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;
98 int sysctl_tcp_min_rtt_wlen __read_mostly
= 300;
100 int sysctl_tcp_thin_dupack __read_mostly
;
102 int sysctl_tcp_moderate_rcvbuf __read_mostly
= 1;
103 int sysctl_tcp_early_retrans __read_mostly
= 3;
104 int sysctl_tcp_invalid_ratelimit __read_mostly
= HZ
/2;
106 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
107 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
108 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
109 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
110 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
111 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
112 #define FLAG_ECE 0x40 /* ECE in this ACK */
113 #define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */
114 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
115 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
116 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
117 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
118 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
119 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
121 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
122 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
123 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
124 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
126 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
127 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
129 #define REXMIT_NONE 0 /* no loss recovery to do */
130 #define REXMIT_LOST 1 /* retransmit packets marked lost */
131 #define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */
133 /* Adapt the MSS value used to make delayed ack decision to the
136 static void tcp_measure_rcv_mss(struct sock
*sk
, const struct sk_buff
*skb
)
138 struct inet_connection_sock
*icsk
= inet_csk(sk
);
139 const unsigned int lss
= icsk
->icsk_ack
.last_seg_size
;
142 icsk
->icsk_ack
.last_seg_size
= 0;
144 /* skb->len may jitter because of SACKs, even if peer
145 * sends good full-sized frames.
147 len
= skb_shinfo(skb
)->gso_size
? : skb
->len
;
148 if (len
>= icsk
->icsk_ack
.rcv_mss
) {
149 icsk
->icsk_ack
.rcv_mss
= len
;
151 /* Otherwise, we make more careful check taking into account,
152 * that SACKs block is variable.
154 * "len" is invariant segment length, including TCP header.
156 len
+= skb
->data
- skb_transport_header(skb
);
157 if (len
>= TCP_MSS_DEFAULT
+ sizeof(struct tcphdr
) ||
158 /* If PSH is not set, packet should be
159 * full sized, provided peer TCP is not badly broken.
160 * This observation (if it is correct 8)) allows
161 * to handle super-low mtu links fairly.
163 (len
>= TCP_MIN_MSS
+ sizeof(struct tcphdr
) &&
164 !(tcp_flag_word(tcp_hdr(skb
)) & TCP_REMNANT
))) {
165 /* Subtract also invariant (if peer is RFC compliant),
166 * tcp header plus fixed timestamp option length.
167 * Resulting "len" is MSS free of SACK jitter.
169 len
-= tcp_sk(sk
)->tcp_header_len
;
170 icsk
->icsk_ack
.last_seg_size
= len
;
172 icsk
->icsk_ack
.rcv_mss
= len
;
176 if (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
)
177 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED2
;
178 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
182 static void tcp_incr_quickack(struct sock
*sk
)
184 struct inet_connection_sock
*icsk
= inet_csk(sk
);
185 unsigned int quickacks
= tcp_sk(sk
)->rcv_wnd
/ (2 * icsk
->icsk_ack
.rcv_mss
);
189 if (quickacks
> icsk
->icsk_ack
.quick
)
190 icsk
->icsk_ack
.quick
= min(quickacks
, TCP_MAX_QUICKACKS
);
193 static void tcp_enter_quickack_mode(struct sock
*sk
)
195 struct inet_connection_sock
*icsk
= inet_csk(sk
);
196 tcp_incr_quickack(sk
);
197 icsk
->icsk_ack
.pingpong
= 0;
198 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
201 /* Send ACKs quickly, if "quick" count is not exhausted
202 * and the session is not interactive.
205 static bool tcp_in_quickack_mode(struct sock
*sk
)
207 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
208 const struct dst_entry
*dst
= __sk_dst_get(sk
);
210 return (dst
&& dst_metric(dst
, RTAX_QUICKACK
)) ||
211 (icsk
->icsk_ack
.quick
&& !icsk
->icsk_ack
.pingpong
);
214 static void tcp_ecn_queue_cwr(struct tcp_sock
*tp
)
216 if (tp
->ecn_flags
& TCP_ECN_OK
)
217 tp
->ecn_flags
|= TCP_ECN_QUEUE_CWR
;
220 static void tcp_ecn_accept_cwr(struct tcp_sock
*tp
, const struct sk_buff
*skb
)
222 if (tcp_hdr(skb
)->cwr
)
223 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
226 static void tcp_ecn_withdraw_cwr(struct tcp_sock
*tp
)
228 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
231 static void __tcp_ecn_check_ce(struct tcp_sock
*tp
, const struct sk_buff
*skb
)
233 switch (TCP_SKB_CB(skb
)->ip_dsfield
& INET_ECN_MASK
) {
234 case INET_ECN_NOT_ECT
:
235 /* Funny extension: if ECT is not set on a segment,
236 * and we already seen ECT on a previous segment,
237 * it is probably a retransmit.
239 if (tp
->ecn_flags
& TCP_ECN_SEEN
)
240 tcp_enter_quickack_mode((struct sock
*)tp
);
243 if (tcp_ca_needs_ecn((struct sock
*)tp
))
244 tcp_ca_event((struct sock
*)tp
, CA_EVENT_ECN_IS_CE
);
246 if (!(tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)) {
247 /* Better not delay acks, sender can have a very low cwnd */
248 tcp_enter_quickack_mode((struct sock
*)tp
);
249 tp
->ecn_flags
|= TCP_ECN_DEMAND_CWR
;
251 tp
->ecn_flags
|= TCP_ECN_SEEN
;
254 if (tcp_ca_needs_ecn((struct sock
*)tp
))
255 tcp_ca_event((struct sock
*)tp
, CA_EVENT_ECN_NO_CE
);
256 tp
->ecn_flags
|= TCP_ECN_SEEN
;
261 static void tcp_ecn_check_ce(struct tcp_sock
*tp
, const struct sk_buff
*skb
)
263 if (tp
->ecn_flags
& TCP_ECN_OK
)
264 __tcp_ecn_check_ce(tp
, skb
);
267 static void tcp_ecn_rcv_synack(struct tcp_sock
*tp
, const struct tcphdr
*th
)
269 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| th
->cwr
))
270 tp
->ecn_flags
&= ~TCP_ECN_OK
;
273 static void tcp_ecn_rcv_syn(struct tcp_sock
*tp
, const struct tcphdr
*th
)
275 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| !th
->cwr
))
276 tp
->ecn_flags
&= ~TCP_ECN_OK
;
279 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock
*tp
, const struct tcphdr
*th
)
281 if (th
->ece
&& !th
->syn
&& (tp
->ecn_flags
& TCP_ECN_OK
))
286 /* Buffer size and advertised window tuning.
288 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
291 static void tcp_sndbuf_expand(struct sock
*sk
)
293 const struct tcp_sock
*tp
= tcp_sk(sk
);
297 /* Worst case is non GSO/TSO : each frame consumes one skb
298 * and skb->head is kmalloced using power of two area of memory
300 per_mss
= max_t(u32
, tp
->rx_opt
.mss_clamp
, tp
->mss_cache
) +
302 SKB_DATA_ALIGN(sizeof(struct skb_shared_info
));
304 per_mss
= roundup_pow_of_two(per_mss
) +
305 SKB_DATA_ALIGN(sizeof(struct sk_buff
));
307 nr_segs
= max_t(u32
, TCP_INIT_CWND
, tp
->snd_cwnd
);
308 nr_segs
= max_t(u32
, nr_segs
, tp
->reordering
+ 1);
310 /* Fast Recovery (RFC 5681 3.2) :
311 * Cubic needs 1.7 factor, rounded to 2 to include
312 * extra cushion (application might react slowly to POLLOUT)
314 sndmem
= 2 * nr_segs
* per_mss
;
316 if (sk
->sk_sndbuf
< sndmem
)
317 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
320 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
322 * All tcp_full_space() is split to two parts: "network" buffer, allocated
323 * forward and advertised in receiver window (tp->rcv_wnd) and
324 * "application buffer", required to isolate scheduling/application
325 * latencies from network.
326 * window_clamp is maximal advertised window. It can be less than
327 * tcp_full_space(), in this case tcp_full_space() - window_clamp
328 * is reserved for "application" buffer. The less window_clamp is
329 * the smoother our behaviour from viewpoint of network, but the lower
330 * throughput and the higher sensitivity of the connection to losses. 8)
332 * rcv_ssthresh is more strict window_clamp used at "slow start"
333 * phase to predict further behaviour of this connection.
334 * It is used for two goals:
335 * - to enforce header prediction at sender, even when application
336 * requires some significant "application buffer". It is check #1.
337 * - to prevent pruning of receive queue because of misprediction
338 * of receiver window. Check #2.
340 * The scheme does not work when sender sends good segments opening
341 * window and then starts to feed us spaghetti. But it should work
342 * in common situations. Otherwise, we have to rely on queue collapsing.
345 /* Slow part of check#2. */
346 static int __tcp_grow_window(const struct sock
*sk
, const struct sk_buff
*skb
)
348 struct tcp_sock
*tp
= tcp_sk(sk
);
350 int truesize
= tcp_win_from_space(skb
->truesize
) >> 1;
351 int window
= tcp_win_from_space(sysctl_tcp_rmem
[2]) >> 1;
353 while (tp
->rcv_ssthresh
<= window
) {
354 if (truesize
<= skb
->len
)
355 return 2 * inet_csk(sk
)->icsk_ack
.rcv_mss
;
363 static void tcp_grow_window(struct sock
*sk
, const struct sk_buff
*skb
)
365 struct tcp_sock
*tp
= tcp_sk(sk
);
368 if (tp
->rcv_ssthresh
< tp
->window_clamp
&&
369 (int)tp
->rcv_ssthresh
< tcp_space(sk
) &&
370 !tcp_under_memory_pressure(sk
)) {
373 /* Check #2. Increase window, if skb with such overhead
374 * will fit to rcvbuf in future.
376 if (tcp_win_from_space(skb
->truesize
) <= skb
->len
)
377 incr
= 2 * tp
->advmss
;
379 incr
= __tcp_grow_window(sk
, skb
);
382 incr
= max_t(int, incr
, 2 * skb
->len
);
383 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
+ incr
,
385 inet_csk(sk
)->icsk_ack
.quick
|= 1;
390 /* 3. Tuning rcvbuf, when connection enters established state. */
391 static void tcp_fixup_rcvbuf(struct sock
*sk
)
393 u32 mss
= tcp_sk(sk
)->advmss
;
396 rcvmem
= 2 * SKB_TRUESIZE(mss
+ MAX_TCP_HEADER
) *
397 tcp_default_init_rwnd(mss
);
399 /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
400 * Allow enough cushion so that sender is not limited by our window
402 if (sysctl_tcp_moderate_rcvbuf
)
405 if (sk
->sk_rcvbuf
< rcvmem
)
406 sk
->sk_rcvbuf
= min(rcvmem
, sysctl_tcp_rmem
[2]);
409 /* 4. Try to fixup all. It is made immediately after connection enters
412 void tcp_init_buffer_space(struct sock
*sk
)
414 struct tcp_sock
*tp
= tcp_sk(sk
);
417 if (!(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
))
418 tcp_fixup_rcvbuf(sk
);
419 if (!(sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
))
420 tcp_sndbuf_expand(sk
);
422 tp
->rcvq_space
.space
= tp
->rcv_wnd
;
423 tp
->rcvq_space
.time
= tcp_time_stamp
;
424 tp
->rcvq_space
.seq
= tp
->copied_seq
;
426 maxwin
= tcp_full_space(sk
);
428 if (tp
->window_clamp
>= maxwin
) {
429 tp
->window_clamp
= maxwin
;
431 if (sysctl_tcp_app_win
&& maxwin
> 4 * tp
->advmss
)
432 tp
->window_clamp
= max(maxwin
-
433 (maxwin
>> sysctl_tcp_app_win
),
437 /* Force reservation of one segment. */
438 if (sysctl_tcp_app_win
&&
439 tp
->window_clamp
> 2 * tp
->advmss
&&
440 tp
->window_clamp
+ tp
->advmss
> maxwin
)
441 tp
->window_clamp
= max(2 * tp
->advmss
, maxwin
- tp
->advmss
);
443 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, tp
->window_clamp
);
444 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
447 /* 5. Recalculate window clamp after socket hit its memory bounds. */
448 static void tcp_clamp_window(struct sock
*sk
)
450 struct tcp_sock
*tp
= tcp_sk(sk
);
451 struct inet_connection_sock
*icsk
= inet_csk(sk
);
453 icsk
->icsk_ack
.quick
= 0;
455 if (sk
->sk_rcvbuf
< sysctl_tcp_rmem
[2] &&
456 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
) &&
457 !tcp_under_memory_pressure(sk
) &&
458 sk_memory_allocated(sk
) < sk_prot_mem_limits(sk
, 0)) {
459 sk
->sk_rcvbuf
= min(atomic_read(&sk
->sk_rmem_alloc
),
462 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
)
463 tp
->rcv_ssthresh
= min(tp
->window_clamp
, 2U * tp
->advmss
);
466 /* Initialize RCV_MSS value.
467 * RCV_MSS is an our guess about MSS used by the peer.
468 * We haven't any direct information about the MSS.
469 * It's better to underestimate the RCV_MSS rather than overestimate.
470 * Overestimations make us ACKing less frequently than needed.
471 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
473 void tcp_initialize_rcv_mss(struct sock
*sk
)
475 const struct tcp_sock
*tp
= tcp_sk(sk
);
476 unsigned int hint
= min_t(unsigned int, tp
->advmss
, tp
->mss_cache
);
478 hint
= min(hint
, tp
->rcv_wnd
/ 2);
479 hint
= min(hint
, TCP_MSS_DEFAULT
);
480 hint
= max(hint
, TCP_MIN_MSS
);
482 inet_csk(sk
)->icsk_ack
.rcv_mss
= hint
;
484 EXPORT_SYMBOL(tcp_initialize_rcv_mss
);
486 /* Receiver "autotuning" code.
488 * The algorithm for RTT estimation w/o timestamps is based on
489 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
490 * <http://public.lanl.gov/radiant/pubs.html#DRS>
492 * More detail on this code can be found at
493 * <http://staff.psc.edu/jheffner/>,
494 * though this reference is out of date. A new paper
497 static void tcp_rcv_rtt_update(struct tcp_sock
*tp
, u32 sample
, int win_dep
)
499 u32 new_sample
= tp
->rcv_rtt_est
.rtt
;
505 if (new_sample
!= 0) {
506 /* If we sample in larger samples in the non-timestamp
507 * case, we could grossly overestimate the RTT especially
508 * with chatty applications or bulk transfer apps which
509 * are stalled on filesystem I/O.
511 * Also, since we are only going for a minimum in the
512 * non-timestamp case, we do not smooth things out
513 * else with timestamps disabled convergence takes too
517 m
-= (new_sample
>> 3);
525 /* No previous measure. */
529 if (tp
->rcv_rtt_est
.rtt
!= new_sample
)
530 tp
->rcv_rtt_est
.rtt
= new_sample
;
533 static inline void tcp_rcv_rtt_measure(struct tcp_sock
*tp
)
535 if (tp
->rcv_rtt_est
.time
== 0)
537 if (before(tp
->rcv_nxt
, tp
->rcv_rtt_est
.seq
))
539 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rcv_rtt_est
.time
, 1);
542 tp
->rcv_rtt_est
.seq
= tp
->rcv_nxt
+ tp
->rcv_wnd
;
543 tp
->rcv_rtt_est
.time
= tcp_time_stamp
;
546 static inline void tcp_rcv_rtt_measure_ts(struct sock
*sk
,
547 const struct sk_buff
*skb
)
549 struct tcp_sock
*tp
= tcp_sk(sk
);
550 if (tp
->rx_opt
.rcv_tsecr
&&
551 (TCP_SKB_CB(skb
)->end_seq
-
552 TCP_SKB_CB(skb
)->seq
>= inet_csk(sk
)->icsk_ack
.rcv_mss
))
553 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
, 0);
557 * This function should be called every time data is copied to user space.
558 * It calculates the appropriate TCP receive buffer space.
560 void tcp_rcv_space_adjust(struct sock
*sk
)
562 struct tcp_sock
*tp
= tcp_sk(sk
);
566 time
= tcp_time_stamp
- tp
->rcvq_space
.time
;
567 if (time
< (tp
->rcv_rtt_est
.rtt
>> 3) || tp
->rcv_rtt_est
.rtt
== 0)
570 /* Number of bytes copied to user in last RTT */
571 copied
= tp
->copied_seq
- tp
->rcvq_space
.seq
;
572 if (copied
<= tp
->rcvq_space
.space
)
576 * copied = bytes received in previous RTT, our base window
577 * To cope with packet losses, we need a 2x factor
578 * To cope with slow start, and sender growing its cwin by 100 %
579 * every RTT, we need a 4x factor, because the ACK we are sending
580 * now is for the next RTT, not the current one :
581 * <prev RTT . ><current RTT .. ><next RTT .... >
584 if (sysctl_tcp_moderate_rcvbuf
&&
585 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
)) {
586 int rcvwin
, rcvmem
, rcvbuf
;
588 /* minimal window to cope with packet losses, assuming
589 * steady state. Add some cushion because of small variations.
591 rcvwin
= (copied
<< 1) + 16 * tp
->advmss
;
593 /* If rate increased by 25%,
594 * assume slow start, rcvwin = 3 * copied
595 * If rate increased by 50%,
596 * assume sender can use 2x growth, rcvwin = 4 * copied
599 tp
->rcvq_space
.space
+ (tp
->rcvq_space
.space
>> 2)) {
601 tp
->rcvq_space
.space
+ (tp
->rcvq_space
.space
>> 1))
604 rcvwin
+= (rcvwin
>> 1);
607 rcvmem
= SKB_TRUESIZE(tp
->advmss
+ MAX_TCP_HEADER
);
608 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
611 rcvbuf
= min(rcvwin
/ tp
->advmss
* rcvmem
, sysctl_tcp_rmem
[2]);
612 if (rcvbuf
> sk
->sk_rcvbuf
) {
613 sk
->sk_rcvbuf
= rcvbuf
;
615 /* Make the window clamp follow along. */
616 tp
->window_clamp
= rcvwin
;
619 tp
->rcvq_space
.space
= copied
;
622 tp
->rcvq_space
.seq
= tp
->copied_seq
;
623 tp
->rcvq_space
.time
= tcp_time_stamp
;
626 /* There is something which you must keep in mind when you analyze the
627 * behavior of the tp->ato delayed ack timeout interval. When a
628 * connection starts up, we want to ack as quickly as possible. The
629 * problem is that "good" TCP's do slow start at the beginning of data
630 * transmission. The means that until we send the first few ACK's the
631 * sender will sit on his end and only queue most of his data, because
632 * he can only send snd_cwnd unacked packets at any given time. For
633 * each ACK we send, he increments snd_cwnd and transmits more of his
636 static void tcp_event_data_recv(struct sock
*sk
, struct sk_buff
*skb
)
638 struct tcp_sock
*tp
= tcp_sk(sk
);
639 struct inet_connection_sock
*icsk
= inet_csk(sk
);
642 inet_csk_schedule_ack(sk
);
644 tcp_measure_rcv_mss(sk
, skb
);
646 tcp_rcv_rtt_measure(tp
);
648 now
= tcp_time_stamp
;
650 if (!icsk
->icsk_ack
.ato
) {
651 /* The _first_ data packet received, initialize
652 * delayed ACK engine.
654 tcp_incr_quickack(sk
);
655 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
657 int m
= now
- icsk
->icsk_ack
.lrcvtime
;
659 if (m
<= TCP_ATO_MIN
/ 2) {
660 /* The fastest case is the first. */
661 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + TCP_ATO_MIN
/ 2;
662 } else if (m
< icsk
->icsk_ack
.ato
) {
663 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + m
;
664 if (icsk
->icsk_ack
.ato
> icsk
->icsk_rto
)
665 icsk
->icsk_ack
.ato
= icsk
->icsk_rto
;
666 } else if (m
> icsk
->icsk_rto
) {
667 /* Too long gap. Apparently sender failed to
668 * restart window, so that we send ACKs quickly.
670 tcp_incr_quickack(sk
);
674 icsk
->icsk_ack
.lrcvtime
= now
;
676 tcp_ecn_check_ce(tp
, skb
);
679 tcp_grow_window(sk
, skb
);
682 /* Called to compute a smoothed rtt estimate. The data fed to this
683 * routine either comes from timestamps, or from segments that were
684 * known _not_ to have been retransmitted [see Karn/Partridge
685 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
686 * piece by Van Jacobson.
687 * NOTE: the next three routines used to be one big routine.
688 * To save cycles in the RFC 1323 implementation it was better to break
689 * it up into three procedures. -- erics
691 static void tcp_rtt_estimator(struct sock
*sk
, long mrtt_us
)
693 struct tcp_sock
*tp
= tcp_sk(sk
);
694 long m
= mrtt_us
; /* RTT */
695 u32 srtt
= tp
->srtt_us
;
697 /* The following amusing code comes from Jacobson's
698 * article in SIGCOMM '88. Note that rtt and mdev
699 * are scaled versions of rtt and mean deviation.
700 * This is designed to be as fast as possible
701 * m stands for "measurement".
703 * On a 1990 paper the rto value is changed to:
704 * RTO = rtt + 4 * mdev
706 * Funny. This algorithm seems to be very broken.
707 * These formulae increase RTO, when it should be decreased, increase
708 * too slowly, when it should be increased quickly, decrease too quickly
709 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
710 * does not matter how to _calculate_ it. Seems, it was trap
711 * that VJ failed to avoid. 8)
714 m
-= (srtt
>> 3); /* m is now error in rtt est */
715 srtt
+= m
; /* rtt = 7/8 rtt + 1/8 new */
717 m
= -m
; /* m is now abs(error) */
718 m
-= (tp
->mdev_us
>> 2); /* similar update on mdev */
719 /* This is similar to one of Eifel findings.
720 * Eifel blocks mdev updates when rtt decreases.
721 * This solution is a bit different: we use finer gain
722 * for mdev in this case (alpha*beta).
723 * Like Eifel it also prevents growth of rto,
724 * but also it limits too fast rto decreases,
725 * happening in pure Eifel.
730 m
-= (tp
->mdev_us
>> 2); /* similar update on mdev */
732 tp
->mdev_us
+= m
; /* mdev = 3/4 mdev + 1/4 new */
733 if (tp
->mdev_us
> tp
->mdev_max_us
) {
734 tp
->mdev_max_us
= tp
->mdev_us
;
735 if (tp
->mdev_max_us
> tp
->rttvar_us
)
736 tp
->rttvar_us
= tp
->mdev_max_us
;
738 if (after(tp
->snd_una
, tp
->rtt_seq
)) {
739 if (tp
->mdev_max_us
< tp
->rttvar_us
)
740 tp
->rttvar_us
-= (tp
->rttvar_us
- tp
->mdev_max_us
) >> 2;
741 tp
->rtt_seq
= tp
->snd_nxt
;
742 tp
->mdev_max_us
= tcp_rto_min_us(sk
);
745 /* no previous measure. */
746 srtt
= m
<< 3; /* take the measured time to be rtt */
747 tp
->mdev_us
= m
<< 1; /* make sure rto = 3*rtt */
748 tp
->rttvar_us
= max(tp
->mdev_us
, tcp_rto_min_us(sk
));
749 tp
->mdev_max_us
= tp
->rttvar_us
;
750 tp
->rtt_seq
= tp
->snd_nxt
;
752 tp
->srtt_us
= max(1U, srtt
);
755 /* Set the sk_pacing_rate to allow proper sizing of TSO packets.
756 * Note: TCP stack does not yet implement pacing.
757 * FQ packet scheduler can be used to implement cheap but effective
758 * TCP pacing, to smooth the burst on large writes when packets
759 * in flight is significantly lower than cwnd (or rwin)
761 int sysctl_tcp_pacing_ss_ratio __read_mostly
= 200;
762 int sysctl_tcp_pacing_ca_ratio __read_mostly
= 120;
764 static void tcp_update_pacing_rate(struct sock
*sk
)
766 const struct tcp_sock
*tp
= tcp_sk(sk
);
769 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
770 rate
= (u64
)tp
->mss_cache
* ((USEC_PER_SEC
/ 100) << 3);
772 /* current rate is (cwnd * mss) / srtt
773 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
774 * In Congestion Avoidance phase, set it to 120 % the current rate.
