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>
73 #include <net/inet_common.h>
74 #include <linux/ipsec.h>
75 #include <asm/unaligned.h>
76 #include <net/netdma.h>
78 int sysctl_tcp_timestamps __read_mostly = 1;
79 int sysctl_tcp_window_scaling __read_mostly = 1;
80 int sysctl_tcp_sack __read_mostly = 1;
81 int sysctl_tcp_fack __read_mostly = 1;
82 int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
83 EXPORT_SYMBOL(sysctl_tcp_reordering);
84 int sysctl_tcp_dsack __read_mostly = 1;
85 int sysctl_tcp_app_win __read_mostly = 31;
86 int sysctl_tcp_adv_win_scale __read_mostly = 1;
87 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
89 /* rfc5961 challenge ack rate limiting */
90 int sysctl_tcp_challenge_ack_limit = 100;
92 int sysctl_tcp_stdurg __read_mostly;
93 int sysctl_tcp_rfc1337 __read_mostly;
94 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
95 int sysctl_tcp_frto __read_mostly = 2;
97 int sysctl_tcp_thin_dupack __read_mostly;
99 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
100 int sysctl_tcp_early_retrans __read_mostly = 3;
102 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
103 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
104 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
105 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
106 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
107 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
108 #define FLAG_ECE 0x40 /* ECE in this ACK */
109 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
110 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
111 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
112 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
113 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
114 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
116 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
117 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
118 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
119 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
121 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
122 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
124 /* Adapt the MSS value used to make delayed ack decision to the
127 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
129 struct inet_connection_sock *icsk = inet_csk(sk);
130 const unsigned int lss = icsk->icsk_ack.last_seg_size;
133 icsk->icsk_ack.last_seg_size = 0;
135 /* skb->len may jitter because of SACKs, even if peer
136 * sends good full-sized frames.
138 len = skb_shinfo(skb)->gso_size ? : skb->len;
139 if (len >= icsk->icsk_ack.rcv_mss) {
140 icsk->icsk_ack.rcv_mss = len;
142 /* Otherwise, we make more careful check taking into account,
143 * that SACKs block is variable.
145 * "len" is invariant segment length, including TCP header.
147 len += skb->data - skb_transport_header(skb);
148 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
149 /* If PSH is not set, packet should be
150 * full sized, provided peer TCP is not badly broken.
151 * This observation (if it is correct 8)) allows
152 * to handle super-low mtu links fairly.
154 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
155 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
156 /* Subtract also invariant (if peer is RFC compliant),
157 * tcp header plus fixed timestamp option length.
158 * Resulting "len" is MSS free of SACK jitter.
160 len -= tcp_sk(sk)->tcp_header_len;
161 icsk->icsk_ack.last_seg_size = len;
163 icsk->icsk_ack.rcv_mss = len;
167 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
168 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
169 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
173 static void tcp_incr_quickack(struct sock *sk)
175 struct inet_connection_sock *icsk = inet_csk(sk);
176 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
180 if (quickacks > icsk->icsk_ack.quick)
181 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
184 static void tcp_enter_quickack_mode(struct sock *sk)
186 struct inet_connection_sock *icsk = inet_csk(sk);
187 tcp_incr_quickack(sk);
188 icsk->icsk_ack.pingpong = 0;
189 icsk->icsk_ack.ato = TCP_ATO_MIN;
192 /* Send ACKs quickly, if "quick" count is not exhausted
193 * and the session is not interactive.
196 static inline bool tcp_in_quickack_mode(const struct sock *sk)
198 const struct inet_connection_sock *icsk = inet_csk(sk);
200 return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
203 static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp)
205 if (tp->ecn_flags & TCP_ECN_OK)
206 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
209 static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
211 if (tcp_hdr(skb)->cwr)
212 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
215 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp)
217 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
220 static inline void TCP_ECN_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
222 if (!(tp->ecn_flags & TCP_ECN_OK))
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 (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
236 /* Better not delay acks, sender can have a very low cwnd */
237 tcp_enter_quickack_mode((struct sock *)tp);
238 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
242 tp->ecn_flags |= TCP_ECN_SEEN;
246 static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
248 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
249 tp->ecn_flags &= ~TCP_ECN_OK;
252 static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
254 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
255 tp->ecn_flags &= ~TCP_ECN_OK;
258 static bool TCP_ECN_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
260 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
265 /* Buffer size and advertised window tuning.
267 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
270 static void tcp_sndbuf_expand(struct sock *sk)
272 const struct tcp_sock *tp = tcp_sk(sk);
276 /* Worst case is non GSO/TSO : each frame consumes one skb
277 * and skb->head is kmalloced using power of two area of memory
279 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
281 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
283 per_mss = roundup_pow_of_two(per_mss) +
284 SKB_DATA_ALIGN(sizeof(struct sk_buff));
286 nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd);
287 nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
289 /* Fast Recovery (RFC 5681 3.2) :
290 * Cubic needs 1.7 factor, rounded to 2 to include
291 * extra cushion (application might react slowly to POLLOUT)
293 sndmem = 2 * nr_segs * per_mss;
295 if (sk->sk_sndbuf < sndmem)
296 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
299 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
301 * All tcp_full_space() is split to two parts: "network" buffer, allocated
302 * forward and advertised in receiver window (tp->rcv_wnd) and
303 * "application buffer", required to isolate scheduling/application
304 * latencies from network.
305 * window_clamp is maximal advertised window. It can be less than
306 * tcp_full_space(), in this case tcp_full_space() - window_clamp
307 * is reserved for "application" buffer. The less window_clamp is
308 * the smoother our behaviour from viewpoint of network, but the lower
309 * throughput and the higher sensitivity of the connection to losses. 8)
311 * rcv_ssthresh is more strict window_clamp used at "slow start"
312 * phase to predict further behaviour of this connection.
313 * It is used for two goals:
314 * - to enforce header prediction at sender, even when application
315 * requires some significant "application buffer". It is check #1.
316 * - to prevent pruning of receive queue because of misprediction
317 * of receiver window. Check #2.
319 * The scheme does not work when sender sends good segments opening
320 * window and then starts to feed us spaghetti. But it should work
321 * in common situations. Otherwise, we have to rely on queue collapsing.
324 /* Slow part of check#2. */
325 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
327 struct tcp_sock *tp = tcp_sk(sk);
329 int truesize = tcp_win_from_space(skb->truesize) >> 1;
330 int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
332 while (tp->rcv_ssthresh <= window) {
333 if (truesize <= skb->len)
334 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
342 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
344 struct tcp_sock *tp = tcp_sk(sk);
347 if (tp->rcv_ssthresh < tp->window_clamp &&
348 (int)tp->rcv_ssthresh < tcp_space(sk) &&
349 !sk_under_memory_pressure(sk)) {
352 /* Check #2. Increase window, if skb with such overhead
353 * will fit to rcvbuf in future.
355 if (tcp_win_from_space(skb->truesize) <= skb->len)
356 incr = 2 * tp->advmss;
358 incr = __tcp_grow_window(sk, skb);
361 incr = max_t(int, incr, 2 * skb->len);
362 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
364 inet_csk(sk)->icsk_ack.quick |= 1;
369 /* 3. Tuning rcvbuf, when connection enters established state. */
370 static void tcp_fixup_rcvbuf(struct sock *sk)
372 u32 mss = tcp_sk(sk)->advmss;
375 rcvmem = 2 * SKB_TRUESIZE(mss + MAX_TCP_HEADER) *
376 tcp_default_init_rwnd(mss);
378 /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
379 * Allow enough cushion so that sender is not limited by our window
381 if (sysctl_tcp_moderate_rcvbuf)
384 if (sk->sk_rcvbuf < rcvmem)
385 sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]);
388 /* 4. Try to fixup all. It is made immediately after connection enters
391 void tcp_init_buffer_space(struct sock *sk)
393 struct tcp_sock *tp = tcp_sk(sk);
396 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
397 tcp_fixup_rcvbuf(sk);
398 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
399 tcp_sndbuf_expand(sk);
401 tp->rcvq_space.space = tp->rcv_wnd;
402 tp->rcvq_space.time = tcp_time_stamp;
403 tp->rcvq_space.seq = tp->copied_seq;
405 maxwin = tcp_full_space(sk);
407 if (tp->window_clamp >= maxwin) {
408 tp->window_clamp = maxwin;
410 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
411 tp->window_clamp = max(maxwin -
412 (maxwin >> sysctl_tcp_app_win),
416 /* Force reservation of one segment. */
417 if (sysctl_tcp_app_win &&
418 tp->window_clamp > 2 * tp->advmss &&
419 tp->window_clamp + tp->advmss > maxwin)
420 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
422 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
423 tp->snd_cwnd_stamp = tcp_time_stamp;
426 /* 5. Recalculate window clamp after socket hit its memory bounds. */
427 static void tcp_clamp_window(struct sock *sk)
429 struct tcp_sock *tp = tcp_sk(sk);
430 struct inet_connection_sock *icsk = inet_csk(sk);
432 icsk->icsk_ack.quick = 0;
434 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
435 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
436 !sk_under_memory_pressure(sk) &&
437 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
438 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
441 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
442 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
445 /* Initialize RCV_MSS value.
446 * RCV_MSS is an our guess about MSS used by the peer.
447 * We haven't any direct information about the MSS.
448 * It's better to underestimate the RCV_MSS rather than overestimate.
449 * Overestimations make us ACKing less frequently than needed.
450 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
452 void tcp_initialize_rcv_mss(struct sock *sk)
454 const struct tcp_sock *tp = tcp_sk(sk);
455 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
457 hint = min(hint, tp->rcv_wnd / 2);
458 hint = min(hint, TCP_MSS_DEFAULT);
459 hint = max(hint, TCP_MIN_MSS);
461 inet_csk(sk)->icsk_ack.rcv_mss = hint;
463 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
465 /* Receiver "autotuning" code.
467 * The algorithm for RTT estimation w/o timestamps is based on
468 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
469 * <http://public.lanl.gov/radiant/pubs.html#DRS>
471 * More detail on this code can be found at
472 * <http://staff.psc.edu/jheffner/>,
473 * though this reference is out of date. A new paper
476 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
478 u32 new_sample = tp->rcv_rtt_est.rtt;
484 if (new_sample != 0) {
485 /* If we sample in larger samples in the non-timestamp
486 * case, we could grossly overestimate the RTT especially
487 * with chatty applications or bulk transfer apps which
488 * are stalled on filesystem I/O.
490 * Also, since we are only going for a minimum in the
491 * non-timestamp case, we do not smooth things out
492 * else with timestamps disabled convergence takes too
496 m -= (new_sample >> 3);
504 /* No previous measure. */
508 if (tp->rcv_rtt_est.rtt != new_sample)
509 tp->rcv_rtt_est.rtt = new_sample;
512 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
514 if (tp->rcv_rtt_est.time == 0)
516 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
518 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rcv_rtt_est.time, 1);
521 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
522 tp->rcv_rtt_est.time = tcp_time_stamp;
525 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
526 const struct sk_buff *skb)
528 struct tcp_sock *tp = tcp_sk(sk);
529 if (tp->rx_opt.rcv_tsecr &&
530 (TCP_SKB_CB(skb)->end_seq -
531 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
532 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
536 * This function should be called every time data is copied to user space.
537 * It calculates the appropriate TCP receive buffer space.
539 void tcp_rcv_space_adjust(struct sock *sk)
541 struct tcp_sock *tp = tcp_sk(sk);
545 time = tcp_time_stamp - tp->rcvq_space.time;
546 if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
549 /* Number of bytes copied to user in last RTT */
550 copied = tp->copied_seq - tp->rcvq_space.seq;
551 if (copied <= tp->rcvq_space.space)
555 * copied = bytes received in previous RTT, our base window
556 * To cope with packet losses, we need a 2x factor
557 * To cope with slow start, and sender growing its cwin by 100 %
558 * every RTT, we need a 4x factor, because the ACK we are sending
559 * now is for the next RTT, not the current one :
560 * <prev RTT . ><current RTT .. ><next RTT .... >
563 if (sysctl_tcp_moderate_rcvbuf &&
564 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
565 int rcvwin, rcvmem, rcvbuf;
567 /* minimal window to cope with packet losses, assuming
568 * steady state. Add some cushion because of small variations.
570 rcvwin = (copied << 1) + 16 * tp->advmss;
572 /* If rate increased by 25%,
573 * assume slow start, rcvwin = 3 * copied
574 * If rate increased by 50%,
575 * assume sender can use 2x growth, rcvwin = 4 * copied
578 tp->rcvq_space.space + (tp->rcvq_space.space >> 2)) {
580 tp->rcvq_space.space + (tp->rcvq_space.space >> 1))
583 rcvwin += (rcvwin >> 1);
586 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
587 while (tcp_win_from_space(rcvmem) < tp->advmss)
590 rcvbuf = min(rcvwin / tp->advmss * rcvmem, sysctl_tcp_rmem[2]);
591 if (rcvbuf > sk->sk_rcvbuf) {
592 sk->sk_rcvbuf = rcvbuf;
594 /* Make the window clamp follow along. */
595 tp->window_clamp = rcvwin;
598 tp->rcvq_space.space = copied;
601 tp->rcvq_space.seq = tp->copied_seq;
602 tp->rcvq_space.time = tcp_time_stamp;
605 /* There is something which you must keep in mind when you analyze the
606 * behavior of the tp->ato delayed ack timeout interval. When a
607 * connection starts up, we want to ack as quickly as possible. The
608 * problem is that "good" TCP's do slow start at the beginning of data
609 * transmission. The means that until we send the first few ACK's the
610 * sender will sit on his end and only queue most of his data, because
611 * he can only send snd_cwnd unacked packets at any given time. For
612 * each ACK we send, he increments snd_cwnd and transmits more of his
615 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
617 struct tcp_sock *tp = tcp_sk(sk);
618 struct inet_connection_sock *icsk = inet_csk(sk);
621 inet_csk_schedule_ack(sk);
623 tcp_measure_rcv_mss(sk, skb);
625 tcp_rcv_rtt_measure(tp);
627 now = tcp_time_stamp;
629 if (!icsk->icsk_ack.ato) {
630 /* The _first_ data packet received, initialize
631 * delayed ACK engine.
633 tcp_incr_quickack(sk);
634 icsk->icsk_ack.ato = TCP_ATO_MIN;
636 int m = now - icsk->icsk_ack.lrcvtime;
638 if (m <= TCP_ATO_MIN / 2) {
639 /* The fastest case is the first. */
640 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
641 } else if (m < icsk->icsk_ack.ato) {
642 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
643 if (icsk->icsk_ack.ato > icsk->icsk_rto)
644 icsk->icsk_ack.ato = icsk->icsk_rto;
645 } else if (m > icsk->icsk_rto) {
646 /* Too long gap. Apparently sender failed to
647 * restart window, so that we send ACKs quickly.
649 tcp_incr_quickack(sk);
653 icsk->icsk_ack.lrcvtime = now;
655 TCP_ECN_check_ce(tp, skb);
658 tcp_grow_window(sk, skb);
661 /* Called to compute a smoothed rtt estimate. The data fed to this
662 * routine either comes from timestamps, or from segments that were
663 * known _not_ to have been retransmitted [see Karn/Partridge
664 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
665 * piece by Van Jacobson.
666 * NOTE: the next three routines used to be one big routine.
667 * To save cycles in the RFC 1323 implementation it was better to break
668 * it up into three procedures. -- erics
670 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
672 struct tcp_sock *tp = tcp_sk(sk);
673 long m = mrtt; /* RTT */
675 /* The following amusing code comes from Jacobson's
676 * article in SIGCOMM '88. Note that rtt and mdev
677 * are scaled versions of rtt and mean deviation.
678 * This is designed to be as fast as possible
679 * m stands for "measurement".
681 * On a 1990 paper the rto value is changed to:
682 * RTO = rtt + 4 * mdev
684 * Funny. This algorithm seems to be very broken.
685 * These formulae increase RTO, when it should be decreased, increase
686 * too slowly, when it should be increased quickly, decrease too quickly
687 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
688 * does not matter how to _calculate_ it. Seems, it was trap
689 * that VJ failed to avoid. 8)
694 m -= (tp->srtt >> 3); /* m is now error in rtt est */
695 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */
697 m = -m; /* m is now abs(error) */
698 m -= (tp->mdev >> 2); /* similar update on mdev */
699 /* This is similar to one of Eifel findings.
700 * Eifel blocks mdev updates when rtt decreases.
701 * This solution is a bit different: we use finer gain
702 * for mdev in this case (alpha*beta).
703 * Like Eifel it also prevents growth of rto,
704 * but also it limits too fast rto decreases,
705 * happening in pure Eifel.
710 m -= (tp->mdev >> 2); /* similar update on mdev */
712 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */
713 if (tp->mdev > tp->mdev_max) {
714 tp->mdev_max = tp->mdev;
715 if (tp->mdev_max > tp->rttvar)
716 tp->rttvar = tp->mdev_max;
718 if (after(tp->snd_una, tp->rtt_seq)) {
719 if (tp->mdev_max < tp->rttvar)
720 tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2;
721 tp->rtt_seq = tp->snd_nxt;
722 tp->mdev_max = tcp_rto_min(sk);
725 /* no previous measure. */
726 tp->srtt = m << 3; /* take the measured time to be rtt */
727 tp->mdev = m << 1; /* make sure rto = 3*rtt */
728 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
729 tp->rtt_seq = tp->snd_nxt;
733 /* Set the sk_pacing_rate to allow proper sizing of TSO packets.
734 * Note: TCP stack does not yet implement pacing.
735 * FQ packet scheduler can be used to implement cheap but effective
736 * TCP pacing, to smooth the burst on large writes when packets
737 * in flight is significantly lower than cwnd (or rwin)
739 static void tcp_update_pacing_rate(struct sock *sk)
741 const struct tcp_sock *tp = tcp_sk(sk);
744 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
745 rate = (u64)tp->mss_cache * 2 * (HZ << 3);
747 rate *= max(tp->snd_cwnd, tp->packets_out);
749 /* Correction for small srtt : minimum srtt being 8 (1 jiffy << 3),
750 * be conservative and assume srtt = 1 (125 us instead of 1.25 ms)
751 * We probably need usec resolution in the future.
