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
, long mrtt_us
)
672 struct tcp_sock
*tp
= tcp_sk(sk
);
673 long m
= mrtt_us
; /* RTT */
674 u32 srtt
= tp
->srtt_us
;
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_us
>> 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_us
>> 2); /* similar update on mdev */
711 tp
->mdev_us
+= m
; /* mdev = 3/4 mdev + 1/4 new */
712 if (tp
->mdev_us
> tp
->mdev_max_us
) {
713 tp
->mdev_max_us
= tp
->mdev_us
;
714 if (tp
->mdev_max_us
> tp
->rttvar_us
)
715 tp
->rttvar_us
= tp
->mdev_max_us
;
717 if (after(tp
->snd_una
, tp
->rtt_seq
)) {
718 if (tp
->mdev_max_us
< tp
->rttvar_us
)
719 tp
->rttvar_us
-= (tp
->rttvar_us
- tp
->mdev_max_us
) >> 2;
720 tp
->rtt_seq
= tp
->snd_nxt
;
721 tp
->mdev_max_us
= tcp_rto_min_us(sk
);
724 /* no previous measure. */
725 srtt
= m
<< 3; /* take the measured time to be rtt */
726 tp
->mdev_us
= m
<< 1; /* make sure rto = 3*rtt */
727 tp
->rttvar_us
= max(tp
->mdev_us
, tcp_rto_min_us(sk
));
728 tp
->mdev_max_us
= tp
->rttvar_us
;
729 tp
->rtt_seq
= tp
->snd_nxt
;
731 tp
->srtt_us
= max(1U, srtt
);
734 /* Set the sk_pacing_rate to allow proper sizing of TSO packets.
735 * Note: TCP stack does not yet implement pacing.
736 * FQ packet scheduler can be used to implement cheap but effective
737 * TCP pacing, to smooth the burst on large writes when packets
738 * in flight is significantly lower than cwnd (or rwin)
740 static void tcp_update_pacing_rate(struct sock
*sk
)
742 const struct tcp_sock
*tp
= tcp_sk(sk
);
745 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
746 rate
= (u64
)tp
->mss_cache
* 2 * (USEC_PER_SEC
<< 3);
748 rate
*= max(tp
->snd_cwnd
, tp
->packets_out
);
750 if (likely(tp
->srtt_us
))
751 do_div(rate
, tp
->srtt_us
);
753 /* ACCESS_ONCE() is needed because sch_fq fetches sk_pacing_rate
754 * without any lock. We want to make sure compiler wont store
755 * intermediate values in this location.
757 ACCESS_ONCE(sk
->sk_pacing_rate
) = min_t(u64
, rate
,
758 sk
->sk_max_pacing_rate
);
761 /* Calculate rto without backoff. This is the second half of Van Jacobson's
762 * routine referred to above.
764 static void tcp_set_rto(struct sock
*sk
)
766 const struct tcp_sock
*tp
= tcp_sk(sk
);
767 /* Old crap is replaced with new one. 8)
770 * 1. If rtt variance happened to be less 50msec, it is hallucination.
771 * It cannot be less due to utterly erratic ACK generation made
772 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
773 * to do with delayed acks, because at cwnd>2 true delack timeout
774 * is invisible. Actually, Linux-2.4 also generates erratic
775 * ACKs in some circumstances.
777 inet_csk(sk
)->icsk_rto
= __tcp_set_rto(tp
);
779 /* 2. Fixups made earlier cannot be right.
780 * If we do not estimate RTO correctly without them,
781 * all the algo is pure shit and should be replaced
782 * with correct one. It is exactly, which we pretend to do.
785 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
786 * guarantees that rto is higher.
791 __u32
tcp_init_cwnd(const struct tcp_sock
*tp
, const struct dst_entry
*dst
)
793 __u32 cwnd
= (dst
? dst_metric(dst
, RTAX_INITCWND
) : 0);
796 cwnd
= TCP_INIT_CWND
;
797 return min_t(__u32
, cwnd
, tp
->snd_cwnd_clamp
);
801 * Packet counting of FACK is based on in-order assumptions, therefore TCP
802 * disables it when reordering is detected
804 void tcp_disable_fack(struct tcp_sock
*tp
)
806 /* RFC3517 uses different metric in lost marker => reset on change */
808 tp
->lost_skb_hint
= NULL
;
809 tp
->rx_opt
.sack_ok
&= ~TCP_FACK_ENABLED
;
812 /* Take a notice that peer is sending D-SACKs */
813 static void tcp_dsack_seen(struct tcp_sock
*tp
)
815 tp
->rx_opt
.sack_ok
|= TCP_DSACK_SEEN
;
818 static void tcp_update_reordering(struct sock
*sk
, const int metric
,
821 struct tcp_sock
*tp
= tcp_sk(sk
);
822 if (metric
> tp
->reordering
) {
825 tp
->reordering
= min(TCP_MAX_REORDERING
, metric
);
827 /* This exciting event is worth to be remembered. 8) */
829 mib_idx
= LINUX_MIB_TCPTSREORDER
;
830 else if (tcp_is_reno(tp
))
831 mib_idx
= LINUX_MIB_TCPRENOREORDER
;
832 else if (tcp_is_fack(tp
))
833 mib_idx
= LINUX_MIB_TCPFACKREORDER
;
835 mib_idx
= LINUX_MIB_TCPSACKREORDER
;
837 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
838 #if FASTRETRANS_DEBUG > 1
839 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
840 tp
->rx_opt
.sack_ok
, inet_csk(sk
)->icsk_ca_state
,
844 tp
->undo_marker
? tp
->undo_retrans
: 0);
846 tcp_disable_fack(tp
);
850 tcp_disable_early_retrans(tp
);
853 /* This must be called before lost_out is incremented */
854 static void tcp_verify_retransmit_hint(struct tcp_sock
*tp
, struct sk_buff
*skb
)
856 if ((tp
->retransmit_skb_hint
== NULL
) ||
857 before(TCP_SKB_CB(skb
)->seq
,
858 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
859 tp
->retransmit_skb_hint
= skb
;
862 after(TCP_SKB_CB(skb
)->end_seq
, tp
->retransmit_high
))
863 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
866 static void tcp_skb_mark_lost(struct tcp_sock
*tp
, struct sk_buff
*skb
)
868 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
869 tcp_verify_retransmit_hint(tp
, skb
);
871 tp
->lost_out
+= tcp_skb_pcount(skb
);
872 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
876 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock
*tp
,
879 tcp_verify_retransmit_hint(tp
, skb
);
881 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
882 tp
->lost_out
+= tcp_skb_pcount(skb
);
883 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
887 /* This procedure tags the retransmission queue when SACKs arrive.
889 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
890 * Packets in queue with these bits set are counted in variables
891 * sacked_out, retrans_out and lost_out, correspondingly.
893 * Valid combinations are:
894 * Tag InFlight Description
895 * 0 1 - orig segment is in flight.
896 * S 0 - nothing flies, orig reached receiver.
897 * L 0 - nothing flies, orig lost by net.
898 * R 2 - both orig and retransmit are in flight.
899 * L|R 1 - orig is lost, retransmit is in flight.
900 * S|R 1 - orig reached receiver, retrans is still in flight.
901 * (L|S|R is logically valid, it could occur when L|R is sacked,
902 * but it is equivalent to plain S and code short-curcuits it to S.
903 * L|S is logically invalid, it would mean -1 packet in flight 8))
905 * These 6 states form finite state machine, controlled by the following events:
906 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
907 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
908 * 3. Loss detection event of two flavors:
909 * A. Scoreboard estimator decided the packet is lost.
910 * A'. Reno "three dupacks" marks head of queue lost.
911 * A''. Its FACK modification, head until snd.fack is lost.
912 * B. SACK arrives sacking SND.NXT at the moment, when the
913 * segment was retransmitted.
914 * 4. D-SACK added new rule: D-SACK changes any tag to S.
916 * It is pleasant to note, that state diagram turns out to be commutative,
917 * so that we are allowed not to be bothered by order of our actions,
918 * when multiple events arrive simultaneously. (see the function below).
920 * Reordering detection.
921 * --------------------
922 * Reordering metric is maximal distance, which a packet can be displaced
923 * in packet stream. With SACKs we can estimate it:
925 * 1. SACK fills old hole and the corresponding segment was not
926 * ever retransmitted -> reordering. Alas, we cannot use it
927 * when segment was retransmitted.
928 * 2. The last flaw is solved with D-SACK. D-SACK arrives
929 * for retransmitted and already SACKed segment -> reordering..
930 * Both of these heuristics are not used in Loss state, when we cannot
931 * account for retransmits accurately.
933 * SACK block validation.
934 * ----------------------
936 * SACK block range validation checks that the received SACK block fits to
937 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
938 * Note that SND.UNA is not included to the range though being valid because
939 * it means that the receiver is rather inconsistent with itself reporting
940 * SACK reneging when it should advance SND.UNA. Such SACK block this is
941 * perfectly valid, however, in light of RFC2018 which explicitly states
942 * that "SACK block MUST reflect the newest segment. Even if the newest
943 * segment is going to be discarded ...", not that it looks very clever
944 * in case of head skb. Due to potentional receiver driven attacks, we
945 * choose to avoid immediate execution of a walk in write queue due to
946 * reneging and defer head skb's loss recovery to standard loss recovery
947 * procedure that will eventually trigger (nothing forbids us doing this).
949 * Implements also blockage to start_seq wrap-around. Problem lies in the
950 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
951 * there's no guarantee that it will be before snd_nxt (n). The problem
952 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
955 * <- outs wnd -> <- wrapzone ->
956 * u e n u_w e_w s n_w
958 * |<------------+------+----- TCP seqno space --------------+---------->|
959 * ...-- <2^31 ->| |<--------...
960 * ...---- >2^31 ------>| |<--------...
962 * Current code wouldn't be vulnerable but it's better still to discard such
963 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
964 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
965 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
966 * equal to the ideal case (infinite seqno space without wrap caused issues).
968 * With D-SACK the lower bound is extended to cover sequence space below
969 * SND.UNA down to undo_marker, which is the last point of interest. Yet
970 * again, D-SACK block must not to go across snd_una (for the same reason as
971 * for the normal SACK blocks, explained above). But there all simplicity
972 * ends, TCP might receive valid D-SACKs below that. As long as they reside
973 * fully below undo_marker they do not affect behavior in anyway and can
974 * therefore be safely ignored. In rare cases (which are more or less
975 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
976 * fragmentation and packet reordering past skb's retransmission. To consider
977 * them correctly, the acceptable range must be extended even more though
978 * the exact amount is rather hard to quantify. However, tp->max_window can
979 * be used as an exaggerated estimate.
981 static bool tcp_is_sackblock_valid(struct tcp_sock
*tp
, bool is_dsack
,
982 u32 start_seq
, u32 end_seq
)
984 /* Too far in future, or reversed (interpretation is ambiguous) */
985 if (after(end_seq
, tp
->snd_nxt
) || !before(start_seq
, end_seq
))
988 /* Nasty start_seq wrap-around check (see comments above) */
989 if (!before(start_seq
, tp
->snd_nxt
))
992 /* In outstanding window? ...This is valid exit for D-SACKs too.
993 * start_seq == snd_una is non-sensical (see comments above)
995 if (after(start_seq
, tp
->snd_una
))
998 if (!is_dsack
|| !tp
->undo_marker
)
1001 /* ...Then it's D-SACK, and must reside below snd_una completely */
1002 if (after(end_seq
, tp
->snd_una
))
1005 if (!before(start_seq
, tp
->undo_marker
))
1009 if (!after(end_seq
, tp
->undo_marker
))
1012 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1013 * start_seq < undo_marker and end_seq >= undo_marker.
1015 return !before(start_seq
, end_seq
- tp
->max_window
);
1018 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1019 * Event "B". Later note: FACK people cheated me again 8), we have to account
1020 * for reordering! Ugly, but should help.
1022 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1023 * less than what is now known to be received by the other end (derived from
1024 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1025 * retransmitted skbs to avoid some costly processing per ACKs.
1027 static void tcp_mark_lost_retrans(struct sock
*sk
)
1029 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1030 struct tcp_sock
*tp
= tcp_sk(sk
);
1031 struct sk_buff
*skb
;
1033 u32 new_low_seq
= tp
->snd_nxt
;
1034 u32 received_upto
= tcp_highest_sack_seq(tp
);
1036 if (!tcp_is_fack(tp
) || !tp
->retrans_out
||
1037 !after(received_upto
, tp
->lost_retrans_low
) ||
1038 icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
1041 tcp_for_write_queue(skb
, sk
) {
1042 u32 ack_seq
= TCP_SKB_CB(skb
)->ack_seq
;
1044 if (skb
== tcp_send_head(sk
))
1046 if (cnt
== tp
->retrans_out
)
1048 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1051 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
))
1054 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1055 * constraint here (see above) but figuring out that at
1056 * least tp->reordering SACK blocks reside between ack_seq
1057 * and received_upto is not easy task to do cheaply with
1058 * the available datastructures.
1060 * Whether FACK should check here for tp->reordering segs
1061 * in-between one could argue for either way (it would be
1062 * rather simple to implement as we could count fack_count
1063 * during the walk and do tp->fackets_out - fack_count).
1065 if (after(received_upto
, ack_seq
)) {
1066 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1067 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1069 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
1070 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSTRETRANSMIT
);
1072 if (before(ack_seq
, new_low_seq
))
1073 new_low_seq
= ack_seq
;
1074 cnt
+= tcp_skb_pcount(skb
);
1078 if (tp
->retrans_out
)
1079 tp
->lost_retrans_low
= new_low_seq
;
1082 static bool tcp_check_dsack(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1083 struct tcp_sack_block_wire
*sp
, int num_sacks
,
1086 struct tcp_sock
*tp
= tcp_sk(sk
);
1087 u32 start_seq_0
= get_unaligned_be32(&sp
[0].start_seq
);
1088 u32 end_seq_0
= get_unaligned_be32(&sp
[0].end_seq
);
1089 bool dup_sack
= false;
1091 if (before(start_seq_0
, TCP_SKB_CB(ack_skb
)->ack_seq
)) {
1094 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPDSACKRECV
);
1095 } else if (num_sacks
> 1) {
1096 u32 end_seq_1
= get_unaligned_be32(&sp
[1].end_seq
);
1097 u32 start_seq_1
= get_unaligned_be32(&sp
[1].start_seq
);
1099 if (!after(end_seq_0
, end_seq_1
) &&
1100 !before(start_seq_0
, start_seq_1
)) {
1103 NET_INC_STATS_BH(sock_net(sk
),
1104 LINUX_MIB_TCPDSACKOFORECV
);
1108 /* D-SACK for already forgotten data... Do dumb counting. */
1109 if (dup_sack
&& tp
->undo_marker
&& tp
->undo_retrans
&&
1110 !after(end_seq_0
, prior_snd_una
) &&
1111 after(end_seq_0
, tp
->undo_marker
))
1117 struct tcp_sacktag_state
{
1120 long rtt_us
; /* RTT measured by SACKing never-retransmitted data */
1124 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1125 * the incoming SACK may not exactly match but we can find smaller MSS
1126 * aligned portion of it that matches. Therefore we might need to fragment
1127 * which may fail and creates some hassle (caller must handle error case
1130 * FIXME: this could be merged to shift decision code
1132 static int tcp_match_skb_to_sack(struct sock
*sk
, struct sk_buff
*skb
,
1133 u32 start_seq
, u32 end_seq
)
1137 unsigned int pkt_len
;
1140 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1141 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1143 if (tcp_skb_pcount(skb
) > 1 && !in_sack
&&
1144 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
1145 mss
= tcp_skb_mss(skb
);
1146 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1149 pkt_len
= start_seq
- TCP_SKB_CB(skb
)->seq
;
1153 pkt_len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1158 /* Round if necessary so that SACKs cover only full MSSes
1159 * and/or the remaining small portion (if present)
1161 if (pkt_len
> mss
) {
1162 unsigned int new_len
= (pkt_len
/ mss
) * mss
;
1163 if (!in_sack
&& new_len
< pkt_len
) {
1165 if (new_len
> skb
->len
)
1170 err
= tcp_fragment(sk
, skb
, pkt_len
, mss
);
1178 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1179 static u8
tcp_sacktag_one(struct sock
*sk
,
1180 struct tcp_sacktag_state
*state
, u8 sacked
,
1181 u32 start_seq
, u32 end_seq
,
1182 int dup_sack
, int pcount
,
1183 const struct skb_mstamp
*xmit_time
)
1185 struct tcp_sock
*tp
= tcp_sk(sk
);
1186 int fack_count
= state
->fack_count
;
1188 /* Account D-SACK for retransmitted packet. */
1189 if (dup_sack
&& (sacked
& TCPCB_RETRANS
)) {
1190 if (tp
->undo_marker
&& tp
->undo_retrans
&&
1191 after(end_seq
, tp
->undo_marker
))
1193 if (sacked
& TCPCB_SACKED_ACKED
)
1194 state
->reord
= min(fack_count
, state
->reord
);
1197 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1198 if (!after(end_seq
, tp
->snd_una
))
1201 if (!(sacked
& TCPCB_SACKED_ACKED
)) {
1202 if (sacked
& TCPCB_SACKED_RETRANS
) {
1203 /* If the segment is not tagged as lost,
1204 * we do not clear RETRANS, believing
1205 * that retransmission is still in flight.
1207 if (sacked
& TCPCB_LOST
) {
1208 sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1209 tp
->lost_out
-= pcount
;
1210 tp
->retrans_out
-= pcount
;
1213 if (!(sacked
& TCPCB_RETRANS
)) {
1214 /* New sack for not retransmitted frame,
1215 * which was in hole. It is reordering.
1217 if (before(start_seq
,
1218 tcp_highest_sack_seq(tp
)))
1219 state
->reord
= min(fack_count
,
1221 if (!after(end_seq
, tp
->high_seq
))
1222 state
->flag
|= FLAG_ORIG_SACK_ACKED
;
1223 /* Pick the earliest sequence sacked for RTT */
1224 if (state
->rtt_us
< 0) {
1225 struct skb_mstamp now
;
1227 skb_mstamp_get(&now
);
1228 state
->rtt_us
= skb_mstamp_us_delta(&now
,
1233 if (sacked
& TCPCB_LOST
) {
1234 sacked
&= ~TCPCB_LOST
;
1235 tp
->lost_out
-= pcount
;
1239 sacked
|= TCPCB_SACKED_ACKED
;
1240 state
->flag
|= FLAG_DATA_SACKED
;
1241 tp
->sacked_out
+= pcount
;
1243 fack_count
+= pcount
;
1245 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1246 if (!tcp_is_fack(tp
) && (tp
->lost_skb_hint
!= NULL
) &&
1247 before(start_seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
))
1248 tp
->lost_cnt_hint
+= pcount
;
1250 if (fack_count
> tp
->fackets_out
)
1251 tp
->fackets_out
= fack_count
;
1254 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1255 * frames and clear it. undo_retrans is decreased above, L|R frames
1256 * are accounted above as well.
