2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
23 * Pedro Roque : Fast Retransmit/Recovery.
25 * Retransmit queue handled by TCP.
26 * Better retransmit timer handling.
27 * New congestion avoidance.
31 * Eric : Fast Retransmit.
32 * Randy Scott : MSS option defines.
33 * Eric Schenk : Fixes to slow start algorithm.
34 * Eric Schenk : Yet another double ACK bug.
35 * Eric Schenk : Delayed ACK bug fixes.
36 * Eric Schenk : Floyd style fast retrans war avoidance.
37 * David S. Miller : Don't allow zero congestion window.
38 * Eric Schenk : Fix retransmitter so that it sends
39 * next packet on ack of previous packet.
40 * Andi Kleen : Moved open_request checking here
41 * and process RSTs for open_requests.
42 * Andi Kleen : Better prune_queue, and other fixes.
43 * Andrey Savochkin: Fix RTT measurements in the presence of
45 * Andrey Savochkin: Check sequence numbers correctly when
46 * removing SACKs due to in sequence incoming
48 * Andi Kleen: Make sure we never ack data there is not
49 * enough room for. Also make this condition
50 * a fatal error if it might still happen.
51 * Andi Kleen: Add tcp_measure_rcv_mss to make
52 * connections with MSS<min(MTU,ann. MSS)
53 * work without delayed acks.
54 * Andi Kleen: Process packets with PSH set in the
56 * J Hadi Salim: ECN support
59 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
60 * engine. Lots of bugs are found.
61 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
64 #define pr_fmt(fmt) "TCP: " fmt
67 #include <linux/slab.h>
68 #include <linux/module.h>
69 #include <linux/sysctl.h>
70 #include <linux/kernel.h>
71 #include <linux/prefetch.h>
74 #include <net/inet_common.h>
75 #include <linux/ipsec.h>
76 #include <asm/unaligned.h>
77 #include <linux/errqueue.h>
79 int sysctl_tcp_timestamps __read_mostly = 1;
80 int sysctl_tcp_window_scaling __read_mostly = 1;
81 int sysctl_tcp_sack __read_mostly = 1;
82 int sysctl_tcp_fack __read_mostly = 1;
83 int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
84 EXPORT_SYMBOL(sysctl_tcp_reordering);
85 int sysctl_tcp_dsack __read_mostly = 1;
86 int sysctl_tcp_app_win __read_mostly = 31;
87 int sysctl_tcp_adv_win_scale __read_mostly = 1;
88 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
90 /* rfc5961 challenge ack rate limiting */
91 int sysctl_tcp_challenge_ack_limit = 100;
93 int sysctl_tcp_stdurg __read_mostly;
94 int sysctl_tcp_rfc1337 __read_mostly;
95 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
96 int sysctl_tcp_frto __read_mostly = 2;
98 int sysctl_tcp_thin_dupack __read_mostly;
100 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
101 int sysctl_tcp_early_retrans __read_mostly = 3;
103 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
104 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
105 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
106 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
107 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
108 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
109 #define FLAG_ECE 0x40 /* ECE in this ACK */
110 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
111 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
112 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
113 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
114 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
115 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
117 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
118 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
119 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
120 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
122 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
123 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
125 /* Adapt the MSS value used to make delayed ack decision to the
128 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
130 struct inet_connection_sock *icsk = inet_csk(sk);
131 const unsigned int lss = icsk->icsk_ack.last_seg_size;
134 icsk->icsk_ack.last_seg_size = 0;
136 /* skb->len may jitter because of SACKs, even if peer
137 * sends good full-sized frames.
139 len = skb_shinfo(skb)->gso_size ? : skb->len;
140 if (len >= icsk->icsk_ack.rcv_mss) {
141 icsk->icsk_ack.rcv_mss = len;
143 /* Otherwise, we make more careful check taking into account,
144 * that SACKs block is variable.
146 * "len" is invariant segment length, including TCP header.
148 len += skb->data - skb_transport_header(skb);
149 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
150 /* If PSH is not set, packet should be
151 * full sized, provided peer TCP is not badly broken.
152 * This observation (if it is correct 8)) allows
153 * to handle super-low mtu links fairly.
155 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
156 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
157 /* Subtract also invariant (if peer is RFC compliant),
158 * tcp header plus fixed timestamp option length.
159 * Resulting "len" is MSS free of SACK jitter.
161 len -= tcp_sk(sk)->tcp_header_len;
162 icsk->icsk_ack.last_seg_size = len;
164 icsk->icsk_ack.rcv_mss = len;
168 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
169 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
170 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
174 static void tcp_incr_quickack(struct sock *sk)
176 struct inet_connection_sock *icsk = inet_csk(sk);
177 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
181 if (quickacks > icsk->icsk_ack.quick)
182 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
185 static void tcp_enter_quickack_mode(struct sock *sk)
187 struct inet_connection_sock *icsk = inet_csk(sk);
188 tcp_incr_quickack(sk);
189 icsk->icsk_ack.pingpong = 0;
190 icsk->icsk_ack.ato = TCP_ATO_MIN;
193 /* Send ACKs quickly, if "quick" count is not exhausted
194 * and the session is not interactive.
197 static inline bool tcp_in_quickack_mode(const struct sock *sk)
199 const struct inet_connection_sock *icsk = inet_csk(sk);
201 return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
204 static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
206 if (tp->ecn_flags & TCP_ECN_OK)
207 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
210 static void tcp_ecn_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
212 if (tcp_hdr(skb)->cwr)
213 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
216 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
218 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
221 static void __tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
223 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
224 case INET_ECN_NOT_ECT:
225 /* Funny extension: if ECT is not set on a segment,
226 * and we already seen ECT on a previous segment,
227 * it is probably a retransmit.
229 if (tp->ecn_flags & TCP_ECN_SEEN)
230 tcp_enter_quickack_mode((struct sock *)tp);
233 if (tcp_ca_needs_ecn((struct sock *)tp))
234 tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_IS_CE);
236 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
237 /* Better not delay acks, sender can have a very low cwnd */
238 tcp_enter_quickack_mode((struct sock *)tp);
239 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
241 tp->ecn_flags |= TCP_ECN_SEEN;
244 if (tcp_ca_needs_ecn((struct sock *)tp))
245 tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_NO_CE);
246 tp->ecn_flags |= TCP_ECN_SEEN;
251 static void tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
253 if (tp->ecn_flags & TCP_ECN_OK)
254 __tcp_ecn_check_ce(tp, skb);
257 static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
259 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
260 tp->ecn_flags &= ~TCP_ECN_OK;
263 static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
265 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
266 tp->ecn_flags &= ~TCP_ECN_OK;
269 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
271 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
276 /* Buffer size and advertised window tuning.
278 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
281 static void tcp_sndbuf_expand(struct sock *sk)
283 const struct tcp_sock *tp = tcp_sk(sk);
287 /* Worst case is non GSO/TSO : each frame consumes one skb
288 * and skb->head is kmalloced using power of two area of memory
290 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
292 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
294 per_mss = roundup_pow_of_two(per_mss) +
295 SKB_DATA_ALIGN(sizeof(struct sk_buff));
297 nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd);
298 nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
300 /* Fast Recovery (RFC 5681 3.2) :
301 * Cubic needs 1.7 factor, rounded to 2 to include
302 * extra cushion (application might react slowly to POLLOUT)
304 sndmem = 2 * nr_segs * per_mss;
306 if (sk->sk_sndbuf < sndmem)
307 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
310 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
312 * All tcp_full_space() is split to two parts: "network" buffer, allocated
313 * forward and advertised in receiver window (tp->rcv_wnd) and
314 * "application buffer", required to isolate scheduling/application
315 * latencies from network.
316 * window_clamp is maximal advertised window. It can be less than
317 * tcp_full_space(), in this case tcp_full_space() - window_clamp
318 * is reserved for "application" buffer. The less window_clamp is
319 * the smoother our behaviour from viewpoint of network, but the lower
320 * throughput and the higher sensitivity of the connection to losses. 8)
322 * rcv_ssthresh is more strict window_clamp used at "slow start"
323 * phase to predict further behaviour of this connection.
324 * It is used for two goals:
325 * - to enforce header prediction at sender, even when application
326 * requires some significant "application buffer". It is check #1.
327 * - to prevent pruning of receive queue because of misprediction
328 * of receiver window. Check #2.
330 * The scheme does not work when sender sends good segments opening
331 * window and then starts to feed us spaghetti. But it should work
332 * in common situations. Otherwise, we have to rely on queue collapsing.
335 /* Slow part of check#2. */
336 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
338 struct tcp_sock *tp = tcp_sk(sk);
340 int truesize = tcp_win_from_space(skb->truesize) >> 1;
341 int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
343 while (tp->rcv_ssthresh <= window) {
344 if (truesize <= skb->len)
345 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
353 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
355 struct tcp_sock *tp = tcp_sk(sk);
358 if (tp->rcv_ssthresh < tp->window_clamp &&
359 (int)tp->rcv_ssthresh < tcp_space(sk) &&
360 !sk_under_memory_pressure(sk)) {
363 /* Check #2. Increase window, if skb with such overhead
364 * will fit to rcvbuf in future.
366 if (tcp_win_from_space(skb->truesize) <= skb->len)
367 incr = 2 * tp->advmss;
369 incr = __tcp_grow_window(sk, skb);
372 incr = max_t(int, incr, 2 * skb->len);
373 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
375 inet_csk(sk)->icsk_ack.quick |= 1;
380 /* 3. Tuning rcvbuf, when connection enters established state. */
381 static void tcp_fixup_rcvbuf(struct sock *sk)
383 u32 mss = tcp_sk(sk)->advmss;
386 rcvmem = 2 * SKB_TRUESIZE(mss + MAX_TCP_HEADER) *
387 tcp_default_init_rwnd(mss);
389 /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
390 * Allow enough cushion so that sender is not limited by our window
392 if (sysctl_tcp_moderate_rcvbuf)
395 if (sk->sk_rcvbuf < rcvmem)
396 sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]);
399 /* 4. Try to fixup all. It is made immediately after connection enters
402 void tcp_init_buffer_space(struct sock *sk)
404 struct tcp_sock *tp = tcp_sk(sk);
407 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
408 tcp_fixup_rcvbuf(sk);
409 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
410 tcp_sndbuf_expand(sk);
412 tp->rcvq_space.space = tp->rcv_wnd;
413 tp->rcvq_space.time = tcp_time_stamp;
414 tp->rcvq_space.seq = tp->copied_seq;
416 maxwin = tcp_full_space(sk);
418 if (tp->window_clamp >= maxwin) {
419 tp->window_clamp = maxwin;
421 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
422 tp->window_clamp = max(maxwin -
423 (maxwin >> sysctl_tcp_app_win),
427 /* Force reservation of one segment. */
428 if (sysctl_tcp_app_win &&
429 tp->window_clamp > 2 * tp->advmss &&
430 tp->window_clamp + tp->advmss > maxwin)
431 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
433 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
434 tp->snd_cwnd_stamp = tcp_time_stamp;
437 /* 5. Recalculate window clamp after socket hit its memory bounds. */
438 static void tcp_clamp_window(struct sock *sk)
440 struct tcp_sock *tp = tcp_sk(sk);
441 struct inet_connection_sock *icsk = inet_csk(sk);
443 icsk->icsk_ack.quick = 0;
445 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
446 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
447 !sk_under_memory_pressure(sk) &&
448 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
449 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
452 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
453 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
456 /* Initialize RCV_MSS value.
457 * RCV_MSS is an our guess about MSS used by the peer.
458 * We haven't any direct information about the MSS.
459 * It's better to underestimate the RCV_MSS rather than overestimate.
460 * Overestimations make us ACKing less frequently than needed.
461 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
463 void tcp_initialize_rcv_mss(struct sock *sk)
465 const struct tcp_sock *tp = tcp_sk(sk);
466 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
468 hint = min(hint, tp->rcv_wnd / 2);
469 hint = min(hint, TCP_MSS_DEFAULT);
470 hint = max(hint, TCP_MIN_MSS);
472 inet_csk(sk)->icsk_ack.rcv_mss = hint;
474 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
476 /* Receiver "autotuning" code.
478 * The algorithm for RTT estimation w/o timestamps is based on
479 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
480 * <http://public.lanl.gov/radiant/pubs.html#DRS>
482 * More detail on this code can be found at
483 * <http://staff.psc.edu/jheffner/>,
484 * though this reference is out of date. A new paper
487 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
489 u32 new_sample = tp->rcv_rtt_est.rtt;
495 if (new_sample != 0) {
496 /* If we sample in larger samples in the non-timestamp
497 * case, we could grossly overestimate the RTT especially
498 * with chatty applications or bulk transfer apps which
499 * are stalled on filesystem I/O.
501 * Also, since we are only going for a minimum in the
502 * non-timestamp case, we do not smooth things out
503 * else with timestamps disabled convergence takes too
507 m -= (new_sample >> 3);
515 /* No previous measure. */
519 if (tp->rcv_rtt_est.rtt != new_sample)
520 tp->rcv_rtt_est.rtt = new_sample;
523 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
525 if (tp->rcv_rtt_est.time == 0)
527 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
529 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rcv_rtt_est.time, 1);
532 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
533 tp->rcv_rtt_est.time = tcp_time_stamp;
536 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
537 const struct sk_buff *skb)
539 struct tcp_sock *tp = tcp_sk(sk);
540 if (tp->rx_opt.rcv_tsecr &&
541 (TCP_SKB_CB(skb)->end_seq -
542 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
543 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
547 * This function should be called every time data is copied to user space.
548 * It calculates the appropriate TCP receive buffer space.
550 void tcp_rcv_space_adjust(struct sock *sk)
552 struct tcp_sock *tp = tcp_sk(sk);
556 time = tcp_time_stamp - tp->rcvq_space.time;
557 if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
560 /* Number of bytes copied to user in last RTT */
561 copied = tp->copied_seq - tp->rcvq_space.seq;
562 if (copied <= tp->rcvq_space.space)
566 * copied = bytes received in previous RTT, our base window
567 * To cope with packet losses, we need a 2x factor
568 * To cope with slow start, and sender growing its cwin by 100 %
569 * every RTT, we need a 4x factor, because the ACK we are sending
570 * now is for the next RTT, not the current one :
571 * <prev RTT . ><current RTT .. ><next RTT .... >
574 if (sysctl_tcp_moderate_rcvbuf &&
575 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
576 int rcvwin, rcvmem, rcvbuf;
578 /* minimal window to cope with packet losses, assuming
579 * steady state. Add some cushion because of small variations.
581 rcvwin = (copied << 1) + 16 * tp->advmss;
583 /* If rate increased by 25%,
584 * assume slow start, rcvwin = 3 * copied
585 * If rate increased by 50%,
586 * assume sender can use 2x growth, rcvwin = 4 * copied
589 tp->rcvq_space.space + (tp->rcvq_space.space >> 2)) {
591 tp->rcvq_space.space + (tp->rcvq_space.space >> 1))
594 rcvwin += (rcvwin >> 1);
597 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
598 while (tcp_win_from_space(rcvmem) < tp->advmss)
601 rcvbuf = min(rcvwin / tp->advmss * rcvmem, sysctl_tcp_rmem[2]);
602 if (rcvbuf > sk->sk_rcvbuf) {
603 sk->sk_rcvbuf = rcvbuf;
605 /* Make the window clamp follow along. */
606 tp->window_clamp = rcvwin;
609 tp->rcvq_space.space = copied;
612 tp->rcvq_space.seq = tp->copied_seq;
613 tp->rcvq_space.time = tcp_time_stamp;
616 /* There is something which you must keep in mind when you analyze the
617 * behavior of the tp->ato delayed ack timeout interval. When a
618 * connection starts up, we want to ack as quickly as possible. The
619 * problem is that "good" TCP's do slow start at the beginning of data
620 * transmission. The means that until we send the first few ACK's the
621 * sender will sit on his end and only queue most of his data, because
622 * he can only send snd_cwnd unacked packets at any given time. For
623 * each ACK we send, he increments snd_cwnd and transmits more of his
626 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
628 struct tcp_sock *tp = tcp_sk(sk);
629 struct inet_connection_sock *icsk = inet_csk(sk);
632 inet_csk_schedule_ack(sk);
634 tcp_measure_rcv_mss(sk, skb);
636 tcp_rcv_rtt_measure(tp);
638 now = tcp_time_stamp;
640 if (!icsk->icsk_ack.ato) {
641 /* The _first_ data packet received, initialize
642 * delayed ACK engine.
644 tcp_incr_quickack(sk);
645 icsk->icsk_ack.ato = TCP_ATO_MIN;
647 int m = now - icsk->icsk_ack.lrcvtime;
649 if (m <= TCP_ATO_MIN / 2) {
650 /* The fastest case is the first. */
651 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
652 } else if (m < icsk->icsk_ack.ato) {
653 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
654 if (icsk->icsk_ack.ato > icsk->icsk_rto)
655 icsk->icsk_ack.ato = icsk->icsk_rto;
656 } else if (m > icsk->icsk_rto) {
657 /* Too long gap. Apparently sender failed to
658 * restart window, so that we send ACKs quickly.
660 tcp_incr_quickack(sk);
664 icsk->icsk_ack.lrcvtime = now;
666 tcp_ecn_check_ce(tp, skb);
669 tcp_grow_window(sk, skb);
672 /* Called to compute a smoothed rtt estimate. The data fed to this
673 * routine either comes from timestamps, or from segments that were
674 * known _not_ to have been retransmitted [see Karn/Partridge
675 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
676 * piece by Van Jacobson.
677 * NOTE: the next three routines used to be one big routine.
678 * To save cycles in the RFC 1323 implementation it was better to break
679 * it up into three procedures. -- erics
681 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
683 struct tcp_sock *tp = tcp_sk(sk);
684 long m = mrtt_us; /* RTT */
685 u32 srtt = tp->srtt_us;
687 /* The following amusing code comes from Jacobson's
688 * article in SIGCOMM '88. Note that rtt and mdev
689 * are scaled versions of rtt and mean deviation.
690 * This is designed to be as fast as possible
691 * m stands for "measurement".
693 * On a 1990 paper the rto value is changed to:
694 * RTO = rtt + 4 * mdev
696 * Funny. This algorithm seems to be very broken.
697 * These formulae increase RTO, when it should be decreased, increase
698 * too slowly, when it should be increased quickly, decrease too quickly
699 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
700 * does not matter how to _calculate_ it. Seems, it was trap
701 * that VJ failed to avoid. 8)
704 m -= (srtt >> 3); /* m is now error in rtt est */
705 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
707 m = -m; /* m is now abs(error) */
708 m -= (tp->mdev_us >> 2); /* similar update on mdev */
709 /* This is similar to one of Eifel findings.
710 * Eifel blocks mdev updates when rtt decreases.
711 * This solution is a bit different: we use finer gain
712 * for mdev in this case (alpha*beta).
713 * Like Eifel it also prevents growth of rto,
714 * but also it limits too fast rto decreases,
715 * happening in pure Eifel.
