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 */
115 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
116 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
117 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
118 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
120 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
121 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
123 /* Adapt the MSS value used to make delayed ack decision to the
126 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
128 struct inet_connection_sock *icsk = inet_csk(sk);
129 const unsigned int lss = icsk->icsk_ack.last_seg_size;
132 icsk->icsk_ack.last_seg_size = 0;
134 /* skb->len may jitter because of SACKs, even if peer
135 * sends good full-sized frames.
137 len = skb_shinfo(skb)->gso_size ? : skb->len;
138 if (len >= icsk->icsk_ack.rcv_mss) {
139 icsk->icsk_ack.rcv_mss = len;
141 /* Otherwise, we make more careful check taking into account,
142 * that SACKs block is variable.
144 * "len" is invariant segment length, including TCP header.
146 len += skb->data - skb_transport_header(skb);
147 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
148 /* If PSH is not set, packet should be
149 * full sized, provided peer TCP is not badly broken.
150 * This observation (if it is correct 8)) allows
151 * to handle super-low mtu links fairly.
153 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
154 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
155 /* Subtract also invariant (if peer is RFC compliant),
156 * tcp header plus fixed timestamp option length.
157 * Resulting "len" is MSS free of SACK jitter.
159 len -= tcp_sk(sk)->tcp_header_len;
160 icsk->icsk_ack.last_seg_size = len;
162 icsk->icsk_ack.rcv_mss = len;
166 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
167 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
168 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
172 static void tcp_incr_quickack(struct sock *sk)
174 struct inet_connection_sock *icsk = inet_csk(sk);
175 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
179 if (quickacks > icsk->icsk_ack.quick)
180 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
183 static void tcp_enter_quickack_mode(struct sock *sk)
185 struct inet_connection_sock *icsk = inet_csk(sk);
186 tcp_incr_quickack(sk);
187 icsk->icsk_ack.pingpong = 0;
188 icsk->icsk_ack.ato = TCP_ATO_MIN;
191 /* Send ACKs quickly, if "quick" count is not exhausted
192 * and the session is not interactive.
195 static inline bool tcp_in_quickack_mode(const struct sock *sk)
197 const struct inet_connection_sock *icsk = inet_csk(sk);
199 return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
202 static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp)
204 if (tp->ecn_flags & TCP_ECN_OK)
205 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
208 static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
210 if (tcp_hdr(skb)->cwr)
211 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
214 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp)
216 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
219 static inline void TCP_ECN_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
221 if (!(tp->ecn_flags & TCP_ECN_OK))
224 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
225 case INET_ECN_NOT_ECT:
226 /* Funny extension: if ECT is not set on a segment,
227 * and we already seen ECT on a previous segment,
228 * it is probably a retransmit.
230 if (tp->ecn_flags & TCP_ECN_SEEN)
231 tcp_enter_quickack_mode((struct sock *)tp);
234 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
235 /* Better not delay acks, sender can have a very low cwnd */
236 tcp_enter_quickack_mode((struct sock *)tp);
237 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
241 tp->ecn_flags |= TCP_ECN_SEEN;
245 static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
247 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
248 tp->ecn_flags &= ~TCP_ECN_OK;
251 static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
253 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
254 tp->ecn_flags &= ~TCP_ECN_OK;
257 static bool TCP_ECN_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
259 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
264 /* Buffer size and advertised window tuning.
266 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
269 static void tcp_fixup_sndbuf(struct sock *sk)
271 int sndmem = SKB_TRUESIZE(tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER);
273 sndmem *= TCP_INIT_CWND;
274 if (sk->sk_sndbuf < sndmem)
275 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
278 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
280 * All tcp_full_space() is split to two parts: "network" buffer, allocated
281 * forward and advertised in receiver window (tp->rcv_wnd) and
282 * "application buffer", required to isolate scheduling/application
283 * latencies from network.
284 * window_clamp is maximal advertised window. It can be less than
285 * tcp_full_space(), in this case tcp_full_space() - window_clamp
286 * is reserved for "application" buffer. The less window_clamp is
287 * the smoother our behaviour from viewpoint of network, but the lower
288 * throughput and the higher sensitivity of the connection to losses. 8)
290 * rcv_ssthresh is more strict window_clamp used at "slow start"
291 * phase to predict further behaviour of this connection.
292 * It is used for two goals:
293 * - to enforce header prediction at sender, even when application
294 * requires some significant "application buffer". It is check #1.
295 * - to prevent pruning of receive queue because of misprediction
296 * of receiver window. Check #2.
298 * The scheme does not work when sender sends good segments opening
299 * window and then starts to feed us spaghetti. But it should work
300 * in common situations. Otherwise, we have to rely on queue collapsing.
303 /* Slow part of check#2. */
304 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
306 struct tcp_sock *tp = tcp_sk(sk);
308 int truesize = tcp_win_from_space(skb->truesize) >> 1;
309 int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
311 while (tp->rcv_ssthresh <= window) {
312 if (truesize <= skb->len)
313 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
321 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
323 struct tcp_sock *tp = tcp_sk(sk);
326 if (tp->rcv_ssthresh < tp->window_clamp &&
327 (int)tp->rcv_ssthresh < tcp_space(sk) &&
328 !sk_under_memory_pressure(sk)) {
331 /* Check #2. Increase window, if skb with such overhead
332 * will fit to rcvbuf in future.
334 if (tcp_win_from_space(skb->truesize) <= skb->len)
335 incr = 2 * tp->advmss;
337 incr = __tcp_grow_window(sk, skb);
340 incr = max_t(int, incr, 2 * skb->len);
341 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
343 inet_csk(sk)->icsk_ack.quick |= 1;
348 /* 3. Tuning rcvbuf, when connection enters established state. */
350 static void tcp_fixup_rcvbuf(struct sock *sk)
352 u32 mss = tcp_sk(sk)->advmss;
353 u32 icwnd = TCP_DEFAULT_INIT_RCVWND;
356 /* Limit to 10 segments if mss <= 1460,
357 * or 14600/mss segments, with a minimum of two segments.
360 icwnd = max_t(u32, (1460 * TCP_DEFAULT_INIT_RCVWND) / mss, 2);
362 rcvmem = SKB_TRUESIZE(mss + MAX_TCP_HEADER);
363 while (tcp_win_from_space(rcvmem) < mss)
368 if (sk->sk_rcvbuf < rcvmem)
369 sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]);
372 /* 4. Try to fixup all. It is made immediately after connection enters
375 void tcp_init_buffer_space(struct sock *sk)
377 struct tcp_sock *tp = tcp_sk(sk);
380 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
381 tcp_fixup_rcvbuf(sk);
382 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
383 tcp_fixup_sndbuf(sk);
385 tp->rcvq_space.space = tp->rcv_wnd;
387 maxwin = tcp_full_space(sk);
389 if (tp->window_clamp >= maxwin) {
390 tp->window_clamp = maxwin;
392 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
393 tp->window_clamp = max(maxwin -
394 (maxwin >> sysctl_tcp_app_win),
398 /* Force reservation of one segment. */
399 if (sysctl_tcp_app_win &&
400 tp->window_clamp > 2 * tp->advmss &&
401 tp->window_clamp + tp->advmss > maxwin)
402 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
404 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
405 tp->snd_cwnd_stamp = tcp_time_stamp;
408 /* 5. Recalculate window clamp after socket hit its memory bounds. */
409 static void tcp_clamp_window(struct sock *sk)
411 struct tcp_sock *tp = tcp_sk(sk);
412 struct inet_connection_sock *icsk = inet_csk(sk);
414 icsk->icsk_ack.quick = 0;
416 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
417 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
418 !sk_under_memory_pressure(sk) &&
419 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
420 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
423 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
424 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
427 /* Initialize RCV_MSS value.
428 * RCV_MSS is an our guess about MSS used by the peer.
429 * We haven't any direct information about the MSS.
430 * It's better to underestimate the RCV_MSS rather than overestimate.
431 * Overestimations make us ACKing less frequently than needed.
432 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
434 void tcp_initialize_rcv_mss(struct sock *sk)
436 const struct tcp_sock *tp = tcp_sk(sk);
437 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
439 hint = min(hint, tp->rcv_wnd / 2);
440 hint = min(hint, TCP_MSS_DEFAULT);
441 hint = max(hint, TCP_MIN_MSS);
443 inet_csk(sk)->icsk_ack.rcv_mss = hint;
445 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
447 /* Receiver "autotuning" code.
449 * The algorithm for RTT estimation w/o timestamps is based on
450 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
451 * <http://public.lanl.gov/radiant/pubs.html#DRS>
453 * More detail on this code can be found at
454 * <http://staff.psc.edu/jheffner/>,
455 * though this reference is out of date. A new paper
458 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
460 u32 new_sample = tp->rcv_rtt_est.rtt;
466 if (new_sample != 0) {
467 /* If we sample in larger samples in the non-timestamp
468 * case, we could grossly overestimate the RTT especially
469 * with chatty applications or bulk transfer apps which
470 * are stalled on filesystem I/O.
472 * Also, since we are only going for a minimum in the
473 * non-timestamp case, we do not smooth things out
474 * else with timestamps disabled convergence takes too
478 m -= (new_sample >> 3);
486 /* No previous measure. */
490 if (tp->rcv_rtt_est.rtt != new_sample)
491 tp->rcv_rtt_est.rtt = new_sample;
494 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
496 if (tp->rcv_rtt_est.time == 0)
498 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
500 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rcv_rtt_est.time, 1);
503 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
504 tp->rcv_rtt_est.time = tcp_time_stamp;
507 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
508 const struct sk_buff *skb)
510 struct tcp_sock *tp = tcp_sk(sk);
511 if (tp->rx_opt.rcv_tsecr &&
512 (TCP_SKB_CB(skb)->end_seq -
513 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
514 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
518 * This function should be called every time data is copied to user space.
519 * It calculates the appropriate TCP receive buffer space.
521 void tcp_rcv_space_adjust(struct sock *sk)
523 struct tcp_sock *tp = tcp_sk(sk);
527 if (tp->rcvq_space.time == 0)
530 time = tcp_time_stamp - tp->rcvq_space.time;
531 if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
534 space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
536 space = max(tp->rcvq_space.space, space);
538 if (tp->rcvq_space.space != space) {
541 tp->rcvq_space.space = space;
543 if (sysctl_tcp_moderate_rcvbuf &&
544 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
545 int new_clamp = space;
547 /* Receive space grows, normalize in order to
548 * take into account packet headers and sk_buff
549 * structure overhead.
554 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
555 while (tcp_win_from_space(rcvmem) < tp->advmss)
558 space = min(space, sysctl_tcp_rmem[2]);
559 if (space > sk->sk_rcvbuf) {
560 sk->sk_rcvbuf = space;
562 /* Make the window clamp follow along. */
563 tp->window_clamp = new_clamp;
569 tp->rcvq_space.seq = tp->copied_seq;
570 tp->rcvq_space.time = tcp_time_stamp;
573 /* There is something which you must keep in mind when you analyze the
574 * behavior of the tp->ato delayed ack timeout interval. When a
575 * connection starts up, we want to ack as quickly as possible. The
576 * problem is that "good" TCP's do slow start at the beginning of data
577 * transmission. The means that until we send the first few ACK's the
578 * sender will sit on his end and only queue most of his data, because
579 * he can only send snd_cwnd unacked packets at any given time. For
580 * each ACK we send, he increments snd_cwnd and transmits more of his
583 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
585 struct tcp_sock *tp = tcp_sk(sk);
586 struct inet_connection_sock *icsk = inet_csk(sk);
589 inet_csk_schedule_ack(sk);
591 tcp_measure_rcv_mss(sk, skb);
593 tcp_rcv_rtt_measure(tp);
595 now = tcp_time_stamp;
597 if (!icsk->icsk_ack.ato) {
598 /* The _first_ data packet received, initialize
599 * delayed ACK engine.
601 tcp_incr_quickack(sk);
602 icsk->icsk_ack.ato = TCP_ATO_MIN;
604 int m = now - icsk->icsk_ack.lrcvtime;
606 if (m <= TCP_ATO_MIN / 2) {
607 /* The fastest case is the first. */
608 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
609 } else if (m < icsk->icsk_ack.ato) {
610 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
611 if (icsk->icsk_ack.ato > icsk->icsk_rto)
612 icsk->icsk_ack.ato = icsk->icsk_rto;
613 } else if (m > icsk->icsk_rto) {
614 /* Too long gap. Apparently sender failed to
615 * restart window, so that we send ACKs quickly.
617 tcp_incr_quickack(sk);
621 icsk->icsk_ack.lrcvtime = now;
623 TCP_ECN_check_ce(tp, skb);
626 tcp_grow_window(sk, skb);
629 /* Called to compute a smoothed rtt estimate. The data fed to this
630 * routine either comes from timestamps, or from segments that were
631 * known _not_ to have been retransmitted [see Karn/Partridge
632 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
633 * piece by Van Jacobson.
634 * NOTE: the next three routines used to be one big routine.
635 * To save cycles in the RFC 1323 implementation it was better to break
636 * it up into three procedures. -- erics
638 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
640 struct tcp_sock *tp = tcp_sk(sk);
641 long m = mrtt; /* RTT */
643 /* The following amusing code comes from Jacobson's
644 * article in SIGCOMM '88. Note that rtt and mdev
645 * are scaled versions of rtt and mean deviation.
646 * This is designed to be as fast as possible
647 * m stands for "measurement".
649 * On a 1990 paper the rto value is changed to:
650 * RTO = rtt + 4 * mdev
652 * Funny. This algorithm seems to be very broken.
653 * These formulae increase RTO, when it should be decreased, increase
654 * too slowly, when it should be increased quickly, decrease too quickly
655 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
656 * does not matter how to _calculate_ it. Seems, it was trap
657 * that VJ failed to avoid. 8)
662 m -= (tp->srtt >> 3); /* m is now error in rtt est */
663 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */
665 m = -m; /* m is now abs(error) */
666 m -= (tp->mdev >> 2); /* similar update on mdev */
667 /* This is similar to one of Eifel findings.
668 * Eifel blocks mdev updates when rtt decreases.
669 * This solution is a bit different: we use finer gain
670 * for mdev in this case (alpha*beta).
671 * Like Eifel it also prevents growth of rto,
672 * but also it limits too fast rto decreases,
673 * happening in pure Eifel.
678 m -= (tp->mdev >> 2); /* similar update on mdev */
680 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */
681 if (tp->mdev > tp->mdev_max) {
682 tp->mdev_max = tp->mdev;
683 if (tp->mdev_max > tp->rttvar)
684 tp->rttvar = tp->mdev_max;
686 if (after(tp->snd_una, tp->rtt_seq)) {
687 if (tp->mdev_max < tp->rttvar)
688 tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2;
689 tp->rtt_seq = tp->snd_nxt;
690 tp->mdev_max = tcp_rto_min(sk);
693 /* no previous measure. */
694 tp->srtt = m << 3; /* take the measured time to be rtt */
695 tp->mdev = m << 1; /* make sure rto = 3*rtt */
696 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
697 tp->rtt_seq = tp->snd_nxt;
701 /* Calculate rto without backoff. This is the second half of Van Jacobson's
702 * routine referred to above.
704 void tcp_set_rto(struct sock *sk)
706 const struct tcp_sock *tp = tcp_sk(sk);
707 /* Old crap is replaced with new one. 8)
710 * 1. If rtt variance happened to be less 50msec, it is hallucination.
711 * It cannot be less due to utterly erratic ACK generation made
712 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
713 * to do with delayed acks, because at cwnd>2 true delack timeout
714 * is invisible. Actually, Linux-2.4 also generates erratic
715 * ACKs in some circumstances.
717 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
719 /* 2. Fixups made earlier cannot be right.
720 * If we do not estimate RTO correctly without them,
721 * all the algo is pure shit and should be replaced
722 * with correct one. It is exactly, which we pretend to do.
725 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
726 * guarantees that rto is higher.
731 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
733 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
736 cwnd = TCP_INIT_CWND;
737 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
741 * Packet counting of FACK is based on in-order assumptions, therefore TCP
742 * disables it when reordering is detected
744 void tcp_disable_fack(struct tcp_sock *tp)
746 /* RFC3517 uses different metric in lost marker => reset on change */
748 tp->lost_skb_hint = NULL;
749 tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED;
752 /* Take a notice that peer is sending D-SACKs */
753 static void tcp_dsack_seen(struct tcp_sock *tp)
755 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
758 static void tcp_update_reordering(struct sock *sk, const int metric,
761 struct tcp_sock *tp = tcp_sk(sk);
762 if (metric > tp->reordering) {
765 tp->reordering = min(TCP_MAX_REORDERING, metric);
767 /* This exciting event is worth to be remembered. 8) */
769 mib_idx = LINUX_MIB_TCPTSREORDER;
770 else if (tcp_is_reno(tp))
771 mib_idx = LINUX_MIB_TCPRENOREORDER;
772 else if (tcp_is_fack(tp))
773 mib_idx = LINUX_MIB_TCPFACKREORDER;
775 mib_idx = LINUX_MIB_TCPSACKREORDER;
777 NET_INC_STATS_BH(sock_net(sk), mib_idx);
778 #if FASTRETRANS_DEBUG > 1
779 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
780 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
784 tp->undo_marker ? tp->undo_retrans : 0);
786 tcp_disable_fack(tp);
790 tcp_disable_early_retrans(tp);
793 /* This must be called before lost_out is incremented */
794 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
796 if ((tp->retransmit_skb_hint == NULL) ||
797 before(TCP_SKB_CB(skb)->seq,
798 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
799 tp->retransmit_skb_hint = skb;
802 after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
803 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
806 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
808 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
809 tcp_verify_retransmit_hint(tp, skb);
811 tp->lost_out += tcp_skb_pcount(skb);
812 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
816 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp,
819 tcp_verify_retransmit_hint(tp, skb);
821 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
822 tp->lost_out += tcp_skb_pcount(skb);
823 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
827 /* This procedure tags the retransmission queue when SACKs arrive.
829 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
830 * Packets in queue with these bits set are counted in variables
831 * sacked_out, retrans_out and lost_out, correspondingly.
