2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
23 * Pedro Roque : Fast Retransmit/Recovery.
25 * Retransmit queue handled by TCP.
26 * Better retransmit timer handling.
27 * New congestion avoidance.
31 * Eric : Fast Retransmit.
32 * Randy Scott : MSS option defines.
33 * Eric Schenk : Fixes to slow start algorithm.
34 * Eric Schenk : Yet another double ACK bug.
35 * Eric Schenk : Delayed ACK bug fixes.
36 * Eric Schenk : Floyd style fast retrans war avoidance.
37 * David S. Miller : Don't allow zero congestion window.
38 * Eric Schenk : Fix retransmitter so that it sends
39 * next packet on ack of previous packet.
40 * Andi Kleen : Moved open_request checking here
41 * and process RSTs for open_requests.
42 * Andi Kleen : Better prune_queue, and other fixes.
43 * Andrey Savochkin: Fix RTT measurements in the presence of
45 * Andrey Savochkin: Check sequence numbers correctly when
46 * removing SACKs due to in sequence incoming
48 * Andi Kleen: Make sure we never ack data there is not
49 * enough room for. Also make this condition
50 * a fatal error if it might still happen.
51 * Andi Kleen: Add tcp_measure_rcv_mss to make
52 * connections with MSS<min(MTU,ann. MSS)
53 * work without delayed acks.
54 * Andi Kleen: Process packets with PSH set in the
56 * J Hadi Salim: ECN support
59 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
60 * engine. Lots of bugs are found.
61 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
64 #define pr_fmt(fmt) "TCP: " fmt
67 #include <linux/slab.h>
68 #include <linux/module.h>
69 #include <linux/sysctl.h>
70 #include <linux/kernel.h>
73 #include <net/inet_common.h>
74 #include <linux/ipsec.h>
75 #include <asm/unaligned.h>
76 #include <net/netdma.h>
78 int sysctl_tcp_timestamps __read_mostly
= 1;
79 int sysctl_tcp_window_scaling __read_mostly
= 1;
80 int sysctl_tcp_sack __read_mostly
= 1;
81 int sysctl_tcp_fack __read_mostly
= 1;
82 int sysctl_tcp_reordering __read_mostly
= TCP_FASTRETRANS_THRESH
;
83 EXPORT_SYMBOL(sysctl_tcp_reordering
);
84 int sysctl_tcp_dsack __read_mostly
= 1;
85 int sysctl_tcp_app_win __read_mostly
= 31;
86 int sysctl_tcp_adv_win_scale __read_mostly
= 1;
87 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale
);
89 /* rfc5961 challenge ack rate limiting */
90 int sysctl_tcp_challenge_ack_limit
= 100;
92 int sysctl_tcp_stdurg __read_mostly
;
93 int sysctl_tcp_rfc1337 __read_mostly
;
94 int sysctl_tcp_max_orphans __read_mostly
= NR_FILE
;
95 int sysctl_tcp_frto __read_mostly
= 2;
97 int sysctl_tcp_thin_dupack __read_mostly
;
99 int sysctl_tcp_moderate_rcvbuf __read_mostly
= 1;
100 int sysctl_tcp_early_retrans __read_mostly
= 3;
102 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
103 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
104 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
105 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
106 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
107 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
108 #define FLAG_ECE 0x40 /* ECE in this ACK */
109 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
110 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
111 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
112 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
113 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
114 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
116 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
117 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
118 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
119 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
121 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
122 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
124 /* Adapt the MSS value used to make delayed ack decision to the
127 static void tcp_measure_rcv_mss(struct sock
*sk
, const struct sk_buff
*skb
)
129 struct inet_connection_sock
*icsk
= inet_csk(sk
);
130 const unsigned int lss
= icsk
->icsk_ack
.last_seg_size
;
133 icsk
->icsk_ack
.last_seg_size
= 0;
135 /* skb->len may jitter because of SACKs, even if peer
136 * sends good full-sized frames.
138 len
= skb_shinfo(skb
)->gso_size
? : skb
->len
;
139 if (len
>= icsk
->icsk_ack
.rcv_mss
) {
140 icsk
->icsk_ack
.rcv_mss
= len
;
142 /* Otherwise, we make more careful check taking into account,
143 * that SACKs block is variable.
145 * "len" is invariant segment length, including TCP header.
147 len
+= skb
->data
- skb_transport_header(skb
);
148 if (len
>= TCP_MSS_DEFAULT
+ sizeof(struct tcphdr
) ||
149 /* If PSH is not set, packet should be
150 * full sized, provided peer TCP is not badly broken.
151 * This observation (if it is correct 8)) allows
152 * to handle super-low mtu links fairly.
154 (len
>= TCP_MIN_MSS
+ sizeof(struct tcphdr
) &&
155 !(tcp_flag_word(tcp_hdr(skb
)) & TCP_REMNANT
))) {
156 /* Subtract also invariant (if peer is RFC compliant),
157 * tcp header plus fixed timestamp option length.
158 * Resulting "len" is MSS free of SACK jitter.
160 len
-= tcp_sk(sk
)->tcp_header_len
;
161 icsk
->icsk_ack
.last_seg_size
= len
;
163 icsk
->icsk_ack
.rcv_mss
= len
;
167 if (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
)
168 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED2
;
169 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
173 static void tcp_incr_quickack(struct sock
*sk
)
175 struct inet_connection_sock
*icsk
= inet_csk(sk
);
176 unsigned int quickacks
= tcp_sk(sk
)->rcv_wnd
/ (2 * icsk
->icsk_ack
.rcv_mss
);
180 if (quickacks
> icsk
->icsk_ack
.quick
)
181 icsk
->icsk_ack
.quick
= min(quickacks
, TCP_MAX_QUICKACKS
);
184 static void tcp_enter_quickack_mode(struct sock
*sk
)
186 struct inet_connection_sock
*icsk
= inet_csk(sk
);
187 tcp_incr_quickack(sk
);
188 icsk
->icsk_ack
.pingpong
= 0;
189 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
192 /* Send ACKs quickly, if "quick" count is not exhausted
193 * and the session is not interactive.
196 static inline bool tcp_in_quickack_mode(const struct sock
*sk
)
198 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
200 return icsk
->icsk_ack
.quick
&& !icsk
->icsk_ack
.pingpong
;
203 static inline void TCP_ECN_queue_cwr(struct tcp_sock
*tp
)
205 if (tp
->ecn_flags
& TCP_ECN_OK
)
206 tp
->ecn_flags
|= TCP_ECN_QUEUE_CWR
;
209 static inline void TCP_ECN_accept_cwr(struct tcp_sock
*tp
, const struct sk_buff
*skb
)
211 if (tcp_hdr(skb
)->cwr
)
212 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
215 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock
*tp
)
217 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
220 static inline void TCP_ECN_check_ce(struct tcp_sock
*tp
, const struct sk_buff
*skb
)
222 if (!(tp
->ecn_flags
& TCP_ECN_OK
))
225 switch (TCP_SKB_CB(skb
)->ip_dsfield
& INET_ECN_MASK
) {
226 case INET_ECN_NOT_ECT
:
227 /* Funny extension: if ECT is not set on a segment,
228 * and we already seen ECT on a previous segment,
229 * it is probably a retransmit.
231 if (tp
->ecn_flags
& TCP_ECN_SEEN
)
232 tcp_enter_quickack_mode((struct sock
*)tp
);
235 if (!(tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)) {
236 /* Better not delay acks, sender can have a very low cwnd */
237 tcp_enter_quickack_mode((struct sock
*)tp
);
238 tp
->ecn_flags
|= TCP_ECN_DEMAND_CWR
;
242 tp
->ecn_flags
|= TCP_ECN_SEEN
;
246 static inline void TCP_ECN_rcv_synack(struct tcp_sock
*tp
, const struct tcphdr
*th
)
248 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| th
->cwr
))
249 tp
->ecn_flags
&= ~TCP_ECN_OK
;
252 static inline void TCP_ECN_rcv_syn(struct tcp_sock
*tp
, const struct tcphdr
*th
)
254 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| !th
->cwr
))
255 tp
->ecn_flags
&= ~TCP_ECN_OK
;
258 static bool TCP_ECN_rcv_ecn_echo(const struct tcp_sock
*tp
, const struct tcphdr
*th
)
260 if (th
->ece
&& !th
->syn
&& (tp
->ecn_flags
& TCP_ECN_OK
))
265 /* Buffer size and advertised window tuning.
267 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
270 static void tcp_sndbuf_expand(struct sock
*sk
)
272 const struct tcp_sock
*tp
= tcp_sk(sk
);
276 /* Worst case is non GSO/TSO : each frame consumes one skb
277 * and skb->head is kmalloced using power of two area of memory
279 per_mss
= max_t(u32
, tp
->rx_opt
.mss_clamp
, tp
->mss_cache
) +
281 SKB_DATA_ALIGN(sizeof(struct skb_shared_info
));
283 per_mss
= roundup_pow_of_two(per_mss
) +
284 SKB_DATA_ALIGN(sizeof(struct sk_buff
));
286 nr_segs
= max_t(u32
, TCP_INIT_CWND
, tp
->snd_cwnd
);
287 nr_segs
= max_t(u32
, nr_segs
, tp
->reordering
+ 1);
289 /* Fast Recovery (RFC 5681 3.2) :
290 * Cubic needs 1.7 factor, rounded to 2 to include
291 * extra cushion (application might react slowly to POLLOUT)
293 sndmem
= 2 * nr_segs
* per_mss
;
295 if (sk
->sk_sndbuf
< sndmem
)
296 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
299 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
301 * All tcp_full_space() is split to two parts: "network" buffer, allocated
302 * forward and advertised in receiver window (tp->rcv_wnd) and
303 * "application buffer", required to isolate scheduling/application
304 * latencies from network.
305 * window_clamp is maximal advertised window. It can be less than
306 * tcp_full_space(), in this case tcp_full_space() - window_clamp
307 * is reserved for "application" buffer. The less window_clamp is
308 * the smoother our behaviour from viewpoint of network, but the lower
309 * throughput and the higher sensitivity of the connection to losses. 8)
311 * rcv_ssthresh is more strict window_clamp used at "slow start"
312 * phase to predict further behaviour of this connection.
313 * It is used for two goals:
314 * - to enforce header prediction at sender, even when application
315 * requires some significant "application buffer". It is check #1.
316 * - to prevent pruning of receive queue because of misprediction
317 * of receiver window. Check #2.
319 * The scheme does not work when sender sends good segments opening
320 * window and then starts to feed us spaghetti. But it should work
321 * in common situations. Otherwise, we have to rely on queue collapsing.
324 /* Slow part of check#2. */
325 static int __tcp_grow_window(const struct sock
*sk
, const struct sk_buff
*skb
)
327 struct tcp_sock
*tp
= tcp_sk(sk
);
329 int truesize
= tcp_win_from_space(skb
->truesize
) >> 1;
330 int window
= tcp_win_from_space(sysctl_tcp_rmem
[2]) >> 1;
332 while (tp
->rcv_ssthresh
<= window
) {
333 if (truesize
<= skb
->len
)
334 return 2 * inet_csk(sk
)->icsk_ack
.rcv_mss
;
342 static void tcp_grow_window(struct sock
*sk
, const struct sk_buff
*skb
)
344 struct tcp_sock
*tp
= tcp_sk(sk
);
347 if (tp
->rcv_ssthresh
< tp
->window_clamp
&&
348 (int)tp
->rcv_ssthresh
< tcp_space(sk
) &&
349 !sk_under_memory_pressure(sk
)) {
352 /* Check #2. Increase window, if skb with such overhead
353 * will fit to rcvbuf in future.
355 if (tcp_win_from_space(skb
->truesize
) <= skb
->len
)
356 incr
= 2 * tp
->advmss
;
358 incr
= __tcp_grow_window(sk
, skb
);
361 incr
= max_t(int, incr
, 2 * skb
->len
);
362 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
+ incr
,
364 inet_csk(sk
)->icsk_ack
.quick
|= 1;
369 /* 3. Tuning rcvbuf, when connection enters established state. */
370 static void tcp_fixup_rcvbuf(struct sock
*sk
)
372 u32 mss
= tcp_sk(sk
)->advmss
;
375 rcvmem
= 2 * SKB_TRUESIZE(mss
+ MAX_TCP_HEADER
) *
376 tcp_default_init_rwnd(mss
);
378 /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
379 * Allow enough cushion so that sender is not limited by our window
381 if (sysctl_tcp_moderate_rcvbuf
)
384 if (sk
->sk_rcvbuf
< rcvmem
)
385 sk
->sk_rcvbuf
= min(rcvmem
, sysctl_tcp_rmem
[2]);
388 /* 4. Try to fixup all. It is made immediately after connection enters
391 void tcp_init_buffer_space(struct sock
*sk
)
393 struct tcp_sock
*tp
= tcp_sk(sk
);
396 if (!(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
))
397 tcp_fixup_rcvbuf(sk
);
398 if (!(sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
))
399 tcp_sndbuf_expand(sk
);
401 tp
->rcvq_space
.space
= tp
->rcv_wnd
;
402 tp
->rcvq_space
.time
= tcp_time_stamp
;
403 tp
->rcvq_space
.seq
= tp
->copied_seq
;
405 maxwin
= tcp_full_space(sk
);
407 if (tp
->window_clamp
>= maxwin
) {
408 tp
->window_clamp
= maxwin
;
410 if (sysctl_tcp_app_win
&& maxwin
> 4 * tp
->advmss
)
411 tp
->window_clamp
= max(maxwin
-
412 (maxwin
>> sysctl_tcp_app_win
),
416 /* Force reservation of one segment. */
417 if (sysctl_tcp_app_win
&&
418 tp
->window_clamp
> 2 * tp
->advmss
&&
419 tp
->window_clamp
+ tp
->advmss
> maxwin
)
420 tp
->window_clamp
= max(2 * tp
->advmss
, maxwin
- tp
->advmss
);
422 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, tp
->window_clamp
);
423 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
426 /* 5. Recalculate window clamp after socket hit its memory bounds. */
427 static void tcp_clamp_window(struct sock
*sk
)
429 struct tcp_sock
*tp
= tcp_sk(sk
);
430 struct inet_connection_sock
*icsk
= inet_csk(sk
);
432 icsk
->icsk_ack
.quick
= 0;
434 if (sk
->sk_rcvbuf
< sysctl_tcp_rmem
[2] &&
435 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
) &&
436 !sk_under_memory_pressure(sk
) &&
437 sk_memory_allocated(sk
) < sk_prot_mem_limits(sk
, 0)) {
438 sk
->sk_rcvbuf
= min(atomic_read(&sk
->sk_rmem_alloc
),
441 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
)
442 tp
->rcv_ssthresh
= min(tp
->window_clamp
, 2U * tp
->advmss
);
445 /* Initialize RCV_MSS value.
446 * RCV_MSS is an our guess about MSS used by the peer.
447 * We haven't any direct information about the MSS.
448 * It's better to underestimate the RCV_MSS rather than overestimate.
449 * Overestimations make us ACKing less frequently than needed.
450 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
452 void tcp_initialize_rcv_mss(struct sock
*sk
)
454 const struct tcp_sock
*tp
= tcp_sk(sk
);
455 unsigned int hint
= min_t(unsigned int, tp
->advmss
, tp
->mss_cache
);
457 hint
= min(hint
, tp
->rcv_wnd
/ 2);
458 hint
= min(hint
, TCP_MSS_DEFAULT
);
459 hint
= max(hint
, TCP_MIN_MSS
);
461 inet_csk(sk
)->icsk_ack
.rcv_mss
= hint
;
463 EXPORT_SYMBOL(tcp_initialize_rcv_mss
);
465 /* Receiver "autotuning" code.
467 * The algorithm for RTT estimation w/o timestamps is based on
468 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
469 * <http://public.lanl.gov/radiant/pubs.html#DRS>
471 * More detail on this code can be found at
472 * <http://staff.psc.edu/jheffner/>,
473 * though this reference is out of date. A new paper
476 static void tcp_rcv_rtt_update(struct tcp_sock
*tp
, u32 sample
, int win_dep
)
478 u32 new_sample
= tp
->rcv_rtt_est
.rtt
;
484 if (new_sample
!= 0) {
485 /* If we sample in larger samples in the non-timestamp
486 * case, we could grossly overestimate the RTT especially
487 * with chatty applications or bulk transfer apps which
488 * are stalled on filesystem I/O.
490 * Also, since we are only going for a minimum in the
491 * non-timestamp case, we do not smooth things out
492 * else with timestamps disabled convergence takes too
496 m
-= (new_sample
>> 3);
504 /* No previous measure. */
508 if (tp
->rcv_rtt_est
.rtt
!= new_sample
)
509 tp
->rcv_rtt_est
.rtt
= new_sample
;
512 static inline void tcp_rcv_rtt_measure(struct tcp_sock
*tp
)
514 if (tp
->rcv_rtt_est
.time
== 0)
516 if (before(tp
->rcv_nxt
, tp
->rcv_rtt_est
.seq
))
518 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rcv_rtt_est
.time
, 1);
521 tp
->rcv_rtt_est
.seq
= tp
->rcv_nxt
+ tp
->rcv_wnd
;
522 tp
->rcv_rtt_est
.time
= tcp_time_stamp
;
525 static inline void tcp_rcv_rtt_measure_ts(struct sock
*sk
,
526 const struct sk_buff
*skb
)
528 struct tcp_sock
*tp
= tcp_sk(sk
);
529 if (tp
->rx_opt
.rcv_tsecr
&&
530 (TCP_SKB_CB(skb
)->end_seq
-
531 TCP_SKB_CB(skb
)->seq
>= inet_csk(sk
)->icsk_ack
.rcv_mss
))
532 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
, 0);
536 * This function should be called every time data is copied to user space.
537 * It calculates the appropriate TCP receive buffer space.
539 void tcp_rcv_space_adjust(struct sock
*sk
)
541 struct tcp_sock
*tp
= tcp_sk(sk
);
545 time
= tcp_time_stamp
- tp
->rcvq_space
.time
;
546 if (time
< (tp
->rcv_rtt_est
.rtt
>> 3) || tp
->rcv_rtt_est
.rtt
== 0)
549 /* Number of bytes copied to user in last RTT */
550 copied
= tp
->copied_seq
- tp
->rcvq_space
.seq
;
551 if (copied
<= tp
->rcvq_space
.space
)
555 * copied = bytes received in previous RTT, our base window
556 * To cope with packet losses, we need a 2x factor
557 * To cope with slow start, and sender growing its cwin by 100 %
558 * every RTT, we need a 4x factor, because the ACK we are sending
559 * now is for the next RTT, not the current one :
560 * <prev RTT . ><current RTT .. ><next RTT .... >
563 if (sysctl_tcp_moderate_rcvbuf
&&
564 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
)) {
565 int rcvwin
, rcvmem
, rcvbuf
;
567 /* minimal window to cope with packet losses, assuming
568 * steady state. Add some cushion because of small variations.
570 rcvwin
= (copied
<< 1) + 16 * tp
->advmss
;
572 /* If rate increased by 25%,
573 * assume slow start, rcvwin = 3 * copied
574 * If rate increased by 50%,
575 * assume sender can use 2x growth, rcvwin = 4 * copied
578 tp
->rcvq_space
.space
+ (tp
->rcvq_space
.space
>> 2)) {
580 tp
->rcvq_space
.space
+ (tp
->rcvq_space
.space
>> 1))
583 rcvwin
+= (rcvwin
>> 1);
586 rcvmem
= SKB_TRUESIZE(tp
->advmss
+ MAX_TCP_HEADER
);
587 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
590 rcvbuf
= min(rcvwin
/ tp
->advmss
* rcvmem
, sysctl_tcp_rmem
[2]);
591 if (rcvbuf
> sk
->sk_rcvbuf
) {
592 sk
->sk_rcvbuf
= rcvbuf
;
594 /* Make the window clamp follow along. */
595 tp
->window_clamp
= rcvwin
;
598 tp
->rcvq_space
.space
= copied
;
601 tp
->rcvq_space
.seq
= tp
->copied_seq
;
602 tp
->rcvq_space
.time
= tcp_time_stamp
;
605 /* There is something which you must keep in mind when you analyze the
606 * behavior of the tp->ato delayed ack timeout interval. When a
607 * connection starts up, we want to ack as quickly as possible. The
608 * problem is that "good" TCP's do slow start at the beginning of data
609 * transmission. The means that until we send the first few ACK's the
610 * sender will sit on his end and only queue most of his data, because
611 * he can only send snd_cwnd unacked packets at any given time. For
612 * each ACK we send, he increments snd_cwnd and transmits more of his
615 static void tcp_event_data_recv(struct sock
*sk
, struct sk_buff
*skb
)
617 struct tcp_sock
*tp
= tcp_sk(sk
);
618 struct inet_connection_sock
*icsk
= inet_csk(sk
);
621 inet_csk_schedule_ack(sk
);
623 tcp_measure_rcv_mss(sk
, skb
);
625 tcp_rcv_rtt_measure(tp
);
627 now
= tcp_time_stamp
;
629 if (!icsk
->icsk_ack
.ato
) {
630 /* The _first_ data packet received, initialize
631 * delayed ACK engine.
