2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
23 * Pedro Roque : Fast Retransmit/Recovery.
25 * Retransmit queue handled by TCP.
26 * Better retransmit timer handling.
27 * New congestion avoidance.
31 * Eric : Fast Retransmit.
32 * Randy Scott : MSS option defines.
33 * Eric Schenk : Fixes to slow start algorithm.
34 * Eric Schenk : Yet another double ACK bug.
35 * Eric Schenk : Delayed ACK bug fixes.
36 * Eric Schenk : Floyd style fast retrans war avoidance.
37 * David S. Miller : Don't allow zero congestion window.
38 * Eric Schenk : Fix retransmitter so that it sends
39 * next packet on ack of previous packet.
40 * Andi Kleen : Moved open_request checking here
41 * and process RSTs for open_requests.
42 * Andi Kleen : Better prune_queue, and other fixes.
43 * Andrey Savochkin: Fix RTT measurements in the presence of
45 * Andrey Savochkin: Check sequence numbers correctly when
46 * removing SACKs due to in sequence incoming
48 * Andi Kleen: Make sure we never ack data there is not
49 * enough room for. Also make this condition
50 * a fatal error if it might still happen.
51 * Andi Kleen: Add tcp_measure_rcv_mss to make
52 * connections with MSS<min(MTU,ann. MSS)
53 * work without delayed acks.
54 * Andi Kleen: Process packets with PSH set in the
56 * J Hadi Salim: ECN support
59 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
60 * engine. Lots of bugs are found.
61 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
64 #define pr_fmt(fmt) "TCP: " fmt
67 #include <linux/slab.h>
68 #include <linux/module.h>
69 #include <linux/sysctl.h>
70 #include <linux/kernel.h>
71 #include <linux/prefetch.h>
74 #include <net/inet_common.h>
75 #include <linux/ipsec.h>
76 #include <asm/unaligned.h>
77 #include <linux/errqueue.h>
79 int sysctl_tcp_timestamps __read_mostly = 1;
80 int sysctl_tcp_window_scaling __read_mostly = 1;
81 int sysctl_tcp_sack __read_mostly = 1;
82 int sysctl_tcp_fack __read_mostly = 1;
83 int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
84 int sysctl_tcp_max_reordering __read_mostly = 300;
85 EXPORT_SYMBOL(sysctl_tcp_reordering);
86 int sysctl_tcp_dsack __read_mostly = 1;
87 int sysctl_tcp_app_win __read_mostly = 31;
88 int sysctl_tcp_adv_win_scale __read_mostly = 1;
89 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
91 /* rfc5961 challenge ack rate limiting */
92 int sysctl_tcp_challenge_ack_limit = 100;
94 int sysctl_tcp_stdurg __read_mostly;
95 int sysctl_tcp_rfc1337 __read_mostly;
96 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
97 int sysctl_tcp_frto __read_mostly = 2;
99 int sysctl_tcp_thin_dupack __read_mostly;
101 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
102 int sysctl_tcp_early_retrans __read_mostly = 3;
104 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
105 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
106 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
107 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
108 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
109 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
110 #define FLAG_ECE 0x40 /* ECE in this ACK */
111 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
112 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
113 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
114 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
115 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
116 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
118 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
119 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
120 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
121 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
123 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
124 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
126 /* Adapt the MSS value used to make delayed ack decision to the
129 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
131 struct inet_connection_sock *icsk = inet_csk(sk);
132 const unsigned int lss = icsk->icsk_ack.last_seg_size;
135 icsk->icsk_ack.last_seg_size = 0;
137 /* skb->len may jitter because of SACKs, even if peer
138 * sends good full-sized frames.
140 len = skb_shinfo(skb)->gso_size ? : skb->len;
141 if (len >= icsk->icsk_ack.rcv_mss) {
142 icsk->icsk_ack.rcv_mss = len;
144 /* Otherwise, we make more careful check taking into account,
145 * that SACKs block is variable.
147 * "len" is invariant segment length, including TCP header.
149 len += skb->data - skb_transport_header(skb);
150 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
151 /* If PSH is not set, packet should be
152 * full sized, provided peer TCP is not badly broken.
153 * This observation (if it is correct 8)) allows
154 * to handle super-low mtu links fairly.
156 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
157 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
158 /* Subtract also invariant (if peer is RFC compliant),
159 * tcp header plus fixed timestamp option length.
160 * Resulting "len" is MSS free of SACK jitter.
162 len -= tcp_sk(sk)->tcp_header_len;
163 icsk->icsk_ack.last_seg_size = len;
165 icsk->icsk_ack.rcv_mss = len;
169 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
170 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
171 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
175 static void tcp_incr_quickack(struct sock *sk)
177 struct inet_connection_sock *icsk = inet_csk(sk);
178 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
182 if (quickacks > icsk->icsk_ack.quick)
183 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
186 static void tcp_enter_quickack_mode(struct sock *sk)
188 struct inet_connection_sock *icsk = inet_csk(sk);
189 tcp_incr_quickack(sk);
190 icsk->icsk_ack.pingpong = 0;
191 icsk->icsk_ack.ato = TCP_ATO_MIN;
194 /* Send ACKs quickly, if "quick" count is not exhausted
195 * and the session is not interactive.
198 static inline bool tcp_in_quickack_mode(const struct sock *sk)
200 const struct inet_connection_sock *icsk = inet_csk(sk);
202 return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
205 static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
207 if (tp->ecn_flags & TCP_ECN_OK)
208 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
211 static void tcp_ecn_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
213 if (tcp_hdr(skb)->cwr)
214 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
217 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
219 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
222 static void __tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
224 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
225 case INET_ECN_NOT_ECT:
226 /* Funny extension: if ECT is not set on a segment,
227 * and we already seen ECT on a previous segment,
228 * it is probably a retransmit.
230 if (tp->ecn_flags & TCP_ECN_SEEN)
231 tcp_enter_quickack_mode((struct sock *)tp);
234 if (tcp_ca_needs_ecn((struct sock *)tp))
235 tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_IS_CE);
237 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
238 /* Better not delay acks, sender can have a very low cwnd */
239 tcp_enter_quickack_mode((struct sock *)tp);
240 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
242 tp->ecn_flags |= TCP_ECN_SEEN;
245 if (tcp_ca_needs_ecn((struct sock *)tp))
246 tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_NO_CE);
247 tp->ecn_flags |= TCP_ECN_SEEN;
252 static void tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
254 if (tp->ecn_flags & TCP_ECN_OK)
255 __tcp_ecn_check_ce(tp, skb);
258 static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
260 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
261 tp->ecn_flags &= ~TCP_ECN_OK;
264 static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
266 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
267 tp->ecn_flags &= ~TCP_ECN_OK;
270 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
272 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
277 /* Buffer size and advertised window tuning.
279 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
282 static void tcp_sndbuf_expand(struct sock *sk)
284 const struct tcp_sock *tp = tcp_sk(sk);
288 /* Worst case is non GSO/TSO : each frame consumes one skb
289 * and skb->head is kmalloced using power of two area of memory
291 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
293 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
295 per_mss = roundup_pow_of_two(per_mss) +
296 SKB_DATA_ALIGN(sizeof(struct sk_buff));
298 nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd);
299 nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
301 /* Fast Recovery (RFC 5681 3.2) :
302 * Cubic needs 1.7 factor, rounded to 2 to include
303 * extra cushion (application might react slowly to POLLOUT)
305 sndmem = 2 * nr_segs * per_mss;
307 if (sk->sk_sndbuf < sndmem)
308 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
311 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
313 * All tcp_full_space() is split to two parts: "network" buffer, allocated
314 * forward and advertised in receiver window (tp->rcv_wnd) and
315 * "application buffer", required to isolate scheduling/application
316 * latencies from network.
317 * window_clamp is maximal advertised window. It can be less than
318 * tcp_full_space(), in this case tcp_full_space() - window_clamp
319 * is reserved for "application" buffer. The less window_clamp is
320 * the smoother our behaviour from viewpoint of network, but the lower
321 * throughput and the higher sensitivity of the connection to losses. 8)
323 * rcv_ssthresh is more strict window_clamp used at "slow start"
324 * phase to predict further behaviour of this connection.
325 * It is used for two goals:
326 * - to enforce header prediction at sender, even when application
327 * requires some significant "application buffer". It is check #1.
328 * - to prevent pruning of receive queue because of misprediction
329 * of receiver window. Check #2.
331 * The scheme does not work when sender sends good segments opening
332 * window and then starts to feed us spaghetti. But it should work
333 * in common situations. Otherwise, we have to rely on queue collapsing.
336 /* Slow part of check#2. */
337 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
339 struct tcp_sock *tp = tcp_sk(sk);
341 int truesize = tcp_win_from_space(skb->truesize) >> 1;
342 int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
344 while (tp->rcv_ssthresh <= window) {
345 if (truesize <= skb->len)
346 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
354 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
356 struct tcp_sock *tp = tcp_sk(sk);
359 if (tp->rcv_ssthresh < tp->window_clamp &&
360 (int)tp->rcv_ssthresh < tcp_space(sk) &&
361 !sk_under_memory_pressure(sk)) {
364 /* Check #2. Increase window, if skb with such overhead
365 * will fit to rcvbuf in future.
367 if (tcp_win_from_space(skb->truesize) <= skb->len)
368 incr = 2 * tp->advmss;
370 incr = __tcp_grow_window(sk, skb);
373 incr = max_t(int, incr, 2 * skb->len);
374 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
376 inet_csk(sk)->icsk_ack.quick |= 1;
381 /* 3. Tuning rcvbuf, when connection enters established state. */
382 static void tcp_fixup_rcvbuf(struct sock *sk)
384 u32 mss = tcp_sk(sk)->advmss;
387 rcvmem = 2 * SKB_TRUESIZE(mss + MAX_TCP_HEADER) *
388 tcp_default_init_rwnd(mss);
390 /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
391 * Allow enough cushion so that sender is not limited by our window
393 if (sysctl_tcp_moderate_rcvbuf)
396 if (sk->sk_rcvbuf < rcvmem)
397 sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]);
400 /* 4. Try to fixup all. It is made immediately after connection enters
403 void tcp_init_buffer_space(struct sock *sk)
405 struct tcp_sock *tp = tcp_sk(sk);
408 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
409 tcp_fixup_rcvbuf(sk);
410 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
411 tcp_sndbuf_expand(sk);
413 tp->rcvq_space.space = tp->rcv_wnd;
414 tp->rcvq_space.time = tcp_time_stamp;
415 tp->rcvq_space.seq = tp->copied_seq;
417 maxwin = tcp_full_space(sk);
419 if (tp->window_clamp >= maxwin) {
420 tp->window_clamp = maxwin;
422 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
423 tp->window_clamp = max(maxwin -
424 (maxwin >> sysctl_tcp_app_win),
428 /* Force reservation of one segment. */
429 if (sysctl_tcp_app_win &&
430 tp->window_clamp > 2 * tp->advmss &&
431 tp->window_clamp + tp->advmss > maxwin)
432 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
434 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
435 tp->snd_cwnd_stamp = tcp_time_stamp;
438 /* 5. Recalculate window clamp after socket hit its memory bounds. */
439 static void tcp_clamp_window(struct sock *sk)
441 struct tcp_sock *tp = tcp_sk(sk);
442 struct inet_connection_sock *icsk = inet_csk(sk);
444 icsk->icsk_ack.quick = 0;
446 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
447 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
448 !sk_under_memory_pressure(sk) &&
449 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
450 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
453 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
454 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
457 /* Initialize RCV_MSS value.
458 * RCV_MSS is an our guess about MSS used by the peer.
459 * We haven't any direct information about the MSS.
460 * It's better to underestimate the RCV_MSS rather than overestimate.
461 * Overestimations make us ACKing less frequently than needed.
462 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
464 void tcp_initialize_rcv_mss(struct sock *sk)
466 const struct tcp_sock *tp = tcp_sk(sk);
467 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
469 hint = min(hint, tp->rcv_wnd / 2);
470 hint = min(hint, TCP_MSS_DEFAULT);
471 hint = max(hint, TCP_MIN_MSS);
473 inet_csk(sk)->icsk_ack.rcv_mss = hint;
475 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
477 /* Receiver "autotuning" code.
479 * The algorithm for RTT estimation w/o timestamps is based on
480 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
481 * <http://public.lanl.gov/radiant/pubs.html#DRS>
483 * More detail on this code can be found at
484 * <http://staff.psc.edu/jheffner/>,
485 * though this reference is out of date. A new paper
488 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
490 u32 new_sample = tp->rcv_rtt_est.rtt;
496 if (new_sample != 0) {
497 /* If we sample in larger samples in the non-timestamp
498 * case, we could grossly overestimate the RTT especially
499 * with chatty applications or bulk transfer apps which
500 * are stalled on filesystem I/O.
502 * Also, since we are only going for a minimum in the
503 * non-timestamp case, we do not smooth things out
504 * else with timestamps disabled convergence takes too
508 m -= (new_sample >> 3);
516 /* No previous measure. */
520 if (tp->rcv_rtt_est.rtt != new_sample)
521 tp->rcv_rtt_est.rtt = new_sample;
524 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
526 if (tp->rcv_rtt_est.time == 0)
528 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
530 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rcv_rtt_est.time, 1);
533 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
534 tp->rcv_rtt_est.time = tcp_time_stamp;
537 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
538 const struct sk_buff *skb)
540 struct tcp_sock *tp = tcp_sk(sk);
541 if (tp->rx_opt.rcv_tsecr &&
542 (TCP_SKB_CB(skb)->end_seq -
543 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
544 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
548 * This function should be called every time data is copied to user space.
549 * It calculates the appropriate TCP receive buffer space.
551 void tcp_rcv_space_adjust(struct sock *sk)
553 struct tcp_sock *tp = tcp_sk(sk);
557 time = tcp_time_stamp - tp->rcvq_space.time;
558 if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
561 /* Number of bytes copied to user in last RTT */
562 copied = tp->copied_seq - tp->rcvq_space.seq;
563 if (copied <= tp->rcvq_space.space)
567 * copied = bytes received in previous RTT, our base window
568 * To cope with packet losses, we need a 2x factor
569 * To cope with slow start, and sender growing its cwin by 100 %
570 * every RTT, we need a 4x factor, because the ACK we are sending
571 * now is for the next RTT, not the current one :
572 * <prev RTT . ><current RTT .. ><next RTT .... >
575 if (sysctl_tcp_moderate_rcvbuf &&
576 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
577 int rcvwin, rcvmem, rcvbuf;
579 /* minimal window to cope with packet losses, assuming
580 * steady state. Add some cushion because of small variations.
582 rcvwin = (copied << 1) + 16 * tp->advmss;
584 /* If rate increased by 25%,
585 * assume slow start, rcvwin = 3 * copied
586 * If rate increased by 50%,
587 * assume sender can use 2x growth, rcvwin = 4 * copied
590 tp->rcvq_space.space + (tp->rcvq_space.space >> 2)) {
592 tp->rcvq_space.space + (tp->rcvq_space.space >> 1))
595 rcvwin += (rcvwin >> 1);
598 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
599 while (tcp_win_from_space(rcvmem) < tp->advmss)
602 rcvbuf = min(rcvwin / tp->advmss * rcvmem, sysctl_tcp_rmem[2]);
603 if (rcvbuf > sk->sk_rcvbuf) {
604 sk->sk_rcvbuf = rcvbuf;
606 /* Make the window clamp follow along. */
607 tp->window_clamp = rcvwin;
610 tp->rcvq_space.space = copied;
613 tp->rcvq_space.seq = tp->copied_seq;
614 tp->rcvq_space.time = tcp_time_stamp;
617 /* There is something which you must keep in mind when you analyze the
618 * behavior of the tp->ato delayed ack timeout interval. When a
619 * connection starts up, we want to ack as quickly as possible. The
620 * problem is that "good" TCP's do slow start at the beginning of data
621 * transmission. The means that until we send the first few ACK's the
622 * sender will sit on his end and only queue most of his data, because
623 * he can only send snd_cwnd unacked packets at any given time. For
624 * each ACK we send, he increments snd_cwnd and transmits more of his
627 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
629 struct tcp_sock *tp = tcp_sk(sk);
630 struct inet_connection_sock *icsk = inet_csk(sk);
633 inet_csk_schedule_ack(sk);
635 tcp_measure_rcv_mss(sk, skb);
637 tcp_rcv_rtt_measure(tp);
639 now = tcp_time_stamp;
641 if (!icsk->icsk_ack.ato) {
642 /* The _first_ data packet received, initialize
643 * delayed ACK engine.
645 tcp_incr_quickack(sk);
646 icsk->icsk_ack.ato = TCP_ATO_MIN;
648 int m = now - icsk->icsk_ack.lrcvtime;
650 if (m <= TCP_ATO_MIN / 2) {
651 /* The fastest case is the first. */
652 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
653 } else if (m < icsk->icsk_ack.ato) {
654 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
655 if (icsk->icsk_ack.ato > icsk->icsk_rto)
656 icsk->icsk_ack.ato = icsk->icsk_rto;
657 } else if (m > icsk->icsk_rto) {
658 /* Too long gap. Apparently sender failed to
659 * restart window, so that we send ACKs quickly.
661 tcp_incr_quickack(sk);
665 icsk->icsk_ack.lrcvtime = now;
667 tcp_ecn_check_ce(tp, skb);
670 tcp_grow_window(sk, skb);
673 /* Called to compute a smoothed rtt estimate. The data fed to this
674 * routine either comes from timestamps, or from segments that were
675 * known _not_ to have been retransmitted [see Karn/Partridge
676 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
677 * piece by Van Jacobson.
678 * NOTE: the next three routines used to be one big routine.
679 * To save cycles in the RFC 1323 implementation it was better to break
680 * it up into three procedures. -- erics
682 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
684 struct tcp_sock *tp = tcp_sk(sk);
685 long m = mrtt_us; /* RTT */
686 u32 srtt = tp->srtt_us;
688 /* The following amusing code comes from Jacobson's
689 * article in SIGCOMM '88. Note that rtt and mdev
690 * are scaled versions of rtt and mean deviation.
691 * This is designed to be as fast as possible
692 * m stands for "measurement".
694 * On a 1990 paper the rto value is changed to:
695 * RTO = rtt + 4 * mdev
697 * Funny. This algorithm seems to be very broken.
698 * These formulae increase RTO, when it should be decreased, increase
699 * too slowly, when it should be increased quickly, decrease too quickly
700 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
701 * does not matter how to _calculate_ it. Seems, it was trap
702 * that VJ failed to avoid. 8)
705 m -= (srtt >> 3); /* m is now error in rtt est */
706 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
708 m = -m; /* m is now abs(error) */
709 m -= (tp->mdev_us >> 2); /* similar update on mdev */
710 /* This is similar to one of Eifel findings.
711 * Eifel blocks mdev updates when rtt decreases.
712 * This solution is a bit different: we use finer gain
713 * for mdev in this case (alpha*beta).
714 * Like Eifel it also prevents growth of rto,
715 * but also it limits too fast rto decreases,
716 * happening in pure Eifel.