1258 if (dup_sack
&& (sacked
& TCPCB_SACKED_RETRANS
)) {
1259 sacked
&= ~TCPCB_SACKED_RETRANS
;
1260 tp
->retrans_out
-= pcount
;
1266 /* Shift newly-SACKed bytes from this skb to the immediately previous
1267 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1269 static bool tcp_shifted_skb(struct sock
*sk
, struct sk_buff
*skb
,
1270 struct tcp_sacktag_state
*state
,
1271 unsigned int pcount
, int shifted
, int mss
,
1274 struct tcp_sock
*tp
= tcp_sk(sk
);
1275 struct sk_buff
*prev
= tcp_write_queue_prev(sk
, skb
);
1276 u32 start_seq
= TCP_SKB_CB(skb
)->seq
; /* start of newly-SACKed */
1277 u32 end_seq
= start_seq
+ shifted
; /* end of newly-SACKed */
1281 /* Adjust counters and hints for the newly sacked sequence
1282 * range but discard the return value since prev is already
1283 * marked. We must tag the range first because the seq
1284 * advancement below implicitly advances
1285 * tcp_highest_sack_seq() when skb is highest_sack.
1287 tcp_sacktag_one(sk
, state
, TCP_SKB_CB(skb
)->sacked
,
1288 start_seq
, end_seq
, dup_sack
, pcount
,
1291 if (skb
== tp
->lost_skb_hint
)
1292 tp
->lost_cnt_hint
+= pcount
;
1294 TCP_SKB_CB(prev
)->end_seq
+= shifted
;
1295 TCP_SKB_CB(skb
)->seq
+= shifted
;
1297 skb_shinfo(prev
)->gso_segs
+= pcount
;
1298 BUG_ON(skb_shinfo(skb
)->gso_segs
< pcount
);
1299 skb_shinfo(skb
)->gso_segs
-= pcount
;
1301 /* When we're adding to gso_segs == 1, gso_size will be zero,
1302 * in theory this shouldn't be necessary but as long as DSACK
1303 * code can come after this skb later on it's better to keep
1304 * setting gso_size to something.
1306 if (!skb_shinfo(prev
)->gso_size
) {
1307 skb_shinfo(prev
)->gso_size
= mss
;
1308 skb_shinfo(prev
)->gso_type
= sk
->sk_gso_type
;
1311 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1312 if (skb_shinfo(skb
)->gso_segs
<= 1) {
1313 skb_shinfo(skb
)->gso_size
= 0;
1314 skb_shinfo(skb
)->gso_type
= 0;
1317 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1318 TCP_SKB_CB(prev
)->sacked
|= (TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
);
1321 BUG_ON(!tcp_skb_pcount(skb
));
1322 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKSHIFTED
);
1326 /* Whole SKB was eaten :-) */
1328 if (skb
== tp
->retransmit_skb_hint
)
1329 tp
->retransmit_skb_hint
= prev
;
1330 if (skb
== tp
->lost_skb_hint
) {
1331 tp
->lost_skb_hint
= prev
;
1332 tp
->lost_cnt_hint
-= tcp_skb_pcount(prev
);
1335 TCP_SKB_CB(prev
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
;
1336 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
1337 TCP_SKB_CB(prev
)->end_seq
++;
1339 if (skb
== tcp_highest_sack(sk
))
1340 tcp_advance_highest_sack(sk
, skb
);
1342 tcp_unlink_write_queue(skb
, sk
);
1343 sk_wmem_free_skb(sk
, skb
);
1345 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKMERGED
);
1350 /* I wish gso_size would have a bit more sane initialization than
1351 * something-or-zero which complicates things
1353 static int tcp_skb_seglen(const struct sk_buff
*skb
)
1355 return tcp_skb_pcount(skb
) == 1 ? skb
->len
: tcp_skb_mss(skb
);
1358 /* Shifting pages past head area doesn't work */
1359 static int skb_can_shift(const struct sk_buff
*skb
)
1361 return !skb_headlen(skb
) && skb_is_nonlinear(skb
);
1364 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1367 static struct sk_buff
*tcp_shift_skb_data(struct sock
*sk
, struct sk_buff
*skb
,
1368 struct tcp_sacktag_state
*state
,
1369 u32 start_seq
, u32 end_seq
,
1372 struct tcp_sock
*tp
= tcp_sk(sk
);
1373 struct sk_buff
*prev
;
1379 if (!sk_can_gso(sk
))
1382 /* Normally R but no L won't result in plain S */
1384 (TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_RETRANS
)) == TCPCB_SACKED_RETRANS
)
1386 if (!skb_can_shift(skb
))
1388 /* This frame is about to be dropped (was ACKed). */
1389 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1392 /* Can only happen with delayed DSACK + discard craziness */
1393 if (unlikely(skb
== tcp_write_queue_head(sk
)))
1395 prev
= tcp_write_queue_prev(sk
, skb
);
1397 if ((TCP_SKB_CB(prev
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
)
1400 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1401 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1405 pcount
= tcp_skb_pcount(skb
);
1406 mss
= tcp_skb_seglen(skb
);
1408 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1409 * drop this restriction as unnecessary
1411 if (mss
!= tcp_skb_seglen(prev
))
1414 if (!after(TCP_SKB_CB(skb
)->end_seq
, start_seq
))
1416 /* CHECKME: This is non-MSS split case only?, this will
1417 * cause skipped skbs due to advancing loop btw, original
1418 * has that feature too
1420 if (tcp_skb_pcount(skb
) <= 1)
1423 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1425 /* TODO: head merge to next could be attempted here
1426 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1427 * though it might not be worth of the additional hassle
1429 * ...we can probably just fallback to what was done
1430 * previously. We could try merging non-SACKed ones
1431 * as well but it probably isn't going to buy off
1432 * because later SACKs might again split them, and
1433 * it would make skb timestamp tracking considerably
1439 len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1441 BUG_ON(len
> skb
->len
);
1443 /* MSS boundaries should be honoured or else pcount will
1444 * severely break even though it makes things bit trickier.
1445 * Optimize common case to avoid most of the divides
1447 mss
= tcp_skb_mss(skb
);
1449 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1450 * drop this restriction as unnecessary
1452 if (mss
!= tcp_skb_seglen(prev
))
1457 } else if (len
< mss
) {
1465 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1466 if (!after(TCP_SKB_CB(skb
)->seq
+ len
, tp
->snd_una
))
1469 if (!skb_shift(prev
, skb
, len
))
1471 if (!tcp_shifted_skb(sk
, skb
, state
, pcount
, len
, mss
, dup_sack
))
1474 /* Hole filled allows collapsing with the next as well, this is very
1475 * useful when hole on every nth skb pattern happens
1477 if (prev
== tcp_write_queue_tail(sk
))
1479 skb
= tcp_write_queue_next(sk
, prev
);
1481 if (!skb_can_shift(skb
) ||
1482 (skb
== tcp_send_head(sk
)) ||
1483 ((TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
) ||
1484 (mss
!= tcp_skb_seglen(skb
)))
1488 if (skb_shift(prev
, skb
, len
)) {
1489 pcount
+= tcp_skb_pcount(skb
);
1490 tcp_shifted_skb(sk
, skb
, state
, tcp_skb_pcount(skb
), len
, mss
, 0);
1494 state
->fack_count
+= pcount
;
1501 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKSHIFTFALLBACK
);
1505 static struct sk_buff
*tcp_sacktag_walk(struct sk_buff
*skb
, struct sock
*sk
,
1506 struct tcp_sack_block
*next_dup
,
1507 struct tcp_sacktag_state
*state
,
1508 u32 start_seq
, u32 end_seq
,
1511 struct tcp_sock
*tp
= tcp_sk(sk
);
1512 struct sk_buff
*tmp
;
1514 tcp_for_write_queue_from(skb
, sk
) {
1516 bool dup_sack
= dup_sack_in
;
1518 if (skb
== tcp_send_head(sk
))
1521 /* queue is in-order => we can short-circuit the walk early */
1522 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1525 if ((next_dup
!= NULL
) &&
1526 before(TCP_SKB_CB(skb
)->seq
, next_dup
->end_seq
)) {
1527 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1528 next_dup
->start_seq
,
1534 /* skb reference here is a bit tricky to get right, since
1535 * shifting can eat and free both this skb and the next,
1536 * so not even _safe variant of the loop is enough.
1539 tmp
= tcp_shift_skb_data(sk
, skb
, state
,
1540 start_seq
, end_seq
, dup_sack
);
1549 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1555 if (unlikely(in_sack
< 0))
1559 TCP_SKB_CB(skb
)->sacked
=
1562 TCP_SKB_CB(skb
)->sacked
,
1563 TCP_SKB_CB(skb
)->seq
,
1564 TCP_SKB_CB(skb
)->end_seq
,
1566 tcp_skb_pcount(skb
),
1569 if (!before(TCP_SKB_CB(skb
)->seq
,
1570 tcp_highest_sack_seq(tp
)))
1571 tcp_advance_highest_sack(sk
, skb
);
1574 state
->fack_count
+= tcp_skb_pcount(skb
);
1579 /* Avoid all extra work that is being done by sacktag while walking in
1582 static struct sk_buff
*tcp_sacktag_skip(struct sk_buff
*skb
, struct sock
*sk
,
1583 struct tcp_sacktag_state
*state
,
1586 tcp_for_write_queue_from(skb
, sk
) {
1587 if (skb
== tcp_send_head(sk
))
1590 if (after(TCP_SKB_CB(skb
)->end_seq
, skip_to_seq
))
1593 state
->fack_count
+= tcp_skb_pcount(skb
);
1598 static struct sk_buff
*tcp_maybe_skipping_dsack(struct sk_buff
*skb
,
1600 struct tcp_sack_block
*next_dup
,
1601 struct tcp_sacktag_state
*state
,
1604 if (next_dup
== NULL
)
1607 if (before(next_dup
->start_seq
, skip_to_seq
)) {
1608 skb
= tcp_sacktag_skip(skb
, sk
, state
, next_dup
->start_seq
);
1609 skb
= tcp_sacktag_walk(skb
, sk
, NULL
, state
,
1610 next_dup
->start_seq
, next_dup
->end_seq
,
1617 static int tcp_sack_cache_ok(const struct tcp_sock
*tp
, const struct tcp_sack_block
*cache
)
1619 return cache
< tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1623 tcp_sacktag_write_queue(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1624 u32 prior_snd_una
, long *sack_rtt_us
)
1626 struct tcp_sock
*tp
= tcp_sk(sk
);
1627 const unsigned char *ptr
= (skb_transport_header(ack_skb
) +
1628 TCP_SKB_CB(ack_skb
)->sacked
);
1629 struct tcp_sack_block_wire
*sp_wire
= (struct tcp_sack_block_wire
*)(ptr
+2);
1630 struct tcp_sack_block sp
[TCP_NUM_SACKS
];
1631 struct tcp_sack_block
*cache
;
1632 struct tcp_sacktag_state state
;
1633 struct sk_buff
*skb
;
1634 int num_sacks
= min(TCP_NUM_SACKS
, (ptr
[1] - TCPOLEN_SACK_BASE
) >> 3);
1636 bool found_dup_sack
= false;
1638 int first_sack_index
;
1641 state
.reord
= tp
->packets_out
;
1644 if (!tp
->sacked_out
) {
1645 if (WARN_ON(tp
->fackets_out
))
1646 tp
->fackets_out
= 0;
1647 tcp_highest_sack_reset(sk
);
1650 found_dup_sack
= tcp_check_dsack(sk
, ack_skb
, sp_wire
,
1651 num_sacks
, prior_snd_una
);
1653 state
.flag
|= FLAG_DSACKING_ACK
;
1655 /* Eliminate too old ACKs, but take into
1656 * account more or less fresh ones, they can
1657 * contain valid SACK info.
1659 if (before(TCP_SKB_CB(ack_skb
)->ack_seq
, prior_snd_una
- tp
->max_window
))
1662 if (!tp
->packets_out
)
1666 first_sack_index
= 0;
1667 for (i
= 0; i
< num_sacks
; i
++) {
1668 bool dup_sack
= !i
&& found_dup_sack
;
1670 sp
[used_sacks
].start_seq
= get_unaligned_be32(&sp_wire
[i
].start_seq
);
1671 sp
[used_sacks
].end_seq
= get_unaligned_be32(&sp_wire
[i
].end_seq
);
1673 if (!tcp_is_sackblock_valid(tp
, dup_sack
,
1674 sp
[used_sacks
].start_seq
,
1675 sp
[used_sacks
].end_seq
)) {
1679 if (!tp
->undo_marker
)
1680 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDNOUNDO
;
1682 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDOLD
;
1684 /* Don't count olds caused by ACK reordering */
1685 if ((TCP_SKB_CB(ack_skb
)->ack_seq
!= tp
->snd_una
) &&
1686 !after(sp
[used_sacks
].end_seq
, tp
->snd_una
))
1688 mib_idx
= LINUX_MIB_TCPSACKDISCARD
;
1691 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
1693 first_sack_index
= -1;
1697 /* Ignore very old stuff early */
1698 if (!after(sp
[used_sacks
].end_seq
, prior_snd_una
))
1704 /* order SACK blocks to allow in order walk of the retrans queue */
1705 for (i
= used_sacks
- 1; i
> 0; i
--) {
1706 for (j
= 0; j
< i
; j
++) {
1707 if (after(sp
[j
].start_seq
, sp
[j
+ 1].start_seq
)) {
1708 swap(sp
[j
], sp
[j
+ 1]);
1710 /* Track where the first SACK block goes to */
1711 if (j
== first_sack_index
)
1712 first_sack_index
= j
+ 1;
1717 skb
= tcp_write_queue_head(sk
);
1718 state
.fack_count
= 0;
1721 if (!tp
->sacked_out
) {
1722 /* It's already past, so skip checking against it */
1723 cache
= tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1725 cache
= tp
->recv_sack_cache
;
1726 /* Skip empty blocks in at head of the cache */
1727 while (tcp_sack_cache_ok(tp
, cache
) && !cache
->start_seq
&&
1732 while (i
< used_sacks
) {
1733 u32 start_seq
= sp
[i
].start_seq
;
1734 u32 end_seq
= sp
[i
].end_seq
;
1735 bool dup_sack
= (found_dup_sack
&& (i
== first_sack_index
));
1736 struct tcp_sack_block
*next_dup
= NULL
;
1738 if (found_dup_sack
&& ((i
+ 1) == first_sack_index
))
1739 next_dup
= &sp
[i
+ 1];
1741 /* Skip too early cached blocks */
1742 while (tcp_sack_cache_ok(tp
, cache
) &&
1743 !before(start_seq
, cache
->end_seq
))
1746 /* Can skip some work by looking recv_sack_cache? */
1747 if (tcp_sack_cache_ok(tp
, cache
) && !dup_sack
&&
1748 after(end_seq
, cache
->start_seq
)) {
1751 if (before(start_seq
, cache
->start_seq
)) {
1752 skb
= tcp_sacktag_skip(skb
, sk
, &state
,
1754 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
,
1761 /* Rest of the block already fully processed? */
1762 if (!after(end_seq
, cache
->end_seq
))
1765 skb
= tcp_maybe_skipping_dsack(skb
, sk
, next_dup
,
1769 /* ...tail remains todo... */
1770 if (tcp_highest_sack_seq(tp
) == cache
->end_seq
) {
1771 /* ...but better entrypoint exists! */
1772 skb
= tcp_highest_sack(sk
);
1775 state
.fack_count
= tp
->fackets_out
;
1780 skb
= tcp_sacktag_skip(skb
, sk
, &state
, cache
->end_seq
);
1781 /* Check overlap against next cached too (past this one already) */
1786 if (!before(start_seq
, tcp_highest_sack_seq(tp
))) {
1787 skb
= tcp_highest_sack(sk
);
1790 state
.fack_count
= tp
->fackets_out
;
1792 skb
= tcp_sacktag_skip(skb
, sk
, &state
, start_seq
);
1795 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
, &state
,
1796 start_seq
, end_seq
, dup_sack
);
1802 /* Clear the head of the cache sack blocks so we can skip it next time */
1803 for (i
= 0; i
< ARRAY_SIZE(tp
->recv_sack_cache
) - used_sacks
; i
++) {
1804 tp
->recv_sack_cache
[i
].start_seq
= 0;
1805 tp
->recv_sack_cache
[i
].end_seq
= 0;
1807 for (j
= 0; j
< used_sacks
; j
++)
1808 tp
->recv_sack_cache
[i
++] = sp
[j
];
1810 tcp_mark_lost_retrans(sk
);
1812 tcp_verify_left_out(tp
);
1814 if ((state
.reord
< tp
->fackets_out
) &&
1815 ((inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Loss
) || tp
->undo_marker
))
1816 tcp_update_reordering(sk
, tp
->fackets_out
- state
.reord
, 0);
1820 #if FASTRETRANS_DEBUG > 0
1821 WARN_ON((int)tp
->sacked_out
< 0);
1822 WARN_ON((int)tp
->lost_out
< 0);
1823 WARN_ON((int)tp
->retrans_out
< 0);
1824 WARN_ON((int)tcp_packets_in_flight(tp
) < 0);
1826 *sack_rtt_us
= state
.rtt_us
;
1830 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1831 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1833 static bool tcp_limit_reno_sacked(struct tcp_sock
*tp
)
1837 holes
= max(tp
->lost_out
, 1U);
1838 holes
= min(holes
, tp
->packets_out
);
1840 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1841 tp
->sacked_out
= tp
->packets_out
- holes
;
1847 /* If we receive more dupacks than we expected counting segments
1848 * in assumption of absent reordering, interpret this as reordering.
1849 * The only another reason could be bug in receiver TCP.
1851 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1853 struct tcp_sock
*tp
= tcp_sk(sk
);
1854 if (tcp_limit_reno_sacked(tp
))
1855 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1858 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1860 static void tcp_add_reno_sack(struct sock
*sk
)
1862 struct tcp_sock
*tp
= tcp_sk(sk
);
1864 tcp_check_reno_reordering(sk
, 0);
1865 tcp_verify_left_out(tp
);
1868 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1870 static void tcp_remove_reno_sacks(struct sock
*sk
, int acked
)
1872 struct tcp_sock
*tp
= tcp_sk(sk
);
1875 /* One ACK acked hole. The rest eat duplicate ACKs. */
1876 if (acked
- 1 >= tp
->sacked_out
)
1879 tp
->sacked_out
-= acked
- 1;
1881 tcp_check_reno_reordering(sk
, acked
);
1882 tcp_verify_left_out(tp
);
1885 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1890 static void tcp_clear_retrans_partial(struct tcp_sock
*tp
)
1892 tp
->retrans_out
= 0;
1895 tp
->undo_marker
= 0;
1896 tp
->undo_retrans
= 0;
1899 void tcp_clear_retrans(struct tcp_sock
*tp
)
1901 tcp_clear_retrans_partial(tp
);
1903 tp
->fackets_out
= 0;
1907 /* Enter Loss state. If "how" is not zero, forget all SACK information
1908 * and reset tags completely, otherwise preserve SACKs. If receiver
1909 * dropped its ofo queue, we will know this due to reneging detection.
