1 // SPDX-License-Identifier: GPL-2.0
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
7 * Implementation of the Transmission Control Protocol(TCP).
10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11 * Mark Evans, <evansmp@uhura.aston.ac.uk>
12 * Corey Minyard <wf-rch!minyard@relay.EU.net>
13 * Florian La Roche, <flla@stud.uni-sb.de>
14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15 * Linus Torvalds, <torvalds@cs.helsinki.fi>
16 * Alan Cox, <gw4pts@gw4pts.ampr.org>
17 * Matthew Dillon, <dillon@apollo.west.oic.com>
18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19 * Jorge Cwik, <jorge@laser.satlink.net>
24 * Pedro Roque : Fast Retransmit/Recovery.
26 * Retransmit queue handled by TCP.
27 * Better retransmit timer handling.
28 * New congestion avoidance.
32 * Eric : Fast Retransmit.
33 * Randy Scott : MSS option defines.
34 * Eric Schenk : Fixes to slow start algorithm.
35 * Eric Schenk : Yet another double ACK bug.
36 * Eric Schenk : Delayed ACK bug fixes.
37 * Eric Schenk : Floyd style fast retrans war avoidance.
38 * David S. Miller : Don't allow zero congestion window.
39 * Eric Schenk : Fix retransmitter so that it sends
40 * next packet on ack of previous packet.
41 * Andi Kleen : Moved open_request checking here
42 * and process RSTs for open_requests.
43 * Andi Kleen : Better prune_queue, and other fixes.
44 * Andrey Savochkin: Fix RTT measurements in the presence of
46 * Andrey Savochkin: Check sequence numbers correctly when
47 * removing SACKs due to in sequence incoming
49 * Andi Kleen: Make sure we never ack data there is not
50 * enough room for. Also make this condition
51 * a fatal error if it might still happen.
52 * Andi Kleen: Add tcp_measure_rcv_mss to make
53 * connections with MSS<min(MTU,ann. MSS)
54 * work without delayed acks.
55 * Andi Kleen: Process packets with PSH set in the
57 * J Hadi Salim: ECN support
60 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
61 * engine. Lots of bugs are found.
62 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
65 #define pr_fmt(fmt) "TCP: " fmt
68 #include <linux/slab.h>
69 #include <linux/module.h>
70 #include <linux/sysctl.h>
71 #include <linux/kernel.h>
72 #include <linux/prefetch.h>
75 #include <net/inet_common.h>
76 #include <linux/ipsec.h>
77 #include <asm/unaligned.h>
78 #include <linux/errqueue.h>
79 #include <trace/events/tcp.h>
80 #include <linux/static_key.h>
82 int sysctl_tcp_max_orphans __read_mostly
= NR_FILE
;
84 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
85 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
86 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
87 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
88 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
89 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
90 #define FLAG_ECE 0x40 /* ECE in this ACK */
91 #define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */
92 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
93 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
94 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
95 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
96 #define FLAG_SET_XMIT_TIMER 0x1000 /* Set TLP or RTO timer */
97 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
98 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
99 #define FLAG_NO_CHALLENGE_ACK 0x8000 /* do not call tcp_send_challenge_ack() */
101 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
102 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
103 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE|FLAG_DSACKING_ACK)
104 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
106 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
107 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
109 #define REXMIT_NONE 0 /* no loss recovery to do */
110 #define REXMIT_LOST 1 /* retransmit packets marked lost */
111 #define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */
113 static void tcp_gro_dev_warn(struct sock
*sk
, const struct sk_buff
*skb
,
116 static bool __once __read_mostly
;
119 struct net_device
*dev
;
124 dev
= dev_get_by_index_rcu(sock_net(sk
), skb
->skb_iif
);
125 if (!dev
|| len
>= dev
->mtu
)
126 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n",
127 dev
? dev
->name
: "Unknown driver");
132 /* Adapt the MSS value used to make delayed ack decision to the
135 static void tcp_measure_rcv_mss(struct sock
*sk
, const struct sk_buff
*skb
)
137 struct inet_connection_sock
*icsk
= inet_csk(sk
);
138 const unsigned int lss
= icsk
->icsk_ack
.last_seg_size
;
141 icsk
->icsk_ack
.last_seg_size
= 0;
143 /* skb->len may jitter because of SACKs, even if peer
144 * sends good full-sized frames.
146 len
= skb_shinfo(skb
)->gso_size
? : skb
->len
;
147 if (len
>= icsk
->icsk_ack
.rcv_mss
) {
148 icsk
->icsk_ack
.rcv_mss
= min_t(unsigned int, len
,
150 /* Account for possibly-removed options */
151 if (unlikely(len
> icsk
->icsk_ack
.rcv_mss
+
152 MAX_TCP_OPTION_SPACE
))
153 tcp_gro_dev_warn(sk
, skb
, len
);
155 /* Otherwise, we make more careful check taking into account,
156 * that SACKs block is variable.
158 * "len" is invariant segment length, including TCP header.
160 len
+= skb
->data
- skb_transport_header(skb
);
161 if (len
>= TCP_MSS_DEFAULT
+ sizeof(struct tcphdr
) ||
162 /* If PSH is not set, packet should be
163 * full sized, provided peer TCP is not badly broken.
164 * This observation (if it is correct 8)) allows
165 * to handle super-low mtu links fairly.
167 (len
>= TCP_MIN_MSS
+ sizeof(struct tcphdr
) &&
168 !(tcp_flag_word(tcp_hdr(skb
)) & TCP_REMNANT
))) {
169 /* Subtract also invariant (if peer is RFC compliant),
170 * tcp header plus fixed timestamp option length.
171 * Resulting "len" is MSS free of SACK jitter.
173 len
-= tcp_sk(sk
)->tcp_header_len
;
174 icsk
->icsk_ack
.last_seg_size
= len
;
176 icsk
->icsk_ack
.rcv_mss
= len
;
180 if (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
)
181 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED2
;
182 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
186 static void tcp_incr_quickack(struct sock
*sk
)
188 struct inet_connection_sock
*icsk
= inet_csk(sk
);
189 unsigned int quickacks
= tcp_sk(sk
)->rcv_wnd
/ (2 * icsk
->icsk_ack
.rcv_mss
);
193 if (quickacks
> icsk
->icsk_ack
.quick
)
194 icsk
->icsk_ack
.quick
= min(quickacks
, TCP_MAX_QUICKACKS
);
197 static void tcp_enter_quickack_mode(struct sock
*sk
)
199 struct inet_connection_sock
*icsk
= inet_csk(sk
);
200 tcp_incr_quickack(sk
);
201 icsk
->icsk_ack
.pingpong
= 0;
202 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
205 /* Send ACKs quickly, if "quick" count is not exhausted
206 * and the session is not interactive.
209 static bool tcp_in_quickack_mode(struct sock
*sk
)
211 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
212 const struct dst_entry
*dst
= __sk_dst_get(sk
);
214 return (dst
&& dst_metric(dst
, RTAX_QUICKACK
)) ||
215 (icsk
->icsk_ack
.quick
&& !icsk
->icsk_ack
.pingpong
);
218 static void tcp_ecn_queue_cwr(struct tcp_sock
*tp
)
220 if (tp
->ecn_flags
& TCP_ECN_OK
)
221 tp
->ecn_flags
|= TCP_ECN_QUEUE_CWR
;
224 static void tcp_ecn_accept_cwr(struct tcp_sock
*tp
, const struct sk_buff
*skb
)
226 if (tcp_hdr(skb
)->cwr
)
227 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
230 static void tcp_ecn_withdraw_cwr(struct tcp_sock
*tp
)
232 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
235 static void __tcp_ecn_check_ce(struct tcp_sock
*tp
, const struct sk_buff
*skb
)
237 switch (TCP_SKB_CB(skb
)->ip_dsfield
& INET_ECN_MASK
) {
238 case INET_ECN_NOT_ECT
:
239 /* Funny extension: if ECT is not set on a segment,
240 * and we already seen ECT on a previous segment,
241 * it is probably a retransmit.
243 if (tp
->ecn_flags
& TCP_ECN_SEEN
)
244 tcp_enter_quickack_mode((struct sock
*)tp
);
247 if (tcp_ca_needs_ecn((struct sock
*)tp
))
248 tcp_ca_event((struct sock
*)tp
, CA_EVENT_ECN_IS_CE
);
250 if (!(tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)) {
251 /* Better not delay acks, sender can have a very low cwnd */
252 tcp_enter_quickack_mode((struct sock
*)tp
);
253 tp
->ecn_flags
|= TCP_ECN_DEMAND_CWR
;
255 tp
->ecn_flags
|= TCP_ECN_SEEN
;
258 if (tcp_ca_needs_ecn((struct sock
*)tp
))
259 tcp_ca_event((struct sock
*)tp
, CA_EVENT_ECN_NO_CE
);
260 tp
->ecn_flags
|= TCP_ECN_SEEN
;
265 static void tcp_ecn_check_ce(struct tcp_sock
*tp
, const struct sk_buff
*skb
)
267 if (tp
->ecn_flags
& TCP_ECN_OK
)
268 __tcp_ecn_check_ce(tp
, skb
);
271 static void tcp_ecn_rcv_synack(struct tcp_sock
*tp
, const struct tcphdr
*th
)
273 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| th
->cwr
))
274 tp
->ecn_flags
&= ~TCP_ECN_OK
;
277 static void tcp_ecn_rcv_syn(struct tcp_sock
*tp
, const struct tcphdr
*th
)
279 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| !th
->cwr
))
280 tp
->ecn_flags
&= ~TCP_ECN_OK
;
283 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock
*tp
, const struct tcphdr
*th
)
285 if (th
->ece
&& !th
->syn
&& (tp
->ecn_flags
& TCP_ECN_OK
))
290 /* Buffer size and advertised window tuning.
292 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
295 static void tcp_sndbuf_expand(struct sock
*sk
)
297 const struct tcp_sock
*tp
= tcp_sk(sk
);
298 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
302 /* Worst case is non GSO/TSO : each frame consumes one skb
303 * and skb->head is kmalloced using power of two area of memory
305 per_mss
= max_t(u32
, tp
->rx_opt
.mss_clamp
, tp
->mss_cache
) +
307 SKB_DATA_ALIGN(sizeof(struct skb_shared_info
));
309 per_mss
= roundup_pow_of_two(per_mss
) +
310 SKB_DATA_ALIGN(sizeof(struct sk_buff
));
312 nr_segs
= max_t(u32
, TCP_INIT_CWND
, tp
->snd_cwnd
);
313 nr_segs
= max_t(u32
, nr_segs
, tp
->reordering
+ 1);
315 /* Fast Recovery (RFC 5681 3.2) :
316 * Cubic needs 1.7 factor, rounded to 2 to include
317 * extra cushion (application might react slowly to POLLOUT)
319 sndmem
= ca_ops
->sndbuf_expand
? ca_ops
->sndbuf_expand(sk
) : 2;
320 sndmem
*= nr_segs
* per_mss
;
322 if (sk
->sk_sndbuf
< sndmem
)
323 sk
->sk_sndbuf
= min(sndmem
, sock_net(sk
)->ipv4
.sysctl_tcp_wmem
[2]);
326 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
328 * All tcp_full_space() is split to two parts: "network" buffer, allocated
329 * forward and advertised in receiver window (tp->rcv_wnd) and
330 * "application buffer", required to isolate scheduling/application
331 * latencies from network.
332 * window_clamp is maximal advertised window. It can be less than
333 * tcp_full_space(), in this case tcp_full_space() - window_clamp
334 * is reserved for "application" buffer. The less window_clamp is
335 * the smoother our behaviour from viewpoint of network, but the lower
336 * throughput and the higher sensitivity of the connection to losses. 8)
338 * rcv_ssthresh is more strict window_clamp used at "slow start"
339 * phase to predict further behaviour of this connection.
340 * It is used for two goals:
341 * - to enforce header prediction at sender, even when application
342 * requires some significant "application buffer". It is check #1.
343 * - to prevent pruning of receive queue because of misprediction
344 * of receiver window. Check #2.
346 * The scheme does not work when sender sends good segments opening
347 * window and then starts to feed us spaghetti. But it should work
348 * in common situations. Otherwise, we have to rely on queue collapsing.
351 /* Slow part of check#2. */
352 static int __tcp_grow_window(const struct sock
*sk
, const struct sk_buff
*skb
)
354 struct tcp_sock
*tp
= tcp_sk(sk
);
356 int truesize
= tcp_win_from_space(sk
, skb
->truesize
) >> 1;
357 int window
= tcp_win_from_space(sk
, sock_net(sk
)->ipv4
.sysctl_tcp_rmem
[2]) >> 1;
359 while (tp
->rcv_ssthresh
<= window
) {
360 if (truesize
<= skb
->len
)
361 return 2 * inet_csk(sk
)->icsk_ack
.rcv_mss
;
369 static void tcp_grow_window(struct sock
*sk
, const struct sk_buff
*skb
)
371 struct tcp_sock
*tp
= tcp_sk(sk
);
374 if (tp
->rcv_ssthresh
< tp
->window_clamp
&&
375 (int)tp
->rcv_ssthresh
< tcp_space(sk
) &&
376 !tcp_under_memory_pressure(sk
)) {
379 /* Check #2. Increase window, if skb with such overhead
380 * will fit to rcvbuf in future.
382 if (tcp_win_from_space(sk
, skb
->truesize
) <= skb
->len
)
383 incr
= 2 * tp
->advmss
;
385 incr
= __tcp_grow_window(sk
, skb
);
388 incr
= max_t(int, incr
, 2 * skb
->len
);
389 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
+ incr
,
391 inet_csk(sk
)->icsk_ack
.quick
|= 1;
396 /* 3. Tuning rcvbuf, when connection enters established state. */
397 static void tcp_fixup_rcvbuf(struct sock
*sk
)
399 u32 mss
= tcp_sk(sk
)->advmss
;
402 rcvmem
= 2 * SKB_TRUESIZE(mss
+ MAX_TCP_HEADER
) *
403 tcp_default_init_rwnd(mss
);
405 /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
406 * Allow enough cushion so that sender is not limited by our window
408 if (sock_net(sk
)->ipv4
.sysctl_tcp_moderate_rcvbuf
)
411 if (sk
->sk_rcvbuf
< rcvmem
)
412 sk
->sk_rcvbuf
= min(rcvmem
, sock_net(sk
)->ipv4
.sysctl_tcp_rmem
[2]);
415 /* 4. Try to fixup all. It is made immediately after connection enters
418 void tcp_init_buffer_space(struct sock
*sk
)
420 int tcp_app_win
= sock_net(sk
)->ipv4
.sysctl_tcp_app_win
;
421 struct tcp_sock
*tp
= tcp_sk(sk
);
424 if (!(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
))
425 tcp_fixup_rcvbuf(sk
);
426 if (!(sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
))
427 tcp_sndbuf_expand(sk
);
429 tp
->rcvq_space
.space
= tp
->rcv_wnd
;
430 tcp_mstamp_refresh(tp
);
431 tp
->rcvq_space
.time
= tp
->tcp_mstamp
;
432 tp
->rcvq_space
.seq
= tp
->copied_seq
;
434 maxwin
= tcp_full_space(sk
);
436 if (tp
->window_clamp
>= maxwin
) {
437 tp
->window_clamp
= maxwin
;
439 if (tcp_app_win
&& maxwin
> 4 * tp
->advmss
)
440 tp
->window_clamp
= max(maxwin
-
441 (maxwin
>> tcp_app_win
),
445 /* Force reservation of one segment. */
447 tp
->window_clamp
> 2 * tp
->advmss
&&
448 tp
->window_clamp
+ tp
->advmss
> maxwin
)
449 tp
->window_clamp
= max(2 * tp
->advmss
, maxwin
- tp
->advmss
);
451 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, tp
->window_clamp
);
452 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
455 /* 5. Recalculate window clamp after socket hit its memory bounds. */
456 static void tcp_clamp_window(struct sock
*sk
)
458 struct tcp_sock
*tp
= tcp_sk(sk
);
459 struct inet_connection_sock
*icsk
= inet_csk(sk
);
460 struct net
*net
= sock_net(sk
);
462 icsk
->icsk_ack
.quick
= 0;
464 if (sk
->sk_rcvbuf
< net
->ipv4
.sysctl_tcp_rmem
[2] &&
465 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
) &&
466 !tcp_under_memory_pressure(sk
) &&
467 sk_memory_allocated(sk
) < sk_prot_mem_limits(sk
, 0)) {
468 sk
->sk_rcvbuf
= min(atomic_read(&sk
->sk_rmem_alloc
),
469 net
->ipv4
.sysctl_tcp_rmem
[2]);
471 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
)
472 tp
->rcv_ssthresh
= min(tp
->window_clamp
, 2U * tp
->advmss
);
475 /* Initialize RCV_MSS value.
476 * RCV_MSS is an our guess about MSS used by the peer.
477 * We haven't any direct information about the MSS.
478 * It's better to underestimate the RCV_MSS rather than overestimate.
479 * Overestimations make us ACKing less frequently than needed.
480 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
482 void tcp_initialize_rcv_mss(struct sock
*sk
)
484 const struct tcp_sock
*tp
= tcp_sk(sk
);
485 unsigned int hint
= min_t(unsigned int, tp
->advmss
, tp
->mss_cache
);
487 hint
= min(hint
, tp
->rcv_wnd
/ 2);
488 hint
= min(hint
, TCP_MSS_DEFAULT
);
489 hint
= max(hint
, TCP_MIN_MSS
);
491 inet_csk(sk
)->icsk_ack
.rcv_mss
= hint
;
493 EXPORT_SYMBOL(tcp_initialize_rcv_mss
);
495 /* Receiver "autotuning" code.
497 * The algorithm for RTT estimation w/o timestamps is based on
498 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
499 * <http://public.lanl.gov/radiant/pubs.html#DRS>
501 * More detail on this code can be found at
502 * <http://staff.psc.edu/jheffner/>,
503 * though this reference is out of date. A new paper
506 static void tcp_rcv_rtt_update(struct tcp_sock
*tp
, u32 sample
, int win_dep
)
508 u32 new_sample
= tp
->rcv_rtt_est
.rtt_us
;
514 if (new_sample
!= 0) {
515 /* If we sample in larger samples in the non-timestamp
516 * case, we could grossly overestimate the RTT especially
517 * with chatty applications or bulk transfer apps which
518 * are stalled on filesystem I/O.
520 * Also, since we are only going for a minimum in the
521 * non-timestamp case, we do not smooth things out
522 * else with timestamps disabled convergence takes too
526 m
-= (new_sample
>> 3);
534 /* No previous measure. */
538 tp
->rcv_rtt_est
.rtt_us
= new_sample
;
541 static inline void tcp_rcv_rtt_measure(struct tcp_sock
*tp
)
545 if (tp
->rcv_rtt_est
.time
== 0)
547 if (before(tp
->rcv_nxt
, tp
->rcv_rtt_est
.seq
))
549 delta_us
= tcp_stamp_us_delta(tp
->tcp_mstamp
, tp
->rcv_rtt_est
.time
);
550 tcp_rcv_rtt_update(tp
, delta_us
, 1);
553 tp
->rcv_rtt_est
.seq
= tp
->rcv_nxt
+ tp
->rcv_wnd
;
554 tp
->rcv_rtt_est
.time
= tp
->tcp_mstamp
;
557 static inline void tcp_rcv_rtt_measure_ts(struct sock
*sk
,
558 const struct sk_buff
*skb
)
560 struct tcp_sock
*tp
= tcp_sk(sk
);
562 if (tp
->rx_opt
.rcv_tsecr
&&
563 (TCP_SKB_CB(skb
)->end_seq
-
564 TCP_SKB_CB(skb
)->seq
>= inet_csk(sk
)->icsk_ack
.rcv_mss
)) {
565 u32 delta
= tcp_time_stamp(tp
) - tp
->rx_opt
.rcv_tsecr
;
566 u32 delta_us
= delta
* (USEC_PER_SEC
/ TCP_TS_HZ
);
568 tcp_rcv_rtt_update(tp
, delta_us
, 0);
573 * This function should be called every time data is copied to user space.
574 * It calculates the appropriate TCP receive buffer space.
576 void tcp_rcv_space_adjust(struct sock
*sk
)
578 struct tcp_sock
*tp
= tcp_sk(sk
);
582 time
= tcp_stamp_us_delta(tp
->tcp_mstamp
, tp
->rcvq_space
.time
);
583 if (time
< (tp
->rcv_rtt_est
.rtt_us
>> 3) || tp
->rcv_rtt_est
.rtt_us
== 0)
586 /* Number of bytes copied to user in last RTT */
587 copied
= tp
->copied_seq
- tp
->rcvq_space
.seq
;
588 if (copied
<= tp
->rcvq_space
.space
)
592 * copied = bytes received in previous RTT, our base window
593 * To cope with packet losses, we need a 2x factor
594 * To cope with slow start, and sender growing its cwin by 100 %
595 * every RTT, we need a 4x factor, because the ACK we are sending
596 * now is for the next RTT, not the current one :
597 * <prev RTT . ><current RTT .. ><next RTT .... >
600 if (sock_net(sk
)->ipv4
.sysctl_tcp_moderate_rcvbuf
&&
601 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
)) {
602 int rcvwin
, rcvmem
, rcvbuf
;
604 /* minimal window to cope with packet losses, assuming
605 * steady state. Add some cushion because of small variations.
607 rcvwin
= (copied
<< 1) + 16 * tp
->advmss
;
609 /* If rate increased by 25%,
610 * assume slow start, rcvwin = 3 * copied
611 * If rate increased by 50%,
612 * assume sender can use 2x growth, rcvwin = 4 * copied
615 tp
->rcvq_space
.space
+ (tp
->rcvq_space
.space
>> 2)) {
617 tp
->rcvq_space
.space
+ (tp
->rcvq_space
.space
>> 1))
620 rcvwin
+= (rcvwin
>> 1);
623 rcvmem
= SKB_TRUESIZE(tp
->advmss
+ MAX_TCP_HEADER
);
624 while (tcp_win_from_space(sk
, rcvmem
) < tp
->advmss
)
627 rcvbuf
= min(rcvwin
/ tp
->advmss
* rcvmem
,
628 sock_net(sk
)->ipv4
.sysctl_tcp_rmem
[2]);
629 if (rcvbuf
> sk
->sk_rcvbuf
) {
630 sk
->sk_rcvbuf
= rcvbuf
;
632 /* Make the window clamp follow along. */
633 tp
->window_clamp
= rcvwin
;
636 tp
->rcvq_space
.space
= copied
;
639 tp
->rcvq_space
.seq
= tp
->copied_seq
;
640 tp
->rcvq_space
.time
= tp
->tcp_mstamp
;
643 /* There is something which you must keep in mind when you analyze the
644 * behavior of the tp->ato delayed ack timeout interval. When a
645 * connection starts up, we want to ack as quickly as possible. The
646 * problem is that "good" TCP's do slow start at the beginning of data
647 * transmission. The means that until we send the first few ACK's the
648 * sender will sit on his end and only queue most of his data, because
649 * he can only send snd_cwnd unacked packets at any given time. For
650 * each ACK we send, he increments snd_cwnd and transmits more of his
653 static void tcp_event_data_recv(struct sock
*sk
, struct sk_buff
*skb
)
655 struct tcp_sock
*tp
= tcp_sk(sk
);
656 struct inet_connection_sock
*icsk
= inet_csk(sk
);
659 inet_csk_schedule_ack(sk
);
661 tcp_measure_rcv_mss(sk
, skb
);
663 tcp_rcv_rtt_measure(tp
);
667 if (!icsk
->icsk_ack
.ato
) {
668 /* The _first_ data packet received, initialize
669 * delayed ACK engine.
671 tcp_incr_quickack(sk
);
672 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
674 int m
= now
- icsk
->icsk_ack
.lrcvtime
;
676 if (m
<= TCP_ATO_MIN
/ 2) {
677 /* The fastest case is the first. */
678 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + TCP_ATO_MIN
/ 2;
679 } else if (m
< icsk
->icsk_ack
.ato
) {
680 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + m
;
681 if (icsk
->icsk_ack
.ato
> icsk
->icsk_rto
)
682 icsk
->icsk_ack
.ato
= icsk
->icsk_rto
;
683 } else if (m
> icsk
->icsk_rto
) {
684 /* Too long gap. Apparently sender failed to
685 * restart window, so that we send ACKs quickly.
687 tcp_incr_quickack(sk
);
691 icsk
->icsk_ack
.lrcvtime
= now
;
693 tcp_ecn_check_ce(tp
, skb
);
696 tcp_grow_window(sk
, skb
);
699 /* Called to compute a smoothed rtt estimate. The data fed to this
700 * routine either comes from timestamps, or from segments that were
701 * known _not_ to have been retransmitted [see Karn/Partridge
702 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
703 * piece by Van Jacobson.
704 * NOTE: the next three routines used to be one big routine.
705 * To save cycles in the RFC 1323 implementation it was better to break
706 * it up into three procedures. -- erics
708 static void tcp_rtt_estimator(struct sock
*sk
, long mrtt_us
)
710 struct tcp_sock
*tp
= tcp_sk(sk
);
711 long m
= mrtt_us
; /* RTT */
712 u32 srtt
= tp
->srtt_us
;
714 /* The following amusing code comes from Jacobson's
715 * article in SIGCOMM '88. Note that rtt and mdev
716 * are scaled versions of rtt and mean deviation.