633 tcp_incr_quickack(sk
);
634 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
636 int m
= now
- icsk
->icsk_ack
.lrcvtime
;
638 if (m
<= TCP_ATO_MIN
/ 2) {
639 /* The fastest case is the first. */
640 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + TCP_ATO_MIN
/ 2;
641 } else if (m
< icsk
->icsk_ack
.ato
) {
642 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + m
;
643 if (icsk
->icsk_ack
.ato
> icsk
->icsk_rto
)
644 icsk
->icsk_ack
.ato
= icsk
->icsk_rto
;
645 } else if (m
> icsk
->icsk_rto
) {
646 /* Too long gap. Apparently sender failed to
647 * restart window, so that we send ACKs quickly.
649 tcp_incr_quickack(sk
);
653 icsk
->icsk_ack
.lrcvtime
= now
;
655 TCP_ECN_check_ce(tp
, skb
);
658 tcp_grow_window(sk
, skb
);
661 /* Called to compute a smoothed rtt estimate. The data fed to this
662 * routine either comes from timestamps, or from segments that were
663 * known _not_ to have been retransmitted [see Karn/Partridge
664 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
665 * piece by Van Jacobson.
666 * NOTE: the next three routines used to be one big routine.
667 * To save cycles in the RFC 1323 implementation it was better to break
668 * it up into three procedures. -- erics
670 static void tcp_rtt_estimator(struct sock
*sk
, const __u32 mrtt
)
672 struct tcp_sock
*tp
= tcp_sk(sk
);
673 long m
= mrtt
; /* RTT */
675 /* The following amusing code comes from Jacobson's
676 * article in SIGCOMM '88. Note that rtt and mdev
677 * are scaled versions of rtt and mean deviation.
678 * This is designed to be as fast as possible
679 * m stands for "measurement".
681 * On a 1990 paper the rto value is changed to:
682 * RTO = rtt + 4 * mdev
684 * Funny. This algorithm seems to be very broken.
685 * These formulae increase RTO, when it should be decreased, increase
686 * too slowly, when it should be increased quickly, decrease too quickly
687 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
688 * does not matter how to _calculate_ it. Seems, it was trap
689 * that VJ failed to avoid. 8)
694 m
-= (tp
->srtt
>> 3); /* m is now error in rtt est */
695 tp
->srtt
+= m
; /* rtt = 7/8 rtt + 1/8 new */
697 m
= -m
; /* m is now abs(error) */
698 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
699 /* This is similar to one of Eifel findings.
700 * Eifel blocks mdev updates when rtt decreases.
701 * This solution is a bit different: we use finer gain
702 * for mdev in this case (alpha*beta).
703 * Like Eifel it also prevents growth of rto,
704 * but also it limits too fast rto decreases,
705 * happening in pure Eifel.
710 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
712 tp
->mdev
+= m
; /* mdev = 3/4 mdev + 1/4 new */
713 if (tp
->mdev
> tp
->mdev_max
) {
714 tp
->mdev_max
= tp
->mdev
;
715 if (tp
->mdev_max
> tp
->rttvar
)
716 tp
->rttvar
= tp
->mdev_max
;
718 if (after(tp
->snd_una
, tp
->rtt_seq
)) {
719 if (tp
->mdev_max
< tp
->rttvar
)
720 tp
->rttvar
-= (tp
->rttvar
- tp
->mdev_max
) >> 2;
721 tp
->rtt_seq
= tp
->snd_nxt
;
722 tp
->mdev_max
= tcp_rto_min(sk
);
725 /* no previous measure. */
726 tp
->srtt
= m
<< 3; /* take the measured time to be rtt */
727 tp
->mdev
= m
<< 1; /* make sure rto = 3*rtt */
728 tp
->mdev_max
= tp
->rttvar
= max(tp
->mdev
, tcp_rto_min(sk
));
729 tp
->rtt_seq
= tp
->snd_nxt
;
733 /* Set the sk_pacing_rate to allow proper sizing of TSO packets.
734 * Note: TCP stack does not yet implement pacing.
735 * FQ packet scheduler can be used to implement cheap but effective
736 * TCP pacing, to smooth the burst on large writes when packets
737 * in flight is significantly lower than cwnd (or rwin)
739 static void tcp_update_pacing_rate(struct sock
*sk
)
741 const struct tcp_sock
*tp
= tcp_sk(sk
);
744 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
745 rate
= (u64
)tp
->mss_cache
* 2 * (HZ
<< 3);
747 rate
*= max(tp
->snd_cwnd
, tp
->packets_out
);
749 /* Correction for small srtt : minimum srtt being 8 (1 jiffy << 3),
750 * be conservative and assume srtt = 1 (125 us instead of 1.25 ms)
751 * We probably need usec resolution in the future.
752 * Note: This also takes care of possible srtt=0 case,
753 * when tcp_rtt_estimator() was not yet called.
755 if (tp
->srtt
> 8 + 2)
756 do_div(rate
, tp
->srtt
);
758 /* ACCESS_ONCE() is needed because sch_fq fetches sk_pacing_rate
759 * without any lock. We want to make sure compiler wont store
760 * intermediate values in this location.
762 ACCESS_ONCE(sk
->sk_pacing_rate
) = min_t(u64
, rate
,
763 sk
->sk_max_pacing_rate
);
766 /* Calculate rto without backoff. This is the second half of Van Jacobson's
767 * routine referred to above.
769 static void tcp_set_rto(struct sock
*sk
)
771 const struct tcp_sock
*tp
= tcp_sk(sk
);
772 /* Old crap is replaced with new one. 8)
775 * 1. If rtt variance happened to be less 50msec, it is hallucination.
776 * It cannot be less due to utterly erratic ACK generation made
777 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
778 * to do with delayed acks, because at cwnd>2 true delack timeout
779 * is invisible. Actually, Linux-2.4 also generates erratic
780 * ACKs in some circumstances.
782 inet_csk(sk
)->icsk_rto
= __tcp_set_rto(tp
);
784 /* 2. Fixups made earlier cannot be right.
785 * If we do not estimate RTO correctly without them,
786 * all the algo is pure shit and should be replaced
787 * with correct one. It is exactly, which we pretend to do.
790 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
791 * guarantees that rto is higher.
796 __u32
tcp_init_cwnd(const struct tcp_sock
*tp
, const struct dst_entry
*dst
)
798 __u32 cwnd
= (dst
? dst_metric(dst
, RTAX_INITCWND
) : 0);
801 cwnd
= TCP_INIT_CWND
;
802 return min_t(__u32
, cwnd
, tp
->snd_cwnd_clamp
);
806 * Packet counting of FACK is based on in-order assumptions, therefore TCP
807 * disables it when reordering is detected
809 void tcp_disable_fack(struct tcp_sock
*tp
)
811 /* RFC3517 uses different metric in lost marker => reset on change */
813 tp
->lost_skb_hint
= NULL
;
814 tp
->rx_opt
.sack_ok
&= ~TCP_FACK_ENABLED
;
817 /* Take a notice that peer is sending D-SACKs */
818 static void tcp_dsack_seen(struct tcp_sock
*tp
)
820 tp
->rx_opt
.sack_ok
|= TCP_DSACK_SEEN
;
823 static void tcp_update_reordering(struct sock
*sk
, const int metric
,
826 struct tcp_sock
*tp
= tcp_sk(sk
);
827 if (metric
> tp
->reordering
) {
830 tp
->reordering
= min(TCP_MAX_REORDERING
, metric
);
832 /* This exciting event is worth to be remembered. 8) */
834 mib_idx
= LINUX_MIB_TCPTSREORDER
;
835 else if (tcp_is_reno(tp
))
836 mib_idx
= LINUX_MIB_TCPRENOREORDER
;
837 else if (tcp_is_fack(tp
))
838 mib_idx
= LINUX_MIB_TCPFACKREORDER
;
840 mib_idx
= LINUX_MIB_TCPSACKREORDER
;
842 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
843 #if FASTRETRANS_DEBUG > 1
844 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
845 tp
->rx_opt
.sack_ok
, inet_csk(sk
)->icsk_ca_state
,
849 tp
->undo_marker
? tp
->undo_retrans
: 0);
851 tcp_disable_fack(tp
);
855 tcp_disable_early_retrans(tp
);
858 /* This must be called before lost_out is incremented */
859 static void tcp_verify_retransmit_hint(struct tcp_sock
*tp
, struct sk_buff
*skb
)
861 if ((tp
->retransmit_skb_hint
== NULL
) ||
862 before(TCP_SKB_CB(skb
)->seq
,
863 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
864 tp
->retransmit_skb_hint
= skb
;
867 after(TCP_SKB_CB(skb
)->end_seq
, tp
->retransmit_high
))
868 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
871 static void tcp_skb_mark_lost(struct tcp_sock
*tp
, struct sk_buff
*skb
)
873 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
874 tcp_verify_retransmit_hint(tp
, skb
);
876 tp
->lost_out
+= tcp_skb_pcount(skb
);
877 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
881 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock
*tp
,
884 tcp_verify_retransmit_hint(tp
, skb
);
886 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
887 tp
->lost_out
+= tcp_skb_pcount(skb
);
888 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
892 /* This procedure tags the retransmission queue when SACKs arrive.
894 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
895 * Packets in queue with these bits set are counted in variables
896 * sacked_out, retrans_out and lost_out, correspondingly.
898 * Valid combinations are:
899 * Tag InFlight Description
900 * 0 1 - orig segment is in flight.
901 * S 0 - nothing flies, orig reached receiver.
902 * L 0 - nothing flies, orig lost by net.
903 * R 2 - both orig and retransmit are in flight.
904 * L|R 1 - orig is lost, retransmit is in flight.
905 * S|R 1 - orig reached receiver, retrans is still in flight.
906 * (L|S|R is logically valid, it could occur when L|R is sacked,
907 * but it is equivalent to plain S and code short-curcuits it to S.
908 * L|S is logically invalid, it would mean -1 packet in flight 8))
910 * These 6 states form finite state machine, controlled by the following events:
911 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
912 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
913 * 3. Loss detection event of two flavors:
914 * A. Scoreboard estimator decided the packet is lost.
915 * A'. Reno "three dupacks" marks head of queue lost.
916 * A''. Its FACK modification, head until snd.fack is lost.
917 * B. SACK arrives sacking SND.NXT at the moment, when the
918 * segment was retransmitted.
919 * 4. D-SACK added new rule: D-SACK changes any tag to S.
921 * It is pleasant to note, that state diagram turns out to be commutative,
922 * so that we are allowed not to be bothered by order of our actions,
923 * when multiple events arrive simultaneously. (see the function below).
925 * Reordering detection.
926 * --------------------
927 * Reordering metric is maximal distance, which a packet can be displaced
928 * in packet stream. With SACKs we can estimate it:
930 * 1. SACK fills old hole and the corresponding segment was not
931 * ever retransmitted -> reordering. Alas, we cannot use it
932 * when segment was retransmitted.
933 * 2. The last flaw is solved with D-SACK. D-SACK arrives
934 * for retransmitted and already SACKed segment -> reordering..
935 * Both of these heuristics are not used in Loss state, when we cannot
936 * account for retransmits accurately.
938 * SACK block validation.
939 * ----------------------
941 * SACK block range validation checks that the received SACK block fits to
942 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
943 * Note that SND.UNA is not included to the range though being valid because
944 * it means that the receiver is rather inconsistent with itself reporting
945 * SACK reneging when it should advance SND.UNA. Such SACK block this is
946 * perfectly valid, however, in light of RFC2018 which explicitly states
947 * that "SACK block MUST reflect the newest segment. Even if the newest
948 * segment is going to be discarded ...", not that it looks very clever
949 * in case of head skb. Due to potentional receiver driven attacks, we
950 * choose to avoid immediate execution of a walk in write queue due to
951 * reneging and defer head skb's loss recovery to standard loss recovery
952 * procedure that will eventually trigger (nothing forbids us doing this).
954 * Implements also blockage to start_seq wrap-around. Problem lies in the
955 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
956 * there's no guarantee that it will be before snd_nxt (n). The problem
957 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
960 * <- outs wnd -> <- wrapzone ->
961 * u e n u_w e_w s n_w
963 * |<------------+------+----- TCP seqno space --------------+---------->|
964 * ...-- <2^31 ->| |<--------...
965 * ...---- >2^31 ------>| |<--------...
967 * Current code wouldn't be vulnerable but it's better still to discard such
968 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
969 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
970 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
971 * equal to the ideal case (infinite seqno space without wrap caused issues).
973 * With D-SACK the lower bound is extended to cover sequence space below
974 * SND.UNA down to undo_marker, which is the last point of interest. Yet
975 * again, D-SACK block must not to go across snd_una (for the same reason as
976 * for the normal SACK blocks, explained above). But there all simplicity
977 * ends, TCP might receive valid D-SACKs below that. As long as they reside
978 * fully below undo_marker they do not affect behavior in anyway and can
979 * therefore be safely ignored. In rare cases (which are more or less
980 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
981 * fragmentation and packet reordering past skb's retransmission. To consider
982 * them correctly, the acceptable range must be extended even more though
983 * the exact amount is rather hard to quantify. However, tp->max_window can
984 * be used as an exaggerated estimate.
986 static bool tcp_is_sackblock_valid(struct tcp_sock
*tp
, bool is_dsack
,
987 u32 start_seq
, u32 end_seq
)
989 /* Too far in future, or reversed (interpretation is ambiguous) */
990 if (after(end_seq
, tp
->snd_nxt
) || !before(start_seq
, end_seq
))
993 /* Nasty start_seq wrap-around check (see comments above) */
994 if (!before(start_seq
, tp
->snd_nxt
))
997 /* In outstanding window? ...This is valid exit for D-SACKs too.
998 * start_seq == snd_una is non-sensical (see comments above)
1000 if (after(start_seq
, tp
->snd_una
))
1003 if (!is_dsack
|| !tp
->undo_marker
)
1006 /* ...Then it's D-SACK, and must reside below snd_una completely */
1007 if (after(end_seq
, tp
->snd_una
))
1010 if (!before(start_seq
, tp
->undo_marker
))
1014 if (!after(end_seq
, tp
->undo_marker
))
1017 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1018 * start_seq < undo_marker and end_seq >= undo_marker.
1020 return !before(start_seq
, end_seq
- tp
->max_window
);
1023 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1024 * Event "B". Later note: FACK people cheated me again 8), we have to account
1025 * for reordering! Ugly, but should help.
1027 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1028 * less than what is now known to be received by the other end (derived from
1029 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1030 * retransmitted skbs to avoid some costly processing per ACKs.
1032 static void tcp_mark_lost_retrans(struct sock
*sk
)
1034 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1035 struct tcp_sock
*tp
= tcp_sk(sk
);
1036 struct sk_buff
*skb
;
1038 u32 new_low_seq
= tp
->snd_nxt
;
1039 u32 received_upto
= tcp_highest_sack_seq(tp
);
1041 if (!tcp_is_fack(tp
) || !tp
->retrans_out
||
1042 !after(received_upto
, tp
->lost_retrans_low
) ||
1043 icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
1046 tcp_for_write_queue(skb
, sk
) {
1047 u32 ack_seq
= TCP_SKB_CB(skb
)->ack_seq
;
1049 if (skb
== tcp_send_head(sk
))
1051 if (cnt
== tp
->retrans_out
)
1053 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1056 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
))
1059 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1060 * constraint here (see above) but figuring out that at
1061 * least tp->reordering SACK blocks reside between ack_seq
1062 * and received_upto is not easy task to do cheaply with
1063 * the available datastructures.
1065 * Whether FACK should check here for tp->reordering segs
1066 * in-between one could argue for either way (it would be
1067 * rather simple to implement as we could count fack_count
1068 * during the walk and do tp->fackets_out - fack_count).
1070 if (after(received_upto
, ack_seq
)) {
1071 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1072 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1074 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
1075 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSTRETRANSMIT
);
1077 if (before(ack_seq
, new_low_seq
))
1078 new_low_seq
= ack_seq
;
1079 cnt
+= tcp_skb_pcount(skb
);
1083 if (tp
->retrans_out
)
1084 tp
->lost_retrans_low
= new_low_seq
;
1087 static bool tcp_check_dsack(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1088 struct tcp_sack_block_wire
*sp
, int num_sacks
,
1091 struct tcp_sock
*tp
= tcp_sk(sk
);
1092 u32 start_seq_0
= get_unaligned_be32(&sp
[0].start_seq
);
1093 u32 end_seq_0
= get_unaligned_be32(&sp
[0].end_seq
);
1094 bool dup_sack
= false;
1096 if (before(start_seq_0
, TCP_SKB_CB(ack_skb
)->ack_seq
)) {
1099 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPDSACKRECV
);
1100 } else if (num_sacks
> 1) {
1101 u32 end_seq_1
= get_unaligned_be32(&sp
[1].end_seq
);
1102 u32 start_seq_1
= get_unaligned_be32(&sp
[1].start_seq
);
1104 if (!after(end_seq_0
, end_seq_1
) &&
1105 !before(start_seq_0
, start_seq_1
)) {
1108 NET_INC_STATS_BH(sock_net(sk
),
1109 LINUX_MIB_TCPDSACKOFORECV
);
1113 /* D-SACK for already forgotten data... Do dumb counting. */
1114 if (dup_sack
&& tp
->undo_marker
&& tp
->undo_retrans
&&
1115 !after(end_seq_0
, prior_snd_una
) &&
1116 after(end_seq_0
, tp
->undo_marker
))
1122 struct tcp_sacktag_state
{
1126 s32 rtt
; /* RTT measured by SACKing never-retransmitted data */
1129 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1130 * the incoming SACK may not exactly match but we can find smaller MSS
1131 * aligned portion of it that matches. Therefore we might need to fragment
1132 * which may fail and creates some hassle (caller must handle error case
1135 * FIXME: this could be merged to shift decision code
1137 static int tcp_match_skb_to_sack(struct sock
*sk
, struct sk_buff
*skb
,
1138 u32 start_seq
, u32 end_seq
)
1142 unsigned int pkt_len
;
1145 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1146 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1148 if (tcp_skb_pcount(skb
) > 1 && !in_sack
&&
1149 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
1150 mss
= tcp_skb_mss(skb
);
1151 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1154 pkt_len
= start_seq
- TCP_SKB_CB(skb
)->seq
;
1158 pkt_len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1163 /* Round if necessary so that SACKs cover only full MSSes
1164 * and/or the remaining small portion (if present)
1166 if (pkt_len
> mss
) {
1167 unsigned int new_len
= (pkt_len
/ mss
) * mss
;
1168 if (!in_sack
&& new_len
< pkt_len
) {
1170 if (new_len
> skb
->len
)
1175 err
= tcp_fragment(sk
, skb
, pkt_len
, mss
);
1183 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1184 static u8
tcp_sacktag_one(struct sock
*sk
,
1185 struct tcp_sacktag_state
*state
, u8 sacked
,
1186 u32 start_seq
, u32 end_seq
,
1187 int dup_sack
, int pcount
, u32 xmit_time
)
1189 struct tcp_sock
*tp
= tcp_sk(sk
);
1190 int fack_count
= state
->fack_count
;
1192 /* Account D-SACK for retransmitted packet. */
1193 if (dup_sack
&& (sacked
& TCPCB_RETRANS
)) {
1194 if (tp
->undo_marker
&& tp
->undo_retrans
&&
1195 after(end_seq
, tp
->undo_marker
))
1197 if (sacked
& TCPCB_SACKED_ACKED
)
1198 state
->reord
= min(fack_count
, state
->reord
);
1201 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1202 if (!after(end_seq
, tp
->snd_una
))
1205 if (!(sacked
& TCPCB_SACKED_ACKED
)) {
1206 if (sacked
& TCPCB_SACKED_RETRANS
) {
1207 /* If the segment is not tagged as lost,
1208 * we do not clear RETRANS, believing
1209 * that retransmission is still in flight.
1211 if (sacked
& TCPCB_LOST
) {
1212 sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1213 tp
->lost_out
-= pcount
;
1214 tp
->retrans_out
-= pcount
;
1217 if (!(sacked
& TCPCB_RETRANS
)) {
1218 /* New sack for not retransmitted frame,
1219 * which was in hole. It is reordering.