720 m -= (tp->mdev_us >> 2); /* similar update on mdev */
722 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
723 if (tp->mdev_us > tp->mdev_max_us) {
724 tp->mdev_max_us = tp->mdev_us;
725 if (tp->mdev_max_us > tp->rttvar_us)
726 tp->rttvar_us = tp->mdev_max_us;
728 if (after(tp->snd_una, tp->rtt_seq)) {
729 if (tp->mdev_max_us < tp->rttvar_us)
730 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
731 tp->rtt_seq = tp->snd_nxt;
732 tp->mdev_max_us = tcp_rto_min_us(sk);
735 /* no previous measure. */
736 srtt = m << 3; /* take the measured time to be rtt */
737 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
738 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
739 tp->mdev_max_us = tp->rttvar_us;
740 tp->rtt_seq = tp->snd_nxt;
742 tp->srtt_us = max(1U, srtt);
745 /* Set the sk_pacing_rate to allow proper sizing of TSO packets.
746 * Note: TCP stack does not yet implement pacing.
747 * FQ packet scheduler can be used to implement cheap but effective
748 * TCP pacing, to smooth the burst on large writes when packets
749 * in flight is significantly lower than cwnd (or rwin)
751 static void tcp_update_pacing_rate(struct sock *sk)
753 const struct tcp_sock *tp = tcp_sk(sk);
756 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
757 rate = (u64)tp->mss_cache * 2 * (USEC_PER_SEC << 3);
759 rate *= max(tp->snd_cwnd, tp->packets_out);
761 if (likely(tp->srtt_us))
762 do_div(rate, tp->srtt_us);
764 /* ACCESS_ONCE() is needed because sch_fq fetches sk_pacing_rate
765 * without any lock. We want to make sure compiler wont store
766 * intermediate values in this location.
768 ACCESS_ONCE(sk->sk_pacing_rate) = min_t(u64, rate,
769 sk->sk_max_pacing_rate);
772 /* Calculate rto without backoff. This is the second half of Van Jacobson's
773 * routine referred to above.
775 static void tcp_set_rto(struct sock *sk)
777 const struct tcp_sock *tp = tcp_sk(sk);
778 /* Old crap is replaced with new one. 8)
781 * 1. If rtt variance happened to be less 50msec, it is hallucination.
782 * It cannot be less due to utterly erratic ACK generation made
783 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
784 * to do with delayed acks, because at cwnd>2 true delack timeout
785 * is invisible. Actually, Linux-2.4 also generates erratic
786 * ACKs in some circumstances.
788 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
790 /* 2. Fixups made earlier cannot be right.
791 * If we do not estimate RTO correctly without them,
792 * all the algo is pure shit and should be replaced
793 * with correct one. It is exactly, which we pretend to do.
796 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
797 * guarantees that rto is higher.
802 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
804 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
807 cwnd = TCP_INIT_CWND;
808 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
812 * Packet counting of FACK is based on in-order assumptions, therefore TCP
813 * disables it when reordering is detected
815 void tcp_disable_fack(struct tcp_sock *tp)
817 /* RFC3517 uses different metric in lost marker => reset on change */
819 tp->lost_skb_hint = NULL;
820 tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED;
823 /* Take a notice that peer is sending D-SACKs */
824 static void tcp_dsack_seen(struct tcp_sock *tp)
826 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
829 static void tcp_update_reordering(struct sock *sk, const int metric,
832 struct tcp_sock *tp = tcp_sk(sk);
833 if (metric > tp->reordering) {
836 tp->reordering = min(TCP_MAX_REORDERING, metric);
838 /* This exciting event is worth to be remembered. 8) */
840 mib_idx = LINUX_MIB_TCPTSREORDER;
841 else if (tcp_is_reno(tp))
842 mib_idx = LINUX_MIB_TCPRENOREORDER;
843 else if (tcp_is_fack(tp))
844 mib_idx = LINUX_MIB_TCPFACKREORDER;
846 mib_idx = LINUX_MIB_TCPSACKREORDER;
848 NET_INC_STATS_BH(sock_net(sk), mib_idx);
849 #if FASTRETRANS_DEBUG > 1
850 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
851 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
855 tp->undo_marker ? tp->undo_retrans : 0);
857 tcp_disable_fack(tp);
861 tcp_disable_early_retrans(tp);
864 /* This must be called before lost_out is incremented */
865 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
867 if ((tp->retransmit_skb_hint == NULL) ||
868 before(TCP_SKB_CB(skb)->seq,
869 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
870 tp->retransmit_skb_hint = skb;
873 after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
874 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
877 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
879 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
880 tcp_verify_retransmit_hint(tp, skb);
882 tp->lost_out += tcp_skb_pcount(skb);
883 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
887 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp,
890 tcp_verify_retransmit_hint(tp, skb);
892 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
893 tp->lost_out += tcp_skb_pcount(skb);
894 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
898 /* This procedure tags the retransmission queue when SACKs arrive.
900 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
901 * Packets in queue with these bits set are counted in variables
902 * sacked_out, retrans_out and lost_out, correspondingly.
904 * Valid combinations are:
905 * Tag InFlight Description
906 * 0 1 - orig segment is in flight.
907 * S 0 - nothing flies, orig reached receiver.
908 * L 0 - nothing flies, orig lost by net.
909 * R 2 - both orig and retransmit are in flight.
910 * L|R 1 - orig is lost, retransmit is in flight.
911 * S|R 1 - orig reached receiver, retrans is still in flight.
912 * (L|S|R is logically valid, it could occur when L|R is sacked,
913 * but it is equivalent to plain S and code short-curcuits it to S.
914 * L|S is logically invalid, it would mean -1 packet in flight 8))
916 * These 6 states form finite state machine, controlled by the following events:
917 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
918 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
919 * 3. Loss detection event of two flavors:
920 * A. Scoreboard estimator decided the packet is lost.
921 * A'. Reno "three dupacks" marks head of queue lost.
922 * A''. Its FACK modification, head until snd.fack is lost.
923 * B. SACK arrives sacking SND.NXT at the moment, when the
924 * segment was retransmitted.
925 * 4. D-SACK added new rule: D-SACK changes any tag to S.
927 * It is pleasant to note, that state diagram turns out to be commutative,
928 * so that we are allowed not to be bothered by order of our actions,
929 * when multiple events arrive simultaneously. (see the function below).
931 * Reordering detection.
932 * --------------------
933 * Reordering metric is maximal distance, which a packet can be displaced
934 * in packet stream. With SACKs we can estimate it:
936 * 1. SACK fills old hole and the corresponding segment was not
937 * ever retransmitted -> reordering. Alas, we cannot use it
938 * when segment was retransmitted.
939 * 2. The last flaw is solved with D-SACK. D-SACK arrives
940 * for retransmitted and already SACKed segment -> reordering..
941 * Both of these heuristics are not used in Loss state, when we cannot
942 * account for retransmits accurately.
944 * SACK block validation.
945 * ----------------------
947 * SACK block range validation checks that the received SACK block fits to
948 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
949 * Note that SND.UNA is not included to the range though being valid because
950 * it means that the receiver is rather inconsistent with itself reporting
951 * SACK reneging when it should advance SND.UNA. Such SACK block this is
952 * perfectly valid, however, in light of RFC2018 which explicitly states
953 * that "SACK block MUST reflect the newest segment. Even if the newest
954 * segment is going to be discarded ...", not that it looks very clever
955 * in case of head skb. Due to potentional receiver driven attacks, we
956 * choose to avoid immediate execution of a walk in write queue due to
957 * reneging and defer head skb's loss recovery to standard loss recovery
958 * procedure that will eventually trigger (nothing forbids us doing this).
960 * Implements also blockage to start_seq wrap-around. Problem lies in the
961 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
962 * there's no guarantee that it will be before snd_nxt (n). The problem
963 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
966 * <- outs wnd -> <- wrapzone ->
967 * u e n u_w e_w s n_w
969 * |<------------+------+----- TCP seqno space --------------+---------->|
970 * ...-- <2^31 ->| |<--------...
971 * ...---- >2^31 ------>| |<--------...
973 * Current code wouldn't be vulnerable but it's better still to discard such
974 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
975 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
976 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
977 * equal to the ideal case (infinite seqno space without wrap caused issues).
979 * With D-SACK the lower bound is extended to cover sequence space below
980 * SND.UNA down to undo_marker, which is the last point of interest. Yet
981 * again, D-SACK block must not to go across snd_una (for the same reason as
982 * for the normal SACK blocks, explained above). But there all simplicity
983 * ends, TCP might receive valid D-SACKs below that. As long as they reside
984 * fully below undo_marker they do not affect behavior in anyway and can
985 * therefore be safely ignored. In rare cases (which are more or less
986 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
987 * fragmentation and packet reordering past skb's retransmission. To consider
988 * them correctly, the acceptable range must be extended even more though
989 * the exact amount is rather hard to quantify. However, tp->max_window can
990 * be used as an exaggerated estimate.
992 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
993 u32 start_seq, u32 end_seq)
995 /* Too far in future, or reversed (interpretation is ambiguous) */
996 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
999 /* Nasty start_seq wrap-around check (see comments above) */
1000 if (!before(start_seq, tp->snd_nxt))
1003 /* In outstanding window? ...This is valid exit for D-SACKs too.
1004 * start_seq == snd_una is non-sensical (see comments above)
1006 if (after(start_seq, tp->snd_una))
1009 if (!is_dsack || !tp->undo_marker)
1012 /* ...Then it's D-SACK, and must reside below snd_una completely */
1013 if (after(end_seq, tp->snd_una))
1016 if (!before(start_seq, tp->undo_marker))
1020 if (!after(end_seq, tp->undo_marker))
1023 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1024 * start_seq < undo_marker and end_seq >= undo_marker.
1026 return !before(start_seq, end_seq - tp->max_window);
1029 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1030 * Event "B". Later note: FACK people cheated me again 8), we have to account
1031 * for reordering! Ugly, but should help.
1033 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1034 * less than what is now known to be received by the other end (derived from
1035 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1036 * retransmitted skbs to avoid some costly processing per ACKs.
1038 static void tcp_mark_lost_retrans(struct sock *sk)
1040 const struct inet_connection_sock *icsk = inet_csk(sk);
1041 struct tcp_sock *tp = tcp_sk(sk);
1042 struct sk_buff *skb;
1044 u32 new_low_seq = tp->snd_nxt;
1045 u32 received_upto = tcp_highest_sack_seq(tp);
1047 if (!tcp_is_fack(tp) || !tp->retrans_out ||
1048 !after(received_upto, tp->lost_retrans_low) ||
1049 icsk->icsk_ca_state != TCP_CA_Recovery)
1052 tcp_for_write_queue(skb, sk) {
1053 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1055 if (skb == tcp_send_head(sk))
1057 if (cnt == tp->retrans_out)
1059 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1062 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1065 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1066 * constraint here (see above) but figuring out that at
1067 * least tp->reordering SACK blocks reside between ack_seq
1068 * and received_upto is not easy task to do cheaply with
1069 * the available datastructures.
1071 * Whether FACK should check here for tp->reordering segs
1072 * in-between one could argue for either way (it would be
1073 * rather simple to implement as we could count fack_count
1074 * during the walk and do tp->fackets_out - fack_count).
1076 if (after(received_upto, ack_seq)) {
1077 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1078 tp->retrans_out -= tcp_skb_pcount(skb);
1080 tcp_skb_mark_lost_uncond_verify(tp, skb);
1081 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
1083 if (before(ack_seq, new_low_seq))
1084 new_low_seq = ack_seq;
1085 cnt += tcp_skb_pcount(skb);
1089 if (tp->retrans_out)
1090 tp->lost_retrans_low = new_low_seq;
1093 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1094 struct tcp_sack_block_wire *sp, int num_sacks,
1097 struct tcp_sock *tp = tcp_sk(sk);
1098 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1099 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1100 bool dup_sack = false;
1102 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1105 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1106 } else if (num_sacks > 1) {
1107 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1108 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1110 if (!after(end_seq_0, end_seq_1) &&
1111 !before(start_seq_0, start_seq_1)) {
1114 NET_INC_STATS_BH(sock_net(sk),
1115 LINUX_MIB_TCPDSACKOFORECV);
1119 /* D-SACK for already forgotten data... Do dumb counting. */
1120 if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 &&
1121 !after(end_seq_0, prior_snd_una) &&
1122 after(end_seq_0, tp->undo_marker))
1128 struct tcp_sacktag_state {
1131 long rtt_us; /* RTT measured by SACKing never-retransmitted data */
1135 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1136 * the incoming SACK may not exactly match but we can find smaller MSS
1137 * aligned portion of it that matches. Therefore we might need to fragment
1138 * which may fail and creates some hassle (caller must handle error case
1141 * FIXME: this could be merged to shift decision code
1143 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1144 u32 start_seq, u32 end_seq)
1148 unsigned int pkt_len;
1151 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1152 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1154 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1155 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1156 mss = tcp_skb_mss(skb);
1157 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1160 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1164 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1169 /* Round if necessary so that SACKs cover only full MSSes
1170 * and/or the remaining small portion (if present)
1172 if (pkt_len > mss) {
1173 unsigned int new_len = (pkt_len / mss) * mss;
1174 if (!in_sack && new_len < pkt_len) {
1176 if (new_len >= skb->len)
1181 err = tcp_fragment(sk, skb, pkt_len, mss, GFP_ATOMIC);
1189 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1190 static u8 tcp_sacktag_one(struct sock *sk,
1191 struct tcp_sacktag_state *state, u8 sacked,
1192 u32 start_seq, u32 end_seq,
1193 int dup_sack, int pcount,
1194 const struct skb_mstamp *xmit_time)
1196 struct tcp_sock *tp = tcp_sk(sk);
1197 int fack_count = state->fack_count;
1199 /* Account D-SACK for retransmitted packet. */
1200 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1201 if (tp->undo_marker && tp->undo_retrans > 0 &&
1202 after(end_seq, tp->undo_marker))
1204 if (sacked & TCPCB_SACKED_ACKED)
1205 state->reord = min(fack_count, state->reord);
1208 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1209 if (!after(end_seq, tp->snd_una))
1212 if (!(sacked & TCPCB_SACKED_ACKED)) {
1213 if (sacked & TCPCB_SACKED_RETRANS) {
1214 /* If the segment is not tagged as lost,
1215 * we do not clear RETRANS, believing
1216 * that retransmission is still in flight.
1218 if (sacked & TCPCB_LOST) {
1219 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1220 tp->lost_out -= pcount;
1221 tp->retrans_out -= pcount;
1224 if (!(sacked & TCPCB_RETRANS)) {
1225 /* New sack for not retransmitted frame,
1226 * which was in hole. It is reordering.
1228 if (before(start_seq,
1229 tcp_highest_sack_seq(tp)))
1230 state->reord = min(fack_count,
1232 if (!after(end_seq, tp->high_seq))
1233 state->flag |= FLAG_ORIG_SACK_ACKED;
1234 /* Pick the earliest sequence sacked for RTT */
1235 if (state->rtt_us < 0) {
1236 struct skb_mstamp now;
1238 skb_mstamp_get(&now);
1239 state->rtt_us = skb_mstamp_us_delta(&now,
1244 if (sacked & TCPCB_LOST) {
1245 sacked &= ~TCPCB_LOST;
1246 tp->lost_out -= pcount;
1250 sacked |= TCPCB_SACKED_ACKED;
1251 state->flag |= FLAG_DATA_SACKED;
1252 tp->sacked_out += pcount;
1254 fack_count += pcount;
1256 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1257 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
1258 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1259 tp->lost_cnt_hint += pcount;
1261 if (fack_count > tp->fackets_out)
1262 tp->fackets_out = fack_count;
1265 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1266 * frames and clear it. undo_retrans is decreased above, L|R frames
1267 * are accounted above as well.
1269 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1270 sacked &= ~TCPCB_SACKED_RETRANS;
1271 tp->retrans_out -= pcount;
1277 /* Shift newly-SACKed bytes from this skb to the immediately previous
1278 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1280 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1281 struct tcp_sacktag_state *state,
1282 unsigned int pcount, int shifted, int mss,
1285 struct tcp_sock *tp = tcp_sk(sk);
1286 struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1287 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1288 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1292 /* Adjust counters and hints for the newly sacked sequence
1293 * range but discard the return value since prev is already
1294 * marked. We must tag the range first because the seq
1295 * advancement below implicitly advances
1296 * tcp_highest_sack_seq() when skb is highest_sack.
1298 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1299 start_seq, end_seq, dup_sack, pcount,
1302 if (skb == tp->lost_skb_hint)
1303 tp->lost_cnt_hint += pcount;
1305 TCP_SKB_CB(prev)->end_seq += shifted;
1306 TCP_SKB_CB(skb)->seq += shifted;
1308 tcp_skb_pcount_add(prev, pcount);
1309 BUG_ON(tcp_skb_pcount(skb) < pcount);
1310 tcp_skb_pcount_add(skb, -pcount);
1312 /* When we're adding to gso_segs == 1, gso_size will be zero,
1313 * in theory this shouldn't be necessary but as long as DSACK
1314 * code can come after this skb later on it's better to keep
1315 * setting gso_size to something.
1317 if (!skb_shinfo(prev)->gso_size) {
1318 skb_shinfo(prev)->gso_size = mss;
1319 skb_shinfo(prev)->gso_type = sk->sk_gso_type;
1322 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1323 if (tcp_skb_pcount(skb) <= 1) {
1324 skb_shinfo(skb)->gso_size = 0;
1325 skb_shinfo(skb)->gso_type = 0;
1328 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1329 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1332 BUG_ON(!tcp_skb_pcount(skb));
1333 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1337 /* Whole SKB was eaten :-) */
1339 if (skb == tp->retransmit_skb_hint)
1340 tp->retransmit_skb_hint = prev;
1341 if (skb == tp->lost_skb_hint) {
1342 tp->lost_skb_hint = prev;
1343 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1346 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1347 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1348 TCP_SKB_CB(prev)->end_seq++;
1350 if (skb == tcp_highest_sack(sk))
1351 tcp_advance_highest_sack(sk, skb);
1353 tcp_unlink_write_queue(skb, sk);
1354 sk_wmem_free_skb(sk, skb);
1356 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1361 /* I wish gso_size would have a bit more sane initialization than
1362 * something-or-zero which complicates things
1364 static int tcp_skb_seglen(const struct sk_buff *skb)
1366 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1369 /* Shifting pages past head area doesn't work */
1370 static int skb_can_shift(const struct sk_buff *skb)
1372 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1375 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1378 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1379 struct tcp_sacktag_state *state,
1380 u32 start_seq, u32 end_seq,
1383 struct tcp_sock *tp = tcp_sk(sk);
1384 struct sk_buff *prev;
1390 if (!sk_can_gso(sk))
1393 /* Normally R but no L won't result in plain S */
1395 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1397 if (!skb_can_shift(skb))
1399 /* This frame is about to be dropped (was ACKed). */
1400 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1403 /* Can only happen with delayed DSACK + discard craziness */
1404 if (unlikely(skb == tcp_write_queue_head(sk)))
1406 prev = tcp_write_queue_prev(sk, skb);
1408 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1411 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1412 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1416 pcount = tcp_skb_pcount(skb);
1417 mss = tcp_skb_seglen(skb);
1419 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1420 * drop this restriction as unnecessary
1422 if (mss != tcp_skb_seglen(prev))
1425 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1427 /* CHECKME: This is non-MSS split case only?, this will
1428 * cause skipped skbs due to advancing loop btw, original
1429 * has that feature too
1431 if (tcp_skb_pcount(skb) <= 1)
1434 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1436 /* TODO: head merge to next could be attempted here
1437 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1438 * though it might not be worth of the additional hassle
1440 * ...we can probably just fallback to what was done
1441 * previously. We could try merging non-SACKed ones
1442 * as well but it probably isn't going to buy off
1443 * because later SACKs might again split them, and
1444 * it would make skb timestamp tracking considerably
1450 len = end_seq - TCP_SKB_CB(skb)->seq;
1452 BUG_ON(len > skb->len);
1454 /* MSS boundaries should be honoured or else pcount will
1455 * severely break even though it makes things bit trickier.