833 * Valid combinations are:
834 * Tag InFlight Description
835 * 0 1 - orig segment is in flight.
836 * S 0 - nothing flies, orig reached receiver.
837 * L 0 - nothing flies, orig lost by net.
838 * R 2 - both orig and retransmit are in flight.
839 * L|R 1 - orig is lost, retransmit is in flight.
840 * S|R 1 - orig reached receiver, retrans is still in flight.
841 * (L|S|R is logically valid, it could occur when L|R is sacked,
842 * but it is equivalent to plain S and code short-curcuits it to S.
843 * L|S is logically invalid, it would mean -1 packet in flight 8))
845 * These 6 states form finite state machine, controlled by the following events:
846 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
847 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
848 * 3. Loss detection event of two flavors:
849 * A. Scoreboard estimator decided the packet is lost.
850 * A'. Reno "three dupacks" marks head of queue lost.
851 * A''. Its FACK modification, head until snd.fack is lost.
852 * B. SACK arrives sacking SND.NXT at the moment, when the
853 * segment was retransmitted.
854 * 4. D-SACK added new rule: D-SACK changes any tag to S.
856 * It is pleasant to note, that state diagram turns out to be commutative,
857 * so that we are allowed not to be bothered by order of our actions,
858 * when multiple events arrive simultaneously. (see the function below).
860 * Reordering detection.
861 * --------------------
862 * Reordering metric is maximal distance, which a packet can be displaced
863 * in packet stream. With SACKs we can estimate it:
865 * 1. SACK fills old hole and the corresponding segment was not
866 * ever retransmitted -> reordering. Alas, we cannot use it
867 * when segment was retransmitted.
868 * 2. The last flaw is solved with D-SACK. D-SACK arrives
869 * for retransmitted and already SACKed segment -> reordering..
870 * Both of these heuristics are not used in Loss state, when we cannot
871 * account for retransmits accurately.
873 * SACK block validation.
874 * ----------------------
876 * SACK block range validation checks that the received SACK block fits to
877 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
878 * Note that SND.UNA is not included to the range though being valid because
879 * it means that the receiver is rather inconsistent with itself reporting
880 * SACK reneging when it should advance SND.UNA. Such SACK block this is
881 * perfectly valid, however, in light of RFC2018 which explicitly states
882 * that "SACK block MUST reflect the newest segment. Even if the newest
883 * segment is going to be discarded ...", not that it looks very clever
884 * in case of head skb. Due to potentional receiver driven attacks, we
885 * choose to avoid immediate execution of a walk in write queue due to
886 * reneging and defer head skb's loss recovery to standard loss recovery
887 * procedure that will eventually trigger (nothing forbids us doing this).
889 * Implements also blockage to start_seq wrap-around. Problem lies in the
890 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
891 * there's no guarantee that it will be before snd_nxt (n). The problem
892 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
895 * <- outs wnd -> <- wrapzone ->
896 * u e n u_w e_w s n_w
898 * |<------------+------+----- TCP seqno space --------------+---------->|
899 * ...-- <2^31 ->| |<--------...
900 * ...---- >2^31 ------>| |<--------...
902 * Current code wouldn't be vulnerable but it's better still to discard such
903 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
904 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
905 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
906 * equal to the ideal case (infinite seqno space without wrap caused issues).
908 * With D-SACK the lower bound is extended to cover sequence space below
909 * SND.UNA down to undo_marker, which is the last point of interest. Yet
910 * again, D-SACK block must not to go across snd_una (for the same reason as
911 * for the normal SACK blocks, explained above). But there all simplicity
912 * ends, TCP might receive valid D-SACKs below that. As long as they reside
913 * fully below undo_marker they do not affect behavior in anyway and can
914 * therefore be safely ignored. In rare cases (which are more or less
915 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
916 * fragmentation and packet reordering past skb's retransmission. To consider
917 * them correctly, the acceptable range must be extended even more though
918 * the exact amount is rather hard to quantify. However, tp->max_window can
919 * be used as an exaggerated estimate.
921 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
922 u32 start_seq, u32 end_seq)
924 /* Too far in future, or reversed (interpretation is ambiguous) */
925 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
928 /* Nasty start_seq wrap-around check (see comments above) */
929 if (!before(start_seq, tp->snd_nxt))
932 /* In outstanding window? ...This is valid exit for D-SACKs too.
933 * start_seq == snd_una is non-sensical (see comments above)
935 if (after(start_seq, tp->snd_una))
938 if (!is_dsack || !tp->undo_marker)
941 /* ...Then it's D-SACK, and must reside below snd_una completely */
942 if (after(end_seq, tp->snd_una))
945 if (!before(start_seq, tp->undo_marker))
949 if (!after(end_seq, tp->undo_marker))
952 /* Undo_marker boundary crossing (overestimates a lot). Known already:
953 * start_seq < undo_marker and end_seq >= undo_marker.
955 return !before(start_seq, end_seq - tp->max_window);
958 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
959 * Event "B". Later note: FACK people cheated me again 8), we have to account
960 * for reordering! Ugly, but should help.
962 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
963 * less than what is now known to be received by the other end (derived from
964 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
965 * retransmitted skbs to avoid some costly processing per ACKs.
967 static void tcp_mark_lost_retrans(struct sock *sk)
969 const struct inet_connection_sock *icsk = inet_csk(sk);
970 struct tcp_sock *tp = tcp_sk(sk);
973 u32 new_low_seq = tp->snd_nxt;
974 u32 received_upto = tcp_highest_sack_seq(tp);
976 if (!tcp_is_fack(tp) || !tp->retrans_out ||
977 !after(received_upto, tp->lost_retrans_low) ||
978 icsk->icsk_ca_state != TCP_CA_Recovery)
981 tcp_for_write_queue(skb, sk) {
982 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
984 if (skb == tcp_send_head(sk))
986 if (cnt == tp->retrans_out)
988 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
991 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
994 /* TODO: We would like to get rid of tcp_is_fack(tp) only
995 * constraint here (see above) but figuring out that at
996 * least tp->reordering SACK blocks reside between ack_seq
997 * and received_upto is not easy task to do cheaply with
998 * the available datastructures.
1000 * Whether FACK should check here for tp->reordering segs
1001 * in-between one could argue for either way (it would be
1002 * rather simple to implement as we could count fack_count
1003 * during the walk and do tp->fackets_out - fack_count).
1005 if (after(received_upto, ack_seq)) {
1006 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1007 tp->retrans_out -= tcp_skb_pcount(skb);
1009 tcp_skb_mark_lost_uncond_verify(tp, skb);
1010 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
1012 if (before(ack_seq, new_low_seq))
1013 new_low_seq = ack_seq;
1014 cnt += tcp_skb_pcount(skb);
1018 if (tp->retrans_out)
1019 tp->lost_retrans_low = new_low_seq;
1022 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1023 struct tcp_sack_block_wire *sp, int num_sacks,
1026 struct tcp_sock *tp = tcp_sk(sk);
1027 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1028 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1029 bool dup_sack = false;
1031 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1034 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1035 } else if (num_sacks > 1) {
1036 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1037 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1039 if (!after(end_seq_0, end_seq_1) &&
1040 !before(start_seq_0, start_seq_1)) {
1043 NET_INC_STATS_BH(sock_net(sk),
1044 LINUX_MIB_TCPDSACKOFORECV);
1048 /* D-SACK for already forgotten data... Do dumb counting. */
1049 if (dup_sack && tp->undo_marker && tp->undo_retrans &&
1050 !after(end_seq_0, prior_snd_una) &&
1051 after(end_seq_0, tp->undo_marker))
1057 struct tcp_sacktag_state {
1063 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1064 * the incoming SACK may not exactly match but we can find smaller MSS
1065 * aligned portion of it that matches. Therefore we might need to fragment
1066 * which may fail and creates some hassle (caller must handle error case
1069 * FIXME: this could be merged to shift decision code
1071 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1072 u32 start_seq, u32 end_seq)
1076 unsigned int pkt_len;
1079 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1080 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1082 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1083 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1084 mss = tcp_skb_mss(skb);
1085 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1088 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1092 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1097 /* Round if necessary so that SACKs cover only full MSSes
1098 * and/or the remaining small portion (if present)
1100 if (pkt_len > mss) {
1101 unsigned int new_len = (pkt_len / mss) * mss;
1102 if (!in_sack && new_len < pkt_len) {
1104 if (new_len > skb->len)
1109 err = tcp_fragment(sk, skb, pkt_len, mss);
1117 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1118 static u8 tcp_sacktag_one(struct sock *sk,
1119 struct tcp_sacktag_state *state, u8 sacked,
1120 u32 start_seq, u32 end_seq,
1121 bool dup_sack, int pcount)
1123 struct tcp_sock *tp = tcp_sk(sk);
1124 int fack_count = state->fack_count;
1126 /* Account D-SACK for retransmitted packet. */
1127 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1128 if (tp->undo_marker && tp->undo_retrans &&
1129 after(end_seq, tp->undo_marker))
1131 if (sacked & TCPCB_SACKED_ACKED)
1132 state->reord = min(fack_count, state->reord);
1135 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1136 if (!after(end_seq, tp->snd_una))
1139 if (!(sacked & TCPCB_SACKED_ACKED)) {
1140 if (sacked & TCPCB_SACKED_RETRANS) {
1141 /* If the segment is not tagged as lost,
1142 * we do not clear RETRANS, believing
1143 * that retransmission is still in flight.
1145 if (sacked & TCPCB_LOST) {
1146 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1147 tp->lost_out -= pcount;
1148 tp->retrans_out -= pcount;
1151 if (!(sacked & TCPCB_RETRANS)) {
1152 /* New sack for not retransmitted frame,
1153 * which was in hole. It is reordering.
1155 if (before(start_seq,
1156 tcp_highest_sack_seq(tp)))
1157 state->reord = min(fack_count,
1159 if (!after(end_seq, tp->high_seq))
1160 state->flag |= FLAG_ORIG_SACK_ACKED;
1163 if (sacked & TCPCB_LOST) {
1164 sacked &= ~TCPCB_LOST;
1165 tp->lost_out -= pcount;
1169 sacked |= TCPCB_SACKED_ACKED;
1170 state->flag |= FLAG_DATA_SACKED;
1171 tp->sacked_out += pcount;
1173 fack_count += pcount;
1175 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1176 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
1177 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1178 tp->lost_cnt_hint += pcount;
1180 if (fack_count > tp->fackets_out)
1181 tp->fackets_out = fack_count;
1184 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1185 * frames and clear it. undo_retrans is decreased above, L|R frames
1186 * are accounted above as well.
1188 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1189 sacked &= ~TCPCB_SACKED_RETRANS;
1190 tp->retrans_out -= pcount;
1196 /* Shift newly-SACKed bytes from this skb to the immediately previous
1197 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1199 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1200 struct tcp_sacktag_state *state,
1201 unsigned int pcount, int shifted, int mss,
1204 struct tcp_sock *tp = tcp_sk(sk);
1205 struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1206 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1207 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1211 /* Adjust counters and hints for the newly sacked sequence
1212 * range but discard the return value since prev is already
1213 * marked. We must tag the range first because the seq
1214 * advancement below implicitly advances
1215 * tcp_highest_sack_seq() when skb is highest_sack.
1217 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1218 start_seq, end_seq, dup_sack, pcount);
1220 if (skb == tp->lost_skb_hint)
1221 tp->lost_cnt_hint += pcount;
1223 TCP_SKB_CB(prev)->end_seq += shifted;
1224 TCP_SKB_CB(skb)->seq += shifted;
1226 skb_shinfo(prev)->gso_segs += pcount;
1227 BUG_ON(skb_shinfo(skb)->gso_segs < pcount);
1228 skb_shinfo(skb)->gso_segs -= pcount;
1230 /* When we're adding to gso_segs == 1, gso_size will be zero,
1231 * in theory this shouldn't be necessary but as long as DSACK
1232 * code can come after this skb later on it's better to keep
1233 * setting gso_size to something.
1235 if (!skb_shinfo(prev)->gso_size) {
1236 skb_shinfo(prev)->gso_size = mss;
1237 skb_shinfo(prev)->gso_type = sk->sk_gso_type;
1240 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1241 if (skb_shinfo(skb)->gso_segs <= 1) {
1242 skb_shinfo(skb)->gso_size = 0;
1243 skb_shinfo(skb)->gso_type = 0;
1246 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1247 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1250 BUG_ON(!tcp_skb_pcount(skb));
1251 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1255 /* Whole SKB was eaten :-) */
1257 if (skb == tp->retransmit_skb_hint)
1258 tp->retransmit_skb_hint = prev;
1259 if (skb == tp->scoreboard_skb_hint)
1260 tp->scoreboard_skb_hint = prev;
1261 if (skb == tp->lost_skb_hint) {
1262 tp->lost_skb_hint = prev;
1263 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1266 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(prev)->tcp_flags;
1267 if (skb == tcp_highest_sack(sk))
1268 tcp_advance_highest_sack(sk, skb);
1270 tcp_unlink_write_queue(skb, sk);
1271 sk_wmem_free_skb(sk, skb);
1273 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1278 /* I wish gso_size would have a bit more sane initialization than
1279 * something-or-zero which complicates things
1281 static int tcp_skb_seglen(const struct sk_buff *skb)
1283 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1286 /* Shifting pages past head area doesn't work */
1287 static int skb_can_shift(const struct sk_buff *skb)
1289 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1292 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1295 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1296 struct tcp_sacktag_state *state,
1297 u32 start_seq, u32 end_seq,
1300 struct tcp_sock *tp = tcp_sk(sk);
1301 struct sk_buff *prev;
1307 if (!sk_can_gso(sk))
1310 /* Normally R but no L won't result in plain S */
1312 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1314 if (!skb_can_shift(skb))
1316 /* This frame is about to be dropped (was ACKed). */
1317 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1320 /* Can only happen with delayed DSACK + discard craziness */
1321 if (unlikely(skb == tcp_write_queue_head(sk)))
1323 prev = tcp_write_queue_prev(sk, skb);
1325 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1328 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1329 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1333 pcount = tcp_skb_pcount(skb);
1334 mss = tcp_skb_seglen(skb);
1336 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1337 * drop this restriction as unnecessary
1339 if (mss != tcp_skb_seglen(prev))
1342 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1344 /* CHECKME: This is non-MSS split case only?, this will
1345 * cause skipped skbs due to advancing loop btw, original
1346 * has that feature too
1348 if (tcp_skb_pcount(skb) <= 1)
1351 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1353 /* TODO: head merge to next could be attempted here
1354 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1355 * though it might not be worth of the additional hassle
1357 * ...we can probably just fallback to what was done
1358 * previously. We could try merging non-SACKed ones
1359 * as well but it probably isn't going to buy off
1360 * because later SACKs might again split them, and
1361 * it would make skb timestamp tracking considerably
1367 len = end_seq - TCP_SKB_CB(skb)->seq;
1369 BUG_ON(len > skb->len);
1371 /* MSS boundaries should be honoured or else pcount will
1372 * severely break even though it makes things bit trickier.
1373 * Optimize common case to avoid most of the divides
1375 mss = tcp_skb_mss(skb);
1377 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1378 * drop this restriction as unnecessary
1380 if (mss != tcp_skb_seglen(prev))
1385 } else if (len < mss) {
1393 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1394 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1397 if (!skb_shift(prev, skb, len))
1399 if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1402 /* Hole filled allows collapsing with the next as well, this is very
1403 * useful when hole on every nth skb pattern happens
1405 if (prev == tcp_write_queue_tail(sk))
1407 skb = tcp_write_queue_next(sk, prev);
1409 if (!skb_can_shift(skb) ||
1410 (skb == tcp_send_head(sk)) ||
1411 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1412 (mss != tcp_skb_seglen(skb)))
1416 if (skb_shift(prev, skb, len)) {
1417 pcount += tcp_skb_pcount(skb);
1418 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1422 state->fack_count += pcount;
1429 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1433 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1434 struct tcp_sack_block *next_dup,
1435 struct tcp_sacktag_state *state,
1436 u32 start_seq, u32 end_seq,
1439 struct tcp_sock *tp = tcp_sk(sk);
1440 struct sk_buff *tmp;
1442 tcp_for_write_queue_from(skb, sk) {
1444 bool dup_sack = dup_sack_in;
1446 if (skb == tcp_send_head(sk))
1449 /* queue is in-order => we can short-circuit the walk early */
1450 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1453 if ((next_dup != NULL) &&
1454 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1455 in_sack = tcp_match_skb_to_sack(sk, skb,
1456 next_dup->start_seq,
1462 /* skb reference here is a bit tricky to get right, since
1463 * shifting can eat and free both this skb and the next,
1464 * so not even _safe variant of the loop is enough.
1467 tmp = tcp_shift_skb_data(sk, skb, state,
1468 start_seq, end_seq, dup_sack);
1477 in_sack = tcp_match_skb_to_sack(sk, skb,
1483 if (unlikely(in_sack < 0))
1487 TCP_SKB_CB(skb)->sacked =
1490 TCP_SKB_CB(skb)->sacked,
1491 TCP_SKB_CB(skb)->seq,
1492 TCP_SKB_CB(skb)->end_seq,
1494 tcp_skb_pcount(skb));
1496 if (!before(TCP_SKB_CB(skb)->seq,
1497 tcp_highest_sack_seq(tp)))
1498 tcp_advance_highest_sack(sk, skb);
1501 state->fack_count += tcp_skb_pcount(skb);
1506 /* Avoid all extra work that is being done by sacktag while walking in
1509 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1510 struct tcp_sacktag_state *state,
1513 tcp_for_write_queue_from(skb, sk) {
1514 if (skb == tcp_send_head(sk))
1517 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1520 state->fack_count += tcp_skb_pcount(skb);
1525 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1527 struct tcp_sack_block *next_dup,
1528 struct tcp_sacktag_state *state,
1531 if (next_dup == NULL)
1534 if (before(next_dup->start_seq, skip_to_seq)) {
1535 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1536 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1537 next_dup->start_seq, next_dup->end_seq,
1544 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1546 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1550 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1553 struct tcp_sock *tp = tcp_sk(sk);
1554 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1555 TCP_SKB_CB(ack_skb)->sacked);
1556 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1557 struct tcp_sack_block sp[TCP_NUM_SACKS];
1558 struct tcp_sack_block *cache;
1559 struct tcp_sacktag_state state;
1560 struct sk_buff *skb;
1561 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1563 bool found_dup_sack = false;
1565 int first_sack_index;
1568 state.reord = tp->packets_out;
1570 if (!tp->sacked_out) {
1571 if (WARN_ON(tp->fackets_out))
1572 tp->fackets_out = 0;
1573 tcp_highest_sack_reset(sk);
1576 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1577 num_sacks, prior_snd_una);
1579 state.flag |= FLAG_DSACKING_ACK;
1581 /* Eliminate too old ACKs, but take into
1582 * account more or less fresh ones, they can
1583 * contain valid SACK info.