721 m -= (tp->mdev_us >> 2); /* similar update on mdev */
723 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
724 if (tp->mdev_us > tp->mdev_max_us) {
725 tp->mdev_max_us = tp->mdev_us;
726 if (tp->mdev_max_us > tp->rttvar_us)
727 tp->rttvar_us = tp->mdev_max_us;
729 if (after(tp->snd_una, tp->rtt_seq)) {
730 if (tp->mdev_max_us < tp->rttvar_us)
731 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
732 tp->rtt_seq = tp->snd_nxt;
733 tp->mdev_max_us = tcp_rto_min_us(sk);
736 /* no previous measure. */
737 srtt = m << 3; /* take the measured time to be rtt */
738 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
739 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
740 tp->mdev_max_us = tp->rttvar_us;
741 tp->rtt_seq = tp->snd_nxt;
743 tp->srtt_us = max(1U, srtt);
746 /* Set the sk_pacing_rate to allow proper sizing of TSO packets.
747 * Note: TCP stack does not yet implement pacing.
748 * FQ packet scheduler can be used to implement cheap but effective
749 * TCP pacing, to smooth the burst on large writes when packets
750 * in flight is significantly lower than cwnd (or rwin)
752 static void tcp_update_pacing_rate(struct sock *sk)
754 const struct tcp_sock *tp = tcp_sk(sk);
757 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
758 rate = (u64)tp->mss_cache * 2 * (USEC_PER_SEC << 3);
760 rate *= max(tp->snd_cwnd, tp->packets_out);
762 if (likely(tp->srtt_us))
763 do_div(rate, tp->srtt_us);
765 /* ACCESS_ONCE() is needed because sch_fq fetches sk_pacing_rate
766 * without any lock. We want to make sure compiler wont store
767 * intermediate values in this location.
769 ACCESS_ONCE(sk->sk_pacing_rate) = min_t(u64, rate,
770 sk->sk_max_pacing_rate);
773 /* Calculate rto without backoff. This is the second half of Van Jacobson's
774 * routine referred to above.
776 static void tcp_set_rto(struct sock *sk)
778 const struct tcp_sock *tp = tcp_sk(sk);
779 /* Old crap is replaced with new one. 8)
782 * 1. If rtt variance happened to be less 50msec, it is hallucination.
783 * It cannot be less due to utterly erratic ACK generation made
784 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
785 * to do with delayed acks, because at cwnd>2 true delack timeout
786 * is invisible. Actually, Linux-2.4 also generates erratic
787 * ACKs in some circumstances.
789 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
791 /* 2. Fixups made earlier cannot be right.
792 * If we do not estimate RTO correctly without them,
793 * all the algo is pure shit and should be replaced
794 * with correct one. It is exactly, which we pretend to do.
797 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
798 * guarantees that rto is higher.
803 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
805 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
808 cwnd = TCP_INIT_CWND;
809 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
813 * Packet counting of FACK is based on in-order assumptions, therefore TCP
814 * disables it when reordering is detected
816 void tcp_disable_fack(struct tcp_sock *tp)
818 /* RFC3517 uses different metric in lost marker => reset on change */
820 tp->lost_skb_hint = NULL;
821 tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED;
824 /* Take a notice that peer is sending D-SACKs */
825 static void tcp_dsack_seen(struct tcp_sock *tp)
827 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
830 static void tcp_update_reordering(struct sock *sk, const int metric,
833 struct tcp_sock *tp = tcp_sk(sk);
834 if (metric > tp->reordering) {
837 tp->reordering = min(sysctl_tcp_max_reordering, metric);
839 /* This exciting event is worth to be remembered. 8) */
841 mib_idx = LINUX_MIB_TCPTSREORDER;
842 else if (tcp_is_reno(tp))
843 mib_idx = LINUX_MIB_TCPRENOREORDER;
844 else if (tcp_is_fack(tp))
845 mib_idx = LINUX_MIB_TCPFACKREORDER;
847 mib_idx = LINUX_MIB_TCPSACKREORDER;
849 NET_INC_STATS_BH(sock_net(sk), mib_idx);
850 #if FASTRETRANS_DEBUG > 1
851 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
852 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
856 tp->undo_marker ? tp->undo_retrans : 0);
858 tcp_disable_fack(tp);
862 tcp_disable_early_retrans(tp);
865 /* This must be called before lost_out is incremented */
866 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
868 if ((tp->retransmit_skb_hint == NULL) ||
869 before(TCP_SKB_CB(skb)->seq,
870 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
871 tp->retransmit_skb_hint = skb;
874 after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
875 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
878 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
880 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
881 tcp_verify_retransmit_hint(tp, skb);
883 tp->lost_out += tcp_skb_pcount(skb);
884 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
888 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp,
891 tcp_verify_retransmit_hint(tp, skb);
893 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
894 tp->lost_out += tcp_skb_pcount(skb);
895 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
899 /* This procedure tags the retransmission queue when SACKs arrive.
901 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
902 * Packets in queue with these bits set are counted in variables
903 * sacked_out, retrans_out and lost_out, correspondingly.
905 * Valid combinations are:
906 * Tag InFlight Description
907 * 0 1 - orig segment is in flight.
908 * S 0 - nothing flies, orig reached receiver.
909 * L 0 - nothing flies, orig lost by net.
910 * R 2 - both orig and retransmit are in flight.
911 * L|R 1 - orig is lost, retransmit is in flight.
912 * S|R 1 - orig reached receiver, retrans is still in flight.
913 * (L|S|R is logically valid, it could occur when L|R is sacked,
914 * but it is equivalent to plain S and code short-curcuits it to S.
915 * L|S is logically invalid, it would mean -1 packet in flight 8))
917 * These 6 states form finite state machine, controlled by the following events:
918 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
919 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
920 * 3. Loss detection event of two flavors:
921 * A. Scoreboard estimator decided the packet is lost.
922 * A'. Reno "three dupacks" marks head of queue lost.
923 * A''. Its FACK modification, head until snd.fack is lost.
924 * B. SACK arrives sacking SND.NXT at the moment, when the
925 * segment was retransmitted.
926 * 4. D-SACK added new rule: D-SACK changes any tag to S.
928 * It is pleasant to note, that state diagram turns out to be commutative,
929 * so that we are allowed not to be bothered by order of our actions,
930 * when multiple events arrive simultaneously. (see the function below).
932 * Reordering detection.
933 * --------------------
934 * Reordering metric is maximal distance, which a packet can be displaced
935 * in packet stream. With SACKs we can estimate it:
937 * 1. SACK fills old hole and the corresponding segment was not
938 * ever retransmitted -> reordering. Alas, we cannot use it
939 * when segment was retransmitted.
940 * 2. The last flaw is solved with D-SACK. D-SACK arrives
941 * for retransmitted and already SACKed segment -> reordering..
942 * Both of these heuristics are not used in Loss state, when we cannot
943 * account for retransmits accurately.
945 * SACK block validation.
946 * ----------------------
948 * SACK block range validation checks that the received SACK block fits to
949 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
950 * Note that SND.UNA is not included to the range though being valid because
951 * it means that the receiver is rather inconsistent with itself reporting
952 * SACK reneging when it should advance SND.UNA. Such SACK block this is
953 * perfectly valid, however, in light of RFC2018 which explicitly states
954 * that "SACK block MUST reflect the newest segment. Even if the newest
955 * segment is going to be discarded ...", not that it looks very clever
956 * in case of head skb. Due to potentional receiver driven attacks, we
957 * choose to avoid immediate execution of a walk in write queue due to
958 * reneging and defer head skb's loss recovery to standard loss recovery
959 * procedure that will eventually trigger (nothing forbids us doing this).
961 * Implements also blockage to start_seq wrap-around. Problem lies in the
962 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
963 * there's no guarantee that it will be before snd_nxt (n). The problem
964 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
967 * <- outs wnd -> <- wrapzone ->
968 * u e n u_w e_w s n_w
970 * |<------------+------+----- TCP seqno space --------------+---------->|
971 * ...-- <2^31 ->| |<--------...
972 * ...---- >2^31 ------>| |<--------...
974 * Current code wouldn't be vulnerable but it's better still to discard such
975 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
976 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
977 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
978 * equal to the ideal case (infinite seqno space without wrap caused issues).
980 * With D-SACK the lower bound is extended to cover sequence space below
981 * SND.UNA down to undo_marker, which is the last point of interest. Yet
982 * again, D-SACK block must not to go across snd_una (for the same reason as
983 * for the normal SACK blocks, explained above). But there all simplicity
984 * ends, TCP might receive valid D-SACKs below that. As long as they reside
985 * fully below undo_marker they do not affect behavior in anyway and can
986 * therefore be safely ignored. In rare cases (which are more or less
987 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
988 * fragmentation and packet reordering past skb's retransmission. To consider
989 * them correctly, the acceptable range must be extended even more though
990 * the exact amount is rather hard to quantify. However, tp->max_window can
991 * be used as an exaggerated estimate.
993 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
994 u32 start_seq, u32 end_seq)
996 /* Too far in future, or reversed (interpretation is ambiguous) */
997 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1000 /* Nasty start_seq wrap-around check (see comments above) */
1001 if (!before(start_seq, tp->snd_nxt))
1004 /* In outstanding window? ...This is valid exit for D-SACKs too.
1005 * start_seq == snd_una is non-sensical (see comments above)
1007 if (after(start_seq, tp->snd_una))
1010 if (!is_dsack || !tp->undo_marker)
1013 /* ...Then it's D-SACK, and must reside below snd_una completely */
1014 if (after(end_seq, tp->snd_una))
1017 if (!before(start_seq, tp->undo_marker))
1021 if (!after(end_seq, tp->undo_marker))
1024 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1025 * start_seq < undo_marker and end_seq >= undo_marker.
1027 return !before(start_seq, end_seq - tp->max_window);
1030 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1031 * Event "B". Later note: FACK people cheated me again 8), we have to account
1032 * for reordering! Ugly, but should help.
1034 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1035 * less than what is now known to be received by the other end (derived from
1036 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1037 * retransmitted skbs to avoid some costly processing per ACKs.
1039 static void tcp_mark_lost_retrans(struct sock *sk)
1041 const struct inet_connection_sock *icsk = inet_csk(sk);
1042 struct tcp_sock *tp = tcp_sk(sk);
1043 struct sk_buff *skb;
1045 u32 new_low_seq = tp->snd_nxt;
1046 u32 received_upto = tcp_highest_sack_seq(tp);
1048 if (!tcp_is_fack(tp) || !tp->retrans_out ||
1049 !after(received_upto, tp->lost_retrans_low) ||
1050 icsk->icsk_ca_state != TCP_CA_Recovery)
1053 tcp_for_write_queue(skb, sk) {
1054 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1056 if (skb == tcp_send_head(sk))
1058 if (cnt == tp->retrans_out)
1060 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1063 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1066 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1067 * constraint here (see above) but figuring out that at
1068 * least tp->reordering SACK blocks reside between ack_seq
1069 * and received_upto is not easy task to do cheaply with
1070 * the available datastructures.
1072 * Whether FACK should check here for tp->reordering segs
1073 * in-between one could argue for either way (it would be
1074 * rather simple to implement as we could count fack_count
1075 * during the walk and do tp->fackets_out - fack_count).
1077 if (after(received_upto, ack_seq)) {
1078 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1079 tp->retrans_out -= tcp_skb_pcount(skb);
1081 tcp_skb_mark_lost_uncond_verify(tp, skb);
1082 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
1084 if (before(ack_seq, new_low_seq))
1085 new_low_seq = ack_seq;
1086 cnt += tcp_skb_pcount(skb);
1090 if (tp->retrans_out)
1091 tp->lost_retrans_low = new_low_seq;
1094 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1095 struct tcp_sack_block_wire *sp, int num_sacks,
1098 struct tcp_sock *tp = tcp_sk(sk);
1099 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1100 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1101 bool dup_sack = false;
1103 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1106 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1107 } else if (num_sacks > 1) {
1108 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1109 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1111 if (!after(end_seq_0, end_seq_1) &&
1112 !before(start_seq_0, start_seq_1)) {
1115 NET_INC_STATS_BH(sock_net(sk),
1116 LINUX_MIB_TCPDSACKOFORECV);
1120 /* D-SACK for already forgotten data... Do dumb counting. */
1121 if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 &&
1122 !after(end_seq_0, prior_snd_una) &&
1123 after(end_seq_0, tp->undo_marker))
1129 struct tcp_sacktag_state {
1132 long rtt_us; /* RTT measured by SACKing never-retransmitted data */
1136 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1137 * the incoming SACK may not exactly match but we can find smaller MSS
1138 * aligned portion of it that matches. Therefore we might need to fragment
1139 * which may fail and creates some hassle (caller must handle error case
1142 * FIXME: this could be merged to shift decision code
1144 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1145 u32 start_seq, u32 end_seq)
1149 unsigned int pkt_len;
1152 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1153 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1155 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1156 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1157 mss = tcp_skb_mss(skb);
1158 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1161 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1165 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1170 /* Round if necessary so that SACKs cover only full MSSes
1171 * and/or the remaining small portion (if present)
1173 if (pkt_len > mss) {
1174 unsigned int new_len = (pkt_len / mss) * mss;
1175 if (!in_sack && new_len < pkt_len) {
1177 if (new_len >= skb->len)
1182 err = tcp_fragment(sk, skb, pkt_len, mss, GFP_ATOMIC);
1190 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1191 static u8 tcp_sacktag_one(struct sock *sk,
1192 struct tcp_sacktag_state *state, u8 sacked,
1193 u32 start_seq, u32 end_seq,
1194 int dup_sack, int pcount,
1195 const struct skb_mstamp *xmit_time)
1197 struct tcp_sock *tp = tcp_sk(sk);
1198 int fack_count = state->fack_count;
1200 /* Account D-SACK for retransmitted packet. */
1201 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1202 if (tp->undo_marker && tp->undo_retrans > 0 &&
1203 after(end_seq, tp->undo_marker))
1205 if (sacked & TCPCB_SACKED_ACKED)
1206 state->reord = min(fack_count, state->reord);
1209 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1210 if (!after(end_seq, tp->snd_una))
1213 if (!(sacked & TCPCB_SACKED_ACKED)) {
1214 if (sacked & TCPCB_SACKED_RETRANS) {
1215 /* If the segment is not tagged as lost,
1216 * we do not clear RETRANS, believing
1217 * that retransmission is still in flight.
1219 if (sacked & TCPCB_LOST) {
1220 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1221 tp->lost_out -= pcount;
1222 tp->retrans_out -= pcount;
1225 if (!(sacked & TCPCB_RETRANS)) {
1226 /* New sack for not retransmitted frame,
1227 * which was in hole. It is reordering.
1229 if (before(start_seq,
1230 tcp_highest_sack_seq(tp)))
1231 state->reord = min(fack_count,
1233 if (!after(end_seq, tp->high_seq))
1234 state->flag |= FLAG_ORIG_SACK_ACKED;
1235 /* Pick the earliest sequence sacked for RTT */
1236 if (state->rtt_us < 0) {
1237 struct skb_mstamp now;
1239 skb_mstamp_get(&now);
1240 state->rtt_us = skb_mstamp_us_delta(&now,
1245 if (sacked & TCPCB_LOST) {
1246 sacked &= ~TCPCB_LOST;
1247 tp->lost_out -= pcount;
1251 sacked |= TCPCB_SACKED_ACKED;
1252 state->flag |= FLAG_DATA_SACKED;
1253 tp->sacked_out += pcount;
1255 fack_count += pcount;
1257 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1258 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
1259 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1260 tp->lost_cnt_hint += pcount;
1262 if (fack_count > tp->fackets_out)
1263 tp->fackets_out = fack_count;
1266 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1267 * frames and clear it. undo_retrans is decreased above, L|R frames
1268 * are accounted above as well.
1270 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1271 sacked &= ~TCPCB_SACKED_RETRANS;
1272 tp->retrans_out -= pcount;
1278 /* Shift newly-SACKed bytes from this skb to the immediately previous
1279 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1281 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1282 struct tcp_sacktag_state *state,
1283 unsigned int pcount, int shifted, int mss,
1286 struct tcp_sock *tp = tcp_sk(sk);
1287 struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1288 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1289 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1293 /* Adjust counters and hints for the newly sacked sequence
1294 * range but discard the return value since prev is already
1295 * marked. We must tag the range first because the seq
1296 * advancement below implicitly advances
1297 * tcp_highest_sack_seq() when skb is highest_sack.
1299 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1300 start_seq, end_seq, dup_sack, pcount,
1303 if (skb == tp->lost_skb_hint)
1304 tp->lost_cnt_hint += pcount;
1306 TCP_SKB_CB(prev)->end_seq += shifted;
1307 TCP_SKB_CB(skb)->seq += shifted;
1309 tcp_skb_pcount_add(prev, pcount);
1310 BUG_ON(tcp_skb_pcount(skb) < pcount);
1311 tcp_skb_pcount_add(skb, -pcount);
1313 /* When we're adding to gso_segs == 1, gso_size will be zero,
1314 * in theory this shouldn't be necessary but as long as DSACK
1315 * code can come after this skb later on it's better to keep
1316 * setting gso_size to something.
1318 if (!skb_shinfo(prev)->gso_size) {
1319 skb_shinfo(prev)->gso_size = mss;
1320 skb_shinfo(prev)->gso_type = sk->sk_gso_type;
1323 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1324 if (tcp_skb_pcount(skb) <= 1) {
1325 skb_shinfo(skb)->gso_size = 0;
1326 skb_shinfo(skb)->gso_type = 0;
1329 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1330 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1333 BUG_ON(!tcp_skb_pcount(skb));
1334 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1338 /* Whole SKB was eaten :-) */
1340 if (skb == tp->retransmit_skb_hint)
1341 tp->retransmit_skb_hint = prev;
1342 if (skb == tp->lost_skb_hint) {
1343 tp->lost_skb_hint = prev;
1344 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1347 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1348 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1349 TCP_SKB_CB(prev)->end_seq++;
1351 if (skb == tcp_highest_sack(sk))
1352 tcp_advance_highest_sack(sk, skb);
1354 tcp_unlink_write_queue(skb, sk);
1355 sk_wmem_free_skb(sk, skb);
1357 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1362 /* I wish gso_size would have a bit more sane initialization than
1363 * something-or-zero which complicates things
1365 static int tcp_skb_seglen(const struct sk_buff *skb)
1367 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1370 /* Shifting pages past head area doesn't work */
1371 static int skb_can_shift(const struct sk_buff *skb)
1373 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1376 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1379 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1380 struct tcp_sacktag_state *state,
1381 u32 start_seq, u32 end_seq,
1384 struct tcp_sock *tp = tcp_sk(sk);
1385 struct sk_buff *prev;
1391 if (!sk_can_gso(sk))
1394 /* Normally R but no L won't result in plain S */
1396 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1398 if (!skb_can_shift(skb))
1400 /* This frame is about to be dropped (was ACKed). */
1401 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1404 /* Can only happen with delayed DSACK + discard craziness */
1405 if (unlikely(skb == tcp_write_queue_head(sk)))
1407 prev = tcp_write_queue_prev(sk, skb);
1409 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1412 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1413 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1417 pcount = tcp_skb_pcount(skb);
1418 mss = tcp_skb_seglen(skb);
1420 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1421 * drop this restriction as unnecessary
1423 if (mss != tcp_skb_seglen(prev))
1426 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1428 /* CHECKME: This is non-MSS split case only?, this will
1429 * cause skipped skbs due to advancing loop btw, original
1430 * has that feature too
1432 if (tcp_skb_pcount(skb) <= 1)
1435 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1437 /* TODO: head merge to next could be attempted here
1438 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1439 * though it might not be worth of the additional hassle
1441 * ...we can probably just fallback to what was done
1442 * previously. We could try merging non-SACKed ones
1443 * as well but it probably isn't going to buy off
1444 * because later SACKs might again split them, and
1445 * it would make skb timestamp tracking considerably
1451 len = end_seq - TCP_SKB_CB(skb)->seq;
1453 BUG_ON(len > skb->len);
1455 /* MSS boundaries should be honoured or else pcount will
1456 * severely break even though it makes things bit trickier.
