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/reciprocal_div.h>
74 #include <net/inet_common.h>
75 #include <linux/ipsec.h>
76 #include <asm/unaligned.h>
77 #include <net/netdma.h>
79 int sysctl_tcp_timestamps __read_mostly
= 1;
80 int sysctl_tcp_window_scaling __read_mostly
= 1;
81 int sysctl_tcp_sack __read_mostly
= 1;
82 int sysctl_tcp_fack __read_mostly
= 1;
83 int sysctl_tcp_reordering __read_mostly
= TCP_FASTRETRANS_THRESH
;
84 EXPORT_SYMBOL(sysctl_tcp_reordering
);
85 int sysctl_tcp_dsack __read_mostly
= 1;
86 int sysctl_tcp_app_win __read_mostly
= 31;
87 int sysctl_tcp_adv_win_scale __read_mostly
= 1;
88 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale
);
90 /* rfc5961 challenge ack rate limiting */
91 int sysctl_tcp_challenge_ack_limit
= 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;
98 int sysctl_tcp_thin_dupack __read_mostly
;
100 int sysctl_tcp_moderate_rcvbuf __read_mostly
= 1;
101 int sysctl_tcp_early_retrans __read_mostly
= 3;
102 int sysctl_tcp_default_init_rwnd __read_mostly
= TCP_DEFAULT_INIT_RCVWND
;
104 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
105 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
106 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
107 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
108 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
109 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
110 #define FLAG_ECE 0x40 /* ECE in this ACK */
111 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
112 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
113 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
114 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
115 #define FLAG_SET_XMIT_TIMER 0x1000 /* Set TLP or RTO timer */
116 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
117 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
119 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
120 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
121 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
122 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
124 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
125 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
127 /* Adapt the MSS value used to make delayed ack decision to the
130 static void tcp_measure_rcv_mss(struct sock
*sk
, const struct sk_buff
*skb
)
132 struct inet_connection_sock
*icsk
= inet_csk(sk
);
133 const unsigned int lss
= icsk
->icsk_ack
.last_seg_size
;
136 icsk
->icsk_ack
.last_seg_size
= 0;
138 /* skb->len may jitter because of SACKs, even if peer
139 * sends good full-sized frames.
141 len
= skb_shinfo(skb
)->gso_size
? : skb
->len
;
142 if (len
>= icsk
->icsk_ack
.rcv_mss
) {
143 icsk
->icsk_ack
.rcv_mss
= len
;
145 /* Otherwise, we make more careful check taking into account,
146 * that SACKs block is variable.
148 * "len" is invariant segment length, including TCP header.
150 len
+= skb
->data
- skb_transport_header(skb
);
151 if (len
>= TCP_MSS_DEFAULT
+ sizeof(struct tcphdr
) ||
152 /* If PSH is not set, packet should be
153 * full sized, provided peer TCP is not badly broken.
154 * This observation (if it is correct 8)) allows
155 * to handle super-low mtu links fairly.
157 (len
>= TCP_MIN_MSS
+ sizeof(struct tcphdr
) &&
158 !(tcp_flag_word(tcp_hdr(skb
)) & TCP_REMNANT
))) {
159 /* Subtract also invariant (if peer is RFC compliant),
160 * tcp header plus fixed timestamp option length.
161 * Resulting "len" is MSS free of SACK jitter.
163 len
-= tcp_sk(sk
)->tcp_header_len
;
164 icsk
->icsk_ack
.last_seg_size
= len
;
166 icsk
->icsk_ack
.rcv_mss
= len
;
170 if (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
)
171 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED2
;
172 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
176 static void tcp_incr_quickack(struct sock
*sk
)
178 struct inet_connection_sock
*icsk
= inet_csk(sk
);
179 unsigned int quickacks
= tcp_sk(sk
)->rcv_wnd
/ (2 * icsk
->icsk_ack
.rcv_mss
);
183 if (quickacks
> icsk
->icsk_ack
.quick
)
184 icsk
->icsk_ack
.quick
= min(quickacks
, TCP_MAX_QUICKACKS
);
187 static void tcp_enter_quickack_mode(struct sock
*sk
)
189 struct inet_connection_sock
*icsk
= inet_csk(sk
);
190 tcp_incr_quickack(sk
);
191 icsk
->icsk_ack
.pingpong
= 0;
192 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
195 /* Send ACKs quickly, if "quick" count is not exhausted
196 * and the session is not interactive.
199 static inline bool tcp_in_quickack_mode(const struct sock
*sk
)
201 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
203 return icsk
->icsk_ack
.quick
&& !icsk
->icsk_ack
.pingpong
;
206 static inline void TCP_ECN_queue_cwr(struct tcp_sock
*tp
)
208 if (tp
->ecn_flags
& TCP_ECN_OK
)
209 tp
->ecn_flags
|= TCP_ECN_QUEUE_CWR
;
212 static inline void TCP_ECN_accept_cwr(struct tcp_sock
*tp
, const struct sk_buff
*skb
)
214 if (tcp_hdr(skb
)->cwr
)
215 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
218 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock
*tp
)
220 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
223 static inline void TCP_ECN_check_ce(struct tcp_sock
*tp
, const struct sk_buff
*skb
)
225 if (!(tp
->ecn_flags
& TCP_ECN_OK
))
228 switch (TCP_SKB_CB(skb
)->ip_dsfield
& INET_ECN_MASK
) {
229 case INET_ECN_NOT_ECT
:
230 /* Funny extension: if ECT is not set on a segment,
231 * and we already seen ECT on a previous segment,
232 * it is probably a retransmit.
234 if (tp
->ecn_flags
& TCP_ECN_SEEN
)
235 tcp_enter_quickack_mode((struct sock
*)tp
);
238 if (!(tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)) {
239 /* Better not delay acks, sender can have a very low cwnd */
240 tcp_enter_quickack_mode((struct sock
*)tp
);
241 tp
->ecn_flags
|= TCP_ECN_DEMAND_CWR
;
245 tp
->ecn_flags
|= TCP_ECN_SEEN
;
249 static inline void TCP_ECN_rcv_synack(struct tcp_sock
*tp
, const struct tcphdr
*th
)
251 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| th
->cwr
))
252 tp
->ecn_flags
&= ~TCP_ECN_OK
;
255 static inline void TCP_ECN_rcv_syn(struct tcp_sock
*tp
, const struct tcphdr
*th
)
257 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| !th
->cwr
))
258 tp
->ecn_flags
&= ~TCP_ECN_OK
;
261 static bool TCP_ECN_rcv_ecn_echo(const struct tcp_sock
*tp
, const struct tcphdr
*th
)
263 if (th
->ece
&& !th
->syn
&& (tp
->ecn_flags
& TCP_ECN_OK
))
268 /* Buffer size and advertised window tuning.
270 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
273 static void tcp_fixup_sndbuf(struct sock
*sk
)
275 int sndmem
= SKB_TRUESIZE(tcp_sk(sk
)->rx_opt
.mss_clamp
+ MAX_TCP_HEADER
);
277 sndmem
*= TCP_INIT_CWND
;
278 if (sk
->sk_sndbuf
< sndmem
)
279 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
282 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
284 * All tcp_full_space() is split to two parts: "network" buffer, allocated
285 * forward and advertised in receiver window (tp->rcv_wnd) and
286 * "application buffer", required to isolate scheduling/application
287 * latencies from network.
288 * window_clamp is maximal advertised window. It can be less than
289 * tcp_full_space(), in this case tcp_full_space() - window_clamp
290 * is reserved for "application" buffer. The less window_clamp is
291 * the smoother our behaviour from viewpoint of network, but the lower
292 * throughput and the higher sensitivity of the connection to losses. 8)
294 * rcv_ssthresh is more strict window_clamp used at "slow start"
295 * phase to predict further behaviour of this connection.
296 * It is used for two goals:
297 * - to enforce header prediction at sender, even when application
298 * requires some significant "application buffer". It is check #1.
299 * - to prevent pruning of receive queue because of misprediction
300 * of receiver window. Check #2.
302 * The scheme does not work when sender sends good segments opening
303 * window and then starts to feed us spaghetti. But it should work
304 * in common situations. Otherwise, we have to rely on queue collapsing.
307 /* Slow part of check#2. */
308 static int __tcp_grow_window(const struct sock
*sk
, const struct sk_buff
*skb
)
310 struct tcp_sock
*tp
= tcp_sk(sk
);
312 int truesize
= tcp_win_from_space(skb
->truesize
) >> 1;
313 int window
= tcp_win_from_space(sysctl_tcp_rmem
[2]) >> 1;
315 while (tp
->rcv_ssthresh
<= window
) {
316 if (truesize
<= skb
->len
)
317 return 2 * inet_csk(sk
)->icsk_ack
.rcv_mss
;
325 static void tcp_grow_window(struct sock
*sk
, const struct sk_buff
*skb
)
327 struct tcp_sock
*tp
= tcp_sk(sk
);
330 if (tp
->rcv_ssthresh
< tp
->window_clamp
&&
331 (int)tp
->rcv_ssthresh
< tcp_space(sk
) &&
332 !sk_under_memory_pressure(sk
)) {
335 /* Check #2. Increase window, if skb with such overhead
336 * will fit to rcvbuf in future.
338 if (tcp_win_from_space(skb
->truesize
) <= skb
->len
)
339 incr
= 2 * tp
->advmss
;
341 incr
= __tcp_grow_window(sk
, skb
);
344 incr
= max_t(int, incr
, 2 * skb
->len
);
345 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
+ incr
,
347 inet_csk(sk
)->icsk_ack
.quick
|= 1;
352 /* 3. Tuning rcvbuf, when connection enters established state. */
354 static void tcp_fixup_rcvbuf(struct sock
*sk
)
356 u32 mss
= tcp_sk(sk
)->advmss
;
357 u32 icwnd
= sysctl_tcp_default_init_rwnd
;
360 /* Limit to 10 segments if mss <= 1460,
361 * or 14600/mss segments, with a minimum of two segments.
364 icwnd
= max_t(u32
, (1460 * icwnd
) / mss
, 2);
366 rcvmem
= SKB_TRUESIZE(mss
+ MAX_TCP_HEADER
);
367 while (tcp_win_from_space(rcvmem
) < mss
)
372 if (sk
->sk_rcvbuf
< rcvmem
)
373 sk
->sk_rcvbuf
= min(rcvmem
, sysctl_tcp_rmem
[2]);
376 /* 4. Try to fixup all. It is made immediately after connection enters
379 void tcp_init_buffer_space(struct sock
*sk
)
381 struct tcp_sock
*tp
= tcp_sk(sk
);
384 if (!(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
))
385 tcp_fixup_rcvbuf(sk
);
386 if (!(sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
))
387 tcp_fixup_sndbuf(sk
);
389 tp
->rcvq_space
.space
= tp
->rcv_wnd
;
391 maxwin
= tcp_full_space(sk
);
393 if (tp
->window_clamp
>= maxwin
) {
394 tp
->window_clamp
= maxwin
;
396 if (sysctl_tcp_app_win
&& maxwin
> 4 * tp
->advmss
)
397 tp
->window_clamp
= max(maxwin
-
398 (maxwin
>> sysctl_tcp_app_win
),
402 /* Force reservation of one segment. */
403 if (sysctl_tcp_app_win
&&
404 tp
->window_clamp
> 2 * tp
->advmss
&&
405 tp
->window_clamp
+ tp
->advmss
> maxwin
)
406 tp
->window_clamp
= max(2 * tp
->advmss
, maxwin
- tp
->advmss
);
408 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, tp
->window_clamp
);
409 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
412 /* 5. Recalculate window clamp after socket hit its memory bounds. */
413 static void tcp_clamp_window(struct sock
*sk
)
415 struct tcp_sock
*tp
= tcp_sk(sk
);
416 struct inet_connection_sock
*icsk
= inet_csk(sk
);
418 icsk
->icsk_ack
.quick
= 0;
420 if (sk
->sk_rcvbuf
< sysctl_tcp_rmem
[2] &&
421 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
) &&
422 !sk_under_memory_pressure(sk
) &&
423 sk_memory_allocated(sk
) < sk_prot_mem_limits(sk
, 0)) {
424 sk
->sk_rcvbuf
= min(atomic_read(&sk
->sk_rmem_alloc
),
427 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
)
428 tp
->rcv_ssthresh
= min(tp
->window_clamp
, 2U * tp
->advmss
);
431 /* Initialize RCV_MSS value.
432 * RCV_MSS is an our guess about MSS used by the peer.
433 * We haven't any direct information about the MSS.
434 * It's better to underestimate the RCV_MSS rather than overestimate.
435 * Overestimations make us ACKing less frequently than needed.
436 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
438 void tcp_initialize_rcv_mss(struct sock
*sk
)
440 const struct tcp_sock
*tp
= tcp_sk(sk
);
441 unsigned int hint
= min_t(unsigned int, tp
->advmss
, tp
->mss_cache
);
443 hint
= min(hint
, tp
->rcv_wnd
/ 2);
444 hint
= min(hint
, TCP_MSS_DEFAULT
);
445 hint
= max(hint
, TCP_MIN_MSS
);
447 inet_csk(sk
)->icsk_ack
.rcv_mss
= hint
;
449 EXPORT_SYMBOL(tcp_initialize_rcv_mss
);
451 /* Receiver "autotuning" code.
453 * The algorithm for RTT estimation w/o timestamps is based on
454 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
455 * <http://public.lanl.gov/radiant/pubs.html#DRS>
457 * More detail on this code can be found at
458 * <http://staff.psc.edu/jheffner/>,
459 * though this reference is out of date. A new paper
462 static void tcp_rcv_rtt_update(struct tcp_sock
*tp
, u32 sample
, int win_dep
)
464 u32 new_sample
= tp
->rcv_rtt_est
.rtt
;
470 if (new_sample
!= 0) {
471 /* If we sample in larger samples in the non-timestamp
472 * case, we could grossly overestimate the RTT especially
473 * with chatty applications or bulk transfer apps which
474 * are stalled on filesystem I/O.
476 * Also, since we are only going for a minimum in the
477 * non-timestamp case, we do not smooth things out
478 * else with timestamps disabled convergence takes too
482 m
-= (new_sample
>> 3);
490 /* No previous measure. */
494 if (tp
->rcv_rtt_est
.rtt
!= new_sample
)
495 tp
->rcv_rtt_est
.rtt
= new_sample
;
498 static inline void tcp_rcv_rtt_measure(struct tcp_sock
*tp
)
500 if (tp
->rcv_rtt_est
.time
== 0)
502 if (before(tp
->rcv_nxt
, tp
->rcv_rtt_est
.seq
))
504 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rcv_rtt_est
.time
, 1);
507 tp
->rcv_rtt_est
.seq
= tp
->rcv_nxt
+ tp
->rcv_wnd
;
508 tp
->rcv_rtt_est
.time
= tcp_time_stamp
;
511 static inline void tcp_rcv_rtt_measure_ts(struct sock
*sk
,
512 const struct sk_buff
*skb
)
514 struct tcp_sock
*tp
= tcp_sk(sk
);
515 if (tp
->rx_opt
.rcv_tsecr
&&
516 (TCP_SKB_CB(skb
)->end_seq
-
517 TCP_SKB_CB(skb
)->seq
>= inet_csk(sk
)->icsk_ack
.rcv_mss
))
518 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
, 0);
522 * This function should be called every time data is copied to user space.
523 * It calculates the appropriate TCP receive buffer space.
525 void tcp_rcv_space_adjust(struct sock
*sk
)
527 struct tcp_sock
*tp
= tcp_sk(sk
);
531 if (tp
->rcvq_space
.time
== 0)
534 time
= tcp_time_stamp
- tp
->rcvq_space
.time
;
535 if (time
< (tp
->rcv_rtt_est
.rtt
>> 3) || tp
->rcv_rtt_est
.rtt
== 0)
538 space
= 2 * (tp
->copied_seq
- tp
->rcvq_space
.seq
);
540 space
= max(tp
->rcvq_space
.space
, space
);
542 if (tp
->rcvq_space
.space
!= space
) {
545 tp
->rcvq_space
.space
= space
;
547 if (sysctl_tcp_moderate_rcvbuf
&&
548 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
)) {
549 int new_clamp
= space
;
551 /* Receive space grows, normalize in order to
552 * take into account packet headers and sk_buff
553 * structure overhead.
558 rcvmem
= SKB_TRUESIZE(tp
->advmss
+ MAX_TCP_HEADER
);
559 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
562 space
= min(space
, sysctl_tcp_rmem
[2]);
563 if (space
> sk
->sk_rcvbuf
) {
564 sk
->sk_rcvbuf
= space
;
566 /* Make the window clamp follow along. */
567 tp
->window_clamp
= new_clamp
;
573 tp
->rcvq_space
.seq
= tp
->copied_seq
;
574 tp
->rcvq_space
.time
= tcp_time_stamp
;
577 /* There is something which you must keep in mind when you analyze the
578 * behavior of the tp->ato delayed ack timeout interval. When a
579 * connection starts up, we want to ack as quickly as possible. The
580 * problem is that "good" TCP's do slow start at the beginning of data
581 * transmission. The means that until we send the first few ACK's the
582 * sender will sit on his end and only queue most of his data, because
583 * he can only send snd_cwnd unacked packets at any given time. For
584 * each ACK we send, he increments snd_cwnd and transmits more of his
587 static void tcp_event_data_recv(struct sock
*sk
, struct sk_buff
*skb
)
589 struct tcp_sock
*tp
= tcp_sk(sk
);
590 struct inet_connection_sock
*icsk
= inet_csk(sk
);
593 inet_csk_schedule_ack(sk
);
595 tcp_measure_rcv_mss(sk
, skb
);
597 tcp_rcv_rtt_measure(tp
);
599 now
= tcp_time_stamp
;
601 if (!icsk
->icsk_ack
.ato
) {
602 /* The _first_ data packet received, initialize
603 * delayed ACK engine.
605 tcp_incr_quickack(sk
);
606 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
608 int m
= now
- icsk
->icsk_ack
.lrcvtime
;
610 if (m
<= TCP_ATO_MIN
/ 2) {
611 /* The fastest case is the first. */
612 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + TCP_ATO_MIN
/ 2;
613 } else if (m
< icsk
->icsk_ack
.ato
) {
614 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + m
;
615 if (icsk
->icsk_ack
.ato
> icsk
->icsk_rto
)
616 icsk
->icsk_ack
.ato
= icsk
->icsk_rto
;
617 } else if (m
> icsk
->icsk_rto
) {
618 /* Too long gap. Apparently sender failed to
619 * restart window, so that we send ACKs quickly.
621 tcp_incr_quickack(sk
);
625 icsk
->icsk_ack
.lrcvtime
= now
;
627 TCP_ECN_check_ce(tp
, skb
);
630 tcp_grow_window(sk
, skb
);
633 /* Called to compute a smoothed rtt estimate. The data fed to this
634 * routine either comes from timestamps, or from segments that were
635 * known _not_ to have been retransmitted [see Karn/Partridge
636 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
637 * piece by Van Jacobson.
638 * NOTE: the next three routines used to be one big routine.
639 * To save cycles in the RFC 1323 implementation it was better to break
640 * it up into three procedures. -- erics
642 static void tcp_rtt_estimator(struct sock
*sk
, const __u32 mrtt
)
644 struct tcp_sock
*tp
= tcp_sk(sk
);
645 long m
= mrtt
; /* RTT */
647 /* The following amusing code comes from Jacobson's
648 * article in SIGCOMM '88. Note that rtt and mdev
649 * are scaled versions of rtt and mean deviation.
650 * This is designed to be as fast as possible
651 * m stands for "measurement".
653 * On a 1990 paper the rto value is changed to:
654 * RTO = rtt + 4 * mdev
656 * Funny. This algorithm seems to be very broken.
657 * These formulae increase RTO, when it should be decreased, increase
658 * too slowly, when it should be increased quickly, decrease too quickly
659 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
660 * does not matter how to _calculate_ it. Seems, it was trap
661 * that VJ failed to avoid. 8)
666 m
-= (tp
->srtt
>> 3); /* m is now error in rtt est */
667 tp
->srtt
+= m
; /* rtt = 7/8 rtt + 1/8 new */
669 m
= -m
; /* m is now abs(error) */
670 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
671 /* This is similar to one of Eifel findings.
672 * Eifel blocks mdev updates when rtt decreases.
673 * This solution is a bit different: we use finer gain
674 * for mdev in this case (alpha*beta).
675 * Like Eifel it also prevents growth of rto,
676 * but also it limits too fast rto decreases,
677 * happening in pure Eifel.
682 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
684 tp
->mdev
+= m
; /* mdev = 3/4 mdev + 1/4 new */
685 if (tp
->mdev
> tp
->mdev_max
) {
686 tp
->mdev_max
= tp
->mdev
;
687 if (tp
->mdev_max
> tp
->rttvar
)
688 tp
->rttvar
= tp
->mdev_max
;
690 if (after(tp
->snd_una
, tp
->rtt_seq
)) {
691 if (tp
->mdev_max
< tp
->rttvar
)
692 tp
->rttvar
-= (tp
->rttvar
- tp
->mdev_max
) >> 2;
693 tp
->rtt_seq
= tp
->snd_nxt
;
694 tp
->mdev_max
= tcp_rto_min(sk
);
697 /* no previous measure. */
698 tp
->srtt
= m
<< 3; /* take the measured time to be rtt */
699 tp
->mdev
= m
<< 1; /* make sure rto = 3*rtt */
700 tp
->mdev_max
= tp
->rttvar
= max(tp
->mdev
, tcp_rto_min(sk
));
701 tp
->rtt_seq
= tp
->snd_nxt
;
705 /* Set the sk_pacing_rate to allow proper sizing of TSO packets.
706 * Note: TCP stack does not yet implement pacing.
