2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
23 * Pedro Roque : Fast Retransmit/Recovery.
25 * Retransmit queue handled by TCP.
26 * Better retransmit timer handling.
27 * New congestion avoidance.
31 * Eric : Fast Retransmit.
32 * Randy Scott : MSS option defines.
33 * Eric Schenk : Fixes to slow start algorithm.
34 * Eric Schenk : Yet another double ACK bug.
35 * Eric Schenk : Delayed ACK bug fixes.
36 * Eric Schenk : Floyd style fast retrans war avoidance.
37 * David S. Miller : Don't allow zero congestion window.
38 * Eric Schenk : Fix retransmitter so that it sends
39 * next packet on ack of previous packet.
40 * Andi Kleen : Moved open_request checking here
41 * and process RSTs for open_requests.
42 * Andi Kleen : Better prune_queue, and other fixes.
43 * Andrey Savochkin: Fix RTT measurements in the presence of
45 * Andrey Savochkin: Check sequence numbers correctly when
46 * removing SACKs due to in sequence incoming
48 * Andi Kleen: Make sure we never ack data there is not
49 * enough room for. Also make this condition
50 * a fatal error if it might still happen.
51 * Andi Kleen: Add tcp_measure_rcv_mss to make
52 * connections with MSS<min(MTU,ann. MSS)
53 * work without delayed acks.
54 * Andi Kleen: Process packets with PSH set in the
56 * J Hadi Salim: ECN support
59 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
60 * engine. Lots of bugs are found.
61 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
64 #define pr_fmt(fmt) "TCP: " fmt
67 #include <linux/slab.h>
68 #include <linux/module.h>
69 #include <linux/sysctl.h>
70 #include <linux/kernel.h>
71 #include <linux/prefetch.h>
74 #include <net/inet_common.h>
75 #include <linux/ipsec.h>
76 #include <asm/unaligned.h>
77 #include <linux/errqueue.h>
79 int sysctl_tcp_timestamps __read_mostly
= 1;
80 int sysctl_tcp_window_scaling __read_mostly
= 1;
81 int sysctl_tcp_sack __read_mostly
= 1;
82 int sysctl_tcp_fack __read_mostly
= 1;
83 int sysctl_tcp_reordering __read_mostly
= TCP_FASTRETRANS_THRESH
;
84 int sysctl_tcp_max_reordering __read_mostly
= 300;
85 EXPORT_SYMBOL(sysctl_tcp_reordering
);
86 int sysctl_tcp_dsack __read_mostly
= 1;
87 int sysctl_tcp_app_win __read_mostly
= 31;
88 int sysctl_tcp_adv_win_scale __read_mostly
= 1;
89 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale
);
91 /* rfc5961 challenge ack rate limiting */
92 int sysctl_tcp_challenge_ack_limit
= 100;
94 int sysctl_tcp_stdurg __read_mostly
;
95 int sysctl_tcp_rfc1337 __read_mostly
;
96 int sysctl_tcp_max_orphans __read_mostly
= NR_FILE
;
97 int sysctl_tcp_frto __read_mostly
= 2;
99 int sysctl_tcp_thin_dupack __read_mostly
;
101 int sysctl_tcp_moderate_rcvbuf __read_mostly
= 1;
102 int sysctl_tcp_early_retrans __read_mostly
= 3;
103 int sysctl_tcp_invalid_ratelimit __read_mostly
= HZ
/2;
105 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
106 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
107 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
108 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
109 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
110 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
111 #define FLAG_ECE 0x40 /* ECE in this ACK */
112 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
113 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
114 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
115 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
116 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
117 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
119 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
120 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
121 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
122 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
124 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
125 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
127 /* Adapt the MSS value used to make delayed ack decision to the
130 static void tcp_measure_rcv_mss(struct sock
*sk
, const struct sk_buff
*skb
)
132 struct inet_connection_sock
*icsk
= inet_csk(sk
);
133 const unsigned int lss
= icsk
->icsk_ack
.last_seg_size
;
136 icsk
->icsk_ack
.last_seg_size
= 0;
138 /* skb->len may jitter because of SACKs, even if peer
139 * sends good full-sized frames.
141 len
= skb_shinfo(skb
)->gso_size
? : skb
->len
;
142 if (len
>= icsk
->icsk_ack
.rcv_mss
) {
143 icsk
->icsk_ack
.rcv_mss
= len
;
145 /* Otherwise, we make more careful check taking into account,
146 * that SACKs block is variable.
148 * "len" is invariant segment length, including TCP header.
150 len
+= skb
->data
- skb_transport_header(skb
);
151 if (len
>= TCP_MSS_DEFAULT
+ sizeof(struct tcphdr
) ||
152 /* If PSH is not set, packet should be
153 * full sized, provided peer TCP is not badly broken.
154 * This observation (if it is correct 8)) allows
155 * to handle super-low mtu links fairly.
157 (len
>= TCP_MIN_MSS
+ sizeof(struct tcphdr
) &&
158 !(tcp_flag_word(tcp_hdr(skb
)) & TCP_REMNANT
))) {
159 /* Subtract also invariant (if peer is RFC compliant),
160 * tcp header plus fixed timestamp option length.
161 * Resulting "len" is MSS free of SACK jitter.
163 len
-= tcp_sk(sk
)->tcp_header_len
;
164 icsk
->icsk_ack
.last_seg_size
= len
;
166 icsk
->icsk_ack
.rcv_mss
= len
;
170 if (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
)
171 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED2
;
172 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
176 static void tcp_incr_quickack(struct sock
*sk
)
178 struct inet_connection_sock
*icsk
= inet_csk(sk
);
179 unsigned int quickacks
= tcp_sk(sk
)->rcv_wnd
/ (2 * icsk
->icsk_ack
.rcv_mss
);
183 if (quickacks
> icsk
->icsk_ack
.quick
)
184 icsk
->icsk_ack
.quick
= min(quickacks
, TCP_MAX_QUICKACKS
);
187 static void tcp_enter_quickack_mode(struct sock
*sk
)
189 struct inet_connection_sock
*icsk
= inet_csk(sk
);
190 tcp_incr_quickack(sk
);
191 icsk
->icsk_ack
.pingpong
= 0;
192 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
195 /* Send ACKs quickly, if "quick" count is not exhausted
196 * and the session is not interactive.
199 static inline bool tcp_in_quickack_mode(const struct sock
*sk
)
201 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
203 return icsk
->icsk_ack
.quick
&& !icsk
->icsk_ack
.pingpong
;
206 static void tcp_ecn_queue_cwr(struct tcp_sock
*tp
)
208 if (tp
->ecn_flags
& TCP_ECN_OK
)
209 tp
->ecn_flags
|= TCP_ECN_QUEUE_CWR
;
212 static void tcp_ecn_accept_cwr(struct tcp_sock
*tp
, const struct sk_buff
*skb
)
214 if (tcp_hdr(skb
)->cwr
)
215 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
218 static void tcp_ecn_withdraw_cwr(struct tcp_sock
*tp
)
220 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
223 static void __tcp_ecn_check_ce(struct tcp_sock
*tp
, const struct sk_buff
*skb
)
225 switch (TCP_SKB_CB(skb
)->ip_dsfield
& INET_ECN_MASK
) {
226 case INET_ECN_NOT_ECT
:
227 /* Funny extension: if ECT is not set on a segment,
228 * and we already seen ECT on a previous segment,
229 * it is probably a retransmit.
231 if (tp
->ecn_flags
& TCP_ECN_SEEN
)
232 tcp_enter_quickack_mode((struct sock
*)tp
);
235 if (tcp_ca_needs_ecn((struct sock
*)tp
))
236 tcp_ca_event((struct sock
*)tp
, CA_EVENT_ECN_IS_CE
);
238 if (!(tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)) {
239 /* Better not delay acks, sender can have a very low cwnd */
240 tcp_enter_quickack_mode((struct sock
*)tp
);
241 tp
->ecn_flags
|= TCP_ECN_DEMAND_CWR
;
243 tp
->ecn_flags
|= TCP_ECN_SEEN
;
246 if (tcp_ca_needs_ecn((struct sock
*)tp
))
247 tcp_ca_event((struct sock
*)tp
, CA_EVENT_ECN_NO_CE
);
248 tp
->ecn_flags
|= TCP_ECN_SEEN
;
253 static void tcp_ecn_check_ce(struct tcp_sock
*tp
, const struct sk_buff
*skb
)
255 if (tp
->ecn_flags
& TCP_ECN_OK
)
256 __tcp_ecn_check_ce(tp
, skb
);
259 static void tcp_ecn_rcv_synack(struct tcp_sock
*tp
, const struct tcphdr
*th
)
261 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| th
->cwr
))
262 tp
->ecn_flags
&= ~TCP_ECN_OK
;
265 static void tcp_ecn_rcv_syn(struct tcp_sock
*tp
, const struct tcphdr
*th
)
267 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| !th
->cwr
))
268 tp
->ecn_flags
&= ~TCP_ECN_OK
;
271 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock
*tp
, const struct tcphdr
*th
)
273 if (th
->ece
&& !th
->syn
&& (tp
->ecn_flags
& TCP_ECN_OK
))
278 /* Buffer size and advertised window tuning.
280 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
283 static void tcp_sndbuf_expand(struct sock
*sk
)
285 const struct tcp_sock
*tp
= tcp_sk(sk
);
289 /* Worst case is non GSO/TSO : each frame consumes one skb
290 * and skb->head is kmalloced using power of two area of memory
292 per_mss
= max_t(u32
, tp
->rx_opt
.mss_clamp
, tp
->mss_cache
) +
294 SKB_DATA_ALIGN(sizeof(struct skb_shared_info
));
296 per_mss
= roundup_pow_of_two(per_mss
) +
297 SKB_DATA_ALIGN(sizeof(struct sk_buff
));
299 nr_segs
= max_t(u32
, TCP_INIT_CWND
, tp
->snd_cwnd
);
300 nr_segs
= max_t(u32
, nr_segs
, tp
->reordering
+ 1);
302 /* Fast Recovery (RFC 5681 3.2) :
303 * Cubic needs 1.7 factor, rounded to 2 to include
304 * extra cushion (application might react slowly to POLLOUT)
306 sndmem
= 2 * nr_segs
* per_mss
;
308 if (sk
->sk_sndbuf
< sndmem
)
309 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
312 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
314 * All tcp_full_space() is split to two parts: "network" buffer, allocated
315 * forward and advertised in receiver window (tp->rcv_wnd) and
316 * "application buffer", required to isolate scheduling/application
317 * latencies from network.
318 * window_clamp is maximal advertised window. It can be less than
319 * tcp_full_space(), in this case tcp_full_space() - window_clamp
320 * is reserved for "application" buffer. The less window_clamp is
321 * the smoother our behaviour from viewpoint of network, but the lower
322 * throughput and the higher sensitivity of the connection to losses. 8)
324 * rcv_ssthresh is more strict window_clamp used at "slow start"
325 * phase to predict further behaviour of this connection.
326 * It is used for two goals:
327 * - to enforce header prediction at sender, even when application
328 * requires some significant "application buffer". It is check #1.
329 * - to prevent pruning of receive queue because of misprediction
330 * of receiver window. Check #2.
332 * The scheme does not work when sender sends good segments opening
333 * window and then starts to feed us spaghetti. But it should work
334 * in common situations. Otherwise, we have to rely on queue collapsing.
337 /* Slow part of check#2. */
338 static int __tcp_grow_window(const struct sock
*sk
, const struct sk_buff
*skb
)
340 struct tcp_sock
*tp
= tcp_sk(sk
);
342 int truesize
= tcp_win_from_space(skb
->truesize
) >> 1;
343 int window
= tcp_win_from_space(sysctl_tcp_rmem
[2]) >> 1;
345 while (tp
->rcv_ssthresh
<= window
) {
346 if (truesize
<= skb
->len
)
347 return 2 * inet_csk(sk
)->icsk_ack
.rcv_mss
;
355 static void tcp_grow_window(struct sock
*sk
, const struct sk_buff
*skb
)
357 struct tcp_sock
*tp
= tcp_sk(sk
);
360 if (tp
->rcv_ssthresh
< tp
->window_clamp
&&
361 (int)tp
->rcv_ssthresh
< tcp_space(sk
) &&
362 !sk_under_memory_pressure(sk
)) {
365 /* Check #2. Increase window, if skb with such overhead
366 * will fit to rcvbuf in future.
368 if (tcp_win_from_space(skb
->truesize
) <= skb
->len
)
369 incr
= 2 * tp
->advmss
;
371 incr
= __tcp_grow_window(sk
, skb
);
374 incr
= max_t(int, incr
, 2 * skb
->len
);
375 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
+ incr
,
377 inet_csk(sk
)->icsk_ack
.quick
|= 1;
382 /* 3. Tuning rcvbuf, when connection enters established state. */
383 static void tcp_fixup_rcvbuf(struct sock
*sk
)
385 u32 mss
= tcp_sk(sk
)->advmss
;
388 rcvmem
= 2 * SKB_TRUESIZE(mss
+ MAX_TCP_HEADER
) *
389 tcp_default_init_rwnd(mss
);
391 /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
392 * Allow enough cushion so that sender is not limited by our window
394 if (sysctl_tcp_moderate_rcvbuf
)
397 if (sk
->sk_rcvbuf
< rcvmem
)
398 sk
->sk_rcvbuf
= min(rcvmem
, sysctl_tcp_rmem
[2]);
401 /* 4. Try to fixup all. It is made immediately after connection enters
404 void tcp_init_buffer_space(struct sock
*sk
)
406 struct tcp_sock
*tp
= tcp_sk(sk
);
409 if (!(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
))
410 tcp_fixup_rcvbuf(sk
);
411 if (!(sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
))
412 tcp_sndbuf_expand(sk
);
414 tp
->rcvq_space
.space
= tp
->rcv_wnd
;
415 tp
->rcvq_space
.time
= tcp_time_stamp
;
416 tp
->rcvq_space
.seq
= tp
->copied_seq
;
418 maxwin
= tcp_full_space(sk
);
420 if (tp
->window_clamp
>= maxwin
) {
421 tp
->window_clamp
= maxwin
;
423 if (sysctl_tcp_app_win
&& maxwin
> 4 * tp
->advmss
)
424 tp
->window_clamp
= max(maxwin
-
425 (maxwin
>> sysctl_tcp_app_win
),
429 /* Force reservation of one segment. */
430 if (sysctl_tcp_app_win
&&
431 tp
->window_clamp
> 2 * tp
->advmss
&&
432 tp
->window_clamp
+ tp
->advmss
> maxwin
)
433 tp
->window_clamp
= max(2 * tp
->advmss
, maxwin
- tp
->advmss
);
435 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, tp
->window_clamp
);
436 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
439 /* 5. Recalculate window clamp after socket hit its memory bounds. */
440 static void tcp_clamp_window(struct sock
*sk
)
442 struct tcp_sock
*tp
= tcp_sk(sk
);
443 struct inet_connection_sock
*icsk
= inet_csk(sk
);
445 icsk
->icsk_ack
.quick
= 0;
447 if (sk
->sk_rcvbuf
< sysctl_tcp_rmem
[2] &&
448 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
) &&
449 !sk_under_memory_pressure(sk
) &&
450 sk_memory_allocated(sk
) < sk_prot_mem_limits(sk
, 0)) {
451 sk
->sk_rcvbuf
= min(atomic_read(&sk
->sk_rmem_alloc
),
454 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
)
455 tp
->rcv_ssthresh
= min(tp
->window_clamp
, 2U * tp
->advmss
);
458 /* Initialize RCV_MSS value.
459 * RCV_MSS is an our guess about MSS used by the peer.
460 * We haven't any direct information about the MSS.
461 * It's better to underestimate the RCV_MSS rather than overestimate.
462 * Overestimations make us ACKing less frequently than needed.
463 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
465 void tcp_initialize_rcv_mss(struct sock
*sk
)
467 const struct tcp_sock
*tp
= tcp_sk(sk
);
468 unsigned int hint
= min_t(unsigned int, tp
->advmss
, tp
->mss_cache
);
470 hint
= min(hint
, tp
->rcv_wnd
/ 2);
471 hint
= min(hint
, TCP_MSS_DEFAULT
);
472 hint
= max(hint
, TCP_MIN_MSS
);
474 inet_csk(sk
)->icsk_ack
.rcv_mss
= hint
;
476 EXPORT_SYMBOL(tcp_initialize_rcv_mss
);
478 /* Receiver "autotuning" code.
480 * The algorithm for RTT estimation w/o timestamps is based on
481 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
482 * <http://public.lanl.gov/radiant/pubs.html#DRS>
484 * More detail on this code can be found at
485 * <http://staff.psc.edu/jheffner/>,
486 * though this reference is out of date. A new paper
489 static void tcp_rcv_rtt_update(struct tcp_sock
*tp
, u32 sample
, int win_dep
)
491 u32 new_sample
= tp
->rcv_rtt_est
.rtt
;
497 if (new_sample
!= 0) {
498 /* If we sample in larger samples in the non-timestamp
499 * case, we could grossly overestimate the RTT especially
500 * with chatty applications or bulk transfer apps which
501 * are stalled on filesystem I/O.
503 * Also, since we are only going for a minimum in the
504 * non-timestamp case, we do not smooth things out
505 * else with timestamps disabled convergence takes too
509 m
-= (new_sample
>> 3);
517 /* No previous measure. */
521 if (tp
->rcv_rtt_est
.rtt
!= new_sample
)
522 tp
->rcv_rtt_est
.rtt
= new_sample
;
525 static inline void tcp_rcv_rtt_measure(struct tcp_sock
*tp
)
527 if (tp
->rcv_rtt_est
.time
== 0)
529 if (before(tp
->rcv_nxt
, tp
->rcv_rtt_est
.seq
))
531 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rcv_rtt_est
.time
, 1);
534 tp
->rcv_rtt_est
.seq
= tp
->rcv_nxt
+ tp
->rcv_wnd
;
535 tp
->rcv_rtt_est
.time
= tcp_time_stamp
;
538 static inline void tcp_rcv_rtt_measure_ts(struct sock
*sk
,
539 const struct sk_buff
*skb
)
541 struct tcp_sock
*tp
= tcp_sk(sk
);
542 if (tp
->rx_opt
.rcv_tsecr
&&
543 (TCP_SKB_CB(skb
)->end_seq
-
544 TCP_SKB_CB(skb
)->seq
>= inet_csk(sk
)->icsk_ack
.rcv_mss
))
545 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
, 0);
549 * This function should be called every time data is copied to user space.
550 * It calculates the appropriate TCP receive buffer space.
552 void tcp_rcv_space_adjust(struct sock
*sk
)
554 struct tcp_sock
*tp
= tcp_sk(sk
);
558 time
= tcp_time_stamp
- tp
->rcvq_space
.time
;
559 if (time
< (tp
->rcv_rtt_est
.rtt
>> 3) || tp
->rcv_rtt_est
.rtt
== 0)
562 /* Number of bytes copied to user in last RTT */
563 copied
= tp
->copied_seq
- tp
->rcvq_space
.seq
;
564 if (copied
<= tp
->rcvq_space
.space
)
568 * copied = bytes received in previous RTT, our base window
569 * To cope with packet losses, we need a 2x factor
570 * To cope with slow start, and sender growing its cwin by 100 %
571 * every RTT, we need a 4x factor, because the ACK we are sending
572 * now is for the next RTT, not the current one :
573 * <prev RTT . ><current RTT .. ><next RTT .... >
576 if (sysctl_tcp_moderate_rcvbuf
&&
577 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
)) {
578 int rcvwin
, rcvmem
, rcvbuf
;
580 /* minimal window to cope with packet losses, assuming
581 * steady state. Add some cushion because of small variations.
583 rcvwin
= (copied
<< 1) + 16 * tp
->advmss
;
585 /* If rate increased by 25%,
586 * assume slow start, rcvwin = 3 * copied
587 * If rate increased by 50%,
588 * assume sender can use 2x growth, rcvwin = 4 * copied
591 tp
->rcvq_space
.space
+ (tp
->rcvq_space
.space
>> 2)) {
593 tp
->rcvq_space
.space
+ (tp
->rcvq_space
.space
>> 1))
596 rcvwin
+= (rcvwin
>> 1);
599 rcvmem
= SKB_TRUESIZE(tp
->advmss
+ MAX_TCP_HEADER
);
600 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
603 rcvbuf
= min(rcvwin
/ tp
->advmss
* rcvmem
, sysctl_tcp_rmem
[2]);
604 if (rcvbuf
> sk
->sk_rcvbuf
) {
605 sk
->sk_rcvbuf
= rcvbuf
;
607 /* Make the window clamp follow along. */
608 tp
->window_clamp
= rcvwin
;
611 tp
->rcvq_space
.space
= copied
;
614 tp
->rcvq_space
.seq
= tp
->copied_seq
;
615 tp
->rcvq_space
.time
= tcp_time_stamp
;
618 /* There is something which you must keep in mind when you analyze the
619 * behavior of the tp->ato delayed ack timeout interval. When a
620 * connection starts up, we want to ack as quickly as possible. The
621 * problem is that "good" TCP's do slow start at the beginning of data
622 * transmission. The means that until we send the first few ACK's the
623 * sender will sit on his end and only queue most of his data, because
624 * he can only send snd_cwnd unacked packets at any given time. For
625 * each ACK we send, he increments snd_cwnd and transmits more of his
628 static void tcp_event_data_recv(struct sock
*sk
, struct sk_buff
*skb
)
630 struct tcp_sock
*tp
= tcp_sk(sk
);
631 struct inet_connection_sock
*icsk
= inet_csk(sk
);
634 inet_csk_schedule_ack(sk
);
636 tcp_measure_rcv_mss(sk
, skb
);
638 tcp_rcv_rtt_measure(tp
);
640 now
= tcp_time_stamp
;
642 if (!icsk
->icsk_ack
.ato
) {
643 /* The _first_ data packet received, initialize
644 * delayed ACK engine.
646 tcp_incr_quickack(sk
);
647 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
649 int m
= now
- icsk
->icsk_ack
.lrcvtime
;
651 if (m
<= TCP_ATO_MIN
/ 2) {
652 /* The fastest case is the first. */
653 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + TCP_ATO_MIN
/ 2;
654 } else if (m
< icsk
->icsk_ack
.ato
) {
655 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + m
;
656 if (icsk
->icsk_ack
.ato
> icsk
->icsk_rto
)
657 icsk
->icsk_ack
.ato
= icsk
->icsk_rto
;
658 } else if (m
> icsk
->icsk_rto
) {
659 /* Too long gap. Apparently sender failed to
660 * restart window, so that we send ACKs quickly.
662 tcp_incr_quickack(sk
);
666 icsk
->icsk_ack
.lrcvtime
= now
;
668 tcp_ecn_check_ce(tp
, skb
);
671 tcp_grow_window(sk
, skb
);
674 /* Called to compute a smoothed rtt estimate. The data fed to this
675 * routine either comes from timestamps, or from segments that were
676 * known _not_ to have been retransmitted [see Karn/Partridge
677 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
678 * piece by Van Jacobson.