1585 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1588 if (!tp->packets_out)
1592 first_sack_index = 0;
1593 for (i = 0; i < num_sacks; i++) {
1594 bool dup_sack = !i && found_dup_sack;
1596 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1597 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1599 if (!tcp_is_sackblock_valid(tp, dup_sack,
1600 sp[used_sacks].start_seq,
1601 sp[used_sacks].end_seq)) {
1605 if (!tp->undo_marker)
1606 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1608 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1610 /* Don't count olds caused by ACK reordering */
1611 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1612 !after(sp[used_sacks].end_seq, tp->snd_una))
1614 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1617 NET_INC_STATS_BH(sock_net(sk), mib_idx);
1619 first_sack_index = -1;
1623 /* Ignore very old stuff early */
1624 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1630 /* order SACK blocks to allow in order walk of the retrans queue */
1631 for (i = used_sacks - 1; i > 0; i--) {
1632 for (j = 0; j < i; j++) {
1633 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1634 swap(sp[j], sp[j + 1]);
1636 /* Track where the first SACK block goes to */
1637 if (j == first_sack_index)
1638 first_sack_index = j + 1;
1643 skb = tcp_write_queue_head(sk);
1644 state.fack_count = 0;
1647 if (!tp->sacked_out) {
1648 /* It's already past, so skip checking against it */
1649 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1651 cache = tp->recv_sack_cache;
1652 /* Skip empty blocks in at head of the cache */
1653 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1658 while (i < used_sacks) {
1659 u32 start_seq = sp[i].start_seq;
1660 u32 end_seq = sp[i].end_seq;
1661 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1662 struct tcp_sack_block *next_dup = NULL;
1664 if (found_dup_sack && ((i + 1) == first_sack_index))
1665 next_dup = &sp[i + 1];
1667 /* Skip too early cached blocks */
1668 while (tcp_sack_cache_ok(tp, cache) &&
1669 !before(start_seq, cache->end_seq))
1672 /* Can skip some work by looking recv_sack_cache? */
1673 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1674 after(end_seq, cache->start_seq)) {
1677 if (before(start_seq, cache->start_seq)) {
1678 skb = tcp_sacktag_skip(skb, sk, &state,
1680 skb = tcp_sacktag_walk(skb, sk, next_dup,
1687 /* Rest of the block already fully processed? */
1688 if (!after(end_seq, cache->end_seq))
1691 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1695 /* ...tail remains todo... */
1696 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1697 /* ...but better entrypoint exists! */
1698 skb = tcp_highest_sack(sk);
1701 state.fack_count = tp->fackets_out;
1706 skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq);
1707 /* Check overlap against next cached too (past this one already) */
1712 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1713 skb = tcp_highest_sack(sk);
1716 state.fack_count = tp->fackets_out;
1718 skb = tcp_sacktag_skip(skb, sk, &state, start_seq);
1721 skb = tcp_sacktag_walk(skb, sk, next_dup, &state,
1722 start_seq, end_seq, dup_sack);
1728 /* Clear the head of the cache sack blocks so we can skip it next time */
1729 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1730 tp->recv_sack_cache[i].start_seq = 0;
1731 tp->recv_sack_cache[i].end_seq = 0;
1733 for (j = 0; j < used_sacks; j++)
1734 tp->recv_sack_cache[i++] = sp[j];
1736 tcp_mark_lost_retrans(sk);
1738 tcp_verify_left_out(tp);
1740 if ((state.reord < tp->fackets_out) &&
1741 ((inet_csk(sk)->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker))
1742 tcp_update_reordering(sk, tp->fackets_out - state.reord, 0);
1746 #if FASTRETRANS_DEBUG > 0
1747 WARN_ON((int)tp->sacked_out < 0);
1748 WARN_ON((int)tp->lost_out < 0);
1749 WARN_ON((int)tp->retrans_out < 0);
1750 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1755 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1756 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1758 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1762 holes = max(tp->lost_out, 1U);
1763 holes = min(holes, tp->packets_out);
1765 if ((tp->sacked_out + holes) > tp->packets_out) {
1766 tp->sacked_out = tp->packets_out - holes;
1772 /* If we receive more dupacks than we expected counting segments
1773 * in assumption of absent reordering, interpret this as reordering.
1774 * The only another reason could be bug in receiver TCP.
1776 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1778 struct tcp_sock *tp = tcp_sk(sk);
1779 if (tcp_limit_reno_sacked(tp))
1780 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1783 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1785 static void tcp_add_reno_sack(struct sock *sk)
1787 struct tcp_sock *tp = tcp_sk(sk);
1789 tcp_check_reno_reordering(sk, 0);
1790 tcp_verify_left_out(tp);
1793 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1795 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1797 struct tcp_sock *tp = tcp_sk(sk);
1800 /* One ACK acked hole. The rest eat duplicate ACKs. */
1801 if (acked - 1 >= tp->sacked_out)
1804 tp->sacked_out -= acked - 1;
1806 tcp_check_reno_reordering(sk, acked);
1807 tcp_verify_left_out(tp);
1810 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1815 static void tcp_clear_retrans_partial(struct tcp_sock *tp)
1817 tp->retrans_out = 0;
1820 tp->undo_marker = 0;
1821 tp->undo_retrans = 0;
1824 void tcp_clear_retrans(struct tcp_sock *tp)
1826 tcp_clear_retrans_partial(tp);
1828 tp->fackets_out = 0;
1832 /* Enter Loss state. If "how" is not zero, forget all SACK information
1833 * and reset tags completely, otherwise preserve SACKs. If receiver
1834 * dropped its ofo queue, we will know this due to reneging detection.
1836 void tcp_enter_loss(struct sock *sk, int how)
1838 const struct inet_connection_sock *icsk = inet_csk(sk);
1839 struct tcp_sock *tp = tcp_sk(sk);
1840 struct sk_buff *skb;
1841 bool new_recovery = false;
1843 /* Reduce ssthresh if it has not yet been made inside this window. */
1844 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1845 !after(tp->high_seq, tp->snd_una) ||
1846 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1847 new_recovery = true;
1848 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1849 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1850 tcp_ca_event(sk, CA_EVENT_LOSS);
1853 tp->snd_cwnd_cnt = 0;
1854 tp->snd_cwnd_stamp = tcp_time_stamp;
1856 tcp_clear_retrans_partial(tp);
1858 if (tcp_is_reno(tp))
1859 tcp_reset_reno_sack(tp);
1862 /* Push undo marker, if it was plain RTO and nothing
1863 * was retransmitted. */
1864 tp->undo_marker = tp->snd_una;
1867 tp->fackets_out = 0;
1869 tcp_clear_all_retrans_hints(tp);
1871 tcp_for_write_queue(skb, sk) {
1872 if (skb == tcp_send_head(sk))
1875 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1876 tp->undo_marker = 0;
1877 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1878 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
1879 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1880 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1881 tp->lost_out += tcp_skb_pcount(skb);
1882 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
1885 tcp_verify_left_out(tp);
1887 tp->reordering = min_t(unsigned int, tp->reordering,
1888 sysctl_tcp_reordering);
1889 tcp_set_ca_state(sk, TCP_CA_Loss);
1890 tp->high_seq = tp->snd_nxt;
1891 TCP_ECN_queue_cwr(tp);
1893 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1894 * loss recovery is underway except recurring timeout(s) on
1895 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1897 tp->frto = sysctl_tcp_frto &&
1898 (new_recovery || icsk->icsk_retransmits) &&
1899 !inet_csk(sk)->icsk_mtup.probe_size;
1902 /* If ACK arrived pointing to a remembered SACK, it means that our
1903 * remembered SACKs do not reflect real state of receiver i.e.
1904 * receiver _host_ is heavily congested (or buggy).
1906 * Do processing similar to RTO timeout.
1908 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
1910 if (flag & FLAG_SACK_RENEGING) {
1911 struct inet_connection_sock *icsk = inet_csk(sk);
1912 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
1914 tcp_enter_loss(sk, 1);
1915 icsk->icsk_retransmits++;
1916 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
1917 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1918 icsk->icsk_rto, TCP_RTO_MAX);
1924 static inline int tcp_fackets_out(const struct tcp_sock *tp)
1926 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
1929 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
1930 * counter when SACK is enabled (without SACK, sacked_out is used for
1933 * Instead, with FACK TCP uses fackets_out that includes both SACKed
1934 * segments up to the highest received SACK block so far and holes in
1937 * With reordering, holes may still be in flight, so RFC3517 recovery
1938 * uses pure sacked_out (total number of SACKed segments) even though
1939 * it violates the RFC that uses duplicate ACKs, often these are equal
1940 * but when e.g. out-of-window ACKs or packet duplication occurs,
1941 * they differ. Since neither occurs due to loss, TCP should really
1944 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
1946 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
1949 static bool tcp_pause_early_retransmit(struct sock *sk, int flag)
1951 struct tcp_sock *tp = tcp_sk(sk);
1952 unsigned long delay;
1954 /* Delay early retransmit and entering fast recovery for
1955 * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
1956 * available, or RTO is scheduled to fire first.
1958 if (sysctl_tcp_early_retrans < 2 || sysctl_tcp_early_retrans > 3 ||
1959 (flag & FLAG_ECE) || !tp->srtt)
1962 delay = max_t(unsigned long, (tp->srtt >> 5), msecs_to_jiffies(2));
1963 if (!time_after(inet_csk(sk)->icsk_timeout, (jiffies + delay)))
1966 inet_csk_reset_xmit_timer(sk, ICSK_TIME_EARLY_RETRANS, delay,
1971 static inline int tcp_skb_timedout(const struct sock *sk,
1972 const struct sk_buff *skb)
1974 return tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto;
1977 static inline int tcp_head_timedout(const struct sock *sk)
1979 const struct tcp_sock *tp = tcp_sk(sk);
1981 return tp->packets_out &&
1982 tcp_skb_timedout(sk, tcp_write_queue_head(sk));
1985 /* Linux NewReno/SACK/FACK/ECN state machine.
1986 * --------------------------------------
1988 * "Open" Normal state, no dubious events, fast path.
1989 * "Disorder" In all the respects it is "Open",
1990 * but requires a bit more attention. It is entered when
1991 * we see some SACKs or dupacks. It is split of "Open"
1992 * mainly to move some processing from fast path to slow one.
1993 * "CWR" CWND was reduced due to some Congestion Notification event.
1994 * It can be ECN, ICMP source quench, local device congestion.
1995 * "Recovery" CWND was reduced, we are fast-retransmitting.
1996 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
1998 * tcp_fastretrans_alert() is entered:
1999 * - each incoming ACK, if state is not "Open"
2000 * - when arrived ACK is unusual, namely:
2005 * Counting packets in flight is pretty simple.
2007 * in_flight = packets_out - left_out + retrans_out
2009 * packets_out is SND.NXT-SND.UNA counted in packets.
2011 * retrans_out is number of retransmitted segments.
2013 * left_out is number of segments left network, but not ACKed yet.
2015 * left_out = sacked_out + lost_out
2017 * sacked_out: Packets, which arrived to receiver out of order
2018 * and hence not ACKed. With SACKs this number is simply
2019 * amount of SACKed data. Even without SACKs
2020 * it is easy to give pretty reliable estimate of this number,
2021 * counting duplicate ACKs.
2023 * lost_out: Packets lost by network. TCP has no explicit
2024 * "loss notification" feedback from network (for now).
2025 * It means that this number can be only _guessed_.
2026 * Actually, it is the heuristics to predict lossage that
2027 * distinguishes different algorithms.
2029 * F.e. after RTO, when all the queue is considered as lost,
2030 * lost_out = packets_out and in_flight = retrans_out.
2032 * Essentially, we have now two algorithms counting
2035 * FACK: It is the simplest heuristics. As soon as we decided
2036 * that something is lost, we decide that _all_ not SACKed
2037 * packets until the most forward SACK are lost. I.e.
2038 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2039 * It is absolutely correct estimate, if network does not reorder
2040 * packets. And it loses any connection to reality when reordering
2041 * takes place. We use FACK by default until reordering
2042 * is suspected on the path to this destination.
2044 * NewReno: when Recovery is entered, we assume that one segment
2045 * is lost (classic Reno). While we are in Recovery and
2046 * a partial ACK arrives, we assume that one more packet
2047 * is lost (NewReno). This heuristics are the same in NewReno
2050 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2051 * deflation etc. CWND is real congestion window, never inflated, changes
2052 * only according to classic VJ rules.
2054 * Really tricky (and requiring careful tuning) part of algorithm
2055 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2056 * The first determines the moment _when_ we should reduce CWND and,
2057 * hence, slow down forward transmission. In fact, it determines the moment
2058 * when we decide that hole is caused by loss, rather than by a reorder.
2060 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2061 * holes, caused by lost packets.
2063 * And the most logically complicated part of algorithm is undo
2064 * heuristics. We detect false retransmits due to both too early
2065 * fast retransmit (reordering) and underestimated RTO, analyzing
2066 * timestamps and D-SACKs. When we detect that some segments were
2067 * retransmitted by mistake and CWND reduction was wrong, we undo
2068 * window reduction and abort recovery phase. This logic is hidden
2069 * inside several functions named tcp_try_undo_<something>.
2072 /* This function decides, when we should leave Disordered state
2073 * and enter Recovery phase, reducing congestion window.
2075 * Main question: may we further continue forward transmission
2076 * with the same cwnd?
2078 static bool tcp_time_to_recover(struct sock *sk, int flag)
2080 struct tcp_sock *tp = tcp_sk(sk);
2083 /* Trick#1: The loss is proven. */
2087 /* Not-A-Trick#2 : Classic rule... */
2088 if (tcp_dupack_heuristics(tp) > tp->reordering)
2091 /* Trick#3 : when we use RFC2988 timer restart, fast
2092 * retransmit can be triggered by timeout of queue head.
2094 if (tcp_is_fack(tp) && tcp_head_timedout(sk))
2097 /* Trick#4: It is still not OK... But will it be useful to delay
2100 packets_out = tp->packets_out;
2101 if (packets_out <= tp->reordering &&
2102 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2103 !tcp_may_send_now(sk)) {
2104 /* We have nothing to send. This connection is limited
2105 * either by receiver window or by application.
2110 /* If a thin stream is detected, retransmit after first
2111 * received dupack. Employ only if SACK is supported in order
2112 * to avoid possible corner-case series of spurious retransmissions
2113 * Use only if there are no unsent data.
2115 if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2116 tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2117 tcp_is_sack(tp) && !tcp_send_head(sk))
2120 /* Trick#6: TCP early retransmit, per RFC5827. To avoid spurious
2121 * retransmissions due to small network reorderings, we implement
2122 * Mitigation A.3 in the RFC and delay the retransmission for a short
2123 * interval if appropriate.
2125 if (tp->do_early_retrans && !tp->retrans_out && tp->sacked_out &&
2126 (tp->packets_out >= (tp->sacked_out + 1) && tp->packets_out < 4) &&
2127 !tcp_may_send_now(sk))
2128 return !tcp_pause_early_retransmit(sk, flag);
2133 /* New heuristics: it is possible only after we switched to restart timer
2134 * each time when something is ACKed. Hence, we can detect timed out packets
2135 * during fast retransmit without falling to slow start.
2137 * Usefulness of this as is very questionable, since we should know which of
2138 * the segments is the next to timeout which is relatively expensive to find
2139 * in general case unless we add some data structure just for that. The
2140 * current approach certainly won't find the right one too often and when it
2141 * finally does find _something_ it usually marks large part of the window
2142 * right away (because a retransmission with a larger timestamp blocks the
2143 * loop from advancing). -ij
2145 static void tcp_timeout_skbs(struct sock *sk)
2147 struct tcp_sock *tp = tcp_sk(sk);
2148 struct sk_buff *skb;
2150 if (!tcp_is_fack(tp) || !tcp_head_timedout(sk))
2153 skb = tp->scoreboard_skb_hint;
2154 if (tp->scoreboard_skb_hint == NULL)
2155 skb = tcp_write_queue_head(sk);
2157 tcp_for_write_queue_from(skb, sk) {
2158 if (skb == tcp_send_head(sk))
2160 if (!tcp_skb_timedout(sk, skb))
2163 tcp_skb_mark_lost(tp, skb);
2166 tp->scoreboard_skb_hint = skb;
2168 tcp_verify_left_out(tp);
2171 /* Detect loss in event "A" above by marking head of queue up as lost.
2172 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2173 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2174 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2175 * the maximum SACKed segments to pass before reaching this limit.
