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
,
132 static bool __once __read_mostly
;
135 struct net_device
*dev
;
140 dev
= dev_get_by_index_rcu(sock_net(sk
), skb
->skb_iif
);
141 if (!dev
|| len
>= dev
->mtu
)
142 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n",
143 dev
? dev
->name
: "Unknown driver");
148 /* Adapt the MSS value used to make delayed ack decision to the
151 static void tcp_measure_rcv_mss(struct sock
*sk
, const struct sk_buff
*skb
)
153 struct inet_connection_sock
*icsk
= inet_csk(sk
);
154 const unsigned int lss
= icsk
->icsk_ack
.last_seg_size
;
157 icsk
->icsk_ack
.last_seg_size
= 0;
159 /* skb->len may jitter because of SACKs, even if peer
160 * sends good full-sized frames.
162 len
= skb_shinfo(skb
)->gso_size
? : skb
->len
;
163 if (len
>= icsk
->icsk_ack
.rcv_mss
) {
164 icsk
->icsk_ack
.rcv_mss
= min_t(unsigned int, len
,
166 /* Account for possibly-removed options */
167 if (unlikely(len
> icsk
->icsk_ack
.rcv_mss
+
168 MAX_TCP_OPTION_SPACE
))
169 tcp_gro_dev_warn(sk
, skb
, len
);
171 /* Otherwise, we make more careful check taking into account,
172 * that SACKs block is variable.
174 * "len" is invariant segment length, including TCP header.
176 len
+= skb
->data
- skb_transport_header(skb
);
177 if (len
>= TCP_MSS_DEFAULT
+ sizeof(struct tcphdr
) ||
178 /* If PSH is not set, packet should be
179 * full sized, provided peer TCP is not badly broken.
180 * This observation (if it is correct 8)) allows
181 * to handle super-low mtu links fairly.
183 (len
>= TCP_MIN_MSS
+ sizeof(struct tcphdr
) &&
184 !(tcp_flag_word(tcp_hdr(skb
)) & TCP_REMNANT
))) {
185 /* Subtract also invariant (if peer is RFC compliant),
186 * tcp header plus fixed timestamp option length.
187 * Resulting "len" is MSS free of SACK jitter.
189 len
-= tcp_sk(sk
)->tcp_header_len
;
190 icsk
->icsk_ack
.last_seg_size
= len
;
192 icsk
->icsk_ack
.rcv_mss
= len
;
196 if (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
)
197 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED2
;
198 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
202 static void tcp_incr_quickack(struct sock
*sk
)
204 struct inet_connection_sock
*icsk
= inet_csk(sk
);
205 unsigned int quickacks
= tcp_sk(sk
)->rcv_wnd
/ (2 * icsk
->icsk_ack
.rcv_mss
);
209 if (quickacks
> icsk
->icsk_ack
.quick
)
210 icsk
->icsk_ack
.quick
= min(quickacks
, TCP_MAX_QUICKACKS
);
213 static void tcp_enter_quickack_mode(struct sock
*sk
)
215 struct inet_connection_sock
*icsk
= inet_csk(sk
);
216 tcp_incr_quickack(sk
);
217 icsk
->icsk_ack
.pingpong
= 0;
218 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
221 /* Send ACKs quickly, if "quick" count is not exhausted
222 * and the session is not interactive.
225 static bool tcp_in_quickack_mode(struct sock
*sk
)
227 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
228 const struct dst_entry
*dst
= __sk_dst_get(sk
);
230 return (dst
&& dst_metric(dst
, RTAX_QUICKACK
)) ||
231 (icsk
->icsk_ack
.quick
&& !icsk
->icsk_ack
.pingpong
);
234 static void tcp_ecn_queue_cwr(struct tcp_sock
*tp
)
236 if (tp
->ecn_flags
& TCP_ECN_OK
)
237 tp
->ecn_flags
|= TCP_ECN_QUEUE_CWR
;
240 static void tcp_ecn_accept_cwr(struct tcp_sock
*tp
, const struct sk_buff
*skb
)
242 if (tcp_hdr(skb
)->cwr
)
243 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
246 static void tcp_ecn_withdraw_cwr(struct tcp_sock
*tp
)
248 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
251 static void __tcp_ecn_check_ce(struct tcp_sock
*tp
, const struct sk_buff
*skb
)
253 switch (TCP_SKB_CB(skb
)->ip_dsfield
& INET_ECN_MASK
) {
254 case INET_ECN_NOT_ECT
:
255 /* Funny extension: if ECT is not set on a segment,
256 * and we already seen ECT on a previous segment,
257 * it is probably a retransmit.
259 if (tp
->ecn_flags
& TCP_ECN_SEEN
)
260 tcp_enter_quickack_mode((struct sock
*)tp
);
263 if (tcp_ca_needs_ecn((struct sock
*)tp
))
264 tcp_ca_event((struct sock
*)tp
, CA_EVENT_ECN_IS_CE
);
266 if (!(tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)) {
267 /* Better not delay acks, sender can have a very low cwnd */
268 tcp_enter_quickack_mode((struct sock
*)tp
);
269 tp
->ecn_flags
|= TCP_ECN_DEMAND_CWR
;
271 tp
->ecn_flags
|= TCP_ECN_SEEN
;
274 if (tcp_ca_needs_ecn((struct sock
*)tp
))
275 tcp_ca_event((struct sock
*)tp
, CA_EVENT_ECN_NO_CE
);
276 tp
->ecn_flags
|= TCP_ECN_SEEN
;
281 static void tcp_ecn_check_ce(struct tcp_sock
*tp
, const struct sk_buff
*skb
)
283 if (tp
->ecn_flags
& TCP_ECN_OK
)
284 __tcp_ecn_check_ce(tp
, skb
);
287 static void tcp_ecn_rcv_synack(struct tcp_sock
*tp
, const struct tcphdr
*th
)
289 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| th
->cwr
))
290 tp
->ecn_flags
&= ~TCP_ECN_OK
;
293 static void tcp_ecn_rcv_syn(struct tcp_sock
*tp
, const struct tcphdr
*th
)
295 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| !th
->cwr
))
296 tp
->ecn_flags
&= ~TCP_ECN_OK
;
299 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock
*tp
, const struct tcphdr
*th
)
301 if (th
->ece
&& !th
->syn
&& (tp
->ecn_flags
& TCP_ECN_OK
))
306 /* Buffer size and advertised window tuning.
308 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
311 static void tcp_sndbuf_expand(struct sock
*sk
)
313 const struct tcp_sock
*tp
= tcp_sk(sk
);
314 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
318 /* Worst case is non GSO/TSO : each frame consumes one skb
319 * and skb->head is kmalloced using power of two area of memory
321 per_mss
= max_t(u32
, tp
->rx_opt
.mss_clamp
, tp
->mss_cache
) +
323 SKB_DATA_ALIGN(sizeof(struct skb_shared_info
));
325 per_mss
= roundup_pow_of_two(per_mss
) +
326 SKB_DATA_ALIGN(sizeof(struct sk_buff
));
328 nr_segs
= max_t(u32
, TCP_INIT_CWND
, tp
->snd_cwnd
);
329 nr_segs
= max_t(u32
, nr_segs
, tp
->reordering
+ 1);
331 /* Fast Recovery (RFC 5681 3.2) :
332 * Cubic needs 1.7 factor, rounded to 2 to include
333 * extra cushion (application might react slowly to POLLOUT)
335 sndmem
= ca_ops
->sndbuf_expand
? ca_ops
->sndbuf_expand(sk
) : 2;
336 sndmem
*= nr_segs
* per_mss
;
338 if (sk
->sk_sndbuf
< sndmem
)
339 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
342 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
344 * All tcp_full_space() is split to two parts: "network" buffer, allocated
345 * forward and advertised in receiver window (tp->rcv_wnd) and
346 * "application buffer", required to isolate scheduling/application
347 * latencies from network.
348 * window_clamp is maximal advertised window. It can be less than
349 * tcp_full_space(), in this case tcp_full_space() - window_clamp
350 * is reserved for "application" buffer. The less window_clamp is
351 * the smoother our behaviour from viewpoint of network, but the lower
352 * throughput and the higher sensitivity of the connection to losses. 8)
354 * rcv_ssthresh is more strict window_clamp used at "slow start"
355 * phase to predict further behaviour of this connection.
356 * It is used for two goals:
357 * - to enforce header prediction at sender, even when application
358 * requires some significant "application buffer". It is check #1.
359 * - to prevent pruning of receive queue because of misprediction
360 * of receiver window. Check #2.
362 * The scheme does not work when sender sends good segments opening
363 * window and then starts to feed us spaghetti. But it should work
364 * in common situations. Otherwise, we have to rely on queue collapsing.
367 /* Slow part of check#2. */
368 static int __tcp_grow_window(const struct sock
*sk
, const struct sk_buff
*skb
)
370 struct tcp_sock
*tp
= tcp_sk(sk
);
372 int truesize
= tcp_win_from_space(skb
->truesize
) >> 1;
373 int window
= tcp_win_from_space(sysctl_tcp_rmem
[2]) >> 1;
375 while (tp
->rcv_ssthresh
<= window
) {
376 if (truesize
<= skb
->len
)
377 return 2 * inet_csk(sk
)->icsk_ack
.rcv_mss
;
385 static void tcp_grow_window(struct sock
*sk
, const struct sk_buff
*skb
)
387 struct tcp_sock
*tp
= tcp_sk(sk
);
390 if (tp
->rcv_ssthresh
< tp
->window_clamp
&&
391 (int)tp
->rcv_ssthresh
< tcp_space(sk
) &&
392 !tcp_under_memory_pressure(sk
)) {
395 /* Check #2. Increase window, if skb with such overhead
396 * will fit to rcvbuf in future.
398 if (tcp_win_from_space(skb
->truesize
) <= skb
->len
)
399 incr
= 2 * tp
->advmss
;
401 incr
= __tcp_grow_window(sk
, skb
);
404 incr
= max_t(int, incr
, 2 * skb
->len
);
405 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
+ incr
,
407 inet_csk(sk
)->icsk_ack
.quick
|= 1;
412 /* 3. Tuning rcvbuf, when connection enters established state. */
413 static void tcp_fixup_rcvbuf(struct sock
*sk
)
415 u32 mss
= tcp_sk(sk
)->advmss
;
418 rcvmem
= 2 * SKB_TRUESIZE(mss
+ MAX_TCP_HEADER
) *
419 tcp_default_init_rwnd(mss
);
421 /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
422 * Allow enough cushion so that sender is not limited by our window
424 if (sysctl_tcp_moderate_rcvbuf
)
427 if (sk
->sk_rcvbuf
< rcvmem
)
428 sk
->sk_rcvbuf
= min(rcvmem
, sysctl_tcp_rmem
[2]);
431 /* 4. Try to fixup all. It is made immediately after connection enters
434 void tcp_init_buffer_space(struct sock
*sk
)
436 struct tcp_sock
*tp
= tcp_sk(sk
);
439 if (!(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
))
440 tcp_fixup_rcvbuf(sk
);
441 if (!(sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
))
442 tcp_sndbuf_expand(sk
);
444 tp
->rcvq_space
.space
= tp
->rcv_wnd
;
445 tp
->rcvq_space
.time
= tcp_time_stamp
;
446 tp
->rcvq_space
.seq
= tp
->copied_seq
;
448 maxwin
= tcp_full_space(sk
);
450 if (tp
->window_clamp
>= maxwin
) {
451 tp
->window_clamp
= maxwin
;
453 if (sysctl_tcp_app_win
&& maxwin
> 4 * tp
->advmss
)
454 tp
->window_clamp
= max(maxwin
-
455 (maxwin
>> sysctl_tcp_app_win
),
459 /* Force reservation of one segment. */
460 if (sysctl_tcp_app_win
&&
461 tp
->window_clamp
> 2 * tp
->advmss
&&
462 tp
->window_clamp
+ tp
->advmss
> maxwin
)
463 tp
->window_clamp
= max(2 * tp
->advmss
, maxwin
- tp
->advmss
);
465 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, tp
->window_clamp
);
466 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
469 /* 5. Recalculate window clamp after socket hit its memory bounds. */
470 static void tcp_clamp_window(struct sock
*sk
)
472 struct tcp_sock
*tp
= tcp_sk(sk
);
473 struct inet_connection_sock
*icsk
= inet_csk(sk
);
475 icsk
->icsk_ack
.quick
= 0;
477 if (sk
->sk_rcvbuf
< sysctl_tcp_rmem
[2] &&
478 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
) &&
479 !tcp_under_memory_pressure(sk
) &&
480 sk_memory_allocated(sk
) < sk_prot_mem_limits(sk
, 0)) {
481 sk
->sk_rcvbuf
= min(atomic_read(&sk
->sk_rmem_alloc
),
484 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
)
485 tp
->rcv_ssthresh
= min(tp
->window_clamp
, 2U * tp
->advmss
);
488 /* Initialize RCV_MSS value.
489 * RCV_MSS is an our guess about MSS used by the peer.
490 * We haven't any direct information about the MSS.
491 * It's better to underestimate the RCV_MSS rather than overestimate.
492 * Overestimations make us ACKing less frequently than needed.
493 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
495 void tcp_initialize_rcv_mss(struct sock
*sk
)
497 const struct tcp_sock
*tp
= tcp_sk(sk
);
498 unsigned int hint
= min_t(unsigned int, tp
->advmss
, tp
->mss_cache
);
500 hint
= min(hint
, tp
->rcv_wnd
/ 2);
501 hint
= min(hint
, TCP_MSS_DEFAULT
);
502 hint
= max(hint
, TCP_MIN_MSS
);
504 inet_csk(sk
)->icsk_ack
.rcv_mss
= hint
;
506 EXPORT_SYMBOL(tcp_initialize_rcv_mss
);
508 /* Receiver "autotuning" code.
510 * The algorithm for RTT estimation w/o timestamps is based on
511 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
512 * <http://public.lanl.gov/radiant/pubs.html#DRS>
514 * More detail on this code can be found at
515 * <http://staff.psc.edu/jheffner/>,
516 * though this reference is out of date. A new paper
519 static void tcp_rcv_rtt_update(struct tcp_sock
*tp
, u32 sample
, int win_dep
)
521 u32 new_sample
= tp
->rcv_rtt_est
.rtt
;
527 if (new_sample
!= 0) {
528 /* If we sample in larger samples in the non-timestamp
529 * case, we could grossly overestimate the RTT especially
530 * with chatty applications or bulk transfer apps which
531 * are stalled on filesystem I/O.
533 * Also, since we are only going for a minimum in the
534 * non-timestamp case, we do not smooth things out
535 * else with timestamps disabled convergence takes too
539 m
-= (new_sample
>> 3);
547 /* No previous measure. */
551 if (tp
->rcv_rtt_est
.rtt
!= new_sample
)
552 tp
->rcv_rtt_est
.rtt
= new_sample
;
555 static inline void tcp_rcv_rtt_measure(struct tcp_sock
*tp
)
557 if (tp
->rcv_rtt_est
.time
== 0)
559 if (before(tp
->rcv_nxt
, tp
->rcv_rtt_est
.seq
))
561 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rcv_rtt_est
.time
, 1);
564 tp
->rcv_rtt_est
.seq
= tp
->rcv_nxt
+ tp
->rcv_wnd
;
565 tp
->rcv_rtt_est
.time
= tcp_time_stamp
;
568 static inline void tcp_rcv_rtt_measure_ts(struct sock
*sk
,
569 const struct sk_buff
*skb
)
571 struct tcp_sock
*tp
= tcp_sk(sk
);
572 if (tp
->rx_opt
.rcv_tsecr
&&
573 (TCP_SKB_CB(skb
)->end_seq
-
574 TCP_SKB_CB(skb
)->seq
>= inet_csk(sk
)->icsk_ack
.rcv_mss
))
575 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
, 0);
579 * This function should be called every time data is copied to user space.
580 * It calculates the appropriate TCP receive buffer space.
582 void tcp_rcv_space_adjust(struct sock
*sk
)
584 struct tcp_sock
*tp
= tcp_sk(sk
);
588 time
= tcp_time_stamp
- tp
->rcvq_space
.time
;
589 if (time
< (tp
->rcv_rtt_est
.rtt
>> 3) || tp
->rcv_rtt_est
.rtt
== 0)
592 /* Number of bytes copied to user in last RTT */
593 copied
= tp
->copied_seq
- tp
->rcvq_space
.seq
;
594 if (copied
<= tp
->rcvq_space
.space
)
598 * copied = bytes received in previous RTT, our base window
599 * To cope with packet losses, we need a 2x factor
600 * To cope with slow start, and sender growing its cwin by 100 %
601 * every RTT, we need a 4x factor, because the ACK we are sending
602 * now is for the next RTT, not the current one :
603 * <prev RTT . ><current RTT .. ><next RTT .... >
606 if (sysctl_tcp_moderate_rcvbuf
&&
607 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
)) {
608 int rcvwin
, rcvmem
, rcvbuf
;
610 /* minimal window to cope with packet losses, assuming
611 * steady state. Add some cushion because of small variations.
613 rcvwin
= (copied
<< 1) + 16 * tp
->advmss
;
615 /* If rate increased by 25%,
616 * assume slow start, rcvwin = 3 * copied
617 * If rate increased by 50%,
618 * assume sender can use 2x growth, rcvwin = 4 * copied
621 tp
->rcvq_space
.space
+ (tp
->rcvq_space
.space
>> 2)) {
623 tp
->rcvq_space
.space
+ (tp
->rcvq_space
.space
>> 1))
626 rcvwin
+= (rcvwin
>> 1);
629 rcvmem
= SKB_TRUESIZE(tp
->advmss
+ MAX_TCP_HEADER
);
630 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
633 rcvbuf
= min(rcvwin
/ tp
->advmss
* rcvmem
, sysctl_tcp_rmem
[2]);
634 if (rcvbuf
> sk
->sk_rcvbuf
) {
635 sk
->sk_rcvbuf
= rcvbuf
;
637 /* Make the window clamp follow along. */
638 tp
->window_clamp
= rcvwin
;
641 tp
->rcvq_space
.space
= copied
;
644 tp
->rcvq_space
.seq
= tp
->copied_seq
;
645 tp
->rcvq_space
.time
= tcp_time_stamp
;
648 /* There is something which you must keep in mind when you analyze the
649 * behavior of the tp->ato delayed ack timeout interval. When a
650 * connection starts up, we want to ack as quickly as possible. The
651 * problem is that "good" TCP's do slow start at the beginning of data
652 * transmission. The means that until we send the first few ACK's the
653 * sender will sit on his end and only queue most of his data, because
654 * he can only send snd_cwnd unacked packets at any given time. For
655 * each ACK we send, he increments snd_cwnd and transmits more of his
658 static void tcp_event_data_recv(struct sock
*sk
, struct sk_buff
*skb
)
660 struct tcp_sock
*tp
= tcp_sk(sk
);
661 struct inet_connection_sock
*icsk
= inet_csk(sk
);
664 inet_csk_schedule_ack(sk
);
666 tcp_measure_rcv_mss(sk
, skb
);
668 tcp_rcv_rtt_measure(tp
);
670 now
= tcp_time_stamp
;
672 if (!icsk
->icsk_ack
.ato
) {
673 /* The _first_ data packet received, initialize
674 * delayed ACK engine.
676 tcp_incr_quickack(sk
);
677 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
679 int m
= now
- icsk
->icsk_ack
.lrcvtime
;
681 if (m
<= TCP_ATO_MIN
/ 2) {
682 /* The fastest case is the first. */
683 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + TCP_ATO_MIN
/ 2;
684 } else if (m
< icsk
->icsk_ack
.ato
) {
685 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + m
;
686 if (icsk
->icsk_ack
.ato
> icsk
->icsk_rto
)
687 icsk
->icsk_ack
.ato
= icsk
->icsk_rto
;
688 } else if (m
> icsk
->icsk_rto
) {
689 /* Too long gap. Apparently sender failed to
690 * restart window, so that we send ACKs quickly.
692 tcp_incr_quickack(sk
);
696 icsk
->icsk_ack
.lrcvtime
= now
;
698 tcp_ecn_check_ce(tp
, skb
);
701 tcp_grow_window(sk
, skb
);
704 /* Called to compute a smoothed rtt estimate. The data fed to this
705 * routine either comes from timestamps, or from segments that were
706 * known _not_ to have been retransmitted [see Karn/Partridge
707 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
708 * piece by Van Jacobson.
709 * NOTE: the next three routines used to be one big routine.
710 * To save cycles in the RFC 1323 implementation it was better to break
711 * it up into three procedures. -- erics
713 static void tcp_rtt_estimator(struct sock
*sk
, long mrtt_us
)
715 struct tcp_sock
*tp
= tcp_sk(sk
);
716 long m
= mrtt_us
; /* RTT */
717 u32 srtt
= tp
->srtt_us
;
719 /* The following amusing code comes from Jacobson's
720 * article in SIGCOMM '88. Note that rtt and mdev
721 * are scaled versions of rtt and mean deviation.
722 * This is designed to be as fast as possible
723 * m stands for "measurement".
725 * On a 1990 paper the rto value is changed to:
726 * RTO = rtt + 4 * mdev
728 * Funny. This algorithm seems to be very broken.
729 * These formulae increase RTO, when it should be decreased, increase
730 * too slowly, when it should be increased quickly, decrease too quickly
731 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
732 * does not matter how to _calculate_ it. Seems, it was trap
733 * that VJ failed to avoid. 8)
736 m
-= (srtt
>> 3); /* m is now error in rtt est */
737 srtt
+= m
; /* rtt = 7/8 rtt + 1/8 new */
739 m
= -m
; /* m is now abs(error) */
740 m
-= (tp
->mdev_us
>> 2); /* similar update on mdev */
741 /* This is similar to one of Eifel findings.
742 * Eifel blocks mdev updates when rtt decreases.
743 * This solution is a bit different: we use finer gain
744 * for mdev in this case (alpha*beta).
745 * Like Eifel it also prevents growth of rto,
746 * but also it limits too fast rto decreases,
747 * happening in pure Eifel.
752 m
-= (tp
->mdev_us
>> 2); /* similar update on mdev */
754 tp
->mdev_us
+= m
; /* mdev = 3/4 mdev + 1/4 new */
755 if (tp
->mdev_us
> tp
->mdev_max_us
) {
756 tp
->mdev_max_us
= tp
->mdev_us
;
757 if (tp
->mdev_max_us
> tp
->rttvar_us
)
758 tp
->rttvar_us
= tp
->mdev_max_us
;
760 if (after(tp
->snd_una
, tp
->rtt_seq
)) {
761 if (tp
->mdev_max_us
< tp
->rttvar_us
)
762 tp
->rttvar_us
-= (tp
->rttvar_us
- tp
->mdev_max_us
) >> 2;
763 tp
->rtt_seq
= tp
->snd_nxt
;
764 tp
->mdev_max_us
= tcp_rto_min_us(sk
);
767 /* no previous measure. */
768 srtt
= m
<< 3; /* take the measured time to be rtt */
769 tp
->mdev_us
= m
<< 1; /* make sure rto = 3*rtt */
770 tp
->rttvar_us
= max(tp
->mdev_us
, tcp_rto_min_us(sk
));
771 tp
->mdev_max_us
= tp
->rttvar_us
;
772 tp
->rtt_seq
= tp
->snd_nxt
;
774 tp
->srtt_us
= max(1U, srtt
);
777 /* Set the sk_pacing_rate to allow proper sizing of TSO packets.
778 * Note: TCP stack does not yet implement pacing.
779 * FQ packet scheduler can be used to implement cheap but effective
780 * TCP pacing, to smooth the burst on large writes when packets
781 * in flight is significantly lower than cwnd (or rwin)
783 int sysctl_tcp_pacing_ss_ratio __read_mostly
= 200;
784 int sysctl_tcp_pacing_ca_ratio __read_mostly
= 120;
786 static void tcp_update_pacing_rate(struct sock
*sk
)
788 const struct tcp_sock
*tp
= tcp_sk(sk
);
791 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
792 rate
= (u64
)tp
->mss_cache
* ((USEC_PER_SEC
/ 100) << 3);
794 /* current rate is (cwnd * mss) / srtt
795 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
796 * In Congestion Avoidance phase, set it to 120 % the current rate.
798 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
799 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
800 * end of slow start and should slow down.
802 if (tp
->snd_cwnd
< tp
->snd_ssthresh
/ 2)
803 rate
*= sysctl_tcp_pacing_ss_ratio
;
805 rate
*= sysctl_tcp_pacing_ca_ratio
;
807 rate
*= max(tp
->snd_cwnd
, tp
->packets_out
);
809 if (likely(tp
->srtt_us
))
810 do_div(rate
, tp
->srtt_us
);
812 /* ACCESS_ONCE() is needed because sch_fq fetches sk_pacing_rate
813 * without any lock. We want to make sure compiler wont store
814 * intermediate values in this location.
816 ACCESS_ONCE(sk
->sk_pacing_rate
) = min_t(u64
, rate
,
817 sk
->sk_max_pacing_rate
);
820 /* Calculate rto without backoff. This is the second half of Van Jacobson's
821 * routine referred to above.