1221 if (before(start_seq
,
1222 tcp_highest_sack_seq(tp
)))
1223 state
->reord
= min(fack_count
,
1225 if (!after(end_seq
, tp
->high_seq
))
1226 state
->flag
|= FLAG_ORIG_SACK_ACKED
;
1227 /* Pick the earliest sequence sacked for RTT */
1229 state
->rtt
= tcp_time_stamp
- xmit_time
;
1232 if (sacked
& TCPCB_LOST
) {
1233 sacked
&= ~TCPCB_LOST
;
1234 tp
->lost_out
-= pcount
;
1238 sacked
|= TCPCB_SACKED_ACKED
;
1239 state
->flag
|= FLAG_DATA_SACKED
;
1240 tp
->sacked_out
+= pcount
;
1242 fack_count
+= pcount
;
1244 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1245 if (!tcp_is_fack(tp
) && (tp
->lost_skb_hint
!= NULL
) &&
1246 before(start_seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
))
1247 tp
->lost_cnt_hint
+= pcount
;
1249 if (fack_count
> tp
->fackets_out
)
1250 tp
->fackets_out
= fack_count
;
1253 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1254 * frames and clear it. undo_retrans is decreased above, L|R frames
1255 * are accounted above as well.
1257 if (dup_sack
&& (sacked
& TCPCB_SACKED_RETRANS
)) {
1258 sacked
&= ~TCPCB_SACKED_RETRANS
;
1259 tp
->retrans_out
-= pcount
;
1265 /* Shift newly-SACKed bytes from this skb to the immediately previous
1266 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1268 static bool tcp_shifted_skb(struct sock
*sk
, struct sk_buff
*skb
,
1269 struct tcp_sacktag_state
*state
,
1270 unsigned int pcount
, int shifted
, int mss
,
1273 struct tcp_sock
*tp
= tcp_sk(sk
);
1274 struct sk_buff
*prev
= tcp_write_queue_prev(sk
, skb
);
1275 u32 start_seq
= TCP_SKB_CB(skb
)->seq
; /* start of newly-SACKed */
1276 u32 end_seq
= start_seq
+ shifted
; /* end of newly-SACKed */
1280 /* Adjust counters and hints for the newly sacked sequence
1281 * range but discard the return value since prev is already
1282 * marked. We must tag the range first because the seq
1283 * advancement below implicitly advances
1284 * tcp_highest_sack_seq() when skb is highest_sack.
1286 tcp_sacktag_one(sk
, state
, TCP_SKB_CB(skb
)->sacked
,
1287 start_seq
, end_seq
, dup_sack
, pcount
,
1288 TCP_SKB_CB(skb
)->when
);
1290 if (skb
== tp
->lost_skb_hint
)
1291 tp
->lost_cnt_hint
+= pcount
;
1293 TCP_SKB_CB(prev
)->end_seq
+= shifted
;
1294 TCP_SKB_CB(skb
)->seq
+= shifted
;
1296 skb_shinfo(prev
)->gso_segs
+= pcount
;
1297 BUG_ON(skb_shinfo(skb
)->gso_segs
< pcount
);
1298 skb_shinfo(skb
)->gso_segs
-= pcount
;
1300 /* When we're adding to gso_segs == 1, gso_size will be zero,
1301 * in theory this shouldn't be necessary but as long as DSACK
1302 * code can come after this skb later on it's better to keep
1303 * setting gso_size to something.
1305 if (!skb_shinfo(prev
)->gso_size
) {
1306 skb_shinfo(prev
)->gso_size
= mss
;
1307 skb_shinfo(prev
)->gso_type
= sk
->sk_gso_type
;
1310 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1311 if (skb_shinfo(skb
)->gso_segs
<= 1) {
1312 skb_shinfo(skb
)->gso_size
= 0;
1313 skb_shinfo(skb
)->gso_type
= 0;
1316 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1317 TCP_SKB_CB(prev
)->sacked
|= (TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
);
1320 BUG_ON(!tcp_skb_pcount(skb
));
1321 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKSHIFTED
);
1325 /* Whole SKB was eaten :-) */
1327 if (skb
== tp
->retransmit_skb_hint
)
1328 tp
->retransmit_skb_hint
= prev
;
1329 if (skb
== tp
->lost_skb_hint
) {
1330 tp
->lost_skb_hint
= prev
;
1331 tp
->lost_cnt_hint
-= tcp_skb_pcount(prev
);
1334 TCP_SKB_CB(prev
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
;
1335 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
1336 TCP_SKB_CB(prev
)->end_seq
++;
1338 if (skb
== tcp_highest_sack(sk
))
1339 tcp_advance_highest_sack(sk
, skb
);
1341 tcp_unlink_write_queue(skb
, sk
);
1342 sk_wmem_free_skb(sk
, skb
);
1344 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKMERGED
);
1349 /* I wish gso_size would have a bit more sane initialization than
1350 * something-or-zero which complicates things
1352 static int tcp_skb_seglen(const struct sk_buff
*skb
)
1354 return tcp_skb_pcount(skb
) == 1 ? skb
->len
: tcp_skb_mss(skb
);
1357 /* Shifting pages past head area doesn't work */
1358 static int skb_can_shift(const struct sk_buff
*skb
)
1360 return !skb_headlen(skb
) && skb_is_nonlinear(skb
);
1363 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1366 static struct sk_buff
*tcp_shift_skb_data(struct sock
*sk
, struct sk_buff
*skb
,
1367 struct tcp_sacktag_state
*state
,
1368 u32 start_seq
, u32 end_seq
,
1371 struct tcp_sock
*tp
= tcp_sk(sk
);
1372 struct sk_buff
*prev
;
1378 if (!sk_can_gso(sk
))
1381 /* Normally R but no L won't result in plain S */
1383 (TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_RETRANS
)) == TCPCB_SACKED_RETRANS
)
1385 if (!skb_can_shift(skb
))
1387 /* This frame is about to be dropped (was ACKed). */
1388 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1391 /* Can only happen with delayed DSACK + discard craziness */
1392 if (unlikely(skb
== tcp_write_queue_head(sk
)))
1394 prev
= tcp_write_queue_prev(sk
, skb
);
1396 if ((TCP_SKB_CB(prev
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
)
1399 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1400 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1404 pcount
= tcp_skb_pcount(skb
);
1405 mss
= tcp_skb_seglen(skb
);
1407 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1408 * drop this restriction as unnecessary
1410 if (mss
!= tcp_skb_seglen(prev
))
1413 if (!after(TCP_SKB_CB(skb
)->end_seq
, start_seq
))
1415 /* CHECKME: This is non-MSS split case only?, this will
1416 * cause skipped skbs due to advancing loop btw, original
1417 * has that feature too
1419 if (tcp_skb_pcount(skb
) <= 1)
1422 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1424 /* TODO: head merge to next could be attempted here
1425 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1426 * though it might not be worth of the additional hassle
1428 * ...we can probably just fallback to what was done
1429 * previously. We could try merging non-SACKed ones
1430 * as well but it probably isn't going to buy off
1431 * because later SACKs might again split them, and
1432 * it would make skb timestamp tracking considerably
1438 len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1440 BUG_ON(len
> skb
->len
);
1442 /* MSS boundaries should be honoured or else pcount will
1443 * severely break even though it makes things bit trickier.
1444 * Optimize common case to avoid most of the divides
1446 mss
= tcp_skb_mss(skb
);
1448 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1449 * drop this restriction as unnecessary
1451 if (mss
!= tcp_skb_seglen(prev
))
1456 } else if (len
< mss
) {
1464 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1465 if (!after(TCP_SKB_CB(skb
)->seq
+ len
, tp
->snd_una
))
1468 if (!skb_shift(prev
, skb
, len
))
1470 if (!tcp_shifted_skb(sk
, skb
, state
, pcount
, len
, mss
, dup_sack
))
1473 /* Hole filled allows collapsing with the next as well, this is very
1474 * useful when hole on every nth skb pattern happens
1476 if (prev
== tcp_write_queue_tail(sk
))
1478 skb
= tcp_write_queue_next(sk
, prev
);
1480 if (!skb_can_shift(skb
) ||
1481 (skb
== tcp_send_head(sk
)) ||
1482 ((TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
) ||
1483 (mss
!= tcp_skb_seglen(skb
)))
1487 if (skb_shift(prev
, skb
, len
)) {
1488 pcount
+= tcp_skb_pcount(skb
);
1489 tcp_shifted_skb(sk
, skb
, state
, tcp_skb_pcount(skb
), len
, mss
, 0);
1493 state
->fack_count
+= pcount
;
1500 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKSHIFTFALLBACK
);
1504 static struct sk_buff
*tcp_sacktag_walk(struct sk_buff
*skb
, struct sock
*sk
,
1505 struct tcp_sack_block
*next_dup
,
1506 struct tcp_sacktag_state
*state
,
1507 u32 start_seq
, u32 end_seq
,
1510 struct tcp_sock
*tp
= tcp_sk(sk
);
1511 struct sk_buff
*tmp
;
1513 tcp_for_write_queue_from(skb
, sk
) {
1515 bool dup_sack
= dup_sack_in
;
1517 if (skb
== tcp_send_head(sk
))
1520 /* queue is in-order => we can short-circuit the walk early */
1521 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1524 if ((next_dup
!= NULL
) &&
1525 before(TCP_SKB_CB(skb
)->seq
, next_dup
->end_seq
)) {
1526 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1527 next_dup
->start_seq
,
1533 /* skb reference here is a bit tricky to get right, since
1534 * shifting can eat and free both this skb and the next,
1535 * so not even _safe variant of the loop is enough.
1538 tmp
= tcp_shift_skb_data(sk
, skb
, state
,
1539 start_seq
, end_seq
, dup_sack
);
1548 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1554 if (unlikely(in_sack
< 0))
1558 TCP_SKB_CB(skb
)->sacked
=
1561 TCP_SKB_CB(skb
)->sacked
,
1562 TCP_SKB_CB(skb
)->seq
,
1563 TCP_SKB_CB(skb
)->end_seq
,
1565 tcp_skb_pcount(skb
),
1566 TCP_SKB_CB(skb
)->when
);
1568 if (!before(TCP_SKB_CB(skb
)->seq
,
1569 tcp_highest_sack_seq(tp
)))
1570 tcp_advance_highest_sack(sk
, skb
);
1573 state
->fack_count
+= tcp_skb_pcount(skb
);
1578 /* Avoid all extra work that is being done by sacktag while walking in
1581 static struct sk_buff
*tcp_sacktag_skip(struct sk_buff
*skb
, struct sock
*sk
,
1582 struct tcp_sacktag_state
*state
,
1585 tcp_for_write_queue_from(skb
, sk
) {
1586 if (skb
== tcp_send_head(sk
))
1589 if (after(TCP_SKB_CB(skb
)->end_seq
, skip_to_seq
))
1592 state
->fack_count
+= tcp_skb_pcount(skb
);
1597 static struct sk_buff
*tcp_maybe_skipping_dsack(struct sk_buff
*skb
,
1599 struct tcp_sack_block
*next_dup
,
1600 struct tcp_sacktag_state
*state
,
1603 if (next_dup
== NULL
)
1606 if (before(next_dup
->start_seq
, skip_to_seq
)) {
1607 skb
= tcp_sacktag_skip(skb
, sk
, state
, next_dup
->start_seq
);
1608 skb
= tcp_sacktag_walk(skb
, sk
, NULL
, state
,
1609 next_dup
->start_seq
, next_dup
->end_seq
,
1616 static int tcp_sack_cache_ok(const struct tcp_sock
*tp
, const struct tcp_sack_block
*cache
)
1618 return cache
< tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1622 tcp_sacktag_write_queue(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1623 u32 prior_snd_una
, s32
*sack_rtt
)
1625 struct tcp_sock
*tp
= tcp_sk(sk
);
1626 const unsigned char *ptr
= (skb_transport_header(ack_skb
) +
1627 TCP_SKB_CB(ack_skb
)->sacked
);
1628 struct tcp_sack_block_wire
*sp_wire
= (struct tcp_sack_block_wire
*)(ptr
+2);
1629 struct tcp_sack_block sp
[TCP_NUM_SACKS
];
1630 struct tcp_sack_block
*cache
;
1631 struct tcp_sacktag_state state
;
1632 struct sk_buff
*skb
;
1633 int num_sacks
= min(TCP_NUM_SACKS
, (ptr
[1] - TCPOLEN_SACK_BASE
) >> 3);
1635 bool found_dup_sack
= false;
1637 int first_sack_index
;
1640 state
.reord
= tp
->packets_out
;
1643 if (!tp
->sacked_out
) {
1644 if (WARN_ON(tp
->fackets_out
))
1645 tp
->fackets_out
= 0;
1646 tcp_highest_sack_reset(sk
);
1649 found_dup_sack
= tcp_check_dsack(sk
, ack_skb
, sp_wire
,
1650 num_sacks
, prior_snd_una
);
1652 state
.flag
|= FLAG_DSACKING_ACK
;
1654 /* Eliminate too old ACKs, but take into
1655 * account more or less fresh ones, they can
1656 * contain valid SACK info.
1658 if (before(TCP_SKB_CB(ack_skb
)->ack_seq
, prior_snd_una
- tp
->max_window
))
1661 if (!tp
->packets_out
)
1665 first_sack_index
= 0;
1666 for (i
= 0; i
< num_sacks
; i
++) {
1667 bool dup_sack
= !i
&& found_dup_sack
;
1669 sp
[used_sacks
].start_seq
= get_unaligned_be32(&sp_wire
[i
].start_seq
);
1670 sp
[used_sacks
].end_seq
= get_unaligned_be32(&sp_wire
[i
].end_seq
);
1672 if (!tcp_is_sackblock_valid(tp
, dup_sack
,
1673 sp
[used_sacks
].start_seq
,
1674 sp
[used_sacks
].end_seq
)) {
1678 if (!tp
->undo_marker
)
1679 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDNOUNDO
;
1681 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDOLD
;
1683 /* Don't count olds caused by ACK reordering */
1684 if ((TCP_SKB_CB(ack_skb
)->ack_seq
!= tp
->snd_una
) &&
1685 !after(sp
[used_sacks
].end_seq
, tp
->snd_una
))
1687 mib_idx
= LINUX_MIB_TCPSACKDISCARD
;
1690 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
1692 first_sack_index
= -1;
1696 /* Ignore very old stuff early */
1697 if (!after(sp
[used_sacks
].end_seq
, prior_snd_una
))
1703 /* order SACK blocks to allow in order walk of the retrans queue */
1704 for (i
= used_sacks
- 1; i
> 0; i
--) {
1705 for (j
= 0; j
< i
; j
++) {
1706 if (after(sp
[j
].start_seq
, sp
[j
+ 1].start_seq
)) {
1707 swap(sp
[j
], sp
[j
+ 1]);
1709 /* Track where the first SACK block goes to */
1710 if (j
== first_sack_index
)
1711 first_sack_index
= j
+ 1;
1716 skb
= tcp_write_queue_head(sk
);
1717 state
.fack_count
= 0;
1720 if (!tp
->sacked_out
) {
1721 /* It's already past, so skip checking against it */
1722 cache
= tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1724 cache
= tp
->recv_sack_cache
;
1725 /* Skip empty blocks in at head of the cache */
1726 while (tcp_sack_cache_ok(tp
, cache
) && !cache
->start_seq
&&
1731 while (i
< used_sacks
) {
1732 u32 start_seq
= sp
[i
].start_seq
;
1733 u32 end_seq
= sp
[i
].end_seq
;
1734 bool dup_sack
= (found_dup_sack
&& (i
== first_sack_index
));
1735 struct tcp_sack_block
*next_dup
= NULL
;
1737 if (found_dup_sack
&& ((i
+ 1) == first_sack_index
))
1738 next_dup
= &sp
[i
+ 1];
1740 /* Skip too early cached blocks */
1741 while (tcp_sack_cache_ok(tp
, cache
) &&
1742 !before(start_seq
, cache
->end_seq
))
1745 /* Can skip some work by looking recv_sack_cache? */
1746 if (tcp_sack_cache_ok(tp
, cache
) && !dup_sack
&&
1747 after(end_seq
, cache
->start_seq
)) {
1750 if (before(start_seq
, cache
->start_seq
)) {
1751 skb
= tcp_sacktag_skip(skb
, sk
, &state
,
1753 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
,
1760 /* Rest of the block already fully processed? */
1761 if (!after(end_seq
, cache
->end_seq
))
1764 skb
= tcp_maybe_skipping_dsack(skb
, sk
, next_dup
,
1768 /* ...tail remains todo... */
1769 if (tcp_highest_sack_seq(tp
) == cache
->end_seq
) {
1770 /* ...but better entrypoint exists! */
1771 skb
= tcp_highest_sack(sk
);
1774 state
.fack_count
= tp
->fackets_out
;
1779 skb
= tcp_sacktag_skip(skb
, sk
, &state
, cache
->end_seq
);
1780 /* Check overlap against next cached too (past this one already) */
1785 if (!before(start_seq
, tcp_highest_sack_seq(tp
))) {
1786 skb
= tcp_highest_sack(sk
);
1789 state
.fack_count
= tp
->fackets_out
;
1791 skb
= tcp_sacktag_skip(skb
, sk
, &state
, start_seq
);
1794 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
, &state
,
1795 start_seq
, end_seq
, dup_sack
);
1801 /* Clear the head of the cache sack blocks so we can skip it next time */
1802 for (i
= 0; i
< ARRAY_SIZE(tp
->recv_sack_cache
) - used_sacks
; i
++) {
1803 tp
->recv_sack_cache
[i
].start_seq
= 0;
1804 tp
->recv_sack_cache
[i
].end_seq
= 0;
1806 for (j
= 0; j
< used_sacks
; j
++)
1807 tp
->recv_sack_cache
[i
++] = sp
[j
];
1809 tcp_mark_lost_retrans(sk
);
1811 tcp_verify_left_out(tp
);
1813 if ((state
.reord
< tp
->fackets_out
) &&
1814 ((inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Loss
) || tp
->undo_marker
))
1815 tcp_update_reordering(sk
, tp
->fackets_out
- state
.reord
, 0);
1819 #if FASTRETRANS_DEBUG > 0
1820 WARN_ON((int)tp
->sacked_out
< 0);
1821 WARN_ON((int)tp
->lost_out
< 0);
1822 WARN_ON((int)tp
->retrans_out
< 0);
1823 WARN_ON((int)tcp_packets_in_flight(tp
) < 0);
1825 *sack_rtt
= state
.rtt
;
1829 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1830 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1832 static bool tcp_limit_reno_sacked(struct tcp_sock
*tp
)
1836 holes
= max(tp
->lost_out
, 1U);
1837 holes
= min(holes
, tp
->packets_out
);
1839 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1840 tp
->sacked_out
= tp
->packets_out
- holes
;
1846 /* If we receive more dupacks than we expected counting segments
1847 * in assumption of absent reordering, interpret this as reordering.
1848 * The only another reason could be bug in receiver TCP.
1850 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1852 struct tcp_sock
*tp
= tcp_sk(sk
);
1853 if (tcp_limit_reno_sacked(tp
))
1854 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1857 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1859 static void tcp_add_reno_sack(struct sock
*sk
)
1861 struct tcp_sock
*tp
= tcp_sk(sk
);
1863 tcp_check_reno_reordering(sk
, 0);
1864 tcp_verify_left_out(tp
);
1867 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1869 static void tcp_remove_reno_sacks(struct sock
*sk
, int acked
)
1871 struct tcp_sock
*tp
= tcp_sk(sk
);
1874 /* One ACK acked hole. The rest eat duplicate ACKs. */
1875 if (acked
- 1 >= tp
->sacked_out
)
1878 tp
->sacked_out
-= acked
- 1;
1880 tcp_check_reno_reordering(sk
, acked
);
1881 tcp_verify_left_out(tp
);
1884 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1889 static void tcp_clear_retrans_partial(struct tcp_sock
*tp
)
1891 tp
->retrans_out
= 0;
1894 tp
->undo_marker
= 0;
1895 tp
->undo_retrans
= 0;
1898 void tcp_clear_retrans(struct tcp_sock
*tp
)
1900 tcp_clear_retrans_partial(tp
);
1902 tp
->fackets_out
= 0;
1906 /* Enter Loss state. If "how" is not zero, forget all SACK information
1907 * and reset tags completely, otherwise preserve SACKs. If receiver
1908 * dropped its ofo queue, we will know this due to reneging detection.