679 * NOTE: the next three routines used to be one big routine.
680 * To save cycles in the RFC 1323 implementation it was better to break
681 * it up into three procedures. -- erics
683 static void tcp_rtt_estimator(struct sock
*sk
, long mrtt_us
)
685 struct tcp_sock
*tp
= tcp_sk(sk
);
686 long m
= mrtt_us
; /* RTT */
687 u32 srtt
= tp
->srtt_us
;
689 /* The following amusing code comes from Jacobson's
690 * article in SIGCOMM '88. Note that rtt and mdev
691 * are scaled versions of rtt and mean deviation.
692 * This is designed to be as fast as possible
693 * m stands for "measurement".
695 * On a 1990 paper the rto value is changed to:
696 * RTO = rtt + 4 * mdev
698 * Funny. This algorithm seems to be very broken.
699 * These formulae increase RTO, when it should be decreased, increase
700 * too slowly, when it should be increased quickly, decrease too quickly
701 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
702 * does not matter how to _calculate_ it. Seems, it was trap
703 * that VJ failed to avoid. 8)
706 m
-= (srtt
>> 3); /* m is now error in rtt est */
707 srtt
+= m
; /* rtt = 7/8 rtt + 1/8 new */
709 m
= -m
; /* m is now abs(error) */
710 m
-= (tp
->mdev_us
>> 2); /* similar update on mdev */
711 /* This is similar to one of Eifel findings.
712 * Eifel blocks mdev updates when rtt decreases.
713 * This solution is a bit different: we use finer gain
714 * for mdev in this case (alpha*beta).
715 * Like Eifel it also prevents growth of rto,
716 * but also it limits too fast rto decreases,
717 * happening in pure Eifel.
722 m
-= (tp
->mdev_us
>> 2); /* similar update on mdev */
724 tp
->mdev_us
+= m
; /* mdev = 3/4 mdev + 1/4 new */
725 if (tp
->mdev_us
> tp
->mdev_max_us
) {
726 tp
->mdev_max_us
= tp
->mdev_us
;
727 if (tp
->mdev_max_us
> tp
->rttvar_us
)
728 tp
->rttvar_us
= tp
->mdev_max_us
;
730 if (after(tp
->snd_una
, tp
->rtt_seq
)) {
731 if (tp
->mdev_max_us
< tp
->rttvar_us
)
732 tp
->rttvar_us
-= (tp
->rttvar_us
- tp
->mdev_max_us
) >> 2;
733 tp
->rtt_seq
= tp
->snd_nxt
;
734 tp
->mdev_max_us
= tcp_rto_min_us(sk
);
737 /* no previous measure. */
738 srtt
= m
<< 3; /* take the measured time to be rtt */
739 tp
->mdev_us
= m
<< 1; /* make sure rto = 3*rtt */
740 tp
->rttvar_us
= max(tp
->mdev_us
, tcp_rto_min_us(sk
));
741 tp
->mdev_max_us
= tp
->rttvar_us
;
742 tp
->rtt_seq
= tp
->snd_nxt
;
744 tp
->srtt_us
= max(1U, srtt
);
747 /* Set the sk_pacing_rate to allow proper sizing of TSO packets.
748 * Note: TCP stack does not yet implement pacing.
749 * FQ packet scheduler can be used to implement cheap but effective
750 * TCP pacing, to smooth the burst on large writes when packets
751 * in flight is significantly lower than cwnd (or rwin)
753 static void tcp_update_pacing_rate(struct sock
*sk
)
755 const struct tcp_sock
*tp
= tcp_sk(sk
);
758 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
759 rate
= (u64
)tp
->mss_cache
* 2 * (USEC_PER_SEC
<< 3);
761 rate
*= max(tp
->snd_cwnd
, tp
->packets_out
);
763 if (likely(tp
->srtt_us
))
764 do_div(rate
, tp
->srtt_us
);
766 /* ACCESS_ONCE() is needed because sch_fq fetches sk_pacing_rate
767 * without any lock. We want to make sure compiler wont store
768 * intermediate values in this location.
770 ACCESS_ONCE(sk
->sk_pacing_rate
) = min_t(u64
, rate
,
771 sk
->sk_max_pacing_rate
);
774 /* Calculate rto without backoff. This is the second half of Van Jacobson's
775 * routine referred to above.
777 static void tcp_set_rto(struct sock
*sk
)
779 const struct tcp_sock
*tp
= tcp_sk(sk
);
780 /* Old crap is replaced with new one. 8)
783 * 1. If rtt variance happened to be less 50msec, it is hallucination.
784 * It cannot be less due to utterly erratic ACK generation made
785 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
786 * to do with delayed acks, because at cwnd>2 true delack timeout
787 * is invisible. Actually, Linux-2.4 also generates erratic
788 * ACKs in some circumstances.
790 inet_csk(sk
)->icsk_rto
= __tcp_set_rto(tp
);
792 /* 2. Fixups made earlier cannot be right.
793 * If we do not estimate RTO correctly without them,
794 * all the algo is pure shit and should be replaced
795 * with correct one. It is exactly, which we pretend to do.
798 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
799 * guarantees that rto is higher.
804 __u32
tcp_init_cwnd(const struct tcp_sock
*tp
, const struct dst_entry
*dst
)
806 __u32 cwnd
= (dst
? dst_metric(dst
, RTAX_INITCWND
) : 0);
809 cwnd
= TCP_INIT_CWND
;
810 return min_t(__u32
, cwnd
, tp
->snd_cwnd_clamp
);
814 * Packet counting of FACK is based on in-order assumptions, therefore TCP
815 * disables it when reordering is detected
817 void tcp_disable_fack(struct tcp_sock
*tp
)
819 /* RFC3517 uses different metric in lost marker => reset on change */
821 tp
->lost_skb_hint
= NULL
;
822 tp
->rx_opt
.sack_ok
&= ~TCP_FACK_ENABLED
;
825 /* Take a notice that peer is sending D-SACKs */
826 static void tcp_dsack_seen(struct tcp_sock
*tp
)
828 tp
->rx_opt
.sack_ok
|= TCP_DSACK_SEEN
;
831 static void tcp_update_reordering(struct sock
*sk
, const int metric
,
834 struct tcp_sock
*tp
= tcp_sk(sk
);
835 if (metric
> tp
->reordering
) {
838 tp
->reordering
= min(sysctl_tcp_max_reordering
, metric
);
840 /* This exciting event is worth to be remembered. 8) */
842 mib_idx
= LINUX_MIB_TCPTSREORDER
;
843 else if (tcp_is_reno(tp
))
844 mib_idx
= LINUX_MIB_TCPRENOREORDER
;
845 else if (tcp_is_fack(tp
))
846 mib_idx
= LINUX_MIB_TCPFACKREORDER
;
848 mib_idx
= LINUX_MIB_TCPSACKREORDER
;
850 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
851 #if FASTRETRANS_DEBUG > 1
852 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
853 tp
->rx_opt
.sack_ok
, inet_csk(sk
)->icsk_ca_state
,
857 tp
->undo_marker
? tp
->undo_retrans
: 0);
859 tcp_disable_fack(tp
);
863 tcp_disable_early_retrans(tp
);
866 /* This must be called before lost_out is incremented */
867 static void tcp_verify_retransmit_hint(struct tcp_sock
*tp
, struct sk_buff
*skb
)
869 if ((tp
->retransmit_skb_hint
== NULL
) ||
870 before(TCP_SKB_CB(skb
)->seq
,
871 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
872 tp
->retransmit_skb_hint
= skb
;
875 after(TCP_SKB_CB(skb
)->end_seq
, tp
->retransmit_high
))
876 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
879 static void tcp_skb_mark_lost(struct tcp_sock
*tp
, struct sk_buff
*skb
)
881 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
882 tcp_verify_retransmit_hint(tp
, skb
);
884 tp
->lost_out
+= tcp_skb_pcount(skb
);
885 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
889 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock
*tp
,
892 tcp_verify_retransmit_hint(tp
, skb
);
894 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
895 tp
->lost_out
+= tcp_skb_pcount(skb
);
896 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
900 /* This procedure tags the retransmission queue when SACKs arrive.
902 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
903 * Packets in queue with these bits set are counted in variables
904 * sacked_out, retrans_out and lost_out, correspondingly.
906 * Valid combinations are:
907 * Tag InFlight Description
908 * 0 1 - orig segment is in flight.
909 * S 0 - nothing flies, orig reached receiver.
910 * L 0 - nothing flies, orig lost by net.
911 * R 2 - both orig and retransmit are in flight.
912 * L|R 1 - orig is lost, retransmit is in flight.
913 * S|R 1 - orig reached receiver, retrans is still in flight.
914 * (L|S|R is logically valid, it could occur when L|R is sacked,
915 * but it is equivalent to plain S and code short-curcuits it to S.
916 * L|S is logically invalid, it would mean -1 packet in flight 8))
918 * These 6 states form finite state machine, controlled by the following events:
919 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
920 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
921 * 3. Loss detection event of two flavors:
922 * A. Scoreboard estimator decided the packet is lost.
923 * A'. Reno "three dupacks" marks head of queue lost.
924 * A''. Its FACK modification, head until snd.fack is lost.
925 * B. SACK arrives sacking SND.NXT at the moment, when the
926 * segment was retransmitted.
927 * 4. D-SACK added new rule: D-SACK changes any tag to S.
929 * It is pleasant to note, that state diagram turns out to be commutative,
930 * so that we are allowed not to be bothered by order of our actions,
931 * when multiple events arrive simultaneously. (see the function below).
933 * Reordering detection.
934 * --------------------
935 * Reordering metric is maximal distance, which a packet can be displaced
936 * in packet stream. With SACKs we can estimate it:
938 * 1. SACK fills old hole and the corresponding segment was not
939 * ever retransmitted -> reordering. Alas, we cannot use it
940 * when segment was retransmitted.
941 * 2. The last flaw is solved with D-SACK. D-SACK arrives
942 * for retransmitted and already SACKed segment -> reordering..
943 * Both of these heuristics are not used in Loss state, when we cannot
944 * account for retransmits accurately.
946 * SACK block validation.
947 * ----------------------
949 * SACK block range validation checks that the received SACK block fits to
950 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
951 * Note that SND.UNA is not included to the range though being valid because
952 * it means that the receiver is rather inconsistent with itself reporting
953 * SACK reneging when it should advance SND.UNA. Such SACK block this is
954 * perfectly valid, however, in light of RFC2018 which explicitly states
955 * that "SACK block MUST reflect the newest segment. Even if the newest
956 * segment is going to be discarded ...", not that it looks very clever
957 * in case of head skb. Due to potentional receiver driven attacks, we
958 * choose to avoid immediate execution of a walk in write queue due to
959 * reneging and defer head skb's loss recovery to standard loss recovery
960 * procedure that will eventually trigger (nothing forbids us doing this).
962 * Implements also blockage to start_seq wrap-around. Problem lies in the
963 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
964 * there's no guarantee that it will be before snd_nxt (n). The problem
965 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
968 * <- outs wnd -> <- wrapzone ->
969 * u e n u_w e_w s n_w
971 * |<------------+------+----- TCP seqno space --------------+---------->|
972 * ...-- <2^31 ->| |<--------...
973 * ...---- >2^31 ------>| |<--------...
975 * Current code wouldn't be vulnerable but it's better still to discard such
976 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
977 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
978 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
979 * equal to the ideal case (infinite seqno space without wrap caused issues).
981 * With D-SACK the lower bound is extended to cover sequence space below
982 * SND.UNA down to undo_marker, which is the last point of interest. Yet
983 * again, D-SACK block must not to go across snd_una (for the same reason as
984 * for the normal SACK blocks, explained above). But there all simplicity
985 * ends, TCP might receive valid D-SACKs below that. As long as they reside
986 * fully below undo_marker they do not affect behavior in anyway and can
987 * therefore be safely ignored. In rare cases (which are more or less
988 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
989 * fragmentation and packet reordering past skb's retransmission. To consider
990 * them correctly, the acceptable range must be extended even more though
991 * the exact amount is rather hard to quantify. However, tp->max_window can
992 * be used as an exaggerated estimate.
994 static bool tcp_is_sackblock_valid(struct tcp_sock
*tp
, bool is_dsack
,
995 u32 start_seq
, u32 end_seq
)
997 /* Too far in future, or reversed (interpretation is ambiguous) */
998 if (after(end_seq
, tp
->snd_nxt
) || !before(start_seq
, end_seq
))
1001 /* Nasty start_seq wrap-around check (see comments above) */
1002 if (!before(start_seq
, tp
->snd_nxt
))
1005 /* In outstanding window? ...This is valid exit for D-SACKs too.
1006 * start_seq == snd_una is non-sensical (see comments above)
1008 if (after(start_seq
, tp
->snd_una
))
1011 if (!is_dsack
|| !tp
->undo_marker
)
1014 /* ...Then it's D-SACK, and must reside below snd_una completely */
1015 if (after(end_seq
, tp
->snd_una
))
1018 if (!before(start_seq
, tp
->undo_marker
))
1022 if (!after(end_seq
, tp
->undo_marker
))
1025 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1026 * start_seq < undo_marker and end_seq >= undo_marker.
1028 return !before(start_seq
, end_seq
- tp
->max_window
);
1031 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1032 * Event "B". Later note: FACK people cheated me again 8), we have to account
1033 * for reordering! Ugly, but should help.
1035 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1036 * less than what is now known to be received by the other end (derived from
1037 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1038 * retransmitted skbs to avoid some costly processing per ACKs.
1040 static void tcp_mark_lost_retrans(struct sock
*sk
)
1042 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1043 struct tcp_sock
*tp
= tcp_sk(sk
);
1044 struct sk_buff
*skb
;
1046 u32 new_low_seq
= tp
->snd_nxt
;
1047 u32 received_upto
= tcp_highest_sack_seq(tp
);
1049 if (!tcp_is_fack(tp
) || !tp
->retrans_out
||
1050 !after(received_upto
, tp
->lost_retrans_low
) ||
1051 icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
1054 tcp_for_write_queue(skb
, sk
) {
1055 u32 ack_seq
= TCP_SKB_CB(skb
)->ack_seq
;
1057 if (skb
== tcp_send_head(sk
))
1059 if (cnt
== tp
->retrans_out
)
1061 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1064 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
))
1067 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1068 * constraint here (see above) but figuring out that at
1069 * least tp->reordering SACK blocks reside between ack_seq
1070 * and received_upto is not easy task to do cheaply with
1071 * the available datastructures.
1073 * Whether FACK should check here for tp->reordering segs
1074 * in-between one could argue for either way (it would be
1075 * rather simple to implement as we could count fack_count
1076 * during the walk and do tp->fackets_out - fack_count).
1078 if (after(received_upto
, ack_seq
)) {
1079 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1080 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1082 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
1083 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSTRETRANSMIT
);
1085 if (before(ack_seq
, new_low_seq
))
1086 new_low_seq
= ack_seq
;
1087 cnt
+= tcp_skb_pcount(skb
);
1091 if (tp
->retrans_out
)
1092 tp
->lost_retrans_low
= new_low_seq
;
1095 static bool tcp_check_dsack(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1096 struct tcp_sack_block_wire
*sp
, int num_sacks
,
1099 struct tcp_sock
*tp
= tcp_sk(sk
);
1100 u32 start_seq_0
= get_unaligned_be32(&sp
[0].start_seq
);
1101 u32 end_seq_0
= get_unaligned_be32(&sp
[0].end_seq
);
1102 bool dup_sack
= false;
1104 if (before(start_seq_0
, TCP_SKB_CB(ack_skb
)->ack_seq
)) {
1107 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPDSACKRECV
);
1108 } else if (num_sacks
> 1) {
1109 u32 end_seq_1
= get_unaligned_be32(&sp
[1].end_seq
);
1110 u32 start_seq_1
= get_unaligned_be32(&sp
[1].start_seq
);
1112 if (!after(end_seq_0
, end_seq_1
) &&
1113 !before(start_seq_0
, start_seq_1
)) {
1116 NET_INC_STATS_BH(sock_net(sk
),
1117 LINUX_MIB_TCPDSACKOFORECV
);
1121 /* D-SACK for already forgotten data... Do dumb counting. */
1122 if (dup_sack
&& tp
->undo_marker
&& tp
->undo_retrans
> 0 &&
1123 !after(end_seq_0
, prior_snd_una
) &&
1124 after(end_seq_0
, tp
->undo_marker
))
1130 struct tcp_sacktag_state
{
1133 long rtt_us
; /* RTT measured by SACKing never-retransmitted data */
1137 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1138 * the incoming SACK may not exactly match but we can find smaller MSS
1139 * aligned portion of it that matches. Therefore we might need to fragment
1140 * which may fail and creates some hassle (caller must handle error case
1143 * FIXME: this could be merged to shift decision code
1145 static int tcp_match_skb_to_sack(struct sock
*sk
, struct sk_buff
*skb
,
1146 u32 start_seq
, u32 end_seq
)
1150 unsigned int pkt_len
;
1153 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1154 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1156 if (tcp_skb_pcount(skb
) > 1 && !in_sack
&&
1157 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
1158 mss
= tcp_skb_mss(skb
);
1159 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1162 pkt_len
= start_seq
- TCP_SKB_CB(skb
)->seq
;
1166 pkt_len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1171 /* Round if necessary so that SACKs cover only full MSSes
1172 * and/or the remaining small portion (if present)
1174 if (pkt_len
> mss
) {
1175 unsigned int new_len
= (pkt_len
/ mss
) * mss
;
1176 if (!in_sack
&& new_len
< pkt_len
) {
1178 if (new_len
>= skb
->len
)
1183 err
= tcp_fragment(sk
, skb
, pkt_len
, mss
, GFP_ATOMIC
);
1191 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1192 static u8
tcp_sacktag_one(struct sock
*sk
,
1193 struct tcp_sacktag_state
*state
, u8 sacked
,
1194 u32 start_seq
, u32 end_seq
,
1195 int dup_sack
, int pcount
,
1196 const struct skb_mstamp
*xmit_time
)
1198 struct tcp_sock
*tp
= tcp_sk(sk
);
1199 int fack_count
= state
->fack_count
;
1201 /* Account D-SACK for retransmitted packet. */
1202 if (dup_sack
&& (sacked
& TCPCB_RETRANS
)) {
1203 if (tp
->undo_marker
&& tp
->undo_retrans
> 0 &&
1204 after(end_seq
, tp
->undo_marker
))
1206 if (sacked
& TCPCB_SACKED_ACKED
)
1207 state
->reord
= min(fack_count
, state
->reord
);
1210 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1211 if (!after(end_seq
, tp
->snd_una
))
1214 if (!(sacked
& TCPCB_SACKED_ACKED
)) {
1215 if (sacked
& TCPCB_SACKED_RETRANS
) {
1216 /* If the segment is not tagged as lost,
1217 * we do not clear RETRANS, believing
1218 * that retransmission is still in flight.
1220 if (sacked
& TCPCB_LOST
) {
1221 sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1222 tp
->lost_out
-= pcount
;
1223 tp
->retrans_out
-= pcount
;
1226 if (!(sacked
& TCPCB_RETRANS
)) {
1227 /* New sack for not retransmitted frame,
1228 * which was in hole. It is reordering.
1230 if (before(start_seq
,
1231 tcp_highest_sack_seq(tp
)))
1232 state
->reord
= min(fack_count
,
1234 if (!after(end_seq
, tp
->high_seq
))
1235 state
->flag
|= FLAG_ORIG_SACK_ACKED
;
1236 /* Pick the earliest sequence sacked for RTT */
1237 if (state
->rtt_us
< 0) {
1238 struct skb_mstamp now
;
1240 skb_mstamp_get(&now
);
1241 state
->rtt_us
= skb_mstamp_us_delta(&now
,
1246 if (sacked
& TCPCB_LOST
) {
1247 sacked
&= ~TCPCB_LOST
;
1248 tp
->lost_out
-= pcount
;
1252 sacked
|= TCPCB_SACKED_ACKED
;
1253 state
->flag
|= FLAG_DATA_SACKED
;
1254 tp
->sacked_out
+= pcount
;
1256 fack_count
+= pcount
;
1258 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1259 if (!tcp_is_fack(tp
) && (tp
->lost_skb_hint
!= NULL
) &&
1260 before(start_seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
))
1261 tp
->lost_cnt_hint
+= pcount
;
1263 if (fack_count
> tp
->fackets_out
)
1264 tp
->fackets_out
= fack_count
;
1267 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1268 * frames and clear it. undo_retrans is decreased above, L|R frames
1269 * are accounted above as well.
1271 if (dup_sack
&& (sacked
& TCPCB_SACKED_RETRANS
)) {
1272 sacked
&= ~TCPCB_SACKED_RETRANS
;
1273 tp
->retrans_out
-= pcount
;
1279 /* Shift newly-SACKed bytes from this skb to the immediately previous
1280 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1282 static bool tcp_shifted_skb(struct sock
*sk
, struct sk_buff
*skb
,
1283 struct tcp_sacktag_state
*state
,
1284 unsigned int pcount
, int shifted
, int mss
,
1287 struct tcp_sock
*tp
= tcp_sk(sk
);
1288 struct sk_buff
*prev
= tcp_write_queue_prev(sk
, skb
);
1289 u32 start_seq
= TCP_SKB_CB(skb
)->seq
; /* start of newly-SACKed */
1290 u32 end_seq
= start_seq
+ shifted
; /* end of newly-SACKed */
1294 /* Adjust counters and hints for the newly sacked sequence
1295 * range but discard the return value since prev is already
1296 * marked. We must tag the range first because the seq
1297 * advancement below implicitly advances
1298 * tcp_highest_sack_seq() when skb is highest_sack.
1300 tcp_sacktag_one(sk
, state
, TCP_SKB_CB(skb
)->sacked
,
1301 start_seq
, end_seq
, dup_sack
, pcount
,
1304 if (skb
== tp
->lost_skb_hint
)
1305 tp
->lost_cnt_hint
+= pcount
;
1307 TCP_SKB_CB(prev
)->end_seq
+= shifted
;
1308 TCP_SKB_CB(skb
)->seq
+= shifted
;
1310 tcp_skb_pcount_add(prev
, pcount
);
1311 BUG_ON(tcp_skb_pcount(skb
) < pcount
);
1312 tcp_skb_pcount_add(skb
, -pcount
);
1314 /* When we're adding to gso_segs == 1, gso_size will be zero,
1315 * in theory this shouldn't be necessary but as long as DSACK
1316 * code can come after this skb later on it's better to keep
1317 * setting gso_size to something.
