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/jump_label_ratelimit.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_DEFERRED_FALSE(clean_acked_data_enabled
, HZ
);
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_deferred_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_slow_dec_deferred(&clean_acked_data_enabled
);
129 icsk
->icsk_clean_acked
= NULL
;
131 EXPORT_SYMBOL_GPL(clean_acked_data_disable
);
133 void clean_acked_data_flush(void)
135 static_key_deferred_flush(&clean_acked_data_enabled
);
137 EXPORT_SYMBOL_GPL(clean_acked_data_flush
);
140 static void tcp_gro_dev_warn(struct sock
*sk
, const struct sk_buff
*skb
,
143 static bool __once __read_mostly
;
146 struct net_device
*dev
;
151 dev
= dev_get_by_index_rcu(sock_net(sk
), skb
->skb_iif
);
152 if (!dev
|| len
>= dev
->mtu
)
153 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n",
154 dev
? dev
->name
: "Unknown driver");
159 /* Adapt the MSS value used to make delayed ack decision to the
162 static void tcp_measure_rcv_mss(struct sock
*sk
, const struct sk_buff
*skb
)
164 struct inet_connection_sock
*icsk
= inet_csk(sk
);
165 const unsigned int lss
= icsk
->icsk_ack
.last_seg_size
;
168 icsk
->icsk_ack
.last_seg_size
= 0;
170 /* skb->len may jitter because of SACKs, even if peer
171 * sends good full-sized frames.
173 len
= skb_shinfo(skb
)->gso_size
? : skb
->len
;
174 if (len
>= icsk
->icsk_ack
.rcv_mss
) {
175 icsk
->icsk_ack
.rcv_mss
= min_t(unsigned int, len
,
177 /* Account for possibly-removed options */
178 if (unlikely(len
> icsk
->icsk_ack
.rcv_mss
+
179 MAX_TCP_OPTION_SPACE
))
180 tcp_gro_dev_warn(sk
, skb
, len
);
182 /* Otherwise, we make more careful check taking into account,
183 * that SACKs block is variable.
185 * "len" is invariant segment length, including TCP header.
187 len
+= skb
->data
- skb_transport_header(skb
);
188 if (len
>= TCP_MSS_DEFAULT
+ sizeof(struct tcphdr
) ||
189 /* If PSH is not set, packet should be
190 * full sized, provided peer TCP is not badly broken.
191 * This observation (if it is correct 8)) allows
192 * to handle super-low mtu links fairly.
194 (len
>= TCP_MIN_MSS
+ sizeof(struct tcphdr
) &&
195 !(tcp_flag_word(tcp_hdr(skb
)) & TCP_REMNANT
))) {
196 /* Subtract also invariant (if peer is RFC compliant),
197 * tcp header plus fixed timestamp option length.
198 * Resulting "len" is MSS free of SACK jitter.
200 len
-= tcp_sk(sk
)->tcp_header_len
;
201 icsk
->icsk_ack
.last_seg_size
= len
;
203 icsk
->icsk_ack
.rcv_mss
= len
;
207 if (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
)
208 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED2
;
209 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
213 static void tcp_incr_quickack(struct sock
*sk
, unsigned int max_quickacks
)
215 struct inet_connection_sock
*icsk
= inet_csk(sk
);
216 unsigned int quickacks
= tcp_sk(sk
)->rcv_wnd
/ (2 * icsk
->icsk_ack
.rcv_mss
);
220 quickacks
= min(quickacks
, max_quickacks
);
221 if (quickacks
> icsk
->icsk_ack
.quick
)
222 icsk
->icsk_ack
.quick
= quickacks
;
225 void tcp_enter_quickack_mode(struct sock
*sk
, unsigned int max_quickacks
)
227 struct inet_connection_sock
*icsk
= inet_csk(sk
);
229 tcp_incr_quickack(sk
, max_quickacks
);
230 inet_csk_exit_pingpong_mode(sk
);
231 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
233 EXPORT_SYMBOL(tcp_enter_quickack_mode
);
235 /* Send ACKs quickly, if "quick" count is not exhausted
236 * and the session is not interactive.
239 static bool tcp_in_quickack_mode(struct sock
*sk
)
241 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
242 const struct dst_entry
*dst
= __sk_dst_get(sk
);
244 return (dst
&& dst_metric(dst
, RTAX_QUICKACK
)) ||
245 (icsk
->icsk_ack
.quick
&& !inet_csk_in_pingpong_mode(sk
));
248 static void tcp_ecn_queue_cwr(struct tcp_sock
*tp
)
250 if (tp
->ecn_flags
& TCP_ECN_OK
)
251 tp
->ecn_flags
|= TCP_ECN_QUEUE_CWR
;
254 static void tcp_ecn_accept_cwr(struct sock
*sk
, const struct sk_buff
*skb
)
256 if (tcp_hdr(skb
)->cwr
) {
257 tcp_sk(sk
)->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
259 /* If the sender is telling us it has entered CWR, then its
260 * cwnd may be very low (even just 1 packet), so we should ACK
263 inet_csk(sk
)->icsk_ack
.pending
|= ICSK_ACK_NOW
;
267 static void tcp_ecn_withdraw_cwr(struct tcp_sock
*tp
)
269 tp
->ecn_flags
&= ~TCP_ECN_QUEUE_CWR
;
272 static void __tcp_ecn_check_ce(struct sock
*sk
, const struct sk_buff
*skb
)
274 struct tcp_sock
*tp
= tcp_sk(sk
);
276 switch (TCP_SKB_CB(skb
)->ip_dsfield
& INET_ECN_MASK
) {
277 case INET_ECN_NOT_ECT
:
278 /* Funny extension: if ECT is not set on a segment,
279 * and we already seen ECT on a previous segment,
280 * it is probably a retransmit.
282 if (tp
->ecn_flags
& TCP_ECN_SEEN
)
283 tcp_enter_quickack_mode(sk
, 2);
286 if (tcp_ca_needs_ecn(sk
))
287 tcp_ca_event(sk
, CA_EVENT_ECN_IS_CE
);
289 if (!(tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)) {
290 /* Better not delay acks, sender can have a very low cwnd */
291 tcp_enter_quickack_mode(sk
, 2);
292 tp
->ecn_flags
|= TCP_ECN_DEMAND_CWR
;
294 tp
->ecn_flags
|= TCP_ECN_SEEN
;
297 if (tcp_ca_needs_ecn(sk
))
298 tcp_ca_event(sk
, CA_EVENT_ECN_NO_CE
);
299 tp
->ecn_flags
|= TCP_ECN_SEEN
;
304 static void tcp_ecn_check_ce(struct sock
*sk
, const struct sk_buff
*skb
)
306 if (tcp_sk(sk
)->ecn_flags
& TCP_ECN_OK
)
307 __tcp_ecn_check_ce(sk
, skb
);
310 static void tcp_ecn_rcv_synack(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 void tcp_ecn_rcv_syn(struct tcp_sock
*tp
, const struct tcphdr
*th
)
318 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| !th
->cwr
))
319 tp
->ecn_flags
&= ~TCP_ECN_OK
;
322 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock
*tp
, const struct tcphdr
*th
)
324 if (th
->ece
&& !th
->syn
&& (tp
->ecn_flags
& TCP_ECN_OK
))
329 /* Buffer size and advertised window tuning.
331 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
334 static void tcp_sndbuf_expand(struct sock
*sk
)
336 const struct tcp_sock
*tp
= tcp_sk(sk
);
337 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
341 /* Worst case is non GSO/TSO : each frame consumes one skb
342 * and skb->head is kmalloced using power of two area of memory
344 per_mss
= max_t(u32
, tp
->rx_opt
.mss_clamp
, tp
->mss_cache
) +
346 SKB_DATA_ALIGN(sizeof(struct skb_shared_info
));
348 per_mss
= roundup_pow_of_two(per_mss
) +
349 SKB_DATA_ALIGN(sizeof(struct sk_buff
));
351 nr_segs
= max_t(u32
, TCP_INIT_CWND
, tp
->snd_cwnd
);
352 nr_segs
= max_t(u32
, nr_segs
, tp
->reordering
+ 1);
354 /* Fast Recovery (RFC 5681 3.2) :
355 * Cubic needs 1.7 factor, rounded to 2 to include
356 * extra cushion (application might react slowly to EPOLLOUT)
358 sndmem
= ca_ops
->sndbuf_expand
? ca_ops
->sndbuf_expand(sk
) : 2;
359 sndmem
*= nr_segs
* per_mss
;
361 if (sk
->sk_sndbuf
< sndmem
)
362 WRITE_ONCE(sk
->sk_sndbuf
,
363 min(sndmem
, sock_net(sk
)->ipv4
.sysctl_tcp_wmem
[2]));
366 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
368 * All tcp_full_space() is split to two parts: "network" buffer, allocated
369 * forward and advertised in receiver window (tp->rcv_wnd) and
370 * "application buffer", required to isolate scheduling/application
371 * latencies from network.
372 * window_clamp is maximal advertised window. It can be less than
373 * tcp_full_space(), in this case tcp_full_space() - window_clamp
374 * is reserved for "application" buffer. The less window_clamp is
375 * the smoother our behaviour from viewpoint of network, but the lower
376 * throughput and the higher sensitivity of the connection to losses. 8)
378 * rcv_ssthresh is more strict window_clamp used at "slow start"
379 * phase to predict further behaviour of this connection.
380 * It is used for two goals:
381 * - to enforce header prediction at sender, even when application
382 * requires some significant "application buffer". It is check #1.
383 * - to prevent pruning of receive queue because of misprediction
384 * of receiver window. Check #2.
386 * The scheme does not work when sender sends good segments opening
387 * window and then starts to feed us spaghetti. But it should work
388 * in common situations. Otherwise, we have to rely on queue collapsing.
391 /* Slow part of check#2. */
392 static int __tcp_grow_window(const struct sock
*sk
, const struct sk_buff
*skb
)
394 struct tcp_sock
*tp
= tcp_sk(sk
);
396 int truesize
= tcp_win_from_space(sk
, skb
->truesize
) >> 1;
397 int window
= tcp_win_from_space(sk
, sock_net(sk
)->ipv4
.sysctl_tcp_rmem
[2]) >> 1;
399 while (tp
->rcv_ssthresh
<= window
) {
400 if (truesize
<= skb
->len
)
401 return 2 * inet_csk(sk
)->icsk_ack
.rcv_mss
;
409 static void tcp_grow_window(struct sock
*sk
, const struct sk_buff
*skb
)
411 struct tcp_sock
*tp
= tcp_sk(sk
);
414 room
= min_t(int, tp
->window_clamp
, tcp_space(sk
)) - tp
->rcv_ssthresh
;
417 if (room
> 0 && !tcp_under_memory_pressure(sk
)) {
420 /* Check #2. Increase window, if skb with such overhead
421 * will fit to rcvbuf in future.
423 if (tcp_win_from_space(sk
, skb
->truesize
) <= skb
->len
)
424 incr
= 2 * tp
->advmss
;
426 incr
= __tcp_grow_window(sk
, skb
);
429 incr
= max_t(int, incr
, 2 * skb
->len
);
430 tp
->rcv_ssthresh
+= min(room
, incr
);
431 inet_csk(sk
)->icsk_ack
.quick
|= 1;
436 /* 3. Try to fixup all. It is made immediately after connection enters
439 void tcp_init_buffer_space(struct sock
*sk
)
441 int tcp_app_win
= sock_net(sk
)->ipv4
.sysctl_tcp_app_win
;
442 struct tcp_sock
*tp
= tcp_sk(sk
);
445 if (!(sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
))
446 tcp_sndbuf_expand(sk
);
448 tp
->rcvq_space
.space
= min_t(u32
, tp
->rcv_wnd
, TCP_INIT_CWND
* tp
->advmss
);
449 tcp_mstamp_refresh(tp
);
450 tp
->rcvq_space
.time
= tp
->tcp_mstamp
;
451 tp
->rcvq_space
.seq
= tp
->copied_seq
;
453 maxwin
= tcp_full_space(sk
);
455 if (tp
->window_clamp
>= maxwin
) {
456 tp
->window_clamp
= maxwin
;
458 if (tcp_app_win
&& maxwin
> 4 * tp
->advmss
)
459 tp
->window_clamp
= max(maxwin
-
460 (maxwin
>> tcp_app_win
),
464 /* Force reservation of one segment. */
466 tp
->window_clamp
> 2 * tp
->advmss
&&
467 tp
->window_clamp
+ tp
->advmss
> maxwin
)
468 tp
->window_clamp
= max(2 * tp
->advmss
, maxwin
- tp
->advmss
);
470 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, tp
->window_clamp
);
471 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
474 /* 4. Recalculate window clamp after socket hit its memory bounds. */
475 static void tcp_clamp_window(struct sock
*sk
)
477 struct tcp_sock
*tp
= tcp_sk(sk
);
478 struct inet_connection_sock
*icsk
= inet_csk(sk
);
479 struct net
*net
= sock_net(sk
);
481 icsk
->icsk_ack
.quick
= 0;
483 if (sk
->sk_rcvbuf
< net
->ipv4
.sysctl_tcp_rmem
[2] &&
484 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
) &&
485 !tcp_under_memory_pressure(sk
) &&
486 sk_memory_allocated(sk
) < sk_prot_mem_limits(sk
, 0)) {
487 WRITE_ONCE(sk
->sk_rcvbuf
,
488 min(atomic_read(&sk
->sk_rmem_alloc
),
489 net
->ipv4
.sysctl_tcp_rmem
[2]));
491 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
)
492 tp
->rcv_ssthresh
= min(tp
->window_clamp
, 2U * tp
->advmss
);
495 /* Initialize RCV_MSS value.
496 * RCV_MSS is an our guess about MSS used by the peer.
497 * We haven't any direct information about the MSS.
498 * It's better to underestimate the RCV_MSS rather than overestimate.
499 * Overestimations make us ACKing less frequently than needed.
500 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
502 void tcp_initialize_rcv_mss(struct sock
*sk
)
504 const struct tcp_sock
*tp
= tcp_sk(sk
);
505 unsigned int hint
= min_t(unsigned int, tp
->advmss
, tp
->mss_cache
);
507 hint
= min(hint
, tp
->rcv_wnd
/ 2);
508 hint
= min(hint
, TCP_MSS_DEFAULT
);
509 hint
= max(hint
, TCP_MIN_MSS
);
511 inet_csk(sk
)->icsk_ack
.rcv_mss
= hint
;
513 EXPORT_SYMBOL(tcp_initialize_rcv_mss
);
515 /* Receiver "autotuning" code.
517 * The algorithm for RTT estimation w/o timestamps is based on
518 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
519 * <http://public.lanl.gov/radiant/pubs.html#DRS>
521 * More detail on this code can be found at
522 * <http://staff.psc.edu/jheffner/>,
523 * though this reference is out of date. A new paper
526 static void tcp_rcv_rtt_update(struct tcp_sock
*tp
, u32 sample
, int win_dep
)
528 u32 new_sample
= tp
->rcv_rtt_est
.rtt_us
;
531 if (new_sample
!= 0) {
532 /* If we sample in larger samples in the non-timestamp
533 * case, we could grossly overestimate the RTT especially
534 * with chatty applications or bulk transfer apps which
535 * are stalled on filesystem I/O.
537 * Also, since we are only going for a minimum in the
538 * non-timestamp case, we do not smooth things out
539 * else with timestamps disabled convergence takes too
543 m
-= (new_sample
>> 3);
551 /* No previous measure. */
555 tp
->rcv_rtt_est
.rtt_us
= new_sample
;
558 static inline void tcp_rcv_rtt_measure(struct tcp_sock
*tp
)
562 if (tp
->rcv_rtt_est
.time
== 0)
564 if (before(tp
->rcv_nxt
, tp
->rcv_rtt_est
.seq
))
566 delta_us
= tcp_stamp_us_delta(tp
->tcp_mstamp
, tp
->rcv_rtt_est
.time
);
569 tcp_rcv_rtt_update(tp
, delta_us
, 1);
572 tp
->rcv_rtt_est
.seq
= tp
->rcv_nxt
+ tp
->rcv_wnd
;
573 tp
->rcv_rtt_est
.time
= tp
->tcp_mstamp
;
576 static inline void tcp_rcv_rtt_measure_ts(struct sock
*sk
,
577 const struct sk_buff
*skb
)
579 struct tcp_sock
*tp
= tcp_sk(sk
);
581 if (tp
->rx_opt
.rcv_tsecr
== tp
->rcv_rtt_last_tsecr
)
583 tp
->rcv_rtt_last_tsecr
= tp
->rx_opt
.rcv_tsecr
;
585 if (TCP_SKB_CB(skb
)->end_seq
-
586 TCP_SKB_CB(skb
)->seq
>= inet_csk(sk
)->icsk_ack
.rcv_mss
) {
587 u32 delta
= tcp_time_stamp(tp
) - tp
->rx_opt
.rcv_tsecr
;
590 if (likely(delta
< INT_MAX
/ (USEC_PER_SEC
/ TCP_TS_HZ
))) {
593 delta_us
= delta
* (USEC_PER_SEC
/ TCP_TS_HZ
);
594 tcp_rcv_rtt_update(tp
, delta_us
, 0);
600 * This function should be called every time data is copied to user space.
601 * It calculates the appropriate TCP receive buffer space.
603 void tcp_rcv_space_adjust(struct sock
*sk
)
605 struct tcp_sock
*tp
= tcp_sk(sk
);
609 trace_tcp_rcv_space_adjust(sk
);
611 tcp_mstamp_refresh(tp
);
612 time
= tcp_stamp_us_delta(tp
->tcp_mstamp
, tp
->rcvq_space
.time
);
613 if (time
< (tp
->rcv_rtt_est
.rtt_us
>> 3) || tp
->rcv_rtt_est
.rtt_us
== 0)
616 /* Number of bytes copied to user in last RTT */
617 copied
= tp
->copied_seq
- tp
->rcvq_space
.seq
;
618 if (copied
<= tp
->rcvq_space
.space
)
622 * copied = bytes received in previous RTT, our base window
623 * To cope with packet losses, we need a 2x factor
624 * To cope with slow start, and sender growing its cwin by 100 %
625 * every RTT, we need a 4x factor, because the ACK we are sending
626 * now is for the next RTT, not the current one :
627 * <prev RTT . ><current RTT .. ><next RTT .... >
630 if (sock_net(sk
)->ipv4
.sysctl_tcp_moderate_rcvbuf
&&
631 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
)) {
635 /* minimal window to cope with packet losses, assuming
636 * steady state. Add some cushion because of small variations.
638 rcvwin
= ((u64
)copied
<< 1) + 16 * tp
->advmss
;
640 /* Accommodate for sender rate increase (eg. slow start) */
641 grow
= rcvwin
* (copied
- tp
->rcvq_space
.space
);
642 do_div(grow
, tp
->rcvq_space
.space
);
643 rcvwin
+= (grow
<< 1);
645 rcvmem
= SKB_TRUESIZE(tp
->advmss
+ MAX_TCP_HEADER
);
646 while (tcp_win_from_space(sk
, rcvmem
) < tp
->advmss
)
649 do_div(rcvwin
, tp
->advmss
);
650 rcvbuf
= min_t(u64
, rcvwin
* rcvmem
,
651 sock_net(sk
)->ipv4
.sysctl_tcp_rmem
[2]);
652 if (rcvbuf
> sk
->sk_rcvbuf
) {
653 WRITE_ONCE(sk
->sk_rcvbuf
, rcvbuf
);
655 /* Make the window clamp follow along. */
656 tp
->window_clamp
= tcp_win_from_space(sk
, rcvbuf
);
659 tp
->rcvq_space
.space
= copied
;
662 tp
->rcvq_space
.seq
= tp
->copied_seq
;
663 tp
->rcvq_space
.time
= tp
->tcp_mstamp
;
666 /* There is something which you must keep in mind when you analyze the
667 * behavior of the tp->ato delayed ack timeout interval. When a
668 * connection starts up, we want to ack as quickly as possible. The
669 * problem is that "good" TCP's do slow start at the beginning of data
670 * transmission. The means that until we send the first few ACK's the
671 * sender will sit on his end and only queue most of his data, because
672 * he can only send snd_cwnd unacked packets at any given time. For
673 * each ACK we send, he increments snd_cwnd and transmits more of his
676 static void tcp_event_data_recv(struct sock
*sk
, struct sk_buff
*skb
)
678 struct tcp_sock
*tp
= tcp_sk(sk
);
679 struct inet_connection_sock
*icsk
= inet_csk(sk
);
682 inet_csk_schedule_ack(sk
);
684 tcp_measure_rcv_mss(sk
, skb
);
686 tcp_rcv_rtt_measure(tp
);
690 if (!icsk
->icsk_ack
.ato
) {
691 /* The _first_ data packet received, initialize
692 * delayed ACK engine.
694 tcp_incr_quickack(sk
, TCP_MAX_QUICKACKS
);
695 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
697 int m
= now
- icsk
->icsk_ack
.lrcvtime
;
699 if (m
<= TCP_ATO_MIN
/ 2) {
700 /* The fastest case is the first. */
701 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + TCP_ATO_MIN
/ 2;
702 } else if (m
< icsk
->icsk_ack
.ato
) {
703 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + m
;
704 if (icsk
->icsk_ack
.ato
> icsk
->icsk_rto
)
705 icsk
->icsk_ack
.ato
= icsk
->icsk_rto
;
706 } else if (m
> icsk
->icsk_rto
) {
707 /* Too long gap. Apparently sender failed to
708 * restart window, so that we send ACKs quickly.
710 tcp_incr_quickack(sk
, TCP_MAX_QUICKACKS
);
714 icsk
->icsk_ack
.lrcvtime
= now
;
716 tcp_ecn_check_ce(sk
, skb
);
719 tcp_grow_window(sk
, skb
);
722 /* Called to compute a smoothed rtt estimate. The data fed to this
723 * routine either comes from timestamps, or from segments that were
724 * known _not_ to have been retransmitted [see Karn/Partridge
725 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
726 * piece by Van Jacobson.
727 * NOTE: the next three routines used to be one big routine.
728 * To save cycles in the RFC 1323 implementation it was better to break
729 * it up into three procedures. -- erics
731 static void tcp_rtt_estimator(struct sock
*sk
, long mrtt_us
)
733 struct tcp_sock
*tp
= tcp_sk(sk
);
734 long m
= mrtt_us
; /* RTT */
735 u32 srtt
= tp
->srtt_us
;
737 /* The following amusing code comes from Jacobson's
738 * article in SIGCOMM '88. Note that rtt and mdev
739 * are scaled versions of rtt and mean deviation.
740 * This is designed to be as fast as possible
741 * m stands for "measurement".
743 * On a 1990 paper the rto value is changed to:
744 * RTO = rtt + 4 * mdev
746 * Funny. This algorithm seems to be very broken.
747 * These formulae increase RTO, when it should be decreased, increase
748 * too slowly, when it should be increased quickly, decrease too quickly
749 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
750 * does not matter how to _calculate_ it. Seems, it was trap
751 * that VJ failed to avoid. 8)
754 m
-= (srtt
>> 3); /* m is now error in rtt est */
755 srtt
+= m
; /* rtt = 7/8 rtt + 1/8 new */
757 m
= -m
; /* m is now abs(error) */
758 m
-= (tp
->mdev_us
>> 2); /* similar update on mdev */
759 /* This is similar to one of Eifel findings.
760 * Eifel blocks mdev updates when rtt decreases.
761 * This solution is a bit different: we use finer gain
762 * for mdev in this case (alpha*beta).
763 * Like Eifel it also prevents growth of rto,
764 * but also it limits too fast rto decreases,
765 * happening in pure Eifel.