707 * FQ packet scheduler can be used to implement cheap but effective
708 * TCP pacing, to smooth the burst on large writes when packets
709 * in flight is significantly lower than cwnd (or rwin)
711 static void tcp_update_pacing_rate(struct sock
*sk
)
713 const struct tcp_sock
*tp
= tcp_sk(sk
);
716 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
717 rate
= (u64
)tp
->mss_cache
* 2 * (HZ
<< 3);
719 rate
*= max(tp
->snd_cwnd
, tp
->packets_out
);
721 /* Correction for small srtt : minimum srtt being 8 (1 jiffy << 3),
722 * be conservative and assume srtt = 1 (125 us instead of 1.25 ms)
723 * We probably need usec resolution in the future.
724 * Note: This also takes care of possible srtt=0 case,
725 * when tcp_rtt_estimator() was not yet called.
727 if (tp
->srtt
> 8 + 2)
728 do_div(rate
, tp
->srtt
);
730 sk
->sk_pacing_rate
= min_t(u64
, rate
, ~0U);
733 /* Calculate rto without backoff. This is the second half of Van Jacobson's
734 * routine referred to above.
736 void tcp_set_rto(struct sock
*sk
)
738 const struct tcp_sock
*tp
= tcp_sk(sk
);
739 /* Old crap is replaced with new one. 8)
742 * 1. If rtt variance happened to be less 50msec, it is hallucination.
743 * It cannot be less due to utterly erratic ACK generation made
744 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
745 * to do with delayed acks, because at cwnd>2 true delack timeout
746 * is invisible. Actually, Linux-2.4 also generates erratic
747 * ACKs in some circumstances.
749 inet_csk(sk
)->icsk_rto
= __tcp_set_rto(tp
);
751 /* 2. Fixups made earlier cannot be right.
752 * If we do not estimate RTO correctly without them,
753 * all the algo is pure shit and should be replaced
754 * with correct one. It is exactly, which we pretend to do.
757 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
758 * guarantees that rto is higher.
763 __u32
tcp_init_cwnd(const struct tcp_sock
*tp
, const struct dst_entry
*dst
)
765 __u32 cwnd
= (dst
? dst_metric(dst
, RTAX_INITCWND
) : 0);
768 cwnd
= TCP_INIT_CWND
;
769 return min_t(__u32
, cwnd
, tp
->snd_cwnd_clamp
);
773 * Packet counting of FACK is based on in-order assumptions, therefore TCP
774 * disables it when reordering is detected
776 void tcp_disable_fack(struct tcp_sock
*tp
)
778 /* RFC3517 uses different metric in lost marker => reset on change */
780 tp
->lost_skb_hint
= NULL
;
781 tp
->rx_opt
.sack_ok
&= ~TCP_FACK_ENABLED
;
784 /* Take a notice that peer is sending D-SACKs */
785 static void tcp_dsack_seen(struct tcp_sock
*tp
)
787 tp
->rx_opt
.sack_ok
|= TCP_DSACK_SEEN
;
790 static void tcp_update_reordering(struct sock
*sk
, const int metric
,
793 struct tcp_sock
*tp
= tcp_sk(sk
);
794 if (metric
> tp
->reordering
) {
797 tp
->reordering
= min(TCP_MAX_REORDERING
, metric
);
799 /* This exciting event is worth to be remembered. 8) */
801 mib_idx
= LINUX_MIB_TCPTSREORDER
;
802 else if (tcp_is_reno(tp
))
803 mib_idx
= LINUX_MIB_TCPRENOREORDER
;
804 else if (tcp_is_fack(tp
))
805 mib_idx
= LINUX_MIB_TCPFACKREORDER
;
807 mib_idx
= LINUX_MIB_TCPSACKREORDER
;
809 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
810 #if FASTRETRANS_DEBUG > 1
811 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
812 tp
->rx_opt
.sack_ok
, inet_csk(sk
)->icsk_ca_state
,
816 tp
->undo_marker
? tp
->undo_retrans
: 0);
818 tcp_disable_fack(tp
);
822 tcp_disable_early_retrans(tp
);
825 /* This must be called before lost_out is incremented */
826 static void tcp_verify_retransmit_hint(struct tcp_sock
*tp
, struct sk_buff
*skb
)
828 if ((tp
->retransmit_skb_hint
== NULL
) ||
829 before(TCP_SKB_CB(skb
)->seq
,
830 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
831 tp
->retransmit_skb_hint
= skb
;
834 after(TCP_SKB_CB(skb
)->end_seq
, tp
->retransmit_high
))
835 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
838 static void tcp_skb_mark_lost(struct tcp_sock
*tp
, struct sk_buff
*skb
)
840 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
841 tcp_verify_retransmit_hint(tp
, skb
);
843 tp
->lost_out
+= tcp_skb_pcount(skb
);
844 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
848 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock
*tp
,
851 tcp_verify_retransmit_hint(tp
, skb
);
853 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
854 tp
->lost_out
+= tcp_skb_pcount(skb
);
855 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
859 /* This procedure tags the retransmission queue when SACKs arrive.
861 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
862 * Packets in queue with these bits set are counted in variables
863 * sacked_out, retrans_out and lost_out, correspondingly.
865 * Valid combinations are:
866 * Tag InFlight Description
867 * 0 1 - orig segment is in flight.
868 * S 0 - nothing flies, orig reached receiver.
869 * L 0 - nothing flies, orig lost by net.
870 * R 2 - both orig and retransmit are in flight.
871 * L|R 1 - orig is lost, retransmit is in flight.
872 * S|R 1 - orig reached receiver, retrans is still in flight.
873 * (L|S|R is logically valid, it could occur when L|R is sacked,
874 * but it is equivalent to plain S and code short-curcuits it to S.
875 * L|S is logically invalid, it would mean -1 packet in flight 8))
877 * These 6 states form finite state machine, controlled by the following events:
878 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
879 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
880 * 3. Loss detection event of two flavors:
881 * A. Scoreboard estimator decided the packet is lost.
882 * A'. Reno "three dupacks" marks head of queue lost.
883 * A''. Its FACK modification, head until snd.fack is lost.
884 * B. SACK arrives sacking SND.NXT at the moment, when the
885 * segment was retransmitted.
886 * 4. D-SACK added new rule: D-SACK changes any tag to S.
888 * It is pleasant to note, that state diagram turns out to be commutative,
889 * so that we are allowed not to be bothered by order of our actions,
890 * when multiple events arrive simultaneously. (see the function below).
892 * Reordering detection.
893 * --------------------
894 * Reordering metric is maximal distance, which a packet can be displaced
895 * in packet stream. With SACKs we can estimate it:
897 * 1. SACK fills old hole and the corresponding segment was not
898 * ever retransmitted -> reordering. Alas, we cannot use it
899 * when segment was retransmitted.
900 * 2. The last flaw is solved with D-SACK. D-SACK arrives
901 * for retransmitted and already SACKed segment -> reordering..
902 * Both of these heuristics are not used in Loss state, when we cannot
903 * account for retransmits accurately.
905 * SACK block validation.
906 * ----------------------
908 * SACK block range validation checks that the received SACK block fits to
909 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
910 * Note that SND.UNA is not included to the range though being valid because
911 * it means that the receiver is rather inconsistent with itself reporting
912 * SACK reneging when it should advance SND.UNA. Such SACK block this is
913 * perfectly valid, however, in light of RFC2018 which explicitly states
914 * that "SACK block MUST reflect the newest segment. Even if the newest
915 * segment is going to be discarded ...", not that it looks very clever
916 * in case of head skb. Due to potentional receiver driven attacks, we
917 * choose to avoid immediate execution of a walk in write queue due to
918 * reneging and defer head skb's loss recovery to standard loss recovery
919 * procedure that will eventually trigger (nothing forbids us doing this).
921 * Implements also blockage to start_seq wrap-around. Problem lies in the
922 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
923 * there's no guarantee that it will be before snd_nxt (n). The problem
924 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
927 * <- outs wnd -> <- wrapzone ->
928 * u e n u_w e_w s n_w
930 * |<------------+------+----- TCP seqno space --------------+---------->|
931 * ...-- <2^31 ->| |<--------...
932 * ...---- >2^31 ------>| |<--------...
934 * Current code wouldn't be vulnerable but it's better still to discard such
935 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
936 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
937 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
938 * equal to the ideal case (infinite seqno space without wrap caused issues).
940 * With D-SACK the lower bound is extended to cover sequence space below
941 * SND.UNA down to undo_marker, which is the last point of interest. Yet
942 * again, D-SACK block must not to go across snd_una (for the same reason as
943 * for the normal SACK blocks, explained above). But there all simplicity
944 * ends, TCP might receive valid D-SACKs below that. As long as they reside
945 * fully below undo_marker they do not affect behavior in anyway and can
946 * therefore be safely ignored. In rare cases (which are more or less
947 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
948 * fragmentation and packet reordering past skb's retransmission. To consider
949 * them correctly, the acceptable range must be extended even more though
950 * the exact amount is rather hard to quantify. However, tp->max_window can
951 * be used as an exaggerated estimate.
953 static bool tcp_is_sackblock_valid(struct tcp_sock
*tp
, bool is_dsack
,
954 u32 start_seq
, u32 end_seq
)
956 /* Too far in future, or reversed (interpretation is ambiguous) */
957 if (after(end_seq
, tp
->snd_nxt
) || !before(start_seq
, end_seq
))
960 /* Nasty start_seq wrap-around check (see comments above) */
961 if (!before(start_seq
, tp
->snd_nxt
))
964 /* In outstanding window? ...This is valid exit for D-SACKs too.
965 * start_seq == snd_una is non-sensical (see comments above)
967 if (after(start_seq
, tp
->snd_una
))
970 if (!is_dsack
|| !tp
->undo_marker
)
973 /* ...Then it's D-SACK, and must reside below snd_una completely */
974 if (after(end_seq
, tp
->snd_una
))
977 if (!before(start_seq
, tp
->undo_marker
))
981 if (!after(end_seq
, tp
->undo_marker
))
984 /* Undo_marker boundary crossing (overestimates a lot). Known already:
985 * start_seq < undo_marker and end_seq >= undo_marker.
987 return !before(start_seq
, end_seq
- tp
->max_window
);
990 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
991 * Event "B". Later note: FACK people cheated me again 8), we have to account
992 * for reordering! Ugly, but should help.
994 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
995 * less than what is now known to be received by the other end (derived from
996 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
997 * retransmitted skbs to avoid some costly processing per ACKs.
999 static void tcp_mark_lost_retrans(struct sock
*sk
)
1001 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1002 struct tcp_sock
*tp
= tcp_sk(sk
);
1003 struct sk_buff
*skb
;
1005 u32 new_low_seq
= tp
->snd_nxt
;
1006 u32 received_upto
= tcp_highest_sack_seq(tp
);
1008 if (!tcp_is_fack(tp
) || !tp
->retrans_out
||
1009 !after(received_upto
, tp
->lost_retrans_low
) ||
1010 icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
1013 tcp_for_write_queue(skb
, sk
) {
1014 u32 ack_seq
= TCP_SKB_CB(skb
)->ack_seq
;
1016 if (skb
== tcp_send_head(sk
))
1018 if (cnt
== tp
->retrans_out
)
1020 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1023 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
))
1026 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1027 * constraint here (see above) but figuring out that at
1028 * least tp->reordering SACK blocks reside between ack_seq
1029 * and received_upto is not easy task to do cheaply with
1030 * the available datastructures.
1032 * Whether FACK should check here for tp->reordering segs
1033 * in-between one could argue for either way (it would be
1034 * rather simple to implement as we could count fack_count
1035 * during the walk and do tp->fackets_out - fack_count).
1037 if (after(received_upto
, ack_seq
)) {
1038 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1039 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1041 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
1042 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSTRETRANSMIT
);
1044 if (before(ack_seq
, new_low_seq
))
1045 new_low_seq
= ack_seq
;
1046 cnt
+= tcp_skb_pcount(skb
);
1050 if (tp
->retrans_out
)
1051 tp
->lost_retrans_low
= new_low_seq
;
1054 static bool tcp_check_dsack(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1055 struct tcp_sack_block_wire
*sp
, int num_sacks
,
1058 struct tcp_sock
*tp
= tcp_sk(sk
);
1059 u32 start_seq_0
= get_unaligned_be32(&sp
[0].start_seq
);
1060 u32 end_seq_0
= get_unaligned_be32(&sp
[0].end_seq
);
1061 bool dup_sack
= false;
1063 if (before(start_seq_0
, TCP_SKB_CB(ack_skb
)->ack_seq
)) {
1066 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPDSACKRECV
);
1067 } else if (num_sacks
> 1) {
1068 u32 end_seq_1
= get_unaligned_be32(&sp
[1].end_seq
);
1069 u32 start_seq_1
= get_unaligned_be32(&sp
[1].start_seq
);
1071 if (!after(end_seq_0
, end_seq_1
) &&
1072 !before(start_seq_0
, start_seq_1
)) {
1075 NET_INC_STATS_BH(sock_net(sk
),
1076 LINUX_MIB_TCPDSACKOFORECV
);
1080 /* D-SACK for already forgotten data... Do dumb counting. */
1081 if (dup_sack
&& tp
->undo_marker
&& tp
->undo_retrans
> 0 &&
1082 !after(end_seq_0
, prior_snd_una
) &&
1083 after(end_seq_0
, tp
->undo_marker
))
1089 struct tcp_sacktag_state
{
1095 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1096 * the incoming SACK may not exactly match but we can find smaller MSS
1097 * aligned portion of it that matches. Therefore we might need to fragment
1098 * which may fail and creates some hassle (caller must handle error case
1101 * FIXME: this could be merged to shift decision code
1103 static int tcp_match_skb_to_sack(struct sock
*sk
, struct sk_buff
*skb
,
1104 u32 start_seq
, u32 end_seq
)
1108 unsigned int pkt_len
;
1111 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1112 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1114 if (tcp_skb_pcount(skb
) > 1 && !in_sack
&&
1115 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
1116 mss
= tcp_skb_mss(skb
);
1117 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1120 pkt_len
= start_seq
- TCP_SKB_CB(skb
)->seq
;
1124 pkt_len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1129 /* Round if necessary so that SACKs cover only full MSSes
1130 * and/or the remaining small portion (if present)
1132 if (pkt_len
> mss
) {
1133 unsigned int new_len
= (pkt_len
/ mss
) * mss
;
1134 if (!in_sack
&& new_len
< pkt_len
) {
1136 if (new_len
>= skb
->len
)
1141 err
= tcp_fragment(sk
, skb
, pkt_len
, mss
);
1149 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1150 static u8
tcp_sacktag_one(struct sock
*sk
,
1151 struct tcp_sacktag_state
*state
, u8 sacked
,
1152 u32 start_seq
, u32 end_seq
,
1153 bool dup_sack
, int pcount
)
1155 struct tcp_sock
*tp
= tcp_sk(sk
);
1156 int fack_count
= state
->fack_count
;
1158 /* Account D-SACK for retransmitted packet. */
1159 if (dup_sack
&& (sacked
& TCPCB_RETRANS
)) {
1160 if (tp
->undo_marker
&& tp
->undo_retrans
> 0 &&
1161 after(end_seq
, tp
->undo_marker
))
1163 if (sacked
& TCPCB_SACKED_ACKED
)
1164 state
->reord
= min(fack_count
, state
->reord
);
1167 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1168 if (!after(end_seq
, tp
->snd_una
))
1171 if (!(sacked
& TCPCB_SACKED_ACKED
)) {
1172 if (sacked
& TCPCB_SACKED_RETRANS
) {
1173 /* If the segment is not tagged as lost,
1174 * we do not clear RETRANS, believing
1175 * that retransmission is still in flight.
1177 if (sacked
& TCPCB_LOST
) {
1178 sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1179 tp
->lost_out
-= pcount
;
1180 tp
->retrans_out
-= pcount
;
1183 if (!(sacked
& TCPCB_RETRANS
)) {
1184 /* New sack for not retransmitted frame,
1185 * which was in hole. It is reordering.
1187 if (before(start_seq
,
1188 tcp_highest_sack_seq(tp
)))
1189 state
->reord
= min(fack_count
,
1191 if (!after(end_seq
, tp
->high_seq
))
1192 state
->flag
|= FLAG_ORIG_SACK_ACKED
;
1195 if (sacked
& TCPCB_LOST
) {
1196 sacked
&= ~TCPCB_LOST
;
1197 tp
->lost_out
-= pcount
;
1201 sacked
|= TCPCB_SACKED_ACKED
;
1202 state
->flag
|= FLAG_DATA_SACKED
;
1203 tp
->sacked_out
+= pcount
;
1205 fack_count
+= pcount
;
1207 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1208 if (!tcp_is_fack(tp
) && (tp
->lost_skb_hint
!= NULL
) &&
1209 before(start_seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
))
1210 tp
->lost_cnt_hint
+= pcount
;
1212 if (fack_count
> tp
->fackets_out
)
1213 tp
->fackets_out
= fack_count
;
1216 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1217 * frames and clear it. undo_retrans is decreased above, L|R frames
1218 * are accounted above as well.
1220 if (dup_sack
&& (sacked
& TCPCB_SACKED_RETRANS
)) {
1221 sacked
&= ~TCPCB_SACKED_RETRANS
;
1222 tp
->retrans_out
-= pcount
;
1228 /* Shift newly-SACKed bytes from this skb to the immediately previous
1229 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1231 static bool tcp_shifted_skb(struct sock
*sk
, struct sk_buff
*skb
,
1232 struct tcp_sacktag_state
*state
,
1233 unsigned int pcount
, int shifted
, int mss
,
1236 struct tcp_sock
*tp
= tcp_sk(sk
);
1237 struct sk_buff
*prev
= tcp_write_queue_prev(sk
, skb
);
1238 u32 start_seq
= TCP_SKB_CB(skb
)->seq
; /* start of newly-SACKed */
1239 u32 end_seq
= start_seq
+ shifted
; /* end of newly-SACKed */
1243 /* Adjust counters and hints for the newly sacked sequence
1244 * range but discard the return value since prev is already
1245 * marked. We must tag the range first because the seq
1246 * advancement below implicitly advances
1247 * tcp_highest_sack_seq() when skb is highest_sack.
1249 tcp_sacktag_one(sk
, state
, TCP_SKB_CB(skb
)->sacked
,
1250 start_seq
, end_seq
, dup_sack
, pcount
);
1252 if (skb
== tp
->lost_skb_hint
)
1253 tp
->lost_cnt_hint
+= pcount
;
1255 TCP_SKB_CB(prev
)->end_seq
+= shifted
;
1256 TCP_SKB_CB(skb
)->seq
+= shifted
;
1258 skb_shinfo(prev
)->gso_segs
+= pcount
;
1259 BUG_ON(skb_shinfo(skb
)->gso_segs
< pcount
);
1260 skb_shinfo(skb
)->gso_segs
-= pcount
;
1262 /* When we're adding to gso_segs == 1, gso_size will be zero,
1263 * in theory this shouldn't be necessary but as long as DSACK
1264 * code can come after this skb later on it's better to keep
1265 * setting gso_size to something.
1267 if (!skb_shinfo(prev
)->gso_size
) {
1268 skb_shinfo(prev
)->gso_size
= mss
;
1269 skb_shinfo(prev
)->gso_type
= sk
->sk_gso_type
;
1272 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1273 if (skb_shinfo(skb
)->gso_segs
<= 1) {
1274 skb_shinfo(skb
)->gso_size
= 0;
1275 skb_shinfo(skb
)->gso_type
= 0;
1278 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1279 TCP_SKB_CB(prev
)->sacked
|= (TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
);
1282 BUG_ON(!tcp_skb_pcount(skb
));
1283 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKSHIFTED
);
1287 /* Whole SKB was eaten :-) */
1289 if (skb
== tp
->retransmit_skb_hint
)
1290 tp
->retransmit_skb_hint
= prev
;
1291 if (skb
== tp
->scoreboard_skb_hint
)
1292 tp
->scoreboard_skb_hint
= prev
;
1293 if (skb
== tp
->lost_skb_hint
) {
1294 tp
->lost_skb_hint
= prev
;
1295 tp
->lost_cnt_hint
-= tcp_skb_pcount(prev
);
1298 TCP_SKB_CB(prev
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
;
1299 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
1300 TCP_SKB_CB(prev
)->end_seq
++;
1302 if (skb
== tcp_highest_sack(sk
))
1303 tcp_advance_highest_sack(sk
, skb
);
1305 tcp_unlink_write_queue(skb
, sk
);
1306 sk_wmem_free_skb(sk
, skb
);
1308 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKMERGED
);
1313 /* I wish gso_size would have a bit more sane initialization than
1314 * something-or-zero which complicates things
1316 static int tcp_skb_seglen(const struct sk_buff
*skb
)
1318 return tcp_skb_pcount(skb
) == 1 ? skb
->len
: tcp_skb_mss(skb
);
1321 /* Shifting pages past head area doesn't work */
1322 static int skb_can_shift(const struct sk_buff
*skb
)
1324 return !skb_headlen(skb
) && skb_is_nonlinear(skb
);
1327 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1330 static struct sk_buff
*tcp_shift_skb_data(struct sock
*sk
, struct sk_buff
*skb
,
1331 struct tcp_sacktag_state
*state
,
1332 u32 start_seq
, u32 end_seq
,
1335 struct tcp_sock
*tp
= tcp_sk(sk
);
1336 struct sk_buff
*prev
;
1342 if (!sk_can_gso(sk
))
1345 /* Normally R but no L won't result in plain S */
1347 (TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_RETRANS
)) == TCPCB_SACKED_RETRANS
)
1349 if (!skb_can_shift(skb
))
1351 /* This frame is about to be dropped (was ACKed). */
1352 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1355 /* Can only happen with delayed DSACK + discard craziness */
1356 if (unlikely(skb
== tcp_write_queue_head(sk
)))
1358 prev
= tcp_write_queue_prev(sk
, skb
);
1360 if ((TCP_SKB_CB(prev
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
)
1363 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1364 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1368 pcount
= tcp_skb_pcount(skb
);
1369 mss
= tcp_skb_seglen(skb
);
1371 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1372 * drop this restriction as unnecessary
1374 if (mss
!= tcp_skb_seglen(prev
))
1377 if (!after(TCP_SKB_CB(skb
)->end_seq
, start_seq
))
1379 /* CHECKME: This is non-MSS split case only?, this will
1380 * cause skipped skbs due to advancing loop btw, original
1381 * has that feature too
1383 if (tcp_skb_pcount(skb
) <= 1)
1386 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1388 /* TODO: head merge to next could be attempted here
1389 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1390 * though it might not be worth of the additional hassle
1392 * ...we can probably just fallback to what was done
1393 * previously. We could try merging non-SACKed ones
1394 * as well but it probably isn't going to buy off
1395 * because later SACKs might again split them, and
1396 * it would make skb timestamp tracking considerably
1402 len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1404 BUG_ON(len
> skb
->len
);
1406 /* MSS boundaries should be honoured or else pcount will
1407 * severely break even though it makes things bit trickier.