2177 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2179 struct tcp_sock *tp = tcp_sk(sk);
2180 struct sk_buff *skb;
2184 /* Use SACK to deduce losses of new sequences sent during recovery */
2185 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq;
2187 WARN_ON(packets > tp->packets_out);
2188 if (tp->lost_skb_hint) {
2189 skb = tp->lost_skb_hint;
2190 cnt = tp->lost_cnt_hint;
2191 /* Head already handled? */
2192 if (mark_head && skb != tcp_write_queue_head(sk))
2195 skb = tcp_write_queue_head(sk);
2199 tcp_for_write_queue_from(skb, sk) {
2200 if (skb == tcp_send_head(sk))
2202 /* TODO: do this better */
2203 /* this is not the most efficient way to do this... */
2204 tp->lost_skb_hint = skb;
2205 tp->lost_cnt_hint = cnt;
2207 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2211 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2212 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2213 cnt += tcp_skb_pcount(skb);
2215 if (cnt > packets) {
2216 if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2217 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2218 (oldcnt >= packets))
2221 mss = skb_shinfo(skb)->gso_size;
2222 err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss);
2228 tcp_skb_mark_lost(tp, skb);
2233 tcp_verify_left_out(tp);
2236 /* Account newly detected lost packet(s) */
2238 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2240 struct tcp_sock *tp = tcp_sk(sk);
2242 if (tcp_is_reno(tp)) {
2243 tcp_mark_head_lost(sk, 1, 1);
2244 } else if (tcp_is_fack(tp)) {
2245 int lost = tp->fackets_out - tp->reordering;
2248 tcp_mark_head_lost(sk, lost, 0);
2250 int sacked_upto = tp->sacked_out - tp->reordering;
2251 if (sacked_upto >= 0)
2252 tcp_mark_head_lost(sk, sacked_upto, 0);
2253 else if (fast_rexmit)
2254 tcp_mark_head_lost(sk, 1, 1);
2257 tcp_timeout_skbs(sk);
2260 /* CWND moderation, preventing bursts due to too big ACKs
2261 * in dubious situations.
2263 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2265 tp->snd_cwnd = min(tp->snd_cwnd,
2266 tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2267 tp->snd_cwnd_stamp = tcp_time_stamp;
2270 /* Nothing was retransmitted or returned timestamp is less
2271 * than timestamp of the first retransmission.
2273 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2275 return !tp->retrans_stamp ||
2276 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2277 before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
2280 /* Undo procedures. */
2282 #if FASTRETRANS_DEBUG > 1
2283 static void DBGUNDO(struct sock *sk, const char *msg)
2285 struct tcp_sock *tp = tcp_sk(sk);
2286 struct inet_sock *inet = inet_sk(sk);
2288 if (sk->sk_family == AF_INET) {
2289 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2291 &inet->inet_daddr, ntohs(inet->inet_dport),
2292 tp->snd_cwnd, tcp_left_out(tp),
2293 tp->snd_ssthresh, tp->prior_ssthresh,
2296 #if IS_ENABLED(CONFIG_IPV6)
2297 else if (sk->sk_family == AF_INET6) {
2298 struct ipv6_pinfo *np = inet6_sk(sk);
2299 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2301 &np->daddr, ntohs(inet->inet_dport),
2302 tp->snd_cwnd, tcp_left_out(tp),
2303 tp->snd_ssthresh, tp->prior_ssthresh,
2309 #define DBGUNDO(x...) do { } while (0)
2312 static void tcp_undo_cwr(struct sock *sk, const bool undo_ssthresh)
2314 struct tcp_sock *tp = tcp_sk(sk);
2316 if (tp->prior_ssthresh) {
2317 const struct inet_connection_sock *icsk = inet_csk(sk);
2319 if (icsk->icsk_ca_ops->undo_cwnd)
2320 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2322 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2324 if (undo_ssthresh && tp->prior_ssthresh > tp->snd_ssthresh) {
2325 tp->snd_ssthresh = tp->prior_ssthresh;
2326 TCP_ECN_withdraw_cwr(tp);
2329 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2331 tp->snd_cwnd_stamp = tcp_time_stamp;
2334 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2336 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2339 /* People celebrate: "We love our President!" */
2340 static bool tcp_try_undo_recovery(struct sock *sk)
2342 struct tcp_sock *tp = tcp_sk(sk);
2344 if (tcp_may_undo(tp)) {
2347 /* Happy end! We did not retransmit anything
2348 * or our original transmission succeeded.
2350 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2351 tcp_undo_cwr(sk, true);
2352 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2353 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2355 mib_idx = LINUX_MIB_TCPFULLUNDO;
2357 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2358 tp->undo_marker = 0;
2360 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2361 /* Hold old state until something *above* high_seq
2362 * is ACKed. For Reno it is MUST to prevent false
2363 * fast retransmits (RFC2582). SACK TCP is safe. */
2364 tcp_moderate_cwnd(tp);
2367 tcp_set_ca_state(sk, TCP_CA_Open);
2371 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2372 static void tcp_try_undo_dsack(struct sock *sk)
2374 struct tcp_sock *tp = tcp_sk(sk);
2376 if (tp->undo_marker && !tp->undo_retrans) {
2377 DBGUNDO(sk, "D-SACK");
2378 tcp_undo_cwr(sk, true);
2379 tp->undo_marker = 0;
2380 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2384 /* We can clear retrans_stamp when there are no retransmissions in the
2385 * window. It would seem that it is trivially available for us in
2386 * tp->retrans_out, however, that kind of assumptions doesn't consider
2387 * what will happen if errors occur when sending retransmission for the
2388 * second time. ...It could the that such segment has only
2389 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2390 * the head skb is enough except for some reneging corner cases that
2391 * are not worth the effort.
2393 * Main reason for all this complexity is the fact that connection dying
2394 * time now depends on the validity of the retrans_stamp, in particular,
2395 * that successive retransmissions of a segment must not advance
2396 * retrans_stamp under any conditions.
2398 static bool tcp_any_retrans_done(const struct sock *sk)
2400 const struct tcp_sock *tp = tcp_sk(sk);
2401 struct sk_buff *skb;
2403 if (tp->retrans_out)
2406 skb = tcp_write_queue_head(sk);
2407 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2413 /* Undo during fast recovery after partial ACK. */
2415 static int tcp_try_undo_partial(struct sock *sk, int acked)
2417 struct tcp_sock *tp = tcp_sk(sk);
2418 /* Partial ACK arrived. Force Hoe's retransmit. */
2419 int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering);
2421 if (tcp_may_undo(tp)) {
2422 /* Plain luck! Hole if filled with delayed
2423 * packet, rather than with a retransmit.
2425 if (!tcp_any_retrans_done(sk))
2426 tp->retrans_stamp = 0;
2428 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2431 tcp_undo_cwr(sk, false);
2432 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2434 /* So... Do not make Hoe's retransmit yet.
2435 * If the first packet was delayed, the rest
2436 * ones are most probably delayed as well.
2443 /* Undo during loss recovery after partial ACK or using F-RTO. */
2444 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2446 struct tcp_sock *tp = tcp_sk(sk);
2448 if (frto_undo || tcp_may_undo(tp)) {
2449 struct sk_buff *skb;
2450 tcp_for_write_queue(skb, sk) {
2451 if (skb == tcp_send_head(sk))
2453 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2456 tcp_clear_all_retrans_hints(tp);
2458 DBGUNDO(sk, "partial loss");
2460 tcp_undo_cwr(sk, true);
2461 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2463 NET_INC_STATS_BH(sock_net(sk),
2464 LINUX_MIB_TCPSPURIOUSRTOS);
2465 inet_csk(sk)->icsk_retransmits = 0;
2466 tp->undo_marker = 0;
2467 if (frto_undo || tcp_is_sack(tp))
2468 tcp_set_ca_state(sk, TCP_CA_Open);
2474 /* The cwnd reduction in CWR and Recovery use the PRR algorithm
2475 * https://datatracker.ietf.org/doc/draft-ietf-tcpm-proportional-rate-reduction/
2476 * It computes the number of packets to send (sndcnt) based on packets newly
2478 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2479 * cwnd reductions across a full RTT.
2480 * 2) If packets in flight is lower than ssthresh (such as due to excess
2481 * losses and/or application stalls), do not perform any further cwnd
2482 * reductions, but instead slow start up to ssthresh.
2484 static void tcp_init_cwnd_reduction(struct sock *sk, const bool set_ssthresh)
2486 struct tcp_sock *tp = tcp_sk(sk);
2488 tp->high_seq = tp->snd_nxt;
2489 tp->tlp_high_seq = 0;
2490 tp->snd_cwnd_cnt = 0;
2491 tp->prior_cwnd = tp->snd_cwnd;
2492 tp->prr_delivered = 0;
2495 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2496 TCP_ECN_queue_cwr(tp);
2499 static void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked,
2502 struct tcp_sock *tp = tcp_sk(sk);
2504 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2506 tp->prr_delivered += newly_acked_sacked;
2507 if (tcp_packets_in_flight(tp) > tp->snd_ssthresh) {
2508 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2510 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2512 sndcnt = min_t(int, delta,
2513 max_t(int, tp->prr_delivered - tp->prr_out,
2514 newly_acked_sacked) + 1);
2517 sndcnt = max(sndcnt, (fast_rexmit ? 1 : 0));
2518 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2521 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2523 struct tcp_sock *tp = tcp_sk(sk);
2525 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2526 if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR ||
2527 (tp->undo_marker && tp->snd_ssthresh < TCP_INFINITE_SSTHRESH)) {
2528 tp->snd_cwnd = tp->snd_ssthresh;
2529 tp->snd_cwnd_stamp = tcp_time_stamp;
2531 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2534 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2535 void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
2537 struct tcp_sock *tp = tcp_sk(sk);
2539 tp->prior_ssthresh = 0;
2540 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2541 tp->undo_marker = 0;
2542 tcp_init_cwnd_reduction(sk, set_ssthresh);
2543 tcp_set_ca_state(sk, TCP_CA_CWR);
2547 static void tcp_try_keep_open(struct sock *sk)
2549 struct tcp_sock *tp = tcp_sk(sk);
2550 int state = TCP_CA_Open;
2552 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2553 state = TCP_CA_Disorder;
2555 if (inet_csk(sk)->icsk_ca_state != state) {
2556 tcp_set_ca_state(sk, state);
2557 tp->high_seq = tp->snd_nxt;
2561 static void tcp_try_to_open(struct sock *sk, int flag, int newly_acked_sacked)
2563 struct tcp_sock *tp = tcp_sk(sk);
2565 tcp_verify_left_out(tp);
2567 if (!tcp_any_retrans_done(sk))
2568 tp->retrans_stamp = 0;
2570 if (flag & FLAG_ECE)
2571 tcp_enter_cwr(sk, 1);
2573 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2574 tcp_try_keep_open(sk);
2575 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
2576 tcp_moderate_cwnd(tp);
2578 tcp_cwnd_reduction(sk, newly_acked_sacked, 0);
2582 static void tcp_mtup_probe_failed(struct sock *sk)
2584 struct inet_connection_sock *icsk = inet_csk(sk);
2586 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2587 icsk->icsk_mtup.probe_size = 0;
2590 static void tcp_mtup_probe_success(struct sock *sk)
2592 struct tcp_sock *tp = tcp_sk(sk);
2593 struct inet_connection_sock *icsk = inet_csk(sk);
2595 /* FIXME: breaks with very large cwnd */
2596 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2597 tp->snd_cwnd = tp->snd_cwnd *
2598 tcp_mss_to_mtu(sk, tp->mss_cache) /
2599 icsk->icsk_mtup.probe_size;
2600 tp->snd_cwnd_cnt = 0;
2601 tp->snd_cwnd_stamp = tcp_time_stamp;
2602 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2604 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2605 icsk->icsk_mtup.probe_size = 0;
2606 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2609 /* Do a simple retransmit without using the backoff mechanisms in
2610 * tcp_timer. This is used for path mtu discovery.
2611 * The socket is already locked here.
2613 void tcp_simple_retransmit(struct sock *sk)
2615 const struct inet_connection_sock *icsk = inet_csk(sk);
2616 struct tcp_sock *tp = tcp_sk(sk);
2617 struct sk_buff *skb;
2618 unsigned int mss = tcp_current_mss(sk);
2619 u32 prior_lost = tp->lost_out;
2621 tcp_for_write_queue(skb, sk) {
2622 if (skb == tcp_send_head(sk))
2624 if (tcp_skb_seglen(skb) > mss &&
2625 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2626 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2627 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2628 tp->retrans_out -= tcp_skb_pcount(skb);
2630 tcp_skb_mark_lost_uncond_verify(tp, skb);
2634 tcp_clear_retrans_hints_partial(tp);
2636 if (prior_lost == tp->lost_out)
2639 if (tcp_is_reno(tp))
2640 tcp_limit_reno_sacked(tp);
2642 tcp_verify_left_out(tp);
2644 /* Don't muck with the congestion window here.
2645 * Reason is that we do not increase amount of _data_
2646 * in network, but units changed and effective
2647 * cwnd/ssthresh really reduced now.
2649 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2650 tp->high_seq = tp->snd_nxt;
2651 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2652 tp->prior_ssthresh = 0;
2653 tp->undo_marker = 0;
2654 tcp_set_ca_state(sk, TCP_CA_Loss);
2656 tcp_xmit_retransmit_queue(sk);
2658 EXPORT_SYMBOL(tcp_simple_retransmit);
2660 static void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2662 struct tcp_sock *tp = tcp_sk(sk);
2665 if (tcp_is_reno(tp))
2666 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2668 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2670 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2672 tp->prior_ssthresh = 0;
2673 tp->undo_marker = tp->snd_una;
2674 tp->undo_retrans = tp->retrans_out;
2676 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2678 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2679 tcp_init_cwnd_reduction(sk, true);
2681 tcp_set_ca_state(sk, TCP_CA_Recovery);
2684 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2685 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2687 static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack)
2689 struct inet_connection_sock *icsk = inet_csk(sk);
2690 struct tcp_sock *tp = tcp_sk(sk);
2691 bool recovered = !before(tp->snd_una, tp->high_seq);
2693 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2694 if (flag & FLAG_ORIG_SACK_ACKED) {
2695 /* Step 3.b. A timeout is spurious if not all data are
2696 * lost, i.e., never-retransmitted data are (s)acked.
2698 tcp_try_undo_loss(sk, true);
2701 if (after(tp->snd_nxt, tp->high_seq) &&
2702 (flag & FLAG_DATA_SACKED || is_dupack)) {
2703 tp->frto = 0; /* Loss was real: 2nd part of step 3.a */
2704 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2705 tp->high_seq = tp->snd_nxt;
2706 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
2708 if (after(tp->snd_nxt, tp->high_seq))
2709 return; /* Step 2.b */
2715 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2716 icsk->icsk_retransmits = 0;
2717 tcp_try_undo_recovery(sk);
2720 if (flag & FLAG_DATA_ACKED)
2721 icsk->icsk_retransmits = 0;
2722 if (tcp_is_reno(tp)) {
2723 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2724 * delivered. Lower inflight to clock out (re)tranmissions.
2726 if (after(tp->snd_nxt, tp->high_seq) && is_dupack)
2727 tcp_add_reno_sack(sk);
2728 else if (flag & FLAG_SND_UNA_ADVANCED)
2729 tcp_reset_reno_sack(tp);
2731 if (tcp_try_undo_loss(sk, false))
2733 tcp_xmit_retransmit_queue(sk);
2736 /* Process an event, which can update packets-in-flight not trivially.
2737 * Main goal of this function is to calculate new estimate for left_out,
2738 * taking into account both packets sitting in receiver's buffer and
2739 * packets lost by network.
2741 * Besides that it does CWND reduction, when packet loss is detected
2742 * and changes state of machine.
2744 * It does _not_ decide what to send, it is made in function
2745 * tcp_xmit_retransmit_queue().
2747 static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked,
2748 int prior_sacked, bool is_dupack,
2751 struct inet_connection_sock *icsk = inet_csk(sk);
2752 struct tcp_sock *tp = tcp_sk(sk);
2753 int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2754 (tcp_fackets_out(tp) > tp->reordering));
2755 int newly_acked_sacked = 0;
2756 int fast_rexmit = 0;
2758 if (WARN_ON(!tp->packets_out && tp->sacked_out))
2760 if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2761 tp->fackets_out = 0;
2763 /* Now state machine starts.
2764 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2765 if (flag & FLAG_ECE)
2766 tp->prior_ssthresh = 0;
2768 /* B. In all the states check for reneging SACKs. */
2769 if (tcp_check_sack_reneging(sk, flag))
2772 /* C. Check consistency of the current state. */
2773 tcp_verify_left_out(tp);
2775 /* D. Check state exit conditions. State can be terminated
2776 * when high_seq is ACKed. */
2777 if (icsk->icsk_ca_state == TCP_CA_Open) {
2778 WARN_ON(tp->retrans_out != 0);
2779 tp->retrans_stamp = 0;
2780 } else if (!before(tp->snd_una, tp->high_seq)) {
2781 switch (icsk->icsk_ca_state) {
2783 /* CWR is to be held something *above* high_seq
2784 * is ACKed for CWR bit to reach receiver. */
2785 if (tp->snd_una != tp->high_seq) {
2786 tcp_end_cwnd_reduction(sk);
2787 tcp_set_ca_state(sk, TCP_CA_Open);
2791 case TCP_CA_Recovery:
2792 if (tcp_is_reno(tp))
2793 tcp_reset_reno_sack(tp);
2794 if (tcp_try_undo_recovery(sk))
2796 tcp_end_cwnd_reduction(sk);
2801 /* E. Process state. */
2802 switch (icsk->icsk_ca_state) {
2803 case TCP_CA_Recovery:
2804 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2805 if (tcp_is_reno(tp) && is_dupack)
2806 tcp_add_reno_sack(sk);
2808 do_lost = tcp_try_undo_partial(sk, pkts_acked);
2809 newly_acked_sacked = pkts_acked + tp->sacked_out - prior_sacked;
2812 tcp_process_loss(sk, flag, is_dupack);
2813 if (icsk->icsk_ca_state != TCP_CA_Open)
2815 /* Fall through to processing in Open state. */
2817 if (tcp_is_reno(tp)) {
2818 if (flag & FLAG_SND_UNA_ADVANCED)
2819 tcp_reset_reno_sack(tp);
2821 tcp_add_reno_sack(sk);
2823 newly_acked_sacked = pkts_acked + tp->sacked_out - prior_sacked;
2825 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2826 tcp_try_undo_dsack(sk);
2828 if (!tcp_time_to_recover(sk, flag)) {
2829 tcp_try_to_open(sk, flag, newly_acked_sacked);
2833 /* MTU probe failure: don't reduce cwnd */
2834 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2835 icsk->icsk_mtup.probe_size &&
2836 tp->snd_una == tp->mtu_probe.probe_seq_start) {
2837 tcp_mtup_probe_failed(sk);
2838 /* Restores the reduction we did in tcp_mtup_probe() */
2840 tcp_simple_retransmit(sk);
2844 /* Otherwise enter Recovery state */
2845 tcp_enter_recovery(sk, (flag & FLAG_ECE));
2849 if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk)))
2850 tcp_update_scoreboard(sk, fast_rexmit);
2851 tcp_cwnd_reduction(sk, newly_acked_sacked, fast_rexmit);
2852 tcp_xmit_retransmit_queue(sk);
2855 void tcp_valid_rtt_meas(struct sock *sk, u32 seq_rtt)
2857 tcp_rtt_estimator(sk, seq_rtt);
2859 inet_csk(sk)->icsk_backoff = 0;
2861 EXPORT_SYMBOL(tcp_valid_rtt_meas);
2863 /* Read draft-ietf-tcplw-high-performance before mucking
2864 * with this code. (Supersedes RFC1323)
2866 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
2868 /* RTTM Rule: A TSecr value received in a segment is used to
2869 * update the averaged RTT measurement only if the segment
2870 * acknowledges some new data, i.e., only if it advances the
2871 * left edge of the send window.