1456 * Optimize common case to avoid most of the divides
1458 mss = tcp_skb_mss(skb);
1460 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1461 * drop this restriction as unnecessary
1463 if (mss != tcp_skb_seglen(prev))
1468 } else if (len < mss) {
1476 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1477 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1480 if (!skb_shift(prev, skb, len))
1482 if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1485 /* Hole filled allows collapsing with the next as well, this is very
1486 * useful when hole on every nth skb pattern happens
1488 if (prev == tcp_write_queue_tail(sk))
1490 skb = tcp_write_queue_next(sk, prev);
1492 if (!skb_can_shift(skb) ||
1493 (skb == tcp_send_head(sk)) ||
1494 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1495 (mss != tcp_skb_seglen(skb)))
1499 if (skb_shift(prev, skb, len)) {
1500 pcount += tcp_skb_pcount(skb);
1501 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1505 state->fack_count += pcount;
1512 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1516 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1517 struct tcp_sack_block *next_dup,
1518 struct tcp_sacktag_state *state,
1519 u32 start_seq, u32 end_seq,
1522 struct tcp_sock *tp = tcp_sk(sk);
1523 struct sk_buff *tmp;
1525 tcp_for_write_queue_from(skb, sk) {
1527 bool dup_sack = dup_sack_in;
1529 if (skb == tcp_send_head(sk))
1532 /* queue is in-order => we can short-circuit the walk early */
1533 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1536 if ((next_dup != NULL) &&
1537 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1538 in_sack = tcp_match_skb_to_sack(sk, skb,
1539 next_dup->start_seq,
1545 /* skb reference here is a bit tricky to get right, since
1546 * shifting can eat and free both this skb and the next,
1547 * so not even _safe variant of the loop is enough.
1550 tmp = tcp_shift_skb_data(sk, skb, state,
1551 start_seq, end_seq, dup_sack);
1560 in_sack = tcp_match_skb_to_sack(sk, skb,
1566 if (unlikely(in_sack < 0))
1570 TCP_SKB_CB(skb)->sacked =
1573 TCP_SKB_CB(skb)->sacked,
1574 TCP_SKB_CB(skb)->seq,
1575 TCP_SKB_CB(skb)->end_seq,
1577 tcp_skb_pcount(skb),
1580 if (!before(TCP_SKB_CB(skb)->seq,
1581 tcp_highest_sack_seq(tp)))
1582 tcp_advance_highest_sack(sk, skb);
1585 state->fack_count += tcp_skb_pcount(skb);
1590 /* Avoid all extra work that is being done by sacktag while walking in
1593 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1594 struct tcp_sacktag_state *state,
1597 tcp_for_write_queue_from(skb, sk) {
1598 if (skb == tcp_send_head(sk))
1601 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1604 state->fack_count += tcp_skb_pcount(skb);
1609 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1611 struct tcp_sack_block *next_dup,
1612 struct tcp_sacktag_state *state,
1615 if (next_dup == NULL)
1618 if (before(next_dup->start_seq, skip_to_seq)) {
1619 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1620 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1621 next_dup->start_seq, next_dup->end_seq,
1628 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1630 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1634 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1635 u32 prior_snd_una, long *sack_rtt_us)
1637 struct tcp_sock *tp = tcp_sk(sk);
1638 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1639 TCP_SKB_CB(ack_skb)->sacked);
1640 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1641 struct tcp_sack_block sp[TCP_NUM_SACKS];
1642 struct tcp_sack_block *cache;
1643 struct tcp_sacktag_state state;
1644 struct sk_buff *skb;
1645 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1647 bool found_dup_sack = false;
1649 int first_sack_index;
1652 state.reord = tp->packets_out;
1655 if (!tp->sacked_out) {
1656 if (WARN_ON(tp->fackets_out))
1657 tp->fackets_out = 0;
1658 tcp_highest_sack_reset(sk);
1661 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1662 num_sacks, prior_snd_una);
1664 state.flag |= FLAG_DSACKING_ACK;
1666 /* Eliminate too old ACKs, but take into
1667 * account more or less fresh ones, they can
1668 * contain valid SACK info.
1670 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1673 if (!tp->packets_out)
1677 first_sack_index = 0;
1678 for (i = 0; i < num_sacks; i++) {
1679 bool dup_sack = !i && found_dup_sack;
1681 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1682 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1684 if (!tcp_is_sackblock_valid(tp, dup_sack,
1685 sp[used_sacks].start_seq,
1686 sp[used_sacks].end_seq)) {
1690 if (!tp->undo_marker)
1691 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1693 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1695 /* Don't count olds caused by ACK reordering */
1696 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1697 !after(sp[used_sacks].end_seq, tp->snd_una))
1699 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1702 NET_INC_STATS_BH(sock_net(sk), mib_idx);
1704 first_sack_index = -1;
1708 /* Ignore very old stuff early */
1709 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1715 /* order SACK blocks to allow in order walk of the retrans queue */
1716 for (i = used_sacks - 1; i > 0; i--) {
1717 for (j = 0; j < i; j++) {
1718 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1719 swap(sp[j], sp[j + 1]);
1721 /* Track where the first SACK block goes to */
1722 if (j == first_sack_index)
1723 first_sack_index = j + 1;
1728 skb = tcp_write_queue_head(sk);
1729 state.fack_count = 0;
1732 if (!tp->sacked_out) {
1733 /* It's already past, so skip checking against it */
1734 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1736 cache = tp->recv_sack_cache;
1737 /* Skip empty blocks in at head of the cache */
1738 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1743 while (i < used_sacks) {
1744 u32 start_seq = sp[i].start_seq;
1745 u32 end_seq = sp[i].end_seq;
1746 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1747 struct tcp_sack_block *next_dup = NULL;
1749 if (found_dup_sack && ((i + 1) == first_sack_index))
1750 next_dup = &sp[i + 1];
1752 /* Skip too early cached blocks */
1753 while (tcp_sack_cache_ok(tp, cache) &&
1754 !before(start_seq, cache->end_seq))
1757 /* Can skip some work by looking recv_sack_cache? */
1758 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1759 after(end_seq, cache->start_seq)) {
1762 if (before(start_seq, cache->start_seq)) {
1763 skb = tcp_sacktag_skip(skb, sk, &state,
1765 skb = tcp_sacktag_walk(skb, sk, next_dup,
1772 /* Rest of the block already fully processed? */
1773 if (!after(end_seq, cache->end_seq))
1776 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1780 /* ...tail remains todo... */
1781 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1782 /* ...but better entrypoint exists! */
1783 skb = tcp_highest_sack(sk);
1786 state.fack_count = tp->fackets_out;
1791 skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq);
1792 /* Check overlap against next cached too (past this one already) */
1797 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1798 skb = tcp_highest_sack(sk);
1801 state.fack_count = tp->fackets_out;
1803 skb = tcp_sacktag_skip(skb, sk, &state, start_seq);
1806 skb = tcp_sacktag_walk(skb, sk, next_dup, &state,
1807 start_seq, end_seq, dup_sack);
1813 /* Clear the head of the cache sack blocks so we can skip it next time */
1814 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1815 tp->recv_sack_cache[i].start_seq = 0;
1816 tp->recv_sack_cache[i].end_seq = 0;
1818 for (j = 0; j < used_sacks; j++)
1819 tp->recv_sack_cache[i++] = sp[j];
1821 tcp_mark_lost_retrans(sk);
1823 tcp_verify_left_out(tp);
1825 if ((state.reord < tp->fackets_out) &&
1826 ((inet_csk(sk)->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker))
1827 tcp_update_reordering(sk, tp->fackets_out - state.reord, 0);
1831 #if FASTRETRANS_DEBUG > 0
1832 WARN_ON((int)tp->sacked_out < 0);
1833 WARN_ON((int)tp->lost_out < 0);
1834 WARN_ON((int)tp->retrans_out < 0);
1835 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1837 *sack_rtt_us = state.rtt_us;
1841 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1842 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1844 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1848 holes = max(tp->lost_out, 1U);
1849 holes = min(holes, tp->packets_out);
1851 if ((tp->sacked_out + holes) > tp->packets_out) {
1852 tp->sacked_out = tp->packets_out - holes;
1858 /* If we receive more dupacks than we expected counting segments
1859 * in assumption of absent reordering, interpret this as reordering.
1860 * The only another reason could be bug in receiver TCP.
1862 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1864 struct tcp_sock *tp = tcp_sk(sk);
1865 if (tcp_limit_reno_sacked(tp))
1866 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1869 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1871 static void tcp_add_reno_sack(struct sock *sk)
1873 struct tcp_sock *tp = tcp_sk(sk);
1875 tcp_check_reno_reordering(sk, 0);
1876 tcp_verify_left_out(tp);
1879 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1881 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1883 struct tcp_sock *tp = tcp_sk(sk);
1886 /* One ACK acked hole. The rest eat duplicate ACKs. */
1887 if (acked - 1 >= tp->sacked_out)
1890 tp->sacked_out -= acked - 1;
1892 tcp_check_reno_reordering(sk, acked);
1893 tcp_verify_left_out(tp);
1896 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1901 void tcp_clear_retrans(struct tcp_sock *tp)
1903 tp->retrans_out = 0;
1905 tp->undo_marker = 0;
1906 tp->undo_retrans = -1;
1907 tp->fackets_out = 0;
1911 static inline void tcp_init_undo(struct tcp_sock *tp)
1913 tp->undo_marker = tp->snd_una;
1914 /* Retransmission still in flight may cause DSACKs later. */
1915 tp->undo_retrans = tp->retrans_out ? : -1;
1918 /* Enter Loss state. If we detect SACK reneging, forget all SACK information
1919 * and reset tags completely, otherwise preserve SACKs. If receiver
1920 * dropped its ofo queue, we will know this due to reneging detection.
1922 void tcp_enter_loss(struct sock *sk)
1924 const struct inet_connection_sock *icsk = inet_csk(sk);
1925 struct tcp_sock *tp = tcp_sk(sk);
1926 struct sk_buff *skb;
1927 bool new_recovery = false;
1928 bool is_reneg; /* is receiver reneging on SACKs? */
1930 /* Reduce ssthresh if it has not yet been made inside this window. */
1931 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1932 !after(tp->high_seq, tp->snd_una) ||
1933 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1934 new_recovery = true;
1935 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1936 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1937 tcp_ca_event(sk, CA_EVENT_LOSS);
1941 tp->snd_cwnd_cnt = 0;
1942 tp->snd_cwnd_stamp = tcp_time_stamp;
1944 tp->retrans_out = 0;
1947 if (tcp_is_reno(tp))
1948 tcp_reset_reno_sack(tp);
1950 skb = tcp_write_queue_head(sk);
1951 is_reneg = skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED);
1953 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
1955 tp->fackets_out = 0;
1957 tcp_clear_all_retrans_hints(tp);
1959 tcp_for_write_queue(skb, sk) {
1960 if (skb == tcp_send_head(sk))
1963 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1964 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || is_reneg) {
1965 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1966 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1967 tp->lost_out += tcp_skb_pcount(skb);
1968 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
1971 tcp_verify_left_out(tp);
1973 /* Timeout in disordered state after receiving substantial DUPACKs
1974 * suggests that the degree of reordering is over-estimated.
1976 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
1977 tp->sacked_out >= sysctl_tcp_reordering)
1978 tp->reordering = min_t(unsigned int, tp->reordering,
1979 sysctl_tcp_reordering);
1980 tcp_set_ca_state(sk, TCP_CA_Loss);
1981 tp->high_seq = tp->snd_nxt;
1982 tcp_ecn_queue_cwr(tp);
1984 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1985 * loss recovery is underway except recurring timeout(s) on
1986 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1988 tp->frto = sysctl_tcp_frto &&
1989 (new_recovery || icsk->icsk_retransmits) &&
1990 !inet_csk(sk)->icsk_mtup.probe_size;
1993 /* If ACK arrived pointing to a remembered SACK, it means that our
1994 * remembered SACKs do not reflect real state of receiver i.e.
1995 * receiver _host_ is heavily congested (or buggy).
1997 * To avoid big spurious retransmission bursts due to transient SACK
1998 * scoreboard oddities that look like reneging, we give the receiver a
1999 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2000 * restore sanity to the SACK scoreboard. If the apparent reneging
2001 * persists until this RTO then we'll clear the SACK scoreboard.
2003 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
2005 if (flag & FLAG_SACK_RENEGING) {
2006 struct tcp_sock *tp = tcp_sk(sk);
2007 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2008 msecs_to_jiffies(10));
2010 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2011 delay, TCP_RTO_MAX);
2017 static inline int tcp_fackets_out(const struct tcp_sock *tp)
2019 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
2022 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2023 * counter when SACK is enabled (without SACK, sacked_out is used for
2026 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2027 * segments up to the highest received SACK block so far and holes in
2030 * With reordering, holes may still be in flight, so RFC3517 recovery
2031 * uses pure sacked_out (total number of SACKed segments) even though
2032 * it violates the RFC that uses duplicate ACKs, often these are equal
2033 * but when e.g. out-of-window ACKs or packet duplication occurs,
2034 * they differ. Since neither occurs due to loss, TCP should really
2037 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2039 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2042 static bool tcp_pause_early_retransmit(struct sock *sk, int flag)
2044 struct tcp_sock *tp = tcp_sk(sk);
2045 unsigned long delay;
2047 /* Delay early retransmit and entering fast recovery for
2048 * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
2049 * available, or RTO is scheduled to fire first.
2051 if (sysctl_tcp_early_retrans < 2 || sysctl_tcp_early_retrans > 3 ||
2052 (flag & FLAG_ECE) || !tp->srtt_us)
2055 delay = max(usecs_to_jiffies(tp->srtt_us >> 5),
2056 msecs_to_jiffies(2));
2058 if (!time_after(inet_csk(sk)->icsk_timeout, (jiffies + delay)))
2061 inet_csk_reset_xmit_timer(sk, ICSK_TIME_EARLY_RETRANS, delay,
2066 /* Linux NewReno/SACK/FACK/ECN state machine.
2067 * --------------------------------------
2069 * "Open" Normal state, no dubious events, fast path.
2070 * "Disorder" In all the respects it is "Open",
2071 * but requires a bit more attention. It is entered when
2072 * we see some SACKs or dupacks. It is split of "Open"
2073 * mainly to move some processing from fast path to slow one.
2074 * "CWR" CWND was reduced due to some Congestion Notification event.
2075 * It can be ECN, ICMP source quench, local device congestion.
2076 * "Recovery" CWND was reduced, we are fast-retransmitting.
2077 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2079 * tcp_fastretrans_alert() is entered:
2080 * - each incoming ACK, if state is not "Open"
2081 * - when arrived ACK is unusual, namely:
2086 * Counting packets in flight is pretty simple.
2088 * in_flight = packets_out - left_out + retrans_out
2090 * packets_out is SND.NXT-SND.UNA counted in packets.
2092 * retrans_out is number of retransmitted segments.
2094 * left_out is number of segments left network, but not ACKed yet.
2096 * left_out = sacked_out + lost_out
2098 * sacked_out: Packets, which arrived to receiver out of order
2099 * and hence not ACKed. With SACKs this number is simply
2100 * amount of SACKed data. Even without SACKs
2101 * it is easy to give pretty reliable estimate of this number,
2102 * counting duplicate ACKs.
2104 * lost_out: Packets lost by network. TCP has no explicit
2105 * "loss notification" feedback from network (for now).
2106 * It means that this number can be only _guessed_.
2107 * Actually, it is the heuristics to predict lossage that
2108 * distinguishes different algorithms.
2110 * F.e. after RTO, when all the queue is considered as lost,
2111 * lost_out = packets_out and in_flight = retrans_out.
2113 * Essentially, we have now two algorithms counting
2116 * FACK: It is the simplest heuristics. As soon as we decided
2117 * that something is lost, we decide that _all_ not SACKed
2118 * packets until the most forward SACK are lost. I.e.
2119 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2120 * It is absolutely correct estimate, if network does not reorder
2121 * packets. And it loses any connection to reality when reordering
2122 * takes place. We use FACK by default until reordering
2123 * is suspected on the path to this destination.
2125 * NewReno: when Recovery is entered, we assume that one segment
2126 * is lost (classic Reno). While we are in Recovery and
2127 * a partial ACK arrives, we assume that one more packet
2128 * is lost (NewReno). This heuristics are the same in NewReno
2131 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2132 * deflation etc. CWND is real congestion window, never inflated, changes
2133 * only according to classic VJ rules.
2135 * Really tricky (and requiring careful tuning) part of algorithm
2136 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2137 * The first determines the moment _when_ we should reduce CWND and,
2138 * hence, slow down forward transmission. In fact, it determines the moment
2139 * when we decide that hole is caused by loss, rather than by a reorder.
2141 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2142 * holes, caused by lost packets.
2144 * And the most logically complicated part of algorithm is undo
2145 * heuristics. We detect false retransmits due to both too early
2146 * fast retransmit (reordering) and underestimated RTO, analyzing
2147 * timestamps and D-SACKs. When we detect that some segments were
2148 * retransmitted by mistake and CWND reduction was wrong, we undo
2149 * window reduction and abort recovery phase. This logic is hidden
2150 * inside several functions named tcp_try_undo_<something>.
2153 /* This function decides, when we should leave Disordered state
2154 * and enter Recovery phase, reducing congestion window.
2156 * Main question: may we further continue forward transmission
2157 * with the same cwnd?
2159 static bool tcp_time_to_recover(struct sock *sk, int flag)
2161 struct tcp_sock *tp = tcp_sk(sk);
2164 /* Trick#1: The loss is proven. */
2168 /* Not-A-Trick#2 : Classic rule... */
2169 if (tcp_dupack_heuristics(tp) > tp->reordering)
2172 /* Trick#4: It is still not OK... But will it be useful to delay
2175 packets_out = tp->packets_out;
2176 if (packets_out <= tp->reordering &&
2177 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2178 !tcp_may_send_now(sk)) {
2179 /* We have nothing to send. This connection is limited
2180 * either by receiver window or by application.
2185 /* If a thin stream is detected, retransmit after first
2186 * received dupack. Employ only if SACK is supported in order
2187 * to avoid possible corner-case series of spurious retransmissions
2188 * Use only if there are no unsent data.