1911 void tcp_enter_loss(struct sock
*sk
, int how
)
1913 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1914 struct tcp_sock
*tp
= tcp_sk(sk
);
1915 struct sk_buff
*skb
;
1916 bool new_recovery
= false;
1918 /* Reduce ssthresh if it has not yet been made inside this window. */
1919 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
||
1920 !after(tp
->high_seq
, tp
->snd_una
) ||
1921 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
1922 new_recovery
= true;
1923 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1924 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1925 tcp_ca_event(sk
, CA_EVENT_LOSS
);
1928 tp
->snd_cwnd_cnt
= 0;
1929 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1931 tcp_clear_retrans_partial(tp
);
1933 if (tcp_is_reno(tp
))
1934 tcp_reset_reno_sack(tp
);
1936 tp
->undo_marker
= tp
->snd_una
;
1939 tp
->fackets_out
= 0;
1941 tcp_clear_all_retrans_hints(tp
);
1943 tcp_for_write_queue(skb
, sk
) {
1944 if (skb
== tcp_send_head(sk
))
1947 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
1948 tp
->undo_marker
= 0;
1950 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
1951 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) || how
) {
1952 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
1953 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1954 tp
->lost_out
+= tcp_skb_pcount(skb
);
1955 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
1958 tcp_verify_left_out(tp
);
1960 /* Timeout in disordered state after receiving substantial DUPACKs
1961 * suggests that the degree of reordering is over-estimated.
1963 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
&&
1964 tp
->sacked_out
>= sysctl_tcp_reordering
)
1965 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1966 sysctl_tcp_reordering
);
1967 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1968 tp
->high_seq
= tp
->snd_nxt
;
1969 TCP_ECN_queue_cwr(tp
);
1971 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1972 * loss recovery is underway except recurring timeout(s) on
1973 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1975 tp
->frto
= sysctl_tcp_frto
&&
1976 (new_recovery
|| icsk
->icsk_retransmits
) &&
1977 !inet_csk(sk
)->icsk_mtup
.probe_size
;
1980 /* If ACK arrived pointing to a remembered SACK, it means that our
1981 * remembered SACKs do not reflect real state of receiver i.e.
1982 * receiver _host_ is heavily congested (or buggy).
1984 * Do processing similar to RTO timeout.
1986 static bool tcp_check_sack_reneging(struct sock
*sk
, int flag
)
1988 if (flag
& FLAG_SACK_RENEGING
) {
1989 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1990 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSACKRENEGING
);
1992 tcp_enter_loss(sk
, 1);
1993 icsk
->icsk_retransmits
++;
1994 tcp_retransmit_skb(sk
, tcp_write_queue_head(sk
));
1995 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1996 icsk
->icsk_rto
, TCP_RTO_MAX
);
2002 static inline int tcp_fackets_out(const struct tcp_sock
*tp
)
2004 return tcp_is_reno(tp
) ? tp
->sacked_out
+ 1 : tp
->fackets_out
;
2007 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2008 * counter when SACK is enabled (without SACK, sacked_out is used for
2011 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2012 * segments up to the highest received SACK block so far and holes in
2015 * With reordering, holes may still be in flight, so RFC3517 recovery
2016 * uses pure sacked_out (total number of SACKed segments) even though
2017 * it violates the RFC that uses duplicate ACKs, often these are equal
2018 * but when e.g. out-of-window ACKs or packet duplication occurs,
2019 * they differ. Since neither occurs due to loss, TCP should really
2022 static inline int tcp_dupack_heuristics(const struct tcp_sock
*tp
)
2024 return tcp_is_fack(tp
) ? tp
->fackets_out
: tp
->sacked_out
+ 1;
2027 static bool tcp_pause_early_retransmit(struct sock
*sk
, int flag
)
2029 struct tcp_sock
*tp
= tcp_sk(sk
);
2030 unsigned long delay
;
2032 /* Delay early retransmit and entering fast recovery for
2033 * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
2034 * available, or RTO is scheduled to fire first.
2036 if (sysctl_tcp_early_retrans
< 2 || sysctl_tcp_early_retrans
> 3 ||
2037 (flag
& FLAG_ECE
) || !tp
->srtt_us
)
2040 delay
= max(usecs_to_jiffies(tp
->srtt_us
>> 5),
2041 msecs_to_jiffies(2));
2043 if (!time_after(inet_csk(sk
)->icsk_timeout
, (jiffies
+ delay
)))
2046 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_EARLY_RETRANS
, delay
,
2051 /* Linux NewReno/SACK/FACK/ECN state machine.
2052 * --------------------------------------
2054 * "Open" Normal state, no dubious events, fast path.
2055 * "Disorder" In all the respects it is "Open",
2056 * but requires a bit more attention. It is entered when
2057 * we see some SACKs or dupacks. It is split of "Open"
2058 * mainly to move some processing from fast path to slow one.
2059 * "CWR" CWND was reduced due to some Congestion Notification event.
2060 * It can be ECN, ICMP source quench, local device congestion.
2061 * "Recovery" CWND was reduced, we are fast-retransmitting.
2062 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2064 * tcp_fastretrans_alert() is entered:
2065 * - each incoming ACK, if state is not "Open"
2066 * - when arrived ACK is unusual, namely:
2071 * Counting packets in flight is pretty simple.
2073 * in_flight = packets_out - left_out + retrans_out
2075 * packets_out is SND.NXT-SND.UNA counted in packets.
2077 * retrans_out is number of retransmitted segments.
2079 * left_out is number of segments left network, but not ACKed yet.
2081 * left_out = sacked_out + lost_out
2083 * sacked_out: Packets, which arrived to receiver out of order
2084 * and hence not ACKed. With SACKs this number is simply
2085 * amount of SACKed data. Even without SACKs
2086 * it is easy to give pretty reliable estimate of this number,
2087 * counting duplicate ACKs.
2089 * lost_out: Packets lost by network. TCP has no explicit
2090 * "loss notification" feedback from network (for now).
2091 * It means that this number can be only _guessed_.
2092 * Actually, it is the heuristics to predict lossage that
2093 * distinguishes different algorithms.
2095 * F.e. after RTO, when all the queue is considered as lost,
2096 * lost_out = packets_out and in_flight = retrans_out.
2098 * Essentially, we have now two algorithms counting
2101 * FACK: It is the simplest heuristics. As soon as we decided
2102 * that something is lost, we decide that _all_ not SACKed
2103 * packets until the most forward SACK are lost. I.e.
2104 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2105 * It is absolutely correct estimate, if network does not reorder
2106 * packets. And it loses any connection to reality when reordering
2107 * takes place. We use FACK by default until reordering
2108 * is suspected on the path to this destination.
2110 * NewReno: when Recovery is entered, we assume that one segment
2111 * is lost (classic Reno). While we are in Recovery and
2112 * a partial ACK arrives, we assume that one more packet
2113 * is lost (NewReno). This heuristics are the same in NewReno
2116 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2117 * deflation etc. CWND is real congestion window, never inflated, changes
2118 * only according to classic VJ rules.
2120 * Really tricky (and requiring careful tuning) part of algorithm
2121 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2122 * The first determines the moment _when_ we should reduce CWND and,
2123 * hence, slow down forward transmission. In fact, it determines the moment
2124 * when we decide that hole is caused by loss, rather than by a reorder.
2126 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2127 * holes, caused by lost packets.
2129 * And the most logically complicated part of algorithm is undo
2130 * heuristics. We detect false retransmits due to both too early
2131 * fast retransmit (reordering) and underestimated RTO, analyzing
2132 * timestamps and D-SACKs. When we detect that some segments were
2133 * retransmitted by mistake and CWND reduction was wrong, we undo
2134 * window reduction and abort recovery phase. This logic is hidden
2135 * inside several functions named tcp_try_undo_<something>.
2138 /* This function decides, when we should leave Disordered state
2139 * and enter Recovery phase, reducing congestion window.
2141 * Main question: may we further continue forward transmission
2142 * with the same cwnd?
2144 static bool tcp_time_to_recover(struct sock
*sk
, int flag
)
2146 struct tcp_sock
*tp
= tcp_sk(sk
);
2149 /* Trick#1: The loss is proven. */
2153 /* Not-A-Trick#2 : Classic rule... */
2154 if (tcp_dupack_heuristics(tp
) > tp
->reordering
)
2157 /* Trick#4: It is still not OK... But will it be useful to delay
2160 packets_out
= tp
->packets_out
;
2161 if (packets_out
<= tp
->reordering
&&
2162 tp
->sacked_out
>= max_t(__u32
, packets_out
/2, sysctl_tcp_reordering
) &&
2163 !tcp_may_send_now(sk
)) {
2164 /* We have nothing to send. This connection is limited
2165 * either by receiver window or by application.
2170 /* If a thin stream is detected, retransmit after first
2171 * received dupack. Employ only if SACK is supported in order
2172 * to avoid possible corner-case series of spurious retransmissions
2173 * Use only if there are no unsent data.
2175 if ((tp
->thin_dupack
|| sysctl_tcp_thin_dupack
) &&
2176 tcp_stream_is_thin(tp
) && tcp_dupack_heuristics(tp
) > 1 &&
2177 tcp_is_sack(tp
) && !tcp_send_head(sk
))
2180 /* Trick#6: TCP early retransmit, per RFC5827. To avoid spurious
2181 * retransmissions due to small network reorderings, we implement
2182 * Mitigation A.3 in the RFC and delay the retransmission for a short
2183 * interval if appropriate.
2185 if (tp
->do_early_retrans
&& !tp
->retrans_out
&& tp
->sacked_out
&&
2186 (tp
->packets_out
>= (tp
->sacked_out
+ 1) && tp
->packets_out
< 4) &&
2187 !tcp_may_send_now(sk
))
2188 return !tcp_pause_early_retransmit(sk
, flag
);
2193 /* Detect loss in event "A" above by marking head of queue up as lost.
2194 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2195 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2196 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2197 * the maximum SACKed segments to pass before reaching this limit.
2199 static void tcp_mark_head_lost(struct sock
*sk
, int packets
, int mark_head
)
2201 struct tcp_sock
*tp
= tcp_sk(sk
);
2202 struct sk_buff
*skb
;
2206 /* Use SACK to deduce losses of new sequences sent during recovery */
2207 const u32 loss_high
= tcp_is_sack(tp
) ? tp
->snd_nxt
: tp
->high_seq
;
2209 WARN_ON(packets
> tp
->packets_out
);
2210 if (tp
->lost_skb_hint
) {
2211 skb
= tp
->lost_skb_hint
;
2212 cnt
= tp
->lost_cnt_hint
;
2213 /* Head already handled? */
2214 if (mark_head
&& skb
!= tcp_write_queue_head(sk
))
2217 skb
= tcp_write_queue_head(sk
);
2221 tcp_for_write_queue_from(skb
, sk
) {
2222 if (skb
== tcp_send_head(sk
))
2224 /* TODO: do this better */
2225 /* this is not the most efficient way to do this... */
2226 tp
->lost_skb_hint
= skb
;
2227 tp
->lost_cnt_hint
= cnt
;
2229 if (after(TCP_SKB_CB(skb
)->end_seq
, loss_high
))
2233 if (tcp_is_fack(tp
) || tcp_is_reno(tp
) ||
2234 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2235 cnt
+= tcp_skb_pcount(skb
);
2237 if (cnt
> packets
) {
2238 if ((tcp_is_sack(tp
) && !tcp_is_fack(tp
)) ||
2239 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
) ||
2240 (oldcnt
>= packets
))
2243 mss
= skb_shinfo(skb
)->gso_size
;
2244 err
= tcp_fragment(sk
, skb
, (packets
- oldcnt
) * mss
, mss
);
2250 tcp_skb_mark_lost(tp
, skb
);
2255 tcp_verify_left_out(tp
);
2258 /* Account newly detected lost packet(s) */
2260 static void tcp_update_scoreboard(struct sock
*sk
, int fast_rexmit
)
2262 struct tcp_sock
*tp
= tcp_sk(sk
);
2264 if (tcp_is_reno(tp
)) {
2265 tcp_mark_head_lost(sk
, 1, 1);
2266 } else if (tcp_is_fack(tp
)) {
2267 int lost
= tp
->fackets_out
- tp
->reordering
;
2270 tcp_mark_head_lost(sk
, lost
, 0);
2272 int sacked_upto
= tp
->sacked_out
- tp
->reordering
;
2273 if (sacked_upto
>= 0)
2274 tcp_mark_head_lost(sk
, sacked_upto
, 0);
2275 else if (fast_rexmit
)
2276 tcp_mark_head_lost(sk
, 1, 1);
2280 /* CWND moderation, preventing bursts due to too big ACKs
2281 * in dubious situations.
2283 static inline void tcp_moderate_cwnd(struct tcp_sock
*tp
)
2285 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
2286 tcp_packets_in_flight(tp
) + tcp_max_burst(tp
));
2287 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2290 /* Nothing was retransmitted or returned timestamp is less
2291 * than timestamp of the first retransmission.
2293 static inline bool tcp_packet_delayed(const struct tcp_sock
*tp
)
2295 return !tp
->retrans_stamp
||
2296 (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2297 before(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
));
2300 /* Undo procedures. */
2302 #if FASTRETRANS_DEBUG > 1
2303 static void DBGUNDO(struct sock
*sk
, const char *msg
)
2305 struct tcp_sock
*tp
= tcp_sk(sk
);
2306 struct inet_sock
*inet
= inet_sk(sk
);
2308 if (sk
->sk_family
== AF_INET
) {
2309 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2311 &inet
->inet_daddr
, ntohs(inet
->inet_dport
),
2312 tp
->snd_cwnd
, tcp_left_out(tp
),
2313 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2316 #if IS_ENABLED(CONFIG_IPV6)
2317 else if (sk
->sk_family
== AF_INET6
) {
2318 struct ipv6_pinfo
*np
= inet6_sk(sk
);
2319 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2321 &np
->daddr
, ntohs(inet
->inet_dport
),
2322 tp
->snd_cwnd
, tcp_left_out(tp
),
2323 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2329 #define DBGUNDO(x...) do { } while (0)
2332 static void tcp_undo_cwnd_reduction(struct sock
*sk
, bool unmark_loss
)
2334 struct tcp_sock
*tp
= tcp_sk(sk
);
2337 struct sk_buff
*skb
;
2339 tcp_for_write_queue(skb
, sk
) {
2340 if (skb
== tcp_send_head(sk
))
2342 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2345 tcp_clear_all_retrans_hints(tp
);
2348 if (tp
->prior_ssthresh
) {
2349 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2351 if (icsk
->icsk_ca_ops
->undo_cwnd
)
2352 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
2354 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
<< 1);
2356 if (tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
2357 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
2358 TCP_ECN_withdraw_cwr(tp
);
2361 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2363 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2364 tp
->undo_marker
= 0;
2367 static inline bool tcp_may_undo(const struct tcp_sock
*tp
)
2369 return tp
->undo_marker
&& (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
2372 /* People celebrate: "We love our President!" */
2373 static bool tcp_try_undo_recovery(struct sock
*sk
)
2375 struct tcp_sock
*tp
= tcp_sk(sk
);
2377 if (tcp_may_undo(tp
)) {
2380 /* Happy end! We did not retransmit anything
2381 * or our original transmission succeeded.
2383 DBGUNDO(sk
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
2384 tcp_undo_cwnd_reduction(sk
, false);
2385 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
2386 mib_idx
= LINUX_MIB_TCPLOSSUNDO
;
2388 mib_idx
= LINUX_MIB_TCPFULLUNDO
;
2390 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2392 if (tp
->snd_una
== tp
->high_seq
&& tcp_is_reno(tp
)) {
2393 /* Hold old state until something *above* high_seq
2394 * is ACKed. For Reno it is MUST to prevent false
2395 * fast retransmits (RFC2582). SACK TCP is safe. */
2396 tcp_moderate_cwnd(tp
);
2399 tcp_set_ca_state(sk
, TCP_CA_Open
);
2403 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2404 static bool tcp_try_undo_dsack(struct sock
*sk
)
2406 struct tcp_sock
*tp
= tcp_sk(sk
);
2408 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
2409 DBGUNDO(sk
, "D-SACK");
2410 tcp_undo_cwnd_reduction(sk
, false);
2411 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPDSACKUNDO
);
2417 /* We can clear retrans_stamp when there are no retransmissions in the
2418 * window. It would seem that it is trivially available for us in
2419 * tp->retrans_out, however, that kind of assumptions doesn't consider
2420 * what will happen if errors occur when sending retransmission for the
2421 * second time. ...It could the that such segment has only
2422 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2423 * the head skb is enough except for some reneging corner cases that
2424 * are not worth the effort.
2426 * Main reason for all this complexity is the fact that connection dying
2427 * time now depends on the validity of the retrans_stamp, in particular,
2428 * that successive retransmissions of a segment must not advance
2429 * retrans_stamp under any conditions.
2431 static bool tcp_any_retrans_done(const struct sock
*sk
)
2433 const struct tcp_sock
*tp
= tcp_sk(sk
);
2434 struct sk_buff
*skb
;
2436 if (tp
->retrans_out
)
2439 skb
= tcp_write_queue_head(sk
);
2440 if (unlikely(skb
&& TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
))
2446 /* Undo during loss recovery after partial ACK or using F-RTO. */
2447 static bool tcp_try_undo_loss(struct sock
*sk
, bool frto_undo
)
2449 struct tcp_sock
*tp
= tcp_sk(sk
);
2451 if (frto_undo
|| tcp_may_undo(tp
)) {
2452 tcp_undo_cwnd_reduction(sk
, true);
2454 DBGUNDO(sk
, "partial loss");
2455 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSSUNDO
);
2457 NET_INC_STATS_BH(sock_net(sk
),
2458 LINUX_MIB_TCPSPURIOUSRTOS
);
2459 inet_csk(sk
)->icsk_retransmits
= 0;
2460 if (frto_undo
|| tcp_is_sack(tp
))
2461 tcp_set_ca_state(sk
, TCP_CA_Open
);
2467 /* The cwnd reduction in CWR and Recovery use the PRR algorithm
2468 * https://datatracker.ietf.org/doc/draft-ietf-tcpm-proportional-rate-reduction/
2469 * It computes the number of packets to send (sndcnt) based on packets newly
2471 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2472 * cwnd reductions across a full RTT.
2473 * 2) If packets in flight is lower than ssthresh (such as due to excess
2474 * losses and/or application stalls), do not perform any further cwnd
2475 * reductions, but instead slow start up to ssthresh.