1319 if (!skb_shinfo(prev
)->gso_size
) {
1320 skb_shinfo(prev
)->gso_size
= mss
;
1321 skb_shinfo(prev
)->gso_type
= sk
->sk_gso_type
;
1324 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1325 if (tcp_skb_pcount(skb
) <= 1) {
1326 skb_shinfo(skb
)->gso_size
= 0;
1327 skb_shinfo(skb
)->gso_type
= 0;
1330 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1331 TCP_SKB_CB(prev
)->sacked
|= (TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
);
1334 BUG_ON(!tcp_skb_pcount(skb
));
1335 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKSHIFTED
);
1339 /* Whole SKB was eaten :-) */
1341 if (skb
== tp
->retransmit_skb_hint
)
1342 tp
->retransmit_skb_hint
= prev
;
1343 if (skb
== tp
->lost_skb_hint
) {
1344 tp
->lost_skb_hint
= prev
;
1345 tp
->lost_cnt_hint
-= tcp_skb_pcount(prev
);
1348 TCP_SKB_CB(prev
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
;
1349 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
1350 TCP_SKB_CB(prev
)->end_seq
++;
1352 if (skb
== tcp_highest_sack(sk
))
1353 tcp_advance_highest_sack(sk
, skb
);
1355 tcp_unlink_write_queue(skb
, sk
);
1356 sk_wmem_free_skb(sk
, skb
);
1358 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKMERGED
);
1363 /* I wish gso_size would have a bit more sane initialization than
1364 * something-or-zero which complicates things
1366 static int tcp_skb_seglen(const struct sk_buff
*skb
)
1368 return tcp_skb_pcount(skb
) == 1 ? skb
->len
: tcp_skb_mss(skb
);
1371 /* Shifting pages past head area doesn't work */
1372 static int skb_can_shift(const struct sk_buff
*skb
)
1374 return !skb_headlen(skb
) && skb_is_nonlinear(skb
);
1377 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1380 static struct sk_buff
*tcp_shift_skb_data(struct sock
*sk
, struct sk_buff
*skb
,
1381 struct tcp_sacktag_state
*state
,
1382 u32 start_seq
, u32 end_seq
,
1385 struct tcp_sock
*tp
= tcp_sk(sk
);
1386 struct sk_buff
*prev
;
1392 if (!sk_can_gso(sk
))
1395 /* Normally R but no L won't result in plain S */
1397 (TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_RETRANS
)) == TCPCB_SACKED_RETRANS
)
1399 if (!skb_can_shift(skb
))
1401 /* This frame is about to be dropped (was ACKed). */
1402 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1405 /* Can only happen with delayed DSACK + discard craziness */
1406 if (unlikely(skb
== tcp_write_queue_head(sk
)))
1408 prev
= tcp_write_queue_prev(sk
, skb
);
1410 if ((TCP_SKB_CB(prev
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
)
1413 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1414 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1418 pcount
= tcp_skb_pcount(skb
);
1419 mss
= tcp_skb_seglen(skb
);
1421 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1422 * drop this restriction as unnecessary
1424 if (mss
!= tcp_skb_seglen(prev
))
1427 if (!after(TCP_SKB_CB(skb
)->end_seq
, start_seq
))
1429 /* CHECKME: This is non-MSS split case only?, this will
1430 * cause skipped skbs due to advancing loop btw, original
1431 * has that feature too
1433 if (tcp_skb_pcount(skb
) <= 1)
1436 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1438 /* TODO: head merge to next could be attempted here
1439 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1440 * though it might not be worth of the additional hassle
1442 * ...we can probably just fallback to what was done
1443 * previously. We could try merging non-SACKed ones
1444 * as well but it probably isn't going to buy off
1445 * because later SACKs might again split them, and
1446 * it would make skb timestamp tracking considerably
1452 len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1454 BUG_ON(len
> skb
->len
);
1456 /* MSS boundaries should be honoured or else pcount will
1457 * severely break even though it makes things bit trickier.
1458 * Optimize common case to avoid most of the divides
1460 mss
= tcp_skb_mss(skb
);
1462 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1463 * drop this restriction as unnecessary
1465 if (mss
!= tcp_skb_seglen(prev
))
1470 } else if (len
< mss
) {
1478 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1479 if (!after(TCP_SKB_CB(skb
)->seq
+ len
, tp
->snd_una
))
1482 if (!skb_shift(prev
, skb
, len
))
1484 if (!tcp_shifted_skb(sk
, skb
, state
, pcount
, len
, mss
, dup_sack
))
1487 /* Hole filled allows collapsing with the next as well, this is very
1488 * useful when hole on every nth skb pattern happens
1490 if (prev
== tcp_write_queue_tail(sk
))
1492 skb
= tcp_write_queue_next(sk
, prev
);
1494 if (!skb_can_shift(skb
) ||
1495 (skb
== tcp_send_head(sk
)) ||
1496 ((TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
) ||
1497 (mss
!= tcp_skb_seglen(skb
)))
1501 if (skb_shift(prev
, skb
, len
)) {
1502 pcount
+= tcp_skb_pcount(skb
);
1503 tcp_shifted_skb(sk
, skb
, state
, tcp_skb_pcount(skb
), len
, mss
, 0);
1507 state
->fack_count
+= pcount
;
1514 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKSHIFTFALLBACK
);
1518 static struct sk_buff
*tcp_sacktag_walk(struct sk_buff
*skb
, struct sock
*sk
,
1519 struct tcp_sack_block
*next_dup
,
1520 struct tcp_sacktag_state
*state
,
1521 u32 start_seq
, u32 end_seq
,
1524 struct tcp_sock
*tp
= tcp_sk(sk
);
1525 struct sk_buff
*tmp
;
1527 tcp_for_write_queue_from(skb
, sk
) {
1529 bool dup_sack
= dup_sack_in
;
1531 if (skb
== tcp_send_head(sk
))
1534 /* queue is in-order => we can short-circuit the walk early */
1535 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1538 if ((next_dup
!= NULL
) &&
1539 before(TCP_SKB_CB(skb
)->seq
, next_dup
->end_seq
)) {
1540 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1541 next_dup
->start_seq
,
1547 /* skb reference here is a bit tricky to get right, since
1548 * shifting can eat and free both this skb and the next,
1549 * so not even _safe variant of the loop is enough.
1552 tmp
= tcp_shift_skb_data(sk
, skb
, state
,
1553 start_seq
, end_seq
, dup_sack
);
1562 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1568 if (unlikely(in_sack
< 0))
1572 TCP_SKB_CB(skb
)->sacked
=
1575 TCP_SKB_CB(skb
)->sacked
,
1576 TCP_SKB_CB(skb
)->seq
,
1577 TCP_SKB_CB(skb
)->end_seq
,
1579 tcp_skb_pcount(skb
),
1582 if (!before(TCP_SKB_CB(skb
)->seq
,
1583 tcp_highest_sack_seq(tp
)))
1584 tcp_advance_highest_sack(sk
, skb
);
1587 state
->fack_count
+= tcp_skb_pcount(skb
);
1592 /* Avoid all extra work that is being done by sacktag while walking in
1595 static struct sk_buff
*tcp_sacktag_skip(struct sk_buff
*skb
, struct sock
*sk
,
1596 struct tcp_sacktag_state
*state
,
1599 tcp_for_write_queue_from(skb
, sk
) {
1600 if (skb
== tcp_send_head(sk
))
1603 if (after(TCP_SKB_CB(skb
)->end_seq
, skip_to_seq
))
1606 state
->fack_count
+= tcp_skb_pcount(skb
);
1611 static struct sk_buff
*tcp_maybe_skipping_dsack(struct sk_buff
*skb
,
1613 struct tcp_sack_block
*next_dup
,
1614 struct tcp_sacktag_state
*state
,
1617 if (next_dup
== NULL
)
1620 if (before(next_dup
->start_seq
, skip_to_seq
)) {
1621 skb
= tcp_sacktag_skip(skb
, sk
, state
, next_dup
->start_seq
);
1622 skb
= tcp_sacktag_walk(skb
, sk
, NULL
, state
,
1623 next_dup
->start_seq
, next_dup
->end_seq
,
1630 static int tcp_sack_cache_ok(const struct tcp_sock
*tp
, const struct tcp_sack_block
*cache
)
1632 return cache
< tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1636 tcp_sacktag_write_queue(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1637 u32 prior_snd_una
, long *sack_rtt_us
)
1639 struct tcp_sock
*tp
= tcp_sk(sk
);
1640 const unsigned char *ptr
= (skb_transport_header(ack_skb
) +
1641 TCP_SKB_CB(ack_skb
)->sacked
);
1642 struct tcp_sack_block_wire
*sp_wire
= (struct tcp_sack_block_wire
*)(ptr
+2);
1643 struct tcp_sack_block sp
[TCP_NUM_SACKS
];
1644 struct tcp_sack_block
*cache
;
1645 struct tcp_sacktag_state state
;
1646 struct sk_buff
*skb
;
1647 int num_sacks
= min(TCP_NUM_SACKS
, (ptr
[1] - TCPOLEN_SACK_BASE
) >> 3);
1649 bool found_dup_sack
= false;
1651 int first_sack_index
;
1654 state
.reord
= tp
->packets_out
;
1657 if (!tp
->sacked_out
) {
1658 if (WARN_ON(tp
->fackets_out
))
1659 tp
->fackets_out
= 0;
1660 tcp_highest_sack_reset(sk
);
1663 found_dup_sack
= tcp_check_dsack(sk
, ack_skb
, sp_wire
,
1664 num_sacks
, prior_snd_una
);
1666 state
.flag
|= FLAG_DSACKING_ACK
;
1668 /* Eliminate too old ACKs, but take into
1669 * account more or less fresh ones, they can
1670 * contain valid SACK info.
1672 if (before(TCP_SKB_CB(ack_skb
)->ack_seq
, prior_snd_una
- tp
->max_window
))
1675 if (!tp
->packets_out
)
1679 first_sack_index
= 0;
1680 for (i
= 0; i
< num_sacks
; i
++) {
1681 bool dup_sack
= !i
&& found_dup_sack
;
1683 sp
[used_sacks
].start_seq
= get_unaligned_be32(&sp_wire
[i
].start_seq
);
1684 sp
[used_sacks
].end_seq
= get_unaligned_be32(&sp_wire
[i
].end_seq
);
1686 if (!tcp_is_sackblock_valid(tp
, dup_sack
,
1687 sp
[used_sacks
].start_seq
,
1688 sp
[used_sacks
].end_seq
)) {
1692 if (!tp
->undo_marker
)
1693 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDNOUNDO
;
1695 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDOLD
;
1697 /* Don't count olds caused by ACK reordering */
1698 if ((TCP_SKB_CB(ack_skb
)->ack_seq
!= tp
->snd_una
) &&
1699 !after(sp
[used_sacks
].end_seq
, tp
->snd_una
))
1701 mib_idx
= LINUX_MIB_TCPSACKDISCARD
;
1704 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
1706 first_sack_index
= -1;
1710 /* Ignore very old stuff early */
1711 if (!after(sp
[used_sacks
].end_seq
, prior_snd_una
))
1717 /* order SACK blocks to allow in order walk of the retrans queue */
1718 for (i
= used_sacks
- 1; i
> 0; i
--) {
1719 for (j
= 0; j
< i
; j
++) {
1720 if (after(sp
[j
].start_seq
, sp
[j
+ 1].start_seq
)) {
1721 swap(sp
[j
], sp
[j
+ 1]);
1723 /* Track where the first SACK block goes to */
1724 if (j
== first_sack_index
)
1725 first_sack_index
= j
+ 1;
1730 skb
= tcp_write_queue_head(sk
);
1731 state
.fack_count
= 0;
1734 if (!tp
->sacked_out
) {
1735 /* It's already past, so skip checking against it */
1736 cache
= tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1738 cache
= tp
->recv_sack_cache
;
1739 /* Skip empty blocks in at head of the cache */
1740 while (tcp_sack_cache_ok(tp
, cache
) && !cache
->start_seq
&&
1745 while (i
< used_sacks
) {
1746 u32 start_seq
= sp
[i
].start_seq
;
1747 u32 end_seq
= sp
[i
].end_seq
;
1748 bool dup_sack
= (found_dup_sack
&& (i
== first_sack_index
));
1749 struct tcp_sack_block
*next_dup
= NULL
;
1751 if (found_dup_sack
&& ((i
+ 1) == first_sack_index
))
1752 next_dup
= &sp
[i
+ 1];
1754 /* Skip too early cached blocks */
1755 while (tcp_sack_cache_ok(tp
, cache
) &&
1756 !before(start_seq
, cache
->end_seq
))
1759 /* Can skip some work by looking recv_sack_cache? */
1760 if (tcp_sack_cache_ok(tp
, cache
) && !dup_sack
&&
1761 after(end_seq
, cache
->start_seq
)) {
1764 if (before(start_seq
, cache
->start_seq
)) {
1765 skb
= tcp_sacktag_skip(skb
, sk
, &state
,
1767 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
,
1774 /* Rest of the block already fully processed? */
1775 if (!after(end_seq
, cache
->end_seq
))
1778 skb
= tcp_maybe_skipping_dsack(skb
, sk
, next_dup
,
1782 /* ...tail remains todo... */
1783 if (tcp_highest_sack_seq(tp
) == cache
->end_seq
) {
1784 /* ...but better entrypoint exists! */
1785 skb
= tcp_highest_sack(sk
);
1788 state
.fack_count
= tp
->fackets_out
;
1793 skb
= tcp_sacktag_skip(skb
, sk
, &state
, cache
->end_seq
);
1794 /* Check overlap against next cached too (past this one already) */
1799 if (!before(start_seq
, tcp_highest_sack_seq(tp
))) {
1800 skb
= tcp_highest_sack(sk
);
1803 state
.fack_count
= tp
->fackets_out
;
1805 skb
= tcp_sacktag_skip(skb
, sk
, &state
, start_seq
);
1808 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
, &state
,
1809 start_seq
, end_seq
, dup_sack
);
1815 /* Clear the head of the cache sack blocks so we can skip it next time */
1816 for (i
= 0; i
< ARRAY_SIZE(tp
->recv_sack_cache
) - used_sacks
; i
++) {
1817 tp
->recv_sack_cache
[i
].start_seq
= 0;
1818 tp
->recv_sack_cache
[i
].end_seq
= 0;
1820 for (j
= 0; j
< used_sacks
; j
++)
1821 tp
->recv_sack_cache
[i
++] = sp
[j
];
1823 tcp_mark_lost_retrans(sk
);
1825 tcp_verify_left_out(tp
);
1827 if ((state
.reord
< tp
->fackets_out
) &&
1828 ((inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Loss
) || tp
->undo_marker
))
1829 tcp_update_reordering(sk
, tp
->fackets_out
- state
.reord
, 0);
1833 #if FASTRETRANS_DEBUG > 0
1834 WARN_ON((int)tp
->sacked_out
< 0);
1835 WARN_ON((int)tp
->lost_out
< 0);
1836 WARN_ON((int)tp
->retrans_out
< 0);
1837 WARN_ON((int)tcp_packets_in_flight(tp
) < 0);
1839 *sack_rtt_us
= state
.rtt_us
;
1843 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1844 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1846 static bool tcp_limit_reno_sacked(struct tcp_sock
*tp
)
1850 holes
= max(tp
->lost_out
, 1U);
1851 holes
= min(holes
, tp
->packets_out
);
1853 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1854 tp
->sacked_out
= tp
->packets_out
- holes
;
1860 /* If we receive more dupacks than we expected counting segments
1861 * in assumption of absent reordering, interpret this as reordering.
1862 * The only another reason could be bug in receiver TCP.
1864 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1866 struct tcp_sock
*tp
= tcp_sk(sk
);
1867 if (tcp_limit_reno_sacked(tp
))
1868 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1871 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1873 static void tcp_add_reno_sack(struct sock
*sk
)
1875 struct tcp_sock
*tp
= tcp_sk(sk
);
1877 tcp_check_reno_reordering(sk
, 0);
1878 tcp_verify_left_out(tp
);
1881 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1883 static void tcp_remove_reno_sacks(struct sock
*sk
, int acked
)
1885 struct tcp_sock
*tp
= tcp_sk(sk
);
1888 /* One ACK acked hole. The rest eat duplicate ACKs. */
1889 if (acked
- 1 >= tp
->sacked_out
)
1892 tp
->sacked_out
-= acked
- 1;
1894 tcp_check_reno_reordering(sk
, acked
);
1895 tcp_verify_left_out(tp
);
1898 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1903 void tcp_clear_retrans(struct tcp_sock
*tp
)
1905 tp
->retrans_out
= 0;
1907 tp
->undo_marker
= 0;
1908 tp
->undo_retrans
= -1;
1909 tp
->fackets_out
= 0;
1913 static inline void tcp_init_undo(struct tcp_sock
*tp
)
1915 tp
->undo_marker
= tp
->snd_una
;
1916 /* Retransmission still in flight may cause DSACKs later. */
1917 tp
->undo_retrans
= tp
->retrans_out
? : -1;
1920 /* Enter Loss state. If we detect SACK reneging, forget all SACK information
1921 * and reset tags completely, otherwise preserve SACKs. If receiver
1922 * dropped its ofo queue, we will know this due to reneging detection.
1924 void tcp_enter_loss(struct sock
*sk
)
1926 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1927 struct tcp_sock
*tp
= tcp_sk(sk
);
1928 struct sk_buff
*skb
;
1929 bool new_recovery
= false;
1930 bool is_reneg
; /* is receiver reneging on SACKs? */
1932 /* Reduce ssthresh if it has not yet been made inside this window. */
1933 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
||
1934 !after(tp
->high_seq
, tp
->snd_una
) ||
1935 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
1936 new_recovery
= true;
1937 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1938 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1939 tcp_ca_event(sk
, CA_EVENT_LOSS
);
1943 tp
->snd_cwnd_cnt
= 0;
1944 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1946 tp
->retrans_out
= 0;
1949 if (tcp_is_reno(tp
))
1950 tcp_reset_reno_sack(tp
);
1952 skb
= tcp_write_queue_head(sk
);
1953 is_reneg
= skb
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
);
1955 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSACKRENEGING
);
1957 tp
->fackets_out
= 0;
1959 tcp_clear_all_retrans_hints(tp
);
1961 tcp_for_write_queue(skb
, sk
) {
1962 if (skb
== tcp_send_head(sk
))
1965 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
1966 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) || is_reneg
) {
1967 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
1968 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1969 tp
->lost_out
+= tcp_skb_pcount(skb
);
1970 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
1973 tcp_verify_left_out(tp
);
1975 /* Timeout in disordered state after receiving substantial DUPACKs
1976 * suggests that the degree of reordering is over-estimated.
1978 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
&&
1979 tp
->sacked_out
>= sysctl_tcp_reordering
)
1980 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1981 sysctl_tcp_reordering
);
1982 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1983 tp
->high_seq
= tp
->snd_nxt
;
1984 tcp_ecn_queue_cwr(tp
);
1986 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1987 * loss recovery is underway except recurring timeout(s) on
1988 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1990 tp
->frto
= sysctl_tcp_frto
&&
1991 (new_recovery
|| icsk
->icsk_retransmits
) &&
1992 !inet_csk(sk
)->icsk_mtup
.probe_size
;
1995 /* If ACK arrived pointing to a remembered SACK, it means that our
1996 * remembered SACKs do not reflect real state of receiver i.e.
1997 * receiver _host_ is heavily congested (or buggy).
1999 * To avoid big spurious retransmission bursts due to transient SACK
2000 * scoreboard oddities that look like reneging, we give the receiver a
2001 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2002 * restore sanity to the SACK scoreboard. If the apparent reneging
2003 * persists until this RTO then we'll clear the SACK scoreboard.
2005 static bool tcp_check_sack_reneging(struct sock
*sk
, int flag
)
2007 if (flag
& FLAG_SACK_RENEGING
) {
2008 struct tcp_sock
*tp
= tcp_sk(sk
);
2009 unsigned long delay
= max(usecs_to_jiffies(tp
->srtt_us
>> 4),
2010 msecs_to_jiffies(10));
2012 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2013 delay
, TCP_RTO_MAX
);
2019 static inline int tcp_fackets_out(const struct tcp_sock
*tp
)
2021 return tcp_is_reno(tp
) ? tp
->sacked_out
+ 1 : tp
->fackets_out
;
2024 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2025 * counter when SACK is enabled (without SACK, sacked_out is used for
2028 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2029 * segments up to the highest received SACK block so far and holes in
2032 * With reordering, holes may still be in flight, so RFC3517 recovery
2033 * uses pure sacked_out (total number of SACKed segments) even though
2034 * it violates the RFC that uses duplicate ACKs, often these are equal
2035 * but when e.g. out-of-window ACKs or packet duplication occurs,
2036 * they differ. Since neither occurs due to loss, TCP should really
2039 static inline int tcp_dupack_heuristics(const struct tcp_sock
*tp
)
2041 return tcp_is_fack(tp
) ? tp
->fackets_out
: tp
->sacked_out
+ 1;
2044 static bool tcp_pause_early_retransmit(struct sock
*sk
, int flag
)
2046 struct tcp_sock
*tp
= tcp_sk(sk
);
2047 unsigned long delay
;
2049 /* Delay early retransmit and entering fast recovery for
2050 * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
2051 * available, or RTO is scheduled to fire first.
2053 if (sysctl_tcp_early_retrans
< 2 || sysctl_tcp_early_retrans
> 3 ||
2054 (flag
& FLAG_ECE
) || !tp
->srtt_us
)
2057 delay
= max(usecs_to_jiffies(tp
->srtt_us
>> 5),
2058 msecs_to_jiffies(2));
2060 if (!time_after(inet_csk(sk
)->icsk_timeout
, (jiffies
+ delay
)))
2063 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_EARLY_RETRANS
, delay
,
2068 /* Linux NewReno/SACK/FACK/ECN state machine.
2069 * --------------------------------------
2071 * "Open" Normal state, no dubious events, fast path.
2072 * "Disorder" In all the respects it is "Open",
2073 * but requires a bit more attention. It is entered when
2074 * we see some SACKs or dupacks. It is split of "Open"
2075 * mainly to move some processing from fast path to slow one.
2076 * "CWR" CWND was reduced due to some Congestion Notification event.
2077 * It can be ECN, ICMP source quench, local device congestion.
2078 * "Recovery" CWND was reduced, we are fast-retransmitting.
2079 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2081 * tcp_fastretrans_alert() is entered:
2082 * - each incoming ACK, if state is not "Open"
2083 * - when arrived ACK is unusual, namely:
2088 * Counting packets in flight is pretty simple.
2090 * in_flight = packets_out - left_out + retrans_out
2092 * packets_out is SND.NXT-SND.UNA counted in packets.
2094 * retrans_out is number of retransmitted segments.
2096 * left_out is number of segments left network, but not ACKed yet.
2098 * left_out = sacked_out + lost_out
2100 * sacked_out: Packets, which arrived to receiver out of order
2101 * and hence not ACKed. With SACKs this number is simply
2102 * amount of SACKed data. Even without SACKs
2103 * it is easy to give pretty reliable estimate of this number,
2104 * counting duplicate ACKs.
2106 * lost_out: Packets lost by network. TCP has no explicit
2107 * "loss notification" feedback from network (for now).
2108 * It means that this number can be only _guessed_.
2109 * Actually, it is the heuristics to predict lossage that
2110 * distinguishes different algorithms.