2190 if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2191 tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2192 tcp_is_sack(tp) && !tcp_send_head(sk))
2195 /* Trick#6: TCP early retransmit, per RFC5827. To avoid spurious
2196 * retransmissions due to small network reorderings, we implement
2197 * Mitigation A.3 in the RFC and delay the retransmission for a short
2198 * interval if appropriate.
2200 if (tp->do_early_retrans && !tp->retrans_out && tp->sacked_out &&
2201 (tp->packets_out >= (tp->sacked_out + 1) && tp->packets_out < 4) &&
2202 !tcp_may_send_now(sk))
2203 return !tcp_pause_early_retransmit(sk, flag);
2208 /* Detect loss in event "A" above by marking head of queue up as lost.
2209 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2210 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2211 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2212 * the maximum SACKed segments to pass before reaching this limit.
2214 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2216 struct tcp_sock *tp = tcp_sk(sk);
2217 struct sk_buff *skb;
2221 /* Use SACK to deduce losses of new sequences sent during recovery */
2222 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq;
2224 WARN_ON(packets > tp->packets_out);
2225 if (tp->lost_skb_hint) {
2226 skb = tp->lost_skb_hint;
2227 cnt = tp->lost_cnt_hint;
2228 /* Head already handled? */
2229 if (mark_head && skb != tcp_write_queue_head(sk))
2232 skb = tcp_write_queue_head(sk);
2236 tcp_for_write_queue_from(skb, sk) {
2237 if (skb == tcp_send_head(sk))
2239 /* TODO: do this better */
2240 /* this is not the most efficient way to do this... */
2241 tp->lost_skb_hint = skb;
2242 tp->lost_cnt_hint = cnt;
2244 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2248 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2249 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2250 cnt += tcp_skb_pcount(skb);
2252 if (cnt > packets) {
2253 if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2254 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2255 (oldcnt >= packets))
2258 mss = skb_shinfo(skb)->gso_size;
2259 err = tcp_fragment(sk, skb, (packets - oldcnt) * mss,
2266 tcp_skb_mark_lost(tp, skb);
2271 tcp_verify_left_out(tp);
2274 /* Account newly detected lost packet(s) */
2276 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2278 struct tcp_sock *tp = tcp_sk(sk);
2280 if (tcp_is_reno(tp)) {
2281 tcp_mark_head_lost(sk, 1, 1);
2282 } else if (tcp_is_fack(tp)) {
2283 int lost = tp->fackets_out - tp->reordering;
2286 tcp_mark_head_lost(sk, lost, 0);
2288 int sacked_upto = tp->sacked_out - tp->reordering;
2289 if (sacked_upto >= 0)
2290 tcp_mark_head_lost(sk, sacked_upto, 0);
2291 else if (fast_rexmit)
2292 tcp_mark_head_lost(sk, 1, 1);
2296 /* CWND moderation, preventing bursts due to too big ACKs
2297 * in dubious situations.
2299 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2301 tp->snd_cwnd = min(tp->snd_cwnd,
2302 tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2303 tp->snd_cwnd_stamp = tcp_time_stamp;
2306 /* Nothing was retransmitted or returned timestamp is less
2307 * than timestamp of the first retransmission.
2309 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2311 return !tp->retrans_stamp ||
2312 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2313 before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
2316 /* Undo procedures. */
2318 /* We can clear retrans_stamp when there are no retransmissions in the
2319 * window. It would seem that it is trivially available for us in
2320 * tp->retrans_out, however, that kind of assumptions doesn't consider
2321 * what will happen if errors occur when sending retransmission for the
2322 * second time. ...It could the that such segment has only
2323 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2324 * the head skb is enough except for some reneging corner cases that
2325 * are not worth the effort.
2327 * Main reason for all this complexity is the fact that connection dying
2328 * time now depends on the validity of the retrans_stamp, in particular,
2329 * that successive retransmissions of a segment must not advance
2330 * retrans_stamp under any conditions.
2332 static bool tcp_any_retrans_done(const struct sock *sk)
2334 const struct tcp_sock *tp = tcp_sk(sk);
2335 struct sk_buff *skb;
2337 if (tp->retrans_out)
2340 skb = tcp_write_queue_head(sk);
2341 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2347 #if FASTRETRANS_DEBUG > 1
2348 static void DBGUNDO(struct sock *sk, const char *msg)
2350 struct tcp_sock *tp = tcp_sk(sk);
2351 struct inet_sock *inet = inet_sk(sk);
2353 if (sk->sk_family == AF_INET) {
2354 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2356 &inet->inet_daddr, ntohs(inet->inet_dport),
2357 tp->snd_cwnd, tcp_left_out(tp),
2358 tp->snd_ssthresh, tp->prior_ssthresh,
2361 #if IS_ENABLED(CONFIG_IPV6)
2362 else if (sk->sk_family == AF_INET6) {
2363 struct ipv6_pinfo *np = inet6_sk(sk);
2364 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2366 &np->daddr, ntohs(inet->inet_dport),
2367 tp->snd_cwnd, tcp_left_out(tp),
2368 tp->snd_ssthresh, tp->prior_ssthresh,
2374 #define DBGUNDO(x...) do { } while (0)
2377 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2379 struct tcp_sock *tp = tcp_sk(sk);
2382 struct sk_buff *skb;
2384 tcp_for_write_queue(skb, sk) {
2385 if (skb == tcp_send_head(sk))
2387 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2390 tcp_clear_all_retrans_hints(tp);
2393 if (tp->prior_ssthresh) {
2394 const struct inet_connection_sock *icsk = inet_csk(sk);
2396 if (icsk->icsk_ca_ops->undo_cwnd)
2397 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2399 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2401 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2402 tp->snd_ssthresh = tp->prior_ssthresh;
2403 tcp_ecn_withdraw_cwr(tp);
2406 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2408 tp->snd_cwnd_stamp = tcp_time_stamp;
2409 tp->undo_marker = 0;
2412 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2414 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2417 /* People celebrate: "We love our President!" */
2418 static bool tcp_try_undo_recovery(struct sock *sk)
2420 struct tcp_sock *tp = tcp_sk(sk);
2422 if (tcp_may_undo(tp)) {
2425 /* Happy end! We did not retransmit anything
2426 * or our original transmission succeeded.
2428 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2429 tcp_undo_cwnd_reduction(sk, false);
2430 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2431 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2433 mib_idx = LINUX_MIB_TCPFULLUNDO;
2435 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2437 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2438 /* Hold old state until something *above* high_seq
2439 * is ACKed. For Reno it is MUST to prevent false
2440 * fast retransmits (RFC2582). SACK TCP is safe. */
2441 tcp_moderate_cwnd(tp);
2442 if (!tcp_any_retrans_done(sk))
2443 tp->retrans_stamp = 0;
2446 tcp_set_ca_state(sk, TCP_CA_Open);
2450 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2451 static bool tcp_try_undo_dsack(struct sock *sk)
2453 struct tcp_sock *tp = tcp_sk(sk);
2455 if (tp->undo_marker && !tp->undo_retrans) {
2456 DBGUNDO(sk, "D-SACK");
2457 tcp_undo_cwnd_reduction(sk, false);
2458 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2464 /* Undo during loss recovery after partial ACK or using F-RTO. */
2465 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2467 struct tcp_sock *tp = tcp_sk(sk);
2469 if (frto_undo || tcp_may_undo(tp)) {
2470 tcp_undo_cwnd_reduction(sk, true);
2472 DBGUNDO(sk, "partial loss");
2473 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2475 NET_INC_STATS_BH(sock_net(sk),
2476 LINUX_MIB_TCPSPURIOUSRTOS);
2477 inet_csk(sk)->icsk_retransmits = 0;
2478 if (frto_undo || tcp_is_sack(tp))
2479 tcp_set_ca_state(sk, TCP_CA_Open);
2485 /* The cwnd reduction in CWR and Recovery use the PRR algorithm
2486 * https://datatracker.ietf.org/doc/draft-ietf-tcpm-proportional-rate-reduction/
2487 * It computes the number of packets to send (sndcnt) based on packets newly
2489 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2490 * cwnd reductions across a full RTT.
2491 * 2) If packets in flight is lower than ssthresh (such as due to excess
2492 * losses and/or application stalls), do not perform any further cwnd
2493 * reductions, but instead slow start up to ssthresh.
2495 static void tcp_init_cwnd_reduction(struct sock *sk)
2497 struct tcp_sock *tp = tcp_sk(sk);
2499 tp->high_seq = tp->snd_nxt;
2500 tp->tlp_high_seq = 0;
2501 tp->snd_cwnd_cnt = 0;
2502 tp->prior_cwnd = tp->snd_cwnd;
2503 tp->prr_delivered = 0;
2505 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2506 tcp_ecn_queue_cwr(tp);
2509 static void tcp_cwnd_reduction(struct sock *sk, const int prior_unsacked,
2512 struct tcp_sock *tp = tcp_sk(sk);
2514 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2515 int newly_acked_sacked = prior_unsacked -
2516 (tp->packets_out - tp->sacked_out);
2518 tp->prr_delivered += newly_acked_sacked;
2519 if (tcp_packets_in_flight(tp) > tp->snd_ssthresh) {
2520 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2522 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2524 sndcnt = min_t(int, delta,
2525 max_t(int, tp->prr_delivered - tp->prr_out,
2526 newly_acked_sacked) + 1);
2529 sndcnt = max(sndcnt, (fast_rexmit ? 1 : 0));
2530 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2533 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2535 struct tcp_sock *tp = tcp_sk(sk);
2537 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2538 if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR ||
2539 (tp->undo_marker && tp->snd_ssthresh < TCP_INFINITE_SSTHRESH)) {
2540 tp->snd_cwnd = tp->snd_ssthresh;
2541 tp->snd_cwnd_stamp = tcp_time_stamp;
2543 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2546 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2547 void tcp_enter_cwr(struct sock *sk)
2549 struct tcp_sock *tp = tcp_sk(sk);
2551 tp->prior_ssthresh = 0;
2552 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2553 tp->undo_marker = 0;
2554 tcp_init_cwnd_reduction(sk);
2555 tcp_set_ca_state(sk, TCP_CA_CWR);
2559 static void tcp_try_keep_open(struct sock *sk)
2561 struct tcp_sock *tp = tcp_sk(sk);
2562 int state = TCP_CA_Open;
2564 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2565 state = TCP_CA_Disorder;
2567 if (inet_csk(sk)->icsk_ca_state != state) {
2568 tcp_set_ca_state(sk, state);
2569 tp->high_seq = tp->snd_nxt;
2573 static void tcp_try_to_open(struct sock *sk, int flag, const int prior_unsacked)
2575 struct tcp_sock *tp = tcp_sk(sk);
2577 tcp_verify_left_out(tp);
2579 if (!tcp_any_retrans_done(sk))
2580 tp->retrans_stamp = 0;
2582 if (flag & FLAG_ECE)
2585 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2586 tcp_try_keep_open(sk);
2588 tcp_cwnd_reduction(sk, prior_unsacked, 0);
2592 static void tcp_mtup_probe_failed(struct sock *sk)
2594 struct inet_connection_sock *icsk = inet_csk(sk);
2596 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2597 icsk->icsk_mtup.probe_size = 0;
2600 static void tcp_mtup_probe_success(struct sock *sk)
2602 struct tcp_sock *tp = tcp_sk(sk);
2603 struct inet_connection_sock *icsk = inet_csk(sk);
2605 /* FIXME: breaks with very large cwnd */
2606 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2607 tp->snd_cwnd = tp->snd_cwnd *
2608 tcp_mss_to_mtu(sk, tp->mss_cache) /
2609 icsk->icsk_mtup.probe_size;
2610 tp->snd_cwnd_cnt = 0;
2611 tp->snd_cwnd_stamp = tcp_time_stamp;
2612 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2614 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2615 icsk->icsk_mtup.probe_size = 0;
2616 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2619 /* Do a simple retransmit without using the backoff mechanisms in
2620 * tcp_timer. This is used for path mtu discovery.
2621 * The socket is already locked here.
2623 void tcp_simple_retransmit(struct sock *sk)
2625 const struct inet_connection_sock *icsk = inet_csk(sk);
2626 struct tcp_sock *tp = tcp_sk(sk);
2627 struct sk_buff *skb;
2628 unsigned int mss = tcp_current_mss(sk);
2629 u32 prior_lost = tp->lost_out;
2631 tcp_for_write_queue(skb, sk) {
2632 if (skb == tcp_send_head(sk))
2634 if (tcp_skb_seglen(skb) > mss &&
2635 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2636 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2637 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2638 tp->retrans_out -= tcp_skb_pcount(skb);
2640 tcp_skb_mark_lost_uncond_verify(tp, skb);
2644 tcp_clear_retrans_hints_partial(tp);
2646 if (prior_lost == tp->lost_out)
2649 if (tcp_is_reno(tp))
2650 tcp_limit_reno_sacked(tp);
2652 tcp_verify_left_out(tp);
2654 /* Don't muck with the congestion window here.
2655 * Reason is that we do not increase amount of _data_
2656 * in network, but units changed and effective
2657 * cwnd/ssthresh really reduced now.
2659 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2660 tp->high_seq = tp->snd_nxt;
2661 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2662 tp->prior_ssthresh = 0;
2663 tp->undo_marker = 0;
2664 tcp_set_ca_state(sk, TCP_CA_Loss);
2666 tcp_xmit_retransmit_queue(sk);
2668 EXPORT_SYMBOL(tcp_simple_retransmit);
2670 static void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2672 struct tcp_sock *tp = tcp_sk(sk);
2675 if (tcp_is_reno(tp))
2676 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2678 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2680 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2682 tp->prior_ssthresh = 0;
2685 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2687 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2688 tcp_init_cwnd_reduction(sk);
2690 tcp_set_ca_state(sk, TCP_CA_Recovery);
2693 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2694 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2696 static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack)
2698 struct tcp_sock *tp = tcp_sk(sk);
2699 bool recovered = !before(tp->snd_una, tp->high_seq);
2701 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2702 /* Step 3.b. A timeout is spurious if not all data are
2703 * lost, i.e., never-retransmitted data are (s)acked.
2705 if (tcp_try_undo_loss(sk, flag & FLAG_ORIG_SACK_ACKED))
2708 if (after(tp->snd_nxt, tp->high_seq) &&
2709 (flag & FLAG_DATA_SACKED || is_dupack)) {
2710 tp->frto = 0; /* Loss was real: 2nd part of step 3.a */
2711 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2712 tp->high_seq = tp->snd_nxt;
2713 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
2715 if (after(tp->snd_nxt, tp->high_seq))
2716 return; /* Step 2.b */
2722 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2723 tcp_try_undo_recovery(sk);
2726 if (tcp_is_reno(tp)) {
2727 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2728 * delivered. Lower inflight to clock out (re)tranmissions.
2730 if (after(tp->snd_nxt, tp->high_seq) && is_dupack)
2731 tcp_add_reno_sack(sk);
2732 else if (flag & FLAG_SND_UNA_ADVANCED)
2733 tcp_reset_reno_sack(tp);
2735 if (tcp_try_undo_loss(sk, false))
2737 tcp_xmit_retransmit_queue(sk);
2740 /* Undo during fast recovery after partial ACK. */
2741 static bool tcp_try_undo_partial(struct sock *sk, const int acked,
2742 const int prior_unsacked)
2744 struct tcp_sock *tp = tcp_sk(sk);
2746 if (tp->undo_marker && tcp_packet_delayed(tp)) {
2747 /* Plain luck! Hole if filled with delayed
2748 * packet, rather than with a retransmit.
2750 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2752 /* We are getting evidence that the reordering degree is higher
2753 * than we realized. If there are no retransmits out then we
2754 * can undo. Otherwise we clock out new packets but do not
2755 * mark more packets lost or retransmit more.
2757 if (tp->retrans_out) {
2758 tcp_cwnd_reduction(sk, prior_unsacked, 0);
2762 if (!tcp_any_retrans_done(sk))
2763 tp->retrans_stamp = 0;
2765 DBGUNDO(sk, "partial recovery");
2766 tcp_undo_cwnd_reduction(sk, true);
2767 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2768 tcp_try_keep_open(sk);
2774 /* Process an event, which can update packets-in-flight not trivially.
2775 * Main goal of this function is to calculate new estimate for left_out,
2776 * taking into account both packets sitting in receiver's buffer and
2777 * packets lost by network.
2779 * Besides that it does CWND reduction, when packet loss is detected
2780 * and changes state of machine.
2782 * It does _not_ decide what to send, it is made in function
2783 * tcp_xmit_retransmit_queue().
2785 static void tcp_fastretrans_alert(struct sock *sk, const int acked,
2786 const int prior_unsacked,
2787 bool is_dupack, int flag)
2789 struct inet_connection_sock *icsk = inet_csk(sk);
2790 struct tcp_sock *tp = tcp_sk(sk);
2791 bool do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2792 (tcp_fackets_out(tp) > tp->reordering));
2793 int fast_rexmit = 0;
2795 if (WARN_ON(!tp->packets_out && tp->sacked_out))
2797 if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2798 tp->fackets_out = 0;
2800 /* Now state machine starts.
2801 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2802 if (flag & FLAG_ECE)
2803 tp->prior_ssthresh = 0;
2805 /* B. In all the states check for reneging SACKs. */
2806 if (tcp_check_sack_reneging(sk, flag))
2809 /* C. Check consistency of the current state. */
2810 tcp_verify_left_out(tp);
2812 /* D. Check state exit conditions. State can be terminated
2813 * when high_seq is ACKed. */
2814 if (icsk->icsk_ca_state == TCP_CA_Open) {
2815 WARN_ON(tp->retrans_out != 0);
2816 tp->retrans_stamp = 0;
2817 } else if (!before(tp->snd_una, tp->high_seq)) {
2818 switch (icsk->icsk_ca_state) {
2820 /* CWR is to be held something *above* high_seq
2821 * is ACKed for CWR bit to reach receiver. */
2822 if (tp->snd_una != tp->high_seq) {
2823 tcp_end_cwnd_reduction(sk);
2824 tcp_set_ca_state(sk, TCP_CA_Open);
2828 case TCP_CA_Recovery:
2829 if (tcp_is_reno(tp))
2830 tcp_reset_reno_sack(tp);
2831 if (tcp_try_undo_recovery(sk))
2833 tcp_end_cwnd_reduction(sk);
2838 /* E. Process state. */
2839 switch (icsk->icsk_ca_state) {
2840 case TCP_CA_Recovery:
2841 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2842 if (tcp_is_reno(tp) && is_dupack)
2843 tcp_add_reno_sack(sk);
2845 if (tcp_try_undo_partial(sk, acked, prior_unsacked))
2847 /* Partial ACK arrived. Force fast retransmit. */
2848 do_lost = tcp_is_reno(tp) ||
2849 tcp_fackets_out(tp) > tp->reordering;
2851 if (tcp_try_undo_dsack(sk)) {
2852 tcp_try_keep_open(sk);
2857 tcp_process_loss(sk, flag, is_dupack);
2858 if (icsk->icsk_ca_state != TCP_CA_Open)
2860 /* Fall through to processing in Open state. */
2862 if (tcp_is_reno(tp)) {
2863 if (flag & FLAG_SND_UNA_ADVANCED)
2864 tcp_reset_reno_sack(tp);
2866 tcp_add_reno_sack(sk);
2869 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2870 tcp_try_undo_dsack(sk);
2872 if (!tcp_time_to_recover(sk, flag)) {
2873 tcp_try_to_open(sk, flag, prior_unsacked);
2877 /* MTU probe failure: don't reduce cwnd */
2878 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2879 icsk->icsk_mtup.probe_size &&
2880 tp->snd_una == tp->mtu_probe.probe_seq_start) {
2881 tcp_mtup_probe_failed(sk);
2882 /* Restores the reduction we did in tcp_mtup_probe() */
2884 tcp_simple_retransmit(sk);
2888 /* Otherwise enter Recovery state */
2889 tcp_enter_recovery(sk, (flag & FLAG_ECE));
2894 tcp_update_scoreboard(sk, fast_rexmit);
2895 tcp_cwnd_reduction(sk, prior_unsacked, fast_rexmit);
2896 tcp_xmit_retransmit_queue(sk);
2899 static inline bool tcp_ack_update_rtt(struct sock *sk, const int flag,
2900 long seq_rtt_us, long sack_rtt_us)
2902 const struct tcp_sock *tp = tcp_sk(sk);
2904 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2905 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2906 * Karn's algorithm forbids taking RTT if some retransmitted data
2907 * is acked (RFC6298).