2477 static void tcp_init_cwnd_reduction(struct sock
*sk
, const bool set_ssthresh
)
2479 struct tcp_sock
*tp
= tcp_sk(sk
);
2481 tp
->high_seq
= tp
->snd_nxt
;
2482 tp
->tlp_high_seq
= 0;
2483 tp
->snd_cwnd_cnt
= 0;
2484 tp
->prior_cwnd
= tp
->snd_cwnd
;
2485 tp
->prr_delivered
= 0;
2488 tp
->snd_ssthresh
= inet_csk(sk
)->icsk_ca_ops
->ssthresh(sk
);
2489 TCP_ECN_queue_cwr(tp
);
2492 static void tcp_cwnd_reduction(struct sock
*sk
, const int prior_unsacked
,
2495 struct tcp_sock
*tp
= tcp_sk(sk
);
2497 int delta
= tp
->snd_ssthresh
- tcp_packets_in_flight(tp
);
2498 int newly_acked_sacked
= prior_unsacked
-
2499 (tp
->packets_out
- tp
->sacked_out
);
2501 tp
->prr_delivered
+= newly_acked_sacked
;
2502 if (tcp_packets_in_flight(tp
) > tp
->snd_ssthresh
) {
2503 u64 dividend
= (u64
)tp
->snd_ssthresh
* tp
->prr_delivered
+
2505 sndcnt
= div_u64(dividend
, tp
->prior_cwnd
) - tp
->prr_out
;
2507 sndcnt
= min_t(int, delta
,
2508 max_t(int, tp
->prr_delivered
- tp
->prr_out
,
2509 newly_acked_sacked
) + 1);
2512 sndcnt
= max(sndcnt
, (fast_rexmit
? 1 : 0));
2513 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + sndcnt
;
2516 static inline void tcp_end_cwnd_reduction(struct sock
*sk
)
2518 struct tcp_sock
*tp
= tcp_sk(sk
);
2520 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2521 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_CWR
||
2522 (tp
->undo_marker
&& tp
->snd_ssthresh
< TCP_INFINITE_SSTHRESH
)) {
2523 tp
->snd_cwnd
= tp
->snd_ssthresh
;
2524 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2526 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
2529 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2530 void tcp_enter_cwr(struct sock
*sk
, const int set_ssthresh
)
2532 struct tcp_sock
*tp
= tcp_sk(sk
);
2534 tp
->prior_ssthresh
= 0;
2535 if (inet_csk(sk
)->icsk_ca_state
< TCP_CA_CWR
) {
2536 tp
->undo_marker
= 0;
2537 tcp_init_cwnd_reduction(sk
, set_ssthresh
);
2538 tcp_set_ca_state(sk
, TCP_CA_CWR
);
2542 static void tcp_try_keep_open(struct sock
*sk
)
2544 struct tcp_sock
*tp
= tcp_sk(sk
);
2545 int state
= TCP_CA_Open
;
2547 if (tcp_left_out(tp
) || tcp_any_retrans_done(sk
))
2548 state
= TCP_CA_Disorder
;
2550 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
2551 tcp_set_ca_state(sk
, state
);
2552 tp
->high_seq
= tp
->snd_nxt
;
2556 static void tcp_try_to_open(struct sock
*sk
, int flag
, const int prior_unsacked
)
2558 struct tcp_sock
*tp
= tcp_sk(sk
);
2560 tcp_verify_left_out(tp
);
2562 if (!tcp_any_retrans_done(sk
))
2563 tp
->retrans_stamp
= 0;
2565 if (flag
& FLAG_ECE
)
2566 tcp_enter_cwr(sk
, 1);
2568 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
2569 tcp_try_keep_open(sk
);
2571 tcp_cwnd_reduction(sk
, prior_unsacked
, 0);
2575 static void tcp_mtup_probe_failed(struct sock
*sk
)
2577 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2579 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
2580 icsk
->icsk_mtup
.probe_size
= 0;
2583 static void tcp_mtup_probe_success(struct sock
*sk
)
2585 struct tcp_sock
*tp
= tcp_sk(sk
);
2586 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2588 /* FIXME: breaks with very large cwnd */
2589 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2590 tp
->snd_cwnd
= tp
->snd_cwnd
*
2591 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
2592 icsk
->icsk_mtup
.probe_size
;
2593 tp
->snd_cwnd_cnt
= 0;
2594 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2595 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2597 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
2598 icsk
->icsk_mtup
.probe_size
= 0;
2599 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
2602 /* Do a simple retransmit without using the backoff mechanisms in
2603 * tcp_timer. This is used for path mtu discovery.
2604 * The socket is already locked here.
2606 void tcp_simple_retransmit(struct sock
*sk
)
2608 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2609 struct tcp_sock
*tp
= tcp_sk(sk
);
2610 struct sk_buff
*skb
;
2611 unsigned int mss
= tcp_current_mss(sk
);
2612 u32 prior_lost
= tp
->lost_out
;
2614 tcp_for_write_queue(skb
, sk
) {
2615 if (skb
== tcp_send_head(sk
))
2617 if (tcp_skb_seglen(skb
) > mss
&&
2618 !(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
2619 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2620 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
2621 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2623 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
2627 tcp_clear_retrans_hints_partial(tp
);
2629 if (prior_lost
== tp
->lost_out
)
2632 if (tcp_is_reno(tp
))
2633 tcp_limit_reno_sacked(tp
);
2635 tcp_verify_left_out(tp
);
2637 /* Don't muck with the congestion window here.
2638 * Reason is that we do not increase amount of _data_
2639 * in network, but units changed and effective
2640 * cwnd/ssthresh really reduced now.
2642 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
2643 tp
->high_seq
= tp
->snd_nxt
;
2644 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2645 tp
->prior_ssthresh
= 0;
2646 tp
->undo_marker
= 0;
2647 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2649 tcp_xmit_retransmit_queue(sk
);
2651 EXPORT_SYMBOL(tcp_simple_retransmit
);
2653 static void tcp_enter_recovery(struct sock
*sk
, bool ece_ack
)
2655 struct tcp_sock
*tp
= tcp_sk(sk
);
2658 if (tcp_is_reno(tp
))
2659 mib_idx
= LINUX_MIB_TCPRENORECOVERY
;
2661 mib_idx
= LINUX_MIB_TCPSACKRECOVERY
;
2663 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2665 tp
->prior_ssthresh
= 0;
2666 tp
->undo_marker
= tp
->snd_una
;
2667 tp
->undo_retrans
= tp
->retrans_out
;
2669 if (inet_csk(sk
)->icsk_ca_state
< TCP_CA_CWR
) {
2671 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2672 tcp_init_cwnd_reduction(sk
, true);
2674 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
2677 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2678 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2680 static void tcp_process_loss(struct sock
*sk
, int flag
, bool is_dupack
)
2682 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2683 struct tcp_sock
*tp
= tcp_sk(sk
);
2684 bool recovered
= !before(tp
->snd_una
, tp
->high_seq
);
2686 if (tp
->frto
) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2687 if (flag
& FLAG_ORIG_SACK_ACKED
) {
2688 /* Step 3.b. A timeout is spurious if not all data are
2689 * lost, i.e., never-retransmitted data are (s)acked.
2691 tcp_try_undo_loss(sk
, true);
2694 if (after(tp
->snd_nxt
, tp
->high_seq
) &&
2695 (flag
& FLAG_DATA_SACKED
|| is_dupack
)) {
2696 tp
->frto
= 0; /* Loss was real: 2nd part of step 3.a */
2697 } else if (flag
& FLAG_SND_UNA_ADVANCED
&& !recovered
) {
2698 tp
->high_seq
= tp
->snd_nxt
;
2699 __tcp_push_pending_frames(sk
, tcp_current_mss(sk
),
2701 if (after(tp
->snd_nxt
, tp
->high_seq
))
2702 return; /* Step 2.b */
2708 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2709 icsk
->icsk_retransmits
= 0;
2710 tcp_try_undo_recovery(sk
);
2713 if (flag
& FLAG_DATA_ACKED
)
2714 icsk
->icsk_retransmits
= 0;
2715 if (tcp_is_reno(tp
)) {
2716 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2717 * delivered. Lower inflight to clock out (re)tranmissions.
2719 if (after(tp
->snd_nxt
, tp
->high_seq
) && is_dupack
)
2720 tcp_add_reno_sack(sk
);
2721 else if (flag
& FLAG_SND_UNA_ADVANCED
)
2722 tcp_reset_reno_sack(tp
);
2724 if (tcp_try_undo_loss(sk
, false))
2726 tcp_xmit_retransmit_queue(sk
);
2729 /* Undo during fast recovery after partial ACK. */
2730 static bool tcp_try_undo_partial(struct sock
*sk
, const int acked
,
2731 const int prior_unsacked
)
2733 struct tcp_sock
*tp
= tcp_sk(sk
);
2735 if (tp
->undo_marker
&& tcp_packet_delayed(tp
)) {
2736 /* Plain luck! Hole if filled with delayed
2737 * packet, rather than with a retransmit.
2739 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
2741 /* We are getting evidence that the reordering degree is higher
2742 * than we realized. If there are no retransmits out then we
2743 * can undo. Otherwise we clock out new packets but do not
2744 * mark more packets lost or retransmit more.
2746 if (tp
->retrans_out
) {
2747 tcp_cwnd_reduction(sk
, prior_unsacked
, 0);
2751 if (!tcp_any_retrans_done(sk
))
2752 tp
->retrans_stamp
= 0;
2754 DBGUNDO(sk
, "partial recovery");
2755 tcp_undo_cwnd_reduction(sk
, true);
2756 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPARTIALUNDO
);
2757 tcp_try_keep_open(sk
);
2763 /* Process an event, which can update packets-in-flight not trivially.
2764 * Main goal of this function is to calculate new estimate for left_out,
2765 * taking into account both packets sitting in receiver's buffer and
2766 * packets lost by network.
2768 * Besides that it does CWND reduction, when packet loss is detected
2769 * and changes state of machine.
2771 * It does _not_ decide what to send, it is made in function
2772 * tcp_xmit_retransmit_queue().
2774 static void tcp_fastretrans_alert(struct sock
*sk
, const int acked
,
2775 const int prior_unsacked
,
2776 bool is_dupack
, int flag
)
2778 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2779 struct tcp_sock
*tp
= tcp_sk(sk
);
2780 bool do_lost
= is_dupack
|| ((flag
& FLAG_DATA_SACKED
) &&
2781 (tcp_fackets_out(tp
) > tp
->reordering
));
2782 int fast_rexmit
= 0;
2784 if (WARN_ON(!tp
->packets_out
&& tp
->sacked_out
))
2786 if (WARN_ON(!tp
->sacked_out
&& tp
->fackets_out
))
2787 tp
->fackets_out
= 0;
2789 /* Now state machine starts.
2790 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2791 if (flag
& FLAG_ECE
)
2792 tp
->prior_ssthresh
= 0;
2794 /* B. In all the states check for reneging SACKs. */
2795 if (tcp_check_sack_reneging(sk
, flag
))
2798 /* C. Check consistency of the current state. */
2799 tcp_verify_left_out(tp
);
2801 /* D. Check state exit conditions. State can be terminated
2802 * when high_seq is ACKed. */
2803 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
2804 WARN_ON(tp
->retrans_out
!= 0);
2805 tp
->retrans_stamp
= 0;
2806 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
2807 switch (icsk
->icsk_ca_state
) {
2809 /* CWR is to be held something *above* high_seq
2810 * is ACKed for CWR bit to reach receiver. */
2811 if (tp
->snd_una
!= tp
->high_seq
) {
2812 tcp_end_cwnd_reduction(sk
);
2813 tcp_set_ca_state(sk
, TCP_CA_Open
);
2817 case TCP_CA_Recovery
:
2818 if (tcp_is_reno(tp
))
2819 tcp_reset_reno_sack(tp
);
2820 if (tcp_try_undo_recovery(sk
))
2822 tcp_end_cwnd_reduction(sk
);
2827 /* E. Process state. */
2828 switch (icsk
->icsk_ca_state
) {
2829 case TCP_CA_Recovery
:
2830 if (!(flag
& FLAG_SND_UNA_ADVANCED
)) {
2831 if (tcp_is_reno(tp
) && is_dupack
)
2832 tcp_add_reno_sack(sk
);
2834 if (tcp_try_undo_partial(sk
, acked
, prior_unsacked
))
2836 /* Partial ACK arrived. Force fast retransmit. */
2837 do_lost
= tcp_is_reno(tp
) ||
2838 tcp_fackets_out(tp
) > tp
->reordering
;
2840 if (tcp_try_undo_dsack(sk
)) {
2841 tcp_try_keep_open(sk
);
2846 tcp_process_loss(sk
, flag
, is_dupack
);
2847 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
2849 /* Fall through to processing in Open state. */
2851 if (tcp_is_reno(tp
)) {
2852 if (flag
& FLAG_SND_UNA_ADVANCED
)
2853 tcp_reset_reno_sack(tp
);
2855 tcp_add_reno_sack(sk
);
2858 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
)
2859 tcp_try_undo_dsack(sk
);
2861 if (!tcp_time_to_recover(sk
, flag
)) {
2862 tcp_try_to_open(sk
, flag
, prior_unsacked
);
2866 /* MTU probe failure: don't reduce cwnd */
2867 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
2868 icsk
->icsk_mtup
.probe_size
&&
2869 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
2870 tcp_mtup_probe_failed(sk
);
2871 /* Restores the reduction we did in tcp_mtup_probe() */
2873 tcp_simple_retransmit(sk
);
2877 /* Otherwise enter Recovery state */
2878 tcp_enter_recovery(sk
, (flag
& FLAG_ECE
));
2883 tcp_update_scoreboard(sk
, fast_rexmit
);
2884 tcp_cwnd_reduction(sk
, prior_unsacked
, fast_rexmit
);
2885 tcp_xmit_retransmit_queue(sk
);
2888 static inline bool tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
2889 long seq_rtt_us
, long sack_rtt_us
)
2891 const struct tcp_sock
*tp
= tcp_sk(sk
);
2893 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2894 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2895 * Karn's algorithm forbids taking RTT if some retransmitted data
2896 * is acked (RFC6298).
2898 if (flag
& FLAG_RETRANS_DATA_ACKED
)
2902 seq_rtt_us
= sack_rtt_us
;
2904 /* RTTM Rule: A TSecr value received in a segment is used to
2905 * update the averaged RTT measurement only if the segment
2906 * acknowledges some new data, i.e., only if it advances the
2907 * left edge of the send window.
2908 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2910 if (seq_rtt_us
< 0 && tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2912 seq_rtt_us
= jiffies_to_usecs(tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
);
2917 tcp_rtt_estimator(sk
, seq_rtt_us
);
2920 /* RFC6298: only reset backoff on valid RTT measurement. */
2921 inet_csk(sk
)->icsk_backoff
= 0;
2925 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2926 static void tcp_synack_rtt_meas(struct sock
*sk
, const u32 synack_stamp
)
2928 struct tcp_sock
*tp
= tcp_sk(sk
);
2929 long seq_rtt_us
= -1L;
2931 if (synack_stamp
&& !tp
->total_retrans
)
2932 seq_rtt_us
= jiffies_to_usecs(tcp_time_stamp
- synack_stamp
);
2934 /* If the ACK acks both the SYNACK and the (Fast Open'd) data packets
2935 * sent in SYN_RECV, SYNACK RTT is the smooth RTT computed in tcp_ack()
2938 tcp_ack_update_rtt(sk
, FLAG_SYN_ACKED
, seq_rtt_us
, -1L);
2941 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
, u32 acked
, u32 in_flight
)
2943 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2944 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, acked
, in_flight
);
2945 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_time_stamp
;
2948 /* Restart timer after forward progress on connection.
2949 * RFC2988 recommends to restart timer to now+rto.
2951 void tcp_rearm_rto(struct sock
*sk
)
2953 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2954 struct tcp_sock
*tp
= tcp_sk(sk
);
2956 /* If the retrans timer is currently being used by Fast Open
2957 * for SYN-ACK retrans purpose, stay put.
2959 if (tp
->fastopen_rsk
)
2962 if (!tp
->packets_out
) {
2963 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
2965 u32 rto
= inet_csk(sk
)->icsk_rto
;
2966 /* Offset the time elapsed after installing regular RTO */
2967 if (icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
||
2968 icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
) {
2969 struct sk_buff
*skb
= tcp_write_queue_head(sk
);
2970 const u32 rto_time_stamp
= TCP_SKB_CB(skb
)->when
+ rto
;
2971 s32 delta
= (s32
)(rto_time_stamp
- tcp_time_stamp
);
2972 /* delta may not be positive if the socket is locked
2973 * when the retrans timer fires and is rescheduled.
2978 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
, rto
,
2983 /* This function is called when the delayed ER timer fires. TCP enters
2984 * fast recovery and performs fast-retransmit.
2986 void tcp_resume_early_retransmit(struct sock
*sk
)
2988 struct tcp_sock
*tp
= tcp_sk(sk
);
2992 /* Stop if ER is disabled after the delayed ER timer is scheduled */
2993 if (!tp
->do_early_retrans
)
2996 tcp_enter_recovery(sk
, false);
2997 tcp_update_scoreboard(sk
, 1);
2998 tcp_xmit_retransmit_queue(sk
);
3001 /* If we get here, the whole TSO packet has not been acked. */
3002 static u32
tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
)
3004 struct tcp_sock
*tp
= tcp_sk(sk
);
3007 BUG_ON(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
));
3009 packets_acked
= tcp_skb_pcount(skb
);
3010 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
3012 packets_acked
-= tcp_skb_pcount(skb
);
3014 if (packets_acked
) {
3015 BUG_ON(tcp_skb_pcount(skb
) == 0);
3016 BUG_ON(!before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
));
3019 return packets_acked
;
3022 /* Remove acknowledged frames from the retransmission queue. If our packet
3023 * is before the ack sequence we can discard it as it's confirmed to have
3024 * arrived at the other end.
3026 static int tcp_clean_rtx_queue(struct sock
*sk
, int prior_fackets
,
3027 u32 prior_snd_una
, long sack_rtt_us
)
3029 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3030 struct skb_mstamp first_ackt
, last_ackt
, now
;
3031 struct tcp_sock
*tp
= tcp_sk(sk
);
3032 u32 prior_sacked
= tp
->sacked_out
;
3033 u32 reord
= tp
->packets_out
;
3034 bool fully_acked
= true;
3035 long ca_seq_rtt_us
= -1L;
3036 long seq_rtt_us
= -1L;
3037 struct sk_buff
*skb
;
3044 while ((skb
= tcp_write_queue_head(sk
)) && skb
!= tcp_send_head(sk
)) {
3045 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
3046 u8 sacked
= scb
->sacked
;
3049 /* Determine how many packets and what bytes were acked, tso and else */
3050 if (after(scb
->end_seq
, tp
->snd_una
)) {
3051 if (tcp_skb_pcount(skb
) == 1 ||
3052 !after(tp
->snd_una
, scb
->seq
))
3055 acked_pcount
= tcp_tso_acked(sk
, skb
);
3059 fully_acked
= false;
3061 acked_pcount
= tcp_skb_pcount(skb
);
3064 if (sacked
& TCPCB_RETRANS
) {
3065 if (sacked
& TCPCB_SACKED_RETRANS
)
3066 tp
->retrans_out
-= acked_pcount
;
3067 flag
|= FLAG_RETRANS_DATA_ACKED
;
3069 last_ackt
= skb
->skb_mstamp
;
3070 WARN_ON_ONCE(last_ackt
.v64
== 0);
3071 if (!first_ackt
.v64
)
3072 first_ackt
= last_ackt
;
3074 if (!(sacked
& TCPCB_SACKED_ACKED
))
3075 reord
= min(pkts_acked
, reord
);
3076 if (!after(scb
->end_seq
, tp
->high_seq
))
3077 flag
|= FLAG_ORIG_SACK_ACKED
;
3080 if (sacked
& TCPCB_SACKED_ACKED
)
3081 tp
->sacked_out
-= acked_pcount
;
3082 if (sacked
& TCPCB_LOST
)
3083 tp
->lost_out
-= acked_pcount
;
3085 tp
->packets_out
-= acked_pcount
;
3086 pkts_acked
+= acked_pcount
;
3088 /* Initial outgoing SYN's get put onto the write_queue
3089 * just like anything else we transmit. It is not
3090 * true data, and if we misinform our callers that
3091 * this ACK acks real data, we will erroneously exit
3092 * connection startup slow start one packet too
3093 * quickly. This is severely frowned upon behavior.
