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_fack __read_mostly
;
80 int sysctl_tcp_max_reordering __read_mostly
= 300;
81 int sysctl_tcp_dsack __read_mostly
= 1;
82 int sysctl_tcp_app_win __read_mostly
= 31;
83 int sysctl_tcp_adv_win_scale __read_mostly
= 1;
84 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale
);
86 /* rfc5961 challenge ack rate limiting */
87 int sysctl_tcp_challenge_ack_limit
= 1000;
89 int sysctl_tcp_stdurg __read_mostly
;
90 int sysctl_tcp_rfc1337 __read_mostly
;
91 int sysctl_tcp_max_orphans __read_mostly
= NR_FILE
;
92 int sysctl_tcp_frto __read_mostly
= 2;
93 int sysctl_tcp_min_rtt_wlen __read_mostly
= 300;
94 int sysctl_tcp_moderate_rcvbuf __read_mostly
= 1;
95 int sysctl_tcp_early_retrans __read_mostly
= 3;
96 int sysctl_tcp_invalid_ratelimit __read_mostly
= HZ
/2;
98 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
99 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
100 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
101 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
102 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
103 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
104 #define FLAG_ECE 0x40 /* ECE in this ACK */
105 #define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */
106 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
107 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
108 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
109 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
110 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
111 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
112 #define FLAG_NO_CHALLENGE_ACK 0x8000 /* do not call tcp_send_challenge_ack() */
114 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
115 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
116 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
117 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
119 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
120 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
122 #define REXMIT_NONE 0 /* no loss recovery to do */
123 #define REXMIT_LOST 1 /* retransmit packets marked lost */
124 #define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */
126 static void tcp_gro_dev_warn(struct sock
*sk
, const struct sk_buff
*skb
,
129 static bool __once __read_mostly
;
132 struct net_device
*dev
;
137 dev
= dev_get_by_index_rcu(sock_net(sk
), skb
->skb_iif
);
138 if (!dev
|| len
>= dev
->mtu
)
139 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n",
140 dev
? dev
->name
: "Unknown driver");
145 /* Adapt the MSS value used to make delayed ack decision to the
148 static void tcp_measure_rcv_mss(struct sock
*sk
, const struct sk_buff
*skb
)
150 struct inet_connection_sock
*icsk
= inet_csk(sk
);
151 const unsigned int lss
= icsk
->icsk_ack
.last_seg_size
;
154 icsk
->icsk_ack
.last_seg_size
= 0;
156 /* skb->len may jitter because of SACKs, even if peer
157 * sends good full-sized frames.
159 len
= skb_shinfo(skb
)->gso_size
? : skb
->len
;
160 if (len
>= icsk
->icsk_ack
.rcv_mss
) {
161 icsk
->icsk_ack
.rcv_mss
= min_t(unsigned int, len
,
163 /* Account for possibly-removed options */
164 if (unlikely(len
> icsk
->icsk_ack
.rcv_mss
+
165 MAX_TCP_OPTION_SPACE
))
166 tcp_gro_dev_warn(sk
, skb
, len
);
168 /* Otherwise, we make more careful check taking into account,
169 * that SACKs block is variable.
171 * "len" is invariant segment length, including TCP header.
173 len
+= skb
->data
- skb_transport_header(skb
);
174 if (len
>= TCP_MSS_DEFAULT
+ sizeof(struct tcphdr
) ||
175 /* If PSH is not set, packet should be
176 * full sized, provided peer TCP is not badly broken.
177 * This observation (if it is correct 8)) allows
178 * to handle super-low mtu links fairly.
180 (len
>= TCP_MIN_MSS
+ sizeof(struct tcphdr
) &&
181 !(tcp_flag_word(tcp_hdr(skb
)) & TCP_REMNANT
))) {
182 /* Subtract also invariant (if peer is RFC compliant),
183 * tcp header plus fixed timestamp option length.
184 * Resulting "len" is MSS free of SACK jitter.
186 len
-= tcp_sk(sk
)->tcp_header_len
;
187 icsk
->icsk_ack
.last_seg_size
= len
;
189 icsk
->icsk_ack
.rcv_mss
= len
;
193 if (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
)
194 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED2
;
195 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
199 static void tcp_incr_quickack(struct sock
*sk
)
201 struct inet_connection_sock
*icsk
= inet_csk(sk
);
202 unsigned int quickacks
= tcp_sk(sk
)->rcv_wnd
/ (2 * icsk
->icsk_ack
.rcv_mss
);
206 if (quickacks
> icsk
->icsk_ack
.quick
)
207 icsk
->icsk_ack
.quick
= min(quickacks
, TCP_MAX_QUICKACKS
);
210 static void tcp_enter_quickack_mode(struct sock
*sk
)
212 struct inet_connection_sock
*icsk
= inet_csk(sk
);
213 tcp_incr_quickack(sk
);
214 icsk
->icsk_ack
.pingpong
= 0;
215 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
218 /* Send ACKs quickly, if "quick" count is not exhausted
219 * and the session is not interactive.
222 static bool tcp_in_quickack_mode(struct sock
*sk
)
224 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
225 const struct dst_entry
*dst
= __sk_dst_get(sk
);
227 return (dst
&& dst_metric(dst
, RTAX_QUICKACK
)) ||
228 (icsk
->icsk_ack
.quick
&& !icsk
->icsk_ack
.pingpong
);
231 static void tcp_ecn_queue_cwr(struct tcp_sock
*tp
)
233 if (tp
->ecn_flags
& TCP_ECN_OK
)
234 tp
->ecn_flags
|= TCP_ECN_QUEUE_CWR
;
237 static void tcp_ecn_accept_cwr(struct tcp_sock
*tp
, const struct sk_buff
*skb
)
239 if (tcp_hdr(skb
)->cwr
)
240 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
243 static void tcp_ecn_withdraw_cwr(struct tcp_sock
*tp
)
245 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
248 static void __tcp_ecn_check_ce(struct tcp_sock
*tp
, const struct sk_buff
*skb
)
250 switch (TCP_SKB_CB(skb
)->ip_dsfield
& INET_ECN_MASK
) {
251 case INET_ECN_NOT_ECT
:
252 /* Funny extension: if ECT is not set on a segment,
253 * and we already seen ECT on a previous segment,
254 * it is probably a retransmit.
256 if (tp
->ecn_flags
& TCP_ECN_SEEN
)
257 tcp_enter_quickack_mode((struct sock
*)tp
);
260 if (tcp_ca_needs_ecn((struct sock
*)tp
))
261 tcp_ca_event((struct sock
*)tp
, CA_EVENT_ECN_IS_CE
);
263 if (!(tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)) {
264 /* Better not delay acks, sender can have a very low cwnd */
265 tcp_enter_quickack_mode((struct sock
*)tp
);
266 tp
->ecn_flags
|= TCP_ECN_DEMAND_CWR
;
268 tp
->ecn_flags
|= TCP_ECN_SEEN
;
271 if (tcp_ca_needs_ecn((struct sock
*)tp
))
272 tcp_ca_event((struct sock
*)tp
, CA_EVENT_ECN_NO_CE
);
273 tp
->ecn_flags
|= TCP_ECN_SEEN
;
278 static void tcp_ecn_check_ce(struct tcp_sock
*tp
, const struct sk_buff
*skb
)
280 if (tp
->ecn_flags
& TCP_ECN_OK
)
281 __tcp_ecn_check_ce(tp
, skb
);
284 static void tcp_ecn_rcv_synack(struct tcp_sock
*tp
, const struct tcphdr
*th
)
286 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| th
->cwr
))
287 tp
->ecn_flags
&= ~TCP_ECN_OK
;
290 static void tcp_ecn_rcv_syn(struct tcp_sock
*tp
, const struct tcphdr
*th
)
292 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| !th
->cwr
))
293 tp
->ecn_flags
&= ~TCP_ECN_OK
;
296 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock
*tp
, const struct tcphdr
*th
)
298 if (th
->ece
&& !th
->syn
&& (tp
->ecn_flags
& TCP_ECN_OK
))
303 /* Buffer size and advertised window tuning.
305 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
308 static void tcp_sndbuf_expand(struct sock
*sk
)
310 const struct tcp_sock
*tp
= tcp_sk(sk
);
311 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
315 /* Worst case is non GSO/TSO : each frame consumes one skb
316 * and skb->head is kmalloced using power of two area of memory
318 per_mss
= max_t(u32
, tp
->rx_opt
.mss_clamp
, tp
->mss_cache
) +
320 SKB_DATA_ALIGN(sizeof(struct skb_shared_info
));
322 per_mss
= roundup_pow_of_two(per_mss
) +
323 SKB_DATA_ALIGN(sizeof(struct sk_buff
));
325 nr_segs
= max_t(u32
, TCP_INIT_CWND
, tp
->snd_cwnd
);
326 nr_segs
= max_t(u32
, nr_segs
, tp
->reordering
+ 1);
328 /* Fast Recovery (RFC 5681 3.2) :
329 * Cubic needs 1.7 factor, rounded to 2 to include
330 * extra cushion (application might react slowly to POLLOUT)
332 sndmem
= ca_ops
->sndbuf_expand
? ca_ops
->sndbuf_expand(sk
) : 2;
333 sndmem
*= nr_segs
* per_mss
;
335 if (sk
->sk_sndbuf
< sndmem
)
336 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
339 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
341 * All tcp_full_space() is split to two parts: "network" buffer, allocated
342 * forward and advertised in receiver window (tp->rcv_wnd) and
343 * "application buffer", required to isolate scheduling/application
344 * latencies from network.
345 * window_clamp is maximal advertised window. It can be less than
346 * tcp_full_space(), in this case tcp_full_space() - window_clamp
347 * is reserved for "application" buffer. The less window_clamp is
348 * the smoother our behaviour from viewpoint of network, but the lower
349 * throughput and the higher sensitivity of the connection to losses. 8)
351 * rcv_ssthresh is more strict window_clamp used at "slow start"
352 * phase to predict further behaviour of this connection.
353 * It is used for two goals:
354 * - to enforce header prediction at sender, even when application
355 * requires some significant "application buffer". It is check #1.
356 * - to prevent pruning of receive queue because of misprediction
357 * of receiver window. Check #2.
359 * The scheme does not work when sender sends good segments opening
360 * window and then starts to feed us spaghetti. But it should work
361 * in common situations. Otherwise, we have to rely on queue collapsing.
364 /* Slow part of check#2. */
365 static int __tcp_grow_window(const struct sock
*sk
, const struct sk_buff
*skb
)
367 struct tcp_sock
*tp
= tcp_sk(sk
);
369 int truesize
= tcp_win_from_space(skb
->truesize
) >> 1;
370 int window
= tcp_win_from_space(sysctl_tcp_rmem
[2]) >> 1;
372 while (tp
->rcv_ssthresh
<= window
) {
373 if (truesize
<= skb
->len
)
374 return 2 * inet_csk(sk
)->icsk_ack
.rcv_mss
;
382 static void tcp_grow_window(struct sock
*sk
, const struct sk_buff
*skb
)
384 struct tcp_sock
*tp
= tcp_sk(sk
);
387 if (tp
->rcv_ssthresh
< tp
->window_clamp
&&
388 (int)tp
->rcv_ssthresh
< tcp_space(sk
) &&
389 !tcp_under_memory_pressure(sk
)) {
392 /* Check #2. Increase window, if skb with such overhead
393 * will fit to rcvbuf in future.
395 if (tcp_win_from_space(skb
->truesize
) <= skb
->len
)
396 incr
= 2 * tp
->advmss
;
398 incr
= __tcp_grow_window(sk
, skb
);
401 incr
= max_t(int, incr
, 2 * skb
->len
);
402 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
+ incr
,
404 inet_csk(sk
)->icsk_ack
.quick
|= 1;
409 /* 3. Tuning rcvbuf, when connection enters established state. */
410 static void tcp_fixup_rcvbuf(struct sock
*sk
)
412 u32 mss
= tcp_sk(sk
)->advmss
;
415 rcvmem
= 2 * SKB_TRUESIZE(mss
+ MAX_TCP_HEADER
) *
416 tcp_default_init_rwnd(mss
);
418 /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
419 * Allow enough cushion so that sender is not limited by our window
421 if (sysctl_tcp_moderate_rcvbuf
)
424 if (sk
->sk_rcvbuf
< rcvmem
)
425 sk
->sk_rcvbuf
= min(rcvmem
, sysctl_tcp_rmem
[2]);
428 /* 4. Try to fixup all. It is made immediately after connection enters
431 void tcp_init_buffer_space(struct sock
*sk
)
433 struct tcp_sock
*tp
= tcp_sk(sk
);
436 if (!(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
))
437 tcp_fixup_rcvbuf(sk
);
438 if (!(sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
))
439 tcp_sndbuf_expand(sk
);
441 tp
->rcvq_space
.space
= tp
->rcv_wnd
;
442 tcp_mstamp_refresh(tp
);
443 tp
->rcvq_space
.time
= tp
->tcp_mstamp
;
444 tp
->rcvq_space
.seq
= tp
->copied_seq
;
446 maxwin
= tcp_full_space(sk
);
448 if (tp
->window_clamp
>= maxwin
) {
449 tp
->window_clamp
= maxwin
;
451 if (sysctl_tcp_app_win
&& maxwin
> 4 * tp
->advmss
)
452 tp
->window_clamp
= max(maxwin
-
453 (maxwin
>> sysctl_tcp_app_win
),
457 /* Force reservation of one segment. */
458 if (sysctl_tcp_app_win
&&
459 tp
->window_clamp
> 2 * tp
->advmss
&&
460 tp
->window_clamp
+ tp
->advmss
> maxwin
)
461 tp
->window_clamp
= max(2 * tp
->advmss
, maxwin
- tp
->advmss
);
463 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, tp
->window_clamp
);
464 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
467 /* 5. Recalculate window clamp after socket hit its memory bounds. */
468 static void tcp_clamp_window(struct sock
*sk
)
470 struct tcp_sock
*tp
= tcp_sk(sk
);
471 struct inet_connection_sock
*icsk
= inet_csk(sk
);
473 icsk
->icsk_ack
.quick
= 0;
475 if (sk
->sk_rcvbuf
< sysctl_tcp_rmem
[2] &&
476 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
) &&
477 !tcp_under_memory_pressure(sk
) &&
478 sk_memory_allocated(sk
) < sk_prot_mem_limits(sk
, 0)) {
479 sk
->sk_rcvbuf
= min(atomic_read(&sk
->sk_rmem_alloc
),
482 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
)
483 tp
->rcv_ssthresh
= min(tp
->window_clamp
, 2U * tp
->advmss
);
486 /* Initialize RCV_MSS value.
487 * RCV_MSS is an our guess about MSS used by the peer.
488 * We haven't any direct information about the MSS.
489 * It's better to underestimate the RCV_MSS rather than overestimate.
490 * Overestimations make us ACKing less frequently than needed.
491 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
493 void tcp_initialize_rcv_mss(struct sock
*sk
)
495 const struct tcp_sock
*tp
= tcp_sk(sk
);
496 unsigned int hint
= min_t(unsigned int, tp
->advmss
, tp
->mss_cache
);
498 hint
= min(hint
, tp
->rcv_wnd
/ 2);
499 hint
= min(hint
, TCP_MSS_DEFAULT
);
500 hint
= max(hint
, TCP_MIN_MSS
);
502 inet_csk(sk
)->icsk_ack
.rcv_mss
= hint
;
504 EXPORT_SYMBOL(tcp_initialize_rcv_mss
);
506 /* Receiver "autotuning" code.
508 * The algorithm for RTT estimation w/o timestamps is based on
509 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
510 * <http://public.lanl.gov/radiant/pubs.html#DRS>
512 * More detail on this code can be found at
513 * <http://staff.psc.edu/jheffner/>,
514 * though this reference is out of date. A new paper
517 static void tcp_rcv_rtt_update(struct tcp_sock
*tp
, u32 sample
, int win_dep
)
519 u32 new_sample
= tp
->rcv_rtt_est
.rtt_us
;
525 if (new_sample
!= 0) {
526 /* If we sample in larger samples in the non-timestamp
527 * case, we could grossly overestimate the RTT especially
528 * with chatty applications or bulk transfer apps which
529 * are stalled on filesystem I/O.
531 * Also, since we are only going for a minimum in the
532 * non-timestamp case, we do not smooth things out
533 * else with timestamps disabled convergence takes too
537 m
-= (new_sample
>> 3);
545 /* No previous measure. */
549 tp
->rcv_rtt_est
.rtt_us
= new_sample
;
552 static inline void tcp_rcv_rtt_measure(struct tcp_sock
*tp
)
556 if (tp
->rcv_rtt_est
.time
== 0)
558 if (before(tp
->rcv_nxt
, tp
->rcv_rtt_est
.seq
))
560 delta_us
= tcp_stamp_us_delta(tp
->tcp_mstamp
, tp
->rcv_rtt_est
.time
);
561 tcp_rcv_rtt_update(tp
, delta_us
, 1);
564 tp
->rcv_rtt_est
.seq
= tp
->rcv_nxt
+ tp
->rcv_wnd
;
565 tp
->rcv_rtt_est
.time
= tp
->tcp_mstamp
;
568 static inline void tcp_rcv_rtt_measure_ts(struct sock
*sk
,
569 const struct sk_buff
*skb
)
571 struct tcp_sock
*tp
= tcp_sk(sk
);
573 if (tp
->rx_opt
.rcv_tsecr
&&
574 (TCP_SKB_CB(skb
)->end_seq
-
575 TCP_SKB_CB(skb
)->seq
>= inet_csk(sk
)->icsk_ack
.rcv_mss
)) {
576 u32 delta
= tcp_time_stamp(tp
) - tp
->rx_opt
.rcv_tsecr
;
577 u32 delta_us
= delta
* (USEC_PER_SEC
/ TCP_TS_HZ
);
579 tcp_rcv_rtt_update(tp
, delta_us
, 0);
584 * This function should be called every time data is copied to user space.
585 * It calculates the appropriate TCP receive buffer space.
587 void tcp_rcv_space_adjust(struct sock
*sk
)
589 struct tcp_sock
*tp
= tcp_sk(sk
);
593 time
= tcp_stamp_us_delta(tp
->tcp_mstamp
, tp
->rcvq_space
.time
);
594 if (time
< (tp
->rcv_rtt_est
.rtt_us
>> 3) || tp
->rcv_rtt_est
.rtt_us
== 0)
597 /* Number of bytes copied to user in last RTT */
598 copied
= tp
->copied_seq
- tp
->rcvq_space
.seq
;
599 if (copied
<= tp
->rcvq_space
.space
)
603 * copied = bytes received in previous RTT, our base window
604 * To cope with packet losses, we need a 2x factor
605 * To cope with slow start, and sender growing its cwin by 100 %
606 * every RTT, we need a 4x factor, because the ACK we are sending
607 * now is for the next RTT, not the current one :
608 * <prev RTT . ><current RTT .. ><next RTT .... >
611 if (sysctl_tcp_moderate_rcvbuf
&&
612 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
)) {
613 int rcvwin
, rcvmem
, rcvbuf
;
615 /* minimal window to cope with packet losses, assuming
616 * steady state. Add some cushion because of small variations.
618 rcvwin
= (copied
<< 1) + 16 * tp
->advmss
;
620 /* If rate increased by 25%,
621 * assume slow start, rcvwin = 3 * copied
622 * If rate increased by 50%,
623 * assume sender can use 2x growth, rcvwin = 4 * copied
626 tp
->rcvq_space
.space
+ (tp
->rcvq_space
.space
>> 2)) {
628 tp
->rcvq_space
.space
+ (tp
->rcvq_space
.space
>> 1))
631 rcvwin
+= (rcvwin
>> 1);
634 rcvmem
= SKB_TRUESIZE(tp
->advmss
+ MAX_TCP_HEADER
);
635 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
638 rcvbuf
= min(rcvwin
/ tp
->advmss
* rcvmem
, sysctl_tcp_rmem
[2]);
639 if (rcvbuf
> sk
->sk_rcvbuf
) {
640 sk
->sk_rcvbuf
= rcvbuf
;
642 /* Make the window clamp follow along. */
643 tp
->window_clamp
= rcvwin
;
646 tp
->rcvq_space
.space
= copied
;
649 tp
->rcvq_space
.seq
= tp
->copied_seq
;
650 tp
->rcvq_space
.time
= tp
->tcp_mstamp
;
653 /* There is something which you must keep in mind when you analyze the
654 * behavior of the tp->ato delayed ack timeout interval. When a
655 * connection starts up, we want to ack as quickly as possible. The
656 * problem is that "good" TCP's do slow start at the beginning of data
657 * transmission. The means that until we send the first few ACK's the
658 * sender will sit on his end and only queue most of his data, because
659 * he can only send snd_cwnd unacked packets at any given time. For
660 * each ACK we send, he increments snd_cwnd and transmits more of his
663 static void tcp_event_data_recv(struct sock
*sk
, struct sk_buff
*skb
)
665 struct tcp_sock
*tp
= tcp_sk(sk
);
666 struct inet_connection_sock
*icsk
= inet_csk(sk
);
669 inet_csk_schedule_ack(sk
);
671 tcp_measure_rcv_mss(sk
, skb
);
673 tcp_rcv_rtt_measure(tp
);
677 if (!icsk
->icsk_ack
.ato
) {
678 /* The _first_ data packet received, initialize
679 * delayed ACK engine.
681 tcp_incr_quickack(sk
);
682 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
684 int m
= now
- icsk
->icsk_ack
.lrcvtime
;
686 if (m
<= TCP_ATO_MIN
/ 2) {
687 /* The fastest case is the first. */
688 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + TCP_ATO_MIN
/ 2;
689 } else if (m
< icsk
->icsk_ack
.ato
) {
690 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + m
;
691 if (icsk
->icsk_ack
.ato
> icsk
->icsk_rto
)
692 icsk
->icsk_ack
.ato
= icsk
->icsk_rto
;
693 } else if (m
> icsk
->icsk_rto
) {
694 /* Too long gap. Apparently sender failed to
695 * restart window, so that we send ACKs quickly.
697 tcp_incr_quickack(sk
);
701 icsk
->icsk_ack
.lrcvtime
= now
;
703 tcp_ecn_check_ce(tp
, skb
);
706 tcp_grow_window(sk
, skb
);
709 /* Called to compute a smoothed rtt estimate. The data fed to this
710 * routine either comes from timestamps, or from segments that were
711 * known _not_ to have been retransmitted [see Karn/Partridge
712 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
713 * piece by Van Jacobson.
714 * NOTE: the next three routines used to be one big routine.
715 * To save cycles in the RFC 1323 implementation it was better to break
716 * it up into three procedures. -- erics
718 static void tcp_rtt_estimator(struct sock
*sk
, long mrtt_us
)
720 struct tcp_sock
*tp
= tcp_sk(sk
);
721 long m
= mrtt_us
; /* RTT */
722 u32 srtt
= tp
->srtt_us
;
724 /* The following amusing code comes from Jacobson's
725 * article in SIGCOMM '88. Note that rtt and mdev
726 * are scaled versions of rtt and mean deviation.
727 * This is designed to be as fast as possible
728 * m stands for "measurement".
730 * On a 1990 paper the rto value is changed to:
731 * RTO = rtt + 4 * mdev
733 * Funny. This algorithm seems to be very broken.
734 * These formulae increase RTO, when it should be decreased, increase
735 * too slowly, when it should be increased quickly, decrease too quickly
736 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
737 * does not matter how to _calculate_ it. Seems, it was trap
738 * that VJ failed to avoid. 8)
741 m
-= (srtt
>> 3); /* m is now error in rtt est */
742 srtt
+= m
; /* rtt = 7/8 rtt + 1/8 new */
744 m
= -m
; /* m is now abs(error) */
745 m
-= (tp
->mdev_us
>> 2); /* similar update on mdev */
746 /* This is similar to one of Eifel findings.
747 * Eifel blocks mdev updates when rtt decreases.
748 * This solution is a bit different: we use finer gain
749 * for mdev in this case (alpha*beta).
750 * Like Eifel it also prevents growth of rto,
751 * but also it limits too fast rto decreases,
752 * happening in pure Eifel.
757 m
-= (tp
->mdev_us
>> 2); /* similar update on mdev */
759 tp
->mdev_us
+= m
; /* mdev = 3/4 mdev + 1/4 new */
760 if (tp
->mdev_us
> tp
->mdev_max_us
) {
761 tp
->mdev_max_us
= tp
->mdev_us
;
762 if (tp
->mdev_max_us
> tp
->rttvar_us
)
763 tp
->rttvar_us
= tp
->mdev_max_us
;
765 if (after(tp
->snd_una
, tp
->rtt_seq
)) {
766 if (tp
->mdev_max_us
< tp
->rttvar_us
)
767 tp
->rttvar_us
-= (tp
->rttvar_us
- tp
->mdev_max_us
) >> 2;
768 tp
->rtt_seq
= tp
->snd_nxt
;
769 tp
->mdev_max_us
= tcp_rto_min_us(sk
);
772 /* no previous measure. */
773 srtt
= m
<< 3; /* take the measured time to be rtt */
774 tp
->mdev_us
= m
<< 1; /* make sure rto = 3*rtt */
775 tp
->rttvar_us
= max(tp
->mdev_us
, tcp_rto_min_us(sk
));
776 tp
->mdev_max_us
= tp
->rttvar_us
;
777 tp
->rtt_seq
= tp
->snd_nxt
;
779 tp
->srtt_us
= max(1U, srtt
);
782 /* Set the sk_pacing_rate to allow proper sizing of TSO packets.