2873 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2874 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2876 * Changed: reset backoff as soon as we see the first valid sample.
2877 * If we do not, we get strongly overestimated rto. With timestamps
2878 * samples are accepted even from very old segments: f.e., when rtt=1
2879 * increases to 8, we retransmit 5 times and after 8 seconds delayed
2880 * answer arrives rto becomes 120 seconds! If at least one of segments
2881 * in window is lost... Voila. --ANK (010210)
2883 struct tcp_sock *tp = tcp_sk(sk);
2885 tcp_valid_rtt_meas(sk, tcp_time_stamp - tp->rx_opt.rcv_tsecr);
2888 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
2890 /* We don't have a timestamp. Can only use
2891 * packets that are not retransmitted to determine
2892 * rtt estimates. Also, we must not reset the
2893 * backoff for rto until we get a non-retransmitted
2894 * packet. This allows us to deal with a situation
2895 * where the network delay has increased suddenly.
2896 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2899 if (flag & FLAG_RETRANS_DATA_ACKED)
2902 tcp_valid_rtt_meas(sk, seq_rtt);
2905 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
2908 const struct tcp_sock *tp = tcp_sk(sk);
2909 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2910 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
2911 tcp_ack_saw_tstamp(sk, flag);
2912 else if (seq_rtt >= 0)
2913 tcp_ack_no_tstamp(sk, seq_rtt, flag);
2916 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
2918 const struct inet_connection_sock *icsk = inet_csk(sk);
2919 icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight);
2920 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
2923 /* Restart timer after forward progress on connection.
2924 * RFC2988 recommends to restart timer to now+rto.
2926 void tcp_rearm_rto(struct sock *sk)
2928 const struct inet_connection_sock *icsk = inet_csk(sk);
2929 struct tcp_sock *tp = tcp_sk(sk);
2931 /* If the retrans timer is currently being used by Fast Open
2932 * for SYN-ACK retrans purpose, stay put.
2934 if (tp->fastopen_rsk)
2937 if (!tp->packets_out) {
2938 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2940 u32 rto = inet_csk(sk)->icsk_rto;
2941 /* Offset the time elapsed after installing regular RTO */
2942 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
2943 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
2944 struct sk_buff *skb = tcp_write_queue_head(sk);
2945 const u32 rto_time_stamp = TCP_SKB_CB(skb)->when + rto;
2946 s32 delta = (s32)(rto_time_stamp - tcp_time_stamp);
2947 /* delta may not be positive if the socket is locked
2948 * when the retrans timer fires and is rescheduled.
2953 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
2958 /* This function is called when the delayed ER timer fires. TCP enters
2959 * fast recovery and performs fast-retransmit.
2961 void tcp_resume_early_retransmit(struct sock *sk)
2963 struct tcp_sock *tp = tcp_sk(sk);
2967 /* Stop if ER is disabled after the delayed ER timer is scheduled */
2968 if (!tp->do_early_retrans)
2971 tcp_enter_recovery(sk, false);
2972 tcp_update_scoreboard(sk, 1);
2973 tcp_xmit_retransmit_queue(sk);
2976 /* If we get here, the whole TSO packet has not been acked. */
2977 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
2979 struct tcp_sock *tp = tcp_sk(sk);
2982 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
2984 packets_acked = tcp_skb_pcount(skb);
2985 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2987 packets_acked -= tcp_skb_pcount(skb);
2989 if (packets_acked) {
2990 BUG_ON(tcp_skb_pcount(skb) == 0);
2991 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
2994 return packets_acked;
2997 /* Remove acknowledged frames from the retransmission queue. If our packet
2998 * is before the ack sequence we can discard it as it's confirmed to have
2999 * arrived at the other end.
3001 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3004 struct tcp_sock *tp = tcp_sk(sk);
3005 const struct inet_connection_sock *icsk = inet_csk(sk);
3006 struct sk_buff *skb;
3007 u32 now = tcp_time_stamp;
3008 int fully_acked = true;
3011 u32 reord = tp->packets_out;
3012 u32 prior_sacked = tp->sacked_out;
3014 s32 ca_seq_rtt = -1;
3015 ktime_t last_ackt = net_invalid_timestamp();
3017 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3018 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3020 u8 sacked = scb->sacked;
3022 /* Determine how many packets and what bytes were acked, tso and else */
3023 if (after(scb->end_seq, tp->snd_una)) {
3024 if (tcp_skb_pcount(skb) == 1 ||
3025 !after(tp->snd_una, scb->seq))
3028 acked_pcount = tcp_tso_acked(sk, skb);
3032 fully_acked = false;
3034 acked_pcount = tcp_skb_pcount(skb);
3037 if (sacked & TCPCB_RETRANS) {
3038 if (sacked & TCPCB_SACKED_RETRANS)
3039 tp->retrans_out -= acked_pcount;
3040 flag |= FLAG_RETRANS_DATA_ACKED;
3044 ca_seq_rtt = now - scb->when;
3045 last_ackt = skb->tstamp;
3047 seq_rtt = ca_seq_rtt;
3049 if (!(sacked & TCPCB_SACKED_ACKED))
3050 reord = min(pkts_acked, reord);
3051 if (!after(scb->end_seq, tp->high_seq))
3052 flag |= FLAG_ORIG_SACK_ACKED;
3055 if (sacked & TCPCB_SACKED_ACKED)
3056 tp->sacked_out -= acked_pcount;
3057 if (sacked & TCPCB_LOST)
3058 tp->lost_out -= acked_pcount;
3060 tp->packets_out -= acked_pcount;
3061 pkts_acked += acked_pcount;
3063 /* Initial outgoing SYN's get put onto the write_queue
3064 * just like anything else we transmit. It is not
3065 * true data, and if we misinform our callers that
3066 * this ACK acks real data, we will erroneously exit
3067 * connection startup slow start one packet too
3068 * quickly. This is severely frowned upon behavior.
3070 if (!(scb->tcp_flags & TCPHDR_SYN)) {
3071 flag |= FLAG_DATA_ACKED;
3073 flag |= FLAG_SYN_ACKED;
3074 tp->retrans_stamp = 0;
3080 tcp_unlink_write_queue(skb, sk);
3081 sk_wmem_free_skb(sk, skb);
3082 tp->scoreboard_skb_hint = NULL;
3083 if (skb == tp->retransmit_skb_hint)
3084 tp->retransmit_skb_hint = NULL;
3085 if (skb == tp->lost_skb_hint)
3086 tp->lost_skb_hint = NULL;
3089 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3090 tp->snd_up = tp->snd_una;
3092 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3093 flag |= FLAG_SACK_RENEGING;
3095 if (flag & FLAG_ACKED) {
3096 const struct tcp_congestion_ops *ca_ops
3097 = inet_csk(sk)->icsk_ca_ops;
3099 if (unlikely(icsk->icsk_mtup.probe_size &&
3100 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3101 tcp_mtup_probe_success(sk);
3104 tcp_ack_update_rtt(sk, flag, seq_rtt);
3107 if (tcp_is_reno(tp)) {
3108 tcp_remove_reno_sacks(sk, pkts_acked);
3112 /* Non-retransmitted hole got filled? That's reordering */
3113 if (reord < prior_fackets)
3114 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3116 delta = tcp_is_fack(tp) ? pkts_acked :
3117 prior_sacked - tp->sacked_out;
3118 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3121 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3123 if (ca_ops->pkts_acked) {
3126 /* Is the ACK triggering packet unambiguous? */
3127 if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
3128 /* High resolution needed and available? */
3129 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
3130 !ktime_equal(last_ackt,
3131 net_invalid_timestamp()))
3132 rtt_us = ktime_us_delta(ktime_get_real(),
3134 else if (ca_seq_rtt >= 0)
3135 rtt_us = jiffies_to_usecs(ca_seq_rtt);
3138 ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
3142 #if FASTRETRANS_DEBUG > 0
3143 WARN_ON((int)tp->sacked_out < 0);
3144 WARN_ON((int)tp->lost_out < 0);
3145 WARN_ON((int)tp->retrans_out < 0);
3146 if (!tp->packets_out && tcp_is_sack(tp)) {
3147 icsk = inet_csk(sk);
3149 pr_debug("Leak l=%u %d\n",
3150 tp->lost_out, icsk->icsk_ca_state);
3153 if (tp->sacked_out) {
3154 pr_debug("Leak s=%u %d\n",
3155 tp->sacked_out, icsk->icsk_ca_state);
3158 if (tp->retrans_out) {
3159 pr_debug("Leak r=%u %d\n",
3160 tp->retrans_out, icsk->icsk_ca_state);
3161 tp->retrans_out = 0;
3168 static void tcp_ack_probe(struct sock *sk)
3170 const struct tcp_sock *tp = tcp_sk(sk);
3171 struct inet_connection_sock *icsk = inet_csk(sk);
3173 /* Was it a usable window open? */
3175 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3176 icsk->icsk_backoff = 0;
3177 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3178 /* Socket must be waked up by subsequent tcp_data_snd_check().
3179 * This function is not for random using!
3182 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3183 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3188 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3190 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3191 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3194 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3196 const struct tcp_sock *tp = tcp_sk(sk);
3197 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
3198 !tcp_in_cwnd_reduction(sk);
3201 /* Check that window update is acceptable.
3202 * The function assumes that snd_una<=ack<=snd_next.
3204 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3205 const u32 ack, const u32 ack_seq,
3208 return after(ack, tp->snd_una) ||
3209 after(ack_seq, tp->snd_wl1) ||
3210 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3213 /* Update our send window.
3215 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3216 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3218 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3221 struct tcp_sock *tp = tcp_sk(sk);
3223 u32 nwin = ntohs(tcp_hdr(skb)->window);
3225 if (likely(!tcp_hdr(skb)->syn))
3226 nwin <<= tp->rx_opt.snd_wscale;
3228 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3229 flag |= FLAG_WIN_UPDATE;
3230 tcp_update_wl(tp, ack_seq);
3232 if (tp->snd_wnd != nwin) {
3235 /* Note, it is the only place, where
3236 * fast path is recovered for sending TCP.
3239 tcp_fast_path_check(sk);
3241 if (nwin > tp->max_window) {
3242 tp->max_window = nwin;
3243 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3253 /* RFC 5961 7 [ACK Throttling] */
3254 static void tcp_send_challenge_ack(struct sock *sk)
3256 /* unprotected vars, we dont care of overwrites */
3257 static u32 challenge_timestamp;
3258 static unsigned int challenge_count;
3259 u32 now = jiffies / HZ;
3261 if (now != challenge_timestamp) {
3262 challenge_timestamp = now;
3263 challenge_count = 0;
3265 if (++challenge_count <= sysctl_tcp_challenge_ack_limit) {
3266 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPCHALLENGEACK);
3271 /* This routine deals with acks during a TLP episode.
3272 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3274 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3276 struct tcp_sock *tp = tcp_sk(sk);
3277 bool is_tlp_dupack = (ack == tp->tlp_high_seq) &&
3278 !(flag & (FLAG_SND_UNA_ADVANCED |
3279 FLAG_NOT_DUP | FLAG_DATA_SACKED));
3281 /* Mark the end of TLP episode on receiving TLP dupack or when
3282 * ack is after tlp_high_seq.
3284 if (is_tlp_dupack) {
3285 tp->tlp_high_seq = 0;
3289 if (after(ack, tp->tlp_high_seq)) {
3290 tp->tlp_high_seq = 0;
3291 /* Don't reduce cwnd if DSACK arrives for TLP retrans. */
3292 if (!(flag & FLAG_DSACKING_ACK)) {
3293 tcp_init_cwnd_reduction(sk, true);
3294 tcp_set_ca_state(sk, TCP_CA_CWR);
3295 tcp_end_cwnd_reduction(sk);
3296 tcp_set_ca_state(sk, TCP_CA_Open);
3297 NET_INC_STATS_BH(sock_net(sk),
3298 LINUX_MIB_TCPLOSSPROBERECOVERY);
3303 /* This routine deals with incoming acks, but not outgoing ones. */
3304 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3306 struct inet_connection_sock *icsk = inet_csk(sk);
3307 struct tcp_sock *tp = tcp_sk(sk);
3308 u32 prior_snd_una = tp->snd_una;
3309 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3310 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3311 bool is_dupack = false;
3312 u32 prior_in_flight;
3315 int prior_sacked = tp->sacked_out;
3318 /* If the ack is older than previous acks
3319 * then we can probably ignore it.
3321 if (before(ack, prior_snd_una)) {
3322 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3323 if (before(ack, prior_snd_una - tp->max_window)) {
3324 tcp_send_challenge_ack(sk);
3330 /* If the ack includes data we haven't sent yet, discard
3331 * this segment (RFC793 Section 3.9).
3333 if (after(ack, tp->snd_nxt))
3336 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
3337 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
3340 if (after(ack, prior_snd_una))
3341 flag |= FLAG_SND_UNA_ADVANCED;
3343 prior_fackets = tp->fackets_out;
3344 prior_in_flight = tcp_packets_in_flight(tp);
3346 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3347 /* Window is constant, pure forward advance.
3348 * No more checks are required.
3349 * Note, we use the fact that SND.UNA>=SND.WL2.
3351 tcp_update_wl(tp, ack_seq);
3353 flag |= FLAG_WIN_UPDATE;
3355 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
3357 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3359 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3362 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3364 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3366 if (TCP_SKB_CB(skb)->sacked)
3367 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3369 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
3372 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
3375 /* We passed data and got it acked, remove any soft error
3376 * log. Something worked...
3378 sk->sk_err_soft = 0;
3379 icsk->icsk_probes_out = 0;
3380 tp->rcv_tstamp = tcp_time_stamp;
3381 prior_packets = tp->packets_out;
3385 /* See if we can take anything off of the retransmit queue. */
3386 flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una);
3388 pkts_acked = prior_packets - tp->packets_out;
3390 if (tcp_ack_is_dubious(sk, flag)) {
3391 /* Advance CWND, if state allows this. */
3392 if ((flag & FLAG_DATA_ACKED) && tcp_may_raise_cwnd(sk, flag))
3393 tcp_cong_avoid(sk, ack, prior_in_flight);
3394 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3395 tcp_fastretrans_alert(sk, pkts_acked, prior_sacked,
3398 if (flag & FLAG_DATA_ACKED)
3399 tcp_cong_avoid(sk, ack, prior_in_flight);
3402 if (tp->tlp_high_seq)
3403 tcp_process_tlp_ack(sk, ack, flag);
3405 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) {
3406 struct dst_entry *dst = __sk_dst_get(sk);
3411 if (icsk->icsk_pending == ICSK_TIME_RETRANS)
3412 tcp_schedule_loss_probe(sk);
3416 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3417 if (flag & FLAG_DSACKING_ACK)
3418 tcp_fastretrans_alert(sk, pkts_acked, prior_sacked,
3420 /* If this ack opens up a zero window, clear backoff. It was
3421 * being used to time the probes, and is probably far higher than
3422 * it needs to be for normal retransmission.
3424 if (tcp_send_head(sk))
3427 if (tp->tlp_high_seq)
3428 tcp_process_tlp_ack(sk, ack, flag);
3432 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3436 /* If data was SACKed, tag it and see if we should send more data.
3437 * If data was DSACKed, see if we can undo a cwnd reduction.
3439 if (TCP_SKB_CB(skb)->sacked) {
3440 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3441 tcp_fastretrans_alert(sk, pkts_acked, prior_sacked,
3445 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3449 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3450 * But, this can also be called on packets in the established flow when
3451 * the fast version below fails.
3453 void tcp_parse_options(const struct sk_buff *skb,
3454 struct tcp_options_received *opt_rx, int estab,
3455 struct tcp_fastopen_cookie *foc)
3457 const unsigned char *ptr;
3458 const struct tcphdr *th = tcp_hdr(skb);
3459 int length = (th->doff * 4) - sizeof(struct tcphdr);
3461 ptr = (const unsigned char *)(th + 1);
3462 opt_rx->saw_tstamp = 0;
3464 while (length > 0) {
3465 int opcode = *ptr++;
3471 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3476 if (opsize < 2) /* "silly options" */
3478 if (opsize > length)
3479 return; /* don't parse partial options */
3482 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3483 u16 in_mss = get_unaligned_be16(ptr);
3485 if (opt_rx->user_mss &&
3486 opt_rx->user_mss < in_mss)
3487 in_mss = opt_rx->user_mss;
3488 opt_rx->mss_clamp = in_mss;
3493 if (opsize == TCPOLEN_WINDOW && th->syn &&
3494 !estab && sysctl_tcp_window_scaling) {
3495 __u8 snd_wscale = *(__u8 *)ptr;
3496 opt_rx->wscale_ok = 1;
3497 if (snd_wscale > 14) {
3498 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3503 opt_rx->snd_wscale = snd_wscale;
3506 case TCPOPT_TIMESTAMP:
3507 if ((opsize == TCPOLEN_TIMESTAMP) &&
3508 ((estab && opt_rx->tstamp_ok) ||
3509 (!estab && sysctl_tcp_timestamps))) {
3510 opt_rx->saw_tstamp = 1;
3511 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3512 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3515 case TCPOPT_SACK_PERM:
3516 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3517 !estab && sysctl_tcp_sack) {
3518 opt_rx->sack_ok = TCP_SACK_SEEN;
3519 tcp_sack_reset(opt_rx);
3524 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3525 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3527 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3530 #ifdef CONFIG_TCP_MD5SIG
3533 * The MD5 Hash has already been
3534 * checked (see tcp_v{4,6}_do_rcv()).