752 * Note: This also takes care of possible srtt=0 case,
753 * when tcp_rtt_estimator() was not yet called.
755 if (tp->srtt > 8 + 2)
756 do_div(rate, tp->srtt);
758 sk->sk_pacing_rate = min_t(u64, rate, sk->sk_max_pacing_rate);
761 /* Calculate rto without backoff. This is the second half of Van Jacobson's
762 * routine referred to above.
764 void tcp_set_rto(struct sock *sk)
766 const struct tcp_sock *tp = tcp_sk(sk);
767 /* Old crap is replaced with new one. 8)
770 * 1. If rtt variance happened to be less 50msec, it is hallucination.
771 * It cannot be less due to utterly erratic ACK generation made
772 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
773 * to do with delayed acks, because at cwnd>2 true delack timeout
774 * is invisible. Actually, Linux-2.4 also generates erratic
775 * ACKs in some circumstances.
777 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
779 /* 2. Fixups made earlier cannot be right.
780 * If we do not estimate RTO correctly without them,
781 * all the algo is pure shit and should be replaced
782 * with correct one. It is exactly, which we pretend to do.
785 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
786 * guarantees that rto is higher.
791 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
793 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
796 cwnd = TCP_INIT_CWND;
797 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
801 * Packet counting of FACK is based on in-order assumptions, therefore TCP
802 * disables it when reordering is detected
804 void tcp_disable_fack(struct tcp_sock *tp)
806 /* RFC3517 uses different metric in lost marker => reset on change */
808 tp->lost_skb_hint = NULL;
809 tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED;
812 /* Take a notice that peer is sending D-SACKs */
813 static void tcp_dsack_seen(struct tcp_sock *tp)
815 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
818 static void tcp_update_reordering(struct sock *sk, const int metric,
821 struct tcp_sock *tp = tcp_sk(sk);
822 if (metric > tp->reordering) {
825 tp->reordering = min(TCP_MAX_REORDERING, metric);
827 /* This exciting event is worth to be remembered. 8) */
829 mib_idx = LINUX_MIB_TCPTSREORDER;
830 else if (tcp_is_reno(tp))
831 mib_idx = LINUX_MIB_TCPRENOREORDER;
832 else if (tcp_is_fack(tp))
833 mib_idx = LINUX_MIB_TCPFACKREORDER;
835 mib_idx = LINUX_MIB_TCPSACKREORDER;
837 NET_INC_STATS_BH(sock_net(sk), mib_idx);
838 #if FASTRETRANS_DEBUG > 1
839 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
840 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
844 tp->undo_marker ? tp->undo_retrans : 0);
846 tcp_disable_fack(tp);
850 tcp_disable_early_retrans(tp);
853 /* This must be called before lost_out is incremented */
854 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
856 if ((tp->retransmit_skb_hint == NULL) ||
857 before(TCP_SKB_CB(skb)->seq,
858 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
859 tp->retransmit_skb_hint = skb;
862 after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
863 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
866 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
868 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
869 tcp_verify_retransmit_hint(tp, skb);
871 tp->lost_out += tcp_skb_pcount(skb);
872 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
876 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp,
879 tcp_verify_retransmit_hint(tp, skb);
881 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
882 tp->lost_out += tcp_skb_pcount(skb);
883 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
887 /* This procedure tags the retransmission queue when SACKs arrive.
889 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
890 * Packets in queue with these bits set are counted in variables
891 * sacked_out, retrans_out and lost_out, correspondingly.
893 * Valid combinations are:
894 * Tag InFlight Description
895 * 0 1 - orig segment is in flight.
896 * S 0 - nothing flies, orig reached receiver.
897 * L 0 - nothing flies, orig lost by net.
898 * R 2 - both orig and retransmit are in flight.
899 * L|R 1 - orig is lost, retransmit is in flight.
900 * S|R 1 - orig reached receiver, retrans is still in flight.
901 * (L|S|R is logically valid, it could occur when L|R is sacked,
902 * but it is equivalent to plain S and code short-curcuits it to S.
903 * L|S is logically invalid, it would mean -1 packet in flight 8))
905 * These 6 states form finite state machine, controlled by the following events:
906 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
907 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
908 * 3. Loss detection event of two flavors:
909 * A. Scoreboard estimator decided the packet is lost.
910 * A'. Reno "three dupacks" marks head of queue lost.
911 * A''. Its FACK modification, head until snd.fack is lost.
912 * B. SACK arrives sacking SND.NXT at the moment, when the
913 * segment was retransmitted.
914 * 4. D-SACK added new rule: D-SACK changes any tag to S.
916 * It is pleasant to note, that state diagram turns out to be commutative,
917 * so that we are allowed not to be bothered by order of our actions,
918 * when multiple events arrive simultaneously. (see the function below).
920 * Reordering detection.
921 * --------------------
922 * Reordering metric is maximal distance, which a packet can be displaced
923 * in packet stream. With SACKs we can estimate it:
925 * 1. SACK fills old hole and the corresponding segment was not
926 * ever retransmitted -> reordering. Alas, we cannot use it
927 * when segment was retransmitted.
928 * 2. The last flaw is solved with D-SACK. D-SACK arrives
929 * for retransmitted and already SACKed segment -> reordering..
930 * Both of these heuristics are not used in Loss state, when we cannot
931 * account for retransmits accurately.
933 * SACK block validation.
934 * ----------------------
936 * SACK block range validation checks that the received SACK block fits to
937 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
938 * Note that SND.UNA is not included to the range though being valid because
939 * it means that the receiver is rather inconsistent with itself reporting
940 * SACK reneging when it should advance SND.UNA. Such SACK block this is
941 * perfectly valid, however, in light of RFC2018 which explicitly states
942 * that "SACK block MUST reflect the newest segment. Even if the newest
943 * segment is going to be discarded ...", not that it looks very clever
944 * in case of head skb. Due to potentional receiver driven attacks, we
945 * choose to avoid immediate execution of a walk in write queue due to
946 * reneging and defer head skb's loss recovery to standard loss recovery
947 * procedure that will eventually trigger (nothing forbids us doing this).
949 * Implements also blockage to start_seq wrap-around. Problem lies in the
950 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
951 * there's no guarantee that it will be before snd_nxt (n). The problem
952 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
955 * <- outs wnd -> <- wrapzone ->
956 * u e n u_w e_w s n_w
958 * |<------------+------+----- TCP seqno space --------------+---------->|
959 * ...-- <2^31 ->| |<--------...
960 * ...---- >2^31 ------>| |<--------...
962 * Current code wouldn't be vulnerable but it's better still to discard such
963 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
964 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
965 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
966 * equal to the ideal case (infinite seqno space without wrap caused issues).
968 * With D-SACK the lower bound is extended to cover sequence space below
969 * SND.UNA down to undo_marker, which is the last point of interest. Yet
970 * again, D-SACK block must not to go across snd_una (for the same reason as
971 * for the normal SACK blocks, explained above). But there all simplicity
972 * ends, TCP might receive valid D-SACKs below that. As long as they reside
973 * fully below undo_marker they do not affect behavior in anyway and can
974 * therefore be safely ignored. In rare cases (which are more or less
975 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
976 * fragmentation and packet reordering past skb's retransmission. To consider
977 * them correctly, the acceptable range must be extended even more though
978 * the exact amount is rather hard to quantify. However, tp->max_window can
979 * be used as an exaggerated estimate.
981 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
982 u32 start_seq, u32 end_seq)
984 /* Too far in future, or reversed (interpretation is ambiguous) */
985 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
988 /* Nasty start_seq wrap-around check (see comments above) */
989 if (!before(start_seq, tp->snd_nxt))
992 /* In outstanding window? ...This is valid exit for D-SACKs too.
993 * start_seq == snd_una is non-sensical (see comments above)
995 if (after(start_seq, tp->snd_una))
998 if (!is_dsack || !tp->undo_marker)
1001 /* ...Then it's D-SACK, and must reside below snd_una completely */
1002 if (after(end_seq, tp->snd_una))
1005 if (!before(start_seq, tp->undo_marker))
1009 if (!after(end_seq, tp->undo_marker))
1012 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1013 * start_seq < undo_marker and end_seq >= undo_marker.
1015 return !before(start_seq, end_seq - tp->max_window);
1018 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1019 * Event "B". Later note: FACK people cheated me again 8), we have to account
1020 * for reordering! Ugly, but should help.
1022 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1023 * less than what is now known to be received by the other end (derived from
1024 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1025 * retransmitted skbs to avoid some costly processing per ACKs.
1027 static void tcp_mark_lost_retrans(struct sock *sk)
1029 const struct inet_connection_sock *icsk = inet_csk(sk);
1030 struct tcp_sock *tp = tcp_sk(sk);
1031 struct sk_buff *skb;
1033 u32 new_low_seq = tp->snd_nxt;
1034 u32 received_upto = tcp_highest_sack_seq(tp);
1036 if (!tcp_is_fack(tp) || !tp->retrans_out ||
1037 !after(received_upto, tp->lost_retrans_low) ||
1038 icsk->icsk_ca_state != TCP_CA_Recovery)
1041 tcp_for_write_queue(skb, sk) {
1042 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1044 if (skb == tcp_send_head(sk))
1046 if (cnt == tp->retrans_out)
1048 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1051 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1054 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1055 * constraint here (see above) but figuring out that at
1056 * least tp->reordering SACK blocks reside between ack_seq
1057 * and received_upto is not easy task to do cheaply with
1058 * the available datastructures.
1060 * Whether FACK should check here for tp->reordering segs
1061 * in-between one could argue for either way (it would be
1062 * rather simple to implement as we could count fack_count
1063 * during the walk and do tp->fackets_out - fack_count).
1065 if (after(received_upto, ack_seq)) {
1066 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1067 tp->retrans_out -= tcp_skb_pcount(skb);
1069 tcp_skb_mark_lost_uncond_verify(tp, skb);
1070 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
1072 if (before(ack_seq, new_low_seq))
1073 new_low_seq = ack_seq;
1074 cnt += tcp_skb_pcount(skb);
1078 if (tp->retrans_out)
1079 tp->lost_retrans_low = new_low_seq;
1082 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1083 struct tcp_sack_block_wire *sp, int num_sacks,
1086 struct tcp_sock *tp = tcp_sk(sk);
1087 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1088 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1089 bool dup_sack = false;
1091 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1094 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1095 } else if (num_sacks > 1) {
1096 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1097 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1099 if (!after(end_seq_0, end_seq_1) &&
1100 !before(start_seq_0, start_seq_1)) {
1103 NET_INC_STATS_BH(sock_net(sk),
1104 LINUX_MIB_TCPDSACKOFORECV);
1108 /* D-SACK for already forgotten data... Do dumb counting. */
1109 if (dup_sack && tp->undo_marker && tp->undo_retrans &&
1110 !after(end_seq_0, prior_snd_una) &&
1111 after(end_seq_0, tp->undo_marker))
1117 struct tcp_sacktag_state {
1121 s32 rtt; /* RTT measured by SACKing never-retransmitted data */
1124 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1125 * the incoming SACK may not exactly match but we can find smaller MSS
1126 * aligned portion of it that matches. Therefore we might need to fragment
1127 * which may fail and creates some hassle (caller must handle error case
1130 * FIXME: this could be merged to shift decision code
1132 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1133 u32 start_seq, u32 end_seq)
1137 unsigned int pkt_len;
1140 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1141 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1143 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1144 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1145 mss = tcp_skb_mss(skb);
1146 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1149 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1153 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1158 /* Round if necessary so that SACKs cover only full MSSes
1159 * and/or the remaining small portion (if present)
1161 if (pkt_len > mss) {
1162 unsigned int new_len = (pkt_len / mss) * mss;
1163 if (!in_sack && new_len < pkt_len) {
1165 if (new_len > skb->len)
1170 err = tcp_fragment(sk, skb, pkt_len, mss);
1178 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1179 static u8 tcp_sacktag_one(struct sock *sk,
1180 struct tcp_sacktag_state *state, u8 sacked,
1181 u32 start_seq, u32 end_seq,
1182 int dup_sack, int pcount, u32 xmit_time)
1184 struct tcp_sock *tp = tcp_sk(sk);
1185 int fack_count = state->fack_count;
1187 /* Account D-SACK for retransmitted packet. */
1188 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1189 if (tp->undo_marker && tp->undo_retrans &&
1190 after(end_seq, tp->undo_marker))
1192 if (sacked & TCPCB_SACKED_ACKED)
1193 state->reord = min(fack_count, state->reord);
1196 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1197 if (!after(end_seq, tp->snd_una))
1200 if (!(sacked & TCPCB_SACKED_ACKED)) {
1201 if (sacked & TCPCB_SACKED_RETRANS) {
1202 /* If the segment is not tagged as lost,
1203 * we do not clear RETRANS, believing
1204 * that retransmission is still in flight.
1206 if (sacked & TCPCB_LOST) {
1207 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1208 tp->lost_out -= pcount;
1209 tp->retrans_out -= pcount;
1212 if (!(sacked & TCPCB_RETRANS)) {
1213 /* New sack for not retransmitted frame,
1214 * which was in hole. It is reordering.
1216 if (before(start_seq,
1217 tcp_highest_sack_seq(tp)))
1218 state->reord = min(fack_count,
1220 if (!after(end_seq, tp->high_seq))
1221 state->flag |= FLAG_ORIG_SACK_ACKED;
1222 /* Pick the earliest sequence sacked for RTT */
1224 state->rtt = tcp_time_stamp - xmit_time;
1227 if (sacked & TCPCB_LOST) {
1228 sacked &= ~TCPCB_LOST;
1229 tp->lost_out -= pcount;
1233 sacked |= TCPCB_SACKED_ACKED;
1234 state->flag |= FLAG_DATA_SACKED;
1235 tp->sacked_out += pcount;
1237 fack_count += pcount;
1239 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1240 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
1241 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1242 tp->lost_cnt_hint += pcount;
1244 if (fack_count > tp->fackets_out)
1245 tp->fackets_out = fack_count;
1248 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1249 * frames and clear it. undo_retrans is decreased above, L|R frames
1250 * are accounted above as well.
1252 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1253 sacked &= ~TCPCB_SACKED_RETRANS;
1254 tp->retrans_out -= pcount;
1260 /* Shift newly-SACKed bytes from this skb to the immediately previous
1261 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1263 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1264 struct tcp_sacktag_state *state,
1265 unsigned int pcount, int shifted, int mss,
1268 struct tcp_sock *tp = tcp_sk(sk);
1269 struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1270 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1271 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1275 /* Adjust counters and hints for the newly sacked sequence
1276 * range but discard the return value since prev is already
1277 * marked. We must tag the range first because the seq
1278 * advancement below implicitly advances
1279 * tcp_highest_sack_seq() when skb is highest_sack.
1281 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1282 start_seq, end_seq, dup_sack, pcount,
1283 TCP_SKB_CB(skb)->when);
1285 if (skb == tp->lost_skb_hint)
1286 tp->lost_cnt_hint += pcount;
1288 TCP_SKB_CB(prev)->end_seq += shifted;
1289 TCP_SKB_CB(skb)->seq += shifted;
1291 skb_shinfo(prev)->gso_segs += pcount;
1292 BUG_ON(skb_shinfo(skb)->gso_segs < pcount);
1293 skb_shinfo(skb)->gso_segs -= pcount;
1295 /* When we're adding to gso_segs == 1, gso_size will be zero,
1296 * in theory this shouldn't be necessary but as long as DSACK
1297 * code can come after this skb later on it's better to keep
1298 * setting gso_size to something.
1300 if (!skb_shinfo(prev)->gso_size) {
1301 skb_shinfo(prev)->gso_size = mss;
1302 skb_shinfo(prev)->gso_type = sk->sk_gso_type;
1305 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1306 if (skb_shinfo(skb)->gso_segs <= 1) {
1307 skb_shinfo(skb)->gso_size = 0;
1308 skb_shinfo(skb)->gso_type = 0;
1311 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1312 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1315 BUG_ON(!tcp_skb_pcount(skb));
1316 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1320 /* Whole SKB was eaten :-) */
1322 if (skb == tp->retransmit_skb_hint)
1323 tp->retransmit_skb_hint = prev;
1324 if (skb == tp->lost_skb_hint) {
1325 tp->lost_skb_hint = prev;
1326 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1329 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(prev)->tcp_flags;
1330 if (skb == tcp_highest_sack(sk))
1331 tcp_advance_highest_sack(sk, skb);
1333 tcp_unlink_write_queue(skb, sk);
1334 sk_wmem_free_skb(sk, skb);
1336 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1341 /* I wish gso_size would have a bit more sane initialization than
1342 * something-or-zero which complicates things
1344 static int tcp_skb_seglen(const struct sk_buff *skb)
1346 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1349 /* Shifting pages past head area doesn't work */
1350 static int skb_can_shift(const struct sk_buff *skb)
1352 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1355 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1358 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1359 struct tcp_sacktag_state *state,
1360 u32 start_seq, u32 end_seq,
1363 struct tcp_sock *tp = tcp_sk(sk);
1364 struct sk_buff *prev;
1370 if (!sk_can_gso(sk))
1373 /* Normally R but no L won't result in plain S */
1375 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1377 if (!skb_can_shift(skb))
1379 /* This frame is about to be dropped (was ACKed). */
1380 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1383 /* Can only happen with delayed DSACK + discard craziness */
1384 if (unlikely(skb == tcp_write_queue_head(sk)))
1386 prev = tcp_write_queue_prev(sk, skb);
1388 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1391 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1392 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1396 pcount = tcp_skb_pcount(skb);
1397 mss = tcp_skb_seglen(skb);
1399 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1400 * drop this restriction as unnecessary
1402 if (mss != tcp_skb_seglen(prev))
1405 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1407 /* CHECKME: This is non-MSS split case only?, this will
1408 * cause skipped skbs due to advancing loop btw, original
1409 * has that feature too
1411 if (tcp_skb_pcount(skb) <= 1)
1414 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1416 /* TODO: head merge to next could be attempted here
1417 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1418 * though it might not be worth of the additional hassle
1420 * ...we can probably just fallback to what was done
1421 * previously. We could try merging non-SACKed ones
1422 * as well but it probably isn't going to buy off
1423 * because later SACKs might again split them, and
1424 * it would make skb timestamp tracking considerably
1430 len = end_seq - TCP_SKB_CB(skb)->seq;
1432 BUG_ON(len > skb->len);
1434 /* MSS boundaries should be honoured or else pcount will
1435 * severely break even though it makes things bit trickier.
