_ts_recent(tp);
5533 tcp_rcv_rtt_measure_ts(sk, skb);
5535 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS);
5537 /* Bulk data transfer: receiver */
5538 eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5541 tcp_event_data_recv(sk, skb);
5543 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5544 /* Well, only one small jumplet in fast path... */
5545 tcp_ack(sk, skb, FLAG_DATA);
5546 tcp_data_snd_check(sk);
5547 if (!inet_csk_ack_scheduled(sk))
5551 __tcp_ack_snd_check(sk, 0);
5554 kfree_skb_partial(skb, fragstolen);
5561 if (len < (th->doff << 2) || tcp_checksum_complete(skb))
5564 if (!th->ack && !th->rst && !th->syn)
5568 * Standard slow path.
5571 if (!tcp_validate_incoming(sk, skb, th, 1))
5575 if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5578 tcp_rcv_rtt_measure_ts(sk, skb);
5580 /* Process urgent data. */
5581 tcp_urg(sk, skb, th);
5583 /* step 7: process the segment text */
5584 tcp_data_queue(sk, skb);
5586 tcp_data_snd_check(sk);
5587 tcp_ack_snd_check(sk);
5591 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
5592 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5597 EXPORT_SYMBOL(tcp_rcv_established);
5599 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5601 struct tcp_sock *tp = tcp_sk(sk);
5602 struct inet_connection_sock *icsk = inet_csk(sk);
5604 tcp_set_state(sk, TCP_ESTABLISHED);
5605 icsk->icsk_ack.lrcvtime = tcp_jiffies32;
5608 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5609 security_inet_conn_established(sk, skb);
5612 tcp_init_transfer(sk, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB);
5614 /* Prevent spurious tcp_cwnd_restart() on first data
5617 tp->lsndtime = tcp_jiffies32;
5619 if (sock_flag(sk, SOCK_KEEPOPEN))
5620 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5622 if (!tp->rx_opt.snd_wscale)
5623 __tcp_fast_path_on(tp, tp->snd_wnd);
5628 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5629 struct tcp_fastopen_cookie *cookie)
5631 struct tcp_sock *tp = tcp_sk(sk);
5632 struct sk_buff *data = tp->syn_data ? tcp_rtx_queue_head(sk) : NULL;
5633 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
5634 bool syn_drop = false;
5636 if (mss == tp->rx_opt.user_mss) {
5637 struct tcp_options_received opt;
5639 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5640 tcp_clear_options(&opt);
5641 opt.user_mss = opt.mss_clamp = 0;
5642 tcp_parse_options(sock_net(sk), synack, &opt, 0, NULL);
5643 mss = opt.mss_clamp;
5646 if (!tp->syn_fastopen) {
5647 /* Ignore an unsolicited cookie */
5649 } else if (tp->total_retrans) {
5650 /* SYN timed out and the SYN-ACK neither has a cookie nor
5651 * acknowledges data. Presumably the remote received only
5652 * the retransmitted (regular) SYNs: either the original
5653 * SYN-data or the corresponding SYN-ACK was dropped.
5655 syn_drop = (cookie->len < 0 && data);
5656 } else if (cookie->len < 0 && !tp->syn_data) {
5657 /* We requested a cookie but didn't get it. If we did not use
5658 * the (old) exp opt format then try so next time (try_exp=1).