776 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
777 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
778 * end of slow start and should slow down.
780 if (tp
->snd_cwnd
< tp
->snd_ssthresh
/ 2)
781 rate
*= sysctl_tcp_pacing_ss_ratio
;
783 rate
*= sysctl_tcp_pacing_ca_ratio
;
785 rate
*= max(tp
->snd_cwnd
, tp
->packets_out
);
787 if (likely(tp
->srtt_us
))
788 do_div(rate
, tp
->srtt_us
);
790 /* ACCESS_ONCE() is needed because sch_fq fetches sk_pacing_rate
791 * without any lock. We want to make sure compiler wont store
792 * intermediate values in this location.
794 ACCESS_ONCE(sk
->sk_pacing_rate
) = min_t(u64
, rate
,
795 sk
->sk_max_pacing_rate
);
798 /* Calculate rto without backoff. This is the second half of Van Jacobson's
799 * routine referred to above.
801 static void tcp_set_rto(struct sock
*sk
)
803 const struct tcp_sock
*tp
= tcp_sk(sk
);
804 /* Old crap is replaced with new one. 8)
807 * 1. If rtt variance happened to be less 50msec, it is hallucination.
808 * It cannot be less due to utterly erratic ACK generation made
809 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
810 * to do with delayed acks, because at cwnd>2 true delack timeout
811 * is invisible. Actually, Linux-2.4 also generates erratic
812 * ACKs in some circumstances.
814 inet_csk(sk
)->icsk_rto
= __tcp_set_rto(tp
);
816 /* 2. Fixups made earlier cannot be right.
817 * If we do not estimate RTO correctly without them,
818 * all the algo is pure shit and should be replaced
819 * with correct one. It is exactly, which we pretend to do.
822 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
823 * guarantees that rto is higher.
828 __u32
tcp_init_cwnd(const struct tcp_sock
*tp
, const struct dst_entry
*dst
)
830 __u32 cwnd
= (dst
? dst_metric(dst
, RTAX_INITCWND
) : 0);
833 cwnd
= TCP_INIT_CWND
;
834 return min_t(__u32
, cwnd
, tp
->snd_cwnd_clamp
);
838 * Packet counting of FACK is based on in-order assumptions, therefore TCP
839 * disables it when reordering is detected
841 void tcp_disable_fack(struct tcp_sock
*tp
)
843 /* RFC3517 uses different metric in lost marker => reset on change */
845 tp
->lost_skb_hint
= NULL
;
846 tp
->rx_opt
.sack_ok
&= ~TCP_FACK_ENABLED
;
849 /* Take a notice that peer is sending D-SACKs */
850 static void tcp_dsack_seen(struct tcp_sock
*tp
)
852 tp
->rx_opt
.sack_ok
|= TCP_DSACK_SEEN
;
855 static void tcp_update_reordering(struct sock
*sk
, const int metric
,
858 struct tcp_sock
*tp
= tcp_sk(sk
);
859 if (metric
> tp
->reordering
) {
862 tp
->reordering
= min(sysctl_tcp_max_reordering
, metric
);
864 /* This exciting event is worth to be remembered. 8) */
866 mib_idx
= LINUX_MIB_TCPTSREORDER
;
867 else if (tcp_is_reno(tp
))
868 mib_idx
= LINUX_MIB_TCPRENOREORDER
;
869 else if (tcp_is_fack(tp
))
870 mib_idx
= LINUX_MIB_TCPFACKREORDER
;
872 mib_idx
= LINUX_MIB_TCPSACKREORDER
;
874 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
875 #if FASTRETRANS_DEBUG > 1
876 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
877 tp
->rx_opt
.sack_ok
, inet_csk(sk
)->icsk_ca_state
,
881 tp
->undo_marker
? tp
->undo_retrans
: 0);
883 tcp_disable_fack(tp
);
887 tcp_disable_early_retrans(tp
);
891 /* This must be called before lost_out is incremented */
892 static void tcp_verify_retransmit_hint(struct tcp_sock
*tp
, struct sk_buff
*skb
)
894 if (!tp
->retransmit_skb_hint
||
895 before(TCP_SKB_CB(skb
)->seq
,
896 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
897 tp
->retransmit_skb_hint
= skb
;
900 after(TCP_SKB_CB(skb
)->end_seq
, tp
->retransmit_high
))
901 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
904 static void tcp_skb_mark_lost(struct tcp_sock
*tp
, struct sk_buff
*skb
)
906 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
907 tcp_verify_retransmit_hint(tp
, skb
);
909 tp
->lost_out
+= tcp_skb_pcount(skb
);
910 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
914 void tcp_skb_mark_lost_uncond_verify(struct tcp_sock
*tp
, struct sk_buff
*skb
)
916 tcp_verify_retransmit_hint(tp
, skb
);
918 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
919 tp
->lost_out
+= tcp_skb_pcount(skb
);
920 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
924 /* This procedure tags the retransmission queue when SACKs arrive.
926 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
927 * Packets in queue with these bits set are counted in variables
928 * sacked_out, retrans_out and lost_out, correspondingly.
930 * Valid combinations are:
931 * Tag InFlight Description
932 * 0 1 - orig segment is in flight.
933 * S 0 - nothing flies, orig reached receiver.
934 * L 0 - nothing flies, orig lost by net.
935 * R 2 - both orig and retransmit are in flight.
936 * L|R 1 - orig is lost, retransmit is in flight.
937 * S|R 1 - orig reached receiver, retrans is still in flight.
938 * (L|S|R is logically valid, it could occur when L|R is sacked,
939 * but it is equivalent to plain S and code short-curcuits it to S.
940 * L|S is logically invalid, it would mean -1 packet in flight 8))
942 * These 6 states form finite state machine, controlled by the following events:
943 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
944 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
945 * 3. Loss detection event of two flavors:
946 * A. Scoreboard estimator decided the packet is lost.
947 * A'. Reno "three dupacks" marks head of queue lost.
948 * A''. Its FACK modification, head until snd.fack is lost.
949 * B. SACK arrives sacking SND.NXT at the moment, when the
950 * segment was retransmitted.
951 * 4. D-SACK added new rule: D-SACK changes any tag to S.
953 * It is pleasant to note, that state diagram turns out to be commutative,
954 * so that we are allowed not to be bothered by order of our actions,
955 * when multiple events arrive simultaneously. (see the function below).
957 * Reordering detection.
958 * --------------------
959 * Reordering metric is maximal distance, which a packet can be displaced
960 * in packet stream. With SACKs we can estimate it:
962 * 1. SACK fills old hole and the corresponding segment was not
963 * ever retransmitted -> reordering. Alas, we cannot use it
964 * when segment was retransmitted.
965 * 2. The last flaw is solved with D-SACK. D-SACK arrives
966 * for retransmitted and already SACKed segment -> reordering..
967 * Both of these heuristics are not used in Loss state, when we cannot
968 * account for retransmits accurately.
970 * SACK block validation.
971 * ----------------------
973 * SACK block range validation checks that the received SACK block fits to
974 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
975 * Note that SND.UNA is not included to the range though being valid because
976 * it means that the receiver is rather inconsistent with itself reporting
977 * SACK reneging when it should advance SND.UNA. Such SACK block this is
978 * perfectly valid, however, in light of RFC2018 which explicitly states
979 * that "SACK block MUST reflect the newest segment. Even if the newest
980 * segment is going to be discarded ...", not that it looks very clever
981 * in case of head skb. Due to potentional receiver driven attacks, we
982 * choose to avoid immediate execution of a walk in write queue due to
983 * reneging and defer head skb's loss recovery to standard loss recovery
984 * procedure that will eventually trigger (nothing forbids us doing this).
986 * Implements also blockage to start_seq wrap-around. Problem lies in the
987 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
988 * there's no guarantee that it will be before snd_nxt (n). The problem
989 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
992 * <- outs wnd -> <- wrapzone ->
993 * u e n u_w e_w s n_w
995 * |<------------+------+----- TCP seqno space --------------+---------->|
996 * ...-- <2^31 ->| |<--------...
997 * ...---- >2^31 ------>| |<--------...
999 * Current code wouldn't be vulnerable but it's better still to discard such
1000 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1001 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1002 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1003 * equal to the ideal case (infinite seqno space without wrap caused issues).
1005 * With D-SACK the lower bound is extended to cover sequence space below
1006 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1007 * again, D-SACK block must not to go across snd_una (for the same reason as
1008 * for the normal SACK blocks, explained above). But there all simplicity
1009 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1010 * fully below undo_marker they do not affect behavior in anyway and can
1011 * therefore be safely ignored. In rare cases (which are more or less
1012 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1013 * fragmentation and packet reordering past skb's retransmission. To consider
1014 * them correctly, the acceptable range must be extended even more though
1015 * the exact amount is rather hard to quantify. However, tp->max_window can
1016 * be used as an exaggerated estimate.
1018 static bool tcp_is_sackblock_valid(struct tcp_sock
*tp
, bool is_dsack
,
1019 u32 start_seq
, u32 end_seq
)
1021 /* Too far in future, or reversed (interpretation is ambiguous) */
1022 if (after(end_seq
, tp
->snd_nxt
) || !before(start_seq
, end_seq
))
1025 /* Nasty start_seq wrap-around check (see comments above) */
1026 if (!before(start_seq
, tp
->snd_nxt
))
1029 /* In outstanding window? ...This is valid exit for D-SACKs too.
1030 * start_seq == snd_una is non-sensical (see comments above)
1032 if (after(start_seq
, tp
->snd_una
))
1035 if (!is_dsack
|| !tp
->undo_marker
)
1038 /* ...Then it's D-SACK, and must reside below snd_una completely */
1039 if (after(end_seq
, tp
->snd_una
))
1042 if (!before(start_seq
, tp
->undo_marker
))
1046 if (!after(end_seq
, tp
->undo_marker
))
1049 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1050 * start_seq < undo_marker and end_seq >= undo_marker.
1052 return !before(start_seq
, end_seq
- tp
->max_window
);
1055 static bool tcp_check_dsack(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1056 struct tcp_sack_block_wire
*sp
, int num_sacks
,
1059 struct tcp_sock
*tp
= tcp_sk(sk
);
1060 u32 start_seq_0
= get_unaligned_be32(&sp
[0].start_seq
);
1061 u32 end_seq_0
= get_unaligned_be32(&sp
[0].end_seq
);
1062 bool dup_sack
= false;
1064 if (before(start_seq_0
, TCP_SKB_CB(ack_skb
)->ack_seq
)) {
1067 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPDSACKRECV
);
1068 } else if (num_sacks
> 1) {
1069 u32 end_seq_1
= get_unaligned_be32(&sp
[1].end_seq
);
1070 u32 start_seq_1
= get_unaligned_be32(&sp
[1].start_seq
);
1072 if (!after(end_seq_0
, end_seq_1
) &&
1073 !before(start_seq_0
, start_seq_1
)) {
1076 NET_INC_STATS_BH(sock_net(sk
),
1077 LINUX_MIB_TCPDSACKOFORECV
);
1081 /* D-SACK for already forgotten data... Do dumb counting. */
1082 if (dup_sack
&& tp
->undo_marker
&& tp
->undo_retrans
> 0 &&
1083 !after(end_seq_0
, prior_snd_una
) &&
1084 after(end_seq_0
, tp
->undo_marker
))
1090 struct tcp_sacktag_state
{
1093 /* Timestamps for earliest and latest never-retransmitted segment
1094 * that was SACKed. RTO needs the earliest RTT to stay conservative,
1095 * but congestion control should still get an accurate delay signal.
1097 struct skb_mstamp first_sackt
;
1098 struct skb_mstamp last_sackt
;
1102 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1103 * the incoming SACK may not exactly match but we can find smaller MSS
1104 * aligned portion of it that matches. Therefore we might need to fragment
1105 * which may fail and creates some hassle (caller must handle error case
1108 * FIXME: this could be merged to shift decision code
1110 static int tcp_match_skb_to_sack(struct sock
*sk
, struct sk_buff
*skb
,
1111 u32 start_seq
, u32 end_seq
)
1115 unsigned int pkt_len
;
1118 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1119 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1121 if (tcp_skb_pcount(skb
) > 1 && !in_sack
&&
1122 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
1123 mss
= tcp_skb_mss(skb
);
1124 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1127 pkt_len
= start_seq
- TCP_SKB_CB(skb
)->seq
;
1131 pkt_len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1136 /* Round if necessary so that SACKs cover only full MSSes
1137 * and/or the remaining small portion (if present)
1139 if (pkt_len
> mss
) {
1140 unsigned int new_len
= (pkt_len
/ mss
) * mss
;
1141 if (!in_sack
&& new_len
< pkt_len
) {
1143 if (new_len
>= skb
->len
)
1148 err
= tcp_fragment(sk
, skb
, pkt_len
, mss
, GFP_ATOMIC
);
1156 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1157 static u8
tcp_sacktag_one(struct sock
*sk
,
1158 struct tcp_sacktag_state
*state
, u8 sacked
,
1159 u32 start_seq
, u32 end_seq
,
1160 int dup_sack
, int pcount
,
1161 const struct skb_mstamp
*xmit_time
)
1163 struct tcp_sock
*tp
= tcp_sk(sk
);
1164 int fack_count
= state
->fack_count
;
1166 /* Account D-SACK for retransmitted packet. */
1167 if (dup_sack
&& (sacked
& TCPCB_RETRANS
)) {
1168 if (tp
->undo_marker
&& tp
->undo_retrans
> 0 &&
1169 after(end_seq
, tp
->undo_marker
))
1171 if (sacked
& TCPCB_SACKED_ACKED
)
1172 state
->reord
= min(fack_count
, state
->reord
);
1175 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1176 if (!after(end_seq
, tp
->snd_una
))
1179 if (!(sacked
& TCPCB_SACKED_ACKED
)) {
1180 tcp_rack_advance(tp
, xmit_time
, sacked
);
1182 if (sacked
& TCPCB_SACKED_RETRANS
) {
1183 /* If the segment is not tagged as lost,
1184 * we do not clear RETRANS, believing
1185 * that retransmission is still in flight.
1187 if (sacked
& TCPCB_LOST
) {
1188 sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1189 tp
->lost_out
-= pcount
;
1190 tp
->retrans_out
-= pcount
;
1193 if (!(sacked
& TCPCB_RETRANS
)) {
1194 /* New sack for not retransmitted frame,
1195 * which was in hole. It is reordering.
1197 if (before(start_seq
,
1198 tcp_highest_sack_seq(tp
)))
1199 state
->reord
= min(fack_count
,
1201 if (!after(end_seq
, tp
->high_seq
))
1202 state
->flag
|= FLAG_ORIG_SACK_ACKED
;
1203 if (state
->first_sackt
.v64
== 0)
1204 state
->first_sackt
= *xmit_time
;
1205 state
->last_sackt
= *xmit_time
;
1208 if (sacked
& TCPCB_LOST
) {
1209 sacked
&= ~TCPCB_LOST
;
1210 tp
->lost_out
-= pcount
;
1214 sacked
|= TCPCB_SACKED_ACKED
;
1215 state
->flag
|= FLAG_DATA_SACKED
;
1216 tp
->sacked_out
+= pcount
;
1217 tp
->delivered
+= pcount
; /* Out-of-order packets delivered */
1219 fack_count
+= pcount
;
1221 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1222 if (!tcp_is_fack(tp
) && tp
->lost_skb_hint
&&
1223 before(start_seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
))
1224 tp
->lost_cnt_hint
+= pcount
;
1226 if (fack_count
> tp
->fackets_out
)
1227 tp
->fackets_out
= fack_count
;
1230 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1231 * frames and clear it. undo_retrans is decreased above, L|R frames
1232 * are accounted above as well.
1234 if (dup_sack
&& (sacked
& TCPCB_SACKED_RETRANS
)) {
1235 sacked
&= ~TCPCB_SACKED_RETRANS
;
1236 tp
->retrans_out
-= pcount
;
1242 /* Shift newly-SACKed bytes from this skb to the immediately previous
1243 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1245 static bool tcp_shifted_skb(struct sock
*sk
, struct sk_buff
*skb
,
1246 struct tcp_sacktag_state
*state
,
1247 unsigned int pcount
, int shifted
, int mss
,
1250 struct tcp_sock
*tp
= tcp_sk(sk
);
1251 struct sk_buff
*prev
= tcp_write_queue_prev(sk
, skb
);
1252 u32 start_seq
= TCP_SKB_CB(skb
)->seq
; /* start of newly-SACKed */
1253 u32 end_seq
= start_seq
+ shifted
; /* end of newly-SACKed */
1257 /* Adjust counters and hints for the newly sacked sequence
1258 * range but discard the return value since prev is already
1259 * marked. We must tag the range first because the seq
1260 * advancement below implicitly advances
1261 * tcp_highest_sack_seq() when skb is highest_sack.
1263 tcp_sacktag_one(sk
, state
, TCP_SKB_CB(skb
)->sacked
,
1264 start_seq
, end_seq
, dup_sack
, pcount
,
1267 if (skb
== tp
->lost_skb_hint
)
1268 tp
->lost_cnt_hint
+= pcount
;
1270 TCP_SKB_CB(prev
)->end_seq
+= shifted
;
1271 TCP_SKB_CB(skb
)->seq
+= shifted
;
1273 tcp_skb_pcount_add(prev
, pcount
);
1274 BUG_ON(tcp_skb_pcount(skb
) < pcount
);
1275 tcp_skb_pcount_add(skb
, -pcount
);
1277 /* When we're adding to gso_segs == 1, gso_size will be zero,
1278 * in theory this shouldn't be necessary but as long as DSACK
1279 * code can come after this skb later on it's better to keep
1280 * setting gso_size to something.
1282 if (!TCP_SKB_CB(prev
)->tcp_gso_size
)
1283 TCP_SKB_CB(prev
)->tcp_gso_size
= mss
;
1285 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1286 if (tcp_skb_pcount(skb
) <= 1)
1287 TCP_SKB_CB(skb
)->tcp_gso_size
= 0;
1289 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1290 TCP_SKB_CB(prev
)->sacked
|= (TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
);
1293 BUG_ON(!tcp_skb_pcount(skb
));
1294 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKSHIFTED
);
1298 /* Whole SKB was eaten :-) */
1300 if (skb
== tp
->retransmit_skb_hint
)
1301 tp
->retransmit_skb_hint
= prev
;
1302 if (skb
== tp
->lost_skb_hint
) {
1303 tp
->lost_skb_hint
= prev
;
1304 tp
->lost_cnt_hint
-= tcp_skb_pcount(prev
);
1307 TCP_SKB_CB(prev
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
;
1308 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
1309 TCP_SKB_CB(prev
)->end_seq
++;
1311 if (skb
== tcp_highest_sack(sk
))
1312 tcp_advance_highest_sack(sk
, skb
);
1314 tcp_unlink_write_queue(skb
, sk
);
1315 sk_wmem_free_skb(sk
, skb
);
1317 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKMERGED
);
1322 /* I wish gso_size would have a bit more sane initialization than
1323 * something-or-zero which complicates things
1325 static int tcp_skb_seglen(const struct sk_buff
*skb
)
1327 return tcp_skb_pcount(skb
) == 1 ? skb
->len
: tcp_skb_mss(skb
);
1330 /* Shifting pages past head area doesn't work */
1331 static int skb_can_shift(const struct sk_buff
*skb
)
1333 return !skb_headlen(skb
) && skb_is_nonlinear(skb
);
1336 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1339 static struct sk_buff
*tcp_shift_skb_data(struct sock
*sk
, struct sk_buff
*skb
,
1340 struct tcp_sacktag_state
*state
,
1341 u32 start_seq
, u32 end_seq
,
1344 struct tcp_sock
*tp
= tcp_sk(sk
);
1345 struct sk_buff
*prev
;
1351 if (!sk_can_gso(sk
))
1354 /* Normally R but no L won't result in plain S */
1356 (TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_RETRANS
)) == TCPCB_SACKED_RETRANS
)
1358 if (!skb_can_shift(skb
))
1360 /* This frame is about to be dropped (was ACKed). */
1361 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1364 /* Can only happen with delayed DSACK + discard craziness */
1365 if (unlikely(skb
== tcp_write_queue_head(sk
)))
1367 prev
= tcp_write_queue_prev(sk
, skb
);
1369 if ((TCP_SKB_CB(prev
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
)
1372 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1373 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1377 pcount
= tcp_skb_pcount(skb
);
1378 mss
= tcp_skb_seglen(skb
);
1380 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1381 * drop this restriction as unnecessary
1383 if (mss
!= tcp_skb_seglen(prev
))
1386 if (!after(TCP_SKB_CB(skb
)->end_seq
, start_seq
))
1388 /* CHECKME: This is non-MSS split case only?, this will
1389 * cause skipped skbs due to advancing loop btw, original
1390 * has that feature too
1392 if (tcp_skb_pcount(skb
) <= 1)
1395 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1397 /* TODO: head merge to next could be attempted here
1398 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1399 * though it might not be worth of the additional hassle
1401 * ...we can probably just fallback to what was done
1402 * previously. We could try merging non-SACKed ones
1403 * as well but it probably isn't going to buy off
1404 * because later SACKs might again split them, and
1405 * it would make skb timestamp tracking considerably
1411 len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1413 BUG_ON(len
> skb
->len
);
1415 /* MSS boundaries should be honoured or else pcount will
1416 * severely break even though it makes things bit trickier.
1417 * Optimize common case to avoid most of the divides
1419 mss
= tcp_skb_mss(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
))
1429 } else if (len
< mss
) {
1437 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1438 if (!after(TCP_SKB_CB(skb
)->seq
+ len
, tp
->snd_una
))
1441 if (!skb_shift(prev
, skb
, len
))
1443 if (!tcp_shifted_skb(sk
, skb
, state
, pcount
, len
, mss
, dup_sack
))
1446 /* Hole filled allows collapsing with the next as well, this is very
1447 * useful when hole on every nth skb pattern happens
1449 if (prev
== tcp_write_queue_tail(sk
))
1451 skb
= tcp_write_queue_next(sk
, prev
);
1453 if (!skb_can_shift(skb
) ||
1454 (skb
== tcp_send_head(sk
)) ||
1455 ((TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
) ||
1456 (mss
!= tcp_skb_seglen(skb
)))
1460 if (skb_shift(prev
, skb
, len
)) {
1461 pcount
+= tcp_skb_pcount(skb
);
1462 tcp_shifted_skb(sk
, skb
, state
, tcp_skb_pcount(skb
), len
, mss
, 0);
1466 state
->fack_count
+= pcount
;
1473 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKSHIFTFALLBACK
);
1477 static struct sk_buff
*tcp_sacktag_walk(struct sk_buff
*skb
, struct sock
*sk
,
1478 struct tcp_sack_block
*next_dup
,
1479 struct tcp_sacktag_state
*state
,
1480 u32 start_seq
, u32 end_seq
,
1483 struct tcp_sock
*tp
= tcp_sk(sk
);
1484 struct sk_buff
*tmp
;
1486 tcp_for_write_queue_from(skb
, sk
) {
1488 bool dup_sack
= dup_sack_in
;
1490 if (skb
== tcp_send_head(sk
))
1493 /* queue is in-order => we can short-circuit the walk early */
1494 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1498 before(TCP_SKB_CB(skb
)->seq
, next_dup
->end_seq
)) {
1499 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1500 next_dup
->start_seq
,
1506 /* skb reference here is a bit tricky to get right, since
1507 * shifting can eat and free both this skb and the next,
1508 * so not even _safe variant of the loop is enough.
