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>
73 #include <net/inet_common.h>
74 #include <linux/ipsec.h>
75 #include <asm/unaligned.h>
76 #include <net/netdma.h>
78 int sysctl_tcp_timestamps __read_mostly
= 1;
79 int sysctl_tcp_window_scaling __read_mostly
= 1;
80 int sysctl_tcp_sack __read_mostly
= 1;
81 int sysctl_tcp_fack __read_mostly
= 1;
82 int sysctl_tcp_reordering __read_mostly
= TCP_FASTRETRANS_THRESH
;
83 EXPORT_SYMBOL(sysctl_tcp_reordering
);
84 int sysctl_tcp_dsack __read_mostly
= 1;
85 int sysctl_tcp_app_win __read_mostly
= 31;
86 int sysctl_tcp_adv_win_scale __read_mostly
= 1;
87 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale
);
89 /* rfc5961 challenge ack rate limiting */
90 int sysctl_tcp_challenge_ack_limit
= 100;
92 int sysctl_tcp_stdurg __read_mostly
;
93 int sysctl_tcp_rfc1337 __read_mostly
;
94 int sysctl_tcp_max_orphans __read_mostly
= NR_FILE
;
95 int sysctl_tcp_frto __read_mostly
= 2;
97 int sysctl_tcp_thin_dupack __read_mostly
;
99 int sysctl_tcp_moderate_rcvbuf __read_mostly
= 1;
100 int sysctl_tcp_early_retrans __read_mostly
= 3;
102 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
103 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
104 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
105 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
106 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
107 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
108 #define FLAG_ECE 0x40 /* ECE in this ACK */
109 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
110 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
111 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
112 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
113 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
114 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
116 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
117 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
118 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
119 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
121 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
122 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
124 /* Adapt the MSS value used to make delayed ack decision to the
127 static void tcp_measure_rcv_mss(struct sock
*sk
, const struct sk_buff
*skb
)
129 struct inet_connection_sock
*icsk
= inet_csk(sk
);
130 const unsigned int lss
= icsk
->icsk_ack
.last_seg_size
;
133 icsk
->icsk_ack
.last_seg_size
= 0;
135 /* skb->len may jitter because of SACKs, even if peer
136 * sends good full-sized frames.
138 len
= skb_shinfo(skb
)->gso_size
? : skb
->len
;
139 if (len
>= icsk
->icsk_ack
.rcv_mss
) {
140 icsk
->icsk_ack
.rcv_mss
= len
;
142 /* Otherwise, we make more careful check taking into account,
143 * that SACKs block is variable.
145 * "len" is invariant segment length, including TCP header.
147 len
+= skb
->data
- skb_transport_header(skb
);
148 if (len
>= TCP_MSS_DEFAULT
+ sizeof(struct tcphdr
) ||
149 /* If PSH is not set, packet should be
150 * full sized, provided peer TCP is not badly broken.
151 * This observation (if it is correct 8)) allows
152 * to handle super-low mtu links fairly.
154 (len
>= TCP_MIN_MSS
+ sizeof(struct tcphdr
) &&
155 !(tcp_flag_word(tcp_hdr(skb
)) & TCP_REMNANT
))) {
156 /* Subtract also invariant (if peer is RFC compliant),
157 * tcp header plus fixed timestamp option length.
158 * Resulting "len" is MSS free of SACK jitter.
160 len
-= tcp_sk(sk
)->tcp_header_len
;
161 icsk
->icsk_ack
.last_seg_size
= len
;
163 icsk
->icsk_ack
.rcv_mss
= len
;
167 if (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
)
168 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED2
;
169 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
173 static void tcp_incr_quickack(struct sock
*sk
)
175 struct inet_connection_sock
*icsk
= inet_csk(sk
);
176 unsigned int quickacks
= tcp_sk(sk
)->rcv_wnd
/ (2 * icsk
->icsk_ack
.rcv_mss
);
180 if (quickacks
> icsk
->icsk_ack
.quick
)
181 icsk
->icsk_ack
.quick
= min(quickacks
, TCP_MAX_QUICKACKS
);
184 static void tcp_enter_quickack_mode(struct sock
*sk
)
186 struct inet_connection_sock
*icsk
= inet_csk(sk
);
187 tcp_incr_quickack(sk
);
188 icsk
->icsk_ack
.pingpong
= 0;
189 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
192 /* Send ACKs quickly, if "quick" count is not exhausted
193 * and the session is not interactive.
196 static inline bool tcp_in_quickack_mode(const struct sock
*sk
)
198 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
200 return icsk
->icsk_ack
.quick
&& !icsk
->icsk_ack
.pingpong
;
203 static inline void TCP_ECN_queue_cwr(struct tcp_sock
*tp
)
205 if (tp
->ecn_flags
& TCP_ECN_OK
)
206 tp
->ecn_flags
|= TCP_ECN_QUEUE_CWR
;
209 static inline void TCP_ECN_accept_cwr(struct tcp_sock
*tp
, const struct sk_buff
*skb
)
211 if (tcp_hdr(skb
)->cwr
)
212 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
215 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock
*tp
)
217 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
220 static inline void TCP_ECN_check_ce(struct tcp_sock
*tp
, const struct sk_buff
*skb
)
222 if (!(tp
->ecn_flags
& TCP_ECN_OK
))
225 switch (TCP_SKB_CB(skb
)->ip_dsfield
& INET_ECN_MASK
) {
226 case INET_ECN_NOT_ECT
:
227 /* Funny extension: if ECT is not set on a segment,
228 * and we already seen ECT on a previous segment,
229 * it is probably a retransmit.
231 if (tp
->ecn_flags
& TCP_ECN_SEEN
)
232 tcp_enter_quickack_mode((struct sock
*)tp
);
235 if (!(tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)) {
236 /* Better not delay acks, sender can have a very low cwnd */
237 tcp_enter_quickack_mode((struct sock
*)tp
);
238 tp
->ecn_flags
|= TCP_ECN_DEMAND_CWR
;
242 tp
->ecn_flags
|= TCP_ECN_SEEN
;
246 static inline void TCP_ECN_rcv_synack(struct tcp_sock
*tp
, const struct tcphdr
*th
)
248 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| th
->cwr
))
249 tp
->ecn_flags
&= ~TCP_ECN_OK
;
252 static inline void TCP_ECN_rcv_syn(struct tcp_sock
*tp
, const struct tcphdr
*th
)
254 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| !th
->cwr
))
255 tp
->ecn_flags
&= ~TCP_ECN_OK
;
258 static bool TCP_ECN_rcv_ecn_echo(const struct tcp_sock
*tp
, const struct tcphdr
*th
)
260 if (th
->ece
&& !th
->syn
&& (tp
->ecn_flags
& TCP_ECN_OK
))
265 /* Buffer size and advertised window tuning.
267 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
270 static void tcp_sndbuf_expand(struct sock
*sk
)
272 const struct tcp_sock
*tp
= tcp_sk(sk
);
276 /* Worst case is non GSO/TSO : each frame consumes one skb
277 * and skb->head is kmalloced using power of two area of memory
279 per_mss
= max_t(u32
, tp
->rx_opt
.mss_clamp
, tp
->mss_cache
) +
281 SKB_DATA_ALIGN(sizeof(struct skb_shared_info
));
283 per_mss
= roundup_pow_of_two(per_mss
) +
284 SKB_DATA_ALIGN(sizeof(struct sk_buff
));
286 nr_segs
= max_t(u32
, TCP_INIT_CWND
, tp
->snd_cwnd
);
287 nr_segs
= max_t(u32
, nr_segs
, tp
->reordering
+ 1);
289 /* Fast Recovery (RFC 5681 3.2) :
290 * Cubic needs 1.7 factor, rounded to 2 to include
291 * extra cushion (application might react slowly to POLLOUT)
293 sndmem
= 2 * nr_segs
* per_mss
;
295 if (sk
->sk_sndbuf
< sndmem
)
296 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
299 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
301 * All tcp_full_space() is split to two parts: "network" buffer, allocated
302 * forward and advertised in receiver window (tp->rcv_wnd) and
303 * "application buffer", required to isolate scheduling/application
304 * latencies from network.
305 * window_clamp is maximal advertised window. It can be less than
306 * tcp_full_space(), in this case tcp_full_space() - window_clamp
307 * is reserved for "application" buffer. The less window_clamp is
308 * the smoother our behaviour from viewpoint of network, but the lower
309 * throughput and the higher sensitivity of the connection to losses. 8)
311 * rcv_ssthresh is more strict window_clamp used at "slow start"
312 * phase to predict further behaviour of this connection.
313 * It is used for two goals:
314 * - to enforce header prediction at sender, even when application
315 * requires some significant "application buffer". It is check #1.
316 * - to prevent pruning of receive queue because of misprediction
317 * of receiver window. Check #2.
319 * The scheme does not work when sender sends good segments opening
320 * window and then starts to feed us spaghetti. But it should work
321 * in common situations. Otherwise, we have to rely on queue collapsing.
324 /* Slow part of check#2. */
325 static int __tcp_grow_window(const struct sock
*sk
, const struct sk_buff
*skb
)
327 struct tcp_sock
*tp
= tcp_sk(sk
);
329 int truesize
= tcp_win_from_space(skb
->truesize
) >> 1;
330 int window
= tcp_win_from_space(sysctl_tcp_rmem
[2]) >> 1;
332 while (tp
->rcv_ssthresh
<= window
) {
333 if (truesize
<= skb
->len
)
334 return 2 * inet_csk(sk
)->icsk_ack
.rcv_mss
;
342 static void tcp_grow_window(struct sock
*sk
, const struct sk_buff
*skb
)
344 struct tcp_sock
*tp
= tcp_sk(sk
);
347 if (tp
->rcv_ssthresh
< tp
->window_clamp
&&
348 (int)tp
->rcv_ssthresh
< tcp_space(sk
) &&
349 !sk_under_memory_pressure(sk
)) {
352 /* Check #2. Increase window, if skb with such overhead
353 * will fit to rcvbuf in future.
355 if (tcp_win_from_space(skb
->truesize
) <= skb
->len
)
356 incr
= 2 * tp
->advmss
;
358 incr
= __tcp_grow_window(sk
, skb
);
361 incr
= max_t(int, incr
, 2 * skb
->len
);
362 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
+ incr
,
364 inet_csk(sk
)->icsk_ack
.quick
|= 1;
369 /* 3. Tuning rcvbuf, when connection enters established state. */
370 static void tcp_fixup_rcvbuf(struct sock
*sk
)
372 u32 mss
= tcp_sk(sk
)->advmss
;
375 rcvmem
= 2 * SKB_TRUESIZE(mss
+ MAX_TCP_HEADER
) *
376 tcp_default_init_rwnd(mss
);
378 /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
379 * Allow enough cushion so that sender is not limited by our window
381 if (sysctl_tcp_moderate_rcvbuf
)
384 if (sk
->sk_rcvbuf
< rcvmem
)
385 sk
->sk_rcvbuf
= min(rcvmem
, sysctl_tcp_rmem
[2]);
388 /* 4. Try to fixup all. It is made immediately after connection enters
391 void tcp_init_buffer_space(struct sock
*sk
)
393 struct tcp_sock
*tp
= tcp_sk(sk
);
396 if (!(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
))
397 tcp_fixup_rcvbuf(sk
);
398 if (!(sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
))
399 tcp_sndbuf_expand(sk
);
401 tp
->rcvq_space
.space
= tp
->rcv_wnd
;
402 tp
->rcvq_space
.time
= tcp_time_stamp
;
403 tp
->rcvq_space
.seq
= tp
->copied_seq
;
405 maxwin
= tcp_full_space(sk
);
407 if (tp
->window_clamp
>= maxwin
) {
408 tp
->window_clamp
= maxwin
;
410 if (sysctl_tcp_app_win
&& maxwin
> 4 * tp
->advmss
)
411 tp
->window_clamp
= max(maxwin
-
412 (maxwin
>> sysctl_tcp_app_win
),
416 /* Force reservation of one segment. */
417 if (sysctl_tcp_app_win
&&
418 tp
->window_clamp
> 2 * tp
->advmss
&&
419 tp
->window_clamp
+ tp
->advmss
> maxwin
)
420 tp
->window_clamp
= max(2 * tp
->advmss
, maxwin
- tp
->advmss
);
422 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, tp
->window_clamp
);
423 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
426 /* 5. Recalculate window clamp after socket hit its memory bounds. */
427 static void tcp_clamp_window(struct sock
*sk
)
429 struct tcp_sock
*tp
= tcp_sk(sk
);
430 struct inet_connection_sock
*icsk
= inet_csk(sk
);
432 icsk
->icsk_ack
.quick
= 0;
434 if (sk
->sk_rcvbuf
< sysctl_tcp_rmem
[2] &&
435 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
) &&
436 !sk_under_memory_pressure(sk
) &&
437 sk_memory_allocated(sk
) < sk_prot_mem_limits(sk
, 0)) {
438 sk
->sk_rcvbuf
= min(atomic_read(&sk
->sk_rmem_alloc
),
441 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
)
442 tp
->rcv_ssthresh
= min(tp
->window_clamp
, 2U * tp
->advmss
);
445 /* Initialize RCV_MSS value.
446 * RCV_MSS is an our guess about MSS used by the peer.
447 * We haven't any direct information about the MSS.
448 * It's better to underestimate the RCV_MSS rather than overestimate.
449 * Overestimations make us ACKing less frequently than needed.
450 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
452 void tcp_initialize_rcv_mss(struct sock
*sk
)
454 const struct tcp_sock
*tp
= tcp_sk(sk
);
455 unsigned int hint
= min_t(unsigned int, tp
->advmss
, tp
->mss_cache
);
457 hint
= min(hint
, tp
->rcv_wnd
/ 2);
458 hint
= min(hint
, TCP_MSS_DEFAULT
);
459 hint
= max(hint
, TCP_MIN_MSS
);
461 inet_csk(sk
)->icsk_ack
.rcv_mss
= hint
;
463 EXPORT_SYMBOL(tcp_initialize_rcv_mss
);
465 /* Receiver "autotuning" code.
467 * The algorithm for RTT estimation w/o timestamps is based on
468 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
469 * <http://public.lanl.gov/radiant/pubs.html#DRS>
471 * More detail on this code can be found at
472 * <http://staff.psc.edu/jheffner/>,
473 * though this reference is out of date. A new paper
476 static void tcp_rcv_rtt_update(struct tcp_sock
*tp
, u32 sample
, int win_dep
)
478 u32 new_sample
= tp
->rcv_rtt_est
.rtt
;
484 if (new_sample
!= 0) {
485 /* If we sample in larger samples in the non-timestamp
486 * case, we could grossly overestimate the RTT especially
487 * with chatty applications or bulk transfer apps which
488 * are stalled on filesystem I/O.
490 * Also, since we are only going for a minimum in the
491 * non-timestamp case, we do not smooth things out
492 * else with timestamps disabled convergence takes too
496 m
-= (new_sample
>> 3);
504 /* No previous measure. */
508 if (tp
->rcv_rtt_est
.rtt
!= new_sample
)
509 tp
->rcv_rtt_est
.rtt
= new_sample
;
512 static inline void tcp_rcv_rtt_measure(struct tcp_sock
*tp
)
514 if (tp
->rcv_rtt_est
.time
== 0)
516 if (before(tp
->rcv_nxt
, tp
->rcv_rtt_est
.seq
))
518 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rcv_rtt_est
.time
, 1);
521 tp
->rcv_rtt_est
.seq
= tp
->rcv_nxt
+ tp
->rcv_wnd
;
522 tp
->rcv_rtt_est
.time
= tcp_time_stamp
;
525 static inline void tcp_rcv_rtt_measure_ts(struct sock
*sk
,
526 const struct sk_buff
*skb
)
528 struct tcp_sock
*tp
= tcp_sk(sk
);
529 if (tp
->rx_opt
.rcv_tsecr
&&
530 (TCP_SKB_CB(skb
)->end_seq
-
531 TCP_SKB_CB(skb
)->seq
>= inet_csk(sk
)->icsk_ack
.rcv_mss
))
532 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
, 0);
536 * This function should be called every time data is copied to user space.
537 * It calculates the appropriate TCP receive buffer space.
539 void tcp_rcv_space_adjust(struct sock
*sk
)
541 struct tcp_sock
*tp
= tcp_sk(sk
);
545 time
= tcp_time_stamp
- tp
->rcvq_space
.time
;
546 if (time
< (tp
->rcv_rtt_est
.rtt
>> 3) || tp
->rcv_rtt_est
.rtt
== 0)
549 /* Number of bytes copied to user in last RTT */
550 copied
= tp
->copied_seq
- tp
->rcvq_space
.seq
;
551 if (copied
<= tp
->rcvq_space
.space
)
555 * copied = bytes received in previous RTT, our base window
556 * To cope with packet losses, we need a 2x factor
557 * To cope with slow start, and sender growing its cwin by 100 %
558 * every RTT, we need a 4x factor, because the ACK we are sending
559 * now is for the next RTT, not the current one :
560 * <prev RTT . ><current RTT .. ><next RTT .... >
563 if (sysctl_tcp_moderate_rcvbuf
&&
564 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
)) {
565 int rcvwin
, rcvmem
, rcvbuf
;
567 /* minimal window to cope with packet losses, assuming
568 * steady state. Add some cushion because of small variations.
570 rcvwin
= (copied
<< 1) + 16 * tp
->advmss
;
572 /* If rate increased by 25%,
573 * assume slow start, rcvwin = 3 * copied
574 * If rate increased by 50%,
575 * assume sender can use 2x growth, rcvwin = 4 * copied
578 tp
->rcvq_space
.space
+ (tp
->rcvq_space
.space
>> 2)) {
580 tp
->rcvq_space
.space
+ (tp
->rcvq_space
.space
>> 1))
583 rcvwin
+= (rcvwin
>> 1);
586 rcvmem
= SKB_TRUESIZE(tp
->advmss
+ MAX_TCP_HEADER
);
587 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
590 rcvbuf
= min(rcvwin
/ tp
->advmss
* rcvmem
, sysctl_tcp_rmem
[2]);
591 if (rcvbuf
> sk
->sk_rcvbuf
) {
592 sk
->sk_rcvbuf
= rcvbuf
;
594 /* Make the window clamp follow along. */
595 tp
->window_clamp
= rcvwin
;
598 tp
->rcvq_space
.space
= copied
;
601 tp
->rcvq_space
.seq
= tp
->copied_seq
;
602 tp
->rcvq_space
.time
= tcp_time_stamp
;
605 /* There is something which you must keep in mind when you analyze the
606 * behavior of the tp->ato delayed ack timeout interval. When a
607 * connection starts up, we want to ack as quickly as possible. The
608 * problem is that "good" TCP's do slow start at the beginning of data
609 * transmission. The means that until we send the first few ACK's the
610 * sender will sit on his end and only queue most of his data, because
611 * he can only send snd_cwnd unacked packets at any given time. For
612 * each ACK we send, he increments snd_cwnd and transmits more of his
615 static void tcp_event_data_recv(struct sock
*sk
, struct sk_buff
*skb
)
617 struct tcp_sock
*tp
= tcp_sk(sk
);
618 struct inet_connection_sock
*icsk
= inet_csk(sk
);
621 inet_csk_schedule_ack(sk
);
623 tcp_measure_rcv_mss(sk
, skb
);
625 tcp_rcv_rtt_measure(tp
);
627 now
= tcp_time_stamp
;
629 if (!icsk
->icsk_ack
.ato
) {
630 /* The _first_ data packet received, initialize
631 * delayed ACK engine.
633 tcp_incr_quickack(sk
);
634 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
636 int m
= now
- icsk
->icsk_ack
.lrcvtime
;
638 if (m
<= TCP_ATO_MIN
/ 2) {
639 /* The fastest case is the first. */
640 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + TCP_ATO_MIN
/ 2;
641 } else if (m
< icsk
->icsk_ack
.ato
) {
642 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + m
;
643 if (icsk
->icsk_ack
.ato
> icsk
->icsk_rto
)
644 icsk
->icsk_ack
.ato
= icsk
->icsk_rto
;
645 } else if (m
> icsk
->icsk_rto
) {
646 /* Too long gap. Apparently sender failed to
647 * restart window, so that we send ACKs quickly.
649 tcp_incr_quickack(sk
);
653 icsk
->icsk_ack
.lrcvtime
= now
;
655 TCP_ECN_check_ce(tp
, skb
);
658 tcp_grow_window(sk
, skb
);
661 /* Called to compute a smoothed rtt estimate. The data fed to this
662 * routine either comes from timestamps, or from segments that were
663 * known _not_ to have been retransmitted [see Karn/Partridge
664 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
665 * piece by Van Jacobson.
666 * NOTE: the next three routines used to be one big routine.
667 * To save cycles in the RFC 1323 implementation it was better to break
668 * it up into three procedures. -- erics
670 static void tcp_rtt_estimator(struct sock
*sk
, const __u32 mrtt
)
672 struct tcp_sock
*tp
= tcp_sk(sk
);
673 long m
= mrtt
; /* RTT */
676 /* The following amusing code comes from Jacobson's
677 * article in SIGCOMM '88. Note that rtt and mdev
678 * are scaled versions of rtt and mean deviation.
679 * This is designed to be as fast as possible
680 * m stands for "measurement".
682 * On a 1990 paper the rto value is changed to:
683 * RTO = rtt + 4 * mdev
685 * Funny. This algorithm seems to be very broken.
686 * These formulae increase RTO, when it should be decreased, increase
687 * too slowly, when it should be increased quickly, decrease too quickly
688 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
689 * does not matter how to _calculate_ it. Seems, it was trap
690 * that VJ failed to avoid. 8)
693 m
-= (srtt
>> 3); /* m is now error in rtt est */
694 srtt
+= m
; /* rtt = 7/8 rtt + 1/8 new */
696 m
= -m
; /* m is now abs(error) */
697 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
698 /* This is similar to one of Eifel findings.