1436 * Optimize common case to avoid most of the divides
1438 mss = tcp_skb_mss(skb);
1440 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1441 * drop this restriction as unnecessary
1443 if (mss != tcp_skb_seglen(prev))
1448 } else if (len < mss) {
1456 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1457 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1460 if (!skb_shift(prev, skb, len))
1462 if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1465 /* Hole filled allows collapsing with the next as well, this is very
1466 * useful when hole on every nth skb pattern happens
1468 if (prev == tcp_write_queue_tail(sk))
1470 skb = tcp_write_queue_next(sk, prev);
1472 if (!skb_can_shift(skb) ||
1473 (skb == tcp_send_head(sk)) ||
1474 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1475 (mss != tcp_skb_seglen(skb)))
1479 if (skb_shift(prev, skb, len)) {
1480 pcount += tcp_skb_pcount(skb);
1481 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1485 state->fack_count += pcount;
1492 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1496 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1497 struct tcp_sack_block *next_dup,
1498 struct tcp_sacktag_state *state,
1499 u32 start_seq, u32 end_seq,
1502 struct tcp_sock *tp = tcp_sk(sk);
1503 struct sk_buff *tmp;
1505 tcp_for_write_queue_from(skb, sk) {
1507 bool dup_sack = dup_sack_in;
1509 if (skb == tcp_send_head(sk))
1512 /* queue is in-order => we can short-circuit the walk early */
1513 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1516 if ((next_dup != NULL) &&
1517 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1518 in_sack = tcp_match_skb_to_sack(sk, skb,
1519 next_dup->start_seq,
1525 /* skb reference here is a bit tricky to get right, since
1526 * shifting can eat and free both this skb and the next,
1527 * so not even _safe variant of the loop is enough.
1530 tmp = tcp_shift_skb_data(sk, skb, state,
1531 start_seq, end_seq, dup_sack);
1540 in_sack = tcp_match_skb_to_sack(sk, skb,
1546 if (unlikely(in_sack < 0))
1550 TCP_SKB_CB(skb)->sacked =
1553 TCP_SKB_CB(skb)->sacked,
1554 TCP_SKB_CB(skb)->seq,
1555 TCP_SKB_CB(skb)->end_seq,
1557 tcp_skb_pcount(skb),
1558 TCP_SKB_CB(skb)->when);
1560 if (!before(TCP_SKB_CB(skb)->seq,
1561 tcp_highest_sack_seq(tp)))
1562 tcp_advance_highest_sack(sk, skb);
1565 state->fack_count += tcp_skb_pcount(skb);
1570 /* Avoid all extra work that is being done by sacktag while walking in
1573 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1574 struct tcp_sacktag_state *state,
1577 tcp_for_write_queue_from(skb, sk) {
1578 if (skb == tcp_send_head(sk))
1581 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1584 state->fack_count += tcp_skb_pcount(skb);
1589 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1591 struct tcp_sack_block *next_dup,
1592 struct tcp_sacktag_state *state,
1595 if (next_dup == NULL)
1598 if (before(next_dup->start_seq, skip_to_seq)) {
1599 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1600 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1601 next_dup->start_seq, next_dup->end_seq,
1608 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1610 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1614 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1615 u32 prior_snd_una, s32 *sack_rtt)
1617 struct tcp_sock *tp = tcp_sk(sk);
1618 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1619 TCP_SKB_CB(ack_skb)->sacked);
1620 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1621 struct tcp_sack_block sp[TCP_NUM_SACKS];
1622 struct tcp_sack_block *cache;
1623 struct tcp_sacktag_state state;
1624 struct sk_buff *skb;
1625 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1627 bool found_dup_sack = false;
1629 int first_sack_index;
1632 state.reord = tp->packets_out;
1635 if (!tp->sacked_out) {
1636 if (WARN_ON(tp->fackets_out))
1637 tp->fackets_out = 0;
1638 tcp_highest_sack_reset(sk);
1641 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1642 num_sacks, prior_snd_una);
1644 state.flag |= FLAG_DSACKING_ACK;
1646 /* Eliminate too old ACKs, but take into
1647 * account more or less fresh ones, they can
1648 * contain valid SACK info.
1650 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1653 if (!tp->packets_out)
1657 first_sack_index = 0;
1658 for (i = 0; i < num_sacks; i++) {
1659 bool dup_sack = !i && found_dup_sack;
1661 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1662 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1664 if (!tcp_is_sackblock_valid(tp, dup_sack,
1665 sp[used_sacks].start_seq,
1666 sp[used_sacks].end_seq)) {
1670 if (!tp->undo_marker)
1671 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1673 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1675 /* Don't count olds caused by ACK reordering */
1676 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1677 !after(sp[used_sacks].end_seq, tp->snd_una))
1679 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1682 NET_INC_STATS_BH(sock_net(sk), mib_idx);
1684 first_sack_index = -1;
1688 /* Ignore very old stuff early */
1689 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1695 /* order SACK blocks to allow in order walk of the retrans queue */
1696 for (i = used_sacks - 1; i > 0; i--) {
1697 for (j = 0; j < i; j++) {
1698 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1699 swap(sp[j], sp[j + 1]);
1701 /* Track where the first SACK block goes to */
1702 if (j == first_sack_index)
1703 first_sack_index = j + 1;
1708 skb = tcp_write_queue_head(sk);
1709 state.fack_count = 0;
1712 if (!tp->sacked_out) {
1713 /* It's already past, so skip checking against it */
1714 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1716 cache = tp->recv_sack_cache;
1717 /* Skip empty blocks in at head of the cache */
1718 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1723 while (i < used_sacks) {
1724 u32 start_seq = sp[i].start_seq;
1725 u32 end_seq = sp[i].end_seq;
1726 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1727 struct tcp_sack_block *next_dup = NULL;
1729 if (found_dup_sack && ((i + 1) == first_sack_index))
1730 next_dup = &sp[i + 1];
1732 /* Skip too early cached blocks */
1733 while (tcp_sack_cache_ok(tp, cache) &&
1734 !before(start_seq, cache->end_seq))
1737 /* Can skip some work by looking recv_sack_cache? */
1738 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1739 after(end_seq, cache->start_seq)) {
1742 if (before(start_seq, cache->start_seq)) {
1743 skb = tcp_sacktag_skip(skb, sk, &state,
1745 skb = tcp_sacktag_walk(skb, sk, next_dup,
1752 /* Rest of the block already fully processed? */
1753 if (!after(end_seq, cache->end_seq))
1756 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1760 /* ...tail remains todo... */
1761 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1762 /* ...but better entrypoint exists! */
1763 skb = tcp_highest_sack(sk);
1766 state.fack_count = tp->fackets_out;
1771 skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq);
1772 /* Check overlap against next cached too (past this one already) */
1777 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1778 skb = tcp_highest_sack(sk);
1781 state.fack_count = tp->fackets_out;
1783 skb = tcp_sacktag_skip(skb, sk, &state, start_seq);
1786 skb = tcp_sacktag_walk(skb, sk, next_dup, &state,
1787 start_seq, end_seq, dup_sack);
1793 /* Clear the head of the cache sack blocks so we can skip it next time */
1794 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1795 tp->recv_sack_cache[i].start_seq = 0;
1796 tp->recv_sack_cache[i].end_seq = 0;
1798 for (j = 0; j < used_sacks; j++)
1799 tp->recv_sack_cache[i++] = sp[j];
1801 tcp_mark_lost_retrans(sk);
1803 tcp_verify_left_out(tp);
1805 if ((state.reord < tp->fackets_out) &&
1806 ((inet_csk(sk)->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker))
1807 tcp_update_reordering(sk, tp->fackets_out - state.reord, 0);
1811 #if FASTRETRANS_DEBUG > 0
1812 WARN_ON((int)tp->sacked_out < 0);
1813 WARN_ON((int)tp->lost_out < 0);
1814 WARN_ON((int)tp->retrans_out < 0);
1815 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1817 *sack_rtt = state.rtt;
1821 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1822 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1824 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1828 holes = max(tp->lost_out, 1U);
1829 holes = min(holes, tp->packets_out);
1831 if ((tp->sacked_out + holes) > tp->packets_out) {
1832 tp->sacked_out = tp->packets_out - holes;
1838 /* If we receive more dupacks than we expected counting segments
1839 * in assumption of absent reordering, interpret this as reordering.
1840 * The only another reason could be bug in receiver TCP.
1842 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1844 struct tcp_sock *tp = tcp_sk(sk);
1845 if (tcp_limit_reno_sacked(tp))
1846 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1849 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1851 static void tcp_add_reno_sack(struct sock *sk)
1853 struct tcp_sock *tp = tcp_sk(sk);
1855 tcp_check_reno_reordering(sk, 0);
1856 tcp_verify_left_out(tp);
1859 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1861 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1863 struct tcp_sock *tp = tcp_sk(sk);
1866 /* One ACK acked hole. The rest eat duplicate ACKs. */
1867 if (acked - 1 >= tp->sacked_out)
1870 tp->sacked_out -= acked - 1;
1872 tcp_check_reno_reordering(sk, acked);
1873 tcp_verify_left_out(tp);
1876 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1881 static void tcp_clear_retrans_partial(struct tcp_sock *tp)
1883 tp->retrans_out = 0;
1886 tp->undo_marker = 0;
1887 tp->undo_retrans = 0;
1890 void tcp_clear_retrans(struct tcp_sock *tp)
1892 tcp_clear_retrans_partial(tp);
1894 tp->fackets_out = 0;
1898 /* Enter Loss state. If "how" is not zero, forget all SACK information
1899 * and reset tags completely, otherwise preserve SACKs. If receiver
1900 * dropped its ofo queue, we will know this due to reneging detection.
1902 void tcp_enter_loss(struct sock *sk, int how)
1904 const struct inet_connection_sock *icsk = inet_csk(sk);
1905 struct tcp_sock *tp = tcp_sk(sk);
1906 struct sk_buff *skb;
1907 bool new_recovery = false;
1909 /* Reduce ssthresh if it has not yet been made inside this window. */
1910 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1911 !after(tp->high_seq, tp->snd_una) ||
1912 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1913 new_recovery = true;
1914 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1915 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1916 tcp_ca_event(sk, CA_EVENT_LOSS);
1919 tp->snd_cwnd_cnt = 0;
1920 tp->snd_cwnd_stamp = tcp_time_stamp;
1922 tcp_clear_retrans_partial(tp);
1924 if (tcp_is_reno(tp))
1925 tcp_reset_reno_sack(tp);
1927 tp->undo_marker = tp->snd_una;
1930 tp->fackets_out = 0;
1932 tcp_clear_all_retrans_hints(tp);
1934 tcp_for_write_queue(skb, sk) {
1935 if (skb == tcp_send_head(sk))
1938 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1939 tp->undo_marker = 0;
1940 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1941 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
1942 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1943 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1944 tp->lost_out += tcp_skb_pcount(skb);
1945 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
1948 tcp_verify_left_out(tp);
1950 /* Timeout in disordered state after receiving substantial DUPACKs
1951 * suggests that the degree of reordering is over-estimated.
1953 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
1954 tp->sacked_out >= sysctl_tcp_reordering)
1955 tp->reordering = min_t(unsigned int, tp->reordering,
1956 sysctl_tcp_reordering);
1957 tcp_set_ca_state(sk, TCP_CA_Loss);
1958 tp->high_seq = tp->snd_nxt;
1959 TCP_ECN_queue_cwr(tp);
1961 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1962 * loss recovery is underway except recurring timeout(s) on
1963 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1965 tp->frto = sysctl_tcp_frto &&
1966 (new_recovery || icsk->icsk_retransmits) &&
1967 !inet_csk(sk)->icsk_mtup.probe_size;
1970 /* If ACK arrived pointing to a remembered SACK, it means that our
1971 * remembered SACKs do not reflect real state of receiver i.e.
1972 * receiver _host_ is heavily congested (or buggy).
1974 * Do processing similar to RTO timeout.
1976 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
1978 if (flag & FLAG_SACK_RENEGING) {
1979 struct inet_connection_sock *icsk = inet_csk(sk);
1980 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
1982 tcp_enter_loss(sk, 1);
1983 icsk->icsk_retransmits++;
1984 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
1985 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1986 icsk->icsk_rto, TCP_RTO_MAX);
1992 static inline int tcp_fackets_out(const struct tcp_sock *tp)
1994 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
1997 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
1998 * counter when SACK is enabled (without SACK, sacked_out is used for
2001 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2002 * segments up to the highest received SACK block so far and holes in
2005 * With reordering, holes may still be in flight, so RFC3517 recovery
2006 * uses pure sacked_out (total number of SACKed segments) even though
2007 * it violates the RFC that uses duplicate ACKs, often these are equal
2008 * but when e.g. out-of-window ACKs or packet duplication occurs,
2009 * they differ. Since neither occurs due to loss, TCP should really
2012 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2014 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2017 static bool tcp_pause_early_retransmit(struct sock *sk, int flag)
2019 struct tcp_sock *tp = tcp_sk(sk);
2020 unsigned long delay;
2022 /* Delay early retransmit and entering fast recovery for
2023 * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
2024 * available, or RTO is scheduled to fire first.
2026 if (sysctl_tcp_early_retrans < 2 || sysctl_tcp_early_retrans > 3 ||
2027 (flag & FLAG_ECE) || !tp->srtt)
2030 delay = max_t(unsigned long, (tp->srtt >> 5), msecs_to_jiffies(2));
2031 if (!time_after(inet_csk(sk)->icsk_timeout, (jiffies + delay)))
2034 inet_csk_reset_xmit_timer(sk, ICSK_TIME_EARLY_RETRANS, delay,
2039 /* Linux NewReno/SACK/FACK/ECN state machine.
2040 * --------------------------------------
2042 * "Open" Normal state, no dubious events, fast path.
2043 * "Disorder" In all the respects it is "Open",
2044 * but requires a bit more attention. It is entered when
2045 * we see some SACKs or dupacks. It is split of "Open"
2046 * mainly to move some processing from fast path to slow one.
2047 * "CWR" CWND was reduced due to some Congestion Notification event.
2048 * It can be ECN, ICMP source quench, local device congestion.
2049 * "Recovery" CWND was reduced, we are fast-retransmitting.
2050 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2052 * tcp_fastretrans_alert() is entered:
2053 * - each incoming ACK, if state is not "Open"
2054 * - when arrived ACK is unusual, namely:
2059 * Counting packets in flight is pretty simple.
2061 * in_flight = packets_out - left_out + retrans_out
2063 * packets_out is SND.NXT-SND.UNA counted in packets.
2065 * retrans_out is number of retransmitted segments.
2067 * left_out is number of segments left network, but not ACKed yet.
2069 * left_out = sacked_out + lost_out
2071 * sacked_out: Packets, which arrived to receiver out of order
2072 * and hence not ACKed. With SACKs this number is simply
2073 * amount of SACKed data. Even without SACKs
2074 * it is easy to give pretty reliable estimate of this number,
2075 * counting duplicate ACKs.
2077 * lost_out: Packets lost by network. TCP has no explicit
2078 * "loss notification" feedback from network (for now).
2079 * It means that this number can be only _guessed_.
2080 * Actually, it is the heuristics to predict lossage that
2081 * distinguishes different algorithms.
2083 * F.e. after RTO, when all the queue is considered as lost,
2084 * lost_out = packets_out and in_flight = retrans_out.
2086 * Essentially, we have now two algorithms counting
2089 * FACK: It is the simplest heuristics. As soon as we decided
2090 * that something is lost, we decide that _all_ not SACKed
2091 * packets until the most forward SACK are lost. I.e.
2092 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2093 * It is absolutely correct estimate, if network does not reorder
2094 * packets. And it loses any connection to reality when reordering
2095 * takes place. We use FACK by default until reordering
2096 * is suspected on the path to this destination.
2098 * NewReno: when Recovery is entered, we assume that one segment
2099 * is lost (classic Reno). While we are in Recovery and
2100 * a partial ACK arrives, we assume that one more packet
2101 * is lost (NewReno). This heuristics are the same in NewReno
2104 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2105 * deflation etc. CWND is real congestion window, never inflated, changes
2106 * only according to classic VJ rules.
2108 * Really tricky (and requiring careful tuning) part of algorithm
2109 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2110 * The first determines the moment _when_ we should reduce CWND and,
2111 * hence, slow down forward transmission. In fact, it determines the moment
2112 * when we decide that hole is caused by loss, rather than by a reorder.
2114 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2115 * holes, caused by lost packets.
2117 * And the most logically complicated part of algorithm is undo
2118 * heuristics. We detect false retransmits due to both too early
2119 * fast retransmit (reordering) and underestimated RTO, analyzing
2120 * timestamps and D-SACKs. When we detect that some segments were
2121 * retransmitted by mistake and CWND reduction was wrong, we undo
2122 * window reduction and abort recovery phase. This logic is hidden
2123 * inside several functions named tcp_try_undo_<something>.
2126 /* This function decides, when we should leave Disordered state
2127 * and enter Recovery phase, reducing congestion window.
2129 * Main question: may we further continue forward transmission
2130 * with the same cwnd?
2132 static bool tcp_time_to_recover(struct sock *sk, int flag)
2134 struct tcp_sock *tp = tcp_sk(sk);
2137 /* Trick#1: The loss is proven. */
2141 /* Not-A-Trick#2 : Classic rule... */
2142 if (tcp_dupack_heuristics(tp) > tp->reordering)
2145 /* Trick#4: It is still not OK... But will it be useful to delay
2148 packets_out = tp->packets_out;
2149 if (packets_out <= tp->reordering &&
2150 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2151 !tcp_may_send_now(sk)) {
2152 /* We have nothing to send. This connection is limited
2153 * either by receiver window or by application.
2158 /* If a thin stream is detected, retransmit after first
2159 * received dupack. Employ only if SACK is supported in order
2160 * to avoid possible corner-case series of spurious retransmissions
2161 * Use only if there are no unsent data.
