2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
23 * Pedro Roque : Fast Retransmit/Recovery.
25 * Retransmit queue handled by TCP.
26 * Better retransmit timer handling.
27 * New congestion avoidance.
31 * Eric : Fast Retransmit.
32 * Randy Scott : MSS option defines.
33 * Eric Schenk : Fixes to slow start algorithm.
34 * Eric Schenk : Yet another double ACK bug.
35 * Eric Schenk : Delayed ACK bug fixes.
36 * Eric Schenk : Floyd style fast retrans war avoidance.
37 * David S. Miller : Don't allow zero congestion window.
38 * Eric Schenk : Fix retransmitter so that it sends
39 * next packet on ack of previous packet.
40 * Andi Kleen : Moved open_request checking here
41 * and process RSTs for open_requests.
42 * Andi Kleen : Better prune_queue, and other fixes.
43 * Andrey Savochkin: Fix RTT measurements in the presence of
45 * Andrey Savochkin: Check sequence numbers correctly when
46 * removing SACKs due to in sequence incoming
48 * Andi Kleen: Make sure we never ack data there is not
49 * enough room for. Also make this condition
50 * a fatal error if it might still happen.
51 * Andi Kleen: Add tcp_measure_rcv_mss to make
52 * connections with MSS<min(MTU,ann. MSS)
53 * work without delayed acks.
54 * Andi Kleen: Process packets with PSH set in the
56 * J Hadi Salim: ECN support
59 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
60 * engine. Lots of bugs are found.
61 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
64 #define pr_fmt(fmt) "TCP: " fmt
67 #include <linux/slab.h>
68 #include <linux/module.h>
69 #include <linux/sysctl.h>
70 #include <linux/kernel.h>
71 #include <linux/prefetch.h>
74 #include <net/inet_common.h>
75 #include <linux/ipsec.h>
76 #include <asm/unaligned.h>
77 #include <linux/errqueue.h>
79 int sysctl_tcp_timestamps __read_mostly
= 1;
80 int sysctl_tcp_window_scaling __read_mostly
= 1;
81 int sysctl_tcp_sack __read_mostly
= 1;
82 int sysctl_tcp_fack __read_mostly
;
83 int sysctl_tcp_max_reordering __read_mostly
= 300;
84 int sysctl_tcp_dsack __read_mostly
= 1;
85 int sysctl_tcp_app_win __read_mostly
= 31;
86 int sysctl_tcp_adv_win_scale __read_mostly
= 1;
87 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale
);
88 EXPORT_SYMBOL(sysctl_tcp_timestamps
);
90 /* rfc5961 challenge ack rate limiting */
91 int sysctl_tcp_challenge_ack_limit
= 1000;
93 int sysctl_tcp_stdurg __read_mostly
;
94 int sysctl_tcp_rfc1337 __read_mostly
;
95 int sysctl_tcp_max_orphans __read_mostly
= NR_FILE
;
96 int sysctl_tcp_frto __read_mostly
= 2;
97 int sysctl_tcp_min_rtt_wlen __read_mostly
= 300;
98 int sysctl_tcp_moderate_rcvbuf __read_mostly
= 1;
99 int sysctl_tcp_early_retrans __read_mostly
= 3;
100 int sysctl_tcp_invalid_ratelimit __read_mostly
= HZ
/2;
102 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
103 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
104 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
105 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
106 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
107 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
108 #define FLAG_ECE 0x40 /* ECE in this ACK */
109 #define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */
110 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
111 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
112 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
113 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
114 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
115 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
117 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
118 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
119 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
120 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
122 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
123 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
125 #define REXMIT_NONE 0 /* no loss recovery to do */
126 #define REXMIT_LOST 1 /* retransmit packets marked lost */
127 #define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */
129 static void tcp_gro_dev_warn(struct sock
*sk
, const struct sk_buff
*skb
)
131 static bool __once __read_mostly
;
134 struct net_device
*dev
;
139 dev
= dev_get_by_index_rcu(sock_net(sk
), skb
->skb_iif
);
140 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n",
141 dev
? dev
->name
: "Unknown driver");
146 /* Adapt the MSS value used to make delayed ack decision to the
149 static void tcp_measure_rcv_mss(struct sock
*sk
, const struct sk_buff
*skb
)
151 struct inet_connection_sock
*icsk
= inet_csk(sk
);
152 const unsigned int lss
= icsk
->icsk_ack
.last_seg_size
;
155 icsk
->icsk_ack
.last_seg_size
= 0;
157 /* skb->len may jitter because of SACKs, even if peer
158 * sends good full-sized frames.
160 len
= skb_shinfo(skb
)->gso_size
? : skb
->len
;
161 if (len
>= icsk
->icsk_ack
.rcv_mss
) {
162 icsk
->icsk_ack
.rcv_mss
= min_t(unsigned int, len
,
164 if (unlikely(icsk
->icsk_ack
.rcv_mss
!= len
))
165 tcp_gro_dev_warn(sk
, skb
);
167 /* Otherwise, we make more careful check taking into account,
168 * that SACKs block is variable.
170 * "len" is invariant segment length, including TCP header.
172 len
+= skb
->data
- skb_transport_header(skb
);
173 if (len
>= TCP_MSS_DEFAULT
+ sizeof(struct tcphdr
) ||
174 /* If PSH is not set, packet should be
175 * full sized, provided peer TCP is not badly broken.
176 * This observation (if it is correct 8)) allows
177 * to handle super-low mtu links fairly.
179 (len
>= TCP_MIN_MSS
+ sizeof(struct tcphdr
) &&
180 !(tcp_flag_word(tcp_hdr(skb
)) & TCP_REMNANT
))) {
181 /* Subtract also invariant (if peer is RFC compliant),
182 * tcp header plus fixed timestamp option length.
183 * Resulting "len" is MSS free of SACK jitter.
185 len
-= tcp_sk(sk
)->tcp_header_len
;
186 icsk
->icsk_ack
.last_seg_size
= len
;
188 icsk
->icsk_ack
.rcv_mss
= len
;
192 if (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
)
193 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED2
;
194 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
198 static void tcp_incr_quickack(struct sock
*sk
)
200 struct inet_connection_sock
*icsk
= inet_csk(sk
);
201 unsigned int quickacks
= tcp_sk(sk
)->rcv_wnd
/ (2 * icsk
->icsk_ack
.rcv_mss
);
205 if (quickacks
> icsk
->icsk_ack
.quick
)
206 icsk
->icsk_ack
.quick
= min(quickacks
, TCP_MAX_QUICKACKS
);
209 static void tcp_enter_quickack_mode(struct sock
*sk
)
211 struct inet_connection_sock
*icsk
= inet_csk(sk
);
212 tcp_incr_quickack(sk
);
213 icsk
->icsk_ack
.pingpong
= 0;
214 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
217 /* Send ACKs quickly, if "quick" count is not exhausted
218 * and the session is not interactive.
221 static bool tcp_in_quickack_mode(struct sock
*sk
)
223 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
224 const struct dst_entry
*dst
= __sk_dst_get(sk
);
226 return (dst
&& dst_metric(dst
, RTAX_QUICKACK
)) ||
227 (icsk
->icsk_ack
.quick
&& !icsk
->icsk_ack
.pingpong
);
230 static void tcp_ecn_queue_cwr(struct tcp_sock
*tp
)
232 if (tp
->ecn_flags
& TCP_ECN_OK
)
233 tp
->ecn_flags
|= TCP_ECN_QUEUE_CWR
;
236 static void tcp_ecn_accept_cwr(struct tcp_sock
*tp
, const struct sk_buff
*skb
)
238 if (tcp_hdr(skb
)->cwr
)
239 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
242 static void tcp_ecn_withdraw_cwr(struct tcp_sock
*tp
)
244 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
247 static void __tcp_ecn_check_ce(struct tcp_sock
*tp
, const struct sk_buff
*skb
)
249 switch (TCP_SKB_CB(skb
)->ip_dsfield
& INET_ECN_MASK
) {
250 case INET_ECN_NOT_ECT
:
251 /* Funny extension: if ECT is not set on a segment,
252 * and we already seen ECT on a previous segment,
253 * it is probably a retransmit.
255 if (tp
->ecn_flags
& TCP_ECN_SEEN
)
256 tcp_enter_quickack_mode((struct sock
*)tp
);
259 if (tcp_ca_needs_ecn((struct sock
*)tp
))
260 tcp_ca_event((struct sock
*)tp
, CA_EVENT_ECN_IS_CE
);
262 if (!(tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)) {
263 /* Better not delay acks, sender can have a very low cwnd */
264 tcp_enter_quickack_mode((struct sock
*)tp
);
265 tp
->ecn_flags
|= TCP_ECN_DEMAND_CWR
;
267 tp
->ecn_flags
|= TCP_ECN_SEEN
;
270 if (tcp_ca_needs_ecn((struct sock
*)tp
))
271 tcp_ca_event((struct sock
*)tp
, CA_EVENT_ECN_NO_CE
);
272 tp
->ecn_flags
|= TCP_ECN_SEEN
;
277 static void tcp_ecn_check_ce(struct tcp_sock
*tp
, const struct sk_buff
*skb
)
279 if (tp
->ecn_flags
& TCP_ECN_OK
)
280 __tcp_ecn_check_ce(tp
, skb
);
283 static void tcp_ecn_rcv_synack(struct tcp_sock
*tp
, const struct tcphdr
*th
)
285 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| th
->cwr
))
286 tp
->ecn_flags
&= ~TCP_ECN_OK
;
289 static void tcp_ecn_rcv_syn(struct tcp_sock
*tp
, const struct tcphdr
*th
)
291 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| !th
->cwr
))
292 tp
->ecn_flags
&= ~TCP_ECN_OK
;
295 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock
*tp
, const struct tcphdr
*th
)
297 if (th
->ece
&& !th
->syn
&& (tp
->ecn_flags
& TCP_ECN_OK
))
302 /* Buffer size and advertised window tuning.
304 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
307 static void tcp_sndbuf_expand(struct sock
*sk
)
309 const struct tcp_sock
*tp
= tcp_sk(sk
);
310 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
314 /* Worst case is non GSO/TSO : each frame consumes one skb
315 * and skb->head is kmalloced using power of two area of memory
317 per_mss
= max_t(u32
, tp
->rx_opt
.mss_clamp
, tp
->mss_cache
) +
319 SKB_DATA_ALIGN(sizeof(struct skb_shared_info
));
321 per_mss
= roundup_pow_of_two(per_mss
) +
322 SKB_DATA_ALIGN(sizeof(struct sk_buff
));
324 nr_segs
= max_t(u32
, TCP_INIT_CWND
, tp
->snd_cwnd
);
325 nr_segs
= max_t(u32
, nr_segs
, tp
->reordering
+ 1);
327 /* Fast Recovery (RFC 5681 3.2) :
328 * Cubic needs 1.7 factor, rounded to 2 to include
329 * extra cushion (application might react slowly to POLLOUT)
331 sndmem
= ca_ops
->sndbuf_expand
? ca_ops
->sndbuf_expand(sk
) : 2;
332 sndmem
*= nr_segs
* per_mss
;
334 if (sk
->sk_sndbuf
< sndmem
)
335 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
338 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
340 * All tcp_full_space() is split to two parts: "network" buffer, allocated
341 * forward and advertised in receiver window (tp->rcv_wnd) and
342 * "application buffer", required to isolate scheduling/application
343 * latencies from network.
344 * window_clamp is maximal advertised window. It can be less than
345 * tcp_full_space(), in this case tcp_full_space() - window_clamp
346 * is reserved for "application" buffer. The less window_clamp is
347 * the smoother our behaviour from viewpoint of network, but the lower
348 * throughput and the higher sensitivity of the connection to losses. 8)
350 * rcv_ssthresh is more strict window_clamp used at "slow start"
351 * phase to predict further behaviour of this connection.
352 * It is used for two goals:
353 * - to enforce header prediction at sender, even when application
354 * requires some significant "application buffer". It is check #1.
355 * - to prevent pruning of receive queue because of misprediction
356 * of receiver window. Check #2.
358 * The scheme does not work when sender sends good segments opening
359 * window and then starts to feed us spaghetti. But it should work
360 * in common situations. Otherwise, we have to rely on queue collapsing.
363 /* Slow part of check#2. */
364 static int __tcp_grow_window(const struct sock
*sk
, const struct sk_buff
*skb
)
366 struct tcp_sock
*tp
= tcp_sk(sk
);
368 int truesize
= tcp_win_from_space(skb
->truesize
) >> 1;
369 int window
= tcp_win_from_space(sysctl_tcp_rmem
[2]) >> 1;
371 while (tp
->rcv_ssthresh
<= window
) {
372 if (truesize
<= skb
->len
)
373 return 2 * inet_csk(sk
)->icsk_ack
.rcv_mss
;
381 static void tcp_grow_window(struct sock
*sk
, const struct sk_buff
*skb
)
383 struct tcp_sock
*tp
= tcp_sk(sk
);
386 if (tp
->rcv_ssthresh
< tp
->window_clamp
&&
387 (int)tp
->rcv_ssthresh
< tcp_space(sk
) &&
388 !tcp_under_memory_pressure(sk
)) {
391 /* Check #2. Increase window, if skb with such overhead
392 * will fit to rcvbuf in future.
394 if (tcp_win_from_space(skb
->truesize
) <= skb
->len
)
395 incr
= 2 * tp
->advmss
;
397 incr
= __tcp_grow_window(sk
, skb
);
400 incr
= max_t(int, incr
, 2 * skb
->len
);
401 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
+ incr
,
403 inet_csk(sk
)->icsk_ack
.quick
|= 1;
408 /* 3. Tuning rcvbuf, when connection enters established state. */
409 static void tcp_fixup_rcvbuf(struct sock
*sk
)
411 u32 mss
= tcp_sk(sk
)->advmss
;
414 rcvmem
= 2 * SKB_TRUESIZE(mss
+ MAX_TCP_HEADER
) *
415 tcp_default_init_rwnd(mss
);
417 /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
418 * Allow enough cushion so that sender is not limited by our window
420 if (sysctl_tcp_moderate_rcvbuf
)
423 if (sk
->sk_rcvbuf
< rcvmem
)
424 sk
->sk_rcvbuf
= min(rcvmem
, sysctl_tcp_rmem
[2]);
427 /* 4. Try to fixup all. It is made immediately after connection enters
430 void tcp_init_buffer_space(struct sock
*sk
)
432 struct tcp_sock
*tp
= tcp_sk(sk
);
435 if (!(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
))
436 tcp_fixup_rcvbuf(sk
);
437 if (!(sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
))
438 tcp_sndbuf_expand(sk
);
440 tp
->rcvq_space
.space
= tp
->rcv_wnd
;
441 tp
->rcvq_space
.time
= tcp_time_stamp
;
442 tp
->rcvq_space
.seq
= tp
->copied_seq
;
444 maxwin
= tcp_full_space(sk
);
446 if (tp
->window_clamp
>= maxwin
) {
447 tp
->window_clamp
= maxwin
;
449 if (sysctl_tcp_app_win
&& maxwin
> 4 * tp
->advmss
)
450 tp
->window_clamp
= max(maxwin
-
451 (maxwin
>> sysctl_tcp_app_win
),
455 /* Force reservation of one segment. */
456 if (sysctl_tcp_app_win
&&
457 tp
->window_clamp
> 2 * tp
->advmss
&&
458 tp
->window_clamp
+ tp
->advmss
> maxwin
)
459 tp
->window_clamp
= max(2 * tp
->advmss
, maxwin
- tp
->advmss
);
461 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, tp
->window_clamp
);
462 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
465 /* 5. Recalculate window clamp after socket hit its memory bounds. */
466 static void tcp_clamp_window(struct sock
*sk
)
468 struct tcp_sock
*tp
= tcp_sk(sk
);
469 struct inet_connection_sock
*icsk
= inet_csk(sk
);
471 icsk
->icsk_ack
.quick
= 0;
473 if (sk
->sk_rcvbuf
< sysctl_tcp_rmem
[2] &&
474 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
) &&
475 !tcp_under_memory_pressure(sk
) &&
476 sk_memory_allocated(sk
) < sk_prot_mem_limits(sk
, 0)) {
477 sk
->sk_rcvbuf
= min(atomic_read(&sk
->sk_rmem_alloc
),
480 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
)
481 tp
->rcv_ssthresh
= min(tp
->window_clamp
, 2U * tp
->advmss
);
484 /* Initialize RCV_MSS value.
485 * RCV_MSS is an our guess about MSS used by the peer.
486 * We haven't any direct information about the MSS.
487 * It's better to underestimate the RCV_MSS rather than overestimate.
488 * Overestimations make us ACKing less frequently than needed.
489 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
491 void tcp_initialize_rcv_mss(struct sock
*sk
)
493 const struct tcp_sock
*tp
= tcp_sk(sk
);
494 unsigned int hint
= min_t(unsigned int, tp
->advmss
, tp
->mss_cache
);
496 hint
= min(hint
, tp
->rcv_wnd
/ 2);
497 hint
= min(hint
, TCP_MSS_DEFAULT
);
498 hint
= max(hint
, TCP_MIN_MSS
);
500 inet_csk(sk
)->icsk_ack
.rcv_mss
= hint
;
502 EXPORT_SYMBOL(tcp_initialize_rcv_mss
);
504 /* Receiver "autotuning" code.
506 * The algorithm for RTT estimation w/o timestamps is based on
507 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
508 * <http://public.lanl.gov/radiant/pubs.html#DRS>
510 * More detail on this code can be found at
511 * <http://staff.psc.edu/jheffner/>,
512 * though this reference is out of date. A new paper
515 static void tcp_rcv_rtt_update(struct tcp_sock
*tp
, u32 sample
, int win_dep
)
517 u32 new_sample
= tp
->rcv_rtt_est
.rtt
;
523 if (new_sample
!= 0) {
524 /* If we sample in larger samples in the non-timestamp
525 * case, we could grossly overestimate the RTT especially
526 * with chatty applications or bulk transfer apps which
527 * are stalled on filesystem I/O.
529 * Also, since we are only going for a minimum in the
530 * non-timestamp case, we do not smooth things out
531 * else with timestamps disabled convergence takes too
535 m
-= (new_sample
>> 3);
543 /* No previous measure. */
547 if (tp
->rcv_rtt_est
.rtt
!= new_sample
)
548 tp
->rcv_rtt_est
.rtt
= new_sample
;
551 static inline void tcp_rcv_rtt_measure(struct tcp_sock
*tp
)
553 if (tp
->rcv_rtt_est
.time
== 0)
555 if (before(tp
->rcv_nxt
, tp
->rcv_rtt_est
.seq
))
557 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rcv_rtt_est
.time
, 1);
560 tp
->rcv_rtt_est
.seq
= tp
->rcv_nxt
+ tp
->rcv_wnd
;
561 tp
->rcv_rtt_est
.time
= tcp_time_stamp
;
564 static inline void tcp_rcv_rtt_measure_ts(struct sock
*sk
,
565 const struct sk_buff
*skb
)
567 struct tcp_sock
*tp
= tcp_sk(sk
);
568 if (tp
->rx_opt
.rcv_tsecr
&&
569 (TCP_SKB_CB(skb
)->end_seq
-
570 TCP_SKB_CB(skb
)->seq
>= inet_csk(sk
)->icsk_ack
.rcv_mss
))
571 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
, 0);
575 * This function should be called every time data is copied to user space.
576 * It calculates the appropriate TCP receive buffer space.
578 void tcp_rcv_space_adjust(struct sock
*sk
)
580 struct tcp_sock
*tp
= tcp_sk(sk
);
584 time
= tcp_time_stamp
- tp
->rcvq_space
.time
;
585 if (time
< (tp
->rcv_rtt_est
.rtt
>> 3) || tp
->rcv_rtt_est
.rtt
== 0)
588 /* Number of bytes copied to user in last RTT */
589 copied
= tp
->copied_seq
- tp
->rcvq_space
.seq
;
590 if (copied
<= tp
->rcvq_space
.space
)
594 * copied = bytes received in previous RTT, our base window
595 * To cope with packet losses, we need a 2x factor
596 * To cope with slow start, and sender growing its cwin by 100 %
597 * every RTT, we need a 4x factor, because the ACK we are sending
598 * now is for the next RTT, not the current one :
599 * <prev RTT . ><current RTT .. ><next RTT .... >
602 if (sysctl_tcp_moderate_rcvbuf
&&
603 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
)) {
604 int rcvwin
, rcvmem
, rcvbuf
;
606 /* minimal window to cope with packet losses, assuming
607 * steady state. Add some cushion because of small variations.
609 rcvwin
= (copied
<< 1) + 16 * tp
->advmss
;
611 /* If rate increased by 25%,
612 * assume slow start, rcvwin = 3 * copied
613 * If rate increased by 50%,
614 * assume sender can use 2x growth, rcvwin = 4 * copied
617 tp
->rcvq_space
.space
+ (tp
->rcvq_space
.space
>> 2)) {
619 tp
->rcvq_space
.space
+ (tp
->rcvq_space
.space
>> 1))
622 rcvwin
+= (rcvwin
>> 1);
625 rcvmem
= SKB_TRUESIZE(tp
->advmss
+ MAX_TCP_HEADER
);
626 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
629 rcvbuf
= min(rcvwin
/ tp
->advmss
* rcvmem
, sysctl_tcp_rmem
[2]);
630 if (rcvbuf
> sk
->sk_rcvbuf
) {
631 sk
->sk_rcvbuf
= rcvbuf
;
633 /* Make the window clamp follow along. */
634 tp
->window_clamp
= rcvwin
;
637 tp
->rcvq_space
.space
= copied
;
640 tp
->rcvq_space
.seq
= tp
->copied_seq
;
641 tp
->rcvq_space
.time
= tcp_time_stamp
;
644 /* There is something which you must keep in mind when you analyze the
645 * behavior of the tp->ato delayed ack timeout interval. When a
646 * connection starts up, we want to ack as quickly as possible. The
647 * problem is that "good" TCP's do slow start at the beginning of data
648 * transmission. The means that until we send the first few ACK's the
649 * sender will sit on his end and only queue most of his data, because
650 * he can only send snd_cwnd unacked packets at any given time. For
651 * each ACK we send, he increments snd_cwnd and transmits more of his
654 static void tcp_event_data_recv(struct sock
*sk
, struct sk_buff
*skb
)
656 struct tcp_sock
*tp
= tcp_sk(sk
);
657 struct inet_connection_sock
*icsk
= inet_csk(sk
);
660 inet_csk_schedule_ack(sk
);
662 tcp_measure_rcv_mss(sk
, skb
);
664 tcp_rcv_rtt_measure(tp
);
666 now
= tcp_time_stamp
;
668 if (!icsk
->icsk_ack
.ato
) {
669 /* The _first_ data packet received, initialize
670 * delayed ACK engine.
672 tcp_incr_quickack(sk
);
673 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
675 int m
= now
- icsk
->icsk_ack
.lrcvtime
;
677 if (m
<= TCP_ATO_MIN
/ 2) {
678 /* The fastest case is the first. */
679 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + TCP_ATO_MIN
/ 2;
680 } else if (m
< icsk
->icsk_ack
.ato
) {
681 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + m
;
682 if (icsk
->icsk_ack
.ato
> icsk
->icsk_rto
)
683 icsk
->icsk_ack
.ato
= icsk
->icsk_rto
;
684 } else if (m
> icsk
->icsk_rto
) {
685 /* Too long gap. Apparently sender failed to
686 * restart window, so that we send ACKs quickly.
688 tcp_incr_quickack(sk
);
692 icsk
->icsk_ack
.lrcvtime
= now
;
694 tcp_ecn_check_ce(tp
, skb
);
697 tcp_grow_window(sk
, skb
);
700 /* Called to compute a smoothed rtt estimate. The data fed to this
701 * routine either comes from timestamps, or from segments that were
702 * known _not_ to have been retransmitted [see Karn/Partridge
703 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
704 * piece by Van Jacobson.
705 * NOTE: the next three routines used to be one big routine.
706 * To save cycles in the RFC 1323 implementation it was better to break
707 * it up into three procedures. -- erics
709 static void tcp_rtt_estimator(struct sock
*sk
, long mrtt_us
)
711 struct tcp_sock
*tp
= tcp_sk(sk
);
712 long m
= mrtt_us
; /* RTT */
713 u32 srtt
= tp
->srtt_us
;
715 /* The following amusing code comes from Jacobson's
716 * article in SIGCOMM '88. Note that rtt and mdev
717 * are scaled versions of rtt and mean deviation.
718 * This is designed to be as fast as possible
719 * m stands for "measurement".
721 * On a 1990 paper the rto value is changed to:
722 * RTO = rtt + 4 * mdev
724 * Funny. This algorithm seems to be very broken.
725 * These formulae increase RTO, when it should be decreased, increase
726 * too slowly, when it should be increased quickly, decrease too quickly
727 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
728 * does not matter how to _calculate_ it. Seems, it was trap
729 * that VJ failed to avoid. 8)
732 m
-= (srtt
>> 3); /* m is now error in rtt est */
733 srtt
+= m
; /* rtt = 7/8 rtt + 1/8 new */
735 m
= -m
; /* m is now abs(error) */
736 m
-= (tp
->mdev_us
>> 2); /* similar update on mdev */
737 /* This is similar to one of Eifel findings.
738 * Eifel blocks mdev updates when rtt decreases.
739 * This solution is a bit different: we use finer gain
740 * for mdev in this case (alpha*beta).