699 * Eifel blocks mdev updates when rtt decreases.
700 * This solution is a bit different: we use finer gain
701 * for mdev in this case (alpha*beta).
702 * Like Eifel it also prevents growth of rto,
703 * but also it limits too fast rto decreases,
704 * happening in pure Eifel.
709 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
711 tp
->mdev
+= m
; /* mdev = 3/4 mdev + 1/4 new */
712 if (tp
->mdev
> tp
->mdev_max
) {
713 tp
->mdev_max
= tp
->mdev
;
714 if (tp
->mdev_max
> tp
->rttvar
)
715 tp
->rttvar
= tp
->mdev_max
;
717 if (after(tp
->snd_una
, tp
->rtt_seq
)) {
718 if (tp
->mdev_max
< tp
->rttvar
)
719 tp
->rttvar
-= (tp
->rttvar
- tp
->mdev_max
) >> 2;
720 tp
->rtt_seq
= tp
->snd_nxt
;
721 tp
->mdev_max
= tcp_rto_min(sk
);
724 /* no previous measure. */
725 srtt
= m
<< 3; /* take the measured time to be rtt */
726 tp
->mdev
= m
<< 1; /* make sure rto = 3*rtt */
727 tp
->mdev_max
= tp
->rttvar
= max(tp
->mdev
, tcp_rto_min(sk
));
728 tp
->rtt_seq
= tp
->snd_nxt
;
730 tp
->srtt
= max(1U, srtt
);
733 /* Set the sk_pacing_rate to allow proper sizing of TSO packets.
734 * Note: TCP stack does not yet implement pacing.
735 * FQ packet scheduler can be used to implement cheap but effective
736 * TCP pacing, to smooth the burst on large writes when packets
737 * in flight is significantly lower than cwnd (or rwin)
739 static void tcp_update_pacing_rate(struct sock
*sk
)
741 const struct tcp_sock
*tp
= tcp_sk(sk
);
744 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
745 rate
= (u64
)tp
->mss_cache
* 2 * (HZ
<< 3);
747 rate
*= max(tp
->snd_cwnd
, tp
->packets_out
);
749 /* Correction for small srtt and scheduling constraints.
750 * For small rtt, consider noise is too high, and use
751 * the minimal value (srtt = 1 -> 125 us for HZ=1000)
753 * We probably need usec resolution in the future.
754 * Note: This also takes care of possible srtt=0 case,
755 * when tcp_rtt_estimator() was not yet called.
757 if (tp
->srtt
> 8 + 2)
758 do_div(rate
, tp
->srtt
);
760 /* ACCESS_ONCE() is needed because sch_fq fetches sk_pacing_rate
761 * without any lock. We want to make sure compiler wont store
762 * intermediate values in this location.
764 ACCESS_ONCE(sk
->sk_pacing_rate
) = min_t(u64
, rate
,
765 sk
->sk_max_pacing_rate
);
768 /* Calculate rto without backoff. This is the second half of Van Jacobson's
769 * routine referred to above.
771 static void tcp_set_rto(struct sock
*sk
)
773 const struct tcp_sock
*tp
= tcp_sk(sk
);
774 /* Old crap is replaced with new one. 8)
777 * 1. If rtt variance happened to be less 50msec, it is hallucination.
778 * It cannot be less due to utterly erratic ACK generation made
779 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
780 * to do with delayed acks, because at cwnd>2 true delack timeout
781 * is invisible. Actually, Linux-2.4 also generates erratic
782 * ACKs in some circumstances.
784 inet_csk(sk
)->icsk_rto
= __tcp_set_rto(tp
);
786 /* 2. Fixups made earlier cannot be right.
787 * If we do not estimate RTO correctly without them,
788 * all the algo is pure shit and should be replaced
789 * with correct one. It is exactly, which we pretend to do.
792 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
793 * guarantees that rto is higher.
798 __u32
tcp_init_cwnd(const struct tcp_sock
*tp
, const struct dst_entry
*dst
)
800 __u32 cwnd
= (dst
? dst_metric(dst
, RTAX_INITCWND
) : 0);
803 cwnd
= TCP_INIT_CWND
;
804 return min_t(__u32
, cwnd
, tp
->snd_cwnd_clamp
);
808 * Packet counting of FACK is based on in-order assumptions, therefore TCP
809 * disables it when reordering is detected
811 void tcp_disable_fack(struct tcp_sock
*tp
)
813 /* RFC3517 uses different metric in lost marker => reset on change */
815 tp
->lost_skb_hint
= NULL
;
816 tp
->rx_opt
.sack_ok
&= ~TCP_FACK_ENABLED
;
819 /* Take a notice that peer is sending D-SACKs */
820 static void tcp_dsack_seen(struct tcp_sock
*tp
)
822 tp
->rx_opt
.sack_ok
|= TCP_DSACK_SEEN
;
825 static void tcp_update_reordering(struct sock
*sk
, const int metric
,
828 struct tcp_sock
*tp
= tcp_sk(sk
);
829 if (metric
> tp
->reordering
) {
832 tp
->reordering
= min(TCP_MAX_REORDERING
, metric
);
834 /* This exciting event is worth to be remembered. 8) */
836 mib_idx
= LINUX_MIB_TCPTSREORDER
;
837 else if (tcp_is_reno(tp
))
838 mib_idx
= LINUX_MIB_TCPRENOREORDER
;
839 else if (tcp_is_fack(tp
))
840 mib_idx
= LINUX_MIB_TCPFACKREORDER
;
842 mib_idx
= LINUX_MIB_TCPSACKREORDER
;
844 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
845 #if FASTRETRANS_DEBUG > 1
846 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
847 tp
->rx_opt
.sack_ok
, inet_csk(sk
)->icsk_ca_state
,
851 tp
->undo_marker
? tp
->undo_retrans
: 0);
853 tcp_disable_fack(tp
);
857 tcp_disable_early_retrans(tp
);
860 /* This must be called before lost_out is incremented */
861 static void tcp_verify_retransmit_hint(struct tcp_sock
*tp
, struct sk_buff
*skb
)
863 if ((tp
->retransmit_skb_hint
== NULL
) ||
864 before(TCP_SKB_CB(skb
)->seq
,
865 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
866 tp
->retransmit_skb_hint
= skb
;
869 after(TCP_SKB_CB(skb
)->end_seq
, tp
->retransmit_high
))
870 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
873 static void tcp_skb_mark_lost(struct tcp_sock
*tp
, struct sk_buff
*skb
)
875 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
876 tcp_verify_retransmit_hint(tp
, skb
);
878 tp
->lost_out
+= tcp_skb_pcount(skb
);
879 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
883 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock
*tp
,
886 tcp_verify_retransmit_hint(tp
, skb
);
888 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
889 tp
->lost_out
+= tcp_skb_pcount(skb
);
890 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
894 /* This procedure tags the retransmission queue when SACKs arrive.
896 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
897 * Packets in queue with these bits set are counted in variables
898 * sacked_out, retrans_out and lost_out, correspondingly.
900 * Valid combinations are:
901 * Tag InFlight Description
902 * 0 1 - orig segment is in flight.
903 * S 0 - nothing flies, orig reached receiver.
904 * L 0 - nothing flies, orig lost by net.
905 * R 2 - both orig and retransmit are in flight.
906 * L|R 1 - orig is lost, retransmit is in flight.
907 * S|R 1 - orig reached receiver, retrans is still in flight.
908 * (L|S|R is logically valid, it could occur when L|R is sacked,
909 * but it is equivalent to plain S and code short-curcuits it to S.
910 * L|S is logically invalid, it would mean -1 packet in flight 8))
912 * These 6 states form finite state machine, controlled by the following events:
913 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
914 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
915 * 3. Loss detection event of two flavors:
916 * A. Scoreboard estimator decided the packet is lost.
917 * A'. Reno "three dupacks" marks head of queue lost.
918 * A''. Its FACK modification, head until snd.fack is lost.
919 * B. SACK arrives sacking SND.NXT at the moment, when the
920 * segment was retransmitted.
921 * 4. D-SACK added new rule: D-SACK changes any tag to S.
923 * It is pleasant to note, that state diagram turns out to be commutative,
924 * so that we are allowed not to be bothered by order of our actions,
925 * when multiple events arrive simultaneously. (see the function below).
927 * Reordering detection.
928 * --------------------
929 * Reordering metric is maximal distance, which a packet can be displaced
930 * in packet stream. With SACKs we can estimate it:
932 * 1. SACK fills old hole and the corresponding segment was not
933 * ever retransmitted -> reordering. Alas, we cannot use it
934 * when segment was retransmitted.
935 * 2. The last flaw is solved with D-SACK. D-SACK arrives
936 * for retransmitted and already SACKed segment -> reordering..
937 * Both of these heuristics are not used in Loss state, when we cannot
938 * account for retransmits accurately.
940 * SACK block validation.
941 * ----------------------
943 * SACK block range validation checks that the received SACK block fits to
944 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
945 * Note that SND.UNA is not included to the range though being valid because
946 * it means that the receiver is rather inconsistent with itself reporting
947 * SACK reneging when it should advance SND.UNA. Such SACK block this is
948 * perfectly valid, however, in light of RFC2018 which explicitly states
949 * that "SACK block MUST reflect the newest segment. Even if the newest
950 * segment is going to be discarded ...", not that it looks very clever
951 * in case of head skb. Due to potentional receiver driven attacks, we
952 * choose to avoid immediate execution of a walk in write queue due to
953 * reneging and defer head skb's loss recovery to standard loss recovery
954 * procedure that will eventually trigger (nothing forbids us doing this).
956 * Implements also blockage to start_seq wrap-around. Problem lies in the
957 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
958 * there's no guarantee that it will be before snd_nxt (n). The problem
959 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
962 * <- outs wnd -> <- wrapzone ->
963 * u e n u_w e_w s n_w
965 * |<------------+------+----- TCP seqno space --------------+---------->|
966 * ...-- <2^31 ->| |<--------...
967 * ...---- >2^31 ------>| |<--------...
969 * Current code wouldn't be vulnerable but it's better still to discard such
970 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
971 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
972 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
973 * equal to the ideal case (infinite seqno space without wrap caused issues).
975 * With D-SACK the lower bound is extended to cover sequence space below
976 * SND.UNA down to undo_marker, which is the last point of interest. Yet
977 * again, D-SACK block must not to go across snd_una (for the same reason as
978 * for the normal SACK blocks, explained above). But there all simplicity
979 * ends, TCP might receive valid D-SACKs below that. As long as they reside
980 * fully below undo_marker they do not affect behavior in anyway and can
981 * therefore be safely ignored. In rare cases (which are more or less
982 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
983 * fragmentation and packet reordering past skb's retransmission. To consider
984 * them correctly, the acceptable range must be extended even more though
985 * the exact amount is rather hard to quantify. However, tp->max_window can
986 * be used as an exaggerated estimate.
988 static bool tcp_is_sackblock_valid(struct tcp_sock
*tp
, bool is_dsack
,
989 u32 start_seq
, u32 end_seq
)
991 /* Too far in future, or reversed (interpretation is ambiguous) */
992 if (after(end_seq
, tp
->snd_nxt
) || !before(start_seq
, end_seq
))
995 /* Nasty start_seq wrap-around check (see comments above) */
996 if (!before(start_seq
, tp
->snd_nxt
))
999 /* In outstanding window? ...This is valid exit for D-SACKs too.
1000 * start_seq == snd_una is non-sensical (see comments above)
1002 if (after(start_seq
, tp
->snd_una
))
1005 if (!is_dsack
|| !tp
->undo_marker
)
1008 /* ...Then it's D-SACK, and must reside below snd_una completely */
1009 if (after(end_seq
, tp
->snd_una
))
1012 if (!before(start_seq
, tp
->undo_marker
))
1016 if (!after(end_seq
, tp
->undo_marker
))
1019 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1020 * start_seq < undo_marker and end_seq >= undo_marker.
1022 return !before(start_seq
, end_seq
- tp
->max_window
);
1025 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1026 * Event "B". Later note: FACK people cheated me again 8), we have to account
1027 * for reordering! Ugly, but should help.
1029 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1030 * less than what is now known to be received by the other end (derived from
1031 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1032 * retransmitted skbs to avoid some costly processing per ACKs.
1034 static void tcp_mark_lost_retrans(struct sock
*sk
)
1036 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1037 struct tcp_sock
*tp
= tcp_sk(sk
);
1038 struct sk_buff
*skb
;
1040 u32 new_low_seq
= tp
->snd_nxt
;
1041 u32 received_upto
= tcp_highest_sack_seq(tp
);
1043 if (!tcp_is_fack(tp
) || !tp
->retrans_out
||
1044 !after(received_upto
, tp
->lost_retrans_low
) ||
1045 icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
1048 tcp_for_write_queue(skb
, sk
) {
1049 u32 ack_seq
= TCP_SKB_CB(skb
)->ack_seq
;
1051 if (skb
== tcp_send_head(sk
))
1053 if (cnt
== tp
->retrans_out
)
1055 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1058 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
))
1061 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1062 * constraint here (see above) but figuring out that at
1063 * least tp->reordering SACK blocks reside between ack_seq
1064 * and received_upto is not easy task to do cheaply with
1065 * the available datastructures.
1067 * Whether FACK should check here for tp->reordering segs
1068 * in-between one could argue for either way (it would be
1069 * rather simple to implement as we could count fack_count
1070 * during the walk and do tp->fackets_out - fack_count).
1072 if (after(received_upto
, ack_seq
)) {
1073 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1074 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1076 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
1077 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSTRETRANSMIT
);
1079 if (before(ack_seq
, new_low_seq
))
1080 new_low_seq
= ack_seq
;
1081 cnt
+= tcp_skb_pcount(skb
);
1085 if (tp
->retrans_out
)
1086 tp
->lost_retrans_low
= new_low_seq
;
1089 static bool tcp_check_dsack(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1090 struct tcp_sack_block_wire
*sp
, int num_sacks
,
1093 struct tcp_sock
*tp
= tcp_sk(sk
);
1094 u32 start_seq_0
= get_unaligned_be32(&sp
[0].start_seq
);
1095 u32 end_seq_0
= get_unaligned_be32(&sp
[0].end_seq
);
1096 bool dup_sack
= false;
1098 if (before(start_seq_0
, TCP_SKB_CB(ack_skb
)->ack_seq
)) {
1101 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPDSACKRECV
);
1102 } else if (num_sacks
> 1) {
1103 u32 end_seq_1
= get_unaligned_be32(&sp
[1].end_seq
);
1104 u32 start_seq_1
= get_unaligned_be32(&sp
[1].start_seq
);
1106 if (!after(end_seq_0
, end_seq_1
) &&
1107 !before(start_seq_0
, start_seq_1
)) {
1110 NET_INC_STATS_BH(sock_net(sk
),
1111 LINUX_MIB_TCPDSACKOFORECV
);
1115 /* D-SACK for already forgotten data... Do dumb counting. */
1116 if (dup_sack
&& tp
->undo_marker
&& tp
->undo_retrans
&&
1117 !after(end_seq_0
, prior_snd_una
) &&
1118 after(end_seq_0
, tp
->undo_marker
))
1124 struct tcp_sacktag_state
{
1128 s32 rtt
; /* RTT measured by SACKing never-retransmitted data */
1131 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1132 * the incoming SACK may not exactly match but we can find smaller MSS
1133 * aligned portion of it that matches. Therefore we might need to fragment
1134 * which may fail and creates some hassle (caller must handle error case
1137 * FIXME: this could be merged to shift decision code
1139 static int tcp_match_skb_to_sack(struct sock
*sk
, struct sk_buff
*skb
,
1140 u32 start_seq
, u32 end_seq
)
1144 unsigned int pkt_len
;
1147 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1148 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1150 if (tcp_skb_pcount(skb
) > 1 && !in_sack
&&
1151 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
1152 mss
= tcp_skb_mss(skb
);
1153 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1156 pkt_len
= start_seq
- TCP_SKB_CB(skb
)->seq
;
1160 pkt_len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1165 /* Round if necessary so that SACKs cover only full MSSes
1166 * and/or the remaining small portion (if present)
1168 if (pkt_len
> mss
) {
1169 unsigned int new_len
= (pkt_len
/ mss
) * mss
;
1170 if (!in_sack
&& new_len
< pkt_len
) {
1172 if (new_len
> skb
->len
)
1177 err
= tcp_fragment(sk
, skb
, pkt_len
, mss
);
1185 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1186 static u8
tcp_sacktag_one(struct sock
*sk
,
1187 struct tcp_sacktag_state
*state
, u8 sacked
,
1188 u32 start_seq
, u32 end_seq
,
1189 int dup_sack
, int pcount
, u32 xmit_time
)
1191 struct tcp_sock
*tp
= tcp_sk(sk
);
1192 int fack_count
= state
->fack_count
;
1194 /* Account D-SACK for retransmitted packet. */
1195 if (dup_sack
&& (sacked
& TCPCB_RETRANS
)) {
1196 if (tp
->undo_marker
&& tp
->undo_retrans
&&
1197 after(end_seq
, tp
->undo_marker
))
1199 if (sacked
& TCPCB_SACKED_ACKED
)
1200 state
->reord
= min(fack_count
, state
->reord
);
1203 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1204 if (!after(end_seq
, tp
->snd_una
))
1207 if (!(sacked
& TCPCB_SACKED_ACKED
)) {
1208 if (sacked
& TCPCB_SACKED_RETRANS
) {
1209 /* If the segment is not tagged as lost,
1210 * we do not clear RETRANS, believing
1211 * that retransmission is still in flight.
1213 if (sacked
& TCPCB_LOST
) {
1214 sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1215 tp
->lost_out
-= pcount
;
1216 tp
->retrans_out
-= pcount
;
1219 if (!(sacked
& TCPCB_RETRANS
)) {
1220 /* New sack for not retransmitted frame,
1221 * which was in hole. It is reordering.
1223 if (before(start_seq
,
1224 tcp_highest_sack_seq(tp
)))
1225 state
->reord
= min(fack_count
,
1227 if (!after(end_seq
, tp
->high_seq
))
1228 state
->flag
|= FLAG_ORIG_SACK_ACKED
;
1229 /* Pick the earliest sequence sacked for RTT */
1231 state
->rtt
= tcp_time_stamp
- xmit_time
;
1234 if (sacked
& TCPCB_LOST
) {
1235 sacked
&= ~TCPCB_LOST
;
1236 tp
->lost_out
-= pcount
;
1240 sacked
|= TCPCB_SACKED_ACKED
;
1241 state
->flag
|= FLAG_DATA_SACKED
;
1242 tp
->sacked_out
+= pcount
;
1244 fack_count
+= pcount
;
1246 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1247 if (!tcp_is_fack(tp
) && (tp
->lost_skb_hint
!= NULL
) &&
1248 before(start_seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
))
1249 tp
->lost_cnt_hint
+= pcount
;
1251 if (fack_count
> tp
->fackets_out
)
1252 tp
->fackets_out
= fack_count
;
1255 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1256 * frames and clear it. undo_retrans is decreased above, L|R frames
1257 * are accounted above as well.
1259 if (dup_sack
&& (sacked
& TCPCB_SACKED_RETRANS
)) {
1260 sacked
&= ~TCPCB_SACKED_RETRANS
;
1261 tp
->retrans_out
-= pcount
;
1267 /* Shift newly-SACKed bytes from this skb to the immediately previous
1268 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1270 static bool tcp_shifted_skb(struct sock
*sk
, struct sk_buff
*skb
,
1271 struct tcp_sacktag_state
*state
,
1272 unsigned int pcount
, int shifted
, int mss
,
1275 struct tcp_sock
*tp
= tcp_sk(sk
);
1276 struct sk_buff
*prev
= tcp_write_queue_prev(sk
, skb
);
1277 u32 start_seq
= TCP_SKB_CB(skb
)->seq
; /* start of newly-SACKed */
1278 u32 end_seq
= start_seq
+ shifted
; /* end of newly-SACKed */
1282 /* Adjust counters and hints for the newly sacked sequence
1283 * range but discard the return value since prev is already
1284 * marked. We must tag the range first because the seq
1285 * advancement below implicitly advances
1286 * tcp_highest_sack_seq() when skb is highest_sack.
1288 tcp_sacktag_one(sk
, state
, TCP_SKB_CB(skb
)->sacked
,
1289 start_seq
, end_seq
, dup_sack
, pcount
,
1290 TCP_SKB_CB(skb
)->when
);
1292 if (skb
== tp
->lost_skb_hint
)
1293 tp
->lost_cnt_hint
+= pcount
;
1295 TCP_SKB_CB(prev
)->end_seq
+= shifted
;
1296 TCP_SKB_CB(skb
)->seq
+= shifted
;
1298 skb_shinfo(prev
)->gso_segs
+= pcount
;
1299 BUG_ON(skb_shinfo(skb
)->gso_segs
< pcount
);
1300 skb_shinfo(skb
)->gso_segs
-= pcount
;
1302 /* When we're adding to gso_segs == 1, gso_size will be zero,
1303 * in theory this shouldn't be necessary but as long as DSACK
1304 * code can come after this skb later on it's better to keep
1305 * setting gso_size to something.