1457 * Optimize common case to avoid most of the divides
1459 mss = tcp_skb_mss(skb);
1461 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1462 * drop this restriction as unnecessary
1464 if (mss != tcp_skb_seglen(prev))
1469 } else if (len < mss) {
1477 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1478 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1481 if (!skb_shift(prev, skb, len))
1483 if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1486 /* Hole filled allows collapsing with the next as well, this is very
1487 * useful when hole on every nth skb pattern happens
1489 if (prev == tcp_write_queue_tail(sk))
1491 skb = tcp_write_queue_next(sk, prev);
1493 if (!skb_can_shift(skb) ||
1494 (skb == tcp_send_head(sk)) ||
1495 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1496 (mss != tcp_skb_seglen(skb)))
1500 if (skb_shift(prev, skb, len)) {
1501 pcount += tcp_skb_pcount(skb);
1502 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1506 state->fack_count += pcount;
1513 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1517 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1518 struct tcp_sack_block *next_dup,
1519 struct tcp_sacktag_state *state,
1520 u32 start_seq, u32 end_seq,
1523 struct tcp_sock *tp = tcp_sk(sk);
1524 struct sk_buff *tmp;
1526 tcp_for_write_queue_from(skb, sk) {
1528 bool dup_sack = dup_sack_in;
1530 if (skb == tcp_send_head(sk))
1533 /* queue is in-order => we can short-circuit the walk early */
1534 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1537 if ((next_dup != NULL) &&
1538 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1539 in_sack = tcp_match_skb_to_sack(sk, skb,
1540 next_dup->start_seq,
1546 /* skb reference here is a bit tricky to get right, since
1547 * shifting can eat and free both this skb and the next,
1548 * so not even _safe variant of the loop is enough.
1551 tmp = tcp_shift_skb_data(sk, skb, state,
1552 start_seq, end_seq, dup_sack);
1561 in_sack = tcp_match_skb_to_sack(sk, skb,
1567 if (unlikely(in_sack < 0))
1571 TCP_SKB_CB(skb)->sacked =
1574 TCP_SKB_CB(skb)->sacked,
1575 TCP_SKB_CB(skb)->seq,
1576 TCP_SKB_CB(skb)->end_seq,
1578 tcp_skb_pcount(skb),
1581 if (!before(TCP_SKB_CB(skb)->seq,
1582 tcp_highest_sack_seq(tp)))
1583 tcp_advance_highest_sack(sk, skb);
1586 state->fack_count += tcp_skb_pcount(skb);
1591 /* Avoid all extra work that is being done by sacktag while walking in
1594 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1595 struct tcp_sacktag_state *state,
1598 tcp_for_write_queue_from(skb, sk) {
1599 if (skb == tcp_send_head(sk))
1602 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1605 state->fack_count += tcp_skb_pcount(skb);
1610 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1612 struct tcp_sack_block *next_dup,
1613 struct tcp_sacktag_state *state,
1616 if (next_dup == NULL)
1619 if (before(next_dup->start_seq, skip_to_seq)) {
1620 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1621 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1622 next_dup->start_seq, next_dup->end_seq,
1629 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1631 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1635 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1636 u32 prior_snd_una, long *sack_rtt_us)
1638 struct tcp_sock *tp = tcp_sk(sk);
1639 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1640 TCP_SKB_CB(ack_skb)->sacked);
1641 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1642 struct tcp_sack_block sp[TCP_NUM_SACKS];
1643 struct tcp_sack_block *cache;
1644 struct tcp_sacktag_state state;
1645 struct sk_buff *skb;
1646 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1648 bool found_dup_sack = false;
1650 int first_sack_index;
1653 state.reord = tp->packets_out;
1656 if (!tp->sacked_out) {
1657 if (WARN_ON(tp->fackets_out))
1658 tp->fackets_out = 0;
1659 tcp_highest_sack_reset(sk);
1662 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1663 num_sacks, prior_snd_una);
1665 state.flag |= FLAG_DSACKING_ACK;
1667 /* Eliminate too old ACKs, but take into
1668 * account more or less fresh ones, they can
1669 * contain valid SACK info.
1671 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1674 if (!tp->packets_out)
1678 first_sack_index = 0;
1679 for (i = 0; i < num_sacks; i++) {
1680 bool dup_sack = !i && found_dup_sack;
1682 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1683 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1685 if (!tcp_is_sackblock_valid(tp, dup_sack,
1686 sp[used_sacks].start_seq,
1687 sp[used_sacks].end_seq)) {
1691 if (!tp->undo_marker)
1692 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1694 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1696 /* Don't count olds caused by ACK reordering */
1697 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1698 !after(sp[used_sacks].end_seq, tp->snd_una))
1700 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1703 NET_INC_STATS_BH(sock_net(sk), mib_idx);
1705 first_sack_index = -1;
1709 /* Ignore very old stuff early */
1710 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1716 /* order SACK blocks to allow in order walk of the retrans queue */
1717 for (i = used_sacks - 1; i > 0; i--) {
1718 for (j = 0; j < i; j++) {
1719 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1720 swap(sp[j], sp[j + 1]);
1722 /* Track where the first SACK block goes to */
1723 if (j == first_sack_index)
1724 first_sack_index = j + 1;
1729 skb = tcp_write_queue_head(sk);
1730 state.fack_count = 0;
1733 if (!tp->sacked_out) {
1734 /* It's already past, so skip checking against it */
1735 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1737 cache = tp->recv_sack_cache;
1738 /* Skip empty blocks in at head of the cache */
1739 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1744 while (i < used_sacks) {
1745 u32 start_seq = sp[i].start_seq;
1746 u32 end_seq = sp[i].end_seq;
1747 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1748 struct tcp_sack_block *next_dup = NULL;
1750 if (found_dup_sack && ((i + 1) == first_sack_index))
1751 next_dup = &sp[i + 1];
1753 /* Skip too early cached blocks */
1754 while (tcp_sack_cache_ok(tp, cache) &&
1755 !before(start_seq, cache->end_seq))
1758 /* Can skip some work by looking recv_sack_cache? */
1759 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1760 after(end_seq, cache->start_seq)) {
1763 if (before(start_seq, cache->start_seq)) {
1764 skb = tcp_sacktag_skip(skb, sk, &state,
1766 skb = tcp_sacktag_walk(skb, sk, next_dup,
1773 /* Rest of the block already fully processed? */
1774 if (!after(end_seq, cache->end_seq))
1777 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1781 /* ...tail remains todo... */
1782 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1783 /* ...but better entrypoint exists! */
1784 skb = tcp_highest_sack(sk);
1787 state.fack_count = tp->fackets_out;
1792 skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq);
1793 /* Check overlap against next cached too (past this one already) */
1798 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1799 skb = tcp_highest_sack(sk);
1802 state.fack_count = tp->fackets_out;
1804 skb = tcp_sacktag_skip(skb, sk, &state, start_seq);
1807 skb = tcp_sacktag_walk(skb, sk, next_dup, &state,
1808 start_seq, end_seq, dup_sack);
1814 /* Clear the head of the cache sack blocks so we can skip it next time */
1815 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1816 tp->recv_sack_cache[i].start_seq = 0;
1817 tp->recv_sack_cache[i].end_seq = 0;
1819 for (j = 0; j < used_sacks; j++)
1820 tp->recv_sack_cache[i++] = sp[j];
1822 tcp_mark_lost_retrans(sk);
1824 tcp_verify_left_out(tp);
1826 if ((state.reord < tp->fackets_out) &&
1827 ((inet_csk(sk)->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker))
1828 tcp_update_reordering(sk, tp->fackets_out - state.reord, 0);
1832 #if FASTRETRANS_DEBUG > 0
1833 WARN_ON((int)tp->sacked_out < 0);
1834 WARN_ON((int)tp->lost_out < 0);
1835 WARN_ON((int)tp->retrans_out < 0);
1836 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1838 *sack_rtt_us = state.rtt_us;
1842 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1843 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1845 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1849 holes = max(tp->lost_out, 1U);
1850 holes = min(holes, tp->packets_out);
1852 if ((tp->sacked_out + holes) > tp->packets_out) {
1853 tp->sacked_out = tp->packets_out - holes;
1859 /* If we receive more dupacks than we expected counting segments
1860 * in assumption of absent reordering, interpret this as reordering.
1861 * The only another reason could be bug in receiver TCP.
1863 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1865 struct tcp_sock *tp = tcp_sk(sk);
1866 if (tcp_limit_reno_sacked(tp))
1867 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1870 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1872 static void tcp_add_reno_sack(struct sock *sk)
1874 struct tcp_sock *tp = tcp_sk(sk);
1876 tcp_check_reno_reordering(sk, 0);
1877 tcp_verify_left_out(tp);
1880 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1882 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1884 struct tcp_sock *tp = tcp_sk(sk);
1887 /* One ACK acked hole. The rest eat duplicate ACKs. */
1888 if (acked - 1 >= tp->sacked_out)
1891 tp->sacked_out -= acked - 1;
1893 tcp_check_reno_reordering(sk, acked);
1894 tcp_verify_left_out(tp);
1897 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1902 void tcp_clear_retrans(struct tcp_sock *tp)
1904 tp->retrans_out = 0;
1906 tp->undo_marker = 0;
1907 tp->undo_retrans = -1;
1908 tp->fackets_out = 0;
1912 static inline void tcp_init_undo(struct tcp_sock *tp)
1914 tp->undo_marker = tp->snd_una;
1915 /* Retransmission still in flight may cause DSACKs later. */
1916 tp->undo_retrans = tp->retrans_out ? : -1;
1919 /* Enter Loss state. If we detect SACK reneging, forget all SACK information
1920 * and reset tags completely, otherwise preserve SACKs. If receiver
1921 * dropped its ofo queue, we will know this due to reneging detection.
1923 void tcp_enter_loss(struct sock *sk)
1925 const struct inet_connection_sock *icsk = inet_csk(sk);
1926 struct tcp_sock *tp = tcp_sk(sk);
1927 struct sk_buff *skb;
1928 bool new_recovery = false;
1929 bool is_reneg; /* is receiver reneging on SACKs? */
1931 /* Reduce ssthresh if it has not yet been made inside this window. */
1932 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1933 !after(tp->high_seq, tp->snd_una) ||
1934 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1935 new_recovery = true;
1936 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1937 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1938 tcp_ca_event(sk, CA_EVENT_LOSS);
1942 tp->snd_cwnd_cnt = 0;
1943 tp->snd_cwnd_stamp = tcp_time_stamp;
1945 tp->retrans_out = 0;
1948 if (tcp_is_reno(tp))
1949 tcp_reset_reno_sack(tp);
1951 skb = tcp_write_queue_head(sk);
1952 is_reneg = skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED);
1954 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
1956 tp->fackets_out = 0;
1958 tcp_clear_all_retrans_hints(tp);
1960 tcp_for_write_queue(skb, sk) {
1961 if (skb == tcp_send_head(sk))
1964 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1965 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || is_reneg) {
1966 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1967 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1968 tp->lost_out += tcp_skb_pcount(skb);
1969 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
1972 tcp_verify_left_out(tp);
1974 /* Timeout in disordered state after receiving substantial DUPACKs
1975 * suggests that the degree of reordering is over-estimated.
1977 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
1978 tp->sacked_out >= sysctl_tcp_reordering)
1979 tp->reordering = min_t(unsigned int, tp->reordering,
1980 sysctl_tcp_reordering);
1981 tcp_set_ca_state(sk, TCP_CA_Loss);
1982 tp->high_seq = tp->snd_nxt;
1983 tcp_ecn_queue_cwr(tp);
1985 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1986 * loss recovery is underway except recurring timeout(s) on
1987 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1989 tp->frto = sysctl_tcp_frto &&
1990 (new_recovery || icsk->icsk_retransmits) &&
1991 !inet_csk(sk)->icsk_mtup.probe_size;
1994 /* If ACK arrived pointing to a remembered SACK, it means that our
1995 * remembered SACKs do not reflect real state of receiver i.e.
1996 * receiver _host_ is heavily congested (or buggy).
1998 * To avoid big spurious retransmission bursts due to transient SACK
1999 * scoreboard oddities that look like reneging, we give the receiver a
2000 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2001 * restore sanity to the SACK scoreboard. If the apparent reneging
2002 * persists until this RTO then we'll clear the SACK scoreboard.
2004 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
2006 if (flag & FLAG_SACK_RENEGING) {
2007 struct tcp_sock *tp = tcp_sk(sk);
2008 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2009 msecs_to_jiffies(10));
2011 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2012 delay, TCP_RTO_MAX);
2018 static inline int tcp_fackets_out(const struct tcp_sock *tp)
2020 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
2023 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2024 * counter when SACK is enabled (without SACK, sacked_out is used for
2027 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2028 * segments up to the highest received SACK block so far and holes in
2031 * With reordering, holes may still be in flight, so RFC3517 recovery
2032 * uses pure sacked_out (total number of SACKed segments) even though
2033 * it violates the RFC that uses duplicate ACKs, often these are equal
2034 * but when e.g. out-of-window ACKs or packet duplication occurs,
2035 * they differ. Since neither occurs due to loss, TCP should really
2038 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2040 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2043 static bool tcp_pause_early_retransmit(struct sock *sk, int flag)
2045 struct tcp_sock *tp = tcp_sk(sk);
2046 unsigned long delay;
2048 /* Delay early retransmit and entering fast recovery for
2049 * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
2050 * available, or RTO is scheduled to fire first.
2052 if (sysctl_tcp_early_retrans < 2 || sysctl_tcp_early_retrans > 3 ||
2053 (flag & FLAG_ECE) || !tp->srtt_us)
2056 delay = max(usecs_to_jiffies(tp->srtt_us >> 5),
2057 msecs_to_jiffies(2));
2059 if (!time_after(inet_csk(sk)->icsk_timeout, (jiffies + delay)))
2062 inet_csk_reset_xmit_timer(sk, ICSK_TIME_EARLY_RETRANS, delay,
2067 /* Linux NewReno/SACK/FACK/ECN state machine.
2068 * --------------------------------------
2070 * "Open" Normal state, no dubious events, fast path.
2071 * "Disorder" In all the respects it is "Open",
2072 * but requires a bit more attention. It is entered when
2073 * we see some SACKs or dupacks. It is split of "Open"
2074 * mainly to move some processing from fast path to slow one.
2075 * "CWR" CWND was reduced due to some Congestion Notification event.
2076 * It can be ECN, ICMP source quench, local device congestion.
2077 * "Recovery" CWND was reduced, we are fast-retransmitting.
2078 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2080 * tcp_fastretrans_alert() is entered:
2081 * - each incoming ACK, if state is not "Open"
2082 * - when arrived ACK is unusual, namely:
2087 * Counting packets in flight is pretty simple.
2089 * in_flight = packets_out - left_out + retrans_out
2091 * packets_out is SND.NXT-SND.UNA counted in packets.
2093 * retrans_out is number of retransmitted segments.
2095 * left_out is number of segments left network, but not ACKed yet.
2097 * left_out = sacked_out + lost_out
2099 * sacked_out: Packets, which arrived to receiver out of order
2100 * and hence not ACKed. With SACKs this number is simply
2101 * amount of SACKed data. Even without SACKs
2102 * it is easy to give pretty reliable estimate of this number,
2103 * counting duplicate ACKs.
2105 * lost_out: Packets lost by network. TCP has no explicit
2106 * "loss notification" feedback from network (for now).
2107 * It means that this number can be only _guessed_.
2108 * Actually, it is the heuristics to predict lossage that
2109 * distinguishes different algorithms.
2111 * F.e. after RTO, when all the queue is considered as lost,
2112 * lost_out = packets_out and in_flight = retrans_out.
2114 * Essentially, we have now two algorithms counting
2117 * FACK: It is the simplest heuristics. As soon as we decided
2118 * that something is lost, we decide that _all_ not SACKed
2119 * packets until the most forward SACK are lost. I.e.
2120 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2121 * It is absolutely correct estimate, if network does not reorder
2122 * packets. And it loses any connection to reality when reordering
2123 * takes place. We use FACK by default until reordering
2124 * is suspected on the path to this destination.
2126 * NewReno: when Recovery is entered, we assume that one segment
2127 * is lost (classic Reno). While we are in Recovery and
2128 * a partial ACK arrives, we assume that one more packet
2129 * is lost (NewReno). This heuristics are the same in NewReno
2132 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2133 * deflation etc. CWND is real congestion window, never inflated, changes
2134 * only according to classic VJ rules.
2136 * Really tricky (and requiring careful tuning) part of algorithm
2137 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2138 * The first determines the moment _when_ we should reduce CWND and,
2139 * hence, slow down forward transmission. In fact, it determines the moment
2140 * when we decide that hole is caused by loss, rather than by a reorder.
2142 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2143 * holes, caused by lost packets.
2145 * And the most logically complicated part of algorithm is undo
2146 * heuristics. We detect false retransmits due to both too early
2147 * fast retransmit (reordering) and underestimated RTO, analyzing
2148 * timestamps and D-SACKs. When we detect that some segments were
2149 * retransmitted by mistake and CWND reduction was wrong, we undo
2150 * window reduction and abort recovery phase. This logic is hidden
2151 * inside several functions named tcp_try_undo_<something>.
2154 /* This function decides, when we should leave Disordered state
2155 * and enter Recovery phase, reducing congestion window.
2157 * Main question: may we further continue forward transmission
2158 * with the same cwnd?
2160 static bool tcp_time_to_recover(struct sock *sk, int flag)
2162 struct tcp_sock *tp = tcp_sk(sk);
2165 /* Trick#1: The loss is proven. */
2169 /* Not-A-Trick#2 : Classic rule... */
2170 if (tcp_dupack_heuristics(tp) > tp->reordering)
2173 /* Trick#4: It is still not OK... But will it be useful to delay
2176 packets_out = tp->packets_out;
2177 if (packets_out <= tp->reordering &&
2178 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2179 !tcp_may_send_now(sk)) {
2180 /* We have nothing to send. This connection is limited
2181 * either by receiver window or by application.
2186 /* If a thin stream is detected, retransmit after first
2187 * received dupack. Employ only if SACK is supported in order
2188 * to avoid possible corner-case series of spurious retransmissions
2189 * Use only if there are no unsent data.