717 * This is designed to be as fast as possible
718 * m stands for "measurement".
720 * On a 1990 paper the rto value is changed to:
721 * RTO = rtt + 4 * mdev
723 * Funny. This algorithm seems to be very broken.
724 * These formulae increase RTO, when it should be decreased, increase
725 * too slowly, when it should be increased quickly, decrease too quickly
726 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
727 * does not matter how to _calculate_ it. Seems, it was trap
728 * that VJ failed to avoid. 8)
731 m
-= (srtt
>> 3); /* m is now error in rtt est */
732 srtt
+= m
; /* rtt = 7/8 rtt + 1/8 new */
734 m
= -m
; /* m is now abs(error) */
735 m
-= (tp
->mdev_us
>> 2); /* similar update on mdev */
736 /* This is similar to one of Eifel findings.
737 * Eifel blocks mdev updates when rtt decreases.
738 * This solution is a bit different: we use finer gain
739 * for mdev in this case (alpha*beta).
740 * Like Eifel it also prevents growth of rto,
741 * but also it limits too fast rto decreases,
742 * happening in pure Eifel.
747 m
-= (tp
->mdev_us
>> 2); /* similar update on mdev */
749 tp
->mdev_us
+= m
; /* mdev = 3/4 mdev + 1/4 new */
750 if (tp
->mdev_us
> tp
->mdev_max_us
) {
751 tp
->mdev_max_us
= tp
->mdev_us
;
752 if (tp
->mdev_max_us
> tp
->rttvar_us
)
753 tp
->rttvar_us
= tp
->mdev_max_us
;
755 if (after(tp
->snd_una
, tp
->rtt_seq
)) {
756 if (tp
->mdev_max_us
< tp
->rttvar_us
)
757 tp
->rttvar_us
-= (tp
->rttvar_us
- tp
->mdev_max_us
) >> 2;
758 tp
->rtt_seq
= tp
->snd_nxt
;
759 tp
->mdev_max_us
= tcp_rto_min_us(sk
);
762 /* no previous measure. */
763 srtt
= m
<< 3; /* take the measured time to be rtt */
764 tp
->mdev_us
= m
<< 1; /* make sure rto = 3*rtt */
765 tp
->rttvar_us
= max(tp
->mdev_us
, tcp_rto_min_us(sk
));
766 tp
->mdev_max_us
= tp
->rttvar_us
;
767 tp
->rtt_seq
= tp
->snd_nxt
;
769 tp
->srtt_us
= max(1U, srtt
);
772 static void tcp_update_pacing_rate(struct sock
*sk
)
774 const struct tcp_sock
*tp
= tcp_sk(sk
);
777 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
778 rate
= (u64
)tp
->mss_cache
* ((USEC_PER_SEC
/ 100) << 3);
780 /* current rate is (cwnd * mss) / srtt
781 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
782 * In Congestion Avoidance phase, set it to 120 % the current rate.
784 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
785 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
786 * end of slow start and should slow down.
788 if (tp
->snd_cwnd
< tp
->snd_ssthresh
/ 2)
789 rate
*= sock_net(sk
)->ipv4
.sysctl_tcp_pacing_ss_ratio
;
791 rate
*= sock_net(sk
)->ipv4
.sysctl_tcp_pacing_ca_ratio
;
793 rate
*= max(tp
->snd_cwnd
, tp
->packets_out
);
795 if (likely(tp
->srtt_us
))
796 do_div(rate
, tp
->srtt_us
);
798 /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate
799 * without any lock. We want to make sure compiler wont store
800 * intermediate values in this location.
802 WRITE_ONCE(sk
->sk_pacing_rate
, min_t(u64
, rate
,
803 sk
->sk_max_pacing_rate
));
806 /* Calculate rto without backoff. This is the second half of Van Jacobson's
807 * routine referred to above.
809 static void tcp_set_rto(struct sock
*sk
)
811 const struct tcp_sock
*tp
= tcp_sk(sk
);
812 /* Old crap is replaced with new one. 8)
815 * 1. If rtt variance happened to be less 50msec, it is hallucination.
816 * It cannot be less due to utterly erratic ACK generation made
817 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
818 * to do with delayed acks, because at cwnd>2 true delack timeout
819 * is invisible. Actually, Linux-2.4 also generates erratic
820 * ACKs in some circumstances.
822 inet_csk(sk
)->icsk_rto
= __tcp_set_rto(tp
);
824 /* 2. Fixups made earlier cannot be right.
825 * If we do not estimate RTO correctly without them,
826 * all the algo is pure shit and should be replaced
827 * with correct one. It is exactly, which we pretend to do.
830 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
831 * guarantees that rto is higher.
836 __u32
tcp_init_cwnd(const struct tcp_sock
*tp
, const struct dst_entry
*dst
)
838 __u32 cwnd
= (dst
? dst_metric(dst
, RTAX_INITCWND
) : 0);
841 cwnd
= TCP_INIT_CWND
;
842 return min_t(__u32
, cwnd
, tp
->snd_cwnd_clamp
);
845 /* Take a notice that peer is sending D-SACKs */
846 static void tcp_dsack_seen(struct tcp_sock
*tp
)
848 tp
->rx_opt
.sack_ok
|= TCP_DSACK_SEEN
;
849 tp
->rack
.dsack_seen
= 1;
852 /* It's reordering when higher sequence was delivered (i.e. sacked) before
853 * some lower never-retransmitted sequence ("low_seq"). The maximum reordering
854 * distance is approximated in full-mss packet distance ("reordering").
856 static void tcp_check_sack_reordering(struct sock
*sk
, const u32 low_seq
,
859 struct tcp_sock
*tp
= tcp_sk(sk
);
860 const u32 mss
= tp
->mss_cache
;
863 fack
= tcp_highest_sack_seq(tp
);
864 if (!before(low_seq
, fack
))
867 metric
= fack
- low_seq
;
868 if ((metric
> tp
->reordering
* mss
) && mss
) {
869 #if FASTRETRANS_DEBUG > 1
870 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
871 tp
->rx_opt
.sack_ok
, inet_csk(sk
)->icsk_ca_state
,
875 tp
->undo_marker
? tp
->undo_retrans
: 0);
877 tp
->reordering
= min_t(u32
, (metric
+ mss
- 1) / mss
,
878 sock_net(sk
)->ipv4
.sysctl_tcp_max_reordering
);
882 /* This exciting event is worth to be remembered. 8) */
883 NET_INC_STATS(sock_net(sk
),
884 ts
? LINUX_MIB_TCPTSREORDER
: LINUX_MIB_TCPSACKREORDER
);
887 /* This must be called before lost_out is incremented */
888 static void tcp_verify_retransmit_hint(struct tcp_sock
*tp
, struct sk_buff
*skb
)
890 if (!tp
->retransmit_skb_hint
||
891 before(TCP_SKB_CB(skb
)->seq
,
892 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
893 tp
->retransmit_skb_hint
= skb
;
896 /* Sum the number of packets on the wire we have marked as lost.
897 * There are two cases we care about here:
898 * a) Packet hasn't been marked lost (nor retransmitted),
899 * and this is the first loss.
900 * b) Packet has been marked both lost and retransmitted,
901 * and this means we think it was lost again.
903 static void tcp_sum_lost(struct tcp_sock
*tp
, struct sk_buff
*skb
)
905 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
907 if (!(sacked
& TCPCB_LOST
) ||
908 ((sacked
& TCPCB_LOST
) && (sacked
& TCPCB_SACKED_RETRANS
)))
909 tp
->lost
+= tcp_skb_pcount(skb
);
912 static void tcp_skb_mark_lost(struct tcp_sock
*tp
, struct sk_buff
*skb
)
914 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
915 tcp_verify_retransmit_hint(tp
, skb
);
917 tp
->lost_out
+= tcp_skb_pcount(skb
);
918 tcp_sum_lost(tp
, skb
);
919 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
923 void tcp_skb_mark_lost_uncond_verify(struct tcp_sock
*tp
, struct sk_buff
*skb
)
925 tcp_verify_retransmit_hint(tp
, skb
);
927 tcp_sum_lost(tp
, skb
);
928 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
929 tp
->lost_out
+= tcp_skb_pcount(skb
);
930 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
934 /* This procedure tags the retransmission queue when SACKs arrive.
936 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
937 * Packets in queue with these bits set are counted in variables
938 * sacked_out, retrans_out and lost_out, correspondingly.
940 * Valid combinations are:
941 * Tag InFlight Description
942 * 0 1 - orig segment is in flight.
943 * S 0 - nothing flies, orig reached receiver.
944 * L 0 - nothing flies, orig lost by net.
945 * R 2 - both orig and retransmit are in flight.
946 * L|R 1 - orig is lost, retransmit is in flight.
947 * S|R 1 - orig reached receiver, retrans is still in flight.
948 * (L|S|R is logically valid, it could occur when L|R is sacked,
949 * but it is equivalent to plain S and code short-curcuits it to S.
950 * L|S is logically invalid, it would mean -1 packet in flight 8))
952 * These 6 states form finite state machine, controlled by the following events:
953 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
954 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
955 * 3. Loss detection event of two flavors:
956 * A. Scoreboard estimator decided the packet is lost.
957 * A'. Reno "three dupacks" marks head of queue lost.
958 * B. SACK arrives sacking SND.NXT at the moment, when the
959 * segment was retransmitted.
960 * 4. D-SACK added new rule: D-SACK changes any tag to S.
962 * It is pleasant to note, that state diagram turns out to be commutative,
963 * so that we are allowed not to be bothered by order of our actions,
964 * when multiple events arrive simultaneously. (see the function below).
966 * Reordering detection.
967 * --------------------
968 * Reordering metric is maximal distance, which a packet can be displaced
969 * in packet stream. With SACKs we can estimate it:
971 * 1. SACK fills old hole and the corresponding segment was not
972 * ever retransmitted -> reordering. Alas, we cannot use it
973 * when segment was retransmitted.
974 * 2. The last flaw is solved with D-SACK. D-SACK arrives
975 * for retransmitted and already SACKed segment -> reordering..
976 * Both of these heuristics are not used in Loss state, when we cannot
977 * account for retransmits accurately.
979 * SACK block validation.
980 * ----------------------
982 * SACK block range validation checks that the received SACK block fits to
983 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
984 * Note that SND.UNA is not included to the range though being valid because
985 * it means that the receiver is rather inconsistent with itself reporting
986 * SACK reneging when it should advance SND.UNA. Such SACK block this is
987 * perfectly valid, however, in light of RFC2018 which explicitly states
988 * that "SACK block MUST reflect the newest segment. Even if the newest
989 * segment is going to be discarded ...", not that it looks very clever
990 * in case of head skb. Due to potentional receiver driven attacks, we
991 * choose to avoid immediate execution of a walk in write queue due to
992 * reneging and defer head skb's loss recovery to standard loss recovery
993 * procedure that will eventually trigger (nothing forbids us doing this).
995 * Implements also blockage to start_seq wrap-around. Problem lies in the
996 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
997 * there's no guarantee that it will be before snd_nxt (n). The problem
998 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1001 * <- outs wnd -> <- wrapzone ->
1002 * u e n u_w e_w s n_w
1004 * |<------------+------+----- TCP seqno space --------------+---------->|
1005 * ...-- <2^31 ->| |<--------...
1006 * ...---- >2^31 ------>| |<--------...
1008 * Current code wouldn't be vulnerable but it's better still to discard such
1009 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1010 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1011 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1012 * equal to the ideal case (infinite seqno space without wrap caused issues).
1014 * With D-SACK the lower bound is extended to cover sequence space below
1015 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1016 * again, D-SACK block must not to go across snd_una (for the same reason as
1017 * for the normal SACK blocks, explained above). But there all simplicity
1018 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1019 * fully below undo_marker they do not affect behavior in anyway and can
1020 * therefore be safely ignored. In rare cases (which are more or less
1021 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1022 * fragmentation and packet reordering past skb's retransmission. To consider
1023 * them correctly, the acceptable range must be extended even more though
1024 * the exact amount is rather hard to quantify. However, tp->max_window can
1025 * be used as an exaggerated estimate.
1027 static bool tcp_is_sackblock_valid(struct tcp_sock
*tp
, bool is_dsack
,
1028 u32 start_seq
, u32 end_seq
)
1030 /* Too far in future, or reversed (interpretation is ambiguous) */
1031 if (after(end_seq
, tp
->snd_nxt
) || !before(start_seq
, end_seq
))
1034 /* Nasty start_seq wrap-around check (see comments above) */
1035 if (!before(start_seq
, tp
->snd_nxt
))
1038 /* In outstanding window? ...This is valid exit for D-SACKs too.
1039 * start_seq == snd_una is non-sensical (see comments above)
1041 if (after(start_seq
, tp
->snd_una
))
1044 if (!is_dsack
|| !tp
->undo_marker
)
1047 /* ...Then it's D-SACK, and must reside below snd_una completely */
1048 if (after(end_seq
, tp
->snd_una
))
1051 if (!before(start_seq
, tp
->undo_marker
))
1055 if (!after(end_seq
, tp
->undo_marker
))
1058 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1059 * start_seq < undo_marker and end_seq >= undo_marker.
1061 return !before(start_seq
, end_seq
- tp
->max_window
);
1064 static bool tcp_check_dsack(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1065 struct tcp_sack_block_wire
*sp
, int num_sacks
,
1068 struct tcp_sock
*tp
= tcp_sk(sk
);
1069 u32 start_seq_0
= get_unaligned_be32(&sp
[0].start_seq
);
1070 u32 end_seq_0
= get_unaligned_be32(&sp
[0].end_seq
);
1071 bool dup_sack
= false;
1073 if (before(start_seq_0
, TCP_SKB_CB(ack_skb
)->ack_seq
)) {
1076 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPDSACKRECV
);
1077 } else if (num_sacks
> 1) {
1078 u32 end_seq_1
= get_unaligned_be32(&sp
[1].end_seq
);
1079 u32 start_seq_1
= get_unaligned_be32(&sp
[1].start_seq
);
1081 if (!after(end_seq_0
, end_seq_1
) &&
1082 !before(start_seq_0
, start_seq_1
)) {
1085 NET_INC_STATS(sock_net(sk
),
1086 LINUX_MIB_TCPDSACKOFORECV
);
1090 /* D-SACK for already forgotten data... Do dumb counting. */
1091 if (dup_sack
&& tp
->undo_marker
&& tp
->undo_retrans
> 0 &&
1092 !after(end_seq_0
, prior_snd_una
) &&
1093 after(end_seq_0
, tp
->undo_marker
))
1099 struct tcp_sacktag_state
{
1101 /* Timestamps for earliest and latest never-retransmitted segment
1102 * that was SACKed. RTO needs the earliest RTT to stay conservative,
1103 * but congestion control should still get an accurate delay signal.
1107 struct rate_sample
*rate
;
1109 unsigned int mss_now
;
1112 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1113 * the incoming SACK may not exactly match but we can find smaller MSS
1114 * aligned portion of it that matches. Therefore we might need to fragment
1115 * which may fail and creates some hassle (caller must handle error case
1118 * FIXME: this could be merged to shift decision code
1120 static int tcp_match_skb_to_sack(struct sock
*sk
, struct sk_buff
*skb
,
1121 u32 start_seq
, u32 end_seq
)
1125 unsigned int pkt_len
;
1128 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1129 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1131 if (tcp_skb_pcount(skb
) > 1 && !in_sack
&&
1132 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
1133 mss
= tcp_skb_mss(skb
);
1134 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1137 pkt_len
= start_seq
- TCP_SKB_CB(skb
)->seq
;
1141 pkt_len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1146 /* Round if necessary so that SACKs cover only full MSSes
1147 * and/or the remaining small portion (if present)
1149 if (pkt_len
> mss
) {
1150 unsigned int new_len
= (pkt_len
/ mss
) * mss
;
1151 if (!in_sack
&& new_len
< pkt_len
)
1156 if (pkt_len
>= skb
->len
&& !in_sack
)
1159 err
= tcp_fragment(sk
, TCP_FRAG_IN_RTX_QUEUE
, skb
,
1160 pkt_len
, mss
, GFP_ATOMIC
);
1168 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1169 static u8
tcp_sacktag_one(struct sock
*sk
,
1170 struct tcp_sacktag_state
*state
, u8 sacked
,
1171 u32 start_seq
, u32 end_seq
,
1172 int dup_sack
, int pcount
,
1175 struct tcp_sock
*tp
= tcp_sk(sk
);
1177 /* Account D-SACK for retransmitted packet. */
1178 if (dup_sack
&& (sacked
& TCPCB_RETRANS
)) {
1179 if (tp
->undo_marker
&& tp
->undo_retrans
> 0 &&
1180 after(end_seq
, tp
->undo_marker
))
1182 if ((sacked
& TCPCB_SACKED_ACKED
) &&
1183 before(start_seq
, state
->reord
))
1184 state
->reord
= start_seq
;
1187 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1188 if (!after(end_seq
, tp
->snd_una
))
1191 if (!(sacked
& TCPCB_SACKED_ACKED
)) {
1192 tcp_rack_advance(tp
, sacked
, end_seq
, xmit_time
);
1194 if (sacked
& TCPCB_SACKED_RETRANS
) {
1195 /* If the segment is not tagged as lost,
1196 * we do not clear RETRANS, believing
1197 * that retransmission is still in flight.
1199 if (sacked
& TCPCB_LOST
) {
1200 sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1201 tp
->lost_out
-= pcount
;
1202 tp
->retrans_out
-= pcount
;
1205 if (!(sacked
& TCPCB_RETRANS
)) {
1206 /* New sack for not retransmitted frame,
1207 * which was in hole. It is reordering.
1209 if (before(start_seq
,
1210 tcp_highest_sack_seq(tp
)) &&
1211 before(start_seq
, state
->reord
))
1212 state
->reord
= start_seq
;
1214 if (!after(end_seq
, tp
->high_seq
))
1215 state
->flag
|= FLAG_ORIG_SACK_ACKED
;
1216 if (state
->first_sackt
== 0)
1217 state
->first_sackt
= xmit_time
;
1218 state
->last_sackt
= xmit_time
;
1221 if (sacked
& TCPCB_LOST
) {
1222 sacked
&= ~TCPCB_LOST
;
1223 tp
->lost_out
-= pcount
;
1227 sacked
|= TCPCB_SACKED_ACKED
;
1228 state
->flag
|= FLAG_DATA_SACKED
;
1229 tp
->sacked_out
+= pcount
;
1230 tp
->delivered
+= pcount
; /* Out-of-order packets delivered */
1232 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1233 if (tp
->lost_skb_hint
&&
1234 before(start_seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
))
1235 tp
->lost_cnt_hint
+= pcount
;
1238 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1239 * frames and clear it. undo_retrans is decreased above, L|R frames
1240 * are accounted above as well.
1242 if (dup_sack
&& (sacked
& TCPCB_SACKED_RETRANS
)) {
1243 sacked
&= ~TCPCB_SACKED_RETRANS
;
1244 tp
->retrans_out
-= pcount
;
1250 /* Shift newly-SACKed bytes from this skb to the immediately previous
1251 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1253 static bool tcp_shifted_skb(struct sock
*sk
, struct sk_buff
*prev
,
1254 struct sk_buff
*skb
,
1255 struct tcp_sacktag_state
*state
,
1256 unsigned int pcount
, int shifted
, int mss
,
1259 struct tcp_sock
*tp
= tcp_sk(sk
);
1260 u32 start_seq
= TCP_SKB_CB(skb
)->seq
; /* start of newly-SACKed */
1261 u32 end_seq
= start_seq
+ shifted
; /* end of newly-SACKed */
1265 /* Adjust counters and hints for the newly sacked sequence
1266 * range but discard the return value since prev is already
1267 * marked. We must tag the range first because the seq
1268 * advancement below implicitly advances
1269 * tcp_highest_sack_seq() when skb is highest_sack.
1271 tcp_sacktag_one(sk
, state
, TCP_SKB_CB(skb
)->sacked
,
1272 start_seq
, end_seq
, dup_sack
, pcount
,
1274 tcp_rate_skb_delivered(sk
, skb
, state
->rate
);
1276 if (skb
== tp
->lost_skb_hint
)
1277 tp
->lost_cnt_hint
+= pcount
;
1279 TCP_SKB_CB(prev
)->end_seq
+= shifted
;
1280 TCP_SKB_CB(skb
)->seq
+= shifted
;
1282 tcp_skb_pcount_add(prev
, pcount
);
1283 BUG_ON(tcp_skb_pcount(skb
) < pcount
);
1284 tcp_skb_pcount_add(skb
, -pcount
);
1286 /* When we're adding to gso_segs == 1, gso_size will be zero,
1287 * in theory this shouldn't be necessary but as long as DSACK
1288 * code can come after this skb later on it's better to keep
1289 * setting gso_size to something.
1291 if (!TCP_SKB_CB(prev
)->tcp_gso_size
)
1292 TCP_SKB_CB(prev
)->tcp_gso_size
= mss
;
1294 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1295 if (tcp_skb_pcount(skb
) <= 1)
1296 TCP_SKB_CB(skb
)->tcp_gso_size
= 0;
1298 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1299 TCP_SKB_CB(prev
)->sacked
|= (TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
);
1302 BUG_ON(!tcp_skb_pcount(skb
));
1303 NET_INC_STATS(sock_net(sk
), LINUX_MIB_SACKSHIFTED
);
1307 /* Whole SKB was eaten :-) */
1309 if (skb
== tp
->retransmit_skb_hint
)
1310 tp
->retransmit_skb_hint
= prev
;
1311 if (skb
== tp
->lost_skb_hint
) {
1312 tp
->lost_skb_hint
= prev
;
1313 tp
->lost_cnt_hint
-= tcp_skb_pcount(prev
);
1316 TCP_SKB_CB(prev
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
;
1317 TCP_SKB_CB(prev
)->eor
= TCP_SKB_CB(skb
)->eor
;
1318 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
1319 TCP_SKB_CB(prev
)->end_seq
++;
1321 if (skb
== tcp_highest_sack(sk
))
1322 tcp_advance_highest_sack(sk
, skb
);
1324 tcp_skb_collapse_tstamp(prev
, skb
);
1325 if (unlikely(TCP_SKB_CB(prev
)->tx
.delivered_mstamp
))
1326 TCP_SKB_CB(prev
)->tx
.delivered_mstamp
= 0;
1328 tcp_rtx_queue_unlink_and_free(skb
, sk
);
1330 NET_INC_STATS(sock_net(sk
), LINUX_MIB_SACKMERGED
);
1335 /* I wish gso_size would have a bit more sane initialization than
1336 * something-or-zero which complicates things
1338 static int tcp_skb_seglen(const struct sk_buff
*skb
)
1340 return tcp_skb_pcount(skb
) == 1 ? skb
->len
: tcp_skb_mss(skb
);
1343 /* Shifting pages past head area doesn't work */
1344 static int skb_can_shift(const struct sk_buff
*skb
)
1346 return !skb_headlen(skb
) && skb_is_nonlinear(skb
);
1349 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1352 static struct sk_buff
*tcp_shift_skb_data(struct sock
*sk
, struct sk_buff
*skb
,
1353 struct tcp_sacktag_state
*state
,
1354 u32 start_seq
, u32 end_seq
,
1357 struct tcp_sock
*tp
= tcp_sk(sk
);
1358 struct sk_buff
*prev
;
1364 if (!sk_can_gso(sk
))
1367 /* Normally R but no L won't result in plain S */
1369 (TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_RETRANS
)) == TCPCB_SACKED_RETRANS
)
1371 if (!skb_can_shift(skb
))
1373 /* This frame is about to be dropped (was ACKed). */
1374 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1377 /* Can only happen with delayed DSACK + discard craziness */
1378 prev
= skb_rb_prev(skb
);
1382 if ((TCP_SKB_CB(prev
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
)
1385 if (!tcp_skb_can_collapse_to(prev
))
1388 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1389 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1393 pcount
= tcp_skb_pcount(skb
);
1394 mss
= tcp_skb_seglen(skb
);
1396 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1397 * drop this restriction as unnecessary
1399 if (mss
!= tcp_skb_seglen(prev
))
1402 if (!after(TCP_SKB_CB(skb
)->end_seq
, start_seq
))
1404 /* CHECKME: This is non-MSS split case only?, this will
1405 * cause skipped skbs due to advancing loop btw, original
1406 * has that feature too
1408 if (tcp_skb_pcount(skb
) <= 1)
1411 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1413 /* TODO: head merge to next could be attempted here
1414 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1415 * though it might not be worth of the additional hassle
1417 * ...we can probably just fallback to what was done
1418 * previously. We could try merging non-SACKed ones
1419 * as well but it probably isn't going to buy off
1420 * because later SACKs might again split them, and
1421 * it would make skb timestamp tracking considerably
1427 len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1429 BUG_ON(len
> skb
->len
);
1431 /* MSS boundaries should be honoured or else pcount will
1432 * severely break even though it makes things bit trickier.
