2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
23 * Pedro Roque : Fast Retransmit/Recovery.
25 * Retransmit queue handled by TCP.
26 * Better retransmit timer handling.
27 * New congestion avoidance.
31 * Eric : Fast Retransmit.
32 * Randy Scott : MSS option defines.
33 * Eric Schenk : Fixes to slow start algorithm.
34 * Eric Schenk : Yet another double ACK bug.
35 * Eric Schenk : Delayed ACK bug fixes.
36 * Eric Schenk : Floyd style fast retrans war avoidance.
37 * David S. Miller : Don't allow zero congestion window.
38 * Eric Schenk : Fix retransmitter so that it sends
39 * next packet on ack of previous packet.
40 * Andi Kleen : Moved open_request checking here
41 * and process RSTs for open_requests.
42 * Andi Kleen : Better prune_queue, and other fixes.
43 * Andrey Savochkin: Fix RTT measurements in the presence of
45 * Andrey Savochkin: Check sequence numbers correctly when
46 * removing SACKs due to in sequence incoming
48 * Andi Kleen: Make sure we never ack data there is not
49 * enough room for. Also make this condition
50 * a fatal error if it might still happen.
51 * Andi Kleen: Add tcp_measure_rcv_mss to make
52 * connections with MSS<min(MTU,ann. MSS)
53 * work without delayed acks.
54 * Andi Kleen: Process packets with PSH set in the
56 * J Hadi Salim: ECN support
59 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
60 * engine. Lots of bugs are found.
61 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
64 #define pr_fmt(fmt) "TCP: " fmt
67 #include <linux/slab.h>
68 #include <linux/module.h>
69 #include <linux/sysctl.h>
70 #include <linux/kernel.h>
71 #include <linux/prefetch.h>
74 #include <net/inet_common.h>
75 #include <linux/ipsec.h>
76 #include <asm/unaligned.h>
77 #include <linux/errqueue.h>
79 int sysctl_tcp_timestamps __read_mostly
= 1;
80 int sysctl_tcp_window_scaling __read_mostly
= 1;
81 int sysctl_tcp_sack __read_mostly
= 1;
82 int sysctl_tcp_fack __read_mostly
= 1;
83 int sysctl_tcp_reordering __read_mostly
= TCP_FASTRETRANS_THRESH
;
84 int sysctl_tcp_max_reordering __read_mostly
= 300;
85 EXPORT_SYMBOL(sysctl_tcp_reordering
);
86 int sysctl_tcp_dsack __read_mostly
= 1;
87 int sysctl_tcp_app_win __read_mostly
= 31;
88 int sysctl_tcp_adv_win_scale __read_mostly
= 1;
89 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale
);
91 /* rfc5961 challenge ack rate limiting */
92 int sysctl_tcp_challenge_ack_limit
= 100;
94 int sysctl_tcp_stdurg __read_mostly
;
95 int sysctl_tcp_rfc1337 __read_mostly
;
96 int sysctl_tcp_max_orphans __read_mostly
= NR_FILE
;
97 int sysctl_tcp_frto __read_mostly
= 2;
99 int sysctl_tcp_thin_dupack __read_mostly
;
101 int sysctl_tcp_moderate_rcvbuf __read_mostly
= 1;
102 int sysctl_tcp_early_retrans __read_mostly
= 3;
103 int sysctl_tcp_invalid_ratelimit __read_mostly
= HZ
/2;
105 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
106 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
107 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
108 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
109 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
110 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
111 #define FLAG_ECE 0x40 /* ECE in this ACK */
112 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
113 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
114 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
115 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
116 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
117 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
119 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
120 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
121 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
122 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
124 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
125 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
127 /* Adapt the MSS value used to make delayed ack decision to the
130 static void tcp_measure_rcv_mss(struct sock
*sk
, const struct sk_buff
*skb
)
132 struct inet_connection_sock
*icsk
= inet_csk(sk
);
133 const unsigned int lss
= icsk
->icsk_ack
.last_seg_size
;
136 icsk
->icsk_ack
.last_seg_size
= 0;
138 /* skb->len may jitter because of SACKs, even if peer
139 * sends good full-sized frames.
141 len
= skb_shinfo(skb
)->gso_size
? : skb
->len
;
142 if (len
>= icsk
->icsk_ack
.rcv_mss
) {
143 icsk
->icsk_ack
.rcv_mss
= len
;
145 /* Otherwise, we make more careful check taking into account,
146 * that SACKs block is variable.
148 * "len" is invariant segment length, including TCP header.
150 len
+= skb
->data
- skb_transport_header(skb
);
151 if (len
>= TCP_MSS_DEFAULT
+ sizeof(struct tcphdr
) ||
152 /* If PSH is not set, packet should be
153 * full sized, provided peer TCP is not badly broken.
154 * This observation (if it is correct 8)) allows
155 * to handle super-low mtu links fairly.
157 (len
>= TCP_MIN_MSS
+ sizeof(struct tcphdr
) &&
158 !(tcp_flag_word(tcp_hdr(skb
)) & TCP_REMNANT
))) {
159 /* Subtract also invariant (if peer is RFC compliant),
160 * tcp header plus fixed timestamp option length.
161 * Resulting "len" is MSS free of SACK jitter.
163 len
-= tcp_sk(sk
)->tcp_header_len
;
164 icsk
->icsk_ack
.last_seg_size
= len
;
166 icsk
->icsk_ack
.rcv_mss
= len
;
170 if (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
)
171 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED2
;
172 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
176 static void tcp_incr_quickack(struct sock
*sk
)
178 struct inet_connection_sock
*icsk
= inet_csk(sk
);
179 unsigned int quickacks
= tcp_sk(sk
)->rcv_wnd
/ (2 * icsk
->icsk_ack
.rcv_mss
);
183 if (quickacks
> icsk
->icsk_ack
.quick
)
184 icsk
->icsk_ack
.quick
= min(quickacks
, TCP_MAX_QUICKACKS
);
187 static void tcp_enter_quickack_mode(struct sock
*sk
)
189 struct inet_connection_sock
*icsk
= inet_csk(sk
);
190 tcp_incr_quickack(sk
);
191 icsk
->icsk_ack
.pingpong
= 0;
192 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
195 /* Send ACKs quickly, if "quick" count is not exhausted
196 * and the session is not interactive.
199 static inline bool tcp_in_quickack_mode(const struct sock
*sk
)
201 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
203 return icsk
->icsk_ack
.quick
&& !icsk
->icsk_ack
.pingpong
;
206 static void tcp_ecn_queue_cwr(struct tcp_sock
*tp
)
208 if (tp
->ecn_flags
& TCP_ECN_OK
)
209 tp
->ecn_flags
|= TCP_ECN_QUEUE_CWR
;
212 static void tcp_ecn_accept_cwr(struct tcp_sock
*tp
, const struct sk_buff
*skb
)
214 if (tcp_hdr(skb
)->cwr
)
215 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
218 static void tcp_ecn_withdraw_cwr(struct tcp_sock
*tp
)
220 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
223 static void __tcp_ecn_check_ce(struct tcp_sock
*tp
, const struct sk_buff
*skb
)
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 (tcp_ca_needs_ecn((struct sock
*)tp
))
236 tcp_ca_event((struct sock
*)tp
, CA_EVENT_ECN_IS_CE
);
238 if (!(tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)) {
239 /* Better not delay acks, sender can have a very low cwnd */
240 tcp_enter_quickack_mode((struct sock
*)tp
);
241 tp
->ecn_flags
|= TCP_ECN_DEMAND_CWR
;
243 tp
->ecn_flags
|= TCP_ECN_SEEN
;
246 if (tcp_ca_needs_ecn((struct sock
*)tp
))
247 tcp_ca_event((struct sock
*)tp
, CA_EVENT_ECN_NO_CE
);
248 tp
->ecn_flags
|= TCP_ECN_SEEN
;
253 static void tcp_ecn_check_ce(struct tcp_sock
*tp
, const struct sk_buff
*skb
)
255 if (tp
->ecn_flags
& TCP_ECN_OK
)
256 __tcp_ecn_check_ce(tp
, skb
);
259 static void tcp_ecn_rcv_synack(struct tcp_sock
*tp
, const struct tcphdr
*th
)
261 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| th
->cwr
))
262 tp
->ecn_flags
&= ~TCP_ECN_OK
;
265 static void tcp_ecn_rcv_syn(struct tcp_sock
*tp
, const struct tcphdr
*th
)
267 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| !th
->cwr
))
268 tp
->ecn_flags
&= ~TCP_ECN_OK
;
271 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock
*tp
, const struct tcphdr
*th
)
273 if (th
->ece
&& !th
->syn
&& (tp
->ecn_flags
& TCP_ECN_OK
))
278 /* Buffer size and advertised window tuning.
280 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
283 static void tcp_sndbuf_expand(struct sock
*sk
)
285 const struct tcp_sock
*tp
= tcp_sk(sk
);
289 /* Worst case is non GSO/TSO : each frame consumes one skb
290 * and skb->head is kmalloced using power of two area of memory
292 per_mss
= max_t(u32
, tp
->rx_opt
.mss_clamp
, tp
->mss_cache
) +
294 SKB_DATA_ALIGN(sizeof(struct skb_shared_info
));
296 per_mss
= roundup_pow_of_two(per_mss
) +
297 SKB_DATA_ALIGN(sizeof(struct sk_buff
));
299 nr_segs
= max_t(u32
, TCP_INIT_CWND
, tp
->snd_cwnd
);
300 nr_segs
= max_t(u32
, nr_segs
, tp
->reordering
+ 1);
302 /* Fast Recovery (RFC 5681 3.2) :
303 * Cubic needs 1.7 factor, rounded to 2 to include
304 * extra cushion (application might react slowly to POLLOUT)
306 sndmem
= 2 * nr_segs
* per_mss
;
308 if (sk
->sk_sndbuf
< sndmem
)
309 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
312 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
314 * All tcp_full_space() is split to two parts: "network" buffer, allocated
315 * forward and advertised in receiver window (tp->rcv_wnd) and
316 * "application buffer", required to isolate scheduling/application
317 * latencies from network.
318 * window_clamp is maximal advertised window. It can be less than
319 * tcp_full_space(), in this case tcp_full_space() - window_clamp
320 * is reserved for "application" buffer. The less window_clamp is
321 * the smoother our behaviour from viewpoint of network, but the lower
322 * throughput and the higher sensitivity of the connection to losses. 8)
324 * rcv_ssthresh is more strict window_clamp used at "slow start"
325 * phase to predict further behaviour of this connection.
326 * It is used for two goals:
327 * - to enforce header prediction at sender, even when application
328 * requires some significant "application buffer". It is check #1.
329 * - to prevent pruning of receive queue because of misprediction
330 * of receiver window. Check #2.
332 * The scheme does not work when sender sends good segments opening
333 * window and then starts to feed us spaghetti. But it should work
334 * in common situations. Otherwise, we have to rely on queue collapsing.
337 /* Slow part of check#2. */
338 static int __tcp_grow_window(const struct sock
*sk
, const struct sk_buff
*skb
)
340 struct tcp_sock
*tp
= tcp_sk(sk
);
342 int truesize
= tcp_win_from_space(skb
->truesize
) >> 1;
343 int window
= tcp_win_from_space(sysctl_tcp_rmem
[2]) >> 1;
345 while (tp
->rcv_ssthresh
<= window
) {
346 if (truesize
<= skb
->len
)
347 return 2 * inet_csk(sk
)->icsk_ack
.rcv_mss
;
355 static void tcp_grow_window(struct sock
*sk
, const struct sk_buff
*skb
)
357 struct tcp_sock
*tp
= tcp_sk(sk
);
360 if (tp
->rcv_ssthresh
< tp
->window_clamp
&&
361 (int)tp
->rcv_ssthresh
< tcp_space(sk
) &&
362 !sk_under_memory_pressure(sk
)) {
365 /* Check #2. Increase window, if skb with such overhead
366 * will fit to rcvbuf in future.
368 if (tcp_win_from_space(skb
->truesize
) <= skb
->len
)
369 incr
= 2 * tp
->advmss
;
371 incr
= __tcp_grow_window(sk
, skb
);
374 incr
= max_t(int, incr
, 2 * skb
->len
);
375 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
+ incr
,
377 inet_csk(sk
)->icsk_ack
.quick
|= 1;
382 /* 3. Tuning rcvbuf, when connection enters established state. */
383 static void tcp_fixup_rcvbuf(struct sock
*sk
)
385 u32 mss
= tcp_sk(sk
)->advmss
;
388 rcvmem
= 2 * SKB_TRUESIZE(mss
+ MAX_TCP_HEADER
) *
389 tcp_default_init_rwnd(mss
);
391 /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
392 * Allow enough cushion so that sender is not limited by our window
394 if (sysctl_tcp_moderate_rcvbuf
)
397 if (sk
->sk_rcvbuf
< rcvmem
)
398 sk
->sk_rcvbuf
= min(rcvmem
, sysctl_tcp_rmem
[2]);
401 /* 4. Try to fixup all. It is made immediately after connection enters
404 void tcp_init_buffer_space(struct sock
*sk
)
406 struct tcp_sock
*tp
= tcp_sk(sk
);
409 if (!(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
))
410 tcp_fixup_rcvbuf(sk
);
411 if (!(sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
))
412 tcp_sndbuf_expand(sk
);
414 tp
->rcvq_space
.space
= tp
->rcv_wnd
;
415 tp
->rcvq_space
.time
= tcp_time_stamp
;
416 tp
->rcvq_space
.seq
= tp
->copied_seq
;
418 maxwin
= tcp_full_space(sk
);
420 if (tp
->window_clamp
>= maxwin
) {
421 tp
->window_clamp
= maxwin
;
423 if (sysctl_tcp_app_win
&& maxwin
> 4 * tp
->advmss
)
424 tp
->window_clamp
= max(maxwin
-
425 (maxwin
>> sysctl_tcp_app_win
),
429 /* Force reservation of one segment. */
430 if (sysctl_tcp_app_win
&&
431 tp
->window_clamp
> 2 * tp
->advmss
&&
432 tp
->window_clamp
+ tp
->advmss
> maxwin
)
433 tp
->window_clamp
= max(2 * tp
->advmss
, maxwin
- tp
->advmss
);
435 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, tp
->window_clamp
);
436 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
439 /* 5. Recalculate window clamp after socket hit its memory bounds. */
440 static void tcp_clamp_window(struct sock
*sk
)
442 struct tcp_sock
*tp
= tcp_sk(sk
);
443 struct inet_connection_sock
*icsk
= inet_csk(sk
);
445 icsk
->icsk_ack
.quick
= 0;
447 if (sk
->sk_rcvbuf
< sysctl_tcp_rmem
[2] &&
448 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
) &&
449 !sk_under_memory_pressure(sk
) &&
450 sk_memory_allocated(sk
) < sk_prot_mem_limits(sk
, 0)) {
451 sk
->sk_rcvbuf
= min(atomic_read(&sk
->sk_rmem_alloc
),
454 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
)
455 tp
->rcv_ssthresh
= min(tp
->window_clamp
, 2U * tp
->advmss
);
458 /* Initialize RCV_MSS value.
459 * RCV_MSS is an our guess about MSS used by the peer.
460 * We haven't any direct information about the MSS.
461 * It's better to underestimate the RCV_MSS rather than overestimate.
462 * Overestimations make us ACKing less frequently than needed.
463 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
465 void tcp_initialize_rcv_mss(struct sock
*sk
)
467 const struct tcp_sock
*tp
= tcp_sk(sk
);
468 unsigned int hint
= min_t(unsigned int, tp
->advmss
, tp
->mss_cache
);
470 hint
= min(hint
, tp
->rcv_wnd
/ 2);
471 hint
= min(hint
, TCP_MSS_DEFAULT
);
472 hint
= max(hint
, TCP_MIN_MSS
);
474 inet_csk(sk
)->icsk_ack
.rcv_mss
= hint
;
476 EXPORT_SYMBOL(tcp_initialize_rcv_mss
);
478 /* Receiver "autotuning" code.
480 * The algorithm for RTT estimation w/o timestamps is based on
481 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
482 * <http://public.lanl.gov/radiant/pubs.html#DRS>
484 * More detail on this code can be found at
485 * <http://staff.psc.edu/jheffner/>,
486 * though this reference is out of date. A new paper
489 static void tcp_rcv_rtt_update(struct tcp_sock
*tp
, u32 sample
, int win_dep
)
491 u32 new_sample
= tp
->rcv_rtt_est
.rtt
;
497 if (new_sample
!= 0) {
498 /* If we sample in larger samples in the non-timestamp
499 * case, we could grossly overestimate the RTT especially
500 * with chatty applications or bulk transfer apps which
501 * are stalled on filesystem I/O.
503 * Also, since we are only going for a minimum in the
504 * non-timestamp case, we do not smooth things out
505 * else with timestamps disabled convergence takes too
509 m
-= (new_sample
>> 3);
517 /* No previous measure. */
521 if (tp
->rcv_rtt_est
.rtt
!= new_sample
)
522 tp
->rcv_rtt_est
.rtt
= new_sample
;
525 static inline void tcp_rcv_rtt_measure(struct tcp_sock
*tp
)
527 if (tp
->rcv_rtt_est
.time
== 0)
529 if (before(tp
->rcv_nxt
, tp
->rcv_rtt_est
.seq
))
531 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rcv_rtt_est
.time
, 1);
534 tp
->rcv_rtt_est
.seq
= tp
->rcv_nxt
+ tp
->rcv_wnd
;
535 tp
->rcv_rtt_est
.time
= tcp_time_stamp
;
538 static inline void tcp_rcv_rtt_measure_ts(struct sock
*sk
,
539 const struct sk_buff
*skb
)
541 struct tcp_sock
*tp
= tcp_sk(sk
);
542 if (tp
->rx_opt
.rcv_tsecr
&&
543 (TCP_SKB_CB(skb
)->end_seq
-
544 TCP_SKB_CB(skb
)->seq
>= inet_csk(sk
)->icsk_ack
.rcv_mss
))
545 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
, 0);
549 * This function should be called every time data is copied to user space.
550 * It calculates the appropriate TCP receive buffer space.
552 void tcp_rcv_space_adjust(struct sock
*sk
)
554 struct tcp_sock
*tp
= tcp_sk(sk
);
558 time
= tcp_time_stamp
- tp
->rcvq_space
.time
;
559 if (time
< (tp
->rcv_rtt_est
.rtt
>> 3) || tp
->rcv_rtt_est
.rtt
== 0)
562 /* Number of bytes copied to user in last RTT */
563 copied
= tp
->copied_seq
- tp
->rcvq_space
.seq
;
564 if (copied
<= tp
->rcvq_space
.space
)
568 * copied = bytes received in previous RTT, our base window
569 * To cope with packet losses, we need a 2x factor
570 * To cope with slow start, and sender growing its cwin by 100 %
571 * every RTT, we need a 4x factor, because the ACK we are sending
572 * now is for the next RTT, not the current one :
573 * <prev RTT . ><current RTT .. ><next RTT .... >
576 if (sysctl_tcp_moderate_rcvbuf
&&
577 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
)) {
578 int rcvwin
, rcvmem
, rcvbuf
;
580 /* minimal window to cope with packet losses, assuming
581 * steady state. Add some cushion because of small variations.
583 rcvwin
= (copied
<< 1) + 16 * tp
->advmss
;
585 /* If rate increased by 25%,
586 * assume slow start, rcvwin = 3 * copied
587 * If rate increased by 50%,
588 * assume sender can use 2x growth, rcvwin = 4 * copied
591 tp
->rcvq_space
.space
+ (tp
->rcvq_space
.space
>> 2)) {
593 tp
->rcvq_space
.space
+ (tp
->rcvq_space
.space
>> 1))
596 rcvwin
+= (rcvwin
>> 1);
599 rcvmem
= SKB_TRUESIZE(tp
->advmss
+ MAX_TCP_HEADER
);
600 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
603 rcvbuf
= min(rcvwin
/ tp
->advmss
* rcvmem
, sysctl_tcp_rmem
[2]);
604 if (rcvbuf
> sk
->sk_rcvbuf
) {
605 sk
->sk_rcvbuf
= rcvbuf
;
607 /* Make the window clamp follow along. */
608 tp
->window_clamp
= rcvwin
;
611 tp
->rcvq_space
.space
= copied
;
614 tp
->rcvq_space
.seq
= tp
->copied_seq
;
615 tp
->rcvq_space
.time
= tcp_time_stamp
;
618 /* There is something which you must keep in mind when you analyze the
619 * behavior of the tp->ato delayed ack timeout interval. When a
620 * connection starts up, we want to ack as quickly as possible. The
621 * problem is that "good" TCP's do slow start at the beginning of data
622 * transmission. The means that until we send the first few ACK's the
623 * sender will sit on his end and only queue most of his data, because
624 * he can only send snd_cwnd unacked packets at any given time. For
625 * each ACK we send, he increments snd_cwnd and transmits more of his
628 static void tcp_event_data_recv(struct sock
*sk
, struct sk_buff
*skb
)
630 struct tcp_sock
*tp
= tcp_sk(sk
);
631 struct inet_connection_sock
*icsk
= inet_csk(sk
);
634 inet_csk_schedule_ack(sk
);
636 tcp_measure_rcv_mss(sk
, skb
);
638 tcp_rcv_rtt_measure(tp
);
640 now
= tcp_time_stamp
;
642 if (!icsk
->icsk_ack
.ato
) {
643 /* The _first_ data packet received, initialize
644 * delayed ACK engine.
646 tcp_incr_quickack(sk
);
647 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
649 int m
= now
- icsk
->icsk_ack
.lrcvtime
;
651 if (m
<= TCP_ATO_MIN
/ 2) {
652 /* The fastest case is the first. */
653 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + TCP_ATO_MIN
/ 2;
654 } else if (m
< icsk
->icsk_ack
.ato
) {
655 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + m
;
656 if (icsk
->icsk_ack
.ato
> icsk
->icsk_rto
)
657 icsk
->icsk_ack
.ato
= icsk
->icsk_rto
;
658 } else if (m
> icsk
->icsk_rto
) {
659 /* Too long gap. Apparently sender failed to
660 * restart window, so that we send ACKs quickly.
662 tcp_incr_quickack(sk
);
666 icsk
->icsk_ack
.lrcvtime
= now
;
668 tcp_ecn_check_ce(tp
, skb
);
671 tcp_grow_window(sk
, skb
);
674 /* Called to compute a smoothed rtt estimate. The data fed to this
675 * routine either comes from timestamps, or from segments that were
676 * known _not_ to have been retransmitted [see Karn/Partridge
677 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
678 * piece by Van Jacobson.
679 * NOTE: the next three routines used to be one big routine.
680 * To save cycles in the RFC 1323 implementation it was better to break
681 * it up into three procedures. -- erics
683 static void tcp_rtt_estimator(struct sock
*sk
, long mrtt_us
)
685 struct tcp_sock
*tp
= tcp_sk(sk
);
686 long m
= mrtt_us
; /* RTT */
687 u32 srtt
= tp
->srtt_us
;
689 /* The following amusing code comes from Jacobson's
690 * article in SIGCOMM '88. Note that rtt and mdev
691 * are scaled versions of rtt and mean deviation.
692 * This is designed to be as fast as possible
693 * m stands for "measurement".
695 * On a 1990 paper the rto value is changed to:
696 * RTO = rtt + 4 * mdev
698 * Funny. This algorithm seems to be very broken.