2191 if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2192 tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2193 tcp_is_sack(tp) && !tcp_send_head(sk))
2196 /* Trick#6: TCP early retransmit, per RFC5827. To avoid spurious
2197 * retransmissions due to small network reorderings, we implement
2198 * Mitigation A.3 in the RFC and delay the retransmission for a short
2199 * interval if appropriate.
2201 if (tp->do_early_retrans && !tp->retrans_out && tp->sacked_out &&
2202 (tp->packets_out >= (tp->sacked_out + 1) && tp->packets_out < 4) &&
2203 !tcp_may_send_now(sk))
2204 return !tcp_pause_early_retransmit(sk, flag);
2209 /* Detect loss in event "A" above by marking head of queue up as lost.
2210 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2211 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2212 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2213 * the maximum SACKed segments to pass before reaching this limit.
2215 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2217 struct tcp_sock *tp = tcp_sk(sk);
2218 struct sk_buff *skb;
2222 /* Use SACK to deduce losses of new sequences sent during recovery */
2223 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq;
2225 WARN_ON(packets > tp->packets_out);
2226 if (tp->lost_skb_hint) {
2227 skb = tp->lost_skb_hint;
2228 cnt = tp->lost_cnt_hint;
2229 /* Head already handled? */
2230 if (mark_head && skb != tcp_write_queue_head(sk))
2233 skb = tcp_write_queue_head(sk);
2237 tcp_for_write_queue_from(skb, sk) {
2238 if (skb == tcp_send_head(sk))
2240 /* TODO: do this better */
2241 /* this is not the most efficient way to do this... */
2242 tp->lost_skb_hint = skb;
2243 tp->lost_cnt_hint = cnt;
2245 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2249 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2250 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2251 cnt += tcp_skb_pcount(skb);
2253 if (cnt > packets) {
2254 if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2255 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2256 (oldcnt >= packets))
2259 mss = skb_shinfo(skb)->gso_size;
2260 err = tcp_fragment(sk, skb, (packets - oldcnt) * mss,
2267 tcp_skb_mark_lost(tp, skb);
2272 tcp_verify_left_out(tp);
2275 /* Account newly detected lost packet(s) */
2277 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2279 struct tcp_sock *tp = tcp_sk(sk);
2281 if (tcp_is_reno(tp)) {
2282 tcp_mark_head_lost(sk, 1, 1);
2283 } else if (tcp_is_fack(tp)) {
2284 int lost = tp->fackets_out - tp->reordering;
2287 tcp_mark_head_lost(sk, lost, 0);
2289 int sacked_upto = tp->sacked_out - tp->reordering;
2290 if (sacked_upto >= 0)
2291 tcp_mark_head_lost(sk, sacked_upto, 0);
2292 else if (fast_rexmit)
2293 tcp_mark_head_lost(sk, 1, 1);
2297 /* CWND moderation, preventing bursts due to too big ACKs
2298 * in dubious situations.
2300 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2302 tp->snd_cwnd = min(tp->snd_cwnd,
2303 tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2304 tp->snd_cwnd_stamp = tcp_time_stamp;
2307 /* Nothing was retransmitted or returned timestamp is less
2308 * than timestamp of the first retransmission.
2310 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2312 return !tp->retrans_stamp ||
2313 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2314 before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
2317 /* Undo procedures. */
2319 #if FASTRETRANS_DEBUG > 1
2320 static void DBGUNDO(struct sock *sk, const char *msg)
2322 struct tcp_sock *tp = tcp_sk(sk);
2323 struct inet_sock *inet = inet_sk(sk);
2325 if (sk->sk_family == AF_INET) {
2326 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2328 &inet->inet_daddr, ntohs(inet->inet_dport),
2329 tp->snd_cwnd, tcp_left_out(tp),
2330 tp->snd_ssthresh, tp->prior_ssthresh,
2333 #if IS_ENABLED(CONFIG_IPV6)
2334 else if (sk->sk_family == AF_INET6) {
2335 struct ipv6_pinfo *np = inet6_sk(sk);
2336 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2338 &np->daddr, ntohs(inet->inet_dport),
2339 tp->snd_cwnd, tcp_left_out(tp),
2340 tp->snd_ssthresh, tp->prior_ssthresh,
2346 #define DBGUNDO(x...) do { } while (0)
2349 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2351 struct tcp_sock *tp = tcp_sk(sk);
2354 struct sk_buff *skb;
2356 tcp_for_write_queue(skb, sk) {
2357 if (skb == tcp_send_head(sk))
2359 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2362 tcp_clear_all_retrans_hints(tp);
2365 if (tp->prior_ssthresh) {
2366 const struct inet_connection_sock *icsk = inet_csk(sk);
2368 if (icsk->icsk_ca_ops->undo_cwnd)
2369 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2371 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2373 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2374 tp->snd_ssthresh = tp->prior_ssthresh;
2375 tcp_ecn_withdraw_cwr(tp);
2378 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2380 tp->snd_cwnd_stamp = tcp_time_stamp;
2381 tp->undo_marker = 0;
2384 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2386 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2389 /* People celebrate: "We love our President!" */
2390 static bool tcp_try_undo_recovery(struct sock *sk)
2392 struct tcp_sock *tp = tcp_sk(sk);
2394 if (tcp_may_undo(tp)) {
2397 /* Happy end! We did not retransmit anything
2398 * or our original transmission succeeded.
2400 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2401 tcp_undo_cwnd_reduction(sk, false);
2402 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2403 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2405 mib_idx = LINUX_MIB_TCPFULLUNDO;
2407 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2409 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2410 /* Hold old state until something *above* high_seq
2411 * is ACKed. For Reno it is MUST to prevent false
2412 * fast retransmits (RFC2582). SACK TCP is safe. */
2413 tcp_moderate_cwnd(tp);
2416 tcp_set_ca_state(sk, TCP_CA_Open);
2420 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2421 static bool tcp_try_undo_dsack(struct sock *sk)
2423 struct tcp_sock *tp = tcp_sk(sk);
2425 if (tp->undo_marker && !tp->undo_retrans) {
2426 DBGUNDO(sk, "D-SACK");
2427 tcp_undo_cwnd_reduction(sk, false);
2428 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2434 /* We can clear retrans_stamp when there are no retransmissions in the
2435 * window. It would seem that it is trivially available for us in
2436 * tp->retrans_out, however, that kind of assumptions doesn't consider
2437 * what will happen if errors occur when sending retransmission for the
2438 * second time. ...It could the that such segment has only
2439 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2440 * the head skb is enough except for some reneging corner cases that
2441 * are not worth the effort.
2443 * Main reason for all this complexity is the fact that connection dying
2444 * time now depends on the validity of the retrans_stamp, in particular,
2445 * that successive retransmissions of a segment must not advance
2446 * retrans_stamp under any conditions.
2448 static bool tcp_any_retrans_done(const struct sock *sk)
2450 const struct tcp_sock *tp = tcp_sk(sk);
2451 struct sk_buff *skb;
2453 if (tp->retrans_out)
2456 skb = tcp_write_queue_head(sk);
2457 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2463 /* Undo during loss recovery after partial ACK or using F-RTO. */
2464 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2466 struct tcp_sock *tp = tcp_sk(sk);
2468 if (frto_undo || tcp_may_undo(tp)) {
2469 tcp_undo_cwnd_reduction(sk, true);
2471 DBGUNDO(sk, "partial loss");
2472 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2474 NET_INC_STATS_BH(sock_net(sk),
2475 LINUX_MIB_TCPSPURIOUSRTOS);
2476 inet_csk(sk)->icsk_retransmits = 0;
2477 if (frto_undo || tcp_is_sack(tp))
2478 tcp_set_ca_state(sk, TCP_CA_Open);
2484 /* The cwnd reduction in CWR and Recovery use the PRR algorithm
2485 * https://datatracker.ietf.org/doc/draft-ietf-tcpm-proportional-rate-reduction/
2486 * It computes the number of packets to send (sndcnt) based on packets newly
2488 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2489 * cwnd reductions across a full RTT.
2490 * 2) If packets in flight is lower than ssthresh (such as due to excess
2491 * losses and/or application stalls), do not perform any further cwnd
2492 * reductions, but instead slow start up to ssthresh.
2494 static void tcp_init_cwnd_reduction(struct sock *sk)
2496 struct tcp_sock *tp = tcp_sk(sk);
2498 tp->high_seq = tp->snd_nxt;
2499 tp->tlp_high_seq = 0;
2500 tp->snd_cwnd_cnt = 0;
2501 tp->prior_cwnd = tp->snd_cwnd;
2502 tp->prr_delivered = 0;
2504 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2505 tcp_ecn_queue_cwr(tp);
2508 static void tcp_cwnd_reduction(struct sock *sk, const int prior_unsacked,
2511 struct tcp_sock *tp = tcp_sk(sk);
2513 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2514 int newly_acked_sacked = prior_unsacked -
2515 (tp->packets_out - tp->sacked_out);
2517 tp->prr_delivered += newly_acked_sacked;
2518 if (tcp_packets_in_flight(tp) > tp->snd_ssthresh) {
2519 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2521 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2523 sndcnt = min_t(int, delta,
2524 max_t(int, tp->prr_delivered - tp->prr_out,
2525 newly_acked_sacked) + 1);
2528 sndcnt = max(sndcnt, (fast_rexmit ? 1 : 0));
2529 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2532 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2534 struct tcp_sock *tp = tcp_sk(sk);
2536 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2537 if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR ||
2538 (tp->undo_marker && tp->snd_ssthresh < TCP_INFINITE_SSTHRESH)) {
2539 tp->snd_cwnd = tp->snd_ssthresh;
2540 tp->snd_cwnd_stamp = tcp_time_stamp;
2542 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2545 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2546 void tcp_enter_cwr(struct sock *sk)
2548 struct tcp_sock *tp = tcp_sk(sk);
2550 tp->prior_ssthresh = 0;
2551 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2552 tp->undo_marker = 0;
2553 tcp_init_cwnd_reduction(sk);
2554 tcp_set_ca_state(sk, TCP_CA_CWR);
2558 static void tcp_try_keep_open(struct sock *sk)
2560 struct tcp_sock *tp = tcp_sk(sk);
2561 int state = TCP_CA_Open;
2563 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2564 state = TCP_CA_Disorder;
2566 if (inet_csk(sk)->icsk_ca_state != state) {
2567 tcp_set_ca_state(sk, state);
2568 tp->high_seq = tp->snd_nxt;
2572 static void tcp_try_to_open(struct sock *sk, int flag, const int prior_unsacked)
2574 struct tcp_sock *tp = tcp_sk(sk);
2576 tcp_verify_left_out(tp);
2578 if (!tcp_any_retrans_done(sk))
2579 tp->retrans_stamp = 0;
2581 if (flag & FLAG_ECE)
2584 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2585 tcp_try_keep_open(sk);
2587 tcp_cwnd_reduction(sk, prior_unsacked, 0);
2591 static void tcp_mtup_probe_failed(struct sock *sk)
2593 struct inet_connection_sock *icsk = inet_csk(sk);
2595 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2596 icsk->icsk_mtup.probe_size = 0;
2599 static void tcp_mtup_probe_success(struct sock *sk)
2601 struct tcp_sock *tp = tcp_sk(sk);
2602 struct inet_connection_sock *icsk = inet_csk(sk);
2604 /* FIXME: breaks with very large cwnd */
2605 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2606 tp->snd_cwnd = tp->snd_cwnd *
2607 tcp_mss_to_mtu(sk, tp->mss_cache) /
2608 icsk->icsk_mtup.probe_size;
2609 tp->snd_cwnd_cnt = 0;
2610 tp->snd_cwnd_stamp = tcp_time_stamp;
2611 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2613 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2614 icsk->icsk_mtup.probe_size = 0;
2615 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2618 /* Do a simple retransmit without using the backoff mechanisms in
2619 * tcp_timer. This is used for path mtu discovery.
2620 * The socket is already locked here.
2622 void tcp_simple_retransmit(struct sock *sk)
2624 const struct inet_connection_sock *icsk = inet_csk(sk);
2625 struct tcp_sock *tp = tcp_sk(sk);
2626 struct sk_buff *skb;
2627 unsigned int mss = tcp_current_mss(sk);
2628 u32 prior_lost = tp->lost_out;
2630 tcp_for_write_queue(skb, sk) {
2631 if (skb == tcp_send_head(sk))
2633 if (tcp_skb_seglen(skb) > mss &&
2634 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2635 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2636 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2637 tp->retrans_out -= tcp_skb_pcount(skb);
2639 tcp_skb_mark_lost_uncond_verify(tp, skb);
2643 tcp_clear_retrans_hints_partial(tp);
2645 if (prior_lost == tp->lost_out)
2648 if (tcp_is_reno(tp))
2649 tcp_limit_reno_sacked(tp);
2651 tcp_verify_left_out(tp);
2653 /* Don't muck with the congestion window here.
2654 * Reason is that we do not increase amount of _data_
2655 * in network, but units changed and effective
2656 * cwnd/ssthresh really reduced now.
2658 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2659 tp->high_seq = tp->snd_nxt;
2660 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2661 tp->prior_ssthresh = 0;
2662 tp->undo_marker = 0;
2663 tcp_set_ca_state(sk, TCP_CA_Loss);
2665 tcp_xmit_retransmit_queue(sk);
2667 EXPORT_SYMBOL(tcp_simple_retransmit);
2669 static void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2671 struct tcp_sock *tp = tcp_sk(sk);
2674 if (tcp_is_reno(tp))
2675 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2677 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2679 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2681 tp->prior_ssthresh = 0;
2684 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2686 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2687 tcp_init_cwnd_reduction(sk);
2689 tcp_set_ca_state(sk, TCP_CA_Recovery);
2692 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2693 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2695 static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack)
2697 struct tcp_sock *tp = tcp_sk(sk);
2698 bool recovered = !before(tp->snd_una, tp->high_seq);
2700 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2701 /* Step 3.b. A timeout is spurious if not all data are
2702 * lost, i.e., never-retransmitted data are (s)acked.
2704 if (tcp_try_undo_loss(sk, flag & FLAG_ORIG_SACK_ACKED))
2707 if (after(tp->snd_nxt, tp->high_seq) &&
2708 (flag & FLAG_DATA_SACKED || is_dupack)) {
2709 tp->frto = 0; /* Loss was real: 2nd part of step 3.a */
2710 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2711 tp->high_seq = tp->snd_nxt;
2712 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
2714 if (after(tp->snd_nxt, tp->high_seq))
2715 return; /* Step 2.b */
2721 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2722 tcp_try_undo_recovery(sk);
2725 if (tcp_is_reno(tp)) {
2726 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2727 * delivered. Lower inflight to clock out (re)tranmissions.
2729 if (after(tp->snd_nxt, tp->high_seq) && is_dupack)
2730 tcp_add_reno_sack(sk);
2731 else if (flag & FLAG_SND_UNA_ADVANCED)
2732 tcp_reset_reno_sack(tp);
2734 if (tcp_try_undo_loss(sk, false))
2736 tcp_xmit_retransmit_queue(sk);
2739 /* Undo during fast recovery after partial ACK. */
2740 static bool tcp_try_undo_partial(struct sock *sk, const int acked,
2741 const int prior_unsacked)
2743 struct tcp_sock *tp = tcp_sk(sk);
2745 if (tp->undo_marker && tcp_packet_delayed(tp)) {
2746 /* Plain luck! Hole if filled with delayed
2747 * packet, rather than with a retransmit.
2749 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2751 /* We are getting evidence that the reordering degree is higher
2752 * than we realized. If there are no retransmits out then we
2753 * can undo. Otherwise we clock out new packets but do not
2754 * mark more packets lost or retransmit more.
2756 if (tp->retrans_out) {
2757 tcp_cwnd_reduction(sk, prior_unsacked, 0);
2761 if (!tcp_any_retrans_done(sk))
2762 tp->retrans_stamp = 0;
2764 DBGUNDO(sk, "partial recovery");
2765 tcp_undo_cwnd_reduction(sk, true);
2766 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2767 tcp_try_keep_open(sk);
2773 /* Process an event, which can update packets-in-flight not trivially.
2774 * Main goal of this function is to calculate new estimate for left_out,
2775 * taking into account both packets sitting in receiver's buffer and
2776 * packets lost by network.
2778 * Besides that it does CWND reduction, when packet loss is detected
2779 * and changes state of machine.
2781 * It does _not_ decide what to send, it is made in function
2782 * tcp_xmit_retransmit_queue().
2784 static void tcp_fastretrans_alert(struct sock *sk, const int acked,
2785 const int prior_unsacked,
2786 bool is_dupack, int flag)
2788 struct inet_connection_sock *icsk = inet_csk(sk);
2789 struct tcp_sock *tp = tcp_sk(sk);
2790 bool do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2791 (tcp_fackets_out(tp) > tp->reordering));
2792 int fast_rexmit = 0;
2794 if (WARN_ON(!tp->packets_out && tp->sacked_out))
2796 if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2797 tp->fackets_out = 0;
2799 /* Now state machine starts.
2800 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2801 if (flag & FLAG_ECE)
2802 tp->prior_ssthresh = 0;
2804 /* B. In all the states check for reneging SACKs. */
2805 if (tcp_check_sack_reneging(sk, flag))
2808 /* C. Check consistency of the current state. */
2809 tcp_verify_left_out(tp);
2811 /* D. Check state exit conditions. State can be terminated
2812 * when high_seq is ACKed. */
2813 if (icsk->icsk_ca_state == TCP_CA_Open) {
2814 WARN_ON(tp->retrans_out != 0);
2815 tp->retrans_stamp = 0;
2816 } else if (!before(tp->snd_una, tp->high_seq)) {
2817 switch (icsk->icsk_ca_state) {
2819 /* CWR is to be held something *above* high_seq
2820 * is ACKed for CWR bit to reach receiver. */
2821 if (tp->snd_una != tp->high_seq) {
2822 tcp_end_cwnd_reduction(sk);
2823 tcp_set_ca_state(sk, TCP_CA_Open);
2827 case TCP_CA_Recovery:
2828 if (tcp_is_reno(tp))
2829 tcp_reset_reno_sack(tp);
2830 if (tcp_try_undo_recovery(sk))
2832 tcp_end_cwnd_reduction(sk);
2837 /* E. Process state. */
2838 switch (icsk->icsk_ca_state) {
2839 case TCP_CA_Recovery:
2840 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2841 if (tcp_is_reno(tp) && is_dupack)
2842 tcp_add_reno_sack(sk);
2844 if (tcp_try_undo_partial(sk, acked, prior_unsacked))
2846 /* Partial ACK arrived. Force fast retransmit. */
2847 do_lost = tcp_is_reno(tp) ||
2848 tcp_fackets_out(tp) > tp->reordering;
2850 if (tcp_try_undo_dsack(sk)) {
2851 tcp_try_keep_open(sk);
2856 tcp_process_loss(sk, flag, is_dupack);
2857 if (icsk->icsk_ca_state != TCP_CA_Open)
2859 /* Fall through to processing in Open state. */
2861 if (tcp_is_reno(tp)) {
2862 if (flag & FLAG_SND_UNA_ADVANCED)
2863 tcp_reset_reno_sack(tp);
2865 tcp_add_reno_sack(sk);
2868 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2869 tcp_try_undo_dsack(sk);
2871 if (!tcp_time_to_recover(sk, flag)) {
2872 tcp_try_to_open(sk, flag, prior_unsacked);
2876 /* MTU probe failure: don't reduce cwnd */
2877 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2878 icsk->icsk_mtup.probe_size &&
2879 tp->snd_una == tp->mtu_probe.probe_seq_start) {
2880 tcp_mtup_probe_failed(sk);
2881 /* Restores the reduction we did in tcp_mtup_probe() */
2883 tcp_simple_retransmit(sk);
2887 /* Otherwise enter Recovery state */
2888 tcp_enter_recovery(sk, (flag & FLAG_ECE));
2893 tcp_update_scoreboard(sk, fast_rexmit);
2894 tcp_cwnd_reduction(sk, prior_unsacked, fast_rexmit);
2895 tcp_xmit_retransmit_queue(sk);
2898 static inline bool tcp_ack_update_rtt(struct sock *sk, const int flag,
2899 long seq_rtt_us, long sack_rtt_us)
2901 const struct tcp_sock *tp = tcp_sk(sk);
2903 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2904 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2905 * Karn's algorithm forbids taking RTT if some retransmitted data
2906 * is acked (RFC6298).
