2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
23 * Pedro Roque : Fast Retransmit/Recovery.
25 * Retransmit queue handled by TCP.
26 * Better retransmit timer handling.
27 * New congestion avoidance.
31 * Eric : Fast Retransmit.
32 * Randy Scott : MSS option defines.
33 * Eric Schenk : Fixes to slow start algorithm.
34 * Eric Schenk : Yet another double ACK bug.
35 * Eric Schenk : Delayed ACK bug fixes.
36 * Eric Schenk : Floyd style fast retrans war avoidance.
37 * David S. Miller : Don't allow zero congestion window.
38 * Eric Schenk : Fix retransmitter so that it sends
39 * next packet on ack of previous packet.
40 * Andi Kleen : Moved open_request checking here
41 * and process RSTs for open_requests.
42 * Andi Kleen : Better prune_queue, and other fixes.
43 * Andrey Savochkin: Fix RTT measurements in the presence of
45 * Andrey Savochkin: Check sequence numbers correctly when
46 * removing SACKs due to in sequence incoming
48 * Andi Kleen: Make sure we never ack data there is not
49 * enough room for. Also make this condition
50 * a fatal error if it might still happen.
51 * Andi Kleen: Add tcp_measure_rcv_mss to make
52 * connections with MSS<min(MTU,ann. MSS)
53 * work without delayed acks.
54 * Andi Kleen: Process packets with PSH set in the
56 * J Hadi Salim: ECN support
59 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
60 * engine. Lots of bugs are found.
61 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
64 #define pr_fmt(fmt) "TCP: " fmt
67 #include <linux/slab.h>
68 #include <linux/module.h>
69 #include <linux/sysctl.h>
70 #include <linux/kernel.h>
71 #include <linux/prefetch.h>
74 #include <net/inet_common.h>
75 #include <linux/ipsec.h>
76 #include <asm/unaligned.h>
77 #include <linux/errqueue.h>
79 int sysctl_tcp_timestamps __read_mostly
= 1;
80 int sysctl_tcp_window_scaling __read_mostly
= 1;
81 int sysctl_tcp_sack __read_mostly
= 1;
82 int sysctl_tcp_fack __read_mostly
= 1;
83 int sysctl_tcp_max_reordering __read_mostly
= 300;
84 int sysctl_tcp_dsack __read_mostly
= 1;
85 int sysctl_tcp_app_win __read_mostly
= 31;
86 int sysctl_tcp_adv_win_scale __read_mostly
= 1;
87 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale
);
89 /* rfc5961 challenge ack rate limiting */
90 int sysctl_tcp_challenge_ack_limit
= 1000;
92 int sysctl_tcp_stdurg __read_mostly
;
93 int sysctl_tcp_rfc1337 __read_mostly
;
94 int sysctl_tcp_max_orphans __read_mostly
= NR_FILE
;
95 int sysctl_tcp_frto __read_mostly
= 2;
96 int sysctl_tcp_min_rtt_wlen __read_mostly
= 300;
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;
102 int sysctl_tcp_invalid_ratelimit __read_mostly
= HZ
/2;
104 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
105 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
106 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
107 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
108 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
109 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
110 #define FLAG_ECE 0x40 /* ECE in this ACK */
111 #define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */
112 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
113 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
114 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
115 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
116 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
117 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
119 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
120 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
121 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
122 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
124 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
125 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
127 #define REXMIT_NONE 0 /* no loss recovery to do */
128 #define REXMIT_LOST 1 /* retransmit packets marked lost */
129 #define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */
131 /* Adapt the MSS value used to make delayed ack decision to the
134 static void tcp_measure_rcv_mss(struct sock
*sk
, const struct sk_buff
*skb
)
136 struct inet_connection_sock
*icsk
= inet_csk(sk
);
137 const unsigned int lss
= icsk
->icsk_ack
.last_seg_size
;
140 icsk
->icsk_ack
.last_seg_size
= 0;
142 /* skb->len may jitter because of SACKs, even if peer
143 * sends good full-sized frames.
145 len
= skb_shinfo(skb
)->gso_size
? : skb
->len
;
146 if (len
>= icsk
->icsk_ack
.rcv_mss
) {
147 icsk
->icsk_ack
.rcv_mss
= len
;
149 /* Otherwise, we make more careful check taking into account,
150 * that SACKs block is variable.
152 * "len" is invariant segment length, including TCP header.
154 len
+= skb
->data
- skb_transport_header(skb
);
155 if (len
>= TCP_MSS_DEFAULT
+ sizeof(struct tcphdr
) ||
156 /* If PSH is not set, packet should be
157 * full sized, provided peer TCP is not badly broken.
158 * This observation (if it is correct 8)) allows
159 * to handle super-low mtu links fairly.
161 (len
>= TCP_MIN_MSS
+ sizeof(struct tcphdr
) &&
162 !(tcp_flag_word(tcp_hdr(skb
)) & TCP_REMNANT
))) {
163 /* Subtract also invariant (if peer is RFC compliant),
164 * tcp header plus fixed timestamp option length.
165 * Resulting "len" is MSS free of SACK jitter.
167 len
-= tcp_sk(sk
)->tcp_header_len
;
168 icsk
->icsk_ack
.last_seg_size
= len
;
170 icsk
->icsk_ack
.rcv_mss
= len
;
174 if (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
)
175 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED2
;
176 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
180 static void tcp_incr_quickack(struct sock
*sk
)
182 struct inet_connection_sock
*icsk
= inet_csk(sk
);
183 unsigned int quickacks
= tcp_sk(sk
)->rcv_wnd
/ (2 * icsk
->icsk_ack
.rcv_mss
);
187 if (quickacks
> icsk
->icsk_ack
.quick
)
188 icsk
->icsk_ack
.quick
= min(quickacks
, TCP_MAX_QUICKACKS
);
191 static void tcp_enter_quickack_mode(struct sock
*sk
)
193 struct inet_connection_sock
*icsk
= inet_csk(sk
);
194 tcp_incr_quickack(sk
);
195 icsk
->icsk_ack
.pingpong
= 0;
196 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
199 /* Send ACKs quickly, if "quick" count is not exhausted
200 * and the session is not interactive.
203 static bool tcp_in_quickack_mode(struct sock
*sk
)
205 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
206 const struct dst_entry
*dst
= __sk_dst_get(sk
);
208 return (dst
&& dst_metric(dst
, RTAX_QUICKACK
)) ||
209 (icsk
->icsk_ack
.quick
&& !icsk
->icsk_ack
.pingpong
);
212 static void tcp_ecn_queue_cwr(struct tcp_sock
*tp
)
214 if (tp
->ecn_flags
& TCP_ECN_OK
)
215 tp
->ecn_flags
|= TCP_ECN_QUEUE_CWR
;
218 static void tcp_ecn_accept_cwr(struct tcp_sock
*tp
, const struct sk_buff
*skb
)
220 if (tcp_hdr(skb
)->cwr
)
221 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
224 static void tcp_ecn_withdraw_cwr(struct tcp_sock
*tp
)
226 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
229 static void __tcp_ecn_check_ce(struct tcp_sock
*tp
, const struct sk_buff
*skb
)
231 switch (TCP_SKB_CB(skb
)->ip_dsfield
& INET_ECN_MASK
) {
232 case INET_ECN_NOT_ECT
:
233 /* Funny extension: if ECT is not set on a segment,
234 * and we already seen ECT on a previous segment,
235 * it is probably a retransmit.
237 if (tp
->ecn_flags
& TCP_ECN_SEEN
)
238 tcp_enter_quickack_mode((struct sock
*)tp
);
241 if (tcp_ca_needs_ecn((struct sock
*)tp
))
242 tcp_ca_event((struct sock
*)tp
, CA_EVENT_ECN_IS_CE
);
244 if (!(tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)) {
245 /* Better not delay acks, sender can have a very low cwnd */
246 tcp_enter_quickack_mode((struct sock
*)tp
);
247 tp
->ecn_flags
|= TCP_ECN_DEMAND_CWR
;
249 tp
->ecn_flags
|= TCP_ECN_SEEN
;
252 if (tcp_ca_needs_ecn((struct sock
*)tp
))
253 tcp_ca_event((struct sock
*)tp
, CA_EVENT_ECN_NO_CE
);
254 tp
->ecn_flags
|= TCP_ECN_SEEN
;
259 static void tcp_ecn_check_ce(struct tcp_sock
*tp
, const struct sk_buff
*skb
)
261 if (tp
->ecn_flags
& TCP_ECN_OK
)
262 __tcp_ecn_check_ce(tp
, skb
);
265 static void tcp_ecn_rcv_synack(struct tcp_sock
*tp
, const struct tcphdr
*th
)
267 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| th
->cwr
))
268 tp
->ecn_flags
&= ~TCP_ECN_OK
;
271 static void tcp_ecn_rcv_syn(struct tcp_sock
*tp
, const struct tcphdr
*th
)
273 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| !th
->cwr
))
274 tp
->ecn_flags
&= ~TCP_ECN_OK
;
277 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock
*tp
, const struct tcphdr
*th
)
279 if (th
->ece
&& !th
->syn
&& (tp
->ecn_flags
& TCP_ECN_OK
))
284 /* Buffer size and advertised window tuning.
286 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
289 static void tcp_sndbuf_expand(struct sock
*sk
)
291 const struct tcp_sock
*tp
= tcp_sk(sk
);
292 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
296 /* Worst case is non GSO/TSO : each frame consumes one skb
297 * and skb->head is kmalloced using power of two area of memory
299 per_mss
= max_t(u32
, tp
->rx_opt
.mss_clamp
, tp
->mss_cache
) +
301 SKB_DATA_ALIGN(sizeof(struct skb_shared_info
));
303 per_mss
= roundup_pow_of_two(per_mss
) +
304 SKB_DATA_ALIGN(sizeof(struct sk_buff
));
306 nr_segs
= max_t(u32
, TCP_INIT_CWND
, tp
->snd_cwnd
);
307 nr_segs
= max_t(u32
, nr_segs
, tp
->reordering
+ 1);
309 /* Fast Recovery (RFC 5681 3.2) :
310 * Cubic needs 1.7 factor, rounded to 2 to include
311 * extra cushion (application might react slowly to POLLOUT)
313 sndmem
= ca_ops
->sndbuf_expand
? ca_ops
->sndbuf_expand(sk
) : 2;
314 sndmem
*= nr_segs
* per_mss
;
316 if (sk
->sk_sndbuf
< sndmem
)
317 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
320 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
322 * All tcp_full_space() is split to two parts: "network" buffer, allocated
323 * forward and advertised in receiver window (tp->rcv_wnd) and
324 * "application buffer", required to isolate scheduling/application
325 * latencies from network.
326 * window_clamp is maximal advertised window. It can be less than
327 * tcp_full_space(), in this case tcp_full_space() - window_clamp
328 * is reserved for "application" buffer. The less window_clamp is
329 * the smoother our behaviour from viewpoint of network, but the lower
330 * throughput and the higher sensitivity of the connection to losses. 8)
332 * rcv_ssthresh is more strict window_clamp used at "slow start"
333 * phase to predict further behaviour of this connection.
334 * It is used for two goals:
335 * - to enforce header prediction at sender, even when application
336 * requires some significant "application buffer". It is check #1.
337 * - to prevent pruning of receive queue because of misprediction
338 * of receiver window. Check #2.
340 * The scheme does not work when sender sends good segments opening
341 * window and then starts to feed us spaghetti. But it should work
342 * in common situations. Otherwise, we have to rely on queue collapsing.
345 /* Slow part of check#2. */
346 static int __tcp_grow_window(const struct sock
*sk
, const struct sk_buff
*skb
)
348 struct tcp_sock
*tp
= tcp_sk(sk
);
350 int truesize
= tcp_win_from_space(skb
->truesize
) >> 1;
351 int window
= tcp_win_from_space(sysctl_tcp_rmem
[2]) >> 1;
353 while (tp
->rcv_ssthresh
<= window
) {
354 if (truesize
<= skb
->len
)
355 return 2 * inet_csk(sk
)->icsk_ack
.rcv_mss
;
363 static void tcp_grow_window(struct sock
*sk
, const struct sk_buff
*skb
)
365 struct tcp_sock
*tp
= tcp_sk(sk
);
368 if (tp
->rcv_ssthresh
< tp
->window_clamp
&&
369 (int)tp
->rcv_ssthresh
< tcp_space(sk
) &&
370 !tcp_under_memory_pressure(sk
)) {
373 /* Check #2. Increase window, if skb with such overhead
374 * will fit to rcvbuf in future.
376 if (tcp_win_from_space(skb
->truesize
) <= skb
->len
)
377 incr
= 2 * tp
->advmss
;
379 incr
= __tcp_grow_window(sk
, skb
);
382 incr
= max_t(int, incr
, 2 * skb
->len
);
383 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
+ incr
,
385 inet_csk(sk
)->icsk_ack
.quick
|= 1;
390 /* 3. Tuning rcvbuf, when connection enters established state. */
391 static void tcp_fixup_rcvbuf(struct sock
*sk
)
393 u32 mss
= tcp_sk(sk
)->advmss
;
396 rcvmem
= 2 * SKB_TRUESIZE(mss
+ MAX_TCP_HEADER
) *
397 tcp_default_init_rwnd(mss
);
399 /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
400 * Allow enough cushion so that sender is not limited by our window
402 if (sysctl_tcp_moderate_rcvbuf
)
405 if (sk
->sk_rcvbuf
< rcvmem
)
406 sk
->sk_rcvbuf
= min(rcvmem
, sysctl_tcp_rmem
[2]);
409 /* 4. Try to fixup all. It is made immediately after connection enters
412 void tcp_init_buffer_space(struct sock
*sk
)
414 struct tcp_sock
*tp
= tcp_sk(sk
);
417 if (!(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
))
418 tcp_fixup_rcvbuf(sk
);
419 if (!(sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
))
420 tcp_sndbuf_expand(sk
);
422 tp
->rcvq_space
.space
= tp
->rcv_wnd
;
423 tp
->rcvq_space
.time
= tcp_time_stamp
;
424 tp
->rcvq_space
.seq
= tp
->copied_seq
;
426 maxwin
= tcp_full_space(sk
);
428 if (tp
->window_clamp
>= maxwin
) {
429 tp
->window_clamp
= maxwin
;
431 if (sysctl_tcp_app_win
&& maxwin
> 4 * tp
->advmss
)
432 tp
->window_clamp
= max(maxwin
-
433 (maxwin
>> sysctl_tcp_app_win
),
437 /* Force reservation of one segment. */
438 if (sysctl_tcp_app_win
&&
439 tp
->window_clamp
> 2 * tp
->advmss
&&
440 tp
->window_clamp
+ tp
->advmss
> maxwin
)
441 tp
->window_clamp
= max(2 * tp
->advmss
, maxwin
- tp
->advmss
);
443 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, tp
->window_clamp
);
444 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
447 /* 5. Recalculate window clamp after socket hit its memory bounds. */
448 static void tcp_clamp_window(struct sock
*sk
)
450 struct tcp_sock
*tp
= tcp_sk(sk
);
451 struct inet_connection_sock
*icsk
= inet_csk(sk
);
453 icsk
->icsk_ack
.quick
= 0;
455 if (sk
->sk_rcvbuf
< sysctl_tcp_rmem
[2] &&
456 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
) &&
457 !tcp_under_memory_pressure(sk
) &&
458 sk_memory_allocated(sk
) < sk_prot_mem_limits(sk
, 0)) {
459 sk
->sk_rcvbuf
= min(atomic_read(&sk
->sk_rmem_alloc
),
462 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
)
463 tp
->rcv_ssthresh
= min(tp
->window_clamp
, 2U * tp
->advmss
);
466 /* Initialize RCV_MSS value.
467 * RCV_MSS is an our guess about MSS used by the peer.
468 * We haven't any direct information about the MSS.
469 * It's better to underestimate the RCV_MSS rather than overestimate.
470 * Overestimations make us ACKing less frequently than needed.
471 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
473 void tcp_initialize_rcv_mss(struct sock
*sk
)
475 const struct tcp_sock
*tp
= tcp_sk(sk
);
476 unsigned int hint
= min_t(unsigned int, tp
->advmss
, tp
->mss_cache
);
478 hint
= min(hint
, tp
->rcv_wnd
/ 2);
479 hint
= min(hint
, TCP_MSS_DEFAULT
);
480 hint
= max(hint
, TCP_MIN_MSS
);
482 inet_csk(sk
)->icsk_ack
.rcv_mss
= hint
;
484 EXPORT_SYMBOL(tcp_initialize_rcv_mss
);
486 /* Receiver "autotuning" code.
488 * The algorithm for RTT estimation w/o timestamps is based on
489 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
490 * <http://public.lanl.gov/radiant/pubs.html#DRS>
492 * More detail on this code can be found at
493 * <http://staff.psc.edu/jheffner/>,
494 * though this reference is out of date. A new paper
497 static void tcp_rcv_rtt_update(struct tcp_sock
*tp
, u32 sample
, int win_dep
)
499 u32 new_sample
= tp
->rcv_rtt_est
.rtt
;
505 if (new_sample
!= 0) {
506 /* If we sample in larger samples in the non-timestamp
507 * case, we could grossly overestimate the RTT especially
508 * with chatty applications or bulk transfer apps which
509 * are stalled on filesystem I/O.
511 * Also, since we are only going for a minimum in the
512 * non-timestamp case, we do not smooth things out
513 * else with timestamps disabled convergence takes too
517 m
-= (new_sample
>> 3);
525 /* No previous measure. */
529 if (tp
->rcv_rtt_est
.rtt
!= new_sample
)
530 tp
->rcv_rtt_est
.rtt
= new_sample
;
533 static inline void tcp_rcv_rtt_measure(struct tcp_sock
*tp
)
535 if (tp
->rcv_rtt_est
.time
== 0)
537 if (before(tp
->rcv_nxt
, tp
->rcv_rtt_est
.seq
))
539 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rcv_rtt_est
.time
, 1);
542 tp
->rcv_rtt_est
.seq
= tp
->rcv_nxt
+ tp
->rcv_wnd
;
543 tp
->rcv_rtt_est
.time
= tcp_time_stamp
;
546 static inline void tcp_rcv_rtt_measure_ts(struct sock
*sk
,
547 const struct sk_buff
*skb
)
549 struct tcp_sock
*tp
= tcp_sk(sk
);
550 if (tp
->rx_opt
.rcv_tsecr
&&
551 (TCP_SKB_CB(skb
)->end_seq
-
552 TCP_SKB_CB(skb
)->seq
>= inet_csk(sk
)->icsk_ack
.rcv_mss
))
553 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
, 0);
557 * This function should be called every time data is copied to user space.
558 * It calculates the appropriate TCP receive buffer space.
560 void tcp_rcv_space_adjust(struct sock
*sk
)
562 struct tcp_sock
*tp
= tcp_sk(sk
);
566 time
= tcp_time_stamp
- tp
->rcvq_space
.time
;
567 if (time
< (tp
->rcv_rtt_est
.rtt
>> 3) || tp
->rcv_rtt_est
.rtt
== 0)
570 /* Number of bytes copied to user in last RTT */
571 copied
= tp
->copied_seq
- tp
->rcvq_space
.seq
;
572 if (copied
<= tp
->rcvq_space
.space
)
576 * copied = bytes received in previous RTT, our base window
577 * To cope with packet losses, we need a 2x factor
578 * To cope with slow start, and sender growing its cwin by 100 %
579 * every RTT, we need a 4x factor, because the ACK we are sending
580 * now is for the next RTT, not the current one :
581 * <prev RTT . ><current RTT .. ><next RTT .... >
584 if (sysctl_tcp_moderate_rcvbuf
&&
585 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
)) {
586 int rcvwin
, rcvmem
, rcvbuf
;
588 /* minimal window to cope with packet losses, assuming
589 * steady state. Add some cushion because of small variations.
591 rcvwin
= (copied
<< 1) + 16 * tp
->advmss
;
593 /* If rate increased by 25%,
594 * assume slow start, rcvwin = 3 * copied
595 * If rate increased by 50%,
596 * assume sender can use 2x growth, rcvwin = 4 * copied
599 tp
->rcvq_space
.space
+ (tp
->rcvq_space
.space
>> 2)) {
601 tp
->rcvq_space
.space
+ (tp
->rcvq_space
.space
>> 1))
604 rcvwin
+= (rcvwin
>> 1);
607 rcvmem
= SKB_TRUESIZE(tp
->advmss
+ MAX_TCP_HEADER
);
608 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
611 rcvbuf
= min(rcvwin
/ tp
->advmss
* rcvmem
, sysctl_tcp_rmem
[2]);
612 if (rcvbuf
> sk
->sk_rcvbuf
) {
613 sk
->sk_rcvbuf
= rcvbuf
;
615 /* Make the window clamp follow along. */
616 tp
->window_clamp
= rcvwin
;
619 tp
->rcvq_space
.space
= copied
;
622 tp
->rcvq_space
.seq
= tp
->copied_seq
;
623 tp
->rcvq_space
.time
= tcp_time_stamp
;
626 /* There is something which you must keep in mind when you analyze the
627 * behavior of the tp->ato delayed ack timeout interval. When a
628 * connection starts up, we want to ack as quickly as possible. The
629 * problem is that "good" TCP's do slow start at the beginning of data
630 * transmission. The means that until we send the first few ACK's the
631 * sender will sit on his end and only queue most of his data, because
632 * he can only send snd_cwnd unacked packets at any given time. For
633 * each ACK we send, he increments snd_cwnd and transmits more of his
636 static void tcp_event_data_recv(struct sock
*sk
, struct sk_buff
*skb
)
638 struct tcp_sock
*tp
= tcp_sk(sk
);
639 struct inet_connection_sock
*icsk
= inet_csk(sk
);
642 inet_csk_schedule_ack(sk
);
644 tcp_measure_rcv_mss(sk
, skb
);
646 tcp_rcv_rtt_measure(tp
);
648 now
= tcp_time_stamp
;
650 if (!icsk
->icsk_ack
.ato
) {
651 /* The _first_ data packet received, initialize
652 * delayed ACK engine.
654 tcp_incr_quickack(sk
);
655 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
657 int m
= now
- icsk
->icsk_ack
.lrcvtime
;
659 if (m
<= TCP_ATO_MIN
/ 2) {
660 /* The fastest case is the first. */
661 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + TCP_ATO_MIN
/ 2;
662 } else if (m
< icsk
->icsk_ack
.ato
) {
663 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + m
;
664 if (icsk
->icsk_ack
.ato
> icsk
->icsk_rto
)
665 icsk
->icsk_ack
.ato
= icsk
->icsk_rto
;
666 } else if (m
> icsk
->icsk_rto
) {
667 /* Too long gap. Apparently sender failed to
668 * restart window, so that we send ACKs quickly.
670 tcp_incr_quickack(sk
);
674 icsk
->icsk_ack
.lrcvtime
= now
;
676 tcp_ecn_check_ce(tp
, skb
);
679 tcp_grow_window(sk
, skb
);
682 /* Called to compute a smoothed rtt estimate. The data fed to this
683 * routine either comes from timestamps, or from segments that were
684 * known _not_ to have been retransmitted [see Karn/Partridge
685 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
686 * piece by Van Jacobson.
687 * NOTE: the next three routines used to be one big routine.
688 * To save cycles in the RFC 1323 implementation it was better to break
689 * it up into three procedures. -- erics
691 static void tcp_rtt_estimator(struct sock
*sk
, long mrtt_us
)
693 struct tcp_sock
*tp
= tcp_sk(sk
);
694 long m
= mrtt_us
; /* RTT */
695 u32 srtt
= tp
->srtt_us
;
697 /* The following amusing code comes from Jacobson's
698 * article in SIGCOMM '88. Note that rtt and mdev
699 * are scaled versions of rtt and mean deviation.
700 * This is designed to be as fast as possible
701 * m stands for "measurement".
703 * On a 1990 paper the rto value is changed to:
704 * RTO = rtt + 4 * mdev
706 * Funny. This algorithm seems to be very broken.
707 * These formulae increase RTO, when it should be decreased, increase
708 * too slowly, when it should be increased quickly, decrease too quickly
709 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
710 * does not matter how to _calculate_ it. Seems, it was trap
711 * that VJ failed to avoid. 8)
714 m
-= (srtt
>> 3); /* m is now error in rtt est */
715 srtt
+= m
; /* rtt = 7/8 rtt + 1/8 new */
717 m
= -m
; /* m is now abs(error) */
718 m
-= (tp
->mdev_us
>> 2); /* similar update on mdev */
719 /* This is similar to one of Eifel findings.