770 m
-= (tp
->mdev_us
>> 2); /* similar update on mdev */
772 tp
->mdev_us
+= m
; /* mdev = 3/4 mdev + 1/4 new */
773 if (tp
->mdev_us
> tp
->mdev_max_us
) {
774 tp
->mdev_max_us
= tp
->mdev_us
;
775 if (tp
->mdev_max_us
> tp
->rttvar_us
)
776 tp
->rttvar_us
= tp
->mdev_max_us
;
778 if (after(tp
->snd_una
, tp
->rtt_seq
)) {
779 if (tp
->mdev_max_us
< tp
->rttvar_us
)
780 tp
->rttvar_us
-= (tp
->rttvar_us
- tp
->mdev_max_us
) >> 2;
781 tp
->rtt_seq
= tp
->snd_nxt
;
782 tp
->mdev_max_us
= tcp_rto_min_us(sk
);
787 /* no previous measure. */
788 srtt
= m
<< 3; /* take the measured time to be rtt */
789 tp
->mdev_us
= m
<< 1; /* make sure rto = 3*rtt */
790 tp
->rttvar_us
= max(tp
->mdev_us
, tcp_rto_min_us(sk
));
791 tp
->mdev_max_us
= tp
->rttvar_us
;
792 tp
->rtt_seq
= tp
->snd_nxt
;
796 tp
->srtt_us
= max(1U, srtt
);
799 static void tcp_update_pacing_rate(struct sock
*sk
)
801 const struct tcp_sock
*tp
= tcp_sk(sk
);
804 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
805 rate
= (u64
)tp
->mss_cache
* ((USEC_PER_SEC
/ 100) << 3);
807 /* current rate is (cwnd * mss) / srtt
808 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
809 * In Congestion Avoidance phase, set it to 120 % the current rate.
811 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
812 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
813 * end of slow start and should slow down.
815 if (tp
->snd_cwnd
< tp
->snd_ssthresh
/ 2)
816 rate
*= sock_net(sk
)->ipv4
.sysctl_tcp_pacing_ss_ratio
;
818 rate
*= sock_net(sk
)->ipv4
.sysctl_tcp_pacing_ca_ratio
;
820 rate
*= max(tp
->snd_cwnd
, tp
->packets_out
);
822 if (likely(tp
->srtt_us
))
823 do_div(rate
, tp
->srtt_us
);
825 /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate
826 * without any lock. We want to make sure compiler wont store
827 * intermediate values in this location.
829 WRITE_ONCE(sk
->sk_pacing_rate
, min_t(u64
, rate
,
830 sk
->sk_max_pacing_rate
));
833 /* Calculate rto without backoff. This is the second half of Van Jacobson's
834 * routine referred to above.
836 static void tcp_set_rto(struct sock
*sk
)
838 const struct tcp_sock
*tp
= tcp_sk(sk
);
839 /* Old crap is replaced with new one. 8)
842 * 1. If rtt variance happened to be less 50msec, it is hallucination.
843 * It cannot be less due to utterly erratic ACK generation made
844 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
845 * to do with delayed acks, because at cwnd>2 true delack timeout
846 * is invisible. Actually, Linux-2.4 also generates erratic
847 * ACKs in some circumstances.
849 inet_csk(sk
)->icsk_rto
= __tcp_set_rto(tp
);
851 /* 2. Fixups made earlier cannot be right.
852 * If we do not estimate RTO correctly without them,
853 * all the algo is pure shit and should be replaced
854 * with correct one. It is exactly, which we pretend to do.
857 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
858 * guarantees that rto is higher.
863 __u32
tcp_init_cwnd(const struct tcp_sock
*tp
, const struct dst_entry
*dst
)
865 __u32 cwnd
= (dst
? dst_metric(dst
, RTAX_INITCWND
) : 0);
868 cwnd
= TCP_INIT_CWND
;
869 return min_t(__u32
, cwnd
, tp
->snd_cwnd_clamp
);
872 /* Take a notice that peer is sending D-SACKs */
873 static void tcp_dsack_seen(struct tcp_sock
*tp
)
875 tp
->rx_opt
.sack_ok
|= TCP_DSACK_SEEN
;
876 tp
->rack
.dsack_seen
= 1;
880 /* It's reordering when higher sequence was delivered (i.e. sacked) before
881 * some lower never-retransmitted sequence ("low_seq"). The maximum reordering
882 * distance is approximated in full-mss packet distance ("reordering").
884 static void tcp_check_sack_reordering(struct sock
*sk
, const u32 low_seq
,
887 struct tcp_sock
*tp
= tcp_sk(sk
);
888 const u32 mss
= tp
->mss_cache
;
891 fack
= tcp_highest_sack_seq(tp
);
892 if (!before(low_seq
, fack
))
895 metric
= fack
- low_seq
;
896 if ((metric
> tp
->reordering
* mss
) && mss
) {
897 #if FASTRETRANS_DEBUG > 1
898 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
899 tp
->rx_opt
.sack_ok
, inet_csk(sk
)->icsk_ca_state
,
903 tp
->undo_marker
? tp
->undo_retrans
: 0);
905 tp
->reordering
= min_t(u32
, (metric
+ mss
- 1) / mss
,
906 sock_net(sk
)->ipv4
.sysctl_tcp_max_reordering
);
909 /* This exciting event is worth to be remembered. 8) */
911 NET_INC_STATS(sock_net(sk
),
912 ts
? LINUX_MIB_TCPTSREORDER
: LINUX_MIB_TCPSACKREORDER
);
915 /* This must be called before lost_out is incremented */
916 static void tcp_verify_retransmit_hint(struct tcp_sock
*tp
, struct sk_buff
*skb
)
918 if (!tp
->retransmit_skb_hint
||
919 before(TCP_SKB_CB(skb
)->seq
,
920 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
921 tp
->retransmit_skb_hint
= skb
;
924 /* Sum the number of packets on the wire we have marked as lost.
925 * There are two cases we care about here:
926 * a) Packet hasn't been marked lost (nor retransmitted),
927 * and this is the first loss.
928 * b) Packet has been marked both lost and retransmitted,
929 * and this means we think it was lost again.
931 static void tcp_sum_lost(struct tcp_sock
*tp
, struct sk_buff
*skb
)
933 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
935 if (!(sacked
& TCPCB_LOST
) ||
936 ((sacked
& TCPCB_LOST
) && (sacked
& TCPCB_SACKED_RETRANS
)))
937 tp
->lost
+= tcp_skb_pcount(skb
);
940 static void tcp_skb_mark_lost(struct tcp_sock
*tp
, struct sk_buff
*skb
)
942 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
943 tcp_verify_retransmit_hint(tp
, skb
);
945 tp
->lost_out
+= tcp_skb_pcount(skb
);
946 tcp_sum_lost(tp
, skb
);
947 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
951 void tcp_skb_mark_lost_uncond_verify(struct tcp_sock
*tp
, struct sk_buff
*skb
)
953 tcp_verify_retransmit_hint(tp
, skb
);
955 tcp_sum_lost(tp
, skb
);
956 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
957 tp
->lost_out
+= tcp_skb_pcount(skb
);
958 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
962 /* This procedure tags the retransmission queue when SACKs arrive.
964 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
965 * Packets in queue with these bits set are counted in variables
966 * sacked_out, retrans_out and lost_out, correspondingly.
968 * Valid combinations are:
969 * Tag InFlight Description
970 * 0 1 - orig segment is in flight.
971 * S 0 - nothing flies, orig reached receiver.
972 * L 0 - nothing flies, orig lost by net.
973 * R 2 - both orig and retransmit are in flight.
974 * L|R 1 - orig is lost, retransmit is in flight.
975 * S|R 1 - orig reached receiver, retrans is still in flight.
976 * (L|S|R is logically valid, it could occur when L|R is sacked,
977 * but it is equivalent to plain S and code short-curcuits it to S.
978 * L|S is logically invalid, it would mean -1 packet in flight 8))
980 * These 6 states form finite state machine, controlled by the following events:
981 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
982 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
983 * 3. Loss detection event of two flavors:
984 * A. Scoreboard estimator decided the packet is lost.
985 * A'. Reno "three dupacks" marks head of queue lost.
986 * B. SACK arrives sacking SND.NXT at the moment, when the
987 * segment was retransmitted.
988 * 4. D-SACK added new rule: D-SACK changes any tag to S.
990 * It is pleasant to note, that state diagram turns out to be commutative,
991 * so that we are allowed not to be bothered by order of our actions,
992 * when multiple events arrive simultaneously. (see the function below).
994 * Reordering detection.
995 * --------------------
996 * Reordering metric is maximal distance, which a packet can be displaced
997 * in packet stream. With SACKs we can estimate it:
999 * 1. SACK fills old hole and the corresponding segment was not
1000 * ever retransmitted -> reordering. Alas, we cannot use it
1001 * when segment was retransmitted.
1002 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1003 * for retransmitted and already SACKed segment -> reordering..
1004 * Both of these heuristics are not used in Loss state, when we cannot
1005 * account for retransmits accurately.
1007 * SACK block validation.
1008 * ----------------------
1010 * SACK block range validation checks that the received SACK block fits to
1011 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1012 * Note that SND.UNA is not included to the range though being valid because
1013 * it means that the receiver is rather inconsistent with itself reporting
1014 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1015 * perfectly valid, however, in light of RFC2018 which explicitly states
1016 * that "SACK block MUST reflect the newest segment. Even if the newest
1017 * segment is going to be discarded ...", not that it looks very clever
1018 * in case of head skb. Due to potentional receiver driven attacks, we
1019 * choose to avoid immediate execution of a walk in write queue due to
1020 * reneging and defer head skb's loss recovery to standard loss recovery
1021 * procedure that will eventually trigger (nothing forbids us doing this).
1023 * Implements also blockage to start_seq wrap-around. Problem lies in the
1024 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1025 * there's no guarantee that it will be before snd_nxt (n). The problem
1026 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1029 * <- outs wnd -> <- wrapzone ->
1030 * u e n u_w e_w s n_w
1032 * |<------------+------+----- TCP seqno space --------------+---------->|
1033 * ...-- <2^31 ->| |<--------...
1034 * ...---- >2^31 ------>| |<--------...
1036 * Current code wouldn't be vulnerable but it's better still to discard such
1037 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1038 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1039 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1040 * equal to the ideal case (infinite seqno space without wrap caused issues).
1042 * With D-SACK the lower bound is extended to cover sequence space below
1043 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1044 * again, D-SACK block must not to go across snd_una (for the same reason as
1045 * for the normal SACK blocks, explained above). But there all simplicity
1046 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1047 * fully below undo_marker they do not affect behavior in anyway and can
1048 * therefore be safely ignored. In rare cases (which are more or less
1049 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1050 * fragmentation and packet reordering past skb's retransmission. To consider
1051 * them correctly, the acceptable range must be extended even more though
1052 * the exact amount is rather hard to quantify. However, tp->max_window can
1053 * be used as an exaggerated estimate.
1055 static bool tcp_is_sackblock_valid(struct tcp_sock
*tp
, bool is_dsack
,
1056 u32 start_seq
, u32 end_seq
)
1058 /* Too far in future, or reversed (interpretation is ambiguous) */
1059 if (after(end_seq
, tp
->snd_nxt
) || !before(start_seq
, end_seq
))
1062 /* Nasty start_seq wrap-around check (see comments above) */
1063 if (!before(start_seq
, tp
->snd_nxt
))
1066 /* In outstanding window? ...This is valid exit for D-SACKs too.
1067 * start_seq == snd_una is non-sensical (see comments above)
1069 if (after(start_seq
, tp
->snd_una
))
1072 if (!is_dsack
|| !tp
->undo_marker
)
1075 /* ...Then it's D-SACK, and must reside below snd_una completely */
1076 if (after(end_seq
, tp
->snd_una
))
1079 if (!before(start_seq
, tp
->undo_marker
))
1083 if (!after(end_seq
, tp
->undo_marker
))
1086 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1087 * start_seq < undo_marker and end_seq >= undo_marker.
1089 return !before(start_seq
, end_seq
- tp
->max_window
);
1092 static bool tcp_check_dsack(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1093 struct tcp_sack_block_wire
*sp
, int num_sacks
,
1096 struct tcp_sock
*tp
= tcp_sk(sk
);
1097 u32 start_seq_0
= get_unaligned_be32(&sp
[0].start_seq
);
1098 u32 end_seq_0
= get_unaligned_be32(&sp
[0].end_seq
);
1099 bool dup_sack
= false;
1101 if (before(start_seq_0
, TCP_SKB_CB(ack_skb
)->ack_seq
)) {
1104 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPDSACKRECV
);
1105 } else if (num_sacks
> 1) {
1106 u32 end_seq_1
= get_unaligned_be32(&sp
[1].end_seq
);
1107 u32 start_seq_1
= get_unaligned_be32(&sp
[1].start_seq
);
1109 if (!after(end_seq_0
, end_seq_1
) &&
1110 !before(start_seq_0
, start_seq_1
)) {
1113 NET_INC_STATS(sock_net(sk
),
1114 LINUX_MIB_TCPDSACKOFORECV
);
1118 /* D-SACK for already forgotten data... Do dumb counting. */
1119 if (dup_sack
&& tp
->undo_marker
&& tp
->undo_retrans
> 0 &&
1120 !after(end_seq_0
, prior_snd_una
) &&
1121 after(end_seq_0
, tp
->undo_marker
))
1127 struct tcp_sacktag_state
{
1129 /* Timestamps for earliest and latest never-retransmitted segment
1130 * that was SACKed. RTO needs the earliest RTT to stay conservative,
1131 * but congestion control should still get an accurate delay signal.
1135 struct rate_sample
*rate
;
1137 unsigned int mss_now
;
1140 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1141 * the incoming SACK may not exactly match but we can find smaller MSS
1142 * aligned portion of it that matches. Therefore we might need to fragment
1143 * which may fail and creates some hassle (caller must handle error case
1146 * FIXME: this could be merged to shift decision code
1148 static int tcp_match_skb_to_sack(struct sock
*sk
, struct sk_buff
*skb
,
1149 u32 start_seq
, u32 end_seq
)
1153 unsigned int pkt_len
;
1156 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1157 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1159 if (tcp_skb_pcount(skb
) > 1 && !in_sack
&&
1160 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
1161 mss
= tcp_skb_mss(skb
);
1162 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1165 pkt_len
= start_seq
- TCP_SKB_CB(skb
)->seq
;
1169 pkt_len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1174 /* Round if necessary so that SACKs cover only full MSSes
1175 * and/or the remaining small portion (if present)
1177 if (pkt_len
> mss
) {
1178 unsigned int new_len
= (pkt_len
/ mss
) * mss
;
1179 if (!in_sack
&& new_len
< pkt_len
)
1184 if (pkt_len
>= skb
->len
&& !in_sack
)
1187 err
= tcp_fragment(sk
, TCP_FRAG_IN_RTX_QUEUE
, skb
,
1188 pkt_len
, mss
, GFP_ATOMIC
);
1196 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1197 static u8
tcp_sacktag_one(struct sock
*sk
,
1198 struct tcp_sacktag_state
*state
, u8 sacked
,
1199 u32 start_seq
, u32 end_seq
,
1200 int dup_sack
, int pcount
,
1203 struct tcp_sock
*tp
= tcp_sk(sk
);
1205 /* Account D-SACK for retransmitted packet. */
1206 if (dup_sack
&& (sacked
& TCPCB_RETRANS
)) {
1207 if (tp
->undo_marker
&& tp
->undo_retrans
> 0 &&
1208 after(end_seq
, tp
->undo_marker
))
1210 if ((sacked
& TCPCB_SACKED_ACKED
) &&
1211 before(start_seq
, state
->reord
))
1212 state
->reord
= start_seq
;
1215 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1216 if (!after(end_seq
, tp
->snd_una
))
1219 if (!(sacked
& TCPCB_SACKED_ACKED
)) {
1220 tcp_rack_advance(tp
, sacked
, end_seq
, xmit_time
);
1222 if (sacked
& TCPCB_SACKED_RETRANS
) {
1223 /* If the segment is not tagged as lost,
1224 * we do not clear RETRANS, believing
1225 * that retransmission is still in flight.
1227 if (sacked
& TCPCB_LOST
) {
1228 sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1229 tp
->lost_out
-= pcount
;
1230 tp
->retrans_out
-= pcount
;
1233 if (!(sacked
& TCPCB_RETRANS
)) {
1234 /* New sack for not retransmitted frame,
1235 * which was in hole. It is reordering.
1237 if (before(start_seq
,
1238 tcp_highest_sack_seq(tp
)) &&
1239 before(start_seq
, state
->reord
))
1240 state
->reord
= start_seq
;
1242 if (!after(end_seq
, tp
->high_seq
))
1243 state
->flag
|= FLAG_ORIG_SACK_ACKED
;
1244 if (state
->first_sackt
== 0)
1245 state
->first_sackt
= xmit_time
;
1246 state
->last_sackt
= xmit_time
;
1249 if (sacked
& TCPCB_LOST
) {
1250 sacked
&= ~TCPCB_LOST
;
1251 tp
->lost_out
-= pcount
;
1255 sacked
|= TCPCB_SACKED_ACKED
;
1256 state
->flag
|= FLAG_DATA_SACKED
;
1257 tp
->sacked_out
+= pcount
;
1258 tp
->delivered
+= pcount
; /* Out-of-order packets delivered */
1260 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1261 if (tp
->lost_skb_hint
&&
1262 before(start_seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
))
1263 tp
->lost_cnt_hint
+= pcount
;
1266 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1267 * frames and clear it. undo_retrans is decreased above, L|R frames
1268 * are accounted above as well.
1270 if (dup_sack
&& (sacked
& TCPCB_SACKED_RETRANS
)) {
1271 sacked
&= ~TCPCB_SACKED_RETRANS
;
1272 tp
->retrans_out
-= pcount
;
1278 /* Shift newly-SACKed bytes from this skb to the immediately previous
1279 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1281 static bool tcp_shifted_skb(struct sock
*sk
, struct sk_buff
*prev
,
1282 struct sk_buff
*skb
,
1283 struct tcp_sacktag_state
*state
,
1284 unsigned int pcount
, int shifted
, int mss
,
1287 struct tcp_sock
*tp
= tcp_sk(sk
);
1288 u32 start_seq
= TCP_SKB_CB(skb
)->seq
; /* start of newly-SACKed */
1289 u32 end_seq
= start_seq
+ shifted
; /* end of newly-SACKed */
1293 /* Adjust counters and hints for the newly sacked sequence
1294 * range but discard the return value since prev is already
1295 * marked. We must tag the range first because the seq
1296 * advancement below implicitly advances
1297 * tcp_highest_sack_seq() when skb is highest_sack.
1299 tcp_sacktag_one(sk
, state
, TCP_SKB_CB(skb
)->sacked
,
1300 start_seq
, end_seq
, dup_sack
, pcount
,
1301 tcp_skb_timestamp_us(skb
));
1302 tcp_rate_skb_delivered(sk
, skb
, state
->rate
);
1304 if (skb
== tp
->lost_skb_hint
)
1305 tp
->lost_cnt_hint
+= pcount
;
1307 TCP_SKB_CB(prev
)->end_seq
+= shifted
;
1308 TCP_SKB_CB(skb
)->seq
+= shifted
;
1310 tcp_skb_pcount_add(prev
, pcount
);
1311 WARN_ON_ONCE(tcp_skb_pcount(skb
) < pcount
);
1312 tcp_skb_pcount_add(skb
, -pcount
);
1314 /* When we're adding to gso_segs == 1, gso_size will be zero,
1315 * in theory this shouldn't be necessary but as long as DSACK
1316 * code can come after this skb later on it's better to keep
1317 * setting gso_size to something.
1319 if (!TCP_SKB_CB(prev
)->tcp_gso_size
)
1320 TCP_SKB_CB(prev
)->tcp_gso_size
= mss
;
1322 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1323 if (tcp_skb_pcount(skb
) <= 1)
1324 TCP_SKB_CB(skb
)->tcp_gso_size
= 0;
1326 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1327 TCP_SKB_CB(prev
)->sacked
|= (TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
);
1330 BUG_ON(!tcp_skb_pcount(skb
));
1331 NET_INC_STATS(sock_net(sk
), LINUX_MIB_SACKSHIFTED
);
1335 /* Whole SKB was eaten :-) */
1337 if (skb
== tp
->retransmit_skb_hint
)
1338 tp
->retransmit_skb_hint
= prev
;
1339 if (skb
== tp
->lost_skb_hint
) {
1340 tp
->lost_skb_hint
= prev
;
1341 tp
->lost_cnt_hint
-= tcp_skb_pcount(prev
);
1344 TCP_SKB_CB(prev
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
;
1345 TCP_SKB_CB(prev
)->eor
= TCP_SKB_CB(skb
)->eor
;
1346 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
1347 TCP_SKB_CB(prev
)->end_seq
++;
1349 if (skb
== tcp_highest_sack(sk
))
1350 tcp_advance_highest_sack(sk
, skb
);
1352 tcp_skb_collapse_tstamp(prev
, skb
);
1353 if (unlikely(TCP_SKB_CB(prev
)->tx
.delivered_mstamp
))
1354 TCP_SKB_CB(prev
)->tx
.delivered_mstamp
= 0;
1356 tcp_rtx_queue_unlink_and_free(skb
, sk
);
1358 NET_INC_STATS(sock_net(sk
), LINUX_MIB_SACKMERGED
);
1363 /* I wish gso_size would have a bit more sane initialization than
1364 * something-or-zero which complicates things
1366 static int tcp_skb_seglen(const struct sk_buff
*skb
)
1368 return tcp_skb_pcount(skb
) == 1 ? skb
->len
: tcp_skb_mss(skb
);
1371 /* Shifting pages past head area doesn't work */
1372 static int skb_can_shift(const struct sk_buff
*skb
)
1374 return !skb_headlen(skb
) && skb_is_nonlinear(skb
);
1377 int tcp_skb_shift(struct sk_buff
*to
, struct sk_buff
*from
,
1378 int pcount
, int shiftlen
)
1380 /* TCP min gso_size is 8 bytes (TCP_MIN_GSO_SIZE)
1381 * Since TCP_SKB_CB(skb)->tcp_gso_segs is 16 bits, we need
1382 * to make sure not storing more than 65535 * 8 bytes per skb,
1383 * even if current MSS is bigger.
1385 if (unlikely(to
->len
+ shiftlen
>= 65535 * TCP_MIN_GSO_SIZE
))
1387 if (unlikely(tcp_skb_pcount(to
) + pcount
> 65535))
1389 return skb_shift(to
, from
, shiftlen
);
1392 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1395 static struct sk_buff
*tcp_shift_skb_data(struct sock
*sk
, struct sk_buff
*skb
,
1396 struct tcp_sacktag_state
*state
,
1397 u32 start_seq
, u32 end_seq
,
1400 struct tcp_sock
*tp
= tcp_sk(sk
);
1401 struct sk_buff
*prev
;
1407 /* Normally R but no L won't result in plain S */
1409 (TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_RETRANS
)) == TCPCB_SACKED_RETRANS
)
1411 if (!skb_can_shift(skb
))
1413 /* This frame is about to be dropped (was ACKed). */
1414 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1417 /* Can only happen with delayed DSACK + discard craziness */
1418 prev
= skb_rb_prev(skb
);
1422 if ((TCP_SKB_CB(prev
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
)
1425 if (!tcp_skb_can_collapse_to(prev
))
1428 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1429 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1433 pcount
= tcp_skb_pcount(skb
);
1434 mss
= tcp_skb_seglen(skb
);
1436 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1437 * drop this restriction as unnecessary
1439 if (mss
!= tcp_skb_seglen(prev
))
1442 if (!after(TCP_SKB_CB(skb
)->end_seq
, start_seq
))
1444 /* CHECKME: This is non-MSS split case only?, this will
1445 * cause skipped skbs due to advancing loop btw, original
1446 * has that feature too
1448 if (tcp_skb_pcount(skb
) <= 1)
1451 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1453 /* TODO: head merge to next could be attempted here
1454 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1455 * though it might not be worth of the additional hassle
1457 * ...we can probably just fallback to what was done
1458 * previously. We could try merging non-SACKed ones
1459 * as well but it probably isn't going to buy off
1460 * because later SACKs might again split them, and
1461 * it would make skb timestamp tracking considerably
1467 len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1469 BUG_ON(len
> skb
->len
);
1471 /* MSS boundaries should be honoured or else pcount will
1472 * severely break even though it makes things bit trickier.