2112 * F.e. after RTO, when all the queue is considered as lost,
2113 * lost_out = packets_out and in_flight = retrans_out.
2115 * Essentially, we have now two algorithms counting
2118 * FACK: It is the simplest heuristics. As soon as we decided
2119 * that something is lost, we decide that _all_ not SACKed
2120 * packets until the most forward SACK are lost. I.e.
2121 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2122 * It is absolutely correct estimate, if network does not reorder
2123 * packets. And it loses any connection to reality when reordering
2124 * takes place. We use FACK by default until reordering
2125 * is suspected on the path to this destination.
2127 * NewReno: when Recovery is entered, we assume that one segment
2128 * is lost (classic Reno). While we are in Recovery and
2129 * a partial ACK arrives, we assume that one more packet
2130 * is lost (NewReno). This heuristics are the same in NewReno
2133 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2134 * deflation etc. CWND is real congestion window, never inflated, changes
2135 * only according to classic VJ rules.
2137 * Really tricky (and requiring careful tuning) part of algorithm
2138 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2139 * The first determines the moment _when_ we should reduce CWND and,
2140 * hence, slow down forward transmission. In fact, it determines the moment
2141 * when we decide that hole is caused by loss, rather than by a reorder.
2143 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2144 * holes, caused by lost packets.
2146 * And the most logically complicated part of algorithm is undo
2147 * heuristics. We detect false retransmits due to both too early
2148 * fast retransmit (reordering) and underestimated RTO, analyzing
2149 * timestamps and D-SACKs. When we detect that some segments were
2150 * retransmitted by mistake and CWND reduction was wrong, we undo
2151 * window reduction and abort recovery phase. This logic is hidden
2152 * inside several functions named tcp_try_undo_<something>.
2155 /* This function decides, when we should leave Disordered state
2156 * and enter Recovery phase, reducing congestion window.
2158 * Main question: may we further continue forward transmission
2159 * with the same cwnd?
2161 static bool tcp_time_to_recover(struct sock
*sk
, int flag
)
2163 struct tcp_sock
*tp
= tcp_sk(sk
);
2166 /* Trick#1: The loss is proven. */
2170 /* Not-A-Trick#2 : Classic rule... */
2171 if (tcp_dupack_heuristics(tp
) > tp
->reordering
)
2174 /* Trick#4: It is still not OK... But will it be useful to delay
2177 packets_out
= tp
->packets_out
;
2178 if (packets_out
<= tp
->reordering
&&
2179 tp
->sacked_out
>= max_t(__u32
, packets_out
/2, sysctl_tcp_reordering
) &&
2180 !tcp_may_send_now(sk
)) {
2181 /* We have nothing to send. This connection is limited
2182 * either by receiver window or by application.
2187 /* If a thin stream is detected, retransmit after first
2188 * received dupack. Employ only if SACK is supported in order
2189 * to avoid possible corner-case series of spurious retransmissions
2190 * Use only if there are no unsent data.
2192 if ((tp
->thin_dupack
|| sysctl_tcp_thin_dupack
) &&
2193 tcp_stream_is_thin(tp
) && tcp_dupack_heuristics(tp
) > 1 &&
2194 tcp_is_sack(tp
) && !tcp_send_head(sk
))
2197 /* Trick#6: TCP early retransmit, per RFC5827. To avoid spurious
2198 * retransmissions due to small network reorderings, we implement
2199 * Mitigation A.3 in the RFC and delay the retransmission for a short
2200 * interval if appropriate.
2202 if (tp
->do_early_retrans
&& !tp
->retrans_out
&& tp
->sacked_out
&&
2203 (tp
->packets_out
>= (tp
->sacked_out
+ 1) && tp
->packets_out
< 4) &&
2204 !tcp_may_send_now(sk
))
2205 return !tcp_pause_early_retransmit(sk
, flag
);
2210 /* Detect loss in event "A" above by marking head of queue up as lost.
2211 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2212 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2213 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2214 * the maximum SACKed segments to pass before reaching this limit.
2216 static void tcp_mark_head_lost(struct sock
*sk
, int packets
, int mark_head
)
2218 struct tcp_sock
*tp
= tcp_sk(sk
);
2219 struct sk_buff
*skb
;
2223 /* Use SACK to deduce losses of new sequences sent during recovery */
2224 const u32 loss_high
= tcp_is_sack(tp
) ? tp
->snd_nxt
: tp
->high_seq
;
2226 WARN_ON(packets
> tp
->packets_out
);
2227 if (tp
->lost_skb_hint
) {
2228 skb
= tp
->lost_skb_hint
;
2229 cnt
= tp
->lost_cnt_hint
;
2230 /* Head already handled? */
2231 if (mark_head
&& skb
!= tcp_write_queue_head(sk
))
2234 skb
= tcp_write_queue_head(sk
);
2238 tcp_for_write_queue_from(skb
, sk
) {
2239 if (skb
== tcp_send_head(sk
))
2241 /* TODO: do this better */
2242 /* this is not the most efficient way to do this... */
2243 tp
->lost_skb_hint
= skb
;
2244 tp
->lost_cnt_hint
= cnt
;
2246 if (after(TCP_SKB_CB(skb
)->end_seq
, loss_high
))
2250 if (tcp_is_fack(tp
) || tcp_is_reno(tp
) ||
2251 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2252 cnt
+= tcp_skb_pcount(skb
);
2254 if (cnt
> packets
) {
2255 if ((tcp_is_sack(tp
) && !tcp_is_fack(tp
)) ||
2256 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
) ||
2257 (oldcnt
>= packets
))
2260 mss
= skb_shinfo(skb
)->gso_size
;
2261 err
= tcp_fragment(sk
, skb
, (packets
- oldcnt
) * mss
,
2268 tcp_skb_mark_lost(tp
, skb
);
2273 tcp_verify_left_out(tp
);
2276 /* Account newly detected lost packet(s) */
2278 static void tcp_update_scoreboard(struct sock
*sk
, int fast_rexmit
)
2280 struct tcp_sock
*tp
= tcp_sk(sk
);
2282 if (tcp_is_reno(tp
)) {
2283 tcp_mark_head_lost(sk
, 1, 1);
2284 } else if (tcp_is_fack(tp
)) {
2285 int lost
= tp
->fackets_out
- tp
->reordering
;
2288 tcp_mark_head_lost(sk
, lost
, 0);
2290 int sacked_upto
= tp
->sacked_out
- tp
->reordering
;
2291 if (sacked_upto
>= 0)
2292 tcp_mark_head_lost(sk
, sacked_upto
, 0);
2293 else if (fast_rexmit
)
2294 tcp_mark_head_lost(sk
, 1, 1);
2298 /* CWND moderation, preventing bursts due to too big ACKs
2299 * in dubious situations.
2301 static inline void tcp_moderate_cwnd(struct tcp_sock
*tp
)
2303 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
2304 tcp_packets_in_flight(tp
) + tcp_max_burst(tp
));
2305 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2308 /* Nothing was retransmitted or returned timestamp is less
2309 * than timestamp of the first retransmission.
2311 static inline bool tcp_packet_delayed(const struct tcp_sock
*tp
)
2313 return !tp
->retrans_stamp
||
2314 (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2315 before(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
));
2318 /* Undo procedures. */
2320 /* We can clear retrans_stamp when there are no retransmissions in the
2321 * window. It would seem that it is trivially available for us in
2322 * tp->retrans_out, however, that kind of assumptions doesn't consider
2323 * what will happen if errors occur when sending retransmission for the
2324 * second time. ...It could the that such segment has only
2325 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2326 * the head skb is enough except for some reneging corner cases that
2327 * are not worth the effort.
2329 * Main reason for all this complexity is the fact that connection dying
2330 * time now depends on the validity of the retrans_stamp, in particular,
2331 * that successive retransmissions of a segment must not advance
2332 * retrans_stamp under any conditions.
2334 static bool tcp_any_retrans_done(const struct sock
*sk
)
2336 const struct tcp_sock
*tp
= tcp_sk(sk
);
2337 struct sk_buff
*skb
;
2339 if (tp
->retrans_out
)
2342 skb
= tcp_write_queue_head(sk
);
2343 if (unlikely(skb
&& TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
))
2349 #if FASTRETRANS_DEBUG > 1
2350 static void DBGUNDO(struct sock
*sk
, const char *msg
)
2352 struct tcp_sock
*tp
= tcp_sk(sk
);
2353 struct inet_sock
*inet
= inet_sk(sk
);
2355 if (sk
->sk_family
== AF_INET
) {
2356 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2358 &inet
->inet_daddr
, ntohs(inet
->inet_dport
),
2359 tp
->snd_cwnd
, tcp_left_out(tp
),
2360 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2363 #if IS_ENABLED(CONFIG_IPV6)
2364 else if (sk
->sk_family
== AF_INET6
) {
2365 struct ipv6_pinfo
*np
= inet6_sk(sk
);
2366 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2368 &np
->daddr
, ntohs(inet
->inet_dport
),
2369 tp
->snd_cwnd
, tcp_left_out(tp
),
2370 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2376 #define DBGUNDO(x...) do { } while (0)
2379 static void tcp_undo_cwnd_reduction(struct sock
*sk
, bool unmark_loss
)
2381 struct tcp_sock
*tp
= tcp_sk(sk
);
2384 struct sk_buff
*skb
;
2386 tcp_for_write_queue(skb
, sk
) {
2387 if (skb
== tcp_send_head(sk
))
2389 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2392 tcp_clear_all_retrans_hints(tp
);
2395 if (tp
->prior_ssthresh
) {
2396 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2398 if (icsk
->icsk_ca_ops
->undo_cwnd
)
2399 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
2401 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
<< 1);
2403 if (tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
2404 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
2405 tcp_ecn_withdraw_cwr(tp
);
2408 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2410 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2411 tp
->undo_marker
= 0;
2414 static inline bool tcp_may_undo(const struct tcp_sock
*tp
)
2416 return tp
->undo_marker
&& (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
2419 /* People celebrate: "We love our President!" */
2420 static bool tcp_try_undo_recovery(struct sock
*sk
)
2422 struct tcp_sock
*tp
= tcp_sk(sk
);
2424 if (tcp_may_undo(tp
)) {
2427 /* Happy end! We did not retransmit anything
2428 * or our original transmission succeeded.
2430 DBGUNDO(sk
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
2431 tcp_undo_cwnd_reduction(sk
, false);
2432 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
2433 mib_idx
= LINUX_MIB_TCPLOSSUNDO
;
2435 mib_idx
= LINUX_MIB_TCPFULLUNDO
;
2437 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2439 if (tp
->snd_una
== tp
->high_seq
&& tcp_is_reno(tp
)) {
2440 /* Hold old state until something *above* high_seq
2441 * is ACKed. For Reno it is MUST to prevent false
2442 * fast retransmits (RFC2582). SACK TCP is safe. */
2443 tcp_moderate_cwnd(tp
);
2444 if (!tcp_any_retrans_done(sk
))
2445 tp
->retrans_stamp
= 0;
2448 tcp_set_ca_state(sk
, TCP_CA_Open
);
2452 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2453 static bool tcp_try_undo_dsack(struct sock
*sk
)
2455 struct tcp_sock
*tp
= tcp_sk(sk
);
2457 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
2458 DBGUNDO(sk
, "D-SACK");
2459 tcp_undo_cwnd_reduction(sk
, false);
2460 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPDSACKUNDO
);
2466 /* Undo during loss recovery after partial ACK or using F-RTO. */
2467 static bool tcp_try_undo_loss(struct sock
*sk
, bool frto_undo
)
2469 struct tcp_sock
*tp
= tcp_sk(sk
);
2471 if (frto_undo
|| tcp_may_undo(tp
)) {
2472 tcp_undo_cwnd_reduction(sk
, true);
2474 DBGUNDO(sk
, "partial loss");
2475 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSSUNDO
);
2477 NET_INC_STATS_BH(sock_net(sk
),
2478 LINUX_MIB_TCPSPURIOUSRTOS
);
2479 inet_csk(sk
)->icsk_retransmits
= 0;
2480 if (frto_undo
|| tcp_is_sack(tp
))
2481 tcp_set_ca_state(sk
, TCP_CA_Open
);
2487 /* The cwnd reduction in CWR and Recovery use the PRR algorithm
2488 * https://datatracker.ietf.org/doc/draft-ietf-tcpm-proportional-rate-reduction/
2489 * It computes the number of packets to send (sndcnt) based on packets newly
2491 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2492 * cwnd reductions across a full RTT.
2493 * 2) If packets in flight is lower than ssthresh (such as due to excess
2494 * losses and/or application stalls), do not perform any further cwnd
2495 * reductions, but instead slow start up to ssthresh.
2497 static void tcp_init_cwnd_reduction(struct sock
*sk
)
2499 struct tcp_sock
*tp
= tcp_sk(sk
);
2501 tp
->high_seq
= tp
->snd_nxt
;
2502 tp
->tlp_high_seq
= 0;
2503 tp
->snd_cwnd_cnt
= 0;
2504 tp
->prior_cwnd
= tp
->snd_cwnd
;
2505 tp
->prr_delivered
= 0;
2507 tp
->snd_ssthresh
= inet_csk(sk
)->icsk_ca_ops
->ssthresh(sk
);
2508 tcp_ecn_queue_cwr(tp
);
2511 static void tcp_cwnd_reduction(struct sock
*sk
, const int prior_unsacked
,
2514 struct tcp_sock
*tp
= tcp_sk(sk
);
2516 int delta
= tp
->snd_ssthresh
- tcp_packets_in_flight(tp
);
2517 int newly_acked_sacked
= prior_unsacked
-
2518 (tp
->packets_out
- tp
->sacked_out
);
2520 tp
->prr_delivered
+= newly_acked_sacked
;
2521 if (tcp_packets_in_flight(tp
) > tp
->snd_ssthresh
) {
2522 u64 dividend
= (u64
)tp
->snd_ssthresh
* tp
->prr_delivered
+
2524 sndcnt
= div_u64(dividend
, tp
->prior_cwnd
) - tp
->prr_out
;
2526 sndcnt
= min_t(int, delta
,
2527 max_t(int, tp
->prr_delivered
- tp
->prr_out
,
2528 newly_acked_sacked
) + 1);
2531 sndcnt
= max(sndcnt
, (fast_rexmit
? 1 : 0));
2532 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + sndcnt
;
2535 static inline void tcp_end_cwnd_reduction(struct sock
*sk
)
2537 struct tcp_sock
*tp
= tcp_sk(sk
);
2539 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2540 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_CWR
||
2541 (tp
->undo_marker
&& tp
->snd_ssthresh
< TCP_INFINITE_SSTHRESH
)) {
2542 tp
->snd_cwnd
= tp
->snd_ssthresh
;
2543 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2545 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
2548 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2549 void tcp_enter_cwr(struct sock
*sk
)
2551 struct tcp_sock
*tp
= tcp_sk(sk
);
2553 tp
->prior_ssthresh
= 0;
2554 if (inet_csk(sk
)->icsk_ca_state
< TCP_CA_CWR
) {
2555 tp
->undo_marker
= 0;
2556 tcp_init_cwnd_reduction(sk
);
2557 tcp_set_ca_state(sk
, TCP_CA_CWR
);
2561 static void tcp_try_keep_open(struct sock
*sk
)
2563 struct tcp_sock
*tp
= tcp_sk(sk
);
2564 int state
= TCP_CA_Open
;
2566 if (tcp_left_out(tp
) || tcp_any_retrans_done(sk
))
2567 state
= TCP_CA_Disorder
;
2569 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
2570 tcp_set_ca_state(sk
, state
);
2571 tp
->high_seq
= tp
->snd_nxt
;
2575 static void tcp_try_to_open(struct sock
*sk
, int flag
, const int prior_unsacked
)
2577 struct tcp_sock
*tp
= tcp_sk(sk
);
2579 tcp_verify_left_out(tp
);
2581 if (!tcp_any_retrans_done(sk
))
2582 tp
->retrans_stamp
= 0;
2584 if (flag
& FLAG_ECE
)
2587 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
2588 tcp_try_keep_open(sk
);
2590 tcp_cwnd_reduction(sk
, prior_unsacked
, 0);
2594 static void tcp_mtup_probe_failed(struct sock
*sk
)
2596 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2598 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
2599 icsk
->icsk_mtup
.probe_size
= 0;
2602 static void tcp_mtup_probe_success(struct sock
*sk
)
2604 struct tcp_sock
*tp
= tcp_sk(sk
);
2605 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2607 /* FIXME: breaks with very large cwnd */
2608 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2609 tp
->snd_cwnd
= tp
->snd_cwnd
*
2610 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
2611 icsk
->icsk_mtup
.probe_size
;
2612 tp
->snd_cwnd_cnt
= 0;
2613 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2614 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2616 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
2617 icsk
->icsk_mtup
.probe_size
= 0;
2618 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
2621 /* Do a simple retransmit without using the backoff mechanisms in
2622 * tcp_timer. This is used for path mtu discovery.
2623 * The socket is already locked here.
2625 void tcp_simple_retransmit(struct sock
*sk
)
2627 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2628 struct tcp_sock
*tp
= tcp_sk(sk
);
2629 struct sk_buff
*skb
;
2630 unsigned int mss
= tcp_current_mss(sk
);
2631 u32 prior_lost
= tp
->lost_out
;
2633 tcp_for_write_queue(skb
, sk
) {
2634 if (skb
== tcp_send_head(sk
))
2636 if (tcp_skb_seglen(skb
) > mss
&&
2637 !(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
2638 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2639 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
2640 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2642 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
2646 tcp_clear_retrans_hints_partial(tp
);
2648 if (prior_lost
== tp
->lost_out
)
2651 if (tcp_is_reno(tp
))
2652 tcp_limit_reno_sacked(tp
);
2654 tcp_verify_left_out(tp
);
2656 /* Don't muck with the congestion window here.
2657 * Reason is that we do not increase amount of _data_
2658 * in network, but units changed and effective
2659 * cwnd/ssthresh really reduced now.
2661 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
2662 tp
->high_seq
= tp
->snd_nxt
;
2663 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2664 tp
->prior_ssthresh
= 0;
2665 tp
->undo_marker
= 0;
2666 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2668 tcp_xmit_retransmit_queue(sk
);
2670 EXPORT_SYMBOL(tcp_simple_retransmit
);
2672 static void tcp_enter_recovery(struct sock
*sk
, bool ece_ack
)
2674 struct tcp_sock
*tp
= tcp_sk(sk
);
2677 if (tcp_is_reno(tp
))
2678 mib_idx
= LINUX_MIB_TCPRENORECOVERY
;
2680 mib_idx
= LINUX_MIB_TCPSACKRECOVERY
;
2682 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2684 tp
->prior_ssthresh
= 0;
2687 if (inet_csk(sk
)->icsk_ca_state
< TCP_CA_CWR
) {
2689 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2690 tcp_init_cwnd_reduction(sk
);
2692 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
2695 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2696 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2698 static void tcp_process_loss(struct sock
*sk
, int flag
, bool is_dupack
)
2700 struct tcp_sock
*tp
= tcp_sk(sk
);
2701 bool recovered
= !before(tp
->snd_una
, tp
->high_seq
);
2703 if (tp
->frto
) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2704 /* Step 3.b. A timeout is spurious if not all data are
2705 * lost, i.e., never-retransmitted data are (s)acked.
2707 if (tcp_try_undo_loss(sk
, flag
& FLAG_ORIG_SACK_ACKED
))
2710 if (after(tp
->snd_nxt
, tp
->high_seq
) &&
2711 (flag
& FLAG_DATA_SACKED
|| is_dupack
)) {
2712 tp
->frto
= 0; /* Loss was real: 2nd part of step 3.a */
2713 } else if (flag
& FLAG_SND_UNA_ADVANCED
&& !recovered
) {
2714 tp
->high_seq
= tp
->snd_nxt
;
2715 __tcp_push_pending_frames(sk
, tcp_current_mss(sk
),
2717 if (after(tp
->snd_nxt
, tp
->high_seq
))
2718 return; /* Step 2.b */
2724 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2725 tcp_try_undo_recovery(sk
);
2728 if (tcp_is_reno(tp
)) {
2729 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2730 * delivered. Lower inflight to clock out (re)tranmissions.
2732 if (after(tp
->snd_nxt
, tp
->high_seq
) && is_dupack
)
2733 tcp_add_reno_sack(sk
);
2734 else if (flag
& FLAG_SND_UNA_ADVANCED
)
2735 tcp_reset_reno_sack(tp
);
2737 if (tcp_try_undo_loss(sk
, false))
2739 tcp_xmit_retransmit_queue(sk
);
2742 /* Undo during fast recovery after partial ACK. */
2743 static bool tcp_try_undo_partial(struct sock
*sk
, const int acked
,
2744 const int prior_unsacked
)
2746 struct tcp_sock
*tp
= tcp_sk(sk
);
2748 if (tp
->undo_marker
&& tcp_packet_delayed(tp
)) {
2749 /* Plain luck! Hole if filled with delayed
2750 * packet, rather than with a retransmit.
2752 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
2754 /* We are getting evidence that the reordering degree is higher
2755 * than we realized. If there are no retransmits out then we
2756 * can undo. Otherwise we clock out new packets but do not
2757 * mark more packets lost or retransmit more.
2759 if (tp
->retrans_out
) {
2760 tcp_cwnd_reduction(sk
, prior_unsacked
, 0);
2764 if (!tcp_any_retrans_done(sk
))
2765 tp
->retrans_stamp
= 0;
2767 DBGUNDO(sk
, "partial recovery");
2768 tcp_undo_cwnd_reduction(sk
, true);
2769 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPARTIALUNDO
);
2770 tcp_try_keep_open(sk
);
2776 /* Process an event, which can update packets-in-flight not trivially.
2777 * Main goal of this function is to calculate new estimate for left_out,
2778 * taking into account both packets sitting in receiver's buffer and
2779 * packets lost by network.
2781 * Besides that it does CWND reduction, when packet loss is detected
2782 * and changes state of machine.
2784 * It does _not_ decide what to send, it is made in function
2785 * tcp_xmit_retransmit_queue().
2787 static void tcp_fastretrans_alert(struct sock
*sk
, const int acked
,
2788 const int prior_unsacked
,
2789 bool is_dupack
, int flag
)
2791 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2792 struct tcp_sock
*tp
= tcp_sk(sk
);
2793 bool do_lost
= is_dupack
|| ((flag
& FLAG_DATA_SACKED
) &&
2794 (tcp_fackets_out(tp
) > tp
->reordering
));
2795 int fast_rexmit
= 0;
2797 if (WARN_ON(!tp
->packets_out
&& tp
->sacked_out
))
2799 if (WARN_ON(!tp
->sacked_out
&& tp
->fackets_out
))
2800 tp
->fackets_out
= 0;
2802 /* Now state machine starts.