783 * Note: TCP stack does not yet implement pacing.
784 * FQ packet scheduler can be used to implement cheap but effective
785 * TCP pacing, to smooth the burst on large writes when packets
786 * in flight is significantly lower than cwnd (or rwin)
788 int sysctl_tcp_pacing_ss_ratio __read_mostly
= 200;
789 int sysctl_tcp_pacing_ca_ratio __read_mostly
= 120;
791 static void tcp_update_pacing_rate(struct sock
*sk
)
793 const struct tcp_sock
*tp
= tcp_sk(sk
);
796 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
797 rate
= (u64
)tp
->mss_cache
* ((USEC_PER_SEC
/ 100) << 3);
799 /* current rate is (cwnd * mss) / srtt
800 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
801 * In Congestion Avoidance phase, set it to 120 % the current rate.
803 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
804 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
805 * end of slow start and should slow down.
807 if (tp
->snd_cwnd
< tp
->snd_ssthresh
/ 2)
808 rate
*= sysctl_tcp_pacing_ss_ratio
;
810 rate
*= sysctl_tcp_pacing_ca_ratio
;
812 rate
*= max(tp
->snd_cwnd
, tp
->packets_out
);
814 if (likely(tp
->srtt_us
))
815 do_div(rate
, tp
->srtt_us
);
817 /* ACCESS_ONCE() is needed because sch_fq fetches sk_pacing_rate
818 * without any lock. We want to make sure compiler wont store
819 * intermediate values in this location.
821 ACCESS_ONCE(sk
->sk_pacing_rate
) = min_t(u64
, rate
,
822 sk
->sk_max_pacing_rate
);
825 /* Calculate rto without backoff. This is the second half of Van Jacobson's
826 * routine referred to above.
828 static void tcp_set_rto(struct sock
*sk
)
830 const struct tcp_sock
*tp
= tcp_sk(sk
);
831 /* Old crap is replaced with new one. 8)
834 * 1. If rtt variance happened to be less 50msec, it is hallucination.
835 * It cannot be less due to utterly erratic ACK generation made
836 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
837 * to do with delayed acks, because at cwnd>2 true delack timeout
838 * is invisible. Actually, Linux-2.4 also generates erratic
839 * ACKs in some circumstances.
841 inet_csk(sk
)->icsk_rto
= __tcp_set_rto(tp
);
843 /* 2. Fixups made earlier cannot be right.
844 * If we do not estimate RTO correctly without them,
845 * all the algo is pure shit and should be replaced
846 * with correct one. It is exactly, which we pretend to do.
849 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
850 * guarantees that rto is higher.
855 __u32
tcp_init_cwnd(const struct tcp_sock
*tp
, const struct dst_entry
*dst
)
857 __u32 cwnd
= (dst
? dst_metric(dst
, RTAX_INITCWND
) : 0);
860 cwnd
= TCP_INIT_CWND
;
861 return min_t(__u32
, cwnd
, tp
->snd_cwnd_clamp
);
865 * Packet counting of FACK is based on in-order assumptions, therefore TCP
866 * disables it when reordering is detected
868 void tcp_disable_fack(struct tcp_sock
*tp
)
870 /* RFC3517 uses different metric in lost marker => reset on change */
872 tp
->lost_skb_hint
= NULL
;
873 tp
->rx_opt
.sack_ok
&= ~TCP_FACK_ENABLED
;
876 /* Take a notice that peer is sending D-SACKs */
877 static void tcp_dsack_seen(struct tcp_sock
*tp
)
879 tp
->rx_opt
.sack_ok
|= TCP_DSACK_SEEN
;
882 static void tcp_update_reordering(struct sock
*sk
, const int metric
,
885 struct tcp_sock
*tp
= tcp_sk(sk
);
888 if (WARN_ON_ONCE(metric
< 0))
891 if (metric
> tp
->reordering
) {
892 tp
->reordering
= min(sysctl_tcp_max_reordering
, metric
);
894 #if FASTRETRANS_DEBUG > 1
895 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
896 tp
->rx_opt
.sack_ok
, inet_csk(sk
)->icsk_ca_state
,
900 tp
->undo_marker
? tp
->undo_retrans
: 0);
902 tcp_disable_fack(tp
);
907 /* This exciting event is worth to be remembered. 8) */
909 mib_idx
= LINUX_MIB_TCPTSREORDER
;
910 else if (tcp_is_reno(tp
))
911 mib_idx
= LINUX_MIB_TCPRENOREORDER
;
912 else if (tcp_is_fack(tp
))
913 mib_idx
= LINUX_MIB_TCPFACKREORDER
;
915 mib_idx
= LINUX_MIB_TCPSACKREORDER
;
917 NET_INC_STATS(sock_net(sk
), mib_idx
);
920 /* This must be called before lost_out is incremented */
921 static void tcp_verify_retransmit_hint(struct tcp_sock
*tp
, struct sk_buff
*skb
)
923 if (!tp
->retransmit_skb_hint
||
924 before(TCP_SKB_CB(skb
)->seq
,
925 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
926 tp
->retransmit_skb_hint
= skb
;
929 /* Sum the number of packets on the wire we have marked as lost.
930 * There are two cases we care about here:
931 * a) Packet hasn't been marked lost (nor retransmitted),
932 * and this is the first loss.
933 * b) Packet has been marked both lost and retransmitted,
934 * and this means we think it was lost again.
936 static void tcp_sum_lost(struct tcp_sock
*tp
, struct sk_buff
*skb
)
938 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
940 if (!(sacked
& TCPCB_LOST
) ||
941 ((sacked
& TCPCB_LOST
) && (sacked
& TCPCB_SACKED_RETRANS
)))
942 tp
->lost
+= tcp_skb_pcount(skb
);
945 static void tcp_skb_mark_lost(struct tcp_sock
*tp
, struct sk_buff
*skb
)
947 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
948 tcp_verify_retransmit_hint(tp
, skb
);
950 tp
->lost_out
+= tcp_skb_pcount(skb
);
951 tcp_sum_lost(tp
, skb
);
952 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
956 void tcp_skb_mark_lost_uncond_verify(struct tcp_sock
*tp
, struct sk_buff
*skb
)
958 tcp_verify_retransmit_hint(tp
, skb
);
960 tcp_sum_lost(tp
, skb
);
961 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
962 tp
->lost_out
+= tcp_skb_pcount(skb
);
963 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
967 /* This procedure tags the retransmission queue when SACKs arrive.
969 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
970 * Packets in queue with these bits set are counted in variables
971 * sacked_out, retrans_out and lost_out, correspondingly.
973 * Valid combinations are:
974 * Tag InFlight Description
975 * 0 1 - orig segment is in flight.
976 * S 0 - nothing flies, orig reached receiver.
977 * L 0 - nothing flies, orig lost by net.
978 * R 2 - both orig and retransmit are in flight.
979 * L|R 1 - orig is lost, retransmit is in flight.
980 * S|R 1 - orig reached receiver, retrans is still in flight.
981 * (L|S|R is logically valid, it could occur when L|R is sacked,
982 * but it is equivalent to plain S and code short-curcuits it to S.
983 * L|S is logically invalid, it would mean -1 packet in flight 8))
985 * These 6 states form finite state machine, controlled by the following events:
986 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
987 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
988 * 3. Loss detection event of two flavors:
989 * A. Scoreboard estimator decided the packet is lost.
990 * A'. Reno "three dupacks" marks head of queue lost.
991 * A''. Its FACK modification, head until snd.fack is lost.
992 * B. SACK arrives sacking SND.NXT at the moment, when the
993 * segment was retransmitted.
994 * 4. D-SACK added new rule: D-SACK changes any tag to S.
996 * It is pleasant to note, that state diagram turns out to be commutative,
997 * so that we are allowed not to be bothered by order of our actions,
998 * when multiple events arrive simultaneously. (see the function below).
1000 * Reordering detection.
1001 * --------------------
1002 * Reordering metric is maximal distance, which a packet can be displaced
1003 * in packet stream. With SACKs we can estimate it:
1005 * 1. SACK fills old hole and the corresponding segment was not
1006 * ever retransmitted -> reordering. Alas, we cannot use it
1007 * when segment was retransmitted.
1008 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1009 * for retransmitted and already SACKed segment -> reordering..
1010 * Both of these heuristics are not used in Loss state, when we cannot
1011 * account for retransmits accurately.
1013 * SACK block validation.
1014 * ----------------------
1016 * SACK block range validation checks that the received SACK block fits to
1017 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1018 * Note that SND.UNA is not included to the range though being valid because
1019 * it means that the receiver is rather inconsistent with itself reporting
1020 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1021 * perfectly valid, however, in light of RFC2018 which explicitly states
1022 * that "SACK block MUST reflect the newest segment. Even if the newest
1023 * segment is going to be discarded ...", not that it looks very clever
1024 * in case of head skb. Due to potentional receiver driven attacks, we
1025 * choose to avoid immediate execution of a walk in write queue due to
1026 * reneging and defer head skb's loss recovery to standard loss recovery
1027 * procedure that will eventually trigger (nothing forbids us doing this).
1029 * Implements also blockage to start_seq wrap-around. Problem lies in the
1030 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1031 * there's no guarantee that it will be before snd_nxt (n). The problem
1032 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1035 * <- outs wnd -> <- wrapzone ->
1036 * u e n u_w e_w s n_w
1038 * |<------------+------+----- TCP seqno space --------------+---------->|
1039 * ...-- <2^31 ->| |<--------...
1040 * ...---- >2^31 ------>| |<--------...
1042 * Current code wouldn't be vulnerable but it's better still to discard such
1043 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1044 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1045 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1046 * equal to the ideal case (infinite seqno space without wrap caused issues).
1048 * With D-SACK the lower bound is extended to cover sequence space below
1049 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1050 * again, D-SACK block must not to go across snd_una (for the same reason as
1051 * for the normal SACK blocks, explained above). But there all simplicity
1052 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1053 * fully below undo_marker they do not affect behavior in anyway and can
1054 * therefore be safely ignored. In rare cases (which are more or less
1055 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1056 * fragmentation and packet reordering past skb's retransmission. To consider
1057 * them correctly, the acceptable range must be extended even more though
1058 * the exact amount is rather hard to quantify. However, tp->max_window can
1059 * be used as an exaggerated estimate.
1061 static bool tcp_is_sackblock_valid(struct tcp_sock
*tp
, bool is_dsack
,
1062 u32 start_seq
, u32 end_seq
)
1064 /* Too far in future, or reversed (interpretation is ambiguous) */
1065 if (after(end_seq
, tp
->snd_nxt
) || !before(start_seq
, end_seq
))
1068 /* Nasty start_seq wrap-around check (see comments above) */
1069 if (!before(start_seq
, tp
->snd_nxt
))
1072 /* In outstanding window? ...This is valid exit for D-SACKs too.
1073 * start_seq == snd_una is non-sensical (see comments above)
1075 if (after(start_seq
, tp
->snd_una
))
1078 if (!is_dsack
|| !tp
->undo_marker
)
1081 /* ...Then it's D-SACK, and must reside below snd_una completely */
1082 if (after(end_seq
, tp
->snd_una
))
1085 if (!before(start_seq
, tp
->undo_marker
))
1089 if (!after(end_seq
, tp
->undo_marker
))
1092 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1093 * start_seq < undo_marker and end_seq >= undo_marker.
1095 return !before(start_seq
, end_seq
- tp
->max_window
);
1098 static bool tcp_check_dsack(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1099 struct tcp_sack_block_wire
*sp
, int num_sacks
,
1102 struct tcp_sock
*tp
= tcp_sk(sk
);
1103 u32 start_seq_0
= get_unaligned_be32(&sp
[0].start_seq
);
1104 u32 end_seq_0
= get_unaligned_be32(&sp
[0].end_seq
);
1105 bool dup_sack
= false;
1107 if (before(start_seq_0
, TCP_SKB_CB(ack_skb
)->ack_seq
)) {
1110 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPDSACKRECV
);
1111 } else if (num_sacks
> 1) {
1112 u32 end_seq_1
= get_unaligned_be32(&sp
[1].end_seq
);
1113 u32 start_seq_1
= get_unaligned_be32(&sp
[1].start_seq
);
1115 if (!after(end_seq_0
, end_seq_1
) &&
1116 !before(start_seq_0
, start_seq_1
)) {
1119 NET_INC_STATS(sock_net(sk
),
1120 LINUX_MIB_TCPDSACKOFORECV
);
1124 /* D-SACK for already forgotten data... Do dumb counting. */
1125 if (dup_sack
&& tp
->undo_marker
&& tp
->undo_retrans
> 0 &&
1126 !after(end_seq_0
, prior_snd_una
) &&
1127 after(end_seq_0
, tp
->undo_marker
))
1133 struct tcp_sacktag_state
{
1136 /* Timestamps for earliest and latest never-retransmitted segment
1137 * that was SACKed. RTO needs the earliest RTT to stay conservative,
1138 * but congestion control should still get an accurate delay signal.
1142 struct rate_sample
*rate
;
1146 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1147 * the incoming SACK may not exactly match but we can find smaller MSS
1148 * aligned portion of it that matches. Therefore we might need to fragment
1149 * which may fail and creates some hassle (caller must handle error case
1152 * FIXME: this could be merged to shift decision code
1154 static int tcp_match_skb_to_sack(struct sock
*sk
, struct sk_buff
*skb
,
1155 u32 start_seq
, u32 end_seq
)
1159 unsigned int pkt_len
;
1162 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1163 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1165 if (tcp_skb_pcount(skb
) > 1 && !in_sack
&&
1166 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
1167 mss
= tcp_skb_mss(skb
);
1168 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1171 pkt_len
= start_seq
- TCP_SKB_CB(skb
)->seq
;
1175 pkt_len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1180 /* Round if necessary so that SACKs cover only full MSSes
1181 * and/or the remaining small portion (if present)
1183 if (pkt_len
> mss
) {
1184 unsigned int new_len
= (pkt_len
/ mss
) * mss
;
1185 if (!in_sack
&& new_len
< pkt_len
)
1190 if (pkt_len
>= skb
->len
&& !in_sack
)
1193 err
= tcp_fragment(sk
, skb
, pkt_len
, mss
, GFP_ATOMIC
);
1201 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1202 static u8
tcp_sacktag_one(struct sock
*sk
,
1203 struct tcp_sacktag_state
*state
, u8 sacked
,
1204 u32 start_seq
, u32 end_seq
,
1205 int dup_sack
, int pcount
,
1208 struct tcp_sock
*tp
= tcp_sk(sk
);
1209 int fack_count
= state
->fack_count
;
1211 /* Account D-SACK for retransmitted packet. */
1212 if (dup_sack
&& (sacked
& TCPCB_RETRANS
)) {
1213 if (tp
->undo_marker
&& tp
->undo_retrans
> 0 &&
1214 after(end_seq
, tp
->undo_marker
))
1216 if (sacked
& TCPCB_SACKED_ACKED
)
1217 state
->reord
= min(fack_count
, state
->reord
);
1220 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1221 if (!after(end_seq
, tp
->snd_una
))
1224 if (!(sacked
& TCPCB_SACKED_ACKED
)) {
1225 tcp_rack_advance(tp
, sacked
, end_seq
, xmit_time
);
1227 if (sacked
& TCPCB_SACKED_RETRANS
) {
1228 /* If the segment is not tagged as lost,
1229 * we do not clear RETRANS, believing
1230 * that retransmission is still in flight.
1232 if (sacked
& TCPCB_LOST
) {
1233 sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1234 tp
->lost_out
-= pcount
;
1235 tp
->retrans_out
-= pcount
;
1238 if (!(sacked
& TCPCB_RETRANS
)) {
1239 /* New sack for not retransmitted frame,
1240 * which was in hole. It is reordering.
1242 if (before(start_seq
,
1243 tcp_highest_sack_seq(tp
)))
1244 state
->reord
= min(fack_count
,
1246 if (!after(end_seq
, tp
->high_seq
))
1247 state
->flag
|= FLAG_ORIG_SACK_ACKED
;
1248 if (state
->first_sackt
== 0)
1249 state
->first_sackt
= xmit_time
;
1250 state
->last_sackt
= xmit_time
;
1253 if (sacked
& TCPCB_LOST
) {
1254 sacked
&= ~TCPCB_LOST
;
1255 tp
->lost_out
-= pcount
;
1259 sacked
|= TCPCB_SACKED_ACKED
;
1260 state
->flag
|= FLAG_DATA_SACKED
;
1261 tp
->sacked_out
+= pcount
;
1262 tp
->delivered
+= pcount
; /* Out-of-order packets delivered */
1264 fack_count
+= pcount
;
1266 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1267 if (!tcp_is_fack(tp
) && tp
->lost_skb_hint
&&
1268 before(start_seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
))
1269 tp
->lost_cnt_hint
+= pcount
;
1271 if (fack_count
> tp
->fackets_out
)
1272 tp
->fackets_out
= fack_count
;
1275 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1276 * frames and clear it. undo_retrans is decreased above, L|R frames
1277 * are accounted above as well.
1279 if (dup_sack
&& (sacked
& TCPCB_SACKED_RETRANS
)) {
1280 sacked
&= ~TCPCB_SACKED_RETRANS
;
1281 tp
->retrans_out
-= pcount
;
1287 /* Shift newly-SACKed bytes from this skb to the immediately previous
1288 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1290 static bool tcp_shifted_skb(struct sock
*sk
, struct sk_buff
*skb
,
1291 struct tcp_sacktag_state
*state
,
1292 unsigned int pcount
, int shifted
, int mss
,
1295 struct tcp_sock
*tp
= tcp_sk(sk
);
1296 struct sk_buff
*prev
= tcp_write_queue_prev(sk
, skb
);
1297 u32 start_seq
= TCP_SKB_CB(skb
)->seq
; /* start of newly-SACKed */
1298 u32 end_seq
= start_seq
+ shifted
; /* end of newly-SACKed */
1302 /* Adjust counters and hints for the newly sacked sequence
1303 * range but discard the return value since prev is already
1304 * marked. We must tag the range first because the seq
1305 * advancement below implicitly advances
1306 * tcp_highest_sack_seq() when skb is highest_sack.
1308 tcp_sacktag_one(sk
, state
, TCP_SKB_CB(skb
)->sacked
,
1309 start_seq
, end_seq
, dup_sack
, pcount
,
1311 tcp_rate_skb_delivered(sk
, skb
, state
->rate
);
1313 if (skb
== tp
->lost_skb_hint
)
1314 tp
->lost_cnt_hint
+= pcount
;
1316 TCP_SKB_CB(prev
)->end_seq
+= shifted
;
1317 TCP_SKB_CB(skb
)->seq
+= shifted
;
1319 tcp_skb_pcount_add(prev
, pcount
);
1320 BUG_ON(tcp_skb_pcount(skb
) < pcount
);
1321 tcp_skb_pcount_add(skb
, -pcount
);
1323 /* When we're adding to gso_segs == 1, gso_size will be zero,
1324 * in theory this shouldn't be necessary but as long as DSACK
1325 * code can come after this skb later on it's better to keep
1326 * setting gso_size to something.
1328 if (!TCP_SKB_CB(prev
)->tcp_gso_size
)
1329 TCP_SKB_CB(prev
)->tcp_gso_size
= mss
;
1331 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1332 if (tcp_skb_pcount(skb
) <= 1)
1333 TCP_SKB_CB(skb
)->tcp_gso_size
= 0;
1335 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1336 TCP_SKB_CB(prev
)->sacked
|= (TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
);
1339 BUG_ON(!tcp_skb_pcount(skb
));
1340 NET_INC_STATS(sock_net(sk
), LINUX_MIB_SACKSHIFTED
);
1344 /* Whole SKB was eaten :-) */
1346 if (skb
== tp
->retransmit_skb_hint
)
1347 tp
->retransmit_skb_hint
= prev
;
1348 if (skb
== tp
->lost_skb_hint
) {
1349 tp
->lost_skb_hint
= prev
;
1350 tp
->lost_cnt_hint
-= tcp_skb_pcount(prev
);
1353 TCP_SKB_CB(prev
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
;
1354 TCP_SKB_CB(prev
)->eor
= TCP_SKB_CB(skb
)->eor
;
1355 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
1356 TCP_SKB_CB(prev
)->end_seq
++;
1358 if (skb
== tcp_highest_sack(sk
))
1359 tcp_advance_highest_sack(sk
, skb
);
1361 tcp_skb_collapse_tstamp(prev
, skb
);
1362 if (unlikely(TCP_SKB_CB(prev
)->tx
.delivered_mstamp
))
1363 TCP_SKB_CB(prev
)->tx
.delivered_mstamp
= 0;
1365 tcp_unlink_write_queue(skb
, sk
);
1366 sk_wmem_free_skb(sk
, skb
);
1368 NET_INC_STATS(sock_net(sk
), LINUX_MIB_SACKMERGED
);
1373 /* I wish gso_size would have a bit more sane initialization than
1374 * something-or-zero which complicates things
1376 static int tcp_skb_seglen(const struct sk_buff
*skb
)
1378 return tcp_skb_pcount(skb
) == 1 ? skb
->len
: tcp_skb_mss(skb
);
1381 /* Shifting pages past head area doesn't work */
1382 static int skb_can_shift(const struct sk_buff
*skb
)
1384 return !skb_headlen(skb
) && skb_is_nonlinear(skb
);
1387 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1390 static struct sk_buff
*tcp_shift_skb_data(struct sock
*sk
, struct sk_buff
*skb
,
1391 struct tcp_sacktag_state
*state
,
1392 u32 start_seq
, u32 end_seq
,
1395 struct tcp_sock
*tp
= tcp_sk(sk
);
1396 struct sk_buff
*prev
;
1402 if (!sk_can_gso(sk
))
1405 /* Normally R but no L won't result in plain S */
1407 (TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_RETRANS
)) == TCPCB_SACKED_RETRANS
)
1409 if (!skb_can_shift(skb
))
1411 /* This frame is about to be dropped (was ACKed). */
1412 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1415 /* Can only happen with delayed DSACK + discard craziness */
1416 if (unlikely(skb
== tcp_write_queue_head(sk
)))
1418 prev
= tcp_write_queue_prev(sk
, skb
);
1420 if ((TCP_SKB_CB(prev
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
)
1423 if (!tcp_skb_can_collapse_to(prev
))
1426 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1427 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1431 pcount
= tcp_skb_pcount(skb
);
1432 mss
= tcp_skb_seglen(skb
);
1434 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1435 * drop this restriction as unnecessary
1437 if (mss
!= tcp_skb_seglen(prev
))
1440 if (!after(TCP_SKB_CB(skb
)->end_seq
, start_seq
))
1442 /* CHECKME: This is non-MSS split case only?, this will
1443 * cause skipped skbs due to advancing loop btw, original
1444 * has that feature too
1446 if (tcp_skb_pcount(skb
) <= 1)
1449 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1451 /* TODO: head merge to next could be attempted here
1452 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1453 * though it might not be worth of the additional hassle
1455 * ...we can probably just fallback to what was done
1456 * previously. We could try merging non-SACKed ones
1457 * as well but it probably isn't going to buy off
1458 * because later SACKs might again split them, and
1459 * it would make skb timestamp tracking considerably
1465 len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1467 BUG_ON(len
> skb
->len
);
1469 /* MSS boundaries should be honoured or else pcount will
1470 * severely break even though it makes things bit trickier.
1471 * Optimize common case to avoid most of the divides
1473 mss
= tcp_skb_mss(skb
);
1475 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1476 * drop this restriction as unnecessary
1478 if (mss
!= tcp_skb_seglen(prev
))
1483 } else if (len
< mss
) {
1491 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1492 if (!after(TCP_SKB_CB(skb
)->seq
+ len
, tp
->snd_una
))
1495 if (!skb_shift(prev
, skb
, len
))
1497 if (!tcp_shifted_skb(sk
, skb
, state
, pcount
, len
, mss
, dup_sack
))
1500 /* Hole filled allows collapsing with the next as well, this is very
1501 * useful when hole on every nth skb pattern happens
1503 if (prev
== tcp_write_queue_tail(sk
))
1505 skb
= tcp_write_queue_next(sk
, prev
);
1507 if (!skb_can_shift(skb
) ||
1508 (skb
== tcp_send_head(sk
)) ||
1509 ((TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
) ||
1510 (mss
!= tcp_skb_seglen(skb
)))
1514 if (skb_shift(prev
, skb
, len
)) {
1515 pcount
+= tcp_skb_pcount(skb
);
1516 tcp_shifted_skb(sk
, skb
, state
, tcp_skb_pcount(skb
), len
, mss
, 0);
1520 state
->fack_count
+= pcount
;
1527 NET_INC_STATS(sock_net(sk
), LINUX_MIB_SACKSHIFTFALLBACK
);
1531 static struct sk_buff
*tcp_sacktag_walk(struct sk_buff
*skb
, struct sock
*sk
,
1532 struct tcp_sack_block
*next_dup
,
1533 struct tcp_sacktag_state
*state
,
1534 u32 start_seq
, u32 end_seq
,
1537 struct tcp_sock
*tp
= tcp_sk(sk
);
1538 struct sk_buff
*tmp
;
1540 tcp_for_write_queue_from(skb
, sk
) {
1542 bool dup_sack
= dup_sack_in
;
1544 if (skb
== tcp_send_head(sk
))
1547 /* queue is in-order => we can short-circuit the walk early */
1548 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1552 before(TCP_SKB_CB(skb
)->seq
, next_dup
->end_seq
)) {
1553 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1554 next_dup
->start_seq
,
1560 /* skb reference here is a bit tricky to get right, since
1561 * shifting can eat and free both this skb and the next,
1562 * so not even _safe variant of the loop is enough.
