2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
23 * Pedro Roque : Fast Retransmit/Recovery.
25 * Retransmit queue handled by TCP.
26 * Better retransmit timer handling.
27 * New congestion avoidance.
31 * Eric : Fast Retransmit.
32 * Randy Scott : MSS option defines.
33 * Eric Schenk : Fixes to slow start algorithm.
34 * Eric Schenk : Yet another double ACK bug.
35 * Eric Schenk : Delayed ACK bug fixes.
36 * Eric Schenk : Floyd style fast retrans war avoidance.
37 * David S. Miller : Don't allow zero congestion window.
38 * Eric Schenk : Fix retransmitter so that it sends
39 * next packet on ack of previous packet.
40 * Andi Kleen : Moved open_request checking here
41 * and process RSTs for open_requests.
42 * Andi Kleen : Better prune_queue, and other fixes.
43 * Andrey Savochkin: Fix RTT measurements in the presence of
45 * Andrey Savochkin: Check sequence numbers correctly when
46 * removing SACKs due to in sequence incoming
48 * Andi Kleen: Make sure we never ack data there is not
49 * enough room for. Also make this condition
50 * a fatal error if it might still happen.
51 * Andi Kleen: Add tcp_measure_rcv_mss to make
52 * connections with MSS<min(MTU,ann. MSS)
53 * work without delayed acks.
54 * Andi Kleen: Process packets with PSH set in the
56 * J Hadi Salim: ECN support
59 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
60 * engine. Lots of bugs are found.
61 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
64 #define pr_fmt(fmt) "TCP: " fmt
67 #include <linux/slab.h>
68 #include <linux/module.h>
69 #include <linux/sysctl.h>
70 #include <linux/kernel.h>
71 #include <linux/prefetch.h>
74 #include <net/inet_common.h>
75 #include <linux/ipsec.h>
76 #include <asm/unaligned.h>
77 #include <linux/errqueue.h>
79 int sysctl_tcp_timestamps __read_mostly = 1;
80 int sysctl_tcp_window_scaling __read_mostly = 1;
81 int sysctl_tcp_sack __read_mostly = 1;
82 int sysctl_tcp_fack __read_mostly = 1;
83 int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
84 int sysctl_tcp_max_reordering __read_mostly = 300;
85 EXPORT_SYMBOL(sysctl_tcp_reordering);
86 int sysctl_tcp_dsack __read_mostly = 1;
87 int sysctl_tcp_app_win __read_mostly = 31;
88 int sysctl_tcp_adv_win_scale __read_mostly = 1;
89 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
91 /* rfc5961 challenge ack rate limiting */
92 int sysctl_tcp_challenge_ack_limit = 100;
94 int sysctl_tcp_stdurg __read_mostly;
95 int sysctl_tcp_rfc1337 __read_mostly;
96 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
97 int sysctl_tcp_frto __read_mostly = 2;
99 int sysctl_tcp_thin_dupack __read_mostly;
101 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
102 int sysctl_tcp_early_retrans __read_mostly = 3;
103 int sysctl_tcp_invalid_ratelimit __read_mostly = HZ/2;
105 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
106 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
107 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
108 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
109 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
110 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
111 #define FLAG_ECE 0x40 /* ECE in this ACK */
112 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
113 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
114 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
115 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
116 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
117 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
119 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
120 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
121 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
122 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
124 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
125 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
127 /* Adapt the MSS value used to make delayed ack decision to the
130 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
132 struct inet_connection_sock *icsk = inet_csk(sk);
133 const unsigned int lss = icsk->icsk_ack.last_seg_size;
136 icsk->icsk_ack.last_seg_size = 0;
138 /* skb->len may jitter because of SACKs, even if peer
139 * sends good full-sized frames.
141 len = skb_shinfo(skb)->gso_size ? : skb->len;
142 if (len >= icsk->icsk_ack.rcv_mss) {
143 icsk->icsk_ack.rcv_mss = len;
145 /* Otherwise, we make more careful check taking into account,
146 * that SACKs block is variable.
148 * "len" is invariant segment length, including TCP header.
150 len += skb->data - skb_transport_header(skb);
151 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
152 /* If PSH is not set, packet should be
153 * full sized, provided peer TCP is not badly broken.
154 * This observation (if it is correct 8)) allows
155 * to handle super-low mtu links fairly.
157 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
158 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
159 /* Subtract also invariant (if peer is RFC compliant),
160 * tcp header plus fixed timestamp option length.
161 * Resulting "len" is MSS free of SACK jitter.
163 len -= tcp_sk(sk)->tcp_header_len;
164 icsk->icsk_ack.last_seg_size = len;
166 icsk->icsk_ack.rcv_mss = len;
170 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
171 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
172 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
176 static void tcp_incr_quickack(struct sock *sk)
178 struct inet_connection_sock *icsk = inet_csk(sk);
179 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
183 if (quickacks > icsk->icsk_ack.quick)
184 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
187 static void tcp_enter_quickack_mode(struct sock *sk)
189 struct inet_connection_sock *icsk = inet_csk(sk);
190 tcp_incr_quickack(sk);
191 icsk->icsk_ack.pingpong = 0;
192 icsk->icsk_ack.ato = TCP_ATO_MIN;
195 /* Send ACKs quickly, if "quick" count is not exhausted
196 * and the session is not interactive.
199 static inline bool tcp_in_quickack_mode(const struct sock *sk)
201 const struct inet_connection_sock *icsk = inet_csk(sk);
203 return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
206 static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
208 if (tp->ecn_flags & TCP_ECN_OK)
209 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
212 static void tcp_ecn_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
214 if (tcp_hdr(skb)->cwr)
215 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
218 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
220 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
223 static void __tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
225 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
226 case INET_ECN_NOT_ECT:
227 /* Funny extension: if ECT is not set on a segment,
228 * and we already seen ECT on a previous segment,
229 * it is probably a retransmit.
231 if (tp->ecn_flags & TCP_ECN_SEEN)
232 tcp_enter_quickack_mode((struct sock *)tp);
235 if (tcp_ca_needs_ecn((struct sock *)tp))
236 tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_IS_CE);
238 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
239 /* Better not delay acks, sender can have a very low cwnd */
240 tcp_enter_quickack_mode((struct sock *)tp);
241 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
243 tp->ecn_flags |= TCP_ECN_SEEN;
246 if (tcp_ca_needs_ecn((struct sock *)tp))
247 tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_NO_CE);
248 tp->ecn_flags |= TCP_ECN_SEEN;
253 static void tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
255 if (tp->ecn_flags & TCP_ECN_OK)
256 __tcp_ecn_check_ce(tp, skb);
259 static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
261 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
262 tp->ecn_flags &= ~TCP_ECN_OK;
265 static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
267 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
268 tp->ecn_flags &= ~TCP_ECN_OK;
271 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
273 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
278 /* Buffer size and advertised window tuning.
280 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
283 static void tcp_sndbuf_expand(struct sock *sk)
285 const struct tcp_sock *tp = tcp_sk(sk);
289 /* Worst case is non GSO/TSO : each frame consumes one skb
290 * and skb->head is kmalloced using power of two area of memory
292 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
294 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
296 per_mss = roundup_pow_of_two(per_mss) +
297 SKB_DATA_ALIGN(sizeof(struct sk_buff));
299 nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd);
300 nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
302 /* Fast Recovery (RFC 5681 3.2) :
303 * Cubic needs 1.7 factor, rounded to 2 to include
304 * extra cushion (application might react slowly to POLLOUT)
306 sndmem = 2 * nr_segs * per_mss;
308 if (sk->sk_sndbuf < sndmem)
309 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
312 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
314 * All tcp_full_space() is split to two parts: "network" buffer, allocated
315 * forward and advertised in receiver window (tp->rcv_wnd) and
316 * "application buffer", required to isolate scheduling/application
317 * latencies from network.
318 * window_clamp is maximal advertised window. It can be less than
319 * tcp_full_space(), in this case tcp_full_space() - window_clamp
320 * is reserved for "application" buffer. The less window_clamp is
321 * the smoother our behaviour from viewpoint of network, but the lower
322 * throughput and the higher sensitivity of the connection to losses. 8)
324 * rcv_ssthresh is more strict window_clamp used at "slow start"
325 * phase to predict further behaviour of this connection.
326 * It is used for two goals:
327 * - to enforce header prediction at sender, even when application
328 * requires some significant "application buffer". It is check #1.
329 * - to prevent pruning of receive queue because of misprediction
330 * of receiver window. Check #2.
332 * The scheme does not work when sender sends good segments opening
333 * window and then starts to feed us spaghetti. But it should work
334 * in common situations. Otherwise, we have to rely on queue collapsing.
337 /* Slow part of check#2. */
338 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
340 struct tcp_sock *tp = tcp_sk(sk);
342 int truesize = tcp_win_from_space(skb->truesize) >> 1;
343 int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
345 while (tp->rcv_ssthresh <= window) {
346 if (truesize <= skb->len)
347 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
355 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
357 struct tcp_sock *tp = tcp_sk(sk);
360 if (tp->rcv_ssthresh < tp->window_clamp &&
361 (int)tp->rcv_ssthresh < tcp_space(sk) &&
362 !tcp_under_memory_pressure(sk)) {
365 /* Check #2. Increase window, if skb with such overhead
366 * will fit to rcvbuf in future.
368 if (tcp_win_from_space(skb->truesize) <= skb->len)
369 incr = 2 * tp->advmss;
371 incr = __tcp_grow_window(sk, skb);
374 incr = max_t(int, incr, 2 * skb->len);
375 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
377 inet_csk(sk)->icsk_ack.quick |= 1;
382 /* 3. Tuning rcvbuf, when connection enters established state. */
383 static void tcp_fixup_rcvbuf(struct sock *sk)
385 u32 mss = tcp_sk(sk)->advmss;
388 rcvmem = 2 * SKB_TRUESIZE(mss + MAX_TCP_HEADER) *
389 tcp_default_init_rwnd(mss);
391 /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
392 * Allow enough cushion so that sender is not limited by our window
394 if (sysctl_tcp_moderate_rcvbuf)
397 if (sk->sk_rcvbuf < rcvmem)
398 sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]);
401 /* 4. Try to fixup all. It is made immediately after connection enters
404 void tcp_init_buffer_space(struct sock *sk)
406 struct tcp_sock *tp = tcp_sk(sk);
409 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
410 tcp_fixup_rcvbuf(sk);
411 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
412 tcp_sndbuf_expand(sk);
414 tp->rcvq_space.space = tp->rcv_wnd;
415 tp->rcvq_space.time = tcp_time_stamp;
416 tp->rcvq_space.seq = tp->copied_seq;
418 maxwin = tcp_full_space(sk);
420 if (tp->window_clamp >= maxwin) {
421 tp->window_clamp = maxwin;
423 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
424 tp->window_clamp = max(maxwin -
425 (maxwin >> sysctl_tcp_app_win),
429 /* Force reservation of one segment. */
430 if (sysctl_tcp_app_win &&
431 tp->window_clamp > 2 * tp->advmss &&
432 tp->window_clamp + tp->advmss > maxwin)
433 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
435 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
436 tp->snd_cwnd_stamp = tcp_time_stamp;
439 /* 5. Recalculate window clamp after socket hit its memory bounds. */
440 static void tcp_clamp_window(struct sock *sk)
442 struct tcp_sock *tp = tcp_sk(sk);
443 struct inet_connection_sock *icsk = inet_csk(sk);
445 icsk->icsk_ack.quick = 0;
447 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
448 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
449 !tcp_under_memory_pressure(sk) &&
450 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
451 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
454 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
455 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
458 /* Initialize RCV_MSS value.
459 * RCV_MSS is an our guess about MSS used by the peer.
460 * We haven't any direct information about the MSS.
461 * It's better to underestimate the RCV_MSS rather than overestimate.
462 * Overestimations make us ACKing less frequently than needed.
463 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
465 void tcp_initialize_rcv_mss(struct sock *sk)
467 const struct tcp_sock *tp = tcp_sk(sk);
468 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
470 hint = min(hint, tp->rcv_wnd / 2);
471 hint = min(hint, TCP_MSS_DEFAULT);
472 hint = max(hint, TCP_MIN_MSS);
474 inet_csk(sk)->icsk_ack.rcv_mss = hint;
476 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
478 /* Receiver "autotuning" code.
480 * The algorithm for RTT estimation w/o timestamps is based on
481 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
482 * <http://public.lanl.gov/radiant/pubs.html#DRS>
484 * More detail on this code can be found at
485 * <http://staff.psc.edu/jheffner/>,
486 * though this reference is out of date. A new paper
489 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
491 u32 new_sample = tp->rcv_rtt_est.rtt;
497 if (new_sample != 0) {
498 /* If we sample in larger samples in the non-timestamp
499 * case, we could grossly overestimate the RTT especially
500 * with chatty applications or bulk transfer apps which
501 * are stalled on filesystem I/O.
503 * Also, since we are only going for a minimum in the
504 * non-timestamp case, we do not smooth things out
505 * else with timestamps disabled convergence takes too
509 m -= (new_sample >> 3);
517 /* No previous measure. */
521 if (tp->rcv_rtt_est.rtt != new_sample)
522 tp->rcv_rtt_est.rtt = new_sample;
525 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
527 if (tp->rcv_rtt_est.time == 0)
529 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
531 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rcv_rtt_est.time, 1);
534 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
535 tp->rcv_rtt_est.time = tcp_time_stamp;
538 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
539 const struct sk_buff *skb)
541 struct tcp_sock *tp = tcp_sk(sk);
542 if (tp->rx_opt.rcv_tsecr &&
543 (TCP_SKB_CB(skb)->end_seq -
544 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
545 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
549 * This function should be called every time data is copied to user space.
550 * It calculates the appropriate TCP receive buffer space.
552 void tcp_rcv_space_adjust(struct sock *sk)
554 struct tcp_sock *tp = tcp_sk(sk);
558 time = tcp_time_stamp - tp->rcvq_space.time;
559 if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
562 /* Number of bytes copied to user in last RTT */
563 copied = tp->copied_seq - tp->rcvq_space.seq;
564 if (copied <= tp->rcvq_space.space)
568 * copied = bytes received in previous RTT, our base window
569 * To cope with packet losses, we need a 2x factor
570 * To cope with slow start, and sender growing its cwin by 100 %
571 * every RTT, we need a 4x factor, because the ACK we are sending
572 * now is for the next RTT, not the current one :
573 * <prev RTT . ><current RTT .. ><next RTT .... >
576 if (sysctl_tcp_moderate_rcvbuf &&
577 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
578 int rcvwin, rcvmem, rcvbuf;
580 /* minimal window to cope with packet losses, assuming
581 * steady state. Add some cushion because of small variations.
583 rcvwin = (copied << 1) + 16 * tp->advmss;
585 /* If rate increased by 25%,
586 * assume slow start, rcvwin = 3 * copied
587 * If rate increased by 50%,
588 * assume sender can use 2x growth, rcvwin = 4 * copied
591 tp->rcvq_space.space + (tp->rcvq_space.space >> 2)) {
593 tp->rcvq_space.space + (tp->rcvq_space.space >> 1))
596 rcvwin += (rcvwin >> 1);
599 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
600 while (tcp_win_from_space(rcvmem) < tp->advmss)
603 rcvbuf = min(rcvwin / tp->advmss * rcvmem, sysctl_tcp_rmem[2]);
604 if (rcvbuf > sk->sk_rcvbuf) {
605 sk->sk_rcvbuf = rcvbuf;
607 /* Make the window clamp follow along. */
608 tp->window_clamp = rcvwin;
611 tp->rcvq_space.space = copied;
614 tp->rcvq_space.seq = tp->copied_seq;
615 tp->rcvq_space.time = tcp_time_stamp;
618 /* There is something which you must keep in mind when you analyze the
619 * behavior of the tp->ato delayed ack timeout interval. When a
620 * connection starts up, we want to ack as quickly as possible. The
621 * problem is that "good" TCP's do slow start at the beginning of data
622 * transmission. The means that until we send the first few ACK's the
623 * sender will sit on his end and only queue most of his data, because
624 * he can only send snd_cwnd unacked packets at any given time. For
625 * each ACK we send, he increments snd_cwnd and transmits more of his
628 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
630 struct tcp_sock *tp = tcp_sk(sk);
631 struct inet_connection_sock *icsk = inet_csk(sk);
634 inet_csk_schedule_ack(sk);
636 tcp_measure_rcv_mss(sk, skb);
638 tcp_rcv_rtt_measure(tp);
640 now = tcp_time_stamp;
642 if (!icsk->icsk_ack.ato) {
643 /* The _first_ data packet received, initialize
644 * delayed ACK engine.
646 tcp_incr_quickack(sk);
647 icsk->icsk_ack.ato = TCP_ATO_MIN;
649 int m = now - icsk->icsk_ack.lrcvtime;
651 if (m <= TCP_ATO_MIN / 2) {
652 /* The fastest case is the first. */
653 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
654 } else if (m < icsk->icsk_ack.ato) {
655 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
656 if (icsk->icsk_ack.ato > icsk->icsk_rto)
657 icsk->icsk_ack.ato = icsk->icsk_rto;
658 } else if (m > icsk->icsk_rto) {
659 /* Too long gap. Apparently sender failed to
660 * restart window, so that we send ACKs quickly.
662 tcp_incr_quickack(sk);
666 icsk->icsk_ack.lrcvtime = now;
668 tcp_ecn_check_ce(tp, skb);
671 tcp_grow_window(sk, skb);
674 /* Called to compute a smoothed rtt estimate. The data fed to this
675 * routine either comes from timestamps, or from segments that were
676 * known _not_ to have been retransmitted [see Karn/Partridge
677 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
678 * piece by Van Jacobson.
679 * NOTE: the next three routines used to be one big routine.
680 * To save cycles in the RFC 1323 implementation it was better to break
681 * it up into three procedures. -- erics
683 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
685 struct tcp_sock *tp = tcp_sk(sk);
686 long m = mrtt_us; /* RTT */
687 u32 srtt = tp->srtt_us;
689 /* The following amusing code comes from Jacobson's
690 * article in SIGCOMM '88. Note that rtt and mdev
691 * are scaled versions of rtt and mean deviation.
692 * This is designed to be as fast as possible
693 * m stands for "measurement".
695 * On a 1990 paper the rto value is changed to:
696 * RTO = rtt + 4 * mdev
698 * Funny. This algorithm seems to be very broken.
699 * These formulae increase RTO, when it should be decreased, increase
700 * too slowly, when it should be increased quickly, decrease too quickly
701 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
702 * does not matter how to _calculate_ it. Seems, it was trap
703 * that VJ failed to avoid. 8)
706 m -= (srtt >> 3); /* m is now error in rtt est */
707 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
709 m = -m; /* m is now abs(error) */
710 m -= (tp->mdev_us >> 2); /* similar update on mdev */
711 /* This is similar to one of Eifel findings.
712 * Eifel blocks mdev updates when rtt decreases.
713 * This solution is a bit different: we use finer gain
714 * for mdev in this case (alpha*beta).
715 * Like Eifel it also prevents growth of rto,
716 * but also it limits too fast rto decreases,
717 * happening in pure Eifel.
722 m -= (tp->mdev_us >> 2); /* similar update on mdev */
724 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
725 if (tp->mdev_us > tp->mdev_max_us) {
726 tp->mdev_max_us = tp->mdev_us;
727 if (tp->mdev_max_us > tp->rttvar_us)
728 tp->rttvar_us = tp->mdev_max_us;
730 if (after(tp->snd_una, tp->rtt_seq)) {
731 if (tp->mdev_max_us < tp->rttvar_us)
732 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
733 tp->rtt_seq = tp->snd_nxt;
734 tp->mdev_max_us = tcp_rto_min_us(sk);
737 /* no previous measure. */
738 srtt = m << 3; /* take the measured time to be rtt */
739 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
740 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
741 tp->mdev_max_us = tp->rttvar_us;
742 tp->rtt_seq = tp->snd_nxt;
744 tp->srtt_us = max(1U, srtt);
747 /* Set the sk_pacing_rate to allow proper sizing of TSO packets.