720 * Eifel blocks mdev updates when rtt decreases.
721 * This solution is a bit different: we use finer gain
722 * for mdev in this case (alpha*beta).
723 * Like Eifel it also prevents growth of rto,
724 * but also it limits too fast rto decreases,
725 * happening in pure Eifel.
730 m
-= (tp
->mdev_us
>> 2); /* similar update on mdev */
732 tp
->mdev_us
+= m
; /* mdev = 3/4 mdev + 1/4 new */
733 if (tp
->mdev_us
> tp
->mdev_max_us
) {
734 tp
->mdev_max_us
= tp
->mdev_us
;
735 if (tp
->mdev_max_us
> tp
->rttvar_us
)
736 tp
->rttvar_us
= tp
->mdev_max_us
;
738 if (after(tp
->snd_una
, tp
->rtt_seq
)) {
739 if (tp
->mdev_max_us
< tp
->rttvar_us
)
740 tp
->rttvar_us
-= (tp
->rttvar_us
- tp
->mdev_max_us
) >> 2;
741 tp
->rtt_seq
= tp
->snd_nxt
;
742 tp
->mdev_max_us
= tcp_rto_min_us(sk
);
745 /* no previous measure. */
746 srtt
= m
<< 3; /* take the measured time to be rtt */
747 tp
->mdev_us
= m
<< 1; /* make sure rto = 3*rtt */
748 tp
->rttvar_us
= max(tp
->mdev_us
, tcp_rto_min_us(sk
));
749 tp
->mdev_max_us
= tp
->rttvar_us
;
750 tp
->rtt_seq
= tp
->snd_nxt
;
752 tp
->srtt_us
= max(1U, srtt
);
755 /* Set the sk_pacing_rate to allow proper sizing of TSO packets.
756 * Note: TCP stack does not yet implement pacing.
757 * FQ packet scheduler can be used to implement cheap but effective
758 * TCP pacing, to smooth the burst on large writes when packets
759 * in flight is significantly lower than cwnd (or rwin)
761 int sysctl_tcp_pacing_ss_ratio __read_mostly
= 200;
762 int sysctl_tcp_pacing_ca_ratio __read_mostly
= 120;
764 static void tcp_update_pacing_rate(struct sock
*sk
)
766 const struct tcp_sock
*tp
= tcp_sk(sk
);
769 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
770 rate
= (u64
)tp
->mss_cache
* ((USEC_PER_SEC
/ 100) << 3);
772 /* current rate is (cwnd * mss) / srtt
773 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
774 * In Congestion Avoidance phase, set it to 120 % the current rate.
776 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
777 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
778 * end of slow start and should slow down.
780 if (tp
->snd_cwnd
< tp
->snd_ssthresh
/ 2)
781 rate
*= sysctl_tcp_pacing_ss_ratio
;
783 rate
*= sysctl_tcp_pacing_ca_ratio
;
785 rate
*= max(tp
->snd_cwnd
, tp
->packets_out
);
787 if (likely(tp
->srtt_us
))
788 do_div(rate
, tp
->srtt_us
);
790 /* ACCESS_ONCE() is needed because sch_fq fetches sk_pacing_rate
791 * without any lock. We want to make sure compiler wont store
792 * intermediate values in this location.
794 ACCESS_ONCE(sk
->sk_pacing_rate
) = min_t(u64
, rate
,
795 sk
->sk_max_pacing_rate
);
798 /* Calculate rto without backoff. This is the second half of Van Jacobson's
799 * routine referred to above.
801 static void tcp_set_rto(struct sock
*sk
)
803 const struct tcp_sock
*tp
= tcp_sk(sk
);
804 /* Old crap is replaced with new one. 8)
807 * 1. If rtt variance happened to be less 50msec, it is hallucination.
808 * It cannot be less due to utterly erratic ACK generation made
809 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
810 * to do with delayed acks, because at cwnd>2 true delack timeout
811 * is invisible. Actually, Linux-2.4 also generates erratic
812 * ACKs in some circumstances.
814 inet_csk(sk
)->icsk_rto
= __tcp_set_rto(tp
);
816 /* 2. Fixups made earlier cannot be right.
817 * If we do not estimate RTO correctly without them,
818 * all the algo is pure shit and should be replaced
819 * with correct one. It is exactly, which we pretend to do.
822 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
823 * guarantees that rto is higher.
828 __u32
tcp_init_cwnd(const struct tcp_sock
*tp
, const struct dst_entry
*dst
)
830 __u32 cwnd
= (dst
? dst_metric(dst
, RTAX_INITCWND
) : 0);
833 cwnd
= TCP_INIT_CWND
;
834 return min_t(__u32
, cwnd
, tp
->snd_cwnd_clamp
);
838 * Packet counting of FACK is based on in-order assumptions, therefore TCP
839 * disables it when reordering is detected
841 void tcp_disable_fack(struct tcp_sock
*tp
)
843 /* RFC3517 uses different metric in lost marker => reset on change */
845 tp
->lost_skb_hint
= NULL
;
846 tp
->rx_opt
.sack_ok
&= ~TCP_FACK_ENABLED
;
849 /* Take a notice that peer is sending D-SACKs */
850 static void tcp_dsack_seen(struct tcp_sock
*tp
)
852 tp
->rx_opt
.sack_ok
|= TCP_DSACK_SEEN
;
855 static void tcp_update_reordering(struct sock
*sk
, const int metric
,
858 struct tcp_sock
*tp
= tcp_sk(sk
);
859 if (metric
> tp
->reordering
) {
862 tp
->reordering
= min(sysctl_tcp_max_reordering
, metric
);
864 /* This exciting event is worth to be remembered. 8) */
866 mib_idx
= LINUX_MIB_TCPTSREORDER
;
867 else if (tcp_is_reno(tp
))
868 mib_idx
= LINUX_MIB_TCPRENOREORDER
;
869 else if (tcp_is_fack(tp
))
870 mib_idx
= LINUX_MIB_TCPFACKREORDER
;
872 mib_idx
= LINUX_MIB_TCPSACKREORDER
;
874 NET_INC_STATS(sock_net(sk
), mib_idx
);
875 #if FASTRETRANS_DEBUG > 1
876 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
877 tp
->rx_opt
.sack_ok
, inet_csk(sk
)->icsk_ca_state
,
881 tp
->undo_marker
? tp
->undo_retrans
: 0);
883 tcp_disable_fack(tp
);
887 tcp_disable_early_retrans(tp
);
891 /* This must be called before lost_out is incremented */
892 static void tcp_verify_retransmit_hint(struct tcp_sock
*tp
, struct sk_buff
*skb
)
894 if (!tp
->retransmit_skb_hint
||
895 before(TCP_SKB_CB(skb
)->seq
,
896 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
897 tp
->retransmit_skb_hint
= skb
;
900 after(TCP_SKB_CB(skb
)->end_seq
, tp
->retransmit_high
))
901 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
904 /* Sum the number of packets on the wire we have marked as lost.
905 * There are two cases we care about here:
906 * a) Packet hasn't been marked lost (nor retransmitted),
907 * and this is the first loss.
908 * b) Packet has been marked both lost and retransmitted,
909 * and this means we think it was lost again.
911 static void tcp_sum_lost(struct tcp_sock
*tp
, struct sk_buff
*skb
)
913 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
915 if (!(sacked
& TCPCB_LOST
) ||
916 ((sacked
& TCPCB_LOST
) && (sacked
& TCPCB_SACKED_RETRANS
)))
917 tp
->lost
+= tcp_skb_pcount(skb
);
920 static void tcp_skb_mark_lost(struct tcp_sock
*tp
, struct sk_buff
*skb
)
922 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
923 tcp_verify_retransmit_hint(tp
, skb
);
925 tp
->lost_out
+= tcp_skb_pcount(skb
);
926 tcp_sum_lost(tp
, skb
);
927 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
931 void tcp_skb_mark_lost_uncond_verify(struct tcp_sock
*tp
, struct sk_buff
*skb
)
933 tcp_verify_retransmit_hint(tp
, skb
);
935 tcp_sum_lost(tp
, skb
);
936 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
937 tp
->lost_out
+= tcp_skb_pcount(skb
);
938 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
942 /* This procedure tags the retransmission queue when SACKs arrive.
944 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
945 * Packets in queue with these bits set are counted in variables
946 * sacked_out, retrans_out and lost_out, correspondingly.
948 * Valid combinations are:
949 * Tag InFlight Description
950 * 0 1 - orig segment is in flight.
951 * S 0 - nothing flies, orig reached receiver.
952 * L 0 - nothing flies, orig lost by net.
953 * R 2 - both orig and retransmit are in flight.
954 * L|R 1 - orig is lost, retransmit is in flight.
955 * S|R 1 - orig reached receiver, retrans is still in flight.
956 * (L|S|R is logically valid, it could occur when L|R is sacked,
957 * but it is equivalent to plain S and code short-curcuits it to S.
958 * L|S is logically invalid, it would mean -1 packet in flight 8))
960 * These 6 states form finite state machine, controlled by the following events:
961 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
962 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
963 * 3. Loss detection event of two flavors:
964 * A. Scoreboard estimator decided the packet is lost.
965 * A'. Reno "three dupacks" marks head of queue lost.
966 * A''. Its FACK modification, head until snd.fack is lost.
967 * B. SACK arrives sacking SND.NXT at the moment, when the
968 * segment was retransmitted.
969 * 4. D-SACK added new rule: D-SACK changes any tag to S.
971 * It is pleasant to note, that state diagram turns out to be commutative,
972 * so that we are allowed not to be bothered by order of our actions,
973 * when multiple events arrive simultaneously. (see the function below).
975 * Reordering detection.
976 * --------------------
977 * Reordering metric is maximal distance, which a packet can be displaced
978 * in packet stream. With SACKs we can estimate it:
980 * 1. SACK fills old hole and the corresponding segment was not
981 * ever retransmitted -> reordering. Alas, we cannot use it
982 * when segment was retransmitted.
983 * 2. The last flaw is solved with D-SACK. D-SACK arrives
984 * for retransmitted and already SACKed segment -> reordering..
985 * Both of these heuristics are not used in Loss state, when we cannot
986 * account for retransmits accurately.
988 * SACK block validation.
989 * ----------------------
991 * SACK block range validation checks that the received SACK block fits to
992 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
993 * Note that SND.UNA is not included to the range though being valid because
994 * it means that the receiver is rather inconsistent with itself reporting
995 * SACK reneging when it should advance SND.UNA. Such SACK block this is
996 * perfectly valid, however, in light of RFC2018 which explicitly states
997 * that "SACK block MUST reflect the newest segment. Even if the newest
998 * segment is going to be discarded ...", not that it looks very clever
999 * in case of head skb. Due to potentional receiver driven attacks, we
1000 * choose to avoid immediate execution of a walk in write queue due to
1001 * reneging and defer head skb's loss recovery to standard loss recovery
1002 * procedure that will eventually trigger (nothing forbids us doing this).
1004 * Implements also blockage to start_seq wrap-around. Problem lies in the
1005 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1006 * there's no guarantee that it will be before snd_nxt (n). The problem
1007 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1010 * <- outs wnd -> <- wrapzone ->
1011 * u e n u_w e_w s n_w
1013 * |<------------+------+----- TCP seqno space --------------+---------->|
1014 * ...-- <2^31 ->| |<--------...
1015 * ...---- >2^31 ------>| |<--------...
1017 * Current code wouldn't be vulnerable but it's better still to discard such
1018 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1019 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1020 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1021 * equal to the ideal case (infinite seqno space without wrap caused issues).
1023 * With D-SACK the lower bound is extended to cover sequence space below
1024 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1025 * again, D-SACK block must not to go across snd_una (for the same reason as
1026 * for the normal SACK blocks, explained above). But there all simplicity
1027 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1028 * fully below undo_marker they do not affect behavior in anyway and can
1029 * therefore be safely ignored. In rare cases (which are more or less
1030 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1031 * fragmentation and packet reordering past skb's retransmission. To consider
1032 * them correctly, the acceptable range must be extended even more though
1033 * the exact amount is rather hard to quantify. However, tp->max_window can
1034 * be used as an exaggerated estimate.
1036 static bool tcp_is_sackblock_valid(struct tcp_sock
*tp
, bool is_dsack
,
1037 u32 start_seq
, u32 end_seq
)
1039 /* Too far in future, or reversed (interpretation is ambiguous) */
1040 if (after(end_seq
, tp
->snd_nxt
) || !before(start_seq
, end_seq
))
1043 /* Nasty start_seq wrap-around check (see comments above) */
1044 if (!before(start_seq
, tp
->snd_nxt
))
1047 /* In outstanding window? ...This is valid exit for D-SACKs too.
1048 * start_seq == snd_una is non-sensical (see comments above)
1050 if (after(start_seq
, tp
->snd_una
))
1053 if (!is_dsack
|| !tp
->undo_marker
)
1056 /* ...Then it's D-SACK, and must reside below snd_una completely */
1057 if (after(end_seq
, tp
->snd_una
))
1060 if (!before(start_seq
, tp
->undo_marker
))
1064 if (!after(end_seq
, tp
->undo_marker
))
1067 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1068 * start_seq < undo_marker and end_seq >= undo_marker.
1070 return !before(start_seq
, end_seq
- tp
->max_window
);
1073 static bool tcp_check_dsack(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1074 struct tcp_sack_block_wire
*sp
, int num_sacks
,
1077 struct tcp_sock
*tp
= tcp_sk(sk
);
1078 u32 start_seq_0
= get_unaligned_be32(&sp
[0].start_seq
);
1079 u32 end_seq_0
= get_unaligned_be32(&sp
[0].end_seq
);
1080 bool dup_sack
= false;
1082 if (before(start_seq_0
, TCP_SKB_CB(ack_skb
)->ack_seq
)) {
1085 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPDSACKRECV
);
1086 } else if (num_sacks
> 1) {
1087 u32 end_seq_1
= get_unaligned_be32(&sp
[1].end_seq
);
1088 u32 start_seq_1
= get_unaligned_be32(&sp
[1].start_seq
);
1090 if (!after(end_seq_0
, end_seq_1
) &&
1091 !before(start_seq_0
, start_seq_1
)) {
1094 NET_INC_STATS(sock_net(sk
),
1095 LINUX_MIB_TCPDSACKOFORECV
);
1099 /* D-SACK for already forgotten data... Do dumb counting. */
1100 if (dup_sack
&& tp
->undo_marker
&& tp
->undo_retrans
> 0 &&
1101 !after(end_seq_0
, prior_snd_una
) &&
1102 after(end_seq_0
, tp
->undo_marker
))
1108 struct tcp_sacktag_state
{
1111 /* Timestamps for earliest and latest never-retransmitted segment
1112 * that was SACKed. RTO needs the earliest RTT to stay conservative,
1113 * but congestion control should still get an accurate delay signal.
1115 struct skb_mstamp first_sackt
;
1116 struct skb_mstamp last_sackt
;
1117 struct rate_sample
*rate
;
1121 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1122 * the incoming SACK may not exactly match but we can find smaller MSS
1123 * aligned portion of it that matches. Therefore we might need to fragment
1124 * which may fail and creates some hassle (caller must handle error case
1127 * FIXME: this could be merged to shift decision code
1129 static int tcp_match_skb_to_sack(struct sock
*sk
, struct sk_buff
*skb
,
1130 u32 start_seq
, u32 end_seq
)
1134 unsigned int pkt_len
;
1137 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1138 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1140 if (tcp_skb_pcount(skb
) > 1 && !in_sack
&&
1141 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
1142 mss
= tcp_skb_mss(skb
);
1143 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1146 pkt_len
= start_seq
- TCP_SKB_CB(skb
)->seq
;
1150 pkt_len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1155 /* Round if necessary so that SACKs cover only full MSSes
1156 * and/or the remaining small portion (if present)
1158 if (pkt_len
> mss
) {
1159 unsigned int new_len
= (pkt_len
/ mss
) * mss
;
1160 if (!in_sack
&& new_len
< pkt_len
) {
1162 if (new_len
>= skb
->len
)
1167 err
= tcp_fragment(sk
, skb
, pkt_len
, mss
, GFP_ATOMIC
);
1175 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1176 static u8
tcp_sacktag_one(struct sock
*sk
,
1177 struct tcp_sacktag_state
*state
, u8 sacked
,
1178 u32 start_seq
, u32 end_seq
,
1179 int dup_sack
, int pcount
,
1180 const struct skb_mstamp
*xmit_time
)
1182 struct tcp_sock
*tp
= tcp_sk(sk
);
1183 int fack_count
= state
->fack_count
;
1185 /* Account D-SACK for retransmitted packet. */
1186 if (dup_sack
&& (sacked
& TCPCB_RETRANS
)) {
1187 if (tp
->undo_marker
&& tp
->undo_retrans
> 0 &&
1188 after(end_seq
, tp
->undo_marker
))
1190 if (sacked
& TCPCB_SACKED_ACKED
)
1191 state
->reord
= min(fack_count
, state
->reord
);
1194 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1195 if (!after(end_seq
, tp
->snd_una
))
1198 if (!(sacked
& TCPCB_SACKED_ACKED
)) {
1199 tcp_rack_advance(tp
, xmit_time
, sacked
);
1201 if (sacked
& TCPCB_SACKED_RETRANS
) {
1202 /* If the segment is not tagged as lost,
1203 * we do not clear RETRANS, believing
1204 * that retransmission is still in flight.
1206 if (sacked
& TCPCB_LOST
) {
1207 sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1208 tp
->lost_out
-= pcount
;
1209 tp
->retrans_out
-= pcount
;
1212 if (!(sacked
& TCPCB_RETRANS
)) {
1213 /* New sack for not retransmitted frame,
1214 * which was in hole. It is reordering.
1216 if (before(start_seq
,
1217 tcp_highest_sack_seq(tp
)))
1218 state
->reord
= min(fack_count
,
1220 if (!after(end_seq
, tp
->high_seq
))
1221 state
->flag
|= FLAG_ORIG_SACK_ACKED
;
1222 if (state
->first_sackt
.v64
== 0)
1223 state
->first_sackt
= *xmit_time
;
1224 state
->last_sackt
= *xmit_time
;
1227 if (sacked
& TCPCB_LOST
) {
1228 sacked
&= ~TCPCB_LOST
;
1229 tp
->lost_out
-= pcount
;
1233 sacked
|= TCPCB_SACKED_ACKED
;
1234 state
->flag
|= FLAG_DATA_SACKED
;
1235 tp
->sacked_out
+= pcount
;
1236 tp
->delivered
+= pcount
; /* Out-of-order packets delivered */
1238 fack_count
+= pcount
;
1240 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1241 if (!tcp_is_fack(tp
) && tp
->lost_skb_hint
&&
1242 before(start_seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
))
1243 tp
->lost_cnt_hint
+= pcount
;
1245 if (fack_count
> tp
->fackets_out
)
1246 tp
->fackets_out
= fack_count
;
1249 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1250 * frames and clear it. undo_retrans is decreased above, L|R frames
1251 * are accounted above as well.
1253 if (dup_sack
&& (sacked
& TCPCB_SACKED_RETRANS
)) {
1254 sacked
&= ~TCPCB_SACKED_RETRANS
;
1255 tp
->retrans_out
-= pcount
;
1261 /* Shift newly-SACKed bytes from this skb to the immediately previous
1262 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1264 static bool tcp_shifted_skb(struct sock
*sk
, struct sk_buff
*skb
,
1265 struct tcp_sacktag_state
*state
,
1266 unsigned int pcount
, int shifted
, int mss
,
1269 struct tcp_sock
*tp
= tcp_sk(sk
);
1270 struct sk_buff
*prev
= tcp_write_queue_prev(sk
, skb
);
1271 u32 start_seq
= TCP_SKB_CB(skb
)->seq
; /* start of newly-SACKed */
1272 u32 end_seq
= start_seq
+ shifted
; /* end of newly-SACKed */
1276 /* Adjust counters and hints for the newly sacked sequence
1277 * range but discard the return value since prev is already
1278 * marked. We must tag the range first because the seq
1279 * advancement below implicitly advances
1280 * tcp_highest_sack_seq() when skb is highest_sack.
1282 tcp_sacktag_one(sk
, state
, TCP_SKB_CB(skb
)->sacked
,
1283 start_seq
, end_seq
, dup_sack
, pcount
,
1285 tcp_rate_skb_delivered(sk
, skb
, state
->rate
);
1287 if (skb
== tp
->lost_skb_hint
)
1288 tp
->lost_cnt_hint
+= pcount
;
1290 TCP_SKB_CB(prev
)->end_seq
+= shifted
;
1291 TCP_SKB_CB(skb
)->seq
+= shifted
;
1293 tcp_skb_pcount_add(prev
, pcount
);
1294 BUG_ON(tcp_skb_pcount(skb
) < pcount
);
1295 tcp_skb_pcount_add(skb
, -pcount
);
1297 /* When we're adding to gso_segs == 1, gso_size will be zero,
1298 * in theory this shouldn't be necessary but as long as DSACK
1299 * code can come after this skb later on it's better to keep
1300 * setting gso_size to something.
1302 if (!TCP_SKB_CB(prev
)->tcp_gso_size
)
1303 TCP_SKB_CB(prev
)->tcp_gso_size
= mss
;
1305 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1306 if (tcp_skb_pcount(skb
) <= 1)
1307 TCP_SKB_CB(skb
)->tcp_gso_size
= 0;
1309 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1310 TCP_SKB_CB(prev
)->sacked
|= (TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
);
1313 BUG_ON(!tcp_skb_pcount(skb
));
1314 NET_INC_STATS(sock_net(sk
), LINUX_MIB_SACKSHIFTED
);
1318 /* Whole SKB was eaten :-) */
1320 if (skb
== tp
->retransmit_skb_hint
)
1321 tp
->retransmit_skb_hint
= prev
;
1322 if (skb
== tp
->lost_skb_hint
) {
1323 tp
->lost_skb_hint
= prev
;
1324 tp
->lost_cnt_hint
-= tcp_skb_pcount(prev
);
1327 TCP_SKB_CB(prev
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
;
1328 TCP_SKB_CB(prev
)->eor
= TCP_SKB_CB(skb
)->eor
;
1329 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
1330 TCP_SKB_CB(prev
)->end_seq
++;
1332 if (skb
== tcp_highest_sack(sk
))
1333 tcp_advance_highest_sack(sk
, skb
);
1335 tcp_skb_collapse_tstamp(prev
, skb
);
1336 if (unlikely(TCP_SKB_CB(prev
)->tx
.delivered_mstamp
.v64
))
1337 TCP_SKB_CB(prev
)->tx
.delivered_mstamp
.v64
= 0;
1339 tcp_unlink_write_queue(skb
, sk
);
1340 sk_wmem_free_skb(sk
, skb
);
1342 NET_INC_STATS(sock_net(sk
), LINUX_MIB_SACKMERGED
);
1347 /* I wish gso_size would have a bit more sane initialization than
1348 * something-or-zero which complicates things
1350 static int tcp_skb_seglen(const struct sk_buff
*skb
)
1352 return tcp_skb_pcount(skb
) == 1 ? skb
->len
: tcp_skb_mss(skb
);
1355 /* Shifting pages past head area doesn't work */
1356 static int skb_can_shift(const struct sk_buff
*skb
)
1358 return !skb_headlen(skb
) && skb_is_nonlinear(skb
);
1361 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1364 static struct sk_buff
*tcp_shift_skb_data(struct sock
*sk
, struct sk_buff
*skb
,
1365 struct tcp_sacktag_state
*state
,
1366 u32 start_seq
, u32 end_seq
,
1369 struct tcp_sock
*tp
= tcp_sk(sk
);
1370 struct sk_buff
*prev
;
1376 if (!sk_can_gso(sk
))
1379 /* Normally R but no L won't result in plain S */
1381 (TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_RETRANS
)) == TCPCB_SACKED_RETRANS
)
1383 if (!skb_can_shift(skb
))
1385 /* This frame is about to be dropped (was ACKed). */
1386 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1389 /* Can only happen with delayed DSACK + discard craziness */
1390 if (unlikely(skb
== tcp_write_queue_head(sk
)))
1392 prev
= tcp_write_queue_prev(sk
, skb
);
1394 if ((TCP_SKB_CB(prev
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
)
1397 if (!tcp_skb_can_collapse_to(prev
))
1400 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1401 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1405 pcount
= tcp_skb_pcount(skb
);
1406 mss
= tcp_skb_seglen(skb
);
1408 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1409 * drop this restriction as unnecessary
1411 if (mss
!= tcp_skb_seglen(prev
))
1414 if (!after(TCP_SKB_CB(skb
)->end_seq
, start_seq
))
1416 /* CHECKME: This is non-MSS split case only?, this will
1417 * cause skipped skbs due to advancing loop btw, original
1418 * has that feature too
1420 if (tcp_skb_pcount(skb
) <= 1)
1423 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1425 /* TODO: head merge to next could be attempted here
1426 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1427 * though it might not be worth of the additional hassle
1429 * ...we can probably just fallback to what was done
1430 * previously. We could try merging non-SACKed ones
1431 * as well but it probably isn't going to buy off
1432 * because later SACKs might again split them, and
1433 * it would make skb timestamp tracking considerably
1439 len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1441 BUG_ON(len
> skb
->len
);
1443 /* MSS boundaries should be honoured or else pcount will
1444 * severely break even though it makes things bit trickier.