1307 if (!skb_shinfo(prev
)->gso_size
) {
1308 skb_shinfo(prev
)->gso_size
= mss
;
1309 skb_shinfo(prev
)->gso_type
= sk
->sk_gso_type
;
1312 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1313 if (skb_shinfo(skb
)->gso_segs
<= 1) {
1314 skb_shinfo(skb
)->gso_size
= 0;
1315 skb_shinfo(skb
)->gso_type
= 0;
1318 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1319 TCP_SKB_CB(prev
)->sacked
|= (TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
);
1322 BUG_ON(!tcp_skb_pcount(skb
));
1323 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKSHIFTED
);
1327 /* Whole SKB was eaten :-) */
1329 if (skb
== tp
->retransmit_skb_hint
)
1330 tp
->retransmit_skb_hint
= prev
;
1331 if (skb
== tp
->lost_skb_hint
) {
1332 tp
->lost_skb_hint
= prev
;
1333 tp
->lost_cnt_hint
-= tcp_skb_pcount(prev
);
1336 TCP_SKB_CB(prev
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
;
1337 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
1338 TCP_SKB_CB(prev
)->end_seq
++;
1340 if (skb
== tcp_highest_sack(sk
))
1341 tcp_advance_highest_sack(sk
, skb
);
1343 tcp_unlink_write_queue(skb
, sk
);
1344 sk_wmem_free_skb(sk
, skb
);
1346 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKMERGED
);
1351 /* I wish gso_size would have a bit more sane initialization than
1352 * something-or-zero which complicates things
1354 static int tcp_skb_seglen(const struct sk_buff
*skb
)
1356 return tcp_skb_pcount(skb
) == 1 ? skb
->len
: tcp_skb_mss(skb
);
1359 /* Shifting pages past head area doesn't work */
1360 static int skb_can_shift(const struct sk_buff
*skb
)
1362 return !skb_headlen(skb
) && skb_is_nonlinear(skb
);
1365 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1368 static struct sk_buff
*tcp_shift_skb_data(struct sock
*sk
, struct sk_buff
*skb
,
1369 struct tcp_sacktag_state
*state
,
1370 u32 start_seq
, u32 end_seq
,
1373 struct tcp_sock
*tp
= tcp_sk(sk
);
1374 struct sk_buff
*prev
;
1380 if (!sk_can_gso(sk
))
1383 /* Normally R but no L won't result in plain S */
1385 (TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_RETRANS
)) == TCPCB_SACKED_RETRANS
)
1387 if (!skb_can_shift(skb
))
1389 /* This frame is about to be dropped (was ACKed). */
1390 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1393 /* Can only happen with delayed DSACK + discard craziness */
1394 if (unlikely(skb
== tcp_write_queue_head(sk
)))
1396 prev
= tcp_write_queue_prev(sk
, skb
);
1398 if ((TCP_SKB_CB(prev
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
)
1401 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1402 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1406 pcount
= tcp_skb_pcount(skb
);
1407 mss
= tcp_skb_seglen(skb
);
1409 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1410 * drop this restriction as unnecessary
1412 if (mss
!= tcp_skb_seglen(prev
))
1415 if (!after(TCP_SKB_CB(skb
)->end_seq
, start_seq
))
1417 /* CHECKME: This is non-MSS split case only?, this will
1418 * cause skipped skbs due to advancing loop btw, original
1419 * has that feature too
1421 if (tcp_skb_pcount(skb
) <= 1)
1424 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1426 /* TODO: head merge to next could be attempted here
1427 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1428 * though it might not be worth of the additional hassle
1430 * ...we can probably just fallback to what was done
1431 * previously. We could try merging non-SACKed ones
1432 * as well but it probably isn't going to buy off
1433 * because later SACKs might again split them, and
1434 * it would make skb timestamp tracking considerably
1440 len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1442 BUG_ON(len
> skb
->len
);
1444 /* MSS boundaries should be honoured or else pcount will
1445 * severely break even though it makes things bit trickier.
1446 * Optimize common case to avoid most of the divides
1448 mss
= tcp_skb_mss(skb
);
1450 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1451 * drop this restriction as unnecessary
1453 if (mss
!= tcp_skb_seglen(prev
))
1458 } else if (len
< mss
) {
1466 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1467 if (!after(TCP_SKB_CB(skb
)->seq
+ len
, tp
->snd_una
))
1470 if (!skb_shift(prev
, skb
, len
))
1472 if (!tcp_shifted_skb(sk
, skb
, state
, pcount
, len
, mss
, dup_sack
))
1475 /* Hole filled allows collapsing with the next as well, this is very
1476 * useful when hole on every nth skb pattern happens
1478 if (prev
== tcp_write_queue_tail(sk
))
1480 skb
= tcp_write_queue_next(sk
, prev
);
1482 if (!skb_can_shift(skb
) ||
1483 (skb
== tcp_send_head(sk
)) ||
1484 ((TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
) ||
1485 (mss
!= tcp_skb_seglen(skb
)))
1489 if (skb_shift(prev
, skb
, len
)) {
1490 pcount
+= tcp_skb_pcount(skb
);
1491 tcp_shifted_skb(sk
, skb
, state
, tcp_skb_pcount(skb
), len
, mss
, 0);
1495 state
->fack_count
+= pcount
;
1502 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKSHIFTFALLBACK
);
1506 static struct sk_buff
*tcp_sacktag_walk(struct sk_buff
*skb
, struct sock
*sk
,
1507 struct tcp_sack_block
*next_dup
,
1508 struct tcp_sacktag_state
*state
,
1509 u32 start_seq
, u32 end_seq
,
1512 struct tcp_sock
*tp
= tcp_sk(sk
);
1513 struct sk_buff
*tmp
;
1515 tcp_for_write_queue_from(skb
, sk
) {
1517 bool dup_sack
= dup_sack_in
;
1519 if (skb
== tcp_send_head(sk
))
1522 /* queue is in-order => we can short-circuit the walk early */
1523 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1526 if ((next_dup
!= NULL
) &&
1527 before(TCP_SKB_CB(skb
)->seq
, next_dup
->end_seq
)) {
1528 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1529 next_dup
->start_seq
,
1535 /* skb reference here is a bit tricky to get right, since
1536 * shifting can eat and free both this skb and the next,
1537 * so not even _safe variant of the loop is enough.
1540 tmp
= tcp_shift_skb_data(sk
, skb
, state
,
1541 start_seq
, end_seq
, dup_sack
);
1550 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1556 if (unlikely(in_sack
< 0))
1560 TCP_SKB_CB(skb
)->sacked
=
1563 TCP_SKB_CB(skb
)->sacked
,
1564 TCP_SKB_CB(skb
)->seq
,
1565 TCP_SKB_CB(skb
)->end_seq
,
1567 tcp_skb_pcount(skb
),
1568 TCP_SKB_CB(skb
)->when
);
1570 if (!before(TCP_SKB_CB(skb
)->seq
,
1571 tcp_highest_sack_seq(tp
)))
1572 tcp_advance_highest_sack(sk
, skb
);
1575 state
->fack_count
+= tcp_skb_pcount(skb
);
1580 /* Avoid all extra work that is being done by sacktag while walking in
1583 static struct sk_buff
*tcp_sacktag_skip(struct sk_buff
*skb
, struct sock
*sk
,
1584 struct tcp_sacktag_state
*state
,
1587 tcp_for_write_queue_from(skb
, sk
) {
1588 if (skb
== tcp_send_head(sk
))
1591 if (after(TCP_SKB_CB(skb
)->end_seq
, skip_to_seq
))
1594 state
->fack_count
+= tcp_skb_pcount(skb
);
1599 static struct sk_buff
*tcp_maybe_skipping_dsack(struct sk_buff
*skb
,
1601 struct tcp_sack_block
*next_dup
,
1602 struct tcp_sacktag_state
*state
,
1605 if (next_dup
== NULL
)
1608 if (before(next_dup
->start_seq
, skip_to_seq
)) {
1609 skb
= tcp_sacktag_skip(skb
, sk
, state
, next_dup
->start_seq
);
1610 skb
= tcp_sacktag_walk(skb
, sk
, NULL
, state
,
1611 next_dup
->start_seq
, next_dup
->end_seq
,
1618 static int tcp_sack_cache_ok(const struct tcp_sock
*tp
, const struct tcp_sack_block
*cache
)
1620 return cache
< tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1624 tcp_sacktag_write_queue(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1625 u32 prior_snd_una
, s32
*sack_rtt
)
1627 struct tcp_sock
*tp
= tcp_sk(sk
);
1628 const unsigned char *ptr
= (skb_transport_header(ack_skb
) +
1629 TCP_SKB_CB(ack_skb
)->sacked
);
1630 struct tcp_sack_block_wire
*sp_wire
= (struct tcp_sack_block_wire
*)(ptr
+2);
1631 struct tcp_sack_block sp
[TCP_NUM_SACKS
];
1632 struct tcp_sack_block
*cache
;
1633 struct tcp_sacktag_state state
;
1634 struct sk_buff
*skb
;
1635 int num_sacks
= min(TCP_NUM_SACKS
, (ptr
[1] - TCPOLEN_SACK_BASE
) >> 3);
1637 bool found_dup_sack
= false;
1639 int first_sack_index
;
1642 state
.reord
= tp
->packets_out
;
1645 if (!tp
->sacked_out
) {
1646 if (WARN_ON(tp
->fackets_out
))
1647 tp
->fackets_out
= 0;
1648 tcp_highest_sack_reset(sk
);
1651 found_dup_sack
= tcp_check_dsack(sk
, ack_skb
, sp_wire
,
1652 num_sacks
, prior_snd_una
);
1654 state
.flag
|= FLAG_DSACKING_ACK
;
1656 /* Eliminate too old ACKs, but take into
1657 * account more or less fresh ones, they can
1658 * contain valid SACK info.
1660 if (before(TCP_SKB_CB(ack_skb
)->ack_seq
, prior_snd_una
- tp
->max_window
))
1663 if (!tp
->packets_out
)
1667 first_sack_index
= 0;
1668 for (i
= 0; i
< num_sacks
; i
++) {
1669 bool dup_sack
= !i
&& found_dup_sack
;
1671 sp
[used_sacks
].start_seq
= get_unaligned_be32(&sp_wire
[i
].start_seq
);
1672 sp
[used_sacks
].end_seq
= get_unaligned_be32(&sp_wire
[i
].end_seq
);
1674 if (!tcp_is_sackblock_valid(tp
, dup_sack
,
1675 sp
[used_sacks
].start_seq
,
1676 sp
[used_sacks
].end_seq
)) {
1680 if (!tp
->undo_marker
)
1681 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDNOUNDO
;
1683 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDOLD
;
1685 /* Don't count olds caused by ACK reordering */
1686 if ((TCP_SKB_CB(ack_skb
)->ack_seq
!= tp
->snd_una
) &&
1687 !after(sp
[used_sacks
].end_seq
, tp
->snd_una
))
1689 mib_idx
= LINUX_MIB_TCPSACKDISCARD
;
1692 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
1694 first_sack_index
= -1;
1698 /* Ignore very old stuff early */
1699 if (!after(sp
[used_sacks
].end_seq
, prior_snd_una
))
1705 /* order SACK blocks to allow in order walk of the retrans queue */
1706 for (i
= used_sacks
- 1; i
> 0; i
--) {
1707 for (j
= 0; j
< i
; j
++) {
1708 if (after(sp
[j
].start_seq
, sp
[j
+ 1].start_seq
)) {
1709 swap(sp
[j
], sp
[j
+ 1]);
1711 /* Track where the first SACK block goes to */
1712 if (j
== first_sack_index
)
1713 first_sack_index
= j
+ 1;
1718 skb
= tcp_write_queue_head(sk
);
1719 state
.fack_count
= 0;
1722 if (!tp
->sacked_out
) {
1723 /* It's already past, so skip checking against it */
1724 cache
= tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1726 cache
= tp
->recv_sack_cache
;
1727 /* Skip empty blocks in at head of the cache */
1728 while (tcp_sack_cache_ok(tp
, cache
) && !cache
->start_seq
&&
1733 while (i
< used_sacks
) {
1734 u32 start_seq
= sp
[i
].start_seq
;
1735 u32 end_seq
= sp
[i
].end_seq
;
1736 bool dup_sack
= (found_dup_sack
&& (i
== first_sack_index
));
1737 struct tcp_sack_block
*next_dup
= NULL
;
1739 if (found_dup_sack
&& ((i
+ 1) == first_sack_index
))
1740 next_dup
= &sp
[i
+ 1];
1742 /* Skip too early cached blocks */
1743 while (tcp_sack_cache_ok(tp
, cache
) &&
1744 !before(start_seq
, cache
->end_seq
))
1747 /* Can skip some work by looking recv_sack_cache? */
1748 if (tcp_sack_cache_ok(tp
, cache
) && !dup_sack
&&
1749 after(end_seq
, cache
->start_seq
)) {
1752 if (before(start_seq
, cache
->start_seq
)) {
1753 skb
= tcp_sacktag_skip(skb
, sk
, &state
,
1755 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
,
1762 /* Rest of the block already fully processed? */
1763 if (!after(end_seq
, cache
->end_seq
))
1766 skb
= tcp_maybe_skipping_dsack(skb
, sk
, next_dup
,
1770 /* ...tail remains todo... */
1771 if (tcp_highest_sack_seq(tp
) == cache
->end_seq
) {
1772 /* ...but better entrypoint exists! */
1773 skb
= tcp_highest_sack(sk
);
1776 state
.fack_count
= tp
->fackets_out
;
1781 skb
= tcp_sacktag_skip(skb
, sk
, &state
, cache
->end_seq
);
1782 /* Check overlap against next cached too (past this one already) */
1787 if (!before(start_seq
, tcp_highest_sack_seq(tp
))) {
1788 skb
= tcp_highest_sack(sk
);
1791 state
.fack_count
= tp
->fackets_out
;
1793 skb
= tcp_sacktag_skip(skb
, sk
, &state
, start_seq
);
1796 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
, &state
,
1797 start_seq
, end_seq
, dup_sack
);
1803 /* Clear the head of the cache sack blocks so we can skip it next time */
1804 for (i
= 0; i
< ARRAY_SIZE(tp
->recv_sack_cache
) - used_sacks
; i
++) {
1805 tp
->recv_sack_cache
[i
].start_seq
= 0;
1806 tp
->recv_sack_cache
[i
].end_seq
= 0;
1808 for (j
= 0; j
< used_sacks
; j
++)
1809 tp
->recv_sack_cache
[i
++] = sp
[j
];
1811 tcp_mark_lost_retrans(sk
);
1813 tcp_verify_left_out(tp
);
1815 if ((state
.reord
< tp
->fackets_out
) &&
1816 ((inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Loss
) || tp
->undo_marker
))
1817 tcp_update_reordering(sk
, tp
->fackets_out
- state
.reord
, 0);
1821 #if FASTRETRANS_DEBUG > 0
1822 WARN_ON((int)tp
->sacked_out
< 0);
1823 WARN_ON((int)tp
->lost_out
< 0);
1824 WARN_ON((int)tp
->retrans_out
< 0);
1825 WARN_ON((int)tcp_packets_in_flight(tp
) < 0);
1827 *sack_rtt
= state
.rtt
;
1831 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1832 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1834 static bool tcp_limit_reno_sacked(struct tcp_sock
*tp
)
1838 holes
= max(tp
->lost_out
, 1U);
1839 holes
= min(holes
, tp
->packets_out
);
1841 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1842 tp
->sacked_out
= tp
->packets_out
- holes
;
1848 /* If we receive more dupacks than we expected counting segments
1849 * in assumption of absent reordering, interpret this as reordering.
1850 * The only another reason could be bug in receiver TCP.
1852 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1854 struct tcp_sock
*tp
= tcp_sk(sk
);
1855 if (tcp_limit_reno_sacked(tp
))
1856 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1859 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1861 static void tcp_add_reno_sack(struct sock
*sk
)
1863 struct tcp_sock
*tp
= tcp_sk(sk
);
1865 tcp_check_reno_reordering(sk
, 0);
1866 tcp_verify_left_out(tp
);
1869 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1871 static void tcp_remove_reno_sacks(struct sock
*sk
, int acked
)
1873 struct tcp_sock
*tp
= tcp_sk(sk
);
1876 /* One ACK acked hole. The rest eat duplicate ACKs. */
1877 if (acked
- 1 >= tp
->sacked_out
)
1880 tp
->sacked_out
-= acked
- 1;
1882 tcp_check_reno_reordering(sk
, acked
);
1883 tcp_verify_left_out(tp
);
1886 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1891 static void tcp_clear_retrans_partial(struct tcp_sock
*tp
)
1893 tp
->retrans_out
= 0;
1896 tp
->undo_marker
= 0;
1897 tp
->undo_retrans
= 0;
1900 void tcp_clear_retrans(struct tcp_sock
*tp
)
1902 tcp_clear_retrans_partial(tp
);
1904 tp
->fackets_out
= 0;
1908 /* Enter Loss state. If "how" is not zero, forget all SACK information
1909 * and reset tags completely, otherwise preserve SACKs. If receiver
1910 * dropped its ofo queue, we will know this due to reneging detection.
1912 void tcp_enter_loss(struct sock
*sk
, int how
)
1914 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1915 struct tcp_sock
*tp
= tcp_sk(sk
);
1916 struct sk_buff
*skb
;
1917 bool new_recovery
= false;
1919 /* Reduce ssthresh if it has not yet been made inside this window. */
1920 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
||
1921 !after(tp
->high_seq
, tp
->snd_una
) ||
1922 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
1923 new_recovery
= true;
1924 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1925 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1926 tcp_ca_event(sk
, CA_EVENT_LOSS
);
1929 tp
->snd_cwnd_cnt
= 0;
1930 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1932 tcp_clear_retrans_partial(tp
);
1934 if (tcp_is_reno(tp
))
1935 tcp_reset_reno_sack(tp
);
1937 tp
->undo_marker
= tp
->snd_una
;
1940 tp
->fackets_out
= 0;
1942 tcp_clear_all_retrans_hints(tp
);
1944 tcp_for_write_queue(skb
, sk
) {
1945 if (skb
== tcp_send_head(sk
))
1948 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
1949 tp
->undo_marker
= 0;
1951 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
1952 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) || how
) {
1953 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
1954 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1955 tp
->lost_out
+= tcp_skb_pcount(skb
);
1956 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
1959 tcp_verify_left_out(tp
);
1961 /* Timeout in disordered state after receiving substantial DUPACKs
1962 * suggests that the degree of reordering is over-estimated.
1964 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
&&
1965 tp
->sacked_out
>= sysctl_tcp_reordering
)
1966 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1967 sysctl_tcp_reordering
);
1968 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1969 tp
->high_seq
= tp
->snd_nxt
;
1970 TCP_ECN_queue_cwr(tp
);
1972 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1973 * loss recovery is underway except recurring timeout(s) on
1974 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1976 tp
->frto
= sysctl_tcp_frto
&&
1977 (new_recovery
|| icsk
->icsk_retransmits
) &&
1978 !inet_csk(sk
)->icsk_mtup
.probe_size
;
1981 /* If ACK arrived pointing to a remembered SACK, it means that our
1982 * remembered SACKs do not reflect real state of receiver i.e.
1983 * receiver _host_ is heavily congested (or buggy).
1985 * Do processing similar to RTO timeout.
1987 static bool tcp_check_sack_reneging(struct sock
*sk
, int flag
)
1989 if (flag
& FLAG_SACK_RENEGING
) {
1990 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1991 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSACKRENEGING
);
1993 tcp_enter_loss(sk
, 1);
1994 icsk
->icsk_retransmits
++;
1995 tcp_retransmit_skb(sk
, tcp_write_queue_head(sk
));
1996 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1997 icsk
->icsk_rto
, TCP_RTO_MAX
);
2003 static inline int tcp_fackets_out(const struct tcp_sock
*tp
)
2005 return tcp_is_reno(tp
) ? tp
->sacked_out
+ 1 : tp
->fackets_out
;
2008 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2009 * counter when SACK is enabled (without SACK, sacked_out is used for
2012 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2013 * segments up to the highest received SACK block so far and holes in
2016 * With reordering, holes may still be in flight, so RFC3517 recovery
2017 * uses pure sacked_out (total number of SACKed segments) even though
2018 * it violates the RFC that uses duplicate ACKs, often these are equal
2019 * but when e.g. out-of-window ACKs or packet duplication occurs,
2020 * they differ. Since neither occurs due to loss, TCP should really
2023 static inline int tcp_dupack_heuristics(const struct tcp_sock
*tp
)
2025 return tcp_is_fack(tp
) ? tp
->fackets_out
: tp
->sacked_out
+ 1;
2028 static bool tcp_pause_early_retransmit(struct sock
*sk
, int flag
)
2030 struct tcp_sock
*tp
= tcp_sk(sk
);
2031 unsigned long delay
;
2033 /* Delay early retransmit and entering fast recovery for
2034 * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
2035 * available, or RTO is scheduled to fire first.
2037 if (sysctl_tcp_early_retrans
< 2 || sysctl_tcp_early_retrans
> 3 ||
2038 (flag
& FLAG_ECE
) || !tp
->srtt
)
2041 delay
= max_t(unsigned long, (tp
->srtt
>> 5), msecs_to_jiffies(2));
2042 if (!time_after(inet_csk(sk
)->icsk_timeout
, (jiffies
+ delay
)))
2045 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_EARLY_RETRANS
, delay
,
2050 /* Linux NewReno/SACK/FACK/ECN state machine.
2051 * --------------------------------------
2053 * "Open" Normal state, no dubious events, fast path.
2054 * "Disorder" In all the respects it is "Open",
2055 * but requires a bit more attention. It is entered when
2056 * we see some SACKs or dupacks. It is split of "Open"
2057 * mainly to move some processing from fast path to slow one.
2058 * "CWR" CWND was reduced due to some Congestion Notification event.
2059 * It can be ECN, ICMP source quench, local device congestion.
2060 * "Recovery" CWND was reduced, we are fast-retransmitting.
2061 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2063 * tcp_fastretrans_alert() is entered:
2064 * - each incoming ACK, if state is not "Open"
2065 * - when arrived ACK is unusual, namely:
2070 * Counting packets in flight is pretty simple.
2072 * in_flight = packets_out - left_out + retrans_out
2074 * packets_out is SND.NXT-SND.UNA counted in packets.
2076 * retrans_out is number of retransmitted segments.
2078 * left_out is number of segments left network, but not ACKed yet.
2080 * left_out = sacked_out + lost_out
2082 * sacked_out: Packets, which arrived to receiver out of order
2083 * and hence not ACKed. With SACKs this number is simply
2084 * amount of SACKed data. Even without SACKs
2085 * it is easy to give pretty reliable estimate of this number,
2086 * counting duplicate ACKs.