699 * These formulae increase RTO, when it should be decreased, increase
700 * too slowly, when it should be increased quickly, decrease too quickly
701 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
702 * does not matter how to _calculate_ it. Seems, it was trap
703 * that VJ failed to avoid. 8)
706 m
-= (srtt
>> 3); /* m is now error in rtt est */
707 srtt
+= m
; /* rtt = 7/8 rtt + 1/8 new */
709 m
= -m
; /* m is now abs(error) */
710 m
-= (tp
->mdev_us
>> 2); /* similar update on mdev */
711 /* This is similar to one of Eifel findings.
712 * Eifel blocks mdev updates when rtt decreases.
713 * This solution is a bit different: we use finer gain
714 * for mdev in this case (alpha*beta).
715 * Like Eifel it also prevents growth of rto,
716 * but also it limits too fast rto decreases,
717 * happening in pure Eifel.
722 m
-= (tp
->mdev_us
>> 2); /* similar update on mdev */
724 tp
->mdev_us
+= m
; /* mdev = 3/4 mdev + 1/4 new */
725 if (tp
->mdev_us
> tp
->mdev_max_us
) {
726 tp
->mdev_max_us
= tp
->mdev_us
;
727 if (tp
->mdev_max_us
> tp
->rttvar_us
)
728 tp
->rttvar_us
= tp
->mdev_max_us
;
730 if (after(tp
->snd_una
, tp
->rtt_seq
)) {
731 if (tp
->mdev_max_us
< tp
->rttvar_us
)
732 tp
->rttvar_us
-= (tp
->rttvar_us
- tp
->mdev_max_us
) >> 2;
733 tp
->rtt_seq
= tp
->snd_nxt
;
734 tp
->mdev_max_us
= tcp_rto_min_us(sk
);
737 /* no previous measure. */
738 srtt
= m
<< 3; /* take the measured time to be rtt */
739 tp
->mdev_us
= m
<< 1; /* make sure rto = 3*rtt */
740 tp
->rttvar_us
= max(tp
->mdev_us
, tcp_rto_min_us(sk
));
741 tp
->mdev_max_us
= tp
->rttvar_us
;
742 tp
->rtt_seq
= tp
->snd_nxt
;
744 tp
->srtt_us
= max(1U, srtt
);
747 /* Set the sk_pacing_rate to allow proper sizing of TSO packets.
748 * Note: TCP stack does not yet implement pacing.
749 * FQ packet scheduler can be used to implement cheap but effective
750 * TCP pacing, to smooth the burst on large writes when packets
751 * in flight is significantly lower than cwnd (or rwin)
753 static void tcp_update_pacing_rate(struct sock
*sk
)
755 const struct tcp_sock
*tp
= tcp_sk(sk
);
758 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
759 rate
= (u64
)tp
->mss_cache
* 2 * (USEC_PER_SEC
<< 3);
761 rate
*= max(tp
->snd_cwnd
, tp
->packets_out
);
763 if (likely(tp
->srtt_us
))
764 do_div(rate
, tp
->srtt_us
);
766 /* ACCESS_ONCE() is needed because sch_fq fetches sk_pacing_rate
767 * without any lock. We want to make sure compiler wont store
768 * intermediate values in this location.
770 ACCESS_ONCE(sk
->sk_pacing_rate
) = min_t(u64
, rate
,
771 sk
->sk_max_pacing_rate
);
774 /* Calculate rto without backoff. This is the second half of Van Jacobson's
775 * routine referred to above.
777 static void tcp_set_rto(struct sock
*sk
)
779 const struct tcp_sock
*tp
= tcp_sk(sk
);
780 /* Old crap is replaced with new one. 8)
783 * 1. If rtt variance happened to be less 50msec, it is hallucination.
784 * It cannot be less due to utterly erratic ACK generation made
785 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
786 * to do with delayed acks, because at cwnd>2 true delack timeout
787 * is invisible. Actually, Linux-2.4 also generates erratic
788 * ACKs in some circumstances.
790 inet_csk(sk
)->icsk_rto
= __tcp_set_rto(tp
);
792 /* 2. Fixups made earlier cannot be right.
793 * If we do not estimate RTO correctly without them,
794 * all the algo is pure shit and should be replaced
795 * with correct one. It is exactly, which we pretend to do.
798 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
799 * guarantees that rto is higher.
804 __u32
tcp_init_cwnd(const struct tcp_sock
*tp
, const struct dst_entry
*dst
)
806 __u32 cwnd
= (dst
? dst_metric(dst
, RTAX_INITCWND
) : 0);
809 cwnd
= TCP_INIT_CWND
;
810 return min_t(__u32
, cwnd
, tp
->snd_cwnd_clamp
);
814 * Packet counting of FACK is based on in-order assumptions, therefore TCP
815 * disables it when reordering is detected
817 void tcp_disable_fack(struct tcp_sock
*tp
)
819 /* RFC3517 uses different metric in lost marker => reset on change */
821 tp
->lost_skb_hint
= NULL
;
822 tp
->rx_opt
.sack_ok
&= ~TCP_FACK_ENABLED
;
825 /* Take a notice that peer is sending D-SACKs */
826 static void tcp_dsack_seen(struct tcp_sock
*tp
)
828 tp
->rx_opt
.sack_ok
|= TCP_DSACK_SEEN
;
831 static void tcp_update_reordering(struct sock
*sk
, const int metric
,
834 struct tcp_sock
*tp
= tcp_sk(sk
);
835 if (metric
> tp
->reordering
) {
838 tp
->reordering
= min(sysctl_tcp_max_reordering
, metric
);
840 /* This exciting event is worth to be remembered. 8) */
842 mib_idx
= LINUX_MIB_TCPTSREORDER
;
843 else if (tcp_is_reno(tp
))
844 mib_idx
= LINUX_MIB_TCPRENOREORDER
;
845 else if (tcp_is_fack(tp
))
846 mib_idx
= LINUX_MIB_TCPFACKREORDER
;
848 mib_idx
= LINUX_MIB_TCPSACKREORDER
;
850 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
851 #if FASTRETRANS_DEBUG > 1
852 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
853 tp
->rx_opt
.sack_ok
, inet_csk(sk
)->icsk_ca_state
,
857 tp
->undo_marker
? tp
->undo_retrans
: 0);
859 tcp_disable_fack(tp
);
863 tcp_disable_early_retrans(tp
);
866 /* This must be called before lost_out is incremented */
867 static void tcp_verify_retransmit_hint(struct tcp_sock
*tp
, struct sk_buff
*skb
)
869 if (!tp
->retransmit_skb_hint
||
870 before(TCP_SKB_CB(skb
)->seq
,
871 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
872 tp
->retransmit_skb_hint
= skb
;
875 after(TCP_SKB_CB(skb
)->end_seq
, tp
->retransmit_high
))
876 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
879 static void tcp_skb_mark_lost(struct tcp_sock
*tp
, struct sk_buff
*skb
)
881 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
882 tcp_verify_retransmit_hint(tp
, skb
);
884 tp
->lost_out
+= tcp_skb_pcount(skb
);
885 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
889 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock
*tp
,
892 tcp_verify_retransmit_hint(tp
, skb
);
894 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
895 tp
->lost_out
+= tcp_skb_pcount(skb
);
896 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
900 /* This procedure tags the retransmission queue when SACKs arrive.
902 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
903 * Packets in queue with these bits set are counted in variables
904 * sacked_out, retrans_out and lost_out, correspondingly.
906 * Valid combinations are:
907 * Tag InFlight Description
908 * 0 1 - orig segment is in flight.
909 * S 0 - nothing flies, orig reached receiver.
910 * L 0 - nothing flies, orig lost by net.
911 * R 2 - both orig and retransmit are in flight.
912 * L|R 1 - orig is lost, retransmit is in flight.
913 * S|R 1 - orig reached receiver, retrans is still in flight.
914 * (L|S|R is logically valid, it could occur when L|R is sacked,
915 * but it is equivalent to plain S and code short-curcuits it to S.
916 * L|S is logically invalid, it would mean -1 packet in flight 8))
918 * These 6 states form finite state machine, controlled by the following events:
919 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
920 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
921 * 3. Loss detection event of two flavors:
922 * A. Scoreboard estimator decided the packet is lost.
923 * A'. Reno "three dupacks" marks head of queue lost.
924 * A''. Its FACK modification, head until snd.fack is lost.
925 * B. SACK arrives sacking SND.NXT at the moment, when the
926 * segment was retransmitted.
927 * 4. D-SACK added new rule: D-SACK changes any tag to S.
929 * It is pleasant to note, that state diagram turns out to be commutative,
930 * so that we are allowed not to be bothered by order of our actions,
931 * when multiple events arrive simultaneously. (see the function below).
933 * Reordering detection.
934 * --------------------
935 * Reordering metric is maximal distance, which a packet can be displaced
936 * in packet stream. With SACKs we can estimate it:
938 * 1. SACK fills old hole and the corresponding segment was not
939 * ever retransmitted -> reordering. Alas, we cannot use it
940 * when segment was retransmitted.
941 * 2. The last flaw is solved with D-SACK. D-SACK arrives
942 * for retransmitted and already SACKed segment -> reordering..
943 * Both of these heuristics are not used in Loss state, when we cannot
944 * account for retransmits accurately.
946 * SACK block validation.
947 * ----------------------
949 * SACK block range validation checks that the received SACK block fits to
950 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
951 * Note that SND.UNA is not included to the range though being valid because
952 * it means that the receiver is rather inconsistent with itself reporting
953 * SACK reneging when it should advance SND.UNA. Such SACK block this is
954 * perfectly valid, however, in light of RFC2018 which explicitly states
955 * that "SACK block MUST reflect the newest segment. Even if the newest
956 * segment is going to be discarded ...", not that it looks very clever
957 * in case of head skb. Due to potentional receiver driven attacks, we
958 * choose to avoid immediate execution of a walk in write queue due to
959 * reneging and defer head skb's loss recovery to standard loss recovery
960 * procedure that will eventually trigger (nothing forbids us doing this).
962 * Implements also blockage to start_seq wrap-around. Problem lies in the
963 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
964 * there's no guarantee that it will be before snd_nxt (n). The problem
965 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
968 * <- outs wnd -> <- wrapzone ->
969 * u e n u_w e_w s n_w
971 * |<------------+------+----- TCP seqno space --------------+---------->|
972 * ...-- <2^31 ->| |<--------...
973 * ...---- >2^31 ------>| |<--------...
975 * Current code wouldn't be vulnerable but it's better still to discard such
976 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
977 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
978 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
979 * equal to the ideal case (infinite seqno space without wrap caused issues).
981 * With D-SACK the lower bound is extended to cover sequence space below
982 * SND.UNA down to undo_marker, which is the last point of interest. Yet
983 * again, D-SACK block must not to go across snd_una (for the same reason as
984 * for the normal SACK blocks, explained above). But there all simplicity
985 * ends, TCP might receive valid D-SACKs below that. As long as they reside
986 * fully below undo_marker they do not affect behavior in anyway and can
987 * therefore be safely ignored. In rare cases (which are more or less
988 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
989 * fragmentation and packet reordering past skb's retransmission. To consider
990 * them correctly, the acceptable range must be extended even more though
991 * the exact amount is rather hard to quantify. However, tp->max_window can
992 * be used as an exaggerated estimate.
994 static bool tcp_is_sackblock_valid(struct tcp_sock
*tp
, bool is_dsack
,
995 u32 start_seq
, u32 end_seq
)
997 /* Too far in future, or reversed (interpretation is ambiguous) */
998 if (after(end_seq
, tp
->snd_nxt
) || !before(start_seq
, end_seq
))
1001 /* Nasty start_seq wrap-around check (see comments above) */
1002 if (!before(start_seq
, tp
->snd_nxt
))
1005 /* In outstanding window? ...This is valid exit for D-SACKs too.
1006 * start_seq == snd_una is non-sensical (see comments above)
1008 if (after(start_seq
, tp
->snd_una
))
1011 if (!is_dsack
|| !tp
->undo_marker
)
1014 /* ...Then it's D-SACK, and must reside below snd_una completely */
1015 if (after(end_seq
, tp
->snd_una
))
1018 if (!before(start_seq
, tp
->undo_marker
))
1022 if (!after(end_seq
, tp
->undo_marker
))
1025 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1026 * start_seq < undo_marker and end_seq >= undo_marker.
1028 return !before(start_seq
, end_seq
- tp
->max_window
);
1031 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1032 * Event "B". Later note: FACK people cheated me again 8), we have to account
1033 * for reordering! Ugly, but should help.
1035 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1036 * less than what is now known to be received by the other end (derived from
1037 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1038 * retransmitted skbs to avoid some costly processing per ACKs.
1040 static void tcp_mark_lost_retrans(struct sock
*sk
)
1042 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1043 struct tcp_sock
*tp
= tcp_sk(sk
);
1044 struct sk_buff
*skb
;
1046 u32 new_low_seq
= tp
->snd_nxt
;
1047 u32 received_upto
= tcp_highest_sack_seq(tp
);
1049 if (!tcp_is_fack(tp
) || !tp
->retrans_out
||
1050 !after(received_upto
, tp
->lost_retrans_low
) ||
1051 icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
1054 tcp_for_write_queue(skb
, sk
) {
1055 u32 ack_seq
= TCP_SKB_CB(skb
)->ack_seq
;
1057 if (skb
== tcp_send_head(sk
))
1059 if (cnt
== tp
->retrans_out
)
1061 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1064 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
))
1067 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1068 * constraint here (see above) but figuring out that at
1069 * least tp->reordering SACK blocks reside between ack_seq
1070 * and received_upto is not easy task to do cheaply with
1071 * the available datastructures.
1073 * Whether FACK should check here for tp->reordering segs
1074 * in-between one could argue for either way (it would be
1075 * rather simple to implement as we could count fack_count
1076 * during the walk and do tp->fackets_out - fack_count).
1078 if (after(received_upto
, ack_seq
)) {
1079 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1080 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1082 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
1083 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSTRETRANSMIT
);
1085 if (before(ack_seq
, new_low_seq
))
1086 new_low_seq
= ack_seq
;
1087 cnt
+= tcp_skb_pcount(skb
);
1091 if (tp
->retrans_out
)
1092 tp
->lost_retrans_low
= new_low_seq
;
1095 static bool tcp_check_dsack(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1096 struct tcp_sack_block_wire
*sp
, int num_sacks
,
1099 struct tcp_sock
*tp
= tcp_sk(sk
);
1100 u32 start_seq_0
= get_unaligned_be32(&sp
[0].start_seq
);
1101 u32 end_seq_0
= get_unaligned_be32(&sp
[0].end_seq
);
1102 bool dup_sack
= false;
1104 if (before(start_seq_0
, TCP_SKB_CB(ack_skb
)->ack_seq
)) {
1107 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPDSACKRECV
);
1108 } else if (num_sacks
> 1) {
1109 u32 end_seq_1
= get_unaligned_be32(&sp
[1].end_seq
);
1110 u32 start_seq_1
= get_unaligned_be32(&sp
[1].start_seq
);
1112 if (!after(end_seq_0
, end_seq_1
) &&
1113 !before(start_seq_0
, start_seq_1
)) {
1116 NET_INC_STATS_BH(sock_net(sk
),
1117 LINUX_MIB_TCPDSACKOFORECV
);
1121 /* D-SACK for already forgotten data... Do dumb counting. */
1122 if (dup_sack
&& tp
->undo_marker
&& tp
->undo_retrans
> 0 &&
1123 !after(end_seq_0
, prior_snd_una
) &&
1124 after(end_seq_0
, tp
->undo_marker
))
1130 struct tcp_sacktag_state
{
1133 long rtt_us
; /* RTT measured by SACKing never-retransmitted data */
1137 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1138 * the incoming SACK may not exactly match but we can find smaller MSS
1139 * aligned portion of it that matches. Therefore we might need to fragment
1140 * which may fail and creates some hassle (caller must handle error case
1143 * FIXME: this could be merged to shift decision code
1145 static int tcp_match_skb_to_sack(struct sock
*sk
, struct sk_buff
*skb
,
1146 u32 start_seq
, u32 end_seq
)
1150 unsigned int pkt_len
;
1153 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1154 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1156 if (tcp_skb_pcount(skb
) > 1 && !in_sack
&&
1157 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
1158 mss
= tcp_skb_mss(skb
);
1159 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1162 pkt_len
= start_seq
- TCP_SKB_CB(skb
)->seq
;
1166 pkt_len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1171 /* Round if necessary so that SACKs cover only full MSSes
1172 * and/or the remaining small portion (if present)
1174 if (pkt_len
> mss
) {
1175 unsigned int new_len
= (pkt_len
/ mss
) * mss
;
1176 if (!in_sack
&& new_len
< pkt_len
) {
1178 if (new_len
>= skb
->len
)
1183 err
= tcp_fragment(sk
, skb
, pkt_len
, mss
, GFP_ATOMIC
);
1191 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1192 static u8
tcp_sacktag_one(struct sock
*sk
,
1193 struct tcp_sacktag_state
*state
, u8 sacked
,
1194 u32 start_seq
, u32 end_seq
,
1195 int dup_sack
, int pcount
,
1196 const struct skb_mstamp
*xmit_time
)
1198 struct tcp_sock
*tp
= tcp_sk(sk
);
1199 int fack_count
= state
->fack_count
;
1201 /* Account D-SACK for retransmitted packet. */
1202 if (dup_sack
&& (sacked
& TCPCB_RETRANS
)) {
1203 if (tp
->undo_marker
&& tp
->undo_retrans
> 0 &&
1204 after(end_seq
, tp
->undo_marker
))
1206 if (sacked
& TCPCB_SACKED_ACKED
)
1207 state
->reord
= min(fack_count
, state
->reord
);
1210 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1211 if (!after(end_seq
, tp
->snd_una
))
1214 if (!(sacked
& TCPCB_SACKED_ACKED
)) {
1215 if (sacked
& TCPCB_SACKED_RETRANS
) {
1216 /* If the segment is not tagged as lost,
1217 * we do not clear RETRANS, believing
1218 * that retransmission is still in flight.
1220 if (sacked
& TCPCB_LOST
) {
1221 sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1222 tp
->lost_out
-= pcount
;
1223 tp
->retrans_out
-= pcount
;
1226 if (!(sacked
& TCPCB_RETRANS
)) {
1227 /* New sack for not retransmitted frame,
1228 * which was in hole. It is reordering.
1230 if (before(start_seq
,
1231 tcp_highest_sack_seq(tp
)))
1232 state
->reord
= min(fack_count
,
1234 if (!after(end_seq
, tp
->high_seq
))
1235 state
->flag
|= FLAG_ORIG_SACK_ACKED
;
1236 /* Pick the earliest sequence sacked for RTT */
1237 if (state
->rtt_us
< 0) {
1238 struct skb_mstamp now
;
1240 skb_mstamp_get(&now
);
1241 state
->rtt_us
= skb_mstamp_us_delta(&now
,
1246 if (sacked
& TCPCB_LOST
) {
1247 sacked
&= ~TCPCB_LOST
;
1248 tp
->lost_out
-= pcount
;
1252 sacked
|= TCPCB_SACKED_ACKED
;
1253 state
->flag
|= FLAG_DATA_SACKED
;
1254 tp
->sacked_out
+= pcount
;
1256 fack_count
+= pcount
;
1258 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1259 if (!tcp_is_fack(tp
) && tp
->lost_skb_hint
&&
1260 before(start_seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
))
1261 tp
->lost_cnt_hint
+= pcount
;
1263 if (fack_count
> tp
->fackets_out
)
1264 tp
->fackets_out
= fack_count
;
1267 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1268 * frames and clear it. undo_retrans is decreased above, L|R frames
1269 * are accounted above as well.
1271 if (dup_sack
&& (sacked
& TCPCB_SACKED_RETRANS
)) {
1272 sacked
&= ~TCPCB_SACKED_RETRANS
;
1273 tp
->retrans_out
-= pcount
;
1279 /* Shift newly-SACKed bytes from this skb to the immediately previous
1280 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1282 static bool tcp_shifted_skb(struct sock
*sk
, struct sk_buff
*skb
,
1283 struct tcp_sacktag_state
*state
,
1284 unsigned int pcount
, int shifted
, int mss
,
1287 struct tcp_sock
*tp
= tcp_sk(sk
);
1288 struct sk_buff
*prev
= tcp_write_queue_prev(sk
, skb
);
1289 u32 start_seq
= TCP_SKB_CB(skb
)->seq
; /* start of newly-SACKed */
1290 u32 end_seq
= start_seq
+ shifted
; /* end of newly-SACKed */
1294 /* Adjust counters and hints for the newly sacked sequence
1295 * range but discard the return value since prev is already
1296 * marked. We must tag the range first because the seq
1297 * advancement below implicitly advances
1298 * tcp_highest_sack_seq() when skb is highest_sack.
1300 tcp_sacktag_one(sk
, state
, TCP_SKB_CB(skb
)->sacked
,
1301 start_seq
, end_seq
, dup_sack
, pcount
,
1304 if (skb
== tp
->lost_skb_hint
)
1305 tp
->lost_cnt_hint
+= pcount
;
1307 TCP_SKB_CB(prev
)->end_seq
+= shifted
;
1308 TCP_SKB_CB(skb
)->seq
+= shifted
;
1310 tcp_skb_pcount_add(prev
, pcount
);
1311 BUG_ON(tcp_skb_pcount(skb
) < pcount
);
1312 tcp_skb_pcount_add(skb
, -pcount
);
1314 /* When we're adding to gso_segs == 1, gso_size will be zero,
1315 * in theory this shouldn't be necessary but as long as DSACK
1316 * code can come after this skb later on it's better to keep
1317 * setting gso_size to something.
1319 if (!skb_shinfo(prev
)->gso_size
) {
1320 skb_shinfo(prev
)->gso_size
= mss
;
1321 skb_shinfo(prev
)->gso_type
= sk
->sk_gso_type
;
1324 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1325 if (tcp_skb_pcount(skb
) <= 1) {
1326 skb_shinfo(skb
)->gso_size
= 0;
1327 skb_shinfo(skb
)->gso_type
= 0;
1330 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1331 TCP_SKB_CB(prev
)->sacked
|= (TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
);
1334 BUG_ON(!tcp_skb_pcount(skb
));
1335 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKSHIFTED
);
1339 /* Whole SKB was eaten :-) */
1341 if (skb
== tp
->retransmit_skb_hint
)
1342 tp
->retransmit_skb_hint
= prev
;
1343 if (skb
== tp
->lost_skb_hint
) {
1344 tp
->lost_skb_hint
= prev
;
1345 tp
->lost_cnt_hint
-= tcp_skb_pcount(prev
);
1348 TCP_SKB_CB(prev
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
;
1349 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
1350 TCP_SKB_CB(prev
)->end_seq
++;
1352 if (skb
== tcp_highest_sack(sk
))
1353 tcp_advance_highest_sack(sk
, skb
);
1355 tcp_unlink_write_queue(skb
, sk
);
1356 sk_wmem_free_skb(sk
, skb
);
1358 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKMERGED
);
1363 /* I wish gso_size would have a bit more sane initialization than
1364 * something-or-zero which complicates things
1366 static int tcp_skb_seglen(const struct sk_buff
*skb
)
1368 return tcp_skb_pcount(skb
) == 1 ? skb
->len
: tcp_skb_mss(skb
);
1371 /* Shifting pages past head area doesn't work */
1372 static int skb_can_shift(const struct sk_buff
*skb
)
1374 return !skb_headlen(skb
) && skb_is_nonlinear(skb
);
1377 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1380 static struct sk_buff
*tcp_shift_skb_data(struct sock
*sk
, struct sk_buff
*skb
,
1381 struct tcp_sacktag_state
*state
,
1382 u32 start_seq
, u32 end_seq
,
1385 struct tcp_sock
*tp
= tcp_sk(sk
);
1386 struct sk_buff
*prev
;
1392 if (!sk_can_gso(sk
))
1395 /* Normally R but no L won't result in plain S */
1397 (TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_RETRANS
)) == TCPCB_SACKED_RETRANS
)
1399 if (!skb_can_shift(skb
))
1401 /* This frame is about to be dropped (was ACKed). */
1402 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1405 /* Can only happen with delayed DSACK + discard craziness */
1406 if (unlikely(skb
== tcp_write_queue_head(sk
)))
1408 prev
= tcp_write_queue_prev(sk
, skb
);
1410 if ((TCP_SKB_CB(prev
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
)
1413 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1414 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1418 pcount
= tcp_skb_pcount(skb
);
1419 mss
= tcp_skb_seglen(skb
);
1421 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1422 * drop this restriction as unnecessary
1424 if (mss
!= tcp_skb_seglen(prev
))
1427 if (!after(TCP_SKB_CB(skb
)->end_seq
, start_seq
))
1429 /* CHECKME: This is non-MSS split case only?, this will
1430 * cause skipped skbs due to advancing loop btw, original
1431 * has that feature too
1433 if (tcp_skb_pcount(skb
) <= 1)
1436 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1438 /* TODO: head merge to next could be attempted here
1439 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1440 * though it might not be worth of the additional hassle
1442 * ...we can probably just fallback to what was done
1443 * previously. We could try merging non-SACKed ones
1444 * as well but it probably isn't going to buy off
1445 * because later SACKs might again split them, and
1446 * it would make skb timestamp tracking considerably
1452 len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1454 BUG_ON(len
> skb
->len
);
1456 /* MSS boundaries should be honoured or else pcount will
1457 * severely break even though it makes things bit trickier.