1433 * Optimize common case to avoid most of the divides
1435 mss
= tcp_skb_mss(skb
);
1437 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1438 * drop this restriction as unnecessary
1440 if (mss
!= tcp_skb_seglen(prev
))
1445 } else if (len
< mss
) {
1453 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1454 if (!after(TCP_SKB_CB(skb
)->seq
+ len
, tp
->snd_una
))
1457 if (!skb_shift(prev
, skb
, len
))
1459 if (!tcp_shifted_skb(sk
, prev
, skb
, state
, pcount
, len
, mss
, dup_sack
))
1462 /* Hole filled allows collapsing with the next as well, this is very
1463 * useful when hole on every nth skb pattern happens
1465 skb
= skb_rb_next(prev
);
1469 if (!skb_can_shift(skb
) ||
1470 ((TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
) ||
1471 (mss
!= tcp_skb_seglen(skb
)))
1475 if (skb_shift(prev
, skb
, len
)) {
1476 pcount
+= tcp_skb_pcount(skb
);
1477 tcp_shifted_skb(sk
, prev
, skb
, state
, tcp_skb_pcount(skb
),
1488 NET_INC_STATS(sock_net(sk
), LINUX_MIB_SACKSHIFTFALLBACK
);
1492 static struct sk_buff
*tcp_sacktag_walk(struct sk_buff
*skb
, struct sock
*sk
,
1493 struct tcp_sack_block
*next_dup
,
1494 struct tcp_sacktag_state
*state
,
1495 u32 start_seq
, u32 end_seq
,
1498 struct tcp_sock
*tp
= tcp_sk(sk
);
1499 struct sk_buff
*tmp
;
1501 skb_rbtree_walk_from(skb
) {
1503 bool dup_sack
= dup_sack_in
;
1505 /* queue is in-order => we can short-circuit the walk early */
1506 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1510 before(TCP_SKB_CB(skb
)->seq
, next_dup
->end_seq
)) {
1511 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1512 next_dup
->start_seq
,
1518 /* skb reference here is a bit tricky to get right, since
1519 * shifting can eat and free both this skb and the next,
1520 * so not even _safe variant of the loop is enough.
1523 tmp
= tcp_shift_skb_data(sk
, skb
, state
,
1524 start_seq
, end_seq
, dup_sack
);
1533 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1539 if (unlikely(in_sack
< 0))
1543 TCP_SKB_CB(skb
)->sacked
=
1546 TCP_SKB_CB(skb
)->sacked
,
1547 TCP_SKB_CB(skb
)->seq
,
1548 TCP_SKB_CB(skb
)->end_seq
,
1550 tcp_skb_pcount(skb
),
1552 tcp_rate_skb_delivered(sk
, skb
, state
->rate
);
1553 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
1554 list_del_init(&skb
->tcp_tsorted_anchor
);
1556 if (!before(TCP_SKB_CB(skb
)->seq
,
1557 tcp_highest_sack_seq(tp
)))
1558 tcp_advance_highest_sack(sk
, skb
);
1564 static struct sk_buff
*tcp_sacktag_bsearch(struct sock
*sk
,
1565 struct tcp_sacktag_state
*state
,
1568 struct rb_node
*parent
, **p
= &sk
->tcp_rtx_queue
.rb_node
;
1569 struct sk_buff
*skb
;
1573 skb
= rb_to_skb(parent
);
1574 if (before(seq
, TCP_SKB_CB(skb
)->seq
)) {
1575 p
= &parent
->rb_left
;
1578 if (!before(seq
, TCP_SKB_CB(skb
)->end_seq
)) {
1579 p
= &parent
->rb_right
;
1587 static struct sk_buff
*tcp_sacktag_skip(struct sk_buff
*skb
, struct sock
*sk
,
1588 struct tcp_sacktag_state
*state
,
1591 if (skb
&& after(TCP_SKB_CB(skb
)->seq
, skip_to_seq
))
1594 return tcp_sacktag_bsearch(sk
, state
, skip_to_seq
);
1597 static struct sk_buff
*tcp_maybe_skipping_dsack(struct sk_buff
*skb
,
1599 struct tcp_sack_block
*next_dup
,
1600 struct tcp_sacktag_state
*state
,
1606 if (before(next_dup
->start_seq
, skip_to_seq
)) {
1607 skb
= tcp_sacktag_skip(skb
, sk
, state
, next_dup
->start_seq
);
1608 skb
= tcp_sacktag_walk(skb
, sk
, NULL
, state
,
1609 next_dup
->start_seq
, next_dup
->end_seq
,
1616 static int tcp_sack_cache_ok(const struct tcp_sock
*tp
, const struct tcp_sack_block
*cache
)
1618 return cache
< tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1622 tcp_sacktag_write_queue(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1623 u32 prior_snd_una
, struct tcp_sacktag_state
*state
)
1625 struct tcp_sock
*tp
= tcp_sk(sk
);
1626 const unsigned char *ptr
= (skb_transport_header(ack_skb
) +
1627 TCP_SKB_CB(ack_skb
)->sacked
);
1628 struct tcp_sack_block_wire
*sp_wire
= (struct tcp_sack_block_wire
*)(ptr
+2);
1629 struct tcp_sack_block sp
[TCP_NUM_SACKS
];
1630 struct tcp_sack_block
*cache
;
1631 struct sk_buff
*skb
;
1632 int num_sacks
= min(TCP_NUM_SACKS
, (ptr
[1] - TCPOLEN_SACK_BASE
) >> 3);
1634 bool found_dup_sack
= false;
1636 int first_sack_index
;
1639 state
->reord
= tp
->snd_nxt
;
1641 if (!tp
->sacked_out
)
1642 tcp_highest_sack_reset(sk
);
1644 found_dup_sack
= tcp_check_dsack(sk
, ack_skb
, sp_wire
,
1645 num_sacks
, prior_snd_una
);
1646 if (found_dup_sack
) {
1647 state
->flag
|= FLAG_DSACKING_ACK
;
1648 tp
->delivered
++; /* A spurious retransmission is delivered */
1651 /* Eliminate too old ACKs, but take into
1652 * account more or less fresh ones, they can
1653 * contain valid SACK info.
1655 if (before(TCP_SKB_CB(ack_skb
)->ack_seq
, prior_snd_una
- tp
->max_window
))
1658 if (!tp
->packets_out
)
1662 first_sack_index
= 0;
1663 for (i
= 0; i
< num_sacks
; i
++) {
1664 bool dup_sack
= !i
&& found_dup_sack
;
1666 sp
[used_sacks
].start_seq
= get_unaligned_be32(&sp_wire
[i
].start_seq
);
1667 sp
[used_sacks
].end_seq
= get_unaligned_be32(&sp_wire
[i
].end_seq
);
1669 if (!tcp_is_sackblock_valid(tp
, dup_sack
,
1670 sp
[used_sacks
].start_seq
,
1671 sp
[used_sacks
].end_seq
)) {
1675 if (!tp
->undo_marker
)
1676 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDNOUNDO
;
1678 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDOLD
;
1680 /* Don't count olds caused by ACK reordering */
1681 if ((TCP_SKB_CB(ack_skb
)->ack_seq
!= tp
->snd_una
) &&
1682 !after(sp
[used_sacks
].end_seq
, tp
->snd_una
))
1684 mib_idx
= LINUX_MIB_TCPSACKDISCARD
;
1687 NET_INC_STATS(sock_net(sk
), mib_idx
);
1689 first_sack_index
= -1;
1693 /* Ignore very old stuff early */
1694 if (!after(sp
[used_sacks
].end_seq
, prior_snd_una
))
1700 /* order SACK blocks to allow in order walk of the retrans queue */
1701 for (i
= used_sacks
- 1; i
> 0; i
--) {
1702 for (j
= 0; j
< i
; j
++) {
1703 if (after(sp
[j
].start_seq
, sp
[j
+ 1].start_seq
)) {
1704 swap(sp
[j
], sp
[j
+ 1]);
1706 /* Track where the first SACK block goes to */
1707 if (j
== first_sack_index
)
1708 first_sack_index
= j
+ 1;
1713 state
->mss_now
= tcp_current_mss(sk
);
1717 if (!tp
->sacked_out
) {
1718 /* It's already past, so skip checking against it */
1719 cache
= tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1721 cache
= tp
->recv_sack_cache
;
1722 /* Skip empty blocks in at head of the cache */
1723 while (tcp_sack_cache_ok(tp
, cache
) && !cache
->start_seq
&&
1728 while (i
< used_sacks
) {
1729 u32 start_seq
= sp
[i
].start_seq
;
1730 u32 end_seq
= sp
[i
].end_seq
;
1731 bool dup_sack
= (found_dup_sack
&& (i
== first_sack_index
));
1732 struct tcp_sack_block
*next_dup
= NULL
;
1734 if (found_dup_sack
&& ((i
+ 1) == first_sack_index
))
1735 next_dup
= &sp
[i
+ 1];
1737 /* Skip too early cached blocks */
1738 while (tcp_sack_cache_ok(tp
, cache
) &&
1739 !before(start_seq
, cache
->end_seq
))
1742 /* Can skip some work by looking recv_sack_cache? */
1743 if (tcp_sack_cache_ok(tp
, cache
) && !dup_sack
&&
1744 after(end_seq
, cache
->start_seq
)) {
1747 if (before(start_seq
, cache
->start_seq
)) {
1748 skb
= tcp_sacktag_skip(skb
, sk
, state
,
1750 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
,
1757 /* Rest of the block already fully processed? */
1758 if (!after(end_seq
, cache
->end_seq
))
1761 skb
= tcp_maybe_skipping_dsack(skb
, sk
, next_dup
,
1765 /* ...tail remains todo... */
1766 if (tcp_highest_sack_seq(tp
) == cache
->end_seq
) {
1767 /* ...but better entrypoint exists! */
1768 skb
= tcp_highest_sack(sk
);
1775 skb
= tcp_sacktag_skip(skb
, sk
, state
, cache
->end_seq
);
1776 /* Check overlap against next cached too (past this one already) */
1781 if (!before(start_seq
, tcp_highest_sack_seq(tp
))) {
1782 skb
= tcp_highest_sack(sk
);
1786 skb
= tcp_sacktag_skip(skb
, sk
, state
, start_seq
);
1789 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
, state
,
1790 start_seq
, end_seq
, dup_sack
);
1796 /* Clear the head of the cache sack blocks so we can skip it next time */
1797 for (i
= 0; i
< ARRAY_SIZE(tp
->recv_sack_cache
) - used_sacks
; i
++) {
1798 tp
->recv_sack_cache
[i
].start_seq
= 0;
1799 tp
->recv_sack_cache
[i
].end_seq
= 0;
1801 for (j
= 0; j
< used_sacks
; j
++)
1802 tp
->recv_sack_cache
[i
++] = sp
[j
];
1804 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Loss
|| tp
->undo_marker
)
1805 tcp_check_sack_reordering(sk
, state
->reord
, 0);
1807 tcp_verify_left_out(tp
);
1810 #if FASTRETRANS_DEBUG > 0
1811 WARN_ON((int)tp
->sacked_out
< 0);
1812 WARN_ON((int)tp
->lost_out
< 0);
1813 WARN_ON((int)tp
->retrans_out
< 0);
1814 WARN_ON((int)tcp_packets_in_flight(tp
) < 0);
1819 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1820 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1822 static bool tcp_limit_reno_sacked(struct tcp_sock
*tp
)
1826 holes
= max(tp
->lost_out
, 1U);
1827 holes
= min(holes
, tp
->packets_out
);
1829 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1830 tp
->sacked_out
= tp
->packets_out
- holes
;
1836 /* If we receive more dupacks than we expected counting segments
1837 * in assumption of absent reordering, interpret this as reordering.
1838 * The only another reason could be bug in receiver TCP.
1840 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1842 struct tcp_sock
*tp
= tcp_sk(sk
);
1844 if (!tcp_limit_reno_sacked(tp
))
1847 tp
->reordering
= min_t(u32
, tp
->packets_out
+ addend
,
1848 sock_net(sk
)->ipv4
.sysctl_tcp_max_reordering
);
1849 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPRENOREORDER
);
1852 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1854 static void tcp_add_reno_sack(struct sock
*sk
)
1856 struct tcp_sock
*tp
= tcp_sk(sk
);
1857 u32 prior_sacked
= tp
->sacked_out
;
1860 tcp_check_reno_reordering(sk
, 0);
1861 if (tp
->sacked_out
> prior_sacked
)
1862 tp
->delivered
++; /* Some out-of-order packet is delivered */
1863 tcp_verify_left_out(tp
);
1866 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1868 static void tcp_remove_reno_sacks(struct sock
*sk
, int acked
)
1870 struct tcp_sock
*tp
= tcp_sk(sk
);
1873 /* One ACK acked hole. The rest eat duplicate ACKs. */
1874 tp
->delivered
+= max_t(int, acked
- tp
->sacked_out
, 1);
1875 if (acked
- 1 >= tp
->sacked_out
)
1878 tp
->sacked_out
-= acked
- 1;
1880 tcp_check_reno_reordering(sk
, acked
);
1881 tcp_verify_left_out(tp
);
1884 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1889 void tcp_clear_retrans(struct tcp_sock
*tp
)
1891 tp
->retrans_out
= 0;
1893 tp
->undo_marker
= 0;
1894 tp
->undo_retrans
= -1;
1898 static inline void tcp_init_undo(struct tcp_sock
*tp
)
1900 tp
->undo_marker
= tp
->snd_una
;
1901 /* Retransmission still in flight may cause DSACKs later. */
1902 tp
->undo_retrans
= tp
->retrans_out
? : -1;
1905 /* Enter Loss state. If we detect SACK reneging, forget all SACK information
1906 * and reset tags completely, otherwise preserve SACKs. If receiver
1907 * dropped its ofo queue, we will know this due to reneging detection.
1909 void tcp_enter_loss(struct sock
*sk
)
1911 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1912 struct tcp_sock
*tp
= tcp_sk(sk
);
1913 struct net
*net
= sock_net(sk
);
1914 struct sk_buff
*skb
;
1915 bool new_recovery
= icsk
->icsk_ca_state
< TCP_CA_Recovery
;
1916 bool is_reneg
; /* is receiver reneging on SACKs? */
1919 /* Reduce ssthresh if it has not yet been made inside this window. */
1920 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
||
1921 !after(tp
->high_seq
, tp
->snd_una
) ||
1922 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
1923 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1924 tp
->prior_cwnd
= tp
->snd_cwnd
;
1925 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1926 tcp_ca_event(sk
, CA_EVENT_LOSS
);
1930 tp
->snd_cwnd_cnt
= 0;
1931 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
1933 tp
->retrans_out
= 0;
1936 if (tcp_is_reno(tp
))
1937 tcp_reset_reno_sack(tp
);
1939 skb
= tcp_rtx_queue_head(sk
);
1940 is_reneg
= skb
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
);
1942 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPSACKRENEGING
);
1945 tcp_clear_all_retrans_hints(tp
);
1947 skb_rbtree_walk_from(skb
) {
1948 mark_lost
= (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
) ||
1951 tcp_sum_lost(tp
, skb
);
1952 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
1954 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
1955 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1956 tp
->lost_out
+= tcp_skb_pcount(skb
);
1959 tcp_verify_left_out(tp
);
1961 /* Timeout in disordered state after receiving substantial DUPACKs
1962 * suggests that the degree of reordering is over-estimated.
1964 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
&&
1965 tp
->sacked_out
>= net
->ipv4
.sysctl_tcp_reordering
)
1966 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1967 net
->ipv4
.sysctl_tcp_reordering
);
1968 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1969 tp
->high_seq
= tp
->snd_nxt
;
1970 tcp_ecn_queue_cwr(tp
);
1972 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1973 * loss recovery is underway except recurring timeout(s) on
1974 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1976 * In theory F-RTO can be used repeatedly during loss recovery.
1977 * In practice this interacts badly with broken middle-boxes that
1978 * falsely raise the receive window, which results in repeated
1979 * timeouts and stop-and-go behavior.
1981 tp
->frto
= net
->ipv4
.sysctl_tcp_frto
&&
1982 (new_recovery
|| icsk
->icsk_retransmits
) &&
1983 !inet_csk(sk
)->icsk_mtup
.probe_size
;
1986 /* If ACK arrived pointing to a remembered SACK, it means that our
1987 * remembered SACKs do not reflect real state of receiver i.e.
1988 * receiver _host_ is heavily congested (or buggy).
1990 * To avoid big spurious retransmission bursts due to transient SACK
1991 * scoreboard oddities that look like reneging, we give the receiver a
1992 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
1993 * restore sanity to the SACK scoreboard. If the apparent reneging
1994 * persists until this RTO then we'll clear the SACK scoreboard.
1996 static bool tcp_check_sack_reneging(struct sock
*sk
, int flag
)
1998 if (flag
& FLAG_SACK_RENEGING
) {
1999 struct tcp_sock
*tp
= tcp_sk(sk
);
2000 unsigned long delay
= max(usecs_to_jiffies(tp
->srtt_us
>> 4),
2001 msecs_to_jiffies(10));
2003 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2004 delay
, TCP_RTO_MAX
);
2010 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2011 * counter when SACK is enabled (without SACK, sacked_out is used for
2014 * With reordering, holes may still be in flight, so RFC3517 recovery
2015 * uses pure sacked_out (total number of SACKed segments) even though
2016 * it violates the RFC that uses duplicate ACKs, often these are equal
2017 * but when e.g. out-of-window ACKs or packet duplication occurs,
2018 * they differ. Since neither occurs due to loss, TCP should really
2021 static inline int tcp_dupack_heuristics(const struct tcp_sock
*tp
)
2023 return tp
->sacked_out
+ 1;
2026 /* Linux NewReno/SACK/ECN state machine.
2027 * --------------------------------------
2029 * "Open" Normal state, no dubious events, fast path.
2030 * "Disorder" In all the respects it is "Open",
2031 * but requires a bit more attention. It is entered when
2032 * we see some SACKs or dupacks. It is split of "Open"
2033 * mainly to move some processing from fast path to slow one.
2034 * "CWR" CWND was reduced due to some Congestion Notification event.
2035 * It can be ECN, ICMP source quench, local device congestion.
2036 * "Recovery" CWND was reduced, we are fast-retransmitting.
2037 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2039 * tcp_fastretrans_alert() is entered:
2040 * - each incoming ACK, if state is not "Open"
2041 * - when arrived ACK is unusual, namely:
2046 * Counting packets in flight is pretty simple.
2048 * in_flight = packets_out - left_out + retrans_out
2050 * packets_out is SND.NXT-SND.UNA counted in packets.
2052 * retrans_out is number of retransmitted segments.
2054 * left_out is number of segments left network, but not ACKed yet.
2056 * left_out = sacked_out + lost_out
2058 * sacked_out: Packets, which arrived to receiver out of order
2059 * and hence not ACKed. With SACKs this number is simply
2060 * amount of SACKed data. Even without SACKs
2061 * it is easy to give pretty reliable estimate of this number,
2062 * counting duplicate ACKs.
2064 * lost_out: Packets lost by network. TCP has no explicit
2065 * "loss notification" feedback from network (for now).
2066 * It means that this number can be only _guessed_.
2067 * Actually, it is the heuristics to predict lossage that
2068 * distinguishes different algorithms.
2070 * F.e. after RTO, when all the queue is considered as lost,
2071 * lost_out = packets_out and in_flight = retrans_out.
2073 * Essentially, we have now a few algorithms detecting
2076 * If the receiver supports SACK:
2078 * RFC6675/3517: It is the conventional algorithm. A packet is
2079 * considered lost if the number of higher sequence packets
2080 * SACKed is greater than or equal the DUPACK thoreshold
2081 * (reordering). This is implemented in tcp_mark_head_lost and
2082 * tcp_update_scoreboard.
2084 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm
2085 * (2017-) that checks timing instead of counting DUPACKs.
2086 * Essentially a packet is considered lost if it's not S/ACKed
2087 * after RTT + reordering_window, where both metrics are
2088 * dynamically measured and adjusted. This is implemented in
2089 * tcp_rack_mark_lost.
2091 * If the receiver does not support SACK:
2093 * NewReno (RFC6582): in Recovery we assume that one segment
2094 * is lost (classic Reno). While we are in Recovery and
2095 * a partial ACK arrives, we assume that one more packet
2096 * is lost (NewReno). This heuristics are the same in NewReno
2099 * Really tricky (and requiring careful tuning) part of algorithm
2100 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2101 * The first determines the moment _when_ we should reduce CWND and,
2102 * hence, slow down forward transmission. In fact, it determines the moment
2103 * when we decide that hole is caused by loss, rather than by a reorder.
2105 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2106 * holes, caused by lost packets.
2108 * And the most logically complicated part of algorithm is undo
2109 * heuristics. We detect false retransmits due to both too early
2110 * fast retransmit (reordering) and underestimated RTO, analyzing
2111 * timestamps and D-SACKs. When we detect that some segments were
2112 * retransmitted by mistake and CWND reduction was wrong, we undo
2113 * window reduction and abort recovery phase. This logic is hidden
2114 * inside several functions named tcp_try_undo_<something>.
2117 /* This function decides, when we should leave Disordered state
2118 * and enter Recovery phase, reducing congestion window.
2120 * Main question: may we further continue forward transmission
2121 * with the same cwnd?
2123 static bool tcp_time_to_recover(struct sock
*sk
, int flag
)
2125 struct tcp_sock
*tp
= tcp_sk(sk
);
2127 /* Trick#1: The loss is proven. */
2131 /* Not-A-Trick#2 : Classic rule... */
2132 if (tcp_dupack_heuristics(tp
) > tp
->reordering
)
2138 /* Detect loss in event "A" above by marking head of queue up as lost.
2139 * For non-SACK(Reno) senders, the first "packets" number of segments
2140 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2141 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2142 * the maximum SACKed segments to pass before reaching this limit.
2144 static void tcp_mark_head_lost(struct sock
*sk
, int packets
, int mark_head
)
2146 struct tcp_sock
*tp
= tcp_sk(sk
);
2147 struct sk_buff
*skb
;
2148 int cnt
, oldcnt
, lost
;
2150 /* Use SACK to deduce losses of new sequences sent during recovery */
2151 const u32 loss_high
= tcp_is_sack(tp
) ? tp
->snd_nxt
: tp
->high_seq
;
2153 WARN_ON(packets
> tp
->packets_out
);
2154 skb
= tp
->lost_skb_hint
;
2156 /* Head already handled? */
2157 if (mark_head
&& after(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
))
2159 cnt
= tp
->lost_cnt_hint
;
2161 skb
= tcp_rtx_queue_head(sk
);
2165 skb_rbtree_walk_from(skb
) {
2166 /* TODO: do this better */
2167 /* this is not the most efficient way to do this... */
2168 tp
->lost_skb_hint
= skb
;
2169 tp
->lost_cnt_hint
= cnt
;
2171 if (after(TCP_SKB_CB(skb
)->end_seq
, loss_high
))
2175 if (tcp_is_reno(tp
) ||
2176 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2177 cnt
+= tcp_skb_pcount(skb
);
2179 if (cnt
> packets
) {
2180 if (tcp_is_sack(tp
) ||
2181 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
) ||
2182 (oldcnt
>= packets
))
2185 mss
= tcp_skb_mss(skb
);
2186 /* If needed, chop off the prefix to mark as lost. */
2187 lost
= (packets
- oldcnt
) * mss
;
2188 if (lost
< skb
->len
&&
2189 tcp_fragment(sk
, TCP_FRAG_IN_RTX_QUEUE
, skb
,
2190 lost
, mss
, GFP_ATOMIC
) < 0)
2195 tcp_skb_mark_lost(tp
, skb
);
2200 tcp_verify_left_out(tp
);
2203 /* Account newly detected lost packet(s) */
2205 static void tcp_update_scoreboard(struct sock
*sk
, int fast_rexmit
)
2207 struct tcp_sock
*tp
= tcp_sk(sk
);
2209 if (tcp_is_reno(tp
)) {
2210 tcp_mark_head_lost(sk
, 1, 1);
2212 int sacked_upto
= tp
->sacked_out
- tp
->reordering
;
2213 if (sacked_upto
>= 0)
2214 tcp_mark_head_lost(sk
, sacked_upto
, 0);
2215 else if (fast_rexmit
)
2216 tcp_mark_head_lost(sk
, 1, 1);
2220 static bool tcp_tsopt_ecr_before(const struct tcp_sock
*tp
, u32 when
)
2222 return tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2223 before(tp
->rx_opt
.rcv_tsecr
, when
);
2226 /* skb is spurious retransmitted if the returned timestamp echo
2227 * reply is prior to the skb transmission time
2229 static bool tcp_skb_spurious_retrans(const struct tcp_sock
*tp
,
2230 const struct sk_buff
*skb
)
2232 return (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
) &&
2233 tcp_tsopt_ecr_before(tp
, tcp_skb_timestamp(skb
));
2236 /* Nothing was retransmitted or returned timestamp is less
2237 * than timestamp of the first retransmission.