2088 * lost_out: Packets lost by network. TCP has no explicit
2089 * "loss notification" feedback from network (for now).
2090 * It means that this number can be only _guessed_.
2091 * Actually, it is the heuristics to predict lossage that
2092 * distinguishes different algorithms.
2094 * F.e. after RTO, when all the queue is considered as lost,
2095 * lost_out = packets_out and in_flight = retrans_out.
2097 * Essentially, we have now two algorithms counting
2100 * FACK: It is the simplest heuristics. As soon as we decided
2101 * that something is lost, we decide that _all_ not SACKed
2102 * packets until the most forward SACK are lost. I.e.
2103 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2104 * It is absolutely correct estimate, if network does not reorder
2105 * packets. And it loses any connection to reality when reordering
2106 * takes place. We use FACK by default until reordering
2107 * is suspected on the path to this destination.
2109 * NewReno: when Recovery is entered, we assume that one segment
2110 * is lost (classic Reno). While we are in Recovery and
2111 * a partial ACK arrives, we assume that one more packet
2112 * is lost (NewReno). This heuristics are the same in NewReno
2115 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2116 * deflation etc. CWND is real congestion window, never inflated, changes
2117 * only according to classic VJ rules.
2119 * Really tricky (and requiring careful tuning) part of algorithm
2120 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2121 * The first determines the moment _when_ we should reduce CWND and,
2122 * hence, slow down forward transmission. In fact, it determines the moment
2123 * when we decide that hole is caused by loss, rather than by a reorder.
2125 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2126 * holes, caused by lost packets.
2128 * And the most logically complicated part of algorithm is undo
2129 * heuristics. We detect false retransmits due to both too early
2130 * fast retransmit (reordering) and underestimated RTO, analyzing
2131 * timestamps and D-SACKs. When we detect that some segments were
2132 * retransmitted by mistake and CWND reduction was wrong, we undo
2133 * window reduction and abort recovery phase. This logic is hidden
2134 * inside several functions named tcp_try_undo_<something>.
2137 /* This function decides, when we should leave Disordered state
2138 * and enter Recovery phase, reducing congestion window.
2140 * Main question: may we further continue forward transmission
2141 * with the same cwnd?
2143 static bool tcp_time_to_recover(struct sock
*sk
, int flag
)
2145 struct tcp_sock
*tp
= tcp_sk(sk
);
2148 /* Trick#1: The loss is proven. */
2152 /* Not-A-Trick#2 : Classic rule... */
2153 if (tcp_dupack_heuristics(tp
) > tp
->reordering
)
2156 /* Trick#4: It is still not OK... But will it be useful to delay
2159 packets_out
= tp
->packets_out
;
2160 if (packets_out
<= tp
->reordering
&&
2161 tp
->sacked_out
>= max_t(__u32
, packets_out
/2, sysctl_tcp_reordering
) &&
2162 !tcp_may_send_now(sk
)) {
2163 /* We have nothing to send. This connection is limited
2164 * either by receiver window or by application.
2169 /* If a thin stream is detected, retransmit after first
2170 * received dupack. Employ only if SACK is supported in order
2171 * to avoid possible corner-case series of spurious retransmissions
2172 * Use only if there are no unsent data.
2174 if ((tp
->thin_dupack
|| sysctl_tcp_thin_dupack
) &&
2175 tcp_stream_is_thin(tp
) && tcp_dupack_heuristics(tp
) > 1 &&
2176 tcp_is_sack(tp
) && !tcp_send_head(sk
))
2179 /* Trick#6: TCP early retransmit, per RFC5827. To avoid spurious
2180 * retransmissions due to small network reorderings, we implement
2181 * Mitigation A.3 in the RFC and delay the retransmission for a short
2182 * interval if appropriate.
2184 if (tp
->do_early_retrans
&& !tp
->retrans_out
&& tp
->sacked_out
&&
2185 (tp
->packets_out
>= (tp
->sacked_out
+ 1) && tp
->packets_out
< 4) &&
2186 !tcp_may_send_now(sk
))
2187 return !tcp_pause_early_retransmit(sk
, flag
);
2192 /* Detect loss in event "A" above by marking head of queue up as lost.
2193 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2194 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2195 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2196 * the maximum SACKed segments to pass before reaching this limit.
2198 static void tcp_mark_head_lost(struct sock
*sk
, int packets
, int mark_head
)
2200 struct tcp_sock
*tp
= tcp_sk(sk
);
2201 struct sk_buff
*skb
;
2205 /* Use SACK to deduce losses of new sequences sent during recovery */
2206 const u32 loss_high
= tcp_is_sack(tp
) ? tp
->snd_nxt
: tp
->high_seq
;
2208 WARN_ON(packets
> tp
->packets_out
);
2209 if (tp
->lost_skb_hint
) {
2210 skb
= tp
->lost_skb_hint
;
2211 cnt
= tp
->lost_cnt_hint
;
2212 /* Head already handled? */
2213 if (mark_head
&& skb
!= tcp_write_queue_head(sk
))
2216 skb
= tcp_write_queue_head(sk
);
2220 tcp_for_write_queue_from(skb
, sk
) {
2221 if (skb
== tcp_send_head(sk
))
2223 /* TODO: do this better */
2224 /* this is not the most efficient way to do this... */
2225 tp
->lost_skb_hint
= skb
;
2226 tp
->lost_cnt_hint
= cnt
;
2228 if (after(TCP_SKB_CB(skb
)->end_seq
, loss_high
))
2232 if (tcp_is_fack(tp
) || tcp_is_reno(tp
) ||
2233 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2234 cnt
+= tcp_skb_pcount(skb
);
2236 if (cnt
> packets
) {
2237 if ((tcp_is_sack(tp
) && !tcp_is_fack(tp
)) ||
2238 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
) ||
2239 (oldcnt
>= packets
))
2242 mss
= skb_shinfo(skb
)->gso_size
;
2243 err
= tcp_fragment(sk
, skb
, (packets
- oldcnt
) * mss
, mss
);
2249 tcp_skb_mark_lost(tp
, skb
);
2254 tcp_verify_left_out(tp
);
2257 /* Account newly detected lost packet(s) */
2259 static void tcp_update_scoreboard(struct sock
*sk
, int fast_rexmit
)
2261 struct tcp_sock
*tp
= tcp_sk(sk
);
2263 if (tcp_is_reno(tp
)) {
2264 tcp_mark_head_lost(sk
, 1, 1);
2265 } else if (tcp_is_fack(tp
)) {
2266 int lost
= tp
->fackets_out
- tp
->reordering
;
2269 tcp_mark_head_lost(sk
, lost
, 0);
2271 int sacked_upto
= tp
->sacked_out
- tp
->reordering
;
2272 if (sacked_upto
>= 0)
2273 tcp_mark_head_lost(sk
, sacked_upto
, 0);
2274 else if (fast_rexmit
)
2275 tcp_mark_head_lost(sk
, 1, 1);
2279 /* CWND moderation, preventing bursts due to too big ACKs
2280 * in dubious situations.
2282 static inline void tcp_moderate_cwnd(struct tcp_sock
*tp
)
2284 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
2285 tcp_packets_in_flight(tp
) + tcp_max_burst(tp
));
2286 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2289 /* Nothing was retransmitted or returned timestamp is less
2290 * than timestamp of the first retransmission.
2292 static inline bool tcp_packet_delayed(const struct tcp_sock
*tp
)
2294 return !tp
->retrans_stamp
||
2295 (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2296 before(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
));
2299 /* Undo procedures. */
2301 #if FASTRETRANS_DEBUG > 1
2302 static void DBGUNDO(struct sock
*sk
, const char *msg
)
2304 struct tcp_sock
*tp
= tcp_sk(sk
);
2305 struct inet_sock
*inet
= inet_sk(sk
);
2307 if (sk
->sk_family
== AF_INET
) {
2308 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2310 &inet
->inet_daddr
, ntohs(inet
->inet_dport
),
2311 tp
->snd_cwnd
, tcp_left_out(tp
),
2312 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2315 #if IS_ENABLED(CONFIG_IPV6)
2316 else if (sk
->sk_family
== AF_INET6
) {
2317 struct ipv6_pinfo
*np
= inet6_sk(sk
);
2318 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2320 &np
->daddr
, ntohs(inet
->inet_dport
),
2321 tp
->snd_cwnd
, tcp_left_out(tp
),
2322 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2328 #define DBGUNDO(x...) do { } while (0)
2331 static void tcp_undo_cwnd_reduction(struct sock
*sk
, bool unmark_loss
)
2333 struct tcp_sock
*tp
= tcp_sk(sk
);
2336 struct sk_buff
*skb
;
2338 tcp_for_write_queue(skb
, sk
) {
2339 if (skb
== tcp_send_head(sk
))
2341 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2344 tcp_clear_all_retrans_hints(tp
);
2347 if (tp
->prior_ssthresh
) {
2348 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2350 if (icsk
->icsk_ca_ops
->undo_cwnd
)
2351 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
2353 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
<< 1);
2355 if (tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
2356 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
2357 TCP_ECN_withdraw_cwr(tp
);
2360 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2362 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2363 tp
->undo_marker
= 0;
2366 static inline bool tcp_may_undo(const struct tcp_sock
*tp
)
2368 return tp
->undo_marker
&& (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
2371 /* People celebrate: "We love our President!" */
2372 static bool tcp_try_undo_recovery(struct sock
*sk
)
2374 struct tcp_sock
*tp
= tcp_sk(sk
);
2376 if (tcp_may_undo(tp
)) {
2379 /* Happy end! We did not retransmit anything
2380 * or our original transmission succeeded.
2382 DBGUNDO(sk
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
2383 tcp_undo_cwnd_reduction(sk
, false);
2384 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
2385 mib_idx
= LINUX_MIB_TCPLOSSUNDO
;
2387 mib_idx
= LINUX_MIB_TCPFULLUNDO
;
2389 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2391 if (tp
->snd_una
== tp
->high_seq
&& tcp_is_reno(tp
)) {
2392 /* Hold old state until something *above* high_seq
2393 * is ACKed. For Reno it is MUST to prevent false
2394 * fast retransmits (RFC2582). SACK TCP is safe. */
2395 tcp_moderate_cwnd(tp
);
2398 tcp_set_ca_state(sk
, TCP_CA_Open
);
2402 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2403 static bool tcp_try_undo_dsack(struct sock
*sk
)
2405 struct tcp_sock
*tp
= tcp_sk(sk
);
2407 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
2408 DBGUNDO(sk
, "D-SACK");
2409 tcp_undo_cwnd_reduction(sk
, false);
2410 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPDSACKUNDO
);
2416 /* We can clear retrans_stamp when there are no retransmissions in the
2417 * window. It would seem that it is trivially available for us in
2418 * tp->retrans_out, however, that kind of assumptions doesn't consider
2419 * what will happen if errors occur when sending retransmission for the
2420 * second time. ...It could the that such segment has only
2421 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2422 * the head skb is enough except for some reneging corner cases that
2423 * are not worth the effort.
2425 * Main reason for all this complexity is the fact that connection dying
2426 * time now depends on the validity of the retrans_stamp, in particular,
2427 * that successive retransmissions of a segment must not advance
2428 * retrans_stamp under any conditions.
2430 static bool tcp_any_retrans_done(const struct sock
*sk
)
2432 const struct tcp_sock
*tp
= tcp_sk(sk
);
2433 struct sk_buff
*skb
;
2435 if (tp
->retrans_out
)
2438 skb
= tcp_write_queue_head(sk
);
2439 if (unlikely(skb
&& TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
))
2445 /* Undo during loss recovery after partial ACK or using F-RTO. */
2446 static bool tcp_try_undo_loss(struct sock
*sk
, bool frto_undo
)
2448 struct tcp_sock
*tp
= tcp_sk(sk
);
2450 if (frto_undo
|| tcp_may_undo(tp
)) {
2451 tcp_undo_cwnd_reduction(sk
, true);
2453 DBGUNDO(sk
, "partial loss");
2454 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSSUNDO
);
2456 NET_INC_STATS_BH(sock_net(sk
),
2457 LINUX_MIB_TCPSPURIOUSRTOS
);
2458 inet_csk(sk
)->icsk_retransmits
= 0;
2459 if (frto_undo
|| tcp_is_sack(tp
))
2460 tcp_set_ca_state(sk
, TCP_CA_Open
);
2466 /* The cwnd reduction in CWR and Recovery use the PRR algorithm
2467 * https://datatracker.ietf.org/doc/draft-ietf-tcpm-proportional-rate-reduction/
2468 * It computes the number of packets to send (sndcnt) based on packets newly
2470 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2471 * cwnd reductions across a full RTT.
2472 * 2) If packets in flight is lower than ssthresh (such as due to excess
2473 * losses and/or application stalls), do not perform any further cwnd
2474 * reductions, but instead slow start up to ssthresh.
2476 static void tcp_init_cwnd_reduction(struct sock
*sk
, const bool set_ssthresh
)
2478 struct tcp_sock
*tp
= tcp_sk(sk
);
2480 tp
->high_seq
= tp
->snd_nxt
;
2481 tp
->tlp_high_seq
= 0;
2482 tp
->snd_cwnd_cnt
= 0;
2483 tp
->prior_cwnd
= tp
->snd_cwnd
;
2484 tp
->prr_delivered
= 0;
2487 tp
->snd_ssthresh
= inet_csk(sk
)->icsk_ca_ops
->ssthresh(sk
);
2488 TCP_ECN_queue_cwr(tp
);
2491 static void tcp_cwnd_reduction(struct sock
*sk
, const int prior_unsacked
,
2494 struct tcp_sock
*tp
= tcp_sk(sk
);
2496 int delta
= tp
->snd_ssthresh
- tcp_packets_in_flight(tp
);
2497 int newly_acked_sacked
= prior_unsacked
-
2498 (tp
->packets_out
- tp
->sacked_out
);
2500 tp
->prr_delivered
+= newly_acked_sacked
;
2501 if (tcp_packets_in_flight(tp
) > tp
->snd_ssthresh
) {
2502 u64 dividend
= (u64
)tp
->snd_ssthresh
* tp
->prr_delivered
+
2504 sndcnt
= div_u64(dividend
, tp
->prior_cwnd
) - tp
->prr_out
;
2506 sndcnt
= min_t(int, delta
,
2507 max_t(int, tp
->prr_delivered
- tp
->prr_out
,
2508 newly_acked_sacked
) + 1);
2511 sndcnt
= max(sndcnt
, (fast_rexmit
? 1 : 0));
2512 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + sndcnt
;
2515 static inline void tcp_end_cwnd_reduction(struct sock
*sk
)
2517 struct tcp_sock
*tp
= tcp_sk(sk
);
2519 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2520 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_CWR
||
2521 (tp
->undo_marker
&& tp
->snd_ssthresh
< TCP_INFINITE_SSTHRESH
)) {
2522 tp
->snd_cwnd
= tp
->snd_ssthresh
;
2523 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2525 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
2528 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2529 void tcp_enter_cwr(struct sock
*sk
, const int set_ssthresh
)
2531 struct tcp_sock
*tp
= tcp_sk(sk
);
2533 tp
->prior_ssthresh
= 0;
2534 if (inet_csk(sk
)->icsk_ca_state
< TCP_CA_CWR
) {
2535 tp
->undo_marker
= 0;
2536 tcp_init_cwnd_reduction(sk
, set_ssthresh
);
2537 tcp_set_ca_state(sk
, TCP_CA_CWR
);
2541 static void tcp_try_keep_open(struct sock
*sk
)
2543 struct tcp_sock
*tp
= tcp_sk(sk
);
2544 int state
= TCP_CA_Open
;
2546 if (tcp_left_out(tp
) || tcp_any_retrans_done(sk
))
2547 state
= TCP_CA_Disorder
;
2549 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
2550 tcp_set_ca_state(sk
, state
);
2551 tp
->high_seq
= tp
->snd_nxt
;
2555 static void tcp_try_to_open(struct sock
*sk
, int flag
, const int prior_unsacked
)
2557 struct tcp_sock
*tp
= tcp_sk(sk
);
2559 tcp_verify_left_out(tp
);
2561 if (!tcp_any_retrans_done(sk
))
2562 tp
->retrans_stamp
= 0;
2564 if (flag
& FLAG_ECE
)
2565 tcp_enter_cwr(sk
, 1);
2567 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
2568 tcp_try_keep_open(sk
);
2570 tcp_cwnd_reduction(sk
, prior_unsacked
, 0);
2574 static void tcp_mtup_probe_failed(struct sock
*sk
)
2576 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2578 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
2579 icsk
->icsk_mtup
.probe_size
= 0;
2582 static void tcp_mtup_probe_success(struct sock
*sk
)
2584 struct tcp_sock
*tp
= tcp_sk(sk
);
2585 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2587 /* FIXME: breaks with very large cwnd */
2588 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2589 tp
->snd_cwnd
= tp
->snd_cwnd
*
2590 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
2591 icsk
->icsk_mtup
.probe_size
;
2592 tp
->snd_cwnd_cnt
= 0;
2593 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2594 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2596 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
2597 icsk
->icsk_mtup
.probe_size
= 0;
2598 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
2601 /* Do a simple retransmit without using the backoff mechanisms in
2602 * tcp_timer. This is used for path mtu discovery.
2603 * The socket is already locked here.
2605 void tcp_simple_retransmit(struct sock
*sk
)
2607 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2608 struct tcp_sock
*tp
= tcp_sk(sk
);
2609 struct sk_buff
*skb
;
2610 unsigned int mss
= tcp_current_mss(sk
);
2611 u32 prior_lost
= tp
->lost_out
;
2613 tcp_for_write_queue(skb
, sk
) {
2614 if (skb
== tcp_send_head(sk
))
2616 if (tcp_skb_seglen(skb
) > mss
&&
2617 !(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
2618 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2619 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
2620 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2622 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
2626 tcp_clear_retrans_hints_partial(tp
);
2628 if (prior_lost
== tp
->lost_out
)
2631 if (tcp_is_reno(tp
))
2632 tcp_limit_reno_sacked(tp
);
2634 tcp_verify_left_out(tp
);
2636 /* Don't muck with the congestion window here.
2637 * Reason is that we do not increase amount of _data_
2638 * in network, but units changed and effective
2639 * cwnd/ssthresh really reduced now.
2641 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
2642 tp
->high_seq
= tp
->snd_nxt
;
2643 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2644 tp
->prior_ssthresh
= 0;
2645 tp
->undo_marker
= 0;
2646 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2648 tcp_xmit_retransmit_queue(sk
);
2650 EXPORT_SYMBOL(tcp_simple_retransmit
);
2652 static void tcp_enter_recovery(struct sock
*sk
, bool ece_ack
)
2654 struct tcp_sock
*tp
= tcp_sk(sk
);
2657 if (tcp_is_reno(tp
))
2658 mib_idx
= LINUX_MIB_TCPRENORECOVERY
;
2660 mib_idx
= LINUX_MIB_TCPSACKRECOVERY
;
2662 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2664 tp
->prior_ssthresh
= 0;
2665 tp
->undo_marker
= tp
->snd_una
;
2666 tp
->undo_retrans
= tp
->retrans_out
;
2668 if (inet_csk(sk
)->icsk_ca_state
< TCP_CA_CWR
) {
2670 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2671 tcp_init_cwnd_reduction(sk
, true);
2673 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
2676 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2677 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2679 static void tcp_process_loss(struct sock
*sk
, int flag
, bool is_dupack
)
2681 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2682 struct tcp_sock
*tp
= tcp_sk(sk
);
2683 bool recovered
= !before(tp
->snd_una
, tp
->high_seq
);
2685 if (tp
->frto
) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2686 if (flag
& FLAG_ORIG_SACK_ACKED
) {
2687 /* Step 3.b. A timeout is spurious if not all data are
2688 * lost, i.e., never-retransmitted data are (s)acked.
2690 tcp_try_undo_loss(sk
, true);
2693 if (after(tp
->snd_nxt
, tp
->high_seq
) &&
2694 (flag
& FLAG_DATA_SACKED
|| is_dupack
)) {
2695 tp
->frto
= 0; /* Loss was real: 2nd part of step 3.a */
2696 } else if (flag
& FLAG_SND_UNA_ADVANCED
&& !recovered
) {
2697 tp
->high_seq
= tp
->snd_nxt
;
2698 __tcp_push_pending_frames(sk
, tcp_current_mss(sk
),
2700 if (after(tp
->snd_nxt
, tp
->high_seq
))
2701 return; /* Step 2.b */
2707 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2708 icsk
->icsk_retransmits
= 0;
2709 tcp_try_undo_recovery(sk
);
2712 if (flag
& FLAG_DATA_ACKED
)
2713 icsk
->icsk_retransmits
= 0;
2714 if (tcp_is_reno(tp
)) {
2715 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2716 * delivered. Lower inflight to clock out (re)tranmissions.
2718 if (after(tp
->snd_nxt
, tp
->high_seq
) && is_dupack
)
2719 tcp_add_reno_sack(sk
);
2720 else if (flag
& FLAG_SND_UNA_ADVANCED
)
2721 tcp_reset_reno_sack(tp
);
2723 if (tcp_try_undo_loss(sk
, false))
2725 tcp_xmit_retransmit_queue(sk
);
2728 /* Undo during fast recovery after partial ACK. */
2729 static bool tcp_try_undo_partial(struct sock
*sk
, const int acked
,
2730 const int prior_unsacked
)
2732 struct tcp_sock
*tp
= tcp_sk(sk
);
2734 if (tp
->undo_marker
&& tcp_packet_delayed(tp
)) {
2735 /* Plain luck! Hole if filled with delayed
2736 * packet, rather than with a retransmit.
2738 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
2740 /* We are getting evidence that the reordering degree is higher
2741 * than we realized. If there are no retransmits out then we
2742 * can undo. Otherwise we clock out new packets but do not
2743 * mark more packets lost or retransmit more.