1408 * Optimize common case to avoid most of the divides
1410 mss
= tcp_skb_mss(skb
);
1412 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1413 * drop this restriction as unnecessary
1415 if (mss
!= tcp_skb_seglen(prev
))
1420 } else if (len
< mss
) {
1428 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1429 if (!after(TCP_SKB_CB(skb
)->seq
+ len
, tp
->snd_una
))
1432 if (!skb_shift(prev
, skb
, len
))
1434 if (!tcp_shifted_skb(sk
, skb
, state
, pcount
, len
, mss
, dup_sack
))
1437 /* Hole filled allows collapsing with the next as well, this is very
1438 * useful when hole on every nth skb pattern happens
1440 if (prev
== tcp_write_queue_tail(sk
))
1442 skb
= tcp_write_queue_next(sk
, prev
);
1444 if (!skb_can_shift(skb
) ||
1445 (skb
== tcp_send_head(sk
)) ||
1446 ((TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
) ||
1447 (mss
!= tcp_skb_seglen(skb
)))
1451 if (skb_shift(prev
, skb
, len
)) {
1452 pcount
+= tcp_skb_pcount(skb
);
1453 tcp_shifted_skb(sk
, skb
, state
, tcp_skb_pcount(skb
), len
, mss
, 0);
1457 state
->fack_count
+= pcount
;
1464 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKSHIFTFALLBACK
);
1468 static struct sk_buff
*tcp_sacktag_walk(struct sk_buff
*skb
, struct sock
*sk
,
1469 struct tcp_sack_block
*next_dup
,
1470 struct tcp_sacktag_state
*state
,
1471 u32 start_seq
, u32 end_seq
,
1474 struct tcp_sock
*tp
= tcp_sk(sk
);
1475 struct sk_buff
*tmp
;
1477 tcp_for_write_queue_from(skb
, sk
) {
1479 bool dup_sack
= dup_sack_in
;
1481 if (skb
== tcp_send_head(sk
))
1484 /* queue is in-order => we can short-circuit the walk early */
1485 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1488 if ((next_dup
!= NULL
) &&
1489 before(TCP_SKB_CB(skb
)->seq
, next_dup
->end_seq
)) {
1490 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1491 next_dup
->start_seq
,
1497 /* skb reference here is a bit tricky to get right, since
1498 * shifting can eat and free both this skb and the next,
1499 * so not even _safe variant of the loop is enough.
1502 tmp
= tcp_shift_skb_data(sk
, skb
, state
,
1503 start_seq
, end_seq
, dup_sack
);
1512 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1518 if (unlikely(in_sack
< 0))
1522 TCP_SKB_CB(skb
)->sacked
=
1525 TCP_SKB_CB(skb
)->sacked
,
1526 TCP_SKB_CB(skb
)->seq
,
1527 TCP_SKB_CB(skb
)->end_seq
,
1529 tcp_skb_pcount(skb
));
1531 if (!before(TCP_SKB_CB(skb
)->seq
,
1532 tcp_highest_sack_seq(tp
)))
1533 tcp_advance_highest_sack(sk
, skb
);
1536 state
->fack_count
+= tcp_skb_pcount(skb
);
1541 /* Avoid all extra work that is being done by sacktag while walking in
1544 static struct sk_buff
*tcp_sacktag_skip(struct sk_buff
*skb
, struct sock
*sk
,
1545 struct tcp_sacktag_state
*state
,
1548 tcp_for_write_queue_from(skb
, sk
) {
1549 if (skb
== tcp_send_head(sk
))
1552 if (after(TCP_SKB_CB(skb
)->end_seq
, skip_to_seq
))
1555 state
->fack_count
+= tcp_skb_pcount(skb
);
1560 static struct sk_buff
*tcp_maybe_skipping_dsack(struct sk_buff
*skb
,
1562 struct tcp_sack_block
*next_dup
,
1563 struct tcp_sacktag_state
*state
,
1566 if (next_dup
== NULL
)
1569 if (before(next_dup
->start_seq
, skip_to_seq
)) {
1570 skb
= tcp_sacktag_skip(skb
, sk
, state
, next_dup
->start_seq
);
1571 skb
= tcp_sacktag_walk(skb
, sk
, NULL
, state
,
1572 next_dup
->start_seq
, next_dup
->end_seq
,
1579 static int tcp_sack_cache_ok(const struct tcp_sock
*tp
, const struct tcp_sack_block
*cache
)
1581 return cache
< tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1585 tcp_sacktag_write_queue(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1588 struct tcp_sock
*tp
= tcp_sk(sk
);
1589 const unsigned char *ptr
= (skb_transport_header(ack_skb
) +
1590 TCP_SKB_CB(ack_skb
)->sacked
);
1591 struct tcp_sack_block_wire
*sp_wire
= (struct tcp_sack_block_wire
*)(ptr
+2);
1592 struct tcp_sack_block sp
[TCP_NUM_SACKS
];
1593 struct tcp_sack_block
*cache
;
1594 struct tcp_sacktag_state state
;
1595 struct sk_buff
*skb
;
1596 int num_sacks
= min(TCP_NUM_SACKS
, (ptr
[1] - TCPOLEN_SACK_BASE
) >> 3);
1598 bool found_dup_sack
= false;
1600 int first_sack_index
;
1603 state
.reord
= tp
->packets_out
;
1605 if (!tp
->sacked_out
) {
1606 if (WARN_ON(tp
->fackets_out
))
1607 tp
->fackets_out
= 0;
1608 tcp_highest_sack_reset(sk
);
1611 found_dup_sack
= tcp_check_dsack(sk
, ack_skb
, sp_wire
,
1612 num_sacks
, prior_snd_una
);
1614 state
.flag
|= FLAG_DSACKING_ACK
;
1616 /* Eliminate too old ACKs, but take into
1617 * account more or less fresh ones, they can
1618 * contain valid SACK info.
1620 if (before(TCP_SKB_CB(ack_skb
)->ack_seq
, prior_snd_una
- tp
->max_window
))
1623 if (!tp
->packets_out
)
1627 first_sack_index
= 0;
1628 for (i
= 0; i
< num_sacks
; i
++) {
1629 bool dup_sack
= !i
&& found_dup_sack
;
1631 sp
[used_sacks
].start_seq
= get_unaligned_be32(&sp_wire
[i
].start_seq
);
1632 sp
[used_sacks
].end_seq
= get_unaligned_be32(&sp_wire
[i
].end_seq
);
1634 if (!tcp_is_sackblock_valid(tp
, dup_sack
,
1635 sp
[used_sacks
].start_seq
,
1636 sp
[used_sacks
].end_seq
)) {
1640 if (!tp
->undo_marker
)
1641 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDNOUNDO
;
1643 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDOLD
;
1645 /* Don't count olds caused by ACK reordering */
1646 if ((TCP_SKB_CB(ack_skb
)->ack_seq
!= tp
->snd_una
) &&
1647 !after(sp
[used_sacks
].end_seq
, tp
->snd_una
))
1649 mib_idx
= LINUX_MIB_TCPSACKDISCARD
;
1652 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
1654 first_sack_index
= -1;
1658 /* Ignore very old stuff early */
1659 if (!after(sp
[used_sacks
].end_seq
, prior_snd_una
))
1665 /* order SACK blocks to allow in order walk of the retrans queue */
1666 for (i
= used_sacks
- 1; i
> 0; i
--) {
1667 for (j
= 0; j
< i
; j
++) {
1668 if (after(sp
[j
].start_seq
, sp
[j
+ 1].start_seq
)) {
1669 swap(sp
[j
], sp
[j
+ 1]);
1671 /* Track where the first SACK block goes to */
1672 if (j
== first_sack_index
)
1673 first_sack_index
= j
+ 1;
1678 skb
= tcp_write_queue_head(sk
);
1679 state
.fack_count
= 0;
1682 if (!tp
->sacked_out
) {
1683 /* It's already past, so skip checking against it */
1684 cache
= tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1686 cache
= tp
->recv_sack_cache
;
1687 /* Skip empty blocks in at head of the cache */
1688 while (tcp_sack_cache_ok(tp
, cache
) && !cache
->start_seq
&&
1693 while (i
< used_sacks
) {
1694 u32 start_seq
= sp
[i
].start_seq
;
1695 u32 end_seq
= sp
[i
].end_seq
;
1696 bool dup_sack
= (found_dup_sack
&& (i
== first_sack_index
));
1697 struct tcp_sack_block
*next_dup
= NULL
;
1699 if (found_dup_sack
&& ((i
+ 1) == first_sack_index
))
1700 next_dup
= &sp
[i
+ 1];
1702 /* Skip too early cached blocks */
1703 while (tcp_sack_cache_ok(tp
, cache
) &&
1704 !before(start_seq
, cache
->end_seq
))
1707 /* Can skip some work by looking recv_sack_cache? */
1708 if (tcp_sack_cache_ok(tp
, cache
) && !dup_sack
&&
1709 after(end_seq
, cache
->start_seq
)) {
1712 if (before(start_seq
, cache
->start_seq
)) {
1713 skb
= tcp_sacktag_skip(skb
, sk
, &state
,
1715 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
,
1722 /* Rest of the block already fully processed? */
1723 if (!after(end_seq
, cache
->end_seq
))
1726 skb
= tcp_maybe_skipping_dsack(skb
, sk
, next_dup
,
1730 /* ...tail remains todo... */
1731 if (tcp_highest_sack_seq(tp
) == cache
->end_seq
) {
1732 /* ...but better entrypoint exists! */
1733 skb
= tcp_highest_sack(sk
);
1736 state
.fack_count
= tp
->fackets_out
;
1741 skb
= tcp_sacktag_skip(skb
, sk
, &state
, cache
->end_seq
);
1742 /* Check overlap against next cached too (past this one already) */
1747 if (!before(start_seq
, tcp_highest_sack_seq(tp
))) {
1748 skb
= tcp_highest_sack(sk
);
1751 state
.fack_count
= tp
->fackets_out
;
1753 skb
= tcp_sacktag_skip(skb
, sk
, &state
, start_seq
);
1756 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
, &state
,
1757 start_seq
, end_seq
, dup_sack
);
1763 /* Clear the head of the cache sack blocks so we can skip it next time */
1764 for (i
= 0; i
< ARRAY_SIZE(tp
->recv_sack_cache
) - used_sacks
; i
++) {
1765 tp
->recv_sack_cache
[i
].start_seq
= 0;
1766 tp
->recv_sack_cache
[i
].end_seq
= 0;
1768 for (j
= 0; j
< used_sacks
; j
++)
1769 tp
->recv_sack_cache
[i
++] = sp
[j
];
1771 tcp_mark_lost_retrans(sk
);
1773 tcp_verify_left_out(tp
);
1775 if ((state
.reord
< tp
->fackets_out
) &&
1776 ((inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Loss
) || tp
->undo_marker
))
1777 tcp_update_reordering(sk
, tp
->fackets_out
- state
.reord
, 0);
1781 #if FASTRETRANS_DEBUG > 0
1782 WARN_ON((int)tp
->sacked_out
< 0);
1783 WARN_ON((int)tp
->lost_out
< 0);
1784 WARN_ON((int)tp
->retrans_out
< 0);
1785 WARN_ON((int)tcp_packets_in_flight(tp
) < 0);
1790 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1791 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1793 static bool tcp_limit_reno_sacked(struct tcp_sock
*tp
)
1797 holes
= max(tp
->lost_out
, 1U);
1798 holes
= min(holes
, tp
->packets_out
);
1800 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1801 tp
->sacked_out
= tp
->packets_out
- holes
;
1807 /* If we receive more dupacks than we expected counting segments
1808 * in assumption of absent reordering, interpret this as reordering.
1809 * The only another reason could be bug in receiver TCP.
1811 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1813 struct tcp_sock
*tp
= tcp_sk(sk
);
1814 if (tcp_limit_reno_sacked(tp
))
1815 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1818 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1820 static void tcp_add_reno_sack(struct sock
*sk
)
1822 struct tcp_sock
*tp
= tcp_sk(sk
);
1824 tcp_check_reno_reordering(sk
, 0);
1825 tcp_verify_left_out(tp
);
1828 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1830 static void tcp_remove_reno_sacks(struct sock
*sk
, int acked
)
1832 struct tcp_sock
*tp
= tcp_sk(sk
);
1835 /* One ACK acked hole. The rest eat duplicate ACKs. */
1836 if (acked
- 1 >= tp
->sacked_out
)
1839 tp
->sacked_out
-= acked
- 1;
1841 tcp_check_reno_reordering(sk
, acked
);
1842 tcp_verify_left_out(tp
);
1845 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1850 static void tcp_clear_retrans_partial(struct tcp_sock
*tp
)
1852 tp
->retrans_out
= 0;
1855 tp
->undo_marker
= 0;
1856 tp
->undo_retrans
= -1;
1859 void tcp_clear_retrans(struct tcp_sock
*tp
)
1861 tcp_clear_retrans_partial(tp
);
1863 tp
->fackets_out
= 0;
1867 /* Enter Loss state. If "how" is not zero, forget all SACK information
1868 * and reset tags completely, otherwise preserve SACKs. If receiver
1869 * dropped its ofo queue, we will know this due to reneging detection.
1871 void tcp_enter_loss(struct sock
*sk
, int how
)
1873 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1874 struct inet_connection_sock
*icsk1
= inet_csk(sk
);
1875 struct tcp_sock
*tp
= tcp_sk(sk
);
1876 struct sk_buff
*skb
;
1877 bool new_recovery
= false;
1879 /* Reduce ssthresh if it has not yet been made inside this window. */
1880 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
||
1881 !after(tp
->high_seq
, tp
->snd_una
) ||
1882 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
1883 new_recovery
= true;
1884 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1885 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1886 tcp_ca_event(sk
, CA_EVENT_LOSS
);
1888 if (icsk
->icsk_MMSRB
== 1)
1890 #ifdef CONFIG_MTK_NET_LOGGING
1891 printk("[mtk_net][mmspb] tcp_enter_loss snd_cwnd=%u, snd_cwnd_cnt=%u\n", tp
->snd_cwnd
, tp
->snd_cwnd_cnt
);
1893 if (tp
->mss_cache
!= 0)
1894 tp
->snd_cwnd
= (tp
->rcv_wnd
/ tp
->mss_cache
);
1897 tp
->snd_cwnd
= (tp
->rcv_wnd
/ tp
->advmss
);
1900 if (tp
->snd_ssthresh
> 16)
1902 tp
->snd_cwnd
= tp
->snd_ssthresh
/ 2;//set snd_cwnd is half of default snd_ssthresh
1906 tp
->snd_cwnd
= tp
->snd_ssthresh
/ 2 + 4;
1908 #ifdef CONFIG_MTK_NET_LOGGING
1909 printk("[mtk_net][mmspb] tcp_enter_loss update snd_cwnd=%u\n", tp
->snd_cwnd
);
1911 icsk1
->icsk_MMSRB
= 0;
1912 #ifdef CONFIG_MTK_NET_LOGGING
1913 printk("[mtk_net][mmspb] tcp_enter_loss set icsk_MMSRB=0\n");
1922 tp
->snd_cwnd_cnt
= 0;
1923 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1925 tcp_clear_retrans_partial(tp
);
1927 if (tcp_is_reno(tp
))
1928 tcp_reset_reno_sack(tp
);
1930 tp
->undo_marker
= tp
->snd_una
;
1933 tp
->fackets_out
= 0;
1935 tcp_clear_all_retrans_hints(tp
);
1937 tcp_for_write_queue(skb
, sk
) {
1938 if (skb
== tcp_send_head(sk
))
1941 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
1942 tp
->undo_marker
= 0;
1943 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
1944 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) || how
) {
1945 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
1946 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1947 tp
->lost_out
+= tcp_skb_pcount(skb
);
1948 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
1951 tcp_verify_left_out(tp
);
1953 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1954 sysctl_tcp_reordering
);
1955 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1956 tp
->high_seq
= tp
->snd_nxt
;
1957 TCP_ECN_queue_cwr(tp
);
1959 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1960 * loss recovery is underway except recurring timeout(s) on
1961 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1963 tp
->frto
= sysctl_tcp_frto
&&
1964 (new_recovery
|| icsk
->icsk_retransmits
) &&
1965 !inet_csk(sk
)->icsk_mtup
.probe_size
;
1968 /* If ACK arrived pointing to a remembered SACK, it means that our
1969 * remembered SACKs do not reflect real state of receiver i.e.
1970 * receiver _host_ is heavily congested (or buggy).
1972 * Do processing similar to RTO timeout.
1974 static bool tcp_check_sack_reneging(struct sock
*sk
, int flag
)
1976 if (flag
& FLAG_SACK_RENEGING
) {
1977 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1978 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSACKRENEGING
);
1980 tcp_enter_loss(sk
, 1);
1981 icsk
->icsk_retransmits
++;
1982 tcp_retransmit_skb(sk
, tcp_write_queue_head(sk
));
1983 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1984 icsk
->icsk_rto
, sysctl_tcp_rto_max
);
1990 static inline int tcp_fackets_out(const struct tcp_sock
*tp
)
1992 return tcp_is_reno(tp
) ? tp
->sacked_out
+ 1 : tp
->fackets_out
;
1995 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
1996 * counter when SACK is enabled (without SACK, sacked_out is used for
1999 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2000 * segments up to the highest received SACK block so far and holes in
2003 * With reordering, holes may still be in flight, so RFC3517 recovery
2004 * uses pure sacked_out (total number of SACKed segments) even though
2005 * it violates the RFC that uses duplicate ACKs, often these are equal
2006 * but when e.g. out-of-window ACKs or packet duplication occurs,
2007 * they differ. Since neither occurs due to loss, TCP should really
2010 static inline int tcp_dupack_heuristics(const struct tcp_sock
*tp
)
2012 return tcp_is_fack(tp
) ? tp
->fackets_out
: tp
->sacked_out
+ 1;
2015 static bool tcp_pause_early_retransmit(struct sock
*sk
, int flag
)
2017 struct tcp_sock
*tp
= tcp_sk(sk
);
2018 unsigned long delay
;
2020 /* Delay early retransmit and entering fast recovery for
2021 * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
2022 * available, or RTO is scheduled to fire first.
2024 if (sysctl_tcp_early_retrans
< 2 || sysctl_tcp_early_retrans
> 3 ||
2025 (flag
& FLAG_ECE
) || !tp
->srtt
)
2028 delay
= max_t(unsigned long, (tp
->srtt
>> 5), msecs_to_jiffies(2));
2029 if (!time_after(inet_csk(sk
)->icsk_timeout
, (jiffies
+ delay
)))
2032 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_EARLY_RETRANS
, delay
,
2033 sysctl_tcp_rto_max
);
2037 static inline int tcp_skb_timedout(const struct sock
*sk
,
2038 const struct sk_buff
*skb
)
2040 return tcp_time_stamp
- TCP_SKB_CB(skb
)->when
> inet_csk(sk
)->icsk_rto
;
2043 static inline int tcp_head_timedout(const struct sock
*sk
)
2045 const struct tcp_sock
*tp
= tcp_sk(sk
);
2047 return tp
->packets_out
&&
2048 tcp_skb_timedout(sk
, tcp_write_queue_head(sk
));
2051 /* Linux NewReno/SACK/FACK/ECN state machine.
2052 * --------------------------------------
2054 * "Open" Normal state, no dubious events, fast path.
2055 * "Disorder" In all the respects it is "Open",
2056 * but requires a bit more attention. It is entered when
2057 * we see some SACKs or dupacks. It is split of "Open"
2058 * mainly to move some processing from fast path to slow one.
2059 * "CWR" CWND was reduced due to some Congestion Notification event.
2060 * It can be ECN, ICMP source quench, local device congestion.
2061 * "Recovery" CWND was reduced, we are fast-retransmitting.
2062 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2064 * tcp_fastretrans_alert() is entered:
2065 * - each incoming ACK, if state is not "Open"
2066 * - when arrived ACK is unusual, namely:
2071 * Counting packets in flight is pretty simple.
2073 * in_flight = packets_out - left_out + retrans_out
2075 * packets_out is SND.NXT-SND.UNA counted in packets.
2077 * retrans_out is number of retransmitted segments.
2079 * left_out is number of segments left network, but not ACKed yet.
2081 * left_out = sacked_out + lost_out
2083 * sacked_out: Packets, which arrived to receiver out of order
2084 * and hence not ACKed. With SACKs this number is simply
2085 * amount of SACKed data. Even without SACKs
2086 * it is easy to give pretty reliable estimate of this number,
2087 * counting duplicate ACKs.
2089 * lost_out: Packets lost by network. TCP has no explicit
2090 * "loss notification" feedback from network (for now).
2091 * It means that this number can be only _guessed_.
2092 * Actually, it is the heuristics to predict lossage that
2093 * distinguishes different algorithms.
2095 * F.e. after RTO, when all the queue is considered as lost,
2096 * lost_out = packets_out and in_flight = retrans_out.
2098 * Essentially, we have now two algorithms counting
2101 * FACK: It is the simplest heuristics. As soon as we decided
2102 * that something is lost, we decide that _all_ not SACKed
2103 * packets until the most forward SACK are lost. I.e.
2104 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2105 * It is absolutely correct estimate, if network does not reorder
2106 * packets. And it loses any connection to reality when reordering
2107 * takes place. We use FACK by default until reordering
2108 * is suspected on the path to this destination.