741 * Like Eifel it also prevents growth of rto,
742 * but also it limits too fast rto decreases,
743 * happening in pure Eifel.
748 m
-= (tp
->mdev_us
>> 2); /* similar update on mdev */
750 tp
->mdev_us
+= m
; /* mdev = 3/4 mdev + 1/4 new */
751 if (tp
->mdev_us
> tp
->mdev_max_us
) {
752 tp
->mdev_max_us
= tp
->mdev_us
;
753 if (tp
->mdev_max_us
> tp
->rttvar_us
)
754 tp
->rttvar_us
= tp
->mdev_max_us
;
756 if (after(tp
->snd_una
, tp
->rtt_seq
)) {
757 if (tp
->mdev_max_us
< tp
->rttvar_us
)
758 tp
->rttvar_us
-= (tp
->rttvar_us
- tp
->mdev_max_us
) >> 2;
759 tp
->rtt_seq
= tp
->snd_nxt
;
760 tp
->mdev_max_us
= tcp_rto_min_us(sk
);
763 /* no previous measure. */
764 srtt
= m
<< 3; /* take the measured time to be rtt */
765 tp
->mdev_us
= m
<< 1; /* make sure rto = 3*rtt */
766 tp
->rttvar_us
= max(tp
->mdev_us
, tcp_rto_min_us(sk
));
767 tp
->mdev_max_us
= tp
->rttvar_us
;
768 tp
->rtt_seq
= tp
->snd_nxt
;
770 tp
->srtt_us
= max(1U, srtt
);
773 /* Set the sk_pacing_rate to allow proper sizing of TSO packets.
774 * Note: TCP stack does not yet implement pacing.
775 * FQ packet scheduler can be used to implement cheap but effective
776 * TCP pacing, to smooth the burst on large writes when packets
777 * in flight is significantly lower than cwnd (or rwin)
779 int sysctl_tcp_pacing_ss_ratio __read_mostly
= 200;
780 int sysctl_tcp_pacing_ca_ratio __read_mostly
= 120;
782 static void tcp_update_pacing_rate(struct sock
*sk
)
784 const struct tcp_sock
*tp
= tcp_sk(sk
);
787 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
788 rate
= (u64
)tp
->mss_cache
* ((USEC_PER_SEC
/ 100) << 3);
790 /* current rate is (cwnd * mss) / srtt
791 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
792 * In Congestion Avoidance phase, set it to 120 % the current rate.
794 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
795 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
796 * end of slow start and should slow down.
798 if (tp
->snd_cwnd
< tp
->snd_ssthresh
/ 2)
799 rate
*= sysctl_tcp_pacing_ss_ratio
;
801 rate
*= sysctl_tcp_pacing_ca_ratio
;
803 rate
*= max(tp
->snd_cwnd
, tp
->packets_out
);
805 if (likely(tp
->srtt_us
))
806 do_div(rate
, tp
->srtt_us
);
808 /* ACCESS_ONCE() is needed because sch_fq fetches sk_pacing_rate
809 * without any lock. We want to make sure compiler wont store
810 * intermediate values in this location.
812 ACCESS_ONCE(sk
->sk_pacing_rate
) = min_t(u64
, rate
,
813 sk
->sk_max_pacing_rate
);
816 /* Calculate rto without backoff. This is the second half of Van Jacobson's
817 * routine referred to above.
819 static void tcp_set_rto(struct sock
*sk
)
821 const struct tcp_sock
*tp
= tcp_sk(sk
);
822 /* Old crap is replaced with new one. 8)
825 * 1. If rtt variance happened to be less 50msec, it is hallucination.
826 * It cannot be less due to utterly erratic ACK generation made
827 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
828 * to do with delayed acks, because at cwnd>2 true delack timeout
829 * is invisible. Actually, Linux-2.4 also generates erratic
830 * ACKs in some circumstances.
832 inet_csk(sk
)->icsk_rto
= __tcp_set_rto(tp
);
834 /* 2. Fixups made earlier cannot be right.
835 * If we do not estimate RTO correctly without them,
836 * all the algo is pure shit and should be replaced
837 * with correct one. It is exactly, which we pretend to do.
840 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
841 * guarantees that rto is higher.
846 __u32
tcp_init_cwnd(const struct tcp_sock
*tp
, const struct dst_entry
*dst
)
848 __u32 cwnd
= (dst
? dst_metric(dst
, RTAX_INITCWND
) : 0);
851 cwnd
= TCP_INIT_CWND
;
852 return min_t(__u32
, cwnd
, tp
->snd_cwnd_clamp
);
856 * Packet counting of FACK is based on in-order assumptions, therefore TCP
857 * disables it when reordering is detected
859 void tcp_disable_fack(struct tcp_sock
*tp
)
861 /* RFC3517 uses different metric in lost marker => reset on change */
863 tp
->lost_skb_hint
= NULL
;
864 tp
->rx_opt
.sack_ok
&= ~TCP_FACK_ENABLED
;
867 /* Take a notice that peer is sending D-SACKs */
868 static void tcp_dsack_seen(struct tcp_sock
*tp
)
870 tp
->rx_opt
.sack_ok
|= TCP_DSACK_SEEN
;
873 static void tcp_update_reordering(struct sock
*sk
, const int metric
,
876 struct tcp_sock
*tp
= tcp_sk(sk
);
877 if (metric
> tp
->reordering
) {
880 tp
->reordering
= min(sysctl_tcp_max_reordering
, metric
);
882 /* This exciting event is worth to be remembered. 8) */
884 mib_idx
= LINUX_MIB_TCPTSREORDER
;
885 else if (tcp_is_reno(tp
))
886 mib_idx
= LINUX_MIB_TCPRENOREORDER
;
887 else if (tcp_is_fack(tp
))
888 mib_idx
= LINUX_MIB_TCPFACKREORDER
;
890 mib_idx
= LINUX_MIB_TCPSACKREORDER
;
892 NET_INC_STATS(sock_net(sk
), mib_idx
);
893 #if FASTRETRANS_DEBUG > 1
894 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
895 tp
->rx_opt
.sack_ok
, inet_csk(sk
)->icsk_ca_state
,
899 tp
->undo_marker
? tp
->undo_retrans
: 0);
901 tcp_disable_fack(tp
);
907 /* This must be called before lost_out is incremented */
908 static void tcp_verify_retransmit_hint(struct tcp_sock
*tp
, struct sk_buff
*skb
)
910 if (!tp
->retransmit_skb_hint
||
911 before(TCP_SKB_CB(skb
)->seq
,
912 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
913 tp
->retransmit_skb_hint
= skb
;
916 /* Sum the number of packets on the wire we have marked as lost.
917 * There are two cases we care about here:
918 * a) Packet hasn't been marked lost (nor retransmitted),
919 * and this is the first loss.
920 * b) Packet has been marked both lost and retransmitted,
921 * and this means we think it was lost again.
923 static void tcp_sum_lost(struct tcp_sock
*tp
, struct sk_buff
*skb
)
925 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
927 if (!(sacked
& TCPCB_LOST
) ||
928 ((sacked
& TCPCB_LOST
) && (sacked
& TCPCB_SACKED_RETRANS
)))
929 tp
->lost
+= tcp_skb_pcount(skb
);
932 static void tcp_skb_mark_lost(struct tcp_sock
*tp
, struct sk_buff
*skb
)
934 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
935 tcp_verify_retransmit_hint(tp
, skb
);
937 tp
->lost_out
+= tcp_skb_pcount(skb
);
938 tcp_sum_lost(tp
, skb
);
939 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
943 void tcp_skb_mark_lost_uncond_verify(struct tcp_sock
*tp
, struct sk_buff
*skb
)
945 tcp_verify_retransmit_hint(tp
, skb
);
947 tcp_sum_lost(tp
, skb
);
948 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
949 tp
->lost_out
+= tcp_skb_pcount(skb
);
950 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
954 /* This procedure tags the retransmission queue when SACKs arrive.
956 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
957 * Packets in queue with these bits set are counted in variables
958 * sacked_out, retrans_out and lost_out, correspondingly.
960 * Valid combinations are:
961 * Tag InFlight Description
962 * 0 1 - orig segment is in flight.
963 * S 0 - nothing flies, orig reached receiver.
964 * L 0 - nothing flies, orig lost by net.
965 * R 2 - both orig and retransmit are in flight.
966 * L|R 1 - orig is lost, retransmit is in flight.
967 * S|R 1 - orig reached receiver, retrans is still in flight.
968 * (L|S|R is logically valid, it could occur when L|R is sacked,
969 * but it is equivalent to plain S and code short-curcuits it to S.
970 * L|S is logically invalid, it would mean -1 packet in flight 8))
972 * These 6 states form finite state machine, controlled by the following events:
973 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
974 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
975 * 3. Loss detection event of two flavors:
976 * A. Scoreboard estimator decided the packet is lost.
977 * A'. Reno "three dupacks" marks head of queue lost.
978 * A''. Its FACK modification, head until snd.fack is lost.
979 * B. SACK arrives sacking SND.NXT at the moment, when the
980 * segment was retransmitted.
981 * 4. D-SACK added new rule: D-SACK changes any tag to S.
983 * It is pleasant to note, that state diagram turns out to be commutative,
984 * so that we are allowed not to be bothered by order of our actions,
985 * when multiple events arrive simultaneously. (see the function below).
987 * Reordering detection.
988 * --------------------
989 * Reordering metric is maximal distance, which a packet can be displaced
990 * in packet stream. With SACKs we can estimate it:
992 * 1. SACK fills old hole and the corresponding segment was not
993 * ever retransmitted -> reordering. Alas, we cannot use it
994 * when segment was retransmitted.
995 * 2. The last flaw is solved with D-SACK. D-SACK arrives
996 * for retransmitted and already SACKed segment -> reordering..
997 * Both of these heuristics are not used in Loss state, when we cannot
998 * account for retransmits accurately.
1000 * SACK block validation.
1001 * ----------------------
1003 * SACK block range validation checks that the received SACK block fits to
1004 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1005 * Note that SND.UNA is not included to the range though being valid because
1006 * it means that the receiver is rather inconsistent with itself reporting
1007 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1008 * perfectly valid, however, in light of RFC2018 which explicitly states
1009 * that "SACK block MUST reflect the newest segment. Even if the newest
1010 * segment is going to be discarded ...", not that it looks very clever
1011 * in case of head skb. Due to potentional receiver driven attacks, we
1012 * choose to avoid immediate execution of a walk in write queue due to
1013 * reneging and defer head skb's loss recovery to standard loss recovery
1014 * procedure that will eventually trigger (nothing forbids us doing this).
1016 * Implements also blockage to start_seq wrap-around. Problem lies in the
1017 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1018 * there's no guarantee that it will be before snd_nxt (n). The problem
1019 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1022 * <- outs wnd -> <- wrapzone ->
1023 * u e n u_w e_w s n_w
1025 * |<------------+------+----- TCP seqno space --------------+---------->|
1026 * ...-- <2^31 ->| |<--------...
1027 * ...---- >2^31 ------>| |<--------...
1029 * Current code wouldn't be vulnerable but it's better still to discard such
1030 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1031 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1032 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1033 * equal to the ideal case (infinite seqno space without wrap caused issues).
1035 * With D-SACK the lower bound is extended to cover sequence space below
1036 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1037 * again, D-SACK block must not to go across snd_una (for the same reason as
1038 * for the normal SACK blocks, explained above). But there all simplicity
1039 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1040 * fully below undo_marker they do not affect behavior in anyway and can
1041 * therefore be safely ignored. In rare cases (which are more or less
1042 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1043 * fragmentation and packet reordering past skb's retransmission. To consider
1044 * them correctly, the acceptable range must be extended even more though
1045 * the exact amount is rather hard to quantify. However, tp->max_window can
1046 * be used as an exaggerated estimate.
1048 static bool tcp_is_sackblock_valid(struct tcp_sock
*tp
, bool is_dsack
,
1049 u32 start_seq
, u32 end_seq
)
1051 /* Too far in future, or reversed (interpretation is ambiguous) */
1052 if (after(end_seq
, tp
->snd_nxt
) || !before(start_seq
, end_seq
))
1055 /* Nasty start_seq wrap-around check (see comments above) */
1056 if (!before(start_seq
, tp
->snd_nxt
))
1059 /* In outstanding window? ...This is valid exit for D-SACKs too.
1060 * start_seq == snd_una is non-sensical (see comments above)
1062 if (after(start_seq
, tp
->snd_una
))
1065 if (!is_dsack
|| !tp
->undo_marker
)
1068 /* ...Then it's D-SACK, and must reside below snd_una completely */
1069 if (after(end_seq
, tp
->snd_una
))
1072 if (!before(start_seq
, tp
->undo_marker
))
1076 if (!after(end_seq
, tp
->undo_marker
))
1079 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1080 * start_seq < undo_marker and end_seq >= undo_marker.
1082 return !before(start_seq
, end_seq
- tp
->max_window
);
1085 static bool tcp_check_dsack(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1086 struct tcp_sack_block_wire
*sp
, int num_sacks
,
1089 struct tcp_sock
*tp
= tcp_sk(sk
);
1090 u32 start_seq_0
= get_unaligned_be32(&sp
[0].start_seq
);
1091 u32 end_seq_0
= get_unaligned_be32(&sp
[0].end_seq
);
1092 bool dup_sack
= false;
1094 if (before(start_seq_0
, TCP_SKB_CB(ack_skb
)->ack_seq
)) {
1097 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPDSACKRECV
);
1098 } else if (num_sacks
> 1) {
1099 u32 end_seq_1
= get_unaligned_be32(&sp
[1].end_seq
);
1100 u32 start_seq_1
= get_unaligned_be32(&sp
[1].start_seq
);
1102 if (!after(end_seq_0
, end_seq_1
) &&
1103 !before(start_seq_0
, start_seq_1
)) {
1106 NET_INC_STATS(sock_net(sk
),
1107 LINUX_MIB_TCPDSACKOFORECV
);
1111 /* D-SACK for already forgotten data... Do dumb counting. */
1112 if (dup_sack
&& tp
->undo_marker
&& tp
->undo_retrans
> 0 &&
1113 !after(end_seq_0
, prior_snd_una
) &&
1114 after(end_seq_0
, tp
->undo_marker
))
1120 struct tcp_sacktag_state
{
1123 /* Timestamps for earliest and latest never-retransmitted segment
1124 * that was SACKed. RTO needs the earliest RTT to stay conservative,
1125 * but congestion control should still get an accurate delay signal.
1127 struct skb_mstamp first_sackt
;
1128 struct skb_mstamp last_sackt
;
1129 struct skb_mstamp ack_time
; /* Timestamp when the S/ACK was received */
1130 struct rate_sample
*rate
;
1134 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1135 * the incoming SACK may not exactly match but we can find smaller MSS
1136 * aligned portion of it that matches. Therefore we might need to fragment
1137 * which may fail and creates some hassle (caller must handle error case
1140 * FIXME: this could be merged to shift decision code
1142 static int tcp_match_skb_to_sack(struct sock
*sk
, struct sk_buff
*skb
,
1143 u32 start_seq
, u32 end_seq
)
1147 unsigned int pkt_len
;
1150 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1151 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1153 if (tcp_skb_pcount(skb
) > 1 && !in_sack
&&
1154 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
1155 mss
= tcp_skb_mss(skb
);
1156 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1159 pkt_len
= start_seq
- TCP_SKB_CB(skb
)->seq
;
1163 pkt_len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1168 /* Round if necessary so that SACKs cover only full MSSes
1169 * and/or the remaining small portion (if present)
1171 if (pkt_len
> mss
) {
1172 unsigned int new_len
= (pkt_len
/ mss
) * mss
;
1173 if (!in_sack
&& new_len
< pkt_len
) {
1175 if (new_len
>= skb
->len
)
1180 err
= tcp_fragment(sk
, skb
, pkt_len
, mss
, GFP_ATOMIC
);
1188 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1189 static u8
tcp_sacktag_one(struct sock
*sk
,
1190 struct tcp_sacktag_state
*state
, u8 sacked
,
1191 u32 start_seq
, u32 end_seq
,
1192 int dup_sack
, int pcount
,
1193 const struct skb_mstamp
*xmit_time
)
1195 struct tcp_sock
*tp
= tcp_sk(sk
);
1196 int fack_count
= state
->fack_count
;
1198 /* Account D-SACK for retransmitted packet. */
1199 if (dup_sack
&& (sacked
& TCPCB_RETRANS
)) {
1200 if (tp
->undo_marker
&& tp
->undo_retrans
> 0 &&
1201 after(end_seq
, tp
->undo_marker
))
1203 if (sacked
& TCPCB_SACKED_ACKED
)
1204 state
->reord
= min(fack_count
, state
->reord
);
1207 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1208 if (!after(end_seq
, tp
->snd_una
))
1211 if (!(sacked
& TCPCB_SACKED_ACKED
)) {
1212 tcp_rack_advance(tp
, sacked
, end_seq
,
1213 xmit_time
, &state
->ack_time
);
1215 if (sacked
& TCPCB_SACKED_RETRANS
) {
1216 /* If the segment is not tagged as lost,
1217 * we do not clear RETRANS, believing
1218 * that retransmission is still in flight.
1220 if (sacked
& TCPCB_LOST
) {
1221 sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1222 tp
->lost_out
-= pcount
;
1223 tp
->retrans_out
-= pcount
;
1226 if (!(sacked
& TCPCB_RETRANS
)) {
1227 /* New sack for not retransmitted frame,
1228 * which was in hole. It is reordering.
1230 if (before(start_seq
,
1231 tcp_highest_sack_seq(tp
)))
1232 state
->reord
= min(fack_count
,
1234 if (!after(end_seq
, tp
->high_seq
))
1235 state
->flag
|= FLAG_ORIG_SACK_ACKED
;
1236 if (state
->first_sackt
.v64
== 0)
1237 state
->first_sackt
= *xmit_time
;
1238 state
->last_sackt
= *xmit_time
;
1241 if (sacked
& TCPCB_LOST
) {
1242 sacked
&= ~TCPCB_LOST
;
1243 tp
->lost_out
-= pcount
;
1247 sacked
|= TCPCB_SACKED_ACKED
;
1248 state
->flag
|= FLAG_DATA_SACKED
;
1249 tp
->sacked_out
+= pcount
;
1250 tp
->delivered
+= pcount
; /* Out-of-order packets delivered */
1252 fack_count
+= pcount
;
1254 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1255 if (!tcp_is_fack(tp
) && tp
->lost_skb_hint
&&
1256 before(start_seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
))
1257 tp
->lost_cnt_hint
+= pcount
;
1259 if (fack_count
> tp
->fackets_out
)
1260 tp
->fackets_out
= fack_count
;
1263 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1264 * frames and clear it. undo_retrans is decreased above, L|R frames
1265 * are accounted above as well.
1267 if (dup_sack
&& (sacked
& TCPCB_SACKED_RETRANS
)) {
1268 sacked
&= ~TCPCB_SACKED_RETRANS
;
1269 tp
->retrans_out
-= pcount
;
1275 /* Shift newly-SACKed bytes from this skb to the immediately previous
1276 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1278 static bool tcp_shifted_skb(struct sock
*sk
, struct sk_buff
*skb
,
1279 struct tcp_sacktag_state
*state
,
1280 unsigned int pcount
, int shifted
, int mss
,
1283 struct tcp_sock
*tp
= tcp_sk(sk
);
1284 struct sk_buff
*prev
= tcp_write_queue_prev(sk
, skb
);
1285 u32 start_seq
= TCP_SKB_CB(skb
)->seq
; /* start of newly-SACKed */
1286 u32 end_seq
= start_seq
+ shifted
; /* end of newly-SACKed */
1290 /* Adjust counters and hints for the newly sacked sequence
1291 * range but discard the return value since prev is already
1292 * marked. We must tag the range first because the seq
1293 * advancement below implicitly advances
1294 * tcp_highest_sack_seq() when skb is highest_sack.
1296 tcp_sacktag_one(sk
, state
, TCP_SKB_CB(skb
)->sacked
,
1297 start_seq
, end_seq
, dup_sack
, pcount
,
1299 tcp_rate_skb_delivered(sk
, skb
, state
->rate
);
1301 if (skb
== tp
->lost_skb_hint
)
1302 tp
->lost_cnt_hint
+= pcount
;
1304 TCP_SKB_CB(prev
)->end_seq
+= shifted
;
1305 TCP_SKB_CB(skb
)->seq
+= shifted
;
1307 tcp_skb_pcount_add(prev
, pcount
);
1308 BUG_ON(tcp_skb_pcount(skb
) < pcount
);
1309 tcp_skb_pcount_add(skb
, -pcount
);
1311 /* When we're adding to gso_segs == 1, gso_size will be zero,
1312 * in theory this shouldn't be necessary but as long as DSACK
1313 * code can come after this skb later on it's better to keep
1314 * setting gso_size to something.
1316 if (!TCP_SKB_CB(prev
)->tcp_gso_size
)
1317 TCP_SKB_CB(prev
)->tcp_gso_size
= mss
;
1319 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1320 if (tcp_skb_pcount(skb
) <= 1)
1321 TCP_SKB_CB(skb
)->tcp_gso_size
= 0;
1323 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1324 TCP_SKB_CB(prev
)->sacked
|= (TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
);
1327 BUG_ON(!tcp_skb_pcount(skb
));
1328 NET_INC_STATS(sock_net(sk
), LINUX_MIB_SACKSHIFTED
);
1332 /* Whole SKB was eaten :-) */
1334 if (skb
== tp
->retransmit_skb_hint
)
1335 tp
->retransmit_skb_hint
= prev
;
1336 if (skb
== tp
->lost_skb_hint
) {
1337 tp
->lost_skb_hint
= prev
;
1338 tp
->lost_cnt_hint
-= tcp_skb_pcount(prev
);
1341 TCP_SKB_CB(prev
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
;
1342 TCP_SKB_CB(prev
)->eor
= TCP_SKB_CB(skb
)->eor
;
1343 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
1344 TCP_SKB_CB(prev
)->end_seq
++;
1346 if (skb
== tcp_highest_sack(sk
))
1347 tcp_advance_highest_sack(sk
, skb
);
1349 tcp_skb_collapse_tstamp(prev
, skb
);
1350 if (unlikely(TCP_SKB_CB(prev
)->tx
.delivered_mstamp
.v64
))
1351 TCP_SKB_CB(prev
)->tx
.delivered_mstamp
.v64
= 0;
1353 tcp_unlink_write_queue(skb
, sk
);
1354 sk_wmem_free_skb(sk
, skb
);
1356 NET_INC_STATS(sock_net(sk
), LINUX_MIB_SACKMERGED
);
1361 /* I wish gso_size would have a bit more sane initialization than
1362 * something-or-zero which complicates things
1364 static int tcp_skb_seglen(const struct sk_buff
*skb
)
1366 return tcp_skb_pcount(skb
) == 1 ? skb
->len
: tcp_skb_mss(skb
);
1369 /* Shifting pages past head area doesn't work */
1370 static int skb_can_shift(const struct sk_buff
*skb
)
1372 return !skb_headlen(skb
) && skb_is_nonlinear(skb
);
1375 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1378 static struct sk_buff
*tcp_shift_skb_data(struct sock
*sk
, struct sk_buff
*skb
,
1379 struct tcp_sacktag_state
*state
,
1380 u32 start_seq
, u32 end_seq
,
1383 struct tcp_sock
*tp
= tcp_sk(sk
);
1384 struct sk_buff
*prev
;
1390 if (!sk_can_gso(sk
))
1393 /* Normally R but no L won't result in plain S */
1395 (TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_RETRANS
)) == TCPCB_SACKED_RETRANS
)
1397 if (!skb_can_shift(skb
))
1399 /* This frame is about to be dropped (was ACKed). */
1400 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1403 /* Can only happen with delayed DSACK + discard craziness */
1404 if (unlikely(skb
== tcp_write_queue_head(sk
)))
1406 prev
= tcp_write_queue_prev(sk
, skb
);
1408 if ((TCP_SKB_CB(prev
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
)
1411 if (!tcp_skb_can_collapse_to(prev
))
1414 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1415 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1419 pcount
= tcp_skb_pcount(skb
);
1420 mss
= tcp_skb_seglen(skb
);
1422 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1423 * drop this restriction as unnecessary
1425 if (mss
!= tcp_skb_seglen(prev
))
1428 if (!after(TCP_SKB_CB(skb
)->end_seq
, start_seq
))
1430 /* CHECKME: This is non-MSS split case only?, this will
1431 * cause skipped skbs due to advancing loop btw, original
1432 * has that feature too
1434 if (tcp_skb_pcount(skb
) <= 1)
1437 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1439 /* TODO: head merge to next could be attempted here
1440 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1441 * though it might not be worth of the additional hassle
1443 * ...we can probably just fallback to what was done
1444 * previously. We could try merging non-SACKed ones
1445 * as well but it probably isn't going to buy off
1446 * because later SACKs might again split them, and
1447 * it would make skb timestamp tracking considerably
1453 len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1455 BUG_ON(len
> skb
->len
);
1457 /* MSS boundaries should be honoured or else pcount will
1458 * severely break even though it makes things bit trickier.