1445 * Optimize common case to avoid most of the divides
1447 mss
= tcp_skb_mss(skb
);
1449 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1450 * drop this restriction as unnecessary
1452 if (mss
!= tcp_skb_seglen(prev
))
1457 } else if (len
< mss
) {
1465 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1466 if (!after(TCP_SKB_CB(skb
)->seq
+ len
, tp
->snd_una
))
1469 if (!skb_shift(prev
, skb
, len
))
1471 if (!tcp_shifted_skb(sk
, skb
, state
, pcount
, len
, mss
, dup_sack
))
1474 /* Hole filled allows collapsing with the next as well, this is very
1475 * useful when hole on every nth skb pattern happens
1477 if (prev
== tcp_write_queue_tail(sk
))
1479 skb
= tcp_write_queue_next(sk
, prev
);
1481 if (!skb_can_shift(skb
) ||
1482 (skb
== tcp_send_head(sk
)) ||
1483 ((TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
) ||
1484 (mss
!= tcp_skb_seglen(skb
)))
1488 if (skb_shift(prev
, skb
, len
)) {
1489 pcount
+= tcp_skb_pcount(skb
);
1490 tcp_shifted_skb(sk
, skb
, state
, tcp_skb_pcount(skb
), len
, mss
, 0);
1494 state
->fack_count
+= pcount
;
1501 NET_INC_STATS(sock_net(sk
), LINUX_MIB_SACKSHIFTFALLBACK
);
1505 static struct sk_buff
*tcp_sacktag_walk(struct sk_buff
*skb
, struct sock
*sk
,
1506 struct tcp_sack_block
*next_dup
,
1507 struct tcp_sacktag_state
*state
,
1508 u32 start_seq
, u32 end_seq
,
1511 struct tcp_sock
*tp
= tcp_sk(sk
);
1512 struct sk_buff
*tmp
;
1514 tcp_for_write_queue_from(skb
, sk
) {
1516 bool dup_sack
= dup_sack_in
;
1518 if (skb
== tcp_send_head(sk
))
1521 /* queue is in-order => we can short-circuit the walk early */
1522 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1526 before(TCP_SKB_CB(skb
)->seq
, next_dup
->end_seq
)) {
1527 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1528 next_dup
->start_seq
,
1534 /* skb reference here is a bit tricky to get right, since
1535 * shifting can eat and free both this skb and the next,
1536 * so not even _safe variant of the loop is enough.
1539 tmp
= tcp_shift_skb_data(sk
, skb
, state
,
1540 start_seq
, end_seq
, dup_sack
);
1549 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1555 if (unlikely(in_sack
< 0))
1559 TCP_SKB_CB(skb
)->sacked
=
1562 TCP_SKB_CB(skb
)->sacked
,
1563 TCP_SKB_CB(skb
)->seq
,
1564 TCP_SKB_CB(skb
)->end_seq
,
1566 tcp_skb_pcount(skb
),
1568 tcp_rate_skb_delivered(sk
, skb
, state
->rate
);
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
,
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
, struct tcp_sacktag_state
*state
)
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 sk_buff
*skb
;
1634 int num_sacks
= min(TCP_NUM_SACKS
, (ptr
[1] - TCPOLEN_SACK_BASE
) >> 3);
1636 bool found_dup_sack
= false;
1638 int first_sack_index
;
1641 state
->reord
= tp
->packets_out
;
1643 if (!tp
->sacked_out
) {
1644 if (WARN_ON(tp
->fackets_out
))
1645 tp
->fackets_out
= 0;
1646 tcp_highest_sack_reset(sk
);
1649 found_dup_sack
= tcp_check_dsack(sk
, ack_skb
, sp_wire
,
1650 num_sacks
, prior_snd_una
);
1651 if (found_dup_sack
) {
1652 state
->flag
|= FLAG_DSACKING_ACK
;
1653 tp
->delivered
++; /* A spurious retransmission is delivered */
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(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 if ((state
->reord
< tp
->fackets_out
) &&
1812 ((inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Loss
) || tp
->undo_marker
))
1813 tcp_update_reordering(sk
, tp
->fackets_out
- state
->reord
, 0);
1815 tcp_verify_left_out(tp
);
1818 #if FASTRETRANS_DEBUG > 0
1819 WARN_ON((int)tp
->sacked_out
< 0);
1820 WARN_ON((int)tp
->lost_out
< 0);
1821 WARN_ON((int)tp
->retrans_out
< 0);
1822 WARN_ON((int)tcp_packets_in_flight(tp
) < 0);
1827 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1828 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1830 static bool tcp_limit_reno_sacked(struct tcp_sock
*tp
)
1834 holes
= max(tp
->lost_out
, 1U);
1835 holes
= min(holes
, tp
->packets_out
);
1837 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1838 tp
->sacked_out
= tp
->packets_out
- holes
;
1844 /* If we receive more dupacks than we expected counting segments
1845 * in assumption of absent reordering, interpret this as reordering.
1846 * The only another reason could be bug in receiver TCP.
1848 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1850 struct tcp_sock
*tp
= tcp_sk(sk
);
1851 if (tcp_limit_reno_sacked(tp
))
1852 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1855 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1857 static void tcp_add_reno_sack(struct sock
*sk
)
1859 struct tcp_sock
*tp
= tcp_sk(sk
);
1860 u32 prior_sacked
= tp
->sacked_out
;
1863 tcp_check_reno_reordering(sk
, 0);
1864 if (tp
->sacked_out
> prior_sacked
)
1865 tp
->delivered
++; /* Some out-of-order packet is delivered */
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 tp
->delivered
+= max_t(int, acked
- tp
->sacked_out
, 1);
1878 if (acked
- 1 >= tp
->sacked_out
)
1881 tp
->sacked_out
-= acked
- 1;
1883 tcp_check_reno_reordering(sk
, acked
);
1884 tcp_verify_left_out(tp
);
1887 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1892 void tcp_clear_retrans(struct tcp_sock
*tp
)
1894 tp
->retrans_out
= 0;
1896 tp
->undo_marker
= 0;
1897 tp
->undo_retrans
= -1;
1898 tp
->fackets_out
= 0;
1902 static inline void tcp_init_undo(struct tcp_sock
*tp
)
1904 tp
->undo_marker
= tp
->snd_una
;
1905 /* Retransmission still in flight may cause DSACKs later. */
1906 tp
->undo_retrans
= tp
->retrans_out
? : -1;
1909 /* Enter Loss state. If we detect SACK reneging, forget all SACK information
1910 * and reset tags completely, otherwise preserve SACKs. If receiver
1911 * dropped its ofo queue, we will know this due to reneging detection.
1913 void tcp_enter_loss(struct sock
*sk
)
1915 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1916 struct tcp_sock
*tp
= tcp_sk(sk
);
1917 struct net
*net
= sock_net(sk
);
1918 struct sk_buff
*skb
;
1919 bool new_recovery
= icsk
->icsk_ca_state
< TCP_CA_Recovery
;
1920 bool is_reneg
; /* is receiver reneging on SACKs? */
1923 /* Reduce ssthresh if it has not yet been made inside this window. */
1924 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
||
1925 !after(tp
->high_seq
, tp
->snd_una
) ||
1926 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
1927 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1928 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1929 tcp_ca_event(sk
, CA_EVENT_LOSS
);
1933 tp
->snd_cwnd_cnt
= 0;
1934 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1936 tp
->retrans_out
= 0;
1939 if (tcp_is_reno(tp
))
1940 tcp_reset_reno_sack(tp
);
1942 skb
= tcp_write_queue_head(sk
);
1943 is_reneg
= skb
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
);
1945 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPSACKRENEGING
);
1947 tp
->fackets_out
= 0;
1949 tcp_clear_all_retrans_hints(tp
);
1951 tcp_for_write_queue(skb
, sk
) {
1952 if (skb
== tcp_send_head(sk
))
1955 mark_lost
= (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
) ||
1958 tcp_sum_lost(tp
, skb
);
1959 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
1961 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
1962 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1963 tp
->lost_out
+= tcp_skb_pcount(skb
);
1964 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
1967 tcp_verify_left_out(tp
);
1969 /* Timeout in disordered state after receiving substantial DUPACKs
1970 * suggests that the degree of reordering is over-estimated.
1972 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
&&
1973 tp
->sacked_out
>= net
->ipv4
.sysctl_tcp_reordering
)
1974 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1975 net
->ipv4
.sysctl_tcp_reordering
);
1976 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1977 tp
->high_seq
= tp
->snd_nxt
;
1978 tcp_ecn_queue_cwr(tp
);
1980 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1981 * loss recovery is underway except recurring timeout(s) on
1982 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1984 tp
->frto
= sysctl_tcp_frto
&&
1985 (new_recovery
|| icsk
->icsk_retransmits
) &&
1986 !inet_csk(sk
)->icsk_mtup
.probe_size
;
1989 /* If ACK arrived pointing to a remembered SACK, it means that our
1990 * remembered SACKs do not reflect real state of receiver i.e.
1991 * receiver _host_ is heavily congested (or buggy).
1993 * To avoid big spurious retransmission bursts due to transient SACK
1994 * scoreboard oddities that look like reneging, we give the receiver a
1995 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
1996 * restore sanity to the SACK scoreboard. If the apparent reneging
1997 * persists until this RTO then we'll clear the SACK scoreboard.
1999 static bool tcp_check_sack_reneging(struct sock
*sk
, int flag
)
2001 if (flag
& FLAG_SACK_RENEGING
) {
2002 struct tcp_sock
*tp
= tcp_sk(sk
);
2003 unsigned long delay
= max(usecs_to_jiffies(tp
->srtt_us
>> 4),
2004 msecs_to_jiffies(10));
2006 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2007 delay
, TCP_RTO_MAX
);
2013 static inline int tcp_fackets_out(const struct tcp_sock
*tp
)
2015 return tcp_is_reno(tp
) ? tp
->sacked_out
+ 1 : tp
->fackets_out
;
2018 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2019 * counter when SACK is enabled (without SACK, sacked_out is used for
2022 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2023 * segments up to the highest received SACK block so far and holes in
2026 * With reordering, holes may still be in flight, so RFC3517 recovery
2027 * uses pure sacked_out (total number of SACKed segments) even though
2028 * it violates the RFC that uses duplicate ACKs, often these are equal
2029 * but when e.g. out-of-window ACKs or packet duplication occurs,
2030 * they differ. Since neither occurs due to loss, TCP should really
2033 static inline int tcp_dupack_heuristics(const struct tcp_sock
*tp
)
2035 return tcp_is_fack(tp
) ? tp
->fackets_out
: tp
->sacked_out
+ 1;
2038 static bool tcp_pause_early_retransmit(struct sock
*sk
, int flag
)
2040 struct tcp_sock
*tp
= tcp_sk(sk
);
2041 unsigned long delay
;
2043 /* Delay early retransmit and entering fast recovery for
2044 * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
2045 * available, or RTO is scheduled to fire first.
2047 if (sysctl_tcp_early_retrans
< 2 || sysctl_tcp_early_retrans
> 3 ||
2048 (flag
& FLAG_ECE
) || !tp
->srtt_us
)
2051 delay
= max(usecs_to_jiffies(tp
->srtt_us
>> 5),
2052 msecs_to_jiffies(2));
2054 if (!time_after(inet_csk(sk
)->icsk_timeout
, (jiffies
+ delay
)))
2057 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_EARLY_RETRANS
, delay
,
2062 /* Linux NewReno/SACK/FACK/ECN state machine.
2063 * --------------------------------------
2065 * "Open" Normal state, no dubious events, fast path.
2066 * "Disorder" In all the respects it is "Open",
2067 * but requires a bit more attention. It is entered when
2068 * we see some SACKs or dupacks. It is split of "Open"
2069 * mainly to move some processing from fast path to slow one.
2070 * "CWR" CWND was reduced due to some Congestion Notification event.
2071 * It can be ECN, ICMP source quench, local device congestion.
2072 * "Recovery" CWND was reduced, we are fast-retransmitting.
2073 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2075 * tcp_fastretrans_alert() is entered:
2076 * - each incoming ACK, if state is not "Open"
2077 * - when arrived ACK is unusual, namely:
2082 * Counting packets in flight is pretty simple.
2084 * in_flight = packets_out - left_out + retrans_out
2086 * packets_out is SND.NXT-SND.UNA counted in packets.
2088 * retrans_out is number of retransmitted segments.
2090 * left_out is number of segments left network, but not ACKed yet.
2092 * left_out = sacked_out + lost_out
2094 * sacked_out: Packets, which arrived to receiver out of order
2095 * and hence not ACKed. With SACKs this number is simply
2096 * amount of SACKed data. Even without SACKs
2097 * it is easy to give pretty reliable estimate of this number,
2098 * counting duplicate ACKs.
2100 * lost_out: Packets lost by network. TCP has no explicit
2101 * "loss notification" feedback from network (for now).
2102 * It means that this number can be only _guessed_.
2103 * Actually, it is the heuristics to predict lossage that
2104 * distinguishes different algorithms.
2106 * F.e. after RTO, when all the queue is considered as lost,
2107 * lost_out = packets_out and in_flight = retrans_out.
2109 * Essentially, we have now two algorithms counting
2112 * FACK: It is the simplest heuristics. As soon as we decided
2113 * that something is lost, we decide that _all_ not SACKed
2114 * packets until the most forward SACK are lost. I.e.
2115 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2116 * It is absolutely correct estimate, if network does not reorder
2117 * packets. And it loses any connection to reality when reordering
2118 * takes place. We use FACK by default until reordering
2119 * is suspected on the path to this destination.
2121 * NewReno: when Recovery is entered, we assume that one segment
2122 * is lost (classic Reno). While we are in Recovery and
2123 * a partial ACK arrives, we assume that one more packet
2124 * is lost (NewReno). This heuristics are the same in NewReno
2127 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2128 * deflation etc. CWND is real congestion window, never inflated, changes
2129 * only according to classic VJ rules.
2131 * Really tricky (and requiring careful tuning) part of algorithm
2132 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2133 * The first determines the moment _when_ we should reduce CWND and,
2134 * hence, slow down forward transmission. In fact, it determines the moment
2135 * when we decide that hole is caused by loss, rather than by a reorder.
2137 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2138 * holes, caused by lost packets.
2140 * And the most logically complicated part of algorithm is undo
2141 * heuristics. We detect false retransmits due to both too early
2142 * fast retransmit (reordering) and underestimated RTO, analyzing
2143 * timestamps and D-SACKs. When we detect that some segments were
2144 * retransmitted by mistake and CWND reduction was wrong, we undo
2145 * window reduction and abort recovery phase. This logic is hidden
2146 * inside several functions named tcp_try_undo_<something>.
2149 /* This function decides, when we should leave Disordered state
2150 * and enter Recovery phase, reducing congestion window.
2152 * Main question: may we further continue forward transmission
2153 * with the same cwnd?
2155 static bool tcp_time_to_recover(struct sock
*sk
, int flag
)
2157 struct tcp_sock
*tp
= tcp_sk(sk
);
2159 int tcp_reordering
= sock_net(sk
)->ipv4
.sysctl_tcp_reordering
;
2161 /* Trick#1: The loss is proven. */
2165 /* Not-A-Trick#2 : Classic rule... */
2166 if (tcp_dupack_heuristics(tp
) > tp
->reordering
)
2169 /* Trick#4: It is still not OK... But will it be useful to delay
2172 packets_out
= tp
->packets_out
;
2173 if (packets_out
<= tp
->reordering
&&
2174 tp
->sacked_out
>= max_t(__u32
, packets_out
/2, tcp_reordering
) &&
2175 !tcp_may_send_now(sk
)) {
2176 /* We have nothing to send. This connection is limited
2177 * either by receiver window or by application.
2182 /* If a thin stream is detected, retransmit after first
2183 * received dupack. Employ only if SACK is supported in order
2184 * to avoid possible corner-case series of spurious retransmissions
2185 * Use only if there are no unsent data.
2187 if ((tp
->thin_dupack
|| sysctl_tcp_thin_dupack
) &&
2188 tcp_stream_is_thin(tp
) && tcp_dupack_heuristics(tp
) > 1 &&
2189 tcp_is_sack(tp
) && !tcp_send_head(sk
))
2192 /* Trick#6: TCP early retransmit, per RFC5827. To avoid spurious
2193 * retransmissions due to small network reorderings, we implement
2194 * Mitigation A.3 in the RFC and delay the retransmission for a short
2195 * interval if appropriate.
2197 if (tp
->do_early_retrans
&& !tp
->retrans_out
&& tp
->sacked_out
&&
2198 (tp
->packets_out
>= (tp
->sacked_out
+ 1) && tp
->packets_out
< 4) &&
2199 !tcp_may_send_now(sk
))
2200 return !tcp_pause_early_retransmit(sk
, flag
);
2205 /* Detect loss in event "A" above by marking head of queue up as lost.
2206 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2207 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2208 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2209 * the maximum SACKed segments to pass before reaching this limit.
2211 static void tcp_mark_head_lost(struct sock
*sk
, int packets
, int mark_head
)
2213 struct tcp_sock
*tp
= tcp_sk(sk
);
2214 struct sk_buff
*skb
;
2215 int cnt
, oldcnt
, lost
;
2217 /* Use SACK to deduce losses of new sequences sent during recovery */
2218 const u32 loss_high
= tcp_is_sack(tp
) ? tp
->snd_nxt
: tp
->high_seq
;
2220 WARN_ON(packets
> tp
->packets_out
);
2221 if (tp
->lost_skb_hint
) {
2222 skb
= tp
->lost_skb_hint
;
2223 cnt
= tp
->lost_cnt_hint
;
2224 /* Head already handled? */
2225 if (mark_head
&& skb
!= tcp_write_queue_head(sk
))
2228 skb
= tcp_write_queue_head(sk
);
2232 tcp_for_write_queue_from(skb
, sk
) {
2233 if (skb
== tcp_send_head(sk
))
2235 /* TODO: do this better */
2236 /* this is not the most efficient way to do this... */
2237 tp
->lost_skb_hint
= skb
;
2238 tp
->lost_cnt_hint
= cnt
;
2240 if (after(TCP_SKB_CB(skb
)->end_seq
, loss_high
))
2244 if (tcp_is_fack(tp
) || tcp_is_reno(tp
) ||
2245 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2246 cnt
+= tcp_skb_pcount(skb
);
2248 if (cnt
> packets
) {
2249 if ((tcp_is_sack(tp
) && !tcp_is_fack(tp
)) ||
2250 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
) ||
2251 (oldcnt
>= packets
))
2254 mss
= tcp_skb_mss(skb
);
2255 /* If needed, chop off the prefix to mark as lost. */
2256 lost
= (packets
- oldcnt
) * mss
;
2257 if (lost
< skb
->len
&&
2258 tcp_fragment(sk
, skb
, lost
, mss
, GFP_ATOMIC
) < 0)
2263 tcp_skb_mark_lost(tp
, skb
);
2268 tcp_verify_left_out(tp
);
2271 /* Account newly detected lost packet(s) */
2273 static void tcp_update_scoreboard(struct sock
*sk
, int fast_rexmit
)
2275 struct tcp_sock
*tp
= tcp_sk(sk
);
2277 if (tcp_is_reno(tp
)) {
2278 tcp_mark_head_lost(sk
, 1, 1);
2279 } else if (tcp_is_fack(tp
)) {
2280 int lost
= tp
->fackets_out
- tp
->reordering
;
2283 tcp_mark_head_lost(sk
, lost
, 0);
2285 int sacked_upto
= tp
->sacked_out
- tp
->reordering
;
2286 if (sacked_upto
>= 0)
2287 tcp_mark_head_lost(sk
, sacked_upto
, 0);
2288 else if (fast_rexmit
)
2289 tcp_mark_head_lost(sk
, 1, 1);
2293 static bool tcp_tsopt_ecr_before(const struct tcp_sock
*tp
, u32 when
)
2295 return tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2296 before(tp
->rx_opt
.rcv_tsecr
, when
);
2299 /* skb is spurious retransmitted if the returned timestamp echo
2300 * reply is prior to the skb transmission time
2302 static bool tcp_skb_spurious_retrans(const struct tcp_sock
*tp
,
2303 const struct sk_buff
*skb
)
2305 return (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
) &&
2306 tcp_tsopt_ecr_before(tp
, tcp_skb_timestamp(skb
));
2309 /* Nothing was retransmitted or returned timestamp is less
2310 * than timestamp of the first retransmission.
2312 static inline bool tcp_packet_delayed(const struct tcp_sock
*tp
)
2314 return !tp
->retrans_stamp
||
2315 tcp_tsopt_ecr_before(tp
, tp
->retrans_stamp
);
2318 /* Undo procedures. */
2320 /* We can clear retrans_stamp when there are no retransmissions in the
2321 * window. It would seem that it is trivially available for us in
2322 * tp->retrans_out, however, that kind of assumptions doesn't consider
2323 * what will happen if errors occur when sending retransmission for the
2324 * second time. ...It could the that such segment has only
2325 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2326 * the head skb is enough except for some reneging corner cases that
2327 * are not worth the effort.
2329 * Main reason for all this complexity is the fact that connection dying
2330 * time now depends on the validity of the retrans_stamp, in particular,
2331 * that successive retransmissions of a segment must not advance
2332 * retrans_stamp under any conditions.
2334 static bool tcp_any_retrans_done(const struct sock
*sk
)
2336 const struct tcp_sock
*tp
= tcp_sk(sk
);
2337 struct sk_buff
*skb
;
2339 if (tp
->retrans_out
)
2342 skb
= tcp_write_queue_head(sk
);
2343 if (unlikely(skb
&& TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
))
2349 #if FASTRETRANS_DEBUG > 1
2350 static void DBGUNDO(struct sock
*sk
, const char *msg
)
2352 struct tcp_sock
*tp
= tcp_sk(sk
);
2353 struct inet_sock
*inet
= inet_sk(sk
);
2355 if (sk
->sk_family
== AF_INET
) {
2356 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2358 &inet
->inet_daddr
, ntohs(inet
->inet_dport
),
2359 tp
->snd_cwnd
, tcp_left_out(tp
),
2360 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2363 #if IS_ENABLED(CONFIG_IPV6)
2364 else if (sk
->sk_family
== AF_INET6
) {
2365 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2367 &sk
->sk_v6_daddr
, ntohs(inet
->inet_dport
),
2368 tp
->snd_cwnd
, tcp_left_out(tp
),
2369 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2375 #define DBGUNDO(x...) do { } while (0)
2378 static void tcp_undo_cwnd_reduction(struct sock
*sk
, bool unmark_loss
)
2380 struct tcp_sock
*tp
= tcp_sk(sk
);
2383 struct sk_buff
*skb
;
2385 tcp_for_write_queue(skb
, sk
) {
2386 if (skb
== tcp_send_head(sk
))
2388 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2391 tcp_clear_all_retrans_hints(tp
);
2394 if (tp
->prior_ssthresh
) {
2395 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2397 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
2399 if (tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
2400 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
2401 tcp_ecn_withdraw_cwr(tp
);
2404 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2405 tp
->undo_marker
= 0;
2408 static inline bool tcp_may_undo(const struct tcp_sock
*tp
)
2410 return tp
->undo_marker
&& (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
2413 /* People celebrate: "We love our President!" */
2414 static bool tcp_try_undo_recovery(struct sock
*sk
)
2416 struct tcp_sock
*tp
= tcp_sk(sk
);
2418 if (tcp_may_undo(tp
)) {
2421 /* Happy end! We did not retransmit anything
2422 * or our original transmission succeeded.
