1 // SPDX-License-Identifier: GPL-2.0
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
7 * Implementation of the Transmission Control Protocol(TCP).
10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11 * Mark Evans, <evansmp@uhura.aston.ac.uk>
12 * Corey Minyard <wf-rch!minyard@relay.EU.net>
13 * Florian La Roche, <flla@stud.uni-sb.de>
14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15 * Linus Torvalds, <torvalds@cs.helsinki.fi>
16 * Alan Cox, <gw4pts@gw4pts.ampr.org>
17 * Matthew Dillon, <dillon@apollo.west.oic.com>
18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19 * Jorge Cwik, <jorge@laser.satlink.net>
24 * Pedro Roque : Fast Retransmit/Recovery.
26 * Retransmit queue handled by TCP.
27 * Better retransmit timer handling.
28 * New congestion avoidance.
32 * Eric : Fast Retransmit.
33 * Randy Scott : MSS option defines.
34 * Eric Schenk : Fixes to slow start algorithm.
35 * Eric Schenk : Yet another double ACK bug.
36 * Eric Schenk : Delayed ACK bug fixes.
37 * Eric Schenk : Floyd style fast retrans war avoidance.
38 * David S. Miller : Don't allow zero congestion window.
39 * Eric Schenk : Fix retransmitter so that it sends
40 * next packet on ack of previous packet.
41 * Andi Kleen : Moved open_request checking here
42 * and process RSTs for open_requests.
43 * Andi Kleen : Better prune_queue, and other fixes.
44 * Andrey Savochkin: Fix RTT measurements in the presence of
46 * Andrey Savochkin: Check sequence numbers correctly when
47 * removing SACKs due to in sequence incoming
49 * Andi Kleen: Make sure we never ack data there is not
50 * enough room for. Also make this condition
51 * a fatal error if it might still happen.
52 * Andi Kleen: Add tcp_measure_rcv_mss to make
53 * connections with MSS<min(MTU,ann. MSS)
54 * work without delayed acks.
55 * Andi Kleen: Process packets with PSH set in the
57 * J Hadi Salim: ECN support
60 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
61 * engine. Lots of bugs are found.
62 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
65 #define pr_fmt(fmt) "TCP: " fmt
68 #include <linux/slab.h>
69 #include <linux/module.h>
70 #include <linux/sysctl.h>
71 #include <linux/kernel.h>
72 #include <linux/prefetch.h>
75 #include <net/inet_common.h>
76 #include <linux/ipsec.h>
77 #include <asm/unaligned.h>
78 #include <linux/errqueue.h>
79 #include <trace/events/tcp.h>
80 #include <linux/static_key.h>
81 #include <net/busy_poll.h>
83 int sysctl_tcp_max_orphans __read_mostly
= NR_FILE
;
85 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
86 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
87 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
88 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
89 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
90 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
91 #define FLAG_ECE 0x40 /* ECE in this ACK */
92 #define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */
93 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
94 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
95 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
96 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
97 #define FLAG_SET_XMIT_TIMER 0x1000 /* Set TLP or RTO timer */
98 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
99 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
100 #define FLAG_NO_CHALLENGE_ACK 0x8000 /* do not call tcp_send_challenge_ack() */
101 #define FLAG_ACK_MAYBE_DELAYED 0x10000 /* Likely a delayed ACK */
103 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
104 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
105 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE|FLAG_DSACKING_ACK)
106 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
108 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
109 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
111 #define REXMIT_NONE 0 /* no loss recovery to do */
112 #define REXMIT_LOST 1 /* retransmit packets marked lost */
113 #define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */
115 #if IS_ENABLED(CONFIG_TLS_DEVICE)
116 static DEFINE_STATIC_KEY_FALSE(clean_acked_data_enabled
);
118 void clean_acked_data_enable(struct inet_connection_sock
*icsk
,
119 void (*cad
)(struct sock
*sk
, u32 ack_seq
))
121 icsk
->icsk_clean_acked
= cad
;
122 static_branch_inc(&clean_acked_data_enabled
);
124 EXPORT_SYMBOL_GPL(clean_acked_data_enable
);
126 void clean_acked_data_disable(struct inet_connection_sock
*icsk
)
128 static_branch_dec(&clean_acked_data_enabled
);
129 icsk
->icsk_clean_acked
= NULL
;
131 EXPORT_SYMBOL_GPL(clean_acked_data_disable
);
134 static void tcp_gro_dev_warn(struct sock
*sk
, const struct sk_buff
*skb
,
137 static bool __once __read_mostly
;
140 struct net_device
*dev
;
145 dev
= dev_get_by_index_rcu(sock_net(sk
), skb
->skb_iif
);
146 if (!dev
|| len
>= dev
->mtu
)
147 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n",
148 dev
? dev
->name
: "Unknown driver");
153 /* Adapt the MSS value used to make delayed ack decision to the
156 static void tcp_measure_rcv_mss(struct sock
*sk
, const struct sk_buff
*skb
)
158 struct inet_connection_sock
*icsk
= inet_csk(sk
);
159 const unsigned int lss
= icsk
->icsk_ack
.last_seg_size
;
162 icsk
->icsk_ack
.last_seg_size
= 0;
164 /* skb->len may jitter because of SACKs, even if peer
165 * sends good full-sized frames.
167 len
= skb_shinfo(skb
)->gso_size
? : skb
->len
;
168 if (len
>= icsk
->icsk_ack
.rcv_mss
) {
169 icsk
->icsk_ack
.rcv_mss
= min_t(unsigned int, len
,
171 /* Account for possibly-removed options */
172 if (unlikely(len
> icsk
->icsk_ack
.rcv_mss
+
173 MAX_TCP_OPTION_SPACE
))
174 tcp_gro_dev_warn(sk
, skb
, len
);
176 /* Otherwise, we make more careful check taking into account,
177 * that SACKs block is variable.
179 * "len" is invariant segment length, including TCP header.
181 len
+= skb
->data
- skb_transport_header(skb
);
182 if (len
>= TCP_MSS_DEFAULT
+ sizeof(struct tcphdr
) ||
183 /* If PSH is not set, packet should be
184 * full sized, provided peer TCP is not badly broken.
185 * This observation (if it is correct 8)) allows
186 * to handle super-low mtu links fairly.
188 (len
>= TCP_MIN_MSS
+ sizeof(struct tcphdr
) &&
189 !(tcp_flag_word(tcp_hdr(skb
)) & TCP_REMNANT
))) {
190 /* Subtract also invariant (if peer is RFC compliant),
191 * tcp header plus fixed timestamp option length.
192 * Resulting "len" is MSS free of SACK jitter.
194 len
-= tcp_sk(sk
)->tcp_header_len
;
195 icsk
->icsk_ack
.last_seg_size
= len
;
197 icsk
->icsk_ack
.rcv_mss
= len
;
201 if (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
)
202 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED2
;
203 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
207 static void tcp_incr_quickack(struct sock
*sk
, unsigned int max_quickacks
)
209 struct inet_connection_sock
*icsk
= inet_csk(sk
);
210 unsigned int quickacks
= tcp_sk(sk
)->rcv_wnd
/ (2 * icsk
->icsk_ack
.rcv_mss
);
214 quickacks
= min(quickacks
, max_quickacks
);
215 if (quickacks
> icsk
->icsk_ack
.quick
)
216 icsk
->icsk_ack
.quick
= quickacks
;
219 void tcp_enter_quickack_mode(struct sock
*sk
, unsigned int max_quickacks
)
221 struct inet_connection_sock
*icsk
= inet_csk(sk
);
223 tcp_incr_quickack(sk
, max_quickacks
);
224 icsk
->icsk_ack
.pingpong
= 0;
225 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
227 EXPORT_SYMBOL(tcp_enter_quickack_mode
);
229 /* Send ACKs quickly, if "quick" count is not exhausted
230 * and the session is not interactive.
233 static bool tcp_in_quickack_mode(struct sock
*sk
)
235 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
236 const struct dst_entry
*dst
= __sk_dst_get(sk
);
238 return (dst
&& dst_metric(dst
, RTAX_QUICKACK
)) ||
239 (icsk
->icsk_ack
.quick
&& !icsk
->icsk_ack
.pingpong
);
242 static void tcp_ecn_queue_cwr(struct tcp_sock
*tp
)
244 if (tp
->ecn_flags
& TCP_ECN_OK
)
245 tp
->ecn_flags
|= TCP_ECN_QUEUE_CWR
;
248 static void tcp_ecn_accept_cwr(struct sock
*sk
, const struct sk_buff
*skb
)
250 if (tcp_hdr(skb
)->cwr
) {
251 tcp_sk(sk
)->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
253 /* If the sender is telling us it has entered CWR, then its
254 * cwnd may be very low (even just 1 packet), so we should ACK
257 inet_csk(sk
)->icsk_ack
.pending
|= ICSK_ACK_NOW
;
261 static void tcp_ecn_withdraw_cwr(struct tcp_sock
*tp
)
263 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
266 static void __tcp_ecn_check_ce(struct sock
*sk
, const struct sk_buff
*skb
)
268 struct tcp_sock
*tp
= tcp_sk(sk
);
270 switch (TCP_SKB_CB(skb
)->ip_dsfield
& INET_ECN_MASK
) {
271 case INET_ECN_NOT_ECT
:
272 /* Funny extension: if ECT is not set on a segment,
273 * and we already seen ECT on a previous segment,
274 * it is probably a retransmit.
276 if (tp
->ecn_flags
& TCP_ECN_SEEN
)
277 tcp_enter_quickack_mode(sk
, 2);
280 if (tcp_ca_needs_ecn(sk
))
281 tcp_ca_event(sk
, CA_EVENT_ECN_IS_CE
);
283 if (!(tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)) {
284 /* Better not delay acks, sender can have a very low cwnd */
285 tcp_enter_quickack_mode(sk
, 2);
286 tp
->ecn_flags
|= TCP_ECN_DEMAND_CWR
;
288 tp
->ecn_flags
|= TCP_ECN_SEEN
;
291 if (tcp_ca_needs_ecn(sk
))
292 tcp_ca_event(sk
, CA_EVENT_ECN_NO_CE
);
293 tp
->ecn_flags
|= TCP_ECN_SEEN
;
298 static void tcp_ecn_check_ce(struct sock
*sk
, const struct sk_buff
*skb
)
300 if (tcp_sk(sk
)->ecn_flags
& TCP_ECN_OK
)
301 __tcp_ecn_check_ce(sk
, skb
);
304 static void tcp_ecn_rcv_synack(struct tcp_sock
*tp
, const struct tcphdr
*th
)
306 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| th
->cwr
))
307 tp
->ecn_flags
&= ~TCP_ECN_OK
;
310 static void tcp_ecn_rcv_syn(struct tcp_sock
*tp
, const struct tcphdr
*th
)
312 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| !th
->cwr
))
313 tp
->ecn_flags
&= ~TCP_ECN_OK
;
316 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock
*tp
, const struct tcphdr
*th
)
318 if (th
->ece
&& !th
->syn
&& (tp
->ecn_flags
& TCP_ECN_OK
))
323 /* Buffer size and advertised window tuning.
325 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
328 static void tcp_sndbuf_expand(struct sock
*sk
)
330 const struct tcp_sock
*tp
= tcp_sk(sk
);
331 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
335 /* Worst case is non GSO/TSO : each frame consumes one skb
336 * and skb->head is kmalloced using power of two area of memory
338 per_mss
= max_t(u32
, tp
->rx_opt
.mss_clamp
, tp
->mss_cache
) +
340 SKB_DATA_ALIGN(sizeof(struct skb_shared_info
));
342 per_mss
= roundup_pow_of_two(per_mss
) +
343 SKB_DATA_ALIGN(sizeof(struct sk_buff
));
345 nr_segs
= max_t(u32
, TCP_INIT_CWND
, tp
->snd_cwnd
);
346 nr_segs
= max_t(u32
, nr_segs
, tp
->reordering
+ 1);
348 /* Fast Recovery (RFC 5681 3.2) :
349 * Cubic needs 1.7 factor, rounded to 2 to include
350 * extra cushion (application might react slowly to EPOLLOUT)
352 sndmem
= ca_ops
->sndbuf_expand
? ca_ops
->sndbuf_expand(sk
) : 2;
353 sndmem
*= nr_segs
* per_mss
;
355 if (sk
->sk_sndbuf
< sndmem
)
356 sk
->sk_sndbuf
= min(sndmem
, sock_net(sk
)->ipv4
.sysctl_tcp_wmem
[2]);
359 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
361 * All tcp_full_space() is split to two parts: "network" buffer, allocated
362 * forward and advertised in receiver window (tp->rcv_wnd) and
363 * "application buffer", required to isolate scheduling/application
364 * latencies from network.
365 * window_clamp is maximal advertised window. It can be less than
366 * tcp_full_space(), in this case tcp_full_space() - window_clamp
367 * is reserved for "application" buffer. The less window_clamp is
368 * the smoother our behaviour from viewpoint of network, but the lower
369 * throughput and the higher sensitivity of the connection to losses. 8)
371 * rcv_ssthresh is more strict window_clamp used at "slow start"
372 * phase to predict further behaviour of this connection.
373 * It is used for two goals:
374 * - to enforce header prediction at sender, even when application
375 * requires some significant "application buffer". It is check #1.
376 * - to prevent pruning of receive queue because of misprediction
377 * of receiver window. Check #2.
379 * The scheme does not work when sender sends good segments opening
380 * window and then starts to feed us spaghetti. But it should work
381 * in common situations. Otherwise, we have to rely on queue collapsing.
384 /* Slow part of check#2. */
385 static int __tcp_grow_window(const struct sock
*sk
, const struct sk_buff
*skb
)
387 struct tcp_sock
*tp
= tcp_sk(sk
);
389 int truesize
= tcp_win_from_space(sk
, skb
->truesize
) >> 1;
390 int window
= tcp_win_from_space(sk
, sock_net(sk
)->ipv4
.sysctl_tcp_rmem
[2]) >> 1;
392 while (tp
->rcv_ssthresh
<= window
) {
393 if (truesize
<= skb
->len
)
394 return 2 * inet_csk(sk
)->icsk_ack
.rcv_mss
;
402 static void tcp_grow_window(struct sock
*sk
, const struct sk_buff
*skb
)
404 struct tcp_sock
*tp
= tcp_sk(sk
);
407 if (tp
->rcv_ssthresh
< tp
->window_clamp
&&
408 (int)tp
->rcv_ssthresh
< tcp_space(sk
) &&
409 !tcp_under_memory_pressure(sk
)) {
412 /* Check #2. Increase window, if skb with such overhead
413 * will fit to rcvbuf in future.
415 if (tcp_win_from_space(sk
, skb
->truesize
) <= skb
->len
)
416 incr
= 2 * tp
->advmss
;
418 incr
= __tcp_grow_window(sk
, skb
);
421 incr
= max_t(int, incr
, 2 * skb
->len
);
422 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
+ incr
,
424 inet_csk(sk
)->icsk_ack
.quick
|= 1;
429 /* 3. Try to fixup all. It is made immediately after connection enters
432 void tcp_init_buffer_space(struct sock
*sk
)
434 int tcp_app_win
= sock_net(sk
)->ipv4
.sysctl_tcp_app_win
;
435 struct tcp_sock
*tp
= tcp_sk(sk
);
438 if (!(sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
))
439 tcp_sndbuf_expand(sk
);
441 tp
->rcvq_space
.space
= min_t(u32
, tp
->rcv_wnd
, TCP_INIT_CWND
* tp
->advmss
);
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 (tcp_app_win
&& maxwin
> 4 * tp
->advmss
)
452 tp
->window_clamp
= max(maxwin
-
453 (maxwin
>> tcp_app_win
),
457 /* Force reservation of one segment. */
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 /* 4. 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
);
472 struct net
*net
= sock_net(sk
);
474 icsk
->icsk_ack
.quick
= 0;
476 if (sk
->sk_rcvbuf
< net
->ipv4
.sysctl_tcp_rmem
[2] &&
477 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
) &&
478 !tcp_under_memory_pressure(sk
) &&
479 sk_memory_allocated(sk
) < sk_prot_mem_limits(sk
, 0)) {
480 sk
->sk_rcvbuf
= min(atomic_read(&sk
->sk_rmem_alloc
),
481 net
->ipv4
.sysctl_tcp_rmem
[2]);
483 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
)
484 tp
->rcv_ssthresh
= min(tp
->window_clamp
, 2U * tp
->advmss
);
487 /* Initialize RCV_MSS value.
488 * RCV_MSS is an our guess about MSS used by the peer.
489 * We haven't any direct information about the MSS.
490 * It's better to underestimate the RCV_MSS rather than overestimate.
491 * Overestimations make us ACKing less frequently than needed.
492 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
494 void tcp_initialize_rcv_mss(struct sock
*sk
)
496 const struct tcp_sock
*tp
= tcp_sk(sk
);
497 unsigned int hint
= min_t(unsigned int, tp
->advmss
, tp
->mss_cache
);
499 hint
= min(hint
, tp
->rcv_wnd
/ 2);
500 hint
= min(hint
, TCP_MSS_DEFAULT
);
501 hint
= max(hint
, TCP_MIN_MSS
);
503 inet_csk(sk
)->icsk_ack
.rcv_mss
= hint
;
505 EXPORT_SYMBOL(tcp_initialize_rcv_mss
);
507 /* Receiver "autotuning" code.
509 * The algorithm for RTT estimation w/o timestamps is based on
510 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
511 * <http://public.lanl.gov/radiant/pubs.html#DRS>
513 * More detail on this code can be found at
514 * <http://staff.psc.edu/jheffner/>,
515 * though this reference is out of date. A new paper
518 static void tcp_rcv_rtt_update(struct tcp_sock
*tp
, u32 sample
, int win_dep
)
520 u32 new_sample
= tp
->rcv_rtt_est
.rtt_us
;
523 if (new_sample
!= 0) {
524 /* If we sample in larger samples in the non-timestamp
525 * case, we could grossly overestimate the RTT especially
526 * with chatty applications or bulk transfer apps which
527 * are stalled on filesystem I/O.
529 * Also, since we are only going for a minimum in the
530 * non-timestamp case, we do not smooth things out
531 * else with timestamps disabled convergence takes too
535 m
-= (new_sample
>> 3);
543 /* No previous measure. */
547 tp
->rcv_rtt_est
.rtt_us
= new_sample
;
550 static inline void tcp_rcv_rtt_measure(struct tcp_sock
*tp
)
554 if (tp
->rcv_rtt_est
.time
== 0)
556 if (before(tp
->rcv_nxt
, tp
->rcv_rtt_est
.seq
))
558 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
== tp
->rcv_rtt_last_tsecr
)
575 tp
->rcv_rtt_last_tsecr
= tp
->rx_opt
.rcv_tsecr
;
577 if (TCP_SKB_CB(skb
)->end_seq
-
578 TCP_SKB_CB(skb
)->seq
>= inet_csk(sk
)->icsk_ack
.rcv_mss
) {
579 u32 delta
= tcp_time_stamp(tp
) - tp
->rx_opt
.rcv_tsecr
;
584 delta_us
= delta
* (USEC_PER_SEC
/ TCP_TS_HZ
);
585 tcp_rcv_rtt_update(tp
, delta_us
, 0);
590 * This function should be called every time data is copied to user space.
591 * It calculates the appropriate TCP receive buffer space.
593 void tcp_rcv_space_adjust(struct sock
*sk
)
595 struct tcp_sock
*tp
= tcp_sk(sk
);
599 trace_tcp_rcv_space_adjust(sk
);
601 tcp_mstamp_refresh(tp
);
602 time
= tcp_stamp_us_delta(tp
->tcp_mstamp
, tp
->rcvq_space
.time
);
603 if (time
< (tp
->rcv_rtt_est
.rtt_us
>> 3) || tp
->rcv_rtt_est
.rtt_us
== 0)
606 /* Number of bytes copied to user in last RTT */
607 copied
= tp
->copied_seq
- tp
->rcvq_space
.seq
;
608 if (copied
<= tp
->rcvq_space
.space
)
612 * copied = bytes received in previous RTT, our base window
613 * To cope with packet losses, we need a 2x factor
614 * To cope with slow start, and sender growing its cwin by 100 %
615 * every RTT, we need a 4x factor, because the ACK we are sending
616 * now is for the next RTT, not the current one :
617 * <prev RTT . ><current RTT .. ><next RTT .... >
620 if (sock_net(sk
)->ipv4
.sysctl_tcp_moderate_rcvbuf
&&
621 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
)) {
625 /* minimal window to cope with packet losses, assuming
626 * steady state. Add some cushion because of small variations.
628 rcvwin
= ((u64
)copied
<< 1) + 16 * tp
->advmss
;
630 /* Accommodate for sender rate increase (eg. slow start) */
631 grow
= rcvwin
* (copied
- tp
->rcvq_space
.space
);
632 do_div(grow
, tp
->rcvq_space
.space
);
633 rcvwin
+= (grow
<< 1);
635 rcvmem
= SKB_TRUESIZE(tp
->advmss
+ MAX_TCP_HEADER
);
636 while (tcp_win_from_space(sk
, rcvmem
) < tp
->advmss
)
639 do_div(rcvwin
, tp
->advmss
);
640 rcvbuf
= min_t(u64
, rcvwin
* rcvmem
,
641 sock_net(sk
)->ipv4
.sysctl_tcp_rmem
[2]);
642 if (rcvbuf
> sk
->sk_rcvbuf
) {
643 sk
->sk_rcvbuf
= rcvbuf
;
645 /* Make the window clamp follow along. */
646 tp
->window_clamp
= tcp_win_from_space(sk
, rcvbuf
);
649 tp
->rcvq_space
.space
= copied
;
652 tp
->rcvq_space
.seq
= tp
->copied_seq
;
653 tp
->rcvq_space
.time
= tp
->tcp_mstamp
;
656 /* There is something which you must keep in mind when you analyze the
657 * behavior of the tp->ato delayed ack timeout interval. When a
658 * connection starts up, we want to ack as quickly as possible. The
659 * problem is that "good" TCP's do slow start at the beginning of data
660 * transmission. The means that until we send the first few ACK's the
661 * sender will sit on his end and only queue most of his data, because
662 * he can only send snd_cwnd unacked packets at any given time. For
663 * each ACK we send, he increments snd_cwnd and transmits more of his
666 static void tcp_event_data_recv(struct sock
*sk
, struct sk_buff
*skb
)
668 struct tcp_sock
*tp
= tcp_sk(sk
);
669 struct inet_connection_sock
*icsk
= inet_csk(sk
);
672 inet_csk_schedule_ack(sk
);
674 tcp_measure_rcv_mss(sk
, skb
);
676 tcp_rcv_rtt_measure(tp
);
680 if (!icsk
->icsk_ack
.ato
) {
681 /* The _first_ data packet received, initialize
682 * delayed ACK engine.
684 tcp_incr_quickack(sk
, TCP_MAX_QUICKACKS
);
685 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
687 int m
= now
- icsk
->icsk_ack
.lrcvtime
;
689 if (m
<= TCP_ATO_MIN
/ 2) {
690 /* The fastest case is the first. */
691 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + TCP_ATO_MIN
/ 2;
692 } else if (m
< icsk
->icsk_ack
.ato
) {
693 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + m
;
694 if (icsk
->icsk_ack
.ato
> icsk
->icsk_rto
)
695 icsk
->icsk_ack
.ato
= icsk
->icsk_rto
;
696 } else if (m
> icsk
->icsk_rto
) {
697 /* Too long gap. Apparently sender failed to
698 * restart window, so that we send ACKs quickly.
700 tcp_incr_quickack(sk
, TCP_MAX_QUICKACKS
);
704 icsk
->icsk_ack
.lrcvtime
= now
;
706 tcp_ecn_check_ce(sk
, skb
);
709 tcp_grow_window(sk
, skb
);
712 /* Called to compute a smoothed rtt estimate. The data fed to this
713 * routine either comes from timestamps, or from segments that were
714 * known _not_ to have been retransmitted [see Karn/Partridge
715 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
716 * piece by Van Jacobson.
717 * NOTE: the next three routines used to be one big routine.
718 * To save cycles in the RFC 1323 implementation it was better to break
719 * it up into three procedures. -- erics
721 static void tcp_rtt_estimator(struct sock
*sk
, long mrtt_us
)
723 struct tcp_sock
*tp
= tcp_sk(sk
);
724 long m
= mrtt_us
; /* RTT */
725 u32 srtt
= tp
->srtt_us
;
727 /* The following amusing code comes from Jacobson's
728 * article in SIGCOMM '88. Note that rtt and mdev
729 * are scaled versions of rtt and mean deviation.
730 * This is designed to be as fast as possible
731 * m stands for "measurement".
733 * On a 1990 paper the rto value is changed to:
734 * RTO = rtt + 4 * mdev
736 * Funny. This algorithm seems to be very broken.
737 * These formulae increase RTO, when it should be decreased, increase
738 * too slowly, when it should be increased quickly, decrease too quickly
739 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
740 * does not matter how to _calculate_ it. Seems, it was trap
741 * that VJ failed to avoid. 8)
744 m
-= (srtt
>> 3); /* m is now error in rtt est */
745 srtt
+= m
; /* rtt = 7/8 rtt + 1/8 new */
747 m
= -m
; /* m is now abs(error) */
748 m
-= (tp
->mdev_us
>> 2); /* similar update on mdev */
749 /* This is similar to one of Eifel findings.
750 * Eifel blocks mdev updates when rtt decreases.
751 * This solution is a bit different: we use finer gain
752 * for mdev in this case (alpha*beta).
753 * Like Eifel it also prevents growth of rto,
754 * but also it limits too fast rto decreases,
755 * happening in pure Eifel.