1458 * Optimize common case to avoid most of the divides
1460 mss
= tcp_skb_mss(skb
);
1462 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1463 * drop this restriction as unnecessary
1465 if (mss
!= tcp_skb_seglen(prev
))
1470 } else if (len
< mss
) {
1478 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1479 if (!after(TCP_SKB_CB(skb
)->seq
+ len
, tp
->snd_una
))
1482 if (!skb_shift(prev
, skb
, len
))
1484 if (!tcp_shifted_skb(sk
, skb
, state
, pcount
, len
, mss
, dup_sack
))
1487 /* Hole filled allows collapsing with the next as well, this is very
1488 * useful when hole on every nth skb pattern happens
1490 if (prev
== tcp_write_queue_tail(sk
))
1492 skb
= tcp_write_queue_next(sk
, prev
);
1494 if (!skb_can_shift(skb
) ||
1495 (skb
== tcp_send_head(sk
)) ||
1496 ((TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
) ||
1497 (mss
!= tcp_skb_seglen(skb
)))
1501 if (skb_shift(prev
, skb
, len
)) {
1502 pcount
+= tcp_skb_pcount(skb
);
1503 tcp_shifted_skb(sk
, skb
, state
, tcp_skb_pcount(skb
), len
, mss
, 0);
1507 state
->fack_count
+= pcount
;
1514 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKSHIFTFALLBACK
);
1518 static struct sk_buff
*tcp_sacktag_walk(struct sk_buff
*skb
, struct sock
*sk
,
1519 struct tcp_sack_block
*next_dup
,
1520 struct tcp_sacktag_state
*state
,
1521 u32 start_seq
, u32 end_seq
,
1524 struct tcp_sock
*tp
= tcp_sk(sk
);
1525 struct sk_buff
*tmp
;
1527 tcp_for_write_queue_from(skb
, sk
) {
1529 bool dup_sack
= dup_sack_in
;
1531 if (skb
== tcp_send_head(sk
))
1534 /* queue is in-order => we can short-circuit the walk early */
1535 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1539 before(TCP_SKB_CB(skb
)->seq
, next_dup
->end_seq
)) {
1540 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1541 next_dup
->start_seq
,
1547 /* skb reference here is a bit tricky to get right, since
1548 * shifting can eat and free both this skb and the next,
1549 * so not even _safe variant of the loop is enough.
1552 tmp
= tcp_shift_skb_data(sk
, skb
, state
,
1553 start_seq
, end_seq
, dup_sack
);
1562 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1568 if (unlikely(in_sack
< 0))
1572 TCP_SKB_CB(skb
)->sacked
=
1575 TCP_SKB_CB(skb
)->sacked
,
1576 TCP_SKB_CB(skb
)->seq
,
1577 TCP_SKB_CB(skb
)->end_seq
,
1579 tcp_skb_pcount(skb
),
1582 if (!before(TCP_SKB_CB(skb
)->seq
,
1583 tcp_highest_sack_seq(tp
)))
1584 tcp_advance_highest_sack(sk
, skb
);
1587 state
->fack_count
+= tcp_skb_pcount(skb
);
1592 /* Avoid all extra work that is being done by sacktag while walking in
1595 static struct sk_buff
*tcp_sacktag_skip(struct sk_buff
*skb
, struct sock
*sk
,
1596 struct tcp_sacktag_state
*state
,
1599 tcp_for_write_queue_from(skb
, sk
) {
1600 if (skb
== tcp_send_head(sk
))
1603 if (after(TCP_SKB_CB(skb
)->end_seq
, skip_to_seq
))
1606 state
->fack_count
+= tcp_skb_pcount(skb
);
1611 static struct sk_buff
*tcp_maybe_skipping_dsack(struct sk_buff
*skb
,
1613 struct tcp_sack_block
*next_dup
,
1614 struct tcp_sacktag_state
*state
,
1620 if (before(next_dup
->start_seq
, skip_to_seq
)) {
1621 skb
= tcp_sacktag_skip(skb
, sk
, state
, next_dup
->start_seq
);
1622 skb
= tcp_sacktag_walk(skb
, sk
, NULL
, state
,
1623 next_dup
->start_seq
, next_dup
->end_seq
,
1630 static int tcp_sack_cache_ok(const struct tcp_sock
*tp
, const struct tcp_sack_block
*cache
)
1632 return cache
< tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1636 tcp_sacktag_write_queue(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1637 u32 prior_snd_una
, long *sack_rtt_us
)
1639 struct tcp_sock
*tp
= tcp_sk(sk
);
1640 const unsigned char *ptr
= (skb_transport_header(ack_skb
) +
1641 TCP_SKB_CB(ack_skb
)->sacked
);
1642 struct tcp_sack_block_wire
*sp_wire
= (struct tcp_sack_block_wire
*)(ptr
+2);
1643 struct tcp_sack_block sp
[TCP_NUM_SACKS
];
1644 struct tcp_sack_block
*cache
;
1645 struct tcp_sacktag_state state
;
1646 struct sk_buff
*skb
;
1647 int num_sacks
= min(TCP_NUM_SACKS
, (ptr
[1] - TCPOLEN_SACK_BASE
) >> 3);
1649 bool found_dup_sack
= false;
1651 int first_sack_index
;
1654 state
.reord
= tp
->packets_out
;
1657 if (!tp
->sacked_out
) {
1658 if (WARN_ON(tp
->fackets_out
))
1659 tp
->fackets_out
= 0;
1660 tcp_highest_sack_reset(sk
);
1663 found_dup_sack
= tcp_check_dsack(sk
, ack_skb
, sp_wire
,
1664 num_sacks
, prior_snd_una
);
1666 state
.flag
|= FLAG_DSACKING_ACK
;
1668 /* Eliminate too old ACKs, but take into
1669 * account more or less fresh ones, they can
1670 * contain valid SACK info.
1672 if (before(TCP_SKB_CB(ack_skb
)->ack_seq
, prior_snd_una
- tp
->max_window
))
1675 if (!tp
->packets_out
)
1679 first_sack_index
= 0;
1680 for (i
= 0; i
< num_sacks
; i
++) {
1681 bool dup_sack
= !i
&& found_dup_sack
;
1683 sp
[used_sacks
].start_seq
= get_unaligned_be32(&sp_wire
[i
].start_seq
);
1684 sp
[used_sacks
].end_seq
= get_unaligned_be32(&sp_wire
[i
].end_seq
);
1686 if (!tcp_is_sackblock_valid(tp
, dup_sack
,
1687 sp
[used_sacks
].start_seq
,
1688 sp
[used_sacks
].end_seq
)) {
1692 if (!tp
->undo_marker
)
1693 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDNOUNDO
;
1695 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDOLD
;
1697 /* Don't count olds caused by ACK reordering */
1698 if ((TCP_SKB_CB(ack_skb
)->ack_seq
!= tp
->snd_una
) &&
1699 !after(sp
[used_sacks
].end_seq
, tp
->snd_una
))
1701 mib_idx
= LINUX_MIB_TCPSACKDISCARD
;
1704 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
1706 first_sack_index
= -1;
1710 /* Ignore very old stuff early */
1711 if (!after(sp
[used_sacks
].end_seq
, prior_snd_una
))
1717 /* order SACK blocks to allow in order walk of the retrans queue */
1718 for (i
= used_sacks
- 1; i
> 0; i
--) {
1719 for (j
= 0; j
< i
; j
++) {
1720 if (after(sp
[j
].start_seq
, sp
[j
+ 1].start_seq
)) {
1721 swap(sp
[j
], sp
[j
+ 1]);
1723 /* Track where the first SACK block goes to */
1724 if (j
== first_sack_index
)
1725 first_sack_index
= j
+ 1;
1730 skb
= tcp_write_queue_head(sk
);
1731 state
.fack_count
= 0;
1734 if (!tp
->sacked_out
) {
1735 /* It's already past, so skip checking against it */
1736 cache
= tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1738 cache
= tp
->recv_sack_cache
;
1739 /* Skip empty blocks in at head of the cache */
1740 while (tcp_sack_cache_ok(tp
, cache
) && !cache
->start_seq
&&
1745 while (i
< used_sacks
) {
1746 u32 start_seq
= sp
[i
].start_seq
;
1747 u32 end_seq
= sp
[i
].end_seq
;
1748 bool dup_sack
= (found_dup_sack
&& (i
== first_sack_index
));
1749 struct tcp_sack_block
*next_dup
= NULL
;
1751 if (found_dup_sack
&& ((i
+ 1) == first_sack_index
))
1752 next_dup
= &sp
[i
+ 1];
1754 /* Skip too early cached blocks */
1755 while (tcp_sack_cache_ok(tp
, cache
) &&
1756 !before(start_seq
, cache
->end_seq
))
1759 /* Can skip some work by looking recv_sack_cache? */
1760 if (tcp_sack_cache_ok(tp
, cache
) && !dup_sack
&&
1761 after(end_seq
, cache
->start_seq
)) {
1764 if (before(start_seq
, cache
->start_seq
)) {
1765 skb
= tcp_sacktag_skip(skb
, sk
, &state
,
1767 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
,
1774 /* Rest of the block already fully processed? */
1775 if (!after(end_seq
, cache
->end_seq
))
1778 skb
= tcp_maybe_skipping_dsack(skb
, sk
, next_dup
,
1782 /* ...tail remains todo... */
1783 if (tcp_highest_sack_seq(tp
) == cache
->end_seq
) {
1784 /* ...but better entrypoint exists! */
1785 skb
= tcp_highest_sack(sk
);
1788 state
.fack_count
= tp
->fackets_out
;
1793 skb
= tcp_sacktag_skip(skb
, sk
, &state
, cache
->end_seq
);
1794 /* Check overlap against next cached too (past this one already) */
1799 if (!before(start_seq
, tcp_highest_sack_seq(tp
))) {
1800 skb
= tcp_highest_sack(sk
);
1803 state
.fack_count
= tp
->fackets_out
;
1805 skb
= tcp_sacktag_skip(skb
, sk
, &state
, start_seq
);
1808 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
, &state
,
1809 start_seq
, end_seq
, dup_sack
);
1815 /* Clear the head of the cache sack blocks so we can skip it next time */
1816 for (i
= 0; i
< ARRAY_SIZE(tp
->recv_sack_cache
) - used_sacks
; i
++) {
1817 tp
->recv_sack_cache
[i
].start_seq
= 0;
1818 tp
->recv_sack_cache
[i
].end_seq
= 0;
1820 for (j
= 0; j
< used_sacks
; j
++)
1821 tp
->recv_sack_cache
[i
++] = sp
[j
];
1823 if ((state
.reord
< tp
->fackets_out
) &&
1824 ((inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Loss
) || tp
->undo_marker
))
1825 tcp_update_reordering(sk
, tp
->fackets_out
- state
.reord
, 0);
1827 tcp_mark_lost_retrans(sk
);
1828 tcp_verify_left_out(tp
);
1831 #if FASTRETRANS_DEBUG > 0
1832 WARN_ON((int)tp
->sacked_out
< 0);
1833 WARN_ON((int)tp
->lost_out
< 0);
1834 WARN_ON((int)tp
->retrans_out
< 0);
1835 WARN_ON((int)tcp_packets_in_flight(tp
) < 0);
1837 *sack_rtt_us
= state
.rtt_us
;
1841 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1842 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1844 static bool tcp_limit_reno_sacked(struct tcp_sock
*tp
)
1848 holes
= max(tp
->lost_out
, 1U);
1849 holes
= min(holes
, tp
->packets_out
);
1851 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1852 tp
->sacked_out
= tp
->packets_out
- holes
;
1858 /* If we receive more dupacks than we expected counting segments
1859 * in assumption of absent reordering, interpret this as reordering.
1860 * The only another reason could be bug in receiver TCP.
1862 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1864 struct tcp_sock
*tp
= tcp_sk(sk
);
1865 if (tcp_limit_reno_sacked(tp
))
1866 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1869 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1871 static void tcp_add_reno_sack(struct sock
*sk
)
1873 struct tcp_sock
*tp
= tcp_sk(sk
);
1875 tcp_check_reno_reordering(sk
, 0);
1876 tcp_verify_left_out(tp
);
1879 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1881 static void tcp_remove_reno_sacks(struct sock
*sk
, int acked
)
1883 struct tcp_sock
*tp
= tcp_sk(sk
);
1886 /* One ACK acked hole. The rest eat duplicate ACKs. */
1887 if (acked
- 1 >= tp
->sacked_out
)
1890 tp
->sacked_out
-= acked
- 1;
1892 tcp_check_reno_reordering(sk
, acked
);
1893 tcp_verify_left_out(tp
);
1896 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1901 void tcp_clear_retrans(struct tcp_sock
*tp
)
1903 tp
->retrans_out
= 0;
1905 tp
->undo_marker
= 0;
1906 tp
->undo_retrans
= -1;
1907 tp
->fackets_out
= 0;
1911 static inline void tcp_init_undo(struct tcp_sock
*tp
)
1913 tp
->undo_marker
= tp
->snd_una
;
1914 /* Retransmission still in flight may cause DSACKs later. */
1915 tp
->undo_retrans
= tp
->retrans_out
? : -1;
1918 /* Enter Loss state. If we detect SACK reneging, forget all SACK information
1919 * and reset tags completely, otherwise preserve SACKs. If receiver
1920 * dropped its ofo queue, we will know this due to reneging detection.
1922 void tcp_enter_loss(struct sock
*sk
)
1924 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1925 struct tcp_sock
*tp
= tcp_sk(sk
);
1926 struct sk_buff
*skb
;
1927 bool new_recovery
= false;
1928 bool is_reneg
; /* is receiver reneging on SACKs? */
1930 /* Reduce ssthresh if it has not yet been made inside this window. */
1931 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
||
1932 !after(tp
->high_seq
, tp
->snd_una
) ||
1933 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
1934 new_recovery
= true;
1935 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1936 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1937 tcp_ca_event(sk
, CA_EVENT_LOSS
);
1941 tp
->snd_cwnd_cnt
= 0;
1942 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1944 tp
->retrans_out
= 0;
1947 if (tcp_is_reno(tp
))
1948 tcp_reset_reno_sack(tp
);
1950 skb
= tcp_write_queue_head(sk
);
1951 is_reneg
= skb
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
);
1953 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSACKRENEGING
);
1955 tp
->fackets_out
= 0;
1957 tcp_clear_all_retrans_hints(tp
);
1959 tcp_for_write_queue(skb
, sk
) {
1960 if (skb
== tcp_send_head(sk
))
1963 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
1964 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) || is_reneg
) {
1965 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
1966 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1967 tp
->lost_out
+= tcp_skb_pcount(skb
);
1968 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
1971 tcp_verify_left_out(tp
);
1973 /* Timeout in disordered state after receiving substantial DUPACKs
1974 * suggests that the degree of reordering is over-estimated.
1976 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
&&
1977 tp
->sacked_out
>= sysctl_tcp_reordering
)
1978 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1979 sysctl_tcp_reordering
);
1980 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1981 tp
->high_seq
= tp
->snd_nxt
;
1982 tcp_ecn_queue_cwr(tp
);
1984 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1985 * loss recovery is underway except recurring timeout(s) on
1986 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1988 tp
->frto
= sysctl_tcp_frto
&&
1989 (new_recovery
|| icsk
->icsk_retransmits
) &&
1990 !inet_csk(sk
)->icsk_mtup
.probe_size
;
1993 /* If ACK arrived pointing to a remembered SACK, it means that our
1994 * remembered SACKs do not reflect real state of receiver i.e.
1995 * receiver _host_ is heavily congested (or buggy).
1997 * To avoid big spurious retransmission bursts due to transient SACK
1998 * scoreboard oddities that look like reneging, we give the receiver a
1999 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2000 * restore sanity to the SACK scoreboard. If the apparent reneging
2001 * persists until this RTO then we'll clear the SACK scoreboard.
2003 static bool tcp_check_sack_reneging(struct sock
*sk
, int flag
)
2005 if (flag
& FLAG_SACK_RENEGING
) {
2006 struct tcp_sock
*tp
= tcp_sk(sk
);
2007 unsigned long delay
= max(usecs_to_jiffies(tp
->srtt_us
>> 4),
2008 msecs_to_jiffies(10));
2010 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2011 delay
, TCP_RTO_MAX
);
2017 static inline int tcp_fackets_out(const struct tcp_sock
*tp
)
2019 return tcp_is_reno(tp
) ? tp
->sacked_out
+ 1 : tp
->fackets_out
;
2022 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2023 * counter when SACK is enabled (without SACK, sacked_out is used for
2026 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2027 * segments up to the highest received SACK block so far and holes in
2030 * With reordering, holes may still be in flight, so RFC3517 recovery
2031 * uses pure sacked_out (total number of SACKed segments) even though
2032 * it violates the RFC that uses duplicate ACKs, often these are equal
2033 * but when e.g. out-of-window ACKs or packet duplication occurs,
2034 * they differ. Since neither occurs due to loss, TCP should really
2037 static inline int tcp_dupack_heuristics(const struct tcp_sock
*tp
)
2039 return tcp_is_fack(tp
) ? tp
->fackets_out
: tp
->sacked_out
+ 1;
2042 static bool tcp_pause_early_retransmit(struct sock
*sk
, int flag
)
2044 struct tcp_sock
*tp
= tcp_sk(sk
);
2045 unsigned long delay
;
2047 /* Delay early retransmit and entering fast recovery for
2048 * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
2049 * available, or RTO is scheduled to fire first.
2051 if (sysctl_tcp_early_retrans
< 2 || sysctl_tcp_early_retrans
> 3 ||
2052 (flag
& FLAG_ECE
) || !tp
->srtt_us
)
2055 delay
= max(usecs_to_jiffies(tp
->srtt_us
>> 5),
2056 msecs_to_jiffies(2));
2058 if (!time_after(inet_csk(sk
)->icsk_timeout
, (jiffies
+ delay
)))
2061 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_EARLY_RETRANS
, delay
,
2066 /* Linux NewReno/SACK/FACK/ECN state machine.
2067 * --------------------------------------
2069 * "Open" Normal state, no dubious events, fast path.
2070 * "Disorder" In all the respects it is "Open",
2071 * but requires a bit more attention. It is entered when
2072 * we see some SACKs or dupacks. It is split of "Open"
2073 * mainly to move some processing from fast path to slow one.
2074 * "CWR" CWND was reduced due to some Congestion Notification event.
2075 * It can be ECN, ICMP source quench, local device congestion.
2076 * "Recovery" CWND was reduced, we are fast-retransmitting.
2077 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2079 * tcp_fastretrans_alert() is entered:
2080 * - each incoming ACK, if state is not "Open"
2081 * - when arrived ACK is unusual, namely:
2086 * Counting packets in flight is pretty simple.
2088 * in_flight = packets_out - left_out + retrans_out
2090 * packets_out is SND.NXT-SND.UNA counted in packets.
2092 * retrans_out is number of retransmitted segments.
2094 * left_out is number of segments left network, but not ACKed yet.
2096 * left_out = sacked_out + lost_out
2098 * sacked_out: Packets, which arrived to receiver out of order
2099 * and hence not ACKed. With SACKs this number is simply
2100 * amount of SACKed data. Even without SACKs
2101 * it is easy to give pretty reliable estimate of this number,
2102 * counting duplicate ACKs.
2104 * lost_out: Packets lost by network. TCP has no explicit
2105 * "loss notification" feedback from network (for now).
2106 * It means that this number can be only _guessed_.
2107 * Actually, it is the heuristics to predict lossage that
2108 * distinguishes different algorithms.
2110 * F.e. after RTO, when all the queue is considered as lost,
2111 * lost_out = packets_out and in_flight = retrans_out.
2113 * Essentially, we have now two algorithms counting
2116 * FACK: It is the simplest heuristics. As soon as we decided
2117 * that something is lost, we decide that _all_ not SACKed
2118 * packets until the most forward SACK are lost. I.e.
2119 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2120 * It is absolutely correct estimate, if network does not reorder
2121 * packets. And it loses any connection to reality when reordering
2122 * takes place. We use FACK by default until reordering
2123 * is suspected on the path to this destination.
2125 * NewReno: when Recovery is entered, we assume that one segment
2126 * is lost (classic Reno). While we are in Recovery and
2127 * a partial ACK arrives, we assume that one more packet
2128 * is lost (NewReno). This heuristics are the same in NewReno
2131 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2132 * deflation etc. CWND is real congestion window, never inflated, changes
2133 * only according to classic VJ rules.
2135 * Really tricky (and requiring careful tuning) part of algorithm
2136 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2137 * The first determines the moment _when_ we should reduce CWND and,
2138 * hence, slow down forward transmission. In fact, it determines the moment
2139 * when we decide that hole is caused by loss, rather than by a reorder.
2141 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2142 * holes, caused by lost packets.
2144 * And the most logically complicated part of algorithm is undo
2145 * heuristics. We detect false retransmits due to both too early
2146 * fast retransmit (reordering) and underestimated RTO, analyzing
2147 * timestamps and D-SACKs. When we detect that some segments were
2148 * retransmitted by mistake and CWND reduction was wrong, we undo
2149 * window reduction and abort recovery phase. This logic is hidden
2150 * inside several functions named tcp_try_undo_<something>.
2153 /* This function decides, when we should leave Disordered state
2154 * and enter Recovery phase, reducing congestion window.
2156 * Main question: may we further continue forward transmission
2157 * with the same cwnd?
2159 static bool tcp_time_to_recover(struct sock
*sk
, int flag
)
2161 struct tcp_sock
*tp
= tcp_sk(sk
);
2164 /* Trick#1: The loss is proven. */
2168 /* Not-A-Trick#2 : Classic rule... */
2169 if (tcp_dupack_heuristics(tp
) > tp
->reordering
)
2172 /* Trick#4: It is still not OK... But will it be useful to delay
2175 packets_out
= tp
->packets_out
;
2176 if (packets_out
<= tp
->reordering
&&
2177 tp
->sacked_out
>= max_t(__u32
, packets_out
/2, sysctl_tcp_reordering
) &&
2178 !tcp_may_send_now(sk
)) {
2179 /* We have nothing to send. This connection is limited
2180 * either by receiver window or by application.