748 * Note: TCP stack does not yet implement pacing.
749 * FQ packet scheduler can be used to implement cheap but effective
750 * TCP pacing, to smooth the burst on large writes when packets
751 * in flight is significantly lower than cwnd (or rwin)
753 static void tcp_update_pacing_rate(struct sock *sk)
755 const struct tcp_sock *tp = tcp_sk(sk);
758 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
759 rate = (u64)tp->mss_cache * 2 * (USEC_PER_SEC << 3);
761 rate *= max(tp->snd_cwnd, tp->packets_out);
763 if (likely(tp->srtt_us))
764 do_div(rate, tp->srtt_us);
766 /* ACCESS_ONCE() is needed because sch_fq fetches sk_pacing_rate
767 * without any lock. We want to make sure compiler wont store
768 * intermediate values in this location.
770 ACCESS_ONCE(sk->sk_pacing_rate) = min_t(u64, rate,
771 sk->sk_max_pacing_rate);
774 /* Calculate rto without backoff. This is the second half of Van Jacobson's
775 * routine referred to above.
777 static void tcp_set_rto(struct sock *sk)
779 const struct tcp_sock *tp = tcp_sk(sk);
780 /* Old crap is replaced with new one. 8)
783 * 1. If rtt variance happened to be less 50msec, it is hallucination.
784 * It cannot be less due to utterly erratic ACK generation made
785 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
786 * to do with delayed acks, because at cwnd>2 true delack timeout
787 * is invisible. Actually, Linux-2.4 also generates erratic
788 * ACKs in some circumstances.
790 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
792 /* 2. Fixups made earlier cannot be right.
793 * If we do not estimate RTO correctly without them,
794 * all the algo is pure shit and should be replaced
795 * with correct one. It is exactly, which we pretend to do.
798 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
799 * guarantees that rto is higher.
804 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
806 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
809 cwnd = TCP_INIT_CWND;
810 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
814 * Packet counting of FACK is based on in-order assumptions, therefore TCP
815 * disables it when reordering is detected
817 void tcp_disable_fack(struct tcp_sock *tp)
819 /* RFC3517 uses different metric in lost marker => reset on change */
821 tp->lost_skb_hint = NULL;
822 tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED;
825 /* Take a notice that peer is sending D-SACKs */
826 static void tcp_dsack_seen(struct tcp_sock *tp)
828 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
831 static void tcp_update_reordering(struct sock *sk, const int metric,
834 struct tcp_sock *tp = tcp_sk(sk);
835 if (metric > tp->reordering) {
838 tp->reordering = min(sysctl_tcp_max_reordering, metric);
840 /* This exciting event is worth to be remembered. 8) */
842 mib_idx = LINUX_MIB_TCPTSREORDER;
843 else if (tcp_is_reno(tp))
844 mib_idx = LINUX_MIB_TCPRENOREORDER;
845 else if (tcp_is_fack(tp))
846 mib_idx = LINUX_MIB_TCPFACKREORDER;
848 mib_idx = LINUX_MIB_TCPSACKREORDER;
850 NET_INC_STATS_BH(sock_net(sk), mib_idx);
851 #if FASTRETRANS_DEBUG > 1
852 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
853 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
857 tp->undo_marker ? tp->undo_retrans : 0);
859 tcp_disable_fack(tp);
863 tcp_disable_early_retrans(tp);
866 /* This must be called before lost_out is incremented */
867 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
869 if (!tp->retransmit_skb_hint ||
870 before(TCP_SKB_CB(skb)->seq,
871 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
872 tp->retransmit_skb_hint = skb;
875 after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
876 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
879 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
881 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
882 tcp_verify_retransmit_hint(tp, skb);
884 tp->lost_out += tcp_skb_pcount(skb);
885 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
889 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp,
892 tcp_verify_retransmit_hint(tp, skb);
894 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
895 tp->lost_out += tcp_skb_pcount(skb);
896 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
900 /* This procedure tags the retransmission queue when SACKs arrive.
902 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
903 * Packets in queue with these bits set are counted in variables
904 * sacked_out, retrans_out and lost_out, correspondingly.
906 * Valid combinations are:
907 * Tag InFlight Description
908 * 0 1 - orig segment is in flight.
909 * S 0 - nothing flies, orig reached receiver.
910 * L 0 - nothing flies, orig lost by net.
911 * R 2 - both orig and retransmit are in flight.
912 * L|R 1 - orig is lost, retransmit is in flight.
913 * S|R 1 - orig reached receiver, retrans is still in flight.
914 * (L|S|R is logically valid, it could occur when L|R is sacked,
915 * but it is equivalent to plain S and code short-curcuits it to S.
916 * L|S is logically invalid, it would mean -1 packet in flight 8))
918 * These 6 states form finite state machine, controlled by the following events:
919 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
920 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
921 * 3. Loss detection event of two flavors:
922 * A. Scoreboard estimator decided the packet is lost.
923 * A'. Reno "three dupacks" marks head of queue lost.
924 * A''. Its FACK modification, head until snd.fack is lost.
925 * B. SACK arrives sacking SND.NXT at the moment, when the
926 * segment was retransmitted.
927 * 4. D-SACK added new rule: D-SACK changes any tag to S.
929 * It is pleasant to note, that state diagram turns out to be commutative,
930 * so that we are allowed not to be bothered by order of our actions,
931 * when multiple events arrive simultaneously. (see the function below).
933 * Reordering detection.
934 * --------------------
935 * Reordering metric is maximal distance, which a packet can be displaced
936 * in packet stream. With SACKs we can estimate it:
938 * 1. SACK fills old hole and the corresponding segment was not
939 * ever retransmitted -> reordering. Alas, we cannot use it
940 * when segment was retransmitted.
941 * 2. The last flaw is solved with D-SACK. D-SACK arrives
942 * for retransmitted and already SACKed segment -> reordering..
943 * Both of these heuristics are not used in Loss state, when we cannot
944 * account for retransmits accurately.
946 * SACK block validation.
947 * ----------------------
949 * SACK block range validation checks that the received SACK block fits to
950 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
951 * Note that SND.UNA is not included to the range though being valid because
952 * it means that the receiver is rather inconsistent with itself reporting
953 * SACK reneging when it should advance SND.UNA. Such SACK block this is
954 * perfectly valid, however, in light of RFC2018 which explicitly states
955 * that "SACK block MUST reflect the newest segment. Even if the newest
956 * segment is going to be discarded ...", not that it looks very clever
957 * in case of head skb. Due to potentional receiver driven attacks, we
958 * choose to avoid immediate execution of a walk in write queue due to
959 * reneging and defer head skb's loss recovery to standard loss recovery
960 * procedure that will eventually trigger (nothing forbids us doing this).
962 * Implements also blockage to start_seq wrap-around. Problem lies in the
963 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
964 * there's no guarantee that it will be before snd_nxt (n). The problem
965 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
968 * <- outs wnd -> <- wrapzone ->
969 * u e n u_w e_w s n_w
971 * |<------------+------+----- TCP seqno space --------------+---------->|
972 * ...-- <2^31 ->| |<--------...
973 * ...---- >2^31 ------>| |<--------...
975 * Current code wouldn't be vulnerable but it's better still to discard such
976 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
977 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
978 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
979 * equal to the ideal case (infinite seqno space without wrap caused issues).
981 * With D-SACK the lower bound is extended to cover sequence space below
982 * SND.UNA down to undo_marker, which is the last point of interest. Yet
983 * again, D-SACK block must not to go across snd_una (for the same reason as
984 * for the normal SACK blocks, explained above). But there all simplicity
985 * ends, TCP might receive valid D-SACKs below that. As long as they reside
986 * fully below undo_marker they do not affect behavior in anyway and can
987 * therefore be safely ignored. In rare cases (which are more or less
988 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
989 * fragmentation and packet reordering past skb's retransmission. To consider
990 * them correctly, the acceptable range must be extended even more though
991 * the exact amount is rather hard to quantify. However, tp->max_window can
992 * be used as an exaggerated estimate.
994 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
995 u32 start_seq, u32 end_seq)
997 /* Too far in future, or reversed (interpretation is ambiguous) */
998 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1001 /* Nasty start_seq wrap-around check (see comments above) */
1002 if (!before(start_seq, tp->snd_nxt))
1005 /* In outstanding window? ...This is valid exit for D-SACKs too.
1006 * start_seq == snd_una is non-sensical (see comments above)
1008 if (after(start_seq, tp->snd_una))
1011 if (!is_dsack || !tp->undo_marker)
1014 /* ...Then it's D-SACK, and must reside below snd_una completely */
1015 if (after(end_seq, tp->snd_una))
1018 if (!before(start_seq, tp->undo_marker))
1022 if (!after(end_seq, tp->undo_marker))
1025 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1026 * start_seq < undo_marker and end_seq >= undo_marker.
1028 return !before(start_seq, end_seq - tp->max_window);
1031 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1032 * Event "B". Later note: FACK people cheated me again 8), we have to account
1033 * for reordering! Ugly, but should help.
1035 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1036 * less than what is now known to be received by the other end (derived from
1037 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1038 * retransmitted skbs to avoid some costly processing per ACKs.
1040 static void tcp_mark_lost_retrans(struct sock *sk)
1042 const struct inet_connection_sock *icsk = inet_csk(sk);
1043 struct tcp_sock *tp = tcp_sk(sk);
1044 struct sk_buff *skb;
1046 u32 new_low_seq = tp->snd_nxt;
1047 u32 received_upto = tcp_highest_sack_seq(tp);
1049 if (!tcp_is_fack(tp) || !tp->retrans_out ||
1050 !after(received_upto, tp->lost_retrans_low) ||
1051 icsk->icsk_ca_state != TCP_CA_Recovery)
1054 tcp_for_write_queue(skb, sk) {
1055 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1057 if (skb == tcp_send_head(sk))
1059 if (cnt == tp->retrans_out)
1061 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1064 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1067 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1068 * constraint here (see above) but figuring out that at
1069 * least tp->reordering SACK blocks reside between ack_seq
1070 * and received_upto is not easy task to do cheaply with
1071 * the available datastructures.
1073 * Whether FACK should check here for tp->reordering segs
1074 * in-between one could argue for either way (it would be
1075 * rather simple to implement as we could count fack_count
1076 * during the walk and do tp->fackets_out - fack_count).
1078 if (after(received_upto, ack_seq)) {
1079 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1080 tp->retrans_out -= tcp_skb_pcount(skb);
1082 tcp_skb_mark_lost_uncond_verify(tp, skb);
1083 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
1085 if (before(ack_seq, new_low_seq))
1086 new_low_seq = ack_seq;
1087 cnt += tcp_skb_pcount(skb);
1091 if (tp->retrans_out)
1092 tp->lost_retrans_low = new_low_seq;
1095 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1096 struct tcp_sack_block_wire *sp, int num_sacks,
1099 struct tcp_sock *tp = tcp_sk(sk);
1100 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1101 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1102 bool dup_sack = false;
1104 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1107 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1108 } else if (num_sacks > 1) {
1109 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1110 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1112 if (!after(end_seq_0, end_seq_1) &&
1113 !before(start_seq_0, start_seq_1)) {
1116 NET_INC_STATS_BH(sock_net(sk),
1117 LINUX_MIB_TCPDSACKOFORECV);
1121 /* D-SACK for already forgotten data... Do dumb counting. */
1122 if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 &&
1123 !after(end_seq_0, prior_snd_una) &&
1124 after(end_seq_0, tp->undo_marker))
1130 struct tcp_sacktag_state {
1133 /* Timestamps for earliest and latest never-retransmitted segment
1134 * that was SACKed. RTO needs the earliest RTT to stay conservative,
1135 * but congestion control should still get an accurate delay signal.
1137 struct skb_mstamp first_sackt;
1138 struct skb_mstamp last_sackt;
1142 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1143 * the incoming SACK may not exactly match but we can find smaller MSS
1144 * aligned portion of it that matches. Therefore we might need to fragment
1145 * which may fail and creates some hassle (caller must handle error case
1148 * FIXME: this could be merged to shift decision code
1150 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1151 u32 start_seq, u32 end_seq)
1155 unsigned int pkt_len;
1158 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1159 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1161 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1162 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1163 mss = tcp_skb_mss(skb);
1164 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1167 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1171 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1176 /* Round if necessary so that SACKs cover only full MSSes
1177 * and/or the remaining small portion (if present)
1179 if (pkt_len > mss) {
1180 unsigned int new_len = (pkt_len / mss) * mss;
1181 if (!in_sack && new_len < pkt_len) {
1183 if (new_len >= skb->len)
1188 err = tcp_fragment(sk, skb, pkt_len, mss, GFP_ATOMIC);
1196 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1197 static u8 tcp_sacktag_one(struct sock *sk,
1198 struct tcp_sacktag_state *state, u8 sacked,
1199 u32 start_seq, u32 end_seq,
1200 int dup_sack, int pcount,
1201 const struct skb_mstamp *xmit_time)
1203 struct tcp_sock *tp = tcp_sk(sk);
1204 int fack_count = state->fack_count;
1206 /* Account D-SACK for retransmitted packet. */
1207 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1208 if (tp->undo_marker && tp->undo_retrans > 0 &&
1209 after(end_seq, tp->undo_marker))
1211 if (sacked & TCPCB_SACKED_ACKED)
1212 state->reord = min(fack_count, state->reord);
1215 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1216 if (!after(end_seq, tp->snd_una))
1219 if (!(sacked & TCPCB_SACKED_ACKED)) {
1220 if (sacked & TCPCB_SACKED_RETRANS) {
1221 /* If the segment is not tagged as lost,
1222 * we do not clear RETRANS, believing
1223 * that retransmission is still in flight.
1225 if (sacked & TCPCB_LOST) {
1226 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1227 tp->lost_out -= pcount;
1228 tp->retrans_out -= pcount;
1231 if (!(sacked & TCPCB_RETRANS)) {
1232 /* New sack for not retransmitted frame,
1233 * which was in hole. It is reordering.
1235 if (before(start_seq,
1236 tcp_highest_sack_seq(tp)))
1237 state->reord = min(fack_count,
1239 if (!after(end_seq, tp->high_seq))
1240 state->flag |= FLAG_ORIG_SACK_ACKED;
1241 if (state->first_sackt.v64 == 0)
1242 state->first_sackt = *xmit_time;
1243 state->last_sackt = *xmit_time;
1246 if (sacked & TCPCB_LOST) {
1247 sacked &= ~TCPCB_LOST;
1248 tp->lost_out -= pcount;
1252 sacked |= TCPCB_SACKED_ACKED;
1253 state->flag |= FLAG_DATA_SACKED;
1254 tp->sacked_out += pcount;
1256 fack_count += pcount;
1258 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1259 if (!tcp_is_fack(tp) && tp->lost_skb_hint &&
1260 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1261 tp->lost_cnt_hint += pcount;
1263 if (fack_count > tp->fackets_out)
1264 tp->fackets_out = fack_count;
1267 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1268 * frames and clear it. undo_retrans is decreased above, L|R frames
1269 * are accounted above as well.
1271 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1272 sacked &= ~TCPCB_SACKED_RETRANS;
1273 tp->retrans_out -= pcount;
1279 /* Shift newly-SACKed bytes from this skb to the immediately previous
1280 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1282 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1283 struct tcp_sacktag_state *state,
1284 unsigned int pcount, int shifted, int mss,
1287 struct tcp_sock *tp = tcp_sk(sk);
1288 struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1289 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1290 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1294 /* Adjust counters and hints for the newly sacked sequence
1295 * range but discard the return value since prev is already
1296 * marked. We must tag the range first because the seq
1297 * advancement below implicitly advances
1298 * tcp_highest_sack_seq() when skb is highest_sack.
1300 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1301 start_seq, end_seq, dup_sack, pcount,
1304 if (skb == tp->lost_skb_hint)
1305 tp->lost_cnt_hint += pcount;
1307 TCP_SKB_CB(prev)->end_seq += shifted;
1308 TCP_SKB_CB(skb)->seq += shifted;
1310 tcp_skb_pcount_add(prev, pcount);
1311 BUG_ON(tcp_skb_pcount(skb) < pcount);
1312 tcp_skb_pcount_add(skb, -pcount);
1314 /* When we're adding to gso_segs == 1, gso_size will be zero,
1315 * in theory this shouldn't be necessary but as long as DSACK
1316 * code can come after this skb later on it's better to keep
1317 * setting gso_size to something.
1319 if (!TCP_SKB_CB(prev)->tcp_gso_size)
1320 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1322 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1323 if (tcp_skb_pcount(skb) <= 1)
1324 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1326 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1327 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1330 BUG_ON(!tcp_skb_pcount(skb));
1331 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1335 /* Whole SKB was eaten :-) */
1337 if (skb == tp->retransmit_skb_hint)
1338 tp->retransmit_skb_hint = prev;
1339 if (skb == tp->lost_skb_hint) {
1340 tp->lost_skb_hint = prev;
1341 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1344 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1345 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1346 TCP_SKB_CB(prev)->end_seq++;
1348 if (skb == tcp_highest_sack(sk))
1349 tcp_advance_highest_sack(sk, skb);
1351 tcp_unlink_write_queue(skb, sk);
1352 sk_wmem_free_skb(sk, skb);
1354 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1359 /* I wish gso_size would have a bit more sane initialization than
1360 * something-or-zero which complicates things
1362 static int tcp_skb_seglen(const struct sk_buff *skb)
1364 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1367 /* Shifting pages past head area doesn't work */
1368 static int skb_can_shift(const struct sk_buff *skb)
1370 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1373 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1376 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1377 struct tcp_sacktag_state *state,
1378 u32 start_seq, u32 end_seq,
1381 struct tcp_sock *tp = tcp_sk(sk);
1382 struct sk_buff *prev;
1388 if (!sk_can_gso(sk))
1391 /* Normally R but no L won't result in plain S */
1393 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1395 if (!skb_can_shift(skb))
1397 /* This frame is about to be dropped (was ACKed). */
1398 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1401 /* Can only happen with delayed DSACK + discard craziness */
1402 if (unlikely(skb == tcp_write_queue_head(sk)))
1404 prev = tcp_write_queue_prev(sk, skb);
1406 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1409 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1410 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1414 pcount = tcp_skb_pcount(skb);
1415 mss = tcp_skb_seglen(skb);
1417 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1418 * drop this restriction as unnecessary
1420 if (mss != tcp_skb_seglen(prev))
1423 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1425 /* CHECKME: This is non-MSS split case only?, this will
1426 * cause skipped skbs due to advancing loop btw, original
1427 * has that feature too
1429 if (tcp_skb_pcount(skb) <= 1)
1432 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1434 /* TODO: head merge to next could be attempted here
1435 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1436 * though it might not be worth of the additional hassle
1438 * ...we can probably just fallback to what was done
1439 * previously. We could try merging non-SACKed ones
1440 * as well but it probably isn't going to buy off
1441 * because later SACKs might again split them, and
1442 * it would make skb timestamp tracking considerably
1448 len = end_seq - TCP_SKB_CB(skb)->seq;
1450 BUG_ON(len > skb->len);
1452 /* MSS boundaries should be honoured or else pcount will
1453 * severely break even though it makes things bit trickier.