1511 tmp
= tcp_shift_skb_data(sk
, skb
, state
,
1512 start_seq
, end_seq
, dup_sack
);
1521 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1527 if (unlikely(in_sack
< 0))
1531 TCP_SKB_CB(skb
)->sacked
=
1534 TCP_SKB_CB(skb
)->sacked
,
1535 TCP_SKB_CB(skb
)->seq
,
1536 TCP_SKB_CB(skb
)->end_seq
,
1538 tcp_skb_pcount(skb
),
1541 if (!before(TCP_SKB_CB(skb
)->seq
,
1542 tcp_highest_sack_seq(tp
)))
1543 tcp_advance_highest_sack(sk
, skb
);
1546 state
->fack_count
+= tcp_skb_pcount(skb
);
1551 /* Avoid all extra work that is being done by sacktag while walking in
1554 static struct sk_buff
*tcp_sacktag_skip(struct sk_buff
*skb
, struct sock
*sk
,
1555 struct tcp_sacktag_state
*state
,
1558 tcp_for_write_queue_from(skb
, sk
) {
1559 if (skb
== tcp_send_head(sk
))
1562 if (after(TCP_SKB_CB(skb
)->end_seq
, skip_to_seq
))
1565 state
->fack_count
+= tcp_skb_pcount(skb
);
1570 static struct sk_buff
*tcp_maybe_skipping_dsack(struct sk_buff
*skb
,
1572 struct tcp_sack_block
*next_dup
,
1573 struct tcp_sacktag_state
*state
,
1579 if (before(next_dup
->start_seq
, skip_to_seq
)) {
1580 skb
= tcp_sacktag_skip(skb
, sk
, state
, next_dup
->start_seq
);
1581 skb
= tcp_sacktag_walk(skb
, sk
, NULL
, state
,
1582 next_dup
->start_seq
, next_dup
->end_seq
,
1589 static int tcp_sack_cache_ok(const struct tcp_sock
*tp
, const struct tcp_sack_block
*cache
)
1591 return cache
< tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1595 tcp_sacktag_write_queue(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1596 u32 prior_snd_una
, struct tcp_sacktag_state
*state
)
1598 struct tcp_sock
*tp
= tcp_sk(sk
);
1599 const unsigned char *ptr
= (skb_transport_header(ack_skb
) +
1600 TCP_SKB_CB(ack_skb
)->sacked
);
1601 struct tcp_sack_block_wire
*sp_wire
= (struct tcp_sack_block_wire
*)(ptr
+2);
1602 struct tcp_sack_block sp
[TCP_NUM_SACKS
];
1603 struct tcp_sack_block
*cache
;
1604 struct sk_buff
*skb
;
1605 int num_sacks
= min(TCP_NUM_SACKS
, (ptr
[1] - TCPOLEN_SACK_BASE
) >> 3);
1607 bool found_dup_sack
= false;
1609 int first_sack_index
;
1612 state
->reord
= tp
->packets_out
;
1614 if (!tp
->sacked_out
) {
1615 if (WARN_ON(tp
->fackets_out
))
1616 tp
->fackets_out
= 0;
1617 tcp_highest_sack_reset(sk
);
1620 found_dup_sack
= tcp_check_dsack(sk
, ack_skb
, sp_wire
,
1621 num_sacks
, prior_snd_una
);
1623 state
->flag
|= FLAG_DSACKING_ACK
;
1625 /* Eliminate too old ACKs, but take into
1626 * account more or less fresh ones, they can
1627 * contain valid SACK info.
1629 if (before(TCP_SKB_CB(ack_skb
)->ack_seq
, prior_snd_una
- tp
->max_window
))
1632 if (!tp
->packets_out
)
1636 first_sack_index
= 0;
1637 for (i
= 0; i
< num_sacks
; i
++) {
1638 bool dup_sack
= !i
&& found_dup_sack
;
1640 sp
[used_sacks
].start_seq
= get_unaligned_be32(&sp_wire
[i
].start_seq
);
1641 sp
[used_sacks
].end_seq
= get_unaligned_be32(&sp_wire
[i
].end_seq
);
1643 if (!tcp_is_sackblock_valid(tp
, dup_sack
,
1644 sp
[used_sacks
].start_seq
,
1645 sp
[used_sacks
].end_seq
)) {
1649 if (!tp
->undo_marker
)
1650 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDNOUNDO
;
1652 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDOLD
;
1654 /* Don't count olds caused by ACK reordering */
1655 if ((TCP_SKB_CB(ack_skb
)->ack_seq
!= tp
->snd_una
) &&
1656 !after(sp
[used_sacks
].end_seq
, tp
->snd_una
))
1658 mib_idx
= LINUX_MIB_TCPSACKDISCARD
;
1661 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
1663 first_sack_index
= -1;
1667 /* Ignore very old stuff early */
1668 if (!after(sp
[used_sacks
].end_seq
, prior_snd_una
))
1674 /* order SACK blocks to allow in order walk of the retrans queue */
1675 for (i
= used_sacks
- 1; i
> 0; i
--) {
1676 for (j
= 0; j
< i
; j
++) {
1677 if (after(sp
[j
].start_seq
, sp
[j
+ 1].start_seq
)) {
1678 swap(sp
[j
], sp
[j
+ 1]);
1680 /* Track where the first SACK block goes to */
1681 if (j
== first_sack_index
)
1682 first_sack_index
= j
+ 1;
1687 skb
= tcp_write_queue_head(sk
);
1688 state
->fack_count
= 0;
1691 if (!tp
->sacked_out
) {
1692 /* It's already past, so skip checking against it */
1693 cache
= tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1695 cache
= tp
->recv_sack_cache
;
1696 /* Skip empty blocks in at head of the cache */
1697 while (tcp_sack_cache_ok(tp
, cache
) && !cache
->start_seq
&&
1702 while (i
< used_sacks
) {
1703 u32 start_seq
= sp
[i
].start_seq
;
1704 u32 end_seq
= sp
[i
].end_seq
;
1705 bool dup_sack
= (found_dup_sack
&& (i
== first_sack_index
));
1706 struct tcp_sack_block
*next_dup
= NULL
;
1708 if (found_dup_sack
&& ((i
+ 1) == first_sack_index
))
1709 next_dup
= &sp
[i
+ 1];
1711 /* Skip too early cached blocks */
1712 while (tcp_sack_cache_ok(tp
, cache
) &&
1713 !before(start_seq
, cache
->end_seq
))
1716 /* Can skip some work by looking recv_sack_cache? */
1717 if (tcp_sack_cache_ok(tp
, cache
) && !dup_sack
&&
1718 after(end_seq
, cache
->start_seq
)) {
1721 if (before(start_seq
, cache
->start_seq
)) {
1722 skb
= tcp_sacktag_skip(skb
, sk
, state
,
1724 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
,
1731 /* Rest of the block already fully processed? */
1732 if (!after(end_seq
, cache
->end_seq
))
1735 skb
= tcp_maybe_skipping_dsack(skb
, sk
, next_dup
,
1739 /* ...tail remains todo... */
1740 if (tcp_highest_sack_seq(tp
) == cache
->end_seq
) {
1741 /* ...but better entrypoint exists! */
1742 skb
= tcp_highest_sack(sk
);
1745 state
->fack_count
= tp
->fackets_out
;
1750 skb
= tcp_sacktag_skip(skb
, sk
, state
, cache
->end_seq
);
1751 /* Check overlap against next cached too (past this one already) */
1756 if (!before(start_seq
, tcp_highest_sack_seq(tp
))) {
1757 skb
= tcp_highest_sack(sk
);
1760 state
->fack_count
= tp
->fackets_out
;
1762 skb
= tcp_sacktag_skip(skb
, sk
, state
, start_seq
);
1765 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
, state
,
1766 start_seq
, end_seq
, dup_sack
);
1772 /* Clear the head of the cache sack blocks so we can skip it next time */
1773 for (i
= 0; i
< ARRAY_SIZE(tp
->recv_sack_cache
) - used_sacks
; i
++) {
1774 tp
->recv_sack_cache
[i
].start_seq
= 0;
1775 tp
->recv_sack_cache
[i
].end_seq
= 0;
1777 for (j
= 0; j
< used_sacks
; j
++)
1778 tp
->recv_sack_cache
[i
++] = sp
[j
];
1780 if ((state
->reord
< tp
->fackets_out
) &&
1781 ((inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Loss
) || tp
->undo_marker
))
1782 tcp_update_reordering(sk
, tp
->fackets_out
- state
->reord
, 0);
1784 tcp_verify_left_out(tp
);
1787 #if FASTRETRANS_DEBUG > 0
1788 WARN_ON((int)tp
->sacked_out
< 0);
1789 WARN_ON((int)tp
->lost_out
< 0);
1790 WARN_ON((int)tp
->retrans_out
< 0);
1791 WARN_ON((int)tcp_packets_in_flight(tp
) < 0);
1796 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1797 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1799 static bool tcp_limit_reno_sacked(struct tcp_sock
*tp
)
1803 holes
= max(tp
->lost_out
, 1U);
1804 holes
= min(holes
, tp
->packets_out
);
1806 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1807 tp
->sacked_out
= tp
->packets_out
- holes
;
1813 /* If we receive more dupacks than we expected counting segments
1814 * in assumption of absent reordering, interpret this as reordering.
1815 * The only another reason could be bug in receiver TCP.
1817 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1819 struct tcp_sock
*tp
= tcp_sk(sk
);
1820 if (tcp_limit_reno_sacked(tp
))
1821 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1824 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1826 static void tcp_add_reno_sack(struct sock
*sk
)
1828 struct tcp_sock
*tp
= tcp_sk(sk
);
1829 u32 prior_sacked
= tp
->sacked_out
;
1832 tcp_check_reno_reordering(sk
, 0);
1833 if (tp
->sacked_out
> prior_sacked
)
1834 tp
->delivered
++; /* Some out-of-order packet is delivered */
1835 tcp_verify_left_out(tp
);
1838 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1840 static void tcp_remove_reno_sacks(struct sock
*sk
, int acked
)
1842 struct tcp_sock
*tp
= tcp_sk(sk
);
1845 /* One ACK acked hole. The rest eat duplicate ACKs. */
1846 tp
->delivered
+= max_t(int, acked
- tp
->sacked_out
, 1);
1847 if (acked
- 1 >= tp
->sacked_out
)
1850 tp
->sacked_out
-= acked
- 1;
1852 tcp_check_reno_reordering(sk
, acked
);
1853 tcp_verify_left_out(tp
);
1856 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1861 void tcp_clear_retrans(struct tcp_sock
*tp
)
1863 tp
->retrans_out
= 0;
1865 tp
->undo_marker
= 0;
1866 tp
->undo_retrans
= -1;
1867 tp
->fackets_out
= 0;
1871 static inline void tcp_init_undo(struct tcp_sock
*tp
)
1873 tp
->undo_marker
= tp
->snd_una
;
1874 /* Retransmission still in flight may cause DSACKs later. */
1875 tp
->undo_retrans
= tp
->retrans_out
? : -1;
1878 /* Enter Loss state. If we detect SACK reneging, forget all SACK information
1879 * and reset tags completely, otherwise preserve SACKs. If receiver
1880 * dropped its ofo queue, we will know this due to reneging detection.
1882 void tcp_enter_loss(struct sock
*sk
)
1884 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1885 struct tcp_sock
*tp
= tcp_sk(sk
);
1886 struct sk_buff
*skb
;
1887 bool new_recovery
= icsk
->icsk_ca_state
< TCP_CA_Recovery
;
1888 bool is_reneg
; /* is receiver reneging on SACKs? */
1890 /* Reduce ssthresh if it has not yet been made inside this window. */
1891 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
||
1892 !after(tp
->high_seq
, tp
->snd_una
) ||
1893 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
1894 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1895 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1896 tcp_ca_event(sk
, CA_EVENT_LOSS
);
1900 tp
->snd_cwnd_cnt
= 0;
1901 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1903 tp
->retrans_out
= 0;
1906 if (tcp_is_reno(tp
))
1907 tcp_reset_reno_sack(tp
);
1909 skb
= tcp_write_queue_head(sk
);
1910 is_reneg
= skb
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
);
1912 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSACKRENEGING
);
1914 tp
->fackets_out
= 0;
1916 tcp_clear_all_retrans_hints(tp
);
1918 tcp_for_write_queue(skb
, sk
) {
1919 if (skb
== tcp_send_head(sk
))
1922 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
1923 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) || is_reneg
) {
1924 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
1925 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1926 tp
->lost_out
+= tcp_skb_pcount(skb
);
1927 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
1930 tcp_verify_left_out(tp
);
1932 /* Timeout in disordered state after receiving substantial DUPACKs
1933 * suggests that the degree of reordering is over-estimated.
1935 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
&&
1936 tp
->sacked_out
>= sysctl_tcp_reordering
)
1937 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1938 sysctl_tcp_reordering
);
1939 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1940 tp
->high_seq
= tp
->snd_nxt
;
1941 tcp_ecn_queue_cwr(tp
);
1943 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1944 * loss recovery is underway except recurring timeout(s) on
1945 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1947 tp
->frto
= sysctl_tcp_frto
&&
1948 (new_recovery
|| icsk
->icsk_retransmits
) &&
1949 !inet_csk(sk
)->icsk_mtup
.probe_size
;
1952 /* If ACK arrived pointing to a remembered SACK, it means that our
1953 * remembered SACKs do not reflect real state of receiver i.e.
1954 * receiver _host_ is heavily congested (or buggy).
1956 * To avoid big spurious retransmission bursts due to transient SACK
1957 * scoreboard oddities that look like reneging, we give the receiver a
1958 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
1959 * restore sanity to the SACK scoreboard. If the apparent reneging
1960 * persists until this RTO then we'll clear the SACK scoreboard.
1962 static bool tcp_check_sack_reneging(struct sock
*sk
, int flag
)
1964 if (flag
& FLAG_SACK_RENEGING
) {
1965 struct tcp_sock
*tp
= tcp_sk(sk
);
1966 unsigned long delay
= max(usecs_to_jiffies(tp
->srtt_us
>> 4),
1967 msecs_to_jiffies(10));
1969 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1970 delay
, TCP_RTO_MAX
);
1976 static inline int tcp_fackets_out(const struct tcp_sock
*tp
)
1978 return tcp_is_reno(tp
) ? tp
->sacked_out
+ 1 : tp
->fackets_out
;
1981 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
1982 * counter when SACK is enabled (without SACK, sacked_out is used for
1985 * Instead, with FACK TCP uses fackets_out that includes both SACKed
1986 * segments up to the highest received SACK block so far and holes in
1989 * With reordering, holes may still be in flight, so RFC3517 recovery
1990 * uses pure sacked_out (total number of SACKed segments) even though
1991 * it violates the RFC that uses duplicate ACKs, often these are equal
1992 * but when e.g. out-of-window ACKs or packet duplication occurs,
1993 * they differ. Since neither occurs due to loss, TCP should really
1996 static inline int tcp_dupack_heuristics(const struct tcp_sock
*tp
)
1998 return tcp_is_fack(tp
) ? tp
->fackets_out
: tp
->sacked_out
+ 1;
2001 static bool tcp_pause_early_retransmit(struct sock
*sk
, int flag
)
2003 struct tcp_sock
*tp
= tcp_sk(sk
);
2004 unsigned long delay
;
2006 /* Delay early retransmit and entering fast recovery for
2007 * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
2008 * available, or RTO is scheduled to fire first.
2010 if (sysctl_tcp_early_retrans
< 2 || sysctl_tcp_early_retrans
> 3 ||
2011 (flag
& FLAG_ECE
) || !tp
->srtt_us
)
2014 delay
= max(usecs_to_jiffies(tp
->srtt_us
>> 5),
2015 msecs_to_jiffies(2));
2017 if (!time_after(inet_csk(sk
)->icsk_timeout
, (jiffies
+ delay
)))
2020 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_EARLY_RETRANS
, delay
,
2025 /* Linux NewReno/SACK/FACK/ECN state machine.
2026 * --------------------------------------
2028 * "Open" Normal state, no dubious events, fast path.
2029 * "Disorder" In all the respects it is "Open",
2030 * but requires a bit more attention. It is entered when
2031 * we see some SACKs or dupacks. It is split of "Open"
2032 * mainly to move some processing from fast path to slow one.
2033 * "CWR" CWND was reduced due to some Congestion Notification event.
2034 * It can be ECN, ICMP source quench, local device congestion.
2035 * "Recovery" CWND was reduced, we are fast-retransmitting.
2036 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2038 * tcp_fastretrans_alert() is entered:
2039 * - each incoming ACK, if state is not "Open"
2040 * - when arrived ACK is unusual, namely:
2045 * Counting packets in flight is pretty simple.
2047 * in_flight = packets_out - left_out + retrans_out
2049 * packets_out is SND.NXT-SND.UNA counted in packets.
2051 * retrans_out is number of retransmitted segments.
2053 * left_out is number of segments left network, but not ACKed yet.
2055 * left_out = sacked_out + lost_out
2057 * sacked_out: Packets, which arrived to receiver out of order
2058 * and hence not ACKed. With SACKs this number is simply
2059 * amount of SACKed data. Even without SACKs
2060 * it is easy to give pretty reliable estimate of this number,
2061 * counting duplicate ACKs.
2063 * lost_out: Packets lost by network. TCP has no explicit
2064 * "loss notification" feedback from network (for now).
2065 * It means that this number can be only _guessed_.
2066 * Actually, it is the heuristics to predict lossage that
2067 * distinguishes different algorithms.
2069 * F.e. after RTO, when all the queue is considered as lost,
2070 * lost_out = packets_out and in_flight = retrans_out.
2072 * Essentially, we have now two algorithms counting
2075 * FACK: It is the simplest heuristics. As soon as we decided
2076 * that something is lost, we decide that _all_ not SACKed
2077 * packets until the most forward SACK are lost. I.e.
2078 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2079 * It is absolutely correct estimate, if network does not reorder
2080 * packets. And it loses any connection to reality when reordering
2081 * takes place. We use FACK by default until reordering
2082 * is suspected on the path to this destination.
2084 * NewReno: when Recovery is entered, we assume that one segment
2085 * is lost (classic Reno). While we are in Recovery and
2086 * a partial ACK arrives, we assume that one more packet
2087 * is lost (NewReno). This heuristics are the same in NewReno
2090 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2091 * deflation etc. CWND is real congestion window, never inflated, changes
2092 * only according to classic VJ rules.
2094 * Really tricky (and requiring careful tuning) part of algorithm
2095 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2096 * The first determines the moment _when_ we should reduce CWND and,
2097 * hence, slow down forward transmission. In fact, it determines the moment
2098 * when we decide that hole is caused by loss, rather than by a reorder.
2100 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2101 * holes, caused by lost packets.
2103 * And the most logically complicated part of algorithm is undo
2104 * heuristics. We detect false retransmits due to both too early
2105 * fast retransmit (reordering) and underestimated RTO, analyzing
2106 * timestamps and D-SACKs. When we detect that some segments were
2107 * retransmitted by mistake and CWND reduction was wrong, we undo
2108 * window reduction and abort recovery phase. This logic is hidden
2109 * inside several functions named tcp_try_undo_<something>.
2112 /* This function decides, when we should leave Disordered state
2113 * and enter Recovery phase, reducing congestion window.
2115 * Main question: may we further continue forward transmission
2116 * with the same cwnd?
2118 static bool tcp_time_to_recover(struct sock
*sk
, int flag
)
2120 struct tcp_sock
*tp
= tcp_sk(sk
);
2123 /* Trick#1: The loss is proven. */
2127 /* Not-A-Trick#2 : Classic rule... */
2128 if (tcp_dupack_heuristics(tp
) > tp
->reordering
)
2131 /* Trick#4: It is still not OK... But will it be useful to delay
2134 packets_out
= tp
->packets_out
;
2135 if (packets_out
<= tp
->reordering
&&
2136 tp
->sacked_out
>= max_t(__u32
, packets_out
/2, sysctl_tcp_reordering
) &&
2137 !tcp_may_send_now(sk
)) {
2138 /* We have nothing to send. This connection is limited
2139 * either by receiver window or by application.
2144 /* If a thin stream is detected, retransmit after first
2145 * received dupack. Employ only if SACK is supported in order
2146 * to avoid possible corner-case series of spurious retransmissions
2147 * Use only if there are no unsent data.
2149 if ((tp
->thin_dupack
|| sysctl_tcp_thin_dupack
) &&
2150 tcp_stream_is_thin(tp
) && tcp_dupack_heuristics(tp
) > 1 &&
2151 tcp_is_sack(tp
) && !tcp_send_head(sk
))
2154 /* Trick#6: TCP early retransmit, per RFC5827. To avoid spurious
2155 * retransmissions due to small network reorderings, we implement
2156 * Mitigation A.3 in the RFC and delay the retransmission for a short
2157 * interval if appropriate.
2159 if (tp
->do_early_retrans
&& !tp
->retrans_out
&& tp
->sacked_out
&&
2160 (tp
->packets_out
>= (tp
->sacked_out
+ 1) && tp
->packets_out
< 4) &&
2161 !tcp_may_send_now(sk
))
2162 return !tcp_pause_early_retransmit(sk
, flag
);
2167 /* Detect loss in event "A" above by marking head of queue up as lost.
2168 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2169 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2170 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2171 * the maximum SACKed segments to pass before reaching this limit.
2173 static void tcp_mark_head_lost(struct sock
*sk
, int packets
, int mark_head
)
2175 struct tcp_sock
*tp
= tcp_sk(sk
);
2176 struct sk_buff
*skb
;
2177 int cnt
, oldcnt
, lost
;
2179 /* Use SACK to deduce losses of new sequences sent during recovery */
2180 const u32 loss_high
= tcp_is_sack(tp
) ? tp
->snd_nxt
: tp
->high_seq
;
2182 WARN_ON(packets
> tp
->packets_out
);
2183 if (tp
->lost_skb_hint
) {
2184 skb
= tp
->lost_skb_hint
;
2185 cnt
= tp
->lost_cnt_hint
;
2186 /* Head already handled? */
2187 if (mark_head
&& skb
!= tcp_write_queue_head(sk
))
2190 skb
= tcp_write_queue_head(sk
);
2194 tcp_for_write_queue_from(skb
, sk
) {
2195 if (skb
== tcp_send_head(sk
))
2197 /* TODO: do this better */
2198 /* this is not the most efficient way to do this... */
2199 tp
->lost_skb_hint
= skb
;
2200 tp
->lost_cnt_hint
= cnt
;
2202 if (after(TCP_SKB_CB(skb
)->end_seq
, loss_high
))
2206 if (tcp_is_fack(tp
) || tcp_is_reno(tp
) ||
2207 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2208 cnt
+= tcp_skb_pcount(skb
);
2210 if (cnt
> packets
) {
2211 if ((tcp_is_sack(tp
) && !tcp_is_fack(tp
)) ||
2212 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
) ||
2213 (oldcnt
>= packets
))
2216 mss
= tcp_skb_mss(skb
);
2217 /* If needed, chop off the prefix to mark as lost. */
2218 lost
= (packets
- oldcnt
) * mss
;
2219 if (lost
< skb
->len
&&
2220 tcp_fragment(sk
, skb
, lost
, mss
, GFP_ATOMIC
) < 0)
2225 tcp_skb_mark_lost(tp
, skb
);
2230 tcp_verify_left_out(tp
);
2233 /* Account newly detected lost packet(s) */
2235 static void tcp_update_scoreboard(struct sock
*sk
, int fast_rexmit
)
2237 struct tcp_sock
*tp
= tcp_sk(sk
);
2239 if (tcp_is_reno(tp
)) {
2240 tcp_mark_head_lost(sk
, 1, 1);
2241 } else if (tcp_is_fack(tp
)) {
2242 int lost
= tp
->fackets_out
- tp
->reordering
;
2245 tcp_mark_head_lost(sk
, lost
, 0);
2247 int sacked_upto
= tp
->sacked_out
- tp
->reordering
;
2248 if (sacked_upto
>= 0)
2249 tcp_mark_head_lost(sk
, sacked_upto
, 0);
2250 else if (fast_rexmit
)
2251 tcp_mark_head_lost(sk
, 1, 1);
2255 /* CWND moderation, preventing bursts due to too big ACKs
2256 * in dubious situations.
2258 static inline void tcp_moderate_cwnd(struct tcp_sock
*tp
)
2260 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
2261 tcp_packets_in_flight(tp
) + tcp_max_burst(tp
));
2262 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2265 static bool tcp_tsopt_ecr_before(const struct tcp_sock
*tp
, u32 when
)
2267 return tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2268 before(tp
->rx_opt
.rcv_tsecr
, when
);
2271 /* skb is spurious retransmitted if the returned timestamp echo
2272 * reply is prior to the skb transmission time
2274 static bool tcp_skb_spurious_retrans(const struct tcp_sock
*tp
,
2275 const struct sk_buff
*skb
)
2277 return (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
) &&
2278 tcp_tsopt_ecr_before(tp
, tcp_skb_timestamp(skb
));
2281 /* Nothing was retransmitted or returned timestamp is less
2282 * than timestamp of the first retransmission.