760 m
-= (tp
->mdev_us
>> 2); /* similar update on mdev */
762 tp
->mdev_us
+= m
; /* mdev = 3/4 mdev + 1/4 new */
763 if (tp
->mdev_us
> tp
->mdev_max_us
) {
764 tp
->mdev_max_us
= tp
->mdev_us
;
765 if (tp
->mdev_max_us
> tp
->rttvar_us
)
766 tp
->rttvar_us
= tp
->mdev_max_us
;
768 if (after(tp
->snd_una
, tp
->rtt_seq
)) {
769 if (tp
->mdev_max_us
< tp
->rttvar_us
)
770 tp
->rttvar_us
-= (tp
->rttvar_us
- tp
->mdev_max_us
) >> 2;
771 tp
->rtt_seq
= tp
->snd_nxt
;
772 tp
->mdev_max_us
= tcp_rto_min_us(sk
);
775 /* no previous measure. */
776 srtt
= m
<< 3; /* take the measured time to be rtt */
777 tp
->mdev_us
= m
<< 1; /* make sure rto = 3*rtt */
778 tp
->rttvar_us
= max(tp
->mdev_us
, tcp_rto_min_us(sk
));
779 tp
->mdev_max_us
= tp
->rttvar_us
;
780 tp
->rtt_seq
= tp
->snd_nxt
;
782 tp
->srtt_us
= max(1U, srtt
);
785 static void tcp_update_pacing_rate(struct sock
*sk
)
787 const struct tcp_sock
*tp
= tcp_sk(sk
);
790 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
791 rate
= (u64
)tp
->mss_cache
* ((USEC_PER_SEC
/ 100) << 3);
793 /* current rate is (cwnd * mss) / srtt
794 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
795 * In Congestion Avoidance phase, set it to 120 % the current rate.
797 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
798 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
799 * end of slow start and should slow down.
801 if (tp
->snd_cwnd
< tp
->snd_ssthresh
/ 2)
802 rate
*= sock_net(sk
)->ipv4
.sysctl_tcp_pacing_ss_ratio
;
804 rate
*= sock_net(sk
)->ipv4
.sysctl_tcp_pacing_ca_ratio
;
806 rate
*= max(tp
->snd_cwnd
, tp
->packets_out
);
808 if (likely(tp
->srtt_us
))
809 do_div(rate
, tp
->srtt_us
);
811 /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate
812 * without any lock. We want to make sure compiler wont store
813 * intermediate values in this location.
815 WRITE_ONCE(sk
->sk_pacing_rate
, min_t(u64
, rate
,
816 sk
->sk_max_pacing_rate
));
819 /* Calculate rto without backoff. This is the second half of Van Jacobson's
820 * routine referred to above.
822 static void tcp_set_rto(struct sock
*sk
)
824 const struct tcp_sock
*tp
= tcp_sk(sk
);
825 /* Old crap is replaced with new one. 8)
828 * 1. If rtt variance happened to be less 50msec, it is hallucination.
829 * It cannot be less due to utterly erratic ACK generation made
830 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
831 * to do with delayed acks, because at cwnd>2 true delack timeout
832 * is invisible. Actually, Linux-2.4 also generates erratic
833 * ACKs in some circumstances.
835 inet_csk(sk
)->icsk_rto
= __tcp_set_rto(tp
);
837 /* 2. Fixups made earlier cannot be right.
838 * If we do not estimate RTO correctly without them,
839 * all the algo is pure shit and should be replaced
840 * with correct one. It is exactly, which we pretend to do.
843 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
844 * guarantees that rto is higher.
849 __u32
tcp_init_cwnd(const struct tcp_sock
*tp
, const struct dst_entry
*dst
)
851 __u32 cwnd
= (dst
? dst_metric(dst
, RTAX_INITCWND
) : 0);
854 cwnd
= TCP_INIT_CWND
;
855 return min_t(__u32
, cwnd
, tp
->snd_cwnd_clamp
);
858 /* Take a notice that peer is sending D-SACKs */
859 static void tcp_dsack_seen(struct tcp_sock
*tp
)
861 tp
->rx_opt
.sack_ok
|= TCP_DSACK_SEEN
;
862 tp
->rack
.dsack_seen
= 1;
866 /* It's reordering when higher sequence was delivered (i.e. sacked) before
867 * some lower never-retransmitted sequence ("low_seq"). The maximum reordering
868 * distance is approximated in full-mss packet distance ("reordering").
870 static void tcp_check_sack_reordering(struct sock
*sk
, const u32 low_seq
,
873 struct tcp_sock
*tp
= tcp_sk(sk
);
874 const u32 mss
= tp
->mss_cache
;
877 fack
= tcp_highest_sack_seq(tp
);
878 if (!before(low_seq
, fack
))
881 metric
= fack
- low_seq
;
882 if ((metric
> tp
->reordering
* mss
) && mss
) {
883 #if FASTRETRANS_DEBUG > 1
884 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
885 tp
->rx_opt
.sack_ok
, inet_csk(sk
)->icsk_ca_state
,
889 tp
->undo_marker
? tp
->undo_retrans
: 0);
891 tp
->reordering
= min_t(u32
, (metric
+ mss
- 1) / mss
,
892 sock_net(sk
)->ipv4
.sysctl_tcp_max_reordering
);
895 /* This exciting event is worth to be remembered. 8) */
897 NET_INC_STATS(sock_net(sk
),
898 ts
? LINUX_MIB_TCPTSREORDER
: LINUX_MIB_TCPSACKREORDER
);
901 /* This must be called before lost_out is incremented */
902 static void tcp_verify_retransmit_hint(struct tcp_sock
*tp
, struct sk_buff
*skb
)
904 if (!tp
->retransmit_skb_hint
||
905 before(TCP_SKB_CB(skb
)->seq
,
906 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
907 tp
->retransmit_skb_hint
= skb
;
910 /* Sum the number of packets on the wire we have marked as lost.
911 * There are two cases we care about here:
912 * a) Packet hasn't been marked lost (nor retransmitted),
913 * and this is the first loss.
914 * b) Packet has been marked both lost and retransmitted,
915 * and this means we think it was lost again.
917 static void tcp_sum_lost(struct tcp_sock
*tp
, struct sk_buff
*skb
)
919 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
921 if (!(sacked
& TCPCB_LOST
) ||
922 ((sacked
& TCPCB_LOST
) && (sacked
& TCPCB_SACKED_RETRANS
)))
923 tp
->lost
+= tcp_skb_pcount(skb
);
926 static void tcp_skb_mark_lost(struct tcp_sock
*tp
, struct sk_buff
*skb
)
928 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
929 tcp_verify_retransmit_hint(tp
, skb
);
931 tp
->lost_out
+= tcp_skb_pcount(skb
);
932 tcp_sum_lost(tp
, skb
);
933 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
937 void tcp_skb_mark_lost_uncond_verify(struct tcp_sock
*tp
, struct sk_buff
*skb
)
939 tcp_verify_retransmit_hint(tp
, skb
);
941 tcp_sum_lost(tp
, skb
);
942 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
943 tp
->lost_out
+= tcp_skb_pcount(skb
);
944 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
948 /* This procedure tags the retransmission queue when SACKs arrive.
950 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
951 * Packets in queue with these bits set are counted in variables
952 * sacked_out, retrans_out and lost_out, correspondingly.
954 * Valid combinations are:
955 * Tag InFlight Description
956 * 0 1 - orig segment is in flight.
957 * S 0 - nothing flies, orig reached receiver.
958 * L 0 - nothing flies, orig lost by net.
959 * R 2 - both orig and retransmit are in flight.
960 * L|R 1 - orig is lost, retransmit is in flight.
961 * S|R 1 - orig reached receiver, retrans is still in flight.
962 * (L|S|R is logically valid, it could occur when L|R is sacked,
963 * but it is equivalent to plain S and code short-curcuits it to S.
964 * L|S is logically invalid, it would mean -1 packet in flight 8))
966 * These 6 states form finite state machine, controlled by the following events:
967 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
968 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
969 * 3. Loss detection event of two flavors:
970 * A. Scoreboard estimator decided the packet is lost.
971 * A'. Reno "three dupacks" marks head of queue lost.
972 * B. SACK arrives sacking SND.NXT at the moment, when the
973 * segment was retransmitted.
974 * 4. D-SACK added new rule: D-SACK changes any tag to S.
976 * It is pleasant to note, that state diagram turns out to be commutative,
977 * so that we are allowed not to be bothered by order of our actions,
978 * when multiple events arrive simultaneously. (see the function below).
980 * Reordering detection.
981 * --------------------
982 * Reordering metric is maximal distance, which a packet can be displaced
983 * in packet stream. With SACKs we can estimate it:
985 * 1. SACK fills old hole and the corresponding segment was not
986 * ever retransmitted -> reordering. Alas, we cannot use it
987 * when segment was retransmitted.
988 * 2. The last flaw is solved with D-SACK. D-SACK arrives
989 * for retransmitted and already SACKed segment -> reordering..
990 * Both of these heuristics are not used in Loss state, when we cannot
991 * account for retransmits accurately.
993 * SACK block validation.
994 * ----------------------
996 * SACK block range validation checks that the received SACK block fits to
997 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
998 * Note that SND.UNA is not included to the range though being valid because
999 * it means that the receiver is rather inconsistent with itself reporting
1000 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1001 * perfectly valid, however, in light of RFC2018 which explicitly states
1002 * that "SACK block MUST reflect the newest segment. Even if the newest
1003 * segment is going to be discarded ...", not that it looks very clever
1004 * in case of head skb. Due to potentional receiver driven attacks, we
1005 * choose to avoid immediate execution of a walk in write queue due to
1006 * reneging and defer head skb's loss recovery to standard loss recovery
1007 * procedure that will eventually trigger (nothing forbids us doing this).
1009 * Implements also blockage to start_seq wrap-around. Problem lies in the
1010 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1011 * there's no guarantee that it will be before snd_nxt (n). The problem
1012 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1015 * <- outs wnd -> <- wrapzone ->
1016 * u e n u_w e_w s n_w
1018 * |<------------+------+----- TCP seqno space --------------+---------->|
1019 * ...-- <2^31 ->| |<--------...
1020 * ...---- >2^31 ------>| |<--------...
1022 * Current code wouldn't be vulnerable but it's better still to discard such
1023 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1024 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1025 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1026 * equal to the ideal case (infinite seqno space without wrap caused issues).
1028 * With D-SACK the lower bound is extended to cover sequence space below
1029 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1030 * again, D-SACK block must not to go across snd_una (for the same reason as
1031 * for the normal SACK blocks, explained above). But there all simplicity
1032 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1033 * fully below undo_marker they do not affect behavior in anyway and can
1034 * therefore be safely ignored. In rare cases (which are more or less
1035 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1036 * fragmentation and packet reordering past skb's retransmission. To consider
1037 * them correctly, the acceptable range must be extended even more though
1038 * the exact amount is rather hard to quantify. However, tp->max_window can
1039 * be used as an exaggerated estimate.
1041 static bool tcp_is_sackblock_valid(struct tcp_sock
*tp
, bool is_dsack
,
1042 u32 start_seq
, u32 end_seq
)
1044 /* Too far in future, or reversed (interpretation is ambiguous) */
1045 if (after(end_seq
, tp
->snd_nxt
) || !before(start_seq
, end_seq
))
1048 /* Nasty start_seq wrap-around check (see comments above) */
1049 if (!before(start_seq
, tp
->snd_nxt
))
1052 /* In outstanding window? ...This is valid exit for D-SACKs too.
1053 * start_seq == snd_una is non-sensical (see comments above)
1055 if (after(start_seq
, tp
->snd_una
))
1058 if (!is_dsack
|| !tp
->undo_marker
)
1061 /* ...Then it's D-SACK, and must reside below snd_una completely */
1062 if (after(end_seq
, tp
->snd_una
))
1065 if (!before(start_seq
, tp
->undo_marker
))
1069 if (!after(end_seq
, tp
->undo_marker
))
1072 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1073 * start_seq < undo_marker and end_seq >= undo_marker.
1075 return !before(start_seq
, end_seq
- tp
->max_window
);
1078 static bool tcp_check_dsack(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1079 struct tcp_sack_block_wire
*sp
, int num_sacks
,
1082 struct tcp_sock
*tp
= tcp_sk(sk
);
1083 u32 start_seq_0
= get_unaligned_be32(&sp
[0].start_seq
);
1084 u32 end_seq_0
= get_unaligned_be32(&sp
[0].end_seq
);
1085 bool dup_sack
= false;
1087 if (before(start_seq_0
, TCP_SKB_CB(ack_skb
)->ack_seq
)) {
1090 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPDSACKRECV
);
1091 } else if (num_sacks
> 1) {
1092 u32 end_seq_1
= get_unaligned_be32(&sp
[1].end_seq
);
1093 u32 start_seq_1
= get_unaligned_be32(&sp
[1].start_seq
);
1095 if (!after(end_seq_0
, end_seq_1
) &&
1096 !before(start_seq_0
, start_seq_1
)) {
1099 NET_INC_STATS(sock_net(sk
),
1100 LINUX_MIB_TCPDSACKOFORECV
);
1104 /* D-SACK for already forgotten data... Do dumb counting. */
1105 if (dup_sack
&& tp
->undo_marker
&& tp
->undo_retrans
> 0 &&
1106 !after(end_seq_0
, prior_snd_una
) &&
1107 after(end_seq_0
, tp
->undo_marker
))
1113 struct tcp_sacktag_state
{
1115 /* Timestamps for earliest and latest never-retransmitted segment
1116 * that was SACKed. RTO needs the earliest RTT to stay conservative,
1117 * but congestion control should still get an accurate delay signal.
1121 struct rate_sample
*rate
;
1123 unsigned int mss_now
;
1126 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1127 * the incoming SACK may not exactly match but we can find smaller MSS
1128 * aligned portion of it that matches. Therefore we might need to fragment
1129 * which may fail and creates some hassle (caller must handle error case
1132 * FIXME: this could be merged to shift decision code
1134 static int tcp_match_skb_to_sack(struct sock
*sk
, struct sk_buff
*skb
,
1135 u32 start_seq
, u32 end_seq
)
1139 unsigned int pkt_len
;
1142 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1143 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1145 if (tcp_skb_pcount(skb
) > 1 && !in_sack
&&
1146 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
1147 mss
= tcp_skb_mss(skb
);
1148 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1151 pkt_len
= start_seq
- TCP_SKB_CB(skb
)->seq
;
1155 pkt_len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1160 /* Round if necessary so that SACKs cover only full MSSes
1161 * and/or the remaining small portion (if present)
1163 if (pkt_len
> mss
) {
1164 unsigned int new_len
= (pkt_len
/ mss
) * mss
;
1165 if (!in_sack
&& new_len
< pkt_len
)
1170 if (pkt_len
>= skb
->len
&& !in_sack
)
1173 err
= tcp_fragment(sk
, TCP_FRAG_IN_RTX_QUEUE
, skb
,
1174 pkt_len
, mss
, GFP_ATOMIC
);
1182 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1183 static u8
tcp_sacktag_one(struct sock
*sk
,
1184 struct tcp_sacktag_state
*state
, u8 sacked
,
1185 u32 start_seq
, u32 end_seq
,
1186 int dup_sack
, int pcount
,
1189 struct tcp_sock
*tp
= tcp_sk(sk
);
1191 /* Account D-SACK for retransmitted packet. */
1192 if (dup_sack
&& (sacked
& TCPCB_RETRANS
)) {
1193 if (tp
->undo_marker
&& tp
->undo_retrans
> 0 &&
1194 after(end_seq
, tp
->undo_marker
))
1196 if ((sacked
& TCPCB_SACKED_ACKED
) &&
1197 before(start_seq
, state
->reord
))
1198 state
->reord
= start_seq
;
1201 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1202 if (!after(end_seq
, tp
->snd_una
))
1205 if (!(sacked
& TCPCB_SACKED_ACKED
)) {
1206 tcp_rack_advance(tp
, sacked
, end_seq
, xmit_time
);
1208 if (sacked
& TCPCB_SACKED_RETRANS
) {
1209 /* If the segment is not tagged as lost,
1210 * we do not clear RETRANS, believing
1211 * that retransmission is still in flight.
1213 if (sacked
& TCPCB_LOST
) {
1214 sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1215 tp
->lost_out
-= pcount
;
1216 tp
->retrans_out
-= pcount
;
1219 if (!(sacked
& TCPCB_RETRANS
)) {
1220 /* New sack for not retransmitted frame,
1221 * which was in hole. It is reordering.
1223 if (before(start_seq
,
1224 tcp_highest_sack_seq(tp
)) &&
1225 before(start_seq
, state
->reord
))
1226 state
->reord
= start_seq
;
1228 if (!after(end_seq
, tp
->high_seq
))
1229 state
->flag
|= FLAG_ORIG_SACK_ACKED
;
1230 if (state
->first_sackt
== 0)
1231 state
->first_sackt
= xmit_time
;
1232 state
->last_sackt
= xmit_time
;
1235 if (sacked
& TCPCB_LOST
) {
1236 sacked
&= ~TCPCB_LOST
;
1237 tp
->lost_out
-= pcount
;
1241 sacked
|= TCPCB_SACKED_ACKED
;
1242 state
->flag
|= FLAG_DATA_SACKED
;
1243 tp
->sacked_out
+= pcount
;
1244 tp
->delivered
+= pcount
; /* Out-of-order packets delivered */
1246 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1247 if (tp
->lost_skb_hint
&&
1248 before(start_seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
))
1249 tp
->lost_cnt_hint
+= pcount
;
1252 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1253 * frames and clear it. undo_retrans is decreased above, L|R frames
1254 * are accounted above as well.
1256 if (dup_sack
&& (sacked
& TCPCB_SACKED_RETRANS
)) {
1257 sacked
&= ~TCPCB_SACKED_RETRANS
;
1258 tp
->retrans_out
-= pcount
;
1264 /* Shift newly-SACKed bytes from this skb to the immediately previous
1265 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1267 static bool tcp_shifted_skb(struct sock
*sk
, struct sk_buff
*prev
,
1268 struct sk_buff
*skb
,
1269 struct tcp_sacktag_state
*state
,
1270 unsigned int pcount
, int shifted
, int mss
,
1273 struct tcp_sock
*tp
= tcp_sk(sk
);
1274 u32 start_seq
= TCP_SKB_CB(skb
)->seq
; /* start of newly-SACKed */
1275 u32 end_seq
= start_seq
+ shifted
; /* end of newly-SACKed */
1279 /* Adjust counters and hints for the newly sacked sequence
1280 * range but discard the return value since prev is already
1281 * marked. We must tag the range first because the seq
1282 * advancement below implicitly advances
1283 * tcp_highest_sack_seq() when skb is highest_sack.
1285 tcp_sacktag_one(sk
, state
, TCP_SKB_CB(skb
)->sacked
,
1286 start_seq
, end_seq
, dup_sack
, pcount
,
1287 tcp_skb_timestamp_us(skb
));
1288 tcp_rate_skb_delivered(sk
, skb
, state
->rate
);
1290 if (skb
== tp
->lost_skb_hint
)
1291 tp
->lost_cnt_hint
+= pcount
;
1293 TCP_SKB_CB(prev
)->end_seq
+= shifted
;
1294 TCP_SKB_CB(skb
)->seq
+= shifted
;
1296 tcp_skb_pcount_add(prev
, pcount
);
1297 BUG_ON(tcp_skb_pcount(skb
) < pcount
);
1298 tcp_skb_pcount_add(skb
, -pcount
);
1300 /* When we're adding to gso_segs == 1, gso_size will be zero,
1301 * in theory this shouldn't be necessary but as long as DSACK
1302 * code can come after this skb later on it's better to keep
1303 * setting gso_size to something.
1305 if (!TCP_SKB_CB(prev
)->tcp_gso_size
)
1306 TCP_SKB_CB(prev
)->tcp_gso_size
= mss
;
1308 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1309 if (tcp_skb_pcount(skb
) <= 1)
1310 TCP_SKB_CB(skb
)->tcp_gso_size
= 0;
1312 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1313 TCP_SKB_CB(prev
)->sacked
|= (TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
);
1316 BUG_ON(!tcp_skb_pcount(skb
));
1317 NET_INC_STATS(sock_net(sk
), LINUX_MIB_SACKSHIFTED
);
1321 /* Whole SKB was eaten :-) */
1323 if (skb
== tp
->retransmit_skb_hint
)
1324 tp
->retransmit_skb_hint
= prev
;
1325 if (skb
== tp
->lost_skb_hint
) {
1326 tp
->lost_skb_hint
= prev
;
1327 tp
->lost_cnt_hint
-= tcp_skb_pcount(prev
);
1330 TCP_SKB_CB(prev
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
;
1331 TCP_SKB_CB(prev
)->eor
= TCP_SKB_CB(skb
)->eor
;
1332 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
1333 TCP_SKB_CB(prev
)->end_seq
++;
1335 if (skb
== tcp_highest_sack(sk
))
1336 tcp_advance_highest_sack(sk
, skb
);
1338 tcp_skb_collapse_tstamp(prev
, skb
);
1339 if (unlikely(TCP_SKB_CB(prev
)->tx
.delivered_mstamp
))
1340 TCP_SKB_CB(prev
)->tx
.delivered_mstamp
= 0;
1342 tcp_rtx_queue_unlink_and_free(skb
, sk
);
1344 NET_INC_STATS(sock_net(sk
), LINUX_MIB_SACKMERGED
);
1349 /* I wish gso_size would have a bit more sane initialization than
1350 * something-or-zero which complicates things
1352 static int tcp_skb_seglen(const struct sk_buff
*skb
)
1354 return tcp_skb_pcount(skb
) == 1 ? skb
->len
: tcp_skb_mss(skb
);
1357 /* Shifting pages past head area doesn't work */
1358 static int skb_can_shift(const struct sk_buff
*skb
)
1360 return !skb_headlen(skb
) && skb_is_nonlinear(skb
);
1363 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1366 static struct sk_buff
*tcp_shift_skb_data(struct sock
*sk
, struct sk_buff
*skb
,
1367 struct tcp_sacktag_state
*state
,
1368 u32 start_seq
, u32 end_seq
,
1371 struct tcp_sock
*tp
= tcp_sk(sk
);
1372 struct sk_buff
*prev
;
1378 /* Normally R but no L won't result in plain S */
1380 (TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_RETRANS
)) == TCPCB_SACKED_RETRANS
)
1382 if (!skb_can_shift(skb
))
1384 /* This frame is about to be dropped (was ACKed). */
1385 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1388 /* Can only happen with delayed DSACK + discard craziness */
1389 prev
= skb_rb_prev(skb
);
1393 if ((TCP_SKB_CB(prev
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
)
1396 if (!tcp_skb_can_collapse_to(prev
))
1399 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1400 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1404 pcount
= tcp_skb_pcount(skb
);
1405 mss
= tcp_skb_seglen(skb
);
1407 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1408 * drop this restriction as unnecessary
1410 if (mss
!= tcp_skb_seglen(prev
))
1413 if (!after(TCP_SKB_CB(skb
)->end_seq
, start_seq
))
1415 /* CHECKME: This is non-MSS split case only?, this will
1416 * cause skipped skbs due to advancing loop btw, original
1417 * has that feature too
1419 if (tcp_skb_pcount(skb
) <= 1)
1422 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1424 /* TODO: head merge to next could be attempted here
1425 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1426 * though it might not be worth of the additional hassle
1428 * ...we can probably just fallback to what was done
1429 * previously. We could try merging non-SACKed ones
1430 * as well but it probably isn't going to buy off
1431 * because later SACKs might again split them, and
1432 * it would make skb timestamp tracking considerably
1438 len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1440 BUG_ON(len
> skb
->len
);
1442 /* MSS boundaries should be honoured or else pcount will
1443 * severely break even though it makes things bit trickier.
1444 * Optimize common case to avoid most of the divides
1446 mss
= tcp_skb_mss(skb
);
1448 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1449 * drop this restriction as unnecessary
1451 if (mss
!= tcp_skb_seglen(prev
))
1456 } else if (len
< mss
) {
1464 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1465 if (!after(TCP_SKB_CB(skb
)->seq
+ len
, tp
->snd_una
))
1468 if (!skb_shift(prev
, skb
, len
))
1470 if (!tcp_shifted_skb(sk
, prev
, skb
, state
, pcount
, len
, mss
, dup_sack
))
1473 /* Hole filled allows collapsing with the next as well, this is very
1474 * useful when hole on every nth skb pattern happens
1476 skb
= skb_rb_next(prev
);
1480 if (!skb_can_shift(skb
) ||
1481 ((TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
) ||
1482 (mss
!= tcp_skb_seglen(skb
)))
1486 if (skb_shift(prev
, skb
, len
)) {
1487 pcount
+= tcp_skb_pcount(skb
);
1488 tcp_shifted_skb(sk
, prev
, skb
, state
, tcp_skb_pcount(skb
),
1499 NET_INC_STATS(sock_net(sk
), LINUX_MIB_SACKSHIFTFALLBACK
);
1503 static struct sk_buff
*tcp_sacktag_walk(struct sk_buff
*skb
, struct sock
*sk
,
1504 struct tcp_sack_block
*next_dup
,
1505 struct tcp_sacktag_state
*state
,
1506 u32 start_seq
, u32 end_seq
,
1509 struct tcp_sock
*tp
= tcp_sk(sk
);
1510 struct sk_buff
*tmp
;
1512 skb_rbtree_walk_from(skb
) {
1514 bool dup_sack
= dup_sack_in
;
1516 /* queue is in-order => we can short-circuit the walk early */
1517 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1521 before(TCP_SKB_CB(skb
)->seq
, next_dup
->end_seq
)) {
1522 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1523 next_dup
->start_seq
,
1529 /* skb reference here is a bit tricky to get right, since
1530 * shifting can eat and free both this skb and the next,
1531 * so not even _safe variant of the loop is enough.