3095 if (!(scb
->tcp_flags
& TCPHDR_SYN
)) {
3096 flag
|= FLAG_DATA_ACKED
;
3098 flag
|= FLAG_SYN_ACKED
;
3099 tp
->retrans_stamp
= 0;
3105 tcp_unlink_write_queue(skb
, sk
);
3106 sk_wmem_free_skb(sk
, skb
);
3107 if (skb
== tp
->retransmit_skb_hint
)
3108 tp
->retransmit_skb_hint
= NULL
;
3109 if (skb
== tp
->lost_skb_hint
)
3110 tp
->lost_skb_hint
= NULL
;
3113 if (likely(between(tp
->snd_up
, prior_snd_una
, tp
->snd_una
)))
3114 tp
->snd_up
= tp
->snd_una
;
3116 if (skb
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
3117 flag
|= FLAG_SACK_RENEGING
;
3119 skb_mstamp_get(&now
);
3120 if (first_ackt
.v64
) {
3121 seq_rtt_us
= skb_mstamp_us_delta(&now
, &first_ackt
);
3122 ca_seq_rtt_us
= skb_mstamp_us_delta(&now
, &last_ackt
);
3125 rtt_update
= tcp_ack_update_rtt(sk
, flag
, seq_rtt_us
, sack_rtt_us
);
3127 if (flag
& FLAG_ACKED
) {
3128 const struct tcp_congestion_ops
*ca_ops
3129 = inet_csk(sk
)->icsk_ca_ops
;
3132 if (unlikely(icsk
->icsk_mtup
.probe_size
&&
3133 !after(tp
->mtu_probe
.probe_seq_end
, tp
->snd_una
))) {
3134 tcp_mtup_probe_success(sk
);
3137 if (tcp_is_reno(tp
)) {
3138 tcp_remove_reno_sacks(sk
, pkts_acked
);
3142 /* Non-retransmitted hole got filled? That's reordering */
3143 if (reord
< prior_fackets
)
3144 tcp_update_reordering(sk
, tp
->fackets_out
- reord
, 0);
3146 delta
= tcp_is_fack(tp
) ? pkts_acked
:
3147 prior_sacked
- tp
->sacked_out
;
3148 tp
->lost_cnt_hint
-= min(tp
->lost_cnt_hint
, delta
);
3151 tp
->fackets_out
-= min(pkts_acked
, tp
->fackets_out
);
3153 if (ca_ops
->pkts_acked
)
3154 ca_ops
->pkts_acked(sk
, pkts_acked
, ca_seq_rtt_us
);
3156 } else if (skb
&& rtt_update
&& sack_rtt_us
>= 0 &&
3157 sack_rtt_us
> skb_mstamp_us_delta(&now
, &skb
->skb_mstamp
)) {
3158 /* Do not re-arm RTO if the sack RTT is measured from data sent
3159 * after when the head was last (re)transmitted. Otherwise the
3160 * timeout may continue to extend in loss recovery.
3165 #if FASTRETRANS_DEBUG > 0
3166 WARN_ON((int)tp
->sacked_out
< 0);
3167 WARN_ON((int)tp
->lost_out
< 0);
3168 WARN_ON((int)tp
->retrans_out
< 0);
3169 if (!tp
->packets_out
&& tcp_is_sack(tp
)) {
3170 icsk
= inet_csk(sk
);
3172 pr_debug("Leak l=%u %d\n",
3173 tp
->lost_out
, icsk
->icsk_ca_state
);
3176 if (tp
->sacked_out
) {
3177 pr_debug("Leak s=%u %d\n",
3178 tp
->sacked_out
, icsk
->icsk_ca_state
);
3181 if (tp
->retrans_out
) {
3182 pr_debug("Leak r=%u %d\n",
3183 tp
->retrans_out
, icsk
->icsk_ca_state
);
3184 tp
->retrans_out
= 0;
3191 static void tcp_ack_probe(struct sock
*sk
)
3193 const struct tcp_sock
*tp
= tcp_sk(sk
);
3194 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3196 /* Was it a usable window open? */
3198 if (!after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
, tcp_wnd_end(tp
))) {
3199 icsk
->icsk_backoff
= 0;
3200 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
3201 /* Socket must be waked up by subsequent tcp_data_snd_check().
3202 * This function is not for random using!
3205 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3206 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
3211 static inline bool tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
3213 return !(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
3214 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
;
3217 /* Decide wheather to run the increase function of congestion control. */
3218 static inline bool tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
3220 if (tcp_in_cwnd_reduction(sk
))
3223 /* If reordering is high then always grow cwnd whenever data is
3224 * delivered regardless of its ordering. Otherwise stay conservative
3225 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3226 * new SACK or ECE mark may first advance cwnd here and later reduce
3227 * cwnd in tcp_fastretrans_alert() based on more states.
3229 if (tcp_sk(sk
)->reordering
> sysctl_tcp_reordering
)
3230 return flag
& FLAG_FORWARD_PROGRESS
;
3232 return flag
& FLAG_DATA_ACKED
;
3235 /* Check that window update is acceptable.
3236 * The function assumes that snd_una<=ack<=snd_next.
3238 static inline bool tcp_may_update_window(const struct tcp_sock
*tp
,
3239 const u32 ack
, const u32 ack_seq
,
3242 return after(ack
, tp
->snd_una
) ||
3243 after(ack_seq
, tp
->snd_wl1
) ||
3244 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
);
3247 /* Update our send window.
3249 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3250 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3252 static int tcp_ack_update_window(struct sock
*sk
, const struct sk_buff
*skb
, u32 ack
,
3255 struct tcp_sock
*tp
= tcp_sk(sk
);
3257 u32 nwin
= ntohs(tcp_hdr(skb
)->window
);
3259 if (likely(!tcp_hdr(skb
)->syn
))
3260 nwin
<<= tp
->rx_opt
.snd_wscale
;
3262 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
3263 flag
|= FLAG_WIN_UPDATE
;
3264 tcp_update_wl(tp
, ack_seq
);
3266 if (tp
->snd_wnd
!= nwin
) {
3269 /* Note, it is the only place, where
3270 * fast path is recovered for sending TCP.
3273 tcp_fast_path_check(sk
);
3275 if (nwin
> tp
->max_window
) {
3276 tp
->max_window
= nwin
;
3277 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
3287 /* RFC 5961 7 [ACK Throttling] */
3288 static void tcp_send_challenge_ack(struct sock
*sk
)
3290 /* unprotected vars, we dont care of overwrites */
3291 static u32 challenge_timestamp
;
3292 static unsigned int challenge_count
;
3293 u32 now
= jiffies
/ HZ
;
3295 if (now
!= challenge_timestamp
) {
3296 challenge_timestamp
= now
;
3297 challenge_count
= 0;
3299 if (++challenge_count
<= sysctl_tcp_challenge_ack_limit
) {
3300 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPCHALLENGEACK
);
3305 static void tcp_store_ts_recent(struct tcp_sock
*tp
)
3307 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
3308 tp
->rx_opt
.ts_recent_stamp
= get_seconds();
3311 static void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
3313 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
3314 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3315 * extra check below makes sure this can only happen
3316 * for pure ACK frames. -DaveM
3318 * Not only, also it occurs for expired timestamps.
3321 if (tcp_paws_check(&tp
->rx_opt
, 0))
3322 tcp_store_ts_recent(tp
);
3326 /* This routine deals with acks during a TLP episode.
3327 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3329 static void tcp_process_tlp_ack(struct sock
*sk
, u32 ack
, int flag
)
3331 struct tcp_sock
*tp
= tcp_sk(sk
);
3332 bool is_tlp_dupack
= (ack
== tp
->tlp_high_seq
) &&
3333 !(flag
& (FLAG_SND_UNA_ADVANCED
|
3334 FLAG_NOT_DUP
| FLAG_DATA_SACKED
));
3336 /* Mark the end of TLP episode on receiving TLP dupack or when
3337 * ack is after tlp_high_seq.
3339 if (is_tlp_dupack
) {
3340 tp
->tlp_high_seq
= 0;
3344 if (after(ack
, tp
->tlp_high_seq
)) {
3345 tp
->tlp_high_seq
= 0;
3346 /* Don't reduce cwnd if DSACK arrives for TLP retrans. */
3347 if (!(flag
& FLAG_DSACKING_ACK
)) {
3348 tcp_init_cwnd_reduction(sk
, true);
3349 tcp_set_ca_state(sk
, TCP_CA_CWR
);
3350 tcp_end_cwnd_reduction(sk
);
3351 tcp_try_keep_open(sk
);
3352 NET_INC_STATS_BH(sock_net(sk
),
3353 LINUX_MIB_TCPLOSSPROBERECOVERY
);
3358 /* This routine deals with incoming acks, but not outgoing ones. */
3359 static int tcp_ack(struct sock
*sk
, const struct sk_buff
*skb
, int flag
)
3361 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3362 struct tcp_sock
*tp
= tcp_sk(sk
);
3363 u32 prior_snd_una
= tp
->snd_una
;
3364 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
3365 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3366 bool is_dupack
= false;
3367 u32 prior_in_flight
;
3369 int prior_packets
= tp
->packets_out
;
3370 const int prior_unsacked
= tp
->packets_out
- tp
->sacked_out
;
3371 int acked
= 0; /* Number of packets newly acked */
3372 long sack_rtt_us
= -1L;
3374 /* If the ack is older than previous acks
3375 * then we can probably ignore it.
3377 if (before(ack
, prior_snd_una
)) {
3378 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3379 if (before(ack
, prior_snd_una
- tp
->max_window
)) {
3380 tcp_send_challenge_ack(sk
);
3386 /* If the ack includes data we haven't sent yet, discard
3387 * this segment (RFC793 Section 3.9).
3389 if (after(ack
, tp
->snd_nxt
))
3392 if (icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
||
3393 icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
)
3396 if (after(ack
, prior_snd_una
))
3397 flag
|= FLAG_SND_UNA_ADVANCED
;
3399 prior_fackets
= tp
->fackets_out
;
3400 prior_in_flight
= tcp_packets_in_flight(tp
);
3402 /* ts_recent update must be made after we are sure that the packet
3405 if (flag
& FLAG_UPDATE_TS_RECENT
)
3406 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
3408 if (!(flag
& FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
3409 /* Window is constant, pure forward advance.
3410 * No more checks are required.
3411 * Note, we use the fact that SND.UNA>=SND.WL2.
3413 tcp_update_wl(tp
, ack_seq
);
3415 flag
|= FLAG_WIN_UPDATE
;
3417 tcp_ca_event(sk
, CA_EVENT_FAST_ACK
);
3419 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPACKS
);
3421 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
3424 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPUREACKS
);
3426 flag
|= tcp_ack_update_window(sk
, skb
, ack
, ack_seq
);
3428 if (TCP_SKB_CB(skb
)->sacked
)
3429 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
,
3432 if (TCP_ECN_rcv_ecn_echo(tp
, tcp_hdr(skb
)))
3435 tcp_ca_event(sk
, CA_EVENT_SLOW_ACK
);
3438 /* We passed data and got it acked, remove any soft error
3439 * log. Something worked...
3441 sk
->sk_err_soft
= 0;
3442 icsk
->icsk_probes_out
= 0;
3443 tp
->rcv_tstamp
= tcp_time_stamp
;
3447 /* See if we can take anything off of the retransmit queue. */
3448 acked
= tp
->packets_out
;
3449 flag
|= tcp_clean_rtx_queue(sk
, prior_fackets
, prior_snd_una
,
3451 acked
-= tp
->packets_out
;
3453 /* Advance cwnd if state allows */
3454 if (tcp_may_raise_cwnd(sk
, flag
))
3455 tcp_cong_avoid(sk
, ack
, acked
, prior_in_flight
);
3457 if (tcp_ack_is_dubious(sk
, flag
)) {
3458 is_dupack
= !(flag
& (FLAG_SND_UNA_ADVANCED
| FLAG_NOT_DUP
));
3459 tcp_fastretrans_alert(sk
, acked
, prior_unsacked
,
3462 if (tp
->tlp_high_seq
)
3463 tcp_process_tlp_ack(sk
, ack
, flag
);
3465 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
& FLAG_NOT_DUP
)) {
3466 struct dst_entry
*dst
= __sk_dst_get(sk
);
3471 if (icsk
->icsk_pending
== ICSK_TIME_RETRANS
)
3472 tcp_schedule_loss_probe(sk
);
3473 tcp_update_pacing_rate(sk
);
3477 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3478 if (flag
& FLAG_DSACKING_ACK
)
3479 tcp_fastretrans_alert(sk
, acked
, prior_unsacked
,
3481 /* If this ack opens up a zero window, clear backoff. It was
3482 * being used to time the probes, and is probably far higher than
3483 * it needs to be for normal retransmission.
3485 if (tcp_send_head(sk
))
3488 if (tp
->tlp_high_seq
)
3489 tcp_process_tlp_ack(sk
, ack
, flag
);
3493 SOCK_DEBUG(sk
, "Ack %u after %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3497 /* If data was SACKed, tag it and see if we should send more data.
3498 * If data was DSACKed, see if we can undo a cwnd reduction.
3500 if (TCP_SKB_CB(skb
)->sacked
) {
3501 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
,
3503 tcp_fastretrans_alert(sk
, acked
, prior_unsacked
,
3507 SOCK_DEBUG(sk
, "Ack %u before %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3511 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3512 * But, this can also be called on packets in the established flow when
3513 * the fast version below fails.
3515 void tcp_parse_options(const struct sk_buff
*skb
,
3516 struct tcp_options_received
*opt_rx
, int estab
,
3517 struct tcp_fastopen_cookie
*foc
)
3519 const unsigned char *ptr
;
3520 const struct tcphdr
*th
= tcp_hdr(skb
);
3521 int length
= (th
->doff
* 4) - sizeof(struct tcphdr
);
3523 ptr
= (const unsigned char *)(th
+ 1);
3524 opt_rx
->saw_tstamp
= 0;
3526 while (length
> 0) {
3527 int opcode
= *ptr
++;
3533 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
3538 if (opsize
< 2) /* "silly options" */
3540 if (opsize
> length
)
3541 return; /* don't parse partial options */
3544 if (opsize
== TCPOLEN_MSS
&& th
->syn
&& !estab
) {
3545 u16 in_mss
= get_unaligned_be16(ptr
);
3547 if (opt_rx
->user_mss
&&
3548 opt_rx
->user_mss
< in_mss
)
3549 in_mss
= opt_rx
->user_mss
;
3550 opt_rx
->mss_clamp
= in_mss
;
3555 if (opsize
== TCPOLEN_WINDOW
&& th
->syn
&&
3556 !estab
&& sysctl_tcp_window_scaling
) {
3557 __u8 snd_wscale
= *(__u8
*)ptr
;
3558 opt_rx
->wscale_ok
= 1;
3559 if (snd_wscale
> 14) {
3560 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3565 opt_rx
->snd_wscale
= snd_wscale
;
3568 case TCPOPT_TIMESTAMP
:
3569 if ((opsize
== TCPOLEN_TIMESTAMP
) &&
3570 ((estab
&& opt_rx
->tstamp_ok
) ||
3571 (!estab
&& sysctl_tcp_timestamps
))) {
3572 opt_rx
->saw_tstamp
= 1;
3573 opt_rx
->rcv_tsval
= get_unaligned_be32(ptr
);
3574 opt_rx
->rcv_tsecr
= get_unaligned_be32(ptr
+ 4);
3577 case TCPOPT_SACK_PERM
:
3578 if (opsize
== TCPOLEN_SACK_PERM
&& th
->syn
&&
3579 !estab
&& sysctl_tcp_sack
) {
3580 opt_rx
->sack_ok
= TCP_SACK_SEEN
;
3581 tcp_sack_reset(opt_rx
);
3586 if ((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
3587 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
3589 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
3592 #ifdef CONFIG_TCP_MD5SIG
3595 * The MD5 Hash has already been
3596 * checked (see tcp_v{4,6}_do_rcv()).
3601 /* Fast Open option shares code 254 using a
3602 * 16 bits magic number. It's valid only in
3603 * SYN or SYN-ACK with an even size.
3605 if (opsize
< TCPOLEN_EXP_FASTOPEN_BASE
||
3606 get_unaligned_be16(ptr
) != TCPOPT_FASTOPEN_MAGIC
||
3607 foc
== NULL
|| !th
->syn
|| (opsize
& 1))
3609 foc
->len
= opsize
- TCPOLEN_EXP_FASTOPEN_BASE
;
3610 if (foc
->len
>= TCP_FASTOPEN_COOKIE_MIN
&&
3611 foc
->len
<= TCP_FASTOPEN_COOKIE_MAX
)
3612 memcpy(foc
->val
, ptr
+ 2, foc
->len
);
3613 else if (foc
->len
!= 0)
3623 EXPORT_SYMBOL(tcp_parse_options
);
3625 static bool tcp_parse_aligned_timestamp(struct tcp_sock
*tp
, const struct tcphdr
*th
)
3627 const __be32
*ptr
= (const __be32
*)(th
+ 1);
3629 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3630 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
3631 tp
->rx_opt
.saw_tstamp
= 1;
3633 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3636 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
) - tp
->tsoffset
;
3638 tp
->rx_opt
.rcv_tsecr
= 0;
3644 /* Fast parse options. This hopes to only see timestamps.
3645 * If it is wrong it falls back on tcp_parse_options().
3647 static bool tcp_fast_parse_options(const struct sk_buff
*skb
,
3648 const struct tcphdr
*th
, struct tcp_sock
*tp
)
3650 /* In the spirit of fast parsing, compare doff directly to constant
3651 * values. Because equality is used, short doff can be ignored here.
3653 if (th
->doff
== (sizeof(*th
) / 4)) {
3654 tp
->rx_opt
.saw_tstamp
= 0;
3656 } else if (tp
->rx_opt
.tstamp_ok
&&
3657 th
->doff
== ((sizeof(*th
) + TCPOLEN_TSTAMP_ALIGNED
) / 4)) {
3658 if (tcp_parse_aligned_timestamp(tp
, th
))
3662 tcp_parse_options(skb
, &tp
->rx_opt
, 1, NULL
);
3663 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
3664 tp
->rx_opt
.rcv_tsecr
-= tp
->tsoffset
;
3669 #ifdef CONFIG_TCP_MD5SIG
3671 * Parse MD5 Signature option
3673 const u8
*tcp_parse_md5sig_option(const struct tcphdr
*th
)
3675 int length
= (th
->doff
<< 2) - sizeof(*th
);
3676 const u8
*ptr
= (const u8
*)(th
+ 1);
3678 /* If the TCP option is too short, we can short cut */
3679 if (length
< TCPOLEN_MD5SIG
)
3682 while (length
> 0) {
3683 int opcode
= *ptr
++;
3694 if (opsize
< 2 || opsize
> length
)
3696 if (opcode
== TCPOPT_MD5SIG
)
3697 return opsize
== TCPOLEN_MD5SIG
? ptr
: NULL
;
3704 EXPORT_SYMBOL(tcp_parse_md5sig_option
);
3707 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3709 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3710 * it can pass through stack. So, the following predicate verifies that
3711 * this segment is not used for anything but congestion avoidance or
3712 * fast retransmit. Moreover, we even are able to eliminate most of such
3713 * second order effects, if we apply some small "replay" window (~RTO)
3714 * to timestamp space.