2803 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2804 if (flag
& FLAG_ECE
)
2805 tp
->prior_ssthresh
= 0;
2807 /* B. In all the states check for reneging SACKs. */
2808 if (tcp_check_sack_reneging(sk
, flag
))
2811 /* C. Check consistency of the current state. */
2812 tcp_verify_left_out(tp
);
2814 /* D. Check state exit conditions. State can be terminated
2815 * when high_seq is ACKed. */
2816 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
2817 WARN_ON(tp
->retrans_out
!= 0);
2818 tp
->retrans_stamp
= 0;
2819 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
2820 switch (icsk
->icsk_ca_state
) {
2822 /* CWR is to be held something *above* high_seq
2823 * is ACKed for CWR bit to reach receiver. */
2824 if (tp
->snd_una
!= tp
->high_seq
) {
2825 tcp_end_cwnd_reduction(sk
);
2826 tcp_set_ca_state(sk
, TCP_CA_Open
);
2830 case TCP_CA_Recovery
:
2831 if (tcp_is_reno(tp
))
2832 tcp_reset_reno_sack(tp
);
2833 if (tcp_try_undo_recovery(sk
))
2835 tcp_end_cwnd_reduction(sk
);
2840 /* E. Process state. */
2841 switch (icsk
->icsk_ca_state
) {
2842 case TCP_CA_Recovery
:
2843 if (!(flag
& FLAG_SND_UNA_ADVANCED
)) {
2844 if (tcp_is_reno(tp
) && is_dupack
)
2845 tcp_add_reno_sack(sk
);
2847 if (tcp_try_undo_partial(sk
, acked
, prior_unsacked
))
2849 /* Partial ACK arrived. Force fast retransmit. */
2850 do_lost
= tcp_is_reno(tp
) ||
2851 tcp_fackets_out(tp
) > tp
->reordering
;
2853 if (tcp_try_undo_dsack(sk
)) {
2854 tcp_try_keep_open(sk
);
2859 tcp_process_loss(sk
, flag
, is_dupack
);
2860 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
2862 /* Fall through to processing in Open state. */
2864 if (tcp_is_reno(tp
)) {
2865 if (flag
& FLAG_SND_UNA_ADVANCED
)
2866 tcp_reset_reno_sack(tp
);
2868 tcp_add_reno_sack(sk
);
2871 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
)
2872 tcp_try_undo_dsack(sk
);
2874 if (!tcp_time_to_recover(sk
, flag
)) {
2875 tcp_try_to_open(sk
, flag
, prior_unsacked
);
2879 /* MTU probe failure: don't reduce cwnd */
2880 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
2881 icsk
->icsk_mtup
.probe_size
&&
2882 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
2883 tcp_mtup_probe_failed(sk
);
2884 /* Restores the reduction we did in tcp_mtup_probe() */
2886 tcp_simple_retransmit(sk
);
2890 /* Otherwise enter Recovery state */
2891 tcp_enter_recovery(sk
, (flag
& FLAG_ECE
));
2896 tcp_update_scoreboard(sk
, fast_rexmit
);
2897 tcp_cwnd_reduction(sk
, prior_unsacked
, fast_rexmit
);
2898 tcp_xmit_retransmit_queue(sk
);
2901 static inline bool tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
2902 long seq_rtt_us
, long sack_rtt_us
)
2904 const struct tcp_sock
*tp
= tcp_sk(sk
);
2906 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2907 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2908 * Karn's algorithm forbids taking RTT if some retransmitted data
2909 * is acked (RFC6298).
2911 if (flag
& FLAG_RETRANS_DATA_ACKED
)
2915 seq_rtt_us
= sack_rtt_us
;
2917 /* RTTM Rule: A TSecr value received in a segment is used to
2918 * update the averaged RTT measurement only if the segment
2919 * acknowledges some new data, i.e., only if it advances the
2920 * left edge of the send window.
2921 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2923 if (seq_rtt_us
< 0 && tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2925 seq_rtt_us
= jiffies_to_usecs(tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
);
2930 tcp_rtt_estimator(sk
, seq_rtt_us
);
2933 /* RFC6298: only reset backoff on valid RTT measurement. */
2934 inet_csk(sk
)->icsk_backoff
= 0;
2938 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2939 static void tcp_synack_rtt_meas(struct sock
*sk
, const u32 synack_stamp
)
2941 struct tcp_sock
*tp
= tcp_sk(sk
);
2942 long seq_rtt_us
= -1L;
2944 if (synack_stamp
&& !tp
->total_retrans
)
2945 seq_rtt_us
= jiffies_to_usecs(tcp_time_stamp
- synack_stamp
);
2947 /* If the ACK acks both the SYNACK and the (Fast Open'd) data packets
2948 * sent in SYN_RECV, SYNACK RTT is the smooth RTT computed in tcp_ack()
2951 tcp_ack_update_rtt(sk
, FLAG_SYN_ACKED
, seq_rtt_us
, -1L);
2954 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
, u32 acked
)
2956 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2958 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, acked
);
2959 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_time_stamp
;
2962 /* Restart timer after forward progress on connection.
2963 * RFC2988 recommends to restart timer to now+rto.
2965 void tcp_rearm_rto(struct sock
*sk
)
2967 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2968 struct tcp_sock
*tp
= tcp_sk(sk
);
2970 /* If the retrans timer is currently being used by Fast Open
2971 * for SYN-ACK retrans purpose, stay put.
2973 if (tp
->fastopen_rsk
)
2976 if (!tp
->packets_out
) {
2977 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
2979 u32 rto
= inet_csk(sk
)->icsk_rto
;
2980 /* Offset the time elapsed after installing regular RTO */
2981 if (icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
||
2982 icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
) {
2983 struct sk_buff
*skb
= tcp_write_queue_head(sk
);
2984 const u32 rto_time_stamp
=
2985 tcp_skb_timestamp(skb
) + rto
;
2986 s32 delta
= (s32
)(rto_time_stamp
- tcp_time_stamp
);
2987 /* delta may not be positive if the socket is locked
2988 * when the retrans timer fires and is rescheduled.
2993 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
, rto
,
2998 /* This function is called when the delayed ER timer fires. TCP enters
2999 * fast recovery and performs fast-retransmit.
3001 void tcp_resume_early_retransmit(struct sock
*sk
)
3003 struct tcp_sock
*tp
= tcp_sk(sk
);
3007 /* Stop if ER is disabled after the delayed ER timer is scheduled */
3008 if (!tp
->do_early_retrans
)
3011 tcp_enter_recovery(sk
, false);
3012 tcp_update_scoreboard(sk
, 1);
3013 tcp_xmit_retransmit_queue(sk
);
3016 /* If we get here, the whole TSO packet has not been acked. */
3017 static u32
tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
)
3019 struct tcp_sock
*tp
= tcp_sk(sk
);
3022 BUG_ON(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
));
3024 packets_acked
= tcp_skb_pcount(skb
);
3025 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
3027 packets_acked
-= tcp_skb_pcount(skb
);
3029 if (packets_acked
) {
3030 BUG_ON(tcp_skb_pcount(skb
) == 0);
3031 BUG_ON(!before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
));
3034 return packets_acked
;
3037 static void tcp_ack_tstamp(struct sock
*sk
, struct sk_buff
*skb
,
3040 const struct skb_shared_info
*shinfo
;
3042 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3043 if (likely(!(sk
->sk_tsflags
& SOF_TIMESTAMPING_TX_ACK
)))
3046 shinfo
= skb_shinfo(skb
);
3047 if ((shinfo
->tx_flags
& SKBTX_ACK_TSTAMP
) &&
3048 between(shinfo
->tskey
, prior_snd_una
, tcp_sk(sk
)->snd_una
- 1))
3049 __skb_tstamp_tx(skb
, NULL
, sk
, SCM_TSTAMP_ACK
);
3052 /* Remove acknowledged frames from the retransmission queue. If our packet
3053 * is before the ack sequence we can discard it as it's confirmed to have
3054 * arrived at the other end.
3056 static int tcp_clean_rtx_queue(struct sock
*sk
, int prior_fackets
,
3057 u32 prior_snd_una
, long sack_rtt_us
)
3059 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3060 struct skb_mstamp first_ackt
, last_ackt
, now
;
3061 struct tcp_sock
*tp
= tcp_sk(sk
);
3062 u32 prior_sacked
= tp
->sacked_out
;
3063 u32 reord
= tp
->packets_out
;
3064 bool fully_acked
= true;
3065 long ca_seq_rtt_us
= -1L;
3066 long seq_rtt_us
= -1L;
3067 struct sk_buff
*skb
;
3074 while ((skb
= tcp_write_queue_head(sk
)) && skb
!= tcp_send_head(sk
)) {
3075 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
3076 u8 sacked
= scb
->sacked
;
3079 tcp_ack_tstamp(sk
, skb
, prior_snd_una
);
3081 /* Determine how many packets and what bytes were acked, tso and else */
3082 if (after(scb
->end_seq
, tp
->snd_una
)) {
3083 if (tcp_skb_pcount(skb
) == 1 ||
3084 !after(tp
->snd_una
, scb
->seq
))
3087 acked_pcount
= tcp_tso_acked(sk
, skb
);
3091 fully_acked
= false;
3093 /* Speedup tcp_unlink_write_queue() and next loop */
3094 prefetchw(skb
->next
);
3095 acked_pcount
= tcp_skb_pcount(skb
);
3098 if (unlikely(sacked
& TCPCB_RETRANS
)) {
3099 if (sacked
& TCPCB_SACKED_RETRANS
)
3100 tp
->retrans_out
-= acked_pcount
;
3101 flag
|= FLAG_RETRANS_DATA_ACKED
;
3103 last_ackt
= skb
->skb_mstamp
;
3104 WARN_ON_ONCE(last_ackt
.v64
== 0);
3105 if (!first_ackt
.v64
)
3106 first_ackt
= last_ackt
;
3108 if (!(sacked
& TCPCB_SACKED_ACKED
))
3109 reord
= min(pkts_acked
, reord
);
3110 if (!after(scb
->end_seq
, tp
->high_seq
))
3111 flag
|= FLAG_ORIG_SACK_ACKED
;
3114 if (sacked
& TCPCB_SACKED_ACKED
)
3115 tp
->sacked_out
-= acked_pcount
;
3116 if (sacked
& TCPCB_LOST
)
3117 tp
->lost_out
-= acked_pcount
;
3119 tp
->packets_out
-= acked_pcount
;
3120 pkts_acked
+= acked_pcount
;
3122 /* Initial outgoing SYN's get put onto the write_queue
3123 * just like anything else we transmit. It is not
3124 * true data, and if we misinform our callers that
3125 * this ACK acks real data, we will erroneously exit
3126 * connection startup slow start one packet too
3127 * quickly. This is severely frowned upon behavior.
3129 if (likely(!(scb
->tcp_flags
& TCPHDR_SYN
))) {
3130 flag
|= FLAG_DATA_ACKED
;
3132 flag
|= FLAG_SYN_ACKED
;
3133 tp
->retrans_stamp
= 0;
3139 tcp_unlink_write_queue(skb
, sk
);
3140 sk_wmem_free_skb(sk
, skb
);
3141 if (unlikely(skb
== tp
->retransmit_skb_hint
))
3142 tp
->retransmit_skb_hint
= NULL
;
3143 if (unlikely(skb
== tp
->lost_skb_hint
))
3144 tp
->lost_skb_hint
= NULL
;
3147 if (likely(between(tp
->snd_up
, prior_snd_una
, tp
->snd_una
)))
3148 tp
->snd_up
= tp
->snd_una
;
3150 if (skb
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
3151 flag
|= FLAG_SACK_RENEGING
;
3153 skb_mstamp_get(&now
);
3154 if (likely(first_ackt
.v64
)) {
3155 seq_rtt_us
= skb_mstamp_us_delta(&now
, &first_ackt
);
3156 ca_seq_rtt_us
= skb_mstamp_us_delta(&now
, &last_ackt
);
3159 rtt_update
= tcp_ack_update_rtt(sk
, flag
, seq_rtt_us
, sack_rtt_us
);
3161 if (flag
& FLAG_ACKED
) {
3162 const struct tcp_congestion_ops
*ca_ops
3163 = inet_csk(sk
)->icsk_ca_ops
;
3166 if (unlikely(icsk
->icsk_mtup
.probe_size
&&
3167 !after(tp
->mtu_probe
.probe_seq_end
, tp
->snd_una
))) {
3168 tcp_mtup_probe_success(sk
);
3171 if (tcp_is_reno(tp
)) {
3172 tcp_remove_reno_sacks(sk
, pkts_acked
);
3176 /* Non-retransmitted hole got filled? That's reordering */
3177 if (reord
< prior_fackets
)
3178 tcp_update_reordering(sk
, tp
->fackets_out
- reord
, 0);
3180 delta
= tcp_is_fack(tp
) ? pkts_acked
:
3181 prior_sacked
- tp
->sacked_out
;
3182 tp
->lost_cnt_hint
-= min(tp
->lost_cnt_hint
, delta
);
3185 tp
->fackets_out
-= min(pkts_acked
, tp
->fackets_out
);
3187 if (ca_ops
->pkts_acked
) {
3188 long rtt_us
= min_t(ulong
, ca_seq_rtt_us
, sack_rtt_us
);
3189 ca_ops
->pkts_acked(sk
, pkts_acked
, rtt_us
);
3192 } else if (skb
&& rtt_update
&& sack_rtt_us
>= 0 &&
3193 sack_rtt_us
> skb_mstamp_us_delta(&now
, &skb
->skb_mstamp
)) {
3194 /* Do not re-arm RTO if the sack RTT is measured from data sent
3195 * after when the head was last (re)transmitted. Otherwise the
3196 * timeout may continue to extend in loss recovery.
3201 #if FASTRETRANS_DEBUG > 0
3202 WARN_ON((int)tp
->sacked_out
< 0);
3203 WARN_ON((int)tp
->lost_out
< 0);
3204 WARN_ON((int)tp
->retrans_out
< 0);
3205 if (!tp
->packets_out
&& tcp_is_sack(tp
)) {
3206 icsk
= inet_csk(sk
);
3208 pr_debug("Leak l=%u %d\n",
3209 tp
->lost_out
, icsk
->icsk_ca_state
);
3212 if (tp
->sacked_out
) {
3213 pr_debug("Leak s=%u %d\n",
3214 tp
->sacked_out
, icsk
->icsk_ca_state
);
3217 if (tp
->retrans_out
) {
3218 pr_debug("Leak r=%u %d\n",
3219 tp
->retrans_out
, icsk
->icsk_ca_state
);
3220 tp
->retrans_out
= 0;
3227 static void tcp_ack_probe(struct sock
*sk
)
3229 const struct tcp_sock
*tp
= tcp_sk(sk
);
3230 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3232 /* Was it a usable window open? */
3234 if (!after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
, tcp_wnd_end(tp
))) {
3235 icsk
->icsk_backoff
= 0;
3236 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
3237 /* Socket must be waked up by subsequent tcp_data_snd_check().
3238 * This function is not for random using!
3241 unsigned long when
= inet_csk_rto_backoff(icsk
, TCP_RTO_MAX
);
3243 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3248 static inline bool tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
3250 return !(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
3251 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
;
3254 /* Decide wheather to run the increase function of congestion control. */
3255 static inline bool tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
3257 if (tcp_in_cwnd_reduction(sk
))
3260 /* If reordering is high then always grow cwnd whenever data is
3261 * delivered regardless of its ordering. Otherwise stay conservative
3262 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3263 * new SACK or ECE mark may first advance cwnd here and later reduce
3264 * cwnd in tcp_fastretrans_alert() based on more states.
3266 if (tcp_sk(sk
)->reordering
> sysctl_tcp_reordering
)
3267 return flag
& FLAG_FORWARD_PROGRESS
;
3269 return flag
& FLAG_DATA_ACKED
;
3272 /* Check that window update is acceptable.
3273 * The function assumes that snd_una<=ack<=snd_next.
3275 static inline bool tcp_may_update_window(const struct tcp_sock
*tp
,
3276 const u32 ack
, const u32 ack_seq
,
3279 return after(ack
, tp
->snd_una
) ||
3280 after(ack_seq
, tp
->snd_wl1
) ||
3281 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
);
3284 /* Update our send window.
3286 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3287 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3289 static int tcp_ack_update_window(struct sock
*sk
, const struct sk_buff
*skb
, u32 ack
,
3292 struct tcp_sock
*tp
= tcp_sk(sk
);
3294 u32 nwin
= ntohs(tcp_hdr(skb
)->window
);
3296 if (likely(!tcp_hdr(skb
)->syn
))
3297 nwin
<<= tp
->rx_opt
.snd_wscale
;
3299 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
3300 flag
|= FLAG_WIN_UPDATE
;
3301 tcp_update_wl(tp
, ack_seq
);
3303 if (tp
->snd_wnd
!= nwin
) {
3306 /* Note, it is the only place, where
3307 * fast path is recovered for sending TCP.
3310 tcp_fast_path_check(sk
);
3312 if (nwin
> tp
->max_window
) {
3313 tp
->max_window
= nwin
;
3314 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
3324 /* RFC 5961 7 [ACK Throttling] */
3325 static void tcp_send_challenge_ack(struct sock
*sk
, const struct sk_buff
*skb
)
3327 /* unprotected vars, we dont care of overwrites */
3328 static u32 challenge_timestamp
;
3329 static unsigned int challenge_count
;
3330 struct tcp_sock
*tp
= tcp_sk(sk
);
3333 /* First check our per-socket dupack rate limit. */
3334 if (tcp_oow_rate_limited(sock_net(sk
), skb
,
3335 LINUX_MIB_TCPACKSKIPPEDCHALLENGE
,
3336 &tp
->last_oow_ack_time
))
3339 /* Then check the check host-wide RFC 5961 rate limit. */
3341 if (now
!= challenge_timestamp
) {
3342 challenge_timestamp
= now
;
3343 challenge_count
= 0;
3345 if (++challenge_count
<= sysctl_tcp_challenge_ack_limit
) {
3346 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPCHALLENGEACK
);
3351 static void tcp_store_ts_recent(struct tcp_sock
*tp
)
3353 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
3354 tp
->rx_opt
.ts_recent_stamp
= get_seconds();
3357 static void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
3359 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
3360 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3361 * extra check below makes sure this can only happen
3362 * for pure ACK frames. -DaveM
3364 * Not only, also it occurs for expired timestamps.
3367 if (tcp_paws_check(&tp
->rx_opt
, 0))
3368 tcp_store_ts_recent(tp
);
3372 /* This routine deals with acks during a TLP episode.
3373 * We mark the end of a TLP episode on receiving TLP dupack or when
3374 * ack is after tlp_high_seq.
3375 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3377 static void tcp_process_tlp_ack(struct sock
*sk
, u32 ack
, int flag
)
3379 struct tcp_sock
*tp
= tcp_sk(sk
);
3381 if (before(ack
, tp
->tlp_high_seq
))
3384 if (flag
& FLAG_DSACKING_ACK
) {
3385 /* This DSACK means original and TLP probe arrived; no loss */
3386 tp
->tlp_high_seq
= 0;
3387 } else if (after(ack
, tp
->tlp_high_seq
)) {
3388 /* ACK advances: there was a loss, so reduce cwnd. Reset
3389 * tlp_high_seq in tcp_init_cwnd_reduction()
3391 tcp_init_cwnd_reduction(sk
);
3392 tcp_set_ca_state(sk
, TCP_CA_CWR
);
3393 tcp_end_cwnd_reduction(sk
);
3394 tcp_try_keep_open(sk
);
3395 NET_INC_STATS_BH(sock_net(sk
),
3396 LINUX_MIB_TCPLOSSPROBERECOVERY
);
3397 } else if (!(flag
& (FLAG_SND_UNA_ADVANCED
|
3398 FLAG_NOT_DUP
| FLAG_DATA_SACKED
))) {
3399 /* Pure dupack: original and TLP probe arrived; no loss */
3400 tp
->tlp_high_seq
= 0;
3404 static inline void tcp_in_ack_event(struct sock
*sk
, u32 flags
)
3406 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3408 if (icsk
->icsk_ca_ops
->in_ack_event
)
3409 icsk
->icsk_ca_ops
->in_ack_event(sk
, flags
);
3412 /* This routine deals with incoming acks, but not outgoing ones. */
3413 static int tcp_ack(struct sock
*sk
, const struct sk_buff
*skb
, int flag
)
3415 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3416 struct tcp_sock
*tp
= tcp_sk(sk
);
3417 u32 prior_snd_una
= tp
->snd_una
;
3418 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
3419 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3420 bool is_dupack
= false;
3422 int prior_packets
= tp
->packets_out
;
3423 const int prior_unsacked
= tp
->packets_out
- tp
->sacked_out
;
3424 int acked
= 0; /* Number of packets newly acked */
3425 long sack_rtt_us
= -1L;
3427 /* We very likely will need to access write queue head. */
3428 prefetchw(sk
->sk_write_queue
.next
);
3430 /* If the ack is older than previous acks
3431 * then we can probably ignore it.
3433 if (before(ack
, prior_snd_una
)) {
3434 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3435 if (before(ack
, prior_snd_una
- tp
->max_window
)) {
3436 tcp_send_challenge_ack(sk
, skb
);
3442 /* If the ack includes data we haven't sent yet, discard
3443 * this segment (RFC793 Section 3.9).
3445 if (after(ack
, tp
->snd_nxt
))
3448 if (icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
||
3449 icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
)
3452 if (after(ack
, prior_snd_una
)) {
3453 flag
|= FLAG_SND_UNA_ADVANCED
;
3454 icsk
->icsk_retransmits
= 0;
3457 prior_fackets
= tp
->fackets_out
;
3459 /* ts_recent update must be made after we are sure that the packet
3462 if (flag
& FLAG_UPDATE_TS_RECENT
)
3463 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
3465 if (!(flag
& FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
3466 /* Window is constant, pure forward advance.
3467 * No more checks are required.
3468 * Note, we use the fact that SND.UNA>=SND.WL2.
3470 tcp_update_wl(tp
, ack_seq
);
3472 flag
|= FLAG_WIN_UPDATE
;
3474 tcp_in_ack_event(sk
, CA_ACK_WIN_UPDATE
);
3476 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPACKS
);
3478 u32 ack_ev_flags
= CA_ACK_SLOWPATH
;
3480 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
3483 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPUREACKS
);
3485 flag
|= tcp_ack_update_window(sk
, skb
, ack
, ack_seq
);
3487 if (TCP_SKB_CB(skb
)->sacked
)
3488 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
,
3491 if (tcp_ecn_rcv_ecn_echo(tp
, tcp_hdr(skb
))) {
3493 ack_ev_flags
|= CA_ACK_ECE
;
3496 if (flag
& FLAG_WIN_UPDATE
)
3497 ack_ev_flags
|= CA_ACK_WIN_UPDATE
;
3499 tcp_in_ack_event(sk
, ack_ev_flags
);
3502 /* We passed data and got it acked, remove any soft error
3503 * log. Something worked...