2909 if (flag & FLAG_RETRANS_DATA_ACKED)
2913 seq_rtt_us = sack_rtt_us;
2915 /* RTTM Rule: A TSecr value received in a segment is used to
2916 * update the averaged RTT measurement only if the segment
2917 * acknowledges some new data, i.e., only if it advances the
2918 * left edge of the send window.
2919 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2921 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2923 seq_rtt_us = jiffies_to_usecs(tcp_time_stamp - tp->rx_opt.rcv_tsecr);
2928 tcp_rtt_estimator(sk, seq_rtt_us);
2931 /* RFC6298: only reset backoff on valid RTT measurement. */
2932 inet_csk(sk)->icsk_backoff = 0;
2936 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2937 static void tcp_synack_rtt_meas(struct sock *sk, const u32 synack_stamp)
2939 struct tcp_sock *tp = tcp_sk(sk);
2940 long seq_rtt_us = -1L;
2942 if (synack_stamp && !tp->total_retrans)
2943 seq_rtt_us = jiffies_to_usecs(tcp_time_stamp - synack_stamp);
2945 /* If the ACK acks both the SYNACK and the (Fast Open'd) data packets
2946 * sent in SYN_RECV, SYNACK RTT is the smooth RTT computed in tcp_ack()
2949 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, seq_rtt_us, -1L);
2952 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
2954 const struct inet_connection_sock *icsk = inet_csk(sk);
2956 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
2957 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
2960 /* Restart timer after forward progress on connection.
2961 * RFC2988 recommends to restart timer to now+rto.
2963 void tcp_rearm_rto(struct sock *sk)
2965 const struct inet_connection_sock *icsk = inet_csk(sk);
2966 struct tcp_sock *tp = tcp_sk(sk);
2968 /* If the retrans timer is currently being used by Fast Open
2969 * for SYN-ACK retrans purpose, stay put.
2971 if (tp->fastopen_rsk)
2974 if (!tp->packets_out) {
2975 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2977 u32 rto = inet_csk(sk)->icsk_rto;
2978 /* Offset the time elapsed after installing regular RTO */
2979 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
2980 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
2981 struct sk_buff *skb = tcp_write_queue_head(sk);
2982 const u32 rto_time_stamp =
2983 tcp_skb_timestamp(skb) + rto;
2984 s32 delta = (s32)(rto_time_stamp - tcp_time_stamp);
2985 /* delta may not be positive if the socket is locked
2986 * when the retrans timer fires and is rescheduled.
2991 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
2996 /* This function is called when the delayed ER timer fires. TCP enters
2997 * fast recovery and performs fast-retransmit.
2999 void tcp_resume_early_retransmit(struct sock *sk)
3001 struct tcp_sock *tp = tcp_sk(sk);
3005 /* Stop if ER is disabled after the delayed ER timer is scheduled */
3006 if (!tp->do_early_retrans)
3009 tcp_enter_recovery(sk, false);
3010 tcp_update_scoreboard(sk, 1);
3011 tcp_xmit_retransmit_queue(sk);
3014 /* If we get here, the whole TSO packet has not been acked. */
3015 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3017 struct tcp_sock *tp = tcp_sk(sk);
3020 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3022 packets_acked = tcp_skb_pcount(skb);
3023 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3025 packets_acked -= tcp_skb_pcount(skb);
3027 if (packets_acked) {
3028 BUG_ON(tcp_skb_pcount(skb) == 0);
3029 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3032 return packets_acked;
3035 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3038 const struct skb_shared_info *shinfo;
3040 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3041 if (likely(!(sk->sk_tsflags & SOF_TIMESTAMPING_TX_ACK)))
3044 shinfo = skb_shinfo(skb);
3045 if ((shinfo->tx_flags & SKBTX_ACK_TSTAMP) &&
3046 between(shinfo->tskey, prior_snd_una, tcp_sk(sk)->snd_una - 1))
3047 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK);
3050 /* Remove acknowledged frames from the retransmission queue. If our packet
3051 * is before the ack sequence we can discard it as it's confirmed to have
3052 * arrived at the other end.
3054 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3055 u32 prior_snd_una, long sack_rtt_us)
3057 const struct inet_connection_sock *icsk = inet_csk(sk);
3058 struct skb_mstamp first_ackt, last_ackt, now;
3059 struct tcp_sock *tp = tcp_sk(sk);
3060 u32 prior_sacked = tp->sacked_out;
3061 u32 reord = tp->packets_out;
3062 bool fully_acked = true;
3063 long ca_seq_rtt_us = -1L;
3064 long seq_rtt_us = -1L;
3065 struct sk_buff *skb;
3072 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3073 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3074 u8 sacked = scb->sacked;
3077 tcp_ack_tstamp(sk, skb, prior_snd_una);
3079 /* Determine how many packets and what bytes were acked, tso and else */
3080 if (after(scb->end_seq, tp->snd_una)) {
3081 if (tcp_skb_pcount(skb) == 1 ||
3082 !after(tp->snd_una, scb->seq))
3085 acked_pcount = tcp_tso_acked(sk, skb);
3089 fully_acked = false;
3091 /* Speedup tcp_unlink_write_queue() and next loop */
3092 prefetchw(skb->next);
3093 acked_pcount = tcp_skb_pcount(skb);
3096 if (unlikely(sacked & TCPCB_RETRANS)) {
3097 if (sacked & TCPCB_SACKED_RETRANS)
3098 tp->retrans_out -= acked_pcount;
3099 flag |= FLAG_RETRANS_DATA_ACKED;
3101 last_ackt = skb->skb_mstamp;
3102 WARN_ON_ONCE(last_ackt.v64 == 0);
3103 if (!first_ackt.v64)
3104 first_ackt = last_ackt;
3106 if (!(sacked & TCPCB_SACKED_ACKED))
3107 reord = min(pkts_acked, reord);
3108 if (!after(scb->end_seq, tp->high_seq))
3109 flag |= FLAG_ORIG_SACK_ACKED;
3112 if (sacked & TCPCB_SACKED_ACKED)
3113 tp->sacked_out -= acked_pcount;
3114 if (sacked & TCPCB_LOST)
3115 tp->lost_out -= acked_pcount;
3117 tp->packets_out -= acked_pcount;
3118 pkts_acked += acked_pcount;
3120 /* Initial outgoing SYN's get put onto the write_queue
3121 * just like anything else we transmit. It is not
3122 * true data, and if we misinform our callers that
3123 * this ACK acks real data, we will erroneously exit
3124 * connection startup slow start one packet too
3125 * quickly. This is severely frowned upon behavior.
3127 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3128 flag |= FLAG_DATA_ACKED;
3130 flag |= FLAG_SYN_ACKED;
3131 tp->retrans_stamp = 0;
3137 tcp_unlink_write_queue(skb, sk);
3138 sk_wmem_free_skb(sk, skb);
3139 if (unlikely(skb == tp->retransmit_skb_hint))
3140 tp->retransmit_skb_hint = NULL;
3141 if (unlikely(skb == tp->lost_skb_hint))
3142 tp->lost_skb_hint = NULL;
3145 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3146 tp->snd_up = tp->snd_una;
3148 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3149 flag |= FLAG_SACK_RENEGING;
3151 skb_mstamp_get(&now);
3152 if (likely(first_ackt.v64)) {
3153 seq_rtt_us = skb_mstamp_us_delta(&now, &first_ackt);
3154 ca_seq_rtt_us = skb_mstamp_us_delta(&now, &last_ackt);
3157 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us);
3159 if (flag & FLAG_ACKED) {
3160 const struct tcp_congestion_ops *ca_ops
3161 = inet_csk(sk)->icsk_ca_ops;
3164 if (unlikely(icsk->icsk_mtup.probe_size &&
3165 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3166 tcp_mtup_probe_success(sk);
3169 if (tcp_is_reno(tp)) {
3170 tcp_remove_reno_sacks(sk, pkts_acked);
3174 /* Non-retransmitted hole got filled? That's reordering */
3175 if (reord < prior_fackets)
3176 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3178 delta = tcp_is_fack(tp) ? pkts_acked :
3179 prior_sacked - tp->sacked_out;
3180 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3183 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3185 if (ca_ops->pkts_acked)
3186 ca_ops->pkts_acked(sk, pkts_acked, ca_seq_rtt_us);
3188 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3189 sack_rtt_us > skb_mstamp_us_delta(&now, &skb->skb_mstamp)) {
3190 /* Do not re-arm RTO if the sack RTT is measured from data sent
3191 * after when the head was last (re)transmitted. Otherwise the
3192 * timeout may continue to extend in loss recovery.
3197 #if FASTRETRANS_DEBUG > 0
3198 WARN_ON((int)tp->sacked_out < 0);
3199 WARN_ON((int)tp->lost_out < 0);
3200 WARN_ON((int)tp->retrans_out < 0);
3201 if (!tp->packets_out && tcp_is_sack(tp)) {
3202 icsk = inet_csk(sk);
3204 pr_debug("Leak l=%u %d\n",
3205 tp->lost_out, icsk->icsk_ca_state);
3208 if (tp->sacked_out) {
3209 pr_debug("Leak s=%u %d\n",
3210 tp->sacked_out, icsk->icsk_ca_state);
3213 if (tp->retrans_out) {
3214 pr_debug("Leak r=%u %d\n",
3215 tp->retrans_out, icsk->icsk_ca_state);
3216 tp->retrans_out = 0;
3223 static void tcp_ack_probe(struct sock *sk)
3225 const struct tcp_sock *tp = tcp_sk(sk);
3226 struct inet_connection_sock *icsk = inet_csk(sk);
3228 /* Was it a usable window open? */
3230 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3231 icsk->icsk_backoff = 0;
3232 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3233 /* Socket must be waked up by subsequent tcp_data_snd_check().
3234 * This function is not for random using!
3237 unsigned long when = inet_csk_rto_backoff(icsk, TCP_RTO_MAX);
3239 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3244 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3246 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3247 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3250 /* Decide wheather to run the increase function of congestion control. */
3251 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3253 if (tcp_in_cwnd_reduction(sk))
3256 /* If reordering is high then always grow cwnd whenever data is
3257 * delivered regardless of its ordering. Otherwise stay conservative
3258 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3259 * new SACK or ECE mark may first advance cwnd here and later reduce
3260 * cwnd in tcp_fastretrans_alert() based on more states.
3262 if (tcp_sk(sk)->reordering > sysctl_tcp_reordering)
3263 return flag & FLAG_FORWARD_PROGRESS;
3265 return flag & FLAG_DATA_ACKED;
3268 /* Check that window update is acceptable.
3269 * The function assumes that snd_una<=ack<=snd_next.
3271 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3272 const u32 ack, const u32 ack_seq,
3275 return after(ack, tp->snd_una) ||
3276 after(ack_seq, tp->snd_wl1) ||
3277 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3280 /* Update our send window.
3282 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3283 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3285 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3288 struct tcp_sock *tp = tcp_sk(sk);
3290 u32 nwin = ntohs(tcp_hdr(skb)->window);
3292 if (likely(!tcp_hdr(skb)->syn))
3293 nwin <<= tp->rx_opt.snd_wscale;
3295 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3296 flag |= FLAG_WIN_UPDATE;
3297 tcp_update_wl(tp, ack_seq);
3299 if (tp->snd_wnd != nwin) {
3302 /* Note, it is the only place, where
3303 * fast path is recovered for sending TCP.
3306 tcp_fast_path_check(sk);
3308 if (nwin > tp->max_window) {
3309 tp->max_window = nwin;
3310 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3320 /* RFC 5961 7 [ACK Throttling] */
3321 static void tcp_send_challenge_ack(struct sock *sk)
3323 /* unprotected vars, we dont care of overwrites */
3324 static u32 challenge_timestamp;
3325 static unsigned int challenge_count;
3326 u32 now = jiffies / HZ;
3328 if (now != challenge_timestamp) {
3329 challenge_timestamp = now;
3330 challenge_count = 0;
3332 if (++challenge_count <= sysctl_tcp_challenge_ack_limit) {
3333 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPCHALLENGEACK);
3338 static void tcp_store_ts_recent(struct tcp_sock *tp)
3340 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3341 tp->rx_opt.ts_recent_stamp = get_seconds();
3344 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3346 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3347 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3348 * extra check below makes sure this can only happen
3349 * for pure ACK frames. -DaveM
3351 * Not only, also it occurs for expired timestamps.
3354 if (tcp_paws_check(&tp->rx_opt, 0))
3355 tcp_store_ts_recent(tp);
3359 /* This routine deals with acks during a TLP episode.
3360 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3362 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3364 struct tcp_sock *tp = tcp_sk(sk);
3365 bool is_tlp_dupack = (ack == tp->tlp_high_seq) &&
3366 !(flag & (FLAG_SND_UNA_ADVANCED |
3367 FLAG_NOT_DUP | FLAG_DATA_SACKED));
3369 /* Mark the end of TLP episode on receiving TLP dupack or when
3370 * ack is after tlp_high_seq.
3372 if (is_tlp_dupack) {
3373 tp->tlp_high_seq = 0;
3377 if (after(ack, tp->tlp_high_seq)) {
3378 tp->tlp_high_seq = 0;
3379 /* Don't reduce cwnd if DSACK arrives for TLP retrans. */
3380 if (!(flag & FLAG_DSACKING_ACK)) {
3381 tcp_init_cwnd_reduction(sk);
3382 tcp_set_ca_state(sk, TCP_CA_CWR);
3383 tcp_end_cwnd_reduction(sk);
3384 tcp_try_keep_open(sk);
3385 NET_INC_STATS_BH(sock_net(sk),
3386 LINUX_MIB_TCPLOSSPROBERECOVERY);
3391 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3393 const struct inet_connection_sock *icsk = inet_csk(sk);
3395 if (icsk->icsk_ca_ops->in_ack_event)
3396 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3399 /* This routine deals with incoming acks, but not outgoing ones. */
3400 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3402 struct inet_connection_sock *icsk = inet_csk(sk);
3403 struct tcp_sock *tp = tcp_sk(sk);
3404 u32 prior_snd_una = tp->snd_una;
3405 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3406 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3407 bool is_dupack = false;
3409 int prior_packets = tp->packets_out;
3410 const int prior_unsacked = tp->packets_out - tp->sacked_out;
3411 int acked = 0; /* Number of packets newly acked */
3412 long sack_rtt_us = -1L;
3414 /* We very likely will need to access write queue head. */
3415 prefetchw(sk->sk_write_queue.next);
3417 /* If the ack is older than previous acks
3418 * then we can probably ignore it.
3420 if (before(ack, prior_snd_una)) {
3421 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3422 if (before(ack, prior_snd_una - tp->max_window)) {
3423 tcp_send_challenge_ack(sk);
3429 /* If the ack includes data we haven't sent yet, discard
3430 * this segment (RFC793 Section 3.9).
3432 if (after(ack, tp->snd_nxt))
3435 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
3436 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
3439 if (after(ack, prior_snd_una)) {
3440 flag |= FLAG_SND_UNA_ADVANCED;
3441 icsk->icsk_retransmits = 0;
3444 prior_fackets = tp->fackets_out;
3446 /* ts_recent update must be made after we are sure that the packet
3449 if (flag & FLAG_UPDATE_TS_RECENT)
3450 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3452 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3453 /* Window is constant, pure forward advance.
3454 * No more checks are required.
3455 * Note, we use the fact that SND.UNA>=SND.WL2.
3457 tcp_update_wl(tp, ack_seq);
3459 flag |= FLAG_WIN_UPDATE;
3461 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3463 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3465 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3467 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3470 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3472 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3474 if (TCP_SKB_CB(skb)->sacked)
3475 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3478 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3480 ack_ev_flags |= CA_ACK_ECE;
3483 if (flag & FLAG_WIN_UPDATE)
3484 ack_ev_flags |= CA_ACK_WIN_UPDATE;
3486 tcp_in_ack_event(sk, ack_ev_flags);
3489 /* We passed data and got it acked, remove any soft error
3490 * log. Something worked...
3492 sk->sk_err_soft = 0;
3493 icsk->icsk_probes_out = 0;
3494 tp->rcv_tstamp = tcp_time_stamp;
3498 /* See if we can take anything off of the retransmit queue. */
3499 acked = tp->packets_out;
3500 flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una,
3502 acked -= tp->packets_out;
3504 /* Advance cwnd if state allows */
3505 if (tcp_may_raise_cwnd(sk, flag))
3506 tcp_cong_avoid(sk, ack, acked);
3508 if (tcp_ack_is_dubious(sk, flag)) {
3509 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3510 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3513 if (tp->tlp_high_seq)
3514 tcp_process_tlp_ack(sk, ack, flag);
3516 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) {
3517 struct dst_entry *dst = __sk_dst_get(sk);
3522 if (icsk->icsk_pending == ICSK_TIME_RETRANS)
3523 tcp_schedule_loss_probe(sk);
3524 tcp_update_pacing_rate(sk);
3528 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3529 if (flag & FLAG_DSACKING_ACK)
3530 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3532 /* If this ack opens up a zero window, clear backoff. It was
3533 * being used to time the probes, and is probably far higher than
3534 * it needs to be for normal retransmission.
3536 if (tcp_send_head(sk))
3539 if (tp->tlp_high_seq)
3540 tcp_process_tlp_ack(sk, ack, flag);
3544 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3548 /* If data was SACKed, tag it and see if we should send more data.
3549 * If data was DSACKed, see if we can undo a cwnd reduction.
3551 if (TCP_SKB_CB(skb)->sacked) {
3552 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3554 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3558 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3562 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3563 * But, this can also be called on packets in the established flow when
3564 * the fast version below fails.