1454 * Optimize common case to avoid most of the divides
1456 mss = tcp_skb_mss(skb);
1458 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1459 * drop this restriction as unnecessary
1461 if (mss != tcp_skb_seglen(prev))
1466 } else if (len < mss) {
1474 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1475 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1478 if (!skb_shift(prev, skb, len))
1480 if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1483 /* Hole filled allows collapsing with the next as well, this is very
1484 * useful when hole on every nth skb pattern happens
1486 if (prev == tcp_write_queue_tail(sk))
1488 skb = tcp_write_queue_next(sk, prev);
1490 if (!skb_can_shift(skb) ||
1491 (skb == tcp_send_head(sk)) ||
1492 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1493 (mss != tcp_skb_seglen(skb)))
1497 if (skb_shift(prev, skb, len)) {
1498 pcount += tcp_skb_pcount(skb);
1499 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1503 state->fack_count += pcount;
1510 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1514 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1515 struct tcp_sack_block *next_dup,
1516 struct tcp_sacktag_state *state,
1517 u32 start_seq, u32 end_seq,
1520 struct tcp_sock *tp = tcp_sk(sk);
1521 struct sk_buff *tmp;
1523 tcp_for_write_queue_from(skb, sk) {
1525 bool dup_sack = dup_sack_in;
1527 if (skb == tcp_send_head(sk))
1530 /* queue is in-order => we can short-circuit the walk early */
1531 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1535 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1536 in_sack = tcp_match_skb_to_sack(sk, skb,
1537 next_dup->start_seq,
1543 /* skb reference here is a bit tricky to get right, since
1544 * shifting can eat and free both this skb and the next,
1545 * so not even _safe variant of the loop is enough.
1548 tmp = tcp_shift_skb_data(sk, skb, state,
1549 start_seq, end_seq, dup_sack);
1558 in_sack = tcp_match_skb_to_sack(sk, skb,
1564 if (unlikely(in_sack < 0))
1568 TCP_SKB_CB(skb)->sacked =
1571 TCP_SKB_CB(skb)->sacked,
1572 TCP_SKB_CB(skb)->seq,
1573 TCP_SKB_CB(skb)->end_seq,
1575 tcp_skb_pcount(skb),
1578 if (!before(TCP_SKB_CB(skb)->seq,
1579 tcp_highest_sack_seq(tp)))
1580 tcp_advance_highest_sack(sk, skb);
1583 state->fack_count += tcp_skb_pcount(skb);
1588 /* Avoid all extra work that is being done by sacktag while walking in
1591 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1592 struct tcp_sacktag_state *state,
1595 tcp_for_write_queue_from(skb, sk) {
1596 if (skb == tcp_send_head(sk))
1599 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1602 state->fack_count += tcp_skb_pcount(skb);
1607 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1609 struct tcp_sack_block *next_dup,
1610 struct tcp_sacktag_state *state,
1616 if (before(next_dup->start_seq, skip_to_seq)) {
1617 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1618 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1619 next_dup->start_seq, next_dup->end_seq,
1626 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1628 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1632 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1633 u32 prior_snd_una, struct tcp_sacktag_state *state)
1635 struct tcp_sock *tp = tcp_sk(sk);
1636 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1637 TCP_SKB_CB(ack_skb)->sacked);
1638 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1639 struct tcp_sack_block sp[TCP_NUM_SACKS];
1640 struct tcp_sack_block *cache;
1641 struct sk_buff *skb;
1642 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1644 bool found_dup_sack = false;
1646 int first_sack_index;
1649 state->reord = tp->packets_out;
1651 if (!tp->sacked_out) {
1652 if (WARN_ON(tp->fackets_out))
1653 tp->fackets_out = 0;
1654 tcp_highest_sack_reset(sk);
1657 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1658 num_sacks, prior_snd_una);
1660 state->flag |= FLAG_DSACKING_ACK;
1662 /* Eliminate too old ACKs, but take into
1663 * account more or less fresh ones, they can
1664 * contain valid SACK info.
1666 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1669 if (!tp->packets_out)
1673 first_sack_index = 0;
1674 for (i = 0; i < num_sacks; i++) {
1675 bool dup_sack = !i && found_dup_sack;
1677 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1678 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1680 if (!tcp_is_sackblock_valid(tp, dup_sack,
1681 sp[used_sacks].start_seq,
1682 sp[used_sacks].end_seq)) {
1686 if (!tp->undo_marker)
1687 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1689 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1691 /* Don't count olds caused by ACK reordering */
1692 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1693 !after(sp[used_sacks].end_seq, tp->snd_una))
1695 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1698 NET_INC_STATS_BH(sock_net(sk), mib_idx);
1700 first_sack_index = -1;
1704 /* Ignore very old stuff early */
1705 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1711 /* order SACK blocks to allow in order walk of the retrans queue */
1712 for (i = used_sacks - 1; i > 0; i--) {
1713 for (j = 0; j < i; j++) {
1714 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1715 swap(sp[j], sp[j + 1]);
1717 /* Track where the first SACK block goes to */
1718 if (j == first_sack_index)
1719 first_sack_index = j + 1;
1724 skb = tcp_write_queue_head(sk);
1725 state->fack_count = 0;
1728 if (!tp->sacked_out) {
1729 /* It's already past, so skip checking against it */
1730 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1732 cache = tp->recv_sack_cache;
1733 /* Skip empty blocks in at head of the cache */
1734 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1739 while (i < used_sacks) {
1740 u32 start_seq = sp[i].start_seq;
1741 u32 end_seq = sp[i].end_seq;
1742 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1743 struct tcp_sack_block *next_dup = NULL;
1745 if (found_dup_sack && ((i + 1) == first_sack_index))
1746 next_dup = &sp[i + 1];
1748 /* Skip too early cached blocks */
1749 while (tcp_sack_cache_ok(tp, cache) &&
1750 !before(start_seq, cache->end_seq))
1753 /* Can skip some work by looking recv_sack_cache? */
1754 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1755 after(end_seq, cache->start_seq)) {
1758 if (before(start_seq, cache->start_seq)) {
1759 skb = tcp_sacktag_skip(skb, sk, state,
1761 skb = tcp_sacktag_walk(skb, sk, next_dup,
1768 /* Rest of the block already fully processed? */
1769 if (!after(end_seq, cache->end_seq))
1772 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1776 /* ...tail remains todo... */
1777 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1778 /* ...but better entrypoint exists! */
1779 skb = tcp_highest_sack(sk);
1782 state->fack_count = tp->fackets_out;
1787 skb = tcp_sacktag_skip(skb, sk, state, cache->end_seq);
1788 /* Check overlap against next cached too (past this one already) */
1793 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1794 skb = tcp_highest_sack(sk);
1797 state->fack_count = tp->fackets_out;
1799 skb = tcp_sacktag_skip(skb, sk, state, start_seq);
1802 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
1803 start_seq, end_seq, dup_sack);
1809 /* Clear the head of the cache sack blocks so we can skip it next time */
1810 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1811 tp->recv_sack_cache[i].start_seq = 0;
1812 tp->recv_sack_cache[i].end_seq = 0;
1814 for (j = 0; j < used_sacks; j++)
1815 tp->recv_sack_cache[i++] = sp[j];
1817 if ((state->reord < tp->fackets_out) &&
1818 ((inet_csk(sk)->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker))
1819 tcp_update_reordering(sk, tp->fackets_out - state->reord, 0);
1821 tcp_mark_lost_retrans(sk);
1822 tcp_verify_left_out(tp);
1825 #if FASTRETRANS_DEBUG > 0
1826 WARN_ON((int)tp->sacked_out < 0);
1827 WARN_ON((int)tp->lost_out < 0);
1828 WARN_ON((int)tp->retrans_out < 0);
1829 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1834 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1835 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1837 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1841 holes = max(tp->lost_out, 1U);
1842 holes = min(holes, tp->packets_out);
1844 if ((tp->sacked_out + holes) > tp->packets_out) {
1845 tp->sacked_out = tp->packets_out - holes;
1851 /* If we receive more dupacks than we expected counting segments
1852 * in assumption of absent reordering, interpret this as reordering.
1853 * The only another reason could be bug in receiver TCP.
1855 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1857 struct tcp_sock *tp = tcp_sk(sk);
1858 if (tcp_limit_reno_sacked(tp))
1859 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1862 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1864 static void tcp_add_reno_sack(struct sock *sk)
1866 struct tcp_sock *tp = tcp_sk(sk);
1868 tcp_check_reno_reordering(sk, 0);
1869 tcp_verify_left_out(tp);
1872 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1874 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1876 struct tcp_sock *tp = tcp_sk(sk);
1879 /* One ACK acked hole. The rest eat duplicate ACKs. */
1880 if (acked - 1 >= tp->sacked_out)
1883 tp->sacked_out -= acked - 1;
1885 tcp_check_reno_reordering(sk, acked);
1886 tcp_verify_left_out(tp);
1889 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1894 void tcp_clear_retrans(struct tcp_sock *tp)
1896 tp->retrans_out = 0;
1898 tp->undo_marker = 0;
1899 tp->undo_retrans = -1;
1900 tp->fackets_out = 0;
1904 static inline void tcp_init_undo(struct tcp_sock *tp)
1906 tp->undo_marker = tp->snd_una;
1907 /* Retransmission still in flight may cause DSACKs later. */
1908 tp->undo_retrans = tp->retrans_out ? : -1;
1911 /* Enter Loss state. If we detect SACK reneging, forget all SACK information
1912 * and reset tags completely, otherwise preserve SACKs. If receiver
1913 * dropped its ofo queue, we will know this due to reneging detection.
1915 void tcp_enter_loss(struct sock *sk)
1917 const struct inet_connection_sock *icsk = inet_csk(sk);
1918 struct tcp_sock *tp = tcp_sk(sk);
1919 struct sk_buff *skb;
1920 bool new_recovery = false;
1921 bool is_reneg; /* is receiver reneging on SACKs? */
1923 /* Reduce ssthresh if it has not yet been made inside this window. */
1924 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1925 !after(tp->high_seq, tp->snd_una) ||
1926 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1927 new_recovery = true;
1928 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1929 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1930 tcp_ca_event(sk, CA_EVENT_LOSS);
1934 tp->snd_cwnd_cnt = 0;
1935 tp->snd_cwnd_stamp = tcp_time_stamp;
1937 tp->retrans_out = 0;
1940 if (tcp_is_reno(tp))
1941 tcp_reset_reno_sack(tp);
1943 skb = tcp_write_queue_head(sk);
1944 is_reneg = skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED);
1946 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
1948 tp->fackets_out = 0;
1950 tcp_clear_all_retrans_hints(tp);
1952 tcp_for_write_queue(skb, sk) {
1953 if (skb == tcp_send_head(sk))
1956 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1957 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || is_reneg) {
1958 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1959 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1960 tp->lost_out += tcp_skb_pcount(skb);
1961 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
1964 tcp_verify_left_out(tp);
1966 /* Timeout in disordered state after receiving substantial DUPACKs
1967 * suggests that the degree of reordering is over-estimated.
1969 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
1970 tp->sacked_out >= sysctl_tcp_reordering)
1971 tp->reordering = min_t(unsigned int, tp->reordering,
1972 sysctl_tcp_reordering);
1973 tcp_set_ca_state(sk, TCP_CA_Loss);
1974 tp->high_seq = tp->snd_nxt;
1975 tcp_ecn_queue_cwr(tp);
1977 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1978 * loss recovery is underway except recurring timeout(s) on
1979 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1981 tp->frto = sysctl_tcp_frto &&
1982 (new_recovery || icsk->icsk_retransmits) &&
1983 !inet_csk(sk)->icsk_mtup.probe_size;
1986 /* If ACK arrived pointing to a remembered SACK, it means that our
1987 * remembered SACKs do not reflect real state of receiver i.e.
1988 * receiver _host_ is heavily congested (or buggy).
1990 * To avoid big spurious retransmission bursts due to transient SACK
1991 * scoreboard oddities that look like reneging, we give the receiver a
1992 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
1993 * restore sanity to the SACK scoreboard. If the apparent reneging
1994 * persists until this RTO then we'll clear the SACK scoreboard.
1996 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
1998 if (flag & FLAG_SACK_RENEGING) {
1999 struct tcp_sock *tp = tcp_sk(sk);
2000 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2001 msecs_to_jiffies(10));
2003 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2004 delay, TCP_RTO_MAX);
2010 static inline int tcp_fackets_out(const struct tcp_sock *tp)
2012 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
2015 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2016 * counter when SACK is enabled (without SACK, sacked_out is used for
2019 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2020 * segments up to the highest received SACK block so far and holes in
2023 * With reordering, holes may still be in flight, so RFC3517 recovery
2024 * uses pure sacked_out (total number of SACKed segments) even though
2025 * it violates the RFC that uses duplicate ACKs, often these are equal
2026 * but when e.g. out-of-window ACKs or packet duplication occurs,
2027 * they differ. Since neither occurs due to loss, TCP should really
2030 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2032 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2035 static bool tcp_pause_early_retransmit(struct sock *sk, int flag)
2037 struct tcp_sock *tp = tcp_sk(sk);
2038 unsigned long delay;
2040 /* Delay early retransmit and entering fast recovery for
2041 * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
2042 * available, or RTO is scheduled to fire first.
2044 if (sysctl_tcp_early_retrans < 2 || sysctl_tcp_early_retrans > 3 ||
2045 (flag & FLAG_ECE) || !tp->srtt_us)
2048 delay = max(usecs_to_jiffies(tp->srtt_us >> 5),
2049 msecs_to_jiffies(2));
2051 if (!time_after(inet_csk(sk)->icsk_timeout, (jiffies + delay)))
2054 inet_csk_reset_xmit_timer(sk, ICSK_TIME_EARLY_RETRANS, delay,
2059 /* Linux NewReno/SACK/FACK/ECN state machine.
2060 * --------------------------------------
2062 * "Open" Normal state, no dubious events, fast path.
2063 * "Disorder" In all the respects it is "Open",
2064 * but requires a bit more attention. It is entered when
2065 * we see some SACKs or dupacks. It is split of "Open"
2066 * mainly to move some processing from fast path to slow one.
2067 * "CWR" CWND was reduced due to some Congestion Notification event.
2068 * It can be ECN, ICMP source quench, local device congestion.
2069 * "Recovery" CWND was reduced, we are fast-retransmitting.
2070 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2072 * tcp_fastretrans_alert() is entered:
2073 * - each incoming ACK, if state is not "Open"
2074 * - when arrived ACK is unusual, namely:
2079 * Counting packets in flight is pretty simple.
2081 * in_flight = packets_out - left_out + retrans_out
2083 * packets_out is SND.NXT-SND.UNA counted in packets.
2085 * retrans_out is number of retransmitted segments.
2087 * left_out is number of segments left network, but not ACKed yet.
2089 * left_out = sacked_out + lost_out
2091 * sacked_out: Packets, which arrived to receiver out of order
2092 * and hence not ACKed. With SACKs this number is simply
2093 * amount of SACKed data. Even without SACKs
2094 * it is easy to give pretty reliable estimate of this number,
2095 * counting duplicate ACKs.
2097 * lost_out: Packets lost by network. TCP has no explicit
2098 * "loss notification" feedback from network (for now).
2099 * It means that this number can be only _guessed_.
2100 * Actually, it is the heuristics to predict lossage that
2101 * distinguishes different algorithms.
2103 * F.e. after RTO, when all the queue is considered as lost,
2104 * lost_out = packets_out and in_flight = retrans_out.
2106 * Essentially, we have now two algorithms counting
2109 * FACK: It is the simplest heuristics. As soon as we decided
2110 * that something is lost, we decide that _all_ not SACKed
2111 * packets until the most forward SACK are lost. I.e.
2112 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2113 * It is absolutely correct estimate, if network does not reorder
2114 * packets. And it loses any connection to reality when reordering
2115 * takes place. We use FACK by default until reordering
2116 * is suspected on the path to this destination.
2118 * NewReno: when Recovery is entered, we assume that one segment
2119 * is lost (classic Reno). While we are in Recovery and
2120 * a partial ACK arrives, we assume that one more packet
2121 * is lost (NewReno). This heuristics are the same in NewReno
2124 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2125 * deflation etc. CWND is real congestion window, never inflated, changes
2126 * only according to classic VJ rules.
2128 * Really tricky (and requiring careful tuning) part of algorithm
2129 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2130 * The first determines the moment _when_ we should reduce CWND and,
2131 * hence, slow down forward transmission. In fact, it determines the moment
2132 * when we decide that hole is caused by loss, rather than by a reorder.
2134 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2135 * holes, caused by lost packets.
2137 * And the most logically complicated part of algorithm is undo
2138 * heuristics. We detect false retransmits due to both too early
2139 * fast retransmit (reordering) and underestimated RTO, analyzing
2140 * timestamps and D-SACKs. When we detect that some segments were
2141 * retransmitted by mistake and CWND reduction was wrong, we undo
2142 * window reduction and abort recovery phase. This logic is hidden
2143 * inside several functions named tcp_try_undo_<something>.
2146 /* This function decides, when we should leave Disordered state
2147 * and enter Recovery phase, reducing congestion window.
2149 * Main question: may we further continue forward transmission
2150 * with the same cwnd?
2152 static bool tcp_time_to_recover(struct sock *sk, int flag)
2154 struct tcp_sock *tp = tcp_sk(sk);
2157 /* Trick#1: The loss is proven. */
2161 /* Not-A-Trick#2 : Classic rule... */
2162 if (tcp_dupack_heuristics(tp) > tp->reordering)
2165 /* Trick#4: It is still not OK... But will it be useful to delay
2168 packets_out = tp->packets_out;
2169 if (packets_out <= tp->reordering &&
2170 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2171 !tcp_may_send_now(sk)) {
2172 /* We have nothing to send. This connection is limited
2173 * either by receiver window or by application.
2178 /* If a thin stream is detected, retransmit after first
2179 * received dupack. Employ only if SACK is supported in order
2180 * to avoid possible corner-case series of spurious retransmissions
2181 * Use only if there are no unsent data.
2183 if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2184 tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2185 tcp_is_sack(tp) && !tcp_send_head(sk))
2188 /* Trick#6: TCP early retransmit, per RFC5827. To avoid spurious
2189 * retransmissions due to small network reorderings, we implement
2190 * Mitigation A.3 in the RFC and delay the retransmission for a short
2191 * interval if appropriate.
2193 if (tp->do_early_retrans && !tp->retrans_out && tp->sacked_out &&
2194 (tp->packets_out >= (tp->sacked_out + 1) && tp->packets_out < 4) &&
2195 !tcp_may_send_now(sk))
2196 return !tcp_pause_early_retransmit(sk, flag);
2201 /* Detect loss in event "A" above by marking head of queue up as lost.
2202 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2203 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2204 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2205 * the maximum SACKed segments to pass before reaching this limit.
2207 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2209 struct tcp_sock *tp = tcp_sk(sk);
2210 struct sk_buff *skb;
2214 /* Use SACK to deduce losses of new sequences sent during recovery */
2215 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq;
2217 WARN_ON(packets > tp->packets_out);
2218 if (tp->lost_skb_hint) {
2219 skb = tp->lost_skb_hint;
2220 cnt = tp->lost_cnt_hint;
2221 /* Head already handled? */
2222 if (mark_head && skb != tcp_write_queue_head(sk))
2225 skb = tcp_write_queue_head(sk);
2229 tcp_for_write_queue_from(skb, sk) {
2230 if (skb == tcp_send_head(sk))
2232 /* TODO: do this better */
2233 /* this is not the most efficient way to do this... */
2234 tp->lost_skb_hint = skb;
2235 tp->lost_cnt_hint = cnt;
2237 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2241 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2242 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2243 cnt += tcp_skb_pcount(skb);
2245 if (cnt > packets) {
2246 if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2247 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2248 (oldcnt >= packets))
2251 mss = tcp_skb_mss(skb);
2252 err = tcp_fragment(sk, skb, (packets - oldcnt) * mss,
2259 tcp_skb_mark_lost(tp, skb);
2264 tcp_verify_left_out(tp);
2267 /* Account newly detected lost packet(s) */
2269 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2271 struct tcp_sock *tp = tcp_sk(sk);
2273 if (tcp_is_reno(tp)) {
2274 tcp_mark_head_lost(sk, 1, 1);
2275 } else if (tcp_is_fack(tp)) {
2276 int lost = tp->fackets_out - tp->reordering;
2279 tcp_mark_head_lost(sk, lost, 0);
2281 int sacked_upto = tp->sacked_out - tp->reordering;
2282 if (sacked_upto >= 0)
2283 tcp_mark_head_lost(sk, sacked_upto, 0);
2284 else if (fast_rexmit)
2285 tcp_mark_head_lost(sk, 1, 1);
2289 /* CWND moderation, preventing bursts due to too big ACKs
2290 * in dubious situations.