2284 static inline bool tcp_packet_delayed(const struct tcp_sock
*tp
)
2286 return !tp
->retrans_stamp
||
2287 tcp_tsopt_ecr_before(tp
, tp
->retrans_stamp
);
2290 /* Undo procedures. */
2292 /* We can clear retrans_stamp when there are no retransmissions in the
2293 * window. It would seem that it is trivially available for us in
2294 * tp->retrans_out, however, that kind of assumptions doesn't consider
2295 * what will happen if errors occur when sending retransmission for the
2296 * second time. ...It could the that such segment has only
2297 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2298 * the head skb is enough except for some reneging corner cases that
2299 * are not worth the effort.
2301 * Main reason for all this complexity is the fact that connection dying
2302 * time now depends on the validity of the retrans_stamp, in particular,
2303 * that successive retransmissions of a segment must not advance
2304 * retrans_stamp under any conditions.
2306 static bool tcp_any_retrans_done(const struct sock
*sk
)
2308 const struct tcp_sock
*tp
= tcp_sk(sk
);
2309 struct sk_buff
*skb
;
2311 if (tp
->retrans_out
)
2314 skb
= tcp_write_queue_head(sk
);
2315 if (unlikely(skb
&& TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
))
2321 #if FASTRETRANS_DEBUG > 1
2322 static void DBGUNDO(struct sock
*sk
, const char *msg
)
2324 struct tcp_sock
*tp
= tcp_sk(sk
);
2325 struct inet_sock
*inet
= inet_sk(sk
);
2327 if (sk
->sk_family
== AF_INET
) {
2328 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2330 &inet
->inet_daddr
, ntohs(inet
->inet_dport
),
2331 tp
->snd_cwnd
, tcp_left_out(tp
),
2332 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2335 #if IS_ENABLED(CONFIG_IPV6)
2336 else if (sk
->sk_family
== AF_INET6
) {
2337 struct ipv6_pinfo
*np
= inet6_sk(sk
);
2338 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2340 &np
->daddr
, ntohs(inet
->inet_dport
),
2341 tp
->snd_cwnd
, tcp_left_out(tp
),
2342 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2348 #define DBGUNDO(x...) do { } while (0)
2351 static void tcp_undo_cwnd_reduction(struct sock
*sk
, bool unmark_loss
)
2353 struct tcp_sock
*tp
= tcp_sk(sk
);
2356 struct sk_buff
*skb
;
2358 tcp_for_write_queue(skb
, sk
) {
2359 if (skb
== tcp_send_head(sk
))
2361 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2364 tcp_clear_all_retrans_hints(tp
);
2367 if (tp
->prior_ssthresh
) {
2368 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2370 if (icsk
->icsk_ca_ops
->undo_cwnd
)
2371 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
2373 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
<< 1);
2375 if (tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
2376 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
2377 tcp_ecn_withdraw_cwr(tp
);
2380 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2381 tp
->undo_marker
= 0;
2384 static inline bool tcp_may_undo(const struct tcp_sock
*tp
)
2386 return tp
->undo_marker
&& (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
2389 /* People celebrate: "We love our President!" */
2390 static bool tcp_try_undo_recovery(struct sock
*sk
)
2392 struct tcp_sock
*tp
= tcp_sk(sk
);
2394 if (tcp_may_undo(tp
)) {
2397 /* Happy end! We did not retransmit anything
2398 * or our original transmission succeeded.
2400 DBGUNDO(sk
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
2401 tcp_undo_cwnd_reduction(sk
, false);
2402 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
2403 mib_idx
= LINUX_MIB_TCPLOSSUNDO
;
2405 mib_idx
= LINUX_MIB_TCPFULLUNDO
;
2407 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2409 if (tp
->snd_una
== tp
->high_seq
&& tcp_is_reno(tp
)) {
2410 /* Hold old state until something *above* high_seq
2411 * is ACKed. For Reno it is MUST to prevent false
2412 * fast retransmits (RFC2582). SACK TCP is safe. */
2413 tcp_moderate_cwnd(tp
);
2414 if (!tcp_any_retrans_done(sk
))
2415 tp
->retrans_stamp
= 0;
2418 tcp_set_ca_state(sk
, TCP_CA_Open
);
2422 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2423 static bool tcp_try_undo_dsack(struct sock
*sk
)
2425 struct tcp_sock
*tp
= tcp_sk(sk
);
2427 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
2428 DBGUNDO(sk
, "D-SACK");
2429 tcp_undo_cwnd_reduction(sk
, false);
2430 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPDSACKUNDO
);
2436 /* Undo during loss recovery after partial ACK or using F-RTO. */
2437 static bool tcp_try_undo_loss(struct sock
*sk
, bool frto_undo
)
2439 struct tcp_sock
*tp
= tcp_sk(sk
);
2441 if (frto_undo
|| tcp_may_undo(tp
)) {
2442 tcp_undo_cwnd_reduction(sk
, true);
2444 DBGUNDO(sk
, "partial loss");
2445 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSSUNDO
);
2447 NET_INC_STATS_BH(sock_net(sk
),
2448 LINUX_MIB_TCPSPURIOUSRTOS
);
2449 inet_csk(sk
)->icsk_retransmits
= 0;
2450 if (frto_undo
|| tcp_is_sack(tp
))
2451 tcp_set_ca_state(sk
, TCP_CA_Open
);
2457 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2458 * It computes the number of packets to send (sndcnt) based on packets newly
2460 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2461 * cwnd reductions across a full RTT.
2462 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2463 * But when the retransmits are acked without further losses, PRR
2464 * slow starts cwnd up to ssthresh to speed up the recovery.
2466 static void tcp_init_cwnd_reduction(struct sock
*sk
)
2468 struct tcp_sock
*tp
= tcp_sk(sk
);
2470 tp
->high_seq
= tp
->snd_nxt
;
2471 tp
->tlp_high_seq
= 0;
2472 tp
->snd_cwnd_cnt
= 0;
2473 tp
->prior_cwnd
= tp
->snd_cwnd
;
2474 tp
->prr_delivered
= 0;
2476 tp
->snd_ssthresh
= inet_csk(sk
)->icsk_ca_ops
->ssthresh(sk
);
2477 tcp_ecn_queue_cwr(tp
);
2480 static void tcp_cwnd_reduction(struct sock
*sk
, int newly_acked_sacked
,
2483 struct tcp_sock
*tp
= tcp_sk(sk
);
2485 int delta
= tp
->snd_ssthresh
- tcp_packets_in_flight(tp
);
2487 if (newly_acked_sacked
<= 0 || WARN_ON_ONCE(!tp
->prior_cwnd
))
2490 tp
->prr_delivered
+= newly_acked_sacked
;
2492 u64 dividend
= (u64
)tp
->snd_ssthresh
* tp
->prr_delivered
+
2494 sndcnt
= div_u64(dividend
, tp
->prior_cwnd
) - tp
->prr_out
;
2495 } else if ((flag
& FLAG_RETRANS_DATA_ACKED
) &&
2496 !(flag
& FLAG_LOST_RETRANS
)) {
2497 sndcnt
= min_t(int, delta
,
2498 max_t(int, tp
->prr_delivered
- tp
->prr_out
,
2499 newly_acked_sacked
) + 1);
2501 sndcnt
= min(delta
, newly_acked_sacked
);
2503 /* Force a fast retransmit upon entering fast recovery */
2504 sndcnt
= max(sndcnt
, (tp
->prr_out
? 0 : 1));
2505 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + sndcnt
;
2508 static inline void tcp_end_cwnd_reduction(struct sock
*sk
)
2510 struct tcp_sock
*tp
= tcp_sk(sk
);
2512 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2513 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_CWR
||
2514 (tp
->undo_marker
&& tp
->snd_ssthresh
< TCP_INFINITE_SSTHRESH
)) {
2515 tp
->snd_cwnd
= tp
->snd_ssthresh
;
2516 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2518 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
2521 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2522 void tcp_enter_cwr(struct sock
*sk
)
2524 struct tcp_sock
*tp
= tcp_sk(sk
);
2526 tp
->prior_ssthresh
= 0;
2527 if (inet_csk(sk
)->icsk_ca_state
< TCP_CA_CWR
) {
2528 tp
->undo_marker
= 0;
2529 tcp_init_cwnd_reduction(sk
);
2530 tcp_set_ca_state(sk
, TCP_CA_CWR
);
2533 EXPORT_SYMBOL(tcp_enter_cwr
);
2535 static void tcp_try_keep_open(struct sock
*sk
)
2537 struct tcp_sock
*tp
= tcp_sk(sk
);
2538 int state
= TCP_CA_Open
;
2540 if (tcp_left_out(tp
) || tcp_any_retrans_done(sk
))
2541 state
= TCP_CA_Disorder
;
2543 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
2544 tcp_set_ca_state(sk
, state
);
2545 tp
->high_seq
= tp
->snd_nxt
;
2549 static void tcp_try_to_open(struct sock
*sk
, int flag
)
2551 struct tcp_sock
*tp
= tcp_sk(sk
);
2553 tcp_verify_left_out(tp
);
2555 if (!tcp_any_retrans_done(sk
))
2556 tp
->retrans_stamp
= 0;
2558 if (flag
& FLAG_ECE
)
2561 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
2562 tcp_try_keep_open(sk
);
2566 static void tcp_mtup_probe_failed(struct sock
*sk
)
2568 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2570 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
2571 icsk
->icsk_mtup
.probe_size
= 0;
2572 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPMTUPFAIL
);
2575 static void tcp_mtup_probe_success(struct sock
*sk
)
2577 struct tcp_sock
*tp
= tcp_sk(sk
);
2578 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2580 /* FIXME: breaks with very large cwnd */
2581 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2582 tp
->snd_cwnd
= tp
->snd_cwnd
*
2583 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
2584 icsk
->icsk_mtup
.probe_size
;
2585 tp
->snd_cwnd_cnt
= 0;
2586 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2587 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2589 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
2590 icsk
->icsk_mtup
.probe_size
= 0;
2591 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
2592 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPMTUPSUCCESS
);
2595 /* Do a simple retransmit without using the backoff mechanisms in
2596 * tcp_timer. This is used for path mtu discovery.
2597 * The socket is already locked here.
2599 void tcp_simple_retransmit(struct sock
*sk
)
2601 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2602 struct tcp_sock
*tp
= tcp_sk(sk
);
2603 struct sk_buff
*skb
;
2604 unsigned int mss
= tcp_current_mss(sk
);
2605 u32 prior_lost
= tp
->lost_out
;
2607 tcp_for_write_queue(skb
, sk
) {
2608 if (skb
== tcp_send_head(sk
))
2610 if (tcp_skb_seglen(skb
) > mss
&&
2611 !(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
2612 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2613 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
2614 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2616 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
2620 tcp_clear_retrans_hints_partial(tp
);
2622 if (prior_lost
== tp
->lost_out
)
2625 if (tcp_is_reno(tp
))
2626 tcp_limit_reno_sacked(tp
);
2628 tcp_verify_left_out(tp
);
2630 /* Don't muck with the congestion window here.
2631 * Reason is that we do not increase amount of _data_
2632 * in network, but units changed and effective
2633 * cwnd/ssthresh really reduced now.
2635 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
2636 tp
->high_seq
= tp
->snd_nxt
;
2637 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2638 tp
->prior_ssthresh
= 0;
2639 tp
->undo_marker
= 0;
2640 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2642 tcp_xmit_retransmit_queue(sk
);
2644 EXPORT_SYMBOL(tcp_simple_retransmit
);
2646 static void tcp_enter_recovery(struct sock
*sk
, bool ece_ack
)
2648 struct tcp_sock
*tp
= tcp_sk(sk
);
2651 if (tcp_is_reno(tp
))
2652 mib_idx
= LINUX_MIB_TCPRENORECOVERY
;
2654 mib_idx
= LINUX_MIB_TCPSACKRECOVERY
;
2656 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2658 tp
->prior_ssthresh
= 0;
2661 if (!tcp_in_cwnd_reduction(sk
)) {
2663 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2664 tcp_init_cwnd_reduction(sk
);
2666 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
2669 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2670 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2672 static void tcp_process_loss(struct sock
*sk
, int flag
, bool is_dupack
,
2675 struct tcp_sock
*tp
= tcp_sk(sk
);
2676 bool recovered
= !before(tp
->snd_una
, tp
->high_seq
);
2678 if ((flag
& FLAG_SND_UNA_ADVANCED
) &&
2679 tcp_try_undo_loss(sk
, false))
2682 if (tp
->frto
) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2683 /* Step 3.b. A timeout is spurious if not all data are
2684 * lost, i.e., never-retransmitted data are (s)acked.
2686 if ((flag
& FLAG_ORIG_SACK_ACKED
) &&
2687 tcp_try_undo_loss(sk
, true))
2690 if (after(tp
->snd_nxt
, tp
->high_seq
)) {
2691 if (flag
& FLAG_DATA_SACKED
|| is_dupack
)
2692 tp
->frto
= 0; /* Step 3.a. loss was real */
2693 } else if (flag
& FLAG_SND_UNA_ADVANCED
&& !recovered
) {
2694 tp
->high_seq
= tp
->snd_nxt
;
2695 /* Step 2.b. Try send new data (but deferred until cwnd
2696 * is updated in tcp_ack()). Otherwise fall back to
2697 * the conventional recovery.
2699 if (tcp_send_head(sk
) &&
2700 after(tcp_wnd_end(tp
), tp
->snd_nxt
)) {
2701 *rexmit
= REXMIT_NEW
;
2709 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2710 tcp_try_undo_recovery(sk
);
2713 if (tcp_is_reno(tp
)) {
2714 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2715 * delivered. Lower inflight to clock out (re)tranmissions.
2717 if (after(tp
->snd_nxt
, tp
->high_seq
) && is_dupack
)
2718 tcp_add_reno_sack(sk
);
2719 else if (flag
& FLAG_SND_UNA_ADVANCED
)
2720 tcp_reset_reno_sack(tp
);
2722 *rexmit
= REXMIT_LOST
;
2725 /* Undo during fast recovery after partial ACK. */
2726 static bool tcp_try_undo_partial(struct sock
*sk
, const int acked
)
2728 struct tcp_sock
*tp
= tcp_sk(sk
);
2730 if (tp
->undo_marker
&& tcp_packet_delayed(tp
)) {
2731 /* Plain luck! Hole if filled with delayed
2732 * packet, rather than with a retransmit.
2734 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
2736 /* We are getting evidence that the reordering degree is higher
2737 * than we realized. If there are no retransmits out then we
2738 * can undo. Otherwise we clock out new packets but do not
2739 * mark more packets lost or retransmit more.
2741 if (tp
->retrans_out
)
2744 if (!tcp_any_retrans_done(sk
))
2745 tp
->retrans_stamp
= 0;
2747 DBGUNDO(sk
, "partial recovery");
2748 tcp_undo_cwnd_reduction(sk
, true);
2749 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPARTIALUNDO
);
2750 tcp_try_keep_open(sk
);
2756 /* Process an event, which can update packets-in-flight not trivially.
2757 * Main goal of this function is to calculate new estimate for left_out,
2758 * taking into account both packets sitting in receiver's buffer and
2759 * packets lost by network.
2761 * Besides that it updates the congestion state when packet loss or ECN
2762 * is detected. But it does not reduce the cwnd, it is done by the
2763 * congestion control later.
2765 * It does _not_ decide what to send, it is made in function
2766 * tcp_xmit_retransmit_queue().
2768 static void tcp_fastretrans_alert(struct sock
*sk
, const int acked
,
2769 bool is_dupack
, int *ack_flag
, int *rexmit
)
2771 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2772 struct tcp_sock
*tp
= tcp_sk(sk
);
2773 int fast_rexmit
= 0, flag
= *ack_flag
;
2774 bool do_lost
= is_dupack
|| ((flag
& FLAG_DATA_SACKED
) &&
2775 (tcp_fackets_out(tp
) > tp
->reordering
));
2777 if (WARN_ON(!tp
->packets_out
&& tp
->sacked_out
))
2779 if (WARN_ON(!tp
->sacked_out
&& tp
->fackets_out
))
2780 tp
->fackets_out
= 0;
2782 /* Now state machine starts.
2783 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2784 if (flag
& FLAG_ECE
)
2785 tp
->prior_ssthresh
= 0;
2787 /* B. In all the states check for reneging SACKs. */
2788 if (tcp_check_sack_reneging(sk
, flag
))
2791 /* C. Check consistency of the current state. */
2792 tcp_verify_left_out(tp
);
2794 /* D. Check state exit conditions. State can be terminated
2795 * when high_seq is ACKed. */
2796 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
2797 WARN_ON(tp
->retrans_out
!= 0);
2798 tp
->retrans_stamp
= 0;
2799 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
2800 switch (icsk
->icsk_ca_state
) {
2802 /* CWR is to be held something *above* high_seq
2803 * is ACKed for CWR bit to reach receiver. */
2804 if (tp
->snd_una
!= tp
->high_seq
) {
2805 tcp_end_cwnd_reduction(sk
);
2806 tcp_set_ca_state(sk
, TCP_CA_Open
);
2810 case TCP_CA_Recovery
:
2811 if (tcp_is_reno(tp
))
2812 tcp_reset_reno_sack(tp
);
2813 if (tcp_try_undo_recovery(sk
))
2815 tcp_end_cwnd_reduction(sk
);
2820 /* Use RACK to detect loss */
2821 if (sysctl_tcp_recovery
& TCP_RACK_LOST_RETRANS
&&
2822 tcp_rack_mark_lost(sk
)) {
2823 flag
|= FLAG_LOST_RETRANS
;
2824 *ack_flag
|= FLAG_LOST_RETRANS
;
2827 /* E. Process state. */
2828 switch (icsk
->icsk_ca_state
) {
2829 case TCP_CA_Recovery
:
2830 if (!(flag
& FLAG_SND_UNA_ADVANCED
)) {
2831 if (tcp_is_reno(tp
) && is_dupack
)
2832 tcp_add_reno_sack(sk
);
2834 if (tcp_try_undo_partial(sk
, acked
))
2836 /* Partial ACK arrived. Force fast retransmit. */
2837 do_lost
= tcp_is_reno(tp
) ||
2838 tcp_fackets_out(tp
) > tp
->reordering
;
2840 if (tcp_try_undo_dsack(sk
)) {
2841 tcp_try_keep_open(sk
);
2846 tcp_process_loss(sk
, flag
, is_dupack
, rexmit
);
2847 if (icsk
->icsk_ca_state
!= TCP_CA_Open
&&
2848 !(flag
& FLAG_LOST_RETRANS
))
2850 /* Change state if cwnd is undone or retransmits are lost */
2852 if (tcp_is_reno(tp
)) {
2853 if (flag
& FLAG_SND_UNA_ADVANCED
)
2854 tcp_reset_reno_sack(tp
);
2856 tcp_add_reno_sack(sk
);
2859 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
)
2860 tcp_try_undo_dsack(sk
);
2862 if (!tcp_time_to_recover(sk
, flag
)) {
2863 tcp_try_to_open(sk
, flag
);
2867 /* MTU probe failure: don't reduce cwnd */
2868 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
2869 icsk
->icsk_mtup
.probe_size
&&
2870 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
2871 tcp_mtup_probe_failed(sk
);
2872 /* Restores the reduction we did in tcp_mtup_probe() */
2874 tcp_simple_retransmit(sk
);
2878 /* Otherwise enter Recovery state */
2879 tcp_enter_recovery(sk
, (flag
& FLAG_ECE
));
2884 tcp_update_scoreboard(sk
, fast_rexmit
);
2885 *rexmit
= REXMIT_LOST
;
2888 /* Kathleen Nichols' algorithm for tracking the minimum value of
2889 * a data stream over some fixed time interval. (E.g., the minimum
2890 * RTT over the past five minutes.) It uses constant space and constant
2891 * time per update yet almost always delivers the same minimum as an
2892 * implementation that has to keep all the data in the window.
2894 * The algorithm keeps track of the best, 2nd best & 3rd best min
2895 * values, maintaining an invariant that the measurement time of the
2896 * n'th best >= n-1'th best. It also makes sure that the three values
2897 * are widely separated in the time window since that bounds the worse
2898 * case error when that data is monotonically increasing over the window.
2900 * Upon getting a new min, we can forget everything earlier because it
2901 * has no value - the new min is <= everything else in the window by
2902 * definition and it's the most recent. So we restart fresh on every new min
2903 * and overwrites 2nd & 3rd choices. The same property holds for 2nd & 3rd
2906 static void tcp_update_rtt_min(struct sock
*sk
, u32 rtt_us
)
2908 const u32 now
= tcp_time_stamp
, wlen
= sysctl_tcp_min_rtt_wlen
* HZ
;
2909 struct rtt_meas
*m
= tcp_sk(sk
)->rtt_min
;
2910 struct rtt_meas rttm
= { .rtt
= (rtt_us
? : 1), .ts
= now
};
2913 /* Check if the new measurement updates the 1st, 2nd, or 3rd choices */
2914 if (unlikely(rttm
.rtt
<= m
[0].rtt
))
2915 m
[0] = m
[1] = m
[2] = rttm
;
2916 else if (rttm
.rtt
<= m
[1].rtt
)
2918 else if (rttm
.rtt
<= m
[2].rtt
)
2921 elapsed
= now
- m
[0].ts
;
2922 if (unlikely(elapsed
> wlen
)) {
2923 /* Passed entire window without a new min so make 2nd choice
2924 * the new min & 3rd choice the new 2nd. So forth and so on.
2929 if (now
- m
[0].ts
> wlen
) {
2932 if (now
- m
[0].ts
> wlen
)
2935 } else if (m
[1].ts
== m
[0].ts
&& elapsed
> wlen
/ 4) {
2936 /* Passed a quarter of the window without a new min so
2937 * take 2nd choice from the 2nd quarter of the window.
2940 } else if (m
[2].ts
== m
[1].ts
&& elapsed
> wlen
/ 2) {
2941 /* Passed half the window without a new min so take the 3rd
2942 * choice from the last half of the window.
2948 static inline bool tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
2949 long seq_rtt_us
, long sack_rtt_us
,
2952 const struct tcp_sock
*tp
= tcp_sk(sk
);
2954 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2955 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2956 * Karn's algorithm forbids taking RTT if some retransmitted data
2957 * is acked (RFC6298).
2960 seq_rtt_us
= sack_rtt_us
;
2962 /* RTTM Rule: A TSecr value received in a segment is used to
2963 * update the averaged RTT measurement only if the segment
2964 * acknowledges some new data, i.e., only if it advances the
2965 * left edge of the send window.
2966 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2968 if (seq_rtt_us
< 0 && tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2970 seq_rtt_us
= ca_rtt_us
= jiffies_to_usecs(tcp_time_stamp
-
2971 tp
->rx_opt
.rcv_tsecr
);
2975 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
2976 * always taken together with ACK, SACK, or TS-opts. Any negative
2977 * values will be skipped with the seq_rtt_us < 0 check above.
2979 tcp_update_rtt_min(sk
, ca_rtt_us
);
2980 tcp_rtt_estimator(sk
, seq_rtt_us
);
2983 /* RFC6298: only reset backoff on valid RTT measurement. */
2984 inet_csk(sk
)->icsk_backoff
= 0;
2988 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2989 void tcp_synack_rtt_meas(struct sock
*sk
, struct request_sock
*req
)
2993 if (req
&& !req
->num_retrans
&& tcp_rsk(req
)->snt_synack
.v64
) {
2994 struct skb_mstamp now
;
2996 skb_mstamp_get(&now
);
2997 rtt_us
= skb_mstamp_us_delta(&now
, &tcp_rsk(req
)->snt_synack
);
3000 tcp_ack_update_rtt(sk
, FLAG_SYN_ACKED
, rtt_us
, -1L, rtt_us
);
3004 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
, u32 acked
)
3006 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3008 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, acked
);
3009 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_time_stamp
;
3012 /* Restart timer after forward progress on connection.
3013 * RFC2988 recommends to restart timer to now+rto.
3015 void tcp_rearm_rto(struct sock
*sk
)
3017 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3018 struct tcp_sock
*tp
= tcp_sk(sk
);
3020 /* If the retrans timer is currently being used by Fast Open
3021 * for SYN-ACK retrans purpose, stay put.