2185 /* If a thin stream is detected, retransmit after first
2186 * received dupack. Employ only if SACK is supported in order
2187 * to avoid possible corner-case series of spurious retransmissions
2188 * Use only if there are no unsent data.
2190 if ((tp
->thin_dupack
|| sysctl_tcp_thin_dupack
) &&
2191 tcp_stream_is_thin(tp
) && tcp_dupack_heuristics(tp
) > 1 &&
2192 tcp_is_sack(tp
) && !tcp_send_head(sk
))
2195 /* Trick#6: TCP early retransmit, per RFC5827. To avoid spurious
2196 * retransmissions due to small network reorderings, we implement
2197 * Mitigation A.3 in the RFC and delay the retransmission for a short
2198 * interval if appropriate.
2200 if (tp
->do_early_retrans
&& !tp
->retrans_out
&& tp
->sacked_out
&&
2201 (tp
->packets_out
>= (tp
->sacked_out
+ 1) && tp
->packets_out
< 4) &&
2202 !tcp_may_send_now(sk
))
2203 return !tcp_pause_early_retransmit(sk
, flag
);
2208 /* Detect loss in event "A" above by marking head of queue up as lost.
2209 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2210 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2211 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2212 * the maximum SACKed segments to pass before reaching this limit.
2214 static void tcp_mark_head_lost(struct sock
*sk
, int packets
, int mark_head
)
2216 struct tcp_sock
*tp
= tcp_sk(sk
);
2217 struct sk_buff
*skb
;
2221 /* Use SACK to deduce losses of new sequences sent during recovery */
2222 const u32 loss_high
= tcp_is_sack(tp
) ? tp
->snd_nxt
: tp
->high_seq
;
2224 WARN_ON(packets
> tp
->packets_out
);
2225 if (tp
->lost_skb_hint
) {
2226 skb
= tp
->lost_skb_hint
;
2227 cnt
= tp
->lost_cnt_hint
;
2228 /* Head already handled? */
2229 if (mark_head
&& skb
!= tcp_write_queue_head(sk
))
2232 skb
= tcp_write_queue_head(sk
);
2236 tcp_for_write_queue_from(skb
, sk
) {
2237 if (skb
== tcp_send_head(sk
))
2239 /* TODO: do this better */
2240 /* this is not the most efficient way to do this... */
2241 tp
->lost_skb_hint
= skb
;
2242 tp
->lost_cnt_hint
= cnt
;
2244 if (after(TCP_SKB_CB(skb
)->end_seq
, loss_high
))
2248 if (tcp_is_fack(tp
) || tcp_is_reno(tp
) ||
2249 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2250 cnt
+= tcp_skb_pcount(skb
);
2252 if (cnt
> packets
) {
2253 if ((tcp_is_sack(tp
) && !tcp_is_fack(tp
)) ||
2254 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
) ||
2255 (oldcnt
>= packets
))
2258 mss
= skb_shinfo(skb
)->gso_size
;
2259 err
= tcp_fragment(sk
, skb
, (packets
- oldcnt
) * mss
,
2266 tcp_skb_mark_lost(tp
, skb
);
2271 tcp_verify_left_out(tp
);
2274 /* Account newly detected lost packet(s) */
2276 static void tcp_update_scoreboard(struct sock
*sk
, int fast_rexmit
)
2278 struct tcp_sock
*tp
= tcp_sk(sk
);
2280 if (tcp_is_reno(tp
)) {
2281 tcp_mark_head_lost(sk
, 1, 1);
2282 } else if (tcp_is_fack(tp
)) {
2283 int lost
= tp
->fackets_out
- tp
->reordering
;
2286 tcp_mark_head_lost(sk
, lost
, 0);
2288 int sacked_upto
= tp
->sacked_out
- tp
->reordering
;
2289 if (sacked_upto
>= 0)
2290 tcp_mark_head_lost(sk
, sacked_upto
, 0);
2291 else if (fast_rexmit
)
2292 tcp_mark_head_lost(sk
, 1, 1);
2296 /* CWND moderation, preventing bursts due to too big ACKs
2297 * in dubious situations.
2299 static inline void tcp_moderate_cwnd(struct tcp_sock
*tp
)
2301 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
2302 tcp_packets_in_flight(tp
) + tcp_max_burst(tp
));
2303 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2306 /* Nothing was retransmitted or returned timestamp is less
2307 * than timestamp of the first retransmission.
2309 static inline bool tcp_packet_delayed(const struct tcp_sock
*tp
)
2311 return !tp
->retrans_stamp
||
2312 (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2313 before(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
));
2316 /* Undo procedures. */
2318 /* We can clear retrans_stamp when there are no retransmissions in the
2319 * window. It would seem that it is trivially available for us in
2320 * tp->retrans_out, however, that kind of assumptions doesn't consider
2321 * what will happen if errors occur when sending retransmission for the
2322 * second time. ...It could the that such segment has only
2323 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2324 * the head skb is enough except for some reneging corner cases that
2325 * are not worth the effort.
2327 * Main reason for all this complexity is the fact that connection dying
2328 * time now depends on the validity of the retrans_stamp, in particular,
2329 * that successive retransmissions of a segment must not advance
2330 * retrans_stamp under any conditions.
2332 static bool tcp_any_retrans_done(const struct sock
*sk
)
2334 const struct tcp_sock
*tp
= tcp_sk(sk
);
2335 struct sk_buff
*skb
;
2337 if (tp
->retrans_out
)
2340 skb
= tcp_write_queue_head(sk
);
2341 if (unlikely(skb
&& TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
))
2347 #if FASTRETRANS_DEBUG > 1
2348 static void DBGUNDO(struct sock
*sk
, const char *msg
)
2350 struct tcp_sock
*tp
= tcp_sk(sk
);
2351 struct inet_sock
*inet
= inet_sk(sk
);
2353 if (sk
->sk_family
== AF_INET
) {
2354 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2356 &inet
->inet_daddr
, ntohs(inet
->inet_dport
),
2357 tp
->snd_cwnd
, tcp_left_out(tp
),
2358 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2361 #if IS_ENABLED(CONFIG_IPV6)
2362 else if (sk
->sk_family
== AF_INET6
) {
2363 struct ipv6_pinfo
*np
= inet6_sk(sk
);
2364 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2366 &np
->daddr
, ntohs(inet
->inet_dport
),
2367 tp
->snd_cwnd
, tcp_left_out(tp
),
2368 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2374 #define DBGUNDO(x...) do { } while (0)
2377 static void tcp_undo_cwnd_reduction(struct sock
*sk
, bool unmark_loss
)
2379 struct tcp_sock
*tp
= tcp_sk(sk
);
2382 struct sk_buff
*skb
;
2384 tcp_for_write_queue(skb
, sk
) {
2385 if (skb
== tcp_send_head(sk
))
2387 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2390 tcp_clear_all_retrans_hints(tp
);
2393 if (tp
->prior_ssthresh
) {
2394 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2396 if (icsk
->icsk_ca_ops
->undo_cwnd
)
2397 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
2399 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
<< 1);
2401 if (tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
2402 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
2403 tcp_ecn_withdraw_cwr(tp
);
2406 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2408 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2409 tp
->undo_marker
= 0;
2412 static inline bool tcp_may_undo(const struct tcp_sock
*tp
)
2414 return tp
->undo_marker
&& (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
2417 /* People celebrate: "We love our President!" */
2418 static bool tcp_try_undo_recovery(struct sock
*sk
)
2420 struct tcp_sock
*tp
= tcp_sk(sk
);
2422 if (tcp_may_undo(tp
)) {
2425 /* Happy end! We did not retransmit anything
2426 * or our original transmission succeeded.
2428 DBGUNDO(sk
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
2429 tcp_undo_cwnd_reduction(sk
, false);
2430 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
2431 mib_idx
= LINUX_MIB_TCPLOSSUNDO
;
2433 mib_idx
= LINUX_MIB_TCPFULLUNDO
;
2435 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2437 if (tp
->snd_una
== tp
->high_seq
&& tcp_is_reno(tp
)) {
2438 /* Hold old state until something *above* high_seq
2439 * is ACKed. For Reno it is MUST to prevent false
2440 * fast retransmits (RFC2582). SACK TCP is safe. */
2441 tcp_moderate_cwnd(tp
);
2442 if (!tcp_any_retrans_done(sk
))
2443 tp
->retrans_stamp
= 0;
2446 tcp_set_ca_state(sk
, TCP_CA_Open
);
2450 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2451 static bool tcp_try_undo_dsack(struct sock
*sk
)
2453 struct tcp_sock
*tp
= tcp_sk(sk
);
2455 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
2456 DBGUNDO(sk
, "D-SACK");
2457 tcp_undo_cwnd_reduction(sk
, false);
2458 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPDSACKUNDO
);
2464 /* Undo during loss recovery after partial ACK or using F-RTO. */
2465 static bool tcp_try_undo_loss(struct sock
*sk
, bool frto_undo
)
2467 struct tcp_sock
*tp
= tcp_sk(sk
);
2469 if (frto_undo
|| tcp_may_undo(tp
)) {
2470 tcp_undo_cwnd_reduction(sk
, true);
2472 DBGUNDO(sk
, "partial loss");
2473 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSSUNDO
);
2475 NET_INC_STATS_BH(sock_net(sk
),
2476 LINUX_MIB_TCPSPURIOUSRTOS
);
2477 inet_csk(sk
)->icsk_retransmits
= 0;
2478 if (frto_undo
|| tcp_is_sack(tp
))
2479 tcp_set_ca_state(sk
, TCP_CA_Open
);
2485 /* The cwnd reduction in CWR and Recovery use the PRR algorithm
2486 * https://datatracker.ietf.org/doc/draft-ietf-tcpm-proportional-rate-reduction/
2487 * It computes the number of packets to send (sndcnt) based on packets newly
2489 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2490 * cwnd reductions across a full RTT.
2491 * 2) If packets in flight is lower than ssthresh (such as due to excess
2492 * losses and/or application stalls), do not perform any further cwnd
2493 * reductions, but instead slow start up to ssthresh.
2495 static void tcp_init_cwnd_reduction(struct sock
*sk
)
2497 struct tcp_sock
*tp
= tcp_sk(sk
);
2499 tp
->high_seq
= tp
->snd_nxt
;
2500 tp
->tlp_high_seq
= 0;
2501 tp
->snd_cwnd_cnt
= 0;
2502 tp
->prior_cwnd
= tp
->snd_cwnd
;
2503 tp
->prr_delivered
= 0;
2505 tp
->snd_ssthresh
= inet_csk(sk
)->icsk_ca_ops
->ssthresh(sk
);
2506 tcp_ecn_queue_cwr(tp
);
2509 static void tcp_cwnd_reduction(struct sock
*sk
, const int prior_unsacked
,
2512 struct tcp_sock
*tp
= tcp_sk(sk
);
2514 int delta
= tp
->snd_ssthresh
- tcp_packets_in_flight(tp
);
2515 int newly_acked_sacked
= prior_unsacked
-
2516 (tp
->packets_out
- tp
->sacked_out
);
2518 tp
->prr_delivered
+= newly_acked_sacked
;
2519 if (tcp_packets_in_flight(tp
) > tp
->snd_ssthresh
) {
2520 u64 dividend
= (u64
)tp
->snd_ssthresh
* tp
->prr_delivered
+
2522 sndcnt
= div_u64(dividend
, tp
->prior_cwnd
) - tp
->prr_out
;
2524 sndcnt
= min_t(int, delta
,
2525 max_t(int, tp
->prr_delivered
- tp
->prr_out
,
2526 newly_acked_sacked
) + 1);
2529 sndcnt
= max(sndcnt
, (fast_rexmit
? 1 : 0));
2530 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + sndcnt
;
2533 static inline void tcp_end_cwnd_reduction(struct sock
*sk
)
2535 struct tcp_sock
*tp
= tcp_sk(sk
);
2537 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2538 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_CWR
||
2539 (tp
->undo_marker
&& tp
->snd_ssthresh
< TCP_INFINITE_SSTHRESH
)) {
2540 tp
->snd_cwnd
= tp
->snd_ssthresh
;
2541 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2543 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
2546 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2547 void tcp_enter_cwr(struct sock
*sk
)
2549 struct tcp_sock
*tp
= tcp_sk(sk
);
2551 tp
->prior_ssthresh
= 0;
2552 if (inet_csk(sk
)->icsk_ca_state
< TCP_CA_CWR
) {
2553 tp
->undo_marker
= 0;
2554 tcp_init_cwnd_reduction(sk
);
2555 tcp_set_ca_state(sk
, TCP_CA_CWR
);
2559 static void tcp_try_keep_open(struct sock
*sk
)
2561 struct tcp_sock
*tp
= tcp_sk(sk
);
2562 int state
= TCP_CA_Open
;
2564 if (tcp_left_out(tp
) || tcp_any_retrans_done(sk
))
2565 state
= TCP_CA_Disorder
;
2567 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
2568 tcp_set_ca_state(sk
, state
);
2569 tp
->high_seq
= tp
->snd_nxt
;
2573 static void tcp_try_to_open(struct sock
*sk
, int flag
, const int prior_unsacked
)
2575 struct tcp_sock
*tp
= tcp_sk(sk
);
2577 tcp_verify_left_out(tp
);
2579 if (!tcp_any_retrans_done(sk
))
2580 tp
->retrans_stamp
= 0;
2582 if (flag
& FLAG_ECE
)
2585 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
2586 tcp_try_keep_open(sk
);
2588 tcp_cwnd_reduction(sk
, prior_unsacked
, 0);
2592 static void tcp_mtup_probe_failed(struct sock
*sk
)
2594 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2596 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
2597 icsk
->icsk_mtup
.probe_size
= 0;
2600 static void tcp_mtup_probe_success(struct sock
*sk
)
2602 struct tcp_sock
*tp
= tcp_sk(sk
);
2603 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2605 /* FIXME: breaks with very large cwnd */
2606 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2607 tp
->snd_cwnd
= tp
->snd_cwnd
*
2608 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
2609 icsk
->icsk_mtup
.probe_size
;
2610 tp
->snd_cwnd_cnt
= 0;
2611 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2612 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2614 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
2615 icsk
->icsk_mtup
.probe_size
= 0;
2616 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
2619 /* Do a simple retransmit without using the backoff mechanisms in
2620 * tcp_timer. This is used for path mtu discovery.
2621 * The socket is already locked here.
2623 void tcp_simple_retransmit(struct sock
*sk
)
2625 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2626 struct tcp_sock
*tp
= tcp_sk(sk
);
2627 struct sk_buff
*skb
;
2628 unsigned int mss
= tcp_current_mss(sk
);
2629 u32 prior_lost
= tp
->lost_out
;
2631 tcp_for_write_queue(skb
, sk
) {
2632 if (skb
== tcp_send_head(sk
))
2634 if (tcp_skb_seglen(skb
) > mss
&&
2635 !(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
2636 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2637 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
2638 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2640 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
2644 tcp_clear_retrans_hints_partial(tp
);
2646 if (prior_lost
== tp
->lost_out
)
2649 if (tcp_is_reno(tp
))
2650 tcp_limit_reno_sacked(tp
);
2652 tcp_verify_left_out(tp
);
2654 /* Don't muck with the congestion window here.
2655 * Reason is that we do not increase amount of _data_
2656 * in network, but units changed and effective
2657 * cwnd/ssthresh really reduced now.
2659 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
2660 tp
->high_seq
= tp
->snd_nxt
;
2661 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2662 tp
->prior_ssthresh
= 0;
2663 tp
->undo_marker
= 0;
2664 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2666 tcp_xmit_retransmit_queue(sk
);
2668 EXPORT_SYMBOL(tcp_simple_retransmit
);
2670 static void tcp_enter_recovery(struct sock
*sk
, bool ece_ack
)
2672 struct tcp_sock
*tp
= tcp_sk(sk
);
2675 if (tcp_is_reno(tp
))
2676 mib_idx
= LINUX_MIB_TCPRENORECOVERY
;
2678 mib_idx
= LINUX_MIB_TCPSACKRECOVERY
;
2680 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2682 tp
->prior_ssthresh
= 0;
2685 if (inet_csk(sk
)->icsk_ca_state
< TCP_CA_CWR
) {
2687 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2688 tcp_init_cwnd_reduction(sk
);
2690 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
2693 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2694 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2696 static void tcp_process_loss(struct sock
*sk
, int flag
, bool is_dupack
)
2698 struct tcp_sock
*tp
= tcp_sk(sk
);
2699 bool recovered
= !before(tp
->snd_una
, tp
->high_seq
);
2701 if (tp
->frto
) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2702 /* Step 3.b. A timeout is spurious if not all data are
2703 * lost, i.e., never-retransmitted data are (s)acked.
2705 if (tcp_try_undo_loss(sk
, flag
& FLAG_ORIG_SACK_ACKED
))
2708 if (after(tp
->snd_nxt
, tp
->high_seq
) &&
2709 (flag
& FLAG_DATA_SACKED
|| is_dupack
)) {
2710 tp
->frto
= 0; /* Loss was real: 2nd part of step 3.a */
2711 } else if (flag
& FLAG_SND_UNA_ADVANCED
&& !recovered
) {
2712 tp
->high_seq
= tp
->snd_nxt
;
2713 __tcp_push_pending_frames(sk
, tcp_current_mss(sk
),
2715 if (after(tp
->snd_nxt
, tp
->high_seq
))
2716 return; /* Step 2.b */
2722 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2723 tcp_try_undo_recovery(sk
);
2726 if (tcp_is_reno(tp
)) {
2727 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2728 * delivered. Lower inflight to clock out (re)tranmissions.
2730 if (after(tp
->snd_nxt
, tp
->high_seq
) && is_dupack
)
2731 tcp_add_reno_sack(sk
);
2732 else if (flag
& FLAG_SND_UNA_ADVANCED
)
2733 tcp_reset_reno_sack(tp
);
2735 if (tcp_try_undo_loss(sk
, false))
2737 tcp_xmit_retransmit_queue(sk
);
2740 /* Undo during fast recovery after partial ACK. */
2741 static bool tcp_try_undo_partial(struct sock
*sk
, const int acked
,
2742 const int prior_unsacked
)
2744 struct tcp_sock
*tp
= tcp_sk(sk
);
2746 if (tp
->undo_marker
&& tcp_packet_delayed(tp
)) {
2747 /* Plain luck! Hole if filled with delayed
2748 * packet, rather than with a retransmit.
2750 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
2752 /* We are getting evidence that the reordering degree is higher
2753 * than we realized. If there are no retransmits out then we
2754 * can undo. Otherwise we clock out new packets but do not
2755 * mark more packets lost or retransmit more.
2757 if (tp
->retrans_out
) {
2758 tcp_cwnd_reduction(sk
, prior_unsacked
, 0);
2762 if (!tcp_any_retrans_done(sk
))
2763 tp
->retrans_stamp
= 0;
2765 DBGUNDO(sk
, "partial recovery");
2766 tcp_undo_cwnd_reduction(sk
, true);
2767 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPARTIALUNDO
);
2768 tcp_try_keep_open(sk
);
2774 /* Process an event, which can update packets-in-flight not trivially.
2775 * Main goal of this function is to calculate new estimate for left_out,
2776 * taking into account both packets sitting in receiver's buffer and
2777 * packets lost by network.
2779 * Besides that it does CWND reduction, when packet loss is detected
2780 * and changes state of machine.
2782 * It does _not_ decide what to send, it is made in function
2783 * tcp_xmit_retransmit_queue().
2785 static void tcp_fastretrans_alert(struct sock
*sk
, const int acked
,
2786 const int prior_unsacked
,
2787 bool is_dupack
, int flag
)
2789 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2790 struct tcp_sock
*tp
= tcp_sk(sk
);
2791 bool do_lost
= is_dupack
|| ((flag
& FLAG_DATA_SACKED
) &&
2792 (tcp_fackets_out(tp
) > tp
->reordering
));
2793 int fast_rexmit
= 0;
2795 if (WARN_ON(!tp
->packets_out
&& tp
->sacked_out
))
2797 if (WARN_ON(!tp
->sacked_out
&& tp
->fackets_out
))
2798 tp
->fackets_out
= 0;
2800 /* Now state machine starts.
2801 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2802 if (flag
& FLAG_ECE
)
2803 tp
->prior_ssthresh
= 0;
2805 /* B. In all the states check for reneging SACKs. */
2806 if (tcp_check_sack_reneging(sk
, flag
))
2809 /* C. Check consistency of the current state. */
2810 tcp_verify_left_out(tp
);
2812 /* D. Check state exit conditions. State can be terminated
2813 * when high_seq is ACKed. */
2814 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
2815 WARN_ON(tp
->retrans_out
!= 0);
2816 tp
->retrans_stamp
= 0;
2817 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
2818 switch (icsk
->icsk_ca_state
) {
2820 /* CWR is to be held something *above* high_seq
2821 * is ACKed for CWR bit to reach receiver. */
2822 if (tp
->snd_una
!= tp
->high_seq
) {
2823 tcp_end_cwnd_reduction(sk
);
2824 tcp_set_ca_state(sk
, TCP_CA_Open
);
2828 case TCP_CA_Recovery
:
2829 if (tcp_is_reno(tp
))
2830 tcp_reset_reno_sack(tp
);
2831 if (tcp_try_undo_recovery(sk
))
2833 tcp_end_cwnd_reduction(sk
);
2838 /* E. Process state. */
2839 switch (icsk
->icsk_ca_state
) {
2840 case TCP_CA_Recovery
:
2841 if (!(flag
& FLAG_SND_UNA_ADVANCED
)) {
2842 if (tcp_is_reno(tp
) && is_dupack
)
2843 tcp_add_reno_sack(sk
);
2845 if (tcp_try_undo_partial(sk
, acked
, prior_unsacked
))
2847 /* Partial ACK arrived. Force fast retransmit. */
2848 do_lost
= tcp_is_reno(tp
) ||
2849 tcp_fackets_out(tp
) > tp
->reordering
;
2851 if (tcp_try_undo_dsack(sk
)) {
2852 tcp_try_keep_open(sk
);
2857 tcp_process_loss(sk
, flag
, is_dupack
);
2858 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
2860 /* Fall through to processing in Open state. */
2862 if (tcp_is_reno(tp
)) {
2863 if (flag
& FLAG_SND_UNA_ADVANCED
)
2864 tcp_reset_reno_sack(tp
);
2866 tcp_add_reno_sack(sk
);
2869 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
)
2870 tcp_try_undo_dsack(sk
);
2872 if (!tcp_time_to_recover(sk
, flag
)) {
2873 tcp_try_to_open(sk
, flag
, prior_unsacked
);
2877 /* MTU probe failure: don't reduce cwnd */
2878 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
2879 icsk
->icsk_mtup
.probe_size
&&
2880 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
2881 tcp_mtup_probe_failed(sk
);
2882 /* Restores the reduction we did in tcp_mtup_probe() */
2884 tcp_simple_retransmit(sk
);
2888 /* Otherwise enter Recovery state */
2889 tcp_enter_recovery(sk
, (flag
& FLAG_ECE
));
2894 tcp_update_scoreboard(sk
, fast_rexmit
);
2895 tcp_cwnd_reduction(sk
, prior_unsacked
, fast_rexmit
);
2896 tcp_xmit_retransmit_queue(sk
);
2899 static inline bool tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
2900 long seq_rtt_us
, long sack_rtt_us
)
2902 const struct tcp_sock
*tp
= tcp_sk(sk
);
2904 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2905 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2906 * Karn's algorithm forbids taking RTT if some retransmitted data
2907 * is acked (RFC6298).