3566 void tcp_parse_options(const struct sk_buff *skb,
3567 struct tcp_options_received *opt_rx, int estab,
3568 struct tcp_fastopen_cookie *foc)
3570 const unsigned char *ptr;
3571 const struct tcphdr *th = tcp_hdr(skb);
3572 int length = (th->doff * 4) - sizeof(struct tcphdr);
3574 ptr = (const unsigned char *)(th + 1);
3575 opt_rx->saw_tstamp = 0;
3577 while (length > 0) {
3578 int opcode = *ptr++;
3584 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3589 if (opsize < 2) /* "silly options" */
3591 if (opsize > length)
3592 return; /* don't parse partial options */
3595 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3596 u16 in_mss = get_unaligned_be16(ptr);
3598 if (opt_rx->user_mss &&
3599 opt_rx->user_mss < in_mss)
3600 in_mss = opt_rx->user_mss;
3601 opt_rx->mss_clamp = in_mss;
3606 if (opsize == TCPOLEN_WINDOW && th->syn &&
3607 !estab && sysctl_tcp_window_scaling) {
3608 __u8 snd_wscale = *(__u8 *)ptr;
3609 opt_rx->wscale_ok = 1;
3610 if (snd_wscale > 14) {
3611 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3616 opt_rx->snd_wscale = snd_wscale;
3619 case TCPOPT_TIMESTAMP:
3620 if ((opsize == TCPOLEN_TIMESTAMP) &&
3621 ((estab && opt_rx->tstamp_ok) ||
3622 (!estab && sysctl_tcp_timestamps))) {
3623 opt_rx->saw_tstamp = 1;
3624 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3625 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3628 case TCPOPT_SACK_PERM:
3629 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3630 !estab && sysctl_tcp_sack) {
3631 opt_rx->sack_ok = TCP_SACK_SEEN;
3632 tcp_sack_reset(opt_rx);
3637 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3638 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3640 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3643 #ifdef CONFIG_TCP_MD5SIG
3646 * The MD5 Hash has already been
3647 * checked (see tcp_v{4,6}_do_rcv()).
3652 /* Fast Open option shares code 254 using a
3653 * 16 bits magic number. It's valid only in
3654 * SYN or SYN-ACK with an even size.
3656 if (opsize < TCPOLEN_EXP_FASTOPEN_BASE ||
3657 get_unaligned_be16(ptr) != TCPOPT_FASTOPEN_MAGIC ||
3658 foc == NULL || !th->syn || (opsize & 1))
3660 foc->len = opsize - TCPOLEN_EXP_FASTOPEN_BASE;
3661 if (foc->len >= TCP_FASTOPEN_COOKIE_MIN &&
3662 foc->len <= TCP_FASTOPEN_COOKIE_MAX)
3663 memcpy(foc->val, ptr + 2, foc->len);
3664 else if (foc->len != 0)
3674 EXPORT_SYMBOL(tcp_parse_options);
3676 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3678 const __be32 *ptr = (const __be32 *)(th + 1);
3680 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3681 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3682 tp->rx_opt.saw_tstamp = 1;
3684 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3687 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3689 tp->rx_opt.rcv_tsecr = 0;
3695 /* Fast parse options. This hopes to only see timestamps.
3696 * If it is wrong it falls back on tcp_parse_options().
3698 static bool tcp_fast_parse_options(const struct sk_buff *skb,
3699 const struct tcphdr *th, struct tcp_sock *tp)
3701 /* In the spirit of fast parsing, compare doff directly to constant
3702 * values. Because equality is used, short doff can be ignored here.
3704 if (th->doff == (sizeof(*th) / 4)) {
3705 tp->rx_opt.saw_tstamp = 0;
3707 } else if (tp->rx_opt.tstamp_ok &&
3708 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3709 if (tcp_parse_aligned_timestamp(tp, th))
3713 tcp_parse_options(skb, &tp->rx_opt, 1, NULL);
3714 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3715 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3720 #ifdef CONFIG_TCP_MD5SIG
3722 * Parse MD5 Signature option
3724 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3726 int length = (th->doff << 2) - sizeof(*th);
3727 const u8 *ptr = (const u8 *)(th + 1);
3729 /* If the TCP option is too short, we can short cut */
3730 if (length < TCPOLEN_MD5SIG)
3733 while (length > 0) {
3734 int opcode = *ptr++;
3745 if (opsize < 2 || opsize > length)
3747 if (opcode == TCPOPT_MD5SIG)
3748 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3755 EXPORT_SYMBOL(tcp_parse_md5sig_option);
3758 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3760 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3761 * it can pass through stack. So, the following predicate verifies that
3762 * this segment is not used for anything but congestion avoidance or
3763 * fast retransmit. Moreover, we even are able to eliminate most of such
3764 * second order effects, if we apply some small "replay" window (~RTO)
3765 * to timestamp space.
3767 * All these measures still do not guarantee that we reject wrapped ACKs
3768 * on networks with high bandwidth, when sequence space is recycled fastly,
3769 * but it guarantees that such events will be very rare and do not affect
3770 * connection seriously. This doesn't look nice, but alas, PAWS is really
3773 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3774 * states that events when retransmit arrives after original data are rare.
3775 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3776 * the biggest problem on large power networks even with minor reordering.
3777 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3778 * up to bandwidth of 18Gigabit/sec. 8) ]
3781 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3783 const struct tcp_sock *tp = tcp_sk(sk);
3784 const struct tcphdr *th = tcp_hdr(skb);
3785 u32 seq = TCP_SKB_CB(skb)->seq;
3786 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3788 return (/* 1. Pure ACK with correct sequence number. */
3789 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3791 /* 2. ... and duplicate ACK. */
3792 ack == tp->snd_una &&
3794 /* 3. ... and does not update window. */
3795 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3797 /* 4. ... and sits in replay window. */
3798 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3801 static inline bool tcp_paws_discard(const struct sock *sk,
3802 const struct sk_buff *skb)
3804 const struct tcp_sock *tp = tcp_sk(sk);
3806 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
3807 !tcp_disordered_ack(sk, skb);
3810 /* Check segment sequence number for validity.
3812 * Segment controls are considered valid, if the segment
3813 * fits to the window after truncation to the window. Acceptability
3814 * of data (and SYN, FIN, of course) is checked separately.
3815 * See tcp_data_queue(), for example.
3817 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3818 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3819 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3820 * (borrowed from freebsd)
3823 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
3825 return !before(end_seq, tp->rcv_wup) &&
3826 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3829 /* When we get a reset we do this. */
3830 void tcp_reset(struct sock *sk)
3832 /* We want the right error as BSD sees it (and indeed as we do). */
3833 switch (sk->sk_state) {
3835 sk->sk_err = ECONNREFUSED;
3837 case TCP_CLOSE_WAIT:
3843 sk->sk_err = ECONNRESET;
3845 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3848 if (!sock_flag(sk, SOCK_DEAD))
3849 sk->sk_error_report(sk);
3855 * Process the FIN bit. This now behaves as it is supposed to work
3856 * and the FIN takes effect when it is validly part of sequence
3857 * space. Not before when we get holes.
3859 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3860 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3863 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3864 * close and we go into CLOSING (and later onto TIME-WAIT)
3866 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3868 static void tcp_fin(struct sock *sk)
3870 struct tcp_sock *tp = tcp_sk(sk);
3871 const struct dst_entry *dst;
3873 inet_csk_schedule_ack(sk);
3875 sk->sk_shutdown |= RCV_SHUTDOWN;
3876 sock_set_flag(sk, SOCK_DONE);
3878 switch (sk->sk_state) {
3880 case TCP_ESTABLISHED:
3881 /* Move to CLOSE_WAIT */
3882 tcp_set_state(sk, TCP_CLOSE_WAIT);
3883 dst = __sk_dst_get(sk);
3884 if (!dst || !dst_metric(dst, RTAX_QUICKACK))
3885 inet_csk(sk)->icsk_ack.pingpong = 1;
3888 case TCP_CLOSE_WAIT:
3890 /* Received a retransmission of the FIN, do
3895 /* RFC793: Remain in the LAST-ACK state. */
3899 /* This case occurs when a simultaneous close
3900 * happens, we must ack the received FIN and
3901 * enter the CLOSING state.
3904 tcp_set_state(sk, TCP_CLOSING);
3907 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3909 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
3912 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3913 * cases we should never reach this piece of code.
3915 pr_err("%s: Impossible, sk->sk_state=%d\n",
3916 __func__, sk->sk_state);
3920 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3921 * Probably, we should reset in this case. For now drop them.
3923 __skb_queue_purge(&tp->out_of_order_queue);
3924 if (tcp_is_sack(tp))
3925 tcp_sack_reset(&tp->rx_opt);
3928 if (!sock_flag(sk, SOCK_DEAD)) {
3929 sk->sk_state_change(sk);
3931 /* Do not send POLL_HUP for half duplex close. */
3932 if (sk->sk_shutdown == SHUTDOWN_MASK ||
3933 sk->sk_state == TCP_CLOSE)
3934 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
3936 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
3940 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
3943 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
3944 if (before(seq, sp->start_seq))
3945 sp->start_seq = seq;
3946 if (after(end_seq, sp->end_seq))
3947 sp->end_seq = end_seq;
3953 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
3955 struct tcp_sock *tp = tcp_sk(sk);
3957 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3960 if (before(seq, tp->rcv_nxt))
3961 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
3963 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
3965 NET_INC_STATS_BH(sock_net(sk), mib_idx);
3967 tp->rx_opt.dsack = 1;
3968 tp->duplicate_sack[0].start_seq = seq;
3969 tp->duplicate_sack[0].end_seq = end_seq;
3973 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
3975 struct tcp_sock *tp = tcp_sk(sk);
3977 if (!tp->rx_opt.dsack)
3978 tcp_dsack_set(sk, seq, end_seq);
3980 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
3983 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
3985 struct tcp_sock *tp = tcp_sk(sk);
3987 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
3988 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3989 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
3990 tcp_enter_quickack_mode(sk);
3992 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3993 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3995 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
3996 end_seq = tp->rcv_nxt;
3997 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4004 /* These routines update the SACK block as out-of-order packets arrive or
4005 * in-order packets close up the sequence space.
4007 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4010 struct tcp_sack_block *sp = &tp->selective_acks[0];
4011 struct tcp_sack_block *swalk = sp + 1;
4013 /* See if the recent change to the first SACK eats into
4014 * or hits the sequence space of other SACK blocks, if so coalesce.
4016 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4017 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4020 /* Zap SWALK, by moving every further SACK up by one slot.
4021 * Decrease num_sacks.
4023 tp->rx_opt.num_sacks--;
4024 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4028 this_sack++, swalk++;
4032 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4034 struct tcp_sock *tp = tcp_sk(sk);
4035 struct tcp_sack_block *sp = &tp->selective_acks[0];
4036 int cur_sacks = tp->rx_opt.num_sacks;
4042 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4043 if (tcp_sack_extend(sp, seq, end_seq)) {
4044 /* Rotate this_sack to the first one. */
4045 for (; this_sack > 0; this_sack--, sp--)
4046 swap(*sp, *(sp - 1));
4048 tcp_sack_maybe_coalesce(tp);
4053 /* Could not find an adjacent existing SACK, build a new one,
4054 * put it at the front, and shift everyone else down. We
4055 * always know there is at least one SACK present already here.
4057 * If the sack array is full, forget about the last one.
4059 if (this_sack >= TCP_NUM_SACKS) {
4061 tp->rx_opt.num_sacks--;
4064 for (; this_sack > 0; this_sack--, sp--)
4068 /* Build the new head SACK, and we're done. */
4069 sp->start_seq = seq;
4070 sp->end_seq = end_seq;
4071 tp->rx_opt.num_sacks++;
4074 /* RCV.NXT advances, some SACKs should be eaten. */
4076 static void tcp_sack_remove(struct tcp_sock *tp)
4078 struct tcp_sack_block *sp = &tp->selective_acks[0];
4079 int num_sacks = tp->rx_opt.num_sacks;
4082 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4083 if (skb_queue_empty(&tp->out_of_order_queue)) {
4084 tp->rx_opt.num_sacks = 0;
4088 for (this_sack = 0; this_sack < num_sacks;) {
4089 /* Check if the start of the sack is covered by RCV.NXT. */
4090 if (!before(tp->rcv_nxt, sp->start_seq)) {
4093 /* RCV.NXT must cover all the block! */
4094 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4096 /* Zap this SACK, by moving forward any other SACKS. */
4097 for (i = this_sack+1; i < num_sacks; i++)
4098 tp->selective_acks[i-1] = tp->selective_acks[i];
4105 tp->rx_opt.num_sacks = num_sacks;
4109 * tcp_try_coalesce - try to merge skb to prior one
4112 * @from: buffer to add in queue
4113 * @fragstolen: pointer to boolean
4115 * Before queueing skb @from after @to, try to merge them
4116 * to reduce overall memory use and queue lengths, if cost is small.
4117 * Packets in ofo or receive queues can stay a long time.
4118 * Better try to coalesce them right now to avoid future collapses.
4119 * Returns true if caller should free @from instead of queueing it
4121 static bool tcp_try_coalesce(struct sock *sk,
4123 struct sk_buff *from,
4128 *fragstolen = false;
4130 /* Its possible this segment overlaps with prior segment in queue */
4131 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4134 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4137 atomic_add(delta, &sk->sk_rmem_alloc);
4138 sk_mem_charge(sk, delta);
4139 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4140 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4141 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4142 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4146 /* This one checks to see if we can put data from the
4147 * out_of_order queue into the receive_queue.
4149 static void tcp_ofo_queue(struct sock *sk)
4151 struct tcp_sock *tp = tcp_sk(sk);
4152 __u32 dsack_high = tp->rcv_nxt;
4153 struct sk_buff *skb, *tail;
4154 bool fragstolen, eaten;
4156 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4157 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4160 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4161 __u32 dsack = dsack_high;
4162 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4163 dsack_high = TCP_SKB_CB(skb)->end_seq;
4164 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4167 __skb_unlink(skb, &tp->out_of_order_queue);
4168 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4169 SOCK_DEBUG(sk, "ofo packet was already received\n");
4173 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4174 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4175 TCP_SKB_CB(skb)->end_seq);
4177 tail = skb_peek_tail(&sk->sk_receive_queue);
4178 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4179 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4181 __skb_queue_tail(&sk->sk_receive_queue, skb);
4182 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4185 kfree_skb_partial(skb, fragstolen);
4189 static bool tcp_prune_ofo_queue(struct sock *sk);
4190 static int tcp_prune_queue(struct sock *sk);
4192 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4195 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4196 !sk_rmem_schedule(sk, skb, size)) {
4198 if (tcp_prune_queue(sk) < 0)
4201 if (!sk_rmem_schedule(sk, skb, size)) {
4202 if (!tcp_prune_ofo_queue(sk))
4205 if (!sk_rmem_schedule(sk, skb, size))
4212 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4214 struct tcp_sock *tp = tcp_sk(sk);
4215 struct sk_buff *skb1;
4218 tcp_ecn_check_ce(tp, skb);
4220 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4221 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFODROP);
4226 /* Disable header prediction. */
4228 inet_csk_schedule_ack(sk);
4230 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4231 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4232 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4234 skb1 = skb_peek_tail(&tp->out_of_order_queue);
4236 /* Initial out of order segment, build 1 SACK. */
4237 if (tcp_is_sack(tp)) {
4238 tp->rx_opt.num_sacks = 1;
4239 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4240 tp->selective_acks[0].end_seq =
4241 TCP_SKB_CB(skb)->end_seq;
4243 __skb_queue_head(&tp->out_of_order_queue, skb);
4247 seq = TCP_SKB_CB(skb)->seq;
4248 end_seq = TCP_SKB_CB(skb)->end_seq;
4250 if (seq == TCP_SKB_CB(skb1)->end_seq) {
4253 if (!tcp_try_coalesce(sk, skb1, skb, &fragstolen)) {
4254 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4256 tcp_grow_window(sk, skb);
4257 kfree_skb_partial(skb, fragstolen);
4261 if (!tp->rx_opt.num_sacks ||
4262 tp->selective_acks[0].end_seq != seq)
4265 /* Common case: data arrive in order after hole. */
4266 tp->selective_acks[0].end_seq = end_seq;
4270 /* Find place to insert this segment. */
4272 if (!after(TCP_SKB_CB(skb1)->seq, seq))
4274 if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
4278 skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
4281 /* Do skb overlap to previous one? */
4282 if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4283 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4284 /* All the bits are present. Drop. */
4285 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4288 tcp_dsack_set(sk, seq, end_seq);
4291 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4292 /* Partial overlap. */
4293 tcp_dsack_set(sk, seq,
4294 TCP_SKB_CB(skb1)->end_seq);
4296 if (skb_queue_is_first(&tp->out_of_order_queue,
4300 skb1 = skb_queue_prev(
4301 &tp->out_of_order_queue,
4306 __skb_queue_head(&tp->out_of_order_queue, skb);
4308 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4310 /* And clean segments covered by new one as whole. */
4311 while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
4312 skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
4314 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4316 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4317 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4321 __skb_unlink(skb1, &tp->out_of_order_queue);
4322 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4323 TCP_SKB_CB(skb1)->end_seq);
4324 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4329 if (tcp_is_sack(tp))
4330 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4333 tcp_grow_window(sk, skb);
4334 skb_set_owner_r(skb, sk);
4338 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
4342 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4344 __skb_pull(skb, hdrlen);
4346 tcp_try_coalesce(sk, tail, skb, fragstolen)) ? 1 : 0;
4347 tcp_sk(sk)->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4349 __skb_queue_tail(&sk->sk_receive_queue, skb);
4350 skb_set_owner_r(skb, sk);
4355 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4357 struct sk_buff *skb;
4363 skb = alloc_skb(size, sk->sk_allocation);
4367 if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4370 if (memcpy_fromiovec(skb_put(skb, size), msg->msg_iov, size))
4373 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4374 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4375 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4377 if (tcp_queue_rcv(sk, skb, 0, &fragstolen)) {
4378 WARN_ON_ONCE(fragstolen); /* should not happen */
4389 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4391 struct tcp_sock *tp = tcp_sk(sk);
4393 bool fragstolen = false;
4395 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4399 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
4401 tcp_ecn_accept_cwr(tp, skb);
4403 tp->rx_opt.dsack = 0;
4405 /* Queue data for delivery to the user.
4406 * Packets in sequence go to the receive queue.
4407 * Out of sequence packets to the out_of_order_queue.
4409 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4410 if (tcp_receive_window(tp) == 0)
4413 /* Ok. In sequence. In window. */
4414 if (tp->ucopy.task == current &&
4415 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4416 sock_owned_by_user(sk) && !tp->urg_data) {
4417 int chunk = min_t(unsigned int, skb->len,
4420 __set_current_state(TASK_RUNNING);
4423 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
4424 tp->ucopy.len -= chunk;
4425 tp->copied_seq += chunk;
4426 eaten = (chunk == skb->len);
4427 tcp_rcv_space_adjust(sk);
4435 tcp_try_rmem_schedule(sk, skb, skb->truesize))
4438 eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
4440 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4442 tcp_event_data_recv(sk, skb);
4443 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4446 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4449 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4450 * gap in queue is filled.