3023 if (tp
->fastopen_rsk
)
3026 if (!tp
->packets_out
) {
3027 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
3029 u32 rto
= inet_csk(sk
)->icsk_rto
;
3030 /* Offset the time elapsed after installing regular RTO */
3031 if (icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
||
3032 icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
) {
3033 struct sk_buff
*skb
= tcp_write_queue_head(sk
);
3034 const u32 rto_time_stamp
=
3035 tcp_skb_timestamp(skb
) + rto
;
3036 s32 delta
= (s32
)(rto_time_stamp
- tcp_time_stamp
);
3037 /* delta may not be positive if the socket is locked
3038 * when the retrans timer fires and is rescheduled.
3043 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
, rto
,
3048 /* This function is called when the delayed ER timer fires. TCP enters
3049 * fast recovery and performs fast-retransmit.
3051 void tcp_resume_early_retransmit(struct sock
*sk
)
3053 struct tcp_sock
*tp
= tcp_sk(sk
);
3057 /* Stop if ER is disabled after the delayed ER timer is scheduled */
3058 if (!tp
->do_early_retrans
)
3061 tcp_enter_recovery(sk
, false);
3062 tcp_update_scoreboard(sk
, 1);
3063 tcp_xmit_retransmit_queue(sk
);
3066 /* If we get here, the whole TSO packet has not been acked. */
3067 static u32
tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
)
3069 struct tcp_sock
*tp
= tcp_sk(sk
);
3072 BUG_ON(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
));
3074 packets_acked
= tcp_skb_pcount(skb
);
3075 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
3077 packets_acked
-= tcp_skb_pcount(skb
);
3079 if (packets_acked
) {
3080 BUG_ON(tcp_skb_pcount(skb
) == 0);
3081 BUG_ON(!before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
));
3084 return packets_acked
;
3087 static void tcp_ack_tstamp(struct sock
*sk
, struct sk_buff
*skb
,
3090 const struct skb_shared_info
*shinfo
;
3092 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3093 if (likely(!(sk
->sk_tsflags
& SOF_TIMESTAMPING_TX_ACK
)))
3096 shinfo
= skb_shinfo(skb
);
3097 if ((shinfo
->tx_flags
& SKBTX_ACK_TSTAMP
) &&
3098 between(shinfo
->tskey
, prior_snd_una
, tcp_sk(sk
)->snd_una
- 1))
3099 __skb_tstamp_tx(skb
, NULL
, sk
, SCM_TSTAMP_ACK
);
3102 /* Remove acknowledged frames from the retransmission queue. If our packet
3103 * is before the ack sequence we can discard it as it's confirmed to have
3104 * arrived at the other end.
3106 static int tcp_clean_rtx_queue(struct sock
*sk
, int prior_fackets
,
3107 u32 prior_snd_una
, int *acked
,
3108 struct tcp_sacktag_state
*sack
)
3110 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3111 struct skb_mstamp first_ackt
, last_ackt
, now
;
3112 struct tcp_sock
*tp
= tcp_sk(sk
);
3113 u32 prior_sacked
= tp
->sacked_out
;
3114 u32 reord
= tp
->packets_out
;
3115 bool fully_acked
= true;
3116 long sack_rtt_us
= -1L;
3117 long seq_rtt_us
= -1L;
3118 long ca_rtt_us
= -1L;
3119 struct sk_buff
*skb
;
3126 while ((skb
= tcp_write_queue_head(sk
)) && skb
!= tcp_send_head(sk
)) {
3127 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
3128 u8 sacked
= scb
->sacked
;
3131 tcp_ack_tstamp(sk
, skb
, prior_snd_una
);
3133 /* Determine how many packets and what bytes were acked, tso and else */
3134 if (after(scb
->end_seq
, tp
->snd_una
)) {
3135 if (tcp_skb_pcount(skb
) == 1 ||
3136 !after(tp
->snd_una
, scb
->seq
))
3139 acked_pcount
= tcp_tso_acked(sk
, skb
);
3143 fully_acked
= false;
3145 /* Speedup tcp_unlink_write_queue() and next loop */
3146 prefetchw(skb
->next
);
3147 acked_pcount
= tcp_skb_pcount(skb
);
3150 if (unlikely(sacked
& TCPCB_RETRANS
)) {
3151 if (sacked
& TCPCB_SACKED_RETRANS
)
3152 tp
->retrans_out
-= acked_pcount
;
3153 flag
|= FLAG_RETRANS_DATA_ACKED
;
3154 } else if (!(sacked
& TCPCB_SACKED_ACKED
)) {
3155 last_ackt
= skb
->skb_mstamp
;
3156 WARN_ON_ONCE(last_ackt
.v64
== 0);
3157 if (!first_ackt
.v64
)
3158 first_ackt
= last_ackt
;
3160 reord
= min(pkts_acked
, reord
);
3161 if (!after(scb
->end_seq
, tp
->high_seq
))
3162 flag
|= FLAG_ORIG_SACK_ACKED
;
3165 if (sacked
& TCPCB_SACKED_ACKED
) {
3166 tp
->sacked_out
-= acked_pcount
;
3167 } else if (tcp_is_sack(tp
)) {
3168 tp
->delivered
+= acked_pcount
;
3169 if (!tcp_skb_spurious_retrans(tp
, skb
))
3170 tcp_rack_advance(tp
, &skb
->skb_mstamp
, sacked
);
3172 if (sacked
& TCPCB_LOST
)
3173 tp
->lost_out
-= acked_pcount
;
3175 tp
->packets_out
-= acked_pcount
;
3176 pkts_acked
+= acked_pcount
;
3178 /* Initial outgoing SYN's get put onto the write_queue
3179 * just like anything else we transmit. It is not
3180 * true data, and if we misinform our callers that
3181 * this ACK acks real data, we will erroneously exit
3182 * connection startup slow start one packet too
3183 * quickly. This is severely frowned upon behavior.
3185 if (likely(!(scb
->tcp_flags
& TCPHDR_SYN
))) {
3186 flag
|= FLAG_DATA_ACKED
;
3188 flag
|= FLAG_SYN_ACKED
;
3189 tp
->retrans_stamp
= 0;
3195 tcp_unlink_write_queue(skb
, sk
);
3196 sk_wmem_free_skb(sk
, skb
);
3197 if (unlikely(skb
== tp
->retransmit_skb_hint
))
3198 tp
->retransmit_skb_hint
= NULL
;
3199 if (unlikely(skb
== tp
->lost_skb_hint
))
3200 tp
->lost_skb_hint
= NULL
;
3203 if (likely(between(tp
->snd_up
, prior_snd_una
, tp
->snd_una
)))
3204 tp
->snd_up
= tp
->snd_una
;
3206 if (skb
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
3207 flag
|= FLAG_SACK_RENEGING
;
3209 skb_mstamp_get(&now
);
3210 if (likely(first_ackt
.v64
) && !(flag
& FLAG_RETRANS_DATA_ACKED
)) {
3211 seq_rtt_us
= skb_mstamp_us_delta(&now
, &first_ackt
);
3212 ca_rtt_us
= skb_mstamp_us_delta(&now
, &last_ackt
);
3214 if (sack
->first_sackt
.v64
) {
3215 sack_rtt_us
= skb_mstamp_us_delta(&now
, &sack
->first_sackt
);
3216 ca_rtt_us
= skb_mstamp_us_delta(&now
, &sack
->last_sackt
);
3219 rtt_update
= tcp_ack_update_rtt(sk
, flag
, seq_rtt_us
, sack_rtt_us
,
3222 if (flag
& FLAG_ACKED
) {
3224 if (unlikely(icsk
->icsk_mtup
.probe_size
&&
3225 !after(tp
->mtu_probe
.probe_seq_end
, tp
->snd_una
))) {
3226 tcp_mtup_probe_success(sk
);
3229 if (tcp_is_reno(tp
)) {
3230 tcp_remove_reno_sacks(sk
, pkts_acked
);
3234 /* Non-retransmitted hole got filled? That's reordering */
3235 if (reord
< prior_fackets
)
3236 tcp_update_reordering(sk
, tp
->fackets_out
- reord
, 0);
3238 delta
= tcp_is_fack(tp
) ? pkts_acked
:
3239 prior_sacked
- tp
->sacked_out
;
3240 tp
->lost_cnt_hint
-= min(tp
->lost_cnt_hint
, delta
);
3243 tp
->fackets_out
-= min(pkts_acked
, tp
->fackets_out
);
3245 } else if (skb
&& rtt_update
&& sack_rtt_us
>= 0 &&
3246 sack_rtt_us
> skb_mstamp_us_delta(&now
, &skb
->skb_mstamp
)) {
3247 /* Do not re-arm RTO if the sack RTT is measured from data sent
3248 * after when the head was last (re)transmitted. Otherwise the
3249 * timeout may continue to extend in loss recovery.
3254 if (icsk
->icsk_ca_ops
->pkts_acked
)
3255 icsk
->icsk_ca_ops
->pkts_acked(sk
, pkts_acked
, ca_rtt_us
);
3257 #if FASTRETRANS_DEBUG > 0
3258 WARN_ON((int)tp
->sacked_out
< 0);
3259 WARN_ON((int)tp
->lost_out
< 0);
3260 WARN_ON((int)tp
->retrans_out
< 0);
3261 if (!tp
->packets_out
&& tcp_is_sack(tp
)) {
3262 icsk
= inet_csk(sk
);
3264 pr_debug("Leak l=%u %d\n",
3265 tp
->lost_out
, icsk
->icsk_ca_state
);
3268 if (tp
->sacked_out
) {
3269 pr_debug("Leak s=%u %d\n",
3270 tp
->sacked_out
, icsk
->icsk_ca_state
);
3273 if (tp
->retrans_out
) {
3274 pr_debug("Leak r=%u %d\n",
3275 tp
->retrans_out
, icsk
->icsk_ca_state
);
3276 tp
->retrans_out
= 0;
3280 *acked
= pkts_acked
;
3284 static void tcp_ack_probe(struct sock
*sk
)
3286 const struct tcp_sock
*tp
= tcp_sk(sk
);
3287 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3289 /* Was it a usable window open? */
3291 if (!after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
, tcp_wnd_end(tp
))) {
3292 icsk
->icsk_backoff
= 0;
3293 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
3294 /* Socket must be waked up by subsequent tcp_data_snd_check().
3295 * This function is not for random using!
3298 unsigned long when
= tcp_probe0_when(sk
, TCP_RTO_MAX
);
3300 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3305 static inline bool tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
3307 return !(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
3308 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
;
3311 /* Decide wheather to run the increase function of congestion control. */
3312 static inline bool tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
3314 /* If reordering is high then always grow cwnd whenever data is
3315 * delivered regardless of its ordering. Otherwise stay conservative
3316 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3317 * new SACK or ECE mark may first advance cwnd here and later reduce
3318 * cwnd in tcp_fastretrans_alert() based on more states.
3320 if (tcp_sk(sk
)->reordering
> sysctl_tcp_reordering
)
3321 return flag
& FLAG_FORWARD_PROGRESS
;
3323 return flag
& FLAG_DATA_ACKED
;
3326 /* The "ultimate" congestion control function that aims to replace the rigid
3327 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3328 * It's called toward the end of processing an ACK with precise rate
3329 * information. All transmission or retransmission are delayed afterwards.
3331 static void tcp_cong_control(struct sock
*sk
, u32 ack
, u32 acked_sacked
,
3334 if (tcp_in_cwnd_reduction(sk
)) {
3335 /* Reduce cwnd if state mandates */
3336 tcp_cwnd_reduction(sk
, acked_sacked
, flag
);
3337 } else if (tcp_may_raise_cwnd(sk
, flag
)) {
3338 /* Advance cwnd if state allows */
3339 tcp_cong_avoid(sk
, ack
, acked_sacked
);
3341 tcp_update_pacing_rate(sk
);
3344 /* Check that window update is acceptable.
3345 * The function assumes that snd_una<=ack<=snd_next.
3347 static inline bool tcp_may_update_window(const struct tcp_sock
*tp
,
3348 const u32 ack
, const u32 ack_seq
,
3351 return after(ack
, tp
->snd_una
) ||
3352 after(ack_seq
, tp
->snd_wl1
) ||
3353 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
);
3356 /* If we update tp->snd_una, also update tp->bytes_acked */
3357 static void tcp_snd_una_update(struct tcp_sock
*tp
, u32 ack
)
3359 u32 delta
= ack
- tp
->snd_una
;
3361 u64_stats_update_begin(&tp
->syncp
);
3362 tp
->bytes_acked
+= delta
;
3363 u64_stats_update_end(&tp
->syncp
);
3367 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3368 static void tcp_rcv_nxt_update(struct tcp_sock
*tp
, u32 seq
)
3370 u32 delta
= seq
- tp
->rcv_nxt
;
3372 u64_stats_update_begin(&tp
->syncp
);
3373 tp
->bytes_received
+= delta
;
3374 u64_stats_update_end(&tp
->syncp
);
3378 /* Update our send window.
3380 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3381 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3383 static int tcp_ack_update_window(struct sock
*sk
, const struct sk_buff
*skb
, u32 ack
,
3386 struct tcp_sock
*tp
= tcp_sk(sk
);
3388 u32 nwin
= ntohs(tcp_hdr(skb
)->window
);
3390 if (likely(!tcp_hdr(skb
)->syn
))
3391 nwin
<<= tp
->rx_opt
.snd_wscale
;
3393 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
3394 flag
|= FLAG_WIN_UPDATE
;
3395 tcp_update_wl(tp
, ack_seq
);
3397 if (tp
->snd_wnd
!= nwin
) {
3400 /* Note, it is the only place, where
3401 * fast path is recovered for sending TCP.
3404 tcp_fast_path_check(sk
);
3406 if (tcp_send_head(sk
))
3407 tcp_slow_start_after_idle_check(sk
);
3409 if (nwin
> tp
->max_window
) {
3410 tp
->max_window
= nwin
;
3411 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
3416 tcp_snd_una_update(tp
, ack
);
3421 /* Return true if we're currently rate-limiting out-of-window ACKs and
3422 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3423 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3424 * attacks that send repeated SYNs or ACKs for the same connection. To
3425 * do this, we do not send a duplicate SYNACK or ACK if the remote
3426 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3428 bool tcp_oow_rate_limited(struct net
*net
, const struct sk_buff
*skb
,
3429 int mib_idx
, u32
*last_oow_ack_time
)
3431 /* Data packets without SYNs are not likely part of an ACK loop. */
3432 if ((TCP_SKB_CB(skb
)->seq
!= TCP_SKB_CB(skb
)->end_seq
) &&
3434 goto not_rate_limited
;
3436 if (*last_oow_ack_time
) {
3437 s32 elapsed
= (s32
)(tcp_time_stamp
- *last_oow_ack_time
);
3439 if (0 <= elapsed
&& elapsed
< sysctl_tcp_invalid_ratelimit
) {
3440 NET_INC_STATS_BH(net
, mib_idx
);
3441 return true; /* rate-limited: don't send yet! */
3445 *last_oow_ack_time
= tcp_time_stamp
;
3448 return false; /* not rate-limited: go ahead, send dupack now! */
3451 /* RFC 5961 7 [ACK Throttling] */
3452 static void tcp_send_challenge_ack(struct sock
*sk
, const struct sk_buff
*skb
)
3454 /* unprotected vars, we dont care of overwrites */
3455 static u32 challenge_timestamp
;
3456 static unsigned int challenge_count
;
3457 struct tcp_sock
*tp
= tcp_sk(sk
);
3460 /* First check our per-socket dupack rate limit. */
3461 if (tcp_oow_rate_limited(sock_net(sk
), skb
,
3462 LINUX_MIB_TCPACKSKIPPEDCHALLENGE
,
3463 &tp
->last_oow_ack_time
))
3466 /* Then check the check host-wide RFC 5961 rate limit. */
3468 if (now
!= challenge_timestamp
) {
3469 challenge_timestamp
= now
;
3470 challenge_count
= 0;
3472 if (++challenge_count
<= sysctl_tcp_challenge_ack_limit
) {
3473 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPCHALLENGEACK
);
3478 static void tcp_store_ts_recent(struct tcp_sock
*tp
)
3480 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
3481 tp
->rx_opt
.ts_recent_stamp
= get_seconds();
3484 static void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
3486 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
3487 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3488 * extra check below makes sure this can only happen
3489 * for pure ACK frames. -DaveM
3491 * Not only, also it occurs for expired timestamps.
3494 if (tcp_paws_check(&tp
->rx_opt
, 0))
3495 tcp_store_ts_recent(tp
);
3499 /* This routine deals with acks during a TLP episode.
3500 * We mark the end of a TLP episode on receiving TLP dupack or when
3501 * ack is after tlp_high_seq.
3502 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3504 static void tcp_process_tlp_ack(struct sock
*sk
, u32 ack
, int flag
)
3506 struct tcp_sock
*tp
= tcp_sk(sk
);
3508 if (before(ack
, tp
->tlp_high_seq
))
3511 if (flag
& FLAG_DSACKING_ACK
) {
3512 /* This DSACK means original and TLP probe arrived; no loss */
3513 tp
->tlp_high_seq
= 0;
3514 } else if (after(ack
, tp
->tlp_high_seq
)) {
3515 /* ACK advances: there was a loss, so reduce cwnd. Reset
3516 * tlp_high_seq in tcp_init_cwnd_reduction()
3518 tcp_init_cwnd_reduction(sk
);
3519 tcp_set_ca_state(sk
, TCP_CA_CWR
);
3520 tcp_end_cwnd_reduction(sk
);
3521 tcp_try_keep_open(sk
);
3522 NET_INC_STATS_BH(sock_net(sk
),
3523 LINUX_MIB_TCPLOSSPROBERECOVERY
);
3524 } else if (!(flag
& (FLAG_SND_UNA_ADVANCED
|
3525 FLAG_NOT_DUP
| FLAG_DATA_SACKED
))) {
3526 /* Pure dupack: original and TLP probe arrived; no loss */
3527 tp
->tlp_high_seq
= 0;
3531 static inline void tcp_in_ack_event(struct sock
*sk
, u32 flags
)
3533 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3535 if (icsk
->icsk_ca_ops
->in_ack_event
)
3536 icsk
->icsk_ca_ops
->in_ack_event(sk
, flags
);
3539 /* Congestion control has updated the cwnd already. So if we're in
3540 * loss recovery then now we do any new sends (for FRTO) or
3541 * retransmits (for CA_Loss or CA_recovery) that make sense.
3543 static void tcp_xmit_recovery(struct sock
*sk
, int rexmit
)
3545 struct tcp_sock
*tp
= tcp_sk(sk
);
3547 if (rexmit
== REXMIT_NONE
)
3550 if (unlikely(rexmit
== 2)) {
3551 __tcp_push_pending_frames(sk
, tcp_current_mss(sk
),
3553 if (after(tp
->snd_nxt
, tp
->high_seq
))
3557 tcp_xmit_retransmit_queue(sk
);
3560 /* This routine deals with incoming acks, but not outgoing ones. */
3561 static int tcp_ack(struct sock
*sk
, const struct sk_buff
*skb
, int flag
)
3563 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3564 struct tcp_sock
*tp
= tcp_sk(sk
);
3565 struct tcp_sacktag_state sack_state
;
3566 u32 prior_snd_una
= tp
->snd_una
;
3567 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
3568 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3569 bool is_dupack
= false;
3571 int prior_packets
= tp
->packets_out
;
3572 u32 prior_delivered
= tp
->delivered
;
3573 int acked
= 0; /* Number of packets newly acked */
3574 int rexmit
= REXMIT_NONE
; /* Flag to (re)transmit to recover losses */
3576 sack_state
.first_sackt
.v64
= 0;
3578 /* We very likely will need to access write queue head. */
3579 prefetchw(sk
->sk_write_queue
.next
);
3581 /* If the ack is older than previous acks
3582 * then we can probably ignore it.
3584 if (before(ack
, prior_snd_una
)) {
3585 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3586 if (before(ack
, prior_snd_una
- tp
->max_window
)) {
3587 tcp_send_challenge_ack(sk
, skb
);
3593 /* If the ack includes data we haven't sent yet, discard
3594 * this segment (RFC793 Section 3.9).
3596 if (after(ack
, tp
->snd_nxt
))
3599 if (icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
||
3600 icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
)
3603 if (after(ack
, prior_snd_una
)) {
3604 flag
|= FLAG_SND_UNA_ADVANCED
;
3605 icsk
->icsk_retransmits
= 0;
3608 prior_fackets
= tp
->fackets_out
;
3610 /* ts_recent update must be made after we are sure that the packet
3613 if (flag
& FLAG_UPDATE_TS_RECENT
)
3614 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
3616 if (!(flag
& FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
3617 /* Window is constant, pure forward advance.
3618 * No more checks are required.
3619 * Note, we use the fact that SND.UNA>=SND.WL2.
3621 tcp_update_wl(tp
, ack_seq
);
3622 tcp_snd_una_update(tp
, ack
);
3623 flag
|= FLAG_WIN_UPDATE
;
3625 tcp_in_ack_event(sk
, CA_ACK_WIN_UPDATE
);
3627 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPACKS
);
3629 u32 ack_ev_flags
= CA_ACK_SLOWPATH
;
3631 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
3634 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPUREACKS
);
3636 flag
|= tcp_ack_update_window(sk
, skb
, ack
, ack_seq
);
3638 if (TCP_SKB_CB(skb
)->sacked
)
3639 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
,
3642 if (tcp_ecn_rcv_ecn_echo(tp
, tcp_hdr(skb
))) {
3644 ack_ev_flags
|= CA_ACK_ECE
;
3647 if (flag
& FLAG_WIN_UPDATE
)
3648 ack_ev_flags
|= CA_ACK_WIN_UPDATE
;
3650 tcp_in_ack_event(sk
, ack_ev_flags
);
3653 /* We passed data and got it acked, remove any soft error
3654 * log. Something worked...
3656 sk
->sk_err_soft
= 0;
3657 icsk
->icsk_probes_out
= 0;
3658 tp
->rcv_tstamp
= tcp_time_stamp
;
3662 /* See if we can take anything off of the retransmit queue. */
3663 flag
|= tcp_clean_rtx_queue(sk
, prior_fackets
, prior_snd_una
, &acked
,
3666 if (tcp_ack_is_dubious(sk
, flag
)) {
3667 is_dupack
= !(flag
& (FLAG_SND_UNA_ADVANCED
| FLAG_NOT_DUP
));
3668 tcp_fastretrans_alert(sk
, acked
, is_dupack
, &flag
, &rexmit
);
3670 if (tp
->tlp_high_seq
)
3671 tcp_process_tlp_ack(sk
, ack
, flag
);
3673 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
& FLAG_NOT_DUP
)) {
3674 struct dst_entry
*dst
= __sk_dst_get(sk
);
3679 if (icsk
->icsk_pending
== ICSK_TIME_RETRANS
)
3680 tcp_schedule_loss_probe(sk
);
3681 tcp_cong_control(sk
, ack
, tp
->delivered
- prior_delivered
, flag
);
3682 tcp_xmit_recovery(sk
, rexmit
);
3686 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3687 if (flag
& FLAG_DSACKING_ACK
)
3688 tcp_fastretrans_alert(sk
, acked
, is_dupack
, &flag
, &rexmit
);
3689 /* If this ack opens up a zero window, clear backoff. It was
3690 * being used to time the probes, and is probably far higher than
3691 * it needs to be for normal retransmission.
3693 if (tcp_send_head(sk
))
3696 if (tp
->tlp_high_seq
)
3697 tcp_process_tlp_ack(sk
, ack
, flag
);
3701 SOCK_DEBUG(sk
, "Ack %u after %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3705 /* If data was SACKed, tag it and see if we should send more data.
3706 * If data was DSACKed, see if we can undo a cwnd reduction.
3708 if (TCP_SKB_CB(skb
)->sacked
) {
3709 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
,
3711 tcp_fastretrans_alert(sk
, acked
, is_dupack
, &flag
, &rexmit
);
3712 tcp_xmit_recovery(sk
, rexmit
);
3715 SOCK_DEBUG(sk
, "Ack %u before %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3719 static void tcp_parse_fastopen_option(int len
, const unsigned char *cookie
,
3720 bool syn
, struct tcp_fastopen_cookie
*foc
,
3723 /* Valid only in SYN or SYN-ACK with an even length. */
3724 if (!foc
|| !syn
|| len
< 0 || (len
& 1))
3727 if (len
>= TCP_FASTOPEN_COOKIE_MIN
&&
3728 len
<= TCP_FASTOPEN_COOKIE_MAX
)
3729 memcpy(foc
->val
, cookie
, len
);
3736 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3737 * But, this can also be called on packets in the established flow when
3738 * the fast version below fails.