1459 * Optimize common case to avoid most of the divides
1461 mss
= tcp_skb_mss(skb
);
1463 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1464 * drop this restriction as unnecessary
1466 if (mss
!= tcp_skb_seglen(prev
))
1471 } else if (len
< mss
) {
1479 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1480 if (!after(TCP_SKB_CB(skb
)->seq
+ len
, tp
->snd_una
))
1483 if (!skb_shift(prev
, skb
, len
))
1485 if (!tcp_shifted_skb(sk
, skb
, state
, pcount
, len
, mss
, dup_sack
))
1488 /* Hole filled allows collapsing with the next as well, this is very
1489 * useful when hole on every nth skb pattern happens
1491 if (prev
== tcp_write_queue_tail(sk
))
1493 skb
= tcp_write_queue_next(sk
, prev
);
1495 if (!skb_can_shift(skb
) ||
1496 (skb
== tcp_send_head(sk
)) ||
1497 ((TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
) ||
1498 (mss
!= tcp_skb_seglen(skb
)))
1502 if (skb_shift(prev
, skb
, len
)) {
1503 pcount
+= tcp_skb_pcount(skb
);
1504 tcp_shifted_skb(sk
, skb
, state
, tcp_skb_pcount(skb
), len
, mss
, 0);
1508 state
->fack_count
+= pcount
;
1515 NET_INC_STATS(sock_net(sk
), LINUX_MIB_SACKSHIFTFALLBACK
);
1519 static struct sk_buff
*tcp_sacktag_walk(struct sk_buff
*skb
, struct sock
*sk
,
1520 struct tcp_sack_block
*next_dup
,
1521 struct tcp_sacktag_state
*state
,
1522 u32 start_seq
, u32 end_seq
,
1525 struct tcp_sock
*tp
= tcp_sk(sk
);
1526 struct sk_buff
*tmp
;
1528 tcp_for_write_queue_from(skb
, sk
) {
1530 bool dup_sack
= dup_sack_in
;
1532 if (skb
== tcp_send_head(sk
))
1535 /* queue is in-order => we can short-circuit the walk early */
1536 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1540 before(TCP_SKB_CB(skb
)->seq
, next_dup
->end_seq
)) {
1541 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1542 next_dup
->start_seq
,
1548 /* skb reference here is a bit tricky to get right, since
1549 * shifting can eat and free both this skb and the next,
1550 * so not even _safe variant of the loop is enough.
1553 tmp
= tcp_shift_skb_data(sk
, skb
, state
,
1554 start_seq
, end_seq
, dup_sack
);
1563 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1569 if (unlikely(in_sack
< 0))
1573 TCP_SKB_CB(skb
)->sacked
=
1576 TCP_SKB_CB(skb
)->sacked
,
1577 TCP_SKB_CB(skb
)->seq
,
1578 TCP_SKB_CB(skb
)->end_seq
,
1580 tcp_skb_pcount(skb
),
1582 tcp_rate_skb_delivered(sk
, skb
, state
->rate
);
1584 if (!before(TCP_SKB_CB(skb
)->seq
,
1585 tcp_highest_sack_seq(tp
)))
1586 tcp_advance_highest_sack(sk
, skb
);
1589 state
->fack_count
+= tcp_skb_pcount(skb
);
1594 /* Avoid all extra work that is being done by sacktag while walking in
1597 static struct sk_buff
*tcp_sacktag_skip(struct sk_buff
*skb
, struct sock
*sk
,
1598 struct tcp_sacktag_state
*state
,
1601 tcp_for_write_queue_from(skb
, sk
) {
1602 if (skb
== tcp_send_head(sk
))
1605 if (after(TCP_SKB_CB(skb
)->end_seq
, skip_to_seq
))
1608 state
->fack_count
+= tcp_skb_pcount(skb
);
1613 static struct sk_buff
*tcp_maybe_skipping_dsack(struct sk_buff
*skb
,
1615 struct tcp_sack_block
*next_dup
,
1616 struct tcp_sacktag_state
*state
,
1622 if (before(next_dup
->start_seq
, skip_to_seq
)) {
1623 skb
= tcp_sacktag_skip(skb
, sk
, state
, next_dup
->start_seq
);
1624 skb
= tcp_sacktag_walk(skb
, sk
, NULL
, state
,
1625 next_dup
->start_seq
, next_dup
->end_seq
,
1632 static int tcp_sack_cache_ok(const struct tcp_sock
*tp
, const struct tcp_sack_block
*cache
)
1634 return cache
< tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1638 tcp_sacktag_write_queue(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1639 u32 prior_snd_una
, struct tcp_sacktag_state
*state
)
1641 struct tcp_sock
*tp
= tcp_sk(sk
);
1642 const unsigned char *ptr
= (skb_transport_header(ack_skb
) +
1643 TCP_SKB_CB(ack_skb
)->sacked
);
1644 struct tcp_sack_block_wire
*sp_wire
= (struct tcp_sack_block_wire
*)(ptr
+2);
1645 struct tcp_sack_block sp
[TCP_NUM_SACKS
];
1646 struct tcp_sack_block
*cache
;
1647 struct sk_buff
*skb
;
1648 int num_sacks
= min(TCP_NUM_SACKS
, (ptr
[1] - TCPOLEN_SACK_BASE
) >> 3);
1650 bool found_dup_sack
= false;
1652 int first_sack_index
;
1655 state
->reord
= tp
->packets_out
;
1657 if (!tp
->sacked_out
) {
1658 if (WARN_ON(tp
->fackets_out
))
1659 tp
->fackets_out
= 0;
1660 tcp_highest_sack_reset(sk
);
1663 found_dup_sack
= tcp_check_dsack(sk
, ack_skb
, sp_wire
,
1664 num_sacks
, prior_snd_una
);
1665 if (found_dup_sack
) {
1666 state
->flag
|= FLAG_DSACKING_ACK
;
1667 tp
->delivered
++; /* A spurious retransmission is delivered */
1670 /* Eliminate too old ACKs, but take into
1671 * account more or less fresh ones, they can
1672 * contain valid SACK info.
1674 if (before(TCP_SKB_CB(ack_skb
)->ack_seq
, prior_snd_una
- tp
->max_window
))
1677 if (!tp
->packets_out
)
1681 first_sack_index
= 0;
1682 for (i
= 0; i
< num_sacks
; i
++) {
1683 bool dup_sack
= !i
&& found_dup_sack
;
1685 sp
[used_sacks
].start_seq
= get_unaligned_be32(&sp_wire
[i
].start_seq
);
1686 sp
[used_sacks
].end_seq
= get_unaligned_be32(&sp_wire
[i
].end_seq
);
1688 if (!tcp_is_sackblock_valid(tp
, dup_sack
,
1689 sp
[used_sacks
].start_seq
,
1690 sp
[used_sacks
].end_seq
)) {
1694 if (!tp
->undo_marker
)
1695 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDNOUNDO
;
1697 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDOLD
;
1699 /* Don't count olds caused by ACK reordering */
1700 if ((TCP_SKB_CB(ack_skb
)->ack_seq
!= tp
->snd_una
) &&
1701 !after(sp
[used_sacks
].end_seq
, tp
->snd_una
))
1703 mib_idx
= LINUX_MIB_TCPSACKDISCARD
;
1706 NET_INC_STATS(sock_net(sk
), mib_idx
);
1708 first_sack_index
= -1;
1712 /* Ignore very old stuff early */
1713 if (!after(sp
[used_sacks
].end_seq
, prior_snd_una
))
1719 /* order SACK blocks to allow in order walk of the retrans queue */
1720 for (i
= used_sacks
- 1; i
> 0; i
--) {
1721 for (j
= 0; j
< i
; j
++) {
1722 if (after(sp
[j
].start_seq
, sp
[j
+ 1].start_seq
)) {
1723 swap(sp
[j
], sp
[j
+ 1]);
1725 /* Track where the first SACK block goes to */
1726 if (j
== first_sack_index
)
1727 first_sack_index
= j
+ 1;
1732 skb
= tcp_write_queue_head(sk
);
1733 state
->fack_count
= 0;
1736 if (!tp
->sacked_out
) {
1737 /* It's already past, so skip checking against it */
1738 cache
= tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1740 cache
= tp
->recv_sack_cache
;
1741 /* Skip empty blocks in at head of the cache */
1742 while (tcp_sack_cache_ok(tp
, cache
) && !cache
->start_seq
&&
1747 while (i
< used_sacks
) {
1748 u32 start_seq
= sp
[i
].start_seq
;
1749 u32 end_seq
= sp
[i
].end_seq
;
1750 bool dup_sack
= (found_dup_sack
&& (i
== first_sack_index
));
1751 struct tcp_sack_block
*next_dup
= NULL
;
1753 if (found_dup_sack
&& ((i
+ 1) == first_sack_index
))
1754 next_dup
= &sp
[i
+ 1];
1756 /* Skip too early cached blocks */
1757 while (tcp_sack_cache_ok(tp
, cache
) &&
1758 !before(start_seq
, cache
->end_seq
))
1761 /* Can skip some work by looking recv_sack_cache? */
1762 if (tcp_sack_cache_ok(tp
, cache
) && !dup_sack
&&
1763 after(end_seq
, cache
->start_seq
)) {
1766 if (before(start_seq
, cache
->start_seq
)) {
1767 skb
= tcp_sacktag_skip(skb
, sk
, state
,
1769 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
,
1776 /* Rest of the block already fully processed? */
1777 if (!after(end_seq
, cache
->end_seq
))
1780 skb
= tcp_maybe_skipping_dsack(skb
, sk
, next_dup
,
1784 /* ...tail remains todo... */
1785 if (tcp_highest_sack_seq(tp
) == cache
->end_seq
) {
1786 /* ...but better entrypoint exists! */
1787 skb
= tcp_highest_sack(sk
);
1790 state
->fack_count
= tp
->fackets_out
;
1795 skb
= tcp_sacktag_skip(skb
, sk
, state
, cache
->end_seq
);
1796 /* Check overlap against next cached too (past this one already) */
1801 if (!before(start_seq
, tcp_highest_sack_seq(tp
))) {
1802 skb
= tcp_highest_sack(sk
);
1805 state
->fack_count
= tp
->fackets_out
;
1807 skb
= tcp_sacktag_skip(skb
, sk
, state
, start_seq
);
1810 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
, state
,
1811 start_seq
, end_seq
, dup_sack
);
1817 /* Clear the head of the cache sack blocks so we can skip it next time */
1818 for (i
= 0; i
< ARRAY_SIZE(tp
->recv_sack_cache
) - used_sacks
; i
++) {
1819 tp
->recv_sack_cache
[i
].start_seq
= 0;
1820 tp
->recv_sack_cache
[i
].end_seq
= 0;
1822 for (j
= 0; j
< used_sacks
; j
++)
1823 tp
->recv_sack_cache
[i
++] = sp
[j
];
1825 if ((state
->reord
< tp
->fackets_out
) &&
1826 ((inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Loss
) || tp
->undo_marker
))
1827 tcp_update_reordering(sk
, tp
->fackets_out
- state
->reord
, 0);
1829 tcp_verify_left_out(tp
);
1832 #if FASTRETRANS_DEBUG > 0
1833 WARN_ON((int)tp
->sacked_out
< 0);
1834 WARN_ON((int)tp
->lost_out
< 0);
1835 WARN_ON((int)tp
->retrans_out
< 0);
1836 WARN_ON((int)tcp_packets_in_flight(tp
) < 0);
1841 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1842 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1844 static bool tcp_limit_reno_sacked(struct tcp_sock
*tp
)
1848 holes
= max(tp
->lost_out
, 1U);
1849 holes
= min(holes
, tp
->packets_out
);
1851 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1852 tp
->sacked_out
= tp
->packets_out
- holes
;
1858 /* If we receive more dupacks than we expected counting segments
1859 * in assumption of absent reordering, interpret this as reordering.
1860 * The only another reason could be bug in receiver TCP.
1862 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1864 struct tcp_sock
*tp
= tcp_sk(sk
);
1865 if (tcp_limit_reno_sacked(tp
))
1866 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1869 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1871 static void tcp_add_reno_sack(struct sock
*sk
)
1873 struct tcp_sock
*tp
= tcp_sk(sk
);
1874 u32 prior_sacked
= tp
->sacked_out
;
1877 tcp_check_reno_reordering(sk
, 0);
1878 if (tp
->sacked_out
> prior_sacked
)
1879 tp
->delivered
++; /* Some out-of-order packet is delivered */
1880 tcp_verify_left_out(tp
);
1883 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1885 static void tcp_remove_reno_sacks(struct sock
*sk
, int acked
)
1887 struct tcp_sock
*tp
= tcp_sk(sk
);
1890 /* One ACK acked hole. The rest eat duplicate ACKs. */
1891 tp
->delivered
+= max_t(int, acked
- tp
->sacked_out
, 1);
1892 if (acked
- 1 >= tp
->sacked_out
)
1895 tp
->sacked_out
-= acked
- 1;
1897 tcp_check_reno_reordering(sk
, acked
);
1898 tcp_verify_left_out(tp
);
1901 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1906 void tcp_clear_retrans(struct tcp_sock
*tp
)
1908 tp
->retrans_out
= 0;
1910 tp
->undo_marker
= 0;
1911 tp
->undo_retrans
= -1;
1912 tp
->fackets_out
= 0;
1916 static inline void tcp_init_undo(struct tcp_sock
*tp
)
1918 tp
->undo_marker
= tp
->snd_una
;
1919 /* Retransmission still in flight may cause DSACKs later. */
1920 tp
->undo_retrans
= tp
->retrans_out
? : -1;
1923 /* Enter Loss state. If we detect SACK reneging, forget all SACK information
1924 * and reset tags completely, otherwise preserve SACKs. If receiver
1925 * dropped its ofo queue, we will know this due to reneging detection.
1927 void tcp_enter_loss(struct sock
*sk
)
1929 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1930 struct tcp_sock
*tp
= tcp_sk(sk
);
1931 struct net
*net
= sock_net(sk
);
1932 struct sk_buff
*skb
;
1933 bool is_reneg
; /* is receiver reneging on SACKs? */
1936 /* Reduce ssthresh if it has not yet been made inside this window. */
1937 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
||
1938 !after(tp
->high_seq
, tp
->snd_una
) ||
1939 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
1940 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1941 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1942 tcp_ca_event(sk
, CA_EVENT_LOSS
);
1946 tp
->snd_cwnd_cnt
= 0;
1947 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1949 tp
->retrans_out
= 0;
1952 if (tcp_is_reno(tp
))
1953 tcp_reset_reno_sack(tp
);
1955 skb
= tcp_write_queue_head(sk
);
1956 is_reneg
= skb
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
);
1958 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPSACKRENEGING
);
1960 tp
->fackets_out
= 0;
1962 tcp_clear_all_retrans_hints(tp
);
1964 tcp_for_write_queue(skb
, sk
) {
1965 if (skb
== tcp_send_head(sk
))
1968 mark_lost
= (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
) ||
1971 tcp_sum_lost(tp
, skb
);
1972 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
1974 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
1975 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1976 tp
->lost_out
+= tcp_skb_pcount(skb
);
1979 tcp_verify_left_out(tp
);
1981 /* Timeout in disordered state after receiving substantial DUPACKs
1982 * suggests that the degree of reordering is over-estimated.
1984 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
&&
1985 tp
->sacked_out
>= net
->ipv4
.sysctl_tcp_reordering
)
1986 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1987 net
->ipv4
.sysctl_tcp_reordering
);
1988 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1989 tp
->high_seq
= tp
->snd_nxt
;
1990 tcp_ecn_queue_cwr(tp
);
1992 /* F-RTO RFC5682 sec 3.1 step 1 mandates to disable F-RTO
1993 * if a previous recovery is underway, otherwise it may incorrectly
1994 * call a timeout spurious if some previously retransmitted packets
1995 * are s/acked (sec 3.2). We do not apply that retriction since
1996 * retransmitted skbs are permanently tagged with TCPCB_EVER_RETRANS
1997 * so FLAG_ORIG_SACK_ACKED is always correct. But we do disable F-RTO
1998 * on PTMU discovery to avoid sending new data.
2000 tp
->frto
= sysctl_tcp_frto
&& !inet_csk(sk
)->icsk_mtup
.probe_size
;
2003 /* If ACK arrived pointing to a remembered SACK, it means that our
2004 * remembered SACKs do not reflect real state of receiver i.e.
2005 * receiver _host_ is heavily congested (or buggy).
2007 * To avoid big spurious retransmission bursts due to transient SACK
2008 * scoreboard oddities that look like reneging, we give the receiver a
2009 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2010 * restore sanity to the SACK scoreboard. If the apparent reneging
2011 * persists until this RTO then we'll clear the SACK scoreboard.
2013 static bool tcp_check_sack_reneging(struct sock
*sk
, int flag
)
2015 if (flag
& FLAG_SACK_RENEGING
) {
2016 struct tcp_sock
*tp
= tcp_sk(sk
);
2017 unsigned long delay
= max(usecs_to_jiffies(tp
->srtt_us
>> 4),
2018 msecs_to_jiffies(10));
2020 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2021 delay
, TCP_RTO_MAX
);
2027 static inline int tcp_fackets_out(const struct tcp_sock
*tp
)
2029 return tcp_is_reno(tp
) ? tp
->sacked_out
+ 1 : tp
->fackets_out
;
2032 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2033 * counter when SACK is enabled (without SACK, sacked_out is used for
2036 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2037 * segments up to the highest received SACK block so far and holes in
2040 * With reordering, holes may still be in flight, so RFC3517 recovery
2041 * uses pure sacked_out (total number of SACKed segments) even though
2042 * it violates the RFC that uses duplicate ACKs, often these are equal
2043 * but when e.g. out-of-window ACKs or packet duplication occurs,
2044 * they differ. Since neither occurs due to loss, TCP should really
2047 static inline int tcp_dupack_heuristics(const struct tcp_sock
*tp
)
2049 return tcp_is_fack(tp
) ? tp
->fackets_out
: tp
->sacked_out
+ 1;
2052 /* Linux NewReno/SACK/FACK/ECN state machine.
2053 * --------------------------------------
2055 * "Open" Normal state, no dubious events, fast path.
2056 * "Disorder" In all the respects it is "Open",
2057 * but requires a bit more attention. It is entered when
2058 * we see some SACKs or dupacks. It is split of "Open"
2059 * mainly to move some processing from fast path to slow one.
2060 * "CWR" CWND was reduced due to some Congestion Notification event.
2061 * It can be ECN, ICMP source quench, local device congestion.
2062 * "Recovery" CWND was reduced, we are fast-retransmitting.
2063 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2065 * tcp_fastretrans_alert() is entered:
2066 * - each incoming ACK, if state is not "Open"
2067 * - when arrived ACK is unusual, namely:
2072 * Counting packets in flight is pretty simple.
2074 * in_flight = packets_out - left_out + retrans_out
2076 * packets_out is SND.NXT-SND.UNA counted in packets.
2078 * retrans_out is number of retransmitted segments.
2080 * left_out is number of segments left network, but not ACKed yet.
2082 * left_out = sacked_out + lost_out
2084 * sacked_out: Packets, which arrived to receiver out of order
2085 * and hence not ACKed. With SACKs this number is simply
2086 * amount of SACKed data. Even without SACKs
2087 * it is easy to give pretty reliable estimate of this number,
2088 * counting duplicate ACKs.
2090 * lost_out: Packets lost by network. TCP has no explicit
2091 * "loss notification" feedback from network (for now).
2092 * It means that this number can be only _guessed_.
2093 * Actually, it is the heuristics to predict lossage that
2094 * distinguishes different algorithms.
2096 * F.e. after RTO, when all the queue is considered as lost,
2097 * lost_out = packets_out and in_flight = retrans_out.
2099 * Essentially, we have now a few algorithms detecting
2102 * If the receiver supports SACK:
2104 * RFC6675/3517: It is the conventional algorithm. A packet is
2105 * considered lost if the number of higher sequence packets
2106 * SACKed is greater than or equal the DUPACK thoreshold
2107 * (reordering). This is implemented in tcp_mark_head_lost and
2108 * tcp_update_scoreboard.
2110 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm
2111 * (2017-) that checks timing instead of counting DUPACKs.
2112 * Essentially a packet is considered lost if it's not S/ACKed
2113 * after RTT + reordering_window, where both metrics are
2114 * dynamically measured and adjusted. This is implemented in
2115 * tcp_rack_mark_lost.
2117 * FACK (Disabled by default. Subsumbed by RACK):
2118 * It is the simplest heuristics. As soon as we decided
2119 * that something is lost, we decide that _all_ not SACKed
2120 * packets until the most forward SACK are lost. I.e.
2121 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2122 * It is absolutely correct estimate, if network does not reorder
2123 * packets. And it loses any connection to reality when reordering
2124 * takes place. We use FACK by default until reordering
2125 * is suspected on the path to this destination.
2127 * If the receiver does not support SACK:
2129 * NewReno (RFC6582): in Recovery we assume that one segment
2130 * is lost (classic Reno). While we are in Recovery and
2131 * a partial ACK arrives, we assume that one more packet
2132 * is lost (NewReno). This heuristics are the same in NewReno
2135 * Really tricky (and requiring careful tuning) part of algorithm
2136 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2137 * The first determines the moment _when_ we should reduce CWND and,
2138 * hence, slow down forward transmission. In fact, it determines the moment
2139 * when we decide that hole is caused by loss, rather than by a reorder.
2141 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2142 * holes, caused by lost packets.
2144 * And the most logically complicated part of algorithm is undo
2145 * heuristics. We detect false retransmits due to both too early
2146 * fast retransmit (reordering) and underestimated RTO, analyzing
2147 * timestamps and D-SACKs. When we detect that some segments were
2148 * retransmitted by mistake and CWND reduction was wrong, we undo
2149 * window reduction and abort recovery phase. This logic is hidden
2150 * inside several functions named tcp_try_undo_<something>.
2153 /* This function decides, when we should leave Disordered state
2154 * and enter Recovery phase, reducing congestion window.
2156 * Main question: may we further continue forward transmission
2157 * with the same cwnd?
2159 static bool tcp_time_to_recover(struct sock
*sk
, int flag
)
2161 struct tcp_sock
*tp
= tcp_sk(sk
);
2163 /* Trick#1: The loss is proven. */
2167 /* Not-A-Trick#2 : Classic rule... */
2168 if (tcp_dupack_heuristics(tp
) > tp
->reordering
)
2174 /* Detect loss in event "A" above by marking head of queue up as lost.
2175 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2176 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2177 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2178 * the maximum SACKed segments to pass before reaching this limit.
2180 static void tcp_mark_head_lost(struct sock
*sk
, int packets
, int mark_head
)
2182 struct tcp_sock
*tp
= tcp_sk(sk
);
2183 struct sk_buff
*skb
;
2184 int cnt
, oldcnt
, lost
;
2186 /* Use SACK to deduce losses of new sequences sent during recovery */
2187 const u32 loss_high
= tcp_is_sack(tp
) ? tp
->snd_nxt
: tp
->high_seq
;
2189 WARN_ON(packets
> tp
->packets_out
);
2190 if (tp
->lost_skb_hint
) {
2191 skb
= tp
->lost_skb_hint
;
2192 cnt
= tp
->lost_cnt_hint
;
2193 /* Head already handled? */
2194 if (mark_head
&& skb
!= tcp_write_queue_head(sk
))
2197 skb
= tcp_write_queue_head(sk
);
2201 tcp_for_write_queue_from(skb
, sk
) {
2202 if (skb
== tcp_send_head(sk
))
2204 /* TODO: do this better */
2205 /* this is not the most efficient way to do this... */
2206 tp
->lost_skb_hint
= skb
;
2207 tp
->lost_cnt_hint
= cnt
;
2209 if (after(TCP_SKB_CB(skb
)->end_seq
, loss_high
))
2213 if (tcp_is_fack(tp
) || tcp_is_reno(tp
) ||
2214 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2215 cnt
+= tcp_skb_pcount(skb
);
2217 if (cnt
> packets
) {
2218 if ((tcp_is_sack(tp
) && !tcp_is_fack(tp
)) ||
2219 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
) ||
2220 (oldcnt
>= packets
))
2223 mss
= tcp_skb_mss(skb
);
2224 /* If needed, chop off the prefix to mark as lost. */
2225 lost
= (packets
- oldcnt
) * mss
;
2226 if (lost
< skb
->len
&&
2227 tcp_fragment(sk
, skb
, lost
, mss
, GFP_ATOMIC
) < 0)
2232 tcp_skb_mark_lost(tp
, skb
);
2237 tcp_verify_left_out(tp
);
2240 /* Account newly detected lost packet(s) */
2242 static void tcp_update_scoreboard(struct sock
*sk
, int fast_rexmit
)
2244 struct tcp_sock
*tp
= tcp_sk(sk
);
2246 if (tcp_is_reno(tp
)) {
2247 tcp_mark_head_lost(sk
, 1, 1);
2248 } else if (tcp_is_fack(tp
)) {
2249 int lost
= tp
->fackets_out
- tp
->reordering
;
2252 tcp_mark_head_lost(sk
, lost
, 0);
2254 int sacked_upto
= tp
->sacked_out
- tp
->reordering
;
2255 if (sacked_upto
>= 0)
2256 tcp_mark_head_lost(sk
, sacked_upto
, 0);
2257 else if (fast_rexmit
)
2258 tcp_mark_head_lost(sk
, 1, 1);
2262 static bool tcp_tsopt_ecr_before(const struct tcp_sock
*tp
, u32 when
)
2264 return tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2265 before(tp
->rx_opt
.rcv_tsecr
, when
);
2268 /* skb is spurious retransmitted if the returned timestamp echo
2269 * reply is prior to the skb transmission time
2271 static bool tcp_skb_spurious_retrans(const struct tcp_sock
*tp
,
2272 const struct sk_buff
*skb
)
2274 return (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
) &&
2275 tcp_tsopt_ecr_before(tp
, tcp_skb_timestamp(skb
));
2278 /* Nothing was retransmitted or returned timestamp is less
2279 * than timestamp of the first retransmission.