1565 tmp
= tcp_shift_skb_data(sk
, skb
, state
,
1566 start_seq
, end_seq
, dup_sack
);
1575 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1581 if (unlikely(in_sack
< 0))
1585 TCP_SKB_CB(skb
)->sacked
=
1588 TCP_SKB_CB(skb
)->sacked
,
1589 TCP_SKB_CB(skb
)->seq
,
1590 TCP_SKB_CB(skb
)->end_seq
,
1592 tcp_skb_pcount(skb
),
1594 tcp_rate_skb_delivered(sk
, skb
, state
->rate
);
1596 if (!before(TCP_SKB_CB(skb
)->seq
,
1597 tcp_highest_sack_seq(tp
)))
1598 tcp_advance_highest_sack(sk
, skb
);
1601 state
->fack_count
+= tcp_skb_pcount(skb
);
1606 /* Avoid all extra work that is being done by sacktag while walking in
1609 static struct sk_buff
*tcp_sacktag_skip(struct sk_buff
*skb
, struct sock
*sk
,
1610 struct tcp_sacktag_state
*state
,
1613 tcp_for_write_queue_from(skb
, sk
) {
1614 if (skb
== tcp_send_head(sk
))
1617 if (after(TCP_SKB_CB(skb
)->end_seq
, skip_to_seq
))
1620 state
->fack_count
+= tcp_skb_pcount(skb
);
1625 static struct sk_buff
*tcp_maybe_skipping_dsack(struct sk_buff
*skb
,
1627 struct tcp_sack_block
*next_dup
,
1628 struct tcp_sacktag_state
*state
,
1634 if (before(next_dup
->start_seq
, skip_to_seq
)) {
1635 skb
= tcp_sacktag_skip(skb
, sk
, state
, next_dup
->start_seq
);
1636 skb
= tcp_sacktag_walk(skb
, sk
, NULL
, state
,
1637 next_dup
->start_seq
, next_dup
->end_seq
,
1644 static int tcp_sack_cache_ok(const struct tcp_sock
*tp
, const struct tcp_sack_block
*cache
)
1646 return cache
< tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1650 tcp_sacktag_write_queue(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1651 u32 prior_snd_una
, struct tcp_sacktag_state
*state
)
1653 struct tcp_sock
*tp
= tcp_sk(sk
);
1654 const unsigned char *ptr
= (skb_transport_header(ack_skb
) +
1655 TCP_SKB_CB(ack_skb
)->sacked
);
1656 struct tcp_sack_block_wire
*sp_wire
= (struct tcp_sack_block_wire
*)(ptr
+2);
1657 struct tcp_sack_block sp
[TCP_NUM_SACKS
];
1658 struct tcp_sack_block
*cache
;
1659 struct sk_buff
*skb
;
1660 int num_sacks
= min(TCP_NUM_SACKS
, (ptr
[1] - TCPOLEN_SACK_BASE
) >> 3);
1662 bool found_dup_sack
= false;
1664 int first_sack_index
;
1667 state
->reord
= tp
->packets_out
;
1669 if (!tp
->sacked_out
) {
1670 if (WARN_ON(tp
->fackets_out
))
1671 tp
->fackets_out
= 0;
1672 tcp_highest_sack_reset(sk
);
1675 found_dup_sack
= tcp_check_dsack(sk
, ack_skb
, sp_wire
,
1676 num_sacks
, prior_snd_una
);
1677 if (found_dup_sack
) {
1678 state
->flag
|= FLAG_DSACKING_ACK
;
1679 tp
->delivered
++; /* A spurious retransmission is delivered */
1682 /* Eliminate too old ACKs, but take into
1683 * account more or less fresh ones, they can
1684 * contain valid SACK info.
1686 if (before(TCP_SKB_CB(ack_skb
)->ack_seq
, prior_snd_una
- tp
->max_window
))
1689 if (!tp
->packets_out
)
1693 first_sack_index
= 0;
1694 for (i
= 0; i
< num_sacks
; i
++) {
1695 bool dup_sack
= !i
&& found_dup_sack
;
1697 sp
[used_sacks
].start_seq
= get_unaligned_be32(&sp_wire
[i
].start_seq
);
1698 sp
[used_sacks
].end_seq
= get_unaligned_be32(&sp_wire
[i
].end_seq
);
1700 if (!tcp_is_sackblock_valid(tp
, dup_sack
,
1701 sp
[used_sacks
].start_seq
,
1702 sp
[used_sacks
].end_seq
)) {
1706 if (!tp
->undo_marker
)
1707 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDNOUNDO
;
1709 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDOLD
;
1711 /* Don't count olds caused by ACK reordering */
1712 if ((TCP_SKB_CB(ack_skb
)->ack_seq
!= tp
->snd_una
) &&
1713 !after(sp
[used_sacks
].end_seq
, tp
->snd_una
))
1715 mib_idx
= LINUX_MIB_TCPSACKDISCARD
;
1718 NET_INC_STATS(sock_net(sk
), mib_idx
);
1720 first_sack_index
= -1;
1724 /* Ignore very old stuff early */
1725 if (!after(sp
[used_sacks
].end_seq
, prior_snd_una
))
1731 /* order SACK blocks to allow in order walk of the retrans queue */
1732 for (i
= used_sacks
- 1; i
> 0; i
--) {
1733 for (j
= 0; j
< i
; j
++) {
1734 if (after(sp
[j
].start_seq
, sp
[j
+ 1].start_seq
)) {
1735 swap(sp
[j
], sp
[j
+ 1]);
1737 /* Track where the first SACK block goes to */
1738 if (j
== first_sack_index
)
1739 first_sack_index
= j
+ 1;
1744 skb
= tcp_write_queue_head(sk
);
1745 state
->fack_count
= 0;
1748 if (!tp
->sacked_out
) {
1749 /* It's already past, so skip checking against it */
1750 cache
= tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1752 cache
= tp
->recv_sack_cache
;
1753 /* Skip empty blocks in at head of the cache */
1754 while (tcp_sack_cache_ok(tp
, cache
) && !cache
->start_seq
&&
1759 while (i
< used_sacks
) {
1760 u32 start_seq
= sp
[i
].start_seq
;
1761 u32 end_seq
= sp
[i
].end_seq
;
1762 bool dup_sack
= (found_dup_sack
&& (i
== first_sack_index
));
1763 struct tcp_sack_block
*next_dup
= NULL
;
1765 if (found_dup_sack
&& ((i
+ 1) == first_sack_index
))
1766 next_dup
= &sp
[i
+ 1];
1768 /* Skip too early cached blocks */
1769 while (tcp_sack_cache_ok(tp
, cache
) &&
1770 !before(start_seq
, cache
->end_seq
))
1773 /* Can skip some work by looking recv_sack_cache? */
1774 if (tcp_sack_cache_ok(tp
, cache
) && !dup_sack
&&
1775 after(end_seq
, cache
->start_seq
)) {
1778 if (before(start_seq
, cache
->start_seq
)) {
1779 skb
= tcp_sacktag_skip(skb
, sk
, state
,
1781 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
,
1788 /* Rest of the block already fully processed? */
1789 if (!after(end_seq
, cache
->end_seq
))
1792 skb
= tcp_maybe_skipping_dsack(skb
, sk
, next_dup
,
1796 /* ...tail remains todo... */
1797 if (tcp_highest_sack_seq(tp
) == cache
->end_seq
) {
1798 /* ...but better entrypoint exists! */
1799 skb
= tcp_highest_sack(sk
);
1802 state
->fack_count
= tp
->fackets_out
;
1807 skb
= tcp_sacktag_skip(skb
, sk
, state
, cache
->end_seq
);
1808 /* Check overlap against next cached too (past this one already) */
1813 if (!before(start_seq
, tcp_highest_sack_seq(tp
))) {
1814 skb
= tcp_highest_sack(sk
);
1817 state
->fack_count
= tp
->fackets_out
;
1819 skb
= tcp_sacktag_skip(skb
, sk
, state
, start_seq
);
1822 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
, state
,
1823 start_seq
, end_seq
, dup_sack
);
1829 /* Clear the head of the cache sack blocks so we can skip it next time */
1830 for (i
= 0; i
< ARRAY_SIZE(tp
->recv_sack_cache
) - used_sacks
; i
++) {
1831 tp
->recv_sack_cache
[i
].start_seq
= 0;
1832 tp
->recv_sack_cache
[i
].end_seq
= 0;
1834 for (j
= 0; j
< used_sacks
; j
++)
1835 tp
->recv_sack_cache
[i
++] = sp
[j
];
1837 if ((state
->reord
< tp
->fackets_out
) &&
1838 ((inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Loss
) || tp
->undo_marker
))
1839 tcp_update_reordering(sk
, tp
->fackets_out
- state
->reord
, 0);
1841 tcp_verify_left_out(tp
);
1844 #if FASTRETRANS_DEBUG > 0
1845 WARN_ON((int)tp
->sacked_out
< 0);
1846 WARN_ON((int)tp
->lost_out
< 0);
1847 WARN_ON((int)tp
->retrans_out
< 0);
1848 WARN_ON((int)tcp_packets_in_flight(tp
) < 0);
1853 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1854 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1856 static bool tcp_limit_reno_sacked(struct tcp_sock
*tp
)
1860 holes
= max(tp
->lost_out
, 1U);
1861 holes
= min(holes
, tp
->packets_out
);
1863 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1864 tp
->sacked_out
= tp
->packets_out
- holes
;
1870 /* If we receive more dupacks than we expected counting segments
1871 * in assumption of absent reordering, interpret this as reordering.
1872 * The only another reason could be bug in receiver TCP.
1874 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1876 struct tcp_sock
*tp
= tcp_sk(sk
);
1877 if (tcp_limit_reno_sacked(tp
))
1878 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1881 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1883 static void tcp_add_reno_sack(struct sock
*sk
)
1885 struct tcp_sock
*tp
= tcp_sk(sk
);
1886 u32 prior_sacked
= tp
->sacked_out
;
1889 tcp_check_reno_reordering(sk
, 0);
1890 if (tp
->sacked_out
> prior_sacked
)
1891 tp
->delivered
++; /* Some out-of-order packet is delivered */
1892 tcp_verify_left_out(tp
);
1895 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1897 static void tcp_remove_reno_sacks(struct sock
*sk
, int acked
)
1899 struct tcp_sock
*tp
= tcp_sk(sk
);
1902 /* One ACK acked hole. The rest eat duplicate ACKs. */
1903 tp
->delivered
+= max_t(int, acked
- tp
->sacked_out
, 1);
1904 if (acked
- 1 >= tp
->sacked_out
)
1907 tp
->sacked_out
-= acked
- 1;
1909 tcp_check_reno_reordering(sk
, acked
);
1910 tcp_verify_left_out(tp
);
1913 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1918 void tcp_clear_retrans(struct tcp_sock
*tp
)
1920 tp
->retrans_out
= 0;
1922 tp
->undo_marker
= 0;
1923 tp
->undo_retrans
= -1;
1924 tp
->fackets_out
= 0;
1928 static inline void tcp_init_undo(struct tcp_sock
*tp
)
1930 tp
->undo_marker
= tp
->snd_una
;
1931 /* Retransmission still in flight may cause DSACKs later. */
1932 tp
->undo_retrans
= tp
->retrans_out
? : -1;
1935 /* Enter Loss state. If we detect SACK reneging, forget all SACK information
1936 * and reset tags completely, otherwise preserve SACKs. If receiver
1937 * dropped its ofo queue, we will know this due to reneging detection.
1939 void tcp_enter_loss(struct sock
*sk
)
1941 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1942 struct tcp_sock
*tp
= tcp_sk(sk
);
1943 struct net
*net
= sock_net(sk
);
1944 struct sk_buff
*skb
;
1945 bool new_recovery
= icsk
->icsk_ca_state
< TCP_CA_Recovery
;
1946 bool is_reneg
; /* is receiver reneging on SACKs? */
1949 /* Reduce ssthresh if it has not yet been made inside this window. */
1950 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
||
1951 !after(tp
->high_seq
, tp
->snd_una
) ||
1952 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
1953 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1954 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1955 tcp_ca_event(sk
, CA_EVENT_LOSS
);
1959 tp
->snd_cwnd_cnt
= 0;
1960 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
1962 tp
->retrans_out
= 0;
1965 if (tcp_is_reno(tp
))
1966 tcp_reset_reno_sack(tp
);
1968 skb
= tcp_write_queue_head(sk
);
1969 is_reneg
= skb
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
);
1971 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPSACKRENEGING
);
1973 tp
->fackets_out
= 0;
1975 tcp_clear_all_retrans_hints(tp
);
1977 tcp_for_write_queue(skb
, sk
) {
1978 if (skb
== tcp_send_head(sk
))
1981 mark_lost
= (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
) ||
1984 tcp_sum_lost(tp
, skb
);
1985 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
1987 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
1988 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1989 tp
->lost_out
+= tcp_skb_pcount(skb
);
1992 tcp_verify_left_out(tp
);
1994 /* Timeout in disordered state after receiving substantial DUPACKs
1995 * suggests that the degree of reordering is over-estimated.
1997 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
&&
1998 tp
->sacked_out
>= net
->ipv4
.sysctl_tcp_reordering
)
1999 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
2000 net
->ipv4
.sysctl_tcp_reordering
);
2001 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2002 tp
->high_seq
= tp
->snd_nxt
;
2003 tcp_ecn_queue_cwr(tp
);
2005 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
2006 * loss recovery is underway except recurring timeout(s) on
2007 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
2009 * In theory F-RTO can be used repeatedly during loss recovery.
2010 * In practice this interacts badly with broken middle-boxes that
2011 * falsely raise the receive window, which results in repeated
2012 * timeouts and stop-and-go behavior.
2014 tp
->frto
= sysctl_tcp_frto
&&
2015 (new_recovery
|| icsk
->icsk_retransmits
) &&
2016 !inet_csk(sk
)->icsk_mtup
.probe_size
;
2019 /* If ACK arrived pointing to a remembered SACK, it means that our
2020 * remembered SACKs do not reflect real state of receiver i.e.
2021 * receiver _host_ is heavily congested (or buggy).
2023 * To avoid big spurious retransmission bursts due to transient SACK
2024 * scoreboard oddities that look like reneging, we give the receiver a
2025 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2026 * restore sanity to the SACK scoreboard. If the apparent reneging
2027 * persists until this RTO then we'll clear the SACK scoreboard.
2029 static bool tcp_check_sack_reneging(struct sock
*sk
, int flag
)
2031 if (flag
& FLAG_SACK_RENEGING
) {
2032 struct tcp_sock
*tp
= tcp_sk(sk
);
2033 unsigned long delay
= max(usecs_to_jiffies(tp
->srtt_us
>> 4),
2034 msecs_to_jiffies(10));
2036 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2037 delay
, TCP_RTO_MAX
);
2043 static inline int tcp_fackets_out(const struct tcp_sock
*tp
)
2045 return tcp_is_reno(tp
) ? tp
->sacked_out
+ 1 : tp
->fackets_out
;
2048 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2049 * counter when SACK is enabled (without SACK, sacked_out is used for
2052 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2053 * segments up to the highest received SACK block so far and holes in
2056 * With reordering, holes may still be in flight, so RFC3517 recovery
2057 * uses pure sacked_out (total number of SACKed segments) even though
2058 * it violates the RFC that uses duplicate ACKs, often these are equal
2059 * but when e.g. out-of-window ACKs or packet duplication occurs,
2060 * they differ. Since neither occurs due to loss, TCP should really
2063 static inline int tcp_dupack_heuristics(const struct tcp_sock
*tp
)
2065 return tcp_is_fack(tp
) ? tp
->fackets_out
: tp
->sacked_out
+ 1;
2068 /* Linux NewReno/SACK/FACK/ECN state machine.
2069 * --------------------------------------
2071 * "Open" Normal state, no dubious events, fast path.
2072 * "Disorder" In all the respects it is "Open",
2073 * but requires a bit more attention. It is entered when
2074 * we see some SACKs or dupacks. It is split of "Open"
2075 * mainly to move some processing from fast path to slow one.
2076 * "CWR" CWND was reduced due to some Congestion Notification event.
2077 * It can be ECN, ICMP source quench, local device congestion.
2078 * "Recovery" CWND was reduced, we are fast-retransmitting.
2079 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2081 * tcp_fastretrans_alert() is entered:
2082 * - each incoming ACK, if state is not "Open"
2083 * - when arrived ACK is unusual, namely:
2088 * Counting packets in flight is pretty simple.
2090 * in_flight = packets_out - left_out + retrans_out
2092 * packets_out is SND.NXT-SND.UNA counted in packets.
2094 * retrans_out is number of retransmitted segments.
2096 * left_out is number of segments left network, but not ACKed yet.
2098 * left_out = sacked_out + lost_out
2100 * sacked_out: Packets, which arrived to receiver out of order
2101 * and hence not ACKed. With SACKs this number is simply
2102 * amount of SACKed data. Even without SACKs
2103 * it is easy to give pretty reliable estimate of this number,
2104 * counting duplicate ACKs.
2106 * lost_out: Packets lost by network. TCP has no explicit
2107 * "loss notification" feedback from network (for now).
2108 * It means that this number can be only _guessed_.
2109 * Actually, it is the heuristics to predict lossage that
2110 * distinguishes different algorithms.
2112 * F.e. after RTO, when all the queue is considered as lost,
2113 * lost_out = packets_out and in_flight = retrans_out.
2115 * Essentially, we have now a few algorithms detecting
2118 * If the receiver supports SACK:
2120 * RFC6675/3517: It is the conventional algorithm. A packet is
2121 * considered lost if the number of higher sequence packets
2122 * SACKed is greater than or equal the DUPACK thoreshold
2123 * (reordering). This is implemented in tcp_mark_head_lost and
2124 * tcp_update_scoreboard.
2126 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm
2127 * (2017-) that checks timing instead of counting DUPACKs.
2128 * Essentially a packet is considered lost if it's not S/ACKed
2129 * after RTT + reordering_window, where both metrics are
2130 * dynamically measured and adjusted. This is implemented in
2131 * tcp_rack_mark_lost.
2133 * FACK (Disabled by default. Subsumbed by RACK):
2134 * It is the simplest heuristics. As soon as we decided
2135 * that something is lost, we decide that _all_ not SACKed
2136 * packets until the most forward SACK are lost. I.e.
2137 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2138 * It is absolutely correct estimate, if network does not reorder
2139 * packets. And it loses any connection to reality when reordering
2140 * takes place. We use FACK by default until reordering
2141 * is suspected on the path to this destination.
2143 * If the receiver does not support SACK:
2145 * NewReno (RFC6582): in Recovery we assume that one segment
2146 * is lost (classic Reno). While we are in Recovery and
2147 * a partial ACK arrives, we assume that one more packet
2148 * is lost (NewReno). This heuristics are the same in NewReno
2151 * Really tricky (and requiring careful tuning) part of algorithm
2152 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2153 * The first determines the moment _when_ we should reduce CWND and,
2154 * hence, slow down forward transmission. In fact, it determines the moment
2155 * when we decide that hole is caused by loss, rather than by a reorder.
2157 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2158 * holes, caused by lost packets.
2160 * And the most logically complicated part of algorithm is undo
2161 * heuristics. We detect false retransmits due to both too early
2162 * fast retransmit (reordering) and underestimated RTO, analyzing
2163 * timestamps and D-SACKs. When we detect that some segments were
2164 * retransmitted by mistake and CWND reduction was wrong, we undo
2165 * window reduction and abort recovery phase. This logic is hidden
2166 * inside several functions named tcp_try_undo_<something>.
2169 /* This function decides, when we should leave Disordered state
2170 * and enter Recovery phase, reducing congestion window.
2172 * Main question: may we further continue forward transmission
2173 * with the same cwnd?
2175 static bool tcp_time_to_recover(struct sock
*sk
, int flag
)
2177 struct tcp_sock
*tp
= tcp_sk(sk
);
2179 /* Trick#1: The loss is proven. */
2183 /* Not-A-Trick#2 : Classic rule... */
2184 if (tcp_dupack_heuristics(tp
) > tp
->reordering
)
2190 /* Detect loss in event "A" above by marking head of queue up as lost.
2191 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2192 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2193 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2194 * the maximum SACKed segments to pass before reaching this limit.
2196 static void tcp_mark_head_lost(struct sock
*sk
, int packets
, int mark_head
)
2198 struct tcp_sock
*tp
= tcp_sk(sk
);
2199 struct sk_buff
*skb
;
2200 int cnt
, oldcnt
, lost
;
2202 /* Use SACK to deduce losses of new sequences sent during recovery */
2203 const u32 loss_high
= tcp_is_sack(tp
) ? tp
->snd_nxt
: tp
->high_seq
;
2205 WARN_ON(packets
> tp
->packets_out
);
2206 if (tp
->lost_skb_hint
) {
2207 skb
= tp
->lost_skb_hint
;
2208 cnt
= tp
->lost_cnt_hint
;
2209 /* Head already handled? */
2210 if (mark_head
&& skb
!= tcp_write_queue_head(sk
))
2213 skb
= tcp_write_queue_head(sk
);
2217 tcp_for_write_queue_from(skb
, sk
) {
2218 if (skb
== tcp_send_head(sk
))
2220 /* TODO: do this better */
2221 /* this is not the most efficient way to do this... */
2222 tp
->lost_skb_hint
= skb
;
2223 tp
->lost_cnt_hint
= cnt
;
2225 if (after(TCP_SKB_CB(skb
)->end_seq
, loss_high
))
2229 if (tcp_is_fack(tp
) || tcp_is_reno(tp
) ||
2230 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2231 cnt
+= tcp_skb_pcount(skb
);
2233 if (cnt
> packets
) {
2234 if ((tcp_is_sack(tp
) && !tcp_is_fack(tp
)) ||
2235 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
) ||
2236 (oldcnt
>= packets
))
2239 mss
= tcp_skb_mss(skb
);
2240 /* If needed, chop off the prefix to mark as lost. */
2241 lost
= (packets
- oldcnt
) * mss
;
2242 if (lost
< skb
->len
&&
2243 tcp_fragment(sk
, skb
, lost
, mss
, GFP_ATOMIC
) < 0)
2248 tcp_skb_mark_lost(tp
, skb
);
2253 tcp_verify_left_out(tp
);
2256 /* Account newly detected lost packet(s) */
2258 static void tcp_update_scoreboard(struct sock
*sk
, int fast_rexmit
)
2260 struct tcp_sock
*tp
= tcp_sk(sk
);
2262 if (tcp_is_reno(tp
)) {
2263 tcp_mark_head_lost(sk
, 1, 1);
2264 } else if (tcp_is_fack(tp
)) {
2265 int lost
= tp
->fackets_out
- tp
->reordering
;
2268 tcp_mark_head_lost(sk
, lost
, 0);
2270 int sacked_upto
= tp
->sacked_out
- tp
->reordering
;
2271 if (sacked_upto
>= 0)
2272 tcp_mark_head_lost(sk
, sacked_upto
, 0);
2273 else if (fast_rexmit
)
2274 tcp_mark_head_lost(sk
, 1, 1);
2278 static bool tcp_tsopt_ecr_before(const struct tcp_sock
*tp
, u32 when
)
2280 return tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2281 before(tp
->rx_opt
.rcv_tsecr
, when
);
2284 /* skb is spurious retransmitted if the returned timestamp echo
2285 * reply is prior to the skb transmission time
2287 static bool tcp_skb_spurious_retrans(const struct tcp_sock
*tp
,
2288 const struct sk_buff
*skb
)
2290 return (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
) &&
2291 tcp_tsopt_ecr_before(tp
, tcp_skb_timestamp(skb
));
2294 /* Nothing was retransmitted or returned timestamp is less
2295 * than timestamp of the first retransmission.
2297 static inline bool tcp_packet_delayed(const struct tcp_sock
*tp
)
2299 return !tp
->retrans_stamp
||
2300 tcp_tsopt_ecr_before(tp
, tp
->retrans_stamp
);
2303 /* Undo procedures. */
2305 /* We can clear retrans_stamp when there are no retransmissions in the
2306 * window. It would seem that it is trivially available for us in
2307 * tp->retrans_out, however, that kind of assumptions doesn't consider
2308 * what will happen if errors occur when sending retransmission for the
2309 * second time. ...It could the that such segment has only
2310 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2311 * the head skb is enough except for some reneging corner cases that
2312 * are not worth the effort.
2314 * Main reason for all this complexity is the fact that connection dying
2315 * time now depends on the validity of the retrans_stamp, in particular,
2316 * that successive retransmissions of a segment must not advance
2317 * retrans_stamp under any conditions.