2909 if (flag
& FLAG_RETRANS_DATA_ACKED
)
2913 seq_rtt_us
= sack_rtt_us
;
2915 /* RTTM Rule: A TSecr value received in a segment is used to
2916 * update the averaged RTT measurement only if the segment
2917 * acknowledges some new data, i.e., only if it advances the
2918 * left edge of the send window.
2919 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2921 if (seq_rtt_us
< 0 && tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2923 seq_rtt_us
= jiffies_to_usecs(tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
);
2928 tcp_rtt_estimator(sk
, seq_rtt_us
);
2931 /* RFC6298: only reset backoff on valid RTT measurement. */
2932 inet_csk(sk
)->icsk_backoff
= 0;
2936 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2937 static void tcp_synack_rtt_meas(struct sock
*sk
, const u32 synack_stamp
)
2939 struct tcp_sock
*tp
= tcp_sk(sk
);
2940 long seq_rtt_us
= -1L;
2942 if (synack_stamp
&& !tp
->total_retrans
)
2943 seq_rtt_us
= jiffies_to_usecs(tcp_time_stamp
- synack_stamp
);
2945 /* If the ACK acks both the SYNACK and the (Fast Open'd) data packets
2946 * sent in SYN_RECV, SYNACK RTT is the smooth RTT computed in tcp_ack()
2949 tcp_ack_update_rtt(sk
, FLAG_SYN_ACKED
, seq_rtt_us
, -1L);
2952 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
, u32 acked
)
2954 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2956 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, acked
);
2957 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_time_stamp
;
2960 /* Restart timer after forward progress on connection.
2961 * RFC2988 recommends to restart timer to now+rto.
2963 void tcp_rearm_rto(struct sock
*sk
)
2965 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2966 struct tcp_sock
*tp
= tcp_sk(sk
);
2968 /* If the retrans timer is currently being used by Fast Open
2969 * for SYN-ACK retrans purpose, stay put.
2971 if (tp
->fastopen_rsk
)
2974 if (!tp
->packets_out
) {
2975 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
2977 u32 rto
= inet_csk(sk
)->icsk_rto
;
2978 /* Offset the time elapsed after installing regular RTO */
2979 if (icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
||
2980 icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
) {
2981 struct sk_buff
*skb
= tcp_write_queue_head(sk
);
2982 const u32 rto_time_stamp
=
2983 tcp_skb_timestamp(skb
) + rto
;
2984 s32 delta
= (s32
)(rto_time_stamp
- tcp_time_stamp
);
2985 /* delta may not be positive if the socket is locked
2986 * when the retrans timer fires and is rescheduled.
2991 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
, rto
,
2996 /* This function is called when the delayed ER timer fires. TCP enters
2997 * fast recovery and performs fast-retransmit.
2999 void tcp_resume_early_retransmit(struct sock
*sk
)
3001 struct tcp_sock
*tp
= tcp_sk(sk
);
3005 /* Stop if ER is disabled after the delayed ER timer is scheduled */
3006 if (!tp
->do_early_retrans
)
3009 tcp_enter_recovery(sk
, false);
3010 tcp_update_scoreboard(sk
, 1);
3011 tcp_xmit_retransmit_queue(sk
);
3014 /* If we get here, the whole TSO packet has not been acked. */
3015 static u32
tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
)
3017 struct tcp_sock
*tp
= tcp_sk(sk
);
3020 BUG_ON(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
));
3022 packets_acked
= tcp_skb_pcount(skb
);
3023 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
3025 packets_acked
-= tcp_skb_pcount(skb
);
3027 if (packets_acked
) {
3028 BUG_ON(tcp_skb_pcount(skb
) == 0);
3029 BUG_ON(!before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
));
3032 return packets_acked
;
3035 static void tcp_ack_tstamp(struct sock
*sk
, struct sk_buff
*skb
,
3038 const struct skb_shared_info
*shinfo
;
3040 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3041 if (likely(!(sk
->sk_tsflags
& SOF_TIMESTAMPING_TX_ACK
)))
3044 shinfo
= skb_shinfo(skb
);
3045 if ((shinfo
->tx_flags
& SKBTX_ACK_TSTAMP
) &&
3046 between(shinfo
->tskey
, prior_snd_una
, tcp_sk(sk
)->snd_una
- 1))
3047 __skb_tstamp_tx(skb
, NULL
, sk
, SCM_TSTAMP_ACK
);
3050 /* Remove acknowledged frames from the retransmission queue. If our packet
3051 * is before the ack sequence we can discard it as it's confirmed to have
3052 * arrived at the other end.
3054 static int tcp_clean_rtx_queue(struct sock
*sk
, int prior_fackets
,
3055 u32 prior_snd_una
, long sack_rtt_us
)
3057 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3058 struct skb_mstamp first_ackt
, last_ackt
, now
;
3059 struct tcp_sock
*tp
= tcp_sk(sk
);
3060 u32 prior_sacked
= tp
->sacked_out
;
3061 u32 reord
= tp
->packets_out
;
3062 bool fully_acked
= true;
3063 long ca_seq_rtt_us
= -1L;
3064 long seq_rtt_us
= -1L;
3065 struct sk_buff
*skb
;
3072 while ((skb
= tcp_write_queue_head(sk
)) && skb
!= tcp_send_head(sk
)) {
3073 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
3074 u8 sacked
= scb
->sacked
;
3077 tcp_ack_tstamp(sk
, skb
, prior_snd_una
);
3079 /* Determine how many packets and what bytes were acked, tso and else */
3080 if (after(scb
->end_seq
, tp
->snd_una
)) {
3081 if (tcp_skb_pcount(skb
) == 1 ||
3082 !after(tp
->snd_una
, scb
->seq
))
3085 acked_pcount
= tcp_tso_acked(sk
, skb
);
3089 fully_acked
= false;
3091 /* Speedup tcp_unlink_write_queue() and next loop */
3092 prefetchw(skb
->next
);
3093 acked_pcount
= tcp_skb_pcount(skb
);
3096 if (unlikely(sacked
& TCPCB_RETRANS
)) {
3097 if (sacked
& TCPCB_SACKED_RETRANS
)
3098 tp
->retrans_out
-= acked_pcount
;
3099 flag
|= FLAG_RETRANS_DATA_ACKED
;
3100 } else if (!(sacked
& TCPCB_SACKED_ACKED
)) {
3101 last_ackt
= skb
->skb_mstamp
;
3102 WARN_ON_ONCE(last_ackt
.v64
== 0);
3103 if (!first_ackt
.v64
)
3104 first_ackt
= last_ackt
;
3106 reord
= min(pkts_acked
, reord
);
3107 if (!after(scb
->end_seq
, tp
->high_seq
))
3108 flag
|= FLAG_ORIG_SACK_ACKED
;
3111 if (sacked
& TCPCB_SACKED_ACKED
)
3112 tp
->sacked_out
-= acked_pcount
;
3113 if (sacked
& TCPCB_LOST
)
3114 tp
->lost_out
-= acked_pcount
;
3116 tp
->packets_out
-= acked_pcount
;
3117 pkts_acked
+= acked_pcount
;
3119 /* Initial outgoing SYN's get put onto the write_queue
3120 * just like anything else we transmit. It is not
3121 * true data, and if we misinform our callers that
3122 * this ACK acks real data, we will erroneously exit
3123 * connection startup slow start one packet too
3124 * quickly. This is severely frowned upon behavior.
3126 if (likely(!(scb
->tcp_flags
& TCPHDR_SYN
))) {
3127 flag
|= FLAG_DATA_ACKED
;
3129 flag
|= FLAG_SYN_ACKED
;
3130 tp
->retrans_stamp
= 0;
3136 tcp_unlink_write_queue(skb
, sk
);
3137 sk_wmem_free_skb(sk
, skb
);
3138 if (unlikely(skb
== tp
->retransmit_skb_hint
))
3139 tp
->retransmit_skb_hint
= NULL
;
3140 if (unlikely(skb
== tp
->lost_skb_hint
))
3141 tp
->lost_skb_hint
= NULL
;
3144 if (likely(between(tp
->snd_up
, prior_snd_una
, tp
->snd_una
)))
3145 tp
->snd_up
= tp
->snd_una
;
3147 if (skb
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
3148 flag
|= FLAG_SACK_RENEGING
;
3150 skb_mstamp_get(&now
);
3151 if (likely(first_ackt
.v64
)) {
3152 seq_rtt_us
= skb_mstamp_us_delta(&now
, &first_ackt
);
3153 ca_seq_rtt_us
= skb_mstamp_us_delta(&now
, &last_ackt
);
3156 rtt_update
= tcp_ack_update_rtt(sk
, flag
, seq_rtt_us
, sack_rtt_us
);
3158 if (flag
& FLAG_ACKED
) {
3159 const struct tcp_congestion_ops
*ca_ops
3160 = inet_csk(sk
)->icsk_ca_ops
;
3163 if (unlikely(icsk
->icsk_mtup
.probe_size
&&
3164 !after(tp
->mtu_probe
.probe_seq_end
, tp
->snd_una
))) {
3165 tcp_mtup_probe_success(sk
);
3168 if (tcp_is_reno(tp
)) {
3169 tcp_remove_reno_sacks(sk
, pkts_acked
);
3173 /* Non-retransmitted hole got filled? That's reordering */
3174 if (reord
< prior_fackets
)
3175 tcp_update_reordering(sk
, tp
->fackets_out
- reord
, 0);
3177 delta
= tcp_is_fack(tp
) ? pkts_acked
:
3178 prior_sacked
- tp
->sacked_out
;
3179 tp
->lost_cnt_hint
-= min(tp
->lost_cnt_hint
, delta
);
3182 tp
->fackets_out
-= min(pkts_acked
, tp
->fackets_out
);
3184 if (ca_ops
->pkts_acked
) {
3185 long rtt_us
= min_t(ulong
, ca_seq_rtt_us
, sack_rtt_us
);
3186 ca_ops
->pkts_acked(sk
, pkts_acked
, rtt_us
);
3189 } else if (skb
&& rtt_update
&& sack_rtt_us
>= 0 &&
3190 sack_rtt_us
> skb_mstamp_us_delta(&now
, &skb
->skb_mstamp
)) {
3191 /* Do not re-arm RTO if the sack RTT is measured from data sent
3192 * after when the head was last (re)transmitted. Otherwise the
3193 * timeout may continue to extend in loss recovery.
3198 #if FASTRETRANS_DEBUG > 0
3199 WARN_ON((int)tp
->sacked_out
< 0);
3200 WARN_ON((int)tp
->lost_out
< 0);
3201 WARN_ON((int)tp
->retrans_out
< 0);
3202 if (!tp
->packets_out
&& tcp_is_sack(tp
)) {
3203 icsk
= inet_csk(sk
);
3205 pr_debug("Leak l=%u %d\n",
3206 tp
->lost_out
, icsk
->icsk_ca_state
);
3209 if (tp
->sacked_out
) {
3210 pr_debug("Leak s=%u %d\n",
3211 tp
->sacked_out
, icsk
->icsk_ca_state
);
3214 if (tp
->retrans_out
) {
3215 pr_debug("Leak r=%u %d\n",
3216 tp
->retrans_out
, icsk
->icsk_ca_state
);
3217 tp
->retrans_out
= 0;
3224 static void tcp_ack_probe(struct sock
*sk
)
3226 const struct tcp_sock
*tp
= tcp_sk(sk
);
3227 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3229 /* Was it a usable window open? */
3231 if (!after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
, tcp_wnd_end(tp
))) {
3232 icsk
->icsk_backoff
= 0;
3233 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
3234 /* Socket must be waked up by subsequent tcp_data_snd_check().
3235 * This function is not for random using!
3238 unsigned long when
= inet_csk_rto_backoff(icsk
, TCP_RTO_MAX
);
3240 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3245 static inline bool tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
3247 return !(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
3248 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
;
3251 /* Decide wheather to run the increase function of congestion control. */
3252 static inline bool tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
3254 if (tcp_in_cwnd_reduction(sk
))
3257 /* If reordering is high then always grow cwnd whenever data is
3258 * delivered regardless of its ordering. Otherwise stay conservative
3259 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3260 * new SACK or ECE mark may first advance cwnd here and later reduce
3261 * cwnd in tcp_fastretrans_alert() based on more states.
3263 if (tcp_sk(sk
)->reordering
> sysctl_tcp_reordering
)
3264 return flag
& FLAG_FORWARD_PROGRESS
;
3266 return flag
& FLAG_DATA_ACKED
;
3269 /* Check that window update is acceptable.
3270 * The function assumes that snd_una<=ack<=snd_next.
3272 static inline bool tcp_may_update_window(const struct tcp_sock
*tp
,
3273 const u32 ack
, const u32 ack_seq
,
3276 return after(ack
, tp
->snd_una
) ||
3277 after(ack_seq
, tp
->snd_wl1
) ||
3278 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
);
3281 /* If we update tp->snd_una, also update tp->bytes_acked */
3282 static void tcp_snd_una_update(struct tcp_sock
*tp
, u32 ack
)
3284 u32 delta
= ack
- tp
->snd_una
;
3286 tp
->bytes_acked
+= delta
;
3290 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3291 static void tcp_rcv_nxt_update(struct tcp_sock
*tp
, u32 seq
)
3293 u32 delta
= seq
- tp
->rcv_nxt
;
3295 tp
->bytes_received
+= delta
;
3299 /* Update our send window.
3301 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3302 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3304 static int tcp_ack_update_window(struct sock
*sk
, const struct sk_buff
*skb
, u32 ack
,
3307 struct tcp_sock
*tp
= tcp_sk(sk
);
3309 u32 nwin
= ntohs(tcp_hdr(skb
)->window
);
3311 if (likely(!tcp_hdr(skb
)->syn
))
3312 nwin
<<= tp
->rx_opt
.snd_wscale
;
3314 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
3315 flag
|= FLAG_WIN_UPDATE
;
3316 tcp_update_wl(tp
, ack_seq
);
3318 if (tp
->snd_wnd
!= nwin
) {
3321 /* Note, it is the only place, where
3322 * fast path is recovered for sending TCP.
3325 tcp_fast_path_check(sk
);
3327 if (nwin
> tp
->max_window
) {
3328 tp
->max_window
= nwin
;
3329 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
3334 tcp_snd_una_update(tp
, ack
);
3339 /* Return true if we're currently rate-limiting out-of-window ACKs and
3340 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3341 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3342 * attacks that send repeated SYNs or ACKs for the same connection. To
3343 * do this, we do not send a duplicate SYNACK or ACK if the remote
3344 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3346 bool tcp_oow_rate_limited(struct net
*net
, const struct sk_buff
*skb
,
3347 int mib_idx
, u32
*last_oow_ack_time
)
3349 /* Data packets without SYNs are not likely part of an ACK loop. */
3350 if ((TCP_SKB_CB(skb
)->seq
!= TCP_SKB_CB(skb
)->end_seq
) &&
3352 goto not_rate_limited
;
3354 if (*last_oow_ack_time
) {
3355 s32 elapsed
= (s32
)(tcp_time_stamp
- *last_oow_ack_time
);
3357 if (0 <= elapsed
&& elapsed
< sysctl_tcp_invalid_ratelimit
) {
3358 NET_INC_STATS_BH(net
, mib_idx
);
3359 return true; /* rate-limited: don't send yet! */
3363 *last_oow_ack_time
= tcp_time_stamp
;
3366 return false; /* not rate-limited: go ahead, send dupack now! */
3369 /* RFC 5961 7 [ACK Throttling] */
3370 static void tcp_send_challenge_ack(struct sock
*sk
, const struct sk_buff
*skb
)
3372 /* unprotected vars, we dont care of overwrites */
3373 static u32 challenge_timestamp
;
3374 static unsigned int challenge_count
;
3375 struct tcp_sock
*tp
= tcp_sk(sk
);
3378 /* First check our per-socket dupack rate limit. */
3379 if (tcp_oow_rate_limited(sock_net(sk
), skb
,
3380 LINUX_MIB_TCPACKSKIPPEDCHALLENGE
,
3381 &tp
->last_oow_ack_time
))
3384 /* Then check the check host-wide RFC 5961 rate limit. */
3386 if (now
!= challenge_timestamp
) {
3387 challenge_timestamp
= now
;
3388 challenge_count
= 0;
3390 if (++challenge_count
<= sysctl_tcp_challenge_ack_limit
) {
3391 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPCHALLENGEACK
);
3396 static void tcp_store_ts_recent(struct tcp_sock
*tp
)
3398 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
3399 tp
->rx_opt
.ts_recent_stamp
= get_seconds();
3402 static void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
3404 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
3405 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3406 * extra check below makes sure this can only happen
3407 * for pure ACK frames. -DaveM
3409 * Not only, also it occurs for expired timestamps.
3412 if (tcp_paws_check(&tp
->rx_opt
, 0))
3413 tcp_store_ts_recent(tp
);
3417 /* This routine deals with acks during a TLP episode.
3418 * We mark the end of a TLP episode on receiving TLP dupack or when
3419 * ack is after tlp_high_seq.
3420 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3422 static void tcp_process_tlp_ack(struct sock
*sk
, u32 ack
, int flag
)
3424 struct tcp_sock
*tp
= tcp_sk(sk
);
3426 if (before(ack
, tp
->tlp_high_seq
))
3429 if (flag
& FLAG_DSACKING_ACK
) {
3430 /* This DSACK means original and TLP probe arrived; no loss */
3431 tp
->tlp_high_seq
= 0;
3432 } else if (after(ack
, tp
->tlp_high_seq
)) {
3433 /* ACK advances: there was a loss, so reduce cwnd. Reset
3434 * tlp_high_seq in tcp_init_cwnd_reduction()
3436 tcp_init_cwnd_reduction(sk
);
3437 tcp_set_ca_state(sk
, TCP_CA_CWR
);
3438 tcp_end_cwnd_reduction(sk
);
3439 tcp_try_keep_open(sk
);
3440 NET_INC_STATS_BH(sock_net(sk
),
3441 LINUX_MIB_TCPLOSSPROBERECOVERY
);
3442 } else if (!(flag
& (FLAG_SND_UNA_ADVANCED
|
3443 FLAG_NOT_DUP
| FLAG_DATA_SACKED
))) {
3444 /* Pure dupack: original and TLP probe arrived; no loss */
3445 tp
->tlp_high_seq
= 0;
3449 static inline void tcp_in_ack_event(struct sock
*sk
, u32 flags
)
3451 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3453 if (icsk
->icsk_ca_ops
->in_ack_event
)
3454 icsk
->icsk_ca_ops
->in_ack_event(sk
, flags
);
3457 /* This routine deals with incoming acks, but not outgoing ones. */
3458 static int tcp_ack(struct sock
*sk
, const struct sk_buff
*skb
, int flag
)
3460 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3461 struct tcp_sock
*tp
= tcp_sk(sk
);
3462 u32 prior_snd_una
= tp
->snd_una
;
3463 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
3464 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3465 bool is_dupack
= false;
3467 int prior_packets
= tp
->packets_out
;
3468 const int prior_unsacked
= tp
->packets_out
- tp
->sacked_out
;
3469 int acked
= 0; /* Number of packets newly acked */
3470 long sack_rtt_us
= -1L;
3472 /* We very likely will need to access write queue head. */
3473 prefetchw(sk
->sk_write_queue
.next
);
3475 /* If the ack is older than previous acks
3476 * then we can probably ignore it.
3478 if (before(ack
, prior_snd_una
)) {
3479 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3480 if (before(ack
, prior_snd_una
- tp
->max_window
)) {
3481 tcp_send_challenge_ack(sk
, skb
);
3487 /* If the ack includes data we haven't sent yet, discard
3488 * this segment (RFC793 Section 3.9).
3490 if (after(ack
, tp
->snd_nxt
))
3493 if (icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
||
3494 icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
)
3497 if (after(ack
, prior_snd_una
)) {
3498 flag
|= FLAG_SND_UNA_ADVANCED
;
3499 icsk
->icsk_retransmits
= 0;
3502 prior_fackets
= tp
->fackets_out
;
3504 /* ts_recent update must be made after we are sure that the packet
3507 if (flag
& FLAG_UPDATE_TS_RECENT
)
3508 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
3510 if (!(flag
& FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
3511 /* Window is constant, pure forward advance.
3512 * No more checks are required.
3513 * Note, we use the fact that SND.UNA>=SND.WL2.
3515 tcp_update_wl(tp
, ack_seq
);
3516 tcp_snd_una_update(tp
, ack
);
3517 flag
|= FLAG_WIN_UPDATE
;
3519 tcp_in_ack_event(sk
, CA_ACK_WIN_UPDATE
);
3521 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPACKS
);
3523 u32 ack_ev_flags
= CA_ACK_SLOWPATH
;
3525 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
3528 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPUREACKS
);
3530 flag
|= tcp_ack_update_window(sk
, skb
, ack
, ack_seq
);
3532 if (TCP_SKB_CB(skb
)->sacked
)
3533 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
,
3536 if (tcp_ecn_rcv_ecn_echo(tp
, tcp_hdr(skb
))) {
3538 ack_ev_flags
|= CA_ACK_ECE
;
3541 if (flag
& FLAG_WIN_UPDATE
)
3542 ack_ev_flags
|= CA_ACK_WIN_UPDATE
;
3544 tcp_in_ack_event(sk
, ack_ev_flags
);
3547 /* We passed data and got it acked, remove any soft error
3548 * log. Something worked...
3550 sk
->sk_err_soft
= 0;
3551 icsk
->icsk_probes_out
= 0;
3552 tp
->rcv_tstamp
= tcp_time_stamp
;
3556 /* See if we can take anything off of the retransmit queue. */
3557 acked
= tp
->packets_out
;
3558 flag
|= tcp_clean_rtx_queue(sk
, prior_fackets
, prior_snd_una
,
3560 acked
-= tp
->packets_out
;
3562 /* Advance cwnd if state allows */
3563 if (tcp_may_raise_cwnd(sk
, flag
))
3564 tcp_cong_avoid(sk
, ack
, acked
);
3566 if (tcp_ack_is_dubious(sk
, flag
)) {
3567 is_dupack
= !(flag
& (FLAG_SND_UNA_ADVANCED
| FLAG_NOT_DUP
));
3568 tcp_fastretrans_alert(sk
, acked
, prior_unsacked
,
3571 if (tp
->tlp_high_seq
)
3572 tcp_process_tlp_ack(sk
, ack
, flag
);
3574 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
& FLAG_NOT_DUP
)) {
3575 struct dst_entry
*dst
= __sk_dst_get(sk
);
3580 if (icsk
->icsk_pending
== ICSK_TIME_RETRANS
)
3581 tcp_schedule_loss_probe(sk
);
3582 tcp_update_pacing_rate(sk
);
3586 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3587 if (flag
& FLAG_DSACKING_ACK
)
3588 tcp_fastretrans_alert(sk
, acked
, prior_unsacked
,
3590 /* If this ack opens up a zero window, clear backoff. It was
3591 * being used to time the probes, and is probably far higher than
3592 * it needs to be for normal retransmission.
3594 if (tcp_send_head(sk
))
3597 if (tp
->tlp_high_seq
)
3598 tcp_process_tlp_ack(sk
, ack
, flag
);
3602 SOCK_DEBUG(sk
, "Ack %u after %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3606 /* If data was SACKed, tag it and see if we should send more data.
3607 * If data was DSACKed, see if we can undo a cwnd reduction.
3609 if (TCP_SKB_CB(skb
)->sacked
) {
3610 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
,
3612 tcp_fastretrans_alert(sk
, acked
, prior_unsacked
,
3616 SOCK_DEBUG(sk
, "Ack %u before %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3620 static void tcp_parse_fastopen_option(int len
, const unsigned char *cookie
,
3621 bool syn
, struct tcp_fastopen_cookie
*foc
,
3624 /* Valid only in SYN or SYN-ACK with an even length. */
3625 if (!foc
|| !syn
|| len
< 0 || (len
& 1))
3628 if (len
>= TCP_FASTOPEN_COOKIE_MIN
&&
3629 len
<= TCP_FASTOPEN_COOKIE_MAX
)
3630 memcpy(foc
->val
, cookie
, len
);
3637 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3638 * But, this can also be called on packets in the established flow when
3639 * the fast version below fails.