3716 * All these measures still do not guarantee that we reject wrapped ACKs
3717 * on networks with high bandwidth, when sequence space is recycled fastly,
3718 * but it guarantees that such events will be very rare and do not affect
3719 * connection seriously. This doesn't look nice, but alas, PAWS is really
3722 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3723 * states that events when retransmit arrives after original data are rare.
3724 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3725 * the biggest problem on large power networks even with minor reordering.
3726 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3727 * up to bandwidth of 18Gigabit/sec. 8) ]
3730 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
3732 const struct tcp_sock
*tp
= tcp_sk(sk
);
3733 const struct tcphdr
*th
= tcp_hdr(skb
);
3734 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3735 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3737 return (/* 1. Pure ACK with correct sequence number. */
3738 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
3740 /* 2. ... and duplicate ACK. */
3741 ack
== tp
->snd_una
&&
3743 /* 3. ... and does not update window. */
3744 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
3746 /* 4. ... and sits in replay window. */
3747 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
3750 static inline bool tcp_paws_discard(const struct sock
*sk
,
3751 const struct sk_buff
*skb
)
3753 const struct tcp_sock
*tp
= tcp_sk(sk
);
3755 return !tcp_paws_check(&tp
->rx_opt
, TCP_PAWS_WINDOW
) &&
3756 !tcp_disordered_ack(sk
, skb
);
3759 /* Check segment sequence number for validity.
3761 * Segment controls are considered valid, if the segment
3762 * fits to the window after truncation to the window. Acceptability
3763 * of data (and SYN, FIN, of course) is checked separately.
3764 * See tcp_data_queue(), for example.
3766 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3767 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3768 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3769 * (borrowed from freebsd)
3772 static inline bool tcp_sequence(const struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3774 return !before(end_seq
, tp
->rcv_wup
) &&
3775 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
3778 /* When we get a reset we do this. */
3779 void tcp_reset(struct sock
*sk
)
3781 /* We want the right error as BSD sees it (and indeed as we do). */
3782 switch (sk
->sk_state
) {
3784 sk
->sk_err
= ECONNREFUSED
;
3786 case TCP_CLOSE_WAIT
:
3792 sk
->sk_err
= ECONNRESET
;
3794 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3797 if (!sock_flag(sk
, SOCK_DEAD
))
3798 sk
->sk_error_report(sk
);
3804 * Process the FIN bit. This now behaves as it is supposed to work
3805 * and the FIN takes effect when it is validly part of sequence
3806 * space. Not before when we get holes.
3808 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3809 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3812 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3813 * close and we go into CLOSING (and later onto TIME-WAIT)
3815 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3817 static void tcp_fin(struct sock
*sk
)
3819 struct tcp_sock
*tp
= tcp_sk(sk
);
3820 const struct dst_entry
*dst
;
3822 inet_csk_schedule_ack(sk
);
3824 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
3825 sock_set_flag(sk
, SOCK_DONE
);
3827 switch (sk
->sk_state
) {
3829 case TCP_ESTABLISHED
:
3830 /* Move to CLOSE_WAIT */
3831 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
3832 dst
= __sk_dst_get(sk
);
3833 if (!dst
|| !dst_metric(dst
, RTAX_QUICKACK
))
3834 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
3837 case TCP_CLOSE_WAIT
:
3839 /* Received a retransmission of the FIN, do
3844 /* RFC793: Remain in the LAST-ACK state. */
3848 /* This case occurs when a simultaneous close
3849 * happens, we must ack the received FIN and
3850 * enter the CLOSING state.
3853 tcp_set_state(sk
, TCP_CLOSING
);
3856 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3858 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
3861 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3862 * cases we should never reach this piece of code.
3864 pr_err("%s: Impossible, sk->sk_state=%d\n",
3865 __func__
, sk
->sk_state
);
3869 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3870 * Probably, we should reset in this case. For now drop them.
3872 __skb_queue_purge(&tp
->out_of_order_queue
);
3873 if (tcp_is_sack(tp
))
3874 tcp_sack_reset(&tp
->rx_opt
);
3877 if (!sock_flag(sk
, SOCK_DEAD
)) {
3878 sk
->sk_state_change(sk
);
3880 /* Do not send POLL_HUP for half duplex close. */
3881 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
3882 sk
->sk_state
== TCP_CLOSE
)
3883 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_HUP
);
3885 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_IN
);
3889 static inline bool tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
,
3892 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
3893 if (before(seq
, sp
->start_seq
))
3894 sp
->start_seq
= seq
;
3895 if (after(end_seq
, sp
->end_seq
))
3896 sp
->end_seq
= end_seq
;
3902 static void tcp_dsack_set(struct sock
*sk
, u32 seq
, u32 end_seq
)
3904 struct tcp_sock
*tp
= tcp_sk(sk
);
3906 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
3909 if (before(seq
, tp
->rcv_nxt
))
3910 mib_idx
= LINUX_MIB_TCPDSACKOLDSENT
;
3912 mib_idx
= LINUX_MIB_TCPDSACKOFOSENT
;
3914 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
3916 tp
->rx_opt
.dsack
= 1;
3917 tp
->duplicate_sack
[0].start_seq
= seq
;
3918 tp
->duplicate_sack
[0].end_seq
= end_seq
;
3922 static void tcp_dsack_extend(struct sock
*sk
, u32 seq
, u32 end_seq
)
3924 struct tcp_sock
*tp
= tcp_sk(sk
);
3926 if (!tp
->rx_opt
.dsack
)
3927 tcp_dsack_set(sk
, seq
, end_seq
);
3929 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
3932 static void tcp_send_dupack(struct sock
*sk
, const struct sk_buff
*skb
)
3934 struct tcp_sock
*tp
= tcp_sk(sk
);
3936 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
3937 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
3938 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
3939 tcp_enter_quickack_mode(sk
);
3941 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
3942 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
3944 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
3945 end_seq
= tp
->rcv_nxt
;
3946 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, end_seq
);
3953 /* These routines update the SACK block as out-of-order packets arrive or
3954 * in-order packets close up the sequence space.
3956 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
3959 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3960 struct tcp_sack_block
*swalk
= sp
+ 1;
3962 /* See if the recent change to the first SACK eats into
3963 * or hits the sequence space of other SACK blocks, if so coalesce.
3965 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
;) {
3966 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
3969 /* Zap SWALK, by moving every further SACK up by one slot.
3970 * Decrease num_sacks.
3972 tp
->rx_opt
.num_sacks
--;
3973 for (i
= this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
3977 this_sack
++, swalk
++;
3981 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
3983 struct tcp_sock
*tp
= tcp_sk(sk
);
3984 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3985 int cur_sacks
= tp
->rx_opt
.num_sacks
;
3991 for (this_sack
= 0; this_sack
< cur_sacks
; this_sack
++, sp
++) {
3992 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
3993 /* Rotate this_sack to the first one. */
3994 for (; this_sack
> 0; this_sack
--, sp
--)
3995 swap(*sp
, *(sp
- 1));
3997 tcp_sack_maybe_coalesce(tp
);
4002 /* Could not find an adjacent existing SACK, build a new one,
4003 * put it at the front, and shift everyone else down. We
4004 * always know there is at least one SACK present already here.
4006 * If the sack array is full, forget about the last one.
4008 if (this_sack
>= TCP_NUM_SACKS
) {
4010 tp
->rx_opt
.num_sacks
--;
4013 for (; this_sack
> 0; this_sack
--, sp
--)
4017 /* Build the new head SACK, and we're done. */
4018 sp
->start_seq
= seq
;
4019 sp
->end_seq
= end_seq
;
4020 tp
->rx_opt
.num_sacks
++;
4023 /* RCV.NXT advances, some SACKs should be eaten. */
4025 static void tcp_sack_remove(struct tcp_sock
*tp
)
4027 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4028 int num_sacks
= tp
->rx_opt
.num_sacks
;
4031 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4032 if (skb_queue_empty(&tp
->out_of_order_queue
)) {
4033 tp
->rx_opt
.num_sacks
= 0;
4037 for (this_sack
= 0; this_sack
< num_sacks
;) {
4038 /* Check if the start of the sack is covered by RCV.NXT. */
4039 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
4042 /* RCV.NXT must cover all the block! */
4043 WARN_ON(before(tp
->rcv_nxt
, sp
->end_seq
));
4045 /* Zap this SACK, by moving forward any other SACKS. */
4046 for (i
= this_sack
+1; i
< num_sacks
; i
++)
4047 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
4054 tp
->rx_opt
.num_sacks
= num_sacks
;
4057 /* This one checks to see if we can put data from the
4058 * out_of_order queue into the receive_queue.
4060 static void tcp_ofo_queue(struct sock
*sk
)
4062 struct tcp_sock
*tp
= tcp_sk(sk
);
4063 __u32 dsack_high
= tp
->rcv_nxt
;
4064 struct sk_buff
*skb
;
4066 while ((skb
= skb_peek(&tp
->out_of_order_queue
)) != NULL
) {
4067 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
4070 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
4071 __u32 dsack
= dsack_high
;
4072 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
4073 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
4074 tcp_dsack_extend(sk
, TCP_SKB_CB(skb
)->seq
, dsack
);
4077 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4078 SOCK_DEBUG(sk
, "ofo packet was already received\n");
4079 __skb_unlink(skb
, &tp
->out_of_order_queue
);
4083 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
4084 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4085 TCP_SKB_CB(skb
)->end_seq
);
4087 __skb_unlink(skb
, &tp
->out_of_order_queue
);
4088 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4089 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4090 if (tcp_hdr(skb
)->fin
)
4095 static bool tcp_prune_ofo_queue(struct sock
*sk
);
4096 static int tcp_prune_queue(struct sock
*sk
);
4098 static int tcp_try_rmem_schedule(struct sock
*sk
, struct sk_buff
*skb
,
4101 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
4102 !sk_rmem_schedule(sk
, skb
, size
)) {
4104 if (tcp_prune_queue(sk
) < 0)
4107 if (!sk_rmem_schedule(sk
, skb
, size
)) {
4108 if (!tcp_prune_ofo_queue(sk
))
4111 if (!sk_rmem_schedule(sk
, skb
, size
))
4119 * tcp_try_coalesce - try to merge skb to prior one
4122 * @from: buffer to add in queue
4123 * @fragstolen: pointer to boolean
4125 * Before queueing skb @from after @to, try to merge them
4126 * to reduce overall memory use and queue lengths, if cost is small.
4127 * Packets in ofo or receive queues can stay a long time.
4128 * Better try to coalesce them right now to avoid future collapses.
4129 * Returns true if caller should free @from instead of queueing it
4131 static bool tcp_try_coalesce(struct sock
*sk
,
4133 struct sk_buff
*from
,
4138 *fragstolen
= false;
4140 if (tcp_hdr(from
)->fin
)
4143 /* Its possible this segment overlaps with prior segment in queue */
4144 if (TCP_SKB_CB(from
)->seq
!= TCP_SKB_CB(to
)->end_seq
)
4147 if (!skb_try_coalesce(to
, from
, fragstolen
, &delta
))
4150 atomic_add(delta
, &sk
->sk_rmem_alloc
);
4151 sk_mem_charge(sk
, delta
);
4152 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRCVCOALESCE
);
4153 TCP_SKB_CB(to
)->end_seq
= TCP_SKB_CB(from
)->end_seq
;
4154 TCP_SKB_CB(to
)->ack_seq
= TCP_SKB_CB(from
)->ack_seq
;
4158 static void tcp_data_queue_ofo(struct sock
*sk
, struct sk_buff
*skb
)
4160 struct tcp_sock
*tp
= tcp_sk(sk
);
4161 struct sk_buff
*skb1
;
4164 TCP_ECN_check_ce(tp
, skb
);
4166 if (unlikely(tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))) {
4167 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFODROP
);
4172 /* Disable header prediction. */
4174 inet_csk_schedule_ack(sk
);
4176 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFOQUEUE
);
4177 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
4178 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4180 skb1
= skb_peek_tail(&tp
->out_of_order_queue
);
4182 /* Initial out of order segment, build 1 SACK. */
4183 if (tcp_is_sack(tp
)) {
4184 tp
->rx_opt
.num_sacks
= 1;
4185 tp
->selective_acks
[0].start_seq
= TCP_SKB_CB(skb
)->seq
;
4186 tp
->selective_acks
[0].end_seq
=
4187 TCP_SKB_CB(skb
)->end_seq
;
4189 __skb_queue_head(&tp
->out_of_order_queue
, skb
);
4193 seq
= TCP_SKB_CB(skb
)->seq
;
4194 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4196 if (seq
== TCP_SKB_CB(skb1
)->end_seq
) {
4199 if (!tcp_try_coalesce(sk
, skb1
, skb
, &fragstolen
)) {
4200 __skb_queue_after(&tp
->out_of_order_queue
, skb1
, skb
);
4202 tcp_grow_window(sk
, skb
);
4203 kfree_skb_partial(skb
, fragstolen
);
4207 if (!tp
->rx_opt
.num_sacks
||
4208 tp
->selective_acks
[0].end_seq
!= seq
)
4211 /* Common case: data arrive in order after hole. */
4212 tp
->selective_acks
[0].end_seq
= end_seq
;
4216 /* Find place to insert this segment. */
4218 if (!after(TCP_SKB_CB(skb1
)->seq
, seq
))
4220 if (skb_queue_is_first(&tp
->out_of_order_queue
, skb1
)) {
4224 skb1
= skb_queue_prev(&tp
->out_of_order_queue
, skb1
);
4227 /* Do skb overlap to previous one? */
4228 if (skb1
&& before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4229 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4230 /* All the bits are present. Drop. */
4231 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFOMERGE
);
4234 tcp_dsack_set(sk
, seq
, end_seq
);
4237 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
4238 /* Partial overlap. */
4239 tcp_dsack_set(sk
, seq
,
4240 TCP_SKB_CB(skb1
)->end_seq
);
4242 if (skb_queue_is_first(&tp
->out_of_order_queue
,
4246 skb1
= skb_queue_prev(
4247 &tp
->out_of_order_queue
,
4252 __skb_queue_head(&tp
->out_of_order_queue
, skb
);
4254 __skb_queue_after(&tp
->out_of_order_queue
, skb1
, skb
);
4256 /* And clean segments covered by new one as whole. */
4257 while (!skb_queue_is_last(&tp
->out_of_order_queue
, skb
)) {
4258 skb1
= skb_queue_next(&tp
->out_of_order_queue
, skb
);
4260 if (!after(end_seq
, TCP_SKB_CB(skb1
)->seq
))
4262 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4263 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4267 __skb_unlink(skb1
, &tp
->out_of_order_queue
);
4268 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4269 TCP_SKB_CB(skb1
)->end_seq
);
4270 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFOMERGE
);
4275 if (tcp_is_sack(tp
))
4276 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
4279 tcp_grow_window(sk
, skb
);
4280 skb_set_owner_r(skb
, sk
);
4284 static int __must_check
tcp_queue_rcv(struct sock
*sk
, struct sk_buff
*skb
, int hdrlen
,
4288 struct sk_buff
*tail
= skb_peek_tail(&sk
->sk_receive_queue
);
4290 __skb_pull(skb
, hdrlen
);
4292 tcp_try_coalesce(sk
, tail
, skb
, fragstolen
)) ? 1 : 0;
4293 tcp_sk(sk
)->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4295 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4296 skb_set_owner_r(skb
, sk
);
4301 int tcp_send_rcvq(struct sock
*sk
, struct msghdr
*msg
, size_t size
)
4303 struct sk_buff
*skb
= NULL
;
4310 skb
= alloc_skb(size
+ sizeof(*th
), sk
->sk_allocation
);
4314 if (tcp_try_rmem_schedule(sk
, skb
, size
+ sizeof(*th
)))
4317 th
= (struct tcphdr
*)skb_put(skb
, sizeof(*th
));
4318 skb_reset_transport_header(skb
);
4319 memset(th
, 0, sizeof(*th
));
4321 if (memcpy_fromiovec(skb_put(skb
, size
), msg
->msg_iov
, size
))
4324 TCP_SKB_CB(skb
)->seq
= tcp_sk(sk
)->rcv_nxt
;
4325 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ size
;
4326 TCP_SKB_CB(skb
)->ack_seq
= tcp_sk(sk
)->snd_una
- 1;
4328 if (tcp_queue_rcv(sk
, skb
, sizeof(*th
), &fragstolen
)) {
4329 WARN_ON_ONCE(fragstolen
); /* should not happen */
4340 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
4342 const struct tcphdr
*th
= tcp_hdr(skb
);
4343 struct tcp_sock
*tp
= tcp_sk(sk
);
4345 bool fragstolen
= false;
4347 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
)
4351 __skb_pull(skb
, th
->doff
* 4);
4353 TCP_ECN_accept_cwr(tp
, skb
);
4355 tp
->rx_opt
.dsack
= 0;
4357 /* Queue data for delivery to the user.
4358 * Packets in sequence go to the receive queue.
4359 * Out of sequence packets to the out_of_order_queue.
4361 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
4362 if (tcp_receive_window(tp
) == 0)
4365 /* Ok. In sequence. In window. */
4366 if (tp
->ucopy
.task
== current
&&
4367 tp
->copied_seq
== tp
->rcv_nxt
&& tp
->ucopy
.len
&&
4368 sock_owned_by_user(sk
) && !tp
->urg_data
) {
4369 int chunk
= min_t(unsigned int, skb
->len
,
4372 __set_current_state(TASK_RUNNING
);
4375 if (!skb_copy_datagram_iovec(skb
, 0, tp
->ucopy
.iov
, chunk
)) {
4376 tp
->ucopy
.len
-= chunk
;
4377 tp
->copied_seq
+= chunk
;
4378 eaten
= (chunk
== skb
->len
);
4379 tcp_rcv_space_adjust(sk
);
4387 tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))
4390 eaten
= tcp_queue_rcv(sk
, skb
, 0, &fragstolen
);
4392 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4394 tcp_event_data_recv(sk
, skb
);
4398 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4401 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4402 * gap in queue is filled.
4404 if (skb_queue_empty(&tp
->out_of_order_queue
))
4405 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
4408 if (tp
->rx_opt
.num_sacks
)
4409 tcp_sack_remove(tp
);
4411 tcp_fast_path_check(sk
);
4414 kfree_skb_partial(skb
, fragstolen
);
4415 if (!sock_flag(sk
, SOCK_DEAD
))
4416 sk
->sk_data_ready(sk
);
4420 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4421 /* A retransmit, 2nd most common case. Force an immediate ack. */
4422 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4423 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4426 tcp_enter_quickack_mode(sk
);
4427 inet_csk_schedule_ack(sk
);
4433 /* Out of window. F.e. zero window probe. */
4434 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
4437 tcp_enter_quickack_mode(sk
);
4439 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4440 /* Partial packet, seq < rcv_next < end_seq */
4441 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
4442 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4443 TCP_SKB_CB(skb
)->end_seq
);
4445 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
4447 /* If window is closed, drop tail of packet. But after
4448 * remembering D-SACK for its head made in previous line.