2281 static inline bool tcp_packet_delayed(const struct tcp_sock
*tp
)
2283 return !tp
->retrans_stamp
||
2284 tcp_tsopt_ecr_before(tp
, tp
->retrans_stamp
);
2287 /* Undo procedures. */
2289 /* We can clear retrans_stamp when there are no retransmissions in the
2290 * window. It would seem that it is trivially available for us in
2291 * tp->retrans_out, however, that kind of assumptions doesn't consider
2292 * what will happen if errors occur when sending retransmission for the
2293 * second time. ...It could the that such segment has only
2294 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2295 * the head skb is enough except for some reneging corner cases that
2296 * are not worth the effort.
2298 * Main reason for all this complexity is the fact that connection dying
2299 * time now depends on the validity of the retrans_stamp, in particular,
2300 * that successive retransmissions of a segment must not advance
2301 * retrans_stamp under any conditions.
2303 static bool tcp_any_retrans_done(const struct sock
*sk
)
2305 const struct tcp_sock
*tp
= tcp_sk(sk
);
2306 struct sk_buff
*skb
;
2308 if (tp
->retrans_out
)
2311 skb
= tcp_write_queue_head(sk
);
2312 if (unlikely(skb
&& TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
))
2318 #if FASTRETRANS_DEBUG > 1
2319 static void DBGUNDO(struct sock
*sk
, const char *msg
)
2321 struct tcp_sock
*tp
= tcp_sk(sk
);
2322 struct inet_sock
*inet
= inet_sk(sk
);
2324 if (sk
->sk_family
== AF_INET
) {
2325 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2327 &inet
->inet_daddr
, ntohs(inet
->inet_dport
),
2328 tp
->snd_cwnd
, tcp_left_out(tp
),
2329 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2332 #if IS_ENABLED(CONFIG_IPV6)
2333 else if (sk
->sk_family
== AF_INET6
) {
2334 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2336 &sk
->sk_v6_daddr
, ntohs(inet
->inet_dport
),
2337 tp
->snd_cwnd
, tcp_left_out(tp
),
2338 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2344 #define DBGUNDO(x...) do { } while (0)
2347 static void tcp_undo_cwnd_reduction(struct sock
*sk
, bool unmark_loss
)
2349 struct tcp_sock
*tp
= tcp_sk(sk
);
2352 struct sk_buff
*skb
;
2354 tcp_for_write_queue(skb
, sk
) {
2355 if (skb
== tcp_send_head(sk
))
2357 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2360 tcp_clear_all_retrans_hints(tp
);
2363 if (tp
->prior_ssthresh
) {
2364 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2366 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
2368 if (tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
2369 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
2370 tcp_ecn_withdraw_cwr(tp
);
2373 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2374 tp
->undo_marker
= 0;
2377 static inline bool tcp_may_undo(const struct tcp_sock
*tp
)
2379 return tp
->undo_marker
&& (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
2382 /* People celebrate: "We love our President!" */
2383 static bool tcp_try_undo_recovery(struct sock
*sk
)
2385 struct tcp_sock
*tp
= tcp_sk(sk
);
2387 if (tcp_may_undo(tp
)) {
2390 /* Happy end! We did not retransmit anything
2391 * or our original transmission succeeded.
2393 DBGUNDO(sk
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
2394 tcp_undo_cwnd_reduction(sk
, false);
2395 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
2396 mib_idx
= LINUX_MIB_TCPLOSSUNDO
;
2398 mib_idx
= LINUX_MIB_TCPFULLUNDO
;
2400 NET_INC_STATS(sock_net(sk
), mib_idx
);
2402 if (tp
->snd_una
== tp
->high_seq
&& tcp_is_reno(tp
)) {
2403 /* Hold old state until something *above* high_seq
2404 * is ACKed. For Reno it is MUST to prevent false
2405 * fast retransmits (RFC2582). SACK TCP is safe. */
2406 if (!tcp_any_retrans_done(sk
))
2407 tp
->retrans_stamp
= 0;
2410 tcp_set_ca_state(sk
, TCP_CA_Open
);
2414 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2415 static bool tcp_try_undo_dsack(struct sock
*sk
)
2417 struct tcp_sock
*tp
= tcp_sk(sk
);
2419 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
2420 DBGUNDO(sk
, "D-SACK");
2421 tcp_undo_cwnd_reduction(sk
, false);
2422 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPDSACKUNDO
);
2428 /* Undo during loss recovery after partial ACK or using F-RTO. */
2429 static bool tcp_try_undo_loss(struct sock
*sk
, bool frto_undo
)
2431 struct tcp_sock
*tp
= tcp_sk(sk
);
2433 if (frto_undo
|| tcp_may_undo(tp
)) {
2434 tcp_undo_cwnd_reduction(sk
, true);
2436 DBGUNDO(sk
, "partial loss");
2437 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPLOSSUNDO
);
2439 NET_INC_STATS(sock_net(sk
),
2440 LINUX_MIB_TCPSPURIOUSRTOS
);
2441 inet_csk(sk
)->icsk_retransmits
= 0;
2442 if (frto_undo
|| tcp_is_sack(tp
))
2443 tcp_set_ca_state(sk
, TCP_CA_Open
);
2449 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2450 * It computes the number of packets to send (sndcnt) based on packets newly
2452 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2453 * cwnd reductions across a full RTT.
2454 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2455 * But when the retransmits are acked without further losses, PRR
2456 * slow starts cwnd up to ssthresh to speed up the recovery.
2458 static void tcp_init_cwnd_reduction(struct sock
*sk
)
2460 struct tcp_sock
*tp
= tcp_sk(sk
);
2462 tp
->high_seq
= tp
->snd_nxt
;
2463 tp
->tlp_high_seq
= 0;
2464 tp
->snd_cwnd_cnt
= 0;
2465 tp
->prior_cwnd
= tp
->snd_cwnd
;
2466 tp
->prr_delivered
= 0;
2468 tp
->snd_ssthresh
= inet_csk(sk
)->icsk_ca_ops
->ssthresh(sk
);
2469 tcp_ecn_queue_cwr(tp
);
2472 void tcp_cwnd_reduction(struct sock
*sk
, int newly_acked_sacked
, int flag
)
2474 struct tcp_sock
*tp
= tcp_sk(sk
);
2476 int delta
= tp
->snd_ssthresh
- tcp_packets_in_flight(tp
);
2478 if (newly_acked_sacked
<= 0 || WARN_ON_ONCE(!tp
->prior_cwnd
))
2481 tp
->prr_delivered
+= newly_acked_sacked
;
2483 u64 dividend
= (u64
)tp
->snd_ssthresh
* tp
->prr_delivered
+
2485 sndcnt
= div_u64(dividend
, tp
->prior_cwnd
) - tp
->prr_out
;
2486 } else if ((flag
& FLAG_RETRANS_DATA_ACKED
) &&
2487 !(flag
& FLAG_LOST_RETRANS
)) {
2488 sndcnt
= min_t(int, delta
,
2489 max_t(int, tp
->prr_delivered
- tp
->prr_out
,
2490 newly_acked_sacked
) + 1);
2492 sndcnt
= min(delta
, newly_acked_sacked
);
2494 /* Force a fast retransmit upon entering fast recovery */
2495 sndcnt
= max(sndcnt
, (tp
->prr_out
? 0 : 1));
2496 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + sndcnt
;
2499 static inline void tcp_end_cwnd_reduction(struct sock
*sk
)
2501 struct tcp_sock
*tp
= tcp_sk(sk
);
2503 if (inet_csk(sk
)->icsk_ca_ops
->cong_control
)
2506 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2507 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_CWR
||
2508 (tp
->undo_marker
&& tp
->snd_ssthresh
< TCP_INFINITE_SSTHRESH
)) {
2509 tp
->snd_cwnd
= tp
->snd_ssthresh
;
2510 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2512 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
2515 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2516 void tcp_enter_cwr(struct sock
*sk
)
2518 struct tcp_sock
*tp
= tcp_sk(sk
);
2520 tp
->prior_ssthresh
= 0;
2521 if (inet_csk(sk
)->icsk_ca_state
< TCP_CA_CWR
) {
2522 tp
->undo_marker
= 0;
2523 tcp_init_cwnd_reduction(sk
);
2524 tcp_set_ca_state(sk
, TCP_CA_CWR
);
2527 EXPORT_SYMBOL(tcp_enter_cwr
);
2529 static void tcp_try_keep_open(struct sock
*sk
)
2531 struct tcp_sock
*tp
= tcp_sk(sk
);
2532 int state
= TCP_CA_Open
;
2534 if (tcp_left_out(tp
) || tcp_any_retrans_done(sk
))
2535 state
= TCP_CA_Disorder
;
2537 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
2538 tcp_set_ca_state(sk
, state
);
2539 tp
->high_seq
= tp
->snd_nxt
;
2543 static void tcp_try_to_open(struct sock
*sk
, int flag
)
2545 struct tcp_sock
*tp
= tcp_sk(sk
);
2547 tcp_verify_left_out(tp
);
2549 if (!tcp_any_retrans_done(sk
))
2550 tp
->retrans_stamp
= 0;
2552 if (flag
& FLAG_ECE
)
2555 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
2556 tcp_try_keep_open(sk
);
2560 static void tcp_mtup_probe_failed(struct sock
*sk
)
2562 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2564 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
2565 icsk
->icsk_mtup
.probe_size
= 0;
2566 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPMTUPFAIL
);
2569 static void tcp_mtup_probe_success(struct sock
*sk
)
2571 struct tcp_sock
*tp
= tcp_sk(sk
);
2572 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2574 /* FIXME: breaks with very large cwnd */
2575 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2576 tp
->snd_cwnd
= tp
->snd_cwnd
*
2577 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
2578 icsk
->icsk_mtup
.probe_size
;
2579 tp
->snd_cwnd_cnt
= 0;
2580 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2581 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2583 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
2584 icsk
->icsk_mtup
.probe_size
= 0;
2585 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
2586 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPMTUPSUCCESS
);
2589 /* Do a simple retransmit without using the backoff mechanisms in
2590 * tcp_timer. This is used for path mtu discovery.
2591 * The socket is already locked here.
2593 void tcp_simple_retransmit(struct sock
*sk
)
2595 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2596 struct tcp_sock
*tp
= tcp_sk(sk
);
2597 struct sk_buff
*skb
;
2598 unsigned int mss
= tcp_current_mss(sk
);
2599 u32 prior_lost
= tp
->lost_out
;
2601 tcp_for_write_queue(skb
, sk
) {
2602 if (skb
== tcp_send_head(sk
))
2604 if (tcp_skb_seglen(skb
) > mss
&&
2605 !(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
2606 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2607 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
2608 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2610 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
2614 tcp_clear_retrans_hints_partial(tp
);
2616 if (prior_lost
== tp
->lost_out
)
2619 if (tcp_is_reno(tp
))
2620 tcp_limit_reno_sacked(tp
);
2622 tcp_verify_left_out(tp
);
2624 /* Don't muck with the congestion window here.
2625 * Reason is that we do not increase amount of _data_
2626 * in network, but units changed and effective
2627 * cwnd/ssthresh really reduced now.
2629 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
2630 tp
->high_seq
= tp
->snd_nxt
;
2631 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2632 tp
->prior_ssthresh
= 0;
2633 tp
->undo_marker
= 0;
2634 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2636 tcp_xmit_retransmit_queue(sk
);
2638 EXPORT_SYMBOL(tcp_simple_retransmit
);
2640 void tcp_enter_recovery(struct sock
*sk
, bool ece_ack
)
2642 struct tcp_sock
*tp
= tcp_sk(sk
);
2645 if (tcp_is_reno(tp
))
2646 mib_idx
= LINUX_MIB_TCPRENORECOVERY
;
2648 mib_idx
= LINUX_MIB_TCPSACKRECOVERY
;
2650 NET_INC_STATS(sock_net(sk
), mib_idx
);
2652 tp
->prior_ssthresh
= 0;
2655 if (!tcp_in_cwnd_reduction(sk
)) {
2657 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2658 tcp_init_cwnd_reduction(sk
);
2660 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
2663 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2664 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2666 static void tcp_process_loss(struct sock
*sk
, int flag
, bool is_dupack
,
2669 struct tcp_sock
*tp
= tcp_sk(sk
);
2670 bool recovered
= !before(tp
->snd_una
, tp
->high_seq
);
2672 if ((flag
& FLAG_SND_UNA_ADVANCED
) &&
2673 tcp_try_undo_loss(sk
, false))
2676 /* The ACK (s)acks some never-retransmitted data meaning not all
2677 * the data packets before the timeout were lost. Therefore we
2678 * undo the congestion window and state. This is essentially
2679 * the operation in F-RTO (RFC5682 section 3.1 step 3.b). Since
2680 * a retransmitted skb is permantly marked, we can apply such an
2681 * operation even if F-RTO was not used.
2683 if ((flag
& FLAG_ORIG_SACK_ACKED
) &&
2684 tcp_try_undo_loss(sk
, tp
->undo_marker
))
2687 if (tp
->frto
) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2688 if (after(tp
->snd_nxt
, tp
->high_seq
)) {
2689 if (flag
& FLAG_DATA_SACKED
|| is_dupack
)
2690 tp
->frto
= 0; /* Step 3.a. loss was real */
2691 } else if (flag
& FLAG_SND_UNA_ADVANCED
&& !recovered
) {
2692 tp
->high_seq
= tp
->snd_nxt
;
2693 /* Step 2.b. Try send new data (but deferred until cwnd
2694 * is updated in tcp_ack()). Otherwise fall back to
2695 * the conventional recovery.
2697 if (tcp_send_head(sk
) &&
2698 after(tcp_wnd_end(tp
), tp
->snd_nxt
)) {
2699 *rexmit
= REXMIT_NEW
;
2707 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2708 tcp_try_undo_recovery(sk
);
2711 if (tcp_is_reno(tp
)) {
2712 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2713 * delivered. Lower inflight to clock out (re)tranmissions.
2715 if (after(tp
->snd_nxt
, tp
->high_seq
) && is_dupack
)
2716 tcp_add_reno_sack(sk
);
2717 else if (flag
& FLAG_SND_UNA_ADVANCED
)
2718 tcp_reset_reno_sack(tp
);
2720 *rexmit
= REXMIT_LOST
;
2723 /* Undo during fast recovery after partial ACK. */
2724 static bool tcp_try_undo_partial(struct sock
*sk
, const int acked
)
2726 struct tcp_sock
*tp
= tcp_sk(sk
);
2728 if (tp
->undo_marker
&& tcp_packet_delayed(tp
)) {
2729 /* Plain luck! Hole if filled with delayed
2730 * packet, rather than with a retransmit.
2732 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
2734 /* We are getting evidence that the reordering degree is higher
2735 * than we realized. If there are no retransmits out then we
2736 * can undo. Otherwise we clock out new packets but do not
2737 * mark more packets lost or retransmit more.
2739 if (tp
->retrans_out
)
2742 if (!tcp_any_retrans_done(sk
))
2743 tp
->retrans_stamp
= 0;
2745 DBGUNDO(sk
, "partial recovery");
2746 tcp_undo_cwnd_reduction(sk
, true);
2747 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPPARTIALUNDO
);
2748 tcp_try_keep_open(sk
);
2754 static void tcp_rack_identify_loss(struct sock
*sk
, int *ack_flag
,
2755 const struct skb_mstamp
*ack_time
)
2757 struct tcp_sock
*tp
= tcp_sk(sk
);
2759 /* Use RACK to detect loss */
2760 if (sysctl_tcp_recovery
& TCP_RACK_LOSS_DETECTION
) {
2761 u32 prior_retrans
= tp
->retrans_out
;
2763 tcp_rack_mark_lost(sk
, ack_time
);
2764 if (prior_retrans
> tp
->retrans_out
)
2765 *ack_flag
|= FLAG_LOST_RETRANS
;
2769 /* Process an event, which can update packets-in-flight not trivially.
2770 * Main goal of this function is to calculate new estimate for left_out,
2771 * taking into account both packets sitting in receiver's buffer and
2772 * packets lost by network.
2774 * Besides that it updates the congestion state when packet loss or ECN
2775 * is detected. But it does not reduce the cwnd, it is done by the
2776 * congestion control later.
2778 * It does _not_ decide what to send, it is made in function
2779 * tcp_xmit_retransmit_queue().
2781 static void tcp_fastretrans_alert(struct sock
*sk
, const int acked
,
2782 bool is_dupack
, int *ack_flag
, int *rexmit
,
2783 const struct skb_mstamp
*ack_time
)
2785 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2786 struct tcp_sock
*tp
= tcp_sk(sk
);
2787 int fast_rexmit
= 0, flag
= *ack_flag
;
2788 bool do_lost
= is_dupack
|| ((flag
& FLAG_DATA_SACKED
) &&
2789 (tcp_fackets_out(tp
) > tp
->reordering
));
2791 if (WARN_ON(!tp
->packets_out
&& tp
->sacked_out
))
2793 if (WARN_ON(!tp
->sacked_out
&& tp
->fackets_out
))
2794 tp
->fackets_out
= 0;
2796 /* Now state machine starts.
2797 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2798 if (flag
& FLAG_ECE
)
2799 tp
->prior_ssthresh
= 0;
2801 /* B. In all the states check for reneging SACKs. */
2802 if (tcp_check_sack_reneging(sk
, flag
))
2805 /* C. Check consistency of the current state. */
2806 tcp_verify_left_out(tp
);
2808 /* D. Check state exit conditions. State can be terminated
2809 * when high_seq is ACKed. */
2810 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
2811 WARN_ON(tp
->retrans_out
!= 0);
2812 tp
->retrans_stamp
= 0;
2813 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
2814 switch (icsk
->icsk_ca_state
) {
2816 /* CWR is to be held something *above* high_seq
2817 * is ACKed for CWR bit to reach receiver. */
2818 if (tp
->snd_una
!= tp
->high_seq
) {
2819 tcp_end_cwnd_reduction(sk
);
2820 tcp_set_ca_state(sk
, TCP_CA_Open
);
2824 case TCP_CA_Recovery
:
2825 if (tcp_is_reno(tp
))
2826 tcp_reset_reno_sack(tp
);
2827 if (tcp_try_undo_recovery(sk
))
2829 tcp_end_cwnd_reduction(sk
);
2834 /* E. Process state. */
2835 switch (icsk
->icsk_ca_state
) {
2836 case TCP_CA_Recovery
:
2837 if (!(flag
& FLAG_SND_UNA_ADVANCED
)) {
2838 if (tcp_is_reno(tp
) && is_dupack
)
2839 tcp_add_reno_sack(sk
);
2841 if (tcp_try_undo_partial(sk
, acked
))
2843 /* Partial ACK arrived. Force fast retransmit. */
2844 do_lost
= tcp_is_reno(tp
) ||
2845 tcp_fackets_out(tp
) > tp
->reordering
;
2847 if (tcp_try_undo_dsack(sk
)) {
2848 tcp_try_keep_open(sk
);
2851 tcp_rack_identify_loss(sk
, ack_flag
, ack_time
);
2854 tcp_process_loss(sk
, flag
, is_dupack
, rexmit
);
2855 tcp_rack_identify_loss(sk
, ack_flag
, ack_time
);
2856 if (!(icsk
->icsk_ca_state
== TCP_CA_Open
||
2857 (*ack_flag
& FLAG_LOST_RETRANS
)))
2859 /* Change state if cwnd is undone or retransmits are lost */
2861 if (tcp_is_reno(tp
)) {
2862 if (flag
& FLAG_SND_UNA_ADVANCED
)
2863 tcp_reset_reno_sack(tp
);
2865 tcp_add_reno_sack(sk
);
2868 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
)
2869 tcp_try_undo_dsack(sk
);
2871 tcp_rack_identify_loss(sk
, ack_flag
, ack_time
);
2872 if (!tcp_time_to_recover(sk
, flag
)) {
2873 tcp_try_to_open(sk
, flag
);
2877 /* MTU probe failure: don't reduce cwnd */
2878 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
2879 icsk
->icsk_mtup
.probe_size
&&
2880 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
2881 tcp_mtup_probe_failed(sk
);
2882 /* Restores the reduction we did in tcp_mtup_probe() */
2884 tcp_simple_retransmit(sk
);
2888 /* Otherwise enter Recovery state */
2889 tcp_enter_recovery(sk
, (flag
& FLAG_ECE
));
2894 tcp_update_scoreboard(sk
, fast_rexmit
);
2895 *rexmit
= REXMIT_LOST
;
2898 static void tcp_update_rtt_min(struct sock
*sk
, u32 rtt_us
)
2900 struct tcp_sock
*tp
= tcp_sk(sk
);
2901 u32 wlen
= sysctl_tcp_min_rtt_wlen
* HZ
;
2903 minmax_running_min(&tp
->rtt_min
, wlen
, tcp_time_stamp
,
2904 rtt_us
? : jiffies_to_usecs(1));
2907 static inline bool tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
2908 long seq_rtt_us
, long sack_rtt_us
,
2911 const struct tcp_sock
*tp
= tcp_sk(sk
);
2913 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2914 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2915 * Karn's algorithm forbids taking RTT if some retransmitted data
2916 * is acked (RFC6298).
2919 seq_rtt_us
= sack_rtt_us
;
2921 /* RTTM Rule: A TSecr value received in a segment is used to
2922 * update the averaged RTT measurement only if the segment
2923 * acknowledges some new data, i.e., only if it advances the
2924 * left edge of the send window.
2925 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2927 if (seq_rtt_us
< 0 && tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2929 seq_rtt_us
= ca_rtt_us
= jiffies_to_usecs(tcp_time_stamp
-
2930 tp
->rx_opt
.rcv_tsecr
);
2934 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
2935 * always taken together with ACK, SACK, or TS-opts. Any negative
2936 * values will be skipped with the seq_rtt_us < 0 check above.
2938 tcp_update_rtt_min(sk
, ca_rtt_us
);
2939 tcp_rtt_estimator(sk
, seq_rtt_us
);
2942 /* RFC6298: only reset backoff on valid RTT measurement. */
2943 inet_csk(sk
)->icsk_backoff
= 0;
2947 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2948 void tcp_synack_rtt_meas(struct sock
*sk
, struct request_sock
*req
)
2952 if (req
&& !req
->num_retrans
&& tcp_rsk(req
)->snt_synack
.v64
) {
2953 struct skb_mstamp now
;
2955 skb_mstamp_get(&now
);
2956 rtt_us
= skb_mstamp_us_delta(&now
, &tcp_rsk(req
)->snt_synack
);
2959 tcp_ack_update_rtt(sk
, FLAG_SYN_ACKED
, rtt_us
, -1L, rtt_us
);
2963 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
, u32 acked
)
2965 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2967 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, acked
);
2968 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_time_stamp
;
2971 /* Restart timer after forward progress on connection.
2972 * RFC2988 recommends to restart timer to now+rto.
2974 void tcp_rearm_rto(struct sock
*sk
)
2976 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2977 struct tcp_sock
*tp
= tcp_sk(sk
);
2979 /* If the retrans timer is currently being used by Fast Open
2980 * for SYN-ACK retrans purpose, stay put.
2982 if (tp
->fastopen_rsk
)
2985 if (!tp
->packets_out
) {
2986 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
2988 u32 rto
= inet_csk(sk
)->icsk_rto
;
2989 /* Offset the time elapsed after installing regular RTO */
2990 if (icsk
->icsk_pending
== ICSK_TIME_REO_TIMEOUT
||
2991 icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
) {
2992 struct sk_buff
*skb
= tcp_write_queue_head(sk
);
2993 const u32 rto_time_stamp
=
2994 tcp_skb_timestamp(skb
) + rto
;
2995 s32 delta
= (s32
)(rto_time_stamp
- tcp_time_stamp
);
2996 /* delta may not be positive if the socket is locked
2997 * when the retrans timer fires and is rescheduled.
3002 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
, rto
,
3007 /* If we get here, the whole TSO packet has not been acked. */
3008 static u32
tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
)
3010 struct tcp_sock
*tp
= tcp_sk(sk
);
3013 BUG_ON(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
));
3015 packets_acked
= tcp_skb_pcount(skb
);
3016 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
3018 packets_acked
-= tcp_skb_pcount(skb
);
3020 if (packets_acked
) {
3021 BUG_ON(tcp_skb_pcount(skb
) == 0);
3022 BUG_ON(!before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
));
3025 return packets_acked
;
3028 static void tcp_ack_tstamp(struct sock
*sk
, struct sk_buff
*skb
,
3031 const struct skb_shared_info
*shinfo
;
3033 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3034 if (likely(!TCP_SKB_CB(skb
)->txstamp_ack
))
3037 shinfo
= skb_shinfo(skb
);
3038 if (!before(shinfo
->tskey
, prior_snd_una
) &&
3039 before(shinfo
->tskey
, tcp_sk(sk
)->snd_una
))
3040 __skb_tstamp_tx(skb
, NULL
, sk
, SCM_TSTAMP_ACK
);
3043 /* Remove acknowledged frames from the retransmission queue. If our packet
3044 * is before the ack sequence we can discard it as it's confirmed to have
3045 * arrived at the other end.