3740 void tcp_parse_options(const struct sk_buff
*skb
,
3741 struct tcp_options_received
*opt_rx
, int estab
,
3742 struct tcp_fastopen_cookie
*foc
)
3744 const unsigned char *ptr
;
3745 const struct tcphdr
*th
= tcp_hdr(skb
);
3746 int length
= (th
->doff
* 4) - sizeof(struct tcphdr
);
3748 ptr
= (const unsigned char *)(th
+ 1);
3749 opt_rx
->saw_tstamp
= 0;
3751 while (length
> 0) {
3752 int opcode
= *ptr
++;
3758 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
3763 if (opsize
< 2) /* "silly options" */
3765 if (opsize
> length
)
3766 return; /* don't parse partial options */
3769 if (opsize
== TCPOLEN_MSS
&& th
->syn
&& !estab
) {
3770 u16 in_mss
= get_unaligned_be16(ptr
);
3772 if (opt_rx
->user_mss
&&
3773 opt_rx
->user_mss
< in_mss
)
3774 in_mss
= opt_rx
->user_mss
;
3775 opt_rx
->mss_clamp
= in_mss
;
3780 if (opsize
== TCPOLEN_WINDOW
&& th
->syn
&&
3781 !estab
&& sysctl_tcp_window_scaling
) {
3782 __u8 snd_wscale
= *(__u8
*)ptr
;
3783 opt_rx
->wscale_ok
= 1;
3784 if (snd_wscale
> 14) {
3785 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3790 opt_rx
->snd_wscale
= snd_wscale
;
3793 case TCPOPT_TIMESTAMP
:
3794 if ((opsize
== TCPOLEN_TIMESTAMP
) &&
3795 ((estab
&& opt_rx
->tstamp_ok
) ||
3796 (!estab
&& sysctl_tcp_timestamps
))) {
3797 opt_rx
->saw_tstamp
= 1;
3798 opt_rx
->rcv_tsval
= get_unaligned_be32(ptr
);
3799 opt_rx
->rcv_tsecr
= get_unaligned_be32(ptr
+ 4);
3802 case TCPOPT_SACK_PERM
:
3803 if (opsize
== TCPOLEN_SACK_PERM
&& th
->syn
&&
3804 !estab
&& sysctl_tcp_sack
) {
3805 opt_rx
->sack_ok
= TCP_SACK_SEEN
;
3806 tcp_sack_reset(opt_rx
);
3811 if ((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
3812 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
3814 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
3817 #ifdef CONFIG_TCP_MD5SIG
3820 * The MD5 Hash has already been
3821 * checked (see tcp_v{4,6}_do_rcv()).
3825 case TCPOPT_FASTOPEN
:
3826 tcp_parse_fastopen_option(
3827 opsize
- TCPOLEN_FASTOPEN_BASE
,
3828 ptr
, th
->syn
, foc
, false);
3832 /* Fast Open option shares code 254 using a
3833 * 16 bits magic number.
3835 if (opsize
>= TCPOLEN_EXP_FASTOPEN_BASE
&&
3836 get_unaligned_be16(ptr
) ==
3837 TCPOPT_FASTOPEN_MAGIC
)
3838 tcp_parse_fastopen_option(opsize
-
3839 TCPOLEN_EXP_FASTOPEN_BASE
,
3840 ptr
+ 2, th
->syn
, foc
, true);
3849 EXPORT_SYMBOL(tcp_parse_options
);
3851 static bool tcp_parse_aligned_timestamp(struct tcp_sock
*tp
, const struct tcphdr
*th
)
3853 const __be32
*ptr
= (const __be32
*)(th
+ 1);
3855 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3856 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
3857 tp
->rx_opt
.saw_tstamp
= 1;
3859 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3862 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
) - tp
->tsoffset
;
3864 tp
->rx_opt
.rcv_tsecr
= 0;
3870 /* Fast parse options. This hopes to only see timestamps.
3871 * If it is wrong it falls back on tcp_parse_options().
3873 static bool tcp_fast_parse_options(const struct sk_buff
*skb
,
3874 const struct tcphdr
*th
, struct tcp_sock
*tp
)
3876 /* In the spirit of fast parsing, compare doff directly to constant
3877 * values. Because equality is used, short doff can be ignored here.
3879 if (th
->doff
== (sizeof(*th
) / 4)) {
3880 tp
->rx_opt
.saw_tstamp
= 0;
3882 } else if (tp
->rx_opt
.tstamp_ok
&&
3883 th
->doff
== ((sizeof(*th
) + TCPOLEN_TSTAMP_ALIGNED
) / 4)) {
3884 if (tcp_parse_aligned_timestamp(tp
, th
))
3888 tcp_parse_options(skb
, &tp
->rx_opt
, 1, NULL
);
3889 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
3890 tp
->rx_opt
.rcv_tsecr
-= tp
->tsoffset
;
3895 #ifdef CONFIG_TCP_MD5SIG
3897 * Parse MD5 Signature option
3899 const u8
*tcp_parse_md5sig_option(const struct tcphdr
*th
)
3901 int length
= (th
->doff
<< 2) - sizeof(*th
);
3902 const u8
*ptr
= (const u8
*)(th
+ 1);
3904 /* If the TCP option is too short, we can short cut */
3905 if (length
< TCPOLEN_MD5SIG
)
3908 while (length
> 0) {
3909 int opcode
= *ptr
++;
3920 if (opsize
< 2 || opsize
> length
)
3922 if (opcode
== TCPOPT_MD5SIG
)
3923 return opsize
== TCPOLEN_MD5SIG
? ptr
: NULL
;
3930 EXPORT_SYMBOL(tcp_parse_md5sig_option
);
3933 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3935 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3936 * it can pass through stack. So, the following predicate verifies that
3937 * this segment is not used for anything but congestion avoidance or
3938 * fast retransmit. Moreover, we even are able to eliminate most of such
3939 * second order effects, if we apply some small "replay" window (~RTO)
3940 * to timestamp space.
3942 * All these measures still do not guarantee that we reject wrapped ACKs
3943 * on networks with high bandwidth, when sequence space is recycled fastly,
3944 * but it guarantees that such events will be very rare and do not affect
3945 * connection seriously. This doesn't look nice, but alas, PAWS is really
3948 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3949 * states that events when retransmit arrives after original data are rare.
3950 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3951 * the biggest problem on large power networks even with minor reordering.
3952 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3953 * up to bandwidth of 18Gigabit/sec. 8) ]
3956 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
3958 const struct tcp_sock
*tp
= tcp_sk(sk
);
3959 const struct tcphdr
*th
= tcp_hdr(skb
);
3960 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3961 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3963 return (/* 1. Pure ACK with correct sequence number. */
3964 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
3966 /* 2. ... and duplicate ACK. */
3967 ack
== tp
->snd_una
&&
3969 /* 3. ... and does not update window. */
3970 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
3972 /* 4. ... and sits in replay window. */
3973 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
3976 static inline bool tcp_paws_discard(const struct sock
*sk
,
3977 const struct sk_buff
*skb
)
3979 const struct tcp_sock
*tp
= tcp_sk(sk
);
3981 return !tcp_paws_check(&tp
->rx_opt
, TCP_PAWS_WINDOW
) &&
3982 !tcp_disordered_ack(sk
, skb
);
3985 /* Check segment sequence number for validity.
3987 * Segment controls are considered valid, if the segment
3988 * fits to the window after truncation to the window. Acceptability
3989 * of data (and SYN, FIN, of course) is checked separately.
3990 * See tcp_data_queue(), for example.
3992 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3993 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3994 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3995 * (borrowed from freebsd)
3998 static inline bool tcp_sequence(const struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
4000 return !before(end_seq
, tp
->rcv_wup
) &&
4001 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
4004 /* When we get a reset we do this. */
4005 void tcp_reset(struct sock
*sk
)
4007 /* We want the right error as BSD sees it (and indeed as we do). */
4008 switch (sk
->sk_state
) {
4010 sk
->sk_err
= ECONNREFUSED
;
4012 case TCP_CLOSE_WAIT
:
4018 sk
->sk_err
= ECONNRESET
;
4020 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4023 if (!sock_flag(sk
, SOCK_DEAD
))
4024 sk
->sk_error_report(sk
);
4030 * Process the FIN bit. This now behaves as it is supposed to work
4031 * and the FIN takes effect when it is validly part of sequence
4032 * space. Not before when we get holes.
4034 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4035 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4038 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4039 * close and we go into CLOSING (and later onto TIME-WAIT)
4041 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4043 void tcp_fin(struct sock
*sk
)
4045 struct tcp_sock
*tp
= tcp_sk(sk
);
4047 inet_csk_schedule_ack(sk
);
4049 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
4050 sock_set_flag(sk
, SOCK_DONE
);
4052 switch (sk
->sk_state
) {
4054 case TCP_ESTABLISHED
:
4055 /* Move to CLOSE_WAIT */
4056 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
4057 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
4060 case TCP_CLOSE_WAIT
:
4062 /* Received a retransmission of the FIN, do
4067 /* RFC793: Remain in the LAST-ACK state. */
4071 /* This case occurs when a simultaneous close
4072 * happens, we must ack the received FIN and
4073 * enter the CLOSING state.
4076 tcp_set_state(sk
, TCP_CLOSING
);
4079 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4081 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
4084 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4085 * cases we should never reach this piece of code.
4087 pr_err("%s: Impossible, sk->sk_state=%d\n",
4088 __func__
, sk
->sk_state
);
4092 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4093 * Probably, we should reset in this case. For now drop them.
4095 __skb_queue_purge(&tp
->out_of_order_queue
);
4096 if (tcp_is_sack(tp
))
4097 tcp_sack_reset(&tp
->rx_opt
);
4100 if (!sock_flag(sk
, SOCK_DEAD
)) {
4101 sk
->sk_state_change(sk
);
4103 /* Do not send POLL_HUP for half duplex close. */
4104 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
4105 sk
->sk_state
== TCP_CLOSE
)
4106 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_HUP
);
4108 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_IN
);
4112 static inline bool tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
,
4115 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
4116 if (before(seq
, sp
->start_seq
))
4117 sp
->start_seq
= seq
;
4118 if (after(end_seq
, sp
->end_seq
))
4119 sp
->end_seq
= end_seq
;
4125 static void tcp_dsack_set(struct sock
*sk
, u32 seq
, u32 end_seq
)
4127 struct tcp_sock
*tp
= tcp_sk(sk
);
4129 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
4132 if (before(seq
, tp
->rcv_nxt
))
4133 mib_idx
= LINUX_MIB_TCPDSACKOLDSENT
;
4135 mib_idx
= LINUX_MIB_TCPDSACKOFOSENT
;
4137 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
4139 tp
->rx_opt
.dsack
= 1;
4140 tp
->duplicate_sack
[0].start_seq
= seq
;
4141 tp
->duplicate_sack
[0].end_seq
= end_seq
;
4145 static void tcp_dsack_extend(struct sock
*sk
, u32 seq
, u32 end_seq
)
4147 struct tcp_sock
*tp
= tcp_sk(sk
);
4149 if (!tp
->rx_opt
.dsack
)
4150 tcp_dsack_set(sk
, seq
, end_seq
);
4152 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
4155 static void tcp_send_dupack(struct sock
*sk
, const struct sk_buff
*skb
)
4157 struct tcp_sock
*tp
= tcp_sk(sk
);
4159 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
4160 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4161 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4162 tcp_enter_quickack_mode(sk
);
4164 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
4165 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4167 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
4168 end_seq
= tp
->rcv_nxt
;
4169 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, end_seq
);
4176 /* These routines update the SACK block as out-of-order packets arrive or
4177 * in-order packets close up the sequence space.
4179 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
4182 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4183 struct tcp_sack_block
*swalk
= sp
+ 1;
4185 /* See if the recent change to the first SACK eats into
4186 * or hits the sequence space of other SACK blocks, if so coalesce.
4188 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
;) {
4189 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
4192 /* Zap SWALK, by moving every further SACK up by one slot.
4193 * Decrease num_sacks.
4195 tp
->rx_opt
.num_sacks
--;
4196 for (i
= this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
4200 this_sack
++, swalk
++;
4204 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
4206 struct tcp_sock
*tp
= tcp_sk(sk
);
4207 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4208 int cur_sacks
= tp
->rx_opt
.num_sacks
;
4214 for (this_sack
= 0; this_sack
< cur_sacks
; this_sack
++, sp
++) {
4215 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
4216 /* Rotate this_sack to the first one. */
4217 for (; this_sack
> 0; this_sack
--, sp
--)
4218 swap(*sp
, *(sp
- 1));
4220 tcp_sack_maybe_coalesce(tp
);
4225 /* Could not find an adjacent existing SACK, build a new one,
4226 * put it at the front, and shift everyone else down. We
4227 * always know there is at least one SACK present already here.
4229 * If the sack array is full, forget about the last one.
4231 if (this_sack
>= TCP_NUM_SACKS
) {
4233 tp
->rx_opt
.num_sacks
--;
4236 for (; this_sack
> 0; this_sack
--, sp
--)
4240 /* Build the new head SACK, and we're done. */
4241 sp
->start_seq
= seq
;
4242 sp
->end_seq
= end_seq
;
4243 tp
->rx_opt
.num_sacks
++;
4246 /* RCV.NXT advances, some SACKs should be eaten. */
4248 static void tcp_sack_remove(struct tcp_sock
*tp
)
4250 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4251 int num_sacks
= tp
->rx_opt
.num_sacks
;
4254 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4255 if (skb_queue_empty(&tp
->out_of_order_queue
)) {
4256 tp
->rx_opt
.num_sacks
= 0;
4260 for (this_sack
= 0; this_sack
< num_sacks
;) {
4261 /* Check if the start of the sack is covered by RCV.NXT. */
4262 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
4265 /* RCV.NXT must cover all the block! */
4266 WARN_ON(before(tp
->rcv_nxt
, sp
->end_seq
));
4268 /* Zap this SACK, by moving forward any other SACKS. */
4269 for (i
= this_sack
+1; i
< num_sacks
; i
++)
4270 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
4277 tp
->rx_opt
.num_sacks
= num_sacks
;
4281 * tcp_try_coalesce - try to merge skb to prior one
4284 * @from: buffer to add in queue
4285 * @fragstolen: pointer to boolean
4287 * Before queueing skb @from after @to, try to merge them
4288 * to reduce overall memory use and queue lengths, if cost is small.
4289 * Packets in ofo or receive queues can stay a long time.
4290 * Better try to coalesce them right now to avoid future collapses.
4291 * Returns true if caller should free @from instead of queueing it
4293 static bool tcp_try_coalesce(struct sock
*sk
,
4295 struct sk_buff
*from
,
4300 *fragstolen
= false;
4302 /* Its possible this segment overlaps with prior segment in queue */
4303 if (TCP_SKB_CB(from
)->seq
!= TCP_SKB_CB(to
)->end_seq
)
4306 if (!skb_try_coalesce(to
, from
, fragstolen
, &delta
))
4309 atomic_add(delta
, &sk
->sk_rmem_alloc
);
4310 sk_mem_charge(sk
, delta
);
4311 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRCVCOALESCE
);
4312 TCP_SKB_CB(to
)->end_seq
= TCP_SKB_CB(from
)->end_seq
;
4313 TCP_SKB_CB(to
)->ack_seq
= TCP_SKB_CB(from
)->ack_seq
;
4314 TCP_SKB_CB(to
)->tcp_flags
|= TCP_SKB_CB(from
)->tcp_flags
;
4318 /* This one checks to see if we can put data from the
4319 * out_of_order queue into the receive_queue.
4321 static void tcp_ofo_queue(struct sock
*sk
)
4323 struct tcp_sock
*tp
= tcp_sk(sk
);
4324 __u32 dsack_high
= tp
->rcv_nxt
;
4325 struct sk_buff
*skb
, *tail
;
4326 bool fragstolen
, eaten
;
4328 while ((skb
= skb_peek(&tp
->out_of_order_queue
)) != NULL
) {
4329 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
4332 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
4333 __u32 dsack
= dsack_high
;
4334 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
4335 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
4336 tcp_dsack_extend(sk
, TCP_SKB_CB(skb
)->seq
, dsack
);
4339 __skb_unlink(skb
, &tp
->out_of_order_queue
);
4340 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4341 SOCK_DEBUG(sk
, "ofo packet was already received\n");
4345 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
4346 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4347 TCP_SKB_CB(skb
)->end_seq
);
4349 tail
= skb_peek_tail(&sk
->sk_receive_queue
);
4350 eaten
= tail
&& tcp_try_coalesce(sk
, tail
, skb
, &fragstolen
);
4351 tcp_rcv_nxt_update(tp
, TCP_SKB_CB(skb
)->end_seq
);
4353 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4354 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
4357 kfree_skb_partial(skb
, fragstolen
);
4361 static bool tcp_prune_ofo_queue(struct sock
*sk
);
4362 static int tcp_prune_queue(struct sock
*sk
);
4364 static int tcp_try_rmem_schedule(struct sock
*sk
, struct sk_buff
*skb
,
4367 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
4368 !sk_rmem_schedule(sk
, skb
, size
)) {
4370 if (tcp_prune_queue(sk
) < 0)
4373 if (!sk_rmem_schedule(sk
, skb
, size
)) {
4374 if (!tcp_prune_ofo_queue(sk
))
4377 if (!sk_rmem_schedule(sk
, skb
, size
))
4384 static void tcp_data_queue_ofo(struct sock
*sk
, struct sk_buff
*skb
)
4386 struct tcp_sock
*tp
= tcp_sk(sk
);
4387 struct sk_buff
*skb1
;
4390 tcp_ecn_check_ce(tp
, skb
);
4392 if (unlikely(tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))) {
4393 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFODROP
);
4398 /* Disable header prediction. */
4400 inet_csk_schedule_ack(sk
);
4402 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFOQUEUE
);
4403 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
4404 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4406 skb1
= skb_peek_tail(&tp
->out_of_order_queue
);
4408 /* Initial out of order segment, build 1 SACK. */
4409 if (tcp_is_sack(tp
)) {
4410 tp
->rx_opt
.num_sacks
= 1;
4411 tp
->selective_acks
[0].start_seq
= TCP_SKB_CB(skb
)->seq
;
4412 tp
->selective_acks
[0].end_seq
=
4413 TCP_SKB_CB(skb
)->end_seq
;
4415 __skb_queue_head(&tp
->out_of_order_queue
, skb
);
4419 seq
= TCP_SKB_CB(skb
)->seq
;
4420 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4422 if (seq
== TCP_SKB_CB(skb1
)->end_seq
) {
4425 if (!tcp_try_coalesce(sk
, skb1
, skb
, &fragstolen
)) {
4426 __skb_queue_after(&tp
->out_of_order_queue
, skb1
, skb
);
4428 tcp_grow_window(sk
, skb
);
4429 kfree_skb_partial(skb
, fragstolen
);
4433 if (!tp
->rx_opt
.num_sacks
||
4434 tp
->selective_acks
[0].end_seq
!= seq
)
4437 /* Common case: data arrive in order after hole. */
4438 tp
->selective_acks
[0].end_seq
= end_seq
;
4442 /* Find place to insert this segment. */
4444 if (!after(TCP_SKB_CB(skb1
)->seq
, seq
))
4446 if (skb_queue_is_first(&tp
->out_of_order_queue
, skb1
)) {
4450 skb1
= skb_queue_prev(&tp
->out_of_order_queue
, skb1
);
4453 /* Do skb overlap to previous one? */
4454 if (skb1
&& before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4455 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4456 /* All the bits are present. Drop. */
4457 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFOMERGE
);
4460 tcp_dsack_set(sk
, seq
, end_seq
);
4463 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
4464 /* Partial overlap. */
4465 tcp_dsack_set(sk
, seq
,
4466 TCP_SKB_CB(skb1
)->end_seq
);
4468 if (skb_queue_is_first(&tp
->out_of_order_queue
,
4472 skb1
= skb_queue_prev(
4473 &tp
->out_of_order_queue
,
4478 __skb_queue_head(&tp
->out_of_order_queue
, skb
);
4480 __skb_queue_after(&tp
->out_of_order_queue
, skb1
, skb
);
4482 /* And clean segments covered by new one as whole. */
4483 while (!skb_queue_is_last(&tp
->out_of_order_queue
, skb
)) {
4484 skb1
= skb_queue_next(&tp
->out_of_order_queue
, skb
);
4486 if (!after(end_seq
, TCP_SKB_CB(skb1
)->seq
))
4488 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4489 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4493 __skb_unlink(skb1
, &tp
->out_of_order_queue
);
4494 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4495 TCP_SKB_CB(skb1
)->end_seq
);
4496 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFOMERGE
);
4501 if (tcp_is_sack(tp
))
4502 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
4505 tcp_grow_window(sk
, skb
);
4506 skb_set_owner_r(skb
, sk
);
4510 static int __must_check
tcp_queue_rcv(struct sock
*sk
, struct sk_buff
*skb
, int hdrlen
,
4514 struct sk_buff
*tail
= skb_peek_tail(&sk
->sk_receive_queue
);
4516 __skb_pull(skb
, hdrlen
);
4518 tcp_try_coalesce(sk
, tail
, skb
, fragstolen
)) ? 1 : 0;
4519 tcp_rcv_nxt_update(tcp_sk(sk
), TCP_SKB_CB(skb
)->end_seq
);
4521 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4522 skb_set_owner_r(skb
, sk
);
4527 int tcp_send_rcvq(struct sock
*sk
, struct msghdr
*msg
, size_t size
)
4529 struct sk_buff
*skb
;
4537 if (size
> PAGE_SIZE
) {
4538 int npages
= min_t(size_t, size
>> PAGE_SHIFT
, MAX_SKB_FRAGS
);
4540 data_len
= npages
<< PAGE_SHIFT
;
4541 size
= data_len
+ (size
& ~PAGE_MASK
);
4543 skb
= alloc_skb_with_frags(size
- data_len
, data_len
,
4544 PAGE_ALLOC_COSTLY_ORDER
,
4545 &err
, sk
->sk_allocation
);
4549 skb_put(skb
, size
- data_len
);
4550 skb
->data_len
= data_len
;
4553 if (tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))
4556 err
= skb_copy_datagram_from_iter(skb
, 0, &msg
->msg_iter
, size
);
4560 TCP_SKB_CB(skb
)->seq
= tcp_sk(sk
)->rcv_nxt
;
4561 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ size
;
4562 TCP_SKB_CB(skb
)->ack_seq
= tcp_sk(sk
)->snd_una
- 1;
4564 if (tcp_queue_rcv(sk
, skb
, 0, &fragstolen
)) {
4565 WARN_ON_ONCE(fragstolen
); /* should not happen */
4577 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
4579 struct tcp_sock
*tp
= tcp_sk(sk
);
4581 bool fragstolen
= false;
4583 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
)
4587 __skb_pull(skb
, tcp_hdr(skb
)->doff
* 4);
4589 tcp_ecn_accept_cwr(tp
, skb
);
4591 tp
->rx_opt
.dsack
= 0;
4593 /* Queue data for delivery to the user.
4594 * Packets in sequence go to the receive queue.
4595 * Out of sequence packets to the out_of_order_queue.
4597 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
4598 if (tcp_receive_window(tp
) == 0)
4601 /* Ok. In sequence. In window. */
4602 if (tp
->ucopy
.task
== current
&&
4603 tp
->copied_seq
== tp
->rcv_nxt
&& tp
->ucopy
.len
&&
4604 sock_owned_by_user(sk
) && !tp
->urg_data
) {
4605 int chunk
= min_t(unsigned int, skb
->len
,
4608 __set_current_state(TASK_RUNNING
);
4611 if (!skb_copy_datagram_msg(skb
, 0, tp
->ucopy
.msg
, chunk
)) {
4612 tp
->ucopy
.len
-= chunk
;
4613 tp
->copied_seq
+= chunk
;
4614 eaten
= (chunk
== skb
->len
);
4615 tcp_rcv_space_adjust(sk
);
4623 if (skb_queue_len(&sk
->sk_receive_queue
) == 0)
4624 sk_forced_mem_schedule(sk
, skb
->truesize
);
4625 else if (tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))
4628 eaten
= tcp_queue_rcv(sk
, skb
, 0, &fragstolen
);
4630 tcp_rcv_nxt_update(tp
, TCP_SKB_CB(skb
)->end_seq
);
4632 tcp_event_data_recv(sk
, skb
);
4633 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
4636 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4639 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4640 * gap in queue is filled.