823 static void tcp_set_rto(struct sock
*sk
)
825 const struct tcp_sock
*tp
= tcp_sk(sk
);
826 /* Old crap is replaced with new one. 8)
829 * 1. If rtt variance happened to be less 50msec, it is hallucination.
830 * It cannot be less due to utterly erratic ACK generation made
831 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
832 * to do with delayed acks, because at cwnd>2 true delack timeout
833 * is invisible. Actually, Linux-2.4 also generates erratic
834 * ACKs in some circumstances.
836 inet_csk(sk
)->icsk_rto
= __tcp_set_rto(tp
);
838 /* 2. Fixups made earlier cannot be right.
839 * If we do not estimate RTO correctly without them,
840 * all the algo is pure shit and should be replaced
841 * with correct one. It is exactly, which we pretend to do.
844 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
845 * guarantees that rto is higher.
850 __u32
tcp_init_cwnd(const struct tcp_sock
*tp
, const struct dst_entry
*dst
)
852 __u32 cwnd
= (dst
? dst_metric(dst
, RTAX_INITCWND
) : 0);
855 cwnd
= TCP_INIT_CWND
;
856 return min_t(__u32
, cwnd
, tp
->snd_cwnd_clamp
);
860 * Packet counting of FACK is based on in-order assumptions, therefore TCP
861 * disables it when reordering is detected
863 void tcp_disable_fack(struct tcp_sock
*tp
)
865 /* RFC3517 uses different metric in lost marker => reset on change */
867 tp
->lost_skb_hint
= NULL
;
868 tp
->rx_opt
.sack_ok
&= ~TCP_FACK_ENABLED
;
871 /* Take a notice that peer is sending D-SACKs */
872 static void tcp_dsack_seen(struct tcp_sock
*tp
)
874 tp
->rx_opt
.sack_ok
|= TCP_DSACK_SEEN
;
877 static void tcp_update_reordering(struct sock
*sk
, const int metric
,
880 struct tcp_sock
*tp
= tcp_sk(sk
);
881 if (metric
> tp
->reordering
) {
884 tp
->reordering
= min(sysctl_tcp_max_reordering
, metric
);
886 /* This exciting event is worth to be remembered. 8) */
888 mib_idx
= LINUX_MIB_TCPTSREORDER
;
889 else if (tcp_is_reno(tp
))
890 mib_idx
= LINUX_MIB_TCPRENOREORDER
;
891 else if (tcp_is_fack(tp
))
892 mib_idx
= LINUX_MIB_TCPFACKREORDER
;
894 mib_idx
= LINUX_MIB_TCPSACKREORDER
;
896 NET_INC_STATS(sock_net(sk
), mib_idx
);
897 #if FASTRETRANS_DEBUG > 1
898 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
899 tp
->rx_opt
.sack_ok
, inet_csk(sk
)->icsk_ca_state
,
903 tp
->undo_marker
? tp
->undo_retrans
: 0);
905 tcp_disable_fack(tp
);
911 /* This must be called before lost_out is incremented */
912 static void tcp_verify_retransmit_hint(struct tcp_sock
*tp
, struct sk_buff
*skb
)
914 if (!tp
->retransmit_skb_hint
||
915 before(TCP_SKB_CB(skb
)->seq
,
916 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
917 tp
->retransmit_skb_hint
= skb
;
920 /* Sum the number of packets on the wire we have marked as lost.
921 * There are two cases we care about here:
922 * a) Packet hasn't been marked lost (nor retransmitted),
923 * and this is the first loss.
924 * b) Packet has been marked both lost and retransmitted,
925 * and this means we think it was lost again.
927 static void tcp_sum_lost(struct tcp_sock
*tp
, struct sk_buff
*skb
)
929 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
931 if (!(sacked
& TCPCB_LOST
) ||
932 ((sacked
& TCPCB_LOST
) && (sacked
& TCPCB_SACKED_RETRANS
)))
933 tp
->lost
+= tcp_skb_pcount(skb
);
936 static void tcp_skb_mark_lost(struct tcp_sock
*tp
, struct sk_buff
*skb
)
938 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
939 tcp_verify_retransmit_hint(tp
, skb
);
941 tp
->lost_out
+= tcp_skb_pcount(skb
);
942 tcp_sum_lost(tp
, skb
);
943 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
947 void tcp_skb_mark_lost_uncond_verify(struct tcp_sock
*tp
, struct sk_buff
*skb
)
949 tcp_verify_retransmit_hint(tp
, skb
);
951 tcp_sum_lost(tp
, skb
);
952 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
953 tp
->lost_out
+= tcp_skb_pcount(skb
);
954 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
958 /* This procedure tags the retransmission queue when SACKs arrive.
960 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
961 * Packets in queue with these bits set are counted in variables
962 * sacked_out, retrans_out and lost_out, correspondingly.
964 * Valid combinations are:
965 * Tag InFlight Description
966 * 0 1 - orig segment is in flight.
967 * S 0 - nothing flies, orig reached receiver.
968 * L 0 - nothing flies, orig lost by net.
969 * R 2 - both orig and retransmit are in flight.
970 * L|R 1 - orig is lost, retransmit is in flight.
971 * S|R 1 - orig reached receiver, retrans is still in flight.
972 * (L|S|R is logically valid, it could occur when L|R is sacked,
973 * but it is equivalent to plain S and code short-curcuits it to S.
974 * L|S is logically invalid, it would mean -1 packet in flight 8))
976 * These 6 states form finite state machine, controlled by the following events:
977 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
978 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
979 * 3. Loss detection event of two flavors:
980 * A. Scoreboard estimator decided the packet is lost.
981 * A'. Reno "three dupacks" marks head of queue lost.
982 * A''. Its FACK modification, head until snd.fack is lost.
983 * B. SACK arrives sacking SND.NXT at the moment, when the
984 * segment was retransmitted.
985 * 4. D-SACK added new rule: D-SACK changes any tag to S.
987 * It is pleasant to note, that state diagram turns out to be commutative,
988 * so that we are allowed not to be bothered by order of our actions,
989 * when multiple events arrive simultaneously. (see the function below).
991 * Reordering detection.
992 * --------------------
993 * Reordering metric is maximal distance, which a packet can be displaced
994 * in packet stream. With SACKs we can estimate it:
996 * 1. SACK fills old hole and the corresponding segment was not
997 * ever retransmitted -> reordering. Alas, we cannot use it
998 * when segment was retransmitted.
999 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1000 * for retransmitted and already SACKed segment -> reordering..
1001 * Both of these heuristics are not used in Loss state, when we cannot
1002 * account for retransmits accurately.
1004 * SACK block validation.
1005 * ----------------------
1007 * SACK block range validation checks that the received SACK block fits to
1008 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1009 * Note that SND.UNA is not included to the range though being valid because
1010 * it means that the receiver is rather inconsistent with itself reporting
1011 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1012 * perfectly valid, however, in light of RFC2018 which explicitly states
1013 * that "SACK block MUST reflect the newest segment. Even if the newest
1014 * segment is going to be discarded ...", not that it looks very clever
1015 * in case of head skb. Due to potentional receiver driven attacks, we
1016 * choose to avoid immediate execution of a walk in write queue due to
1017 * reneging and defer head skb's loss recovery to standard loss recovery
1018 * procedure that will eventually trigger (nothing forbids us doing this).
1020 * Implements also blockage to start_seq wrap-around. Problem lies in the
1021 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1022 * there's no guarantee that it will be before snd_nxt (n). The problem
1023 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1026 * <- outs wnd -> <- wrapzone ->
1027 * u e n u_w e_w s n_w
1029 * |<------------+------+----- TCP seqno space --------------+---------->|
1030 * ...-- <2^31 ->| |<--------...
1031 * ...---- >2^31 ------>| |<--------...
1033 * Current code wouldn't be vulnerable but it's better still to discard such
1034 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1035 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1036 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1037 * equal to the ideal case (infinite seqno space without wrap caused issues).
1039 * With D-SACK the lower bound is extended to cover sequence space below
1040 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1041 * again, D-SACK block must not to go across snd_una (for the same reason as
1042 * for the normal SACK blocks, explained above). But there all simplicity
1043 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1044 * fully below undo_marker they do not affect behavior in anyway and can
1045 * therefore be safely ignored. In rare cases (which are more or less
1046 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1047 * fragmentation and packet reordering past skb's retransmission. To consider
1048 * them correctly, the acceptable range must be extended even more though
1049 * the exact amount is rather hard to quantify. However, tp->max_window can
1050 * be used as an exaggerated estimate.
1052 static bool tcp_is_sackblock_valid(struct tcp_sock
*tp
, bool is_dsack
,
1053 u32 start_seq
, u32 end_seq
)
1055 /* Too far in future, or reversed (interpretation is ambiguous) */
1056 if (after(end_seq
, tp
->snd_nxt
) || !before(start_seq
, end_seq
))
1059 /* Nasty start_seq wrap-around check (see comments above) */
1060 if (!before(start_seq
, tp
->snd_nxt
))
1063 /* In outstanding window? ...This is valid exit for D-SACKs too.
1064 * start_seq == snd_una is non-sensical (see comments above)
1066 if (after(start_seq
, tp
->snd_una
))
1069 if (!is_dsack
|| !tp
->undo_marker
)
1072 /* ...Then it's D-SACK, and must reside below snd_una completely */
1073 if (after(end_seq
, tp
->snd_una
))
1076 if (!before(start_seq
, tp
->undo_marker
))
1080 if (!after(end_seq
, tp
->undo_marker
))
1083 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1084 * start_seq < undo_marker and end_seq >= undo_marker.
1086 return !before(start_seq
, end_seq
- tp
->max_window
);
1089 static bool tcp_check_dsack(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1090 struct tcp_sack_block_wire
*sp
, int num_sacks
,
1093 struct tcp_sock
*tp
= tcp_sk(sk
);
1094 u32 start_seq_0
= get_unaligned_be32(&sp
[0].start_seq
);
1095 u32 end_seq_0
= get_unaligned_be32(&sp
[0].end_seq
);
1096 bool dup_sack
= false;
1098 if (before(start_seq_0
, TCP_SKB_CB(ack_skb
)->ack_seq
)) {
1101 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPDSACKRECV
);
1102 } else if (num_sacks
> 1) {
1103 u32 end_seq_1
= get_unaligned_be32(&sp
[1].end_seq
);
1104 u32 start_seq_1
= get_unaligned_be32(&sp
[1].start_seq
);
1106 if (!after(end_seq_0
, end_seq_1
) &&
1107 !before(start_seq_0
, start_seq_1
)) {
1110 NET_INC_STATS(sock_net(sk
),
1111 LINUX_MIB_TCPDSACKOFORECV
);
1115 /* D-SACK for already forgotten data... Do dumb counting. */
1116 if (dup_sack
&& tp
->undo_marker
&& tp
->undo_retrans
> 0 &&
1117 !after(end_seq_0
, prior_snd_una
) &&
1118 after(end_seq_0
, tp
->undo_marker
))
1124 struct tcp_sacktag_state
{
1127 /* Timestamps for earliest and latest never-retransmitted segment
1128 * that was SACKed. RTO needs the earliest RTT to stay conservative,
1129 * but congestion control should still get an accurate delay signal.
1131 struct skb_mstamp first_sackt
;
1132 struct skb_mstamp last_sackt
;
1133 struct skb_mstamp ack_time
; /* Timestamp when the S/ACK was received */
1134 struct rate_sample
*rate
;
1138 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1139 * the incoming SACK may not exactly match but we can find smaller MSS
1140 * aligned portion of it that matches. Therefore we might need to fragment
1141 * which may fail and creates some hassle (caller must handle error case
1144 * FIXME: this could be merged to shift decision code
1146 static int tcp_match_skb_to_sack(struct sock
*sk
, struct sk_buff
*skb
,
1147 u32 start_seq
, u32 end_seq
)
1151 unsigned int pkt_len
;
1154 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1155 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1157 if (tcp_skb_pcount(skb
) > 1 && !in_sack
&&
1158 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
1159 mss
= tcp_skb_mss(skb
);
1160 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1163 pkt_len
= start_seq
- TCP_SKB_CB(skb
)->seq
;
1167 pkt_len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1172 /* Round if necessary so that SACKs cover only full MSSes
1173 * and/or the remaining small portion (if present)
1175 if (pkt_len
> mss
) {
1176 unsigned int new_len
= (pkt_len
/ mss
) * mss
;
1177 if (!in_sack
&& new_len
< pkt_len
) {
1179 if (new_len
>= skb
->len
)
1184 err
= tcp_fragment(sk
, skb
, pkt_len
, mss
, GFP_ATOMIC
);
1192 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1193 static u8
tcp_sacktag_one(struct sock
*sk
,
1194 struct tcp_sacktag_state
*state
, u8 sacked
,
1195 u32 start_seq
, u32 end_seq
,
1196 int dup_sack
, int pcount
,
1197 const struct skb_mstamp
*xmit_time
)
1199 struct tcp_sock
*tp
= tcp_sk(sk
);
1200 int fack_count
= state
->fack_count
;
1202 /* Account D-SACK for retransmitted packet. */
1203 if (dup_sack
&& (sacked
& TCPCB_RETRANS
)) {
1204 if (tp
->undo_marker
&& tp
->undo_retrans
> 0 &&
1205 after(end_seq
, tp
->undo_marker
))
1207 if (sacked
& TCPCB_SACKED_ACKED
)
1208 state
->reord
= min(fack_count
, state
->reord
);
1211 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1212 if (!after(end_seq
, tp
->snd_una
))
1215 if (!(sacked
& TCPCB_SACKED_ACKED
)) {
1216 tcp_rack_advance(tp
, sacked
, end_seq
,
1217 xmit_time
, &state
->ack_time
);
1219 if (sacked
& TCPCB_SACKED_RETRANS
) {
1220 /* If the segment is not tagged as lost,
1221 * we do not clear RETRANS, believing
1222 * that retransmission is still in flight.
1224 if (sacked
& TCPCB_LOST
) {
1225 sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1226 tp
->lost_out
-= pcount
;
1227 tp
->retrans_out
-= pcount
;
1230 if (!(sacked
& TCPCB_RETRANS
)) {
1231 /* New sack for not retransmitted frame,
1232 * which was in hole. It is reordering.
1234 if (before(start_seq
,
1235 tcp_highest_sack_seq(tp
)))
1236 state
->reord
= min(fack_count
,
1238 if (!after(end_seq
, tp
->high_seq
))
1239 state
->flag
|= FLAG_ORIG_SACK_ACKED
;
1240 if (state
->first_sackt
.v64
== 0)
1241 state
->first_sackt
= *xmit_time
;
1242 state
->last_sackt
= *xmit_time
;
1245 if (sacked
& TCPCB_LOST
) {
1246 sacked
&= ~TCPCB_LOST
;
1247 tp
->lost_out
-= pcount
;
1251 sacked
|= TCPCB_SACKED_ACKED
;
1252 state
->flag
|= FLAG_DATA_SACKED
;
1253 tp
->sacked_out
+= pcount
;
1254 tp
->delivered
+= pcount
; /* Out-of-order packets delivered */
1256 fack_count
+= pcount
;
1258 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1259 if (!tcp_is_fack(tp
) && tp
->lost_skb_hint
&&
1260 before(start_seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
))
1261 tp
->lost_cnt_hint
+= pcount
;
1263 if (fack_count
> tp
->fackets_out
)
1264 tp
->fackets_out
= fack_count
;
1267 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1268 * frames and clear it. undo_retrans is decreased above, L|R frames
1269 * are accounted above as well.
1271 if (dup_sack
&& (sacked
& TCPCB_SACKED_RETRANS
)) {
1272 sacked
&= ~TCPCB_SACKED_RETRANS
;
1273 tp
->retrans_out
-= pcount
;
1279 /* Shift newly-SACKed bytes from this skb to the immediately previous
1280 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1282 static bool tcp_shifted_skb(struct sock
*sk
, struct sk_buff
*skb
,
1283 struct tcp_sacktag_state
*state
,
1284 unsigned int pcount
, int shifted
, int mss
,
1287 struct tcp_sock
*tp
= tcp_sk(sk
);
1288 struct sk_buff
*prev
= tcp_write_queue_prev(sk
, skb
);
1289 u32 start_seq
= TCP_SKB_CB(skb
)->seq
; /* start of newly-SACKed */
1290 u32 end_seq
= start_seq
+ shifted
; /* end of newly-SACKed */
1294 /* Adjust counters and hints for the newly sacked sequence
1295 * range but discard the return value since prev is already
1296 * marked. We must tag the range first because the seq
1297 * advancement below implicitly advances
1298 * tcp_highest_sack_seq() when skb is highest_sack.
1300 tcp_sacktag_one(sk
, state
, TCP_SKB_CB(skb
)->sacked
,
1301 start_seq
, end_seq
, dup_sack
, pcount
,
1303 tcp_rate_skb_delivered(sk
, skb
, state
->rate
);
1305 if (skb
== tp
->lost_skb_hint
)
1306 tp
->lost_cnt_hint
+= pcount
;
1308 TCP_SKB_CB(prev
)->end_seq
+= shifted
;
1309 TCP_SKB_CB(skb
)->seq
+= shifted
;
1311 tcp_skb_pcount_add(prev
, pcount
);
1312 BUG_ON(tcp_skb_pcount(skb
) < pcount
);
1313 tcp_skb_pcount_add(skb
, -pcount
);
1315 /* When we're adding to gso_segs == 1, gso_size will be zero,
1316 * in theory this shouldn't be necessary but as long as DSACK
1317 * code can come after this skb later on it's better to keep
1318 * setting gso_size to something.
1320 if (!TCP_SKB_CB(prev
)->tcp_gso_size
)
1321 TCP_SKB_CB(prev
)->tcp_gso_size
= mss
;
1323 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1324 if (tcp_skb_pcount(skb
) <= 1)
1325 TCP_SKB_CB(skb
)->tcp_gso_size
= 0;
1327 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1328 TCP_SKB_CB(prev
)->sacked
|= (TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
);
1331 BUG_ON(!tcp_skb_pcount(skb
));
1332 NET_INC_STATS(sock_net(sk
), LINUX_MIB_SACKSHIFTED
);
1336 /* Whole SKB was eaten :-) */
1338 if (skb
== tp
->retransmit_skb_hint
)
1339 tp
->retransmit_skb_hint
= prev
;
1340 if (skb
== tp
->lost_skb_hint
) {
1341 tp
->lost_skb_hint
= prev
;
1342 tp
->lost_cnt_hint
-= tcp_skb_pcount(prev
);
1345 TCP_SKB_CB(prev
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
;
1346 TCP_SKB_CB(prev
)->eor
= TCP_SKB_CB(skb
)->eor
;
1347 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
1348 TCP_SKB_CB(prev
)->end_seq
++;
1350 if (skb
== tcp_highest_sack(sk
))
1351 tcp_advance_highest_sack(sk
, skb
);
1353 tcp_skb_collapse_tstamp(prev
, skb
);
1354 if (unlikely(TCP_SKB_CB(prev
)->tx
.delivered_mstamp
.v64
))
1355 TCP_SKB_CB(prev
)->tx
.delivered_mstamp
.v64
= 0;
1357 tcp_unlink_write_queue(skb
, sk
);
1358 sk_wmem_free_skb(sk
, skb
);
1360 NET_INC_STATS(sock_net(sk
), LINUX_MIB_SACKMERGED
);
1365 /* I wish gso_size would have a bit more sane initialization than
1366 * something-or-zero which complicates things
1368 static int tcp_skb_seglen(const struct sk_buff
*skb
)
1370 return tcp_skb_pcount(skb
) == 1 ? skb
->len
: tcp_skb_mss(skb
);
1373 /* Shifting pages past head area doesn't work */
1374 static int skb_can_shift(const struct sk_buff
*skb
)
1376 return !skb_headlen(skb
) && skb_is_nonlinear(skb
);
1379 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1382 static struct sk_buff
*tcp_shift_skb_data(struct sock
*sk
, struct sk_buff
*skb
,
1383 struct tcp_sacktag_state
*state
,
1384 u32 start_seq
, u32 end_seq
,
1387 struct tcp_sock
*tp
= tcp_sk(sk
);
1388 struct sk_buff
*prev
;
1394 if (!sk_can_gso(sk
))
1397 /* Normally R but no L won't result in plain S */
1399 (TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_RETRANS
)) == TCPCB_SACKED_RETRANS
)
1401 if (!skb_can_shift(skb
))
1403 /* This frame is about to be dropped (was ACKed). */
1404 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1407 /* Can only happen with delayed DSACK + discard craziness */
1408 if (unlikely(skb
== tcp_write_queue_head(sk
)))
1410 prev
= tcp_write_queue_prev(sk
, skb
);
1412 if ((TCP_SKB_CB(prev
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
)
1415 if (!tcp_skb_can_collapse_to(prev
))
1418 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1419 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1423 pcount
= tcp_skb_pcount(skb
);
1424 mss
= tcp_skb_seglen(skb
);
1426 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1427 * drop this restriction as unnecessary
1429 if (mss
!= tcp_skb_seglen(prev
))
1432 if (!after(TCP_SKB_CB(skb
)->end_seq
, start_seq
))
1434 /* CHECKME: This is non-MSS split case only?, this will
1435 * cause skipped skbs due to advancing loop btw, original
1436 * has that feature too
1438 if (tcp_skb_pcount(skb
) <= 1)
1441 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1443 /* TODO: head merge to next could be attempted here
1444 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1445 * though it might not be worth of the additional hassle
1447 * ...we can probably just fallback to what was done
1448 * previously. We could try merging non-SACKed ones
1449 * as well but it probably isn't going to buy off
1450 * because later SACKs might again split them, and
1451 * it would make skb timestamp tracking considerably
1457 len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1459 BUG_ON(len
> skb
->len
);
1461 /* MSS boundaries should be honoured or else pcount will
1462 * severely break even though it makes things bit trickier.
1463 * Optimize common case to avoid most of the divides
1465 mss
= tcp_skb_mss(skb
);
1467 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1468 * drop this restriction as unnecessary
1470 if (mss
!= tcp_skb_seglen(prev
))
1475 } else if (len
< mss
) {
1483 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1484 if (!after(TCP_SKB_CB(skb
)->seq
+ len
, tp
->snd_una
))
1487 if (!skb_shift(prev
, skb
, len
))
1489 if (!tcp_shifted_skb(sk
, skb
, state
, pcount
, len
, mss
, dup_sack
))
1492 /* Hole filled allows collapsing with the next as well, this is very
1493 * useful when hole on every nth skb pattern happens
1495 if (prev
== tcp_write_queue_tail(sk
))
1497 skb
= tcp_write_queue_next(sk
, prev
);
1499 if (!skb_can_shift(skb
) ||
1500 (skb
== tcp_send_head(sk
)) ||
1501 ((TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
) ||
1502 (mss
!= tcp_skb_seglen(skb
)))
1506 if (skb_shift(prev
, skb
, len
)) {
1507 pcount
+= tcp_skb_pcount(skb
);
1508 tcp_shifted_skb(sk
, skb
, state
, tcp_skb_pcount(skb
), len
, mss
, 0);
1512 state
->fack_count
+= pcount
;
1519 NET_INC_STATS(sock_net(sk
), LINUX_MIB_SACKSHIFTFALLBACK
);
1523 static struct sk_buff
*tcp_sacktag_walk(struct sk_buff
*skb
, struct sock
*sk
,
1524 struct tcp_sack_block
*next_dup
,
1525 struct tcp_sacktag_state
*state
,
1526 u32 start_seq
, u32 end_seq
,
1529 struct tcp_sock
*tp
= tcp_sk(sk
);
1530 struct sk_buff
*tmp
;
1532 tcp_for_write_queue_from(skb
, sk
) {
1534 bool dup_sack
= dup_sack_in
;
1536 if (skb
== tcp_send_head(sk
))
1539 /* queue is in-order => we can short-circuit the walk early */
1540 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1544 before(TCP_SKB_CB(skb
)->seq
, next_dup
->end_seq
)) {
1545 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1546 next_dup
->start_seq
,
1552 /* skb reference here is a bit tricky to get right, since
1553 * shifting can eat and free both this skb and the next,
1554 * so not even _safe variant of the loop is enough.