3047 static int tcp_clean_rtx_queue(struct sock
*sk
, int prior_fackets
,
3048 u32 prior_snd_una
, int *acked
,
3049 struct tcp_sacktag_state
*sack
)
3051 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3052 struct skb_mstamp first_ackt
, last_ackt
;
3053 struct skb_mstamp
*now
= &sack
->ack_time
;
3054 struct tcp_sock
*tp
= tcp_sk(sk
);
3055 u32 prior_sacked
= tp
->sacked_out
;
3056 u32 reord
= tp
->packets_out
;
3057 bool fully_acked
= true;
3058 long sack_rtt_us
= -1L;
3059 long seq_rtt_us
= -1L;
3060 long ca_rtt_us
= -1L;
3061 struct sk_buff
*skb
;
3063 u32 last_in_flight
= 0;
3069 while ((skb
= tcp_write_queue_head(sk
)) && skb
!= tcp_send_head(sk
)) {
3070 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
3071 u8 sacked
= scb
->sacked
;
3074 tcp_ack_tstamp(sk
, skb
, prior_snd_una
);
3076 /* Determine how many packets and what bytes were acked, tso and else */
3077 if (after(scb
->end_seq
, tp
->snd_una
)) {
3078 if (tcp_skb_pcount(skb
) == 1 ||
3079 !after(tp
->snd_una
, scb
->seq
))
3082 acked_pcount
= tcp_tso_acked(sk
, skb
);
3085 fully_acked
= false;
3087 /* Speedup tcp_unlink_write_queue() and next loop */
3088 prefetchw(skb
->next
);
3089 acked_pcount
= tcp_skb_pcount(skb
);
3092 if (unlikely(sacked
& TCPCB_RETRANS
)) {
3093 if (sacked
& TCPCB_SACKED_RETRANS
)
3094 tp
->retrans_out
-= acked_pcount
;
3095 flag
|= FLAG_RETRANS_DATA_ACKED
;
3096 } else if (!(sacked
& TCPCB_SACKED_ACKED
)) {
3097 last_ackt
= skb
->skb_mstamp
;
3098 WARN_ON_ONCE(last_ackt
.v64
== 0);
3099 if (!first_ackt
.v64
)
3100 first_ackt
= last_ackt
;
3102 last_in_flight
= TCP_SKB_CB(skb
)->tx
.in_flight
;
3103 reord
= min(pkts_acked
, reord
);
3104 if (!after(scb
->end_seq
, tp
->high_seq
))
3105 flag
|= FLAG_ORIG_SACK_ACKED
;
3108 if (sacked
& TCPCB_SACKED_ACKED
) {
3109 tp
->sacked_out
-= acked_pcount
;
3110 } else if (tcp_is_sack(tp
)) {
3111 tp
->delivered
+= acked_pcount
;
3112 if (!tcp_skb_spurious_retrans(tp
, skb
))
3113 tcp_rack_advance(tp
, sacked
, scb
->end_seq
,
3117 if (sacked
& TCPCB_LOST
)
3118 tp
->lost_out
-= acked_pcount
;
3120 tp
->packets_out
-= acked_pcount
;
3121 pkts_acked
+= acked_pcount
;
3122 tcp_rate_skb_delivered(sk
, skb
, sack
->rate
);
3124 /* Initial outgoing SYN's get put onto the write_queue
3125 * just like anything else we transmit. It is not
3126 * true data, and if we misinform our callers that
3127 * this ACK acks real data, we will erroneously exit
3128 * connection startup slow start one packet too
3129 * quickly. This is severely frowned upon behavior.
3131 if (likely(!(scb
->tcp_flags
& TCPHDR_SYN
))) {
3132 flag
|= FLAG_DATA_ACKED
;
3134 flag
|= FLAG_SYN_ACKED
;
3135 tp
->retrans_stamp
= 0;
3141 tcp_unlink_write_queue(skb
, sk
);
3142 sk_wmem_free_skb(sk
, skb
);
3143 if (unlikely(skb
== tp
->retransmit_skb_hint
))
3144 tp
->retransmit_skb_hint
= NULL
;
3145 if (unlikely(skb
== tp
->lost_skb_hint
))
3146 tp
->lost_skb_hint
= NULL
;
3150 tcp_chrono_stop(sk
, TCP_CHRONO_BUSY
);
3152 if (likely(between(tp
->snd_up
, prior_snd_una
, tp
->snd_una
)))
3153 tp
->snd_up
= tp
->snd_una
;
3155 if (skb
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
3156 flag
|= FLAG_SACK_RENEGING
;
3158 if (likely(first_ackt
.v64
) && !(flag
& FLAG_RETRANS_DATA_ACKED
)) {
3159 seq_rtt_us
= skb_mstamp_us_delta(now
, &first_ackt
);
3160 ca_rtt_us
= skb_mstamp_us_delta(now
, &last_ackt
);
3162 if (sack
->first_sackt
.v64
) {
3163 sack_rtt_us
= skb_mstamp_us_delta(now
, &sack
->first_sackt
);
3164 ca_rtt_us
= skb_mstamp_us_delta(now
, &sack
->last_sackt
);
3166 sack
->rate
->rtt_us
= ca_rtt_us
; /* RTT of last (S)ACKed packet, or -1 */
3167 rtt_update
= tcp_ack_update_rtt(sk
, flag
, seq_rtt_us
, sack_rtt_us
,
3170 if (flag
& FLAG_ACKED
) {
3172 if (unlikely(icsk
->icsk_mtup
.probe_size
&&
3173 !after(tp
->mtu_probe
.probe_seq_end
, tp
->snd_una
))) {
3174 tcp_mtup_probe_success(sk
);
3177 if (tcp_is_reno(tp
)) {
3178 tcp_remove_reno_sacks(sk
, pkts_acked
);
3182 /* Non-retransmitted hole got filled? That's reordering */
3183 if (reord
< prior_fackets
)
3184 tcp_update_reordering(sk
, tp
->fackets_out
- reord
, 0);
3186 delta
= tcp_is_fack(tp
) ? pkts_acked
:
3187 prior_sacked
- tp
->sacked_out
;
3188 tp
->lost_cnt_hint
-= min(tp
->lost_cnt_hint
, delta
);
3191 tp
->fackets_out
-= min(pkts_acked
, tp
->fackets_out
);
3193 } else if (skb
&& rtt_update
&& sack_rtt_us
>= 0 &&
3194 sack_rtt_us
> skb_mstamp_us_delta(now
, &skb
->skb_mstamp
)) {
3195 /* Do not re-arm RTO if the sack RTT is measured from data sent
3196 * after when the head was last (re)transmitted. Otherwise the
3197 * timeout may continue to extend in loss recovery.
3202 if (icsk
->icsk_ca_ops
->pkts_acked
) {
3203 struct ack_sample sample
= { .pkts_acked
= pkts_acked
,
3204 .rtt_us
= ca_rtt_us
,
3205 .in_flight
= last_in_flight
};
3207 icsk
->icsk_ca_ops
->pkts_acked(sk
, &sample
);
3210 #if FASTRETRANS_DEBUG > 0
3211 WARN_ON((int)tp
->sacked_out
< 0);
3212 WARN_ON((int)tp
->lost_out
< 0);
3213 WARN_ON((int)tp
->retrans_out
< 0);
3214 if (!tp
->packets_out
&& tcp_is_sack(tp
)) {
3215 icsk
= inet_csk(sk
);
3217 pr_debug("Leak l=%u %d\n",
3218 tp
->lost_out
, icsk
->icsk_ca_state
);
3221 if (tp
->sacked_out
) {
3222 pr_debug("Leak s=%u %d\n",
3223 tp
->sacked_out
, icsk
->icsk_ca_state
);
3226 if (tp
->retrans_out
) {
3227 pr_debug("Leak r=%u %d\n",
3228 tp
->retrans_out
, icsk
->icsk_ca_state
);
3229 tp
->retrans_out
= 0;
3233 *acked
= pkts_acked
;
3237 static void tcp_ack_probe(struct sock
*sk
)
3239 const struct tcp_sock
*tp
= tcp_sk(sk
);
3240 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3242 /* Was it a usable window open? */
3244 if (!after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
, tcp_wnd_end(tp
))) {
3245 icsk
->icsk_backoff
= 0;
3246 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
3247 /* Socket must be waked up by subsequent tcp_data_snd_check().
3248 * This function is not for random using!
3251 unsigned long when
= tcp_probe0_when(sk
, TCP_RTO_MAX
);
3253 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3258 static inline bool tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
3260 return !(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
3261 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
;
3264 /* Decide wheather to run the increase function of congestion control. */
3265 static inline bool tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
3267 /* If reordering is high then always grow cwnd whenever data is
3268 * delivered regardless of its ordering. Otherwise stay conservative
3269 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3270 * new SACK or ECE mark may first advance cwnd here and later reduce
3271 * cwnd in tcp_fastretrans_alert() based on more states.
3273 if (tcp_sk(sk
)->reordering
> sock_net(sk
)->ipv4
.sysctl_tcp_reordering
)
3274 return flag
& FLAG_FORWARD_PROGRESS
;
3276 return flag
& FLAG_DATA_ACKED
;
3279 /* The "ultimate" congestion control function that aims to replace the rigid
3280 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3281 * It's called toward the end of processing an ACK with precise rate
3282 * information. All transmission or retransmission are delayed afterwards.
3284 static void tcp_cong_control(struct sock
*sk
, u32 ack
, u32 acked_sacked
,
3285 int flag
, const struct rate_sample
*rs
)
3287 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3289 if (icsk
->icsk_ca_ops
->cong_control
) {
3290 icsk
->icsk_ca_ops
->cong_control(sk
, rs
);
3294 if (tcp_in_cwnd_reduction(sk
)) {
3295 /* Reduce cwnd if state mandates */
3296 tcp_cwnd_reduction(sk
, acked_sacked
, flag
);
3297 } else if (tcp_may_raise_cwnd(sk
, flag
)) {
3298 /* Advance cwnd if state allows */
3299 tcp_cong_avoid(sk
, ack
, acked_sacked
);
3301 tcp_update_pacing_rate(sk
);
3304 /* Check that window update is acceptable.
3305 * The function assumes that snd_una<=ack<=snd_next.
3307 static inline bool tcp_may_update_window(const struct tcp_sock
*tp
,
3308 const u32 ack
, const u32 ack_seq
,
3311 return after(ack
, tp
->snd_una
) ||
3312 after(ack_seq
, tp
->snd_wl1
) ||
3313 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
);
3316 /* If we update tp->snd_una, also update tp->bytes_acked */
3317 static void tcp_snd_una_update(struct tcp_sock
*tp
, u32 ack
)
3319 u32 delta
= ack
- tp
->snd_una
;
3321 sock_owned_by_me((struct sock
*)tp
);
3322 tp
->bytes_acked
+= delta
;
3326 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3327 static void tcp_rcv_nxt_update(struct tcp_sock
*tp
, u32 seq
)
3329 u32 delta
= seq
- tp
->rcv_nxt
;
3331 sock_owned_by_me((struct sock
*)tp
);
3332 tp
->bytes_received
+= delta
;
3336 /* Update our send window.
3338 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3339 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3341 static int tcp_ack_update_window(struct sock
*sk
, const struct sk_buff
*skb
, u32 ack
,
3344 struct tcp_sock
*tp
= tcp_sk(sk
);
3346 u32 nwin
= ntohs(tcp_hdr(skb
)->window
);
3348 if (likely(!tcp_hdr(skb
)->syn
))
3349 nwin
<<= tp
->rx_opt
.snd_wscale
;
3351 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
3352 flag
|= FLAG_WIN_UPDATE
;
3353 tcp_update_wl(tp
, ack_seq
);
3355 if (tp
->snd_wnd
!= nwin
) {
3358 /* Note, it is the only place, where
3359 * fast path is recovered for sending TCP.
3362 tcp_fast_path_check(sk
);
3364 if (tcp_send_head(sk
))
3365 tcp_slow_start_after_idle_check(sk
);
3367 if (nwin
> tp
->max_window
) {
3368 tp
->max_window
= nwin
;
3369 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
3374 tcp_snd_una_update(tp
, ack
);
3379 static bool __tcp_oow_rate_limited(struct net
*net
, int mib_idx
,
3380 u32
*last_oow_ack_time
)
3382 if (*last_oow_ack_time
) {
3383 s32 elapsed
= (s32
)(tcp_time_stamp
- *last_oow_ack_time
);
3385 if (0 <= elapsed
&& elapsed
< sysctl_tcp_invalid_ratelimit
) {
3386 NET_INC_STATS(net
, mib_idx
);
3387 return true; /* rate-limited: don't send yet! */
3391 *last_oow_ack_time
= tcp_time_stamp
;
3393 return false; /* not rate-limited: go ahead, send dupack now! */
3396 /* Return true if we're currently rate-limiting out-of-window ACKs and
3397 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3398 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3399 * attacks that send repeated SYNs or ACKs for the same connection. To
3400 * do this, we do not send a duplicate SYNACK or ACK if the remote
3401 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3403 bool tcp_oow_rate_limited(struct net
*net
, const struct sk_buff
*skb
,
3404 int mib_idx
, u32
*last_oow_ack_time
)
3406 /* Data packets without SYNs are not likely part of an ACK loop. */
3407 if ((TCP_SKB_CB(skb
)->seq
!= TCP_SKB_CB(skb
)->end_seq
) &&
3411 return __tcp_oow_rate_limited(net
, mib_idx
, last_oow_ack_time
);
3414 /* RFC 5961 7 [ACK Throttling] */
3415 static void tcp_send_challenge_ack(struct sock
*sk
, const struct sk_buff
*skb
)
3417 /* unprotected vars, we dont care of overwrites */
3418 static u32 challenge_timestamp
;
3419 static unsigned int challenge_count
;
3420 struct tcp_sock
*tp
= tcp_sk(sk
);
3423 /* First check our per-socket dupack rate limit. */
3424 if (__tcp_oow_rate_limited(sock_net(sk
),
3425 LINUX_MIB_TCPACKSKIPPEDCHALLENGE
,
3426 &tp
->last_oow_ack_time
))
3429 /* Then check host-wide RFC 5961 rate limit. */
3431 if (now
!= challenge_timestamp
) {
3432 u32 half
= (sysctl_tcp_challenge_ack_limit
+ 1) >> 1;
3434 challenge_timestamp
= now
;
3435 WRITE_ONCE(challenge_count
, half
+
3436 prandom_u32_max(sysctl_tcp_challenge_ack_limit
));
3438 count
= READ_ONCE(challenge_count
);
3440 WRITE_ONCE(challenge_count
, count
- 1);
3441 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPCHALLENGEACK
);
3446 static void tcp_store_ts_recent(struct tcp_sock
*tp
)
3448 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
3449 tp
->rx_opt
.ts_recent_stamp
= get_seconds();
3452 static void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
3454 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
3455 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3456 * extra check below makes sure this can only happen
3457 * for pure ACK frames. -DaveM
3459 * Not only, also it occurs for expired timestamps.
3462 if (tcp_paws_check(&tp
->rx_opt
, 0))
3463 tcp_store_ts_recent(tp
);
3467 /* This routine deals with acks during a TLP episode.
3468 * We mark the end of a TLP episode on receiving TLP dupack or when
3469 * ack is after tlp_high_seq.
3470 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3472 static void tcp_process_tlp_ack(struct sock
*sk
, u32 ack
, int flag
)
3474 struct tcp_sock
*tp
= tcp_sk(sk
);
3476 if (before(ack
, tp
->tlp_high_seq
))
3479 if (flag
& FLAG_DSACKING_ACK
) {
3480 /* This DSACK means original and TLP probe arrived; no loss */
3481 tp
->tlp_high_seq
= 0;
3482 } else if (after(ack
, tp
->tlp_high_seq
)) {
3483 /* ACK advances: there was a loss, so reduce cwnd. Reset
3484 * tlp_high_seq in tcp_init_cwnd_reduction()
3486 tcp_init_cwnd_reduction(sk
);
3487 tcp_set_ca_state(sk
, TCP_CA_CWR
);
3488 tcp_end_cwnd_reduction(sk
);
3489 tcp_try_keep_open(sk
);
3490 NET_INC_STATS(sock_net(sk
),
3491 LINUX_MIB_TCPLOSSPROBERECOVERY
);
3492 } else if (!(flag
& (FLAG_SND_UNA_ADVANCED
|
3493 FLAG_NOT_DUP
| FLAG_DATA_SACKED
))) {
3494 /* Pure dupack: original and TLP probe arrived; no loss */
3495 tp
->tlp_high_seq
= 0;
3499 static inline void tcp_in_ack_event(struct sock
*sk
, u32 flags
)
3501 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3503 if (icsk
->icsk_ca_ops
->in_ack_event
)
3504 icsk
->icsk_ca_ops
->in_ack_event(sk
, flags
);
3507 /* Congestion control has updated the cwnd already. So if we're in
3508 * loss recovery then now we do any new sends (for FRTO) or
3509 * retransmits (for CA_Loss or CA_recovery) that make sense.
3511 static void tcp_xmit_recovery(struct sock
*sk
, int rexmit
)
3513 struct tcp_sock
*tp
= tcp_sk(sk
);
3515 if (rexmit
== REXMIT_NONE
)
3518 if (unlikely(rexmit
== 2)) {
3519 __tcp_push_pending_frames(sk
, tcp_current_mss(sk
),
3521 if (after(tp
->snd_nxt
, tp
->high_seq
))
3525 tcp_xmit_retransmit_queue(sk
);
3528 /* This routine deals with incoming acks, but not outgoing ones. */
3529 static int tcp_ack(struct sock
*sk
, const struct sk_buff
*skb
, int flag
)
3531 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3532 struct tcp_sock
*tp
= tcp_sk(sk
);
3533 struct tcp_sacktag_state sack_state
;
3534 struct rate_sample rs
= { .prior_delivered
= 0 };
3535 u32 prior_snd_una
= tp
->snd_una
;
3536 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
3537 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3538 bool is_dupack
= false;
3540 int prior_packets
= tp
->packets_out
;
3541 u32 delivered
= tp
->delivered
;
3542 u32 lost
= tp
->lost
;
3543 int acked
= 0; /* Number of packets newly acked */
3544 int rexmit
= REXMIT_NONE
; /* Flag to (re)transmit to recover losses */
3546 sack_state
.first_sackt
.v64
= 0;
3547 sack_state
.rate
= &rs
;
3549 /* We very likely will need to access write queue head. */
3550 prefetchw(sk
->sk_write_queue
.next
);
3552 /* If the ack is older than previous acks
3553 * then we can probably ignore it.
3555 if (before(ack
, prior_snd_una
)) {
3556 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3557 if (before(ack
, prior_snd_una
- tp
->max_window
)) {
3558 tcp_send_challenge_ack(sk
, skb
);
3564 /* If the ack includes data we haven't sent yet, discard
3565 * this segment (RFC793 Section 3.9).
3567 if (after(ack
, tp
->snd_nxt
))
3570 skb_mstamp_get(&sack_state
.ack_time
);
3572 if (icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
)
3575 if (after(ack
, prior_snd_una
)) {
3576 flag
|= FLAG_SND_UNA_ADVANCED
;
3577 icsk
->icsk_retransmits
= 0;
3580 prior_fackets
= tp
->fackets_out
;
3581 rs
.prior_in_flight
= tcp_packets_in_flight(tp
);
3583 /* ts_recent update must be made after we are sure that the packet
3586 if (flag
& FLAG_UPDATE_TS_RECENT
)
3587 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
3589 if (!(flag
& FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
3590 /* Window is constant, pure forward advance.
3591 * No more checks are required.
3592 * Note, we use the fact that SND.UNA>=SND.WL2.
3594 tcp_update_wl(tp
, ack_seq
);
3595 tcp_snd_una_update(tp
, ack
);
3596 flag
|= FLAG_WIN_UPDATE
;
3598 tcp_in_ack_event(sk
, CA_ACK_WIN_UPDATE
);
3600 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPHPACKS
);
3602 u32 ack_ev_flags
= CA_ACK_SLOWPATH
;
3604 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
3607 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPPUREACKS
);
3609 flag
|= tcp_ack_update_window(sk
, skb
, ack
, ack_seq
);
3611 if (TCP_SKB_CB(skb
)->sacked
)
3612 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
,
3615 if (tcp_ecn_rcv_ecn_echo(tp
, tcp_hdr(skb
))) {
3617 ack_ev_flags
|= CA_ACK_ECE
;
3620 if (flag
& FLAG_WIN_UPDATE
)
3621 ack_ev_flags
|= CA_ACK_WIN_UPDATE
;
3623 tcp_in_ack_event(sk
, ack_ev_flags
);
3626 /* We passed data and got it acked, remove any soft error
3627 * log. Something worked...
3629 sk
->sk_err_soft
= 0;
3630 icsk
->icsk_probes_out
= 0;
3631 tp
->rcv_tstamp
= tcp_time_stamp
;
3635 /* See if we can take anything off of the retransmit queue. */
3636 flag
|= tcp_clean_rtx_queue(sk
, prior_fackets
, prior_snd_una
, &acked
,
3639 if (tcp_ack_is_dubious(sk
, flag
)) {
3640 is_dupack
= !(flag
& (FLAG_SND_UNA_ADVANCED
| FLAG_NOT_DUP
));
3641 tcp_fastretrans_alert(sk
, acked
, is_dupack
, &flag
, &rexmit
,
3642 &sack_state
.ack_time
);
3644 if (tp
->tlp_high_seq
)
3645 tcp_process_tlp_ack(sk
, ack
, flag
);
3647 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
& FLAG_NOT_DUP
))
3650 if (icsk
->icsk_pending
== ICSK_TIME_RETRANS
)
3651 tcp_schedule_loss_probe(sk
);
3652 delivered
= tp
->delivered
- delivered
; /* freshly ACKed or SACKed */
3653 lost
= tp
->lost
- lost
; /* freshly marked lost */
3654 tcp_rate_gen(sk
, delivered
, lost
, &sack_state
.ack_time
,
3656 tcp_cong_control(sk
, ack
, delivered
, flag
, sack_state
.rate
);
3657 tcp_xmit_recovery(sk
, rexmit
);
3661 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3662 if (flag
& FLAG_DSACKING_ACK
)
3663 tcp_fastretrans_alert(sk
, acked
, is_dupack
, &flag
, &rexmit
,
3664 &sack_state
.ack_time
);
3665 /* If this ack opens up a zero window, clear backoff. It was
3666 * being used to time the probes, and is probably far higher than
3667 * it needs to be for normal retransmission.
3669 if (tcp_send_head(sk
))
3672 if (tp
->tlp_high_seq
)
3673 tcp_process_tlp_ack(sk
, ack
, flag
);
3677 SOCK_DEBUG(sk
, "Ack %u after %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3681 /* If data was SACKed, tag it and see if we should send more data.
3682 * If data was DSACKed, see if we can undo a cwnd reduction.
3684 if (TCP_SKB_CB(skb
)->sacked
) {
3685 skb_mstamp_get(&sack_state
.ack_time
);
3686 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
,
3688 tcp_fastretrans_alert(sk
, acked
, is_dupack
, &flag
, &rexmit
,
3689 &sack_state
.ack_time
);
3690 tcp_xmit_recovery(sk
, rexmit
);
3693 SOCK_DEBUG(sk
, "Ack %u before %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3697 static void tcp_parse_fastopen_option(int len
, const unsigned char *cookie
,
3698 bool syn
, struct tcp_fastopen_cookie
*foc
,
3701 /* Valid only in SYN or SYN-ACK with an even length. */
3702 if (!foc
|| !syn
|| len
< 0 || (len
& 1))
3705 if (len
>= TCP_FASTOPEN_COOKIE_MIN
&&
3706 len
<= TCP_FASTOPEN_COOKIE_MAX
)
3707 memcpy(foc
->val
, cookie
, len
);
3714 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3715 * But, this can also be called on packets in the established flow when
3716 * the fast version below fails.
3718 void tcp_parse_options(const struct sk_buff
*skb
,
3719 struct tcp_options_received
*opt_rx
, int estab
,
3720 struct tcp_fastopen_cookie
*foc
)
3722 const unsigned char *ptr
;
3723 const struct tcphdr
*th
= tcp_hdr(skb
);
3724 int length
= (th
->doff
* 4) - sizeof(struct tcphdr
);
3726 ptr
= (const unsigned char *)(th
+ 1);
3727 opt_rx
->saw_tstamp
= 0;
3729 while (length
> 0) {
3730 int opcode
= *ptr
++;
3736 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
3741 if (opsize
< 2) /* "silly options" */
3743 if (opsize
> length
)
3744 return; /* don't parse partial options */
3747 if (opsize
== TCPOLEN_MSS
&& th
->syn
&& !estab
) {
3748 u16 in_mss
= get_unaligned_be16(ptr
);
3750 if (opt_rx
->user_mss
&&
3751 opt_rx
->user_mss
< in_mss
)
3752 in_mss
= opt_rx
->user_mss
;
3753 opt_rx
->mss_clamp
= in_mss
;
3758 if (opsize
== TCPOLEN_WINDOW
&& th
->syn
&&
3759 !estab
&& sysctl_tcp_window_scaling
) {
3760 __u8 snd_wscale
= *(__u8
*)ptr
;
3761 opt_rx
->wscale_ok
= 1;
3762 if (snd_wscale
> 14) {
3763 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3768 opt_rx
->snd_wscale
= snd_wscale
;
3771 case TCPOPT_TIMESTAMP
:
3772 if ((opsize
== TCPOLEN_TIMESTAMP
) &&
3773 ((estab
&& opt_rx
->tstamp_ok
) ||
3774 (!estab
&& sysctl_tcp_timestamps
))) {
3775 opt_rx
->saw_tstamp
= 1;
3776 opt_rx
->rcv_tsval
= get_unaligned_be32(ptr
);
3777 opt_rx
->rcv_tsecr
= get_unaligned_be32(ptr
+ 4);
3780 case TCPOPT_SACK_PERM
:
3781 if (opsize
== TCPOLEN_SACK_PERM
&& th
->syn
&&
3782 !estab
&& sysctl_tcp_sack
) {
3783 opt_rx
->sack_ok
= TCP_SACK_SEEN
;
3784 tcp_sack_reset(opt_rx
);
3789 if ((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
3790 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
3792 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
3795 #ifdef CONFIG_TCP_MD5SIG
3798 * The MD5 Hash has already been
3799 * checked (see tcp_v{4,6}_do_rcv()).