1473 * Optimize common case to avoid most of the divides
1475 mss
= tcp_skb_mss(skb
);
1477 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1478 * drop this restriction as unnecessary
1480 if (mss
!= tcp_skb_seglen(prev
))
1485 } else if (len
< mss
) {
1493 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1494 if (!after(TCP_SKB_CB(skb
)->seq
+ len
, tp
->snd_una
))
1497 if (!tcp_skb_shift(prev
, skb
, pcount
, len
))
1499 if (!tcp_shifted_skb(sk
, prev
, skb
, state
, pcount
, len
, mss
, dup_sack
))
1502 /* Hole filled allows collapsing with the next as well, this is very
1503 * useful when hole on every nth skb pattern happens
1505 skb
= skb_rb_next(prev
);
1509 if (!skb_can_shift(skb
) ||
1510 ((TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
) ||
1511 (mss
!= tcp_skb_seglen(skb
)))
1515 pcount
= tcp_skb_pcount(skb
);
1516 if (tcp_skb_shift(prev
, skb
, pcount
, len
))
1517 tcp_shifted_skb(sk
, prev
, skb
, state
, pcount
,
1527 NET_INC_STATS(sock_net(sk
), LINUX_MIB_SACKSHIFTFALLBACK
);
1531 static struct sk_buff
*tcp_sacktag_walk(struct sk_buff
*skb
, struct sock
*sk
,
1532 struct tcp_sack_block
*next_dup
,
1533 struct tcp_sacktag_state
*state
,
1534 u32 start_seq
, u32 end_seq
,
1537 struct tcp_sock
*tp
= tcp_sk(sk
);
1538 struct sk_buff
*tmp
;
1540 skb_rbtree_walk_from(skb
) {
1542 bool dup_sack
= dup_sack_in
;
1544 /* queue is in-order => we can short-circuit the walk early */
1545 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1549 before(TCP_SKB_CB(skb
)->seq
, next_dup
->end_seq
)) {
1550 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1551 next_dup
->start_seq
,
1557 /* skb reference here is a bit tricky to get right, since
1558 * shifting can eat and free both this skb and the next,
1559 * so not even _safe variant of the loop is enough.
1562 tmp
= tcp_shift_skb_data(sk
, skb
, state
,
1563 start_seq
, end_seq
, dup_sack
);
1572 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1578 if (unlikely(in_sack
< 0))
1582 TCP_SKB_CB(skb
)->sacked
=
1585 TCP_SKB_CB(skb
)->sacked
,
1586 TCP_SKB_CB(skb
)->seq
,
1587 TCP_SKB_CB(skb
)->end_seq
,
1589 tcp_skb_pcount(skb
),
1590 tcp_skb_timestamp_us(skb
));
1591 tcp_rate_skb_delivered(sk
, skb
, state
->rate
);
1592 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
1593 list_del_init(&skb
->tcp_tsorted_anchor
);
1595 if (!before(TCP_SKB_CB(skb
)->seq
,
1596 tcp_highest_sack_seq(tp
)))
1597 tcp_advance_highest_sack(sk
, skb
);
1603 static struct sk_buff
*tcp_sacktag_bsearch(struct sock
*sk
, u32 seq
)
1605 struct rb_node
*parent
, **p
= &sk
->tcp_rtx_queue
.rb_node
;
1606 struct sk_buff
*skb
;
1610 skb
= rb_to_skb(parent
);
1611 if (before(seq
, TCP_SKB_CB(skb
)->seq
)) {
1612 p
= &parent
->rb_left
;
1615 if (!before(seq
, TCP_SKB_CB(skb
)->end_seq
)) {
1616 p
= &parent
->rb_right
;
1624 static struct sk_buff
*tcp_sacktag_skip(struct sk_buff
*skb
, struct sock
*sk
,
1627 if (skb
&& after(TCP_SKB_CB(skb
)->seq
, skip_to_seq
))
1630 return tcp_sacktag_bsearch(sk
, skip_to_seq
);
1633 static struct sk_buff
*tcp_maybe_skipping_dsack(struct sk_buff
*skb
,
1635 struct tcp_sack_block
*next_dup
,
1636 struct tcp_sacktag_state
*state
,
1642 if (before(next_dup
->start_seq
, skip_to_seq
)) {
1643 skb
= tcp_sacktag_skip(skb
, sk
, next_dup
->start_seq
);
1644 skb
= tcp_sacktag_walk(skb
, sk
, NULL
, state
,
1645 next_dup
->start_seq
, next_dup
->end_seq
,
1652 static int tcp_sack_cache_ok(const struct tcp_sock
*tp
, const struct tcp_sack_block
*cache
)
1654 return cache
< tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1658 tcp_sacktag_write_queue(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1659 u32 prior_snd_una
, struct tcp_sacktag_state
*state
)
1661 struct tcp_sock
*tp
= tcp_sk(sk
);
1662 const unsigned char *ptr
= (skb_transport_header(ack_skb
) +
1663 TCP_SKB_CB(ack_skb
)->sacked
);
1664 struct tcp_sack_block_wire
*sp_wire
= (struct tcp_sack_block_wire
*)(ptr
+2);
1665 struct tcp_sack_block sp
[TCP_NUM_SACKS
];
1666 struct tcp_sack_block
*cache
;
1667 struct sk_buff
*skb
;
1668 int num_sacks
= min(TCP_NUM_SACKS
, (ptr
[1] - TCPOLEN_SACK_BASE
) >> 3);
1670 bool found_dup_sack
= false;
1672 int first_sack_index
;
1675 state
->reord
= tp
->snd_nxt
;
1677 if (!tp
->sacked_out
)
1678 tcp_highest_sack_reset(sk
);
1680 found_dup_sack
= tcp_check_dsack(sk
, ack_skb
, sp_wire
,
1681 num_sacks
, prior_snd_una
);
1682 if (found_dup_sack
) {
1683 state
->flag
|= FLAG_DSACKING_ACK
;
1684 tp
->delivered
++; /* A spurious retransmission is delivered */
1687 /* Eliminate too old ACKs, but take into
1688 * account more or less fresh ones, they can
1689 * contain valid SACK info.
1691 if (before(TCP_SKB_CB(ack_skb
)->ack_seq
, prior_snd_una
- tp
->max_window
))
1694 if (!tp
->packets_out
)
1698 first_sack_index
= 0;
1699 for (i
= 0; i
< num_sacks
; i
++) {
1700 bool dup_sack
= !i
&& found_dup_sack
;
1702 sp
[used_sacks
].start_seq
= get_unaligned_be32(&sp_wire
[i
].start_seq
);
1703 sp
[used_sacks
].end_seq
= get_unaligned_be32(&sp_wire
[i
].end_seq
);
1705 if (!tcp_is_sackblock_valid(tp
, dup_sack
,
1706 sp
[used_sacks
].start_seq
,
1707 sp
[used_sacks
].end_seq
)) {
1711 if (!tp
->undo_marker
)
1712 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDNOUNDO
;
1714 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDOLD
;
1716 /* Don't count olds caused by ACK reordering */
1717 if ((TCP_SKB_CB(ack_skb
)->ack_seq
!= tp
->snd_una
) &&
1718 !after(sp
[used_sacks
].end_seq
, tp
->snd_una
))
1720 mib_idx
= LINUX_MIB_TCPSACKDISCARD
;
1723 NET_INC_STATS(sock_net(sk
), mib_idx
);
1725 first_sack_index
= -1;
1729 /* Ignore very old stuff early */
1730 if (!after(sp
[used_sacks
].end_seq
, prior_snd_una
))
1736 /* order SACK blocks to allow in order walk of the retrans queue */
1737 for (i
= used_sacks
- 1; i
> 0; i
--) {
1738 for (j
= 0; j
< i
; j
++) {
1739 if (after(sp
[j
].start_seq
, sp
[j
+ 1].start_seq
)) {
1740 swap(sp
[j
], sp
[j
+ 1]);
1742 /* Track where the first SACK block goes to */
1743 if (j
== first_sack_index
)
1744 first_sack_index
= j
+ 1;
1749 state
->mss_now
= tcp_current_mss(sk
);
1753 if (!tp
->sacked_out
) {
1754 /* It's already past, so skip checking against it */
1755 cache
= tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1757 cache
= tp
->recv_sack_cache
;
1758 /* Skip empty blocks in at head of the cache */
1759 while (tcp_sack_cache_ok(tp
, cache
) && !cache
->start_seq
&&
1764 while (i
< used_sacks
) {
1765 u32 start_seq
= sp
[i
].start_seq
;
1766 u32 end_seq
= sp
[i
].end_seq
;
1767 bool dup_sack
= (found_dup_sack
&& (i
== first_sack_index
));
1768 struct tcp_sack_block
*next_dup
= NULL
;
1770 if (found_dup_sack
&& ((i
+ 1) == first_sack_index
))
1771 next_dup
= &sp
[i
+ 1];
1773 /* Skip too early cached blocks */
1774 while (tcp_sack_cache_ok(tp
, cache
) &&
1775 !before(start_seq
, cache
->end_seq
))
1778 /* Can skip some work by looking recv_sack_cache? */
1779 if (tcp_sack_cache_ok(tp
, cache
) && !dup_sack
&&
1780 after(end_seq
, cache
->start_seq
)) {
1783 if (before(start_seq
, cache
->start_seq
)) {
1784 skb
= tcp_sacktag_skip(skb
, sk
, start_seq
);
1785 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
,
1792 /* Rest of the block already fully processed? */
1793 if (!after(end_seq
, cache
->end_seq
))
1796 skb
= tcp_maybe_skipping_dsack(skb
, sk
, next_dup
,
1800 /* ...tail remains todo... */
1801 if (tcp_highest_sack_seq(tp
) == cache
->end_seq
) {
1802 /* ...but better entrypoint exists! */
1803 skb
= tcp_highest_sack(sk
);
1810 skb
= tcp_sacktag_skip(skb
, sk
, cache
->end_seq
);
1811 /* Check overlap against next cached too (past this one already) */
1816 if (!before(start_seq
, tcp_highest_sack_seq(tp
))) {
1817 skb
= tcp_highest_sack(sk
);
1821 skb
= tcp_sacktag_skip(skb
, sk
, start_seq
);
1824 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
, state
,
1825 start_seq
, end_seq
, dup_sack
);
1831 /* Clear the head of the cache sack blocks so we can skip it next time */
1832 for (i
= 0; i
< ARRAY_SIZE(tp
->recv_sack_cache
) - used_sacks
; i
++) {
1833 tp
->recv_sack_cache
[i
].start_seq
= 0;
1834 tp
->recv_sack_cache
[i
].end_seq
= 0;
1836 for (j
= 0; j
< used_sacks
; j
++)
1837 tp
->recv_sack_cache
[i
++] = sp
[j
];
1839 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Loss
|| tp
->undo_marker
)
1840 tcp_check_sack_reordering(sk
, state
->reord
, 0);
1842 tcp_verify_left_out(tp
);
1845 #if FASTRETRANS_DEBUG > 0
1846 WARN_ON((int)tp
->sacked_out
< 0);
1847 WARN_ON((int)tp
->lost_out
< 0);
1848 WARN_ON((int)tp
->retrans_out
< 0);
1849 WARN_ON((int)tcp_packets_in_flight(tp
) < 0);
1854 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1855 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1857 static bool tcp_limit_reno_sacked(struct tcp_sock
*tp
)
1861 holes
= max(tp
->lost_out
, 1U);
1862 holes
= min(holes
, tp
->packets_out
);
1864 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1865 tp
->sacked_out
= tp
->packets_out
- holes
;
1871 /* If we receive more dupacks than we expected counting segments
1872 * in assumption of absent reordering, interpret this as reordering.
1873 * The only another reason could be bug in receiver TCP.
1875 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1877 struct tcp_sock
*tp
= tcp_sk(sk
);
1879 if (!tcp_limit_reno_sacked(tp
))
1882 tp
->reordering
= min_t(u32
, tp
->packets_out
+ addend
,
1883 sock_net(sk
)->ipv4
.sysctl_tcp_max_reordering
);
1885 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPRENOREORDER
);
1888 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1890 static void tcp_add_reno_sack(struct sock
*sk
, int num_dupack
)
1893 struct tcp_sock
*tp
= tcp_sk(sk
);
1894 u32 prior_sacked
= tp
->sacked_out
;
1897 tp
->sacked_out
+= num_dupack
;
1898 tcp_check_reno_reordering(sk
, 0);
1899 delivered
= tp
->sacked_out
- prior_sacked
;
1901 tp
->delivered
+= delivered
;
1902 tcp_verify_left_out(tp
);
1906 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1908 static void tcp_remove_reno_sacks(struct sock
*sk
, int acked
)
1910 struct tcp_sock
*tp
= tcp_sk(sk
);
1913 /* One ACK acked hole. The rest eat duplicate ACKs. */
1914 tp
->delivered
+= max_t(int, acked
- tp
->sacked_out
, 1);
1915 if (acked
- 1 >= tp
->sacked_out
)
1918 tp
->sacked_out
-= acked
- 1;
1920 tcp_check_reno_reordering(sk
, acked
);
1921 tcp_verify_left_out(tp
);
1924 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1929 void tcp_clear_retrans(struct tcp_sock
*tp
)
1931 tp
->retrans_out
= 0;
1933 tp
->undo_marker
= 0;
1934 tp
->undo_retrans
= -1;
1938 static inline void tcp_init_undo(struct tcp_sock
*tp
)
1940 tp
->undo_marker
= tp
->snd_una
;
1941 /* Retransmission still in flight may cause DSACKs later. */
1942 tp
->undo_retrans
= tp
->retrans_out
? : -1;
1945 static bool tcp_is_rack(const struct sock
*sk
)
1947 return sock_net(sk
)->ipv4
.sysctl_tcp_recovery
& TCP_RACK_LOSS_DETECTION
;
1950 /* If we detect SACK reneging, forget all SACK information
1951 * and reset tags completely, otherwise preserve SACKs. If receiver
1952 * dropped its ofo queue, we will know this due to reneging detection.
1954 static void tcp_timeout_mark_lost(struct sock
*sk
)
1956 struct tcp_sock
*tp
= tcp_sk(sk
);
1957 struct sk_buff
*skb
, *head
;
1958 bool is_reneg
; /* is receiver reneging on SACKs? */
1960 head
= tcp_rtx_queue_head(sk
);
1961 is_reneg
= head
&& (TCP_SKB_CB(head
)->sacked
& TCPCB_SACKED_ACKED
);
1963 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPSACKRENEGING
);
1965 /* Mark SACK reneging until we recover from this loss event. */
1966 tp
->is_sack_reneg
= 1;
1967 } else if (tcp_is_reno(tp
)) {
1968 tcp_reset_reno_sack(tp
);
1972 skb_rbtree_walk_from(skb
) {
1974 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
1975 else if (tcp_is_rack(sk
) && skb
!= head
&&
1976 tcp_rack_skb_timeout(tp
, skb
, 0) > 0)
1977 continue; /* Don't mark recently sent ones lost yet */
1978 tcp_mark_skb_lost(sk
, skb
);
1980 tcp_verify_left_out(tp
);
1981 tcp_clear_all_retrans_hints(tp
);
1984 /* Enter Loss state. */
1985 void tcp_enter_loss(struct sock
*sk
)
1987 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1988 struct tcp_sock
*tp
= tcp_sk(sk
);
1989 struct net
*net
= sock_net(sk
);
1990 bool new_recovery
= icsk
->icsk_ca_state
< TCP_CA_Recovery
;
1992 tcp_timeout_mark_lost(sk
);
1994 /* Reduce ssthresh if it has not yet been made inside this window. */
1995 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
||
1996 !after(tp
->high_seq
, tp
->snd_una
) ||
1997 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
1998 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1999 tp
->prior_cwnd
= tp
->snd_cwnd
;
2000 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
2001 tcp_ca_event(sk
, CA_EVENT_LOSS
);
2004 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + 1;
2005 tp
->snd_cwnd_cnt
= 0;
2006 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
2008 /* Timeout in disordered state after receiving substantial DUPACKs
2009 * suggests that the degree of reordering is over-estimated.
2011 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
&&
2012 tp
->sacked_out
>= net
->ipv4
.sysctl_tcp_reordering
)
2013 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
2014 net
->ipv4
.sysctl_tcp_reordering
);
2015 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2016 tp
->high_seq
= tp
->snd_nxt
;
2017 tcp_ecn_queue_cwr(tp
);
2019 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
2020 * loss recovery is underway except recurring timeout(s) on
2021 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
2023 tp
->frto
= net
->ipv4
.sysctl_tcp_frto
&&
2024 (new_recovery
|| icsk
->icsk_retransmits
) &&
2025 !inet_csk(sk
)->icsk_mtup
.probe_size
;
2028 /* If ACK arrived pointing to a remembered SACK, it means that our
2029 * remembered SACKs do not reflect real state of receiver i.e.
2030 * receiver _host_ is heavily congested (or buggy).
2032 * To avoid big spurious retransmission bursts due to transient SACK
2033 * scoreboard oddities that look like reneging, we give the receiver a
2034 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2035 * restore sanity to the SACK scoreboard. If the apparent reneging
2036 * persists until this RTO then we'll clear the SACK scoreboard.
2038 static bool tcp_check_sack_reneging(struct sock
*sk
, int flag
)
2040 if (flag
& FLAG_SACK_RENEGING
) {
2041 struct tcp_sock
*tp
= tcp_sk(sk
);
2042 unsigned long delay
= max(usecs_to_jiffies(tp
->srtt_us
>> 4),
2043 msecs_to_jiffies(10));
2045 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2046 delay
, TCP_RTO_MAX
);
2052 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2053 * counter when SACK is enabled (without SACK, sacked_out is used for
2056 * With reordering, holes may still be in flight, so RFC3517 recovery
2057 * uses pure sacked_out (total number of SACKed segments) even though
2058 * it violates the RFC that uses duplicate ACKs, often these are equal
2059 * but when e.g. out-of-window ACKs or packet duplication occurs,
2060 * they differ. Since neither occurs due to loss, TCP should really
2063 static inline int tcp_dupack_heuristics(const struct tcp_sock
*tp
)
2065 return tp
->sacked_out
+ 1;
2068 /* Linux NewReno/SACK/ECN state machine.
2069 * --------------------------------------
2071 * "Open" Normal state, no dubious events, fast path.
2072 * "Disorder" In all the respects it is "Open",
2073 * but requires a bit more attention. It is entered when
2074 * we see some SACKs or dupacks. It is split of "Open"
2075 * mainly to move some processing from fast path to slow one.
2076 * "CWR" CWND was reduced due to some Congestion Notification event.
2077 * It can be ECN, ICMP source quench, local device congestion.
2078 * "Recovery" CWND was reduced, we are fast-retransmitting.
2079 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2081 * tcp_fastretrans_alert() is entered:
2082 * - each incoming ACK, if state is not "Open"
2083 * - when arrived ACK is unusual, namely:
2088 * Counting packets in flight is pretty simple.
2090 * in_flight = packets_out - left_out + retrans_out
2092 * packets_out is SND.NXT-SND.UNA counted in packets.
2094 * retrans_out is number of retransmitted segments.
2096 * left_out is number of segments left network, but not ACKed yet.
2098 * left_out = sacked_out + lost_out
2100 * sacked_out: Packets, which arrived to receiver out of order
2101 * and hence not ACKed. With SACKs this number is simply
2102 * amount of SACKed data. Even without SACKs
2103 * it is easy to give pretty reliable estimate of this number,
2104 * counting duplicate ACKs.
2106 * lost_out: Packets lost by network. TCP has no explicit
2107 * "loss notification" feedback from network (for now).
2108 * It means that this number can be only _guessed_.
2109 * Actually, it is the heuristics to predict lossage that
2110 * distinguishes different algorithms.
2112 * F.e. after RTO, when all the queue is considered as lost,
2113 * lost_out = packets_out and in_flight = retrans_out.
2115 * Essentially, we have now a few algorithms detecting
2118 * If the receiver supports SACK:
2120 * RFC6675/3517: It is the conventional algorithm. A packet is
2121 * considered lost if the number of higher sequence packets
2122 * SACKed is greater than or equal the DUPACK thoreshold
2123 * (reordering). This is implemented in tcp_mark_head_lost and
2124 * tcp_update_scoreboard.
2126 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm
2127 * (2017-) that checks timing instead of counting DUPACKs.
2128 * Essentially a packet is considered lost if it's not S/ACKed
2129 * after RTT + reordering_window, where both metrics are
2130 * dynamically measured and adjusted. This is implemented in
2131 * tcp_rack_mark_lost.
2133 * If the receiver does not support SACK:
2135 * NewReno (RFC6582): in Recovery we assume that one segment
2136 * is lost (classic Reno). While we are in Recovery and
2137 * a partial ACK arrives, we assume that one more packet
2138 * is lost (NewReno). This heuristics are the same in NewReno
2141 * Really tricky (and requiring careful tuning) part of algorithm
2142 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2143 * The first determines the moment _when_ we should reduce CWND and,
2144 * hence, slow down forward transmission. In fact, it determines the moment
2145 * when we decide that hole is caused by loss, rather than by a reorder.
2147 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2148 * holes, caused by lost packets.
2150 * And the most logically complicated part of algorithm is undo
2151 * heuristics. We detect false retransmits due to both too early
2152 * fast retransmit (reordering) and underestimated RTO, analyzing
2153 * timestamps and D-SACKs. When we detect that some segments were
2154 * retransmitted by mistake and CWND reduction was wrong, we undo
2155 * window reduction and abort recovery phase. This logic is hidden
2156 * inside several functions named tcp_try_undo_<something>.
2159 /* This function decides, when we should leave Disordered state
2160 * and enter Recovery phase, reducing congestion window.
2162 * Main question: may we further continue forward transmission
2163 * with the same cwnd?
2165 static bool tcp_time_to_recover(struct sock
*sk
, int flag
)
2167 struct tcp_sock
*tp
= tcp_sk(sk
);
2169 /* Trick#1: The loss is proven. */
2173 /* Not-A-Trick#2 : Classic rule... */
2174 if (!tcp_is_rack(sk
) && tcp_dupack_heuristics(tp
) > tp
->reordering
)
2180 /* Detect loss in event "A" above by marking head of queue up as lost.
2181 * For non-SACK(Reno) senders, the first "packets" number of segments
2182 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2183 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2184 * the maximum SACKed segments to pass before reaching this limit.