4450 if (!tcp_receive_window(tp
))
4455 tcp_data_queue_ofo(sk
, skb
);
4458 static struct sk_buff
*tcp_collapse_one(struct sock
*sk
, struct sk_buff
*skb
,
4459 struct sk_buff_head
*list
)
4461 struct sk_buff
*next
= NULL
;
4463 if (!skb_queue_is_last(list
, skb
))
4464 next
= skb_queue_next(list
, skb
);
4466 __skb_unlink(skb
, list
);
4468 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRCVCOLLAPSED
);
4473 /* Collapse contiguous sequence of skbs head..tail with
4474 * sequence numbers start..end.
4476 * If tail is NULL, this means until the end of the list.
4478 * Segments with FIN/SYN are not collapsed (only because this
4482 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
,
4483 struct sk_buff
*head
, struct sk_buff
*tail
,
4486 struct sk_buff
*skb
, *n
;
4489 /* First, check that queue is collapsible and find
4490 * the point where collapsing can be useful. */
4494 skb_queue_walk_from_safe(list
, skb
, n
) {
4497 /* No new bits? It is possible on ofo queue. */
4498 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4499 skb
= tcp_collapse_one(sk
, skb
, list
);
4505 /* The first skb to collapse is:
4507 * - bloated or contains data before "start" or
4508 * overlaps to the next one.
4510 if (!tcp_hdr(skb
)->syn
&& !tcp_hdr(skb
)->fin
&&
4511 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
4512 before(TCP_SKB_CB(skb
)->seq
, start
))) {
4513 end_of_skbs
= false;
4517 if (!skb_queue_is_last(list
, skb
)) {
4518 struct sk_buff
*next
= skb_queue_next(list
, skb
);
4520 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(next
)->seq
) {
4521 end_of_skbs
= false;
4526 /* Decided to skip this, advance start seq. */
4527 start
= TCP_SKB_CB(skb
)->end_seq
;
4529 if (end_of_skbs
|| tcp_hdr(skb
)->syn
|| tcp_hdr(skb
)->fin
)
4532 while (before(start
, end
)) {
4533 struct sk_buff
*nskb
;
4534 unsigned int header
= skb_headroom(skb
);
4535 int copy
= SKB_MAX_ORDER(header
, 0);
4537 /* Too big header? This can happen with IPv6. */
4540 if (end
- start
< copy
)
4542 nskb
= alloc_skb(copy
+ header
, GFP_ATOMIC
);
4546 skb_set_mac_header(nskb
, skb_mac_header(skb
) - skb
->head
);
4547 skb_set_network_header(nskb
, (skb_network_header(skb
) -
4549 skb_set_transport_header(nskb
, (skb_transport_header(skb
) -
4551 skb_reserve(nskb
, header
);
4552 memcpy(nskb
->head
, skb
->head
, header
);
4553 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
4554 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
4555 __skb_queue_before(list
, skb
, nskb
);
4556 skb_set_owner_r(nskb
, sk
);
4558 /* Copy data, releasing collapsed skbs. */
4560 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
4561 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
4565 size
= min(copy
, size
);
4566 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
4568 TCP_SKB_CB(nskb
)->end_seq
+= size
;
4572 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4573 skb
= tcp_collapse_one(sk
, skb
, list
);
4576 tcp_hdr(skb
)->syn
||
4584 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4585 * and tcp_collapse() them until all the queue is collapsed.
4587 static void tcp_collapse_ofo_queue(struct sock
*sk
)
4589 struct tcp_sock
*tp
= tcp_sk(sk
);
4590 struct sk_buff
*skb
= skb_peek(&tp
->out_of_order_queue
);
4591 struct sk_buff
*head
;
4597 start
= TCP_SKB_CB(skb
)->seq
;
4598 end
= TCP_SKB_CB(skb
)->end_seq
;
4602 struct sk_buff
*next
= NULL
;
4604 if (!skb_queue_is_last(&tp
->out_of_order_queue
, skb
))
4605 next
= skb_queue_next(&tp
->out_of_order_queue
, skb
);
4608 /* Segment is terminated when we see gap or when
4609 * we are at the end of all the queue. */
4611 after(TCP_SKB_CB(skb
)->seq
, end
) ||
4612 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
4613 tcp_collapse(sk
, &tp
->out_of_order_queue
,
4614 head
, skb
, start
, end
);
4618 /* Start new segment */
4619 start
= TCP_SKB_CB(skb
)->seq
;
4620 end
= TCP_SKB_CB(skb
)->end_seq
;
4622 if (before(TCP_SKB_CB(skb
)->seq
, start
))
4623 start
= TCP_SKB_CB(skb
)->seq
;
4624 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
4625 end
= TCP_SKB_CB(skb
)->end_seq
;
4631 * Purge the out-of-order queue.
4632 * Return true if queue was pruned.
4634 static bool tcp_prune_ofo_queue(struct sock
*sk
)
4636 struct tcp_sock
*tp
= tcp_sk(sk
);
4639 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4640 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_OFOPRUNED
);
4641 __skb_queue_purge(&tp
->out_of_order_queue
);
4643 /* Reset SACK state. A conforming SACK implementation will
4644 * do the same at a timeout based retransmit. When a connection
4645 * is in a sad state like this, we care only about integrity
4646 * of the connection not performance.
4648 if (tp
->rx_opt
.sack_ok
)
4649 tcp_sack_reset(&tp
->rx_opt
);
4656 /* Reduce allocated memory if we can, trying to get
4657 * the socket within its memory limits again.
4659 * Return less than zero if we should start dropping frames
4660 * until the socket owning process reads some of the data
4661 * to stabilize the situation.
4663 static int tcp_prune_queue(struct sock
*sk
)
4665 struct tcp_sock
*tp
= tcp_sk(sk
);
4667 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
4669 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PRUNECALLED
);
4671 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
4672 tcp_clamp_window(sk
);
4673 else if (sk_under_memory_pressure(sk
))
4674 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
4676 tcp_collapse_ofo_queue(sk
);
4677 if (!skb_queue_empty(&sk
->sk_receive_queue
))
4678 tcp_collapse(sk
, &sk
->sk_receive_queue
,
4679 skb_peek(&sk
->sk_receive_queue
),
4681 tp
->copied_seq
, tp
->rcv_nxt
);
4684 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4687 /* Collapsing did not help, destructive actions follow.
4688 * This must not ever occur. */
4690 tcp_prune_ofo_queue(sk
);
4692 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4695 /* If we are really being abused, tell the caller to silently
4696 * drop receive data on the floor. It will get retransmitted
4697 * and hopefully then we'll have sufficient space.
4699 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_RCVPRUNED
);
4701 /* Massive buffer overcommit. */
4706 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4707 * As additional protections, we do not touch cwnd in retransmission phases,
4708 * and if application hit its sndbuf limit recently.
4710 void tcp_cwnd_application_limited(struct sock
*sk
)
4712 struct tcp_sock
*tp
= tcp_sk(sk
);
4714 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
4715 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
4716 /* Limited by application or receiver window. */
4717 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
4718 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
4719 if (win_used
< tp
->snd_cwnd
) {
4720 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
4721 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
4723 tp
->snd_cwnd_used
= 0;
4725 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4728 static bool tcp_should_expand_sndbuf(const struct sock
*sk
)
4730 const struct tcp_sock
*tp
= tcp_sk(sk
);
4732 /* If the user specified a specific send buffer setting, do
4735 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
4738 /* If we are under global TCP memory pressure, do not expand. */
4739 if (sk_under_memory_pressure(sk
))
4742 /* If we are under soft global TCP memory pressure, do not expand. */
4743 if (sk_memory_allocated(sk
) >= sk_prot_mem_limits(sk
, 0))
4746 /* If we filled the congestion window, do not expand. */
4747 if (tp
->packets_out
>= tp
->snd_cwnd
)
4753 /* When incoming ACK allowed to free some skb from write_queue,
4754 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4755 * on the exit from tcp input handler.
4757 * PROBLEM: sndbuf expansion does not work well with largesend.
4759 static void tcp_new_space(struct sock
*sk
)
4761 struct tcp_sock
*tp
= tcp_sk(sk
);
4763 if (tcp_should_expand_sndbuf(sk
)) {
4764 tcp_sndbuf_expand(sk
);
4765 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4768 sk
->sk_write_space(sk
);
4771 static void tcp_check_space(struct sock
*sk
)
4773 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
4774 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
4775 if (sk
->sk_socket
&&
4776 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
4781 static inline void tcp_data_snd_check(struct sock
*sk
)
4783 tcp_push_pending_frames(sk
);
4784 tcp_check_space(sk
);
4788 * Check if sending an ack is needed.
4790 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
4792 struct tcp_sock
*tp
= tcp_sk(sk
);
4794 /* More than one full frame received... */
4795 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
&&
4796 /* ... and right edge of window advances far enough.
4797 * (tcp_recvmsg() will send ACK otherwise). Or...
4799 __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
4800 /* We ACK each frame or... */
4801 tcp_in_quickack_mode(sk
) ||
4802 /* We have out of order data. */
4803 (ofo_possible
&& skb_peek(&tp
->out_of_order_queue
))) {
4804 /* Then ack it now */
4807 /* Else, send delayed ack. */
4808 tcp_send_delayed_ack(sk
);
4812 static inline void tcp_ack_snd_check(struct sock
*sk
)
4814 if (!inet_csk_ack_scheduled(sk
)) {
4815 /* We sent a data segment already. */
4818 __tcp_ack_snd_check(sk
, 1);
4822 * This routine is only called when we have urgent data
4823 * signaled. Its the 'slow' part of tcp_urg. It could be
4824 * moved inline now as tcp_urg is only called from one
4825 * place. We handle URGent data wrong. We have to - as
4826 * BSD still doesn't use the correction from RFC961.
4827 * For 1003.1g we should support a new option TCP_STDURG to permit
4828 * either form (or just set the sysctl tcp_stdurg).
4831 static void tcp_check_urg(struct sock
*sk
, const struct tcphdr
*th
)
4833 struct tcp_sock
*tp
= tcp_sk(sk
);
4834 u32 ptr
= ntohs(th
->urg_ptr
);
4836 if (ptr
&& !sysctl_tcp_stdurg
)
4838 ptr
+= ntohl(th
->seq
);
4840 /* Ignore urgent data that we've already seen and read. */
4841 if (after(tp
->copied_seq
, ptr
))
4844 /* Do not replay urg ptr.
4846 * NOTE: interesting situation not covered by specs.
4847 * Misbehaving sender may send urg ptr, pointing to segment,
4848 * which we already have in ofo queue. We are not able to fetch
4849 * such data and will stay in TCP_URG_NOTYET until will be eaten
4850 * by recvmsg(). Seems, we are not obliged to handle such wicked
4851 * situations. But it is worth to think about possibility of some
4852 * DoSes using some hypothetical application level deadlock.
4854 if (before(ptr
, tp
->rcv_nxt
))
4857 /* Do we already have a newer (or duplicate) urgent pointer? */
4858 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
4861 /* Tell the world about our new urgent pointer. */
4864 /* We may be adding urgent data when the last byte read was
4865 * urgent. To do this requires some care. We cannot just ignore
4866 * tp->copied_seq since we would read the last urgent byte again
4867 * as data, nor can we alter copied_seq until this data arrives
4868 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4870 * NOTE. Double Dutch. Rendering to plain English: author of comment
4871 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4872 * and expect that both A and B disappear from stream. This is _wrong_.
4873 * Though this happens in BSD with high probability, this is occasional.
4874 * Any application relying on this is buggy. Note also, that fix "works"
4875 * only in this artificial test. Insert some normal data between A and B and we will
4876 * decline of BSD again. Verdict: it is better to remove to trap
4879 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
4880 !sock_flag(sk
, SOCK_URGINLINE
) && tp
->copied_seq
!= tp
->rcv_nxt
) {
4881 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
4883 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4884 __skb_unlink(skb
, &sk
->sk_receive_queue
);
4889 tp
->urg_data
= TCP_URG_NOTYET
;
4892 /* Disable header prediction. */
4896 /* This is the 'fast' part of urgent handling. */
4897 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, const struct tcphdr
*th
)
4899 struct tcp_sock
*tp
= tcp_sk(sk
);
4901 /* Check if we get a new urgent pointer - normally not. */
4903 tcp_check_urg(sk
, th
);
4905 /* Do we wait for any urgent data? - normally not... */
4906 if (tp
->urg_data
== TCP_URG_NOTYET
) {
4907 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
4910 /* Is the urgent pointer pointing into this packet? */
4911 if (ptr
< skb
->len
) {
4913 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
4915 tp
->urg_data
= TCP_URG_VALID
| tmp
;
4916 if (!sock_flag(sk
, SOCK_DEAD
))
4917 sk
->sk_data_ready(sk
);
4922 static int tcp_copy_to_iovec(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
4924 struct tcp_sock
*tp
= tcp_sk(sk
);
4925 int chunk
= skb
->len
- hlen
;
4929 if (skb_csum_unnecessary(skb
))
4930 err
= skb_copy_datagram_iovec(skb
, hlen
, tp
->ucopy
.iov
, chunk
);
4932 err
= skb_copy_and_csum_datagram_iovec(skb
, hlen
,
4936 tp
->ucopy
.len
-= chunk
;
4937 tp
->copied_seq
+= chunk
;
4938 tcp_rcv_space_adjust(sk
);
4945 static __sum16
__tcp_checksum_complete_user(struct sock
*sk
,
4946 struct sk_buff
*skb
)
4950 if (sock_owned_by_user(sk
)) {
4952 result
= __tcp_checksum_complete(skb
);
4955 result
= __tcp_checksum_complete(skb
);
4960 static inline bool tcp_checksum_complete_user(struct sock
*sk
,
4961 struct sk_buff
*skb
)
4963 return !skb_csum_unnecessary(skb
) &&
4964 __tcp_checksum_complete_user(sk
, skb
);
4967 #ifdef CONFIG_NET_DMA
4968 static bool tcp_dma_try_early_copy(struct sock
*sk
, struct sk_buff
*skb
,
4971 struct tcp_sock
*tp
= tcp_sk(sk
);
4972 int chunk
= skb
->len
- hlen
;
4974 bool copied_early
= false;
4976 if (tp
->ucopy
.wakeup
)
4979 if (!tp
->ucopy
.dma_chan
&& tp
->ucopy
.pinned_list
)
4980 tp
->ucopy
.dma_chan
= net_dma_find_channel();
4982 if (tp
->ucopy
.dma_chan
&& skb_csum_unnecessary(skb
)) {
4984 dma_cookie
= dma_skb_copy_datagram_iovec(tp
->ucopy
.dma_chan
,
4986 tp
->ucopy
.iov
, chunk
,
4987 tp
->ucopy
.pinned_list
);
4992 tp
->ucopy
.dma_cookie
= dma_cookie
;
4993 copied_early
= true;
4995 tp
->ucopy
.len
-= chunk
;
4996 tp
->copied_seq
+= chunk
;
4997 tcp_rcv_space_adjust(sk
);
4999 if ((tp
->ucopy
.len
== 0) ||
5000 (tcp_flag_word(tcp_hdr(skb
)) & TCP_FLAG_PSH
) ||
5001 (atomic_read(&sk
->sk_rmem_alloc
) > (sk
->sk_rcvbuf
>> 1))) {
5002 tp
->ucopy
.wakeup
= 1;
5003 sk
->sk_data_ready(sk
);
5005 } else if (chunk
> 0) {
5006 tp
->ucopy
.wakeup
= 1;
5007 sk
->sk_data_ready(sk
);
5010 return copied_early
;
5012 #endif /* CONFIG_NET_DMA */
5014 /* Does PAWS and seqno based validation of an incoming segment, flags will
5015 * play significant role here.
5017 static bool tcp_validate_incoming(struct sock
*sk
, struct sk_buff
*skb
,
5018 const struct tcphdr
*th
, int syn_inerr
)
5020 struct tcp_sock
*tp
= tcp_sk(sk
);
5022 /* RFC1323: H1. Apply PAWS check first. */
5023 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
5024 tcp_paws_discard(sk
, skb
)) {
5026 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSESTABREJECTED
);
5027 tcp_send_dupack(sk
, skb
);
5030 /* Reset is accepted even if it did not pass PAWS. */
5033 /* Step 1: check sequence number */
5034 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5035 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5036 * (RST) segments are validated by checking their SEQ-fields."
5037 * And page 69: "If an incoming segment is not acceptable,
5038 * an acknowledgment should be sent in reply (unless the RST
5039 * bit is set, if so drop the segment and return)".
5044 tcp_send_dupack(sk
, skb
);
5049 /* Step 2: check RST bit */
5052 * If sequence number exactly matches RCV.NXT, then
5053 * RESET the connection
5055 * Send a challenge ACK
5057 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
)
5060 tcp_send_challenge_ack(sk
);
5064 /* step 3: check security and precedence [ignored] */
5066 /* step 4: Check for a SYN
5067 * RFC 5691 4.2 : Send a challenge ack
5072 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5073 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSYNCHALLENGE
);
5074 tcp_send_challenge_ack(sk
);
5086 * TCP receive function for the ESTABLISHED state.
5088 * It is split into a fast path and a slow path. The fast path is
5090 * - A zero window was announced from us - zero window probing
5091 * is only handled properly in the slow path.
5092 * - Out of order segments arrived.
5093 * - Urgent data is expected.
5094 * - There is no buffer space left
5095 * - Unexpected TCP flags/window values/header lengths are received
5096 * (detected by checking the TCP header against pred_flags)
5097 * - Data is sent in both directions. Fast path only supports pure senders
5098 * or pure receivers (this means either the sequence number or the ack
5099 * value must stay constant)
5100 * - Unexpected TCP option.
5102 * When these conditions are not satisfied it drops into a standard
5103 * receive procedure patterned after RFC793 to handle all cases.
5104 * The first three cases are guaranteed by proper pred_flags setting,
5105 * the rest is checked inline. Fast processing is turned on in
5106 * tcp_data_queue when everything is OK.
5108 void tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
5109 const struct tcphdr
*th
, unsigned int len
)
5111 struct tcp_sock
*tp
= tcp_sk(sk
);
5113 if (unlikely(sk
->sk_rx_dst
== NULL
))
5114 inet_csk(sk
)->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
5116 * Header prediction.
5117 * The code loosely follows the one in the famous
5118 * "30 instruction TCP receive" Van Jacobson mail.
5120 * Van's trick is to deposit buffers into socket queue
5121 * on a device interrupt, to call tcp_recv function
5122 * on the receive process context and checksum and copy
5123 * the buffer to user space. smart...
5125 * Our current scheme is not silly either but we take the
5126 * extra cost of the net_bh soft interrupt processing...
5127 * We do checksum and copy also but from device to kernel.