2745 if (tp
->retrans_out
) {
2746 tcp_cwnd_reduction(sk
, prior_unsacked
, 0);
2750 if (!tcp_any_retrans_done(sk
))
2751 tp
->retrans_stamp
= 0;
2753 DBGUNDO(sk
, "partial recovery");
2754 tcp_undo_cwnd_reduction(sk
, true);
2755 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPARTIALUNDO
);
2756 tcp_try_keep_open(sk
);
2762 /* Process an event, which can update packets-in-flight not trivially.
2763 * Main goal of this function is to calculate new estimate for left_out,
2764 * taking into account both packets sitting in receiver's buffer and
2765 * packets lost by network.
2767 * Besides that it does CWND reduction, when packet loss is detected
2768 * and changes state of machine.
2770 * It does _not_ decide what to send, it is made in function
2771 * tcp_xmit_retransmit_queue().
2773 static void tcp_fastretrans_alert(struct sock
*sk
, const int acked
,
2774 const int prior_unsacked
,
2775 bool is_dupack
, int flag
)
2777 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2778 struct tcp_sock
*tp
= tcp_sk(sk
);
2779 bool do_lost
= is_dupack
|| ((flag
& FLAG_DATA_SACKED
) &&
2780 (tcp_fackets_out(tp
) > tp
->reordering
));
2781 int fast_rexmit
= 0;
2783 if (WARN_ON(!tp
->packets_out
&& tp
->sacked_out
))
2785 if (WARN_ON(!tp
->sacked_out
&& tp
->fackets_out
))
2786 tp
->fackets_out
= 0;
2788 /* Now state machine starts.
2789 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2790 if (flag
& FLAG_ECE
)
2791 tp
->prior_ssthresh
= 0;
2793 /* B. In all the states check for reneging SACKs. */
2794 if (tcp_check_sack_reneging(sk
, flag
))
2797 /* C. Check consistency of the current state. */
2798 tcp_verify_left_out(tp
);
2800 /* D. Check state exit conditions. State can be terminated
2801 * when high_seq is ACKed. */
2802 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
2803 WARN_ON(tp
->retrans_out
!= 0);
2804 tp
->retrans_stamp
= 0;
2805 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
2806 switch (icsk
->icsk_ca_state
) {
2808 /* CWR is to be held something *above* high_seq
2809 * is ACKed for CWR bit to reach receiver. */
2810 if (tp
->snd_una
!= tp
->high_seq
) {
2811 tcp_end_cwnd_reduction(sk
);
2812 tcp_set_ca_state(sk
, TCP_CA_Open
);
2816 case TCP_CA_Recovery
:
2817 if (tcp_is_reno(tp
))
2818 tcp_reset_reno_sack(tp
);
2819 if (tcp_try_undo_recovery(sk
))
2821 tcp_end_cwnd_reduction(sk
);
2826 /* E. Process state. */
2827 switch (icsk
->icsk_ca_state
) {
2828 case TCP_CA_Recovery
:
2829 if (!(flag
& FLAG_SND_UNA_ADVANCED
)) {
2830 if (tcp_is_reno(tp
) && is_dupack
)
2831 tcp_add_reno_sack(sk
);
2833 if (tcp_try_undo_partial(sk
, acked
, prior_unsacked
))
2835 /* Partial ACK arrived. Force fast retransmit. */
2836 do_lost
= tcp_is_reno(tp
) ||
2837 tcp_fackets_out(tp
) > tp
->reordering
;
2839 if (tcp_try_undo_dsack(sk
)) {
2840 tcp_try_keep_open(sk
);
2845 tcp_process_loss(sk
, flag
, is_dupack
);
2846 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
2848 /* Fall through to processing in Open state. */
2850 if (tcp_is_reno(tp
)) {
2851 if (flag
& FLAG_SND_UNA_ADVANCED
)
2852 tcp_reset_reno_sack(tp
);
2854 tcp_add_reno_sack(sk
);
2857 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
)
2858 tcp_try_undo_dsack(sk
);
2860 if (!tcp_time_to_recover(sk
, flag
)) {
2861 tcp_try_to_open(sk
, flag
, prior_unsacked
);
2865 /* MTU probe failure: don't reduce cwnd */
2866 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
2867 icsk
->icsk_mtup
.probe_size
&&
2868 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
2869 tcp_mtup_probe_failed(sk
);
2870 /* Restores the reduction we did in tcp_mtup_probe() */
2872 tcp_simple_retransmit(sk
);
2876 /* Otherwise enter Recovery state */
2877 tcp_enter_recovery(sk
, (flag
& FLAG_ECE
));
2882 tcp_update_scoreboard(sk
, fast_rexmit
);
2883 tcp_cwnd_reduction(sk
, prior_unsacked
, fast_rexmit
);
2884 tcp_xmit_retransmit_queue(sk
);
2887 static inline bool tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
2888 s32 seq_rtt
, s32 sack_rtt
)
2890 const struct tcp_sock
*tp
= tcp_sk(sk
);
2892 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2893 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2894 * Karn's algorithm forbids taking RTT if some retransmitted data
2895 * is acked (RFC6298).
2897 if (flag
& FLAG_RETRANS_DATA_ACKED
)
2903 /* RTTM Rule: A TSecr value received in a segment is used to
2904 * update the averaged RTT measurement only if the segment
2905 * acknowledges some new data, i.e., only if it advances the
2906 * left edge of the send window.
2907 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2909 if (seq_rtt
< 0 && tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2911 seq_rtt
= tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
;
2916 tcp_rtt_estimator(sk
, seq_rtt
);
2919 /* RFC6298: only reset backoff on valid RTT measurement. */
2920 inet_csk(sk
)->icsk_backoff
= 0;
2924 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2925 static void tcp_synack_rtt_meas(struct sock
*sk
, const u32 synack_stamp
)
2927 struct tcp_sock
*tp
= tcp_sk(sk
);
2930 if (synack_stamp
&& !tp
->total_retrans
)
2931 seq_rtt
= tcp_time_stamp
- synack_stamp
;
2933 /* If the ACK acks both the SYNACK and the (Fast Open'd) data packets
2934 * sent in SYN_RECV, SYNACK RTT is the smooth RTT computed in tcp_ack()
2937 tcp_ack_update_rtt(sk
, FLAG_SYN_ACKED
, seq_rtt
, -1);
2940 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
, u32 acked
, u32 in_flight
)
2942 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2943 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, acked
, in_flight
);
2944 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_time_stamp
;
2947 /* Restart timer after forward progress on connection.
2948 * RFC2988 recommends to restart timer to now+rto.
2950 void tcp_rearm_rto(struct sock
*sk
)
2952 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2953 struct tcp_sock
*tp
= tcp_sk(sk
);
2955 /* If the retrans timer is currently being used by Fast Open
2956 * for SYN-ACK retrans purpose, stay put.
2958 if (tp
->fastopen_rsk
)
2961 if (!tp
->packets_out
) {
2962 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
2964 u32 rto
= inet_csk(sk
)->icsk_rto
;
2965 /* Offset the time elapsed after installing regular RTO */
2966 if (icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
||
2967 icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
) {
2968 struct sk_buff
*skb
= tcp_write_queue_head(sk
);
2969 const u32 rto_time_stamp
= TCP_SKB_CB(skb
)->when
+ rto
;
2970 s32 delta
= (s32
)(rto_time_stamp
- tcp_time_stamp
);
2971 /* delta may not be positive if the socket is locked
2972 * when the retrans timer fires and is rescheduled.
2977 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
, rto
,
2982 /* This function is called when the delayed ER timer fires. TCP enters
2983 * fast recovery and performs fast-retransmit.
2985 void tcp_resume_early_retransmit(struct sock
*sk
)
2987 struct tcp_sock
*tp
= tcp_sk(sk
);
2991 /* Stop if ER is disabled after the delayed ER timer is scheduled */
2992 if (!tp
->do_early_retrans
)
2995 tcp_enter_recovery(sk
, false);
2996 tcp_update_scoreboard(sk
, 1);
2997 tcp_xmit_retransmit_queue(sk
);
3000 /* If we get here, the whole TSO packet has not been acked. */
3001 static u32
tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
)
3003 struct tcp_sock
*tp
= tcp_sk(sk
);
3006 BUG_ON(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
));
3008 packets_acked
= tcp_skb_pcount(skb
);
3009 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
3011 packets_acked
-= tcp_skb_pcount(skb
);
3013 if (packets_acked
) {
3014 BUG_ON(tcp_skb_pcount(skb
) == 0);
3015 BUG_ON(!before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
));
3018 return packets_acked
;
3021 /* Remove acknowledged frames from the retransmission queue. If our packet
3022 * is before the ack sequence we can discard it as it's confirmed to have
3023 * arrived at the other end.
3025 static int tcp_clean_rtx_queue(struct sock
*sk
, int prior_fackets
,
3026 u32 prior_snd_una
, s32 sack_rtt
)
3028 struct tcp_sock
*tp
= tcp_sk(sk
);
3029 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3030 struct sk_buff
*skb
;
3031 u32 now
= tcp_time_stamp
;
3032 bool fully_acked
= true;
3035 u32 reord
= tp
->packets_out
;
3036 u32 prior_sacked
= tp
->sacked_out
;
3038 s32 ca_seq_rtt
= -1;
3039 ktime_t last_ackt
= net_invalid_timestamp();
3042 while ((skb
= tcp_write_queue_head(sk
)) && skb
!= tcp_send_head(sk
)) {
3043 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
3045 u8 sacked
= scb
->sacked
;
3047 /* Determine how many packets and what bytes were acked, tso and else */
3048 if (after(scb
->end_seq
, tp
->snd_una
)) {
3049 if (tcp_skb_pcount(skb
) == 1 ||
3050 !after(tp
->snd_una
, scb
->seq
))
3053 acked_pcount
= tcp_tso_acked(sk
, skb
);
3057 fully_acked
= false;
3059 acked_pcount
= tcp_skb_pcount(skb
);
3062 if (sacked
& TCPCB_RETRANS
) {
3063 if (sacked
& TCPCB_SACKED_RETRANS
)
3064 tp
->retrans_out
-= acked_pcount
;
3065 flag
|= FLAG_RETRANS_DATA_ACKED
;
3067 ca_seq_rtt
= now
- scb
->when
;
3068 last_ackt
= skb
->tstamp
;
3070 seq_rtt
= ca_seq_rtt
;
3072 if (!(sacked
& TCPCB_SACKED_ACKED
))
3073 reord
= min(pkts_acked
, reord
);
3074 if (!after(scb
->end_seq
, tp
->high_seq
))
3075 flag
|= FLAG_ORIG_SACK_ACKED
;
3078 if (sacked
& TCPCB_SACKED_ACKED
)
3079 tp
->sacked_out
-= acked_pcount
;
3080 if (sacked
& TCPCB_LOST
)
3081 tp
->lost_out
-= acked_pcount
;
3083 tp
->packets_out
-= acked_pcount
;
3084 pkts_acked
+= acked_pcount
;
3086 /* Initial outgoing SYN's get put onto the write_queue
3087 * just like anything else we transmit. It is not
3088 * true data, and if we misinform our callers that
3089 * this ACK acks real data, we will erroneously exit
3090 * connection startup slow start one packet too
3091 * quickly. This is severely frowned upon behavior.
3093 if (!(scb
->tcp_flags
& TCPHDR_SYN
)) {
3094 flag
|= FLAG_DATA_ACKED
;
3096 flag
|= FLAG_SYN_ACKED
;
3097 tp
->retrans_stamp
= 0;
3103 tcp_unlink_write_queue(skb
, sk
);
3104 sk_wmem_free_skb(sk
, skb
);
3105 if (skb
== tp
->retransmit_skb_hint
)
3106 tp
->retransmit_skb_hint
= NULL
;
3107 if (skb
== tp
->lost_skb_hint
)
3108 tp
->lost_skb_hint
= NULL
;
3111 if (likely(between(tp
->snd_up
, prior_snd_una
, tp
->snd_una
)))
3112 tp
->snd_up
= tp
->snd_una
;
3114 if (skb
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
3115 flag
|= FLAG_SACK_RENEGING
;
3117 rtt_update
= tcp_ack_update_rtt(sk
, flag
, seq_rtt
, sack_rtt
);
3119 if (flag
& FLAG_ACKED
) {
3120 const struct tcp_congestion_ops
*ca_ops
3121 = inet_csk(sk
)->icsk_ca_ops
;
3124 if (unlikely(icsk
->icsk_mtup
.probe_size
&&
3125 !after(tp
->mtu_probe
.probe_seq_end
, tp
->snd_una
))) {
3126 tcp_mtup_probe_success(sk
);
3129 if (tcp_is_reno(tp
)) {
3130 tcp_remove_reno_sacks(sk
, pkts_acked
);
3134 /* Non-retransmitted hole got filled? That's reordering */
3135 if (reord
< prior_fackets
)
3136 tcp_update_reordering(sk
, tp
->fackets_out
- reord
, 0);
3138 delta
= tcp_is_fack(tp
) ? pkts_acked
:
3139 prior_sacked
- tp
->sacked_out
;
3140 tp
->lost_cnt_hint
-= min(tp
->lost_cnt_hint
, delta
);
3143 tp
->fackets_out
-= min(pkts_acked
, tp
->fackets_out
);
3145 if (ca_ops
->pkts_acked
) {
3148 /* Is the ACK triggering packet unambiguous? */
3149 if (!(flag
& FLAG_RETRANS_DATA_ACKED
)) {
3150 /* High resolution needed and available? */
3151 if (ca_ops
->flags
& TCP_CONG_RTT_STAMP
&&
3152 !ktime_equal(last_ackt
,
3153 net_invalid_timestamp()))
3154 rtt_us
= ktime_us_delta(ktime_get_real(),
3156 else if (ca_seq_rtt
>= 0)
3157 rtt_us
= jiffies_to_usecs(ca_seq_rtt
);
3160 ca_ops
->pkts_acked(sk
, pkts_acked
, rtt_us
);
3162 } else if (skb
&& rtt_update
&& sack_rtt
>= 0 &&
3163 sack_rtt
> (s32
)(now
- TCP_SKB_CB(skb
)->when
)) {
3164 /* Do not re-arm RTO if the sack RTT is measured from data sent
3165 * after when the head was last (re)transmitted. Otherwise the
3166 * timeout may continue to extend in loss recovery.
3171 #if FASTRETRANS_DEBUG > 0
3172 WARN_ON((int)tp
->sacked_out
< 0);
3173 WARN_ON((int)tp
->lost_out
< 0);
3174 WARN_ON((int)tp
->retrans_out
< 0);
3175 if (!tp
->packets_out
&& tcp_is_sack(tp
)) {
3176 icsk
= inet_csk(sk
);
3178 pr_debug("Leak l=%u %d\n",
3179 tp
->lost_out
, icsk
->icsk_ca_state
);
3182 if (tp
->sacked_out
) {
3183 pr_debug("Leak s=%u %d\n",
3184 tp
->sacked_out
, icsk
->icsk_ca_state
);
3187 if (tp
->retrans_out
) {
3188 pr_debug("Leak r=%u %d\n",
3189 tp
->retrans_out
, icsk
->icsk_ca_state
);
3190 tp
->retrans_out
= 0;
3197 static void tcp_ack_probe(struct sock
*sk
)
3199 const struct tcp_sock
*tp
= tcp_sk(sk
);
3200 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3202 /* Was it a usable window open? */
3204 if (!after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
, tcp_wnd_end(tp
))) {
3205 icsk
->icsk_backoff
= 0;
3206 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
3207 /* Socket must be waked up by subsequent tcp_data_snd_check().
3208 * This function is not for random using!
3211 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3212 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
3217 static inline bool tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
3219 return !(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
3220 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
;
3223 /* Decide wheather to run the increase function of congestion control. */
3224 static inline bool tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
3226 if (tcp_in_cwnd_reduction(sk
))
3229 /* If reordering is high then always grow cwnd whenever data is
3230 * delivered regardless of its ordering. Otherwise stay conservative
3231 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3232 * new SACK or ECE mark may first advance cwnd here and later reduce
3233 * cwnd in tcp_fastretrans_alert() based on more states.
3235 if (tcp_sk(sk
)->reordering
> sysctl_tcp_reordering
)
3236 return flag
& FLAG_FORWARD_PROGRESS
;
3238 return flag
& FLAG_DATA_ACKED
;
3241 /* Check that window update is acceptable.
3242 * The function assumes that snd_una<=ack<=snd_next.
3244 static inline bool tcp_may_update_window(const struct tcp_sock
*tp
,
3245 const u32 ack
, const u32 ack_seq
,
3248 return after(ack
, tp
->snd_una
) ||
3249 after(ack_seq
, tp
->snd_wl1
) ||
3250 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
);
3253 /* Update our send window.
3255 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3256 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3258 static int tcp_ack_update_window(struct sock
*sk
, const struct sk_buff
*skb
, u32 ack
,
3261 struct tcp_sock
*tp
= tcp_sk(sk
);
3263 u32 nwin
= ntohs(tcp_hdr(skb
)->window
);
3265 if (likely(!tcp_hdr(skb
)->syn
))
3266 nwin
<<= tp
->rx_opt
.snd_wscale
;
3268 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
3269 flag
|= FLAG_WIN_UPDATE
;
3270 tcp_update_wl(tp
, ack_seq
);
3272 if (tp
->snd_wnd
!= nwin
) {
3275 /* Note, it is the only place, where
3276 * fast path is recovered for sending TCP.
3279 tcp_fast_path_check(sk
);
3281 if (nwin
> tp
->max_window
) {
3282 tp
->max_window
= nwin
;
3283 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
3293 /* RFC 5961 7 [ACK Throttling] */
3294 static void tcp_send_challenge_ack(struct sock
*sk
)
3296 /* unprotected vars, we dont care of overwrites */
3297 static u32 challenge_timestamp
;
3298 static unsigned int challenge_count
;
3299 u32 now
= jiffies
/ HZ
;
3301 if (now
!= challenge_timestamp
) {
3302 challenge_timestamp
= now
;
3303 challenge_count
= 0;
3305 if (++challenge_count
<= sysctl_tcp_challenge_ack_limit
) {
3306 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPCHALLENGEACK
);
3311 static void tcp_store_ts_recent(struct tcp_sock
*tp
)
3313 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
3314 tp
->rx_opt
.ts_recent_stamp
= get_seconds();
3317 static void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
3319 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
3320 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3321 * extra check below makes sure this can only happen
3322 * for pure ACK frames. -DaveM
3324 * Not only, also it occurs for expired timestamps.
3327 if (tcp_paws_check(&tp
->rx_opt
, 0))
3328 tcp_store_ts_recent(tp
);
3332 /* This routine deals with acks during a TLP episode.
3333 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3335 static void tcp_process_tlp_ack(struct sock
*sk
, u32 ack
, int flag
)
3337 struct tcp_sock
*tp
= tcp_sk(sk
);
3338 bool is_tlp_dupack
= (ack
== tp
->tlp_high_seq
) &&
3339 !(flag
& (FLAG_SND_UNA_ADVANCED
|
3340 FLAG_NOT_DUP
| FLAG_DATA_SACKED
));
3342 /* Mark the end of TLP episode on receiving TLP dupack or when
3343 * ack is after tlp_high_seq.
3345 if (is_tlp_dupack
) {
3346 tp
->tlp_high_seq
= 0;
3350 if (after(ack
, tp
->tlp_high_seq
)) {
3351 tp
->tlp_high_seq
= 0;
3352 /* Don't reduce cwnd if DSACK arrives for TLP retrans. */
3353 if (!(flag
& FLAG_DSACKING_ACK
)) {
3354 tcp_init_cwnd_reduction(sk
, true);
3355 tcp_set_ca_state(sk
, TCP_CA_CWR
);
3356 tcp_end_cwnd_reduction(sk
);
3357 tcp_try_keep_open(sk
);
3358 NET_INC_STATS_BH(sock_net(sk
),
3359 LINUX_MIB_TCPLOSSPROBERECOVERY
);
3364 /* This routine deals with incoming acks, but not outgoing ones. */
3365 static int tcp_ack(struct sock
*sk
, const struct sk_buff
*skb
, int flag
)
3367 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3368 struct tcp_sock
*tp
= tcp_sk(sk
);
3369 u32 prior_snd_una
= tp
->snd_una
;
3370 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
3371 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3372 bool is_dupack
= false;
3373 u32 prior_in_flight
, prior_cwnd
= tp
->snd_cwnd
, prior_rtt
= tp
->srtt
;
3375 int prior_packets
= tp
->packets_out
;
3376 const int prior_unsacked
= tp
->packets_out
- tp
->sacked_out
;
3377 int acked
= 0; /* Number of packets newly acked */
3380 /* If the ack is older than previous acks
3381 * then we can probably ignore it.
3383 if (before(ack
, prior_snd_una
)) {
3384 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3385 if (before(ack
, prior_snd_una
- tp
->max_window
)) {
3386 tcp_send_challenge_ack(sk
);
3392 /* If the ack includes data we haven't sent yet, discard
3393 * this segment (RFC793 Section 3.9).
3395 if (after(ack
, tp
->snd_nxt
))
3398 if (icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
||
3399 icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
)
3402 if (after(ack
, prior_snd_una
))
3403 flag
|= FLAG_SND_UNA_ADVANCED
;
3405 prior_fackets
= tp
->fackets_out
;
3406 prior_in_flight
= tcp_packets_in_flight(tp
);
3408 /* ts_recent update must be made after we are sure that the packet
3411 if (flag
& FLAG_UPDATE_TS_RECENT
)
3412 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
3414 if (!(flag
& FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
3415 /* Window is constant, pure forward advance.
3416 * No more checks are required.
3417 * Note, we use the fact that SND.UNA>=SND.WL2.