2239 static inline bool tcp_packet_delayed(const struct tcp_sock
*tp
)
2241 return !tp
->retrans_stamp
||
2242 tcp_tsopt_ecr_before(tp
, tp
->retrans_stamp
);
2245 /* Undo procedures. */
2247 /* We can clear retrans_stamp when there are no retransmissions in the
2248 * window. It would seem that it is trivially available for us in
2249 * tp->retrans_out, however, that kind of assumptions doesn't consider
2250 * what will happen if errors occur when sending retransmission for the
2251 * second time. ...It could the that such segment has only
2252 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2253 * the head skb is enough except for some reneging corner cases that
2254 * are not worth the effort.
2256 * Main reason for all this complexity is the fact that connection dying
2257 * time now depends on the validity of the retrans_stamp, in particular,
2258 * that successive retransmissions of a segment must not advance
2259 * retrans_stamp under any conditions.
2261 static bool tcp_any_retrans_done(const struct sock
*sk
)
2263 const struct tcp_sock
*tp
= tcp_sk(sk
);
2264 struct sk_buff
*skb
;
2266 if (tp
->retrans_out
)
2269 skb
= tcp_rtx_queue_head(sk
);
2270 if (unlikely(skb
&& TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
))
2276 static void DBGUNDO(struct sock
*sk
, const char *msg
)
2278 #if FASTRETRANS_DEBUG > 1
2279 struct tcp_sock
*tp
= tcp_sk(sk
);
2280 struct inet_sock
*inet
= inet_sk(sk
);
2282 if (sk
->sk_family
== AF_INET
) {
2283 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2285 &inet
->inet_daddr
, ntohs(inet
->inet_dport
),
2286 tp
->snd_cwnd
, tcp_left_out(tp
),
2287 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2290 #if IS_ENABLED(CONFIG_IPV6)
2291 else if (sk
->sk_family
== AF_INET6
) {
2292 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2294 &sk
->sk_v6_daddr
, ntohs(inet
->inet_dport
),
2295 tp
->snd_cwnd
, tcp_left_out(tp
),
2296 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2303 static void tcp_undo_cwnd_reduction(struct sock
*sk
, bool unmark_loss
)
2305 struct tcp_sock
*tp
= tcp_sk(sk
);
2308 struct sk_buff
*skb
;
2310 skb_rbtree_walk(skb
, &sk
->tcp_rtx_queue
) {
2311 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2314 tcp_clear_all_retrans_hints(tp
);
2317 if (tp
->prior_ssthresh
) {
2318 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2320 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
2322 if (tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
2323 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
2324 tcp_ecn_withdraw_cwr(tp
);
2327 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
2328 tp
->undo_marker
= 0;
2331 static inline bool tcp_may_undo(const struct tcp_sock
*tp
)
2333 return tp
->undo_marker
&& (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
2336 /* People celebrate: "We love our President!" */
2337 static bool tcp_try_undo_recovery(struct sock
*sk
)
2339 struct tcp_sock
*tp
= tcp_sk(sk
);
2341 if (tcp_may_undo(tp
)) {
2344 /* Happy end! We did not retransmit anything
2345 * or our original transmission succeeded.
2347 DBGUNDO(sk
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
2348 tcp_undo_cwnd_reduction(sk
, false);
2349 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
2350 mib_idx
= LINUX_MIB_TCPLOSSUNDO
;
2352 mib_idx
= LINUX_MIB_TCPFULLUNDO
;
2354 NET_INC_STATS(sock_net(sk
), mib_idx
);
2355 } else if (tp
->rack
.reo_wnd_persist
) {
2356 tp
->rack
.reo_wnd_persist
--;
2358 if (tp
->snd_una
== tp
->high_seq
&& tcp_is_reno(tp
)) {
2359 /* Hold old state until something *above* high_seq
2360 * is ACKed. For Reno it is MUST to prevent false
2361 * fast retransmits (RFC2582). SACK TCP is safe. */
2362 if (!tcp_any_retrans_done(sk
))
2363 tp
->retrans_stamp
= 0;
2366 tcp_set_ca_state(sk
, TCP_CA_Open
);
2370 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2371 static bool tcp_try_undo_dsack(struct sock
*sk
)
2373 struct tcp_sock
*tp
= tcp_sk(sk
);
2375 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
2376 tp
->rack
.reo_wnd_persist
= min(TCP_RACK_RECOVERY_THRESH
,
2377 tp
->rack
.reo_wnd_persist
+ 1);
2378 DBGUNDO(sk
, "D-SACK");
2379 tcp_undo_cwnd_reduction(sk
, false);
2380 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPDSACKUNDO
);
2386 /* Undo during loss recovery after partial ACK or using F-RTO. */
2387 static bool tcp_try_undo_loss(struct sock
*sk
, bool frto_undo
)
2389 struct tcp_sock
*tp
= tcp_sk(sk
);
2391 if (frto_undo
|| tcp_may_undo(tp
)) {
2392 tcp_undo_cwnd_reduction(sk
, true);
2394 DBGUNDO(sk
, "partial loss");
2395 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPLOSSUNDO
);
2397 NET_INC_STATS(sock_net(sk
),
2398 LINUX_MIB_TCPSPURIOUSRTOS
);
2399 inet_csk(sk
)->icsk_retransmits
= 0;
2400 if (frto_undo
|| tcp_is_sack(tp
))
2401 tcp_set_ca_state(sk
, TCP_CA_Open
);
2407 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2408 * It computes the number of packets to send (sndcnt) based on packets newly
2410 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2411 * cwnd reductions across a full RTT.
2412 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2413 * But when the retransmits are acked without further losses, PRR
2414 * slow starts cwnd up to ssthresh to speed up the recovery.
2416 static void tcp_init_cwnd_reduction(struct sock
*sk
)
2418 struct tcp_sock
*tp
= tcp_sk(sk
);
2420 tp
->high_seq
= tp
->snd_nxt
;
2421 tp
->tlp_high_seq
= 0;
2422 tp
->snd_cwnd_cnt
= 0;
2423 tp
->prior_cwnd
= tp
->snd_cwnd
;
2424 tp
->prr_delivered
= 0;
2426 tp
->snd_ssthresh
= inet_csk(sk
)->icsk_ca_ops
->ssthresh(sk
);
2427 tcp_ecn_queue_cwr(tp
);
2430 void tcp_cwnd_reduction(struct sock
*sk
, int newly_acked_sacked
, int flag
)
2432 struct tcp_sock
*tp
= tcp_sk(sk
);
2434 int delta
= tp
->snd_ssthresh
- tcp_packets_in_flight(tp
);
2436 if (newly_acked_sacked
<= 0 || WARN_ON_ONCE(!tp
->prior_cwnd
))
2439 tp
->prr_delivered
+= newly_acked_sacked
;
2441 u64 dividend
= (u64
)tp
->snd_ssthresh
* tp
->prr_delivered
+
2443 sndcnt
= div_u64(dividend
, tp
->prior_cwnd
) - tp
->prr_out
;
2444 } else if ((flag
& FLAG_RETRANS_DATA_ACKED
) &&
2445 !(flag
& FLAG_LOST_RETRANS
)) {
2446 sndcnt
= min_t(int, delta
,
2447 max_t(int, tp
->prr_delivered
- tp
->prr_out
,
2448 newly_acked_sacked
) + 1);
2450 sndcnt
= min(delta
, newly_acked_sacked
);
2452 /* Force a fast retransmit upon entering fast recovery */
2453 sndcnt
= max(sndcnt
, (tp
->prr_out
? 0 : 1));
2454 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + sndcnt
;
2457 static inline void tcp_end_cwnd_reduction(struct sock
*sk
)
2459 struct tcp_sock
*tp
= tcp_sk(sk
);
2461 if (inet_csk(sk
)->icsk_ca_ops
->cong_control
)
2464 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2465 if (tp
->snd_ssthresh
< TCP_INFINITE_SSTHRESH
&&
2466 (inet_csk(sk
)->icsk_ca_state
== TCP_CA_CWR
|| tp
->undo_marker
)) {
2467 tp
->snd_cwnd
= tp
->snd_ssthresh
;
2468 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
2470 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
2473 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2474 void tcp_enter_cwr(struct sock
*sk
)
2476 struct tcp_sock
*tp
= tcp_sk(sk
);
2478 tp
->prior_ssthresh
= 0;
2479 if (inet_csk(sk
)->icsk_ca_state
< TCP_CA_CWR
) {
2480 tp
->undo_marker
= 0;
2481 tcp_init_cwnd_reduction(sk
);
2482 tcp_set_ca_state(sk
, TCP_CA_CWR
);
2485 EXPORT_SYMBOL(tcp_enter_cwr
);
2487 static void tcp_try_keep_open(struct sock
*sk
)
2489 struct tcp_sock
*tp
= tcp_sk(sk
);
2490 int state
= TCP_CA_Open
;
2492 if (tcp_left_out(tp
) || tcp_any_retrans_done(sk
))
2493 state
= TCP_CA_Disorder
;
2495 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
2496 tcp_set_ca_state(sk
, state
);
2497 tp
->high_seq
= tp
->snd_nxt
;
2501 static void tcp_try_to_open(struct sock
*sk
, int flag
)
2503 struct tcp_sock
*tp
= tcp_sk(sk
);
2505 tcp_verify_left_out(tp
);
2507 if (!tcp_any_retrans_done(sk
))
2508 tp
->retrans_stamp
= 0;
2510 if (flag
& FLAG_ECE
)
2513 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
2514 tcp_try_keep_open(sk
);
2518 static void tcp_mtup_probe_failed(struct sock
*sk
)
2520 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2522 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
2523 icsk
->icsk_mtup
.probe_size
= 0;
2524 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPMTUPFAIL
);
2527 static void tcp_mtup_probe_success(struct sock
*sk
)
2529 struct tcp_sock
*tp
= tcp_sk(sk
);
2530 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2532 /* FIXME: breaks with very large cwnd */
2533 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2534 tp
->snd_cwnd
= tp
->snd_cwnd
*
2535 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
2536 icsk
->icsk_mtup
.probe_size
;
2537 tp
->snd_cwnd_cnt
= 0;
2538 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
2539 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2541 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
2542 icsk
->icsk_mtup
.probe_size
= 0;
2543 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
2544 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPMTUPSUCCESS
);
2547 /* Do a simple retransmit without using the backoff mechanisms in
2548 * tcp_timer. This is used for path mtu discovery.
2549 * The socket is already locked here.
2551 void tcp_simple_retransmit(struct sock
*sk
)
2553 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2554 struct tcp_sock
*tp
= tcp_sk(sk
);
2555 struct sk_buff
*skb
;
2556 unsigned int mss
= tcp_current_mss(sk
);
2558 skb_rbtree_walk(skb
, &sk
->tcp_rtx_queue
) {
2559 if (tcp_skb_seglen(skb
) > mss
&&
2560 !(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
2561 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2562 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
2563 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2565 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
2569 tcp_clear_retrans_hints_partial(tp
);
2574 if (tcp_is_reno(tp
))
2575 tcp_limit_reno_sacked(tp
);
2577 tcp_verify_left_out(tp
);
2579 /* Don't muck with the congestion window here.
2580 * Reason is that we do not increase amount of _data_
2581 * in network, but units changed and effective
2582 * cwnd/ssthresh really reduced now.
2584 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
2585 tp
->high_seq
= tp
->snd_nxt
;
2586 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2587 tp
->prior_ssthresh
= 0;
2588 tp
->undo_marker
= 0;
2589 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2591 tcp_xmit_retransmit_queue(sk
);
2593 EXPORT_SYMBOL(tcp_simple_retransmit
);
2595 void tcp_enter_recovery(struct sock
*sk
, bool ece_ack
)
2597 struct tcp_sock
*tp
= tcp_sk(sk
);
2600 if (tcp_is_reno(tp
))
2601 mib_idx
= LINUX_MIB_TCPRENORECOVERY
;
2603 mib_idx
= LINUX_MIB_TCPSACKRECOVERY
;
2605 NET_INC_STATS(sock_net(sk
), mib_idx
);
2607 tp
->prior_ssthresh
= 0;
2610 if (!tcp_in_cwnd_reduction(sk
)) {
2612 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2613 tcp_init_cwnd_reduction(sk
);
2615 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
2618 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2619 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2621 static void tcp_process_loss(struct sock
*sk
, int flag
, bool is_dupack
,
2624 struct tcp_sock
*tp
= tcp_sk(sk
);
2625 bool recovered
= !before(tp
->snd_una
, tp
->high_seq
);
2627 if ((flag
& FLAG_SND_UNA_ADVANCED
) &&
2628 tcp_try_undo_loss(sk
, false))
2631 /* The ACK (s)acks some never-retransmitted data meaning not all
2632 * the data packets before the timeout were lost. Therefore we
2633 * undo the congestion window and state. This is essentially
2634 * the operation in F-RTO (RFC5682 section 3.1 step 3.b). Since
2635 * a retransmitted skb is permantly marked, we can apply such an
2636 * operation even if F-RTO was not used.
2638 if ((flag
& FLAG_ORIG_SACK_ACKED
) &&
2639 tcp_try_undo_loss(sk
, tp
->undo_marker
))
2642 if (tp
->frto
) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2643 if (after(tp
->snd_nxt
, tp
->high_seq
)) {
2644 if (flag
& FLAG_DATA_SACKED
|| is_dupack
)
2645 tp
->frto
= 0; /* Step 3.a. loss was real */
2646 } else if (flag
& FLAG_SND_UNA_ADVANCED
&& !recovered
) {
2647 tp
->high_seq
= tp
->snd_nxt
;
2648 /* Step 2.b. Try send new data (but deferred until cwnd
2649 * is updated in tcp_ack()). Otherwise fall back to
2650 * the conventional recovery.
2652 if (!tcp_write_queue_empty(sk
) &&
2653 after(tcp_wnd_end(tp
), tp
->snd_nxt
)) {
2654 *rexmit
= REXMIT_NEW
;
2662 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2663 tcp_try_undo_recovery(sk
);
2666 if (tcp_is_reno(tp
)) {
2667 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2668 * delivered. Lower inflight to clock out (re)tranmissions.
2670 if (after(tp
->snd_nxt
, tp
->high_seq
) && is_dupack
)
2671 tcp_add_reno_sack(sk
);
2672 else if (flag
& FLAG_SND_UNA_ADVANCED
)
2673 tcp_reset_reno_sack(tp
);
2675 *rexmit
= REXMIT_LOST
;
2678 /* Undo during fast recovery after partial ACK. */
2679 static bool tcp_try_undo_partial(struct sock
*sk
, u32 prior_snd_una
)
2681 struct tcp_sock
*tp
= tcp_sk(sk
);
2683 if (tp
->undo_marker
&& tcp_packet_delayed(tp
)) {
2684 /* Plain luck! Hole if filled with delayed
2685 * packet, rather than with a retransmit. Check reordering.
2687 tcp_check_sack_reordering(sk
, prior_snd_una
, 1);
2689 /* We are getting evidence that the reordering degree is higher
2690 * than we realized. If there are no retransmits out then we
2691 * can undo. Otherwise we clock out new packets but do not
2692 * mark more packets lost or retransmit more.
2694 if (tp
->retrans_out
)
2697 if (!tcp_any_retrans_done(sk
))
2698 tp
->retrans_stamp
= 0;
2700 DBGUNDO(sk
, "partial recovery");
2701 tcp_undo_cwnd_reduction(sk
, true);
2702 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPPARTIALUNDO
);
2703 tcp_try_keep_open(sk
);
2709 static void tcp_rack_identify_loss(struct sock
*sk
, int *ack_flag
)
2711 struct tcp_sock
*tp
= tcp_sk(sk
);
2713 /* Use RACK to detect loss */
2714 if (sock_net(sk
)->ipv4
.sysctl_tcp_recovery
& TCP_RACK_LOSS_DETECTION
) {
2715 u32 prior_retrans
= tp
->retrans_out
;
2717 tcp_rack_mark_lost(sk
);
2718 if (prior_retrans
> tp
->retrans_out
)
2719 *ack_flag
|= FLAG_LOST_RETRANS
;
2723 static bool tcp_force_fast_retransmit(struct sock
*sk
)
2725 struct tcp_sock
*tp
= tcp_sk(sk
);
2727 return after(tcp_highest_sack_seq(tp
),
2728 tp
->snd_una
+ tp
->reordering
* tp
->mss_cache
);
2731 /* Process an event, which can update packets-in-flight not trivially.
2732 * Main goal of this function is to calculate new estimate for left_out,
2733 * taking into account both packets sitting in receiver's buffer and
2734 * packets lost by network.
2736 * Besides that it updates the congestion state when packet loss or ECN
2737 * is detected. But it does not reduce the cwnd, it is done by the
2738 * congestion control later.
2740 * It does _not_ decide what to send, it is made in function
2741 * tcp_xmit_retransmit_queue().
2743 static void tcp_fastretrans_alert(struct sock
*sk
, const u32 prior_snd_una
,
2744 bool is_dupack
, int *ack_flag
, int *rexmit
)
2746 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2747 struct tcp_sock
*tp
= tcp_sk(sk
);
2748 int fast_rexmit
= 0, flag
= *ack_flag
;
2749 bool do_lost
= is_dupack
|| ((flag
& FLAG_DATA_SACKED
) &&
2750 tcp_force_fast_retransmit(sk
));
2752 if (!tp
->packets_out
&& tp
->sacked_out
)
2755 /* Now state machine starts.
2756 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2757 if (flag
& FLAG_ECE
)
2758 tp
->prior_ssthresh
= 0;
2760 /* B. In all the states check for reneging SACKs. */
2761 if (tcp_check_sack_reneging(sk
, flag
))
2764 /* C. Check consistency of the current state. */
2765 tcp_verify_left_out(tp
);
2767 /* D. Check state exit conditions. State can be terminated
2768 * when high_seq is ACKed. */
2769 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
2770 WARN_ON(tp
->retrans_out
!= 0);
2771 tp
->retrans_stamp
= 0;
2772 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
2773 switch (icsk
->icsk_ca_state
) {
2775 /* CWR is to be held something *above* high_seq
2776 * is ACKed for CWR bit to reach receiver. */
2777 if (tp
->snd_una
!= tp
->high_seq
) {
2778 tcp_end_cwnd_reduction(sk
);
2779 tcp_set_ca_state(sk
, TCP_CA_Open
);
2783 case TCP_CA_Recovery
:
2784 if (tcp_is_reno(tp
))
2785 tcp_reset_reno_sack(tp
);
2786 if (tcp_try_undo_recovery(sk
))
2788 tcp_end_cwnd_reduction(sk
);
2793 /* E. Process state. */
2794 switch (icsk
->icsk_ca_state
) {
2795 case TCP_CA_Recovery
:
2796 if (!(flag
& FLAG_SND_UNA_ADVANCED
)) {
2797 if (tcp_is_reno(tp
) && is_dupack
)
2798 tcp_add_reno_sack(sk
);
2800 if (tcp_try_undo_partial(sk
, prior_snd_una
))
2802 /* Partial ACK arrived. Force fast retransmit. */
2803 do_lost
= tcp_is_reno(tp
) ||
2804 tcp_force_fast_retransmit(sk
);
2806 if (tcp_try_undo_dsack(sk
)) {
2807 tcp_try_keep_open(sk
);
2810 tcp_rack_identify_loss(sk
, ack_flag
);
2813 tcp_process_loss(sk
, flag
, is_dupack
, rexmit
);
2814 tcp_rack_identify_loss(sk
, ack_flag
);
2815 if (!(icsk
->icsk_ca_state
== TCP_CA_Open
||
2816 (*ack_flag
& FLAG_LOST_RETRANS
)))
2818 /* Change state if cwnd is undone or retransmits are lost */
2821 if (tcp_is_reno(tp
)) {
2822 if (flag
& FLAG_SND_UNA_ADVANCED
)
2823 tcp_reset_reno_sack(tp
);
2825 tcp_add_reno_sack(sk
);
2828 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
)
2829 tcp_try_undo_dsack(sk
);
2831 tcp_rack_identify_loss(sk
, ack_flag
);
2832 if (!tcp_time_to_recover(sk
, flag
)) {
2833 tcp_try_to_open(sk
, flag
);
2837 /* MTU probe failure: don't reduce cwnd */
2838 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
2839 icsk
->icsk_mtup
.probe_size
&&
2840 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
2841 tcp_mtup_probe_failed(sk
);
2842 /* Restores the reduction we did in tcp_mtup_probe() */
2844 tcp_simple_retransmit(sk
);
2848 /* Otherwise enter Recovery state */
2849 tcp_enter_recovery(sk
, (flag
& FLAG_ECE
));
2854 tcp_update_scoreboard(sk
, fast_rexmit
);
2855 *rexmit
= REXMIT_LOST
;
2858 static void tcp_update_rtt_min(struct sock
*sk
, u32 rtt_us
)
2860 u32 wlen
= sock_net(sk
)->ipv4
.sysctl_tcp_min_rtt_wlen
* HZ
;
2861 struct tcp_sock
*tp
= tcp_sk(sk
);
2863 minmax_running_min(&tp
->rtt_min
, wlen
, tcp_jiffies32
,
2864 rtt_us
? : jiffies_to_usecs(1));
2867 static bool tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
2868 long seq_rtt_us
, long sack_rtt_us
,
2869 long ca_rtt_us
, struct rate_sample
*rs
)
2871 const struct tcp_sock
*tp
= tcp_sk(sk
);
2873 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2874 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2875 * Karn's algorithm forbids taking RTT if some retransmitted data
2876 * is acked (RFC6298).
2879 seq_rtt_us
= sack_rtt_us
;
2881 /* RTTM Rule: A TSecr value received in a segment is used to
2882 * update the averaged RTT measurement only if the segment
2883 * acknowledges some new data, i.e., only if it advances the
2884 * left edge of the send window.
2885 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2887 if (seq_rtt_us
< 0 && tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2888 flag
& FLAG_ACKED
) {
2889 u32 delta
= tcp_time_stamp(tp
) - tp
->rx_opt
.rcv_tsecr
;
2890 u32 delta_us
= delta
* (USEC_PER_SEC
/ TCP_TS_HZ
);
2892 seq_rtt_us
= ca_rtt_us
= delta_us
;
2894 rs
->rtt_us
= ca_rtt_us
; /* RTT of last (S)ACKed packet (or -1) */
2898 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
2899 * always taken together with ACK, SACK, or TS-opts. Any negative
2900 * values will be skipped with the seq_rtt_us < 0 check above.
2902 tcp_update_rtt_min(sk
, ca_rtt_us
);
2903 tcp_rtt_estimator(sk
, seq_rtt_us
);
2906 /* RFC6298: only reset backoff on valid RTT measurement. */
2907 inet_csk(sk
)->icsk_backoff
= 0;
2911 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2912 void tcp_synack_rtt_meas(struct sock
*sk
, struct request_sock
*req
)
2914 struct rate_sample rs
;
2917 if (req
&& !req
->num_retrans
&& tcp_rsk(req
)->snt_synack
)
2918 rtt_us
= tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req
)->snt_synack
);
2920 tcp_ack_update_rtt(sk
, FLAG_SYN_ACKED
, rtt_us
, -1L, rtt_us
, &rs
);
2924 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
, u32 acked
)
2926 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2928 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, acked
);
2929 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_jiffies32
;
2932 /* Restart timer after forward progress on connection.
2933 * RFC2988 recommends to restart timer to now+rto.
2935 void tcp_rearm_rto(struct sock
*sk
)
2937 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2938 struct tcp_sock
*tp
= tcp_sk(sk
);
2940 /* If the retrans timer is currently being used by Fast Open
2941 * for SYN-ACK retrans purpose, stay put.
2943 if (tp
->fastopen_rsk
)
2946 if (!tp
->packets_out
) {
2947 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
2949 u32 rto
= inet_csk(sk
)->icsk_rto
;
2950 /* Offset the time elapsed after installing regular RTO */
2951 if (icsk
->icsk_pending
== ICSK_TIME_REO_TIMEOUT
||
2952 icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
) {
2953 s64 delta_us
= tcp_rto_delta_us(sk
);
2954 /* delta_us may not be positive if the socket is locked
2955 * when the retrans timer fires and is rescheduled.
2957 rto
= usecs_to_jiffies(max_t(int, delta_us
, 1));
2959 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
, rto
,
2964 /* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */
2965 static void tcp_set_xmit_timer(struct sock
*sk
)
2967 if (!tcp_schedule_loss_probe(sk
, true))
2971 /* If we get here, the whole TSO packet has not been acked. */
2972 static u32
tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
)
2974 struct tcp_sock
*tp
= tcp_sk(sk
);
2977 BUG_ON(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
));
2979 packets_acked
= tcp_skb_pcount(skb
);
2980 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
2982 packets_acked
-= tcp_skb_pcount(skb
);
2984 if (packets_acked
) {
2985 BUG_ON(tcp_skb_pcount(skb
) == 0);
2986 BUG_ON(!before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
));
2989 return packets_acked
;
2992 static void tcp_ack_tstamp(struct sock
*sk
, struct sk_buff
*skb
,
2995 const struct skb_shared_info
*shinfo
;
2997 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
2998 if (likely(!TCP_SKB_CB(skb
)->txstamp_ack
))
3001 shinfo
= skb_shinfo(skb
);
3002 if (!before(shinfo
->tskey
, prior_snd_una
) &&
3003 before(shinfo
->tskey
, tcp_sk(sk
)->snd_una
)) {
3004 tcp_skb_tsorted_save(skb
) {
3005 __skb_tstamp_tx(skb
, NULL
, sk
, SCM_TSTAMP_ACK
);
3006 } tcp_skb_tsorted_restore(skb
);
3010 /* Remove acknowledged frames from the retransmission queue. If our packet
3011 * is before the ack sequence we can discard it as it's confirmed to have
3012 * arrived at the other end.