3803 case TCPOPT_FASTOPEN
:
3804 tcp_parse_fastopen_option(
3805 opsize
- TCPOLEN_FASTOPEN_BASE
,
3806 ptr
, th
->syn
, foc
, false);
3810 /* Fast Open option shares code 254 using a
3811 * 16 bits magic number.
3813 if (opsize
>= TCPOLEN_EXP_FASTOPEN_BASE
&&
3814 get_unaligned_be16(ptr
) ==
3815 TCPOPT_FASTOPEN_MAGIC
)
3816 tcp_parse_fastopen_option(opsize
-
3817 TCPOLEN_EXP_FASTOPEN_BASE
,
3818 ptr
+ 2, th
->syn
, foc
, true);
3827 EXPORT_SYMBOL(tcp_parse_options
);
3829 static bool tcp_parse_aligned_timestamp(struct tcp_sock
*tp
, const struct tcphdr
*th
)
3831 const __be32
*ptr
= (const __be32
*)(th
+ 1);
3833 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3834 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
3835 tp
->rx_opt
.saw_tstamp
= 1;
3837 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3840 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
) - tp
->tsoffset
;
3842 tp
->rx_opt
.rcv_tsecr
= 0;
3848 /* Fast parse options. This hopes to only see timestamps.
3849 * If it is wrong it falls back on tcp_parse_options().
3851 static bool tcp_fast_parse_options(const struct sk_buff
*skb
,
3852 const struct tcphdr
*th
, struct tcp_sock
*tp
)
3854 /* In the spirit of fast parsing, compare doff directly to constant
3855 * values. Because equality is used, short doff can be ignored here.
3857 if (th
->doff
== (sizeof(*th
) / 4)) {
3858 tp
->rx_opt
.saw_tstamp
= 0;
3860 } else if (tp
->rx_opt
.tstamp_ok
&&
3861 th
->doff
== ((sizeof(*th
) + TCPOLEN_TSTAMP_ALIGNED
) / 4)) {
3862 if (tcp_parse_aligned_timestamp(tp
, th
))
3866 tcp_parse_options(skb
, &tp
->rx_opt
, 1, NULL
);
3867 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
3868 tp
->rx_opt
.rcv_tsecr
-= tp
->tsoffset
;
3873 #ifdef CONFIG_TCP_MD5SIG
3875 * Parse MD5 Signature option
3877 const u8
*tcp_parse_md5sig_option(const struct tcphdr
*th
)
3879 int length
= (th
->doff
<< 2) - sizeof(*th
);
3880 const u8
*ptr
= (const u8
*)(th
+ 1);
3882 /* If the TCP option is too short, we can short cut */
3883 if (length
< TCPOLEN_MD5SIG
)
3886 while (length
> 0) {
3887 int opcode
= *ptr
++;
3898 if (opsize
< 2 || opsize
> length
)
3900 if (opcode
== TCPOPT_MD5SIG
)
3901 return opsize
== TCPOLEN_MD5SIG
? ptr
: NULL
;
3908 EXPORT_SYMBOL(tcp_parse_md5sig_option
);
3911 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3913 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3914 * it can pass through stack. So, the following predicate verifies that
3915 * this segment is not used for anything but congestion avoidance or
3916 * fast retransmit. Moreover, we even are able to eliminate most of such
3917 * second order effects, if we apply some small "replay" window (~RTO)
3918 * to timestamp space.
3920 * All these measures still do not guarantee that we reject wrapped ACKs
3921 * on networks with high bandwidth, when sequence space is recycled fastly,
3922 * but it guarantees that such events will be very rare and do not affect
3923 * connection seriously. This doesn't look nice, but alas, PAWS is really
3926 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3927 * states that events when retransmit arrives after original data are rare.
3928 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3929 * the biggest problem on large power networks even with minor reordering.
3930 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3931 * up to bandwidth of 18Gigabit/sec. 8) ]
3934 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
3936 const struct tcp_sock
*tp
= tcp_sk(sk
);
3937 const struct tcphdr
*th
= tcp_hdr(skb
);
3938 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3939 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3941 return (/* 1. Pure ACK with correct sequence number. */
3942 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
3944 /* 2. ... and duplicate ACK. */
3945 ack
== tp
->snd_una
&&
3947 /* 3. ... and does not update window. */
3948 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
3950 /* 4. ... and sits in replay window. */
3951 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
3954 static inline bool tcp_paws_discard(const struct sock
*sk
,
3955 const struct sk_buff
*skb
)
3957 const struct tcp_sock
*tp
= tcp_sk(sk
);
3959 return !tcp_paws_check(&tp
->rx_opt
, TCP_PAWS_WINDOW
) &&
3960 !tcp_disordered_ack(sk
, skb
);
3963 /* Check segment sequence number for validity.
3965 * Segment controls are considered valid, if the segment
3966 * fits to the window after truncation to the window. Acceptability
3967 * of data (and SYN, FIN, of course) is checked separately.
3968 * See tcp_data_queue(), for example.
3970 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3971 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3972 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3973 * (borrowed from freebsd)
3976 static inline bool tcp_sequence(const struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3978 return !before(end_seq
, tp
->rcv_wup
) &&
3979 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
3982 /* When we get a reset we do this. */
3983 void tcp_reset(struct sock
*sk
)
3985 /* We want the right error as BSD sees it (and indeed as we do). */
3986 switch (sk
->sk_state
) {
3988 sk
->sk_err
= ECONNREFUSED
;
3990 case TCP_CLOSE_WAIT
:
3996 sk
->sk_err
= ECONNRESET
;
3998 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4001 if (!sock_flag(sk
, SOCK_DEAD
))
4002 sk
->sk_error_report(sk
);
4008 * Process the FIN bit. This now behaves as it is supposed to work
4009 * and the FIN takes effect when it is validly part of sequence
4010 * space. Not before when we get holes.
4012 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4013 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4016 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4017 * close and we go into CLOSING (and later onto TIME-WAIT)
4019 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4021 void tcp_fin(struct sock
*sk
)
4023 struct tcp_sock
*tp
= tcp_sk(sk
);
4025 inet_csk_schedule_ack(sk
);
4027 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
4028 sock_set_flag(sk
, SOCK_DONE
);
4030 switch (sk
->sk_state
) {
4032 case TCP_ESTABLISHED
:
4033 /* Move to CLOSE_WAIT */
4034 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
4035 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
4038 case TCP_CLOSE_WAIT
:
4040 /* Received a retransmission of the FIN, do
4045 /* RFC793: Remain in the LAST-ACK state. */
4049 /* This case occurs when a simultaneous close
4050 * happens, we must ack the received FIN and
4051 * enter the CLOSING state.
4054 tcp_set_state(sk
, TCP_CLOSING
);
4057 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4059 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
4062 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4063 * cases we should never reach this piece of code.
4065 pr_err("%s: Impossible, sk->sk_state=%d\n",
4066 __func__
, sk
->sk_state
);
4070 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4071 * Probably, we should reset in this case. For now drop them.
4073 skb_rbtree_purge(&tp
->out_of_order_queue
);
4074 if (tcp_is_sack(tp
))
4075 tcp_sack_reset(&tp
->rx_opt
);
4078 if (!sock_flag(sk
, SOCK_DEAD
)) {
4079 sk
->sk_state_change(sk
);
4081 /* Do not send POLL_HUP for half duplex close. */
4082 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
4083 sk
->sk_state
== TCP_CLOSE
)
4084 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_HUP
);
4086 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_IN
);
4090 static inline bool tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
,
4093 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
4094 if (before(seq
, sp
->start_seq
))
4095 sp
->start_seq
= seq
;
4096 if (after(end_seq
, sp
->end_seq
))
4097 sp
->end_seq
= end_seq
;
4103 static void tcp_dsack_set(struct sock
*sk
, u32 seq
, u32 end_seq
)
4105 struct tcp_sock
*tp
= tcp_sk(sk
);
4107 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
4110 if (before(seq
, tp
->rcv_nxt
))
4111 mib_idx
= LINUX_MIB_TCPDSACKOLDSENT
;
4113 mib_idx
= LINUX_MIB_TCPDSACKOFOSENT
;
4115 NET_INC_STATS(sock_net(sk
), mib_idx
);
4117 tp
->rx_opt
.dsack
= 1;
4118 tp
->duplicate_sack
[0].start_seq
= seq
;
4119 tp
->duplicate_sack
[0].end_seq
= end_seq
;
4123 static void tcp_dsack_extend(struct sock
*sk
, u32 seq
, u32 end_seq
)
4125 struct tcp_sock
*tp
= tcp_sk(sk
);
4127 if (!tp
->rx_opt
.dsack
)
4128 tcp_dsack_set(sk
, seq
, end_seq
);
4130 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
4133 static void tcp_send_dupack(struct sock
*sk
, const struct sk_buff
*skb
)
4135 struct tcp_sock
*tp
= tcp_sk(sk
);
4137 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
4138 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4139 NET_INC_STATS(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4140 tcp_enter_quickack_mode(sk
);
4142 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
4143 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4145 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
4146 end_seq
= tp
->rcv_nxt
;
4147 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, end_seq
);
4154 /* These routines update the SACK block as out-of-order packets arrive or
4155 * in-order packets close up the sequence space.
4157 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
4160 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4161 struct tcp_sack_block
*swalk
= sp
+ 1;
4163 /* See if the recent change to the first SACK eats into
4164 * or hits the sequence space of other SACK blocks, if so coalesce.
4166 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
;) {
4167 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
4170 /* Zap SWALK, by moving every further SACK up by one slot.
4171 * Decrease num_sacks.
4173 tp
->rx_opt
.num_sacks
--;
4174 for (i
= this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
4178 this_sack
++, swalk
++;
4182 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
4184 struct tcp_sock
*tp
= tcp_sk(sk
);
4185 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4186 int cur_sacks
= tp
->rx_opt
.num_sacks
;
4192 for (this_sack
= 0; this_sack
< cur_sacks
; this_sack
++, sp
++) {
4193 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
4194 /* Rotate this_sack to the first one. */
4195 for (; this_sack
> 0; this_sack
--, sp
--)
4196 swap(*sp
, *(sp
- 1));
4198 tcp_sack_maybe_coalesce(tp
);
4203 /* Could not find an adjacent existing SACK, build a new one,
4204 * put it at the front, and shift everyone else down. We
4205 * always know there is at least one SACK present already here.
4207 * If the sack array is full, forget about the last one.
4209 if (this_sack
>= TCP_NUM_SACKS
) {
4211 tp
->rx_opt
.num_sacks
--;
4214 for (; this_sack
> 0; this_sack
--, sp
--)
4218 /* Build the new head SACK, and we're done. */
4219 sp
->start_seq
= seq
;
4220 sp
->end_seq
= end_seq
;
4221 tp
->rx_opt
.num_sacks
++;
4224 /* RCV.NXT advances, some SACKs should be eaten. */
4226 static void tcp_sack_remove(struct tcp_sock
*tp
)
4228 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4229 int num_sacks
= tp
->rx_opt
.num_sacks
;
4232 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4233 if (RB_EMPTY_ROOT(&tp
->out_of_order_queue
)) {
4234 tp
->rx_opt
.num_sacks
= 0;
4238 for (this_sack
= 0; this_sack
< num_sacks
;) {
4239 /* Check if the start of the sack is covered by RCV.NXT. */
4240 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
4243 /* RCV.NXT must cover all the block! */
4244 WARN_ON(before(tp
->rcv_nxt
, sp
->end_seq
));
4246 /* Zap this SACK, by moving forward any other SACKS. */
4247 for (i
= this_sack
+1; i
< num_sacks
; i
++)
4248 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
4255 tp
->rx_opt
.num_sacks
= num_sacks
;
4259 * tcp_try_coalesce - try to merge skb to prior one
4262 * @from: buffer to add in queue
4263 * @fragstolen: pointer to boolean
4265 * Before queueing skb @from after @to, try to merge them
4266 * to reduce overall memory use and queue lengths, if cost is small.
4267 * Packets in ofo or receive queues can stay a long time.
4268 * Better try to coalesce them right now to avoid future collapses.
4269 * Returns true if caller should free @from instead of queueing it
4271 static bool tcp_try_coalesce(struct sock
*sk
,
4273 struct sk_buff
*from
,
4278 *fragstolen
= false;
4280 /* Its possible this segment overlaps with prior segment in queue */
4281 if (TCP_SKB_CB(from
)->seq
!= TCP_SKB_CB(to
)->end_seq
)
4284 if (!skb_try_coalesce(to
, from
, fragstolen
, &delta
))
4287 atomic_add(delta
, &sk
->sk_rmem_alloc
);
4288 sk_mem_charge(sk
, delta
);
4289 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPRCVCOALESCE
);
4290 TCP_SKB_CB(to
)->end_seq
= TCP_SKB_CB(from
)->end_seq
;
4291 TCP_SKB_CB(to
)->ack_seq
= TCP_SKB_CB(from
)->ack_seq
;
4292 TCP_SKB_CB(to
)->tcp_flags
|= TCP_SKB_CB(from
)->tcp_flags
;
4296 static void tcp_drop(struct sock
*sk
, struct sk_buff
*skb
)
4298 sk_drops_add(sk
, skb
);
4302 /* This one checks to see if we can put data from the
4303 * out_of_order queue into the receive_queue.
4305 static void tcp_ofo_queue(struct sock
*sk
)
4307 struct tcp_sock
*tp
= tcp_sk(sk
);
4308 __u32 dsack_high
= tp
->rcv_nxt
;
4309 bool fin
, fragstolen
, eaten
;
4310 struct sk_buff
*skb
, *tail
;
4313 p
= rb_first(&tp
->out_of_order_queue
);
4315 skb
= rb_entry(p
, struct sk_buff
, rbnode
);
4316 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
4319 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
4320 __u32 dsack
= dsack_high
;
4321 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
4322 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
4323 tcp_dsack_extend(sk
, TCP_SKB_CB(skb
)->seq
, dsack
);
4326 rb_erase(&skb
->rbnode
, &tp
->out_of_order_queue
);
4328 if (unlikely(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))) {
4329 SOCK_DEBUG(sk
, "ofo packet was already received\n");
4333 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
4334 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4335 TCP_SKB_CB(skb
)->end_seq
);
4337 tail
= skb_peek_tail(&sk
->sk_receive_queue
);
4338 eaten
= tail
&& tcp_try_coalesce(sk
, tail
, skb
, &fragstolen
);
4339 tcp_rcv_nxt_update(tp
, TCP_SKB_CB(skb
)->end_seq
);
4340 fin
= TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
;
4342 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4344 kfree_skb_partial(skb
, fragstolen
);
4346 if (unlikely(fin
)) {
4348 /* tcp_fin() purges tp->out_of_order_queue,
4349 * so we must end this loop right now.
4356 static bool tcp_prune_ofo_queue(struct sock
*sk
);
4357 static int tcp_prune_queue(struct sock
*sk
);
4359 static int tcp_try_rmem_schedule(struct sock
*sk
, struct sk_buff
*skb
,
4362 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
4363 !sk_rmem_schedule(sk
, skb
, size
)) {
4365 if (tcp_prune_queue(sk
) < 0)
4368 while (!sk_rmem_schedule(sk
, skb
, size
)) {
4369 if (!tcp_prune_ofo_queue(sk
))
4376 static void tcp_data_queue_ofo(struct sock
*sk
, struct sk_buff
*skb
)
4378 struct tcp_sock
*tp
= tcp_sk(sk
);
4379 struct rb_node
**p
, *q
, *parent
;
4380 struct sk_buff
*skb1
;
4384 tcp_ecn_check_ce(tp
, skb
);
4386 if (unlikely(tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))) {
4387 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPOFODROP
);
4392 /* Disable header prediction. */
4394 inet_csk_schedule_ack(sk
);
4396 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPOFOQUEUE
);
4397 seq
= TCP_SKB_CB(skb
)->seq
;
4398 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4399 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
4400 tp
->rcv_nxt
, seq
, end_seq
);
4402 p
= &tp
->out_of_order_queue
.rb_node
;
4403 if (RB_EMPTY_ROOT(&tp
->out_of_order_queue
)) {
4404 /* Initial out of order segment, build 1 SACK. */
4405 if (tcp_is_sack(tp
)) {
4406 tp
->rx_opt
.num_sacks
= 1;
4407 tp
->selective_acks
[0].start_seq
= seq
;
4408 tp
->selective_acks
[0].end_seq
= end_seq
;
4410 rb_link_node(&skb
->rbnode
, NULL
, p
);
4411 rb_insert_color(&skb
->rbnode
, &tp
->out_of_order_queue
);
4412 tp
->ooo_last_skb
= skb
;
4416 /* In the typical case, we are adding an skb to the end of the list.
4417 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4419 if (tcp_try_coalesce(sk
, tp
->ooo_last_skb
, skb
, &fragstolen
)) {
4421 tcp_grow_window(sk
, skb
);
4422 kfree_skb_partial(skb
, fragstolen
);
4426 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4427 if (!before(seq
, TCP_SKB_CB(tp
->ooo_last_skb
)->end_seq
)) {
4428 parent
= &tp
->ooo_last_skb
->rbnode
;
4429 p
= &parent
->rb_right
;
4433 /* Find place to insert this segment. Handle overlaps on the way. */
4437 skb1
= rb_entry(parent
, struct sk_buff
, rbnode
);
4438 if (before(seq
, TCP_SKB_CB(skb1
)->seq
)) {
4439 p
= &parent
->rb_left
;
4442 if (before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4443 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4444 /* All the bits are present. Drop. */
4445 NET_INC_STATS(sock_net(sk
),
4446 LINUX_MIB_TCPOFOMERGE
);
4449 tcp_dsack_set(sk
, seq
, end_seq
);
4452 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
4453 /* Partial overlap. */
4454 tcp_dsack_set(sk
, seq
, TCP_SKB_CB(skb1
)->end_seq
);
4456 /* skb's seq == skb1's seq and skb covers skb1.
4457 * Replace skb1 with skb.
4459 rb_replace_node(&skb1
->rbnode
, &skb
->rbnode
,
4460 &tp
->out_of_order_queue
);
4461 tcp_dsack_extend(sk
,
4462 TCP_SKB_CB(skb1
)->seq
,
4463 TCP_SKB_CB(skb1
)->end_seq
);
4464 NET_INC_STATS(sock_net(sk
),
4465 LINUX_MIB_TCPOFOMERGE
);
4469 } else if (tcp_try_coalesce(sk
, skb1
, skb
, &fragstolen
)) {
4472 p
= &parent
->rb_right
;
4475 /* Insert segment into RB tree. */
4476 rb_link_node(&skb
->rbnode
, parent
, p
);
4477 rb_insert_color(&skb
->rbnode
, &tp
->out_of_order_queue
);
4480 /* Remove other segments covered by skb. */
4481 while ((q
= rb_next(&skb
->rbnode
)) != NULL
) {
4482 skb1
= rb_entry(q
, struct sk_buff
, rbnode
);
4484 if (!after(end_seq
, TCP_SKB_CB(skb1
)->seq
))
4486 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4487 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4491 rb_erase(&skb1
->rbnode
, &tp
->out_of_order_queue
);
4492 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4493 TCP_SKB_CB(skb1
)->end_seq
);
4494 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPOFOMERGE
);
4497 /* If there is no skb after us, we are the last_skb ! */
4499 tp
->ooo_last_skb
= skb
;
4502 if (tcp_is_sack(tp
))
4503 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
4506 tcp_grow_window(sk
, skb
);
4508 skb_set_owner_r(skb
, sk
);
4512 static int __must_check
tcp_queue_rcv(struct sock
*sk
, struct sk_buff
*skb
, int hdrlen
,
4516 struct sk_buff
*tail
= skb_peek_tail(&sk
->sk_receive_queue
);
4518 __skb_pull(skb
, hdrlen
);
4520 tcp_try_coalesce(sk
, tail
, skb
, fragstolen
)) ? 1 : 0;
4521 tcp_rcv_nxt_update(tcp_sk(sk
), TCP_SKB_CB(skb
)->end_seq
);
4523 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4524 skb_set_owner_r(skb
, sk
);
4529 int tcp_send_rcvq(struct sock
*sk
, struct msghdr
*msg
, size_t size
)
4531 struct sk_buff
*skb
;
4539 if (size
> PAGE_SIZE
) {
4540 int npages
= min_t(size_t, size
>> PAGE_SHIFT
, MAX_SKB_FRAGS
);
4542 data_len
= npages
<< PAGE_SHIFT
;
4543 size
= data_len
+ (size
& ~PAGE_MASK
);
4545 skb
= alloc_skb_with_frags(size
- data_len
, data_len
,
4546 PAGE_ALLOC_COSTLY_ORDER
,
4547 &err
, sk
->sk_allocation
);
4551 skb_put(skb
, size
- data_len
);
4552 skb
->data_len
= data_len
;
4555 if (tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))
4558 err
= skb_copy_datagram_from_iter(skb
, 0, &msg
->msg_iter
, size
);
4562 TCP_SKB_CB(skb
)->seq
= tcp_sk(sk
)->rcv_nxt
;
4563 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ size
;
4564 TCP_SKB_CB(skb
)->ack_seq
= tcp_sk(sk
)->snd_una
- 1;
4566 if (tcp_queue_rcv(sk
, skb
, 0, &fragstolen
)) {
4567 WARN_ON_ONCE(fragstolen
); /* should not happen */
4579 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
4581 struct tcp_sock
*tp
= tcp_sk(sk
);
4582 bool fragstolen
= false;
4585 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
) {
4590 __skb_pull(skb
, tcp_hdr(skb
)->doff
* 4);
4592 tcp_ecn_accept_cwr(tp
, skb
);
4594 tp
->rx_opt
.dsack
= 0;
4596 /* Queue data for delivery to the user.
4597 * Packets in sequence go to the receive queue.
4598 * Out of sequence packets to the out_of_order_queue.
4600 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
4601 if (tcp_receive_window(tp
) == 0)
4604 /* Ok. In sequence. In window. */
4605 if (tp
->ucopy
.task
== current
&&
4606 tp
->copied_seq
== tp
->rcv_nxt
&& tp
->ucopy
.len
&&
4607 sock_owned_by_user(sk
) && !tp
->urg_data
) {
4608 int chunk
= min_t(unsigned int, skb
->len
,
4611 __set_current_state(TASK_RUNNING
);
4613 if (!skb_copy_datagram_msg(skb
, 0, tp
->ucopy
.msg
, chunk
)) {
4614 tp
->ucopy
.len
-= chunk
;
4615 tp
->copied_seq
+= chunk
;
4616 eaten
= (chunk
== skb
->len
);
4617 tcp_rcv_space_adjust(sk
);
4624 if (skb_queue_len(&sk
->sk_receive_queue
) == 0)
4625 sk_forced_mem_schedule(sk
, skb
->truesize
);
4626 else if (tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))
4629 eaten
= tcp_queue_rcv(sk
, skb
, 0, &fragstolen
);
4631 tcp_rcv_nxt_update(tp
, TCP_SKB_CB(skb
)->end_seq
);
4633 tcp_event_data_recv(sk
, skb
);
4634 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
4637 if (!RB_EMPTY_ROOT(&tp
->out_of_order_queue
)) {
4640 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4641 * gap in queue is filled.
4643 if (RB_EMPTY_ROOT(&tp
->out_of_order_queue
))
4644 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
4647 if (tp
->rx_opt
.num_sacks
)
4648 tcp_sack_remove(tp
);
4650 tcp_fast_path_check(sk
);
4653 kfree_skb_partial(skb
, fragstolen
);
4654 if (!sock_flag(sk
, SOCK_DEAD
))
4655 sk
->sk_data_ready(sk
);
4659 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4660 /* A retransmit, 2nd most common case. Force an immediate ack. */
4661 NET_INC_STATS(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4662 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4665 tcp_enter_quickack_mode(sk
);
4666 inet_csk_schedule_ack(sk
);
4672 /* Out of window. F.e. zero window probe. */
4673 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
4676 tcp_enter_quickack_mode(sk
);
4678 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4679 /* Partial packet, seq < rcv_next < end_seq */
4680 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
4681 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4682 TCP_SKB_CB(skb
)->end_seq
);
4684 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
4686 /* If window is closed, drop tail of packet. But after
4687 * remembering D-SACK for its head made in previous line.