5659 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5661 try_exp = tp->syn_fastopen_exp ? 2 : 1;
5664 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
5666 if (data) { /* Retransmit unacked data in SYN */
5667 skb_rbtree_walk_from(data) {
5668 if (__tcp_retransmit_skb(sk, data, 1))
5672 NET_INC_STATS(sock_net(sk),
5673 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
5676 tp->syn_data_acked = tp->syn_data;
5677 if (tp->syn_data_acked) {
5678 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
5679 /* SYN-data is counted as two separate packets in tcp_ack() */
5680 if (tp->delivered > 1)
5684 tcp_fastopen_add_skb(sk, synack);
5689 static void smc_check_reset_syn(struct tcp_sock *tp)
5691 #if IS_ENABLED(CONFIG_SMC)
5692 if (static_branch_unlikely(&tcp_have_smc)) {
5693 if (tp->syn_smc && !tp->rx_opt.smc_ok)
5699 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5700 const struct tcphdr *th)
5702 struct inet_connection_sock *icsk = inet_csk(sk);
5703 struct tcp_sock *tp = tcp_sk(sk);
5704 struct tcp_fastopen_cookie foc = { .len = -1 };
5705 int saved_clamp = tp->rx_opt.mss_clamp;
5708 tcp_parse_options(sock_net(sk), skb, &tp->rx_opt, 0, &foc);
5709 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
5710 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
5714 * "If the state is SYN-SENT then
5715 * first check the ACK bit
5716 * If the ACK bit is set
5717 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5718 * a reset (unless the RST bit is set, if so drop
5719 * the segment and return)"
5721 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5722 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
5723 goto reset_and_undo;
5725 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5726 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5727 tcp_time_stamp(tp))) {
5728 NET_INC_STATS(sock_net(sk),
5729 LINUX_MIB_PAWSACTIVEREJECTED);
5730 goto reset_and_undo;
5733 /* Now ACK is acceptable.
5735 * "If the RST bit is set
5736 * If the ACK was acceptable then signal the user "error:
5737 * connection reset", drop the segment, enter CLOSED state,
5738 * delete TCB, and return."
5747 * "fifth, if neither of the SYN or RST bits is set then
5748 * drop the segment and return."
5754 goto discard_and_undo;
5757 * "If the SYN bit is on ...
5758 * are acceptable then ...
5759 * (our SYN has been ACKed), change the connection
5760 * state to ESTABLISHED..."
5763 tcp_ecn_rcv_synack(tp, th);
5765 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5766 tcp_ack(sk, skb, FLAG_SLOWPATH);
5768 /* Ok.. it's good. Set up sequence numbers and
5769 * move to established.
5771 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5772 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5774 /* RFC1323: The window in SYN & SYN/ACK segments is
5777 tp->snd_wnd = ntohs(th->window);
5779 if (!tp->rx_opt.wscale_ok) {
5780 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5781 tp->window_clamp = min(tp->window_clamp, 65535U);
5784 if (tp->rx_opt.saw_tstamp) {
5785 tp->rx_opt.tstamp_ok = 1;
5786 tp->tcp_header_len =
5787 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5788 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5789 tcp_store_ts_recent(tp);
5791 tp->tcp_header_len = sizeof(struct tcphdr);
5794 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5795 tcp_initialize_rcv_mss(sk);
5797 /* Remember, tcp_poll() does not lock socket!
5798 * Change state from SYN-SENT only after copied_seq
5799 * is initialized. */
5800 tp->copied_seq = tp->rcv_nxt;
5802 smc_check_reset_syn(tp);
5806 tcp_finish_connect(sk, skb);
5808 fastopen_fail = (tp->syn_fastopen || tp->syn_data) &&
5809 tcp_rcv_fastopen_synack(sk, skb, &foc);
5811 if (!sock_flag(sk, SOCK_DEAD)) {
5812 sk->sk_state_change(sk);
5813 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5817 if (sk->sk_write_pending ||
5818 icsk->icsk_accept_queue.rskq_defer_accept ||
5819 icsk->icsk_ack.pingpong) {
5820 /* Save one ACK. Data will be ready after
5821 * several ticks, if write_pending is set.
5823 * It may be deleted, but with this feature tcpdumps
5824 * look so _wonderfully_ clever, that I was not able
5825 * to stand against the temptation 8) --ANK
5827 inet_csk_schedule_ack(sk);
5828 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
5829 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5830 TCP_DELACK_MAX, TCP_RTO_MAX);
5841 /* No ACK in the segment */
5845 * "If the RST bit is set
5847 * Otherwise (no ACK) drop the segment and return."
5850 goto discard_and_undo;
5854 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5855 tcp_paws_reject(&tp->rx_opt, 0))
5856 goto discard_and_undo;
5859 /* We see SYN without ACK. It is attempt of
5860 * simultaneous connect with crossed SYNs.