3419 tcp_update_wl(tp
, ack_seq
);
3421 flag
|= FLAG_WIN_UPDATE
;
3423 tcp_ca_event(sk
, CA_EVENT_FAST_ACK
);
3425 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPACKS
);
3427 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
3430 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPUREACKS
);
3432 flag
|= tcp_ack_update_window(sk
, skb
, ack
, ack_seq
);
3434 if (TCP_SKB_CB(skb
)->sacked
)
3435 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
,
3438 if (TCP_ECN_rcv_ecn_echo(tp
, tcp_hdr(skb
)))
3441 tcp_ca_event(sk
, CA_EVENT_SLOW_ACK
);
3444 /* We passed data and got it acked, remove any soft error
3445 * log. Something worked...
3447 sk
->sk_err_soft
= 0;
3448 icsk
->icsk_probes_out
= 0;
3449 tp
->rcv_tstamp
= tcp_time_stamp
;
3453 /* See if we can take anything off of the retransmit queue. */
3454 acked
= tp
->packets_out
;
3455 flag
|= tcp_clean_rtx_queue(sk
, prior_fackets
, prior_snd_una
, sack_rtt
);
3456 acked
-= tp
->packets_out
;
3458 /* Advance cwnd if state allows */
3459 if (tcp_may_raise_cwnd(sk
, flag
))
3460 tcp_cong_avoid(sk
, ack
, acked
, prior_in_flight
);
3462 if (tcp_ack_is_dubious(sk
, flag
)) {
3463 is_dupack
= !(flag
& (FLAG_SND_UNA_ADVANCED
| FLAG_NOT_DUP
));
3464 tcp_fastretrans_alert(sk
, acked
, prior_unsacked
,
3467 if (tp
->tlp_high_seq
)
3468 tcp_process_tlp_ack(sk
, ack
, flag
);
3470 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
& FLAG_NOT_DUP
)) {
3471 struct dst_entry
*dst
= __sk_dst_get(sk
);
3476 if (icsk
->icsk_pending
== ICSK_TIME_RETRANS
)
3477 tcp_schedule_loss_probe(sk
);
3478 if (tp
->srtt
!= prior_rtt
|| tp
->snd_cwnd
!= prior_cwnd
)
3479 tcp_update_pacing_rate(sk
);
3483 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3484 if (flag
& FLAG_DSACKING_ACK
)
3485 tcp_fastretrans_alert(sk
, acked
, prior_unsacked
,
3487 /* If this ack opens up a zero window, clear backoff. It was
3488 * being used to time the probes, and is probably far higher than
3489 * it needs to be for normal retransmission.
3491 if (tcp_send_head(sk
))
3494 if (tp
->tlp_high_seq
)
3495 tcp_process_tlp_ack(sk
, ack
, flag
);
3499 SOCK_DEBUG(sk
, "Ack %u after %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3503 /* If data was SACKed, tag it and see if we should send more data.
3504 * If data was DSACKed, see if we can undo a cwnd reduction.
3506 if (TCP_SKB_CB(skb
)->sacked
) {
3507 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
,
3509 tcp_fastretrans_alert(sk
, acked
, prior_unsacked
,
3513 SOCK_DEBUG(sk
, "Ack %u before %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3517 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3518 * But, this can also be called on packets in the established flow when
3519 * the fast version below fails.
3521 void tcp_parse_options(const struct sk_buff
*skb
,
3522 struct tcp_options_received
*opt_rx
, int estab
,
3523 struct tcp_fastopen_cookie
*foc
)
3525 const unsigned char *ptr
;
3526 const struct tcphdr
*th
= tcp_hdr(skb
);
3527 int length
= (th
->doff
* 4) - sizeof(struct tcphdr
);
3529 ptr
= (const unsigned char *)(th
+ 1);
3530 opt_rx
->saw_tstamp
= 0;
3532 while (length
> 0) {
3533 int opcode
= *ptr
++;
3539 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
3544 if (opsize
< 2) /* "silly options" */
3546 if (opsize
> length
)
3547 return; /* don't parse partial options */
3550 if (opsize
== TCPOLEN_MSS
&& th
->syn
&& !estab
) {
3551 u16 in_mss
= get_unaligned_be16(ptr
);
3553 if (opt_rx
->user_mss
&&
3554 opt_rx
->user_mss
< in_mss
)
3555 in_mss
= opt_rx
->user_mss
;
3556 opt_rx
->mss_clamp
= in_mss
;
3561 if (opsize
== TCPOLEN_WINDOW
&& th
->syn
&&
3562 !estab
&& sysctl_tcp_window_scaling
) {
3563 __u8 snd_wscale
= *(__u8
*)ptr
;
3564 opt_rx
->wscale_ok
= 1;
3565 if (snd_wscale
> 14) {
3566 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3571 opt_rx
->snd_wscale
= snd_wscale
;
3574 case TCPOPT_TIMESTAMP
:
3575 if ((opsize
== TCPOLEN_TIMESTAMP
) &&
3576 ((estab
&& opt_rx
->tstamp_ok
) ||
3577 (!estab
&& sysctl_tcp_timestamps
))) {
3578 opt_rx
->saw_tstamp
= 1;
3579 opt_rx
->rcv_tsval
= get_unaligned_be32(ptr
);
3580 opt_rx
->rcv_tsecr
= get_unaligned_be32(ptr
+ 4);
3583 case TCPOPT_SACK_PERM
:
3584 if (opsize
== TCPOLEN_SACK_PERM
&& th
->syn
&&
3585 !estab
&& sysctl_tcp_sack
) {
3586 opt_rx
->sack_ok
= TCP_SACK_SEEN
;
3587 tcp_sack_reset(opt_rx
);
3592 if ((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
3593 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
3595 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
3598 #ifdef CONFIG_TCP_MD5SIG
3601 * The MD5 Hash has already been
3602 * checked (see tcp_v{4,6}_do_rcv()).
3607 /* Fast Open option shares code 254 using a
3608 * 16 bits magic number. It's valid only in
3609 * SYN or SYN-ACK with an even size.
3611 if (opsize
< TCPOLEN_EXP_FASTOPEN_BASE
||
3612 get_unaligned_be16(ptr
) != TCPOPT_FASTOPEN_MAGIC
||
3613 foc
== NULL
|| !th
->syn
|| (opsize
& 1))
3615 foc
->len
= opsize
- TCPOLEN_EXP_FASTOPEN_BASE
;
3616 if (foc
->len
>= TCP_FASTOPEN_COOKIE_MIN
&&
3617 foc
->len
<= TCP_FASTOPEN_COOKIE_MAX
)
3618 memcpy(foc
->val
, ptr
+ 2, foc
->len
);
3619 else if (foc
->len
!= 0)
3629 EXPORT_SYMBOL(tcp_parse_options
);
3631 static bool tcp_parse_aligned_timestamp(struct tcp_sock
*tp
, const struct tcphdr
*th
)
3633 const __be32
*ptr
= (const __be32
*)(th
+ 1);
3635 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3636 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
3637 tp
->rx_opt
.saw_tstamp
= 1;
3639 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3642 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
) - tp
->tsoffset
;
3644 tp
->rx_opt
.rcv_tsecr
= 0;
3650 /* Fast parse options. This hopes to only see timestamps.
3651 * If it is wrong it falls back on tcp_parse_options().
3653 static bool tcp_fast_parse_options(const struct sk_buff
*skb
,
3654 const struct tcphdr
*th
, struct tcp_sock
*tp
)
3656 /* In the spirit of fast parsing, compare doff directly to constant
3657 * values. Because equality is used, short doff can be ignored here.
3659 if (th
->doff
== (sizeof(*th
) / 4)) {
3660 tp
->rx_opt
.saw_tstamp
= 0;
3662 } else if (tp
->rx_opt
.tstamp_ok
&&
3663 th
->doff
== ((sizeof(*th
) + TCPOLEN_TSTAMP_ALIGNED
) / 4)) {
3664 if (tcp_parse_aligned_timestamp(tp
, th
))
3668 tcp_parse_options(skb
, &tp
->rx_opt
, 1, NULL
);
3669 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
3670 tp
->rx_opt
.rcv_tsecr
-= tp
->tsoffset
;
3675 #ifdef CONFIG_TCP_MD5SIG
3677 * Parse MD5 Signature option
3679 const u8
*tcp_parse_md5sig_option(const struct tcphdr
*th
)
3681 int length
= (th
->doff
<< 2) - sizeof(*th
);
3682 const u8
*ptr
= (const u8
*)(th
+ 1);
3684 /* If the TCP option is too short, we can short cut */
3685 if (length
< TCPOLEN_MD5SIG
)
3688 while (length
> 0) {
3689 int opcode
= *ptr
++;
3700 if (opsize
< 2 || opsize
> length
)
3702 if (opcode
== TCPOPT_MD5SIG
)
3703 return opsize
== TCPOLEN_MD5SIG
? ptr
: NULL
;
3710 EXPORT_SYMBOL(tcp_parse_md5sig_option
);
3713 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3715 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3716 * it can pass through stack. So, the following predicate verifies that
3717 * this segment is not used for anything but congestion avoidance or
3718 * fast retransmit. Moreover, we even are able to eliminate most of such
3719 * second order effects, if we apply some small "replay" window (~RTO)
3720 * to timestamp space.
3722 * All these measures still do not guarantee that we reject wrapped ACKs
3723 * on networks with high bandwidth, when sequence space is recycled fastly,
3724 * but it guarantees that such events will be very rare and do not affect
3725 * connection seriously. This doesn't look nice, but alas, PAWS is really
3728 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3729 * states that events when retransmit arrives after original data are rare.
3730 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3731 * the biggest problem on large power networks even with minor reordering.
3732 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3733 * up to bandwidth of 18Gigabit/sec. 8) ]
3736 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
3738 const struct tcp_sock
*tp
= tcp_sk(sk
);
3739 const struct tcphdr
*th
= tcp_hdr(skb
);
3740 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3741 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3743 return (/* 1. Pure ACK with correct sequence number. */
3744 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
3746 /* 2. ... and duplicate ACK. */
3747 ack
== tp
->snd_una
&&
3749 /* 3. ... and does not update window. */
3750 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
3752 /* 4. ... and sits in replay window. */
3753 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
3756 static inline bool tcp_paws_discard(const struct sock
*sk
,
3757 const struct sk_buff
*skb
)
3759 const struct tcp_sock
*tp
= tcp_sk(sk
);
3761 return !tcp_paws_check(&tp
->rx_opt
, TCP_PAWS_WINDOW
) &&
3762 !tcp_disordered_ack(sk
, skb
);
3765 /* Check segment sequence number for validity.
3767 * Segment controls are considered valid, if the segment
3768 * fits to the window after truncation to the window. Acceptability
3769 * of data (and SYN, FIN, of course) is checked separately.
3770 * See tcp_data_queue(), for example.
3772 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3773 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3774 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3775 * (borrowed from freebsd)
3778 static inline bool tcp_sequence(const struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3780 return !before(end_seq
, tp
->rcv_wup
) &&
3781 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
3784 /* When we get a reset we do this. */
3785 void tcp_reset(struct sock
*sk
)
3787 /* We want the right error as BSD sees it (and indeed as we do). */
3788 switch (sk
->sk_state
) {
3790 sk
->sk_err
= ECONNREFUSED
;
3792 case TCP_CLOSE_WAIT
:
3798 sk
->sk_err
= ECONNRESET
;
3800 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3803 if (!sock_flag(sk
, SOCK_DEAD
))
3804 sk
->sk_error_report(sk
);
3810 * Process the FIN bit. This now behaves as it is supposed to work
3811 * and the FIN takes effect when it is validly part of sequence
3812 * space. Not before when we get holes.
3814 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3815 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3818 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3819 * close and we go into CLOSING (and later onto TIME-WAIT)
3821 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3823 static void tcp_fin(struct sock
*sk
)
3825 struct tcp_sock
*tp
= tcp_sk(sk
);
3826 const struct dst_entry
*dst
;
3828 inet_csk_schedule_ack(sk
);
3830 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
3831 sock_set_flag(sk
, SOCK_DONE
);
3833 switch (sk
->sk_state
) {
3835 case TCP_ESTABLISHED
:
3836 /* Move to CLOSE_WAIT */
3837 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
3838 dst
= __sk_dst_get(sk
);
3839 if (!dst
|| !dst_metric(dst
, RTAX_QUICKACK
))
3840 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
3843 case TCP_CLOSE_WAIT
:
3845 /* Received a retransmission of the FIN, do
3850 /* RFC793: Remain in the LAST-ACK state. */
3854 /* This case occurs when a simultaneous close
3855 * happens, we must ack the received FIN and
3856 * enter the CLOSING state.
3859 tcp_set_state(sk
, TCP_CLOSING
);
3862 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3864 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
3867 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3868 * cases we should never reach this piece of code.
3870 pr_err("%s: Impossible, sk->sk_state=%d\n",
3871 __func__
, sk
->sk_state
);
3875 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3876 * Probably, we should reset in this case. For now drop them.
3878 __skb_queue_purge(&tp
->out_of_order_queue
);
3879 if (tcp_is_sack(tp
))
3880 tcp_sack_reset(&tp
->rx_opt
);
3883 if (!sock_flag(sk
, SOCK_DEAD
)) {
3884 sk
->sk_state_change(sk
);
3886 /* Do not send POLL_HUP for half duplex close. */
3887 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
3888 sk
->sk_state
== TCP_CLOSE
)
3889 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_HUP
);
3891 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_IN
);
3895 static inline bool tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
,
3898 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
3899 if (before(seq
, sp
->start_seq
))
3900 sp
->start_seq
= seq
;
3901 if (after(end_seq
, sp
->end_seq
))
3902 sp
->end_seq
= end_seq
;
3908 static void tcp_dsack_set(struct sock
*sk
, u32 seq
, u32 end_seq
)
3910 struct tcp_sock
*tp
= tcp_sk(sk
);
3912 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
3915 if (before(seq
, tp
->rcv_nxt
))
3916 mib_idx
= LINUX_MIB_TCPDSACKOLDSENT
;
3918 mib_idx
= LINUX_MIB_TCPDSACKOFOSENT
;
3920 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
3922 tp
->rx_opt
.dsack
= 1;
3923 tp
->duplicate_sack
[0].start_seq
= seq
;
3924 tp
->duplicate_sack
[0].end_seq
= end_seq
;
3928 static void tcp_dsack_extend(struct sock
*sk
, u32 seq
, u32 end_seq
)
3930 struct tcp_sock
*tp
= tcp_sk(sk
);
3932 if (!tp
->rx_opt
.dsack
)
3933 tcp_dsack_set(sk
, seq
, end_seq
);
3935 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
3938 static void tcp_send_dupack(struct sock
*sk
, const struct sk_buff
*skb
)
3940 struct tcp_sock
*tp
= tcp_sk(sk
);
3942 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
3943 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
3944 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
3945 tcp_enter_quickack_mode(sk
);
3947 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
3948 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
3950 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
3951 end_seq
= tp
->rcv_nxt
;
3952 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, end_seq
);
3959 /* These routines update the SACK block as out-of-order packets arrive or
3960 * in-order packets close up the sequence space.
3962 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
3965 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3966 struct tcp_sack_block
*swalk
= sp
+ 1;
3968 /* See if the recent change to the first SACK eats into
3969 * or hits the sequence space of other SACK blocks, if so coalesce.
3971 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
;) {
3972 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
3975 /* Zap SWALK, by moving every further SACK up by one slot.
3976 * Decrease num_sacks.
3978 tp
->rx_opt
.num_sacks
--;
3979 for (i
= this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
3983 this_sack
++, swalk
++;
3987 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
3989 struct tcp_sock
*tp
= tcp_sk(sk
);
3990 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3991 int cur_sacks
= tp
->rx_opt
.num_sacks
;
3997 for (this_sack
= 0; this_sack
< cur_sacks
; this_sack
++, sp
++) {
3998 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
3999 /* Rotate this_sack to the first one. */
4000 for (; this_sack
> 0; this_sack
--, sp
--)
4001 swap(*sp
, *(sp
- 1));
4003 tcp_sack_maybe_coalesce(tp
);
4008 /* Could not find an adjacent existing SACK, build a new one,
4009 * put it at the front, and shift everyone else down. We
4010 * always know there is at least one SACK present already here.
4012 * If the sack array is full, forget about the last one.
4014 if (this_sack
>= TCP_NUM_SACKS
) {
4016 tp
->rx_opt
.num_sacks
--;
4019 for (; this_sack
> 0; this_sack
--, sp
--)
4023 /* Build the new head SACK, and we're done. */
4024 sp
->start_seq
= seq
;
4025 sp
->end_seq
= end_seq
;
4026 tp
->rx_opt
.num_sacks
++;
4029 /* RCV.NXT advances, some SACKs should be eaten. */
4031 static void tcp_sack_remove(struct tcp_sock
*tp
)
4033 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4034 int num_sacks
= tp
->rx_opt
.num_sacks
;
4037 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4038 if (skb_queue_empty(&tp
->out_of_order_queue
)) {
4039 tp
->rx_opt
.num_sacks
= 0;
4043 for (this_sack
= 0; this_sack
< num_sacks
;) {
4044 /* Check if the start of the sack is covered by RCV.NXT. */
4045 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
4048 /* RCV.NXT must cover all the block! */
4049 WARN_ON(before(tp
->rcv_nxt
, sp
->end_seq
));
4051 /* Zap this SACK, by moving forward any other SACKS. */
4052 for (i
= this_sack
+1; i
< num_sacks
; i
++)
4053 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
4060 tp
->rx_opt
.num_sacks
= num_sacks
;
4063 /* This one checks to see if we can put data from the
4064 * out_of_order queue into the receive_queue.
4066 static void tcp_ofo_queue(struct sock
*sk
)
4068 struct tcp_sock
*tp
= tcp_sk(sk
);
4069 __u32 dsack_high
= tp
->rcv_nxt
;
4070 struct sk_buff
*skb
;
4072 while ((skb
= skb_peek(&tp
->out_of_order_queue
)) != NULL
) {
4073 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
4076 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
4077 __u32 dsack
= dsack_high
;
4078 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
4079 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
4080 tcp_dsack_extend(sk
, TCP_SKB_CB(skb
)->seq
, dsack
);
4083 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4084 SOCK_DEBUG(sk
, "ofo packet was already received\n");
4085 __skb_unlink(skb
, &tp
->out_of_order_queue
);
4089 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
4090 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4091 TCP_SKB_CB(skb
)->end_seq
);
4093 __skb_unlink(skb
, &tp
->out_of_order_queue
);
4094 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4095 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4096 if (tcp_hdr(skb
)->fin
)
4101 static bool tcp_prune_ofo_queue(struct sock
*sk
);
4102 static int tcp_prune_queue(struct sock
*sk
);
4104 static int tcp_try_rmem_schedule(struct sock
*sk
, struct sk_buff
*skb
,
4107 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
4108 !sk_rmem_schedule(sk
, skb
, size
)) {
4110 if (tcp_prune_queue(sk
) < 0)
4113 if (!sk_rmem_schedule(sk
, skb
, size
)) {
4114 if (!tcp_prune_ofo_queue(sk
))
4117 if (!sk_rmem_schedule(sk
, skb
, size
))
4125 * tcp_try_coalesce - try to merge skb to prior one
4128 * @from: buffer to add in queue
4129 * @fragstolen: pointer to boolean
4131 * Before queueing skb @from after @to, try to merge them
4132 * to reduce overall memory use and queue lengths, if cost is small.
4133 * Packets in ofo or receive queues can stay a long time.
4134 * Better try to coalesce them right now to avoid future collapses.