3014 static int tcp_clean_rtx_queue(struct sock
*sk
, u32 prior_fack
,
3016 struct tcp_sacktag_state
*sack
)
3018 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3019 u64 first_ackt
, last_ackt
;
3020 struct tcp_sock
*tp
= tcp_sk(sk
);
3021 u32 prior_sacked
= tp
->sacked_out
;
3022 u32 reord
= tp
->snd_nxt
; /* lowest acked un-retx un-sacked seq */
3023 struct sk_buff
*skb
, *next
;
3024 bool fully_acked
= true;
3025 long sack_rtt_us
= -1L;
3026 long seq_rtt_us
= -1L;
3027 long ca_rtt_us
= -1L;
3029 u32 last_in_flight
= 0;
3035 for (skb
= skb_rb_first(&sk
->tcp_rtx_queue
); skb
; skb
= next
) {
3036 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
3037 const u32 start_seq
= scb
->seq
;
3038 u8 sacked
= scb
->sacked
;
3041 tcp_ack_tstamp(sk
, skb
, prior_snd_una
);
3043 /* Determine how many packets and what bytes were acked, tso and else */
3044 if (after(scb
->end_seq
, tp
->snd_una
)) {
3045 if (tcp_skb_pcount(skb
) == 1 ||
3046 !after(tp
->snd_una
, scb
->seq
))
3049 acked_pcount
= tcp_tso_acked(sk
, skb
);
3052 fully_acked
= false;
3054 acked_pcount
= tcp_skb_pcount(skb
);
3057 if (unlikely(sacked
& TCPCB_RETRANS
)) {
3058 if (sacked
& TCPCB_SACKED_RETRANS
)
3059 tp
->retrans_out
-= acked_pcount
;
3060 flag
|= FLAG_RETRANS_DATA_ACKED
;
3061 } else if (!(sacked
& TCPCB_SACKED_ACKED
)) {
3062 last_ackt
= skb
->skb_mstamp
;
3063 WARN_ON_ONCE(last_ackt
== 0);
3065 first_ackt
= last_ackt
;
3067 last_in_flight
= TCP_SKB_CB(skb
)->tx
.in_flight
;
3068 if (before(start_seq
, reord
))
3070 if (!after(scb
->end_seq
, tp
->high_seq
))
3071 flag
|= FLAG_ORIG_SACK_ACKED
;
3074 if (sacked
& TCPCB_SACKED_ACKED
) {
3075 tp
->sacked_out
-= acked_pcount
;
3076 } else if (tcp_is_sack(tp
)) {
3077 tp
->delivered
+= acked_pcount
;
3078 if (!tcp_skb_spurious_retrans(tp
, skb
))
3079 tcp_rack_advance(tp
, sacked
, scb
->end_seq
,
3082 if (sacked
& TCPCB_LOST
)
3083 tp
->lost_out
-= acked_pcount
;
3085 tp
->packets_out
-= acked_pcount
;
3086 pkts_acked
+= acked_pcount
;
3087 tcp_rate_skb_delivered(sk
, skb
, sack
->rate
);
3089 /* Initial outgoing SYN's get put onto the write_queue
3090 * just like anything else we transmit. It is not
3091 * true data, and if we misinform our callers that
3092 * this ACK acks real data, we will erroneously exit
3093 * connection startup slow start one packet too
3094 * quickly. This is severely frowned upon behavior.
3096 if (likely(!(scb
->tcp_flags
& TCPHDR_SYN
))) {
3097 flag
|= FLAG_DATA_ACKED
;
3099 flag
|= FLAG_SYN_ACKED
;
3100 tp
->retrans_stamp
= 0;
3106 next
= skb_rb_next(skb
);
3107 if (unlikely(skb
== tp
->retransmit_skb_hint
))
3108 tp
->retransmit_skb_hint
= NULL
;
3109 if (unlikely(skb
== tp
->lost_skb_hint
))
3110 tp
->lost_skb_hint
= NULL
;
3111 tcp_rtx_queue_unlink_and_free(skb
, sk
);
3115 tcp_chrono_stop(sk
, TCP_CHRONO_BUSY
);
3117 if (likely(between(tp
->snd_up
, prior_snd_una
, tp
->snd_una
)))
3118 tp
->snd_up
= tp
->snd_una
;
3120 if (skb
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
3121 flag
|= FLAG_SACK_RENEGING
;
3123 if (likely(first_ackt
) && !(flag
& FLAG_RETRANS_DATA_ACKED
)) {
3124 seq_rtt_us
= tcp_stamp_us_delta(tp
->tcp_mstamp
, first_ackt
);
3125 ca_rtt_us
= tcp_stamp_us_delta(tp
->tcp_mstamp
, last_ackt
);
3127 if (sack
->first_sackt
) {
3128 sack_rtt_us
= tcp_stamp_us_delta(tp
->tcp_mstamp
, sack
->first_sackt
);
3129 ca_rtt_us
= tcp_stamp_us_delta(tp
->tcp_mstamp
, sack
->last_sackt
);
3131 rtt_update
= tcp_ack_update_rtt(sk
, flag
, seq_rtt_us
, sack_rtt_us
,
3132 ca_rtt_us
, sack
->rate
);
3134 if (flag
& FLAG_ACKED
) {
3135 flag
|= FLAG_SET_XMIT_TIMER
; /* set TLP or RTO timer */
3136 if (unlikely(icsk
->icsk_mtup
.probe_size
&&
3137 !after(tp
->mtu_probe
.probe_seq_end
, tp
->snd_una
))) {
3138 tcp_mtup_probe_success(sk
);
3141 if (tcp_is_reno(tp
)) {
3142 tcp_remove_reno_sacks(sk
, pkts_acked
);
3146 /* Non-retransmitted hole got filled? That's reordering */
3147 if (before(reord
, prior_fack
))
3148 tcp_check_sack_reordering(sk
, reord
, 0);
3150 delta
= prior_sacked
- tp
->sacked_out
;
3151 tp
->lost_cnt_hint
-= min(tp
->lost_cnt_hint
, delta
);
3153 } else if (skb
&& rtt_update
&& sack_rtt_us
>= 0 &&
3154 sack_rtt_us
> tcp_stamp_us_delta(tp
->tcp_mstamp
, skb
->skb_mstamp
)) {
3155 /* Do not re-arm RTO if the sack RTT is measured from data sent
3156 * after when the head was last (re)transmitted. Otherwise the
3157 * timeout may continue to extend in loss recovery.
3159 flag
|= FLAG_SET_XMIT_TIMER
; /* set TLP or RTO timer */
3162 if (icsk
->icsk_ca_ops
->pkts_acked
) {
3163 struct ack_sample sample
= { .pkts_acked
= pkts_acked
,
3164 .rtt_us
= sack
->rate
->rtt_us
,
3165 .in_flight
= last_in_flight
};
3167 icsk
->icsk_ca_ops
->pkts_acked(sk
, &sample
);
3170 #if FASTRETRANS_DEBUG > 0
3171 WARN_ON((int)tp
->sacked_out
< 0);
3172 WARN_ON((int)tp
->lost_out
< 0);
3173 WARN_ON((int)tp
->retrans_out
< 0);
3174 if (!tp
->packets_out
&& tcp_is_sack(tp
)) {
3175 icsk
= inet_csk(sk
);
3177 pr_debug("Leak l=%u %d\n",
3178 tp
->lost_out
, icsk
->icsk_ca_state
);
3181 if (tp
->sacked_out
) {
3182 pr_debug("Leak s=%u %d\n",
3183 tp
->sacked_out
, icsk
->icsk_ca_state
);
3186 if (tp
->retrans_out
) {
3187 pr_debug("Leak r=%u %d\n",
3188 tp
->retrans_out
, icsk
->icsk_ca_state
);
3189 tp
->retrans_out
= 0;
3196 static void tcp_ack_probe(struct sock
*sk
)
3198 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3199 struct sk_buff
*head
= tcp_send_head(sk
);
3200 const struct tcp_sock
*tp
= tcp_sk(sk
);
3202 /* Was it a usable window open? */
3205 if (!after(TCP_SKB_CB(head
)->end_seq
, tcp_wnd_end(tp
))) {
3206 icsk
->icsk_backoff
= 0;
3207 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
3208 /* Socket must be waked up by subsequent tcp_data_snd_check().
3209 * This function is not for random using!
3212 unsigned long when
= tcp_probe0_when(sk
, TCP_RTO_MAX
);
3214 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3219 static inline bool tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
3221 return !(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
3222 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
;
3225 /* Decide wheather to run the increase function of congestion control. */
3226 static inline bool tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
3228 /* If reordering is high then always grow cwnd whenever data is
3229 * delivered regardless of its ordering. Otherwise stay conservative
3230 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3231 * new SACK or ECE mark may first advance cwnd here and later reduce
3232 * cwnd in tcp_fastretrans_alert() based on more states.
3234 if (tcp_sk(sk
)->reordering
> sock_net(sk
)->ipv4
.sysctl_tcp_reordering
)
3235 return flag
& FLAG_FORWARD_PROGRESS
;
3237 return flag
& FLAG_DATA_ACKED
;
3240 /* The "ultimate" congestion control function that aims to replace the rigid
3241 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3242 * It's called toward the end of processing an ACK with precise rate
3243 * information. All transmission or retransmission are delayed afterwards.
3245 static void tcp_cong_control(struct sock
*sk
, u32 ack
, u32 acked_sacked
,
3246 int flag
, const struct rate_sample
*rs
)
3248 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3250 if (icsk
->icsk_ca_ops
->cong_control
) {
3251 icsk
->icsk_ca_ops
->cong_control(sk
, rs
);
3255 if (tcp_in_cwnd_reduction(sk
)) {
3256 /* Reduce cwnd if state mandates */
3257 tcp_cwnd_reduction(sk
, acked_sacked
, flag
);
3258 } else if (tcp_may_raise_cwnd(sk
, flag
)) {
3259 /* Advance cwnd if state allows */
3260 tcp_cong_avoid(sk
, ack
, acked_sacked
);
3262 tcp_update_pacing_rate(sk
);
3265 /* Check that window update is acceptable.
3266 * The function assumes that snd_una<=ack<=snd_next.
3268 static inline bool tcp_may_update_window(const struct tcp_sock
*tp
,
3269 const u32 ack
, const u32 ack_seq
,
3272 return after(ack
, tp
->snd_una
) ||
3273 after(ack_seq
, tp
->snd_wl1
) ||
3274 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
);
3277 /* If we update tp->snd_una, also update tp->bytes_acked */
3278 static void tcp_snd_una_update(struct tcp_sock
*tp
, u32 ack
)
3280 u32 delta
= ack
- tp
->snd_una
;
3282 sock_owned_by_me((struct sock
*)tp
);
3283 tp
->bytes_acked
+= delta
;
3287 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3288 static void tcp_rcv_nxt_update(struct tcp_sock
*tp
, u32 seq
)
3290 u32 delta
= seq
- tp
->rcv_nxt
;
3292 sock_owned_by_me((struct sock
*)tp
);
3293 tp
->bytes_received
+= delta
;
3297 /* Update our send window.
3299 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3300 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3302 static int tcp_ack_update_window(struct sock
*sk
, const struct sk_buff
*skb
, u32 ack
,
3305 struct tcp_sock
*tp
= tcp_sk(sk
);
3307 u32 nwin
= ntohs(tcp_hdr(skb
)->window
);
3309 if (likely(!tcp_hdr(skb
)->syn
))
3310 nwin
<<= tp
->rx_opt
.snd_wscale
;
3312 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
3313 flag
|= FLAG_WIN_UPDATE
;
3314 tcp_update_wl(tp
, ack_seq
);
3316 if (tp
->snd_wnd
!= nwin
) {
3319 /* Note, it is the only place, where
3320 * fast path is recovered for sending TCP.
3323 tcp_fast_path_check(sk
);
3325 if (!tcp_write_queue_empty(sk
))
3326 tcp_slow_start_after_idle_check(sk
);
3328 if (nwin
> tp
->max_window
) {
3329 tp
->max_window
= nwin
;
3330 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
3335 tcp_snd_una_update(tp
, ack
);
3340 static bool __tcp_oow_rate_limited(struct net
*net
, int mib_idx
,
3341 u32
*last_oow_ack_time
)
3343 if (*last_oow_ack_time
) {
3344 s32 elapsed
= (s32
)(tcp_jiffies32
- *last_oow_ack_time
);
3346 if (0 <= elapsed
&& elapsed
< net
->ipv4
.sysctl_tcp_invalid_ratelimit
) {
3347 NET_INC_STATS(net
, mib_idx
);
3348 return true; /* rate-limited: don't send yet! */
3352 *last_oow_ack_time
= tcp_jiffies32
;
3354 return false; /* not rate-limited: go ahead, send dupack now! */
3357 /* Return true if we're currently rate-limiting out-of-window ACKs and
3358 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3359 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3360 * attacks that send repeated SYNs or ACKs for the same connection. To
3361 * do this, we do not send a duplicate SYNACK or ACK if the remote
3362 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3364 bool tcp_oow_rate_limited(struct net
*net
, const struct sk_buff
*skb
,
3365 int mib_idx
, u32
*last_oow_ack_time
)
3367 /* Data packets without SYNs are not likely part of an ACK loop. */
3368 if ((TCP_SKB_CB(skb
)->seq
!= TCP_SKB_CB(skb
)->end_seq
) &&
3372 return __tcp_oow_rate_limited(net
, mib_idx
, last_oow_ack_time
);
3375 /* RFC 5961 7 [ACK Throttling] */
3376 static void tcp_send_challenge_ack(struct sock
*sk
, const struct sk_buff
*skb
)
3378 /* unprotected vars, we dont care of overwrites */
3379 static u32 challenge_timestamp
;
3380 static unsigned int challenge_count
;
3381 struct tcp_sock
*tp
= tcp_sk(sk
);
3382 struct net
*net
= sock_net(sk
);
3385 /* First check our per-socket dupack rate limit. */
3386 if (__tcp_oow_rate_limited(net
,
3387 LINUX_MIB_TCPACKSKIPPEDCHALLENGE
,
3388 &tp
->last_oow_ack_time
))
3391 /* Then check host-wide RFC 5961 rate limit. */
3393 if (now
!= challenge_timestamp
) {
3394 u32 ack_limit
= net
->ipv4
.sysctl_tcp_challenge_ack_limit
;
3395 u32 half
= (ack_limit
+ 1) >> 1;
3397 challenge_timestamp
= now
;
3398 WRITE_ONCE(challenge_count
, half
+ prandom_u32_max(ack_limit
));
3400 count
= READ_ONCE(challenge_count
);
3402 WRITE_ONCE(challenge_count
, count
- 1);
3403 NET_INC_STATS(net
, LINUX_MIB_TCPCHALLENGEACK
);
3408 static void tcp_store_ts_recent(struct tcp_sock
*tp
)
3410 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
3411 tp
->rx_opt
.ts_recent_stamp
= get_seconds();
3414 static void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
3416 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
3417 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3418 * extra check below makes sure this can only happen
3419 * for pure ACK frames. -DaveM
3421 * Not only, also it occurs for expired timestamps.
3424 if (tcp_paws_check(&tp
->rx_opt
, 0))
3425 tcp_store_ts_recent(tp
);
3429 /* This routine deals with acks during a TLP episode.
3430 * We mark the end of a TLP episode on receiving TLP dupack or when
3431 * ack is after tlp_high_seq.
3432 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3434 static void tcp_process_tlp_ack(struct sock
*sk
, u32 ack
, int flag
)
3436 struct tcp_sock
*tp
= tcp_sk(sk
);
3438 if (before(ack
, tp
->tlp_high_seq
))
3441 if (flag
& FLAG_DSACKING_ACK
) {
3442 /* This DSACK means original and TLP probe arrived; no loss */
3443 tp
->tlp_high_seq
= 0;
3444 } else if (after(ack
, tp
->tlp_high_seq
)) {
3445 /* ACK advances: there was a loss, so reduce cwnd. Reset
3446 * tlp_high_seq in tcp_init_cwnd_reduction()
3448 tcp_init_cwnd_reduction(sk
);
3449 tcp_set_ca_state(sk
, TCP_CA_CWR
);
3450 tcp_end_cwnd_reduction(sk
);
3451 tcp_try_keep_open(sk
);
3452 NET_INC_STATS(sock_net(sk
),
3453 LINUX_MIB_TCPLOSSPROBERECOVERY
);
3454 } else if (!(flag
& (FLAG_SND_UNA_ADVANCED
|
3455 FLAG_NOT_DUP
| FLAG_DATA_SACKED
))) {
3456 /* Pure dupack: original and TLP probe arrived; no loss */
3457 tp
->tlp_high_seq
= 0;
3461 static inline void tcp_in_ack_event(struct sock
*sk
, u32 flags
)
3463 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3465 if (icsk
->icsk_ca_ops
->in_ack_event
)
3466 icsk
->icsk_ca_ops
->in_ack_event(sk
, flags
);
3469 /* Congestion control has updated the cwnd already. So if we're in
3470 * loss recovery then now we do any new sends (for FRTO) or
3471 * retransmits (for CA_Loss or CA_recovery) that make sense.
3473 static void tcp_xmit_recovery(struct sock
*sk
, int rexmit
)
3475 struct tcp_sock
*tp
= tcp_sk(sk
);
3477 if (rexmit
== REXMIT_NONE
)
3480 if (unlikely(rexmit
== 2)) {
3481 __tcp_push_pending_frames(sk
, tcp_current_mss(sk
),
3483 if (after(tp
->snd_nxt
, tp
->high_seq
))
3487 tcp_xmit_retransmit_queue(sk
);
3490 /* This routine deals with incoming acks, but not outgoing ones. */
3491 static int tcp_ack(struct sock
*sk
, const struct sk_buff
*skb
, int flag
)
3493 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3494 struct tcp_sock
*tp
= tcp_sk(sk
);
3495 struct tcp_sacktag_state sack_state
;
3496 struct rate_sample rs
= { .prior_delivered
= 0 };
3497 u32 prior_snd_una
= tp
->snd_una
;
3498 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
3499 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3500 bool is_dupack
= false;
3501 int prior_packets
= tp
->packets_out
;
3502 u32 delivered
= tp
->delivered
;
3503 u32 lost
= tp
->lost
;
3504 int rexmit
= REXMIT_NONE
; /* Flag to (re)transmit to recover losses */
3507 sack_state
.first_sackt
= 0;
3508 sack_state
.rate
= &rs
;
3510 /* We very likely will need to access rtx queue. */
3511 prefetch(sk
->tcp_rtx_queue
.rb_node
);
3513 /* If the ack is older than previous acks
3514 * then we can probably ignore it.
3516 if (before(ack
, prior_snd_una
)) {
3517 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3518 if (before(ack
, prior_snd_una
- tp
->max_window
)) {
3519 if (!(flag
& FLAG_NO_CHALLENGE_ACK
))
3520 tcp_send_challenge_ack(sk
, skb
);
3526 /* If the ack includes data we haven't sent yet, discard
3527 * this segment (RFC793 Section 3.9).
3529 if (after(ack
, tp
->snd_nxt
))
3532 if (after(ack
, prior_snd_una
)) {
3533 flag
|= FLAG_SND_UNA_ADVANCED
;
3534 icsk
->icsk_retransmits
= 0;
3537 prior_fack
= tcp_is_sack(tp
) ? tcp_highest_sack_seq(tp
) : tp
->snd_una
;
3538 rs
.prior_in_flight
= tcp_packets_in_flight(tp
);
3540 /* ts_recent update must be made after we are sure that the packet
3543 if (flag
& FLAG_UPDATE_TS_RECENT
)
3544 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
3546 if (!(flag
& FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
3547 /* Window is constant, pure forward advance.
3548 * No more checks are required.
3549 * Note, we use the fact that SND.UNA>=SND.WL2.
3551 tcp_update_wl(tp
, ack_seq
);
3552 tcp_snd_una_update(tp
, ack
);
3553 flag
|= FLAG_WIN_UPDATE
;
3555 tcp_in_ack_event(sk
, CA_ACK_WIN_UPDATE
);
3557 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPHPACKS
);
3559 u32 ack_ev_flags
= CA_ACK_SLOWPATH
;
3561 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
3564 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPPUREACKS
);
3566 flag
|= tcp_ack_update_window(sk
, skb
, ack
, ack_seq
);
3568 if (TCP_SKB_CB(skb
)->sacked
)
3569 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
,
3572 if (tcp_ecn_rcv_ecn_echo(tp
, tcp_hdr(skb
))) {
3574 ack_ev_flags
|= CA_ACK_ECE
;
3577 if (flag
& FLAG_WIN_UPDATE
)
3578 ack_ev_flags
|= CA_ACK_WIN_UPDATE
;
3580 tcp_in_ack_event(sk
, ack_ev_flags
);
3583 /* We passed data and got it acked, remove any soft error
3584 * log. Something worked...
3586 sk
->sk_err_soft
= 0;
3587 icsk
->icsk_probes_out
= 0;
3588 tp
->rcv_tstamp
= tcp_jiffies32
;
3592 /* See if we can take anything off of the retransmit queue. */
3593 flag
|= tcp_clean_rtx_queue(sk
, prior_fack
, prior_snd_una
, &sack_state
);
3595 tcp_rack_update_reo_wnd(sk
, &rs
);
3597 if (tp
->tlp_high_seq
)
3598 tcp_process_tlp_ack(sk
, ack
, flag
);
3599 /* If needed, reset TLP/RTO timer; RACK may later override this. */
3600 if (flag
& FLAG_SET_XMIT_TIMER
)
3601 tcp_set_xmit_timer(sk
);
3603 if (tcp_ack_is_dubious(sk
, flag
)) {
3604 is_dupack
= !(flag
& (FLAG_SND_UNA_ADVANCED
| FLAG_NOT_DUP
));
3605 tcp_fastretrans_alert(sk
, prior_snd_una
, is_dupack
, &flag
,
3609 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
& FLAG_NOT_DUP
))
3612 delivered
= tp
->delivered
- delivered
; /* freshly ACKed or SACKed */
3613 lost
= tp
->lost
- lost
; /* freshly marked lost */
3614 tcp_rate_gen(sk
, delivered
, lost
, sack_state
.rate
);
3615 tcp_cong_control(sk
, ack
, delivered
, flag
, sack_state
.rate
);
3616 tcp_xmit_recovery(sk
, rexmit
);
3620 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3621 if (flag
& FLAG_DSACKING_ACK
)
3622 tcp_fastretrans_alert(sk
, prior_snd_una
, is_dupack
, &flag
,
3624 /* If this ack opens up a zero window, clear backoff. It was
3625 * being used to time the probes, and is probably far higher than
3626 * it needs to be for normal retransmission.
3630 if (tp
->tlp_high_seq
)
3631 tcp_process_tlp_ack(sk
, ack
, flag
);
3635 SOCK_DEBUG(sk
, "Ack %u after %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3639 /* If data was SACKed, tag it and see if we should send more data.
3640 * If data was DSACKed, see if we can undo a cwnd reduction.
3642 if (TCP_SKB_CB(skb
)->sacked
) {
3643 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
,
3645 tcp_fastretrans_alert(sk
, prior_snd_una
, is_dupack
, &flag
,
3647 tcp_xmit_recovery(sk
, rexmit
);
3650 SOCK_DEBUG(sk
, "Ack %u before %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3654 static void tcp_parse_fastopen_option(int len
, const unsigned char *cookie
,
3655 bool syn
, struct tcp_fastopen_cookie
*foc
,
3658 /* Valid only in SYN or SYN-ACK with an even length. */
3659 if (!foc
|| !syn
|| len
< 0 || (len
& 1))
3662 if (len
>= TCP_FASTOPEN_COOKIE_MIN
&&
3663 len
<= TCP_FASTOPEN_COOKIE_MAX
)
3664 memcpy(foc
->val
, cookie
, len
);
3671 static void smc_parse_options(const struct tcphdr
*th
,
3672 struct tcp_options_received
*opt_rx
,
3673 const unsigned char *ptr
,
3676 #if IS_ENABLED(CONFIG_SMC)
3677 if (static_branch_unlikely(&tcp_have_smc
)) {
3678 if (th
->syn
&& !(opsize
& 1) &&
3679 opsize
>= TCPOLEN_EXP_SMC_BASE
&&
3680 get_unaligned_be32(ptr
) == TCPOPT_SMC_MAGIC
)
3686 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3687 * But, this can also be called on packets in the established flow when
3688 * the fast version below fails.