1557 tmp
= tcp_shift_skb_data(sk
, skb
, state
,
1558 start_seq
, end_seq
, dup_sack
);
1567 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1573 if (unlikely(in_sack
< 0))
1577 TCP_SKB_CB(skb
)->sacked
=
1580 TCP_SKB_CB(skb
)->sacked
,
1581 TCP_SKB_CB(skb
)->seq
,
1582 TCP_SKB_CB(skb
)->end_seq
,
1584 tcp_skb_pcount(skb
),
1586 tcp_rate_skb_delivered(sk
, skb
, state
->rate
);
1588 if (!before(TCP_SKB_CB(skb
)->seq
,
1589 tcp_highest_sack_seq(tp
)))
1590 tcp_advance_highest_sack(sk
, skb
);
1593 state
->fack_count
+= tcp_skb_pcount(skb
);
1598 /* Avoid all extra work that is being done by sacktag while walking in
1601 static struct sk_buff
*tcp_sacktag_skip(struct sk_buff
*skb
, struct sock
*sk
,
1602 struct tcp_sacktag_state
*state
,
1605 tcp_for_write_queue_from(skb
, sk
) {
1606 if (skb
== tcp_send_head(sk
))
1609 if (after(TCP_SKB_CB(skb
)->end_seq
, skip_to_seq
))
1612 state
->fack_count
+= tcp_skb_pcount(skb
);
1617 static struct sk_buff
*tcp_maybe_skipping_dsack(struct sk_buff
*skb
,
1619 struct tcp_sack_block
*next_dup
,
1620 struct tcp_sacktag_state
*state
,
1626 if (before(next_dup
->start_seq
, skip_to_seq
)) {
1627 skb
= tcp_sacktag_skip(skb
, sk
, state
, next_dup
->start_seq
);
1628 skb
= tcp_sacktag_walk(skb
, sk
, NULL
, state
,
1629 next_dup
->start_seq
, next_dup
->end_seq
,
1636 static int tcp_sack_cache_ok(const struct tcp_sock
*tp
, const struct tcp_sack_block
*cache
)
1638 return cache
< tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1642 tcp_sacktag_write_queue(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1643 u32 prior_snd_una
, struct tcp_sacktag_state
*state
)
1645 struct tcp_sock
*tp
= tcp_sk(sk
);
1646 const unsigned char *ptr
= (skb_transport_header(ack_skb
) +
1647 TCP_SKB_CB(ack_skb
)->sacked
);
1648 struct tcp_sack_block_wire
*sp_wire
= (struct tcp_sack_block_wire
*)(ptr
+2);
1649 struct tcp_sack_block sp
[TCP_NUM_SACKS
];
1650 struct tcp_sack_block
*cache
;
1651 struct sk_buff
*skb
;
1652 int num_sacks
= min(TCP_NUM_SACKS
, (ptr
[1] - TCPOLEN_SACK_BASE
) >> 3);
1654 bool found_dup_sack
= false;
1656 int first_sack_index
;
1659 state
->reord
= tp
->packets_out
;
1661 if (!tp
->sacked_out
) {
1662 if (WARN_ON(tp
->fackets_out
))
1663 tp
->fackets_out
= 0;
1664 tcp_highest_sack_reset(sk
);
1667 found_dup_sack
= tcp_check_dsack(sk
, ack_skb
, sp_wire
,
1668 num_sacks
, prior_snd_una
);
1669 if (found_dup_sack
) {
1670 state
->flag
|= FLAG_DSACKING_ACK
;
1671 tp
->delivered
++; /* A spurious retransmission is delivered */
1674 /* Eliminate too old ACKs, but take into
1675 * account more or less fresh ones, they can
1676 * contain valid SACK info.
1678 if (before(TCP_SKB_CB(ack_skb
)->ack_seq
, prior_snd_una
- tp
->max_window
))
1681 if (!tp
->packets_out
)
1685 first_sack_index
= 0;
1686 for (i
= 0; i
< num_sacks
; i
++) {
1687 bool dup_sack
= !i
&& found_dup_sack
;
1689 sp
[used_sacks
].start_seq
= get_unaligned_be32(&sp_wire
[i
].start_seq
);
1690 sp
[used_sacks
].end_seq
= get_unaligned_be32(&sp_wire
[i
].end_seq
);
1692 if (!tcp_is_sackblock_valid(tp
, dup_sack
,
1693 sp
[used_sacks
].start_seq
,
1694 sp
[used_sacks
].end_seq
)) {
1698 if (!tp
->undo_marker
)
1699 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDNOUNDO
;
1701 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDOLD
;
1703 /* Don't count olds caused by ACK reordering */
1704 if ((TCP_SKB_CB(ack_skb
)->ack_seq
!= tp
->snd_una
) &&
1705 !after(sp
[used_sacks
].end_seq
, tp
->snd_una
))
1707 mib_idx
= LINUX_MIB_TCPSACKDISCARD
;
1710 NET_INC_STATS(sock_net(sk
), mib_idx
);
1712 first_sack_index
= -1;
1716 /* Ignore very old stuff early */
1717 if (!after(sp
[used_sacks
].end_seq
, prior_snd_una
))
1723 /* order SACK blocks to allow in order walk of the retrans queue */
1724 for (i
= used_sacks
- 1; i
> 0; i
--) {
1725 for (j
= 0; j
< i
; j
++) {
1726 if (after(sp
[j
].start_seq
, sp
[j
+ 1].start_seq
)) {
1727 swap(sp
[j
], sp
[j
+ 1]);
1729 /* Track where the first SACK block goes to */
1730 if (j
== first_sack_index
)
1731 first_sack_index
= j
+ 1;
1736 skb
= tcp_write_queue_head(sk
);
1737 state
->fack_count
= 0;
1740 if (!tp
->sacked_out
) {
1741 /* It's already past, so skip checking against it */
1742 cache
= tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1744 cache
= tp
->recv_sack_cache
;
1745 /* Skip empty blocks in at head of the cache */
1746 while (tcp_sack_cache_ok(tp
, cache
) && !cache
->start_seq
&&
1751 while (i
< used_sacks
) {
1752 u32 start_seq
= sp
[i
].start_seq
;
1753 u32 end_seq
= sp
[i
].end_seq
;
1754 bool dup_sack
= (found_dup_sack
&& (i
== first_sack_index
));
1755 struct tcp_sack_block
*next_dup
= NULL
;
1757 if (found_dup_sack
&& ((i
+ 1) == first_sack_index
))
1758 next_dup
= &sp
[i
+ 1];
1760 /* Skip too early cached blocks */
1761 while (tcp_sack_cache_ok(tp
, cache
) &&
1762 !before(start_seq
, cache
->end_seq
))
1765 /* Can skip some work by looking recv_sack_cache? */
1766 if (tcp_sack_cache_ok(tp
, cache
) && !dup_sack
&&
1767 after(end_seq
, cache
->start_seq
)) {
1770 if (before(start_seq
, cache
->start_seq
)) {
1771 skb
= tcp_sacktag_skip(skb
, sk
, state
,
1773 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
,
1780 /* Rest of the block already fully processed? */
1781 if (!after(end_seq
, cache
->end_seq
))
1784 skb
= tcp_maybe_skipping_dsack(skb
, sk
, next_dup
,
1788 /* ...tail remains todo... */
1789 if (tcp_highest_sack_seq(tp
) == cache
->end_seq
) {
1790 /* ...but better entrypoint exists! */
1791 skb
= tcp_highest_sack(sk
);
1794 state
->fack_count
= tp
->fackets_out
;
1799 skb
= tcp_sacktag_skip(skb
, sk
, state
, cache
->end_seq
);
1800 /* Check overlap against next cached too (past this one already) */
1805 if (!before(start_seq
, tcp_highest_sack_seq(tp
))) {
1806 skb
= tcp_highest_sack(sk
);
1809 state
->fack_count
= tp
->fackets_out
;
1811 skb
= tcp_sacktag_skip(skb
, sk
, state
, start_seq
);
1814 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
, state
,
1815 start_seq
, end_seq
, dup_sack
);
1821 /* Clear the head of the cache sack blocks so we can skip it next time */
1822 for (i
= 0; i
< ARRAY_SIZE(tp
->recv_sack_cache
) - used_sacks
; i
++) {
1823 tp
->recv_sack_cache
[i
].start_seq
= 0;
1824 tp
->recv_sack_cache
[i
].end_seq
= 0;
1826 for (j
= 0; j
< used_sacks
; j
++)
1827 tp
->recv_sack_cache
[i
++] = sp
[j
];
1829 if ((state
->reord
< tp
->fackets_out
) &&
1830 ((inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Loss
) || tp
->undo_marker
))
1831 tcp_update_reordering(sk
, tp
->fackets_out
- state
->reord
, 0);
1833 tcp_verify_left_out(tp
);
1836 #if FASTRETRANS_DEBUG > 0
1837 WARN_ON((int)tp
->sacked_out
< 0);
1838 WARN_ON((int)tp
->lost_out
< 0);
1839 WARN_ON((int)tp
->retrans_out
< 0);
1840 WARN_ON((int)tcp_packets_in_flight(tp
) < 0);
1845 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1846 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1848 static bool tcp_limit_reno_sacked(struct tcp_sock
*tp
)
1852 holes
= max(tp
->lost_out
, 1U);
1853 holes
= min(holes
, tp
->packets_out
);
1855 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1856 tp
->sacked_out
= tp
->packets_out
- holes
;
1862 /* If we receive more dupacks than we expected counting segments
1863 * in assumption of absent reordering, interpret this as reordering.
1864 * The only another reason could be bug in receiver TCP.
1866 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1868 struct tcp_sock
*tp
= tcp_sk(sk
);
1869 if (tcp_limit_reno_sacked(tp
))
1870 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1873 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1875 static void tcp_add_reno_sack(struct sock
*sk
)
1877 struct tcp_sock
*tp
= tcp_sk(sk
);
1878 u32 prior_sacked
= tp
->sacked_out
;
1881 tcp_check_reno_reordering(sk
, 0);
1882 if (tp
->sacked_out
> prior_sacked
)
1883 tp
->delivered
++; /* Some out-of-order packet is delivered */
1884 tcp_verify_left_out(tp
);
1887 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1889 static void tcp_remove_reno_sacks(struct sock
*sk
, int acked
)
1891 struct tcp_sock
*tp
= tcp_sk(sk
);
1894 /* One ACK acked hole. The rest eat duplicate ACKs. */
1895 tp
->delivered
+= max_t(int, acked
- tp
->sacked_out
, 1);
1896 if (acked
- 1 >= tp
->sacked_out
)
1899 tp
->sacked_out
-= acked
- 1;
1901 tcp_check_reno_reordering(sk
, acked
);
1902 tcp_verify_left_out(tp
);
1905 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1910 void tcp_clear_retrans(struct tcp_sock
*tp
)
1912 tp
->retrans_out
= 0;
1914 tp
->undo_marker
= 0;
1915 tp
->undo_retrans
= -1;
1916 tp
->fackets_out
= 0;
1920 static inline void tcp_init_undo(struct tcp_sock
*tp
)
1922 tp
->undo_marker
= tp
->snd_una
;
1923 /* Retransmission still in flight may cause DSACKs later. */
1924 tp
->undo_retrans
= tp
->retrans_out
? : -1;
1927 /* Enter Loss state. If we detect SACK reneging, forget all SACK information
1928 * and reset tags completely, otherwise preserve SACKs. If receiver
1929 * dropped its ofo queue, we will know this due to reneging detection.
1931 void tcp_enter_loss(struct sock
*sk
)
1933 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1934 struct tcp_sock
*tp
= tcp_sk(sk
);
1935 struct net
*net
= sock_net(sk
);
1936 struct sk_buff
*skb
;
1937 bool is_reneg
; /* is receiver reneging on SACKs? */
1940 /* Reduce ssthresh if it has not yet been made inside this window. */
1941 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
||
1942 !after(tp
->high_seq
, tp
->snd_una
) ||
1943 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
1944 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1945 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1946 tcp_ca_event(sk
, CA_EVENT_LOSS
);
1950 tp
->snd_cwnd_cnt
= 0;
1951 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1953 tp
->retrans_out
= 0;
1956 if (tcp_is_reno(tp
))
1957 tcp_reset_reno_sack(tp
);
1959 skb
= tcp_write_queue_head(sk
);
1960 is_reneg
= skb
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
);
1962 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPSACKRENEGING
);
1964 tp
->fackets_out
= 0;
1966 tcp_clear_all_retrans_hints(tp
);
1968 tcp_for_write_queue(skb
, sk
) {
1969 if (skb
== tcp_send_head(sk
))
1972 mark_lost
= (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
) ||
1975 tcp_sum_lost(tp
, skb
);
1976 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
1978 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
1979 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1980 tp
->lost_out
+= tcp_skb_pcount(skb
);
1983 tcp_verify_left_out(tp
);
1985 /* Timeout in disordered state after receiving substantial DUPACKs
1986 * suggests that the degree of reordering is over-estimated.
1988 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
&&
1989 tp
->sacked_out
>= net
->ipv4
.sysctl_tcp_reordering
)
1990 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1991 net
->ipv4
.sysctl_tcp_reordering
);
1992 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1993 tp
->high_seq
= tp
->snd_nxt
;
1994 tcp_ecn_queue_cwr(tp
);
1996 /* F-RTO RFC5682 sec 3.1 step 1 mandates to disable F-RTO
1997 * if a previous recovery is underway, otherwise it may incorrectly
1998 * call a timeout spurious if some previously retransmitted packets
1999 * are s/acked (sec 3.2). We do not apply that retriction since
2000 * retransmitted skbs are permanently tagged with TCPCB_EVER_RETRANS
2001 * so FLAG_ORIG_SACK_ACKED is always correct. But we do disable F-RTO
2002 * on PTMU discovery to avoid sending new data.
2004 tp
->frto
= sysctl_tcp_frto
&& !inet_csk(sk
)->icsk_mtup
.probe_size
;
2007 /* If ACK arrived pointing to a remembered SACK, it means that our
2008 * remembered SACKs do not reflect real state of receiver i.e.
2009 * receiver _host_ is heavily congested (or buggy).
2011 * To avoid big spurious retransmission bursts due to transient SACK
2012 * scoreboard oddities that look like reneging, we give the receiver a
2013 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2014 * restore sanity to the SACK scoreboard. If the apparent reneging
2015 * persists until this RTO then we'll clear the SACK scoreboard.
2017 static bool tcp_check_sack_reneging(struct sock
*sk
, int flag
)
2019 if (flag
& FLAG_SACK_RENEGING
) {
2020 struct tcp_sock
*tp
= tcp_sk(sk
);
2021 unsigned long delay
= max(usecs_to_jiffies(tp
->srtt_us
>> 4),
2022 msecs_to_jiffies(10));
2024 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2025 delay
, TCP_RTO_MAX
);
2031 static inline int tcp_fackets_out(const struct tcp_sock
*tp
)
2033 return tcp_is_reno(tp
) ? tp
->sacked_out
+ 1 : tp
->fackets_out
;
2036 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2037 * counter when SACK is enabled (without SACK, sacked_out is used for
2040 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2041 * segments up to the highest received SACK block so far and holes in
2044 * With reordering, holes may still be in flight, so RFC3517 recovery
2045 * uses pure sacked_out (total number of SACKed segments) even though
2046 * it violates the RFC that uses duplicate ACKs, often these are equal
2047 * but when e.g. out-of-window ACKs or packet duplication occurs,
2048 * they differ. Since neither occurs due to loss, TCP should really
2051 static inline int tcp_dupack_heuristics(const struct tcp_sock
*tp
)
2053 return tcp_is_fack(tp
) ? tp
->fackets_out
: tp
->sacked_out
+ 1;
2056 /* Linux NewReno/SACK/FACK/ECN state machine.
2057 * --------------------------------------
2059 * "Open" Normal state, no dubious events, fast path.
2060 * "Disorder" In all the respects it is "Open",
2061 * but requires a bit more attention. It is entered when
2062 * we see some SACKs or dupacks. It is split of "Open"
2063 * mainly to move some processing from fast path to slow one.
2064 * "CWR" CWND was reduced due to some Congestion Notification event.
2065 * It can be ECN, ICMP source quench, local device congestion.
2066 * "Recovery" CWND was reduced, we are fast-retransmitting.
2067 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2069 * tcp_fastretrans_alert() is entered:
2070 * - each incoming ACK, if state is not "Open"
2071 * - when arrived ACK is unusual, namely:
2076 * Counting packets in flight is pretty simple.
2078 * in_flight = packets_out - left_out + retrans_out
2080 * packets_out is SND.NXT-SND.UNA counted in packets.
2082 * retrans_out is number of retransmitted segments.
2084 * left_out is number of segments left network, but not ACKed yet.
2086 * left_out = sacked_out + lost_out
2088 * sacked_out: Packets, which arrived to receiver out of order
2089 * and hence not ACKed. With SACKs this number is simply
2090 * amount of SACKed data. Even without SACKs
2091 * it is easy to give pretty reliable estimate of this number,
2092 * counting duplicate ACKs.
2094 * lost_out: Packets lost by network. TCP has no explicit
2095 * "loss notification" feedback from network (for now).
2096 * It means that this number can be only _guessed_.
2097 * Actually, it is the heuristics to predict lossage that
2098 * distinguishes different algorithms.
2100 * F.e. after RTO, when all the queue is considered as lost,
2101 * lost_out = packets_out and in_flight = retrans_out.
2103 * Essentially, we have now a few algorithms detecting
2106 * If the receiver supports SACK:
2108 * RFC6675/3517: It is the conventional algorithm. A packet is
2109 * considered lost if the number of higher sequence packets
2110 * SACKed is greater than or equal the DUPACK thoreshold
2111 * (reordering). This is implemented in tcp_mark_head_lost and
2112 * tcp_update_scoreboard.
2114 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm
2115 * (2017-) that checks timing instead of counting DUPACKs.
2116 * Essentially a packet is considered lost if it's not S/ACKed
2117 * after RTT + reordering_window, where both metrics are
2118 * dynamically measured and adjusted. This is implemented in
2119 * tcp_rack_mark_lost.
2121 * FACK (Disabled by default. Subsumbed by RACK):
2122 * It is the simplest heuristics. As soon as we decided
2123 * that something is lost, we decide that _all_ not SACKed
2124 * packets until the most forward SACK are lost. I.e.
2125 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2126 * It is absolutely correct estimate, if network does not reorder
2127 * packets. And it loses any connection to reality when reordering
2128 * takes place. We use FACK by default until reordering
2129 * is suspected on the path to this destination.
2131 * If the receiver does not support SACK:
2133 * NewReno (RFC6582): in Recovery we assume that one segment
2134 * is lost (classic Reno). While we are in Recovery and
2135 * a partial ACK arrives, we assume that one more packet
2136 * is lost (NewReno). This heuristics are the same in NewReno
2139 * Really tricky (and requiring careful tuning) part of algorithm
2140 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2141 * The first determines the moment _when_ we should reduce CWND and,
2142 * hence, slow down forward transmission. In fact, it determines the moment
2143 * when we decide that hole is caused by loss, rather than by a reorder.
2145 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2146 * holes, caused by lost packets.
2148 * And the most logically complicated part of algorithm is undo
2149 * heuristics. We detect false retransmits due to both too early
2150 * fast retransmit (reordering) and underestimated RTO, analyzing
2151 * timestamps and D-SACKs. When we detect that some segments were
2152 * retransmitted by mistake and CWND reduction was wrong, we undo
2153 * window reduction and abort recovery phase. This logic is hidden
2154 * inside several functions named tcp_try_undo_<something>.
2157 /* This function decides, when we should leave Disordered state
2158 * and enter Recovery phase, reducing congestion window.
2160 * Main question: may we further continue forward transmission
2161 * with the same cwnd?
2163 static bool tcp_time_to_recover(struct sock
*sk
, int flag
)
2165 struct tcp_sock
*tp
= tcp_sk(sk
);
2167 /* Trick#1: The loss is proven. */
2171 /* Not-A-Trick#2 : Classic rule... */
2172 if (tcp_dupack_heuristics(tp
) > tp
->reordering
)
2178 /* Detect loss in event "A" above by marking head of queue up as lost.
2179 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2180 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2181 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2182 * the maximum SACKed segments to pass before reaching this limit.
2184 static void tcp_mark_head_lost(struct sock
*sk
, int packets
, int mark_head
)
2186 struct tcp_sock
*tp
= tcp_sk(sk
);
2187 struct sk_buff
*skb
;
2188 int cnt
, oldcnt
, lost
;
2190 /* Use SACK to deduce losses of new sequences sent during recovery */
2191 const u32 loss_high
= tcp_is_sack(tp
) ? tp
->snd_nxt
: tp
->high_seq
;
2193 WARN_ON(packets
> tp
->packets_out
);
2194 if (tp
->lost_skb_hint
) {
2195 skb
= tp
->lost_skb_hint
;
2196 cnt
= tp
->lost_cnt_hint
;
2197 /* Head already handled? */
2198 if (mark_head
&& skb
!= tcp_write_queue_head(sk
))
2201 skb
= tcp_write_queue_head(sk
);
2205 tcp_for_write_queue_from(skb
, sk
) {
2206 if (skb
== tcp_send_head(sk
))
2208 /* TODO: do this better */
2209 /* this is not the most efficient way to do this... */
2210 tp
->lost_skb_hint
= skb
;
2211 tp
->lost_cnt_hint
= cnt
;
2213 if (after(TCP_SKB_CB(skb
)->end_seq
, loss_high
))
2217 if (tcp_is_fack(tp
) || tcp_is_reno(tp
) ||
2218 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2219 cnt
+= tcp_skb_pcount(skb
);
2221 if (cnt
> packets
) {
2222 if ((tcp_is_sack(tp
) && !tcp_is_fack(tp
)) ||
2223 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
) ||
2224 (oldcnt
>= packets
))
2227 mss
= tcp_skb_mss(skb
);
2228 /* If needed, chop off the prefix to mark as lost. */
2229 lost
= (packets
- oldcnt
) * mss
;
2230 if (lost
< skb
->len
&&
2231 tcp_fragment(sk
, skb
, lost
, mss
, GFP_ATOMIC
) < 0)
2236 tcp_skb_mark_lost(tp
, skb
);
2241 tcp_verify_left_out(tp
);
2244 /* Account newly detected lost packet(s) */
2246 static void tcp_update_scoreboard(struct sock
*sk
, int fast_rexmit
)
2248 struct tcp_sock
*tp
= tcp_sk(sk
);
2250 if (tcp_is_reno(tp
)) {
2251 tcp_mark_head_lost(sk
, 1, 1);
2252 } else if (tcp_is_fack(tp
)) {
2253 int lost
= tp
->fackets_out
- tp
->reordering
;
2256 tcp_mark_head_lost(sk
, lost
, 0);
2258 int sacked_upto
= tp
->sacked_out
- tp
->reordering
;
2259 if (sacked_upto
>= 0)
2260 tcp_mark_head_lost(sk
, sacked_upto
, 0);
2261 else if (fast_rexmit
)
2262 tcp_mark_head_lost(sk
, 1, 1);
2266 static bool tcp_tsopt_ecr_before(const struct tcp_sock
*tp
, u32 when
)
2268 return tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2269 before(tp
->rx_opt
.rcv_tsecr
, when
);
2272 /* skb is spurious retransmitted if the returned timestamp echo
2273 * reply is prior to the skb transmission time
2275 static bool tcp_skb_spurious_retrans(const struct tcp_sock
*tp
,
2276 const struct sk_buff
*skb
)
2278 return (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
) &&
2279 tcp_tsopt_ecr_before(tp
, tcp_skb_timestamp(skb
));
2282 /* Nothing was retransmitted or returned timestamp is less
2283 * than timestamp of the first retransmission.
2285 static inline bool tcp_packet_delayed(const struct tcp_sock
*tp
)
2287 return !tp
->retrans_stamp
||
2288 tcp_tsopt_ecr_before(tp
, tp
->retrans_stamp
);
2291 /* Undo procedures. */
2293 /* We can clear retrans_stamp when there are no retransmissions in the
2294 * window. It would seem that it is trivially available for us in
2295 * tp->retrans_out, however, that kind of assumptions doesn't consider
2296 * what will happen if errors occur when sending retransmission for the
2297 * second time. ...It could the that such segment has only
2298 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2299 * the head skb is enough except for some reneging corner cases that
2300 * are not worth the effort.
2302 * Main reason for all this complexity is the fact that connection dying
2303 * time now depends on the validity of the retrans_stamp, in particular,
2304 * that successive retransmissions of a segment must not advance
2305 * retrans_stamp under any conditions.
2307 static bool tcp_any_retrans_done(const struct sock
*sk
)
2309 const struct tcp_sock
*tp
= tcp_sk(sk
);
2310 struct sk_buff
*skb
;
2312 if (tp
->retrans_out
)
2315 skb
= tcp_write_queue_head(sk
);
2316 if (unlikely(skb
&& TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
))
2322 #if FASTRETRANS_DEBUG > 1
2323 static void DBGUNDO(struct sock
*sk
, const char *msg
)
2325 struct tcp_sock
*tp
= tcp_sk(sk
);
2326 struct inet_sock
*inet
= inet_sk(sk
);
2328 if (sk
->sk_family
== AF_INET
) {
2329 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2331 &inet
->inet_daddr
, ntohs(inet
->inet_dport
),
2332 tp
->snd_cwnd
, tcp_left_out(tp
),
2333 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2336 #if IS_ENABLED(CONFIG_IPV6)
2337 else if (sk
->sk_family
== AF_INET6
) {
2338 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2340 &sk
->sk_v6_daddr
, ntohs(inet
->inet_dport
),
2341 tp
->snd_cwnd
, tcp_left_out(tp
),
2342 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2348 #define DBGUNDO(x...) do { } while (0)
2351 static void tcp_undo_cwnd_reduction(struct sock
*sk
, bool unmark_loss
)
2353 struct tcp_sock
*tp
= tcp_sk(sk
);
2356 struct sk_buff
*skb
;
2358 tcp_for_write_queue(skb
, sk
) {
2359 if (skb
== tcp_send_head(sk
))
2361 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2364 tcp_clear_all_retrans_hints(tp
);
2367 if (tp
->prior_ssthresh
) {
2368 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2370 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
2372 if (tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
2373 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
2374 tcp_ecn_withdraw_cwr(tp
);
2377 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2378 tp
->undo_marker
= 0;
2381 static inline bool tcp_may_undo(const struct tcp_sock
*tp
)
2383 return tp
->undo_marker
&& (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
2386 /* People celebrate: "We love our President!" */
2387 static bool tcp_try_undo_recovery(struct sock
*sk
)
2389 struct tcp_sock
*tp
= tcp_sk(sk
);
2391 if (tcp_may_undo(tp
)) {
2394 /* Happy end! We did not retransmit anything
2395 * or our original transmission succeeded.