4135 * Returns true if caller should free @from instead of queueing it
4137 static bool tcp_try_coalesce(struct sock
*sk
,
4139 struct sk_buff
*from
,
4144 *fragstolen
= false;
4146 if (tcp_hdr(from
)->fin
)
4149 /* Its possible this segment overlaps with prior segment in queue */
4150 if (TCP_SKB_CB(from
)->seq
!= TCP_SKB_CB(to
)->end_seq
)
4153 if (!skb_try_coalesce(to
, from
, fragstolen
, &delta
))
4156 atomic_add(delta
, &sk
->sk_rmem_alloc
);
4157 sk_mem_charge(sk
, delta
);
4158 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRCVCOALESCE
);
4159 TCP_SKB_CB(to
)->end_seq
= TCP_SKB_CB(from
)->end_seq
;
4160 TCP_SKB_CB(to
)->ack_seq
= TCP_SKB_CB(from
)->ack_seq
;
4164 static void tcp_data_queue_ofo(struct sock
*sk
, struct sk_buff
*skb
)
4166 struct tcp_sock
*tp
= tcp_sk(sk
);
4167 struct sk_buff
*skb1
;
4170 TCP_ECN_check_ce(tp
, skb
);
4172 if (unlikely(tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))) {
4173 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFODROP
);
4178 /* Disable header prediction. */
4180 inet_csk_schedule_ack(sk
);
4182 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFOQUEUE
);
4183 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
4184 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4186 skb1
= skb_peek_tail(&tp
->out_of_order_queue
);
4188 /* Initial out of order segment, build 1 SACK. */
4189 if (tcp_is_sack(tp
)) {
4190 tp
->rx_opt
.num_sacks
= 1;
4191 tp
->selective_acks
[0].start_seq
= TCP_SKB_CB(skb
)->seq
;
4192 tp
->selective_acks
[0].end_seq
=
4193 TCP_SKB_CB(skb
)->end_seq
;
4195 __skb_queue_head(&tp
->out_of_order_queue
, skb
);
4199 seq
= TCP_SKB_CB(skb
)->seq
;
4200 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4202 if (seq
== TCP_SKB_CB(skb1
)->end_seq
) {
4205 if (!tcp_try_coalesce(sk
, skb1
, skb
, &fragstolen
)) {
4206 __skb_queue_after(&tp
->out_of_order_queue
, skb1
, skb
);
4208 tcp_grow_window(sk
, skb
);
4209 kfree_skb_partial(skb
, fragstolen
);
4213 if (!tp
->rx_opt
.num_sacks
||
4214 tp
->selective_acks
[0].end_seq
!= seq
)
4217 /* Common case: data arrive in order after hole. */
4218 tp
->selective_acks
[0].end_seq
= end_seq
;
4222 /* Find place to insert this segment. */
4224 if (!after(TCP_SKB_CB(skb1
)->seq
, seq
))
4226 if (skb_queue_is_first(&tp
->out_of_order_queue
, skb1
)) {
4230 skb1
= skb_queue_prev(&tp
->out_of_order_queue
, skb1
);
4233 /* Do skb overlap to previous one? */
4234 if (skb1
&& before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4235 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4236 /* All the bits are present. Drop. */
4237 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFOMERGE
);
4240 tcp_dsack_set(sk
, seq
, end_seq
);
4243 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
4244 /* Partial overlap. */
4245 tcp_dsack_set(sk
, seq
,
4246 TCP_SKB_CB(skb1
)->end_seq
);
4248 if (skb_queue_is_first(&tp
->out_of_order_queue
,
4252 skb1
= skb_queue_prev(
4253 &tp
->out_of_order_queue
,
4258 __skb_queue_head(&tp
->out_of_order_queue
, skb
);
4260 __skb_queue_after(&tp
->out_of_order_queue
, skb1
, skb
);
4262 /* And clean segments covered by new one as whole. */
4263 while (!skb_queue_is_last(&tp
->out_of_order_queue
, skb
)) {
4264 skb1
= skb_queue_next(&tp
->out_of_order_queue
, skb
);
4266 if (!after(end_seq
, TCP_SKB_CB(skb1
)->seq
))
4268 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4269 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4273 __skb_unlink(skb1
, &tp
->out_of_order_queue
);
4274 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4275 TCP_SKB_CB(skb1
)->end_seq
);
4276 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFOMERGE
);
4281 if (tcp_is_sack(tp
))
4282 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
4285 tcp_grow_window(sk
, skb
);
4286 skb_set_owner_r(skb
, sk
);
4290 static int __must_check
tcp_queue_rcv(struct sock
*sk
, struct sk_buff
*skb
, int hdrlen
,
4294 struct sk_buff
*tail
= skb_peek_tail(&sk
->sk_receive_queue
);
4296 __skb_pull(skb
, hdrlen
);
4298 tcp_try_coalesce(sk
, tail
, skb
, fragstolen
)) ? 1 : 0;
4299 tcp_sk(sk
)->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4301 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4302 skb_set_owner_r(skb
, sk
);
4307 int tcp_send_rcvq(struct sock
*sk
, struct msghdr
*msg
, size_t size
)
4309 struct sk_buff
*skb
= NULL
;
4316 skb
= alloc_skb(size
+ sizeof(*th
), sk
->sk_allocation
);
4320 if (tcp_try_rmem_schedule(sk
, skb
, size
+ sizeof(*th
)))
4323 th
= (struct tcphdr
*)skb_put(skb
, sizeof(*th
));
4324 skb_reset_transport_header(skb
);
4325 memset(th
, 0, sizeof(*th
));
4327 if (memcpy_fromiovec(skb_put(skb
, size
), msg
->msg_iov
, size
))
4330 TCP_SKB_CB(skb
)->seq
= tcp_sk(sk
)->rcv_nxt
;
4331 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ size
;
4332 TCP_SKB_CB(skb
)->ack_seq
= tcp_sk(sk
)->snd_una
- 1;
4334 if (tcp_queue_rcv(sk
, skb
, sizeof(*th
), &fragstolen
)) {
4335 WARN_ON_ONCE(fragstolen
); /* should not happen */
4346 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
4348 const struct tcphdr
*th
= tcp_hdr(skb
);
4349 struct tcp_sock
*tp
= tcp_sk(sk
);
4351 bool fragstolen
= false;
4353 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
)
4357 __skb_pull(skb
, th
->doff
* 4);
4359 TCP_ECN_accept_cwr(tp
, skb
);
4361 tp
->rx_opt
.dsack
= 0;
4363 /* Queue data for delivery to the user.
4364 * Packets in sequence go to the receive queue.
4365 * Out of sequence packets to the out_of_order_queue.
4367 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
4368 if (tcp_receive_window(tp
) == 0)
4371 /* Ok. In sequence. In window. */
4372 if (tp
->ucopy
.task
== current
&&
4373 tp
->copied_seq
== tp
->rcv_nxt
&& tp
->ucopy
.len
&&
4374 sock_owned_by_user(sk
) && !tp
->urg_data
) {
4375 int chunk
= min_t(unsigned int, skb
->len
,
4378 __set_current_state(TASK_RUNNING
);
4381 if (!skb_copy_datagram_iovec(skb
, 0, tp
->ucopy
.iov
, chunk
)) {
4382 tp
->ucopy
.len
-= chunk
;
4383 tp
->copied_seq
+= chunk
;
4384 eaten
= (chunk
== skb
->len
);
4385 tcp_rcv_space_adjust(sk
);
4393 tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))
4396 eaten
= tcp_queue_rcv(sk
, skb
, 0, &fragstolen
);
4398 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4400 tcp_event_data_recv(sk
, skb
);
4404 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4407 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4408 * gap in queue is filled.
4410 if (skb_queue_empty(&tp
->out_of_order_queue
))
4411 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
4414 if (tp
->rx_opt
.num_sacks
)
4415 tcp_sack_remove(tp
);
4417 tcp_fast_path_check(sk
);
4420 kfree_skb_partial(skb
, fragstolen
);
4421 if (!sock_flag(sk
, SOCK_DEAD
))
4422 sk
->sk_data_ready(sk
, 0);
4426 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4427 /* A retransmit, 2nd most common case. Force an immediate ack. */
4428 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4429 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4432 tcp_enter_quickack_mode(sk
);
4433 inet_csk_schedule_ack(sk
);
4439 /* Out of window. F.e. zero window probe. */
4440 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
4443 tcp_enter_quickack_mode(sk
);
4445 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4446 /* Partial packet, seq < rcv_next < end_seq */
4447 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
4448 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4449 TCP_SKB_CB(skb
)->end_seq
);
4451 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
4453 /* If window is closed, drop tail of packet. But after
4454 * remembering D-SACK for its head made in previous line.
4456 if (!tcp_receive_window(tp
))
4461 tcp_data_queue_ofo(sk
, skb
);
4464 static struct sk_buff
*tcp_collapse_one(struct sock
*sk
, struct sk_buff
*skb
,
4465 struct sk_buff_head
*list
)
4467 struct sk_buff
*next
= NULL
;
4469 if (!skb_queue_is_last(list
, skb
))
4470 next
= skb_queue_next(list
, skb
);
4472 __skb_unlink(skb
, list
);
4474 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRCVCOLLAPSED
);
4479 /* Collapse contiguous sequence of skbs head..tail with
4480 * sequence numbers start..end.
4482 * If tail is NULL, this means until the end of the list.
4484 * Segments with FIN/SYN are not collapsed (only because this
4488 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
,
4489 struct sk_buff
*head
, struct sk_buff
*tail
,
4492 struct sk_buff
*skb
, *n
;
4495 /* First, check that queue is collapsible and find
4496 * the point where collapsing can be useful. */
4500 skb_queue_walk_from_safe(list
, skb
, n
) {
4503 /* No new bits? It is possible on ofo queue. */
4504 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4505 skb
= tcp_collapse_one(sk
, skb
, list
);
4511 /* The first skb to collapse is:
4513 * - bloated or contains data before "start" or
4514 * overlaps to the next one.
4516 if (!tcp_hdr(skb
)->syn
&& !tcp_hdr(skb
)->fin
&&
4517 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
4518 before(TCP_SKB_CB(skb
)->seq
, start
))) {
4519 end_of_skbs
= false;
4523 if (!skb_queue_is_last(list
, skb
)) {
4524 struct sk_buff
*next
= skb_queue_next(list
, skb
);
4526 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(next
)->seq
) {
4527 end_of_skbs
= false;
4532 /* Decided to skip this, advance start seq. */
4533 start
= TCP_SKB_CB(skb
)->end_seq
;
4535 if (end_of_skbs
|| tcp_hdr(skb
)->syn
|| tcp_hdr(skb
)->fin
)
4538 while (before(start
, end
)) {
4539 struct sk_buff
*nskb
;
4540 unsigned int header
= skb_headroom(skb
);
4541 int copy
= SKB_MAX_ORDER(header
, 0);
4543 /* Too big header? This can happen with IPv6. */
4546 if (end
- start
< copy
)
4548 nskb
= alloc_skb(copy
+ header
, GFP_ATOMIC
);
4552 skb_set_mac_header(nskb
, skb_mac_header(skb
) - skb
->head
);
4553 skb_set_network_header(nskb
, (skb_network_header(skb
) -
4555 skb_set_transport_header(nskb
, (skb_transport_header(skb
) -
4557 skb_reserve(nskb
, header
);
4558 memcpy(nskb
->head
, skb
->head
, header
);
4559 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
4560 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
4561 __skb_queue_before(list
, skb
, nskb
);
4562 skb_set_owner_r(nskb
, sk
);
4564 /* Copy data, releasing collapsed skbs. */
4566 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
4567 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
4571 size
= min(copy
, size
);
4572 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
4574 TCP_SKB_CB(nskb
)->end_seq
+= size
;
4578 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4579 skb
= tcp_collapse_one(sk
, skb
, list
);
4582 tcp_hdr(skb
)->syn
||
4590 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4591 * and tcp_collapse() them until all the queue is collapsed.
4593 static void tcp_collapse_ofo_queue(struct sock
*sk
)
4595 struct tcp_sock
*tp
= tcp_sk(sk
);
4596 struct sk_buff
*skb
= skb_peek(&tp
->out_of_order_queue
);
4597 struct sk_buff
*head
;
4603 start
= TCP_SKB_CB(skb
)->seq
;
4604 end
= TCP_SKB_CB(skb
)->end_seq
;
4608 struct sk_buff
*next
= NULL
;
4610 if (!skb_queue_is_last(&tp
->out_of_order_queue
, skb
))
4611 next
= skb_queue_next(&tp
->out_of_order_queue
, skb
);
4614 /* Segment is terminated when we see gap or when
4615 * we are at the end of all the queue. */
4617 after(TCP_SKB_CB(skb
)->seq
, end
) ||
4618 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
4619 tcp_collapse(sk
, &tp
->out_of_order_queue
,
4620 head
, skb
, start
, end
);
4624 /* Start new segment */
4625 start
= TCP_SKB_CB(skb
)->seq
;
4626 end
= TCP_SKB_CB(skb
)->end_seq
;
4628 if (before(TCP_SKB_CB(skb
)->seq
, start
))
4629 start
= TCP_SKB_CB(skb
)->seq
;
4630 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
4631 end
= TCP_SKB_CB(skb
)->end_seq
;
4637 * Purge the out-of-order queue.
4638 * Return true if queue was pruned.
4640 static bool tcp_prune_ofo_queue(struct sock
*sk
)
4642 struct tcp_sock
*tp
= tcp_sk(sk
);
4645 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4646 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_OFOPRUNED
);
4647 __skb_queue_purge(&tp
->out_of_order_queue
);
4649 /* Reset SACK state. A conforming SACK implementation will
4650 * do the same at a timeout based retransmit. When a connection
4651 * is in a sad state like this, we care only about integrity
4652 * of the connection not performance.
4654 if (tp
->rx_opt
.sack_ok
)
4655 tcp_sack_reset(&tp
->rx_opt
);
4662 /* Reduce allocated memory if we can, trying to get
4663 * the socket within its memory limits again.
4665 * Return less than zero if we should start dropping frames
4666 * until the socket owning process reads some of the data
4667 * to stabilize the situation.
4669 static int tcp_prune_queue(struct sock
*sk
)
4671 struct tcp_sock
*tp
= tcp_sk(sk
);
4673 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
4675 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PRUNECALLED
);
4677 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
4678 tcp_clamp_window(sk
);
4679 else if (sk_under_memory_pressure(sk
))
4680 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
4682 tcp_collapse_ofo_queue(sk
);
4683 if (!skb_queue_empty(&sk
->sk_receive_queue
))
4684 tcp_collapse(sk
, &sk
->sk_receive_queue
,
4685 skb_peek(&sk
->sk_receive_queue
),
4687 tp
->copied_seq
, tp
->rcv_nxt
);
4690 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4693 /* Collapsing did not help, destructive actions follow.
4694 * This must not ever occur. */
4696 tcp_prune_ofo_queue(sk
);
4698 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4701 /* If we are really being abused, tell the caller to silently
4702 * drop receive data on the floor. It will get retransmitted
4703 * and hopefully then we'll have sufficient space.
4705 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_RCVPRUNED
);
4707 /* Massive buffer overcommit. */
4712 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4713 * As additional protections, we do not touch cwnd in retransmission phases,
4714 * and if application hit its sndbuf limit recently.
4716 void tcp_cwnd_application_limited(struct sock
*sk
)
4718 struct tcp_sock
*tp
= tcp_sk(sk
);
4720 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
4721 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
4722 /* Limited by application or receiver window. */
4723 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
4724 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
4725 if (win_used
< tp
->snd_cwnd
) {
4726 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
4727 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
4729 tp
->snd_cwnd_used
= 0;
4731 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4734 static bool tcp_should_expand_sndbuf(const struct sock
*sk
)
4736 const struct tcp_sock
*tp
= tcp_sk(sk
);
4738 /* If the user specified a specific send buffer setting, do
4741 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
4744 /* If we are under global TCP memory pressure, do not expand. */
4745 if (sk_under_memory_pressure(sk
))
4748 /* If we are under soft global TCP memory pressure, do not expand. */
4749 if (sk_memory_allocated(sk
) >= sk_prot_mem_limits(sk
, 0))
4752 /* If we filled the congestion window, do not expand. */
4753 if (tp
->packets_out
>= tp
->snd_cwnd
)
4759 /* When incoming ACK allowed to free some skb from write_queue,
4760 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4761 * on the exit from tcp input handler.
4763 * PROBLEM: sndbuf expansion does not work well with largesend.
4765 static void tcp_new_space(struct sock
*sk
)
4767 struct tcp_sock
*tp
= tcp_sk(sk
);
4769 if (tcp_should_expand_sndbuf(sk
)) {
4770 tcp_sndbuf_expand(sk
);
4771 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4774 sk
->sk_write_space(sk
);
4777 static void tcp_check_space(struct sock
*sk
)
4779 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
4780 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
4781 if (sk
->sk_socket
&&
4782 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
4787 static inline void tcp_data_snd_check(struct sock
*sk
)
4789 tcp_push_pending_frames(sk
);
4790 tcp_check_space(sk
);
4794 * Check if sending an ack is needed.
4796 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
4798 struct tcp_sock
*tp
= tcp_sk(sk
);
4800 /* More than one full frame received... */
4801 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
&&
4802 /* ... and right edge of window advances far enough.
4803 * (tcp_recvmsg() will send ACK otherwise). Or...
4805 __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
4806 /* We ACK each frame or... */
4807 tcp_in_quickack_mode(sk
) ||
4808 /* We have out of order data. */
4809 (ofo_possible
&& skb_peek(&tp
->out_of_order_queue
))) {
4810 /* Then ack it now */
4813 /* Else, send delayed ack. */
4814 tcp_send_delayed_ack(sk
);
4818 static inline void tcp_ack_snd_check(struct sock
*sk
)
4820 if (!inet_csk_ack_scheduled(sk
)) {
4821 /* We sent a data segment already. */
4824 __tcp_ack_snd_check(sk
, 1);
4828 * This routine is only called when we have urgent data
4829 * signaled. Its the 'slow' part of tcp_urg. It could be
4830 * moved inline now as tcp_urg is only called from one
4831 * place. We handle URGent data wrong. We have to - as
4832 * BSD still doesn't use the correction from RFC961.
4833 * For 1003.1g we should support a new option TCP_STDURG to permit
4834 * either form (or just set the sysctl tcp_stdurg).
4837 static void tcp_check_urg(struct sock
*sk
, const struct tcphdr
*th
)
4839 struct tcp_sock
*tp
= tcp_sk(sk
);
4840 u32 ptr
= ntohs(th
->urg_ptr
);
4842 if (ptr
&& !sysctl_tcp_stdurg
)
4844 ptr
+= ntohl(th
->seq
);
4846 /* Ignore urgent data that we've already seen and read. */
4847 if (after(tp
->copied_seq
, ptr
))
4850 /* Do not replay urg ptr.
4852 * NOTE: interesting situation not covered by specs.
4853 * Misbehaving sender may send urg ptr, pointing to segment,
4854 * which we already have in ofo queue. We are not able to fetch
4855 * such data and will stay in TCP_URG_NOTYET until will be eaten
4856 * by recvmsg(). Seems, we are not obliged to handle such wicked
4857 * situations. But it is worth to think about possibility of some
4858 * DoSes using some hypothetical application level deadlock.
4860 if (before(ptr
, tp
->rcv_nxt
))
4863 /* Do we already have a newer (or duplicate) urgent pointer? */
4864 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
4867 /* Tell the world about our new urgent pointer. */
4870 /* We may be adding urgent data when the last byte read was
4871 * urgent. To do this requires some care. We cannot just ignore
4872 * tp->copied_seq since we would read the last urgent byte again
4873 * as data, nor can we alter copied_seq until this data arrives
4874 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4876 * NOTE. Double Dutch. Rendering to plain English: author of comment
4877 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4878 * and expect that both A and B disappear from stream. This is _wrong_.
4879 * Though this happens in BSD with high probability, this is occasional.
4880 * Any application relying on this is buggy. Note also, that fix "works"
4881 * only in this artificial test. Insert some normal data between A and B and we will
4882 * decline of BSD again. Verdict: it is better to remove to trap
4885 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
4886 !sock_flag(sk
, SOCK_URGINLINE
) && tp
->copied_seq
!= tp
->rcv_nxt
) {
4887 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
4889 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4890 __skb_unlink(skb
, &sk
->sk_receive_queue
);
4895 tp
->urg_data
= TCP_URG_NOTYET
;
4898 /* Disable header prediction. */
4902 /* This is the 'fast' part of urgent handling. */
4903 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, const struct tcphdr
*th
)
4905 struct tcp_sock
*tp
= tcp_sk(sk
);
4907 /* Check if we get a new urgent pointer - normally not. */
4909 tcp_check_urg(sk
, th
);
4911 /* Do we wait for any urgent data? - normally not... */
4912 if (tp
->urg_data
== TCP_URG_NOTYET
) {
4913 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
4916 /* Is the urgent pointer pointing into this packet? */
4917 if (ptr
< skb
->len
) {
4919 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
4921 tp
->urg_data
= TCP_URG_VALID
| tmp
;
4922 if (!sock_flag(sk
, SOCK_DEAD
))
4923 sk
->sk_data_ready(sk
, 0);
4928 static int tcp_copy_to_iovec(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
4930 struct tcp_sock
*tp
= tcp_sk(sk
);
4931 int chunk
= skb
->len
- hlen
;
4935 if (skb_csum_unnecessary(skb
))
4936 err
= skb_copy_datagram_iovec(skb
, hlen
, tp
->ucopy
.iov
, chunk
);
4938 err
= skb_copy_and_csum_datagram_iovec(skb
, hlen
,
4942 tp
->ucopy
.len
-= chunk
;
4943 tp
->copied_seq
+= chunk
;
4944 tcp_rcv_space_adjust(sk
);
4951 static __sum16
__tcp_checksum_complete_user(struct sock
*sk
,
4952 struct sk_buff
*skb
)
4956 if (sock_owned_by_user(sk
)) {
4958 result
= __tcp_checksum_complete(skb
);
4961 result
= __tcp_checksum_complete(skb
);
4966 static inline bool tcp_checksum_complete_user(struct sock
*sk
,
4967 struct sk_buff
*skb
)
4969 return !skb_csum_unnecessary(skb
) &&
4970 __tcp_checksum_complete_user(sk
, skb
);
4973 #ifdef CONFIG_NET_DMA
4974 static bool tcp_dma_try_early_copy(struct sock
*sk
, struct sk_buff
*skb
,
4977 struct tcp_sock
*tp
= tcp_sk(sk
);
4978 int chunk
= skb
->len
- hlen
;
4980 bool copied_early
= false;
4982 if (tp
->ucopy
.wakeup
)
4985 if (!tp
->ucopy
.dma_chan
&& tp
->ucopy
.pinned_list
)
4986 tp
->ucopy
.dma_chan
= net_dma_find_channel();
4988 if (tp
->ucopy
.dma_chan
&& skb_csum_unnecessary(skb
)) {
4990 dma_cookie
= dma_skb_copy_datagram_iovec(tp
->ucopy
.dma_chan
,
4992 tp
->ucopy
.iov
, chunk
,
4993 tp
->ucopy
.pinned_list
);
4998 tp
->ucopy
.dma_cookie
= dma_cookie
;
4999 copied_early
= true;
5001 tp
->ucopy
.len
-= chunk
;
5002 tp
->copied_seq
+= chunk
;
5003 tcp_rcv_space_adjust(sk
);
5005 if ((tp
->ucopy
.len
== 0) ||
5006 (tcp_flag_word(tcp_hdr(skb
)) & TCP_FLAG_PSH
) ||
5007 (atomic_read(&sk
->sk_rmem_alloc
) > (sk
->sk_rcvbuf
>> 1))) {
5008 tp
->ucopy
.wakeup
= 1;
5009 sk
->sk_data_ready(sk
, 0);
5011 } else if (chunk
> 0) {
5012 tp
->ucopy
.wakeup
= 1;
5013 sk
->sk_data_ready(sk
, 0);
5016 return copied_early
;
5018 #endif /* CONFIG_NET_DMA */
5020 /* Does PAWS and seqno based validation of an incoming segment, flags will
5021 * play significant role here.