2319 static bool tcp_any_retrans_done(const struct sock
*sk
)
2321 const struct tcp_sock
*tp
= tcp_sk(sk
);
2322 struct sk_buff
*skb
;
2324 if (tp
->retrans_out
)
2327 skb
= tcp_write_queue_head(sk
);
2328 if (unlikely(skb
&& TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
))
2334 #if FASTRETRANS_DEBUG > 1
2335 static void DBGUNDO(struct sock
*sk
, const char *msg
)
2337 struct tcp_sock
*tp
= tcp_sk(sk
);
2338 struct inet_sock
*inet
= inet_sk(sk
);
2340 if (sk
->sk_family
== AF_INET
) {
2341 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2343 &inet
->inet_daddr
, ntohs(inet
->inet_dport
),
2344 tp
->snd_cwnd
, tcp_left_out(tp
),
2345 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2348 #if IS_ENABLED(CONFIG_IPV6)
2349 else if (sk
->sk_family
== AF_INET6
) {
2350 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2352 &sk
->sk_v6_daddr
, ntohs(inet
->inet_dport
),
2353 tp
->snd_cwnd
, tcp_left_out(tp
),
2354 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2360 #define DBGUNDO(x...) do { } while (0)
2363 static void tcp_undo_cwnd_reduction(struct sock
*sk
, bool unmark_loss
)
2365 struct tcp_sock
*tp
= tcp_sk(sk
);
2368 struct sk_buff
*skb
;
2370 tcp_for_write_queue(skb
, sk
) {
2371 if (skb
== tcp_send_head(sk
))
2373 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2376 tcp_clear_all_retrans_hints(tp
);
2379 if (tp
->prior_ssthresh
) {
2380 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2382 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
2384 if (tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
2385 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
2386 tcp_ecn_withdraw_cwr(tp
);
2389 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
2390 tp
->undo_marker
= 0;
2393 static inline bool tcp_may_undo(const struct tcp_sock
*tp
)
2395 return tp
->undo_marker
&& (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
2398 /* People celebrate: "We love our President!" */
2399 static bool tcp_try_undo_recovery(struct sock
*sk
)
2401 struct tcp_sock
*tp
= tcp_sk(sk
);
2403 if (tcp_may_undo(tp
)) {
2406 /* Happy end! We did not retransmit anything
2407 * or our original transmission succeeded.
2409 DBGUNDO(sk
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
2410 tcp_undo_cwnd_reduction(sk
, false);
2411 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
2412 mib_idx
= LINUX_MIB_TCPLOSSUNDO
;
2414 mib_idx
= LINUX_MIB_TCPFULLUNDO
;
2416 NET_INC_STATS(sock_net(sk
), mib_idx
);
2418 if (tp
->snd_una
== tp
->high_seq
&& tcp_is_reno(tp
)) {
2419 /* Hold old state until something *above* high_seq
2420 * is ACKed. For Reno it is MUST to prevent false
2421 * fast retransmits (RFC2582). SACK TCP is safe. */
2422 if (!tcp_any_retrans_done(sk
))
2423 tp
->retrans_stamp
= 0;
2426 tcp_set_ca_state(sk
, TCP_CA_Open
);
2430 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2431 static bool tcp_try_undo_dsack(struct sock
*sk
)
2433 struct tcp_sock
*tp
= tcp_sk(sk
);
2435 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
2436 DBGUNDO(sk
, "D-SACK");
2437 tcp_undo_cwnd_reduction(sk
, false);
2438 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPDSACKUNDO
);
2444 /* Undo during loss recovery after partial ACK or using F-RTO. */
2445 static bool tcp_try_undo_loss(struct sock
*sk
, bool frto_undo
)
2447 struct tcp_sock
*tp
= tcp_sk(sk
);
2449 if (frto_undo
|| tcp_may_undo(tp
)) {
2450 tcp_undo_cwnd_reduction(sk
, true);
2452 DBGUNDO(sk
, "partial loss");
2453 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPLOSSUNDO
);
2455 NET_INC_STATS(sock_net(sk
),
2456 LINUX_MIB_TCPSPURIOUSRTOS
);
2457 inet_csk(sk
)->icsk_retransmits
= 0;
2458 if (frto_undo
|| tcp_is_sack(tp
))
2459 tcp_set_ca_state(sk
, TCP_CA_Open
);
2465 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2466 * It computes the number of packets to send (sndcnt) based on packets newly
2468 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2469 * cwnd reductions across a full RTT.
2470 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2471 * But when the retransmits are acked without further losses, PRR
2472 * slow starts cwnd up to ssthresh to speed up the recovery.
2474 static void tcp_init_cwnd_reduction(struct sock
*sk
)
2476 struct tcp_sock
*tp
= tcp_sk(sk
);
2478 tp
->high_seq
= tp
->snd_nxt
;
2479 tp
->tlp_high_seq
= 0;
2480 tp
->snd_cwnd_cnt
= 0;
2481 tp
->prior_cwnd
= tp
->snd_cwnd
;
2482 tp
->prr_delivered
= 0;
2484 tp
->snd_ssthresh
= inet_csk(sk
)->icsk_ca_ops
->ssthresh(sk
);
2485 tcp_ecn_queue_cwr(tp
);
2488 void tcp_cwnd_reduction(struct sock
*sk
, int newly_acked_sacked
, int flag
)
2490 struct tcp_sock
*tp
= tcp_sk(sk
);
2492 int delta
= tp
->snd_ssthresh
- tcp_packets_in_flight(tp
);
2494 if (newly_acked_sacked
<= 0 || WARN_ON_ONCE(!tp
->prior_cwnd
))
2497 tp
->prr_delivered
+= newly_acked_sacked
;
2499 u64 dividend
= (u64
)tp
->snd_ssthresh
* tp
->prr_delivered
+
2501 sndcnt
= div_u64(dividend
, tp
->prior_cwnd
) - tp
->prr_out
;
2502 } else if ((flag
& FLAG_RETRANS_DATA_ACKED
) &&
2503 !(flag
& FLAG_LOST_RETRANS
)) {
2504 sndcnt
= min_t(int, delta
,
2505 max_t(int, tp
->prr_delivered
- tp
->prr_out
,
2506 newly_acked_sacked
) + 1);
2508 sndcnt
= min(delta
, newly_acked_sacked
);
2510 /* Force a fast retransmit upon entering fast recovery */
2511 sndcnt
= max(sndcnt
, (tp
->prr_out
? 0 : 1));
2512 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + sndcnt
;
2515 static inline void tcp_end_cwnd_reduction(struct sock
*sk
)
2517 struct tcp_sock
*tp
= tcp_sk(sk
);
2519 if (inet_csk(sk
)->icsk_ca_ops
->cong_control
)
2522 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2523 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_CWR
||
2524 (tp
->undo_marker
&& tp
->snd_ssthresh
< TCP_INFINITE_SSTHRESH
)) {
2525 tp
->snd_cwnd
= tp
->snd_ssthresh
;
2526 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
2528 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
2531 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2532 void tcp_enter_cwr(struct sock
*sk
)
2534 struct tcp_sock
*tp
= tcp_sk(sk
);
2536 tp
->prior_ssthresh
= 0;
2537 if (inet_csk(sk
)->icsk_ca_state
< TCP_CA_CWR
) {
2538 tp
->undo_marker
= 0;
2539 tcp_init_cwnd_reduction(sk
);
2540 tcp_set_ca_state(sk
, TCP_CA_CWR
);
2543 EXPORT_SYMBOL(tcp_enter_cwr
);
2545 static void tcp_try_keep_open(struct sock
*sk
)
2547 struct tcp_sock
*tp
= tcp_sk(sk
);
2548 int state
= TCP_CA_Open
;
2550 if (tcp_left_out(tp
) || tcp_any_retrans_done(sk
))
2551 state
= TCP_CA_Disorder
;
2553 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
2554 tcp_set_ca_state(sk
, state
);
2555 tp
->high_seq
= tp
->snd_nxt
;
2559 static void tcp_try_to_open(struct sock
*sk
, int flag
)
2561 struct tcp_sock
*tp
= tcp_sk(sk
);
2563 tcp_verify_left_out(tp
);
2565 if (!tcp_any_retrans_done(sk
))
2566 tp
->retrans_stamp
= 0;
2568 if (flag
& FLAG_ECE
)
2571 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
2572 tcp_try_keep_open(sk
);
2576 static void tcp_mtup_probe_failed(struct sock
*sk
)
2578 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2580 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
2581 icsk
->icsk_mtup
.probe_size
= 0;
2582 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPMTUPFAIL
);
2585 static void tcp_mtup_probe_success(struct sock
*sk
)
2587 struct tcp_sock
*tp
= tcp_sk(sk
);
2588 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2590 /* FIXME: breaks with very large cwnd */
2591 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2592 tp
->snd_cwnd
= tp
->snd_cwnd
*
2593 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
2594 icsk
->icsk_mtup
.probe_size
;
2595 tp
->snd_cwnd_cnt
= 0;
2596 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
2597 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2599 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
2600 icsk
->icsk_mtup
.probe_size
= 0;
2601 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
2602 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPMTUPSUCCESS
);
2605 /* Do a simple retransmit without using the backoff mechanisms in
2606 * tcp_timer. This is used for path mtu discovery.
2607 * The socket is already locked here.
2609 void tcp_simple_retransmit(struct sock
*sk
)
2611 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2612 struct tcp_sock
*tp
= tcp_sk(sk
);
2613 struct sk_buff
*skb
;
2614 unsigned int mss
= tcp_current_mss(sk
);
2615 u32 prior_lost
= tp
->lost_out
;
2617 tcp_for_write_queue(skb
, sk
) {
2618 if (skb
== tcp_send_head(sk
))
2620 if (tcp_skb_seglen(skb
) > mss
&&
2621 !(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
2622 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2623 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
2624 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2626 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
2630 tcp_clear_retrans_hints_partial(tp
);
2632 if (prior_lost
== tp
->lost_out
)
2635 if (tcp_is_reno(tp
))
2636 tcp_limit_reno_sacked(tp
);
2638 tcp_verify_left_out(tp
);
2640 /* Don't muck with the congestion window here.
2641 * Reason is that we do not increase amount of _data_
2642 * in network, but units changed and effective
2643 * cwnd/ssthresh really reduced now.
2645 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
2646 tp
->high_seq
= tp
->snd_nxt
;
2647 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2648 tp
->prior_ssthresh
= 0;
2649 tp
->undo_marker
= 0;
2650 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2652 tcp_xmit_retransmit_queue(sk
);
2654 EXPORT_SYMBOL(tcp_simple_retransmit
);
2656 void tcp_enter_recovery(struct sock
*sk
, bool ece_ack
)
2658 struct tcp_sock
*tp
= tcp_sk(sk
);
2661 if (tcp_is_reno(tp
))
2662 mib_idx
= LINUX_MIB_TCPRENORECOVERY
;
2664 mib_idx
= LINUX_MIB_TCPSACKRECOVERY
;
2666 NET_INC_STATS(sock_net(sk
), mib_idx
);
2668 tp
->prior_ssthresh
= 0;
2671 if (!tcp_in_cwnd_reduction(sk
)) {
2673 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2674 tcp_init_cwnd_reduction(sk
);
2676 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
2679 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2680 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2682 static void tcp_process_loss(struct sock
*sk
, int flag
, bool is_dupack
,
2685 struct tcp_sock
*tp
= tcp_sk(sk
);
2686 bool recovered
= !before(tp
->snd_una
, tp
->high_seq
);
2688 if ((flag
& FLAG_SND_UNA_ADVANCED
) &&
2689 tcp_try_undo_loss(sk
, false))
2692 /* The ACK (s)acks some never-retransmitted data meaning not all
2693 * the data packets before the timeout were lost. Therefore we
2694 * undo the congestion window and state. This is essentially
2695 * the operation in F-RTO (RFC5682 section 3.1 step 3.b). Since
2696 * a retransmitted skb is permantly marked, we can apply such an
2697 * operation even if F-RTO was not used.
2699 if ((flag
& FLAG_ORIG_SACK_ACKED
) &&
2700 tcp_try_undo_loss(sk
, tp
->undo_marker
))
2703 if (tp
->frto
) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2704 if (after(tp
->snd_nxt
, tp
->high_seq
)) {
2705 if (flag
& FLAG_DATA_SACKED
|| is_dupack
)
2706 tp
->frto
= 0; /* Step 3.a. loss was real */
2707 } else if (flag
& FLAG_SND_UNA_ADVANCED
&& !recovered
) {
2708 tp
->high_seq
= tp
->snd_nxt
;
2709 /* Step 2.b. Try send new data (but deferred until cwnd
2710 * is updated in tcp_ack()). Otherwise fall back to
2711 * the conventional recovery.
2713 if (tcp_send_head(sk
) &&
2714 after(tcp_wnd_end(tp
), tp
->snd_nxt
)) {
2715 *rexmit
= REXMIT_NEW
;
2723 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2724 tcp_try_undo_recovery(sk
);
2727 if (tcp_is_reno(tp
)) {
2728 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2729 * delivered. Lower inflight to clock out (re)tranmissions.
2731 if (after(tp
->snd_nxt
, tp
->high_seq
) && is_dupack
)
2732 tcp_add_reno_sack(sk
);
2733 else if (flag
& FLAG_SND_UNA_ADVANCED
)
2734 tcp_reset_reno_sack(tp
);
2736 *rexmit
= REXMIT_LOST
;
2739 /* Undo during fast recovery after partial ACK. */
2740 static bool tcp_try_undo_partial(struct sock
*sk
, const int acked
)
2742 struct tcp_sock
*tp
= tcp_sk(sk
);
2744 if (tp
->undo_marker
&& tcp_packet_delayed(tp
)) {
2745 /* Plain luck! Hole if filled with delayed
2746 * packet, rather than with a retransmit.
2748 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
2750 /* We are getting evidence that the reordering degree is higher
2751 * than we realized. If there are no retransmits out then we
2752 * can undo. Otherwise we clock out new packets but do not
2753 * mark more packets lost or retransmit more.
2755 if (tp
->retrans_out
)
2758 if (!tcp_any_retrans_done(sk
))
2759 tp
->retrans_stamp
= 0;
2761 DBGUNDO(sk
, "partial recovery");
2762 tcp_undo_cwnd_reduction(sk
, true);
2763 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPPARTIALUNDO
);
2764 tcp_try_keep_open(sk
);
2770 static void tcp_rack_identify_loss(struct sock
*sk
, int *ack_flag
)
2772 struct tcp_sock
*tp
= tcp_sk(sk
);
2774 /* Use RACK to detect loss */
2775 if (sysctl_tcp_recovery
& TCP_RACK_LOSS_DETECTION
) {
2776 u32 prior_retrans
= tp
->retrans_out
;
2778 tcp_rack_mark_lost(sk
);
2779 if (prior_retrans
> tp
->retrans_out
)
2780 *ack_flag
|= FLAG_LOST_RETRANS
;
2784 /* Process an event, which can update packets-in-flight not trivially.
2785 * Main goal of this function is to calculate new estimate for left_out,
2786 * taking into account both packets sitting in receiver's buffer and
2787 * packets lost by network.
2789 * Besides that it updates the congestion state when packet loss or ECN
2790 * is detected. But it does not reduce the cwnd, it is done by the
2791 * congestion control later.
2793 * It does _not_ decide what to send, it is made in function
2794 * tcp_xmit_retransmit_queue().
2796 static void tcp_fastretrans_alert(struct sock
*sk
, const int acked
,
2797 bool is_dupack
, int *ack_flag
, int *rexmit
)
2799 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2800 struct tcp_sock
*tp
= tcp_sk(sk
);
2801 int fast_rexmit
= 0, flag
= *ack_flag
;
2802 bool do_lost
= is_dupack
|| ((flag
& FLAG_DATA_SACKED
) &&
2803 (tcp_fackets_out(tp
) > tp
->reordering
));
2805 if (WARN_ON(!tp
->packets_out
&& tp
->sacked_out
))
2807 if (WARN_ON(!tp
->sacked_out
&& tp
->fackets_out
))
2808 tp
->fackets_out
= 0;
2810 /* Now state machine starts.
2811 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2812 if (flag
& FLAG_ECE
)
2813 tp
->prior_ssthresh
= 0;
2815 /* B. In all the states check for reneging SACKs. */
2816 if (tcp_check_sack_reneging(sk
, flag
))
2819 /* C. Check consistency of the current state. */
2820 tcp_verify_left_out(tp
);
2822 /* D. Check state exit conditions. State can be terminated
2823 * when high_seq is ACKed. */
2824 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
2825 WARN_ON(tp
->retrans_out
!= 0);
2826 tp
->retrans_stamp
= 0;
2827 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
2828 switch (icsk
->icsk_ca_state
) {
2830 /* CWR is to be held something *above* high_seq
2831 * is ACKed for CWR bit to reach receiver. */
2832 if (tp
->snd_una
!= tp
->high_seq
) {
2833 tcp_end_cwnd_reduction(sk
);
2834 tcp_set_ca_state(sk
, TCP_CA_Open
);
2838 case TCP_CA_Recovery
:
2839 if (tcp_is_reno(tp
))
2840 tcp_reset_reno_sack(tp
);
2841 if (tcp_try_undo_recovery(sk
))
2843 tcp_end_cwnd_reduction(sk
);
2848 /* E. Process state. */
2849 switch (icsk
->icsk_ca_state
) {
2850 case TCP_CA_Recovery
:
2851 if (!(flag
& FLAG_SND_UNA_ADVANCED
)) {
2852 if (tcp_is_reno(tp
) && is_dupack
)
2853 tcp_add_reno_sack(sk
);
2855 if (tcp_try_undo_partial(sk
, acked
))
2857 /* Partial ACK arrived. Force fast retransmit. */
2858 do_lost
= tcp_is_reno(tp
) ||
2859 tcp_fackets_out(tp
) > tp
->reordering
;
2861 if (tcp_try_undo_dsack(sk
)) {
2862 tcp_try_keep_open(sk
);
2865 tcp_rack_identify_loss(sk
, ack_flag
);
2868 tcp_process_loss(sk
, flag
, is_dupack
, rexmit
);
2869 tcp_rack_identify_loss(sk
, ack_flag
);
2870 if (!(icsk
->icsk_ca_state
== TCP_CA_Open
||
2871 (*ack_flag
& FLAG_LOST_RETRANS
)))
2873 /* Change state if cwnd is undone or retransmits are lost */
2875 if (tcp_is_reno(tp
)) {
2876 if (flag
& FLAG_SND_UNA_ADVANCED
)
2877 tcp_reset_reno_sack(tp
);
2879 tcp_add_reno_sack(sk
);
2882 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
)
2883 tcp_try_undo_dsack(sk
);
2885 tcp_rack_identify_loss(sk
, ack_flag
);
2886 if (!tcp_time_to_recover(sk
, flag
)) {
2887 tcp_try_to_open(sk
, flag
);
2891 /* MTU probe failure: don't reduce cwnd */
2892 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
2893 icsk
->icsk_mtup
.probe_size
&&
2894 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
2895 tcp_mtup_probe_failed(sk
);
2896 /* Restores the reduction we did in tcp_mtup_probe() */
2898 tcp_simple_retransmit(sk
);
2902 /* Otherwise enter Recovery state */
2903 tcp_enter_recovery(sk
, (flag
& FLAG_ECE
));
2908 tcp_update_scoreboard(sk
, fast_rexmit
);
2909 *rexmit
= REXMIT_LOST
;
2912 static void tcp_update_rtt_min(struct sock
*sk
, u32 rtt_us
)
2914 struct tcp_sock
*tp
= tcp_sk(sk
);
2915 u32 wlen
= sysctl_tcp_min_rtt_wlen
* HZ
;
2917 minmax_running_min(&tp
->rtt_min
, wlen
, tcp_jiffies32
,
2918 rtt_us
? : jiffies_to_usecs(1));
2921 static bool tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
2922 long seq_rtt_us
, long sack_rtt_us
,
2923 long ca_rtt_us
, struct rate_sample
*rs
)
2925 const struct tcp_sock
*tp
= tcp_sk(sk
);
2927 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2928 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2929 * Karn's algorithm forbids taking RTT if some retransmitted data
2930 * is acked (RFC6298).
2933 seq_rtt_us
= sack_rtt_us
;
2935 /* RTTM Rule: A TSecr value received in a segment is used to
2936 * update the averaged RTT measurement only if the segment
2937 * acknowledges some new data, i.e., only if it advances the
2938 * left edge of the send window.
2939 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2941 if (seq_rtt_us
< 0 && tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2942 flag
& FLAG_ACKED
) {
2943 u32 delta
= tcp_time_stamp(tp
) - tp
->rx_opt
.rcv_tsecr
;
2944 u32 delta_us
= delta
* (USEC_PER_SEC
/ TCP_TS_HZ
);
2946 seq_rtt_us
= ca_rtt_us
= delta_us
;
2948 rs
->rtt_us
= ca_rtt_us
; /* RTT of last (S)ACKed packet (or -1) */
2952 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
2953 * always taken together with ACK, SACK, or TS-opts. Any negative
2954 * values will be skipped with the seq_rtt_us < 0 check above.
2956 tcp_update_rtt_min(sk
, ca_rtt_us
);
2957 tcp_rtt_estimator(sk
, seq_rtt_us
);
2960 /* RFC6298: only reset backoff on valid RTT measurement. */
2961 inet_csk(sk
)->icsk_backoff
= 0;
2965 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2966 void tcp_synack_rtt_meas(struct sock
*sk
, struct request_sock
*req
)
2968 struct rate_sample rs
;
2971 if (req
&& !req
->num_retrans
&& tcp_rsk(req
)->snt_synack
)
2972 rtt_us
= tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req
)->snt_synack
);
2974 tcp_ack_update_rtt(sk
, FLAG_SYN_ACKED
, rtt_us
, -1L, rtt_us
, &rs
);
2978 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
, u32 acked
)
2980 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2982 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, acked
);
2983 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_jiffies32
;
2986 /* Restart timer after forward progress on connection.
2987 * RFC2988 recommends to restart timer to now+rto.
2989 void tcp_rearm_rto(struct sock
*sk
)
2991 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2992 struct tcp_sock
*tp
= tcp_sk(sk
);
2994 /* If the retrans timer is currently being used by Fast Open
2995 * for SYN-ACK retrans purpose, stay put.
2997 if (tp
->fastopen_rsk
)
3000 if (!tp
->packets_out
) {
3001 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
3003 u32 rto
= inet_csk(sk
)->icsk_rto
;
3004 /* Offset the time elapsed after installing regular RTO */
3005 if (icsk
->icsk_pending
== ICSK_TIME_REO_TIMEOUT
||
3006 icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
) {
3007 struct sk_buff
*skb
= tcp_write_queue_head(sk
);
3008 u64 rto_time_stamp
= skb
->skb_mstamp
+
3009 jiffies_to_usecs(rto
);
3010 s64 delta_us
= rto_time_stamp
- tp
->tcp_mstamp
;
3011 /* delta_us may not be positive if the socket is locked
3012 * when the retrans timer fires and is rescheduled.
3015 rto
= usecs_to_jiffies(delta_us
);
3017 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
, rto
,
3022 /* If we get here, the whole TSO packet has not been acked. */
3023 static u32
tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
)
3025 struct tcp_sock
*tp
= tcp_sk(sk
);
3028 BUG_ON(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
));
3030 packets_acked
= tcp_skb_pcount(skb
);
3031 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
3033 packets_acked
-= tcp_skb_pcount(skb
);
3035 if (packets_acked
) {
3036 BUG_ON(tcp_skb_pcount(skb
) == 0);
3037 BUG_ON(!before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
));
3040 return packets_acked
;
3043 static void tcp_ack_tstamp(struct sock
*sk
, struct sk_buff
*skb
,
3046 const struct skb_shared_info
*shinfo
;
3048 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3049 if (likely(!TCP_SKB_CB(skb
)->txstamp_ack
))
3052 shinfo
= skb_shinfo(skb
);
3053 if (!before(shinfo
->tskey
, prior_snd_una
) &&
3054 before(shinfo
->tskey
, tcp_sk(sk
)->snd_una
))
3055 __skb_tstamp_tx(skb
, NULL
, sk
, SCM_TSTAMP_ACK
);
3058 /* Remove acknowledged frames from the retransmission queue. If our packet
3059 * is before the ack sequence we can discard it as it's confirmed to have
3060 * arrived at the other end.