2424 DBGUNDO(sk
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
2425 tcp_undo_cwnd_reduction(sk
, false);
2426 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
2427 mib_idx
= LINUX_MIB_TCPLOSSUNDO
;
2429 mib_idx
= LINUX_MIB_TCPFULLUNDO
;
2431 NET_INC_STATS(sock_net(sk
), mib_idx
);
2433 if (tp
->snd_una
== tp
->high_seq
&& tcp_is_reno(tp
)) {
2434 /* Hold old state until something *above* high_seq
2435 * is ACKed. For Reno it is MUST to prevent false
2436 * fast retransmits (RFC2582). SACK TCP is safe. */
2437 if (!tcp_any_retrans_done(sk
))
2438 tp
->retrans_stamp
= 0;
2441 tcp_set_ca_state(sk
, TCP_CA_Open
);
2445 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2446 static bool tcp_try_undo_dsack(struct sock
*sk
)
2448 struct tcp_sock
*tp
= tcp_sk(sk
);
2450 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
2451 DBGUNDO(sk
, "D-SACK");
2452 tcp_undo_cwnd_reduction(sk
, false);
2453 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPDSACKUNDO
);
2459 /* Undo during loss recovery after partial ACK or using F-RTO. */
2460 static bool tcp_try_undo_loss(struct sock
*sk
, bool frto_undo
)
2462 struct tcp_sock
*tp
= tcp_sk(sk
);
2464 if (frto_undo
|| tcp_may_undo(tp
)) {
2465 tcp_undo_cwnd_reduction(sk
, true);
2467 DBGUNDO(sk
, "partial loss");
2468 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPLOSSUNDO
);
2470 NET_INC_STATS(sock_net(sk
),
2471 LINUX_MIB_TCPSPURIOUSRTOS
);
2472 inet_csk(sk
)->icsk_retransmits
= 0;
2473 if (frto_undo
|| tcp_is_sack(tp
))
2474 tcp_set_ca_state(sk
, TCP_CA_Open
);
2480 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2481 * It computes the number of packets to send (sndcnt) based on packets newly
2483 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2484 * cwnd reductions across a full RTT.
2485 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2486 * But when the retransmits are acked without further losses, PRR
2487 * slow starts cwnd up to ssthresh to speed up the recovery.
2489 static void tcp_init_cwnd_reduction(struct sock
*sk
)
2491 struct tcp_sock
*tp
= tcp_sk(sk
);
2493 tp
->high_seq
= tp
->snd_nxt
;
2494 tp
->tlp_high_seq
= 0;
2495 tp
->snd_cwnd_cnt
= 0;
2496 tp
->prior_cwnd
= tp
->snd_cwnd
;
2497 tp
->prr_delivered
= 0;
2499 tp
->snd_ssthresh
= inet_csk(sk
)->icsk_ca_ops
->ssthresh(sk
);
2500 tcp_ecn_queue_cwr(tp
);
2503 static void tcp_cwnd_reduction(struct sock
*sk
, int newly_acked_sacked
,
2506 struct tcp_sock
*tp
= tcp_sk(sk
);
2508 int delta
= tp
->snd_ssthresh
- tcp_packets_in_flight(tp
);
2510 if (newly_acked_sacked
<= 0 || WARN_ON_ONCE(!tp
->prior_cwnd
))
2513 tp
->prr_delivered
+= newly_acked_sacked
;
2515 u64 dividend
= (u64
)tp
->snd_ssthresh
* tp
->prr_delivered
+
2517 sndcnt
= div_u64(dividend
, tp
->prior_cwnd
) - tp
->prr_out
;
2518 } else if ((flag
& FLAG_RETRANS_DATA_ACKED
) &&
2519 !(flag
& FLAG_LOST_RETRANS
)) {
2520 sndcnt
= min_t(int, delta
,
2521 max_t(int, tp
->prr_delivered
- tp
->prr_out
,
2522 newly_acked_sacked
) + 1);
2524 sndcnt
= min(delta
, newly_acked_sacked
);
2526 /* Force a fast retransmit upon entering fast recovery */
2527 sndcnt
= max(sndcnt
, (tp
->prr_out
? 0 : 1));
2528 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + sndcnt
;
2531 static inline void tcp_end_cwnd_reduction(struct sock
*sk
)
2533 struct tcp_sock
*tp
= tcp_sk(sk
);
2535 if (inet_csk(sk
)->icsk_ca_ops
->cong_control
)
2538 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2539 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_CWR
||
2540 (tp
->undo_marker
&& tp
->snd_ssthresh
< TCP_INFINITE_SSTHRESH
)) {
2541 tp
->snd_cwnd
= tp
->snd_ssthresh
;
2542 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2544 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
2547 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2548 void tcp_enter_cwr(struct sock
*sk
)
2550 struct tcp_sock
*tp
= tcp_sk(sk
);
2552 tp
->prior_ssthresh
= 0;
2553 if (inet_csk(sk
)->icsk_ca_state
< TCP_CA_CWR
) {
2554 tp
->undo_marker
= 0;
2555 tcp_init_cwnd_reduction(sk
);
2556 tcp_set_ca_state(sk
, TCP_CA_CWR
);
2559 EXPORT_SYMBOL(tcp_enter_cwr
);
2561 static void tcp_try_keep_open(struct sock
*sk
)
2563 struct tcp_sock
*tp
= tcp_sk(sk
);
2564 int state
= TCP_CA_Open
;
2566 if (tcp_left_out(tp
) || tcp_any_retrans_done(sk
))
2567 state
= TCP_CA_Disorder
;
2569 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
2570 tcp_set_ca_state(sk
, state
);
2571 tp
->high_seq
= tp
->snd_nxt
;
2575 static void tcp_try_to_open(struct sock
*sk
, int flag
)
2577 struct tcp_sock
*tp
= tcp_sk(sk
);
2579 tcp_verify_left_out(tp
);
2581 if (!tcp_any_retrans_done(sk
))
2582 tp
->retrans_stamp
= 0;
2584 if (flag
& FLAG_ECE
)
2587 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
2588 tcp_try_keep_open(sk
);
2592 static void tcp_mtup_probe_failed(struct sock
*sk
)
2594 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2596 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
2597 icsk
->icsk_mtup
.probe_size
= 0;
2598 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPMTUPFAIL
);
2601 static void tcp_mtup_probe_success(struct sock
*sk
)
2603 struct tcp_sock
*tp
= tcp_sk(sk
);
2604 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2606 /* FIXME: breaks with very large cwnd */
2607 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2608 tp
->snd_cwnd
= tp
->snd_cwnd
*
2609 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
2610 icsk
->icsk_mtup
.probe_size
;
2611 tp
->snd_cwnd_cnt
= 0;
2612 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2613 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2615 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
2616 icsk
->icsk_mtup
.probe_size
= 0;
2617 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
2618 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPMTUPSUCCESS
);
2621 /* Do a simple retransmit without using the backoff mechanisms in
2622 * tcp_timer. This is used for path mtu discovery.
2623 * The socket is already locked here.
2625 void tcp_simple_retransmit(struct sock
*sk
)
2627 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2628 struct tcp_sock
*tp
= tcp_sk(sk
);
2629 struct sk_buff
*skb
;
2630 unsigned int mss
= tcp_current_mss(sk
);
2631 u32 prior_lost
= tp
->lost_out
;
2633 tcp_for_write_queue(skb
, sk
) {
2634 if (skb
== tcp_send_head(sk
))
2636 if (tcp_skb_seglen(skb
) > mss
&&
2637 !(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
2638 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2639 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
2640 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2642 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
2646 tcp_clear_retrans_hints_partial(tp
);
2648 if (prior_lost
== tp
->lost_out
)
2651 if (tcp_is_reno(tp
))
2652 tcp_limit_reno_sacked(tp
);
2654 tcp_verify_left_out(tp
);
2656 /* Don't muck with the congestion window here.
2657 * Reason is that we do not increase amount of _data_
2658 * in network, but units changed and effective
2659 * cwnd/ssthresh really reduced now.
2661 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
2662 tp
->high_seq
= tp
->snd_nxt
;
2663 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2664 tp
->prior_ssthresh
= 0;
2665 tp
->undo_marker
= 0;
2666 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2668 tcp_xmit_retransmit_queue(sk
);
2670 EXPORT_SYMBOL(tcp_simple_retransmit
);
2672 static void tcp_enter_recovery(struct sock
*sk
, bool ece_ack
)
2674 struct tcp_sock
*tp
= tcp_sk(sk
);
2677 if (tcp_is_reno(tp
))
2678 mib_idx
= LINUX_MIB_TCPRENORECOVERY
;
2680 mib_idx
= LINUX_MIB_TCPSACKRECOVERY
;
2682 NET_INC_STATS(sock_net(sk
), mib_idx
);
2684 tp
->prior_ssthresh
= 0;
2687 if (!tcp_in_cwnd_reduction(sk
)) {
2689 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2690 tcp_init_cwnd_reduction(sk
);
2692 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
2695 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2696 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2698 static void tcp_process_loss(struct sock
*sk
, int flag
, bool is_dupack
,
2701 struct tcp_sock
*tp
= tcp_sk(sk
);
2702 bool recovered
= !before(tp
->snd_una
, tp
->high_seq
);
2704 if ((flag
& FLAG_SND_UNA_ADVANCED
) &&
2705 tcp_try_undo_loss(sk
, false))
2708 if (tp
->frto
) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2709 /* Step 3.b. A timeout is spurious if not all data are
2710 * lost, i.e., never-retransmitted data are (s)acked.
2712 if ((flag
& FLAG_ORIG_SACK_ACKED
) &&
2713 tcp_try_undo_loss(sk
, true))
2716 if (after(tp
->snd_nxt
, tp
->high_seq
)) {
2717 if (flag
& FLAG_DATA_SACKED
|| is_dupack
)
2718 tp
->frto
= 0; /* Step 3.a. loss was real */
2719 } else if (flag
& FLAG_SND_UNA_ADVANCED
&& !recovered
) {
2720 tp
->high_seq
= tp
->snd_nxt
;
2721 /* Step 2.b. Try send new data (but deferred until cwnd
2722 * is updated in tcp_ack()). Otherwise fall back to
2723 * the conventional recovery.
2725 if (tcp_send_head(sk
) &&
2726 after(tcp_wnd_end(tp
), tp
->snd_nxt
)) {
2727 *rexmit
= REXMIT_NEW
;
2735 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2736 tcp_try_undo_recovery(sk
);
2739 if (tcp_is_reno(tp
)) {
2740 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2741 * delivered. Lower inflight to clock out (re)tranmissions.
2743 if (after(tp
->snd_nxt
, tp
->high_seq
) && is_dupack
)
2744 tcp_add_reno_sack(sk
);
2745 else if (flag
& FLAG_SND_UNA_ADVANCED
)
2746 tcp_reset_reno_sack(tp
);
2748 *rexmit
= REXMIT_LOST
;
2751 /* Undo during fast recovery after partial ACK. */
2752 static bool tcp_try_undo_partial(struct sock
*sk
, const int acked
)
2754 struct tcp_sock
*tp
= tcp_sk(sk
);
2756 if (tp
->undo_marker
&& tcp_packet_delayed(tp
)) {
2757 /* Plain luck! Hole if filled with delayed
2758 * packet, rather than with a retransmit.
2760 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
2762 /* We are getting evidence that the reordering degree is higher
2763 * than we realized. If there are no retransmits out then we
2764 * can undo. Otherwise we clock out new packets but do not
2765 * mark more packets lost or retransmit more.
2767 if (tp
->retrans_out
)
2770 if (!tcp_any_retrans_done(sk
))
2771 tp
->retrans_stamp
= 0;
2773 DBGUNDO(sk
, "partial recovery");
2774 tcp_undo_cwnd_reduction(sk
, true);
2775 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPPARTIALUNDO
);
2776 tcp_try_keep_open(sk
);
2782 /* Process an event, which can update packets-in-flight not trivially.
2783 * Main goal of this function is to calculate new estimate for left_out,
2784 * taking into account both packets sitting in receiver's buffer and
2785 * packets lost by network.
2787 * Besides that it updates the congestion state when packet loss or ECN
2788 * is detected. But it does not reduce the cwnd, it is done by the
2789 * congestion control later.
2791 * It does _not_ decide what to send, it is made in function
2792 * tcp_xmit_retransmit_queue().
2794 static void tcp_fastretrans_alert(struct sock
*sk
, const int acked
,
2795 bool is_dupack
, int *ack_flag
, int *rexmit
)
2797 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2798 struct tcp_sock
*tp
= tcp_sk(sk
);
2799 int fast_rexmit
= 0, flag
= *ack_flag
;
2800 bool do_lost
= is_dupack
|| ((flag
& FLAG_DATA_SACKED
) &&
2801 (tcp_fackets_out(tp
) > tp
->reordering
));
2803 if (WARN_ON(!tp
->packets_out
&& tp
->sacked_out
))
2805 if (WARN_ON(!tp
->sacked_out
&& tp
->fackets_out
))
2806 tp
->fackets_out
= 0;
2808 /* Now state machine starts.
2809 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2810 if (flag
& FLAG_ECE
)
2811 tp
->prior_ssthresh
= 0;
2813 /* B. In all the states check for reneging SACKs. */
2814 if (tcp_check_sack_reneging(sk
, flag
))
2817 /* C. Check consistency of the current state. */
2818 tcp_verify_left_out(tp
);
2820 /* D. Check state exit conditions. State can be terminated
2821 * when high_seq is ACKed. */
2822 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
2823 WARN_ON(tp
->retrans_out
!= 0);
2824 tp
->retrans_stamp
= 0;
2825 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
2826 switch (icsk
->icsk_ca_state
) {
2828 /* CWR is to be held something *above* high_seq
2829 * is ACKed for CWR bit to reach receiver. */
2830 if (tp
->snd_una
!= tp
->high_seq
) {
2831 tcp_end_cwnd_reduction(sk
);
2832 tcp_set_ca_state(sk
, TCP_CA_Open
);
2836 case TCP_CA_Recovery
:
2837 if (tcp_is_reno(tp
))
2838 tcp_reset_reno_sack(tp
);
2839 if (tcp_try_undo_recovery(sk
))
2841 tcp_end_cwnd_reduction(sk
);
2846 /* Use RACK to detect loss */
2847 if (sysctl_tcp_recovery
& TCP_RACK_LOST_RETRANS
&&
2848 tcp_rack_mark_lost(sk
)) {
2849 flag
|= FLAG_LOST_RETRANS
;
2850 *ack_flag
|= FLAG_LOST_RETRANS
;
2853 /* E. Process state. */
2854 switch (icsk
->icsk_ca_state
) {
2855 case TCP_CA_Recovery
:
2856 if (!(flag
& FLAG_SND_UNA_ADVANCED
)) {
2857 if (tcp_is_reno(tp
) && is_dupack
)
2858 tcp_add_reno_sack(sk
);
2860 if (tcp_try_undo_partial(sk
, acked
))
2862 /* Partial ACK arrived. Force fast retransmit. */
2863 do_lost
= tcp_is_reno(tp
) ||
2864 tcp_fackets_out(tp
) > tp
->reordering
;
2866 if (tcp_try_undo_dsack(sk
)) {
2867 tcp_try_keep_open(sk
);
2872 tcp_process_loss(sk
, flag
, is_dupack
, rexmit
);
2873 if (icsk
->icsk_ca_state
!= TCP_CA_Open
&&
2874 !(flag
& FLAG_LOST_RETRANS
))
2876 /* Change state if cwnd is undone or retransmits are lost */
2878 if (tcp_is_reno(tp
)) {
2879 if (flag
& FLAG_SND_UNA_ADVANCED
)
2880 tcp_reset_reno_sack(tp
);
2882 tcp_add_reno_sack(sk
);
2885 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
)
2886 tcp_try_undo_dsack(sk
);
2888 if (!tcp_time_to_recover(sk
, flag
)) {
2889 tcp_try_to_open(sk
, flag
);
2893 /* MTU probe failure: don't reduce cwnd */
2894 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
2895 icsk
->icsk_mtup
.probe_size
&&
2896 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
2897 tcp_mtup_probe_failed(sk
);
2898 /* Restores the reduction we did in tcp_mtup_probe() */
2900 tcp_simple_retransmit(sk
);
2904 /* Otherwise enter Recovery state */
2905 tcp_enter_recovery(sk
, (flag
& FLAG_ECE
));
2910 tcp_update_scoreboard(sk
, fast_rexmit
);
2911 *rexmit
= REXMIT_LOST
;
2914 static void tcp_update_rtt_min(struct sock
*sk
, u32 rtt_us
)
2916 struct tcp_sock
*tp
= tcp_sk(sk
);
2917 u32 wlen
= sysctl_tcp_min_rtt_wlen
* HZ
;
2919 minmax_running_min(&tp
->rtt_min
, wlen
, tcp_time_stamp
,
2920 rtt_us
? : jiffies_to_usecs(1));
2923 static inline bool tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
2924 long seq_rtt_us
, long sack_rtt_us
,
2927 const struct tcp_sock
*tp
= tcp_sk(sk
);
2929 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2930 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2931 * Karn's algorithm forbids taking RTT if some retransmitted data
2932 * is acked (RFC6298).
2935 seq_rtt_us
= sack_rtt_us
;
2937 /* RTTM Rule: A TSecr value received in a segment is used to
2938 * update the averaged RTT measurement only if the segment
2939 * acknowledges some new data, i.e., only if it advances the
2940 * left edge of the send window.
2941 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2943 if (seq_rtt_us
< 0 && tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2945 seq_rtt_us
= ca_rtt_us
= jiffies_to_usecs(tcp_time_stamp
-
2946 tp
->rx_opt
.rcv_tsecr
);
2950 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
2951 * always taken together with ACK, SACK, or TS-opts. Any negative
2952 * values will be skipped with the seq_rtt_us < 0 check above.
2954 tcp_update_rtt_min(sk
, ca_rtt_us
);
2955 tcp_rtt_estimator(sk
, seq_rtt_us
);
2958 /* RFC6298: only reset backoff on valid RTT measurement. */
2959 inet_csk(sk
)->icsk_backoff
= 0;
2963 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2964 void tcp_synack_rtt_meas(struct sock
*sk
, struct request_sock
*req
)
2968 if (req
&& !req
->num_retrans
&& tcp_rsk(req
)->snt_synack
.v64
) {
2969 struct skb_mstamp now
;
2971 skb_mstamp_get(&now
);
2972 rtt_us
= skb_mstamp_us_delta(&now
, &tcp_rsk(req
)->snt_synack
);
2975 tcp_ack_update_rtt(sk
, FLAG_SYN_ACKED
, rtt_us
, -1L, rtt_us
);
2979 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
, u32 acked
)
2981 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2983 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, acked
);
2984 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_time_stamp
;
2987 /* Restart timer after forward progress on connection.
2988 * RFC2988 recommends to restart timer to now+rto.
2990 void tcp_rearm_rto(struct sock
*sk
)
2992 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2993 struct tcp_sock
*tp
= tcp_sk(sk
);
2995 /* If the retrans timer is currently being used by Fast Open
2996 * for SYN-ACK retrans purpose, stay put.
2998 if (tp
->fastopen_rsk
)
3001 if (!tp
->packets_out
) {
3002 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
3004 u32 rto
= inet_csk(sk
)->icsk_rto
;
3005 /* Offset the time elapsed after installing regular RTO */
3006 if (icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
||
3007 icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
) {
3008 struct sk_buff
*skb
= tcp_write_queue_head(sk
);
3009 const u32 rto_time_stamp
=
3010 tcp_skb_timestamp(skb
) + rto
;
3011 s32 delta
= (s32
)(rto_time_stamp
- tcp_time_stamp
);
3012 /* delta may not be positive if the socket is locked
3013 * when the retrans timer fires and is rescheduled.
3018 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
, rto
,
3023 /* This function is called when the delayed ER timer fires. TCP enters
3024 * fast recovery and performs fast-retransmit.
3026 void tcp_resume_early_retransmit(struct sock
*sk
)
3028 struct tcp_sock
*tp
= tcp_sk(sk
);
3032 /* Stop if ER is disabled after the delayed ER timer is scheduled */
3033 if (!tp
->do_early_retrans
)
3036 tcp_enter_recovery(sk
, false);
3037 tcp_update_scoreboard(sk
, 1);
3038 tcp_xmit_retransmit_queue(sk
);
3041 /* If we get here, the whole TSO packet has not been acked. */
3042 static u32
tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
)
3044 struct tcp_sock
*tp
= tcp_sk(sk
);
3047 BUG_ON(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
));
3049 packets_acked
= tcp_skb_pcount(skb
);
3050 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
3052 packets_acked
-= tcp_skb_pcount(skb
);
3054 if (packets_acked
) {
3055 BUG_ON(tcp_skb_pcount(skb
) == 0);
3056 BUG_ON(!before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
));
3059 return packets_acked
;
3062 static void tcp_ack_tstamp(struct sock
*sk
, struct sk_buff
*skb
,
3065 const struct skb_shared_info
*shinfo
;
3067 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3068 if (likely(!TCP_SKB_CB(skb
)->txstamp_ack
))
3071 shinfo
= skb_shinfo(skb
);
3072 if (!before(shinfo
->tskey
, prior_snd_una
) &&
3073 before(shinfo
->tskey
, tcp_sk(sk
)->snd_una
))
3074 __skb_tstamp_tx(skb
, NULL
, sk
, SCM_TSTAMP_ACK
);
3077 /* Remove acknowledged frames from the retransmission queue. If our packet
3078 * is before the ack sequence we can discard it as it's confirmed to have
3079 * arrived at the other end.
3081 static int tcp_clean_rtx_queue(struct sock
*sk
, int prior_fackets
,
3082 u32 prior_snd_una
, int *acked
,
3083 struct tcp_sacktag_state
*sack
,
3084 struct skb_mstamp
*now
)
3086 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3087 struct skb_mstamp first_ackt
, last_ackt
;
3088 struct tcp_sock
*tp
= tcp_sk(sk
);
3089 u32 prior_sacked
= tp
->sacked_out
;
3090 u32 reord
= tp
->packets_out
;
3091 bool fully_acked
= true;
3092 long sack_rtt_us
= -1L;
3093 long seq_rtt_us
= -1L;
3094 long ca_rtt_us
= -1L;
3095 struct sk_buff
*skb
;
3097 u32 last_in_flight
= 0;
3103 while ((skb
= tcp_write_queue_head(sk
)) && skb
!= tcp_send_head(sk
)) {
3104 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
3105 u8 sacked
= scb
->sacked
;
3108 tcp_ack_tstamp(sk
, skb
, prior_snd_una
);
3110 /* Determine how many packets and what bytes were acked, tso and else */
3111 if (after(scb
->end_seq
, tp
->snd_una
)) {
3112 if (tcp_skb_pcount(skb
) == 1 ||
3113 !after(tp
->snd_una
, scb
->seq
))
3116 acked_pcount
= tcp_tso_acked(sk
, skb
);
3119 fully_acked
= false;
3121 /* Speedup tcp_unlink_write_queue() and next loop */
3122 prefetchw(skb
->next
);
3123 acked_pcount
= tcp_skb_pcount(skb
);
3126 if (unlikely(sacked
& TCPCB_RETRANS
)) {
3127 if (sacked
& TCPCB_SACKED_RETRANS
)
3128 tp
->retrans_out
-= acked_pcount
;
3129 flag
|= FLAG_RETRANS_DATA_ACKED
;
3130 } else if (!(sacked
& TCPCB_SACKED_ACKED
)) {
3131 last_ackt
= skb
->skb_mstamp
;
3132 WARN_ON_ONCE(last_ackt
.v64
== 0);
3133 if (!first_ackt
.v64
)
3134 first_ackt
= last_ackt
;
3136 last_in_flight
= TCP_SKB_CB(skb
)->tx
.in_flight
;
3137 reord
= min(pkts_acked
, reord
);
3138 if (!after(scb
->end_seq
, tp
->high_seq
))
3139 flag
|= FLAG_ORIG_SACK_ACKED
;
3142 if (sacked
& TCPCB_SACKED_ACKED
) {
3143 tp
->sacked_out
-= acked_pcount
;
3144 } else if (tcp_is_sack(tp
)) {
3145 tp
->delivered
+= acked_pcount
;
3146 if (!tcp_skb_spurious_retrans(tp
, skb
))
3147 tcp_rack_advance(tp
, &skb
->skb_mstamp
, sacked
);
3149 if (sacked
& TCPCB_LOST
)
3150 tp
->lost_out
-= acked_pcount
;
3152 tp
->packets_out
-= acked_pcount
;
3153 pkts_acked
+= acked_pcount
;
3154 tcp_rate_skb_delivered(sk
, skb
, sack
->rate
);
3156 /* Initial outgoing SYN's get put onto the write_queue
3157 * just like anything else we transmit. It is not
3158 * true data, and if we misinform our callers that
3159 * this ACK acks real data, we will erroneously exit
3160 * connection startup slow start one packet too
3161 * quickly. This is severely frowned upon behavior.