1910 void tcp_enter_loss(struct sock
*sk
, int how
)
1912 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1913 struct tcp_sock
*tp
= tcp_sk(sk
);
1914 struct sk_buff
*skb
;
1915 bool new_recovery
= false;
1917 /* Reduce ssthresh if it has not yet been made inside this window. */
1918 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
||
1919 !after(tp
->high_seq
, tp
->snd_una
) ||
1920 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
1921 new_recovery
= true;
1922 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1923 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1924 tcp_ca_event(sk
, CA_EVENT_LOSS
);
1927 tp
->snd_cwnd_cnt
= 0;
1928 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1930 tcp_clear_retrans_partial(tp
);
1932 if (tcp_is_reno(tp
))
1933 tcp_reset_reno_sack(tp
);
1935 tp
->undo_marker
= tp
->snd_una
;
1938 tp
->fackets_out
= 0;
1940 tcp_clear_all_retrans_hints(tp
);
1942 tcp_for_write_queue(skb
, sk
) {
1943 if (skb
== tcp_send_head(sk
))
1946 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
1947 tp
->undo_marker
= 0;
1948 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
1949 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) || how
) {
1950 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
1951 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1952 tp
->lost_out
+= tcp_skb_pcount(skb
);
1953 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
1956 tcp_verify_left_out(tp
);
1958 /* Timeout in disordered state after receiving substantial DUPACKs
1959 * suggests that the degree of reordering is over-estimated.
1961 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
&&
1962 tp
->sacked_out
>= sysctl_tcp_reordering
)
1963 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1964 sysctl_tcp_reordering
);
1965 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1966 tp
->high_seq
= tp
->snd_nxt
;
1967 TCP_ECN_queue_cwr(tp
);
1969 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1970 * loss recovery is underway except recurring timeout(s) on
1971 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1973 tp
->frto
= sysctl_tcp_frto
&&
1974 (new_recovery
|| icsk
->icsk_retransmits
) &&
1975 !inet_csk(sk
)->icsk_mtup
.probe_size
;
1978 /* If ACK arrived pointing to a remembered SACK, it means that our
1979 * remembered SACKs do not reflect real state of receiver i.e.
1980 * receiver _host_ is heavily congested (or buggy).
1982 * Do processing similar to RTO timeout.
1984 static bool tcp_check_sack_reneging(struct sock
*sk
, int flag
)
1986 if (flag
& FLAG_SACK_RENEGING
) {
1987 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1988 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSACKRENEGING
);
1990 tcp_enter_loss(sk
, 1);
1991 icsk
->icsk_retransmits
++;
1992 tcp_retransmit_skb(sk
, tcp_write_queue_head(sk
));
1993 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1994 icsk
->icsk_rto
, TCP_RTO_MAX
);
2000 static inline int tcp_fackets_out(const struct tcp_sock
*tp
)
2002 return tcp_is_reno(tp
) ? tp
->sacked_out
+ 1 : tp
->fackets_out
;
2005 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2006 * counter when SACK is enabled (without SACK, sacked_out is used for
2009 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2010 * segments up to the highest received SACK block so far and holes in
2013 * With reordering, holes may still be in flight, so RFC3517 recovery
2014 * uses pure sacked_out (total number of SACKed segments) even though
2015 * it violates the RFC that uses duplicate ACKs, often these are equal
2016 * but when e.g. out-of-window ACKs or packet duplication occurs,
2017 * they differ. Since neither occurs due to loss, TCP should really
2020 static inline int tcp_dupack_heuristics(const struct tcp_sock
*tp
)
2022 return tcp_is_fack(tp
) ? tp
->fackets_out
: tp
->sacked_out
+ 1;
2025 static bool tcp_pause_early_retransmit(struct sock
*sk
, int flag
)
2027 struct tcp_sock
*tp
= tcp_sk(sk
);
2028 unsigned long delay
;
2030 /* Delay early retransmit and entering fast recovery for
2031 * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
2032 * available, or RTO is scheduled to fire first.
2034 if (sysctl_tcp_early_retrans
< 2 || sysctl_tcp_early_retrans
> 3 ||
2035 (flag
& FLAG_ECE
) || !tp
->srtt
)
2038 delay
= max_t(unsigned long, (tp
->srtt
>> 5), msecs_to_jiffies(2));
2039 if (!time_after(inet_csk(sk
)->icsk_timeout
, (jiffies
+ delay
)))
2042 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_EARLY_RETRANS
, delay
,
2047 /* Linux NewReno/SACK/FACK/ECN state machine.
2048 * --------------------------------------
2050 * "Open" Normal state, no dubious events, fast path.
2051 * "Disorder" In all the respects it is "Open",
2052 * but requires a bit more attention. It is entered when
2053 * we see some SACKs or dupacks. It is split of "Open"
2054 * mainly to move some processing from fast path to slow one.
2055 * "CWR" CWND was reduced due to some Congestion Notification event.
2056 * It can be ECN, ICMP source quench, local device congestion.
2057 * "Recovery" CWND was reduced, we are fast-retransmitting.
2058 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2060 * tcp_fastretrans_alert() is entered:
2061 * - each incoming ACK, if state is not "Open"
2062 * - when arrived ACK is unusual, namely:
2067 * Counting packets in flight is pretty simple.
2069 * in_flight = packets_out - left_out + retrans_out
2071 * packets_out is SND.NXT-SND.UNA counted in packets.
2073 * retrans_out is number of retransmitted segments.
2075 * left_out is number of segments left network, but not ACKed yet.
2077 * left_out = sacked_out + lost_out
2079 * sacked_out: Packets, which arrived to receiver out of order
2080 * and hence not ACKed. With SACKs this number is simply
2081 * amount of SACKed data. Even without SACKs
2082 * it is easy to give pretty reliable estimate of this number,
2083 * counting duplicate ACKs.
2085 * lost_out: Packets lost by network. TCP has no explicit
2086 * "loss notification" feedback from network (for now).
2087 * It means that this number can be only _guessed_.
2088 * Actually, it is the heuristics to predict lossage that
2089 * distinguishes different algorithms.
2091 * F.e. after RTO, when all the queue is considered as lost,
2092 * lost_out = packets_out and in_flight = retrans_out.
2094 * Essentially, we have now two algorithms counting
2097 * FACK: It is the simplest heuristics. As soon as we decided
2098 * that something is lost, we decide that _all_ not SACKed
2099 * packets until the most forward SACK are lost. I.e.
2100 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2101 * It is absolutely correct estimate, if network does not reorder
2102 * packets. And it loses any connection to reality when reordering
2103 * takes place. We use FACK by default until reordering
2104 * is suspected on the path to this destination.
2106 * NewReno: when Recovery is entered, we assume that one segment
2107 * is lost (classic Reno). While we are in Recovery and
2108 * a partial ACK arrives, we assume that one more packet
2109 * is lost (NewReno). This heuristics are the same in NewReno
2112 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2113 * deflation etc. CWND is real congestion window, never inflated, changes
2114 * only according to classic VJ rules.
2116 * Really tricky (and requiring careful tuning) part of algorithm
2117 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2118 * The first determines the moment _when_ we should reduce CWND and,
2119 * hence, slow down forward transmission. In fact, it determines the moment
2120 * when we decide that hole is caused by loss, rather than by a reorder.
2122 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2123 * holes, caused by lost packets.
2125 * And the most logically complicated part of algorithm is undo
2126 * heuristics. We detect false retransmits due to both too early
2127 * fast retransmit (reordering) and underestimated RTO, analyzing
2128 * timestamps and D-SACKs. When we detect that some segments were
2129 * retransmitted by mistake and CWND reduction was wrong, we undo
2130 * window reduction and abort recovery phase. This logic is hidden
2131 * inside several functions named tcp_try_undo_<something>.
2134 /* This function decides, when we should leave Disordered state
2135 * and enter Recovery phase, reducing congestion window.
2137 * Main question: may we further continue forward transmission
2138 * with the same cwnd?
2140 static bool tcp_time_to_recover(struct sock
*sk
, int flag
)
2142 struct tcp_sock
*tp
= tcp_sk(sk
);
2145 /* Trick#1: The loss is proven. */
2149 /* Not-A-Trick#2 : Classic rule... */
2150 if (tcp_dupack_heuristics(tp
) > tp
->reordering
)
2153 /* Trick#4: It is still not OK... But will it be useful to delay
2156 packets_out
= tp
->packets_out
;
2157 if (packets_out
<= tp
->reordering
&&
2158 tp
->sacked_out
>= max_t(__u32
, packets_out
/2, sysctl_tcp_reordering
) &&
2159 !tcp_may_send_now(sk
)) {
2160 /* We have nothing to send. This connection is limited
2161 * either by receiver window or by application.
2166 /* If a thin stream is detected, retransmit after first
2167 * received dupack. Employ only if SACK is supported in order
2168 * to avoid possible corner-case series of spurious retransmissions
2169 * Use only if there are no unsent data.
2171 if ((tp
->thin_dupack
|| sysctl_tcp_thin_dupack
) &&
2172 tcp_stream_is_thin(tp
) && tcp_dupack_heuristics(tp
) > 1 &&
2173 tcp_is_sack(tp
) && !tcp_send_head(sk
))
2176 /* Trick#6: TCP early retransmit, per RFC5827. To avoid spurious
2177 * retransmissions due to small network reorderings, we implement
2178 * Mitigation A.3 in the RFC and delay the retransmission for a short
2179 * interval if appropriate.
2181 if (tp
->do_early_retrans
&& !tp
->retrans_out
&& tp
->sacked_out
&&
2182 (tp
->packets_out
>= (tp
->sacked_out
+ 1) && tp
->packets_out
< 4) &&
2183 !tcp_may_send_now(sk
))
2184 return !tcp_pause_early_retransmit(sk
, flag
);
2189 /* Detect loss in event "A" above by marking head of queue up as lost.
2190 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2191 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2192 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2193 * the maximum SACKed segments to pass before reaching this limit.
2195 static void tcp_mark_head_lost(struct sock
*sk
, int packets
, int mark_head
)
2197 struct tcp_sock
*tp
= tcp_sk(sk
);
2198 struct sk_buff
*skb
;
2202 /* Use SACK to deduce losses of new sequences sent during recovery */
2203 const u32 loss_high
= tcp_is_sack(tp
) ? tp
->snd_nxt
: tp
->high_seq
;
2205 WARN_ON(packets
> tp
->packets_out
);
2206 if (tp
->lost_skb_hint
) {
2207 skb
= tp
->lost_skb_hint
;
2208 cnt
= tp
->lost_cnt_hint
;
2209 /* Head already handled? */
2210 if (mark_head
&& skb
!= tcp_write_queue_head(sk
))
2213 skb
= tcp_write_queue_head(sk
);
2217 tcp_for_write_queue_from(skb
, sk
) {
2218 if (skb
== tcp_send_head(sk
))
2220 /* TODO: do this better */
2221 /* this is not the most efficient way to do this... */
2222 tp
->lost_skb_hint
= skb
;
2223 tp
->lost_cnt_hint
= cnt
;
2225 if (after(TCP_SKB_CB(skb
)->end_seq
, loss_high
))
2229 if (tcp_is_fack(tp
) || tcp_is_reno(tp
) ||
2230 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2231 cnt
+= tcp_skb_pcount(skb
);
2233 if (cnt
> packets
) {
2234 if ((tcp_is_sack(tp
) && !tcp_is_fack(tp
)) ||
2235 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
) ||
2236 (oldcnt
>= packets
))
2239 mss
= skb_shinfo(skb
)->gso_size
;
2240 err
= tcp_fragment(sk
, skb
, (packets
- oldcnt
) * mss
, mss
);
2246 tcp_skb_mark_lost(tp
, skb
);
2251 tcp_verify_left_out(tp
);
2254 /* Account newly detected lost packet(s) */
2256 static void tcp_update_scoreboard(struct sock
*sk
, int fast_rexmit
)
2258 struct tcp_sock
*tp
= tcp_sk(sk
);
2260 if (tcp_is_reno(tp
)) {
2261 tcp_mark_head_lost(sk
, 1, 1);
2262 } else if (tcp_is_fack(tp
)) {
2263 int lost
= tp
->fackets_out
- tp
->reordering
;
2266 tcp_mark_head_lost(sk
, lost
, 0);
2268 int sacked_upto
= tp
->sacked_out
- tp
->reordering
;
2269 if (sacked_upto
>= 0)
2270 tcp_mark_head_lost(sk
, sacked_upto
, 0);
2271 else if (fast_rexmit
)
2272 tcp_mark_head_lost(sk
, 1, 1);
2276 /* CWND moderation, preventing bursts due to too big ACKs
2277 * in dubious situations.
2279 static inline void tcp_moderate_cwnd(struct tcp_sock
*tp
)
2281 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
2282 tcp_packets_in_flight(tp
) + tcp_max_burst(tp
));
2283 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2286 /* Nothing was retransmitted or returned timestamp is less
2287 * than timestamp of the first retransmission.
2289 static inline bool tcp_packet_delayed(const struct tcp_sock
*tp
)
2291 return !tp
->retrans_stamp
||
2292 (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2293 before(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
));
2296 /* Undo procedures. */
2298 #if FASTRETRANS_DEBUG > 1
2299 static void DBGUNDO(struct sock
*sk
, const char *msg
)
2301 struct tcp_sock
*tp
= tcp_sk(sk
);
2302 struct inet_sock
*inet
= inet_sk(sk
);
2304 if (sk
->sk_family
== AF_INET
) {
2305 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2307 &inet
->inet_daddr
, ntohs(inet
->inet_dport
),
2308 tp
->snd_cwnd
, tcp_left_out(tp
),
2309 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2312 #if IS_ENABLED(CONFIG_IPV6)
2313 else if (sk
->sk_family
== AF_INET6
) {
2314 struct ipv6_pinfo
*np
= inet6_sk(sk
);
2315 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2317 &np
->daddr
, ntohs(inet
->inet_dport
),
2318 tp
->snd_cwnd
, tcp_left_out(tp
),
2319 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2325 #define DBGUNDO(x...) do { } while (0)
2328 static void tcp_undo_cwnd_reduction(struct sock
*sk
, bool unmark_loss
)
2330 struct tcp_sock
*tp
= tcp_sk(sk
);
2333 struct sk_buff
*skb
;
2335 tcp_for_write_queue(skb
, sk
) {
2336 if (skb
== tcp_send_head(sk
))
2338 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2341 tcp_clear_all_retrans_hints(tp
);
2344 if (tp
->prior_ssthresh
) {
2345 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2347 if (icsk
->icsk_ca_ops
->undo_cwnd
)
2348 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
2350 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
<< 1);
2352 if (tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
2353 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
2354 TCP_ECN_withdraw_cwr(tp
);
2357 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2359 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2360 tp
->undo_marker
= 0;
2363 static inline bool tcp_may_undo(const struct tcp_sock
*tp
)
2365 return tp
->undo_marker
&& (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
2368 /* People celebrate: "We love our President!" */
2369 static bool tcp_try_undo_recovery(struct sock
*sk
)
2371 struct tcp_sock
*tp
= tcp_sk(sk
);
2373 if (tcp_may_undo(tp
)) {
2376 /* Happy end! We did not retransmit anything
2377 * or our original transmission succeeded.
2379 DBGUNDO(sk
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
2380 tcp_undo_cwnd_reduction(sk
, false);
2381 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
2382 mib_idx
= LINUX_MIB_TCPLOSSUNDO
;
2384 mib_idx
= LINUX_MIB_TCPFULLUNDO
;
2386 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2388 if (tp
->snd_una
== tp
->high_seq
&& tcp_is_reno(tp
)) {
2389 /* Hold old state until something *above* high_seq
2390 * is ACKed. For Reno it is MUST to prevent false
2391 * fast retransmits (RFC2582). SACK TCP is safe. */
2392 tcp_moderate_cwnd(tp
);
2395 tcp_set_ca_state(sk
, TCP_CA_Open
);
2399 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2400 static bool tcp_try_undo_dsack(struct sock
*sk
)
2402 struct tcp_sock
*tp
= tcp_sk(sk
);
2404 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
2405 DBGUNDO(sk
, "D-SACK");
2406 tcp_undo_cwnd_reduction(sk
, false);
2407 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPDSACKUNDO
);
2413 /* We can clear retrans_stamp when there are no retransmissions in the
2414 * window. It would seem that it is trivially available for us in
2415 * tp->retrans_out, however, that kind of assumptions doesn't consider
2416 * what will happen if errors occur when sending retransmission for the
2417 * second time. ...It could the that such segment has only
2418 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2419 * the head skb is enough except for some reneging corner cases that
2420 * are not worth the effort.
2422 * Main reason for all this complexity is the fact that connection dying
2423 * time now depends on the validity of the retrans_stamp, in particular,
2424 * that successive retransmissions of a segment must not advance
2425 * retrans_stamp under any conditions.
2427 static bool tcp_any_retrans_done(const struct sock
*sk
)
2429 const struct tcp_sock
*tp
= tcp_sk(sk
);
2430 struct sk_buff
*skb
;
2432 if (tp
->retrans_out
)
2435 skb
= tcp_write_queue_head(sk
);
2436 if (unlikely(skb
&& TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
))
2442 /* Undo during loss recovery after partial ACK or using F-RTO. */
2443 static bool tcp_try_undo_loss(struct sock
*sk
, bool frto_undo
)
2445 struct tcp_sock
*tp
= tcp_sk(sk
);
2447 if (frto_undo
|| tcp_may_undo(tp
)) {
2448 tcp_undo_cwnd_reduction(sk
, true);
2450 DBGUNDO(sk
, "partial loss");
2451 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSSUNDO
);
2453 NET_INC_STATS_BH(sock_net(sk
),
2454 LINUX_MIB_TCPSPURIOUSRTOS
);
2455 inet_csk(sk
)->icsk_retransmits
= 0;
2456 if (frto_undo
|| tcp_is_sack(tp
))
2457 tcp_set_ca_state(sk
, TCP_CA_Open
);
2463 /* The cwnd reduction in CWR and Recovery use the PRR algorithm
2464 * https://datatracker.ietf.org/doc/draft-ietf-tcpm-proportional-rate-reduction/
2465 * It computes the number of packets to send (sndcnt) based on packets newly
2467 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2468 * cwnd reductions across a full RTT.
2469 * 2) If packets in flight is lower than ssthresh (such as due to excess
2470 * losses and/or application stalls), do not perform any further cwnd
2471 * reductions, but instead slow start up to ssthresh.
2473 static void tcp_init_cwnd_reduction(struct sock
*sk
, const bool set_ssthresh
)
2475 struct tcp_sock
*tp
= tcp_sk(sk
);
2477 tp
->high_seq
= tp
->snd_nxt
;
2478 tp
->tlp_high_seq
= 0;
2479 tp
->snd_cwnd_cnt
= 0;
2480 tp
->prior_cwnd
= tp
->snd_cwnd
;
2481 tp
->prr_delivered
= 0;
2484 tp
->snd_ssthresh
= inet_csk(sk
)->icsk_ca_ops
->ssthresh(sk
);
2485 TCP_ECN_queue_cwr(tp
);
2488 static void tcp_cwnd_reduction(struct sock
*sk
, const int prior_unsacked
,
2491 struct tcp_sock
*tp
= tcp_sk(sk
);
2493 int delta
= tp
->snd_ssthresh
- tcp_packets_in_flight(tp
);
2494 int newly_acked_sacked
= prior_unsacked
-
2495 (tp
->packets_out
- tp
->sacked_out
);
2497 tp
->prr_delivered
+= newly_acked_sacked
;
2498 if (tcp_packets_in_flight(tp
) > tp
->snd_ssthresh
) {
2499 u64 dividend
= (u64
)tp
->snd_ssthresh
* tp
->prr_delivered
+
2501 sndcnt
= div_u64(dividend
, tp
->prior_cwnd
) - tp
->prr_out
;
2503 sndcnt
= min_t(int, delta
,
2504 max_t(int, tp
->prr_delivered
- tp
->prr_out
,
2505 newly_acked_sacked
) + 1);
2508 sndcnt
= max(sndcnt
, (fast_rexmit
? 1 : 0));
2509 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + sndcnt
;
2512 static inline void tcp_end_cwnd_reduction(struct sock
*sk
)
2514 struct tcp_sock
*tp
= tcp_sk(sk
);
2516 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2517 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_CWR
||
2518 (tp
->undo_marker
&& tp
->snd_ssthresh
< TCP_INFINITE_SSTHRESH
)) {
2519 tp
->snd_cwnd
= tp
->snd_ssthresh
;
2520 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2522 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
2525 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2526 void tcp_enter_cwr(struct sock
*sk
, const int set_ssthresh
)
2528 struct tcp_sock
*tp
= tcp_sk(sk
);
2530 tp
->prior_ssthresh
= 0;
2531 if (inet_csk(sk
)->icsk_ca_state
< TCP_CA_CWR
) {
2532 tp
->undo_marker
= 0;
2533 tcp_init_cwnd_reduction(sk
, set_ssthresh
);
2534 tcp_set_ca_state(sk
, TCP_CA_CWR
);
2538 static void tcp_try_keep_open(struct sock
*sk
)
2540 struct tcp_sock
*tp
= tcp_sk(sk
);
2541 int state
= TCP_CA_Open
;
2543 if (tcp_left_out(tp
) || tcp_any_retrans_done(sk
))
2544 state
= TCP_CA_Disorder
;
2546 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
2547 tcp_set_ca_state(sk
, state
);
2548 tp
->high_seq
= tp
->snd_nxt
;
2552 static void tcp_try_to_open(struct sock
*sk
, int flag
, const int prior_unsacked
)
2554 struct tcp_sock
*tp
= tcp_sk(sk
);
2556 tcp_verify_left_out(tp
);
2558 if (!tcp_any_retrans_done(sk
))
2559 tp
->retrans_stamp
= 0;
2561 if (flag
& FLAG_ECE
)
2562 tcp_enter_cwr(sk
, 1);
2564 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
2565 tcp_try_keep_open(sk
);
2567 tcp_cwnd_reduction(sk
, prior_unsacked
, 0);
2571 static void tcp_mtup_probe_failed(struct sock
*sk
)
2573 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2575 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
2576 icsk
->icsk_mtup
.probe_size
= 0;
2579 static void tcp_mtup_probe_success(struct sock
*sk
)
2581 struct tcp_sock
*tp
= tcp_sk(sk
);
2582 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2584 /* FIXME: breaks with very large cwnd */
2585 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2586 tp
->snd_cwnd
= tp
->snd_cwnd
*
2587 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
2588 icsk
->icsk_mtup
.probe_size
;
2589 tp
->snd_cwnd_cnt
= 0;
2590 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2591 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2593 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
2594 icsk
->icsk_mtup
.probe_size
= 0;
2595 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
2598 /* Do a simple retransmit without using the backoff mechanisms in
2599 * tcp_timer. This is used for path mtu discovery.