5130 tp
->rx_opt
.saw_tstamp
= 0;
5132 /* pred_flags is 0xS?10 << 16 + snd_wnd
5133 * if header_prediction is to be made
5134 * 'S' will always be tp->tcp_header_len >> 2
5135 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5136 * turn it off (when there are holes in the receive
5137 * space for instance)
5138 * PSH flag is ignored.
5141 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
5142 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
&&
5143 !after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
)) {
5144 int tcp_header_len
= tp
->tcp_header_len
;
5146 /* Timestamp header prediction: tcp_header_len
5147 * is automatically equal to th->doff*4 due to pred_flags
5151 /* Check timestamp */
5152 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
5153 /* No? Slow path! */
5154 if (!tcp_parse_aligned_timestamp(tp
, th
))
5157 /* If PAWS failed, check it more carefully in slow path */
5158 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
5161 /* DO NOT update ts_recent here, if checksum fails
5162 * and timestamp was corrupted part, it will result
5163 * in a hung connection since we will drop all
5164 * future packets due to the PAWS test.
5168 if (len
<= tcp_header_len
) {
5169 /* Bulk data transfer: sender */
5170 if (len
== tcp_header_len
) {
5171 /* Predicted packet is in window by definition.
5172 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5173 * Hence, check seq<=rcv_wup reduces to:
5175 if (tcp_header_len
==
5176 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5177 tp
->rcv_nxt
== tp
->rcv_wup
)
5178 tcp_store_ts_recent(tp
);
5180 /* We know that such packets are checksummed
5183 tcp_ack(sk
, skb
, 0);
5185 tcp_data_snd_check(sk
);
5187 } else { /* Header too small */
5188 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5193 int copied_early
= 0;
5194 bool fragstolen
= false;
5196 if (tp
->copied_seq
== tp
->rcv_nxt
&&
5197 len
- tcp_header_len
<= tp
->ucopy
.len
) {
5198 #ifdef CONFIG_NET_DMA
5199 if (tp
->ucopy
.task
== current
&&
5200 sock_owned_by_user(sk
) &&
5201 tcp_dma_try_early_copy(sk
, skb
, tcp_header_len
)) {
5206 if (tp
->ucopy
.task
== current
&&
5207 sock_owned_by_user(sk
) && !copied_early
) {
5208 __set_current_state(TASK_RUNNING
);
5210 if (!tcp_copy_to_iovec(sk
, skb
, tcp_header_len
))
5214 /* Predicted packet is in window by definition.
5215 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5216 * Hence, check seq<=rcv_wup reduces to:
5218 if (tcp_header_len
==
5219 (sizeof(struct tcphdr
) +
5220 TCPOLEN_TSTAMP_ALIGNED
) &&
5221 tp
->rcv_nxt
== tp
->rcv_wup
)
5222 tcp_store_ts_recent(tp
);
5224 tcp_rcv_rtt_measure_ts(sk
, skb
);
5226 __skb_pull(skb
, tcp_header_len
);
5227 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
5228 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITSTOUSER
);
5231 tcp_cleanup_rbuf(sk
, skb
->len
);
5234 if (tcp_checksum_complete_user(sk
, skb
))
5237 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
5240 /* Predicted packet is in window by definition.
5241 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5242 * Hence, check seq<=rcv_wup reduces to:
5244 if (tcp_header_len
==
5245 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5246 tp
->rcv_nxt
== tp
->rcv_wup
)
5247 tcp_store_ts_recent(tp
);
5249 tcp_rcv_rtt_measure_ts(sk
, skb
);
5251 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITS
);
5253 /* Bulk data transfer: receiver */
5254 eaten
= tcp_queue_rcv(sk
, skb
, tcp_header_len
,
5258 tcp_event_data_recv(sk
, skb
);
5260 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
5261 /* Well, only one small jumplet in fast path... */
5262 tcp_ack(sk
, skb
, FLAG_DATA
);
5263 tcp_data_snd_check(sk
);
5264 if (!inet_csk_ack_scheduled(sk
))
5268 if (!copied_early
|| tp
->rcv_nxt
!= tp
->rcv_wup
)
5269 __tcp_ack_snd_check(sk
, 0);
5271 #ifdef CONFIG_NET_DMA
5273 __skb_queue_tail(&sk
->sk_async_wait_queue
, skb
);
5277 kfree_skb_partial(skb
, fragstolen
);
5278 sk
->sk_data_ready(sk
);
5284 if (len
< (th
->doff
<< 2) || tcp_checksum_complete_user(sk
, skb
))
5287 if (!th
->ack
&& !th
->rst
)
5291 * Standard slow path.
5294 if (!tcp_validate_incoming(sk
, skb
, th
, 1))
5298 if (tcp_ack(sk
, skb
, FLAG_SLOWPATH
| FLAG_UPDATE_TS_RECENT
) < 0)
5301 tcp_rcv_rtt_measure_ts(sk
, skb
);
5303 /* Process urgent data. */
5304 tcp_urg(sk
, skb
, th
);
5306 /* step 7: process the segment text */
5307 tcp_data_queue(sk
, skb
);
5309 tcp_data_snd_check(sk
);
5310 tcp_ack_snd_check(sk
);
5314 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_CSUMERRORS
);
5315 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5320 EXPORT_SYMBOL(tcp_rcv_established
);
5322 void tcp_finish_connect(struct sock
*sk
, struct sk_buff
*skb
)
5324 struct tcp_sock
*tp
= tcp_sk(sk
);
5325 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5327 tcp_set_state(sk
, TCP_ESTABLISHED
);
5330 icsk
->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
5331 security_inet_conn_established(sk
, skb
);
5334 /* Make sure socket is routed, for correct metrics. */
5335 icsk
->icsk_af_ops
->rebuild_header(sk
);
5337 tcp_init_metrics(sk
);
5339 tcp_init_congestion_control(sk
);
5341 /* Prevent spurious tcp_cwnd_restart() on first data
5344 tp
->lsndtime
= tcp_time_stamp
;
5346 tcp_init_buffer_space(sk
);
5348 if (sock_flag(sk
, SOCK_KEEPOPEN
))
5349 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
5351 if (!tp
->rx_opt
.snd_wscale
)
5352 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
5356 if (!sock_flag(sk
, SOCK_DEAD
)) {
5357 sk
->sk_state_change(sk
);
5358 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5362 static bool tcp_rcv_fastopen_synack(struct sock
*sk
, struct sk_buff
*synack
,
5363 struct tcp_fastopen_cookie
*cookie
)
5365 struct tcp_sock
*tp
= tcp_sk(sk
);
5366 struct sk_buff
*data
= tp
->syn_data
? tcp_write_queue_head(sk
) : NULL
;
5367 u16 mss
= tp
->rx_opt
.mss_clamp
;
5370 if (mss
== tp
->rx_opt
.user_mss
) {
5371 struct tcp_options_received opt
;
5373 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5374 tcp_clear_options(&opt
);
5375 opt
.user_mss
= opt
.mss_clamp
= 0;
5376 tcp_parse_options(synack
, &opt
, 0, NULL
);
5377 mss
= opt
.mss_clamp
;
5380 if (!tp
->syn_fastopen
) /* Ignore an unsolicited cookie */
5383 /* The SYN-ACK neither has cookie nor acknowledges the data. Presumably
5384 * the remote receives only the retransmitted (regular) SYNs: either
5385 * the original SYN-data or the corresponding SYN-ACK is lost.
5387 syn_drop
= (cookie
->len
<= 0 && data
&& tp
->total_retrans
);
5389 tcp_fastopen_cache_set(sk
, mss
, cookie
, syn_drop
);
5391 if (data
) { /* Retransmit unacked data in SYN */
5392 tcp_for_write_queue_from(data
, sk
) {
5393 if (data
== tcp_send_head(sk
) ||
5394 __tcp_retransmit_skb(sk
, data
))
5398 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPFASTOPENACTIVEFAIL
);
5401 tp
->syn_data_acked
= tp
->syn_data
;
5402 if (tp
->syn_data_acked
)
5403 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPFASTOPENACTIVE
);
5407 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5408 const struct tcphdr
*th
, unsigned int len
)
5410 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5411 struct tcp_sock
*tp
= tcp_sk(sk
);
5412 struct tcp_fastopen_cookie foc
= { .len
= -1 };
5413 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
5415 tcp_parse_options(skb
, &tp
->rx_opt
, 0, &foc
);
5416 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
5417 tp
->rx_opt
.rcv_tsecr
-= tp
->tsoffset
;
5421 * "If the state is SYN-SENT then
5422 * first check the ACK bit
5423 * If the ACK bit is set
5424 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5425 * a reset (unless the RST bit is set, if so drop
5426 * the segment and return)"
5428 if (!after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_una
) ||
5429 after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
))
5430 goto reset_and_undo
;
5432 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
5433 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
5435 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSACTIVEREJECTED
);
5436 goto reset_and_undo
;
5439 /* Now ACK is acceptable.
5441 * "If the RST bit is set
5442 * If the ACK was acceptable then signal the user "error:
5443 * connection reset", drop the segment, enter CLOSED state,
5444 * delete TCB, and return."
5453 * "fifth, if neither of the SYN or RST bits is set then
5454 * drop the segment and return."
5460 goto discard_and_undo
;
5463 * "If the SYN bit is on ...
5464 * are acceptable then ...
5465 * (our SYN has been ACKed), change the connection
5466 * state to ESTABLISHED..."
5469 TCP_ECN_rcv_synack(tp
, th
);
5471 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5472 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
5474 /* Ok.. it's good. Set up sequence numbers and
5475 * move to established.
5477 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5478 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5480 /* RFC1323: The window in SYN & SYN/ACK segments is
5483 tp
->snd_wnd
= ntohs(th
->window
);
5485 if (!tp
->rx_opt
.wscale_ok
) {
5486 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
5487 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
5490 if (tp
->rx_opt
.saw_tstamp
) {
5491 tp
->rx_opt
.tstamp_ok
= 1;
5492 tp
->tcp_header_len
=
5493 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5494 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5495 tcp_store_ts_recent(tp
);
5497 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5500 if (tcp_is_sack(tp
) && sysctl_tcp_fack
)
5501 tcp_enable_fack(tp
);
5504 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5505 tcp_initialize_rcv_mss(sk
);
5507 /* Remember, tcp_poll() does not lock socket!
5508 * Change state from SYN-SENT only after copied_seq
5509 * is initialized. */
5510 tp
->copied_seq
= tp
->rcv_nxt
;
5514 tcp_finish_connect(sk
, skb
);
5516 if ((tp
->syn_fastopen
|| tp
->syn_data
) &&
5517 tcp_rcv_fastopen_synack(sk
, skb
, &foc
))
5520 if (sk
->sk_write_pending
||
5521 icsk
->icsk_accept_queue
.rskq_defer_accept
||
5522 icsk
->icsk_ack
.pingpong
) {
5523 /* Save one ACK. Data will be ready after
5524 * several ticks, if write_pending is set.
5526 * It may be deleted, but with this feature tcpdumps
5527 * look so _wonderfully_ clever, that I was not able
5528 * to stand against the temptation 8) --ANK
5530 inet_csk_schedule_ack(sk
);
5531 icsk
->icsk_ack
.lrcvtime
= tcp_time_stamp
;
5532 tcp_enter_quickack_mode(sk
);
5533 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
5534 TCP_DELACK_MAX
, TCP_RTO_MAX
);
5545 /* No ACK in the segment */
5549 * "If the RST bit is set
5551 * Otherwise (no ACK) drop the segment and return."
5554 goto discard_and_undo
;
5558 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&&
5559 tcp_paws_reject(&tp
->rx_opt
, 0))
5560 goto discard_and_undo
;
5563 /* We see SYN without ACK. It is attempt of
5564 * simultaneous connect with crossed SYNs.
5565 * Particularly, it can be connect to self.
5567 tcp_set_state(sk
, TCP_SYN_RECV
);
5569 if (tp
->rx_opt
.saw_tstamp
) {
5570 tp
->rx_opt
.tstamp_ok
= 1;
5571 tcp_store_ts_recent(tp
);
5572 tp
->tcp_header_len
=
5573 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5575 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5578 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5579 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5581 /* RFC1323: The window in SYN & SYN/ACK segments is
5584 tp
->snd_wnd
= ntohs(th
->window
);
5585 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
5586 tp
->max_window
= tp
->snd_wnd
;
5588 TCP_ECN_rcv_syn(tp
, th
);
5591 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5592 tcp_initialize_rcv_mss(sk
);
5594 tcp_send_synack(sk
);
5596 /* Note, we could accept data and URG from this segment.
5597 * There are no obstacles to make this (except that we must
5598 * either change tcp_recvmsg() to prevent it from returning data
5599 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5601 * However, if we ignore data in ACKless segments sometimes,
5602 * we have no reasons to accept it sometimes.
5603 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5604 * is not flawless. So, discard packet for sanity.
5605 * Uncomment this return to process the data.
5612 /* "fifth, if neither of the SYN or RST bits is set then
5613 * drop the segment and return."
5617 tcp_clear_options(&tp
->rx_opt
);
5618 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5622 tcp_clear_options(&tp
->rx_opt
);
5623 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5628 * This function implements the receiving procedure of RFC 793 for
5629 * all states except ESTABLISHED and TIME_WAIT.
5630 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5631 * address independent.
5634 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5635 const struct tcphdr
*th
, unsigned int len
)
5637 struct tcp_sock
*tp
= tcp_sk(sk
);
5638 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5639 struct request_sock
*req
;
5644 tp
->rx_opt
.saw_tstamp
= 0;
5646 switch (sk
->sk_state
) {
5660 if (icsk
->icsk_af_ops
->conn_request(sk
, skb
) < 0)
5663 /* Now we have several options: In theory there is
5664 * nothing else in the frame. KA9Q has an option to
5665 * send data with the syn, BSD accepts data with the
5666 * syn up to the [to be] advertised window and
5667 * Solaris 2.1 gives you a protocol error. For now
5668 * we just ignore it, that fits the spec precisely
5669 * and avoids incompatibilities. It would be nice in
5670 * future to drop through and process the data.
5672 * Now that TTCP is starting to be used we ought to
5674 * But, this leaves one open to an easy denial of
5675 * service attack, and SYN cookies can't defend
5676 * against this problem. So, we drop the data
5677 * in the interest of security over speed unless
5678 * it's still in use.
5686 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
, len
);
5690 /* Do step6 onward by hand. */
5691 tcp_urg(sk
, skb
, th
);
5693 tcp_data_snd_check(sk
);
5697 req
= tp
->fastopen_rsk
;
5699 WARN_ON_ONCE(sk
->sk_state
!= TCP_SYN_RECV
&&
5700 sk
->sk_state
!= TCP_FIN_WAIT1
);
5702 if (tcp_check_req(sk
, skb
, req
, NULL
, true) == NULL
)
5706 if (!th
->ack
&& !th
->rst
)
5709 if (!tcp_validate_incoming(sk
, skb
, th
, 0))
5712 /* step 5: check the ACK field */
5713 acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
|
5714 FLAG_UPDATE_TS_RECENT
) > 0;
5716 switch (sk
->sk_state
) {
5721 /* Once we leave TCP_SYN_RECV, we no longer need req
5725 synack_stamp
= tcp_rsk(req
)->snt_synack
;
5726 tp
->total_retrans
= req
->num_retrans
;
5727 reqsk_fastopen_remove(sk
, req
, false);
5729 synack_stamp
= tp
->lsndtime
;
5730 /* Make sure socket is routed, for correct metrics. */
5731 icsk
->icsk_af_ops
->rebuild_header(sk
);
5732 tcp_init_congestion_control(sk
);
5735 tp
->copied_seq
= tp
->rcv_nxt
;
5736 tcp_init_buffer_space(sk
);
5739 tcp_set_state(sk
, TCP_ESTABLISHED
);
5740 sk
->sk_state_change(sk
);
5742 /* Note, that this wakeup is only for marginal crossed SYN case.
5743 * Passively open sockets are not waked up, because
5744 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5747 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5749 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
5750 tp
->snd_wnd
= ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
;
5751 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5752 tcp_synack_rtt_meas(sk
, synack_stamp
);
5754 if (tp
->rx_opt
.tstamp_ok
)
5755 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5758 /* Re-arm the timer because data may have been sent out.
5759 * This is similar to the regular data transmission case
5760 * when new data has just been ack'ed.
5762 * (TFO) - we could try to be more aggressive and
5763 * retransmitting any data sooner based on when they
5768 tcp_init_metrics(sk
);
5770 tcp_update_pacing_rate(sk
);
5772 /* Prevent spurious tcp_cwnd_restart() on first data packet */
5773 tp
->lsndtime
= tcp_time_stamp
;
5775 tcp_initialize_rcv_mss(sk
);
5776 tcp_fast_path_on(tp
);
5779 case TCP_FIN_WAIT1
: {
5780 struct dst_entry
*dst
;
5783 /* If we enter the TCP_FIN_WAIT1 state and we are a
5784 * Fast Open socket and this is the first acceptable
5785 * ACK we have received, this would have acknowledged
5786 * our SYNACK so stop the SYNACK timer.
5789 /* Return RST if ack_seq is invalid.
5790 * Note that RFC793 only says to generate a
5791 * DUPACK for it but for TCP Fast Open it seems
5792 * better to treat this case like TCP_SYN_RECV
5797 /* We no longer need the request sock. */
5798 reqsk_fastopen_remove(sk
, req
, false);
5801 if (tp
->snd_una
!= tp
->write_seq
)
5804 tcp_set_state(sk
, TCP_FIN_WAIT2
);
5805 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
5807 dst
= __sk_dst_get(sk
);
5811 if (!sock_flag(sk
, SOCK_DEAD
)) {
5812 /* Wake up lingering close() */
5813 sk
->sk_state_change(sk
);
5817 if (tp
->linger2
< 0 ||
5818 (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5819 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
))) {
5821 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
5825 tmo
= tcp_fin_time(sk
);
5826 if (tmo
> TCP_TIMEWAIT_LEN
) {
5827 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
5828 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
5829 /* Bad case. We could lose such FIN otherwise.
5830 * It is not a big problem, but it looks confusing
5831 * and not so rare event. We still can lose it now,
5832 * if it spins in bh_lock_sock(), but it is really
5835 inet_csk_reset_keepalive_timer(sk
, tmo
);
5837 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
5844 if (tp
->snd_una
== tp
->write_seq
) {
5845 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
5851 if (tp
->snd_una
== tp
->write_seq
) {
5852 tcp_update_metrics(sk
);
5859 /* step 6: check the URG bit */
5860 tcp_urg(sk
, skb
, th
);
5862 /* step 7: process the segment text */
5863 switch (sk
->sk_state
) {
5864 case TCP_CLOSE_WAIT
:
5867 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
5871 /* RFC 793 says to queue data in these states,
5872 * RFC 1122 says we MUST send a reset.
5873 * BSD 4.4 also does reset.
5875 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
5876 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5877 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
5878 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
5884 case TCP_ESTABLISHED
:
5885 tcp_data_queue(sk
, skb
);
5890 /* tcp_data could move socket to TIME-WAIT */
5891 if (sk
->sk_state
!= TCP_CLOSE
) {
5892 tcp_data_snd_check(sk
);
5893 tcp_ack_snd_check(sk
);
5902 EXPORT_SYMBOL(tcp_rcv_state_process
);