2397 DBGUNDO(sk
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
2398 tcp_undo_cwnd_reduction(sk
, false);
2399 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
2400 mib_idx
= LINUX_MIB_TCPLOSSUNDO
;
2402 mib_idx
= LINUX_MIB_TCPFULLUNDO
;
2404 NET_INC_STATS(sock_net(sk
), mib_idx
);
2406 if (tp
->snd_una
== tp
->high_seq
&& tcp_is_reno(tp
)) {
2407 /* Hold old state until something *above* high_seq
2408 * is ACKed. For Reno it is MUST to prevent false
2409 * fast retransmits (RFC2582). SACK TCP is safe. */
2410 if (!tcp_any_retrans_done(sk
))
2411 tp
->retrans_stamp
= 0;
2414 tcp_set_ca_state(sk
, TCP_CA_Open
);
2418 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2419 static bool tcp_try_undo_dsack(struct sock
*sk
)
2421 struct tcp_sock
*tp
= tcp_sk(sk
);
2423 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
2424 DBGUNDO(sk
, "D-SACK");
2425 tcp_undo_cwnd_reduction(sk
, false);
2426 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPDSACKUNDO
);
2432 /* Undo during loss recovery after partial ACK or using F-RTO. */
2433 static bool tcp_try_undo_loss(struct sock
*sk
, bool frto_undo
)
2435 struct tcp_sock
*tp
= tcp_sk(sk
);
2437 if (frto_undo
|| tcp_may_undo(tp
)) {
2438 tcp_undo_cwnd_reduction(sk
, true);
2440 DBGUNDO(sk
, "partial loss");
2441 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPLOSSUNDO
);
2443 NET_INC_STATS(sock_net(sk
),
2444 LINUX_MIB_TCPSPURIOUSRTOS
);
2445 inet_csk(sk
)->icsk_retransmits
= 0;
2446 if (frto_undo
|| tcp_is_sack(tp
))
2447 tcp_set_ca_state(sk
, TCP_CA_Open
);
2453 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2454 * It computes the number of packets to send (sndcnt) based on packets newly
2456 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2457 * cwnd reductions across a full RTT.
2458 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2459 * But when the retransmits are acked without further losses, PRR
2460 * slow starts cwnd up to ssthresh to speed up the recovery.
2462 static void tcp_init_cwnd_reduction(struct sock
*sk
)
2464 struct tcp_sock
*tp
= tcp_sk(sk
);
2466 tp
->high_seq
= tp
->snd_nxt
;
2467 tp
->tlp_high_seq
= 0;
2468 tp
->snd_cwnd_cnt
= 0;
2469 tp
->prior_cwnd
= tp
->snd_cwnd
;
2470 tp
->prr_delivered
= 0;
2472 tp
->snd_ssthresh
= inet_csk(sk
)->icsk_ca_ops
->ssthresh(sk
);
2473 tcp_ecn_queue_cwr(tp
);
2476 void tcp_cwnd_reduction(struct sock
*sk
, int newly_acked_sacked
, int flag
)
2478 struct tcp_sock
*tp
= tcp_sk(sk
);
2480 int delta
= tp
->snd_ssthresh
- tcp_packets_in_flight(tp
);
2482 if (newly_acked_sacked
<= 0 || WARN_ON_ONCE(!tp
->prior_cwnd
))
2485 tp
->prr_delivered
+= newly_acked_sacked
;
2487 u64 dividend
= (u64
)tp
->snd_ssthresh
* tp
->prr_delivered
+
2489 sndcnt
= div_u64(dividend
, tp
->prior_cwnd
) - tp
->prr_out
;
2490 } else if ((flag
& FLAG_RETRANS_DATA_ACKED
) &&
2491 !(flag
& FLAG_LOST_RETRANS
)) {
2492 sndcnt
= min_t(int, delta
,
2493 max_t(int, tp
->prr_delivered
- tp
->prr_out
,
2494 newly_acked_sacked
) + 1);
2496 sndcnt
= min(delta
, newly_acked_sacked
);
2498 /* Force a fast retransmit upon entering fast recovery */
2499 sndcnt
= max(sndcnt
, (tp
->prr_out
? 0 : 1));
2500 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + sndcnt
;
2503 static inline void tcp_end_cwnd_reduction(struct sock
*sk
)
2505 struct tcp_sock
*tp
= tcp_sk(sk
);
2507 if (inet_csk(sk
)->icsk_ca_ops
->cong_control
)
2510 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2511 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_CWR
||
2512 (tp
->undo_marker
&& tp
->snd_ssthresh
< TCP_INFINITE_SSTHRESH
)) {
2513 tp
->snd_cwnd
= tp
->snd_ssthresh
;
2514 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2516 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
2519 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2520 void tcp_enter_cwr(struct sock
*sk
)
2522 struct tcp_sock
*tp
= tcp_sk(sk
);
2524 tp
->prior_ssthresh
= 0;
2525 if (inet_csk(sk
)->icsk_ca_state
< TCP_CA_CWR
) {
2526 tp
->undo_marker
= 0;
2527 tcp_init_cwnd_reduction(sk
);
2528 tcp_set_ca_state(sk
, TCP_CA_CWR
);
2531 EXPORT_SYMBOL(tcp_enter_cwr
);
2533 static void tcp_try_keep_open(struct sock
*sk
)
2535 struct tcp_sock
*tp
= tcp_sk(sk
);
2536 int state
= TCP_CA_Open
;
2538 if (tcp_left_out(tp
) || tcp_any_retrans_done(sk
))
2539 state
= TCP_CA_Disorder
;
2541 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
2542 tcp_set_ca_state(sk
, state
);
2543 tp
->high_seq
= tp
->snd_nxt
;
2547 static void tcp_try_to_open(struct sock
*sk
, int flag
)
2549 struct tcp_sock
*tp
= tcp_sk(sk
);
2551 tcp_verify_left_out(tp
);
2553 if (!tcp_any_retrans_done(sk
))
2554 tp
->retrans_stamp
= 0;
2556 if (flag
& FLAG_ECE
)
2559 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
2560 tcp_try_keep_open(sk
);
2564 static void tcp_mtup_probe_failed(struct sock
*sk
)
2566 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2568 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
2569 icsk
->icsk_mtup
.probe_size
= 0;
2570 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPMTUPFAIL
);
2573 static void tcp_mtup_probe_success(struct sock
*sk
)
2575 struct tcp_sock
*tp
= tcp_sk(sk
);
2576 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2578 /* FIXME: breaks with very large cwnd */
2579 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2580 tp
->snd_cwnd
= tp
->snd_cwnd
*
2581 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
2582 icsk
->icsk_mtup
.probe_size
;
2583 tp
->snd_cwnd_cnt
= 0;
2584 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2585 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2587 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
2588 icsk
->icsk_mtup
.probe_size
= 0;
2589 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
2590 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPMTUPSUCCESS
);
2593 /* Do a simple retransmit without using the backoff mechanisms in
2594 * tcp_timer. This is used for path mtu discovery.
2595 * The socket is already locked here.
2597 void tcp_simple_retransmit(struct sock
*sk
)
2599 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2600 struct tcp_sock
*tp
= tcp_sk(sk
);
2601 struct sk_buff
*skb
;
2602 unsigned int mss
= tcp_current_mss(sk
);
2603 u32 prior_lost
= tp
->lost_out
;
2605 tcp_for_write_queue(skb
, sk
) {
2606 if (skb
== tcp_send_head(sk
))
2608 if (tcp_skb_seglen(skb
) > mss
&&
2609 !(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
2610 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2611 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
2612 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2614 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
2618 tcp_clear_retrans_hints_partial(tp
);
2620 if (prior_lost
== tp
->lost_out
)
2623 if (tcp_is_reno(tp
))
2624 tcp_limit_reno_sacked(tp
);
2626 tcp_verify_left_out(tp
);
2628 /* Don't muck with the congestion window here.
2629 * Reason is that we do not increase amount of _data_
2630 * in network, but units changed and effective
2631 * cwnd/ssthresh really reduced now.
2633 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
2634 tp
->high_seq
= tp
->snd_nxt
;
2635 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2636 tp
->prior_ssthresh
= 0;
2637 tp
->undo_marker
= 0;
2638 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2640 tcp_xmit_retransmit_queue(sk
);
2642 EXPORT_SYMBOL(tcp_simple_retransmit
);
2644 void tcp_enter_recovery(struct sock
*sk
, bool ece_ack
)
2646 struct tcp_sock
*tp
= tcp_sk(sk
);
2649 if (tcp_is_reno(tp
))
2650 mib_idx
= LINUX_MIB_TCPRENORECOVERY
;
2652 mib_idx
= LINUX_MIB_TCPSACKRECOVERY
;
2654 NET_INC_STATS(sock_net(sk
), mib_idx
);
2656 tp
->prior_ssthresh
= 0;
2659 if (!tcp_in_cwnd_reduction(sk
)) {
2661 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2662 tcp_init_cwnd_reduction(sk
);
2664 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
2667 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2668 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2670 static void tcp_process_loss(struct sock
*sk
, int flag
, bool is_dupack
,
2673 struct tcp_sock
*tp
= tcp_sk(sk
);
2674 bool recovered
= !before(tp
->snd_una
, tp
->high_seq
);
2676 if ((flag
& FLAG_SND_UNA_ADVANCED
) &&
2677 tcp_try_undo_loss(sk
, false))
2680 /* The ACK (s)acks some never-retransmitted data meaning not all
2681 * the data packets before the timeout were lost. Therefore we
2682 * undo the congestion window and state. This is essentially
2683 * the operation in F-RTO (RFC5682 section 3.1 step 3.b). Since
2684 * a retransmitted skb is permantly marked, we can apply such an
2685 * operation even if F-RTO was not used.
2687 if ((flag
& FLAG_ORIG_SACK_ACKED
) &&
2688 tcp_try_undo_loss(sk
, tp
->undo_marker
))
2691 if (tp
->frto
) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2692 if (after(tp
->snd_nxt
, tp
->high_seq
)) {
2693 if (flag
& FLAG_DATA_SACKED
|| is_dupack
)
2694 tp
->frto
= 0; /* Step 3.a. loss was real */
2695 } else if (flag
& FLAG_SND_UNA_ADVANCED
&& !recovered
) {
2696 tp
->high_seq
= tp
->snd_nxt
;
2697 /* Step 2.b. Try send new data (but deferred until cwnd
2698 * is updated in tcp_ack()). Otherwise fall back to
2699 * the conventional recovery.
2701 if (tcp_send_head(sk
) &&
2702 after(tcp_wnd_end(tp
), tp
->snd_nxt
)) {
2703 *rexmit
= REXMIT_NEW
;
2711 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2712 tcp_try_undo_recovery(sk
);
2715 if (tcp_is_reno(tp
)) {
2716 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2717 * delivered. Lower inflight to clock out (re)tranmissions.
2719 if (after(tp
->snd_nxt
, tp
->high_seq
) && is_dupack
)
2720 tcp_add_reno_sack(sk
);
2721 else if (flag
& FLAG_SND_UNA_ADVANCED
)
2722 tcp_reset_reno_sack(tp
);
2724 *rexmit
= REXMIT_LOST
;
2727 /* Undo during fast recovery after partial ACK. */
2728 static bool tcp_try_undo_partial(struct sock
*sk
, const int acked
)
2730 struct tcp_sock
*tp
= tcp_sk(sk
);
2732 if (tp
->undo_marker
&& tcp_packet_delayed(tp
)) {
2733 /* Plain luck! Hole if filled with delayed
2734 * packet, rather than with a retransmit.
2736 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
2738 /* We are getting evidence that the reordering degree is higher
2739 * than we realized. If there are no retransmits out then we
2740 * can undo. Otherwise we clock out new packets but do not
2741 * mark more packets lost or retransmit more.
2743 if (tp
->retrans_out
)
2746 if (!tcp_any_retrans_done(sk
))
2747 tp
->retrans_stamp
= 0;
2749 DBGUNDO(sk
, "partial recovery");
2750 tcp_undo_cwnd_reduction(sk
, true);
2751 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPPARTIALUNDO
);
2752 tcp_try_keep_open(sk
);
2758 static void tcp_rack_identify_loss(struct sock
*sk
, int *ack_flag
,
2759 const struct skb_mstamp
*ack_time
)
2761 struct tcp_sock
*tp
= tcp_sk(sk
);
2763 /* Use RACK to detect loss */
2764 if (sysctl_tcp_recovery
& TCP_RACK_LOSS_DETECTION
) {
2765 u32 prior_retrans
= tp
->retrans_out
;
2767 tcp_rack_mark_lost(sk
, ack_time
);
2768 if (prior_retrans
> tp
->retrans_out
)
2769 *ack_flag
|= FLAG_LOST_RETRANS
;
2773 /* Process an event, which can update packets-in-flight not trivially.
2774 * Main goal of this function is to calculate new estimate for left_out,
2775 * taking into account both packets sitting in receiver's buffer and
2776 * packets lost by network.
2778 * Besides that it updates the congestion state when packet loss or ECN
2779 * is detected. But it does not reduce the cwnd, it is done by the
2780 * congestion control later.
2782 * It does _not_ decide what to send, it is made in function
2783 * tcp_xmit_retransmit_queue().
2785 static void tcp_fastretrans_alert(struct sock
*sk
, const int acked
,
2786 bool is_dupack
, int *ack_flag
, int *rexmit
,
2787 const struct skb_mstamp
*ack_time
)
2789 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2790 struct tcp_sock
*tp
= tcp_sk(sk
);
2791 int fast_rexmit
= 0, flag
= *ack_flag
;
2792 bool do_lost
= is_dupack
|| ((flag
& FLAG_DATA_SACKED
) &&
2793 (tcp_fackets_out(tp
) > tp
->reordering
));
2795 if (WARN_ON(!tp
->packets_out
&& tp
->sacked_out
))
2797 if (WARN_ON(!tp
->sacked_out
&& tp
->fackets_out
))
2798 tp
->fackets_out
= 0;
2800 /* Now state machine starts.
2801 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2802 if (flag
& FLAG_ECE
)
2803 tp
->prior_ssthresh
= 0;
2805 /* B. In all the states check for reneging SACKs. */
2806 if (tcp_check_sack_reneging(sk
, flag
))
2809 /* C. Check consistency of the current state. */
2810 tcp_verify_left_out(tp
);
2812 /* D. Check state exit conditions. State can be terminated
2813 * when high_seq is ACKed. */
2814 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
2815 WARN_ON(tp
->retrans_out
!= 0);
2816 tp
->retrans_stamp
= 0;
2817 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
2818 switch (icsk
->icsk_ca_state
) {
2820 /* CWR is to be held something *above* high_seq
2821 * is ACKed for CWR bit to reach receiver. */
2822 if (tp
->snd_una
!= tp
->high_seq
) {
2823 tcp_end_cwnd_reduction(sk
);
2824 tcp_set_ca_state(sk
, TCP_CA_Open
);
2828 case TCP_CA_Recovery
:
2829 if (tcp_is_reno(tp
))
2830 tcp_reset_reno_sack(tp
);
2831 if (tcp_try_undo_recovery(sk
))
2833 tcp_end_cwnd_reduction(sk
);
2838 /* E. Process state. */
2839 switch (icsk
->icsk_ca_state
) {
2840 case TCP_CA_Recovery
:
2841 if (!(flag
& FLAG_SND_UNA_ADVANCED
)) {
2842 if (tcp_is_reno(tp
) && is_dupack
)
2843 tcp_add_reno_sack(sk
);
2845 if (tcp_try_undo_partial(sk
, acked
))
2847 /* Partial ACK arrived. Force fast retransmit. */
2848 do_lost
= tcp_is_reno(tp
) ||
2849 tcp_fackets_out(tp
) > tp
->reordering
;
2851 if (tcp_try_undo_dsack(sk
)) {
2852 tcp_try_keep_open(sk
);
2855 tcp_rack_identify_loss(sk
, ack_flag
, ack_time
);
2858 tcp_process_loss(sk
, flag
, is_dupack
, rexmit
);
2859 tcp_rack_identify_loss(sk
, ack_flag
, ack_time
);
2860 if (!(icsk
->icsk_ca_state
== TCP_CA_Open
||
2861 (*ack_flag
& FLAG_LOST_RETRANS
)))
2863 /* Change state if cwnd is undone or retransmits are lost */
2865 if (tcp_is_reno(tp
)) {
2866 if (flag
& FLAG_SND_UNA_ADVANCED
)
2867 tcp_reset_reno_sack(tp
);
2869 tcp_add_reno_sack(sk
);
2872 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
)
2873 tcp_try_undo_dsack(sk
);
2875 tcp_rack_identify_loss(sk
, ack_flag
, ack_time
);
2876 if (!tcp_time_to_recover(sk
, flag
)) {
2877 tcp_try_to_open(sk
, flag
);
2881 /* MTU probe failure: don't reduce cwnd */
2882 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
2883 icsk
->icsk_mtup
.probe_size
&&
2884 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
2885 tcp_mtup_probe_failed(sk
);
2886 /* Restores the reduction we did in tcp_mtup_probe() */
2888 tcp_simple_retransmit(sk
);
2892 /* Otherwise enter Recovery state */
2893 tcp_enter_recovery(sk
, (flag
& FLAG_ECE
));
2898 tcp_update_scoreboard(sk
, fast_rexmit
);
2899 *rexmit
= REXMIT_LOST
;
2902 static void tcp_update_rtt_min(struct sock
*sk
, u32 rtt_us
)
2904 struct tcp_sock
*tp
= tcp_sk(sk
);
2905 u32 wlen
= sysctl_tcp_min_rtt_wlen
* HZ
;
2907 minmax_running_min(&tp
->rtt_min
, wlen
, tcp_time_stamp
,
2908 rtt_us
? : jiffies_to_usecs(1));
2911 static inline bool tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
2912 long seq_rtt_us
, long sack_rtt_us
,
2915 const struct tcp_sock
*tp
= tcp_sk(sk
);
2917 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2918 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2919 * Karn's algorithm forbids taking RTT if some retransmitted data
2920 * is acked (RFC6298).
2923 seq_rtt_us
= sack_rtt_us
;
2925 /* RTTM Rule: A TSecr value received in a segment is used to
2926 * update the averaged RTT measurement only if the segment
2927 * acknowledges some new data, i.e., only if it advances the
2928 * left edge of the send window.
2929 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2931 if (seq_rtt_us
< 0 && tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2933 seq_rtt_us
= ca_rtt_us
= jiffies_to_usecs(tcp_time_stamp
-
2934 tp
->rx_opt
.rcv_tsecr
);
2938 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
2939 * always taken together with ACK, SACK, or TS-opts. Any negative
2940 * values will be skipped with the seq_rtt_us < 0 check above.
2942 tcp_update_rtt_min(sk
, ca_rtt_us
);
2943 tcp_rtt_estimator(sk
, seq_rtt_us
);
2946 /* RFC6298: only reset backoff on valid RTT measurement. */
2947 inet_csk(sk
)->icsk_backoff
= 0;
2951 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2952 void tcp_synack_rtt_meas(struct sock
*sk
, struct request_sock
*req
)
2956 if (req
&& !req
->num_retrans
&& tcp_rsk(req
)->snt_synack
.v64
) {
2957 struct skb_mstamp now
;
2959 skb_mstamp_get(&now
);
2960 rtt_us
= skb_mstamp_us_delta(&now
, &tcp_rsk(req
)->snt_synack
);
2963 tcp_ack_update_rtt(sk
, FLAG_SYN_ACKED
, rtt_us
, -1L, rtt_us
);
2967 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
, u32 acked
)
2969 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2971 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, acked
);
2972 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_time_stamp
;
2975 /* Restart timer after forward progress on connection.
2976 * RFC2988 recommends to restart timer to now+rto.
2978 void tcp_rearm_rto(struct sock
*sk
)
2980 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2981 struct tcp_sock
*tp
= tcp_sk(sk
);
2983 /* If the retrans timer is currently being used by Fast Open
2984 * for SYN-ACK retrans purpose, stay put.
2986 if (tp
->fastopen_rsk
)
2989 if (!tp
->packets_out
) {
2990 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
2992 u32 rto
= inet_csk(sk
)->icsk_rto
;
2993 /* Offset the time elapsed after installing regular RTO */
2994 if (icsk
->icsk_pending
== ICSK_TIME_REO_TIMEOUT
||
2995 icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
) {
2996 struct sk_buff
*skb
= tcp_write_queue_head(sk
);
2997 const u32 rto_time_stamp
=
2998 tcp_skb_timestamp(skb
) + rto
;
2999 s32 delta
= (s32
)(rto_time_stamp
- tcp_time_stamp
);
3000 /* delta may not be positive if the socket is locked
3001 * when the retrans timer fires and is rescheduled.
3006 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
, rto
,
3011 /* If we get here, the whole TSO packet has not been acked. */
3012 static u32
tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
)
3014 struct tcp_sock
*tp
= tcp_sk(sk
);
3017 BUG_ON(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
));
3019 packets_acked
= tcp_skb_pcount(skb
);
3020 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
3022 packets_acked
-= tcp_skb_pcount(skb
);
3024 if (packets_acked
) {
3025 BUG_ON(tcp_skb_pcount(skb
) == 0);
3026 BUG_ON(!before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
));
3029 return packets_acked
;
3032 static void tcp_ack_tstamp(struct sock
*sk
, struct sk_buff
*skb
,
3035 const struct skb_shared_info
*shinfo
;
3037 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3038 if (likely(!TCP_SKB_CB(skb
)->txstamp_ack
))
3041 shinfo
= skb_shinfo(skb
);
3042 if (!before(shinfo
->tskey
, prior_snd_una
) &&
3043 before(shinfo
->tskey
, tcp_sk(sk
)->snd_una
))
3044 __skb_tstamp_tx(skb
, NULL
, sk
, SCM_TSTAMP_ACK
);
3047 /* Remove acknowledged frames from the retransmission queue. If our packet
3048 * is before the ack sequence we can discard it as it's confirmed to have
3049 * arrived at the other end.
3051 static int tcp_clean_rtx_queue(struct sock
*sk
, int prior_fackets
,
3052 u32 prior_snd_una
, int *acked
,
3053 struct tcp_sacktag_state
*sack
)
3055 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3056 struct skb_mstamp first_ackt
, last_ackt
;
3057 struct skb_mstamp
*now
= &sack
->ack_time
;
3058 struct tcp_sock
*tp
= tcp_sk(sk
);
3059 u32 prior_sacked
= tp
->sacked_out
;
3060 u32 reord
= tp
->packets_out
;
3061 bool fully_acked
= true;
3062 long sack_rtt_us
= -1L;
3063 long seq_rtt_us
= -1L;
3064 long ca_rtt_us
= -1L;
3065 struct sk_buff
*skb
;
3067 u32 last_in_flight
= 0;
3073 while ((skb
= tcp_write_queue_head(sk
)) && skb
!= tcp_send_head(sk
)) {
3074 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
3075 u8 sacked
= scb
->sacked
;
3078 tcp_ack_tstamp(sk
, skb
, prior_snd_una
);
3080 /* Determine how many packets and what bytes were acked, tso and else */
3081 if (after(scb
->end_seq
, tp
->snd_una
)) {
3082 if (tcp_skb_pcount(skb
) == 1 ||
3083 !after(tp
->snd_una
, scb
->seq
))
3086 acked_pcount
= tcp_tso_acked(sk
, skb
);
3089 fully_acked
= false;
3091 /* Speedup tcp_unlink_write_queue() and next loop */
3092 prefetchw(skb
->next
);
3093 acked_pcount
= tcp_skb_pcount(skb
);
3096 if (unlikely(sacked
& TCPCB_RETRANS
)) {
3097 if (sacked
& TCPCB_SACKED_RETRANS
)
3098 tp
->retrans_out
-= acked_pcount
;
3099 flag
|= FLAG_RETRANS_DATA_ACKED
;
3100 } else if (!(sacked
& TCPCB_SACKED_ACKED
)) {
3101 last_ackt
= skb
->skb_mstamp
;
3102 WARN_ON_ONCE(last_ackt
.v64
== 0);
3103 if (!first_ackt
.v64
)
3104 first_ackt
= last_ackt
;
3106 last_in_flight
= TCP_SKB_CB(skb
)->tx
.in_flight
;
3107 reord
= min(pkts_acked
, reord
);
3108 if (!after(scb
->end_seq
, tp
->high_seq
))
3109 flag
|= FLAG_ORIG_SACK_ACKED
;
3112 if (sacked
& TCPCB_SACKED_ACKED
) {
3113 tp
->sacked_out
-= acked_pcount
;
3114 } else if (tcp_is_sack(tp
)) {
3115 tp
->delivered
+= acked_pcount
;
3116 if (!tcp_skb_spurious_retrans(tp
, skb
))
3117 tcp_rack_advance(tp
, sacked
, scb
->end_seq
,
3121 if (sacked
& TCPCB_LOST
)
3122 tp
->lost_out
-= acked_pcount
;
3124 tp
->packets_out
-= acked_pcount
;
3125 pkts_acked
+= acked_pcount
;
3126 tcp_rate_skb_delivered(sk
, skb
, sack
->rate
);
3128 /* Initial outgoing SYN's get put onto the write_queue
3129 * just like anything else we transmit. It is not
3130 * true data, and if we misinform our callers that
3131 * this ACK acks real data, we will erroneously exit
3132 * connection startup slow start one packet too
3133 * quickly. This is severely frowned upon behavior.