3690 void tcp_parse_options(const struct net
*net
,
3691 const struct sk_buff
*skb
,
3692 struct tcp_options_received
*opt_rx
, int estab
,
3693 struct tcp_fastopen_cookie
*foc
)
3695 const unsigned char *ptr
;
3696 const struct tcphdr
*th
= tcp_hdr(skb
);
3697 int length
= (th
->doff
* 4) - sizeof(struct tcphdr
);
3699 ptr
= (const unsigned char *)(th
+ 1);
3700 opt_rx
->saw_tstamp
= 0;
3702 while (length
> 0) {
3703 int opcode
= *ptr
++;
3709 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
3714 if (opsize
< 2) /* "silly options" */
3716 if (opsize
> length
)
3717 return; /* don't parse partial options */
3720 if (opsize
== TCPOLEN_MSS
&& th
->syn
&& !estab
) {
3721 u16 in_mss
= get_unaligned_be16(ptr
);
3723 if (opt_rx
->user_mss
&&
3724 opt_rx
->user_mss
< in_mss
)
3725 in_mss
= opt_rx
->user_mss
;
3726 opt_rx
->mss_clamp
= in_mss
;
3731 if (opsize
== TCPOLEN_WINDOW
&& th
->syn
&&
3732 !estab
&& net
->ipv4
.sysctl_tcp_window_scaling
) {
3733 __u8 snd_wscale
= *(__u8
*)ptr
;
3734 opt_rx
->wscale_ok
= 1;
3735 if (snd_wscale
> TCP_MAX_WSCALE
) {
3736 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n",
3740 snd_wscale
= TCP_MAX_WSCALE
;
3742 opt_rx
->snd_wscale
= snd_wscale
;
3745 case TCPOPT_TIMESTAMP
:
3746 if ((opsize
== TCPOLEN_TIMESTAMP
) &&
3747 ((estab
&& opt_rx
->tstamp_ok
) ||
3748 (!estab
&& net
->ipv4
.sysctl_tcp_timestamps
))) {
3749 opt_rx
->saw_tstamp
= 1;
3750 opt_rx
->rcv_tsval
= get_unaligned_be32(ptr
);
3751 opt_rx
->rcv_tsecr
= get_unaligned_be32(ptr
+ 4);
3754 case TCPOPT_SACK_PERM
:
3755 if (opsize
== TCPOLEN_SACK_PERM
&& th
->syn
&&
3756 !estab
&& net
->ipv4
.sysctl_tcp_sack
) {
3757 opt_rx
->sack_ok
= TCP_SACK_SEEN
;
3758 tcp_sack_reset(opt_rx
);
3763 if ((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
3764 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
3766 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
3769 #ifdef CONFIG_TCP_MD5SIG
3772 * The MD5 Hash has already been
3773 * checked (see tcp_v{4,6}_do_rcv()).
3777 case TCPOPT_FASTOPEN
:
3778 tcp_parse_fastopen_option(
3779 opsize
- TCPOLEN_FASTOPEN_BASE
,
3780 ptr
, th
->syn
, foc
, false);
3784 /* Fast Open option shares code 254 using a
3785 * 16 bits magic number.
3787 if (opsize
>= TCPOLEN_EXP_FASTOPEN_BASE
&&
3788 get_unaligned_be16(ptr
) ==
3789 TCPOPT_FASTOPEN_MAGIC
)
3790 tcp_parse_fastopen_option(opsize
-
3791 TCPOLEN_EXP_FASTOPEN_BASE
,
3792 ptr
+ 2, th
->syn
, foc
, true);
3794 smc_parse_options(th
, opt_rx
, ptr
,
3804 EXPORT_SYMBOL(tcp_parse_options
);
3806 static bool tcp_parse_aligned_timestamp(struct tcp_sock
*tp
, const struct tcphdr
*th
)
3808 const __be32
*ptr
= (const __be32
*)(th
+ 1);
3810 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3811 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
3812 tp
->rx_opt
.saw_tstamp
= 1;
3814 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3817 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
) - tp
->tsoffset
;
3819 tp
->rx_opt
.rcv_tsecr
= 0;
3825 /* Fast parse options. This hopes to only see timestamps.
3826 * If it is wrong it falls back on tcp_parse_options().
3828 static bool tcp_fast_parse_options(const struct net
*net
,
3829 const struct sk_buff
*skb
,
3830 const struct tcphdr
*th
, struct tcp_sock
*tp
)
3832 /* In the spirit of fast parsing, compare doff directly to constant
3833 * values. Because equality is used, short doff can be ignored here.
3835 if (th
->doff
== (sizeof(*th
) / 4)) {
3836 tp
->rx_opt
.saw_tstamp
= 0;
3838 } else if (tp
->rx_opt
.tstamp_ok
&&
3839 th
->doff
== ((sizeof(*th
) + TCPOLEN_TSTAMP_ALIGNED
) / 4)) {
3840 if (tcp_parse_aligned_timestamp(tp
, th
))
3844 tcp_parse_options(net
, skb
, &tp
->rx_opt
, 1, NULL
);
3845 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
3846 tp
->rx_opt
.rcv_tsecr
-= tp
->tsoffset
;
3851 #ifdef CONFIG_TCP_MD5SIG
3853 * Parse MD5 Signature option
3855 const u8
*tcp_parse_md5sig_option(const struct tcphdr
*th
)
3857 int length
= (th
->doff
<< 2) - sizeof(*th
);
3858 const u8
*ptr
= (const u8
*)(th
+ 1);
3860 /* If the TCP option is too short, we can short cut */
3861 if (length
< TCPOLEN_MD5SIG
)
3864 while (length
> 0) {
3865 int opcode
= *ptr
++;
3876 if (opsize
< 2 || opsize
> length
)
3878 if (opcode
== TCPOPT_MD5SIG
)
3879 return opsize
== TCPOLEN_MD5SIG
? ptr
: NULL
;
3886 EXPORT_SYMBOL(tcp_parse_md5sig_option
);
3889 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3891 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3892 * it can pass through stack. So, the following predicate verifies that
3893 * this segment is not used for anything but congestion avoidance or
3894 * fast retransmit. Moreover, we even are able to eliminate most of such
3895 * second order effects, if we apply some small "replay" window (~RTO)
3896 * to timestamp space.
3898 * All these measures still do not guarantee that we reject wrapped ACKs
3899 * on networks with high bandwidth, when sequence space is recycled fastly,
3900 * but it guarantees that such events will be very rare and do not affect
3901 * connection seriously. This doesn't look nice, but alas, PAWS is really
3904 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3905 * states that events when retransmit arrives after original data are rare.
3906 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3907 * the biggest problem on large power networks even with minor reordering.
3908 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3909 * up to bandwidth of 18Gigabit/sec. 8) ]
3912 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
3914 const struct tcp_sock
*tp
= tcp_sk(sk
);
3915 const struct tcphdr
*th
= tcp_hdr(skb
);
3916 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3917 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3919 return (/* 1. Pure ACK with correct sequence number. */
3920 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
3922 /* 2. ... and duplicate ACK. */
3923 ack
== tp
->snd_una
&&
3925 /* 3. ... and does not update window. */
3926 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
3928 /* 4. ... and sits in replay window. */
3929 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
3932 static inline bool tcp_paws_discard(const struct sock
*sk
,
3933 const struct sk_buff
*skb
)
3935 const struct tcp_sock
*tp
= tcp_sk(sk
);
3937 return !tcp_paws_check(&tp
->rx_opt
, TCP_PAWS_WINDOW
) &&
3938 !tcp_disordered_ack(sk
, skb
);
3941 /* Check segment sequence number for validity.
3943 * Segment controls are considered valid, if the segment
3944 * fits to the window after truncation to the window. Acceptability
3945 * of data (and SYN, FIN, of course) is checked separately.
3946 * See tcp_data_queue(), for example.
3948 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3949 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3950 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3951 * (borrowed from freebsd)
3954 static inline bool tcp_sequence(const struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3956 return !before(end_seq
, tp
->rcv_wup
) &&
3957 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
3960 /* When we get a reset we do this. */
3961 void tcp_reset(struct sock
*sk
)
3963 trace_tcp_receive_reset(sk
);
3965 /* We want the right error as BSD sees it (and indeed as we do). */
3966 switch (sk
->sk_state
) {
3968 sk
->sk_err
= ECONNREFUSED
;
3970 case TCP_CLOSE_WAIT
:
3976 sk
->sk_err
= ECONNRESET
;
3978 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3983 if (!sock_flag(sk
, SOCK_DEAD
))
3984 sk
->sk_error_report(sk
);
3988 * Process the FIN bit. This now behaves as it is supposed to work
3989 * and the FIN takes effect when it is validly part of sequence
3990 * space. Not before when we get holes.
3992 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3993 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3996 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3997 * close and we go into CLOSING (and later onto TIME-WAIT)
3999 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4001 void tcp_fin(struct sock
*sk
)
4003 struct tcp_sock
*tp
= tcp_sk(sk
);
4005 inet_csk_schedule_ack(sk
);
4007 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
4008 sock_set_flag(sk
, SOCK_DONE
);
4010 switch (sk
->sk_state
) {
4012 case TCP_ESTABLISHED
:
4013 /* Move to CLOSE_WAIT */
4014 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
4015 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
4018 case TCP_CLOSE_WAIT
:
4020 /* Received a retransmission of the FIN, do
4025 /* RFC793: Remain in the LAST-ACK state. */
4029 /* This case occurs when a simultaneous close
4030 * happens, we must ack the received FIN and
4031 * enter the CLOSING state.
4034 tcp_set_state(sk
, TCP_CLOSING
);
4037 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4039 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
4042 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4043 * cases we should never reach this piece of code.
4045 pr_err("%s: Impossible, sk->sk_state=%d\n",
4046 __func__
, sk
->sk_state
);
4050 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4051 * Probably, we should reset in this case. For now drop them.
4053 skb_rbtree_purge(&tp
->out_of_order_queue
);
4054 if (tcp_is_sack(tp
))
4055 tcp_sack_reset(&tp
->rx_opt
);
4058 if (!sock_flag(sk
, SOCK_DEAD
)) {
4059 sk
->sk_state_change(sk
);
4061 /* Do not send POLL_HUP for half duplex close. */
4062 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
4063 sk
->sk_state
== TCP_CLOSE
)
4064 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_HUP
);
4066 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_IN
);
4070 static inline bool tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
,
4073 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
4074 if (before(seq
, sp
->start_seq
))
4075 sp
->start_seq
= seq
;
4076 if (after(end_seq
, sp
->end_seq
))
4077 sp
->end_seq
= end_seq
;
4083 static void tcp_dsack_set(struct sock
*sk
, u32 seq
, u32 end_seq
)
4085 struct tcp_sock
*tp
= tcp_sk(sk
);
4087 if (tcp_is_sack(tp
) && sock_net(sk
)->ipv4
.sysctl_tcp_dsack
) {
4090 if (before(seq
, tp
->rcv_nxt
))
4091 mib_idx
= LINUX_MIB_TCPDSACKOLDSENT
;
4093 mib_idx
= LINUX_MIB_TCPDSACKOFOSENT
;
4095 NET_INC_STATS(sock_net(sk
), mib_idx
);
4097 tp
->rx_opt
.dsack
= 1;
4098 tp
->duplicate_sack
[0].start_seq
= seq
;
4099 tp
->duplicate_sack
[0].end_seq
= end_seq
;
4103 static void tcp_dsack_extend(struct sock
*sk
, u32 seq
, u32 end_seq
)
4105 struct tcp_sock
*tp
= tcp_sk(sk
);
4107 if (!tp
->rx_opt
.dsack
)
4108 tcp_dsack_set(sk
, seq
, end_seq
);
4110 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
4113 static void tcp_send_dupack(struct sock
*sk
, const struct sk_buff
*skb
)
4115 struct tcp_sock
*tp
= tcp_sk(sk
);
4117 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
4118 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4119 NET_INC_STATS(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4120 tcp_enter_quickack_mode(sk
);
4122 if (tcp_is_sack(tp
) && sock_net(sk
)->ipv4
.sysctl_tcp_dsack
) {
4123 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4125 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
4126 end_seq
= tp
->rcv_nxt
;
4127 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, end_seq
);
4134 /* These routines update the SACK block as out-of-order packets arrive or
4135 * in-order packets close up the sequence space.
4137 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
4140 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4141 struct tcp_sack_block
*swalk
= sp
+ 1;
4143 /* See if the recent change to the first SACK eats into
4144 * or hits the sequence space of other SACK blocks, if so coalesce.
4146 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
;) {
4147 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
4150 /* Zap SWALK, by moving every further SACK up by one slot.
4151 * Decrease num_sacks.
4153 tp
->rx_opt
.num_sacks
--;
4154 for (i
= this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
4158 this_sack
++, swalk
++;
4162 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
4164 struct tcp_sock
*tp
= tcp_sk(sk
);
4165 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4166 int cur_sacks
= tp
->rx_opt
.num_sacks
;
4172 for (this_sack
= 0; this_sack
< cur_sacks
; this_sack
++, sp
++) {
4173 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
4174 /* Rotate this_sack to the first one. */
4175 for (; this_sack
> 0; this_sack
--, sp
--)
4176 swap(*sp
, *(sp
- 1));
4178 tcp_sack_maybe_coalesce(tp
);
4183 /* Could not find an adjacent existing SACK, build a new one,
4184 * put it at the front, and shift everyone else down. We
4185 * always know there is at least one SACK present already here.
4187 * If the sack array is full, forget about the last one.
4189 if (this_sack
>= TCP_NUM_SACKS
) {
4191 tp
->rx_opt
.num_sacks
--;
4194 for (; this_sack
> 0; this_sack
--, sp
--)
4198 /* Build the new head SACK, and we're done. */
4199 sp
->start_seq
= seq
;
4200 sp
->end_seq
= end_seq
;
4201 tp
->rx_opt
.num_sacks
++;
4204 /* RCV.NXT advances, some SACKs should be eaten. */
4206 static void tcp_sack_remove(struct tcp_sock
*tp
)
4208 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4209 int num_sacks
= tp
->rx_opt
.num_sacks
;
4212 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4213 if (RB_EMPTY_ROOT(&tp
->out_of_order_queue
)) {
4214 tp
->rx_opt
.num_sacks
= 0;
4218 for (this_sack
= 0; this_sack
< num_sacks
;) {
4219 /* Check if the start of the sack is covered by RCV.NXT. */
4220 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
4223 /* RCV.NXT must cover all the block! */
4224 WARN_ON(before(tp
->rcv_nxt
, sp
->end_seq
));
4226 /* Zap this SACK, by moving forward any other SACKS. */
4227 for (i
= this_sack
+1; i
< num_sacks
; i
++)
4228 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
4235 tp
->rx_opt
.num_sacks
= num_sacks
;
4239 * tcp_try_coalesce - try to merge skb to prior one
4241 * @dest: destination queue
4243 * @from: buffer to add in queue
4244 * @fragstolen: pointer to boolean
4246 * Before queueing skb @from after @to, try to merge them
4247 * to reduce overall memory use and queue lengths, if cost is small.
4248 * Packets in ofo or receive queues can stay a long time.
4249 * Better try to coalesce them right now to avoid future collapses.
4250 * Returns true if caller should free @from instead of queueing it
4252 static bool tcp_try_coalesce(struct sock
*sk
,
4254 struct sk_buff
*from
,
4259 *fragstolen
= false;
4261 /* Its possible this segment overlaps with prior segment in queue */
4262 if (TCP_SKB_CB(from
)->seq
!= TCP_SKB_CB(to
)->end_seq
)
4265 if (!skb_try_coalesce(to
, from
, fragstolen
, &delta
))
4268 atomic_add(delta
, &sk
->sk_rmem_alloc
);
4269 sk_mem_charge(sk
, delta
);
4270 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPRCVCOALESCE
);
4271 TCP_SKB_CB(to
)->end_seq
= TCP_SKB_CB(from
)->end_seq
;
4272 TCP_SKB_CB(to
)->ack_seq
= TCP_SKB_CB(from
)->ack_seq
;
4273 TCP_SKB_CB(to
)->tcp_flags
|= TCP_SKB_CB(from
)->tcp_flags
;
4275 if (TCP_SKB_CB(from
)->has_rxtstamp
) {
4276 TCP_SKB_CB(to
)->has_rxtstamp
= true;
4277 to
->tstamp
= from
->tstamp
;
4283 static void tcp_drop(struct sock
*sk
, struct sk_buff
*skb
)
4285 sk_drops_add(sk
, skb
);
4289 /* This one checks to see if we can put data from the
4290 * out_of_order queue into the receive_queue.
4292 static void tcp_ofo_queue(struct sock
*sk
)
4294 struct tcp_sock
*tp
= tcp_sk(sk
);
4295 __u32 dsack_high
= tp
->rcv_nxt
;
4296 bool fin
, fragstolen
, eaten
;
4297 struct sk_buff
*skb
, *tail
;
4300 p
= rb_first(&tp
->out_of_order_queue
);
4303 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
4306 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
4307 __u32 dsack
= dsack_high
;
4308 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
4309 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
4310 tcp_dsack_extend(sk
, TCP_SKB_CB(skb
)->seq
, dsack
);
4313 rb_erase(&skb
->rbnode
, &tp
->out_of_order_queue
);
4315 if (unlikely(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))) {
4316 SOCK_DEBUG(sk
, "ofo packet was already received\n");
4320 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
4321 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4322 TCP_SKB_CB(skb
)->end_seq
);
4324 tail
= skb_peek_tail(&sk
->sk_receive_queue
);
4325 eaten
= tail
&& tcp_try_coalesce(sk
, tail
, skb
, &fragstolen
);
4326 tcp_rcv_nxt_update(tp
, TCP_SKB_CB(skb
)->end_seq
);
4327 fin
= TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
;
4329 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4331 kfree_skb_partial(skb
, fragstolen
);
4333 if (unlikely(fin
)) {
4335 /* tcp_fin() purges tp->out_of_order_queue,
4336 * so we must end this loop right now.
4343 static bool tcp_prune_ofo_queue(struct sock
*sk
);
4344 static int tcp_prune_queue(struct sock
*sk
);
4346 static int tcp_try_rmem_schedule(struct sock
*sk
, struct sk_buff
*skb
,
4349 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
4350 !sk_rmem_schedule(sk
, skb
, size
)) {
4352 if (tcp_prune_queue(sk
) < 0)
4355 while (!sk_rmem_schedule(sk
, skb
, size
)) {
4356 if (!tcp_prune_ofo_queue(sk
))
4363 static void tcp_data_queue_ofo(struct sock
*sk
, struct sk_buff
*skb
)
4365 struct tcp_sock
*tp
= tcp_sk(sk
);
4366 struct rb_node
**p
, *parent
;
4367 struct sk_buff
*skb1
;
4371 tcp_ecn_check_ce(tp
, skb
);
4373 if (unlikely(tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))) {
4374 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPOFODROP
);
4379 /* Disable header prediction. */
4381 inet_csk_schedule_ack(sk
);
4383 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPOFOQUEUE
);
4384 seq
= TCP_SKB_CB(skb
)->seq
;
4385 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4386 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
4387 tp
->rcv_nxt
, seq
, end_seq
);
4389 p
= &tp
->out_of_order_queue
.rb_node
;
4390 if (RB_EMPTY_ROOT(&tp
->out_of_order_queue
)) {
4391 /* Initial out of order segment, build 1 SACK. */
4392 if (tcp_is_sack(tp
)) {
4393 tp
->rx_opt
.num_sacks
= 1;
4394 tp
->selective_acks
[0].start_seq
= seq
;
4395 tp
->selective_acks
[0].end_seq
= end_seq
;
4397 rb_link_node(&skb
->rbnode
, NULL
, p
);
4398 rb_insert_color(&skb
->rbnode
, &tp
->out_of_order_queue
);
4399 tp
->ooo_last_skb
= skb
;
4403 /* In the typical case, we are adding an skb to the end of the list.
4404 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4406 if (tcp_try_coalesce(sk
, tp
->ooo_last_skb
,
4407 skb
, &fragstolen
)) {
4409 tcp_grow_window(sk
, skb
);
4410 kfree_skb_partial(skb
, fragstolen
);
4414 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4415 if (!before(seq
, TCP_SKB_CB(tp
->ooo_last_skb
)->end_seq
)) {
4416 parent
= &tp
->ooo_last_skb
->rbnode
;
4417 p
= &parent
->rb_right
;
4421 /* Find place to insert this segment. Handle overlaps on the way. */
4425 skb1
= rb_to_skb(parent
);
4426 if (before(seq
, TCP_SKB_CB(skb1
)->seq
)) {
4427 p
= &parent
->rb_left
;
4430 if (before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4431 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4432 /* All the bits are present. Drop. */
4433 NET_INC_STATS(sock_net(sk
),
4434 LINUX_MIB_TCPOFOMERGE
);
4437 tcp_dsack_set(sk
, seq
, end_seq
);
4440 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
4441 /* Partial overlap. */
4442 tcp_dsack_set(sk
, seq
, TCP_SKB_CB(skb1
)->end_seq
);
4444 /* skb's seq == skb1's seq and skb covers skb1.
4445 * Replace skb1 with skb.
4447 rb_replace_node(&skb1
->rbnode
, &skb
->rbnode
,
4448 &tp
->out_of_order_queue
);
4449 tcp_dsack_extend(sk
,
4450 TCP_SKB_CB(skb1
)->seq
,
4451 TCP_SKB_CB(skb1
)->end_seq
);
4452 NET_INC_STATS(sock_net(sk
),
4453 LINUX_MIB_TCPOFOMERGE
);
4457 } else if (tcp_try_coalesce(sk
, skb1
,
4458 skb
, &fragstolen
)) {
4461 p
= &parent
->rb_right
;
4464 /* Insert segment into RB tree. */
4465 rb_link_node(&skb
->rbnode
, parent
, p
);
4466 rb_insert_color(&skb
->rbnode
, &tp
->out_of_order_queue
);
4469 /* Remove other segments covered by skb. */
4470 while ((skb1
= skb_rb_next(skb
)) != NULL
) {
4471 if (!after(end_seq
, TCP_SKB_CB(skb1
)->seq
))
4473 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4474 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4478 rb_erase(&skb1
->rbnode
, &tp
->out_of_order_queue
);
4479 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4480 TCP_SKB_CB(skb1
)->end_seq
);
4481 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPOFOMERGE
);
4484 /* If there is no skb after us, we are the last_skb ! */
4486 tp
->ooo_last_skb
= skb
;
4489 if (tcp_is_sack(tp
))
4490 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
4493 tcp_grow_window(sk
, skb
);
4495 skb_set_owner_r(skb
, sk
);
4499 static int __must_check
tcp_queue_rcv(struct sock
*sk
, struct sk_buff
*skb
, int hdrlen
,
4503 struct sk_buff
*tail
= skb_peek_tail(&sk
->sk_receive_queue
);
4505 __skb_pull(skb
, hdrlen
);
4507 tcp_try_coalesce(sk
, tail
,
4508 skb
, fragstolen
)) ? 1 : 0;
4509 tcp_rcv_nxt_update(tcp_sk(sk
), TCP_SKB_CB(skb
)->end_seq
);
4511 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4512 skb_set_owner_r(skb
, sk
);
4517 int tcp_send_rcvq(struct sock
*sk
, struct msghdr
*msg
, size_t size
)
4519 struct sk_buff
*skb
;
4527 if (size
> PAGE_SIZE
) {
4528 int npages
= min_t(size_t, size
>> PAGE_SHIFT
, MAX_SKB_FRAGS
);
4530 data_len
= npages
<< PAGE_SHIFT
;
4531 size
= data_len
+ (size
& ~PAGE_MASK
);
4533 skb
= alloc_skb_with_frags(size
- data_len
, data_len
,
4534 PAGE_ALLOC_COSTLY_ORDER
,
4535 &err
, sk
->sk_allocation
);
4539 skb_put(skb
, size
- data_len
);
4540 skb
->data_len
= data_len
;
4543 if (tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))
4546 err
= skb_copy_datagram_from_iter(skb
, 0, &msg
->msg_iter
, size
);
4550 TCP_SKB_CB(skb
)->seq
= tcp_sk(sk
)->rcv_nxt
;
4551 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ size
;
4552 TCP_SKB_CB(skb
)->ack_seq
= tcp_sk(sk
)->snd_una
- 1;
4554 if (tcp_queue_rcv(sk
, skb
, 0, &fragstolen
)) {
4555 WARN_ON_ONCE(fragstolen
); /* should not happen */
4567 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
4569 struct tcp_sock
*tp
= tcp_sk(sk
);
4573 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
) {
4578 __skb_pull(skb
, tcp_hdr(skb
)->doff
* 4);
4580 tcp_ecn_accept_cwr(tp
, skb
);
4582 tp
->rx_opt
.dsack
= 0;
4584 /* Queue data for delivery to the user.
4585 * Packets in sequence go to the receive queue.
4586 * Out of sequence packets to the out_of_order_queue.
4588 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
4589 if (tcp_receive_window(tp
) == 0)
4592 /* Ok. In sequence. In window. */
4594 if (skb_queue_len(&sk
->sk_receive_queue
) == 0)
4595 sk_forced_mem_schedule(sk
, skb
->truesize
);
4596 else if (tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))
4599 eaten
= tcp_queue_rcv(sk
, skb
, 0, &fragstolen
);
4600 tcp_rcv_nxt_update(tp
, TCP_SKB_CB(skb
)->end_seq
);
4602 tcp_event_data_recv(sk
, skb
);
4603 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
4606 if (!RB_EMPTY_ROOT(&tp
->out_of_order_queue
)) {
4609 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4610 * gap in queue is filled.