3505 sk
->sk_err_soft
= 0;
3506 icsk
->icsk_probes_out
= 0;
3507 tp
->rcv_tstamp
= tcp_time_stamp
;
3511 /* See if we can take anything off of the retransmit queue. */
3512 acked
= tp
->packets_out
;
3513 flag
|= tcp_clean_rtx_queue(sk
, prior_fackets
, prior_snd_una
,
3515 acked
-= tp
->packets_out
;
3517 /* Advance cwnd if state allows */
3518 if (tcp_may_raise_cwnd(sk
, flag
))
3519 tcp_cong_avoid(sk
, ack
, acked
);
3521 if (tcp_ack_is_dubious(sk
, flag
)) {
3522 is_dupack
= !(flag
& (FLAG_SND_UNA_ADVANCED
| FLAG_NOT_DUP
));
3523 tcp_fastretrans_alert(sk
, acked
, prior_unsacked
,
3526 if (tp
->tlp_high_seq
)
3527 tcp_process_tlp_ack(sk
, ack
, flag
);
3529 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
& FLAG_NOT_DUP
)) {
3530 struct dst_entry
*dst
= __sk_dst_get(sk
);
3535 if (icsk
->icsk_pending
== ICSK_TIME_RETRANS
)
3536 tcp_schedule_loss_probe(sk
);
3537 tcp_update_pacing_rate(sk
);
3541 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3542 if (flag
& FLAG_DSACKING_ACK
)
3543 tcp_fastretrans_alert(sk
, acked
, prior_unsacked
,
3545 /* If this ack opens up a zero window, clear backoff. It was
3546 * being used to time the probes, and is probably far higher than
3547 * it needs to be for normal retransmission.
3549 if (tcp_send_head(sk
))
3552 if (tp
->tlp_high_seq
)
3553 tcp_process_tlp_ack(sk
, ack
, flag
);
3557 SOCK_DEBUG(sk
, "Ack %u after %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3561 /* If data was SACKed, tag it and see if we should send more data.
3562 * If data was DSACKed, see if we can undo a cwnd reduction.
3564 if (TCP_SKB_CB(skb
)->sacked
) {
3565 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
,
3567 tcp_fastretrans_alert(sk
, acked
, prior_unsacked
,
3571 SOCK_DEBUG(sk
, "Ack %u before %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3575 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3576 * But, this can also be called on packets in the established flow when
3577 * the fast version below fails.
3579 void tcp_parse_options(const struct sk_buff
*skb
,
3580 struct tcp_options_received
*opt_rx
, int estab
,
3581 struct tcp_fastopen_cookie
*foc
)
3583 const unsigned char *ptr
;
3584 const struct tcphdr
*th
= tcp_hdr(skb
);
3585 int length
= (th
->doff
* 4) - sizeof(struct tcphdr
);
3587 ptr
= (const unsigned char *)(th
+ 1);
3588 opt_rx
->saw_tstamp
= 0;
3590 while (length
> 0) {
3591 int opcode
= *ptr
++;
3597 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
3602 if (opsize
< 2) /* "silly options" */
3604 if (opsize
> length
)
3605 return; /* don't parse partial options */
3608 if (opsize
== TCPOLEN_MSS
&& th
->syn
&& !estab
) {
3609 u16 in_mss
= get_unaligned_be16(ptr
);
3611 if (opt_rx
->user_mss
&&
3612 opt_rx
->user_mss
< in_mss
)
3613 in_mss
= opt_rx
->user_mss
;
3614 opt_rx
->mss_clamp
= in_mss
;
3619 if (opsize
== TCPOLEN_WINDOW
&& th
->syn
&&
3620 !estab
&& sysctl_tcp_window_scaling
) {
3621 __u8 snd_wscale
= *(__u8
*)ptr
;
3622 opt_rx
->wscale_ok
= 1;
3623 if (snd_wscale
> 14) {
3624 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3629 opt_rx
->snd_wscale
= snd_wscale
;
3632 case TCPOPT_TIMESTAMP
:
3633 if ((opsize
== TCPOLEN_TIMESTAMP
) &&
3634 ((estab
&& opt_rx
->tstamp_ok
) ||
3635 (!estab
&& sysctl_tcp_timestamps
))) {
3636 opt_rx
->saw_tstamp
= 1;
3637 opt_rx
->rcv_tsval
= get_unaligned_be32(ptr
);
3638 opt_rx
->rcv_tsecr
= get_unaligned_be32(ptr
+ 4);
3641 case TCPOPT_SACK_PERM
:
3642 if (opsize
== TCPOLEN_SACK_PERM
&& th
->syn
&&
3643 !estab
&& sysctl_tcp_sack
) {
3644 opt_rx
->sack_ok
= TCP_SACK_SEEN
;
3645 tcp_sack_reset(opt_rx
);
3650 if ((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
3651 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
3653 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
3656 #ifdef CONFIG_TCP_MD5SIG
3659 * The MD5 Hash has already been
3660 * checked (see tcp_v{4,6}_do_rcv()).
3665 /* Fast Open option shares code 254 using a
3666 * 16 bits magic number. It's valid only in
3667 * SYN or SYN-ACK with an even size.
3669 if (opsize
< TCPOLEN_EXP_FASTOPEN_BASE
||
3670 get_unaligned_be16(ptr
) != TCPOPT_FASTOPEN_MAGIC
||
3671 foc
== NULL
|| !th
->syn
|| (opsize
& 1))
3673 foc
->len
= opsize
- TCPOLEN_EXP_FASTOPEN_BASE
;
3674 if (foc
->len
>= TCP_FASTOPEN_COOKIE_MIN
&&
3675 foc
->len
<= TCP_FASTOPEN_COOKIE_MAX
)
3676 memcpy(foc
->val
, ptr
+ 2, foc
->len
);
3677 else if (foc
->len
!= 0)
3687 EXPORT_SYMBOL(tcp_parse_options
);
3689 static bool tcp_parse_aligned_timestamp(struct tcp_sock
*tp
, const struct tcphdr
*th
)
3691 const __be32
*ptr
= (const __be32
*)(th
+ 1);
3693 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3694 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
3695 tp
->rx_opt
.saw_tstamp
= 1;
3697 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3700 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
) - tp
->tsoffset
;
3702 tp
->rx_opt
.rcv_tsecr
= 0;
3708 /* Fast parse options. This hopes to only see timestamps.
3709 * If it is wrong it falls back on tcp_parse_options().
3711 static bool tcp_fast_parse_options(const struct sk_buff
*skb
,
3712 const struct tcphdr
*th
, struct tcp_sock
*tp
)
3714 /* In the spirit of fast parsing, compare doff directly to constant
3715 * values. Because equality is used, short doff can be ignored here.
3717 if (th
->doff
== (sizeof(*th
) / 4)) {
3718 tp
->rx_opt
.saw_tstamp
= 0;
3720 } else if (tp
->rx_opt
.tstamp_ok
&&
3721 th
->doff
== ((sizeof(*th
) + TCPOLEN_TSTAMP_ALIGNED
) / 4)) {
3722 if (tcp_parse_aligned_timestamp(tp
, th
))
3726 tcp_parse_options(skb
, &tp
->rx_opt
, 1, NULL
);
3727 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
3728 tp
->rx_opt
.rcv_tsecr
-= tp
->tsoffset
;
3733 #ifdef CONFIG_TCP_MD5SIG
3735 * Parse MD5 Signature option
3737 const u8
*tcp_parse_md5sig_option(const struct tcphdr
*th
)
3739 int length
= (th
->doff
<< 2) - sizeof(*th
);
3740 const u8
*ptr
= (const u8
*)(th
+ 1);
3742 /* If the TCP option is too short, we can short cut */
3743 if (length
< TCPOLEN_MD5SIG
)
3746 while (length
> 0) {
3747 int opcode
= *ptr
++;
3758 if (opsize
< 2 || opsize
> length
)
3760 if (opcode
== TCPOPT_MD5SIG
)
3761 return opsize
== TCPOLEN_MD5SIG
? ptr
: NULL
;
3768 EXPORT_SYMBOL(tcp_parse_md5sig_option
);
3771 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3773 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3774 * it can pass through stack. So, the following predicate verifies that
3775 * this segment is not used for anything but congestion avoidance or
3776 * fast retransmit. Moreover, we even are able to eliminate most of such
3777 * second order effects, if we apply some small "replay" window (~RTO)
3778 * to timestamp space.
3780 * All these measures still do not guarantee that we reject wrapped ACKs
3781 * on networks with high bandwidth, when sequence space is recycled fastly,
3782 * but it guarantees that such events will be very rare and do not affect
3783 * connection seriously. This doesn't look nice, but alas, PAWS is really
3786 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3787 * states that events when retransmit arrives after original data are rare.
3788 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3789 * the biggest problem on large power networks even with minor reordering.
3790 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3791 * up to bandwidth of 18Gigabit/sec. 8) ]
3794 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
3796 const struct tcp_sock
*tp
= tcp_sk(sk
);
3797 const struct tcphdr
*th
= tcp_hdr(skb
);
3798 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3799 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3801 return (/* 1. Pure ACK with correct sequence number. */
3802 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
3804 /* 2. ... and duplicate ACK. */
3805 ack
== tp
->snd_una
&&
3807 /* 3. ... and does not update window. */
3808 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
3810 /* 4. ... and sits in replay window. */
3811 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
3814 static inline bool tcp_paws_discard(const struct sock
*sk
,
3815 const struct sk_buff
*skb
)
3817 const struct tcp_sock
*tp
= tcp_sk(sk
);
3819 return !tcp_paws_check(&tp
->rx_opt
, TCP_PAWS_WINDOW
) &&
3820 !tcp_disordered_ack(sk
, skb
);
3823 /* Check segment sequence number for validity.
3825 * Segment controls are considered valid, if the segment
3826 * fits to the window after truncation to the window. Acceptability
3827 * of data (and SYN, FIN, of course) is checked separately.
3828 * See tcp_data_queue(), for example.
3830 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3831 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3832 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3833 * (borrowed from freebsd)
3836 static inline bool tcp_sequence(const struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3838 return !before(end_seq
, tp
->rcv_wup
) &&
3839 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
3842 /* When we get a reset we do this. */
3843 void tcp_reset(struct sock
*sk
)
3845 /* We want the right error as BSD sees it (and indeed as we do). */
3846 switch (sk
->sk_state
) {
3848 sk
->sk_err
= ECONNREFUSED
;
3850 case TCP_CLOSE_WAIT
:
3856 sk
->sk_err
= ECONNRESET
;
3858 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3861 if (!sock_flag(sk
, SOCK_DEAD
))
3862 sk
->sk_error_report(sk
);
3868 * Process the FIN bit. This now behaves as it is supposed to work
3869 * and the FIN takes effect when it is validly part of sequence
3870 * space. Not before when we get holes.
3872 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3873 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3876 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3877 * close and we go into CLOSING (and later onto TIME-WAIT)
3879 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3881 static void tcp_fin(struct sock
*sk
)
3883 struct tcp_sock
*tp
= tcp_sk(sk
);
3884 const struct dst_entry
*dst
;
3886 inet_csk_schedule_ack(sk
);
3888 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
3889 sock_set_flag(sk
, SOCK_DONE
);
3891 switch (sk
->sk_state
) {
3893 case TCP_ESTABLISHED
:
3894 /* Move to CLOSE_WAIT */
3895 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
3896 dst
= __sk_dst_get(sk
);
3897 if (!dst
|| !dst_metric(dst
, RTAX_QUICKACK
))
3898 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
3901 case TCP_CLOSE_WAIT
:
3903 /* Received a retransmission of the FIN, do
3908 /* RFC793: Remain in the LAST-ACK state. */
3912 /* This case occurs when a simultaneous close
3913 * happens, we must ack the received FIN and
3914 * enter the CLOSING state.
3917 tcp_set_state(sk
, TCP_CLOSING
);
3920 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3922 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
3925 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3926 * cases we should never reach this piece of code.
3928 pr_err("%s: Impossible, sk->sk_state=%d\n",
3929 __func__
, sk
->sk_state
);
3933 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3934 * Probably, we should reset in this case. For now drop them.
3936 __skb_queue_purge(&tp
->out_of_order_queue
);
3937 if (tcp_is_sack(tp
))
3938 tcp_sack_reset(&tp
->rx_opt
);
3941 if (!sock_flag(sk
, SOCK_DEAD
)) {
3942 sk
->sk_state_change(sk
);
3944 /* Do not send POLL_HUP for half duplex close. */
3945 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
3946 sk
->sk_state
== TCP_CLOSE
)
3947 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_HUP
);
3949 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_IN
);
3953 static inline bool tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
,
3956 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
3957 if (before(seq
, sp
->start_seq
))
3958 sp
->start_seq
= seq
;
3959 if (after(end_seq
, sp
->end_seq
))
3960 sp
->end_seq
= end_seq
;
3966 static void tcp_dsack_set(struct sock
*sk
, u32 seq
, u32 end_seq
)
3968 struct tcp_sock
*tp
= tcp_sk(sk
);
3970 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
3973 if (before(seq
, tp
->rcv_nxt
))
3974 mib_idx
= LINUX_MIB_TCPDSACKOLDSENT
;
3976 mib_idx
= LINUX_MIB_TCPDSACKOFOSENT
;
3978 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
3980 tp
->rx_opt
.dsack
= 1;
3981 tp
->duplicate_sack
[0].start_seq
= seq
;
3982 tp
->duplicate_sack
[0].end_seq
= end_seq
;
3986 static void tcp_dsack_extend(struct sock
*sk
, u32 seq
, u32 end_seq
)
3988 struct tcp_sock
*tp
= tcp_sk(sk
);
3990 if (!tp
->rx_opt
.dsack
)
3991 tcp_dsack_set(sk
, seq
, end_seq
);
3993 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
3996 static void tcp_send_dupack(struct sock
*sk
, const struct sk_buff
*skb
)
3998 struct tcp_sock
*tp
= tcp_sk(sk
);
4000 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
4001 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4002 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4003 tcp_enter_quickack_mode(sk
);
4005 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
4006 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4008 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
4009 end_seq
= tp
->rcv_nxt
;
4010 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, end_seq
);
4017 /* These routines update the SACK block as out-of-order packets arrive or
4018 * in-order packets close up the sequence space.
4020 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
4023 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4024 struct tcp_sack_block
*swalk
= sp
+ 1;
4026 /* See if the recent change to the first SACK eats into
4027 * or hits the sequence space of other SACK blocks, if so coalesce.
4029 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
;) {
4030 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
4033 /* Zap SWALK, by moving every further SACK up by one slot.
4034 * Decrease num_sacks.
4036 tp
->rx_opt
.num_sacks
--;
4037 for (i
= this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
4041 this_sack
++, swalk
++;
4045 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
4047 struct tcp_sock
*tp
= tcp_sk(sk
);
4048 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4049 int cur_sacks
= tp
->rx_opt
.num_sacks
;
4055 for (this_sack
= 0; this_sack
< cur_sacks
; this_sack
++, sp
++) {
4056 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
4057 /* Rotate this_sack to the first one. */
4058 for (; this_sack
> 0; this_sack
--, sp
--)
4059 swap(*sp
, *(sp
- 1));
4061 tcp_sack_maybe_coalesce(tp
);
4066 /* Could not find an adjacent existing SACK, build a new one,
4067 * put it at the front, and shift everyone else down. We
4068 * always know there is at least one SACK present already here.
4070 * If the sack array is full, forget about the last one.
4072 if (this_sack
>= TCP_NUM_SACKS
) {
4074 tp
->rx_opt
.num_sacks
--;
4077 for (; this_sack
> 0; this_sack
--, sp
--)
4081 /* Build the new head SACK, and we're done. */
4082 sp
->start_seq
= seq
;
4083 sp
->end_seq
= end_seq
;
4084 tp
->rx_opt
.num_sacks
++;
4087 /* RCV.NXT advances, some SACKs should be eaten. */
4089 static void tcp_sack_remove(struct tcp_sock
*tp
)
4091 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4092 int num_sacks
= tp
->rx_opt
.num_sacks
;
4095 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4096 if (skb_queue_empty(&tp
->out_of_order_queue
)) {
4097 tp
->rx_opt
.num_sacks
= 0;
4101 for (this_sack
= 0; this_sack
< num_sacks
;) {
4102 /* Check if the start of the sack is covered by RCV.NXT. */
4103 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
4106 /* RCV.NXT must cover all the block! */
4107 WARN_ON(before(tp
->rcv_nxt
, sp
->end_seq
));
4109 /* Zap this SACK, by moving forward any other SACKS. */
4110 for (i
= this_sack
+1; i
< num_sacks
; i
++)
4111 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
4118 tp
->rx_opt
.num_sacks
= num_sacks
;
4122 * tcp_try_coalesce - try to merge skb to prior one
4125 * @from: buffer to add in queue
4126 * @fragstolen: pointer to boolean
4128 * Before queueing skb @from after @to, try to merge them
4129 * to reduce overall memory use and queue lengths, if cost is small.
4130 * Packets in ofo or receive queues can stay a long time.
4131 * Better try to coalesce them right now to avoid future collapses.
4132 * Returns true if caller should free @from instead of queueing it
4134 static bool tcp_try_coalesce(struct sock
*sk
,
4136 struct sk_buff
*from
,
4141 *fragstolen
= false;
4143 /* Its possible this segment overlaps with prior segment in queue */
4144 if (TCP_SKB_CB(from
)->seq
!= TCP_SKB_CB(to
)->end_seq
)
4147 if (!skb_try_coalesce(to
, from
, fragstolen
, &delta
))
4150 atomic_add(delta
, &sk
->sk_rmem_alloc
);
4151 sk_mem_charge(sk
, delta
);
4152 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRCVCOALESCE
);
4153 TCP_SKB_CB(to
)->end_seq
= TCP_SKB_CB(from
)->end_seq
;
4154 TCP_SKB_CB(to
)->ack_seq
= TCP_SKB_CB(from
)->ack_seq
;
4155 TCP_SKB_CB(to
)->tcp_flags
|= TCP_SKB_CB(from
)->tcp_flags
;
4159 /* This one checks to see if we can put data from the
4160 * out_of_order queue into the receive_queue.
4162 static void tcp_ofo_queue(struct sock
*sk
)
4164 struct tcp_sock
*tp
= tcp_sk(sk
);
4165 __u32 dsack_high
= tp
->rcv_nxt
;
4166 struct sk_buff
*skb
, *tail
;
4167 bool fragstolen
, eaten
;
4169 while ((skb
= skb_peek(&tp
->out_of_order_queue
)) != NULL
) {
4170 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
4173 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
4174 __u32 dsack
= dsack_high
;
4175 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
4176 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
4177 tcp_dsack_extend(sk
, TCP_SKB_CB(skb
)->seq
, dsack
);
4180 __skb_unlink(skb
, &tp
->out_of_order_queue
);
4181 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4182 SOCK_DEBUG(sk
, "ofo packet was already received\n");
4186 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
4187 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4188 TCP_SKB_CB(skb
)->end_seq
);
4190 tail
= skb_peek_tail(&sk
->sk_receive_queue
);
4191 eaten
= tail
&& tcp_try_coalesce(sk
, tail
, skb
, &fragstolen
);
4192 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4194 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4195 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
4198 kfree_skb_partial(skb
, fragstolen
);
4202 static bool tcp_prune_ofo_queue(struct sock
*sk
);
4203 static int tcp_prune_queue(struct sock
*sk
);
4205 static int tcp_try_rmem_schedule(struct sock
*sk
, struct sk_buff
*skb
,
4208 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
4209 !sk_rmem_schedule(sk
, skb
, size
)) {
4211 if (tcp_prune_queue(sk
) < 0)
4214 if (!sk_rmem_schedule(sk
, skb
, size
)) {
4215 if (!tcp_prune_ofo_queue(sk
))
4218 if (!sk_rmem_schedule(sk
, skb
, size
))
4225 static void tcp_data_queue_ofo(struct sock
*sk
, struct sk_buff
*skb
)
4227 struct tcp_sock
*tp
= tcp_sk(sk
);
4228 struct sk_buff
*skb1
;
4231 tcp_ecn_check_ce(tp
, skb
);
4233 if (unlikely(tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))) {
4234 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFODROP
);
4239 /* Disable header prediction. */
4241 inet_csk_schedule_ack(sk
);
4243 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFOQUEUE
);
4244 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
4245 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4247 skb1
= skb_peek_tail(&tp
->out_of_order_queue
);
4249 /* Initial out of order segment, build 1 SACK. */
4250 if (tcp_is_sack(tp
)) {
4251 tp
->rx_opt
.num_sacks
= 1;
4252 tp
->selective_acks
[0].start_seq
= TCP_SKB_CB(skb
)->seq
;
4253 tp
->selective_acks
[0].end_seq
=
4254 TCP_SKB_CB(skb
)->end_seq
;
4256 __skb_queue_head(&tp
->out_of_order_queue
, skb
);
4260 seq
= TCP_SKB_CB(skb
)->seq
;
4261 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4263 if (seq
== TCP_SKB_CB(skb1
)->end_seq
) {
4266 if (!tcp_try_coalesce(sk
, skb1
, skb
, &fragstolen
)) {
4267 __skb_queue_after(&tp
->out_of_order_queue
, skb1
, skb
);
4269 tcp_grow_window(sk
, skb
);
4270 kfree_skb_partial(skb
, fragstolen
);
4274 if (!tp
->rx_opt
.num_sacks
||
4275 tp
->selective_acks
[0].end_seq
!= seq
)
4278 /* Common case: data arrive in order after hole. */
4279 tp
->selective_acks
[0].end_seq
= end_seq
;
4283 /* Find place to insert this segment. */
4285 if (!after(TCP_SKB_CB(skb1
)->seq
, seq
))
4287 if (skb_queue_is_first(&tp
->out_of_order_queue
, skb1
)) {
4291 skb1
= skb_queue_prev(&tp
->out_of_order_queue
, skb1
);
4294 /* Do skb overlap to previous one? */
4295 if (skb1
&& before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4296 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4297 /* All the bits are present. Drop. */
4298 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFOMERGE
);
4301 tcp_dsack_set(sk
, seq
, end_seq
);
4304 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
4305 /* Partial overlap. */
4306 tcp_dsack_set(sk
, seq
,
4307 TCP_SKB_CB(skb1
)->end_seq
);
4309 if (skb_queue_is_first(&tp
->out_of_order_queue
,
4313 skb1
= skb_queue_prev(
4314 &tp
->out_of_order_queue
,
4319 __skb_queue_head(&tp
->out_of_order_queue
, skb
);
4321 __skb_queue_after(&tp
->out_of_order_queue
, skb1
, skb
);
4323 /* And clean segments covered by new one as whole. */
4324 while (!skb_queue_is_last(&tp
->out_of_order_queue
, skb
)) {
4325 skb1
= skb_queue_next(&tp
->out_of_order_queue
, skb
);
4327 if (!after(end_seq
, TCP_SKB_CB(skb1
)->seq
))
4329 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4330 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4334 __skb_unlink(skb1
, &tp
->out_of_order_queue
);
4335 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4336 TCP_SKB_CB(skb1
)->end_seq
);
4337 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFOMERGE
);
4342 if (tcp_is_sack(tp
))
4343 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
4346 tcp_grow_window(sk
, skb
);
4347 skb_set_owner_r(skb
, sk
);
4351 static int __must_check
tcp_queue_rcv(struct sock
*sk
, struct sk_buff
*skb
, int hdrlen
,
4355 struct sk_buff
*tail
= skb_peek_tail(&sk
->sk_receive_queue
);
4357 __skb_pull(skb
, hdrlen
);
4359 tcp_try_coalesce(sk
, tail
, skb
, fragstolen
)) ? 1 : 0;
4360 tcp_sk(sk
)->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4362 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4363 skb_set_owner_r(skb
, sk
);
4368 int tcp_send_rcvq(struct sock
*sk
, struct msghdr
*msg
, size_t size
)
4370 struct sk_buff
*skb
;
4376 skb
= alloc_skb(size
, sk
->sk_allocation
);
4380 if (tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))
4383 if (memcpy_from_msg(skb_put(skb
, size
), msg
, size
))
4386 TCP_SKB_CB(skb
)->seq
= tcp_sk(sk
)->rcv_nxt
;
4387 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ size
;
4388 TCP_SKB_CB(skb
)->ack_seq
= tcp_sk(sk
)->snd_una
- 1;
4390 if (tcp_queue_rcv(sk
, skb
, 0, &fragstolen
)) {
4391 WARN_ON_ONCE(fragstolen
); /* should not happen */
4402 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
4404 struct tcp_sock
*tp
= tcp_sk(sk
);
4406 bool fragstolen
= false;
4408 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
)
4412 __skb_pull(skb
, tcp_hdr(skb
)->doff
* 4);
4414 tcp_ecn_accept_cwr(tp
, skb
);
4416 tp
->rx_opt
.dsack
= 0;
4418 /* Queue data for delivery to the user.