2292 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2294 tp->snd_cwnd = min(tp->snd_cwnd,
2295 tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2296 tp->snd_cwnd_stamp = tcp_time_stamp;
2299 /* Nothing was retransmitted or returned timestamp is less
2300 * than timestamp of the first retransmission.
2302 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2304 return !tp->retrans_stamp ||
2305 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2306 before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
2309 /* Undo procedures. */
2311 /* We can clear retrans_stamp when there are no retransmissions in the
2312 * window. It would seem that it is trivially available for us in
2313 * tp->retrans_out, however, that kind of assumptions doesn't consider
2314 * what will happen if errors occur when sending retransmission for the
2315 * second time. ...It could the that such segment has only
2316 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2317 * the head skb is enough except for some reneging corner cases that
2318 * are not worth the effort.
2320 * Main reason for all this complexity is the fact that connection dying
2321 * time now depends on the validity of the retrans_stamp, in particular,
2322 * that successive retransmissions of a segment must not advance
2323 * retrans_stamp under any conditions.
2325 static bool tcp_any_retrans_done(const struct sock *sk)
2327 const struct tcp_sock *tp = tcp_sk(sk);
2328 struct sk_buff *skb;
2330 if (tp->retrans_out)
2333 skb = tcp_write_queue_head(sk);
2334 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2340 #if FASTRETRANS_DEBUG > 1
2341 static void DBGUNDO(struct sock *sk, const char *msg)
2343 struct tcp_sock *tp = tcp_sk(sk);
2344 struct inet_sock *inet = inet_sk(sk);
2346 if (sk->sk_family == AF_INET) {
2347 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2349 &inet->inet_daddr, ntohs(inet->inet_dport),
2350 tp->snd_cwnd, tcp_left_out(tp),
2351 tp->snd_ssthresh, tp->prior_ssthresh,
2354 #if IS_ENABLED(CONFIG_IPV6)
2355 else if (sk->sk_family == AF_INET6) {
2356 struct ipv6_pinfo *np = inet6_sk(sk);
2357 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2359 &np->daddr, ntohs(inet->inet_dport),
2360 tp->snd_cwnd, tcp_left_out(tp),
2361 tp->snd_ssthresh, tp->prior_ssthresh,
2367 #define DBGUNDO(x...) do { } while (0)
2370 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2372 struct tcp_sock *tp = tcp_sk(sk);
2375 struct sk_buff *skb;
2377 tcp_for_write_queue(skb, sk) {
2378 if (skb == tcp_send_head(sk))
2380 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2383 tcp_clear_all_retrans_hints(tp);
2386 if (tp->prior_ssthresh) {
2387 const struct inet_connection_sock *icsk = inet_csk(sk);
2389 if (icsk->icsk_ca_ops->undo_cwnd)
2390 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2392 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2394 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2395 tp->snd_ssthresh = tp->prior_ssthresh;
2396 tcp_ecn_withdraw_cwr(tp);
2399 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2401 tp->snd_cwnd_stamp = tcp_time_stamp;
2402 tp->undo_marker = 0;
2405 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2407 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2410 /* People celebrate: "We love our President!" */
2411 static bool tcp_try_undo_recovery(struct sock *sk)
2413 struct tcp_sock *tp = tcp_sk(sk);
2415 if (tcp_may_undo(tp)) {
2418 /* Happy end! We did not retransmit anything
2419 * or our original transmission succeeded.
2421 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2422 tcp_undo_cwnd_reduction(sk, false);
2423 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2424 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2426 mib_idx = LINUX_MIB_TCPFULLUNDO;
2428 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2430 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2431 /* Hold old state until something *above* high_seq
2432 * is ACKed. For Reno it is MUST to prevent false
2433 * fast retransmits (RFC2582). SACK TCP is safe. */
2434 tcp_moderate_cwnd(tp);
2435 if (!tcp_any_retrans_done(sk))
2436 tp->retrans_stamp = 0;
2439 tcp_set_ca_state(sk, TCP_CA_Open);
2443 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2444 static bool tcp_try_undo_dsack(struct sock *sk)
2446 struct tcp_sock *tp = tcp_sk(sk);
2448 if (tp->undo_marker && !tp->undo_retrans) {
2449 DBGUNDO(sk, "D-SACK");
2450 tcp_undo_cwnd_reduction(sk, false);
2451 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2457 /* Undo during loss recovery after partial ACK or using F-RTO. */
2458 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2460 struct tcp_sock *tp = tcp_sk(sk);
2462 if (frto_undo || tcp_may_undo(tp)) {
2463 tcp_undo_cwnd_reduction(sk, true);
2465 DBGUNDO(sk, "partial loss");
2466 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2468 NET_INC_STATS_BH(sock_net(sk),
2469 LINUX_MIB_TCPSPURIOUSRTOS);
2470 inet_csk(sk)->icsk_retransmits = 0;
2471 if (frto_undo || tcp_is_sack(tp))
2472 tcp_set_ca_state(sk, TCP_CA_Open);
2478 /* The cwnd reduction in CWR and Recovery use the PRR algorithm
2479 * https://datatracker.ietf.org/doc/draft-ietf-tcpm-proportional-rate-reduction/
2480 * It computes the number of packets to send (sndcnt) based on packets newly
2482 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2483 * cwnd reductions across a full RTT.
2484 * 2) If packets in flight is lower than ssthresh (such as due to excess
2485 * losses and/or application stalls), do not perform any further cwnd
2486 * reductions, but instead slow start up to ssthresh.
2488 static void tcp_init_cwnd_reduction(struct sock *sk)
2490 struct tcp_sock *tp = tcp_sk(sk);
2492 tp->high_seq = tp->snd_nxt;
2493 tp->tlp_high_seq = 0;
2494 tp->snd_cwnd_cnt = 0;
2495 tp->prior_cwnd = tp->snd_cwnd;
2496 tp->prr_delivered = 0;
2498 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2499 tcp_ecn_queue_cwr(tp);
2502 static void tcp_cwnd_reduction(struct sock *sk, const int prior_unsacked,
2505 struct tcp_sock *tp = tcp_sk(sk);
2507 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2508 int newly_acked_sacked = prior_unsacked -
2509 (tp->packets_out - tp->sacked_out);
2511 tp->prr_delivered += newly_acked_sacked;
2512 if (tcp_packets_in_flight(tp) > tp->snd_ssthresh) {
2513 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2515 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2517 sndcnt = min_t(int, delta,
2518 max_t(int, tp->prr_delivered - tp->prr_out,
2519 newly_acked_sacked) + 1);
2522 sndcnt = max(sndcnt, (fast_rexmit ? 1 : 0));
2523 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2526 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2528 struct tcp_sock *tp = tcp_sk(sk);
2530 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2531 if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR ||
2532 (tp->undo_marker && tp->snd_ssthresh < TCP_INFINITE_SSTHRESH)) {
2533 tp->snd_cwnd = tp->snd_ssthresh;
2534 tp->snd_cwnd_stamp = tcp_time_stamp;
2536 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2539 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2540 void tcp_enter_cwr(struct sock *sk)
2542 struct tcp_sock *tp = tcp_sk(sk);
2544 tp->prior_ssthresh = 0;
2545 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2546 tp->undo_marker = 0;
2547 tcp_init_cwnd_reduction(sk);
2548 tcp_set_ca_state(sk, TCP_CA_CWR);
2551 EXPORT_SYMBOL(tcp_enter_cwr);
2553 static void tcp_try_keep_open(struct sock *sk)
2555 struct tcp_sock *tp = tcp_sk(sk);
2556 int state = TCP_CA_Open;
2558 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2559 state = TCP_CA_Disorder;
2561 if (inet_csk(sk)->icsk_ca_state != state) {
2562 tcp_set_ca_state(sk, state);
2563 tp->high_seq = tp->snd_nxt;
2567 static void tcp_try_to_open(struct sock *sk, int flag, const int prior_unsacked)
2569 struct tcp_sock *tp = tcp_sk(sk);
2571 tcp_verify_left_out(tp);
2573 if (!tcp_any_retrans_done(sk))
2574 tp->retrans_stamp = 0;
2576 if (flag & FLAG_ECE)
2579 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2580 tcp_try_keep_open(sk);
2582 tcp_cwnd_reduction(sk, prior_unsacked, 0);
2586 static void tcp_mtup_probe_failed(struct sock *sk)
2588 struct inet_connection_sock *icsk = inet_csk(sk);
2590 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2591 icsk->icsk_mtup.probe_size = 0;
2594 static void tcp_mtup_probe_success(struct sock *sk)
2596 struct tcp_sock *tp = tcp_sk(sk);
2597 struct inet_connection_sock *icsk = inet_csk(sk);
2599 /* FIXME: breaks with very large cwnd */
2600 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2601 tp->snd_cwnd = tp->snd_cwnd *
2602 tcp_mss_to_mtu(sk, tp->mss_cache) /
2603 icsk->icsk_mtup.probe_size;
2604 tp->snd_cwnd_cnt = 0;
2605 tp->snd_cwnd_stamp = tcp_time_stamp;
2606 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2608 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2609 icsk->icsk_mtup.probe_size = 0;
2610 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2613 /* Do a simple retransmit without using the backoff mechanisms in
2614 * tcp_timer. This is used for path mtu discovery.
2615 * The socket is already locked here.
2617 void tcp_simple_retransmit(struct sock *sk)
2619 const struct inet_connection_sock *icsk = inet_csk(sk);
2620 struct tcp_sock *tp = tcp_sk(sk);
2621 struct sk_buff *skb;
2622 unsigned int mss = tcp_current_mss(sk);
2623 u32 prior_lost = tp->lost_out;
2625 tcp_for_write_queue(skb, sk) {
2626 if (skb == tcp_send_head(sk))
2628 if (tcp_skb_seglen(skb) > mss &&
2629 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2630 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2631 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2632 tp->retrans_out -= tcp_skb_pcount(skb);
2634 tcp_skb_mark_lost_uncond_verify(tp, skb);
2638 tcp_clear_retrans_hints_partial(tp);
2640 if (prior_lost == tp->lost_out)
2643 if (tcp_is_reno(tp))
2644 tcp_limit_reno_sacked(tp);
2646 tcp_verify_left_out(tp);
2648 /* Don't muck with the congestion window here.
2649 * Reason is that we do not increase amount of _data_
2650 * in network, but units changed and effective
2651 * cwnd/ssthresh really reduced now.
2653 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2654 tp->high_seq = tp->snd_nxt;
2655 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2656 tp->prior_ssthresh = 0;
2657 tp->undo_marker = 0;
2658 tcp_set_ca_state(sk, TCP_CA_Loss);
2660 tcp_xmit_retransmit_queue(sk);
2662 EXPORT_SYMBOL(tcp_simple_retransmit);
2664 static void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2666 struct tcp_sock *tp = tcp_sk(sk);
2669 if (tcp_is_reno(tp))
2670 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2672 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2674 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2676 tp->prior_ssthresh = 0;
2679 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2681 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2682 tcp_init_cwnd_reduction(sk);
2684 tcp_set_ca_state(sk, TCP_CA_Recovery);
2687 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2688 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2690 static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack)
2692 struct tcp_sock *tp = tcp_sk(sk);
2693 bool recovered = !before(tp->snd_una, tp->high_seq);
2695 if ((flag & FLAG_SND_UNA_ADVANCED) &&
2696 tcp_try_undo_loss(sk, false))
2699 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2700 /* Step 3.b. A timeout is spurious if not all data are
2701 * lost, i.e., never-retransmitted data are (s)acked.
2703 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2704 tcp_try_undo_loss(sk, true))
2707 if (after(tp->snd_nxt, tp->high_seq)) {
2708 if (flag & FLAG_DATA_SACKED || is_dupack)
2709 tp->frto = 0; /* Step 3.a. loss was real */
2710 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2711 tp->high_seq = tp->snd_nxt;
2712 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
2714 if (after(tp->snd_nxt, tp->high_seq))
2715 return; /* Step 2.b */
2721 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2722 tcp_try_undo_recovery(sk);
2725 if (tcp_is_reno(tp)) {
2726 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2727 * delivered. Lower inflight to clock out (re)tranmissions.
2729 if (after(tp->snd_nxt, tp->high_seq) && is_dupack)
2730 tcp_add_reno_sack(sk);
2731 else if (flag & FLAG_SND_UNA_ADVANCED)
2732 tcp_reset_reno_sack(tp);
2734 tcp_xmit_retransmit_queue(sk);
2737 /* Undo during fast recovery after partial ACK. */
2738 static bool tcp_try_undo_partial(struct sock *sk, const int acked,
2739 const int prior_unsacked)
2741 struct tcp_sock *tp = tcp_sk(sk);
2743 if (tp->undo_marker && tcp_packet_delayed(tp)) {
2744 /* Plain luck! Hole if filled with delayed
2745 * packet, rather than with a retransmit.
2747 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2749 /* We are getting evidence that the reordering degree is higher
2750 * than we realized. If there are no retransmits out then we
2751 * can undo. Otherwise we clock out new packets but do not
2752 * mark more packets lost or retransmit more.
2754 if (tp->retrans_out) {
2755 tcp_cwnd_reduction(sk, prior_unsacked, 0);
2759 if (!tcp_any_retrans_done(sk))
2760 tp->retrans_stamp = 0;
2762 DBGUNDO(sk, "partial recovery");
2763 tcp_undo_cwnd_reduction(sk, true);
2764 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2765 tcp_try_keep_open(sk);
2771 /* Process an event, which can update packets-in-flight not trivially.
2772 * Main goal of this function is to calculate new estimate for left_out,
2773 * taking into account both packets sitting in receiver's buffer and
2774 * packets lost by network.
2776 * Besides that it does CWND reduction, when packet loss is detected
2777 * and changes state of machine.
2779 * It does _not_ decide what to send, it is made in function
2780 * tcp_xmit_retransmit_queue().
2782 static void tcp_fastretrans_alert(struct sock *sk, const int acked,
2783 const int prior_unsacked,
2784 bool is_dupack, int flag)
2786 struct inet_connection_sock *icsk = inet_csk(sk);
2787 struct tcp_sock *tp = tcp_sk(sk);
2788 bool do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2789 (tcp_fackets_out(tp) > tp->reordering));
2790 int fast_rexmit = 0;
2792 if (WARN_ON(!tp->packets_out && tp->sacked_out))
2794 if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2795 tp->fackets_out = 0;
2797 /* Now state machine starts.
2798 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2799 if (flag & FLAG_ECE)
2800 tp->prior_ssthresh = 0;
2802 /* B. In all the states check for reneging SACKs. */
2803 if (tcp_check_sack_reneging(sk, flag))
2806 /* C. Check consistency of the current state. */
2807 tcp_verify_left_out(tp);
2809 /* D. Check state exit conditions. State can be terminated
2810 * when high_seq is ACKed. */
2811 if (icsk->icsk_ca_state == TCP_CA_Open) {
2812 WARN_ON(tp->retrans_out != 0);
2813 tp->retrans_stamp = 0;
2814 } else if (!before(tp->snd_una, tp->high_seq)) {
2815 switch (icsk->icsk_ca_state) {
2817 /* CWR is to be held something *above* high_seq
2818 * is ACKed for CWR bit to reach receiver. */
2819 if (tp->snd_una != tp->high_seq) {
2820 tcp_end_cwnd_reduction(sk);
2821 tcp_set_ca_state(sk, TCP_CA_Open);
2825 case TCP_CA_Recovery:
2826 if (tcp_is_reno(tp))
2827 tcp_reset_reno_sack(tp);
2828 if (tcp_try_undo_recovery(sk))
2830 tcp_end_cwnd_reduction(sk);
2835 /* E. Process state. */
2836 switch (icsk->icsk_ca_state) {
2837 case TCP_CA_Recovery:
2838 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2839 if (tcp_is_reno(tp) && is_dupack)
2840 tcp_add_reno_sack(sk);
2842 if (tcp_try_undo_partial(sk, acked, prior_unsacked))
2844 /* Partial ACK arrived. Force fast retransmit. */
2845 do_lost = tcp_is_reno(tp) ||
2846 tcp_fackets_out(tp) > tp->reordering;
2848 if (tcp_try_undo_dsack(sk)) {
2849 tcp_try_keep_open(sk);
2854 tcp_process_loss(sk, flag, is_dupack);
2855 if (icsk->icsk_ca_state != TCP_CA_Open)
2857 /* Fall through to processing in Open state. */
2859 if (tcp_is_reno(tp)) {
2860 if (flag & FLAG_SND_UNA_ADVANCED)
2861 tcp_reset_reno_sack(tp);
2863 tcp_add_reno_sack(sk);
2866 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2867 tcp_try_undo_dsack(sk);
2869 if (!tcp_time_to_recover(sk, flag)) {
2870 tcp_try_to_open(sk, flag, prior_unsacked);
2874 /* MTU probe failure: don't reduce cwnd */
2875 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2876 icsk->icsk_mtup.probe_size &&
2877 tp->snd_una == tp->mtu_probe.probe_seq_start) {
2878 tcp_mtup_probe_failed(sk);
2879 /* Restores the reduction we did in tcp_mtup_probe() */
2881 tcp_simple_retransmit(sk);
2885 /* Otherwise enter Recovery state */
2886 tcp_enter_recovery(sk, (flag & FLAG_ECE));
2891 tcp_update_scoreboard(sk, fast_rexmit);
2892 tcp_cwnd_reduction(sk, prior_unsacked, fast_rexmit);
2893 tcp_xmit_retransmit_queue(sk);
2896 static inline bool tcp_ack_update_rtt(struct sock *sk, const int flag,
2897 long seq_rtt_us, long sack_rtt_us)
2899 const struct tcp_sock *tp = tcp_sk(sk);
2901 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2902 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2903 * Karn's algorithm forbids taking RTT if some retransmitted data
2904 * is acked (RFC6298).
2906 if (flag & FLAG_RETRANS_DATA_ACKED)
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 &&
2920 seq_rtt_us = jiffies_to_usecs(tcp_time_stamp - tp->rx_opt.rcv_tsecr);
2925 tcp_rtt_estimator(sk, seq_rtt_us);
2928 /* RFC6298: only reset backoff on valid RTT measurement. */
2929 inet_csk(sk)->icsk_backoff = 0;
2933 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2934 static void tcp_synack_rtt_meas(struct sock *sk, const u32 synack_stamp)
2936 struct tcp_sock *tp = tcp_sk(sk);
2937 long seq_rtt_us = -1L;
2939 if (synack_stamp && !tp->total_retrans)
2940 seq_rtt_us = jiffies_to_usecs(tcp_time_stamp - synack_stamp);
2942 /* If the ACK acks both the SYNACK and the (Fast Open'd) data packets
2943 * sent in SYN_RECV, SYNACK RTT is the smooth RTT computed in tcp_ack()
2946 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, seq_rtt_us, -1L);
2949 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
2951 const struct inet_connection_sock *icsk = inet_csk(sk);
2953 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
2954 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
2957 /* Restart timer after forward progress on connection.
2958 * RFC2988 recommends to restart timer to now+rto.