3641 void tcp_parse_options(const struct sk_buff
*skb
,
3642 struct tcp_options_received
*opt_rx
, int estab
,
3643 struct tcp_fastopen_cookie
*foc
)
3645 const unsigned char *ptr
;
3646 const struct tcphdr
*th
= tcp_hdr(skb
);
3647 int length
= (th
->doff
* 4) - sizeof(struct tcphdr
);
3649 ptr
= (const unsigned char *)(th
+ 1);
3650 opt_rx
->saw_tstamp
= 0;
3652 while (length
> 0) {
3653 int opcode
= *ptr
++;
3659 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
3664 if (opsize
< 2) /* "silly options" */
3666 if (opsize
> length
)
3667 return; /* don't parse partial options */
3670 if (opsize
== TCPOLEN_MSS
&& th
->syn
&& !estab
) {
3671 u16 in_mss
= get_unaligned_be16(ptr
);
3673 if (opt_rx
->user_mss
&&
3674 opt_rx
->user_mss
< in_mss
)
3675 in_mss
= opt_rx
->user_mss
;
3676 opt_rx
->mss_clamp
= in_mss
;
3681 if (opsize
== TCPOLEN_WINDOW
&& th
->syn
&&
3682 !estab
&& sysctl_tcp_window_scaling
) {
3683 __u8 snd_wscale
= *(__u8
*)ptr
;
3684 opt_rx
->wscale_ok
= 1;
3685 if (snd_wscale
> 14) {
3686 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3691 opt_rx
->snd_wscale
= snd_wscale
;
3694 case TCPOPT_TIMESTAMP
:
3695 if ((opsize
== TCPOLEN_TIMESTAMP
) &&
3696 ((estab
&& opt_rx
->tstamp_ok
) ||
3697 (!estab
&& sysctl_tcp_timestamps
))) {
3698 opt_rx
->saw_tstamp
= 1;
3699 opt_rx
->rcv_tsval
= get_unaligned_be32(ptr
);
3700 opt_rx
->rcv_tsecr
= get_unaligned_be32(ptr
+ 4);
3703 case TCPOPT_SACK_PERM
:
3704 if (opsize
== TCPOLEN_SACK_PERM
&& th
->syn
&&
3705 !estab
&& sysctl_tcp_sack
) {
3706 opt_rx
->sack_ok
= TCP_SACK_SEEN
;
3707 tcp_sack_reset(opt_rx
);
3712 if ((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
3713 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
3715 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
3718 #ifdef CONFIG_TCP_MD5SIG
3721 * The MD5 Hash has already been
3722 * checked (see tcp_v{4,6}_do_rcv()).
3726 case TCPOPT_FASTOPEN
:
3727 tcp_parse_fastopen_option(
3728 opsize
- TCPOLEN_FASTOPEN_BASE
,
3729 ptr
, th
->syn
, foc
, false);
3733 /* Fast Open option shares code 254 using a
3734 * 16 bits magic number.
3736 if (opsize
>= TCPOLEN_EXP_FASTOPEN_BASE
&&
3737 get_unaligned_be16(ptr
) ==
3738 TCPOPT_FASTOPEN_MAGIC
)
3739 tcp_parse_fastopen_option(opsize
-
3740 TCPOLEN_EXP_FASTOPEN_BASE
,
3741 ptr
+ 2, th
->syn
, foc
, true);
3750 EXPORT_SYMBOL(tcp_parse_options
);
3752 static bool tcp_parse_aligned_timestamp(struct tcp_sock
*tp
, const struct tcphdr
*th
)
3754 const __be32
*ptr
= (const __be32
*)(th
+ 1);
3756 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3757 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
3758 tp
->rx_opt
.saw_tstamp
= 1;
3760 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3763 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
) - tp
->tsoffset
;
3765 tp
->rx_opt
.rcv_tsecr
= 0;
3771 /* Fast parse options. This hopes to only see timestamps.
3772 * If it is wrong it falls back on tcp_parse_options().
3774 static bool tcp_fast_parse_options(const struct sk_buff
*skb
,
3775 const struct tcphdr
*th
, struct tcp_sock
*tp
)
3777 /* In the spirit of fast parsing, compare doff directly to constant
3778 * values. Because equality is used, short doff can be ignored here.
3780 if (th
->doff
== (sizeof(*th
) / 4)) {
3781 tp
->rx_opt
.saw_tstamp
= 0;
3783 } else if (tp
->rx_opt
.tstamp_ok
&&
3784 th
->doff
== ((sizeof(*th
) + TCPOLEN_TSTAMP_ALIGNED
) / 4)) {
3785 if (tcp_parse_aligned_timestamp(tp
, th
))
3789 tcp_parse_options(skb
, &tp
->rx_opt
, 1, NULL
);
3790 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
3791 tp
->rx_opt
.rcv_tsecr
-= tp
->tsoffset
;
3796 #ifdef CONFIG_TCP_MD5SIG
3798 * Parse MD5 Signature option
3800 const u8
*tcp_parse_md5sig_option(const struct tcphdr
*th
)
3802 int length
= (th
->doff
<< 2) - sizeof(*th
);
3803 const u8
*ptr
= (const u8
*)(th
+ 1);
3805 /* If the TCP option is too short, we can short cut */
3806 if (length
< TCPOLEN_MD5SIG
)
3809 while (length
> 0) {
3810 int opcode
= *ptr
++;
3821 if (opsize
< 2 || opsize
> length
)
3823 if (opcode
== TCPOPT_MD5SIG
)
3824 return opsize
== TCPOLEN_MD5SIG
? ptr
: NULL
;
3831 EXPORT_SYMBOL(tcp_parse_md5sig_option
);
3834 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3836 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3837 * it can pass through stack. So, the following predicate verifies that
3838 * this segment is not used for anything but congestion avoidance or
3839 * fast retransmit. Moreover, we even are able to eliminate most of such
3840 * second order effects, if we apply some small "replay" window (~RTO)
3841 * to timestamp space.
3843 * All these measures still do not guarantee that we reject wrapped ACKs
3844 * on networks with high bandwidth, when sequence space is recycled fastly,
3845 * but it guarantees that such events will be very rare and do not affect
3846 * connection seriously. This doesn't look nice, but alas, PAWS is really
3849 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3850 * states that events when retransmit arrives after original data are rare.
3851 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3852 * the biggest problem on large power networks even with minor reordering.
3853 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3854 * up to bandwidth of 18Gigabit/sec. 8) ]
3857 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
3859 const struct tcp_sock
*tp
= tcp_sk(sk
);
3860 const struct tcphdr
*th
= tcp_hdr(skb
);
3861 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3862 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3864 return (/* 1. Pure ACK with correct sequence number. */
3865 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
3867 /* 2. ... and duplicate ACK. */
3868 ack
== tp
->snd_una
&&
3870 /* 3. ... and does not update window. */
3871 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
3873 /* 4. ... and sits in replay window. */
3874 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
3877 static inline bool tcp_paws_discard(const struct sock
*sk
,
3878 const struct sk_buff
*skb
)
3880 const struct tcp_sock
*tp
= tcp_sk(sk
);
3882 return !tcp_paws_check(&tp
->rx_opt
, TCP_PAWS_WINDOW
) &&
3883 !tcp_disordered_ack(sk
, skb
);
3886 /* Check segment sequence number for validity.
3888 * Segment controls are considered valid, if the segment
3889 * fits to the window after truncation to the window. Acceptability
3890 * of data (and SYN, FIN, of course) is checked separately.
3891 * See tcp_data_queue(), for example.
3893 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3894 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3895 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3896 * (borrowed from freebsd)
3899 static inline bool tcp_sequence(const struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3901 return !before(end_seq
, tp
->rcv_wup
) &&
3902 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
3905 /* When we get a reset we do this. */
3906 void tcp_reset(struct sock
*sk
)
3908 /* We want the right error as BSD sees it (and indeed as we do). */
3909 switch (sk
->sk_state
) {
3911 sk
->sk_err
= ECONNREFUSED
;
3913 case TCP_CLOSE_WAIT
:
3919 sk
->sk_err
= ECONNRESET
;
3921 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3924 if (!sock_flag(sk
, SOCK_DEAD
))
3925 sk
->sk_error_report(sk
);
3931 * Process the FIN bit. This now behaves as it is supposed to work
3932 * and the FIN takes effect when it is validly part of sequence
3933 * space. Not before when we get holes.
3935 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3936 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3939 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3940 * close and we go into CLOSING (and later onto TIME-WAIT)
3942 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3944 static void tcp_fin(struct sock
*sk
)
3946 struct tcp_sock
*tp
= tcp_sk(sk
);
3947 const struct dst_entry
*dst
;
3949 inet_csk_schedule_ack(sk
);
3951 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
3952 sock_set_flag(sk
, SOCK_DONE
);
3954 switch (sk
->sk_state
) {
3956 case TCP_ESTABLISHED
:
3957 /* Move to CLOSE_WAIT */
3958 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
3959 dst
= __sk_dst_get(sk
);
3960 if (!dst
|| !dst_metric(dst
, RTAX_QUICKACK
))
3961 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
3964 case TCP_CLOSE_WAIT
:
3966 /* Received a retransmission of the FIN, do
3971 /* RFC793: Remain in the LAST-ACK state. */
3975 /* This case occurs when a simultaneous close
3976 * happens, we must ack the received FIN and
3977 * enter the CLOSING state.
3980 tcp_set_state(sk
, TCP_CLOSING
);
3983 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3985 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
3988 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3989 * cases we should never reach this piece of code.
3991 pr_err("%s: Impossible, sk->sk_state=%d\n",
3992 __func__
, sk
->sk_state
);
3996 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3997 * Probably, we should reset in this case. For now drop them.
3999 __skb_queue_purge(&tp
->out_of_order_queue
);
4000 if (tcp_is_sack(tp
))
4001 tcp_sack_reset(&tp
->rx_opt
);
4004 if (!sock_flag(sk
, SOCK_DEAD
)) {
4005 sk
->sk_state_change(sk
);
4007 /* Do not send POLL_HUP for half duplex close. */
4008 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
4009 sk
->sk_state
== TCP_CLOSE
)
4010 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_HUP
);
4012 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_IN
);
4016 static inline bool tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
,
4019 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
4020 if (before(seq
, sp
->start_seq
))
4021 sp
->start_seq
= seq
;
4022 if (after(end_seq
, sp
->end_seq
))
4023 sp
->end_seq
= end_seq
;
4029 static void tcp_dsack_set(struct sock
*sk
, u32 seq
, u32 end_seq
)
4031 struct tcp_sock
*tp
= tcp_sk(sk
);
4033 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
4036 if (before(seq
, tp
->rcv_nxt
))
4037 mib_idx
= LINUX_MIB_TCPDSACKOLDSENT
;
4039 mib_idx
= LINUX_MIB_TCPDSACKOFOSENT
;
4041 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
4043 tp
->rx_opt
.dsack
= 1;
4044 tp
->duplicate_sack
[0].start_seq
= seq
;
4045 tp
->duplicate_sack
[0].end_seq
= end_seq
;
4049 static void tcp_dsack_extend(struct sock
*sk
, u32 seq
, u32 end_seq
)
4051 struct tcp_sock
*tp
= tcp_sk(sk
);
4053 if (!tp
->rx_opt
.dsack
)
4054 tcp_dsack_set(sk
, seq
, end_seq
);
4056 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
4059 static void tcp_send_dupack(struct sock
*sk
, const struct sk_buff
*skb
)
4061 struct tcp_sock
*tp
= tcp_sk(sk
);
4063 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
4064 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4065 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4066 tcp_enter_quickack_mode(sk
);
4068 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
4069 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4071 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
4072 end_seq
= tp
->rcv_nxt
;
4073 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, end_seq
);
4080 /* These routines update the SACK block as out-of-order packets arrive or
4081 * in-order packets close up the sequence space.
4083 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
4086 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4087 struct tcp_sack_block
*swalk
= sp
+ 1;
4089 /* See if the recent change to the first SACK eats into
4090 * or hits the sequence space of other SACK blocks, if so coalesce.
4092 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
;) {
4093 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
4096 /* Zap SWALK, by moving every further SACK up by one slot.
4097 * Decrease num_sacks.
4099 tp
->rx_opt
.num_sacks
--;
4100 for (i
= this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
4104 this_sack
++, swalk
++;
4108 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
4110 struct tcp_sock
*tp
= tcp_sk(sk
);
4111 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4112 int cur_sacks
= tp
->rx_opt
.num_sacks
;
4118 for (this_sack
= 0; this_sack
< cur_sacks
; this_sack
++, sp
++) {
4119 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
4120 /* Rotate this_sack to the first one. */
4121 for (; this_sack
> 0; this_sack
--, sp
--)
4122 swap(*sp
, *(sp
- 1));
4124 tcp_sack_maybe_coalesce(tp
);
4129 /* Could not find an adjacent existing SACK, build a new one,
4130 * put it at the front, and shift everyone else down. We
4131 * always know there is at least one SACK present already here.
4133 * If the sack array is full, forget about the last one.
4135 if (this_sack
>= TCP_NUM_SACKS
) {
4137 tp
->rx_opt
.num_sacks
--;
4140 for (; this_sack
> 0; this_sack
--, sp
--)
4144 /* Build the new head SACK, and we're done. */
4145 sp
->start_seq
= seq
;
4146 sp
->end_seq
= end_seq
;
4147 tp
->rx_opt
.num_sacks
++;
4150 /* RCV.NXT advances, some SACKs should be eaten. */
4152 static void tcp_sack_remove(struct tcp_sock
*tp
)
4154 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4155 int num_sacks
= tp
->rx_opt
.num_sacks
;
4158 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4159 if (skb_queue_empty(&tp
->out_of_order_queue
)) {
4160 tp
->rx_opt
.num_sacks
= 0;
4164 for (this_sack
= 0; this_sack
< num_sacks
;) {
4165 /* Check if the start of the sack is covered by RCV.NXT. */
4166 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
4169 /* RCV.NXT must cover all the block! */
4170 WARN_ON(before(tp
->rcv_nxt
, sp
->end_seq
));
4172 /* Zap this SACK, by moving forward any other SACKS. */
4173 for (i
= this_sack
+1; i
< num_sacks
; i
++)
4174 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
4181 tp
->rx_opt
.num_sacks
= num_sacks
;
4185 * tcp_try_coalesce - try to merge skb to prior one
4188 * @from: buffer to add in queue
4189 * @fragstolen: pointer to boolean
4191 * Before queueing skb @from after @to, try to merge them
4192 * to reduce overall memory use and queue lengths, if cost is small.
4193 * Packets in ofo or receive queues can stay a long time.
4194 * Better try to coalesce them right now to avoid future collapses.
4195 * Returns true if caller should free @from instead of queueing it
4197 static bool tcp_try_coalesce(struct sock
*sk
,
4199 struct sk_buff
*from
,
4204 *fragstolen
= false;
4206 /* Its possible this segment overlaps with prior segment in queue */
4207 if (TCP_SKB_CB(from
)->seq
!= TCP_SKB_CB(to
)->end_seq
)
4210 if (!skb_try_coalesce(to
, from
, fragstolen
, &delta
))
4213 atomic_add(delta
, &sk
->sk_rmem_alloc
);
4214 sk_mem_charge(sk
, delta
);
4215 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRCVCOALESCE
);
4216 TCP_SKB_CB(to
)->end_seq
= TCP_SKB_CB(from
)->end_seq
;
4217 TCP_SKB_CB(to
)->ack_seq
= TCP_SKB_CB(from
)->ack_seq
;
4218 TCP_SKB_CB(to
)->tcp_flags
|= TCP_SKB_CB(from
)->tcp_flags
;
4222 /* This one checks to see if we can put data from the
4223 * out_of_order queue into the receive_queue.
4225 static void tcp_ofo_queue(struct sock
*sk
)
4227 struct tcp_sock
*tp
= tcp_sk(sk
);
4228 __u32 dsack_high
= tp
->rcv_nxt
;
4229 struct sk_buff
*skb
, *tail
;
4230 bool fragstolen
, eaten
;
4232 while ((skb
= skb_peek(&tp
->out_of_order_queue
)) != NULL
) {
4233 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
4236 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
4237 __u32 dsack
= dsack_high
;
4238 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
4239 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
4240 tcp_dsack_extend(sk
, TCP_SKB_CB(skb
)->seq
, dsack
);
4243 __skb_unlink(skb
, &tp
->out_of_order_queue
);
4244 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4245 SOCK_DEBUG(sk
, "ofo packet was already received\n");
4249 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
4250 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4251 TCP_SKB_CB(skb
)->end_seq
);
4253 tail
= skb_peek_tail(&sk
->sk_receive_queue
);
4254 eaten
= tail
&& tcp_try_coalesce(sk
, tail
, skb
, &fragstolen
);
4255 tcp_rcv_nxt_update(tp
, TCP_SKB_CB(skb
)->end_seq
);
4257 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4258 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
4261 kfree_skb_partial(skb
, fragstolen
);
4265 static bool tcp_prune_ofo_queue(struct sock
*sk
);
4266 static int tcp_prune_queue(struct sock
*sk
);
4268 static int tcp_try_rmem_schedule(struct sock
*sk
, struct sk_buff
*skb
,
4271 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
4272 !sk_rmem_schedule(sk
, skb
, size
)) {
4274 if (tcp_prune_queue(sk
) < 0)
4277 if (!sk_rmem_schedule(sk
, skb
, size
)) {
4278 if (!tcp_prune_ofo_queue(sk
))
4281 if (!sk_rmem_schedule(sk
, skb
, size
))
4288 static void tcp_data_queue_ofo(struct sock
*sk
, struct sk_buff
*skb
)
4290 struct tcp_sock
*tp
= tcp_sk(sk
);
4291 struct sk_buff
*skb1
;
4294 tcp_ecn_check_ce(tp
, skb
);
4296 if (unlikely(tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))) {
4297 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFODROP
);
4302 /* Disable header prediction. */
4304 inet_csk_schedule_ack(sk
);
4306 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFOQUEUE
);
4307 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
4308 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4310 skb1
= skb_peek_tail(&tp
->out_of_order_queue
);
4312 /* Initial out of order segment, build 1 SACK. */
4313 if (tcp_is_sack(tp
)) {
4314 tp
->rx_opt
.num_sacks
= 1;
4315 tp
->selective_acks
[0].start_seq
= TCP_SKB_CB(skb
)->seq
;
4316 tp
->selective_acks
[0].end_seq
=
4317 TCP_SKB_CB(skb
)->end_seq
;
4319 __skb_queue_head(&tp
->out_of_order_queue
, skb
);
4323 seq
= TCP_SKB_CB(skb
)->seq
;
4324 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4326 if (seq
== TCP_SKB_CB(skb1
)->end_seq
) {
4329 if (!tcp_try_coalesce(sk
, skb1
, skb
, &fragstolen
)) {
4330 __skb_queue_after(&tp
->out_of_order_queue
, skb1
, skb
);
4332 tcp_grow_window(sk
, skb
);
4333 kfree_skb_partial(skb
, fragstolen
);
4337 if (!tp
->rx_opt
.num_sacks
||
4338 tp
->selective_acks
[0].end_seq
!= seq
)
4341 /* Common case: data arrive in order after hole. */
4342 tp
->selective_acks
[0].end_seq
= end_seq
;
4346 /* Find place to insert this segment. */
4348 if (!after(TCP_SKB_CB(skb1
)->seq
, seq
))
4350 if (skb_queue_is_first(&tp
->out_of_order_queue
, skb1
)) {
4354 skb1
= skb_queue_prev(&tp
->out_of_order_queue
, skb1
);
4357 /* Do skb overlap to previous one? */
4358 if (skb1
&& before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4359 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4360 /* All the bits are present. Drop. */
4361 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFOMERGE
);
4364 tcp_dsack_set(sk
, seq
, end_seq
);
4367 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
4368 /* Partial overlap. */
4369 tcp_dsack_set(sk
, seq
,
4370 TCP_SKB_CB(skb1
)->end_seq
);
4372 if (skb_queue_is_first(&tp
->out_of_order_queue
,
4376 skb1
= skb_queue_prev(
4377 &tp
->out_of_order_queue
,
4382 __skb_queue_head(&tp
->out_of_order_queue
, skb
);
4384 __skb_queue_after(&tp
->out_of_order_queue
, skb1
, skb
);
4386 /* And clean segments covered by new one as whole. */
4387 while (!skb_queue_is_last(&tp
->out_of_order_queue
, skb
)) {
4388 skb1
= skb_queue_next(&tp
->out_of_order_queue
, skb
);
4390 if (!after(end_seq
, TCP_SKB_CB(skb1
)->seq
))
4392 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4393 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4397 __skb_unlink(skb1
, &tp
->out_of_order_queue
);
4398 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4399 TCP_SKB_CB(skb1
)->end_seq
);
4400 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFOMERGE
);
4405 if (tcp_is_sack(tp
))
4406 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
4409 tcp_grow_window(sk
, skb
);
4410 skb_set_owner_r(skb
, sk
);
4414 static int __must_check
tcp_queue_rcv(struct sock
*sk
, struct sk_buff
*skb
, int hdrlen
,
4418 struct sk_buff
*tail
= skb_peek_tail(&sk
->sk_receive_queue
);
4420 __skb_pull(skb
, hdrlen
);
4422 tcp_try_coalesce(sk
, tail
, skb
, fragstolen
)) ? 1 : 0;
4423 tcp_rcv_nxt_update(tcp_sk(sk
), TCP_SKB_CB(skb
)->end_seq
);
4425 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4426 skb_set_owner_r(skb
, sk
);
4431 int tcp_send_rcvq(struct sock
*sk
, struct msghdr
*msg
, size_t size
)
4433 struct sk_buff
*skb
;
4439 skb
= alloc_skb(size
, sk
->sk_allocation
);
4443 if (tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))
4446 if (memcpy_from_msg(skb_put(skb
, size
), msg
, size
))
4449 TCP_SKB_CB(skb
)->seq
= tcp_sk(sk
)->rcv_nxt
;
4450 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ size
;
4451 TCP_SKB_CB(skb
)->ack_seq
= tcp_sk(sk
)->snd_una
- 1;
4453 if (tcp_queue_rcv(sk
, skb
, 0, &fragstolen
)) {
4454 WARN_ON_ONCE(fragstolen
); /* should not happen */
4465 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
4467 struct tcp_sock
*tp
= tcp_sk(sk
);
4469 bool fragstolen
= false;
4471 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
)
4475 __skb_pull(skb
, tcp_hdr(skb
)->doff
* 4);
4477 tcp_ecn_accept_cwr(tp
, skb
);
4479 tp
->rx_opt
.dsack
= 0;
4481 /* Queue data for delivery to the user.
4482 * Packets in sequence go to the receive queue.
4483 * Out of sequence packets to the out_of_order_queue.
