2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
23 * Pedro Roque : Fast Retransmit/Recovery.
25 * Retransmit queue handled by TCP.
26 * Better retransmit timer handling.
27 * New congestion avoidance.
31 * Eric : Fast Retransmit.
32 * Randy Scott : MSS option defines.
33 * Eric Schenk : Fixes to slow start algorithm.
34 * Eric Schenk : Yet another double ACK bug.
35 * Eric Schenk : Delayed ACK bug fixes.
36 * Eric Schenk : Floyd style fast retrans war avoidance.
37 * David S. Miller : Don't allow zero congestion window.
38 * Eric Schenk : Fix retransmitter so that it sends
39 * next packet on ack of previous packet.
40 * Andi Kleen : Moved open_request checking here
41 * and process RSTs for open_requests.
42 * Andi Kleen : Better prune_queue, and other fixes.
43 * Andrey Savochkin: Fix RTT measurements in the presence of
45 * Andrey Savochkin: Check sequence numbers correctly when
46 * removing SACKs due to in sequence incoming
48 * Andi Kleen: Make sure we never ack data there is not
49 * enough room for. Also make this condition
50 * a fatal error if it might still happen.
51 * Andi Kleen: Add tcp_measure_rcv_mss to make
52 * connections with MSS<min(MTU,ann. MSS)
53 * work without delayed acks.
54 * Andi Kleen: Process packets with PSH set in the
56 * J Hadi Salim: ECN support
59 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
60 * engine. Lots of bugs are found.
61 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
64 #define pr_fmt(fmt) "TCP: " fmt
67 #include <linux/slab.h>
68 #include <linux/module.h>
69 #include <linux/sysctl.h>
70 #include <linux/kernel.h>
73 #include <net/inet_common.h>
74 #include <linux/ipsec.h>
75 #include <asm/unaligned.h>
76 #include <net/netdma.h>
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 EXPORT_SYMBOL(sysctl_tcp_reordering
);
85 int sysctl_tcp_dsack __read_mostly
= 1;
86 int sysctl_tcp_app_win __read_mostly
= 31;
87 int sysctl_tcp_adv_win_scale __read_mostly
= 1;
88 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale
);
90 /* rfc5961 challenge ack rate limiting */
91 int sysctl_tcp_challenge_ack_limit
= 100;
93 int sysctl_tcp_stdurg __read_mostly
;
94 int sysctl_tcp_rfc1337 __read_mostly
;
95 int sysctl_tcp_max_orphans __read_mostly
= NR_FILE
;
96 int sysctl_tcp_frto __read_mostly
= 2;
98 int sysctl_tcp_thin_dupack __read_mostly
;
100 int sysctl_tcp_moderate_rcvbuf __read_mostly
= 1;
101 int sysctl_tcp_early_retrans __read_mostly
= 3;
103 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
104 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
105 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
106 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
107 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
108 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
109 #define FLAG_ECE 0x40 /* ECE in this ACK */
110 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
111 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
112 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
113 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
114 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
115 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
117 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
118 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
119 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
120 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
122 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
123 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
125 /* Adapt the MSS value used to make delayed ack decision to the
128 static void tcp_measure_rcv_mss(struct sock
*sk
, const struct sk_buff
*skb
)
130 struct inet_connection_sock
*icsk
= inet_csk(sk
);
131 const unsigned int lss
= icsk
->icsk_ack
.last_seg_size
;
134 icsk
->icsk_ack
.last_seg_size
= 0;
136 /* skb->len may jitter because of SACKs, even if peer
137 * sends good full-sized frames.
139 len
= skb_shinfo(skb
)->gso_size
? : skb
->len
;
140 if (len
>= icsk
->icsk_ack
.rcv_mss
) {
141 icsk
->icsk_ack
.rcv_mss
= len
;
143 /* Otherwise, we make more careful check taking into account,
144 * that SACKs block is variable.
146 * "len" is invariant segment length, including TCP header.
148 len
+= skb
->data
- skb_transport_header(skb
);
149 if (len
>= TCP_MSS_DEFAULT
+ sizeof(struct tcphdr
) ||
150 /* If PSH is not set, packet should be
151 * full sized, provided peer TCP is not badly broken.
152 * This observation (if it is correct 8)) allows
153 * to handle super-low mtu links fairly.
155 (len
>= TCP_MIN_MSS
+ sizeof(struct tcphdr
) &&
156 !(tcp_flag_word(tcp_hdr(skb
)) & TCP_REMNANT
))) {
157 /* Subtract also invariant (if peer is RFC compliant),
158 * tcp header plus fixed timestamp option length.
159 * Resulting "len" is MSS free of SACK jitter.
161 len
-= tcp_sk(sk
)->tcp_header_len
;
162 icsk
->icsk_ack
.last_seg_size
= len
;
164 icsk
->icsk_ack
.rcv_mss
= len
;
168 if (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
)
169 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED2
;
170 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
174 static void tcp_incr_quickack(struct sock
*sk
)
176 struct inet_connection_sock
*icsk
= inet_csk(sk
);
177 unsigned int quickacks
= tcp_sk(sk
)->rcv_wnd
/ (2 * icsk
->icsk_ack
.rcv_mss
);
181 if (quickacks
> icsk
->icsk_ack
.quick
)
182 icsk
->icsk_ack
.quick
= min(quickacks
, TCP_MAX_QUICKACKS
);
185 static void tcp_enter_quickack_mode(struct sock
*sk
)
187 struct inet_connection_sock
*icsk
= inet_csk(sk
);
188 tcp_incr_quickack(sk
);
189 icsk
->icsk_ack
.pingpong
= 0;
190 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
193 /* Send ACKs quickly, if "quick" count is not exhausted
194 * and the session is not interactive.
197 static inline bool tcp_in_quickack_mode(const struct sock
*sk
)
199 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
201 return icsk
->icsk_ack
.quick
&& !icsk
->icsk_ack
.pingpong
;
204 static inline void TCP_ECN_queue_cwr(struct tcp_sock
*tp
)
206 if (tp
->ecn_flags
& TCP_ECN_OK
)
207 tp
->ecn_flags
|= TCP_ECN_QUEUE_CWR
;
210 static inline void TCP_ECN_accept_cwr(struct tcp_sock
*tp
, const struct sk_buff
*skb
)
212 if (tcp_hdr(skb
)->cwr
)
213 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
216 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock
*tp
)
218 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
221 static inline void TCP_ECN_check_ce(struct tcp_sock
*tp
, const struct sk_buff
*skb
)
223 if (!(tp
->ecn_flags
& TCP_ECN_OK
))
226 switch (TCP_SKB_CB(skb
)->ip_dsfield
& INET_ECN_MASK
) {
227 case INET_ECN_NOT_ECT
:
228 /* Funny extension: if ECT is not set on a segment,
229 * and we already seen ECT on a previous segment,
230 * it is probably a retransmit.
232 if (tp
->ecn_flags
& TCP_ECN_SEEN
)
233 tcp_enter_quickack_mode((struct sock
*)tp
);
236 if (!(tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)) {
237 /* Better not delay acks, sender can have a very low cwnd */
238 tcp_enter_quickack_mode((struct sock
*)tp
);
239 tp
->ecn_flags
|= TCP_ECN_DEMAND_CWR
;
243 tp
->ecn_flags
|= TCP_ECN_SEEN
;
247 static inline void TCP_ECN_rcv_synack(struct tcp_sock
*tp
, const struct tcphdr
*th
)
249 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| th
->cwr
))
250 tp
->ecn_flags
&= ~TCP_ECN_OK
;
253 static inline void TCP_ECN_rcv_syn(struct tcp_sock
*tp
, const struct tcphdr
*th
)
255 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| !th
->cwr
))
256 tp
->ecn_flags
&= ~TCP_ECN_OK
;
259 static bool TCP_ECN_rcv_ecn_echo(const struct tcp_sock
*tp
, const struct tcphdr
*th
)
261 if (th
->ece
&& !th
->syn
&& (tp
->ecn_flags
& TCP_ECN_OK
))
266 /* Buffer size and advertised window tuning.
268 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
271 static void tcp_sndbuf_expand(struct sock
*sk
)
273 const struct tcp_sock
*tp
= tcp_sk(sk
);
277 /* Worst case is non GSO/TSO : each frame consumes one skb
278 * and skb->head is kmalloced using power of two area of memory
280 per_mss
= max_t(u32
, tp
->rx_opt
.mss_clamp
, tp
->mss_cache
) +
282 SKB_DATA_ALIGN(sizeof(struct skb_shared_info
));
284 per_mss
= roundup_pow_of_two(per_mss
) +
285 SKB_DATA_ALIGN(sizeof(struct sk_buff
));
287 nr_segs
= max_t(u32
, TCP_INIT_CWND
, tp
->snd_cwnd
);
288 nr_segs
= max_t(u32
, nr_segs
, tp
->reordering
+ 1);
290 /* Fast Recovery (RFC 5681 3.2) :
291 * Cubic needs 1.7 factor, rounded to 2 to include
292 * extra cushion (application might react slowly to POLLOUT)
294 sndmem
= 2 * nr_segs
* per_mss
;
296 if (sk
->sk_sndbuf
< sndmem
)
297 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
300 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
302 * All tcp_full_space() is split to two parts: "network" buffer, allocated
303 * forward and advertised in receiver window (tp->rcv_wnd) and
304 * "application buffer", required to isolate scheduling/application
305 * latencies from network.
306 * window_clamp is maximal advertised window. It can be less than
307 * tcp_full_space(), in this case tcp_full_space() - window_clamp
308 * is reserved for "application" buffer. The less window_clamp is
309 * the smoother our behaviour from viewpoint of network, but the lower
310 * throughput and the higher sensitivity of the connection to losses. 8)
312 * rcv_ssthresh is more strict window_clamp used at "slow start"
313 * phase to predict further behaviour of this connection.
314 * It is used for two goals:
315 * - to enforce header prediction at sender, even when application
316 * requires some significant "application buffer". It is check #1.
317 * - to prevent pruning of receive queue because of misprediction
318 * of receiver window. Check #2.
320 * The scheme does not work when sender sends good segments opening
321 * window and then starts to feed us spaghetti. But it should work
322 * in common situations. Otherwise, we have to rely on queue collapsing.
325 /* Slow part of check#2. */
326 static int __tcp_grow_window(const struct sock
*sk
, const struct sk_buff
*skb
)
328 struct tcp_sock
*tp
= tcp_sk(sk
);
330 int truesize
= tcp_win_from_space(skb
->truesize
) >> 1;
331 int window
= tcp_win_from_space(sysctl_tcp_rmem
[2]) >> 1;
333 while (tp
->rcv_ssthresh
<= window
) {
334 if (truesize
<= skb
->len
)
335 return 2 * inet_csk(sk
)->icsk_ack
.rcv_mss
;
343 static void tcp_grow_window(struct sock
*sk
, const struct sk_buff
*skb
)
345 struct tcp_sock
*tp
= tcp_sk(sk
);
348 if (tp
->rcv_ssthresh
< tp
->window_clamp
&&
349 (int)tp
->rcv_ssthresh
< tcp_space(sk
) &&
350 !sk_under_memory_pressure(sk
)) {
353 /* Check #2. Increase window, if skb with such overhead
354 * will fit to rcvbuf in future.
356 if (tcp_win_from_space(skb
->truesize
) <= skb
->len
)
357 incr
= 2 * tp
->advmss
;
359 incr
= __tcp_grow_window(sk
, skb
);
362 incr
= max_t(int, incr
, 2 * skb
->len
);
363 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
+ incr
,
365 inet_csk(sk
)->icsk_ack
.quick
|= 1;
370 /* 3. Tuning rcvbuf, when connection enters established state. */
371 static void tcp_fixup_rcvbuf(struct sock
*sk
)
373 u32 mss
= tcp_sk(sk
)->advmss
;
376 rcvmem
= 2 * SKB_TRUESIZE(mss
+ MAX_TCP_HEADER
) *
377 tcp_default_init_rwnd(mss
);
379 /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
380 * Allow enough cushion so that sender is not limited by our window
382 if (sysctl_tcp_moderate_rcvbuf
)
385 if (sk
->sk_rcvbuf
< rcvmem
)
386 sk
->sk_rcvbuf
= min(rcvmem
, sysctl_tcp_rmem
[2]);
389 /* 4. Try to fixup all. It is made immediately after connection enters
392 void tcp_init_buffer_space(struct sock
*sk
)
394 struct tcp_sock
*tp
= tcp_sk(sk
);
397 if (!(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
))
398 tcp_fixup_rcvbuf(sk
);
399 if (!(sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
))
400 tcp_sndbuf_expand(sk
);
402 tp
->rcvq_space
.space
= tp
->rcv_wnd
;
403 tp
->rcvq_space
.time
= tcp_time_stamp
;
404 tp
->rcvq_space
.seq
= tp
->copied_seq
;
406 maxwin
= tcp_full_space(sk
);
408 if (tp
->window_clamp
>= maxwin
) {
409 tp
->window_clamp
= maxwin
;
411 if (sysctl_tcp_app_win
&& maxwin
> 4 * tp
->advmss
)
412 tp
->window_clamp
= max(maxwin
-
413 (maxwin
>> sysctl_tcp_app_win
),
417 /* Force reservation of one segment. */
418 if (sysctl_tcp_app_win
&&
419 tp
->window_clamp
> 2 * tp
->advmss
&&
420 tp
->window_clamp
+ tp
->advmss
> maxwin
)
421 tp
->window_clamp
= max(2 * tp
->advmss
, maxwin
- tp
->advmss
);
423 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, tp
->window_clamp
);
424 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
427 /* 5. Recalculate window clamp after socket hit its memory bounds. */
428 static void tcp_clamp_window(struct sock
*sk
)
430 struct tcp_sock
*tp
= tcp_sk(sk
);
431 struct inet_connection_sock
*icsk
= inet_csk(sk
);
433 icsk
->icsk_ack
.quick
= 0;
435 if (sk
->sk_rcvbuf
< sysctl_tcp_rmem
[2] &&
436 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
) &&
437 !sk_under_memory_pressure(sk
) &&
438 sk_memory_allocated(sk
) < sk_prot_mem_limits(sk
, 0)) {
439 sk
->sk_rcvbuf
= min(atomic_read(&sk
->sk_rmem_alloc
),
442 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
)
443 tp
->rcv_ssthresh
= min(tp
->window_clamp
, 2U * tp
->advmss
);
446 /* Initialize RCV_MSS value.
447 * RCV_MSS is an our guess about MSS used by the peer.
448 * We haven't any direct information about the MSS.
449 * It's better to underestimate the RCV_MSS rather than overestimate.
450 * Overestimations make us ACKing less frequently than needed.
451 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
453 void tcp_initialize_rcv_mss(struct sock
*sk
)
455 const struct tcp_sock
*tp
= tcp_sk(sk
);
456 unsigned int hint
= min_t(unsigned int, tp
->advmss
, tp
->mss_cache
);
458 hint
= min(hint
, tp
->rcv_wnd
/ 2);
459 hint
= min(hint
, TCP_MSS_DEFAULT
);
460 hint
= max(hint
, TCP_MIN_MSS
);
462 inet_csk(sk
)->icsk_ack
.rcv_mss
= hint
;
464 EXPORT_SYMBOL(tcp_initialize_rcv_mss
);
466 /* Receiver "autotuning" code.
468 * The algorithm for RTT estimation w/o timestamps is based on
469 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
470 * <http://public.lanl.gov/radiant/pubs.html#DRS>
472 * More detail on this code can be found at
473 * <http://staff.psc.edu/jheffner/>,
474 * though this reference is out of date. A new paper
477 static void tcp_rcv_rtt_update(struct tcp_sock
*tp
, u32 sample
, int win_dep
)
479 u32 new_sample
= tp
->rcv_rtt_est
.rtt
;
485 if (new_sample
!= 0) {
486 /* If we sample in larger samples in the non-timestamp
487 * case, we could grossly overestimate the RTT especially
488 * with chatty applications or bulk transfer apps which
489 * are stalled on filesystem I/O.
491 * Also, since we are only going for a minimum in the
492 * non-timestamp case, we do not smooth things out
493 * else with timestamps disabled convergence takes too
497 m
-= (new_sample
>> 3);
505 /* No previous measure. */
509 if (tp
->rcv_rtt_est
.rtt
!= new_sample
)
510 tp
->rcv_rtt_est
.rtt
= new_sample
;
513 static inline void tcp_rcv_rtt_measure(struct tcp_sock
*tp
)
515 if (tp
->rcv_rtt_est
.time
== 0)
517 if (before(tp
->rcv_nxt
, tp
->rcv_rtt_est
.seq
))
519 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rcv_rtt_est
.time
, 1);
522 tp
->rcv_rtt_est
.seq
= tp
->rcv_nxt
+ tp
->rcv_wnd
;
523 tp
->rcv_rtt_est
.time
= tcp_time_stamp
;
526 static inline void tcp_rcv_rtt_measure_ts(struct sock
*sk
,
527 const struct sk_buff
*skb
)
529 struct tcp_sock
*tp
= tcp_sk(sk
);
530 if (tp
->rx_opt
.rcv_tsecr
&&
531 (TCP_SKB_CB(skb
)->end_seq
-
532 TCP_SKB_CB(skb
)->seq
>= inet_csk(sk
)->icsk_ack
.rcv_mss
))
533 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
, 0);
537 * This function should be called every time data is copied to user space.
538 * It calculates the appropriate TCP receive buffer space.
540 void tcp_rcv_space_adjust(struct sock
*sk
)
542 struct tcp_sock
*tp
= tcp_sk(sk
);
546 time
= tcp_time_stamp
- tp
->rcvq_space
.time
;
547 if (time
< (tp
->rcv_rtt_est
.rtt
>> 3) || tp
->rcv_rtt_est
.rtt
== 0)
550 /* Number of bytes copied to user in last RTT */
551 copied
= tp
->copied_seq
- tp
->rcvq_space
.seq
;
552 if (copied
<= tp
->rcvq_space
.space
)
556 * copied = bytes received in previous RTT, our base window
557 * To cope with packet losses, we need a 2x factor
558 * To cope with slow start, and sender growing its cwin by 100 %
559 * every RTT, we need a 4x factor, because the ACK we are sending
560 * now is for the next RTT, not the current one :
561 * <prev RTT . ><current RTT .. ><next RTT .... >
564 if (sysctl_tcp_moderate_rcvbuf
&&
565 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
)) {
566 int rcvwin
, rcvmem
, rcvbuf
;
568 /* minimal window to cope with packet losses, assuming
569 * steady state. Add some cushion because of small variations.
571 rcvwin
= (copied
<< 1) + 16 * tp
->advmss
;
573 /* If rate increased by 25%,
574 * assume slow start, rcvwin = 3 * copied
575 * If rate increased by 50%,
576 * assume sender can use 2x growth, rcvwin = 4 * copied
579 tp
->rcvq_space
.space
+ (tp
->rcvq_space
.space
>> 2)) {
581 tp
->rcvq_space
.space
+ (tp
->rcvq_space
.space
>> 1))
584 rcvwin
+= (rcvwin
>> 1);
587 rcvmem
= SKB_TRUESIZE(tp
->advmss
+ MAX_TCP_HEADER
);
588 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
591 rcvbuf
= min(rcvwin
/ tp
->advmss
* rcvmem
, sysctl_tcp_rmem
[2]);
592 if (rcvbuf
> sk
->sk_rcvbuf
) {
593 sk
->sk_rcvbuf
= rcvbuf
;
595 /* Make the window clamp follow along. */
596 tp
->window_clamp
= rcvwin
;
599 tp
->rcvq_space
.space
= copied
;
602 tp
->rcvq_space
.seq
= tp
->copied_seq
;
603 tp
->rcvq_space
.time
= tcp_time_stamp
;
606 /* There is something which you must keep in mind when you analyze the
607 * behavior of the tp->ato delayed ack timeout interval. When a
608 * connection starts up, we want to ack as quickly as possible. The
609 * problem is that "good" TCP's do slow start at the beginning of data
610 * transmission. The means that until we send the first few ACK's the
611 * sender will sit on his end and only queue most of his data, because
612 * he can only send snd_cwnd unacked packets at any given time. For
613 * each ACK we send, he increments snd_cwnd and transmits more of his
616 static void tcp_event_data_recv(struct sock
*sk
, struct sk_buff
*skb
)
618 struct tcp_sock
*tp
= tcp_sk(sk
);
619 struct inet_connection_sock
*icsk
= inet_csk(sk
);
622 inet_csk_schedule_ack(sk
);
624 tcp_measure_rcv_mss(sk
, skb
);
626 tcp_rcv_rtt_measure(tp
);
628 now
= tcp_time_stamp
;
630 if (!icsk
->icsk_ack
.ato
) {
631 /* The _first_ data packet received, initialize
632 * delayed ACK engine.
634 tcp_incr_quickack(sk
);
635 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
637 int m
= now
- icsk
->icsk_ack
.lrcvtime
;
639 if (m
<= TCP_ATO_MIN
/ 2) {
640 /* The fastest case is the first. */
641 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + TCP_ATO_MIN
/ 2;
642 } else if (m
< icsk
->icsk_ack
.ato
) {
643 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + m
;
644 if (icsk
->icsk_ack
.ato
> icsk
->icsk_rto
)
645 icsk
->icsk_ack
.ato
= icsk
->icsk_rto
;
646 } else if (m
> icsk
->icsk_rto
) {
647 /* Too long gap. Apparently sender failed to
648 * restart window, so that we send ACKs quickly.
650 tcp_incr_quickack(sk
);
654 icsk
->icsk_ack
.lrcvtime
= now
;
656 TCP_ECN_check_ce(tp
, skb
);
659 tcp_grow_window(sk
, skb
);
662 /* Called to compute a smoothed rtt estimate. The data fed to this
663 * routine either comes from timestamps, or from segments that were
664 * known _not_ to have been retransmitted [see Karn/Partridge
665 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
666 * piece by Van Jacobson.
667 * NOTE: the next three routines used to be one big routine.
668 * To save cycles in the RFC 1323 implementation it was better to break
669 * it up into three procedures. -- erics
671 static void tcp_rtt_estimator(struct sock
*sk
, long mrtt_us
)
673 struct tcp_sock
*tp
= tcp_sk(sk
);
674 long m
= mrtt_us
; /* RTT */
675 u32 srtt
= tp
->srtt_us
;
677 /* The following amusing code comes from Jacobson's
678 * article in SIGCOMM '88. Note that rtt and mdev
679 * are scaled versions of rtt and mean deviation.
680 * This is designed to be as fast as possible
681 * m stands for "measurement".
683 * On a 1990 paper the rto value is changed to:
684 * RTO = rtt + 4 * mdev
686 * Funny. This algorithm seems to be very broken.