2960 void tcp_rearm_rto(struct sock *sk)
2962 const struct inet_connection_sock *icsk = inet_csk(sk);
2963 struct tcp_sock *tp = tcp_sk(sk);
2965 /* If the retrans timer is currently being used by Fast Open
2966 * for SYN-ACK retrans purpose, stay put.
2968 if (tp->fastopen_rsk)
2971 if (!tp->packets_out) {
2972 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2974 u32 rto = inet_csk(sk)->icsk_rto;
2975 /* Offset the time elapsed after installing regular RTO */
2976 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
2977 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
2978 struct sk_buff *skb = tcp_write_queue_head(sk);
2979 const u32 rto_time_stamp =
2980 tcp_skb_timestamp(skb) + rto;
2981 s32 delta = (s32)(rto_time_stamp - tcp_time_stamp);
2982 /* delta may not be positive if the socket is locked
2983 * when the retrans timer fires and is rescheduled.
2988 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
2993 /* This function is called when the delayed ER timer fires. TCP enters
2994 * fast recovery and performs fast-retransmit.
2996 void tcp_resume_early_retransmit(struct sock *sk)
2998 struct tcp_sock *tp = tcp_sk(sk);
3002 /* Stop if ER is disabled after the delayed ER timer is scheduled */
3003 if (!tp->do_early_retrans)
3006 tcp_enter_recovery(sk, false);
3007 tcp_update_scoreboard(sk, 1);
3008 tcp_xmit_retransmit_queue(sk);
3011 /* If we get here, the whole TSO packet has not been acked. */
3012 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3014 struct tcp_sock *tp = tcp_sk(sk);
3017 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3019 packets_acked = tcp_skb_pcount(skb);
3020 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3022 packets_acked -= tcp_skb_pcount(skb);
3024 if (packets_acked) {
3025 BUG_ON(tcp_skb_pcount(skb) == 0);
3026 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3029 return packets_acked;
3032 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3035 const struct skb_shared_info *shinfo;
3037 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3038 if (likely(!(sk->sk_tsflags & SOF_TIMESTAMPING_TX_ACK)))
3041 shinfo = skb_shinfo(skb);
3042 if ((shinfo->tx_flags & SKBTX_ACK_TSTAMP) &&
3043 between(shinfo->tskey, prior_snd_una, tcp_sk(sk)->snd_una - 1))
3044 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK);
3047 /* Remove acknowledged frames from the retransmission queue. If our packet
3048 * is before the ack sequence we can discard it as it's confirmed to have
3049 * arrived at the other end.
3051 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3053 struct tcp_sacktag_state *sack)
3055 const struct inet_connection_sock *icsk = inet_csk(sk);
3056 struct skb_mstamp first_ackt, last_ackt, now;
3057 struct tcp_sock *tp = tcp_sk(sk);
3058 u32 prior_sacked = tp->sacked_out;
3059 u32 reord = tp->packets_out;
3060 bool fully_acked = true;
3061 long sack_rtt_us = -1L;
3062 long seq_rtt_us = -1L;
3063 long ca_rtt_us = -1L;
3064 struct sk_buff *skb;
3071 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3072 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3073 u8 sacked = scb->sacked;
3076 tcp_ack_tstamp(sk, skb, prior_snd_una);
3078 /* Determine how many packets and what bytes were acked, tso and else */
3079 if (after(scb->end_seq, tp->snd_una)) {
3080 if (tcp_skb_pcount(skb) == 1 ||
3081 !after(tp->snd_una, scb->seq))
3084 acked_pcount = tcp_tso_acked(sk, skb);
3088 fully_acked = false;
3090 /* Speedup tcp_unlink_write_queue() and next loop */
3091 prefetchw(skb->next);
3092 acked_pcount = tcp_skb_pcount(skb);
3095 if (unlikely(sacked & TCPCB_RETRANS)) {
3096 if (sacked & TCPCB_SACKED_RETRANS)
3097 tp->retrans_out -= acked_pcount;
3098 flag |= FLAG_RETRANS_DATA_ACKED;
3099 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3100 last_ackt = skb->skb_mstamp;
3101 WARN_ON_ONCE(last_ackt.v64 == 0);
3102 if (!first_ackt.v64)
3103 first_ackt = last_ackt;
3105 reord = min(pkts_acked, reord);
3106 if (!after(scb->end_seq, tp->high_seq))
3107 flag |= FLAG_ORIG_SACK_ACKED;
3110 if (sacked & TCPCB_SACKED_ACKED)
3111 tp->sacked_out -= acked_pcount;
3112 if (sacked & TCPCB_LOST)
3113 tp->lost_out -= acked_pcount;
3115 tp->packets_out -= acked_pcount;
3116 pkts_acked += acked_pcount;
3118 /* Initial outgoing SYN's get put onto the write_queue
3119 * just like anything else we transmit. It is not
3120 * true data, and if we misinform our callers that
3121 * this ACK acks real data, we will erroneously exit
3122 * connection startup slow start one packet too
3123 * quickly. This is severely frowned upon behavior.
3125 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3126 flag |= FLAG_DATA_ACKED;
3128 flag |= FLAG_SYN_ACKED;
3129 tp->retrans_stamp = 0;
3135 tcp_unlink_write_queue(skb, sk);
3136 sk_wmem_free_skb(sk, skb);
3137 if (unlikely(skb == tp->retransmit_skb_hint))
3138 tp->retransmit_skb_hint = NULL;
3139 if (unlikely(skb == tp->lost_skb_hint))
3140 tp->lost_skb_hint = NULL;
3143 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3144 tp->snd_up = tp->snd_una;
3146 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3147 flag |= FLAG_SACK_RENEGING;
3149 skb_mstamp_get(&now);
3150 if (likely(first_ackt.v64)) {
3151 seq_rtt_us = skb_mstamp_us_delta(&now, &first_ackt);
3152 ca_rtt_us = skb_mstamp_us_delta(&now, &last_ackt);
3154 if (sack->first_sackt.v64) {
3155 sack_rtt_us = skb_mstamp_us_delta(&now, &sack->first_sackt);
3156 ca_rtt_us = skb_mstamp_us_delta(&now, &sack->last_sackt);
3159 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us);
3161 if (flag & FLAG_ACKED) {
3163 if (unlikely(icsk->icsk_mtup.probe_size &&
3164 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3165 tcp_mtup_probe_success(sk);
3168 if (tcp_is_reno(tp)) {
3169 tcp_remove_reno_sacks(sk, pkts_acked);
3173 /* Non-retransmitted hole got filled? That's reordering */
3174 if (reord < prior_fackets)
3175 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3177 delta = tcp_is_fack(tp) ? pkts_acked :
3178 prior_sacked - tp->sacked_out;
3179 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3182 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3184 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3185 sack_rtt_us > skb_mstamp_us_delta(&now, &skb->skb_mstamp)) {
3186 /* Do not re-arm RTO if the sack RTT is measured from data sent
3187 * after when the head was last (re)transmitted. Otherwise the
3188 * timeout may continue to extend in loss recovery.
3193 if (icsk->icsk_ca_ops->pkts_acked)
3194 icsk->icsk_ca_ops->pkts_acked(sk, pkts_acked, ca_rtt_us);
3196 #if FASTRETRANS_DEBUG > 0
3197 WARN_ON((int)tp->sacked_out < 0);
3198 WARN_ON((int)tp->lost_out < 0);
3199 WARN_ON((int)tp->retrans_out < 0);
3200 if (!tp->packets_out && tcp_is_sack(tp)) {
3201 icsk = inet_csk(sk);
3203 pr_debug("Leak l=%u %d\n",
3204 tp->lost_out, icsk->icsk_ca_state);
3207 if (tp->sacked_out) {
3208 pr_debug("Leak s=%u %d\n",
3209 tp->sacked_out, icsk->icsk_ca_state);
3212 if (tp->retrans_out) {
3213 pr_debug("Leak r=%u %d\n",
3214 tp->retrans_out, icsk->icsk_ca_state);
3215 tp->retrans_out = 0;
3222 static void tcp_ack_probe(struct sock *sk)
3224 const struct tcp_sock *tp = tcp_sk(sk);
3225 struct inet_connection_sock *icsk = inet_csk(sk);
3227 /* Was it a usable window open? */
3229 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3230 icsk->icsk_backoff = 0;
3231 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3232 /* Socket must be waked up by subsequent tcp_data_snd_check().
3233 * This function is not for random using!
3236 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
3238 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3243 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3245 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3246 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3249 /* Decide wheather to run the increase function of congestion control. */
3250 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3252 if (tcp_in_cwnd_reduction(sk))
3255 /* If reordering is high then always grow cwnd whenever data is
3256 * delivered regardless of its ordering. Otherwise stay conservative
3257 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3258 * new SACK or ECE mark may first advance cwnd here and later reduce
3259 * cwnd in tcp_fastretrans_alert() based on more states.
3261 if (tcp_sk(sk)->reordering > sysctl_tcp_reordering)
3262 return flag & FLAG_FORWARD_PROGRESS;
3264 return flag & FLAG_DATA_ACKED;
3267 /* Check that window update is acceptable.
3268 * The function assumes that snd_una<=ack<=snd_next.
3270 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3271 const u32 ack, const u32 ack_seq,
3274 return after(ack, tp->snd_una) ||
3275 after(ack_seq, tp->snd_wl1) ||
3276 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3279 /* If we update tp->snd_una, also update tp->bytes_acked */
3280 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3282 u32 delta = ack - tp->snd_una;
3284 u64_stats_update_begin(&tp->syncp);
3285 tp->bytes_acked += delta;
3286 u64_stats_update_end(&tp->syncp);
3290 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3291 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3293 u32 delta = seq - tp->rcv_nxt;
3295 u64_stats_update_begin(&tp->syncp);
3296 tp->bytes_received += delta;
3297 u64_stats_update_end(&tp->syncp);
3301 /* Update our send window.
3303 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3304 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3306 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3309 struct tcp_sock *tp = tcp_sk(sk);
3311 u32 nwin = ntohs(tcp_hdr(skb)->window);
3313 if (likely(!tcp_hdr(skb)->syn))
3314 nwin <<= tp->rx_opt.snd_wscale;
3316 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3317 flag |= FLAG_WIN_UPDATE;
3318 tcp_update_wl(tp, ack_seq);
3320 if (tp->snd_wnd != nwin) {
3323 /* Note, it is the only place, where
3324 * fast path is recovered for sending TCP.
3327 tcp_fast_path_check(sk);
3329 if (nwin > tp->max_window) {
3330 tp->max_window = nwin;
3331 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3336 tcp_snd_una_update(tp, ack);
3341 /* Return true if we're currently rate-limiting out-of-window ACKs and
3342 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3343 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3344 * attacks that send repeated SYNs or ACKs for the same connection. To
3345 * do this, we do not send a duplicate SYNACK or ACK if the remote
3346 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3348 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3349 int mib_idx, u32 *last_oow_ack_time)
3351 /* Data packets without SYNs are not likely part of an ACK loop. */
3352 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3354 goto not_rate_limited;
3356 if (*last_oow_ack_time) {
3357 s32 elapsed = (s32)(tcp_time_stamp - *last_oow_ack_time);
3359 if (0 <= elapsed && elapsed < sysctl_tcp_invalid_ratelimit) {
3360 NET_INC_STATS_BH(net, mib_idx);
3361 return true; /* rate-limited: don't send yet! */
3365 *last_oow_ack_time = tcp_time_stamp;
3368 return false; /* not rate-limited: go ahead, send dupack now! */
3371 /* RFC 5961 7 [ACK Throttling] */
3372 static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb)
3374 /* unprotected vars, we dont care of overwrites */
3375 static u32 challenge_timestamp;
3376 static unsigned int challenge_count;
3377 struct tcp_sock *tp = tcp_sk(sk);
3380 /* First check our per-socket dupack rate limit. */
3381 if (tcp_oow_rate_limited(sock_net(sk), skb,
3382 LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3383 &tp->last_oow_ack_time))
3386 /* Then check the check host-wide RFC 5961 rate limit. */
3388 if (now != challenge_timestamp) {
3389 challenge_timestamp = now;
3390 challenge_count = 0;
3392 if (++challenge_count <= sysctl_tcp_challenge_ack_limit) {
3393 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPCHALLENGEACK);
3398 static void tcp_store_ts_recent(struct tcp_sock *tp)
3400 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3401 tp->rx_opt.ts_recent_stamp = get_seconds();
3404 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3406 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3407 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3408 * extra check below makes sure this can only happen
3409 * for pure ACK frames. -DaveM
3411 * Not only, also it occurs for expired timestamps.
3414 if (tcp_paws_check(&tp->rx_opt, 0))
3415 tcp_store_ts_recent(tp);
3419 /* This routine deals with acks during a TLP episode.
3420 * We mark the end of a TLP episode on receiving TLP dupack or when
3421 * ack is after tlp_high_seq.
3422 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3424 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3426 struct tcp_sock *tp = tcp_sk(sk);
3428 if (before(ack, tp->tlp_high_seq))
3431 if (flag & FLAG_DSACKING_ACK) {
3432 /* This DSACK means original and TLP probe arrived; no loss */
3433 tp->tlp_high_seq = 0;
3434 } else if (after(ack, tp->tlp_high_seq)) {
3435 /* ACK advances: there was a loss, so reduce cwnd. Reset
3436 * tlp_high_seq in tcp_init_cwnd_reduction()
3438 tcp_init_cwnd_reduction(sk);
3439 tcp_set_ca_state(sk, TCP_CA_CWR);
3440 tcp_end_cwnd_reduction(sk);
3441 tcp_try_keep_open(sk);
3442 NET_INC_STATS_BH(sock_net(sk),
3443 LINUX_MIB_TCPLOSSPROBERECOVERY);
3444 } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3445 FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3446 /* Pure dupack: original and TLP probe arrived; no loss */
3447 tp->tlp_high_seq = 0;
3451 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3453 const struct inet_connection_sock *icsk = inet_csk(sk);
3455 if (icsk->icsk_ca_ops->in_ack_event)
3456 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3459 /* This routine deals with incoming acks, but not outgoing ones. */
3460 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3462 struct inet_connection_sock *icsk = inet_csk(sk);
3463 struct tcp_sock *tp = tcp_sk(sk);
3464 struct tcp_sacktag_state sack_state;
3465 u32 prior_snd_una = tp->snd_una;
3466 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3467 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3468 bool is_dupack = false;
3470 int prior_packets = tp->packets_out;
3471 const int prior_unsacked = tp->packets_out - tp->sacked_out;
3472 int acked = 0; /* Number of packets newly acked */
3474 sack_state.first_sackt.v64 = 0;
3476 /* We very likely will need to access write queue head. */
3477 prefetchw(sk->sk_write_queue.next);
3479 /* If the ack is older than previous acks
3480 * then we can probably ignore it.
3482 if (before(ack, prior_snd_una)) {
3483 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3484 if (before(ack, prior_snd_una - tp->max_window)) {
3485 tcp_send_challenge_ack(sk, skb);
3491 /* If the ack includes data we haven't sent yet, discard
3492 * this segment (RFC793 Section 3.9).
3494 if (after(ack, tp->snd_nxt))
3497 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
3498 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
3501 if (after(ack, prior_snd_una)) {
3502 flag |= FLAG_SND_UNA_ADVANCED;
3503 icsk->icsk_retransmits = 0;
3506 prior_fackets = tp->fackets_out;
3508 /* ts_recent update must be made after we are sure that the packet
3511 if (flag & FLAG_UPDATE_TS_RECENT)
3512 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3514 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3515 /* Window is constant, pure forward advance.
3516 * No more checks are required.
3517 * Note, we use the fact that SND.UNA>=SND.WL2.
3519 tcp_update_wl(tp, ack_seq);
3520 tcp_snd_una_update(tp, ack);
3521 flag |= FLAG_WIN_UPDATE;
3523 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3525 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3527 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3529 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3532 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3534 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3536 if (TCP_SKB_CB(skb)->sacked)
3537 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3540 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3542 ack_ev_flags |= CA_ACK_ECE;
3545 if (flag & FLAG_WIN_UPDATE)
3546 ack_ev_flags |= CA_ACK_WIN_UPDATE;
3548 tcp_in_ack_event(sk, ack_ev_flags);
3551 /* We passed data and got it acked, remove any soft error
3552 * log. Something worked...
3554 sk->sk_err_soft = 0;
3555 icsk->icsk_probes_out = 0;
3556 tp->rcv_tstamp = tcp_time_stamp;
3560 /* See if we can take anything off of the retransmit queue. */
3561 acked = tp->packets_out;
3562 flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una,
3564 acked -= tp->packets_out;
3566 /* Advance cwnd if state allows */
3567 if (tcp_may_raise_cwnd(sk, flag))
3568 tcp_cong_avoid(sk, ack, acked);
3570 if (tcp_ack_is_dubious(sk, flag)) {
3571 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3572 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3575 if (tp->tlp_high_seq)
3576 tcp_process_tlp_ack(sk, ack, flag);
3578 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) {
3579 struct dst_entry *dst = __sk_dst_get(sk);
3584 if (icsk->icsk_pending == ICSK_TIME_RETRANS)
3585 tcp_schedule_loss_probe(sk);
3586 tcp_update_pacing_rate(sk);
3590 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3591 if (flag & FLAG_DSACKING_ACK)
3592 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3594 /* If this ack opens up a zero window, clear backoff. It was
3595 * being used to time the probes, and is probably far higher than
3596 * it needs to be for normal retransmission.
3598 if (tcp_send_head(sk))
3601 if (tp->tlp_high_seq)
3602 tcp_process_tlp_ack(sk, ack, flag);
3606 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3610 /* If data was SACKed, tag it and see if we should send more data.
3611 * If data was DSACKed, see if we can undo a cwnd reduction.
3613 if (TCP_SKB_CB(skb)->sacked) {
3614 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3616 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3620 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3624 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3625 bool syn, struct tcp_fastopen_cookie *foc,
3628 /* Valid only in SYN or SYN-ACK with an even length. */
3629 if (!foc || !syn || len < 0 || (len & 1))
3632 if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3633 len <= TCP_FASTOPEN_COOKIE_MAX)
3634 memcpy(foc->val, cookie, len);
3641 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3642 * But, this can also be called on packets in the established flow when
3643 * the fast version below fails.
3645 void tcp_parse_options(const struct sk_buff *skb,
3646 struct tcp_options_received *opt_rx, int estab,
3647 struct tcp_fastopen_cookie *foc)
3649 const unsigned char *ptr;
3650 const struct tcphdr *th = tcp_hdr(skb);
3651 int length = (th->doff * 4) - sizeof(struct tcphdr);
3653 ptr = (const unsigned char *)(th + 1);
3654 opt_rx->saw_tstamp = 0;
3656 while (length > 0) {
3657 int opcode = *ptr++;
3663 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3668 if (opsize < 2) /* "silly options" */
3670 if (opsize > length)
3671 return; /* don't parse partial options */
3674 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3675 u16 in_mss = get_unaligned_be16(ptr);
3677 if (opt_rx->user_mss &&
3678 opt_rx->user_mss < in_mss)
3679 in_mss = opt_rx->user_mss;
3680 opt_rx->mss_clamp = in_mss;
3685 if (opsize == TCPOLEN_WINDOW && th->syn &&
3686 !estab && sysctl_tcp_window_scaling) {
3687 __u8 snd_wscale = *(__u8 *)ptr;
3688 opt_rx->wscale_ok = 1;
3689 if (snd_wscale > 14) {
3690 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3695 opt_rx->snd_wscale = snd_wscale;
3698 case TCPOPT_TIMESTAMP:
3699 if ((opsize == TCPOLEN_TIMESTAMP) &&
3700 ((estab && opt_rx->tstamp_ok) ||
3701 (!estab && sysctl_tcp_timestamps))) {
3702 opt_rx->saw_tstamp = 1;
3703 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3704 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3707 case TCPOPT_SACK_PERM:
3708 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3709 !estab && sysctl_tcp_sack) {
3710 opt_rx->sack_ok = TCP_SACK_SEEN;
3711 tcp_sack_reset(opt_rx);
3716 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3717 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3719 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3722 #ifdef CONFIG_TCP_MD5SIG
3725 * The MD5 Hash has already been
3726 * checked (see tcp_v{4,6}_do_rcv()).