3539 /* Fast Open option shares code 254 using a
3540 * 16 bits magic number. It's valid only in
3541 * SYN or SYN-ACK with an even size.
3543 if (opsize < TCPOLEN_EXP_FASTOPEN_BASE ||
3544 get_unaligned_be16(ptr) != TCPOPT_FASTOPEN_MAGIC ||
3545 foc == NULL || !th->syn || (opsize & 1))
3547 foc->len = opsize - TCPOLEN_EXP_FASTOPEN_BASE;
3548 if (foc->len >= TCP_FASTOPEN_COOKIE_MIN &&
3549 foc->len <= TCP_FASTOPEN_COOKIE_MAX)
3550 memcpy(foc->val, ptr + 2, foc->len);
3551 else if (foc->len != 0)
3561 EXPORT_SYMBOL(tcp_parse_options);
3563 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3565 const __be32 *ptr = (const __be32 *)(th + 1);
3567 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3568 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3569 tp->rx_opt.saw_tstamp = 1;
3571 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3573 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3579 /* Fast parse options. This hopes to only see timestamps.
3580 * If it is wrong it falls back on tcp_parse_options().
3582 static bool tcp_fast_parse_options(const struct sk_buff *skb,
3583 const struct tcphdr *th, struct tcp_sock *tp)
3585 /* In the spirit of fast parsing, compare doff directly to constant
3586 * values. Because equality is used, short doff can be ignored here.
3588 if (th->doff == (sizeof(*th) / 4)) {
3589 tp->rx_opt.saw_tstamp = 0;
3591 } else if (tp->rx_opt.tstamp_ok &&
3592 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3593 if (tcp_parse_aligned_timestamp(tp, th))
3597 tcp_parse_options(skb, &tp->rx_opt, 1, NULL);
3598 if (tp->rx_opt.saw_tstamp)
3599 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3604 #ifdef CONFIG_TCP_MD5SIG
3606 * Parse MD5 Signature option
3608 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3610 int length = (th->doff << 2) - sizeof(*th);
3611 const u8 *ptr = (const u8 *)(th + 1);
3613 /* If the TCP option is too short, we can short cut */
3614 if (length < TCPOLEN_MD5SIG)
3617 while (length > 0) {
3618 int opcode = *ptr++;
3629 if (opsize < 2 || opsize > length)
3631 if (opcode == TCPOPT_MD5SIG)
3632 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3639 EXPORT_SYMBOL(tcp_parse_md5sig_option);
3642 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
3644 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3645 tp->rx_opt.ts_recent_stamp = get_seconds();
3648 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3650 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3651 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3652 * extra check below makes sure this can only happen
3653 * for pure ACK frames. -DaveM
3655 * Not only, also it occurs for expired timestamps.
3658 if (tcp_paws_check(&tp->rx_opt, 0))
3659 tcp_store_ts_recent(tp);
3663 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3665 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3666 * it can pass through stack. So, the following predicate verifies that
3667 * this segment is not used for anything but congestion avoidance or
3668 * fast retransmit. Moreover, we even are able to eliminate most of such
3669 * second order effects, if we apply some small "replay" window (~RTO)
3670 * to timestamp space.
3672 * All these measures still do not guarantee that we reject wrapped ACKs
3673 * on networks with high bandwidth, when sequence space is recycled fastly,
3674 * but it guarantees that such events will be very rare and do not affect
3675 * connection seriously. This doesn't look nice, but alas, PAWS is really
3678 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3679 * states that events when retransmit arrives after original data are rare.
3680 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3681 * the biggest problem on large power networks even with minor reordering.
3682 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3683 * up to bandwidth of 18Gigabit/sec. 8) ]
3686 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3688 const struct tcp_sock *tp = tcp_sk(sk);
3689 const struct tcphdr *th = tcp_hdr(skb);
3690 u32 seq = TCP_SKB_CB(skb)->seq;
3691 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3693 return (/* 1. Pure ACK with correct sequence number. */
3694 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3696 /* 2. ... and duplicate ACK. */
3697 ack == tp->snd_una &&
3699 /* 3. ... and does not update window. */
3700 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3702 /* 4. ... and sits in replay window. */
3703 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3706 static inline bool tcp_paws_discard(const struct sock *sk,
3707 const struct sk_buff *skb)
3709 const struct tcp_sock *tp = tcp_sk(sk);
3711 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
3712 !tcp_disordered_ack(sk, skb);
3715 /* Check segment sequence number for validity.
3717 * Segment controls are considered valid, if the segment
3718 * fits to the window after truncation to the window. Acceptability
3719 * of data (and SYN, FIN, of course) is checked separately.
3720 * See tcp_data_queue(), for example.
3722 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3723 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3724 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3725 * (borrowed from freebsd)
3728 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
3730 return !before(end_seq, tp->rcv_wup) &&
3731 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3734 /* When we get a reset we do this. */
3735 void tcp_reset(struct sock *sk)
3737 /* We want the right error as BSD sees it (and indeed as we do). */
3738 switch (sk->sk_state) {
3740 sk->sk_err = ECONNREFUSED;
3742 case TCP_CLOSE_WAIT:
3748 sk->sk_err = ECONNRESET;
3750 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3753 if (!sock_flag(sk, SOCK_DEAD))
3754 sk->sk_error_report(sk);
3760 * Process the FIN bit. This now behaves as it is supposed to work
3761 * and the FIN takes effect when it is validly part of sequence
3762 * space. Not before when we get holes.
3764 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3765 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3768 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3769 * close and we go into CLOSING (and later onto TIME-WAIT)
3771 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3773 static void tcp_fin(struct sock *sk)
3775 struct tcp_sock *tp = tcp_sk(sk);
3777 inet_csk_schedule_ack(sk);
3779 sk->sk_shutdown |= RCV_SHUTDOWN;
3780 sock_set_flag(sk, SOCK_DONE);
3782 switch (sk->sk_state) {
3784 case TCP_ESTABLISHED:
3785 /* Move to CLOSE_WAIT */
3786 tcp_set_state(sk, TCP_CLOSE_WAIT);
3787 inet_csk(sk)->icsk_ack.pingpong = 1;
3790 case TCP_CLOSE_WAIT:
3792 /* Received a retransmission of the FIN, do
3797 /* RFC793: Remain in the LAST-ACK state. */
3801 /* This case occurs when a simultaneous close
3802 * happens, we must ack the received FIN and
3803 * enter the CLOSING state.
3806 tcp_set_state(sk, TCP_CLOSING);
3809 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3811 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
3814 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3815 * cases we should never reach this piece of code.
3817 pr_err("%s: Impossible, sk->sk_state=%d\n",
3818 __func__, sk->sk_state);
3822 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3823 * Probably, we should reset in this case. For now drop them.
3825 __skb_queue_purge(&tp->out_of_order_queue);
3826 if (tcp_is_sack(tp))
3827 tcp_sack_reset(&tp->rx_opt);
3830 if (!sock_flag(sk, SOCK_DEAD)) {
3831 sk->sk_state_change(sk);
3833 /* Do not send POLL_HUP for half duplex close. */
3834 if (sk->sk_shutdown == SHUTDOWN_MASK ||
3835 sk->sk_state == TCP_CLOSE)
3836 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
3838 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
3842 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
3845 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
3846 if (before(seq, sp->start_seq))
3847 sp->start_seq = seq;
3848 if (after(end_seq, sp->end_seq))
3849 sp->end_seq = end_seq;
3855 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
3857 struct tcp_sock *tp = tcp_sk(sk);
3859 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3862 if (before(seq, tp->rcv_nxt))
3863 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
3865 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
3867 NET_INC_STATS_BH(sock_net(sk), mib_idx);
3869 tp->rx_opt.dsack = 1;
3870 tp->duplicate_sack[0].start_seq = seq;
3871 tp->duplicate_sack[0].end_seq = end_seq;
3875 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
3877 struct tcp_sock *tp = tcp_sk(sk);
3879 if (!tp->rx_opt.dsack)
3880 tcp_dsack_set(sk, seq, end_seq);
3882 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
3885 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
3887 struct tcp_sock *tp = tcp_sk(sk);
3889 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
3890 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3891 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
3892 tcp_enter_quickack_mode(sk);
3894 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3895 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3897 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
3898 end_seq = tp->rcv_nxt;
3899 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
3906 /* These routines update the SACK block as out-of-order packets arrive or
3907 * in-order packets close up the sequence space.
3909 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
3912 struct tcp_sack_block *sp = &tp->selective_acks[0];
3913 struct tcp_sack_block *swalk = sp + 1;
3915 /* See if the recent change to the first SACK eats into
3916 * or hits the sequence space of other SACK blocks, if so coalesce.
3918 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
3919 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
3922 /* Zap SWALK, by moving every further SACK up by one slot.
3923 * Decrease num_sacks.
3925 tp->rx_opt.num_sacks--;
3926 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
3930 this_sack++, swalk++;
3934 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
3936 struct tcp_sock *tp = tcp_sk(sk);
3937 struct tcp_sack_block *sp = &tp->selective_acks[0];
3938 int cur_sacks = tp->rx_opt.num_sacks;
3944 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
3945 if (tcp_sack_extend(sp, seq, end_seq)) {
3946 /* Rotate this_sack to the first one. */
3947 for (; this_sack > 0; this_sack--, sp--)
3948 swap(*sp, *(sp - 1));
3950 tcp_sack_maybe_coalesce(tp);
3955 /* Could not find an adjacent existing SACK, build a new one,
3956 * put it at the front, and shift everyone else down. We
3957 * always know there is at least one SACK present already here.
3959 * If the sack array is full, forget about the last one.
3961 if (this_sack >= TCP_NUM_SACKS) {
3963 tp->rx_opt.num_sacks--;
3966 for (; this_sack > 0; this_sack--, sp--)
3970 /* Build the new head SACK, and we're done. */
3971 sp->start_seq = seq;
3972 sp->end_seq = end_seq;
3973 tp->rx_opt.num_sacks++;
3976 /* RCV.NXT advances, some SACKs should be eaten. */
3978 static void tcp_sack_remove(struct tcp_sock *tp)
3980 struct tcp_sack_block *sp = &tp->selective_acks[0];
3981 int num_sacks = tp->rx_opt.num_sacks;
3984 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3985 if (skb_queue_empty(&tp->out_of_order_queue)) {
3986 tp->rx_opt.num_sacks = 0;
3990 for (this_sack = 0; this_sack < num_sacks;) {
3991 /* Check if the start of the sack is covered by RCV.NXT. */
3992 if (!before(tp->rcv_nxt, sp->start_seq)) {
3995 /* RCV.NXT must cover all the block! */
3996 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
3998 /* Zap this SACK, by moving forward any other SACKS. */
3999 for (i=this_sack+1; i < num_sacks; i++)
4000 tp->selective_acks[i-1] = tp->selective_acks[i];
4007 tp->rx_opt.num_sacks = num_sacks;
4010 /* This one checks to see if we can put data from the
4011 * out_of_order queue into the receive_queue.
4013 static void tcp_ofo_queue(struct sock *sk)
4015 struct tcp_sock *tp = tcp_sk(sk);
4016 __u32 dsack_high = tp->rcv_nxt;
4017 struct sk_buff *skb;
4019 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4020 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4023 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4024 __u32 dsack = dsack_high;
4025 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4026 dsack_high = TCP_SKB_CB(skb)->end_seq;
4027 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4030 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4031 SOCK_DEBUG(sk, "ofo packet was already received\n");
4032 __skb_unlink(skb, &tp->out_of_order_queue);
4036 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4037 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4038 TCP_SKB_CB(skb)->end_seq);
4040 __skb_unlink(skb, &tp->out_of_order_queue);
4041 __skb_queue_tail(&sk->sk_receive_queue, skb);
4042 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4043 if (tcp_hdr(skb)->fin)
4048 static bool tcp_prune_ofo_queue(struct sock *sk);
4049 static int tcp_prune_queue(struct sock *sk);
4051 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4054 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4055 !sk_rmem_schedule(sk, skb, size)) {
4057 if (tcp_prune_queue(sk) < 0)
4060 if (!sk_rmem_schedule(sk, skb, size)) {
4061 if (!tcp_prune_ofo_queue(sk))
4064 if (!sk_rmem_schedule(sk, skb, size))
4072 * tcp_try_coalesce - try to merge skb to prior one
4075 * @from: buffer to add in queue
4076 * @fragstolen: pointer to boolean
4078 * Before queueing skb @from after @to, try to merge them
4079 * to reduce overall memory use and queue lengths, if cost is small.
4080 * Packets in ofo or receive queues can stay a long time.
4081 * Better try to coalesce them right now to avoid future collapses.
4082 * Returns true if caller should free @from instead of queueing it
4084 static bool tcp_try_coalesce(struct sock *sk,
4086 struct sk_buff *from,
4091 *fragstolen = false;
4093 if (tcp_hdr(from)->fin)
4096 /* Its possible this segment overlaps with prior segment in queue */
4097 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4100 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4103 atomic_add(delta, &sk->sk_rmem_alloc);
4104 sk_mem_charge(sk, delta);
4105 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4106 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4107 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4111 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4113 struct tcp_sock *tp = tcp_sk(sk);
4114 struct sk_buff *skb1;
4117 TCP_ECN_check_ce(tp, skb);
4119 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4120 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFODROP);
4125 /* Disable header prediction. */
4127 inet_csk_schedule_ack(sk);
4129 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4130 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4131 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4133 skb1 = skb_peek_tail(&tp->out_of_order_queue);
4135 /* Initial out of order segment, build 1 SACK. */
4136 if (tcp_is_sack(tp)) {
4137 tp->rx_opt.num_sacks = 1;
4138 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4139 tp->selective_acks[0].end_seq =
4140 TCP_SKB_CB(skb)->end_seq;
4142 __skb_queue_head(&tp->out_of_order_queue, skb);
4146 seq = TCP_SKB_CB(skb)->seq;
4147 end_seq = TCP_SKB_CB(skb)->end_seq;
4149 if (seq == TCP_SKB_CB(skb1)->end_seq) {
4152 if (!tcp_try_coalesce(sk, skb1, skb, &fragstolen)) {
4153 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4155 kfree_skb_partial(skb, fragstolen);
4159 if (!tp->rx_opt.num_sacks ||
4160 tp->selective_acks[0].end_seq != seq)
4163 /* Common case: data arrive in order after hole. */
4164 tp->selective_acks[0].end_seq = end_seq;
4168 /* Find place to insert this segment. */
4170 if (!after(TCP_SKB_CB(skb1)->seq, seq))
4172 if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
4176 skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
4179 /* Do skb overlap to previous one? */
4180 if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4181 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4182 /* All the bits are present. Drop. */
4183 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4186 tcp_dsack_set(sk, seq, end_seq);
4189 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4190 /* Partial overlap. */
4191 tcp_dsack_set(sk, seq,
4192 TCP_SKB_CB(skb1)->end_seq);
4194 if (skb_queue_is_first(&tp->out_of_order_queue,
4198 skb1 = skb_queue_prev(
4199 &tp->out_of_order_queue,
4204 __skb_queue_head(&tp->out_of_order_queue, skb);
4206 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4208 /* And clean segments covered by new one as whole. */
4209 while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
4210 skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
4212 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4214 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4215 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4219 __skb_unlink(skb1, &tp->out_of_order_queue);
4220 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4221 TCP_SKB_CB(skb1)->end_seq);
4222 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4227 if (tcp_is_sack(tp))
4228 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4231 skb_set_owner_r(skb, sk);
4234 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
4238 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4240 __skb_pull(skb, hdrlen);
4242 tcp_try_coalesce(sk, tail, skb, fragstolen)) ? 1 : 0;
4243 tcp_sk(sk)->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4245 __skb_queue_tail(&sk->sk_receive_queue, skb);
4246 skb_set_owner_r(skb, sk);
4251 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4253 struct sk_buff *skb = NULL;
4260 skb = alloc_skb(size + sizeof(*th), sk->sk_allocation);
4264 if (tcp_try_rmem_schedule(sk, skb, size + sizeof(*th)))
4267 th = (struct tcphdr *)skb_put(skb, sizeof(*th));
4268 skb_reset_transport_header(skb);
4269 memset(th, 0, sizeof(*th));
4271 if (memcpy_fromiovec(skb_put(skb, size), msg->msg_iov, size))
4274 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4275 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4276 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4278 if (tcp_queue_rcv(sk, skb, sizeof(*th), &fragstolen)) {
4279 WARN_ON_ONCE(fragstolen); /* should not happen */
4290 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4292 const struct tcphdr *th = tcp_hdr(skb);
4293 struct tcp_sock *tp = tcp_sk(sk);
4295 bool fragstolen = false;
4297 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4301 __skb_pull(skb, th->doff * 4);
4303 TCP_ECN_accept_cwr(tp, skb);
4305 tp->rx_opt.dsack = 0;
4307 /* Queue data for delivery to the user.
4308 * Packets in sequence go to the receive queue.
4309 * Out of sequence packets to the out_of_order_queue.