687 * These formulae increase RTO, when it should be decreased, increase
688 * too slowly, when it should be increased quickly, decrease too quickly
689 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
690 * does not matter how to _calculate_ it. Seems, it was trap
691 * that VJ failed to avoid. 8)
694 m
-= (srtt
>> 3); /* m is now error in rtt est */
695 srtt
+= m
; /* rtt = 7/8 rtt + 1/8 new */
697 m
= -m
; /* m is now abs(error) */
698 m
-= (tp
->mdev_us
>> 2); /* similar update on mdev */
699 /* This is similar to one of Eifel findings.
700 * Eifel blocks mdev updates when rtt decreases.
701 * This solution is a bit different: we use finer gain
702 * for mdev in this case (alpha*beta).
703 * Like Eifel it also prevents growth of rto,
704 * but also it limits too fast rto decreases,
705 * happening in pure Eifel.
710 m
-= (tp
->mdev_us
>> 2); /* similar update on mdev */
712 tp
->mdev_us
+= m
; /* mdev = 3/4 mdev + 1/4 new */
713 if (tp
->mdev_us
> tp
->mdev_max_us
) {
714 tp
->mdev_max_us
= tp
->mdev_us
;
715 if (tp
->mdev_max_us
> tp
->rttvar_us
)
716 tp
->rttvar_us
= tp
->mdev_max_us
;
718 if (after(tp
->snd_una
, tp
->rtt_seq
)) {
719 if (tp
->mdev_max_us
< tp
->rttvar_us
)
720 tp
->rttvar_us
-= (tp
->rttvar_us
- tp
->mdev_max_us
) >> 2;
721 tp
->rtt_seq
= tp
->snd_nxt
;
722 tp
->mdev_max_us
= tcp_rto_min_us(sk
);
725 /* no previous measure. */
726 srtt
= m
<< 3; /* take the measured time to be rtt */
727 tp
->mdev_us
= m
<< 1; /* make sure rto = 3*rtt */
728 tp
->rttvar_us
= max(tp
->mdev_us
, tcp_rto_min_us(sk
));
729 tp
->mdev_max_us
= tp
->rttvar_us
;
730 tp
->rtt_seq
= tp
->snd_nxt
;
732 tp
->srtt_us
= max(1U, srtt
);
735 /* Set the sk_pacing_rate to allow proper sizing of TSO packets.
736 * Note: TCP stack does not yet implement pacing.
737 * FQ packet scheduler can be used to implement cheap but effective
738 * TCP pacing, to smooth the burst on large writes when packets
739 * in flight is significantly lower than cwnd (or rwin)
741 static void tcp_update_pacing_rate(struct sock
*sk
)
743 const struct tcp_sock
*tp
= tcp_sk(sk
);
746 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
747 rate
= (u64
)tp
->mss_cache
* 2 * (USEC_PER_SEC
<< 3);
749 rate
*= max(tp
->snd_cwnd
, tp
->packets_out
);
751 if (likely(tp
->srtt_us
))
752 do_div(rate
, tp
->srtt_us
);
754 /* ACCESS_ONCE() is needed because sch_fq fetches sk_pacing_rate
755 * without any lock. We want to make sure compiler wont store
756 * intermediate values in this location.
758 ACCESS_ONCE(sk
->sk_pacing_rate
) = min_t(u64
, rate
,
759 sk
->sk_max_pacing_rate
);
762 /* Calculate rto without backoff. This is the second half of Van Jacobson's
763 * routine referred to above.
765 static void tcp_set_rto(struct sock
*sk
)
767 const struct tcp_sock
*tp
= tcp_sk(sk
);
768 /* Old crap is replaced with new one. 8)
771 * 1. If rtt variance happened to be less 50msec, it is hallucination.
772 * It cannot be less due to utterly erratic ACK generation made
773 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
774 * to do with delayed acks, because at cwnd>2 true delack timeout
775 * is invisible. Actually, Linux-2.4 also generates erratic
776 * ACKs in some circumstances.
778 inet_csk(sk
)->icsk_rto
= __tcp_set_rto(tp
);
780 /* 2. Fixups made earlier cannot be right.
781 * If we do not estimate RTO correctly without them,
782 * all the algo is pure shit and should be replaced
783 * with correct one. It is exactly, which we pretend to do.
786 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
787 * guarantees that rto is higher.
792 __u32
tcp_init_cwnd(const struct tcp_sock
*tp
, const struct dst_entry
*dst
)
794 __u32 cwnd
= (dst
? dst_metric(dst
, RTAX_INITCWND
) : 0);
797 cwnd
= TCP_INIT_CWND
;
798 return min_t(__u32
, cwnd
, tp
->snd_cwnd_clamp
);
802 * Packet counting of FACK is based on in-order assumptions, therefore TCP
803 * disables it when reordering is detected
805 void tcp_disable_fack(struct tcp_sock
*tp
)
807 /* RFC3517 uses different metric in lost marker => reset on change */
809 tp
->lost_skb_hint
= NULL
;
810 tp
->rx_opt
.sack_ok
&= ~TCP_FACK_ENABLED
;
813 /* Take a notice that peer is sending D-SACKs */
814 static void tcp_dsack_seen(struct tcp_sock
*tp
)
816 tp
->rx_opt
.sack_ok
|= TCP_DSACK_SEEN
;
819 static void tcp_update_reordering(struct sock
*sk
, const int metric
,
822 struct tcp_sock
*tp
= tcp_sk(sk
);
823 if (metric
> tp
->reordering
) {
826 tp
->reordering
= min(TCP_MAX_REORDERING
, metric
);
828 /* This exciting event is worth to be remembered. 8) */
830 mib_idx
= LINUX_MIB_TCPTSREORDER
;
831 else if (tcp_is_reno(tp
))
832 mib_idx
= LINUX_MIB_TCPRENOREORDER
;
833 else if (tcp_is_fack(tp
))
834 mib_idx
= LINUX_MIB_TCPFACKREORDER
;
836 mib_idx
= LINUX_MIB_TCPSACKREORDER
;
838 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
839 #if FASTRETRANS_DEBUG > 1
840 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
841 tp
->rx_opt
.sack_ok
, inet_csk(sk
)->icsk_ca_state
,
845 tp
->undo_marker
? tp
->undo_retrans
: 0);
847 tcp_disable_fack(tp
);
851 tcp_disable_early_retrans(tp
);
854 /* This must be called before lost_out is incremented */
855 static void tcp_verify_retransmit_hint(struct tcp_sock
*tp
, struct sk_buff
*skb
)
857 if ((tp
->retransmit_skb_hint
== NULL
) ||
858 before(TCP_SKB_CB(skb
)->seq
,
859 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
860 tp
->retransmit_skb_hint
= skb
;
863 after(TCP_SKB_CB(skb
)->end_seq
, tp
->retransmit_high
))
864 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
867 static void tcp_skb_mark_lost(struct tcp_sock
*tp
, struct sk_buff
*skb
)
869 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
870 tcp_verify_retransmit_hint(tp
, skb
);
872 tp
->lost_out
+= tcp_skb_pcount(skb
);
873 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
877 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock
*tp
,
880 tcp_verify_retransmit_hint(tp
, skb
);
882 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
883 tp
->lost_out
+= tcp_skb_pcount(skb
);
884 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
888 /* This procedure tags the retransmission queue when SACKs arrive.
890 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
891 * Packets in queue with these bits set are counted in variables
892 * sacked_out, retrans_out and lost_out, correspondingly.
894 * Valid combinations are:
895 * Tag InFlight Description
896 * 0 1 - orig segment is in flight.
897 * S 0 - nothing flies, orig reached receiver.
898 * L 0 - nothing flies, orig lost by net.
899 * R 2 - both orig and retransmit are in flight.
900 * L|R 1 - orig is lost, retransmit is in flight.
901 * S|R 1 - orig reached receiver, retrans is still in flight.
902 * (L|S|R is logically valid, it could occur when L|R is sacked,
903 * but it is equivalent to plain S and code short-curcuits it to S.
904 * L|S is logically invalid, it would mean -1 packet in flight 8))
906 * These 6 states form finite state machine, controlled by the following events:
907 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
908 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
909 * 3. Loss detection event of two flavors:
910 * A. Scoreboard estimator decided the packet is lost.
911 * A'. Reno "three dupacks" marks head of queue lost.
912 * A''. Its FACK modification, head until snd.fack is lost.
913 * B. SACK arrives sacking SND.NXT at the moment, when the
914 * segment was retransmitted.
915 * 4. D-SACK added new rule: D-SACK changes any tag to S.
917 * It is pleasant to note, that state diagram turns out to be commutative,
918 * so that we are allowed not to be bothered by order of our actions,
919 * when multiple events arrive simultaneously. (see the function below).
921 * Reordering detection.
922 * --------------------
923 * Reordering metric is maximal distance, which a packet can be displaced
924 * in packet stream. With SACKs we can estimate it:
926 * 1. SACK fills old hole and the corresponding segment was not
927 * ever retransmitted -> reordering. Alas, we cannot use it
928 * when segment was retransmitted.
929 * 2. The last flaw is solved with D-SACK. D-SACK arrives
930 * for retransmitted and already SACKed segment -> reordering..
931 * Both of these heuristics are not used in Loss state, when we cannot
932 * account for retransmits accurately.
934 * SACK block validation.
935 * ----------------------
937 * SACK block range validation checks that the received SACK block fits to
938 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
939 * Note that SND.UNA is not included to the range though being valid because
940 * it means that the receiver is rather inconsistent with itself reporting
941 * SACK reneging when it should advance SND.UNA. Such SACK block this is
942 * perfectly valid, however, in light of RFC2018 which explicitly states
943 * that "SACK block MUST reflect the newest segment. Even if the newest
944 * segment is going to be discarded ...", not that it looks very clever
945 * in case of head skb. Due to potentional receiver driven attacks, we
946 * choose to avoid immediate execution of a walk in write queue due to
947 * reneging and defer head skb's loss recovery to standard loss recovery
948 * procedure that will eventually trigger (nothing forbids us doing this).
950 * Implements also blockage to start_seq wrap-around. Problem lies in the
951 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
952 * there's no guarantee that it will be before snd_nxt (n). The problem
953 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
956 * <- outs wnd -> <- wrapzone ->
957 * u e n u_w e_w s n_w
959 * |<------------+------+----- TCP seqno space --------------+---------->|
960 * ...-- <2^31 ->| |<--------...
961 * ...---- >2^31 ------>| |<--------...
963 * Current code wouldn't be vulnerable but it's better still to discard such
964 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
965 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
966 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
967 * equal to the ideal case (infinite seqno space without wrap caused issues).
969 * With D-SACK the lower bound is extended to cover sequence space below
970 * SND.UNA down to undo_marker, which is the last point of interest. Yet
971 * again, D-SACK block must not to go across snd_una (for the same reason as
972 * for the normal SACK blocks, explained above). But there all simplicity
973 * ends, TCP might receive valid D-SACKs below that. As long as they reside
974 * fully below undo_marker they do not affect behavior in anyway and can
975 * therefore be safely ignored. In rare cases (which are more or less
976 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
977 * fragmentation and packet reordering past skb's retransmission. To consider
978 * them correctly, the acceptable range must be extended even more though
979 * the exact amount is rather hard to quantify. However, tp->max_window can
980 * be used as an exaggerated estimate.
982 static bool tcp_is_sackblock_valid(struct tcp_sock
*tp
, bool is_dsack
,
983 u32 start_seq
, u32 end_seq
)
985 /* Too far in future, or reversed (interpretation is ambiguous) */
986 if (after(end_seq
, tp
->snd_nxt
) || !before(start_seq
, end_seq
))
989 /* Nasty start_seq wrap-around check (see comments above) */
990 if (!before(start_seq
, tp
->snd_nxt
))
993 /* In outstanding window? ...This is valid exit for D-SACKs too.
994 * start_seq == snd_una is non-sensical (see comments above)
996 if (after(start_seq
, tp
->snd_una
))
999 if (!is_dsack
|| !tp
->undo_marker
)
1002 /* ...Then it's D-SACK, and must reside below snd_una completely */
1003 if (after(end_seq
, tp
->snd_una
))
1006 if (!before(start_seq
, tp
->undo_marker
))
1010 if (!after(end_seq
, tp
->undo_marker
))
1013 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1014 * start_seq < undo_marker and end_seq >= undo_marker.
1016 return !before(start_seq
, end_seq
- tp
->max_window
);
1019 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1020 * Event "B". Later note: FACK people cheated me again 8), we have to account
1021 * for reordering! Ugly, but should help.
1023 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1024 * less than what is now known to be received by the other end (derived from
1025 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1026 * retransmitted skbs to avoid some costly processing per ACKs.
1028 static void tcp_mark_lost_retrans(struct sock
*sk
)
1030 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1031 struct tcp_sock
*tp
= tcp_sk(sk
);
1032 struct sk_buff
*skb
;
1034 u32 new_low_seq
= tp
->snd_nxt
;
1035 u32 received_upto
= tcp_highest_sack_seq(tp
);
1037 if (!tcp_is_fack(tp
) || !tp
->retrans_out
||
1038 !after(received_upto
, tp
->lost_retrans_low
) ||
1039 icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
1042 tcp_for_write_queue(skb
, sk
) {
1043 u32 ack_seq
= TCP_SKB_CB(skb
)->ack_seq
;
1045 if (skb
== tcp_send_head(sk
))
1047 if (cnt
== tp
->retrans_out
)
1049 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1052 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
))
1055 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1056 * constraint here (see above) but figuring out that at
1057 * least tp->reordering SACK blocks reside between ack_seq
1058 * and received_upto is not easy task to do cheaply with
1059 * the available datastructures.
1061 * Whether FACK should check here for tp->reordering segs
1062 * in-between one could argue for either way (it would be
1063 * rather simple to implement as we could count fack_count
1064 * during the walk and do tp->fackets_out - fack_count).
1066 if (after(received_upto
, ack_seq
)) {
1067 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1068 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1070 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
1071 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSTRETRANSMIT
);
1073 if (before(ack_seq
, new_low_seq
))
1074 new_low_seq
= ack_seq
;
1075 cnt
+= tcp_skb_pcount(skb
);
1079 if (tp
->retrans_out
)
1080 tp
->lost_retrans_low
= new_low_seq
;
1083 static bool tcp_check_dsack(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1084 struct tcp_sack_block_wire
*sp
, int num_sacks
,
1087 struct tcp_sock
*tp
= tcp_sk(sk
);
1088 u32 start_seq_0
= get_unaligned_be32(&sp
[0].start_seq
);
1089 u32 end_seq_0
= get_unaligned_be32(&sp
[0].end_seq
);
1090 bool dup_sack
= false;
1092 if (before(start_seq_0
, TCP_SKB_CB(ack_skb
)->ack_seq
)) {
1095 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPDSACKRECV
);
1096 } else if (num_sacks
> 1) {
1097 u32 end_seq_1
= get_unaligned_be32(&sp
[1].end_seq
);
1098 u32 start_seq_1
= get_unaligned_be32(&sp
[1].start_seq
);
1100 if (!after(end_seq_0
, end_seq_1
) &&
1101 !before(start_seq_0
, start_seq_1
)) {
1104 NET_INC_STATS_BH(sock_net(sk
),
1105 LINUX_MIB_TCPDSACKOFORECV
);
1109 /* D-SACK for already forgotten data... Do dumb counting. */
1110 if (dup_sack
&& tp
->undo_marker
&& tp
->undo_retrans
> 0 &&
1111 !after(end_seq_0
, prior_snd_una
) &&
1112 after(end_seq_0
, tp
->undo_marker
))
1118 struct tcp_sacktag_state
{
1121 long rtt_us
; /* RTT measured by SACKing never-retransmitted data */
1125 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1126 * the incoming SACK may not exactly match but we can find smaller MSS
1127 * aligned portion of it that matches. Therefore we might need to fragment
1128 * which may fail and creates some hassle (caller must handle error case
1131 * FIXME: this could be merged to shift decision code
1133 static int tcp_match_skb_to_sack(struct sock
*sk
, struct sk_buff
*skb
,
1134 u32 start_seq
, u32 end_seq
)
1138 unsigned int pkt_len
;
1141 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1142 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1144 if (tcp_skb_pcount(skb
) > 1 && !in_sack
&&
1145 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
1146 mss
= tcp_skb_mss(skb
);
1147 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1150 pkt_len
= start_seq
- TCP_SKB_CB(skb
)->seq
;
1154 pkt_len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1159 /* Round if necessary so that SACKs cover only full MSSes
1160 * and/or the remaining small portion (if present)
1162 if (pkt_len
> mss
) {
1163 unsigned int new_len
= (pkt_len
/ mss
) * mss
;
1164 if (!in_sack
&& new_len
< pkt_len
) {
1166 if (new_len
>= skb
->len
)
1171 err
= tcp_fragment(sk
, skb
, pkt_len
, mss
, GFP_ATOMIC
);
1179 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1180 static u8
tcp_sacktag_one(struct sock
*sk
,
1181 struct tcp_sacktag_state
*state
, u8 sacked
,
1182 u32 start_seq
, u32 end_seq
,
1183 int dup_sack
, int pcount
,
1184 const struct skb_mstamp
*xmit_time
)
1186 struct tcp_sock
*tp
= tcp_sk(sk
);
1187 int fack_count
= state
->fack_count
;
1189 /* Account D-SACK for retransmitted packet. */
1190 if (dup_sack
&& (sacked
& TCPCB_RETRANS
)) {
1191 if (tp
->undo_marker
&& tp
->undo_retrans
> 0 &&
1192 after(end_seq
, tp
->undo_marker
))
1194 if (sacked
& TCPCB_SACKED_ACKED
)
1195 state
->reord
= min(fack_count
, state
->reord
);
1198 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1199 if (!after(end_seq
, tp
->snd_una
))
1202 if (!(sacked
& TCPCB_SACKED_ACKED
)) {
1203 if (sacked
& TCPCB_SACKED_RETRANS
) {
1204 /* If the segment is not tagged as lost,
1205 * we do not clear RETRANS, believing
1206 * that retransmission is still in flight.
1208 if (sacked
& TCPCB_LOST
) {
1209 sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1210 tp
->lost_out
-= pcount
;
1211 tp
->retrans_out
-= pcount
;
1214 if (!(sacked
& TCPCB_RETRANS
)) {
1215 /* New sack for not retransmitted frame,
1216 * which was in hole. It is reordering.
1218 if (before(start_seq
,
1219 tcp_highest_sack_seq(tp
)))
1220 state
->reord
= min(fack_count
,
1222 if (!after(end_seq
, tp
->high_seq
))
1223 state
->flag
|= FLAG_ORIG_SACK_ACKED
;
1224 /* Pick the earliest sequence sacked for RTT */
1225 if (state
->rtt_us
< 0) {
1226 struct skb_mstamp now
;
1228 skb_mstamp_get(&now
);
1229 state
->rtt_us
= skb_mstamp_us_delta(&now
,
1234 if (sacked
& TCPCB_LOST
) {
1235 sacked
&= ~TCPCB_LOST
;
1236 tp
->lost_out
-= pcount
;
1240 sacked
|= TCPCB_SACKED_ACKED
;
1241 state
->flag
|= FLAG_DATA_SACKED
;
1242 tp
->sacked_out
+= pcount
;
1244 fack_count
+= pcount
;
1246 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1247 if (!tcp_is_fack(tp
) && (tp
->lost_skb_hint
!= NULL
) &&
1248 before(start_seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
))
1249 tp
->lost_cnt_hint
+= pcount
;
1251 if (fack_count
> tp
->fackets_out
)
1252 tp
->fackets_out
= fack_count
;
1255 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1256 * frames and clear it. undo_retrans is decreased above, L|R frames
1257 * are accounted above as well.
1259 if (dup_sack
&& (sacked
& TCPCB_SACKED_RETRANS
)) {
1260 sacked
&= ~TCPCB_SACKED_RETRANS
;
1261 tp
->retrans_out
-= pcount
;
1267 /* Shift newly-SACKed bytes from this skb to the immediately previous
1268 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1270 static bool tcp_shifted_skb(struct sock
*sk
, struct sk_buff
*skb
,
1271 struct tcp_sacktag_state
*state
,
1272 unsigned int pcount
, int shifted
, int mss
,
1275 struct tcp_sock
*tp
= tcp_sk(sk
);
1276 struct sk_buff
*prev
= tcp_write_queue_prev(sk
, skb
);
1277 u32 start_seq
= TCP_SKB_CB(skb
)->seq
; /* start of newly-SACKed */
1278 u32 end_seq
= start_seq
+ shifted
; /* end of newly-SACKed */
1282 /* Adjust counters and hints for the newly sacked sequence
1283 * range but discard the return value since prev is already
1284 * marked. We must tag the range first because the seq
1285 * advancement below implicitly advances
1286 * tcp_highest_sack_seq() when skb is highest_sack.
1288 tcp_sacktag_one(sk
, state
, TCP_SKB_CB(skb
)->sacked
,
1289 start_seq
, end_seq
, dup_sack
, pcount
,
1292 if (skb
== tp
->lost_skb_hint
)
1293 tp
->lost_cnt_hint
+= pcount
;
1295 TCP_SKB_CB(prev
)->end_seq
+= shifted
;
1296 TCP_SKB_CB(skb
)->seq
+= shifted
;
1298 skb_shinfo(prev
)->gso_segs
+= pcount
;
1299 BUG_ON(skb_shinfo(skb
)->gso_segs
< pcount
);
1300 skb_shinfo(skb
)->gso_segs
-= pcount
;
1302 /* When we're adding to gso_segs == 1, gso_size will be zero,
1303 * in theory this shouldn't be necessary but as long as DSACK
1304 * code can come after this skb later on it's better to keep
1305 * setting gso_size to something.