4612 if (RB_EMPTY_ROOT(&tp
->out_of_order_queue
))
4613 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
4616 if (tp
->rx_opt
.num_sacks
)
4617 tcp_sack_remove(tp
);
4619 tcp_fast_path_check(sk
);
4622 kfree_skb_partial(skb
, fragstolen
);
4623 if (!sock_flag(sk
, SOCK_DEAD
))
4624 sk
->sk_data_ready(sk
);
4628 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4629 /* A retransmit, 2nd most common case. Force an immediate ack. */
4630 NET_INC_STATS(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4631 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4634 tcp_enter_quickack_mode(sk
);
4635 inet_csk_schedule_ack(sk
);
4641 /* Out of window. F.e. zero window probe. */
4642 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
4645 tcp_enter_quickack_mode(sk
);
4647 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4648 /* Partial packet, seq < rcv_next < end_seq */
4649 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
4650 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4651 TCP_SKB_CB(skb
)->end_seq
);
4653 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
4655 /* If window is closed, drop tail of packet. But after
4656 * remembering D-SACK for its head made in previous line.
4658 if (!tcp_receive_window(tp
))
4663 tcp_data_queue_ofo(sk
, skb
);
4666 static struct sk_buff
*tcp_skb_next(struct sk_buff
*skb
, struct sk_buff_head
*list
)
4669 return !skb_queue_is_last(list
, skb
) ? skb
->next
: NULL
;
4671 return skb_rb_next(skb
);
4674 static struct sk_buff
*tcp_collapse_one(struct sock
*sk
, struct sk_buff
*skb
,
4675 struct sk_buff_head
*list
,
4676 struct rb_root
*root
)
4678 struct sk_buff
*next
= tcp_skb_next(skb
, list
);
4681 __skb_unlink(skb
, list
);
4683 rb_erase(&skb
->rbnode
, root
);
4686 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPRCVCOLLAPSED
);
4691 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
4692 void tcp_rbtree_insert(struct rb_root
*root
, struct sk_buff
*skb
)
4694 struct rb_node
**p
= &root
->rb_node
;
4695 struct rb_node
*parent
= NULL
;
4696 struct sk_buff
*skb1
;
4700 skb1
= rb_to_skb(parent
);
4701 if (before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb1
)->seq
))
4702 p
= &parent
->rb_left
;
4704 p
= &parent
->rb_right
;
4706 rb_link_node(&skb
->rbnode
, parent
, p
);
4707 rb_insert_color(&skb
->rbnode
, root
);
4710 /* Collapse contiguous sequence of skbs head..tail with
4711 * sequence numbers start..end.
4713 * If tail is NULL, this means until the end of the queue.
4715 * Segments with FIN/SYN are not collapsed (only because this
4719 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
, struct rb_root
*root
,
4720 struct sk_buff
*head
, struct sk_buff
*tail
, u32 start
, u32 end
)
4722 struct sk_buff
*skb
= head
, *n
;
4723 struct sk_buff_head tmp
;
4726 /* First, check that queue is collapsible and find
4727 * the point where collapsing can be useful.
4730 for (end_of_skbs
= true; skb
!= NULL
&& skb
!= tail
; skb
= n
) {
4731 n
= tcp_skb_next(skb
, list
);
4733 /* No new bits? It is possible on ofo queue. */
4734 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4735 skb
= tcp_collapse_one(sk
, skb
, list
, root
);
4741 /* The first skb to collapse is:
4743 * - bloated or contains data before "start" or
4744 * overlaps to the next one.
4746 if (!(TCP_SKB_CB(skb
)->tcp_flags
& (TCPHDR_SYN
| TCPHDR_FIN
)) &&
4747 (tcp_win_from_space(sk
, skb
->truesize
) > skb
->len
||
4748 before(TCP_SKB_CB(skb
)->seq
, start
))) {
4749 end_of_skbs
= false;
4753 if (n
&& n
!= tail
&&
4754 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(n
)->seq
) {
4755 end_of_skbs
= false;
4759 /* Decided to skip this, advance start seq. */
4760 start
= TCP_SKB_CB(skb
)->end_seq
;
4763 (TCP_SKB_CB(skb
)->tcp_flags
& (TCPHDR_SYN
| TCPHDR_FIN
)))
4766 __skb_queue_head_init(&tmp
);
4768 while (before(start
, end
)) {
4769 int copy
= min_t(int, SKB_MAX_ORDER(0, 0), end
- start
);
4770 struct sk_buff
*nskb
;
4772 nskb
= alloc_skb(copy
, GFP_ATOMIC
);
4776 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
4777 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
4779 __skb_queue_before(list
, skb
, nskb
);
4781 __skb_queue_tail(&tmp
, nskb
); /* defer rbtree insertion */
4782 skb_set_owner_r(nskb
, sk
);
4784 /* Copy data, releasing collapsed skbs. */
4786 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
4787 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
4791 size
= min(copy
, size
);
4792 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
4794 TCP_SKB_CB(nskb
)->end_seq
+= size
;
4798 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4799 skb
= tcp_collapse_one(sk
, skb
, list
, root
);
4802 (TCP_SKB_CB(skb
)->tcp_flags
& (TCPHDR_SYN
| TCPHDR_FIN
)))
4808 skb_queue_walk_safe(&tmp
, skb
, n
)
4809 tcp_rbtree_insert(root
, skb
);
4812 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4813 * and tcp_collapse() them until all the queue is collapsed.
4815 static void tcp_collapse_ofo_queue(struct sock
*sk
)
4817 struct tcp_sock
*tp
= tcp_sk(sk
);
4818 struct sk_buff
*skb
, *head
;
4821 skb
= skb_rb_first(&tp
->out_of_order_queue
);
4824 tp
->ooo_last_skb
= skb_rb_last(&tp
->out_of_order_queue
);
4827 start
= TCP_SKB_CB(skb
)->seq
;
4828 end
= TCP_SKB_CB(skb
)->end_seq
;
4830 for (head
= skb
;;) {
4831 skb
= skb_rb_next(skb
);
4833 /* Range is terminated when we see a gap or when
4834 * we are at the queue end.
4837 after(TCP_SKB_CB(skb
)->seq
, end
) ||
4838 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
4839 tcp_collapse(sk
, NULL
, &tp
->out_of_order_queue
,
4840 head
, skb
, start
, end
);
4844 if (unlikely(before(TCP_SKB_CB(skb
)->seq
, start
)))
4845 start
= TCP_SKB_CB(skb
)->seq
;
4846 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
4847 end
= TCP_SKB_CB(skb
)->end_seq
;
4852 * Clean the out-of-order queue to make room.
4853 * We drop high sequences packets to :
4854 * 1) Let a chance for holes to be filled.
4855 * 2) not add too big latencies if thousands of packets sit there.
4856 * (But if application shrinks SO_RCVBUF, we could still end up
4857 * freeing whole queue here)
4859 * Return true if queue has shrunk.
4861 static bool tcp_prune_ofo_queue(struct sock
*sk
)
4863 struct tcp_sock
*tp
= tcp_sk(sk
);
4864 struct rb_node
*node
, *prev
;
4866 if (RB_EMPTY_ROOT(&tp
->out_of_order_queue
))
4869 NET_INC_STATS(sock_net(sk
), LINUX_MIB_OFOPRUNED
);
4870 node
= &tp
->ooo_last_skb
->rbnode
;
4872 prev
= rb_prev(node
);
4873 rb_erase(node
, &tp
->out_of_order_queue
);
4874 tcp_drop(sk
, rb_to_skb(node
));
4876 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
&&
4877 !tcp_under_memory_pressure(sk
))
4881 tp
->ooo_last_skb
= rb_to_skb(prev
);
4883 /* Reset SACK state. A conforming SACK implementation will
4884 * do the same at a timeout based retransmit. When a connection
4885 * is in a sad state like this, we care only about integrity
4886 * of the connection not performance.
4888 if (tp
->rx_opt
.sack_ok
)
4889 tcp_sack_reset(&tp
->rx_opt
);
4893 /* Reduce allocated memory if we can, trying to get
4894 * the socket within its memory limits again.
4896 * Return less than zero if we should start dropping frames
4897 * until the socket owning process reads some of the data
4898 * to stabilize the situation.
4900 static int tcp_prune_queue(struct sock
*sk
)
4902 struct tcp_sock
*tp
= tcp_sk(sk
);
4904 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
4906 NET_INC_STATS(sock_net(sk
), LINUX_MIB_PRUNECALLED
);
4908 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
4909 tcp_clamp_window(sk
);
4910 else if (tcp_under_memory_pressure(sk
))
4911 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
4913 tcp_collapse_ofo_queue(sk
);
4914 if (!skb_queue_empty(&sk
->sk_receive_queue
))
4915 tcp_collapse(sk
, &sk
->sk_receive_queue
, NULL
,
4916 skb_peek(&sk
->sk_receive_queue
),
4918 tp
->copied_seq
, tp
->rcv_nxt
);
4921 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4924 /* Collapsing did not help, destructive actions follow.
4925 * This must not ever occur. */
4927 tcp_prune_ofo_queue(sk
);
4929 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4932 /* If we are really being abused, tell the caller to silently
4933 * drop receive data on the floor. It will get retransmitted
4934 * and hopefully then we'll have sufficient space.
4936 NET_INC_STATS(sock_net(sk
), LINUX_MIB_RCVPRUNED
);
4938 /* Massive buffer overcommit. */
4943 static bool tcp_should_expand_sndbuf(const struct sock
*sk
)
4945 const struct tcp_sock
*tp
= tcp_sk(sk
);
4947 /* If the user specified a specific send buffer setting, do
4950 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
4953 /* If we are under global TCP memory pressure, do not expand. */
4954 if (tcp_under_memory_pressure(sk
))
4957 /* If we are under soft global TCP memory pressure, do not expand. */
4958 if (sk_memory_allocated(sk
) >= sk_prot_mem_limits(sk
, 0))
4961 /* If we filled the congestion window, do not expand. */
4962 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
4968 /* When incoming ACK allowed to free some skb from write_queue,
4969 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4970 * on the exit from tcp input handler.
4972 * PROBLEM: sndbuf expansion does not work well with largesend.
4974 static void tcp_new_space(struct sock
*sk
)
4976 struct tcp_sock
*tp
= tcp_sk(sk
);
4978 if (tcp_should_expand_sndbuf(sk
)) {
4979 tcp_sndbuf_expand(sk
);
4980 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
4983 sk
->sk_write_space(sk
);
4986 static void tcp_check_space(struct sock
*sk
)
4988 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
4989 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
4990 /* pairs with tcp_poll() */
4992 if (sk
->sk_socket
&&
4993 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
4995 if (!test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
4996 tcp_chrono_stop(sk
, TCP_CHRONO_SNDBUF_LIMITED
);
5001 static inline void tcp_data_snd_check(struct sock
*sk
)
5003 tcp_push_pending_frames(sk
);
5004 tcp_check_space(sk
);
5008 * Check if sending an ack is needed.
5010 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
5012 struct tcp_sock
*tp
= tcp_sk(sk
);
5014 /* More than one full frame received... */
5015 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
&&
5016 /* ... and right edge of window advances far enough.
5017 * (tcp_recvmsg() will send ACK otherwise). Or...
5019 __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
5020 /* We ACK each frame or... */
5021 tcp_in_quickack_mode(sk
) ||
5022 /* We have out of order data. */
5023 (ofo_possible
&& !RB_EMPTY_ROOT(&tp
->out_of_order_queue
))) {
5024 /* Then ack it now */
5027 /* Else, send delayed ack. */
5028 tcp_send_delayed_ack(sk
);
5032 static inline void tcp_ack_snd_check(struct sock
*sk
)
5034 if (!inet_csk_ack_scheduled(sk
)) {
5035 /* We sent a data segment already. */
5038 __tcp_ack_snd_check(sk
, 1);
5042 * This routine is only called when we have urgent data
5043 * signaled. Its the 'slow' part of tcp_urg. It could be
5044 * moved inline now as tcp_urg is only called from one
5045 * place. We handle URGent data wrong. We have to - as
5046 * BSD still doesn't use the correction from RFC961.
5047 * For 1003.1g we should support a new option TCP_STDURG to permit
5048 * either form (or just set the sysctl tcp_stdurg).
5051 static void tcp_check_urg(struct sock
*sk
, const struct tcphdr
*th
)
5053 struct tcp_sock
*tp
= tcp_sk(sk
);
5054 u32 ptr
= ntohs(th
->urg_ptr
);
5056 if (ptr
&& !sock_net(sk
)->ipv4
.sysctl_tcp_stdurg
)
5058 ptr
+= ntohl(th
->seq
);
5060 /* Ignore urgent data that we've already seen and read. */
5061 if (after(tp
->copied_seq
, ptr
))
5064 /* Do not replay urg ptr.
5066 * NOTE: interesting situation not covered by specs.
5067 * Misbehaving sender may send urg ptr, pointing to segment,
5068 * which we already have in ofo queue. We are not able to fetch
5069 * such data and will stay in TCP_URG_NOTYET until will be eaten
5070 * by recvmsg(). Seems, we are not obliged to handle such wicked
5071 * situations. But it is worth to think about possibility of some
5072 * DoSes using some hypothetical application level deadlock.
5074 if (before(ptr
, tp
->rcv_nxt
))
5077 /* Do we already have a newer (or duplicate) urgent pointer? */
5078 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
5081 /* Tell the world about our new urgent pointer. */
5084 /* We may be adding urgent data when the last byte read was
5085 * urgent. To do this requires some care. We cannot just ignore
5086 * tp->copied_seq since we would read the last urgent byte again
5087 * as data, nor can we alter copied_seq until this data arrives
5088 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5090 * NOTE. Double Dutch. Rendering to plain English: author of comment
5091 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5092 * and expect that both A and B disappear from stream. This is _wrong_.
5093 * Though this happens in BSD with high probability, this is occasional.
5094 * Any application relying on this is buggy. Note also, that fix "works"
5095 * only in this artificial test. Insert some normal data between A and B and we will
5096 * decline of BSD again. Verdict: it is better to remove to trap
5099 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
5100 !sock_flag(sk
, SOCK_URGINLINE
) && tp
->copied_seq
!= tp
->rcv_nxt
) {
5101 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
5103 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5104 __skb_unlink(skb
, &sk
->sk_receive_queue
);
5109 tp
->urg_data
= TCP_URG_NOTYET
;
5112 /* Disable header prediction. */
5116 /* This is the 'fast' part of urgent handling. */
5117 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, const struct tcphdr
*th
)
5119 struct tcp_sock
*tp
= tcp_sk(sk
);
5121 /* Check if we get a new urgent pointer - normally not. */
5123 tcp_check_urg(sk
, th
);
5125 /* Do we wait for any urgent data? - normally not... */
5126 if (tp
->urg_data
== TCP_URG_NOTYET
) {
5127 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
5130 /* Is the urgent pointer pointing into this packet? */
5131 if (ptr
< skb
->len
) {
5133 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
5135 tp
->urg_data
= TCP_URG_VALID
| tmp
;
5136 if (!sock_flag(sk
, SOCK_DEAD
))
5137 sk
->sk_data_ready(sk
);
5142 /* Accept RST for rcv_nxt - 1 after a FIN.
5143 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a
5144 * FIN is sent followed by a RST packet. The RST is sent with the same
5145 * sequence number as the FIN, and thus according to RFC 5961 a challenge
5146 * ACK should be sent. However, Mac OSX rate limits replies to challenge
5147 * ACKs on the closed socket. In addition middleboxes can drop either the
5148 * challenge ACK or a subsequent RST.
5150 static bool tcp_reset_check(const struct sock
*sk
, const struct sk_buff
*skb
)
5152 struct tcp_sock
*tp
= tcp_sk(sk
);
5154 return unlikely(TCP_SKB_CB(skb
)->seq
== (tp
->rcv_nxt
- 1) &&
5155 (1 << sk
->sk_state
) & (TCPF_CLOSE_WAIT
| TCPF_LAST_ACK
|
5159 /* Does PAWS and seqno based validation of an incoming segment, flags will
5160 * play significant role here.
5162 static bool tcp_validate_incoming(struct sock
*sk
, struct sk_buff
*skb
,
5163 const struct tcphdr
*th
, int syn_inerr
)
5165 struct tcp_sock
*tp
= tcp_sk(sk
);
5166 bool rst_seq_match
= false;
5168 /* RFC1323: H1. Apply PAWS check first. */
5169 if (tcp_fast_parse_options(sock_net(sk
), skb
, th
, tp
) &&
5170 tp
->rx_opt
.saw_tstamp
&&
5171 tcp_paws_discard(sk
, skb
)) {
5173 NET_INC_STATS(sock_net(sk
), LINUX_MIB_PAWSESTABREJECTED
);
5174 if (!tcp_oow_rate_limited(sock_net(sk
), skb
,
5175 LINUX_MIB_TCPACKSKIPPEDPAWS
,
5176 &tp
->last_oow_ack_time
))
5177 tcp_send_dupack(sk
, skb
);
5180 /* Reset is accepted even if it did not pass PAWS. */
5183 /* Step 1: check sequence number */
5184 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5185 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5186 * (RST) segments are validated by checking their SEQ-fields."
5187 * And page 69: "If an incoming segment is not acceptable,
5188 * an acknowledgment should be sent in reply (unless the RST
5189 * bit is set, if so drop the segment and return)".
5194 if (!tcp_oow_rate_limited(sock_net(sk
), skb
,
5195 LINUX_MIB_TCPACKSKIPPEDSEQ
,
5196 &tp
->last_oow_ack_time
))
5197 tcp_send_dupack(sk
, skb
);
5198 } else if (tcp_reset_check(sk
, skb
)) {
5204 /* Step 2: check RST bit */
5206 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a
5207 * FIN and SACK too if available):
5208 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or
5209 * the right-most SACK block,
5211 * RESET the connection
5213 * Send a challenge ACK
5215 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
||
5216 tcp_reset_check(sk
, skb
)) {
5217 rst_seq_match
= true;
5218 } else if (tcp_is_sack(tp
) && tp
->rx_opt
.num_sacks
> 0) {
5219 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
5220 int max_sack
= sp
[0].end_seq
;
5223 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
;
5225 max_sack
= after(sp
[this_sack
].end_seq
,
5227 sp
[this_sack
].end_seq
: max_sack
;
5230 if (TCP_SKB_CB(skb
)->seq
== max_sack
)
5231 rst_seq_match
= true;
5237 /* Disable TFO if RST is out-of-order
5238 * and no data has been received
5239 * for current active TFO socket
5241 if (tp
->syn_fastopen
&& !tp
->data_segs_in
&&
5242 sk
->sk_state
== TCP_ESTABLISHED
)
5243 tcp_fastopen_active_disable(sk
);
5244 tcp_send_challenge_ack(sk
, skb
);
5249 /* step 3: check security and precedence [ignored] */
5251 /* step 4: Check for a SYN
5252 * RFC 5961 4.2 : Send a challenge ack
5257 TCP_INC_STATS(sock_net(sk
), TCP_MIB_INERRS
);
5258 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPSYNCHALLENGE
);
5259 tcp_send_challenge_ack(sk
, skb
);
5271 * TCP receive function for the ESTABLISHED state.
5273 * It is split into a fast path and a slow path. The fast path is
5275 * - A zero window was announced from us - zero window probing
5276 * is only handled properly in the slow path.
5277 * - Out of order segments arrived.
5278 * - Urgent data is expected.
5279 * - There is no buffer space left
5280 * - Unexpected TCP flags/window values/header lengths are received
5281 * (detected by checking the TCP header against pred_flags)
5282 * - Data is sent in both directions. Fast path only supports pure senders
5283 * or pure receivers (this means either the sequence number or the ack
5284 * value must stay constant)
5285 * - Unexpected TCP option.
5287 * When these conditions are not satisfied it drops into a standard
5288 * receive procedure patterned after RFC793 to handle all cases.
5289 * The first three cases are guaranteed by proper pred_flags setting,
5290 * the rest is checked inline. Fast processing is turned on in
5291 * tcp_data_queue when everything is OK.
5293 void tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
5294 const struct tcphdr
*th
)
5296 unsigned int len
= skb
->len
;
5297 struct tcp_sock
*tp
= tcp_sk(sk
);
5299 tcp_mstamp_refresh(tp
);
5300 if (unlikely(!sk
->sk_rx_dst
))
5301 inet_csk(sk
)->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
5303 * Header prediction.
5304 * The code loosely follows the one in the famous
5305 * "30 instruction TCP receive" Van Jacobson mail.
5307 * Van's trick is to deposit buffers into socket queue
5308 * on a device interrupt, to call tcp_recv function
5309 * on the receive process context and checksum and copy
5310 * the buffer to user space. smart...
5312 * Our current scheme is not silly either but we take the
5313 * extra cost of the net_bh soft interrupt processing...
5314 * We do checksum and copy also but from device to kernel.
5317 tp
->rx_opt
.saw_tstamp
= 0;
5319 /* pred_flags is 0xS?10 << 16 + snd_wnd
5320 * if header_prediction is to be made
5321 * 'S' will always be tp->tcp_header_len >> 2
5322 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5323 * turn it off (when there are holes in the receive
5324 * space for instance)
5325 * PSH flag is ignored.
5328 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
5329 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
&&
5330 !after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
)) {
5331 int tcp_header_len
= tp
->tcp_header_len
;
5333 /* Timestamp header prediction: tcp_header_len
5334 * is automatically equal to th->doff*4 due to pred_flags
5338 /* Check timestamp */
5339 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
5340 /* No? Slow path! */
5341 if (!tcp_parse_aligned_timestamp(tp
, th
))
5344 /* If PAWS failed, check it more carefully in slow path */
5345 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
5348 /* DO NOT update ts_recent here, if checksum fails
5349 * and timestamp was corrupted part, it will result
5350 * in a hung connection since we will drop all
5351 * future packets due to the PAWS test.
5355 if (len
<= tcp_header_len
) {
5356 /* Bulk data transfer: sender */
5357 if (len
== tcp_header_len
) {
5358 /* Predicted packet is in window by definition.
5359 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5360 * Hence, check seq<=rcv_wup reduces to:
5362 if (tcp_header_len
==
5363 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5364 tp
->rcv_nxt
== tp
->rcv_wup
)
5365 tcp_store_ts_recent(tp
);
5367 /* We know that such packets are checksummed
5370 tcp_ack(sk
, skb
, 0);
5372 tcp_data_snd_check(sk
);
5374 } else { /* Header too small */
5375 TCP_INC_STATS(sock_net(sk
), TCP_MIB_INERRS
);
5380 bool fragstolen
= false;
5382 if (tcp_checksum_complete(skb
))
5385 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
5388 /* Predicted packet is in window by definition.
5389 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5390 * Hence, check seq<=rcv_wup reduces to:
5392 if (tcp_header_len
==
5393 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5394 tp
->rcv_nxt
== tp
->rcv_wup
)
5395 tcp_store_ts_recent(tp
);
5397 tcp_rcv_rtt_measure_ts(sk
, skb
);
5399 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPHPHITS
);
5401 /* Bulk data transfer: receiver */
5402 eaten
= tcp_queue_rcv(sk
, skb
, tcp_header_len
,
5405 tcp_event_data_recv(sk
, skb
);
5407 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
5408 /* Well, only one small jumplet in fast path... */
5409 tcp_ack(sk
, skb
, FLAG_DATA
);
5410 tcp_data_snd_check(sk
);
5411 if (!inet_csk_ack_scheduled(sk
))
5415 __tcp_ack_snd_check(sk
, 0);
5418 kfree_skb_partial(skb
, fragstolen
);
5419 sk
->sk_data_ready(sk
);
5425 if (len
< (th
->doff
<< 2) || tcp_checksum_complete(skb
))
5428 if (!th
->ack
&& !th
->rst
&& !th
->syn
)
5432 * Standard slow path.