4642 if (skb_queue_empty(&tp
->out_of_order_queue
))
4643 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
4646 if (tp
->rx_opt
.num_sacks
)
4647 tcp_sack_remove(tp
);
4649 tcp_fast_path_check(sk
);
4652 kfree_skb_partial(skb
, fragstolen
);
4653 if (!sock_flag(sk
, SOCK_DEAD
))
4654 sk
->sk_data_ready(sk
);
4658 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4659 /* A retransmit, 2nd most common case. Force an immediate ack. */
4660 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4661 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4664 tcp_enter_quickack_mode(sk
);
4665 inet_csk_schedule_ack(sk
);
4671 /* Out of window. F.e. zero window probe. */
4672 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
4675 tcp_enter_quickack_mode(sk
);
4677 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4678 /* Partial packet, seq < rcv_next < end_seq */
4679 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
4680 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4681 TCP_SKB_CB(skb
)->end_seq
);
4683 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
4685 /* If window is closed, drop tail of packet. But after
4686 * remembering D-SACK for its head made in previous line.
4688 if (!tcp_receive_window(tp
))
4693 tcp_data_queue_ofo(sk
, skb
);
4696 static struct sk_buff
*tcp_collapse_one(struct sock
*sk
, struct sk_buff
*skb
,
4697 struct sk_buff_head
*list
)
4699 struct sk_buff
*next
= NULL
;
4701 if (!skb_queue_is_last(list
, skb
))
4702 next
= skb_queue_next(list
, skb
);
4704 __skb_unlink(skb
, list
);
4706 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRCVCOLLAPSED
);
4711 /* Collapse contiguous sequence of skbs head..tail with
4712 * sequence numbers start..end.
4714 * If tail is NULL, this means until the end of the list.
4716 * Segments with FIN/SYN are not collapsed (only because this
4720 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
,
4721 struct sk_buff
*head
, struct sk_buff
*tail
,
4724 struct sk_buff
*skb
, *n
;
4727 /* First, check that queue is collapsible and find
4728 * the point where collapsing can be useful. */
4732 skb_queue_walk_from_safe(list
, skb
, n
) {
4735 /* No new bits? It is possible on ofo queue. */
4736 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4737 skb
= tcp_collapse_one(sk
, skb
, list
);
4743 /* The first skb to collapse is:
4745 * - bloated or contains data before "start" or
4746 * overlaps to the next one.
4748 if (!(TCP_SKB_CB(skb
)->tcp_flags
& (TCPHDR_SYN
| TCPHDR_FIN
)) &&
4749 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
4750 before(TCP_SKB_CB(skb
)->seq
, start
))) {
4751 end_of_skbs
= false;
4755 if (!skb_queue_is_last(list
, skb
)) {
4756 struct sk_buff
*next
= skb_queue_next(list
, skb
);
4758 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(next
)->seq
) {
4759 end_of_skbs
= false;
4764 /* Decided to skip this, advance start seq. */
4765 start
= TCP_SKB_CB(skb
)->end_seq
;
4768 (TCP_SKB_CB(skb
)->tcp_flags
& (TCPHDR_SYN
| TCPHDR_FIN
)))
4771 while (before(start
, end
)) {
4772 int copy
= min_t(int, SKB_MAX_ORDER(0, 0), end
- start
);
4773 struct sk_buff
*nskb
;
4775 nskb
= alloc_skb(copy
, GFP_ATOMIC
);
4779 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
4780 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
4781 __skb_queue_before(list
, skb
, nskb
);
4782 skb_set_owner_r(nskb
, sk
);
4784 /* Copy data, releasing collapsed skbs. */
4786 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
4787 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
4791 size
= min(copy
, size
);
4792 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
4794 TCP_SKB_CB(nskb
)->end_seq
+= size
;
4798 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4799 skb
= tcp_collapse_one(sk
, skb
, list
);
4802 (TCP_SKB_CB(skb
)->tcp_flags
& (TCPHDR_SYN
| TCPHDR_FIN
)))
4809 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4810 * and tcp_collapse() them until all the queue is collapsed.
4812 static void tcp_collapse_ofo_queue(struct sock
*sk
)
4814 struct tcp_sock
*tp
= tcp_sk(sk
);
4815 struct sk_buff
*skb
= skb_peek(&tp
->out_of_order_queue
);
4816 struct sk_buff
*head
;
4822 start
= TCP_SKB_CB(skb
)->seq
;
4823 end
= TCP_SKB_CB(skb
)->end_seq
;
4827 struct sk_buff
*next
= NULL
;
4829 if (!skb_queue_is_last(&tp
->out_of_order_queue
, skb
))
4830 next
= skb_queue_next(&tp
->out_of_order_queue
, skb
);
4833 /* Segment is terminated when we see gap or when
4834 * we are at the end of all the queue. */
4836 after(TCP_SKB_CB(skb
)->seq
, end
) ||
4837 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
4838 tcp_collapse(sk
, &tp
->out_of_order_queue
,
4839 head
, skb
, start
, end
);
4843 /* Start new segment */
4844 start
= TCP_SKB_CB(skb
)->seq
;
4845 end
= TCP_SKB_CB(skb
)->end_seq
;
4847 if (before(TCP_SKB_CB(skb
)->seq
, start
))
4848 start
= TCP_SKB_CB(skb
)->seq
;
4849 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
4850 end
= TCP_SKB_CB(skb
)->end_seq
;
4856 * Purge the out-of-order queue.
4857 * Return true if queue was pruned.
4859 static bool tcp_prune_ofo_queue(struct sock
*sk
)
4861 struct tcp_sock
*tp
= tcp_sk(sk
);
4864 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4865 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_OFOPRUNED
);
4866 __skb_queue_purge(&tp
->out_of_order_queue
);
4868 /* Reset SACK state. A conforming SACK implementation will
4869 * do the same at a timeout based retransmit. When a connection
4870 * is in a sad state like this, we care only about integrity
4871 * of the connection not performance.
4873 if (tp
->rx_opt
.sack_ok
)
4874 tcp_sack_reset(&tp
->rx_opt
);
4881 /* Reduce allocated memory if we can, trying to get
4882 * the socket within its memory limits again.
4884 * Return less than zero if we should start dropping frames
4885 * until the socket owning process reads some of the data
4886 * to stabilize the situation.
4888 static int tcp_prune_queue(struct sock
*sk
)
4890 struct tcp_sock
*tp
= tcp_sk(sk
);
4892 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
4894 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PRUNECALLED
);
4896 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
4897 tcp_clamp_window(sk
);
4898 else if (tcp_under_memory_pressure(sk
))
4899 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
4901 tcp_collapse_ofo_queue(sk
);
4902 if (!skb_queue_empty(&sk
->sk_receive_queue
))
4903 tcp_collapse(sk
, &sk
->sk_receive_queue
,
4904 skb_peek(&sk
->sk_receive_queue
),
4906 tp
->copied_seq
, tp
->rcv_nxt
);
4909 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4912 /* Collapsing did not help, destructive actions follow.
4913 * This must not ever occur. */
4915 tcp_prune_ofo_queue(sk
);
4917 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4920 /* If we are really being abused, tell the caller to silently
4921 * drop receive data on the floor. It will get retransmitted
4922 * and hopefully then we'll have sufficient space.
4924 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_RCVPRUNED
);
4926 /* Massive buffer overcommit. */
4931 static bool tcp_should_expand_sndbuf(const struct sock
*sk
)
4933 const struct tcp_sock
*tp
= tcp_sk(sk
);
4935 /* If the user specified a specific send buffer setting, do
4938 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
4941 /* If we are under global TCP memory pressure, do not expand. */
4942 if (tcp_under_memory_pressure(sk
))
4945 /* If we are under soft global TCP memory pressure, do not expand. */
4946 if (sk_memory_allocated(sk
) >= sk_prot_mem_limits(sk
, 0))
4949 /* If we filled the congestion window, do not expand. */
4950 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
4956 /* When incoming ACK allowed to free some skb from write_queue,
4957 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4958 * on the exit from tcp input handler.
4960 * PROBLEM: sndbuf expansion does not work well with largesend.
4962 static void tcp_new_space(struct sock
*sk
)
4964 struct tcp_sock
*tp
= tcp_sk(sk
);
4966 if (tcp_should_expand_sndbuf(sk
)) {
4967 tcp_sndbuf_expand(sk
);
4968 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4971 sk
->sk_write_space(sk
);
4974 static void tcp_check_space(struct sock
*sk
)
4976 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
4977 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
4978 /* pairs with tcp_poll() */
4979 smp_mb__after_atomic();
4980 if (sk
->sk_socket
&&
4981 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
4986 static inline void tcp_data_snd_check(struct sock
*sk
)
4988 tcp_push_pending_frames(sk
);
4989 tcp_check_space(sk
);
4993 * Check if sending an ack is needed.
4995 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
4997 struct tcp_sock
*tp
= tcp_sk(sk
);
4999 /* More than one full frame received... */
5000 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
&&
5001 /* ... and right edge of window advances far enough.
5002 * (tcp_recvmsg() will send ACK otherwise). Or...
5004 __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
5005 /* We ACK each frame or... */
5006 tcp_in_quickack_mode(sk
) ||
5007 /* We have out of order data. */
5008 (ofo_possible
&& skb_peek(&tp
->out_of_order_queue
))) {
5009 /* Then ack it now */
5012 /* Else, send delayed ack. */
5013 tcp_send_delayed_ack(sk
);
5017 static inline void tcp_ack_snd_check(struct sock
*sk
)
5019 if (!inet_csk_ack_scheduled(sk
)) {
5020 /* We sent a data segment already. */
5023 __tcp_ack_snd_check(sk
, 1);
5027 * This routine is only called when we have urgent data
5028 * signaled. Its the 'slow' part of tcp_urg. It could be
5029 * moved inline now as tcp_urg is only called from one
5030 * place. We handle URGent data wrong. We have to - as
5031 * BSD still doesn't use the correction from RFC961.
5032 * For 1003.1g we should support a new option TCP_STDURG to permit
5033 * either form (or just set the sysctl tcp_stdurg).
5036 static void tcp_check_urg(struct sock
*sk
, const struct tcphdr
*th
)
5038 struct tcp_sock
*tp
= tcp_sk(sk
);
5039 u32 ptr
= ntohs(th
->urg_ptr
);
5041 if (ptr
&& !sysctl_tcp_stdurg
)
5043 ptr
+= ntohl(th
->seq
);
5045 /* Ignore urgent data that we've already seen and read. */
5046 if (after(tp
->copied_seq
, ptr
))
5049 /* Do not replay urg ptr.
5051 * NOTE: interesting situation not covered by specs.
5052 * Misbehaving sender may send urg ptr, pointing to segment,
5053 * which we already have in ofo queue. We are not able to fetch
5054 * such data and will stay in TCP_URG_NOTYET until will be eaten
5055 * by recvmsg(). Seems, we are not obliged to handle such wicked
5056 * situations. But it is worth to think about possibility of some
5057 * DoSes using some hypothetical application level deadlock.
5059 if (before(ptr
, tp
->rcv_nxt
))
5062 /* Do we already have a newer (or duplicate) urgent pointer? */
5063 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
5066 /* Tell the world about our new urgent pointer. */
5069 /* We may be adding urgent data when the last byte read was
5070 * urgent. To do this requires some care. We cannot just ignore
5071 * tp->copied_seq since we would read the last urgent byte again
5072 * as data, nor can we alter copied_seq until this data arrives
5073 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5075 * NOTE. Double Dutch. Rendering to plain English: author of comment
5076 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5077 * and expect that both A and B disappear from stream. This is _wrong_.
5078 * Though this happens in BSD with high probability, this is occasional.
5079 * Any application relying on this is buggy. Note also, that fix "works"
5080 * only in this artificial test. Insert some normal data between A and B and we will
5081 * decline of BSD again. Verdict: it is better to remove to trap
5084 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
5085 !sock_flag(sk
, SOCK_URGINLINE
) && tp
->copied_seq
!= tp
->rcv_nxt
) {
5086 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
5088 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5089 __skb_unlink(skb
, &sk
->sk_receive_queue
);
5094 tp
->urg_data
= TCP_URG_NOTYET
;
5097 /* Disable header prediction. */
5101 /* This is the 'fast' part of urgent handling. */
5102 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, const struct tcphdr
*th
)
5104 struct tcp_sock
*tp
= tcp_sk(sk
);
5106 /* Check if we get a new urgent pointer - normally not. */
5108 tcp_check_urg(sk
, th
);
5110 /* Do we wait for any urgent data? - normally not... */
5111 if (tp
->urg_data
== TCP_URG_NOTYET
) {
5112 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
5115 /* Is the urgent pointer pointing into this packet? */
5116 if (ptr
< skb
->len
) {
5118 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
5120 tp
->urg_data
= TCP_URG_VALID
| tmp
;
5121 if (!sock_flag(sk
, SOCK_DEAD
))
5122 sk
->sk_data_ready(sk
);
5127 static int tcp_copy_to_iovec(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
5129 struct tcp_sock
*tp
= tcp_sk(sk
);
5130 int chunk
= skb
->len
- hlen
;
5134 if (skb_csum_unnecessary(skb
))
5135 err
= skb_copy_datagram_msg(skb
, hlen
, tp
->ucopy
.msg
, chunk
);
5137 err
= skb_copy_and_csum_datagram_msg(skb
, hlen
, tp
->ucopy
.msg
);
5140 tp
->ucopy
.len
-= chunk
;
5141 tp
->copied_seq
+= chunk
;
5142 tcp_rcv_space_adjust(sk
);
5149 static __sum16
__tcp_checksum_complete_user(struct sock
*sk
,
5150 struct sk_buff
*skb
)
5154 if (sock_owned_by_user(sk
)) {
5156 result
= __tcp_checksum_complete(skb
);
5159 result
= __tcp_checksum_complete(skb
);
5164 static inline bool tcp_checksum_complete_user(struct sock
*sk
,
5165 struct sk_buff
*skb
)
5167 return !skb_csum_unnecessary(skb
) &&
5168 __tcp_checksum_complete_user(sk
, skb
);
5171 /* Does PAWS and seqno based validation of an incoming segment, flags will
5172 * play significant role here.
5174 static bool tcp_validate_incoming(struct sock
*sk
, struct sk_buff
*skb
,
5175 const struct tcphdr
*th
, int syn_inerr
)
5177 struct tcp_sock
*tp
= tcp_sk(sk
);
5179 /* RFC1323: H1. Apply PAWS check first. */
5180 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
5181 tcp_paws_discard(sk
, skb
)) {
5183 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSESTABREJECTED
);
5184 if (!tcp_oow_rate_limited(sock_net(sk
), skb
,
5185 LINUX_MIB_TCPACKSKIPPEDPAWS
,
5186 &tp
->last_oow_ack_time
))
5187 tcp_send_dupack(sk
, skb
);
5190 /* Reset is accepted even if it did not pass PAWS. */
5193 /* Step 1: check sequence number */
5194 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5195 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5196 * (RST) segments are validated by checking their SEQ-fields."
5197 * And page 69: "If an incoming segment is not acceptable,
5198 * an acknowledgment should be sent in reply (unless the RST
5199 * bit is set, if so drop the segment and return)".
5204 if (!tcp_oow_rate_limited(sock_net(sk
), skb
,
5205 LINUX_MIB_TCPACKSKIPPEDSEQ
,
5206 &tp
->last_oow_ack_time
))
5207 tcp_send_dupack(sk
, skb
);
5212 /* Step 2: check RST bit */
5215 * If sequence number exactly matches RCV.NXT, then
5216 * RESET the connection
5218 * Send a challenge ACK
5220 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
)
5223 tcp_send_challenge_ack(sk
, skb
);
5227 /* step 3: check security and precedence [ignored] */
5229 /* step 4: Check for a SYN
5230 * RFC 5961 4.2 : Send a challenge ack
5235 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5236 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSYNCHALLENGE
);
5237 tcp_send_challenge_ack(sk
, skb
);
5249 * TCP receive function for the ESTABLISHED state.
5251 * It is split into a fast path and a slow path. The fast path is
5253 * - A zero window was announced from us - zero window probing
5254 * is only handled properly in the slow path.
5255 * - Out of order segments arrived.
5256 * - Urgent data is expected.
5257 * - There is no buffer space left
5258 * - Unexpected TCP flags/window values/header lengths are received
5259 * (detected by checking the TCP header against pred_flags)
5260 * - Data is sent in both directions. Fast path only supports pure senders
5261 * or pure receivers (this means either the sequence number or the ack
5262 * value must stay constant)
5263 * - Unexpected TCP option.
5265 * When these conditions are not satisfied it drops into a standard
5266 * receive procedure patterned after RFC793 to handle all cases.
5267 * The first three cases are guaranteed by proper pred_flags setting,
5268 * the rest is checked inline. Fast processing is turned on in
5269 * tcp_data_queue when everything is OK.
5271 void tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
5272 const struct tcphdr
*th
, unsigned int len
)
5274 struct tcp_sock
*tp
= tcp_sk(sk
);
5276 if (unlikely(!sk
->sk_rx_dst
))
5277 inet_csk(sk
)->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
5279 * Header prediction.
5280 * The code loosely follows the one in the famous
5281 * "30 instruction TCP receive" Van Jacobson mail.
5283 * Van's trick is to deposit buffers into socket queue
5284 * on a device interrupt, to call tcp_recv function
5285 * on the receive process context and checksum and copy
5286 * the buffer to user space. smart...
5288 * Our current scheme is not silly either but we take the
5289 * extra cost of the net_bh soft interrupt processing...
5290 * We do checksum and copy also but from device to kernel.
5293 tp
->rx_opt
.saw_tstamp
= 0;
5295 /* pred_flags is 0xS?10 << 16 + snd_wnd
5296 * if header_prediction is to be made
5297 * 'S' will always be tp->tcp_header_len >> 2
5298 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5299 * turn it off (when there are holes in the receive
5300 * space for instance)
5301 * PSH flag is ignored.
5304 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
5305 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
&&
5306 !after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
)) {
5307 int tcp_header_len
= tp
->tcp_header_len
;
5309 /* Timestamp header prediction: tcp_header_len
5310 * is automatically equal to th->doff*4 due to pred_flags
5314 /* Check timestamp */
5315 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
5316 /* No? Slow path! */
5317 if (!tcp_parse_aligned_timestamp(tp
, th
))
5320 /* If PAWS failed, check it more carefully in slow path */
5321 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
5324 /* DO NOT update ts_recent here, if checksum fails
5325 * and timestamp was corrupted part, it will result
5326 * in a hung connection since we will drop all
5327 * future packets due to the PAWS test.
5331 if (len
<= tcp_header_len
) {
5332 /* Bulk data transfer: sender */
5333 if (len
== tcp_header_len
) {
5334 /* Predicted packet is in window by definition.
5335 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5336 * Hence, check seq<=rcv_wup reduces to:
5338 if (tcp_header_len
==
5339 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5340 tp
->rcv_nxt
== tp
->rcv_wup
)
5341 tcp_store_ts_recent(tp
);
5343 /* We know that such packets are checksummed
5346 tcp_ack(sk
, skb
, 0);
5348 tcp_data_snd_check(sk
);
5350 } else { /* Header too small */
5351 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5356 bool fragstolen
= false;
5358 if (tp
->ucopy
.task
== current
&&
5359 tp
->copied_seq
== tp
->rcv_nxt
&&
5360 len
- tcp_header_len
<= tp
->ucopy
.len
&&
5361 sock_owned_by_user(sk
)) {
5362 __set_current_state(TASK_RUNNING
);
5364 if (!tcp_copy_to_iovec(sk
, skb
, tcp_header_len
)) {
5365 /* Predicted packet is in window by definition.
5366 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5367 * Hence, check seq<=rcv_wup reduces to:
5369 if (tcp_header_len
==
5370 (sizeof(struct tcphdr
) +
5371 TCPOLEN_TSTAMP_ALIGNED
) &&
5372 tp
->rcv_nxt
== tp
->rcv_wup
)
5373 tcp_store_ts_recent(tp
);
5375 tcp_rcv_rtt_measure_ts(sk
, skb
);
5377 __skb_pull(skb
, tcp_header_len
);
5378 tcp_rcv_nxt_update(tp
, TCP_SKB_CB(skb
)->end_seq
);
5379 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITSTOUSER
);
5384 if (tcp_checksum_complete_user(sk
, skb
))
5387 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
5390 /* Predicted packet is in window by definition.
5391 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5392 * Hence, check seq<=rcv_wup reduces to:
5394 if (tcp_header_len
==
5395 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5396 tp
->rcv_nxt
== tp
->rcv_wup
)
5397 tcp_store_ts_recent(tp
);
5399 tcp_rcv_rtt_measure_ts(sk
, skb
);
5401 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITS
);
5403 /* Bulk data transfer: receiver */
5404 eaten
= tcp_queue_rcv(sk
, skb
, tcp_header_len
,
5408 tcp_event_data_recv(sk
, skb
);
5410 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
5411 /* Well, only one small jumplet in fast path... */
5412 tcp_ack(sk
, skb
, FLAG_DATA
);
5413 tcp_data_snd_check(sk
);
5414 if (!inet_csk_ack_scheduled(sk
))
5418 __tcp_ack_snd_check(sk
, 0);
5421 kfree_skb_partial(skb
, fragstolen
);
5422 sk
->sk_data_ready(sk
);
5428 if (len
< (th
->doff
<< 2) || tcp_checksum_complete_user(sk
, skb
))
5431 if (!th
->ack
&& !th
->rst
&& !th
->syn
)
5435 * Standard slow path.
5438 if (!tcp_validate_incoming(sk
, skb
, th
, 1))
5442 if (tcp_ack(sk
, skb
, FLAG_SLOWPATH
| FLAG_UPDATE_TS_RECENT
) < 0)
5445 tcp_rcv_rtt_measure_ts(sk
, skb
);
5447 /* Process urgent data. */
5448 tcp_urg(sk
, skb
, th
);
5450 /* step 7: process the segment text */
5451 tcp_data_queue(sk
, skb
);
5453 tcp_data_snd_check(sk
);
5454 tcp_ack_snd_check(sk
);
5458 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_CSUMERRORS
);
5459 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5464 EXPORT_SYMBOL(tcp_rcv_established
);
5466 void tcp_finish_connect(struct sock
*sk
, struct sk_buff
*skb
)
5468 struct tcp_sock
*tp
= tcp_sk(sk
);
5469 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5471 tcp_set_state(sk
, TCP_ESTABLISHED
);
5474 icsk
->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
5475 security_inet_conn_established(sk
, skb
);
5478 /* Make sure socket is routed, for correct metrics. */
5479 icsk
->icsk_af_ops
->rebuild_header(sk
);
5481 tcp_init_metrics(sk
);
5483 tcp_init_congestion_control(sk
);
5485 /* Prevent spurious tcp_cwnd_restart() on first data
5488 tp
->lsndtime
= tcp_time_stamp
;
5490 tcp_init_buffer_space(sk
);
5492 if (sock_flag(sk
, SOCK_KEEPOPEN
))
5493 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
5495 if (!tp
->rx_opt
.snd_wscale
)
5496 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
5500 if (!sock_flag(sk
, SOCK_DEAD
)) {
5501 sk
->sk_state_change(sk
);
5502 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5506 static bool tcp_rcv_fastopen_synack(struct sock
*sk
, struct sk_buff
*synack
,
5507 struct tcp_fastopen_cookie
*cookie
)
5509 struct tcp_sock
*tp
= tcp_sk(sk
);
5510 struct sk_buff
*data
= tp
->syn_data
? tcp_write_queue_head(sk
) : NULL
;
5511 u16 mss
= tp
->rx_opt
.mss_clamp
, try_exp
= 0;
5512 bool syn_drop
= false;
5514 if (mss
== tp
->rx_opt
.user_mss
) {
5515 struct tcp_options_received opt
;
5517 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5518 tcp_clear_options(&opt
);
5519 opt
.user_mss
= opt
.mss_clamp
= 0;
5520 tcp_parse_options(synack
, &opt
, 0, NULL
);
5521 mss
= opt
.mss_clamp
;
5524 if (!tp
->syn_fastopen
) {
5525 /* Ignore an unsolicited cookie */
5527 } else if (tp
->total_retrans
) {
5528 /* SYN timed out and the SYN-ACK neither has a cookie nor
5529 * acknowledges data. Presumably the remote received only
5530 * the retransmitted (regular) SYNs: either the original
5531 * SYN-data or the corresponding SYN-ACK was dropped.