1307 if (!skb_shinfo(prev
)->gso_size
) {
1308 skb_shinfo(prev
)->gso_size
= mss
;
1309 skb_shinfo(prev
)->gso_type
= sk
->sk_gso_type
;
1312 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1313 if (skb_shinfo(skb
)->gso_segs
<= 1) {
1314 skb_shinfo(skb
)->gso_size
= 0;
1315 skb_shinfo(skb
)->gso_type
= 0;
1318 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1319 TCP_SKB_CB(prev
)->sacked
|= (TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
);
1322 BUG_ON(!tcp_skb_pcount(skb
));
1323 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKSHIFTED
);
1327 /* Whole SKB was eaten :-) */
1329 if (skb
== tp
->retransmit_skb_hint
)
1330 tp
->retransmit_skb_hint
= prev
;
1331 if (skb
== tp
->lost_skb_hint
) {
1332 tp
->lost_skb_hint
= prev
;
1333 tp
->lost_cnt_hint
-= tcp_skb_pcount(prev
);
1336 TCP_SKB_CB(prev
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
;
1337 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
1338 TCP_SKB_CB(prev
)->end_seq
++;
1340 if (skb
== tcp_highest_sack(sk
))
1341 tcp_advance_highest_sack(sk
, skb
);
1343 tcp_unlink_write_queue(skb
, sk
);
1344 sk_wmem_free_skb(sk
, skb
);
1346 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKMERGED
);
1351 /* I wish gso_size would have a bit more sane initialization than
1352 * something-or-zero which complicates things
1354 static int tcp_skb_seglen(const struct sk_buff
*skb
)
1356 return tcp_skb_pcount(skb
) == 1 ? skb
->len
: tcp_skb_mss(skb
);
1359 /* Shifting pages past head area doesn't work */
1360 static int skb_can_shift(const struct sk_buff
*skb
)
1362 return !skb_headlen(skb
) && skb_is_nonlinear(skb
);
1365 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1368 static struct sk_buff
*tcp_shift_skb_data(struct sock
*sk
, struct sk_buff
*skb
,
1369 struct tcp_sacktag_state
*state
,
1370 u32 start_seq
, u32 end_seq
,
1373 struct tcp_sock
*tp
= tcp_sk(sk
);
1374 struct sk_buff
*prev
;
1380 if (!sk_can_gso(sk
))
1383 /* Normally R but no L won't result in plain S */
1385 (TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_RETRANS
)) == TCPCB_SACKED_RETRANS
)
1387 if (!skb_can_shift(skb
))
1389 /* This frame is about to be dropped (was ACKed). */
1390 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1393 /* Can only happen with delayed DSACK + discard craziness */
1394 if (unlikely(skb
== tcp_write_queue_head(sk
)))
1396 prev
= tcp_write_queue_prev(sk
, skb
);
1398 if ((TCP_SKB_CB(prev
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
)
1401 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1402 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1406 pcount
= tcp_skb_pcount(skb
);
1407 mss
= tcp_skb_seglen(skb
);
1409 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1410 * drop this restriction as unnecessary
1412 if (mss
!= tcp_skb_seglen(prev
))
1415 if (!after(TCP_SKB_CB(skb
)->end_seq
, start_seq
))
1417 /* CHECKME: This is non-MSS split case only?, this will
1418 * cause skipped skbs due to advancing loop btw, original
1419 * has that feature too
1421 if (tcp_skb_pcount(skb
) <= 1)
1424 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1426 /* TODO: head merge to next could be attempted here
1427 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1428 * though it might not be worth of the additional hassle
1430 * ...we can probably just fallback to what was done
1431 * previously. We could try merging non-SACKed ones
1432 * as well but it probably isn't going to buy off
1433 * because later SACKs might again split them, and
1434 * it would make skb timestamp tracking considerably
1440 len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1442 BUG_ON(len
> skb
->len
);
1444 /* MSS boundaries should be honoured or else pcount will
1445 * severely break even though it makes things bit trickier.
1446 * Optimize common case to avoid most of the divides
1448 mss
= tcp_skb_mss(skb
);
1450 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1451 * drop this restriction as unnecessary
1453 if (mss
!= tcp_skb_seglen(prev
))
1458 } else if (len
< mss
) {
1466 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1467 if (!after(TCP_SKB_CB(skb
)->seq
+ len
, tp
->snd_una
))
1470 if (!skb_shift(prev
, skb
, len
))
1472 if (!tcp_shifted_skb(sk
, skb
, state
, pcount
, len
, mss
, dup_sack
))
1475 /* Hole filled allows collapsing with the next as well, this is very
1476 * useful when hole on every nth skb pattern happens
1478 if (prev
== tcp_write_queue_tail(sk
))
1480 skb
= tcp_write_queue_next(sk
, prev
);
1482 if (!skb_can_shift(skb
) ||
1483 (skb
== tcp_send_head(sk
)) ||
1484 ((TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
) ||
1485 (mss
!= tcp_skb_seglen(skb
)))
1489 if (skb_shift(prev
, skb
, len
)) {
1490 pcount
+= tcp_skb_pcount(skb
);
1491 tcp_shifted_skb(sk
, skb
, state
, tcp_skb_pcount(skb
), len
, mss
, 0);
1495 state
->fack_count
+= pcount
;
1502 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKSHIFTFALLBACK
);
1506 static struct sk_buff
*tcp_sacktag_walk(struct sk_buff
*skb
, struct sock
*sk
,
1507 struct tcp_sack_block
*next_dup
,
1508 struct tcp_sacktag_state
*state
,
1509 u32 start_seq
, u32 end_seq
,
1512 struct tcp_sock
*tp
= tcp_sk(sk
);
1513 struct sk_buff
*tmp
;
1515 tcp_for_write_queue_from(skb
, sk
) {
1517 bool dup_sack
= dup_sack_in
;
1519 if (skb
== tcp_send_head(sk
))
1522 /* queue is in-order => we can short-circuit the walk early */
1523 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1526 if ((next_dup
!= NULL
) &&
1527 before(TCP_SKB_CB(skb
)->seq
, next_dup
->end_seq
)) {
1528 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1529 next_dup
->start_seq
,
1535 /* skb reference here is a bit tricky to get right, since
1536 * shifting can eat and free both this skb and the next,
1537 * so not even _safe variant of the loop is enough.
1540 tmp
= tcp_shift_skb_data(sk
, skb
, state
,
1541 start_seq
, end_seq
, dup_sack
);
1550 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1556 if (unlikely(in_sack
< 0))
1560 TCP_SKB_CB(skb
)->sacked
=
1563 TCP_SKB_CB(skb
)->sacked
,
1564 TCP_SKB_CB(skb
)->seq
,
1565 TCP_SKB_CB(skb
)->end_seq
,
1567 tcp_skb_pcount(skb
),
1570 if (!before(TCP_SKB_CB(skb
)->seq
,
1571 tcp_highest_sack_seq(tp
)))
1572 tcp_advance_highest_sack(sk
, skb
);
1575 state
->fack_count
+= tcp_skb_pcount(skb
);
1580 /* Avoid all extra work that is being done by sacktag while walking in
1583 static struct sk_buff
*tcp_sacktag_skip(struct sk_buff
*skb
, struct sock
*sk
,
1584 struct tcp_sacktag_state
*state
,
1587 tcp_for_write_queue_from(skb
, sk
) {
1588 if (skb
== tcp_send_head(sk
))
1591 if (after(TCP_SKB_CB(skb
)->end_seq
, skip_to_seq
))
1594 state
->fack_count
+= tcp_skb_pcount(skb
);
1599 static struct sk_buff
*tcp_maybe_skipping_dsack(struct sk_buff
*skb
,
1601 struct tcp_sack_block
*next_dup
,
1602 struct tcp_sacktag_state
*state
,
1605 if (next_dup
== NULL
)
1608 if (before(next_dup
->start_seq
, skip_to_seq
)) {
1609 skb
= tcp_sacktag_skip(skb
, sk
, state
, next_dup
->start_seq
);
1610 skb
= tcp_sacktag_walk(skb
, sk
, NULL
, state
,
1611 next_dup
->start_seq
, next_dup
->end_seq
,
1618 static int tcp_sack_cache_ok(const struct tcp_sock
*tp
, const struct tcp_sack_block
*cache
)
1620 return cache
< tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1624 tcp_sacktag_write_queue(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1625 u32 prior_snd_una
, long *sack_rtt_us
)
1627 struct tcp_sock
*tp
= tcp_sk(sk
);
1628 const unsigned char *ptr
= (skb_transport_header(ack_skb
) +
1629 TCP_SKB_CB(ack_skb
)->sacked
);
1630 struct tcp_sack_block_wire
*sp_wire
= (struct tcp_sack_block_wire
*)(ptr
+2);
1631 struct tcp_sack_block sp
[TCP_NUM_SACKS
];
1632 struct tcp_sack_block
*cache
;
1633 struct tcp_sacktag_state state
;
1634 struct sk_buff
*skb
;
1635 int num_sacks
= min(TCP_NUM_SACKS
, (ptr
[1] - TCPOLEN_SACK_BASE
) >> 3);
1637 bool found_dup_sack
= false;
1639 int first_sack_index
;
1642 state
.reord
= tp
->packets_out
;
1645 if (!tp
->sacked_out
) {
1646 if (WARN_ON(tp
->fackets_out
))
1647 tp
->fackets_out
= 0;
1648 tcp_highest_sack_reset(sk
);
1651 found_dup_sack
= tcp_check_dsack(sk
, ack_skb
, sp_wire
,
1652 num_sacks
, prior_snd_una
);
1654 state
.flag
|= FLAG_DSACKING_ACK
;
1656 /* Eliminate too old ACKs, but take into
1657 * account more or less fresh ones, they can
1658 * contain valid SACK info.
1660 if (before(TCP_SKB_CB(ack_skb
)->ack_seq
, prior_snd_una
- tp
->max_window
))
1663 if (!tp
->packets_out
)
1667 first_sack_index
= 0;
1668 for (i
= 0; i
< num_sacks
; i
++) {
1669 bool dup_sack
= !i
&& found_dup_sack
;
1671 sp
[used_sacks
].start_seq
= get_unaligned_be32(&sp_wire
[i
].start_seq
);
1672 sp
[used_sacks
].end_seq
= get_unaligned_be32(&sp_wire
[i
].end_seq
);
1674 if (!tcp_is_sackblock_valid(tp
, dup_sack
,
1675 sp
[used_sacks
].start_seq
,
1676 sp
[used_sacks
].end_seq
)) {
1680 if (!tp
->undo_marker
)
1681 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDNOUNDO
;
1683 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDOLD
;
1685 /* Don't count olds caused by ACK reordering */
1686 if ((TCP_SKB_CB(ack_skb
)->ack_seq
!= tp
->snd_una
) &&
1687 !after(sp
[used_sacks
].end_seq
, tp
->snd_una
))
1689 mib_idx
= LINUX_MIB_TCPSACKDISCARD
;
1692 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
1694 first_sack_index
= -1;
1698 /* Ignore very old stuff early */
1699 if (!after(sp
[used_sacks
].end_seq
, prior_snd_una
))
1705 /* order SACK blocks to allow in order walk of the retrans queue */
1706 for (i
= used_sacks
- 1; i
> 0; i
--) {
1707 for (j
= 0; j
< i
; j
++) {
1708 if (after(sp
[j
].start_seq
, sp
[j
+ 1].start_seq
)) {
1709 swap(sp
[j
], sp
[j
+ 1]);
1711 /* Track where the first SACK block goes to */
1712 if (j
== first_sack_index
)
1713 first_sack_index
= j
+ 1;
1718 skb
= tcp_write_queue_head(sk
);
1719 state
.fack_count
= 0;
1722 if (!tp
->sacked_out
) {
1723 /* It's already past, so skip checking against it */
1724 cache
= tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1726 cache
= tp
->recv_sack_cache
;
1727 /* Skip empty blocks in at head of the cache */
1728 while (tcp_sack_cache_ok(tp
, cache
) && !cache
->start_seq
&&
1733 while (i
< used_sacks
) {
1734 u32 start_seq
= sp
[i
].start_seq
;
1735 u32 end_seq
= sp
[i
].end_seq
;
1736 bool dup_sack
= (found_dup_sack
&& (i
== first_sack_index
));
1737 struct tcp_sack_block
*next_dup
= NULL
;
1739 if (found_dup_sack
&& ((i
+ 1) == first_sack_index
))
1740 next_dup
= &sp
[i
+ 1];
1742 /* Skip too early cached blocks */
1743 while (tcp_sack_cache_ok(tp
, cache
) &&
1744 !before(start_seq
, cache
->end_seq
))
1747 /* Can skip some work by looking recv_sack_cache? */
1748 if (tcp_sack_cache_ok(tp
, cache
) && !dup_sack
&&
1749 after(end_seq
, cache
->start_seq
)) {
1752 if (before(start_seq
, cache
->start_seq
)) {
1753 skb
= tcp_sacktag_skip(skb
, sk
, &state
,
1755 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
,
1762 /* Rest of the block already fully processed? */
1763 if (!after(end_seq
, cache
->end_seq
))
1766 skb
= tcp_maybe_skipping_dsack(skb
, sk
, next_dup
,
1770 /* ...tail remains todo... */
1771 if (tcp_highest_sack_seq(tp
) == cache
->end_seq
) {
1772 /* ...but better entrypoint exists! */
1773 skb
= tcp_highest_sack(sk
);
1776 state
.fack_count
= tp
->fackets_out
;
1781 skb
= tcp_sacktag_skip(skb
, sk
, &state
, cache
->end_seq
);
1782 /* Check overlap against next cached too (past this one already) */
1787 if (!before(start_seq
, tcp_highest_sack_seq(tp
))) {
1788 skb
= tcp_highest_sack(sk
);
1791 state
.fack_count
= tp
->fackets_out
;
1793 skb
= tcp_sacktag_skip(skb
, sk
, &state
, start_seq
);
1796 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
, &state
,
1797 start_seq
, end_seq
, dup_sack
);
1803 /* Clear the head of the cache sack blocks so we can skip it next time */
1804 for (i
= 0; i
< ARRAY_SIZE(tp
->recv_sack_cache
) - used_sacks
; i
++) {
1805 tp
->recv_sack_cache
[i
].start_seq
= 0;
1806 tp
->recv_sack_cache
[i
].end_seq
= 0;
1808 for (j
= 0; j
< used_sacks
; j
++)
1809 tp
->recv_sack_cache
[i
++] = sp
[j
];
1811 tcp_mark_lost_retrans(sk
);
1813 tcp_verify_left_out(tp
);
1815 if ((state
.reord
< tp
->fackets_out
) &&
1816 ((inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Loss
) || tp
->undo_marker
))
1817 tcp_update_reordering(sk
, tp
->fackets_out
- state
.reord
, 0);
1821 #if FASTRETRANS_DEBUG > 0
1822 WARN_ON((int)tp
->sacked_out
< 0);
1823 WARN_ON((int)tp
->lost_out
< 0);
1824 WARN_ON((int)tp
->retrans_out
< 0);
1825 WARN_ON((int)tcp_packets_in_flight(tp
) < 0);
1827 *sack_rtt_us
= state
.rtt_us
;
1831 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1832 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1834 static bool tcp_limit_reno_sacked(struct tcp_sock
*tp
)
1838 holes
= max(tp
->lost_out
, 1U);
1839 holes
= min(holes
, tp
->packets_out
);
1841 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1842 tp
->sacked_out
= tp
->packets_out
- holes
;
1848 /* If we receive more dupacks than we expected counting segments
1849 * in assumption of absent reordering, interpret this as reordering.
1850 * The only another reason could be bug in receiver TCP.
1852 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1854 struct tcp_sock
*tp
= tcp_sk(sk
);
1855 if (tcp_limit_reno_sacked(tp
))
1856 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1859 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1861 static void tcp_add_reno_sack(struct sock
*sk
)
1863 struct tcp_sock
*tp
= tcp_sk(sk
);
1865 tcp_check_reno_reordering(sk
, 0);
1866 tcp_verify_left_out(tp
);
1869 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1871 static void tcp_remove_reno_sacks(struct sock
*sk
, int acked
)
1873 struct tcp_sock
*tp
= tcp_sk(sk
);
1876 /* One ACK acked hole. The rest eat duplicate ACKs. */
1877 if (acked
- 1 >= tp
->sacked_out
)
1880 tp
->sacked_out
-= acked
- 1;
1882 tcp_check_reno_reordering(sk
, acked
);
1883 tcp_verify_left_out(tp
);
1886 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1891 static void tcp_clear_retrans_partial(struct tcp_sock
*tp
)
1893 tp
->retrans_out
= 0;
1896 tp
->undo_marker
= 0;
1897 tp
->undo_retrans
= -1;
1900 void tcp_clear_retrans(struct tcp_sock
*tp
)
1902 tcp_clear_retrans_partial(tp
);
1904 tp
->fackets_out
= 0;
1908 /* Enter Loss state. If we detect SACK reneging, forget all SACK information
1909 * and reset tags completely, otherwise preserve SACKs. If receiver
1910 * dropped its ofo queue, we will know this due to reneging detection.
1912 void tcp_enter_loss(struct sock
*sk
)
1914 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1915 struct tcp_sock
*tp
= tcp_sk(sk
);
1916 struct sk_buff
*skb
;
1917 bool new_recovery
= false;
1918 bool is_reneg
; /* is receiver reneging on SACKs? */
1920 /* Reduce ssthresh if it has not yet been made inside this window. */
1921 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
||
1922 !after(tp
->high_seq
, tp
->snd_una
) ||
1923 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
1924 new_recovery
= true;
1925 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1926 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1927 tcp_ca_event(sk
, CA_EVENT_LOSS
);
1930 tp
->snd_cwnd_cnt
= 0;
1931 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1933 tcp_clear_retrans_partial(tp
);
1935 if (tcp_is_reno(tp
))
1936 tcp_reset_reno_sack(tp
);
1938 tp
->undo_marker
= tp
->snd_una
;
1940 skb
= tcp_write_queue_head(sk
);
1941 is_reneg
= skb
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
);
1943 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSACKRENEGING
);
1945 tp
->fackets_out
= 0;
1947 tcp_clear_all_retrans_hints(tp
);
1949 tcp_for_write_queue(skb
, sk
) {
1950 if (skb
== tcp_send_head(sk
))
1953 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
1954 tp
->undo_marker
= 0;
1956 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
1957 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) || is_reneg
) {
1958 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
1959 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1960 tp
->lost_out
+= tcp_skb_pcount(skb
);
1961 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
1964 tcp_verify_left_out(tp
);
1966 /* Timeout in disordered state after receiving substantial DUPACKs
1967 * suggests that the degree of reordering is over-estimated.
1969 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
&&
1970 tp
->sacked_out
>= sysctl_tcp_reordering
)
1971 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1972 sysctl_tcp_reordering
);
1973 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1974 tp
->high_seq
= tp
->snd_nxt
;
1975 TCP_ECN_queue_cwr(tp
);
1977 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1978 * loss recovery is underway except recurring timeout(s) on
1979 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1981 tp
->frto
= sysctl_tcp_frto
&&
1982 (new_recovery
|| icsk
->icsk_retransmits
) &&
1983 !inet_csk(sk
)->icsk_mtup
.probe_size
;
1986 /* If ACK arrived pointing to a remembered SACK, it means that our
1987 * remembered SACKs do not reflect real state of receiver i.e.
1988 * receiver _host_ is heavily congested (or buggy).
1990 * To avoid big spurious retransmission bursts due to transient SACK
1991 * scoreboard oddities that look like reneging, we give the receiver a
1992 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
1993 * restore sanity to the SACK scoreboard. If the apparent reneging
1994 * persists until this RTO then we'll clear the SACK scoreboard.
1996 static bool tcp_check_sack_reneging(struct sock
*sk
, int flag
)
1998 if (flag
& FLAG_SACK_RENEGING
) {
1999 struct tcp_sock
*tp
= tcp_sk(sk
);
2000 unsigned long delay
= max(usecs_to_jiffies(tp
->srtt_us
>> 4),
2001 msecs_to_jiffies(10));
2003 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2004 delay
, TCP_RTO_MAX
);
2010 static inline int tcp_fackets_out(const struct tcp_sock
*tp
)
2012 return tcp_is_reno(tp
) ? tp
->sacked_out
+ 1 : tp
->fackets_out
;
2015 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2016 * counter when SACK is enabled (without SACK, sacked_out is used for
2019 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2020 * segments up to the highest received SACK block so far and holes in
2023 * With reordering, holes may still be in flight, so RFC3517 recovery
2024 * uses pure sacked_out (total number of SACKed segments) even though
2025 * it violates the RFC that uses duplicate ACKs, often these are equal
2026 * but when e.g. out-of-window ACKs or packet duplication occurs,
2027 * they differ. Since neither occurs due to loss, TCP should really
2030 static inline int tcp_dupack_heuristics(const struct tcp_sock
*tp
)
2032 return tcp_is_fack(tp
) ? tp
->fackets_out
: tp
->sacked_out
+ 1;
2035 static bool tcp_pause_early_retransmit(struct sock
*sk
, int flag
)
2037 struct tcp_sock
*tp
= tcp_sk(sk
);
2038 unsigned long delay
;
2040 /* Delay early retransmit and entering fast recovery for
2041 * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
2042 * available, or RTO is scheduled to fire first.
2044 if (sysctl_tcp_early_retrans
< 2 || sysctl_tcp_early_retrans
> 3 ||
2045 (flag
& FLAG_ECE
) || !tp
->srtt_us
)
2048 delay
= max(usecs_to_jiffies(tp
->srtt_us
>> 5),
2049 msecs_to_jiffies(2));
2051 if (!time_after(inet_csk(sk
)->icsk_timeout
, (jiffies
+ delay
)))
2054 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_EARLY_RETRANS
, delay
,
2059 /* Linux NewReno/SACK/FACK/ECN state machine.
2060 * --------------------------------------
2062 * "Open" Normal state, no dubious events, fast path.
2063 * "Disorder" In all the respects it is "Open",
2064 * but requires a bit more attention. It is entered when
2065 * we see some SACKs or dupacks. It is split of "Open"
2066 * mainly to move some processing from fast path to slow one.
2067 * "CWR" CWND was reduced due to some Congestion Notification event.
2068 * It can be ECN, ICMP source quench, local device congestion.
2069 * "Recovery" CWND was reduced, we are fast-retransmitting.
2070 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2072 * tcp_fastretrans_alert() is entered:
2073 * - each incoming ACK, if state is not "Open"
2074 * - when arrived ACK is unusual, namely:
2079 * Counting packets in flight is pretty simple.
2081 * in_flight = packets_out - left_out + retrans_out
2083 * packets_out is SND.NXT-SND.UNA counted in packets.
2085 * retrans_out is number of retransmitted segments.
2087 * left_out is number of segments left network, but not ACKed yet.
2089 * left_out = sacked_out + lost_out
2091 * sacked_out: Packets, which arrived to receiver out of order
2092 * and hence not ACKed. With SACKs this number is simply
2093 * amount of SACKed data. Even without SACKs
2094 * it is easy to give pretty reliable estimate of this number,
2095 * counting duplicate ACKs.
2097 * lost_out: Packets lost by network. TCP has no explicit
2098 * "loss notification" feedback from network (for now).
2099 * It means that this number can be only _guessed_.
2100 * Actually, it is the heuristics to predict lossage that
2101 * distinguishes different algorithms.
2103 * F.e. after RTO, when all the queue is considered as lost,
2104 * lost_out = packets_out and in_flight = retrans_out.
2106 * Essentially, we have now two algorithms counting
2109 * FACK: It is the simplest heuristics. As soon as we decided
2110 * that something is lost, we decide that _all_ not SACKed
2111 * packets until the most forward SACK are lost. I.e.