3135 if (likely(!(scb
->tcp_flags
& TCPHDR_SYN
))) {
3136 flag
|= FLAG_DATA_ACKED
;
3138 flag
|= FLAG_SYN_ACKED
;
3139 tp
->retrans_stamp
= 0;
3145 tcp_unlink_write_queue(skb
, sk
);
3146 sk_wmem_free_skb(sk
, skb
);
3147 if (unlikely(skb
== tp
->retransmit_skb_hint
))
3148 tp
->retransmit_skb_hint
= NULL
;
3149 if (unlikely(skb
== tp
->lost_skb_hint
))
3150 tp
->lost_skb_hint
= NULL
;
3154 tcp_chrono_stop(sk
, TCP_CHRONO_BUSY
);
3156 if (likely(between(tp
->snd_up
, prior_snd_una
, tp
->snd_una
)))
3157 tp
->snd_up
= tp
->snd_una
;
3159 if (skb
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
3160 flag
|= FLAG_SACK_RENEGING
;
3162 if (likely(first_ackt
.v64
) && !(flag
& FLAG_RETRANS_DATA_ACKED
)) {
3163 seq_rtt_us
= skb_mstamp_us_delta(now
, &first_ackt
);
3164 ca_rtt_us
= skb_mstamp_us_delta(now
, &last_ackt
);
3166 if (sack
->first_sackt
.v64
) {
3167 sack_rtt_us
= skb_mstamp_us_delta(now
, &sack
->first_sackt
);
3168 ca_rtt_us
= skb_mstamp_us_delta(now
, &sack
->last_sackt
);
3170 sack
->rate
->rtt_us
= ca_rtt_us
; /* RTT of last (S)ACKed packet, or -1 */
3171 rtt_update
= tcp_ack_update_rtt(sk
, flag
, seq_rtt_us
, sack_rtt_us
,
3174 if (flag
& FLAG_ACKED
) {
3176 if (unlikely(icsk
->icsk_mtup
.probe_size
&&
3177 !after(tp
->mtu_probe
.probe_seq_end
, tp
->snd_una
))) {
3178 tcp_mtup_probe_success(sk
);
3181 if (tcp_is_reno(tp
)) {
3182 tcp_remove_reno_sacks(sk
, pkts_acked
);
3186 /* Non-retransmitted hole got filled? That's reordering */
3187 if (reord
< prior_fackets
)
3188 tcp_update_reordering(sk
, tp
->fackets_out
- reord
, 0);
3190 delta
= tcp_is_fack(tp
) ? pkts_acked
:
3191 prior_sacked
- tp
->sacked_out
;
3192 tp
->lost_cnt_hint
-= min(tp
->lost_cnt_hint
, delta
);
3195 tp
->fackets_out
-= min(pkts_acked
, tp
->fackets_out
);
3197 } else if (skb
&& rtt_update
&& sack_rtt_us
>= 0 &&
3198 sack_rtt_us
> skb_mstamp_us_delta(now
, &skb
->skb_mstamp
)) {
3199 /* Do not re-arm RTO if the sack RTT is measured from data sent
3200 * after when the head was last (re)transmitted. Otherwise the
3201 * timeout may continue to extend in loss recovery.
3206 if (icsk
->icsk_ca_ops
->pkts_acked
) {
3207 struct ack_sample sample
= { .pkts_acked
= pkts_acked
,
3208 .rtt_us
= ca_rtt_us
,
3209 .in_flight
= last_in_flight
};
3211 icsk
->icsk_ca_ops
->pkts_acked(sk
, &sample
);
3214 #if FASTRETRANS_DEBUG > 0
3215 WARN_ON((int)tp
->sacked_out
< 0);
3216 WARN_ON((int)tp
->lost_out
< 0);
3217 WARN_ON((int)tp
->retrans_out
< 0);
3218 if (!tp
->packets_out
&& tcp_is_sack(tp
)) {
3219 icsk
= inet_csk(sk
);
3221 pr_debug("Leak l=%u %d\n",
3222 tp
->lost_out
, icsk
->icsk_ca_state
);
3225 if (tp
->sacked_out
) {
3226 pr_debug("Leak s=%u %d\n",
3227 tp
->sacked_out
, icsk
->icsk_ca_state
);
3230 if (tp
->retrans_out
) {
3231 pr_debug("Leak r=%u %d\n",
3232 tp
->retrans_out
, icsk
->icsk_ca_state
);
3233 tp
->retrans_out
= 0;
3237 *acked
= pkts_acked
;
3241 static void tcp_ack_probe(struct sock
*sk
)
3243 const struct tcp_sock
*tp
= tcp_sk(sk
);
3244 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3246 /* Was it a usable window open? */
3248 if (!after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
, tcp_wnd_end(tp
))) {
3249 icsk
->icsk_backoff
= 0;
3250 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
3251 /* Socket must be waked up by subsequent tcp_data_snd_check().
3252 * This function is not for random using!
3255 unsigned long when
= tcp_probe0_when(sk
, TCP_RTO_MAX
);
3257 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3262 static inline bool tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
3264 return !(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
3265 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
;
3268 /* Decide wheather to run the increase function of congestion control. */
3269 static inline bool tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
3271 /* If reordering is high then always grow cwnd whenever data is
3272 * delivered regardless of its ordering. Otherwise stay conservative
3273 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3274 * new SACK or ECE mark may first advance cwnd here and later reduce
3275 * cwnd in tcp_fastretrans_alert() based on more states.
3277 if (tcp_sk(sk
)->reordering
> sock_net(sk
)->ipv4
.sysctl_tcp_reordering
)
3278 return flag
& FLAG_FORWARD_PROGRESS
;
3280 return flag
& FLAG_DATA_ACKED
;
3283 /* The "ultimate" congestion control function that aims to replace the rigid
3284 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3285 * It's called toward the end of processing an ACK with precise rate
3286 * information. All transmission or retransmission are delayed afterwards.
3288 static void tcp_cong_control(struct sock
*sk
, u32 ack
, u32 acked_sacked
,
3289 int flag
, const struct rate_sample
*rs
)
3291 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3293 if (icsk
->icsk_ca_ops
->cong_control
) {
3294 icsk
->icsk_ca_ops
->cong_control(sk
, rs
);
3298 if (tcp_in_cwnd_reduction(sk
)) {
3299 /* Reduce cwnd if state mandates */
3300 tcp_cwnd_reduction(sk
, acked_sacked
, flag
);
3301 } else if (tcp_may_raise_cwnd(sk
, flag
)) {
3302 /* Advance cwnd if state allows */
3303 tcp_cong_avoid(sk
, ack
, acked_sacked
);
3305 tcp_update_pacing_rate(sk
);
3308 /* Check that window update is acceptable.
3309 * The function assumes that snd_una<=ack<=snd_next.
3311 static inline bool tcp_may_update_window(const struct tcp_sock
*tp
,
3312 const u32 ack
, const u32 ack_seq
,
3315 return after(ack
, tp
->snd_una
) ||
3316 after(ack_seq
, tp
->snd_wl1
) ||
3317 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
);
3320 /* If we update tp->snd_una, also update tp->bytes_acked */
3321 static void tcp_snd_una_update(struct tcp_sock
*tp
, u32 ack
)
3323 u32 delta
= ack
- tp
->snd_una
;
3325 sock_owned_by_me((struct sock
*)tp
);
3326 tp
->bytes_acked
+= delta
;
3330 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3331 static void tcp_rcv_nxt_update(struct tcp_sock
*tp
, u32 seq
)
3333 u32 delta
= seq
- tp
->rcv_nxt
;
3335 sock_owned_by_me((struct sock
*)tp
);
3336 tp
->bytes_received
+= delta
;
3340 /* Update our send window.
3342 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3343 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3345 static int tcp_ack_update_window(struct sock
*sk
, const struct sk_buff
*skb
, u32 ack
,
3348 struct tcp_sock
*tp
= tcp_sk(sk
);
3350 u32 nwin
= ntohs(tcp_hdr(skb
)->window
);
3352 if (likely(!tcp_hdr(skb
)->syn
))
3353 nwin
<<= tp
->rx_opt
.snd_wscale
;
3355 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
3356 flag
|= FLAG_WIN_UPDATE
;
3357 tcp_update_wl(tp
, ack_seq
);
3359 if (tp
->snd_wnd
!= nwin
) {
3362 /* Note, it is the only place, where
3363 * fast path is recovered for sending TCP.
3366 tcp_fast_path_check(sk
);
3368 if (tcp_send_head(sk
))
3369 tcp_slow_start_after_idle_check(sk
);
3371 if (nwin
> tp
->max_window
) {
3372 tp
->max_window
= nwin
;
3373 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
3378 tcp_snd_una_update(tp
, ack
);
3383 static bool __tcp_oow_rate_limited(struct net
*net
, int mib_idx
,
3384 u32
*last_oow_ack_time
)
3386 if (*last_oow_ack_time
) {
3387 s32 elapsed
= (s32
)(tcp_time_stamp
- *last_oow_ack_time
);
3389 if (0 <= elapsed
&& elapsed
< sysctl_tcp_invalid_ratelimit
) {
3390 NET_INC_STATS(net
, mib_idx
);
3391 return true; /* rate-limited: don't send yet! */
3395 *last_oow_ack_time
= tcp_time_stamp
;
3397 return false; /* not rate-limited: go ahead, send dupack now! */
3400 /* Return true if we're currently rate-limiting out-of-window ACKs and
3401 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3402 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3403 * attacks that send repeated SYNs or ACKs for the same connection. To
3404 * do this, we do not send a duplicate SYNACK or ACK if the remote
3405 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3407 bool tcp_oow_rate_limited(struct net
*net
, const struct sk_buff
*skb
,
3408 int mib_idx
, u32
*last_oow_ack_time
)
3410 /* Data packets without SYNs are not likely part of an ACK loop. */
3411 if ((TCP_SKB_CB(skb
)->seq
!= TCP_SKB_CB(skb
)->end_seq
) &&
3415 return __tcp_oow_rate_limited(net
, mib_idx
, last_oow_ack_time
);
3418 /* RFC 5961 7 [ACK Throttling] */
3419 static void tcp_send_challenge_ack(struct sock
*sk
, const struct sk_buff
*skb
)
3421 /* unprotected vars, we dont care of overwrites */
3422 static u32 challenge_timestamp
;
3423 static unsigned int challenge_count
;
3424 struct tcp_sock
*tp
= tcp_sk(sk
);
3427 /* First check our per-socket dupack rate limit. */
3428 if (__tcp_oow_rate_limited(sock_net(sk
),
3429 LINUX_MIB_TCPACKSKIPPEDCHALLENGE
,
3430 &tp
->last_oow_ack_time
))
3433 /* Then check host-wide RFC 5961 rate limit. */
3435 if (now
!= challenge_timestamp
) {
3436 u32 half
= (sysctl_tcp_challenge_ack_limit
+ 1) >> 1;
3438 challenge_timestamp
= now
;
3439 WRITE_ONCE(challenge_count
, half
+
3440 prandom_u32_max(sysctl_tcp_challenge_ack_limit
));
3442 count
= READ_ONCE(challenge_count
);
3444 WRITE_ONCE(challenge_count
, count
- 1);
3445 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPCHALLENGEACK
);
3450 static void tcp_store_ts_recent(struct tcp_sock
*tp
)
3452 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
3453 tp
->rx_opt
.ts_recent_stamp
= get_seconds();
3456 static void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
3458 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
3459 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3460 * extra check below makes sure this can only happen
3461 * for pure ACK frames. -DaveM
3463 * Not only, also it occurs for expired timestamps.
3466 if (tcp_paws_check(&tp
->rx_opt
, 0))
3467 tcp_store_ts_recent(tp
);
3471 /* This routine deals with acks during a TLP episode.
3472 * We mark the end of a TLP episode on receiving TLP dupack or when
3473 * ack is after tlp_high_seq.
3474 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3476 static void tcp_process_tlp_ack(struct sock
*sk
, u32 ack
, int flag
)
3478 struct tcp_sock
*tp
= tcp_sk(sk
);
3480 if (before(ack
, tp
->tlp_high_seq
))
3483 if (flag
& FLAG_DSACKING_ACK
) {
3484 /* This DSACK means original and TLP probe arrived; no loss */
3485 tp
->tlp_high_seq
= 0;
3486 } else if (after(ack
, tp
->tlp_high_seq
)) {
3487 /* ACK advances: there was a loss, so reduce cwnd. Reset
3488 * tlp_high_seq in tcp_init_cwnd_reduction()
3490 tcp_init_cwnd_reduction(sk
);
3491 tcp_set_ca_state(sk
, TCP_CA_CWR
);
3492 tcp_end_cwnd_reduction(sk
);
3493 tcp_try_keep_open(sk
);
3494 NET_INC_STATS(sock_net(sk
),
3495 LINUX_MIB_TCPLOSSPROBERECOVERY
);
3496 } else if (!(flag
& (FLAG_SND_UNA_ADVANCED
|
3497 FLAG_NOT_DUP
| FLAG_DATA_SACKED
))) {
3498 /* Pure dupack: original and TLP probe arrived; no loss */
3499 tp
->tlp_high_seq
= 0;
3503 static inline void tcp_in_ack_event(struct sock
*sk
, u32 flags
)
3505 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3507 if (icsk
->icsk_ca_ops
->in_ack_event
)
3508 icsk
->icsk_ca_ops
->in_ack_event(sk
, flags
);
3511 /* Congestion control has updated the cwnd already. So if we're in
3512 * loss recovery then now we do any new sends (for FRTO) or
3513 * retransmits (for CA_Loss or CA_recovery) that make sense.
3515 static void tcp_xmit_recovery(struct sock
*sk
, int rexmit
)
3517 struct tcp_sock
*tp
= tcp_sk(sk
);
3519 if (rexmit
== REXMIT_NONE
)
3522 if (unlikely(rexmit
== 2)) {
3523 __tcp_push_pending_frames(sk
, tcp_current_mss(sk
),
3525 if (after(tp
->snd_nxt
, tp
->high_seq
))
3529 tcp_xmit_retransmit_queue(sk
);
3532 /* This routine deals with incoming acks, but not outgoing ones. */
3533 static int tcp_ack(struct sock
*sk
, const struct sk_buff
*skb
, int flag
)
3535 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3536 struct tcp_sock
*tp
= tcp_sk(sk
);
3537 struct tcp_sacktag_state sack_state
;
3538 struct rate_sample rs
= { .prior_delivered
= 0 };
3539 u32 prior_snd_una
= tp
->snd_una
;
3540 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
3541 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3542 bool is_dupack
= false;
3544 int prior_packets
= tp
->packets_out
;
3545 u32 delivered
= tp
->delivered
;
3546 u32 lost
= tp
->lost
;
3547 int acked
= 0; /* Number of packets newly acked */
3548 int rexmit
= REXMIT_NONE
; /* Flag to (re)transmit to recover losses */
3550 sack_state
.first_sackt
.v64
= 0;
3551 sack_state
.rate
= &rs
;
3553 /* We very likely will need to access write queue head. */
3554 prefetchw(sk
->sk_write_queue
.next
);
3556 /* If the ack is older than previous acks
3557 * then we can probably ignore it.
3559 if (before(ack
, prior_snd_una
)) {
3560 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3561 if (before(ack
, prior_snd_una
- tp
->max_window
)) {
3562 tcp_send_challenge_ack(sk
, skb
);
3568 /* If the ack includes data we haven't sent yet, discard
3569 * this segment (RFC793 Section 3.9).
3571 if (after(ack
, tp
->snd_nxt
))
3574 skb_mstamp_get(&sack_state
.ack_time
);
3576 if (icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
)
3579 if (after(ack
, prior_snd_una
)) {
3580 flag
|= FLAG_SND_UNA_ADVANCED
;
3581 icsk
->icsk_retransmits
= 0;
3584 prior_fackets
= tp
->fackets_out
;
3585 rs
.prior_in_flight
= tcp_packets_in_flight(tp
);
3587 /* ts_recent update must be made after we are sure that the packet
3590 if (flag
& FLAG_UPDATE_TS_RECENT
)
3591 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
3593 if (!(flag
& FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
3594 /* Window is constant, pure forward advance.
3595 * No more checks are required.
3596 * Note, we use the fact that SND.UNA>=SND.WL2.
3598 tcp_update_wl(tp
, ack_seq
);
3599 tcp_snd_una_update(tp
, ack
);
3600 flag
|= FLAG_WIN_UPDATE
;
3602 tcp_in_ack_event(sk
, CA_ACK_WIN_UPDATE
);
3604 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPHPACKS
);
3606 u32 ack_ev_flags
= CA_ACK_SLOWPATH
;
3608 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
3611 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPPUREACKS
);
3613 flag
|= tcp_ack_update_window(sk
, skb
, ack
, ack_seq
);
3615 if (TCP_SKB_CB(skb
)->sacked
)
3616 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
,
3619 if (tcp_ecn_rcv_ecn_echo(tp
, tcp_hdr(skb
))) {
3621 ack_ev_flags
|= CA_ACK_ECE
;
3624 if (flag
& FLAG_WIN_UPDATE
)
3625 ack_ev_flags
|= CA_ACK_WIN_UPDATE
;
3627 tcp_in_ack_event(sk
, ack_ev_flags
);
3630 /* We passed data and got it acked, remove any soft error
3631 * log. Something worked...
3633 sk
->sk_err_soft
= 0;
3634 icsk
->icsk_probes_out
= 0;
3635 tp
->rcv_tstamp
= tcp_time_stamp
;
3639 /* See if we can take anything off of the retransmit queue. */
3640 flag
|= tcp_clean_rtx_queue(sk
, prior_fackets
, prior_snd_una
, &acked
,
3643 if (tcp_ack_is_dubious(sk
, flag
)) {
3644 is_dupack
= !(flag
& (FLAG_SND_UNA_ADVANCED
| FLAG_NOT_DUP
));
3645 tcp_fastretrans_alert(sk
, acked
, is_dupack
, &flag
, &rexmit
,
3646 &sack_state
.ack_time
);
3648 if (tp
->tlp_high_seq
)
3649 tcp_process_tlp_ack(sk
, ack
, flag
);
3651 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
& FLAG_NOT_DUP
))
3654 if (icsk
->icsk_pending
== ICSK_TIME_RETRANS
)
3655 tcp_schedule_loss_probe(sk
);
3656 delivered
= tp
->delivered
- delivered
; /* freshly ACKed or SACKed */
3657 lost
= tp
->lost
- lost
; /* freshly marked lost */
3658 tcp_rate_gen(sk
, delivered
, lost
, &sack_state
.ack_time
,
3660 tcp_cong_control(sk
, ack
, delivered
, flag
, sack_state
.rate
);
3661 tcp_xmit_recovery(sk
, rexmit
);
3665 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3666 if (flag
& FLAG_DSACKING_ACK
)
3667 tcp_fastretrans_alert(sk
, acked
, is_dupack
, &flag
, &rexmit
,
3668 &sack_state
.ack_time
);
3669 /* If this ack opens up a zero window, clear backoff. It was
3670 * being used to time the probes, and is probably far higher than
3671 * it needs to be for normal retransmission.
3673 if (tcp_send_head(sk
))
3676 if (tp
->tlp_high_seq
)
3677 tcp_process_tlp_ack(sk
, ack
, flag
);
3681 SOCK_DEBUG(sk
, "Ack %u after %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3685 /* If data was SACKed, tag it and see if we should send more data.
3686 * If data was DSACKed, see if we can undo a cwnd reduction.
3688 if (TCP_SKB_CB(skb
)->sacked
) {
3689 skb_mstamp_get(&sack_state
.ack_time
);
3690 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
,
3692 tcp_fastretrans_alert(sk
, acked
, is_dupack
, &flag
, &rexmit
,
3693 &sack_state
.ack_time
);
3694 tcp_xmit_recovery(sk
, rexmit
);
3697 SOCK_DEBUG(sk
, "Ack %u before %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3701 static void tcp_parse_fastopen_option(int len
, const unsigned char *cookie
,
3702 bool syn
, struct tcp_fastopen_cookie
*foc
,
3705 /* Valid only in SYN or SYN-ACK with an even length. */
3706 if (!foc
|| !syn
|| len
< 0 || (len
& 1))
3709 if (len
>= TCP_FASTOPEN_COOKIE_MIN
&&
3710 len
<= TCP_FASTOPEN_COOKIE_MAX
)
3711 memcpy(foc
->val
, cookie
, len
);
3718 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3719 * But, this can also be called on packets in the established flow when
3720 * the fast version below fails.
3722 void tcp_parse_options(const struct sk_buff
*skb
,
3723 struct tcp_options_received
*opt_rx
, int estab
,
3724 struct tcp_fastopen_cookie
*foc
)
3726 const unsigned char *ptr
;
3727 const struct tcphdr
*th
= tcp_hdr(skb
);
3728 int length
= (th
->doff
* 4) - sizeof(struct tcphdr
);
3730 ptr
= (const unsigned char *)(th
+ 1);
3731 opt_rx
->saw_tstamp
= 0;
3733 while (length
> 0) {
3734 int opcode
= *ptr
++;
3740 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
3745 if (opsize
< 2) /* "silly options" */
3747 if (opsize
> length
)
3748 return; /* don't parse partial options */
3751 if (opsize
== TCPOLEN_MSS
&& th
->syn
&& !estab
) {
3752 u16 in_mss
= get_unaligned_be16(ptr
);
3754 if (opt_rx
->user_mss
&&
3755 opt_rx
->user_mss
< in_mss
)
3756 in_mss
= opt_rx
->user_mss
;
3757 opt_rx
->mss_clamp
= in_mss
;
3762 if (opsize
== TCPOLEN_WINDOW
&& th
->syn
&&
3763 !estab
&& sysctl_tcp_window_scaling
) {
3764 __u8 snd_wscale
= *(__u8
*)ptr
;
3765 opt_rx
->wscale_ok
= 1;
3766 if (snd_wscale
> 14) {
3767 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3772 opt_rx
->snd_wscale
= snd_wscale
;
3775 case TCPOPT_TIMESTAMP
:
3776 if ((opsize
== TCPOLEN_TIMESTAMP
) &&
3777 ((estab
&& opt_rx
->tstamp_ok
) ||
3778 (!estab
&& sysctl_tcp_timestamps
))) {
3779 opt_rx
->saw_tstamp
= 1;
3780 opt_rx
->rcv_tsval
= get_unaligned_be32(ptr
);
3781 opt_rx
->rcv_tsecr
= get_unaligned_be32(ptr
+ 4);
3784 case TCPOPT_SACK_PERM
:
3785 if (opsize
== TCPOLEN_SACK_PERM
&& th
->syn
&&
3786 !estab
&& sysctl_tcp_sack
) {
3787 opt_rx
->sack_ok
= TCP_SACK_SEEN
;
3788 tcp_sack_reset(opt_rx
);
3793 if ((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
3794 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
3796 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
3799 #ifdef CONFIG_TCP_MD5SIG
3802 * The MD5 Hash has already been
3803 * checked (see tcp_v{4,6}_do_rcv()).
3807 case TCPOPT_FASTOPEN
:
3808 tcp_parse_fastopen_option(
3809 opsize
- TCPOLEN_FASTOPEN_BASE
,
3810 ptr
, th
->syn
, foc
, false);
3814 /* Fast Open option shares code 254 using a
3815 * 16 bits magic number.