2908 if (flag & FLAG_RETRANS_DATA_ACKED)
2912 seq_rtt_us = sack_rtt_us;
2914 /* RTTM Rule: A TSecr value received in a segment is used to
2915 * update the averaged RTT measurement only if the segment
2916 * acknowledges some new data, i.e., only if it advances the
2917 * left edge of the send window.
2918 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2920 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2922 seq_rtt_us = jiffies_to_usecs(tcp_time_stamp - tp->rx_opt.rcv_tsecr);
2927 tcp_rtt_estimator(sk, seq_rtt_us);
2930 /* RFC6298: only reset backoff on valid RTT measurement. */
2931 inet_csk(sk)->icsk_backoff = 0;
2935 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2936 static void tcp_synack_rtt_meas(struct sock *sk, const u32 synack_stamp)
2938 struct tcp_sock *tp = tcp_sk(sk);
2939 long seq_rtt_us = -1L;
2941 if (synack_stamp && !tp->total_retrans)
2942 seq_rtt_us = jiffies_to_usecs(tcp_time_stamp - synack_stamp);
2944 /* If the ACK acks both the SYNACK and the (Fast Open'd) data packets
2945 * sent in SYN_RECV, SYNACK RTT is the smooth RTT computed in tcp_ack()
2948 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, seq_rtt_us, -1L);
2951 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
2953 const struct inet_connection_sock *icsk = inet_csk(sk);
2955 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
2956 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
2959 /* Restart timer after forward progress on connection.
2960 * RFC2988 recommends to restart timer to now+rto.
2962 void tcp_rearm_rto(struct sock *sk)
2964 const struct inet_connection_sock *icsk = inet_csk(sk);
2965 struct tcp_sock *tp = tcp_sk(sk);
2967 /* If the retrans timer is currently being used by Fast Open
2968 * for SYN-ACK retrans purpose, stay put.
2970 if (tp->fastopen_rsk)
2973 if (!tp->packets_out) {
2974 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2976 u32 rto = inet_csk(sk)->icsk_rto;
2977 /* Offset the time elapsed after installing regular RTO */
2978 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
2979 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
2980 struct sk_buff *skb = tcp_write_queue_head(sk);
2981 const u32 rto_time_stamp =
2982 tcp_skb_timestamp(skb) + rto;
2983 s32 delta = (s32)(rto_time_stamp - tcp_time_stamp);
2984 /* delta may not be positive if the socket is locked
2985 * when the retrans timer fires and is rescheduled.
2990 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
2995 /* This function is called when the delayed ER timer fires. TCP enters
2996 * fast recovery and performs fast-retransmit.
2998 void tcp_resume_early_retransmit(struct sock *sk)
3000 struct tcp_sock *tp = tcp_sk(sk);
3004 /* Stop if ER is disabled after the delayed ER timer is scheduled */
3005 if (!tp->do_early_retrans)
3008 tcp_enter_recovery(sk, false);
3009 tcp_update_scoreboard(sk, 1);
3010 tcp_xmit_retransmit_queue(sk);
3013 /* If we get here, the whole TSO packet has not been acked. */
3014 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3016 struct tcp_sock *tp = tcp_sk(sk);
3019 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3021 packets_acked = tcp_skb_pcount(skb);
3022 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3024 packets_acked -= tcp_skb_pcount(skb);
3026 if (packets_acked) {
3027 BUG_ON(tcp_skb_pcount(skb) == 0);
3028 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3031 return packets_acked;
3034 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3037 const struct skb_shared_info *shinfo;
3039 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3040 if (likely(!(sk->sk_tsflags & SOF_TIMESTAMPING_TX_ACK)))
3043 shinfo = skb_shinfo(skb);
3044 if ((shinfo->tx_flags & SKBTX_ACK_TSTAMP) &&
3045 between(shinfo->tskey, prior_snd_una, tcp_sk(sk)->snd_una - 1))
3046 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK);
3049 /* Remove acknowledged frames from the retransmission queue. If our packet
3050 * is before the ack sequence we can discard it as it's confirmed to have
3051 * arrived at the other end.
3053 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3054 u32 prior_snd_una, long sack_rtt_us)
3056 const struct inet_connection_sock *icsk = inet_csk(sk);
3057 struct skb_mstamp first_ackt, last_ackt, now;
3058 struct tcp_sock *tp = tcp_sk(sk);
3059 u32 prior_sacked = tp->sacked_out;
3060 u32 reord = tp->packets_out;
3061 bool fully_acked = true;
3062 long ca_seq_rtt_us = -1L;
3063 long seq_rtt_us = -1L;
3064 struct sk_buff *skb;
3071 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3072 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3073 u8 sacked = scb->sacked;
3076 tcp_ack_tstamp(sk, skb, prior_snd_una);
3078 /* Determine how many packets and what bytes were acked, tso and else */
3079 if (after(scb->end_seq, tp->snd_una)) {
3080 if (tcp_skb_pcount(skb) == 1 ||
3081 !after(tp->snd_una, scb->seq))
3084 acked_pcount = tcp_tso_acked(sk, skb);
3088 fully_acked = false;
3090 /* Speedup tcp_unlink_write_queue() and next loop */
3091 prefetchw(skb->next);
3092 acked_pcount = tcp_skb_pcount(skb);
3095 if (unlikely(sacked & TCPCB_RETRANS)) {
3096 if (sacked & TCPCB_SACKED_RETRANS)
3097 tp->retrans_out -= acked_pcount;
3098 flag |= FLAG_RETRANS_DATA_ACKED;
3100 last_ackt = skb->skb_mstamp;
3101 WARN_ON_ONCE(last_ackt.v64 == 0);
3102 if (!first_ackt.v64)
3103 first_ackt = last_ackt;
3105 if (!(sacked & TCPCB_SACKED_ACKED))
3106 reord = min(pkts_acked, reord);
3107 if (!after(scb->end_seq, tp->high_seq))
3108 flag |= FLAG_ORIG_SACK_ACKED;
3111 if (sacked & TCPCB_SACKED_ACKED)
3112 tp->sacked_out -= acked_pcount;
3113 if (sacked & TCPCB_LOST)
3114 tp->lost_out -= acked_pcount;
3116 tp->packets_out -= acked_pcount;
3117 pkts_acked += acked_pcount;
3119 /* Initial outgoing SYN's get put onto the write_queue
3120 * just like anything else we transmit. It is not
3121 * true data, and if we misinform our callers that
3122 * this ACK acks real data, we will erroneously exit
3123 * connection startup slow start one packet too
3124 * quickly. This is severely frowned upon behavior.
3126 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3127 flag |= FLAG_DATA_ACKED;
3129 flag |= FLAG_SYN_ACKED;
3130 tp->retrans_stamp = 0;
3136 tcp_unlink_write_queue(skb, sk);
3137 sk_wmem_free_skb(sk, skb);
3138 if (unlikely(skb == tp->retransmit_skb_hint))
3139 tp->retransmit_skb_hint = NULL;
3140 if (unlikely(skb == tp->lost_skb_hint))
3141 tp->lost_skb_hint = NULL;
3144 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3145 tp->snd_up = tp->snd_una;
3147 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3148 flag |= FLAG_SACK_RENEGING;
3150 skb_mstamp_get(&now);
3151 if (likely(first_ackt.v64)) {
3152 seq_rtt_us = skb_mstamp_us_delta(&now, &first_ackt);
3153 ca_seq_rtt_us = skb_mstamp_us_delta(&now, &last_ackt);
3156 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us);
3158 if (flag & FLAG_ACKED) {
3159 const struct tcp_congestion_ops *ca_ops
3160 = inet_csk(sk)->icsk_ca_ops;
3163 if (unlikely(icsk->icsk_mtup.probe_size &&
3164 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3165 tcp_mtup_probe_success(sk);
3168 if (tcp_is_reno(tp)) {
3169 tcp_remove_reno_sacks(sk, pkts_acked);
3173 /* Non-retransmitted hole got filled? That's reordering */
3174 if (reord < prior_fackets)
3175 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3177 delta = tcp_is_fack(tp) ? pkts_acked :
3178 prior_sacked - tp->sacked_out;
3179 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3182 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3184 if (ca_ops->pkts_acked)
3185 ca_ops->pkts_acked(sk, pkts_acked, ca_seq_rtt_us);
3187 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3188 sack_rtt_us > skb_mstamp_us_delta(&now, &skb->skb_mstamp)) {
3189 /* Do not re-arm RTO if the sack RTT is measured from data sent
3190 * after when the head was last (re)transmitted. Otherwise the
3191 * timeout may continue to extend in loss recovery.
3196 #if FASTRETRANS_DEBUG > 0
3197 WARN_ON((int)tp->sacked_out < 0);
3198 WARN_ON((int)tp->lost_out < 0);
3199 WARN_ON((int)tp->retrans_out < 0);
3200 if (!tp->packets_out && tcp_is_sack(tp)) {
3201 icsk = inet_csk(sk);
3203 pr_debug("Leak l=%u %d\n",
3204 tp->lost_out, icsk->icsk_ca_state);
3207 if (tp->sacked_out) {
3208 pr_debug("Leak s=%u %d\n",
3209 tp->sacked_out, icsk->icsk_ca_state);
3212 if (tp->retrans_out) {
3213 pr_debug("Leak r=%u %d\n",
3214 tp->retrans_out, icsk->icsk_ca_state);
3215 tp->retrans_out = 0;
3222 static void tcp_ack_probe(struct sock *sk)
3224 const struct tcp_sock *tp = tcp_sk(sk);
3225 struct inet_connection_sock *icsk = inet_csk(sk);
3227 /* Was it a usable window open? */
3229 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3230 icsk->icsk_backoff = 0;
3231 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3232 /* Socket must be waked up by subsequent tcp_data_snd_check().
3233 * This function is not for random using!
3236 unsigned long when = inet_csk_rto_backoff(icsk, TCP_RTO_MAX);
3238 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3243 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3245 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3246 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3249 /* Decide wheather to run the increase function of congestion control. */
3250 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3252 if (tcp_in_cwnd_reduction(sk))
3255 /* If reordering is high then always grow cwnd whenever data is
3256 * delivered regardless of its ordering. Otherwise stay conservative
3257 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3258 * new SACK or ECE mark may first advance cwnd here and later reduce
3259 * cwnd in tcp_fastretrans_alert() based on more states.
3261 if (tcp_sk(sk)->reordering > sysctl_tcp_reordering)
3262 return flag & FLAG_FORWARD_PROGRESS;
3264 return flag & FLAG_DATA_ACKED;
3267 /* Check that window update is acceptable.
3268 * The function assumes that snd_una<=ack<=snd_next.
3270 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3271 const u32 ack, const u32 ack_seq,
3274 return after(ack, tp->snd_una) ||
3275 after(ack_seq, tp->snd_wl1) ||
3276 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3279 /* Update our send window.
3281 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3282 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3284 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3287 struct tcp_sock *tp = tcp_sk(sk);
3289 u32 nwin = ntohs(tcp_hdr(skb)->window);
3291 if (likely(!tcp_hdr(skb)->syn))
3292 nwin <<= tp->rx_opt.snd_wscale;
3294 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3295 flag |= FLAG_WIN_UPDATE;
3296 tcp_update_wl(tp, ack_seq);
3298 if (tp->snd_wnd != nwin) {
3301 /* Note, it is the only place, where
3302 * fast path is recovered for sending TCP.
3305 tcp_fast_path_check(sk);
3307 if (nwin > tp->max_window) {
3308 tp->max_window = nwin;
3309 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3319 /* RFC 5961 7 [ACK Throttling] */
3320 static void tcp_send_challenge_ack(struct sock *sk)
3322 /* unprotected vars, we dont care of overwrites */
3323 static u32 challenge_timestamp;
3324 static unsigned int challenge_count;
3325 u32 now = jiffies / HZ;
3327 if (now != challenge_timestamp) {
3328 challenge_timestamp = now;
3329 challenge_count = 0;
3331 if (++challenge_count <= sysctl_tcp_challenge_ack_limit) {
3332 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPCHALLENGEACK);
3337 static void tcp_store_ts_recent(struct tcp_sock *tp)
3339 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3340 tp->rx_opt.ts_recent_stamp = get_seconds();
3343 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3345 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3346 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3347 * extra check below makes sure this can only happen
3348 * for pure ACK frames. -DaveM
3350 * Not only, also it occurs for expired timestamps.
3353 if (tcp_paws_check(&tp->rx_opt, 0))
3354 tcp_store_ts_recent(tp);
3358 /* This routine deals with acks during a TLP episode.
3359 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3361 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3363 struct tcp_sock *tp = tcp_sk(sk);
3364 bool is_tlp_dupack = (ack == tp->tlp_high_seq) &&
3365 !(flag & (FLAG_SND_UNA_ADVANCED |
3366 FLAG_NOT_DUP | FLAG_DATA_SACKED));
3368 /* Mark the end of TLP episode on receiving TLP dupack or when
3369 * ack is after tlp_high_seq.
3371 if (is_tlp_dupack) {
3372 tp->tlp_high_seq = 0;
3376 if (after(ack, tp->tlp_high_seq)) {
3377 tp->tlp_high_seq = 0;
3378 /* Don't reduce cwnd if DSACK arrives for TLP retrans. */
3379 if (!(flag & FLAG_DSACKING_ACK)) {
3380 tcp_init_cwnd_reduction(sk);
3381 tcp_set_ca_state(sk, TCP_CA_CWR);
3382 tcp_end_cwnd_reduction(sk);
3383 tcp_try_keep_open(sk);
3384 NET_INC_STATS_BH(sock_net(sk),
3385 LINUX_MIB_TCPLOSSPROBERECOVERY);
3390 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3392 const struct inet_connection_sock *icsk = inet_csk(sk);
3394 if (icsk->icsk_ca_ops->in_ack_event)
3395 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3398 /* This routine deals with incoming acks, but not outgoing ones. */
3399 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3401 struct inet_connection_sock *icsk = inet_csk(sk);
3402 struct tcp_sock *tp = tcp_sk(sk);
3403 u32 prior_snd_una = tp->snd_una;
3404 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3405 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3406 bool is_dupack = false;
3408 int prior_packets = tp->packets_out;
3409 const int prior_unsacked = tp->packets_out - tp->sacked_out;
3410 int acked = 0; /* Number of packets newly acked */
3411 long sack_rtt_us = -1L;
3413 /* We very likely will need to access write queue head. */
3414 prefetchw(sk->sk_write_queue.next);
3416 /* If the ack is older than previous acks
3417 * then we can probably ignore it.
3419 if (before(ack, prior_snd_una)) {
3420 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3421 if (before(ack, prior_snd_una - tp->max_window)) {
3422 tcp_send_challenge_ack(sk);
3428 /* If the ack includes data we haven't sent yet, discard
3429 * this segment (RFC793 Section 3.9).
3431 if (after(ack, tp->snd_nxt))
3434 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
3435 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
3438 if (after(ack, prior_snd_una)) {
3439 flag |= FLAG_SND_UNA_ADVANCED;
3440 icsk->icsk_retransmits = 0;
3443 prior_fackets = tp->fackets_out;
3445 /* ts_recent update must be made after we are sure that the packet
3448 if (flag & FLAG_UPDATE_TS_RECENT)
3449 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3451 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3452 /* Window is constant, pure forward advance.
3453 * No more checks are required.
3454 * Note, we use the fact that SND.UNA>=SND.WL2.
3456 tcp_update_wl(tp, ack_seq);
3458 flag |= FLAG_WIN_UPDATE;
3460 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3462 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3464 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3466 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3469 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3471 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3473 if (TCP_SKB_CB(skb)->sacked)
3474 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3477 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3479 ack_ev_flags |= CA_ACK_ECE;
3482 if (flag & FLAG_WIN_UPDATE)
3483 ack_ev_flags |= CA_ACK_WIN_UPDATE;
3485 tcp_in_ack_event(sk, ack_ev_flags);
3488 /* We passed data and got it acked, remove any soft error
3489 * log. Something worked...
3491 sk->sk_err_soft = 0;
3492 icsk->icsk_probes_out = 0;
3493 tp->rcv_tstamp = tcp_time_stamp;
3497 /* See if we can take anything off of the retransmit queue. */
3498 acked = tp->packets_out;
3499 flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una,
3501 acked -= tp->packets_out;
3503 /* Advance cwnd if state allows */
3504 if (tcp_may_raise_cwnd(sk, flag))
3505 tcp_cong_avoid(sk, ack, acked);
3507 if (tcp_ack_is_dubious(sk, flag)) {
3508 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3509 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3512 if (tp->tlp_high_seq)
3513 tcp_process_tlp_ack(sk, ack, flag);
3515 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) {
3516 struct dst_entry *dst = __sk_dst_get(sk);
3521 if (icsk->icsk_pending == ICSK_TIME_RETRANS)
3522 tcp_schedule_loss_probe(sk);
3523 tcp_update_pacing_rate(sk);
3527 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3528 if (flag & FLAG_DSACKING_ACK)
3529 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3531 /* If this ack opens up a zero window, clear backoff. It was
3532 * being used to time the probes, and is probably far higher than
3533 * it needs to be for normal retransmission.
3535 if (tcp_send_head(sk))
3538 if (tp->tlp_high_seq)
3539 tcp_process_tlp_ack(sk, ack, flag);
3543 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3547 /* If data was SACKed, tag it and see if we should send more data.
3548 * If data was DSACKed, see if we can undo a cwnd reduction.
3550 if (TCP_SKB_CB(skb)->sacked) {
3551 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3553 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3557 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3561 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3562 * But, this can also be called on packets in the established flow when
3563 * the fast version below fails.