1534 tmp
= tcp_shift_skb_data(sk
, skb
, state
,
1535 start_seq
, end_seq
, dup_sack
);
1544 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1550 if (unlikely(in_sack
< 0))
1554 TCP_SKB_CB(skb
)->sacked
=
1557 TCP_SKB_CB(skb
)->sacked
,
1558 TCP_SKB_CB(skb
)->seq
,
1559 TCP_SKB_CB(skb
)->end_seq
,
1561 tcp_skb_pcount(skb
),
1562 tcp_skb_timestamp_us(skb
));
1563 tcp_rate_skb_delivered(sk
, skb
, state
->rate
);
1564 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
1565 list_del_init(&skb
->tcp_tsorted_anchor
);
1567 if (!before(TCP_SKB_CB(skb
)->seq
,
1568 tcp_highest_sack_seq(tp
)))
1569 tcp_advance_highest_sack(sk
, skb
);
1575 static struct sk_buff
*tcp_sacktag_bsearch(struct sock
*sk
,
1576 struct tcp_sacktag_state
*state
,
1579 struct rb_node
*parent
, **p
= &sk
->tcp_rtx_queue
.rb_node
;
1580 struct sk_buff
*skb
;
1584 skb
= rb_to_skb(parent
);
1585 if (before(seq
, TCP_SKB_CB(skb
)->seq
)) {
1586 p
= &parent
->rb_left
;
1589 if (!before(seq
, TCP_SKB_CB(skb
)->end_seq
)) {
1590 p
= &parent
->rb_right
;
1598 static struct sk_buff
*tcp_sacktag_skip(struct sk_buff
*skb
, struct sock
*sk
,
1599 struct tcp_sacktag_state
*state
,
1602 if (skb
&& after(TCP_SKB_CB(skb
)->seq
, skip_to_seq
))
1605 return tcp_sacktag_bsearch(sk
, state
, skip_to_seq
);
1608 static struct sk_buff
*tcp_maybe_skipping_dsack(struct sk_buff
*skb
,
1610 struct tcp_sack_block
*next_dup
,
1611 struct tcp_sacktag_state
*state
,
1617 if (before(next_dup
->start_seq
, skip_to_seq
)) {
1618 skb
= tcp_sacktag_skip(skb
, sk
, state
, next_dup
->start_seq
);
1619 skb
= tcp_sacktag_walk(skb
, sk
, NULL
, state
,
1620 next_dup
->start_seq
, next_dup
->end_seq
,
1627 static int tcp_sack_cache_ok(const struct tcp_sock
*tp
, const struct tcp_sack_block
*cache
)
1629 return cache
< tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1633 tcp_sacktag_write_queue(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1634 u32 prior_snd_una
, struct tcp_sacktag_state
*state
)
1636 struct tcp_sock
*tp
= tcp_sk(sk
);
1637 const unsigned char *ptr
= (skb_transport_header(ack_skb
) +
1638 TCP_SKB_CB(ack_skb
)->sacked
);
1639 struct tcp_sack_block_wire
*sp_wire
= (struct tcp_sack_block_wire
*)(ptr
+2);
1640 struct tcp_sack_block sp
[TCP_NUM_SACKS
];
1641 struct tcp_sack_block
*cache
;
1642 struct sk_buff
*skb
;
1643 int num_sacks
= min(TCP_NUM_SACKS
, (ptr
[1] - TCPOLEN_SACK_BASE
) >> 3);
1645 bool found_dup_sack
= false;
1647 int first_sack_index
;
1650 state
->reord
= tp
->snd_nxt
;
1652 if (!tp
->sacked_out
)
1653 tcp_highest_sack_reset(sk
);
1655 found_dup_sack
= tcp_check_dsack(sk
, ack_skb
, sp_wire
,
1656 num_sacks
, prior_snd_una
);
1657 if (found_dup_sack
) {
1658 state
->flag
|= FLAG_DSACKING_ACK
;
1659 tp
->delivered
++; /* A spurious retransmission is delivered */
1662 /* Eliminate too old ACKs, but take into
1663 * account more or less fresh ones, they can
1664 * contain valid SACK info.
1666 if (before(TCP_SKB_CB(ack_skb
)->ack_seq
, prior_snd_una
- tp
->max_window
))
1669 if (!tp
->packets_out
)
1673 first_sack_index
= 0;
1674 for (i
= 0; i
< num_sacks
; i
++) {
1675 bool dup_sack
= !i
&& found_dup_sack
;
1677 sp
[used_sacks
].start_seq
= get_unaligned_be32(&sp_wire
[i
].start_seq
);
1678 sp
[used_sacks
].end_seq
= get_unaligned_be32(&sp_wire
[i
].end_seq
);
1680 if (!tcp_is_sackblock_valid(tp
, dup_sack
,
1681 sp
[used_sacks
].start_seq
,
1682 sp
[used_sacks
].end_seq
)) {
1686 if (!tp
->undo_marker
)
1687 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDNOUNDO
;
1689 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDOLD
;
1691 /* Don't count olds caused by ACK reordering */
1692 if ((TCP_SKB_CB(ack_skb
)->ack_seq
!= tp
->snd_una
) &&
1693 !after(sp
[used_sacks
].end_seq
, tp
->snd_una
))
1695 mib_idx
= LINUX_MIB_TCPSACKDISCARD
;
1698 NET_INC_STATS(sock_net(sk
), mib_idx
);
1700 first_sack_index
= -1;
1704 /* Ignore very old stuff early */
1705 if (!after(sp
[used_sacks
].end_seq
, prior_snd_una
))
1711 /* order SACK blocks to allow in order walk of the retrans queue */
1712 for (i
= used_sacks
- 1; i
> 0; i
--) {
1713 for (j
= 0; j
< i
; j
++) {
1714 if (after(sp
[j
].start_seq
, sp
[j
+ 1].start_seq
)) {
1715 swap(sp
[j
], sp
[j
+ 1]);
1717 /* Track where the first SACK block goes to */
1718 if (j
== first_sack_index
)
1719 first_sack_index
= j
+ 1;
1724 state
->mss_now
= tcp_current_mss(sk
);
1728 if (!tp
->sacked_out
) {
1729 /* It's already past, so skip checking against it */
1730 cache
= tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1732 cache
= tp
->recv_sack_cache
;
1733 /* Skip empty blocks in at head of the cache */
1734 while (tcp_sack_cache_ok(tp
, cache
) && !cache
->start_seq
&&
1739 while (i
< used_sacks
) {
1740 u32 start_seq
= sp
[i
].start_seq
;
1741 u32 end_seq
= sp
[i
].end_seq
;
1742 bool dup_sack
= (found_dup_sack
&& (i
== first_sack_index
));
1743 struct tcp_sack_block
*next_dup
= NULL
;
1745 if (found_dup_sack
&& ((i
+ 1) == first_sack_index
))
1746 next_dup
= &sp
[i
+ 1];
1748 /* Skip too early cached blocks */
1749 while (tcp_sack_cache_ok(tp
, cache
) &&
1750 !before(start_seq
, cache
->end_seq
))
1753 /* Can skip some work by looking recv_sack_cache? */
1754 if (tcp_sack_cache_ok(tp
, cache
) && !dup_sack
&&
1755 after(end_seq
, cache
->start_seq
)) {
1758 if (before(start_seq
, cache
->start_seq
)) {
1759 skb
= tcp_sacktag_skip(skb
, sk
, state
,
1761 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
,
1768 /* Rest of the block already fully processed? */
1769 if (!after(end_seq
, cache
->end_seq
))
1772 skb
= tcp_maybe_skipping_dsack(skb
, sk
, next_dup
,
1776 /* ...tail remains todo... */
1777 if (tcp_highest_sack_seq(tp
) == cache
->end_seq
) {
1778 /* ...but better entrypoint exists! */
1779 skb
= tcp_highest_sack(sk
);
1786 skb
= tcp_sacktag_skip(skb
, sk
, state
, cache
->end_seq
);
1787 /* Check overlap against next cached too (past this one already) */
1792 if (!before(start_seq
, tcp_highest_sack_seq(tp
))) {
1793 skb
= tcp_highest_sack(sk
);
1797 skb
= tcp_sacktag_skip(skb
, sk
, state
, start_seq
);
1800 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
, state
,
1801 start_seq
, end_seq
, dup_sack
);
1807 /* Clear the head of the cache sack blocks so we can skip it next time */
1808 for (i
= 0; i
< ARRAY_SIZE(tp
->recv_sack_cache
) - used_sacks
; i
++) {
1809 tp
->recv_sack_cache
[i
].start_seq
= 0;
1810 tp
->recv_sack_cache
[i
].end_seq
= 0;
1812 for (j
= 0; j
< used_sacks
; j
++)
1813 tp
->recv_sack_cache
[i
++] = sp
[j
];
1815 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Loss
|| tp
->undo_marker
)
1816 tcp_check_sack_reordering(sk
, state
->reord
, 0);
1818 tcp_verify_left_out(tp
);
1821 #if FASTRETRANS_DEBUG > 0
1822 WARN_ON((int)tp
->sacked_out
< 0);
1823 WARN_ON((int)tp
->lost_out
< 0);
1824 WARN_ON((int)tp
->retrans_out
< 0);
1825 WARN_ON((int)tcp_packets_in_flight(tp
) < 0);
1830 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1831 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1833 static bool tcp_limit_reno_sacked(struct tcp_sock
*tp
)
1837 holes
= max(tp
->lost_out
, 1U);
1838 holes
= min(holes
, tp
->packets_out
);
1840 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1841 tp
->sacked_out
= tp
->packets_out
- holes
;
1847 /* If we receive more dupacks than we expected counting segments
1848 * in assumption of absent reordering, interpret this as reordering.
1849 * The only another reason could be bug in receiver TCP.
1851 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1853 struct tcp_sock
*tp
= tcp_sk(sk
);
1855 if (!tcp_limit_reno_sacked(tp
))
1858 tp
->reordering
= min_t(u32
, tp
->packets_out
+ addend
,
1859 sock_net(sk
)->ipv4
.sysctl_tcp_max_reordering
);
1861 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPRENOREORDER
);
1864 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1866 static void tcp_add_reno_sack(struct sock
*sk
)
1868 struct tcp_sock
*tp
= tcp_sk(sk
);
1869 u32 prior_sacked
= tp
->sacked_out
;
1872 tcp_check_reno_reordering(sk
, 0);
1873 if (tp
->sacked_out
> prior_sacked
)
1874 tp
->delivered
++; /* Some out-of-order packet is delivered */
1875 tcp_verify_left_out(tp
);
1878 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1880 static void tcp_remove_reno_sacks(struct sock
*sk
, int acked
)
1882 struct tcp_sock
*tp
= tcp_sk(sk
);
1885 /* One ACK acked hole. The rest eat duplicate ACKs. */
1886 tp
->delivered
+= max_t(int, acked
- tp
->sacked_out
, 1);
1887 if (acked
- 1 >= tp
->sacked_out
)
1890 tp
->sacked_out
-= acked
- 1;
1892 tcp_check_reno_reordering(sk
, acked
);
1893 tcp_verify_left_out(tp
);
1896 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1901 void tcp_clear_retrans(struct tcp_sock
*tp
)
1903 tp
->retrans_out
= 0;
1905 tp
->undo_marker
= 0;
1906 tp
->undo_retrans
= -1;
1910 static inline void tcp_init_undo(struct tcp_sock
*tp
)
1912 tp
->undo_marker
= tp
->snd_una
;
1913 /* Retransmission still in flight may cause DSACKs later. */
1914 tp
->undo_retrans
= tp
->retrans_out
? : -1;
1917 static bool tcp_is_rack(const struct sock
*sk
)
1919 return sock_net(sk
)->ipv4
.sysctl_tcp_recovery
& TCP_RACK_LOSS_DETECTION
;
1922 /* If we detect SACK reneging, forget all SACK information
1923 * and reset tags completely, otherwise preserve SACKs. If receiver
1924 * dropped its ofo queue, we will know this due to reneging detection.
1926 static void tcp_timeout_mark_lost(struct sock
*sk
)
1928 struct tcp_sock
*tp
= tcp_sk(sk
);
1929 struct sk_buff
*skb
, *head
;
1930 bool is_reneg
; /* is receiver reneging on SACKs? */
1932 head
= tcp_rtx_queue_head(sk
);
1933 is_reneg
= head
&& (TCP_SKB_CB(head
)->sacked
& TCPCB_SACKED_ACKED
);
1935 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPSACKRENEGING
);
1937 /* Mark SACK reneging until we recover from this loss event. */
1938 tp
->is_sack_reneg
= 1;
1939 } else if (tcp_is_reno(tp
)) {
1940 tcp_reset_reno_sack(tp
);
1944 skb_rbtree_walk_from(skb
) {
1946 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
1947 else if (tcp_is_rack(sk
) && skb
!= head
&&
1948 tcp_rack_skb_timeout(tp
, skb
, 0) > 0)
1949 continue; /* Don't mark recently sent ones lost yet */
1950 tcp_mark_skb_lost(sk
, skb
);
1952 tcp_verify_left_out(tp
);
1953 tcp_clear_all_retrans_hints(tp
);
1956 /* Enter Loss state. */
1957 void tcp_enter_loss(struct sock
*sk
)
1959 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1960 struct tcp_sock
*tp
= tcp_sk(sk
);
1961 struct net
*net
= sock_net(sk
);
1962 bool new_recovery
= icsk
->icsk_ca_state
< TCP_CA_Recovery
;
1964 tcp_timeout_mark_lost(sk
);
1966 /* Reduce ssthresh if it has not yet been made inside this window. */
1967 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
||
1968 !after(tp
->high_seq
, tp
->snd_una
) ||
1969 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
1970 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1971 tp
->prior_cwnd
= tp
->snd_cwnd
;
1972 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1973 tcp_ca_event(sk
, CA_EVENT_LOSS
);
1976 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + 1;
1977 tp
->snd_cwnd_cnt
= 0;
1978 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
1980 /* Timeout in disordered state after receiving substantial DUPACKs
1981 * suggests that the degree of reordering is over-estimated.
1983 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
&&
1984 tp
->sacked_out
>= net
->ipv4
.sysctl_tcp_reordering
)
1985 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1986 net
->ipv4
.sysctl_tcp_reordering
);
1987 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1988 tp
->high_seq
= tp
->snd_nxt
;
1989 tcp_ecn_queue_cwr(tp
);
1991 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1992 * loss recovery is underway except recurring timeout(s) on
1993 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1995 tp
->frto
= net
->ipv4
.sysctl_tcp_frto
&&
1996 (new_recovery
|| icsk
->icsk_retransmits
) &&
1997 !inet_csk(sk
)->icsk_mtup
.probe_size
;
2000 /* If ACK arrived pointing to a remembered SACK, it means that our
2001 * remembered SACKs do not reflect real state of receiver i.e.
2002 * receiver _host_ is heavily congested (or buggy).
2004 * To avoid big spurious retransmission bursts due to transient SACK
2005 * scoreboard oddities that look like reneging, we give the receiver a
2006 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2007 * restore sanity to the SACK scoreboard. If the apparent reneging
2008 * persists until this RTO then we'll clear the SACK scoreboard.
2010 static bool tcp_check_sack_reneging(struct sock
*sk
, int flag
)
2012 if (flag
& FLAG_SACK_RENEGING
) {
2013 struct tcp_sock
*tp
= tcp_sk(sk
);
2014 unsigned long delay
= max(usecs_to_jiffies(tp
->srtt_us
>> 4),
2015 msecs_to_jiffies(10));
2017 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2018 delay
, TCP_RTO_MAX
);
2024 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2025 * counter when SACK is enabled (without SACK, sacked_out is used for
2028 * With reordering, holes may still be in flight, so RFC3517 recovery
2029 * uses pure sacked_out (total number of SACKed segments) even though
2030 * it violates the RFC that uses duplicate ACKs, often these are equal
2031 * but when e.g. out-of-window ACKs or packet duplication occurs,
2032 * they differ. Since neither occurs due to loss, TCP should really
2035 static inline int tcp_dupack_heuristics(const struct tcp_sock
*tp
)
2037 return tp
->sacked_out
+ 1;
2040 /* Linux NewReno/SACK/ECN state machine.
2041 * --------------------------------------
2043 * "Open" Normal state, no dubious events, fast path.
2044 * "Disorder" In all the respects it is "Open",
2045 * but requires a bit more attention. It is entered when
2046 * we see some SACKs or dupacks. It is split of "Open"
2047 * mainly to move some processing from fast path to slow one.
2048 * "CWR" CWND was reduced due to some Congestion Notification event.
2049 * It can be ECN, ICMP source quench, local device congestion.
2050 * "Recovery" CWND was reduced, we are fast-retransmitting.
2051 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2053 * tcp_fastretrans_alert() is entered:
2054 * - each incoming ACK, if state is not "Open"
2055 * - when arrived ACK is unusual, namely:
2060 * Counting packets in flight is pretty simple.
2062 * in_flight = packets_out - left_out + retrans_out
2064 * packets_out is SND.NXT-SND.UNA counted in packets.
2066 * retrans_out is number of retransmitted segments.
2068 * left_out is number of segments left network, but not ACKed yet.
2070 * left_out = sacked_out + lost_out
2072 * sacked_out: Packets, which arrived to receiver out of order
2073 * and hence not ACKed. With SACKs this number is simply
2074 * amount of SACKed data. Even without SACKs
2075 * it is easy to give pretty reliable estimate of this number,
2076 * counting duplicate ACKs.
2078 * lost_out: Packets lost by network. TCP has no explicit
2079 * "loss notification" feedback from network (for now).
2080 * It means that this number can be only _guessed_.
2081 * Actually, it is the heuristics to predict lossage that
2082 * distinguishes different algorithms.
2084 * F.e. after RTO, when all the queue is considered as lost,
2085 * lost_out = packets_out and in_flight = retrans_out.
2087 * Essentially, we have now a few algorithms detecting
2090 * If the receiver supports SACK:
2092 * RFC6675/3517: It is the conventional algorithm. A packet is
2093 * considered lost if the number of higher sequence packets
2094 * SACKed is greater than or equal the DUPACK thoreshold
2095 * (reordering). This is implemented in tcp_mark_head_lost and
2096 * tcp_update_scoreboard.
2098 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm
2099 * (2017-) that checks timing instead of counting DUPACKs.
2100 * Essentially a packet is considered lost if it's not S/ACKed
2101 * after RTT + reordering_window, where both metrics are
2102 * dynamically measured and adjusted. This is implemented in
2103 * tcp_rack_mark_lost.
2105 * If the receiver does not support SACK:
2107 * NewReno (RFC6582): in Recovery we assume that one segment
2108 * is lost (classic Reno). While we are in Recovery and
2109 * a partial ACK arrives, we assume that one more packet
2110 * is lost (NewReno). This heuristics are the same in NewReno
2113 * Really tricky (and requiring careful tuning) part of algorithm
2114 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2115 * The first determines the moment _when_ we should reduce CWND and,
2116 * hence, slow down forward transmission. In fact, it determines the moment
2117 * when we decide that hole is caused by loss, rather than by a reorder.
2119 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2120 * holes, caused by lost packets.
2122 * And the most logically complicated part of algorithm is undo
2123 * heuristics. We detect false retransmits due to both too early
2124 * fast retransmit (reordering) and underestimated RTO, analyzing
2125 * timestamps and D-SACKs. When we detect that some segments were
2126 * retransmitted by mistake and CWND reduction was wrong, we undo
2127 * window reduction and abort recovery phase. This logic is hidden
2128 * inside several functions named tcp_try_undo_<something>.
2131 /* This function decides, when we should leave Disordered state
2132 * and enter Recovery phase, reducing congestion window.
2134 * Main question: may we further continue forward transmission
2135 * with the same cwnd?
2137 static bool tcp_time_to_recover(struct sock
*sk
, int flag
)
2139 struct tcp_sock
*tp
= tcp_sk(sk
);
2141 /* Trick#1: The loss is proven. */
2145 /* Not-A-Trick#2 : Classic rule... */
2146 if (!tcp_is_rack(sk
) && tcp_dupack_heuristics(tp
) > tp
->reordering
)
2152 /* Detect loss in event "A" above by marking head of queue up as lost.
2153 * For non-SACK(Reno) senders, the first "packets" number of segments
2154 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2155 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2156 * the maximum SACKed segments to pass before reaching this limit.
2158 static void tcp_mark_head_lost(struct sock
*sk
, int packets
, int mark_head
)
2160 struct tcp_sock
*tp
= tcp_sk(sk
);
2161 struct sk_buff
*skb
;
2162 int cnt
, oldcnt
, lost
;
2164 /* Use SACK to deduce losses of new sequences sent during recovery */
2165 const u32 loss_high
= tcp_is_sack(tp
) ? tp
->snd_nxt
: tp
->high_seq
;
2167 WARN_ON(packets
> tp
->packets_out
);
2168 skb
= tp
->lost_skb_hint
;
2170 /* Head already handled? */
2171 if (mark_head
&& after(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
))
2173 cnt
= tp
->lost_cnt_hint
;
2175 skb
= tcp_rtx_queue_head(sk
);
2179 skb_rbtree_walk_from(skb
) {
2180 /* TODO: do this better */
2181 /* this is not the most efficient way to do this... */
2182 tp
->lost_skb_hint
= skb
;
2183 tp
->lost_cnt_hint
= cnt
;
2185 if (after(TCP_SKB_CB(skb
)->end_seq
, loss_high
))
2189 if (tcp_is_reno(tp
) ||
2190 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2191 cnt
+= tcp_skb_pcount(skb
);
2193 if (cnt
> packets
) {
2194 if (tcp_is_sack(tp
) ||
2195 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
) ||
2196 (oldcnt
>= packets
))
2199 mss
= tcp_skb_mss(skb
);
2200 /* If needed, chop off the prefix to mark as lost. */
2201 lost
= (packets
- oldcnt
) * mss
;
2202 if (lost
< skb
->len
&&
2203 tcp_fragment(sk
, TCP_FRAG_IN_RTX_QUEUE
, skb
,
2204 lost
, mss
, GFP_ATOMIC
) < 0)
2209 tcp_skb_mark_lost(tp
, skb
);
2214 tcp_verify_left_out(tp
);
2217 /* Account newly detected lost packet(s) */
2219 static void tcp_update_scoreboard(struct sock
*sk
, int fast_rexmit
)
2221 struct tcp_sock
*tp
= tcp_sk(sk
);
2223 if (tcp_is_sack(tp
)) {
2224 int sacked_upto
= tp
->sacked_out
- tp
->reordering
;
2225 if (sacked_upto
>= 0)
2226 tcp_mark_head_lost(sk
, sacked_upto
, 0);
2227 else if (fast_rexmit
)
2228 tcp_mark_head_lost(sk
, 1, 1);
2232 static bool tcp_tsopt_ecr_before(const struct tcp_sock
*tp
, u32 when
)
2234 return tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2235 before(tp
->rx_opt
.rcv_tsecr
, when
);
2238 /* skb is spurious retransmitted if the returned timestamp echo
2239 * reply is prior to the skb transmission time
2241 static bool tcp_skb_spurious_retrans(const struct tcp_sock
*tp
,
2242 const struct sk_buff
*skb
)
2244 return (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
) &&
2245 tcp_tsopt_ecr_before(tp
, tcp_skb_timestamp(skb
));
2248 /* Nothing was retransmitted or returned timestamp is less
2249 * than timestamp of the first retransmission.
2251 static inline bool tcp_packet_delayed(const struct tcp_sock
*tp
)
2253 return !tp
->retrans_stamp
||
2254 tcp_tsopt_ecr_before(tp
, tp
->retrans_stamp
);
2257 /* Undo procedures. */
2259 /* We can clear retrans_stamp when there are no retransmissions in the
2260 * window. It would seem that it is trivially available for us in
2261 * tp->retrans_out, however, that kind of assumptions doesn't consider
2262 * what will happen if errors occur when sending retransmission for the
2263 * second time. ...It could the that such segment has only
2264 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2265 * the head skb is enough except for some reneging corner cases that
2266 * are not worth the effort.
2268 * Main reason for all this complexity is the fact that connection dying
2269 * time now depends on the validity of the retrans_stamp, in particular,
2270 * that successive retransmissions of a segment must not advance
2271 * retrans_stamp under any conditions.
2273 static bool tcp_any_retrans_done(const struct sock
*sk
)
2275 const struct tcp_sock
*tp
= tcp_sk(sk
);
2276 struct sk_buff
*skb
;
2278 if (tp
->retrans_out
)
2281 skb
= tcp_rtx_queue_head(sk
);
2282 if (unlikely(skb
&& TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
))
2288 static void DBGUNDO(struct sock
*sk
, const char *msg
)
2290 #if FASTRETRANS_DEBUG > 1
2291 struct tcp_sock
*tp
= tcp_sk(sk
);
2292 struct inet_sock
*inet
= inet_sk(sk
);
2294 if (sk
->sk_family
== AF_INET
) {
2295 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2297 &inet
->inet_daddr
, ntohs(inet
->inet_dport
),
2298 tp
->snd_cwnd
, tcp_left_out(tp
),
2299 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2302 #if IS_ENABLED(CONFIG_IPV6)
2303 else if (sk
->sk_family
== AF_INET6
) {
2304 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2306 &sk
->sk_v6_daddr
, ntohs(inet
->inet_dport
),
2307 tp
->snd_cwnd
, tcp_left_out(tp
),
2308 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2315 static void tcp_undo_cwnd_reduction(struct sock
*sk
, bool unmark_loss
)
2317 struct tcp_sock
*tp
= tcp_sk(sk
);
2320 struct sk_buff
*skb
;
2322 skb_rbtree_walk(skb
, &sk
->tcp_rtx_queue
) {
2323 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2326 tcp_clear_all_retrans_hints(tp
);
2329 if (tp
->prior_ssthresh
) {
2330 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2332 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
2334 if (tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
2335 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
2336 tcp_ecn_withdraw_cwr(tp
);
2339 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
2340 tp
->undo_marker
= 0;
2341 tp
->rack
.advanced
= 1; /* Force RACK to re-exam losses */
2344 static inline bool tcp_may_undo(const struct tcp_sock
*tp
)
2346 return tp
->undo_marker
&& (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
2349 /* People celebrate: "We love our President!" */
2350 static bool tcp_try_undo_recovery(struct sock
*sk
)
2352 struct tcp_sock
*tp
= tcp_sk(sk
);
2354 if (tcp_may_undo(tp
)) {
2357 /* Happy end! We did not retransmit anything
2358 * or our original transmission succeeded.