4452 if (skb_queue_empty(&tp->out_of_order_queue))
4453 inet_csk(sk)->icsk_ack.pingpong = 0;
4456 if (tp->rx_opt.num_sacks)
4457 tcp_sack_remove(tp);
4459 tcp_fast_path_check(sk);
4462 kfree_skb_partial(skb, fragstolen);
4463 if (!sock_flag(sk, SOCK_DEAD))
4464 sk->sk_data_ready(sk);
4468 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4469 /* A retransmit, 2nd most common case. Force an immediate ack. */
4470 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4471 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4474 tcp_enter_quickack_mode(sk);
4475 inet_csk_schedule_ack(sk);
4481 /* Out of window. F.e. zero window probe. */
4482 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4485 tcp_enter_quickack_mode(sk);
4487 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4488 /* Partial packet, seq < rcv_next < end_seq */
4489 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4490 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4491 TCP_SKB_CB(skb)->end_seq);
4493 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4495 /* If window is closed, drop tail of packet. But after
4496 * remembering D-SACK for its head made in previous line.
4498 if (!tcp_receive_window(tp))
4503 tcp_data_queue_ofo(sk, skb);
4506 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4507 struct sk_buff_head *list)
4509 struct sk_buff *next = NULL;
4511 if (!skb_queue_is_last(list, skb))
4512 next = skb_queue_next(list, skb);
4514 __skb_unlink(skb, list);
4516 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4521 /* Collapse contiguous sequence of skbs head..tail with
4522 * sequence numbers start..end.
4524 * If tail is NULL, this means until the end of the list.
4526 * Segments with FIN/SYN are not collapsed (only because this
4530 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4531 struct sk_buff *head, struct sk_buff *tail,
4534 struct sk_buff *skb, *n;
4537 /* First, check that queue is collapsible and find
4538 * the point where collapsing can be useful. */
4542 skb_queue_walk_from_safe(list, skb, n) {
4545 /* No new bits? It is possible on ofo queue. */
4546 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4547 skb = tcp_collapse_one(sk, skb, list);
4553 /* The first skb to collapse is:
4555 * - bloated or contains data before "start" or
4556 * overlaps to the next one.
4558 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
4559 (tcp_win_from_space(skb->truesize) > skb->len ||
4560 before(TCP_SKB_CB(skb)->seq, start))) {
4561 end_of_skbs = false;
4565 if (!skb_queue_is_last(list, skb)) {
4566 struct sk_buff *next = skb_queue_next(list, skb);
4568 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
4569 end_of_skbs = false;
4574 /* Decided to skip this, advance start seq. */
4575 start = TCP_SKB_CB(skb)->end_seq;
4578 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4581 while (before(start, end)) {
4582 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
4583 struct sk_buff *nskb;
4585 nskb = alloc_skb(copy, GFP_ATOMIC);
4589 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4590 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4591 __skb_queue_before(list, skb, nskb);
4592 skb_set_owner_r(nskb, sk);
4594 /* Copy data, releasing collapsed skbs. */
4596 int offset = start - TCP_SKB_CB(skb)->seq;
4597 int size = TCP_SKB_CB(skb)->end_seq - start;
4601 size = min(copy, size);
4602 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4604 TCP_SKB_CB(nskb)->end_seq += size;
4608 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4609 skb = tcp_collapse_one(sk, skb, list);
4612 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4619 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4620 * and tcp_collapse() them until all the queue is collapsed.
4622 static void tcp_collapse_ofo_queue(struct sock *sk)
4624 struct tcp_sock *tp = tcp_sk(sk);
4625 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4626 struct sk_buff *head;
4632 start = TCP_SKB_CB(skb)->seq;
4633 end = TCP_SKB_CB(skb)->end_seq;
4637 struct sk_buff *next = NULL;
4639 if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
4640 next = skb_queue_next(&tp->out_of_order_queue, skb);
4643 /* Segment is terminated when we see gap or when
4644 * we are at the end of all the queue. */
4646 after(TCP_SKB_CB(skb)->seq, end) ||
4647 before(TCP_SKB_CB(skb)->end_seq, start)) {
4648 tcp_collapse(sk, &tp->out_of_order_queue,
4649 head, skb, start, end);
4653 /* Start new segment */
4654 start = TCP_SKB_CB(skb)->seq;
4655 end = TCP_SKB_CB(skb)->end_seq;
4657 if (before(TCP_SKB_CB(skb)->seq, start))
4658 start = TCP_SKB_CB(skb)->seq;
4659 if (after(TCP_SKB_CB(skb)->end_seq, end))
4660 end = TCP_SKB_CB(skb)->end_seq;
4666 * Purge the out-of-order queue.
4667 * Return true if queue was pruned.
4669 static bool tcp_prune_ofo_queue(struct sock *sk)
4671 struct tcp_sock *tp = tcp_sk(sk);
4674 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4675 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
4676 __skb_queue_purge(&tp->out_of_order_queue);
4678 /* Reset SACK state. A conforming SACK implementation will
4679 * do the same at a timeout based retransmit. When a connection
4680 * is in a sad state like this, we care only about integrity
4681 * of the connection not performance.
4683 if (tp->rx_opt.sack_ok)
4684 tcp_sack_reset(&tp->rx_opt);
4691 /* Reduce allocated memory if we can, trying to get
4692 * the socket within its memory limits again.
4694 * Return less than zero if we should start dropping frames
4695 * until the socket owning process reads some of the data
4696 * to stabilize the situation.
4698 static int tcp_prune_queue(struct sock *sk)
4700 struct tcp_sock *tp = tcp_sk(sk);
4702 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4704 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
4706 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4707 tcp_clamp_window(sk);
4708 else if (sk_under_memory_pressure(sk))
4709 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4711 tcp_collapse_ofo_queue(sk);
4712 if (!skb_queue_empty(&sk->sk_receive_queue))
4713 tcp_collapse(sk, &sk->sk_receive_queue,
4714 skb_peek(&sk->sk_receive_queue),
4716 tp->copied_seq, tp->rcv_nxt);
4719 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4722 /* Collapsing did not help, destructive actions follow.
4723 * This must not ever occur. */
4725 tcp_prune_ofo_queue(sk);
4727 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4730 /* If we are really being abused, tell the caller to silently
4731 * drop receive data on the floor. It will get retransmitted
4732 * and hopefully then we'll have sufficient space.
4734 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
4736 /* Massive buffer overcommit. */
4741 static bool tcp_should_expand_sndbuf(const struct sock *sk)
4743 const struct tcp_sock *tp = tcp_sk(sk);
4745 /* If the user specified a specific send buffer setting, do
4748 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4751 /* If we are under global TCP memory pressure, do not expand. */
4752 if (sk_under_memory_pressure(sk))
4755 /* If we are under soft global TCP memory pressure, do not expand. */
4756 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
4759 /* If we filled the congestion window, do not expand. */
4760 if (tp->packets_out >= tp->snd_cwnd)
4766 /* When incoming ACK allowed to free some skb from write_queue,
4767 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4768 * on the exit from tcp input handler.
4770 * PROBLEM: sndbuf expansion does not work well with largesend.
4772 static void tcp_new_space(struct sock *sk)
4774 struct tcp_sock *tp = tcp_sk(sk);
4776 if (tcp_should_expand_sndbuf(sk)) {
4777 tcp_sndbuf_expand(sk);
4778 tp->snd_cwnd_stamp = tcp_time_stamp;
4781 sk->sk_write_space(sk);
4784 static void tcp_check_space(struct sock *sk)
4786 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4787 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4788 if (sk->sk_socket &&
4789 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4794 static inline void tcp_data_snd_check(struct sock *sk)
4796 tcp_push_pending_frames(sk);
4797 tcp_check_space(sk);
4801 * Check if sending an ack is needed.
4803 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4805 struct tcp_sock *tp = tcp_sk(sk);
4807 /* More than one full frame received... */
4808 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
4809 /* ... and right edge of window advances far enough.
4810 * (tcp_recvmsg() will send ACK otherwise). Or...
4812 __tcp_select_window(sk) >= tp->rcv_wnd) ||
4813 /* We ACK each frame or... */
4814 tcp_in_quickack_mode(sk) ||
4815 /* We have out of order data. */
4816 (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
4817 /* Then ack it now */
4820 /* Else, send delayed ack. */
4821 tcp_send_delayed_ack(sk);
4825 static inline void tcp_ack_snd_check(struct sock *sk)
4827 if (!inet_csk_ack_scheduled(sk)) {
4828 /* We sent a data segment already. */
4831 __tcp_ack_snd_check(sk, 1);
4835 * This routine is only called when we have urgent data
4836 * signaled. Its the 'slow' part of tcp_urg. It could be
4837 * moved inline now as tcp_urg is only called from one
4838 * place. We handle URGent data wrong. We have to - as
4839 * BSD still doesn't use the correction from RFC961.
4840 * For 1003.1g we should support a new option TCP_STDURG to permit
4841 * either form (or just set the sysctl tcp_stdurg).
4844 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
4846 struct tcp_sock *tp = tcp_sk(sk);
4847 u32 ptr = ntohs(th->urg_ptr);
4849 if (ptr && !sysctl_tcp_stdurg)
4851 ptr += ntohl(th->seq);
4853 /* Ignore urgent data that we've already seen and read. */
4854 if (after(tp->copied_seq, ptr))
4857 /* Do not replay urg ptr.
4859 * NOTE: interesting situation not covered by specs.
4860 * Misbehaving sender may send urg ptr, pointing to segment,
4861 * which we already have in ofo queue. We are not able to fetch
4862 * such data and will stay in TCP_URG_NOTYET until will be eaten
4863 * by recvmsg(). Seems, we are not obliged to handle such wicked
4864 * situations. But it is worth to think about possibility of some
4865 * DoSes using some hypothetical application level deadlock.
4867 if (before(ptr, tp->rcv_nxt))
4870 /* Do we already have a newer (or duplicate) urgent pointer? */
4871 if (tp->urg_data && !after(ptr, tp->urg_seq))
4874 /* Tell the world about our new urgent pointer. */
4877 /* We may be adding urgent data when the last byte read was
4878 * urgent. To do this requires some care. We cannot just ignore
4879 * tp->copied_seq since we would read the last urgent byte again
4880 * as data, nor can we alter copied_seq until this data arrives
4881 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4883 * NOTE. Double Dutch. Rendering to plain English: author of comment
4884 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4885 * and expect that both A and B disappear from stream. This is _wrong_.
4886 * Though this happens in BSD with high probability, this is occasional.
4887 * Any application relying on this is buggy. Note also, that fix "works"
4888 * only in this artificial test. Insert some normal data between A and B and we will
4889 * decline of BSD again. Verdict: it is better to remove to trap
4892 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
4893 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
4894 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4896 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
4897 __skb_unlink(skb, &sk->sk_receive_queue);
4902 tp->urg_data = TCP_URG_NOTYET;
4905 /* Disable header prediction. */
4909 /* This is the 'fast' part of urgent handling. */
4910 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
4912 struct tcp_sock *tp = tcp_sk(sk);
4914 /* Check if we get a new urgent pointer - normally not. */
4916 tcp_check_urg(sk, th);
4918 /* Do we wait for any urgent data? - normally not... */
4919 if (tp->urg_data == TCP_URG_NOTYET) {
4920 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
4923 /* Is the urgent pointer pointing into this packet? */
4924 if (ptr < skb->len) {
4926 if (skb_copy_bits(skb, ptr, &tmp, 1))
4928 tp->urg_data = TCP_URG_VALID | tmp;
4929 if (!sock_flag(sk, SOCK_DEAD))
4930 sk->sk_data_ready(sk);
4935 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4937 struct tcp_sock *tp = tcp_sk(sk);
4938 int chunk = skb->len - hlen;
4942 if (skb_csum_unnecessary(skb))
4943 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
4945 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
4949 tp->ucopy.len -= chunk;
4950 tp->copied_seq += chunk;
4951 tcp_rcv_space_adjust(sk);
4958 static __sum16 __tcp_checksum_complete_user(struct sock *sk,
4959 struct sk_buff *skb)
4963 if (sock_owned_by_user(sk)) {
4965 result = __tcp_checksum_complete(skb);
4968 result = __tcp_checksum_complete(skb);
4973 static inline bool tcp_checksum_complete_user(struct sock *sk,
4974 struct sk_buff *skb)
4976 return !skb_csum_unnecessary(skb) &&
4977 __tcp_checksum_complete_user(sk, skb);
4980 /* Does PAWS and seqno based validation of an incoming segment, flags will
4981 * play significant role here.
4983 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
4984 const struct tcphdr *th, int syn_inerr)
4986 struct tcp_sock *tp = tcp_sk(sk);
4988 /* RFC1323: H1. Apply PAWS check first. */
4989 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4990 tcp_paws_discard(sk, skb)) {
4992 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
4993 tcp_send_dupack(sk, skb);
4996 /* Reset is accepted even if it did not pass PAWS. */
4999 /* Step 1: check sequence number */
5000 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5001 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5002 * (RST) segments are validated by checking their SEQ-fields."
5003 * And page 69: "If an incoming segment is not acceptable,
5004 * an acknowledgment should be sent in reply (unless the RST
5005 * bit is set, if so drop the segment and return)".
5010 tcp_send_dupack(sk, skb);
5015 /* Step 2: check RST bit */
5018 * If sequence number exactly matches RCV.NXT, then
5019 * RESET the connection
5021 * Send a challenge ACK
5023 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt)
5026 tcp_send_challenge_ack(sk);
5030 /* step 3: check security and precedence [ignored] */
5032 /* step 4: Check for a SYN
5033 * RFC 5691 4.2 : Send a challenge ack
5038 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5039 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5040 tcp_send_challenge_ack(sk);
5052 * TCP receive function for the ESTABLISHED state.
5054 * It is split into a fast path and a slow path. The fast path is
5056 * - A zero window was announced from us - zero window probing
5057 * is only handled properly in the slow path.
5058 * - Out of order segments arrived.
5059 * - Urgent data is expected.
5060 * - There is no buffer space left
5061 * - Unexpected TCP flags/window values/header lengths are received
5062 * (detected by checking the TCP header against pred_flags)
5063 * - Data is sent in both directions. Fast path only supports pure senders
5064 * or pure receivers (this means either the sequence number or the ack
5065 * value must stay constant)
5066 * - Unexpected TCP option.
5068 * When these conditions are not satisfied it drops into a standard
5069 * receive procedure patterned after RFC793 to handle all cases.
5070 * The first three cases are guaranteed by proper pred_flags setting,
5071 * the rest is checked inline. Fast processing is turned on in
5072 * tcp_data_queue when everything is OK.
5074 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5075 const struct tcphdr *th, unsigned int len)
5077 struct tcp_sock *tp = tcp_sk(sk);
5079 if (unlikely(sk->sk_rx_dst == NULL))
5080 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5082 * Header prediction.
5083 * The code loosely follows the one in the famous
5084 * "30 instruction TCP receive" Van Jacobson mail.
5086 * Van's trick is to deposit buffers into socket queue
5087 * on a device interrupt, to call tcp_recv function
5088 * on the receive process context and checksum and copy
5089 * the buffer to user space. smart...
5091 * Our current scheme is not silly either but we take the
5092 * extra cost of the net_bh soft interrupt processing...
5093 * We do checksum and copy also but from device to kernel.
5096 tp->rx_opt.saw_tstamp = 0;
5098 /* pred_flags is 0xS?10 << 16 + snd_wnd
5099 * if header_prediction is to be made
5100 * 'S' will always be tp->tcp_header_len >> 2
5101 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5102 * turn it off (when there are holes in the receive
5103 * space for instance)
5104 * PSH flag is ignored.
5107 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5108 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5109 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5110 int tcp_header_len = tp->tcp_header_len;
5112 /* Timestamp header prediction: tcp_header_len
5113 * is automatically equal to th->doff*4 due to pred_flags
5117 /* Check timestamp */
5118 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5119 /* No? Slow path! */
5120 if (!tcp_parse_aligned_timestamp(tp, th))
5123 /* If PAWS failed, check it more carefully in slow path */
5124 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5127 /* DO NOT update ts_recent here, if checksum fails
5128 * and timestamp was corrupted part, it will result
5129 * in a hung connection since we will drop all
5130 * future packets due to the PAWS test.
5134 if (len <= tcp_header_len) {
5135 /* Bulk data transfer: sender */
5136 if (len == tcp_header_len) {
5137 /* Predicted packet is in window by definition.
5138 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5139 * Hence, check seq<=rcv_wup reduces to:
5141 if (tcp_header_len ==
5142 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5143 tp->rcv_nxt == tp->rcv_wup)
5144 tcp_store_ts_recent(tp);
5146 /* We know that such packets are checksummed
5149 tcp_ack(sk, skb, 0);
5151 tcp_data_snd_check(sk);
5153 } else { /* Header too small */
5154 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5159 bool fragstolen = false;
5161 if (tp->ucopy.task == current &&
5162 tp->copied_seq == tp->rcv_nxt &&
5163 len - tcp_header_len <= tp->ucopy.len &&
5164 sock_owned_by_user(sk)) {
5165 __set_current_state(TASK_RUNNING);
5167 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) {
5168 /* Predicted packet is in window by definition.
5169 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5170 * Hence, check seq<=rcv_wup reduces to:
5172 if (tcp_header_len ==
5173 (sizeof(struct tcphdr) +
5174 TCPOLEN_TSTAMP_ALIGNED) &&
5175 tp->rcv_nxt == tp->rcv_wup)
5176 tcp_store_ts_recent(tp);
5178 tcp_rcv_rtt_measure_ts(sk, skb);
5180 __skb_pull(skb, tcp_header_len);
5181 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5182 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
5187 if (tcp_checksum_complete_user(sk, skb))
5190 if ((int)skb->truesize > sk->sk_forward_alloc)
5193 /* Predicted packet is in window by definition.
5194 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5195 * Hence, check seq<=rcv_wup reduces to:
5197 if (tcp_header_len ==
5198 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5199 tp->rcv_nxt == tp->rcv_wup)
5200 tcp_store_ts_recent(tp);
5202 tcp_rcv_rtt_measure_ts(sk, skb);
5204 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
5206 /* Bulk data transfer: receiver */
5207 eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5211 tcp_event_data_recv(sk, skb);
5213 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5214 /* Well, only one small jumplet in fast path... */
5215 tcp_ack(sk, skb, FLAG_DATA);
5216 tcp_data_snd_check(sk);
5217 if (!inet_csk_ack_scheduled(sk))
5221 __tcp_ack_snd_check(sk, 0);
5224 kfree_skb_partial(skb, fragstolen);
5225 sk->sk_data_ready(sk);
5231 if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
5234 if (!th->ack && !th->rst)
5238 * Standard slow path.