3163 if (likely(!(scb
->tcp_flags
& TCPHDR_SYN
))) {
3164 flag
|= FLAG_DATA_ACKED
;
3166 flag
|= FLAG_SYN_ACKED
;
3167 tp
->retrans_stamp
= 0;
3173 tcp_unlink_write_queue(skb
, sk
);
3174 sk_wmem_free_skb(sk
, skb
);
3175 if (unlikely(skb
== tp
->retransmit_skb_hint
))
3176 tp
->retransmit_skb_hint
= NULL
;
3177 if (unlikely(skb
== tp
->lost_skb_hint
))
3178 tp
->lost_skb_hint
= NULL
;
3182 tcp_chrono_stop(sk
, TCP_CHRONO_BUSY
);
3184 if (likely(between(tp
->snd_up
, prior_snd_una
, tp
->snd_una
)))
3185 tp
->snd_up
= tp
->snd_una
;
3187 if (skb
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
3188 flag
|= FLAG_SACK_RENEGING
;
3190 if (likely(first_ackt
.v64
) && !(flag
& FLAG_RETRANS_DATA_ACKED
)) {
3191 seq_rtt_us
= skb_mstamp_us_delta(now
, &first_ackt
);
3192 ca_rtt_us
= skb_mstamp_us_delta(now
, &last_ackt
);
3194 if (sack
->first_sackt
.v64
) {
3195 sack_rtt_us
= skb_mstamp_us_delta(now
, &sack
->first_sackt
);
3196 ca_rtt_us
= skb_mstamp_us_delta(now
, &sack
->last_sackt
);
3198 sack
->rate
->rtt_us
= ca_rtt_us
; /* RTT of last (S)ACKed packet, or -1 */
3199 rtt_update
= tcp_ack_update_rtt(sk
, flag
, seq_rtt_us
, sack_rtt_us
,
3202 if (flag
& FLAG_ACKED
) {
3204 if (unlikely(icsk
->icsk_mtup
.probe_size
&&
3205 !after(tp
->mtu_probe
.probe_seq_end
, tp
->snd_una
))) {
3206 tcp_mtup_probe_success(sk
);
3209 if (tcp_is_reno(tp
)) {
3210 tcp_remove_reno_sacks(sk
, pkts_acked
);
3214 /* Non-retransmitted hole got filled? That's reordering */
3215 if (reord
< prior_fackets
)
3216 tcp_update_reordering(sk
, tp
->fackets_out
- reord
, 0);
3218 delta
= tcp_is_fack(tp
) ? pkts_acked
:
3219 prior_sacked
- tp
->sacked_out
;
3220 tp
->lost_cnt_hint
-= min(tp
->lost_cnt_hint
, delta
);
3223 tp
->fackets_out
-= min(pkts_acked
, tp
->fackets_out
);
3225 } else if (skb
&& rtt_update
&& sack_rtt_us
>= 0 &&
3226 sack_rtt_us
> skb_mstamp_us_delta(now
, &skb
->skb_mstamp
)) {
3227 /* Do not re-arm RTO if the sack RTT is measured from data sent
3228 * after when the head was last (re)transmitted. Otherwise the
3229 * timeout may continue to extend in loss recovery.
3234 if (icsk
->icsk_ca_ops
->pkts_acked
) {
3235 struct ack_sample sample
= { .pkts_acked
= pkts_acked
,
3236 .rtt_us
= ca_rtt_us
,
3237 .in_flight
= last_in_flight
};
3239 icsk
->icsk_ca_ops
->pkts_acked(sk
, &sample
);
3242 #if FASTRETRANS_DEBUG > 0
3243 WARN_ON((int)tp
->sacked_out
< 0);
3244 WARN_ON((int)tp
->lost_out
< 0);
3245 WARN_ON((int)tp
->retrans_out
< 0);
3246 if (!tp
->packets_out
&& tcp_is_sack(tp
)) {
3247 icsk
= inet_csk(sk
);
3249 pr_debug("Leak l=%u %d\n",
3250 tp
->lost_out
, icsk
->icsk_ca_state
);
3253 if (tp
->sacked_out
) {
3254 pr_debug("Leak s=%u %d\n",
3255 tp
->sacked_out
, icsk
->icsk_ca_state
);
3258 if (tp
->retrans_out
) {
3259 pr_debug("Leak r=%u %d\n",
3260 tp
->retrans_out
, icsk
->icsk_ca_state
);
3261 tp
->retrans_out
= 0;
3265 *acked
= pkts_acked
;
3269 static void tcp_ack_probe(struct sock
*sk
)
3271 const struct tcp_sock
*tp
= tcp_sk(sk
);
3272 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3274 /* Was it a usable window open? */
3276 if (!after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
, tcp_wnd_end(tp
))) {
3277 icsk
->icsk_backoff
= 0;
3278 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
3279 /* Socket must be waked up by subsequent tcp_data_snd_check().
3280 * This function is not for random using!
3283 unsigned long when
= tcp_probe0_when(sk
, TCP_RTO_MAX
);
3285 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3290 static inline bool tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
3292 return !(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
3293 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
;
3296 /* Decide wheather to run the increase function of congestion control. */
3297 static inline bool tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
3299 /* If reordering is high then always grow cwnd whenever data is
3300 * delivered regardless of its ordering. Otherwise stay conservative
3301 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3302 * new SACK or ECE mark may first advance cwnd here and later reduce
3303 * cwnd in tcp_fastretrans_alert() based on more states.
3305 if (tcp_sk(sk
)->reordering
> sock_net(sk
)->ipv4
.sysctl_tcp_reordering
)
3306 return flag
& FLAG_FORWARD_PROGRESS
;
3308 return flag
& FLAG_DATA_ACKED
;
3311 /* The "ultimate" congestion control function that aims to replace the rigid
3312 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3313 * It's called toward the end of processing an ACK with precise rate
3314 * information. All transmission or retransmission are delayed afterwards.
3316 static void tcp_cong_control(struct sock
*sk
, u32 ack
, u32 acked_sacked
,
3317 int flag
, const struct rate_sample
*rs
)
3319 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3321 if (icsk
->icsk_ca_ops
->cong_control
) {
3322 icsk
->icsk_ca_ops
->cong_control(sk
, rs
);
3326 if (tcp_in_cwnd_reduction(sk
)) {
3327 /* Reduce cwnd if state mandates */
3328 tcp_cwnd_reduction(sk
, acked_sacked
, flag
);
3329 } else if (tcp_may_raise_cwnd(sk
, flag
)) {
3330 /* Advance cwnd if state allows */
3331 tcp_cong_avoid(sk
, ack
, acked_sacked
);
3333 tcp_update_pacing_rate(sk
);
3336 /* Check that window update is acceptable.
3337 * The function assumes that snd_una<=ack<=snd_next.
3339 static inline bool tcp_may_update_window(const struct tcp_sock
*tp
,
3340 const u32 ack
, const u32 ack_seq
,
3343 return after(ack
, tp
->snd_una
) ||
3344 after(ack_seq
, tp
->snd_wl1
) ||
3345 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
);
3348 /* If we update tp->snd_una, also update tp->bytes_acked */
3349 static void tcp_snd_una_update(struct tcp_sock
*tp
, u32 ack
)
3351 u32 delta
= ack
- tp
->snd_una
;
3353 sock_owned_by_me((struct sock
*)tp
);
3354 tp
->bytes_acked
+= delta
;
3358 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3359 static void tcp_rcv_nxt_update(struct tcp_sock
*tp
, u32 seq
)
3361 u32 delta
= seq
- tp
->rcv_nxt
;
3363 sock_owned_by_me((struct sock
*)tp
);
3364 tp
->bytes_received
+= delta
;
3368 /* Update our send window.
3370 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3371 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3373 static int tcp_ack_update_window(struct sock
*sk
, const struct sk_buff
*skb
, u32 ack
,
3376 struct tcp_sock
*tp
= tcp_sk(sk
);
3378 u32 nwin
= ntohs(tcp_hdr(skb
)->window
);
3380 if (likely(!tcp_hdr(skb
)->syn
))
3381 nwin
<<= tp
->rx_opt
.snd_wscale
;
3383 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
3384 flag
|= FLAG_WIN_UPDATE
;
3385 tcp_update_wl(tp
, ack_seq
);
3387 if (tp
->snd_wnd
!= nwin
) {
3390 /* Note, it is the only place, where
3391 * fast path is recovered for sending TCP.
3394 tcp_fast_path_check(sk
);
3396 if (tcp_send_head(sk
))
3397 tcp_slow_start_after_idle_check(sk
);
3399 if (nwin
> tp
->max_window
) {
3400 tp
->max_window
= nwin
;
3401 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
3406 tcp_snd_una_update(tp
, ack
);
3411 static bool __tcp_oow_rate_limited(struct net
*net
, int mib_idx
,
3412 u32
*last_oow_ack_time
)
3414 if (*last_oow_ack_time
) {
3415 s32 elapsed
= (s32
)(tcp_time_stamp
- *last_oow_ack_time
);
3417 if (0 <= elapsed
&& elapsed
< sysctl_tcp_invalid_ratelimit
) {
3418 NET_INC_STATS(net
, mib_idx
);
3419 return true; /* rate-limited: don't send yet! */
3423 *last_oow_ack_time
= tcp_time_stamp
;
3425 return false; /* not rate-limited: go ahead, send dupack now! */
3428 /* Return true if we're currently rate-limiting out-of-window ACKs and
3429 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3430 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3431 * attacks that send repeated SYNs or ACKs for the same connection. To
3432 * do this, we do not send a duplicate SYNACK or ACK if the remote
3433 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3435 bool tcp_oow_rate_limited(struct net
*net
, const struct sk_buff
*skb
,
3436 int mib_idx
, u32
*last_oow_ack_time
)
3438 /* Data packets without SYNs are not likely part of an ACK loop. */
3439 if ((TCP_SKB_CB(skb
)->seq
!= TCP_SKB_CB(skb
)->end_seq
) &&
3443 return __tcp_oow_rate_limited(net
, mib_idx
, last_oow_ack_time
);
3446 /* RFC 5961 7 [ACK Throttling] */
3447 static void tcp_send_challenge_ack(struct sock
*sk
, const struct sk_buff
*skb
)
3449 /* unprotected vars, we dont care of overwrites */
3450 static u32 challenge_timestamp
;
3451 static unsigned int challenge_count
;
3452 struct tcp_sock
*tp
= tcp_sk(sk
);
3455 /* First check our per-socket dupack rate limit. */
3456 if (__tcp_oow_rate_limited(sock_net(sk
),
3457 LINUX_MIB_TCPACKSKIPPEDCHALLENGE
,
3458 &tp
->last_oow_ack_time
))
3461 /* Then check host-wide RFC 5961 rate limit. */
3463 if (now
!= challenge_timestamp
) {
3464 u32 half
= (sysctl_tcp_challenge_ack_limit
+ 1) >> 1;
3466 challenge_timestamp
= now
;
3467 WRITE_ONCE(challenge_count
, half
+
3468 prandom_u32_max(sysctl_tcp_challenge_ack_limit
));
3470 count
= READ_ONCE(challenge_count
);
3472 WRITE_ONCE(challenge_count
, count
- 1);
3473 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPCHALLENGEACK
);
3478 static void tcp_store_ts_recent(struct tcp_sock
*tp
)
3480 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
3481 tp
->rx_opt
.ts_recent_stamp
= get_seconds();
3484 static void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
3486 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
3487 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3488 * extra check below makes sure this can only happen
3489 * for pure ACK frames. -DaveM
3491 * Not only, also it occurs for expired timestamps.
3494 if (tcp_paws_check(&tp
->rx_opt
, 0))
3495 tcp_store_ts_recent(tp
);
3499 /* This routine deals with acks during a TLP episode.
3500 * We mark the end of a TLP episode on receiving TLP dupack or when
3501 * ack is after tlp_high_seq.
3502 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3504 static void tcp_process_tlp_ack(struct sock
*sk
, u32 ack
, int flag
)
3506 struct tcp_sock
*tp
= tcp_sk(sk
);
3508 if (before(ack
, tp
->tlp_high_seq
))
3511 if (flag
& FLAG_DSACKING_ACK
) {
3512 /* This DSACK means original and TLP probe arrived; no loss */
3513 tp
->tlp_high_seq
= 0;
3514 } else if (after(ack
, tp
->tlp_high_seq
)) {
3515 /* ACK advances: there was a loss, so reduce cwnd. Reset
3516 * tlp_high_seq in tcp_init_cwnd_reduction()
3518 tcp_init_cwnd_reduction(sk
);
3519 tcp_set_ca_state(sk
, TCP_CA_CWR
);
3520 tcp_end_cwnd_reduction(sk
);
3521 tcp_try_keep_open(sk
);
3522 NET_INC_STATS(sock_net(sk
),
3523 LINUX_MIB_TCPLOSSPROBERECOVERY
);
3524 } else if (!(flag
& (FLAG_SND_UNA_ADVANCED
|
3525 FLAG_NOT_DUP
| FLAG_DATA_SACKED
))) {
3526 /* Pure dupack: original and TLP probe arrived; no loss */
3527 tp
->tlp_high_seq
= 0;
3531 static inline void tcp_in_ack_event(struct sock
*sk
, u32 flags
)
3533 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3535 if (icsk
->icsk_ca_ops
->in_ack_event
)
3536 icsk
->icsk_ca_ops
->in_ack_event(sk
, flags
);
3539 /* Congestion control has updated the cwnd already. So if we're in
3540 * loss recovery then now we do any new sends (for FRTO) or
3541 * retransmits (for CA_Loss or CA_recovery) that make sense.
3543 static void tcp_xmit_recovery(struct sock
*sk
, int rexmit
)
3545 struct tcp_sock
*tp
= tcp_sk(sk
);
3547 if (rexmit
== REXMIT_NONE
)
3550 if (unlikely(rexmit
== 2)) {
3551 __tcp_push_pending_frames(sk
, tcp_current_mss(sk
),
3553 if (after(tp
->snd_nxt
, tp
->high_seq
))
3557 tcp_xmit_retransmit_queue(sk
);
3560 /* This routine deals with incoming acks, but not outgoing ones. */
3561 static int tcp_ack(struct sock
*sk
, const struct sk_buff
*skb
, int flag
)
3563 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3564 struct tcp_sock
*tp
= tcp_sk(sk
);
3565 struct tcp_sacktag_state sack_state
;
3566 struct rate_sample rs
= { .prior_delivered
= 0 };
3567 u32 prior_snd_una
= tp
->snd_una
;
3568 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
3569 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3570 bool is_dupack
= false;
3572 int prior_packets
= tp
->packets_out
;
3573 u32 delivered
= tp
->delivered
;
3574 u32 lost
= tp
->lost
;
3575 int acked
= 0; /* Number of packets newly acked */
3576 int rexmit
= REXMIT_NONE
; /* Flag to (re)transmit to recover losses */
3577 struct skb_mstamp now
;
3579 sack_state
.first_sackt
.v64
= 0;
3580 sack_state
.rate
= &rs
;
3582 /* We very likely will need to access write queue head. */
3583 prefetchw(sk
->sk_write_queue
.next
);
3585 /* If the ack is older than previous acks
3586 * then we can probably ignore it.
3588 if (before(ack
, prior_snd_una
)) {
3589 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3590 if (before(ack
, prior_snd_una
- tp
->max_window
)) {
3591 tcp_send_challenge_ack(sk
, skb
);
3597 /* If the ack includes data we haven't sent yet, discard
3598 * this segment (RFC793 Section 3.9).
3600 if (after(ack
, tp
->snd_nxt
))
3603 skb_mstamp_get(&now
);
3605 if (icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
||
3606 icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
)
3609 if (after(ack
, prior_snd_una
)) {
3610 flag
|= FLAG_SND_UNA_ADVANCED
;
3611 icsk
->icsk_retransmits
= 0;
3614 prior_fackets
= tp
->fackets_out
;
3615 rs
.prior_in_flight
= tcp_packets_in_flight(tp
);
3617 /* ts_recent update must be made after we are sure that the packet
3620 if (flag
& FLAG_UPDATE_TS_RECENT
)
3621 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
3623 if (!(flag
& FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
3624 /* Window is constant, pure forward advance.
3625 * No more checks are required.
3626 * Note, we use the fact that SND.UNA>=SND.WL2.
3628 tcp_update_wl(tp
, ack_seq
);
3629 tcp_snd_una_update(tp
, ack
);
3630 flag
|= FLAG_WIN_UPDATE
;
3632 tcp_in_ack_event(sk
, CA_ACK_WIN_UPDATE
);
3634 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPHPACKS
);
3636 u32 ack_ev_flags
= CA_ACK_SLOWPATH
;
3638 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
3641 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPPUREACKS
);
3643 flag
|= tcp_ack_update_window(sk
, skb
, ack
, ack_seq
);
3645 if (TCP_SKB_CB(skb
)->sacked
)
3646 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
,
3649 if (tcp_ecn_rcv_ecn_echo(tp
, tcp_hdr(skb
))) {
3651 ack_ev_flags
|= CA_ACK_ECE
;
3654 if (flag
& FLAG_WIN_UPDATE
)
3655 ack_ev_flags
|= CA_ACK_WIN_UPDATE
;
3657 tcp_in_ack_event(sk
, ack_ev_flags
);
3660 /* We passed data and got it acked, remove any soft error
3661 * log. Something worked...
3663 sk
->sk_err_soft
= 0;
3664 icsk
->icsk_probes_out
= 0;
3665 tp
->rcv_tstamp
= tcp_time_stamp
;
3669 /* See if we can take anything off of the retransmit queue. */
3670 flag
|= tcp_clean_rtx_queue(sk
, prior_fackets
, prior_snd_una
, &acked
,
3673 if (tcp_ack_is_dubious(sk
, flag
)) {
3674 is_dupack
= !(flag
& (FLAG_SND_UNA_ADVANCED
| FLAG_NOT_DUP
));
3675 tcp_fastretrans_alert(sk
, acked
, is_dupack
, &flag
, &rexmit
);
3677 if (tp
->tlp_high_seq
)
3678 tcp_process_tlp_ack(sk
, ack
, flag
);
3680 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
& FLAG_NOT_DUP
)) {
3681 struct dst_entry
*dst
= __sk_dst_get(sk
);
3686 if (icsk
->icsk_pending
== ICSK_TIME_RETRANS
)
3687 tcp_schedule_loss_probe(sk
);
3688 delivered
= tp
->delivered
- delivered
; /* freshly ACKed or SACKed */
3689 lost
= tp
->lost
- lost
; /* freshly marked lost */
3690 tcp_rate_gen(sk
, delivered
, lost
, &now
, &rs
);
3691 tcp_cong_control(sk
, ack
, delivered
, flag
, &rs
);
3692 tcp_xmit_recovery(sk
, rexmit
);
3696 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3697 if (flag
& FLAG_DSACKING_ACK
)
3698 tcp_fastretrans_alert(sk
, acked
, is_dupack
, &flag
, &rexmit
);
3699 /* If this ack opens up a zero window, clear backoff. It was
3700 * being used to time the probes, and is probably far higher than
3701 * it needs to be for normal retransmission.
3703 if (tcp_send_head(sk
))
3706 if (tp
->tlp_high_seq
)
3707 tcp_process_tlp_ack(sk
, ack
, flag
);
3711 SOCK_DEBUG(sk
, "Ack %u after %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3715 /* If data was SACKed, tag it and see if we should send more data.
3716 * If data was DSACKed, see if we can undo a cwnd reduction.
3718 if (TCP_SKB_CB(skb
)->sacked
) {
3719 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
,
3721 tcp_fastretrans_alert(sk
, acked
, is_dupack
, &flag
, &rexmit
);
3722 tcp_xmit_recovery(sk
, rexmit
);
3725 SOCK_DEBUG(sk
, "Ack %u before %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3729 static void tcp_parse_fastopen_option(int len
, const unsigned char *cookie
,
3730 bool syn
, struct tcp_fastopen_cookie
*foc
,
3733 /* Valid only in SYN or SYN-ACK with an even length. */
3734 if (!foc
|| !syn
|| len
< 0 || (len
& 1))
3737 if (len
>= TCP_FASTOPEN_COOKIE_MIN
&&
3738 len
<= TCP_FASTOPEN_COOKIE_MAX
)
3739 memcpy(foc
->val
, cookie
, len
);
3746 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3747 * But, this can also be called on packets in the established flow when
3748 * the fast version below fails.
3750 void tcp_parse_options(const struct sk_buff
*skb
,
3751 struct tcp_options_received
*opt_rx
, int estab
,
3752 struct tcp_fastopen_cookie
*foc
)
3754 const unsigned char *ptr
;
3755 const struct tcphdr
*th
= tcp_hdr(skb
);
3756 int length
= (th
->doff
* 4) - sizeof(struct tcphdr
);
3758 ptr
= (const unsigned char *)(th
+ 1);
3759 opt_rx
->saw_tstamp
= 0;
3761 while (length
> 0) {
3762 int opcode
= *ptr
++;
3768 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
3773 if (opsize
< 2) /* "silly options" */
3775 if (opsize
> length
)
3776 return; /* don't parse partial options */
3779 if (opsize
== TCPOLEN_MSS
&& th
->syn
&& !estab
) {
3780 u16 in_mss
= get_unaligned_be16(ptr
);
3782 if (opt_rx
->user_mss
&&
3783 opt_rx
->user_mss
< in_mss
)
3784 in_mss
= opt_rx
->user_mss
;
3785 opt_rx
->mss_clamp
= in_mss
;
3790 if (opsize
== TCPOLEN_WINDOW
&& th
->syn
&&
3791 !estab
&& sysctl_tcp_window_scaling
) {
3792 __u8 snd_wscale
= *(__u8
*)ptr
;
3793 opt_rx
->wscale_ok
= 1;
3794 if (snd_wscale
> 14) {
3795 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3800 opt_rx
->snd_wscale
= snd_wscale
;
3803 case TCPOPT_TIMESTAMP
:
3804 if ((opsize
== TCPOLEN_TIMESTAMP
) &&
3805 ((estab
&& opt_rx
->tstamp_ok
) ||
3806 (!estab
&& sysctl_tcp_timestamps
))) {
3807 opt_rx
->saw_tstamp
= 1;
3808 opt_rx
->rcv_tsval
= get_unaligned_be32(ptr
);
3809 opt_rx
->rcv_tsecr
= get_unaligned_be32(ptr
+ 4);
3812 case TCPOPT_SACK_PERM
:
3813 if (opsize
== TCPOLEN_SACK_PERM
&& th
->syn
&&
3814 !estab
&& sysctl_tcp_sack
) {
3815 opt_rx
->sack_ok
= TCP_SACK_SEEN
;
3816 tcp_sack_reset(opt_rx
);
3821 if ((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
3822 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
3824 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
3827 #ifdef CONFIG_TCP_MD5SIG
3830 * The MD5 Hash has already been
3831 * checked (see tcp_v{4,6}_do_rcv()).
3835 case TCPOPT_FASTOPEN
:
3836 tcp_parse_fastopen_option(
3837 opsize
- TCPOLEN_FASTOPEN_BASE
,
3838 ptr
, th
->syn
, foc
, false);
3842 /* Fast Open option shares code 254 using a
3843 * 16 bits magic number.
3845 if (opsize
>= TCPOLEN_EXP_FASTOPEN_BASE
&&
3846 get_unaligned_be16(ptr
) ==
3847 TCPOPT_FASTOPEN_MAGIC
)
3848 tcp_parse_fastopen_option(opsize
-
3849 TCPOLEN_EXP_FASTOPEN_BASE
,
3850 ptr
+ 2, th
->syn
, foc
, true);
3859 EXPORT_SYMBOL(tcp_parse_options
);
3861 static bool tcp_parse_aligned_timestamp(struct tcp_sock
*tp
, const struct tcphdr
*th
)
3863 const __be32
*ptr
= (const __be32
*)(th
+ 1);
3865 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3866 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
3867 tp
->rx_opt
.saw_tstamp
= 1;
3869 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3872 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
) - tp
->tsoffset
;
3874 tp
->rx_opt
.rcv_tsecr
= 0;
3880 /* Fast parse options. This hopes to only see timestamps.