2360 DBGUNDO(sk
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
2361 tcp_undo_cwnd_reduction(sk
, false);
2362 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
2363 mib_idx
= LINUX_MIB_TCPLOSSUNDO
;
2365 mib_idx
= LINUX_MIB_TCPFULLUNDO
;
2367 NET_INC_STATS(sock_net(sk
), mib_idx
);
2368 } else if (tp
->rack
.reo_wnd_persist
) {
2369 tp
->rack
.reo_wnd_persist
--;
2371 if (tp
->snd_una
== tp
->high_seq
&& tcp_is_reno(tp
)) {
2372 /* Hold old state until something *above* high_seq
2373 * is ACKed. For Reno it is MUST to prevent false
2374 * fast retransmits (RFC2582). SACK TCP is safe. */
2375 if (!tcp_any_retrans_done(sk
))
2376 tp
->retrans_stamp
= 0;
2379 tcp_set_ca_state(sk
, TCP_CA_Open
);
2380 tp
->is_sack_reneg
= 0;
2384 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2385 static bool tcp_try_undo_dsack(struct sock
*sk
)
2387 struct tcp_sock
*tp
= tcp_sk(sk
);
2389 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
2390 tp
->rack
.reo_wnd_persist
= min(TCP_RACK_RECOVERY_THRESH
,
2391 tp
->rack
.reo_wnd_persist
+ 1);
2392 DBGUNDO(sk
, "D-SACK");
2393 tcp_undo_cwnd_reduction(sk
, false);
2394 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPDSACKUNDO
);
2400 /* Undo during loss recovery after partial ACK or using F-RTO. */
2401 static bool tcp_try_undo_loss(struct sock
*sk
, bool frto_undo
)
2403 struct tcp_sock
*tp
= tcp_sk(sk
);
2405 if (frto_undo
|| tcp_may_undo(tp
)) {
2406 tcp_undo_cwnd_reduction(sk
, true);
2408 DBGUNDO(sk
, "partial loss");
2409 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPLOSSUNDO
);
2411 NET_INC_STATS(sock_net(sk
),
2412 LINUX_MIB_TCPSPURIOUSRTOS
);
2413 inet_csk(sk
)->icsk_retransmits
= 0;
2414 if (frto_undo
|| tcp_is_sack(tp
)) {
2415 tcp_set_ca_state(sk
, TCP_CA_Open
);
2416 tp
->is_sack_reneg
= 0;
2423 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2424 * It computes the number of packets to send (sndcnt) based on packets newly
2426 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2427 * cwnd reductions across a full RTT.
2428 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2429 * But when the retransmits are acked without further losses, PRR
2430 * slow starts cwnd up to ssthresh to speed up the recovery.
2432 static void tcp_init_cwnd_reduction(struct sock
*sk
)
2434 struct tcp_sock
*tp
= tcp_sk(sk
);
2436 tp
->high_seq
= tp
->snd_nxt
;
2437 tp
->tlp_high_seq
= 0;
2438 tp
->snd_cwnd_cnt
= 0;
2439 tp
->prior_cwnd
= tp
->snd_cwnd
;
2440 tp
->prr_delivered
= 0;
2442 tp
->snd_ssthresh
= inet_csk(sk
)->icsk_ca_ops
->ssthresh(sk
);
2443 tcp_ecn_queue_cwr(tp
);
2446 void tcp_cwnd_reduction(struct sock
*sk
, int newly_acked_sacked
, int flag
)
2448 struct tcp_sock
*tp
= tcp_sk(sk
);
2450 int delta
= tp
->snd_ssthresh
- tcp_packets_in_flight(tp
);
2452 if (newly_acked_sacked
<= 0 || WARN_ON_ONCE(!tp
->prior_cwnd
))
2455 tp
->prr_delivered
+= newly_acked_sacked
;
2457 u64 dividend
= (u64
)tp
->snd_ssthresh
* tp
->prr_delivered
+
2459 sndcnt
= div_u64(dividend
, tp
->prior_cwnd
) - tp
->prr_out
;
2460 } else if ((flag
& FLAG_RETRANS_DATA_ACKED
) &&
2461 !(flag
& FLAG_LOST_RETRANS
)) {
2462 sndcnt
= min_t(int, delta
,
2463 max_t(int, tp
->prr_delivered
- tp
->prr_out
,
2464 newly_acked_sacked
) + 1);
2466 sndcnt
= min(delta
, newly_acked_sacked
);
2468 /* Force a fast retransmit upon entering fast recovery */
2469 sndcnt
= max(sndcnt
, (tp
->prr_out
? 0 : 1));
2470 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + sndcnt
;
2473 static inline void tcp_end_cwnd_reduction(struct sock
*sk
)
2475 struct tcp_sock
*tp
= tcp_sk(sk
);
2477 if (inet_csk(sk
)->icsk_ca_ops
->cong_control
)
2480 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2481 if (tp
->snd_ssthresh
< TCP_INFINITE_SSTHRESH
&&
2482 (inet_csk(sk
)->icsk_ca_state
== TCP_CA_CWR
|| tp
->undo_marker
)) {
2483 tp
->snd_cwnd
= tp
->snd_ssthresh
;
2484 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
2486 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
2489 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2490 void tcp_enter_cwr(struct sock
*sk
)
2492 struct tcp_sock
*tp
= tcp_sk(sk
);
2494 tp
->prior_ssthresh
= 0;
2495 if (inet_csk(sk
)->icsk_ca_state
< TCP_CA_CWR
) {
2496 tp
->undo_marker
= 0;
2497 tcp_init_cwnd_reduction(sk
);
2498 tcp_set_ca_state(sk
, TCP_CA_CWR
);
2501 EXPORT_SYMBOL(tcp_enter_cwr
);
2503 static void tcp_try_keep_open(struct sock
*sk
)
2505 struct tcp_sock
*tp
= tcp_sk(sk
);
2506 int state
= TCP_CA_Open
;
2508 if (tcp_left_out(tp
) || tcp_any_retrans_done(sk
))
2509 state
= TCP_CA_Disorder
;
2511 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
2512 tcp_set_ca_state(sk
, state
);
2513 tp
->high_seq
= tp
->snd_nxt
;
2517 static void tcp_try_to_open(struct sock
*sk
, int flag
)
2519 struct tcp_sock
*tp
= tcp_sk(sk
);
2521 tcp_verify_left_out(tp
);
2523 if (!tcp_any_retrans_done(sk
))
2524 tp
->retrans_stamp
= 0;
2526 if (flag
& FLAG_ECE
)
2529 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
2530 tcp_try_keep_open(sk
);
2534 static void tcp_mtup_probe_failed(struct sock
*sk
)
2536 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2538 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
2539 icsk
->icsk_mtup
.probe_size
= 0;
2540 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPMTUPFAIL
);
2543 static void tcp_mtup_probe_success(struct sock
*sk
)
2545 struct tcp_sock
*tp
= tcp_sk(sk
);
2546 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2548 /* FIXME: breaks with very large cwnd */
2549 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2550 tp
->snd_cwnd
= tp
->snd_cwnd
*
2551 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
2552 icsk
->icsk_mtup
.probe_size
;
2553 tp
->snd_cwnd_cnt
= 0;
2554 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
2555 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2557 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
2558 icsk
->icsk_mtup
.probe_size
= 0;
2559 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
2560 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPMTUPSUCCESS
);
2563 /* Do a simple retransmit without using the backoff mechanisms in
2564 * tcp_timer. This is used for path mtu discovery.
2565 * The socket is already locked here.
2567 void tcp_simple_retransmit(struct sock
*sk
)
2569 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2570 struct tcp_sock
*tp
= tcp_sk(sk
);
2571 struct sk_buff
*skb
;
2572 unsigned int mss
= tcp_current_mss(sk
);
2574 skb_rbtree_walk(skb
, &sk
->tcp_rtx_queue
) {
2575 if (tcp_skb_seglen(skb
) > mss
&&
2576 !(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
2577 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2578 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
2579 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2581 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
2585 tcp_clear_retrans_hints_partial(tp
);
2590 if (tcp_is_reno(tp
))
2591 tcp_limit_reno_sacked(tp
);
2593 tcp_verify_left_out(tp
);
2595 /* Don't muck with the congestion window here.
2596 * Reason is that we do not increase amount of _data_
2597 * in network, but units changed and effective
2598 * cwnd/ssthresh really reduced now.
2600 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
2601 tp
->high_seq
= tp
->snd_nxt
;
2602 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2603 tp
->prior_ssthresh
= 0;
2604 tp
->undo_marker
= 0;
2605 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2607 tcp_xmit_retransmit_queue(sk
);
2609 EXPORT_SYMBOL(tcp_simple_retransmit
);
2611 void tcp_enter_recovery(struct sock
*sk
, bool ece_ack
)
2613 struct tcp_sock
*tp
= tcp_sk(sk
);
2616 if (tcp_is_reno(tp
))
2617 mib_idx
= LINUX_MIB_TCPRENORECOVERY
;
2619 mib_idx
= LINUX_MIB_TCPSACKRECOVERY
;
2621 NET_INC_STATS(sock_net(sk
), mib_idx
);
2623 tp
->prior_ssthresh
= 0;
2626 if (!tcp_in_cwnd_reduction(sk
)) {
2628 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2629 tcp_init_cwnd_reduction(sk
);
2631 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
2634 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2635 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2637 static void tcp_process_loss(struct sock
*sk
, int flag
, bool is_dupack
,
2640 struct tcp_sock
*tp
= tcp_sk(sk
);
2641 bool recovered
= !before(tp
->snd_una
, tp
->high_seq
);
2643 if ((flag
& FLAG_SND_UNA_ADVANCED
) &&
2644 tcp_try_undo_loss(sk
, false))
2647 if (tp
->frto
) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2648 /* Step 3.b. A timeout is spurious if not all data are
2649 * lost, i.e., never-retransmitted data are (s)acked.
2651 if ((flag
& FLAG_ORIG_SACK_ACKED
) &&
2652 tcp_try_undo_loss(sk
, true))
2655 if (after(tp
->snd_nxt
, tp
->high_seq
)) {
2656 if (flag
& FLAG_DATA_SACKED
|| is_dupack
)
2657 tp
->frto
= 0; /* Step 3.a. loss was real */
2658 } else if (flag
& FLAG_SND_UNA_ADVANCED
&& !recovered
) {
2659 tp
->high_seq
= tp
->snd_nxt
;
2660 /* Step 2.b. Try send new data (but deferred until cwnd
2661 * is updated in tcp_ack()). Otherwise fall back to
2662 * the conventional recovery.
2664 if (!tcp_write_queue_empty(sk
) &&
2665 after(tcp_wnd_end(tp
), tp
->snd_nxt
)) {
2666 *rexmit
= REXMIT_NEW
;
2674 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2675 tcp_try_undo_recovery(sk
);
2678 if (tcp_is_reno(tp
)) {
2679 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2680 * delivered. Lower inflight to clock out (re)tranmissions.
2682 if (after(tp
->snd_nxt
, tp
->high_seq
) && is_dupack
)
2683 tcp_add_reno_sack(sk
);
2684 else if (flag
& FLAG_SND_UNA_ADVANCED
)
2685 tcp_reset_reno_sack(tp
);
2687 *rexmit
= REXMIT_LOST
;
2690 /* Undo during fast recovery after partial ACK. */
2691 static bool tcp_try_undo_partial(struct sock
*sk
, u32 prior_snd_una
)
2693 struct tcp_sock
*tp
= tcp_sk(sk
);
2695 if (tp
->undo_marker
&& tcp_packet_delayed(tp
)) {
2696 /* Plain luck! Hole if filled with delayed
2697 * packet, rather than with a retransmit. Check reordering.
2699 tcp_check_sack_reordering(sk
, prior_snd_una
, 1);
2701 /* We are getting evidence that the reordering degree is higher
2702 * than we realized. If there are no retransmits out then we
2703 * can undo. Otherwise we clock out new packets but do not
2704 * mark more packets lost or retransmit more.
2706 if (tp
->retrans_out
)
2709 if (!tcp_any_retrans_done(sk
))
2710 tp
->retrans_stamp
= 0;
2712 DBGUNDO(sk
, "partial recovery");
2713 tcp_undo_cwnd_reduction(sk
, true);
2714 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPPARTIALUNDO
);
2715 tcp_try_keep_open(sk
);
2721 static void tcp_identify_packet_loss(struct sock
*sk
, int *ack_flag
)
2723 struct tcp_sock
*tp
= tcp_sk(sk
);
2725 if (tcp_rtx_queue_empty(sk
))
2728 if (unlikely(tcp_is_reno(tp
))) {
2729 tcp_newreno_mark_lost(sk
, *ack_flag
& FLAG_SND_UNA_ADVANCED
);
2730 } else if (tcp_is_rack(sk
)) {
2731 u32 prior_retrans
= tp
->retrans_out
;
2733 tcp_rack_mark_lost(sk
);
2734 if (prior_retrans
> tp
->retrans_out
)
2735 *ack_flag
|= FLAG_LOST_RETRANS
;
2739 static bool tcp_force_fast_retransmit(struct sock
*sk
)
2741 struct tcp_sock
*tp
= tcp_sk(sk
);
2743 return after(tcp_highest_sack_seq(tp
),
2744 tp
->snd_una
+ tp
->reordering
* tp
->mss_cache
);
2747 /* Process an event, which can update packets-in-flight not trivially.
2748 * Main goal of this function is to calculate new estimate for left_out,
2749 * taking into account both packets sitting in receiver's buffer and
2750 * packets lost by network.
2752 * Besides that it updates the congestion state when packet loss or ECN
2753 * is detected. But it does not reduce the cwnd, it is done by the
2754 * congestion control later.
2756 * It does _not_ decide what to send, it is made in function
2757 * tcp_xmit_retransmit_queue().
2759 static void tcp_fastretrans_alert(struct sock
*sk
, const u32 prior_snd_una
,
2760 bool is_dupack
, int *ack_flag
, int *rexmit
)
2762 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2763 struct tcp_sock
*tp
= tcp_sk(sk
);
2764 int fast_rexmit
= 0, flag
= *ack_flag
;
2765 bool do_lost
= is_dupack
|| ((flag
& FLAG_DATA_SACKED
) &&
2766 tcp_force_fast_retransmit(sk
));
2768 if (!tp
->packets_out
&& tp
->sacked_out
)
2771 /* Now state machine starts.
2772 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2773 if (flag
& FLAG_ECE
)
2774 tp
->prior_ssthresh
= 0;
2776 /* B. In all the states check for reneging SACKs. */
2777 if (tcp_check_sack_reneging(sk
, flag
))
2780 /* C. Check consistency of the current state. */
2781 tcp_verify_left_out(tp
);
2783 /* D. Check state exit conditions. State can be terminated
2784 * when high_seq is ACKed. */
2785 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
2786 WARN_ON(tp
->retrans_out
!= 0);
2787 tp
->retrans_stamp
= 0;
2788 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
2789 switch (icsk
->icsk_ca_state
) {
2791 /* CWR is to be held something *above* high_seq
2792 * is ACKed for CWR bit to reach receiver. */
2793 if (tp
->snd_una
!= tp
->high_seq
) {
2794 tcp_end_cwnd_reduction(sk
);
2795 tcp_set_ca_state(sk
, TCP_CA_Open
);
2799 case TCP_CA_Recovery
:
2800 if (tcp_is_reno(tp
))
2801 tcp_reset_reno_sack(tp
);
2802 if (tcp_try_undo_recovery(sk
))
2804 tcp_end_cwnd_reduction(sk
);
2809 /* E. Process state. */
2810 switch (icsk
->icsk_ca_state
) {
2811 case TCP_CA_Recovery
:
2812 if (!(flag
& FLAG_SND_UNA_ADVANCED
)) {
2813 if (tcp_is_reno(tp
) && is_dupack
)
2814 tcp_add_reno_sack(sk
);
2816 if (tcp_try_undo_partial(sk
, prior_snd_una
))
2818 /* Partial ACK arrived. Force fast retransmit. */
2819 do_lost
= tcp_is_reno(tp
) ||
2820 tcp_force_fast_retransmit(sk
);
2822 if (tcp_try_undo_dsack(sk
)) {
2823 tcp_try_keep_open(sk
);
2826 tcp_identify_packet_loss(sk
, ack_flag
);
2829 tcp_process_loss(sk
, flag
, is_dupack
, rexmit
);
2830 tcp_identify_packet_loss(sk
, ack_flag
);
2831 if (!(icsk
->icsk_ca_state
== TCP_CA_Open
||
2832 (*ack_flag
& FLAG_LOST_RETRANS
)))
2834 /* Change state if cwnd is undone or retransmits are lost */
2837 if (tcp_is_reno(tp
)) {
2838 if (flag
& FLAG_SND_UNA_ADVANCED
)
2839 tcp_reset_reno_sack(tp
);
2841 tcp_add_reno_sack(sk
);
2844 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
)
2845 tcp_try_undo_dsack(sk
);
2847 tcp_identify_packet_loss(sk
, ack_flag
);
2848 if (!tcp_time_to_recover(sk
, flag
)) {
2849 tcp_try_to_open(sk
, flag
);
2853 /* MTU probe failure: don't reduce cwnd */
2854 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
2855 icsk
->icsk_mtup
.probe_size
&&
2856 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
2857 tcp_mtup_probe_failed(sk
);
2858 /* Restores the reduction we did in tcp_mtup_probe() */
2860 tcp_simple_retransmit(sk
);
2864 /* Otherwise enter Recovery state */
2865 tcp_enter_recovery(sk
, (flag
& FLAG_ECE
));
2869 if (!tcp_is_rack(sk
) && do_lost
)
2870 tcp_update_scoreboard(sk
, fast_rexmit
);
2871 *rexmit
= REXMIT_LOST
;
2874 static void tcp_update_rtt_min(struct sock
*sk
, u32 rtt_us
, const int flag
)
2876 u32 wlen
= sock_net(sk
)->ipv4
.sysctl_tcp_min_rtt_wlen
* HZ
;
2877 struct tcp_sock
*tp
= tcp_sk(sk
);
2879 if ((flag
& FLAG_ACK_MAYBE_DELAYED
) && rtt_us
> tcp_min_rtt(tp
)) {
2880 /* If the remote keeps returning delayed ACKs, eventually
2881 * the min filter would pick it up and overestimate the
2882 * prop. delay when it expires. Skip suspected delayed ACKs.
2886 minmax_running_min(&tp
->rtt_min
, wlen
, tcp_jiffies32
,
2887 rtt_us
? : jiffies_to_usecs(1));
2890 static bool tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
2891 long seq_rtt_us
, long sack_rtt_us
,
2892 long ca_rtt_us
, struct rate_sample
*rs
)
2894 const struct tcp_sock
*tp
= tcp_sk(sk
);
2896 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2897 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2898 * Karn's algorithm forbids taking RTT if some retransmitted data
2899 * is acked (RFC6298).
2902 seq_rtt_us
= sack_rtt_us
;
2904 /* RTTM Rule: A TSecr value received in a segment is used to
2905 * update the averaged RTT measurement only if the segment
2906 * acknowledges some new data, i.e., only if it advances the
2907 * left edge of the send window.
2908 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2910 if (seq_rtt_us
< 0 && tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2911 flag
& FLAG_ACKED
) {
2912 u32 delta
= tcp_time_stamp(tp
) - tp
->rx_opt
.rcv_tsecr
;
2913 u32 delta_us
= delta
* (USEC_PER_SEC
/ TCP_TS_HZ
);
2915 seq_rtt_us
= ca_rtt_us
= delta_us
;
2917 rs
->rtt_us
= ca_rtt_us
; /* RTT of last (S)ACKed packet (or -1) */
2921 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
2922 * always taken together with ACK, SACK, or TS-opts. Any negative
2923 * values will be skipped with the seq_rtt_us < 0 check above.
2925 tcp_update_rtt_min(sk
, ca_rtt_us
, flag
);
2926 tcp_rtt_estimator(sk
, seq_rtt_us
);
2929 /* RFC6298: only reset backoff on valid RTT measurement. */
2930 inet_csk(sk
)->icsk_backoff
= 0;
2934 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2935 void tcp_synack_rtt_meas(struct sock
*sk
, struct request_sock
*req
)
2937 struct rate_sample rs
;
2940 if (req
&& !req
->num_retrans
&& tcp_rsk(req
)->snt_synack
)
2941 rtt_us
= tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req
)->snt_synack
);
2943 tcp_ack_update_rtt(sk
, FLAG_SYN_ACKED
, rtt_us
, -1L, rtt_us
, &rs
);
2947 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
, u32 acked
)
2949 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2951 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, acked
);
2952 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_jiffies32
;
2955 /* Restart timer after forward progress on connection.
2956 * RFC2988 recommends to restart timer to now+rto.
2958 void tcp_rearm_rto(struct sock
*sk
)
2960 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2961 struct tcp_sock
*tp
= tcp_sk(sk
);
2963 /* If the retrans timer is currently being used by Fast Open
2964 * for SYN-ACK retrans purpose, stay put.
2966 if (tp
->fastopen_rsk
)
2969 if (!tp
->packets_out
) {
2970 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
2972 u32 rto
= inet_csk(sk
)->icsk_rto
;
2973 /* Offset the time elapsed after installing regular RTO */
2974 if (icsk
->icsk_pending
== ICSK_TIME_REO_TIMEOUT
||
2975 icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
) {
2976 s64 delta_us
= tcp_rto_delta_us(sk
);
2977 /* delta_us may not be positive if the socket is locked
2978 * when the retrans timer fires and is rescheduled.
2980 rto
= usecs_to_jiffies(max_t(int, delta_us
, 1));
2982 tcp_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
, rto
,
2983 TCP_RTO_MAX
, tcp_rtx_queue_head(sk
));
2987 /* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */
2988 static void tcp_set_xmit_timer(struct sock
*sk
)
2990 if (!tcp_schedule_loss_probe(sk
, true))
2994 /* If we get here, the whole TSO packet has not been acked. */
2995 static u32
tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
)
2997 struct tcp_sock
*tp
= tcp_sk(sk
);
3000 BUG_ON(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
));
3002 packets_acked
= tcp_skb_pcount(skb
);
3003 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
3005 packets_acked
-= tcp_skb_pcount(skb
);
3007 if (packets_acked
) {
3008 BUG_ON(tcp_skb_pcount(skb
) == 0);
3009 BUG_ON(!before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
));
3012 return packets_acked
;
3015 static void tcp_ack_tstamp(struct sock
*sk
, struct sk_buff
*skb
,
3018 const struct skb_shared_info
*shinfo
;
3020 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3021 if (likely(!TCP_SKB_CB(skb
)->txstamp_ack
))
3024 shinfo
= skb_shinfo(skb
);
3025 if (!before(shinfo
->tskey
, prior_snd_una
) &&
3026 before(shinfo
->tskey
, tcp_sk(sk
)->snd_una
)) {
3027 tcp_skb_tsorted_save(skb
) {
3028 __skb_tstamp_tx(skb
, NULL
, sk
, SCM_TSTAMP_ACK
);
3029 } tcp_skb_tsorted_restore(skb
);
3033 /* Remove acknowledged frames from the retransmission queue. If our packet
3034 * is before the ack sequence we can discard it as it's confirmed to have
3035 * arrived at the other end.
3037 static int tcp_clean_rtx_queue(struct sock
*sk
, u32 prior_fack
,
3039 struct tcp_sacktag_state
*sack
)
3041 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3042 u64 first_ackt
, last_ackt
;
3043 struct tcp_sock
*tp
= tcp_sk(sk
);
3044 u32 prior_sacked
= tp
->sacked_out
;
3045 u32 reord
= tp
->snd_nxt
; /* lowest acked un-retx un-sacked seq */
3046 struct sk_buff
*skb
, *next
;
3047 bool fully_acked
= true;
3048 long sack_rtt_us
= -1L;
3049 long seq_rtt_us
= -1L;
3050 long ca_rtt_us
= -1L;
3052 u32 last_in_flight
= 0;
3058 for (skb
= skb_rb_first(&sk
->tcp_rtx_queue
); skb
; skb
= next
) {
3059 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
3060 const u32 start_seq
= scb
->seq
;
3061 u8 sacked
= scb
->sacked
;
3064 tcp_ack_tstamp(sk
, skb
, prior_snd_una
);
3066 /* Determine how many packets and what bytes were acked, tso and else */
3067 if (after(scb
->end_seq
, tp
->snd_una
)) {
3068 if (tcp_skb_pcount(skb
) == 1 ||
3069 !after(tp
->snd_una
, scb
->seq
))
3072 acked_pcount
= tcp_tso_acked(sk
, skb
);
3075 fully_acked
= false;
3077 acked_pcount
= tcp_skb_pcount(skb
);
3080 if (unlikely(sacked
& TCPCB_RETRANS
)) {
3081 if (sacked
& TCPCB_SACKED_RETRANS
)
3082 tp
->retrans_out
-= acked_pcount
;
3083 flag
|= FLAG_RETRANS_DATA_ACKED
;
3084 } else if (!(sacked
& TCPCB_SACKED_ACKED
)) {
3085 last_ackt
= tcp_skb_timestamp_us(skb
);
3086 WARN_ON_ONCE(last_ackt
== 0);
3088 first_ackt
= last_ackt
;
3090 last_in_flight
= TCP_SKB_CB(skb
)->tx
.in_flight
;
3091 if (before(start_seq
, reord
))
3093 if (!after(scb
->end_seq
, tp
->high_seq
))
3094 flag
|= FLAG_ORIG_SACK_ACKED
;
3097 if (sacked
& TCPCB_SACKED_ACKED
) {
3098 tp
->sacked_out
-= acked_pcount
;
3099 } else if (tcp_is_sack(tp
)) {
3100 tp
->delivered
+= acked_pcount
;
3101 if (!tcp_skb_spurious_retrans(tp
, skb
))
3102 tcp_rack_advance(tp
, sacked
, scb
->end_seq
,
3103 tcp_skb_timestamp_us(skb
));
3105 if (sacked
& TCPCB_LOST
)
3106 tp
->lost_out
-= acked_pcount
;
3108 tp
->packets_out
-= acked_pcount
;
3109 pkts_acked
+= acked_pcount
;
3110 tcp_rate_skb_delivered(sk
, skb
, sack
->rate
);
3112 /* Initial outgoing SYN's get put onto the write_queue
3113 * just like anything else we transmit. It is not
3114 * true data, and if we misinform our callers that
3115 * this ACK acks real data, we will erroneously exit
3116 * connection startup slow start one packet too
3117 * quickly. This is severely frowned upon behavior.