2163 if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2164 tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2165 tcp_is_sack(tp) && !tcp_send_head(sk))
2168 /* Trick#6: TCP early retransmit, per RFC5827. To avoid spurious
2169 * retransmissions due to small network reorderings, we implement
2170 * Mitigation A.3 in the RFC and delay the retransmission for a short
2171 * interval if appropriate.
2173 if (tp->do_early_retrans && !tp->retrans_out && tp->sacked_out &&
2174 (tp->packets_out >= (tp->sacked_out + 1) && tp->packets_out < 4) &&
2175 !tcp_may_send_now(sk))
2176 return !tcp_pause_early_retransmit(sk, flag);
2181 /* Detect loss in event "A" above by marking head of queue up as lost.
2182 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2183 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2184 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2185 * the maximum SACKed segments to pass before reaching this limit.
2187 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2189 struct tcp_sock *tp = tcp_sk(sk);
2190 struct sk_buff *skb;
2194 /* Use SACK to deduce losses of new sequences sent during recovery */
2195 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq;
2197 WARN_ON(packets > tp->packets_out);
2198 if (tp->lost_skb_hint) {
2199 skb = tp->lost_skb_hint;
2200 cnt = tp->lost_cnt_hint;
2201 /* Head already handled? */
2202 if (mark_head && skb != tcp_write_queue_head(sk))
2205 skb = tcp_write_queue_head(sk);
2209 tcp_for_write_queue_from(skb, sk) {
2210 if (skb == tcp_send_head(sk))
2212 /* TODO: do this better */
2213 /* this is not the most efficient way to do this... */
2214 tp->lost_skb_hint = skb;
2215 tp->lost_cnt_hint = cnt;
2217 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2221 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2222 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2223 cnt += tcp_skb_pcount(skb);
2225 if (cnt > packets) {
2226 if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2227 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2228 (oldcnt >= packets))
2231 mss = skb_shinfo(skb)->gso_size;
2232 err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss);
2238 tcp_skb_mark_lost(tp, skb);
2243 tcp_verify_left_out(tp);
2246 /* Account newly detected lost packet(s) */
2248 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2250 struct tcp_sock *tp = tcp_sk(sk);
2252 if (tcp_is_reno(tp)) {
2253 tcp_mark_head_lost(sk, 1, 1);
2254 } else if (tcp_is_fack(tp)) {
2255 int lost = tp->fackets_out - tp->reordering;
2258 tcp_mark_head_lost(sk, lost, 0);
2260 int sacked_upto = tp->sacked_out - tp->reordering;
2261 if (sacked_upto >= 0)
2262 tcp_mark_head_lost(sk, sacked_upto, 0);
2263 else if (fast_rexmit)
2264 tcp_mark_head_lost(sk, 1, 1);
2268 /* CWND moderation, preventing bursts due to too big ACKs
2269 * in dubious situations.
2271 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2273 tp->snd_cwnd = min(tp->snd_cwnd,
2274 tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2275 tp->snd_cwnd_stamp = tcp_time_stamp;
2278 /* Nothing was retransmitted or returned timestamp is less
2279 * than timestamp of the first retransmission.
2281 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2283 return !tp->retrans_stamp ||
2284 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2285 before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
2288 /* Undo procedures. */
2290 #if FASTRETRANS_DEBUG > 1
2291 static void DBGUNDO(struct sock *sk, const char *msg)
2293 struct tcp_sock *tp = tcp_sk(sk);
2294 struct inet_sock *inet = inet_sk(sk);
2296 if (sk->sk_family == AF_INET) {
2297 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2299 &inet->inet_daddr, ntohs(inet->inet_dport),
2300 tp->snd_cwnd, tcp_left_out(tp),
2301 tp->snd_ssthresh, tp->prior_ssthresh,
2304 #if IS_ENABLED(CONFIG_IPV6)
2305 else if (sk->sk_family == AF_INET6) {
2306 struct ipv6_pinfo *np = inet6_sk(sk);
2307 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2309 &np->daddr, ntohs(inet->inet_dport),
2310 tp->snd_cwnd, tcp_left_out(tp),
2311 tp->snd_ssthresh, tp->prior_ssthresh,
2317 #define DBGUNDO(x...) do { } while (0)
2320 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2322 struct tcp_sock *tp = tcp_sk(sk);
2325 struct sk_buff *skb;
2327 tcp_for_write_queue(skb, sk) {
2328 if (skb == tcp_send_head(sk))
2330 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2333 tcp_clear_all_retrans_hints(tp);
2336 if (tp->prior_ssthresh) {
2337 const struct inet_connection_sock *icsk = inet_csk(sk);
2339 if (icsk->icsk_ca_ops->undo_cwnd)
2340 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2342 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2344 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2345 tp->snd_ssthresh = tp->prior_ssthresh;
2346 TCP_ECN_withdraw_cwr(tp);
2349 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2351 tp->snd_cwnd_stamp = tcp_time_stamp;
2352 tp->undo_marker = 0;
2355 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2357 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2360 /* People celebrate: "We love our President!" */
2361 static bool tcp_try_undo_recovery(struct sock *sk)
2363 struct tcp_sock *tp = tcp_sk(sk);
2365 if (tcp_may_undo(tp)) {
2368 /* Happy end! We did not retransmit anything
2369 * or our original transmission succeeded.
2371 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2372 tcp_undo_cwnd_reduction(sk, false);
2373 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2374 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2376 mib_idx = LINUX_MIB_TCPFULLUNDO;
2378 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2380 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2381 /* Hold old state until something *above* high_seq
2382 * is ACKed. For Reno it is MUST to prevent false
2383 * fast retransmits (RFC2582). SACK TCP is safe. */
2384 tcp_moderate_cwnd(tp);
2387 tcp_set_ca_state(sk, TCP_CA_Open);
2391 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2392 static bool tcp_try_undo_dsack(struct sock *sk)
2394 struct tcp_sock *tp = tcp_sk(sk);
2396 if (tp->undo_marker && !tp->undo_retrans) {
2397 DBGUNDO(sk, "D-SACK");
2398 tcp_undo_cwnd_reduction(sk, false);
2399 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2405 /* We can clear retrans_stamp when there are no retransmissions in the
2406 * window. It would seem that it is trivially available for us in
2407 * tp->retrans_out, however, that kind of assumptions doesn't consider
2408 * what will happen if errors occur when sending retransmission for the
2409 * second time. ...It could the that such segment has only
2410 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2411 * the head skb is enough except for some reneging corner cases that
2412 * are not worth the effort.
2414 * Main reason for all this complexity is the fact that connection dying
2415 * time now depends on the validity of the retrans_stamp, in particular,
2416 * that successive retransmissions of a segment must not advance
2417 * retrans_stamp under any conditions.
2419 static bool tcp_any_retrans_done(const struct sock *sk)
2421 const struct tcp_sock *tp = tcp_sk(sk);
2422 struct sk_buff *skb;
2424 if (tp->retrans_out)
2427 skb = tcp_write_queue_head(sk);
2428 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2434 /* Undo during loss recovery after partial ACK or using F-RTO. */
2435 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2437 struct tcp_sock *tp = tcp_sk(sk);
2439 if (frto_undo || tcp_may_undo(tp)) {
2440 tcp_undo_cwnd_reduction(sk, true);
2442 DBGUNDO(sk, "partial loss");
2443 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2445 NET_INC_STATS_BH(sock_net(sk),
2446 LINUX_MIB_TCPSPURIOUSRTOS);
2447 inet_csk(sk)->icsk_retransmits = 0;
2448 if (frto_undo || tcp_is_sack(tp))
2449 tcp_set_ca_state(sk, TCP_CA_Open);
2455 /* The cwnd reduction in CWR and Recovery use the PRR algorithm
2456 * https://datatracker.ietf.org/doc/draft-ietf-tcpm-proportional-rate-reduction/
2457 * It computes the number of packets to send (sndcnt) based on packets newly
2459 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2460 * cwnd reductions across a full RTT.
2461 * 2) If packets in flight is lower than ssthresh (such as due to excess
2462 * losses and/or application stalls), do not perform any further cwnd
2463 * reductions, but instead slow start up to ssthresh.
2465 static void tcp_init_cwnd_reduction(struct sock *sk, const bool set_ssthresh)
2467 struct tcp_sock *tp = tcp_sk(sk);
2469 tp->high_seq = tp->snd_nxt;
2470 tp->tlp_high_seq = 0;
2471 tp->snd_cwnd_cnt = 0;
2472 tp->prior_cwnd = tp->snd_cwnd;
2473 tp->prr_delivered = 0;
2476 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2477 TCP_ECN_queue_cwr(tp);
2480 static void tcp_cwnd_reduction(struct sock *sk, const int prior_unsacked,
2483 struct tcp_sock *tp = tcp_sk(sk);
2485 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2486 int newly_acked_sacked = prior_unsacked -
2487 (tp->packets_out - tp->sacked_out);
2489 tp->prr_delivered += newly_acked_sacked;
2490 if (tcp_packets_in_flight(tp) > tp->snd_ssthresh) {
2491 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2493 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2495 sndcnt = min_t(int, delta,
2496 max_t(int, tp->prr_delivered - tp->prr_out,
2497 newly_acked_sacked) + 1);
2500 sndcnt = max(sndcnt, (fast_rexmit ? 1 : 0));
2501 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2504 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2506 struct tcp_sock *tp = tcp_sk(sk);
2508 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2509 if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR ||
2510 (tp->undo_marker && tp->snd_ssthresh < TCP_INFINITE_SSTHRESH)) {
2511 tp->snd_cwnd = tp->snd_ssthresh;
2512 tp->snd_cwnd_stamp = tcp_time_stamp;
2514 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2517 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2518 void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
2520 struct tcp_sock *tp = tcp_sk(sk);
2522 tp->prior_ssthresh = 0;
2523 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2524 tp->undo_marker = 0;
2525 tcp_init_cwnd_reduction(sk, set_ssthresh);
2526 tcp_set_ca_state(sk, TCP_CA_CWR);
2530 static void tcp_try_keep_open(struct sock *sk)
2532 struct tcp_sock *tp = tcp_sk(sk);
2533 int state = TCP_CA_Open;
2535 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2536 state = TCP_CA_Disorder;
2538 if (inet_csk(sk)->icsk_ca_state != state) {
2539 tcp_set_ca_state(sk, state);
2540 tp->high_seq = tp->snd_nxt;
2544 static void tcp_try_to_open(struct sock *sk, int flag, const int prior_unsacked)
2546 struct tcp_sock *tp = tcp_sk(sk);
2548 tcp_verify_left_out(tp);
2550 if (!tcp_any_retrans_done(sk))
2551 tp->retrans_stamp = 0;
2553 if (flag & FLAG_ECE)
2554 tcp_enter_cwr(sk, 1);
2556 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2557 tcp_try_keep_open(sk);
2559 tcp_cwnd_reduction(sk, prior_unsacked, 0);
2563 static void tcp_mtup_probe_failed(struct sock *sk)
2565 struct inet_connection_sock *icsk = inet_csk(sk);
2567 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2568 icsk->icsk_mtup.probe_size = 0;
2571 static void tcp_mtup_probe_success(struct sock *sk)
2573 struct tcp_sock *tp = tcp_sk(sk);
2574 struct inet_connection_sock *icsk = inet_csk(sk);
2576 /* FIXME: breaks with very large cwnd */
2577 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2578 tp->snd_cwnd = tp->snd_cwnd *
2579 tcp_mss_to_mtu(sk, tp->mss_cache) /
2580 icsk->icsk_mtup.probe_size;
2581 tp->snd_cwnd_cnt = 0;
2582 tp->snd_cwnd_stamp = tcp_time_stamp;
2583 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2585 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2586 icsk->icsk_mtup.probe_size = 0;
2587 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2590 /* Do a simple retransmit without using the backoff mechanisms in
2591 * tcp_timer. This is used for path mtu discovery.
2592 * The socket is already locked here.
2594 void tcp_simple_retransmit(struct sock *sk)
2596 const struct inet_connection_sock *icsk = inet_csk(sk);
2597 struct tcp_sock *tp = tcp_sk(sk);
2598 struct sk_buff *skb;
2599 unsigned int mss = tcp_current_mss(sk);
2600 u32 prior_lost = tp->lost_out;
2602 tcp_for_write_queue(skb, sk) {
2603 if (skb == tcp_send_head(sk))
2605 if (tcp_skb_seglen(skb) > mss &&
2606 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2607 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2608 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2609 tp->retrans_out -= tcp_skb_pcount(skb);
2611 tcp_skb_mark_lost_uncond_verify(tp, skb);
2615 tcp_clear_retrans_hints_partial(tp);
2617 if (prior_lost == tp->lost_out)
2620 if (tcp_is_reno(tp))
2621 tcp_limit_reno_sacked(tp);
2623 tcp_verify_left_out(tp);
2625 /* Don't muck with the congestion window here.
2626 * Reason is that we do not increase amount of _data_
2627 * in network, but units changed and effective
2628 * cwnd/ssthresh really reduced now.
2630 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2631 tp->high_seq = tp->snd_nxt;
2632 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2633 tp->prior_ssthresh = 0;
2634 tp->undo_marker = 0;
2635 tcp_set_ca_state(sk, TCP_CA_Loss);
2637 tcp_xmit_retransmit_queue(sk);
2639 EXPORT_SYMBOL(tcp_simple_retransmit);
2641 static void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2643 struct tcp_sock *tp = tcp_sk(sk);
2646 if (tcp_is_reno(tp))
2647 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2649 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2651 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2653 tp->prior_ssthresh = 0;
2654 tp->undo_marker = tp->snd_una;
2655 tp->undo_retrans = tp->retrans_out;
2657 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2659 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2660 tcp_init_cwnd_reduction(sk, true);
2662 tcp_set_ca_state(sk, TCP_CA_Recovery);
2665 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2666 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2668 static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack)
2670 struct inet_connection_sock *icsk = inet_csk(sk);
2671 struct tcp_sock *tp = tcp_sk(sk);
2672 bool recovered = !before(tp->snd_una, tp->high_seq);
2674 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2675 if (flag & FLAG_ORIG_SACK_ACKED) {
2676 /* Step 3.b. A timeout is spurious if not all data are
2677 * lost, i.e., never-retransmitted data are (s)acked.
2679 tcp_try_undo_loss(sk, true);
2682 if (after(tp->snd_nxt, tp->high_seq) &&
2683 (flag & FLAG_DATA_SACKED || is_dupack)) {
2684 tp->frto = 0; /* Loss was real: 2nd part of step 3.a */
2685 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2686 tp->high_seq = tp->snd_nxt;
2687 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
2689 if (after(tp->snd_nxt, tp->high_seq))
2690 return; /* Step 2.b */
2696 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2697 icsk->icsk_retransmits = 0;
2698 tcp_try_undo_recovery(sk);
2701 if (flag & FLAG_DATA_ACKED)
2702 icsk->icsk_retransmits = 0;
2703 if (tcp_is_reno(tp)) {
2704 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2705 * delivered. Lower inflight to clock out (re)tranmissions.
2707 if (after(tp->snd_nxt, tp->high_seq) && is_dupack)
2708 tcp_add_reno_sack(sk);
2709 else if (flag & FLAG_SND_UNA_ADVANCED)
2710 tcp_reset_reno_sack(tp);
2712 if (tcp_try_undo_loss(sk, false))
2714 tcp_xmit_retransmit_queue(sk);
2717 /* Undo during fast recovery after partial ACK. */
2718 static bool tcp_try_undo_partial(struct sock *sk, const int acked,
2719 const int prior_unsacked)
2721 struct tcp_sock *tp = tcp_sk(sk);
2723 if (tp->undo_marker && tcp_packet_delayed(tp)) {
2724 /* Plain luck! Hole if filled with delayed
2725 * packet, rather than with a retransmit.
2727 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2729 /* We are getting evidence that the reordering degree is higher
2730 * than we realized. If there are no retransmits out then we
2731 * can undo. Otherwise we clock out new packets but do not
2732 * mark more packets lost or retransmit more.
2734 if (tp->retrans_out) {
2735 tcp_cwnd_reduction(sk, prior_unsacked, 0);
2739 if (!tcp_any_retrans_done(sk))
2740 tp->retrans_stamp = 0;
2742 DBGUNDO(sk, "partial recovery");
2743 tcp_undo_cwnd_reduction(sk, true);
2744 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2745 tcp_try_keep_open(sk);
2751 /* Process an event, which can update packets-in-flight not trivially.
2752 * Main goal of this function is to calculate new estimate for left_out,
2753 * taking into account both packets sitting in receiver's buffer and
2754 * packets lost by network.
2756 * Besides that it does CWND reduction, when packet loss is detected
2757 * and changes state of machine.
2759 * It does _not_ decide what to send, it is made in function
2760 * tcp_xmit_retransmit_queue().
2762 static void tcp_fastretrans_alert(struct sock *sk, const int acked,
2763 const int prior_unsacked,
2764 bool is_dupack, int flag)
2766 struct inet_connection_sock *icsk = inet_csk(sk);
2767 struct tcp_sock *tp = tcp_sk(sk);
2768 bool do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2769 (tcp_fackets_out(tp) > tp->reordering));
2770 int fast_rexmit = 0;
2772 if (WARN_ON(!tp->packets_out && tp->sacked_out))
2774 if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2775 tp->fackets_out = 0;
2777 /* Now state machine starts.