4485 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
4486 if (tcp_receive_window(tp
) == 0)
4489 /* Ok. In sequence. In window. */
4490 if (tp
->ucopy
.task
== current
&&
4491 tp
->copied_seq
== tp
->rcv_nxt
&& tp
->ucopy
.len
&&
4492 sock_owned_by_user(sk
) && !tp
->urg_data
) {
4493 int chunk
= min_t(unsigned int, skb
->len
,
4496 __set_current_state(TASK_RUNNING
);
4499 if (!skb_copy_datagram_msg(skb
, 0, tp
->ucopy
.msg
, chunk
)) {
4500 tp
->ucopy
.len
-= chunk
;
4501 tp
->copied_seq
+= chunk
;
4502 eaten
= (chunk
== skb
->len
);
4503 tcp_rcv_space_adjust(sk
);
4511 tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))
4514 eaten
= tcp_queue_rcv(sk
, skb
, 0, &fragstolen
);
4516 tcp_rcv_nxt_update(tp
, TCP_SKB_CB(skb
)->end_seq
);
4518 tcp_event_data_recv(sk
, skb
);
4519 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
4522 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4525 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4526 * gap in queue is filled.
4528 if (skb_queue_empty(&tp
->out_of_order_queue
))
4529 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
4532 if (tp
->rx_opt
.num_sacks
)
4533 tcp_sack_remove(tp
);
4535 tcp_fast_path_check(sk
);
4538 kfree_skb_partial(skb
, fragstolen
);
4539 if (!sock_flag(sk
, SOCK_DEAD
))
4540 sk
->sk_data_ready(sk
);
4544 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4545 /* A retransmit, 2nd most common case. Force an immediate ack. */
4546 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4547 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4550 tcp_enter_quickack_mode(sk
);
4551 inet_csk_schedule_ack(sk
);
4557 /* Out of window. F.e. zero window probe. */
4558 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
4561 tcp_enter_quickack_mode(sk
);
4563 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4564 /* Partial packet, seq < rcv_next < end_seq */
4565 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
4566 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4567 TCP_SKB_CB(skb
)->end_seq
);
4569 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
4571 /* If window is closed, drop tail of packet. But after
4572 * remembering D-SACK for its head made in previous line.
4574 if (!tcp_receive_window(tp
))
4579 tcp_data_queue_ofo(sk
, skb
);
4582 static struct sk_buff
*tcp_collapse_one(struct sock
*sk
, struct sk_buff
*skb
,
4583 struct sk_buff_head
*list
)
4585 struct sk_buff
*next
= NULL
;
4587 if (!skb_queue_is_last(list
, skb
))
4588 next
= skb_queue_next(list
, skb
);
4590 __skb_unlink(skb
, list
);
4592 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRCVCOLLAPSED
);
4597 /* Collapse contiguous sequence of skbs head..tail with
4598 * sequence numbers start..end.
4600 * If tail is NULL, this means until the end of the list.
4602 * Segments with FIN/SYN are not collapsed (only because this
4606 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
,
4607 struct sk_buff
*head
, struct sk_buff
*tail
,
4610 struct sk_buff
*skb
, *n
;
4613 /* First, check that queue is collapsible and find
4614 * the point where collapsing can be useful. */
4618 skb_queue_walk_from_safe(list
, skb
, n
) {
4621 /* No new bits? It is possible on ofo queue. */
4622 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4623 skb
= tcp_collapse_one(sk
, skb
, list
);
4629 /* The first skb to collapse is:
4631 * - bloated or contains data before "start" or
4632 * overlaps to the next one.
4634 if (!(TCP_SKB_CB(skb
)->tcp_flags
& (TCPHDR_SYN
| TCPHDR_FIN
)) &&
4635 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
4636 before(TCP_SKB_CB(skb
)->seq
, start
))) {
4637 end_of_skbs
= false;
4641 if (!skb_queue_is_last(list
, skb
)) {
4642 struct sk_buff
*next
= skb_queue_next(list
, skb
);
4644 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(next
)->seq
) {
4645 end_of_skbs
= false;
4650 /* Decided to skip this, advance start seq. */
4651 start
= TCP_SKB_CB(skb
)->end_seq
;
4654 (TCP_SKB_CB(skb
)->tcp_flags
& (TCPHDR_SYN
| TCPHDR_FIN
)))
4657 while (before(start
, end
)) {
4658 int copy
= min_t(int, SKB_MAX_ORDER(0, 0), end
- start
);
4659 struct sk_buff
*nskb
;
4661 nskb
= alloc_skb(copy
, GFP_ATOMIC
);
4665 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
4666 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
4667 __skb_queue_before(list
, skb
, nskb
);
4668 skb_set_owner_r(nskb
, sk
);
4670 /* Copy data, releasing collapsed skbs. */
4672 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
4673 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
4677 size
= min(copy
, size
);
4678 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
4680 TCP_SKB_CB(nskb
)->end_seq
+= size
;
4684 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4685 skb
= tcp_collapse_one(sk
, skb
, list
);
4688 (TCP_SKB_CB(skb
)->tcp_flags
& (TCPHDR_SYN
| TCPHDR_FIN
)))
4695 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4696 * and tcp_collapse() them until all the queue is collapsed.
4698 static void tcp_collapse_ofo_queue(struct sock
*sk
)
4700 struct tcp_sock
*tp
= tcp_sk(sk
);
4701 struct sk_buff
*skb
= skb_peek(&tp
->out_of_order_queue
);
4702 struct sk_buff
*head
;
4708 start
= TCP_SKB_CB(skb
)->seq
;
4709 end
= TCP_SKB_CB(skb
)->end_seq
;
4713 struct sk_buff
*next
= NULL
;
4715 if (!skb_queue_is_last(&tp
->out_of_order_queue
, skb
))
4716 next
= skb_queue_next(&tp
->out_of_order_queue
, skb
);
4719 /* Segment is terminated when we see gap or when
4720 * we are at the end of all the queue. */
4722 after(TCP_SKB_CB(skb
)->seq
, end
) ||
4723 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
4724 tcp_collapse(sk
, &tp
->out_of_order_queue
,
4725 head
, skb
, start
, end
);
4729 /* Start new segment */
4730 start
= TCP_SKB_CB(skb
)->seq
;
4731 end
= TCP_SKB_CB(skb
)->end_seq
;
4733 if (before(TCP_SKB_CB(skb
)->seq
, start
))
4734 start
= TCP_SKB_CB(skb
)->seq
;
4735 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
4736 end
= TCP_SKB_CB(skb
)->end_seq
;
4742 * Purge the out-of-order queue.
4743 * Return true if queue was pruned.
4745 static bool tcp_prune_ofo_queue(struct sock
*sk
)
4747 struct tcp_sock
*tp
= tcp_sk(sk
);
4750 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4751 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_OFOPRUNED
);
4752 __skb_queue_purge(&tp
->out_of_order_queue
);
4754 /* Reset SACK state. A conforming SACK implementation will
4755 * do the same at a timeout based retransmit. When a connection
4756 * is in a sad state like this, we care only about integrity
4757 * of the connection not performance.
4759 if (tp
->rx_opt
.sack_ok
)
4760 tcp_sack_reset(&tp
->rx_opt
);
4767 /* Reduce allocated memory if we can, trying to get
4768 * the socket within its memory limits again.
4770 * Return less than zero if we should start dropping frames
4771 * until the socket owning process reads some of the data
4772 * to stabilize the situation.
4774 static int tcp_prune_queue(struct sock
*sk
)
4776 struct tcp_sock
*tp
= tcp_sk(sk
);
4778 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
4780 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PRUNECALLED
);
4782 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
4783 tcp_clamp_window(sk
);
4784 else if (sk_under_memory_pressure(sk
))
4785 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
4787 tcp_collapse_ofo_queue(sk
);
4788 if (!skb_queue_empty(&sk
->sk_receive_queue
))
4789 tcp_collapse(sk
, &sk
->sk_receive_queue
,
4790 skb_peek(&sk
->sk_receive_queue
),
4792 tp
->copied_seq
, tp
->rcv_nxt
);
4795 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4798 /* Collapsing did not help, destructive actions follow.
4799 * This must not ever occur. */
4801 tcp_prune_ofo_queue(sk
);
4803 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4806 /* If we are really being abused, tell the caller to silently
4807 * drop receive data on the floor. It will get retransmitted
4808 * and hopefully then we'll have sufficient space.
4810 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_RCVPRUNED
);
4812 /* Massive buffer overcommit. */
4817 static bool tcp_should_expand_sndbuf(const struct sock
*sk
)
4819 const struct tcp_sock
*tp
= tcp_sk(sk
);
4821 /* If the user specified a specific send buffer setting, do
4824 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
4827 /* If we are under global TCP memory pressure, do not expand. */
4828 if (sk_under_memory_pressure(sk
))
4831 /* If we are under soft global TCP memory pressure, do not expand. */
4832 if (sk_memory_allocated(sk
) >= sk_prot_mem_limits(sk
, 0))
4835 /* If we filled the congestion window, do not expand. */
4836 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
4842 /* When incoming ACK allowed to free some skb from write_queue,
4843 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4844 * on the exit from tcp input handler.
4846 * PROBLEM: sndbuf expansion does not work well with largesend.
4848 static void tcp_new_space(struct sock
*sk
)
4850 struct tcp_sock
*tp
= tcp_sk(sk
);
4852 if (tcp_should_expand_sndbuf(sk
)) {
4853 tcp_sndbuf_expand(sk
);
4854 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4857 sk
->sk_write_space(sk
);
4860 static void tcp_check_space(struct sock
*sk
)
4862 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
4863 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
4864 /* pairs with tcp_poll() */
4865 smp_mb__after_atomic();
4866 if (sk
->sk_socket
&&
4867 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
4872 static inline void tcp_data_snd_check(struct sock
*sk
)
4874 tcp_push_pending_frames(sk
);
4875 tcp_check_space(sk
);
4879 * Check if sending an ack is needed.
4881 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
4883 struct tcp_sock
*tp
= tcp_sk(sk
);
4885 /* More than one full frame received... */
4886 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
&&
4887 /* ... and right edge of window advances far enough.
4888 * (tcp_recvmsg() will send ACK otherwise). Or...
4890 __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
4891 /* We ACK each frame or... */
4892 tcp_in_quickack_mode(sk
) ||
4893 /* We have out of order data. */
4894 (ofo_possible
&& skb_peek(&tp
->out_of_order_queue
))) {
4895 /* Then ack it now */
4898 /* Else, send delayed ack. */
4899 tcp_send_delayed_ack(sk
);
4903 static inline void tcp_ack_snd_check(struct sock
*sk
)
4905 if (!inet_csk_ack_scheduled(sk
)) {
4906 /* We sent a data segment already. */
4909 __tcp_ack_snd_check(sk
, 1);
4913 * This routine is only called when we have urgent data
4914 * signaled. Its the 'slow' part of tcp_urg. It could be
4915 * moved inline now as tcp_urg is only called from one
4916 * place. We handle URGent data wrong. We have to - as
4917 * BSD still doesn't use the correction from RFC961.
4918 * For 1003.1g we should support a new option TCP_STDURG to permit
4919 * either form (or just set the sysctl tcp_stdurg).
4922 static void tcp_check_urg(struct sock
*sk
, const struct tcphdr
*th
)
4924 struct tcp_sock
*tp
= tcp_sk(sk
);
4925 u32 ptr
= ntohs(th
->urg_ptr
);
4927 if (ptr
&& !sysctl_tcp_stdurg
)
4929 ptr
+= ntohl(th
->seq
);
4931 /* Ignore urgent data that we've already seen and read. */
4932 if (after(tp
->copied_seq
, ptr
))
4935 /* Do not replay urg ptr.
4937 * NOTE: interesting situation not covered by specs.
4938 * Misbehaving sender may send urg ptr, pointing to segment,
4939 * which we already have in ofo queue. We are not able to fetch
4940 * such data and will stay in TCP_URG_NOTYET until will be eaten
4941 * by recvmsg(). Seems, we are not obliged to handle such wicked
4942 * situations. But it is worth to think about possibility of some
4943 * DoSes using some hypothetical application level deadlock.
4945 if (before(ptr
, tp
->rcv_nxt
))
4948 /* Do we already have a newer (or duplicate) urgent pointer? */
4949 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
4952 /* Tell the world about our new urgent pointer. */
4955 /* We may be adding urgent data when the last byte read was
4956 * urgent. To do this requires some care. We cannot just ignore
4957 * tp->copied_seq since we would read the last urgent byte again
4958 * as data, nor can we alter copied_seq until this data arrives
4959 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4961 * NOTE. Double Dutch. Rendering to plain English: author of comment
4962 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4963 * and expect that both A and B disappear from stream. This is _wrong_.
4964 * Though this happens in BSD with high probability, this is occasional.
4965 * Any application relying on this is buggy. Note also, that fix "works"
4966 * only in this artificial test. Insert some normal data between A and B and we will
4967 * decline of BSD again. Verdict: it is better to remove to trap
4970 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
4971 !sock_flag(sk
, SOCK_URGINLINE
) && tp
->copied_seq
!= tp
->rcv_nxt
) {
4972 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
4974 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4975 __skb_unlink(skb
, &sk
->sk_receive_queue
);
4980 tp
->urg_data
= TCP_URG_NOTYET
;
4983 /* Disable header prediction. */
4987 /* This is the 'fast' part of urgent handling. */
4988 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, const struct tcphdr
*th
)
4990 struct tcp_sock
*tp
= tcp_sk(sk
);
4992 /* Check if we get a new urgent pointer - normally not. */
4994 tcp_check_urg(sk
, th
);
4996 /* Do we wait for any urgent data? - normally not... */
4997 if (tp
->urg_data
== TCP_URG_NOTYET
) {
4998 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
5001 /* Is the urgent pointer pointing into this packet? */
5002 if (ptr
< skb
->len
) {
5004 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
5006 tp
->urg_data
= TCP_URG_VALID
| tmp
;
5007 if (!sock_flag(sk
, SOCK_DEAD
))
5008 sk
->sk_data_ready(sk
);
5013 static int tcp_copy_to_iovec(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
5015 struct tcp_sock
*tp
= tcp_sk(sk
);
5016 int chunk
= skb
->len
- hlen
;
5020 if (skb_csum_unnecessary(skb
))
5021 err
= skb_copy_datagram_msg(skb
, hlen
, tp
->ucopy
.msg
, chunk
);
5023 err
= skb_copy_and_csum_datagram_msg(skb
, hlen
, tp
->ucopy
.msg
);
5026 tp
->ucopy
.len
-= chunk
;
5027 tp
->copied_seq
+= chunk
;
5028 tcp_rcv_space_adjust(sk
);
5035 static __sum16
__tcp_checksum_complete_user(struct sock
*sk
,
5036 struct sk_buff
*skb
)
5040 if (sock_owned_by_user(sk
)) {
5042 result
= __tcp_checksum_complete(skb
);
5045 result
= __tcp_checksum_complete(skb
);
5050 static inline bool tcp_checksum_complete_user(struct sock
*sk
,
5051 struct sk_buff
*skb
)
5053 return !skb_csum_unnecessary(skb
) &&
5054 __tcp_checksum_complete_user(sk
, skb
);
5057 /* Does PAWS and seqno based validation of an incoming segment, flags will
5058 * play significant role here.
5060 static bool tcp_validate_incoming(struct sock
*sk
, struct sk_buff
*skb
,
5061 const struct tcphdr
*th
, int syn_inerr
)
5063 struct tcp_sock
*tp
= tcp_sk(sk
);
5065 /* RFC1323: H1. Apply PAWS check first. */
5066 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
5067 tcp_paws_discard(sk
, skb
)) {
5069 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSESTABREJECTED
);
5070 if (!tcp_oow_rate_limited(sock_net(sk
), skb
,
5071 LINUX_MIB_TCPACKSKIPPEDPAWS
,
5072 &tp
->last_oow_ack_time
))
5073 tcp_send_dupack(sk
, skb
);
5076 /* Reset is accepted even if it did not pass PAWS. */
5079 /* Step 1: check sequence number */
5080 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5081 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5082 * (RST) segments are validated by checking their SEQ-fields."
5083 * And page 69: "If an incoming segment is not acceptable,
5084 * an acknowledgment should be sent in reply (unless the RST
5085 * bit is set, if so drop the segment and return)".
5090 if (!tcp_oow_rate_limited(sock_net(sk
), skb
,
5091 LINUX_MIB_TCPACKSKIPPEDSEQ
,
5092 &tp
->last_oow_ack_time
))
5093 tcp_send_dupack(sk
, skb
);
5098 /* Step 2: check RST bit */
5101 * If sequence number exactly matches RCV.NXT, then
5102 * RESET the connection
5104 * Send a challenge ACK
5106 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
)
5109 tcp_send_challenge_ack(sk
, skb
);
5113 /* step 3: check security and precedence [ignored] */
5115 /* step 4: Check for a SYN
5116 * RFC 5961 4.2 : Send a challenge ack
5121 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5122 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSYNCHALLENGE
);
5123 tcp_send_challenge_ack(sk
, skb
);
5135 * TCP receive function for the ESTABLISHED state.
5137 * It is split into a fast path and a slow path. The fast path is
5139 * - A zero window was announced from us - zero window probing
5140 * is only handled properly in the slow path.
5141 * - Out of order segments arrived.
5142 * - Urgent data is expected.
5143 * - There is no buffer space left
5144 * - Unexpected TCP flags/window values/header lengths are received
5145 * (detected by checking the TCP header against pred_flags)
5146 * - Data is sent in both directions. Fast path only supports pure senders
5147 * or pure receivers (this means either the sequence number or the ack
5148 * value must stay constant)
5149 * - Unexpected TCP option.
5151 * When these conditions are not satisfied it drops into a standard
5152 * receive procedure patterned after RFC793 to handle all cases.
5153 * The first three cases are guaranteed by proper pred_flags setting,
5154 * the rest is checked inline. Fast processing is turned on in
5155 * tcp_data_queue when everything is OK.
5157 void tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
5158 const struct tcphdr
*th
, unsigned int len
)
5160 struct tcp_sock
*tp
= tcp_sk(sk
);
5162 if (unlikely(!sk
->sk_rx_dst
))
5163 inet_csk(sk
)->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
5165 * Header prediction.
5166 * The code loosely follows the one in the famous
5167 * "30 instruction TCP receive" Van Jacobson mail.
5169 * Van's trick is to deposit buffers into socket queue
5170 * on a device interrupt, to call tcp_recv function
5171 * on the receive process context and checksum and copy
5172 * the buffer to user space. smart...
5174 * Our current scheme is not silly either but we take the
5175 * extra cost of the net_bh soft interrupt processing...
5176 * We do checksum and copy also but from device to kernel.
5179 tp
->rx_opt
.saw_tstamp
= 0;
5181 /* pred_flags is 0xS?10 << 16 + snd_wnd
5182 * if header_prediction is to be made
5183 * 'S' will always be tp->tcp_header_len >> 2
5184 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5185 * turn it off (when there are holes in the receive
5186 * space for instance)
5187 * PSH flag is ignored.
5190 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
5191 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
&&
5192 !after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
)) {
5193 int tcp_header_len
= tp
->tcp_header_len
;
5195 /* Timestamp header prediction: tcp_header_len
5196 * is automatically equal to th->doff*4 due to pred_flags
5200 /* Check timestamp */
5201 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
5202 /* No? Slow path! */
5203 if (!tcp_parse_aligned_timestamp(tp
, th
))
5206 /* If PAWS failed, check it more carefully in slow path */
5207 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
5210 /* DO NOT update ts_recent here, if checksum fails
5211 * and timestamp was corrupted part, it will result
5212 * in a hung connection since we will drop all
5213 * future packets due to the PAWS test.
5217 if (len
<= tcp_header_len
) {
5218 /* Bulk data transfer: sender */
5219 if (len
== tcp_header_len
) {
5220 /* Predicted packet is in window by definition.
5221 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5222 * Hence, check seq<=rcv_wup reduces to:
5224 if (tcp_header_len
==
5225 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5226 tp
->rcv_nxt
== tp
->rcv_wup
)
5227 tcp_store_ts_recent(tp
);
5229 /* We know that such packets are checksummed
5232 tcp_ack(sk
, skb
, 0);
5234 tcp_data_snd_check(sk
);
5236 } else { /* Header too small */
5237 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5242 bool fragstolen
= false;
5244 if (tp
->ucopy
.task
== current
&&
5245 tp
->copied_seq
== tp
->rcv_nxt
&&
5246 len
- tcp_header_len
<= tp
->ucopy
.len
&&
5247 sock_owned_by_user(sk
)) {
5248 __set_current_state(TASK_RUNNING
);
5250 if (!tcp_copy_to_iovec(sk
, skb
, tcp_header_len
)) {
5251 /* Predicted packet is in window by definition.
5252 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5253 * Hence, check seq<=rcv_wup reduces to:
5255 if (tcp_header_len
==
5256 (sizeof(struct tcphdr
) +
5257 TCPOLEN_TSTAMP_ALIGNED
) &&
5258 tp
->rcv_nxt
== tp
->rcv_wup
)
5259 tcp_store_ts_recent(tp
);
5261 tcp_rcv_rtt_measure_ts(sk
, skb
);
5263 __skb_pull(skb
, tcp_header_len
);
5264 tcp_rcv_nxt_update(tp
, TCP_SKB_CB(skb
)->end_seq
);
5265 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITSTOUSER
);
5270 if (tcp_checksum_complete_user(sk
, skb
))
5273 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
5276 /* Predicted packet is in window by definition.
5277 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5278 * Hence, check seq<=rcv_wup reduces to:
5280 if (tcp_header_len
==
5281 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5282 tp
->rcv_nxt
== tp
->rcv_wup
)
5283 tcp_store_ts_recent(tp
);
5285 tcp_rcv_rtt_measure_ts(sk
, skb
);
5287 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITS
);
5289 /* Bulk data transfer: receiver */
5290 eaten
= tcp_queue_rcv(sk
, skb
, tcp_header_len
,
5294 tcp_event_data_recv(sk
, skb
);
5296 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
5297 /* Well, only one small jumplet in fast path... */
5298 tcp_ack(sk
, skb
, FLAG_DATA
);
5299 tcp_data_snd_check(sk
);
5300 if (!inet_csk_ack_scheduled(sk
))
5304 __tcp_ack_snd_check(sk
, 0);
5307 kfree_skb_partial(skb
, fragstolen
);
5308 sk
->sk_data_ready(sk
);
5314 if (len
< (th
->doff
<< 2) || tcp_checksum_complete_user(sk
, skb
))
5317 if (!th
->ack
&& !th
->rst
&& !th
->syn
)
5321 * Standard slow path.
5324 if (!tcp_validate_incoming(sk
, skb
, th
, 1))
5328 if (tcp_ack(sk
, skb
, FLAG_SLOWPATH
| FLAG_UPDATE_TS_RECENT
) < 0)
5331 tcp_rcv_rtt_measure_ts(sk
, skb
);
5333 /* Process urgent data. */
5334 tcp_urg(sk
, skb
, th
);
5336 /* step 7: process the segment text */
5337 tcp_data_queue(sk
, skb
);
5339 tcp_data_snd_check(sk
);
5340 tcp_ack_snd_check(sk
);
5344 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_CSUMERRORS
);
5345 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5350 EXPORT_SYMBOL(tcp_rcv_established
);
5352 void tcp_finish_connect(struct sock
*sk
, struct sk_buff
*skb
)
5354 struct tcp_sock
*tp
= tcp_sk(sk
);
5355 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5357 tcp_set_state(sk
, TCP_ESTABLISHED
);
5360 icsk
->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
5361 security_inet_conn_established(sk
, skb
);
5364 /* Make sure socket is routed, for correct metrics. */
5365 icsk
->icsk_af_ops
->rebuild_header(sk
);
5367 tcp_init_metrics(sk
);
5369 tcp_init_congestion_control(sk
);
5371 /* Prevent spurious tcp_cwnd_restart() on first data
5374 tp
->lsndtime
= tcp_time_stamp
;
5376 tcp_init_buffer_space(sk
);
5378 if (sock_flag(sk
, SOCK_KEEPOPEN
))
5379 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
5381 if (!tp
->rx_opt
.snd_wscale
)
5382 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
5386 if (!sock_flag(sk
, SOCK_DEAD
)) {
5387 sk
->sk_state_change(sk
);
5388 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5392 static bool tcp_rcv_fastopen_synack(struct sock
*sk
, struct sk_buff
*synack
,
5393 struct tcp_fastopen_cookie
*cookie
)
5395 struct tcp_sock
*tp
= tcp_sk(sk
);
5396 struct sk_buff
*data
= tp
->syn_data
? tcp_write_queue_head(sk
) : NULL
;
5397 u16 mss
= tp
->rx_opt
.mss_clamp
, try_exp
= 0;
5398 bool syn_drop
= false;
5400 if (mss
== tp
->rx_opt
.user_mss
) {
5401 struct tcp_options_received opt
;
5403 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5404 tcp_clear_options(&opt
);
5405 opt
.user_mss
= opt
.mss_clamp
= 0;
5406 tcp_parse_options(synack
, &opt
, 0, NULL
);
5407 mss
= opt
.mss_clamp
;
5410 if (!tp
->syn_fastopen
) {
5411 /* Ignore an unsolicited cookie */
5413 } else if (tp
->total_retrans
) {
5414 /* SYN timed out and the SYN-ACK neither has a cookie nor
5415 * acknowledges data. Presumably the remote received only
5416 * the retransmitted (regular) SYNs: either the original
5417 * SYN-data or the corresponding SYN-ACK was dropped.