3565 void tcp_parse_options(const struct sk_buff *skb,
3566 struct tcp_options_received *opt_rx, int estab,
3567 struct tcp_fastopen_cookie *foc)
3569 const unsigned char *ptr;
3570 const struct tcphdr *th = tcp_hdr(skb);
3571 int length = (th->doff * 4) - sizeof(struct tcphdr);
3573 ptr = (const unsigned char *)(th + 1);
3574 opt_rx->saw_tstamp = 0;
3576 while (length > 0) {
3577 int opcode = *ptr++;
3583 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3588 if (opsize < 2) /* "silly options" */
3590 if (opsize > length)
3591 return; /* don't parse partial options */
3594 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3595 u16 in_mss = get_unaligned_be16(ptr);
3597 if (opt_rx->user_mss &&
3598 opt_rx->user_mss < in_mss)
3599 in_mss = opt_rx->user_mss;
3600 opt_rx->mss_clamp = in_mss;
3605 if (opsize == TCPOLEN_WINDOW && th->syn &&
3606 !estab && sysctl_tcp_window_scaling) {
3607 __u8 snd_wscale = *(__u8 *)ptr;
3608 opt_rx->wscale_ok = 1;
3609 if (snd_wscale > 14) {
3610 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3615 opt_rx->snd_wscale = snd_wscale;
3618 case TCPOPT_TIMESTAMP:
3619 if ((opsize == TCPOLEN_TIMESTAMP) &&
3620 ((estab && opt_rx->tstamp_ok) ||
3621 (!estab && sysctl_tcp_timestamps))) {
3622 opt_rx->saw_tstamp = 1;
3623 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3624 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3627 case TCPOPT_SACK_PERM:
3628 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3629 !estab && sysctl_tcp_sack) {
3630 opt_rx->sack_ok = TCP_SACK_SEEN;
3631 tcp_sack_reset(opt_rx);
3636 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3637 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3639 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3642 #ifdef CONFIG_TCP_MD5SIG
3645 * The MD5 Hash has already been
3646 * checked (see tcp_v{4,6}_do_rcv()).
3651 /* Fast Open option shares code 254 using a
3652 * 16 bits magic number. It's valid only in
3653 * SYN or SYN-ACK with an even size.
3655 if (opsize < TCPOLEN_EXP_FASTOPEN_BASE ||
3656 get_unaligned_be16(ptr) != TCPOPT_FASTOPEN_MAGIC ||
3657 foc == NULL || !th->syn || (opsize & 1))
3659 foc->len = opsize - TCPOLEN_EXP_FASTOPEN_BASE;
3660 if (foc->len >= TCP_FASTOPEN_COOKIE_MIN &&
3661 foc->len <= TCP_FASTOPEN_COOKIE_MAX)
3662 memcpy(foc->val, ptr + 2, foc->len);
3663 else if (foc->len != 0)
3673 EXPORT_SYMBOL(tcp_parse_options);
3675 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3677 const __be32 *ptr = (const __be32 *)(th + 1);
3679 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3680 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3681 tp->rx_opt.saw_tstamp = 1;
3683 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3686 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3688 tp->rx_opt.rcv_tsecr = 0;
3694 /* Fast parse options. This hopes to only see timestamps.
3695 * If it is wrong it falls back on tcp_parse_options().
3697 static bool tcp_fast_parse_options(const struct sk_buff *skb,
3698 const struct tcphdr *th, struct tcp_sock *tp)
3700 /* In the spirit of fast parsing, compare doff directly to constant
3701 * values. Because equality is used, short doff can be ignored here.
3703 if (th->doff == (sizeof(*th) / 4)) {
3704 tp->rx_opt.saw_tstamp = 0;
3706 } else if (tp->rx_opt.tstamp_ok &&
3707 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3708 if (tcp_parse_aligned_timestamp(tp, th))
3712 tcp_parse_options(skb, &tp->rx_opt, 1, NULL);
3713 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3714 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3719 #ifdef CONFIG_TCP_MD5SIG
3721 * Parse MD5 Signature option
3723 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3725 int length = (th->doff << 2) - sizeof(*th);
3726 const u8 *ptr = (const u8 *)(th + 1);
3728 /* If the TCP option is too short, we can short cut */
3729 if (length < TCPOLEN_MD5SIG)
3732 while (length > 0) {
3733 int opcode = *ptr++;
3744 if (opsize < 2 || opsize > length)
3746 if (opcode == TCPOPT_MD5SIG)
3747 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3754 EXPORT_SYMBOL(tcp_parse_md5sig_option);
3757 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3759 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3760 * it can pass through stack. So, the following predicate verifies that
3761 * this segment is not used for anything but congestion avoidance or
3762 * fast retransmit. Moreover, we even are able to eliminate most of such
3763 * second order effects, if we apply some small "replay" window (~RTO)
3764 * to timestamp space.
3766 * All these measures still do not guarantee that we reject wrapped ACKs
3767 * on networks with high bandwidth, when sequence space is recycled fastly,
3768 * but it guarantees that such events will be very rare and do not affect
3769 * connection seriously. This doesn't look nice, but alas, PAWS is really
3772 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3773 * states that events when retransmit arrives after original data are rare.
3774 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3775 * the biggest problem on large power networks even with minor reordering.
3776 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3777 * up to bandwidth of 18Gigabit/sec. 8) ]
3780 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3782 const struct tcp_sock *tp = tcp_sk(sk);
3783 const struct tcphdr *th = tcp_hdr(skb);
3784 u32 seq = TCP_SKB_CB(skb)->seq;
3785 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3787 return (/* 1. Pure ACK with correct sequence number. */
3788 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3790 /* 2. ... and duplicate ACK. */
3791 ack == tp->snd_una &&
3793 /* 3. ... and does not update window. */
3794 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3796 /* 4. ... and sits in replay window. */
3797 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3800 static inline bool tcp_paws_discard(const struct sock *sk,
3801 const struct sk_buff *skb)
3803 const struct tcp_sock *tp = tcp_sk(sk);
3805 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
3806 !tcp_disordered_ack(sk, skb);
3809 /* Check segment sequence number for validity.
3811 * Segment controls are considered valid, if the segment
3812 * fits to the window after truncation to the window. Acceptability
3813 * of data (and SYN, FIN, of course) is checked separately.
3814 * See tcp_data_queue(), for example.
3816 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3817 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3818 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3819 * (borrowed from freebsd)
3822 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
3824 return !before(end_seq, tp->rcv_wup) &&
3825 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3828 /* When we get a reset we do this. */
3829 void tcp_reset(struct sock *sk)
3831 /* We want the right error as BSD sees it (and indeed as we do). */
3832 switch (sk->sk_state) {
3834 sk->sk_err = ECONNREFUSED;
3836 case TCP_CLOSE_WAIT:
3842 sk->sk_err = ECONNRESET;
3844 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3847 if (!sock_flag(sk, SOCK_DEAD))
3848 sk->sk_error_report(sk);
3854 * Process the FIN bit. This now behaves as it is supposed to work
3855 * and the FIN takes effect when it is validly part of sequence
3856 * space. Not before when we get holes.
3858 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3859 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3862 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3863 * close and we go into CLOSING (and later onto TIME-WAIT)
3865 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3867 static void tcp_fin(struct sock *sk)
3869 struct tcp_sock *tp = tcp_sk(sk);
3870 const struct dst_entry *dst;
3872 inet_csk_schedule_ack(sk);
3874 sk->sk_shutdown |= RCV_SHUTDOWN;
3875 sock_set_flag(sk, SOCK_DONE);
3877 switch (sk->sk_state) {
3879 case TCP_ESTABLISHED:
3880 /* Move to CLOSE_WAIT */
3881 tcp_set_state(sk, TCP_CLOSE_WAIT);
3882 dst = __sk_dst_get(sk);
3883 if (!dst || !dst_metric(dst, RTAX_QUICKACK))
3884 inet_csk(sk)->icsk_ack.pingpong = 1;
3887 case TCP_CLOSE_WAIT:
3889 /* Received a retransmission of the FIN, do
3894 /* RFC793: Remain in the LAST-ACK state. */
3898 /* This case occurs when a simultaneous close
3899 * happens, we must ack the received FIN and
3900 * enter the CLOSING state.
3903 tcp_set_state(sk, TCP_CLOSING);
3906 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3908 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
3911 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3912 * cases we should never reach this piece of code.
3914 pr_err("%s: Impossible, sk->sk_state=%d\n",
3915 __func__, sk->sk_state);
3919 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3920 * Probably, we should reset in this case. For now drop them.
3922 __skb_queue_purge(&tp->out_of_order_queue);
3923 if (tcp_is_sack(tp))
3924 tcp_sack_reset(&tp->rx_opt);
3927 if (!sock_flag(sk, SOCK_DEAD)) {
3928 sk->sk_state_change(sk);
3930 /* Do not send POLL_HUP for half duplex close. */
3931 if (sk->sk_shutdown == SHUTDOWN_MASK ||
3932 sk->sk_state == TCP_CLOSE)
3933 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
3935 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
3939 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
3942 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
3943 if (before(seq, sp->start_seq))
3944 sp->start_seq = seq;
3945 if (after(end_seq, sp->end_seq))
3946 sp->end_seq = end_seq;
3952 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
3954 struct tcp_sock *tp = tcp_sk(sk);
3956 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3959 if (before(seq, tp->rcv_nxt))
3960 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
3962 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
3964 NET_INC_STATS_BH(sock_net(sk), mib_idx);
3966 tp->rx_opt.dsack = 1;
3967 tp->duplicate_sack[0].start_seq = seq;
3968 tp->duplicate_sack[0].end_seq = end_seq;
3972 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
3974 struct tcp_sock *tp = tcp_sk(sk);
3976 if (!tp->rx_opt.dsack)
3977 tcp_dsack_set(sk, seq, end_seq);
3979 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
3982 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
3984 struct tcp_sock *tp = tcp_sk(sk);
3986 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
3987 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3988 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
3989 tcp_enter_quickack_mode(sk);
3991 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3992 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3994 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
3995 end_seq = tp->rcv_nxt;
3996 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4003 /* These routines update the SACK block as out-of-order packets arrive or
4004 * in-order packets close up the sequence space.
4006 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4009 struct tcp_sack_block *sp = &tp->selective_acks[0];
4010 struct tcp_sack_block *swalk = sp + 1;
4012 /* See if the recent change to the first SACK eats into
4013 * or hits the sequence space of other SACK blocks, if so coalesce.
4015 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4016 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4019 /* Zap SWALK, by moving every further SACK up by one slot.
4020 * Decrease num_sacks.
4022 tp->rx_opt.num_sacks--;
4023 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4027 this_sack++, swalk++;
4031 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4033 struct tcp_sock *tp = tcp_sk(sk);
4034 struct tcp_sack_block *sp = &tp->selective_acks[0];
4035 int cur_sacks = tp->rx_opt.num_sacks;
4041 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4042 if (tcp_sack_extend(sp, seq, end_seq)) {
4043 /* Rotate this_sack to the first one. */
4044 for (; this_sack > 0; this_sack--, sp--)
4045 swap(*sp, *(sp - 1));
4047 tcp_sack_maybe_coalesce(tp);
4052 /* Could not find an adjacent existing SACK, build a new one,
4053 * put it at the front, and shift everyone else down. We
4054 * always know there is at least one SACK present already here.
4056 * If the sack array is full, forget about the last one.
4058 if (this_sack >= TCP_NUM_SACKS) {
4060 tp->rx_opt.num_sacks--;
4063 for (; this_sack > 0; this_sack--, sp--)
4067 /* Build the new head SACK, and we're done. */
4068 sp->start_seq = seq;
4069 sp->end_seq = end_seq;
4070 tp->rx_opt.num_sacks++;
4073 /* RCV.NXT advances, some SACKs should be eaten. */
4075 static void tcp_sack_remove(struct tcp_sock *tp)
4077 struct tcp_sack_block *sp = &tp->selective_acks[0];
4078 int num_sacks = tp->rx_opt.num_sacks;
4081 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4082 if (skb_queue_empty(&tp->out_of_order_queue)) {
4083 tp->rx_opt.num_sacks = 0;
4087 for (this_sack = 0; this_sack < num_sacks;) {
4088 /* Check if the start of the sack is covered by RCV.NXT. */
4089 if (!before(tp->rcv_nxt, sp->start_seq)) {
4092 /* RCV.NXT must cover all the block! */
4093 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4095 /* Zap this SACK, by moving forward any other SACKS. */
4096 for (i = this_sack+1; i < num_sacks; i++)
4097 tp->selective_acks[i-1] = tp->selective_acks[i];
4104 tp->rx_opt.num_sacks = num_sacks;
4108 * tcp_try_coalesce - try to merge skb to prior one
4111 * @from: buffer to add in queue
4112 * @fragstolen: pointer to boolean
4114 * Before queueing skb @from after @to, try to merge them
4115 * to reduce overall memory use and queue lengths, if cost is small.
4116 * Packets in ofo or receive queues can stay a long time.
4117 * Better try to coalesce them right now to avoid future collapses.
4118 * Returns true if caller should free @from instead of queueing it
4120 static bool tcp_try_coalesce(struct sock *sk,
4122 struct sk_buff *from,
4127 *fragstolen = false;
4129 /* Its possible this segment overlaps with prior segment in queue */
4130 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4133 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4136 atomic_add(delta, &sk->sk_rmem_alloc);
4137 sk_mem_charge(sk, delta);
4138 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4139 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4140 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4141 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4145 /* This one checks to see if we can put data from the
4146 * out_of_order queue into the receive_queue.
4148 static void tcp_ofo_queue(struct sock *sk)
4150 struct tcp_sock *tp = tcp_sk(sk);
4151 __u32 dsack_high = tp->rcv_nxt;
4152 struct sk_buff *skb, *tail;
4153 bool fragstolen, eaten;
4155 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4156 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4159 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4160 __u32 dsack = dsack_high;
4161 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4162 dsack_high = TCP_SKB_CB(skb)->end_seq;
4163 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4166 __skb_unlink(skb, &tp->out_of_order_queue);
4167 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4168 SOCK_DEBUG(sk, "ofo packet was already received\n");
4172 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4173 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4174 TCP_SKB_CB(skb)->end_seq);
4176 tail = skb_peek_tail(&sk->sk_receive_queue);
4177 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4178 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4180 __skb_queue_tail(&sk->sk_receive_queue, skb);
4181 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4184 kfree_skb_partial(skb, fragstolen);
4188 static bool tcp_prune_ofo_queue(struct sock *sk);
4189 static int tcp_prune_queue(struct sock *sk);
4191 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4194 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4195 !sk_rmem_schedule(sk, skb, size)) {
4197 if (tcp_prune_queue(sk) < 0)
4200 if (!sk_rmem_schedule(sk, skb, size)) {
4201 if (!tcp_prune_ofo_queue(sk))
4204 if (!sk_rmem_schedule(sk, skb, size))
4211 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4213 struct tcp_sock *tp = tcp_sk(sk);
4214 struct sk_buff *skb1;
4217 tcp_ecn_check_ce(tp, skb);
4219 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4220 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFODROP);
4225 /* Disable header prediction. */
4227 inet_csk_schedule_ack(sk);
4229 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4230 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4231 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4233 skb1 = skb_peek_tail(&tp->out_of_order_queue);
4235 /* Initial out of order segment, build 1 SACK. */
4236 if (tcp_is_sack(tp)) {
4237 tp->rx_opt.num_sacks = 1;
4238 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4239 tp->selective_acks[0].end_seq =
4240 TCP_SKB_CB(skb)->end_seq;
4242 __skb_queue_head(&tp->out_of_order_queue, skb);
4246 seq = TCP_SKB_CB(skb)->seq;
4247 end_seq = TCP_SKB_CB(skb)->end_seq;
4249 if (seq == TCP_SKB_CB(skb1)->end_seq) {
4252 if (!tcp_try_coalesce(sk, skb1, skb, &fragstolen)) {
4253 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4255 tcp_grow_window(sk, skb);
4256 kfree_skb_partial(skb, fragstolen);
4260 if (!tp->rx_opt.num_sacks ||
4261 tp->selective_acks[0].end_seq != seq)
4264 /* Common case: data arrive in order after hole. */
4265 tp->selective_acks[0].end_seq = end_seq;
4269 /* Find place to insert this segment. */
4271 if (!after(TCP_SKB_CB(skb1)->seq, seq))
4273 if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
4277 skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
4280 /* Do skb overlap to previous one? */
4281 if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4282 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4283 /* All the bits are present. Drop. */
4284 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4287 tcp_dsack_set(sk, seq, end_seq);
4290 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4291 /* Partial overlap. */
4292 tcp_dsack_set(sk, seq,
4293 TCP_SKB_CB(skb1)->end_seq);
4295 if (skb_queue_is_first(&tp->out_of_order_queue,
4299 skb1 = skb_queue_prev(
4300 &tp->out_of_order_queue,
4305 __skb_queue_head(&tp->out_of_order_queue, skb);
4307 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4309 /* And clean segments covered by new one as whole. */
4310 while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
4311 skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
4313 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4315 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4316 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4320 __skb_unlink(skb1, &tp->out_of_order_queue);
4321 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4322 TCP_SKB_CB(skb1)->end_seq);
4323 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4328 if (tcp_is_sack(tp))
4329 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4332 tcp_grow_window(sk, skb);
4333 skb_set_owner_r(skb, sk);
4337 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
4341 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4343 __skb_pull(skb, hdrlen);
4345 tcp_try_coalesce(sk, tail, skb, fragstolen)) ? 1 : 0;
4346 tcp_sk(sk)->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4348 __skb_queue_tail(&sk->sk_receive_queue, skb);
4349 skb_set_owner_r(skb, sk);
4354 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4356 struct sk_buff *skb;
4362 skb = alloc_skb(size, sk->sk_allocation);
4366 if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4369 if (memcpy_fromiovec(skb_put(skb, size), msg->msg_iov, size))
4372 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4373 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4374 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4376 if (tcp_queue_rcv(sk, skb, 0, &fragstolen)) {
4377 WARN_ON_ONCE(fragstolen); /* should not happen */
4388 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4390 struct tcp_sock *tp = tcp_sk(sk);
4392 bool fragstolen = false;
4394 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4398 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
4400 tcp_ecn_accept_cwr(tp, skb);
4402 tp->rx_opt.dsack = 0;
4404 /* Queue data for delivery to the user.
4405 * Packets in sequence go to the receive queue.
4406 * Out of sequence packets to the out_of_order_queue.
4408 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4409 if (tcp_receive_window(tp) == 0)
4412 /* Ok. In sequence. In window. */
4413 if (tp->ucopy.task == current &&
4414 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4415 sock_owned_by_user(sk) && !tp->urg_data) {
4416 int chunk = min_t(unsigned int, skb->len,
4419 __set_current_state(TASK_RUNNING);
4422 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
4423 tp->ucopy.len -= chunk;
4424 tp->copied_seq += chunk;
4425 eaten = (chunk == skb->len);
4426 tcp_rcv_space_adjust(sk);
4434 tcp_try_rmem_schedule(sk, skb, skb->truesize))
4437 eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
4439 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4441 tcp_event_data_recv(sk, skb);
4442 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4445 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4448 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4449 * gap in queue is filled.