2110 * NewReno: when Recovery is entered, we assume that one segment
2111 * is lost (classic Reno). While we are in Recovery and
2112 * a partial ACK arrives, we assume that one more packet
2113 * is lost (NewReno). This heuristics are the same in NewReno
2116 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2117 * deflation etc. CWND is real congestion window, never inflated, changes
2118 * only according to classic VJ rules.
2120 * Really tricky (and requiring careful tuning) part of algorithm
2121 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2122 * The first determines the moment _when_ we should reduce CWND and,
2123 * hence, slow down forward transmission. In fact, it determines the moment
2124 * when we decide that hole is caused by loss, rather than by a reorder.
2126 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2127 * holes, caused by lost packets.
2129 * And the most logically complicated part of algorithm is undo
2130 * heuristics. We detect false retransmits due to both too early
2131 * fast retransmit (reordering) and underestimated RTO, analyzing
2132 * timestamps and D-SACKs. When we detect that some segments were
2133 * retransmitted by mistake and CWND reduction was wrong, we undo
2134 * window reduction and abort recovery phase. This logic is hidden
2135 * inside several functions named tcp_try_undo_<something>.
2138 /* This function decides, when we should leave Disordered state
2139 * and enter Recovery phase, reducing congestion window.
2141 * Main question: may we further continue forward transmission
2142 * with the same cwnd?
2144 static bool tcp_time_to_recover(struct sock
*sk
, int flag
)
2146 struct tcp_sock
*tp
= tcp_sk(sk
);
2149 /* Trick#1: The loss is proven. */
2153 /* Not-A-Trick#2 : Classic rule... */
2154 if (tcp_dupack_heuristics(tp
) > tp
->reordering
)
2157 /* Trick#3 : when we use RFC2988 timer restart, fast
2158 * retransmit can be triggered by timeout of queue head.
2160 if (tcp_is_fack(tp
) && tcp_head_timedout(sk
))
2163 /* Trick#4: It is still not OK... But will it be useful to delay
2166 packets_out
= tp
->packets_out
;
2167 if (packets_out
<= tp
->reordering
&&
2168 tp
->sacked_out
>= max_t(__u32
, packets_out
/2, sysctl_tcp_reordering
) &&
2169 !tcp_may_send_now(sk
)) {
2170 /* We have nothing to send. This connection is limited
2171 * either by receiver window or by application.
2176 /* If a thin stream is detected, retransmit after first
2177 * received dupack. Employ only if SACK is supported in order
2178 * to avoid possible corner-case series of spurious retransmissions
2179 * Use only if there are no unsent data.
2181 if ((tp
->thin_dupack
|| sysctl_tcp_thin_dupack
) &&
2182 tcp_stream_is_thin(tp
) && tcp_dupack_heuristics(tp
) > 1 &&
2183 tcp_is_sack(tp
) && !tcp_send_head(sk
))
2186 /* Trick#6: TCP early retransmit, per RFC5827. To avoid spurious
2187 * retransmissions due to small network reorderings, we implement
2188 * Mitigation A.3 in the RFC and delay the retransmission for a short
2189 * interval if appropriate.
2191 if (tp
->do_early_retrans
&& !tp
->retrans_out
&& tp
->sacked_out
&&
2192 (tp
->packets_out
>= (tp
->sacked_out
+ 1) && tp
->packets_out
< 4) &&
2193 !tcp_may_send_now(sk
))
2194 return !tcp_pause_early_retransmit(sk
, flag
);
2199 /* New heuristics: it is possible only after we switched to restart timer
2200 * each time when something is ACKed. Hence, we can detect timed out packets
2201 * during fast retransmit without falling to slow start.
2203 * Usefulness of this as is very questionable, since we should know which of
2204 * the segments is the next to timeout which is relatively expensive to find
2205 * in general case unless we add some data structure just for that. The
2206 * current approach certainly won't find the right one too often and when it
2207 * finally does find _something_ it usually marks large part of the window
2208 * right away (because a retransmission with a larger timestamp blocks the
2209 * loop from advancing). -ij
2211 static void tcp_timeout_skbs(struct sock
*sk
)
2213 struct tcp_sock
*tp
= tcp_sk(sk
);
2214 struct sk_buff
*skb
;
2216 if (!tcp_is_fack(tp
) || !tcp_head_timedout(sk
))
2219 skb
= tp
->scoreboard_skb_hint
;
2220 if (tp
->scoreboard_skb_hint
== NULL
)
2221 skb
= tcp_write_queue_head(sk
);
2223 tcp_for_write_queue_from(skb
, sk
) {
2224 if (skb
== tcp_send_head(sk
))
2226 if (!tcp_skb_timedout(sk
, skb
))
2229 tcp_skb_mark_lost(tp
, skb
);
2232 tp
->scoreboard_skb_hint
= skb
;
2234 tcp_verify_left_out(tp
);
2237 /* Detect loss in event "A" above by marking head of queue up as lost.
2238 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2239 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2240 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2241 * the maximum SACKed segments to pass before reaching this limit.
2243 static void tcp_mark_head_lost(struct sock
*sk
, int packets
, int mark_head
)
2245 struct tcp_sock
*tp
= tcp_sk(sk
);
2246 struct sk_buff
*skb
;
2250 /* Use SACK to deduce losses of new sequences sent during recovery */
2251 const u32 loss_high
= tcp_is_sack(tp
) ? tp
->snd_nxt
: tp
->high_seq
;
2253 WARN_ON(packets
> tp
->packets_out
);
2254 if (tp
->lost_skb_hint
) {
2255 skb
= tp
->lost_skb_hint
;
2256 cnt
= tp
->lost_cnt_hint
;
2257 /* Head already handled? */
2258 if (mark_head
&& skb
!= tcp_write_queue_head(sk
))
2261 skb
= tcp_write_queue_head(sk
);
2265 tcp_for_write_queue_from(skb
, sk
) {
2266 if (skb
== tcp_send_head(sk
))
2268 /* TODO: do this better */
2269 /* this is not the most efficient way to do this... */
2270 tp
->lost_skb_hint
= skb
;
2271 tp
->lost_cnt_hint
= cnt
;
2273 if (after(TCP_SKB_CB(skb
)->end_seq
, loss_high
))
2277 if (tcp_is_fack(tp
) || tcp_is_reno(tp
) ||
2278 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2279 cnt
+= tcp_skb_pcount(skb
);
2281 if (cnt
> packets
) {
2282 if ((tcp_is_sack(tp
) && !tcp_is_fack(tp
)) ||
2283 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
) ||
2284 (oldcnt
>= packets
))
2287 mss
= skb_shinfo(skb
)->gso_size
;
2288 err
= tcp_fragment(sk
, skb
, (packets
- oldcnt
) * mss
, mss
);
2294 tcp_skb_mark_lost(tp
, skb
);
2299 tcp_verify_left_out(tp
);
2302 /* Account newly detected lost packet(s) */
2304 static void tcp_update_scoreboard(struct sock
*sk
, int fast_rexmit
)
2306 struct tcp_sock
*tp
= tcp_sk(sk
);
2308 if (tcp_is_reno(tp
)) {
2309 tcp_mark_head_lost(sk
, 1, 1);
2310 } else if (tcp_is_fack(tp
)) {
2311 int lost
= tp
->fackets_out
- tp
->reordering
;
2314 tcp_mark_head_lost(sk
, lost
, 0);
2316 int sacked_upto
= tp
->sacked_out
- tp
->reordering
;
2317 if (sacked_upto
>= 0)
2318 tcp_mark_head_lost(sk
, sacked_upto
, 0);
2319 else if (fast_rexmit
)
2320 tcp_mark_head_lost(sk
, 1, 1);
2323 tcp_timeout_skbs(sk
);
2326 /* CWND moderation, preventing bursts due to too big ACKs
2327 * in dubious situations.
2329 static inline void tcp_moderate_cwnd(struct tcp_sock
*tp
)
2331 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
2332 tcp_packets_in_flight(tp
) + tcp_max_burst(tp
));
2333 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2336 /* Nothing was retransmitted or returned timestamp is less
2337 * than timestamp of the first retransmission.
2339 static inline bool tcp_packet_delayed(const struct tcp_sock
*tp
)
2341 return !tp
->retrans_stamp
||
2342 (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2343 before(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
));
2346 /* Undo procedures. */
2348 #if FASTRETRANS_DEBUG > 1
2349 static void DBGUNDO(struct sock
*sk
, const char *msg
)
2351 struct tcp_sock
*tp
= tcp_sk(sk
);
2352 struct inet_sock
*inet
= inet_sk(sk
);
2354 if (sk
->sk_family
== AF_INET
) {
2355 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2357 &inet
->inet_daddr
, ntohs(inet
->inet_dport
),
2358 tp
->snd_cwnd
, tcp_left_out(tp
),
2359 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2362 #if IS_ENABLED(CONFIG_IPV6)
2363 else if (sk
->sk_family
== AF_INET6
) {
2364 struct ipv6_pinfo
*np
= inet6_sk(sk
);
2365 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2367 &np
->daddr
, ntohs(inet
->inet_dport
),
2368 tp
->snd_cwnd
, tcp_left_out(tp
),
2369 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2375 #define DBGUNDO(x...) do { } while (0)
2378 static void tcp_undo_cwr(struct sock
*sk
, const bool undo_ssthresh
)
2380 struct tcp_sock
*tp
= tcp_sk(sk
);
2382 if (tp
->prior_ssthresh
) {
2383 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2385 if (icsk
->icsk_ca_ops
->undo_cwnd
)
2386 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
2388 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
<< 1);
2390 if (undo_ssthresh
&& tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
2391 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
2392 TCP_ECN_withdraw_cwr(tp
);
2395 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2397 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2400 static inline bool tcp_may_undo(const struct tcp_sock
*tp
)
2402 return tp
->undo_marker
&& (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
2405 /* People celebrate: "We love our President!" */
2406 static bool tcp_try_undo_recovery(struct sock
*sk
)
2408 struct tcp_sock
*tp
= tcp_sk(sk
);
2410 if (tcp_may_undo(tp
)) {
2413 /* Happy end! We did not retransmit anything
2414 * or our original transmission succeeded.
2416 DBGUNDO(sk
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
2417 tcp_undo_cwr(sk
, true);
2418 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
2419 mib_idx
= LINUX_MIB_TCPLOSSUNDO
;
2421 mib_idx
= LINUX_MIB_TCPFULLUNDO
;
2423 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2424 tp
->undo_marker
= 0;
2426 if (tp
->snd_una
== tp
->high_seq
&& tcp_is_reno(tp
)) {
2427 /* Hold old state until something *above* high_seq
2428 * is ACKed. For Reno it is MUST to prevent false
2429 * fast retransmits (RFC2582). SACK TCP is safe. */
2430 tcp_moderate_cwnd(tp
);
2433 tcp_set_ca_state(sk
, TCP_CA_Open
);
2437 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2438 static void tcp_try_undo_dsack(struct sock
*sk
)
2440 struct tcp_sock
*tp
= tcp_sk(sk
);
2442 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
2443 DBGUNDO(sk
, "D-SACK");
2444 tcp_undo_cwr(sk
, true);
2445 tp
->undo_marker
= 0;
2446 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPDSACKUNDO
);
2450 /* We can clear retrans_stamp when there are no retransmissions in the
2451 * window. It would seem that it is trivially available for us in
2452 * tp->retrans_out, however, that kind of assumptions doesn't consider
2453 * what will happen if errors occur when sending retransmission for the
2454 * second time. ...It could the that such segment has only
2455 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2456 * the head skb is enough except for some reneging corner cases that
2457 * are not worth the effort.
2459 * Main reason for all this complexity is the fact that connection dying
2460 * time now depends on the validity of the retrans_stamp, in particular,
2461 * that successive retransmissions of a segment must not advance
2462 * retrans_stamp under any conditions.
2464 static bool tcp_any_retrans_done(const struct sock
*sk
)
2466 const struct tcp_sock
*tp
= tcp_sk(sk
);
2467 struct sk_buff
*skb
;
2469 if (tp
->retrans_out
)
2472 skb
= tcp_write_queue_head(sk
);
2473 if (unlikely(skb
&& TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
))
2479 /* Undo during fast recovery after partial ACK. */
2481 static int tcp_try_undo_partial(struct sock
*sk
, int acked
)
2483 struct tcp_sock
*tp
= tcp_sk(sk
);
2484 /* Partial ACK arrived. Force Hoe's retransmit. */
2485 int failed
= tcp_is_reno(tp
) || (tcp_fackets_out(tp
) > tp
->reordering
);
2487 if (tcp_may_undo(tp
)) {
2488 /* Plain luck! Hole if filled with delayed
2489 * packet, rather than with a retransmit.
2491 if (!tcp_any_retrans_done(sk
))
2492 tp
->retrans_stamp
= 0;
2494 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
2497 tcp_undo_cwr(sk
, false);
2498 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPARTIALUNDO
);
2500 /* So... Do not make Hoe's retransmit yet.
2501 * If the first packet was delayed, the rest
2502 * ones are most probably delayed as well.
2509 /* Undo during loss recovery after partial ACK or using F-RTO. */
2510 static bool tcp_try_undo_loss(struct sock
*sk
, bool frto_undo
)
2512 struct tcp_sock
*tp
= tcp_sk(sk
);
2514 if (frto_undo
|| tcp_may_undo(tp
)) {
2515 struct sk_buff
*skb
;
2516 tcp_for_write_queue(skb
, sk
) {
2517 if (skb
== tcp_send_head(sk
))
2519 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2522 tcp_clear_all_retrans_hints(tp
);
2524 DBGUNDO(sk
, "partial loss");
2526 tcp_undo_cwr(sk
, true);
2527 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSSUNDO
);
2529 NET_INC_STATS_BH(sock_net(sk
),
2530 LINUX_MIB_TCPSPURIOUSRTOS
);
2531 inet_csk(sk
)->icsk_retransmits
= 0;
2532 tp
->undo_marker
= 0;
2533 if (frto_undo
|| tcp_is_sack(tp
))
2534 tcp_set_ca_state(sk
, TCP_CA_Open
);
2540 /* The cwnd reduction in CWR and Recovery use the PRR algorithm
2541 * https://datatracker.ietf.org/doc/draft-ietf-tcpm-proportional-rate-reduction/
2542 * It computes the number of packets to send (sndcnt) based on packets newly
2544 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2545 * cwnd reductions across a full RTT.
2546 * 2) If packets in flight is lower than ssthresh (such as due to excess
2547 * losses and/or application stalls), do not perform any further cwnd
2548 * reductions, but instead slow start up to ssthresh.
2550 static void tcp_init_cwnd_reduction(struct sock
*sk
, const bool set_ssthresh
)
2552 struct tcp_sock
*tp
= tcp_sk(sk
);
2554 tp
->high_seq
= tp
->snd_nxt
;
2555 tp
->tlp_high_seq
= 0;
2556 tp
->snd_cwnd_cnt
= 0;
2557 tp
->prior_cwnd
= tp
->snd_cwnd
;
2558 tp
->prr_delivered
= 0;
2561 tp
->snd_ssthresh
= inet_csk(sk
)->icsk_ca_ops
->ssthresh(sk
);
2562 TCP_ECN_queue_cwr(tp
);
2565 static void tcp_cwnd_reduction(struct sock
*sk
, int newly_acked_sacked
,
2568 struct tcp_sock
*tp
= tcp_sk(sk
);
2570 int delta
= tp
->snd_ssthresh
- tcp_packets_in_flight(tp
);
2572 tp
->prr_delivered
+= newly_acked_sacked
;
2573 if (tcp_packets_in_flight(tp
) > tp
->snd_ssthresh
) {
2574 u64 dividend
= (u64
)tp
->snd_ssthresh
* tp
->prr_delivered
+
2576 sndcnt
= div_u64(dividend
, tp
->prior_cwnd
) - tp
->prr_out
;
2578 sndcnt
= min_t(int, delta
,
2579 max_t(int, tp
->prr_delivered
- tp
->prr_out
,
2580 newly_acked_sacked
) + 1);
2583 sndcnt
= max(sndcnt
, (fast_rexmit
? 1 : 0));
2584 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + sndcnt
;
2587 static inline void tcp_end_cwnd_reduction(struct sock
*sk
)
2589 struct tcp_sock
*tp
= tcp_sk(sk
);
2591 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2592 if (tp
->snd_ssthresh
< TCP_INFINITE_SSTHRESH
&&
2593 (inet_csk(sk
)->icsk_ca_state
== TCP_CA_CWR
|| tp
->undo_marker
)) {
2594 tp
->snd_cwnd
= tp
->snd_ssthresh
;
2595 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2597 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
2600 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2601 void tcp_enter_cwr(struct sock
*sk
, const int set_ssthresh
)
2603 struct tcp_sock
*tp
= tcp_sk(sk
);
2605 tp
->prior_ssthresh
= 0;
2606 if (inet_csk(sk
)->icsk_ca_state
< TCP_CA_CWR
) {
2607 tp
->undo_marker
= 0;
2608 tcp_init_cwnd_reduction(sk
, set_ssthresh
);
2609 tcp_set_ca_state(sk
, TCP_CA_CWR
);
2613 static void tcp_try_keep_open(struct sock
*sk
)
2615 struct tcp_sock
*tp
= tcp_sk(sk
);
2616 int state
= TCP_CA_Open
;
2618 if (tcp_left_out(tp
) || tcp_any_retrans_done(sk
))
2619 state
= TCP_CA_Disorder
;
2621 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
2622 tcp_set_ca_state(sk
, state
);
2623 tp
->high_seq
= tp
->snd_nxt
;
2627 static void tcp_try_to_open(struct sock
*sk
, int flag
, int newly_acked_sacked
)
2629 struct tcp_sock
*tp
= tcp_sk(sk
);
2631 tcp_verify_left_out(tp
);
2633 if (!tcp_any_retrans_done(sk
))
2634 tp
->retrans_stamp
= 0;
2636 if (flag
& FLAG_ECE
)
2637 tcp_enter_cwr(sk
, 1);
2639 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
2640 tcp_try_keep_open(sk
);
2641 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
)
2642 tcp_moderate_cwnd(tp
);
2644 tcp_cwnd_reduction(sk
, newly_acked_sacked
, 0);
2648 static void tcp_mtup_probe_failed(struct sock
*sk
)
2650 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2652 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
2653 icsk
->icsk_mtup
.probe_size
= 0;
2656 static void tcp_mtup_probe_success(struct sock
*sk
)
2658 struct tcp_sock
*tp
= tcp_sk(sk
);
2659 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2661 /* FIXME: breaks with very large cwnd */
2662 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2663 tp
->snd_cwnd
= tp
->snd_cwnd
*
2664 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
2665 icsk
->icsk_mtup
.probe_size
;
2666 tp
->snd_cwnd_cnt
= 0;
2667 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2668 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2670 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
2671 icsk
->icsk_mtup
.probe_size
= 0;
2672 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
2675 /* Do a simple retransmit without using the backoff mechanisms in
2676 * tcp_timer. This is used for path mtu discovery.
2677 * The socket is already locked here.
2679 void tcp_simple_retransmit(struct sock
*sk
)
2681 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2682 struct tcp_sock
*tp
= tcp_sk(sk
);
2683 struct sk_buff
*skb
;
2684 unsigned int mss
= tcp_current_mss(sk
);
2685 u32 prior_lost
= tp
->lost_out
;
2687 tcp_for_write_queue(skb
, sk
) {
2688 if (skb
== tcp_send_head(sk
))
2690 if (tcp_skb_seglen(skb
) > mss
&&
2691 !(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
2692 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2693 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
2694 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2696 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
2700 tcp_clear_retrans_hints_partial(tp
);
2702 if (prior_lost
== tp
->lost_out
)
2705 if (tcp_is_reno(tp
))
2706 tcp_limit_reno_sacked(tp
);
2708 tcp_verify_left_out(tp
);
2710 /* Don't muck with the congestion window here.
2711 * Reason is that we do not increase amount of _data_
2712 * in network, but units changed and effective
2713 * cwnd/ssthresh really reduced now.
2715 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
2716 tp
->high_seq
= tp
->snd_nxt
;
2717 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2718 tp
->prior_ssthresh
= 0;
2719 tp
->undo_marker
= 0;
2720 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2722 tcp_xmit_retransmit_queue(sk
);
2724 EXPORT_SYMBOL(tcp_simple_retransmit
);
2726 static void tcp_enter_recovery(struct sock
*sk
, bool ece_ack
)
2728 struct tcp_sock
*tp
= tcp_sk(sk
);
2731 if (tcp_is_reno(tp
))
2732 mib_idx
= LINUX_MIB_TCPRENORECOVERY
;
2734 mib_idx
= LINUX_MIB_TCPSACKRECOVERY
;
2736 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2738 tp
->prior_ssthresh
= 0;
2739 tp
->undo_marker
= tp
->snd_una
;
2740 tp
->undo_retrans
= tp
->retrans_out
? : -1;
2742 if (inet_csk(sk
)->icsk_ca_state
< TCP_CA_CWR
) {
2744 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2745 tcp_init_cwnd_reduction(sk
, true);
2747 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
2750 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2751 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2753 static void tcp_process_loss(struct sock
*sk
, int flag
, bool is_dupack
)
2755 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2756 struct tcp_sock
*tp
= tcp_sk(sk
);
2757 bool recovered
= !before(tp
->snd_una
, tp
->high_seq
);
2759 if (tp
->frto
) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2760 /* Step 3.b. A timeout is spurious if not all data are
2761 * lost, i.e., never-retransmitted data are (s)acked.
2763 if (tcp_try_undo_loss(sk
, flag
& FLAG_ORIG_SACK_ACKED
))
2766 if (after(tp
->snd_nxt
, tp
->high_seq
) &&
2767 (flag
& FLAG_DATA_SACKED
|| is_dupack
)) {
2768 tp
->frto
= 0; /* Loss was real: 2nd part of step 3.a */
2769 } else if (flag
& FLAG_SND_UNA_ADVANCED
&& !recovered
) {
2770 tp
->high_seq
= tp
->snd_nxt
;
2771 __tcp_push_pending_frames(sk
, tcp_current_mss(sk
),
2773 if (after(tp
->snd_nxt
, tp
->high_seq
))
2774 return; /* Step 2.b */
2780 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2781 icsk
->icsk_retransmits
= 0;
2782 tcp_try_undo_recovery(sk
);
2785 if (flag
& FLAG_DATA_ACKED
)
2786 icsk
->icsk_retransmits
= 0;
2787 if (tcp_is_reno(tp
)) {
2788 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2789 * delivered. Lower inflight to clock out (re)tranmissions.