5419 syn_drop
= (cookie
->len
< 0 && data
);
5420 } else if (cookie
->len
< 0 && !tp
->syn_data
) {
5421 /* We requested a cookie but didn't get it. If we did not use
5422 * the (old) exp opt format then try so next time (try_exp=1).
5423 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5425 try_exp
= tp
->syn_fastopen_exp
? 2 : 1;
5428 tcp_fastopen_cache_set(sk
, mss
, cookie
, syn_drop
, try_exp
);
5430 if (data
) { /* Retransmit unacked data in SYN */
5431 tcp_for_write_queue_from(data
, sk
) {
5432 if (data
== tcp_send_head(sk
) ||
5433 __tcp_retransmit_skb(sk
, data
))
5437 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPFASTOPENACTIVEFAIL
);
5440 tp
->syn_data_acked
= tp
->syn_data
;
5441 if (tp
->syn_data_acked
)
5442 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPFASTOPENACTIVE
);
5446 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5447 const struct tcphdr
*th
, unsigned int len
)
5449 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5450 struct tcp_sock
*tp
= tcp_sk(sk
);
5451 struct tcp_fastopen_cookie foc
= { .len
= -1 };
5452 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
5454 tcp_parse_options(skb
, &tp
->rx_opt
, 0, &foc
);
5455 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
5456 tp
->rx_opt
.rcv_tsecr
-= tp
->tsoffset
;
5460 * "If the state is SYN-SENT then
5461 * first check the ACK bit
5462 * If the ACK bit is set
5463 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5464 * a reset (unless the RST bit is set, if so drop
5465 * the segment and return)"
5467 if (!after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_una
) ||
5468 after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
))
5469 goto reset_and_undo
;
5471 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
5472 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
5474 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSACTIVEREJECTED
);
5475 goto reset_and_undo
;
5478 /* Now ACK is acceptable.
5480 * "If the RST bit is set
5481 * If the ACK was acceptable then signal the user "error:
5482 * connection reset", drop the segment, enter CLOSED state,
5483 * delete TCB, and return."
5492 * "fifth, if neither of the SYN or RST bits is set then
5493 * drop the segment and return."
5499 goto discard_and_undo
;
5502 * "If the SYN bit is on ...
5503 * are acceptable then ...
5504 * (our SYN has been ACKed), change the connection
5505 * state to ESTABLISHED..."
5508 tcp_ecn_rcv_synack(tp
, th
);
5510 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5511 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
5513 /* Ok.. it's good. Set up sequence numbers and
5514 * move to established.
5516 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5517 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5519 /* RFC1323: The window in SYN & SYN/ACK segments is
5522 tp
->snd_wnd
= ntohs(th
->window
);
5524 if (!tp
->rx_opt
.wscale_ok
) {
5525 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
5526 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
5529 if (tp
->rx_opt
.saw_tstamp
) {
5530 tp
->rx_opt
.tstamp_ok
= 1;
5531 tp
->tcp_header_len
=
5532 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5533 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5534 tcp_store_ts_recent(tp
);
5536 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5539 if (tcp_is_sack(tp
) && sysctl_tcp_fack
)
5540 tcp_enable_fack(tp
);
5543 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5544 tcp_initialize_rcv_mss(sk
);
5546 /* Remember, tcp_poll() does not lock socket!
5547 * Change state from SYN-SENT only after copied_seq
5548 * is initialized. */
5549 tp
->copied_seq
= tp
->rcv_nxt
;
5553 tcp_finish_connect(sk
, skb
);
5555 if ((tp
->syn_fastopen
|| tp
->syn_data
) &&
5556 tcp_rcv_fastopen_synack(sk
, skb
, &foc
))
5559 if (sk
->sk_write_pending
||
5560 icsk
->icsk_accept_queue
.rskq_defer_accept
||
5561 icsk
->icsk_ack
.pingpong
) {
5562 /* Save one ACK. Data will be ready after
5563 * several ticks, if write_pending is set.
5565 * It may be deleted, but with this feature tcpdumps
5566 * look so _wonderfully_ clever, that I was not able
5567 * to stand against the temptation 8) --ANK
5569 inet_csk_schedule_ack(sk
);
5570 icsk
->icsk_ack
.lrcvtime
= tcp_time_stamp
;
5571 tcp_enter_quickack_mode(sk
);
5572 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
5573 TCP_DELACK_MAX
, TCP_RTO_MAX
);
5584 /* No ACK in the segment */
5588 * "If the RST bit is set
5590 * Otherwise (no ACK) drop the segment and return."
5593 goto discard_and_undo
;
5597 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&&
5598 tcp_paws_reject(&tp
->rx_opt
, 0))
5599 goto discard_and_undo
;
5602 /* We see SYN without ACK. It is attempt of
5603 * simultaneous connect with crossed SYNs.
5604 * Particularly, it can be connect to self.
5606 tcp_set_state(sk
, TCP_SYN_RECV
);
5608 if (tp
->rx_opt
.saw_tstamp
) {
5609 tp
->rx_opt
.tstamp_ok
= 1;
5610 tcp_store_ts_recent(tp
);
5611 tp
->tcp_header_len
=
5612 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5614 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5617 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5618 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5620 /* RFC1323: The window in SYN & SYN/ACK segments is
5623 tp
->snd_wnd
= ntohs(th
->window
);
5624 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
5625 tp
->max_window
= tp
->snd_wnd
;
5627 tcp_ecn_rcv_syn(tp
, th
);
5630 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5631 tcp_initialize_rcv_mss(sk
);
5633 tcp_send_synack(sk
);
5635 /* Note, we could accept data and URG from this segment.
5636 * There are no obstacles to make this (except that we must
5637 * either change tcp_recvmsg() to prevent it from returning data
5638 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5640 * However, if we ignore data in ACKless segments sometimes,
5641 * we have no reasons to accept it sometimes.
5642 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5643 * is not flawless. So, discard packet for sanity.
5644 * Uncomment this return to process the data.
5651 /* "fifth, if neither of the SYN or RST bits is set then
5652 * drop the segment and return."
5656 tcp_clear_options(&tp
->rx_opt
);
5657 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5661 tcp_clear_options(&tp
->rx_opt
);
5662 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5667 * This function implements the receiving procedure of RFC 793 for
5668 * all states except ESTABLISHED and TIME_WAIT.
5669 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5670 * address independent.
5673 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5674 const struct tcphdr
*th
, unsigned int len
)
5676 struct tcp_sock
*tp
= tcp_sk(sk
);
5677 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5678 struct request_sock
*req
;
5683 tp
->rx_opt
.saw_tstamp
= 0;
5685 switch (sk
->sk_state
) {
5699 if (icsk
->icsk_af_ops
->conn_request(sk
, skb
) < 0)
5702 /* Now we have several options: In theory there is
5703 * nothing else in the frame. KA9Q has an option to
5704 * send data with the syn, BSD accepts data with the
5705 * syn up to the [to be] advertised window and
5706 * Solaris 2.1 gives you a protocol error. For now
5707 * we just ignore it, that fits the spec precisely
5708 * and avoids incompatibilities. It would be nice in
5709 * future to drop through and process the data.
5711 * Now that TTCP is starting to be used we ought to
5713 * But, this leaves one open to an easy denial of
5714 * service attack, and SYN cookies can't defend
5715 * against this problem. So, we drop the data
5716 * in the interest of security over speed unless
5717 * it's still in use.
5725 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
, len
);
5729 /* Do step6 onward by hand. */
5730 tcp_urg(sk
, skb
, th
);
5732 tcp_data_snd_check(sk
);
5736 req
= tp
->fastopen_rsk
;
5738 WARN_ON_ONCE(sk
->sk_state
!= TCP_SYN_RECV
&&
5739 sk
->sk_state
!= TCP_FIN_WAIT1
);
5741 if (!tcp_check_req(sk
, skb
, req
, true))
5745 if (!th
->ack
&& !th
->rst
&& !th
->syn
)
5748 if (!tcp_validate_incoming(sk
, skb
, th
, 0))
5751 /* step 5: check the ACK field */
5752 acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
|
5753 FLAG_UPDATE_TS_RECENT
) > 0;
5755 switch (sk
->sk_state
) {
5760 /* Once we leave TCP_SYN_RECV, we no longer need req
5764 synack_stamp
= tcp_rsk(req
)->snt_synack
;
5765 tp
->total_retrans
= req
->num_retrans
;
5766 reqsk_fastopen_remove(sk
, req
, false);
5768 synack_stamp
= tp
->lsndtime
;
5769 /* Make sure socket is routed, for correct metrics. */
5770 icsk
->icsk_af_ops
->rebuild_header(sk
);
5771 tcp_init_congestion_control(sk
);
5774 tp
->copied_seq
= tp
->rcv_nxt
;
5775 tcp_init_buffer_space(sk
);
5778 tcp_set_state(sk
, TCP_ESTABLISHED
);
5779 sk
->sk_state_change(sk
);
5781 /* Note, that this wakeup is only for marginal crossed SYN case.
5782 * Passively open sockets are not waked up, because
5783 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5786 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5788 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
5789 tp
->snd_wnd
= ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
;
5790 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5791 tcp_synack_rtt_meas(sk
, synack_stamp
);
5793 if (tp
->rx_opt
.tstamp_ok
)
5794 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5797 /* Re-arm the timer because data may have been sent out.
5798 * This is similar to the regular data transmission case
5799 * when new data has just been ack'ed.
5801 * (TFO) - we could try to be more aggressive and
5802 * retransmitting any data sooner based on when they
5807 tcp_init_metrics(sk
);
5809 tcp_update_pacing_rate(sk
);
5811 /* Prevent spurious tcp_cwnd_restart() on first data packet */
5812 tp
->lsndtime
= tcp_time_stamp
;
5814 tcp_initialize_rcv_mss(sk
);
5815 tcp_fast_path_on(tp
);
5818 case TCP_FIN_WAIT1
: {
5819 struct dst_entry
*dst
;
5822 /* If we enter the TCP_FIN_WAIT1 state and we are a
5823 * Fast Open socket and this is the first acceptable
5824 * ACK we have received, this would have acknowledged
5825 * our SYNACK so stop the SYNACK timer.
5828 /* Return RST if ack_seq is invalid.
5829 * Note that RFC793 only says to generate a
5830 * DUPACK for it but for TCP Fast Open it seems
5831 * better to treat this case like TCP_SYN_RECV
5836 /* We no longer need the request sock. */
5837 reqsk_fastopen_remove(sk
, req
, false);
5840 if (tp
->snd_una
!= tp
->write_seq
)
5843 tcp_set_state(sk
, TCP_FIN_WAIT2
);
5844 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
5846 dst
= __sk_dst_get(sk
);
5850 if (!sock_flag(sk
, SOCK_DEAD
)) {
5851 /* Wake up lingering close() */
5852 sk
->sk_state_change(sk
);
5856 if (tp
->linger2
< 0 ||
5857 (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5858 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
))) {
5860 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
5864 tmo
= tcp_fin_time(sk
);
5865 if (tmo
> TCP_TIMEWAIT_LEN
) {
5866 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
5867 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
5868 /* Bad case. We could lose such FIN otherwise.
5869 * It is not a big problem, but it looks confusing
5870 * and not so rare event. We still can lose it now,
5871 * if it spins in bh_lock_sock(), but it is really
5874 inet_csk_reset_keepalive_timer(sk
, tmo
);
5876 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
5883 if (tp
->snd_una
== tp
->write_seq
) {
5884 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
5890 if (tp
->snd_una
== tp
->write_seq
) {
5891 tcp_update_metrics(sk
);
5898 /* step 6: check the URG bit */
5899 tcp_urg(sk
, skb
, th
);
5901 /* step 7: process the segment text */
5902 switch (sk
->sk_state
) {
5903 case TCP_CLOSE_WAIT
:
5906 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
5910 /* RFC 793 says to queue data in these states,
5911 * RFC 1122 says we MUST send a reset.
5912 * BSD 4.4 also does reset.
5914 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
5915 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5916 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
5917 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
5923 case TCP_ESTABLISHED
:
5924 tcp_data_queue(sk
, skb
);
5929 /* tcp_data could move socket to TIME-WAIT */
5930 if (sk
->sk_state
!= TCP_CLOSE
) {
5931 tcp_data_snd_check(sk
);
5932 tcp_ack_snd_check(sk
);
5941 EXPORT_SYMBOL(tcp_rcv_state_process
);
5943 static inline void pr_drop_req(struct request_sock
*req
, __u16 port
, int family
)
5945 struct inet_request_sock
*ireq
= inet_rsk(req
);
5947 if (family
== AF_INET
)
5948 net_dbg_ratelimited("drop open request from %pI4/%u\n",
5949 &ireq
->ir_rmt_addr
, port
);
5950 #if IS_ENABLED(CONFIG_IPV6)
5951 else if (family
== AF_INET6
)
5952 net_dbg_ratelimited("drop open request from %pI6/%u\n",
5953 &ireq
->ir_v6_rmt_addr
, port
);
5957 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
5959 * If we receive a SYN packet with these bits set, it means a
5960 * network is playing bad games with TOS bits. In order to
5961 * avoid possible false congestion notifications, we disable
5962 * TCP ECN negotiation.
5964 * Exception: tcp_ca wants ECN. This is required for DCTCP
5965 * congestion control: Linux DCTCP asserts ECT on all packets,
5966 * including SYN, which is most optimal solution; however,
5967 * others, such as FreeBSD do not.
5969 static void tcp_ecn_create_request(struct request_sock
*req
,
5970 const struct sk_buff
*skb
,
5971 const struct sock
*listen_sk
,
5972 const struct dst_entry
*dst
)
5974 const struct tcphdr
*th
= tcp_hdr(skb
);
5975 const struct net
*net
= sock_net(listen_sk
);
5976 bool th_ecn
= th
->ece
&& th
->cwr
;
5982 ect
= !INET_ECN_is_not_ect(TCP_SKB_CB(skb
)->ip_dsfield
);
5983 ecn_ok
= net
->ipv4
.sysctl_tcp_ecn
|| dst_feature(dst
, RTAX_FEATURE_ECN
);
5985 if ((!ect
&& ecn_ok
) || tcp_ca_needs_ecn(listen_sk
))
5986 inet_rsk(req
)->ecn_ok
= 1;
5989 static void tcp_openreq_init(struct request_sock
*req
,
5990 const struct tcp_options_received
*rx_opt
,
5991 struct sk_buff
*skb
, const struct sock
*sk
)
5993 struct inet_request_sock
*ireq
= inet_rsk(req
);
5995 req
->rcv_wnd
= 0; /* So that tcp_send_synack() knows! */
5997 tcp_rsk(req
)->rcv_isn
= TCP_SKB_CB(skb
)->seq
;
5998 tcp_rsk(req
)->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5999 tcp_rsk(req
)->snt_synack
= tcp_time_stamp
;
6000 tcp_rsk(req
)->last_oow_ack_time
= 0;
6001 req
->mss
= rx_opt
->mss_clamp
;
6002 req
->ts_recent
= rx_opt
->saw_tstamp
? rx_opt
->rcv_tsval
: 0;
6003 ireq
->tstamp_ok
= rx_opt
->tstamp_ok
;
6004 ireq
->sack_ok
= rx_opt
->sack_ok
;
6005 ireq
->snd_wscale
= rx_opt
->snd_wscale
;
6006 ireq
->wscale_ok
= rx_opt
->wscale_ok
;
6009 ireq
->ir_rmt_port
= tcp_hdr(skb
)->source
;
6010 ireq
->ir_num
= ntohs(tcp_hdr(skb
)->dest
);
6011 ireq
->ir_mark
= inet_request_mark(sk
, skb
);
6014 struct request_sock
*inet_reqsk_alloc(const struct request_sock_ops
*ops
,
6015 struct sock
*sk_listener
)
6017 struct request_sock
*req
= reqsk_alloc(ops
, sk_listener
);
6020 struct inet_request_sock
*ireq
= inet_rsk(req
);
6022 kmemcheck_annotate_bitfield(ireq
, flags
);
6024 atomic64_set(&ireq
->ir_cookie
, 0);
6025 ireq
->ireq_state
= TCP_NEW_SYN_RECV
;
6026 write_pnet(&ireq
->ireq_net
, sock_net(sk_listener
));
6027 ireq
->ireq_family
= sk_listener
->sk_family
;
6032 EXPORT_SYMBOL(inet_reqsk_alloc
);
6035 * Return true if a syncookie should be sent
6037 static bool tcp_syn_flood_action(struct sock
*sk
,
6038 const struct sk_buff
*skb
,
6041 const char *msg
= "Dropping request";
6042 bool want_cookie
= false;
6043 struct listen_sock
*lopt
;
6045 #ifdef CONFIG_SYN_COOKIES
6046 if (sysctl_tcp_syncookies
) {
6047 msg
= "Sending cookies";
6049 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPREQQFULLDOCOOKIES
);
6052 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPREQQFULLDROP
);
6054 lopt
= inet_csk(sk
)->icsk_accept_queue
.listen_opt
;
6055 if (!lopt
->synflood_warned
&& sysctl_tcp_syncookies
!= 2) {
6056 lopt
->synflood_warned
= 1;
6057 pr_info("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6058 proto
, ntohs(tcp_hdr(skb
)->dest
), msg
);
6063 int tcp_conn_request(struct request_sock_ops
*rsk_ops
,
6064 const struct tcp_request_sock_ops
*af_ops
,
6065 struct sock
*sk
, struct sk_buff
*skb
)
6067 struct tcp_options_received tmp_opt
;
6068 struct request_sock
*req
;
6069 struct tcp_sock
*tp
= tcp_sk(sk
);
6070 struct dst_entry
*dst
= NULL
;
6071 __u32 isn
= TCP_SKB_CB(skb
)->tcp_tw_isn
;
6072 bool want_cookie
= false, fastopen
;
6074 struct tcp_fastopen_cookie foc
= { .len
= -1 };
6078 /* TW buckets are converted to open requests without
6079 * limitations, they conserve resources and peer is
6080 * evidently real one.
6082 if ((sysctl_tcp_syncookies
== 2 ||
6083 inet_csk_reqsk_queue_is_full(sk
)) && !isn
) {
6084 want_cookie
= tcp_syn_flood_action(sk
, skb
, rsk_ops
->slab_name
);
6090 /* Accept backlog is full. If we have already queued enough
6091 * of warm entries in syn queue, drop request. It is better than
6092 * clogging syn queue with openreqs with exponentially increasing
6095 if (sk_acceptq_is_full(sk
) && inet_csk_reqsk_queue_young(sk
) > 1) {
6096 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_LISTENOVERFLOWS
);
6100 req
= inet_reqsk_alloc(rsk_ops
, sk
);
6104 tcp_rsk(req
)->af_specific
= af_ops
;
6106 tcp_clear_options(&tmp_opt
);
6107 tmp_opt
.mss_clamp
= af_ops
->mss_clamp
;
6108 tmp_opt
.user_mss
= tp
->rx_opt
.user_mss
;
6109 tcp_parse_options(skb
, &tmp_opt
, 0, want_cookie
? NULL
: &foc
);
6111 if (want_cookie
&& !tmp_opt
.saw_tstamp
)
6112 tcp_clear_options(&tmp_opt
);
6114 tmp_opt
.tstamp_ok
= tmp_opt
.saw_tstamp
;
6115 tcp_openreq_init(req
, &tmp_opt
, skb
, sk
);
6117 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6118 inet_rsk(req
)->ir_iif
= sk
->sk_bound_dev_if
;
6120 af_ops
->init_req(req
, sk
, skb
);
6122 if (security_inet_conn_request(sk
, skb
, req
))
6125 if (!want_cookie
&& !isn
) {
6126 /* VJ's idea. We save last timestamp seen
6127 * from the destination in peer table, when entering
6128 * state TIME-WAIT, and check against it before
6129 * accepting new connection request.
6131 * If "isn" is not zero, this request hit alive
6132 * timewait bucket, so that all the necessary checks
6133 * are made in the function processing timewait state.
6135 if (tcp_death_row
.sysctl_tw_recycle
) {
6138 dst
= af_ops
->route_req(sk
, &fl
, req
, &strict
);
6140 if (dst
&& strict
&&
6141 !tcp_peer_is_proven(req
, dst
, true,
6142 tmp_opt
.saw_tstamp
)) {
6143 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSPASSIVEREJECTED
);
6144 goto drop_and_release
;
6147 /* Kill the following clause, if you dislike this way. */
6148 else if (!sysctl_tcp_syncookies
&&
6149 (sysctl_max_syn_backlog
- inet_csk_reqsk_queue_len(sk
) <
6150 (sysctl_max_syn_backlog
>> 2)) &&
6151 !tcp_peer_is_proven(req
, dst
, false,
6152 tmp_opt
.saw_tstamp
)) {
6153 /* Without syncookies last quarter of
6154 * backlog is filled with destinations,
6155 * proven to be alive.
6156 * It means that we continue to communicate
6157 * to destinations, already remembered
6158 * to the moment of synflood.
6160 pr_drop_req(req
, ntohs(tcp_hdr(skb
)->source
),
6162 goto drop_and_release
;
6165 isn
= af_ops
->init_seq(skb
);
6168 dst
= af_ops
->route_req(sk
, &fl
, req
, NULL
);
6173 tcp_ecn_create_request(req
, skb
, sk
, dst
);
6176 isn
= cookie_init_sequence(af_ops
, sk
, skb
, &req
->mss
);
6177 req
->cookie_ts
= tmp_opt
.tstamp_ok
;
6178 if (!tmp_opt
.tstamp_ok
)
6179 inet_rsk(req
)->ecn_ok
= 0;
6182 tcp_rsk(req
)->snt_isn
= isn
;
6183 tcp_openreq_init_rwin(req
, sk
, dst
);
6184 fastopen
= !want_cookie
&&
6185 tcp_try_fastopen(sk
, skb
, req
, &foc
, dst
);
6186 err
= af_ops
->send_synack(sk
, dst
, &fl
, req
,
6187 skb_get_queue_mapping(skb
), &foc
);
6189 if (err
|| want_cookie
)
6192 tcp_rsk(req
)->tfo_listener
= false;
6193 af_ops
->queue_hash_add(sk
, req
, TCP_TIMEOUT_INIT
);
6203 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_LISTENDROPS
);
6206 EXPORT_SYMBOL(tcp_conn_request
);