2186 static void tcp_mark_head_lost(struct sock
*sk
, int packets
, int mark_head
)
2188 struct tcp_sock
*tp
= tcp_sk(sk
);
2189 struct sk_buff
*skb
;
2190 int cnt
, oldcnt
, lost
;
2192 /* Use SACK to deduce losses of new sequences sent during recovery */
2193 const u32 loss_high
= tcp_is_sack(tp
) ? tp
->snd_nxt
: tp
->high_seq
;
2195 WARN_ON(packets
> tp
->packets_out
);
2196 skb
= tp
->lost_skb_hint
;
2198 /* Head already handled? */
2199 if (mark_head
&& after(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
))
2201 cnt
= tp
->lost_cnt_hint
;
2203 skb
= tcp_rtx_queue_head(sk
);
2207 skb_rbtree_walk_from(skb
) {
2208 /* TODO: do this better */
2209 /* this is not the most efficient way to do this... */
2210 tp
->lost_skb_hint
= skb
;
2211 tp
->lost_cnt_hint
= cnt
;
2213 if (after(TCP_SKB_CB(skb
)->end_seq
, loss_high
))
2217 if (tcp_is_reno(tp
) ||
2218 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2219 cnt
+= tcp_skb_pcount(skb
);
2221 if (cnt
> packets
) {
2222 if (tcp_is_sack(tp
) ||
2223 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
) ||
2224 (oldcnt
>= packets
))
2227 mss
= tcp_skb_mss(skb
);
2228 /* If needed, chop off the prefix to mark as lost. */
2229 lost
= (packets
- oldcnt
) * mss
;
2230 if (lost
< skb
->len
&&
2231 tcp_fragment(sk
, TCP_FRAG_IN_RTX_QUEUE
, skb
,
2232 lost
, mss
, GFP_ATOMIC
) < 0)
2237 tcp_skb_mark_lost(tp
, skb
);
2242 tcp_verify_left_out(tp
);
2245 /* Account newly detected lost packet(s) */
2247 static void tcp_update_scoreboard(struct sock
*sk
, int fast_rexmit
)
2249 struct tcp_sock
*tp
= tcp_sk(sk
);
2251 if (tcp_is_sack(tp
)) {
2252 int sacked_upto
= tp
->sacked_out
- tp
->reordering
;
2253 if (sacked_upto
>= 0)
2254 tcp_mark_head_lost(sk
, sacked_upto
, 0);
2255 else if (fast_rexmit
)
2256 tcp_mark_head_lost(sk
, 1, 1);
2260 static bool tcp_tsopt_ecr_before(const struct tcp_sock
*tp
, u32 when
)
2262 return tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2263 before(tp
->rx_opt
.rcv_tsecr
, when
);
2266 /* skb is spurious retransmitted if the returned timestamp echo
2267 * reply is prior to the skb transmission time
2269 static bool tcp_skb_spurious_retrans(const struct tcp_sock
*tp
,
2270 const struct sk_buff
*skb
)
2272 return (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
) &&
2273 tcp_tsopt_ecr_before(tp
, tcp_skb_timestamp(skb
));
2276 /* Nothing was retransmitted or returned timestamp is less
2277 * than timestamp of the first retransmission.
2279 static inline bool tcp_packet_delayed(const struct tcp_sock
*tp
)
2281 return tp
->retrans_stamp
&&
2282 tcp_tsopt_ecr_before(tp
, tp
->retrans_stamp
);
2285 /* Undo procedures. */
2287 /* We can clear retrans_stamp when there are no retransmissions in the
2288 * window. It would seem that it is trivially available for us in
2289 * tp->retrans_out, however, that kind of assumptions doesn't consider
2290 * what will happen if errors occur when sending retransmission for the
2291 * second time. ...It could the that such segment has only
2292 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2293 * the head skb is enough except for some reneging corner cases that
2294 * are not worth the effort.
2296 * Main reason for all this complexity is the fact that connection dying
2297 * time now depends on the validity of the retrans_stamp, in particular,
2298 * that successive retransmissions of a segment must not advance
2299 * retrans_stamp under any conditions.
2301 static bool tcp_any_retrans_done(const struct sock
*sk
)
2303 const struct tcp_sock
*tp
= tcp_sk(sk
);
2304 struct sk_buff
*skb
;
2306 if (tp
->retrans_out
)
2309 skb
= tcp_rtx_queue_head(sk
);
2310 if (unlikely(skb
&& TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
))
2316 static void DBGUNDO(struct sock
*sk
, const char *msg
)
2318 #if FASTRETRANS_DEBUG > 1
2319 struct tcp_sock
*tp
= tcp_sk(sk
);
2320 struct inet_sock
*inet
= inet_sk(sk
);
2322 if (sk
->sk_family
== AF_INET
) {
2323 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2325 &inet
->inet_daddr
, ntohs(inet
->inet_dport
),
2326 tp
->snd_cwnd
, tcp_left_out(tp
),
2327 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2330 #if IS_ENABLED(CONFIG_IPV6)
2331 else if (sk
->sk_family
== AF_INET6
) {
2332 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2334 &sk
->sk_v6_daddr
, ntohs(inet
->inet_dport
),
2335 tp
->snd_cwnd
, tcp_left_out(tp
),
2336 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2343 static void tcp_undo_cwnd_reduction(struct sock
*sk
, bool unmark_loss
)
2345 struct tcp_sock
*tp
= tcp_sk(sk
);
2348 struct sk_buff
*skb
;
2350 skb_rbtree_walk(skb
, &sk
->tcp_rtx_queue
) {
2351 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2354 tcp_clear_all_retrans_hints(tp
);
2357 if (tp
->prior_ssthresh
) {
2358 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2360 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
2362 if (tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
2363 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
2364 tcp_ecn_withdraw_cwr(tp
);
2367 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
2368 tp
->undo_marker
= 0;
2369 tp
->rack
.advanced
= 1; /* Force RACK to re-exam losses */
2372 static inline bool tcp_may_undo(const struct tcp_sock
*tp
)
2374 return tp
->undo_marker
&& (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
2377 /* People celebrate: "We love our President!" */
2378 static bool tcp_try_undo_recovery(struct sock
*sk
)
2380 struct tcp_sock
*tp
= tcp_sk(sk
);
2382 if (tcp_may_undo(tp
)) {
2385 /* Happy end! We did not retransmit anything
2386 * or our original transmission succeeded.
2388 DBGUNDO(sk
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
2389 tcp_undo_cwnd_reduction(sk
, false);
2390 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
2391 mib_idx
= LINUX_MIB_TCPLOSSUNDO
;
2393 mib_idx
= LINUX_MIB_TCPFULLUNDO
;
2395 NET_INC_STATS(sock_net(sk
), mib_idx
);
2396 } else if (tp
->rack
.reo_wnd_persist
) {
2397 tp
->rack
.reo_wnd_persist
--;
2399 if (tp
->snd_una
== tp
->high_seq
&& tcp_is_reno(tp
)) {
2400 /* Hold old state until something *above* high_seq
2401 * is ACKed. For Reno it is MUST to prevent false
2402 * fast retransmits (RFC2582). SACK TCP is safe. */
2403 if (!tcp_any_retrans_done(sk
))
2404 tp
->retrans_stamp
= 0;
2407 tcp_set_ca_state(sk
, TCP_CA_Open
);
2408 tp
->is_sack_reneg
= 0;
2412 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2413 static bool tcp_try_undo_dsack(struct sock
*sk
)
2415 struct tcp_sock
*tp
= tcp_sk(sk
);
2417 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
2418 tp
->rack
.reo_wnd_persist
= min(TCP_RACK_RECOVERY_THRESH
,
2419 tp
->rack
.reo_wnd_persist
+ 1);
2420 DBGUNDO(sk
, "D-SACK");
2421 tcp_undo_cwnd_reduction(sk
, false);
2422 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPDSACKUNDO
);
2428 /* Undo during loss recovery after partial ACK or using F-RTO. */
2429 static bool tcp_try_undo_loss(struct sock
*sk
, bool frto_undo
)
2431 struct tcp_sock
*tp
= tcp_sk(sk
);
2433 if (frto_undo
|| tcp_may_undo(tp
)) {
2434 tcp_undo_cwnd_reduction(sk
, true);
2436 DBGUNDO(sk
, "partial loss");
2437 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPLOSSUNDO
);
2439 NET_INC_STATS(sock_net(sk
),
2440 LINUX_MIB_TCPSPURIOUSRTOS
);
2441 inet_csk(sk
)->icsk_retransmits
= 0;
2442 if (frto_undo
|| tcp_is_sack(tp
)) {
2443 tcp_set_ca_state(sk
, TCP_CA_Open
);
2444 tp
->is_sack_reneg
= 0;
2451 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2452 * It computes the number of packets to send (sndcnt) based on packets newly
2454 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2455 * cwnd reductions across a full RTT.
2456 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2457 * But when the retransmits are acked without further losses, PRR
2458 * slow starts cwnd up to ssthresh to speed up the recovery.
2460 static void tcp_init_cwnd_reduction(struct sock
*sk
)
2462 struct tcp_sock
*tp
= tcp_sk(sk
);
2464 tp
->high_seq
= tp
->snd_nxt
;
2465 tp
->tlp_high_seq
= 0;
2466 tp
->snd_cwnd_cnt
= 0;
2467 tp
->prior_cwnd
= tp
->snd_cwnd
;
2468 tp
->prr_delivered
= 0;
2470 tp
->snd_ssthresh
= inet_csk(sk
)->icsk_ca_ops
->ssthresh(sk
);
2471 tcp_ecn_queue_cwr(tp
);
2474 void tcp_cwnd_reduction(struct sock
*sk
, int newly_acked_sacked
, int flag
)
2476 struct tcp_sock
*tp
= tcp_sk(sk
);
2478 int delta
= tp
->snd_ssthresh
- tcp_packets_in_flight(tp
);
2480 if (newly_acked_sacked
<= 0 || WARN_ON_ONCE(!tp
->prior_cwnd
))
2483 tp
->prr_delivered
+= newly_acked_sacked
;
2485 u64 dividend
= (u64
)tp
->snd_ssthresh
* tp
->prr_delivered
+
2487 sndcnt
= div_u64(dividend
, tp
->prior_cwnd
) - tp
->prr_out
;
2488 } else if ((flag
& (FLAG_RETRANS_DATA_ACKED
| FLAG_LOST_RETRANS
)) ==
2489 FLAG_RETRANS_DATA_ACKED
) {
2490 sndcnt
= min_t(int, delta
,
2491 max_t(int, tp
->prr_delivered
- tp
->prr_out
,
2492 newly_acked_sacked
) + 1);
2494 sndcnt
= min(delta
, newly_acked_sacked
);
2496 /* Force a fast retransmit upon entering fast recovery */
2497 sndcnt
= max(sndcnt
, (tp
->prr_out
? 0 : 1));
2498 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + sndcnt
;
2501 static inline void tcp_end_cwnd_reduction(struct sock
*sk
)
2503 struct tcp_sock
*tp
= tcp_sk(sk
);
2505 if (inet_csk(sk
)->icsk_ca_ops
->cong_control
)
2508 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2509 if (tp
->snd_ssthresh
< TCP_INFINITE_SSTHRESH
&&
2510 (inet_csk(sk
)->icsk_ca_state
== TCP_CA_CWR
|| tp
->undo_marker
)) {
2511 tp
->snd_cwnd
= tp
->snd_ssthresh
;
2512 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
2514 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
2517 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2518 void tcp_enter_cwr(struct sock
*sk
)
2520 struct tcp_sock
*tp
= tcp_sk(sk
);
2522 tp
->prior_ssthresh
= 0;
2523 if (inet_csk(sk
)->icsk_ca_state
< TCP_CA_CWR
) {
2524 tp
->undo_marker
= 0;
2525 tcp_init_cwnd_reduction(sk
);
2526 tcp_set_ca_state(sk
, TCP_CA_CWR
);
2529 EXPORT_SYMBOL(tcp_enter_cwr
);
2531 static void tcp_try_keep_open(struct sock
*sk
)
2533 struct tcp_sock
*tp
= tcp_sk(sk
);
2534 int state
= TCP_CA_Open
;
2536 if (tcp_left_out(tp
) || tcp_any_retrans_done(sk
))
2537 state
= TCP_CA_Disorder
;
2539 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
2540 tcp_set_ca_state(sk
, state
);
2541 tp
->high_seq
= tp
->snd_nxt
;
2545 static void tcp_try_to_open(struct sock
*sk
, int flag
)
2547 struct tcp_sock
*tp
= tcp_sk(sk
);
2549 tcp_verify_left_out(tp
);
2551 if (!tcp_any_retrans_done(sk
))
2552 tp
->retrans_stamp
= 0;
2554 if (flag
& FLAG_ECE
)
2557 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
2558 tcp_try_keep_open(sk
);
2562 static void tcp_mtup_probe_failed(struct sock
*sk
)
2564 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2566 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
2567 icsk
->icsk_mtup
.probe_size
= 0;
2568 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPMTUPFAIL
);
2571 static void tcp_mtup_probe_success(struct sock
*sk
)
2573 struct tcp_sock
*tp
= tcp_sk(sk
);
2574 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2576 /* FIXME: breaks with very large cwnd */
2577 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2578 tp
->snd_cwnd
= tp
->snd_cwnd
*
2579 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
2580 icsk
->icsk_mtup
.probe_size
;
2581 tp
->snd_cwnd_cnt
= 0;
2582 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
2583 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2585 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
2586 icsk
->icsk_mtup
.probe_size
= 0;
2587 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
2588 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPMTUPSUCCESS
);
2591 /* Do a simple retransmit without using the backoff mechanisms in
2592 * tcp_timer. This is used for path mtu discovery.
2593 * The socket is already locked here.
2595 void tcp_simple_retransmit(struct sock
*sk
)
2597 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2598 struct tcp_sock
*tp
= tcp_sk(sk
);
2599 struct sk_buff
*skb
;
2600 unsigned int mss
= tcp_current_mss(sk
);
2602 skb_rbtree_walk(skb
, &sk
->tcp_rtx_queue
) {
2603 if (tcp_skb_seglen(skb
) > mss
&&
2604 !(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
2605 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2606 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
2607 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2609 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
2613 tcp_clear_retrans_hints_partial(tp
);
2618 if (tcp_is_reno(tp
))
2619 tcp_limit_reno_sacked(tp
);
2621 tcp_verify_left_out(tp
);
2623 /* Don't muck with the congestion window here.
2624 * Reason is that we do not increase amount of _data_
2625 * in network, but units changed and effective
2626 * cwnd/ssthresh really reduced now.
2628 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
2629 tp
->high_seq
= tp
->snd_nxt
;
2630 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2631 tp
->prior_ssthresh
= 0;
2632 tp
->undo_marker
= 0;
2633 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2635 tcp_xmit_retransmit_queue(sk
);
2637 EXPORT_SYMBOL(tcp_simple_retransmit
);
2639 void tcp_enter_recovery(struct sock
*sk
, bool ece_ack
)
2641 struct tcp_sock
*tp
= tcp_sk(sk
);
2644 if (tcp_is_reno(tp
))
2645 mib_idx
= LINUX_MIB_TCPRENORECOVERY
;
2647 mib_idx
= LINUX_MIB_TCPSACKRECOVERY
;
2649 NET_INC_STATS(sock_net(sk
), mib_idx
);
2651 tp
->prior_ssthresh
= 0;
2654 if (!tcp_in_cwnd_reduction(sk
)) {
2656 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2657 tcp_init_cwnd_reduction(sk
);
2659 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
2662 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2663 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2665 static void tcp_process_loss(struct sock
*sk
, int flag
, int num_dupack
,
2668 struct tcp_sock
*tp
= tcp_sk(sk
);
2669 bool recovered
= !before(tp
->snd_una
, tp
->high_seq
);
2671 if ((flag
& FLAG_SND_UNA_ADVANCED
|| rcu_access_pointer(tp
->fastopen_rsk
)) &&
2672 tcp_try_undo_loss(sk
, false))
2675 if (tp
->frto
) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2676 /* Step 3.b. A timeout is spurious if not all data are
2677 * lost, i.e., never-retransmitted data are (s)acked.
2679 if ((flag
& FLAG_ORIG_SACK_ACKED
) &&
2680 tcp_try_undo_loss(sk
, true))
2683 if (after(tp
->snd_nxt
, tp
->high_seq
)) {
2684 if (flag
& FLAG_DATA_SACKED
|| num_dupack
)
2685 tp
->frto
= 0; /* Step 3.a. loss was real */
2686 } else if (flag
& FLAG_SND_UNA_ADVANCED
&& !recovered
) {
2687 tp
->high_seq
= tp
->snd_nxt
;
2688 /* Step 2.b. Try send new data (but deferred until cwnd
2689 * is updated in tcp_ack()). Otherwise fall back to
2690 * the conventional recovery.
2692 if (!tcp_write_queue_empty(sk
) &&
2693 after(tcp_wnd_end(tp
), tp
->snd_nxt
)) {
2694 *rexmit
= REXMIT_NEW
;
2702 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2703 tcp_try_undo_recovery(sk
);
2706 if (tcp_is_reno(tp
)) {
2707 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2708 * delivered. Lower inflight to clock out (re)tranmissions.
2710 if (after(tp
->snd_nxt
, tp
->high_seq
) && num_dupack
)
2711 tcp_add_reno_sack(sk
, num_dupack
);
2712 else if (flag
& FLAG_SND_UNA_ADVANCED
)
2713 tcp_reset_reno_sack(tp
);
2715 *rexmit
= REXMIT_LOST
;
2718 /* Undo during fast recovery after partial ACK. */
2719 static bool tcp_try_undo_partial(struct sock
*sk
, u32 prior_snd_una
)
2721 struct tcp_sock
*tp
= tcp_sk(sk
);
2723 if (tp
->undo_marker
&& tcp_packet_delayed(tp
)) {
2724 /* Plain luck! Hole if filled with delayed
2725 * packet, rather than with a retransmit. Check reordering.
2727 tcp_check_sack_reordering(sk
, prior_snd_una
, 1);
2729 /* We are getting evidence that the reordering degree is higher
2730 * than we realized. If there are no retransmits out then we
2731 * can undo. Otherwise we clock out new packets but do not
2732 * mark more packets lost or retransmit more.
2734 if (tp
->retrans_out
)
2737 if (!tcp_any_retrans_done(sk
))
2738 tp
->retrans_stamp
= 0;
2740 DBGUNDO(sk
, "partial recovery");
2741 tcp_undo_cwnd_reduction(sk
, true);
2742 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPPARTIALUNDO
);
2743 tcp_try_keep_open(sk
);
2749 static void tcp_identify_packet_loss(struct sock
*sk
, int *ack_flag
)
2751 struct tcp_sock
*tp
= tcp_sk(sk
);
2753 if (tcp_rtx_queue_empty(sk
))
2756 if (unlikely(tcp_is_reno(tp
))) {
2757 tcp_newreno_mark_lost(sk
, *ack_flag
& FLAG_SND_UNA_ADVANCED
);
2758 } else if (tcp_is_rack(sk
)) {
2759 u32 prior_retrans
= tp
->retrans_out
;
2761 tcp_rack_mark_lost(sk
);
2762 if (prior_retrans
> tp
->retrans_out
)
2763 *ack_flag
|= FLAG_LOST_RETRANS
;
2767 static bool tcp_force_fast_retransmit(struct sock
*sk
)
2769 struct tcp_sock
*tp
= tcp_sk(sk
);
2771 return after(tcp_highest_sack_seq(tp
),
2772 tp
->snd_una
+ tp
->reordering
* tp
->mss_cache
);
2775 /* Process an event, which can update packets-in-flight not trivially.
2776 * Main goal of this function is to calculate new estimate for left_out,
2777 * taking into account both packets sitting in receiver's buffer and
2778 * packets lost by network.
2780 * Besides that it updates the congestion state when packet loss or ECN
2781 * is detected. But it does not reduce the cwnd, it is done by the
2782 * congestion control later.
2784 * It does _not_ decide what to send, it is made in function
2785 * tcp_xmit_retransmit_queue().
2787 static void tcp_fastretrans_alert(struct sock
*sk
, const u32 prior_snd_una
,
2788 int num_dupack
, int *ack_flag
, int *rexmit
)
2790 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2791 struct tcp_sock
*tp
= tcp_sk(sk
);
2792 int fast_rexmit
= 0, flag
= *ack_flag
;
2793 bool do_lost
= num_dupack
|| ((flag
& FLAG_DATA_SACKED
) &&
2794 tcp_force_fast_retransmit(sk
));
2796 if (!tp
->packets_out
&& tp
->sacked_out
)
2799 /* Now state machine starts.
2800 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2801 if (flag
& FLAG_ECE
)
2802 tp
->prior_ssthresh
= 0;
2804 /* B. In all the states check for reneging SACKs. */
2805 if (tcp_check_sack_reneging(sk
, flag
))
2808 /* C. Check consistency of the current state. */
2809 tcp_verify_left_out(tp
);
2811 /* D. Check state exit conditions. State can be terminated
2812 * when high_seq is ACKed. */
2813 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
2814 WARN_ON(tp
->retrans_out
!= 0);
2815 tp
->retrans_stamp
= 0;
2816 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
2817 switch (icsk
->icsk_ca_state
) {
2819 /* CWR is to be held something *above* high_seq
2820 * is ACKed for CWR bit to reach receiver. */
2821 if (tp
->snd_una
!= tp
->high_seq
) {
2822 tcp_end_cwnd_reduction(sk
);
2823 tcp_set_ca_state(sk
, TCP_CA_Open
);
2827 case TCP_CA_Recovery
:
2828 if (tcp_is_reno(tp
))
2829 tcp_reset_reno_sack(tp
);
2830 if (tcp_try_undo_recovery(sk
))
2832 tcp_end_cwnd_reduction(sk
);
2837 /* E. Process state. */
2838 switch (icsk
->icsk_ca_state
) {
2839 case TCP_CA_Recovery
:
2840 if (!(flag
& FLAG_SND_UNA_ADVANCED
)) {
2841 if (tcp_is_reno(tp
))
2842 tcp_add_reno_sack(sk
, num_dupack
);
2844 if (tcp_try_undo_partial(sk
, prior_snd_una
))
2846 /* Partial ACK arrived. Force fast retransmit. */
2847 do_lost
= tcp_is_reno(tp
) ||
2848 tcp_force_fast_retransmit(sk
);
2850 if (tcp_try_undo_dsack(sk
)) {
2851 tcp_try_keep_open(sk
);
2854 tcp_identify_packet_loss(sk
, ack_flag
);
2857 tcp_process_loss(sk
, flag
, num_dupack
, rexmit
);
2858 tcp_identify_packet_loss(sk
, ack_flag
);
2859 if (!(icsk
->icsk_ca_state
== TCP_CA_Open
||
2860 (*ack_flag
& FLAG_LOST_RETRANS
)))
2862 /* Change state if cwnd is undone or retransmits are lost */
2865 if (tcp_is_reno(tp
)) {
2866 if (flag
& FLAG_SND_UNA_ADVANCED
)
2867 tcp_reset_reno_sack(tp
);
2868 tcp_add_reno_sack(sk
, num_dupack
);
2871 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
)
2872 tcp_try_undo_dsack(sk
);
2874 tcp_identify_packet_loss(sk
, ack_flag
);
2875 if (!tcp_time_to_recover(sk
, flag
)) {
2876 tcp_try_to_open(sk
, flag
);
2880 /* MTU probe failure: don't reduce cwnd */
2881 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
2882 icsk
->icsk_mtup
.probe_size
&&
2883 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
2884 tcp_mtup_probe_failed(sk
);
2885 /* Restores the reduction we did in tcp_mtup_probe() */
2887 tcp_simple_retransmit(sk
);
2891 /* Otherwise enter Recovery state */
2892 tcp_enter_recovery(sk
, (flag
& FLAG_ECE
));
2896 if (!tcp_is_rack(sk
) && do_lost
)
2897 tcp_update_scoreboard(sk
, fast_rexmit
);
2898 *rexmit
= REXMIT_LOST
;
2901 static void tcp_update_rtt_min(struct sock
*sk
, u32 rtt_us
, const int flag
)
2903 u32 wlen
= sock_net(sk
)->ipv4
.sysctl_tcp_min_rtt_wlen
* HZ
;
2904 struct tcp_sock
*tp
= tcp_sk(sk
);
2906 if ((flag
& FLAG_ACK_MAYBE_DELAYED
) && rtt_us
> tcp_min_rtt(tp
)) {
2907 /* If the remote keeps returning delayed ACKs, eventually
2908 * the min filter would pick it up and overestimate the
2909 * prop. delay when it expires. Skip suspected delayed ACKs.