2791 if (after(tp
->snd_nxt
, tp
->high_seq
) && is_dupack
)
2792 tcp_add_reno_sack(sk
);
2793 else if (flag
& FLAG_SND_UNA_ADVANCED
)
2794 tcp_reset_reno_sack(tp
);
2796 if (tcp_try_undo_loss(sk
, false))
2798 tcp_xmit_retransmit_queue(sk
);
2801 /* Process an event, which can update packets-in-flight not trivially.
2802 * Main goal of this function is to calculate new estimate for left_out,
2803 * taking into account both packets sitting in receiver's buffer and
2804 * packets lost by network.
2806 * Besides that it does CWND reduction, when packet loss is detected
2807 * and changes state of machine.
2809 * It does _not_ decide what to send, it is made in function
2810 * tcp_xmit_retransmit_queue().
2812 static void tcp_fastretrans_alert(struct sock
*sk
, int pkts_acked
,
2813 int prior_sacked
, int prior_packets
,
2814 bool is_dupack
, int flag
)
2816 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2817 struct tcp_sock
*tp
= tcp_sk(sk
);
2818 int do_lost
= is_dupack
|| ((flag
& FLAG_DATA_SACKED
) &&
2819 (tcp_fackets_out(tp
) > tp
->reordering
));
2820 int newly_acked_sacked
= 0;
2821 int fast_rexmit
= 0;
2823 if (WARN_ON(!tp
->packets_out
&& tp
->sacked_out
))
2825 if (WARN_ON(!tp
->sacked_out
&& tp
->fackets_out
))
2826 tp
->fackets_out
= 0;
2828 /* Now state machine starts.
2829 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2830 if (flag
& FLAG_ECE
)
2831 tp
->prior_ssthresh
= 0;
2833 /* B. In all the states check for reneging SACKs. */
2834 if (tcp_check_sack_reneging(sk
, flag
))
2837 /* C. Check consistency of the current state. */
2838 tcp_verify_left_out(tp
);
2840 /* D. Check state exit conditions. State can be terminated
2841 * when high_seq is ACKed. */
2842 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
2843 WARN_ON(tp
->retrans_out
!= 0);
2844 tp
->retrans_stamp
= 0;
2845 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
2846 switch (icsk
->icsk_ca_state
) {
2848 /* CWR is to be held something *above* high_seq
2849 * is ACKed for CWR bit to reach receiver. */
2850 if (tp
->snd_una
!= tp
->high_seq
) {
2851 tcp_end_cwnd_reduction(sk
);
2852 tcp_set_ca_state(sk
, TCP_CA_Open
);
2856 case TCP_CA_Recovery
:
2857 if (tcp_is_reno(tp
))
2858 tcp_reset_reno_sack(tp
);
2859 if (tcp_try_undo_recovery(sk
))
2861 tcp_end_cwnd_reduction(sk
);
2866 /* E. Process state. */
2867 switch (icsk
->icsk_ca_state
) {
2868 case TCP_CA_Recovery
:
2869 if (!(flag
& FLAG_SND_UNA_ADVANCED
)) {
2870 if (tcp_is_reno(tp
) && is_dupack
)
2871 tcp_add_reno_sack(sk
);
2873 do_lost
= tcp_try_undo_partial(sk
, pkts_acked
);
2874 newly_acked_sacked
= prior_packets
- tp
->packets_out
+
2875 tp
->sacked_out
- prior_sacked
;
2878 tcp_process_loss(sk
, flag
, is_dupack
);
2879 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
2881 /* Fall through to processing in Open state. */
2883 if (tcp_is_reno(tp
)) {
2884 if (flag
& FLAG_SND_UNA_ADVANCED
)
2885 tcp_reset_reno_sack(tp
);
2887 tcp_add_reno_sack(sk
);
2889 newly_acked_sacked
= prior_packets
- tp
->packets_out
+
2890 tp
->sacked_out
- prior_sacked
;
2892 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
)
2893 tcp_try_undo_dsack(sk
);
2895 if (!tcp_time_to_recover(sk
, flag
)) {
2896 tcp_try_to_open(sk
, flag
, newly_acked_sacked
);
2900 /* MTU probe failure: don't reduce cwnd */
2901 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
2902 icsk
->icsk_mtup
.probe_size
&&
2903 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
2904 tcp_mtup_probe_failed(sk
);
2905 /* Restores the reduction we did in tcp_mtup_probe() */
2907 tcp_simple_retransmit(sk
);
2911 /* Otherwise enter Recovery state */
2912 tcp_enter_recovery(sk
, (flag
& FLAG_ECE
));
2916 if (do_lost
|| (tcp_is_fack(tp
) && tcp_head_timedout(sk
)))
2917 tcp_update_scoreboard(sk
, fast_rexmit
);
2918 tcp_cwnd_reduction(sk
, newly_acked_sacked
, fast_rexmit
);
2919 tcp_xmit_retransmit_queue(sk
);
2922 void tcp_valid_rtt_meas(struct sock
*sk
, u32 seq_rtt
)
2924 tcp_rtt_estimator(sk
, seq_rtt
);
2926 inet_csk(sk
)->icsk_backoff
= 0;
2928 EXPORT_SYMBOL(tcp_valid_rtt_meas
);
2930 /* Read draft-ietf-tcplw-high-performance before mucking
2931 * with this code. (Supersedes RFC1323)
2933 static void tcp_ack_saw_tstamp(struct sock
*sk
, int flag
)
2935 /* RTTM Rule: A TSecr value received in a segment is used to
2936 * update the averaged RTT measurement only if the segment
2937 * acknowledges some new data, i.e., only if it advances the
2938 * left edge of the send window.
2940 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2941 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2943 * Changed: reset backoff as soon as we see the first valid sample.
2944 * If we do not, we get strongly overestimated rto. With timestamps
2945 * samples are accepted even from very old segments: f.e., when rtt=1
2946 * increases to 8, we retransmit 5 times and after 8 seconds delayed
2947 * answer arrives rto becomes 120 seconds! If at least one of segments
2948 * in window is lost... Voila. --ANK (010210)
2950 struct tcp_sock
*tp
= tcp_sk(sk
);
2952 tcp_valid_rtt_meas(sk
, tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
);
2955 static void tcp_ack_no_tstamp(struct sock
*sk
, u32 seq_rtt
, int flag
)
2957 /* We don't have a timestamp. Can only use
2958 * packets that are not retransmitted to determine
2959 * rtt estimates. Also, we must not reset the
2960 * backoff for rto until we get a non-retransmitted
2961 * packet. This allows us to deal with a situation
2962 * where the network delay has increased suddenly.
2963 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2966 if (flag
& FLAG_RETRANS_DATA_ACKED
)
2969 tcp_valid_rtt_meas(sk
, seq_rtt
);
2972 static inline void tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
2975 const struct tcp_sock
*tp
= tcp_sk(sk
);
2976 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2977 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
2978 tcp_ack_saw_tstamp(sk
, flag
);
2979 else if (seq_rtt
>= 0)
2980 tcp_ack_no_tstamp(sk
, seq_rtt
, flag
);
2983 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
, u32 in_flight
)
2985 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2986 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, in_flight
);
2987 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_time_stamp
;
2990 /* Restart timer after forward progress on connection.
2991 * RFC2988 recommends to restart timer to now+rto.
2993 void tcp_rearm_rto(struct sock
*sk
)
2995 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2996 struct tcp_sock
*tp
= tcp_sk(sk
);
2998 /* If the retrans timer is currently being used by Fast Open
2999 * for SYN-ACK retrans purpose, stay put.
3001 if (tp
->fastopen_rsk
)
3004 if (!tp
->packets_out
) {
3005 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
3007 u32 rto
= inet_csk(sk
)->icsk_rto
;
3008 /* Offset the time elapsed after installing regular RTO */
3009 if (icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
||
3010 icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
) {
3011 s32 delta
= tcp_rto_delta(sk
);
3012 /* delta may not be positive if the socket is locked
3013 * when the retrans timer fires and is rescheduled.
3015 rto
= max_t(int, delta
, 1);
3017 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
, rto
,
3018 sysctl_tcp_rto_max
);
3022 /* This function is called when the delayed ER timer fires. TCP enters
3023 * fast recovery and performs fast-retransmit.
3025 void tcp_resume_early_retransmit(struct sock
*sk
)
3027 struct tcp_sock
*tp
= tcp_sk(sk
);
3031 /* Stop if ER is disabled after the delayed ER timer is scheduled */
3032 if (!tp
->do_early_retrans
)
3035 tcp_enter_recovery(sk
, false);
3036 tcp_update_scoreboard(sk
, 1);
3037 tcp_xmit_retransmit_queue(sk
);
3040 /* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */
3041 static void tcp_set_xmit_timer(struct sock
*sk
)
3043 if (!tcp_schedule_loss_probe(sk
))
3047 /* If we get here, the whole TSO packet has not been acked. */
3048 static u32
tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
)
3050 struct tcp_sock
*tp
= tcp_sk(sk
);
3053 BUG_ON(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
));
3055 packets_acked
= tcp_skb_pcount(skb
);
3056 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
3058 packets_acked
-= tcp_skb_pcount(skb
);
3060 if (packets_acked
) {
3061 BUG_ON(tcp_skb_pcount(skb
) == 0);
3062 BUG_ON(!before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
));
3065 return packets_acked
;
3068 /* Remove acknowledged frames from the retransmission queue. If our packet
3069 * is before the ack sequence we can discard it as it's confirmed to have
3070 * arrived at the other end.
3072 static int tcp_clean_rtx_queue(struct sock
*sk
, int prior_fackets
,
3075 struct tcp_sock
*tp
= tcp_sk(sk
);
3076 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3077 struct sk_buff
*skb
;
3078 u32 now
= tcp_time_stamp
;
3079 int fully_acked
= true;
3082 u32 reord
= tp
->packets_out
;
3083 u32 prior_sacked
= tp
->sacked_out
;
3085 s32 ca_seq_rtt
= -1;
3086 ktime_t last_ackt
= net_invalid_timestamp();
3088 while ((skb
= tcp_write_queue_head(sk
)) && skb
!= tcp_send_head(sk
)) {
3089 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
3091 u8 sacked
= scb
->sacked
;
3093 /* Determine how many packets and what bytes were acked, tso and else */
3094 if (after(scb
->end_seq
, tp
->snd_una
)) {
3095 if (tcp_skb_pcount(skb
) == 1 ||
3096 !after(tp
->snd_una
, scb
->seq
))
3099 acked_pcount
= tcp_tso_acked(sk
, skb
);
3103 fully_acked
= false;
3105 acked_pcount
= tcp_skb_pcount(skb
);
3108 if (sacked
& TCPCB_RETRANS
) {
3109 if (sacked
& TCPCB_SACKED_RETRANS
)
3110 tp
->retrans_out
-= acked_pcount
;
3111 flag
|= FLAG_RETRANS_DATA_ACKED
;
3115 ca_seq_rtt
= now
- scb
->when
;
3116 last_ackt
= skb
->tstamp
;
3118 seq_rtt
= ca_seq_rtt
;
3120 if (!(sacked
& TCPCB_SACKED_ACKED
)) {
3121 reord
= min(pkts_acked
, reord
);
3122 if (!after(scb
->end_seq
, tp
->high_seq
))
3123 flag
|= FLAG_ORIG_SACK_ACKED
;
3127 if (sacked
& TCPCB_SACKED_ACKED
)
3128 tp
->sacked_out
-= acked_pcount
;
3129 if (sacked
& TCPCB_LOST
)
3130 tp
->lost_out
-= acked_pcount
;
3132 tp
->packets_out
-= acked_pcount
;
3133 pkts_acked
+= acked_pcount
;
3135 /* Initial outgoing SYN's get put onto the write_queue
3136 * just like anything else we transmit. It is not
3137 * true data, and if we misinform our callers that
3138 * this ACK acks real data, we will erroneously exit
3139 * connection startup slow start one packet too
3140 * quickly. This is severely frowned upon behavior.
3142 if (!(scb
->tcp_flags
& TCPHDR_SYN
)) {
3143 flag
|= FLAG_DATA_ACKED
;
3145 flag
|= FLAG_SYN_ACKED
;
3146 tp
->retrans_stamp
= 0;
3152 tcp_unlink_write_queue(skb
, sk
);
3153 sk_wmem_free_skb(sk
, skb
);
3154 tp
->scoreboard_skb_hint
= NULL
;
3155 if (skb
== tp
->retransmit_skb_hint
)
3156 tp
->retransmit_skb_hint
= NULL
;
3157 if (skb
== tp
->lost_skb_hint
)
3158 tp
->lost_skb_hint
= NULL
;
3161 if (likely(between(tp
->snd_up
, prior_snd_una
, tp
->snd_una
)))
3162 tp
->snd_up
= tp
->snd_una
;
3164 if (skb
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
3165 flag
|= FLAG_SACK_RENEGING
;
3167 if (flag
& FLAG_ACKED
) {
3168 const struct tcp_congestion_ops
*ca_ops
3169 = inet_csk(sk
)->icsk_ca_ops
;
3171 if (unlikely(icsk
->icsk_mtup
.probe_size
&&
3172 !after(tp
->mtu_probe
.probe_seq_end
, tp
->snd_una
))) {
3173 tcp_mtup_probe_success(sk
);
3176 tcp_ack_update_rtt(sk
, flag
, seq_rtt
);
3177 flag
|= FLAG_SET_XMIT_TIMER
; /* set TLP or RTO timer */
3179 if (tcp_is_reno(tp
)) {
3180 tcp_remove_reno_sacks(sk
, pkts_acked
);
3184 /* Non-retransmitted hole got filled? That's reordering */
3185 if (reord
< prior_fackets
)
3186 tcp_update_reordering(sk
, tp
->fackets_out
- reord
, 0);
3188 delta
= tcp_is_fack(tp
) ? pkts_acked
:
3189 prior_sacked
- tp
->sacked_out
;
3190 tp
->lost_cnt_hint
-= min(tp
->lost_cnt_hint
, delta
);
3193 tp
->fackets_out
-= min(pkts_acked
, tp
->fackets_out
);
3195 if (ca_ops
->pkts_acked
) {
3198 /* Is the ACK triggering packet unambiguous? */
3199 if (!(flag
& FLAG_RETRANS_DATA_ACKED
)) {
3200 /* High resolution needed and available? */
3201 if (ca_ops
->flags
& TCP_CONG_RTT_STAMP
&&
3202 !ktime_equal(last_ackt
,
3203 net_invalid_timestamp()))
3204 rtt_us
= ktime_us_delta(ktime_get_real(),
3206 else if (ca_seq_rtt
>= 0)
3207 rtt_us
= jiffies_to_usecs(ca_seq_rtt
);
3210 ca_ops
->pkts_acked(sk
, pkts_acked
, rtt_us
);
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;
3240 static void tcp_ack_probe(struct sock
*sk
)
3242 const struct tcp_sock
*tp
= tcp_sk(sk
);
3243 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3245 /* Was it a usable window open? */
3247 if (!after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
, tcp_wnd_end(tp
))) {
3248 icsk
->icsk_backoff
= 0;
3249 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
3250 /* Socket must be waked up by subsequent tcp_data_snd_check().
3251 * This function is not for random using!
3254 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3255 min_t(unsigned int, icsk
->icsk_rto
<< icsk
->icsk_backoff
, sysctl_tcp_rto_max
),
3256 sysctl_tcp_rto_max
);
3260 static inline bool tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
3262 return !(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
3263 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
;
3266 static inline bool tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
3268 const struct tcp_sock
*tp
= tcp_sk(sk
);
3269 return (!(flag
& FLAG_ECE
) || tp
->snd_cwnd
< tp
->snd_ssthresh
) &&
3270 !tcp_in_cwnd_reduction(sk
);
3273 /* Check that window update is acceptable.
3274 * The function assumes that snd_una<=ack<=snd_next.
3276 static inline bool tcp_may_update_window(const struct tcp_sock
*tp
,
3277 const u32 ack
, const u32 ack_seq
,
3280 return after(ack
, tp
->snd_una
) ||
3281 after(ack_seq
, tp
->snd_wl1
) ||
3282 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
);
3285 /* Update our send window.
3287 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3288 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3290 static int tcp_ack_update_window(struct sock
*sk
, const struct sk_buff
*skb
, u32 ack
,
3293 struct tcp_sock
*tp
= tcp_sk(sk
);
3295 u32 nwin
= ntohs(tcp_hdr(skb
)->window
);
3297 if (likely(!tcp_hdr(skb
)->syn
))
3298 nwin
<<= tp
->rx_opt
.snd_wscale
;
3300 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
3301 flag
|= FLAG_WIN_UPDATE
;
3302 tcp_update_wl(tp
, ack_seq
);
3304 if (tp
->snd_wnd
!= nwin
) {
3307 /* Note, it is the only place, where
3308 * fast path is recovered for sending TCP.
3311 tcp_fast_path_check(sk
);
3313 if (nwin
> tp
->max_window
) {
3314 tp
->max_window
= nwin
;
3315 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
3325 /* RFC 5961 7 [ACK Throttling] */
3326 static void tcp_send_challenge_ack(struct sock
*sk
)
3328 /* unprotected vars, we dont care of overwrites */
3329 static u32 challenge_timestamp
;
3330 static unsigned int challenge_count
;
3331 u32 now
= jiffies
/ HZ
;
3334 if (now
!= challenge_timestamp
) {
3335 u32 half
= (sysctl_tcp_challenge_ack_limit
+ 1) >> 1;
3337 challenge_timestamp
= now
;
3338 ACCESS_ONCE(challenge_count
) = half
+
3339 reciprocal_divide(prandom_u32(),
3340 sysctl_tcp_challenge_ack_limit
);
3342 count
= ACCESS_ONCE(challenge_count
);
3344 ACCESS_ONCE(challenge_count
) = count
- 1;
3345 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPCHALLENGEACK
);
3350 static void tcp_store_ts_recent(struct tcp_sock
*tp
)
3352 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
3353 tp
->rx_opt
.ts_recent_stamp
= get_seconds();
3356 static void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
3358 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
3359 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3360 * extra check below makes sure this can only happen
3361 * for pure ACK frames. -DaveM
3363 * Not only, also it occurs for expired timestamps.
3366 if (tcp_paws_check(&tp
->rx_opt
, 0))
3367 tcp_store_ts_recent(tp
);
3371 /* This routine deals with acks during a TLP episode.
3372 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3374 static void tcp_process_tlp_ack(struct sock
*sk
, u32 ack
, int flag
)
3376 struct tcp_sock
*tp
= tcp_sk(sk
);
3377 bool is_tlp_dupack
= (ack
== tp
->tlp_high_seq
) &&
3378 !(flag
& (FLAG_SND_UNA_ADVANCED
|
3379 FLAG_NOT_DUP
| FLAG_DATA_SACKED
));
3381 /* Mark the end of TLP episode on receiving TLP dupack or when
3382 * ack is after tlp_high_seq.
3384 if (is_tlp_dupack
) {
3385 tp
->tlp_high_seq
= 0;
3389 if (after(ack
, tp
->tlp_high_seq
)) {
3390 tp
->tlp_high_seq
= 0;
3391 /* Don't reduce cwnd if DSACK arrives for TLP retrans. */
3392 if (!(flag
& FLAG_DSACKING_ACK
)) {
3393 tcp_init_cwnd_reduction(sk
, true);
3394 tcp_set_ca_state(sk
, TCP_CA_CWR
);
3395 tcp_end_cwnd_reduction(sk
);
3396 tcp_try_keep_open(sk
);
3397 NET_INC_STATS_BH(sock_net(sk
),
3398 LINUX_MIB_TCPLOSSPROBERECOVERY
);
3403 /* This routine deals with incoming acks, but not outgoing ones. */
3404 static int tcp_ack(struct sock
*sk
, const struct sk_buff
*skb
, int flag
)
3406 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3407 struct tcp_sock
*tp
= tcp_sk(sk
);
3408 u32 prior_snd_una
= tp
->snd_una
;
3409 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
3410 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3411 bool is_dupack
= false;
3412 u32 prior_in_flight
, prior_cwnd
= tp
->snd_cwnd
, prior_rtt
= tp
->srtt
;
3414 int prior_packets
= tp
->packets_out
;
3415 int prior_sacked
= tp
->sacked_out
;
3417 int previous_packets_out
= 0;
3419 /* If the ack is older than previous acks
3420 * then we can probably ignore it.
3422 if (before(ack
, prior_snd_una
)) {
3423 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3424 if (before(ack
, prior_snd_una
- tp
->max_window
)) {
3425 tcp_send_challenge_ack(sk
);
3431 /* If the ack includes data we haven't sent yet, discard
3432 * this segment (RFC793 Section 3.9).
3434 if (after(ack
, tp
->snd_nxt
))
3437 if (after(ack
, prior_snd_una
))
3438 flag
|= FLAG_SND_UNA_ADVANCED
;
3440 prior_fackets
= tp
->fackets_out
;
3441 prior_in_flight
= tcp_packets_in_flight(tp
);
3443 /* ts_recent update must be made after we are sure that the packet
3446 if (flag
& FLAG_UPDATE_TS_RECENT
)
3447 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
3449 if (!(flag
& FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
3450 /* Window is constant, pure forward advance.
3451 * No more checks are required.
3452 * Note, we use the fact that SND.UNA>=SND.WL2.
3454 tcp_update_wl(tp
, ack_seq
);
3456 flag
|= FLAG_WIN_UPDATE
;
3458 tcp_ca_event(sk
, CA_EVENT_FAST_ACK
);
3460 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPACKS
);
3462 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
3465 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPUREACKS
);
3467 flag
|= tcp_ack_update_window(sk
, skb
, ack
, ack_seq
);
3469 if (TCP_SKB_CB(skb
)->sacked
)
3470 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
3472 if (TCP_ECN_rcv_ecn_echo(tp
, tcp_hdr(skb
)))
3475 tcp_ca_event(sk
, CA_EVENT_SLOW_ACK
);
3478 /* We passed data and got it acked, remove any soft error
3479 * log. Something worked...