5861 * Particularly, it can be connect to self.
5863 tcp_set_state(sk, TCP_SYN_RECV);
5865 if (tp->rx_opt.saw_tstamp) {
5866 tp->rx_opt.tstamp_ok = 1;
5867 tcp_store_ts_recent(tp);
5868 tp->tcp_header_len =
5869 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5871 tp->tcp_header_len = sizeof(struct tcphdr);
5874 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5875 tp->copied_seq = tp->rcv_nxt;
5876 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5878 /* RFC1323: The window in SYN & SYN/ACK segments is
5881 tp->snd_wnd = ntohs(th->window);
5882 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5883 tp->max_window = tp->snd_wnd;
5885 tcp_ecn_rcv_syn(tp, th);
5888 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5889 tcp_initialize_rcv_mss(sk);
5891 tcp_send_synack(sk);
5893 /* Note, we could accept data and URG from this segment.
5894 * There are no obstacles to make this (except that we must
5895 * either change tcp_recvmsg() to prevent it from returning data
5896 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5898 * However, if we ignore data in ACKless segments sometimes,
5899 * we have no reasons to accept it sometimes.
5900 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5901 * is not flawless. So, discard packet for sanity.
5902 * Uncomment this return to process the data.
5909 /* "fifth, if neither of the SYN or RST bits is set then
5910 * drop the segment and return."
5914 tcp_clear_options(&tp->rx_opt);
5915 tp->rx_opt.mss_clamp = saved_clamp;
5919 tcp_clear_options(&tp->rx_opt);
5920 tp->rx_opt.mss_clamp = saved_clamp;
5925 * This function implements the receiving procedure of RFC 793 for
5926 * all states except ESTABLISHED and TIME_WAIT.
5927 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5928 * address independent.
5931 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
5933 struct tcp_sock *tp = tcp_sk(sk);
5934 struct inet_connection_sock *icsk = inet_csk(sk);
5935 const struct tcphdr *th = tcp_hdr(skb);
5936 struct request_sock *req;
5940 switch (sk->sk_state) {
5954 /* It is possible that we process SYN packets from backlog,
5955 * so we need to make sure to disable BH right there.
5958 acceptable = icsk->icsk_af_ops->conn_request(sk, skb) >= 0;
5969 tp->rx_opt.saw_tstamp = 0;
5970 tcp_mstamp_refresh(tp);
5971 queued = tcp_rcv_synsent_state_process(sk, skb, th);
5975 /* Do step6 onward by hand. */
5976 tcp_urg(sk, skb, th);
5978 tcp_data_snd_check(sk);
5982 tcp_mstamp_refresh(tp);
5983 tp->rx_opt.saw_tstamp = 0;
5984 req = tp->fastopen_rsk;
5988 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
5989 sk->sk_state != TCP_FIN_WAIT1);
5991 if (!tcp_check_req(sk, skb, req, true, &req_stolen))
5995 if (!th->ack && !th->rst && !th->syn)
5998 if (!tcp_validate_incoming(sk, skb, th, 0))
6001 /* step 5: check the ACK field */
6002 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
6003 FLAG_UPDATE_TS_RECENT |
6004 FLAG_NO_CHALLENGE_ACK) > 0;
6007 if (sk->sk_state == TCP_SYN_RECV)
6008 return 1; /* send one RST */
6009 tcp_send_challenge_ack(sk, skb);
6012 switch (sk->sk_state) {
6014 tp->delivered++; /* SYN-ACK delivery isn't tracked in tcp_ack */
6016 tcp_synack_rtt_meas(sk, req);
6018 /* Once we leave TCP_SYN_RECV, we no longer need req
6022 inet_csk(sk)->icsk_retransmits = 0;
6023 reqsk_fastopen_remove(sk, req, false);
6024 /* Re-arm the timer because data may have been sent out.
6025 * This is similar to the regular data transmission case
6026 * when new data has just been ack'ed.