4689 if (!tcp_receive_window(tp
))
4694 tcp_data_queue_ofo(sk
, skb
);
4697 static struct sk_buff
*tcp_skb_next(struct sk_buff
*skb
, struct sk_buff_head
*list
)
4700 return !skb_queue_is_last(list
, skb
) ? skb
->next
: NULL
;
4702 return rb_entry_safe(rb_next(&skb
->rbnode
), struct sk_buff
, rbnode
);
4705 static struct sk_buff
*tcp_collapse_one(struct sock
*sk
, struct sk_buff
*skb
,
4706 struct sk_buff_head
*list
,
4707 struct rb_root
*root
)
4709 struct sk_buff
*next
= tcp_skb_next(skb
, list
);
4712 __skb_unlink(skb
, list
);
4714 rb_erase(&skb
->rbnode
, root
);
4717 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPRCVCOLLAPSED
);
4722 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
4723 static void tcp_rbtree_insert(struct rb_root
*root
, struct sk_buff
*skb
)
4725 struct rb_node
**p
= &root
->rb_node
;
4726 struct rb_node
*parent
= NULL
;
4727 struct sk_buff
*skb1
;
4731 skb1
= rb_entry(parent
, struct sk_buff
, rbnode
);
4732 if (before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb1
)->seq
))
4733 p
= &parent
->rb_left
;
4735 p
= &parent
->rb_right
;
4737 rb_link_node(&skb
->rbnode
, parent
, p
);
4738 rb_insert_color(&skb
->rbnode
, root
);
4741 /* Collapse contiguous sequence of skbs head..tail with
4742 * sequence numbers start..end.
4744 * If tail is NULL, this means until the end of the queue.
4746 * Segments with FIN/SYN are not collapsed (only because this
4750 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
, struct rb_root
*root
,
4751 struct sk_buff
*head
, struct sk_buff
*tail
, u32 start
, u32 end
)
4753 struct sk_buff
*skb
= head
, *n
;
4754 struct sk_buff_head tmp
;
4757 /* First, check that queue is collapsible and find
4758 * the point where collapsing can be useful.
4761 for (end_of_skbs
= true; skb
!= NULL
&& skb
!= tail
; skb
= n
) {
4762 n
= tcp_skb_next(skb
, list
);
4764 /* No new bits? It is possible on ofo queue. */
4765 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4766 skb
= tcp_collapse_one(sk
, skb
, list
, root
);
4772 /* The first skb to collapse is:
4774 * - bloated or contains data before "start" or
4775 * overlaps to the next one.
4777 if (!(TCP_SKB_CB(skb
)->tcp_flags
& (TCPHDR_SYN
| TCPHDR_FIN
)) &&
4778 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
4779 before(TCP_SKB_CB(skb
)->seq
, start
))) {
4780 end_of_skbs
= false;
4784 if (n
&& n
!= tail
&&
4785 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(n
)->seq
) {
4786 end_of_skbs
= false;
4790 /* Decided to skip this, advance start seq. */
4791 start
= TCP_SKB_CB(skb
)->end_seq
;
4794 (TCP_SKB_CB(skb
)->tcp_flags
& (TCPHDR_SYN
| TCPHDR_FIN
)))
4797 __skb_queue_head_init(&tmp
);
4799 while (before(start
, end
)) {
4800 int copy
= min_t(int, SKB_MAX_ORDER(0, 0), end
- start
);
4801 struct sk_buff
*nskb
;
4803 nskb
= alloc_skb(copy
, GFP_ATOMIC
);
4807 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
4808 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
4810 __skb_queue_before(list
, skb
, nskb
);
4812 __skb_queue_tail(&tmp
, nskb
); /* defer rbtree insertion */
4813 skb_set_owner_r(nskb
, sk
);
4815 /* Copy data, releasing collapsed skbs. */
4817 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
4818 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
4822 size
= min(copy
, size
);
4823 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
4825 TCP_SKB_CB(nskb
)->end_seq
+= size
;
4829 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4830 skb
= tcp_collapse_one(sk
, skb
, list
, root
);
4833 (TCP_SKB_CB(skb
)->tcp_flags
& (TCPHDR_SYN
| TCPHDR_FIN
)))
4839 skb_queue_walk_safe(&tmp
, skb
, n
)
4840 tcp_rbtree_insert(root
, skb
);
4843 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4844 * and tcp_collapse() them until all the queue is collapsed.
4846 static void tcp_collapse_ofo_queue(struct sock
*sk
)
4848 struct tcp_sock
*tp
= tcp_sk(sk
);
4849 struct sk_buff
*skb
, *head
;
4853 p
= rb_first(&tp
->out_of_order_queue
);
4854 skb
= rb_entry_safe(p
, struct sk_buff
, rbnode
);
4857 p
= rb_last(&tp
->out_of_order_queue
);
4858 /* Note: This is possible p is NULL here. We do not
4859 * use rb_entry_safe(), as ooo_last_skb is valid only
4860 * if rbtree is not empty.
4862 tp
->ooo_last_skb
= rb_entry(p
, struct sk_buff
, rbnode
);
4865 start
= TCP_SKB_CB(skb
)->seq
;
4866 end
= TCP_SKB_CB(skb
)->end_seq
;
4868 for (head
= skb
;;) {
4869 skb
= tcp_skb_next(skb
, NULL
);
4871 /* Range is terminated when we see a gap or when
4872 * we are at the queue end.
4875 after(TCP_SKB_CB(skb
)->seq
, end
) ||
4876 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
4877 tcp_collapse(sk
, NULL
, &tp
->out_of_order_queue
,
4878 head
, skb
, start
, end
);
4882 if (unlikely(before(TCP_SKB_CB(skb
)->seq
, start
)))
4883 start
= TCP_SKB_CB(skb
)->seq
;
4884 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
4885 end
= TCP_SKB_CB(skb
)->end_seq
;
4890 * Clean the out-of-order queue to make room.
4891 * We drop high sequences packets to :
4892 * 1) Let a chance for holes to be filled.
4893 * 2) not add too big latencies if thousands of packets sit there.
4894 * (But if application shrinks SO_RCVBUF, we could still end up
4895 * freeing whole queue here)
4897 * Return true if queue has shrunk.
4899 static bool tcp_prune_ofo_queue(struct sock
*sk
)
4901 struct tcp_sock
*tp
= tcp_sk(sk
);
4902 struct rb_node
*node
, *prev
;
4904 if (RB_EMPTY_ROOT(&tp
->out_of_order_queue
))
4907 NET_INC_STATS(sock_net(sk
), LINUX_MIB_OFOPRUNED
);
4908 node
= &tp
->ooo_last_skb
->rbnode
;
4910 prev
= rb_prev(node
);
4911 rb_erase(node
, &tp
->out_of_order_queue
);
4912 tcp_drop(sk
, rb_entry(node
, struct sk_buff
, rbnode
));
4914 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
&&
4915 !tcp_under_memory_pressure(sk
))
4919 tp
->ooo_last_skb
= rb_entry(prev
, struct sk_buff
, rbnode
);
4921 /* Reset SACK state. A conforming SACK implementation will
4922 * do the same at a timeout based retransmit. When a connection
4923 * is in a sad state like this, we care only about integrity
4924 * of the connection not performance.
4926 if (tp
->rx_opt
.sack_ok
)
4927 tcp_sack_reset(&tp
->rx_opt
);
4931 /* Reduce allocated memory if we can, trying to get
4932 * the socket within its memory limits again.
4934 * Return less than zero if we should start dropping frames
4935 * until the socket owning process reads some of the data
4936 * to stabilize the situation.
4938 static int tcp_prune_queue(struct sock
*sk
)
4940 struct tcp_sock
*tp
= tcp_sk(sk
);
4942 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
4944 NET_INC_STATS(sock_net(sk
), LINUX_MIB_PRUNECALLED
);
4946 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
4947 tcp_clamp_window(sk
);
4948 else if (tcp_under_memory_pressure(sk
))
4949 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
4951 tcp_collapse_ofo_queue(sk
);
4952 if (!skb_queue_empty(&sk
->sk_receive_queue
))
4953 tcp_collapse(sk
, &sk
->sk_receive_queue
, NULL
,
4954 skb_peek(&sk
->sk_receive_queue
),
4956 tp
->copied_seq
, tp
->rcv_nxt
);
4959 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4962 /* Collapsing did not help, destructive actions follow.
4963 * This must not ever occur. */
4965 tcp_prune_ofo_queue(sk
);
4967 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4970 /* If we are really being abused, tell the caller to silently
4971 * drop receive data on the floor. It will get retransmitted
4972 * and hopefully then we'll have sufficient space.
4974 NET_INC_STATS(sock_net(sk
), LINUX_MIB_RCVPRUNED
);
4976 /* Massive buffer overcommit. */
4981 static bool tcp_should_expand_sndbuf(const struct sock
*sk
)
4983 const struct tcp_sock
*tp
= tcp_sk(sk
);
4985 /* If the user specified a specific send buffer setting, do
4988 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
4991 /* If we are under global TCP memory pressure, do not expand. */
4992 if (tcp_under_memory_pressure(sk
))
4995 /* If we are under soft global TCP memory pressure, do not expand. */
4996 if (sk_memory_allocated(sk
) >= sk_prot_mem_limits(sk
, 0))
4999 /* If we filled the congestion window, do not expand. */
5000 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
5006 /* When incoming ACK allowed to free some skb from write_queue,
5007 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
5008 * on the exit from tcp input handler.
5010 * PROBLEM: sndbuf expansion does not work well with largesend.
5012 static void tcp_new_space(struct sock
*sk
)
5014 struct tcp_sock
*tp
= tcp_sk(sk
);
5016 if (tcp_should_expand_sndbuf(sk
)) {
5017 tcp_sndbuf_expand(sk
);
5018 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
5021 sk
->sk_write_space(sk
);
5024 static void tcp_check_space(struct sock
*sk
)
5026 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
5027 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
5028 /* pairs with tcp_poll() */
5030 if (sk
->sk_socket
&&
5031 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
5033 if (!test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
5034 tcp_chrono_stop(sk
, TCP_CHRONO_SNDBUF_LIMITED
);
5039 static inline void tcp_data_snd_check(struct sock
*sk
)
5041 tcp_push_pending_frames(sk
);
5042 tcp_check_space(sk
);
5046 * Check if sending an ack is needed.
5048 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
5050 struct tcp_sock
*tp
= tcp_sk(sk
);
5052 /* More than one full frame received... */
5053 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
&&
5054 /* ... and right edge of window advances far enough.
5055 * (tcp_recvmsg() will send ACK otherwise). Or...
5057 __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
5058 /* We ACK each frame or... */
5059 tcp_in_quickack_mode(sk
) ||
5060 /* We have out of order data. */
5061 (ofo_possible
&& !RB_EMPTY_ROOT(&tp
->out_of_order_queue
))) {
5062 /* Then ack it now */
5065 /* Else, send delayed ack. */
5066 tcp_send_delayed_ack(sk
);
5070 static inline void tcp_ack_snd_check(struct sock
*sk
)
5072 if (!inet_csk_ack_scheduled(sk
)) {
5073 /* We sent a data segment already. */
5076 __tcp_ack_snd_check(sk
, 1);
5080 * This routine is only called when we have urgent data
5081 * signaled. Its the 'slow' part of tcp_urg. It could be
5082 * moved inline now as tcp_urg is only called from one
5083 * place. We handle URGent data wrong. We have to - as
5084 * BSD still doesn't use the correction from RFC961.
5085 * For 1003.1g we should support a new option TCP_STDURG to permit
5086 * either form (or just set the sysctl tcp_stdurg).
5089 static void tcp_check_urg(struct sock
*sk
, const struct tcphdr
*th
)
5091 struct tcp_sock
*tp
= tcp_sk(sk
);
5092 u32 ptr
= ntohs(th
->urg_ptr
);
5094 if (ptr
&& !sysctl_tcp_stdurg
)
5096 ptr
+= ntohl(th
->seq
);
5098 /* Ignore urgent data that we've already seen and read. */
5099 if (after(tp
->copied_seq
, ptr
))
5102 /* Do not replay urg ptr.
5104 * NOTE: interesting situation not covered by specs.
5105 * Misbehaving sender may send urg ptr, pointing to segment,
5106 * which we already have in ofo queue. We are not able to fetch
5107 * such data and will stay in TCP_URG_NOTYET until will be eaten
5108 * by recvmsg(). Seems, we are not obliged to handle such wicked
5109 * situations. But it is worth to think about possibility of some
5110 * DoSes using some hypothetical application level deadlock.
5112 if (before(ptr
, tp
->rcv_nxt
))
5115 /* Do we already have a newer (or duplicate) urgent pointer? */
5116 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
5119 /* Tell the world about our new urgent pointer. */
5122 /* We may be adding urgent data when the last byte read was
5123 * urgent. To do this requires some care. We cannot just ignore
5124 * tp->copied_seq since we would read the last urgent byte again
5125 * as data, nor can we alter copied_seq until this data arrives
5126 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5128 * NOTE. Double Dutch. Rendering to plain English: author of comment
5129 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5130 * and expect that both A and B disappear from stream. This is _wrong_.
5131 * Though this happens in BSD with high probability, this is occasional.
5132 * Any application relying on this is buggy. Note also, that fix "works"
5133 * only in this artificial test. Insert some normal data between A and B and we will
5134 * decline of BSD again. Verdict: it is better to remove to trap
5137 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
5138 !sock_flag(sk
, SOCK_URGINLINE
) && tp
->copied_seq
!= tp
->rcv_nxt
) {
5139 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
5141 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5142 __skb_unlink(skb
, &sk
->sk_receive_queue
);
5147 tp
->urg_data
= TCP_URG_NOTYET
;
5150 /* Disable header prediction. */
5154 /* This is the 'fast' part of urgent handling. */
5155 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, const struct tcphdr
*th
)
5157 struct tcp_sock
*tp
= tcp_sk(sk
);
5159 /* Check if we get a new urgent pointer - normally not. */
5161 tcp_check_urg(sk
, th
);
5163 /* Do we wait for any urgent data? - normally not... */
5164 if (tp
->urg_data
== TCP_URG_NOTYET
) {
5165 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
5168 /* Is the urgent pointer pointing into this packet? */
5169 if (ptr
< skb
->len
) {
5171 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
5173 tp
->urg_data
= TCP_URG_VALID
| tmp
;
5174 if (!sock_flag(sk
, SOCK_DEAD
))
5175 sk
->sk_data_ready(sk
);
5180 static int tcp_copy_to_iovec(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
5182 struct tcp_sock
*tp
= tcp_sk(sk
);
5183 int chunk
= skb
->len
- hlen
;
5186 if (skb_csum_unnecessary(skb
))
5187 err
= skb_copy_datagram_msg(skb
, hlen
, tp
->ucopy
.msg
, chunk
);
5189 err
= skb_copy_and_csum_datagram_msg(skb
, hlen
, tp
->ucopy
.msg
);
5192 tp
->ucopy
.len
-= chunk
;
5193 tp
->copied_seq
+= chunk
;
5194 tcp_rcv_space_adjust(sk
);
5200 /* Accept RST for rcv_nxt - 1 after a FIN.
5201 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a
5202 * FIN is sent followed by a RST packet. The RST is sent with the same
5203 * sequence number as the FIN, and thus according to RFC 5961 a challenge
5204 * ACK should be sent. However, Mac OSX rate limits replies to challenge
5205 * ACKs on the closed socket. In addition middleboxes can drop either the
5206 * challenge ACK or a subsequent RST.
5208 static bool tcp_reset_check(const struct sock
*sk
, const struct sk_buff
*skb
)
5210 struct tcp_sock
*tp
= tcp_sk(sk
);
5212 return unlikely(TCP_SKB_CB(skb
)->seq
== (tp
->rcv_nxt
- 1) &&
5213 (1 << sk
->sk_state
) & (TCPF_CLOSE_WAIT
| TCPF_LAST_ACK
|
5217 /* Does PAWS and seqno based validation of an incoming segment, flags will
5218 * play significant role here.
5220 static bool tcp_validate_incoming(struct sock
*sk
, struct sk_buff
*skb
,
5221 const struct tcphdr
*th
, int syn_inerr
)
5223 struct tcp_sock
*tp
= tcp_sk(sk
);
5224 bool rst_seq_match
= false;
5226 /* RFC1323: H1. Apply PAWS check first. */
5227 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
5228 tcp_paws_discard(sk
, skb
)) {
5230 NET_INC_STATS(sock_net(sk
), LINUX_MIB_PAWSESTABREJECTED
);
5231 if (!tcp_oow_rate_limited(sock_net(sk
), skb
,
5232 LINUX_MIB_TCPACKSKIPPEDPAWS
,
5233 &tp
->last_oow_ack_time
))
5234 tcp_send_dupack(sk
, skb
);
5237 /* Reset is accepted even if it did not pass PAWS. */
5240 /* Step 1: check sequence number */
5241 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5242 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5243 * (RST) segments are validated by checking their SEQ-fields."
5244 * And page 69: "If an incoming segment is not acceptable,
5245 * an acknowledgment should be sent in reply (unless the RST
5246 * bit is set, if so drop the segment and return)".
5251 if (!tcp_oow_rate_limited(sock_net(sk
), skb
,
5252 LINUX_MIB_TCPACKSKIPPEDSEQ
,
5253 &tp
->last_oow_ack_time
))
5254 tcp_send_dupack(sk
, skb
);
5255 } else if (tcp_reset_check(sk
, skb
)) {
5261 /* Step 2: check RST bit */
5263 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a
5264 * FIN and SACK too if available):
5265 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or
5266 * the right-most SACK block,
5268 * RESET the connection
5270 * Send a challenge ACK
5272 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
||
5273 tcp_reset_check(sk
, skb
)) {
5274 rst_seq_match
= true;
5275 } else if (tcp_is_sack(tp
) && tp
->rx_opt
.num_sacks
> 0) {
5276 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
5277 int max_sack
= sp
[0].end_seq
;
5280 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
;
5282 max_sack
= after(sp
[this_sack
].end_seq
,
5284 sp
[this_sack
].end_seq
: max_sack
;
5287 if (TCP_SKB_CB(skb
)->seq
== max_sack
)
5288 rst_seq_match
= true;
5294 tcp_send_challenge_ack(sk
, skb
);
5298 /* step 3: check security and precedence [ignored] */
5300 /* step 4: Check for a SYN
5301 * RFC 5961 4.2 : Send a challenge ack
5306 TCP_INC_STATS(sock_net(sk
), TCP_MIB_INERRS
);
5307 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPSYNCHALLENGE
);
5308 tcp_send_challenge_ack(sk
, skb
);
5320 * TCP receive function for the ESTABLISHED state.
5322 * It is split into a fast path and a slow path. The fast path is
5324 * - A zero window was announced from us - zero window probing
5325 * is only handled properly in the slow path.
5326 * - Out of order segments arrived.
5327 * - Urgent data is expected.
5328 * - There is no buffer space left
5329 * - Unexpected TCP flags/window values/header lengths are received
5330 * (detected by checking the TCP header against pred_flags)
5331 * - Data is sent in both directions. Fast path only supports pure senders
5332 * or pure receivers (this means either the sequence number or the ack
5333 * value must stay constant)
5334 * - Unexpected TCP option.
5336 * When these conditions are not satisfied it drops into a standard
5337 * receive procedure patterned after RFC793 to handle all cases.
5338 * The first three cases are guaranteed by proper pred_flags setting,
5339 * the rest is checked inline. Fast processing is turned on in
5340 * tcp_data_queue when everything is OK.
5342 void tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
5343 const struct tcphdr
*th
, unsigned int len
)
5345 struct tcp_sock
*tp
= tcp_sk(sk
);
5347 if (unlikely(!sk
->sk_rx_dst
))
5348 inet_csk(sk
)->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
5350 * Header prediction.
5351 * The code loosely follows the one in the famous
5352 * "30 instruction TCP receive" Van Jacobson mail.
5354 * Van's trick is to deposit buffers into socket queue
5355 * on a device interrupt, to call tcp_recv function
5356 * on the receive process context and checksum and copy
5357 * the buffer to user space. smart...
5359 * Our current scheme is not silly either but we take the
5360 * extra cost of the net_bh soft interrupt processing...
5361 * We do checksum and copy also but from device to kernel.
5364 tp
->rx_opt
.saw_tstamp
= 0;
5366 /* pred_flags is 0xS?10 << 16 + snd_wnd
5367 * if header_prediction is to be made
5368 * 'S' will always be tp->tcp_header_len >> 2
5369 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5370 * turn it off (when there are holes in the receive
5371 * space for instance)
5372 * PSH flag is ignored.
5375 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
5376 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
&&
5377 !after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
)) {
5378 int tcp_header_len
= tp
->tcp_header_len
;
5380 /* Timestamp header prediction: tcp_header_len
5381 * is automatically equal to th->doff*4 due to pred_flags
5385 /* Check timestamp */
5386 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
5387 /* No? Slow path! */
5388 if (!tcp_parse_aligned_timestamp(tp
, th
))
5391 /* If PAWS failed, check it more carefully in slow path */
5392 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
5395 /* DO NOT update ts_recent here, if checksum fails
5396 * and timestamp was corrupted part, it will result
5397 * in a hung connection since we will drop all
5398 * future packets due to the PAWS test.
5402 if (len
<= tcp_header_len
) {
5403 /* Bulk data transfer: sender */
5404 if (len
== tcp_header_len
) {
5405 /* Predicted packet is in window by definition.
5406 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5407 * Hence, check seq<=rcv_wup reduces to:
5409 if (tcp_header_len
==
5410 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5411 tp
->rcv_nxt
== tp
->rcv_wup
)
5412 tcp_store_ts_recent(tp
);
5414 /* We know that such packets are checksummed
5417 tcp_ack(sk
, skb
, 0);
5419 tcp_data_snd_check(sk
);
5421 } else { /* Header too small */
5422 TCP_INC_STATS(sock_net(sk
), TCP_MIB_INERRS
);
5427 bool fragstolen
= false;
5429 if (tp
->ucopy
.task
== current
&&
5430 tp
->copied_seq
== tp
->rcv_nxt
&&
5431 len
- tcp_header_len
<= tp
->ucopy
.len
&&
5432 sock_owned_by_user(sk
)) {
5433 __set_current_state(TASK_RUNNING
);
5435 if (!tcp_copy_to_iovec(sk
, skb
, tcp_header_len
)) {
5436 /* Predicted packet is in window by definition.
5437 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5438 * Hence, check seq<=rcv_wup reduces to:
5440 if (tcp_header_len
==
5441 (sizeof(struct tcphdr
) +
5442 TCPOLEN_TSTAMP_ALIGNED
) &&
5443 tp
->rcv_nxt
== tp
->rcv_wup
)
5444 tcp_store_ts_recent(tp
);
5446 tcp_rcv_rtt_measure_ts(sk
, skb
);
5448 __skb_pull(skb
, tcp_header_len
);
5449 tcp_rcv_nxt_update(tp
, TCP_SKB_CB(skb
)->end_seq
);
5450 NET_INC_STATS(sock_net(sk
),
5451 LINUX_MIB_TCPHPHITSTOUSER
);
5456 if (tcp_checksum_complete(skb
))
5459 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
5462 /* Predicted packet is in window by definition.
5463 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5464 * Hence, check seq<=rcv_wup reduces to:
5466 if (tcp_header_len
==
5467 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5468 tp
->rcv_nxt
== tp
->rcv_wup
)
5469 tcp_store_ts_recent(tp
);
5471 tcp_rcv_rtt_measure_ts(sk
, skb
);
5473 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPHPHITS
);
5475 /* Bulk data transfer: receiver */
5476 eaten
= tcp_queue_rcv(sk
, skb
, tcp_header_len
,
5480 tcp_event_data_recv(sk
, skb
);
5482 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
5483 /* Well, only one small jumplet in fast path... */
5484 tcp_ack(sk
, skb
, FLAG_DATA
);
5485 tcp_data_snd_check(sk
);
5486 if (!inet_csk_ack_scheduled(sk
))
5490 __tcp_ack_snd_check(sk
, 0);
5493 kfree_skb_partial(skb
, fragstolen
);
5494 sk
->sk_data_ready(sk
);
5500 if (len
< (th
->doff
<< 2) || tcp_checksum_complete(skb
))
5503 if (!th
->ack
&& !th
->rst
&& !th
->syn
)
5507 * Standard slow path.