2600 * The socket is already locked here.
2602 void tcp_simple_retransmit(struct sock
*sk
)
2604 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2605 struct tcp_sock
*tp
= tcp_sk(sk
);
2606 struct sk_buff
*skb
;
2607 unsigned int mss
= tcp_current_mss(sk
);
2608 u32 prior_lost
= tp
->lost_out
;
2610 tcp_for_write_queue(skb
, sk
) {
2611 if (skb
== tcp_send_head(sk
))
2613 if (tcp_skb_seglen(skb
) > mss
&&
2614 !(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
2615 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2616 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
2617 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2619 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
2623 tcp_clear_retrans_hints_partial(tp
);
2625 if (prior_lost
== tp
->lost_out
)
2628 if (tcp_is_reno(tp
))
2629 tcp_limit_reno_sacked(tp
);
2631 tcp_verify_left_out(tp
);
2633 /* Don't muck with the congestion window here.
2634 * Reason is that we do not increase amount of _data_
2635 * in network, but units changed and effective
2636 * cwnd/ssthresh really reduced now.
2638 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
2639 tp
->high_seq
= tp
->snd_nxt
;
2640 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2641 tp
->prior_ssthresh
= 0;
2642 tp
->undo_marker
= 0;
2643 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2645 tcp_xmit_retransmit_queue(sk
);
2647 EXPORT_SYMBOL(tcp_simple_retransmit
);
2649 static void tcp_enter_recovery(struct sock
*sk
, bool ece_ack
)
2651 struct tcp_sock
*tp
= tcp_sk(sk
);
2654 if (tcp_is_reno(tp
))
2655 mib_idx
= LINUX_MIB_TCPRENORECOVERY
;
2657 mib_idx
= LINUX_MIB_TCPSACKRECOVERY
;
2659 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2661 tp
->prior_ssthresh
= 0;
2662 tp
->undo_marker
= tp
->snd_una
;
2663 tp
->undo_retrans
= tp
->retrans_out
;
2665 if (inet_csk(sk
)->icsk_ca_state
< TCP_CA_CWR
) {
2667 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2668 tcp_init_cwnd_reduction(sk
, true);
2670 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
2673 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2674 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2676 static void tcp_process_loss(struct sock
*sk
, int flag
, bool is_dupack
)
2678 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2679 struct tcp_sock
*tp
= tcp_sk(sk
);
2680 bool recovered
= !before(tp
->snd_una
, tp
->high_seq
);
2682 if (tp
->frto
) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2683 if (flag
& FLAG_ORIG_SACK_ACKED
) {
2684 /* Step 3.b. A timeout is spurious if not all data are
2685 * lost, i.e., never-retransmitted data are (s)acked.
2687 tcp_try_undo_loss(sk
, true);
2690 if (after(tp
->snd_nxt
, tp
->high_seq
) &&
2691 (flag
& FLAG_DATA_SACKED
|| is_dupack
)) {
2692 tp
->frto
= 0; /* Loss was real: 2nd part of step 3.a */
2693 } else if (flag
& FLAG_SND_UNA_ADVANCED
&& !recovered
) {
2694 tp
->high_seq
= tp
->snd_nxt
;
2695 __tcp_push_pending_frames(sk
, tcp_current_mss(sk
),
2697 if (after(tp
->snd_nxt
, tp
->high_seq
))
2698 return; /* Step 2.b */
2704 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2705 icsk
->icsk_retransmits
= 0;
2706 tcp_try_undo_recovery(sk
);
2709 if (flag
& FLAG_DATA_ACKED
)
2710 icsk
->icsk_retransmits
= 0;
2711 if (tcp_is_reno(tp
)) {
2712 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2713 * delivered. Lower inflight to clock out (re)tranmissions.
2715 if (after(tp
->snd_nxt
, tp
->high_seq
) && is_dupack
)
2716 tcp_add_reno_sack(sk
);
2717 else if (flag
& FLAG_SND_UNA_ADVANCED
)
2718 tcp_reset_reno_sack(tp
);
2720 if (tcp_try_undo_loss(sk
, false))
2722 tcp_xmit_retransmit_queue(sk
);
2725 /* Undo during fast recovery after partial ACK. */
2726 static bool tcp_try_undo_partial(struct sock
*sk
, const int acked
,
2727 const int prior_unsacked
)
2729 struct tcp_sock
*tp
= tcp_sk(sk
);
2731 if (tp
->undo_marker
&& tcp_packet_delayed(tp
)) {
2732 /* Plain luck! Hole if filled with delayed
2733 * packet, rather than with a retransmit.
2735 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
2737 /* We are getting evidence that the reordering degree is higher
2738 * than we realized. If there are no retransmits out then we
2739 * can undo. Otherwise we clock out new packets but do not
2740 * mark more packets lost or retransmit more.
2742 if (tp
->retrans_out
) {
2743 tcp_cwnd_reduction(sk
, prior_unsacked
, 0);
2747 if (!tcp_any_retrans_done(sk
))
2748 tp
->retrans_stamp
= 0;
2750 DBGUNDO(sk
, "partial recovery");
2751 tcp_undo_cwnd_reduction(sk
, true);
2752 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPARTIALUNDO
);
2753 tcp_try_keep_open(sk
);
2759 /* Process an event, which can update packets-in-flight not trivially.
2760 * Main goal of this function is to calculate new estimate for left_out,
2761 * taking into account both packets sitting in receiver's buffer and
2762 * packets lost by network.
2764 * Besides that it does CWND reduction, when packet loss is detected
2765 * and changes state of machine.
2767 * It does _not_ decide what to send, it is made in function
2768 * tcp_xmit_retransmit_queue().
2770 static void tcp_fastretrans_alert(struct sock
*sk
, const int acked
,
2771 const int prior_unsacked
,
2772 bool is_dupack
, int flag
)
2774 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2775 struct tcp_sock
*tp
= tcp_sk(sk
);
2776 bool do_lost
= is_dupack
|| ((flag
& FLAG_DATA_SACKED
) &&
2777 (tcp_fackets_out(tp
) > tp
->reordering
));
2778 int fast_rexmit
= 0;
2780 if (WARN_ON(!tp
->packets_out
&& tp
->sacked_out
))
2782 if (WARN_ON(!tp
->sacked_out
&& tp
->fackets_out
))
2783 tp
->fackets_out
= 0;
2785 /* Now state machine starts.
2786 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2787 if (flag
& FLAG_ECE
)
2788 tp
->prior_ssthresh
= 0;
2790 /* B. In all the states check for reneging SACKs. */
2791 if (tcp_check_sack_reneging(sk
, flag
))
2794 /* C. Check consistency of the current state. */
2795 tcp_verify_left_out(tp
);
2797 /* D. Check state exit conditions. State can be terminated
2798 * when high_seq is ACKed. */
2799 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
2800 WARN_ON(tp
->retrans_out
!= 0);
2801 tp
->retrans_stamp
= 0;
2802 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
2803 switch (icsk
->icsk_ca_state
) {
2805 /* CWR is to be held something *above* high_seq
2806 * is ACKed for CWR bit to reach receiver. */
2807 if (tp
->snd_una
!= tp
->high_seq
) {
2808 tcp_end_cwnd_reduction(sk
);
2809 tcp_set_ca_state(sk
, TCP_CA_Open
);
2813 case TCP_CA_Recovery
:
2814 if (tcp_is_reno(tp
))
2815 tcp_reset_reno_sack(tp
);
2816 if (tcp_try_undo_recovery(sk
))
2818 tcp_end_cwnd_reduction(sk
);
2823 /* E. Process state. */
2824 switch (icsk
->icsk_ca_state
) {
2825 case TCP_CA_Recovery
:
2826 if (!(flag
& FLAG_SND_UNA_ADVANCED
)) {
2827 if (tcp_is_reno(tp
) && is_dupack
)
2828 tcp_add_reno_sack(sk
);
2830 if (tcp_try_undo_partial(sk
, acked
, prior_unsacked
))
2832 /* Partial ACK arrived. Force fast retransmit. */
2833 do_lost
= tcp_is_reno(tp
) ||
2834 tcp_fackets_out(tp
) > tp
->reordering
;
2836 if (tcp_try_undo_dsack(sk
)) {
2837 tcp_try_keep_open(sk
);
2842 tcp_process_loss(sk
, flag
, is_dupack
);
2843 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
2845 /* Fall through to processing in Open state. */
2847 if (tcp_is_reno(tp
)) {
2848 if (flag
& FLAG_SND_UNA_ADVANCED
)
2849 tcp_reset_reno_sack(tp
);
2851 tcp_add_reno_sack(sk
);
2854 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
)
2855 tcp_try_undo_dsack(sk
);
2857 if (!tcp_time_to_recover(sk
, flag
)) {
2858 tcp_try_to_open(sk
, flag
, prior_unsacked
);
2862 /* MTU probe failure: don't reduce cwnd */
2863 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
2864 icsk
->icsk_mtup
.probe_size
&&
2865 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
2866 tcp_mtup_probe_failed(sk
);
2867 /* Restores the reduction we did in tcp_mtup_probe() */
2869 tcp_simple_retransmit(sk
);
2873 /* Otherwise enter Recovery state */
2874 tcp_enter_recovery(sk
, (flag
& FLAG_ECE
));
2879 tcp_update_scoreboard(sk
, fast_rexmit
);
2880 tcp_cwnd_reduction(sk
, prior_unsacked
, fast_rexmit
);
2881 tcp_xmit_retransmit_queue(sk
);
2884 static inline bool tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
2885 s32 seq_rtt
, s32 sack_rtt
)
2887 const struct tcp_sock
*tp
= tcp_sk(sk
);
2889 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2890 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2891 * Karn's algorithm forbids taking RTT if some retransmitted data
2892 * is acked (RFC6298).
2894 if (flag
& FLAG_RETRANS_DATA_ACKED
)
2900 /* RTTM Rule: A TSecr value received in a segment is used to
2901 * update the averaged RTT measurement only if the segment
2902 * acknowledges some new data, i.e., only if it advances the
2903 * left edge of the send window.
2904 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2906 if (seq_rtt
< 0 && tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2908 seq_rtt
= tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
;
2913 tcp_rtt_estimator(sk
, seq_rtt
);
2916 /* RFC6298: only reset backoff on valid RTT measurement. */
2917 inet_csk(sk
)->icsk_backoff
= 0;
2921 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2922 static void tcp_synack_rtt_meas(struct sock
*sk
, const u32 synack_stamp
)
2924 struct tcp_sock
*tp
= tcp_sk(sk
);
2927 if (synack_stamp
&& !tp
->total_retrans
)
2928 seq_rtt
= tcp_time_stamp
- synack_stamp
;
2930 /* If the ACK acks both the SYNACK and the (Fast Open'd) data packets
2931 * sent in SYN_RECV, SYNACK RTT is the smooth RTT computed in tcp_ack()
2934 tcp_ack_update_rtt(sk
, FLAG_SYN_ACKED
, seq_rtt
, -1);
2937 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
, u32 acked
, u32 in_flight
)
2939 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2940 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, acked
, in_flight
);
2941 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_time_stamp
;
2944 /* Restart timer after forward progress on connection.
2945 * RFC2988 recommends to restart timer to now+rto.
2947 void tcp_rearm_rto(struct sock
*sk
)
2949 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2950 struct tcp_sock
*tp
= tcp_sk(sk
);
2952 /* If the retrans timer is currently being used by Fast Open
2953 * for SYN-ACK retrans purpose, stay put.
2955 if (tp
->fastopen_rsk
)
2958 if (!tp
->packets_out
) {
2959 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
2961 u32 rto
= inet_csk(sk
)->icsk_rto
;
2962 /* Offset the time elapsed after installing regular RTO */
2963 if (icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
||
2964 icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
) {
2965 struct sk_buff
*skb
= tcp_write_queue_head(sk
);
2966 const u32 rto_time_stamp
= TCP_SKB_CB(skb
)->when
+ rto
;
2967 s32 delta
= (s32
)(rto_time_stamp
- tcp_time_stamp
);
2968 /* delta may not be positive if the socket is locked
2969 * when the retrans timer fires and is rescheduled.
2974 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
, rto
,
2979 /* This function is called when the delayed ER timer fires. TCP enters
2980 * fast recovery and performs fast-retransmit.
2982 void tcp_resume_early_retransmit(struct sock
*sk
)
2984 struct tcp_sock
*tp
= tcp_sk(sk
);
2988 /* Stop if ER is disabled after the delayed ER timer is scheduled */
2989 if (!tp
->do_early_retrans
)
2992 tcp_enter_recovery(sk
, false);
2993 tcp_update_scoreboard(sk
, 1);
2994 tcp_xmit_retransmit_queue(sk
);
2997 /* If we get here, the whole TSO packet has not been acked. */
2998 static u32
tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
)
3000 struct tcp_sock
*tp
= tcp_sk(sk
);
3003 BUG_ON(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
));
3005 packets_acked
= tcp_skb_pcount(skb
);
3006 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
3008 packets_acked
-= tcp_skb_pcount(skb
);
3010 if (packets_acked
) {
3011 BUG_ON(tcp_skb_pcount(skb
) == 0);
3012 BUG_ON(!before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
));
3015 return packets_acked
;
3018 /* Remove acknowledged frames from the retransmission queue. If our packet
3019 * is before the ack sequence we can discard it as it's confirmed to have
3020 * arrived at the other end.
3022 static int tcp_clean_rtx_queue(struct sock
*sk
, int prior_fackets
,
3023 u32 prior_snd_una
, s32 sack_rtt
)
3025 struct tcp_sock
*tp
= tcp_sk(sk
);
3026 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3027 struct sk_buff
*skb
;
3028 u32 now
= tcp_time_stamp
;
3029 bool fully_acked
= true;
3032 u32 reord
= tp
->packets_out
;
3033 u32 prior_sacked
= tp
->sacked_out
;
3035 s32 ca_seq_rtt
= -1;
3036 ktime_t last_ackt
= net_invalid_timestamp();
3039 while ((skb
= tcp_write_queue_head(sk
)) && skb
!= tcp_send_head(sk
)) {
3040 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
3042 u8 sacked
= scb
->sacked
;
3044 /* Determine how many packets and what bytes were acked, tso and else */
3045 if (after(scb
->end_seq
, tp
->snd_una
)) {
3046 if (tcp_skb_pcount(skb
) == 1 ||
3047 !after(tp
->snd_una
, scb
->seq
))
3050 acked_pcount
= tcp_tso_acked(sk
, skb
);
3054 fully_acked
= false;
3056 acked_pcount
= tcp_skb_pcount(skb
);
3059 if (sacked
& TCPCB_RETRANS
) {
3060 if (sacked
& TCPCB_SACKED_RETRANS
)
3061 tp
->retrans_out
-= acked_pcount
;
3062 flag
|= FLAG_RETRANS_DATA_ACKED
;
3064 ca_seq_rtt
= now
- scb
->when
;
3065 last_ackt
= skb
->tstamp
;
3067 seq_rtt
= ca_seq_rtt
;
3069 if (!(sacked
& TCPCB_SACKED_ACKED
))
3070 reord
= min(pkts_acked
, reord
);
3071 if (!after(scb
->end_seq
, tp
->high_seq
))
3072 flag
|= FLAG_ORIG_SACK_ACKED
;
3075 if (sacked
& TCPCB_SACKED_ACKED
)
3076 tp
->sacked_out
-= acked_pcount
;
3077 if (sacked
& TCPCB_LOST
)
3078 tp
->lost_out
-= acked_pcount
;
3080 tp
->packets_out
-= acked_pcount
;
3081 pkts_acked
+= acked_pcount
;
3083 /* Initial outgoing SYN's get put onto the write_queue
3084 * just like anything else we transmit. It is not
3085 * true data, and if we misinform our callers that
3086 * this ACK acks real data, we will erroneously exit
3087 * connection startup slow start one packet too
3088 * quickly. This is severely frowned upon behavior.
3090 if (!(scb
->tcp_flags
& TCPHDR_SYN
)) {
3091 flag
|= FLAG_DATA_ACKED
;
3093 flag
|= FLAG_SYN_ACKED
;
3094 tp
->retrans_stamp
= 0;
3100 tcp_unlink_write_queue(skb
, sk
);
3101 sk_wmem_free_skb(sk
, skb
);
3102 if (skb
== tp
->retransmit_skb_hint
)
3103 tp
->retransmit_skb_hint
= NULL
;
3104 if (skb
== tp
->lost_skb_hint
)
3105 tp
->lost_skb_hint
= NULL
;
3108 if (likely(between(tp
->snd_up
, prior_snd_una
, tp
->snd_una
)))
3109 tp
->snd_up
= tp
->snd_una
;
3111 if (skb
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
3112 flag
|= FLAG_SACK_RENEGING
;
3114 rtt_update
= tcp_ack_update_rtt(sk
, flag
, seq_rtt
, sack_rtt
);
3116 if (flag
& FLAG_ACKED
) {
3117 const struct tcp_congestion_ops
*ca_ops
3118 = inet_csk(sk
)->icsk_ca_ops
;
3121 if (unlikely(icsk
->icsk_mtup
.probe_size
&&
3122 !after(tp
->mtu_probe
.probe_seq_end
, tp
->snd_una
))) {
3123 tcp_mtup_probe_success(sk
);
3126 if (tcp_is_reno(tp
)) {
3127 tcp_remove_reno_sacks(sk
, pkts_acked
);
3131 /* Non-retransmitted hole got filled? That's reordering */
3132 if (reord
< prior_fackets
)
3133 tcp_update_reordering(sk
, tp
->fackets_out
- reord
, 0);
3135 delta
= tcp_is_fack(tp
) ? pkts_acked
:
3136 prior_sacked
- tp
->sacked_out
;
3137 tp
->lost_cnt_hint
-= min(tp
->lost_cnt_hint
, delta
);
3140 tp
->fackets_out
-= min(pkts_acked
, tp
->fackets_out
);
3142 if (ca_ops
->pkts_acked
) {
3145 /* Is the ACK triggering packet unambiguous? */
3146 if (!(flag
& FLAG_RETRANS_DATA_ACKED
)) {
3147 /* High resolution needed and available? */
3148 if (ca_ops
->flags
& TCP_CONG_RTT_STAMP
&&
3149 !ktime_equal(last_ackt
,
3150 net_invalid_timestamp()))
3151 rtt_us
= ktime_us_delta(ktime_get_real(),
3153 else if (ca_seq_rtt
>= 0)
3154 rtt_us
= jiffies_to_usecs(ca_seq_rtt
);
3157 ca_ops
->pkts_acked(sk
, pkts_acked
, rtt_us
);
3159 } else if (skb
&& rtt_update
&& sack_rtt
>= 0 &&
3160 sack_rtt
> (s32
)(now
- TCP_SKB_CB(skb
)->when
)) {
3161 /* Do not re-arm RTO if the sack RTT is measured from data sent
3162 * after when the head was last (re)transmitted. Otherwise the
3163 * timeout may continue to extend in loss recovery.