2112 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2113 * It is absolutely correct estimate, if network does not reorder
2114 * packets. And it loses any connection to reality when reordering
2115 * takes place. We use FACK by default until reordering
2116 * is suspected on the path to this destination.
2118 * NewReno: when Recovery is entered, we assume that one segment
2119 * is lost (classic Reno). While we are in Recovery and
2120 * a partial ACK arrives, we assume that one more packet
2121 * is lost (NewReno). This heuristics are the same in NewReno
2124 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2125 * deflation etc. CWND is real congestion window, never inflated, changes
2126 * only according to classic VJ rules.
2128 * Really tricky (and requiring careful tuning) part of algorithm
2129 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2130 * The first determines the moment _when_ we should reduce CWND and,
2131 * hence, slow down forward transmission. In fact, it determines the moment
2132 * when we decide that hole is caused by loss, rather than by a reorder.
2134 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2135 * holes, caused by lost packets.
2137 * And the most logically complicated part of algorithm is undo
2138 * heuristics. We detect false retransmits due to both too early
2139 * fast retransmit (reordering) and underestimated RTO, analyzing
2140 * timestamps and D-SACKs. When we detect that some segments were
2141 * retransmitted by mistake and CWND reduction was wrong, we undo
2142 * window reduction and abort recovery phase. This logic is hidden
2143 * inside several functions named tcp_try_undo_<something>.
2146 /* This function decides, when we should leave Disordered state
2147 * and enter Recovery phase, reducing congestion window.
2149 * Main question: may we further continue forward transmission
2150 * with the same cwnd?
2152 static bool tcp_time_to_recover(struct sock
*sk
, int flag
)
2154 struct tcp_sock
*tp
= tcp_sk(sk
);
2157 /* Trick#1: The loss is proven. */
2161 /* Not-A-Trick#2 : Classic rule... */
2162 if (tcp_dupack_heuristics(tp
) > tp
->reordering
)
2165 /* Trick#4: It is still not OK... But will it be useful to delay
2168 packets_out
= tp
->packets_out
;
2169 if (packets_out
<= tp
->reordering
&&
2170 tp
->sacked_out
>= max_t(__u32
, packets_out
/2, sysctl_tcp_reordering
) &&
2171 !tcp_may_send_now(sk
)) {
2172 /* We have nothing to send. This connection is limited
2173 * either by receiver window or by application.
2178 /* If a thin stream is detected, retransmit after first
2179 * received dupack. Employ only if SACK is supported in order
2180 * to avoid possible corner-case series of spurious retransmissions
2181 * Use only if there are no unsent data.
2183 if ((tp
->thin_dupack
|| sysctl_tcp_thin_dupack
) &&
2184 tcp_stream_is_thin(tp
) && tcp_dupack_heuristics(tp
) > 1 &&
2185 tcp_is_sack(tp
) && !tcp_send_head(sk
))
2188 /* Trick#6: TCP early retransmit, per RFC5827. To avoid spurious
2189 * retransmissions due to small network reorderings, we implement
2190 * Mitigation A.3 in the RFC and delay the retransmission for a short
2191 * interval if appropriate.
2193 if (tp
->do_early_retrans
&& !tp
->retrans_out
&& tp
->sacked_out
&&
2194 (tp
->packets_out
>= (tp
->sacked_out
+ 1) && tp
->packets_out
< 4) &&
2195 !tcp_may_send_now(sk
))
2196 return !tcp_pause_early_retransmit(sk
, flag
);
2201 /* Detect loss in event "A" above by marking head of queue up as lost.
2202 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2203 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2204 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2205 * the maximum SACKed segments to pass before reaching this limit.
2207 static void tcp_mark_head_lost(struct sock
*sk
, int packets
, int mark_head
)
2209 struct tcp_sock
*tp
= tcp_sk(sk
);
2210 struct sk_buff
*skb
;
2214 /* Use SACK to deduce losses of new sequences sent during recovery */
2215 const u32 loss_high
= tcp_is_sack(tp
) ? tp
->snd_nxt
: tp
->high_seq
;
2217 WARN_ON(packets
> tp
->packets_out
);
2218 if (tp
->lost_skb_hint
) {
2219 skb
= tp
->lost_skb_hint
;
2220 cnt
= tp
->lost_cnt_hint
;
2221 /* Head already handled? */
2222 if (mark_head
&& skb
!= tcp_write_queue_head(sk
))
2225 skb
= tcp_write_queue_head(sk
);
2229 tcp_for_write_queue_from(skb
, sk
) {
2230 if (skb
== tcp_send_head(sk
))
2232 /* TODO: do this better */
2233 /* this is not the most efficient way to do this... */
2234 tp
->lost_skb_hint
= skb
;
2235 tp
->lost_cnt_hint
= cnt
;
2237 if (after(TCP_SKB_CB(skb
)->end_seq
, loss_high
))
2241 if (tcp_is_fack(tp
) || tcp_is_reno(tp
) ||
2242 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2243 cnt
+= tcp_skb_pcount(skb
);
2245 if (cnt
> packets
) {
2246 if ((tcp_is_sack(tp
) && !tcp_is_fack(tp
)) ||
2247 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
) ||
2248 (oldcnt
>= packets
))
2251 mss
= skb_shinfo(skb
)->gso_size
;
2252 err
= tcp_fragment(sk
, skb
, (packets
- oldcnt
) * mss
,
2259 tcp_skb_mark_lost(tp
, skb
);
2264 tcp_verify_left_out(tp
);
2267 /* Account newly detected lost packet(s) */
2269 static void tcp_update_scoreboard(struct sock
*sk
, int fast_rexmit
)
2271 struct tcp_sock
*tp
= tcp_sk(sk
);
2273 if (tcp_is_reno(tp
)) {
2274 tcp_mark_head_lost(sk
, 1, 1);
2275 } else if (tcp_is_fack(tp
)) {
2276 int lost
= tp
->fackets_out
- tp
->reordering
;
2279 tcp_mark_head_lost(sk
, lost
, 0);
2281 int sacked_upto
= tp
->sacked_out
- tp
->reordering
;
2282 if (sacked_upto
>= 0)
2283 tcp_mark_head_lost(sk
, sacked_upto
, 0);
2284 else if (fast_rexmit
)
2285 tcp_mark_head_lost(sk
, 1, 1);
2289 /* CWND moderation, preventing bursts due to too big ACKs
2290 * in dubious situations.
2292 static inline void tcp_moderate_cwnd(struct tcp_sock
*tp
)
2294 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
2295 tcp_packets_in_flight(tp
) + tcp_max_burst(tp
));
2296 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2299 /* Nothing was retransmitted or returned timestamp is less
2300 * than timestamp of the first retransmission.
2302 static inline bool tcp_packet_delayed(const struct tcp_sock
*tp
)
2304 return !tp
->retrans_stamp
||
2305 (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2306 before(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
));
2309 /* Undo procedures. */
2311 #if FASTRETRANS_DEBUG > 1
2312 static void DBGUNDO(struct sock
*sk
, const char *msg
)
2314 struct tcp_sock
*tp
= tcp_sk(sk
);
2315 struct inet_sock
*inet
= inet_sk(sk
);
2317 if (sk
->sk_family
== AF_INET
) {
2318 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2320 &inet
->inet_daddr
, ntohs(inet
->inet_dport
),
2321 tp
->snd_cwnd
, tcp_left_out(tp
),
2322 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2325 #if IS_ENABLED(CONFIG_IPV6)
2326 else if (sk
->sk_family
== AF_INET6
) {
2327 struct ipv6_pinfo
*np
= inet6_sk(sk
);
2328 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2330 &np
->daddr
, ntohs(inet
->inet_dport
),
2331 tp
->snd_cwnd
, tcp_left_out(tp
),
2332 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2338 #define DBGUNDO(x...) do { } while (0)
2341 static void tcp_undo_cwnd_reduction(struct sock
*sk
, bool unmark_loss
)
2343 struct tcp_sock
*tp
= tcp_sk(sk
);
2346 struct sk_buff
*skb
;
2348 tcp_for_write_queue(skb
, sk
) {
2349 if (skb
== tcp_send_head(sk
))
2351 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2354 tcp_clear_all_retrans_hints(tp
);
2357 if (tp
->prior_ssthresh
) {
2358 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2360 if (icsk
->icsk_ca_ops
->undo_cwnd
)
2361 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
2363 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
<< 1);
2365 if (tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
2366 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
2367 TCP_ECN_withdraw_cwr(tp
);
2370 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2372 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2373 tp
->undo_marker
= 0;
2376 static inline bool tcp_may_undo(const struct tcp_sock
*tp
)
2378 return tp
->undo_marker
&& (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
2381 /* People celebrate: "We love our President!" */
2382 static bool tcp_try_undo_recovery(struct sock
*sk
)
2384 struct tcp_sock
*tp
= tcp_sk(sk
);
2386 if (tcp_may_undo(tp
)) {
2389 /* Happy end! We did not retransmit anything
2390 * or our original transmission succeeded.
2392 DBGUNDO(sk
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
2393 tcp_undo_cwnd_reduction(sk
, false);
2394 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
2395 mib_idx
= LINUX_MIB_TCPLOSSUNDO
;
2397 mib_idx
= LINUX_MIB_TCPFULLUNDO
;
2399 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2401 if (tp
->snd_una
== tp
->high_seq
&& tcp_is_reno(tp
)) {
2402 /* Hold old state until something *above* high_seq
2403 * is ACKed. For Reno it is MUST to prevent false
2404 * fast retransmits (RFC2582). SACK TCP is safe. */
2405 tcp_moderate_cwnd(tp
);
2408 tcp_set_ca_state(sk
, TCP_CA_Open
);
2412 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2413 static bool tcp_try_undo_dsack(struct sock
*sk
)
2415 struct tcp_sock
*tp
= tcp_sk(sk
);
2417 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
2418 DBGUNDO(sk
, "D-SACK");
2419 tcp_undo_cwnd_reduction(sk
, false);
2420 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPDSACKUNDO
);
2426 /* We can clear retrans_stamp when there are no retransmissions in the
2427 * window. It would seem that it is trivially available for us in
2428 * tp->retrans_out, however, that kind of assumptions doesn't consider
2429 * what will happen if errors occur when sending retransmission for the
2430 * second time. ...It could the that such segment has only
2431 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2432 * the head skb is enough except for some reneging corner cases that
2433 * are not worth the effort.
2435 * Main reason for all this complexity is the fact that connection dying
2436 * time now depends on the validity of the retrans_stamp, in particular,
2437 * that successive retransmissions of a segment must not advance
2438 * retrans_stamp under any conditions.
2440 static bool tcp_any_retrans_done(const struct sock
*sk
)
2442 const struct tcp_sock
*tp
= tcp_sk(sk
);
2443 struct sk_buff
*skb
;
2445 if (tp
->retrans_out
)
2448 skb
= tcp_write_queue_head(sk
);
2449 if (unlikely(skb
&& TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
))
2455 /* Undo during loss recovery after partial ACK or using F-RTO. */
2456 static bool tcp_try_undo_loss(struct sock
*sk
, bool frto_undo
)
2458 struct tcp_sock
*tp
= tcp_sk(sk
);
2460 if (frto_undo
|| tcp_may_undo(tp
)) {
2461 tcp_undo_cwnd_reduction(sk
, true);
2463 DBGUNDO(sk
, "partial loss");
2464 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSSUNDO
);
2466 NET_INC_STATS_BH(sock_net(sk
),
2467 LINUX_MIB_TCPSPURIOUSRTOS
);
2468 inet_csk(sk
)->icsk_retransmits
= 0;
2469 if (frto_undo
|| tcp_is_sack(tp
))
2470 tcp_set_ca_state(sk
, TCP_CA_Open
);
2476 /* The cwnd reduction in CWR and Recovery use the PRR algorithm
2477 * https://datatracker.ietf.org/doc/draft-ietf-tcpm-proportional-rate-reduction/
2478 * It computes the number of packets to send (sndcnt) based on packets newly
2480 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2481 * cwnd reductions across a full RTT.
2482 * 2) If packets in flight is lower than ssthresh (such as due to excess
2483 * losses and/or application stalls), do not perform any further cwnd
2484 * reductions, but instead slow start up to ssthresh.
2486 static void tcp_init_cwnd_reduction(struct sock
*sk
)
2488 struct tcp_sock
*tp
= tcp_sk(sk
);
2490 tp
->high_seq
= tp
->snd_nxt
;
2491 tp
->tlp_high_seq
= 0;
2492 tp
->snd_cwnd_cnt
= 0;
2493 tp
->prior_cwnd
= tp
->snd_cwnd
;
2494 tp
->prr_delivered
= 0;
2496 tp
->snd_ssthresh
= inet_csk(sk
)->icsk_ca_ops
->ssthresh(sk
);
2497 TCP_ECN_queue_cwr(tp
);
2500 static void tcp_cwnd_reduction(struct sock
*sk
, const int prior_unsacked
,
2503 struct tcp_sock
*tp
= tcp_sk(sk
);
2505 int delta
= tp
->snd_ssthresh
- tcp_packets_in_flight(tp
);
2506 int newly_acked_sacked
= prior_unsacked
-
2507 (tp
->packets_out
- tp
->sacked_out
);
2509 tp
->prr_delivered
+= newly_acked_sacked
;
2510 if (tcp_packets_in_flight(tp
) > tp
->snd_ssthresh
) {
2511 u64 dividend
= (u64
)tp
->snd_ssthresh
* tp
->prr_delivered
+
2513 sndcnt
= div_u64(dividend
, tp
->prior_cwnd
) - tp
->prr_out
;
2515 sndcnt
= min_t(int, delta
,
2516 max_t(int, tp
->prr_delivered
- tp
->prr_out
,
2517 newly_acked_sacked
) + 1);
2520 sndcnt
= max(sndcnt
, (fast_rexmit
? 1 : 0));
2521 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + sndcnt
;
2524 static inline void tcp_end_cwnd_reduction(struct sock
*sk
)
2526 struct tcp_sock
*tp
= tcp_sk(sk
);
2528 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2529 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_CWR
||
2530 (tp
->undo_marker
&& tp
->snd_ssthresh
< TCP_INFINITE_SSTHRESH
)) {
2531 tp
->snd_cwnd
= tp
->snd_ssthresh
;
2532 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2534 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
2537 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2538 void tcp_enter_cwr(struct sock
*sk
)
2540 struct tcp_sock
*tp
= tcp_sk(sk
);
2542 tp
->prior_ssthresh
= 0;
2543 if (inet_csk(sk
)->icsk_ca_state
< TCP_CA_CWR
) {
2544 tp
->undo_marker
= 0;
2545 tcp_init_cwnd_reduction(sk
);
2546 tcp_set_ca_state(sk
, TCP_CA_CWR
);
2550 static void tcp_try_keep_open(struct sock
*sk
)
2552 struct tcp_sock
*tp
= tcp_sk(sk
);
2553 int state
= TCP_CA_Open
;
2555 if (tcp_left_out(tp
) || tcp_any_retrans_done(sk
))
2556 state
= TCP_CA_Disorder
;
2558 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
2559 tcp_set_ca_state(sk
, state
);
2560 tp
->high_seq
= tp
->snd_nxt
;
2564 static void tcp_try_to_open(struct sock
*sk
, int flag
, const int prior_unsacked
)
2566 struct tcp_sock
*tp
= tcp_sk(sk
);
2568 tcp_verify_left_out(tp
);
2570 if (!tcp_any_retrans_done(sk
))
2571 tp
->retrans_stamp
= 0;
2573 if (flag
& FLAG_ECE
)
2576 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
2577 tcp_try_keep_open(sk
);
2579 tcp_cwnd_reduction(sk
, prior_unsacked
, 0);
2583 static void tcp_mtup_probe_failed(struct sock
*sk
)
2585 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2587 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
2588 icsk
->icsk_mtup
.probe_size
= 0;
2591 static void tcp_mtup_probe_success(struct sock
*sk
)
2593 struct tcp_sock
*tp
= tcp_sk(sk
);
2594 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2596 /* FIXME: breaks with very large cwnd */
2597 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2598 tp
->snd_cwnd
= tp
->snd_cwnd
*
2599 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
2600 icsk
->icsk_mtup
.probe_size
;
2601 tp
->snd_cwnd_cnt
= 0;
2602 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2603 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2605 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
2606 icsk
->icsk_mtup
.probe_size
= 0;
2607 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
2610 /* Do a simple retransmit without using the backoff mechanisms in
2611 * tcp_timer. This is used for path mtu discovery.
2612 * The socket is already locked here.
2614 void tcp_simple_retransmit(struct sock
*sk
)
2616 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2617 struct tcp_sock
*tp
= tcp_sk(sk
);
2618 struct sk_buff
*skb
;
2619 unsigned int mss
= tcp_current_mss(sk
);
2620 u32 prior_lost
= tp
->lost_out
;
2622 tcp_for_write_queue(skb
, sk
) {
2623 if (skb
== tcp_send_head(sk
))
2625 if (tcp_skb_seglen(skb
) > mss
&&
2626 !(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
2627 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2628 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
2629 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2631 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
2635 tcp_clear_retrans_hints_partial(tp
);
2637 if (prior_lost
== tp
->lost_out
)
2640 if (tcp_is_reno(tp
))
2641 tcp_limit_reno_sacked(tp
);
2643 tcp_verify_left_out(tp
);
2645 /* Don't muck with the congestion window here.
2646 * Reason is that we do not increase amount of _data_
2647 * in network, but units changed and effective
2648 * cwnd/ssthresh really reduced now.
2650 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
2651 tp
->high_seq
= tp
->snd_nxt
;
2652 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2653 tp
->prior_ssthresh
= 0;
2654 tp
->undo_marker
= 0;
2655 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2657 tcp_xmit_retransmit_queue(sk
);
2659 EXPORT_SYMBOL(tcp_simple_retransmit
);
2661 static void tcp_enter_recovery(struct sock
*sk
, bool ece_ack
)
2663 struct tcp_sock
*tp
= tcp_sk(sk
);
2666 if (tcp_is_reno(tp
))
2667 mib_idx
= LINUX_MIB_TCPRENORECOVERY
;
2669 mib_idx
= LINUX_MIB_TCPSACKRECOVERY
;
2671 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2673 tp
->prior_ssthresh
= 0;
2674 tp
->undo_marker
= tp
->snd_una
;
2675 tp
->undo_retrans
= tp
->retrans_out
? : -1;
2677 if (inet_csk(sk
)->icsk_ca_state
< TCP_CA_CWR
) {
2679 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2680 tcp_init_cwnd_reduction(sk
);
2682 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
2685 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2686 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2688 static void tcp_process_loss(struct sock
*sk
, int flag
, bool is_dupack
)
2690 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2691 struct tcp_sock
*tp
= tcp_sk(sk
);
2692 bool recovered
= !before(tp
->snd_una
, tp
->high_seq
);
2694 if (tp
->frto
) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2695 /* Step 3.b. A timeout is spurious if not all data are
2696 * lost, i.e., never-retransmitted data are (s)acked.
2698 if (tcp_try_undo_loss(sk
, flag
& FLAG_ORIG_SACK_ACKED
))
2701 if (after(tp
->snd_nxt
, tp
->high_seq
) &&
2702 (flag
& FLAG_DATA_SACKED
|| is_dupack
)) {
2703 tp
->frto
= 0; /* Loss was real: 2nd part of step 3.a */
2704 } else if (flag
& FLAG_SND_UNA_ADVANCED
&& !recovered
) {
2705 tp
->high_seq
= tp
->snd_nxt
;
2706 __tcp_push_pending_frames(sk
, tcp_current_mss(sk
),
2708 if (after(tp
->snd_nxt
, tp
->high_seq
))
2709 return; /* Step 2.b */
2715 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2716 icsk
->icsk_retransmits
= 0;
2717 tcp_try_undo_recovery(sk
);
2720 if (flag
& FLAG_DATA_ACKED
)
2721 icsk
->icsk_retransmits
= 0;
2722 if (tcp_is_reno(tp
)) {
2723 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2724 * delivered. Lower inflight to clock out (re)tranmissions.
2726 if (after(tp
->snd_nxt
, tp
->high_seq
) && is_dupack
)
2727 tcp_add_reno_sack(sk
);
2728 else if (flag
& FLAG_SND_UNA_ADVANCED
)
2729 tcp_reset_reno_sack(tp
);
2731 if (tcp_try_undo_loss(sk
, false))
2733 tcp_xmit_retransmit_queue(sk
);
2736 /* Undo during fast recovery after partial ACK. */
2737 static bool tcp_try_undo_partial(struct sock
*sk
, const int acked
,
2738 const int prior_unsacked
)
2740 struct tcp_sock
*tp
= tcp_sk(sk
);
2742 if (tp
->undo_marker
&& tcp_packet_delayed(tp
)) {
2743 /* Plain luck! Hole if filled with delayed
2744 * packet, rather than with a retransmit.
2746 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
2748 /* We are getting evidence that the reordering degree is higher
2749 * than we realized. If there are no retransmits out then we
2750 * can undo. Otherwise we clock out new packets but do not
2751 * mark more packets lost or retransmit more.
2753 if (tp
->retrans_out
) {
2754 tcp_cwnd_reduction(sk
, prior_unsacked
, 0);
2758 if (!tcp_any_retrans_done(sk
))
2759 tp
->retrans_stamp
= 0;
2761 DBGUNDO(sk
, "partial recovery");
2762 tcp_undo_cwnd_reduction(sk
, true);
2763 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPARTIALUNDO
);
2764 tcp_try_keep_open(sk
);
2770 /* Process an event, which can update packets-in-flight not trivially.
2771 * Main goal of this function is to calculate new estimate for left_out,
2772 * taking into account both packets sitting in receiver's buffer and
2773 * packets lost by network.
2775 * Besides that it does CWND reduction, when packet loss is detected
2776 * and changes state of machine.
2778 * It does _not_ decide what to send, it is made in function
2779 * tcp_xmit_retransmit_queue().
2781 static void tcp_fastretrans_alert(struct sock
*sk
, const int acked
,
2782 const int prior_unsacked
,
2783 bool is_dupack
, int flag
)
2785 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2786 struct tcp_sock
*tp
= tcp_sk(sk
);
2787 bool do_lost
= is_dupack
|| ((flag
& FLAG_DATA_SACKED
) &&
2788 (tcp_fackets_out(tp
) > tp
->reordering
));
2789 int fast_rexmit
= 0;
2791 if (WARN_ON(!tp
->packets_out
&& tp
->sacked_out
))
2793 if (WARN_ON(!tp
->sacked_out
&& tp
->fackets_out
))
2794 tp
->fackets_out
= 0;
2796 /* Now state machine starts.