2778 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2779 if (flag & FLAG_ECE)
2780 tp->prior_ssthresh = 0;
2782 /* B. In all the states check for reneging SACKs. */
2783 if (tcp_check_sack_reneging(sk, flag))
2786 /* C. Check consistency of the current state. */
2787 tcp_verify_left_out(tp);
2789 /* D. Check state exit conditions. State can be terminated
2790 * when high_seq is ACKed. */
2791 if (icsk->icsk_ca_state == TCP_CA_Open) {
2792 WARN_ON(tp->retrans_out != 0);
2793 tp->retrans_stamp = 0;
2794 } else if (!before(tp->snd_una, tp->high_seq)) {
2795 switch (icsk->icsk_ca_state) {
2797 /* CWR is to be held something *above* high_seq
2798 * is ACKed for CWR bit to reach receiver. */
2799 if (tp->snd_una != tp->high_seq) {
2800 tcp_end_cwnd_reduction(sk);
2801 tcp_set_ca_state(sk, TCP_CA_Open);
2805 case TCP_CA_Recovery:
2806 if (tcp_is_reno(tp))
2807 tcp_reset_reno_sack(tp);
2808 if (tcp_try_undo_recovery(sk))
2810 tcp_end_cwnd_reduction(sk);
2815 /* E. Process state. */
2816 switch (icsk->icsk_ca_state) {
2817 case TCP_CA_Recovery:
2818 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2819 if (tcp_is_reno(tp) && is_dupack)
2820 tcp_add_reno_sack(sk);
2822 if (tcp_try_undo_partial(sk, acked, prior_unsacked))
2824 /* Partial ACK arrived. Force fast retransmit. */
2825 do_lost = tcp_is_reno(tp) ||
2826 tcp_fackets_out(tp) > tp->reordering;
2828 if (tcp_try_undo_dsack(sk)) {
2829 tcp_try_keep_open(sk);
2834 tcp_process_loss(sk, flag, is_dupack);
2835 if (icsk->icsk_ca_state != TCP_CA_Open)
2837 /* Fall through to processing in Open state. */
2839 if (tcp_is_reno(tp)) {
2840 if (flag & FLAG_SND_UNA_ADVANCED)
2841 tcp_reset_reno_sack(tp);
2843 tcp_add_reno_sack(sk);
2846 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2847 tcp_try_undo_dsack(sk);
2849 if (!tcp_time_to_recover(sk, flag)) {
2850 tcp_try_to_open(sk, flag, prior_unsacked);
2854 /* MTU probe failure: don't reduce cwnd */
2855 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2856 icsk->icsk_mtup.probe_size &&
2857 tp->snd_una == tp->mtu_probe.probe_seq_start) {
2858 tcp_mtup_probe_failed(sk);
2859 /* Restores the reduction we did in tcp_mtup_probe() */
2861 tcp_simple_retransmit(sk);
2865 /* Otherwise enter Recovery state */
2866 tcp_enter_recovery(sk, (flag & FLAG_ECE));
2871 tcp_update_scoreboard(sk, fast_rexmit);
2872 tcp_cwnd_reduction(sk, prior_unsacked, fast_rexmit);
2873 tcp_xmit_retransmit_queue(sk);
2876 static inline bool tcp_ack_update_rtt(struct sock *sk, const int flag,
2877 s32 seq_rtt, s32 sack_rtt)
2879 const struct tcp_sock *tp = tcp_sk(sk);
2881 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2882 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2883 * Karn's algorithm forbids taking RTT if some retransmitted data
2884 * is acked (RFC6298).
2886 if (flag & FLAG_RETRANS_DATA_ACKED)
2892 /* RTTM Rule: A TSecr value received in a segment is used to
2893 * update the averaged RTT measurement only if the segment
2894 * acknowledges some new data, i.e., only if it advances the
2895 * left edge of the send window.
2896 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2898 if (seq_rtt < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
2899 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr;
2904 tcp_rtt_estimator(sk, seq_rtt);
2907 /* RFC6298: only reset backoff on valid RTT measurement. */
2908 inet_csk(sk)->icsk_backoff = 0;
2912 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2913 static void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
2915 struct tcp_sock *tp = tcp_sk(sk);
2918 if (tp->lsndtime && !tp->total_retrans)
2919 seq_rtt = tcp_time_stamp - tp->lsndtime;
2920 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, seq_rtt, -1);
2923 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
2925 const struct inet_connection_sock *icsk = inet_csk(sk);
2926 icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight);
2927 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
2930 /* Restart timer after forward progress on connection.
2931 * RFC2988 recommends to restart timer to now+rto.
2933 void tcp_rearm_rto(struct sock *sk)
2935 const struct inet_connection_sock *icsk = inet_csk(sk);
2936 struct tcp_sock *tp = tcp_sk(sk);
2938 /* If the retrans timer is currently being used by Fast Open
2939 * for SYN-ACK retrans purpose, stay put.
2941 if (tp->fastopen_rsk)
2944 if (!tp->packets_out) {
2945 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2947 u32 rto = inet_csk(sk)->icsk_rto;
2948 /* Offset the time elapsed after installing regular RTO */
2949 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
2950 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
2951 struct sk_buff *skb = tcp_write_queue_head(sk);
2952 const u32 rto_time_stamp = TCP_SKB_CB(skb)->when + rto;
2953 s32 delta = (s32)(rto_time_stamp - tcp_time_stamp);
2954 /* delta may not be positive if the socket is locked
2955 * when the retrans timer fires and is rescheduled.
2960 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
2965 /* This function is called when the delayed ER timer fires. TCP enters
2966 * fast recovery and performs fast-retransmit.
2968 void tcp_resume_early_retransmit(struct sock *sk)
2970 struct tcp_sock *tp = tcp_sk(sk);
2974 /* Stop if ER is disabled after the delayed ER timer is scheduled */
2975 if (!tp->do_early_retrans)
2978 tcp_enter_recovery(sk, false);
2979 tcp_update_scoreboard(sk, 1);
2980 tcp_xmit_retransmit_queue(sk);
2983 /* If we get here, the whole TSO packet has not been acked. */
2984 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
2986 struct tcp_sock *tp = tcp_sk(sk);
2989 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
2991 packets_acked = tcp_skb_pcount(skb);
2992 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2994 packets_acked -= tcp_skb_pcount(skb);
2996 if (packets_acked) {
2997 BUG_ON(tcp_skb_pcount(skb) == 0);
2998 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3001 return packets_acked;
3004 /* Remove acknowledged frames from the retransmission queue. If our packet
3005 * is before the ack sequence we can discard it as it's confirmed to have
3006 * arrived at the other end.
3008 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3009 u32 prior_snd_una, s32 sack_rtt)
3011 struct tcp_sock *tp = tcp_sk(sk);
3012 const struct inet_connection_sock *icsk = inet_csk(sk);
3013 struct sk_buff *skb;
3014 u32 now = tcp_time_stamp;
3015 bool fully_acked = true;
3018 u32 reord = tp->packets_out;
3019 u32 prior_sacked = tp->sacked_out;
3021 s32 ca_seq_rtt = -1;
3022 ktime_t last_ackt = net_invalid_timestamp();
3024 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3025 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3027 u8 sacked = scb->sacked;
3029 /* Determine how many packets and what bytes were acked, tso and else */
3030 if (after(scb->end_seq, tp->snd_una)) {
3031 if (tcp_skb_pcount(skb) == 1 ||
3032 !after(tp->snd_una, scb->seq))
3035 acked_pcount = tcp_tso_acked(sk, skb);
3039 fully_acked = false;
3041 acked_pcount = tcp_skb_pcount(skb);
3044 if (sacked & TCPCB_RETRANS) {
3045 if (sacked & TCPCB_SACKED_RETRANS)
3046 tp->retrans_out -= acked_pcount;
3047 flag |= FLAG_RETRANS_DATA_ACKED;
3049 ca_seq_rtt = now - scb->when;
3050 last_ackt = skb->tstamp;
3052 seq_rtt = ca_seq_rtt;
3054 if (!(sacked & TCPCB_SACKED_ACKED))
3055 reord = min(pkts_acked, reord);
3056 if (!after(scb->end_seq, tp->high_seq))
3057 flag |= FLAG_ORIG_SACK_ACKED;
3060 if (sacked & TCPCB_SACKED_ACKED)
3061 tp->sacked_out -= acked_pcount;
3062 if (sacked & TCPCB_LOST)
3063 tp->lost_out -= acked_pcount;
3065 tp->packets_out -= acked_pcount;
3066 pkts_acked += acked_pcount;
3068 /* Initial outgoing SYN's get put onto the write_queue
3069 * just like anything else we transmit. It is not
3070 * true data, and if we misinform our callers that
3071 * this ACK acks real data, we will erroneously exit
3072 * connection startup slow start one packet too
3073 * quickly. This is severely frowned upon behavior.
3075 if (!(scb->tcp_flags & TCPHDR_SYN)) {
3076 flag |= FLAG_DATA_ACKED;
3078 flag |= FLAG_SYN_ACKED;
3079 tp->retrans_stamp = 0;
3085 tcp_unlink_write_queue(skb, sk);
3086 sk_wmem_free_skb(sk, skb);
3087 if (skb == tp->retransmit_skb_hint)
3088 tp->retransmit_skb_hint = NULL;
3089 if (skb == tp->lost_skb_hint)
3090 tp->lost_skb_hint = NULL;
3093 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3094 tp->snd_up = tp->snd_una;
3096 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3097 flag |= FLAG_SACK_RENEGING;
3099 if (tcp_ack_update_rtt(sk, flag, seq_rtt, sack_rtt) ||
3100 (flag & FLAG_ACKED))
3103 if (flag & FLAG_ACKED) {
3104 const struct tcp_congestion_ops *ca_ops
3105 = inet_csk(sk)->icsk_ca_ops;
3107 if (unlikely(icsk->icsk_mtup.probe_size &&
3108 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3109 tcp_mtup_probe_success(sk);
3112 if (tcp_is_reno(tp)) {
3113 tcp_remove_reno_sacks(sk, pkts_acked);
3117 /* Non-retransmitted hole got filled? That's reordering */
3118 if (reord < prior_fackets)
3119 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3121 delta = tcp_is_fack(tp) ? pkts_acked :
3122 prior_sacked - tp->sacked_out;
3123 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3126 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3128 if (ca_ops->pkts_acked) {
3131 /* Is the ACK triggering packet unambiguous? */
3132 if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
3133 /* High resolution needed and available? */
3134 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
3135 !ktime_equal(last_ackt,
3136 net_invalid_timestamp()))
3137 rtt_us = ktime_us_delta(ktime_get_real(),
3139 else if (ca_seq_rtt >= 0)
3140 rtt_us = jiffies_to_usecs(ca_seq_rtt);
3143 ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
3147 #if FASTRETRANS_DEBUG > 0
3148 WARN_ON((int)tp->sacked_out < 0);
3149 WARN_ON((int)tp->lost_out < 0);
3150 WARN_ON((int)tp->retrans_out < 0);
3151 if (!tp->packets_out && tcp_is_sack(tp)) {
3152 icsk = inet_csk(sk);
3154 pr_debug("Leak l=%u %d\n",
3155 tp->lost_out, icsk->icsk_ca_state);
3158 if (tp->sacked_out) {
3159 pr_debug("Leak s=%u %d\n",
3160 tp->sacked_out, icsk->icsk_ca_state);
3163 if (tp->retrans_out) {
3164 pr_debug("Leak r=%u %d\n",
3165 tp->retrans_out, icsk->icsk_ca_state);
3166 tp->retrans_out = 0;
3173 static void tcp_ack_probe(struct sock *sk)
3175 const struct tcp_sock *tp = tcp_sk(sk);
3176 struct inet_connection_sock *icsk = inet_csk(sk);
3178 /* Was it a usable window open? */
3180 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3181 icsk->icsk_backoff = 0;
3182 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3183 /* Socket must be waked up by subsequent tcp_data_snd_check().
3184 * This function is not for random using!
3187 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3188 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3193 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3195 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3196 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3199 /* Decide wheather to run the increase function of congestion control. */
3200 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3202 if (tcp_in_cwnd_reduction(sk))
3205 /* If reordering is high then always grow cwnd whenever data is
3206 * delivered regardless of its ordering. Otherwise stay conservative
3207 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3208 * new SACK or ECE mark may first advance cwnd here and later reduce
3209 * cwnd in tcp_fastretrans_alert() based on more states.
3211 if (tcp_sk(sk)->reordering > sysctl_tcp_reordering)
3212 return flag & FLAG_FORWARD_PROGRESS;
3214 return flag & FLAG_DATA_ACKED;
3217 /* Check that window update is acceptable.
3218 * The function assumes that snd_una<=ack<=snd_next.
3220 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3221 const u32 ack, const u32 ack_seq,
3224 return after(ack, tp->snd_una) ||
3225 after(ack_seq, tp->snd_wl1) ||
3226 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3229 /* Update our send window.
3231 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3232 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3234 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3237 struct tcp_sock *tp = tcp_sk(sk);
3239 u32 nwin = ntohs(tcp_hdr(skb)->window);
3241 if (likely(!tcp_hdr(skb)->syn))
3242 nwin <<= tp->rx_opt.snd_wscale;
3244 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3245 flag |= FLAG_WIN_UPDATE;
3246 tcp_update_wl(tp, ack_seq);
3248 if (tp->snd_wnd != nwin) {
3251 /* Note, it is the only place, where
3252 * fast path is recovered for sending TCP.
3255 tcp_fast_path_check(sk);
3257 if (nwin > tp->max_window) {
3258 tp->max_window = nwin;
3259 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3269 /* RFC 5961 7 [ACK Throttling] */
3270 static void tcp_send_challenge_ack(struct sock *sk)
3272 /* unprotected vars, we dont care of overwrites */
3273 static u32 challenge_timestamp;
3274 static unsigned int challenge_count;
3275 u32 now = jiffies / HZ;
3277 if (now != challenge_timestamp) {
3278 challenge_timestamp = now;
3279 challenge_count = 0;
3281 if (++challenge_count <= sysctl_tcp_challenge_ack_limit) {
3282 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPCHALLENGEACK);
3287 static void tcp_store_ts_recent(struct tcp_sock *tp)
3289 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3290 tp->rx_opt.ts_recent_stamp = get_seconds();
3293 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3295 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3296 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3297 * extra check below makes sure this can only happen
3298 * for pure ACK frames. -DaveM
3300 * Not only, also it occurs for expired timestamps.
3303 if (tcp_paws_check(&tp->rx_opt, 0))
3304 tcp_store_ts_recent(tp);
3308 /* This routine deals with acks during a TLP episode.
3309 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3311 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3313 struct tcp_sock *tp = tcp_sk(sk);
3314 bool is_tlp_dupack = (ack == tp->tlp_high_seq) &&
3315 !(flag & (FLAG_SND_UNA_ADVANCED |
3316 FLAG_NOT_DUP | FLAG_DATA_SACKED));
3318 /* Mark the end of TLP episode on receiving TLP dupack or when
3319 * ack is after tlp_high_seq.
3321 if (is_tlp_dupack) {
3322 tp->tlp_high_seq = 0;
3326 if (after(ack, tp->tlp_high_seq)) {
3327 tp->tlp_high_seq = 0;
3328 /* Don't reduce cwnd if DSACK arrives for TLP retrans. */
3329 if (!(flag & FLAG_DSACKING_ACK)) {
3330 tcp_init_cwnd_reduction(sk, true);
3331 tcp_set_ca_state(sk, TCP_CA_CWR);
3332 tcp_end_cwnd_reduction(sk);
3333 tcp_set_ca_state(sk, TCP_CA_Open);
3334 NET_INC_STATS_BH(sock_net(sk),
3335 LINUX_MIB_TCPLOSSPROBERECOVERY);
3340 /* This routine deals with incoming acks, but not outgoing ones. */
3341 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3343 struct inet_connection_sock *icsk = inet_csk(sk);
3344 struct tcp_sock *tp = tcp_sk(sk);
3345 u32 prior_snd_una = tp->snd_una;
3346 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3347 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3348 bool is_dupack = false;
3349 u32 prior_in_flight, prior_cwnd = tp->snd_cwnd, prior_rtt = tp->srtt;
3351 int prior_packets = tp->packets_out;
3352 const int prior_unsacked = tp->packets_out - tp->sacked_out;
3353 int acked = 0; /* Number of packets newly acked */
3356 /* If the ack is older than previous acks
3357 * then we can probably ignore it.
3359 if (before(ack, prior_snd_una)) {
3360 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3361 if (before(ack, prior_snd_una - tp->max_window)) {
3362 tcp_send_challenge_ack(sk);
3368 /* If the ack includes data we haven't sent yet, discard
3369 * this segment (RFC793 Section 3.9).
3371 if (after(ack, tp->snd_nxt))
3374 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
3375 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
3378 if (after(ack, prior_snd_una))
3379 flag |= FLAG_SND_UNA_ADVANCED;
3381 prior_fackets = tp->fackets_out;
3382 prior_in_flight = tcp_packets_in_flight(tp);
3384 /* ts_recent update must be made after we are sure that the packet
3387 if (flag & FLAG_UPDATE_TS_RECENT)
3388 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3390 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3391 /* Window is constant, pure forward advance.
3392 * No more checks are required.
3393 * Note, we use the fact that SND.UNA>=SND.WL2.
3395 tcp_update_wl(tp, ack_seq);
3397 flag |= FLAG_WIN_UPDATE;
3399 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
3401 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3403 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3406 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3408 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3410 if (TCP_SKB_CB(skb)->sacked)
3411 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3414 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
3417 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
3420 /* We passed data and got it acked, remove any soft error
3421 * log. Something worked...
3423 sk->sk_err_soft = 0;
3424 icsk->icsk_probes_out = 0;
3425 tp->rcv_tstamp = tcp_time_stamp;
3429 /* See if we can take anything off of the retransmit queue. */
3430 acked = tp->packets_out;
3431 flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una, sack_rtt);
3432 acked -= tp->packets_out;
3434 /* Advance cwnd if state allows */
3435 if (tcp_may_raise_cwnd(sk, flag))
3436 tcp_cong_avoid(sk, ack, prior_in_flight);
3438 if (tcp_ack_is_dubious(sk, flag)) {
3439 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3440 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3443 if (tp->tlp_high_seq)
3444 tcp_process_tlp_ack(sk, ack, flag);
3446 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) {
3447 struct dst_entry *dst = __sk_dst_get(sk);
3452 if (icsk->icsk_pending == ICSK_TIME_RETRANS)
3453 tcp_schedule_loss_probe(sk);
3454 if (tp->srtt != prior_rtt || tp->snd_cwnd != prior_cwnd)
3455 tcp_update_pacing_rate(sk);
3459 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3460 if (flag & FLAG_DSACKING_ACK)
3461 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3463 /* If this ack opens up a zero window, clear backoff. It was
3464 * being used to time the probes, and is probably far higher than
3465 * it needs to be for normal retransmission.
3467 if (tcp_send_head(sk))
3470 if (tp->tlp_high_seq)
3471 tcp_process_tlp_ack(sk, ack, flag);
3475 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3479 /* If data was SACKed, tag it and see if we should send more data.
3480 * If data was DSACKed, see if we can undo a cwnd reduction.