2913 minmax_running_min(&tp
->rtt_min
, wlen
, tcp_jiffies32
,
2914 rtt_us
? : jiffies_to_usecs(1));
2917 static bool tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
2918 long seq_rtt_us
, long sack_rtt_us
,
2919 long ca_rtt_us
, struct rate_sample
*rs
)
2921 const struct tcp_sock
*tp
= tcp_sk(sk
);
2923 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2924 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2925 * Karn's algorithm forbids taking RTT if some retransmitted data
2926 * is acked (RFC6298).
2929 seq_rtt_us
= sack_rtt_us
;
2931 /* RTTM Rule: A TSecr value received in a segment is used to
2932 * update the averaged RTT measurement only if the segment
2933 * acknowledges some new data, i.e., only if it advances the
2934 * left edge of the send window.
2935 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2937 if (seq_rtt_us
< 0 && tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2938 flag
& FLAG_ACKED
) {
2939 u32 delta
= tcp_time_stamp(tp
) - tp
->rx_opt
.rcv_tsecr
;
2941 if (likely(delta
< INT_MAX
/ (USEC_PER_SEC
/ TCP_TS_HZ
))) {
2942 seq_rtt_us
= delta
* (USEC_PER_SEC
/ TCP_TS_HZ
);
2943 ca_rtt_us
= seq_rtt_us
;
2946 rs
->rtt_us
= ca_rtt_us
; /* RTT of last (S)ACKed packet (or -1) */
2950 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
2951 * always taken together with ACK, SACK, or TS-opts. Any negative
2952 * values will be skipped with the seq_rtt_us < 0 check above.
2954 tcp_update_rtt_min(sk
, ca_rtt_us
, flag
);
2955 tcp_rtt_estimator(sk
, seq_rtt_us
);
2958 /* RFC6298: only reset backoff on valid RTT measurement. */
2959 inet_csk(sk
)->icsk_backoff
= 0;
2963 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2964 void tcp_synack_rtt_meas(struct sock
*sk
, struct request_sock
*req
)
2966 struct rate_sample rs
;
2969 if (req
&& !req
->num_retrans
&& tcp_rsk(req
)->snt_synack
)
2970 rtt_us
= tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req
)->snt_synack
);
2972 tcp_ack_update_rtt(sk
, FLAG_SYN_ACKED
, rtt_us
, -1L, rtt_us
, &rs
);
2976 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
, u32 acked
)
2978 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2980 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, acked
);
2981 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_jiffies32
;
2984 /* Restart timer after forward progress on connection.
2985 * RFC2988 recommends to restart timer to now+rto.
2987 void tcp_rearm_rto(struct sock
*sk
)
2989 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2990 struct tcp_sock
*tp
= tcp_sk(sk
);
2992 /* If the retrans timer is currently being used by Fast Open
2993 * for SYN-ACK retrans purpose, stay put.
2995 if (rcu_access_pointer(tp
->fastopen_rsk
))
2998 if (!tp
->packets_out
) {
2999 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
3001 u32 rto
= inet_csk(sk
)->icsk_rto
;
3002 /* Offset the time elapsed after installing regular RTO */
3003 if (icsk
->icsk_pending
== ICSK_TIME_REO_TIMEOUT
||
3004 icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
) {
3005 s64 delta_us
= tcp_rto_delta_us(sk
);
3006 /* delta_us may not be positive if the socket is locked
3007 * when the retrans timer fires and is rescheduled.
3009 rto
= usecs_to_jiffies(max_t(int, delta_us
, 1));
3011 tcp_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
, rto
,
3012 TCP_RTO_MAX
, tcp_rtx_queue_head(sk
));
3016 /* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */
3017 static void tcp_set_xmit_timer(struct sock
*sk
)
3019 if (!tcp_schedule_loss_probe(sk
, true))
3023 /* If we get here, the whole TSO packet has not been acked. */
3024 static u32
tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
)
3026 struct tcp_sock
*tp
= tcp_sk(sk
);
3029 BUG_ON(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
));
3031 packets_acked
= tcp_skb_pcount(skb
);
3032 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
3034 packets_acked
-= tcp_skb_pcount(skb
);
3036 if (packets_acked
) {
3037 BUG_ON(tcp_skb_pcount(skb
) == 0);
3038 BUG_ON(!before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
));
3041 return packets_acked
;
3044 static void tcp_ack_tstamp(struct sock
*sk
, struct sk_buff
*skb
,
3047 const struct skb_shared_info
*shinfo
;
3049 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3050 if (likely(!TCP_SKB_CB(skb
)->txstamp_ack
))
3053 shinfo
= skb_shinfo(skb
);
3054 if (!before(shinfo
->tskey
, prior_snd_una
) &&
3055 before(shinfo
->tskey
, tcp_sk(sk
)->snd_una
)) {
3056 tcp_skb_tsorted_save(skb
) {
3057 __skb_tstamp_tx(skb
, NULL
, sk
, SCM_TSTAMP_ACK
);
3058 } tcp_skb_tsorted_restore(skb
);
3062 /* Remove acknowledged frames from the retransmission queue. If our packet
3063 * is before the ack sequence we can discard it as it's confirmed to have
3064 * arrived at the other end.
3066 static int tcp_clean_rtx_queue(struct sock
*sk
, u32 prior_fack
,
3068 struct tcp_sacktag_state
*sack
)
3070 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3071 u64 first_ackt
, last_ackt
;
3072 struct tcp_sock
*tp
= tcp_sk(sk
);
3073 u32 prior_sacked
= tp
->sacked_out
;
3074 u32 reord
= tp
->snd_nxt
; /* lowest acked un-retx un-sacked seq */
3075 struct sk_buff
*skb
, *next
;
3076 bool fully_acked
= true;
3077 long sack_rtt_us
= -1L;
3078 long seq_rtt_us
= -1L;
3079 long ca_rtt_us
= -1L;
3081 u32 last_in_flight
= 0;
3087 for (skb
= skb_rb_first(&sk
->tcp_rtx_queue
); skb
; skb
= next
) {
3088 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
3089 const u32 start_seq
= scb
->seq
;
3090 u8 sacked
= scb
->sacked
;
3093 tcp_ack_tstamp(sk
, skb
, prior_snd_una
);
3095 /* Determine how many packets and what bytes were acked, tso and else */
3096 if (after(scb
->end_seq
, tp
->snd_una
)) {
3097 if (tcp_skb_pcount(skb
) == 1 ||
3098 !after(tp
->snd_una
, scb
->seq
))
3101 acked_pcount
= tcp_tso_acked(sk
, skb
);
3104 fully_acked
= false;
3106 acked_pcount
= tcp_skb_pcount(skb
);
3109 if (unlikely(sacked
& TCPCB_RETRANS
)) {
3110 if (sacked
& TCPCB_SACKED_RETRANS
)
3111 tp
->retrans_out
-= acked_pcount
;
3112 flag
|= FLAG_RETRANS_DATA_ACKED
;
3113 } else if (!(sacked
& TCPCB_SACKED_ACKED
)) {
3114 last_ackt
= tcp_skb_timestamp_us(skb
);
3115 WARN_ON_ONCE(last_ackt
== 0);
3117 first_ackt
= last_ackt
;
3119 last_in_flight
= TCP_SKB_CB(skb
)->tx
.in_flight
;
3120 if (before(start_seq
, reord
))
3122 if (!after(scb
->end_seq
, tp
->high_seq
))
3123 flag
|= FLAG_ORIG_SACK_ACKED
;
3126 if (sacked
& TCPCB_SACKED_ACKED
) {
3127 tp
->sacked_out
-= acked_pcount
;
3128 } else if (tcp_is_sack(tp
)) {
3129 tp
->delivered
+= acked_pcount
;
3130 if (!tcp_skb_spurious_retrans(tp
, skb
))
3131 tcp_rack_advance(tp
, sacked
, scb
->end_seq
,
3132 tcp_skb_timestamp_us(skb
));
3134 if (sacked
& TCPCB_LOST
)
3135 tp
->lost_out
-= acked_pcount
;
3137 tp
->packets_out
-= acked_pcount
;
3138 pkts_acked
+= acked_pcount
;
3139 tcp_rate_skb_delivered(sk
, skb
, sack
->rate
);
3141 /* Initial outgoing SYN's get put onto the write_queue
3142 * just like anything else we transmit. It is not
3143 * true data, and if we misinform our callers that
3144 * this ACK acks real data, we will erroneously exit
3145 * connection startup slow start one packet too
3146 * quickly. This is severely frowned upon behavior.
3148 if (likely(!(scb
->tcp_flags
& TCPHDR_SYN
))) {
3149 flag
|= FLAG_DATA_ACKED
;
3151 flag
|= FLAG_SYN_ACKED
;
3152 tp
->retrans_stamp
= 0;
3158 next
= skb_rb_next(skb
);
3159 if (unlikely(skb
== tp
->retransmit_skb_hint
))
3160 tp
->retransmit_skb_hint
= NULL
;
3161 if (unlikely(skb
== tp
->lost_skb_hint
))
3162 tp
->lost_skb_hint
= NULL
;
3163 tcp_rtx_queue_unlink_and_free(skb
, sk
);
3167 tcp_chrono_stop(sk
, TCP_CHRONO_BUSY
);
3169 if (likely(between(tp
->snd_up
, prior_snd_una
, tp
->snd_una
)))
3170 tp
->snd_up
= tp
->snd_una
;
3172 if (skb
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
3173 flag
|= FLAG_SACK_RENEGING
;
3175 if (likely(first_ackt
) && !(flag
& FLAG_RETRANS_DATA_ACKED
)) {
3176 seq_rtt_us
= tcp_stamp_us_delta(tp
->tcp_mstamp
, first_ackt
);
3177 ca_rtt_us
= tcp_stamp_us_delta(tp
->tcp_mstamp
, last_ackt
);
3179 if (pkts_acked
== 1 && last_in_flight
< tp
->mss_cache
&&
3180 last_in_flight
&& !prior_sacked
&& fully_acked
&&
3181 sack
->rate
->prior_delivered
+ 1 == tp
->delivered
&&
3182 !(flag
& (FLAG_CA_ALERT
| FLAG_SYN_ACKED
))) {
3183 /* Conservatively mark a delayed ACK. It's typically
3184 * from a lone runt packet over the round trip to
3185 * a receiver w/o out-of-order or CE events.
3187 flag
|= FLAG_ACK_MAYBE_DELAYED
;
3190 if (sack
->first_sackt
) {
3191 sack_rtt_us
= tcp_stamp_us_delta(tp
->tcp_mstamp
, sack
->first_sackt
);
3192 ca_rtt_us
= tcp_stamp_us_delta(tp
->tcp_mstamp
, sack
->last_sackt
);
3194 rtt_update
= tcp_ack_update_rtt(sk
, flag
, seq_rtt_us
, sack_rtt_us
,
3195 ca_rtt_us
, sack
->rate
);
3197 if (flag
& FLAG_ACKED
) {
3198 flag
|= FLAG_SET_XMIT_TIMER
; /* set TLP or RTO timer */
3199 if (unlikely(icsk
->icsk_mtup
.probe_size
&&
3200 !after(tp
->mtu_probe
.probe_seq_end
, tp
->snd_una
))) {
3201 tcp_mtup_probe_success(sk
);
3204 if (tcp_is_reno(tp
)) {
3205 tcp_remove_reno_sacks(sk
, pkts_acked
);
3207 /* If any of the cumulatively ACKed segments was
3208 * retransmitted, non-SACK case cannot confirm that
3209 * progress was due to original transmission due to
3210 * lack of TCPCB_SACKED_ACKED bits even if some of
3211 * the packets may have been never retransmitted.
3213 if (flag
& FLAG_RETRANS_DATA_ACKED
)
3214 flag
&= ~FLAG_ORIG_SACK_ACKED
;
3218 /* Non-retransmitted hole got filled? That's reordering */
3219 if (before(reord
, prior_fack
))
3220 tcp_check_sack_reordering(sk
, reord
, 0);
3222 delta
= prior_sacked
- tp
->sacked_out
;
3223 tp
->lost_cnt_hint
-= min(tp
->lost_cnt_hint
, delta
);
3225 } else if (skb
&& rtt_update
&& sack_rtt_us
>= 0 &&
3226 sack_rtt_us
> tcp_stamp_us_delta(tp
->tcp_mstamp
,
3227 tcp_skb_timestamp_us(skb
))) {
3228 /* Do not re-arm RTO if the sack RTT is measured from data sent
3229 * after when the head was last (re)transmitted. Otherwise the
3230 * timeout may continue to extend in loss recovery.
3232 flag
|= FLAG_SET_XMIT_TIMER
; /* set TLP or RTO timer */
3235 if (icsk
->icsk_ca_ops
->pkts_acked
) {
3236 struct ack_sample sample
= { .pkts_acked
= pkts_acked
,
3237 .rtt_us
= sack
->rate
->rtt_us
,
3238 .in_flight
= last_in_flight
};
3240 icsk
->icsk_ca_ops
->pkts_acked(sk
, &sample
);
3243 #if FASTRETRANS_DEBUG > 0
3244 WARN_ON((int)tp
->sacked_out
< 0);
3245 WARN_ON((int)tp
->lost_out
< 0);
3246 WARN_ON((int)tp
->retrans_out
< 0);
3247 if (!tp
->packets_out
&& tcp_is_sack(tp
)) {
3248 icsk
= inet_csk(sk
);
3250 pr_debug("Leak l=%u %d\n",
3251 tp
->lost_out
, icsk
->icsk_ca_state
);
3254 if (tp
->sacked_out
) {
3255 pr_debug("Leak s=%u %d\n",
3256 tp
->sacked_out
, icsk
->icsk_ca_state
);
3259 if (tp
->retrans_out
) {
3260 pr_debug("Leak r=%u %d\n",
3261 tp
->retrans_out
, icsk
->icsk_ca_state
);
3262 tp
->retrans_out
= 0;
3269 static void tcp_ack_probe(struct sock
*sk
)
3271 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3272 struct sk_buff
*head
= tcp_send_head(sk
);
3273 const struct tcp_sock
*tp
= tcp_sk(sk
);
3275 /* Was it a usable window open? */
3278 if (!after(TCP_SKB_CB(head
)->end_seq
, tcp_wnd_end(tp
))) {
3279 icsk
->icsk_backoff
= 0;
3280 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
3281 /* Socket must be waked up by subsequent tcp_data_snd_check().
3282 * This function is not for random using!
3285 unsigned long when
= tcp_probe0_when(sk
, TCP_RTO_MAX
);
3287 tcp_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3288 when
, TCP_RTO_MAX
, NULL
);
3292 static inline bool tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
3294 return !(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
3295 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
;
3298 /* Decide wheather to run the increase function of congestion control. */
3299 static inline bool tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
3301 /* If reordering is high then always grow cwnd whenever data is
3302 * delivered regardless of its ordering. Otherwise stay conservative
3303 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3304 * new SACK or ECE mark may first advance cwnd here and later reduce
3305 * cwnd in tcp_fastretrans_alert() based on more states.
3307 if (tcp_sk(sk
)->reordering
> sock_net(sk
)->ipv4
.sysctl_tcp_reordering
)
3308 return flag
& FLAG_FORWARD_PROGRESS
;
3310 return flag
& FLAG_DATA_ACKED
;
3313 /* The "ultimate" congestion control function that aims to replace the rigid
3314 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3315 * It's called toward the end of processing an ACK with precise rate
3316 * information. All transmission or retransmission are delayed afterwards.
3318 static void tcp_cong_control(struct sock
*sk
, u32 ack
, u32 acked_sacked
,
3319 int flag
, const struct rate_sample
*rs
)
3321 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3323 if (icsk
->icsk_ca_ops
->cong_control
) {
3324 icsk
->icsk_ca_ops
->cong_control(sk
, rs
);
3328 if (tcp_in_cwnd_reduction(sk
)) {
3329 /* Reduce cwnd if state mandates */
3330 tcp_cwnd_reduction(sk
, acked_sacked
, flag
);
3331 } else if (tcp_may_raise_cwnd(sk
, flag
)) {
3332 /* Advance cwnd if state allows */
3333 tcp_cong_avoid(sk
, ack
, acked_sacked
);
3335 tcp_update_pacing_rate(sk
);
3338 /* Check that window update is acceptable.
3339 * The function assumes that snd_una<=ack<=snd_next.
3341 static inline bool tcp_may_update_window(const struct tcp_sock
*tp
,
3342 const u32 ack
, const u32 ack_seq
,
3345 return after(ack
, tp
->snd_una
) ||
3346 after(ack_seq
, tp
->snd_wl1
) ||
3347 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
);
3350 /* If we update tp->snd_una, also update tp->bytes_acked */
3351 static void tcp_snd_una_update(struct tcp_sock
*tp
, u32 ack
)
3353 u32 delta
= ack
- tp
->snd_una
;
3355 sock_owned_by_me((struct sock
*)tp
);
3356 tp
->bytes_acked
+= delta
;
3360 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3361 static void tcp_rcv_nxt_update(struct tcp_sock
*tp
, u32 seq
)
3363 u32 delta
= seq
- tp
->rcv_nxt
;
3365 sock_owned_by_me((struct sock
*)tp
);
3366 tp
->bytes_received
+= delta
;
3367 WRITE_ONCE(tp
->rcv_nxt
, seq
);
3370 /* Update our send window.
3372 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3373 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3375 static int tcp_ack_update_window(struct sock
*sk
, const struct sk_buff
*skb
, u32 ack
,
3378 struct tcp_sock
*tp
= tcp_sk(sk
);
3380 u32 nwin
= ntohs(tcp_hdr(skb
)->window
);
3382 if (likely(!tcp_hdr(skb
)->syn
))
3383 nwin
<<= tp
->rx_opt
.snd_wscale
;
3385 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
3386 flag
|= FLAG_WIN_UPDATE
;
3387 tcp_update_wl(tp
, ack_seq
);
3389 if (tp
->snd_wnd
!= nwin
) {
3392 /* Note, it is the only place, where
3393 * fast path is recovered for sending TCP.
3396 tcp_fast_path_check(sk
);
3398 if (!tcp_write_queue_empty(sk
))
3399 tcp_slow_start_after_idle_check(sk
);
3401 if (nwin
> tp
->max_window
) {
3402 tp
->max_window
= nwin
;
3403 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
3408 tcp_snd_una_update(tp
, ack
);
3413 static bool __tcp_oow_rate_limited(struct net
*net
, int mib_idx
,
3414 u32
*last_oow_ack_time
)
3416 if (*last_oow_ack_time
) {
3417 s32 elapsed
= (s32
)(tcp_jiffies32
- *last_oow_ack_time
);
3419 if (0 <= elapsed
&& elapsed
< net
->ipv4
.sysctl_tcp_invalid_ratelimit
) {
3420 NET_INC_STATS(net
, mib_idx
);
3421 return true; /* rate-limited: don't send yet! */
3425 *last_oow_ack_time
= tcp_jiffies32
;
3427 return false; /* not rate-limited: go ahead, send dupack now! */
3430 /* Return true if we're currently rate-limiting out-of-window ACKs and
3431 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3432 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3433 * attacks that send repeated SYNs or ACKs for the same connection. To
3434 * do this, we do not send a duplicate SYNACK or ACK if the remote
3435 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3437 bool tcp_oow_rate_limited(struct net
*net
, const struct sk_buff
*skb
,
3438 int mib_idx
, u32
*last_oow_ack_time
)
3440 /* Data packets without SYNs are not likely part of an ACK loop. */
3441 if ((TCP_SKB_CB(skb
)->seq
!= TCP_SKB_CB(skb
)->end_seq
) &&
3445 return __tcp_oow_rate_limited(net
, mib_idx
, last_oow_ack_time
);
3448 /* RFC 5961 7 [ACK Throttling] */
3449 static void tcp_send_challenge_ack(struct sock
*sk
, const struct sk_buff
*skb
)
3451 /* unprotected vars, we dont care of overwrites */
3452 static u32 challenge_timestamp
;
3453 static unsigned int challenge_count
;
3454 struct tcp_sock
*tp
= tcp_sk(sk
);
3455 struct net
*net
= sock_net(sk
);
3458 /* First check our per-socket dupack rate limit. */
3459 if (__tcp_oow_rate_limited(net
,
3460 LINUX_MIB_TCPACKSKIPPEDCHALLENGE
,
3461 &tp
->last_oow_ack_time
))
3464 /* Then check host-wide RFC 5961 rate limit. */
3466 if (now
!= challenge_timestamp
) {
3467 u32 ack_limit
= net
->ipv4
.sysctl_tcp_challenge_ack_limit
;
3468 u32 half
= (ack_limit
+ 1) >> 1;
3470 challenge_timestamp
= now
;
3471 WRITE_ONCE(challenge_count
, half
+ prandom_u32_max(ack_limit
));
3473 count
= READ_ONCE(challenge_count
);
3475 WRITE_ONCE(challenge_count
, count
- 1);
3476 NET_INC_STATS(net
, LINUX_MIB_TCPCHALLENGEACK
);
3481 static void tcp_store_ts_recent(struct tcp_sock
*tp
)
3483 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
3484 tp
->rx_opt
.ts_recent_stamp
= ktime_get_seconds();
3487 static void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
3489 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
3490 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3491 * extra check below makes sure this can only happen
3492 * for pure ACK frames. -DaveM
3494 * Not only, also it occurs for expired timestamps.
3497 if (tcp_paws_check(&tp
->rx_opt
, 0))
3498 tcp_store_ts_recent(tp
);
3502 /* This routine deals with acks during a TLP episode.
3503 * We mark the end of a TLP episode on receiving TLP dupack or when
3504 * ack is after tlp_high_seq.
3505 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3507 static void tcp_process_tlp_ack(struct sock
*sk
, u32 ack
, int flag
)
3509 struct tcp_sock
*tp
= tcp_sk(sk
);
3511 if (before(ack
, tp
->tlp_high_seq
))
3514 if (flag
& FLAG_DSACKING_ACK
) {
3515 /* This DSACK means original and TLP probe arrived; no loss */
3516 tp
->tlp_high_seq
= 0;
3517 } else if (after(ack
, tp
->tlp_high_seq
)) {
3518 /* ACK advances: there was a loss, so reduce cwnd. Reset
3519 * tlp_high_seq in tcp_init_cwnd_reduction()
3521 tcp_init_cwnd_reduction(sk
);
3522 tcp_set_ca_state(sk
, TCP_CA_CWR
);
3523 tcp_end_cwnd_reduction(sk
);
3524 tcp_try_keep_open(sk
);
3525 NET_INC_STATS(sock_net(sk
),
3526 LINUX_MIB_TCPLOSSPROBERECOVERY
);
3527 } else if (!(flag
& (FLAG_SND_UNA_ADVANCED
|
3528 FLAG_NOT_DUP
| FLAG_DATA_SACKED
))) {
3529 /* Pure dupack: original and TLP probe arrived; no loss */
3530 tp
->tlp_high_seq
= 0;
3534 static inline void tcp_in_ack_event(struct sock
*sk
, u32 flags
)
3536 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3538 if (icsk
->icsk_ca_ops
->in_ack_event
)
3539 icsk
->icsk_ca_ops
->in_ack_event(sk
, flags
);
3542 /* Congestion control has updated the cwnd already. So if we're in
3543 * loss recovery then now we do any new sends (for FRTO) or
3544 * retransmits (for CA_Loss or CA_recovery) that make sense.