2797 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2798 if (flag
& FLAG_ECE
)
2799 tp
->prior_ssthresh
= 0;
2801 /* B. In all the states check for reneging SACKs. */
2802 if (tcp_check_sack_reneging(sk
, flag
))
2805 /* C. Check consistency of the current state. */
2806 tcp_verify_left_out(tp
);
2808 /* D. Check state exit conditions. State can be terminated
2809 * when high_seq is ACKed. */
2810 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
2811 WARN_ON(tp
->retrans_out
!= 0);
2812 tp
->retrans_stamp
= 0;
2813 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
2814 switch (icsk
->icsk_ca_state
) {
2816 /* CWR is to be held something *above* high_seq
2817 * is ACKed for CWR bit to reach receiver. */
2818 if (tp
->snd_una
!= tp
->high_seq
) {
2819 tcp_end_cwnd_reduction(sk
);
2820 tcp_set_ca_state(sk
, TCP_CA_Open
);
2824 case TCP_CA_Recovery
:
2825 if (tcp_is_reno(tp
))
2826 tcp_reset_reno_sack(tp
);
2827 if (tcp_try_undo_recovery(sk
))
2829 tcp_end_cwnd_reduction(sk
);
2834 /* E. Process state. */
2835 switch (icsk
->icsk_ca_state
) {
2836 case TCP_CA_Recovery
:
2837 if (!(flag
& FLAG_SND_UNA_ADVANCED
)) {
2838 if (tcp_is_reno(tp
) && is_dupack
)
2839 tcp_add_reno_sack(sk
);
2841 if (tcp_try_undo_partial(sk
, acked
, prior_unsacked
))
2843 /* Partial ACK arrived. Force fast retransmit. */
2844 do_lost
= tcp_is_reno(tp
) ||
2845 tcp_fackets_out(tp
) > tp
->reordering
;
2847 if (tcp_try_undo_dsack(sk
)) {
2848 tcp_try_keep_open(sk
);
2853 tcp_process_loss(sk
, flag
, is_dupack
);
2854 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
2856 /* Fall through to processing in Open state. */
2858 if (tcp_is_reno(tp
)) {
2859 if (flag
& FLAG_SND_UNA_ADVANCED
)
2860 tcp_reset_reno_sack(tp
);
2862 tcp_add_reno_sack(sk
);
2865 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
)
2866 tcp_try_undo_dsack(sk
);
2868 if (!tcp_time_to_recover(sk
, flag
)) {
2869 tcp_try_to_open(sk
, flag
, prior_unsacked
);
2873 /* MTU probe failure: don't reduce cwnd */
2874 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
2875 icsk
->icsk_mtup
.probe_size
&&
2876 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
2877 tcp_mtup_probe_failed(sk
);
2878 /* Restores the reduction we did in tcp_mtup_probe() */
2880 tcp_simple_retransmit(sk
);
2884 /* Otherwise enter Recovery state */
2885 tcp_enter_recovery(sk
, (flag
& FLAG_ECE
));
2890 tcp_update_scoreboard(sk
, fast_rexmit
);
2891 tcp_cwnd_reduction(sk
, prior_unsacked
, fast_rexmit
);
2892 tcp_xmit_retransmit_queue(sk
);
2895 static inline bool tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
2896 long seq_rtt_us
, long sack_rtt_us
)
2898 const struct tcp_sock
*tp
= tcp_sk(sk
);
2900 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2901 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2902 * Karn's algorithm forbids taking RTT if some retransmitted data
2903 * is acked (RFC6298).
2905 if (flag
& FLAG_RETRANS_DATA_ACKED
)
2909 seq_rtt_us
= sack_rtt_us
;
2911 /* RTTM Rule: A TSecr value received in a segment is used to
2912 * update the averaged RTT measurement only if the segment
2913 * acknowledges some new data, i.e., only if it advances the
2914 * left edge of the send window.
2915 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2917 if (seq_rtt_us
< 0 && tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2919 seq_rtt_us
= jiffies_to_usecs(tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
);
2924 tcp_rtt_estimator(sk
, seq_rtt_us
);
2927 /* RFC6298: only reset backoff on valid RTT measurement. */
2928 inet_csk(sk
)->icsk_backoff
= 0;
2932 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2933 static void tcp_synack_rtt_meas(struct sock
*sk
, const u32 synack_stamp
)
2935 struct tcp_sock
*tp
= tcp_sk(sk
);
2936 long seq_rtt_us
= -1L;
2938 if (synack_stamp
&& !tp
->total_retrans
)
2939 seq_rtt_us
= jiffies_to_usecs(tcp_time_stamp
- synack_stamp
);
2941 /* If the ACK acks both the SYNACK and the (Fast Open'd) data packets
2942 * sent in SYN_RECV, SYNACK RTT is the smooth RTT computed in tcp_ack()
2945 tcp_ack_update_rtt(sk
, FLAG_SYN_ACKED
, seq_rtt_us
, -1L);
2948 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
, u32 acked
)
2950 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2952 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, acked
);
2953 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_time_stamp
;
2956 /* Restart timer after forward progress on connection.
2957 * RFC2988 recommends to restart timer to now+rto.
2959 void tcp_rearm_rto(struct sock
*sk
)
2961 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2962 struct tcp_sock
*tp
= tcp_sk(sk
);
2964 /* If the retrans timer is currently being used by Fast Open
2965 * for SYN-ACK retrans purpose, stay put.
2967 if (tp
->fastopen_rsk
)
2970 if (!tp
->packets_out
) {
2971 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
2973 u32 rto
= inet_csk(sk
)->icsk_rto
;
2974 /* Offset the time elapsed after installing regular RTO */
2975 if (icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
||
2976 icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
) {
2977 struct sk_buff
*skb
= tcp_write_queue_head(sk
);
2978 const u32 rto_time_stamp
= TCP_SKB_CB(skb
)->when
+ rto
;
2979 s32 delta
= (s32
)(rto_time_stamp
- tcp_time_stamp
);
2980 /* delta may not be positive if the socket is locked
2981 * when the retrans timer fires and is rescheduled.
2986 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
, rto
,
2991 /* This function is called when the delayed ER timer fires. TCP enters
2992 * fast recovery and performs fast-retransmit.
2994 void tcp_resume_early_retransmit(struct sock
*sk
)
2996 struct tcp_sock
*tp
= tcp_sk(sk
);
3000 /* Stop if ER is disabled after the delayed ER timer is scheduled */
3001 if (!tp
->do_early_retrans
)
3004 tcp_enter_recovery(sk
, false);
3005 tcp_update_scoreboard(sk
, 1);
3006 tcp_xmit_retransmit_queue(sk
);
3009 /* If we get here, the whole TSO packet has not been acked. */
3010 static u32
tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
)
3012 struct tcp_sock
*tp
= tcp_sk(sk
);
3015 BUG_ON(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
));
3017 packets_acked
= tcp_skb_pcount(skb
);
3018 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
3020 packets_acked
-= tcp_skb_pcount(skb
);
3022 if (packets_acked
) {
3023 BUG_ON(tcp_skb_pcount(skb
) == 0);
3024 BUG_ON(!before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
));
3027 return packets_acked
;
3030 /* Remove acknowledged frames from the retransmission queue. If our packet
3031 * is before the ack sequence we can discard it as it's confirmed to have
3032 * arrived at the other end.
3034 static int tcp_clean_rtx_queue(struct sock
*sk
, int prior_fackets
,
3035 u32 prior_snd_una
, long sack_rtt_us
)
3037 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3038 struct skb_mstamp first_ackt
, last_ackt
, now
;
3039 struct tcp_sock
*tp
= tcp_sk(sk
);
3040 u32 prior_sacked
= tp
->sacked_out
;
3041 u32 reord
= tp
->packets_out
;
3042 bool fully_acked
= true;
3043 long ca_seq_rtt_us
= -1L;
3044 long seq_rtt_us
= -1L;
3045 struct sk_buff
*skb
;
3052 while ((skb
= tcp_write_queue_head(sk
)) && skb
!= tcp_send_head(sk
)) {
3053 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
3054 u8 sacked
= scb
->sacked
;
3057 /* Determine how many packets and what bytes were acked, tso and else */
3058 if (after(scb
->end_seq
, tp
->snd_una
)) {
3059 if (tcp_skb_pcount(skb
) == 1 ||
3060 !after(tp
->snd_una
, scb
->seq
))
3063 acked_pcount
= tcp_tso_acked(sk
, skb
);
3067 fully_acked
= false;
3069 acked_pcount
= tcp_skb_pcount(skb
);
3072 if (sacked
& TCPCB_RETRANS
) {
3073 if (sacked
& TCPCB_SACKED_RETRANS
)
3074 tp
->retrans_out
-= acked_pcount
;
3075 flag
|= FLAG_RETRANS_DATA_ACKED
;
3077 last_ackt
= skb
->skb_mstamp
;
3078 WARN_ON_ONCE(last_ackt
.v64
== 0);
3079 if (!first_ackt
.v64
)
3080 first_ackt
= last_ackt
;
3082 if (!(sacked
& TCPCB_SACKED_ACKED
))
3083 reord
= min(pkts_acked
, reord
);
3084 if (!after(scb
->end_seq
, tp
->high_seq
))
3085 flag
|= FLAG_ORIG_SACK_ACKED
;
3088 if (sacked
& TCPCB_SACKED_ACKED
)
3089 tp
->sacked_out
-= acked_pcount
;
3090 if (sacked
& TCPCB_LOST
)
3091 tp
->lost_out
-= acked_pcount
;
3093 tp
->packets_out
-= acked_pcount
;
3094 pkts_acked
+= acked_pcount
;
3096 /* Initial outgoing SYN's get put onto the write_queue
3097 * just like anything else we transmit. It is not
3098 * true data, and if we misinform our callers that
3099 * this ACK acks real data, we will erroneously exit
3100 * connection startup slow start one packet too
3101 * quickly. This is severely frowned upon behavior.
3103 if (!(scb
->tcp_flags
& TCPHDR_SYN
)) {
3104 flag
|= FLAG_DATA_ACKED
;
3106 flag
|= FLAG_SYN_ACKED
;
3107 tp
->retrans_stamp
= 0;
3110 if (unlikely(skb_shinfo(skb
)->tx_flags
& SKBTX_ACK_TSTAMP
) &&
3111 between(skb_shinfo(skb
)->tskey
, prior_snd_una
,
3113 __skb_tstamp_tx(skb
, NULL
, sk
, SCM_TSTAMP_ACK
);
3118 tcp_unlink_write_queue(skb
, sk
);
3119 sk_wmem_free_skb(sk
, skb
);
3120 if (skb
== tp
->retransmit_skb_hint
)
3121 tp
->retransmit_skb_hint
= NULL
;
3122 if (skb
== tp
->lost_skb_hint
)
3123 tp
->lost_skb_hint
= NULL
;
3126 if (likely(between(tp
->snd_up
, prior_snd_una
, tp
->snd_una
)))
3127 tp
->snd_up
= tp
->snd_una
;
3129 if (skb
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
3130 flag
|= FLAG_SACK_RENEGING
;
3132 skb_mstamp_get(&now
);
3133 if (first_ackt
.v64
) {
3134 seq_rtt_us
= skb_mstamp_us_delta(&now
, &first_ackt
);
3135 ca_seq_rtt_us
= skb_mstamp_us_delta(&now
, &last_ackt
);
3138 rtt_update
= tcp_ack_update_rtt(sk
, flag
, seq_rtt_us
, sack_rtt_us
);
3140 if (flag
& FLAG_ACKED
) {
3141 const struct tcp_congestion_ops
*ca_ops
3142 = inet_csk(sk
)->icsk_ca_ops
;
3145 if (unlikely(icsk
->icsk_mtup
.probe_size
&&
3146 !after(tp
->mtu_probe
.probe_seq_end
, tp
->snd_una
))) {
3147 tcp_mtup_probe_success(sk
);
3150 if (tcp_is_reno(tp
)) {
3151 tcp_remove_reno_sacks(sk
, pkts_acked
);
3155 /* Non-retransmitted hole got filled? That's reordering */
3156 if (reord
< prior_fackets
)
3157 tcp_update_reordering(sk
, tp
->fackets_out
- reord
, 0);
3159 delta
= tcp_is_fack(tp
) ? pkts_acked
:
3160 prior_sacked
- tp
->sacked_out
;
3161 tp
->lost_cnt_hint
-= min(tp
->lost_cnt_hint
, delta
);
3164 tp
->fackets_out
-= min(pkts_acked
, tp
->fackets_out
);
3166 if (ca_ops
->pkts_acked
)
3167 ca_ops
->pkts_acked(sk
, pkts_acked
, ca_seq_rtt_us
);
3169 } else if (skb
&& rtt_update
&& sack_rtt_us
>= 0 &&
3170 sack_rtt_us
> skb_mstamp_us_delta(&now
, &skb
->skb_mstamp
)) {
3171 /* Do not re-arm RTO if the sack RTT is measured from data sent
3172 * after when the head was last (re)transmitted. Otherwise the
3173 * timeout may continue to extend in loss recovery.
3178 #if FASTRETRANS_DEBUG > 0
3179 WARN_ON((int)tp
->sacked_out
< 0);
3180 WARN_ON((int)tp
->lost_out
< 0);
3181 WARN_ON((int)tp
->retrans_out
< 0);
3182 if (!tp
->packets_out
&& tcp_is_sack(tp
)) {
3183 icsk
= inet_csk(sk
);
3185 pr_debug("Leak l=%u %d\n",
3186 tp
->lost_out
, icsk
->icsk_ca_state
);
3189 if (tp
->sacked_out
) {
3190 pr_debug("Leak s=%u %d\n",
3191 tp
->sacked_out
, icsk
->icsk_ca_state
);
3194 if (tp
->retrans_out
) {
3195 pr_debug("Leak r=%u %d\n",
3196 tp
->retrans_out
, icsk
->icsk_ca_state
);
3197 tp
->retrans_out
= 0;
3204 static void tcp_ack_probe(struct sock
*sk
)
3206 const struct tcp_sock
*tp
= tcp_sk(sk
);
3207 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3209 /* Was it a usable window open? */
3211 if (!after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
, tcp_wnd_end(tp
))) {
3212 icsk
->icsk_backoff
= 0;
3213 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
3214 /* Socket must be waked up by subsequent tcp_data_snd_check().
3215 * This function is not for random using!
3218 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3219 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
3224 static inline bool tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
3226 return !(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
3227 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
;
3230 /* Decide wheather to run the increase function of congestion control. */
3231 static inline bool tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
3233 if (tcp_in_cwnd_reduction(sk
))
3236 /* If reordering is high then always grow cwnd whenever data is
3237 * delivered regardless of its ordering. Otherwise stay conservative
3238 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3239 * new SACK or ECE mark may first advance cwnd here and later reduce
3240 * cwnd in tcp_fastretrans_alert() based on more states.
3242 if (tcp_sk(sk
)->reordering
> sysctl_tcp_reordering
)
3243 return flag
& FLAG_FORWARD_PROGRESS
;
3245 return flag
& FLAG_DATA_ACKED
;
3248 /* Check that window update is acceptable.
3249 * The function assumes that snd_una<=ack<=snd_next.
3251 static inline bool tcp_may_update_window(const struct tcp_sock
*tp
,
3252 const u32 ack
, const u32 ack_seq
,
3255 return after(ack
, tp
->snd_una
) ||
3256 after(ack_seq
, tp
->snd_wl1
) ||
3257 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
);
3260 /* Update our send window.
3262 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3263 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3265 static int tcp_ack_update_window(struct sock
*sk
, const struct sk_buff
*skb
, u32 ack
,
3268 struct tcp_sock
*tp
= tcp_sk(sk
);
3270 u32 nwin
= ntohs(tcp_hdr(skb
)->window
);
3272 if (likely(!tcp_hdr(skb
)->syn
))
3273 nwin
<<= tp
->rx_opt
.snd_wscale
;
3275 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
3276 flag
|= FLAG_WIN_UPDATE
;
3277 tcp_update_wl(tp
, ack_seq
);
3279 if (tp
->snd_wnd
!= nwin
) {
3282 /* Note, it is the only place, where
3283 * fast path is recovered for sending TCP.
3286 tcp_fast_path_check(sk
);
3288 if (nwin
> tp
->max_window
) {
3289 tp
->max_window
= nwin
;
3290 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
3300 /* RFC 5961 7 [ACK Throttling] */
3301 static void tcp_send_challenge_ack(struct sock
*sk
)
3303 /* unprotected vars, we dont care of overwrites */
3304 static u32 challenge_timestamp
;
3305 static unsigned int challenge_count
;
3306 u32 now
= jiffies
/ HZ
;
3308 if (now
!= challenge_timestamp
) {
3309 challenge_timestamp
= now
;
3310 challenge_count
= 0;
3312 if (++challenge_count
<= sysctl_tcp_challenge_ack_limit
) {
3313 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPCHALLENGEACK
);
3318 static void tcp_store_ts_recent(struct tcp_sock
*tp
)
3320 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
3321 tp
->rx_opt
.ts_recent_stamp
= get_seconds();
3324 static void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
3326 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
3327 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3328 * extra check below makes sure this can only happen
3329 * for pure ACK frames. -DaveM
3331 * Not only, also it occurs for expired timestamps.
3334 if (tcp_paws_check(&tp
->rx_opt
, 0))
3335 tcp_store_ts_recent(tp
);
3339 /* This routine deals with acks during a TLP episode.
3340 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3342 static void tcp_process_tlp_ack(struct sock
*sk
, u32 ack
, int flag
)
3344 struct tcp_sock
*tp
= tcp_sk(sk
);
3345 bool is_tlp_dupack
= (ack
== tp
->tlp_high_seq
) &&
3346 !(flag
& (FLAG_SND_UNA_ADVANCED
|
3347 FLAG_NOT_DUP
| FLAG_DATA_SACKED
));
3349 /* Mark the end of TLP episode on receiving TLP dupack or when
3350 * ack is after tlp_high_seq.
3352 if (is_tlp_dupack
) {
3353 tp
->tlp_high_seq
= 0;
3357 if (after(ack
, tp
->tlp_high_seq
)) {
3358 tp
->tlp_high_seq
= 0;
3359 /* Don't reduce cwnd if DSACK arrives for TLP retrans. */
3360 if (!(flag
& FLAG_DSACKING_ACK
)) {
3361 tcp_init_cwnd_reduction(sk
);
3362 tcp_set_ca_state(sk
, TCP_CA_CWR
);
3363 tcp_end_cwnd_reduction(sk
);
3364 tcp_try_keep_open(sk
);
3365 NET_INC_STATS_BH(sock_net(sk
),
3366 LINUX_MIB_TCPLOSSPROBERECOVERY
);
3371 /* This routine deals with incoming acks, but not outgoing ones. */
3372 static int tcp_ack(struct sock
*sk
, const struct sk_buff
*skb
, int flag
)
3374 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3375 struct tcp_sock
*tp
= tcp_sk(sk
);
3376 u32 prior_snd_una
= tp
->snd_una
;
3377 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
3378 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3379 bool is_dupack
= false;
3381 int prior_packets
= tp
->packets_out
;
3382 const int prior_unsacked
= tp
->packets_out
- tp
->sacked_out
;
3383 int acked
= 0; /* Number of packets newly acked */
3384 long sack_rtt_us
= -1L;
3386 /* If the ack is older than previous acks
3387 * then we can probably ignore it.
3389 if (before(ack
, prior_snd_una
)) {
3390 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3391 if (before(ack
, prior_snd_una
- tp
->max_window
)) {
3392 tcp_send_challenge_ack(sk
);
3398 /* If the ack includes data we haven't sent yet, discard
3399 * this segment (RFC793 Section 3.9).
3401 if (after(ack
, tp
->snd_nxt
))
3404 if (icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
||
3405 icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
)
3408 if (after(ack
, prior_snd_una
))
3409 flag
|= FLAG_SND_UNA_ADVANCED
;
3411 prior_fackets
= tp
->fackets_out
;
3413 /* ts_recent update must be made after we are sure that the packet
3416 if (flag
& FLAG_UPDATE_TS_RECENT
)
3417 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
3419 if (!(flag
& FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
3420 /* Window is constant, pure forward advance.
3421 * No more checks are required.
3422 * Note, we use the fact that SND.UNA>=SND.WL2.
3424 tcp_update_wl(tp
, ack_seq
);
3426 flag
|= FLAG_WIN_UPDATE
;
3428 tcp_ca_event(sk
, CA_EVENT_FAST_ACK
);
3430 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPACKS
);
3432 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
3435 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPUREACKS
);
3437 flag
|= tcp_ack_update_window(sk
, skb
, ack
, ack_seq
);
3439 if (TCP_SKB_CB(skb
)->sacked
)
3440 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
,
3443 if (TCP_ECN_rcv_ecn_echo(tp
, tcp_hdr(skb
)))
3446 tcp_ca_event(sk
, CA_EVENT_SLOW_ACK
);
3449 /* We passed data and got it acked, remove any soft error
3450 * log. Something worked...
3452 sk
->sk_err_soft
= 0;
3453 icsk
->icsk_probes_out
= 0;
3454 tp
->rcv_tstamp
= tcp_time_stamp
;
3458 /* See if we can take anything off of the retransmit queue. */
3459 acked
= tp
->packets_out
;
3460 flag
|= tcp_clean_rtx_queue(sk
, prior_fackets
, prior_snd_una
,
3462 acked
-= tp
->packets_out
;
3464 /* Advance cwnd if state allows */
3465 if (tcp_may_raise_cwnd(sk
, flag
))
3466 tcp_cong_avoid(sk
, ack
, acked
);
3468 if (tcp_ack_is_dubious(sk
, flag
)) {
3469 is_dupack
= !(flag
& (FLAG_SND_UNA_ADVANCED
| FLAG_NOT_DUP
));
3470 tcp_fastretrans_alert(sk
, acked
, prior_unsacked
,
3473 if (tp
->tlp_high_seq
)
3474 tcp_process_tlp_ack(sk
, ack
, flag
);
3476 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
& FLAG_NOT_DUP
)) {
3477 struct dst_entry
*dst
= __sk_dst_get(sk
);
3482 if (icsk
->icsk_pending
== ICSK_TIME_RETRANS
)
3483 tcp_schedule_loss_probe(sk
);
3484 tcp_update_pacing_rate(sk
);
3488 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3489 if (flag
& FLAG_DSACKING_ACK
)
3490 tcp_fastretrans_alert(sk
, acked
, prior_unsacked
,
3492 /* If this ack opens up a zero window, clear backoff. It was
3493 * being used to time the probes, and is probably far higher than
3494 * it needs to be for normal retransmission.
3496 if (tcp_send_head(sk
))
3499 if (tp
->tlp_high_seq
)
3500 tcp_process_tlp_ack(sk
, ack
, flag
);
3504 SOCK_DEBUG(sk
, "Ack %u after %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3508 /* If data was SACKed, tag it and see if we should send more data.