3481 sk
->sk_err_soft
= 0;
3482 icsk
->icsk_probes_out
= 0;
3483 tp
->rcv_tstamp
= tcp_time_stamp
;
3487 /* See if we can take anything off of the retransmit queue. */
3488 previous_packets_out
= tp
->packets_out
;
3489 flag
|= tcp_clean_rtx_queue(sk
, prior_fackets
, prior_snd_una
);
3491 pkts_acked
= previous_packets_out
- tp
->packets_out
;
3493 if (tp
->tlp_high_seq
)
3494 tcp_process_tlp_ack(sk
, ack
, flag
);
3495 /* If needed, reset TLP/RTO timer; RACK may later override this. */
3496 if (flag
& FLAG_SET_XMIT_TIMER
)
3497 tcp_set_xmit_timer(sk
);
3499 if (tcp_ack_is_dubious(sk
, flag
)) {
3500 /* Advance CWND, if state allows this. */
3501 if ((flag
& FLAG_DATA_ACKED
) && tcp_may_raise_cwnd(sk
, flag
))
3502 tcp_cong_avoid(sk
, ack
, prior_in_flight
);
3503 is_dupack
= !(flag
& (FLAG_SND_UNA_ADVANCED
| FLAG_NOT_DUP
));
3504 tcp_fastretrans_alert(sk
, pkts_acked
, prior_sacked
,
3505 prior_packets
, is_dupack
, flag
);
3507 if (flag
& FLAG_DATA_ACKED
)
3508 tcp_cong_avoid(sk
, ack
, prior_in_flight
);
3511 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
& FLAG_NOT_DUP
)) {
3512 struct dst_entry
*dst
= __sk_dst_get(sk
);
3517 if (tp
->srtt
!= prior_rtt
|| tp
->snd_cwnd
!= prior_cwnd
)
3518 tcp_update_pacing_rate(sk
);
3522 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3523 if (flag
& FLAG_DSACKING_ACK
)
3524 tcp_fastretrans_alert(sk
, pkts_acked
, prior_sacked
,
3525 prior_packets
, is_dupack
, flag
);
3526 /* If this ack opens up a zero window, clear backoff. It was
3527 * being used to time the probes, and is probably far higher than
3528 * it needs to be for normal retransmission.
3530 if (tcp_send_head(sk
))
3533 if (tp
->tlp_high_seq
)
3534 tcp_process_tlp_ack(sk
, ack
, flag
);
3538 SOCK_DEBUG(sk
, "Ack %u after %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3542 /* If data was SACKed, tag it and see if we should send more data.
3543 * If data was DSACKed, see if we can undo a cwnd reduction.
3545 if (TCP_SKB_CB(skb
)->sacked
) {
3546 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
3547 tcp_fastretrans_alert(sk
, pkts_acked
, prior_sacked
,
3548 prior_packets
, is_dupack
, flag
);
3551 SOCK_DEBUG(sk
, "Ack %u before %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3555 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3556 * But, this can also be called on packets in the established flow when
3557 * the fast version below fails.
3559 void tcp_parse_options(const struct sk_buff
*skb
,
3560 struct tcp_options_received
*opt_rx
, int estab
,
3561 struct tcp_fastopen_cookie
*foc
)
3563 const unsigned char *ptr
;
3564 const struct tcphdr
*th
= tcp_hdr(skb
);
3565 int length
= (th
->doff
* 4) - sizeof(struct tcphdr
);
3567 ptr
= (const unsigned char *)(th
+ 1);
3568 opt_rx
->saw_tstamp
= 0;
3570 while (length
> 0) {
3571 int opcode
= *ptr
++;
3577 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
3582 if (opsize
< 2) /* "silly options" */
3584 if (opsize
> length
)
3585 return; /* don't parse partial options */
3588 if (opsize
== TCPOLEN_MSS
&& th
->syn
&& !estab
) {
3589 u16 in_mss
= get_unaligned_be16(ptr
);
3591 if (opt_rx
->user_mss
&&
3592 opt_rx
->user_mss
< in_mss
)
3593 in_mss
= opt_rx
->user_mss
;
3594 opt_rx
->mss_clamp
= in_mss
;
3599 if (opsize
== TCPOLEN_WINDOW
&& th
->syn
&&
3600 !estab
&& sysctl_tcp_window_scaling
) {
3601 __u8 snd_wscale
= *(__u8
*)ptr
;
3602 opt_rx
->wscale_ok
= 1;
3603 if (snd_wscale
> 14) {
3604 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3609 opt_rx
->snd_wscale
= snd_wscale
;
3612 case TCPOPT_TIMESTAMP
:
3613 if ((opsize
== TCPOLEN_TIMESTAMP
) &&
3614 ((estab
&& opt_rx
->tstamp_ok
) ||
3615 (!estab
&& sysctl_tcp_timestamps
))) {
3616 opt_rx
->saw_tstamp
= 1;
3617 opt_rx
->rcv_tsval
= get_unaligned_be32(ptr
);
3618 opt_rx
->rcv_tsecr
= get_unaligned_be32(ptr
+ 4);
3621 case TCPOPT_SACK_PERM
:
3622 if (opsize
== TCPOLEN_SACK_PERM
&& th
->syn
&&
3623 !estab
&& sysctl_tcp_sack
) {
3624 opt_rx
->sack_ok
= TCP_SACK_SEEN
;
3625 tcp_sack_reset(opt_rx
);
3630 if ((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
3631 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
3633 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
3636 #ifdef CONFIG_TCP_MD5SIG
3639 * The MD5 Hash has already been
3640 * checked (see tcp_v{4,6}_do_rcv()).
3645 /* Fast Open option shares code 254 using a
3646 * 16 bits magic number. It's valid only in
3647 * SYN or SYN-ACK with an even size.
3649 if (opsize
< TCPOLEN_EXP_FASTOPEN_BASE
||
3650 get_unaligned_be16(ptr
) != TCPOPT_FASTOPEN_MAGIC
||
3651 foc
== NULL
|| !th
->syn
|| (opsize
& 1))
3653 foc
->len
= opsize
- TCPOLEN_EXP_FASTOPEN_BASE
;
3654 if (foc
->len
>= TCP_FASTOPEN_COOKIE_MIN
&&
3655 foc
->len
<= TCP_FASTOPEN_COOKIE_MAX
)
3656 memcpy(foc
->val
, ptr
+ 2, foc
->len
);
3657 else if (foc
->len
!= 0)
3667 EXPORT_SYMBOL(tcp_parse_options
);
3669 static bool tcp_parse_aligned_timestamp(struct tcp_sock
*tp
, const struct tcphdr
*th
)
3671 const __be32
*ptr
= (const __be32
*)(th
+ 1);
3673 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3674 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
3675 tp
->rx_opt
.saw_tstamp
= 1;
3677 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3680 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
) - tp
->tsoffset
;
3682 tp
->rx_opt
.rcv_tsecr
= 0;
3688 /* Fast parse options. This hopes to only see timestamps.
3689 * If it is wrong it falls back on tcp_parse_options().
3691 static bool tcp_fast_parse_options(const struct sk_buff
*skb
,
3692 const struct tcphdr
*th
, struct tcp_sock
*tp
)
3694 /* In the spirit of fast parsing, compare doff directly to constant
3695 * values. Because equality is used, short doff can be ignored here.
3697 if (th
->doff
== (sizeof(*th
) / 4)) {
3698 tp
->rx_opt
.saw_tstamp
= 0;
3700 } else if (tp
->rx_opt
.tstamp_ok
&&
3701 th
->doff
== ((sizeof(*th
) + TCPOLEN_TSTAMP_ALIGNED
) / 4)) {
3702 if (tcp_parse_aligned_timestamp(tp
, th
))
3706 tcp_parse_options(skb
, &tp
->rx_opt
, 1, NULL
);
3707 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
3708 tp
->rx_opt
.rcv_tsecr
-= tp
->tsoffset
;
3713 #ifdef CONFIG_TCP_MD5SIG
3715 * Parse MD5 Signature option
3717 const u8
*tcp_parse_md5sig_option(const struct tcphdr
*th
)
3719 int length
= (th
->doff
<< 2) - sizeof(*th
);
3720 const u8
*ptr
= (const u8
*)(th
+ 1);
3722 /* If the TCP option is too short, we can short cut */
3723 if (length
< TCPOLEN_MD5SIG
)
3726 while (length
> 0) {
3727 int opcode
= *ptr
++;
3738 if (opsize
< 2 || opsize
> length
)
3740 if (opcode
== TCPOPT_MD5SIG
)
3741 return opsize
== TCPOLEN_MD5SIG
? ptr
: NULL
;
3748 EXPORT_SYMBOL(tcp_parse_md5sig_option
);
3751 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3753 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3754 * it can pass through stack. So, the following predicate verifies that
3755 * this segment is not used for anything but congestion avoidance or
3756 * fast retransmit. Moreover, we even are able to eliminate most of such
3757 * second order effects, if we apply some small "replay" window (~RTO)
3758 * to timestamp space.
3760 * All these measures still do not guarantee that we reject wrapped ACKs
3761 * on networks with high bandwidth, when sequence space is recycled fastly,
3762 * but it guarantees that such events will be very rare and do not affect
3763 * connection seriously. This doesn't look nice, but alas, PAWS is really
3766 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3767 * states that events when retransmit arrives after original data are rare.
3768 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3769 * the biggest problem on large power networks even with minor reordering.
3770 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3771 * up to bandwidth of 18Gigabit/sec. 8) ]
3774 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
3776 const struct tcp_sock
*tp
= tcp_sk(sk
);
3777 const struct tcphdr
*th
= tcp_hdr(skb
);
3778 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3779 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3781 return (/* 1. Pure ACK with correct sequence number. */
3782 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
3784 /* 2. ... and duplicate ACK. */
3785 ack
== tp
->snd_una
&&
3787 /* 3. ... and does not update window. */
3788 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
3790 /* 4. ... and sits in replay window. */
3791 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
3794 static inline bool tcp_paws_discard(const struct sock
*sk
,
3795 const struct sk_buff
*skb
)
3797 const struct tcp_sock
*tp
= tcp_sk(sk
);
3799 return !tcp_paws_check(&tp
->rx_opt
, TCP_PAWS_WINDOW
) &&
3800 !tcp_disordered_ack(sk
, skb
);
3803 /* Check segment sequence number for validity.
3805 * Segment controls are considered valid, if the segment
3806 * fits to the window after truncation to the window. Acceptability
3807 * of data (and SYN, FIN, of course) is checked separately.
3808 * See tcp_data_queue(), for example.
3810 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3811 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3812 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3813 * (borrowed from freebsd)
3816 static inline bool tcp_sequence(const struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3818 return !before(end_seq
, tp
->rcv_wup
) &&
3819 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
3822 /* When we get a reset we do this. */
3823 void tcp_reset(struct sock
*sk
)
3825 /* We want the right error as BSD sees it (and indeed as we do). */
3826 switch (sk
->sk_state
) {
3828 sk
->sk_err
= ECONNREFUSED
;
3830 case TCP_CLOSE_WAIT
:
3836 sk
->sk_err
= ECONNRESET
;
3838 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3841 if (!sock_flag(sk
, SOCK_DEAD
))
3842 sk
->sk_error_report(sk
);
3848 * Process the FIN bit. This now behaves as it is supposed to work
3849 * and the FIN takes effect when it is validly part of sequence
3850 * space. Not before when we get holes.
3852 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3853 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3856 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3857 * close and we go into CLOSING (and later onto TIME-WAIT)
3859 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3861 static void tcp_fin(struct sock
*sk
)
3863 struct tcp_sock
*tp
= tcp_sk(sk
);
3865 inet_csk_schedule_ack(sk
);
3867 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
3868 sock_set_flag(sk
, SOCK_DONE
);
3870 switch (sk
->sk_state
) {
3872 case TCP_ESTABLISHED
:
3873 /* Move to CLOSE_WAIT */
3874 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
3875 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
3878 case TCP_CLOSE_WAIT
:
3880 /* Received a retransmission of the FIN, do
3885 /* RFC793: Remain in the LAST-ACK state. */
3889 /* This case occurs when a simultaneous close
3890 * happens, we must ack the received FIN and
3891 * enter the CLOSING state.
3894 tcp_set_state(sk
, TCP_CLOSING
);
3897 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3899 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
3902 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3903 * cases we should never reach this piece of code.
3905 pr_err("%s: Impossible, sk->sk_state=%d\n",
3906 __func__
, sk
->sk_state
);
3910 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3911 * Probably, we should reset in this case. For now drop them.
3913 __skb_queue_purge(&tp
->out_of_order_queue
);
3914 if (tcp_is_sack(tp
))
3915 tcp_sack_reset(&tp
->rx_opt
);
3918 if (!sock_flag(sk
, SOCK_DEAD
)) {
3919 sk
->sk_state_change(sk
);
3921 /* Do not send POLL_HUP for half duplex close. */
3922 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
3923 sk
->sk_state
== TCP_CLOSE
)
3924 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_HUP
);
3926 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_IN
);
3930 static inline bool tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
,
3933 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
3934 if (before(seq
, sp
->start_seq
))
3935 sp
->start_seq
= seq
;
3936 if (after(end_seq
, sp
->end_seq
))
3937 sp
->end_seq
= end_seq
;
3943 static void tcp_dsack_set(struct sock
*sk
, u32 seq
, u32 end_seq
)
3945 struct tcp_sock
*tp
= tcp_sk(sk
);
3947 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
3950 if (before(seq
, tp
->rcv_nxt
))
3951 mib_idx
= LINUX_MIB_TCPDSACKOLDSENT
;
3953 mib_idx
= LINUX_MIB_TCPDSACKOFOSENT
;
3955 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
3957 tp
->rx_opt
.dsack
= 1;
3958 tp
->duplicate_sack
[0].start_seq
= seq
;
3959 tp
->duplicate_sack
[0].end_seq
= end_seq
;
3963 static void tcp_dsack_extend(struct sock
*sk
, u32 seq
, u32 end_seq
)
3965 struct tcp_sock
*tp
= tcp_sk(sk
);
3967 if (!tp
->rx_opt
.dsack
)
3968 tcp_dsack_set(sk
, seq
, end_seq
);
3970 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
3973 static void tcp_send_dupack(struct sock
*sk
, const struct sk_buff
*skb
)
3975 struct tcp_sock
*tp
= tcp_sk(sk
);
3977 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
3978 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
3979 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
3980 tcp_enter_quickack_mode(sk
);
3982 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
3983 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
3985 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
3986 end_seq
= tp
->rcv_nxt
;
3987 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, end_seq
);
3994 /* These routines update the SACK block as out-of-order packets arrive or
3995 * in-order packets close up the sequence space.
3997 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
4000 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4001 struct tcp_sack_block
*swalk
= sp
+ 1;
4003 /* See if the recent change to the first SACK eats into
4004 * or hits the sequence space of other SACK blocks, if so coalesce.
4006 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
;) {
4007 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
4010 /* Zap SWALK, by moving every further SACK up by one slot.
4011 * Decrease num_sacks.
4013 tp
->rx_opt
.num_sacks
--;
4014 for (i
= this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
4018 this_sack
++, swalk
++;
4022 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
4024 struct tcp_sock
*tp
= tcp_sk(sk
);
4025 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4026 int cur_sacks
= tp
->rx_opt
.num_sacks
;
4032 for (this_sack
= 0; this_sack
< cur_sacks
; this_sack
++, sp
++) {
4033 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
4034 /* Rotate this_sack to the first one. */
4035 for (; this_sack
> 0; this_sack
--, sp
--)
4036 swap(*sp
, *(sp
- 1));
4038 tcp_sack_maybe_coalesce(tp
);
4043 /* Could not find an adjacent existing SACK, build a new one,
4044 * put it at the front, and shift everyone else down. We
4045 * always know there is at least one SACK present already here.
4047 * If the sack array is full, forget about the last one.
4049 if (this_sack
>= TCP_NUM_SACKS
) {
4051 tp
->rx_opt
.num_sacks
--;
4054 for (; this_sack
> 0; this_sack
--, sp
--)
4058 /* Build the new head SACK, and we're done. */
4059 sp
->start_seq
= seq
;
4060 sp
->end_seq
= end_seq
;
4061 tp
->rx_opt
.num_sacks
++;
4064 /* RCV.NXT advances, some SACKs should be eaten. */
4066 static void tcp_sack_remove(struct tcp_sock
*tp
)
4068 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4069 int num_sacks
= tp
->rx_opt
.num_sacks
;
4072 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4073 if (skb_queue_empty(&tp
->out_of_order_queue
)) {
4074 tp
->rx_opt
.num_sacks
= 0;
4078 for (this_sack
= 0; this_sack
< num_sacks
;) {
4079 /* Check if the start of the sack is covered by RCV.NXT. */
4080 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
4083 /* RCV.NXT must cover all the block! */
4084 WARN_ON(before(tp
->rcv_nxt
, sp
->end_seq
));
4086 /* Zap this SACK, by moving forward any other SACKS. */
4087 for (i
=this_sack
+1; i
< num_sacks
; i
++)
4088 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
4095 tp
->rx_opt
.num_sacks
= num_sacks
;
4098 /* This one checks to see if we can put data from the
4099 * out_of_order queue into the receive_queue.
4101 static void tcp_ofo_queue(struct sock
*sk
)
4103 struct tcp_sock
*tp
= tcp_sk(sk
);
4104 __u32 dsack_high
= tp
->rcv_nxt
;
4105 struct sk_buff
*skb
;
4107 while ((skb
= skb_peek(&tp
->out_of_order_queue
)) != NULL
) {
4108 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
4111 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
4112 __u32 dsack
= dsack_high
;
4113 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
4114 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
4115 tcp_dsack_extend(sk
, TCP_SKB_CB(skb
)->seq
, dsack
);
4118 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4119 SOCK_DEBUG(sk
, "ofo packet was already received\n");
4120 __skb_unlink(skb
, &tp
->out_of_order_queue
);
4124 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
4125 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4126 TCP_SKB_CB(skb
)->end_seq
);
4128 __skb_unlink(skb
, &tp
->out_of_order_queue
);
4129 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4130 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4131 if (tcp_hdr(skb
)->fin
)
4136 static bool tcp_prune_ofo_queue(struct sock
*sk
);
4137 static int tcp_prune_queue(struct sock
*sk
);
4139 static int tcp_try_rmem_schedule(struct sock
*sk
, struct sk_buff
*skb
,
4142 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
4143 !sk_rmem_schedule(sk
, skb
, size
)) {
4145 if (tcp_prune_queue(sk
) < 0)
4148 if (!sk_rmem_schedule(sk
, skb
, size
)) {
4149 if (!tcp_prune_ofo_queue(sk
))
4152 if (!sk_rmem_schedule(sk
, skb
, size
))
4160 * tcp_try_coalesce - try to merge skb to prior one
4163 * @from: buffer to add in queue
4164 * @fragstolen: pointer to boolean
4166 * Before queueing skb @from after @to, try to merge them
4167 * to reduce overall memory use and queue lengths, if cost is small.
4168 * Packets in ofo or receive queues can stay a long time.
4169 * Better try to coalesce them right now to avoid future collapses.