4311 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4312 if (tcp_receive_window(tp) == 0)
4315 /* Ok. In sequence. In window. */
4316 if (tp->ucopy.task == current &&
4317 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4318 sock_owned_by_user(sk) && !tp->urg_data) {
4319 int chunk = min_t(unsigned int, skb->len,
4322 __set_current_state(TASK_RUNNING);
4325 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
4326 tp->ucopy.len -= chunk;
4327 tp->copied_seq += chunk;
4328 eaten = (chunk == skb->len);
4329 tcp_rcv_space_adjust(sk);
4337 tcp_try_rmem_schedule(sk, skb, skb->truesize))
4340 eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
4342 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4344 tcp_event_data_recv(sk, skb);
4348 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4351 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4352 * gap in queue is filled.
4354 if (skb_queue_empty(&tp->out_of_order_queue))
4355 inet_csk(sk)->icsk_ack.pingpong = 0;
4358 if (tp->rx_opt.num_sacks)
4359 tcp_sack_remove(tp);
4361 tcp_fast_path_check(sk);
4364 kfree_skb_partial(skb, fragstolen);
4365 if (!sock_flag(sk, SOCK_DEAD))
4366 sk->sk_data_ready(sk, 0);
4370 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4371 /* A retransmit, 2nd most common case. Force an immediate ack. */
4372 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4373 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4376 tcp_enter_quickack_mode(sk);
4377 inet_csk_schedule_ack(sk);
4383 /* Out of window. F.e. zero window probe. */
4384 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4387 tcp_enter_quickack_mode(sk);
4389 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4390 /* Partial packet, seq < rcv_next < end_seq */
4391 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4392 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4393 TCP_SKB_CB(skb)->end_seq);
4395 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4397 /* If window is closed, drop tail of packet. But after
4398 * remembering D-SACK for its head made in previous line.
4400 if (!tcp_receive_window(tp))
4405 tcp_data_queue_ofo(sk, skb);
4408 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4409 struct sk_buff_head *list)
4411 struct sk_buff *next = NULL;
4413 if (!skb_queue_is_last(list, skb))
4414 next = skb_queue_next(list, skb);
4416 __skb_unlink(skb, list);
4418 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4423 /* Collapse contiguous sequence of skbs head..tail with
4424 * sequence numbers start..end.
4426 * If tail is NULL, this means until the end of the list.
4428 * Segments with FIN/SYN are not collapsed (only because this
4432 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4433 struct sk_buff *head, struct sk_buff *tail,
4436 struct sk_buff *skb, *n;
4439 /* First, check that queue is collapsible and find
4440 * the point where collapsing can be useful. */
4444 skb_queue_walk_from_safe(list, skb, n) {
4447 /* No new bits? It is possible on ofo queue. */
4448 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4449 skb = tcp_collapse_one(sk, skb, list);
4455 /* The first skb to collapse is:
4457 * - bloated or contains data before "start" or
4458 * overlaps to the next one.
4460 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
4461 (tcp_win_from_space(skb->truesize) > skb->len ||
4462 before(TCP_SKB_CB(skb)->seq, start))) {
4463 end_of_skbs = false;
4467 if (!skb_queue_is_last(list, skb)) {
4468 struct sk_buff *next = skb_queue_next(list, skb);
4470 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
4471 end_of_skbs = false;
4476 /* Decided to skip this, advance start seq. */
4477 start = TCP_SKB_CB(skb)->end_seq;
4479 if (end_of_skbs || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
4482 while (before(start, end)) {
4483 struct sk_buff *nskb;
4484 unsigned int header = skb_headroom(skb);
4485 int copy = SKB_MAX_ORDER(header, 0);
4487 /* Too big header? This can happen with IPv6. */
4490 if (end - start < copy)
4492 nskb = alloc_skb(copy + header, GFP_ATOMIC);
4496 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
4497 skb_set_network_header(nskb, (skb_network_header(skb) -
4499 skb_set_transport_header(nskb, (skb_transport_header(skb) -
4501 skb_reserve(nskb, header);
4502 memcpy(nskb->head, skb->head, header);
4503 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4504 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4505 __skb_queue_before(list, skb, nskb);
4506 skb_set_owner_r(nskb, sk);
4508 /* Copy data, releasing collapsed skbs. */
4510 int offset = start - TCP_SKB_CB(skb)->seq;
4511 int size = TCP_SKB_CB(skb)->end_seq - start;
4515 size = min(copy, size);
4516 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4518 TCP_SKB_CB(nskb)->end_seq += size;
4522 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4523 skb = tcp_collapse_one(sk, skb, list);
4526 tcp_hdr(skb)->syn ||
4534 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4535 * and tcp_collapse() them until all the queue is collapsed.
4537 static void tcp_collapse_ofo_queue(struct sock *sk)
4539 struct tcp_sock *tp = tcp_sk(sk);
4540 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4541 struct sk_buff *head;
4547 start = TCP_SKB_CB(skb)->seq;
4548 end = TCP_SKB_CB(skb)->end_seq;
4552 struct sk_buff *next = NULL;
4554 if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
4555 next = skb_queue_next(&tp->out_of_order_queue, skb);
4558 /* Segment is terminated when we see gap or when
4559 * we are at the end of all the queue. */
4561 after(TCP_SKB_CB(skb)->seq, end) ||
4562 before(TCP_SKB_CB(skb)->end_seq, start)) {
4563 tcp_collapse(sk, &tp->out_of_order_queue,
4564 head, skb, start, end);
4568 /* Start new segment */
4569 start = TCP_SKB_CB(skb)->seq;
4570 end = TCP_SKB_CB(skb)->end_seq;
4572 if (before(TCP_SKB_CB(skb)->seq, start))
4573 start = TCP_SKB_CB(skb)->seq;
4574 if (after(TCP_SKB_CB(skb)->end_seq, end))
4575 end = TCP_SKB_CB(skb)->end_seq;
4581 * Purge the out-of-order queue.
4582 * Return true if queue was pruned.
4584 static bool tcp_prune_ofo_queue(struct sock *sk)
4586 struct tcp_sock *tp = tcp_sk(sk);
4589 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4590 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
4591 __skb_queue_purge(&tp->out_of_order_queue);
4593 /* Reset SACK state. A conforming SACK implementation will
4594 * do the same at a timeout based retransmit. When a connection
4595 * is in a sad state like this, we care only about integrity
4596 * of the connection not performance.
4598 if (tp->rx_opt.sack_ok)
4599 tcp_sack_reset(&tp->rx_opt);
4606 /* Reduce allocated memory if we can, trying to get
4607 * the socket within its memory limits again.
4609 * Return less than zero if we should start dropping frames
4610 * until the socket owning process reads some of the data
4611 * to stabilize the situation.
4613 static int tcp_prune_queue(struct sock *sk)
4615 struct tcp_sock *tp = tcp_sk(sk);
4617 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4619 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
4621 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4622 tcp_clamp_window(sk);
4623 else if (sk_under_memory_pressure(sk))
4624 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4626 tcp_collapse_ofo_queue(sk);
4627 if (!skb_queue_empty(&sk->sk_receive_queue))
4628 tcp_collapse(sk, &sk->sk_receive_queue,
4629 skb_peek(&sk->sk_receive_queue),
4631 tp->copied_seq, tp->rcv_nxt);
4634 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4637 /* Collapsing did not help, destructive actions follow.
4638 * This must not ever occur. */
4640 tcp_prune_ofo_queue(sk);
4642 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4645 /* If we are really being abused, tell the caller to silently
4646 * drop receive data on the floor. It will get retransmitted
4647 * and hopefully then we'll have sufficient space.
4649 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
4651 /* Massive buffer overcommit. */
4656 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4657 * As additional protections, we do not touch cwnd in retransmission phases,
4658 * and if application hit its sndbuf limit recently.
4660 void tcp_cwnd_application_limited(struct sock *sk)
4662 struct tcp_sock *tp = tcp_sk(sk);
4664 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
4665 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4666 /* Limited by application or receiver window. */
4667 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4668 u32 win_used = max(tp->snd_cwnd_used, init_win);
4669 if (win_used < tp->snd_cwnd) {
4670 tp->snd_ssthresh = tcp_current_ssthresh(sk);
4671 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4673 tp->snd_cwnd_used = 0;
4675 tp->snd_cwnd_stamp = tcp_time_stamp;
4678 static bool tcp_should_expand_sndbuf(const struct sock *sk)
4680 const struct tcp_sock *tp = tcp_sk(sk);
4682 /* If the user specified a specific send buffer setting, do
4685 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4688 /* If we are under global TCP memory pressure, do not expand. */
4689 if (sk_under_memory_pressure(sk))
4692 /* If we are under soft global TCP memory pressure, do not expand. */
4693 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
4696 /* If we filled the congestion window, do not expand. */
4697 if (tp->packets_out >= tp->snd_cwnd)
4703 /* When incoming ACK allowed to free some skb from write_queue,
4704 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4705 * on the exit from tcp input handler.
4707 * PROBLEM: sndbuf expansion does not work well with largesend.
4709 static void tcp_new_space(struct sock *sk)
4711 struct tcp_sock *tp = tcp_sk(sk);
4713 if (tcp_should_expand_sndbuf(sk)) {
4714 int sndmem = SKB_TRUESIZE(max_t(u32,
4715 tp->rx_opt.mss_clamp,
4718 int demanded = max_t(unsigned int, tp->snd_cwnd,
4719 tp->reordering + 1);
4720 sndmem *= 2 * demanded;
4721 if (sndmem > sk->sk_sndbuf)
4722 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
4723 tp->snd_cwnd_stamp = tcp_time_stamp;
4726 sk->sk_write_space(sk);
4729 static void tcp_check_space(struct sock *sk)
4731 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4732 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4733 if (sk->sk_socket &&
4734 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4739 static inline void tcp_data_snd_check(struct sock *sk)
4741 tcp_push_pending_frames(sk);
4742 tcp_check_space(sk);
4746 * Check if sending an ack is needed.
4748 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4750 struct tcp_sock *tp = tcp_sk(sk);
4752 /* More than one full frame received... */
4753 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
4754 /* ... and right edge of window advances far enough.
4755 * (tcp_recvmsg() will send ACK otherwise). Or...
4757 __tcp_select_window(sk) >= tp->rcv_wnd) ||
4758 /* We ACK each frame or... */
4759 tcp_in_quickack_mode(sk) ||
4760 /* We have out of order data. */
4761 (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
4762 /* Then ack it now */
4765 /* Else, send delayed ack. */
4766 tcp_send_delayed_ack(sk);
4770 static inline void tcp_ack_snd_check(struct sock *sk)
4772 if (!inet_csk_ack_scheduled(sk)) {
4773 /* We sent a data segment already. */
4776 __tcp_ack_snd_check(sk, 1);
4780 * This routine is only called when we have urgent data
4781 * signaled. Its the 'slow' part of tcp_urg. It could be
4782 * moved inline now as tcp_urg is only called from one
4783 * place. We handle URGent data wrong. We have to - as
4784 * BSD still doesn't use the correction from RFC961.
4785 * For 1003.1g we should support a new option TCP_STDURG to permit
4786 * either form (or just set the sysctl tcp_stdurg).
4789 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
4791 struct tcp_sock *tp = tcp_sk(sk);
4792 u32 ptr = ntohs(th->urg_ptr);
4794 if (ptr && !sysctl_tcp_stdurg)
4796 ptr += ntohl(th->seq);
4798 /* Ignore urgent data that we've already seen and read. */
4799 if (after(tp->copied_seq, ptr))
4802 /* Do not replay urg ptr.
4804 * NOTE: interesting situation not covered by specs.
4805 * Misbehaving sender may send urg ptr, pointing to segment,
4806 * which we already have in ofo queue. We are not able to fetch
4807 * such data and will stay in TCP_URG_NOTYET until will be eaten
4808 * by recvmsg(). Seems, we are not obliged to handle such wicked
4809 * situations. But it is worth to think about possibility of some
4810 * DoSes using some hypothetical application level deadlock.
4812 if (before(ptr, tp->rcv_nxt))
4815 /* Do we already have a newer (or duplicate) urgent pointer? */
4816 if (tp->urg_data && !after(ptr, tp->urg_seq))
4819 /* Tell the world about our new urgent pointer. */
4822 /* We may be adding urgent data when the last byte read was
4823 * urgent. To do this requires some care. We cannot just ignore
4824 * tp->copied_seq since we would read the last urgent byte again
4825 * as data, nor can we alter copied_seq until this data arrives
4826 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4828 * NOTE. Double Dutch. Rendering to plain English: author of comment
4829 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4830 * and expect that both A and B disappear from stream. This is _wrong_.
4831 * Though this happens in BSD with high probability, this is occasional.
4832 * Any application relying on this is buggy. Note also, that fix "works"
4833 * only in this artificial test. Insert some normal data between A and B and we will
4834 * decline of BSD again. Verdict: it is better to remove to trap
4837 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
4838 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
4839 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4841 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
4842 __skb_unlink(skb, &sk->sk_receive_queue);
4847 tp->urg_data = TCP_URG_NOTYET;
4850 /* Disable header prediction. */
4854 /* This is the 'fast' part of urgent handling. */
4855 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
4857 struct tcp_sock *tp = tcp_sk(sk);
4859 /* Check if we get a new urgent pointer - normally not. */
4861 tcp_check_urg(sk, th);
4863 /* Do we wait for any urgent data? - normally not... */
4864 if (tp->urg_data == TCP_URG_NOTYET) {
4865 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
4868 /* Is the urgent pointer pointing into this packet? */
4869 if (ptr < skb->len) {
4871 if (skb_copy_bits(skb, ptr, &tmp, 1))
4873 tp->urg_data = TCP_URG_VALID | tmp;
4874 if (!sock_flag(sk, SOCK_DEAD))
4875 sk->sk_data_ready(sk, 0);
4880 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4882 struct tcp_sock *tp = tcp_sk(sk);
4883 int chunk = skb->len - hlen;
4887 if (skb_csum_unnecessary(skb))
4888 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
4890 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
4894 tp->ucopy.len -= chunk;
4895 tp->copied_seq += chunk;
4896 tcp_rcv_space_adjust(sk);
4903 static __sum16 __tcp_checksum_complete_user(struct sock *sk,
4904 struct sk_buff *skb)
4908 if (sock_owned_by_user(sk)) {
4910 result = __tcp_checksum_complete(skb);
4913 result = __tcp_checksum_complete(skb);
4918 static inline bool tcp_checksum_complete_user(struct sock *sk,
4919 struct sk_buff *skb)
4921 return !skb_csum_unnecessary(skb) &&
4922 __tcp_checksum_complete_user(sk, skb);
4925 #ifdef CONFIG_NET_DMA
4926 static bool tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb,
4929 struct tcp_sock *tp = tcp_sk(sk);
4930 int chunk = skb->len - hlen;
4932 bool copied_early = false;
4934 if (tp->ucopy.wakeup)
4937 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
4938 tp->ucopy.dma_chan = net_dma_find_channel();
4940 if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
4942 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
4944 tp->ucopy.iov, chunk,
4945 tp->ucopy.pinned_list);
4950 tp->ucopy.dma_cookie = dma_cookie;
4951 copied_early = true;
4953 tp->ucopy.len -= chunk;
4954 tp->copied_seq += chunk;
4955 tcp_rcv_space_adjust(sk);
4957 if ((tp->ucopy.len == 0) ||
4958 (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
4959 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
4960 tp->ucopy.wakeup = 1;
4961 sk->sk_data_ready(sk, 0);
4963 } else if (chunk > 0) {
4964 tp->ucopy.wakeup = 1;
4965 sk->sk_data_ready(sk, 0);
4968 return copied_early;
4970 #endif /* CONFIG_NET_DMA */
4972 /* Does PAWS and seqno based validation of an incoming segment, flags will
4973 * play significant role here.
4975 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
4976 const struct tcphdr *th, int syn_inerr)
4978 struct tcp_sock *tp = tcp_sk(sk);
4980 /* RFC1323: H1. Apply PAWS check first. */
4981 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4982 tcp_paws_discard(sk, skb)) {
4984 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
4985 tcp_send_dupack(sk, skb);
4988 /* Reset is accepted even if it did not pass PAWS. */
4991 /* Step 1: check sequence number */
4992 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4993 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4994 * (RST) segments are validated by checking their SEQ-fields."
4995 * And page 69: "If an incoming segment is not acceptable,
4996 * an acknowledgment should be sent in reply (unless the RST
4997 * bit is set, if so drop the segment and return)".
5002 tcp_send_dupack(sk, skb);
5007 /* Step 2: check RST bit */
5010 * If sequence number exactly matches RCV.NXT, then
5011 * RESET the connection
5013 * Send a challenge ACK
5015 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt)
5018 tcp_send_challenge_ack(sk);
5022 /* step 3: check security and precedence [ignored] */
5024 /* step 4: Check for a SYN
5025 * RFC 5691 4.2 : Send a challenge ack
5030 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5031 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5032 tcp_send_challenge_ack(sk);
5044 * TCP receive function for the ESTABLISHED state.
5046 * It is split into a fast path and a slow path. The fast path is
5048 * - A zero window was announced from us - zero window probing
5049 * is only handled properly in the slow path.
5050 * - Out of order segments arrived.
5051 * - Urgent data is expected.
5052 * - There is no buffer space left
5053 * - Unexpected TCP flags/window values/header lengths are received
5054 * (detected by checking the TCP header against pred_flags)
5055 * - Data is sent in both directions. Fast path only supports pure senders
5056 * or pure receivers (this means either the sequence number or the ack
5057 * value must stay constant)
5058 * - Unexpected TCP option.
5060 * When these conditions are not satisfied it drops into a standard
5061 * receive procedure patterned after RFC793 to handle all cases.
5062 * The first three cases are guaranteed by proper pred_flags setting,
5063 * the rest is checked inline. Fast processing is turned on in
5064 * tcp_data_queue when everything is OK.
5066 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5067 const struct tcphdr *th, unsigned int len)
5069 struct tcp_sock *tp = tcp_sk(sk);
5071 if (unlikely(sk->sk_rx_dst == NULL))
5072 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5074 * Header prediction.
5075 * The code loosely follows the one in the famous
5076 * "30 instruction TCP receive" Van Jacobson mail.
5078 * Van's trick is to deposit buffers into socket queue
5079 * on a device interrupt, to call tcp_recv function
5080 * on the receive process context and checksum and copy
5081 * the buffer to user space. smart...