3509 * If data was DSACKed, see if we can undo a cwnd reduction.
3511 if (TCP_SKB_CB(skb
)->sacked
) {
3512 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
,
3514 tcp_fastretrans_alert(sk
, acked
, prior_unsacked
,
3518 SOCK_DEBUG(sk
, "Ack %u before %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3522 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3523 * But, this can also be called on packets in the established flow when
3524 * the fast version below fails.
3526 void tcp_parse_options(const struct sk_buff
*skb
,
3527 struct tcp_options_received
*opt_rx
, int estab
,
3528 struct tcp_fastopen_cookie
*foc
)
3530 const unsigned char *ptr
;
3531 const struct tcphdr
*th
= tcp_hdr(skb
);
3532 int length
= (th
->doff
* 4) - sizeof(struct tcphdr
);
3534 ptr
= (const unsigned char *)(th
+ 1);
3535 opt_rx
->saw_tstamp
= 0;
3537 while (length
> 0) {
3538 int opcode
= *ptr
++;
3544 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
3549 if (opsize
< 2) /* "silly options" */
3551 if (opsize
> length
)
3552 return; /* don't parse partial options */
3555 if (opsize
== TCPOLEN_MSS
&& th
->syn
&& !estab
) {
3556 u16 in_mss
= get_unaligned_be16(ptr
);
3558 if (opt_rx
->user_mss
&&
3559 opt_rx
->user_mss
< in_mss
)
3560 in_mss
= opt_rx
->user_mss
;
3561 opt_rx
->mss_clamp
= in_mss
;
3566 if (opsize
== TCPOLEN_WINDOW
&& th
->syn
&&
3567 !estab
&& sysctl_tcp_window_scaling
) {
3568 __u8 snd_wscale
= *(__u8
*)ptr
;
3569 opt_rx
->wscale_ok
= 1;
3570 if (snd_wscale
> 14) {
3571 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3576 opt_rx
->snd_wscale
= snd_wscale
;
3579 case TCPOPT_TIMESTAMP
:
3580 if ((opsize
== TCPOLEN_TIMESTAMP
) &&
3581 ((estab
&& opt_rx
->tstamp_ok
) ||
3582 (!estab
&& sysctl_tcp_timestamps
))) {
3583 opt_rx
->saw_tstamp
= 1;
3584 opt_rx
->rcv_tsval
= get_unaligned_be32(ptr
);
3585 opt_rx
->rcv_tsecr
= get_unaligned_be32(ptr
+ 4);
3588 case TCPOPT_SACK_PERM
:
3589 if (opsize
== TCPOLEN_SACK_PERM
&& th
->syn
&&
3590 !estab
&& sysctl_tcp_sack
) {
3591 opt_rx
->sack_ok
= TCP_SACK_SEEN
;
3592 tcp_sack_reset(opt_rx
);
3597 if ((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
3598 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
3600 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
3603 #ifdef CONFIG_TCP_MD5SIG
3606 * The MD5 Hash has already been
3607 * checked (see tcp_v{4,6}_do_rcv()).
3612 /* Fast Open option shares code 254 using a
3613 * 16 bits magic number. It's valid only in
3614 * SYN or SYN-ACK with an even size.
3616 if (opsize
< TCPOLEN_EXP_FASTOPEN_BASE
||
3617 get_unaligned_be16(ptr
) != TCPOPT_FASTOPEN_MAGIC
||
3618 foc
== NULL
|| !th
->syn
|| (opsize
& 1))
3620 foc
->len
= opsize
- TCPOLEN_EXP_FASTOPEN_BASE
;
3621 if (foc
->len
>= TCP_FASTOPEN_COOKIE_MIN
&&
3622 foc
->len
<= TCP_FASTOPEN_COOKIE_MAX
)
3623 memcpy(foc
->val
, ptr
+ 2, foc
->len
);
3624 else if (foc
->len
!= 0)
3634 EXPORT_SYMBOL(tcp_parse_options
);
3636 static bool tcp_parse_aligned_timestamp(struct tcp_sock
*tp
, const struct tcphdr
*th
)
3638 const __be32
*ptr
= (const __be32
*)(th
+ 1);
3640 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3641 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
3642 tp
->rx_opt
.saw_tstamp
= 1;
3644 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3647 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
) - tp
->tsoffset
;
3649 tp
->rx_opt
.rcv_tsecr
= 0;
3655 /* Fast parse options. This hopes to only see timestamps.
3656 * If it is wrong it falls back on tcp_parse_options().
3658 static bool tcp_fast_parse_options(const struct sk_buff
*skb
,
3659 const struct tcphdr
*th
, struct tcp_sock
*tp
)
3661 /* In the spirit of fast parsing, compare doff directly to constant
3662 * values. Because equality is used, short doff can be ignored here.
3664 if (th
->doff
== (sizeof(*th
) / 4)) {
3665 tp
->rx_opt
.saw_tstamp
= 0;
3667 } else if (tp
->rx_opt
.tstamp_ok
&&
3668 th
->doff
== ((sizeof(*th
) + TCPOLEN_TSTAMP_ALIGNED
) / 4)) {
3669 if (tcp_parse_aligned_timestamp(tp
, th
))
3673 tcp_parse_options(skb
, &tp
->rx_opt
, 1, NULL
);
3674 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
3675 tp
->rx_opt
.rcv_tsecr
-= tp
->tsoffset
;
3680 #ifdef CONFIG_TCP_MD5SIG
3682 * Parse MD5 Signature option
3684 const u8
*tcp_parse_md5sig_option(const struct tcphdr
*th
)
3686 int length
= (th
->doff
<< 2) - sizeof(*th
);
3687 const u8
*ptr
= (const u8
*)(th
+ 1);
3689 /* If the TCP option is too short, we can short cut */
3690 if (length
< TCPOLEN_MD5SIG
)
3693 while (length
> 0) {
3694 int opcode
= *ptr
++;
3705 if (opsize
< 2 || opsize
> length
)
3707 if (opcode
== TCPOPT_MD5SIG
)
3708 return opsize
== TCPOLEN_MD5SIG
? ptr
: NULL
;
3715 EXPORT_SYMBOL(tcp_parse_md5sig_option
);
3718 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3720 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3721 * it can pass through stack. So, the following predicate verifies that
3722 * this segment is not used for anything but congestion avoidance or
3723 * fast retransmit. Moreover, we even are able to eliminate most of such
3724 * second order effects, if we apply some small "replay" window (~RTO)
3725 * to timestamp space.
3727 * All these measures still do not guarantee that we reject wrapped ACKs
3728 * on networks with high bandwidth, when sequence space is recycled fastly,
3729 * but it guarantees that such events will be very rare and do not affect
3730 * connection seriously. This doesn't look nice, but alas, PAWS is really
3733 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3734 * states that events when retransmit arrives after original data are rare.
3735 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3736 * the biggest problem on large power networks even with minor reordering.
3737 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3738 * up to bandwidth of 18Gigabit/sec. 8) ]
3741 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
3743 const struct tcp_sock
*tp
= tcp_sk(sk
);
3744 const struct tcphdr
*th
= tcp_hdr(skb
);
3745 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3746 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3748 return (/* 1. Pure ACK with correct sequence number. */
3749 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
3751 /* 2. ... and duplicate ACK. */
3752 ack
== tp
->snd_una
&&
3754 /* 3. ... and does not update window. */
3755 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
3757 /* 4. ... and sits in replay window. */
3758 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
3761 static inline bool tcp_paws_discard(const struct sock
*sk
,
3762 const struct sk_buff
*skb
)
3764 const struct tcp_sock
*tp
= tcp_sk(sk
);
3766 return !tcp_paws_check(&tp
->rx_opt
, TCP_PAWS_WINDOW
) &&
3767 !tcp_disordered_ack(sk
, skb
);
3770 /* Check segment sequence number for validity.
3772 * Segment controls are considered valid, if the segment
3773 * fits to the window after truncation to the window. Acceptability
3774 * of data (and SYN, FIN, of course) is checked separately.
3775 * See tcp_data_queue(), for example.
3777 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3778 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3779 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3780 * (borrowed from freebsd)
3783 static inline bool tcp_sequence(const struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3785 return !before(end_seq
, tp
->rcv_wup
) &&
3786 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
3789 /* When we get a reset we do this. */
3790 void tcp_reset(struct sock
*sk
)
3792 /* We want the right error as BSD sees it (and indeed as we do). */
3793 switch (sk
->sk_state
) {
3795 sk
->sk_err
= ECONNREFUSED
;
3797 case TCP_CLOSE_WAIT
:
3803 sk
->sk_err
= ECONNRESET
;
3805 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3808 if (!sock_flag(sk
, SOCK_DEAD
))
3809 sk
->sk_error_report(sk
);
3815 * Process the FIN bit. This now behaves as it is supposed to work
3816 * and the FIN takes effect when it is validly part of sequence
3817 * space. Not before when we get holes.
3819 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3820 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3823 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3824 * close and we go into CLOSING (and later onto TIME-WAIT)
3826 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3828 static void tcp_fin(struct sock
*sk
)
3830 struct tcp_sock
*tp
= tcp_sk(sk
);
3831 const struct dst_entry
*dst
;
3833 inet_csk_schedule_ack(sk
);
3835 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
3836 sock_set_flag(sk
, SOCK_DONE
);
3838 switch (sk
->sk_state
) {
3840 case TCP_ESTABLISHED
:
3841 /* Move to CLOSE_WAIT */
3842 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
3843 dst
= __sk_dst_get(sk
);
3844 if (!dst
|| !dst_metric(dst
, RTAX_QUICKACK
))
3845 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
3848 case TCP_CLOSE_WAIT
:
3850 /* Received a retransmission of the FIN, do
3855 /* RFC793: Remain in the LAST-ACK state. */
3859 /* This case occurs when a simultaneous close
3860 * happens, we must ack the received FIN and
3861 * enter the CLOSING state.
3864 tcp_set_state(sk
, TCP_CLOSING
);
3867 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3869 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
3872 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3873 * cases we should never reach this piece of code.
3875 pr_err("%s: Impossible, sk->sk_state=%d\n",
3876 __func__
, sk
->sk_state
);
3880 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3881 * Probably, we should reset in this case. For now drop them.
3883 __skb_queue_purge(&tp
->out_of_order_queue
);
3884 if (tcp_is_sack(tp
))
3885 tcp_sack_reset(&tp
->rx_opt
);
3888 if (!sock_flag(sk
, SOCK_DEAD
)) {
3889 sk
->sk_state_change(sk
);
3891 /* Do not send POLL_HUP for half duplex close. */
3892 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
3893 sk
->sk_state
== TCP_CLOSE
)
3894 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_HUP
);
3896 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_IN
);
3900 static inline bool tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
,
3903 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
3904 if (before(seq
, sp
->start_seq
))
3905 sp
->start_seq
= seq
;
3906 if (after(end_seq
, sp
->end_seq
))
3907 sp
->end_seq
= end_seq
;
3913 static void tcp_dsack_set(struct sock
*sk
, u32 seq
, u32 end_seq
)
3915 struct tcp_sock
*tp
= tcp_sk(sk
);
3917 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
3920 if (before(seq
, tp
->rcv_nxt
))
3921 mib_idx
= LINUX_MIB_TCPDSACKOLDSENT
;
3923 mib_idx
= LINUX_MIB_TCPDSACKOFOSENT
;
3925 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
3927 tp
->rx_opt
.dsack
= 1;
3928 tp
->duplicate_sack
[0].start_seq
= seq
;
3929 tp
->duplicate_sack
[0].end_seq
= end_seq
;
3933 static void tcp_dsack_extend(struct sock
*sk
, u32 seq
, u32 end_seq
)
3935 struct tcp_sock
*tp
= tcp_sk(sk
);
3937 if (!tp
->rx_opt
.dsack
)
3938 tcp_dsack_set(sk
, seq
, end_seq
);
3940 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
3943 static void tcp_send_dupack(struct sock
*sk
, const struct sk_buff
*skb
)
3945 struct tcp_sock
*tp
= tcp_sk(sk
);
3947 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
3948 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
3949 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
3950 tcp_enter_quickack_mode(sk
);
3952 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
3953 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
3955 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
3956 end_seq
= tp
->rcv_nxt
;
3957 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, end_seq
);
3964 /* These routines update the SACK block as out-of-order packets arrive or
3965 * in-order packets close up the sequence space.
3967 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
3970 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3971 struct tcp_sack_block
*swalk
= sp
+ 1;
3973 /* See if the recent change to the first SACK eats into
3974 * or hits the sequence space of other SACK blocks, if so coalesce.
3976 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
;) {
3977 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
3980 /* Zap SWALK, by moving every further SACK up by one slot.
3981 * Decrease num_sacks.
3983 tp
->rx_opt
.num_sacks
--;
3984 for (i
= this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
3988 this_sack
++, swalk
++;
3992 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
3994 struct tcp_sock
*tp
= tcp_sk(sk
);
3995 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3996 int cur_sacks
= tp
->rx_opt
.num_sacks
;
4002 for (this_sack
= 0; this_sack
< cur_sacks
; this_sack
++, sp
++) {
4003 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
4004 /* Rotate this_sack to the first one. */
4005 for (; this_sack
> 0; this_sack
--, sp
--)
4006 swap(*sp
, *(sp
- 1));
4008 tcp_sack_maybe_coalesce(tp
);
4013 /* Could not find an adjacent existing SACK, build a new one,
4014 * put it at the front, and shift everyone else down. We
4015 * always know there is at least one SACK present already here.
4017 * If the sack array is full, forget about the last one.
4019 if (this_sack
>= TCP_NUM_SACKS
) {
4021 tp
->rx_opt
.num_sacks
--;
4024 for (; this_sack
> 0; this_sack
--, sp
--)
4028 /* Build the new head SACK, and we're done. */
4029 sp
->start_seq
= seq
;
4030 sp
->end_seq
= end_seq
;
4031 tp
->rx_opt
.num_sacks
++;
4034 /* RCV.NXT advances, some SACKs should be eaten. */
4036 static void tcp_sack_remove(struct tcp_sock
*tp
)
4038 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4039 int num_sacks
= tp
->rx_opt
.num_sacks
;
4042 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4043 if (skb_queue_empty(&tp
->out_of_order_queue
)) {
4044 tp
->rx_opt
.num_sacks
= 0;
4048 for (this_sack
= 0; this_sack
< num_sacks
;) {
4049 /* Check if the start of the sack is covered by RCV.NXT. */
4050 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
4053 /* RCV.NXT must cover all the block! */
4054 WARN_ON(before(tp
->rcv_nxt
, sp
->end_seq
));
4056 /* Zap this SACK, by moving forward any other SACKS. */
4057 for (i
= this_sack
+1; i
< num_sacks
; i
++)
4058 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
4065 tp
->rx_opt
.num_sacks
= num_sacks
;
4068 /* This one checks to see if we can put data from the
4069 * out_of_order queue into the receive_queue.
4071 static void tcp_ofo_queue(struct sock
*sk
)
4073 struct tcp_sock
*tp
= tcp_sk(sk
);
4074 __u32 dsack_high
= tp
->rcv_nxt
;
4075 struct sk_buff
*skb
;
4077 while ((skb
= skb_peek(&tp
->out_of_order_queue
)) != NULL
) {
4078 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
4081 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
4082 __u32 dsack
= dsack_high
;
4083 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
4084 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
4085 tcp_dsack_extend(sk
, TCP_SKB_CB(skb
)->seq
, dsack
);
4088 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4089 SOCK_DEBUG(sk
, "ofo packet was already received\n");
4090 __skb_unlink(skb
, &tp
->out_of_order_queue
);
4094 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
4095 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4096 TCP_SKB_CB(skb
)->end_seq
);
4098 __skb_unlink(skb
, &tp
->out_of_order_queue
);
4099 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4100 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4101 if (tcp_hdr(skb
)->fin
)
4106 static bool tcp_prune_ofo_queue(struct sock
*sk
);
4107 static int tcp_prune_queue(struct sock
*sk
);
4109 static int tcp_try_rmem_schedule(struct sock
*sk
, struct sk_buff
*skb
,
4112 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
4113 !sk_rmem_schedule(sk
, skb
, size
)) {
4115 if (tcp_prune_queue(sk
) < 0)
4118 if (!sk_rmem_schedule(sk
, skb
, size
)) {
4119 if (!tcp_prune_ofo_queue(sk
))
4122 if (!sk_rmem_schedule(sk
, skb
, size
))
4130 * tcp_try_coalesce - try to merge skb to prior one
4133 * @from: buffer to add in queue
4134 * @fragstolen: pointer to boolean
4136 * Before queueing skb @from after @to, try to merge them
4137 * to reduce overall memory use and queue lengths, if cost is small.
4138 * Packets in ofo or receive queues can stay a long time.
4139 * Better try to coalesce them right now to avoid future collapses.
4140 * Returns true if caller should free @from instead of queueing it
4142 static bool tcp_try_coalesce(struct sock
*sk
,
4144 struct sk_buff
*from
,
4149 *fragstolen
= false;
4151 if (tcp_hdr(from
)->fin
)
4154 /* Its possible this segment overlaps with prior segment in queue */
4155 if (TCP_SKB_CB(from
)->seq
!= TCP_SKB_CB(to
)->end_seq
)
4158 if (!skb_try_coalesce(to
, from
, fragstolen
, &delta
))
4161 atomic_add(delta
, &sk
->sk_rmem_alloc
);
4162 sk_mem_charge(sk
, delta
);
4163 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRCVCOALESCE
);
4164 TCP_SKB_CB(to
)->end_seq
= TCP_SKB_CB(from
)->end_seq
;
4165 TCP_SKB_CB(to
)->ack_seq
= TCP_SKB_CB(from
)->ack_seq
;
4169 static void tcp_data_queue_ofo(struct sock
*sk
, struct sk_buff
*skb
)
4171 struct tcp_sock
*tp
= tcp_sk(sk
);
4172 struct sk_buff
*skb1
;
4175 TCP_ECN_check_ce(tp
, skb
);
4177 if (unlikely(tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))) {
4178 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFODROP
);
4183 /* Disable header prediction. */
4185 inet_csk_schedule_ack(sk
);
4187 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFOQUEUE
);
4188 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
4189 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4191 skb1
= skb_peek_tail(&tp
->out_of_order_queue
);
4193 /* Initial out of order segment, build 1 SACK. */
4194 if (tcp_is_sack(tp
)) {
4195 tp
->rx_opt
.num_sacks
= 1;
4196 tp
->selective_acks
[0].start_seq
= TCP_SKB_CB(skb
)->seq
;
4197 tp
->selective_acks
[0].end_seq
=
4198 TCP_SKB_CB(skb
)->end_seq
;
4200 __skb_queue_head(&tp
->out_of_order_queue
, skb
);
4204 seq
= TCP_SKB_CB(skb
)->seq
;
4205 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4207 if (seq
== TCP_SKB_CB(skb1
)->end_seq
) {
4210 if (!tcp_try_coalesce(sk
, skb1
, skb
, &fragstolen
)) {
4211 __skb_queue_after(&tp
->out_of_order_queue
, skb1
, skb
);
4213 tcp_grow_window(sk
, skb
);
4214 kfree_skb_partial(skb
, fragstolen
);
4218 if (!tp
->rx_opt
.num_sacks
||
4219 tp
->selective_acks
[0].end_seq
!= seq
)
4222 /* Common case: data arrive in order after hole. */
4223 tp
->selective_acks
[0].end_seq
= end_seq
;
4227 /* Find place to insert this segment. */
4229 if (!after(TCP_SKB_CB(skb1
)->seq
, seq
))
4231 if (skb_queue_is_first(&tp
->out_of_order_queue
, skb1
)) {
4235 skb1
= skb_queue_prev(&tp
->out_of_order_queue
, skb1
);
4238 /* Do skb overlap to previous one? */
4239 if (skb1
&& before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4240 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4241 /* All the bits are present. Drop. */
4242 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFOMERGE
);
4245 tcp_dsack_set(sk
, seq
, end_seq
);
4248 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
4249 /* Partial overlap. */
4250 tcp_dsack_set(sk
, seq
,
4251 TCP_SKB_CB(skb1
)->end_seq
);
4253 if (skb_queue_is_first(&tp
->out_of_order_queue
,
4257 skb1
= skb_queue_prev(
4258 &tp
->out_of_order_queue
,
4263 __skb_queue_head(&tp
->out_of_order_queue
, skb
);
4265 __skb_queue_after(&tp
->out_of_order_queue
, skb1
, skb
);
4267 /* And clean segments covered by new one as whole. */
4268 while (!skb_queue_is_last(&tp
->out_of_order_queue
, skb
)) {
4269 skb1
= skb_queue_next(&tp
->out_of_order_queue
, skb
);
4271 if (!after(end_seq
, TCP_SKB_CB(skb1
)->seq
))
4273 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4274 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4278 __skb_unlink(skb1
, &tp
->out_of_order_queue
);
4279 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4280 TCP_SKB_CB(skb1
)->end_seq
);
4281 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFOMERGE
);
4286 if (tcp_is_sack(tp
))
4287 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
4290 tcp_grow_window(sk
, skb
);
4291 skb_set_owner_r(skb
, sk
);
4295 static int __must_check
tcp_queue_rcv(struct sock
*sk
, struct sk_buff
*skb
, int hdrlen
,
4299 struct sk_buff
*tail
= skb_peek_tail(&sk
->sk_receive_queue
);
4301 __skb_pull(skb
, hdrlen
);
4303 tcp_try_coalesce(sk
, tail
, skb
, fragstolen
)) ? 1 : 0;
4304 tcp_sk(sk
)->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4306 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4307 skb_set_owner_r(skb
, sk
);
4312 int tcp_send_rcvq(struct sock
*sk
, struct msghdr
*msg
, size_t size
)
4314 struct sk_buff
*skb
= NULL
;
4321 skb
= alloc_skb(size
+ sizeof(*th
), sk
->sk_allocation
);
4325 if (tcp_try_rmem_schedule(sk
, skb
, size
+ sizeof(*th
)))
4328 th
= (struct tcphdr
*)skb_put(skb
, sizeof(*th
));
4329 skb_reset_transport_header(skb
);
4330 memset(th
, 0, sizeof(*th
));
4332 if (memcpy_fromiovec(skb_put(skb
, size
), msg
->msg_iov
, size
))
4335 TCP_SKB_CB(skb
)->seq
= tcp_sk(sk
)->rcv_nxt
;
4336 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ size
;
4337 TCP_SKB_CB(skb
)->ack_seq
= tcp_sk(sk
)->snd_una
- 1;
4339 if (tcp_queue_rcv(sk
, skb
, sizeof(*th
), &fragstolen
)) {
4340 WARN_ON_ONCE(fragstolen
); /* should not happen */
4351 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
4353 const struct tcphdr
*th
= tcp_hdr(skb
);
4354 struct tcp_sock
*tp
= tcp_sk(sk
);
4356 bool fragstolen
= false;
4358 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
)
4362 __skb_pull(skb
, th
->doff
* 4);
4364 TCP_ECN_accept_cwr(tp
, skb
);
4366 tp
->rx_opt
.dsack
= 0;
4368 /* Queue data for delivery to the user.