3119 if (likely(!(scb
->tcp_flags
& TCPHDR_SYN
))) {
3120 flag
|= FLAG_DATA_ACKED
;
3122 flag
|= FLAG_SYN_ACKED
;
3123 tp
->retrans_stamp
= 0;
3129 next
= skb_rb_next(skb
);
3130 if (unlikely(skb
== tp
->retransmit_skb_hint
))
3131 tp
->retransmit_skb_hint
= NULL
;
3132 if (unlikely(skb
== tp
->lost_skb_hint
))
3133 tp
->lost_skb_hint
= NULL
;
3134 tcp_rtx_queue_unlink_and_free(skb
, sk
);
3138 tcp_chrono_stop(sk
, TCP_CHRONO_BUSY
);
3140 if (likely(between(tp
->snd_up
, prior_snd_una
, tp
->snd_una
)))
3141 tp
->snd_up
= tp
->snd_una
;
3143 if (skb
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
3144 flag
|= FLAG_SACK_RENEGING
;
3146 if (likely(first_ackt
) && !(flag
& FLAG_RETRANS_DATA_ACKED
)) {
3147 seq_rtt_us
= tcp_stamp_us_delta(tp
->tcp_mstamp
, first_ackt
);
3148 ca_rtt_us
= tcp_stamp_us_delta(tp
->tcp_mstamp
, last_ackt
);
3150 if (pkts_acked
== 1 && last_in_flight
< tp
->mss_cache
&&
3151 last_in_flight
&& !prior_sacked
&& fully_acked
&&
3152 sack
->rate
->prior_delivered
+ 1 == tp
->delivered
&&
3153 !(flag
& (FLAG_CA_ALERT
| FLAG_SYN_ACKED
))) {
3154 /* Conservatively mark a delayed ACK. It's typically
3155 * from a lone runt packet over the round trip to
3156 * a receiver w/o out-of-order or CE events.
3158 flag
|= FLAG_ACK_MAYBE_DELAYED
;
3161 if (sack
->first_sackt
) {
3162 sack_rtt_us
= tcp_stamp_us_delta(tp
->tcp_mstamp
, sack
->first_sackt
);
3163 ca_rtt_us
= tcp_stamp_us_delta(tp
->tcp_mstamp
, sack
->last_sackt
);
3165 rtt_update
= tcp_ack_update_rtt(sk
, flag
, seq_rtt_us
, sack_rtt_us
,
3166 ca_rtt_us
, sack
->rate
);
3168 if (flag
& FLAG_ACKED
) {
3169 flag
|= FLAG_SET_XMIT_TIMER
; /* set TLP or RTO timer */
3170 if (unlikely(icsk
->icsk_mtup
.probe_size
&&
3171 !after(tp
->mtu_probe
.probe_seq_end
, tp
->snd_una
))) {
3172 tcp_mtup_probe_success(sk
);
3175 if (tcp_is_reno(tp
)) {
3176 tcp_remove_reno_sacks(sk
, pkts_acked
);
3178 /* If any of the cumulatively ACKed segments was
3179 * retransmitted, non-SACK case cannot confirm that
3180 * progress was due to original transmission due to
3181 * lack of TCPCB_SACKED_ACKED bits even if some of
3182 * the packets may have been never retransmitted.
3184 if (flag
& FLAG_RETRANS_DATA_ACKED
)
3185 flag
&= ~FLAG_ORIG_SACK_ACKED
;
3189 /* Non-retransmitted hole got filled? That's reordering */
3190 if (before(reord
, prior_fack
))
3191 tcp_check_sack_reordering(sk
, reord
, 0);
3193 delta
= prior_sacked
- tp
->sacked_out
;
3194 tp
->lost_cnt_hint
-= min(tp
->lost_cnt_hint
, delta
);
3196 } else if (skb
&& rtt_update
&& sack_rtt_us
>= 0 &&
3197 sack_rtt_us
> tcp_stamp_us_delta(tp
->tcp_mstamp
,
3198 tcp_skb_timestamp_us(skb
))) {
3199 /* Do not re-arm RTO if the sack RTT is measured from data sent
3200 * after when the head was last (re)transmitted. Otherwise the
3201 * timeout may continue to extend in loss recovery.
3203 flag
|= FLAG_SET_XMIT_TIMER
; /* set TLP or RTO timer */
3206 if (icsk
->icsk_ca_ops
->pkts_acked
) {
3207 struct ack_sample sample
= { .pkts_acked
= pkts_acked
,
3208 .rtt_us
= sack
->rate
->rtt_us
,
3209 .in_flight
= last_in_flight
};
3211 icsk
->icsk_ca_ops
->pkts_acked(sk
, &sample
);
3214 #if FASTRETRANS_DEBUG > 0
3215 WARN_ON((int)tp
->sacked_out
< 0);
3216 WARN_ON((int)tp
->lost_out
< 0);
3217 WARN_ON((int)tp
->retrans_out
< 0);
3218 if (!tp
->packets_out
&& tcp_is_sack(tp
)) {
3219 icsk
= inet_csk(sk
);
3221 pr_debug("Leak l=%u %d\n",
3222 tp
->lost_out
, icsk
->icsk_ca_state
);
3225 if (tp
->sacked_out
) {
3226 pr_debug("Leak s=%u %d\n",
3227 tp
->sacked_out
, icsk
->icsk_ca_state
);
3230 if (tp
->retrans_out
) {
3231 pr_debug("Leak r=%u %d\n",
3232 tp
->retrans_out
, icsk
->icsk_ca_state
);
3233 tp
->retrans_out
= 0;
3240 static void tcp_ack_probe(struct sock
*sk
)
3242 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3243 struct sk_buff
*head
= tcp_send_head(sk
);
3244 const struct tcp_sock
*tp
= tcp_sk(sk
);
3246 /* Was it a usable window open? */
3249 if (!after(TCP_SKB_CB(head
)->end_seq
, tcp_wnd_end(tp
))) {
3250 icsk
->icsk_backoff
= 0;
3251 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
3252 /* Socket must be waked up by subsequent tcp_data_snd_check().
3253 * This function is not for random using!
3256 unsigned long when
= tcp_probe0_when(sk
, TCP_RTO_MAX
);
3258 tcp_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3259 when
, TCP_RTO_MAX
, NULL
);
3263 static inline bool tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
3265 return !(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
3266 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
;
3269 /* Decide wheather to run the increase function of congestion control. */
3270 static inline bool tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
3272 /* If reordering is high then always grow cwnd whenever data is
3273 * delivered regardless of its ordering. Otherwise stay conservative
3274 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3275 * new SACK or ECE mark may first advance cwnd here and later reduce
3276 * cwnd in tcp_fastretrans_alert() based on more states.
3278 if (tcp_sk(sk
)->reordering
> sock_net(sk
)->ipv4
.sysctl_tcp_reordering
)
3279 return flag
& FLAG_FORWARD_PROGRESS
;
3281 return flag
& FLAG_DATA_ACKED
;
3284 /* The "ultimate" congestion control function that aims to replace the rigid
3285 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3286 * It's called toward the end of processing an ACK with precise rate
3287 * information. All transmission or retransmission are delayed afterwards.
3289 static void tcp_cong_control(struct sock
*sk
, u32 ack
, u32 acked_sacked
,
3290 int flag
, const struct rate_sample
*rs
)
3292 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3294 if (icsk
->icsk_ca_ops
->cong_control
) {
3295 icsk
->icsk_ca_ops
->cong_control(sk
, rs
);
3299 if (tcp_in_cwnd_reduction(sk
)) {
3300 /* Reduce cwnd if state mandates */
3301 tcp_cwnd_reduction(sk
, acked_sacked
, flag
);
3302 } else if (tcp_may_raise_cwnd(sk
, flag
)) {
3303 /* Advance cwnd if state allows */
3304 tcp_cong_avoid(sk
, ack
, acked_sacked
);
3306 tcp_update_pacing_rate(sk
);
3309 /* Check that window update is acceptable.
3310 * The function assumes that snd_una<=ack<=snd_next.
3312 static inline bool tcp_may_update_window(const struct tcp_sock
*tp
,
3313 const u32 ack
, const u32 ack_seq
,
3316 return after(ack
, tp
->snd_una
) ||
3317 after(ack_seq
, tp
->snd_wl1
) ||
3318 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
);
3321 /* If we update tp->snd_una, also update tp->bytes_acked */
3322 static void tcp_snd_una_update(struct tcp_sock
*tp
, u32 ack
)
3324 u32 delta
= ack
- tp
->snd_una
;
3326 sock_owned_by_me((struct sock
*)tp
);
3327 tp
->bytes_acked
+= delta
;
3331 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3332 static void tcp_rcv_nxt_update(struct tcp_sock
*tp
, u32 seq
)
3334 u32 delta
= seq
- tp
->rcv_nxt
;
3336 sock_owned_by_me((struct sock
*)tp
);
3337 tp
->bytes_received
+= delta
;
3341 /* Update our send window.
3343 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3344 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3346 static int tcp_ack_update_window(struct sock
*sk
, const struct sk_buff
*skb
, u32 ack
,
3349 struct tcp_sock
*tp
= tcp_sk(sk
);
3351 u32 nwin
= ntohs(tcp_hdr(skb
)->window
);
3353 if (likely(!tcp_hdr(skb
)->syn
))
3354 nwin
<<= tp
->rx_opt
.snd_wscale
;
3356 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
3357 flag
|= FLAG_WIN_UPDATE
;
3358 tcp_update_wl(tp
, ack_seq
);
3360 if (tp
->snd_wnd
!= nwin
) {
3363 /* Note, it is the only place, where
3364 * fast path is recovered for sending TCP.
3367 tcp_fast_path_check(sk
);
3369 if (!tcp_write_queue_empty(sk
))
3370 tcp_slow_start_after_idle_check(sk
);
3372 if (nwin
> tp
->max_window
) {
3373 tp
->max_window
= nwin
;
3374 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
3379 tcp_snd_una_update(tp
, ack
);
3384 static bool __tcp_oow_rate_limited(struct net
*net
, int mib_idx
,
3385 u32
*last_oow_ack_time
)
3387 if (*last_oow_ack_time
) {
3388 s32 elapsed
= (s32
)(tcp_jiffies32
- *last_oow_ack_time
);
3390 if (0 <= elapsed
&& elapsed
< net
->ipv4
.sysctl_tcp_invalid_ratelimit
) {
3391 NET_INC_STATS(net
, mib_idx
);
3392 return true; /* rate-limited: don't send yet! */
3396 *last_oow_ack_time
= tcp_jiffies32
;
3398 return false; /* not rate-limited: go ahead, send dupack now! */
3401 /* Return true if we're currently rate-limiting out-of-window ACKs and
3402 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3403 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3404 * attacks that send repeated SYNs or ACKs for the same connection. To
3405 * do this, we do not send a duplicate SYNACK or ACK if the remote
3406 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3408 bool tcp_oow_rate_limited(struct net
*net
, const struct sk_buff
*skb
,
3409 int mib_idx
, u32
*last_oow_ack_time
)
3411 /* Data packets without SYNs are not likely part of an ACK loop. */
3412 if ((TCP_SKB_CB(skb
)->seq
!= TCP_SKB_CB(skb
)->end_seq
) &&
3416 return __tcp_oow_rate_limited(net
, mib_idx
, last_oow_ack_time
);
3419 /* RFC 5961 7 [ACK Throttling] */
3420 static void tcp_send_challenge_ack(struct sock
*sk
, const struct sk_buff
*skb
)
3422 /* unprotected vars, we dont care of overwrites */
3423 static u32 challenge_timestamp
;
3424 static unsigned int challenge_count
;
3425 struct tcp_sock
*tp
= tcp_sk(sk
);
3426 struct net
*net
= sock_net(sk
);
3429 /* First check our per-socket dupack rate limit. */
3430 if (__tcp_oow_rate_limited(net
,
3431 LINUX_MIB_TCPACKSKIPPEDCHALLENGE
,
3432 &tp
->last_oow_ack_time
))
3435 /* Then check host-wide RFC 5961 rate limit. */
3437 if (now
!= challenge_timestamp
) {
3438 u32 ack_limit
= net
->ipv4
.sysctl_tcp_challenge_ack_limit
;
3439 u32 half
= (ack_limit
+ 1) >> 1;
3441 challenge_timestamp
= now
;
3442 WRITE_ONCE(challenge_count
, half
+ prandom_u32_max(ack_limit
));
3444 count
= READ_ONCE(challenge_count
);
3446 WRITE_ONCE(challenge_count
, count
- 1);
3447 NET_INC_STATS(net
, LINUX_MIB_TCPCHALLENGEACK
);
3452 static void tcp_store_ts_recent(struct tcp_sock
*tp
)
3454 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
3455 tp
->rx_opt
.ts_recent_stamp
= ktime_get_seconds();
3458 static void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
3460 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
3461 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3462 * extra check below makes sure this can only happen
3463 * for pure ACK frames. -DaveM
3465 * Not only, also it occurs for expired timestamps.
3468 if (tcp_paws_check(&tp
->rx_opt
, 0))
3469 tcp_store_ts_recent(tp
);
3473 /* This routine deals with acks during a TLP episode.
3474 * We mark the end of a TLP episode on receiving TLP dupack or when
3475 * ack is after tlp_high_seq.
3476 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3478 static void tcp_process_tlp_ack(struct sock
*sk
, u32 ack
, int flag
)
3480 struct tcp_sock
*tp
= tcp_sk(sk
);
3482 if (before(ack
, tp
->tlp_high_seq
))
3485 if (flag
& FLAG_DSACKING_ACK
) {
3486 /* This DSACK means original and TLP probe arrived; no loss */
3487 tp
->tlp_high_seq
= 0;
3488 } else if (after(ack
, tp
->tlp_high_seq
)) {
3489 /* ACK advances: there was a loss, so reduce cwnd. Reset
3490 * tlp_high_seq in tcp_init_cwnd_reduction()
3492 tcp_init_cwnd_reduction(sk
);
3493 tcp_set_ca_state(sk
, TCP_CA_CWR
);
3494 tcp_end_cwnd_reduction(sk
);
3495 tcp_try_keep_open(sk
);
3496 NET_INC_STATS(sock_net(sk
),
3497 LINUX_MIB_TCPLOSSPROBERECOVERY
);
3498 } else if (!(flag
& (FLAG_SND_UNA_ADVANCED
|
3499 FLAG_NOT_DUP
| FLAG_DATA_SACKED
))) {
3500 /* Pure dupack: original and TLP probe arrived; no loss */
3501 tp
->tlp_high_seq
= 0;
3505 static inline void tcp_in_ack_event(struct sock
*sk
, u32 flags
)
3507 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3509 if (icsk
->icsk_ca_ops
->in_ack_event
)
3510 icsk
->icsk_ca_ops
->in_ack_event(sk
, flags
);
3513 /* Congestion control has updated the cwnd already. So if we're in
3514 * loss recovery then now we do any new sends (for FRTO) or
3515 * retransmits (for CA_Loss or CA_recovery) that make sense.
3517 static void tcp_xmit_recovery(struct sock
*sk
, int rexmit
)
3519 struct tcp_sock
*tp
= tcp_sk(sk
);
3521 if (rexmit
== REXMIT_NONE
)
3524 if (unlikely(rexmit
== 2)) {
3525 __tcp_push_pending_frames(sk
, tcp_current_mss(sk
),
3527 if (after(tp
->snd_nxt
, tp
->high_seq
))
3531 tcp_xmit_retransmit_queue(sk
);
3534 /* Returns the number of packets newly acked or sacked by the current ACK */
3535 static u32
tcp_newly_delivered(struct sock
*sk
, u32 prior_delivered
, int flag
)
3537 const struct net
*net
= sock_net(sk
);
3538 struct tcp_sock
*tp
= tcp_sk(sk
);
3541 delivered
= tp
->delivered
- prior_delivered
;
3542 NET_ADD_STATS(net
, LINUX_MIB_TCPDELIVERED
, delivered
);
3543 if (flag
& FLAG_ECE
) {
3544 tp
->delivered_ce
+= delivered
;
3545 NET_ADD_STATS(net
, LINUX_MIB_TCPDELIVEREDCE
, delivered
);
3550 /* This routine deals with incoming acks, but not outgoing ones. */
3551 static int tcp_ack(struct sock
*sk
, const struct sk_buff
*skb
, int flag
)
3553 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3554 struct tcp_sock
*tp
= tcp_sk(sk
);
3555 struct tcp_sacktag_state sack_state
;
3556 struct rate_sample rs
= { .prior_delivered
= 0 };
3557 u32 prior_snd_una
= tp
->snd_una
;
3558 bool is_sack_reneg
= tp
->is_sack_reneg
;
3559 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
3560 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3561 bool is_dupack
= false;
3562 int prior_packets
= tp
->packets_out
;
3563 u32 delivered
= tp
->delivered
;
3564 u32 lost
= tp
->lost
;
3565 int rexmit
= REXMIT_NONE
; /* Flag to (re)transmit to recover losses */
3568 sack_state
.first_sackt
= 0;
3569 sack_state
.rate
= &rs
;
3571 /* We very likely will need to access rtx queue. */
3572 prefetch(sk
->tcp_rtx_queue
.rb_node
);
3574 /* If the ack is older than previous acks
3575 * then we can probably ignore it.
3577 if (before(ack
, prior_snd_una
)) {
3578 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3579 if (before(ack
, prior_snd_una
- tp
->max_window
)) {
3580 if (!(flag
& FLAG_NO_CHALLENGE_ACK
))
3581 tcp_send_challenge_ack(sk
, skb
);
3587 /* If the ack includes data we haven't sent yet, discard
3588 * this segment (RFC793 Section 3.9).
3590 if (after(ack
, tp
->snd_nxt
))
3593 if (after(ack
, prior_snd_una
)) {
3594 flag
|= FLAG_SND_UNA_ADVANCED
;
3595 icsk
->icsk_retransmits
= 0;
3597 #if IS_ENABLED(CONFIG_TLS_DEVICE)
3598 if (static_branch_unlikely(&clean_acked_data_enabled
))
3599 if (icsk
->icsk_clean_acked
)
3600 icsk
->icsk_clean_acked(sk
, ack
);
3604 prior_fack
= tcp_is_sack(tp
) ? tcp_highest_sack_seq(tp
) : tp
->snd_una
;
3605 rs
.prior_in_flight
= tcp_packets_in_flight(tp
);
3607 /* ts_recent update must be made after we are sure that the packet
3610 if (flag
& FLAG_UPDATE_TS_RECENT
)
3611 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
3613 if (!(flag
& FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
3614 /* Window is constant, pure forward advance.
3615 * No more checks are required.
3616 * Note, we use the fact that SND.UNA>=SND.WL2.
3618 tcp_update_wl(tp
, ack_seq
);
3619 tcp_snd_una_update(tp
, ack
);
3620 flag
|= FLAG_WIN_UPDATE
;
3622 tcp_in_ack_event(sk
, CA_ACK_WIN_UPDATE
);
3624 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPHPACKS
);
3626 u32 ack_ev_flags
= CA_ACK_SLOWPATH
;
3628 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
3631 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPPUREACKS
);
3633 flag
|= tcp_ack_update_window(sk
, skb
, ack
, ack_seq
);
3635 if (TCP_SKB_CB(skb
)->sacked
)
3636 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
,
3639 if (tcp_ecn_rcv_ecn_echo(tp
, tcp_hdr(skb
))) {
3641 ack_ev_flags
|= CA_ACK_ECE
;
3644 if (flag
& FLAG_WIN_UPDATE
)
3645 ack_ev_flags
|= CA_ACK_WIN_UPDATE
;
3647 tcp_in_ack_event(sk
, ack_ev_flags
);
3650 /* We passed data and got it acked, remove any soft error
3651 * log. Something worked...
3653 sk
->sk_err_soft
= 0;
3654 icsk
->icsk_probes_out
= 0;
3655 tp
->rcv_tstamp
= tcp_jiffies32
;
3659 /* See if we can take anything off of the retransmit queue. */
3660 flag
|= tcp_clean_rtx_queue(sk
, prior_fack
, prior_snd_una
, &sack_state
);
3662 tcp_rack_update_reo_wnd(sk
, &rs
);
3664 if (tp
->tlp_high_seq
)
3665 tcp_process_tlp_ack(sk
, ack
, flag
);
3666 /* If needed, reset TLP/RTO timer; RACK may later override this. */
3667 if (flag
& FLAG_SET_XMIT_TIMER
)
3668 tcp_set_xmit_timer(sk
);
3670 if (tcp_ack_is_dubious(sk
, flag
)) {
3671 is_dupack
= !(flag
& (FLAG_SND_UNA_ADVANCED
| FLAG_NOT_DUP
));
3672 tcp_fastretrans_alert(sk
, prior_snd_una
, is_dupack
, &flag
,
3676 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
& FLAG_NOT_DUP
))
3679 delivered
= tcp_newly_delivered(sk
, delivered
, flag
);
3680 lost
= tp
->lost
- lost
; /* freshly marked lost */
3681 rs
.is_ack_delayed
= !!(flag
& FLAG_ACK_MAYBE_DELAYED
);
3682 tcp_rate_gen(sk
, delivered
, lost
, is_sack_reneg
, sack_state
.rate
);
3683 tcp_cong_control(sk
, ack
, delivered
, flag
, sack_state
.rate
);
3684 tcp_xmit_recovery(sk
, rexmit
);
3688 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3689 if (flag
& FLAG_DSACKING_ACK
) {
3690 tcp_fastretrans_alert(sk
, prior_snd_una
, is_dupack
, &flag
,
3692 tcp_newly_delivered(sk
, delivered
, flag
);
3694 /* If this ack opens up a zero window, clear backoff. It was
3695 * being used to time the probes, and is probably far higher than
3696 * it needs to be for normal retransmission.
3700 if (tp
->tlp_high_seq
)
3701 tcp_process_tlp_ack(sk
, ack
, flag
);
3705 SOCK_DEBUG(sk
, "Ack %u after %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3709 /* If data was SACKed, tag it and see if we should send more data.
3710 * If data was DSACKed, see if we can undo a cwnd reduction.
3712 if (TCP_SKB_CB(skb
)->sacked
) {
3713 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
,
3715 tcp_fastretrans_alert(sk
, prior_snd_una
, is_dupack
, &flag
,
3717 tcp_newly_delivered(sk
, delivered
, flag
);
3718 tcp_xmit_recovery(sk
, rexmit
);
3721 SOCK_DEBUG(sk
, "Ack %u before %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3725 static void tcp_parse_fastopen_option(int len
, const unsigned char *cookie
,
3726 bool syn
, struct tcp_fastopen_cookie
*foc
,
3729 /* Valid only in SYN or SYN-ACK with an even length. */
3730 if (!foc
|| !syn
|| len
< 0 || (len
& 1))
3733 if (len
>= TCP_FASTOPEN_COOKIE_MIN
&&
3734 len
<= TCP_FASTOPEN_COOKIE_MAX
)
3735 memcpy(foc
->val
, cookie
, len
);
3742 static void smc_parse_options(const struct tcphdr
*th
,
3743 struct tcp_options_received
*opt_rx
,
3744 const unsigned char *ptr
,
3747 #if IS_ENABLED(CONFIG_SMC)
3748 if (static_branch_unlikely(&tcp_have_smc
)) {
3749 if (th
->syn
&& !(opsize
& 1) &&
3750 opsize
>= TCPOLEN_EXP_SMC_BASE
&&
3751 get_unaligned_be32(ptr
) == TCPOPT_SMC_MAGIC
)
3757 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3758 * But, this can also be called on packets in the established flow when
3759 * the fast version below fails.
3761 void tcp_parse_options(const struct net
*net
,
3762 const struct sk_buff
*skb
,
3763 struct tcp_options_received
*opt_rx
, int estab
,
3764 struct tcp_fastopen_cookie
*foc
)
3766 const unsigned char *ptr
;
3767 const struct tcphdr
*th
= tcp_hdr(skb
);
3768 int length
= (th
->doff
* 4) - sizeof(struct tcphdr
);
3770 ptr
= (const unsigned char *)(th
+ 1);
3771 opt_rx
->saw_tstamp
= 0;
3773 while (length
> 0) {
3774 int opcode
= *ptr
++;
3780 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
3785 if (opsize
< 2) /* "silly options" */
3787 if (opsize
> length
)
3788 return; /* don't parse partial options */
3791 if (opsize
== TCPOLEN_MSS
&& th
->syn
&& !estab
) {
3792 u16 in_mss
= get_unaligned_be16(ptr
);
3794 if (opt_rx
->user_mss
&&
3795 opt_rx
->user_mss
< in_mss
)
3796 in_mss
= opt_rx
->user_mss
;
3797 opt_rx
->mss_clamp
= in_mss
;
3802 if (opsize
== TCPOLEN_WINDOW
&& th
->syn
&&
3803 !estab
&& net
->ipv4
.sysctl_tcp_window_scaling
) {
3804 __u8 snd_wscale
= *(__u8
*)ptr
;
3805 opt_rx
->wscale_ok
= 1;
3806 if (snd_wscale
> TCP_MAX_WSCALE
) {
3807 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n",
3811 snd_wscale
= TCP_MAX_WSCALE
;
3813 opt_rx
->snd_wscale
= snd_wscale
;
3816 case TCPOPT_TIMESTAMP
:
3817 if ((opsize
== TCPOLEN_TIMESTAMP
) &&
3818 ((estab
&& opt_rx
->tstamp_ok
) ||
3819 (!estab
&& net
->ipv4
.sysctl_tcp_timestamps
))) {
3820 opt_rx
->saw_tstamp
= 1;
3821 opt_rx
->rcv_tsval
= get_unaligned_be32(ptr
);
3822 opt_rx
->rcv_tsecr
= get_unaligned_be32(ptr
+ 4);
3825 case TCPOPT_SACK_PERM
:
3826 if (opsize
== TCPOLEN_SACK_PERM
&& th
->syn
&&
3827 !estab
&& net
->ipv4
.sysctl_tcp_sack
) {
3828 opt_rx
->sack_ok
= TCP_SACK_SEEN
;
3829 tcp_sack_reset(opt_rx
);
3834 if ((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
3835 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
3837 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
3840 #ifdef CONFIG_TCP_MD5SIG
3843 * The MD5 Hash has already been
3844 * checked (see tcp_v{4,6}_do_rcv()).
3848 case TCPOPT_FASTOPEN
:
3849 tcp_parse_fastopen_option(
3850 opsize
- TCPOLEN_FASTOPEN_BASE
,
3851 ptr
, th
->syn
, foc
, false);
3855 /* Fast Open option shares code 254 using a
3856 * 16 bits magic number.