4170 * Returns true if caller should free @from instead of queueing it
4172 static bool tcp_try_coalesce(struct sock
*sk
,
4174 struct sk_buff
*from
,
4179 *fragstolen
= false;
4181 if (tcp_hdr(from
)->fin
)
4184 /* Its possible this segment overlaps with prior segment in queue */
4185 if (TCP_SKB_CB(from
)->seq
!= TCP_SKB_CB(to
)->end_seq
)
4188 if (!skb_try_coalesce(to
, from
, fragstolen
, &delta
))
4191 atomic_add(delta
, &sk
->sk_rmem_alloc
);
4192 sk_mem_charge(sk
, delta
);
4193 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRCVCOALESCE
);
4194 TCP_SKB_CB(to
)->end_seq
= TCP_SKB_CB(from
)->end_seq
;
4195 TCP_SKB_CB(to
)->ack_seq
= TCP_SKB_CB(from
)->ack_seq
;
4199 static void tcp_data_queue_ofo(struct sock
*sk
, struct sk_buff
*skb
)
4201 struct tcp_sock
*tp
= tcp_sk(sk
);
4202 struct sk_buff
*skb1
;
4205 TCP_ECN_check_ce(tp
, skb
);
4207 if (unlikely(tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))) {
4208 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFODROP
);
4213 /* Disable header prediction. */
4215 inet_csk_schedule_ack(sk
);
4217 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFOQUEUE
);
4218 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
4219 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4221 skb1
= skb_peek_tail(&tp
->out_of_order_queue
);
4223 /* Initial out of order segment, build 1 SACK. */
4224 if (tcp_is_sack(tp
)) {
4225 tp
->rx_opt
.num_sacks
= 1;
4226 tp
->selective_acks
[0].start_seq
= TCP_SKB_CB(skb
)->seq
;
4227 tp
->selective_acks
[0].end_seq
=
4228 TCP_SKB_CB(skb
)->end_seq
;
4230 __skb_queue_head(&tp
->out_of_order_queue
, skb
);
4234 seq
= TCP_SKB_CB(skb
)->seq
;
4235 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4237 if (seq
== TCP_SKB_CB(skb1
)->end_seq
) {
4240 if (!tcp_try_coalesce(sk
, skb1
, skb
, &fragstolen
)) {
4241 __skb_queue_after(&tp
->out_of_order_queue
, skb1
, skb
);
4243 kfree_skb_partial(skb
, fragstolen
);
4247 if (!tp
->rx_opt
.num_sacks
||
4248 tp
->selective_acks
[0].end_seq
!= seq
)
4251 /* Common case: data arrive in order after hole. */
4252 tp
->selective_acks
[0].end_seq
= end_seq
;
4256 /* Find place to insert this segment. */
4258 if (!after(TCP_SKB_CB(skb1
)->seq
, seq
))
4260 if (skb_queue_is_first(&tp
->out_of_order_queue
, skb1
)) {
4264 skb1
= skb_queue_prev(&tp
->out_of_order_queue
, skb1
);
4267 /* Do skb overlap to previous one? */
4268 if (skb1
&& before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4269 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4270 /* All the bits are present. Drop. */
4271 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFOMERGE
);
4274 tcp_dsack_set(sk
, seq
, end_seq
);
4277 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
4278 /* Partial overlap. */
4279 tcp_dsack_set(sk
, seq
,
4280 TCP_SKB_CB(skb1
)->end_seq
);
4282 if (skb_queue_is_first(&tp
->out_of_order_queue
,
4286 skb1
= skb_queue_prev(
4287 &tp
->out_of_order_queue
,
4292 __skb_queue_head(&tp
->out_of_order_queue
, skb
);
4294 __skb_queue_after(&tp
->out_of_order_queue
, skb1
, skb
);
4296 /* And clean segments covered by new one as whole. */
4297 while (!skb_queue_is_last(&tp
->out_of_order_queue
, skb
)) {
4298 skb1
= skb_queue_next(&tp
->out_of_order_queue
, skb
);
4300 if (!after(end_seq
, TCP_SKB_CB(skb1
)->seq
))
4302 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4303 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4307 __skb_unlink(skb1
, &tp
->out_of_order_queue
);
4308 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4309 TCP_SKB_CB(skb1
)->end_seq
);
4310 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFOMERGE
);
4315 if (tcp_is_sack(tp
))
4316 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
4319 skb_set_owner_r(skb
, sk
);
4322 static int __must_check
tcp_queue_rcv(struct sock
*sk
, struct sk_buff
*skb
, int hdrlen
,
4326 struct sk_buff
*tail
= skb_peek_tail(&sk
->sk_receive_queue
);
4328 __skb_pull(skb
, hdrlen
);
4330 tcp_try_coalesce(sk
, tail
, skb
, fragstolen
)) ? 1 : 0;
4331 tcp_sk(sk
)->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4333 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4334 skb_set_owner_r(skb
, sk
);
4339 int tcp_send_rcvq(struct sock
*sk
, struct msghdr
*msg
, size_t size
)
4341 struct sk_buff
*skb
= NULL
;
4348 skb
= alloc_skb(size
+ sizeof(*th
), sk
->sk_allocation
);
4352 if (tcp_try_rmem_schedule(sk
, skb
, size
+ sizeof(*th
)))
4355 th
= (struct tcphdr
*)skb_put(skb
, sizeof(*th
));
4356 skb_reset_transport_header(skb
);
4357 memset(th
, 0, sizeof(*th
));
4359 if (memcpy_fromiovec(skb_put(skb
, size
), msg
->msg_iov
, size
))
4362 TCP_SKB_CB(skb
)->seq
= tcp_sk(sk
)->rcv_nxt
;
4363 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ size
;
4364 TCP_SKB_CB(skb
)->ack_seq
= tcp_sk(sk
)->snd_una
- 1;
4366 if (tcp_queue_rcv(sk
, skb
, sizeof(*th
), &fragstolen
)) {
4367 WARN_ON_ONCE(fragstolen
); /* should not happen */
4378 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
4380 const struct tcphdr
*th
= tcp_hdr(skb
);
4381 struct tcp_sock
*tp
= tcp_sk(sk
);
4383 bool fragstolen
= false;
4385 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
)
4389 __skb_pull(skb
, th
->doff
* 4);
4391 TCP_ECN_accept_cwr(tp
, skb
);
4393 tp
->rx_opt
.dsack
= 0;
4395 /* Queue data for delivery to the user.
4396 * Packets in sequence go to the receive queue.
4397 * Out of sequence packets to the out_of_order_queue.
4399 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
4400 if (tcp_receive_window(tp
) == 0)
4403 /* Ok. In sequence. In window. */
4404 if (tp
->ucopy
.task
== current
&&
4405 tp
->copied_seq
== tp
->rcv_nxt
&& tp
->ucopy
.len
&&
4406 sock_owned_by_user(sk
) && !tp
->urg_data
) {
4407 int chunk
= min_t(unsigned int, skb
->len
,
4410 __set_current_state(TASK_RUNNING
);
4413 if (!skb_copy_datagram_iovec(skb
, 0, tp
->ucopy
.iov
, chunk
)) {
4414 tp
->ucopy
.len
-= chunk
;
4415 tp
->copied_seq
+= chunk
;
4416 eaten
= (chunk
== skb
->len
);
4417 tcp_rcv_space_adjust(sk
);
4425 tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))
4428 eaten
= tcp_queue_rcv(sk
, skb
, 0, &fragstolen
);
4430 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4432 tcp_event_data_recv(sk
, skb
);
4436 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4439 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4440 * gap in queue is filled.
4442 if (skb_queue_empty(&tp
->out_of_order_queue
))
4443 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
4446 if (tp
->rx_opt
.num_sacks
)
4447 tcp_sack_remove(tp
);
4449 tcp_fast_path_check(sk
);
4452 kfree_skb_partial(skb
, fragstolen
);
4453 if (!sock_flag(sk
, SOCK_DEAD
))
4454 sk
->sk_data_ready(sk
, 0);
4458 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4459 /* A retransmit, 2nd most common case. Force an immediate ack. */
4460 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4461 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4464 tcp_enter_quickack_mode(sk
);
4465 inet_csk_schedule_ack(sk
);
4471 /* Out of window. F.e. zero window probe. */
4472 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
4475 tcp_enter_quickack_mode(sk
);
4477 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4478 /* Partial packet, seq < rcv_next < end_seq */
4479 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
4480 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4481 TCP_SKB_CB(skb
)->end_seq
);
4483 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
4485 /* If window is closed, drop tail of packet. But after
4486 * remembering D-SACK for its head made in previous line.
4488 if (!tcp_receive_window(tp
))
4493 tcp_data_queue_ofo(sk
, skb
);
4496 static struct sk_buff
*tcp_collapse_one(struct sock
*sk
, struct sk_buff
*skb
,
4497 struct sk_buff_head
*list
)
4499 struct sk_buff
*next
= NULL
;
4501 if (!skb_queue_is_last(list
, skb
))
4502 next
= skb_queue_next(list
, skb
);
4504 __skb_unlink(skb
, list
);
4506 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRCVCOLLAPSED
);
4511 /* Collapse contiguous sequence of skbs head..tail with
4512 * sequence numbers start..end.
4514 * If tail is NULL, this means until the end of the list.
4516 * Segments with FIN/SYN are not collapsed (only because this
4520 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
,
4521 struct sk_buff
*head
, struct sk_buff
*tail
,
4524 struct sk_buff
*skb
, *n
;
4527 /* First, check that queue is collapsible and find
4528 * the point where collapsing can be useful. */
4532 skb_queue_walk_from_safe(list
, skb
, n
) {
4535 /* No new bits? It is possible on ofo queue. */
4536 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4537 skb
= tcp_collapse_one(sk
, skb
, list
);
4543 /* The first skb to collapse is:
4545 * - bloated or contains data before "start" or
4546 * overlaps to the next one.
4548 if (!tcp_hdr(skb
)->syn
&& !tcp_hdr(skb
)->fin
&&
4549 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
4550 before(TCP_SKB_CB(skb
)->seq
, start
))) {
4551 end_of_skbs
= false;
4555 if (!skb_queue_is_last(list
, skb
)) {
4556 struct sk_buff
*next
= skb_queue_next(list
, skb
);
4558 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(next
)->seq
) {
4559 end_of_skbs
= false;
4564 /* Decided to skip this, advance start seq. */
4565 start
= TCP_SKB_CB(skb
)->end_seq
;
4567 if (end_of_skbs
|| tcp_hdr(skb
)->syn
|| tcp_hdr(skb
)->fin
)
4570 while (before(start
, end
)) {
4571 struct sk_buff
*nskb
;
4572 unsigned int header
= skb_headroom(skb
);
4573 int copy
= SKB_MAX_ORDER(header
, 0);
4575 /* Too big header? This can happen with IPv6. */
4578 if (end
- start
< copy
)
4580 nskb
= alloc_skb(copy
+ header
, GFP_ATOMIC
);
4584 skb_set_mac_header(nskb
, skb_mac_header(skb
) - skb
->head
);
4585 skb_set_network_header(nskb
, (skb_network_header(skb
) -
4587 skb_set_transport_header(nskb
, (skb_transport_header(skb
) -
4589 skb_reserve(nskb
, header
);
4590 memcpy(nskb
->head
, skb
->head
, header
);
4591 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
4592 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
4593 __skb_queue_before(list
, skb
, nskb
);
4594 skb_set_owner_r(nskb
, sk
);
4596 /* Copy data, releasing collapsed skbs. */
4598 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
4599 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
4603 size
= min(copy
, size
);
4604 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
4606 TCP_SKB_CB(nskb
)->end_seq
+= size
;
4610 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4611 skb
= tcp_collapse_one(sk
, skb
, list
);
4614 tcp_hdr(skb
)->syn
||
4622 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4623 * and tcp_collapse() them until all the queue is collapsed.
4625 static void tcp_collapse_ofo_queue(struct sock
*sk
)
4627 struct tcp_sock
*tp
= tcp_sk(sk
);
4628 struct sk_buff
*skb
= skb_peek(&tp
->out_of_order_queue
);
4629 struct sk_buff
*head
;
4635 start
= TCP_SKB_CB(skb
)->seq
;
4636 end
= TCP_SKB_CB(skb
)->end_seq
;
4640 struct sk_buff
*next
= NULL
;
4642 if (!skb_queue_is_last(&tp
->out_of_order_queue
, skb
))
4643 next
= skb_queue_next(&tp
->out_of_order_queue
, skb
);
4646 /* Segment is terminated when we see gap or when
4647 * we are at the end of all the queue. */
4649 after(TCP_SKB_CB(skb
)->seq
, end
) ||
4650 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
4651 tcp_collapse(sk
, &tp
->out_of_order_queue
,
4652 head
, skb
, start
, end
);
4656 /* Start new segment */
4657 start
= TCP_SKB_CB(skb
)->seq
;
4658 end
= TCP_SKB_CB(skb
)->end_seq
;
4660 if (before(TCP_SKB_CB(skb
)->seq
, start
))
4661 start
= TCP_SKB_CB(skb
)->seq
;
4662 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
4663 end
= TCP_SKB_CB(skb
)->end_seq
;
4669 * Purge the out-of-order queue.
4670 * Return true if queue was pruned.
4672 static bool tcp_prune_ofo_queue(struct sock
*sk
)
4674 struct tcp_sock
*tp
= tcp_sk(sk
);
4677 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4678 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_OFOPRUNED
);
4679 __skb_queue_purge(&tp
->out_of_order_queue
);
4681 /* Reset SACK state. A conforming SACK implementation will
4682 * do the same at a timeout based retransmit. When a connection
4683 * is in a sad state like this, we care only about integrity
4684 * of the connection not performance.
4686 if (tp
->rx_opt
.sack_ok
)
4687 tcp_sack_reset(&tp
->rx_opt
);
4694 /* Reduce allocated memory if we can, trying to get
4695 * the socket within its memory limits again.
4697 * Return less than zero if we should start dropping frames
4698 * until the socket owning process reads some of the data
4699 * to stabilize the situation.
4701 static int tcp_prune_queue(struct sock
*sk
)
4703 struct tcp_sock
*tp
= tcp_sk(sk
);
4705 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
4707 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PRUNECALLED
);
4709 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
4710 tcp_clamp_window(sk
);
4711 else if (sk_under_memory_pressure(sk
))
4712 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
4714 tcp_collapse_ofo_queue(sk
);
4715 if (!skb_queue_empty(&sk
->sk_receive_queue
))
4716 tcp_collapse(sk
, &sk
->sk_receive_queue
,
4717 skb_peek(&sk
->sk_receive_queue
),
4719 tp
->copied_seq
, tp
->rcv_nxt
);
4722 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4725 /* Collapsing did not help, destructive actions follow.
4726 * This must not ever occur. */
4728 tcp_prune_ofo_queue(sk
);
4730 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4733 /* If we are really being abused, tell the caller to silently
4734 * drop receive data on the floor. It will get retransmitted
4735 * and hopefully then we'll have sufficient space.
4737 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_RCVPRUNED
);
4739 /* Massive buffer overcommit. */
4744 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4745 * As additional protections, we do not touch cwnd in retransmission phases,
4746 * and if application hit its sndbuf limit recently.
4748 void tcp_cwnd_application_limited(struct sock
*sk
)
4750 struct tcp_sock
*tp
= tcp_sk(sk
);
4752 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
4753 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
4754 /* Limited by application or receiver window. */
4755 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
4756 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
4757 if (win_used
< tp
->snd_cwnd
) {
4758 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
4759 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
4761 tp
->snd_cwnd_used
= 0;
4763 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4766 static bool tcp_should_expand_sndbuf(const struct sock
*sk
)
4768 const struct tcp_sock
*tp
= tcp_sk(sk
);
4770 /* If the user specified a specific send buffer setting, do
4773 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
4776 /* If we are under global TCP memory pressure, do not expand. */
4777 if (sk_under_memory_pressure(sk
))
4780 /* If we are under soft global TCP memory pressure, do not expand. */
4781 if (sk_memory_allocated(sk
) >= sk_prot_mem_limits(sk
, 0))
4784 /* If we filled the congestion window, do not expand. */
4785 if (tp
->packets_out
>= tp
->snd_cwnd
)
4791 /* When incoming ACK allowed to free some skb from write_queue,
4792 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4793 * on the exit from tcp input handler.
4795 * PROBLEM: sndbuf expansion does not work well with largesend.
4797 static void tcp_new_space(struct sock
*sk
)
4799 struct tcp_sock
*tp
= tcp_sk(sk
);
4801 if (tcp_should_expand_sndbuf(sk
)) {
4802 int sndmem
= SKB_TRUESIZE(max_t(u32
,
4803 tp
->rx_opt
.mss_clamp
,
4806 int demanded
= max_t(unsigned int, tp
->snd_cwnd
,
4807 tp
->reordering
+ 1);
4808 sndmem
*= 2 * demanded
;
4809 if (sndmem
> sk
->sk_sndbuf
)
4810 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
4811 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4814 sk
->sk_write_space(sk
);
4817 static void tcp_check_space(struct sock
*sk
)
4819 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
4820 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
4821 if (sk
->sk_socket
&&
4822 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
4827 static inline void tcp_data_snd_check(struct sock
*sk
)
4829 tcp_push_pending_frames(sk
);
4830 tcp_check_space(sk
);
4834 * Check if sending an ack is needed.
4836 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
4838 struct tcp_sock
*tp
= tcp_sk(sk
);
4840 /* More than one full frame received... */
4841 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
&&
4842 /* ... and right edge of window advances far enough.
4843 * (tcp_recvmsg() will send ACK otherwise). Or...
4845 __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
4846 /* We ACK each frame or... */
4847 tcp_in_quickack_mode(sk
) ||
4848 /* We have out of order data. */
4849 (ofo_possible
&& skb_peek(&tp
->out_of_order_queue
))) {
4850 /* Then ack it now */
4853 /* Else, send delayed ack. */
4854 tcp_send_delayed_ack(sk
);
4858 static inline void tcp_ack_snd_check(struct sock
*sk
)
4860 if (!inet_csk_ack_scheduled(sk
)) {
4861 /* We sent a data segment already. */
4864 __tcp_ack_snd_check(sk
, 1);
4868 * This routine is only called when we have urgent data
4869 * signaled. Its the 'slow' part of tcp_urg. It could be
4870 * moved inline now as tcp_urg is only called from one
4871 * place. We handle URGent data wrong. We have to - as
4872 * BSD still doesn't use the correction from RFC961.
4873 * For 1003.1g we should support a new option TCP_STDURG to permit
4874 * either form (or just set the sysctl tcp_stdurg).
4877 static void tcp_check_urg(struct sock
*sk
, const struct tcphdr
*th
)
4879 struct tcp_sock
*tp
= tcp_sk(sk
);
4880 u32 ptr
= ntohs(th
->urg_ptr
);
4882 if (ptr
&& !sysctl_tcp_stdurg
)
4884 ptr
+= ntohl(th
->seq
);
4886 /* Ignore urgent data that we've already seen and read. */
4887 if (after(tp
->copied_seq
, ptr
))
4890 /* Do not replay urg ptr.
4892 * NOTE: interesting situation not covered by specs.
4893 * Misbehaving sender may send urg ptr, pointing to segment,
4894 * which we already have in ofo queue. We are not able to fetch
4895 * such data and will stay in TCP_URG_NOTYET until will be eaten
4896 * by recvmsg(). Seems, we are not obliged to handle such wicked
4897 * situations. But it is worth to think about possibility of some
4898 * DoSes using some hypothetical application level deadlock.
4900 if (before(ptr
, tp
->rcv_nxt
))
4903 /* Do we already have a newer (or duplicate) urgent pointer? */
4904 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
4907 /* Tell the world about our new urgent pointer. */
4910 /* We may be adding urgent data when the last byte read was
4911 * urgent. To do this requires some care. We cannot just ignore
4912 * tp->copied_seq since we would read the last urgent byte again
4913 * as data, nor can we alter copied_seq until this data arrives
4914 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4916 * NOTE. Double Dutch. Rendering to plain English: author of comment
4917 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4918 * and expect that both A and B disappear from stream. This is _wrong_.
4919 * Though this happens in BSD with high probability, this is occasional.
4920 * Any application relying on this is buggy. Note also, that fix "works"
4921 * only in this artificial test. Insert some normal data between A and B and we will
4922 * decline of BSD again. Verdict: it is better to remove to trap
4925 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
4926 !sock_flag(sk
, SOCK_URGINLINE
) && tp
->copied_seq
!= tp
->rcv_nxt
) {
4927 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
4929 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4930 __skb_unlink(skb
, &sk
->sk_receive_queue
);
4935 tp
->urg_data
= TCP_URG_NOTYET
;
4938 /* Disable header prediction. */
4942 /* This is the 'fast' part of urgent handling. */
4943 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, const struct tcphdr
*th
)
4945 struct tcp_sock
*tp
= tcp_sk(sk
);
4947 /* Check if we get a new urgent pointer - normally not. */
4949 tcp_check_urg(sk
, th
);
4951 /* Do we wait for any urgent data? - normally not... */
4952 if (tp
->urg_data
== TCP_URG_NOTYET
) {
4953 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
4956 /* Is the urgent pointer pointing into this packet? */
4957 if (ptr
< skb
->len
) {
4959 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
4961 tp
->urg_data
= TCP_URG_VALID
| tmp
;
4962 if (!sock_flag(sk
, SOCK_DEAD
))
4963 sk
->sk_data_ready(sk
, 0);
4968 static int tcp_copy_to_iovec(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
4970 struct tcp_sock
*tp
= tcp_sk(sk
);
4971 int chunk
= skb
->len
- hlen
;
4975 if (skb_csum_unnecessary(skb
))
4976 err
= skb_copy_datagram_iovec(skb
, hlen
, tp
->ucopy
.iov
, chunk
);
4978 err
= skb_copy_and_csum_datagram_iovec(skb
, hlen
,
4982 tp
->ucopy
.len
-= chunk
;
4983 tp
->copied_seq
+= chunk
;
4984 tcp_rcv_space_adjust(sk
);
4991 static __sum16
__tcp_checksum_complete_user(struct sock
*sk
,
4992 struct sk_buff
*skb
)
4996 if (sock_owned_by_user(sk
)) {
4998 result
= __tcp_checksum_complete(skb
);
5001 result
= __tcp_checksum_complete(skb
);
5006 static inline bool tcp_checksum_complete_user(struct sock
*sk
,
5007 struct sk_buff
*skb
)
5009 return !skb_csum_unnecessary(skb
) &&
5010 __tcp_checksum_complete_user(sk
, skb
);
5013 #ifdef CONFIG_NET_DMA
5014 static bool tcp_dma_try_early_copy(struct sock
*sk
, struct sk_buff
*skb
,
5017 struct tcp_sock
*tp
= tcp_sk(sk
);
5018 int chunk
= skb
->len
- hlen
;
5020 bool copied_early
= false;
5022 if (tp
->ucopy
.wakeup
)
5025 if (!tp
->ucopy
.dma_chan
&& tp
->ucopy
.pinned_list
)
5026 tp
->ucopy
.dma_chan
= net_dma_find_channel();
5028 if (tp
->ucopy
.dma_chan
&& skb_csum_unnecessary(skb
)) {
5030 dma_cookie
= dma_skb_copy_datagram_iovec(tp
->ucopy
.dma_chan
,
5032 tp
->ucopy
.iov
, chunk
,
5033 tp
->ucopy
.pinned_list
);
5038 tp
->ucopy
.dma_cookie
= dma_cookie
;
5039 copied_early
= true;
5041 tp
->ucopy
.len
-= chunk
;
5042 tp
->copied_seq
+= chunk
;
5043 tcp_rcv_space_adjust(sk
);
5045 if ((tp
->ucopy
.len
== 0) ||
5046 (tcp_flag_word(tcp_hdr(skb
)) & TCP_FLAG_PSH
) ||
5047 (atomic_read(&sk
->sk_rmem_alloc
) > (sk
->sk_rcvbuf
>> 1))) {
5048 tp
->ucopy
.wakeup
= 1;
5049 sk
->sk_data_ready(sk
, 0);
5051 } else if (chunk
> 0) {
5052 tp
->ucopy
.wakeup
= 1;
5053 sk
->sk_data_ready(sk
, 0);
5056 return copied_early
;
5058 #endif /* CONFIG_NET_DMA */
5060 /* Does PAWS and seqno based validation of an incoming segment, flags will
5061 * play significant role here.