4369 * Packets in sequence go to the receive queue.
4370 * Out of sequence packets to the out_of_order_queue.
4372 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
4373 if (tcp_receive_window(tp
) == 0)
4376 /* Ok. In sequence. In window. */
4377 if (tp
->ucopy
.task
== current
&&
4378 tp
->copied_seq
== tp
->rcv_nxt
&& tp
->ucopy
.len
&&
4379 sock_owned_by_user(sk
) && !tp
->urg_data
) {
4380 int chunk
= min_t(unsigned int, skb
->len
,
4383 __set_current_state(TASK_RUNNING
);
4386 if (!skb_copy_datagram_iovec(skb
, 0, tp
->ucopy
.iov
, chunk
)) {
4387 tp
->ucopy
.len
-= chunk
;
4388 tp
->copied_seq
+= chunk
;
4389 eaten
= (chunk
== skb
->len
);
4390 tcp_rcv_space_adjust(sk
);
4398 tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))
4401 eaten
= tcp_queue_rcv(sk
, skb
, 0, &fragstolen
);
4403 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4405 tcp_event_data_recv(sk
, skb
);
4409 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4412 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4413 * gap in queue is filled.
4415 if (skb_queue_empty(&tp
->out_of_order_queue
))
4416 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
4419 if (tp
->rx_opt
.num_sacks
)
4420 tcp_sack_remove(tp
);
4422 tcp_fast_path_check(sk
);
4425 kfree_skb_partial(skb
, fragstolen
);
4426 if (!sock_flag(sk
, SOCK_DEAD
))
4427 sk
->sk_data_ready(sk
);
4431 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4432 /* A retransmit, 2nd most common case. Force an immediate ack. */
4433 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4434 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4437 tcp_enter_quickack_mode(sk
);
4438 inet_csk_schedule_ack(sk
);
4444 /* Out of window. F.e. zero window probe. */
4445 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
4448 tcp_enter_quickack_mode(sk
);
4450 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4451 /* Partial packet, seq < rcv_next < end_seq */
4452 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
4453 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4454 TCP_SKB_CB(skb
)->end_seq
);
4456 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
4458 /* If window is closed, drop tail of packet. But after
4459 * remembering D-SACK for its head made in previous line.
4461 if (!tcp_receive_window(tp
))
4466 tcp_data_queue_ofo(sk
, skb
);
4469 static struct sk_buff
*tcp_collapse_one(struct sock
*sk
, struct sk_buff
*skb
,
4470 struct sk_buff_head
*list
)
4472 struct sk_buff
*next
= NULL
;
4474 if (!skb_queue_is_last(list
, skb
))
4475 next
= skb_queue_next(list
, skb
);
4477 __skb_unlink(skb
, list
);
4479 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRCVCOLLAPSED
);
4484 /* Collapse contiguous sequence of skbs head..tail with
4485 * sequence numbers start..end.
4487 * If tail is NULL, this means until the end of the list.
4489 * Segments with FIN/SYN are not collapsed (only because this
4493 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
,
4494 struct sk_buff
*head
, struct sk_buff
*tail
,
4497 struct sk_buff
*skb
, *n
;
4500 /* First, check that queue is collapsible and find
4501 * the point where collapsing can be useful. */
4505 skb_queue_walk_from_safe(list
, skb
, n
) {
4508 /* No new bits? It is possible on ofo queue. */
4509 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4510 skb
= tcp_collapse_one(sk
, skb
, list
);
4516 /* The first skb to collapse is:
4518 * - bloated or contains data before "start" or
4519 * overlaps to the next one.
4521 if (!tcp_hdr(skb
)->syn
&& !tcp_hdr(skb
)->fin
&&
4522 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
4523 before(TCP_SKB_CB(skb
)->seq
, start
))) {
4524 end_of_skbs
= false;
4528 if (!skb_queue_is_last(list
, skb
)) {
4529 struct sk_buff
*next
= skb_queue_next(list
, skb
);
4531 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(next
)->seq
) {
4532 end_of_skbs
= false;
4537 /* Decided to skip this, advance start seq. */
4538 start
= TCP_SKB_CB(skb
)->end_seq
;
4540 if (end_of_skbs
|| tcp_hdr(skb
)->syn
|| tcp_hdr(skb
)->fin
)
4543 while (before(start
, end
)) {
4544 struct sk_buff
*nskb
;
4545 unsigned int header
= skb_headroom(skb
);
4546 int copy
= SKB_MAX_ORDER(header
, 0);
4548 /* Too big header? This can happen with IPv6. */
4551 if (end
- start
< copy
)
4553 nskb
= alloc_skb(copy
+ header
, GFP_ATOMIC
);
4557 skb_set_mac_header(nskb
, skb_mac_header(skb
) - skb
->head
);
4558 skb_set_network_header(nskb
, (skb_network_header(skb
) -
4560 skb_set_transport_header(nskb
, (skb_transport_header(skb
) -
4562 skb_reserve(nskb
, header
);
4563 memcpy(nskb
->head
, skb
->head
, header
);
4564 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
4565 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
4566 __skb_queue_before(list
, skb
, nskb
);
4567 skb_set_owner_r(nskb
, sk
);
4569 /* Copy data, releasing collapsed skbs. */
4571 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
4572 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
4576 size
= min(copy
, size
);
4577 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
4579 TCP_SKB_CB(nskb
)->end_seq
+= size
;
4583 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4584 skb
= tcp_collapse_one(sk
, skb
, list
);
4587 tcp_hdr(skb
)->syn
||
4595 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4596 * and tcp_collapse() them until all the queue is collapsed.
4598 static void tcp_collapse_ofo_queue(struct sock
*sk
)
4600 struct tcp_sock
*tp
= tcp_sk(sk
);
4601 struct sk_buff
*skb
= skb_peek(&tp
->out_of_order_queue
);
4602 struct sk_buff
*head
;
4608 start
= TCP_SKB_CB(skb
)->seq
;
4609 end
= TCP_SKB_CB(skb
)->end_seq
;
4613 struct sk_buff
*next
= NULL
;
4615 if (!skb_queue_is_last(&tp
->out_of_order_queue
, skb
))
4616 next
= skb_queue_next(&tp
->out_of_order_queue
, skb
);
4619 /* Segment is terminated when we see gap or when
4620 * we are at the end of all the queue. */
4622 after(TCP_SKB_CB(skb
)->seq
, end
) ||
4623 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
4624 tcp_collapse(sk
, &tp
->out_of_order_queue
,
4625 head
, skb
, start
, end
);
4629 /* Start new segment */
4630 start
= TCP_SKB_CB(skb
)->seq
;
4631 end
= TCP_SKB_CB(skb
)->end_seq
;
4633 if (before(TCP_SKB_CB(skb
)->seq
, start
))
4634 start
= TCP_SKB_CB(skb
)->seq
;
4635 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
4636 end
= TCP_SKB_CB(skb
)->end_seq
;
4642 * Purge the out-of-order queue.
4643 * Return true if queue was pruned.
4645 static bool tcp_prune_ofo_queue(struct sock
*sk
)
4647 struct tcp_sock
*tp
= tcp_sk(sk
);
4650 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4651 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_OFOPRUNED
);
4652 __skb_queue_purge(&tp
->out_of_order_queue
);
4654 /* Reset SACK state. A conforming SACK implementation will
4655 * do the same at a timeout based retransmit. When a connection
4656 * is in a sad state like this, we care only about integrity
4657 * of the connection not performance.
4659 if (tp
->rx_opt
.sack_ok
)
4660 tcp_sack_reset(&tp
->rx_opt
);
4667 /* Reduce allocated memory if we can, trying to get
4668 * the socket within its memory limits again.
4670 * Return less than zero if we should start dropping frames
4671 * until the socket owning process reads some of the data
4672 * to stabilize the situation.
4674 static int tcp_prune_queue(struct sock
*sk
)
4676 struct tcp_sock
*tp
= tcp_sk(sk
);
4678 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
4680 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PRUNECALLED
);
4682 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
4683 tcp_clamp_window(sk
);
4684 else if (sk_under_memory_pressure(sk
))
4685 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
4687 tcp_collapse_ofo_queue(sk
);
4688 if (!skb_queue_empty(&sk
->sk_receive_queue
))
4689 tcp_collapse(sk
, &sk
->sk_receive_queue
,
4690 skb_peek(&sk
->sk_receive_queue
),
4692 tp
->copied_seq
, tp
->rcv_nxt
);
4695 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4698 /* Collapsing did not help, destructive actions follow.
4699 * This must not ever occur. */
4701 tcp_prune_ofo_queue(sk
);
4703 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4706 /* If we are really being abused, tell the caller to silently
4707 * drop receive data on the floor. It will get retransmitted
4708 * and hopefully then we'll have sufficient space.
4710 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_RCVPRUNED
);
4712 /* Massive buffer overcommit. */
4717 static bool tcp_should_expand_sndbuf(const struct sock
*sk
)
4719 const struct tcp_sock
*tp
= tcp_sk(sk
);
4721 /* If the user specified a specific send buffer setting, do
4724 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
4727 /* If we are under global TCP memory pressure, do not expand. */
4728 if (sk_under_memory_pressure(sk
))
4731 /* If we are under soft global TCP memory pressure, do not expand. */
4732 if (sk_memory_allocated(sk
) >= sk_prot_mem_limits(sk
, 0))
4735 /* If we filled the congestion window, do not expand. */
4736 if (tp
->packets_out
>= tp
->snd_cwnd
)
4742 /* When incoming ACK allowed to free some skb from write_queue,
4743 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4744 * on the exit from tcp input handler.
4746 * PROBLEM: sndbuf expansion does not work well with largesend.
4748 static void tcp_new_space(struct sock
*sk
)
4750 struct tcp_sock
*tp
= tcp_sk(sk
);
4752 if (tcp_should_expand_sndbuf(sk
)) {
4753 tcp_sndbuf_expand(sk
);
4754 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4757 sk
->sk_write_space(sk
);
4760 static void tcp_check_space(struct sock
*sk
)
4762 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
4763 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
4764 if (sk
->sk_socket
&&
4765 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
4770 static inline void tcp_data_snd_check(struct sock
*sk
)
4772 tcp_push_pending_frames(sk
);
4773 tcp_check_space(sk
);
4777 * Check if sending an ack is needed.
4779 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
4781 struct tcp_sock
*tp
= tcp_sk(sk
);
4783 /* More than one full frame received... */
4784 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
&&
4785 /* ... and right edge of window advances far enough.
4786 * (tcp_recvmsg() will send ACK otherwise). Or...
4788 __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
4789 /* We ACK each frame or... */
4790 tcp_in_quickack_mode(sk
) ||
4791 /* We have out of order data. */
4792 (ofo_possible
&& skb_peek(&tp
->out_of_order_queue
))) {
4793 /* Then ack it now */
4796 /* Else, send delayed ack. */
4797 tcp_send_delayed_ack(sk
);
4801 static inline void tcp_ack_snd_check(struct sock
*sk
)
4803 if (!inet_csk_ack_scheduled(sk
)) {
4804 /* We sent a data segment already. */
4807 __tcp_ack_snd_check(sk
, 1);
4811 * This routine is only called when we have urgent data
4812 * signaled. Its the 'slow' part of tcp_urg. It could be
4813 * moved inline now as tcp_urg is only called from one
4814 * place. We handle URGent data wrong. We have to - as
4815 * BSD still doesn't use the correction from RFC961.
4816 * For 1003.1g we should support a new option TCP_STDURG to permit
4817 * either form (or just set the sysctl tcp_stdurg).
4820 static void tcp_check_urg(struct sock
*sk
, const struct tcphdr
*th
)
4822 struct tcp_sock
*tp
= tcp_sk(sk
);
4823 u32 ptr
= ntohs(th
->urg_ptr
);
4825 if (ptr
&& !sysctl_tcp_stdurg
)
4827 ptr
+= ntohl(th
->seq
);
4829 /* Ignore urgent data that we've already seen and read. */
4830 if (after(tp
->copied_seq
, ptr
))
4833 /* Do not replay urg ptr.
4835 * NOTE: interesting situation not covered by specs.
4836 * Misbehaving sender may send urg ptr, pointing to segment,
4837 * which we already have in ofo queue. We are not able to fetch
4838 * such data and will stay in TCP_URG_NOTYET until will be eaten
4839 * by recvmsg(). Seems, we are not obliged to handle such wicked
4840 * situations. But it is worth to think about possibility of some
4841 * DoSes using some hypothetical application level deadlock.
4843 if (before(ptr
, tp
->rcv_nxt
))
4846 /* Do we already have a newer (or duplicate) urgent pointer? */
4847 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
4850 /* Tell the world about our new urgent pointer. */
4853 /* We may be adding urgent data when the last byte read was
4854 * urgent. To do this requires some care. We cannot just ignore
4855 * tp->copied_seq since we would read the last urgent byte again
4856 * as data, nor can we alter copied_seq until this data arrives
4857 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4859 * NOTE. Double Dutch. Rendering to plain English: author of comment
4860 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4861 * and expect that both A and B disappear from stream. This is _wrong_.
4862 * Though this happens in BSD with high probability, this is occasional.
4863 * Any application relying on this is buggy. Note also, that fix "works"
4864 * only in this artificial test. Insert some normal data between A and B and we will
4865 * decline of BSD again. Verdict: it is better to remove to trap
4868 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
4869 !sock_flag(sk
, SOCK_URGINLINE
) && tp
->copied_seq
!= tp
->rcv_nxt
) {
4870 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
4872 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4873 __skb_unlink(skb
, &sk
->sk_receive_queue
);
4878 tp
->urg_data
= TCP_URG_NOTYET
;
4881 /* Disable header prediction. */
4885 /* This is the 'fast' part of urgent handling. */
4886 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, const struct tcphdr
*th
)
4888 struct tcp_sock
*tp
= tcp_sk(sk
);
4890 /* Check if we get a new urgent pointer - normally not. */
4892 tcp_check_urg(sk
, th
);
4894 /* Do we wait for any urgent data? - normally not... */
4895 if (tp
->urg_data
== TCP_URG_NOTYET
) {
4896 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
4899 /* Is the urgent pointer pointing into this packet? */
4900 if (ptr
< skb
->len
) {
4902 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
4904 tp
->urg_data
= TCP_URG_VALID
| tmp
;
4905 if (!sock_flag(sk
, SOCK_DEAD
))
4906 sk
->sk_data_ready(sk
);
4911 static int tcp_copy_to_iovec(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
4913 struct tcp_sock
*tp
= tcp_sk(sk
);
4914 int chunk
= skb
->len
- hlen
;
4918 if (skb_csum_unnecessary(skb
))
4919 err
= skb_copy_datagram_iovec(skb
, hlen
, tp
->ucopy
.iov
, chunk
);
4921 err
= skb_copy_and_csum_datagram_iovec(skb
, hlen
,
4925 tp
->ucopy
.len
-= chunk
;
4926 tp
->copied_seq
+= chunk
;
4927 tcp_rcv_space_adjust(sk
);
4934 static __sum16
__tcp_checksum_complete_user(struct sock
*sk
,
4935 struct sk_buff
*skb
)
4939 if (sock_owned_by_user(sk
)) {
4941 result
= __tcp_checksum_complete(skb
);
4944 result
= __tcp_checksum_complete(skb
);
4949 static inline bool tcp_checksum_complete_user(struct sock
*sk
,
4950 struct sk_buff
*skb
)
4952 return !skb_csum_unnecessary(skb
) &&
4953 __tcp_checksum_complete_user(sk
, skb
);
4956 #ifdef CONFIG_NET_DMA
4957 static bool tcp_dma_try_early_copy(struct sock
*sk
, struct sk_buff
*skb
,
4960 struct tcp_sock
*tp
= tcp_sk(sk
);
4961 int chunk
= skb
->len
- hlen
;
4963 bool copied_early
= false;
4965 if (tp
->ucopy
.wakeup
)
4968 if (!tp
->ucopy
.dma_chan
&& tp
->ucopy
.pinned_list
)
4969 tp
->ucopy
.dma_chan
= net_dma_find_channel();
4971 if (tp
->ucopy
.dma_chan
&& skb_csum_unnecessary(skb
)) {
4973 dma_cookie
= dma_skb_copy_datagram_iovec(tp
->ucopy
.dma_chan
,
4975 tp
->ucopy
.iov
, chunk
,
4976 tp
->ucopy
.pinned_list
);
4981 tp
->ucopy
.dma_cookie
= dma_cookie
;
4982 copied_early
= true;
4984 tp
->ucopy
.len
-= chunk
;
4985 tp
->copied_seq
+= chunk
;
4986 tcp_rcv_space_adjust(sk
);
4988 if ((tp
->ucopy
.len
== 0) ||
4989 (tcp_flag_word(tcp_hdr(skb
)) & TCP_FLAG_PSH
) ||
4990 (atomic_read(&sk
->sk_rmem_alloc
) > (sk
->sk_rcvbuf
>> 1))) {
4991 tp
->ucopy
.wakeup
= 1;
4992 sk
->sk_data_ready(sk
);
4994 } else if (chunk
> 0) {
4995 tp
->ucopy
.wakeup
= 1;
4996 sk
->sk_data_ready(sk
);
4999 return copied_early
;
5001 #endif /* CONFIG_NET_DMA */
5003 /* Does PAWS and seqno based validation of an incoming segment, flags will
5004 * play significant role here.
5006 static bool tcp_validate_incoming(struct sock
*sk
, struct sk_buff
*skb
,
5007 const struct tcphdr
*th
, int syn_inerr
)
5009 struct tcp_sock
*tp
= tcp_sk(sk
);
5011 /* RFC1323: H1. Apply PAWS check first. */
5012 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
5013 tcp_paws_discard(sk
, skb
)) {
5015 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSESTABREJECTED
);
5016 tcp_send_dupack(sk
, skb
);
5019 /* Reset is accepted even if it did not pass PAWS. */
5022 /* Step 1: check sequence number */
5023 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5024 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5025 * (RST) segments are validated by checking their SEQ-fields."
5026 * And page 69: "If an incoming segment is not acceptable,
5027 * an acknowledgment should be sent in reply (unless the RST
5028 * bit is set, if so drop the segment and return)".
5033 tcp_send_dupack(sk
, skb
);
5038 /* Step 2: check RST bit */
5041 * If sequence number exactly matches RCV.NXT, then
5042 * RESET the connection
5044 * Send a challenge ACK
5046 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
)
5049 tcp_send_challenge_ack(sk
);
5053 /* step 3: check security and precedence [ignored] */
5055 /* step 4: Check for a SYN
5056 * RFC 5691 4.2 : Send a challenge ack
5061 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5062 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSYNCHALLENGE
);
5063 tcp_send_challenge_ack(sk
);
5075 * TCP receive function for the ESTABLISHED state.
5077 * It is split into a fast path and a slow path. The fast path is
5079 * - A zero window was announced from us - zero window probing
5080 * is only handled properly in the slow path.
5081 * - Out of order segments arrived.
5082 * - Urgent data is expected.
5083 * - There is no buffer space left
5084 * - Unexpected TCP flags/window values/header lengths are received
5085 * (detected by checking the TCP header against pred_flags)
5086 * - Data is sent in both directions. Fast path only supports pure senders
5087 * or pure receivers (this means either the sequence number or the ack
5088 * value must stay constant)
5089 * - Unexpected TCP option.
5091 * When these conditions are not satisfied it drops into a standard
5092 * receive procedure patterned after RFC793 to handle all cases.
5093 * The first three cases are guaranteed by proper pred_flags setting,
5094 * the rest is checked inline. Fast processing is turned on in
5095 * tcp_data_queue when everything is OK.
5097 void tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
5098 const struct tcphdr
*th
, unsigned int len
)
5100 struct tcp_sock
*tp
= tcp_sk(sk
);
5102 if (unlikely(sk
->sk_rx_dst
== NULL
))
5103 inet_csk(sk
)->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
5105 * Header prediction.
5106 * The code loosely follows the one in the famous
5107 * "30 instruction TCP receive" Van Jacobson mail.
5109 * Van's trick is to deposit buffers into socket queue
5110 * on a device interrupt, to call tcp_recv function
5111 * on the receive process context and checksum and copy
5112 * the buffer to user space. smart...
5114 * Our current scheme is not silly either but we take the
5115 * extra cost of the net_bh soft interrupt processing...
5116 * We do checksum and copy also but from device to kernel.
5119 tp
->rx_opt
.saw_tstamp
= 0;
5121 /* pred_flags is 0xS?10 << 16 + snd_wnd
5122 * if header_prediction is to be made
5123 * 'S' will always be tp->tcp_header_len >> 2
5124 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5125 * turn it off (when there are holes in the receive
5126 * space for instance)
5127 * PSH flag is ignored.
5130 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
5131 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
&&
5132 !after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
)) {
5133 int tcp_header_len
= tp
->tcp_header_len
;
5135 /* Timestamp header prediction: tcp_header_len
5136 * is automatically equal to th->doff*4 due to pred_flags
5140 /* Check timestamp */
5141 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
5142 /* No? Slow path! */
5143 if (!tcp_parse_aligned_timestamp(tp
, th
))
5146 /* If PAWS failed, check it more carefully in slow path */
5147 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
5150 /* DO NOT update ts_recent here, if checksum fails
5151 * and timestamp was corrupted part, it will result
5152 * in a hung connection since we will drop all
5153 * future packets due to the PAWS test.
5157 if (len
<= tcp_header_len
) {
5158 /* Bulk data transfer: sender */
5159 if (len
== tcp_header_len
) {
5160 /* Predicted packet is in window by definition.
5161 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5162 * Hence, check seq<=rcv_wup reduces to:
5164 if (tcp_header_len
==
5165 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5166 tp
->rcv_nxt
== tp
->rcv_wup
)
5167 tcp_store_ts_recent(tp
);
5169 /* We know that such packets are checksummed
5172 tcp_ack(sk
, skb
, 0);
5174 tcp_data_snd_check(sk
);
5176 } else { /* Header too small */
5177 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5182 int copied_early
= 0;
5183 bool fragstolen
= false;
5185 if (tp
->copied_seq
== tp
->rcv_nxt
&&
5186 len
- tcp_header_len
<= tp
->ucopy
.len
) {
5187 #ifdef CONFIG_NET_DMA
5188 if (tp
->ucopy
.task
== current
&&
5189 sock_owned_by_user(sk
) &&
5190 tcp_dma_try_early_copy(sk
, skb
, tcp_header_len
)) {
5195 if (tp
->ucopy
.task
== current
&&
5196 sock_owned_by_user(sk
) && !copied_early
) {
5197 __set_current_state(TASK_RUNNING
);
5199 if (!tcp_copy_to_iovec(sk
, skb
, tcp_header_len
))
5203 /* Predicted packet is in window by definition.