3881 * If it is wrong it falls back on tcp_parse_options().
3883 static bool tcp_fast_parse_options(const struct sk_buff
*skb
,
3884 const struct tcphdr
*th
, struct tcp_sock
*tp
)
3886 /* In the spirit of fast parsing, compare doff directly to constant
3887 * values. Because equality is used, short doff can be ignored here.
3889 if (th
->doff
== (sizeof(*th
) / 4)) {
3890 tp
->rx_opt
.saw_tstamp
= 0;
3892 } else if (tp
->rx_opt
.tstamp_ok
&&
3893 th
->doff
== ((sizeof(*th
) + TCPOLEN_TSTAMP_ALIGNED
) / 4)) {
3894 if (tcp_parse_aligned_timestamp(tp
, th
))
3898 tcp_parse_options(skb
, &tp
->rx_opt
, 1, NULL
);
3899 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
3900 tp
->rx_opt
.rcv_tsecr
-= tp
->tsoffset
;
3905 #ifdef CONFIG_TCP_MD5SIG
3907 * Parse MD5 Signature option
3909 const u8
*tcp_parse_md5sig_option(const struct tcphdr
*th
)
3911 int length
= (th
->doff
<< 2) - sizeof(*th
);
3912 const u8
*ptr
= (const u8
*)(th
+ 1);
3914 /* If the TCP option is too short, we can short cut */
3915 if (length
< TCPOLEN_MD5SIG
)
3918 while (length
> 0) {
3919 int opcode
= *ptr
++;
3930 if (opsize
< 2 || opsize
> length
)
3932 if (opcode
== TCPOPT_MD5SIG
)
3933 return opsize
== TCPOLEN_MD5SIG
? ptr
: NULL
;
3940 EXPORT_SYMBOL(tcp_parse_md5sig_option
);
3943 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3945 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3946 * it can pass through stack. So, the following predicate verifies that
3947 * this segment is not used for anything but congestion avoidance or
3948 * fast retransmit. Moreover, we even are able to eliminate most of such
3949 * second order effects, if we apply some small "replay" window (~RTO)
3950 * to timestamp space.
3952 * All these measures still do not guarantee that we reject wrapped ACKs
3953 * on networks with high bandwidth, when sequence space is recycled fastly,
3954 * but it guarantees that such events will be very rare and do not affect
3955 * connection seriously. This doesn't look nice, but alas, PAWS is really
3958 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3959 * states that events when retransmit arrives after original data are rare.
3960 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3961 * the biggest problem on large power networks even with minor reordering.
3962 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3963 * up to bandwidth of 18Gigabit/sec. 8) ]
3966 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
3968 const struct tcp_sock
*tp
= tcp_sk(sk
);
3969 const struct tcphdr
*th
= tcp_hdr(skb
);
3970 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3971 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3973 return (/* 1. Pure ACK with correct sequence number. */
3974 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
3976 /* 2. ... and duplicate ACK. */
3977 ack
== tp
->snd_una
&&
3979 /* 3. ... and does not update window. */
3980 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
3982 /* 4. ... and sits in replay window. */
3983 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
3986 static inline bool tcp_paws_discard(const struct sock
*sk
,
3987 const struct sk_buff
*skb
)
3989 const struct tcp_sock
*tp
= tcp_sk(sk
);
3991 return !tcp_paws_check(&tp
->rx_opt
, TCP_PAWS_WINDOW
) &&
3992 !tcp_disordered_ack(sk
, skb
);
3995 /* Check segment sequence number for validity.
3997 * Segment controls are considered valid, if the segment
3998 * fits to the window after truncation to the window. Acceptability
3999 * of data (and SYN, FIN, of course) is checked separately.
4000 * See tcp_data_queue(), for example.
4002 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4003 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4004 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4005 * (borrowed from freebsd)
4008 static inline bool tcp_sequence(const struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
4010 return !before(end_seq
, tp
->rcv_wup
) &&
4011 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
4014 /* When we get a reset we do this. */
4015 void tcp_reset(struct sock
*sk
)
4017 /* We want the right error as BSD sees it (and indeed as we do). */
4018 switch (sk
->sk_state
) {
4020 sk
->sk_err
= ECONNREFUSED
;
4022 case TCP_CLOSE_WAIT
:
4028 sk
->sk_err
= ECONNRESET
;
4030 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4033 if (!sock_flag(sk
, SOCK_DEAD
))
4034 sk
->sk_error_report(sk
);
4040 * Process the FIN bit. This now behaves as it is supposed to work
4041 * and the FIN takes effect when it is validly part of sequence
4042 * space. Not before when we get holes.
4044 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4045 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4048 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4049 * close and we go into CLOSING (and later onto TIME-WAIT)
4051 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4053 void tcp_fin(struct sock
*sk
)
4055 struct tcp_sock
*tp
= tcp_sk(sk
);
4057 inet_csk_schedule_ack(sk
);
4059 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
4060 sock_set_flag(sk
, SOCK_DONE
);
4062 switch (sk
->sk_state
) {
4064 case TCP_ESTABLISHED
:
4065 /* Move to CLOSE_WAIT */
4066 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
4067 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
4070 case TCP_CLOSE_WAIT
:
4072 /* Received a retransmission of the FIN, do
4077 /* RFC793: Remain in the LAST-ACK state. */
4081 /* This case occurs when a simultaneous close
4082 * happens, we must ack the received FIN and
4083 * enter the CLOSING state.
4086 tcp_set_state(sk
, TCP_CLOSING
);
4089 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4091 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
4094 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4095 * cases we should never reach this piece of code.
4097 pr_err("%s: Impossible, sk->sk_state=%d\n",
4098 __func__
, sk
->sk_state
);
4102 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4103 * Probably, we should reset in this case. For now drop them.
4105 skb_rbtree_purge(&tp
->out_of_order_queue
);
4106 if (tcp_is_sack(tp
))
4107 tcp_sack_reset(&tp
->rx_opt
);
4110 if (!sock_flag(sk
, SOCK_DEAD
)) {
4111 sk
->sk_state_change(sk
);
4113 /* Do not send POLL_HUP for half duplex close. */
4114 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
4115 sk
->sk_state
== TCP_CLOSE
)
4116 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_HUP
);
4118 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_IN
);
4122 static inline bool tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
,
4125 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
4126 if (before(seq
, sp
->start_seq
))
4127 sp
->start_seq
= seq
;
4128 if (after(end_seq
, sp
->end_seq
))
4129 sp
->end_seq
= end_seq
;
4135 static void tcp_dsack_set(struct sock
*sk
, u32 seq
, u32 end_seq
)
4137 struct tcp_sock
*tp
= tcp_sk(sk
);
4139 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
4142 if (before(seq
, tp
->rcv_nxt
))
4143 mib_idx
= LINUX_MIB_TCPDSACKOLDSENT
;
4145 mib_idx
= LINUX_MIB_TCPDSACKOFOSENT
;
4147 NET_INC_STATS(sock_net(sk
), mib_idx
);
4149 tp
->rx_opt
.dsack
= 1;
4150 tp
->duplicate_sack
[0].start_seq
= seq
;
4151 tp
->duplicate_sack
[0].end_seq
= end_seq
;
4155 static void tcp_dsack_extend(struct sock
*sk
, u32 seq
, u32 end_seq
)
4157 struct tcp_sock
*tp
= tcp_sk(sk
);
4159 if (!tp
->rx_opt
.dsack
)
4160 tcp_dsack_set(sk
, seq
, end_seq
);
4162 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
4165 static void tcp_send_dupack(struct sock
*sk
, const struct sk_buff
*skb
)
4167 struct tcp_sock
*tp
= tcp_sk(sk
);
4169 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
4170 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4171 NET_INC_STATS(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4172 tcp_enter_quickack_mode(sk
);
4174 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
4175 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4177 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
4178 end_seq
= tp
->rcv_nxt
;
4179 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, end_seq
);
4186 /* These routines update the SACK block as out-of-order packets arrive or
4187 * in-order packets close up the sequence space.
4189 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
4192 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4193 struct tcp_sack_block
*swalk
= sp
+ 1;
4195 /* See if the recent change to the first SACK eats into
4196 * or hits the sequence space of other SACK blocks, if so coalesce.
4198 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
;) {
4199 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
4202 /* Zap SWALK, by moving every further SACK up by one slot.
4203 * Decrease num_sacks.
4205 tp
->rx_opt
.num_sacks
--;
4206 for (i
= this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
4210 this_sack
++, swalk
++;
4214 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
4216 struct tcp_sock
*tp
= tcp_sk(sk
);
4217 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4218 int cur_sacks
= tp
->rx_opt
.num_sacks
;
4224 for (this_sack
= 0; this_sack
< cur_sacks
; this_sack
++, sp
++) {
4225 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
4226 /* Rotate this_sack to the first one. */
4227 for (; this_sack
> 0; this_sack
--, sp
--)
4228 swap(*sp
, *(sp
- 1));
4230 tcp_sack_maybe_coalesce(tp
);
4235 /* Could not find an adjacent existing SACK, build a new one,
4236 * put it at the front, and shift everyone else down. We
4237 * always know there is at least one SACK present already here.
4239 * If the sack array is full, forget about the last one.
4241 if (this_sack
>= TCP_NUM_SACKS
) {
4243 tp
->rx_opt
.num_sacks
--;
4246 for (; this_sack
> 0; this_sack
--, sp
--)
4250 /* Build the new head SACK, and we're done. */
4251 sp
->start_seq
= seq
;
4252 sp
->end_seq
= end_seq
;
4253 tp
->rx_opt
.num_sacks
++;
4256 /* RCV.NXT advances, some SACKs should be eaten. */
4258 static void tcp_sack_remove(struct tcp_sock
*tp
)
4260 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4261 int num_sacks
= tp
->rx_opt
.num_sacks
;
4264 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4265 if (RB_EMPTY_ROOT(&tp
->out_of_order_queue
)) {
4266 tp
->rx_opt
.num_sacks
= 0;
4270 for (this_sack
= 0; this_sack
< num_sacks
;) {
4271 /* Check if the start of the sack is covered by RCV.NXT. */
4272 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
4275 /* RCV.NXT must cover all the block! */
4276 WARN_ON(before(tp
->rcv_nxt
, sp
->end_seq
));
4278 /* Zap this SACK, by moving forward any other SACKS. */
4279 for (i
= this_sack
+1; i
< num_sacks
; i
++)
4280 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
4287 tp
->rx_opt
.num_sacks
= num_sacks
;
4291 * tcp_try_coalesce - try to merge skb to prior one
4294 * @from: buffer to add in queue
4295 * @fragstolen: pointer to boolean
4297 * Before queueing skb @from after @to, try to merge them
4298 * to reduce overall memory use and queue lengths, if cost is small.
4299 * Packets in ofo or receive queues can stay a long time.
4300 * Better try to coalesce them right now to avoid future collapses.
4301 * Returns true if caller should free @from instead of queueing it
4303 static bool tcp_try_coalesce(struct sock
*sk
,
4305 struct sk_buff
*from
,
4310 *fragstolen
= false;
4312 /* Its possible this segment overlaps with prior segment in queue */
4313 if (TCP_SKB_CB(from
)->seq
!= TCP_SKB_CB(to
)->end_seq
)
4316 if (!skb_try_coalesce(to
, from
, fragstolen
, &delta
))
4319 atomic_add(delta
, &sk
->sk_rmem_alloc
);
4320 sk_mem_charge(sk
, delta
);
4321 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPRCVCOALESCE
);
4322 TCP_SKB_CB(to
)->end_seq
= TCP_SKB_CB(from
)->end_seq
;
4323 TCP_SKB_CB(to
)->ack_seq
= TCP_SKB_CB(from
)->ack_seq
;
4324 TCP_SKB_CB(to
)->tcp_flags
|= TCP_SKB_CB(from
)->tcp_flags
;
4328 static void tcp_drop(struct sock
*sk
, struct sk_buff
*skb
)
4330 sk_drops_add(sk
, skb
);
4334 /* This one checks to see if we can put data from the
4335 * out_of_order queue into the receive_queue.
4337 static void tcp_ofo_queue(struct sock
*sk
)
4339 struct tcp_sock
*tp
= tcp_sk(sk
);
4340 __u32 dsack_high
= tp
->rcv_nxt
;
4341 bool fin
, fragstolen
, eaten
;
4342 struct sk_buff
*skb
, *tail
;
4345 p
= rb_first(&tp
->out_of_order_queue
);
4347 skb
= rb_entry(p
, struct sk_buff
, rbnode
);
4348 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
4351 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
4352 __u32 dsack
= dsack_high
;
4353 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
4354 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
4355 tcp_dsack_extend(sk
, TCP_SKB_CB(skb
)->seq
, dsack
);
4358 rb_erase(&skb
->rbnode
, &tp
->out_of_order_queue
);
4360 if (unlikely(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))) {
4361 SOCK_DEBUG(sk
, "ofo packet was already received\n");
4365 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
4366 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4367 TCP_SKB_CB(skb
)->end_seq
);
4369 tail
= skb_peek_tail(&sk
->sk_receive_queue
);
4370 eaten
= tail
&& tcp_try_coalesce(sk
, tail
, skb
, &fragstolen
);
4371 tcp_rcv_nxt_update(tp
, TCP_SKB_CB(skb
)->end_seq
);
4372 fin
= TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
;
4374 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4376 kfree_skb_partial(skb
, fragstolen
);
4378 if (unlikely(fin
)) {
4380 /* tcp_fin() purges tp->out_of_order_queue,
4381 * so we must end this loop right now.
4388 static bool tcp_prune_ofo_queue(struct sock
*sk
);
4389 static int tcp_prune_queue(struct sock
*sk
);
4391 static int tcp_try_rmem_schedule(struct sock
*sk
, struct sk_buff
*skb
,
4394 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
4395 !sk_rmem_schedule(sk
, skb
, size
)) {
4397 if (tcp_prune_queue(sk
) < 0)
4400 while (!sk_rmem_schedule(sk
, skb
, size
)) {
4401 if (!tcp_prune_ofo_queue(sk
))
4408 static void tcp_data_queue_ofo(struct sock
*sk
, struct sk_buff
*skb
)
4410 struct tcp_sock
*tp
= tcp_sk(sk
);
4411 struct rb_node
**p
, *q
, *parent
;
4412 struct sk_buff
*skb1
;
4416 tcp_ecn_check_ce(tp
, skb
);
4418 if (unlikely(tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))) {
4419 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPOFODROP
);
4424 /* Disable header prediction. */
4426 inet_csk_schedule_ack(sk
);
4428 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPOFOQUEUE
);
4429 seq
= TCP_SKB_CB(skb
)->seq
;
4430 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4431 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
4432 tp
->rcv_nxt
, seq
, end_seq
);
4434 p
= &tp
->out_of_order_queue
.rb_node
;
4435 if (RB_EMPTY_ROOT(&tp
->out_of_order_queue
)) {
4436 /* Initial out of order segment, build 1 SACK. */
4437 if (tcp_is_sack(tp
)) {
4438 tp
->rx_opt
.num_sacks
= 1;
4439 tp
->selective_acks
[0].start_seq
= seq
;
4440 tp
->selective_acks
[0].end_seq
= end_seq
;
4442 rb_link_node(&skb
->rbnode
, NULL
, p
);
4443 rb_insert_color(&skb
->rbnode
, &tp
->out_of_order_queue
);
4444 tp
->ooo_last_skb
= skb
;
4448 /* In the typical case, we are adding an skb to the end of the list.
4449 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4451 if (tcp_try_coalesce(sk
, tp
->ooo_last_skb
, skb
, &fragstolen
)) {
4453 tcp_grow_window(sk
, skb
);
4454 kfree_skb_partial(skb
, fragstolen
);
4458 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4459 if (!before(seq
, TCP_SKB_CB(tp
->ooo_last_skb
)->end_seq
)) {
4460 parent
= &tp
->ooo_last_skb
->rbnode
;
4461 p
= &parent
->rb_right
;
4465 /* Find place to insert this segment. Handle overlaps on the way. */
4469 skb1
= rb_entry(parent
, struct sk_buff
, rbnode
);
4470 if (before(seq
, TCP_SKB_CB(skb1
)->seq
)) {
4471 p
= &parent
->rb_left
;
4474 if (before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4475 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4476 /* All the bits are present. Drop. */
4477 NET_INC_STATS(sock_net(sk
),
4478 LINUX_MIB_TCPOFOMERGE
);
4481 tcp_dsack_set(sk
, seq
, end_seq
);
4484 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
4485 /* Partial overlap. */
4486 tcp_dsack_set(sk
, seq
, TCP_SKB_CB(skb1
)->end_seq
);
4488 /* skb's seq == skb1's seq and skb covers skb1.
4489 * Replace skb1 with skb.
4491 rb_replace_node(&skb1
->rbnode
, &skb
->rbnode
,
4492 &tp
->out_of_order_queue
);
4493 tcp_dsack_extend(sk
,
4494 TCP_SKB_CB(skb1
)->seq
,
4495 TCP_SKB_CB(skb1
)->end_seq
);
4496 NET_INC_STATS(sock_net(sk
),
4497 LINUX_MIB_TCPOFOMERGE
);
4501 } else if (tcp_try_coalesce(sk
, skb1
, skb
, &fragstolen
)) {
4504 p
= &parent
->rb_right
;
4507 /* Insert segment into RB tree. */
4508 rb_link_node(&skb
->rbnode
, parent
, p
);
4509 rb_insert_color(&skb
->rbnode
, &tp
->out_of_order_queue
);
4512 /* Remove other segments covered by skb. */
4513 while ((q
= rb_next(&skb
->rbnode
)) != NULL
) {
4514 skb1
= rb_entry(q
, struct sk_buff
, rbnode
);
4516 if (!after(end_seq
, TCP_SKB_CB(skb1
)->seq
))
4518 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4519 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4523 rb_erase(&skb1
->rbnode
, &tp
->out_of_order_queue
);
4524 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4525 TCP_SKB_CB(skb1
)->end_seq
);
4526 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPOFOMERGE
);
4529 /* If there is no skb after us, we are the last_skb ! */
4531 tp
->ooo_last_skb
= skb
;
4534 if (tcp_is_sack(tp
))
4535 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
4538 tcp_grow_window(sk
, skb
);
4539 skb_set_owner_r(skb
, sk
);
4543 static int __must_check
tcp_queue_rcv(struct sock
*sk
, struct sk_buff
*skb
, int hdrlen
,
4547 struct sk_buff
*tail
= skb_peek_tail(&sk
->sk_receive_queue
);
4549 __skb_pull(skb
, hdrlen
);
4551 tcp_try_coalesce(sk
, tail
, skb
, fragstolen
)) ? 1 : 0;
4552 tcp_rcv_nxt_update(tcp_sk(sk
), TCP_SKB_CB(skb
)->end_seq
);
4554 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4555 skb_set_owner_r(skb
, sk
);
4560 int tcp_send_rcvq(struct sock
*sk
, struct msghdr
*msg
, size_t size
)
4562 struct sk_buff
*skb
;
4570 if (size
> PAGE_SIZE
) {
4571 int npages
= min_t(size_t, size
>> PAGE_SHIFT
, MAX_SKB_FRAGS
);
4573 data_len
= npages
<< PAGE_SHIFT
;
4574 size
= data_len
+ (size
& ~PAGE_MASK
);
4576 skb
= alloc_skb_with_frags(size
- data_len
, data_len
,
4577 PAGE_ALLOC_COSTLY_ORDER
,
4578 &err
, sk
->sk_allocation
);
4582 skb_put(skb
, size
- data_len
);
4583 skb
->data_len
= data_len
;
4586 if (tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))
4589 err
= skb_copy_datagram_from_iter(skb
, 0, &msg
->msg_iter
, size
);
4593 TCP_SKB_CB(skb
)->seq
= tcp_sk(sk
)->rcv_nxt
;
4594 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ size
;
4595 TCP_SKB_CB(skb
)->ack_seq
= tcp_sk(sk
)->snd_una
- 1;
4597 if (tcp_queue_rcv(sk
, skb
, 0, &fragstolen
)) {
4598 WARN_ON_ONCE(fragstolen
); /* should not happen */
4610 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
4612 struct tcp_sock
*tp
= tcp_sk(sk
);
4613 bool fragstolen
= false;
4616 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
) {
4621 __skb_pull(skb
, tcp_hdr(skb
)->doff
* 4);
4623 tcp_ecn_accept_cwr(tp
, skb
);
4625 tp
->rx_opt
.dsack
= 0;
4627 /* Queue data for delivery to the user.
4628 * Packets in sequence go to the receive queue.
4629 * Out of sequence packets to the out_of_order_queue.
4631 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
4632 if (tcp_receive_window(tp
) == 0)
4635 /* Ok. In sequence. In window. */
4636 if (tp
->ucopy
.task
== current
&&
4637 tp
->copied_seq
== tp
->rcv_nxt
&& tp
->ucopy
.len
&&
4638 sock_owned_by_user(sk
) && !tp
->urg_data
) {
4639 int chunk
= min_t(unsigned int, skb
->len
,
4642 __set_current_state(TASK_RUNNING
);
4644 if (!skb_copy_datagram_msg(skb
, 0, tp
->ucopy
.msg
, chunk
)) {
4645 tp
->ucopy
.len
-= chunk
;
4646 tp
->copied_seq
+= chunk
;
4647 eaten
= (chunk
== skb
->len
);
4648 tcp_rcv_space_adjust(sk
);
4655 if (skb_queue_len(&sk
->sk_receive_queue
) == 0)
4656 sk_forced_mem_schedule(sk
, skb
->truesize
);
4657 else if (tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))
4660 eaten
= tcp_queue_rcv(sk
, skb
, 0, &fragstolen
);
4662 tcp_rcv_nxt_update(tp
, TCP_SKB_CB(skb
)->end_seq
);
4664 tcp_event_data_recv(sk
, skb
);
4665 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
4668 if (!RB_EMPTY_ROOT(&tp
->out_of_order_queue
)) {
4671 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4672 * gap in queue is filled.
4674 if (RB_EMPTY_ROOT(&tp
->out_of_order_queue
))
4675 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
4678 if (tp
->rx_opt
.num_sacks
)
4679 tcp_sack_remove(tp
);
4681 tcp_fast_path_check(sk
);
4684 kfree_skb_partial(skb
, fragstolen
);
4685 if (!sock_flag(sk
, SOCK_DEAD
))
4686 sk
->sk_data_ready(sk
);
4690 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4691 /* A retransmit, 2nd most common case. Force an immediate ack. */
4692 NET_INC_STATS(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4693 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4696 tcp_enter_quickack_mode(sk
);
4697 inet_csk_schedule_ack(sk
);
4703 /* Out of window. F.e. zero window probe. */
4704 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
4707 tcp_enter_quickack_mode(sk
);
4709 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4710 /* Partial packet, seq < rcv_next < end_seq */
4711 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
4712 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4713 TCP_SKB_CB(skb
)->end_seq
);
4715 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
4717 /* If window is closed, drop tail of packet. But after
4718 * remembering D-SACK for its head made in previous line.
4720 if (!tcp_receive_window(tp
))
4725 tcp_data_queue_ofo(sk
, skb
);
4728 static struct sk_buff
*tcp_skb_next(struct sk_buff
*skb
, struct sk_buff_head
*list
)
4731 return !skb_queue_is_last(list
, skb
) ? skb
->next
: NULL
;
4733 return rb_entry_safe(rb_next(&skb
->rbnode
), struct sk_buff
, rbnode
);
4736 static struct sk_buff
*tcp_collapse_one(struct sock
*sk
, struct sk_buff
*skb
,
4737 struct sk_buff_head
*list
,
4738 struct rb_root
*root
)
4740 struct sk_buff
*next
= tcp_skb_next(skb
, list
);
4743 __skb_unlink(skb
, list
);
4745 rb_erase(&skb
->rbnode
, root
);
4748 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPRCVCOLLAPSED
);
4753 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
4754 static void tcp_rbtree_insert(struct rb_root
*root
, struct sk_buff
*skb
)
4756 struct rb_node
**p
= &root
->rb_node
;
4757 struct rb_node
*parent
= NULL
;
4758 struct sk_buff
*skb1
;
4762 skb1
= rb_entry(parent
, struct sk_buff
, rbnode
);
4763 if (before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb1
)->seq
))
4764 p
= &parent
->rb_left
;
4766 p
= &parent
->rb_right
;
4768 rb_link_node(&skb
->rbnode
, parent
, p
);
4769 rb_insert_color(&skb
->rbnode
, root
);
4772 /* Collapse contiguous sequence of skbs head..tail with
4773 * sequence numbers start..end.