3062 static int tcp_clean_rtx_queue(struct sock
*sk
, int prior_fackets
,
3063 u32 prior_snd_una
, int *acked
,
3064 struct tcp_sacktag_state
*sack
)
3066 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3067 u64 first_ackt
, last_ackt
;
3068 struct tcp_sock
*tp
= tcp_sk(sk
);
3069 u32 prior_sacked
= tp
->sacked_out
;
3070 u32 reord
= tp
->packets_out
;
3071 bool fully_acked
= true;
3072 long sack_rtt_us
= -1L;
3073 long seq_rtt_us
= -1L;
3074 long ca_rtt_us
= -1L;
3075 struct sk_buff
*skb
;
3077 u32 last_in_flight
= 0;
3083 while ((skb
= tcp_write_queue_head(sk
)) && skb
!= tcp_send_head(sk
)) {
3084 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
3085 u8 sacked
= scb
->sacked
;
3088 tcp_ack_tstamp(sk
, skb
, prior_snd_una
);
3090 /* Determine how many packets and what bytes were acked, tso and else */
3091 if (after(scb
->end_seq
, tp
->snd_una
)) {
3092 if (tcp_skb_pcount(skb
) == 1 ||
3093 !after(tp
->snd_una
, scb
->seq
))
3096 acked_pcount
= tcp_tso_acked(sk
, skb
);
3099 fully_acked
= false;
3101 /* Speedup tcp_unlink_write_queue() and next loop */
3102 prefetchw(skb
->next
);
3103 acked_pcount
= tcp_skb_pcount(skb
);
3106 if (unlikely(sacked
& TCPCB_RETRANS
)) {
3107 if (sacked
& TCPCB_SACKED_RETRANS
)
3108 tp
->retrans_out
-= acked_pcount
;
3109 flag
|= FLAG_RETRANS_DATA_ACKED
;
3110 } else if (!(sacked
& TCPCB_SACKED_ACKED
)) {
3111 last_ackt
= skb
->skb_mstamp
;
3112 WARN_ON_ONCE(last_ackt
== 0);
3114 first_ackt
= last_ackt
;
3116 last_in_flight
= TCP_SKB_CB(skb
)->tx
.in_flight
;
3117 reord
= min(pkts_acked
, reord
);
3118 if (!after(scb
->end_seq
, tp
->high_seq
))
3119 flag
|= FLAG_ORIG_SACK_ACKED
;
3122 if (sacked
& TCPCB_SACKED_ACKED
) {
3123 tp
->sacked_out
-= acked_pcount
;
3124 } else if (tcp_is_sack(tp
)) {
3125 tp
->delivered
+= acked_pcount
;
3126 if (!tcp_skb_spurious_retrans(tp
, skb
))
3127 tcp_rack_advance(tp
, sacked
, scb
->end_seq
,
3130 if (sacked
& TCPCB_LOST
)
3131 tp
->lost_out
-= acked_pcount
;
3133 tp
->packets_out
-= acked_pcount
;
3134 pkts_acked
+= acked_pcount
;
3135 tcp_rate_skb_delivered(sk
, skb
, sack
->rate
);
3137 /* Initial outgoing SYN's get put onto the write_queue
3138 * just like anything else we transmit. It is not
3139 * true data, and if we misinform our callers that
3140 * this ACK acks real data, we will erroneously exit
3141 * connection startup slow start one packet too
3142 * quickly. This is severely frowned upon behavior.
3144 if (likely(!(scb
->tcp_flags
& TCPHDR_SYN
))) {
3145 flag
|= FLAG_DATA_ACKED
;
3147 flag
|= FLAG_SYN_ACKED
;
3148 tp
->retrans_stamp
= 0;
3154 tcp_unlink_write_queue(skb
, sk
);
3155 sk_wmem_free_skb(sk
, skb
);
3156 if (unlikely(skb
== tp
->retransmit_skb_hint
))
3157 tp
->retransmit_skb_hint
= NULL
;
3158 if (unlikely(skb
== tp
->lost_skb_hint
))
3159 tp
->lost_skb_hint
= NULL
;
3163 tcp_chrono_stop(sk
, TCP_CHRONO_BUSY
);
3165 if (likely(between(tp
->snd_up
, prior_snd_una
, tp
->snd_una
)))
3166 tp
->snd_up
= tp
->snd_una
;
3168 if (skb
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
3169 flag
|= FLAG_SACK_RENEGING
;
3171 if (likely(first_ackt
) && !(flag
& FLAG_RETRANS_DATA_ACKED
)) {
3172 seq_rtt_us
= tcp_stamp_us_delta(tp
->tcp_mstamp
, first_ackt
);
3173 ca_rtt_us
= tcp_stamp_us_delta(tp
->tcp_mstamp
, last_ackt
);
3175 if (sack
->first_sackt
) {
3176 sack_rtt_us
= tcp_stamp_us_delta(tp
->tcp_mstamp
, sack
->first_sackt
);
3177 ca_rtt_us
= tcp_stamp_us_delta(tp
->tcp_mstamp
, sack
->last_sackt
);
3179 rtt_update
= tcp_ack_update_rtt(sk
, flag
, seq_rtt_us
, sack_rtt_us
,
3180 ca_rtt_us
, sack
->rate
);
3182 if (flag
& FLAG_ACKED
) {
3184 if (unlikely(icsk
->icsk_mtup
.probe_size
&&
3185 !after(tp
->mtu_probe
.probe_seq_end
, tp
->snd_una
))) {
3186 tcp_mtup_probe_success(sk
);
3189 if (tcp_is_reno(tp
)) {
3190 tcp_remove_reno_sacks(sk
, pkts_acked
);
3194 /* Non-retransmitted hole got filled? That's reordering */
3195 if (reord
< prior_fackets
&& reord
<= tp
->fackets_out
)
3196 tcp_update_reordering(sk
, tp
->fackets_out
- reord
, 0);
3198 delta
= tcp_is_fack(tp
) ? pkts_acked
:
3199 prior_sacked
- tp
->sacked_out
;
3200 tp
->lost_cnt_hint
-= min(tp
->lost_cnt_hint
, delta
);
3203 tp
->fackets_out
-= min(pkts_acked
, tp
->fackets_out
);
3205 } else if (skb
&& rtt_update
&& sack_rtt_us
>= 0 &&
3206 sack_rtt_us
> tcp_stamp_us_delta(tp
->tcp_mstamp
, skb
->skb_mstamp
)) {
3207 /* Do not re-arm RTO if the sack RTT is measured from data sent
3208 * after when the head was last (re)transmitted. Otherwise the
3209 * timeout may continue to extend in loss recovery.
3214 if (icsk
->icsk_ca_ops
->pkts_acked
) {
3215 struct ack_sample sample
= { .pkts_acked
= pkts_acked
,
3216 .rtt_us
= sack
->rate
->rtt_us
,
3217 .in_flight
= last_in_flight
};
3219 icsk
->icsk_ca_ops
->pkts_acked(sk
, &sample
);
3222 #if FASTRETRANS_DEBUG > 0
3223 WARN_ON((int)tp
->sacked_out
< 0);
3224 WARN_ON((int)tp
->lost_out
< 0);
3225 WARN_ON((int)tp
->retrans_out
< 0);
3226 if (!tp
->packets_out
&& tcp_is_sack(tp
)) {
3227 icsk
= inet_csk(sk
);
3229 pr_debug("Leak l=%u %d\n",
3230 tp
->lost_out
, icsk
->icsk_ca_state
);
3233 if (tp
->sacked_out
) {
3234 pr_debug("Leak s=%u %d\n",
3235 tp
->sacked_out
, icsk
->icsk_ca_state
);
3238 if (tp
->retrans_out
) {
3239 pr_debug("Leak r=%u %d\n",
3240 tp
->retrans_out
, icsk
->icsk_ca_state
);
3241 tp
->retrans_out
= 0;
3245 *acked
= pkts_acked
;
3249 static void tcp_ack_probe(struct sock
*sk
)
3251 const struct tcp_sock
*tp
= tcp_sk(sk
);
3252 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3254 /* Was it a usable window open? */
3256 if (!after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
, tcp_wnd_end(tp
))) {
3257 icsk
->icsk_backoff
= 0;
3258 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
3259 /* Socket must be waked up by subsequent tcp_data_snd_check().
3260 * This function is not for random using!
3263 unsigned long when
= tcp_probe0_when(sk
, TCP_RTO_MAX
);
3265 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3270 static inline bool tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
3272 return !(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
3273 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
;
3276 /* Decide wheather to run the increase function of congestion control. */
3277 static inline bool tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
3279 /* If reordering is high then always grow cwnd whenever data is
3280 * delivered regardless of its ordering. Otherwise stay conservative
3281 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3282 * new SACK or ECE mark may first advance cwnd here and later reduce
3283 * cwnd in tcp_fastretrans_alert() based on more states.
3285 if (tcp_sk(sk
)->reordering
> sock_net(sk
)->ipv4
.sysctl_tcp_reordering
)
3286 return flag
& FLAG_FORWARD_PROGRESS
;
3288 return flag
& FLAG_DATA_ACKED
;
3291 /* The "ultimate" congestion control function that aims to replace the rigid
3292 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3293 * It's called toward the end of processing an ACK with precise rate
3294 * information. All transmission or retransmission are delayed afterwards.
3296 static void tcp_cong_control(struct sock
*sk
, u32 ack
, u32 acked_sacked
,
3297 int flag
, const struct rate_sample
*rs
)
3299 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3301 if (icsk
->icsk_ca_ops
->cong_control
) {
3302 icsk
->icsk_ca_ops
->cong_control(sk
, rs
);
3306 if (tcp_in_cwnd_reduction(sk
)) {
3307 /* Reduce cwnd if state mandates */
3308 tcp_cwnd_reduction(sk
, acked_sacked
, flag
);
3309 } else if (tcp_may_raise_cwnd(sk
, flag
)) {
3310 /* Advance cwnd if state allows */
3311 tcp_cong_avoid(sk
, ack
, acked_sacked
);
3313 tcp_update_pacing_rate(sk
);
3316 /* Check that window update is acceptable.
3317 * The function assumes that snd_una<=ack<=snd_next.
3319 static inline bool tcp_may_update_window(const struct tcp_sock
*tp
,
3320 const u32 ack
, const u32 ack_seq
,
3323 return after(ack
, tp
->snd_una
) ||
3324 after(ack_seq
, tp
->snd_wl1
) ||
3325 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
);
3328 /* If we update tp->snd_una, also update tp->bytes_acked */
3329 static void tcp_snd_una_update(struct tcp_sock
*tp
, u32 ack
)
3331 u32 delta
= ack
- tp
->snd_una
;
3333 sock_owned_by_me((struct sock
*)tp
);
3334 tp
->bytes_acked
+= delta
;
3338 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3339 static void tcp_rcv_nxt_update(struct tcp_sock
*tp
, u32 seq
)
3341 u32 delta
= seq
- tp
->rcv_nxt
;
3343 sock_owned_by_me((struct sock
*)tp
);
3344 tp
->bytes_received
+= delta
;
3348 /* Update our send window.
3350 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3351 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3353 static int tcp_ack_update_window(struct sock
*sk
, const struct sk_buff
*skb
, u32 ack
,
3356 struct tcp_sock
*tp
= tcp_sk(sk
);
3358 u32 nwin
= ntohs(tcp_hdr(skb
)->window
);
3360 if (likely(!tcp_hdr(skb
)->syn
))
3361 nwin
<<= tp
->rx_opt
.snd_wscale
;
3363 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
3364 flag
|= FLAG_WIN_UPDATE
;
3365 tcp_update_wl(tp
, ack_seq
);
3367 if (tp
->snd_wnd
!= nwin
) {
3370 /* Note, it is the only place, where
3371 * fast path is recovered for sending TCP.
3374 tcp_fast_path_check(sk
);
3376 if (tcp_send_head(sk
))
3377 tcp_slow_start_after_idle_check(sk
);
3379 if (nwin
> tp
->max_window
) {
3380 tp
->max_window
= nwin
;
3381 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
3386 tcp_snd_una_update(tp
, ack
);
3391 static bool __tcp_oow_rate_limited(struct net
*net
, int mib_idx
,
3392 u32
*last_oow_ack_time
)
3394 if (*last_oow_ack_time
) {
3395 s32 elapsed
= (s32
)(tcp_jiffies32
- *last_oow_ack_time
);
3397 if (0 <= elapsed
&& elapsed
< sysctl_tcp_invalid_ratelimit
) {
3398 NET_INC_STATS(net
, mib_idx
);
3399 return true; /* rate-limited: don't send yet! */
3403 *last_oow_ack_time
= tcp_jiffies32
;
3405 return false; /* not rate-limited: go ahead, send dupack now! */
3408 /* Return true if we're currently rate-limiting out-of-window ACKs and
3409 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3410 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3411 * attacks that send repeated SYNs or ACKs for the same connection. To
3412 * do this, we do not send a duplicate SYNACK or ACK if the remote
3413 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3415 bool tcp_oow_rate_limited(struct net
*net
, const struct sk_buff
*skb
,
3416 int mib_idx
, u32
*last_oow_ack_time
)
3418 /* Data packets without SYNs are not likely part of an ACK loop. */
3419 if ((TCP_SKB_CB(skb
)->seq
!= TCP_SKB_CB(skb
)->end_seq
) &&
3423 return __tcp_oow_rate_limited(net
, mib_idx
, last_oow_ack_time
);
3426 /* RFC 5961 7 [ACK Throttling] */
3427 static void tcp_send_challenge_ack(struct sock
*sk
, const struct sk_buff
*skb
)
3429 /* unprotected vars, we dont care of overwrites */
3430 static u32 challenge_timestamp
;
3431 static unsigned int challenge_count
;
3432 struct tcp_sock
*tp
= tcp_sk(sk
);
3435 /* First check our per-socket dupack rate limit. */
3436 if (__tcp_oow_rate_limited(sock_net(sk
),
3437 LINUX_MIB_TCPACKSKIPPEDCHALLENGE
,
3438 &tp
->last_oow_ack_time
))
3441 /* Then check host-wide RFC 5961 rate limit. */
3443 if (now
!= challenge_timestamp
) {
3444 u32 half
= (sysctl_tcp_challenge_ack_limit
+ 1) >> 1;
3446 challenge_timestamp
= now
;
3447 WRITE_ONCE(challenge_count
, half
+
3448 prandom_u32_max(sysctl_tcp_challenge_ack_limit
));
3450 count
= READ_ONCE(challenge_count
);
3452 WRITE_ONCE(challenge_count
, count
- 1);
3453 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPCHALLENGEACK
);
3458 static void tcp_store_ts_recent(struct tcp_sock
*tp
)
3460 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
3461 tp
->rx_opt
.ts_recent_stamp
= get_seconds();
3464 static void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
3466 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
3467 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3468 * extra check below makes sure this can only happen
3469 * for pure ACK frames. -DaveM
3471 * Not only, also it occurs for expired timestamps.
3474 if (tcp_paws_check(&tp
->rx_opt
, 0))
3475 tcp_store_ts_recent(tp
);
3479 /* This routine deals with acks during a TLP episode.
3480 * We mark the end of a TLP episode on receiving TLP dupack or when
3481 * ack is after tlp_high_seq.
3482 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3484 static void tcp_process_tlp_ack(struct sock
*sk
, u32 ack
, int flag
)
3486 struct tcp_sock
*tp
= tcp_sk(sk
);
3488 if (before(ack
, tp
->tlp_high_seq
))
3491 if (flag
& FLAG_DSACKING_ACK
) {
3492 /* This DSACK means original and TLP probe arrived; no loss */
3493 tp
->tlp_high_seq
= 0;
3494 } else if (after(ack
, tp
->tlp_high_seq
)) {
3495 /* ACK advances: there was a loss, so reduce cwnd. Reset
3496 * tlp_high_seq in tcp_init_cwnd_reduction()
3498 tcp_init_cwnd_reduction(sk
);
3499 tcp_set_ca_state(sk
, TCP_CA_CWR
);
3500 tcp_end_cwnd_reduction(sk
);
3501 tcp_try_keep_open(sk
);
3502 NET_INC_STATS(sock_net(sk
),
3503 LINUX_MIB_TCPLOSSPROBERECOVERY
);
3504 } else if (!(flag
& (FLAG_SND_UNA_ADVANCED
|
3505 FLAG_NOT_DUP
| FLAG_DATA_SACKED
))) {
3506 /* Pure dupack: original and TLP probe arrived; no loss */
3507 tp
->tlp_high_seq
= 0;
3511 static inline void tcp_in_ack_event(struct sock
*sk
, u32 flags
)
3513 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3515 if (icsk
->icsk_ca_ops
->in_ack_event
)
3516 icsk
->icsk_ca_ops
->in_ack_event(sk
, flags
);
3519 /* Congestion control has updated the cwnd already. So if we're in
3520 * loss recovery then now we do any new sends (for FRTO) or
3521 * retransmits (for CA_Loss or CA_recovery) that make sense.
3523 static void tcp_xmit_recovery(struct sock
*sk
, int rexmit
)
3525 struct tcp_sock
*tp
= tcp_sk(sk
);
3527 if (rexmit
== REXMIT_NONE
)
3530 if (unlikely(rexmit
== 2)) {
3531 __tcp_push_pending_frames(sk
, tcp_current_mss(sk
),
3533 if (after(tp
->snd_nxt
, tp
->high_seq
))
3537 tcp_xmit_retransmit_queue(sk
);
3540 /* This routine deals with incoming acks, but not outgoing ones. */
3541 static int tcp_ack(struct sock
*sk
, const struct sk_buff
*skb
, int flag
)
3543 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3544 struct tcp_sock
*tp
= tcp_sk(sk
);
3545 struct tcp_sacktag_state sack_state
;
3546 struct rate_sample rs
= { .prior_delivered
= 0 };
3547 u32 prior_snd_una
= tp
->snd_una
;
3548 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
3549 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3550 bool is_dupack
= false;
3552 int prior_packets
= tp
->packets_out
;
3553 u32 delivered
= tp
->delivered
;
3554 u32 lost
= tp
->lost
;
3555 int acked
= 0; /* Number of packets newly acked */
3556 int rexmit
= REXMIT_NONE
; /* Flag to (re)transmit to recover losses */
3558 sack_state
.first_sackt
= 0;
3559 sack_state
.rate
= &rs
;
3561 /* We very likely will need to access write queue head. */
3562 prefetchw(sk
->sk_write_queue
.next
);
3564 /* If the ack is older than previous acks
3565 * then we can probably ignore it.
3567 if (before(ack
, prior_snd_una
)) {
3568 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3569 if (before(ack
, prior_snd_una
- tp
->max_window
)) {
3570 if (!(flag
& FLAG_NO_CHALLENGE_ACK
))
3571 tcp_send_challenge_ack(sk
, skb
);
3577 /* If the ack includes data we haven't sent yet, discard
3578 * this segment (RFC793 Section 3.9).
3580 if (after(ack
, tp
->snd_nxt
))
3583 if (icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
)
3586 if (after(ack
, prior_snd_una
)) {
3587 flag
|= FLAG_SND_UNA_ADVANCED
;
3588 icsk
->icsk_retransmits
= 0;
3591 prior_fackets
= tp
->fackets_out
;
3592 rs
.prior_in_flight
= tcp_packets_in_flight(tp
);
3594 /* ts_recent update must be made after we are sure that the packet
3597 if (flag
& FLAG_UPDATE_TS_RECENT
)
3598 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
3600 if (!(flag
& FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
3601 /* Window is constant, pure forward advance.
3602 * No more checks are required.
3603 * Note, we use the fact that SND.UNA>=SND.WL2.
3605 tcp_update_wl(tp
, ack_seq
);
3606 tcp_snd_una_update(tp
, ack
);
3607 flag
|= FLAG_WIN_UPDATE
;
3609 tcp_in_ack_event(sk
, CA_ACK_WIN_UPDATE
);
3611 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPHPACKS
);
3613 u32 ack_ev_flags
= CA_ACK_SLOWPATH
;
3615 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
3618 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPPUREACKS
);
3620 flag
|= tcp_ack_update_window(sk
, skb
, ack
, ack_seq
);
3622 if (TCP_SKB_CB(skb
)->sacked
)
3623 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
,
3626 if (tcp_ecn_rcv_ecn_echo(tp
, tcp_hdr(skb
))) {
3628 ack_ev_flags
|= CA_ACK_ECE
;
3631 if (flag
& FLAG_WIN_UPDATE
)
3632 ack_ev_flags
|= CA_ACK_WIN_UPDATE
;
3634 tcp_in_ack_event(sk
, ack_ev_flags
);
3637 /* We passed data and got it acked, remove any soft error
3638 * log. Something worked...
3640 sk
->sk_err_soft
= 0;
3641 icsk
->icsk_probes_out
= 0;
3642 tp
->rcv_tstamp
= tcp_jiffies32
;
3646 /* See if we can take anything off of the retransmit queue. */
3647 flag
|= tcp_clean_rtx_queue(sk
, prior_fackets
, prior_snd_una
, &acked
,
3650 if (tcp_ack_is_dubious(sk
, flag
)) {
3651 is_dupack
= !(flag
& (FLAG_SND_UNA_ADVANCED
| FLAG_NOT_DUP
));
3652 tcp_fastretrans_alert(sk
, acked
, is_dupack
, &flag
, &rexmit
);
3654 if (tp
->tlp_high_seq
)
3655 tcp_process_tlp_ack(sk
, ack
, flag
);
3657 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
& FLAG_NOT_DUP
))
3660 if (icsk
->icsk_pending
== ICSK_TIME_RETRANS
)
3661 tcp_schedule_loss_probe(sk
);
3662 delivered
= tp
->delivered
- delivered
; /* freshly ACKed or SACKed */
3663 lost
= tp
->lost
- lost
; /* freshly marked lost */
3664 tcp_rate_gen(sk
, delivered
, lost
, sack_state
.rate
);
3665 tcp_cong_control(sk
, ack
, delivered
, flag
, sack_state
.rate
);
3666 tcp_xmit_recovery(sk
, rexmit
);
3670 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3671 if (flag
& FLAG_DSACKING_ACK
)
3672 tcp_fastretrans_alert(sk
, acked
, is_dupack
, &flag
, &rexmit
);
3673 /* If this ack opens up a zero window, clear backoff. It was
3674 * being used to time the probes, and is probably far higher than
3675 * it needs to be for normal retransmission.
3677 if (tcp_send_head(sk
))
3680 if (tp
->tlp_high_seq
)
3681 tcp_process_tlp_ack(sk
, ack
, flag
);
3685 SOCK_DEBUG(sk
, "Ack %u after %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3689 /* If data was SACKed, tag it and see if we should send more data.
3690 * If data was DSACKed, see if we can undo a cwnd reduction.
3692 if (TCP_SKB_CB(skb
)->sacked
) {
3693 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
,
3695 tcp_fastretrans_alert(sk
, acked
, is_dupack
, &flag
, &rexmit
);
3696 tcp_xmit_recovery(sk
, rexmit
);
3699 SOCK_DEBUG(sk
, "Ack %u before %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3703 static void tcp_parse_fastopen_option(int len
, const unsigned char *cookie
,
3704 bool syn
, struct tcp_fastopen_cookie
*foc
,
3707 /* Valid only in SYN or SYN-ACK with an even length. */
3708 if (!foc
|| !syn
|| len
< 0 || (len
& 1))
3711 if (len
>= TCP_FASTOPEN_COOKIE_MIN
&&
3712 len
<= TCP_FASTOPEN_COOKIE_MAX
)
3713 memcpy(foc
->val
, cookie
, len
);
3720 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3721 * But, this can also be called on packets in the established flow when
3722 * the fast version below fails.
3724 void tcp_parse_options(const struct net
*net
,
3725 const struct sk_buff
*skb
,
3726 struct tcp_options_received
*opt_rx
, int estab
,
3727 struct tcp_fastopen_cookie
*foc
)
3729 const unsigned char *ptr
;
3730 const struct tcphdr
*th
= tcp_hdr(skb
);
3731 int length
= (th
->doff
* 4) - sizeof(struct tcphdr
);
3733 ptr
= (const unsigned char *)(th
+ 1);
3734 opt_rx
->saw_tstamp
= 0;
3736 while (length
> 0) {
3737 int opcode
= *ptr
++;
3743 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
3748 if (opsize
< 2) /* "silly options" */
3750 if (opsize
> length
)
3751 return; /* don't parse partial options */
3754 if (opsize
== TCPOLEN_MSS
&& th
->syn
&& !estab
) {
3755 u16 in_mss
= get_unaligned_be16(ptr
);
3757 if (opt_rx
->user_mss
&&
3758 opt_rx
->user_mss
< in_mss
)
3759 in_mss
= opt_rx
->user_mss
;
3760 opt_rx
->mss_clamp
= in_mss
;
3765 if (opsize
== TCPOLEN_WINDOW
&& th
->syn
&&
3766 !estab
&& net
->ipv4
.sysctl_tcp_window_scaling
) {
3767 __u8 snd_wscale
= *(__u8
*)ptr
;
3768 opt_rx
->wscale_ok
= 1;
3769 if (snd_wscale
> TCP_MAX_WSCALE
) {
3770 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n",
3774 snd_wscale
= TCP_MAX_WSCALE
;
3776 opt_rx
->snd_wscale
= snd_wscale
;
3779 case TCPOPT_TIMESTAMP
:
3780 if ((opsize
== TCPOLEN_TIMESTAMP
) &&
3781 ((estab
&& opt_rx
->tstamp_ok
) ||
3782 (!estab
&& net
->ipv4
.sysctl_tcp_timestamps
))) {
3783 opt_rx
->saw_tstamp
= 1;
3784 opt_rx
->rcv_tsval
= get_unaligned_be32(ptr
);
3785 opt_rx
->rcv_tsecr
= get_unaligned_be32(ptr
+ 4);
3788 case TCPOPT_SACK_PERM
:
3789 if (opsize
== TCPOLEN_SACK_PERM
&& th
->syn
&&
3790 !estab
&& net
->ipv4
.sysctl_tcp_sack
) {
3791 opt_rx
->sack_ok
= TCP_SACK_SEEN
;
3792 tcp_sack_reset(opt_rx
);
3797 if ((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
3798 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
3800 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
3803 #ifdef CONFIG_TCP_MD5SIG
3806 * The MD5 Hash has already been
3807 * checked (see tcp_v{4,6}_do_rcv()).