3730 case TCPOPT_FASTOPEN:
3731 tcp_parse_fastopen_option(
3732 opsize - TCPOLEN_FASTOPEN_BASE,
3733 ptr, th->syn, foc, false);
3737 /* Fast Open option shares code 254 using a
3738 * 16 bits magic number.
3740 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
3741 get_unaligned_be16(ptr) ==
3742 TCPOPT_FASTOPEN_MAGIC)
3743 tcp_parse_fastopen_option(opsize -
3744 TCPOLEN_EXP_FASTOPEN_BASE,
3745 ptr + 2, th->syn, foc, true);
3754 EXPORT_SYMBOL(tcp_parse_options);
3756 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3758 const __be32 *ptr = (const __be32 *)(th + 1);
3760 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3761 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3762 tp->rx_opt.saw_tstamp = 1;
3764 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3767 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3769 tp->rx_opt.rcv_tsecr = 0;
3775 /* Fast parse options. This hopes to only see timestamps.
3776 * If it is wrong it falls back on tcp_parse_options().
3778 static bool tcp_fast_parse_options(const struct sk_buff *skb,
3779 const struct tcphdr *th, struct tcp_sock *tp)
3781 /* In the spirit of fast parsing, compare doff directly to constant
3782 * values. Because equality is used, short doff can be ignored here.
3784 if (th->doff == (sizeof(*th) / 4)) {
3785 tp->rx_opt.saw_tstamp = 0;
3787 } else if (tp->rx_opt.tstamp_ok &&
3788 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3789 if (tcp_parse_aligned_timestamp(tp, th))
3793 tcp_parse_options(skb, &tp->rx_opt, 1, NULL);
3794 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3795 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3800 #ifdef CONFIG_TCP_MD5SIG
3802 * Parse MD5 Signature option
3804 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3806 int length = (th->doff << 2) - sizeof(*th);
3807 const u8 *ptr = (const u8 *)(th + 1);
3809 /* If the TCP option is too short, we can short cut */
3810 if (length < TCPOLEN_MD5SIG)
3813 while (length > 0) {
3814 int opcode = *ptr++;
3825 if (opsize < 2 || opsize > length)
3827 if (opcode == TCPOPT_MD5SIG)
3828 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3835 EXPORT_SYMBOL(tcp_parse_md5sig_option);
3838 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3840 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3841 * it can pass through stack. So, the following predicate verifies that
3842 * this segment is not used for anything but congestion avoidance or
3843 * fast retransmit. Moreover, we even are able to eliminate most of such
3844 * second order effects, if we apply some small "replay" window (~RTO)
3845 * to timestamp space.
3847 * All these measures still do not guarantee that we reject wrapped ACKs
3848 * on networks with high bandwidth, when sequence space is recycled fastly,
3849 * but it guarantees that such events will be very rare and do not affect
3850 * connection seriously. This doesn't look nice, but alas, PAWS is really
3853 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3854 * states that events when retransmit arrives after original data are rare.
3855 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3856 * the biggest problem on large power networks even with minor reordering.
3857 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3858 * up to bandwidth of 18Gigabit/sec. 8) ]
3861 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3863 const struct tcp_sock *tp = tcp_sk(sk);
3864 const struct tcphdr *th = tcp_hdr(skb);
3865 u32 seq = TCP_SKB_CB(skb)->seq;
3866 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3868 return (/* 1. Pure ACK with correct sequence number. */
3869 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3871 /* 2. ... and duplicate ACK. */
3872 ack == tp->snd_una &&
3874 /* 3. ... and does not update window. */
3875 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3877 /* 4. ... and sits in replay window. */
3878 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3881 static inline bool tcp_paws_discard(const struct sock *sk,
3882 const struct sk_buff *skb)
3884 const struct tcp_sock *tp = tcp_sk(sk);
3886 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
3887 !tcp_disordered_ack(sk, skb);
3890 /* Check segment sequence number for validity.
3892 * Segment controls are considered valid, if the segment
3893 * fits to the window after truncation to the window. Acceptability
3894 * of data (and SYN, FIN, of course) is checked separately.
3895 * See tcp_data_queue(), for example.
3897 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3898 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3899 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3900 * (borrowed from freebsd)
3903 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
3905 return !before(end_seq, tp->rcv_wup) &&
3906 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3909 /* When we get a reset we do this. */
3910 void tcp_reset(struct sock *sk)
3912 /* We want the right error as BSD sees it (and indeed as we do). */
3913 switch (sk->sk_state) {
3915 sk->sk_err = ECONNREFUSED;
3917 case TCP_CLOSE_WAIT:
3923 sk->sk_err = ECONNRESET;
3925 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3928 if (!sock_flag(sk, SOCK_DEAD))
3929 sk->sk_error_report(sk);
3935 * Process the FIN bit. This now behaves as it is supposed to work
3936 * and the FIN takes effect when it is validly part of sequence
3937 * space. Not before when we get holes.
3939 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3940 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3943 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3944 * close and we go into CLOSING (and later onto TIME-WAIT)
3946 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3948 static void tcp_fin(struct sock *sk)
3950 struct tcp_sock *tp = tcp_sk(sk);
3951 const struct dst_entry *dst;
3953 inet_csk_schedule_ack(sk);
3955 sk->sk_shutdown |= RCV_SHUTDOWN;
3956 sock_set_flag(sk, SOCK_DONE);
3958 switch (sk->sk_state) {
3960 case TCP_ESTABLISHED:
3961 /* Move to CLOSE_WAIT */
3962 tcp_set_state(sk, TCP_CLOSE_WAIT);
3963 dst = __sk_dst_get(sk);
3964 if (!dst || !dst_metric(dst, RTAX_QUICKACK))
3965 inet_csk(sk)->icsk_ack.pingpong = 1;
3968 case TCP_CLOSE_WAIT:
3970 /* Received a retransmission of the FIN, do
3975 /* RFC793: Remain in the LAST-ACK state. */
3979 /* This case occurs when a simultaneous close
3980 * happens, we must ack the received FIN and
3981 * enter the CLOSING state.
3984 tcp_set_state(sk, TCP_CLOSING);
3987 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3989 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
3992 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3993 * cases we should never reach this piece of code.
3995 pr_err("%s: Impossible, sk->sk_state=%d\n",
3996 __func__, sk->sk_state);
4000 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4001 * Probably, we should reset in this case. For now drop them.
4003 __skb_queue_purge(&tp->out_of_order_queue);
4004 if (tcp_is_sack(tp))
4005 tcp_sack_reset(&tp->rx_opt);
4008 if (!sock_flag(sk, SOCK_DEAD)) {
4009 sk->sk_state_change(sk);
4011 /* Do not send POLL_HUP for half duplex close. */
4012 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4013 sk->sk_state == TCP_CLOSE)
4014 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4016 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4020 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4023 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4024 if (before(seq, sp->start_seq))
4025 sp->start_seq = seq;
4026 if (after(end_seq, sp->end_seq))
4027 sp->end_seq = end_seq;
4033 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4035 struct tcp_sock *tp = tcp_sk(sk);
4037 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4040 if (before(seq, tp->rcv_nxt))
4041 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4043 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4045 NET_INC_STATS_BH(sock_net(sk), mib_idx);
4047 tp->rx_opt.dsack = 1;
4048 tp->duplicate_sack[0].start_seq = seq;
4049 tp->duplicate_sack[0].end_seq = end_seq;
4053 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4055 struct tcp_sock *tp = tcp_sk(sk);
4057 if (!tp->rx_opt.dsack)
4058 tcp_dsack_set(sk, seq, end_seq);
4060 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4063 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4065 struct tcp_sock *tp = tcp_sk(sk);
4067 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4068 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4069 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4070 tcp_enter_quickack_mode(sk);
4072 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4073 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4075 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4076 end_seq = tp->rcv_nxt;
4077 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4084 /* These routines update the SACK block as out-of-order packets arrive or
4085 * in-order packets close up the sequence space.
4087 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4090 struct tcp_sack_block *sp = &tp->selective_acks[0];
4091 struct tcp_sack_block *swalk = sp + 1;
4093 /* See if the recent change to the first SACK eats into
4094 * or hits the sequence space of other SACK blocks, if so coalesce.
4096 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4097 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4100 /* Zap SWALK, by moving every further SACK up by one slot.
4101 * Decrease num_sacks.
4103 tp->rx_opt.num_sacks--;
4104 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4108 this_sack++, swalk++;
4112 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4114 struct tcp_sock *tp = tcp_sk(sk);
4115 struct tcp_sack_block *sp = &tp->selective_acks[0];
4116 int cur_sacks = tp->rx_opt.num_sacks;
4122 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4123 if (tcp_sack_extend(sp, seq, end_seq)) {
4124 /* Rotate this_sack to the first one. */
4125 for (; this_sack > 0; this_sack--, sp--)
4126 swap(*sp, *(sp - 1));
4128 tcp_sack_maybe_coalesce(tp);
4133 /* Could not find an adjacent existing SACK, build a new one,
4134 * put it at the front, and shift everyone else down. We
4135 * always know there is at least one SACK present already here.
4137 * If the sack array is full, forget about the last one.
4139 if (this_sack >= TCP_NUM_SACKS) {
4141 tp->rx_opt.num_sacks--;
4144 for (; this_sack > 0; this_sack--, sp--)
4148 /* Build the new head SACK, and we're done. */
4149 sp->start_seq = seq;
4150 sp->end_seq = end_seq;
4151 tp->rx_opt.num_sacks++;
4154 /* RCV.NXT advances, some SACKs should be eaten. */
4156 static void tcp_sack_remove(struct tcp_sock *tp)
4158 struct tcp_sack_block *sp = &tp->selective_acks[0];
4159 int num_sacks = tp->rx_opt.num_sacks;
4162 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4163 if (skb_queue_empty(&tp->out_of_order_queue)) {
4164 tp->rx_opt.num_sacks = 0;
4168 for (this_sack = 0; this_sack < num_sacks;) {
4169 /* Check if the start of the sack is covered by RCV.NXT. */
4170 if (!before(tp->rcv_nxt, sp->start_seq)) {
4173 /* RCV.NXT must cover all the block! */
4174 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4176 /* Zap this SACK, by moving forward any other SACKS. */
4177 for (i = this_sack+1; i < num_sacks; i++)
4178 tp->selective_acks[i-1] = tp->selective_acks[i];
4185 tp->rx_opt.num_sacks = num_sacks;
4189 * tcp_try_coalesce - try to merge skb to prior one
4192 * @from: buffer to add in queue
4193 * @fragstolen: pointer to boolean
4195 * Before queueing skb @from after @to, try to merge them
4196 * to reduce overall memory use and queue lengths, if cost is small.
4197 * Packets in ofo or receive queues can stay a long time.
4198 * Better try to coalesce them right now to avoid future collapses.
4199 * Returns true if caller should free @from instead of queueing it
4201 static bool tcp_try_coalesce(struct sock *sk,
4203 struct sk_buff *from,
4208 *fragstolen = false;
4210 /* Its possible this segment overlaps with prior segment in queue */
4211 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4214 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4217 atomic_add(delta, &sk->sk_rmem_alloc);
4218 sk_mem_charge(sk, delta);
4219 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4220 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4221 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4222 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4226 /* This one checks to see if we can put data from the
4227 * out_of_order queue into the receive_queue.
4229 static void tcp_ofo_queue(struct sock *sk)
4231 struct tcp_sock *tp = tcp_sk(sk);
4232 __u32 dsack_high = tp->rcv_nxt;
4233 struct sk_buff *skb, *tail;
4234 bool fragstolen, eaten;
4236 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4237 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4240 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4241 __u32 dsack = dsack_high;
4242 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4243 dsack_high = TCP_SKB_CB(skb)->end_seq;
4244 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4247 __skb_unlink(skb, &tp->out_of_order_queue);
4248 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4249 SOCK_DEBUG(sk, "ofo packet was already received\n");
4253 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4254 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4255 TCP_SKB_CB(skb)->end_seq);
4257 tail = skb_peek_tail(&sk->sk_receive_queue);
4258 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4259 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4261 __skb_queue_tail(&sk->sk_receive_queue, skb);
4262 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4265 kfree_skb_partial(skb, fragstolen);
4269 static bool tcp_prune_ofo_queue(struct sock *sk);
4270 static int tcp_prune_queue(struct sock *sk);
4272 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4275 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4276 !sk_rmem_schedule(sk, skb, size)) {
4278 if (tcp_prune_queue(sk) < 0)
4281 if (!sk_rmem_schedule(sk, skb, size)) {
4282 if (!tcp_prune_ofo_queue(sk))
4285 if (!sk_rmem_schedule(sk, skb, size))
4292 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4294 struct tcp_sock *tp = tcp_sk(sk);
4295 struct sk_buff *skb1;
4298 tcp_ecn_check_ce(tp, skb);
4300 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4301 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFODROP);
4306 /* Disable header prediction. */
4308 inet_csk_schedule_ack(sk);
4310 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4311 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4312 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4314 skb1 = skb_peek_tail(&tp->out_of_order_queue);
4316 /* Initial out of order segment, build 1 SACK. */
4317 if (tcp_is_sack(tp)) {
4318 tp->rx_opt.num_sacks = 1;
4319 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4320 tp->selective_acks[0].end_seq =
4321 TCP_SKB_CB(skb)->end_seq;
4323 __skb_queue_head(&tp->out_of_order_queue, skb);
4327 seq = TCP_SKB_CB(skb)->seq;
4328 end_seq = TCP_SKB_CB(skb)->end_seq;
4330 if (seq == TCP_SKB_CB(skb1)->end_seq) {
4333 if (!tcp_try_coalesce(sk, skb1, skb, &fragstolen)) {
4334 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4336 tcp_grow_window(sk, skb);
4337 kfree_skb_partial(skb, fragstolen);
4341 if (!tp->rx_opt.num_sacks ||
4342 tp->selective_acks[0].end_seq != seq)
4345 /* Common case: data arrive in order after hole. */
4346 tp->selective_acks[0].end_seq = end_seq;
4350 /* Find place to insert this segment. */
4352 if (!after(TCP_SKB_CB(skb1)->seq, seq))
4354 if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
4358 skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
4361 /* Do skb overlap to previous one? */
4362 if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4363 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4364 /* All the bits are present. Drop. */
4365 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4368 tcp_dsack_set(sk, seq, end_seq);
4371 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4372 /* Partial overlap. */
4373 tcp_dsack_set(sk, seq,
4374 TCP_SKB_CB(skb1)->end_seq);
4376 if (skb_queue_is_first(&tp->out_of_order_queue,
4380 skb1 = skb_queue_prev(
4381 &tp->out_of_order_queue,
4386 __skb_queue_head(&tp->out_of_order_queue, skb);
4388 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4390 /* And clean segments covered by new one as whole. */
4391 while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
4392 skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
4394 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4396 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4397 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4401 __skb_unlink(skb1, &tp->out_of_order_queue);
4402 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4403 TCP_SKB_CB(skb1)->end_seq);
4404 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4409 if (tcp_is_sack(tp))
4410 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4413 tcp_grow_window(sk, skb);
4414 skb_set_owner_r(skb, sk);
4418 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
4422 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4424 __skb_pull(skb, hdrlen);
4426 tcp_try_coalesce(sk, tail, skb, fragstolen)) ? 1 : 0;
4427 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
4429 __skb_queue_tail(&sk->sk_receive_queue, skb);
4430 skb_set_owner_r(skb, sk);
4435 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4437 struct sk_buff *skb;
4443 skb = alloc_skb(size, sk->sk_allocation);
4447 if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4450 if (memcpy_from_msg(skb_put(skb, size), msg, size))
4453 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4454 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4455 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4457 if (tcp_queue_rcv(sk, skb, 0, &fragstolen)) {
4458 WARN_ON_ONCE(fragstolen); /* should not happen */
4469 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4471 struct tcp_sock *tp = tcp_sk(sk);
4473 bool fragstolen = false;
4475 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4479 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
4481 tcp_ecn_accept_cwr(tp, skb);
4483 tp->rx_opt.dsack = 0;
4485 /* Queue data for delivery to the user.
4486 * Packets in sequence go to the receive queue.
4487 * Out of sequence packets to the out_of_order_queue.
4489 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4490 if (tcp_receive_window(tp) == 0)
4493 /* Ok. In sequence. In window. */
4494 if (tp->ucopy.task == current &&
4495 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4496 sock_owned_by_user(sk) && !tp->urg_data) {
4497 int chunk = min_t(unsigned int, skb->len,
4500 __set_current_state(TASK_RUNNING);
4503 if (!skb_copy_datagram_msg(skb, 0, tp->ucopy.msg, chunk)) {
4504 tp->ucopy.len -= chunk;
4505 tp->copied_seq += chunk;
4506 eaten = (chunk == skb->len);
4507 tcp_rcv_space_adjust(sk);
4515 if (skb_queue_len(&sk->sk_receive_queue) == 0)
4516 sk_forced_mem_schedule(sk, skb->truesize);
4517 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4520 eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
4522 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4524 tcp_event_data_recv(sk, skb);
4525 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4528 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4531 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4532 * gap in queue is filled.
4534 if (skb_queue_empty(&tp->out_of_order_queue))
4535 inet_csk(sk)->icsk_ack.pingpong = 0;
4538 if (tp->rx_opt.num_sacks)
4539 tcp_sack_remove(tp);
4541 tcp_fast_path_check(sk);
4544 kfree_skb_partial(skb, fragstolen);
4545 if (!sock_flag(sk, SOCK_DEAD))
4546 sk->sk_data_ready(sk);
4550 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4551 /* A retransmit, 2nd most common case. Force an immediate ack. */
4552 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4553 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4556 tcp_enter_quickack_mode(sk);
4557 inet_csk_schedule_ack(sk);
4563 /* Out of window. F.e. zero window probe. */
4564 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4567 tcp_enter_quickack_mode(sk);
4569 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4570 /* Partial packet, seq < rcv_next < end_seq */
4571 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4572 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4573 TCP_SKB_CB(skb)->end_seq);
4575 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4577 /* If window is closed, drop tail of packet. But after
4578 * remembering D-SACK for its head made in previous line.
4580 if (!tcp_receive_window(tp))
4585 tcp_data_queue_ofo(sk, skb);
4588 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4589 struct sk_buff_head *list)
4591 struct sk_buff *next = NULL;
4593 if (!skb_queue_is_last(list, skb))
4594 next = skb_queue_next(list, skb);
4596 __skb_unlink(skb, list);
4598 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4603 /* Collapse contiguous sequence of skbs head..tail with
4604 * sequence numbers start..end.
4606 * If tail is NULL, this means until the end of the list.
4608 * Segments with FIN/SYN are not collapsed (only because this
4612 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4613 struct sk_buff *head, struct sk_buff *tail,
4616 struct sk_buff *skb, *n;
4619 /* First, check that queue is collapsible and find
4620 * the point where collapsing can be useful. */
4624 skb_queue_walk_from_safe(list, skb, n) {
4627 /* No new bits? It is possible on ofo queue. */
4628 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4629 skb = tcp_collapse_one(sk, skb, list);
4635 /* The first skb to collapse is:
4637 * - bloated or contains data before "start" or
4638 * overlaps to the next one.