4775 * If tail is NULL, this means until the end of the queue.
4777 * Segments with FIN/SYN are not collapsed (only because this
4781 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
, struct rb_root
*root
,
4782 struct sk_buff
*head
, struct sk_buff
*tail
, u32 start
, u32 end
)
4784 struct sk_buff
*skb
= head
, *n
;
4785 struct sk_buff_head tmp
;
4788 /* First, check that queue is collapsible and find
4789 * the point where collapsing can be useful.
4792 for (end_of_skbs
= true; skb
!= NULL
&& skb
!= tail
; skb
= n
) {
4793 n
= tcp_skb_next(skb
, list
);
4795 /* No new bits? It is possible on ofo queue. */
4796 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4797 skb
= tcp_collapse_one(sk
, skb
, list
, root
);
4803 /* The first skb to collapse is:
4805 * - bloated or contains data before "start" or
4806 * overlaps to the next one.
4808 if (!(TCP_SKB_CB(skb
)->tcp_flags
& (TCPHDR_SYN
| TCPHDR_FIN
)) &&
4809 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
4810 before(TCP_SKB_CB(skb
)->seq
, start
))) {
4811 end_of_skbs
= false;
4815 if (n
&& n
!= tail
&&
4816 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(n
)->seq
) {
4817 end_of_skbs
= false;
4821 /* Decided to skip this, advance start seq. */
4822 start
= TCP_SKB_CB(skb
)->end_seq
;
4825 (TCP_SKB_CB(skb
)->tcp_flags
& (TCPHDR_SYN
| TCPHDR_FIN
)))
4828 __skb_queue_head_init(&tmp
);
4830 while (before(start
, end
)) {
4831 int copy
= min_t(int, SKB_MAX_ORDER(0, 0), end
- start
);
4832 struct sk_buff
*nskb
;
4834 nskb
= alloc_skb(copy
, GFP_ATOMIC
);
4838 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
4839 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
4841 __skb_queue_before(list
, skb
, nskb
);
4843 __skb_queue_tail(&tmp
, nskb
); /* defer rbtree insertion */
4844 skb_set_owner_r(nskb
, sk
);
4846 /* Copy data, releasing collapsed skbs. */
4848 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
4849 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
4853 size
= min(copy
, size
);
4854 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
4856 TCP_SKB_CB(nskb
)->end_seq
+= size
;
4860 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4861 skb
= tcp_collapse_one(sk
, skb
, list
, root
);
4864 (TCP_SKB_CB(skb
)->tcp_flags
& (TCPHDR_SYN
| TCPHDR_FIN
)))
4870 skb_queue_walk_safe(&tmp
, skb
, n
)
4871 tcp_rbtree_insert(root
, skb
);
4874 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4875 * and tcp_collapse() them until all the queue is collapsed.
4877 static void tcp_collapse_ofo_queue(struct sock
*sk
)
4879 struct tcp_sock
*tp
= tcp_sk(sk
);
4880 struct sk_buff
*skb
, *head
;
4884 p
= rb_first(&tp
->out_of_order_queue
);
4885 skb
= rb_entry_safe(p
, struct sk_buff
, rbnode
);
4888 p
= rb_last(&tp
->out_of_order_queue
);
4889 /* Note: This is possible p is NULL here. We do not
4890 * use rb_entry_safe(), as ooo_last_skb is valid only
4891 * if rbtree is not empty.
4893 tp
->ooo_last_skb
= rb_entry(p
, struct sk_buff
, rbnode
);
4896 start
= TCP_SKB_CB(skb
)->seq
;
4897 end
= TCP_SKB_CB(skb
)->end_seq
;
4899 for (head
= skb
;;) {
4900 skb
= tcp_skb_next(skb
, NULL
);
4902 /* Range is terminated when we see a gap or when
4903 * we are at the queue end.
4906 after(TCP_SKB_CB(skb
)->seq
, end
) ||
4907 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
4908 tcp_collapse(sk
, NULL
, &tp
->out_of_order_queue
,
4909 head
, skb
, start
, end
);
4913 if (unlikely(before(TCP_SKB_CB(skb
)->seq
, start
)))
4914 start
= TCP_SKB_CB(skb
)->seq
;
4915 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
4916 end
= TCP_SKB_CB(skb
)->end_seq
;
4921 * Clean the out-of-order queue to make room.
4922 * We drop high sequences packets to :
4923 * 1) Let a chance for holes to be filled.
4924 * 2) not add too big latencies if thousands of packets sit there.
4925 * (But if application shrinks SO_RCVBUF, we could still end up
4926 * freeing whole queue here)
4928 * Return true if queue has shrunk.
4930 static bool tcp_prune_ofo_queue(struct sock
*sk
)
4932 struct tcp_sock
*tp
= tcp_sk(sk
);
4933 struct rb_node
*node
, *prev
;
4935 if (RB_EMPTY_ROOT(&tp
->out_of_order_queue
))
4938 NET_INC_STATS(sock_net(sk
), LINUX_MIB_OFOPRUNED
);
4939 node
= &tp
->ooo_last_skb
->rbnode
;
4941 prev
= rb_prev(node
);
4942 rb_erase(node
, &tp
->out_of_order_queue
);
4943 tcp_drop(sk
, rb_entry(node
, struct sk_buff
, rbnode
));
4945 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
&&
4946 !tcp_under_memory_pressure(sk
))
4950 tp
->ooo_last_skb
= rb_entry(prev
, struct sk_buff
, rbnode
);
4952 /* Reset SACK state. A conforming SACK implementation will
4953 * do the same at a timeout based retransmit. When a connection
4954 * is in a sad state like this, we care only about integrity
4955 * of the connection not performance.
4957 if (tp
->rx_opt
.sack_ok
)
4958 tcp_sack_reset(&tp
->rx_opt
);
4962 /* Reduce allocated memory if we can, trying to get
4963 * the socket within its memory limits again.
4965 * Return less than zero if we should start dropping frames
4966 * until the socket owning process reads some of the data
4967 * to stabilize the situation.
4969 static int tcp_prune_queue(struct sock
*sk
)
4971 struct tcp_sock
*tp
= tcp_sk(sk
);
4973 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
4975 NET_INC_STATS(sock_net(sk
), LINUX_MIB_PRUNECALLED
);
4977 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
4978 tcp_clamp_window(sk
);
4979 else if (tcp_under_memory_pressure(sk
))
4980 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
4982 tcp_collapse_ofo_queue(sk
);
4983 if (!skb_queue_empty(&sk
->sk_receive_queue
))
4984 tcp_collapse(sk
, &sk
->sk_receive_queue
, NULL
,
4985 skb_peek(&sk
->sk_receive_queue
),
4987 tp
->copied_seq
, tp
->rcv_nxt
);
4990 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4993 /* Collapsing did not help, destructive actions follow.
4994 * This must not ever occur. */
4996 tcp_prune_ofo_queue(sk
);
4998 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
5001 /* If we are really being abused, tell the caller to silently
5002 * drop receive data on the floor. It will get retransmitted
5003 * and hopefully then we'll have sufficient space.
5005 NET_INC_STATS(sock_net(sk
), LINUX_MIB_RCVPRUNED
);
5007 /* Massive buffer overcommit. */
5012 static bool tcp_should_expand_sndbuf(const struct sock
*sk
)
5014 const struct tcp_sock
*tp
= tcp_sk(sk
);
5016 /* If the user specified a specific send buffer setting, do
5019 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
5022 /* If we are under global TCP memory pressure, do not expand. */
5023 if (tcp_under_memory_pressure(sk
))
5026 /* If we are under soft global TCP memory pressure, do not expand. */
5027 if (sk_memory_allocated(sk
) >= sk_prot_mem_limits(sk
, 0))
5030 /* If we filled the congestion window, do not expand. */
5031 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
5037 /* When incoming ACK allowed to free some skb from write_queue,
5038 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
5039 * on the exit from tcp input handler.
5041 * PROBLEM: sndbuf expansion does not work well with largesend.
5043 static void tcp_new_space(struct sock
*sk
)
5045 struct tcp_sock
*tp
= tcp_sk(sk
);
5047 if (tcp_should_expand_sndbuf(sk
)) {
5048 tcp_sndbuf_expand(sk
);
5049 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
5052 sk
->sk_write_space(sk
);
5055 static void tcp_check_space(struct sock
*sk
)
5057 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
5058 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
5059 /* pairs with tcp_poll() */
5060 smp_mb__after_atomic();
5061 if (sk
->sk_socket
&&
5062 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
5064 if (!test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
5065 tcp_chrono_stop(sk
, TCP_CHRONO_SNDBUF_LIMITED
);
5070 static inline void tcp_data_snd_check(struct sock
*sk
)
5072 tcp_push_pending_frames(sk
);
5073 tcp_check_space(sk
);
5077 * Check if sending an ack is needed.
5079 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
5081 struct tcp_sock
*tp
= tcp_sk(sk
);
5083 /* More than one full frame received... */
5084 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
&&
5085 /* ... and right edge of window advances far enough.
5086 * (tcp_recvmsg() will send ACK otherwise). Or...
5088 __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
5089 /* We ACK each frame or... */
5090 tcp_in_quickack_mode(sk
) ||
5091 /* We have out of order data. */
5092 (ofo_possible
&& !RB_EMPTY_ROOT(&tp
->out_of_order_queue
))) {
5093 /* Then ack it now */
5096 /* Else, send delayed ack. */
5097 tcp_send_delayed_ack(sk
);
5101 static inline void tcp_ack_snd_check(struct sock
*sk
)
5103 if (!inet_csk_ack_scheduled(sk
)) {
5104 /* We sent a data segment already. */
5107 __tcp_ack_snd_check(sk
, 1);
5111 * This routine is only called when we have urgent data
5112 * signaled. Its the 'slow' part of tcp_urg. It could be
5113 * moved inline now as tcp_urg is only called from one
5114 * place. We handle URGent data wrong. We have to - as
5115 * BSD still doesn't use the correction from RFC961.
5116 * For 1003.1g we should support a new option TCP_STDURG to permit
5117 * either form (or just set the sysctl tcp_stdurg).
5120 static void tcp_check_urg(struct sock
*sk
, const struct tcphdr
*th
)
5122 struct tcp_sock
*tp
= tcp_sk(sk
);
5123 u32 ptr
= ntohs(th
->urg_ptr
);
5125 if (ptr
&& !sysctl_tcp_stdurg
)
5127 ptr
+= ntohl(th
->seq
);
5129 /* Ignore urgent data that we've already seen and read. */
5130 if (after(tp
->copied_seq
, ptr
))
5133 /* Do not replay urg ptr.
5135 * NOTE: interesting situation not covered by specs.
5136 * Misbehaving sender may send urg ptr, pointing to segment,
5137 * which we already have in ofo queue. We are not able to fetch
5138 * such data and will stay in TCP_URG_NOTYET until will be eaten
5139 * by recvmsg(). Seems, we are not obliged to handle such wicked
5140 * situations. But it is worth to think about possibility of some
5141 * DoSes using some hypothetical application level deadlock.
5143 if (before(ptr
, tp
->rcv_nxt
))
5146 /* Do we already have a newer (or duplicate) urgent pointer? */
5147 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
5150 /* Tell the world about our new urgent pointer. */
5153 /* We may be adding urgent data when the last byte read was
5154 * urgent. To do this requires some care. We cannot just ignore
5155 * tp->copied_seq since we would read the last urgent byte again
5156 * as data, nor can we alter copied_seq until this data arrives
5157 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5159 * NOTE. Double Dutch. Rendering to plain English: author of comment
5160 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5161 * and expect that both A and B disappear from stream. This is _wrong_.
5162 * Though this happens in BSD with high probability, this is occasional.
5163 * Any application relying on this is buggy. Note also, that fix "works"
5164 * only in this artificial test. Insert some normal data between A and B and we will
5165 * decline of BSD again. Verdict: it is better to remove to trap
5168 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
5169 !sock_flag(sk
, SOCK_URGINLINE
) && tp
->copied_seq
!= tp
->rcv_nxt
) {
5170 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
5172 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5173 __skb_unlink(skb
, &sk
->sk_receive_queue
);
5178 tp
->urg_data
= TCP_URG_NOTYET
;
5181 /* Disable header prediction. */
5185 /* This is the 'fast' part of urgent handling. */
5186 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, const struct tcphdr
*th
)
5188 struct tcp_sock
*tp
= tcp_sk(sk
);
5190 /* Check if we get a new urgent pointer - normally not. */
5192 tcp_check_urg(sk
, th
);
5194 /* Do we wait for any urgent data? - normally not... */
5195 if (tp
->urg_data
== TCP_URG_NOTYET
) {
5196 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
5199 /* Is the urgent pointer pointing into this packet? */
5200 if (ptr
< skb
->len
) {
5202 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
5204 tp
->urg_data
= TCP_URG_VALID
| tmp
;
5205 if (!sock_flag(sk
, SOCK_DEAD
))
5206 sk
->sk_data_ready(sk
);
5211 static int tcp_copy_to_iovec(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
5213 struct tcp_sock
*tp
= tcp_sk(sk
);
5214 int chunk
= skb
->len
- hlen
;
5217 if (skb_csum_unnecessary(skb
))
5218 err
= skb_copy_datagram_msg(skb
, hlen
, tp
->ucopy
.msg
, chunk
);
5220 err
= skb_copy_and_csum_datagram_msg(skb
, hlen
, tp
->ucopy
.msg
);
5223 tp
->ucopy
.len
-= chunk
;
5224 tp
->copied_seq
+= chunk
;
5225 tcp_rcv_space_adjust(sk
);
5231 /* Does PAWS and seqno based validation of an incoming segment, flags will
5232 * play significant role here.
5234 static bool tcp_validate_incoming(struct sock
*sk
, struct sk_buff
*skb
,
5235 const struct tcphdr
*th
, int syn_inerr
)
5237 struct tcp_sock
*tp
= tcp_sk(sk
);
5238 bool rst_seq_match
= false;
5240 /* RFC1323: H1. Apply PAWS check first. */
5241 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
5242 tcp_paws_discard(sk
, skb
)) {
5244 NET_INC_STATS(sock_net(sk
), LINUX_MIB_PAWSESTABREJECTED
);
5245 if (!tcp_oow_rate_limited(sock_net(sk
), skb
,
5246 LINUX_MIB_TCPACKSKIPPEDPAWS
,
5247 &tp
->last_oow_ack_time
))
5248 tcp_send_dupack(sk
, skb
);
5251 /* Reset is accepted even if it did not pass PAWS. */
5254 /* Step 1: check sequence number */
5255 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5256 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5257 * (RST) segments are validated by checking their SEQ-fields."
5258 * And page 69: "If an incoming segment is not acceptable,
5259 * an acknowledgment should be sent in reply (unless the RST
5260 * bit is set, if so drop the segment and return)".
5265 if (!tcp_oow_rate_limited(sock_net(sk
), skb
,
5266 LINUX_MIB_TCPACKSKIPPEDSEQ
,
5267 &tp
->last_oow_ack_time
))
5268 tcp_send_dupack(sk
, skb
);
5273 /* Step 2: check RST bit */
5275 /* RFC 5961 3.2 (extend to match against SACK too if available):
5276 * If seq num matches RCV.NXT or the right-most SACK block,
5278 * RESET the connection
5280 * Send a challenge ACK
5282 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
5283 rst_seq_match
= true;
5284 } else if (tcp_is_sack(tp
) && tp
->rx_opt
.num_sacks
> 0) {
5285 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
5286 int max_sack
= sp
[0].end_seq
;
5289 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
;
5291 max_sack
= after(sp
[this_sack
].end_seq
,
5293 sp
[this_sack
].end_seq
: max_sack
;
5296 if (TCP_SKB_CB(skb
)->seq
== max_sack
)
5297 rst_seq_match
= true;
5303 tcp_send_challenge_ack(sk
, skb
);
5307 /* step 3: check security and precedence [ignored] */
5309 /* step 4: Check for a SYN
5310 * RFC 5961 4.2 : Send a challenge ack
5315 TCP_INC_STATS(sock_net(sk
), TCP_MIB_INERRS
);
5316 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPSYNCHALLENGE
);
5317 tcp_send_challenge_ack(sk
, skb
);
5329 * TCP receive function for the ESTABLISHED state.
5331 * It is split into a fast path and a slow path. The fast path is
5333 * - A zero window was announced from us - zero window probing
5334 * is only handled properly in the slow path.
5335 * - Out of order segments arrived.
5336 * - Urgent data is expected.
5337 * - There is no buffer space left
5338 * - Unexpected TCP flags/window values/header lengths are received
5339 * (detected by checking the TCP header against pred_flags)
5340 * - Data is sent in both directions. Fast path only supports pure senders
5341 * or pure receivers (this means either the sequence number or the ack
5342 * value must stay constant)
5343 * - Unexpected TCP option.
5345 * When these conditions are not satisfied it drops into a standard
5346 * receive procedure patterned after RFC793 to handle all cases.
5347 * The first three cases are guaranteed by proper pred_flags setting,
5348 * the rest is checked inline. Fast processing is turned on in
5349 * tcp_data_queue when everything is OK.
5351 void tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
5352 const struct tcphdr
*th
, unsigned int len
)
5354 struct tcp_sock
*tp
= tcp_sk(sk
);
5356 if (unlikely(!sk
->sk_rx_dst
))
5357 inet_csk(sk
)->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
5359 * Header prediction.
5360 * The code loosely follows the one in the famous
5361 * "30 instruction TCP receive" Van Jacobson mail.
5363 * Van's trick is to deposit buffers into socket queue
5364 * on a device interrupt, to call tcp_recv function
5365 * on the receive process context and checksum and copy
5366 * the buffer to user space. smart...
5368 * Our current scheme is not silly either but we take the
5369 * extra cost of the net_bh soft interrupt processing...
5370 * We do checksum and copy also but from device to kernel.
5373 tp
->rx_opt
.saw_tstamp
= 0;
5375 /* pred_flags is 0xS?10 << 16 + snd_wnd
5376 * if header_prediction is to be made
5377 * 'S' will always be tp->tcp_header_len >> 2
5378 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5379 * turn it off (when there are holes in the receive
5380 * space for instance)
5381 * PSH flag is ignored.
5384 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
5385 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
&&
5386 !after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
)) {
5387 int tcp_header_len
= tp
->tcp_header_len
;
5389 /* Timestamp header prediction: tcp_header_len
5390 * is automatically equal to th->doff*4 due to pred_flags
5394 /* Check timestamp */
5395 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
5396 /* No? Slow path! */
5397 if (!tcp_parse_aligned_timestamp(tp
, th
))
5400 /* If PAWS failed, check it more carefully in slow path */
5401 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
5404 /* DO NOT update ts_recent here, if checksum fails
5405 * and timestamp was corrupted part, it will result
5406 * in a hung connection since we will drop all
5407 * future packets due to the PAWS test.
5411 if (len
<= tcp_header_len
) {
5412 /* Bulk data transfer: sender */
5413 if (len
== tcp_header_len
) {
5414 /* Predicted packet is in window by definition.
5415 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5416 * Hence, check seq<=rcv_wup reduces to:
5418 if (tcp_header_len
==
5419 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5420 tp
->rcv_nxt
== tp
->rcv_wup
)
5421 tcp_store_ts_recent(tp
);
5423 /* We know that such packets are checksummed
5426 tcp_ack(sk
, skb
, 0);
5428 tcp_data_snd_check(sk
);
5430 } else { /* Header too small */
5431 TCP_INC_STATS(sock_net(sk
), TCP_MIB_INERRS
);
5436 bool fragstolen
= false;
5438 if (tp
->ucopy
.task
== current
&&
5439 tp
->copied_seq
== tp
->rcv_nxt
&&
5440 len
- tcp_header_len
<= tp
->ucopy
.len
&&
5441 sock_owned_by_user(sk
)) {
5442 __set_current_state(TASK_RUNNING
);
5444 if (!tcp_copy_to_iovec(sk
, skb
, tcp_header_len
)) {
5445 /* Predicted packet is in window by definition.
5446 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5447 * Hence, check seq<=rcv_wup reduces to:
5449 if (tcp_header_len
==
5450 (sizeof(struct tcphdr
) +
5451 TCPOLEN_TSTAMP_ALIGNED
) &&
5452 tp
->rcv_nxt
== tp
->rcv_wup
)
5453 tcp_store_ts_recent(tp
);
5455 tcp_rcv_rtt_measure_ts(sk
, skb
);
5457 __skb_pull(skb
, tcp_header_len
);
5458 tcp_rcv_nxt_update(tp
, TCP_SKB_CB(skb
)->end_seq
);
5459 NET_INC_STATS(sock_net(sk
),
5460 LINUX_MIB_TCPHPHITSTOUSER
);
5465 if (tcp_checksum_complete(skb
))
5468 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
5471 /* Predicted packet is in window by definition.
5472 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5473 * Hence, check seq<=rcv_wup reduces to:
5475 if (tcp_header_len
==
5476 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5477 tp
->rcv_nxt
== tp
->rcv_wup
)
5478 tcp_store_ts_recent(tp
);
5480 tcp_rcv_rtt_measure_ts(sk
, skb
);
5482 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPHPHITS
);
5484 /* Bulk data transfer: receiver */
5485 eaten
= tcp_queue_rcv(sk
, skb
, tcp_header_len
,
5489 tcp_event_data_recv(sk
, skb
);
5491 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
5492 /* Well, only one small jumplet in fast path... */
5493 tcp_ack(sk
, skb
, FLAG_DATA
);
5494 tcp_data_snd_check(sk
);
5495 if (!inet_csk_ack_scheduled(sk
))
5499 __tcp_ack_snd_check(sk
, 0);
5502 kfree_skb_partial(skb
, fragstolen
);
5503 sk
->sk_data_ready(sk
);
5509 if (len
< (th
->doff
<< 2) || tcp_checksum_complete(skb
))
5512 if (!th
->ack
&& !th
->rst
&& !th
->syn
)
5516 * Standard slow path.
5519 if (!tcp_validate_incoming(sk
, skb
, th
, 1))
5523 if (tcp_ack(sk
, skb
, FLAG_SLOWPATH
| FLAG_UPDATE_TS_RECENT
) < 0)
5526 tcp_rcv_rtt_measure_ts(sk
, skb
);
5528 /* Process urgent data. */
5529 tcp_urg(sk
, skb
, th
);
5531 /* step 7: process the segment text */
5532 tcp_data_queue(sk
, skb
);
5534 tcp_data_snd_check(sk
);
5535 tcp_ack_snd_check(sk
);
5539 TCP_INC_STATS(sock_net(sk
), TCP_MIB_CSUMERRORS
);
5540 TCP_INC_STATS(sock_net(sk
), TCP_MIB_INERRS
);
5545 EXPORT_SYMBOL(tcp_rcv_established
);
5547 void tcp_finish_connect(struct sock
*sk
, struct sk_buff
*skb
)
5549 struct tcp_sock
*tp
= tcp_sk(sk
);
5550 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5552 tcp_set_state(sk
, TCP_ESTABLISHED
);
5555 icsk
->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
5556 security_inet_conn_established(sk
, skb
);
5559 /* Make sure socket is routed, for correct metrics. */
5560 icsk
->icsk_af_ops
->rebuild_header(sk
);
5562 tcp_init_metrics(sk
);
5564 tcp_init_congestion_control(sk
);
5566 /* Prevent spurious tcp_cwnd_restart() on first data
5569 tp
->lsndtime
= tcp_time_stamp
;
5571 tcp_init_buffer_space(sk
);
5573 if (sock_flag(sk
, SOCK_KEEPOPEN
))
5574 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
5576 if (!tp
->rx_opt
.snd_wscale
)
5577 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
5581 if (!sock_flag(sk
, SOCK_DEAD
)) {
5582 sk
->sk_state_change(sk
);
5583 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5587 static bool tcp_rcv_fastopen_synack(struct sock
*sk
, struct sk_buff
*synack
,
5588 struct tcp_fastopen_cookie
*cookie
)
5590 struct tcp_sock
*tp
= tcp_sk(sk
);
5591 struct sk_buff
*data
= tp
->syn_data
? tcp_write_queue_head(sk
) : NULL
;
5592 u16 mss
= tp
->rx_opt
.mss_clamp
, try_exp
= 0;
5593 bool syn_drop
= false;
5595 if (mss
== tp
->rx_opt
.user_mss
) {
5596 struct tcp_options_received opt
;
5598 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5599 tcp_clear_options(&opt
);
5600 opt
.user_mss
= opt
.mss_clamp
= 0;
5601 tcp_parse_options(synack
, &opt
, 0, NULL
);
5602 mss
= opt
.mss_clamp
;
5605 if (!tp
->syn_fastopen
) {
5606 /* Ignore an unsolicited cookie */
5608 } else if (tp
->total_retrans
) {
5609 /* SYN timed out and the SYN-ACK neither has a cookie nor
5610 * acknowledges data. Presumably the remote received only
5611 * the retransmitted (regular) SYNs: either the original
5612 * SYN-data or the corresponding SYN-ACK was dropped.