3482 if (TCP_SKB_CB(skb)->sacked) {
3483 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3485 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3489 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3493 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3494 * But, this can also be called on packets in the established flow when
3495 * the fast version below fails.
3497 void tcp_parse_options(const struct sk_buff *skb,
3498 struct tcp_options_received *opt_rx, int estab,
3499 struct tcp_fastopen_cookie *foc)
3501 const unsigned char *ptr;
3502 const struct tcphdr *th = tcp_hdr(skb);
3503 int length = (th->doff * 4) - sizeof(struct tcphdr);
3505 ptr = (const unsigned char *)(th + 1);
3506 opt_rx->saw_tstamp = 0;
3508 while (length > 0) {
3509 int opcode = *ptr++;
3515 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3520 if (opsize < 2) /* "silly options" */
3522 if (opsize > length)
3523 return; /* don't parse partial options */
3526 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3527 u16 in_mss = get_unaligned_be16(ptr);
3529 if (opt_rx->user_mss &&
3530 opt_rx->user_mss < in_mss)
3531 in_mss = opt_rx->user_mss;
3532 opt_rx->mss_clamp = in_mss;
3537 if (opsize == TCPOLEN_WINDOW && th->syn &&
3538 !estab && sysctl_tcp_window_scaling) {
3539 __u8 snd_wscale = *(__u8 *)ptr;
3540 opt_rx->wscale_ok = 1;
3541 if (snd_wscale > 14) {
3542 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3547 opt_rx->snd_wscale = snd_wscale;
3550 case TCPOPT_TIMESTAMP:
3551 if ((opsize == TCPOLEN_TIMESTAMP) &&
3552 ((estab && opt_rx->tstamp_ok) ||
3553 (!estab && sysctl_tcp_timestamps))) {
3554 opt_rx->saw_tstamp = 1;
3555 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3556 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3559 case TCPOPT_SACK_PERM:
3560 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3561 !estab && sysctl_tcp_sack) {
3562 opt_rx->sack_ok = TCP_SACK_SEEN;
3563 tcp_sack_reset(opt_rx);
3568 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3569 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3571 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3574 #ifdef CONFIG_TCP_MD5SIG
3577 * The MD5 Hash has already been
3578 * checked (see tcp_v{4,6}_do_rcv()).
3583 /* Fast Open option shares code 254 using a
3584 * 16 bits magic number. It's valid only in
3585 * SYN or SYN-ACK with an even size.
3587 if (opsize < TCPOLEN_EXP_FASTOPEN_BASE ||
3588 get_unaligned_be16(ptr) != TCPOPT_FASTOPEN_MAGIC ||
3589 foc == NULL || !th->syn || (opsize & 1))
3591 foc->len = opsize - TCPOLEN_EXP_FASTOPEN_BASE;
3592 if (foc->len >= TCP_FASTOPEN_COOKIE_MIN &&
3593 foc->len <= TCP_FASTOPEN_COOKIE_MAX)
3594 memcpy(foc->val, ptr + 2, foc->len);
3595 else if (foc->len != 0)
3605 EXPORT_SYMBOL(tcp_parse_options);
3607 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3609 const __be32 *ptr = (const __be32 *)(th + 1);
3611 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3612 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3613 tp->rx_opt.saw_tstamp = 1;
3615 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3618 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3620 tp->rx_opt.rcv_tsecr = 0;
3626 /* Fast parse options. This hopes to only see timestamps.
3627 * If it is wrong it falls back on tcp_parse_options().
3629 static bool tcp_fast_parse_options(const struct sk_buff *skb,
3630 const struct tcphdr *th, struct tcp_sock *tp)
3632 /* In the spirit of fast parsing, compare doff directly to constant
3633 * values. Because equality is used, short doff can be ignored here.
3635 if (th->doff == (sizeof(*th) / 4)) {
3636 tp->rx_opt.saw_tstamp = 0;
3638 } else if (tp->rx_opt.tstamp_ok &&
3639 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3640 if (tcp_parse_aligned_timestamp(tp, th))
3644 tcp_parse_options(skb, &tp->rx_opt, 1, NULL);
3645 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3646 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3651 #ifdef CONFIG_TCP_MD5SIG
3653 * Parse MD5 Signature option
3655 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3657 int length = (th->doff << 2) - sizeof(*th);
3658 const u8 *ptr = (const u8 *)(th + 1);
3660 /* If the TCP option is too short, we can short cut */
3661 if (length < TCPOLEN_MD5SIG)
3664 while (length > 0) {
3665 int opcode = *ptr++;
3676 if (opsize < 2 || opsize > length)
3678 if (opcode == TCPOPT_MD5SIG)
3679 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3686 EXPORT_SYMBOL(tcp_parse_md5sig_option);
3689 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3691 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3692 * it can pass through stack. So, the following predicate verifies that
3693 * this segment is not used for anything but congestion avoidance or
3694 * fast retransmit. Moreover, we even are able to eliminate most of such
3695 * second order effects, if we apply some small "replay" window (~RTO)
3696 * to timestamp space.
3698 * All these measures still do not guarantee that we reject wrapped ACKs
3699 * on networks with high bandwidth, when sequence space is recycled fastly,
3700 * but it guarantees that such events will be very rare and do not affect
3701 * connection seriously. This doesn't look nice, but alas, PAWS is really
3704 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3705 * states that events when retransmit arrives after original data are rare.
3706 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3707 * the biggest problem on large power networks even with minor reordering.
3708 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3709 * up to bandwidth of 18Gigabit/sec. 8) ]
3712 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3714 const struct tcp_sock *tp = tcp_sk(sk);
3715 const struct tcphdr *th = tcp_hdr(skb);
3716 u32 seq = TCP_SKB_CB(skb)->seq;
3717 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3719 return (/* 1. Pure ACK with correct sequence number. */
3720 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3722 /* 2. ... and duplicate ACK. */
3723 ack == tp->snd_una &&
3725 /* 3. ... and does not update window. */
3726 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3728 /* 4. ... and sits in replay window. */
3729 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3732 static inline bool tcp_paws_discard(const struct sock *sk,
3733 const struct sk_buff *skb)
3735 const struct tcp_sock *tp = tcp_sk(sk);
3737 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
3738 !tcp_disordered_ack(sk, skb);
3741 /* Check segment sequence number for validity.
3743 * Segment controls are considered valid, if the segment
3744 * fits to the window after truncation to the window. Acceptability
3745 * of data (and SYN, FIN, of course) is checked separately.
3746 * See tcp_data_queue(), for example.
3748 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3749 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3750 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3751 * (borrowed from freebsd)
3754 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
3756 return !before(end_seq, tp->rcv_wup) &&
3757 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3760 /* When we get a reset we do this. */
3761 void tcp_reset(struct sock *sk)
3763 /* We want the right error as BSD sees it (and indeed as we do). */
3764 switch (sk->sk_state) {
3766 sk->sk_err = ECONNREFUSED;
3768 case TCP_CLOSE_WAIT:
3774 sk->sk_err = ECONNRESET;
3776 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3779 if (!sock_flag(sk, SOCK_DEAD))
3780 sk->sk_error_report(sk);
3786 * Process the FIN bit. This now behaves as it is supposed to work
3787 * and the FIN takes effect when it is validly part of sequence
3788 * space. Not before when we get holes.
3790 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3791 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3794 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3795 * close and we go into CLOSING (and later onto TIME-WAIT)
3797 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3799 static void tcp_fin(struct sock *sk)
3801 struct tcp_sock *tp = tcp_sk(sk);
3802 const struct dst_entry *dst;
3804 inet_csk_schedule_ack(sk);
3806 sk->sk_shutdown |= RCV_SHUTDOWN;
3807 sock_set_flag(sk, SOCK_DONE);
3809 switch (sk->sk_state) {
3811 case TCP_ESTABLISHED:
3812 /* Move to CLOSE_WAIT */
3813 tcp_set_state(sk, TCP_CLOSE_WAIT);
3814 dst = __sk_dst_get(sk);
3815 if (!dst || !dst_metric(dst, RTAX_QUICKACK))
3816 inet_csk(sk)->icsk_ack.pingpong = 1;
3819 case TCP_CLOSE_WAIT:
3821 /* Received a retransmission of the FIN, do
3826 /* RFC793: Remain in the LAST-ACK state. */
3830 /* This case occurs when a simultaneous close
3831 * happens, we must ack the received FIN and
3832 * enter the CLOSING state.
3835 tcp_set_state(sk, TCP_CLOSING);
3838 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3840 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
3843 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3844 * cases we should never reach this piece of code.
3846 pr_err("%s: Impossible, sk->sk_state=%d\n",
3847 __func__, sk->sk_state);
3851 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3852 * Probably, we should reset in this case. For now drop them.
3854 __skb_queue_purge(&tp->out_of_order_queue);
3855 if (tcp_is_sack(tp))
3856 tcp_sack_reset(&tp->rx_opt);
3859 if (!sock_flag(sk, SOCK_DEAD)) {
3860 sk->sk_state_change(sk);
3862 /* Do not send POLL_HUP for half duplex close. */
3863 if (sk->sk_shutdown == SHUTDOWN_MASK ||
3864 sk->sk_state == TCP_CLOSE)
3865 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
3867 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
3871 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
3874 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
3875 if (before(seq, sp->start_seq))
3876 sp->start_seq = seq;
3877 if (after(end_seq, sp->end_seq))
3878 sp->end_seq = end_seq;
3884 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
3886 struct tcp_sock *tp = tcp_sk(sk);
3888 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3891 if (before(seq, tp->rcv_nxt))
3892 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
3894 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
3896 NET_INC_STATS_BH(sock_net(sk), mib_idx);
3898 tp->rx_opt.dsack = 1;
3899 tp->duplicate_sack[0].start_seq = seq;
3900 tp->duplicate_sack[0].end_seq = end_seq;
3904 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
3906 struct tcp_sock *tp = tcp_sk(sk);
3908 if (!tp->rx_opt.dsack)
3909 tcp_dsack_set(sk, seq, end_seq);
3911 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
3914 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
3916 struct tcp_sock *tp = tcp_sk(sk);
3918 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
3919 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3920 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
3921 tcp_enter_quickack_mode(sk);
3923 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3924 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3926 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
3927 end_seq = tp->rcv_nxt;
3928 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
3935 /* These routines update the SACK block as out-of-order packets arrive or
3936 * in-order packets close up the sequence space.
3938 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
3941 struct tcp_sack_block *sp = &tp->selective_acks[0];
3942 struct tcp_sack_block *swalk = sp + 1;
3944 /* See if the recent change to the first SACK eats into
3945 * or hits the sequence space of other SACK blocks, if so coalesce.
3947 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
3948 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
3951 /* Zap SWALK, by moving every further SACK up by one slot.
3952 * Decrease num_sacks.
3954 tp->rx_opt.num_sacks--;
3955 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
3959 this_sack++, swalk++;
3963 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
3965 struct tcp_sock *tp = tcp_sk(sk);
3966 struct tcp_sack_block *sp = &tp->selective_acks[0];
3967 int cur_sacks = tp->rx_opt.num_sacks;
3973 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
3974 if (tcp_sack_extend(sp, seq, end_seq)) {
3975 /* Rotate this_sack to the first one. */
3976 for (; this_sack > 0; this_sack--, sp--)
3977 swap(*sp, *(sp - 1));
3979 tcp_sack_maybe_coalesce(tp);
3984 /* Could not find an adjacent existing SACK, build a new one,
3985 * put it at the front, and shift everyone else down. We
3986 * always know there is at least one SACK present already here.
3988 * If the sack array is full, forget about the last one.
3990 if (this_sack >= TCP_NUM_SACKS) {
3992 tp->rx_opt.num_sacks--;
3995 for (; this_sack > 0; this_sack--, sp--)
3999 /* Build the new head SACK, and we're done. */
4000 sp->start_seq = seq;
4001 sp->end_seq = end_seq;
4002 tp->rx_opt.num_sacks++;
4005 /* RCV.NXT advances, some SACKs should be eaten. */
4007 static void tcp_sack_remove(struct tcp_sock *tp)
4009 struct tcp_sack_block *sp = &tp->selective_acks[0];
4010 int num_sacks = tp->rx_opt.num_sacks;
4013 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4014 if (skb_queue_empty(&tp->out_of_order_queue)) {
4015 tp->rx_opt.num_sacks = 0;
4019 for (this_sack = 0; this_sack < num_sacks;) {
4020 /* Check if the start of the sack is covered by RCV.NXT. */
4021 if (!before(tp->rcv_nxt, sp->start_seq)) {
4024 /* RCV.NXT must cover all the block! */
4025 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4027 /* Zap this SACK, by moving forward any other SACKS. */
4028 for (i=this_sack+1; i < num_sacks; i++)
4029 tp->selective_acks[i-1] = tp->selective_acks[i];
4036 tp->rx_opt.num_sacks = num_sacks;
4039 /* This one checks to see if we can put data from the
4040 * out_of_order queue into the receive_queue.
4042 static void tcp_ofo_queue(struct sock *sk)
4044 struct tcp_sock *tp = tcp_sk(sk);
4045 __u32 dsack_high = tp->rcv_nxt;
4046 struct sk_buff *skb;
4048 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4049 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4052 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4053 __u32 dsack = dsack_high;
4054 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4055 dsack_high = TCP_SKB_CB(skb)->end_seq;
4056 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4059 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4060 SOCK_DEBUG(sk, "ofo packet was already received\n");
4061 __skb_unlink(skb, &tp->out_of_order_queue);
4065 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4066 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4067 TCP_SKB_CB(skb)->end_seq);
4069 __skb_unlink(skb, &tp->out_of_order_queue);
4070 __skb_queue_tail(&sk->sk_receive_queue, skb);
4071 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4072 if (tcp_hdr(skb)->fin)
4077 static bool tcp_prune_ofo_queue(struct sock *sk);
4078 static int tcp_prune_queue(struct sock *sk);
4080 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4083 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4084 !sk_rmem_schedule(sk, skb, size)) {
4086 if (tcp_prune_queue(sk) < 0)
4089 if (!sk_rmem_schedule(sk, skb, size)) {
4090 if (!tcp_prune_ofo_queue(sk))
4093 if (!sk_rmem_schedule(sk, skb, size))
4101 * tcp_try_coalesce - try to merge skb to prior one
4104 * @from: buffer to add in queue
4105 * @fragstolen: pointer to boolean
4107 * Before queueing skb @from after @to, try to merge them
4108 * to reduce overall memory use and queue lengths, if cost is small.
4109 * Packets in ofo or receive queues can stay a long time.
4110 * Better try to coalesce them right now to avoid future collapses.
4111 * Returns true if caller should free @from instead of queueing it
4113 static bool tcp_try_coalesce(struct sock *sk,
4115 struct sk_buff *from,
4120 *fragstolen = false;
4122 if (tcp_hdr(from)->fin)
4125 /* Its possible this segment overlaps with prior segment in queue */
4126 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4129 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4132 atomic_add(delta, &sk->sk_rmem_alloc);
4133 sk_mem_charge(sk, delta);
4134 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4135 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4136 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4140 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4142 struct tcp_sock *tp = tcp_sk(sk);
4143 struct sk_buff *skb1;
4146 TCP_ECN_check_ce(tp, skb);
4148 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4149 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFODROP);
4154 /* Disable header prediction. */
4156 inet_csk_schedule_ack(sk);
4158 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4159 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4160 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4162 skb1 = skb_peek_tail(&tp->out_of_order_queue);
4164 /* Initial out of order segment, build 1 SACK. */
4165 if (tcp_is_sack(tp)) {
4166 tp->rx_opt.num_sacks = 1;
4167 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4168 tp->selective_acks[0].end_seq =
4169 TCP_SKB_CB(skb)->end_seq;
4171 __skb_queue_head(&tp->out_of_order_queue, skb);
4175 seq = TCP_SKB_CB(skb)->seq;
4176 end_seq = TCP_SKB_CB(skb)->end_seq;
4178 if (seq == TCP_SKB_CB(skb1)->end_seq) {
4181 if (!tcp_try_coalesce(sk, skb1, skb, &fragstolen)) {
4182 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4184 tcp_grow_window(sk, skb);
4185 kfree_skb_partial(skb, fragstolen);
4189 if (!tp->rx_opt.num_sacks ||
4190 tp->selective_acks[0].end_seq != seq)
4193 /* Common case: data arrive in order after hole. */
4194 tp->selective_acks[0].end_seq = end_seq;
4198 /* Find place to insert this segment. */
4200 if (!after(TCP_SKB_CB(skb1)->seq, seq))
4202 if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
4206 skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
4209 /* Do skb overlap to previous one? */
4210 if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4211 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4212 /* All the bits are present. Drop. */
4213 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4216 tcp_dsack_set(sk, seq, end_seq);
4219 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4220 /* Partial overlap. */
4221 tcp_dsack_set(sk, seq,
4222 TCP_SKB_CB(skb1)->end_seq);
4224 if (skb_queue_is_first(&tp->out_of_order_queue,
4228 skb1 = skb_queue_prev(
4229 &tp->out_of_order_queue,
4234 __skb_queue_head(&tp->out_of_order_queue, skb);
4236 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4238 /* And clean segments covered by new one as whole. */
4239 while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
4240 skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
4242 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4244 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4245 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4249 __skb_unlink(skb1, &tp->out_of_order_queue);
4250 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4251 TCP_SKB_CB(skb1)->end_seq);
4252 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4257 if (tcp_is_sack(tp))
4258 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4261 tcp_grow_window(sk, skb);
4262 skb_set_owner_r(skb, sk);
4266 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
4270 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4272 __skb_pull(skb, hdrlen);
4274 tcp_try_coalesce(sk, tail, skb, fragstolen)) ? 1 : 0;
4275 tcp_sk(sk)->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4277 __skb_queue_tail(&sk->sk_receive_queue, skb);
4278 skb_set_owner_r(skb, sk);
4283 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4285 struct sk_buff *skb = NULL;
4292 skb = alloc_skb(size + sizeof(*th), sk->sk_allocation);
4296 if (tcp_try_rmem_schedule(sk, skb, size + sizeof(*th)))
4299 th = (struct tcphdr *)skb_put(skb, sizeof(*th));
4300 skb_reset_transport_header(skb);
4301 memset(th, 0, sizeof(*th));
4303 if (memcpy_fromiovec(skb_put(skb, size), msg->msg_iov, size))
4306 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4307 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4308 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4310 if (tcp_queue_rcv(sk, skb, sizeof(*th), &fragstolen)) {
4311 WARN_ON_ONCE(fragstolen); /* should not happen */
4322 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4324 const struct tcphdr *th = tcp_hdr(skb);
4325 struct tcp_sock *tp = tcp_sk(sk);
4327 bool fragstolen = false;
4329 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4333 __skb_pull(skb, th->doff * 4);
4335 TCP_ECN_accept_cwr(tp, skb);
4337 tp->rx_opt.dsack = 0;
4339 /* Queue data for delivery to the user.