6028 * (TFO) - we could try to be more aggressive and
6029 * retransmitting any data sooner based on when they
6034 tcp_init_transfer(sk, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB);
6035 tp->copied_seq = tp->rcv_nxt;
6038 tcp_set_state(sk, TCP_ESTABLISHED);
6039 sk->sk_state_change(sk);
6041 /* Note, that this wakeup is only for marginal crossed SYN case.
6042 * Passively open sockets are not waked up, because
6043 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
6046 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6048 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
6049 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
6050 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6052 if (tp->rx_opt.tstamp_ok)
6053 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6055 if (!inet_csk(sk)->icsk_ca_ops->cong_control)
6056 tcp_update_pacing_rate(sk);
6058 /* Prevent spurious tcp_cwnd_restart() on first data packet */
6059 tp->lsndtime = tcp_jiffies32;
6061 tcp_initialize_rcv_mss(sk);
6062 tcp_fast_path_on(tp);
6065 case TCP_FIN_WAIT1: {
6068 /* If we enter the TCP_FIN_WAIT1 state and we are a
6069 * Fast Open socket and this is the first acceptable
6070 * ACK we have received, this would have acknowledged
6071 * our SYNACK so stop the SYNACK timer.
6074 /* We no longer need the request sock. */
6075 reqsk_fastopen_remove(sk, req, false);
6078 if (tp->snd_una != tp->write_seq)
6081 tcp_set_state(sk, TCP_FIN_WAIT2);
6082 sk->sk_shutdown |= SEND_SHUTDOWN;
6086 if (!sock_flag(sk, SOCK_DEAD)) {
6087 /* Wake up lingering close() */
6088 sk->sk_state_change(sk);
6092 if (tp->linger2 < 0) {
6094 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6097 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6098 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6099 /* Receive out of order FIN after close() */
6100 if (tp->syn_fastopen && th->fin)
6101 tcp_fastopen_active_disable(sk);
6103 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6107 tmo = tcp_fin_time(sk);
6108 if (tmo > TCP_TIMEWAIT_LEN) {
6109 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
6110 } else if (th->fin || sock_owned_by_user(sk)) {
6111 /* Bad case. We could lose such FIN otherwise.
6112 * It is not a big problem, but it looks confusing
6113 * and not so rare event. We still can lose it now,
6114 * if it spins in bh_lock_sock(), but it is really
6117 inet_csk_reset_keepalive_timer(sk, tmo);
6119 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
6126 if (tp->snd_una == tp->write_seq) {
6127 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6133 if (tp->snd_una == tp->write_seq) {
6134 tcp_update_metrics(sk);
6141 /* step 6: check the URG bit */
6142 tcp_urg(sk, skb, th);
6144 /* step 7: process the segment text */
6145 switch (sk->sk_state) {
6146 case TCP_CLOSE_WAIT:
6149 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
6154 /* RFC 793 says to queue data in these states,
6155 * RFC 1122 says we MUST send a reset.
6156 * BSD 4.4 also does reset.
6158 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6159 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6160 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6161 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6167 case TCP_ESTABLISHED:
6168 tcp_data_queue(sk, skb);
6173 /* tcp_data could move socket to TIME-WAIT */
6174 if (sk->sk_state != TCP_CLOSE) {
6175 tcp_data_snd_check(sk);
6176 tcp_ack_snd_check(sk);
6185 EXPORT_SYMBOL(tcp_rcv_state_process);
6187 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
6189 struct inet_request_sock *ireq = inet_rsk(req);
6191 if (family == AF_INET)
6192 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6193 &ireq->ir_rmt_addr, port);
6194 #if IS_ENABLED(CONFIG_IPV6)
6195 else if (family == AF_INET6)
6196 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6197 &ireq->ir_v6_rmt_addr, port);
6201 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6203 * If we receive a SYN packet with these bits set, it means a
6204 * network is playing bad games with TOS bits. In order to
6205 * avoid possible false congestion notifications, we disable
6206 * TCP ECN negotiation.
6208 * Exception: tcp_ca wants ECN. This is required for DCTCP
6209 * congestion control: Linux DCTCP asserts ECT on all packets,
6210 * including SYN, which is most optimal solution; however,
6211 * others, such as FreeBSD do not.