3546 static void tcp_xmit_recovery(struct sock
*sk
, int rexmit
)
3548 struct tcp_sock
*tp
= tcp_sk(sk
);
3550 if (rexmit
== REXMIT_NONE
|| sk
->sk_state
== TCP_SYN_SENT
)
3553 if (unlikely(rexmit
== 2)) {
3554 __tcp_push_pending_frames(sk
, tcp_current_mss(sk
),
3556 if (after(tp
->snd_nxt
, tp
->high_seq
))
3560 tcp_xmit_retransmit_queue(sk
);
3563 /* Returns the number of packets newly acked or sacked by the current ACK */
3564 static u32
tcp_newly_delivered(struct sock
*sk
, u32 prior_delivered
, int flag
)
3566 const struct net
*net
= sock_net(sk
);
3567 struct tcp_sock
*tp
= tcp_sk(sk
);
3570 delivered
= tp
->delivered
- prior_delivered
;
3571 NET_ADD_STATS(net
, LINUX_MIB_TCPDELIVERED
, delivered
);
3572 if (flag
& FLAG_ECE
) {
3573 tp
->delivered_ce
+= delivered
;
3574 NET_ADD_STATS(net
, LINUX_MIB_TCPDELIVEREDCE
, delivered
);
3579 /* This routine deals with incoming acks, but not outgoing ones. */
3580 static int tcp_ack(struct sock
*sk
, const struct sk_buff
*skb
, int flag
)
3582 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3583 struct tcp_sock
*tp
= tcp_sk(sk
);
3584 struct tcp_sacktag_state sack_state
;
3585 struct rate_sample rs
= { .prior_delivered
= 0 };
3586 u32 prior_snd_una
= tp
->snd_una
;
3587 bool is_sack_reneg
= tp
->is_sack_reneg
;
3588 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
3589 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3591 int prior_packets
= tp
->packets_out
;
3592 u32 delivered
= tp
->delivered
;
3593 u32 lost
= tp
->lost
;
3594 int rexmit
= REXMIT_NONE
; /* Flag to (re)transmit to recover losses */
3597 sack_state
.first_sackt
= 0;
3598 sack_state
.rate
= &rs
;
3600 /* We very likely will need to access rtx queue. */
3601 prefetch(sk
->tcp_rtx_queue
.rb_node
);
3603 /* If the ack is older than previous acks
3604 * then we can probably ignore it.
3606 if (before(ack
, prior_snd_una
)) {
3607 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3608 if (before(ack
, prior_snd_una
- tp
->max_window
)) {
3609 if (!(flag
& FLAG_NO_CHALLENGE_ACK
))
3610 tcp_send_challenge_ack(sk
, skb
);
3616 /* If the ack includes data we haven't sent yet, discard
3617 * this segment (RFC793 Section 3.9).
3619 if (after(ack
, tp
->snd_nxt
))
3622 if (after(ack
, prior_snd_una
)) {
3623 flag
|= FLAG_SND_UNA_ADVANCED
;
3624 icsk
->icsk_retransmits
= 0;
3626 #if IS_ENABLED(CONFIG_TLS_DEVICE)
3627 if (static_branch_unlikely(&clean_acked_data_enabled
.key
))
3628 if (icsk
->icsk_clean_acked
)
3629 icsk
->icsk_clean_acked(sk
, ack
);
3633 prior_fack
= tcp_is_sack(tp
) ? tcp_highest_sack_seq(tp
) : tp
->snd_una
;
3634 rs
.prior_in_flight
= tcp_packets_in_flight(tp
);
3636 /* ts_recent update must be made after we are sure that the packet
3639 if (flag
& FLAG_UPDATE_TS_RECENT
)
3640 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
3642 if ((flag
& (FLAG_SLOWPATH
| FLAG_SND_UNA_ADVANCED
)) ==
3643 FLAG_SND_UNA_ADVANCED
) {
3644 /* Window is constant, pure forward advance.
3645 * No more checks are required.
3646 * Note, we use the fact that SND.UNA>=SND.WL2.
3648 tcp_update_wl(tp
, ack_seq
);
3649 tcp_snd_una_update(tp
, ack
);
3650 flag
|= FLAG_WIN_UPDATE
;
3652 tcp_in_ack_event(sk
, CA_ACK_WIN_UPDATE
);
3654 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPHPACKS
);
3656 u32 ack_ev_flags
= CA_ACK_SLOWPATH
;
3658 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
3661 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPPUREACKS
);
3663 flag
|= tcp_ack_update_window(sk
, skb
, ack
, ack_seq
);
3665 if (TCP_SKB_CB(skb
)->sacked
)
3666 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
,
3669 if (tcp_ecn_rcv_ecn_echo(tp
, tcp_hdr(skb
))) {
3671 ack_ev_flags
|= CA_ACK_ECE
;
3674 if (flag
& FLAG_WIN_UPDATE
)
3675 ack_ev_flags
|= CA_ACK_WIN_UPDATE
;
3677 tcp_in_ack_event(sk
, ack_ev_flags
);
3680 /* We passed data and got it acked, remove any soft error
3681 * log. Something worked...
3683 sk
->sk_err_soft
= 0;
3684 icsk
->icsk_probes_out
= 0;
3685 tp
->rcv_tstamp
= tcp_jiffies32
;
3689 /* See if we can take anything off of the retransmit queue. */
3690 flag
|= tcp_clean_rtx_queue(sk
, prior_fack
, prior_snd_una
, &sack_state
);
3692 tcp_rack_update_reo_wnd(sk
, &rs
);
3694 if (tp
->tlp_high_seq
)
3695 tcp_process_tlp_ack(sk
, ack
, flag
);
3696 /* If needed, reset TLP/RTO timer; RACK may later override this. */
3697 if (flag
& FLAG_SET_XMIT_TIMER
)
3698 tcp_set_xmit_timer(sk
);
3700 if (tcp_ack_is_dubious(sk
, flag
)) {
3701 if (!(flag
& (FLAG_SND_UNA_ADVANCED
| FLAG_NOT_DUP
))) {
3703 /* Consider if pure acks were aggregated in tcp_add_backlog() */
3704 if (!(flag
& FLAG_DATA
))
3705 num_dupack
= max_t(u16
, 1, skb_shinfo(skb
)->gso_segs
);
3707 tcp_fastretrans_alert(sk
, prior_snd_una
, num_dupack
, &flag
,
3711 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
& FLAG_NOT_DUP
))
3714 delivered
= tcp_newly_delivered(sk
, delivered
, flag
);
3715 lost
= tp
->lost
- lost
; /* freshly marked lost */
3716 rs
.is_ack_delayed
= !!(flag
& FLAG_ACK_MAYBE_DELAYED
);
3717 tcp_rate_gen(sk
, delivered
, lost
, is_sack_reneg
, sack_state
.rate
);
3718 tcp_cong_control(sk
, ack
, delivered
, flag
, sack_state
.rate
);
3719 tcp_xmit_recovery(sk
, rexmit
);
3723 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3724 if (flag
& FLAG_DSACKING_ACK
) {
3725 tcp_fastretrans_alert(sk
, prior_snd_una
, num_dupack
, &flag
,
3727 tcp_newly_delivered(sk
, delivered
, flag
);
3729 /* If this ack opens up a zero window, clear backoff. It was
3730 * being used to time the probes, and is probably far higher than
3731 * it needs to be for normal retransmission.
3735 if (tp
->tlp_high_seq
)
3736 tcp_process_tlp_ack(sk
, ack
, flag
);
3740 /* If data was SACKed, tag it and see if we should send more data.
3741 * If data was DSACKed, see if we can undo a cwnd reduction.
3743 if (TCP_SKB_CB(skb
)->sacked
) {
3744 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
,
3746 tcp_fastretrans_alert(sk
, prior_snd_una
, num_dupack
, &flag
,
3748 tcp_newly_delivered(sk
, delivered
, flag
);
3749 tcp_xmit_recovery(sk
, rexmit
);
3755 static void tcp_parse_fastopen_option(int len
, const unsigned char *cookie
,
3756 bool syn
, struct tcp_fastopen_cookie
*foc
,
3759 /* Valid only in SYN or SYN-ACK with an even length. */
3760 if (!foc
|| !syn
|| len
< 0 || (len
& 1))
3763 if (len
>= TCP_FASTOPEN_COOKIE_MIN
&&
3764 len
<= TCP_FASTOPEN_COOKIE_MAX
)
3765 memcpy(foc
->val
, cookie
, len
);
3772 static void smc_parse_options(const struct tcphdr
*th
,
3773 struct tcp_options_received
*opt_rx
,
3774 const unsigned char *ptr
,
3777 #if IS_ENABLED(CONFIG_SMC)
3778 if (static_branch_unlikely(&tcp_have_smc
)) {
3779 if (th
->syn
&& !(opsize
& 1) &&
3780 opsize
>= TCPOLEN_EXP_SMC_BASE
&&
3781 get_unaligned_be32(ptr
) == TCPOPT_SMC_MAGIC
)
3787 /* Try to parse the MSS option from the TCP header. Return 0 on failure, clamped
3790 static u16
tcp_parse_mss_option(const struct tcphdr
*th
, u16 user_mss
)
3792 const unsigned char *ptr
= (const unsigned char *)(th
+ 1);
3793 int length
= (th
->doff
* 4) - sizeof(struct tcphdr
);
3796 while (length
> 0) {
3797 int opcode
= *ptr
++;
3803 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
3810 if (opsize
< 2) /* "silly options" */
3812 if (opsize
> length
)
3813 return mss
; /* fail on partial options */
3814 if (opcode
== TCPOPT_MSS
&& opsize
== TCPOLEN_MSS
) {
3815 u16 in_mss
= get_unaligned_be16(ptr
);
3818 if (user_mss
&& user_mss
< in_mss
)
3830 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3831 * But, this can also be called on packets in the established flow when
3832 * the fast version below fails.
3834 void tcp_parse_options(const struct net
*net
,
3835 const struct sk_buff
*skb
,
3836 struct tcp_options_received
*opt_rx
, int estab
,
3837 struct tcp_fastopen_cookie
*foc
)
3839 const unsigned char *ptr
;
3840 const struct tcphdr
*th
= tcp_hdr(skb
);
3841 int length
= (th
->doff
* 4) - sizeof(struct tcphdr
);
3843 ptr
= (const unsigned char *)(th
+ 1);
3844 opt_rx
->saw_tstamp
= 0;
3846 while (length
> 0) {
3847 int opcode
= *ptr
++;
3853 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
3860 if (opsize
< 2) /* "silly options" */
3862 if (opsize
> length
)
3863 return; /* don't parse partial options */
3866 if (opsize
== TCPOLEN_MSS
&& th
->syn
&& !estab
) {
3867 u16 in_mss
= get_unaligned_be16(ptr
);
3869 if (opt_rx
->user_mss
&&
3870 opt_rx
->user_mss
< in_mss
)
3871 in_mss
= opt_rx
->user_mss
;
3872 opt_rx
->mss_clamp
= in_mss
;
3877 if (opsize
== TCPOLEN_WINDOW
&& th
->syn
&&
3878 !estab
&& net
->ipv4
.sysctl_tcp_window_scaling
) {
3879 __u8 snd_wscale
= *(__u8
*)ptr
;
3880 opt_rx
->wscale_ok
= 1;
3881 if (snd_wscale
> TCP_MAX_WSCALE
) {
3882 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n",
3886 snd_wscale
= TCP_MAX_WSCALE
;
3888 opt_rx
->snd_wscale
= snd_wscale
;
3891 case TCPOPT_TIMESTAMP
:
3892 if ((opsize
== TCPOLEN_TIMESTAMP
) &&
3893 ((estab
&& opt_rx
->tstamp_ok
) ||
3894 (!estab
&& net
->ipv4
.sysctl_tcp_timestamps
))) {
3895 opt_rx
->saw_tstamp
= 1;
3896 opt_rx
->rcv_tsval
= get_unaligned_be32(ptr
);
3897 opt_rx
->rcv_tsecr
= get_unaligned_be32(ptr
+ 4);
3900 case TCPOPT_SACK_PERM
:
3901 if (opsize
== TCPOLEN_SACK_PERM
&& th
->syn
&&
3902 !estab
&& net
->ipv4
.sysctl_tcp_sack
) {
3903 opt_rx
->sack_ok
= TCP_SACK_SEEN
;
3904 tcp_sack_reset(opt_rx
);
3909 if ((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
3910 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
3912 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
3915 #ifdef CONFIG_TCP_MD5SIG
3918 * The MD5 Hash has already been
3919 * checked (see tcp_v{4,6}_do_rcv()).
3923 case TCPOPT_FASTOPEN
:
3924 tcp_parse_fastopen_option(
3925 opsize
- TCPOLEN_FASTOPEN_BASE
,
3926 ptr
, th
->syn
, foc
, false);
3930 /* Fast Open option shares code 254 using a
3931 * 16 bits magic number.
3933 if (opsize
>= TCPOLEN_EXP_FASTOPEN_BASE
&&
3934 get_unaligned_be16(ptr
) ==
3935 TCPOPT_FASTOPEN_MAGIC
)
3936 tcp_parse_fastopen_option(opsize
-
3937 TCPOLEN_EXP_FASTOPEN_BASE
,
3938 ptr
+ 2, th
->syn
, foc
, true);
3940 smc_parse_options(th
, opt_rx
, ptr
,
3950 EXPORT_SYMBOL(tcp_parse_options
);
3952 static bool tcp_parse_aligned_timestamp(struct tcp_sock
*tp
, const struct tcphdr
*th
)
3954 const __be32
*ptr
= (const __be32
*)(th
+ 1);
3956 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3957 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
3958 tp
->rx_opt
.saw_tstamp
= 1;
3960 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3963 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
) - tp
->tsoffset
;
3965 tp
->rx_opt
.rcv_tsecr
= 0;
3971 /* Fast parse options. This hopes to only see timestamps.
3972 * If it is wrong it falls back on tcp_parse_options().
3974 static bool tcp_fast_parse_options(const struct net
*net
,
3975 const struct sk_buff
*skb
,
3976 const struct tcphdr
*th
, struct tcp_sock
*tp
)
3978 /* In the spirit of fast parsing, compare doff directly to constant
3979 * values. Because equality is used, short doff can be ignored here.
3981 if (th
->doff
== (sizeof(*th
) / 4)) {
3982 tp
->rx_opt
.saw_tstamp
= 0;
3984 } else if (tp
->rx_opt
.tstamp_ok
&&
3985 th
->doff
== ((sizeof(*th
) + TCPOLEN_TSTAMP_ALIGNED
) / 4)) {
3986 if (tcp_parse_aligned_timestamp(tp
, th
))
3990 tcp_parse_options(net
, skb
, &tp
->rx_opt
, 1, NULL
);
3991 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
3992 tp
->rx_opt
.rcv_tsecr
-= tp
->tsoffset
;
3997 #ifdef CONFIG_TCP_MD5SIG
3999 * Parse MD5 Signature option
4001 const u8
*tcp_parse_md5sig_option(const struct tcphdr
*th
)
4003 int length
= (th
->doff
<< 2) - sizeof(*th
);
4004 const u8
*ptr
= (const u8
*)(th
+ 1);
4006 /* If not enough data remaining, we can short cut */
4007 while (length
>= TCPOLEN_MD5SIG
) {
4008 int opcode
= *ptr
++;
4019 if (opsize
< 2 || opsize
> length
)
4021 if (opcode
== TCPOPT_MD5SIG
)
4022 return opsize
== TCPOLEN_MD5SIG
? ptr
: NULL
;
4029 EXPORT_SYMBOL(tcp_parse_md5sig_option
);
4032 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
4034 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
4035 * it can pass through stack. So, the following predicate verifies that
4036 * this segment is not used for anything but congestion avoidance or
4037 * fast retransmit. Moreover, we even are able to eliminate most of such
4038 * second order effects, if we apply some small "replay" window (~RTO)
4039 * to timestamp space.
4041 * All these measures still do not guarantee that we reject wrapped ACKs
4042 * on networks with high bandwidth, when sequence space is recycled fastly,
4043 * but it guarantees that such events will be very rare and do not affect
4044 * connection seriously. This doesn't look nice, but alas, PAWS is really
4047 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
4048 * states that events when retransmit arrives after original data are rare.
4049 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
4050 * the biggest problem on large power networks even with minor reordering.
4051 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
4052 * up to bandwidth of 18Gigabit/sec. 8) ]
4055 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
4057 const struct tcp_sock
*tp
= tcp_sk(sk
);
4058 const struct tcphdr
*th
= tcp_hdr(skb
);
4059 u32 seq
= TCP_SKB_CB(skb
)->seq
;
4060 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
4062 return (/* 1. Pure ACK with correct sequence number. */
4063 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
4065 /* 2. ... and duplicate ACK. */
4066 ack
== tp
->snd_una
&&
4068 /* 3. ... and does not update window. */
4069 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
4071 /* 4. ... and sits in replay window. */
4072 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
4075 static inline bool tcp_paws_discard(const struct sock
*sk
,
4076 const struct sk_buff
*skb
)
4078 const struct tcp_sock
*tp
= tcp_sk(sk
);
4080 return !tcp_paws_check(&tp
->rx_opt
, TCP_PAWS_WINDOW
) &&
4081 !tcp_disordered_ack(sk
, skb
);
4084 /* Check segment sequence number for validity.
4086 * Segment controls are considered valid, if the segment
4087 * fits to the window after truncation to the window. Acceptability
4088 * of data (and SYN, FIN, of course) is checked separately.
4089 * See tcp_data_queue(), for example.
4091 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4092 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4093 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4094 * (borrowed from freebsd)
4097 static inline bool tcp_sequence(const struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
4099 return !before(end_seq
, tp
->rcv_wup
) &&
4100 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
4103 /* When we get a reset we do this. */
4104 void tcp_reset(struct sock
*sk
)
4106 trace_tcp_receive_reset(sk
);
4108 /* We want the right error as BSD sees it (and indeed as we do). */
4109 switch (sk
->sk_state
) {
4111 sk
->sk_err
= ECONNREFUSED
;
4113 case TCP_CLOSE_WAIT
:
4119 sk
->sk_err
= ECONNRESET
;
4121 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4124 tcp_write_queue_purge(sk
);
4127 if (!sock_flag(sk
, SOCK_DEAD
))
4128 sk
->sk_error_report(sk
);
4132 * Process the FIN bit. This now behaves as it is supposed to work
4133 * and the FIN takes effect when it is validly part of sequence
4134 * space. Not before when we get holes.
4136 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4137 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4140 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4141 * close and we go into CLOSING (and later onto TIME-WAIT)
4143 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4145 void tcp_fin(struct sock
*sk
)
4147 struct tcp_sock
*tp
= tcp_sk(sk
);
4149 inet_csk_schedule_ack(sk
);
4151 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
4152 sock_set_flag(sk
, SOCK_DONE
);
4154 switch (sk
->sk_state
) {
4156 case TCP_ESTABLISHED
:
4157 /* Move to CLOSE_WAIT */
4158 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
4159 inet_csk_enter_pingpong_mode(sk
);
4162 case TCP_CLOSE_WAIT
:
4164 /* Received a retransmission of the FIN, do
4169 /* RFC793: Remain in the LAST-ACK state. */
4173 /* This case occurs when a simultaneous close
4174 * happens, we must ack the received FIN and
4175 * enter the CLOSING state.
4178 tcp_set_state(sk
, TCP_CLOSING
);
4181 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4183 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
4186 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4187 * cases we should never reach this piece of code.
4189 pr_err("%s: Impossible, sk->sk_state=%d\n",
4190 __func__
, sk
->sk_state
);
4194 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4195 * Probably, we should reset in this case. For now drop them.
4197 skb_rbtree_purge(&tp
->out_of_order_queue
);
4198 if (tcp_is_sack(tp
))
4199 tcp_sack_reset(&tp
->rx_opt
);
4202 if (!sock_flag(sk
, SOCK_DEAD
)) {
4203 sk
->sk_state_change(sk
);
4205 /* Do not send POLL_HUP for half duplex close. */
4206 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
4207 sk
->sk_state
== TCP_CLOSE
)
4208 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_HUP
);
4210 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_IN
);
4214 static inline bool tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
,
4217 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
4218 if (before(seq
, sp
->start_seq
))
4219 sp
->start_seq
= seq
;
4220 if (after(end_seq
, sp
->end_seq
))
4221 sp
->end_seq
= end_seq
;
4227 static void tcp_dsack_set(struct sock
*sk
, u32 seq
, u32 end_seq
)
4229 struct tcp_sock
*tp
= tcp_sk(sk
);
4231 if (tcp_is_sack(tp
) && sock_net(sk
)->ipv4
.sysctl_tcp_dsack
) {
4234 if (before(seq
, tp
->rcv_nxt
))
4235 mib_idx
= LINUX_MIB_TCPDSACKOLDSENT
;
4237 mib_idx
= LINUX_MIB_TCPDSACKOFOSENT
;
4239 NET_INC_STATS(sock_net(sk
), mib_idx
);
4241 tp
->rx_opt
.dsack
= 1;
4242 tp
->duplicate_sack
[0].start_seq
= seq
;
4243 tp
->duplicate_sack
[0].end_seq
= end_seq
;
4247 static void tcp_dsack_extend(struct sock
*sk
, u32 seq
, u32 end_seq
)
4249 struct tcp_sock
*tp
= tcp_sk(sk
);
4251 if (!tp
->rx_opt
.dsack
)
4252 tcp_dsack_set(sk
, seq
, end_seq
);
4254 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
4257 static void tcp_rcv_spurious_retrans(struct sock
*sk
, const struct sk_buff
*skb
)
4259 /* When the ACK path fails or drops most ACKs, the sender would
4260 * timeout and spuriously retransmit the same segment repeatedly.
4261 * The receiver remembers and reflects via DSACKs. Leverage the
4262 * DSACK state and change the txhash to re-route speculatively.
4264 if (TCP_SKB_CB(skb
)->seq
== tcp_sk(sk
)->duplicate_sack
[0].start_seq
)
4265 sk_rethink_txhash(sk
);
4268 static void tcp_send_dupack(struct sock
*sk
, const struct sk_buff
*skb
)
4270 struct tcp_sock
*tp
= tcp_sk(sk
);
4272 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
4273 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4274 NET_INC_STATS(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4275 tcp_enter_quickack_mode(sk
, TCP_MAX_QUICKACKS
);
4277 if (tcp_is_sack(tp
) && sock_net(sk
)->ipv4
.sysctl_tcp_dsack
) {
4278 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4280 tcp_rcv_spurious_retrans(sk
, skb
);
4281 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
4282 end_seq
= tp
->rcv_nxt
;
4283 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, end_seq
);
4290 /* These routines update the SACK block as out-of-order packets arrive or
4291 * in-order packets close up the sequence space.
4293 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
4296 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4297 struct tcp_sack_block
*swalk
= sp
+ 1;
4299 /* See if the recent change to the first SACK eats into
4300 * or hits the sequence space of other SACK blocks, if so coalesce.
4302 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
;) {
4303 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
4306 /* Zap SWALK, by moving every further SACK up by one slot.
4307 * Decrease num_sacks.