4419 * Packets in sequence go to the receive queue.
4420 * Out of sequence packets to the out_of_order_queue.
4422 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
4423 if (tcp_receive_window(tp
) == 0)
4426 /* Ok. In sequence. In window. */
4427 if (tp
->ucopy
.task
== current
&&
4428 tp
->copied_seq
== tp
->rcv_nxt
&& tp
->ucopy
.len
&&
4429 sock_owned_by_user(sk
) && !tp
->urg_data
) {
4430 int chunk
= min_t(unsigned int, skb
->len
,
4433 __set_current_state(TASK_RUNNING
);
4436 if (!skb_copy_datagram_msg(skb
, 0, tp
->ucopy
.msg
, chunk
)) {
4437 tp
->ucopy
.len
-= chunk
;
4438 tp
->copied_seq
+= chunk
;
4439 eaten
= (chunk
== skb
->len
);
4440 tcp_rcv_space_adjust(sk
);
4448 tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))
4451 eaten
= tcp_queue_rcv(sk
, skb
, 0, &fragstolen
);
4453 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4455 tcp_event_data_recv(sk
, skb
);
4456 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
4459 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4462 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4463 * gap in queue is filled.
4465 if (skb_queue_empty(&tp
->out_of_order_queue
))
4466 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
4469 if (tp
->rx_opt
.num_sacks
)
4470 tcp_sack_remove(tp
);
4472 tcp_fast_path_check(sk
);
4475 kfree_skb_partial(skb
, fragstolen
);
4476 if (!sock_flag(sk
, SOCK_DEAD
))
4477 sk
->sk_data_ready(sk
);
4481 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4482 /* A retransmit, 2nd most common case. Force an immediate ack. */
4483 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4484 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4487 tcp_enter_quickack_mode(sk
);
4488 inet_csk_schedule_ack(sk
);
4494 /* Out of window. F.e. zero window probe. */
4495 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
4498 tcp_enter_quickack_mode(sk
);
4500 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4501 /* Partial packet, seq < rcv_next < end_seq */
4502 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
4503 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4504 TCP_SKB_CB(skb
)->end_seq
);
4506 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
4508 /* If window is closed, drop tail of packet. But after
4509 * remembering D-SACK for its head made in previous line.
4511 if (!tcp_receive_window(tp
))
4516 tcp_data_queue_ofo(sk
, skb
);
4519 static struct sk_buff
*tcp_collapse_one(struct sock
*sk
, struct sk_buff
*skb
,
4520 struct sk_buff_head
*list
)
4522 struct sk_buff
*next
= NULL
;
4524 if (!skb_queue_is_last(list
, skb
))
4525 next
= skb_queue_next(list
, skb
);
4527 __skb_unlink(skb
, list
);
4529 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRCVCOLLAPSED
);
4534 /* Collapse contiguous sequence of skbs head..tail with
4535 * sequence numbers start..end.
4537 * If tail is NULL, this means until the end of the list.
4539 * Segments with FIN/SYN are not collapsed (only because this
4543 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
,
4544 struct sk_buff
*head
, struct sk_buff
*tail
,
4547 struct sk_buff
*skb
, *n
;
4550 /* First, check that queue is collapsible and find
4551 * the point where collapsing can be useful. */
4555 skb_queue_walk_from_safe(list
, skb
, n
) {
4558 /* No new bits? It is possible on ofo queue. */
4559 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4560 skb
= tcp_collapse_one(sk
, skb
, list
);
4566 /* The first skb to collapse is:
4568 * - bloated or contains data before "start" or
4569 * overlaps to the next one.
4571 if (!(TCP_SKB_CB(skb
)->tcp_flags
& (TCPHDR_SYN
| TCPHDR_FIN
)) &&
4572 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
4573 before(TCP_SKB_CB(skb
)->seq
, start
))) {
4574 end_of_skbs
= false;
4578 if (!skb_queue_is_last(list
, skb
)) {
4579 struct sk_buff
*next
= skb_queue_next(list
, skb
);
4581 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(next
)->seq
) {
4582 end_of_skbs
= false;
4587 /* Decided to skip this, advance start seq. */
4588 start
= TCP_SKB_CB(skb
)->end_seq
;
4591 (TCP_SKB_CB(skb
)->tcp_flags
& (TCPHDR_SYN
| TCPHDR_FIN
)))
4594 while (before(start
, end
)) {
4595 int copy
= min_t(int, SKB_MAX_ORDER(0, 0), end
- start
);
4596 struct sk_buff
*nskb
;
4598 nskb
= alloc_skb(copy
, GFP_ATOMIC
);
4602 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
4603 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
4604 __skb_queue_before(list
, skb
, nskb
);
4605 skb_set_owner_r(nskb
, sk
);
4607 /* Copy data, releasing collapsed skbs. */
4609 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
4610 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
4614 size
= min(copy
, size
);
4615 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
4617 TCP_SKB_CB(nskb
)->end_seq
+= size
;
4621 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4622 skb
= tcp_collapse_one(sk
, skb
, list
);
4625 (TCP_SKB_CB(skb
)->tcp_flags
& (TCPHDR_SYN
| TCPHDR_FIN
)))
4632 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4633 * and tcp_collapse() them until all the queue is collapsed.
4635 static void tcp_collapse_ofo_queue(struct sock
*sk
)
4637 struct tcp_sock
*tp
= tcp_sk(sk
);
4638 struct sk_buff
*skb
= skb_peek(&tp
->out_of_order_queue
);
4639 struct sk_buff
*head
;
4645 start
= TCP_SKB_CB(skb
)->seq
;
4646 end
= TCP_SKB_CB(skb
)->end_seq
;
4650 struct sk_buff
*next
= NULL
;
4652 if (!skb_queue_is_last(&tp
->out_of_order_queue
, skb
))
4653 next
= skb_queue_next(&tp
->out_of_order_queue
, skb
);
4656 /* Segment is terminated when we see gap or when
4657 * we are at the end of all the queue. */
4659 after(TCP_SKB_CB(skb
)->seq
, end
) ||
4660 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
4661 tcp_collapse(sk
, &tp
->out_of_order_queue
,
4662 head
, skb
, start
, end
);
4666 /* Start new segment */
4667 start
= TCP_SKB_CB(skb
)->seq
;
4668 end
= TCP_SKB_CB(skb
)->end_seq
;
4670 if (before(TCP_SKB_CB(skb
)->seq
, start
))
4671 start
= TCP_SKB_CB(skb
)->seq
;
4672 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
4673 end
= TCP_SKB_CB(skb
)->end_seq
;
4679 * Purge the out-of-order queue.
4680 * Return true if queue was pruned.
4682 static bool tcp_prune_ofo_queue(struct sock
*sk
)
4684 struct tcp_sock
*tp
= tcp_sk(sk
);
4687 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4688 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_OFOPRUNED
);
4689 __skb_queue_purge(&tp
->out_of_order_queue
);
4691 /* Reset SACK state. A conforming SACK implementation will
4692 * do the same at a timeout based retransmit. When a connection
4693 * is in a sad state like this, we care only about integrity
4694 * of the connection not performance.
4696 if (tp
->rx_opt
.sack_ok
)
4697 tcp_sack_reset(&tp
->rx_opt
);
4704 /* Reduce allocated memory if we can, trying to get
4705 * the socket within its memory limits again.
4707 * Return less than zero if we should start dropping frames
4708 * until the socket owning process reads some of the data
4709 * to stabilize the situation.
4711 static int tcp_prune_queue(struct sock
*sk
)
4713 struct tcp_sock
*tp
= tcp_sk(sk
);
4715 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
4717 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PRUNECALLED
);
4719 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
4720 tcp_clamp_window(sk
);
4721 else if (sk_under_memory_pressure(sk
))
4722 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
4724 tcp_collapse_ofo_queue(sk
);
4725 if (!skb_queue_empty(&sk
->sk_receive_queue
))
4726 tcp_collapse(sk
, &sk
->sk_receive_queue
,
4727 skb_peek(&sk
->sk_receive_queue
),
4729 tp
->copied_seq
, tp
->rcv_nxt
);
4732 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4735 /* Collapsing did not help, destructive actions follow.
4736 * This must not ever occur. */
4738 tcp_prune_ofo_queue(sk
);
4740 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4743 /* If we are really being abused, tell the caller to silently
4744 * drop receive data on the floor. It will get retransmitted
4745 * and hopefully then we'll have sufficient space.
4747 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_RCVPRUNED
);
4749 /* Massive buffer overcommit. */
4754 static bool tcp_should_expand_sndbuf(const struct sock
*sk
)
4756 const struct tcp_sock
*tp
= tcp_sk(sk
);
4758 /* If the user specified a specific send buffer setting, do
4761 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
4764 /* If we are under global TCP memory pressure, do not expand. */
4765 if (sk_under_memory_pressure(sk
))
4768 /* If we are under soft global TCP memory pressure, do not expand. */
4769 if (sk_memory_allocated(sk
) >= sk_prot_mem_limits(sk
, 0))
4772 /* If we filled the congestion window, do not expand. */
4773 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
4779 /* When incoming ACK allowed to free some skb from write_queue,
4780 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4781 * on the exit from tcp input handler.
4783 * PROBLEM: sndbuf expansion does not work well with largesend.
4785 static void tcp_new_space(struct sock
*sk
)
4787 struct tcp_sock
*tp
= tcp_sk(sk
);
4789 if (tcp_should_expand_sndbuf(sk
)) {
4790 tcp_sndbuf_expand(sk
);
4791 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4794 sk
->sk_write_space(sk
);
4797 static void tcp_check_space(struct sock
*sk
)
4799 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
4800 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
4801 if (sk
->sk_socket
&&
4802 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
4807 static inline void tcp_data_snd_check(struct sock
*sk
)
4809 tcp_push_pending_frames(sk
);
4810 tcp_check_space(sk
);
4814 * Check if sending an ack is needed.
4816 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
4818 struct tcp_sock
*tp
= tcp_sk(sk
);
4820 /* More than one full frame received... */
4821 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
&&
4822 /* ... and right edge of window advances far enough.
4823 * (tcp_recvmsg() will send ACK otherwise). Or...
4825 __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
4826 /* We ACK each frame or... */
4827 tcp_in_quickack_mode(sk
) ||
4828 /* We have out of order data. */
4829 (ofo_possible
&& skb_peek(&tp
->out_of_order_queue
))) {
4830 /* Then ack it now */
4833 /* Else, send delayed ack. */
4834 tcp_send_delayed_ack(sk
);
4838 static inline void tcp_ack_snd_check(struct sock
*sk
)
4840 if (!inet_csk_ack_scheduled(sk
)) {
4841 /* We sent a data segment already. */
4844 __tcp_ack_snd_check(sk
, 1);
4848 * This routine is only called when we have urgent data
4849 * signaled. Its the 'slow' part of tcp_urg. It could be
4850 * moved inline now as tcp_urg is only called from one
4851 * place. We handle URGent data wrong. We have to - as
4852 * BSD still doesn't use the correction from RFC961.
4853 * For 1003.1g we should support a new option TCP_STDURG to permit
4854 * either form (or just set the sysctl tcp_stdurg).
4857 static void tcp_check_urg(struct sock
*sk
, const struct tcphdr
*th
)
4859 struct tcp_sock
*tp
= tcp_sk(sk
);
4860 u32 ptr
= ntohs(th
->urg_ptr
);
4862 if (ptr
&& !sysctl_tcp_stdurg
)
4864 ptr
+= ntohl(th
->seq
);
4866 /* Ignore urgent data that we've already seen and read. */
4867 if (after(tp
->copied_seq
, ptr
))
4870 /* Do not replay urg ptr.
4872 * NOTE: interesting situation not covered by specs.
4873 * Misbehaving sender may send urg ptr, pointing to segment,
4874 * which we already have in ofo queue. We are not able to fetch
4875 * such data and will stay in TCP_URG_NOTYET until will be eaten
4876 * by recvmsg(). Seems, we are not obliged to handle such wicked
4877 * situations. But it is worth to think about possibility of some
4878 * DoSes using some hypothetical application level deadlock.
4880 if (before(ptr
, tp
->rcv_nxt
))
4883 /* Do we already have a newer (or duplicate) urgent pointer? */
4884 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
4887 /* Tell the world about our new urgent pointer. */
4890 /* We may be adding urgent data when the last byte read was
4891 * urgent. To do this requires some care. We cannot just ignore
4892 * tp->copied_seq since we would read the last urgent byte again
4893 * as data, nor can we alter copied_seq until this data arrives
4894 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4896 * NOTE. Double Dutch. Rendering to plain English: author of comment
4897 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4898 * and expect that both A and B disappear from stream. This is _wrong_.
4899 * Though this happens in BSD with high probability, this is occasional.
4900 * Any application relying on this is buggy. Note also, that fix "works"
4901 * only in this artificial test. Insert some normal data between A and B and we will
4902 * decline of BSD again. Verdict: it is better to remove to trap
4905 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
4906 !sock_flag(sk
, SOCK_URGINLINE
) && tp
->copied_seq
!= tp
->rcv_nxt
) {
4907 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
4909 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4910 __skb_unlink(skb
, &sk
->sk_receive_queue
);
4915 tp
->urg_data
= TCP_URG_NOTYET
;
4918 /* Disable header prediction. */
4922 /* This is the 'fast' part of urgent handling. */
4923 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, const struct tcphdr
*th
)
4925 struct tcp_sock
*tp
= tcp_sk(sk
);
4927 /* Check if we get a new urgent pointer - normally not. */
4929 tcp_check_urg(sk
, th
);
4931 /* Do we wait for any urgent data? - normally not... */
4932 if (tp
->urg_data
== TCP_URG_NOTYET
) {
4933 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
4936 /* Is the urgent pointer pointing into this packet? */
4937 if (ptr
< skb
->len
) {
4939 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
4941 tp
->urg_data
= TCP_URG_VALID
| tmp
;
4942 if (!sock_flag(sk
, SOCK_DEAD
))
4943 sk
->sk_data_ready(sk
);
4948 static int tcp_copy_to_iovec(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
4950 struct tcp_sock
*tp
= tcp_sk(sk
);
4951 int chunk
= skb
->len
- hlen
;
4955 if (skb_csum_unnecessary(skb
))
4956 err
= skb_copy_datagram_msg(skb
, hlen
, tp
->ucopy
.msg
, chunk
);
4958 err
= skb_copy_and_csum_datagram_msg(skb
, hlen
, tp
->ucopy
.msg
);
4961 tp
->ucopy
.len
-= chunk
;
4962 tp
->copied_seq
+= chunk
;
4963 tcp_rcv_space_adjust(sk
);
4970 static __sum16
__tcp_checksum_complete_user(struct sock
*sk
,
4971 struct sk_buff
*skb
)
4975 if (sock_owned_by_user(sk
)) {
4977 result
= __tcp_checksum_complete(skb
);
4980 result
= __tcp_checksum_complete(skb
);
4985 static inline bool tcp_checksum_complete_user(struct sock
*sk
,
4986 struct sk_buff
*skb
)
4988 return !skb_csum_unnecessary(skb
) &&
4989 __tcp_checksum_complete_user(sk
, skb
);
4992 /* Does PAWS and seqno based validation of an incoming segment, flags will
4993 * play significant role here.
4995 static bool tcp_validate_incoming(struct sock
*sk
, struct sk_buff
*skb
,
4996 const struct tcphdr
*th
, int syn_inerr
)
4998 struct tcp_sock
*tp
= tcp_sk(sk
);
5000 /* RFC1323: H1. Apply PAWS check first. */
5001 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
5002 tcp_paws_discard(sk
, skb
)) {
5004 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSESTABREJECTED
);
5005 if (!tcp_oow_rate_limited(sock_net(sk
), skb
,
5006 LINUX_MIB_TCPACKSKIPPEDPAWS
,
5007 &tp
->last_oow_ack_time
))
5008 tcp_send_dupack(sk
, skb
);
5011 /* Reset is accepted even if it did not pass PAWS. */
5014 /* Step 1: check sequence number */
5015 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5016 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5017 * (RST) segments are validated by checking their SEQ-fields."
5018 * And page 69: "If an incoming segment is not acceptable,
5019 * an acknowledgment should be sent in reply (unless the RST
5020 * bit is set, if so drop the segment and return)".
5025 if (!tcp_oow_rate_limited(sock_net(sk
), skb
,
5026 LINUX_MIB_TCPACKSKIPPEDSEQ
,
5027 &tp
->last_oow_ack_time
))
5028 tcp_send_dupack(sk
, skb
);
5033 /* Step 2: check RST bit */
5036 * If sequence number exactly matches RCV.NXT, then
5037 * RESET the connection
5039 * Send a challenge ACK
5041 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
)
5044 tcp_send_challenge_ack(sk
, skb
);
5048 /* step 3: check security and precedence [ignored] */
5050 /* step 4: Check for a SYN
5051 * RFC 5961 4.2 : Send a challenge ack
5056 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5057 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSYNCHALLENGE
);
5058 tcp_send_challenge_ack(sk
, skb
);
5070 * TCP receive function for the ESTABLISHED state.
5072 * It is split into a fast path and a slow path. The fast path is
5074 * - A zero window was announced from us - zero window probing
5075 * is only handled properly in the slow path.
5076 * - Out of order segments arrived.
5077 * - Urgent data is expected.
5078 * - There is no buffer space left
5079 * - Unexpected TCP flags/window values/header lengths are received
5080 * (detected by checking the TCP header against pred_flags)
5081 * - Data is sent in both directions. Fast path only supports pure senders
5082 * or pure receivers (this means either the sequence number or the ack
5083 * value must stay constant)
5084 * - Unexpected TCP option.
5086 * When these conditions are not satisfied it drops into a standard
5087 * receive procedure patterned after RFC793 to handle all cases.
5088 * The first three cases are guaranteed by proper pred_flags setting,
5089 * the rest is checked inline. Fast processing is turned on in
5090 * tcp_data_queue when everything is OK.
5092 void tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
5093 const struct tcphdr
*th
, unsigned int len
)
5095 struct tcp_sock
*tp
= tcp_sk(sk
);
5097 if (unlikely(sk
->sk_rx_dst
== NULL
))
5098 inet_csk(sk
)->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
5100 * Header prediction.
5101 * The code loosely follows the one in the famous
5102 * "30 instruction TCP receive" Van Jacobson mail.
5104 * Van's trick is to deposit buffers into socket queue
5105 * on a device interrupt, to call tcp_recv function
5106 * on the receive process context and checksum and copy
5107 * the buffer to user space. smart...
5109 * Our current scheme is not silly either but we take the
5110 * extra cost of the net_bh soft interrupt processing...
5111 * We do checksum and copy also but from device to kernel.
5114 tp
->rx_opt
.saw_tstamp
= 0;
5116 /* pred_flags is 0xS?10 << 16 + snd_wnd
5117 * if header_prediction is to be made
5118 * 'S' will always be tp->tcp_header_len >> 2
5119 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5120 * turn it off (when there are holes in the receive
5121 * space for instance)
5122 * PSH flag is ignored.
5125 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
5126 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
&&
5127 !after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
)) {
5128 int tcp_header_len
= tp
->tcp_header_len
;
5130 /* Timestamp header prediction: tcp_header_len
5131 * is automatically equal to th->doff*4 due to pred_flags
5135 /* Check timestamp */
5136 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
5137 /* No? Slow path! */
5138 if (!tcp_parse_aligned_timestamp(tp
, th
))
5141 /* If PAWS failed, check it more carefully in slow path */
5142 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
5145 /* DO NOT update ts_recent here, if checksum fails
5146 * and timestamp was corrupted part, it will result
5147 * in a hung connection since we will drop all
5148 * future packets due to the PAWS test.
5152 if (len
<= tcp_header_len
) {
5153 /* Bulk data transfer: sender */
5154 if (len
== tcp_header_len
) {
5155 /* Predicted packet is in window by definition.
5156 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5157 * Hence, check seq<=rcv_wup reduces to:
5159 if (tcp_header_len
==
5160 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5161 tp
->rcv_nxt
== tp
->rcv_wup
)
5162 tcp_store_ts_recent(tp
);
5164 /* We know that such packets are checksummed
5167 tcp_ack(sk
, skb
, 0);
5169 tcp_data_snd_check(sk
);
5171 } else { /* Header too small */
5172 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5177 bool fragstolen
= false;
5179 if (tp
->ucopy
.task
== current
&&
5180 tp
->copied_seq
== tp
->rcv_nxt
&&
5181 len
- tcp_header_len
<= tp
->ucopy
.len
&&
5182 sock_owned_by_user(sk
)) {
5183 __set_current_state(TASK_RUNNING
);
5185 if (!tcp_copy_to_iovec(sk
, skb
, tcp_header_len
)) {
5186 /* Predicted packet is in window by definition.
5187 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5188 * Hence, check seq<=rcv_wup reduces to:
5190 if (tcp_header_len
==
5191 (sizeof(struct tcphdr
) +
5192 TCPOLEN_TSTAMP_ALIGNED
) &&
5193 tp
->rcv_nxt
== tp
->rcv_wup
)
5194 tcp_store_ts_recent(tp
);
5196 tcp_rcv_rtt_measure_ts(sk
, skb
);
5198 __skb_pull(skb
, tcp_header_len
);
5199 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
5200 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITSTOUSER
);
5205 if (tcp_checksum_complete_user(sk
, skb
))
5208 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
5211 /* Predicted packet is in window by definition.