4640 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
4641 (tcp_win_from_space(skb->truesize) > skb->len ||
4642 before(TCP_SKB_CB(skb)->seq, start))) {
4643 end_of_skbs = false;
4647 if (!skb_queue_is_last(list, skb)) {
4648 struct sk_buff *next = skb_queue_next(list, skb);
4650 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
4651 end_of_skbs = false;
4656 /* Decided to skip this, advance start seq. */
4657 start = TCP_SKB_CB(skb)->end_seq;
4660 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4663 while (before(start, end)) {
4664 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
4665 struct sk_buff *nskb;
4667 nskb = alloc_skb(copy, GFP_ATOMIC);
4671 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4672 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4673 __skb_queue_before(list, skb, nskb);
4674 skb_set_owner_r(nskb, sk);
4676 /* Copy data, releasing collapsed skbs. */
4678 int offset = start - TCP_SKB_CB(skb)->seq;
4679 int size = TCP_SKB_CB(skb)->end_seq - start;
4683 size = min(copy, size);
4684 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4686 TCP_SKB_CB(nskb)->end_seq += size;
4690 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4691 skb = tcp_collapse_one(sk, skb, list);
4694 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4701 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4702 * and tcp_collapse() them until all the queue is collapsed.
4704 static void tcp_collapse_ofo_queue(struct sock *sk)
4706 struct tcp_sock *tp = tcp_sk(sk);
4707 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4708 struct sk_buff *head;
4714 start = TCP_SKB_CB(skb)->seq;
4715 end = TCP_SKB_CB(skb)->end_seq;
4719 struct sk_buff *next = NULL;
4721 if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
4722 next = skb_queue_next(&tp->out_of_order_queue, skb);
4725 /* Segment is terminated when we see gap or when
4726 * we are at the end of all the queue. */
4728 after(TCP_SKB_CB(skb)->seq, end) ||
4729 before(TCP_SKB_CB(skb)->end_seq, start)) {
4730 tcp_collapse(sk, &tp->out_of_order_queue,
4731 head, skb, start, end);
4735 /* Start new segment */
4736 start = TCP_SKB_CB(skb)->seq;
4737 end = TCP_SKB_CB(skb)->end_seq;
4739 if (before(TCP_SKB_CB(skb)->seq, start))
4740 start = TCP_SKB_CB(skb)->seq;
4741 if (after(TCP_SKB_CB(skb)->end_seq, end))
4742 end = TCP_SKB_CB(skb)->end_seq;
4748 * Purge the out-of-order queue.
4749 * Return true if queue was pruned.
4751 static bool tcp_prune_ofo_queue(struct sock *sk)
4753 struct tcp_sock *tp = tcp_sk(sk);
4756 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4757 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
4758 __skb_queue_purge(&tp->out_of_order_queue);
4760 /* Reset SACK state. A conforming SACK implementation will
4761 * do the same at a timeout based retransmit. When a connection
4762 * is in a sad state like this, we care only about integrity
4763 * of the connection not performance.
4765 if (tp->rx_opt.sack_ok)
4766 tcp_sack_reset(&tp->rx_opt);
4773 /* Reduce allocated memory if we can, trying to get
4774 * the socket within its memory limits again.
4776 * Return less than zero if we should start dropping frames
4777 * until the socket owning process reads some of the data
4778 * to stabilize the situation.
4780 static int tcp_prune_queue(struct sock *sk)
4782 struct tcp_sock *tp = tcp_sk(sk);
4784 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4786 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
4788 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4789 tcp_clamp_window(sk);
4790 else if (tcp_under_memory_pressure(sk))
4791 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4793 tcp_collapse_ofo_queue(sk);
4794 if (!skb_queue_empty(&sk->sk_receive_queue))
4795 tcp_collapse(sk, &sk->sk_receive_queue,
4796 skb_peek(&sk->sk_receive_queue),
4798 tp->copied_seq, tp->rcv_nxt);
4801 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4804 /* Collapsing did not help, destructive actions follow.
4805 * This must not ever occur. */
4807 tcp_prune_ofo_queue(sk);
4809 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4812 /* If we are really being abused, tell the caller to silently
4813 * drop receive data on the floor. It will get retransmitted
4814 * and hopefully then we'll have sufficient space.
4816 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
4818 /* Massive buffer overcommit. */
4823 static bool tcp_should_expand_sndbuf(const struct sock *sk)
4825 const struct tcp_sock *tp = tcp_sk(sk);
4827 /* If the user specified a specific send buffer setting, do
4830 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4833 /* If we are under global TCP memory pressure, do not expand. */
4834 if (tcp_under_memory_pressure(sk))
4837 /* If we are under soft global TCP memory pressure, do not expand. */
4838 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
4841 /* If we filled the congestion window, do not expand. */
4842 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
4848 /* When incoming ACK allowed to free some skb from write_queue,
4849 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4850 * on the exit from tcp input handler.
4852 * PROBLEM: sndbuf expansion does not work well with largesend.
4854 static void tcp_new_space(struct sock *sk)
4856 struct tcp_sock *tp = tcp_sk(sk);
4858 if (tcp_should_expand_sndbuf(sk)) {
4859 tcp_sndbuf_expand(sk);
4860 tp->snd_cwnd_stamp = tcp_time_stamp;
4863 sk->sk_write_space(sk);
4866 static void tcp_check_space(struct sock *sk)
4868 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4869 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4870 /* pairs with tcp_poll() */
4871 smp_mb__after_atomic();
4872 if (sk->sk_socket &&
4873 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4878 static inline void tcp_data_snd_check(struct sock *sk)
4880 tcp_push_pending_frames(sk);
4881 tcp_check_space(sk);
4885 * Check if sending an ack is needed.
4887 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4889 struct tcp_sock *tp = tcp_sk(sk);
4891 /* More than one full frame received... */
4892 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
4893 /* ... and right edge of window advances far enough.
4894 * (tcp_recvmsg() will send ACK otherwise). Or...
4896 __tcp_select_window(sk) >= tp->rcv_wnd) ||
4897 /* We ACK each frame or... */
4898 tcp_in_quickack_mode(sk) ||
4899 /* We have out of order data. */
4900 (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
4901 /* Then ack it now */
4904 /* Else, send delayed ack. */
4905 tcp_send_delayed_ack(sk);
4909 static inline void tcp_ack_snd_check(struct sock *sk)
4911 if (!inet_csk_ack_scheduled(sk)) {
4912 /* We sent a data segment already. */
4915 __tcp_ack_snd_check(sk, 1);
4919 * This routine is only called when we have urgent data
4920 * signaled. Its the 'slow' part of tcp_urg. It could be
4921 * moved inline now as tcp_urg is only called from one
4922 * place. We handle URGent data wrong. We have to - as
4923 * BSD still doesn't use the correction from RFC961.
4924 * For 1003.1g we should support a new option TCP_STDURG to permit
4925 * either form (or just set the sysctl tcp_stdurg).
4928 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
4930 struct tcp_sock *tp = tcp_sk(sk);
4931 u32 ptr = ntohs(th->urg_ptr);
4933 if (ptr && !sysctl_tcp_stdurg)
4935 ptr += ntohl(th->seq);
4937 /* Ignore urgent data that we've already seen and read. */
4938 if (after(tp->copied_seq, ptr))
4941 /* Do not replay urg ptr.
4943 * NOTE: interesting situation not covered by specs.
4944 * Misbehaving sender may send urg ptr, pointing to segment,
4945 * which we already have in ofo queue. We are not able to fetch
4946 * such data and will stay in TCP_URG_NOTYET until will be eaten
4947 * by recvmsg(). Seems, we are not obliged to handle such wicked
4948 * situations. But it is worth to think about possibility of some
4949 * DoSes using some hypothetical application level deadlock.
4951 if (before(ptr, tp->rcv_nxt))
4954 /* Do we already have a newer (or duplicate) urgent pointer? */
4955 if (tp->urg_data && !after(ptr, tp->urg_seq))
4958 /* Tell the world about our new urgent pointer. */
4961 /* We may be adding urgent data when the last byte read was
4962 * urgent. To do this requires some care. We cannot just ignore
4963 * tp->copied_seq since we would read the last urgent byte again
4964 * as data, nor can we alter copied_seq until this data arrives
4965 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4967 * NOTE. Double Dutch. Rendering to plain English: author of comment
4968 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4969 * and expect that both A and B disappear from stream. This is _wrong_.
4970 * Though this happens in BSD with high probability, this is occasional.
4971 * Any application relying on this is buggy. Note also, that fix "works"
4972 * only in this artificial test. Insert some normal data between A and B and we will
4973 * decline of BSD again. Verdict: it is better to remove to trap
4976 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
4977 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
4978 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4980 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
4981 __skb_unlink(skb, &sk->sk_receive_queue);
4986 tp->urg_data = TCP_URG_NOTYET;
4989 /* Disable header prediction. */
4993 /* This is the 'fast' part of urgent handling. */
4994 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
4996 struct tcp_sock *tp = tcp_sk(sk);
4998 /* Check if we get a new urgent pointer - normally not. */
5000 tcp_check_urg(sk, th);
5002 /* Do we wait for any urgent data? - normally not... */
5003 if (tp->urg_data == TCP_URG_NOTYET) {
5004 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5007 /* Is the urgent pointer pointing into this packet? */
5008 if (ptr < skb->len) {
5010 if (skb_copy_bits(skb, ptr, &tmp, 1))
5012 tp->urg_data = TCP_URG_VALID | tmp;
5013 if (!sock_flag(sk, SOCK_DEAD))
5014 sk->sk_data_ready(sk);
5019 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
5021 struct tcp_sock *tp = tcp_sk(sk);
5022 int chunk = skb->len - hlen;
5026 if (skb_csum_unnecessary(skb))
5027 err = skb_copy_datagram_msg(skb, hlen, tp->ucopy.msg, chunk);
5029 err = skb_copy_and_csum_datagram_msg(skb, hlen, tp->ucopy.msg);
5032 tp->ucopy.len -= chunk;
5033 tp->copied_seq += chunk;
5034 tcp_rcv_space_adjust(sk);
5041 static __sum16 __tcp_checksum_complete_user(struct sock *sk,
5042 struct sk_buff *skb)
5046 if (sock_owned_by_user(sk)) {
5048 result = __tcp_checksum_complete(skb);
5051 result = __tcp_checksum_complete(skb);
5056 static inline bool tcp_checksum_complete_user(struct sock *sk,
5057 struct sk_buff *skb)
5059 return !skb_csum_unnecessary(skb) &&
5060 __tcp_checksum_complete_user(sk, skb);
5063 /* Does PAWS and seqno based validation of an incoming segment, flags will
5064 * play significant role here.
5066 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5067 const struct tcphdr *th, int syn_inerr)
5069 struct tcp_sock *tp = tcp_sk(sk);
5071 /* RFC1323: H1. Apply PAWS check first. */
5072 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
5073 tcp_paws_discard(sk, skb)) {
5075 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5076 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5077 LINUX_MIB_TCPACKSKIPPEDPAWS,
5078 &tp->last_oow_ack_time))
5079 tcp_send_dupack(sk, skb);
5082 /* Reset is accepted even if it did not pass PAWS. */
5085 /* Step 1: check sequence number */
5086 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5087 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5088 * (RST) segments are validated by checking their SEQ-fields."
5089 * And page 69: "If an incoming segment is not acceptable,
5090 * an acknowledgment should be sent in reply (unless the RST
5091 * bit is set, if so drop the segment and return)".
5096 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5097 LINUX_MIB_TCPACKSKIPPEDSEQ,
5098 &tp->last_oow_ack_time))
5099 tcp_send_dupack(sk, skb);
5104 /* Step 2: check RST bit */
5107 * If sequence number exactly matches RCV.NXT, then
5108 * RESET the connection
5110 * Send a challenge ACK
5112 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt)
5115 tcp_send_challenge_ack(sk, skb);
5119 /* step 3: check security and precedence [ignored] */
5121 /* step 4: Check for a SYN
5122 * RFC 5961 4.2 : Send a challenge ack
5127 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5128 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5129 tcp_send_challenge_ack(sk, skb);
5141 * TCP receive function for the ESTABLISHED state.
5143 * It is split into a fast path and a slow path. The fast path is
5145 * - A zero window was announced from us - zero window probing
5146 * is only handled properly in the slow path.
5147 * - Out of order segments arrived.
5148 * - Urgent data is expected.
5149 * - There is no buffer space left
5150 * - Unexpected TCP flags/window values/header lengths are received
5151 * (detected by checking the TCP header against pred_flags)
5152 * - Data is sent in both directions. Fast path only supports pure senders
5153 * or pure receivers (this means either the sequence number or the ack
5154 * value must stay constant)
5155 * - Unexpected TCP option.
5157 * When these conditions are not satisfied it drops into a standard
5158 * receive procedure patterned after RFC793 to handle all cases.
5159 * The first three cases are guaranteed by proper pred_flags setting,
5160 * the rest is checked inline. Fast processing is turned on in
5161 * tcp_data_queue when everything is OK.
5163 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5164 const struct tcphdr *th, unsigned int len)
5166 struct tcp_sock *tp = tcp_sk(sk);
5168 if (unlikely(!sk->sk_rx_dst))
5169 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5171 * Header prediction.
5172 * The code loosely follows the one in the famous
5173 * "30 instruction TCP receive" Van Jacobson mail.
5175 * Van's trick is to deposit buffers into socket queue
5176 * on a device interrupt, to call tcp_recv function
5177 * on the receive process context and checksum and copy
5178 * the buffer to user space. smart...
5180 * Our current scheme is not silly either but we take the
5181 * extra cost of the net_bh soft interrupt processing...
5182 * We do checksum and copy also but from device to kernel.
5185 tp->rx_opt.saw_tstamp = 0;
5187 /* pred_flags is 0xS?10 << 16 + snd_wnd
5188 * if header_prediction is to be made
5189 * 'S' will always be tp->tcp_header_len >> 2
5190 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5191 * turn it off (when there are holes in the receive
5192 * space for instance)
5193 * PSH flag is ignored.
5196 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5197 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5198 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5199 int tcp_header_len = tp->tcp_header_len;
5201 /* Timestamp header prediction: tcp_header_len
5202 * is automatically equal to th->doff*4 due to pred_flags
5206 /* Check timestamp */
5207 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5208 /* No? Slow path! */
5209 if (!tcp_parse_aligned_timestamp(tp, th))
5212 /* If PAWS failed, check it more carefully in slow path */
5213 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5216 /* DO NOT update ts_recent here, if checksum fails
5217 * and timestamp was corrupted part, it will result
5218 * in a hung connection since we will drop all
5219 * future packets due to the PAWS test.
5223 if (len <= tcp_header_len) {
5224 /* Bulk data transfer: sender */
5225 if (len == tcp_header_len) {
5226 /* Predicted packet is in window by definition.
5227 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5228 * Hence, check seq<=rcv_wup reduces to:
5230 if (tcp_header_len ==
5231 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5232 tp->rcv_nxt == tp->rcv_wup)
5233 tcp_store_ts_recent(tp);
5235 /* We know that such packets are checksummed
5238 tcp_ack(sk, skb, 0);
5240 tcp_data_snd_check(sk);
5242 } else { /* Header too small */
5243 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5248 bool fragstolen = false;
5250 if (tp->ucopy.task == current &&
5251 tp->copied_seq == tp->rcv_nxt &&
5252 len - tcp_header_len <= tp->ucopy.len &&
5253 sock_owned_by_user(sk)) {
5254 __set_current_state(TASK_RUNNING);
5256 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) {
5257 /* Predicted packet is in window by definition.
5258 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5259 * Hence, check seq<=rcv_wup reduces to:
5261 if (tcp_header_len ==
5262 (sizeof(struct tcphdr) +
5263 TCPOLEN_TSTAMP_ALIGNED) &&
5264 tp->rcv_nxt == tp->rcv_wup)
5265 tcp_store_ts_recent(tp);
5267 tcp_rcv_rtt_measure_ts(sk, skb);
5269 __skb_pull(skb, tcp_header_len);
5270 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
5271 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
5276 if (tcp_checksum_complete_user(sk, skb))
5279 if ((int)skb->truesize > sk->sk_forward_alloc)
5282 /* Predicted packet is in window by definition.
5283 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5284 * Hence, check seq<=rcv_wup reduces to:
5286 if (tcp_header_len ==
5287 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5288 tp->rcv_nxt == tp->rcv_wup)
5289 tcp_store_ts_recent(tp);
5291 tcp_rcv_rtt_measure_ts(sk, skb);
5293 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
5295 /* Bulk data transfer: receiver */
5296 eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5300 tcp_event_data_recv(sk, skb);
5302 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5303 /* Well, only one small jumplet in fast path... */
5304 tcp_ack(sk, skb, FLAG_DATA);
5305 tcp_data_snd_check(sk);
5306 if (!inet_csk_ack_scheduled(sk))
5310 __tcp_ack_snd_check(sk, 0);
5313 kfree_skb_partial(skb, fragstolen);
5314 sk->sk_data_ready(sk);
5320 if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
5323 if (!th->ack && !th->rst && !th->syn)
5327 * Standard slow path.
5330 if (!tcp_validate_incoming(sk, skb, th, 1))
5334 if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5337 tcp_rcv_rtt_measure_ts(sk, skb);
5339 /* Process urgent data. */
5340 tcp_urg(sk, skb, th);
5342 /* step 7: process the segment text */
5343 tcp_data_queue(sk, skb);
5345 tcp_data_snd_check(sk);
5346 tcp_ack_snd_check(sk);
5350 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_CSUMERRORS);
5351 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5356 EXPORT_SYMBOL(tcp_rcv_established);
5358 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5360 struct tcp_sock *tp = tcp_sk(sk);
5361 struct inet_connection_sock *icsk = inet_csk(sk);
5363 tcp_set_state(sk, TCP_ESTABLISHED);
5366 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5367 security_inet_conn_established(sk, skb);
5370 /* Make sure socket is routed, for correct metrics. */
5371 icsk->icsk_af_ops->rebuild_header(sk);
5373 tcp_init_metrics(sk);
5375 tcp_init_congestion_control(sk);
5377 /* Prevent spurious tcp_cwnd_restart() on first data
5380 tp->lsndtime = tcp_time_stamp;
5382 tcp_init_buffer_space(sk);
5384 if (sock_flag(sk, SOCK_KEEPOPEN))
5385 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5387 if (!tp->rx_opt.snd_wscale)
5388 __tcp_fast_path_on(tp, tp->snd_wnd);
5392 if (!sock_flag(sk, SOCK_DEAD)) {
5393 sk->sk_state_change(sk);
5394 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5398 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5399 struct tcp_fastopen_cookie *cookie)
5401 struct tcp_sock *tp = tcp_sk(sk);
5402 struct sk_buff *data = tp->syn_data ? tcp_write_queue_head(sk) : NULL;
5403 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
5404 bool syn_drop = false;
5406 if (mss == tp->rx_opt.user_mss) {
5407 struct tcp_options_received opt;
5409 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5410 tcp_clear_options(&opt);
5411 opt.user_mss = opt.mss_clamp = 0;
5412 tcp_parse_options(synack, &opt, 0, NULL);
5413 mss = opt.mss_clamp;
5416 if (!tp->syn_fastopen) {
5417 /* Ignore an unsolicited cookie */
5419 } else if (tp->total_retrans) {
5420 /* SYN timed out and the SYN-ACK neither has a cookie nor
5421 * acknowledges data. Presumably the remote received only
5422 * the retransmitted (regular) SYNs: either the original
5423 * SYN-data or the corresponding SYN-ACK was dropped.
5425 syn_drop = (cookie->len < 0 && data);
5426 } else if (cookie->len < 0 && !tp->syn_data) {
5427 /* We requested a cookie but didn't get it. If we did not use
5428 * the (old) exp opt format then try so next time (try_exp=1).