3817 if (opsize
>= TCPOLEN_EXP_FASTOPEN_BASE
&&
3818 get_unaligned_be16(ptr
) ==
3819 TCPOPT_FASTOPEN_MAGIC
)
3820 tcp_parse_fastopen_option(opsize
-
3821 TCPOLEN_EXP_FASTOPEN_BASE
,
3822 ptr
+ 2, th
->syn
, foc
, true);
3831 EXPORT_SYMBOL(tcp_parse_options
);
3833 static bool tcp_parse_aligned_timestamp(struct tcp_sock
*tp
, const struct tcphdr
*th
)
3835 const __be32
*ptr
= (const __be32
*)(th
+ 1);
3837 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3838 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
3839 tp
->rx_opt
.saw_tstamp
= 1;
3841 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3844 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
) - tp
->tsoffset
;
3846 tp
->rx_opt
.rcv_tsecr
= 0;
3852 /* Fast parse options. This hopes to only see timestamps.
3853 * If it is wrong it falls back on tcp_parse_options().
3855 static bool tcp_fast_parse_options(const struct sk_buff
*skb
,
3856 const struct tcphdr
*th
, struct tcp_sock
*tp
)
3858 /* In the spirit of fast parsing, compare doff directly to constant
3859 * values. Because equality is used, short doff can be ignored here.
3861 if (th
->doff
== (sizeof(*th
) / 4)) {
3862 tp
->rx_opt
.saw_tstamp
= 0;
3864 } else if (tp
->rx_opt
.tstamp_ok
&&
3865 th
->doff
== ((sizeof(*th
) + TCPOLEN_TSTAMP_ALIGNED
) / 4)) {
3866 if (tcp_parse_aligned_timestamp(tp
, th
))
3870 tcp_parse_options(skb
, &tp
->rx_opt
, 1, NULL
);
3871 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
3872 tp
->rx_opt
.rcv_tsecr
-= tp
->tsoffset
;
3877 #ifdef CONFIG_TCP_MD5SIG
3879 * Parse MD5 Signature option
3881 const u8
*tcp_parse_md5sig_option(const struct tcphdr
*th
)
3883 int length
= (th
->doff
<< 2) - sizeof(*th
);
3884 const u8
*ptr
= (const u8
*)(th
+ 1);
3886 /* If the TCP option is too short, we can short cut */
3887 if (length
< TCPOLEN_MD5SIG
)
3890 while (length
> 0) {
3891 int opcode
= *ptr
++;
3902 if (opsize
< 2 || opsize
> length
)
3904 if (opcode
== TCPOPT_MD5SIG
)
3905 return opsize
== TCPOLEN_MD5SIG
? ptr
: NULL
;
3912 EXPORT_SYMBOL(tcp_parse_md5sig_option
);
3915 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3917 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3918 * it can pass through stack. So, the following predicate verifies that
3919 * this segment is not used for anything but congestion avoidance or
3920 * fast retransmit. Moreover, we even are able to eliminate most of such
3921 * second order effects, if we apply some small "replay" window (~RTO)
3922 * to timestamp space.
3924 * All these measures still do not guarantee that we reject wrapped ACKs
3925 * on networks with high bandwidth, when sequence space is recycled fastly,
3926 * but it guarantees that such events will be very rare and do not affect
3927 * connection seriously. This doesn't look nice, but alas, PAWS is really
3930 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3931 * states that events when retransmit arrives after original data are rare.
3932 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3933 * the biggest problem on large power networks even with minor reordering.
3934 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3935 * up to bandwidth of 18Gigabit/sec. 8) ]
3938 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
3940 const struct tcp_sock
*tp
= tcp_sk(sk
);
3941 const struct tcphdr
*th
= tcp_hdr(skb
);
3942 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3943 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3945 return (/* 1. Pure ACK with correct sequence number. */
3946 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
3948 /* 2. ... and duplicate ACK. */
3949 ack
== tp
->snd_una
&&
3951 /* 3. ... and does not update window. */
3952 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
3954 /* 4. ... and sits in replay window. */
3955 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
3958 static inline bool tcp_paws_discard(const struct sock
*sk
,
3959 const struct sk_buff
*skb
)
3961 const struct tcp_sock
*tp
= tcp_sk(sk
);
3963 return !tcp_paws_check(&tp
->rx_opt
, TCP_PAWS_WINDOW
) &&
3964 !tcp_disordered_ack(sk
, skb
);
3967 /* Check segment sequence number for validity.
3969 * Segment controls are considered valid, if the segment
3970 * fits to the window after truncation to the window. Acceptability
3971 * of data (and SYN, FIN, of course) is checked separately.
3972 * See tcp_data_queue(), for example.
3974 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3975 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3976 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3977 * (borrowed from freebsd)
3980 static inline bool tcp_sequence(const struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3982 return !before(end_seq
, tp
->rcv_wup
) &&
3983 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
3986 /* When we get a reset we do this. */
3987 void tcp_reset(struct sock
*sk
)
3989 /* We want the right error as BSD sees it (and indeed as we do). */
3990 switch (sk
->sk_state
) {
3992 sk
->sk_err
= ECONNREFUSED
;
3994 case TCP_CLOSE_WAIT
:
4000 sk
->sk_err
= ECONNRESET
;
4002 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4005 if (!sock_flag(sk
, SOCK_DEAD
))
4006 sk
->sk_error_report(sk
);
4012 * Process the FIN bit. This now behaves as it is supposed to work
4013 * and the FIN takes effect when it is validly part of sequence
4014 * space. Not before when we get holes.
4016 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4017 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4020 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4021 * close and we go into CLOSING (and later onto TIME-WAIT)
4023 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4025 void tcp_fin(struct sock
*sk
)
4027 struct tcp_sock
*tp
= tcp_sk(sk
);
4029 inet_csk_schedule_ack(sk
);
4031 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
4032 sock_set_flag(sk
, SOCK_DONE
);
4034 switch (sk
->sk_state
) {
4036 case TCP_ESTABLISHED
:
4037 /* Move to CLOSE_WAIT */
4038 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
4039 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
4042 case TCP_CLOSE_WAIT
:
4044 /* Received a retransmission of the FIN, do
4049 /* RFC793: Remain in the LAST-ACK state. */
4053 /* This case occurs when a simultaneous close
4054 * happens, we must ack the received FIN and
4055 * enter the CLOSING state.
4058 tcp_set_state(sk
, TCP_CLOSING
);
4061 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4063 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
4066 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4067 * cases we should never reach this piece of code.
4069 pr_err("%s: Impossible, sk->sk_state=%d\n",
4070 __func__
, sk
->sk_state
);
4074 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4075 * Probably, we should reset in this case. For now drop them.
4077 skb_rbtree_purge(&tp
->out_of_order_queue
);
4078 if (tcp_is_sack(tp
))
4079 tcp_sack_reset(&tp
->rx_opt
);
4082 if (!sock_flag(sk
, SOCK_DEAD
)) {
4083 sk
->sk_state_change(sk
);
4085 /* Do not send POLL_HUP for half duplex close. */
4086 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
4087 sk
->sk_state
== TCP_CLOSE
)
4088 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_HUP
);
4090 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_IN
);
4094 static inline bool tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
,
4097 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
4098 if (before(seq
, sp
->start_seq
))
4099 sp
->start_seq
= seq
;
4100 if (after(end_seq
, sp
->end_seq
))
4101 sp
->end_seq
= end_seq
;
4107 static void tcp_dsack_set(struct sock
*sk
, u32 seq
, u32 end_seq
)
4109 struct tcp_sock
*tp
= tcp_sk(sk
);
4111 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
4114 if (before(seq
, tp
->rcv_nxt
))
4115 mib_idx
= LINUX_MIB_TCPDSACKOLDSENT
;
4117 mib_idx
= LINUX_MIB_TCPDSACKOFOSENT
;
4119 NET_INC_STATS(sock_net(sk
), mib_idx
);
4121 tp
->rx_opt
.dsack
= 1;
4122 tp
->duplicate_sack
[0].start_seq
= seq
;
4123 tp
->duplicate_sack
[0].end_seq
= end_seq
;
4127 static void tcp_dsack_extend(struct sock
*sk
, u32 seq
, u32 end_seq
)
4129 struct tcp_sock
*tp
= tcp_sk(sk
);
4131 if (!tp
->rx_opt
.dsack
)
4132 tcp_dsack_set(sk
, seq
, end_seq
);
4134 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
4137 static void tcp_send_dupack(struct sock
*sk
, const struct sk_buff
*skb
)
4139 struct tcp_sock
*tp
= tcp_sk(sk
);
4141 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
4142 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4143 NET_INC_STATS(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4144 tcp_enter_quickack_mode(sk
);
4146 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
4147 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4149 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
4150 end_seq
= tp
->rcv_nxt
;
4151 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, end_seq
);
4158 /* These routines update the SACK block as out-of-order packets arrive or
4159 * in-order packets close up the sequence space.
4161 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
4164 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4165 struct tcp_sack_block
*swalk
= sp
+ 1;
4167 /* See if the recent change to the first SACK eats into
4168 * or hits the sequence space of other SACK blocks, if so coalesce.
4170 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
;) {
4171 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
4174 /* Zap SWALK, by moving every further SACK up by one slot.
4175 * Decrease num_sacks.
4177 tp
->rx_opt
.num_sacks
--;
4178 for (i
= this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
4182 this_sack
++, swalk
++;
4186 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
4188 struct tcp_sock
*tp
= tcp_sk(sk
);
4189 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4190 int cur_sacks
= tp
->rx_opt
.num_sacks
;
4196 for (this_sack
= 0; this_sack
< cur_sacks
; this_sack
++, sp
++) {
4197 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
4198 /* Rotate this_sack to the first one. */
4199 for (; this_sack
> 0; this_sack
--, sp
--)
4200 swap(*sp
, *(sp
- 1));
4202 tcp_sack_maybe_coalesce(tp
);
4207 /* Could not find an adjacent existing SACK, build a new one,
4208 * put it at the front, and shift everyone else down. We
4209 * always know there is at least one SACK present already here.
4211 * If the sack array is full, forget about the last one.
4213 if (this_sack
>= TCP_NUM_SACKS
) {
4215 tp
->rx_opt
.num_sacks
--;
4218 for (; this_sack
> 0; this_sack
--, sp
--)
4222 /* Build the new head SACK, and we're done. */
4223 sp
->start_seq
= seq
;
4224 sp
->end_seq
= end_seq
;
4225 tp
->rx_opt
.num_sacks
++;
4228 /* RCV.NXT advances, some SACKs should be eaten. */
4230 static void tcp_sack_remove(struct tcp_sock
*tp
)
4232 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4233 int num_sacks
= tp
->rx_opt
.num_sacks
;
4236 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4237 if (RB_EMPTY_ROOT(&tp
->out_of_order_queue
)) {
4238 tp
->rx_opt
.num_sacks
= 0;
4242 for (this_sack
= 0; this_sack
< num_sacks
;) {
4243 /* Check if the start of the sack is covered by RCV.NXT. */
4244 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
4247 /* RCV.NXT must cover all the block! */
4248 WARN_ON(before(tp
->rcv_nxt
, sp
->end_seq
));
4250 /* Zap this SACK, by moving forward any other SACKS. */
4251 for (i
= this_sack
+1; i
< num_sacks
; i
++)
4252 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
4259 tp
->rx_opt
.num_sacks
= num_sacks
;
4263 * tcp_try_coalesce - try to merge skb to prior one
4266 * @from: buffer to add in queue
4267 * @fragstolen: pointer to boolean
4269 * Before queueing skb @from after @to, try to merge them
4270 * to reduce overall memory use and queue lengths, if cost is small.
4271 * Packets in ofo or receive queues can stay a long time.
4272 * Better try to coalesce them right now to avoid future collapses.
4273 * Returns true if caller should free @from instead of queueing it
4275 static bool tcp_try_coalesce(struct sock
*sk
,
4277 struct sk_buff
*from
,
4282 *fragstolen
= false;
4284 /* Its possible this segment overlaps with prior segment in queue */
4285 if (TCP_SKB_CB(from
)->seq
!= TCP_SKB_CB(to
)->end_seq
)
4288 if (!skb_try_coalesce(to
, from
, fragstolen
, &delta
))
4291 atomic_add(delta
, &sk
->sk_rmem_alloc
);
4292 sk_mem_charge(sk
, delta
);
4293 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPRCVCOALESCE
);
4294 TCP_SKB_CB(to
)->end_seq
= TCP_SKB_CB(from
)->end_seq
;
4295 TCP_SKB_CB(to
)->ack_seq
= TCP_SKB_CB(from
)->ack_seq
;
4296 TCP_SKB_CB(to
)->tcp_flags
|= TCP_SKB_CB(from
)->tcp_flags
;
4300 static void tcp_drop(struct sock
*sk
, struct sk_buff
*skb
)
4302 sk_drops_add(sk
, skb
);
4306 /* This one checks to see if we can put data from the
4307 * out_of_order queue into the receive_queue.
4309 static void tcp_ofo_queue(struct sock
*sk
)
4311 struct tcp_sock
*tp
= tcp_sk(sk
);
4312 __u32 dsack_high
= tp
->rcv_nxt
;
4313 bool fin
, fragstolen
, eaten
;
4314 struct sk_buff
*skb
, *tail
;
4317 p
= rb_first(&tp
->out_of_order_queue
);
4319 skb
= rb_entry(p
, struct sk_buff
, rbnode
);
4320 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
4323 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
4324 __u32 dsack
= dsack_high
;
4325 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
4326 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
4327 tcp_dsack_extend(sk
, TCP_SKB_CB(skb
)->seq
, dsack
);
4330 rb_erase(&skb
->rbnode
, &tp
->out_of_order_queue
);
4332 if (unlikely(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))) {
4333 SOCK_DEBUG(sk
, "ofo packet was already received\n");
4337 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
4338 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4339 TCP_SKB_CB(skb
)->end_seq
);
4341 tail
= skb_peek_tail(&sk
->sk_receive_queue
);
4342 eaten
= tail
&& tcp_try_coalesce(sk
, tail
, skb
, &fragstolen
);
4343 tcp_rcv_nxt_update(tp
, TCP_SKB_CB(skb
)->end_seq
);
4344 fin
= TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
;
4346 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4348 kfree_skb_partial(skb
, fragstolen
);
4350 if (unlikely(fin
)) {
4352 /* tcp_fin() purges tp->out_of_order_queue,
4353 * so we must end this loop right now.
4360 static bool tcp_prune_ofo_queue(struct sock
*sk
);
4361 static int tcp_prune_queue(struct sock
*sk
);
4363 static int tcp_try_rmem_schedule(struct sock
*sk
, struct sk_buff
*skb
,
4366 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
4367 !sk_rmem_schedule(sk
, skb
, size
)) {
4369 if (tcp_prune_queue(sk
) < 0)
4372 while (!sk_rmem_schedule(sk
, skb
, size
)) {
4373 if (!tcp_prune_ofo_queue(sk
))
4380 static void tcp_data_queue_ofo(struct sock
*sk
, struct sk_buff
*skb
)
4382 struct tcp_sock
*tp
= tcp_sk(sk
);
4383 struct rb_node
**p
, *q
, *parent
;
4384 struct sk_buff
*skb1
;
4388 tcp_ecn_check_ce(tp
, skb
);
4390 if (unlikely(tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))) {
4391 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPOFODROP
);
4396 /* Disable header prediction. */
4398 inet_csk_schedule_ack(sk
);
4400 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPOFOQUEUE
);
4401 seq
= TCP_SKB_CB(skb
)->seq
;
4402 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4403 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
4404 tp
->rcv_nxt
, seq
, end_seq
);
4406 p
= &tp
->out_of_order_queue
.rb_node
;
4407 if (RB_EMPTY_ROOT(&tp
->out_of_order_queue
)) {
4408 /* Initial out of order segment, build 1 SACK. */
4409 if (tcp_is_sack(tp
)) {
4410 tp
->rx_opt
.num_sacks
= 1;
4411 tp
->selective_acks
[0].start_seq
= seq
;
4412 tp
->selective_acks
[0].end_seq
= end_seq
;
4414 rb_link_node(&skb
->rbnode
, NULL
, p
);
4415 rb_insert_color(&skb
->rbnode
, &tp
->out_of_order_queue
);
4416 tp
->ooo_last_skb
= skb
;
4420 /* In the typical case, we are adding an skb to the end of the list.
4421 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4423 if (tcp_try_coalesce(sk
, tp
->ooo_last_skb
, skb
, &fragstolen
)) {
4425 tcp_grow_window(sk
, skb
);
4426 kfree_skb_partial(skb
, fragstolen
);
4430 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4431 if (!before(seq
, TCP_SKB_CB(tp
->ooo_last_skb
)->end_seq
)) {
4432 parent
= &tp
->ooo_last_skb
->rbnode
;
4433 p
= &parent
->rb_right
;
4437 /* Find place to insert this segment. Handle overlaps on the way. */
4441 skb1
= rb_entry(parent
, struct sk_buff
, rbnode
);
4442 if (before(seq
, TCP_SKB_CB(skb1
)->seq
)) {
4443 p
= &parent
->rb_left
;
4446 if (before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4447 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4448 /* All the bits are present. Drop. */
4449 NET_INC_STATS(sock_net(sk
),
4450 LINUX_MIB_TCPOFOMERGE
);
4453 tcp_dsack_set(sk
, seq
, end_seq
);
4456 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
4457 /* Partial overlap. */
4458 tcp_dsack_set(sk
, seq
, TCP_SKB_CB(skb1
)->end_seq
);
4460 /* skb's seq == skb1's seq and skb covers skb1.
4461 * Replace skb1 with skb.
4463 rb_replace_node(&skb1
->rbnode
, &skb
->rbnode
,
4464 &tp
->out_of_order_queue
);
4465 tcp_dsack_extend(sk
,
4466 TCP_SKB_CB(skb1
)->seq
,
4467 TCP_SKB_CB(skb1
)->end_seq
);
4468 NET_INC_STATS(sock_net(sk
),
4469 LINUX_MIB_TCPOFOMERGE
);
4473 } else if (tcp_try_coalesce(sk
, skb1
, skb
, &fragstolen
)) {
4476 p
= &parent
->rb_right
;
4479 /* Insert segment into RB tree. */
4480 rb_link_node(&skb
->rbnode
, parent
, p
);
4481 rb_insert_color(&skb
->rbnode
, &tp
->out_of_order_queue
);
4484 /* Remove other segments covered by skb. */
4485 while ((q
= rb_next(&skb
->rbnode
)) != NULL
) {
4486 skb1
= rb_entry(q
, struct sk_buff
, rbnode
);
4488 if (!after(end_seq
, TCP_SKB_CB(skb1
)->seq
))
4490 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4491 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4495 rb_erase(&skb1
->rbnode
, &tp
->out_of_order_queue
);
4496 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4497 TCP_SKB_CB(skb1
)->end_seq
);
4498 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPOFOMERGE
);
4501 /* If there is no skb after us, we are the last_skb ! */
4503 tp
->ooo_last_skb
= skb
;
4506 if (tcp_is_sack(tp
))
4507 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
4510 tcp_grow_window(sk
, skb
);
4512 skb_set_owner_r(skb
, sk
);
4516 static int __must_check
tcp_queue_rcv(struct sock
*sk
, struct sk_buff
*skb
, int hdrlen
,
4520 struct sk_buff
*tail
= skb_peek_tail(&sk
->sk_receive_queue
);
4522 __skb_pull(skb
, hdrlen
);
4524 tcp_try_coalesce(sk
, tail
, skb
, fragstolen
)) ? 1 : 0;
4525 tcp_rcv_nxt_update(tcp_sk(sk
), TCP_SKB_CB(skb
)->end_seq
);
4527 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4528 skb_set_owner_r(skb
, sk
);
4533 int tcp_send_rcvq(struct sock
*sk
, struct msghdr
*msg
, size_t size
)
4535 struct sk_buff
*skb
;
4543 if (size
> PAGE_SIZE
) {
4544 int npages
= min_t(size_t, size
>> PAGE_SHIFT
, MAX_SKB_FRAGS
);
4546 data_len
= npages
<< PAGE_SHIFT
;
4547 size
= data_len
+ (size
& ~PAGE_MASK
);
4549 skb
= alloc_skb_with_frags(size
- data_len
, data_len
,
4550 PAGE_ALLOC_COSTLY_ORDER
,
4551 &err
, sk
->sk_allocation
);
4555 skb_put(skb
, size
- data_len
);
4556 skb
->data_len
= data_len
;
4559 if (tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))
4562 err
= skb_copy_datagram_from_iter(skb
, 0, &msg
->msg_iter
, size
);
4566 TCP_SKB_CB(skb
)->seq
= tcp_sk(sk
)->rcv_nxt
;
4567 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ size
;
4568 TCP_SKB_CB(skb
)->ack_seq
= tcp_sk(sk
)->snd_una
- 1;
4570 if (tcp_queue_rcv(sk
, skb
, 0, &fragstolen
)) {
4571 WARN_ON_ONCE(fragstolen
); /* should not happen */
4583 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
4585 struct tcp_sock
*tp
= tcp_sk(sk
);
4586 bool fragstolen
= false;
4589 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
) {
4594 __skb_pull(skb
, tcp_hdr(skb
)->doff
* 4);
4596 tcp_ecn_accept_cwr(tp
, skb
);
4598 tp
->rx_opt
.dsack
= 0;
4600 /* Queue data for delivery to the user.
4601 * Packets in sequence go to the receive queue.
4602 * Out of sequence packets to the out_of_order_queue.
4604 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
4605 if (tcp_receive_window(tp
) == 0)
4608 /* Ok. In sequence. In window. */
4609 if (tp
->ucopy
.task
== current
&&
4610 tp
->copied_seq
== tp
->rcv_nxt
&& tp
->ucopy
.len
&&
4611 sock_owned_by_user(sk
) && !tp
->urg_data
) {
4612 int chunk
= min_t(unsigned int, skb
->len
,
4615 __set_current_state(TASK_RUNNING
);
4617 if (!skb_copy_datagram_msg(skb
, 0, tp
->ucopy
.msg
, chunk
)) {
4618 tp
->ucopy
.len
-= chunk
;
4619 tp
->copied_seq
+= chunk
;
4620 eaten
= (chunk
== skb
->len
);
4621 tcp_rcv_space_adjust(sk
);
4628 if (skb_queue_len(&sk
->sk_receive_queue
) == 0)
4629 sk_forced_mem_schedule(sk
, skb
->truesize
);
4630 else if (tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))
4633 eaten
= tcp_queue_rcv(sk
, skb
, 0, &fragstolen
);
4635 tcp_rcv_nxt_update(tp
, TCP_SKB_CB(skb
)->end_seq
);
4637 tcp_event_data_recv(sk
, skb
);
4638 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
4641 if (!RB_EMPTY_ROOT(&tp
->out_of_order_queue
)) {
4644 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4645 * gap in queue is filled.
4647 if (RB_EMPTY_ROOT(&tp
->out_of_order_queue
))
4648 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
4651 if (tp
->rx_opt
.num_sacks
)
4652 tcp_sack_remove(tp
);
4654 tcp_fast_path_check(sk
);
4657 kfree_skb_partial(skb
, fragstolen
);
4658 if (!sock_flag(sk
, SOCK_DEAD
))
4659 sk
->sk_data_ready(sk
);
4663 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4664 /* A retransmit, 2nd most common case. Force an immediate ack. */
4665 NET_INC_STATS(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4666 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4669 tcp_enter_quickack_mode(sk
);
4670 inet_csk_schedule_ack(sk
);
4676 /* Out of window. F.e. zero window probe. */
4677 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
4680 tcp_enter_quickack_mode(sk
);
4682 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4683 /* Partial packet, seq < rcv_next < end_seq */
4684 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
4685 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4686 TCP_SKB_CB(skb
)->end_seq
);
4688 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
4690 /* If window is closed, drop tail of packet. But after
4691 * remembering D-SACK for its head made in previous line.
4693 if (!tcp_receive_window(tp
))
4698 tcp_data_queue_ofo(sk
, skb
);
4701 static struct sk_buff
*tcp_skb_next(struct sk_buff
*skb
, struct sk_buff_head
*list
)
4704 return !skb_queue_is_last(list
, skb
) ? skb
->next
: NULL
;
4706 return rb_entry_safe(rb_next(&skb
->rbnode
), struct sk_buff
, rbnode
);
4709 static struct sk_buff
*tcp_collapse_one(struct sock
*sk
, struct sk_buff
*skb
,
4710 struct sk_buff_head
*list
,
4711 struct rb_root
*root
)
4713 struct sk_buff
*next
= tcp_skb_next(skb
, list
);
4716 __skb_unlink(skb
, list
);
4718 rb_erase(&skb
->rbnode
, root
);
4721 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPRCVCOLLAPSED
);
4726 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
4727 static void tcp_rbtree_insert(struct rb_root
*root
, struct sk_buff
*skb
)
4729 struct rb_node
**p
= &root
->rb_node
;
4730 struct rb_node
*parent
= NULL
;
4731 struct sk_buff
*skb1
;
4735 skb1
= rb_entry(parent
, struct sk_buff
, rbnode
);
4736 if (before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb1
)->seq
))
4737 p
= &parent
->rb_left
;
4739 p
= &parent
->rb_right
;
4741 rb_link_node(&skb
->rbnode
, parent
, p
);
4742 rb_insert_color(&skb
->rbnode
, root
);
4745 /* Collapse contiguous sequence of skbs head..tail with
4746 * sequence numbers start..end.