3168 #if FASTRETRANS_DEBUG > 0
3169 WARN_ON((int)tp
->sacked_out
< 0);
3170 WARN_ON((int)tp
->lost_out
< 0);
3171 WARN_ON((int)tp
->retrans_out
< 0);
3172 if (!tp
->packets_out
&& tcp_is_sack(tp
)) {
3173 icsk
= inet_csk(sk
);
3175 pr_debug("Leak l=%u %d\n",
3176 tp
->lost_out
, icsk
->icsk_ca_state
);
3179 if (tp
->sacked_out
) {
3180 pr_debug("Leak s=%u %d\n",
3181 tp
->sacked_out
, icsk
->icsk_ca_state
);
3184 if (tp
->retrans_out
) {
3185 pr_debug("Leak r=%u %d\n",
3186 tp
->retrans_out
, icsk
->icsk_ca_state
);
3187 tp
->retrans_out
= 0;
3194 static void tcp_ack_probe(struct sock
*sk
)
3196 const struct tcp_sock
*tp
= tcp_sk(sk
);
3197 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3199 /* Was it a usable window open? */
3201 if (!after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
, tcp_wnd_end(tp
))) {
3202 icsk
->icsk_backoff
= 0;
3203 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
3204 /* Socket must be waked up by subsequent tcp_data_snd_check().
3205 * This function is not for random using!
3208 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3209 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
3214 static inline bool tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
3216 return !(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
3217 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
;
3220 /* Decide wheather to run the increase function of congestion control. */
3221 static inline bool tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
3223 if (tcp_in_cwnd_reduction(sk
))
3226 /* If reordering is high then always grow cwnd whenever data is
3227 * delivered regardless of its ordering. Otherwise stay conservative
3228 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3229 * new SACK or ECE mark may first advance cwnd here and later reduce
3230 * cwnd in tcp_fastretrans_alert() based on more states.
3232 if (tcp_sk(sk
)->reordering
> sysctl_tcp_reordering
)
3233 return flag
& FLAG_FORWARD_PROGRESS
;
3235 return flag
& FLAG_DATA_ACKED
;
3238 /* Check that window update is acceptable.
3239 * The function assumes that snd_una<=ack<=snd_next.
3241 static inline bool tcp_may_update_window(const struct tcp_sock
*tp
,
3242 const u32 ack
, const u32 ack_seq
,
3245 return after(ack
, tp
->snd_una
) ||
3246 after(ack_seq
, tp
->snd_wl1
) ||
3247 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
);
3250 /* Update our send window.
3252 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3253 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3255 static int tcp_ack_update_window(struct sock
*sk
, const struct sk_buff
*skb
, u32 ack
,
3258 struct tcp_sock
*tp
= tcp_sk(sk
);
3260 u32 nwin
= ntohs(tcp_hdr(skb
)->window
);
3262 if (likely(!tcp_hdr(skb
)->syn
))
3263 nwin
<<= tp
->rx_opt
.snd_wscale
;
3265 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
3266 flag
|= FLAG_WIN_UPDATE
;
3267 tcp_update_wl(tp
, ack_seq
);
3269 if (tp
->snd_wnd
!= nwin
) {
3272 /* Note, it is the only place, where
3273 * fast path is recovered for sending TCP.
3276 tcp_fast_path_check(sk
);
3278 if (nwin
> tp
->max_window
) {
3279 tp
->max_window
= nwin
;
3280 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
3290 /* RFC 5961 7 [ACK Throttling] */
3291 static void tcp_send_challenge_ack(struct sock
*sk
)
3293 /* unprotected vars, we dont care of overwrites */
3294 static u32 challenge_timestamp
;
3295 static unsigned int challenge_count
;
3296 u32 now
= jiffies
/ HZ
;
3298 if (now
!= challenge_timestamp
) {
3299 challenge_timestamp
= now
;
3300 challenge_count
= 0;
3302 if (++challenge_count
<= sysctl_tcp_challenge_ack_limit
) {
3303 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPCHALLENGEACK
);
3308 static void tcp_store_ts_recent(struct tcp_sock
*tp
)
3310 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
3311 tp
->rx_opt
.ts_recent_stamp
= get_seconds();
3314 static void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
3316 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
3317 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3318 * extra check below makes sure this can only happen
3319 * for pure ACK frames. -DaveM
3321 * Not only, also it occurs for expired timestamps.
3324 if (tcp_paws_check(&tp
->rx_opt
, 0))
3325 tcp_store_ts_recent(tp
);
3329 /* This routine deals with acks during a TLP episode.
3330 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3332 static void tcp_process_tlp_ack(struct sock
*sk
, u32 ack
, int flag
)
3334 struct tcp_sock
*tp
= tcp_sk(sk
);
3335 bool is_tlp_dupack
= (ack
== tp
->tlp_high_seq
) &&
3336 !(flag
& (FLAG_SND_UNA_ADVANCED
|
3337 FLAG_NOT_DUP
| FLAG_DATA_SACKED
));
3339 /* Mark the end of TLP episode on receiving TLP dupack or when
3340 * ack is after tlp_high_seq.
3342 if (is_tlp_dupack
) {
3343 tp
->tlp_high_seq
= 0;
3347 if (after(ack
, tp
->tlp_high_seq
)) {
3348 tp
->tlp_high_seq
= 0;
3349 /* Don't reduce cwnd if DSACK arrives for TLP retrans. */
3350 if (!(flag
& FLAG_DSACKING_ACK
)) {
3351 tcp_init_cwnd_reduction(sk
, true);
3352 tcp_set_ca_state(sk
, TCP_CA_CWR
);
3353 tcp_end_cwnd_reduction(sk
);
3354 tcp_try_keep_open(sk
);
3355 NET_INC_STATS_BH(sock_net(sk
),
3356 LINUX_MIB_TCPLOSSPROBERECOVERY
);
3361 /* This routine deals with incoming acks, but not outgoing ones. */
3362 static int tcp_ack(struct sock
*sk
, const struct sk_buff
*skb
, int flag
)
3364 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3365 struct tcp_sock
*tp
= tcp_sk(sk
);
3366 u32 prior_snd_una
= tp
->snd_una
;
3367 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
3368 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3369 bool is_dupack
= false;
3370 u32 prior_in_flight
, prior_cwnd
= tp
->snd_cwnd
, prior_rtt
= tp
->srtt
;
3372 int prior_packets
= tp
->packets_out
;
3373 const int prior_unsacked
= tp
->packets_out
- tp
->sacked_out
;
3374 int acked
= 0; /* Number of packets newly acked */
3377 /* If the ack is older than previous acks
3378 * then we can probably ignore it.
3380 if (before(ack
, prior_snd_una
)) {
3381 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3382 if (before(ack
, prior_snd_una
- tp
->max_window
)) {
3383 tcp_send_challenge_ack(sk
);
3389 /* If the ack includes data we haven't sent yet, discard
3390 * this segment (RFC793 Section 3.9).
3392 if (after(ack
, tp
->snd_nxt
))
3395 if (icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
||
3396 icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
)
3399 if (after(ack
, prior_snd_una
))
3400 flag
|= FLAG_SND_UNA_ADVANCED
;
3402 prior_fackets
= tp
->fackets_out
;
3403 prior_in_flight
= tcp_packets_in_flight(tp
);
3405 /* ts_recent update must be made after we are sure that the packet
3408 if (flag
& FLAG_UPDATE_TS_RECENT
)
3409 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
3411 if (!(flag
& FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
3412 /* Window is constant, pure forward advance.
3413 * No more checks are required.
3414 * Note, we use the fact that SND.UNA>=SND.WL2.
3416 tcp_update_wl(tp
, ack_seq
);
3418 flag
|= FLAG_WIN_UPDATE
;
3420 tcp_ca_event(sk
, CA_EVENT_FAST_ACK
);
3422 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPACKS
);
3424 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
3427 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPUREACKS
);
3429 flag
|= tcp_ack_update_window(sk
, skb
, ack
, ack_seq
);
3431 if (TCP_SKB_CB(skb
)->sacked
)
3432 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
,
3435 if (TCP_ECN_rcv_ecn_echo(tp
, tcp_hdr(skb
)))
3438 tcp_ca_event(sk
, CA_EVENT_SLOW_ACK
);
3441 /* We passed data and got it acked, remove any soft error
3442 * log. Something worked...
3444 sk
->sk_err_soft
= 0;
3445 icsk
->icsk_probes_out
= 0;
3446 tp
->rcv_tstamp
= tcp_time_stamp
;
3450 /* See if we can take anything off of the retransmit queue. */
3451 acked
= tp
->packets_out
;
3452 flag
|= tcp_clean_rtx_queue(sk
, prior_fackets
, prior_snd_una
, sack_rtt
);
3453 acked
-= tp
->packets_out
;
3455 /* Advance cwnd if state allows */
3456 if (tcp_may_raise_cwnd(sk
, flag
))
3457 tcp_cong_avoid(sk
, ack
, acked
, prior_in_flight
);
3459 if (tcp_ack_is_dubious(sk
, flag
)) {
3460 is_dupack
= !(flag
& (FLAG_SND_UNA_ADVANCED
| FLAG_NOT_DUP
));
3461 tcp_fastretrans_alert(sk
, acked
, prior_unsacked
,
3464 if (tp
->tlp_high_seq
)
3465 tcp_process_tlp_ack(sk
, ack
, flag
);
3467 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
& FLAG_NOT_DUP
)) {
3468 struct dst_entry
*dst
= __sk_dst_get(sk
);
3473 if (icsk
->icsk_pending
== ICSK_TIME_RETRANS
)
3474 tcp_schedule_loss_probe(sk
);
3475 if (tp
->srtt
!= prior_rtt
|| tp
->snd_cwnd
!= prior_cwnd
)
3476 tcp_update_pacing_rate(sk
);
3480 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3481 if (flag
& FLAG_DSACKING_ACK
)
3482 tcp_fastretrans_alert(sk
, acked
, prior_unsacked
,
3484 /* If this ack opens up a zero window, clear backoff. It was
3485 * being used to time the probes, and is probably far higher than
3486 * it needs to be for normal retransmission.
3488 if (tcp_send_head(sk
))
3491 if (tp
->tlp_high_seq
)
3492 tcp_process_tlp_ack(sk
, ack
, flag
);
3496 SOCK_DEBUG(sk
, "Ack %u after %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3500 /* If data was SACKed, tag it and see if we should send more data.
3501 * If data was DSACKed, see if we can undo a cwnd reduction.
3503 if (TCP_SKB_CB(skb
)->sacked
) {
3504 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
,
3506 tcp_fastretrans_alert(sk
, acked
, prior_unsacked
,
3510 SOCK_DEBUG(sk
, "Ack %u before %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3514 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3515 * But, this can also be called on packets in the established flow when
3516 * the fast version below fails.
3518 void tcp_parse_options(const struct sk_buff
*skb
,
3519 struct tcp_options_received
*opt_rx
, int estab
,
3520 struct tcp_fastopen_cookie
*foc
)
3522 const unsigned char *ptr
;
3523 const struct tcphdr
*th
= tcp_hdr(skb
);
3524 int length
= (th
->doff
* 4) - sizeof(struct tcphdr
);
3526 ptr
= (const unsigned char *)(th
+ 1);
3527 opt_rx
->saw_tstamp
= 0;
3529 while (length
> 0) {
3530 int opcode
= *ptr
++;
3536 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
3541 if (opsize
< 2) /* "silly options" */
3543 if (opsize
> length
)
3544 return; /* don't parse partial options */
3547 if (opsize
== TCPOLEN_MSS
&& th
->syn
&& !estab
) {
3548 u16 in_mss
= get_unaligned_be16(ptr
);
3550 if (opt_rx
->user_mss
&&
3551 opt_rx
->user_mss
< in_mss
)
3552 in_mss
= opt_rx
->user_mss
;
3553 opt_rx
->mss_clamp
= in_mss
;
3558 if (opsize
== TCPOLEN_WINDOW
&& th
->syn
&&
3559 !estab
&& sysctl_tcp_window_scaling
) {
3560 __u8 snd_wscale
= *(__u8
*)ptr
;
3561 opt_rx
->wscale_ok
= 1;
3562 if (snd_wscale
> 14) {
3563 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3568 opt_rx
->snd_wscale
= snd_wscale
;
3571 case TCPOPT_TIMESTAMP
:
3572 if ((opsize
== TCPOLEN_TIMESTAMP
) &&
3573 ((estab
&& opt_rx
->tstamp_ok
) ||
3574 (!estab
&& sysctl_tcp_timestamps
))) {
3575 opt_rx
->saw_tstamp
= 1;
3576 opt_rx
->rcv_tsval
= get_unaligned_be32(ptr
);
3577 opt_rx
->rcv_tsecr
= get_unaligned_be32(ptr
+ 4);
3580 case TCPOPT_SACK_PERM
:
3581 if (opsize
== TCPOLEN_SACK_PERM
&& th
->syn
&&
3582 !estab
&& sysctl_tcp_sack
) {
3583 opt_rx
->sack_ok
= TCP_SACK_SEEN
;
3584 tcp_sack_reset(opt_rx
);
3589 if ((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
3590 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
3592 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
3595 #ifdef CONFIG_TCP_MD5SIG
3598 * The MD5 Hash has already been
3599 * checked (see tcp_v{4,6}_do_rcv()).
3604 /* Fast Open option shares code 254 using a
3605 * 16 bits magic number. It's valid only in
3606 * SYN or SYN-ACK with an even size.
3608 if (opsize
< TCPOLEN_EXP_FASTOPEN_BASE
||
3609 get_unaligned_be16(ptr
) != TCPOPT_FASTOPEN_MAGIC
||
3610 foc
== NULL
|| !th
->syn
|| (opsize
& 1))
3612 foc
->len
= opsize
- TCPOLEN_EXP_FASTOPEN_BASE
;
3613 if (foc
->len
>= TCP_FASTOPEN_COOKIE_MIN
&&
3614 foc
->len
<= TCP_FASTOPEN_COOKIE_MAX
)
3615 memcpy(foc
->val
, ptr
+ 2, foc
->len
);
3616 else if (foc
->len
!= 0)
3626 EXPORT_SYMBOL(tcp_parse_options
);
3628 static bool tcp_parse_aligned_timestamp(struct tcp_sock
*tp
, const struct tcphdr
*th
)
3630 const __be32
*ptr
= (const __be32
*)(th
+ 1);
3632 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3633 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
3634 tp
->rx_opt
.saw_tstamp
= 1;
3636 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3639 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
) - tp
->tsoffset
;
3641 tp
->rx_opt
.rcv_tsecr
= 0;
3647 /* Fast parse options. This hopes to only see timestamps.
3648 * If it is wrong it falls back on tcp_parse_options().
3650 static bool tcp_fast_parse_options(const struct sk_buff
*skb
,
3651 const struct tcphdr
*th
, struct tcp_sock
*tp
)
3653 /* In the spirit of fast parsing, compare doff directly to constant
3654 * values. Because equality is used, short doff can be ignored here.
3656 if (th
->doff
== (sizeof(*th
) / 4)) {
3657 tp
->rx_opt
.saw_tstamp
= 0;
3659 } else if (tp
->rx_opt
.tstamp_ok
&&
3660 th
->doff
== ((sizeof(*th
) + TCPOLEN_TSTAMP_ALIGNED
) / 4)) {
3661 if (tcp_parse_aligned_timestamp(tp
, th
))
3665 tcp_parse_options(skb
, &tp
->rx_opt
, 1, NULL
);
3666 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
3667 tp
->rx_opt
.rcv_tsecr
-= tp
->tsoffset
;
3672 #ifdef CONFIG_TCP_MD5SIG
3674 * Parse MD5 Signature option
3676 const u8
*tcp_parse_md5sig_option(const struct tcphdr
*th
)
3678 int length
= (th
->doff
<< 2) - sizeof(*th
);
3679 const u8
*ptr
= (const u8
*)(th
+ 1);
3681 /* If the TCP option is too short, we can short cut */
3682 if (length
< TCPOLEN_MD5SIG
)
3685 while (length
> 0) {
3686 int opcode
= *ptr
++;
3697 if (opsize
< 2 || opsize
> length
)
3699 if (opcode
== TCPOPT_MD5SIG
)
3700 return opsize
== TCPOLEN_MD5SIG
? ptr
: NULL
;
3707 EXPORT_SYMBOL(tcp_parse_md5sig_option
);
3710 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3712 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3713 * it can pass through stack. So, the following predicate verifies that
3714 * this segment is not used for anything but congestion avoidance or
3715 * fast retransmit. Moreover, we even are able to eliminate most of such
3716 * second order effects, if we apply some small "replay" window (~RTO)
3717 * to timestamp space.
3719 * All these measures still do not guarantee that we reject wrapped ACKs
3720 * on networks with high bandwidth, when sequence space is recycled fastly,
3721 * but it guarantees that such events will be very rare and do not affect
3722 * connection seriously. This doesn't look nice, but alas, PAWS is really
3725 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3726 * states that events when retransmit arrives after original data are rare.
3727 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3728 * the biggest problem on large power networks even with minor reordering.
3729 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3730 * up to bandwidth of 18Gigabit/sec. 8) ]
3733 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
3735 const struct tcp_sock
*tp
= tcp_sk(sk
);
3736 const struct tcphdr
*th
= tcp_hdr(skb
);
3737 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3738 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3740 return (/* 1. Pure ACK with correct sequence number. */
3741 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
3743 /* 2. ... and duplicate ACK. */
3744 ack
== tp
->snd_una
&&
3746 /* 3. ... and does not update window. */
3747 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
3749 /* 4. ... and sits in replay window. */
3750 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
3753 static inline bool tcp_paws_discard(const struct sock
*sk
,
3754 const struct sk_buff
*skb
)
3756 const struct tcp_sock
*tp
= tcp_sk(sk
);
3758 return !tcp_paws_check(&tp
->rx_opt
, TCP_PAWS_WINDOW
) &&
3759 !tcp_disordered_ack(sk
, skb
);
3762 /* Check segment sequence number for validity.
3764 * Segment controls are considered valid, if the segment
3765 * fits to the window after truncation to the window. Acceptability
3766 * of data (and SYN, FIN, of course) is checked separately.
3767 * See tcp_data_queue(), for example.
3769 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3770 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3771 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3772 * (borrowed from freebsd)
3775 static inline bool tcp_sequence(const struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3777 return !before(end_seq
, tp
->rcv_wup
) &&
3778 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
3781 /* When we get a reset we do this. */
3782 void tcp_reset(struct sock
*sk
)
3784 /* We want the right error as BSD sees it (and indeed as we do). */
3785 switch (sk
->sk_state
) {
3787 sk
->sk_err
= ECONNREFUSED
;
3789 case TCP_CLOSE_WAIT
:
3795 sk
->sk_err
= ECONNRESET
;
3797 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3800 if (!sock_flag(sk
, SOCK_DEAD
))
3801 sk
->sk_error_report(sk
);
3807 * Process the FIN bit. This now behaves as it is supposed to work
3808 * and the FIN takes effect when it is validly part of sequence
3809 * space. Not before when we get holes.
3811 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3812 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3815 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3816 * close and we go into CLOSING (and later onto TIME-WAIT)
3818 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3820 static void tcp_fin(struct sock
*sk
)
3822 struct tcp_sock
*tp
= tcp_sk(sk
);
3823 const struct dst_entry
*dst
;
3825 inet_csk_schedule_ack(sk
);
3827 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
3828 sock_set_flag(sk
, SOCK_DONE
);
3830 switch (sk
->sk_state
) {
3832 case TCP_ESTABLISHED
:
3833 /* Move to CLOSE_WAIT */
3834 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
3835 dst
= __sk_dst_get(sk
);
3836 if (!dst
|| !dst_metric(dst
, RTAX_QUICKACK
))
3837 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
3840 case TCP_CLOSE_WAIT
:
3842 /* Received a retransmission of the FIN, do
3847 /* RFC793: Remain in the LAST-ACK state. */
3851 /* This case occurs when a simultaneous close
3852 * happens, we must ack the received FIN and
3853 * enter the CLOSING state.
3856 tcp_set_state(sk
, TCP_CLOSING
);
3859 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3861 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
3864 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3865 * cases we should never reach this piece of code.
3867 pr_err("%s: Impossible, sk->sk_state=%d\n",
3868 __func__
, sk
->sk_state
);
3872 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3873 * Probably, we should reset in this case. For now drop them.