6213 static void tcp_ecn_create_request(struct request_sock *req,
6214 const struct sk_buff *skb,
6215 const struct sock *listen_sk,
6216 const struct dst_entry *dst)
6218 const struct tcphdr *th = tcp_hdr(skb);
6219 const struct net *net = sock_net(listen_sk);
6220 bool th_ecn = th->ece && th->cwr;
6227 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
6228 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK);
6229 ecn_ok = net->ipv4.sysctl_tcp_ecn || ecn_ok_dst;
6231 if ((!ect && ecn_ok) || tcp_ca_needs_ecn(listen_sk) ||
6232 (ecn_ok_dst & DST_FEATURE_ECN_CA) ||
6233 tcp_bpf_ca_needs_ecn((struct sock *)req))
6234 inet_rsk(req)->ecn_ok = 1;
6237 static void tcp_openreq_init(struct request_sock *req,
6238 const struct tcp_options_received *rx_opt,
6239 struct sk_buff *skb, const struct sock *sk)
6241 struct inet_request_sock *ireq = inet_rsk(req);
6243 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */
6245 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6246 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
6247 tcp_rsk(req)->snt_synack = tcp_clock_us();
6248 tcp_rsk(req)->last_oow_ack_time = 0;
6249 req->mss = rx_opt->mss_clamp;
6250 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6251 ireq->tstamp_ok = rx_opt->tstamp_ok;
6252 ireq->sack_ok = rx_opt->sack_ok;
6253 ireq->snd_wscale = rx_opt->snd_wscale;
6254 ireq->wscale_ok = rx_opt->wscale_ok;
6257 ireq->ir_rmt_port = tcp_hdr(skb)->source;
6258 ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6259 ireq->ir_mark = inet_request_mark(sk, skb);
6260 #if IS_ENABLED(CONFIG_SMC)
6261 ireq->smc_ok = rx_opt->smc_ok;
6265 struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
6266 struct sock *sk_listener,
6267 bool attach_listener)
6269 struct request_sock *req = reqsk_alloc(ops, sk_listener,
6273 struct inet_request_sock *ireq = inet_rsk(req);
6275 ireq->ireq_opt = NULL;
6276 #if IS_ENABLED(CONFIG_IPV6)
6277 ireq->pktopts = NULL;
6279 atomic64_set(&ireq->ir_cookie, 0);
6280 ireq->ireq_state = TCP_NEW_SYN_RECV;
6281 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
6282 ireq->ireq_family = sk_listener->sk_family;
6287 EXPORT_SYMBOL(inet_reqsk_alloc);
6290 * Return true if a syncookie should be sent
6292 static bool tcp_syn_flood_action(const struct sock *sk,
6293 const struct sk_buff *skb,
6296 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
6297 const char *msg = "Dropping request";
6298 bool want_cookie = false;
6299 struct net *net = sock_net(sk);
6301 #ifdef CONFIG_SYN_COOKIES
6302 if (net->ipv4.sysctl_tcp_syncookies) {
6303 msg = "Sending cookies";
6305 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
6308 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
6310 if (!queue->synflood_warned &&
6311 net->ipv4.sysctl_tcp_syncookies != 2 &&
6312 xchg(&queue->synflood_warned, 1) == 0)
6313 pr_info("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6314 proto, ntohs(tcp_hdr(skb)->dest), msg);
6319 static void tcp_reqsk_record_syn(const struct sock *sk,
6320 struct request_sock *req,
6321 const struct sk_buff *skb)
6323 if (tcp_sk(sk)->save_syn) {
6324 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
6327 copy = kmalloc(len + sizeof(u32), GFP_ATOMIC);
6330 memcpy(©[1], skb_network_header(skb), len);
6331 req->saved_syn = copy;
6336 int tcp_conn_request(struct request_sock_ops *rsk_ops,
6337 const struct tcp_request_sock_ops *af_ops,
6338 struct sock *sk, struct sk_buff *skb)
6340 struct tcp_fastopen_cookie foc = { .len = -1 };
6341 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
6342 struct tcp_options_received tmp_opt;
6343 struct tcp_sock *tp = tcp_sk(sk);
6344 struct net *net = sock_net(sk);
6345 struct sock *fastopen_sk = NULL;
6346 struct request_sock *req;
6347 bool want_cookie = false;
6348 struct dst_entry *dst;
6351 /* TW buckets are converted to open requests without
6352 * limitations, they conserve resources and peer is
6353 * evidently real one.