3858 if (opsize
>= TCPOLEN_EXP_FASTOPEN_BASE
&&
3859 get_unaligned_be16(ptr
) ==
3860 TCPOPT_FASTOPEN_MAGIC
)
3861 tcp_parse_fastopen_option(opsize
-
3862 TCPOLEN_EXP_FASTOPEN_BASE
,
3863 ptr
+ 2, th
->syn
, foc
, true);
3865 smc_parse_options(th
, opt_rx
, ptr
,
3875 EXPORT_SYMBOL(tcp_parse_options
);
3877 static bool tcp_parse_aligned_timestamp(struct tcp_sock
*tp
, const struct tcphdr
*th
)
3879 const __be32
*ptr
= (const __be32
*)(th
+ 1);
3881 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3882 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
3883 tp
->rx_opt
.saw_tstamp
= 1;
3885 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3888 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
) - tp
->tsoffset
;
3890 tp
->rx_opt
.rcv_tsecr
= 0;
3896 /* Fast parse options. This hopes to only see timestamps.
3897 * If it is wrong it falls back on tcp_parse_options().
3899 static bool tcp_fast_parse_options(const struct net
*net
,
3900 const struct sk_buff
*skb
,
3901 const struct tcphdr
*th
, struct tcp_sock
*tp
)
3903 /* In the spirit of fast parsing, compare doff directly to constant
3904 * values. Because equality is used, short doff can be ignored here.
3906 if (th
->doff
== (sizeof(*th
) / 4)) {
3907 tp
->rx_opt
.saw_tstamp
= 0;
3909 } else if (tp
->rx_opt
.tstamp_ok
&&
3910 th
->doff
== ((sizeof(*th
) + TCPOLEN_TSTAMP_ALIGNED
) / 4)) {
3911 if (tcp_parse_aligned_timestamp(tp
, th
))
3915 tcp_parse_options(net
, skb
, &tp
->rx_opt
, 1, NULL
);
3916 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
3917 tp
->rx_opt
.rcv_tsecr
-= tp
->tsoffset
;
3922 #ifdef CONFIG_TCP_MD5SIG
3924 * Parse MD5 Signature option
3926 const u8
*tcp_parse_md5sig_option(const struct tcphdr
*th
)
3928 int length
= (th
->doff
<< 2) - sizeof(*th
);
3929 const u8
*ptr
= (const u8
*)(th
+ 1);
3931 /* If not enough data remaining, we can short cut */
3932 while (length
>= TCPOLEN_MD5SIG
) {
3933 int opcode
= *ptr
++;
3944 if (opsize
< 2 || opsize
> length
)
3946 if (opcode
== TCPOPT_MD5SIG
)
3947 return opsize
== TCPOLEN_MD5SIG
? ptr
: NULL
;
3954 EXPORT_SYMBOL(tcp_parse_md5sig_option
);
3957 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3959 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3960 * it can pass through stack. So, the following predicate verifies that
3961 * this segment is not used for anything but congestion avoidance or
3962 * fast retransmit. Moreover, we even are able to eliminate most of such
3963 * second order effects, if we apply some small "replay" window (~RTO)
3964 * to timestamp space.
3966 * All these measures still do not guarantee that we reject wrapped ACKs
3967 * on networks with high bandwidth, when sequence space is recycled fastly,
3968 * but it guarantees that such events will be very rare and do not affect
3969 * connection seriously. This doesn't look nice, but alas, PAWS is really
3972 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3973 * states that events when retransmit arrives after original data are rare.
3974 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3975 * the biggest problem on large power networks even with minor reordering.
3976 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3977 * up to bandwidth of 18Gigabit/sec. 8) ]
3980 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
3982 const struct tcp_sock
*tp
= tcp_sk(sk
);
3983 const struct tcphdr
*th
= tcp_hdr(skb
);
3984 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3985 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3987 return (/* 1. Pure ACK with correct sequence number. */
3988 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
3990 /* 2. ... and duplicate ACK. */
3991 ack
== tp
->snd_una
&&
3993 /* 3. ... and does not update window. */
3994 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
3996 /* 4. ... and sits in replay window. */
3997 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
4000 static inline bool tcp_paws_discard(const struct sock
*sk
,
4001 const struct sk_buff
*skb
)
4003 const struct tcp_sock
*tp
= tcp_sk(sk
);
4005 return !tcp_paws_check(&tp
->rx_opt
, TCP_PAWS_WINDOW
) &&
4006 !tcp_disordered_ack(sk
, skb
);
4009 /* Check segment sequence number for validity.
4011 * Segment controls are considered valid, if the segment
4012 * fits to the window after truncation to the window. Acceptability
4013 * of data (and SYN, FIN, of course) is checked separately.
4014 * See tcp_data_queue(), for example.
4016 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4017 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4018 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4019 * (borrowed from freebsd)
4022 static inline bool tcp_sequence(const struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
4024 return !before(end_seq
, tp
->rcv_wup
) &&
4025 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
4028 /* When we get a reset we do this. */
4029 void tcp_reset(struct sock
*sk
)
4031 trace_tcp_receive_reset(sk
);
4033 /* We want the right error as BSD sees it (and indeed as we do). */
4034 switch (sk
->sk_state
) {
4036 sk
->sk_err
= ECONNREFUSED
;
4038 case TCP_CLOSE_WAIT
:
4044 sk
->sk_err
= ECONNRESET
;
4046 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4049 tcp_write_queue_purge(sk
);
4052 if (!sock_flag(sk
, SOCK_DEAD
))
4053 sk
->sk_error_report(sk
);
4057 * Process the FIN bit. This now behaves as it is supposed to work
4058 * and the FIN takes effect when it is validly part of sequence
4059 * space. Not before when we get holes.
4061 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4062 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4065 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4066 * close and we go into CLOSING (and later onto TIME-WAIT)
4068 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4070 void tcp_fin(struct sock
*sk
)
4072 struct tcp_sock
*tp
= tcp_sk(sk
);
4074 inet_csk_schedule_ack(sk
);
4076 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
4077 sock_set_flag(sk
, SOCK_DONE
);
4079 switch (sk
->sk_state
) {
4081 case TCP_ESTABLISHED
:
4082 /* Move to CLOSE_WAIT */
4083 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
4084 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
4087 case TCP_CLOSE_WAIT
:
4089 /* Received a retransmission of the FIN, do
4094 /* RFC793: Remain in the LAST-ACK state. */
4098 /* This case occurs when a simultaneous close
4099 * happens, we must ack the received FIN and
4100 * enter the CLOSING state.
4103 tcp_set_state(sk
, TCP_CLOSING
);
4106 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4108 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
4111 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4112 * cases we should never reach this piece of code.
4114 pr_err("%s: Impossible, sk->sk_state=%d\n",
4115 __func__
, sk
->sk_state
);
4119 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4120 * Probably, we should reset in this case. For now drop them.
4122 skb_rbtree_purge(&tp
->out_of_order_queue
);
4123 if (tcp_is_sack(tp
))
4124 tcp_sack_reset(&tp
->rx_opt
);
4127 if (!sock_flag(sk
, SOCK_DEAD
)) {
4128 sk
->sk_state_change(sk
);
4130 /* Do not send POLL_HUP for half duplex close. */
4131 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
4132 sk
->sk_state
== TCP_CLOSE
)
4133 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_HUP
);
4135 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_IN
);
4139 static inline bool tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
,
4142 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
4143 if (before(seq
, sp
->start_seq
))
4144 sp
->start_seq
= seq
;
4145 if (after(end_seq
, sp
->end_seq
))
4146 sp
->end_seq
= end_seq
;
4152 static void tcp_dsack_set(struct sock
*sk
, u32 seq
, u32 end_seq
)
4154 struct tcp_sock
*tp
= tcp_sk(sk
);
4156 if (tcp_is_sack(tp
) && sock_net(sk
)->ipv4
.sysctl_tcp_dsack
) {
4159 if (before(seq
, tp
->rcv_nxt
))
4160 mib_idx
= LINUX_MIB_TCPDSACKOLDSENT
;
4162 mib_idx
= LINUX_MIB_TCPDSACKOFOSENT
;
4164 NET_INC_STATS(sock_net(sk
), mib_idx
);
4166 tp
->rx_opt
.dsack
= 1;
4167 tp
->duplicate_sack
[0].start_seq
= seq
;
4168 tp
->duplicate_sack
[0].end_seq
= end_seq
;
4172 static void tcp_dsack_extend(struct sock
*sk
, u32 seq
, u32 end_seq
)
4174 struct tcp_sock
*tp
= tcp_sk(sk
);
4176 if (!tp
->rx_opt
.dsack
)
4177 tcp_dsack_set(sk
, seq
, end_seq
);
4179 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
4182 static void tcp_rcv_spurious_retrans(struct sock
*sk
, const struct sk_buff
*skb
)
4184 /* When the ACK path fails or drops most ACKs, the sender would
4185 * timeout and spuriously retransmit the same segment repeatedly.
4186 * The receiver remembers and reflects via DSACKs. Leverage the
4187 * DSACK state and change the txhash to re-route speculatively.
4189 if (TCP_SKB_CB(skb
)->seq
== tcp_sk(sk
)->duplicate_sack
[0].start_seq
)
4190 sk_rethink_txhash(sk
);
4193 static void tcp_send_dupack(struct sock
*sk
, const struct sk_buff
*skb
)
4195 struct tcp_sock
*tp
= tcp_sk(sk
);
4197 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
4198 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4199 NET_INC_STATS(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4200 tcp_enter_quickack_mode(sk
, TCP_MAX_QUICKACKS
);
4202 if (tcp_is_sack(tp
) && sock_net(sk
)->ipv4
.sysctl_tcp_dsack
) {
4203 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4205 tcp_rcv_spurious_retrans(sk
, skb
);
4206 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
4207 end_seq
= tp
->rcv_nxt
;
4208 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, end_seq
);
4215 /* These routines update the SACK block as out-of-order packets arrive or
4216 * in-order packets close up the sequence space.
4218 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
4221 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4222 struct tcp_sack_block
*swalk
= sp
+ 1;
4224 /* See if the recent change to the first SACK eats into
4225 * or hits the sequence space of other SACK blocks, if so coalesce.
4227 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
;) {
4228 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
4231 /* Zap SWALK, by moving every further SACK up by one slot.
4232 * Decrease num_sacks.
4234 tp
->rx_opt
.num_sacks
--;
4235 for (i
= this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
4239 this_sack
++, swalk
++;
4243 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
4245 struct tcp_sock
*tp
= tcp_sk(sk
);
4246 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4247 int cur_sacks
= tp
->rx_opt
.num_sacks
;
4253 for (this_sack
= 0; this_sack
< cur_sacks
; this_sack
++, sp
++) {
4254 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
4255 /* Rotate this_sack to the first one. */
4256 for (; this_sack
> 0; this_sack
--, sp
--)
4257 swap(*sp
, *(sp
- 1));
4259 tcp_sack_maybe_coalesce(tp
);
4264 /* Could not find an adjacent existing SACK, build a new one,
4265 * put it at the front, and shift everyone else down. We
4266 * always know there is at least one SACK present already here.
4268 * If the sack array is full, forget about the last one.
4270 if (this_sack
>= TCP_NUM_SACKS
) {
4271 if (tp
->compressed_ack
> TCP_FASTRETRANS_THRESH
)
4274 tp
->rx_opt
.num_sacks
--;
4277 for (; this_sack
> 0; this_sack
--, sp
--)
4281 /* Build the new head SACK, and we're done. */
4282 sp
->start_seq
= seq
;
4283 sp
->end_seq
= end_seq
;
4284 tp
->rx_opt
.num_sacks
++;
4287 /* RCV.NXT advances, some SACKs should be eaten. */
4289 static void tcp_sack_remove(struct tcp_sock
*tp
)
4291 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4292 int num_sacks
= tp
->rx_opt
.num_sacks
;
4295 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4296 if (RB_EMPTY_ROOT(&tp
->out_of_order_queue
)) {
4297 tp
->rx_opt
.num_sacks
= 0;
4301 for (this_sack
= 0; this_sack
< num_sacks
;) {
4302 /* Check if the start of the sack is covered by RCV.NXT. */
4303 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
4306 /* RCV.NXT must cover all the block! */
4307 WARN_ON(before(tp
->rcv_nxt
, sp
->end_seq
));
4309 /* Zap this SACK, by moving forward any other SACKS. */
4310 for (i
= this_sack
+1; i
< num_sacks
; i
++)
4311 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
4318 tp
->rx_opt
.num_sacks
= num_sacks
;
4322 * tcp_try_coalesce - try to merge skb to prior one
4324 * @dest: destination queue
4326 * @from: buffer to add in queue
4327 * @fragstolen: pointer to boolean
4329 * Before queueing skb @from after @to, try to merge them
4330 * to reduce overall memory use and queue lengths, if cost is small.
4331 * Packets in ofo or receive queues can stay a long time.
4332 * Better try to coalesce them right now to avoid future collapses.
4333 * Returns true if caller should free @from instead of queueing it
4335 static bool tcp_try_coalesce(struct sock
*sk
,
4337 struct sk_buff
*from
,
4342 *fragstolen
= false;
4344 /* Its possible this segment overlaps with prior segment in queue */
4345 if (TCP_SKB_CB(from
)->seq
!= TCP_SKB_CB(to
)->end_seq
)
4348 #ifdef CONFIG_TLS_DEVICE
4349 if (from
->decrypted
!= to
->decrypted
)
4353 if (!skb_try_coalesce(to
, from
, fragstolen
, &delta
))
4356 atomic_add(delta
, &sk
->sk_rmem_alloc
);
4357 sk_mem_charge(sk
, delta
);
4358 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPRCVCOALESCE
);
4359 TCP_SKB_CB(to
)->end_seq
= TCP_SKB_CB(from
)->end_seq
;
4360 TCP_SKB_CB(to
)->ack_seq
= TCP_SKB_CB(from
)->ack_seq
;
4361 TCP_SKB_CB(to
)->tcp_flags
|= TCP_SKB_CB(from
)->tcp_flags
;
4363 if (TCP_SKB_CB(from
)->has_rxtstamp
) {
4364 TCP_SKB_CB(to
)->has_rxtstamp
= true;
4365 to
->tstamp
= from
->tstamp
;
4366 skb_hwtstamps(to
)->hwtstamp
= skb_hwtstamps(from
)->hwtstamp
;
4372 static bool tcp_ooo_try_coalesce(struct sock
*sk
,
4374 struct sk_buff
*from
,
4377 bool res
= tcp_try_coalesce(sk
, to
, from
, fragstolen
);
4379 /* In case tcp_drop() is called later, update to->gso_segs */
4381 u32 gso_segs
= max_t(u16
, 1, skb_shinfo(to
)->gso_segs
) +
4382 max_t(u16
, 1, skb_shinfo(from
)->gso_segs
);
4384 skb_shinfo(to
)->gso_segs
= min_t(u32
, gso_segs
, 0xFFFF);
4389 static void tcp_drop(struct sock
*sk
, struct sk_buff
*skb
)
4391 sk_drops_add(sk
, skb
);
4395 /* This one checks to see if we can put data from the
4396 * out_of_order queue into the receive_queue.
4398 static void tcp_ofo_queue(struct sock
*sk
)
4400 struct tcp_sock
*tp
= tcp_sk(sk
);
4401 __u32 dsack_high
= tp
->rcv_nxt
;
4402 bool fin
, fragstolen
, eaten
;
4403 struct sk_buff
*skb
, *tail
;
4406 p
= rb_first(&tp
->out_of_order_queue
);
4409 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
4412 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
4413 __u32 dsack
= dsack_high
;
4414 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
4415 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
4416 tcp_dsack_extend(sk
, TCP_SKB_CB(skb
)->seq
, dsack
);
4419 rb_erase(&skb
->rbnode
, &tp
->out_of_order_queue
);
4421 if (unlikely(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))) {
4422 SOCK_DEBUG(sk
, "ofo packet was already received\n");
4426 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
4427 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4428 TCP_SKB_CB(skb
)->end_seq
);
4430 tail
= skb_peek_tail(&sk
->sk_receive_queue
);
4431 eaten
= tail
&& tcp_try_coalesce(sk
, tail
, skb
, &fragstolen
);
4432 tcp_rcv_nxt_update(tp
, TCP_SKB_CB(skb
)->end_seq
);
4433 fin
= TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
;
4435 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4437 kfree_skb_partial(skb
, fragstolen
);
4439 if (unlikely(fin
)) {
4441 /* tcp_fin() purges tp->out_of_order_queue,
4442 * so we must end this loop right now.
4449 static bool tcp_prune_ofo_queue(struct sock
*sk
);
4450 static int tcp_prune_queue(struct sock
*sk
);
4452 static int tcp_try_rmem_schedule(struct sock
*sk
, struct sk_buff
*skb
,
4455 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
4456 !sk_rmem_schedule(sk
, skb
, size
)) {
4458 if (tcp_prune_queue(sk
) < 0)
4461 while (!sk_rmem_schedule(sk
, skb
, size
)) {
4462 if (!tcp_prune_ofo_queue(sk
))
4469 static void tcp_data_queue_ofo(struct sock
*sk
, struct sk_buff
*skb
)
4471 struct tcp_sock
*tp
= tcp_sk(sk
);
4472 struct rb_node
**p
, *parent
;
4473 struct sk_buff
*skb1
;
4477 tcp_ecn_check_ce(sk
, skb
);
4479 if (unlikely(tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))) {
4480 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPOFODROP
);
4485 /* Disable header prediction. */
4487 inet_csk_schedule_ack(sk
);
4489 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPOFOQUEUE
);
4490 seq
= TCP_SKB_CB(skb
)->seq
;
4491 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4492 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
4493 tp
->rcv_nxt
, seq
, end_seq
);
4495 p
= &tp
->out_of_order_queue
.rb_node
;
4496 if (RB_EMPTY_ROOT(&tp
->out_of_order_queue
)) {
4497 /* Initial out of order segment, build 1 SACK. */
4498 if (tcp_is_sack(tp
)) {
4499 tp
->rx_opt
.num_sacks
= 1;
4500 tp
->selective_acks
[0].start_seq
= seq
;
4501 tp
->selective_acks
[0].end_seq
= end_seq
;
4503 rb_link_node(&skb
->rbnode
, NULL
, p
);
4504 rb_insert_color(&skb
->rbnode
, &tp
->out_of_order_queue
);
4505 tp
->ooo_last_skb
= skb
;
4509 /* In the typical case, we are adding an skb to the end of the list.
4510 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4512 if (tcp_ooo_try_coalesce(sk
, tp
->ooo_last_skb
,
4513 skb
, &fragstolen
)) {
4515 tcp_grow_window(sk
, skb
);
4516 kfree_skb_partial(skb
, fragstolen
);
4520 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4521 if (!before(seq
, TCP_SKB_CB(tp
->ooo_last_skb
)->end_seq
)) {
4522 parent
= &tp
->ooo_last_skb
->rbnode
;
4523 p
= &parent
->rb_right
;
4527 /* Find place to insert this segment. Handle overlaps on the way. */
4531 skb1
= rb_to_skb(parent
);
4532 if (before(seq
, TCP_SKB_CB(skb1
)->seq
)) {
4533 p
= &parent
->rb_left
;
4536 if (before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4537 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4538 /* All the bits are present. Drop. */
4539 NET_INC_STATS(sock_net(sk
),
4540 LINUX_MIB_TCPOFOMERGE
);
4543 tcp_dsack_set(sk
, seq
, end_seq
);
4546 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
4547 /* Partial overlap. */
4548 tcp_dsack_set(sk
, seq
, TCP_SKB_CB(skb1
)->end_seq
);
4550 /* skb's seq == skb1's seq and skb covers skb1.
4551 * Replace skb1 with skb.
4553 rb_replace_node(&skb1
->rbnode
, &skb
->rbnode
,
4554 &tp
->out_of_order_queue
);
4555 tcp_dsack_extend(sk
,
4556 TCP_SKB_CB(skb1
)->seq
,
4557 TCP_SKB_CB(skb1
)->end_seq
);
4558 NET_INC_STATS(sock_net(sk
),
4559 LINUX_MIB_TCPOFOMERGE
);
4563 } else if (tcp_ooo_try_coalesce(sk
, skb1
,
4564 skb
, &fragstolen
)) {
4567 p
= &parent
->rb_right
;
4570 /* Insert segment into RB tree. */
4571 rb_link_node(&skb
->rbnode
, parent
, p
);
4572 rb_insert_color(&skb
->rbnode
, &tp
->out_of_order_queue
);
4575 /* Remove other segments covered by skb. */
4576 while ((skb1
= skb_rb_next(skb
)) != NULL
) {
4577 if (!after(end_seq
, TCP_SKB_CB(skb1
)->seq
))
4579 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4580 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4584 rb_erase(&skb1
->rbnode
, &tp
->out_of_order_queue
);
4585 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4586 TCP_SKB_CB(skb1
)->end_seq
);
4587 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPOFOMERGE
);
4590 /* If there is no skb after us, we are the last_skb ! */
4592 tp
->ooo_last_skb
= skb
;
4595 if (tcp_is_sack(tp
))
4596 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
4599 tcp_grow_window(sk
, skb
);
4601 skb_set_owner_r(skb
, sk
);
4605 static int __must_check
tcp_queue_rcv(struct sock
*sk
, struct sk_buff
*skb
, int hdrlen
,
4609 struct sk_buff
*tail
= skb_peek_tail(&sk
->sk_receive_queue
);
4611 __skb_pull(skb
, hdrlen
);
4613 tcp_try_coalesce(sk
, tail
,
4614 skb
, fragstolen
)) ? 1 : 0;
4615 tcp_rcv_nxt_update(tcp_sk(sk
), TCP_SKB_CB(skb
)->end_seq
);
4617 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4618 skb_set_owner_r(skb
, sk
);
4623 int tcp_send_rcvq(struct sock
*sk
, struct msghdr
*msg
, size_t size
)
4625 struct sk_buff
*skb
;
4633 if (size
> PAGE_SIZE
) {
4634 int npages
= min_t(size_t, size
>> PAGE_SHIFT
, MAX_SKB_FRAGS
);
4636 data_len
= npages
<< PAGE_SHIFT
;
4637 size
= data_len
+ (size
& ~PAGE_MASK
);
4639 skb
= alloc_skb_with_frags(size
- data_len
, data_len
,
4640 PAGE_ALLOC_COSTLY_ORDER
,
4641 &err
, sk
->sk_allocation
);
4645 skb_put(skb
, size
- data_len
);
4646 skb
->data_len
= data_len
;
4649 if (tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
)) {
4650 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPRCVQDROP
);
4654 err
= skb_copy_datagram_from_iter(skb
, 0, &msg
->msg_iter
, size
);
4658 TCP_SKB_CB(skb
)->seq
= tcp_sk(sk
)->rcv_nxt
;
4659 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ size
;
4660 TCP_SKB_CB(skb
)->ack_seq
= tcp_sk(sk
)->snd_una
- 1;
4662 if (tcp_queue_rcv(sk
, skb
, 0, &fragstolen
)) {
4663 WARN_ON_ONCE(fragstolen
); /* should not happen */
4675 void tcp_data_ready(struct sock
*sk
)
4677 const struct tcp_sock
*tp
= tcp_sk(sk
);
4678 int avail
= tp
->rcv_nxt
- tp
->copied_seq
;
4680 if (avail
< sk
->sk_rcvlowat
&& !sock_flag(sk
, SOCK_DONE
))
4683 sk
->sk_data_ready(sk
);
4686 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
4688 struct tcp_sock
*tp
= tcp_sk(sk
);
4692 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
) {
4697 __skb_pull(skb
, tcp_hdr(skb
)->doff
* 4);
4699 tcp_ecn_accept_cwr(sk
, skb
);
4701 tp
->rx_opt
.dsack
= 0;
4703 /* Queue data for delivery to the user.
4704 * Packets in sequence go to the receive queue.
4705 * Out of sequence packets to the out_of_order_queue.
4707 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
4708 if (tcp_receive_window(tp
) == 0) {
4709 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPZEROWINDOWDROP
);
4713 /* Ok. In sequence. In window. */
4715 if (skb_queue_len(&sk
->sk_receive_queue
) == 0)
4716 sk_forced_mem_schedule(sk
, skb
->truesize
);
4717 else if (tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
)) {
4718 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPRCVQDROP
);
4722 eaten
= tcp_queue_rcv(sk
, skb
, 0, &fragstolen
);
4724 tcp_event_data_recv(sk
, skb
);
4725 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
4728 if (!RB_EMPTY_ROOT(&tp
->out_of_order_queue
)) {
4731 /* RFC5681. 4.2. SHOULD send immediate ACK, when
4732 * gap in queue is filled.
4734 if (RB_EMPTY_ROOT(&tp
->out_of_order_queue
))
4735 inet_csk(sk
)->icsk_ack
.pending
|= ICSK_ACK_NOW
;
4738 if (tp
->rx_opt
.num_sacks
)
4739 tcp_sack_remove(tp
);
4741 tcp_fast_path_check(sk
);
4744 kfree_skb_partial(skb
, fragstolen
);
4745 if (!sock_flag(sk
, SOCK_DEAD
))
4750 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4751 tcp_rcv_spurious_retrans(sk
, skb
);
4752 /* A retransmit, 2nd most common case. Force an immediate ack. */
4753 NET_INC_STATS(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4754 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4757 tcp_enter_quickack_mode(sk
, TCP_MAX_QUICKACKS
);
4758 inet_csk_schedule_ack(sk
);
4764 /* Out of window. F.e. zero window probe. */
4765 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
4768 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4769 /* Partial packet, seq < rcv_next < end_seq */
4770 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
4771 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4772 TCP_SKB_CB(skb
)->end_seq
);
4774 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
4776 /* If window is closed, drop tail of packet. But after
4777 * remembering D-SACK for its head made in previous line.