4451 if (skb_queue_empty(&tp->out_of_order_queue))
4452 inet_csk(sk)->icsk_ack.pingpong = 0;
4455 if (tp->rx_opt.num_sacks)
4456 tcp_sack_remove(tp);
4458 tcp_fast_path_check(sk);
4461 kfree_skb_partial(skb, fragstolen);
4462 if (!sock_flag(sk, SOCK_DEAD))
4463 sk->sk_data_ready(sk);
4467 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4468 /* A retransmit, 2nd most common case. Force an immediate ack. */
4469 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4470 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4473 tcp_enter_quickack_mode(sk);
4474 inet_csk_schedule_ack(sk);
4480 /* Out of window. F.e. zero window probe. */
4481 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4484 tcp_enter_quickack_mode(sk);
4486 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4487 /* Partial packet, seq < rcv_next < end_seq */
4488 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4489 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4490 TCP_SKB_CB(skb)->end_seq);
4492 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4494 /* If window is closed, drop tail of packet. But after
4495 * remembering D-SACK for its head made in previous line.
4497 if (!tcp_receive_window(tp))
4502 tcp_data_queue_ofo(sk, skb);
4505 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4506 struct sk_buff_head *list)
4508 struct sk_buff *next = NULL;
4510 if (!skb_queue_is_last(list, skb))
4511 next = skb_queue_next(list, skb);
4513 __skb_unlink(skb, list);
4515 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4520 /* Collapse contiguous sequence of skbs head..tail with
4521 * sequence numbers start..end.
4523 * If tail is NULL, this means until the end of the list.
4525 * Segments with FIN/SYN are not collapsed (only because this
4529 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4530 struct sk_buff *head, struct sk_buff *tail,
4533 struct sk_buff *skb, *n;
4536 /* First, check that queue is collapsible and find
4537 * the point where collapsing can be useful. */
4541 skb_queue_walk_from_safe(list, skb, n) {
4544 /* No new bits? It is possible on ofo queue. */
4545 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4546 skb = tcp_collapse_one(sk, skb, list);
4552 /* The first skb to collapse is:
4554 * - bloated or contains data before "start" or
4555 * overlaps to the next one.
4557 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
4558 (tcp_win_from_space(skb->truesize) > skb->len ||
4559 before(TCP_SKB_CB(skb)->seq, start))) {
4560 end_of_skbs = false;
4564 if (!skb_queue_is_last(list, skb)) {
4565 struct sk_buff *next = skb_queue_next(list, skb);
4567 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
4568 end_of_skbs = false;
4573 /* Decided to skip this, advance start seq. */
4574 start = TCP_SKB_CB(skb)->end_seq;
4577 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4580 while (before(start, end)) {
4581 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
4582 struct sk_buff *nskb;
4584 nskb = alloc_skb(copy, GFP_ATOMIC);
4588 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4589 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4590 __skb_queue_before(list, skb, nskb);
4591 skb_set_owner_r(nskb, sk);
4593 /* Copy data, releasing collapsed skbs. */
4595 int offset = start - TCP_SKB_CB(skb)->seq;
4596 int size = TCP_SKB_CB(skb)->end_seq - start;
4600 size = min(copy, size);
4601 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4603 TCP_SKB_CB(nskb)->end_seq += size;
4607 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4608 skb = tcp_collapse_one(sk, skb, list);
4611 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4618 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4619 * and tcp_collapse() them until all the queue is collapsed.
4621 static void tcp_collapse_ofo_queue(struct sock *sk)
4623 struct tcp_sock *tp = tcp_sk(sk);
4624 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4625 struct sk_buff *head;
4631 start = TCP_SKB_CB(skb)->seq;
4632 end = TCP_SKB_CB(skb)->end_seq;
4636 struct sk_buff *next = NULL;
4638 if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
4639 next = skb_queue_next(&tp->out_of_order_queue, skb);
4642 /* Segment is terminated when we see gap or when
4643 * we are at the end of all the queue. */
4645 after(TCP_SKB_CB(skb)->seq, end) ||
4646 before(TCP_SKB_CB(skb)->end_seq, start)) {
4647 tcp_collapse(sk, &tp->out_of_order_queue,
4648 head, skb, start, end);
4652 /* Start new segment */
4653 start = TCP_SKB_CB(skb)->seq;
4654 end = TCP_SKB_CB(skb)->end_seq;
4656 if (before(TCP_SKB_CB(skb)->seq, start))
4657 start = TCP_SKB_CB(skb)->seq;
4658 if (after(TCP_SKB_CB(skb)->end_seq, end))
4659 end = TCP_SKB_CB(skb)->end_seq;
4665 * Purge the out-of-order queue.
4666 * Return true if queue was pruned.
4668 static bool tcp_prune_ofo_queue(struct sock *sk)
4670 struct tcp_sock *tp = tcp_sk(sk);
4673 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4674 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
4675 __skb_queue_purge(&tp->out_of_order_queue);
4677 /* Reset SACK state. A conforming SACK implementation will
4678 * do the same at a timeout based retransmit. When a connection
4679 * is in a sad state like this, we care only about integrity
4680 * of the connection not performance.
4682 if (tp->rx_opt.sack_ok)
4683 tcp_sack_reset(&tp->rx_opt);
4690 /* Reduce allocated memory if we can, trying to get
4691 * the socket within its memory limits again.
4693 * Return less than zero if we should start dropping frames
4694 * until the socket owning process reads some of the data
4695 * to stabilize the situation.
4697 static int tcp_prune_queue(struct sock *sk)
4699 struct tcp_sock *tp = tcp_sk(sk);
4701 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4703 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
4705 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4706 tcp_clamp_window(sk);
4707 else if (sk_under_memory_pressure(sk))
4708 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4710 tcp_collapse_ofo_queue(sk);
4711 if (!skb_queue_empty(&sk->sk_receive_queue))
4712 tcp_collapse(sk, &sk->sk_receive_queue,
4713 skb_peek(&sk->sk_receive_queue),
4715 tp->copied_seq, tp->rcv_nxt);
4718 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4721 /* Collapsing did not help, destructive actions follow.
4722 * This must not ever occur. */
4724 tcp_prune_ofo_queue(sk);
4726 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4729 /* If we are really being abused, tell the caller to silently
4730 * drop receive data on the floor. It will get retransmitted
4731 * and hopefully then we'll have sufficient space.
4733 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
4735 /* Massive buffer overcommit. */
4740 static bool tcp_should_expand_sndbuf(const struct sock *sk)
4742 const struct tcp_sock *tp = tcp_sk(sk);
4744 /* If the user specified a specific send buffer setting, do
4747 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4750 /* If we are under global TCP memory pressure, do not expand. */
4751 if (sk_under_memory_pressure(sk))
4754 /* If we are under soft global TCP memory pressure, do not expand. */
4755 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
4758 /* If we filled the congestion window, do not expand. */
4759 if (tp->packets_out >= tp->snd_cwnd)
4765 /* When incoming ACK allowed to free some skb from write_queue,
4766 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4767 * on the exit from tcp input handler.
4769 * PROBLEM: sndbuf expansion does not work well with largesend.
4771 static void tcp_new_space(struct sock *sk)
4773 struct tcp_sock *tp = tcp_sk(sk);
4775 if (tcp_should_expand_sndbuf(sk)) {
4776 tcp_sndbuf_expand(sk);
4777 tp->snd_cwnd_stamp = tcp_time_stamp;
4780 sk->sk_write_space(sk);
4783 static void tcp_check_space(struct sock *sk)
4785 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4786 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4787 if (sk->sk_socket &&
4788 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4793 static inline void tcp_data_snd_check(struct sock *sk)
4795 tcp_push_pending_frames(sk);
4796 tcp_check_space(sk);
4800 * Check if sending an ack is needed.
4802 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4804 struct tcp_sock *tp = tcp_sk(sk);
4806 /* More than one full frame received... */
4807 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
4808 /* ... and right edge of window advances far enough.
4809 * (tcp_recvmsg() will send ACK otherwise). Or...
4811 __tcp_select_window(sk) >= tp->rcv_wnd) ||
4812 /* We ACK each frame or... */
4813 tcp_in_quickack_mode(sk) ||
4814 /* We have out of order data. */
4815 (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
4816 /* Then ack it now */
4819 /* Else, send delayed ack. */
4820 tcp_send_delayed_ack(sk);
4824 static inline void tcp_ack_snd_check(struct sock *sk)
4826 if (!inet_csk_ack_scheduled(sk)) {
4827 /* We sent a data segment already. */
4830 __tcp_ack_snd_check(sk, 1);
4834 * This routine is only called when we have urgent data
4835 * signaled. Its the 'slow' part of tcp_urg. It could be
4836 * moved inline now as tcp_urg is only called from one
4837 * place. We handle URGent data wrong. We have to - as
4838 * BSD still doesn't use the correction from RFC961.
4839 * For 1003.1g we should support a new option TCP_STDURG to permit
4840 * either form (or just set the sysctl tcp_stdurg).
4843 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
4845 struct tcp_sock *tp = tcp_sk(sk);
4846 u32 ptr = ntohs(th->urg_ptr);
4848 if (ptr && !sysctl_tcp_stdurg)
4850 ptr += ntohl(th->seq);
4852 /* Ignore urgent data that we've already seen and read. */
4853 if (after(tp->copied_seq, ptr))
4856 /* Do not replay urg ptr.
4858 * NOTE: interesting situation not covered by specs.
4859 * Misbehaving sender may send urg ptr, pointing to segment,
4860 * which we already have in ofo queue. We are not able to fetch
4861 * such data and will stay in TCP_URG_NOTYET until will be eaten
4862 * by recvmsg(). Seems, we are not obliged to handle such wicked
4863 * situations. But it is worth to think about possibility of some
4864 * DoSes using some hypothetical application level deadlock.
4866 if (before(ptr, tp->rcv_nxt))
4869 /* Do we already have a newer (or duplicate) urgent pointer? */
4870 if (tp->urg_data && !after(ptr, tp->urg_seq))
4873 /* Tell the world about our new urgent pointer. */
4876 /* We may be adding urgent data when the last byte read was
4877 * urgent. To do this requires some care. We cannot just ignore
4878 * tp->copied_seq since we would read the last urgent byte again
4879 * as data, nor can we alter copied_seq until this data arrives
4880 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4882 * NOTE. Double Dutch. Rendering to plain English: author of comment
4883 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4884 * and expect that both A and B disappear from stream. This is _wrong_.
4885 * Though this happens in BSD with high probability, this is occasional.
4886 * Any application relying on this is buggy. Note also, that fix "works"
4887 * only in this artificial test. Insert some normal data between A and B and we will
4888 * decline of BSD again. Verdict: it is better to remove to trap
4891 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
4892 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
4893 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4895 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
4896 __skb_unlink(skb, &sk->sk_receive_queue);
4901 tp->urg_data = TCP_URG_NOTYET;
4904 /* Disable header prediction. */
4908 /* This is the 'fast' part of urgent handling. */
4909 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
4911 struct tcp_sock *tp = tcp_sk(sk);
4913 /* Check if we get a new urgent pointer - normally not. */
4915 tcp_check_urg(sk, th);
4917 /* Do we wait for any urgent data? - normally not... */
4918 if (tp->urg_data == TCP_URG_NOTYET) {
4919 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
4922 /* Is the urgent pointer pointing into this packet? */
4923 if (ptr < skb->len) {
4925 if (skb_copy_bits(skb, ptr, &tmp, 1))
4927 tp->urg_data = TCP_URG_VALID | tmp;
4928 if (!sock_flag(sk, SOCK_DEAD))
4929 sk->sk_data_ready(sk);
4934 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4936 struct tcp_sock *tp = tcp_sk(sk);
4937 int chunk = skb->len - hlen;
4941 if (skb_csum_unnecessary(skb))
4942 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
4944 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
4948 tp->ucopy.len -= chunk;
4949 tp->copied_seq += chunk;
4950 tcp_rcv_space_adjust(sk);
4957 static __sum16 __tcp_checksum_complete_user(struct sock *sk,
4958 struct sk_buff *skb)
4962 if (sock_owned_by_user(sk)) {
4964 result = __tcp_checksum_complete(skb);
4967 result = __tcp_checksum_complete(skb);
4972 static inline bool tcp_checksum_complete_user(struct sock *sk,
4973 struct sk_buff *skb)
4975 return !skb_csum_unnecessary(skb) &&
4976 __tcp_checksum_complete_user(sk, skb);
4979 /* Does PAWS and seqno based validation of an incoming segment, flags will
4980 * play significant role here.
4982 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
4983 const struct tcphdr *th, int syn_inerr)
4985 struct tcp_sock *tp = tcp_sk(sk);
4987 /* RFC1323: H1. Apply PAWS check first. */
4988 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4989 tcp_paws_discard(sk, skb)) {
4991 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
4992 tcp_send_dupack(sk, skb);
4995 /* Reset is accepted even if it did not pass PAWS. */
4998 /* Step 1: check sequence number */
4999 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5000 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5001 * (RST) segments are validated by checking their SEQ-fields."
5002 * And page 69: "If an incoming segment is not acceptable,
5003 * an acknowledgment should be sent in reply (unless the RST
5004 * bit is set, if so drop the segment and return)".
5009 tcp_send_dupack(sk, skb);
5014 /* Step 2: check RST bit */
5017 * If sequence number exactly matches RCV.NXT, then
5018 * RESET the connection
5020 * Send a challenge ACK
5022 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt)
5025 tcp_send_challenge_ack(sk);
5029 /* step 3: check security and precedence [ignored] */
5031 /* step 4: Check for a SYN
5032 * RFC 5691 4.2 : Send a challenge ack
5037 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5038 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5039 tcp_send_challenge_ack(sk);
5051 * TCP receive function for the ESTABLISHED state.
5053 * It is split into a fast path and a slow path. The fast path is
5055 * - A zero window was announced from us - zero window probing
5056 * is only handled properly in the slow path.
5057 * - Out of order segments arrived.
5058 * - Urgent data is expected.
5059 * - There is no buffer space left
5060 * - Unexpected TCP flags/window values/header lengths are received
5061 * (detected by checking the TCP header against pred_flags)
5062 * - Data is sent in both directions. Fast path only supports pure senders
5063 * or pure receivers (this means either the sequence number or the ack
5064 * value must stay constant)
5065 * - Unexpected TCP option.
5067 * When these conditions are not satisfied it drops into a standard
5068 * receive procedure patterned after RFC793 to handle all cases.
5069 * The first three cases are guaranteed by proper pred_flags setting,
5070 * the rest is checked inline. Fast processing is turned on in
5071 * tcp_data_queue when everything is OK.
5073 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5074 const struct tcphdr *th, unsigned int len)
5076 struct tcp_sock *tp = tcp_sk(sk);
5078 if (unlikely(sk->sk_rx_dst == NULL))
5079 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5081 * Header prediction.
5082 * The code loosely follows the one in the famous
5083 * "30 instruction TCP receive" Van Jacobson mail.
5085 * Van's trick is to deposit buffers into socket queue
5086 * on a device interrupt, to call tcp_recv function
5087 * on the receive process context and checksum and copy
5088 * the buffer to user space. smart...
5090 * Our current scheme is not silly either but we take the
5091 * extra cost of the net_bh soft interrupt processing...
5092 * We do checksum and copy also but from device to kernel.
5095 tp->rx_opt.saw_tstamp = 0;
5097 /* pred_flags is 0xS?10 << 16 + snd_wnd
5098 * if header_prediction is to be made
5099 * 'S' will always be tp->tcp_header_len >> 2
5100 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5101 * turn it off (when there are holes in the receive
5102 * space for instance)
5103 * PSH flag is ignored.
5106 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5107 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5108 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5109 int tcp_header_len = tp->tcp_header_len;
5111 /* Timestamp header prediction: tcp_header_len
5112 * is automatically equal to th->doff*4 due to pred_flags
5116 /* Check timestamp */
5117 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5118 /* No? Slow path! */
5119 if (!tcp_parse_aligned_timestamp(tp, th))
5122 /* If PAWS failed, check it more carefully in slow path */
5123 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5126 /* DO NOT update ts_recent here, if checksum fails
5127 * and timestamp was corrupted part, it will result
5128 * in a hung connection since we will drop all
5129 * future packets due to the PAWS test.
5133 if (len <= tcp_header_len) {
5134 /* Bulk data transfer: sender */
5135 if (len == tcp_header_len) {
5136 /* Predicted packet is in window by definition.
5137 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5138 * Hence, check seq<=rcv_wup reduces to:
5140 if (tcp_header_len ==
5141 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5142 tp->rcv_nxt == tp->rcv_wup)
5143 tcp_store_ts_recent(tp);
5145 /* We know that such packets are checksummed
5148 tcp_ack(sk, skb, 0);
5150 tcp_data_snd_check(sk);
5152 } else { /* Header too small */
5153 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5158 bool fragstolen = false;
5160 if (tp->ucopy.task == current &&
5161 tp->copied_seq == tp->rcv_nxt &&
5162 len - tcp_header_len <= tp->ucopy.len &&
5163 sock_owned_by_user(sk)) {
5164 __set_current_state(TASK_RUNNING);
5166 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) {
5167 /* Predicted packet is in window by definition.
5168 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5169 * Hence, check seq<=rcv_wup reduces to:
5171 if (tcp_header_len ==
5172 (sizeof(struct tcphdr) +
5173 TCPOLEN_TSTAMP_ALIGNED) &&
5174 tp->rcv_nxt == tp->rcv_wup)
5175 tcp_store_ts_recent(tp);
5177 tcp_rcv_rtt_measure_ts(sk, skb);
5179 __skb_pull(skb, tcp_header_len);
5180 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5181 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
5186 if (tcp_checksum_complete_user(sk, skb))
5189 if ((int)skb->truesize > sk->sk_forward_alloc)
5192 /* Predicted packet is in window by definition.
5193 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5194 * Hence, check seq<=rcv_wup reduces to:
5196 if (tcp_header_len ==
5197 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5198 tp->rcv_nxt == tp->rcv_wup)
5199 tcp_store_ts_recent(tp);
5201 tcp_rcv_rtt_measure_ts(sk, skb);
5203 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
5205 /* Bulk data transfer: receiver */
5206 eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5210 tcp_event_data_recv(sk, skb);
5212 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5213 /* Well, only one small jumplet in fast path... */
5214 tcp_ack(sk, skb, FLAG_DATA);
5215 tcp_data_snd_check(sk);
5216 if (!inet_csk_ack_scheduled(sk))
5220 __tcp_ack_snd_check(sk, 0);
5223 kfree_skb_partial(skb, fragstolen);
5224 sk->sk_data_ready(sk);
5230 if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
5233 if (!th->ack && !th->rst)
5237 * Standard slow path.