3875 __skb_queue_purge(&tp
->out_of_order_queue
);
3876 if (tcp_is_sack(tp
))
3877 tcp_sack_reset(&tp
->rx_opt
);
3880 if (!sock_flag(sk
, SOCK_DEAD
)) {
3881 sk
->sk_state_change(sk
);
3883 /* Do not send POLL_HUP for half duplex close. */
3884 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
3885 sk
->sk_state
== TCP_CLOSE
)
3886 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_HUP
);
3888 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_IN
);
3892 static inline bool tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
,
3895 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
3896 if (before(seq
, sp
->start_seq
))
3897 sp
->start_seq
= seq
;
3898 if (after(end_seq
, sp
->end_seq
))
3899 sp
->end_seq
= end_seq
;
3905 static void tcp_dsack_set(struct sock
*sk
, u32 seq
, u32 end_seq
)
3907 struct tcp_sock
*tp
= tcp_sk(sk
);
3909 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
3912 if (before(seq
, tp
->rcv_nxt
))
3913 mib_idx
= LINUX_MIB_TCPDSACKOLDSENT
;
3915 mib_idx
= LINUX_MIB_TCPDSACKOFOSENT
;
3917 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
3919 tp
->rx_opt
.dsack
= 1;
3920 tp
->duplicate_sack
[0].start_seq
= seq
;
3921 tp
->duplicate_sack
[0].end_seq
= end_seq
;
3925 static void tcp_dsack_extend(struct sock
*sk
, u32 seq
, u32 end_seq
)
3927 struct tcp_sock
*tp
= tcp_sk(sk
);
3929 if (!tp
->rx_opt
.dsack
)
3930 tcp_dsack_set(sk
, seq
, end_seq
);
3932 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
3935 static void tcp_send_dupack(struct sock
*sk
, const struct sk_buff
*skb
)
3937 struct tcp_sock
*tp
= tcp_sk(sk
);
3939 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
3940 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
3941 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
3942 tcp_enter_quickack_mode(sk
);
3944 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
3945 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
3947 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
3948 end_seq
= tp
->rcv_nxt
;
3949 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, end_seq
);
3956 /* These routines update the SACK block as out-of-order packets arrive or
3957 * in-order packets close up the sequence space.
3959 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
3962 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3963 struct tcp_sack_block
*swalk
= sp
+ 1;
3965 /* See if the recent change to the first SACK eats into
3966 * or hits the sequence space of other SACK blocks, if so coalesce.
3968 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
;) {
3969 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
3972 /* Zap SWALK, by moving every further SACK up by one slot.
3973 * Decrease num_sacks.
3975 tp
->rx_opt
.num_sacks
--;
3976 for (i
= this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
3980 this_sack
++, swalk
++;
3984 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
3986 struct tcp_sock
*tp
= tcp_sk(sk
);
3987 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3988 int cur_sacks
= tp
->rx_opt
.num_sacks
;
3994 for (this_sack
= 0; this_sack
< cur_sacks
; this_sack
++, sp
++) {
3995 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
3996 /* Rotate this_sack to the first one. */
3997 for (; this_sack
> 0; this_sack
--, sp
--)
3998 swap(*sp
, *(sp
- 1));
4000 tcp_sack_maybe_coalesce(tp
);
4005 /* Could not find an adjacent existing SACK, build a new one,
4006 * put it at the front, and shift everyone else down. We
4007 * always know there is at least one SACK present already here.
4009 * If the sack array is full, forget about the last one.
4011 if (this_sack
>= TCP_NUM_SACKS
) {
4013 tp
->rx_opt
.num_sacks
--;
4016 for (; this_sack
> 0; this_sack
--, sp
--)
4020 /* Build the new head SACK, and we're done. */
4021 sp
->start_seq
= seq
;
4022 sp
->end_seq
= end_seq
;
4023 tp
->rx_opt
.num_sacks
++;
4026 /* RCV.NXT advances, some SACKs should be eaten. */
4028 static void tcp_sack_remove(struct tcp_sock
*tp
)
4030 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4031 int num_sacks
= tp
->rx_opt
.num_sacks
;
4034 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4035 if (skb_queue_empty(&tp
->out_of_order_queue
)) {
4036 tp
->rx_opt
.num_sacks
= 0;
4040 for (this_sack
= 0; this_sack
< num_sacks
;) {
4041 /* Check if the start of the sack is covered by RCV.NXT. */
4042 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
4045 /* RCV.NXT must cover all the block! */
4046 WARN_ON(before(tp
->rcv_nxt
, sp
->end_seq
));
4048 /* Zap this SACK, by moving forward any other SACKS. */
4049 for (i
= this_sack
+1; i
< num_sacks
; i
++)
4050 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
4057 tp
->rx_opt
.num_sacks
= num_sacks
;
4060 /* This one checks to see if we can put data from the
4061 * out_of_order queue into the receive_queue.
4063 static void tcp_ofo_queue(struct sock
*sk
)
4065 struct tcp_sock
*tp
= tcp_sk(sk
);
4066 __u32 dsack_high
= tp
->rcv_nxt
;
4067 struct sk_buff
*skb
;
4069 while ((skb
= skb_peek(&tp
->out_of_order_queue
)) != NULL
) {
4070 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
4073 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
4074 __u32 dsack
= dsack_high
;
4075 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
4076 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
4077 tcp_dsack_extend(sk
, TCP_SKB_CB(skb
)->seq
, dsack
);
4080 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4081 SOCK_DEBUG(sk
, "ofo packet was already received\n");
4082 __skb_unlink(skb
, &tp
->out_of_order_queue
);
4086 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
4087 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4088 TCP_SKB_CB(skb
)->end_seq
);
4090 __skb_unlink(skb
, &tp
->out_of_order_queue
);
4091 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4092 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4093 if (tcp_hdr(skb
)->fin
)
4098 static bool tcp_prune_ofo_queue(struct sock
*sk
);
4099 static int tcp_prune_queue(struct sock
*sk
);
4101 static int tcp_try_rmem_schedule(struct sock
*sk
, struct sk_buff
*skb
,
4104 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
4105 !sk_rmem_schedule(sk
, skb
, size
)) {
4107 if (tcp_prune_queue(sk
) < 0)
4110 if (!sk_rmem_schedule(sk
, skb
, size
)) {
4111 if (!tcp_prune_ofo_queue(sk
))
4114 if (!sk_rmem_schedule(sk
, skb
, size
))
4122 * tcp_try_coalesce - try to merge skb to prior one
4125 * @from: buffer to add in queue
4126 * @fragstolen: pointer to boolean
4128 * Before queueing skb @from after @to, try to merge them
4129 * to reduce overall memory use and queue lengths, if cost is small.
4130 * Packets in ofo or receive queues can stay a long time.
4131 * Better try to coalesce them right now to avoid future collapses.
4132 * Returns true if caller should free @from instead of queueing it
4134 static bool tcp_try_coalesce(struct sock
*sk
,
4136 struct sk_buff
*from
,
4141 *fragstolen
= false;
4143 if (tcp_hdr(from
)->fin
)
4146 /* Its possible this segment overlaps with prior segment in queue */
4147 if (TCP_SKB_CB(from
)->seq
!= TCP_SKB_CB(to
)->end_seq
)
4150 if (!skb_try_coalesce(to
, from
, fragstolen
, &delta
))
4153 atomic_add(delta
, &sk
->sk_rmem_alloc
);
4154 sk_mem_charge(sk
, delta
);
4155 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRCVCOALESCE
);
4156 TCP_SKB_CB(to
)->end_seq
= TCP_SKB_CB(from
)->end_seq
;
4157 TCP_SKB_CB(to
)->ack_seq
= TCP_SKB_CB(from
)->ack_seq
;
4161 static void tcp_data_queue_ofo(struct sock
*sk
, struct sk_buff
*skb
)
4163 struct tcp_sock
*tp
= tcp_sk(sk
);
4164 struct sk_buff
*skb1
;
4167 TCP_ECN_check_ce(tp
, skb
);
4169 if (unlikely(tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))) {
4170 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFODROP
);
4175 /* Disable header prediction. */
4177 inet_csk_schedule_ack(sk
);
4179 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFOQUEUE
);
4180 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
4181 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4183 skb1
= skb_peek_tail(&tp
->out_of_order_queue
);
4185 /* Initial out of order segment, build 1 SACK. */
4186 if (tcp_is_sack(tp
)) {
4187 tp
->rx_opt
.num_sacks
= 1;
4188 tp
->selective_acks
[0].start_seq
= TCP_SKB_CB(skb
)->seq
;
4189 tp
->selective_acks
[0].end_seq
=
4190 TCP_SKB_CB(skb
)->end_seq
;
4192 __skb_queue_head(&tp
->out_of_order_queue
, skb
);
4196 seq
= TCP_SKB_CB(skb
)->seq
;
4197 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4199 if (seq
== TCP_SKB_CB(skb1
)->end_seq
) {
4202 if (!tcp_try_coalesce(sk
, skb1
, skb
, &fragstolen
)) {
4203 __skb_queue_after(&tp
->out_of_order_queue
, skb1
, skb
);
4205 tcp_grow_window(sk
, skb
);
4206 kfree_skb_partial(skb
, fragstolen
);
4210 if (!tp
->rx_opt
.num_sacks
||
4211 tp
->selective_acks
[0].end_seq
!= seq
)
4214 /* Common case: data arrive in order after hole. */
4215 tp
->selective_acks
[0].end_seq
= end_seq
;
4219 /* Find place to insert this segment. */
4221 if (!after(TCP_SKB_CB(skb1
)->seq
, seq
))
4223 if (skb_queue_is_first(&tp
->out_of_order_queue
, skb1
)) {
4227 skb1
= skb_queue_prev(&tp
->out_of_order_queue
, skb1
);
4230 /* Do skb overlap to previous one? */
4231 if (skb1
&& before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4232 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4233 /* All the bits are present. Drop. */
4234 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFOMERGE
);
4237 tcp_dsack_set(sk
, seq
, end_seq
);
4240 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
4241 /* Partial overlap. */
4242 tcp_dsack_set(sk
, seq
,
4243 TCP_SKB_CB(skb1
)->end_seq
);
4245 if (skb_queue_is_first(&tp
->out_of_order_queue
,
4249 skb1
= skb_queue_prev(
4250 &tp
->out_of_order_queue
,
4255 __skb_queue_head(&tp
->out_of_order_queue
, skb
);
4257 __skb_queue_after(&tp
->out_of_order_queue
, skb1
, skb
);
4259 /* And clean segments covered by new one as whole. */
4260 while (!skb_queue_is_last(&tp
->out_of_order_queue
, skb
)) {
4261 skb1
= skb_queue_next(&tp
->out_of_order_queue
, skb
);
4263 if (!after(end_seq
, TCP_SKB_CB(skb1
)->seq
))
4265 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4266 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4270 __skb_unlink(skb1
, &tp
->out_of_order_queue
);
4271 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4272 TCP_SKB_CB(skb1
)->end_seq
);
4273 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFOMERGE
);
4278 if (tcp_is_sack(tp
))
4279 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
4282 tcp_grow_window(sk
, skb
);
4283 skb_set_owner_r(skb
, sk
);
4287 static int __must_check
tcp_queue_rcv(struct sock
*sk
, struct sk_buff
*skb
, int hdrlen
,
4291 struct sk_buff
*tail
= skb_peek_tail(&sk
->sk_receive_queue
);
4293 __skb_pull(skb
, hdrlen
);
4295 tcp_try_coalesce(sk
, tail
, skb
, fragstolen
)) ? 1 : 0;
4296 tcp_sk(sk
)->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4298 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4299 skb_set_owner_r(skb
, sk
);
4304 int tcp_send_rcvq(struct sock
*sk
, struct msghdr
*msg
, size_t size
)
4306 struct sk_buff
*skb
= NULL
;
4313 skb
= alloc_skb(size
+ sizeof(*th
), sk
->sk_allocation
);
4317 if (tcp_try_rmem_schedule(sk
, skb
, size
+ sizeof(*th
)))
4320 th
= (struct tcphdr
*)skb_put(skb
, sizeof(*th
));
4321 skb_reset_transport_header(skb
);
4322 memset(th
, 0, sizeof(*th
));
4324 if (memcpy_fromiovec(skb_put(skb
, size
), msg
->msg_iov
, size
))
4327 TCP_SKB_CB(skb
)->seq
= tcp_sk(sk
)->rcv_nxt
;
4328 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ size
;
4329 TCP_SKB_CB(skb
)->ack_seq
= tcp_sk(sk
)->snd_una
- 1;
4331 if (tcp_queue_rcv(sk
, skb
, sizeof(*th
), &fragstolen
)) {
4332 WARN_ON_ONCE(fragstolen
); /* should not happen */
4343 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
4345 const struct tcphdr
*th
= tcp_hdr(skb
);
4346 struct tcp_sock
*tp
= tcp_sk(sk
);
4348 bool fragstolen
= false;
4350 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
)
4354 __skb_pull(skb
, th
->doff
* 4);
4356 TCP_ECN_accept_cwr(tp
, skb
);
4358 tp
->rx_opt
.dsack
= 0;
4360 /* Queue data for delivery to the user.
4361 * Packets in sequence go to the receive queue.
4362 * Out of sequence packets to the out_of_order_queue.
4364 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
4365 if (tcp_receive_window(tp
) == 0)
4368 /* Ok. In sequence. In window. */
4369 if (tp
->ucopy
.task
== current
&&
4370 tp
->copied_seq
== tp
->rcv_nxt
&& tp
->ucopy
.len
&&
4371 sock_owned_by_user(sk
) && !tp
->urg_data
) {
4372 int chunk
= min_t(unsigned int, skb
->len
,
4375 __set_current_state(TASK_RUNNING
);
4378 if (!skb_copy_datagram_iovec(skb
, 0, tp
->ucopy
.iov
, chunk
)) {
4379 tp
->ucopy
.len
-= chunk
;
4380 tp
->copied_seq
+= chunk
;
4381 eaten
= (chunk
== skb
->len
);
4382 tcp_rcv_space_adjust(sk
);
4390 tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))
4393 eaten
= tcp_queue_rcv(sk
, skb
, 0, &fragstolen
);
4395 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4397 tcp_event_data_recv(sk
, skb
);
4401 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4404 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4405 * gap in queue is filled.
4407 if (skb_queue_empty(&tp
->out_of_order_queue
))
4408 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
4411 if (tp
->rx_opt
.num_sacks
)
4412 tcp_sack_remove(tp
);
4414 tcp_fast_path_check(sk
);
4417 kfree_skb_partial(skb
, fragstolen
);
4418 if (!sock_flag(sk
, SOCK_DEAD
))
4419 sk
->sk_data_ready(sk
, 0);
4423 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4424 /* A retransmit, 2nd most common case. Force an immediate ack. */
4425 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4426 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4429 tcp_enter_quickack_mode(sk
);
4430 inet_csk_schedule_ack(sk
);
4436 /* Out of window. F.e. zero window probe. */
4437 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
4440 tcp_enter_quickack_mode(sk
);
4442 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4443 /* Partial packet, seq < rcv_next < end_seq */
4444 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
4445 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4446 TCP_SKB_CB(skb
)->end_seq
);
4448 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
4450 /* If window is closed, drop tail of packet. But after
4451 * remembering D-SACK for its head made in previous line.
4453 if (!tcp_receive_window(tp
))
4458 tcp_data_queue_ofo(sk
, skb
);
4461 static struct sk_buff
*tcp_collapse_one(struct sock
*sk
, struct sk_buff
*skb
,
4462 struct sk_buff_head
*list
)
4464 struct sk_buff
*next
= NULL
;
4466 if (!skb_queue_is_last(list
, skb
))
4467 next
= skb_queue_next(list
, skb
);
4469 __skb_unlink(skb
, list
);
4471 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRCVCOLLAPSED
);
4476 /* Collapse contiguous sequence of skbs head..tail with
4477 * sequence numbers start..end.
4479 * If tail is NULL, this means until the end of the list.
4481 * Segments with FIN/SYN are not collapsed (only because this
4485 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
,
4486 struct sk_buff
*head
, struct sk_buff
*tail
,
4489 struct sk_buff
*skb
, *n
;
4492 /* First, check that queue is collapsible and find
4493 * the point where collapsing can be useful. */
4497 skb_queue_walk_from_safe(list
, skb
, n
) {
4500 /* No new bits? It is possible on ofo queue. */
4501 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4502 skb
= tcp_collapse_one(sk
, skb
, list
);
4508 /* The first skb to collapse is:
4510 * - bloated or contains data before "start" or
4511 * overlaps to the next one.
4513 if (!tcp_hdr(skb
)->syn
&& !tcp_hdr(skb
)->fin
&&
4514 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
4515 before(TCP_SKB_CB(skb
)->seq
, start
))) {
4516 end_of_skbs
= false;
4520 if (!skb_queue_is_last(list
, skb
)) {
4521 struct sk_buff
*next
= skb_queue_next(list
, skb
);
4523 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(next
)->seq
) {
4524 end_of_skbs
= false;
4529 /* Decided to skip this, advance start seq. */
4530 start
= TCP_SKB_CB(skb
)->end_seq
;
4532 if (end_of_skbs
|| tcp_hdr(skb
)->syn
|| tcp_hdr(skb
)->fin
)
4535 while (before(start
, end
)) {
4536 struct sk_buff
*nskb
;
4537 unsigned int header
= skb_headroom(skb
);
4538 int copy
= SKB_MAX_ORDER(header
, 0);
4540 /* Too big header? This can happen with IPv6. */
4543 if (end
- start
< copy
)
4545 nskb
= alloc_skb(copy
+ header
, GFP_ATOMIC
);
4549 skb_set_mac_header(nskb
, skb_mac_header(skb
) - skb
->head
);
4550 skb_set_network_header(nskb
, (skb_network_header(skb
) -
4552 skb_set_transport_header(nskb
, (skb_transport_header(skb
) -
4554 skb_reserve(nskb
, header
);
4555 memcpy(nskb
->head
, skb
->head
, header
);
4556 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
4557 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
4558 __skb_queue_before(list
, skb
, nskb
);
4559 skb_set_owner_r(nskb
, sk
);
4561 /* Copy data, releasing collapsed skbs. */
4563 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
4564 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
4568 size
= min(copy
, size
);
4569 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
4571 TCP_SKB_CB(nskb
)->end_seq
+= size
;
4575 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4576 skb
= tcp_collapse_one(sk
, skb
, list
);
4579 tcp_hdr(skb
)->syn
||
4587 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4588 * and tcp_collapse() them until all the queue is collapsed.
4590 static void tcp_collapse_ofo_queue(struct sock
*sk
)
4592 struct tcp_sock
*tp
= tcp_sk(sk
);
4593 struct sk_buff
*skb
= skb_peek(&tp
->out_of_order_queue
);
4594 struct sk_buff
*head
;
4600 start
= TCP_SKB_CB(skb
)->seq
;
4601 end
= TCP_SKB_CB(skb
)->end_seq
;
4605 struct sk_buff
*next
= NULL
;
4607 if (!skb_queue_is_last(&tp
->out_of_order_queue
, skb
))
4608 next
= skb_queue_next(&tp
->out_of_order_queue
, skb
);
4611 /* Segment is terminated when we see gap or when
4612 * we are at the end of all the queue. */
4614 after(TCP_SKB_CB(skb
)->seq
, end
) ||
4615 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
4616 tcp_collapse(sk
, &tp
->out_of_order_queue
,
4617 head
, skb
, start
, end
);
4621 /* Start new segment */
4622 start
= TCP_SKB_CB(skb
)->seq
;
4623 end
= TCP_SKB_CB(skb
)->end_seq
;
4625 if (before(TCP_SKB_CB(skb
)->seq
, start
))
4626 start
= TCP_SKB_CB(skb
)->seq
;
4627 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
4628 end
= TCP_SKB_CB(skb
)->end_seq
;
4634 * Purge the out-of-order queue.
4635 * Return true if queue was pruned.
4637 static bool tcp_prune_ofo_queue(struct sock
*sk
)
4639 struct tcp_sock
*tp
= tcp_sk(sk
);
4642 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4643 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_OFOPRUNED
);
4644 __skb_queue_purge(&tp
->out_of_order_queue
);
4646 /* Reset SACK state. A conforming SACK implementation will
4647 * do the same at a timeout based retransmit. When a connection
4648 * is in a sad state like this, we care only about integrity
4649 * of the connection not performance.
4651 if (tp
->rx_opt
.sack_ok
)
4652 tcp_sack_reset(&tp
->rx_opt
);
4659 /* Reduce allocated memory if we can, trying to get
4660 * the socket within its memory limits again.
4662 * Return less than zero if we should start dropping frames
4663 * until the socket owning process reads some of the data
4664 * to stabilize the situation.