3811 case TCPOPT_FASTOPEN
:
3812 tcp_parse_fastopen_option(
3813 opsize
- TCPOLEN_FASTOPEN_BASE
,
3814 ptr
, th
->syn
, foc
, false);
3818 /* Fast Open option shares code 254 using a
3819 * 16 bits magic number.
3821 if (opsize
>= TCPOLEN_EXP_FASTOPEN_BASE
&&
3822 get_unaligned_be16(ptr
) ==
3823 TCPOPT_FASTOPEN_MAGIC
)
3824 tcp_parse_fastopen_option(opsize
-
3825 TCPOLEN_EXP_FASTOPEN_BASE
,
3826 ptr
+ 2, th
->syn
, foc
, true);
3835 EXPORT_SYMBOL(tcp_parse_options
);
3837 static bool tcp_parse_aligned_timestamp(struct tcp_sock
*tp
, const struct tcphdr
*th
)
3839 const __be32
*ptr
= (const __be32
*)(th
+ 1);
3841 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3842 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
3843 tp
->rx_opt
.saw_tstamp
= 1;
3845 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3848 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
) - tp
->tsoffset
;
3850 tp
->rx_opt
.rcv_tsecr
= 0;
3856 /* Fast parse options. This hopes to only see timestamps.
3857 * If it is wrong it falls back on tcp_parse_options().
3859 static bool tcp_fast_parse_options(const struct net
*net
,
3860 const struct sk_buff
*skb
,
3861 const struct tcphdr
*th
, struct tcp_sock
*tp
)
3863 /* In the spirit of fast parsing, compare doff directly to constant
3864 * values. Because equality is used, short doff can be ignored here.
3866 if (th
->doff
== (sizeof(*th
) / 4)) {
3867 tp
->rx_opt
.saw_tstamp
= 0;
3869 } else if (tp
->rx_opt
.tstamp_ok
&&
3870 th
->doff
== ((sizeof(*th
) + TCPOLEN_TSTAMP_ALIGNED
) / 4)) {
3871 if (tcp_parse_aligned_timestamp(tp
, th
))
3875 tcp_parse_options(net
, skb
, &tp
->rx_opt
, 1, NULL
);
3876 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
3877 tp
->rx_opt
.rcv_tsecr
-= tp
->tsoffset
;
3882 #ifdef CONFIG_TCP_MD5SIG
3884 * Parse MD5 Signature option
3886 const u8
*tcp_parse_md5sig_option(const struct tcphdr
*th
)
3888 int length
= (th
->doff
<< 2) - sizeof(*th
);
3889 const u8
*ptr
= (const u8
*)(th
+ 1);
3891 /* If the TCP option is too short, we can short cut */
3892 if (length
< TCPOLEN_MD5SIG
)
3895 while (length
> 0) {
3896 int opcode
= *ptr
++;
3907 if (opsize
< 2 || opsize
> length
)
3909 if (opcode
== TCPOPT_MD5SIG
)
3910 return opsize
== TCPOLEN_MD5SIG
? ptr
: NULL
;
3917 EXPORT_SYMBOL(tcp_parse_md5sig_option
);
3920 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3922 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3923 * it can pass through stack. So, the following predicate verifies that
3924 * this segment is not used for anything but congestion avoidance or
3925 * fast retransmit. Moreover, we even are able to eliminate most of such
3926 * second order effects, if we apply some small "replay" window (~RTO)
3927 * to timestamp space.
3929 * All these measures still do not guarantee that we reject wrapped ACKs
3930 * on networks with high bandwidth, when sequence space is recycled fastly,
3931 * but it guarantees that such events will be very rare and do not affect
3932 * connection seriously. This doesn't look nice, but alas, PAWS is really
3935 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3936 * states that events when retransmit arrives after original data are rare.
3937 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3938 * the biggest problem on large power networks even with minor reordering.
3939 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3940 * up to bandwidth of 18Gigabit/sec. 8) ]
3943 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
3945 const struct tcp_sock
*tp
= tcp_sk(sk
);
3946 const struct tcphdr
*th
= tcp_hdr(skb
);
3947 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3948 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3950 return (/* 1. Pure ACK with correct sequence number. */
3951 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
3953 /* 2. ... and duplicate ACK. */
3954 ack
== tp
->snd_una
&&
3956 /* 3. ... and does not update window. */
3957 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
3959 /* 4. ... and sits in replay window. */
3960 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
3963 static inline bool tcp_paws_discard(const struct sock
*sk
,
3964 const struct sk_buff
*skb
)
3966 const struct tcp_sock
*tp
= tcp_sk(sk
);
3968 return !tcp_paws_check(&tp
->rx_opt
, TCP_PAWS_WINDOW
) &&
3969 !tcp_disordered_ack(sk
, skb
);
3972 /* Check segment sequence number for validity.
3974 * Segment controls are considered valid, if the segment
3975 * fits to the window after truncation to the window. Acceptability
3976 * of data (and SYN, FIN, of course) is checked separately.
3977 * See tcp_data_queue(), for example.
3979 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3980 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3981 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3982 * (borrowed from freebsd)
3985 static inline bool tcp_sequence(const struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3987 return !before(end_seq
, tp
->rcv_wup
) &&
3988 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
3991 /* When we get a reset we do this. */
3992 void tcp_reset(struct sock
*sk
)
3994 /* We want the right error as BSD sees it (and indeed as we do). */
3995 switch (sk
->sk_state
) {
3997 sk
->sk_err
= ECONNREFUSED
;
3999 case TCP_CLOSE_WAIT
:
4005 sk
->sk_err
= ECONNRESET
;
4007 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4012 if (!sock_flag(sk
, SOCK_DEAD
))
4013 sk
->sk_error_report(sk
);
4017 * Process the FIN bit. This now behaves as it is supposed to work
4018 * and the FIN takes effect when it is validly part of sequence
4019 * space. Not before when we get holes.
4021 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4022 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4025 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4026 * close and we go into CLOSING (and later onto TIME-WAIT)
4028 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4030 void tcp_fin(struct sock
*sk
)
4032 struct tcp_sock
*tp
= tcp_sk(sk
);
4034 inet_csk_schedule_ack(sk
);
4036 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
4037 sock_set_flag(sk
, SOCK_DONE
);
4039 switch (sk
->sk_state
) {
4041 case TCP_ESTABLISHED
:
4042 /* Move to CLOSE_WAIT */
4043 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
4044 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
4047 case TCP_CLOSE_WAIT
:
4049 /* Received a retransmission of the FIN, do
4054 /* RFC793: Remain in the LAST-ACK state. */
4058 /* This case occurs when a simultaneous close
4059 * happens, we must ack the received FIN and
4060 * enter the CLOSING state.
4063 tcp_set_state(sk
, TCP_CLOSING
);
4066 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4068 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
4071 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4072 * cases we should never reach this piece of code.
4074 pr_err("%s: Impossible, sk->sk_state=%d\n",
4075 __func__
, sk
->sk_state
);
4079 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4080 * Probably, we should reset in this case. For now drop them.
4082 skb_rbtree_purge(&tp
->out_of_order_queue
);
4083 if (tcp_is_sack(tp
))
4084 tcp_sack_reset(&tp
->rx_opt
);
4087 if (!sock_flag(sk
, SOCK_DEAD
)) {
4088 sk
->sk_state_change(sk
);
4090 /* Do not send POLL_HUP for half duplex close. */
4091 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
4092 sk
->sk_state
== TCP_CLOSE
)
4093 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_HUP
);
4095 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_IN
);
4099 static inline bool tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
,
4102 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
4103 if (before(seq
, sp
->start_seq
))
4104 sp
->start_seq
= seq
;
4105 if (after(end_seq
, sp
->end_seq
))
4106 sp
->end_seq
= end_seq
;
4112 static void tcp_dsack_set(struct sock
*sk
, u32 seq
, u32 end_seq
)
4114 struct tcp_sock
*tp
= tcp_sk(sk
);
4116 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
4119 if (before(seq
, tp
->rcv_nxt
))
4120 mib_idx
= LINUX_MIB_TCPDSACKOLDSENT
;
4122 mib_idx
= LINUX_MIB_TCPDSACKOFOSENT
;
4124 NET_INC_STATS(sock_net(sk
), mib_idx
);
4126 tp
->rx_opt
.dsack
= 1;
4127 tp
->duplicate_sack
[0].start_seq
= seq
;
4128 tp
->duplicate_sack
[0].end_seq
= end_seq
;
4132 static void tcp_dsack_extend(struct sock
*sk
, u32 seq
, u32 end_seq
)
4134 struct tcp_sock
*tp
= tcp_sk(sk
);
4136 if (!tp
->rx_opt
.dsack
)
4137 tcp_dsack_set(sk
, seq
, end_seq
);
4139 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
4142 static void tcp_send_dupack(struct sock
*sk
, const struct sk_buff
*skb
)
4144 struct tcp_sock
*tp
= tcp_sk(sk
);
4146 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
4147 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4148 NET_INC_STATS(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4149 tcp_enter_quickack_mode(sk
);
4151 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
4152 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4154 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
4155 end_seq
= tp
->rcv_nxt
;
4156 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, end_seq
);
4163 /* These routines update the SACK block as out-of-order packets arrive or
4164 * in-order packets close up the sequence space.
4166 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
4169 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4170 struct tcp_sack_block
*swalk
= sp
+ 1;
4172 /* See if the recent change to the first SACK eats into
4173 * or hits the sequence space of other SACK blocks, if so coalesce.
4175 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
;) {
4176 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
4179 /* Zap SWALK, by moving every further SACK up by one slot.
4180 * Decrease num_sacks.
4182 tp
->rx_opt
.num_sacks
--;
4183 for (i
= this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
4187 this_sack
++, swalk
++;
4191 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
4193 struct tcp_sock
*tp
= tcp_sk(sk
);
4194 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4195 int cur_sacks
= tp
->rx_opt
.num_sacks
;
4201 for (this_sack
= 0; this_sack
< cur_sacks
; this_sack
++, sp
++) {
4202 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
4203 /* Rotate this_sack to the first one. */
4204 for (; this_sack
> 0; this_sack
--, sp
--)
4205 swap(*sp
, *(sp
- 1));
4207 tcp_sack_maybe_coalesce(tp
);
4212 /* Could not find an adjacent existing SACK, build a new one,
4213 * put it at the front, and shift everyone else down. We
4214 * always know there is at least one SACK present already here.
4216 * If the sack array is full, forget about the last one.
4218 if (this_sack
>= TCP_NUM_SACKS
) {
4220 tp
->rx_opt
.num_sacks
--;
4223 for (; this_sack
> 0; this_sack
--, sp
--)
4227 /* Build the new head SACK, and we're done. */
4228 sp
->start_seq
= seq
;
4229 sp
->end_seq
= end_seq
;
4230 tp
->rx_opt
.num_sacks
++;
4233 /* RCV.NXT advances, some SACKs should be eaten. */
4235 static void tcp_sack_remove(struct tcp_sock
*tp
)
4237 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4238 int num_sacks
= tp
->rx_opt
.num_sacks
;
4241 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4242 if (RB_EMPTY_ROOT(&tp
->out_of_order_queue
)) {
4243 tp
->rx_opt
.num_sacks
= 0;
4247 for (this_sack
= 0; this_sack
< num_sacks
;) {
4248 /* Check if the start of the sack is covered by RCV.NXT. */
4249 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
4252 /* RCV.NXT must cover all the block! */
4253 WARN_ON(before(tp
->rcv_nxt
, sp
->end_seq
));
4255 /* Zap this SACK, by moving forward any other SACKS. */
4256 for (i
= this_sack
+1; i
< num_sacks
; i
++)
4257 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
4264 tp
->rx_opt
.num_sacks
= num_sacks
;
4268 * tcp_try_coalesce - try to merge skb to prior one
4271 * @from: buffer to add in queue
4272 * @fragstolen: pointer to boolean
4274 * Before queueing skb @from after @to, try to merge them
4275 * to reduce overall memory use and queue lengths, if cost is small.
4276 * Packets in ofo or receive queues can stay a long time.
4277 * Better try to coalesce them right now to avoid future collapses.
4278 * Returns true if caller should free @from instead of queueing it
4280 static bool tcp_try_coalesce(struct sock
*sk
,
4282 struct sk_buff
*from
,
4287 *fragstolen
= false;
4289 /* Its possible this segment overlaps with prior segment in queue */
4290 if (TCP_SKB_CB(from
)->seq
!= TCP_SKB_CB(to
)->end_seq
)
4293 if (!skb_try_coalesce(to
, from
, fragstolen
, &delta
))
4296 atomic_add(delta
, &sk
->sk_rmem_alloc
);
4297 sk_mem_charge(sk
, delta
);
4298 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPRCVCOALESCE
);
4299 TCP_SKB_CB(to
)->end_seq
= TCP_SKB_CB(from
)->end_seq
;
4300 TCP_SKB_CB(to
)->ack_seq
= TCP_SKB_CB(from
)->ack_seq
;
4301 TCP_SKB_CB(to
)->tcp_flags
|= TCP_SKB_CB(from
)->tcp_flags
;
4305 static void tcp_drop(struct sock
*sk
, struct sk_buff
*skb
)
4307 sk_drops_add(sk
, skb
);
4311 /* This one checks to see if we can put data from the
4312 * out_of_order queue into the receive_queue.
4314 static void tcp_ofo_queue(struct sock
*sk
)
4316 struct tcp_sock
*tp
= tcp_sk(sk
);
4317 __u32 dsack_high
= tp
->rcv_nxt
;
4318 bool fin
, fragstolen
, eaten
;
4319 struct sk_buff
*skb
, *tail
;
4322 p
= rb_first(&tp
->out_of_order_queue
);
4324 skb
= rb_entry(p
, struct sk_buff
, rbnode
);
4325 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
4328 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
4329 __u32 dsack
= dsack_high
;
4330 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
4331 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
4332 tcp_dsack_extend(sk
, TCP_SKB_CB(skb
)->seq
, dsack
);
4335 rb_erase(&skb
->rbnode
, &tp
->out_of_order_queue
);
4337 if (unlikely(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))) {
4338 SOCK_DEBUG(sk
, "ofo packet was already received\n");
4342 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
4343 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4344 TCP_SKB_CB(skb
)->end_seq
);
4346 tail
= skb_peek_tail(&sk
->sk_receive_queue
);
4347 eaten
= tail
&& tcp_try_coalesce(sk
, tail
, skb
, &fragstolen
);
4348 tcp_rcv_nxt_update(tp
, TCP_SKB_CB(skb
)->end_seq
);
4349 fin
= TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
;
4351 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4353 kfree_skb_partial(skb
, fragstolen
);
4355 if (unlikely(fin
)) {
4357 /* tcp_fin() purges tp->out_of_order_queue,
4358 * so we must end this loop right now.
4365 static bool tcp_prune_ofo_queue(struct sock
*sk
);
4366 static int tcp_prune_queue(struct sock
*sk
);
4368 static int tcp_try_rmem_schedule(struct sock
*sk
, struct sk_buff
*skb
,
4371 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
4372 !sk_rmem_schedule(sk
, skb
, size
)) {
4374 if (tcp_prune_queue(sk
) < 0)
4377 while (!sk_rmem_schedule(sk
, skb
, size
)) {
4378 if (!tcp_prune_ofo_queue(sk
))
4385 static void tcp_data_queue_ofo(struct sock
*sk
, struct sk_buff
*skb
)
4387 struct tcp_sock
*tp
= tcp_sk(sk
);
4388 struct rb_node
**p
, *q
, *parent
;
4389 struct sk_buff
*skb1
;
4393 tcp_ecn_check_ce(tp
, skb
);
4395 if (unlikely(tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))) {
4396 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPOFODROP
);
4401 /* Disable header prediction. */
4403 inet_csk_schedule_ack(sk
);
4405 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPOFOQUEUE
);
4406 seq
= TCP_SKB_CB(skb
)->seq
;
4407 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4408 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
4409 tp
->rcv_nxt
, seq
, end_seq
);
4411 p
= &tp
->out_of_order_queue
.rb_node
;
4412 if (RB_EMPTY_ROOT(&tp
->out_of_order_queue
)) {
4413 /* Initial out of order segment, build 1 SACK. */
4414 if (tcp_is_sack(tp
)) {
4415 tp
->rx_opt
.num_sacks
= 1;
4416 tp
->selective_acks
[0].start_seq
= seq
;
4417 tp
->selective_acks
[0].end_seq
= end_seq
;
4419 rb_link_node(&skb
->rbnode
, NULL
, p
);
4420 rb_insert_color(&skb
->rbnode
, &tp
->out_of_order_queue
);
4421 tp
->ooo_last_skb
= skb
;
4425 /* In the typical case, we are adding an skb to the end of the list.
4426 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4428 if (tcp_try_coalesce(sk
, tp
->ooo_last_skb
, skb
, &fragstolen
)) {
4430 tcp_grow_window(sk
, skb
);
4431 kfree_skb_partial(skb
, fragstolen
);
4435 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4436 if (!before(seq
, TCP_SKB_CB(tp
->ooo_last_skb
)->end_seq
)) {
4437 parent
= &tp
->ooo_last_skb
->rbnode
;
4438 p
= &parent
->rb_right
;
4442 /* Find place to insert this segment. Handle overlaps on the way. */
4446 skb1
= rb_entry(parent
, struct sk_buff
, rbnode
);
4447 if (before(seq
, TCP_SKB_CB(skb1
)->seq
)) {
4448 p
= &parent
->rb_left
;
4451 if (before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4452 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4453 /* All the bits are present. Drop. */
4454 NET_INC_STATS(sock_net(sk
),
4455 LINUX_MIB_TCPOFOMERGE
);
4458 tcp_dsack_set(sk
, seq
, end_seq
);
4461 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
4462 /* Partial overlap. */
4463 tcp_dsack_set(sk
, seq
, TCP_SKB_CB(skb1
)->end_seq
);
4465 /* skb's seq == skb1's seq and skb covers skb1.
4466 * Replace skb1 with skb.
4468 rb_replace_node(&skb1
->rbnode
, &skb
->rbnode
,
4469 &tp
->out_of_order_queue
);
4470 tcp_dsack_extend(sk
,
4471 TCP_SKB_CB(skb1
)->seq
,
4472 TCP_SKB_CB(skb1
)->end_seq
);
4473 NET_INC_STATS(sock_net(sk
),
4474 LINUX_MIB_TCPOFOMERGE
);
4478 } else if (tcp_try_coalesce(sk
, skb1
, skb
, &fragstolen
)) {
4481 p
= &parent
->rb_right
;
4484 /* Insert segment into RB tree. */
4485 rb_link_node(&skb
->rbnode
, parent
, p
);
4486 rb_insert_color(&skb
->rbnode
, &tp
->out_of_order_queue
);
4489 /* Remove other segments covered by skb. */
4490 while ((q
= rb_next(&skb
->rbnode
)) != NULL
) {
4491 skb1
= rb_entry(q
, struct sk_buff
, rbnode
);
4493 if (!after(end_seq
, TCP_SKB_CB(skb1
)->seq
))
4495 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4496 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4500 rb_erase(&skb1
->rbnode
, &tp
->out_of_order_queue
);
4501 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4502 TCP_SKB_CB(skb1
)->end_seq
);
4503 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPOFOMERGE
);
4506 /* If there is no skb after us, we are the last_skb ! */
4508 tp
->ooo_last_skb
= skb
;
4511 if (tcp_is_sack(tp
))
4512 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
4515 tcp_grow_window(sk
, skb
);
4517 skb_set_owner_r(skb
, sk
);
4521 static int __must_check
tcp_queue_rcv(struct sock
*sk
, struct sk_buff
*skb
, int hdrlen
,
4525 struct sk_buff
*tail
= skb_peek_tail(&sk
->sk_receive_queue
);
4527 __skb_pull(skb
, hdrlen
);
4529 tcp_try_coalesce(sk
, tail
, skb
, fragstolen
)) ? 1 : 0;
4530 tcp_rcv_nxt_update(tcp_sk(sk
), TCP_SKB_CB(skb
)->end_seq
);
4532 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4533 skb_set_owner_r(skb
, sk
);
4538 int tcp_send_rcvq(struct sock
*sk
, struct msghdr
*msg
, size_t size
)
4540 struct sk_buff
*skb
;
4548 if (size
> PAGE_SIZE
) {
4549 int npages
= min_t(size_t, size
>> PAGE_SHIFT
, MAX_SKB_FRAGS
);
4551 data_len
= npages
<< PAGE_SHIFT
;
4552 size
= data_len
+ (size
& ~PAGE_MASK
);
4554 skb
= alloc_skb_with_frags(size
- data_len
, data_len
,
4555 PAGE_ALLOC_COSTLY_ORDER
,
4556 &err
, sk
->sk_allocation
);
4560 skb_put(skb
, size
- data_len
);
4561 skb
->data_len
= data_len
;
4564 if (tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))
4567 err
= skb_copy_datagram_from_iter(skb
, 0, &msg
->msg_iter
, size
);
4571 TCP_SKB_CB(skb
)->seq
= tcp_sk(sk
)->rcv_nxt
;
4572 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ size
;
4573 TCP_SKB_CB(skb
)->ack_seq
= tcp_sk(sk
)->snd_una
- 1;
4575 if (tcp_queue_rcv(sk
, skb
, 0, &fragstolen
)) {
4576 WARN_ON_ONCE(fragstolen
); /* should not happen */
4588 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
4590 struct tcp_sock
*tp
= tcp_sk(sk
);
4591 bool fragstolen
= false;
4594 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
) {
4599 __skb_pull(skb
, tcp_hdr(skb
)->doff
* 4);
4601 tcp_ecn_accept_cwr(tp
, skb
);
4603 tp
->rx_opt
.dsack
= 0;
4605 /* Queue data for delivery to the user.
4606 * Packets in sequence go to the receive queue.
4607 * Out of sequence packets to the out_of_order_queue.
4609 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
4610 if (tcp_receive_window(tp
) == 0)
4613 /* Ok. In sequence. In window. */
4614 if (tp
->ucopy
.task
== current
&&
4615 tp
->copied_seq
== tp
->rcv_nxt
&& tp
->ucopy
.len
&&
4616 sock_owned_by_user(sk
) && !tp
->urg_data
) {
4617 int chunk
= min_t(unsigned int, skb
->len
,
4620 __set_current_state(TASK_RUNNING
);
4622 if (!skb_copy_datagram_msg(skb
, 0, tp
->ucopy
.msg
, chunk
)) {
4623 tp
->ucopy
.len
-= chunk
;
4624 tp
->copied_seq
+= chunk
;
4625 eaten
= (chunk
== skb
->len
);
4626 tcp_rcv_space_adjust(sk
);
4633 if (skb_queue_len(&sk
->sk_receive_queue
) == 0)
4634 sk_forced_mem_schedule(sk
, skb
->truesize
);
4635 else if (tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))
4638 eaten
= tcp_queue_rcv(sk
, skb
, 0, &fragstolen
);
4640 tcp_rcv_nxt_update(tp
, TCP_SKB_CB(skb
)->end_seq
);
4642 tcp_event_data_recv(sk
, skb
);
4643 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
4646 if (!RB_EMPTY_ROOT(&tp
->out_of_order_queue
)) {
4649 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4650 * gap in queue is filled.
4652 if (RB_EMPTY_ROOT(&tp
->out_of_order_queue
))
4653 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
4656 if (tp
->rx_opt
.num_sacks
)
4657 tcp_sack_remove(tp
);
4659 tcp_fast_path_check(sk
);
4662 kfree_skb_partial(skb
, fragstolen
);
4663 if (!sock_flag(sk
, SOCK_DEAD
))
4664 sk
->sk_data_ready(sk
);
4668 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4669 /* A retransmit, 2nd most common case. Force an immediate ack. */
4670 NET_INC_STATS(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4671 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4674 tcp_enter_quickack_mode(sk
);
4675 inet_csk_schedule_ack(sk
);
4681 /* Out of window. F.e. zero window probe. */
4682 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
4685 tcp_enter_quickack_mode(sk
);
4687 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4688 /* Partial packet, seq < rcv_next < end_seq */
4689 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
4690 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4691 TCP_SKB_CB(skb
)->end_seq
);
4693 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
4695 /* If window is closed, drop tail of packet. But after
4696 * remembering D-SACK for its head made in previous line.
4698 if (!tcp_receive_window(tp
))
4703 tcp_data_queue_ofo(sk
, skb
);
4706 static struct sk_buff
*tcp_skb_next(struct sk_buff
*skb
, struct sk_buff_head
*list
)
4709 return !skb_queue_is_last(list
, skb
) ? skb
->next
: NULL
;
4711 return rb_entry_safe(rb_next(&skb
->rbnode
), struct sk_buff
, rbnode
);
4714 static struct sk_buff
*tcp_collapse_one(struct sock
*sk
, struct sk_buff
*skb
,
4715 struct sk_buff_head
*list
,
4716 struct rb_root
*root
)
4718 struct sk_buff
*next
= tcp_skb_next(skb
, list
);
4721 __skb_unlink(skb
, list
);
4723 rb_erase(&skb
->rbnode
, root
);
4726 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPRCVCOLLAPSED
);
4731 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
4732 static void tcp_rbtree_insert(struct rb_root
*root
, struct sk_buff
*skb
)
4734 struct rb_node
**p
= &root
->rb_node
;
4735 struct rb_node
*parent
= NULL
;
4736 struct sk_buff
*skb1
;
4740 skb1
= rb_entry(parent
, struct sk_buff
, rbnode
);
4741 if (before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb1
)->seq
))
4742 p
= &parent
->rb_left
;
4744 p
= &parent
->rb_right
;
4746 rb_link_node(&skb
->rbnode
, parent
, p
);
4747 rb_insert_color(&skb
->rbnode
, root
);
4750 /* Collapse contiguous sequence of skbs head..tail with
4751 * sequence numbers start..end.