4340 * Packets in sequence go to the receive queue.
4341 * Out of sequence packets to the out_of_order_queue.
4343 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4344 if (tcp_receive_window(tp) == 0)
4347 /* Ok. In sequence. In window. */
4348 if (tp->ucopy.task == current &&
4349 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4350 sock_owned_by_user(sk) && !tp->urg_data) {
4351 int chunk = min_t(unsigned int, skb->len,
4354 __set_current_state(TASK_RUNNING);
4357 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
4358 tp->ucopy.len -= chunk;
4359 tp->copied_seq += chunk;
4360 eaten = (chunk == skb->len);
4361 tcp_rcv_space_adjust(sk);
4369 tcp_try_rmem_schedule(sk, skb, skb->truesize))
4372 eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
4374 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4376 tcp_event_data_recv(sk, skb);
4380 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4383 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4384 * gap in queue is filled.
4386 if (skb_queue_empty(&tp->out_of_order_queue))
4387 inet_csk(sk)->icsk_ack.pingpong = 0;
4390 if (tp->rx_opt.num_sacks)
4391 tcp_sack_remove(tp);
4393 tcp_fast_path_check(sk);
4396 kfree_skb_partial(skb, fragstolen);
4397 if (!sock_flag(sk, SOCK_DEAD))
4398 sk->sk_data_ready(sk, 0);
4402 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4403 /* A retransmit, 2nd most common case. Force an immediate ack. */
4404 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4405 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4408 tcp_enter_quickack_mode(sk);
4409 inet_csk_schedule_ack(sk);
4415 /* Out of window. F.e. zero window probe. */
4416 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4419 tcp_enter_quickack_mode(sk);
4421 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4422 /* Partial packet, seq < rcv_next < end_seq */
4423 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4424 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4425 TCP_SKB_CB(skb)->end_seq);
4427 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4429 /* If window is closed, drop tail of packet. But after
4430 * remembering D-SACK for its head made in previous line.
4432 if (!tcp_receive_window(tp))
4437 tcp_data_queue_ofo(sk, skb);
4440 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4441 struct sk_buff_head *list)
4443 struct sk_buff *next = NULL;
4445 if (!skb_queue_is_last(list, skb))
4446 next = skb_queue_next(list, skb);
4448 __skb_unlink(skb, list);
4450 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4455 /* Collapse contiguous sequence of skbs head..tail with
4456 * sequence numbers start..end.
4458 * If tail is NULL, this means until the end of the list.
4460 * Segments with FIN/SYN are not collapsed (only because this
4464 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4465 struct sk_buff *head, struct sk_buff *tail,
4468 struct sk_buff *skb, *n;
4471 /* First, check that queue is collapsible and find
4472 * the point where collapsing can be useful. */
4476 skb_queue_walk_from_safe(list, skb, n) {
4479 /* No new bits? It is possible on ofo queue. */
4480 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4481 skb = tcp_collapse_one(sk, skb, list);
4487 /* The first skb to collapse is:
4489 * - bloated or contains data before "start" or
4490 * overlaps to the next one.
4492 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
4493 (tcp_win_from_space(skb->truesize) > skb->len ||
4494 before(TCP_SKB_CB(skb)->seq, start))) {
4495 end_of_skbs = false;
4499 if (!skb_queue_is_last(list, skb)) {
4500 struct sk_buff *next = skb_queue_next(list, skb);
4502 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
4503 end_of_skbs = false;
4508 /* Decided to skip this, advance start seq. */
4509 start = TCP_SKB_CB(skb)->end_seq;
4511 if (end_of_skbs || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
4514 while (before(start, end)) {
4515 struct sk_buff *nskb;
4516 unsigned int header = skb_headroom(skb);
4517 int copy = SKB_MAX_ORDER(header, 0);
4519 /* Too big header? This can happen with IPv6. */
4522 if (end - start < copy)
4524 nskb = alloc_skb(copy + header, GFP_ATOMIC);
4528 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
4529 skb_set_network_header(nskb, (skb_network_header(skb) -
4531 skb_set_transport_header(nskb, (skb_transport_header(skb) -
4533 skb_reserve(nskb, header);
4534 memcpy(nskb->head, skb->head, header);
4535 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4536 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4537 __skb_queue_before(list, skb, nskb);
4538 skb_set_owner_r(nskb, sk);
4540 /* Copy data, releasing collapsed skbs. */
4542 int offset = start - TCP_SKB_CB(skb)->seq;
4543 int size = TCP_SKB_CB(skb)->end_seq - start;
4547 size = min(copy, size);
4548 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4550 TCP_SKB_CB(nskb)->end_seq += size;
4554 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4555 skb = tcp_collapse_one(sk, skb, list);
4558 tcp_hdr(skb)->syn ||
4566 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4567 * and tcp_collapse() them until all the queue is collapsed.
4569 static void tcp_collapse_ofo_queue(struct sock *sk)
4571 struct tcp_sock *tp = tcp_sk(sk);
4572 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4573 struct sk_buff *head;
4579 start = TCP_SKB_CB(skb)->seq;
4580 end = TCP_SKB_CB(skb)->end_seq;
4584 struct sk_buff *next = NULL;
4586 if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
4587 next = skb_queue_next(&tp->out_of_order_queue, skb);
4590 /* Segment is terminated when we see gap or when
4591 * we are at the end of all the queue. */
4593 after(TCP_SKB_CB(skb)->seq, end) ||
4594 before(TCP_SKB_CB(skb)->end_seq, start)) {
4595 tcp_collapse(sk, &tp->out_of_order_queue,
4596 head, skb, start, end);
4600 /* Start new segment */
4601 start = TCP_SKB_CB(skb)->seq;
4602 end = TCP_SKB_CB(skb)->end_seq;
4604 if (before(TCP_SKB_CB(skb)->seq, start))
4605 start = TCP_SKB_CB(skb)->seq;
4606 if (after(TCP_SKB_CB(skb)->end_seq, end))
4607 end = TCP_SKB_CB(skb)->end_seq;
4613 * Purge the out-of-order queue.
4614 * Return true if queue was pruned.
4616 static bool tcp_prune_ofo_queue(struct sock *sk)
4618 struct tcp_sock *tp = tcp_sk(sk);
4621 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4622 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
4623 __skb_queue_purge(&tp->out_of_order_queue);
4625 /* Reset SACK state. A conforming SACK implementation will
4626 * do the same at a timeout based retransmit. When a connection
4627 * is in a sad state like this, we care only about integrity
4628 * of the connection not performance.
4630 if (tp->rx_opt.sack_ok)
4631 tcp_sack_reset(&tp->rx_opt);
4638 /* Reduce allocated memory if we can, trying to get
4639 * the socket within its memory limits again.
4641 * Return less than zero if we should start dropping frames
4642 * until the socket owning process reads some of the data
4643 * to stabilize the situation.
4645 static int tcp_prune_queue(struct sock *sk)
4647 struct tcp_sock *tp = tcp_sk(sk);
4649 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4651 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
4653 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4654 tcp_clamp_window(sk);
4655 else if (sk_under_memory_pressure(sk))
4656 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4658 tcp_collapse_ofo_queue(sk);
4659 if (!skb_queue_empty(&sk->sk_receive_queue))
4660 tcp_collapse(sk, &sk->sk_receive_queue,
4661 skb_peek(&sk->sk_receive_queue),
4663 tp->copied_seq, tp->rcv_nxt);
4666 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4669 /* Collapsing did not help, destructive actions follow.
4670 * This must not ever occur. */
4672 tcp_prune_ofo_queue(sk);
4674 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4677 /* If we are really being abused, tell the caller to silently
4678 * drop receive data on the floor. It will get retransmitted
4679 * and hopefully then we'll have sufficient space.
4681 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
4683 /* Massive buffer overcommit. */
4688 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4689 * As additional protections, we do not touch cwnd in retransmission phases,
4690 * and if application hit its sndbuf limit recently.
4692 void tcp_cwnd_application_limited(struct sock *sk)
4694 struct tcp_sock *tp = tcp_sk(sk);
4696 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
4697 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4698 /* Limited by application or receiver window. */
4699 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4700 u32 win_used = max(tp->snd_cwnd_used, init_win);
4701 if (win_used < tp->snd_cwnd) {
4702 tp->snd_ssthresh = tcp_current_ssthresh(sk);
4703 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4705 tp->snd_cwnd_used = 0;
4707 tp->snd_cwnd_stamp = tcp_time_stamp;
4710 static bool tcp_should_expand_sndbuf(const struct sock *sk)
4712 const struct tcp_sock *tp = tcp_sk(sk);
4714 /* If the user specified a specific send buffer setting, do
4717 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4720 /* If we are under global TCP memory pressure, do not expand. */
4721 if (sk_under_memory_pressure(sk))
4724 /* If we are under soft global TCP memory pressure, do not expand. */
4725 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
4728 /* If we filled the congestion window, do not expand. */
4729 if (tp->packets_out >= tp->snd_cwnd)
4735 /* When incoming ACK allowed to free some skb from write_queue,
4736 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4737 * on the exit from tcp input handler.
4739 * PROBLEM: sndbuf expansion does not work well with largesend.
4741 static void tcp_new_space(struct sock *sk)
4743 struct tcp_sock *tp = tcp_sk(sk);
4745 if (tcp_should_expand_sndbuf(sk)) {
4746 tcp_sndbuf_expand(sk);
4747 tp->snd_cwnd_stamp = tcp_time_stamp;
4750 sk->sk_write_space(sk);
4753 static void tcp_check_space(struct sock *sk)
4755 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4756 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4757 if (sk->sk_socket &&
4758 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4763 static inline void tcp_data_snd_check(struct sock *sk)
4765 tcp_push_pending_frames(sk);
4766 tcp_check_space(sk);
4770 * Check if sending an ack is needed.
4772 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4774 struct tcp_sock *tp = tcp_sk(sk);
4776 /* More than one full frame received... */
4777 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
4778 /* ... and right edge of window advances far enough.
4779 * (tcp_recvmsg() will send ACK otherwise). Or...
4781 __tcp_select_window(sk) >= tp->rcv_wnd) ||
4782 /* We ACK each frame or... */
4783 tcp_in_quickack_mode(sk) ||
4784 /* We have out of order data. */
4785 (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
4786 /* Then ack it now */
4789 /* Else, send delayed ack. */
4790 tcp_send_delayed_ack(sk);
4794 static inline void tcp_ack_snd_check(struct sock *sk)
4796 if (!inet_csk_ack_scheduled(sk)) {
4797 /* We sent a data segment already. */
4800 __tcp_ack_snd_check(sk, 1);
4804 * This routine is only called when we have urgent data
4805 * signaled. Its the 'slow' part of tcp_urg. It could be
4806 * moved inline now as tcp_urg is only called from one
4807 * place. We handle URGent data wrong. We have to - as
4808 * BSD still doesn't use the correction from RFC961.
4809 * For 1003.1g we should support a new option TCP_STDURG to permit
4810 * either form (or just set the sysctl tcp_stdurg).
4813 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
4815 struct tcp_sock *tp = tcp_sk(sk);
4816 u32 ptr = ntohs(th->urg_ptr);
4818 if (ptr && !sysctl_tcp_stdurg)
4820 ptr += ntohl(th->seq);
4822 /* Ignore urgent data that we've already seen and read. */
4823 if (after(tp->copied_seq, ptr))
4826 /* Do not replay urg ptr.
4828 * NOTE: interesting situation not covered by specs.
4829 * Misbehaving sender may send urg ptr, pointing to segment,
4830 * which we already have in ofo queue. We are not able to fetch
4831 * such data and will stay in TCP_URG_NOTYET until will be eaten
4832 * by recvmsg(). Seems, we are not obliged to handle such wicked
4833 * situations. But it is worth to think about possibility of some
4834 * DoSes using some hypothetical application level deadlock.
4836 if (before(ptr, tp->rcv_nxt))
4839 /* Do we already have a newer (or duplicate) urgent pointer? */
4840 if (tp->urg_data && !after(ptr, tp->urg_seq))
4843 /* Tell the world about our new urgent pointer. */
4846 /* We may be adding urgent data when the last byte read was
4847 * urgent. To do this requires some care. We cannot just ignore
4848 * tp->copied_seq since we would read the last urgent byte again
4849 * as data, nor can we alter copied_seq until this data arrives
4850 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4852 * NOTE. Double Dutch. Rendering to plain English: author of comment
4853 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4854 * and expect that both A and B disappear from stream. This is _wrong_.
4855 * Though this happens in BSD with high probability, this is occasional.
4856 * Any application relying on this is buggy. Note also, that fix "works"
4857 * only in this artificial test. Insert some normal data between A and B and we will
4858 * decline of BSD again. Verdict: it is better to remove to trap
4861 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
4862 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
4863 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4865 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
4866 __skb_unlink(skb, &sk->sk_receive_queue);
4871 tp->urg_data = TCP_URG_NOTYET;
4874 /* Disable header prediction. */
4878 /* This is the 'fast' part of urgent handling. */
4879 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
4881 struct tcp_sock *tp = tcp_sk(sk);
4883 /* Check if we get a new urgent pointer - normally not. */
4885 tcp_check_urg(sk, th);
4887 /* Do we wait for any urgent data? - normally not... */
4888 if (tp->urg_data == TCP_URG_NOTYET) {
4889 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
4892 /* Is the urgent pointer pointing into this packet? */
4893 if (ptr < skb->len) {
4895 if (skb_copy_bits(skb, ptr, &tmp, 1))
4897 tp->urg_data = TCP_URG_VALID | tmp;
4898 if (!sock_flag(sk, SOCK_DEAD))
4899 sk->sk_data_ready(sk, 0);
4904 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4906 struct tcp_sock *tp = tcp_sk(sk);
4907 int chunk = skb->len - hlen;
4911 if (skb_csum_unnecessary(skb))
4912 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
4914 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
4918 tp->ucopy.len -= chunk;
4919 tp->copied_seq += chunk;
4920 tcp_rcv_space_adjust(sk);
4927 static __sum16 __tcp_checksum_complete_user(struct sock *sk,
4928 struct sk_buff *skb)
4932 if (sock_owned_by_user(sk)) {
4934 result = __tcp_checksum_complete(skb);
4937 result = __tcp_checksum_complete(skb);
4942 static inline bool tcp_checksum_complete_user(struct sock *sk,
4943 struct sk_buff *skb)
4945 return !skb_csum_unnecessary(skb) &&
4946 __tcp_checksum_complete_user(sk, skb);
4949 #ifdef CONFIG_NET_DMA
4950 static bool tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb,
4953 struct tcp_sock *tp = tcp_sk(sk);
4954 int chunk = skb->len - hlen;
4956 bool copied_early = false;
4958 if (tp->ucopy.wakeup)
4961 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
4962 tp->ucopy.dma_chan = net_dma_find_channel();
4964 if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
4966 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
4968 tp->ucopy.iov, chunk,
4969 tp->ucopy.pinned_list);
4974 tp->ucopy.dma_cookie = dma_cookie;
4975 copied_early = true;
4977 tp->ucopy.len -= chunk;
4978 tp->copied_seq += chunk;
4979 tcp_rcv_space_adjust(sk);
4981 if ((tp->ucopy.len == 0) ||
4982 (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
4983 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
4984 tp->ucopy.wakeup = 1;
4985 sk->sk_data_ready(sk, 0);
4987 } else if (chunk > 0) {
4988 tp->ucopy.wakeup = 1;
4989 sk->sk_data_ready(sk, 0);
4992 return copied_early;
4994 #endif /* CONFIG_NET_DMA */
4996 /* Does PAWS and seqno based validation of an incoming segment, flags will
4997 * play significant role here.
4999 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5000 const struct tcphdr *th, int syn_inerr)
5002 struct tcp_sock *tp = tcp_sk(sk);
5004 /* RFC1323: H1. Apply PAWS check first. */
5005 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
5006 tcp_paws_discard(sk, skb)) {
5008 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5009 tcp_send_dupack(sk, skb);
5012 /* Reset is accepted even if it did not pass PAWS. */
5015 /* Step 1: check sequence number */
5016 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5017 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5018 * (RST) segments are validated by checking their SEQ-fields."
5019 * And page 69: "If an incoming segment is not acceptable,
5020 * an acknowledgment should be sent in reply (unless the RST
5021 * bit is set, if so drop the segment and return)".
5026 tcp_send_dupack(sk, skb);
5031 /* Step 2: check RST bit */
5034 * If sequence number exactly matches RCV.NXT, then
5035 * RESET the connection
5037 * Send a challenge ACK
5039 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt)
5042 tcp_send_challenge_ack(sk);
5046 /* step 3: check security and precedence [ignored] */
5048 /* step 4: Check for a SYN
5049 * RFC 5691 4.2 : Send a challenge ack
5054 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5055 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5056 tcp_send_challenge_ack(sk);
5068 * TCP receive function for the ESTABLISHED state.
5070 * It is split into a fast path and a slow path. The fast path is
5072 * - A zero window was announced from us - zero window probing
5073 * is only handled properly in the slow path.
5074 * - Out of order segments arrived.
5075 * - Urgent data is expected.
5076 * - There is no buffer space left
5077 * - Unexpected TCP flags/window values/header lengths are received
5078 * (detected by checking the TCP header against pred_flags)
5079 * - Data is sent in both directions. Fast path only supports pure senders
5080 * or pure receivers (this means either the sequence number or the ack
5081 * value must stay constant)
5082 * - Unexpected TCP option.
5084 * When these conditions are not satisfied it drops into a standard
5085 * receive procedure patterned after RFC793 to handle all cases.
5086 * The first three cases are guaranteed by proper pred_flags setting,
5087 * the rest is checked inline. Fast processing is turned on in
5088 * tcp_data_queue when everything is OK.