4309 tp
->rx_opt
.num_sacks
--;
4310 for (i
= this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
4314 this_sack
++, swalk
++;
4318 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
4320 struct tcp_sock
*tp
= tcp_sk(sk
);
4321 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4322 int cur_sacks
= tp
->rx_opt
.num_sacks
;
4328 for (this_sack
= 0; this_sack
< cur_sacks
; this_sack
++, sp
++) {
4329 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
4330 /* Rotate this_sack to the first one. */
4331 for (; this_sack
> 0; this_sack
--, sp
--)
4332 swap(*sp
, *(sp
- 1));
4334 tcp_sack_maybe_coalesce(tp
);
4339 /* Could not find an adjacent existing SACK, build a new one,
4340 * put it at the front, and shift everyone else down. We
4341 * always know there is at least one SACK present already here.
4343 * If the sack array is full, forget about the last one.
4345 if (this_sack
>= TCP_NUM_SACKS
) {
4346 if (tp
->compressed_ack
> TCP_FASTRETRANS_THRESH
)
4349 tp
->rx_opt
.num_sacks
--;
4352 for (; this_sack
> 0; this_sack
--, sp
--)
4356 /* Build the new head SACK, and we're done. */
4357 sp
->start_seq
= seq
;
4358 sp
->end_seq
= end_seq
;
4359 tp
->rx_opt
.num_sacks
++;
4362 /* RCV.NXT advances, some SACKs should be eaten. */
4364 static void tcp_sack_remove(struct tcp_sock
*tp
)
4366 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4367 int num_sacks
= tp
->rx_opt
.num_sacks
;
4370 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4371 if (RB_EMPTY_ROOT(&tp
->out_of_order_queue
)) {
4372 tp
->rx_opt
.num_sacks
= 0;
4376 for (this_sack
= 0; this_sack
< num_sacks
;) {
4377 /* Check if the start of the sack is covered by RCV.NXT. */
4378 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
4381 /* RCV.NXT must cover all the block! */
4382 WARN_ON(before(tp
->rcv_nxt
, sp
->end_seq
));
4384 /* Zap this SACK, by moving forward any other SACKS. */
4385 for (i
= this_sack
+1; i
< num_sacks
; i
++)
4386 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
4393 tp
->rx_opt
.num_sacks
= num_sacks
;
4397 * tcp_try_coalesce - try to merge skb to prior one
4399 * @dest: destination queue
4401 * @from: buffer to add in queue
4402 * @fragstolen: pointer to boolean
4404 * Before queueing skb @from after @to, try to merge them
4405 * to reduce overall memory use and queue lengths, if cost is small.
4406 * Packets in ofo or receive queues can stay a long time.
4407 * Better try to coalesce them right now to avoid future collapses.
4408 * Returns true if caller should free @from instead of queueing it
4410 static bool tcp_try_coalesce(struct sock
*sk
,
4412 struct sk_buff
*from
,
4417 *fragstolen
= false;
4419 /* Its possible this segment overlaps with prior segment in queue */
4420 if (TCP_SKB_CB(from
)->seq
!= TCP_SKB_CB(to
)->end_seq
)
4423 #ifdef CONFIG_TLS_DEVICE
4424 if (from
->decrypted
!= to
->decrypted
)
4428 if (!skb_try_coalesce(to
, from
, fragstolen
, &delta
))
4431 atomic_add(delta
, &sk
->sk_rmem_alloc
);
4432 sk_mem_charge(sk
, delta
);
4433 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPRCVCOALESCE
);
4434 TCP_SKB_CB(to
)->end_seq
= TCP_SKB_CB(from
)->end_seq
;
4435 TCP_SKB_CB(to
)->ack_seq
= TCP_SKB_CB(from
)->ack_seq
;
4436 TCP_SKB_CB(to
)->tcp_flags
|= TCP_SKB_CB(from
)->tcp_flags
;
4438 if (TCP_SKB_CB(from
)->has_rxtstamp
) {
4439 TCP_SKB_CB(to
)->has_rxtstamp
= true;
4440 to
->tstamp
= from
->tstamp
;
4441 skb_hwtstamps(to
)->hwtstamp
= skb_hwtstamps(from
)->hwtstamp
;
4447 static bool tcp_ooo_try_coalesce(struct sock
*sk
,
4449 struct sk_buff
*from
,
4452 bool res
= tcp_try_coalesce(sk
, to
, from
, fragstolen
);
4454 /* In case tcp_drop() is called later, update to->gso_segs */
4456 u32 gso_segs
= max_t(u16
, 1, skb_shinfo(to
)->gso_segs
) +
4457 max_t(u16
, 1, skb_shinfo(from
)->gso_segs
);
4459 skb_shinfo(to
)->gso_segs
= min_t(u32
, gso_segs
, 0xFFFF);
4464 static void tcp_drop(struct sock
*sk
, struct sk_buff
*skb
)
4466 sk_drops_add(sk
, skb
);
4470 /* This one checks to see if we can put data from the
4471 * out_of_order queue into the receive_queue.
4473 static void tcp_ofo_queue(struct sock
*sk
)
4475 struct tcp_sock
*tp
= tcp_sk(sk
);
4476 __u32 dsack_high
= tp
->rcv_nxt
;
4477 bool fin
, fragstolen
, eaten
;
4478 struct sk_buff
*skb
, *tail
;
4481 p
= rb_first(&tp
->out_of_order_queue
);
4484 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
4487 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
4488 __u32 dsack
= dsack_high
;
4489 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
4490 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
4491 tcp_dsack_extend(sk
, TCP_SKB_CB(skb
)->seq
, dsack
);
4494 rb_erase(&skb
->rbnode
, &tp
->out_of_order_queue
);
4496 if (unlikely(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))) {
4501 tail
= skb_peek_tail(&sk
->sk_receive_queue
);
4502 eaten
= tail
&& tcp_try_coalesce(sk
, tail
, skb
, &fragstolen
);
4503 tcp_rcv_nxt_update(tp
, TCP_SKB_CB(skb
)->end_seq
);
4504 fin
= TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
;
4506 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4508 kfree_skb_partial(skb
, fragstolen
);
4510 if (unlikely(fin
)) {
4512 /* tcp_fin() purges tp->out_of_order_queue,
4513 * so we must end this loop right now.
4520 static bool tcp_prune_ofo_queue(struct sock
*sk
);
4521 static int tcp_prune_queue(struct sock
*sk
);
4523 static int tcp_try_rmem_schedule(struct sock
*sk
, struct sk_buff
*skb
,
4526 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
4527 !sk_rmem_schedule(sk
, skb
, size
)) {
4529 if (tcp_prune_queue(sk
) < 0)
4532 while (!sk_rmem_schedule(sk
, skb
, size
)) {
4533 if (!tcp_prune_ofo_queue(sk
))
4540 static void tcp_data_queue_ofo(struct sock
*sk
, struct sk_buff
*skb
)
4542 struct tcp_sock
*tp
= tcp_sk(sk
);
4543 struct rb_node
**p
, *parent
;
4544 struct sk_buff
*skb1
;
4548 tcp_ecn_check_ce(sk
, skb
);
4550 if (unlikely(tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))) {
4551 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPOFODROP
);
4556 /* Disable header prediction. */
4558 inet_csk_schedule_ack(sk
);
4560 tp
->rcv_ooopack
+= max_t(u16
, 1, skb_shinfo(skb
)->gso_segs
);
4561 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPOFOQUEUE
);
4562 seq
= TCP_SKB_CB(skb
)->seq
;
4563 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4565 p
= &tp
->out_of_order_queue
.rb_node
;
4566 if (RB_EMPTY_ROOT(&tp
->out_of_order_queue
)) {
4567 /* Initial out of order segment, build 1 SACK. */
4568 if (tcp_is_sack(tp
)) {
4569 tp
->rx_opt
.num_sacks
= 1;
4570 tp
->selective_acks
[0].start_seq
= seq
;
4571 tp
->selective_acks
[0].end_seq
= end_seq
;
4573 rb_link_node(&skb
->rbnode
, NULL
, p
);
4574 rb_insert_color(&skb
->rbnode
, &tp
->out_of_order_queue
);
4575 tp
->ooo_last_skb
= skb
;
4579 /* In the typical case, we are adding an skb to the end of the list.
4580 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4582 if (tcp_ooo_try_coalesce(sk
, tp
->ooo_last_skb
,
4583 skb
, &fragstolen
)) {
4585 tcp_grow_window(sk
, skb
);
4586 kfree_skb_partial(skb
, fragstolen
);
4590 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4591 if (!before(seq
, TCP_SKB_CB(tp
->ooo_last_skb
)->end_seq
)) {
4592 parent
= &tp
->ooo_last_skb
->rbnode
;
4593 p
= &parent
->rb_right
;
4597 /* Find place to insert this segment. Handle overlaps on the way. */
4601 skb1
= rb_to_skb(parent
);
4602 if (before(seq
, TCP_SKB_CB(skb1
)->seq
)) {
4603 p
= &parent
->rb_left
;
4606 if (before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4607 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4608 /* All the bits are present. Drop. */
4609 NET_INC_STATS(sock_net(sk
),
4610 LINUX_MIB_TCPOFOMERGE
);
4613 tcp_dsack_set(sk
, seq
, end_seq
);
4616 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
4617 /* Partial overlap. */
4618 tcp_dsack_set(sk
, seq
, TCP_SKB_CB(skb1
)->end_seq
);
4620 /* skb's seq == skb1's seq and skb covers skb1.
4621 * Replace skb1 with skb.
4623 rb_replace_node(&skb1
->rbnode
, &skb
->rbnode
,
4624 &tp
->out_of_order_queue
);
4625 tcp_dsack_extend(sk
,
4626 TCP_SKB_CB(skb1
)->seq
,
4627 TCP_SKB_CB(skb1
)->end_seq
);
4628 NET_INC_STATS(sock_net(sk
),
4629 LINUX_MIB_TCPOFOMERGE
);
4633 } else if (tcp_ooo_try_coalesce(sk
, skb1
,
4634 skb
, &fragstolen
)) {
4637 p
= &parent
->rb_right
;
4640 /* Insert segment into RB tree. */
4641 rb_link_node(&skb
->rbnode
, parent
, p
);
4642 rb_insert_color(&skb
->rbnode
, &tp
->out_of_order_queue
);
4645 /* Remove other segments covered by skb. */
4646 while ((skb1
= skb_rb_next(skb
)) != NULL
) {
4647 if (!after(end_seq
, TCP_SKB_CB(skb1
)->seq
))
4649 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4650 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4654 rb_erase(&skb1
->rbnode
, &tp
->out_of_order_queue
);
4655 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4656 TCP_SKB_CB(skb1
)->end_seq
);
4657 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPOFOMERGE
);
4660 /* If there is no skb after us, we are the last_skb ! */
4662 tp
->ooo_last_skb
= skb
;
4665 if (tcp_is_sack(tp
))
4666 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
4669 tcp_grow_window(sk
, skb
);
4671 skb_set_owner_r(skb
, sk
);
4675 static int __must_check
tcp_queue_rcv(struct sock
*sk
, struct sk_buff
*skb
,
4679 struct sk_buff
*tail
= skb_peek_tail(&sk
->sk_receive_queue
);
4682 tcp_try_coalesce(sk
, tail
,
4683 skb
, fragstolen
)) ? 1 : 0;
4684 tcp_rcv_nxt_update(tcp_sk(sk
), TCP_SKB_CB(skb
)->end_seq
);
4686 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4687 skb_set_owner_r(skb
, sk
);
4692 int tcp_send_rcvq(struct sock
*sk
, struct msghdr
*msg
, size_t size
)
4694 struct sk_buff
*skb
;
4702 if (size
> PAGE_SIZE
) {
4703 int npages
= min_t(size_t, size
>> PAGE_SHIFT
, MAX_SKB_FRAGS
);
4705 data_len
= npages
<< PAGE_SHIFT
;
4706 size
= data_len
+ (size
& ~PAGE_MASK
);
4708 skb
= alloc_skb_with_frags(size
- data_len
, data_len
,
4709 PAGE_ALLOC_COSTLY_ORDER
,
4710 &err
, sk
->sk_allocation
);
4714 skb_put(skb
, size
- data_len
);
4715 skb
->data_len
= data_len
;
4718 if (tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
)) {
4719 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPRCVQDROP
);
4723 err
= skb_copy_datagram_from_iter(skb
, 0, &msg
->msg_iter
, size
);
4727 TCP_SKB_CB(skb
)->seq
= tcp_sk(sk
)->rcv_nxt
;
4728 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ size
;
4729 TCP_SKB_CB(skb
)->ack_seq
= tcp_sk(sk
)->snd_una
- 1;
4731 if (tcp_queue_rcv(sk
, skb
, &fragstolen
)) {
4732 WARN_ON_ONCE(fragstolen
); /* should not happen */
4744 void tcp_data_ready(struct sock
*sk
)
4746 const struct tcp_sock
*tp
= tcp_sk(sk
);
4747 int avail
= tp
->rcv_nxt
- tp
->copied_seq
;
4749 if (avail
< sk
->sk_rcvlowat
&& !sock_flag(sk
, SOCK_DONE
))
4752 sk
->sk_data_ready(sk
);
4755 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
4757 struct tcp_sock
*tp
= tcp_sk(sk
);
4761 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
) {
4766 __skb_pull(skb
, tcp_hdr(skb
)->doff
* 4);
4768 tcp_ecn_accept_cwr(sk
, skb
);
4770 tp
->rx_opt
.dsack
= 0;
4772 /* Queue data for delivery to the user.
4773 * Packets in sequence go to the receive queue.
4774 * Out of sequence packets to the out_of_order_queue.
4776 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
4777 if (tcp_receive_window(tp
) == 0) {
4778 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPZEROWINDOWDROP
);
4782 /* Ok. In sequence. In window. */
4784 if (skb_queue_len(&sk
->sk_receive_queue
) == 0)
4785 sk_forced_mem_schedule(sk
, skb
->truesize
);
4786 else if (tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
)) {
4787 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPRCVQDROP
);
4791 eaten
= tcp_queue_rcv(sk
, skb
, &fragstolen
);
4793 tcp_event_data_recv(sk
, skb
);
4794 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
4797 if (!RB_EMPTY_ROOT(&tp
->out_of_order_queue
)) {
4800 /* RFC5681. 4.2. SHOULD send immediate ACK, when
4801 * gap in queue is filled.
4803 if (RB_EMPTY_ROOT(&tp
->out_of_order_queue
))
4804 inet_csk(sk
)->icsk_ack
.pending
|= ICSK_ACK_NOW
;
4807 if (tp
->rx_opt
.num_sacks
)
4808 tcp_sack_remove(tp
);
4810 tcp_fast_path_check(sk
);
4813 kfree_skb_partial(skb
, fragstolen
);
4814 if (!sock_flag(sk
, SOCK_DEAD
))
4819 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4820 tcp_rcv_spurious_retrans(sk
, skb
);
4821 /* A retransmit, 2nd most common case. Force an immediate ack. */
4822 NET_INC_STATS(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4823 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4826 tcp_enter_quickack_mode(sk
, TCP_MAX_QUICKACKS
);
4827 inet_csk_schedule_ack(sk
);
4833 /* Out of window. F.e. zero window probe. */
4834 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
4837 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4838 /* Partial packet, seq < rcv_next < end_seq */
4839 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
4841 /* If window is closed, drop tail of packet. But after
4842 * remembering D-SACK for its head made in previous line.
4844 if (!tcp_receive_window(tp
)) {
4845 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPZEROWINDOWDROP
);
4851 tcp_data_queue_ofo(sk
, skb
);
4854 static struct sk_buff
*tcp_skb_next(struct sk_buff
*skb
, struct sk_buff_head
*list
)
4857 return !skb_queue_is_last(list
, skb
) ? skb
->next
: NULL
;
4859 return skb_rb_next(skb
);
4862 static struct sk_buff
*tcp_collapse_one(struct sock
*sk
, struct sk_buff
*skb
,
4863 struct sk_buff_head
*list
,
4864 struct rb_root
*root
)
4866 struct sk_buff
*next
= tcp_skb_next(skb
, list
);
4869 __skb_unlink(skb
, list
);
4871 rb_erase(&skb
->rbnode
, root
);
4874 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPRCVCOLLAPSED
);
4879 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
4880 void tcp_rbtree_insert(struct rb_root
*root
, struct sk_buff
*skb
)
4882 struct rb_node
**p
= &root
->rb_node
;
4883 struct rb_node
*parent
= NULL
;
4884 struct sk_buff
*skb1
;
4888 skb1
= rb_to_skb(parent
);
4889 if (before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb1
)->seq
))
4890 p
= &parent
->rb_left
;
4892 p
= &parent
->rb_right
;
4894 rb_link_node(&skb
->rbnode
, parent
, p
);
4895 rb_insert_color(&skb
->rbnode
, root
);
4898 /* Collapse contiguous sequence of skbs head..tail with
4899 * sequence numbers start..end.
4901 * If tail is NULL, this means until the end of the queue.
4903 * Segments with FIN/SYN are not collapsed (only because this
4907 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
, struct rb_root
*root
,
4908 struct sk_buff
*head
, struct sk_buff
*tail
, u32 start
, u32 end
)
4910 struct sk_buff
*skb
= head
, *n
;
4911 struct sk_buff_head tmp
;
4914 /* First, check that queue is collapsible and find
4915 * the point where collapsing can be useful.
4918 for (end_of_skbs
= true; skb
!= NULL
&& skb
!= tail
; skb
= n
) {
4919 n
= tcp_skb_next(skb
, list
);
4921 /* No new bits? It is possible on ofo queue. */
4922 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4923 skb
= tcp_collapse_one(sk
, skb
, list
, root
);
4929 /* The first skb to collapse is:
4931 * - bloated or contains data before "start" or
4932 * overlaps to the next one.
4934 if (!(TCP_SKB_CB(skb
)->tcp_flags
& (TCPHDR_SYN
| TCPHDR_FIN
)) &&
4935 (tcp_win_from_space(sk
, skb
->truesize
) > skb
->len
||
4936 before(TCP_SKB_CB(skb
)->seq
, start
))) {
4937 end_of_skbs
= false;
4941 if (n
&& n
!= tail
&&
4942 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(n
)->seq
) {
4943 end_of_skbs
= false;
4947 /* Decided to skip this, advance start seq. */
4948 start
= TCP_SKB_CB(skb
)->end_seq
;
4951 (TCP_SKB_CB(skb
)->tcp_flags
& (TCPHDR_SYN
| TCPHDR_FIN
)))
4954 __skb_queue_head_init(&tmp
);
4956 while (before(start
, end
)) {
4957 int copy
= min_t(int, SKB_MAX_ORDER(0, 0), end
- start
);
4958 struct sk_buff
*nskb
;
4960 nskb
= alloc_skb(copy
, GFP_ATOMIC
);
4964 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
4965 #ifdef CONFIG_TLS_DEVICE
4966 nskb
->decrypted
= skb
->decrypted
;
4968 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
4970 __skb_queue_before(list
, skb
, nskb
);
4972 __skb_queue_tail(&tmp
, nskb
); /* defer rbtree insertion */
4973 skb_set_owner_r(nskb
, sk
);
4975 /* Copy data, releasing collapsed skbs. */
4977 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
4978 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
4982 size
= min(copy
, size
);
4983 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
4985 TCP_SKB_CB(nskb
)->end_seq
+= size
;
4989 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4990 skb
= tcp_collapse_one(sk
, skb
, list
, root
);
4993 (TCP_SKB_CB(skb
)->tcp_flags
& (TCPHDR_SYN
| TCPHDR_FIN
)))
4995 #ifdef CONFIG_TLS_DEVICE
4996 if (skb
->decrypted
!= nskb
->decrypted
)
5003 skb_queue_walk_safe(&tmp
, skb
, n
)
5004 tcp_rbtree_insert(root
, skb
);
5007 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
5008 * and tcp_collapse() them until all the queue is collapsed.
5010 static void tcp_collapse_ofo_queue(struct sock
*sk
)
5012 struct tcp_sock
*tp
= tcp_sk(sk
);
5013 u32 range_truesize
, sum_tiny
= 0;
5014 struct sk_buff
*skb
, *head
;
5017 skb
= skb_rb_first(&tp
->out_of_order_queue
);
5020 tp
->ooo_last_skb
= skb_rb_last(&tp
->out_of_order_queue
);
5023 start
= TCP_SKB_CB(skb
)->seq
;
5024 end
= TCP_SKB_CB(skb
)->end_seq
;
5025 range_truesize
= skb
->truesize
;
5027 for (head
= skb
;;) {
5028 skb
= skb_rb_next(skb
);
5030 /* Range is terminated when we see a gap or when
5031 * we are at the queue end.
5034 after(TCP_SKB_CB(skb
)->seq
, end
) ||
5035 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
5036 /* Do not attempt collapsing tiny skbs */
5037 if (range_truesize
!= head
->truesize
||
5038 end
- start
>= SKB_WITH_OVERHEAD(SK_MEM_QUANTUM
)) {
5039 tcp_collapse(sk
, NULL
, &tp
->out_of_order_queue
,
5040 head
, skb
, start
, end
);
5042 sum_tiny
+= range_truesize
;
5043 if (sum_tiny
> sk
->sk_rcvbuf
>> 3)
5049 range_truesize
+= skb
->truesize
;
5050 if (unlikely(before(TCP_SKB_CB(skb
)->seq
, start
)))
5051 start
= TCP_SKB_CB(skb
)->seq
;
5052 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
5053 end
= TCP_SKB_CB(skb
)->end_seq
;
5058 * Clean the out-of-order queue to make room.
5059 * We drop high sequences packets to :
5060 * 1) Let a chance for holes to be filled.
5061 * 2) not add too big latencies if thousands of packets sit there.
5062 * (But if application shrinks SO_RCVBUF, we could still end up
5063 * freeing whole queue here)
5064 * 3) Drop at least 12.5 % of sk_rcvbuf to avoid malicious attacks.
5066 * Return true if queue has shrunk.
5068 static bool tcp_prune_ofo_queue(struct sock
*sk
)
5070 struct tcp_sock
*tp
= tcp_sk(sk
);
5071 struct rb_node
*node
, *prev
;
5074 if (RB_EMPTY_ROOT(&tp
->out_of_order_queue
))
5077 NET_INC_STATS(sock_net(sk
), LINUX_MIB_OFOPRUNED
);
5078 goal
= sk
->sk_rcvbuf
>> 3;
5079 node
= &tp
->ooo_last_skb
->rbnode
;
5081 prev
= rb_prev(node
);
5082 rb_erase(node
, &tp
->out_of_order_queue
);
5083 goal
-= rb_to_skb(node
)->truesize
;
5084 tcp_drop(sk
, rb_to_skb(node
));
5085 if (!prev
|| goal
<= 0) {
5087 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
&&
5088 !tcp_under_memory_pressure(sk
))
5090 goal
= sk
->sk_rcvbuf
>> 3;
5094 tp
->ooo_last_skb
= rb_to_skb(prev
);
5096 /* Reset SACK state. A conforming SACK implementation will
5097 * do the same at a timeout based retransmit. When a connection
5098 * is in a sad state like this, we care only about integrity
5099 * of the connection not performance.
5101 if (tp
->rx_opt
.sack_ok
)
5102 tcp_sack_reset(&tp
->rx_opt
);
5106 /* Reduce allocated memory if we can, trying to get
5107 * the socket within its memory limits again.
5109 * Return less than zero if we should start dropping frames
5110 * until the socket owning process reads some of the data
5111 * to stabilize the situation.