5510 if (!tcp_validate_incoming(sk
, skb
, th
, 1))
5514 if (tcp_ack(sk
, skb
, FLAG_SLOWPATH
| FLAG_UPDATE_TS_RECENT
) < 0)
5517 tcp_rcv_rtt_measure_ts(sk
, skb
);
5519 /* Process urgent data. */
5520 tcp_urg(sk
, skb
, th
);
5522 /* step 7: process the segment text */
5523 tcp_data_queue(sk
, skb
);
5525 tcp_data_snd_check(sk
);
5526 tcp_ack_snd_check(sk
);
5530 TCP_INC_STATS(sock_net(sk
), TCP_MIB_CSUMERRORS
);
5531 TCP_INC_STATS(sock_net(sk
), TCP_MIB_INERRS
);
5536 EXPORT_SYMBOL(tcp_rcv_established
);
5538 void tcp_finish_connect(struct sock
*sk
, struct sk_buff
*skb
)
5540 struct tcp_sock
*tp
= tcp_sk(sk
);
5541 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5543 tcp_set_state(sk
, TCP_ESTABLISHED
);
5544 icsk
->icsk_ack
.lrcvtime
= tcp_time_stamp
;
5547 icsk
->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
5548 security_inet_conn_established(sk
, skb
);
5551 /* Make sure socket is routed, for correct metrics. */
5552 icsk
->icsk_af_ops
->rebuild_header(sk
);
5554 tcp_init_metrics(sk
);
5556 tcp_init_congestion_control(sk
);
5558 /* Prevent spurious tcp_cwnd_restart() on first data
5561 tp
->lsndtime
= tcp_time_stamp
;
5563 tcp_init_buffer_space(sk
);
5565 if (sock_flag(sk
, SOCK_KEEPOPEN
))
5566 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
5568 if (!tp
->rx_opt
.snd_wscale
)
5569 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
5573 if (!sock_flag(sk
, SOCK_DEAD
)) {
5574 sk
->sk_state_change(sk
);
5575 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5579 static bool tcp_rcv_fastopen_synack(struct sock
*sk
, struct sk_buff
*synack
,
5580 struct tcp_fastopen_cookie
*cookie
)
5582 struct tcp_sock
*tp
= tcp_sk(sk
);
5583 struct sk_buff
*data
= tp
->syn_data
? tcp_write_queue_head(sk
) : NULL
;
5584 u16 mss
= tp
->rx_opt
.mss_clamp
, try_exp
= 0;
5585 bool syn_drop
= false;
5587 if (mss
== tp
->rx_opt
.user_mss
) {
5588 struct tcp_options_received opt
;
5590 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5591 tcp_clear_options(&opt
);
5592 opt
.user_mss
= opt
.mss_clamp
= 0;
5593 tcp_parse_options(synack
, &opt
, 0, NULL
);
5594 mss
= opt
.mss_clamp
;
5597 if (!tp
->syn_fastopen
) {
5598 /* Ignore an unsolicited cookie */
5600 } else if (tp
->total_retrans
) {
5601 /* SYN timed out and the SYN-ACK neither has a cookie nor
5602 * acknowledges data. Presumably the remote received only
5603 * the retransmitted (regular) SYNs: either the original
5604 * SYN-data or the corresponding SYN-ACK was dropped.
5606 syn_drop
= (cookie
->len
< 0 && data
);
5607 } else if (cookie
->len
< 0 && !tp
->syn_data
) {
5608 /* We requested a cookie but didn't get it. If we did not use
5609 * the (old) exp opt format then try so next time (try_exp=1).
5610 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5612 try_exp
= tp
->syn_fastopen_exp
? 2 : 1;
5615 tcp_fastopen_cache_set(sk
, mss
, cookie
, syn_drop
, try_exp
);
5617 if (data
) { /* Retransmit unacked data in SYN */
5618 tcp_for_write_queue_from(data
, sk
) {
5619 if (data
== tcp_send_head(sk
) ||
5620 __tcp_retransmit_skb(sk
, data
, 1))
5624 NET_INC_STATS(sock_net(sk
),
5625 LINUX_MIB_TCPFASTOPENACTIVEFAIL
);
5628 tp
->syn_data_acked
= tp
->syn_data
;
5629 if (tp
->syn_data_acked
)
5630 NET_INC_STATS(sock_net(sk
),
5631 LINUX_MIB_TCPFASTOPENACTIVE
);
5633 tcp_fastopen_add_skb(sk
, synack
);
5638 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5639 const struct tcphdr
*th
)
5641 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5642 struct tcp_sock
*tp
= tcp_sk(sk
);
5643 struct tcp_fastopen_cookie foc
= { .len
= -1 };
5644 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
5646 tcp_parse_options(skb
, &tp
->rx_opt
, 0, &foc
);
5647 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
5648 tp
->rx_opt
.rcv_tsecr
-= tp
->tsoffset
;
5652 * "If the state is SYN-SENT then
5653 * first check the ACK bit
5654 * If the ACK bit is set
5655 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5656 * a reset (unless the RST bit is set, if so drop
5657 * the segment and return)"
5659 if (!after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_una
) ||
5660 after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
))
5661 goto reset_and_undo
;
5663 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
5664 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
5666 NET_INC_STATS(sock_net(sk
),
5667 LINUX_MIB_PAWSACTIVEREJECTED
);
5668 goto reset_and_undo
;
5671 /* Now ACK is acceptable.
5673 * "If the RST bit is set
5674 * If the ACK was acceptable then signal the user "error:
5675 * connection reset", drop the segment, enter CLOSED state,
5676 * delete TCB, and return."
5685 * "fifth, if neither of the SYN or RST bits is set then
5686 * drop the segment and return."
5692 goto discard_and_undo
;
5695 * "If the SYN bit is on ...
5696 * are acceptable then ...
5697 * (our SYN has been ACKed), change the connection
5698 * state to ESTABLISHED..."
5701 tcp_ecn_rcv_synack(tp
, th
);
5703 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5704 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
5706 /* Ok.. it's good. Set up sequence numbers and
5707 * move to established.
5709 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5710 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5712 /* RFC1323: The window in SYN & SYN/ACK segments is
5715 tp
->snd_wnd
= ntohs(th
->window
);
5717 if (!tp
->rx_opt
.wscale_ok
) {
5718 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
5719 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
5722 if (tp
->rx_opt
.saw_tstamp
) {
5723 tp
->rx_opt
.tstamp_ok
= 1;
5724 tp
->tcp_header_len
=
5725 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5726 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5727 tcp_store_ts_recent(tp
);
5729 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5732 if (tcp_is_sack(tp
) && sysctl_tcp_fack
)
5733 tcp_enable_fack(tp
);
5736 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5737 tcp_initialize_rcv_mss(sk
);
5739 /* Remember, tcp_poll() does not lock socket!
5740 * Change state from SYN-SENT only after copied_seq
5741 * is initialized. */
5742 tp
->copied_seq
= tp
->rcv_nxt
;
5746 tcp_finish_connect(sk
, skb
);
5748 if ((tp
->syn_fastopen
|| tp
->syn_data
) &&
5749 tcp_rcv_fastopen_synack(sk
, skb
, &foc
))
5752 if (sk
->sk_write_pending
||
5753 icsk
->icsk_accept_queue
.rskq_defer_accept
||
5754 icsk
->icsk_ack
.pingpong
) {
5755 /* Save one ACK. Data will be ready after
5756 * several ticks, if write_pending is set.
5758 * It may be deleted, but with this feature tcpdumps
5759 * look so _wonderfully_ clever, that I was not able
5760 * to stand against the temptation 8) --ANK
5762 inet_csk_schedule_ack(sk
);
5763 tcp_enter_quickack_mode(sk
);
5764 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
5765 TCP_DELACK_MAX
, TCP_RTO_MAX
);
5776 /* No ACK in the segment */
5780 * "If the RST bit is set
5782 * Otherwise (no ACK) drop the segment and return."
5785 goto discard_and_undo
;
5789 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&&
5790 tcp_paws_reject(&tp
->rx_opt
, 0))
5791 goto discard_and_undo
;
5794 /* We see SYN without ACK. It is attempt of
5795 * simultaneous connect with crossed SYNs.
5796 * Particularly, it can be connect to self.
5798 tcp_set_state(sk
, TCP_SYN_RECV
);
5800 if (tp
->rx_opt
.saw_tstamp
) {
5801 tp
->rx_opt
.tstamp_ok
= 1;
5802 tcp_store_ts_recent(tp
);
5803 tp
->tcp_header_len
=
5804 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5806 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5809 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5810 tp
->copied_seq
= tp
->rcv_nxt
;
5811 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5813 /* RFC1323: The window in SYN & SYN/ACK segments is
5816 tp
->snd_wnd
= ntohs(th
->window
);
5817 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
5818 tp
->max_window
= tp
->snd_wnd
;
5820 tcp_ecn_rcv_syn(tp
, th
);
5823 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5824 tcp_initialize_rcv_mss(sk
);
5826 tcp_send_synack(sk
);
5828 /* Note, we could accept data and URG from this segment.
5829 * There are no obstacles to make this (except that we must
5830 * either change tcp_recvmsg() to prevent it from returning data
5831 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5833 * However, if we ignore data in ACKless segments sometimes,
5834 * we have no reasons to accept it sometimes.
5835 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5836 * is not flawless. So, discard packet for sanity.
5837 * Uncomment this return to process the data.
5844 /* "fifth, if neither of the SYN or RST bits is set then
5845 * drop the segment and return."
5849 tcp_clear_options(&tp
->rx_opt
);
5850 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5854 tcp_clear_options(&tp
->rx_opt
);
5855 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5860 * This function implements the receiving procedure of RFC 793 for
5861 * all states except ESTABLISHED and TIME_WAIT.
5862 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5863 * address independent.
5866 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
)
5868 struct tcp_sock
*tp
= tcp_sk(sk
);
5869 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5870 const struct tcphdr
*th
= tcp_hdr(skb
);
5871 struct request_sock
*req
;
5875 switch (sk
->sk_state
) {
5889 /* It is possible that we process SYN packets from backlog,
5890 * so we need to make sure to disable BH right there.
5893 acceptable
= icsk
->icsk_af_ops
->conn_request(sk
, skb
) >= 0;
5904 tp
->rx_opt
.saw_tstamp
= 0;
5905 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
);
5909 /* Do step6 onward by hand. */
5910 tcp_urg(sk
, skb
, th
);
5912 tcp_data_snd_check(sk
);
5916 tp
->rx_opt
.saw_tstamp
= 0;
5917 req
= tp
->fastopen_rsk
;
5919 WARN_ON_ONCE(sk
->sk_state
!= TCP_SYN_RECV
&&
5920 sk
->sk_state
!= TCP_FIN_WAIT1
);
5922 if (!tcp_check_req(sk
, skb
, req
, true))
5926 if (!th
->ack
&& !th
->rst
&& !th
->syn
)
5929 if (!tcp_validate_incoming(sk
, skb
, th
, 0))
5932 /* step 5: check the ACK field */
5933 acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
|
5934 FLAG_UPDATE_TS_RECENT
) > 0;
5936 switch (sk
->sk_state
) {
5942 tcp_synack_rtt_meas(sk
, req
);
5944 /* Once we leave TCP_SYN_RECV, we no longer need req
5948 inet_csk(sk
)->icsk_retransmits
= 0;
5949 reqsk_fastopen_remove(sk
, req
, false);
5951 /* Make sure socket is routed, for correct metrics. */
5952 icsk
->icsk_af_ops
->rebuild_header(sk
);
5953 tcp_init_congestion_control(sk
);
5956 tp
->copied_seq
= tp
->rcv_nxt
;
5957 tcp_init_buffer_space(sk
);
5960 tcp_set_state(sk
, TCP_ESTABLISHED
);
5961 sk
->sk_state_change(sk
);
5963 /* Note, that this wakeup is only for marginal crossed SYN case.
5964 * Passively open sockets are not waked up, because
5965 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5968 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5970 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
5971 tp
->snd_wnd
= ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
;
5972 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5974 if (tp
->rx_opt
.tstamp_ok
)
5975 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5978 /* Re-arm the timer because data may have been sent out.
5979 * This is similar to the regular data transmission case
5980 * when new data has just been ack'ed.
5982 * (TFO) - we could try to be more aggressive and
5983 * retransmitting any data sooner based on when they
5988 tcp_init_metrics(sk
);
5990 if (!inet_csk(sk
)->icsk_ca_ops
->cong_control
)
5991 tcp_update_pacing_rate(sk
);
5993 /* Prevent spurious tcp_cwnd_restart() on first data packet */
5994 tp
->lsndtime
= tcp_time_stamp
;
5996 tcp_initialize_rcv_mss(sk
);
5997 tcp_fast_path_on(tp
);
6000 case TCP_FIN_WAIT1
: {
6003 /* If we enter the TCP_FIN_WAIT1 state and we are a
6004 * Fast Open socket and this is the first acceptable
6005 * ACK we have received, this would have acknowledged
6006 * our SYNACK so stop the SYNACK timer.
6009 /* Return RST if ack_seq is invalid.
6010 * Note that RFC793 only says to generate a
6011 * DUPACK for it but for TCP Fast Open it seems
6012 * better to treat this case like TCP_SYN_RECV
6017 /* We no longer need the request sock. */
6018 reqsk_fastopen_remove(sk
, req
, false);
6021 if (tp
->snd_una
!= tp
->write_seq
)
6024 tcp_set_state(sk
, TCP_FIN_WAIT2
);
6025 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
6029 if (!sock_flag(sk
, SOCK_DEAD
)) {
6030 /* Wake up lingering close() */
6031 sk
->sk_state_change(sk
);
6035 if (tp
->linger2
< 0 ||
6036 (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
6037 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
))) {
6039 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
6043 tmo
= tcp_fin_time(sk
);
6044 if (tmo
> TCP_TIMEWAIT_LEN
) {
6045 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
6046 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
6047 /* Bad case. We could lose such FIN otherwise.
6048 * It is not a big problem, but it looks confusing
6049 * and not so rare event. We still can lose it now,
6050 * if it spins in bh_lock_sock(), but it is really
6053 inet_csk_reset_keepalive_timer(sk
, tmo
);
6055 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
6062 if (tp
->snd_una
== tp
->write_seq
) {
6063 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
6069 if (tp
->snd_una
== tp
->write_seq
) {
6070 tcp_update_metrics(sk
);
6077 /* step 6: check the URG bit */
6078 tcp_urg(sk
, skb
, th
);
6080 /* step 7: process the segment text */
6081 switch (sk
->sk_state
) {
6082 case TCP_CLOSE_WAIT
:
6085 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
6089 /* RFC 793 says to queue data in these states,
6090 * RFC 1122 says we MUST send a reset.
6091 * BSD 4.4 also does reset.
6093 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
6094 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
6095 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
6096 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
6102 case TCP_ESTABLISHED
:
6103 tcp_data_queue(sk
, skb
);
6108 /* tcp_data could move socket to TIME-WAIT */
6109 if (sk
->sk_state
!= TCP_CLOSE
) {
6110 tcp_data_snd_check(sk
);
6111 tcp_ack_snd_check(sk
);
6120 EXPORT_SYMBOL(tcp_rcv_state_process
);
6122 static inline void pr_drop_req(struct request_sock
*req
, __u16 port
, int family
)
6124 struct inet_request_sock
*ireq
= inet_rsk(req
);
6126 if (family
== AF_INET
)
6127 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6128 &ireq
->ir_rmt_addr
, port
);
6129 #if IS_ENABLED(CONFIG_IPV6)
6130 else if (family
== AF_INET6
)
6131 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6132 &ireq
->ir_v6_rmt_addr
, port
);
6136 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6138 * If we receive a SYN packet with these bits set, it means a
6139 * network is playing bad games with TOS bits. In order to
6140 * avoid possible false congestion notifications, we disable
6141 * TCP ECN negotiation.
6143 * Exception: tcp_ca wants ECN. This is required for DCTCP
6144 * congestion control: Linux DCTCP asserts ECT on all packets,
6145 * including SYN, which is most optimal solution; however,
6146 * others, such as FreeBSD do not.
6148 static void tcp_ecn_create_request(struct request_sock
*req
,
6149 const struct sk_buff
*skb
,
6150 const struct sock
*listen_sk
,
6151 const struct dst_entry
*dst
)
6153 const struct tcphdr
*th
= tcp_hdr(skb
);
6154 const struct net
*net
= sock_net(listen_sk
);
6155 bool th_ecn
= th
->ece
&& th
->cwr
;
6162 ect
= !INET_ECN_is_not_ect(TCP_SKB_CB(skb
)->ip_dsfield
);
6163 ecn_ok_dst
= dst_feature(dst
, DST_FEATURE_ECN_MASK
);
6164 ecn_ok
= net
->ipv4
.sysctl_tcp_ecn
|| ecn_ok_dst
;
6166 if ((!ect
&& ecn_ok
) || tcp_ca_needs_ecn(listen_sk
) ||
6167 (ecn_ok_dst
& DST_FEATURE_ECN_CA
))
6168 inet_rsk(req
)->ecn_ok
= 1;
6171 static void tcp_openreq_init(struct request_sock
*req
,
6172 const struct tcp_options_received
*rx_opt
,
6173 struct sk_buff
*skb
, const struct sock
*sk
)
6175 struct inet_request_sock
*ireq
= inet_rsk(req
);
6177 req
->rsk_rcv_wnd
= 0; /* So that tcp_send_synack() knows! */
6179 tcp_rsk(req
)->rcv_isn
= TCP_SKB_CB(skb
)->seq
;
6180 tcp_rsk(req
)->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
6181 skb_mstamp_get(&tcp_rsk(req
)->snt_synack
);
6182 tcp_rsk(req
)->last_oow_ack_time
= 0;
6183 req
->mss
= rx_opt
->mss_clamp
;
6184 req
->ts_recent
= rx_opt
->saw_tstamp
? rx_opt
->rcv_tsval
: 0;
6185 ireq
->tstamp_ok
= rx_opt
->tstamp_ok
;
6186 ireq
->sack_ok
= rx_opt
->sack_ok
;
6187 ireq
->snd_wscale
= rx_opt
->snd_wscale
;
6188 ireq
->wscale_ok
= rx_opt
->wscale_ok
;
6191 ireq
->ir_rmt_port
= tcp_hdr(skb
)->source
;
6192 ireq
->ir_num
= ntohs(tcp_hdr(skb
)->dest
);
6193 ireq
->ir_mark
= inet_request_mark(sk
, skb
);
6196 struct request_sock
*inet_reqsk_alloc(const struct request_sock_ops
*ops
,
6197 struct sock
*sk_listener
,
6198 bool attach_listener
)
6200 struct request_sock
*req
= reqsk_alloc(ops
, sk_listener
,
6204 struct inet_request_sock
*ireq
= inet_rsk(req
);
6206 kmemcheck_annotate_bitfield(ireq
, flags
);
6208 #if IS_ENABLED(CONFIG_IPV6)
6209 ireq
->pktopts
= NULL
;
6211 atomic64_set(&ireq
->ir_cookie
, 0);
6212 ireq
->ireq_state
= TCP_NEW_SYN_RECV
;
6213 write_pnet(&ireq
->ireq_net
, sock_net(sk_listener
));
6214 ireq
->ireq_family
= sk_listener
->sk_family
;
6219 EXPORT_SYMBOL(inet_reqsk_alloc
);
6222 * Return true if a syncookie should be sent
6224 static bool tcp_syn_flood_action(const struct sock
*sk
,
6225 const struct sk_buff
*skb
,
6228 struct request_sock_queue
*queue
= &inet_csk(sk
)->icsk_accept_queue
;
6229 const char *msg
= "Dropping request";
6230 bool want_cookie
= false;
6231 struct net
*net
= sock_net(sk
);
6233 #ifdef CONFIG_SYN_COOKIES
6234 if (net
->ipv4
.sysctl_tcp_syncookies
) {
6235 msg
= "Sending cookies";
6237 __NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPREQQFULLDOCOOKIES
);
6240 __NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPREQQFULLDROP
);
6242 if (!queue
->synflood_warned
&&
6243 net
->ipv4
.sysctl_tcp_syncookies
!= 2 &&
6244 xchg(&queue
->synflood_warned
, 1) == 0)
6245 pr_info("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6246 proto
, ntohs(tcp_hdr(skb
)->dest
), msg
);
6251 static void tcp_reqsk_record_syn(const struct sock
*sk
,
6252 struct request_sock
*req
,
6253 const struct sk_buff
*skb
)
6255 if (tcp_sk(sk
)->save_syn
) {
6256 u32 len
= skb_network_header_len(skb
) + tcp_hdrlen(skb
);
6259 copy
= kmalloc(len
+ sizeof(u32
), GFP_ATOMIC
);
6262 memcpy(©
[1], skb_network_header(skb
), len
);
6263 req
->saved_syn
= copy
;
6268 int tcp_conn_request(struct request_sock_ops
*rsk_ops
,
6269 const struct tcp_request_sock_ops
*af_ops
,
6270 struct sock
*sk
, struct sk_buff
*skb
)
6272 struct tcp_fastopen_cookie foc
= { .len
= -1 };
6273 __u32 isn
= TCP_SKB_CB(skb
)->tcp_tw_isn
;
6274 struct tcp_options_received tmp_opt
;
6275 struct tcp_sock
*tp
= tcp_sk(sk
);
6276 struct net
*net
= sock_net(sk
);
6277 struct sock
*fastopen_sk
= NULL
;
6278 struct dst_entry
*dst
= NULL
;
6279 struct request_sock
*req
;
6280 bool want_cookie
= false;
6283 /* TW buckets are converted to open requests without
6284 * limitations, they conserve resources and peer is
6285 * evidently real one.
6287 if ((net
->ipv4
.sysctl_tcp_syncookies
== 2 ||
6288 inet_csk_reqsk_queue_is_full(sk
)) && !isn
) {
6289 want_cookie
= tcp_syn_flood_action(sk
, skb
, rsk_ops
->slab_name
);
6294 if (sk_acceptq_is_full(sk
)) {
6295 NET_INC_STATS(sock_net(sk
), LINUX_MIB_LISTENOVERFLOWS
);
6299 req
= inet_reqsk_alloc(rsk_ops
, sk
, !want_cookie
);
6303 tcp_rsk(req
)->af_specific
= af_ops
;
6304 tcp_rsk(req
)->ts_off
= 0;
6306 tcp_clear_options(&tmp_opt
);
6307 tmp_opt
.mss_clamp
= af_ops
->mss_clamp
;
6308 tmp_opt
.user_mss
= tp
->rx_opt
.user_mss
;
6309 tcp_parse_options(skb
, &tmp_opt
, 0, want_cookie
? NULL
: &foc
);
6311 if (want_cookie
&& !tmp_opt
.saw_tstamp
)
6312 tcp_clear_options(&tmp_opt
);
6314 tmp_opt
.tstamp_ok
= tmp_opt
.saw_tstamp
;
6315 tcp_openreq_init(req
, &tmp_opt
, skb
, sk
);
6316 inet_rsk(req
)->no_srccheck
= inet_sk(sk
)->transparent
;
6318 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6319 inet_rsk(req
)->ir_iif
= inet_request_bound_dev_if(sk
, skb
);
6321 af_ops
->init_req(req
, sk
, skb
);
6323 if (security_inet_conn_request(sk
, skb
, req
))
6326 if (isn
&& tmp_opt
.tstamp_ok
)
6327 af_ops
->init_seq(skb
, &tcp_rsk(req
)->ts_off
);
6329 if (!want_cookie
&& !isn
) {
6330 /* VJ's idea. We save last timestamp seen
6331 * from the destination in peer table, when entering
6332 * state TIME-WAIT, and check against it before
6333 * accepting new connection request.
6335 * If "isn" is not zero, this request hit alive
6336 * timewait bucket, so that all the necessary checks
6337 * are made in the function processing timewait state.
6339 if (net
->ipv4
.tcp_death_row
.sysctl_tw_recycle
) {
6342 dst
= af_ops
->route_req(sk
, &fl
, req
, &strict
);
6344 if (dst
&& strict
&&
6345 !tcp_peer_is_proven(req
, dst
, true,
6346 tmp_opt
.saw_tstamp
)) {
6347 NET_INC_STATS(sock_net(sk
), LINUX_MIB_PAWSPASSIVEREJECTED
);
6348 goto drop_and_release
;
6351 /* Kill the following clause, if you dislike this way. */
6352 else if (!net
->ipv4
.sysctl_tcp_syncookies
&&
6353 (net
->ipv4
.sysctl_max_syn_backlog
- inet_csk_reqsk_queue_len(sk
) <
6354 (net
->ipv4
.sysctl_max_syn_backlog
>> 2)) &&
6355 !tcp_peer_is_proven(req
, dst
, false,
6356 tmp_opt
.saw_tstamp
)) {
6357 /* Without syncookies last quarter of
6358 * backlog is filled with destinations,
6359 * proven to be alive.
6360 * It means that we continue to communicate
6361 * to destinations, already remembered
6362 * to the moment of synflood.
6364 pr_drop_req(req
, ntohs(tcp_hdr(skb
)->source
),
6366 goto drop_and_release
;
6369 isn
= af_ops
->init_seq(skb
, &tcp_rsk(req
)->ts_off
);
6372 dst
= af_ops
->route_req(sk
, &fl
, req
, NULL
);
6377 tcp_ecn_create_request(req
, skb
, sk
, dst
);
6380 isn
= cookie_init_sequence(af_ops
, sk
, skb
, &req
->mss
);
6381 tcp_rsk(req
)->ts_off
= 0;
6382 req
->cookie_ts
= tmp_opt
.tstamp_ok
;
6383 if (!tmp_opt
.tstamp_ok
)
6384 inet_rsk(req
)->ecn_ok
= 0;
6387 tcp_rsk(req
)->snt_isn
= isn
;
6388 tcp_rsk(req
)->txhash
= net_tx_rndhash();
6389 tcp_openreq_init_rwin(req
, sk
, dst
);
6391 tcp_reqsk_record_syn(sk
, req
, skb
);
6392 fastopen_sk
= tcp_try_fastopen(sk
, skb
, req
, &foc
, dst
);
6395 af_ops
->send_synack(fastopen_sk
, dst
, &fl
, req
,
6396 &foc
, TCP_SYNACK_FASTOPEN
);
6397 /* Add the child socket directly into the accept queue */
6398 inet_csk_reqsk_queue_add(sk
, req
, fastopen_sk
);
6399 sk
->sk_data_ready(sk
);
6400 bh_unlock_sock(fastopen_sk
);
6401 sock_put(fastopen_sk
);
6403 tcp_rsk(req
)->tfo_listener
= false;
6405 inet_csk_reqsk_queue_hash_add(sk
, req
, TCP_TIMEOUT_INIT
);
6406 af_ops
->send_synack(sk
, dst
, &fl
, req
, &foc
,
6407 !want_cookie
? TCP_SYNACK_NORMAL
:
6425 EXPORT_SYMBOL(tcp_conn_request
);