4779 if (!tcp_receive_window(tp
)) {
4780 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPZEROWINDOWDROP
);
4786 tcp_data_queue_ofo(sk
, skb
);
4789 static struct sk_buff
*tcp_skb_next(struct sk_buff
*skb
, struct sk_buff_head
*list
)
4792 return !skb_queue_is_last(list
, skb
) ? skb
->next
: NULL
;
4794 return skb_rb_next(skb
);
4797 static struct sk_buff
*tcp_collapse_one(struct sock
*sk
, struct sk_buff
*skb
,
4798 struct sk_buff_head
*list
,
4799 struct rb_root
*root
)
4801 struct sk_buff
*next
= tcp_skb_next(skb
, list
);
4804 __skb_unlink(skb
, list
);
4806 rb_erase(&skb
->rbnode
, root
);
4809 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPRCVCOLLAPSED
);
4814 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
4815 void tcp_rbtree_insert(struct rb_root
*root
, struct sk_buff
*skb
)
4817 struct rb_node
**p
= &root
->rb_node
;
4818 struct rb_node
*parent
= NULL
;
4819 struct sk_buff
*skb1
;
4823 skb1
= rb_to_skb(parent
);
4824 if (before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb1
)->seq
))
4825 p
= &parent
->rb_left
;
4827 p
= &parent
->rb_right
;
4829 rb_link_node(&skb
->rbnode
, parent
, p
);
4830 rb_insert_color(&skb
->rbnode
, root
);
4833 /* Collapse contiguous sequence of skbs head..tail with
4834 * sequence numbers start..end.
4836 * If tail is NULL, this means until the end of the queue.
4838 * Segments with FIN/SYN are not collapsed (only because this
4842 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
, struct rb_root
*root
,
4843 struct sk_buff
*head
, struct sk_buff
*tail
, u32 start
, u32 end
)
4845 struct sk_buff
*skb
= head
, *n
;
4846 struct sk_buff_head tmp
;
4849 /* First, check that queue is collapsible and find
4850 * the point where collapsing can be useful.
4853 for (end_of_skbs
= true; skb
!= NULL
&& skb
!= tail
; skb
= n
) {
4854 n
= tcp_skb_next(skb
, list
);
4856 /* No new bits? It is possible on ofo queue. */
4857 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4858 skb
= tcp_collapse_one(sk
, skb
, list
, root
);
4864 /* The first skb to collapse is:
4866 * - bloated or contains data before "start" or
4867 * overlaps to the next one.
4869 if (!(TCP_SKB_CB(skb
)->tcp_flags
& (TCPHDR_SYN
| TCPHDR_FIN
)) &&
4870 (tcp_win_from_space(sk
, skb
->truesize
) > skb
->len
||
4871 before(TCP_SKB_CB(skb
)->seq
, start
))) {
4872 end_of_skbs
= false;
4876 if (n
&& n
!= tail
&&
4877 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(n
)->seq
) {
4878 end_of_skbs
= false;
4882 /* Decided to skip this, advance start seq. */
4883 start
= TCP_SKB_CB(skb
)->end_seq
;
4886 (TCP_SKB_CB(skb
)->tcp_flags
& (TCPHDR_SYN
| TCPHDR_FIN
)))
4889 __skb_queue_head_init(&tmp
);
4891 while (before(start
, end
)) {
4892 int copy
= min_t(int, SKB_MAX_ORDER(0, 0), end
- start
);
4893 struct sk_buff
*nskb
;
4895 nskb
= alloc_skb(copy
, GFP_ATOMIC
);
4899 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
4900 #ifdef CONFIG_TLS_DEVICE
4901 nskb
->decrypted
= skb
->decrypted
;
4903 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
4905 __skb_queue_before(list
, skb
, nskb
);
4907 __skb_queue_tail(&tmp
, nskb
); /* defer rbtree insertion */
4908 skb_set_owner_r(nskb
, sk
);
4910 /* Copy data, releasing collapsed skbs. */
4912 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
4913 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
4917 size
= min(copy
, size
);
4918 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
4920 TCP_SKB_CB(nskb
)->end_seq
+= size
;
4924 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4925 skb
= tcp_collapse_one(sk
, skb
, list
, root
);
4928 (TCP_SKB_CB(skb
)->tcp_flags
& (TCPHDR_SYN
| TCPHDR_FIN
)))
4930 #ifdef CONFIG_TLS_DEVICE
4931 if (skb
->decrypted
!= nskb
->decrypted
)
4938 skb_queue_walk_safe(&tmp
, skb
, n
)
4939 tcp_rbtree_insert(root
, skb
);
4942 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4943 * and tcp_collapse() them until all the queue is collapsed.
4945 static void tcp_collapse_ofo_queue(struct sock
*sk
)
4947 struct tcp_sock
*tp
= tcp_sk(sk
);
4948 u32 range_truesize
, sum_tiny
= 0;
4949 struct sk_buff
*skb
, *head
;
4952 skb
= skb_rb_first(&tp
->out_of_order_queue
);
4955 tp
->ooo_last_skb
= skb_rb_last(&tp
->out_of_order_queue
);
4958 start
= TCP_SKB_CB(skb
)->seq
;
4959 end
= TCP_SKB_CB(skb
)->end_seq
;
4960 range_truesize
= skb
->truesize
;
4962 for (head
= skb
;;) {
4963 skb
= skb_rb_next(skb
);
4965 /* Range is terminated when we see a gap or when
4966 * we are at the queue end.
4969 after(TCP_SKB_CB(skb
)->seq
, end
) ||
4970 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
4971 /* Do not attempt collapsing tiny skbs */
4972 if (range_truesize
!= head
->truesize
||
4973 end
- start
>= SKB_WITH_OVERHEAD(SK_MEM_QUANTUM
)) {
4974 tcp_collapse(sk
, NULL
, &tp
->out_of_order_queue
,
4975 head
, skb
, start
, end
);
4977 sum_tiny
+= range_truesize
;
4978 if (sum_tiny
> sk
->sk_rcvbuf
>> 3)
4984 range_truesize
+= skb
->truesize
;
4985 if (unlikely(before(TCP_SKB_CB(skb
)->seq
, start
)))
4986 start
= TCP_SKB_CB(skb
)->seq
;
4987 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
4988 end
= TCP_SKB_CB(skb
)->end_seq
;
4993 * Clean the out-of-order queue to make room.
4994 * We drop high sequences packets to :
4995 * 1) Let a chance for holes to be filled.
4996 * 2) not add too big latencies if thousands of packets sit there.
4997 * (But if application shrinks SO_RCVBUF, we could still end up
4998 * freeing whole queue here)
4999 * 3) Drop at least 12.5 % of sk_rcvbuf to avoid malicious attacks.
5001 * Return true if queue has shrunk.
5003 static bool tcp_prune_ofo_queue(struct sock
*sk
)
5005 struct tcp_sock
*tp
= tcp_sk(sk
);
5006 struct rb_node
*node
, *prev
;
5009 if (RB_EMPTY_ROOT(&tp
->out_of_order_queue
))
5012 NET_INC_STATS(sock_net(sk
), LINUX_MIB_OFOPRUNED
);
5013 goal
= sk
->sk_rcvbuf
>> 3;
5014 node
= &tp
->ooo_last_skb
->rbnode
;
5016 prev
= rb_prev(node
);
5017 rb_erase(node
, &tp
->out_of_order_queue
);
5018 goal
-= rb_to_skb(node
)->truesize
;
5019 tcp_drop(sk
, rb_to_skb(node
));
5020 if (!prev
|| goal
<= 0) {
5022 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
&&
5023 !tcp_under_memory_pressure(sk
))
5025 goal
= sk
->sk_rcvbuf
>> 3;
5029 tp
->ooo_last_skb
= rb_to_skb(prev
);
5031 /* Reset SACK state. A conforming SACK implementation will
5032 * do the same at a timeout based retransmit. When a connection
5033 * is in a sad state like this, we care only about integrity
5034 * of the connection not performance.
5036 if (tp
->rx_opt
.sack_ok
)
5037 tcp_sack_reset(&tp
->rx_opt
);
5041 /* Reduce allocated memory if we can, trying to get
5042 * the socket within its memory limits again.
5044 * Return less than zero if we should start dropping frames
5045 * until the socket owning process reads some of the data
5046 * to stabilize the situation.
5048 static int tcp_prune_queue(struct sock
*sk
)
5050 struct tcp_sock
*tp
= tcp_sk(sk
);
5052 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
5054 NET_INC_STATS(sock_net(sk
), LINUX_MIB_PRUNECALLED
);
5056 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
5057 tcp_clamp_window(sk
);
5058 else if (tcp_under_memory_pressure(sk
))
5059 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
5061 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
5064 tcp_collapse_ofo_queue(sk
);
5065 if (!skb_queue_empty(&sk
->sk_receive_queue
))
5066 tcp_collapse(sk
, &sk
->sk_receive_queue
, NULL
,
5067 skb_peek(&sk
->sk_receive_queue
),
5069 tp
->copied_seq
, tp
->rcv_nxt
);
5072 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
5075 /* Collapsing did not help, destructive actions follow.
5076 * This must not ever occur. */
5078 tcp_prune_ofo_queue(sk
);
5080 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
5083 /* If we are really being abused, tell the caller to silently
5084 * drop receive data on the floor. It will get retransmitted
5085 * and hopefully then we'll have sufficient space.
5087 NET_INC_STATS(sock_net(sk
), LINUX_MIB_RCVPRUNED
);
5089 /* Massive buffer overcommit. */
5094 static bool tcp_should_expand_sndbuf(const struct sock
*sk
)
5096 const struct tcp_sock
*tp
= tcp_sk(sk
);
5098 /* If the user specified a specific send buffer setting, do
5101 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
5104 /* If we are under global TCP memory pressure, do not expand. */
5105 if (tcp_under_memory_pressure(sk
))
5108 /* If we are under soft global TCP memory pressure, do not expand. */
5109 if (sk_memory_allocated(sk
) >= sk_prot_mem_limits(sk
, 0))
5112 /* If we filled the congestion window, do not expand. */
5113 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
5119 /* When incoming ACK allowed to free some skb from write_queue,
5120 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
5121 * on the exit from tcp input handler.
5123 * PROBLEM: sndbuf expansion does not work well with largesend.
5125 static void tcp_new_space(struct sock
*sk
)
5127 struct tcp_sock
*tp
= tcp_sk(sk
);
5129 if (tcp_should_expand_sndbuf(sk
)) {
5130 tcp_sndbuf_expand(sk
);
5131 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
5134 sk
->sk_write_space(sk
);
5137 static void tcp_check_space(struct sock
*sk
)
5139 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
5140 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
5141 /* pairs with tcp_poll() */
5143 if (sk
->sk_socket
&&
5144 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
5146 if (!test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
5147 tcp_chrono_stop(sk
, TCP_CHRONO_SNDBUF_LIMITED
);
5152 static inline void tcp_data_snd_check(struct sock
*sk
)
5154 tcp_push_pending_frames(sk
);
5155 tcp_check_space(sk
);
5159 * Check if sending an ack is needed.
5161 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
5163 struct tcp_sock
*tp
= tcp_sk(sk
);
5164 unsigned long rtt
, delay
;
5166 /* More than one full frame received... */
5167 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
&&
5168 /* ... and right edge of window advances far enough.
5169 * (tcp_recvmsg() will send ACK otherwise).
5170 * If application uses SO_RCVLOWAT, we want send ack now if
5171 * we have not received enough bytes to satisfy the condition.
5173 (tp
->rcv_nxt
- tp
->copied_seq
< sk
->sk_rcvlowat
||
5174 __tcp_select_window(sk
) >= tp
->rcv_wnd
)) ||
5175 /* We ACK each frame or... */
5176 tcp_in_quickack_mode(sk
) ||
5177 /* Protocol state mandates a one-time immediate ACK */
5178 inet_csk(sk
)->icsk_ack
.pending
& ICSK_ACK_NOW
) {
5184 if (!ofo_possible
|| RB_EMPTY_ROOT(&tp
->out_of_order_queue
)) {
5185 tcp_send_delayed_ack(sk
);
5189 if (!tcp_is_sack(tp
) ||
5190 tp
->compressed_ack
>= sock_net(sk
)->ipv4
.sysctl_tcp_comp_sack_nr
)
5193 if (tp
->compressed_ack_rcv_nxt
!= tp
->rcv_nxt
) {
5194 tp
->compressed_ack_rcv_nxt
= tp
->rcv_nxt
;
5195 if (tp
->compressed_ack
> TCP_FASTRETRANS_THRESH
)
5196 NET_ADD_STATS(sock_net(sk
), LINUX_MIB_TCPACKCOMPRESSED
,
5197 tp
->compressed_ack
- TCP_FASTRETRANS_THRESH
);
5198 tp
->compressed_ack
= 0;
5201 if (++tp
->compressed_ack
<= TCP_FASTRETRANS_THRESH
)
5204 if (hrtimer_is_queued(&tp
->compressed_ack_timer
))
5207 /* compress ack timer : 5 % of rtt, but no more than tcp_comp_sack_delay_ns */
5209 rtt
= tp
->rcv_rtt_est
.rtt_us
;
5210 if (tp
->srtt_us
&& tp
->srtt_us
< rtt
)
5213 delay
= min_t(unsigned long, sock_net(sk
)->ipv4
.sysctl_tcp_comp_sack_delay_ns
,
5214 rtt
* (NSEC_PER_USEC
>> 3)/20);
5216 hrtimer_start(&tp
->compressed_ack_timer
, ns_to_ktime(delay
),
5217 HRTIMER_MODE_REL_PINNED_SOFT
);
5220 static inline void tcp_ack_snd_check(struct sock
*sk
)
5222 if (!inet_csk_ack_scheduled(sk
)) {
5223 /* We sent a data segment already. */
5226 __tcp_ack_snd_check(sk
, 1);
5230 * This routine is only called when we have urgent data
5231 * signaled. Its the 'slow' part of tcp_urg. It could be
5232 * moved inline now as tcp_urg is only called from one
5233 * place. We handle URGent data wrong. We have to - as
5234 * BSD still doesn't use the correction from RFC961.
5235 * For 1003.1g we should support a new option TCP_STDURG to permit
5236 * either form (or just set the sysctl tcp_stdurg).
5239 static void tcp_check_urg(struct sock
*sk
, const struct tcphdr
*th
)
5241 struct tcp_sock
*tp
= tcp_sk(sk
);
5242 u32 ptr
= ntohs(th
->urg_ptr
);
5244 if (ptr
&& !sock_net(sk
)->ipv4
.sysctl_tcp_stdurg
)
5246 ptr
+= ntohl(th
->seq
);
5248 /* Ignore urgent data that we've already seen and read. */
5249 if (after(tp
->copied_seq
, ptr
))
5252 /* Do not replay urg ptr.
5254 * NOTE: interesting situation not covered by specs.
5255 * Misbehaving sender may send urg ptr, pointing to segment,
5256 * which we already have in ofo queue. We are not able to fetch
5257 * such data and will stay in TCP_URG_NOTYET until will be eaten
5258 * by recvmsg(). Seems, we are not obliged to handle such wicked
5259 * situations. But it is worth to think about possibility of some
5260 * DoSes using some hypothetical application level deadlock.
5262 if (before(ptr
, tp
->rcv_nxt
))
5265 /* Do we already have a newer (or duplicate) urgent pointer? */
5266 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
5269 /* Tell the world about our new urgent pointer. */
5272 /* We may be adding urgent data when the last byte read was
5273 * urgent. To do this requires some care. We cannot just ignore
5274 * tp->copied_seq since we would read the last urgent byte again
5275 * as data, nor can we alter copied_seq until this data arrives
5276 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5278 * NOTE. Double Dutch. Rendering to plain English: author of comment
5279 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5280 * and expect that both A and B disappear from stream. This is _wrong_.
5281 * Though this happens in BSD with high probability, this is occasional.
5282 * Any application relying on this is buggy. Note also, that fix "works"
5283 * only in this artificial test. Insert some normal data between A and B and we will
5284 * decline of BSD again. Verdict: it is better to remove to trap
5287 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
5288 !sock_flag(sk
, SOCK_URGINLINE
) && tp
->copied_seq
!= tp
->rcv_nxt
) {
5289 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
5291 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5292 __skb_unlink(skb
, &sk
->sk_receive_queue
);
5297 tp
->urg_data
= TCP_URG_NOTYET
;
5300 /* Disable header prediction. */
5304 /* This is the 'fast' part of urgent handling. */
5305 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, const struct tcphdr
*th
)
5307 struct tcp_sock
*tp
= tcp_sk(sk
);
5309 /* Check if we get a new urgent pointer - normally not. */
5311 tcp_check_urg(sk
, th
);
5313 /* Do we wait for any urgent data? - normally not... */
5314 if (tp
->urg_data
== TCP_URG_NOTYET
) {
5315 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
5318 /* Is the urgent pointer pointing into this packet? */
5319 if (ptr
< skb
->len
) {
5321 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
5323 tp
->urg_data
= TCP_URG_VALID
| tmp
;
5324 if (!sock_flag(sk
, SOCK_DEAD
))
5325 sk
->sk_data_ready(sk
);
5330 /* Accept RST for rcv_nxt - 1 after a FIN.
5331 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a
5332 * FIN is sent followed by a RST packet. The RST is sent with the same
5333 * sequence number as the FIN, and thus according to RFC 5961 a challenge
5334 * ACK should be sent. However, Mac OSX rate limits replies to challenge
5335 * ACKs on the closed socket. In addition middleboxes can drop either the
5336 * challenge ACK or a subsequent RST.
5338 static bool tcp_reset_check(const struct sock
*sk
, const struct sk_buff
*skb
)
5340 struct tcp_sock
*tp
= tcp_sk(sk
);
5342 return unlikely(TCP_SKB_CB(skb
)->seq
== (tp
->rcv_nxt
- 1) &&
5343 (1 << sk
->sk_state
) & (TCPF_CLOSE_WAIT
| TCPF_LAST_ACK
|
5347 /* Does PAWS and seqno based validation of an incoming segment, flags will
5348 * play significant role here.
5350 static bool tcp_validate_incoming(struct sock
*sk
, struct sk_buff
*skb
,
5351 const struct tcphdr
*th
, int syn_inerr
)
5353 struct tcp_sock
*tp
= tcp_sk(sk
);
5354 bool rst_seq_match
= false;
5356 /* RFC1323: H1. Apply PAWS check first. */
5357 if (tcp_fast_parse_options(sock_net(sk
), skb
, th
, tp
) &&
5358 tp
->rx_opt
.saw_tstamp
&&
5359 tcp_paws_discard(sk
, skb
)) {
5361 NET_INC_STATS(sock_net(sk
), LINUX_MIB_PAWSESTABREJECTED
);
5362 if (!tcp_oow_rate_limited(sock_net(sk
), skb
,
5363 LINUX_MIB_TCPACKSKIPPEDPAWS
,
5364 &tp
->last_oow_ack_time
))
5365 tcp_send_dupack(sk
, skb
);
5368 /* Reset is accepted even if it did not pass PAWS. */
5371 /* Step 1: check sequence number */
5372 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5373 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5374 * (RST) segments are validated by checking their SEQ-fields."
5375 * And page 69: "If an incoming segment is not acceptable,
5376 * an acknowledgment should be sent in reply (unless the RST
5377 * bit is set, if so drop the segment and return)".
5382 if (!tcp_oow_rate_limited(sock_net(sk
), skb
,
5383 LINUX_MIB_TCPACKSKIPPEDSEQ
,
5384 &tp
->last_oow_ack_time
))
5385 tcp_send_dupack(sk
, skb
);
5386 } else if (tcp_reset_check(sk
, skb
)) {
5392 /* Step 2: check RST bit */
5394 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a
5395 * FIN and SACK too if available):
5396 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or
5397 * the right-most SACK block,
5399 * RESET the connection
5401 * Send a challenge ACK
5403 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
||
5404 tcp_reset_check(sk
, skb
)) {
5405 rst_seq_match
= true;
5406 } else if (tcp_is_sack(tp
) && tp
->rx_opt
.num_sacks
> 0) {
5407 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
5408 int max_sack
= sp
[0].end_seq
;
5411 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
;
5413 max_sack
= after(sp
[this_sack
].end_seq
,
5415 sp
[this_sack
].end_seq
: max_sack
;
5418 if (TCP_SKB_CB(skb
)->seq
== max_sack
)
5419 rst_seq_match
= true;
5425 /* Disable TFO if RST is out-of-order
5426 * and no data has been received
5427 * for current active TFO socket
5429 if (tp
->syn_fastopen
&& !tp
->data_segs_in
&&
5430 sk
->sk_state
== TCP_ESTABLISHED
)
5431 tcp_fastopen_active_disable(sk
);
5432 tcp_send_challenge_ack(sk
, skb
);
5437 /* step 3: check security and precedence [ignored] */
5439 /* step 4: Check for a SYN
5440 * RFC 5961 4.2 : Send a challenge ack
5445 TCP_INC_STATS(sock_net(sk
), TCP_MIB_INERRS
);
5446 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPSYNCHALLENGE
);
5447 tcp_send_challenge_ack(sk
, skb
);
5459 * TCP receive function for the ESTABLISHED state.
5461 * It is split into a fast path and a slow path. The fast path is
5463 * - A zero window was announced from us - zero window probing
5464 * is only handled properly in the slow path.
5465 * - Out of order segments arrived.
5466 * - Urgent data is expected.
5467 * - There is no buffer space left
5468 * - Unexpected TCP flags/window values/header lengths are received
5469 * (detected by checking the TCP header against pred_flags)
5470 * - Data is sent in both directions. Fast path only supports pure senders
5471 * or pure receivers (this means either the sequence number or the ack
5472 * value must stay constant)
5473 * - Unexpected TCP option.
5475 * When these conditions are not satisfied it drops into a standard
5476 * receive procedure patterned after RFC793 to handle all cases.
5477 * The first three cases are guaranteed by proper pred_flags setting,
5478 * the rest is checked inline. Fast processing is turned on in
5479 * tcp_data_queue when everything is OK.
5481 void tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
)
5483 const struct tcphdr
*th
= (const struct tcphdr
*)skb
->data
;
5484 struct tcp_sock
*tp
= tcp_sk(sk
);
5485 unsigned int len
= skb
->len
;
5487 /* TCP congestion window tracking */
5488 trace_tcp_probe(sk
, skb
);
5490 tcp_mstamp_refresh(tp
);
5491 if (unlikely(!sk
->sk_rx_dst
))
5492 inet_csk(sk
)->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
5494 * Header prediction.
5495 * The code loosely follows the one in the famous
5496 * "30 instruction TCP receive" Van Jacobson mail.
5498 * Van's trick is to deposit buffers into socket queue
5499 * on a device interrupt, to call tcp_recv function
5500 * on the receive process context and checksum and copy
5501 * the buffer to user space. smart...
5503 * Our current scheme is not silly either but we take the
5504 * extra cost of the net_bh soft interrupt processing...
5505 * We do checksum and copy also but from device to kernel.
5508 tp
->rx_opt
.saw_tstamp
= 0;
5510 /* pred_flags is 0xS?10 << 16 + snd_wnd
5511 * if header_prediction is to be made
5512 * 'S' will always be tp->tcp_header_len >> 2
5513 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5514 * turn it off (when there are holes in the receive
5515 * space for instance)
5516 * PSH flag is ignored.
5519 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
5520 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
&&
5521 !after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
)) {
5522 int tcp_header_len
= tp
->tcp_header_len
;
5524 /* Timestamp header prediction: tcp_header_len
5525 * is automatically equal to th->doff*4 due to pred_flags
5529 /* Check timestamp */
5530 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
5531 /* No? Slow path! */
5532 if (!tcp_parse_aligned_timestamp(tp
, th
))
5535 /* If PAWS failed, check it more carefully in slow path */
5536 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
5539 /* DO NOT update ts_recent here, if checksum fails
5540 * and timestamp was corrupted part, it will result
5541 * in a hung connection since we will drop all
5542 * future packets due to the PAWS test.
5546 if (len
<= tcp_header_len
) {
5547 /* Bulk data transfer: sender */
5548 if (len
== tcp_header_len
) {
5549 /* Predicted packet is in window by definition.
5550 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5551 * Hence, check seq<=rcv_wup reduces to:
5553 if (tcp_header_len
==
5554 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5555 tp
->rcv_nxt
== tp
->rcv_wup
)
5556 tcp_store_ts_recent(tp
);
5558 /* We know that such packets are checksummed
5561 tcp_ack(sk
, skb
, 0);
5563 tcp_data_snd_check(sk
);
5564 /* When receiving pure ack in fast path, update
5565 * last ts ecr directly instead of calling
5566 * tcp_rcv_rtt_measure_ts()
5568 tp
->rcv_rtt_last_tsecr
= tp
->rx_opt
.rcv_tsecr
;
5570 } else { /* Header too small */
5571 TCP_INC_STATS(sock_net(sk
), TCP_MIB_INERRS
);
5576 bool fragstolen
= false;
5578 if (tcp_checksum_complete(skb
))
5581 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
5584 /* Predicted packet is in window by definition.
5585 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5586 * Hence, check seq<=rcv_wup reduces to:
5588 if (tcp_header_len
==
5589 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5590 tp
->rcv_nxt
== tp
->rcv_wup
)
5591 tcp_store_ts_recent(tp
);
5593 tcp_rcv_rtt_measure_ts(sk
, skb
);
5595 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPHPHITS
);
5597 /* Bulk data transfer: receiver */
5598 eaten
= tcp_queue_rcv(sk
, skb
, tcp_header_len
,
5601 tcp_event_data_recv(sk
, skb
);
5603 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
5604 /* Well, only one small jumplet in fast path... */
5605 tcp_ack(sk
, skb
, FLAG_DATA
);
5606 tcp_data_snd_check(sk
);
5607 if (!inet_csk_ack_scheduled(sk
))
5611 __tcp_ack_snd_check(sk
, 0);
5614 kfree_skb_partial(skb
, fragstolen
);
5621 if (len
< (th
->doff
<< 2) || tcp_checksum_complete(skb
))
5624 if (!th
->ack
&& !th
->rst
&& !th
->syn
)
5628 * Standard slow path.