5533 syn_drop
= (cookie
->len
< 0 && data
);
5534 } else if (cookie
->len
< 0 && !tp
->syn_data
) {
5535 /* We requested a cookie but didn't get it. If we did not use
5536 * the (old) exp opt format then try so next time (try_exp=1).
5537 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5539 try_exp
= tp
->syn_fastopen_exp
? 2 : 1;
5542 tcp_fastopen_cache_set(sk
, mss
, cookie
, syn_drop
, try_exp
);
5544 if (data
) { /* Retransmit unacked data in SYN */
5545 tcp_for_write_queue_from(data
, sk
) {
5546 if (data
== tcp_send_head(sk
) ||
5547 __tcp_retransmit_skb(sk
, data
))
5551 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPFASTOPENACTIVEFAIL
);
5554 tp
->syn_data_acked
= tp
->syn_data
;
5555 if (tp
->syn_data_acked
)
5556 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPFASTOPENACTIVE
);
5558 tcp_fastopen_add_skb(sk
, synack
);
5563 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5564 const struct tcphdr
*th
)
5566 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5567 struct tcp_sock
*tp
= tcp_sk(sk
);
5568 struct tcp_fastopen_cookie foc
= { .len
= -1 };
5569 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
5571 tcp_parse_options(skb
, &tp
->rx_opt
, 0, &foc
);
5572 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
5573 tp
->rx_opt
.rcv_tsecr
-= tp
->tsoffset
;
5577 * "If the state is SYN-SENT then
5578 * first check the ACK bit
5579 * If the ACK bit is set
5580 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5581 * a reset (unless the RST bit is set, if so drop
5582 * the segment and return)"
5584 if (!after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_una
) ||
5585 after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
))
5586 goto reset_and_undo
;
5588 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
5589 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
5591 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSACTIVEREJECTED
);
5592 goto reset_and_undo
;
5595 /* Now ACK is acceptable.
5597 * "If the RST bit is set
5598 * If the ACK was acceptable then signal the user "error:
5599 * connection reset", drop the segment, enter CLOSED state,
5600 * delete TCB, and return."
5609 * "fifth, if neither of the SYN or RST bits is set then
5610 * drop the segment and return."
5616 goto discard_and_undo
;
5619 * "If the SYN bit is on ...
5620 * are acceptable then ...
5621 * (our SYN has been ACKed), change the connection
5622 * state to ESTABLISHED..."
5625 tcp_ecn_rcv_synack(tp
, th
);
5627 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5628 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
5630 /* Ok.. it's good. Set up sequence numbers and
5631 * move to established.
5633 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5634 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5636 /* RFC1323: The window in SYN & SYN/ACK segments is
5639 tp
->snd_wnd
= ntohs(th
->window
);
5641 if (!tp
->rx_opt
.wscale_ok
) {
5642 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
5643 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
5646 if (tp
->rx_opt
.saw_tstamp
) {
5647 tp
->rx_opt
.tstamp_ok
= 1;
5648 tp
->tcp_header_len
=
5649 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5650 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5651 tcp_store_ts_recent(tp
);
5653 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5656 if (tcp_is_sack(tp
) && sysctl_tcp_fack
)
5657 tcp_enable_fack(tp
);
5660 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5661 tcp_initialize_rcv_mss(sk
);
5663 /* Remember, tcp_poll() does not lock socket!
5664 * Change state from SYN-SENT only after copied_seq
5665 * is initialized. */
5666 tp
->copied_seq
= tp
->rcv_nxt
;
5670 tcp_finish_connect(sk
, skb
);
5672 if ((tp
->syn_fastopen
|| tp
->syn_data
) &&
5673 tcp_rcv_fastopen_synack(sk
, skb
, &foc
))
5676 if (sk
->sk_write_pending
||
5677 icsk
->icsk_accept_queue
.rskq_defer_accept
||
5678 icsk
->icsk_ack
.pingpong
) {
5679 /* Save one ACK. Data will be ready after
5680 * several ticks, if write_pending is set.
5682 * It may be deleted, but with this feature tcpdumps
5683 * look so _wonderfully_ clever, that I was not able
5684 * to stand against the temptation 8) --ANK
5686 inet_csk_schedule_ack(sk
);
5687 icsk
->icsk_ack
.lrcvtime
= tcp_time_stamp
;
5688 tcp_enter_quickack_mode(sk
);
5689 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
5690 TCP_DELACK_MAX
, TCP_RTO_MAX
);
5701 /* No ACK in the segment */
5705 * "If the RST bit is set
5707 * Otherwise (no ACK) drop the segment and return."
5710 goto discard_and_undo
;
5714 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&&
5715 tcp_paws_reject(&tp
->rx_opt
, 0))
5716 goto discard_and_undo
;
5719 /* We see SYN without ACK. It is attempt of
5720 * simultaneous connect with crossed SYNs.
5721 * Particularly, it can be connect to self.
5723 tcp_set_state(sk
, TCP_SYN_RECV
);
5725 if (tp
->rx_opt
.saw_tstamp
) {
5726 tp
->rx_opt
.tstamp_ok
= 1;
5727 tcp_store_ts_recent(tp
);
5728 tp
->tcp_header_len
=
5729 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5731 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5734 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5735 tp
->copied_seq
= tp
->rcv_nxt
;
5736 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5738 /* RFC1323: The window in SYN & SYN/ACK segments is
5741 tp
->snd_wnd
= ntohs(th
->window
);
5742 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
5743 tp
->max_window
= tp
->snd_wnd
;
5745 tcp_ecn_rcv_syn(tp
, th
);
5748 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5749 tcp_initialize_rcv_mss(sk
);
5751 tcp_send_synack(sk
);
5753 /* Note, we could accept data and URG from this segment.
5754 * There are no obstacles to make this (except that we must
5755 * either change tcp_recvmsg() to prevent it from returning data
5756 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5758 * However, if we ignore data in ACKless segments sometimes,
5759 * we have no reasons to accept it sometimes.
5760 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5761 * is not flawless. So, discard packet for sanity.
5762 * Uncomment this return to process the data.
5769 /* "fifth, if neither of the SYN or RST bits is set then
5770 * drop the segment and return."
5774 tcp_clear_options(&tp
->rx_opt
);
5775 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5779 tcp_clear_options(&tp
->rx_opt
);
5780 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5785 * This function implements the receiving procedure of RFC 793 for
5786 * all states except ESTABLISHED and TIME_WAIT.
5787 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5788 * address independent.
5791 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
)
5793 struct tcp_sock
*tp
= tcp_sk(sk
);
5794 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5795 const struct tcphdr
*th
= tcp_hdr(skb
);
5796 struct request_sock
*req
;
5800 tp
->rx_opt
.saw_tstamp
= 0;
5802 switch (sk
->sk_state
) {
5816 if (icsk
->icsk_af_ops
->conn_request(sk
, skb
) < 0)
5819 /* Now we have several options: In theory there is
5820 * nothing else in the frame. KA9Q has an option to
5821 * send data with the syn, BSD accepts data with the
5822 * syn up to the [to be] advertised window and
5823 * Solaris 2.1 gives you a protocol error. For now
5824 * we just ignore it, that fits the spec precisely
5825 * and avoids incompatibilities. It would be nice in
5826 * future to drop through and process the data.
5828 * Now that TTCP is starting to be used we ought to
5830 * But, this leaves one open to an easy denial of
5831 * service attack, and SYN cookies can't defend
5832 * against this problem. So, we drop the data
5833 * in the interest of security over speed unless
5834 * it's still in use.
5842 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
);
5846 /* Do step6 onward by hand. */
5847 tcp_urg(sk
, skb
, th
);
5849 tcp_data_snd_check(sk
);
5853 req
= tp
->fastopen_rsk
;
5855 WARN_ON_ONCE(sk
->sk_state
!= TCP_SYN_RECV
&&
5856 sk
->sk_state
!= TCP_FIN_WAIT1
);
5858 if (!tcp_check_req(sk
, skb
, req
, true))
5862 if (!th
->ack
&& !th
->rst
&& !th
->syn
)
5865 if (!tcp_validate_incoming(sk
, skb
, th
, 0))
5868 /* step 5: check the ACK field */
5869 acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
|
5870 FLAG_UPDATE_TS_RECENT
) > 0;
5872 switch (sk
->sk_state
) {
5878 tcp_synack_rtt_meas(sk
, req
);
5880 /* Once we leave TCP_SYN_RECV, we no longer need req
5884 tp
->total_retrans
= req
->num_retrans
;
5885 reqsk_fastopen_remove(sk
, req
, false);
5887 /* Make sure socket is routed, for correct metrics. */
5888 icsk
->icsk_af_ops
->rebuild_header(sk
);
5889 tcp_init_congestion_control(sk
);
5892 tp
->copied_seq
= tp
->rcv_nxt
;
5893 tcp_init_buffer_space(sk
);
5896 tcp_set_state(sk
, TCP_ESTABLISHED
);
5897 sk
->sk_state_change(sk
);
5899 /* Note, that this wakeup is only for marginal crossed SYN case.
5900 * Passively open sockets are not waked up, because
5901 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5904 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5906 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
5907 tp
->snd_wnd
= ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
;
5908 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5910 if (tp
->rx_opt
.tstamp_ok
)
5911 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5914 /* Re-arm the timer because data may have been sent out.
5915 * This is similar to the regular data transmission case
5916 * when new data has just been ack'ed.
5918 * (TFO) - we could try to be more aggressive and
5919 * retransmitting any data sooner based on when they
5924 tcp_init_metrics(sk
);
5926 tcp_update_pacing_rate(sk
);
5928 /* Prevent spurious tcp_cwnd_restart() on first data packet */
5929 tp
->lsndtime
= tcp_time_stamp
;
5931 tcp_initialize_rcv_mss(sk
);
5932 tcp_fast_path_on(tp
);
5935 case TCP_FIN_WAIT1
: {
5936 struct dst_entry
*dst
;
5939 /* If we enter the TCP_FIN_WAIT1 state and we are a
5940 * Fast Open socket and this is the first acceptable
5941 * ACK we have received, this would have acknowledged
5942 * our SYNACK so stop the SYNACK timer.
5945 /* Return RST if ack_seq is invalid.
5946 * Note that RFC793 only says to generate a
5947 * DUPACK for it but for TCP Fast Open it seems
5948 * better to treat this case like TCP_SYN_RECV
5953 /* We no longer need the request sock. */
5954 reqsk_fastopen_remove(sk
, req
, false);
5957 if (tp
->snd_una
!= tp
->write_seq
)
5960 tcp_set_state(sk
, TCP_FIN_WAIT2
);
5961 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
5963 dst
= __sk_dst_get(sk
);
5967 if (!sock_flag(sk
, SOCK_DEAD
)) {
5968 /* Wake up lingering close() */
5969 sk
->sk_state_change(sk
);
5973 if (tp
->linger2
< 0 ||
5974 (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5975 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
))) {
5977 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
5981 tmo
= tcp_fin_time(sk
);
5982 if (tmo
> TCP_TIMEWAIT_LEN
) {
5983 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
5984 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
5985 /* Bad case. We could lose such FIN otherwise.
5986 * It is not a big problem, but it looks confusing
5987 * and not so rare event. We still can lose it now,
5988 * if it spins in bh_lock_sock(), but it is really
5991 inet_csk_reset_keepalive_timer(sk
, tmo
);
5993 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
6000 if (tp
->snd_una
== tp
->write_seq
) {
6001 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
6007 if (tp
->snd_una
== tp
->write_seq
) {
6008 tcp_update_metrics(sk
);
6015 /* step 6: check the URG bit */
6016 tcp_urg(sk
, skb
, th
);
6018 /* step 7: process the segment text */
6019 switch (sk
->sk_state
) {
6020 case TCP_CLOSE_WAIT
:
6023 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
6027 /* RFC 793 says to queue data in these states,
6028 * RFC 1122 says we MUST send a reset.
6029 * BSD 4.4 also does reset.
6031 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
6032 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
6033 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
6034 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
6040 case TCP_ESTABLISHED
:
6041 tcp_data_queue(sk
, skb
);
6046 /* tcp_data could move socket to TIME-WAIT */
6047 if (sk
->sk_state
!= TCP_CLOSE
) {
6048 tcp_data_snd_check(sk
);
6049 tcp_ack_snd_check(sk
);
6058 EXPORT_SYMBOL(tcp_rcv_state_process
);
6060 static inline void pr_drop_req(struct request_sock
*req
, __u16 port
, int family
)
6062 struct inet_request_sock
*ireq
= inet_rsk(req
);
6064 if (family
== AF_INET
)
6065 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6066 &ireq
->ir_rmt_addr
, port
);
6067 #if IS_ENABLED(CONFIG_IPV6)
6068 else if (family
== AF_INET6
)
6069 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6070 &ireq
->ir_v6_rmt_addr
, port
);
6074 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6076 * If we receive a SYN packet with these bits set, it means a
6077 * network is playing bad games with TOS bits. In order to
6078 * avoid possible false congestion notifications, we disable
6079 * TCP ECN negotiation.
6081 * Exception: tcp_ca wants ECN. This is required for DCTCP
6082 * congestion control: Linux DCTCP asserts ECT on all packets,
6083 * including SYN, which is most optimal solution; however,
6084 * others, such as FreeBSD do not.
6086 static void tcp_ecn_create_request(struct request_sock
*req
,
6087 const struct sk_buff
*skb
,
6088 const struct sock
*listen_sk
,
6089 const struct dst_entry
*dst
)
6091 const struct tcphdr
*th
= tcp_hdr(skb
);
6092 const struct net
*net
= sock_net(listen_sk
);
6093 bool th_ecn
= th
->ece
&& th
->cwr
;
6100 ect
= !INET_ECN_is_not_ect(TCP_SKB_CB(skb
)->ip_dsfield
);
6101 ecn_ok_dst
= dst_feature(dst
, DST_FEATURE_ECN_MASK
);
6102 ecn_ok
= net
->ipv4
.sysctl_tcp_ecn
|| ecn_ok_dst
;
6104 if ((!ect
&& ecn_ok
) || tcp_ca_needs_ecn(listen_sk
) ||
6105 (ecn_ok_dst
& DST_FEATURE_ECN_CA
))
6106 inet_rsk(req
)->ecn_ok
= 1;
6109 static void tcp_openreq_init(struct request_sock
*req
,
6110 const struct tcp_options_received
*rx_opt
,
6111 struct sk_buff
*skb
, const struct sock
*sk
)
6113 struct inet_request_sock
*ireq
= inet_rsk(req
);
6115 req
->rsk_rcv_wnd
= 0; /* So that tcp_send_synack() knows! */
6117 tcp_rsk(req
)->rcv_isn
= TCP_SKB_CB(skb
)->seq
;
6118 tcp_rsk(req
)->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
6119 skb_mstamp_get(&tcp_rsk(req
)->snt_synack
);
6120 tcp_rsk(req
)->last_oow_ack_time
= 0;
6121 req
->mss
= rx_opt
->mss_clamp
;
6122 req
->ts_recent
= rx_opt
->saw_tstamp
? rx_opt
->rcv_tsval
: 0;
6123 ireq
->tstamp_ok
= rx_opt
->tstamp_ok
;
6124 ireq
->sack_ok
= rx_opt
->sack_ok
;
6125 ireq
->snd_wscale
= rx_opt
->snd_wscale
;
6126 ireq
->wscale_ok
= rx_opt
->wscale_ok
;
6129 ireq
->ir_rmt_port
= tcp_hdr(skb
)->source
;
6130 ireq
->ir_num
= ntohs(tcp_hdr(skb
)->dest
);
6131 ireq
->ir_mark
= inet_request_mark(sk
, skb
);
6134 struct request_sock
*inet_reqsk_alloc(const struct request_sock_ops
*ops
,
6135 struct sock
*sk_listener
,
6136 bool attach_listener
)
6138 struct request_sock
*req
= reqsk_alloc(ops
, sk_listener
,
6142 struct inet_request_sock
*ireq
= inet_rsk(req
);
6144 kmemcheck_annotate_bitfield(ireq
, flags
);
6146 atomic64_set(&ireq
->ir_cookie
, 0);
6147 ireq
->ireq_state
= TCP_NEW_SYN_RECV
;
6148 write_pnet(&ireq
->ireq_net
, sock_net(sk_listener
));
6149 ireq
->ireq_family
= sk_listener
->sk_family
;
6154 EXPORT_SYMBOL(inet_reqsk_alloc
);
6157 * Return true if a syncookie should be sent
6159 static bool tcp_syn_flood_action(const struct sock
*sk
,
6160 const struct sk_buff
*skb
,
6163 struct request_sock_queue
*queue
= &inet_csk(sk
)->icsk_accept_queue
;
6164 const char *msg
= "Dropping request";
6165 bool want_cookie
= false;
6166 struct net
*net
= sock_net(sk
);
6168 #ifdef CONFIG_SYN_COOKIES
6169 if (net
->ipv4
.sysctl_tcp_syncookies
) {
6170 msg
= "Sending cookies";
6172 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPREQQFULLDOCOOKIES
);
6175 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPREQQFULLDROP
);
6177 if (!queue
->synflood_warned
&&
6178 net
->ipv4
.sysctl_tcp_syncookies
!= 2 &&
6179 xchg(&queue
->synflood_warned
, 1) == 0)
6180 pr_info("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6181 proto
, ntohs(tcp_hdr(skb
)->dest
), msg
);
6186 static void tcp_reqsk_record_syn(const struct sock
*sk
,
6187 struct request_sock
*req
,
6188 const struct sk_buff
*skb
)
6190 if (tcp_sk(sk
)->save_syn
) {
6191 u32 len
= skb_network_header_len(skb
) + tcp_hdrlen(skb
);
6194 copy
= kmalloc(len
+ sizeof(u32
), GFP_ATOMIC
);
6197 memcpy(©
[1], skb_network_header(skb
), len
);
6198 req
->saved_syn
= copy
;
6203 int tcp_conn_request(struct request_sock_ops
*rsk_ops
,
6204 const struct tcp_request_sock_ops
*af_ops
,
6205 struct sock
*sk
, struct sk_buff
*skb
)
6207 struct tcp_fastopen_cookie foc
= { .len
= -1 };
6208 __u32 isn
= TCP_SKB_CB(skb
)->tcp_tw_isn
;
6209 struct tcp_options_received tmp_opt
;
6210 struct tcp_sock
*tp
= tcp_sk(sk
);
6211 struct net
*net
= sock_net(sk
);
6212 struct sock
*fastopen_sk
= NULL
;
6213 struct dst_entry
*dst
= NULL
;
6214 struct request_sock
*req
;
6215 bool want_cookie
= false;
6218 /* TW buckets are converted to open requests without
6219 * limitations, they conserve resources and peer is
6220 * evidently real one.
6222 if ((net
->ipv4
.sysctl_tcp_syncookies
== 2 ||
6223 inet_csk_reqsk_queue_is_full(sk
)) && !isn
) {
6224 want_cookie
= tcp_syn_flood_action(sk
, skb
, rsk_ops
->slab_name
);
6230 /* Accept backlog is full. If we have already queued enough
6231 * of warm entries in syn queue, drop request. It is better than
6232 * clogging syn queue with openreqs with exponentially increasing
6235 if (sk_acceptq_is_full(sk
) && inet_csk_reqsk_queue_young(sk
) > 1) {
6236 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_LISTENOVERFLOWS
);
6240 req
= inet_reqsk_alloc(rsk_ops
, sk
, !want_cookie
);
6244 tcp_rsk(req
)->af_specific
= af_ops
;
6246 tcp_clear_options(&tmp_opt
);
6247 tmp_opt
.mss_clamp
= af_ops
->mss_clamp
;
6248 tmp_opt
.user_mss
= tp
->rx_opt
.user_mss
;
6249 tcp_parse_options(skb
, &tmp_opt
, 0, want_cookie
? NULL
: &foc
);
6251 if (want_cookie
&& !tmp_opt
.saw_tstamp
)
6252 tcp_clear_options(&tmp_opt
);
6254 tmp_opt
.tstamp_ok
= tmp_opt
.saw_tstamp
;
6255 tcp_openreq_init(req
, &tmp_opt
, skb
, sk
);
6257 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6258 inet_rsk(req
)->ir_iif
= inet_request_bound_dev_if(sk
, skb
);
6260 af_ops
->init_req(req
, sk
, skb
);
6262 if (security_inet_conn_request(sk
, skb
, req
))
6265 if (!want_cookie
&& !isn
) {
6266 /* VJ's idea. We save last timestamp seen
6267 * from the destination in peer table, when entering
6268 * state TIME-WAIT, and check against it before
6269 * accepting new connection request.
6271 * If "isn" is not zero, this request hit alive
6272 * timewait bucket, so that all the necessary checks
6273 * are made in the function processing timewait state.
6275 if (tcp_death_row
.sysctl_tw_recycle
) {
6278 dst
= af_ops
->route_req(sk
, &fl
, req
, &strict
);
6280 if (dst
&& strict
&&
6281 !tcp_peer_is_proven(req
, dst
, true,
6282 tmp_opt
.saw_tstamp
)) {
6283 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSPASSIVEREJECTED
);
6284 goto drop_and_release
;
6287 /* Kill the following clause, if you dislike this way. */
6288 else if (!net
->ipv4
.sysctl_tcp_syncookies
&&
6289 (sysctl_max_syn_backlog
- inet_csk_reqsk_queue_len(sk
) <
6290 (sysctl_max_syn_backlog
>> 2)) &&
6291 !tcp_peer_is_proven(req
, dst
, false,
6292 tmp_opt
.saw_tstamp
)) {
6293 /* Without syncookies last quarter of
6294 * backlog is filled with destinations,
6295 * proven to be alive.
6296 * It means that we continue to communicate
6297 * to destinations, already remembered
6298 * to the moment of synflood.
6300 pr_drop_req(req
, ntohs(tcp_hdr(skb
)->source
),
6302 goto drop_and_release
;
6305 isn
= af_ops
->init_seq(skb
);
6308 dst
= af_ops
->route_req(sk
, &fl
, req
, NULL
);
6313 tcp_ecn_create_request(req
, skb
, sk
, dst
);
6316 isn
= cookie_init_sequence(af_ops
, sk
, skb
, &req
->mss
);
6317 req
->cookie_ts
= tmp_opt
.tstamp_ok
;
6318 if (!tmp_opt
.tstamp_ok
)
6319 inet_rsk(req
)->ecn_ok
= 0;
6322 tcp_rsk(req
)->snt_isn
= isn
;
6323 tcp_rsk(req
)->txhash
= net_tx_rndhash();
6324 tcp_openreq_init_rwin(req
, sk
, dst
);
6326 tcp_reqsk_record_syn(sk
, req
, skb
);
6327 fastopen_sk
= tcp_try_fastopen(sk
, skb
, req
, &foc
, dst
);
6330 af_ops
->send_synack(fastopen_sk
, dst
, &fl
, req
,
6332 /* Add the child socket directly into the accept queue */
6333 inet_csk_reqsk_queue_add(sk
, req
, fastopen_sk
);
6334 sk
->sk_data_ready(sk
);
6335 bh_unlock_sock(fastopen_sk
);
6336 sock_put(fastopen_sk
);
6338 tcp_rsk(req
)->tfo_listener
= false;
6340 inet_csk_reqsk_queue_hash_add(sk
, req
, TCP_TIMEOUT_INIT
);
6341 af_ops
->send_synack(sk
, dst
, &fl
, req
,
6342 &foc
, !want_cookie
);
6354 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_LISTENDROPS
);
6357 EXPORT_SYMBOL(tcp_conn_request
);