5614 syn_drop
= (cookie
->len
< 0 && data
);
5615 } else if (cookie
->len
< 0 && !tp
->syn_data
) {
5616 /* We requested a cookie but didn't get it. If we did not use
5617 * the (old) exp opt format then try so next time (try_exp=1).
5618 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5620 try_exp
= tp
->syn_fastopen_exp
? 2 : 1;
5623 tcp_fastopen_cache_set(sk
, mss
, cookie
, syn_drop
, try_exp
);
5625 if (data
) { /* Retransmit unacked data in SYN */
5626 tcp_for_write_queue_from(data
, sk
) {
5627 if (data
== tcp_send_head(sk
) ||
5628 __tcp_retransmit_skb(sk
, data
, 1))
5632 NET_INC_STATS(sock_net(sk
),
5633 LINUX_MIB_TCPFASTOPENACTIVEFAIL
);
5636 tp
->syn_data_acked
= tp
->syn_data
;
5637 if (tp
->syn_data_acked
)
5638 NET_INC_STATS(sock_net(sk
),
5639 LINUX_MIB_TCPFASTOPENACTIVE
);
5641 tcp_fastopen_add_skb(sk
, synack
);
5646 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5647 const struct tcphdr
*th
)
5649 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5650 struct tcp_sock
*tp
= tcp_sk(sk
);
5651 struct tcp_fastopen_cookie foc
= { .len
= -1 };
5652 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
5654 tcp_parse_options(skb
, &tp
->rx_opt
, 0, &foc
);
5655 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
5656 tp
->rx_opt
.rcv_tsecr
-= tp
->tsoffset
;
5660 * "If the state is SYN-SENT then
5661 * first check the ACK bit
5662 * If the ACK bit is set
5663 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5664 * a reset (unless the RST bit is set, if so drop
5665 * the segment and return)"
5667 if (!after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_una
) ||
5668 after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
))
5669 goto reset_and_undo
;
5671 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
5672 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
5674 NET_INC_STATS(sock_net(sk
),
5675 LINUX_MIB_PAWSACTIVEREJECTED
);
5676 goto reset_and_undo
;
5679 /* Now ACK is acceptable.
5681 * "If the RST bit is set
5682 * If the ACK was acceptable then signal the user "error:
5683 * connection reset", drop the segment, enter CLOSED state,
5684 * delete TCB, and return."
5693 * "fifth, if neither of the SYN or RST bits is set then
5694 * drop the segment and return."
5700 goto discard_and_undo
;
5703 * "If the SYN bit is on ...
5704 * are acceptable then ...
5705 * (our SYN has been ACKed), change the connection
5706 * state to ESTABLISHED..."
5709 tcp_ecn_rcv_synack(tp
, th
);
5711 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5712 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
5714 /* Ok.. it's good. Set up sequence numbers and
5715 * move to established.
5717 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5718 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5720 /* RFC1323: The window in SYN & SYN/ACK segments is
5723 tp
->snd_wnd
= ntohs(th
->window
);
5725 if (!tp
->rx_opt
.wscale_ok
) {
5726 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
5727 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
5730 if (tp
->rx_opt
.saw_tstamp
) {
5731 tp
->rx_opt
.tstamp_ok
= 1;
5732 tp
->tcp_header_len
=
5733 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5734 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5735 tcp_store_ts_recent(tp
);
5737 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5740 if (tcp_is_sack(tp
) && sysctl_tcp_fack
)
5741 tcp_enable_fack(tp
);
5744 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5745 tcp_initialize_rcv_mss(sk
);
5747 /* Remember, tcp_poll() does not lock socket!
5748 * Change state from SYN-SENT only after copied_seq
5749 * is initialized. */
5750 tp
->copied_seq
= tp
->rcv_nxt
;
5754 tcp_finish_connect(sk
, skb
);
5756 if ((tp
->syn_fastopen
|| tp
->syn_data
) &&
5757 tcp_rcv_fastopen_synack(sk
, skb
, &foc
))
5760 if (sk
->sk_write_pending
||
5761 icsk
->icsk_accept_queue
.rskq_defer_accept
||
5762 icsk
->icsk_ack
.pingpong
) {
5763 /* Save one ACK. Data will be ready after
5764 * several ticks, if write_pending is set.
5766 * It may be deleted, but with this feature tcpdumps
5767 * look so _wonderfully_ clever, that I was not able
5768 * to stand against the temptation 8) --ANK
5770 inet_csk_schedule_ack(sk
);
5771 icsk
->icsk_ack
.lrcvtime
= tcp_time_stamp
;
5772 tcp_enter_quickack_mode(sk
);
5773 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
5774 TCP_DELACK_MAX
, TCP_RTO_MAX
);
5785 /* No ACK in the segment */
5789 * "If the RST bit is set
5791 * Otherwise (no ACK) drop the segment and return."
5794 goto discard_and_undo
;
5798 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&&
5799 tcp_paws_reject(&tp
->rx_opt
, 0))
5800 goto discard_and_undo
;
5803 /* We see SYN without ACK. It is attempt of
5804 * simultaneous connect with crossed SYNs.
5805 * Particularly, it can be connect to self.
5807 tcp_set_state(sk
, TCP_SYN_RECV
);
5809 if (tp
->rx_opt
.saw_tstamp
) {
5810 tp
->rx_opt
.tstamp_ok
= 1;
5811 tcp_store_ts_recent(tp
);
5812 tp
->tcp_header_len
=
5813 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5815 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5818 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5819 tp
->copied_seq
= tp
->rcv_nxt
;
5820 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5822 /* RFC1323: The window in SYN & SYN/ACK segments is
5825 tp
->snd_wnd
= ntohs(th
->window
);
5826 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
5827 tp
->max_window
= tp
->snd_wnd
;
5829 tcp_ecn_rcv_syn(tp
, th
);
5832 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5833 tcp_initialize_rcv_mss(sk
);
5835 tcp_send_synack(sk
);
5837 /* Note, we could accept data and URG from this segment.
5838 * There are no obstacles to make this (except that we must
5839 * either change tcp_recvmsg() to prevent it from returning data
5840 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5842 * However, if we ignore data in ACKless segments sometimes,
5843 * we have no reasons to accept it sometimes.
5844 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5845 * is not flawless. So, discard packet for sanity.
5846 * Uncomment this return to process the data.
5853 /* "fifth, if neither of the SYN or RST bits is set then
5854 * drop the segment and return."
5858 tcp_clear_options(&tp
->rx_opt
);
5859 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5863 tcp_clear_options(&tp
->rx_opt
);
5864 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5869 * This function implements the receiving procedure of RFC 793 for
5870 * all states except ESTABLISHED and TIME_WAIT.
5871 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5872 * address independent.
5875 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
)
5877 struct tcp_sock
*tp
= tcp_sk(sk
);
5878 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5879 const struct tcphdr
*th
= tcp_hdr(skb
);
5880 struct request_sock
*req
;
5884 switch (sk
->sk_state
) {
5898 if (icsk
->icsk_af_ops
->conn_request(sk
, skb
) < 0)
5907 tp
->rx_opt
.saw_tstamp
= 0;
5908 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
);
5912 /* Do step6 onward by hand. */
5913 tcp_urg(sk
, skb
, th
);
5915 tcp_data_snd_check(sk
);
5919 tp
->rx_opt
.saw_tstamp
= 0;
5920 req
= tp
->fastopen_rsk
;
5922 WARN_ON_ONCE(sk
->sk_state
!= TCP_SYN_RECV
&&
5923 sk
->sk_state
!= TCP_FIN_WAIT1
);
5925 if (!tcp_check_req(sk
, skb
, req
, true))
5929 if (!th
->ack
&& !th
->rst
&& !th
->syn
)
5932 if (!tcp_validate_incoming(sk
, skb
, th
, 0))
5935 /* step 5: check the ACK field */
5936 acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
|
5937 FLAG_UPDATE_TS_RECENT
) > 0;
5939 switch (sk
->sk_state
) {
5945 tcp_synack_rtt_meas(sk
, req
);
5947 /* Once we leave TCP_SYN_RECV, we no longer need req
5951 inet_csk(sk
)->icsk_retransmits
= 0;
5952 reqsk_fastopen_remove(sk
, req
, false);
5954 /* Make sure socket is routed, for correct metrics. */
5955 icsk
->icsk_af_ops
->rebuild_header(sk
);
5956 tcp_init_congestion_control(sk
);
5959 tp
->copied_seq
= tp
->rcv_nxt
;
5960 tcp_init_buffer_space(sk
);
5963 tcp_set_state(sk
, TCP_ESTABLISHED
);
5964 sk
->sk_state_change(sk
);
5966 /* Note, that this wakeup is only for marginal crossed SYN case.
5967 * Passively open sockets are not waked up, because
5968 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5971 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5973 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
5974 tp
->snd_wnd
= ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
;
5975 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5977 if (tp
->rx_opt
.tstamp_ok
)
5978 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5981 /* Re-arm the timer because data may have been sent out.
5982 * This is similar to the regular data transmission case
5983 * when new data has just been ack'ed.
5985 * (TFO) - we could try to be more aggressive and
5986 * retransmitting any data sooner based on when they
5991 tcp_init_metrics(sk
);
5993 if (!inet_csk(sk
)->icsk_ca_ops
->cong_control
)
5994 tcp_update_pacing_rate(sk
);
5996 /* Prevent spurious tcp_cwnd_restart() on first data packet */
5997 tp
->lsndtime
= tcp_time_stamp
;
5999 tcp_initialize_rcv_mss(sk
);
6000 tcp_fast_path_on(tp
);
6003 case TCP_FIN_WAIT1
: {
6004 struct dst_entry
*dst
;
6007 /* If we enter the TCP_FIN_WAIT1 state and we are a
6008 * Fast Open socket and this is the first acceptable
6009 * ACK we have received, this would have acknowledged
6010 * our SYNACK so stop the SYNACK timer.
6013 /* Return RST if ack_seq is invalid.
6014 * Note that RFC793 only says to generate a
6015 * DUPACK for it but for TCP Fast Open it seems
6016 * better to treat this case like TCP_SYN_RECV
6021 /* We no longer need the request sock. */
6022 reqsk_fastopen_remove(sk
, req
, false);
6025 if (tp
->snd_una
!= tp
->write_seq
)
6028 tcp_set_state(sk
, TCP_FIN_WAIT2
);
6029 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
6031 dst
= __sk_dst_get(sk
);
6035 if (!sock_flag(sk
, SOCK_DEAD
)) {
6036 /* Wake up lingering close() */
6037 sk
->sk_state_change(sk
);
6041 if (tp
->linger2
< 0 ||
6042 (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
6043 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
))) {
6045 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
6049 tmo
= tcp_fin_time(sk
);
6050 if (tmo
> TCP_TIMEWAIT_LEN
) {
6051 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
6052 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
6053 /* Bad case. We could lose such FIN otherwise.
6054 * It is not a big problem, but it looks confusing
6055 * and not so rare event. We still can lose it now,
6056 * if it spins in bh_lock_sock(), but it is really
6059 inet_csk_reset_keepalive_timer(sk
, tmo
);
6061 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
6068 if (tp
->snd_una
== tp
->write_seq
) {
6069 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
6075 if (tp
->snd_una
== tp
->write_seq
) {
6076 tcp_update_metrics(sk
);
6083 /* step 6: check the URG bit */
6084 tcp_urg(sk
, skb
, th
);
6086 /* step 7: process the segment text */
6087 switch (sk
->sk_state
) {
6088 case TCP_CLOSE_WAIT
:
6091 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
6095 /* RFC 793 says to queue data in these states,
6096 * RFC 1122 says we MUST send a reset.
6097 * BSD 4.4 also does reset.
6099 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
6100 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
6101 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
6102 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
6108 case TCP_ESTABLISHED
:
6109 tcp_data_queue(sk
, skb
);
6114 /* tcp_data could move socket to TIME-WAIT */
6115 if (sk
->sk_state
!= TCP_CLOSE
) {
6116 tcp_data_snd_check(sk
);
6117 tcp_ack_snd_check(sk
);
6126 EXPORT_SYMBOL(tcp_rcv_state_process
);
6128 static inline void pr_drop_req(struct request_sock
*req
, __u16 port
, int family
)
6130 struct inet_request_sock
*ireq
= inet_rsk(req
);
6132 if (family
== AF_INET
)
6133 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6134 &ireq
->ir_rmt_addr
, port
);
6135 #if IS_ENABLED(CONFIG_IPV6)
6136 else if (family
== AF_INET6
)
6137 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6138 &ireq
->ir_v6_rmt_addr
, port
);
6142 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6144 * If we receive a SYN packet with these bits set, it means a
6145 * network is playing bad games with TOS bits. In order to
6146 * avoid possible false congestion notifications, we disable
6147 * TCP ECN negotiation.
6149 * Exception: tcp_ca wants ECN. This is required for DCTCP
6150 * congestion control: Linux DCTCP asserts ECT on all packets,
6151 * including SYN, which is most optimal solution; however,
6152 * others, such as FreeBSD do not.
6154 static void tcp_ecn_create_request(struct request_sock
*req
,
6155 const struct sk_buff
*skb
,
6156 const struct sock
*listen_sk
,
6157 const struct dst_entry
*dst
)
6159 const struct tcphdr
*th
= tcp_hdr(skb
);
6160 const struct net
*net
= sock_net(listen_sk
);
6161 bool th_ecn
= th
->ece
&& th
->cwr
;
6168 ect
= !INET_ECN_is_not_ect(TCP_SKB_CB(skb
)->ip_dsfield
);
6169 ecn_ok_dst
= dst_feature(dst
, DST_FEATURE_ECN_MASK
);
6170 ecn_ok
= net
->ipv4
.sysctl_tcp_ecn
|| ecn_ok_dst
;
6172 if ((!ect
&& ecn_ok
) || tcp_ca_needs_ecn(listen_sk
) ||
6173 (ecn_ok_dst
& DST_FEATURE_ECN_CA
))
6174 inet_rsk(req
)->ecn_ok
= 1;
6177 static void tcp_openreq_init(struct request_sock
*req
,
6178 const struct tcp_options_received
*rx_opt
,
6179 struct sk_buff
*skb
, const struct sock
*sk
)
6181 struct inet_request_sock
*ireq
= inet_rsk(req
);
6183 req
->rsk_rcv_wnd
= 0; /* So that tcp_send_synack() knows! */
6185 tcp_rsk(req
)->rcv_isn
= TCP_SKB_CB(skb
)->seq
;
6186 tcp_rsk(req
)->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
6187 skb_mstamp_get(&tcp_rsk(req
)->snt_synack
);
6188 tcp_rsk(req
)->last_oow_ack_time
= 0;
6189 req
->mss
= rx_opt
->mss_clamp
;
6190 req
->ts_recent
= rx_opt
->saw_tstamp
? rx_opt
->rcv_tsval
: 0;
6191 ireq
->tstamp_ok
= rx_opt
->tstamp_ok
;
6192 ireq
->sack_ok
= rx_opt
->sack_ok
;
6193 ireq
->snd_wscale
= rx_opt
->snd_wscale
;
6194 ireq
->wscale_ok
= rx_opt
->wscale_ok
;
6197 ireq
->ir_rmt_port
= tcp_hdr(skb
)->source
;
6198 ireq
->ir_num
= ntohs(tcp_hdr(skb
)->dest
);
6199 ireq
->ir_mark
= inet_request_mark(sk
, skb
);
6202 struct request_sock
*inet_reqsk_alloc(const struct request_sock_ops
*ops
,
6203 struct sock
*sk_listener
,
6204 bool attach_listener
)
6206 struct request_sock
*req
= reqsk_alloc(ops
, sk_listener
,
6210 struct inet_request_sock
*ireq
= inet_rsk(req
);
6212 kmemcheck_annotate_bitfield(ireq
, flags
);
6214 #if IS_ENABLED(CONFIG_IPV6)
6215 ireq
->pktopts
= NULL
;
6217 atomic64_set(&ireq
->ir_cookie
, 0);
6218 ireq
->ireq_state
= TCP_NEW_SYN_RECV
;
6219 write_pnet(&ireq
->ireq_net
, sock_net(sk_listener
));
6220 ireq
->ireq_family
= sk_listener
->sk_family
;
6225 EXPORT_SYMBOL(inet_reqsk_alloc
);
6228 * Return true if a syncookie should be sent
6230 static bool tcp_syn_flood_action(const struct sock
*sk
,
6231 const struct sk_buff
*skb
,
6234 struct request_sock_queue
*queue
= &inet_csk(sk
)->icsk_accept_queue
;
6235 const char *msg
= "Dropping request";
6236 bool want_cookie
= false;
6237 struct net
*net
= sock_net(sk
);
6239 #ifdef CONFIG_SYN_COOKIES
6240 if (net
->ipv4
.sysctl_tcp_syncookies
) {
6241 msg
= "Sending cookies";
6243 __NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPREQQFULLDOCOOKIES
);
6246 __NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPREQQFULLDROP
);
6248 if (!queue
->synflood_warned
&&
6249 net
->ipv4
.sysctl_tcp_syncookies
!= 2 &&
6250 xchg(&queue
->synflood_warned
, 1) == 0)
6251 pr_info("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6252 proto
, ntohs(tcp_hdr(skb
)->dest
), msg
);
6257 static void tcp_reqsk_record_syn(const struct sock
*sk
,
6258 struct request_sock
*req
,
6259 const struct sk_buff
*skb
)
6261 if (tcp_sk(sk
)->save_syn
) {
6262 u32 len
= skb_network_header_len(skb
) + tcp_hdrlen(skb
);
6265 copy
= kmalloc(len
+ sizeof(u32
), GFP_ATOMIC
);
6268 memcpy(©
[1], skb_network_header(skb
), len
);
6269 req
->saved_syn
= copy
;
6274 int tcp_conn_request(struct request_sock_ops
*rsk_ops
,
6275 const struct tcp_request_sock_ops
*af_ops
,
6276 struct sock
*sk
, struct sk_buff
*skb
)
6278 struct tcp_fastopen_cookie foc
= { .len
= -1 };
6279 __u32 isn
= TCP_SKB_CB(skb
)->tcp_tw_isn
;
6280 struct tcp_options_received tmp_opt
;
6281 struct tcp_sock
*tp
= tcp_sk(sk
);
6282 struct net
*net
= sock_net(sk
);
6283 struct sock
*fastopen_sk
= NULL
;
6284 struct dst_entry
*dst
= NULL
;
6285 struct request_sock
*req
;
6286 bool want_cookie
= false;
6289 /* TW buckets are converted to open requests without
6290 * limitations, they conserve resources and peer is
6291 * evidently real one.
6293 if ((net
->ipv4
.sysctl_tcp_syncookies
== 2 ||
6294 inet_csk_reqsk_queue_is_full(sk
)) && !isn
) {
6295 want_cookie
= tcp_syn_flood_action(sk
, skb
, rsk_ops
->slab_name
);
6300 if (sk_acceptq_is_full(sk
)) {
6301 NET_INC_STATS(sock_net(sk
), LINUX_MIB_LISTENOVERFLOWS
);
6305 req
= inet_reqsk_alloc(rsk_ops
, sk
, !want_cookie
);
6309 tcp_rsk(req
)->af_specific
= af_ops
;
6311 tcp_clear_options(&tmp_opt
);
6312 tmp_opt
.mss_clamp
= af_ops
->mss_clamp
;
6313 tmp_opt
.user_mss
= tp
->rx_opt
.user_mss
;
6314 tcp_parse_options(skb
, &tmp_opt
, 0, want_cookie
? NULL
: &foc
);
6316 if (want_cookie
&& !tmp_opt
.saw_tstamp
)
6317 tcp_clear_options(&tmp_opt
);
6319 tmp_opt
.tstamp_ok
= tmp_opt
.saw_tstamp
;
6320 tcp_openreq_init(req
, &tmp_opt
, skb
, sk
);
6321 inet_rsk(req
)->no_srccheck
= inet_sk(sk
)->transparent
;
6323 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6324 inet_rsk(req
)->ir_iif
= inet_request_bound_dev_if(sk
, skb
);
6326 af_ops
->init_req(req
, sk
, skb
);
6328 if (security_inet_conn_request(sk
, skb
, req
))
6331 if (!want_cookie
&& !isn
) {
6332 /* VJ's idea. We save last timestamp seen
6333 * from the destination in peer table, when entering
6334 * state TIME-WAIT, and check against it before
6335 * accepting new connection request.
6337 * If "isn" is not zero, this request hit alive
6338 * timewait bucket, so that all the necessary checks
6339 * are made in the function processing timewait state.
6341 if (tcp_death_row
.sysctl_tw_recycle
) {
6344 dst
= af_ops
->route_req(sk
, &fl
, req
, &strict
);
6346 if (dst
&& strict
&&
6347 !tcp_peer_is_proven(req
, dst
, true,
6348 tmp_opt
.saw_tstamp
)) {
6349 NET_INC_STATS(sock_net(sk
), LINUX_MIB_PAWSPASSIVEREJECTED
);
6350 goto drop_and_release
;
6353 /* Kill the following clause, if you dislike this way. */
6354 else if (!net
->ipv4
.sysctl_tcp_syncookies
&&
6355 (sysctl_max_syn_backlog
- inet_csk_reqsk_queue_len(sk
) <
6356 (sysctl_max_syn_backlog
>> 2)) &&
6357 !tcp_peer_is_proven(req
, dst
, false,
6358 tmp_opt
.saw_tstamp
)) {
6359 /* Without syncookies last quarter of
6360 * backlog is filled with destinations,
6361 * proven to be alive.
6362 * It means that we continue to communicate
6363 * to destinations, already remembered
6364 * to the moment of synflood.
6366 pr_drop_req(req
, ntohs(tcp_hdr(skb
)->source
),
6368 goto drop_and_release
;
6371 isn
= af_ops
->init_seq(skb
);
6374 dst
= af_ops
->route_req(sk
, &fl
, req
, NULL
);
6379 tcp_ecn_create_request(req
, skb
, sk
, dst
);
6382 isn
= cookie_init_sequence(af_ops
, sk
, skb
, &req
->mss
);
6383 req
->cookie_ts
= tmp_opt
.tstamp_ok
;
6384 if (!tmp_opt
.tstamp_ok
)
6385 inet_rsk(req
)->ecn_ok
= 0;
6388 tcp_rsk(req
)->snt_isn
= isn
;
6389 tcp_rsk(req
)->txhash
= net_tx_rndhash();
6390 tcp_openreq_init_rwin(req
, sk
, dst
);
6392 tcp_reqsk_record_syn(sk
, req
, skb
);
6393 fastopen_sk
= tcp_try_fastopen(sk
, skb
, req
, &foc
, dst
);
6396 af_ops
->send_synack(fastopen_sk
, dst
, &fl
, req
,
6397 &foc
, TCP_SYNACK_FASTOPEN
);
6398 /* Add the child socket directly into the accept queue */
6399 inet_csk_reqsk_queue_add(sk
, req
, fastopen_sk
);
6400 sk
->sk_data_ready(sk
);
6401 bh_unlock_sock(fastopen_sk
);
6402 sock_put(fastopen_sk
);
6404 tcp_rsk(req
)->tfo_listener
= false;
6406 inet_csk_reqsk_queue_hash_add(sk
, req
, TCP_TIMEOUT_INIT
);
6407 af_ops
->send_synack(sk
, dst
, &fl
, req
, &foc
,
6408 !want_cookie
? TCP_SYNACK_NORMAL
:
6426 EXPORT_SYMBOL(tcp_conn_request
);