4753 * If tail is NULL, this means until the end of the queue.
4755 * Segments with FIN/SYN are not collapsed (only because this
4759 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
, struct rb_root
*root
,
4760 struct sk_buff
*head
, struct sk_buff
*tail
, u32 start
, u32 end
)
4762 struct sk_buff
*skb
= head
, *n
;
4763 struct sk_buff_head tmp
;
4766 /* First, check that queue is collapsible and find
4767 * the point where collapsing can be useful.
4770 for (end_of_skbs
= true; skb
!= NULL
&& skb
!= tail
; skb
= n
) {
4771 n
= tcp_skb_next(skb
, list
);
4773 /* No new bits? It is possible on ofo queue. */
4774 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4775 skb
= tcp_collapse_one(sk
, skb
, list
, root
);
4781 /* The first skb to collapse is:
4783 * - bloated or contains data before "start" or
4784 * overlaps to the next one.
4786 if (!(TCP_SKB_CB(skb
)->tcp_flags
& (TCPHDR_SYN
| TCPHDR_FIN
)) &&
4787 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
4788 before(TCP_SKB_CB(skb
)->seq
, start
))) {
4789 end_of_skbs
= false;
4793 if (n
&& n
!= tail
&&
4794 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(n
)->seq
) {
4795 end_of_skbs
= false;
4799 /* Decided to skip this, advance start seq. */
4800 start
= TCP_SKB_CB(skb
)->end_seq
;
4803 (TCP_SKB_CB(skb
)->tcp_flags
& (TCPHDR_SYN
| TCPHDR_FIN
)))
4806 __skb_queue_head_init(&tmp
);
4808 while (before(start
, end
)) {
4809 int copy
= min_t(int, SKB_MAX_ORDER(0, 0), end
- start
);
4810 struct sk_buff
*nskb
;
4812 nskb
= alloc_skb(copy
, GFP_ATOMIC
);
4816 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
4817 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
4819 __skb_queue_before(list
, skb
, nskb
);
4821 __skb_queue_tail(&tmp
, nskb
); /* defer rbtree insertion */
4822 skb_set_owner_r(nskb
, sk
);
4824 /* Copy data, releasing collapsed skbs. */
4826 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
4827 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
4831 size
= min(copy
, size
);
4832 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
4834 TCP_SKB_CB(nskb
)->end_seq
+= size
;
4838 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4839 skb
= tcp_collapse_one(sk
, skb
, list
, root
);
4842 (TCP_SKB_CB(skb
)->tcp_flags
& (TCPHDR_SYN
| TCPHDR_FIN
)))
4848 skb_queue_walk_safe(&tmp
, skb
, n
)
4849 tcp_rbtree_insert(root
, skb
);
4852 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4853 * and tcp_collapse() them until all the queue is collapsed.
4855 static void tcp_collapse_ofo_queue(struct sock
*sk
)
4857 struct tcp_sock
*tp
= tcp_sk(sk
);
4858 struct sk_buff
*skb
, *head
;
4862 p
= rb_first(&tp
->out_of_order_queue
);
4863 skb
= rb_entry_safe(p
, struct sk_buff
, rbnode
);
4866 p
= rb_last(&tp
->out_of_order_queue
);
4867 /* Note: This is possible p is NULL here. We do not
4868 * use rb_entry_safe(), as ooo_last_skb is valid only
4869 * if rbtree is not empty.
4871 tp
->ooo_last_skb
= rb_entry(p
, struct sk_buff
, rbnode
);
4874 start
= TCP_SKB_CB(skb
)->seq
;
4875 end
= TCP_SKB_CB(skb
)->end_seq
;
4877 for (head
= skb
;;) {
4878 skb
= tcp_skb_next(skb
, NULL
);
4880 /* Range is terminated when we see a gap or when
4881 * we are at the queue end.
4884 after(TCP_SKB_CB(skb
)->seq
, end
) ||
4885 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
4886 tcp_collapse(sk
, NULL
, &tp
->out_of_order_queue
,
4887 head
, skb
, start
, end
);
4891 if (unlikely(before(TCP_SKB_CB(skb
)->seq
, start
)))
4892 start
= TCP_SKB_CB(skb
)->seq
;
4893 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
4894 end
= TCP_SKB_CB(skb
)->end_seq
;
4899 * Clean the out-of-order queue to make room.
4900 * We drop high sequences packets to :
4901 * 1) Let a chance for holes to be filled.
4902 * 2) not add too big latencies if thousands of packets sit there.
4903 * (But if application shrinks SO_RCVBUF, we could still end up
4904 * freeing whole queue here)
4906 * Return true if queue has shrunk.
4908 static bool tcp_prune_ofo_queue(struct sock
*sk
)
4910 struct tcp_sock
*tp
= tcp_sk(sk
);
4911 struct rb_node
*node
, *prev
;
4913 if (RB_EMPTY_ROOT(&tp
->out_of_order_queue
))
4916 NET_INC_STATS(sock_net(sk
), LINUX_MIB_OFOPRUNED
);
4917 node
= &tp
->ooo_last_skb
->rbnode
;
4919 prev
= rb_prev(node
);
4920 rb_erase(node
, &tp
->out_of_order_queue
);
4921 tcp_drop(sk
, rb_entry(node
, struct sk_buff
, rbnode
));
4923 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
&&
4924 !tcp_under_memory_pressure(sk
))
4928 tp
->ooo_last_skb
= rb_entry(prev
, struct sk_buff
, rbnode
);
4930 /* Reset SACK state. A conforming SACK implementation will
4931 * do the same at a timeout based retransmit. When a connection
4932 * is in a sad state like this, we care only about integrity
4933 * of the connection not performance.
4935 if (tp
->rx_opt
.sack_ok
)
4936 tcp_sack_reset(&tp
->rx_opt
);
4940 /* Reduce allocated memory if we can, trying to get
4941 * the socket within its memory limits again.
4943 * Return less than zero if we should start dropping frames
4944 * until the socket owning process reads some of the data
4945 * to stabilize the situation.
4947 static int tcp_prune_queue(struct sock
*sk
)
4949 struct tcp_sock
*tp
= tcp_sk(sk
);
4951 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
4953 NET_INC_STATS(sock_net(sk
), LINUX_MIB_PRUNECALLED
);
4955 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
4956 tcp_clamp_window(sk
);
4957 else if (tcp_under_memory_pressure(sk
))
4958 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
4960 tcp_collapse_ofo_queue(sk
);
4961 if (!skb_queue_empty(&sk
->sk_receive_queue
))
4962 tcp_collapse(sk
, &sk
->sk_receive_queue
, NULL
,
4963 skb_peek(&sk
->sk_receive_queue
),
4965 tp
->copied_seq
, tp
->rcv_nxt
);
4968 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4971 /* Collapsing did not help, destructive actions follow.
4972 * This must not ever occur. */
4974 tcp_prune_ofo_queue(sk
);
4976 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4979 /* If we are really being abused, tell the caller to silently
4980 * drop receive data on the floor. It will get retransmitted
4981 * and hopefully then we'll have sufficient space.
4983 NET_INC_STATS(sock_net(sk
), LINUX_MIB_RCVPRUNED
);
4985 /* Massive buffer overcommit. */
4990 static bool tcp_should_expand_sndbuf(const struct sock
*sk
)
4992 const struct tcp_sock
*tp
= tcp_sk(sk
);
4994 /* If the user specified a specific send buffer setting, do
4997 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
5000 /* If we are under global TCP memory pressure, do not expand. */
5001 if (tcp_under_memory_pressure(sk
))
5004 /* If we are under soft global TCP memory pressure, do not expand. */
5005 if (sk_memory_allocated(sk
) >= sk_prot_mem_limits(sk
, 0))
5008 /* If we filled the congestion window, do not expand. */
5009 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
5015 /* When incoming ACK allowed to free some skb from write_queue,
5016 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
5017 * on the exit from tcp input handler.
5019 * PROBLEM: sndbuf expansion does not work well with largesend.
5021 static void tcp_new_space(struct sock
*sk
)
5023 struct tcp_sock
*tp
= tcp_sk(sk
);
5025 if (tcp_should_expand_sndbuf(sk
)) {
5026 tcp_sndbuf_expand(sk
);
5027 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
5030 sk
->sk_write_space(sk
);
5033 static void tcp_check_space(struct sock
*sk
)
5035 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
5036 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
5037 /* pairs with tcp_poll() */
5039 if (sk
->sk_socket
&&
5040 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
5042 if (!test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
5043 tcp_chrono_stop(sk
, TCP_CHRONO_SNDBUF_LIMITED
);
5048 static inline void tcp_data_snd_check(struct sock
*sk
)
5050 tcp_push_pending_frames(sk
);
5051 tcp_check_space(sk
);
5055 * Check if sending an ack is needed.
5057 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
5059 struct tcp_sock
*tp
= tcp_sk(sk
);
5061 /* More than one full frame received... */
5062 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
&&
5063 /* ... and right edge of window advances far enough.
5064 * (tcp_recvmsg() will send ACK otherwise). Or...
5066 __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
5067 /* We ACK each frame or... */
5068 tcp_in_quickack_mode(sk
) ||
5069 /* We have out of order data. */
5070 (ofo_possible
&& !RB_EMPTY_ROOT(&tp
->out_of_order_queue
))) {
5071 /* Then ack it now */
5074 /* Else, send delayed ack. */
5075 tcp_send_delayed_ack(sk
);
5079 static inline void tcp_ack_snd_check(struct sock
*sk
)
5081 if (!inet_csk_ack_scheduled(sk
)) {
5082 /* We sent a data segment already. */
5085 __tcp_ack_snd_check(sk
, 1);
5089 * This routine is only called when we have urgent data
5090 * signaled. Its the 'slow' part of tcp_urg. It could be
5091 * moved inline now as tcp_urg is only called from one
5092 * place. We handle URGent data wrong. We have to - as
5093 * BSD still doesn't use the correction from RFC961.
5094 * For 1003.1g we should support a new option TCP_STDURG to permit
5095 * either form (or just set the sysctl tcp_stdurg).
5098 static void tcp_check_urg(struct sock
*sk
, const struct tcphdr
*th
)
5100 struct tcp_sock
*tp
= tcp_sk(sk
);
5101 u32 ptr
= ntohs(th
->urg_ptr
);
5103 if (ptr
&& !sysctl_tcp_stdurg
)
5105 ptr
+= ntohl(th
->seq
);
5107 /* Ignore urgent data that we've already seen and read. */
5108 if (after(tp
->copied_seq
, ptr
))
5111 /* Do not replay urg ptr.
5113 * NOTE: interesting situation not covered by specs.
5114 * Misbehaving sender may send urg ptr, pointing to segment,
5115 * which we already have in ofo queue. We are not able to fetch
5116 * such data and will stay in TCP_URG_NOTYET until will be eaten
5117 * by recvmsg(). Seems, we are not obliged to handle such wicked
5118 * situations. But it is worth to think about possibility of some
5119 * DoSes using some hypothetical application level deadlock.
5121 if (before(ptr
, tp
->rcv_nxt
))
5124 /* Do we already have a newer (or duplicate) urgent pointer? */
5125 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
5128 /* Tell the world about our new urgent pointer. */
5131 /* We may be adding urgent data when the last byte read was
5132 * urgent. To do this requires some care. We cannot just ignore
5133 * tp->copied_seq since we would read the last urgent byte again
5134 * as data, nor can we alter copied_seq until this data arrives
5135 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5137 * NOTE. Double Dutch. Rendering to plain English: author of comment
5138 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5139 * and expect that both A and B disappear from stream. This is _wrong_.
5140 * Though this happens in BSD with high probability, this is occasional.
5141 * Any application relying on this is buggy. Note also, that fix "works"
5142 * only in this artificial test. Insert some normal data between A and B and we will
5143 * decline of BSD again. Verdict: it is better to remove to trap
5146 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
5147 !sock_flag(sk
, SOCK_URGINLINE
) && tp
->copied_seq
!= tp
->rcv_nxt
) {
5148 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
5150 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5151 __skb_unlink(skb
, &sk
->sk_receive_queue
);
5156 tp
->urg_data
= TCP_URG_NOTYET
;
5159 /* Disable header prediction. */
5163 /* This is the 'fast' part of urgent handling. */
5164 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, const struct tcphdr
*th
)
5166 struct tcp_sock
*tp
= tcp_sk(sk
);
5168 /* Check if we get a new urgent pointer - normally not. */
5170 tcp_check_urg(sk
, th
);
5172 /* Do we wait for any urgent data? - normally not... */
5173 if (tp
->urg_data
== TCP_URG_NOTYET
) {
5174 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
5177 /* Is the urgent pointer pointing into this packet? */
5178 if (ptr
< skb
->len
) {
5180 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
5182 tp
->urg_data
= TCP_URG_VALID
| tmp
;
5183 if (!sock_flag(sk
, SOCK_DEAD
))
5184 sk
->sk_data_ready(sk
);
5189 static int tcp_copy_to_iovec(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
5191 struct tcp_sock
*tp
= tcp_sk(sk
);
5192 int chunk
= skb
->len
- hlen
;
5195 if (skb_csum_unnecessary(skb
))
5196 err
= skb_copy_datagram_msg(skb
, hlen
, tp
->ucopy
.msg
, chunk
);
5198 err
= skb_copy_and_csum_datagram_msg(skb
, hlen
, tp
->ucopy
.msg
);
5201 tp
->ucopy
.len
-= chunk
;
5202 tp
->copied_seq
+= chunk
;
5203 tcp_rcv_space_adjust(sk
);
5209 /* Accept RST for rcv_nxt - 1 after a FIN.
5210 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a
5211 * FIN is sent followed by a RST packet. The RST is sent with the same
5212 * sequence number as the FIN, and thus according to RFC 5961 a challenge
5213 * ACK should be sent. However, Mac OSX rate limits replies to challenge
5214 * ACKs on the closed socket. In addition middleboxes can drop either the
5215 * challenge ACK or a subsequent RST.
5217 static bool tcp_reset_check(const struct sock
*sk
, const struct sk_buff
*skb
)
5219 struct tcp_sock
*tp
= tcp_sk(sk
);
5221 return unlikely(TCP_SKB_CB(skb
)->seq
== (tp
->rcv_nxt
- 1) &&
5222 (1 << sk
->sk_state
) & (TCPF_CLOSE_WAIT
| TCPF_LAST_ACK
|
5226 /* Does PAWS and seqno based validation of an incoming segment, flags will
5227 * play significant role here.
5229 static bool tcp_validate_incoming(struct sock
*sk
, struct sk_buff
*skb
,
5230 const struct tcphdr
*th
, int syn_inerr
)
5232 struct tcp_sock
*tp
= tcp_sk(sk
);
5233 bool rst_seq_match
= false;
5235 /* RFC1323: H1. Apply PAWS check first. */
5236 if (tcp_fast_parse_options(sock_net(sk
), skb
, th
, tp
) &&
5237 tp
->rx_opt
.saw_tstamp
&&
5238 tcp_paws_discard(sk
, skb
)) {
5240 NET_INC_STATS(sock_net(sk
), LINUX_MIB_PAWSESTABREJECTED
);
5241 if (!tcp_oow_rate_limited(sock_net(sk
), skb
,
5242 LINUX_MIB_TCPACKSKIPPEDPAWS
,
5243 &tp
->last_oow_ack_time
))
5244 tcp_send_dupack(sk
, skb
);
5247 /* Reset is accepted even if it did not pass PAWS. */
5250 /* Step 1: check sequence number */
5251 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5252 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5253 * (RST) segments are validated by checking their SEQ-fields."
5254 * And page 69: "If an incoming segment is not acceptable,
5255 * an acknowledgment should be sent in reply (unless the RST
5256 * bit is set, if so drop the segment and return)".
5261 if (!tcp_oow_rate_limited(sock_net(sk
), skb
,
5262 LINUX_MIB_TCPACKSKIPPEDSEQ
,
5263 &tp
->last_oow_ack_time
))
5264 tcp_send_dupack(sk
, skb
);
5265 } else if (tcp_reset_check(sk
, skb
)) {
5271 /* Step 2: check RST bit */
5273 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a
5274 * FIN and SACK too if available):
5275 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or
5276 * 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 tcp_reset_check(sk
, skb
)) {
5284 rst_seq_match
= true;
5285 } else if (tcp_is_sack(tp
) && tp
->rx_opt
.num_sacks
> 0) {
5286 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
5287 int max_sack
= sp
[0].end_seq
;
5290 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
;
5292 max_sack
= after(sp
[this_sack
].end_seq
,
5294 sp
[this_sack
].end_seq
: max_sack
;
5297 if (TCP_SKB_CB(skb
)->seq
== max_sack
)
5298 rst_seq_match
= true;
5304 /* Disable TFO if RST is out-of-order
5305 * and no data has been received
5306 * for current active TFO socket
5308 if (tp
->syn_fastopen
&& !tp
->data_segs_in
&&
5309 sk
->sk_state
== TCP_ESTABLISHED
)
5310 tcp_fastopen_active_disable(sk
);
5311 tcp_send_challenge_ack(sk
, skb
);
5316 /* step 3: check security and precedence [ignored] */
5318 /* step 4: Check for a SYN
5319 * RFC 5961 4.2 : Send a challenge ack
5324 TCP_INC_STATS(sock_net(sk
), TCP_MIB_INERRS
);
5325 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPSYNCHALLENGE
);
5326 tcp_send_challenge_ack(sk
, skb
);
5338 * TCP receive function for the ESTABLISHED state.
5340 * It is split into a fast path and a slow path. The fast path is
5342 * - A zero window was announced from us - zero window probing
5343 * is only handled properly in the slow path.
5344 * - Out of order segments arrived.
5345 * - Urgent data is expected.
5346 * - There is no buffer space left
5347 * - Unexpected TCP flags/window values/header lengths are received
5348 * (detected by checking the TCP header against pred_flags)
5349 * - Data is sent in both directions. Fast path only supports pure senders
5350 * or pure receivers (this means either the sequence number or the ack
5351 * value must stay constant)
5352 * - Unexpected TCP option.
5354 * When these conditions are not satisfied it drops into a standard
5355 * receive procedure patterned after RFC793 to handle all cases.
5356 * The first three cases are guaranteed by proper pred_flags setting,
5357 * the rest is checked inline. Fast processing is turned on in
5358 * tcp_data_queue when everything is OK.
5360 void tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
5361 const struct tcphdr
*th
)
5363 unsigned int len
= skb
->len
;
5364 struct tcp_sock
*tp
= tcp_sk(sk
);
5366 tcp_mstamp_refresh(tp
);
5367 if (unlikely(!sk
->sk_rx_dst
))
5368 inet_csk(sk
)->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
5370 * Header prediction.
5371 * The code loosely follows the one in the famous
5372 * "30 instruction TCP receive" Van Jacobson mail.
5374 * Van's trick is to deposit buffers into socket queue
5375 * on a device interrupt, to call tcp_recv function
5376 * on the receive process context and checksum and copy
5377 * the buffer to user space. smart...
5379 * Our current scheme is not silly either but we take the
5380 * extra cost of the net_bh soft interrupt processing...
5381 * We do checksum and copy also but from device to kernel.
5384 tp
->rx_opt
.saw_tstamp
= 0;
5386 /* pred_flags is 0xS?10 << 16 + snd_wnd
5387 * if header_prediction is to be made
5388 * 'S' will always be tp->tcp_header_len >> 2
5389 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5390 * turn it off (when there are holes in the receive
5391 * space for instance)
5392 * PSH flag is ignored.
5395 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
5396 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
&&
5397 !after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
)) {
5398 int tcp_header_len
= tp
->tcp_header_len
;
5400 /* Timestamp header prediction: tcp_header_len
5401 * is automatically equal to th->doff*4 due to pred_flags
5405 /* Check timestamp */
5406 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
5407 /* No? Slow path! */
5408 if (!tcp_parse_aligned_timestamp(tp
, th
))
5411 /* If PAWS failed, check it more carefully in slow path */
5412 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
5415 /* DO NOT update ts_recent here, if checksum fails
5416 * and timestamp was corrupted part, it will result
5417 * in a hung connection since we will drop all
5418 * future packets due to the PAWS test.
5422 if (len
<= tcp_header_len
) {
5423 /* Bulk data transfer: sender */
5424 if (len
== tcp_header_len
) {
5425 /* Predicted packet is in window by definition.
5426 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5427 * Hence, check seq<=rcv_wup reduces to:
5429 if (tcp_header_len
==
5430 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5431 tp
->rcv_nxt
== tp
->rcv_wup
)
5432 tcp_store_ts_recent(tp
);
5434 /* We know that such packets are checksummed
5437 tcp_ack(sk
, skb
, 0);
5439 tcp_data_snd_check(sk
);
5441 } else { /* Header too small */
5442 TCP_INC_STATS(sock_net(sk
), TCP_MIB_INERRS
);
5447 bool fragstolen
= false;
5449 if (tp
->ucopy
.task
== current
&&
5450 tp
->copied_seq
== tp
->rcv_nxt
&&
5451 len
- tcp_header_len
<= tp
->ucopy
.len
&&
5452 sock_owned_by_user(sk
)) {
5453 __set_current_state(TASK_RUNNING
);
5455 if (!tcp_copy_to_iovec(sk
, skb
, tcp_header_len
)) {
5456 /* Predicted packet is in window by definition.
5457 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5458 * Hence, check seq<=rcv_wup reduces to:
5460 if (tcp_header_len
==
5461 (sizeof(struct tcphdr
) +
5462 TCPOLEN_TSTAMP_ALIGNED
) &&
5463 tp
->rcv_nxt
== tp
->rcv_wup
)
5464 tcp_store_ts_recent(tp
);
5466 tcp_rcv_rtt_measure_ts(sk
, skb
);
5468 __skb_pull(skb
, tcp_header_len
);
5469 tcp_rcv_nxt_update(tp
, TCP_SKB_CB(skb
)->end_seq
);
5470 NET_INC_STATS(sock_net(sk
),
5471 LINUX_MIB_TCPHPHITSTOUSER
);
5476 if (tcp_checksum_complete(skb
))
5479 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
5482 /* Predicted packet is in window by definition.
5483 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5484 * Hence, check seq<=rcv_wup reduces to:
5486 if (tcp_header_len
==
5487 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5488 tp
->rcv_nxt
== tp
->rcv_wup
)
5489 tcp_store_ts_recent(tp
);
5491 tcp_rcv_rtt_measure_ts(sk
, skb
);
5493 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPHPHITS
);
5495 /* Bulk data transfer: receiver */
5496 eaten
= tcp_queue_rcv(sk
, skb
, tcp_header_len
,
5500 tcp_event_data_recv(sk
, skb
);
5502 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
5503 /* Well, only one small jumplet in fast path... */
5504 tcp_ack(sk
, skb
, FLAG_DATA
);
5505 tcp_data_snd_check(sk
);
5506 if (!inet_csk_ack_scheduled(sk
))
5510 __tcp_ack_snd_check(sk
, 0);
5513 kfree_skb_partial(skb
, fragstolen
);
5514 sk
->sk_data_ready(sk
);
5520 if (len
< (th
->doff
<< 2) || tcp_checksum_complete(skb
))
5523 if (!th
->ack
&& !th
->rst
&& !th
->syn
)
5527 * Standard slow path.