5212 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5213 * Hence, check seq<=rcv_wup reduces to:
5215 if (tcp_header_len
==
5216 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5217 tp
->rcv_nxt
== tp
->rcv_wup
)
5218 tcp_store_ts_recent(tp
);
5220 tcp_rcv_rtt_measure_ts(sk
, skb
);
5222 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITS
);
5224 /* Bulk data transfer: receiver */
5225 eaten
= tcp_queue_rcv(sk
, skb
, tcp_header_len
,
5229 tcp_event_data_recv(sk
, skb
);
5231 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
5232 /* Well, only one small jumplet in fast path... */
5233 tcp_ack(sk
, skb
, FLAG_DATA
);
5234 tcp_data_snd_check(sk
);
5235 if (!inet_csk_ack_scheduled(sk
))
5239 __tcp_ack_snd_check(sk
, 0);
5242 kfree_skb_partial(skb
, fragstolen
);
5243 sk
->sk_data_ready(sk
);
5249 if (len
< (th
->doff
<< 2) || tcp_checksum_complete_user(sk
, skb
))
5252 if (!th
->ack
&& !th
->rst
&& !th
->syn
)
5256 * Standard slow path.
5259 if (!tcp_validate_incoming(sk
, skb
, th
, 1))
5263 if (tcp_ack(sk
, skb
, FLAG_SLOWPATH
| FLAG_UPDATE_TS_RECENT
) < 0)
5266 tcp_rcv_rtt_measure_ts(sk
, skb
);
5268 /* Process urgent data. */
5269 tcp_urg(sk
, skb
, th
);
5271 /* step 7: process the segment text */
5272 tcp_data_queue(sk
, skb
);
5274 tcp_data_snd_check(sk
);
5275 tcp_ack_snd_check(sk
);
5279 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_CSUMERRORS
);
5280 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5285 EXPORT_SYMBOL(tcp_rcv_established
);
5287 void tcp_finish_connect(struct sock
*sk
, struct sk_buff
*skb
)
5289 struct tcp_sock
*tp
= tcp_sk(sk
);
5290 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5292 tcp_set_state(sk
, TCP_ESTABLISHED
);
5295 icsk
->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
5296 security_inet_conn_established(sk
, skb
);
5299 /* Make sure socket is routed, for correct metrics. */
5300 icsk
->icsk_af_ops
->rebuild_header(sk
);
5302 tcp_init_metrics(sk
);
5304 tcp_init_congestion_control(sk
);
5306 /* Prevent spurious tcp_cwnd_restart() on first data
5309 tp
->lsndtime
= tcp_time_stamp
;
5311 tcp_init_buffer_space(sk
);
5313 if (sock_flag(sk
, SOCK_KEEPOPEN
))
5314 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
5316 if (!tp
->rx_opt
.snd_wscale
)
5317 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
5321 if (!sock_flag(sk
, SOCK_DEAD
)) {
5322 sk
->sk_state_change(sk
);
5323 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5327 static bool tcp_rcv_fastopen_synack(struct sock
*sk
, struct sk_buff
*synack
,
5328 struct tcp_fastopen_cookie
*cookie
)
5330 struct tcp_sock
*tp
= tcp_sk(sk
);
5331 struct sk_buff
*data
= tp
->syn_data
? tcp_write_queue_head(sk
) : NULL
;
5332 u16 mss
= tp
->rx_opt
.mss_clamp
;
5335 if (mss
== tp
->rx_opt
.user_mss
) {
5336 struct tcp_options_received opt
;
5338 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5339 tcp_clear_options(&opt
);
5340 opt
.user_mss
= opt
.mss_clamp
= 0;
5341 tcp_parse_options(synack
, &opt
, 0, NULL
);
5342 mss
= opt
.mss_clamp
;
5345 if (!tp
->syn_fastopen
) /* Ignore an unsolicited cookie */
5348 /* The SYN-ACK neither has cookie nor acknowledges the data. Presumably
5349 * the remote receives only the retransmitted (regular) SYNs: either
5350 * the original SYN-data or the corresponding SYN-ACK is lost.
5352 syn_drop
= (cookie
->len
<= 0 && data
&& tp
->total_retrans
);
5354 tcp_fastopen_cache_set(sk
, mss
, cookie
, syn_drop
);
5356 if (data
) { /* Retransmit unacked data in SYN */
5357 tcp_for_write_queue_from(data
, sk
) {
5358 if (data
== tcp_send_head(sk
) ||
5359 __tcp_retransmit_skb(sk
, data
))
5363 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPFASTOPENACTIVEFAIL
);
5366 tp
->syn_data_acked
= tp
->syn_data
;
5367 if (tp
->syn_data_acked
)
5368 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPFASTOPENACTIVE
);
5372 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5373 const struct tcphdr
*th
, unsigned int len
)
5375 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5376 struct tcp_sock
*tp
= tcp_sk(sk
);
5377 struct tcp_fastopen_cookie foc
= { .len
= -1 };
5378 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
5380 tcp_parse_options(skb
, &tp
->rx_opt
, 0, &foc
);
5381 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
5382 tp
->rx_opt
.rcv_tsecr
-= tp
->tsoffset
;
5386 * "If the state is SYN-SENT then
5387 * first check the ACK bit
5388 * If the ACK bit is set
5389 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5390 * a reset (unless the RST bit is set, if so drop
5391 * the segment and return)"
5393 if (!after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_una
) ||
5394 after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
))
5395 goto reset_and_undo
;
5397 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
5398 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
5400 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSACTIVEREJECTED
);
5401 goto reset_and_undo
;
5404 /* Now ACK is acceptable.
5406 * "If the RST bit is set
5407 * If the ACK was acceptable then signal the user "error:
5408 * connection reset", drop the segment, enter CLOSED state,
5409 * delete TCB, and return."
5418 * "fifth, if neither of the SYN or RST bits is set then
5419 * drop the segment and return."
5425 goto discard_and_undo
;
5428 * "If the SYN bit is on ...
5429 * are acceptable then ...
5430 * (our SYN has been ACKed), change the connection
5431 * state to ESTABLISHED..."
5434 tcp_ecn_rcv_synack(tp
, th
);
5436 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5437 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
5439 /* Ok.. it's good. Set up sequence numbers and
5440 * move to established.
5442 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5443 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5445 /* RFC1323: The window in SYN & SYN/ACK segments is
5448 tp
->snd_wnd
= ntohs(th
->window
);
5450 if (!tp
->rx_opt
.wscale_ok
) {
5451 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
5452 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
5455 if (tp
->rx_opt
.saw_tstamp
) {
5456 tp
->rx_opt
.tstamp_ok
= 1;
5457 tp
->tcp_header_len
=
5458 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5459 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5460 tcp_store_ts_recent(tp
);
5462 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5465 if (tcp_is_sack(tp
) && sysctl_tcp_fack
)
5466 tcp_enable_fack(tp
);
5469 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5470 tcp_initialize_rcv_mss(sk
);
5472 /* Remember, tcp_poll() does not lock socket!
5473 * Change state from SYN-SENT only after copied_seq
5474 * is initialized. */
5475 tp
->copied_seq
= tp
->rcv_nxt
;
5479 tcp_finish_connect(sk
, skb
);
5481 if ((tp
->syn_fastopen
|| tp
->syn_data
) &&
5482 tcp_rcv_fastopen_synack(sk
, skb
, &foc
))
5485 if (sk
->sk_write_pending
||
5486 icsk
->icsk_accept_queue
.rskq_defer_accept
||
5487 icsk
->icsk_ack
.pingpong
) {
5488 /* Save one ACK. Data will be ready after
5489 * several ticks, if write_pending is set.
5491 * It may be deleted, but with this feature tcpdumps
5492 * look so _wonderfully_ clever, that I was not able
5493 * to stand against the temptation 8) --ANK
5495 inet_csk_schedule_ack(sk
);
5496 icsk
->icsk_ack
.lrcvtime
= tcp_time_stamp
;
5497 tcp_enter_quickack_mode(sk
);
5498 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
5499 TCP_DELACK_MAX
, TCP_RTO_MAX
);
5510 /* No ACK in the segment */
5514 * "If the RST bit is set
5516 * Otherwise (no ACK) drop the segment and return."
5519 goto discard_and_undo
;
5523 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&&
5524 tcp_paws_reject(&tp
->rx_opt
, 0))
5525 goto discard_and_undo
;
5528 /* We see SYN without ACK. It is attempt of
5529 * simultaneous connect with crossed SYNs.
5530 * Particularly, it can be connect to self.
5532 tcp_set_state(sk
, TCP_SYN_RECV
);
5534 if (tp
->rx_opt
.saw_tstamp
) {
5535 tp
->rx_opt
.tstamp_ok
= 1;
5536 tcp_store_ts_recent(tp
);
5537 tp
->tcp_header_len
=
5538 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5540 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5543 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5544 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5546 /* RFC1323: The window in SYN & SYN/ACK segments is
5549 tp
->snd_wnd
= ntohs(th
->window
);
5550 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
5551 tp
->max_window
= tp
->snd_wnd
;
5553 tcp_ecn_rcv_syn(tp
, th
);
5556 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5557 tcp_initialize_rcv_mss(sk
);
5559 tcp_send_synack(sk
);
5561 /* Note, we could accept data and URG from this segment.
5562 * There are no obstacles to make this (except that we must
5563 * either change tcp_recvmsg() to prevent it from returning data
5564 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5566 * However, if we ignore data in ACKless segments sometimes,
5567 * we have no reasons to accept it sometimes.
5568 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5569 * is not flawless. So, discard packet for sanity.
5570 * Uncomment this return to process the data.
5577 /* "fifth, if neither of the SYN or RST bits is set then
5578 * drop the segment and return."
5582 tcp_clear_options(&tp
->rx_opt
);
5583 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5587 tcp_clear_options(&tp
->rx_opt
);
5588 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5593 * This function implements the receiving procedure of RFC 793 for
5594 * all states except ESTABLISHED and TIME_WAIT.
5595 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5596 * address independent.
5599 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5600 const struct tcphdr
*th
, unsigned int len
)
5602 struct tcp_sock
*tp
= tcp_sk(sk
);
5603 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5604 struct request_sock
*req
;
5609 tp
->rx_opt
.saw_tstamp
= 0;
5611 switch (sk
->sk_state
) {
5625 if (icsk
->icsk_af_ops
->conn_request(sk
, skb
) < 0)
5628 /* Now we have several options: In theory there is
5629 * nothing else in the frame. KA9Q has an option to
5630 * send data with the syn, BSD accepts data with the
5631 * syn up to the [to be] advertised window and
5632 * Solaris 2.1 gives you a protocol error. For now
5633 * we just ignore it, that fits the spec precisely
5634 * and avoids incompatibilities. It would be nice in
5635 * future to drop through and process the data.
5637 * Now that TTCP is starting to be used we ought to
5639 * But, this leaves one open to an easy denial of
5640 * service attack, and SYN cookies can't defend
5641 * against this problem. So, we drop the data
5642 * in the interest of security over speed unless
5643 * it's still in use.
5651 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
, len
);
5655 /* Do step6 onward by hand. */
5656 tcp_urg(sk
, skb
, th
);
5658 tcp_data_snd_check(sk
);
5662 req
= tp
->fastopen_rsk
;
5664 WARN_ON_ONCE(sk
->sk_state
!= TCP_SYN_RECV
&&
5665 sk
->sk_state
!= TCP_FIN_WAIT1
);
5667 if (tcp_check_req(sk
, skb
, req
, NULL
, true) == NULL
)
5671 if (!th
->ack
&& !th
->rst
&& !th
->syn
)
5674 if (!tcp_validate_incoming(sk
, skb
, th
, 0))
5677 /* step 5: check the ACK field */
5678 acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
|
5679 FLAG_UPDATE_TS_RECENT
) > 0;
5681 switch (sk
->sk_state
) {
5686 /* Once we leave TCP_SYN_RECV, we no longer need req
5690 synack_stamp
= tcp_rsk(req
)->snt_synack
;
5691 tp
->total_retrans
= req
->num_retrans
;
5692 reqsk_fastopen_remove(sk
, req
, false);
5694 synack_stamp
= tp
->lsndtime
;
5695 /* Make sure socket is routed, for correct metrics. */
5696 icsk
->icsk_af_ops
->rebuild_header(sk
);
5697 tcp_init_congestion_control(sk
);
5700 tp
->copied_seq
= tp
->rcv_nxt
;
5701 tcp_init_buffer_space(sk
);
5704 tcp_set_state(sk
, TCP_ESTABLISHED
);
5705 sk
->sk_state_change(sk
);
5707 /* Note, that this wakeup is only for marginal crossed SYN case.
5708 * Passively open sockets are not waked up, because
5709 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5712 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5714 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
5715 tp
->snd_wnd
= ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
;
5716 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5717 tcp_synack_rtt_meas(sk
, synack_stamp
);
5719 if (tp
->rx_opt
.tstamp_ok
)
5720 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5723 /* Re-arm the timer because data may have been sent out.
5724 * This is similar to the regular data transmission case
5725 * when new data has just been ack'ed.
5727 * (TFO) - we could try to be more aggressive and
5728 * retransmitting any data sooner based on when they
5733 tcp_init_metrics(sk
);
5735 tcp_update_pacing_rate(sk
);
5737 /* Prevent spurious tcp_cwnd_restart() on first data packet */
5738 tp
->lsndtime
= tcp_time_stamp
;
5740 tcp_initialize_rcv_mss(sk
);
5741 tcp_fast_path_on(tp
);
5744 case TCP_FIN_WAIT1
: {
5745 struct dst_entry
*dst
;
5748 /* If we enter the TCP_FIN_WAIT1 state and we are a
5749 * Fast Open socket and this is the first acceptable
5750 * ACK we have received, this would have acknowledged
5751 * our SYNACK so stop the SYNACK timer.
5754 /* Return RST if ack_seq is invalid.
5755 * Note that RFC793 only says to generate a
5756 * DUPACK for it but for TCP Fast Open it seems
5757 * better to treat this case like TCP_SYN_RECV
5762 /* We no longer need the request sock. */
5763 reqsk_fastopen_remove(sk
, req
, false);
5766 if (tp
->snd_una
!= tp
->write_seq
)
5769 tcp_set_state(sk
, TCP_FIN_WAIT2
);
5770 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
5772 dst
= __sk_dst_get(sk
);
5776 if (!sock_flag(sk
, SOCK_DEAD
)) {
5777 /* Wake up lingering close() */
5778 sk
->sk_state_change(sk
);
5782 if (tp
->linger2
< 0 ||
5783 (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5784 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
))) {
5786 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
5790 tmo
= tcp_fin_time(sk
);
5791 if (tmo
> TCP_TIMEWAIT_LEN
) {
5792 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
5793 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
5794 /* Bad case. We could lose such FIN otherwise.
5795 * It is not a big problem, but it looks confusing
5796 * and not so rare event. We still can lose it now,
5797 * if it spins in bh_lock_sock(), but it is really
5800 inet_csk_reset_keepalive_timer(sk
, tmo
);
5802 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
5809 if (tp
->snd_una
== tp
->write_seq
) {
5810 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
5816 if (tp
->snd_una
== tp
->write_seq
) {
5817 tcp_update_metrics(sk
);
5824 /* step 6: check the URG bit */
5825 tcp_urg(sk
, skb
, th
);
5827 /* step 7: process the segment text */
5828 switch (sk
->sk_state
) {
5829 case TCP_CLOSE_WAIT
:
5832 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
5836 /* RFC 793 says to queue data in these states,
5837 * RFC 1122 says we MUST send a reset.
5838 * BSD 4.4 also does reset.
5840 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
5841 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5842 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
5843 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
5849 case TCP_ESTABLISHED
:
5850 tcp_data_queue(sk
, skb
);
5855 /* tcp_data could move socket to TIME-WAIT */
5856 if (sk
->sk_state
!= TCP_CLOSE
) {
5857 tcp_data_snd_check(sk
);
5858 tcp_ack_snd_check(sk
);
5867 EXPORT_SYMBOL(tcp_rcv_state_process
);
5869 static inline void pr_drop_req(struct request_sock
*req
, __u16 port
, int family
)
5871 struct inet_request_sock
*ireq
= inet_rsk(req
);
5873 if (family
== AF_INET
)
5874 net_dbg_ratelimited("drop open request from %pI4/%u\n",
5875 &ireq
->ir_rmt_addr
, port
);
5876 #if IS_ENABLED(CONFIG_IPV6)
5877 else if (family
== AF_INET6
)
5878 net_dbg_ratelimited("drop open request from %pI6/%u\n",
5879 &ireq
->ir_v6_rmt_addr
, port
);
5883 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
5885 * If we receive a SYN packet with these bits set, it means a
5886 * network is playing bad games with TOS bits. In order to
5887 * avoid possible false congestion notifications, we disable
5888 * TCP ECN negotiation.
5890 * Exception: tcp_ca wants ECN. This is required for DCTCP
5891 * congestion control: Linux DCTCP asserts ECT on all packets,
5892 * including SYN, which is most optimal solution; however,
5893 * others, such as FreeBSD do not.
5895 static void tcp_ecn_create_request(struct request_sock
*req
,
5896 const struct sk_buff
*skb
,
5897 const struct sock
*listen_sk
,
5898 const struct dst_entry
*dst
)
5900 const struct tcphdr
*th
= tcp_hdr(skb
);
5901 const struct net
*net
= sock_net(listen_sk
);
5902 bool th_ecn
= th
->ece
&& th
->cwr
;
5908 ect
= !INET_ECN_is_not_ect(TCP_SKB_CB(skb
)->ip_dsfield
);
5909 ecn_ok
= net
->ipv4
.sysctl_tcp_ecn
|| dst_feature(dst
, RTAX_FEATURE_ECN
);
5911 if ((!ect
&& ecn_ok
) || tcp_ca_needs_ecn(listen_sk
))
5912 inet_rsk(req
)->ecn_ok
= 1;
5915 int tcp_conn_request(struct request_sock_ops
*rsk_ops
,
5916 const struct tcp_request_sock_ops
*af_ops
,
5917 struct sock
*sk
, struct sk_buff
*skb
)
5919 struct tcp_options_received tmp_opt
;
5920 struct request_sock
*req
;
5921 struct tcp_sock
*tp
= tcp_sk(sk
);
5922 struct dst_entry
*dst
= NULL
;
5923 __u32 isn
= TCP_SKB_CB(skb
)->tcp_tw_isn
;
5924 bool want_cookie
= false, fastopen
;
5926 struct tcp_fastopen_cookie foc
= { .len
= -1 };
5930 /* TW buckets are converted to open requests without
5931 * limitations, they conserve resources and peer is
5932 * evidently real one.
5934 if ((sysctl_tcp_syncookies
== 2 ||
5935 inet_csk_reqsk_queue_is_full(sk
)) && !isn
) {
5936 want_cookie
= tcp_syn_flood_action(sk
, skb
, rsk_ops
->slab_name
);
5942 /* Accept backlog is full. If we have already queued enough
5943 * of warm entries in syn queue, drop request. It is better than
5944 * clogging syn queue with openreqs with exponentially increasing
5947 if (sk_acceptq_is_full(sk
) && inet_csk_reqsk_queue_young(sk
) > 1) {
5948 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_LISTENOVERFLOWS
);
5952 req
= inet_reqsk_alloc(rsk_ops
);
5956 tcp_rsk(req
)->af_specific
= af_ops
;
5958 tcp_clear_options(&tmp_opt
);
5959 tmp_opt
.mss_clamp
= af_ops
->mss_clamp
;
5960 tmp_opt
.user_mss
= tp
->rx_opt
.user_mss
;
5961 tcp_parse_options(skb
, &tmp_opt
, 0, want_cookie
? NULL
: &foc
);
5963 if (want_cookie
&& !tmp_opt
.saw_tstamp
)
5964 tcp_clear_options(&tmp_opt
);
5966 tmp_opt
.tstamp_ok
= tmp_opt
.saw_tstamp
;
5967 tcp_openreq_init(req
, &tmp_opt
, skb
, sk
);
5969 af_ops
->init_req(req
, sk
, skb
);
5971 if (security_inet_conn_request(sk
, skb
, req
))
5974 if (!want_cookie
&& !isn
) {
5975 /* VJ's idea. We save last timestamp seen
5976 * from the destination in peer table, when entering
5977 * state TIME-WAIT, and check against it before
5978 * accepting new connection request.
5980 * If "isn" is not zero, this request hit alive
5981 * timewait bucket, so that all the necessary checks
5982 * are made in the function processing timewait state.
5984 if (tcp_death_row
.sysctl_tw_recycle
) {
5987 dst
= af_ops
->route_req(sk
, &fl
, req
, &strict
);
5989 if (dst
&& strict
&&
5990 !tcp_peer_is_proven(req
, dst
, true,
5991 tmp_opt
.saw_tstamp
)) {
5992 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSPASSIVEREJECTED
);
5993 goto drop_and_release
;
5996 /* Kill the following clause, if you dislike this way. */
5997 else if (!sysctl_tcp_syncookies
&&
5998 (sysctl_max_syn_backlog
- inet_csk_reqsk_queue_len(sk
) <
5999 (sysctl_max_syn_backlog
>> 2)) &&
6000 !tcp_peer_is_proven(req
, dst
, false,
6001 tmp_opt
.saw_tstamp
)) {
6002 /* Without syncookies last quarter of
6003 * backlog is filled with destinations,
6004 * proven to be alive.
6005 * It means that we continue to communicate
6006 * to destinations, already remembered
6007 * to the moment of synflood.
6009 pr_drop_req(req
, ntohs(tcp_hdr(skb
)->source
),
6011 goto drop_and_release
;
6014 isn
= af_ops
->init_seq(skb
);
6017 dst
= af_ops
->route_req(sk
, &fl
, req
, NULL
);
6022 tcp_ecn_create_request(req
, skb
, sk
, dst
);
6025 isn
= cookie_init_sequence(af_ops
, sk
, skb
, &req
->mss
);
6026 req
->cookie_ts
= tmp_opt
.tstamp_ok
;
6027 if (!tmp_opt
.tstamp_ok
)
6028 inet_rsk(req
)->ecn_ok
= 0;
6031 tcp_rsk(req
)->snt_isn
= isn
;
6032 tcp_openreq_init_rwin(req
, sk
, dst
);
6033 fastopen
= !want_cookie
&&
6034 tcp_try_fastopen(sk
, skb
, req
, &foc
, dst
);
6035 err
= af_ops
->send_synack(sk
, dst
, &fl
, req
,
6036 skb_get_queue_mapping(skb
), &foc
);
6038 if (err
|| want_cookie
)
6041 tcp_rsk(req
)->listener
= NULL
;
6042 af_ops
->queue_hash_add(sk
, req
, TCP_TIMEOUT_INIT
);
6052 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_LISTENDROPS
);
6055 EXPORT_SYMBOL(tcp_conn_request
);