5631 if (!tcp_validate_incoming(sk
, skb
, th
, 1))
5635 if (tcp_ack(sk
, skb
, FLAG_SLOWPATH
| FLAG_UPDATE_TS_RECENT
) < 0)
5638 tcp_rcv_rtt_measure_ts(sk
, skb
);
5640 /* Process urgent data. */
5641 tcp_urg(sk
, skb
, th
);
5643 /* step 7: process the segment text */
5644 tcp_data_queue(sk
, skb
);
5646 tcp_data_snd_check(sk
);
5647 tcp_ack_snd_check(sk
);
5651 TCP_INC_STATS(sock_net(sk
), TCP_MIB_CSUMERRORS
);
5652 TCP_INC_STATS(sock_net(sk
), TCP_MIB_INERRS
);
5657 EXPORT_SYMBOL(tcp_rcv_established
);
5659 void tcp_finish_connect(struct sock
*sk
, struct sk_buff
*skb
)
5661 struct tcp_sock
*tp
= tcp_sk(sk
);
5662 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5664 tcp_set_state(sk
, TCP_ESTABLISHED
);
5665 icsk
->icsk_ack
.lrcvtime
= tcp_jiffies32
;
5668 icsk
->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
5669 security_inet_conn_established(sk
, skb
);
5670 sk_mark_napi_id(sk
, skb
);
5673 tcp_init_transfer(sk
, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB
);
5675 /* Prevent spurious tcp_cwnd_restart() on first data
5678 tp
->lsndtime
= tcp_jiffies32
;
5680 if (sock_flag(sk
, SOCK_KEEPOPEN
))
5681 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
5683 if (!tp
->rx_opt
.snd_wscale
)
5684 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
5689 static bool tcp_rcv_fastopen_synack(struct sock
*sk
, struct sk_buff
*synack
,
5690 struct tcp_fastopen_cookie
*cookie
)
5692 struct tcp_sock
*tp
= tcp_sk(sk
);
5693 struct sk_buff
*data
= tp
->syn_data
? tcp_rtx_queue_head(sk
) : NULL
;
5694 u16 mss
= tp
->rx_opt
.mss_clamp
, try_exp
= 0;
5695 bool syn_drop
= false;
5697 if (mss
== tp
->rx_opt
.user_mss
) {
5698 struct tcp_options_received opt
;
5700 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5701 tcp_clear_options(&opt
);
5702 opt
.user_mss
= opt
.mss_clamp
= 0;
5703 tcp_parse_options(sock_net(sk
), synack
, &opt
, 0, NULL
);
5704 mss
= opt
.mss_clamp
;
5707 if (!tp
->syn_fastopen
) {
5708 /* Ignore an unsolicited cookie */
5710 } else if (tp
->total_retrans
) {
5711 /* SYN timed out and the SYN-ACK neither has a cookie nor
5712 * acknowledges data. Presumably the remote received only
5713 * the retransmitted (regular) SYNs: either the original
5714 * SYN-data or the corresponding SYN-ACK was dropped.
5716 syn_drop
= (cookie
->len
< 0 && data
);
5717 } else if (cookie
->len
< 0 && !tp
->syn_data
) {
5718 /* We requested a cookie but didn't get it. If we did not use
5719 * the (old) exp opt format then try so next time (try_exp=1).
5720 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5722 try_exp
= tp
->syn_fastopen_exp
? 2 : 1;
5725 tcp_fastopen_cache_set(sk
, mss
, cookie
, syn_drop
, try_exp
);
5727 if (data
) { /* Retransmit unacked data in SYN */
5728 skb_rbtree_walk_from(data
) {
5729 if (__tcp_retransmit_skb(sk
, data
, 1))
5733 NET_INC_STATS(sock_net(sk
),
5734 LINUX_MIB_TCPFASTOPENACTIVEFAIL
);
5737 tp
->syn_data_acked
= tp
->syn_data
;
5738 if (tp
->syn_data_acked
) {
5739 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPFASTOPENACTIVE
);
5740 /* SYN-data is counted as two separate packets in tcp_ack() */
5741 if (tp
->delivered
> 1)
5745 tcp_fastopen_add_skb(sk
, synack
);
5750 static void smc_check_reset_syn(struct tcp_sock
*tp
)
5752 #if IS_ENABLED(CONFIG_SMC)
5753 if (static_branch_unlikely(&tcp_have_smc
)) {
5754 if (tp
->syn_smc
&& !tp
->rx_opt
.smc_ok
)
5760 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5761 const struct tcphdr
*th
)
5763 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5764 struct tcp_sock
*tp
= tcp_sk(sk
);
5765 struct tcp_fastopen_cookie foc
= { .len
= -1 };
5766 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
5769 tcp_parse_options(sock_net(sk
), skb
, &tp
->rx_opt
, 0, &foc
);
5770 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
5771 tp
->rx_opt
.rcv_tsecr
-= tp
->tsoffset
;
5775 * "If the state is SYN-SENT then
5776 * first check the ACK bit
5777 * If the ACK bit is set
5778 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5779 * a reset (unless the RST bit is set, if so drop
5780 * the segment and return)"
5782 if (!after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_una
) ||
5783 after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
))
5784 goto reset_and_undo
;
5786 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
5787 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
5788 tcp_time_stamp(tp
))) {
5789 NET_INC_STATS(sock_net(sk
),
5790 LINUX_MIB_PAWSACTIVEREJECTED
);
5791 goto reset_and_undo
;
5794 /* Now ACK is acceptable.
5796 * "If the RST bit is set
5797 * If the ACK was acceptable then signal the user "error:
5798 * connection reset", drop the segment, enter CLOSED state,
5799 * delete TCB, and return."
5808 * "fifth, if neither of the SYN or RST bits is set then
5809 * drop the segment and return."
5815 goto discard_and_undo
;
5818 * "If the SYN bit is on ...
5819 * are acceptable then ...
5820 * (our SYN has been ACKed), change the connection
5821 * state to ESTABLISHED..."
5824 tcp_ecn_rcv_synack(tp
, th
);
5826 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5827 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
5829 /* Ok.. it's good. Set up sequence numbers and
5830 * move to established.
5832 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
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
);
5840 if (!tp
->rx_opt
.wscale_ok
) {
5841 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
5842 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
5845 if (tp
->rx_opt
.saw_tstamp
) {
5846 tp
->rx_opt
.tstamp_ok
= 1;
5847 tp
->tcp_header_len
=
5848 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5849 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5850 tcp_store_ts_recent(tp
);
5852 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5855 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5856 tcp_initialize_rcv_mss(sk
);
5858 /* Remember, tcp_poll() does not lock socket!
5859 * Change state from SYN-SENT only after copied_seq
5860 * is initialized. */
5861 tp
->copied_seq
= tp
->rcv_nxt
;
5863 smc_check_reset_syn(tp
);
5867 tcp_finish_connect(sk
, skb
);
5869 fastopen_fail
= (tp
->syn_fastopen
|| tp
->syn_data
) &&
5870 tcp_rcv_fastopen_synack(sk
, skb
, &foc
);
5872 if (!sock_flag(sk
, SOCK_DEAD
)) {
5873 sk
->sk_state_change(sk
);
5874 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5878 if (sk
->sk_write_pending
||
5879 icsk
->icsk_accept_queue
.rskq_defer_accept
||
5880 icsk
->icsk_ack
.pingpong
) {
5881 /* Save one ACK. Data will be ready after
5882 * several ticks, if write_pending is set.
5884 * It may be deleted, but with this feature tcpdumps
5885 * look so _wonderfully_ clever, that I was not able
5886 * to stand against the temptation 8) --ANK
5888 inet_csk_schedule_ack(sk
);
5889 tcp_enter_quickack_mode(sk
, TCP_MAX_QUICKACKS
);
5890 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
5891 TCP_DELACK_MAX
, TCP_RTO_MAX
);
5902 /* No ACK in the segment */
5906 * "If the RST bit is set
5908 * Otherwise (no ACK) drop the segment and return."
5911 goto discard_and_undo
;
5915 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&&
5916 tcp_paws_reject(&tp
->rx_opt
, 0))
5917 goto discard_and_undo
;
5920 /* We see SYN without ACK. It is attempt of
5921 * simultaneous connect with crossed SYNs.
5922 * Particularly, it can be connect to self.
5924 tcp_set_state(sk
, TCP_SYN_RECV
);
5926 if (tp
->rx_opt
.saw_tstamp
) {
5927 tp
->rx_opt
.tstamp_ok
= 1;
5928 tcp_store_ts_recent(tp
);
5929 tp
->tcp_header_len
=
5930 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5932 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5935 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5936 tp
->copied_seq
= tp
->rcv_nxt
;
5937 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5939 /* RFC1323: The window in SYN & SYN/ACK segments is
5942 tp
->snd_wnd
= ntohs(th
->window
);
5943 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
5944 tp
->max_window
= tp
->snd_wnd
;
5946 tcp_ecn_rcv_syn(tp
, th
);
5949 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5950 tcp_initialize_rcv_mss(sk
);
5952 tcp_send_synack(sk
);
5954 /* Note, we could accept data and URG from this segment.
5955 * There are no obstacles to make this (except that we must
5956 * either change tcp_recvmsg() to prevent it from returning data
5957 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5959 * However, if we ignore data in ACKless segments sometimes,
5960 * we have no reasons to accept it sometimes.
5961 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5962 * is not flawless. So, discard packet for sanity.
5963 * Uncomment this return to process the data.
5970 /* "fifth, if neither of the SYN or RST bits is set then
5971 * drop the segment and return."
5975 tcp_clear_options(&tp
->rx_opt
);
5976 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5980 tcp_clear_options(&tp
->rx_opt
);
5981 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5986 * This function implements the receiving procedure of RFC 793 for
5987 * all states except ESTABLISHED and TIME_WAIT.
5988 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5989 * address independent.
5992 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
)
5994 struct tcp_sock
*tp
= tcp_sk(sk
);
5995 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5996 const struct tcphdr
*th
= tcp_hdr(skb
);
5997 struct request_sock
*req
;
6001 switch (sk
->sk_state
) {
6015 /* It is possible that we process SYN packets from backlog,
6016 * so we need to make sure to disable BH and RCU right there.
6020 acceptable
= icsk
->icsk_af_ops
->conn_request(sk
, skb
) >= 0;
6032 tp
->rx_opt
.saw_tstamp
= 0;
6033 tcp_mstamp_refresh(tp
);
6034 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
);
6038 /* Do step6 onward by hand. */
6039 tcp_urg(sk
, skb
, th
);
6041 tcp_data_snd_check(sk
);
6045 tcp_mstamp_refresh(tp
);
6046 tp
->rx_opt
.saw_tstamp
= 0;
6047 req
= tp
->fastopen_rsk
;
6051 WARN_ON_ONCE(sk
->sk_state
!= TCP_SYN_RECV
&&
6052 sk
->sk_state
!= TCP_FIN_WAIT1
);
6054 if (!tcp_check_req(sk
, skb
, req
, true, &req_stolen
))
6058 if (!th
->ack
&& !th
->rst
&& !th
->syn
)
6061 if (!tcp_validate_incoming(sk
, skb
, th
, 0))
6064 /* step 5: check the ACK field */
6065 acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
|
6066 FLAG_UPDATE_TS_RECENT
|
6067 FLAG_NO_CHALLENGE_ACK
) > 0;
6070 if (sk
->sk_state
== TCP_SYN_RECV
)
6071 return 1; /* send one RST */
6072 tcp_send_challenge_ack(sk
, skb
);
6075 switch (sk
->sk_state
) {
6077 tp
->delivered
++; /* SYN-ACK delivery isn't tracked in tcp_ack */
6079 tcp_synack_rtt_meas(sk
, req
);
6081 /* Once we leave TCP_SYN_RECV, we no longer need req
6085 inet_csk(sk
)->icsk_retransmits
= 0;
6086 reqsk_fastopen_remove(sk
, req
, false);
6087 /* Re-arm the timer because data may have been sent out.
6088 * This is similar to the regular data transmission case
6089 * when new data has just been ack'ed.
6091 * (TFO) - we could try to be more aggressive and
6092 * retransmitting any data sooner based on when they
6097 tcp_init_transfer(sk
, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB
);
6098 tp
->copied_seq
= tp
->rcv_nxt
;
6101 tcp_set_state(sk
, TCP_ESTABLISHED
);
6102 sk
->sk_state_change(sk
);
6104 /* Note, that this wakeup is only for marginal crossed SYN case.
6105 * Passively open sockets are not waked up, because
6106 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
6109 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
6111 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
6112 tp
->snd_wnd
= ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
;
6113 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
6115 if (tp
->rx_opt
.tstamp_ok
)
6116 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
6118 if (!inet_csk(sk
)->icsk_ca_ops
->cong_control
)
6119 tcp_update_pacing_rate(sk
);
6121 /* Prevent spurious tcp_cwnd_restart() on first data packet */
6122 tp
->lsndtime
= tcp_jiffies32
;
6124 tcp_initialize_rcv_mss(sk
);
6125 tcp_fast_path_on(tp
);
6128 case TCP_FIN_WAIT1
: {
6131 /* If we enter the TCP_FIN_WAIT1 state and we are a
6132 * Fast Open socket and this is the first acceptable
6133 * ACK we have received, this would have acknowledged
6134 * our SYNACK so stop the SYNACK timer.
6137 /* We no longer need the request sock. */
6138 reqsk_fastopen_remove(sk
, req
, false);
6141 if (tp
->snd_una
!= tp
->write_seq
)
6144 tcp_set_state(sk
, TCP_FIN_WAIT2
);
6145 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
6149 if (!sock_flag(sk
, SOCK_DEAD
)) {
6150 /* Wake up lingering close() */
6151 sk
->sk_state_change(sk
);
6155 if (tp
->linger2
< 0) {
6157 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
6160 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
6161 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
6162 /* Receive out of order FIN after close() */
6163 if (tp
->syn_fastopen
&& th
->fin
)
6164 tcp_fastopen_active_disable(sk
);
6166 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
6170 tmo
= tcp_fin_time(sk
);
6171 if (tmo
> TCP_TIMEWAIT_LEN
) {
6172 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
6173 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
6174 /* Bad case. We could lose such FIN otherwise.
6175 * It is not a big problem, but it looks confusing
6176 * and not so rare event. We still can lose it now,
6177 * if it spins in bh_lock_sock(), but it is really
6180 inet_csk_reset_keepalive_timer(sk
, tmo
);
6182 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
6189 if (tp
->snd_una
== tp
->write_seq
) {
6190 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
6196 if (tp
->snd_una
== tp
->write_seq
) {
6197 tcp_update_metrics(sk
);
6204 /* step 6: check the URG bit */
6205 tcp_urg(sk
, skb
, th
);
6207 /* step 7: process the segment text */
6208 switch (sk
->sk_state
) {
6209 case TCP_CLOSE_WAIT
:
6212 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
6217 /* RFC 793 says to queue data in these states,
6218 * RFC 1122 says we MUST send a reset.
6219 * BSD 4.4 also does reset.
6221 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
6222 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
6223 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
6224 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
6230 case TCP_ESTABLISHED
:
6231 tcp_data_queue(sk
, skb
);
6236 /* tcp_data could move socket to TIME-WAIT */
6237 if (sk
->sk_state
!= TCP_CLOSE
) {
6238 tcp_data_snd_check(sk
);
6239 tcp_ack_snd_check(sk
);
6248 EXPORT_SYMBOL(tcp_rcv_state_process
);
6250 static inline void pr_drop_req(struct request_sock
*req
, __u16 port
, int family
)
6252 struct inet_request_sock
*ireq
= inet_rsk(req
);
6254 if (family
== AF_INET
)
6255 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6256 &ireq
->ir_rmt_addr
, port
);
6257 #if IS_ENABLED(CONFIG_IPV6)
6258 else if (family
== AF_INET6
)
6259 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6260 &ireq
->ir_v6_rmt_addr
, port
);
6264 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6266 * If we receive a SYN packet with these bits set, it means a
6267 * network is playing bad games with TOS bits. In order to
6268 * avoid possible false congestion notifications, we disable
6269 * TCP ECN negotiation.
6271 * Exception: tcp_ca wants ECN. This is required for DCTCP
6272 * congestion control: Linux DCTCP asserts ECT on all packets,
6273 * including SYN, which is most optimal solution; however,
6274 * others, such as FreeBSD do not.
6276 static void tcp_ecn_create_request(struct request_sock
*req
,
6277 const struct sk_buff
*skb
,
6278 const struct sock
*listen_sk
,
6279 const struct dst_entry
*dst
)
6281 const struct tcphdr
*th
= tcp_hdr(skb
);
6282 const struct net
*net
= sock_net(listen_sk
);
6283 bool th_ecn
= th
->ece
&& th
->cwr
;
6290 ect
= !INET_ECN_is_not_ect(TCP_SKB_CB(skb
)->ip_dsfield
);
6291 ecn_ok_dst
= dst_feature(dst
, DST_FEATURE_ECN_MASK
);
6292 ecn_ok
= net
->ipv4
.sysctl_tcp_ecn
|| ecn_ok_dst
;
6294 if ((!ect
&& ecn_ok
) || tcp_ca_needs_ecn(listen_sk
) ||
6295 (ecn_ok_dst
& DST_FEATURE_ECN_CA
) ||
6296 tcp_bpf_ca_needs_ecn((struct sock
*)req
))
6297 inet_rsk(req
)->ecn_ok
= 1;
6300 static void tcp_openreq_init(struct request_sock
*req
,
6301 const struct tcp_options_received
*rx_opt
,
6302 struct sk_buff
*skb
, const struct sock
*sk
)
6304 struct inet_request_sock
*ireq
= inet_rsk(req
);
6306 req
->rsk_rcv_wnd
= 0; /* So that tcp_send_synack() knows! */
6308 tcp_rsk(req
)->rcv_isn
= TCP_SKB_CB(skb
)->seq
;
6309 tcp_rsk(req
)->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
6310 tcp_rsk(req
)->snt_synack
= tcp_clock_us();
6311 tcp_rsk(req
)->last_oow_ack_time
= 0;
6312 req
->mss
= rx_opt
->mss_clamp
;
6313 req
->ts_recent
= rx_opt
->saw_tstamp
? rx_opt
->rcv_tsval
: 0;
6314 ireq
->tstamp_ok
= rx_opt
->tstamp_ok
;
6315 ireq
->sack_ok
= rx_opt
->sack_ok
;
6316 ireq
->snd_wscale
= rx_opt
->snd_wscale
;
6317 ireq
->wscale_ok
= rx_opt
->wscale_ok
;
6320 ireq
->ir_rmt_port
= tcp_hdr(skb
)->source
;
6321 ireq
->ir_num
= ntohs(tcp_hdr(skb
)->dest
);
6322 ireq
->ir_mark
= inet_request_mark(sk
, skb
);
6323 #if IS_ENABLED(CONFIG_SMC)
6324 ireq
->smc_ok
= rx_opt
->smc_ok
;
6328 struct request_sock
*inet_reqsk_alloc(const struct request_sock_ops
*ops
,
6329 struct sock
*sk_listener
,
6330 bool attach_listener
)
6332 struct request_sock
*req
= reqsk_alloc(ops
, sk_listener
,
6336 struct inet_request_sock
*ireq
= inet_rsk(req
);
6338 ireq
->ireq_opt
= NULL
;
6339 #if IS_ENABLED(CONFIG_IPV6)
6340 ireq
->pktopts
= NULL
;
6342 atomic64_set(&ireq
->ir_cookie
, 0);
6343 ireq
->ireq_state
= TCP_NEW_SYN_RECV
;
6344 write_pnet(&ireq
->ireq_net
, sock_net(sk_listener
));
6345 ireq
->ireq_family
= sk_listener
->sk_family
;
6350 EXPORT_SYMBOL(inet_reqsk_alloc
);
6353 * Return true if a syncookie should be sent
6355 static bool tcp_syn_flood_action(const struct sock
*sk
,
6356 const struct sk_buff
*skb
,
6359 struct request_sock_queue
*queue
= &inet_csk(sk
)->icsk_accept_queue
;
6360 const char *msg
= "Dropping request";
6361 bool want_cookie
= false;
6362 struct net
*net
= sock_net(sk
);
6364 #ifdef CONFIG_SYN_COOKIES
6365 if (net
->ipv4
.sysctl_tcp_syncookies
) {
6366 msg
= "Sending cookies";
6368 __NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPREQQFULLDOCOOKIES
);
6371 __NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPREQQFULLDROP
);
6373 if (!queue
->synflood_warned
&&
6374 net
->ipv4
.sysctl_tcp_syncookies
!= 2 &&
6375 xchg(&queue
->synflood_warned
, 1) == 0)
6376 net_info_ratelimited("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6377 proto
, ntohs(tcp_hdr(skb
)->dest
), msg
);
6382 static void tcp_reqsk_record_syn(const struct sock
*sk
,
6383 struct request_sock
*req
,
6384 const struct sk_buff
*skb
)
6386 if (tcp_sk(sk
)->save_syn
) {
6387 u32 len
= skb_network_header_len(skb
) + tcp_hdrlen(skb
);
6390 copy
= kmalloc(len
+ sizeof(u32
), GFP_ATOMIC
);
6393 memcpy(©
[1], skb_network_header(skb
), len
);
6394 req
->saved_syn
= copy
;
6399 int tcp_conn_request(struct request_sock_ops
*rsk_ops
,
6400 const struct tcp_request_sock_ops
*af_ops
,
6401 struct sock
*sk
, struct sk_buff
*skb
)
6403 struct tcp_fastopen_cookie foc
= { .len
= -1 };
6404 __u32 isn
= TCP_SKB_CB(skb
)->tcp_tw_isn
;
6405 struct tcp_options_received tmp_opt
;
6406 struct tcp_sock
*tp
= tcp_sk(sk
);
6407 struct net
*net
= sock_net(sk
);
6408 struct sock
*fastopen_sk
= NULL
;
6409 struct request_sock
*req
;
6410 bool want_cookie
= false;
6411 struct dst_entry
*dst
;
6414 /* TW buckets are converted to open requests without
6415 * limitations, they conserve resources and peer is
6416 * evidently real one.
6418 if ((net
->ipv4
.sysctl_tcp_syncookies
== 2 ||
6419 inet_csk_reqsk_queue_is_full(sk
)) && !isn
) {
6420 want_cookie
= tcp_syn_flood_action(sk
, skb
, rsk_ops
->slab_name
);
6425 if (sk_acceptq_is_full(sk
)) {
6426 NET_INC_STATS(sock_net(sk
), LINUX_MIB_LISTENOVERFLOWS
);
6430 req
= inet_reqsk_alloc(rsk_ops
, sk
, !want_cookie
);
6434 tcp_rsk(req
)->af_specific
= af_ops
;
6435 tcp_rsk(req
)->ts_off
= 0;
6437 tcp_clear_options(&tmp_opt
);
6438 tmp_opt
.mss_clamp
= af_ops
->mss_clamp
;
6439 tmp_opt
.user_mss
= tp
->rx_opt
.user_mss
;
6440 tcp_parse_options(sock_net(sk
), skb
, &tmp_opt
, 0,
6441 want_cookie
? NULL
: &foc
);
6443 if (want_cookie
&& !tmp_opt
.saw_tstamp
)
6444 tcp_clear_options(&tmp_opt
);
6446 if (IS_ENABLED(CONFIG_SMC
) && want_cookie
)
6449 tmp_opt
.tstamp_ok
= tmp_opt
.saw_tstamp
;
6450 tcp_openreq_init(req
, &tmp_opt
, skb
, sk
);
6451 inet_rsk(req
)->no_srccheck
= inet_sk(sk
)->transparent
;
6453 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6454 inet_rsk(req
)->ir_iif
= inet_request_bound_dev_if(sk
, skb
);
6456 af_ops
->init_req(req
, sk
, skb
);
6458 if (security_inet_conn_request(sk
, skb
, req
))
6461 if (tmp_opt
.tstamp_ok
)
6462 tcp_rsk(req
)->ts_off
= af_ops
->init_ts_off(net
, skb
);
6464 dst
= af_ops
->route_req(sk
, &fl
, req
);
6468 if (!want_cookie
&& !isn
) {
6469 /* Kill the following clause, if you dislike this way. */
6470 if (!net
->ipv4
.sysctl_tcp_syncookies
&&
6471 (net
->ipv4
.sysctl_max_syn_backlog
- inet_csk_reqsk_queue_len(sk
) <
6472 (net
->ipv4
.sysctl_max_syn_backlog
>> 2)) &&
6473 !tcp_peer_is_proven(req
, dst
)) {
6474 /* Without syncookies last quarter of
6475 * backlog is filled with destinations,
6476 * proven to be alive.
6477 * It means that we continue to communicate
6478 * to destinations, already remembered
6479 * to the moment of synflood.
6481 pr_drop_req(req
, ntohs(tcp_hdr(skb
)->source
),
6483 goto drop_and_release
;
6486 isn
= af_ops
->init_seq(skb
);
6489 tcp_ecn_create_request(req
, skb
, sk
, dst
);
6492 isn
= cookie_init_sequence(af_ops
, sk
, skb
, &req
->mss
);
6493 req
->cookie_ts
= tmp_opt
.tstamp_ok
;
6494 if (!tmp_opt
.tstamp_ok
)
6495 inet_rsk(req
)->ecn_ok
= 0;
6498 tcp_rsk(req
)->snt_isn
= isn
;
6499 tcp_rsk(req
)->txhash
= net_tx_rndhash();
6500 tcp_openreq_init_rwin(req
, sk
, dst
);
6501 sk_rx_queue_set(req_to_sk(req
), skb
);
6503 tcp_reqsk_record_syn(sk
, req
, skb
);
6504 fastopen_sk
= tcp_try_fastopen(sk
, skb
, req
, &foc
, dst
);
6507 af_ops
->send_synack(fastopen_sk
, dst
, &fl
, req
,
6508 &foc
, TCP_SYNACK_FASTOPEN
);
6509 /* Add the child socket directly into the accept queue */
6510 inet_csk_reqsk_queue_add(sk
, req
, fastopen_sk
);
6511 sk
->sk_data_ready(sk
);
6512 bh_unlock_sock(fastopen_sk
);
6513 sock_put(fastopen_sk
);
6515 tcp_rsk(req
)->tfo_listener
= false;
6517 inet_csk_reqsk_queue_hash_add(sk
, req
,
6518 tcp_timeout_init((struct sock
*)req
));
6519 af_ops
->send_synack(sk
, dst
, &fl
, req
, &foc
,
6520 !want_cookie
? TCP_SYNACK_NORMAL
:
6538 EXPORT_SYMBOL(tcp_conn_request
);