5530 if (!tcp_validate_incoming(sk
, skb
, th
, 1))
5534 if (tcp_ack(sk
, skb
, FLAG_SLOWPATH
| FLAG_UPDATE_TS_RECENT
) < 0)
5537 tcp_rcv_rtt_measure_ts(sk
, skb
);
5539 /* Process urgent data. */
5540 tcp_urg(sk
, skb
, th
);
5542 /* step 7: process the segment text */
5543 tcp_data_queue(sk
, skb
);
5545 tcp_data_snd_check(sk
);
5546 tcp_ack_snd_check(sk
);
5550 TCP_INC_STATS(sock_net(sk
), TCP_MIB_CSUMERRORS
);
5551 TCP_INC_STATS(sock_net(sk
), TCP_MIB_INERRS
);
5556 EXPORT_SYMBOL(tcp_rcv_established
);
5558 void tcp_finish_connect(struct sock
*sk
, struct sk_buff
*skb
)
5560 struct tcp_sock
*tp
= tcp_sk(sk
);
5561 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5563 tcp_set_state(sk
, TCP_ESTABLISHED
);
5564 icsk
->icsk_ack
.lrcvtime
= tcp_jiffies32
;
5567 icsk
->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
5568 security_inet_conn_established(sk
, skb
);
5571 /* Make sure socket is routed, for correct metrics. */
5572 icsk
->icsk_af_ops
->rebuild_header(sk
);
5574 tcp_init_metrics(sk
);
5575 tcp_call_bpf(sk
, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB
);
5576 tcp_init_congestion_control(sk
);
5578 /* Prevent spurious tcp_cwnd_restart() on first data
5581 tp
->lsndtime
= tcp_jiffies32
;
5583 tcp_init_buffer_space(sk
);
5585 if (sock_flag(sk
, SOCK_KEEPOPEN
))
5586 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
5588 if (!tp
->rx_opt
.snd_wscale
)
5589 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
5595 static bool tcp_rcv_fastopen_synack(struct sock
*sk
, struct sk_buff
*synack
,
5596 struct tcp_fastopen_cookie
*cookie
)
5598 struct tcp_sock
*tp
= tcp_sk(sk
);
5599 struct sk_buff
*data
= tp
->syn_data
? tcp_write_queue_head(sk
) : NULL
;
5600 u16 mss
= tp
->rx_opt
.mss_clamp
, try_exp
= 0;
5601 bool syn_drop
= false;
5603 if (mss
== tp
->rx_opt
.user_mss
) {
5604 struct tcp_options_received opt
;
5606 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5607 tcp_clear_options(&opt
);
5608 opt
.user_mss
= opt
.mss_clamp
= 0;
5609 tcp_parse_options(sock_net(sk
), synack
, &opt
, 0, NULL
);
5610 mss
= opt
.mss_clamp
;
5613 if (!tp
->syn_fastopen
) {
5614 /* Ignore an unsolicited cookie */
5616 } else if (tp
->total_retrans
) {
5617 /* SYN timed out and the SYN-ACK neither has a cookie nor
5618 * acknowledges data. Presumably the remote received only
5619 * the retransmitted (regular) SYNs: either the original
5620 * SYN-data or the corresponding SYN-ACK was dropped.
5622 syn_drop
= (cookie
->len
< 0 && data
);
5623 } else if (cookie
->len
< 0 && !tp
->syn_data
) {
5624 /* We requested a cookie but didn't get it. If we did not use
5625 * the (old) exp opt format then try so next time (try_exp=1).
5626 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5628 try_exp
= tp
->syn_fastopen_exp
? 2 : 1;
5631 tcp_fastopen_cache_set(sk
, mss
, cookie
, syn_drop
, try_exp
);
5633 if (data
) { /* Retransmit unacked data in SYN */
5634 tcp_for_write_queue_from(data
, sk
) {
5635 if (data
== tcp_send_head(sk
) ||
5636 __tcp_retransmit_skb(sk
, data
, 1))
5640 NET_INC_STATS(sock_net(sk
),
5641 LINUX_MIB_TCPFASTOPENACTIVEFAIL
);
5644 tp
->syn_data_acked
= tp
->syn_data
;
5645 if (tp
->syn_data_acked
)
5646 NET_INC_STATS(sock_net(sk
),
5647 LINUX_MIB_TCPFASTOPENACTIVE
);
5649 tcp_fastopen_add_skb(sk
, synack
);
5654 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5655 const struct tcphdr
*th
)
5657 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5658 struct tcp_sock
*tp
= tcp_sk(sk
);
5659 struct tcp_fastopen_cookie foc
= { .len
= -1 };
5660 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
5663 tcp_parse_options(sock_net(sk
), skb
, &tp
->rx_opt
, 0, &foc
);
5664 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
5665 tp
->rx_opt
.rcv_tsecr
-= tp
->tsoffset
;
5669 * "If the state is SYN-SENT then
5670 * first check the ACK bit
5671 * If the ACK bit is set
5672 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5673 * a reset (unless the RST bit is set, if so drop
5674 * the segment and return)"
5676 if (!after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_una
) ||
5677 after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
))
5678 goto reset_and_undo
;
5680 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
5681 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
5682 tcp_time_stamp(tp
))) {
5683 NET_INC_STATS(sock_net(sk
),
5684 LINUX_MIB_PAWSACTIVEREJECTED
);
5685 goto reset_and_undo
;
5688 /* Now ACK is acceptable.
5690 * "If the RST bit is set
5691 * If the ACK was acceptable then signal the user "error:
5692 * connection reset", drop the segment, enter CLOSED state,
5693 * delete TCB, and return."
5702 * "fifth, if neither of the SYN or RST bits is set then
5703 * drop the segment and return."
5709 goto discard_and_undo
;
5712 * "If the SYN bit is on ...
5713 * are acceptable then ...
5714 * (our SYN has been ACKed), change the connection
5715 * state to ESTABLISHED..."
5718 tcp_ecn_rcv_synack(tp
, th
);
5720 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5721 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
5723 /* Ok.. it's good. Set up sequence numbers and
5724 * move to established.
5726 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5727 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5729 /* RFC1323: The window in SYN & SYN/ACK segments is
5732 tp
->snd_wnd
= ntohs(th
->window
);
5734 if (!tp
->rx_opt
.wscale_ok
) {
5735 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
5736 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
5739 if (tp
->rx_opt
.saw_tstamp
) {
5740 tp
->rx_opt
.tstamp_ok
= 1;
5741 tp
->tcp_header_len
=
5742 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5743 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5744 tcp_store_ts_recent(tp
);
5746 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5749 if (tcp_is_sack(tp
) && sysctl_tcp_fack
)
5750 tcp_enable_fack(tp
);
5753 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5754 tcp_initialize_rcv_mss(sk
);
5756 /* Remember, tcp_poll() does not lock socket!
5757 * Change state from SYN-SENT only after copied_seq
5758 * is initialized. */
5759 tp
->copied_seq
= tp
->rcv_nxt
;
5763 tcp_finish_connect(sk
, skb
);
5765 fastopen_fail
= (tp
->syn_fastopen
|| tp
->syn_data
) &&
5766 tcp_rcv_fastopen_synack(sk
, skb
, &foc
);
5768 if (!sock_flag(sk
, SOCK_DEAD
)) {
5769 sk
->sk_state_change(sk
);
5770 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5774 if (sk
->sk_write_pending
||
5775 icsk
->icsk_accept_queue
.rskq_defer_accept
||
5776 icsk
->icsk_ack
.pingpong
) {
5777 /* Save one ACK. Data will be ready after
5778 * several ticks, if write_pending is set.
5780 * It may be deleted, but with this feature tcpdumps
5781 * look so _wonderfully_ clever, that I was not able
5782 * to stand against the temptation 8) --ANK
5784 inet_csk_schedule_ack(sk
);
5785 tcp_enter_quickack_mode(sk
);
5786 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
5787 TCP_DELACK_MAX
, TCP_RTO_MAX
);
5798 /* No ACK in the segment */
5802 * "If the RST bit is set
5804 * Otherwise (no ACK) drop the segment and return."
5807 goto discard_and_undo
;
5811 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&&
5812 tcp_paws_reject(&tp
->rx_opt
, 0))
5813 goto discard_and_undo
;
5816 /* We see SYN without ACK. It is attempt of
5817 * simultaneous connect with crossed SYNs.
5818 * Particularly, it can be connect to self.
5820 tcp_set_state(sk
, TCP_SYN_RECV
);
5822 if (tp
->rx_opt
.saw_tstamp
) {
5823 tp
->rx_opt
.tstamp_ok
= 1;
5824 tcp_store_ts_recent(tp
);
5825 tp
->tcp_header_len
=
5826 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5828 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5831 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5832 tp
->copied_seq
= tp
->rcv_nxt
;
5833 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5835 /* RFC1323: The window in SYN & SYN/ACK segments is
5838 tp
->snd_wnd
= ntohs(th
->window
);
5839 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
5840 tp
->max_window
= tp
->snd_wnd
;
5842 tcp_ecn_rcv_syn(tp
, th
);
5845 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5846 tcp_initialize_rcv_mss(sk
);
5848 tcp_send_synack(sk
);
5850 /* Note, we could accept data and URG from this segment.
5851 * There are no obstacles to make this (except that we must
5852 * either change tcp_recvmsg() to prevent it from returning data
5853 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5855 * However, if we ignore data in ACKless segments sometimes,
5856 * we have no reasons to accept it sometimes.
5857 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5858 * is not flawless. So, discard packet for sanity.
5859 * Uncomment this return to process the data.
5866 /* "fifth, if neither of the SYN or RST bits is set then
5867 * drop the segment and return."
5871 tcp_clear_options(&tp
->rx_opt
);
5872 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5876 tcp_clear_options(&tp
->rx_opt
);
5877 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5882 * This function implements the receiving procedure of RFC 793 for
5883 * all states except ESTABLISHED and TIME_WAIT.
5884 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5885 * address independent.
5888 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
)
5890 struct tcp_sock
*tp
= tcp_sk(sk
);
5891 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5892 const struct tcphdr
*th
= tcp_hdr(skb
);
5893 struct request_sock
*req
;
5897 switch (sk
->sk_state
) {
5911 /* It is possible that we process SYN packets from backlog,
5912 * so we need to make sure to disable BH right there.
5915 acceptable
= icsk
->icsk_af_ops
->conn_request(sk
, skb
) >= 0;
5926 tp
->rx_opt
.saw_tstamp
= 0;
5927 tcp_mstamp_refresh(tp
);
5928 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
);
5932 /* Do step6 onward by hand. */
5933 tcp_urg(sk
, skb
, th
);
5935 tcp_data_snd_check(sk
);
5939 tcp_mstamp_refresh(tp
);
5940 tp
->rx_opt
.saw_tstamp
= 0;
5941 req
= tp
->fastopen_rsk
;
5943 WARN_ON_ONCE(sk
->sk_state
!= TCP_SYN_RECV
&&
5944 sk
->sk_state
!= TCP_FIN_WAIT1
);
5946 if (!tcp_check_req(sk
, skb
, req
, true))
5950 if (!th
->ack
&& !th
->rst
&& !th
->syn
)
5953 if (!tcp_validate_incoming(sk
, skb
, th
, 0))
5956 /* step 5: check the ACK field */
5957 acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
|
5958 FLAG_UPDATE_TS_RECENT
|
5959 FLAG_NO_CHALLENGE_ACK
) > 0;
5962 if (sk
->sk_state
== TCP_SYN_RECV
)
5963 return 1; /* send one RST */
5964 tcp_send_challenge_ack(sk
, skb
);
5967 switch (sk
->sk_state
) {
5970 tcp_synack_rtt_meas(sk
, req
);
5972 /* Once we leave TCP_SYN_RECV, we no longer need req
5976 inet_csk(sk
)->icsk_retransmits
= 0;
5977 reqsk_fastopen_remove(sk
, req
, false);
5979 /* Make sure socket is routed, for correct metrics. */
5980 icsk
->icsk_af_ops
->rebuild_header(sk
);
5981 tcp_call_bpf(sk
, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB
);
5982 tcp_init_congestion_control(sk
);
5985 tp
->copied_seq
= tp
->rcv_nxt
;
5986 tcp_init_buffer_space(sk
);
5989 tcp_set_state(sk
, TCP_ESTABLISHED
);
5990 sk
->sk_state_change(sk
);
5992 /* Note, that this wakeup is only for marginal crossed SYN case.
5993 * Passively open sockets are not waked up, because
5994 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5997 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5999 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
6000 tp
->snd_wnd
= ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
;
6001 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
6003 if (tp
->rx_opt
.tstamp_ok
)
6004 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
6007 /* Re-arm the timer because data may have been sent out.
6008 * This is similar to the regular data transmission case
6009 * when new data has just been ack'ed.
6011 * (TFO) - we could try to be more aggressive and
6012 * retransmitting any data sooner based on when they
6017 tcp_init_metrics(sk
);
6019 if (!inet_csk(sk
)->icsk_ca_ops
->cong_control
)
6020 tcp_update_pacing_rate(sk
);
6022 /* Prevent spurious tcp_cwnd_restart() on first data packet */
6023 tp
->lsndtime
= tcp_jiffies32
;
6025 tcp_initialize_rcv_mss(sk
);
6026 tcp_fast_path_on(tp
);
6029 case TCP_FIN_WAIT1
: {
6032 /* If we enter the TCP_FIN_WAIT1 state and we are a
6033 * Fast Open socket and this is the first acceptable
6034 * ACK we have received, this would have acknowledged
6035 * our SYNACK so stop the SYNACK timer.
6038 /* We no longer need the request sock. */
6039 reqsk_fastopen_remove(sk
, req
, false);
6042 if (tp
->snd_una
!= tp
->write_seq
)
6045 tcp_set_state(sk
, TCP_FIN_WAIT2
);
6046 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
6050 if (!sock_flag(sk
, SOCK_DEAD
)) {
6051 /* Wake up lingering close() */
6052 sk
->sk_state_change(sk
);
6056 if (tp
->linger2
< 0) {
6058 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
6061 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
6062 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
6063 /* Receive out of order FIN after close() */
6064 if (tp
->syn_fastopen
&& th
->fin
)
6065 tcp_fastopen_active_disable(sk
);
6067 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
6071 tmo
= tcp_fin_time(sk
);
6072 if (tmo
> TCP_TIMEWAIT_LEN
) {
6073 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
6074 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
6075 /* Bad case. We could lose such FIN otherwise.
6076 * It is not a big problem, but it looks confusing
6077 * and not so rare event. We still can lose it now,
6078 * if it spins in bh_lock_sock(), but it is really
6081 inet_csk_reset_keepalive_timer(sk
, tmo
);
6083 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
6090 if (tp
->snd_una
== tp
->write_seq
) {
6091 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
6097 if (tp
->snd_una
== tp
->write_seq
) {
6098 tcp_update_metrics(sk
);
6105 /* step 6: check the URG bit */
6106 tcp_urg(sk
, skb
, th
);
6108 /* step 7: process the segment text */
6109 switch (sk
->sk_state
) {
6110 case TCP_CLOSE_WAIT
:
6113 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
6117 /* RFC 793 says to queue data in these states,
6118 * RFC 1122 says we MUST send a reset.
6119 * BSD 4.4 also does reset.
6121 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
6122 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
6123 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
6124 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
6130 case TCP_ESTABLISHED
:
6131 tcp_data_queue(sk
, skb
);
6136 /* tcp_data could move socket to TIME-WAIT */
6137 if (sk
->sk_state
!= TCP_CLOSE
) {
6138 tcp_data_snd_check(sk
);
6139 tcp_ack_snd_check(sk
);
6148 EXPORT_SYMBOL(tcp_rcv_state_process
);
6150 static inline void pr_drop_req(struct request_sock
*req
, __u16 port
, int family
)
6152 struct inet_request_sock
*ireq
= inet_rsk(req
);
6154 if (family
== AF_INET
)
6155 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6156 &ireq
->ir_rmt_addr
, port
);
6157 #if IS_ENABLED(CONFIG_IPV6)
6158 else if (family
== AF_INET6
)
6159 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6160 &ireq
->ir_v6_rmt_addr
, port
);
6164 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6166 * If we receive a SYN packet with these bits set, it means a
6167 * network is playing bad games with TOS bits. In order to
6168 * avoid possible false congestion notifications, we disable
6169 * TCP ECN negotiation.
6171 * Exception: tcp_ca wants ECN. This is required for DCTCP
6172 * congestion control: Linux DCTCP asserts ECT on all packets,
6173 * including SYN, which is most optimal solution; however,
6174 * others, such as FreeBSD do not.
6176 static void tcp_ecn_create_request(struct request_sock
*req
,
6177 const struct sk_buff
*skb
,
6178 const struct sock
*listen_sk
,
6179 const struct dst_entry
*dst
)
6181 const struct tcphdr
*th
= tcp_hdr(skb
);
6182 const struct net
*net
= sock_net(listen_sk
);
6183 bool th_ecn
= th
->ece
&& th
->cwr
;
6190 ect
= !INET_ECN_is_not_ect(TCP_SKB_CB(skb
)->ip_dsfield
);
6191 ecn_ok_dst
= dst_feature(dst
, DST_FEATURE_ECN_MASK
);
6192 ecn_ok
= net
->ipv4
.sysctl_tcp_ecn
|| ecn_ok_dst
;
6194 if ((!ect
&& ecn_ok
) || tcp_ca_needs_ecn(listen_sk
) ||
6195 (ecn_ok_dst
& DST_FEATURE_ECN_CA
) ||
6196 tcp_bpf_ca_needs_ecn((struct sock
*)req
))
6197 inet_rsk(req
)->ecn_ok
= 1;
6200 static void tcp_openreq_init(struct request_sock
*req
,
6201 const struct tcp_options_received
*rx_opt
,
6202 struct sk_buff
*skb
, const struct sock
*sk
)
6204 struct inet_request_sock
*ireq
= inet_rsk(req
);
6206 req
->rsk_rcv_wnd
= 0; /* So that tcp_send_synack() knows! */
6208 tcp_rsk(req
)->rcv_isn
= TCP_SKB_CB(skb
)->seq
;
6209 tcp_rsk(req
)->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
6210 tcp_rsk(req
)->snt_synack
= tcp_clock_us();
6211 tcp_rsk(req
)->last_oow_ack_time
= 0;
6212 req
->mss
= rx_opt
->mss_clamp
;
6213 req
->ts_recent
= rx_opt
->saw_tstamp
? rx_opt
->rcv_tsval
: 0;
6214 ireq
->tstamp_ok
= rx_opt
->tstamp_ok
;
6215 ireq
->sack_ok
= rx_opt
->sack_ok
;
6216 ireq
->snd_wscale
= rx_opt
->snd_wscale
;
6217 ireq
->wscale_ok
= rx_opt
->wscale_ok
;
6220 ireq
->ir_rmt_port
= tcp_hdr(skb
)->source
;
6221 ireq
->ir_num
= ntohs(tcp_hdr(skb
)->dest
);
6222 ireq
->ir_mark
= inet_request_mark(sk
, skb
);
6225 struct request_sock
*inet_reqsk_alloc(const struct request_sock_ops
*ops
,
6226 struct sock
*sk_listener
,
6227 bool attach_listener
)
6229 struct request_sock
*req
= reqsk_alloc(ops
, sk_listener
,
6233 struct inet_request_sock
*ireq
= inet_rsk(req
);
6235 kmemcheck_annotate_bitfield(ireq
, flags
);
6237 #if IS_ENABLED(CONFIG_IPV6)
6238 ireq
->pktopts
= NULL
;
6240 atomic64_set(&ireq
->ir_cookie
, 0);
6241 ireq
->ireq_state
= TCP_NEW_SYN_RECV
;
6242 write_pnet(&ireq
->ireq_net
, sock_net(sk_listener
));
6243 ireq
->ireq_family
= sk_listener
->sk_family
;
6248 EXPORT_SYMBOL(inet_reqsk_alloc
);
6251 * Return true if a syncookie should be sent
6253 static bool tcp_syn_flood_action(const struct sock
*sk
,
6254 const struct sk_buff
*skb
,
6257 struct request_sock_queue
*queue
= &inet_csk(sk
)->icsk_accept_queue
;
6258 const char *msg
= "Dropping request";
6259 bool want_cookie
= false;
6260 struct net
*net
= sock_net(sk
);
6262 #ifdef CONFIG_SYN_COOKIES
6263 if (net
->ipv4
.sysctl_tcp_syncookies
) {
6264 msg
= "Sending cookies";
6266 __NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPREQQFULLDOCOOKIES
);
6269 __NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPREQQFULLDROP
);
6271 if (!queue
->synflood_warned
&&
6272 net
->ipv4
.sysctl_tcp_syncookies
!= 2 &&
6273 xchg(&queue
->synflood_warned
, 1) == 0)
6274 pr_info("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6275 proto
, ntohs(tcp_hdr(skb
)->dest
), msg
);
6280 static void tcp_reqsk_record_syn(const struct sock
*sk
,
6281 struct request_sock
*req
,
6282 const struct sk_buff
*skb
)
6284 if (tcp_sk(sk
)->save_syn
) {
6285 u32 len
= skb_network_header_len(skb
) + tcp_hdrlen(skb
);
6288 copy
= kmalloc(len
+ sizeof(u32
), GFP_ATOMIC
);
6291 memcpy(©
[1], skb_network_header(skb
), len
);
6292 req
->saved_syn
= copy
;
6297 int tcp_conn_request(struct request_sock_ops
*rsk_ops
,
6298 const struct tcp_request_sock_ops
*af_ops
,
6299 struct sock
*sk
, struct sk_buff
*skb
)
6301 struct tcp_fastopen_cookie foc
= { .len
= -1 };
6302 __u32 isn
= TCP_SKB_CB(skb
)->tcp_tw_isn
;
6303 struct tcp_options_received tmp_opt
;
6304 struct tcp_sock
*tp
= tcp_sk(sk
);
6305 struct net
*net
= sock_net(sk
);
6306 struct sock
*fastopen_sk
= NULL
;
6307 struct dst_entry
*dst
= NULL
;
6308 struct request_sock
*req
;
6309 bool want_cookie
= false;
6312 /* TW buckets are converted to open requests without
6313 * limitations, they conserve resources and peer is
6314 * evidently real one.
6316 if ((net
->ipv4
.sysctl_tcp_syncookies
== 2 ||
6317 inet_csk_reqsk_queue_is_full(sk
)) && !isn
) {
6318 want_cookie
= tcp_syn_flood_action(sk
, skb
, rsk_ops
->slab_name
);
6323 if (sk_acceptq_is_full(sk
)) {
6324 NET_INC_STATS(sock_net(sk
), LINUX_MIB_LISTENOVERFLOWS
);
6328 req
= inet_reqsk_alloc(rsk_ops
, sk
, !want_cookie
);
6332 tcp_rsk(req
)->af_specific
= af_ops
;
6333 tcp_rsk(req
)->ts_off
= 0;
6335 tcp_clear_options(&tmp_opt
);
6336 tmp_opt
.mss_clamp
= af_ops
->mss_clamp
;
6337 tmp_opt
.user_mss
= tp
->rx_opt
.user_mss
;
6338 tcp_parse_options(sock_net(sk
), skb
, &tmp_opt
, 0,
6339 want_cookie
? NULL
: &foc
);
6341 if (want_cookie
&& !tmp_opt
.saw_tstamp
)
6342 tcp_clear_options(&tmp_opt
);
6344 tmp_opt
.tstamp_ok
= tmp_opt
.saw_tstamp
;
6345 tcp_openreq_init(req
, &tmp_opt
, skb
, sk
);
6346 inet_rsk(req
)->no_srccheck
= inet_sk(sk
)->transparent
;
6348 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6349 inet_rsk(req
)->ir_iif
= inet_request_bound_dev_if(sk
, skb
);
6351 af_ops
->init_req(req
, sk
, skb
);
6353 if (security_inet_conn_request(sk
, skb
, req
))
6356 if (tmp_opt
.tstamp_ok
)
6357 tcp_rsk(req
)->ts_off
= af_ops
->init_ts_off(net
, skb
);
6359 if (!want_cookie
&& !isn
) {
6360 /* Kill the following clause, if you dislike this way. */
6361 if (!net
->ipv4
.sysctl_tcp_syncookies
&&
6362 (net
->ipv4
.sysctl_max_syn_backlog
- inet_csk_reqsk_queue_len(sk
) <
6363 (net
->ipv4
.sysctl_max_syn_backlog
>> 2)) &&
6364 !tcp_peer_is_proven(req
, dst
)) {
6365 /* Without syncookies last quarter of
6366 * backlog is filled with destinations,
6367 * proven to be alive.
6368 * It means that we continue to communicate
6369 * to destinations, already remembered
6370 * to the moment of synflood.
6372 pr_drop_req(req
, ntohs(tcp_hdr(skb
)->source
),
6374 goto drop_and_release
;
6377 isn
= af_ops
->init_seq(skb
);
6380 dst
= af_ops
->route_req(sk
, &fl
, req
);
6385 tcp_ecn_create_request(req
, skb
, sk
, dst
);
6388 isn
= cookie_init_sequence(af_ops
, sk
, skb
, &req
->mss
);
6389 req
->cookie_ts
= tmp_opt
.tstamp_ok
;
6390 if (!tmp_opt
.tstamp_ok
)
6391 inet_rsk(req
)->ecn_ok
= 0;
6394 tcp_rsk(req
)->snt_isn
= isn
;
6395 tcp_rsk(req
)->txhash
= net_tx_rndhash();
6396 tcp_openreq_init_rwin(req
, sk
, dst
);
6398 tcp_reqsk_record_syn(sk
, req
, skb
);
6399 fastopen_sk
= tcp_try_fastopen(sk
, skb
, req
, &foc
, dst
);
6402 af_ops
->send_synack(fastopen_sk
, dst
, &fl
, req
,
6403 &foc
, TCP_SYNACK_FASTOPEN
);
6404 /* Add the child socket directly into the accept queue */
6405 inet_csk_reqsk_queue_add(sk
, req
, fastopen_sk
);
6406 sk
->sk_data_ready(sk
);
6407 bh_unlock_sock(fastopen_sk
);
6408 sock_put(fastopen_sk
);
6410 tcp_rsk(req
)->tfo_listener
= false;
6412 inet_csk_reqsk_queue_hash_add(sk
, req
,
6413 tcp_timeout_init((struct sock
*)req
));
6414 af_ops
->send_synack(sk
, dst
, &fl
, req
, &foc
,
6415 !want_cookie
? TCP_SYNACK_NORMAL
:
6433 EXPORT_SYMBOL(tcp_conn_request
);