4666 static int tcp_prune_queue(struct sock
*sk
)
4668 struct tcp_sock
*tp
= tcp_sk(sk
);
4670 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
4672 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PRUNECALLED
);
4674 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
4675 tcp_clamp_window(sk
);
4676 else if (sk_under_memory_pressure(sk
))
4677 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
4679 tcp_collapse_ofo_queue(sk
);
4680 if (!skb_queue_empty(&sk
->sk_receive_queue
))
4681 tcp_collapse(sk
, &sk
->sk_receive_queue
,
4682 skb_peek(&sk
->sk_receive_queue
),
4684 tp
->copied_seq
, tp
->rcv_nxt
);
4687 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4690 /* Collapsing did not help, destructive actions follow.
4691 * This must not ever occur. */
4693 tcp_prune_ofo_queue(sk
);
4695 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4698 /* If we are really being abused, tell the caller to silently
4699 * drop receive data on the floor. It will get retransmitted
4700 * and hopefully then we'll have sufficient space.
4702 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_RCVPRUNED
);
4704 /* Massive buffer overcommit. */
4709 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4710 * As additional protections, we do not touch cwnd in retransmission phases,
4711 * and if application hit its sndbuf limit recently.
4713 void tcp_cwnd_application_limited(struct sock
*sk
)
4715 struct tcp_sock
*tp
= tcp_sk(sk
);
4717 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
4718 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
4719 /* Limited by application or receiver window. */
4720 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
4721 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
4722 if (win_used
< tp
->snd_cwnd
) {
4723 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
4724 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
4726 tp
->snd_cwnd_used
= 0;
4728 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4731 static bool tcp_should_expand_sndbuf(const struct sock
*sk
)
4733 const struct tcp_sock
*tp
= tcp_sk(sk
);
4735 /* If the user specified a specific send buffer setting, do
4738 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
4741 /* If we are under global TCP memory pressure, do not expand. */
4742 if (sk_under_memory_pressure(sk
))
4745 /* If we are under soft global TCP memory pressure, do not expand. */
4746 if (sk_memory_allocated(sk
) >= sk_prot_mem_limits(sk
, 0))
4749 /* If we filled the congestion window, do not expand. */
4750 if (tp
->packets_out
>= tp
->snd_cwnd
)
4756 /* When incoming ACK allowed to free some skb from write_queue,
4757 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4758 * on the exit from tcp input handler.
4760 * PROBLEM: sndbuf expansion does not work well with largesend.
4762 static void tcp_new_space(struct sock
*sk
)
4764 struct tcp_sock
*tp
= tcp_sk(sk
);
4766 if (tcp_should_expand_sndbuf(sk
)) {
4767 tcp_sndbuf_expand(sk
);
4768 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4771 sk
->sk_write_space(sk
);
4774 static void tcp_check_space(struct sock
*sk
)
4776 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
4777 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
4778 if (sk
->sk_socket
&&
4779 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
4784 static inline void tcp_data_snd_check(struct sock
*sk
)
4786 tcp_push_pending_frames(sk
);
4787 tcp_check_space(sk
);
4791 * Check if sending an ack is needed.
4793 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
4795 struct tcp_sock
*tp
= tcp_sk(sk
);
4797 /* More than one full frame received... */
4798 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
&&
4799 /* ... and right edge of window advances far enough.
4800 * (tcp_recvmsg() will send ACK otherwise). Or...
4802 __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
4803 /* We ACK each frame or... */
4804 tcp_in_quickack_mode(sk
) ||
4805 /* We have out of order data. */
4806 (ofo_possible
&& skb_peek(&tp
->out_of_order_queue
))) {
4807 /* Then ack it now */
4810 /* Else, send delayed ack. */
4811 tcp_send_delayed_ack(sk
);
4815 static inline void tcp_ack_snd_check(struct sock
*sk
)
4817 if (!inet_csk_ack_scheduled(sk
)) {
4818 /* We sent a data segment already. */
4821 __tcp_ack_snd_check(sk
, 1);
4825 * This routine is only called when we have urgent data
4826 * signaled. Its the 'slow' part of tcp_urg. It could be
4827 * moved inline now as tcp_urg is only called from one
4828 * place. We handle URGent data wrong. We have to - as
4829 * BSD still doesn't use the correction from RFC961.
4830 * For 1003.1g we should support a new option TCP_STDURG to permit
4831 * either form (or just set the sysctl tcp_stdurg).
4834 static void tcp_check_urg(struct sock
*sk
, const struct tcphdr
*th
)
4836 struct tcp_sock
*tp
= tcp_sk(sk
);
4837 u32 ptr
= ntohs(th
->urg_ptr
);
4839 if (ptr
&& !sysctl_tcp_stdurg
)
4841 ptr
+= ntohl(th
->seq
);
4843 /* Ignore urgent data that we've already seen and read. */
4844 if (after(tp
->copied_seq
, ptr
))
4847 /* Do not replay urg ptr.
4849 * NOTE: interesting situation not covered by specs.
4850 * Misbehaving sender may send urg ptr, pointing to segment,
4851 * which we already have in ofo queue. We are not able to fetch
4852 * such data and will stay in TCP_URG_NOTYET until will be eaten
4853 * by recvmsg(). Seems, we are not obliged to handle such wicked
4854 * situations. But it is worth to think about possibility of some
4855 * DoSes using some hypothetical application level deadlock.
4857 if (before(ptr
, tp
->rcv_nxt
))
4860 /* Do we already have a newer (or duplicate) urgent pointer? */
4861 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
4864 /* Tell the world about our new urgent pointer. */
4867 /* We may be adding urgent data when the last byte read was
4868 * urgent. To do this requires some care. We cannot just ignore
4869 * tp->copied_seq since we would read the last urgent byte again
4870 * as data, nor can we alter copied_seq until this data arrives
4871 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4873 * NOTE. Double Dutch. Rendering to plain English: author of comment
4874 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4875 * and expect that both A and B disappear from stream. This is _wrong_.
4876 * Though this happens in BSD with high probability, this is occasional.
4877 * Any application relying on this is buggy. Note also, that fix "works"
4878 * only in this artificial test. Insert some normal data between A and B and we will
4879 * decline of BSD again. Verdict: it is better to remove to trap
4882 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
4883 !sock_flag(sk
, SOCK_URGINLINE
) && tp
->copied_seq
!= tp
->rcv_nxt
) {
4884 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
4886 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4887 __skb_unlink(skb
, &sk
->sk_receive_queue
);
4892 tp
->urg_data
= TCP_URG_NOTYET
;
4895 /* Disable header prediction. */
4899 /* This is the 'fast' part of urgent handling. */
4900 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, const struct tcphdr
*th
)
4902 struct tcp_sock
*tp
= tcp_sk(sk
);
4904 /* Check if we get a new urgent pointer - normally not. */
4906 tcp_check_urg(sk
, th
);
4908 /* Do we wait for any urgent data? - normally not... */
4909 if (tp
->urg_data
== TCP_URG_NOTYET
) {
4910 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
4913 /* Is the urgent pointer pointing into this packet? */
4914 if (ptr
< skb
->len
) {
4916 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
4918 tp
->urg_data
= TCP_URG_VALID
| tmp
;
4919 if (!sock_flag(sk
, SOCK_DEAD
))
4920 sk
->sk_data_ready(sk
, 0);
4925 static int tcp_copy_to_iovec(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
4927 struct tcp_sock
*tp
= tcp_sk(sk
);
4928 int chunk
= skb
->len
- hlen
;
4932 if (skb_csum_unnecessary(skb
))
4933 err
= skb_copy_datagram_iovec(skb
, hlen
, tp
->ucopy
.iov
, chunk
);
4935 err
= skb_copy_and_csum_datagram_iovec(skb
, hlen
,
4939 tp
->ucopy
.len
-= chunk
;
4940 tp
->copied_seq
+= chunk
;
4941 tcp_rcv_space_adjust(sk
);
4948 static __sum16
__tcp_checksum_complete_user(struct sock
*sk
,
4949 struct sk_buff
*skb
)
4953 if (sock_owned_by_user(sk
)) {
4955 result
= __tcp_checksum_complete(skb
);
4958 result
= __tcp_checksum_complete(skb
);
4963 static inline bool tcp_checksum_complete_user(struct sock
*sk
,
4964 struct sk_buff
*skb
)
4966 return !skb_csum_unnecessary(skb
) &&
4967 __tcp_checksum_complete_user(sk
, skb
);
4970 #ifdef CONFIG_NET_DMA
4971 static bool tcp_dma_try_early_copy(struct sock
*sk
, struct sk_buff
*skb
,
4974 struct tcp_sock
*tp
= tcp_sk(sk
);
4975 int chunk
= skb
->len
- hlen
;
4977 bool copied_early
= false;
4979 if (tp
->ucopy
.wakeup
)
4982 if (!tp
->ucopy
.dma_chan
&& tp
->ucopy
.pinned_list
)
4983 tp
->ucopy
.dma_chan
= net_dma_find_channel();
4985 if (tp
->ucopy
.dma_chan
&& skb_csum_unnecessary(skb
)) {
4987 dma_cookie
= dma_skb_copy_datagram_iovec(tp
->ucopy
.dma_chan
,
4989 tp
->ucopy
.iov
, chunk
,
4990 tp
->ucopy
.pinned_list
);
4995 tp
->ucopy
.dma_cookie
= dma_cookie
;
4996 copied_early
= true;
4998 tp
->ucopy
.len
-= chunk
;
4999 tp
->copied_seq
+= chunk
;
5000 tcp_rcv_space_adjust(sk
);
5002 if ((tp
->ucopy
.len
== 0) ||
5003 (tcp_flag_word(tcp_hdr(skb
)) & TCP_FLAG_PSH
) ||
5004 (atomic_read(&sk
->sk_rmem_alloc
) > (sk
->sk_rcvbuf
>> 1))) {
5005 tp
->ucopy
.wakeup
= 1;
5006 sk
->sk_data_ready(sk
, 0);
5008 } else if (chunk
> 0) {
5009 tp
->ucopy
.wakeup
= 1;
5010 sk
->sk_data_ready(sk
, 0);
5013 return copied_early
;
5015 #endif /* CONFIG_NET_DMA */
5017 /* Does PAWS and seqno based validation of an incoming segment, flags will
5018 * play significant role here.
5020 static bool tcp_validate_incoming(struct sock
*sk
, struct sk_buff
*skb
,
5021 const struct tcphdr
*th
, int syn_inerr
)
5023 struct tcp_sock
*tp
= tcp_sk(sk
);
5025 /* RFC1323: H1. Apply PAWS check first. */
5026 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
5027 tcp_paws_discard(sk
, skb
)) {
5029 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSESTABREJECTED
);
5030 tcp_send_dupack(sk
, skb
);
5033 /* Reset is accepted even if it did not pass PAWS. */
5036 /* Step 1: check sequence number */
5037 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5038 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5039 * (RST) segments are validated by checking their SEQ-fields."
5040 * And page 69: "If an incoming segment is not acceptable,
5041 * an acknowledgment should be sent in reply (unless the RST
5042 * bit is set, if so drop the segment and return)".
5047 tcp_send_dupack(sk
, skb
);
5052 /* Step 2: check RST bit */
5055 * If sequence number exactly matches RCV.NXT, then
5056 * RESET the connection
5058 * Send a challenge ACK
5060 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
)
5063 tcp_send_challenge_ack(sk
);
5067 /* step 3: check security and precedence [ignored] */
5069 /* step 4: Check for a SYN
5070 * RFC 5691 4.2 : Send a challenge ack
5075 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5076 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSYNCHALLENGE
);
5077 tcp_send_challenge_ack(sk
);
5089 * TCP receive function for the ESTABLISHED state.
5091 * It is split into a fast path and a slow path. The fast path is
5093 * - A zero window was announced from us - zero window probing
5094 * is only handled properly in the slow path.
5095 * - Out of order segments arrived.
5096 * - Urgent data is expected.
5097 * - There is no buffer space left
5098 * - Unexpected TCP flags/window values/header lengths are received
5099 * (detected by checking the TCP header against pred_flags)
5100 * - Data is sent in both directions. Fast path only supports pure senders
5101 * or pure receivers (this means either the sequence number or the ack
5102 * value must stay constant)
5103 * - Unexpected TCP option.
5105 * When these conditions are not satisfied it drops into a standard
5106 * receive procedure patterned after RFC793 to handle all cases.
5107 * The first three cases are guaranteed by proper pred_flags setting,
5108 * the rest is checked inline. Fast processing is turned on in
5109 * tcp_data_queue when everything is OK.
5111 void tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
5112 const struct tcphdr
*th
, unsigned int len
)
5114 struct tcp_sock
*tp
= tcp_sk(sk
);
5116 if (unlikely(sk
->sk_rx_dst
== NULL
))
5117 inet_csk(sk
)->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
5119 * Header prediction.
5120 * The code loosely follows the one in the famous
5121 * "30 instruction TCP receive" Van Jacobson mail.
5123 * Van's trick is to deposit buffers into socket queue
5124 * on a device interrupt, to call tcp_recv function
5125 * on the receive process context and checksum and copy
5126 * the buffer to user space. smart...
5128 * Our current scheme is not silly either but we take the
5129 * extra cost of the net_bh soft interrupt processing...
5130 * We do checksum and copy also but from device to kernel.
5133 tp
->rx_opt
.saw_tstamp
= 0;
5135 /* pred_flags is 0xS?10 << 16 + snd_wnd
5136 * if header_prediction is to be made
5137 * 'S' will always be tp->tcp_header_len >> 2
5138 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5139 * turn it off (when there are holes in the receive
5140 * space for instance)
5141 * PSH flag is ignored.
5144 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
5145 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
&&
5146 !after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
)) {
5147 int tcp_header_len
= tp
->tcp_header_len
;
5149 /* Timestamp header prediction: tcp_header_len
5150 * is automatically equal to th->doff*4 due to pred_flags
5154 /* Check timestamp */
5155 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
5156 /* No? Slow path! */
5157 if (!tcp_parse_aligned_timestamp(tp
, th
))
5160 /* If PAWS failed, check it more carefully in slow path */
5161 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
5164 /* DO NOT update ts_recent here, if checksum fails
5165 * and timestamp was corrupted part, it will result
5166 * in a hung connection since we will drop all
5167 * future packets due to the PAWS test.
5171 if (len
<= tcp_header_len
) {
5172 /* Bulk data transfer: sender */
5173 if (len
== tcp_header_len
) {
5174 /* Predicted packet is in window by definition.
5175 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5176 * Hence, check seq<=rcv_wup reduces to:
5178 if (tcp_header_len
==
5179 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5180 tp
->rcv_nxt
== tp
->rcv_wup
)
5181 tcp_store_ts_recent(tp
);
5183 /* We know that such packets are checksummed
5186 tcp_ack(sk
, skb
, 0);
5188 tcp_data_snd_check(sk
);
5190 } else { /* Header too small */
5191 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5196 int copied_early
= 0;
5197 bool fragstolen
= false;
5199 if (tp
->copied_seq
== tp
->rcv_nxt
&&
5200 len
- tcp_header_len
<= tp
->ucopy
.len
) {
5201 #ifdef CONFIG_NET_DMA
5202 if (tp
->ucopy
.task
== current
&&
5203 sock_owned_by_user(sk
) &&
5204 tcp_dma_try_early_copy(sk
, skb
, tcp_header_len
)) {
5209 if (tp
->ucopy
.task
== current
&&
5210 sock_owned_by_user(sk
) && !copied_early
) {
5211 __set_current_state(TASK_RUNNING
);
5213 if (!tcp_copy_to_iovec(sk
, skb
, tcp_header_len
))
5217 /* Predicted packet is in window by definition.
5218 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5219 * Hence, check seq<=rcv_wup reduces to:
5221 if (tcp_header_len
==
5222 (sizeof(struct tcphdr
) +
5223 TCPOLEN_TSTAMP_ALIGNED
) &&
5224 tp
->rcv_nxt
== tp
->rcv_wup
)
5225 tcp_store_ts_recent(tp
);
5227 tcp_rcv_rtt_measure_ts(sk
, skb
);
5229 __skb_pull(skb
, tcp_header_len
);
5230 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
5231 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITSTOUSER
);
5234 tcp_cleanup_rbuf(sk
, skb
->len
);
5237 if (tcp_checksum_complete_user(sk
, skb
))
5240 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
5243 /* Predicted packet is in window by definition.
5244 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5245 * Hence, check seq<=rcv_wup reduces to:
5247 if (tcp_header_len
==
5248 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5249 tp
->rcv_nxt
== tp
->rcv_wup
)
5250 tcp_store_ts_recent(tp
);
5252 tcp_rcv_rtt_measure_ts(sk
, skb
);
5254 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITS
);
5256 /* Bulk data transfer: receiver */
5257 eaten
= tcp_queue_rcv(sk
, skb
, tcp_header_len
,
5261 tcp_event_data_recv(sk
, skb
);
5263 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
5264 /* Well, only one small jumplet in fast path... */
5265 tcp_ack(sk
, skb
, FLAG_DATA
);
5266 tcp_data_snd_check(sk
);
5267 if (!inet_csk_ack_scheduled(sk
))
5271 if (!copied_early
|| tp
->rcv_nxt
!= tp
->rcv_wup
)
5272 __tcp_ack_snd_check(sk
, 0);
5274 #ifdef CONFIG_NET_DMA
5276 __skb_queue_tail(&sk
->sk_async_wait_queue
, skb
);
5280 kfree_skb_partial(skb
, fragstolen
);
5281 sk
->sk_data_ready(sk
, 0);
5287 if (len
< (th
->doff
<< 2) || tcp_checksum_complete_user(sk
, skb
))
5290 if (!th
->ack
&& !th
->rst
)
5294 * Standard slow path.
5297 if (!tcp_validate_incoming(sk
, skb
, th
, 1))
5301 if (tcp_ack(sk
, skb
, FLAG_SLOWPATH
| FLAG_UPDATE_TS_RECENT
) < 0)
5304 tcp_rcv_rtt_measure_ts(sk
, skb
);
5306 /* Process urgent data. */
5307 tcp_urg(sk
, skb
, th
);
5309 /* step 7: process the segment text */
5310 tcp_data_queue(sk
, skb
);
5312 tcp_data_snd_check(sk
);
5313 tcp_ack_snd_check(sk
);
5317 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_CSUMERRORS
);
5318 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5323 EXPORT_SYMBOL(tcp_rcv_established
);
5325 void tcp_finish_connect(struct sock
*sk
, struct sk_buff
*skb
)
5327 struct tcp_sock
*tp
= tcp_sk(sk
);
5328 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5330 tcp_set_state(sk
, TCP_ESTABLISHED
);
5333 icsk
->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
5334 security_inet_conn_established(sk
, skb
);
5337 /* Make sure socket is routed, for correct metrics. */
5338 icsk
->icsk_af_ops
->rebuild_header(sk
);
5340 tcp_init_metrics(sk
);
5342 tcp_init_congestion_control(sk
);
5344 /* Prevent spurious tcp_cwnd_restart() on first data
5347 tp
->lsndtime
= tcp_time_stamp
;
5349 tcp_init_buffer_space(sk
);
5351 if (sock_flag(sk
, SOCK_KEEPOPEN
))
5352 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
5354 if (!tp
->rx_opt
.snd_wscale
)
5355 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
5359 if (!sock_flag(sk
, SOCK_DEAD
)) {
5360 sk
->sk_state_change(sk
);
5361 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5365 static bool tcp_rcv_fastopen_synack(struct sock
*sk
, struct sk_buff
*synack
,
5366 struct tcp_fastopen_cookie
*cookie
)
5368 struct tcp_sock
*tp
= tcp_sk(sk
);
5369 struct sk_buff
*data
= tp
->syn_data
? tcp_write_queue_head(sk
) : NULL
;
5370 u16 mss
= tp
->rx_opt
.mss_clamp
;
5373 if (mss
== tp
->rx_opt
.user_mss
) {
5374 struct tcp_options_received opt
;
5376 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5377 tcp_clear_options(&opt
);
5378 opt
.user_mss
= opt
.mss_clamp
= 0;
5379 tcp_parse_options(synack
, &opt
, 0, NULL
);
5380 mss
= opt
.mss_clamp
;
5383 if (!tp
->syn_fastopen
) /* Ignore an unsolicited cookie */
5386 /* The SYN-ACK neither has cookie nor acknowledges the data. Presumably
5387 * the remote receives only the retransmitted (regular) SYNs: either
5388 * the original SYN-data or the corresponding SYN-ACK is lost.
5390 syn_drop
= (cookie
->len
<= 0 && data
&& tp
->total_retrans
);
5392 tcp_fastopen_cache_set(sk
, mss
, cookie
, syn_drop
);
5394 if (data
) { /* Retransmit unacked data in SYN */
5395 tcp_for_write_queue_from(data
, sk
) {
5396 if (data
== tcp_send_head(sk
) ||
5397 __tcp_retransmit_skb(sk
, data
))
5403 tp
->syn_data_acked
= tp
->syn_data
;
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
);