5090 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5091 const struct tcphdr *th, unsigned int len)
5093 struct tcp_sock *tp = tcp_sk(sk);
5095 if (unlikely(sk->sk_rx_dst == NULL))
5096 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5098 * Header prediction.
5099 * The code loosely follows the one in the famous
5100 * "30 instruction TCP receive" Van Jacobson mail.
5102 * Van's trick is to deposit buffers into socket queue
5103 * on a device interrupt, to call tcp_recv function
5104 * on the receive process context and checksum and copy
5105 * the buffer to user space. smart...
5107 * Our current scheme is not silly either but we take the
5108 * extra cost of the net_bh soft interrupt processing...
5109 * We do checksum and copy also but from device to kernel.
5112 tp->rx_opt.saw_tstamp = 0;
5114 /* pred_flags is 0xS?10 << 16 + snd_wnd
5115 * if header_prediction is to be made
5116 * 'S' will always be tp->tcp_header_len >> 2
5117 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5118 * turn it off (when there are holes in the receive
5119 * space for instance)
5120 * PSH flag is ignored.
5123 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5124 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5125 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5126 int tcp_header_len = tp->tcp_header_len;
5128 /* Timestamp header prediction: tcp_header_len
5129 * is automatically equal to th->doff*4 due to pred_flags
5133 /* Check timestamp */
5134 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5135 /* No? Slow path! */
5136 if (!tcp_parse_aligned_timestamp(tp, th))
5139 /* If PAWS failed, check it more carefully in slow path */
5140 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5143 /* DO NOT update ts_recent here, if checksum fails
5144 * and timestamp was corrupted part, it will result
5145 * in a hung connection since we will drop all
5146 * future packets due to the PAWS test.
5150 if (len <= tcp_header_len) {
5151 /* Bulk data transfer: sender */
5152 if (len == tcp_header_len) {
5153 /* Predicted packet is in window by definition.
5154 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5155 * Hence, check seq<=rcv_wup reduces to:
5157 if (tcp_header_len ==
5158 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5159 tp->rcv_nxt == tp->rcv_wup)
5160 tcp_store_ts_recent(tp);
5162 /* We know that such packets are checksummed
5165 tcp_ack(sk, skb, 0);
5167 tcp_data_snd_check(sk);
5169 } else { /* Header too small */
5170 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5175 int copied_early = 0;
5176 bool fragstolen = false;
5178 if (tp->copied_seq == tp->rcv_nxt &&
5179 len - tcp_header_len <= tp->ucopy.len) {
5180 #ifdef CONFIG_NET_DMA
5181 if (tp->ucopy.task == current &&
5182 sock_owned_by_user(sk) &&
5183 tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
5188 if (tp->ucopy.task == current &&
5189 sock_owned_by_user(sk) && !copied_early) {
5190 __set_current_state(TASK_RUNNING);
5192 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
5196 /* Predicted packet is in window by definition.
5197 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5198 * Hence, check seq<=rcv_wup reduces to:
5200 if (tcp_header_len ==
5201 (sizeof(struct tcphdr) +
5202 TCPOLEN_TSTAMP_ALIGNED) &&
5203 tp->rcv_nxt == tp->rcv_wup)
5204 tcp_store_ts_recent(tp);
5206 tcp_rcv_rtt_measure_ts(sk, skb);
5208 __skb_pull(skb, tcp_header_len);
5209 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5210 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
5213 tcp_cleanup_rbuf(sk, skb->len);
5216 if (tcp_checksum_complete_user(sk, skb))
5219 if ((int)skb->truesize > sk->sk_forward_alloc)
5222 /* Predicted packet is in window by definition.
5223 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5224 * Hence, check seq<=rcv_wup reduces to:
5226 if (tcp_header_len ==
5227 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5228 tp->rcv_nxt == tp->rcv_wup)
5229 tcp_store_ts_recent(tp);
5231 tcp_rcv_rtt_measure_ts(sk, skb);
5233 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
5235 /* Bulk data transfer: receiver */
5236 eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5240 tcp_event_data_recv(sk, skb);
5242 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5243 /* Well, only one small jumplet in fast path... */
5244 tcp_ack(sk, skb, FLAG_DATA);
5245 tcp_data_snd_check(sk);
5246 if (!inet_csk_ack_scheduled(sk))
5250 if (!copied_early || tp->rcv_nxt != tp->rcv_wup)
5251 __tcp_ack_snd_check(sk, 0);
5253 #ifdef CONFIG_NET_DMA
5255 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
5259 kfree_skb_partial(skb, fragstolen);
5260 sk->sk_data_ready(sk, 0);
5266 if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
5269 if (!th->ack && !th->rst)
5273 * Standard slow path.
5276 if (!tcp_validate_incoming(sk, skb, th, 1))
5280 if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5283 tcp_rcv_rtt_measure_ts(sk, skb);
5285 /* Process urgent data. */
5286 tcp_urg(sk, skb, th);
5288 /* step 7: process the segment text */
5289 tcp_data_queue(sk, skb);
5291 tcp_data_snd_check(sk);
5292 tcp_ack_snd_check(sk);
5296 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_CSUMERRORS);
5297 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5302 EXPORT_SYMBOL(tcp_rcv_established);
5304 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5306 struct tcp_sock *tp = tcp_sk(sk);
5307 struct inet_connection_sock *icsk = inet_csk(sk);
5309 tcp_set_state(sk, TCP_ESTABLISHED);
5312 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5313 security_inet_conn_established(sk, skb);
5316 /* Make sure socket is routed, for correct metrics. */
5317 icsk->icsk_af_ops->rebuild_header(sk);
5319 tcp_init_metrics(sk);
5321 tcp_init_congestion_control(sk);
5323 /* Prevent spurious tcp_cwnd_restart() on first data
5326 tp->lsndtime = tcp_time_stamp;
5328 tcp_init_buffer_space(sk);
5330 if (sock_flag(sk, SOCK_KEEPOPEN))
5331 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5333 if (!tp->rx_opt.snd_wscale)
5334 __tcp_fast_path_on(tp, tp->snd_wnd);
5338 if (!sock_flag(sk, SOCK_DEAD)) {
5339 sk->sk_state_change(sk);
5340 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5344 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5345 struct tcp_fastopen_cookie *cookie)
5347 struct tcp_sock *tp = tcp_sk(sk);
5348 struct sk_buff *data = tp->syn_data ? tcp_write_queue_head(sk) : NULL;
5349 u16 mss = tp->rx_opt.mss_clamp;
5352 if (mss == tp->rx_opt.user_mss) {
5353 struct tcp_options_received opt;
5355 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5356 tcp_clear_options(&opt);
5357 opt.user_mss = opt.mss_clamp = 0;
5358 tcp_parse_options(synack, &opt, 0, NULL);
5359 mss = opt.mss_clamp;
5362 if (!tp->syn_fastopen) /* Ignore an unsolicited cookie */
5365 /* The SYN-ACK neither has cookie nor acknowledges the data. Presumably
5366 * the remote receives only the retransmitted (regular) SYNs: either
5367 * the original SYN-data or the corresponding SYN-ACK is lost.
5369 syn_drop = (cookie->len <= 0 && data && tp->total_retrans);
5371 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop);
5373 if (data) { /* Retransmit unacked data in SYN */
5374 tcp_for_write_queue_from(data, sk) {
5375 if (data == tcp_send_head(sk) ||
5376 __tcp_retransmit_skb(sk, data))
5382 tp->syn_data_acked = tp->syn_data;
5386 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5387 const struct tcphdr *th, unsigned int len)
5389 struct inet_connection_sock *icsk = inet_csk(sk);
5390 struct tcp_sock *tp = tcp_sk(sk);
5391 struct tcp_fastopen_cookie foc = { .len = -1 };
5392 int saved_clamp = tp->rx_opt.mss_clamp;
5394 tcp_parse_options(skb, &tp->rx_opt, 0, &foc);
5395 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
5396 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
5400 * "If the state is SYN-SENT then
5401 * first check the ACK bit
5402 * If the ACK bit is set
5403 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5404 * a reset (unless the RST bit is set, if so drop
5405 * the segment and return)"
5407 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5408 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
5409 goto reset_and_undo;
5411 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5412 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5414 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
5415 goto reset_and_undo;
5418 /* Now ACK is acceptable.
5420 * "If the RST bit is set
5421 * If the ACK was acceptable then signal the user "error:
5422 * connection reset", drop the segment, enter CLOSED state,
5423 * delete TCB, and return."
5432 * "fifth, if neither of the SYN or RST bits is set then
5433 * drop the segment and return."
5439 goto discard_and_undo;
5442 * "If the SYN bit is on ...
5443 * are acceptable then ...
5444 * (our SYN has been ACKed), change the connection
5445 * state to ESTABLISHED..."
5448 TCP_ECN_rcv_synack(tp, th);
5450 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5451 tcp_ack(sk, skb, FLAG_SLOWPATH);
5453 /* Ok.. it's good. Set up sequence numbers and
5454 * move to established.
5456 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5457 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5459 /* RFC1323: The window in SYN & SYN/ACK segments is
5462 tp->snd_wnd = ntohs(th->window);
5464 if (!tp->rx_opt.wscale_ok) {
5465 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5466 tp->window_clamp = min(tp->window_clamp, 65535U);
5469 if (tp->rx_opt.saw_tstamp) {
5470 tp->rx_opt.tstamp_ok = 1;
5471 tp->tcp_header_len =
5472 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5473 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5474 tcp_store_ts_recent(tp);
5476 tp->tcp_header_len = sizeof(struct tcphdr);
5479 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5480 tcp_enable_fack(tp);
5483 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5484 tcp_initialize_rcv_mss(sk);
5486 /* Remember, tcp_poll() does not lock socket!
5487 * Change state from SYN-SENT only after copied_seq
5488 * is initialized. */
5489 tp->copied_seq = tp->rcv_nxt;
5493 tcp_finish_connect(sk, skb);
5495 if ((tp->syn_fastopen || tp->syn_data) &&
5496 tcp_rcv_fastopen_synack(sk, skb, &foc))
5499 if (sk->sk_write_pending ||
5500 icsk->icsk_accept_queue.rskq_defer_accept ||
5501 icsk->icsk_ack.pingpong) {
5502 /* Save one ACK. Data will be ready after
5503 * several ticks, if write_pending is set.
5505 * It may be deleted, but with this feature tcpdumps
5506 * look so _wonderfully_ clever, that I was not able
5507 * to stand against the temptation 8) --ANK
5509 inet_csk_schedule_ack(sk);
5510 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5511 tcp_enter_quickack_mode(sk);
5512 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5513 TCP_DELACK_MAX, TCP_RTO_MAX);
5524 /* No ACK in the segment */
5528 * "If the RST bit is set
5530 * Otherwise (no ACK) drop the segment and return."
5533 goto discard_and_undo;
5537 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5538 tcp_paws_reject(&tp->rx_opt, 0))
5539 goto discard_and_undo;
5542 /* We see SYN without ACK. It is attempt of
5543 * simultaneous connect with crossed SYNs.
5544 * Particularly, it can be connect to self.
5546 tcp_set_state(sk, TCP_SYN_RECV);
5548 if (tp->rx_opt.saw_tstamp) {
5549 tp->rx_opt.tstamp_ok = 1;
5550 tcp_store_ts_recent(tp);
5551 tp->tcp_header_len =
5552 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5554 tp->tcp_header_len = sizeof(struct tcphdr);
5557 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5558 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5560 /* RFC1323: The window in SYN & SYN/ACK segments is
5563 tp->snd_wnd = ntohs(th->window);
5564 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5565 tp->max_window = tp->snd_wnd;
5567 TCP_ECN_rcv_syn(tp, th);
5570 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5571 tcp_initialize_rcv_mss(sk);
5573 tcp_send_synack(sk);
5575 /* Note, we could accept data and URG from this segment.
5576 * There are no obstacles to make this (except that we must
5577 * either change tcp_recvmsg() to prevent it from returning data
5578 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5580 * However, if we ignore data in ACKless segments sometimes,
5581 * we have no reasons to accept it sometimes.
5582 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5583 * is not flawless. So, discard packet for sanity.
5584 * Uncomment this return to process the data.
5591 /* "fifth, if neither of the SYN or RST bits is set then
5592 * drop the segment and return."
5596 tcp_clear_options(&tp->rx_opt);
5597 tp->rx_opt.mss_clamp = saved_clamp;
5601 tcp_clear_options(&tp->rx_opt);
5602 tp->rx_opt.mss_clamp = saved_clamp;
5607 * This function implements the receiving procedure of RFC 793 for
5608 * all states except ESTABLISHED and TIME_WAIT.
5609 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5610 * address independent.
5613 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5614 const struct tcphdr *th, unsigned int len)
5616 struct tcp_sock *tp = tcp_sk(sk);
5617 struct inet_connection_sock *icsk = inet_csk(sk);
5618 struct request_sock *req;
5622 tp->rx_opt.saw_tstamp = 0;
5624 switch (sk->sk_state) {
5638 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5641 /* Now we have several options: In theory there is
5642 * nothing else in the frame. KA9Q has an option to
5643 * send data with the syn, BSD accepts data with the
5644 * syn up to the [to be] advertised window and
5645 * Solaris 2.1 gives you a protocol error. For now
5646 * we just ignore it, that fits the spec precisely
5647 * and avoids incompatibilities. It would be nice in
5648 * future to drop through and process the data.
5650 * Now that TTCP is starting to be used we ought to
5652 * But, this leaves one open to an easy denial of
5653 * service attack, and SYN cookies can't defend
5654 * against this problem. So, we drop the data
5655 * in the interest of security over speed unless
5656 * it's still in use.
5664 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5668 /* Do step6 onward by hand. */
5669 tcp_urg(sk, skb, th);
5671 tcp_data_snd_check(sk);
5675 req = tp->fastopen_rsk;
5677 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
5678 sk->sk_state != TCP_FIN_WAIT1);
5680 if (tcp_check_req(sk, skb, req, NULL, true) == NULL)
5684 if (!th->ack && !th->rst)
5687 if (!tcp_validate_incoming(sk, skb, th, 0))
5690 /* step 5: check the ACK field */
5691 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
5692 FLAG_UPDATE_TS_RECENT) > 0;
5694 switch (sk->sk_state) {
5699 /* Once we leave TCP_SYN_RECV, we no longer need req
5703 tp->total_retrans = req->num_retrans;
5704 reqsk_fastopen_remove(sk, req, false);
5706 /* Make sure socket is routed, for correct metrics. */
5707 icsk->icsk_af_ops->rebuild_header(sk);
5708 tcp_init_congestion_control(sk);
5711 tp->copied_seq = tp->rcv_nxt;
5712 tcp_init_buffer_space(sk);
5715 tcp_set_state(sk, TCP_ESTABLISHED);
5716 sk->sk_state_change(sk);
5718 /* Note, that this wakeup is only for marginal crossed SYN case.
5719 * Passively open sockets are not waked up, because
5720 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5723 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5725 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5726 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
5727 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5728 tcp_synack_rtt_meas(sk, req);
5730 if (tp->rx_opt.tstamp_ok)
5731 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5734 /* Re-arm the timer because data may have been sent out.
5735 * This is similar to the regular data transmission case
5736 * when new data has just been ack'ed.
5738 * (TFO) - we could try to be more aggressive and
5739 * retransmitting any data sooner based on when they
5744 tcp_init_metrics(sk);
5746 /* Prevent spurious tcp_cwnd_restart() on first data packet */
5747 tp->lsndtime = tcp_time_stamp;
5749 tcp_initialize_rcv_mss(sk);
5750 tcp_fast_path_on(tp);
5753 case TCP_FIN_WAIT1: {
5754 struct dst_entry *dst;
5757 /* If we enter the TCP_FIN_WAIT1 state and we are a
5758 * Fast Open socket and this is the first acceptable
5759 * ACK we have received, this would have acknowledged
5760 * our SYNACK so stop the SYNACK timer.
5763 /* Return RST if ack_seq is invalid.
5764 * Note that RFC793 only says to generate a
5765 * DUPACK for it but for TCP Fast Open it seems
5766 * better to treat this case like TCP_SYN_RECV
5771 /* We no longer need the request sock. */
5772 reqsk_fastopen_remove(sk, req, false);
5775 if (tp->snd_una != tp->write_seq)
5778 tcp_set_state(sk, TCP_FIN_WAIT2);
5779 sk->sk_shutdown |= SEND_SHUTDOWN;
5781 dst = __sk_dst_get(sk);
5785 if (!sock_flag(sk, SOCK_DEAD)) {
5786 /* Wake up lingering close() */
5787 sk->sk_state_change(sk);
5791 if (tp->linger2 < 0 ||
5792 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5793 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5795 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5799 tmo = tcp_fin_time(sk);
5800 if (tmo > TCP_TIMEWAIT_LEN) {
5801 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5802 } else if (th->fin || sock_owned_by_user(sk)) {
5803 /* Bad case. We could lose such FIN otherwise.
5804 * It is not a big problem, but it looks confusing
5805 * and not so rare event. We still can lose it now,
5806 * if it spins in bh_lock_sock(), but it is really
5809 inet_csk_reset_keepalive_timer(sk, tmo);
5811 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5818 if (tp->snd_una == tp->write_seq) {
5819 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5825 if (tp->snd_una == tp->write_seq) {
5826 tcp_update_metrics(sk);
5833 /* step 6: check the URG bit */
5834 tcp_urg(sk, skb, th);
5836 /* step 7: process the segment text */
5837 switch (sk->sk_state) {
5838 case TCP_CLOSE_WAIT:
5841 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5845 /* RFC 793 says to queue data in these states,
5846 * RFC 1122 says we MUST send a reset.
5847 * BSD 4.4 also does reset.
5849 if (sk->sk_shutdown & RCV_SHUTDOWN) {
5850 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5851 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5852 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5858 case TCP_ESTABLISHED:
5859 tcp_data_queue(sk, skb);
5864 /* tcp_data could move socket to TIME-WAIT */
5865 if (sk->sk_state != TCP_CLOSE) {
5866 tcp_data_snd_check(sk);
5867 tcp_ack_snd_check(sk);
5876 EXPORT_SYMBOL(tcp_rcv_state_process);