5204 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5205 * Hence, check seq<=rcv_wup reduces to:
5207 if (tcp_header_len
==
5208 (sizeof(struct tcphdr
) +
5209 TCPOLEN_TSTAMP_ALIGNED
) &&
5210 tp
->rcv_nxt
== tp
->rcv_wup
)
5211 tcp_store_ts_recent(tp
);
5213 tcp_rcv_rtt_measure_ts(sk
, skb
);
5215 __skb_pull(skb
, tcp_header_len
);
5216 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
5217 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITSTOUSER
);
5220 tcp_cleanup_rbuf(sk
, skb
->len
);
5223 if (tcp_checksum_complete_user(sk
, skb
))
5226 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
5229 /* Predicted packet is in window by definition.
5230 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5231 * Hence, check seq<=rcv_wup reduces to:
5233 if (tcp_header_len
==
5234 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5235 tp
->rcv_nxt
== tp
->rcv_wup
)
5236 tcp_store_ts_recent(tp
);
5238 tcp_rcv_rtt_measure_ts(sk
, skb
);
5240 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITS
);
5242 /* Bulk data transfer: receiver */
5243 eaten
= tcp_queue_rcv(sk
, skb
, tcp_header_len
,
5247 tcp_event_data_recv(sk
, skb
);
5249 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
5250 /* Well, only one small jumplet in fast path... */
5251 tcp_ack(sk
, skb
, FLAG_DATA
);
5252 tcp_data_snd_check(sk
);
5253 if (!inet_csk_ack_scheduled(sk
))
5257 if (!copied_early
|| tp
->rcv_nxt
!= tp
->rcv_wup
)
5258 __tcp_ack_snd_check(sk
, 0);
5260 #ifdef CONFIG_NET_DMA
5262 __skb_queue_tail(&sk
->sk_async_wait_queue
, skb
);
5266 kfree_skb_partial(skb
, fragstolen
);
5267 sk
->sk_data_ready(sk
);
5273 if (len
< (th
->doff
<< 2) || tcp_checksum_complete_user(sk
, skb
))
5276 if (!th
->ack
&& !th
->rst
)
5280 * Standard slow path.
5283 if (!tcp_validate_incoming(sk
, skb
, th
, 1))
5287 if (tcp_ack(sk
, skb
, FLAG_SLOWPATH
| FLAG_UPDATE_TS_RECENT
) < 0)
5290 tcp_rcv_rtt_measure_ts(sk
, skb
);
5292 /* Process urgent data. */
5293 tcp_urg(sk
, skb
, th
);
5295 /* step 7: process the segment text */
5296 tcp_data_queue(sk
, skb
);
5298 tcp_data_snd_check(sk
);
5299 tcp_ack_snd_check(sk
);
5303 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_CSUMERRORS
);
5304 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5309 EXPORT_SYMBOL(tcp_rcv_established
);
5311 void tcp_finish_connect(struct sock
*sk
, struct sk_buff
*skb
)
5313 struct tcp_sock
*tp
= tcp_sk(sk
);
5314 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5316 tcp_set_state(sk
, TCP_ESTABLISHED
);
5319 icsk
->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
5320 security_inet_conn_established(sk
, skb
);
5323 /* Make sure socket is routed, for correct metrics. */
5324 icsk
->icsk_af_ops
->rebuild_header(sk
);
5326 tcp_init_metrics(sk
);
5328 tcp_init_congestion_control(sk
);
5330 /* Prevent spurious tcp_cwnd_restart() on first data
5333 tp
->lsndtime
= tcp_time_stamp
;
5335 tcp_init_buffer_space(sk
);
5337 if (sock_flag(sk
, SOCK_KEEPOPEN
))
5338 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
5340 if (!tp
->rx_opt
.snd_wscale
)
5341 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
5345 if (!sock_flag(sk
, SOCK_DEAD
)) {
5346 sk
->sk_state_change(sk
);
5347 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5351 static bool tcp_rcv_fastopen_synack(struct sock
*sk
, struct sk_buff
*synack
,
5352 struct tcp_fastopen_cookie
*cookie
)
5354 struct tcp_sock
*tp
= tcp_sk(sk
);
5355 struct sk_buff
*data
= tp
->syn_data
? tcp_write_queue_head(sk
) : NULL
;
5356 u16 mss
= tp
->rx_opt
.mss_clamp
;
5359 if (mss
== tp
->rx_opt
.user_mss
) {
5360 struct tcp_options_received opt
;
5362 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5363 tcp_clear_options(&opt
);
5364 opt
.user_mss
= opt
.mss_clamp
= 0;
5365 tcp_parse_options(synack
, &opt
, 0, NULL
);
5366 mss
= opt
.mss_clamp
;
5369 if (!tp
->syn_fastopen
) /* Ignore an unsolicited cookie */
5372 /* The SYN-ACK neither has cookie nor acknowledges the data. Presumably
5373 * the remote receives only the retransmitted (regular) SYNs: either
5374 * the original SYN-data or the corresponding SYN-ACK is lost.
5376 syn_drop
= (cookie
->len
<= 0 && data
&& tp
->total_retrans
);
5378 tcp_fastopen_cache_set(sk
, mss
, cookie
, syn_drop
);
5380 if (data
) { /* Retransmit unacked data in SYN */
5381 tcp_for_write_queue_from(data
, sk
) {
5382 if (data
== tcp_send_head(sk
) ||
5383 __tcp_retransmit_skb(sk
, data
))
5387 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPFASTOPENACTIVEFAIL
);
5390 tp
->syn_data_acked
= tp
->syn_data
;
5391 if (tp
->syn_data_acked
)
5392 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPFASTOPENACTIVE
);
5396 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5397 const struct tcphdr
*th
, unsigned int len
)
5399 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5400 struct tcp_sock
*tp
= tcp_sk(sk
);
5401 struct tcp_fastopen_cookie foc
= { .len
= -1 };
5402 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
5404 tcp_parse_options(skb
, &tp
->rx_opt
, 0, &foc
);
5405 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
5406 tp
->rx_opt
.rcv_tsecr
-= tp
->tsoffset
;
5410 * "If the state is SYN-SENT then
5411 * first check the ACK bit
5412 * If the ACK bit is set
5413 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5414 * a reset (unless the RST bit is set, if so drop
5415 * the segment and return)"
5417 if (!after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_una
) ||
5418 after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
))
5419 goto reset_and_undo
;
5421 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
5422 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
5424 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSACTIVEREJECTED
);
5425 goto reset_and_undo
;
5428 /* Now ACK is acceptable.
5430 * "If the RST bit is set
5431 * If the ACK was acceptable then signal the user "error:
5432 * connection reset", drop the segment, enter CLOSED state,
5433 * delete TCB, and return."
5442 * "fifth, if neither of the SYN or RST bits is set then
5443 * drop the segment and return."
5449 goto discard_and_undo
;
5452 * "If the SYN bit is on ...
5453 * are acceptable then ...
5454 * (our SYN has been ACKed), change the connection
5455 * state to ESTABLISHED..."
5458 TCP_ECN_rcv_synack(tp
, th
);
5460 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5461 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
5463 /* Ok.. it's good. Set up sequence numbers and
5464 * move to established.
5466 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5467 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5469 /* RFC1323: The window in SYN & SYN/ACK segments is
5472 tp
->snd_wnd
= ntohs(th
->window
);
5474 if (!tp
->rx_opt
.wscale_ok
) {
5475 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
5476 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
5479 if (tp
->rx_opt
.saw_tstamp
) {
5480 tp
->rx_opt
.tstamp_ok
= 1;
5481 tp
->tcp_header_len
=
5482 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5483 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5484 tcp_store_ts_recent(tp
);
5486 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5489 if (tcp_is_sack(tp
) && sysctl_tcp_fack
)
5490 tcp_enable_fack(tp
);
5493 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5494 tcp_initialize_rcv_mss(sk
);
5496 /* Remember, tcp_poll() does not lock socket!
5497 * Change state from SYN-SENT only after copied_seq
5498 * is initialized. */
5499 tp
->copied_seq
= tp
->rcv_nxt
;
5503 tcp_finish_connect(sk
, skb
);
5505 if ((tp
->syn_fastopen
|| tp
->syn_data
) &&
5506 tcp_rcv_fastopen_synack(sk
, skb
, &foc
))
5509 if (sk
->sk_write_pending
||
5510 icsk
->icsk_accept_queue
.rskq_defer_accept
||
5511 icsk
->icsk_ack
.pingpong
) {
5512 /* Save one ACK. Data will be ready after
5513 * several ticks, if write_pending is set.
5515 * It may be deleted, but with this feature tcpdumps
5516 * look so _wonderfully_ clever, that I was not able
5517 * to stand against the temptation 8) --ANK
5519 inet_csk_schedule_ack(sk
);
5520 icsk
->icsk_ack
.lrcvtime
= tcp_time_stamp
;
5521 tcp_enter_quickack_mode(sk
);
5522 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
5523 TCP_DELACK_MAX
, TCP_RTO_MAX
);
5534 /* No ACK in the segment */
5538 * "If the RST bit is set
5540 * Otherwise (no ACK) drop the segment and return."
5543 goto discard_and_undo
;
5547 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&&
5548 tcp_paws_reject(&tp
->rx_opt
, 0))
5549 goto discard_and_undo
;
5552 /* We see SYN without ACK. It is attempt of
5553 * simultaneous connect with crossed SYNs.
5554 * Particularly, it can be connect to self.
5556 tcp_set_state(sk
, TCP_SYN_RECV
);
5558 if (tp
->rx_opt
.saw_tstamp
) {
5559 tp
->rx_opt
.tstamp_ok
= 1;
5560 tcp_store_ts_recent(tp
);
5561 tp
->tcp_header_len
=
5562 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5564 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5567 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5568 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5570 /* RFC1323: The window in SYN & SYN/ACK segments is
5573 tp
->snd_wnd
= ntohs(th
->window
);
5574 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
5575 tp
->max_window
= tp
->snd_wnd
;
5577 TCP_ECN_rcv_syn(tp
, th
);
5580 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5581 tcp_initialize_rcv_mss(sk
);
5583 tcp_send_synack(sk
);
5585 /* Note, we could accept data and URG from this segment.
5586 * There are no obstacles to make this (except that we must
5587 * either change tcp_recvmsg() to prevent it from returning data
5588 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5590 * However, if we ignore data in ACKless segments sometimes,
5591 * we have no reasons to accept it sometimes.
5592 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5593 * is not flawless. So, discard packet for sanity.
5594 * Uncomment this return to process the data.
5601 /* "fifth, if neither of the SYN or RST bits is set then
5602 * drop the segment and return."
5606 tcp_clear_options(&tp
->rx_opt
);
5607 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5611 tcp_clear_options(&tp
->rx_opt
);
5612 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5617 * This function implements the receiving procedure of RFC 793 for
5618 * all states except ESTABLISHED and TIME_WAIT.
5619 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5620 * address independent.
5623 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5624 const struct tcphdr
*th
, unsigned int len
)
5626 struct tcp_sock
*tp
= tcp_sk(sk
);
5627 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5628 struct request_sock
*req
;
5633 tp
->rx_opt
.saw_tstamp
= 0;
5635 switch (sk
->sk_state
) {
5649 if (icsk
->icsk_af_ops
->conn_request(sk
, skb
) < 0)
5652 /* Now we have several options: In theory there is
5653 * nothing else in the frame. KA9Q has an option to
5654 * send data with the syn, BSD accepts data with the
5655 * syn up to the [to be] advertised window and
5656 * Solaris 2.1 gives you a protocol error. For now
5657 * we just ignore it, that fits the spec precisely
5658 * and avoids incompatibilities. It would be nice in
5659 * future to drop through and process the data.
5661 * Now that TTCP is starting to be used we ought to
5663 * But, this leaves one open to an easy denial of
5664 * service attack, and SYN cookies can't defend
5665 * against this problem. So, we drop the data
5666 * in the interest of security over speed unless
5667 * it's still in use.
5675 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
, len
);
5679 /* Do step6 onward by hand. */
5680 tcp_urg(sk
, skb
, th
);
5682 tcp_data_snd_check(sk
);
5686 req
= tp
->fastopen_rsk
;
5688 WARN_ON_ONCE(sk
->sk_state
!= TCP_SYN_RECV
&&
5689 sk
->sk_state
!= TCP_FIN_WAIT1
);
5691 if (tcp_check_req(sk
, skb
, req
, NULL
, true) == NULL
)
5695 if (!th
->ack
&& !th
->rst
)
5698 if (!tcp_validate_incoming(sk
, skb
, th
, 0))
5701 /* step 5: check the ACK field */
5702 acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
|
5703 FLAG_UPDATE_TS_RECENT
) > 0;
5705 switch (sk
->sk_state
) {
5710 /* Once we leave TCP_SYN_RECV, we no longer need req
5714 synack_stamp
= tcp_rsk(req
)->snt_synack
;
5715 tp
->total_retrans
= req
->num_retrans
;
5716 reqsk_fastopen_remove(sk
, req
, false);
5718 synack_stamp
= tp
->lsndtime
;
5719 /* Make sure socket is routed, for correct metrics. */
5720 icsk
->icsk_af_ops
->rebuild_header(sk
);
5721 tcp_init_congestion_control(sk
);
5724 tp
->copied_seq
= tp
->rcv_nxt
;
5725 tcp_init_buffer_space(sk
);
5728 tcp_set_state(sk
, TCP_ESTABLISHED
);
5729 sk
->sk_state_change(sk
);
5731 /* Note, that this wakeup is only for marginal crossed SYN case.
5732 * Passively open sockets are not waked up, because
5733 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5736 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5738 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
5739 tp
->snd_wnd
= ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
;
5740 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5741 tcp_synack_rtt_meas(sk
, synack_stamp
);
5743 if (tp
->rx_opt
.tstamp_ok
)
5744 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5747 /* Re-arm the timer because data may have been sent out.
5748 * This is similar to the regular data transmission case
5749 * when new data has just been ack'ed.
5751 * (TFO) - we could try to be more aggressive and
5752 * retransmitting any data sooner based on when they
5757 tcp_init_metrics(sk
);
5759 tcp_update_pacing_rate(sk
);
5761 /* Prevent spurious tcp_cwnd_restart() on first data packet */
5762 tp
->lsndtime
= tcp_time_stamp
;
5764 tcp_initialize_rcv_mss(sk
);
5765 tcp_fast_path_on(tp
);
5768 case TCP_FIN_WAIT1
: {
5769 struct dst_entry
*dst
;
5772 /* If we enter the TCP_FIN_WAIT1 state and we are a
5773 * Fast Open socket and this is the first acceptable
5774 * ACK we have received, this would have acknowledged
5775 * our SYNACK so stop the SYNACK timer.
5778 /* Return RST if ack_seq is invalid.
5779 * Note that RFC793 only says to generate a
5780 * DUPACK for it but for TCP Fast Open it seems
5781 * better to treat this case like TCP_SYN_RECV
5786 /* We no longer need the request sock. */
5787 reqsk_fastopen_remove(sk
, req
, false);
5790 if (tp
->snd_una
!= tp
->write_seq
)
5793 tcp_set_state(sk
, TCP_FIN_WAIT2
);
5794 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
5796 dst
= __sk_dst_get(sk
);
5800 if (!sock_flag(sk
, SOCK_DEAD
)) {
5801 /* Wake up lingering close() */
5802 sk
->sk_state_change(sk
);
5806 if (tp
->linger2
< 0 ||
5807 (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5808 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
))) {
5810 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
5814 tmo
= tcp_fin_time(sk
);
5815 if (tmo
> TCP_TIMEWAIT_LEN
) {
5816 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
5817 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
5818 /* Bad case. We could lose such FIN otherwise.
5819 * It is not a big problem, but it looks confusing
5820 * and not so rare event. We still can lose it now,
5821 * if it spins in bh_lock_sock(), but it is really
5824 inet_csk_reset_keepalive_timer(sk
, tmo
);
5826 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
5833 if (tp
->snd_una
== tp
->write_seq
) {
5834 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
5840 if (tp
->snd_una
== tp
->write_seq
) {
5841 tcp_update_metrics(sk
);
5848 /* step 6: check the URG bit */
5849 tcp_urg(sk
, skb
, th
);
5851 /* step 7: process the segment text */
5852 switch (sk
->sk_state
) {
5853 case TCP_CLOSE_WAIT
:
5856 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
5860 /* RFC 793 says to queue data in these states,
5861 * RFC 1122 says we MUST send a reset.
5862 * BSD 4.4 also does reset.
5864 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
5865 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5866 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
5867 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
5873 case TCP_ESTABLISHED
:
5874 tcp_data_queue(sk
, skb
);
5879 /* tcp_data could move socket to TIME-WAIT */
5880 if (sk
->sk_state
!= TCP_CLOSE
) {
5881 tcp_data_snd_check(sk
);
5882 tcp_ack_snd_check(sk
);
5891 EXPORT_SYMBOL(tcp_rcv_state_process
);
5893 static inline void pr_drop_req(struct request_sock
*req
, __u16 port
, int family
)
5895 struct inet_request_sock
*ireq
= inet_rsk(req
);
5897 if (family
== AF_INET
)
5898 LIMIT_NETDEBUG(KERN_DEBUG
pr_fmt("drop open request from %pI4/%u\n"),
5899 &ireq
->ir_rmt_addr
, port
);
5900 #if IS_ENABLED(CONFIG_IPV6)
5901 else if (family
== AF_INET6
)
5902 LIMIT_NETDEBUG(KERN_DEBUG
pr_fmt("drop open request from %pI6/%u\n"),
5903 &ireq
->ir_v6_rmt_addr
, port
);
5907 int tcp_conn_request(struct request_sock_ops
*rsk_ops
,
5908 const struct tcp_request_sock_ops
*af_ops
,
5909 struct sock
*sk
, struct sk_buff
*skb
)
5911 struct tcp_options_received tmp_opt
;
5912 struct request_sock
*req
;
5913 struct tcp_sock
*tp
= tcp_sk(sk
);
5914 struct dst_entry
*dst
= NULL
;
5915 __u32 isn
= TCP_SKB_CB(skb
)->when
;
5916 bool want_cookie
= false, fastopen
;
5918 struct tcp_fastopen_cookie foc
= { .len
= -1 };
5922 /* TW buckets are converted to open requests without
5923 * limitations, they conserve resources and peer is
5924 * evidently real one.
5926 if ((sysctl_tcp_syncookies
== 2 ||
5927 inet_csk_reqsk_queue_is_full(sk
)) && !isn
) {
5928 want_cookie
= tcp_syn_flood_action(sk
, skb
, rsk_ops
->slab_name
);
5934 /* Accept backlog is full. If we have already queued enough
5935 * of warm entries in syn queue, drop request. It is better than
5936 * clogging syn queue with openreqs with exponentially increasing
5939 if (sk_acceptq_is_full(sk
) && inet_csk_reqsk_queue_young(sk
) > 1) {
5940 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_LISTENOVERFLOWS
);
5944 req
= inet_reqsk_alloc(rsk_ops
);
5948 tcp_rsk(req
)->af_specific
= af_ops
;
5950 tcp_clear_options(&tmp_opt
);
5951 tmp_opt
.mss_clamp
= af_ops
->mss_clamp
;
5952 tmp_opt
.user_mss
= tp
->rx_opt
.user_mss
;
5953 tcp_parse_options(skb
, &tmp_opt
, 0, want_cookie
? NULL
: &foc
);
5955 if (want_cookie
&& !tmp_opt
.saw_tstamp
)
5956 tcp_clear_options(&tmp_opt
);
5958 tmp_opt
.tstamp_ok
= tmp_opt
.saw_tstamp
;
5959 tcp_openreq_init(req
, &tmp_opt
, skb
, sk
);
5961 af_ops
->init_req(req
, sk
, skb
);
5963 if (security_inet_conn_request(sk
, skb
, req
))
5966 if (!want_cookie
|| tmp_opt
.tstamp_ok
)
5967 TCP_ECN_create_request(req
, skb
, sock_net(sk
));
5970 isn
= cookie_init_sequence(af_ops
, sk
, skb
, &req
->mss
);
5971 req
->cookie_ts
= tmp_opt
.tstamp_ok
;
5973 /* VJ's idea. We save last timestamp seen
5974 * from the destination in peer table, when entering
5975 * state TIME-WAIT, and check against it before
5976 * accepting new connection request.
5978 * If "isn" is not zero, this request hit alive
5979 * timewait bucket, so that all the necessary checks
5980 * are made in the function processing timewait state.
5982 if (tmp_opt
.saw_tstamp
&& tcp_death_row
.sysctl_tw_recycle
) {
5985 dst
= af_ops
->route_req(sk
, &fl
, req
, &strict
);
5986 if (dst
&& strict
&&
5987 !tcp_peer_is_proven(req
, dst
, true)) {
5988 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSPASSIVEREJECTED
);
5989 goto drop_and_release
;
5992 /* Kill the following clause, if you dislike this way. */
5993 else if (!sysctl_tcp_syncookies
&&
5994 (sysctl_max_syn_backlog
- inet_csk_reqsk_queue_len(sk
) <
5995 (sysctl_max_syn_backlog
>> 2)) &&
5996 !tcp_peer_is_proven(req
, dst
, false)) {
5997 /* Without syncookies last quarter of
5998 * backlog is filled with destinations,
5999 * proven to be alive.
6000 * It means that we continue to communicate
6001 * to destinations, already remembered
6002 * to the moment of synflood.
6004 pr_drop_req(req
, ntohs(tcp_hdr(skb
)->source
),
6006 goto drop_and_release
;
6009 isn
= af_ops
->init_seq(skb
);
6012 dst
= af_ops
->route_req(sk
, &fl
, req
, NULL
);
6017 tcp_rsk(req
)->snt_isn
= isn
;
6018 tcp_openreq_init_rwin(req
, sk
, dst
);
6019 fastopen
= !want_cookie
&&
6020 tcp_try_fastopen(sk
, skb
, req
, &foc
, dst
);
6021 err
= af_ops
->send_synack(sk
, dst
, &fl
, req
,
6022 skb_get_queue_mapping(skb
), &foc
);
6024 if (err
|| want_cookie
)
6027 tcp_rsk(req
)->listener
= NULL
;
6028 af_ops
->queue_hash_add(sk
, req
, TCP_TIMEOUT_INIT
);
6038 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_LISTENDROPS
);
6041 EXPORT_SYMBOL(tcp_conn_request
);