5083 * Our current scheme is not silly either but we take the
5084 * extra cost of the net_bh soft interrupt processing...
5085 * We do checksum and copy also but from device to kernel.
5088 tp->rx_opt.saw_tstamp = 0;
5090 /* pred_flags is 0xS?10 << 16 + snd_wnd
5091 * if header_prediction is to be made
5092 * 'S' will always be tp->tcp_header_len >> 2
5093 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5094 * turn it off (when there are holes in the receive
5095 * space for instance)
5096 * PSH flag is ignored.
5099 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5100 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5101 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5102 int tcp_header_len = tp->tcp_header_len;
5104 /* Timestamp header prediction: tcp_header_len
5105 * is automatically equal to th->doff*4 due to pred_flags
5109 /* Check timestamp */
5110 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5111 /* No? Slow path! */
5112 if (!tcp_parse_aligned_timestamp(tp, th))
5115 /* If PAWS failed, check it more carefully in slow path */
5116 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5119 /* DO NOT update ts_recent here, if checksum fails
5120 * and timestamp was corrupted part, it will result
5121 * in a hung connection since we will drop all
5122 * future packets due to the PAWS test.
5126 if (len <= tcp_header_len) {
5127 /* Bulk data transfer: sender */
5128 if (len == tcp_header_len) {
5129 /* Predicted packet is in window by definition.
5130 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5131 * Hence, check seq<=rcv_wup reduces to:
5133 if (tcp_header_len ==
5134 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5135 tp->rcv_nxt == tp->rcv_wup)
5136 tcp_store_ts_recent(tp);
5138 /* We know that such packets are checksummed
5141 tcp_ack(sk, skb, 0);
5143 tcp_data_snd_check(sk);
5145 } else { /* Header too small */
5146 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5151 int copied_early = 0;
5152 bool fragstolen = false;
5154 if (tp->copied_seq == tp->rcv_nxt &&
5155 len - tcp_header_len <= tp->ucopy.len) {
5156 #ifdef CONFIG_NET_DMA
5157 if (tp->ucopy.task == current &&
5158 sock_owned_by_user(sk) &&
5159 tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
5164 if (tp->ucopy.task == current &&
5165 sock_owned_by_user(sk) && !copied_early) {
5166 __set_current_state(TASK_RUNNING);
5168 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
5172 /* Predicted packet is in window by definition.
5173 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5174 * Hence, check seq<=rcv_wup reduces to:
5176 if (tcp_header_len ==
5177 (sizeof(struct tcphdr) +
5178 TCPOLEN_TSTAMP_ALIGNED) &&
5179 tp->rcv_nxt == tp->rcv_wup)
5180 tcp_store_ts_recent(tp);
5182 tcp_rcv_rtt_measure_ts(sk, skb);
5184 __skb_pull(skb, tcp_header_len);
5185 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5186 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
5189 tcp_cleanup_rbuf(sk, skb->len);
5192 if (tcp_checksum_complete_user(sk, skb))
5195 if ((int)skb->truesize > sk->sk_forward_alloc)
5198 /* Predicted packet is in window by definition.
5199 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5200 * Hence, check seq<=rcv_wup reduces to:
5202 if (tcp_header_len ==
5203 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5204 tp->rcv_nxt == tp->rcv_wup)
5205 tcp_store_ts_recent(tp);
5207 tcp_rcv_rtt_measure_ts(sk, skb);
5209 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
5211 /* Bulk data transfer: receiver */
5212 eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5216 tcp_event_data_recv(sk, skb);
5218 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5219 /* Well, only one small jumplet in fast path... */
5220 tcp_ack(sk, skb, FLAG_DATA);
5221 tcp_data_snd_check(sk);
5222 if (!inet_csk_ack_scheduled(sk))
5226 if (!copied_early || tp->rcv_nxt != tp->rcv_wup)
5227 __tcp_ack_snd_check(sk, 0);
5229 #ifdef CONFIG_NET_DMA
5231 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
5235 kfree_skb_partial(skb, fragstolen);
5236 sk->sk_data_ready(sk, 0);
5242 if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
5245 if (!th->ack && !th->rst)
5249 * Standard slow path.
5252 if (!tcp_validate_incoming(sk, skb, th, 1))
5256 if (tcp_ack(sk, skb, FLAG_SLOWPATH) < 0)
5259 /* ts_recent update must be made after we are sure that the packet
5262 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
5264 tcp_rcv_rtt_measure_ts(sk, skb);
5266 /* Process urgent data. */
5267 tcp_urg(sk, skb, th);
5269 /* step 7: process the segment text */
5270 tcp_data_queue(sk, skb);
5272 tcp_data_snd_check(sk);
5273 tcp_ack_snd_check(sk);
5277 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5283 EXPORT_SYMBOL(tcp_rcv_established);
5285 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5287 struct tcp_sock *tp = tcp_sk(sk);
5288 struct inet_connection_sock *icsk = inet_csk(sk);
5290 tcp_set_state(sk, TCP_ESTABLISHED);
5293 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5294 security_inet_conn_established(sk, skb);
5297 /* Make sure socket is routed, for correct metrics. */
5298 icsk->icsk_af_ops->rebuild_header(sk);
5300 tcp_init_metrics(sk);
5302 tcp_init_congestion_control(sk);
5304 /* Prevent spurious tcp_cwnd_restart() on first data
5307 tp->lsndtime = tcp_time_stamp;
5309 tcp_init_buffer_space(sk);
5311 if (sock_flag(sk, SOCK_KEEPOPEN))
5312 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5314 if (!tp->rx_opt.snd_wscale)
5315 __tcp_fast_path_on(tp, tp->snd_wnd);
5319 if (!sock_flag(sk, SOCK_DEAD)) {
5320 sk->sk_state_change(sk);
5321 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5325 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5326 struct tcp_fastopen_cookie *cookie)
5328 struct tcp_sock *tp = tcp_sk(sk);
5329 struct sk_buff *data = tp->syn_data ? tcp_write_queue_head(sk) : NULL;
5330 u16 mss = tp->rx_opt.mss_clamp;
5333 if (mss == tp->rx_opt.user_mss) {
5334 struct tcp_options_received opt;
5336 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5337 tcp_clear_options(&opt);
5338 opt.user_mss = opt.mss_clamp = 0;
5339 tcp_parse_options(synack, &opt, 0, NULL);
5340 mss = opt.mss_clamp;
5343 if (!tp->syn_fastopen) /* Ignore an unsolicited cookie */
5346 /* The SYN-ACK neither has cookie nor acknowledges the data. Presumably
5347 * the remote receives only the retransmitted (regular) SYNs: either
5348 * the original SYN-data or the corresponding SYN-ACK is lost.
5350 syn_drop = (cookie->len <= 0 && data && tp->total_retrans);
5352 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop);
5354 if (data) { /* Retransmit unacked data in SYN */
5355 tcp_for_write_queue_from(data, sk) {
5356 if (data == tcp_send_head(sk) ||
5357 __tcp_retransmit_skb(sk, data))
5363 tp->syn_data_acked = tp->syn_data;
5367 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5368 const struct tcphdr *th, unsigned int len)
5370 struct inet_connection_sock *icsk = inet_csk(sk);
5371 struct tcp_sock *tp = tcp_sk(sk);
5372 struct tcp_fastopen_cookie foc = { .len = -1 };
5373 int saved_clamp = tp->rx_opt.mss_clamp;
5375 tcp_parse_options(skb, &tp->rx_opt, 0, &foc);
5376 if (tp->rx_opt.saw_tstamp)
5377 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
5381 * "If the state is SYN-SENT then
5382 * first check the ACK bit
5383 * If the ACK bit is set
5384 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5385 * a reset (unless the RST bit is set, if so drop
5386 * the segment and return)"
5388 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5389 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
5390 goto reset_and_undo;
5392 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5393 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5395 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
5396 goto reset_and_undo;
5399 /* Now ACK is acceptable.
5401 * "If the RST bit is set
5402 * If the ACK was acceptable then signal the user "error:
5403 * connection reset", drop the segment, enter CLOSED state,
5404 * delete TCB, and return."
5413 * "fifth, if neither of the SYN or RST bits is set then
5414 * drop the segment and return."
5420 goto discard_and_undo;
5423 * "If the SYN bit is on ...
5424 * are acceptable then ...
5425 * (our SYN has been ACKed), change the connection
5426 * state to ESTABLISHED..."
5429 TCP_ECN_rcv_synack(tp, th);
5431 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5432 tcp_ack(sk, skb, FLAG_SLOWPATH);
5434 /* Ok.. it's good. Set up sequence numbers and
5435 * move to established.
5437 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5438 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5440 /* RFC1323: The window in SYN & SYN/ACK segments is
5443 tp->snd_wnd = ntohs(th->window);
5445 if (!tp->rx_opt.wscale_ok) {
5446 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5447 tp->window_clamp = min(tp->window_clamp, 65535U);
5450 if (tp->rx_opt.saw_tstamp) {
5451 tp->rx_opt.tstamp_ok = 1;
5452 tp->tcp_header_len =
5453 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5454 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5455 tcp_store_ts_recent(tp);
5457 tp->tcp_header_len = sizeof(struct tcphdr);
5460 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5461 tcp_enable_fack(tp);
5464 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5465 tcp_initialize_rcv_mss(sk);
5467 /* Remember, tcp_poll() does not lock socket!
5468 * Change state from SYN-SENT only after copied_seq
5469 * is initialized. */
5470 tp->copied_seq = tp->rcv_nxt;
5474 tcp_finish_connect(sk, skb);
5476 if ((tp->syn_fastopen || tp->syn_data) &&
5477 tcp_rcv_fastopen_synack(sk, skb, &foc))
5480 if (sk->sk_write_pending ||
5481 icsk->icsk_accept_queue.rskq_defer_accept ||
5482 icsk->icsk_ack.pingpong) {
5483 /* Save one ACK. Data will be ready after
5484 * several ticks, if write_pending is set.
5486 * It may be deleted, but with this feature tcpdumps
5487 * look so _wonderfully_ clever, that I was not able
5488 * to stand against the temptation 8) --ANK
5490 inet_csk_schedule_ack(sk);
5491 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5492 tcp_enter_quickack_mode(sk);
5493 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5494 TCP_DELACK_MAX, TCP_RTO_MAX);
5505 /* No ACK in the segment */
5509 * "If the RST bit is set
5511 * Otherwise (no ACK) drop the segment and return."
5514 goto discard_and_undo;
5518 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5519 tcp_paws_reject(&tp->rx_opt, 0))
5520 goto discard_and_undo;
5523 /* We see SYN without ACK. It is attempt of
5524 * simultaneous connect with crossed SYNs.
5525 * Particularly, it can be connect to self.
5527 tcp_set_state(sk, TCP_SYN_RECV);
5529 if (tp->rx_opt.saw_tstamp) {
5530 tp->rx_opt.tstamp_ok = 1;
5531 tcp_store_ts_recent(tp);
5532 tp->tcp_header_len =
5533 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5535 tp->tcp_header_len = sizeof(struct tcphdr);
5538 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5539 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5541 /* RFC1323: The window in SYN & SYN/ACK segments is
5544 tp->snd_wnd = ntohs(th->window);
5545 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5546 tp->max_window = tp->snd_wnd;
5548 TCP_ECN_rcv_syn(tp, th);
5551 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5552 tcp_initialize_rcv_mss(sk);
5554 tcp_send_synack(sk);
5556 /* Note, we could accept data and URG from this segment.
5557 * There are no obstacles to make this (except that we must
5558 * either change tcp_recvmsg() to prevent it from returning data
5559 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5561 * However, if we ignore data in ACKless segments sometimes,
5562 * we have no reasons to accept it sometimes.
5563 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5564 * is not flawless. So, discard packet for sanity.
5565 * Uncomment this return to process the data.
5572 /* "fifth, if neither of the SYN or RST bits is set then
5573 * drop the segment and return."
5577 tcp_clear_options(&tp->rx_opt);
5578 tp->rx_opt.mss_clamp = saved_clamp;
5582 tcp_clear_options(&tp->rx_opt);
5583 tp->rx_opt.mss_clamp = saved_clamp;
5588 * This function implements the receiving procedure of RFC 793 for
5589 * all states except ESTABLISHED and TIME_WAIT.
5590 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5591 * address independent.
5594 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5595 const struct tcphdr *th, unsigned int len)
5597 struct tcp_sock *tp = tcp_sk(sk);
5598 struct inet_connection_sock *icsk = inet_csk(sk);
5599 struct request_sock *req;
5602 tp->rx_opt.saw_tstamp = 0;
5604 switch (sk->sk_state) {
5618 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5621 /* Now we have several options: In theory there is
5622 * nothing else in the frame. KA9Q has an option to
5623 * send data with the syn, BSD accepts data with the
5624 * syn up to the [to be] advertised window and
5625 * Solaris 2.1 gives you a protocol error. For now
5626 * we just ignore it, that fits the spec precisely
5627 * and avoids incompatibilities. It would be nice in
5628 * future to drop through and process the data.
5630 * Now that TTCP is starting to be used we ought to
5632 * But, this leaves one open to an easy denial of
5633 * service attack, and SYN cookies can't defend
5634 * against this problem. So, we drop the data
5635 * in the interest of security over speed unless
5636 * it's still in use.
5644 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5648 /* Do step6 onward by hand. */
5649 tcp_urg(sk, skb, th);
5651 tcp_data_snd_check(sk);
5655 req = tp->fastopen_rsk;
5657 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
5658 sk->sk_state != TCP_FIN_WAIT1);
5660 if (tcp_check_req(sk, skb, req, NULL, true) == NULL)
5664 if (!th->ack && !th->rst)
5667 if (!tcp_validate_incoming(sk, skb, th, 0))
5670 /* step 5: check the ACK field */
5672 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH) > 0;
5674 switch (sk->sk_state) {
5677 /* Once we leave TCP_SYN_RECV, we no longer
5678 * need req so release it.
5681 tcp_synack_rtt_meas(sk, req);
5682 tp->total_retrans = req->num_retrans;
5684 reqsk_fastopen_remove(sk, req, false);
5686 /* Make sure socket is routed, for
5689 icsk->icsk_af_ops->rebuild_header(sk);
5690 tcp_init_congestion_control(sk);
5693 tcp_init_buffer_space(sk);
5694 tp->copied_seq = tp->rcv_nxt;
5697 tcp_set_state(sk, TCP_ESTABLISHED);
5698 sk->sk_state_change(sk);
5700 /* Note, that this wakeup is only for marginal
5701 * crossed SYN case. Passively open sockets
5702 * are not waked up, because sk->sk_sleep ==
5703 * NULL and sk->sk_socket == NULL.
5707 SOCK_WAKE_IO, POLL_OUT);
5709 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5710 tp->snd_wnd = ntohs(th->window) <<
5711 tp->rx_opt.snd_wscale;
5712 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5714 if (tp->rx_opt.tstamp_ok)
5715 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5718 /* Re-arm the timer because data may
5719 * have been sent out. This is similar
5720 * to the regular data transmission case
5721 * when new data has just been ack'ed.
5723 * (TFO) - we could try to be more
5724 * aggressive and retranmitting any data
5725 * sooner based on when they were sent
5730 tcp_init_metrics(sk);
5732 /* Prevent spurious tcp_cwnd_restart() on
5733 * first data packet.
5735 tp->lsndtime = tcp_time_stamp;
5737 tcp_initialize_rcv_mss(sk);
5738 tcp_fast_path_on(tp);
5745 /* If we enter the TCP_FIN_WAIT1 state and we are a
5746 * Fast Open socket and this is the first acceptable
5747 * ACK we have received, this would have acknowledged
5748 * our SYNACK so stop the SYNACK timer.
5751 /* Return RST if ack_seq is invalid.
5752 * Note that RFC793 only says to generate a
5753 * DUPACK for it but for TCP Fast Open it seems
5754 * better to treat this case like TCP_SYN_RECV
5759 /* We no longer need the request sock. */
5760 reqsk_fastopen_remove(sk, req, false);
5763 if (tp->snd_una == tp->write_seq) {
5764 struct dst_entry *dst;
5766 tcp_set_state(sk, TCP_FIN_WAIT2);
5767 sk->sk_shutdown |= SEND_SHUTDOWN;
5769 dst = __sk_dst_get(sk);
5773 if (!sock_flag(sk, SOCK_DEAD))
5774 /* Wake up lingering close() */
5775 sk->sk_state_change(sk);
5779 if (tp->linger2 < 0 ||
5780 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5781 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5783 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5787 tmo = tcp_fin_time(sk);
5788 if (tmo > TCP_TIMEWAIT_LEN) {
5789 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5790 } else if (th->fin || sock_owned_by_user(sk)) {
5791 /* Bad case. We could lose such FIN otherwise.
5792 * It is not a big problem, but it looks confusing
5793 * and not so rare event. We still can lose it now,
5794 * if it spins in bh_lock_sock(), but it is really
5797 inet_csk_reset_keepalive_timer(sk, tmo);
5799 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5807 if (tp->snd_una == tp->write_seq) {
5808 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5814 if (tp->snd_una == tp->write_seq) {
5815 tcp_update_metrics(sk);
5823 /* ts_recent update must be made after we are sure that the packet
5826 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
5828 /* step 6: check the URG bit */
5829 tcp_urg(sk, skb, th);
5831 /* step 7: process the segment text */
5832 switch (sk->sk_state) {
5833 case TCP_CLOSE_WAIT:
5836 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5840 /* RFC 793 says to queue data in these states,
5841 * RFC 1122 says we MUST send a reset.
5842 * BSD 4.4 also does reset.
5844 if (sk->sk_shutdown & RCV_SHUTDOWN) {
5845 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5846 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5847 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5853 case TCP_ESTABLISHED:
5854 tcp_data_queue(sk, skb);
5859 /* tcp_data could move socket to TIME-WAIT */
5860 if (sk->sk_state != TCP_CLOSE) {
5861 tcp_data_snd_check(sk);
5862 tcp_ack_snd_check(sk);
5871 EXPORT_SYMBOL(tcp_rcv_state_process);