5429 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5431 try_exp = tp->syn_fastopen_exp ? 2 : 1;
5434 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
5436 if (data) { /* Retransmit unacked data in SYN */
5437 tcp_for_write_queue_from(data, sk) {
5438 if (data == tcp_send_head(sk) ||
5439 __tcp_retransmit_skb(sk, data))
5443 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVEFAIL);
5446 tp->syn_data_acked = tp->syn_data;
5447 if (tp->syn_data_acked)
5448 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
5452 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5453 const struct tcphdr *th, unsigned int len)
5455 struct inet_connection_sock *icsk = inet_csk(sk);
5456 struct tcp_sock *tp = tcp_sk(sk);
5457 struct tcp_fastopen_cookie foc = { .len = -1 };
5458 int saved_clamp = tp->rx_opt.mss_clamp;
5460 tcp_parse_options(skb, &tp->rx_opt, 0, &foc);
5461 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
5462 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
5466 * "If the state is SYN-SENT then
5467 * first check the ACK bit
5468 * If the ACK bit is set
5469 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5470 * a reset (unless the RST bit is set, if so drop
5471 * the segment and return)"
5473 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5474 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
5475 goto reset_and_undo;
5477 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5478 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5480 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
5481 goto reset_and_undo;
5484 /* Now ACK is acceptable.
5486 * "If the RST bit is set
5487 * If the ACK was acceptable then signal the user "error:
5488 * connection reset", drop the segment, enter CLOSED state,
5489 * delete TCB, and return."
5498 * "fifth, if neither of the SYN or RST bits is set then
5499 * drop the segment and return."
5505 goto discard_and_undo;
5508 * "If the SYN bit is on ...
5509 * are acceptable then ...
5510 * (our SYN has been ACKed), change the connection
5511 * state to ESTABLISHED..."
5514 tcp_ecn_rcv_synack(tp, th);
5516 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5517 tcp_ack(sk, skb, FLAG_SLOWPATH);
5519 /* Ok.. it's good. Set up sequence numbers and
5520 * move to established.
5522 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5523 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5525 /* RFC1323: The window in SYN & SYN/ACK segments is
5528 tp->snd_wnd = ntohs(th->window);
5530 if (!tp->rx_opt.wscale_ok) {
5531 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5532 tp->window_clamp = min(tp->window_clamp, 65535U);
5535 if (tp->rx_opt.saw_tstamp) {
5536 tp->rx_opt.tstamp_ok = 1;
5537 tp->tcp_header_len =
5538 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5539 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5540 tcp_store_ts_recent(tp);
5542 tp->tcp_header_len = sizeof(struct tcphdr);
5545 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5546 tcp_enable_fack(tp);
5549 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5550 tcp_initialize_rcv_mss(sk);
5552 /* Remember, tcp_poll() does not lock socket!
5553 * Change state from SYN-SENT only after copied_seq
5554 * is initialized. */
5555 tp->copied_seq = tp->rcv_nxt;
5559 tcp_finish_connect(sk, skb);
5561 if ((tp->syn_fastopen || tp->syn_data) &&
5562 tcp_rcv_fastopen_synack(sk, skb, &foc))
5565 if (sk->sk_write_pending ||
5566 icsk->icsk_accept_queue.rskq_defer_accept ||
5567 icsk->icsk_ack.pingpong) {
5568 /* Save one ACK. Data will be ready after
5569 * several ticks, if write_pending is set.
5571 * It may be deleted, but with this feature tcpdumps
5572 * look so _wonderfully_ clever, that I was not able
5573 * to stand against the temptation 8) --ANK
5575 inet_csk_schedule_ack(sk);
5576 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5577 tcp_enter_quickack_mode(sk);
5578 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5579 TCP_DELACK_MAX, TCP_RTO_MAX);
5590 /* No ACK in the segment */
5594 * "If the RST bit is set
5596 * Otherwise (no ACK) drop the segment and return."
5599 goto discard_and_undo;
5603 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5604 tcp_paws_reject(&tp->rx_opt, 0))
5605 goto discard_and_undo;
5608 /* We see SYN without ACK. It is attempt of
5609 * simultaneous connect with crossed SYNs.
5610 * Particularly, it can be connect to self.
5612 tcp_set_state(sk, TCP_SYN_RECV);
5614 if (tp->rx_opt.saw_tstamp) {
5615 tp->rx_opt.tstamp_ok = 1;
5616 tcp_store_ts_recent(tp);
5617 tp->tcp_header_len =
5618 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5620 tp->tcp_header_len = sizeof(struct tcphdr);
5623 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5624 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5626 /* RFC1323: The window in SYN & SYN/ACK segments is
5629 tp->snd_wnd = ntohs(th->window);
5630 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5631 tp->max_window = tp->snd_wnd;
5633 tcp_ecn_rcv_syn(tp, th);
5636 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5637 tcp_initialize_rcv_mss(sk);
5639 tcp_send_synack(sk);
5641 /* Note, we could accept data and URG from this segment.
5642 * There are no obstacles to make this (except that we must
5643 * either change tcp_recvmsg() to prevent it from returning data
5644 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5646 * However, if we ignore data in ACKless segments sometimes,
5647 * we have no reasons to accept it sometimes.
5648 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5649 * is not flawless. So, discard packet for sanity.
5650 * Uncomment this return to process the data.
5657 /* "fifth, if neither of the SYN or RST bits is set then
5658 * drop the segment and return."
5662 tcp_clear_options(&tp->rx_opt);
5663 tp->rx_opt.mss_clamp = saved_clamp;
5667 tcp_clear_options(&tp->rx_opt);
5668 tp->rx_opt.mss_clamp = saved_clamp;
5673 * This function implements the receiving procedure of RFC 793 for
5674 * all states except ESTABLISHED and TIME_WAIT.
5675 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5676 * address independent.
5679 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5680 const struct tcphdr *th, unsigned int len)
5682 struct tcp_sock *tp = tcp_sk(sk);
5683 struct inet_connection_sock *icsk = inet_csk(sk);
5684 struct request_sock *req;
5689 tp->rx_opt.saw_tstamp = 0;
5691 switch (sk->sk_state) {
5705 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5708 /* Now we have several options: In theory there is
5709 * nothing else in the frame. KA9Q has an option to
5710 * send data with the syn, BSD accepts data with the
5711 * syn up to the [to be] advertised window and
5712 * Solaris 2.1 gives you a protocol error. For now
5713 * we just ignore it, that fits the spec precisely
5714 * and avoids incompatibilities. It would be nice in
5715 * future to drop through and process the data.
5717 * Now that TTCP is starting to be used we ought to
5719 * But, this leaves one open to an easy denial of
5720 * service attack, and SYN cookies can't defend
5721 * against this problem. So, we drop the data
5722 * in the interest of security over speed unless
5723 * it's still in use.
5731 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5735 /* Do step6 onward by hand. */
5736 tcp_urg(sk, skb, th);
5738 tcp_data_snd_check(sk);
5742 req = tp->fastopen_rsk;
5744 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
5745 sk->sk_state != TCP_FIN_WAIT1);
5747 if (!tcp_check_req(sk, skb, req, true))
5751 if (!th->ack && !th->rst && !th->syn)
5754 if (!tcp_validate_incoming(sk, skb, th, 0))
5757 /* step 5: check the ACK field */
5758 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
5759 FLAG_UPDATE_TS_RECENT) > 0;
5761 switch (sk->sk_state) {
5766 /* Once we leave TCP_SYN_RECV, we no longer need req
5770 synack_stamp = tcp_rsk(req)->snt_synack;
5771 tp->total_retrans = req->num_retrans;
5772 reqsk_fastopen_remove(sk, req, false);
5774 synack_stamp = tp->lsndtime;
5775 /* Make sure socket is routed, for correct metrics. */
5776 icsk->icsk_af_ops->rebuild_header(sk);
5777 tcp_init_congestion_control(sk);
5780 tp->copied_seq = tp->rcv_nxt;
5781 tcp_init_buffer_space(sk);
5784 tcp_set_state(sk, TCP_ESTABLISHED);
5785 sk->sk_state_change(sk);
5787 /* Note, that this wakeup is only for marginal crossed SYN case.
5788 * Passively open sockets are not waked up, because
5789 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5792 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5794 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5795 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
5796 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5797 tcp_synack_rtt_meas(sk, synack_stamp);
5799 if (tp->rx_opt.tstamp_ok)
5800 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5803 /* Re-arm the timer because data may have been sent out.
5804 * This is similar to the regular data transmission case
5805 * when new data has just been ack'ed.
5807 * (TFO) - we could try to be more aggressive and
5808 * retransmitting any data sooner based on when they
5813 tcp_init_metrics(sk);
5815 tcp_update_pacing_rate(sk);
5817 /* Prevent spurious tcp_cwnd_restart() on first data packet */
5818 tp->lsndtime = tcp_time_stamp;
5820 tcp_initialize_rcv_mss(sk);
5821 tcp_fast_path_on(tp);
5824 case TCP_FIN_WAIT1: {
5825 struct dst_entry *dst;
5828 /* If we enter the TCP_FIN_WAIT1 state and we are a
5829 * Fast Open socket and this is the first acceptable
5830 * ACK we have received, this would have acknowledged
5831 * our SYNACK so stop the SYNACK timer.
5834 /* Return RST if ack_seq is invalid.
5835 * Note that RFC793 only says to generate a
5836 * DUPACK for it but for TCP Fast Open it seems
5837 * better to treat this case like TCP_SYN_RECV
5842 /* We no longer need the request sock. */
5843 reqsk_fastopen_remove(sk, req, false);
5846 if (tp->snd_una != tp->write_seq)
5849 tcp_set_state(sk, TCP_FIN_WAIT2);
5850 sk->sk_shutdown |= SEND_SHUTDOWN;
5852 dst = __sk_dst_get(sk);
5856 if (!sock_flag(sk, SOCK_DEAD)) {
5857 /* Wake up lingering close() */
5858 sk->sk_state_change(sk);
5862 if (tp->linger2 < 0 ||
5863 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5864 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5866 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5870 tmo = tcp_fin_time(sk);
5871 if (tmo > TCP_TIMEWAIT_LEN) {
5872 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5873 } else if (th->fin || sock_owned_by_user(sk)) {
5874 /* Bad case. We could lose such FIN otherwise.
5875 * It is not a big problem, but it looks confusing
5876 * and not so rare event. We still can lose it now,
5877 * if it spins in bh_lock_sock(), but it is really
5880 inet_csk_reset_keepalive_timer(sk, tmo);
5882 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5889 if (tp->snd_una == tp->write_seq) {
5890 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5896 if (tp->snd_una == tp->write_seq) {
5897 tcp_update_metrics(sk);
5904 /* step 6: check the URG bit */
5905 tcp_urg(sk, skb, th);
5907 /* step 7: process the segment text */
5908 switch (sk->sk_state) {
5909 case TCP_CLOSE_WAIT:
5912 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5916 /* RFC 793 says to queue data in these states,
5917 * RFC 1122 says we MUST send a reset.
5918 * BSD 4.4 also does reset.
5920 if (sk->sk_shutdown & RCV_SHUTDOWN) {
5921 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5922 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5923 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5929 case TCP_ESTABLISHED:
5930 tcp_data_queue(sk, skb);
5935 /* tcp_data could move socket to TIME-WAIT */
5936 if (sk->sk_state != TCP_CLOSE) {
5937 tcp_data_snd_check(sk);
5938 tcp_ack_snd_check(sk);
5947 EXPORT_SYMBOL(tcp_rcv_state_process);
5949 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
5951 struct inet_request_sock *ireq = inet_rsk(req);
5953 if (family == AF_INET)
5954 net_dbg_ratelimited("drop open request from %pI4/%u\n",
5955 &ireq->ir_rmt_addr, port);
5956 #if IS_ENABLED(CONFIG_IPV6)
5957 else if (family == AF_INET6)
5958 net_dbg_ratelimited("drop open request from %pI6/%u\n",
5959 &ireq->ir_v6_rmt_addr, port);
5963 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
5965 * If we receive a SYN packet with these bits set, it means a
5966 * network is playing bad games with TOS bits. In order to
5967 * avoid possible false congestion notifications, we disable
5968 * TCP ECN negotiation.
5970 * Exception: tcp_ca wants ECN. This is required for DCTCP
5971 * congestion control: Linux DCTCP asserts ECT on all packets,
5972 * including SYN, which is most optimal solution; however,
5973 * others, such as FreeBSD do not.
5975 static void tcp_ecn_create_request(struct request_sock *req,
5976 const struct sk_buff *skb,
5977 const struct sock *listen_sk,
5978 const struct dst_entry *dst)
5980 const struct tcphdr *th = tcp_hdr(skb);
5981 const struct net *net = sock_net(listen_sk);
5982 bool th_ecn = th->ece && th->cwr;
5988 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
5989 ecn_ok = net->ipv4.sysctl_tcp_ecn || dst_feature(dst, RTAX_FEATURE_ECN);
5991 if ((!ect && ecn_ok) || tcp_ca_needs_ecn(listen_sk))
5992 inet_rsk(req)->ecn_ok = 1;
5995 static void tcp_openreq_init(struct request_sock *req,
5996 const struct tcp_options_received *rx_opt,
5997 struct sk_buff *skb, const struct sock *sk)
5999 struct inet_request_sock *ireq = inet_rsk(req);
6001 req->rcv_wnd = 0; /* So that tcp_send_synack() knows! */
6003 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6004 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
6005 tcp_rsk(req)->snt_synack = tcp_time_stamp;
6006 tcp_rsk(req)->last_oow_ack_time = 0;
6007 req->mss = rx_opt->mss_clamp;
6008 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6009 ireq->tstamp_ok = rx_opt->tstamp_ok;
6010 ireq->sack_ok = rx_opt->sack_ok;
6011 ireq->snd_wscale = rx_opt->snd_wscale;
6012 ireq->wscale_ok = rx_opt->wscale_ok;
6015 ireq->ir_rmt_port = tcp_hdr(skb)->source;
6016 ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6017 ireq->ir_mark = inet_request_mark(sk, skb);
6020 struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
6021 struct sock *sk_listener)
6023 struct request_sock *req = reqsk_alloc(ops, sk_listener);
6026 struct inet_request_sock *ireq = inet_rsk(req);
6028 kmemcheck_annotate_bitfield(ireq, flags);
6030 atomic64_set(&ireq->ir_cookie, 0);
6031 ireq->ireq_state = TCP_NEW_SYN_RECV;
6032 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
6033 ireq->ireq_family = sk_listener->sk_family;
6038 EXPORT_SYMBOL(inet_reqsk_alloc);
6041 * Return true if a syncookie should be sent
6043 static bool tcp_syn_flood_action(struct sock *sk,
6044 const struct sk_buff *skb,
6047 const char *msg = "Dropping request";
6048 bool want_cookie = false;
6049 struct listen_sock *lopt;
6051 #ifdef CONFIG_SYN_COOKIES
6052 if (sysctl_tcp_syncookies) {
6053 msg = "Sending cookies";
6055 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
6058 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
6060 lopt = inet_csk(sk)->icsk_accept_queue.listen_opt;
6061 if (!lopt->synflood_warned && sysctl_tcp_syncookies != 2) {
6062 lopt->synflood_warned = 1;
6063 pr_info("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6064 proto, ntohs(tcp_hdr(skb)->dest), msg);
6069 static void tcp_reqsk_record_syn(const struct sock *sk,
6070 struct request_sock *req,
6071 const struct sk_buff *skb)
6073 if (tcp_sk(sk)->save_syn) {
6074 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
6077 copy = kmalloc(len + sizeof(u32), GFP_ATOMIC);
6080 memcpy(©[1], skb_network_header(skb), len);
6081 req->saved_syn = copy;
6086 int tcp_conn_request(struct request_sock_ops *rsk_ops,
6087 const struct tcp_request_sock_ops *af_ops,
6088 struct sock *sk, struct sk_buff *skb)
6090 struct tcp_options_received tmp_opt;
6091 struct request_sock *req;
6092 struct tcp_sock *tp = tcp_sk(sk);
6093 struct dst_entry *dst = NULL;
6094 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
6095 bool want_cookie = false, fastopen;
6097 struct tcp_fastopen_cookie foc = { .len = -1 };
6101 /* TW buckets are converted to open requests without
6102 * limitations, they conserve resources and peer is
6103 * evidently real one.
6105 if ((sysctl_tcp_syncookies == 2 ||
6106 inet_csk_reqsk_queue_is_full(sk)) && !isn) {
6107 want_cookie = tcp_syn_flood_action(sk, skb, rsk_ops->slab_name);
6113 /* Accept backlog is full. If we have already queued enough
6114 * of warm entries in syn queue, drop request. It is better than
6115 * clogging syn queue with openreqs with exponentially increasing
6118 if (sk_acceptq_is_full(sk) && inet_csk_reqsk_queue_young(sk) > 1) {
6119 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6123 req = inet_reqsk_alloc(rsk_ops, sk);
6127 tcp_rsk(req)->af_specific = af_ops;
6129 tcp_clear_options(&tmp_opt);
6130 tmp_opt.mss_clamp = af_ops->mss_clamp;
6131 tmp_opt.user_mss = tp->rx_opt.user_mss;
6132 tcp_parse_options(skb, &tmp_opt, 0, want_cookie ? NULL : &foc);
6134 if (want_cookie && !tmp_opt.saw_tstamp)
6135 tcp_clear_options(&tmp_opt);
6137 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
6138 tcp_openreq_init(req, &tmp_opt, skb, sk);
6140 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6141 inet_rsk(req)->ir_iif = sk->sk_bound_dev_if;
6143 af_ops->init_req(req, sk, skb);
6145 if (security_inet_conn_request(sk, skb, req))
6148 if (!want_cookie && !isn) {
6149 /* VJ's idea. We save last timestamp seen
6150 * from the destination in peer table, when entering
6151 * state TIME-WAIT, and check against it before
6152 * accepting new connection request.
6154 * If "isn" is not zero, this request hit alive
6155 * timewait bucket, so that all the necessary checks
6156 * are made in the function processing timewait state.
6158 if (tcp_death_row.sysctl_tw_recycle) {
6161 dst = af_ops->route_req(sk, &fl, req, &strict);
6163 if (dst && strict &&
6164 !tcp_peer_is_proven(req, dst, true,
6165 tmp_opt.saw_tstamp)) {
6166 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSPASSIVEREJECTED);
6167 goto drop_and_release;
6170 /* Kill the following clause, if you dislike this way. */
6171 else if (!sysctl_tcp_syncookies &&
6172 (sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
6173 (sysctl_max_syn_backlog >> 2)) &&
6174 !tcp_peer_is_proven(req, dst, false,
6175 tmp_opt.saw_tstamp)) {
6176 /* Without syncookies last quarter of
6177 * backlog is filled with destinations,
6178 * proven to be alive.
6179 * It means that we continue to communicate
6180 * to destinations, already remembered
6181 * to the moment of synflood.
6183 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
6185 goto drop_and_release;
6188 isn = af_ops->init_seq(skb);
6191 dst = af_ops->route_req(sk, &fl, req, NULL);
6196 tcp_ecn_create_request(req, skb, sk, dst);
6199 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
6200 req->cookie_ts = tmp_opt.tstamp_ok;
6201 if (!tmp_opt.tstamp_ok)
6202 inet_rsk(req)->ecn_ok = 0;
6205 tcp_rsk(req)->snt_isn = isn;
6206 tcp_openreq_init_rwin(req, sk, dst);
6207 fastopen = !want_cookie &&
6208 tcp_try_fastopen(sk, skb, req, &foc, dst);
6209 err = af_ops->send_synack(sk, dst, &fl, req,
6210 skb_get_queue_mapping(skb), &foc);
6212 if (err || want_cookie)
6215 tcp_rsk(req)->tfo_listener = false;
6216 af_ops->queue_hash_add(sk, req, TCP_TIMEOUT_INIT);
6218 tcp_reqsk_record_syn(sk, req, skb);
6227 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_LISTENDROPS);
6230 EXPORT_SYMBOL(tcp_conn_request);