5023 static bool tcp_validate_incoming(struct sock
*sk
, struct sk_buff
*skb
,
5024 const struct tcphdr
*th
, int syn_inerr
)
5026 struct tcp_sock
*tp
= tcp_sk(sk
);
5028 /* RFC1323: H1. Apply PAWS check first. */
5029 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
5030 tcp_paws_discard(sk
, skb
)) {
5032 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSESTABREJECTED
);
5033 tcp_send_dupack(sk
, skb
);
5036 /* Reset is accepted even if it did not pass PAWS. */
5039 /* Step 1: check sequence number */
5040 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5041 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5042 * (RST) segments are validated by checking their SEQ-fields."
5043 * And page 69: "If an incoming segment is not acceptable,
5044 * an acknowledgment should be sent in reply (unless the RST
5045 * bit is set, if so drop the segment and return)".
5050 tcp_send_dupack(sk
, skb
);
5055 /* Step 2: check RST bit */
5058 * If sequence number exactly matches RCV.NXT, then
5059 * RESET the connection
5061 * Send a challenge ACK
5063 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
)
5066 tcp_send_challenge_ack(sk
);
5070 /* step 3: check security and precedence [ignored] */
5072 /* step 4: Check for a SYN
5073 * RFC 5691 4.2 : Send a challenge ack
5078 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5079 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSYNCHALLENGE
);
5080 tcp_send_challenge_ack(sk
);
5092 * TCP receive function for the ESTABLISHED state.
5094 * It is split into a fast path and a slow path. The fast path is
5096 * - A zero window was announced from us - zero window probing
5097 * is only handled properly in the slow path.
5098 * - Out of order segments arrived.
5099 * - Urgent data is expected.
5100 * - There is no buffer space left
5101 * - Unexpected TCP flags/window values/header lengths are received
5102 * (detected by checking the TCP header against pred_flags)
5103 * - Data is sent in both directions. Fast path only supports pure senders
5104 * or pure receivers (this means either the sequence number or the ack
5105 * value must stay constant)
5106 * - Unexpected TCP option.
5108 * When these conditions are not satisfied it drops into a standard
5109 * receive procedure patterned after RFC793 to handle all cases.
5110 * The first three cases are guaranteed by proper pred_flags setting,
5111 * the rest is checked inline. Fast processing is turned on in
5112 * tcp_data_queue when everything is OK.
5114 void tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
5115 const struct tcphdr
*th
, unsigned int len
)
5117 struct tcp_sock
*tp
= tcp_sk(sk
);
5119 if (unlikely(sk
->sk_rx_dst
== NULL
))
5120 inet_csk(sk
)->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
5122 * Header prediction.
5123 * The code loosely follows the one in the famous
5124 * "30 instruction TCP receive" Van Jacobson mail.
5126 * Van's trick is to deposit buffers into socket queue
5127 * on a device interrupt, to call tcp_recv function
5128 * on the receive process context and checksum and copy
5129 * the buffer to user space. smart...
5131 * Our current scheme is not silly either but we take the
5132 * extra cost of the net_bh soft interrupt processing...
5133 * We do checksum and copy also but from device to kernel.
5136 tp
->rx_opt
.saw_tstamp
= 0;
5138 /* pred_flags is 0xS?10 << 16 + snd_wnd
5139 * if header_prediction is to be made
5140 * 'S' will always be tp->tcp_header_len >> 2
5141 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5142 * turn it off (when there are holes in the receive
5143 * space for instance)
5144 * PSH flag is ignored.
5147 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
5148 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
&&
5149 !after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
)) {
5150 int tcp_header_len
= tp
->tcp_header_len
;
5152 /* Timestamp header prediction: tcp_header_len
5153 * is automatically equal to th->doff*4 due to pred_flags
5157 /* Check timestamp */
5158 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
5159 /* No? Slow path! */
5160 if (!tcp_parse_aligned_timestamp(tp
, th
))
5163 /* If PAWS failed, check it more carefully in slow path */
5164 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
5167 /* DO NOT update ts_recent here, if checksum fails
5168 * and timestamp was corrupted part, it will result
5169 * in a hung connection since we will drop all
5170 * future packets due to the PAWS test.
5174 if (len
<= tcp_header_len
) {
5175 /* Bulk data transfer: sender */
5176 if (len
== tcp_header_len
) {
5177 /* Predicted packet is in window by definition.
5178 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5179 * Hence, check seq<=rcv_wup reduces to:
5181 if (tcp_header_len
==
5182 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5183 tp
->rcv_nxt
== tp
->rcv_wup
)
5184 tcp_store_ts_recent(tp
);
5186 /* We know that such packets are checksummed
5189 tcp_ack(sk
, skb
, 0);
5191 tcp_data_snd_check(sk
);
5193 } else { /* Header too small */
5194 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5199 int copied_early
= 0;
5200 bool fragstolen
= false;
5202 if (tp
->copied_seq
== tp
->rcv_nxt
&&
5203 len
- tcp_header_len
<= tp
->ucopy
.len
) {
5204 #ifdef CONFIG_NET_DMA
5205 if (tp
->ucopy
.task
== current
&&
5206 sock_owned_by_user(sk
) &&
5207 tcp_dma_try_early_copy(sk
, skb
, tcp_header_len
)) {
5212 if (tp
->ucopy
.task
== current
&&
5213 sock_owned_by_user(sk
) && !copied_early
) {
5214 __set_current_state(TASK_RUNNING
);
5216 if (!tcp_copy_to_iovec(sk
, skb
, tcp_header_len
))
5220 /* Predicted packet is in window by definition.
5221 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5222 * Hence, check seq<=rcv_wup reduces to:
5224 if (tcp_header_len
==
5225 (sizeof(struct tcphdr
) +
5226 TCPOLEN_TSTAMP_ALIGNED
) &&
5227 tp
->rcv_nxt
== tp
->rcv_wup
)
5228 tcp_store_ts_recent(tp
);
5230 tcp_rcv_rtt_measure_ts(sk
, skb
);
5232 __skb_pull(skb
, tcp_header_len
);
5233 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
5234 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITSTOUSER
);
5237 tcp_cleanup_rbuf(sk
, skb
->len
);
5240 if (tcp_checksum_complete_user(sk
, skb
))
5243 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
5246 /* Predicted packet is in window by definition.
5247 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5248 * Hence, check seq<=rcv_wup reduces to:
5250 if (tcp_header_len
==
5251 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5252 tp
->rcv_nxt
== tp
->rcv_wup
)
5253 tcp_store_ts_recent(tp
);
5255 tcp_rcv_rtt_measure_ts(sk
, skb
);
5257 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITS
);
5259 /* Bulk data transfer: receiver */
5260 eaten
= tcp_queue_rcv(sk
, skb
, tcp_header_len
,
5264 tcp_event_data_recv(sk
, skb
);
5266 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
5267 /* Well, only one small jumplet in fast path... */
5268 tcp_ack(sk
, skb
, FLAG_DATA
);
5269 tcp_data_snd_check(sk
);
5270 if (!inet_csk_ack_scheduled(sk
))
5274 if (!copied_early
|| tp
->rcv_nxt
!= tp
->rcv_wup
)
5275 __tcp_ack_snd_check(sk
, 0);
5277 #ifdef CONFIG_NET_DMA
5279 __skb_queue_tail(&sk
->sk_async_wait_queue
, skb
);
5283 kfree_skb_partial(skb
, fragstolen
);
5284 sk
->sk_data_ready(sk
, 0);
5290 if (len
< (th
->doff
<< 2) || tcp_checksum_complete_user(sk
, skb
))
5293 if (!th
->ack
&& !th
->rst
)
5297 * Standard slow path.
5300 if (!tcp_validate_incoming(sk
, skb
, th
, 1))
5304 if (tcp_ack(sk
, skb
, FLAG_SLOWPATH
| FLAG_UPDATE_TS_RECENT
) < 0)
5307 tcp_rcv_rtt_measure_ts(sk
, skb
);
5309 /* Process urgent data. */
5310 tcp_urg(sk
, skb
, th
);
5312 /* step 7: process the segment text */
5313 tcp_data_queue(sk
, skb
);
5315 tcp_data_snd_check(sk
);
5316 tcp_ack_snd_check(sk
);
5320 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_CSUMERRORS
);
5321 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5326 EXPORT_SYMBOL(tcp_rcv_established
);
5328 void tcp_finish_connect(struct sock
*sk
, struct sk_buff
*skb
)
5330 struct tcp_sock
*tp
= tcp_sk(sk
);
5331 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5333 tcp_set_state(sk
, TCP_ESTABLISHED
);
5336 icsk
->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
5337 security_inet_conn_established(sk
, skb
);
5340 /* Make sure socket is routed, for correct metrics. */
5341 icsk
->icsk_af_ops
->rebuild_header(sk
);
5343 tcp_init_metrics(sk
);
5345 tcp_init_congestion_control(sk
);
5347 /* Prevent spurious tcp_cwnd_restart() on first data
5350 tp
->lsndtime
= tcp_time_stamp
;
5352 tcp_init_buffer_space(sk
);
5354 if (sock_flag(sk
, SOCK_KEEPOPEN
))
5355 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
5357 if (!tp
->rx_opt
.snd_wscale
)
5358 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
5362 if (!sock_flag(sk
, SOCK_DEAD
)) {
5363 sk
->sk_state_change(sk
);
5364 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5368 static bool tcp_rcv_fastopen_synack(struct sock
*sk
, struct sk_buff
*synack
,
5369 struct tcp_fastopen_cookie
*cookie
)
5371 struct tcp_sock
*tp
= tcp_sk(sk
);
5372 struct sk_buff
*data
= tp
->syn_data
? tcp_write_queue_head(sk
) : NULL
;
5373 u16 mss
= tp
->rx_opt
.mss_clamp
;
5376 if (mss
== tp
->rx_opt
.user_mss
) {
5377 struct tcp_options_received opt
;
5379 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5380 tcp_clear_options(&opt
);
5381 opt
.user_mss
= opt
.mss_clamp
= 0;
5382 tcp_parse_options(synack
, &opt
, 0, NULL
);
5383 mss
= opt
.mss_clamp
;
5386 if (!tp
->syn_fastopen
) /* Ignore an unsolicited cookie */
5389 /* The SYN-ACK neither has cookie nor acknowledges the data. Presumably
5390 * the remote receives only the retransmitted (regular) SYNs: either
5391 * the original SYN-data or the corresponding SYN-ACK is lost.
5393 syn_drop
= (cookie
->len
<= 0 && data
&& tp
->total_retrans
);
5395 tcp_fastopen_cache_set(sk
, mss
, cookie
, syn_drop
);
5397 if (data
) { /* Retransmit unacked data in SYN */
5398 tcp_for_write_queue_from(data
, sk
) {
5399 if (data
== tcp_send_head(sk
) ||
5400 __tcp_retransmit_skb(sk
, data
))
5406 tp
->syn_data_acked
= tp
->syn_data
;
5410 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5411 const struct tcphdr
*th
, unsigned int len
)
5413 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5414 struct tcp_sock
*tp
= tcp_sk(sk
);
5415 struct tcp_fastopen_cookie foc
= { .len
= -1 };
5416 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
5418 tcp_parse_options(skb
, &tp
->rx_opt
, 0, &foc
);
5419 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
5420 tp
->rx_opt
.rcv_tsecr
-= tp
->tsoffset
;
5424 * "If the state is SYN-SENT then
5425 * first check the ACK bit
5426 * If the ACK bit is set
5427 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5428 * a reset (unless the RST bit is set, if so drop
5429 * the segment and return)"
5431 if (!after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_una
) ||
5432 after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
))
5433 goto reset_and_undo
;
5435 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
5436 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
5438 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSACTIVEREJECTED
);
5439 goto reset_and_undo
;
5442 /* Now ACK is acceptable.
5444 * "If the RST bit is set
5445 * If the ACK was acceptable then signal the user "error:
5446 * connection reset", drop the segment, enter CLOSED state,
5447 * delete TCB, and return."
5456 * "fifth, if neither of the SYN or RST bits is set then
5457 * drop the segment and return."
5463 goto discard_and_undo
;
5466 * "If the SYN bit is on ...
5467 * are acceptable then ...
5468 * (our SYN has been ACKed), change the connection
5469 * state to ESTABLISHED..."
5472 TCP_ECN_rcv_synack(tp
, th
);
5474 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5475 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
5477 /* Ok.. it's good. Set up sequence numbers and
5478 * move to established.
5480 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5481 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5483 /* RFC1323: The window in SYN & SYN/ACK segments is
5486 tp
->snd_wnd
= ntohs(th
->window
);
5488 if (!tp
->rx_opt
.wscale_ok
) {
5489 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
5490 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
5493 if (tp
->rx_opt
.saw_tstamp
) {
5494 tp
->rx_opt
.tstamp_ok
= 1;
5495 tp
->tcp_header_len
=
5496 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5497 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5498 tcp_store_ts_recent(tp
);
5500 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5503 if (tcp_is_sack(tp
) && sysctl_tcp_fack
)
5504 tcp_enable_fack(tp
);
5507 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5508 tcp_initialize_rcv_mss(sk
);
5510 /* Remember, tcp_poll() does not lock socket!
5511 * Change state from SYN-SENT only after copied_seq
5512 * is initialized. */
5513 tp
->copied_seq
= tp
->rcv_nxt
;
5517 tcp_finish_connect(sk
, skb
);
5519 if ((tp
->syn_fastopen
|| tp
->syn_data
) &&
5520 tcp_rcv_fastopen_synack(sk
, skb
, &foc
))
5523 if (sk
->sk_write_pending
||
5524 icsk
->icsk_accept_queue
.rskq_defer_accept
||
5525 icsk
->icsk_ack
.pingpong
) {
5526 /* Save one ACK. Data will be ready after
5527 * several ticks, if write_pending is set.
5529 * It may be deleted, but with this feature tcpdumps
5530 * look so _wonderfully_ clever, that I was not able
5531 * to stand against the temptation 8) --ANK
5533 inet_csk_schedule_ack(sk
);
5534 icsk
->icsk_ack
.lrcvtime
= tcp_time_stamp
;
5535 tcp_enter_quickack_mode(sk
);
5536 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
5537 TCP_DELACK_MAX
, TCP_RTO_MAX
);
5548 /* No ACK in the segment */
5552 * "If the RST bit is set
5554 * Otherwise (no ACK) drop the segment and return."
5557 goto discard_and_undo
;
5561 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&&
5562 tcp_paws_reject(&tp
->rx_opt
, 0))
5563 goto discard_and_undo
;
5566 /* We see SYN without ACK. It is attempt of
5567 * simultaneous connect with crossed SYNs.
5568 * Particularly, it can be connect to self.
5570 tcp_set_state(sk
, TCP_SYN_RECV
);
5572 if (tp
->rx_opt
.saw_tstamp
) {
5573 tp
->rx_opt
.tstamp_ok
= 1;
5574 tcp_store_ts_recent(tp
);
5575 tp
->tcp_header_len
=
5576 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5578 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5581 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5582 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5584 /* RFC1323: The window in SYN & SYN/ACK segments is
5587 tp
->snd_wnd
= ntohs(th
->window
);
5588 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
5589 tp
->max_window
= tp
->snd_wnd
;
5591 TCP_ECN_rcv_syn(tp
, th
);
5594 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5595 tcp_initialize_rcv_mss(sk
);
5597 tcp_send_synack(sk
);
5599 /* Note, we could accept data and URG from this segment.
5600 * There are no obstacles to make this (except that we must
5601 * either change tcp_recvmsg() to prevent it from returning data
5602 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5604 * However, if we ignore data in ACKless segments sometimes,
5605 * we have no reasons to accept it sometimes.
5606 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5607 * is not flawless. So, discard packet for sanity.
5608 * Uncomment this return to process the data.
5615 /* "fifth, if neither of the SYN or RST bits is set then
5616 * drop the segment and return."
5620 tcp_clear_options(&tp
->rx_opt
);
5621 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5625 tcp_clear_options(&tp
->rx_opt
);
5626 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5631 * This function implements the receiving procedure of RFC 793 for
5632 * all states except ESTABLISHED and TIME_WAIT.
5633 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5634 * address independent.
5637 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5638 const struct tcphdr
*th
, unsigned int len
)
5640 struct tcp_sock
*tp
= tcp_sk(sk
);
5641 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5642 struct request_sock
*req
;
5647 tp
->rx_opt
.saw_tstamp
= 0;
5649 switch (sk
->sk_state
) {
5663 if (icsk
->icsk_af_ops
->conn_request(sk
, skb
) < 0)
5666 /* Now we have several options: In theory there is
5667 * nothing else in the frame. KA9Q has an option to
5668 * send data with the syn, BSD accepts data with the
5669 * syn up to the [to be] advertised window and
5670 * Solaris 2.1 gives you a protocol error. For now
5671 * we just ignore it, that fits the spec precisely
5672 * and avoids incompatibilities. It would be nice in
5673 * future to drop through and process the data.
5675 * Now that TTCP is starting to be used we ought to
5677 * But, this leaves one open to an easy denial of
5678 * service attack, and SYN cookies can't defend
5679 * against this problem. So, we drop the data
5680 * in the interest of security over speed unless
5681 * it's still in use.
5689 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
, len
);
5693 /* Do step6 onward by hand. */
5694 tcp_urg(sk
, skb
, th
);
5696 tcp_data_snd_check(sk
);
5700 req
= tp
->fastopen_rsk
;
5702 WARN_ON_ONCE(sk
->sk_state
!= TCP_SYN_RECV
&&
5703 sk
->sk_state
!= TCP_FIN_WAIT1
);
5705 if (tcp_check_req(sk
, skb
, req
, NULL
, true) == NULL
)
5709 if (!th
->ack
&& !th
->rst
)
5712 if (!tcp_validate_incoming(sk
, skb
, th
, 0))
5715 /* step 5: check the ACK field */
5716 acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
|
5717 FLAG_UPDATE_TS_RECENT
) > 0;
5719 switch (sk
->sk_state
) {
5724 /* Once we leave TCP_SYN_RECV, we no longer need req
5728 synack_stamp
= tcp_rsk(req
)->snt_synack
;
5729 tp
->total_retrans
= req
->num_retrans
;
5730 reqsk_fastopen_remove(sk
, req
, false);
5732 synack_stamp
= tp
->lsndtime
;
5733 /* Make sure socket is routed, for correct metrics. */
5734 icsk
->icsk_af_ops
->rebuild_header(sk
);
5735 tcp_init_congestion_control(sk
);
5738 tp
->copied_seq
= tp
->rcv_nxt
;
5739 tcp_init_buffer_space(sk
);
5742 tcp_set_state(sk
, TCP_ESTABLISHED
);
5743 sk
->sk_state_change(sk
);
5745 /* Note, that this wakeup is only for marginal crossed SYN case.
5746 * Passively open sockets are not waked up, because
5747 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5750 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5752 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
5753 tp
->snd_wnd
= ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
;
5754 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5755 tcp_synack_rtt_meas(sk
, synack_stamp
);
5757 if (tp
->rx_opt
.tstamp_ok
)
5758 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5761 /* Re-arm the timer because data may have been sent out.
5762 * This is similar to the regular data transmission case
5763 * when new data has just been ack'ed.
5765 * (TFO) - we could try to be more aggressive and
5766 * retransmitting any data sooner based on when they
5771 tcp_init_metrics(sk
);
5773 tcp_update_pacing_rate(sk
);
5775 /* Prevent spurious tcp_cwnd_restart() on first data packet */
5776 tp
->lsndtime
= tcp_time_stamp
;
5778 tcp_initialize_rcv_mss(sk
);
5779 tcp_fast_path_on(tp
);
5782 case TCP_FIN_WAIT1
: {
5783 struct dst_entry
*dst
;
5786 /* If we enter the TCP_FIN_WAIT1 state and we are a
5787 * Fast Open socket and this is the first acceptable
5788 * ACK we have received, this would have acknowledged
5789 * our SYNACK so stop the SYNACK timer.
5792 /* Return RST if ack_seq is invalid.
5793 * Note that RFC793 only says to generate a
5794 * DUPACK for it but for TCP Fast Open it seems
5795 * better to treat this case like TCP_SYN_RECV
5800 /* We no longer need the request sock. */
5801 reqsk_fastopen_remove(sk
, req
, false);
5804 if (tp
->snd_una
!= tp
->write_seq
)
5807 tcp_set_state(sk
, TCP_FIN_WAIT2
);
5808 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
5810 dst
= __sk_dst_get(sk
);
5814 if (!sock_flag(sk
, SOCK_DEAD
)) {
5815 /* Wake up lingering close() */
5816 sk
->sk_state_change(sk
);
5820 if (tp
->linger2
< 0 ||
5821 (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5822 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
))) {
5824 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
5828 tmo
= tcp_fin_time(sk
);
5829 if (tmo
> TCP_TIMEWAIT_LEN
) {
5830 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
5831 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
5832 /* Bad case. We could lose such FIN otherwise.
5833 * It is not a big problem, but it looks confusing
5834 * and not so rare event. We still can lose it now,
5835 * if it spins in bh_lock_sock(), but it is really
5838 inet_csk_reset_keepalive_timer(sk
, tmo
);
5840 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
5847 if (tp
->snd_una
== tp
->write_seq
) {
5848 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
5854 if (tp
->snd_una
== tp
->write_seq
) {
5855 tcp_update_metrics(sk
);
5862 /* step 6: check the URG bit */
5863 tcp_urg(sk
, skb
, th
);
5865 /* step 7: process the segment text */
5866 switch (sk
->sk_state
) {
5867 case TCP_CLOSE_WAIT
:
5870 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
5874 /* RFC 793 says to queue data in these states,
5875 * RFC 1122 says we MUST send a reset.
5876 * BSD 4.4 also does reset.
5878 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
5879 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5880 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
5881 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
5887 case TCP_ESTABLISHED
:
5888 tcp_data_queue(sk
, skb
);
5893 /* tcp_data could move socket to TIME-WAIT */
5894 if (sk
->sk_state
!= TCP_CLOSE
) {
5895 tcp_data_snd_check(sk
);
5896 tcp_ack_snd_check(sk
);
5905 EXPORT_SYMBOL(tcp_rcv_state_process
);