4748 * If tail is NULL, this means until the end of the queue.
4750 * Segments with FIN/SYN are not collapsed (only because this
4754 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
, struct rb_root
*root
,
4755 struct sk_buff
*head
, struct sk_buff
*tail
, u32 start
, u32 end
)
4757 struct sk_buff
*skb
= head
, *n
;
4758 struct sk_buff_head tmp
;
4761 /* First, check that queue is collapsible and find
4762 * the point where collapsing can be useful.
4765 for (end_of_skbs
= true; skb
!= NULL
&& skb
!= tail
; skb
= n
) {
4766 n
= tcp_skb_next(skb
, list
);
4768 /* No new bits? It is possible on ofo queue. */
4769 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4770 skb
= tcp_collapse_one(sk
, skb
, list
, root
);
4776 /* The first skb to collapse is:
4778 * - bloated or contains data before "start" or
4779 * overlaps to the next one.
4781 if (!(TCP_SKB_CB(skb
)->tcp_flags
& (TCPHDR_SYN
| TCPHDR_FIN
)) &&
4782 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
4783 before(TCP_SKB_CB(skb
)->seq
, start
))) {
4784 end_of_skbs
= false;
4788 if (n
&& n
!= tail
&&
4789 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(n
)->seq
) {
4790 end_of_skbs
= false;
4794 /* Decided to skip this, advance start seq. */
4795 start
= TCP_SKB_CB(skb
)->end_seq
;
4798 (TCP_SKB_CB(skb
)->tcp_flags
& (TCPHDR_SYN
| TCPHDR_FIN
)))
4801 __skb_queue_head_init(&tmp
);
4803 while (before(start
, end
)) {
4804 int copy
= min_t(int, SKB_MAX_ORDER(0, 0), end
- start
);
4805 struct sk_buff
*nskb
;
4807 nskb
= alloc_skb(copy
, GFP_ATOMIC
);
4811 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
4812 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
4814 __skb_queue_before(list
, skb
, nskb
);
4816 __skb_queue_tail(&tmp
, nskb
); /* defer rbtree insertion */
4817 skb_set_owner_r(nskb
, sk
);
4819 /* Copy data, releasing collapsed skbs. */
4821 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
4822 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
4826 size
= min(copy
, size
);
4827 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
4829 TCP_SKB_CB(nskb
)->end_seq
+= size
;
4833 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4834 skb
= tcp_collapse_one(sk
, skb
, list
, root
);
4837 (TCP_SKB_CB(skb
)->tcp_flags
& (TCPHDR_SYN
| TCPHDR_FIN
)))
4843 skb_queue_walk_safe(&tmp
, skb
, n
)
4844 tcp_rbtree_insert(root
, skb
);
4847 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4848 * and tcp_collapse() them until all the queue is collapsed.
4850 static void tcp_collapse_ofo_queue(struct sock
*sk
)
4852 struct tcp_sock
*tp
= tcp_sk(sk
);
4853 struct sk_buff
*skb
, *head
;
4857 p
= rb_first(&tp
->out_of_order_queue
);
4858 skb
= rb_entry_safe(p
, struct sk_buff
, rbnode
);
4861 p
= rb_last(&tp
->out_of_order_queue
);
4862 /* Note: This is possible p is NULL here. We do not
4863 * use rb_entry_safe(), as ooo_last_skb is valid only
4864 * if rbtree is not empty.
4866 tp
->ooo_last_skb
= rb_entry(p
, struct sk_buff
, rbnode
);
4869 start
= TCP_SKB_CB(skb
)->seq
;
4870 end
= TCP_SKB_CB(skb
)->end_seq
;
4872 for (head
= skb
;;) {
4873 skb
= tcp_skb_next(skb
, NULL
);
4875 /* Range is terminated when we see a gap or when
4876 * we are at the queue end.
4879 after(TCP_SKB_CB(skb
)->seq
, end
) ||
4880 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
4881 tcp_collapse(sk
, NULL
, &tp
->out_of_order_queue
,
4882 head
, skb
, start
, end
);
4886 if (unlikely(before(TCP_SKB_CB(skb
)->seq
, start
)))
4887 start
= TCP_SKB_CB(skb
)->seq
;
4888 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
4889 end
= TCP_SKB_CB(skb
)->end_seq
;
4894 * Clean the out-of-order queue to make room.
4895 * We drop high sequences packets to :
4896 * 1) Let a chance for holes to be filled.
4897 * 2) not add too big latencies if thousands of packets sit there.
4898 * (But if application shrinks SO_RCVBUF, we could still end up
4899 * freeing whole queue here)
4901 * Return true if queue has shrunk.
4903 static bool tcp_prune_ofo_queue(struct sock
*sk
)
4905 struct tcp_sock
*tp
= tcp_sk(sk
);
4906 struct rb_node
*node
, *prev
;
4908 if (RB_EMPTY_ROOT(&tp
->out_of_order_queue
))
4911 NET_INC_STATS(sock_net(sk
), LINUX_MIB_OFOPRUNED
);
4912 node
= &tp
->ooo_last_skb
->rbnode
;
4914 prev
= rb_prev(node
);
4915 rb_erase(node
, &tp
->out_of_order_queue
);
4916 tcp_drop(sk
, rb_entry(node
, struct sk_buff
, rbnode
));
4918 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
&&
4919 !tcp_under_memory_pressure(sk
))
4923 tp
->ooo_last_skb
= rb_entry(prev
, struct sk_buff
, rbnode
);
4925 /* Reset SACK state. A conforming SACK implementation will
4926 * do the same at a timeout based retransmit. When a connection
4927 * is in a sad state like this, we care only about integrity
4928 * of the connection not performance.
4930 if (tp
->rx_opt
.sack_ok
)
4931 tcp_sack_reset(&tp
->rx_opt
);
4935 /* Reduce allocated memory if we can, trying to get
4936 * the socket within its memory limits again.
4938 * Return less than zero if we should start dropping frames
4939 * until the socket owning process reads some of the data
4940 * to stabilize the situation.
4942 static int tcp_prune_queue(struct sock
*sk
)
4944 struct tcp_sock
*tp
= tcp_sk(sk
);
4946 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
4948 NET_INC_STATS(sock_net(sk
), LINUX_MIB_PRUNECALLED
);
4950 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
4951 tcp_clamp_window(sk
);
4952 else if (tcp_under_memory_pressure(sk
))
4953 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
4955 tcp_collapse_ofo_queue(sk
);
4956 if (!skb_queue_empty(&sk
->sk_receive_queue
))
4957 tcp_collapse(sk
, &sk
->sk_receive_queue
, NULL
,
4958 skb_peek(&sk
->sk_receive_queue
),
4960 tp
->copied_seq
, tp
->rcv_nxt
);
4963 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4966 /* Collapsing did not help, destructive actions follow.
4967 * This must not ever occur. */
4969 tcp_prune_ofo_queue(sk
);
4971 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4974 /* If we are really being abused, tell the caller to silently
4975 * drop receive data on the floor. It will get retransmitted
4976 * and hopefully then we'll have sufficient space.
4978 NET_INC_STATS(sock_net(sk
), LINUX_MIB_RCVPRUNED
);
4980 /* Massive buffer overcommit. */
4985 static bool tcp_should_expand_sndbuf(const struct sock
*sk
)
4987 const struct tcp_sock
*tp
= tcp_sk(sk
);
4989 /* If the user specified a specific send buffer setting, do
4992 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
4995 /* If we are under global TCP memory pressure, do not expand. */
4996 if (tcp_under_memory_pressure(sk
))
4999 /* If we are under soft global TCP memory pressure, do not expand. */
5000 if (sk_memory_allocated(sk
) >= sk_prot_mem_limits(sk
, 0))
5003 /* If we filled the congestion window, do not expand. */
5004 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
5010 /* When incoming ACK allowed to free some skb from write_queue,
5011 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
5012 * on the exit from tcp input handler.
5014 * PROBLEM: sndbuf expansion does not work well with largesend.
5016 static void tcp_new_space(struct sock
*sk
)
5018 struct tcp_sock
*tp
= tcp_sk(sk
);
5020 if (tcp_should_expand_sndbuf(sk
)) {
5021 tcp_sndbuf_expand(sk
);
5022 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
5025 sk
->sk_write_space(sk
);
5028 static void tcp_check_space(struct sock
*sk
)
5030 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
5031 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
5032 /* pairs with tcp_poll() */
5034 if (sk
->sk_socket
&&
5035 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
5037 if (!test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
5038 tcp_chrono_stop(sk
, TCP_CHRONO_SNDBUF_LIMITED
);
5043 static inline void tcp_data_snd_check(struct sock
*sk
)
5045 tcp_push_pending_frames(sk
);
5046 tcp_check_space(sk
);
5050 * Check if sending an ack is needed.
5052 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
5054 struct tcp_sock
*tp
= tcp_sk(sk
);
5056 /* More than one full frame received... */
5057 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
&&
5058 /* ... and right edge of window advances far enough.
5059 * (tcp_recvmsg() will send ACK otherwise). Or...
5061 __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
5062 /* We ACK each frame or... */
5063 tcp_in_quickack_mode(sk
) ||
5064 /* We have out of order data. */
5065 (ofo_possible
&& !RB_EMPTY_ROOT(&tp
->out_of_order_queue
))) {
5066 /* Then ack it now */
5069 /* Else, send delayed ack. */
5070 tcp_send_delayed_ack(sk
);
5074 static inline void tcp_ack_snd_check(struct sock
*sk
)
5076 if (!inet_csk_ack_scheduled(sk
)) {
5077 /* We sent a data segment already. */
5080 __tcp_ack_snd_check(sk
, 1);
5084 * This routine is only called when we have urgent data
5085 * signaled. Its the 'slow' part of tcp_urg. It could be
5086 * moved inline now as tcp_urg is only called from one
5087 * place. We handle URGent data wrong. We have to - as
5088 * BSD still doesn't use the correction from RFC961.
5089 * For 1003.1g we should support a new option TCP_STDURG to permit
5090 * either form (or just set the sysctl tcp_stdurg).
5093 static void tcp_check_urg(struct sock
*sk
, const struct tcphdr
*th
)
5095 struct tcp_sock
*tp
= tcp_sk(sk
);
5096 u32 ptr
= ntohs(th
->urg_ptr
);
5098 if (ptr
&& !sysctl_tcp_stdurg
)
5100 ptr
+= ntohl(th
->seq
);
5102 /* Ignore urgent data that we've already seen and read. */
5103 if (after(tp
->copied_seq
, ptr
))
5106 /* Do not replay urg ptr.
5108 * NOTE: interesting situation not covered by specs.
5109 * Misbehaving sender may send urg ptr, pointing to segment,
5110 * which we already have in ofo queue. We are not able to fetch
5111 * such data and will stay in TCP_URG_NOTYET until will be eaten
5112 * by recvmsg(). Seems, we are not obliged to handle such wicked
5113 * situations. But it is worth to think about possibility of some
5114 * DoSes using some hypothetical application level deadlock.
5116 if (before(ptr
, tp
->rcv_nxt
))
5119 /* Do we already have a newer (or duplicate) urgent pointer? */
5120 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
5123 /* Tell the world about our new urgent pointer. */
5126 /* We may be adding urgent data when the last byte read was
5127 * urgent. To do this requires some care. We cannot just ignore
5128 * tp->copied_seq since we would read the last urgent byte again
5129 * as data, nor can we alter copied_seq until this data arrives
5130 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5132 * NOTE. Double Dutch. Rendering to plain English: author of comment
5133 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5134 * and expect that both A and B disappear from stream. This is _wrong_.
5135 * Though this happens in BSD with high probability, this is occasional.
5136 * Any application relying on this is buggy. Note also, that fix "works"
5137 * only in this artificial test. Insert some normal data between A and B and we will
5138 * decline of BSD again. Verdict: it is better to remove to trap
5141 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
5142 !sock_flag(sk
, SOCK_URGINLINE
) && tp
->copied_seq
!= tp
->rcv_nxt
) {
5143 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
5145 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5146 __skb_unlink(skb
, &sk
->sk_receive_queue
);
5151 tp
->urg_data
= TCP_URG_NOTYET
;
5154 /* Disable header prediction. */
5158 /* This is the 'fast' part of urgent handling. */
5159 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, const struct tcphdr
*th
)
5161 struct tcp_sock
*tp
= tcp_sk(sk
);
5163 /* Check if we get a new urgent pointer - normally not. */
5165 tcp_check_urg(sk
, th
);
5167 /* Do we wait for any urgent data? - normally not... */
5168 if (tp
->urg_data
== TCP_URG_NOTYET
) {
5169 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
5172 /* Is the urgent pointer pointing into this packet? */
5173 if (ptr
< skb
->len
) {
5175 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
5177 tp
->urg_data
= TCP_URG_VALID
| tmp
;
5178 if (!sock_flag(sk
, SOCK_DEAD
))
5179 sk
->sk_data_ready(sk
);
5184 static int tcp_copy_to_iovec(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
5186 struct tcp_sock
*tp
= tcp_sk(sk
);
5187 int chunk
= skb
->len
- hlen
;
5190 if (skb_csum_unnecessary(skb
))
5191 err
= skb_copy_datagram_msg(skb
, hlen
, tp
->ucopy
.msg
, chunk
);
5193 err
= skb_copy_and_csum_datagram_msg(skb
, hlen
, tp
->ucopy
.msg
);
5196 tp
->ucopy
.len
-= chunk
;
5197 tp
->copied_seq
+= chunk
;
5198 tcp_rcv_space_adjust(sk
);
5204 /* Accept RST for rcv_nxt - 1 after a FIN.
5205 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a
5206 * FIN is sent followed by a RST packet. The RST is sent with the same
5207 * sequence number as the FIN, and thus according to RFC 5961 a challenge
5208 * ACK should be sent. However, Mac OSX rate limits replies to challenge
5209 * ACKs on the closed socket. In addition middleboxes can drop either the
5210 * challenge ACK or a subsequent RST.
5212 static bool tcp_reset_check(const struct sock
*sk
, const struct sk_buff
*skb
)
5214 struct tcp_sock
*tp
= tcp_sk(sk
);
5216 return unlikely(TCP_SKB_CB(skb
)->seq
== (tp
->rcv_nxt
- 1) &&
5217 (1 << sk
->sk_state
) & (TCPF_CLOSE_WAIT
| TCPF_LAST_ACK
|
5221 /* Does PAWS and seqno based validation of an incoming segment, flags will
5222 * play significant role here.
5224 static bool tcp_validate_incoming(struct sock
*sk
, struct sk_buff
*skb
,
5225 const struct tcphdr
*th
, int syn_inerr
)
5227 struct tcp_sock
*tp
= tcp_sk(sk
);
5228 bool rst_seq_match
= false;
5230 /* RFC1323: H1. Apply PAWS check first. */
5231 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
5232 tcp_paws_discard(sk
, skb
)) {
5234 NET_INC_STATS(sock_net(sk
), LINUX_MIB_PAWSESTABREJECTED
);
5235 if (!tcp_oow_rate_limited(sock_net(sk
), skb
,
5236 LINUX_MIB_TCPACKSKIPPEDPAWS
,
5237 &tp
->last_oow_ack_time
))
5238 tcp_send_dupack(sk
, skb
);
5241 /* Reset is accepted even if it did not pass PAWS. */
5244 /* Step 1: check sequence number */
5245 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5246 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5247 * (RST) segments are validated by checking their SEQ-fields."
5248 * And page 69: "If an incoming segment is not acceptable,
5249 * an acknowledgment should be sent in reply (unless the RST
5250 * bit is set, if so drop the segment and return)".
5255 if (!tcp_oow_rate_limited(sock_net(sk
), skb
,
5256 LINUX_MIB_TCPACKSKIPPEDSEQ
,
5257 &tp
->last_oow_ack_time
))
5258 tcp_send_dupack(sk
, skb
);
5259 } else if (tcp_reset_check(sk
, skb
)) {
5265 /* Step 2: check RST bit */
5267 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a
5268 * FIN and SACK too if available):
5269 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or
5270 * the right-most SACK block,
5272 * RESET the connection
5274 * Send a challenge ACK
5276 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
||
5277 tcp_reset_check(sk
, skb
)) {
5278 rst_seq_match
= true;
5279 } else if (tcp_is_sack(tp
) && tp
->rx_opt
.num_sacks
> 0) {
5280 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
5281 int max_sack
= sp
[0].end_seq
;
5284 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
;
5286 max_sack
= after(sp
[this_sack
].end_seq
,
5288 sp
[this_sack
].end_seq
: max_sack
;
5291 if (TCP_SKB_CB(skb
)->seq
== max_sack
)
5292 rst_seq_match
= true;
5298 tcp_send_challenge_ack(sk
, skb
);
5302 /* step 3: check security and precedence [ignored] */
5304 /* step 4: Check for a SYN
5305 * RFC 5961 4.2 : Send a challenge ack
5310 TCP_INC_STATS(sock_net(sk
), TCP_MIB_INERRS
);
5311 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPSYNCHALLENGE
);
5312 tcp_send_challenge_ack(sk
, skb
);
5324 * TCP receive function for the ESTABLISHED state.
5326 * It is split into a fast path and a slow path. The fast path is
5328 * - A zero window was announced from us - zero window probing
5329 * is only handled properly in the slow path.
5330 * - Out of order segments arrived.
5331 * - Urgent data is expected.
5332 * - There is no buffer space left
5333 * - Unexpected TCP flags/window values/header lengths are received
5334 * (detected by checking the TCP header against pred_flags)
5335 * - Data is sent in both directions. Fast path only supports pure senders
5336 * or pure receivers (this means either the sequence number or the ack
5337 * value must stay constant)
5338 * - Unexpected TCP option.
5340 * When these conditions are not satisfied it drops into a standard
5341 * receive procedure patterned after RFC793 to handle all cases.
5342 * The first three cases are guaranteed by proper pred_flags setting,
5343 * the rest is checked inline. Fast processing is turned on in
5344 * tcp_data_queue when everything is OK.
5346 void tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
5347 const struct tcphdr
*th
, unsigned int len
)
5349 struct tcp_sock
*tp
= tcp_sk(sk
);
5351 if (unlikely(!sk
->sk_rx_dst
))
5352 inet_csk(sk
)->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
5354 * Header prediction.
5355 * The code loosely follows the one in the famous
5356 * "30 instruction TCP receive" Van Jacobson mail.
5358 * Van's trick is to deposit buffers into socket queue
5359 * on a device interrupt, to call tcp_recv function
5360 * on the receive process context and checksum and copy
5361 * the buffer to user space. smart...
5363 * Our current scheme is not silly either but we take the
5364 * extra cost of the net_bh soft interrupt processing...
5365 * We do checksum and copy also but from device to kernel.
5368 tp
->rx_opt
.saw_tstamp
= 0;
5370 /* pred_flags is 0xS?10 << 16 + snd_wnd
5371 * if header_prediction is to be made
5372 * 'S' will always be tp->tcp_header_len >> 2
5373 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5374 * turn it off (when there are holes in the receive
5375 * space for instance)
5376 * PSH flag is ignored.
5379 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
5380 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
&&
5381 !after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
)) {
5382 int tcp_header_len
= tp
->tcp_header_len
;
5384 /* Timestamp header prediction: tcp_header_len
5385 * is automatically equal to th->doff*4 due to pred_flags
5389 /* Check timestamp */
5390 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
5391 /* No? Slow path! */
5392 if (!tcp_parse_aligned_timestamp(tp
, th
))
5395 /* If PAWS failed, check it more carefully in slow path */
5396 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
5399 /* DO NOT update ts_recent here, if checksum fails
5400 * and timestamp was corrupted part, it will result
5401 * in a hung connection since we will drop all
5402 * future packets due to the PAWS test.
5406 if (len
<= tcp_header_len
) {
5407 /* Bulk data transfer: sender */
5408 if (len
== tcp_header_len
) {
5409 /* Predicted packet is in window by definition.
5410 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5411 * Hence, check seq<=rcv_wup reduces to:
5413 if (tcp_header_len
==
5414 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5415 tp
->rcv_nxt
== tp
->rcv_wup
)
5416 tcp_store_ts_recent(tp
);
5418 /* We know that such packets are checksummed
5421 tcp_ack(sk
, skb
, 0);
5423 tcp_data_snd_check(sk
);
5425 } else { /* Header too small */
5426 TCP_INC_STATS(sock_net(sk
), TCP_MIB_INERRS
);
5431 bool fragstolen
= false;
5433 if (tp
->ucopy
.task
== current
&&
5434 tp
->copied_seq
== tp
->rcv_nxt
&&
5435 len
- tcp_header_len
<= tp
->ucopy
.len
&&
5436 sock_owned_by_user(sk
)) {
5437 __set_current_state(TASK_RUNNING
);
5439 if (!tcp_copy_to_iovec(sk
, skb
, tcp_header_len
)) {
5440 /* Predicted packet is in window by definition.
5441 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5442 * Hence, check seq<=rcv_wup reduces to:
5444 if (tcp_header_len
==
5445 (sizeof(struct tcphdr
) +
5446 TCPOLEN_TSTAMP_ALIGNED
) &&
5447 tp
->rcv_nxt
== tp
->rcv_wup
)
5448 tcp_store_ts_recent(tp
);
5450 tcp_rcv_rtt_measure_ts(sk
, skb
);
5452 __skb_pull(skb
, tcp_header_len
);
5453 tcp_rcv_nxt_update(tp
, TCP_SKB_CB(skb
)->end_seq
);
5454 NET_INC_STATS(sock_net(sk
),
5455 LINUX_MIB_TCPHPHITSTOUSER
);
5460 if (tcp_checksum_complete(skb
))
5463 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
5466 /* Predicted packet is in window by definition.
5467 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5468 * Hence, check seq<=rcv_wup reduces to:
5470 if (tcp_header_len
==
5471 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5472 tp
->rcv_nxt
== tp
->rcv_wup
)
5473 tcp_store_ts_recent(tp
);
5475 tcp_rcv_rtt_measure_ts(sk
, skb
);
5477 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPHPHITS
);
5479 /* Bulk data transfer: receiver */
5480 eaten
= tcp_queue_rcv(sk
, skb
, tcp_header_len
,
5484 tcp_event_data_recv(sk
, skb
);
5486 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
5487 /* Well, only one small jumplet in fast path... */
5488 tcp_ack(sk
, skb
, FLAG_DATA
);
5489 tcp_data_snd_check(sk
);
5490 if (!inet_csk_ack_scheduled(sk
))
5494 __tcp_ack_snd_check(sk
, 0);
5497 kfree_skb_partial(skb
, fragstolen
);
5498 sk
->sk_data_ready(sk
);
5504 if (len
< (th
->doff
<< 2) || tcp_checksum_complete(skb
))
5507 if (!th
->ack
&& !th
->rst
&& !th
->syn
)
5511 * Standard slow path.
5514 if (!tcp_validate_incoming(sk
, skb
, th
, 1))
5518 if (tcp_ack(sk
, skb
, FLAG_SLOWPATH
| FLAG_UPDATE_TS_RECENT
) < 0)
5521 tcp_rcv_rtt_measure_ts(sk
, skb
);
5523 /* Process urgent data. */
5524 tcp_urg(sk
, skb
, th
);
5526 /* step 7: process the segment text */
5527 tcp_data_queue(sk
, skb
);
5529 tcp_data_snd_check(sk
);
5530 tcp_ack_snd_check(sk
);
5534 TCP_INC_STATS(sock_net(sk
), TCP_MIB_CSUMERRORS
);
5535 TCP_INC_STATS(sock_net(sk
), TCP_MIB_INERRS
);
5540 EXPORT_SYMBOL(tcp_rcv_established
);
5542 void tcp_finish_connect(struct sock
*sk
, struct sk_buff
*skb
)
5544 struct tcp_sock
*tp
= tcp_sk(sk
);
5545 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5547 tcp_set_state(sk
, TCP_ESTABLISHED
);
5548 icsk
->icsk_ack
.lrcvtime
= tcp_time_stamp
;
5551 icsk
->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
5552 security_inet_conn_established(sk
, skb
);
5555 /* Make sure socket is routed, for correct metrics. */
5556 icsk
->icsk_af_ops
->rebuild_header(sk
);
5558 tcp_init_metrics(sk
);
5560 tcp_init_congestion_control(sk
);
5562 /* Prevent spurious tcp_cwnd_restart() on first data
5565 tp
->lsndtime
= tcp_time_stamp
;
5567 tcp_init_buffer_space(sk
);
5569 if (sock_flag(sk
, SOCK_KEEPOPEN
))
5570 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
5572 if (!tp
->rx_opt
.snd_wscale
)
5573 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
5577 if (!sock_flag(sk
, SOCK_DEAD
)) {
5578 sk
->sk_state_change(sk
);
5579 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5583 static bool tcp_rcv_fastopen_synack(struct sock
*sk
, struct sk_buff
*synack
,
5584 struct tcp_fastopen_cookie
*cookie
)
5586 struct tcp_sock
*tp
= tcp_sk(sk
);
5587 struct sk_buff
*data
= tp
->syn_data
? tcp_write_queue_head(sk
) : NULL
;
5588 u16 mss
= tp
->rx_opt
.mss_clamp
, try_exp
= 0;
5589 bool syn_drop
= false;
5591 if (mss
== tp
->rx_opt
.user_mss
) {
5592 struct tcp_options_received opt
;
5594 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5595 tcp_clear_options(&opt
);
5596 opt
.user_mss
= opt
.mss_clamp
= 0;
5597 tcp_parse_options(synack
, &opt
, 0, NULL
);
5598 mss
= opt
.mss_clamp
;
5601 if (!tp
->syn_fastopen
) {
5602 /* Ignore an unsolicited cookie */
5604 } else if (tp
->total_retrans
) {
5605 /* SYN timed out and the SYN-ACK neither has a cookie nor
5606 * acknowledges data. Presumably the remote received only
5607 * the retransmitted (regular) SYNs: either the original
5608 * SYN-data or the corresponding SYN-ACK was dropped.