5240 if (!tcp_validate_incoming(sk, skb, th, 1))
5244 if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5247 tcp_rcv_rtt_measure_ts(sk, skb);
5249 /* Process urgent data. */
5250 tcp_urg(sk, skb, th);
5252 /* step 7: process the segment text */
5253 tcp_data_queue(sk, skb);
5255 tcp_data_snd_check(sk);
5256 tcp_ack_snd_check(sk);
5260 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_CSUMERRORS);
5261 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5266 EXPORT_SYMBOL(tcp_rcv_established);
5268 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5270 struct tcp_sock *tp = tcp_sk(sk);
5271 struct inet_connection_sock *icsk = inet_csk(sk);
5273 tcp_set_state(sk, TCP_ESTABLISHED);
5276 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5277 security_inet_conn_established(sk, skb);
5280 /* Make sure socket is routed, for correct metrics. */
5281 icsk->icsk_af_ops->rebuild_header(sk);
5283 tcp_init_metrics(sk);
5285 tcp_init_congestion_control(sk);
5287 /* Prevent spurious tcp_cwnd_restart() on first data
5290 tp->lsndtime = tcp_time_stamp;
5292 tcp_init_buffer_space(sk);
5294 if (sock_flag(sk, SOCK_KEEPOPEN))
5295 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5297 if (!tp->rx_opt.snd_wscale)
5298 __tcp_fast_path_on(tp, tp->snd_wnd);
5302 if (!sock_flag(sk, SOCK_DEAD)) {
5303 sk->sk_state_change(sk);
5304 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5308 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5309 struct tcp_fastopen_cookie *cookie)
5311 struct tcp_sock *tp = tcp_sk(sk);
5312 struct sk_buff *data = tp->syn_data ? tcp_write_queue_head(sk) : NULL;
5313 u16 mss = tp->rx_opt.mss_clamp;
5316 if (mss == tp->rx_opt.user_mss) {
5317 struct tcp_options_received opt;
5319 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5320 tcp_clear_options(&opt);
5321 opt.user_mss = opt.mss_clamp = 0;
5322 tcp_parse_options(synack, &opt, 0, NULL);
5323 mss = opt.mss_clamp;
5326 if (!tp->syn_fastopen) /* Ignore an unsolicited cookie */
5329 /* The SYN-ACK neither has cookie nor acknowledges the data. Presumably
5330 * the remote receives only the retransmitted (regular) SYNs: either
5331 * the original SYN-data or the corresponding SYN-ACK is lost.
5333 syn_drop = (cookie->len <= 0 && data && tp->total_retrans);
5335 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop);
5337 if (data) { /* Retransmit unacked data in SYN */
5338 tcp_for_write_queue_from(data, sk) {
5339 if (data == tcp_send_head(sk) ||
5340 __tcp_retransmit_skb(sk, data))
5344 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVEFAIL);
5347 tp->syn_data_acked = tp->syn_data;
5348 if (tp->syn_data_acked)
5349 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
5353 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5354 const struct tcphdr *th, unsigned int len)
5356 struct inet_connection_sock *icsk = inet_csk(sk);
5357 struct tcp_sock *tp = tcp_sk(sk);
5358 struct tcp_fastopen_cookie foc = { .len = -1 };
5359 int saved_clamp = tp->rx_opt.mss_clamp;
5361 tcp_parse_options(skb, &tp->rx_opt, 0, &foc);
5362 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
5363 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
5367 * "If the state is SYN-SENT then
5368 * first check the ACK bit
5369 * If the ACK bit is set
5370 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5371 * a reset (unless the RST bit is set, if so drop
5372 * the segment and return)"
5374 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5375 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
5376 goto reset_and_undo;
5378 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5379 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5381 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
5382 goto reset_and_undo;
5385 /* Now ACK is acceptable.
5387 * "If the RST bit is set
5388 * If the ACK was acceptable then signal the user "error:
5389 * connection reset", drop the segment, enter CLOSED state,
5390 * delete TCB, and return."
5399 * "fifth, if neither of the SYN or RST bits is set then
5400 * drop the segment and return."
5406 goto discard_and_undo;
5409 * "If the SYN bit is on ...
5410 * are acceptable then ...
5411 * (our SYN has been ACKed), change the connection
5412 * state to ESTABLISHED..."
5415 tcp_ecn_rcv_synack(tp, th);
5417 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5418 tcp_ack(sk, skb, FLAG_SLOWPATH);
5420 /* Ok.. it's good. Set up sequence numbers and
5421 * move to established.
5423 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5424 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5426 /* RFC1323: The window in SYN & SYN/ACK segments is
5429 tp->snd_wnd = ntohs(th->window);
5431 if (!tp->rx_opt.wscale_ok) {
5432 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5433 tp->window_clamp = min(tp->window_clamp, 65535U);
5436 if (tp->rx_opt.saw_tstamp) {
5437 tp->rx_opt.tstamp_ok = 1;
5438 tp->tcp_header_len =
5439 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5440 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5441 tcp_store_ts_recent(tp);
5443 tp->tcp_header_len = sizeof(struct tcphdr);
5446 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5447 tcp_enable_fack(tp);
5450 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5451 tcp_initialize_rcv_mss(sk);
5453 /* Remember, tcp_poll() does not lock socket!
5454 * Change state from SYN-SENT only after copied_seq
5455 * is initialized. */
5456 tp->copied_seq = tp->rcv_nxt;
5460 tcp_finish_connect(sk, skb);
5462 if ((tp->syn_fastopen || tp->syn_data) &&
5463 tcp_rcv_fastopen_synack(sk, skb, &foc))
5466 if (sk->sk_write_pending ||
5467 icsk->icsk_accept_queue.rskq_defer_accept ||
5468 icsk->icsk_ack.pingpong) {
5469 /* Save one ACK. Data will be ready after
5470 * several ticks, if write_pending is set.
5472 * It may be deleted, but with this feature tcpdumps
5473 * look so _wonderfully_ clever, that I was not able
5474 * to stand against the temptation 8) --ANK
5476 inet_csk_schedule_ack(sk);
5477 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5478 tcp_enter_quickack_mode(sk);
5479 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5480 TCP_DELACK_MAX, TCP_RTO_MAX);
5491 /* No ACK in the segment */
5495 * "If the RST bit is set
5497 * Otherwise (no ACK) drop the segment and return."
5500 goto discard_and_undo;
5504 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5505 tcp_paws_reject(&tp->rx_opt, 0))
5506 goto discard_and_undo;
5509 /* We see SYN without ACK. It is attempt of
5510 * simultaneous connect with crossed SYNs.
5511 * Particularly, it can be connect to self.
5513 tcp_set_state(sk, TCP_SYN_RECV);
5515 if (tp->rx_opt.saw_tstamp) {
5516 tp->rx_opt.tstamp_ok = 1;
5517 tcp_store_ts_recent(tp);
5518 tp->tcp_header_len =
5519 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5521 tp->tcp_header_len = sizeof(struct tcphdr);
5524 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5525 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5527 /* RFC1323: The window in SYN & SYN/ACK segments is
5530 tp->snd_wnd = ntohs(th->window);
5531 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5532 tp->max_window = tp->snd_wnd;
5534 tcp_ecn_rcv_syn(tp, th);
5537 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5538 tcp_initialize_rcv_mss(sk);
5540 tcp_send_synack(sk);
5542 /* Note, we could accept data and URG from this segment.
5543 * There are no obstacles to make this (except that we must
5544 * either change tcp_recvmsg() to prevent it from returning data
5545 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5547 * However, if we ignore data in ACKless segments sometimes,
5548 * we have no reasons to accept it sometimes.
5549 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5550 * is not flawless. So, discard packet for sanity.
5551 * Uncomment this return to process the data.
5558 /* "fifth, if neither of the SYN or RST bits is set then
5559 * drop the segment and return."
5563 tcp_clear_options(&tp->rx_opt);
5564 tp->rx_opt.mss_clamp = saved_clamp;
5568 tcp_clear_options(&tp->rx_opt);
5569 tp->rx_opt.mss_clamp = saved_clamp;
5574 * This function implements the receiving procedure of RFC 793 for
5575 * all states except ESTABLISHED and TIME_WAIT.
5576 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5577 * address independent.
5580 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5581 const struct tcphdr *th, unsigned int len)
5583 struct tcp_sock *tp = tcp_sk(sk);
5584 struct inet_connection_sock *icsk = inet_csk(sk);
5585 struct request_sock *req;
5590 tp->rx_opt.saw_tstamp = 0;
5592 switch (sk->sk_state) {
5606 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5609 /* Now we have several options: In theory there is
5610 * nothing else in the frame. KA9Q has an option to
5611 * send data with the syn, BSD accepts data with the
5612 * syn up to the [to be] advertised window and
5613 * Solaris 2.1 gives you a protocol error. For now
5614 * we just ignore it, that fits the spec precisely
5615 * and avoids incompatibilities. It would be nice in
5616 * future to drop through and process the data.
5618 * Now that TTCP is starting to be used we ought to
5620 * But, this leaves one open to an easy denial of
5621 * service attack, and SYN cookies can't defend
5622 * against this problem. So, we drop the data
5623 * in the interest of security over speed unless
5624 * it's still in use.
5632 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5636 /* Do step6 onward by hand. */
5637 tcp_urg(sk, skb, th);
5639 tcp_data_snd_check(sk);
5643 req = tp->fastopen_rsk;
5645 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
5646 sk->sk_state != TCP_FIN_WAIT1);
5648 if (tcp_check_req(sk, skb, req, NULL, true) == NULL)
5652 if (!th->ack && !th->rst)
5655 if (!tcp_validate_incoming(sk, skb, th, 0))
5658 /* step 5: check the ACK field */
5659 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
5660 FLAG_UPDATE_TS_RECENT) > 0;
5662 switch (sk->sk_state) {
5667 /* Once we leave TCP_SYN_RECV, we no longer need req
5671 synack_stamp = tcp_rsk(req)->snt_synack;
5672 tp->total_retrans = req->num_retrans;
5673 reqsk_fastopen_remove(sk, req, false);
5675 synack_stamp = tp->lsndtime;
5676 /* Make sure socket is routed, for correct metrics. */
5677 icsk->icsk_af_ops->rebuild_header(sk);
5678 tcp_init_congestion_control(sk);
5681 tp->copied_seq = tp->rcv_nxt;
5682 tcp_init_buffer_space(sk);
5685 tcp_set_state(sk, TCP_ESTABLISHED);
5686 sk->sk_state_change(sk);
5688 /* Note, that this wakeup is only for marginal crossed SYN case.
5689 * Passively open sockets are not waked up, because
5690 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5693 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5695 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5696 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
5697 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5698 tcp_synack_rtt_meas(sk, synack_stamp);
5700 if (tp->rx_opt.tstamp_ok)
5701 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5704 /* Re-arm the timer because data may have been sent out.
5705 * This is similar to the regular data transmission case
5706 * when new data has just been ack'ed.
5708 * (TFO) - we could try to be more aggressive and
5709 * retransmitting any data sooner based on when they
5714 tcp_init_metrics(sk);
5716 tcp_update_pacing_rate(sk);
5718 /* Prevent spurious tcp_cwnd_restart() on first data packet */
5719 tp->lsndtime = tcp_time_stamp;
5721 tcp_initialize_rcv_mss(sk);
5722 tcp_fast_path_on(tp);
5725 case TCP_FIN_WAIT1: {
5726 struct dst_entry *dst;
5729 /* If we enter the TCP_FIN_WAIT1 state and we are a
5730 * Fast Open socket and this is the first acceptable
5731 * ACK we have received, this would have acknowledged
5732 * our SYNACK so stop the SYNACK timer.
5735 /* Return RST if ack_seq is invalid.
5736 * Note that RFC793 only says to generate a
5737 * DUPACK for it but for TCP Fast Open it seems
5738 * better to treat this case like TCP_SYN_RECV
5743 /* We no longer need the request sock. */
5744 reqsk_fastopen_remove(sk, req, false);
5747 if (tp->snd_una != tp->write_seq)
5750 tcp_set_state(sk, TCP_FIN_WAIT2);
5751 sk->sk_shutdown |= SEND_SHUTDOWN;
5753 dst = __sk_dst_get(sk);
5757 if (!sock_flag(sk, SOCK_DEAD)) {
5758 /* Wake up lingering close() */
5759 sk->sk_state_change(sk);
5763 if (tp->linger2 < 0 ||
5764 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5765 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5767 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5771 tmo = tcp_fin_time(sk);
5772 if (tmo > TCP_TIMEWAIT_LEN) {
5773 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5774 } else if (th->fin || sock_owned_by_user(sk)) {
5775 /* Bad case. We could lose such FIN otherwise.
5776 * It is not a big problem, but it looks confusing
5777 * and not so rare event. We still can lose it now,
5778 * if it spins in bh_lock_sock(), but it is really
5781 inet_csk_reset_keepalive_timer(sk, tmo);
5783 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5790 if (tp->snd_una == tp->write_seq) {
5791 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5797 if (tp->snd_una == tp->write_seq) {
5798 tcp_update_metrics(sk);
5805 /* step 6: check the URG bit */
5806 tcp_urg(sk, skb, th);
5808 /* step 7: process the segment text */
5809 switch (sk->sk_state) {
5810 case TCP_CLOSE_WAIT:
5813 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5817 /* RFC 793 says to queue data in these states,
5818 * RFC 1122 says we MUST send a reset.
5819 * BSD 4.4 also does reset.
5821 if (sk->sk_shutdown & RCV_SHUTDOWN) {
5822 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5823 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5824 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5830 case TCP_ESTABLISHED:
5831 tcp_data_queue(sk, skb);
5836 /* tcp_data could move socket to TIME-WAIT */
5837 if (sk->sk_state != TCP_CLOSE) {
5838 tcp_data_snd_check(sk);
5839 tcp_ack_snd_check(sk);
5848 EXPORT_SYMBOL(tcp_rcv_state_process);
5850 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
5852 struct inet_request_sock *ireq = inet_rsk(req);
5854 if (family == AF_INET)
5855 LIMIT_NETDEBUG(KERN_DEBUG pr_fmt("drop open request from %pI4/%u\n"),
5856 &ireq->ir_rmt_addr, port);
5857 #if IS_ENABLED(CONFIG_IPV6)
5858 else if (family == AF_INET6)
5859 LIMIT_NETDEBUG(KERN_DEBUG pr_fmt("drop open request from %pI6/%u\n"),
5860 &ireq->ir_v6_rmt_addr, port);
5864 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
5866 * If we receive a SYN packet with these bits set, it means a
5867 * network is playing bad games with TOS bits. In order to
5868 * avoid possible false congestion notifications, we disable
5869 * TCP ECN negotiation.
5871 * Exception: tcp_ca wants ECN. This is required for DCTCP
5872 * congestion control; it requires setting ECT on all packets,
5873 * including SYN. We inverse the test in this case: If our
5874 * local socket wants ECN, but peer only set ece/cwr (but not
5875 * ECT in IP header) its probably a non-DCTCP aware sender.
5877 static void tcp_ecn_create_request(struct request_sock *req,
5878 const struct sk_buff *skb,
5879 const struct sock *listen_sk)
5881 const struct tcphdr *th = tcp_hdr(skb);
5882 const struct net *net = sock_net(listen_sk);
5883 bool th_ecn = th->ece && th->cwr;
5889 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
5890 need_ecn = tcp_ca_needs_ecn(listen_sk);
5892 if (!ect && !need_ecn && net->ipv4.sysctl_tcp_ecn)
5893 inet_rsk(req)->ecn_ok = 1;
5894 else if (ect && need_ecn)
5895 inet_rsk(req)->ecn_ok = 1;
5898 int tcp_conn_request(struct request_sock_ops *rsk_ops,
5899 const struct tcp_request_sock_ops *af_ops,
5900 struct sock *sk, struct sk_buff *skb)
5902 struct tcp_options_received tmp_opt;
5903 struct request_sock *req;
5904 struct tcp_sock *tp = tcp_sk(sk);
5905 struct dst_entry *dst = NULL;
5906 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
5907 bool want_cookie = false, fastopen;
5909 struct tcp_fastopen_cookie foc = { .len = -1 };
5913 /* TW buckets are converted to open requests without
5914 * limitations, they conserve resources and peer is
5915 * evidently real one.
5917 if ((sysctl_tcp_syncookies == 2 ||
5918 inet_csk_reqsk_queue_is_full(sk)) && !isn) {
5919 want_cookie = tcp_syn_flood_action(sk, skb, rsk_ops->slab_name);
5925 /* Accept backlog is full. If we have already queued enough
5926 * of warm entries in syn queue, drop request. It is better than
5927 * clogging syn queue with openreqs with exponentially increasing
5930 if (sk_acceptq_is_full(sk) && inet_csk_reqsk_queue_young(sk) > 1) {
5931 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
5935 req = inet_reqsk_alloc(rsk_ops);
5939 tcp_rsk(req)->af_specific = af_ops;
5941 tcp_clear_options(&tmp_opt);
5942 tmp_opt.mss_clamp = af_ops->mss_clamp;
5943 tmp_opt.user_mss = tp->rx_opt.user_mss;
5944 tcp_parse_options(skb, &tmp_opt, 0, want_cookie ? NULL : &foc);
5946 if (want_cookie && !tmp_opt.saw_tstamp)
5947 tcp_clear_options(&tmp_opt);
5949 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
5950 tcp_openreq_init(req, &tmp_opt, skb, sk);
5952 af_ops->init_req(req, sk, skb);
5954 if (security_inet_conn_request(sk, skb, req))
5957 if (!want_cookie || tmp_opt.tstamp_ok)
5958 tcp_ecn_create_request(req, skb, sk);
5961 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
5962 req->cookie_ts = tmp_opt.tstamp_ok;
5964 /* VJ's idea. We save last timestamp seen
5965 * from the destination in peer table, when entering
5966 * state TIME-WAIT, and check against it before
5967 * accepting new connection request.
5969 * If "isn" is not zero, this request hit alive
5970 * timewait bucket, so that all the necessary checks
5971 * are made in the function processing timewait state.
5973 if (tcp_death_row.sysctl_tw_recycle) {
5976 dst = af_ops->route_req(sk, &fl, req, &strict);
5978 if (dst && strict &&
5979 !tcp_peer_is_proven(req, dst, true,
5980 tmp_opt.saw_tstamp)) {
5981 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSPASSIVEREJECTED);
5982 goto drop_and_release;
5985 /* Kill the following clause, if you dislike this way. */
5986 else if (!sysctl_tcp_syncookies &&
5987 (sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
5988 (sysctl_max_syn_backlog >> 2)) &&
5989 !tcp_peer_is_proven(req, dst, false,
5990 tmp_opt.saw_tstamp)) {
5991 /* Without syncookies last quarter of
5992 * backlog is filled with destinations,
5993 * proven to be alive.
5994 * It means that we continue to communicate
5995 * to destinations, already remembered
5996 * to the moment of synflood.
5998 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
6000 goto drop_and_release;
6003 isn = af_ops->init_seq(skb);
6006 dst = af_ops->route_req(sk, &fl, req, NULL);
6011 tcp_rsk(req)->snt_isn = isn;
6012 tcp_openreq_init_rwin(req, sk, dst);
6013 fastopen = !want_cookie &&
6014 tcp_try_fastopen(sk, skb, req, &foc, dst);
6015 err = af_ops->send_synack(sk, dst, &fl, req,
6016 skb_get_queue_mapping(skb), &foc);
6018 if (err || want_cookie)
6021 tcp_rsk(req)->listener = NULL;
6022 af_ops->queue_hash_add(sk, req, TCP_TIMEOUT_INIT);
6032 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_LISTENDROPS);
6035 EXPORT_SYMBOL(tcp_conn_request);