5241 if (!tcp_validate_incoming(sk, skb, th, 1))
5245 if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5248 tcp_rcv_rtt_measure_ts(sk, skb);
5250 /* Process urgent data. */
5251 tcp_urg(sk, skb, th);
5253 /* step 7: process the segment text */
5254 tcp_data_queue(sk, skb);
5256 tcp_data_snd_check(sk);
5257 tcp_ack_snd_check(sk);
5261 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_CSUMERRORS);
5262 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5267 EXPORT_SYMBOL(tcp_rcv_established);
5269 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5271 struct tcp_sock *tp = tcp_sk(sk);
5272 struct inet_connection_sock *icsk = inet_csk(sk);
5274 tcp_set_state(sk, TCP_ESTABLISHED);
5277 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5278 security_inet_conn_established(sk, skb);
5281 /* Make sure socket is routed, for correct metrics. */
5282 icsk->icsk_af_ops->rebuild_header(sk);
5284 tcp_init_metrics(sk);
5286 tcp_init_congestion_control(sk);
5288 /* Prevent spurious tcp_cwnd_restart() on first data
5291 tp->lsndtime = tcp_time_stamp;
5293 tcp_init_buffer_space(sk);
5295 if (sock_flag(sk, SOCK_KEEPOPEN))
5296 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5298 if (!tp->rx_opt.snd_wscale)
5299 __tcp_fast_path_on(tp, tp->snd_wnd);
5303 if (!sock_flag(sk, SOCK_DEAD)) {
5304 sk->sk_state_change(sk);
5305 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5309 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5310 struct tcp_fastopen_cookie *cookie)
5312 struct tcp_sock *tp = tcp_sk(sk);
5313 struct sk_buff *data = tp->syn_data ? tcp_write_queue_head(sk) : NULL;
5314 u16 mss = tp->rx_opt.mss_clamp;
5317 if (mss == tp->rx_opt.user_mss) {
5318 struct tcp_options_received opt;
5320 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5321 tcp_clear_options(&opt);
5322 opt.user_mss = opt.mss_clamp = 0;
5323 tcp_parse_options(synack, &opt, 0, NULL);
5324 mss = opt.mss_clamp;
5327 if (!tp->syn_fastopen) /* Ignore an unsolicited cookie */
5330 /* The SYN-ACK neither has cookie nor acknowledges the data. Presumably
5331 * the remote receives only the retransmitted (regular) SYNs: either
5332 * the original SYN-data or the corresponding SYN-ACK is lost.
5334 syn_drop = (cookie->len <= 0 && data && tp->total_retrans);
5336 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop);
5338 if (data) { /* Retransmit unacked data in SYN */
5339 tcp_for_write_queue_from(data, sk) {
5340 if (data == tcp_send_head(sk) ||
5341 __tcp_retransmit_skb(sk, data))
5345 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVEFAIL);
5348 tp->syn_data_acked = tp->syn_data;
5349 if (tp->syn_data_acked)
5350 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
5354 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5355 const struct tcphdr *th, unsigned int len)
5357 struct inet_connection_sock *icsk = inet_csk(sk);
5358 struct tcp_sock *tp = tcp_sk(sk);
5359 struct tcp_fastopen_cookie foc = { .len = -1 };
5360 int saved_clamp = tp->rx_opt.mss_clamp;
5362 tcp_parse_options(skb, &tp->rx_opt, 0, &foc);
5363 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
5364 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
5368 * "If the state is SYN-SENT then
5369 * first check the ACK bit
5370 * If the ACK bit is set
5371 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5372 * a reset (unless the RST bit is set, if so drop
5373 * the segment and return)"
5375 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5376 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
5377 goto reset_and_undo;
5379 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5380 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5382 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
5383 goto reset_and_undo;
5386 /* Now ACK is acceptable.
5388 * "If the RST bit is set
5389 * If the ACK was acceptable then signal the user "error:
5390 * connection reset", drop the segment, enter CLOSED state,
5391 * delete TCB, and return."
5400 * "fifth, if neither of the SYN or RST bits is set then
5401 * drop the segment and return."
5407 goto discard_and_undo;
5410 * "If the SYN bit is on ...
5411 * are acceptable then ...
5412 * (our SYN has been ACKed), change the connection
5413 * state to ESTABLISHED..."
5416 tcp_ecn_rcv_synack(tp, th);
5418 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5419 tcp_ack(sk, skb, FLAG_SLOWPATH);
5421 /* Ok.. it's good. Set up sequence numbers and
5422 * move to established.
5424 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5425 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5427 /* RFC1323: The window in SYN & SYN/ACK segments is
5430 tp->snd_wnd = ntohs(th->window);
5432 if (!tp->rx_opt.wscale_ok) {
5433 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5434 tp->window_clamp = min(tp->window_clamp, 65535U);
5437 if (tp->rx_opt.saw_tstamp) {
5438 tp->rx_opt.tstamp_ok = 1;
5439 tp->tcp_header_len =
5440 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5441 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5442 tcp_store_ts_recent(tp);
5444 tp->tcp_header_len = sizeof(struct tcphdr);
5447 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5448 tcp_enable_fack(tp);
5451 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5452 tcp_initialize_rcv_mss(sk);
5454 /* Remember, tcp_poll() does not lock socket!
5455 * Change state from SYN-SENT only after copied_seq
5456 * is initialized. */
5457 tp->copied_seq = tp->rcv_nxt;
5461 tcp_finish_connect(sk, skb);
5463 if ((tp->syn_fastopen || tp->syn_data) &&
5464 tcp_rcv_fastopen_synack(sk, skb, &foc))
5467 if (sk->sk_write_pending ||
5468 icsk->icsk_accept_queue.rskq_defer_accept ||
5469 icsk->icsk_ack.pingpong) {
5470 /* Save one ACK. Data will be ready after
5471 * several ticks, if write_pending is set.
5473 * It may be deleted, but with this feature tcpdumps
5474 * look so _wonderfully_ clever, that I was not able
5475 * to stand against the temptation 8) --ANK
5477 inet_csk_schedule_ack(sk);
5478 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5479 tcp_enter_quickack_mode(sk);
5480 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5481 TCP_DELACK_MAX, TCP_RTO_MAX);
5492 /* No ACK in the segment */
5496 * "If the RST bit is set
5498 * Otherwise (no ACK) drop the segment and return."
5501 goto discard_and_undo;
5505 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5506 tcp_paws_reject(&tp->rx_opt, 0))
5507 goto discard_and_undo;
5510 /* We see SYN without ACK. It is attempt of
5511 * simultaneous connect with crossed SYNs.
5512 * Particularly, it can be connect to self.
5514 tcp_set_state(sk, TCP_SYN_RECV);
5516 if (tp->rx_opt.saw_tstamp) {
5517 tp->rx_opt.tstamp_ok = 1;
5518 tcp_store_ts_recent(tp);
5519 tp->tcp_header_len =
5520 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5522 tp->tcp_header_len = sizeof(struct tcphdr);
5525 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5526 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5528 /* RFC1323: The window in SYN & SYN/ACK segments is
5531 tp->snd_wnd = ntohs(th->window);
5532 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5533 tp->max_window = tp->snd_wnd;
5535 tcp_ecn_rcv_syn(tp, th);
5538 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5539 tcp_initialize_rcv_mss(sk);
5541 tcp_send_synack(sk);
5543 /* Note, we could accept data and URG from this segment.
5544 * There are no obstacles to make this (except that we must
5545 * either change tcp_recvmsg() to prevent it from returning data
5546 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5548 * However, if we ignore data in ACKless segments sometimes,
5549 * we have no reasons to accept it sometimes.
5550 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5551 * is not flawless. So, discard packet for sanity.
5552 * Uncomment this return to process the data.
5559 /* "fifth, if neither of the SYN or RST bits is set then
5560 * drop the segment and return."
5564 tcp_clear_options(&tp->rx_opt);
5565 tp->rx_opt.mss_clamp = saved_clamp;
5569 tcp_clear_options(&tp->rx_opt);
5570 tp->rx_opt.mss_clamp = saved_clamp;
5575 * This function implements the receiving procedure of RFC 793 for
5576 * all states except ESTABLISHED and TIME_WAIT.
5577 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5578 * address independent.
5581 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5582 const struct tcphdr *th, unsigned int len)
5584 struct tcp_sock *tp = tcp_sk(sk);
5585 struct inet_connection_sock *icsk = inet_csk(sk);
5586 struct request_sock *req;
5591 tp->rx_opt.saw_tstamp = 0;
5593 switch (sk->sk_state) {
5607 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5610 /* Now we have several options: In theory there is
5611 * nothing else in the frame. KA9Q has an option to
5612 * send data with the syn, BSD accepts data with the
5613 * syn up to the [to be] advertised window and
5614 * Solaris 2.1 gives you a protocol error. For now
5615 * we just ignore it, that fits the spec precisely
5616 * and avoids incompatibilities. It would be nice in
5617 * future to drop through and process the data.
5619 * Now that TTCP is starting to be used we ought to
5621 * But, this leaves one open to an easy denial of
5622 * service attack, and SYN cookies can't defend
5623 * against this problem. So, we drop the data
5624 * in the interest of security over speed unless
5625 * it's still in use.
5633 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5637 /* Do step6 onward by hand. */
5638 tcp_urg(sk, skb, th);
5640 tcp_data_snd_check(sk);
5644 req = tp->fastopen_rsk;
5646 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
5647 sk->sk_state != TCP_FIN_WAIT1);
5649 if (tcp_check_req(sk, skb, req, NULL, true) == NULL)
5653 if (!th->ack && !th->rst)
5656 if (!tcp_validate_incoming(sk, skb, th, 0))
5659 /* step 5: check the ACK field */
5660 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
5661 FLAG_UPDATE_TS_RECENT) > 0;
5663 switch (sk->sk_state) {
5668 /* Once we leave TCP_SYN_RECV, we no longer need req
5672 synack_stamp = tcp_rsk(req)->snt_synack;
5673 tp->total_retrans = req->num_retrans;
5674 reqsk_fastopen_remove(sk, req, false);
5676 synack_stamp = tp->lsndtime;
5677 /* Make sure socket is routed, for correct metrics. */
5678 icsk->icsk_af_ops->rebuild_header(sk);
5679 tcp_init_congestion_control(sk);
5682 tp->copied_seq = tp->rcv_nxt;
5683 tcp_init_buffer_space(sk);
5686 tcp_set_state(sk, TCP_ESTABLISHED);
5687 sk->sk_state_change(sk);
5689 /* Note, that this wakeup is only for marginal crossed SYN case.
5690 * Passively open sockets are not waked up, because
5691 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5694 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5696 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5697 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
5698 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5699 tcp_synack_rtt_meas(sk, synack_stamp);
5701 if (tp->rx_opt.tstamp_ok)
5702 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5705 /* Re-arm the timer because data may have been sent out.
5706 * This is similar to the regular data transmission case
5707 * when new data has just been ack'ed.
5709 * (TFO) - we could try to be more aggressive and
5710 * retransmitting any data sooner based on when they
5715 tcp_init_metrics(sk);
5717 tcp_update_pacing_rate(sk);
5719 /* Prevent spurious tcp_cwnd_restart() on first data packet */
5720 tp->lsndtime = tcp_time_stamp;
5722 tcp_initialize_rcv_mss(sk);
5723 tcp_fast_path_on(tp);
5726 case TCP_FIN_WAIT1: {
5727 struct dst_entry *dst;
5730 /* If we enter the TCP_FIN_WAIT1 state and we are a
5731 * Fast Open socket and this is the first acceptable
5732 * ACK we have received, this would have acknowledged
5733 * our SYNACK so stop the SYNACK timer.
5736 /* Return RST if ack_seq is invalid.
5737 * Note that RFC793 only says to generate a
5738 * DUPACK for it but for TCP Fast Open it seems
5739 * better to treat this case like TCP_SYN_RECV
5744 /* We no longer need the request sock. */
5745 reqsk_fastopen_remove(sk, req, false);
5748 if (tp->snd_una != tp->write_seq)
5751 tcp_set_state(sk, TCP_FIN_WAIT2);
5752 sk->sk_shutdown |= SEND_SHUTDOWN;
5754 dst = __sk_dst_get(sk);
5758 if (!sock_flag(sk, SOCK_DEAD)) {
5759 /* Wake up lingering close() */
5760 sk->sk_state_change(sk);
5764 if (tp->linger2 < 0 ||
5765 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5766 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5768 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5772 tmo = tcp_fin_time(sk);
5773 if (tmo > TCP_TIMEWAIT_LEN) {
5774 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5775 } else if (th->fin || sock_owned_by_user(sk)) {
5776 /* Bad case. We could lose such FIN otherwise.
5777 * It is not a big problem, but it looks confusing
5778 * and not so rare event. We still can lose it now,
5779 * if it spins in bh_lock_sock(), but it is really
5782 inet_csk_reset_keepalive_timer(sk, tmo);
5784 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5791 if (tp->snd_una == tp->write_seq) {
5792 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5798 if (tp->snd_una == tp->write_seq) {
5799 tcp_update_metrics(sk);
5806 /* step 6: check the URG bit */
5807 tcp_urg(sk, skb, th);
5809 /* step 7: process the segment text */
5810 switch (sk->sk_state) {
5811 case TCP_CLOSE_WAIT:
5814 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5818 /* RFC 793 says to queue data in these states,
5819 * RFC 1122 says we MUST send a reset.
5820 * BSD 4.4 also does reset.
5822 if (sk->sk_shutdown & RCV_SHUTDOWN) {
5823 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5824 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5825 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5831 case TCP_ESTABLISHED:
5832 tcp_data_queue(sk, skb);
5837 /* tcp_data could move socket to TIME-WAIT */
5838 if (sk->sk_state != TCP_CLOSE) {
5839 tcp_data_snd_check(sk);
5840 tcp_ack_snd_check(sk);
5849 EXPORT_SYMBOL(tcp_rcv_state_process);
5851 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
5853 struct inet_request_sock *ireq = inet_rsk(req);
5855 if (family == AF_INET)
5856 LIMIT_NETDEBUG(KERN_DEBUG pr_fmt("drop open request from %pI4/%u\n"),
5857 &ireq->ir_rmt_addr, port);
5858 #if IS_ENABLED(CONFIG_IPV6)
5859 else if (family == AF_INET6)
5860 LIMIT_NETDEBUG(KERN_DEBUG pr_fmt("drop open request from %pI6/%u\n"),
5861 &ireq->ir_v6_rmt_addr, port);
5865 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
5867 * If we receive a SYN packet with these bits set, it means a
5868 * network is playing bad games with TOS bits. In order to
5869 * avoid possible false congestion notifications, we disable
5870 * TCP ECN negociation.
5872 * Exception: tcp_ca wants ECN. This is required for DCTCP
5873 * congestion control; it requires setting ECT on all packets,
5874 * including SYN. We inverse the test in this case: If our
5875 * local socket wants ECN, but peer only set ece/cwr (but not
5876 * ECT in IP header) its probably a non-DCTCP aware sender.
5878 static void tcp_ecn_create_request(struct request_sock *req,
5879 const struct sk_buff *skb,
5880 const struct sock *listen_sk)
5882 const struct tcphdr *th = tcp_hdr(skb);
5883 const struct net *net = sock_net(listen_sk);
5884 bool th_ecn = th->ece && th->cwr;
5890 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
5891 need_ecn = tcp_ca_needs_ecn(listen_sk);
5893 if (!ect && !need_ecn && net->ipv4.sysctl_tcp_ecn)
5894 inet_rsk(req)->ecn_ok = 1;
5895 else if (ect && need_ecn)
5896 inet_rsk(req)->ecn_ok = 1;
5899 int tcp_conn_request(struct request_sock_ops *rsk_ops,
5900 const struct tcp_request_sock_ops *af_ops,
5901 struct sock *sk, struct sk_buff *skb)
5903 struct tcp_options_received tmp_opt;
5904 struct request_sock *req;
5905 struct tcp_sock *tp = tcp_sk(sk);
5906 struct dst_entry *dst = NULL;
5907 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
5908 bool want_cookie = false, fastopen;
5910 struct tcp_fastopen_cookie foc = { .len = -1 };
5914 /* TW buckets are converted to open requests without
5915 * limitations, they conserve resources and peer is
5916 * evidently real one.
5918 if ((sysctl_tcp_syncookies == 2 ||
5919 inet_csk_reqsk_queue_is_full(sk)) && !isn) {
5920 want_cookie = tcp_syn_flood_action(sk, skb, rsk_ops->slab_name);
5926 /* Accept backlog is full. If we have already queued enough
5927 * of warm entries in syn queue, drop request. It is better than
5928 * clogging syn queue with openreqs with exponentially increasing
5931 if (sk_acceptq_is_full(sk) && inet_csk_reqsk_queue_young(sk) > 1) {
5932 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
5936 req = inet_reqsk_alloc(rsk_ops);
5940 tcp_rsk(req)->af_specific = af_ops;
5942 tcp_clear_options(&tmp_opt);
5943 tmp_opt.mss_clamp = af_ops->mss_clamp;
5944 tmp_opt.user_mss = tp->rx_opt.user_mss;
5945 tcp_parse_options(skb, &tmp_opt, 0, want_cookie ? NULL : &foc);
5947 if (want_cookie && !tmp_opt.saw_tstamp)
5948 tcp_clear_options(&tmp_opt);
5950 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
5951 tcp_openreq_init(req, &tmp_opt, skb, sk);
5953 af_ops->init_req(req, sk, skb);
5955 if (security_inet_conn_request(sk, skb, req))
5958 if (!want_cookie || tmp_opt.tstamp_ok)
5959 tcp_ecn_create_request(req, skb, sk);
5962 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
5963 req->cookie_ts = tmp_opt.tstamp_ok;
5965 /* VJ's idea. We save last timestamp seen
5966 * from the destination in peer table, when entering
5967 * state TIME-WAIT, and check against it before
5968 * accepting new connection request.
5970 * If "isn" is not zero, this request hit alive
5971 * timewait bucket, so that all the necessary checks
5972 * are made in the function processing timewait state.
5974 if (tcp_death_row.sysctl_tw_recycle) {
5977 dst = af_ops->route_req(sk, &fl, req, &strict);
5979 if (dst && strict &&
5980 !tcp_peer_is_proven(req, dst, true,
5981 tmp_opt.saw_tstamp)) {
5982 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSPASSIVEREJECTED);
5983 goto drop_and_release;
5986 /* Kill the following clause, if you dislike this way. */
5987 else if (!sysctl_tcp_syncookies &&
5988 (sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
5989 (sysctl_max_syn_backlog >> 2)) &&
5990 !tcp_peer_is_proven(req, dst, false,
5991 tmp_opt.saw_tstamp)) {
5992 /* Without syncookies last quarter of
5993 * backlog is filled with destinations,
5994 * proven to be alive.
5995 * It means that we continue to communicate
5996 * to destinations, already remembered
5997 * to the moment of synflood.
5999 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
6001 goto drop_and_release;
6004 isn = af_ops->init_seq(skb);
6007 dst = af_ops->route_req(sk, &fl, req, NULL);
6012 tcp_rsk(req)->snt_isn = isn;
6013 tcp_openreq_init_rwin(req, sk, dst);
6014 fastopen = !want_cookie &&
6015 tcp_try_fastopen(sk, skb, req, &foc, dst);
6016 err = af_ops->send_synack(sk, dst, &fl, req,
6017 skb_get_queue_mapping(skb), &foc);
6019 if (err || want_cookie)
6022 tcp_rsk(req)->listener = NULL;
6023 af_ops->queue_hash_add(sk, req, TCP_TIMEOUT_INIT);
6033 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_LISTENDROPS);
6036 EXPORT_SYMBOL(tcp_conn_request);