5063 static bool tcp_validate_incoming(struct sock
*sk
, struct sk_buff
*skb
,
5064 const struct tcphdr
*th
, int syn_inerr
)
5066 struct tcp_sock
*tp
= tcp_sk(sk
);
5068 /* RFC1323: H1. Apply PAWS check first. */
5069 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
5070 tcp_paws_discard(sk
, skb
)) {
5072 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSESTABREJECTED
);
5073 tcp_send_dupack(sk
, skb
);
5076 /* Reset is accepted even if it did not pass PAWS. */
5079 /* Step 1: check sequence number */
5080 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5081 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5082 * (RST) segments are validated by checking their SEQ-fields."
5083 * And page 69: "If an incoming segment is not acceptable,
5084 * an acknowledgment should be sent in reply (unless the RST
5085 * bit is set, if so drop the segment and return)".
5090 tcp_send_dupack(sk
, skb
);
5095 /* Step 2: check RST bit */
5098 * If sequence number exactly matches RCV.NXT, then
5099 * RESET the connection
5101 * Send a challenge ACK
5103 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
)
5106 tcp_send_challenge_ack(sk
);
5110 /* step 3: check security and precedence [ignored] */
5112 /* step 4: Check for a SYN
5113 * RFC 5691 4.2 : Send a challenge ack
5118 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5119 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSYNCHALLENGE
);
5120 tcp_send_challenge_ack(sk
);
5132 * TCP receive function for the ESTABLISHED state.
5134 * It is split into a fast path and a slow path. The fast path is
5136 * - A zero window was announced from us - zero window probing
5137 * is only handled properly in the slow path.
5138 * - Out of order segments arrived.
5139 * - Urgent data is expected.
5140 * - There is no buffer space left
5141 * - Unexpected TCP flags/window values/header lengths are received
5142 * (detected by checking the TCP header against pred_flags)
5143 * - Data is sent in both directions. Fast path only supports pure senders
5144 * or pure receivers (this means either the sequence number or the ack
5145 * value must stay constant)
5146 * - Unexpected TCP option.
5148 * When these conditions are not satisfied it drops into a standard
5149 * receive procedure patterned after RFC793 to handle all cases.
5150 * The first three cases are guaranteed by proper pred_flags setting,
5151 * the rest is checked inline. Fast processing is turned on in
5152 * tcp_data_queue when everything is OK.
5154 int tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
5155 const struct tcphdr
*th
, unsigned int len
)
5157 struct tcp_sock
*tp
= tcp_sk(sk
);
5159 if (unlikely(sk
->sk_rx_dst
== NULL
))
5160 inet_csk(sk
)->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
5162 * Header prediction.
5163 * The code loosely follows the one in the famous
5164 * "30 instruction TCP receive" Van Jacobson mail.
5166 * Van's trick is to deposit buffers into socket queue
5167 * on a device interrupt, to call tcp_recv function
5168 * on the receive process context and checksum and copy
5169 * the buffer to user space. smart...
5171 * Our current scheme is not silly either but we take the
5172 * extra cost of the net_bh soft interrupt processing...
5173 * We do checksum and copy also but from device to kernel.
5176 tp
->rx_opt
.saw_tstamp
= 0;
5178 /* pred_flags is 0xS?10 << 16 + snd_wnd
5179 * if header_prediction is to be made
5180 * 'S' will always be tp->tcp_header_len >> 2
5181 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5182 * turn it off (when there are holes in the receive
5183 * space for instance)
5184 * PSH flag is ignored.
5187 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
5188 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
&&
5189 !after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
)) {
5190 int tcp_header_len
= tp
->tcp_header_len
;
5192 /* Timestamp header prediction: tcp_header_len
5193 * is automatically equal to th->doff*4 due to pred_flags
5197 /* Check timestamp */
5198 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
5199 /* No? Slow path! */
5200 if (!tcp_parse_aligned_timestamp(tp
, th
))
5203 /* If PAWS failed, check it more carefully in slow path */
5204 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
5207 /* DO NOT update ts_recent here, if checksum fails
5208 * and timestamp was corrupted part, it will result
5209 * in a hung connection since we will drop all
5210 * future packets due to the PAWS test.
5214 if (len
<= tcp_header_len
) {
5215 /* Bulk data transfer: sender */
5216 if (len
== tcp_header_len
) {
5217 /* Predicted packet is in window by definition.
5218 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5219 * Hence, check seq<=rcv_wup reduces to:
5221 if (tcp_header_len
==
5222 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5223 tp
->rcv_nxt
== tp
->rcv_wup
)
5224 tcp_store_ts_recent(tp
);
5226 /* We know that such packets are checksummed
5229 tcp_ack(sk
, skb
, 0);
5231 tcp_data_snd_check(sk
);
5233 } else { /* Header too small */
5234 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5239 int copied_early
= 0;
5240 bool fragstolen
= false;
5242 if (tp
->copied_seq
== tp
->rcv_nxt
&&
5243 len
- tcp_header_len
<= tp
->ucopy
.len
) {
5244 #ifdef CONFIG_NET_DMA
5245 if (tp
->ucopy
.task
== current
&&
5246 sock_owned_by_user(sk
) &&
5247 tcp_dma_try_early_copy(sk
, skb
, tcp_header_len
)) {
5252 if (tp
->ucopy
.task
== current
&&
5253 sock_owned_by_user(sk
) && !copied_early
) {
5254 __set_current_state(TASK_RUNNING
);
5256 if (!tcp_copy_to_iovec(sk
, skb
, tcp_header_len
))
5260 /* Predicted packet is in window by definition.
5261 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5262 * Hence, check seq<=rcv_wup reduces to:
5264 if (tcp_header_len
==
5265 (sizeof(struct tcphdr
) +
5266 TCPOLEN_TSTAMP_ALIGNED
) &&
5267 tp
->rcv_nxt
== tp
->rcv_wup
)
5268 tcp_store_ts_recent(tp
);
5270 tcp_rcv_rtt_measure_ts(sk
, skb
);
5272 __skb_pull(skb
, tcp_header_len
);
5273 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
5274 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITSTOUSER
);
5277 tcp_cleanup_rbuf(sk
, skb
->len
);
5280 if (tcp_checksum_complete_user(sk
, skb
))
5283 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
5286 /* Predicted packet is in window by definition.
5287 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5288 * Hence, check seq<=rcv_wup reduces to:
5290 if (tcp_header_len
==
5291 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5292 tp
->rcv_nxt
== tp
->rcv_wup
)
5293 tcp_store_ts_recent(tp
);
5295 tcp_rcv_rtt_measure_ts(sk
, skb
);
5297 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITS
);
5299 /* Bulk data transfer: receiver */
5300 eaten
= tcp_queue_rcv(sk
, skb
, tcp_header_len
,
5304 tcp_event_data_recv(sk
, skb
);
5306 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
5307 /* Well, only one small jumplet in fast path... */
5308 tcp_ack(sk
, skb
, FLAG_DATA
);
5309 tcp_data_snd_check(sk
);
5310 if (!inet_csk_ack_scheduled(sk
))
5314 if (!copied_early
|| tp
->rcv_nxt
!= tp
->rcv_wup
)
5315 __tcp_ack_snd_check(sk
, 0);
5317 #ifdef CONFIG_NET_DMA
5319 __skb_queue_tail(&sk
->sk_async_wait_queue
, skb
);
5323 kfree_skb_partial(skb
, fragstolen
);
5324 sk
->sk_data_ready(sk
, 0);
5330 if (len
< (th
->doff
<< 2) || tcp_checksum_complete_user(sk
, skb
))
5333 if (!th
->ack
&& !th
->rst
)
5337 * Standard slow path.
5340 if (!tcp_validate_incoming(sk
, skb
, th
, 1))
5344 if (tcp_ack(sk
, skb
, FLAG_SLOWPATH
| FLAG_UPDATE_TS_RECENT
) < 0)
5347 tcp_rcv_rtt_measure_ts(sk
, skb
);
5349 /* Process urgent data. */
5350 tcp_urg(sk
, skb
, th
);
5352 /* step 7: process the segment text */
5353 tcp_data_queue(sk
, skb
);
5355 tcp_data_snd_check(sk
);
5356 tcp_ack_snd_check(sk
);
5360 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_CSUMERRORS
);
5361 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5367 EXPORT_SYMBOL(tcp_rcv_established
);
5369 void tcp_finish_connect(struct sock
*sk
, struct sk_buff
*skb
)
5371 struct tcp_sock
*tp
= tcp_sk(sk
);
5372 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5374 tcp_set_state(sk
, TCP_ESTABLISHED
);
5375 icsk
->icsk_ack
.lrcvtime
= tcp_time_stamp
;
5378 icsk
->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
5379 security_inet_conn_established(sk
, skb
);
5382 /* Make sure socket is routed, for correct metrics. */
5383 icsk
->icsk_af_ops
->rebuild_header(sk
);
5385 tcp_init_metrics(sk
);
5387 tcp_init_congestion_control(sk
);
5389 /* Prevent spurious tcp_cwnd_restart() on first data
5392 tp
->lsndtime
= tcp_time_stamp
;
5394 tcp_init_buffer_space(sk
);
5396 if (sock_flag(sk
, SOCK_KEEPOPEN
))
5397 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
5399 if (!tp
->rx_opt
.snd_wscale
)
5400 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
5404 if (!sock_flag(sk
, SOCK_DEAD
)) {
5405 sk
->sk_state_change(sk
);
5406 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5410 static bool tcp_rcv_fastopen_synack(struct sock
*sk
, struct sk_buff
*synack
,
5411 struct tcp_fastopen_cookie
*cookie
)
5413 struct tcp_sock
*tp
= tcp_sk(sk
);
5414 struct sk_buff
*data
= tp
->syn_data
? tcp_write_queue_head(sk
) : NULL
;
5415 u16 mss
= tp
->rx_opt
.mss_clamp
;
5418 if (mss
== tp
->rx_opt
.user_mss
) {
5419 struct tcp_options_received opt
;
5421 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5422 tcp_clear_options(&opt
);
5423 opt
.user_mss
= opt
.mss_clamp
= 0;
5424 tcp_parse_options(synack
, &opt
, 0, NULL
);
5425 mss
= opt
.mss_clamp
;
5428 if (!tp
->syn_fastopen
) /* Ignore an unsolicited cookie */
5431 /* The SYN-ACK neither has cookie nor acknowledges the data. Presumably
5432 * the remote receives only the retransmitted (regular) SYNs: either
5433 * the original SYN-data or the corresponding SYN-ACK is lost.
5435 syn_drop
= (cookie
->len
<= 0 && data
&& tp
->total_retrans
);
5437 tcp_fastopen_cache_set(sk
, mss
, cookie
, syn_drop
);
5439 if (data
) { /* Retransmit unacked data in SYN */
5440 tcp_for_write_queue_from(data
, sk
) {
5441 if (data
== tcp_send_head(sk
) ||
5442 __tcp_retransmit_skb(sk
, data
))
5448 tp
->syn_data_acked
= tp
->syn_data
;
5452 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5453 const struct tcphdr
*th
, unsigned int len
)
5455 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5456 struct tcp_sock
*tp
= tcp_sk(sk
);
5457 struct tcp_fastopen_cookie foc
= { .len
= -1 };
5458 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
5460 tcp_parse_options(skb
, &tp
->rx_opt
, 0, &foc
);
5461 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
5462 tp
->rx_opt
.rcv_tsecr
-= tp
->tsoffset
;
5466 * "If the state is SYN-SENT then
5467 * first check the ACK bit
5468 * If the ACK bit is set
5469 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5470 * a reset (unless the RST bit is set, if so drop
5471 * the segment and return)"
5473 if (!after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_una
) ||
5474 after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
))
5475 goto reset_and_undo
;
5477 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
5478 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
5480 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSACTIVEREJECTED
);
5481 goto reset_and_undo
;
5484 /* Now ACK is acceptable.
5486 * "If the RST bit is set
5487 * If the ACK was acceptable then signal the user "error:
5488 * connection reset", drop the segment, enter CLOSED state,
5489 * delete TCB, and return."
5498 * "fifth, if neither of the SYN or RST bits is set then
5499 * drop the segment and return."
5505 goto discard_and_undo
;
5508 * "If the SYN bit is on ...
5509 * are acceptable then ...
5510 * (our SYN has been ACKed), change the connection
5511 * state to ESTABLISHED..."
5514 TCP_ECN_rcv_synack(tp
, th
);
5516 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5517 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
5519 /* Ok.. it's good. Set up sequence numbers and
5520 * move to established.
5522 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5523 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5525 /* RFC1323: The window in SYN & SYN/ACK segments is
5528 tp
->snd_wnd
= ntohs(th
->window
);
5530 if (!tp
->rx_opt
.wscale_ok
) {
5531 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
5532 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
5535 if (tp
->rx_opt
.saw_tstamp
) {
5536 tp
->rx_opt
.tstamp_ok
= 1;
5537 tp
->tcp_header_len
=
5538 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5539 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5540 tcp_store_ts_recent(tp
);
5542 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5545 if (tcp_is_sack(tp
) && sysctl_tcp_fack
)
5546 tcp_enable_fack(tp
);
5549 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5550 tcp_initialize_rcv_mss(sk
);
5552 /* Remember, tcp_poll() does not lock socket!
5553 * Change state from SYN-SENT only after copied_seq
5554 * is initialized. */
5555 tp
->copied_seq
= tp
->rcv_nxt
;
5559 tcp_finish_connect(sk
, skb
);
5561 if ((tp
->syn_fastopen
|| tp
->syn_data
) &&
5562 tcp_rcv_fastopen_synack(sk
, skb
, &foc
))
5565 if (sk
->sk_write_pending
||
5566 icsk
->icsk_accept_queue
.rskq_defer_accept
||
5567 icsk
->icsk_ack
.pingpong
) {
5568 /* Save one ACK. Data will be ready after
5569 * several ticks, if write_pending is set.
5571 * It may be deleted, but with this feature tcpdumps
5572 * look so _wonderfully_ clever, that I was not able
5573 * to stand against the temptation 8) --ANK
5575 inet_csk_schedule_ack(sk
);
5576 tcp_enter_quickack_mode(sk
);
5577 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
5578 TCP_DELACK_MAX
, sysctl_tcp_rto_max
);
5589 /* No ACK in the segment */
5593 * "If the RST bit is set
5595 * Otherwise (no ACK) drop the segment and return."
5598 goto discard_and_undo
;
5602 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&&
5603 tcp_paws_reject(&tp
->rx_opt
, 0))
5604 goto discard_and_undo
;
5607 /* We see SYN without ACK. It is attempt of
5608 * simultaneous connect with crossed SYNs.
5609 * Particularly, it can be connect to self.
5611 tcp_set_state(sk
, TCP_SYN_RECV
);
5613 if (tp
->rx_opt
.saw_tstamp
) {
5614 tp
->rx_opt
.tstamp_ok
= 1;
5615 tcp_store_ts_recent(tp
);
5616 tp
->tcp_header_len
=
5617 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5619 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5622 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5623 tp
->copied_seq
= tp
->rcv_nxt
;
5624 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5626 /* RFC1323: The window in SYN & SYN/ACK segments is
5629 tp
->snd_wnd
= ntohs(th
->window
);
5630 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
5631 tp
->max_window
= tp
->snd_wnd
;
5633 TCP_ECN_rcv_syn(tp
, th
);
5636 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5637 tcp_initialize_rcv_mss(sk
);
5639 tcp_send_synack(sk
);
5641 /* Note, we could accept data and URG from this segment.
5642 * There are no obstacles to make this (except that we must
5643 * either change tcp_recvmsg() to prevent it from returning data
5644 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5646 * However, if we ignore data in ACKless segments sometimes,
5647 * we have no reasons to accept it sometimes.
5648 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5649 * is not flawless. So, discard packet for sanity.
5650 * Uncomment this return to process the data.
5657 /* "fifth, if neither of the SYN or RST bits is set then
5658 * drop the segment and return."
5662 tcp_clear_options(&tp
->rx_opt
);
5663 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5667 tcp_clear_options(&tp
->rx_opt
);
5668 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5673 * This function implements the receiving procedure of RFC 793 for
5674 * all states except ESTABLISHED and TIME_WAIT.
5675 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5676 * address independent.
5679 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5680 const struct tcphdr
*th
, unsigned int len
)
5682 struct tcp_sock
*tp
= tcp_sk(sk
);
5683 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5684 struct request_sock
*req
;
5687 tp
->rx_opt
.saw_tstamp
= 0;
5689 switch (sk
->sk_state
) {
5703 if (icsk
->icsk_af_ops
->conn_request(sk
, skb
) < 0)
5706 /* Now we have several options: In theory there is
5707 * nothing else in the frame. KA9Q has an option to
5708 * send data with the syn, BSD accepts data with the
5709 * syn up to the [to be] advertised window and
5710 * Solaris 2.1 gives you a protocol error. For now
5711 * we just ignore it, that fits the spec precisely
5712 * and avoids incompatibilities. It would be nice in
5713 * future to drop through and process the data.
5715 * Now that TTCP is starting to be used we ought to
5717 * But, this leaves one open to an easy denial of
5718 * service attack, and SYN cookies can't defend
5719 * against this problem. So, we drop the data
5720 * in the interest of security over speed unless
5721 * it's still in use.
5729 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
, len
);
5733 /* Do step6 onward by hand. */
5734 tcp_urg(sk
, skb
, th
);
5736 tcp_data_snd_check(sk
);
5740 req
= tp
->fastopen_rsk
;
5742 WARN_ON_ONCE(sk
->sk_state
!= TCP_SYN_RECV
&&
5743 sk
->sk_state
!= TCP_FIN_WAIT1
);
5745 if (tcp_check_req(sk
, skb
, req
, NULL
, true) == NULL
)
5749 if (!th
->ack
&& !th
->rst
)
5752 if (!tcp_validate_incoming(sk
, skb
, th
, 0))
5755 /* step 5: check the ACK field */
5757 int acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
|
5758 FLAG_UPDATE_TS_RECENT
) > 0;
5760 switch (sk
->sk_state
) {
5763 /* Once we leave TCP_SYN_RECV, we no longer
5764 * need req so release it.
5767 tcp_synack_rtt_meas(sk
, req
);
5768 tp
->total_retrans
= req
->num_retrans
;
5770 reqsk_fastopen_remove(sk
, req
, false);
5772 /* Make sure socket is routed, for
5775 icsk
->icsk_af_ops
->rebuild_header(sk
);
5776 tcp_init_congestion_control(sk
);
5779 tcp_init_buffer_space(sk
);
5780 tp
->copied_seq
= tp
->rcv_nxt
;
5783 tcp_set_state(sk
, TCP_ESTABLISHED
);
5784 sk
->sk_state_change(sk
);
5786 /* Note, that this wakeup is only for marginal
5787 * crossed SYN case. Passively open sockets
5788 * are not waked up, because sk->sk_sleep ==
5789 * NULL and sk->sk_socket == NULL.
5793 SOCK_WAKE_IO
, POLL_OUT
);
5795 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
5796 tp
->snd_wnd
= ntohs(th
->window
) <<
5797 tp
->rx_opt
.snd_wscale
;
5798 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5800 if (tp
->rx_opt
.tstamp_ok
)
5801 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5804 /* Re-arm the timer because data may
5805 * have been sent out. This is similar
5806 * to the regular data transmission case
5807 * when new data has just been ack'ed.
5809 * (TFO) - we could try to be more
5810 * aggressive and retranmitting any data
5811 * sooner based on when they were sent
5816 tcp_init_metrics(sk
);
5818 tcp_update_pacing_rate(sk
);
5820 /* Prevent spurious tcp_cwnd_restart() on
5821 * first data packet.
5823 tp
->lsndtime
= tcp_time_stamp
;
5825 tcp_initialize_rcv_mss(sk
);
5826 tcp_fast_path_on(tp
);
5833 /* If we enter the TCP_FIN_WAIT1 state and we are a
5834 * Fast Open socket and this is the first acceptable
5835 * ACK we have received, this would have acknowledged
5836 * our SYNACK so stop the SYNACK timer.
5839 /* Return RST if ack_seq is invalid.
5840 * Note that RFC793 only says to generate a
5841 * DUPACK for it but for TCP Fast Open it seems
5842 * better to treat this case like TCP_SYN_RECV
5847 /* We no longer need the request sock. */
5848 reqsk_fastopen_remove(sk
, req
, false);
5851 if (tp
->snd_una
== tp
->write_seq
) {
5852 struct dst_entry
*dst
;
5854 tcp_set_state(sk
, TCP_FIN_WAIT2
);
5855 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
5857 dst
= __sk_dst_get(sk
);
5861 if (!sock_flag(sk
, SOCK_DEAD
))
5862 /* Wake up lingering close() */
5863 sk
->sk_state_change(sk
);
5867 if (tp
->linger2
< 0 ||
5868 (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5869 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
))) {
5871 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
5875 tmo
= tcp_fin_time(sk
);
5876 if (tmo
> TCP_TIMEWAIT_LEN
) {
5877 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
5878 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
5879 /* Bad case. We could lose such FIN otherwise.
5880 * It is not a big problem, but it looks confusing
5881 * and not so rare event. We still can lose it now,
5882 * if it spins in bh_lock_sock(), but it is really
5885 inet_csk_reset_keepalive_timer(sk
, tmo
);
5887 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
5895 if (tp
->snd_una
== tp
->write_seq
) {
5896 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
5902 if (tp
->snd_una
== tp
->write_seq
) {
5903 tcp_update_metrics(sk
);
5911 /* step 6: check the URG bit */
5912 tcp_urg(sk
, skb
, th
);
5914 /* step 7: process the segment text */
5915 switch (sk
->sk_state
) {
5916 case TCP_CLOSE_WAIT
:
5919 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
5923 /* RFC 793 says to queue data in these states,
5924 * RFC 1122 says we MUST send a reset.
5925 * BSD 4.4 also does reset.
5927 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
5928 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5929 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
5930 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
5936 case TCP_ESTABLISHED
:
5937 tcp_data_queue(sk
, skb
);
5942 /* tcp_data could move socket to TIME-WAIT */
5943 if (sk
->sk_state
!= TCP_CLOSE
) {
5944 tcp_data_snd_check(sk
);
5945 tcp_ack_snd_check(sk
);
5954 EXPORT_SYMBOL(tcp_rcv_state_process
);