5435 if (!tcp_validate_incoming(sk
, skb
, th
, 1))
5439 if (tcp_ack(sk
, skb
, FLAG_SLOWPATH
| FLAG_UPDATE_TS_RECENT
) < 0)
5442 tcp_rcv_rtt_measure_ts(sk
, skb
);
5444 /* Process urgent data. */
5445 tcp_urg(sk
, skb
, th
);
5447 /* step 7: process the segment text */
5448 tcp_data_queue(sk
, skb
);
5450 tcp_data_snd_check(sk
);
5451 tcp_ack_snd_check(sk
);
5455 TCP_INC_STATS(sock_net(sk
), TCP_MIB_CSUMERRORS
);
5456 TCP_INC_STATS(sock_net(sk
), TCP_MIB_INERRS
);
5461 EXPORT_SYMBOL(tcp_rcv_established
);
5463 void tcp_finish_connect(struct sock
*sk
, struct sk_buff
*skb
)
5465 struct tcp_sock
*tp
= tcp_sk(sk
);
5466 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5468 tcp_set_state(sk
, TCP_ESTABLISHED
);
5469 icsk
->icsk_ack
.lrcvtime
= tcp_jiffies32
;
5472 icsk
->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
5473 security_inet_conn_established(sk
, skb
);
5476 tcp_init_transfer(sk
, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB
);
5478 /* Prevent spurious tcp_cwnd_restart() on first data
5481 tp
->lsndtime
= tcp_jiffies32
;
5483 if (sock_flag(sk
, SOCK_KEEPOPEN
))
5484 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
5486 if (!tp
->rx_opt
.snd_wscale
)
5487 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
5492 static bool tcp_rcv_fastopen_synack(struct sock
*sk
, struct sk_buff
*synack
,
5493 struct tcp_fastopen_cookie
*cookie
)
5495 struct tcp_sock
*tp
= tcp_sk(sk
);
5496 struct sk_buff
*data
= tp
->syn_data
? tcp_rtx_queue_head(sk
) : NULL
;
5497 u16 mss
= tp
->rx_opt
.mss_clamp
, try_exp
= 0;
5498 bool syn_drop
= false;
5500 if (mss
== tp
->rx_opt
.user_mss
) {
5501 struct tcp_options_received opt
;
5503 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5504 tcp_clear_options(&opt
);
5505 opt
.user_mss
= opt
.mss_clamp
= 0;
5506 tcp_parse_options(sock_net(sk
), synack
, &opt
, 0, NULL
);
5507 mss
= opt
.mss_clamp
;
5510 if (!tp
->syn_fastopen
) {
5511 /* Ignore an unsolicited cookie */
5513 } else if (tp
->total_retrans
) {
5514 /* SYN timed out and the SYN-ACK neither has a cookie nor
5515 * acknowledges data. Presumably the remote received only
5516 * the retransmitted (regular) SYNs: either the original
5517 * SYN-data or the corresponding SYN-ACK was dropped.
5519 syn_drop
= (cookie
->len
< 0 && data
);
5520 } else if (cookie
->len
< 0 && !tp
->syn_data
) {
5521 /* We requested a cookie but didn't get it. If we did not use
5522 * the (old) exp opt format then try so next time (try_exp=1).
5523 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5525 try_exp
= tp
->syn_fastopen_exp
? 2 : 1;
5528 tcp_fastopen_cache_set(sk
, mss
, cookie
, syn_drop
, try_exp
);
5530 if (data
) { /* Retransmit unacked data in SYN */
5531 skb_rbtree_walk_from(data
) {
5532 if (__tcp_retransmit_skb(sk
, data
, 1))
5536 NET_INC_STATS(sock_net(sk
),
5537 LINUX_MIB_TCPFASTOPENACTIVEFAIL
);
5540 tp
->syn_data_acked
= tp
->syn_data
;
5541 if (tp
->syn_data_acked
)
5542 NET_INC_STATS(sock_net(sk
),
5543 LINUX_MIB_TCPFASTOPENACTIVE
);
5545 tcp_fastopen_add_skb(sk
, synack
);
5550 static void smc_check_reset_syn(struct tcp_sock
*tp
)
5552 #if IS_ENABLED(CONFIG_SMC)
5553 if (static_branch_unlikely(&tcp_have_smc
)) {
5554 if (tp
->syn_smc
&& !tp
->rx_opt
.smc_ok
)
5560 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5561 const struct tcphdr
*th
)
5563 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5564 struct tcp_sock
*tp
= tcp_sk(sk
);
5565 struct tcp_fastopen_cookie foc
= { .len
= -1 };
5566 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
5569 tcp_parse_options(sock_net(sk
), skb
, &tp
->rx_opt
, 0, &foc
);
5570 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
5571 tp
->rx_opt
.rcv_tsecr
-= tp
->tsoffset
;
5575 * "If the state is SYN-SENT then
5576 * first check the ACK bit
5577 * If the ACK bit is set
5578 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5579 * a reset (unless the RST bit is set, if so drop
5580 * the segment and return)"
5582 if (!after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_una
) ||
5583 after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
))
5584 goto reset_and_undo
;
5586 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
5587 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
5588 tcp_time_stamp(tp
))) {
5589 NET_INC_STATS(sock_net(sk
),
5590 LINUX_MIB_PAWSACTIVEREJECTED
);
5591 goto reset_and_undo
;
5594 /* Now ACK is acceptable.
5596 * "If the RST bit is set
5597 * If the ACK was acceptable then signal the user "error:
5598 * connection reset", drop the segment, enter CLOSED state,
5599 * delete TCB, and return."
5608 * "fifth, if neither of the SYN or RST bits is set then
5609 * drop the segment and return."
5615 goto discard_and_undo
;
5618 * "If the SYN bit is on ...
5619 * are acceptable then ...
5620 * (our SYN has been ACKed), change the connection
5621 * state to ESTABLISHED..."
5624 tcp_ecn_rcv_synack(tp
, th
);
5626 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5627 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
5629 /* Ok.. it's good. Set up sequence numbers and
5630 * move to established.
5632 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5633 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5635 /* RFC1323: The window in SYN & SYN/ACK segments is
5638 tp
->snd_wnd
= ntohs(th
->window
);
5640 if (!tp
->rx_opt
.wscale_ok
) {
5641 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
5642 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
5645 if (tp
->rx_opt
.saw_tstamp
) {
5646 tp
->rx_opt
.tstamp_ok
= 1;
5647 tp
->tcp_header_len
=
5648 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5649 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5650 tcp_store_ts_recent(tp
);
5652 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5655 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5656 tcp_initialize_rcv_mss(sk
);
5658 /* Remember, tcp_poll() does not lock socket!
5659 * Change state from SYN-SENT only after copied_seq
5660 * is initialized. */
5661 tp
->copied_seq
= tp
->rcv_nxt
;
5663 smc_check_reset_syn(tp
);
5667 tcp_finish_connect(sk
, skb
);
5669 fastopen_fail
= (tp
->syn_fastopen
|| tp
->syn_data
) &&
5670 tcp_rcv_fastopen_synack(sk
, skb
, &foc
);
5672 if (!sock_flag(sk
, SOCK_DEAD
)) {
5673 sk
->sk_state_change(sk
);
5674 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5678 if (sk
->sk_write_pending
||
5679 icsk
->icsk_accept_queue
.rskq_defer_accept
||
5680 icsk
->icsk_ack
.pingpong
) {
5681 /* Save one ACK. Data will be ready after
5682 * several ticks, if write_pending is set.
5684 * It may be deleted, but with this feature tcpdumps
5685 * look so _wonderfully_ clever, that I was not able
5686 * to stand against the temptation 8) --ANK
5688 inet_csk_schedule_ack(sk
);
5689 tcp_enter_quickack_mode(sk
);
5690 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
5691 TCP_DELACK_MAX
, TCP_RTO_MAX
);
5702 /* No ACK in the segment */
5706 * "If the RST bit is set
5708 * Otherwise (no ACK) drop the segment and return."
5711 goto discard_and_undo
;
5715 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&&
5716 tcp_paws_reject(&tp
->rx_opt
, 0))
5717 goto discard_and_undo
;
5720 /* We see SYN without ACK. It is attempt of
5721 * simultaneous connect with crossed SYNs.
5722 * Particularly, it can be connect to self.
5724 tcp_set_state(sk
, TCP_SYN_RECV
);
5726 if (tp
->rx_opt
.saw_tstamp
) {
5727 tp
->rx_opt
.tstamp_ok
= 1;
5728 tcp_store_ts_recent(tp
);
5729 tp
->tcp_header_len
=
5730 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5732 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5735 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5736 tp
->copied_seq
= tp
->rcv_nxt
;
5737 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5739 /* RFC1323: The window in SYN & SYN/ACK segments is
5742 tp
->snd_wnd
= ntohs(th
->window
);
5743 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
5744 tp
->max_window
= tp
->snd_wnd
;
5746 tcp_ecn_rcv_syn(tp
, th
);
5749 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5750 tcp_initialize_rcv_mss(sk
);
5752 tcp_send_synack(sk
);
5754 /* Note, we could accept data and URG from this segment.
5755 * There are no obstacles to make this (except that we must
5756 * either change tcp_recvmsg() to prevent it from returning data
5757 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5759 * However, if we ignore data in ACKless segments sometimes,
5760 * we have no reasons to accept it sometimes.
5761 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5762 * is not flawless. So, discard packet for sanity.
5763 * Uncomment this return to process the data.
5770 /* "fifth, if neither of the SYN or RST bits is set then
5771 * drop the segment and return."
5775 tcp_clear_options(&tp
->rx_opt
);
5776 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5780 tcp_clear_options(&tp
->rx_opt
);
5781 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5786 * This function implements the receiving procedure of RFC 793 for
5787 * all states except ESTABLISHED and TIME_WAIT.
5788 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5789 * address independent.
5792 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
)
5794 struct tcp_sock
*tp
= tcp_sk(sk
);
5795 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5796 const struct tcphdr
*th
= tcp_hdr(skb
);
5797 struct request_sock
*req
;
5801 switch (sk
->sk_state
) {
5815 /* It is possible that we process SYN packets from backlog,
5816 * so we need to make sure to disable BH right there.
5819 acceptable
= icsk
->icsk_af_ops
->conn_request(sk
, skb
) >= 0;
5830 tp
->rx_opt
.saw_tstamp
= 0;
5831 tcp_mstamp_refresh(tp
);
5832 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
);
5836 /* Do step6 onward by hand. */
5837 tcp_urg(sk
, skb
, th
);
5839 tcp_data_snd_check(sk
);
5843 tcp_mstamp_refresh(tp
);
5844 tp
->rx_opt
.saw_tstamp
= 0;
5845 req
= tp
->fastopen_rsk
;
5847 WARN_ON_ONCE(sk
->sk_state
!= TCP_SYN_RECV
&&
5848 sk
->sk_state
!= TCP_FIN_WAIT1
);
5850 if (!tcp_check_req(sk
, skb
, req
, true))
5854 if (!th
->ack
&& !th
->rst
&& !th
->syn
)
5857 if (!tcp_validate_incoming(sk
, skb
, th
, 0))
5860 /* step 5: check the ACK field */
5861 acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
|
5862 FLAG_UPDATE_TS_RECENT
|
5863 FLAG_NO_CHALLENGE_ACK
) > 0;
5866 if (sk
->sk_state
== TCP_SYN_RECV
)
5867 return 1; /* send one RST */
5868 tcp_send_challenge_ack(sk
, skb
);
5871 switch (sk
->sk_state
) {
5874 tcp_synack_rtt_meas(sk
, req
);
5876 /* Once we leave TCP_SYN_RECV, we no longer need req
5880 inet_csk(sk
)->icsk_retransmits
= 0;
5881 reqsk_fastopen_remove(sk
, req
, false);
5882 /* Re-arm the timer because data may have been sent out.
5883 * This is similar to the regular data transmission case
5884 * when new data has just been ack'ed.
5886 * (TFO) - we could try to be more aggressive and
5887 * retransmitting any data sooner based on when they
5892 tcp_init_transfer(sk
, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB
);
5893 tp
->copied_seq
= tp
->rcv_nxt
;
5896 tcp_set_state(sk
, TCP_ESTABLISHED
);
5897 sk
->sk_state_change(sk
);
5899 /* Note, that this wakeup is only for marginal crossed SYN case.
5900 * Passively open sockets are not waked up, because
5901 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5904 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5906 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
5907 tp
->snd_wnd
= ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
;
5908 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5910 if (tp
->rx_opt
.tstamp_ok
)
5911 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5913 if (!inet_csk(sk
)->icsk_ca_ops
->cong_control
)
5914 tcp_update_pacing_rate(sk
);
5916 /* Prevent spurious tcp_cwnd_restart() on first data packet */
5917 tp
->lsndtime
= tcp_jiffies32
;
5919 tcp_initialize_rcv_mss(sk
);
5920 tcp_fast_path_on(tp
);
5923 case TCP_FIN_WAIT1
: {
5926 /* If we enter the TCP_FIN_WAIT1 state and we are a
5927 * Fast Open socket and this is the first acceptable
5928 * ACK we have received, this would have acknowledged
5929 * our SYNACK so stop the SYNACK timer.
5932 /* We no longer need the request sock. */
5933 reqsk_fastopen_remove(sk
, req
, false);
5936 if (tp
->snd_una
!= tp
->write_seq
)
5939 tcp_set_state(sk
, TCP_FIN_WAIT2
);
5940 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
5944 if (!sock_flag(sk
, SOCK_DEAD
)) {
5945 /* Wake up lingering close() */
5946 sk
->sk_state_change(sk
);
5950 if (tp
->linger2
< 0) {
5952 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
5955 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5956 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
5957 /* Receive out of order FIN after close() */
5958 if (tp
->syn_fastopen
&& th
->fin
)
5959 tcp_fastopen_active_disable(sk
);
5961 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
5965 tmo
= tcp_fin_time(sk
);
5966 if (tmo
> TCP_TIMEWAIT_LEN
) {
5967 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
5968 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
5969 /* Bad case. We could lose such FIN otherwise.
5970 * It is not a big problem, but it looks confusing
5971 * and not so rare event. We still can lose it now,
5972 * if it spins in bh_lock_sock(), but it is really
5975 inet_csk_reset_keepalive_timer(sk
, tmo
);
5977 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
5984 if (tp
->snd_una
== tp
->write_seq
) {
5985 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
5991 if (tp
->snd_una
== tp
->write_seq
) {
5992 tcp_update_metrics(sk
);
5999 /* step 6: check the URG bit */
6000 tcp_urg(sk
, skb
, th
);
6002 /* step 7: process the segment text */
6003 switch (sk
->sk_state
) {
6004 case TCP_CLOSE_WAIT
:
6007 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
6012 /* RFC 793 says to queue data in these states,
6013 * RFC 1122 says we MUST send a reset.
6014 * BSD 4.4 also does reset.
6016 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
6017 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
6018 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
6019 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
6025 case TCP_ESTABLISHED
:
6026 tcp_data_queue(sk
, skb
);
6031 /* tcp_data could move socket to TIME-WAIT */
6032 if (sk
->sk_state
!= TCP_CLOSE
) {
6033 tcp_data_snd_check(sk
);
6034 tcp_ack_snd_check(sk
);
6043 EXPORT_SYMBOL(tcp_rcv_state_process
);
6045 static inline void pr_drop_req(struct request_sock
*req
, __u16 port
, int family
)
6047 struct inet_request_sock
*ireq
= inet_rsk(req
);
6049 if (family
== AF_INET
)
6050 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6051 &ireq
->ir_rmt_addr
, port
);
6052 #if IS_ENABLED(CONFIG_IPV6)
6053 else if (family
== AF_INET6
)
6054 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6055 &ireq
->ir_v6_rmt_addr
, port
);
6059 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6061 * If we receive a SYN packet with these bits set, it means a
6062 * network is playing bad games with TOS bits. In order to
6063 * avoid possible false congestion notifications, we disable
6064 * TCP ECN negotiation.
6066 * Exception: tcp_ca wants ECN. This is required for DCTCP
6067 * congestion control: Linux DCTCP asserts ECT on all packets,
6068 * including SYN, which is most optimal solution; however,
6069 * others, such as FreeBSD do not.
6071 static void tcp_ecn_create_request(struct request_sock
*req
,
6072 const struct sk_buff
*skb
,
6073 const struct sock
*listen_sk
,
6074 const struct dst_entry
*dst
)
6076 const struct tcphdr
*th
= tcp_hdr(skb
);
6077 const struct net
*net
= sock_net(listen_sk
);
6078 bool th_ecn
= th
->ece
&& th
->cwr
;
6085 ect
= !INET_ECN_is_not_ect(TCP_SKB_CB(skb
)->ip_dsfield
);
6086 ecn_ok_dst
= dst_feature(dst
, DST_FEATURE_ECN_MASK
);
6087 ecn_ok
= net
->ipv4
.sysctl_tcp_ecn
|| ecn_ok_dst
;
6089 if ((!ect
&& ecn_ok
) || tcp_ca_needs_ecn(listen_sk
) ||
6090 (ecn_ok_dst
& DST_FEATURE_ECN_CA
) ||
6091 tcp_bpf_ca_needs_ecn((struct sock
*)req
))
6092 inet_rsk(req
)->ecn_ok
= 1;
6095 static void tcp_openreq_init(struct request_sock
*req
,
6096 const struct tcp_options_received
*rx_opt
,
6097 struct sk_buff
*skb
, const struct sock
*sk
)
6099 struct inet_request_sock
*ireq
= inet_rsk(req
);
6101 req
->rsk_rcv_wnd
= 0; /* So that tcp_send_synack() knows! */
6103 tcp_rsk(req
)->rcv_isn
= TCP_SKB_CB(skb
)->seq
;
6104 tcp_rsk(req
)->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
6105 tcp_rsk(req
)->snt_synack
= tcp_clock_us();
6106 tcp_rsk(req
)->last_oow_ack_time
= 0;
6107 req
->mss
= rx_opt
->mss_clamp
;
6108 req
->ts_recent
= rx_opt
->saw_tstamp
? rx_opt
->rcv_tsval
: 0;
6109 ireq
->tstamp_ok
= rx_opt
->tstamp_ok
;
6110 ireq
->sack_ok
= rx_opt
->sack_ok
;
6111 ireq
->snd_wscale
= rx_opt
->snd_wscale
;
6112 ireq
->wscale_ok
= rx_opt
->wscale_ok
;
6115 ireq
->ir_rmt_port
= tcp_hdr(skb
)->source
;
6116 ireq
->ir_num
= ntohs(tcp_hdr(skb
)->dest
);
6117 ireq
->ir_mark
= inet_request_mark(sk
, skb
);
6118 #if IS_ENABLED(CONFIG_SMC)
6119 ireq
->smc_ok
= rx_opt
->smc_ok
;
6123 struct request_sock
*inet_reqsk_alloc(const struct request_sock_ops
*ops
,
6124 struct sock
*sk_listener
,
6125 bool attach_listener
)
6127 struct request_sock
*req
= reqsk_alloc(ops
, sk_listener
,
6131 struct inet_request_sock
*ireq
= inet_rsk(req
);
6133 ireq
->ireq_opt
= NULL
;
6134 #if IS_ENABLED(CONFIG_IPV6)
6135 ireq
->pktopts
= NULL
;
6137 atomic64_set(&ireq
->ir_cookie
, 0);
6138 ireq
->ireq_state
= TCP_NEW_SYN_RECV
;
6139 write_pnet(&ireq
->ireq_net
, sock_net(sk_listener
));
6140 ireq
->ireq_family
= sk_listener
->sk_family
;
6145 EXPORT_SYMBOL(inet_reqsk_alloc
);
6148 * Return true if a syncookie should be sent
6150 static bool tcp_syn_flood_action(const struct sock
*sk
,
6151 const struct sk_buff
*skb
,
6154 struct request_sock_queue
*queue
= &inet_csk(sk
)->icsk_accept_queue
;
6155 const char *msg
= "Dropping request";
6156 bool want_cookie
= false;
6157 struct net
*net
= sock_net(sk
);
6159 #ifdef CONFIG_SYN_COOKIES
6160 if (net
->ipv4
.sysctl_tcp_syncookies
) {
6161 msg
= "Sending cookies";
6163 __NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPREQQFULLDOCOOKIES
);
6166 __NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPREQQFULLDROP
);
6168 if (!queue
->synflood_warned
&&
6169 net
->ipv4
.sysctl_tcp_syncookies
!= 2 &&
6170 xchg(&queue
->synflood_warned
, 1) == 0)
6171 pr_info("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6172 proto
, ntohs(tcp_hdr(skb
)->dest
), msg
);
6177 static void tcp_reqsk_record_syn(const struct sock
*sk
,
6178 struct request_sock
*req
,
6179 const struct sk_buff
*skb
)
6181 if (tcp_sk(sk
)->save_syn
) {
6182 u32 len
= skb_network_header_len(skb
) + tcp_hdrlen(skb
);
6185 copy
= kmalloc(len
+ sizeof(u32
), GFP_ATOMIC
);
6188 memcpy(©
[1], skb_network_header(skb
), len
);
6189 req
->saved_syn
= copy
;
6194 int tcp_conn_request(struct request_sock_ops
*rsk_ops
,
6195 const struct tcp_request_sock_ops
*af_ops
,
6196 struct sock
*sk
, struct sk_buff
*skb
)
6198 struct tcp_fastopen_cookie foc
= { .len
= -1 };
6199 __u32 isn
= TCP_SKB_CB(skb
)->tcp_tw_isn
;
6200 struct tcp_options_received tmp_opt
;
6201 struct tcp_sock
*tp
= tcp_sk(sk
);
6202 struct net
*net
= sock_net(sk
);
6203 struct sock
*fastopen_sk
= NULL
;
6204 struct request_sock
*req
;
6205 bool want_cookie
= false;
6206 struct dst_entry
*dst
;
6209 /* TW buckets are converted to open requests without
6210 * limitations, they conserve resources and peer is
6211 * evidently real one.
6213 if ((net
->ipv4
.sysctl_tcp_syncookies
== 2 ||
6214 inet_csk_reqsk_queue_is_full(sk
)) && !isn
) {
6215 want_cookie
= tcp_syn_flood_action(sk
, skb
, rsk_ops
->slab_name
);
6220 if (sk_acceptq_is_full(sk
)) {
6221 NET_INC_STATS(sock_net(sk
), LINUX_MIB_LISTENOVERFLOWS
);
6225 req
= inet_reqsk_alloc(rsk_ops
, sk
, !want_cookie
);
6229 tcp_rsk(req
)->af_specific
= af_ops
;
6230 tcp_rsk(req
)->ts_off
= 0;
6232 tcp_clear_options(&tmp_opt
);
6233 tmp_opt
.mss_clamp
= af_ops
->mss_clamp
;
6234 tmp_opt
.user_mss
= tp
->rx_opt
.user_mss
;
6235 tcp_parse_options(sock_net(sk
), skb
, &tmp_opt
, 0,
6236 want_cookie
? NULL
: &foc
);
6238 if (want_cookie
&& !tmp_opt
.saw_tstamp
)
6239 tcp_clear_options(&tmp_opt
);
6241 tmp_opt
.tstamp_ok
= tmp_opt
.saw_tstamp
;
6242 tcp_openreq_init(req
, &tmp_opt
, skb
, sk
);
6243 inet_rsk(req
)->no_srccheck
= inet_sk(sk
)->transparent
;
6245 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6246 inet_rsk(req
)->ir_iif
= inet_request_bound_dev_if(sk
, skb
);
6248 af_ops
->init_req(req
, sk
, skb
);
6250 if (security_inet_conn_request(sk
, skb
, req
))
6253 if (tmp_opt
.tstamp_ok
)
6254 tcp_rsk(req
)->ts_off
= af_ops
->init_ts_off(net
, skb
);
6256 dst
= af_ops
->route_req(sk
, &fl
, req
);
6260 if (!want_cookie
&& !isn
) {
6261 /* Kill the following clause, if you dislike this way. */
6262 if (!net
->ipv4
.sysctl_tcp_syncookies
&&
6263 (net
->ipv4
.sysctl_max_syn_backlog
- inet_csk_reqsk_queue_len(sk
) <
6264 (net
->ipv4
.sysctl_max_syn_backlog
>> 2)) &&
6265 !tcp_peer_is_proven(req
, dst
)) {
6266 /* Without syncookies last quarter of
6267 * backlog is filled with destinations,
6268 * proven to be alive.
6269 * It means that we continue to communicate
6270 * to destinations, already remembered
6271 * to the moment of synflood.
6273 pr_drop_req(req
, ntohs(tcp_hdr(skb
)->source
),
6275 goto drop_and_release
;
6278 isn
= af_ops
->init_seq(skb
);
6281 tcp_ecn_create_request(req
, skb
, sk
, dst
);
6284 isn
= cookie_init_sequence(af_ops
, sk
, skb
, &req
->mss
);
6285 req
->cookie_ts
= tmp_opt
.tstamp_ok
;
6286 if (!tmp_opt
.tstamp_ok
)
6287 inet_rsk(req
)->ecn_ok
= 0;
6290 tcp_rsk(req
)->snt_isn
= isn
;
6291 tcp_rsk(req
)->txhash
= net_tx_rndhash();
6292 tcp_openreq_init_rwin(req
, sk
, dst
);
6294 tcp_reqsk_record_syn(sk
, req
, skb
);
6295 fastopen_sk
= tcp_try_fastopen(sk
, skb
, req
, &foc
, dst
);
6298 af_ops
->send_synack(fastopen_sk
, dst
, &fl
, req
,
6299 &foc
, TCP_SYNACK_FASTOPEN
);
6300 /* Add the child socket directly into the accept queue */
6301 inet_csk_reqsk_queue_add(sk
, req
, fastopen_sk
);
6302 sk
->sk_data_ready(sk
);
6303 bh_unlock_sock(fastopen_sk
);
6304 sock_put(fastopen_sk
);
6306 tcp_rsk(req
)->tfo_listener
= false;
6308 inet_csk_reqsk_queue_hash_add(sk
, req
,
6309 tcp_timeout_init((struct sock
*)req
));
6310 af_ops
->send_synack(sk
, dst
, &fl
, req
, &foc
,
6311 !want_cookie
? TCP_SYNACK_NORMAL
:
6329 EXPORT_SYMBOL(tcp_conn_request
);