6355 if ((net->ipv4.sysctl_tcp_syncookies == 2 ||
6356 inet_csk_reqsk_queue_is_full(sk)) && !isn) {
6357 want_cookie = tcp_syn_flood_action(sk, skb, rsk_ops->slab_name);
6362 if (sk_acceptq_is_full(sk)) {
6363 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6367 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie);
6371 tcp_rsk(req)->af_specific = af_ops;
6372 tcp_rsk(req)->ts_off = 0;
6374 tcp_clear_options(&tmp_opt);
6375 tmp_opt.mss_clamp = af_ops->mss_clamp;
6376 tmp_opt.user_mss = tp->rx_opt.user_mss;
6377 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0,
6378 want_cookie ? NULL : &foc);
6380 if (want_cookie && !tmp_opt.saw_tstamp)
6381 tcp_clear_options(&tmp_opt);
6383 if (IS_ENABLED(CONFIG_SMC) && want_cookie)
6386 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
6387 tcp_openreq_init(req, &tmp_opt, skb, sk);
6388 inet_rsk(req)->no_srccheck = inet_sk(sk)->transparent;
6390 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6391 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb);
6393 af_ops->init_req(req, sk, skb);
6395 if (security_inet_conn_request(sk, skb, req))
6398 if (tmp_opt.tstamp_ok)
6399 tcp_rsk(req)->ts_off = af_ops->init_ts_off(net, skb);
6401 dst = af_ops->route_req(sk, &fl, req);
6405 if (!want_cookie && !isn) {
6406 /* Kill the following clause, if you dislike this way. */
6407 if (!net->ipv4.sysctl_tcp_syncookies &&
6408 (net->ipv4.sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
6409 (net->ipv4.sysctl_max_syn_backlog >> 2)) &&
6410 !tcp_peer_is_proven(req, dst)) {
6411 /* Without syncookies last quarter of
6412 * backlog is filled with destinations,
6413 * proven to be alive.
6414 * It means that we continue to communicate
6415 * to destinations, already remembered
6416 * to the moment of synflood.
6418 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
6420 goto drop_and_release;
6423 isn = af_ops->init_seq(skb);
6426 tcp_ecn_create_request(req, skb, sk, dst);
6429 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
6430 req->cookie_ts = tmp_opt.tstamp_ok;
6431 if (!tmp_opt.tstamp_ok)
6432 inet_rsk(req)->ecn_ok = 0;
6435 tcp_rsk(req)->snt_isn = isn;
6436 tcp_rsk(req)->txhash = net_tx_rndhash();
6437 tcp_openreq_init_rwin(req, sk, dst);
6439 tcp_reqsk_record_syn(sk, req, skb);
6440 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst);
6443 af_ops->send_synack(fastopen_sk, dst, &fl, req,
6444 &foc, TCP_SYNACK_FASTOPEN);
6445 /* Add the child socket directly into the accept queue */
6446 inet_csk_reqsk_queue_add(sk, req, fastopen_sk);
6447 sk->sk_data_ready(sk);
6448 bh_unlock_sock(fastopen_sk);
6449 sock_put(fastopen_sk);
6451 tcp_rsk(req)->tfo_listener = false;
6453 inet_csk_reqsk_queue_hash_add(sk, req,
6454 tcp_timeout_init((struct sock *)req));
6455 af_ops->send_synack(sk, dst, &fl, req, &foc,
6456 !want_cookie ? TCP_SYNACK_NORMAL :
6474 EXPORT_SYMBOL(tcp_conn_request);