5610 syn_drop
= (cookie
->len
< 0 && data
);
5611 } else if (cookie
->len
< 0 && !tp
->syn_data
) {
5612 /* We requested a cookie but didn't get it. If we did not use
5613 * the (old) exp opt format then try so next time (try_exp=1).
5614 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5616 try_exp
= tp
->syn_fastopen_exp
? 2 : 1;
5619 tcp_fastopen_cache_set(sk
, mss
, cookie
, syn_drop
, try_exp
);
5621 if (data
) { /* Retransmit unacked data in SYN */
5622 tcp_for_write_queue_from(data
, sk
) {
5623 if (data
== tcp_send_head(sk
) ||
5624 __tcp_retransmit_skb(sk
, data
, 1))
5628 NET_INC_STATS(sock_net(sk
),
5629 LINUX_MIB_TCPFASTOPENACTIVEFAIL
);
5632 tp
->syn_data_acked
= tp
->syn_data
;
5633 if (tp
->syn_data_acked
)
5634 NET_INC_STATS(sock_net(sk
),
5635 LINUX_MIB_TCPFASTOPENACTIVE
);
5637 tcp_fastopen_add_skb(sk
, synack
);
5642 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5643 const struct tcphdr
*th
)
5645 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5646 struct tcp_sock
*tp
= tcp_sk(sk
);
5647 struct tcp_fastopen_cookie foc
= { .len
= -1 };
5648 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
5650 tcp_parse_options(skb
, &tp
->rx_opt
, 0, &foc
);
5651 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
5652 tp
->rx_opt
.rcv_tsecr
-= tp
->tsoffset
;
5656 * "If the state is SYN-SENT then
5657 * first check the ACK bit
5658 * If the ACK bit is set
5659 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5660 * a reset (unless the RST bit is set, if so drop
5661 * the segment and return)"
5663 if (!after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_una
) ||
5664 after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
))
5665 goto reset_and_undo
;
5667 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
5668 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
5670 NET_INC_STATS(sock_net(sk
),
5671 LINUX_MIB_PAWSACTIVEREJECTED
);
5672 goto reset_and_undo
;
5675 /* Now ACK is acceptable.
5677 * "If the RST bit is set
5678 * If the ACK was acceptable then signal the user "error:
5679 * connection reset", drop the segment, enter CLOSED state,
5680 * delete TCB, and return."
5689 * "fifth, if neither of the SYN or RST bits is set then
5690 * drop the segment and return."
5696 goto discard_and_undo
;
5699 * "If the SYN bit is on ...
5700 * are acceptable then ...
5701 * (our SYN has been ACKed), change the connection
5702 * state to ESTABLISHED..."
5705 tcp_ecn_rcv_synack(tp
, th
);
5707 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5708 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
5710 /* Ok.. it's good. Set up sequence numbers and
5711 * move to established.
5713 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5714 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5716 /* RFC1323: The window in SYN & SYN/ACK segments is
5719 tp
->snd_wnd
= ntohs(th
->window
);
5721 if (!tp
->rx_opt
.wscale_ok
) {
5722 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
5723 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
5726 if (tp
->rx_opt
.saw_tstamp
) {
5727 tp
->rx_opt
.tstamp_ok
= 1;
5728 tp
->tcp_header_len
=
5729 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5730 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5731 tcp_store_ts_recent(tp
);
5733 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5736 if (tcp_is_sack(tp
) && sysctl_tcp_fack
)
5737 tcp_enable_fack(tp
);
5740 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5741 tcp_initialize_rcv_mss(sk
);
5743 /* Remember, tcp_poll() does not lock socket!
5744 * Change state from SYN-SENT only after copied_seq
5745 * is initialized. */
5746 tp
->copied_seq
= tp
->rcv_nxt
;
5750 tcp_finish_connect(sk
, skb
);
5752 if ((tp
->syn_fastopen
|| tp
->syn_data
) &&
5753 tcp_rcv_fastopen_synack(sk
, skb
, &foc
))
5756 if (sk
->sk_write_pending
||
5757 icsk
->icsk_accept_queue
.rskq_defer_accept
||
5758 icsk
->icsk_ack
.pingpong
) {
5759 /* Save one ACK. Data will be ready after
5760 * several ticks, if write_pending is set.
5762 * It may be deleted, but with this feature tcpdumps
5763 * look so _wonderfully_ clever, that I was not able
5764 * to stand against the temptation 8) --ANK
5766 inet_csk_schedule_ack(sk
);
5767 tcp_enter_quickack_mode(sk
);
5768 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
5769 TCP_DELACK_MAX
, TCP_RTO_MAX
);
5780 /* No ACK in the segment */
5784 * "If the RST bit is set
5786 * Otherwise (no ACK) drop the segment and return."
5789 goto discard_and_undo
;
5793 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&&
5794 tcp_paws_reject(&tp
->rx_opt
, 0))
5795 goto discard_and_undo
;
5798 /* We see SYN without ACK. It is attempt of
5799 * simultaneous connect with crossed SYNs.
5800 * Particularly, it can be connect to self.
5802 tcp_set_state(sk
, TCP_SYN_RECV
);
5804 if (tp
->rx_opt
.saw_tstamp
) {
5805 tp
->rx_opt
.tstamp_ok
= 1;
5806 tcp_store_ts_recent(tp
);
5807 tp
->tcp_header_len
=
5808 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5810 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5813 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5814 tp
->copied_seq
= tp
->rcv_nxt
;
5815 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5817 /* RFC1323: The window in SYN & SYN/ACK segments is
5820 tp
->snd_wnd
= ntohs(th
->window
);
5821 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
5822 tp
->max_window
= tp
->snd_wnd
;
5824 tcp_ecn_rcv_syn(tp
, th
);
5827 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5828 tcp_initialize_rcv_mss(sk
);
5830 tcp_send_synack(sk
);
5832 /* Note, we could accept data and URG from this segment.
5833 * There are no obstacles to make this (except that we must
5834 * either change tcp_recvmsg() to prevent it from returning data
5835 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5837 * However, if we ignore data in ACKless segments sometimes,
5838 * we have no reasons to accept it sometimes.
5839 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5840 * is not flawless. So, discard packet for sanity.
5841 * Uncomment this return to process the data.
5848 /* "fifth, if neither of the SYN or RST bits is set then
5849 * drop the segment and return."
5853 tcp_clear_options(&tp
->rx_opt
);
5854 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5858 tcp_clear_options(&tp
->rx_opt
);
5859 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5864 * This function implements the receiving procedure of RFC 793 for
5865 * all states except ESTABLISHED and TIME_WAIT.
5866 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5867 * address independent.
5870 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
)
5872 struct tcp_sock
*tp
= tcp_sk(sk
);
5873 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5874 const struct tcphdr
*th
= tcp_hdr(skb
);
5875 struct request_sock
*req
;
5879 switch (sk
->sk_state
) {
5893 /* It is possible that we process SYN packets from backlog,
5894 * so we need to make sure to disable BH right there.
5897 acceptable
= icsk
->icsk_af_ops
->conn_request(sk
, skb
) >= 0;
5908 tp
->rx_opt
.saw_tstamp
= 0;
5909 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
);
5913 /* Do step6 onward by hand. */
5914 tcp_urg(sk
, skb
, th
);
5916 tcp_data_snd_check(sk
);
5920 tp
->rx_opt
.saw_tstamp
= 0;
5921 req
= tp
->fastopen_rsk
;
5923 WARN_ON_ONCE(sk
->sk_state
!= TCP_SYN_RECV
&&
5924 sk
->sk_state
!= TCP_FIN_WAIT1
);
5926 if (!tcp_check_req(sk
, skb
, req
, true))
5930 if (!th
->ack
&& !th
->rst
&& !th
->syn
)
5933 if (!tcp_validate_incoming(sk
, skb
, th
, 0))
5936 /* step 5: check the ACK field */
5937 acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
|
5938 FLAG_UPDATE_TS_RECENT
) > 0;
5940 switch (sk
->sk_state
) {
5946 tcp_synack_rtt_meas(sk
, req
);
5948 /* Once we leave TCP_SYN_RECV, we no longer need req
5952 inet_csk(sk
)->icsk_retransmits
= 0;
5953 reqsk_fastopen_remove(sk
, req
, false);
5955 /* Make sure socket is routed, for correct metrics. */
5956 icsk
->icsk_af_ops
->rebuild_header(sk
);
5957 tcp_init_congestion_control(sk
);
5960 tp
->copied_seq
= tp
->rcv_nxt
;
5961 tcp_init_buffer_space(sk
);
5964 tcp_set_state(sk
, TCP_ESTABLISHED
);
5965 sk
->sk_state_change(sk
);
5967 /* Note, that this wakeup is only for marginal crossed SYN case.
5968 * Passively open sockets are not waked up, because
5969 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5972 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5974 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
5975 tp
->snd_wnd
= ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
;
5976 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5978 if (tp
->rx_opt
.tstamp_ok
)
5979 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5982 /* Re-arm the timer because data may have been sent out.
5983 * This is similar to the regular data transmission case
5984 * when new data has just been ack'ed.
5986 * (TFO) - we could try to be more aggressive and
5987 * retransmitting any data sooner based on when they
5992 tcp_init_metrics(sk
);
5994 if (!inet_csk(sk
)->icsk_ca_ops
->cong_control
)
5995 tcp_update_pacing_rate(sk
);
5997 /* Prevent spurious tcp_cwnd_restart() on first data packet */
5998 tp
->lsndtime
= tcp_time_stamp
;
6000 tcp_initialize_rcv_mss(sk
);
6001 tcp_fast_path_on(tp
);
6004 case TCP_FIN_WAIT1
: {
6007 /* If we enter the TCP_FIN_WAIT1 state and we are a
6008 * Fast Open socket and this is the first acceptable
6009 * ACK we have received, this would have acknowledged
6010 * our SYNACK so stop the SYNACK timer.
6013 /* Return RST if ack_seq is invalid.
6014 * Note that RFC793 only says to generate a
6015 * DUPACK for it but for TCP Fast Open it seems
6016 * better to treat this case like TCP_SYN_RECV
6021 /* We no longer need the request sock. */
6022 reqsk_fastopen_remove(sk
, req
, false);
6025 if (tp
->snd_una
!= tp
->write_seq
)
6028 tcp_set_state(sk
, TCP_FIN_WAIT2
);
6029 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
6033 if (!sock_flag(sk
, SOCK_DEAD
)) {
6034 /* Wake up lingering close() */
6035 sk
->sk_state_change(sk
);
6039 if (tp
->linger2
< 0 ||
6040 (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
6041 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
))) {
6043 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
6047 tmo
= tcp_fin_time(sk
);
6048 if (tmo
> TCP_TIMEWAIT_LEN
) {
6049 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
6050 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
6051 /* Bad case. We could lose such FIN otherwise.
6052 * It is not a big problem, but it looks confusing
6053 * and not so rare event. We still can lose it now,
6054 * if it spins in bh_lock_sock(), but it is really
6057 inet_csk_reset_keepalive_timer(sk
, tmo
);
6059 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
6066 if (tp
->snd_una
== tp
->write_seq
) {
6067 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
6073 if (tp
->snd_una
== tp
->write_seq
) {
6074 tcp_update_metrics(sk
);
6081 /* step 6: check the URG bit */
6082 tcp_urg(sk
, skb
, th
);
6084 /* step 7: process the segment text */
6085 switch (sk
->sk_state
) {
6086 case TCP_CLOSE_WAIT
:
6089 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
6093 /* RFC 793 says to queue data in these states,
6094 * RFC 1122 says we MUST send a reset.
6095 * BSD 4.4 also does reset.
6097 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
6098 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
6099 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
6100 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
6106 case TCP_ESTABLISHED
:
6107 tcp_data_queue(sk
, skb
);
6112 /* tcp_data could move socket to TIME-WAIT */
6113 if (sk
->sk_state
!= TCP_CLOSE
) {
6114 tcp_data_snd_check(sk
);
6115 tcp_ack_snd_check(sk
);
6124 EXPORT_SYMBOL(tcp_rcv_state_process
);
6126 static inline void pr_drop_req(struct request_sock
*req
, __u16 port
, int family
)
6128 struct inet_request_sock
*ireq
= inet_rsk(req
);
6130 if (family
== AF_INET
)
6131 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6132 &ireq
->ir_rmt_addr
, port
);
6133 #if IS_ENABLED(CONFIG_IPV6)
6134 else if (family
== AF_INET6
)
6135 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6136 &ireq
->ir_v6_rmt_addr
, port
);
6140 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6142 * If we receive a SYN packet with these bits set, it means a
6143 * network is playing bad games with TOS bits. In order to
6144 * avoid possible false congestion notifications, we disable
6145 * TCP ECN negotiation.
6147 * Exception: tcp_ca wants ECN. This is required for DCTCP
6148 * congestion control: Linux DCTCP asserts ECT on all packets,
6149 * including SYN, which is most optimal solution; however,
6150 * others, such as FreeBSD do not.
6152 static void tcp_ecn_create_request(struct request_sock
*req
,
6153 const struct sk_buff
*skb
,
6154 const struct sock
*listen_sk
,
6155 const struct dst_entry
*dst
)
6157 const struct tcphdr
*th
= tcp_hdr(skb
);
6158 const struct net
*net
= sock_net(listen_sk
);
6159 bool th_ecn
= th
->ece
&& th
->cwr
;
6166 ect
= !INET_ECN_is_not_ect(TCP_SKB_CB(skb
)->ip_dsfield
);
6167 ecn_ok_dst
= dst_feature(dst
, DST_FEATURE_ECN_MASK
);
6168 ecn_ok
= net
->ipv4
.sysctl_tcp_ecn
|| ecn_ok_dst
;
6170 if ((!ect
&& ecn_ok
) || tcp_ca_needs_ecn(listen_sk
) ||
6171 (ecn_ok_dst
& DST_FEATURE_ECN_CA
))
6172 inet_rsk(req
)->ecn_ok
= 1;
6175 static void tcp_openreq_init(struct request_sock
*req
,
6176 const struct tcp_options_received
*rx_opt
,
6177 struct sk_buff
*skb
, const struct sock
*sk
)
6179 struct inet_request_sock
*ireq
= inet_rsk(req
);
6181 req
->rsk_rcv_wnd
= 0; /* So that tcp_send_synack() knows! */
6183 tcp_rsk(req
)->rcv_isn
= TCP_SKB_CB(skb
)->seq
;
6184 tcp_rsk(req
)->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
6185 skb_mstamp_get(&tcp_rsk(req
)->snt_synack
);
6186 tcp_rsk(req
)->last_oow_ack_time
= 0;
6187 req
->mss
= rx_opt
->mss_clamp
;
6188 req
->ts_recent
= rx_opt
->saw_tstamp
? rx_opt
->rcv_tsval
: 0;
6189 ireq
->tstamp_ok
= rx_opt
->tstamp_ok
;
6190 ireq
->sack_ok
= rx_opt
->sack_ok
;
6191 ireq
->snd_wscale
= rx_opt
->snd_wscale
;
6192 ireq
->wscale_ok
= rx_opt
->wscale_ok
;
6195 ireq
->ir_rmt_port
= tcp_hdr(skb
)->source
;
6196 ireq
->ir_num
= ntohs(tcp_hdr(skb
)->dest
);
6197 ireq
->ir_mark
= inet_request_mark(sk
, skb
);
6200 struct request_sock
*inet_reqsk_alloc(const struct request_sock_ops
*ops
,
6201 struct sock
*sk_listener
,
6202 bool attach_listener
)
6204 struct request_sock
*req
= reqsk_alloc(ops
, sk_listener
,
6208 struct inet_request_sock
*ireq
= inet_rsk(req
);
6210 kmemcheck_annotate_bitfield(ireq
, flags
);
6212 #if IS_ENABLED(CONFIG_IPV6)
6213 ireq
->pktopts
= NULL
;
6215 atomic64_set(&ireq
->ir_cookie
, 0);
6216 ireq
->ireq_state
= TCP_NEW_SYN_RECV
;
6217 write_pnet(&ireq
->ireq_net
, sock_net(sk_listener
));
6218 ireq
->ireq_family
= sk_listener
->sk_family
;
6223 EXPORT_SYMBOL(inet_reqsk_alloc
);
6226 * Return true if a syncookie should be sent
6228 static bool tcp_syn_flood_action(const struct sock
*sk
,
6229 const struct sk_buff
*skb
,
6232 struct request_sock_queue
*queue
= &inet_csk(sk
)->icsk_accept_queue
;
6233 const char *msg
= "Dropping request";
6234 bool want_cookie
= false;
6235 struct net
*net
= sock_net(sk
);
6237 #ifdef CONFIG_SYN_COOKIES
6238 if (net
->ipv4
.sysctl_tcp_syncookies
) {
6239 msg
= "Sending cookies";
6241 __NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPREQQFULLDOCOOKIES
);
6244 __NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPREQQFULLDROP
);
6246 if (!queue
->synflood_warned
&&
6247 net
->ipv4
.sysctl_tcp_syncookies
!= 2 &&
6248 xchg(&queue
->synflood_warned
, 1) == 0)
6249 pr_info("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6250 proto
, ntohs(tcp_hdr(skb
)->dest
), msg
);
6255 static void tcp_reqsk_record_syn(const struct sock
*sk
,
6256 struct request_sock
*req
,
6257 const struct sk_buff
*skb
)
6259 if (tcp_sk(sk
)->save_syn
) {
6260 u32 len
= skb_network_header_len(skb
) + tcp_hdrlen(skb
);
6263 copy
= kmalloc(len
+ sizeof(u32
), GFP_ATOMIC
);
6266 memcpy(©
[1], skb_network_header(skb
), len
);
6267 req
->saved_syn
= copy
;
6272 int tcp_conn_request(struct request_sock_ops
*rsk_ops
,
6273 const struct tcp_request_sock_ops
*af_ops
,
6274 struct sock
*sk
, struct sk_buff
*skb
)
6276 struct tcp_fastopen_cookie foc
= { .len
= -1 };
6277 __u32 isn
= TCP_SKB_CB(skb
)->tcp_tw_isn
;
6278 struct tcp_options_received tmp_opt
;
6279 struct tcp_sock
*tp
= tcp_sk(sk
);
6280 struct net
*net
= sock_net(sk
);
6281 struct sock
*fastopen_sk
= NULL
;
6282 struct dst_entry
*dst
= NULL
;
6283 struct request_sock
*req
;
6284 bool want_cookie
= false;
6287 /* TW buckets are converted to open requests without
6288 * limitations, they conserve resources and peer is
6289 * evidently real one.
6291 if ((net
->ipv4
.sysctl_tcp_syncookies
== 2 ||
6292 inet_csk_reqsk_queue_is_full(sk
)) && !isn
) {
6293 want_cookie
= tcp_syn_flood_action(sk
, skb
, rsk_ops
->slab_name
);
6298 if (sk_acceptq_is_full(sk
)) {
6299 NET_INC_STATS(sock_net(sk
), LINUX_MIB_LISTENOVERFLOWS
);
6303 req
= inet_reqsk_alloc(rsk_ops
, sk
, !want_cookie
);
6307 tcp_rsk(req
)->af_specific
= af_ops
;
6308 tcp_rsk(req
)->ts_off
= 0;
6310 tcp_clear_options(&tmp_opt
);
6311 tmp_opt
.mss_clamp
= af_ops
->mss_clamp
;
6312 tmp_opt
.user_mss
= tp
->rx_opt
.user_mss
;
6313 tcp_parse_options(skb
, &tmp_opt
, 0, want_cookie
? NULL
: &foc
);
6315 if (want_cookie
&& !tmp_opt
.saw_tstamp
)
6316 tcp_clear_options(&tmp_opt
);
6318 tmp_opt
.tstamp_ok
= tmp_opt
.saw_tstamp
;
6319 tcp_openreq_init(req
, &tmp_opt
, skb
, sk
);
6320 inet_rsk(req
)->no_srccheck
= inet_sk(sk
)->transparent
;
6322 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6323 inet_rsk(req
)->ir_iif
= inet_request_bound_dev_if(sk
, skb
);
6325 af_ops
->init_req(req
, sk
, skb
);
6327 if (security_inet_conn_request(sk
, skb
, req
))
6330 if (isn
&& tmp_opt
.tstamp_ok
)
6331 af_ops
->init_seq(skb
, &tcp_rsk(req
)->ts_off
);
6333 if (!want_cookie
&& !isn
) {
6334 /* VJ's idea. We save last timestamp seen
6335 * from the destination in peer table, when entering
6336 * state TIME-WAIT, and check against it before
6337 * accepting new connection request.
6339 * If "isn" is not zero, this request hit alive
6340 * timewait bucket, so that all the necessary checks
6341 * are made in the function processing timewait state.
6343 if (net
->ipv4
.tcp_death_row
.sysctl_tw_recycle
) {
6346 dst
= af_ops
->route_req(sk
, &fl
, req
, &strict
);
6348 if (dst
&& strict
&&
6349 !tcp_peer_is_proven(req
, dst
, true,
6350 tmp_opt
.saw_tstamp
)) {
6351 NET_INC_STATS(sock_net(sk
), LINUX_MIB_PAWSPASSIVEREJECTED
);
6352 goto drop_and_release
;
6355 /* Kill the following clause, if you dislike this way. */
6356 else if (!net
->ipv4
.sysctl_tcp_syncookies
&&
6357 (net
->ipv4
.sysctl_max_syn_backlog
- inet_csk_reqsk_queue_len(sk
) <
6358 (net
->ipv4
.sysctl_max_syn_backlog
>> 2)) &&
6359 !tcp_peer_is_proven(req
, dst
, false,
6360 tmp_opt
.saw_tstamp
)) {
6361 /* Without syncookies last quarter of
6362 * backlog is filled with destinations,
6363 * proven to be alive.
6364 * It means that we continue to communicate
6365 * to destinations, already remembered
6366 * to the moment of synflood.
6368 pr_drop_req(req
, ntohs(tcp_hdr(skb
)->source
),
6370 goto drop_and_release
;
6373 isn
= af_ops
->init_seq(skb
, &tcp_rsk(req
)->ts_off
);
6376 dst
= af_ops
->route_req(sk
, &fl
, req
, NULL
);
6381 tcp_ecn_create_request(req
, skb
, sk
, dst
);
6384 isn
= cookie_init_sequence(af_ops
, sk
, skb
, &req
->mss
);
6385 tcp_rsk(req
)->ts_off
= 0;
6386 req
->cookie_ts
= tmp_opt
.tstamp_ok
;
6387 if (!tmp_opt
.tstamp_ok
)
6388 inet_rsk(req
)->ecn_ok
= 0;
6391 tcp_rsk(req
)->snt_isn
= isn
;
6392 tcp_rsk(req
)->txhash
= net_tx_rndhash();
6393 tcp_openreq_init_rwin(req
, sk
, dst
);
6395 tcp_reqsk_record_syn(sk
, req
, skb
);
6396 fastopen_sk
= tcp_try_fastopen(sk
, skb
, req
, &foc
, dst
);
6399 af_ops
->send_synack(fastopen_sk
, dst
, &fl
, req
,
6400 &foc
, TCP_SYNACK_FASTOPEN
);
6401 /* Add the child socket directly into the accept queue */
6402 inet_csk_reqsk_queue_add(sk
, req
, fastopen_sk
);
6403 sk
->sk_data_ready(sk
);
6404 bh_unlock_sock(fastopen_sk
);
6405 sock_put(fastopen_sk
);
6407 tcp_rsk(req
)->tfo_listener
= false;
6409 inet_csk_reqsk_queue_hash_add(sk
, req
, TCP_TIMEOUT_INIT
);
6410 af_ops
->send_synack(sk
, dst
, &fl
, req
, &foc
,
6411 !want_cookie
? TCP_SYNACK_NORMAL
:
6429 EXPORT_SYMBOL(tcp_conn_request
);