5113 static int tcp_prune_queue(struct sock
*sk
)
5115 struct tcp_sock
*tp
= tcp_sk(sk
);
5117 NET_INC_STATS(sock_net(sk
), LINUX_MIB_PRUNECALLED
);
5119 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
5120 tcp_clamp_window(sk
);
5121 else if (tcp_under_memory_pressure(sk
))
5122 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
5124 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
5127 tcp_collapse_ofo_queue(sk
);
5128 if (!skb_queue_empty(&sk
->sk_receive_queue
))
5129 tcp_collapse(sk
, &sk
->sk_receive_queue
, NULL
,
5130 skb_peek(&sk
->sk_receive_queue
),
5132 tp
->copied_seq
, tp
->rcv_nxt
);
5135 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
5138 /* Collapsing did not help, destructive actions follow.
5139 * This must not ever occur. */
5141 tcp_prune_ofo_queue(sk
);
5143 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
5146 /* If we are really being abused, tell the caller to silently
5147 * drop receive data on the floor. It will get retransmitted
5148 * and hopefully then we'll have sufficient space.
5150 NET_INC_STATS(sock_net(sk
), LINUX_MIB_RCVPRUNED
);
5152 /* Massive buffer overcommit. */
5157 static bool tcp_should_expand_sndbuf(const struct sock
*sk
)
5159 const struct tcp_sock
*tp
= tcp_sk(sk
);
5161 /* If the user specified a specific send buffer setting, do
5164 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
5167 /* If we are under global TCP memory pressure, do not expand. */
5168 if (tcp_under_memory_pressure(sk
))
5171 /* If we are under soft global TCP memory pressure, do not expand. */
5172 if (sk_memory_allocated(sk
) >= sk_prot_mem_limits(sk
, 0))
5175 /* If we filled the congestion window, do not expand. */
5176 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
5182 /* When incoming ACK allowed to free some skb from write_queue,
5183 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
5184 * on the exit from tcp input handler.
5186 * PROBLEM: sndbuf expansion does not work well with largesend.
5188 static void tcp_new_space(struct sock
*sk
)
5190 struct tcp_sock
*tp
= tcp_sk(sk
);
5192 if (tcp_should_expand_sndbuf(sk
)) {
5193 tcp_sndbuf_expand(sk
);
5194 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
5197 sk
->sk_write_space(sk
);
5200 static void tcp_check_space(struct sock
*sk
)
5202 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
5203 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
5204 /* pairs with tcp_poll() */
5206 if (sk
->sk_socket
&&
5207 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
5209 if (!test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
5210 tcp_chrono_stop(sk
, TCP_CHRONO_SNDBUF_LIMITED
);
5215 static inline void tcp_data_snd_check(struct sock
*sk
)
5217 tcp_push_pending_frames(sk
);
5218 tcp_check_space(sk
);
5222 * Check if sending an ack is needed.
5224 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
5226 struct tcp_sock
*tp
= tcp_sk(sk
);
5227 unsigned long rtt
, delay
;
5229 /* More than one full frame received... */
5230 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
&&
5231 /* ... and right edge of window advances far enough.
5232 * (tcp_recvmsg() will send ACK otherwise).
5233 * If application uses SO_RCVLOWAT, we want send ack now if
5234 * we have not received enough bytes to satisfy the condition.
5236 (tp
->rcv_nxt
- tp
->copied_seq
< sk
->sk_rcvlowat
||
5237 __tcp_select_window(sk
) >= tp
->rcv_wnd
)) ||
5238 /* We ACK each frame or... */
5239 tcp_in_quickack_mode(sk
) ||
5240 /* Protocol state mandates a one-time immediate ACK */
5241 inet_csk(sk
)->icsk_ack
.pending
& ICSK_ACK_NOW
) {
5247 if (!ofo_possible
|| RB_EMPTY_ROOT(&tp
->out_of_order_queue
)) {
5248 tcp_send_delayed_ack(sk
);
5252 if (!tcp_is_sack(tp
) ||
5253 tp
->compressed_ack
>= sock_net(sk
)->ipv4
.sysctl_tcp_comp_sack_nr
)
5256 if (tp
->compressed_ack_rcv_nxt
!= tp
->rcv_nxt
) {
5257 tp
->compressed_ack_rcv_nxt
= tp
->rcv_nxt
;
5258 if (tp
->compressed_ack
> TCP_FASTRETRANS_THRESH
)
5259 NET_ADD_STATS(sock_net(sk
), LINUX_MIB_TCPACKCOMPRESSED
,
5260 tp
->compressed_ack
- TCP_FASTRETRANS_THRESH
);
5261 tp
->compressed_ack
= 0;
5264 if (++tp
->compressed_ack
<= TCP_FASTRETRANS_THRESH
)
5267 if (hrtimer_is_queued(&tp
->compressed_ack_timer
))
5270 /* compress ack timer : 5 % of rtt, but no more than tcp_comp_sack_delay_ns */
5272 rtt
= tp
->rcv_rtt_est
.rtt_us
;
5273 if (tp
->srtt_us
&& tp
->srtt_us
< rtt
)
5276 delay
= min_t(unsigned long, sock_net(sk
)->ipv4
.sysctl_tcp_comp_sack_delay_ns
,
5277 rtt
* (NSEC_PER_USEC
>> 3)/20);
5279 hrtimer_start(&tp
->compressed_ack_timer
, ns_to_ktime(delay
),
5280 HRTIMER_MODE_REL_PINNED_SOFT
);
5283 static inline void tcp_ack_snd_check(struct sock
*sk
)
5285 if (!inet_csk_ack_scheduled(sk
)) {
5286 /* We sent a data segment already. */
5289 __tcp_ack_snd_check(sk
, 1);
5293 * This routine is only called when we have urgent data
5294 * signaled. Its the 'slow' part of tcp_urg. It could be
5295 * moved inline now as tcp_urg is only called from one
5296 * place. We handle URGent data wrong. We have to - as
5297 * BSD still doesn't use the correction from RFC961.
5298 * For 1003.1g we should support a new option TCP_STDURG to permit
5299 * either form (or just set the sysctl tcp_stdurg).
5302 static void tcp_check_urg(struct sock
*sk
, const struct tcphdr
*th
)
5304 struct tcp_sock
*tp
= tcp_sk(sk
);
5305 u32 ptr
= ntohs(th
->urg_ptr
);
5307 if (ptr
&& !sock_net(sk
)->ipv4
.sysctl_tcp_stdurg
)
5309 ptr
+= ntohl(th
->seq
);
5311 /* Ignore urgent data that we've already seen and read. */
5312 if (after(tp
->copied_seq
, ptr
))
5315 /* Do not replay urg ptr.
5317 * NOTE: interesting situation not covered by specs.
5318 * Misbehaving sender may send urg ptr, pointing to segment,
5319 * which we already have in ofo queue. We are not able to fetch
5320 * such data and will stay in TCP_URG_NOTYET until will be eaten
5321 * by recvmsg(). Seems, we are not obliged to handle such wicked
5322 * situations. But it is worth to think about possibility of some
5323 * DoSes using some hypothetical application level deadlock.
5325 if (before(ptr
, tp
->rcv_nxt
))
5328 /* Do we already have a newer (or duplicate) urgent pointer? */
5329 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
5332 /* Tell the world about our new urgent pointer. */
5335 /* We may be adding urgent data when the last byte read was
5336 * urgent. To do this requires some care. We cannot just ignore
5337 * tp->copied_seq since we would read the last urgent byte again
5338 * as data, nor can we alter copied_seq until this data arrives
5339 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5341 * NOTE. Double Dutch. Rendering to plain English: author of comment
5342 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5343 * and expect that both A and B disappear from stream. This is _wrong_.
5344 * Though this happens in BSD with high probability, this is occasional.
5345 * Any application relying on this is buggy. Note also, that fix "works"
5346 * only in this artificial test. Insert some normal data between A and B and we will
5347 * decline of BSD again. Verdict: it is better to remove to trap
5350 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
5351 !sock_flag(sk
, SOCK_URGINLINE
) && tp
->copied_seq
!= tp
->rcv_nxt
) {
5352 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
5354 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5355 __skb_unlink(skb
, &sk
->sk_receive_queue
);
5360 tp
->urg_data
= TCP_URG_NOTYET
;
5361 WRITE_ONCE(tp
->urg_seq
, ptr
);
5363 /* Disable header prediction. */
5367 /* This is the 'fast' part of urgent handling. */
5368 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, const struct tcphdr
*th
)
5370 struct tcp_sock
*tp
= tcp_sk(sk
);
5372 /* Check if we get a new urgent pointer - normally not. */
5374 tcp_check_urg(sk
, th
);
5376 /* Do we wait for any urgent data? - normally not... */
5377 if (tp
->urg_data
== TCP_URG_NOTYET
) {
5378 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
5381 /* Is the urgent pointer pointing into this packet? */
5382 if (ptr
< skb
->len
) {
5384 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
5386 tp
->urg_data
= TCP_URG_VALID
| tmp
;
5387 if (!sock_flag(sk
, SOCK_DEAD
))
5388 sk
->sk_data_ready(sk
);
5393 /* Accept RST for rcv_nxt - 1 after a FIN.
5394 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a
5395 * FIN is sent followed by a RST packet. The RST is sent with the same
5396 * sequence number as the FIN, and thus according to RFC 5961 a challenge
5397 * ACK should be sent. However, Mac OSX rate limits replies to challenge
5398 * ACKs on the closed socket. In addition middleboxes can drop either the
5399 * challenge ACK or a subsequent RST.
5401 static bool tcp_reset_check(const struct sock
*sk
, const struct sk_buff
*skb
)
5403 struct tcp_sock
*tp
= tcp_sk(sk
);
5405 return unlikely(TCP_SKB_CB(skb
)->seq
== (tp
->rcv_nxt
- 1) &&
5406 (1 << sk
->sk_state
) & (TCPF_CLOSE_WAIT
| TCPF_LAST_ACK
|
5410 /* Does PAWS and seqno based validation of an incoming segment, flags will
5411 * play significant role here.
5413 static bool tcp_validate_incoming(struct sock
*sk
, struct sk_buff
*skb
,
5414 const struct tcphdr
*th
, int syn_inerr
)
5416 struct tcp_sock
*tp
= tcp_sk(sk
);
5417 bool rst_seq_match
= false;
5419 /* RFC1323: H1. Apply PAWS check first. */
5420 if (tcp_fast_parse_options(sock_net(sk
), skb
, th
, tp
) &&
5421 tp
->rx_opt
.saw_tstamp
&&
5422 tcp_paws_discard(sk
, skb
)) {
5424 NET_INC_STATS(sock_net(sk
), LINUX_MIB_PAWSESTABREJECTED
);
5425 if (!tcp_oow_rate_limited(sock_net(sk
), skb
,
5426 LINUX_MIB_TCPACKSKIPPEDPAWS
,
5427 &tp
->last_oow_ack_time
))
5428 tcp_send_dupack(sk
, skb
);
5431 /* Reset is accepted even if it did not pass PAWS. */
5434 /* Step 1: check sequence number */
5435 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5436 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5437 * (RST) segments are validated by checking their SEQ-fields."
5438 * And page 69: "If an incoming segment is not acceptable,
5439 * an acknowledgment should be sent in reply (unless the RST
5440 * bit is set, if so drop the segment and return)".
5445 if (!tcp_oow_rate_limited(sock_net(sk
), skb
,
5446 LINUX_MIB_TCPACKSKIPPEDSEQ
,
5447 &tp
->last_oow_ack_time
))
5448 tcp_send_dupack(sk
, skb
);
5449 } else if (tcp_reset_check(sk
, skb
)) {
5455 /* Step 2: check RST bit */
5457 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a
5458 * FIN and SACK too if available):
5459 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or
5460 * the right-most SACK block,
5462 * RESET the connection
5464 * Send a challenge ACK
5466 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
||
5467 tcp_reset_check(sk
, skb
)) {
5468 rst_seq_match
= true;
5469 } else if (tcp_is_sack(tp
) && tp
->rx_opt
.num_sacks
> 0) {
5470 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
5471 int max_sack
= sp
[0].end_seq
;
5474 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
;
5476 max_sack
= after(sp
[this_sack
].end_seq
,
5478 sp
[this_sack
].end_seq
: max_sack
;
5481 if (TCP_SKB_CB(skb
)->seq
== max_sack
)
5482 rst_seq_match
= true;
5488 /* Disable TFO if RST is out-of-order
5489 * and no data has been received
5490 * for current active TFO socket
5492 if (tp
->syn_fastopen
&& !tp
->data_segs_in
&&
5493 sk
->sk_state
== TCP_ESTABLISHED
)
5494 tcp_fastopen_active_disable(sk
);
5495 tcp_send_challenge_ack(sk
, skb
);
5500 /* step 3: check security and precedence [ignored] */
5502 /* step 4: Check for a SYN
5503 * RFC 5961 4.2 : Send a challenge ack
5508 TCP_INC_STATS(sock_net(sk
), TCP_MIB_INERRS
);
5509 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPSYNCHALLENGE
);
5510 tcp_send_challenge_ack(sk
, skb
);
5522 * TCP receive function for the ESTABLISHED state.
5524 * It is split into a fast path and a slow path. The fast path is
5526 * - A zero window was announced from us - zero window probing
5527 * is only handled properly in the slow path.
5528 * - Out of order segments arrived.
5529 * - Urgent data is expected.
5530 * - There is no buffer space left
5531 * - Unexpected TCP flags/window values/header lengths are received
5532 * (detected by checking the TCP header against pred_flags)
5533 * - Data is sent in both directions. Fast path only supports pure senders
5534 * or pure receivers (this means either the sequence number or the ack
5535 * value must stay constant)
5536 * - Unexpected TCP option.
5538 * When these conditions are not satisfied it drops into a standard
5539 * receive procedure patterned after RFC793 to handle all cases.
5540 * The first three cases are guaranteed by proper pred_flags setting,
5541 * the rest is checked inline. Fast processing is turned on in
5542 * tcp_data_queue when everything is OK.
5544 void tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
)
5546 const struct tcphdr
*th
= (const struct tcphdr
*)skb
->data
;
5547 struct tcp_sock
*tp
= tcp_sk(sk
);
5548 unsigned int len
= skb
->len
;
5550 /* TCP congestion window tracking */
5551 trace_tcp_probe(sk
, skb
);
5553 tcp_mstamp_refresh(tp
);
5554 if (unlikely(!sk
->sk_rx_dst
))
5555 inet_csk(sk
)->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
5557 * Header prediction.
5558 * The code loosely follows the one in the famous
5559 * "30 instruction TCP receive" Van Jacobson mail.
5561 * Van's trick is to deposit buffers into socket queue
5562 * on a device interrupt, to call tcp_recv function
5563 * on the receive process context and checksum and copy
5564 * the buffer to user space. smart...
5566 * Our current scheme is not silly either but we take the
5567 * extra cost of the net_bh soft interrupt processing...
5568 * We do checksum and copy also but from device to kernel.
5571 tp
->rx_opt
.saw_tstamp
= 0;
5573 /* pred_flags is 0xS?10 << 16 + snd_wnd
5574 * if header_prediction is to be made
5575 * 'S' will always be tp->tcp_header_len >> 2
5576 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5577 * turn it off (when there are holes in the receive
5578 * space for instance)
5579 * PSH flag is ignored.
5582 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
5583 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
&&
5584 !after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
)) {
5585 int tcp_header_len
= tp
->tcp_header_len
;
5587 /* Timestamp header prediction: tcp_header_len
5588 * is automatically equal to th->doff*4 due to pred_flags
5592 /* Check timestamp */
5593 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
5594 /* No? Slow path! */
5595 if (!tcp_parse_aligned_timestamp(tp
, th
))
5598 /* If PAWS failed, check it more carefully in slow path */
5599 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
5602 /* DO NOT update ts_recent here, if checksum fails
5603 * and timestamp was corrupted part, it will result
5604 * in a hung connection since we will drop all
5605 * future packets due to the PAWS test.
5609 if (len
<= tcp_header_len
) {
5610 /* Bulk data transfer: sender */
5611 if (len
== tcp_header_len
) {
5612 /* Predicted packet is in window by definition.
5613 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5614 * Hence, check seq<=rcv_wup reduces to:
5616 if (tcp_header_len
==
5617 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5618 tp
->rcv_nxt
== tp
->rcv_wup
)
5619 tcp_store_ts_recent(tp
);
5621 /* We know that such packets are checksummed
5624 tcp_ack(sk
, skb
, 0);
5626 tcp_data_snd_check(sk
);
5627 /* When receiving pure ack in fast path, update
5628 * last ts ecr directly instead of calling
5629 * tcp_rcv_rtt_measure_ts()
5631 tp
->rcv_rtt_last_tsecr
= tp
->rx_opt
.rcv_tsecr
;
5633 } else { /* Header too small */
5634 TCP_INC_STATS(sock_net(sk
), TCP_MIB_INERRS
);
5639 bool fragstolen
= false;
5641 if (tcp_checksum_complete(skb
))
5644 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
5647 /* Predicted packet is in window by definition.
5648 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5649 * Hence, check seq<=rcv_wup reduces to:
5651 if (tcp_header_len
==
5652 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5653 tp
->rcv_nxt
== tp
->rcv_wup
)
5654 tcp_store_ts_recent(tp
);
5656 tcp_rcv_rtt_measure_ts(sk
, skb
);
5658 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPHPHITS
);
5660 /* Bulk data transfer: receiver */
5661 __skb_pull(skb
, tcp_header_len
);
5662 eaten
= tcp_queue_rcv(sk
, skb
, &fragstolen
);
5664 tcp_event_data_recv(sk
, skb
);
5666 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
5667 /* Well, only one small jumplet in fast path... */
5668 tcp_ack(sk
, skb
, FLAG_DATA
);
5669 tcp_data_snd_check(sk
);
5670 if (!inet_csk_ack_scheduled(sk
))
5674 __tcp_ack_snd_check(sk
, 0);
5677 kfree_skb_partial(skb
, fragstolen
);
5684 if (len
< (th
->doff
<< 2) || tcp_checksum_complete(skb
))
5687 if (!th
->ack
&& !th
->rst
&& !th
->syn
)
5691 * Standard slow path.
5694 if (!tcp_validate_incoming(sk
, skb
, th
, 1))
5698 if (tcp_ack(sk
, skb
, FLAG_SLOWPATH
| FLAG_UPDATE_TS_RECENT
) < 0)
5701 tcp_rcv_rtt_measure_ts(sk
, skb
);
5703 /* Process urgent data. */
5704 tcp_urg(sk
, skb
, th
);
5706 /* step 7: process the segment text */
5707 tcp_data_queue(sk
, skb
);
5709 tcp_data_snd_check(sk
);
5710 tcp_ack_snd_check(sk
);
5714 TCP_INC_STATS(sock_net(sk
), TCP_MIB_CSUMERRORS
);
5715 TCP_INC_STATS(sock_net(sk
), TCP_MIB_INERRS
);
5720 EXPORT_SYMBOL(tcp_rcv_established
);
5722 void tcp_init_transfer(struct sock
*sk
, int bpf_op
)
5724 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5725 struct tcp_sock
*tp
= tcp_sk(sk
);
5728 icsk
->icsk_af_ops
->rebuild_header(sk
);
5729 tcp_init_metrics(sk
);
5731 /* Initialize the congestion window to start the transfer.
5732 * Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
5733 * retransmitted. In light of RFC6298 more aggressive 1sec
5734 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
5735 * retransmission has occurred.
5737 if (tp
->total_retrans
> 1 && tp
->undo_marker
)
5740 tp
->snd_cwnd
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
5741 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
5743 tcp_call_bpf(sk
, bpf_op
, 0, NULL
);
5744 tcp_init_congestion_control(sk
);
5745 tcp_init_buffer_space(sk
);
5748 void tcp_finish_connect(struct sock
*sk
, struct sk_buff
*skb
)
5750 struct tcp_sock
*tp
= tcp_sk(sk
);
5751 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5753 tcp_set_state(sk
, TCP_ESTABLISHED
);
5754 icsk
->icsk_ack
.lrcvtime
= tcp_jiffies32
;
5757 icsk
->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
5758 security_inet_conn_established(sk
, skb
);
5759 sk_mark_napi_id(sk
, skb
);
5762 tcp_init_transfer(sk
, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB
);
5764 /* Prevent spurious tcp_cwnd_restart() on first data
5767 tp
->lsndtime
= tcp_jiffies32
;
5769 if (sock_flag(sk
, SOCK_KEEPOPEN
))
5770 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
5772 if (!tp
->rx_opt
.snd_wscale
)
5773 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
5778 static bool tcp_rcv_fastopen_synack(struct sock
*sk
, struct sk_buff
*synack
,
5779 struct tcp_fastopen_cookie
*cookie
)
5781 struct tcp_sock
*tp
= tcp_sk(sk
);
5782 struct sk_buff
*data
= tp
->syn_data
? tcp_rtx_queue_head(sk
) : NULL
;
5783 u16 mss
= tp
->rx_opt
.mss_clamp
, try_exp
= 0;
5784 bool syn_drop
= false;
5786 if (mss
== tp
->rx_opt
.user_mss
) {
5787 struct tcp_options_received opt
;
5789 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5790 tcp_clear_options(&opt
);
5791 opt
.user_mss
= opt
.mss_clamp
= 0;
5792 tcp_parse_options(sock_net(sk
), synack
, &opt
, 0, NULL
);
5793 mss
= opt
.mss_clamp
;
5796 if (!tp
->syn_fastopen
) {
5797 /* Ignore an unsolicited cookie */
5799 } else if (tp
->total_retrans
) {
5800 /* SYN timed out and the SYN-ACK neither has a cookie nor
5801 * acknowledges data. Presumably the remote received only
5802 * the retransmitted (regular) SYNs: either the original
5803 * SYN-data or the corresponding SYN-ACK was dropped.
5805 syn_drop
= (cookie
->len
< 0 && data
);
5806 } else if (cookie
->len
< 0 && !tp
->syn_data
) {
5807 /* We requested a cookie but didn't get it. If we did not use
5808 * the (old) exp opt format then try so next time (try_exp=1).
5809 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5811 try_exp
= tp
->syn_fastopen_exp
? 2 : 1;
5814 tcp_fastopen_cache_set(sk
, mss
, cookie
, syn_drop
, try_exp
);
5816 if (data
) { /* Retransmit unacked data in SYN */
5817 skb_rbtree_walk_from(data
) {
5818 if (__tcp_retransmit_skb(sk
, data
, 1))
5822 NET_INC_STATS(sock_net(sk
),
5823 LINUX_MIB_TCPFASTOPENACTIVEFAIL
);
5826 tp
->syn_data_acked
= tp
->syn_data
;
5827 if (tp
->syn_data_acked
) {
5828 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPFASTOPENACTIVE
);
5829 /* SYN-data is counted as two separate packets in tcp_ack() */
5830 if (tp
->delivered
> 1)
5834 tcp_fastopen_add_skb(sk
, synack
);
5839 static void smc_check_reset_syn(struct tcp_sock
*tp
)
5841 #if IS_ENABLED(CONFIG_SMC)
5842 if (static_branch_unlikely(&tcp_have_smc
)) {
5843 if (tp
->syn_smc
&& !tp
->rx_opt
.smc_ok
)
5849 static void tcp_try_undo_spurious_syn(struct sock
*sk
)
5851 struct tcp_sock
*tp
= tcp_sk(sk
);
5854 /* undo_marker is set when SYN or SYNACK times out. The timeout is
5855 * spurious if the ACK's timestamp option echo value matches the
5856 * original SYN timestamp.
5858 syn_stamp
= tp
->retrans_stamp
;
5859 if (tp
->undo_marker
&& syn_stamp
&& tp
->rx_opt
.saw_tstamp
&&
5860 syn_stamp
== tp
->rx_opt
.rcv_tsecr
)
5861 tp
->undo_marker
= 0;
5864 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5865 const struct tcphdr
*th
)
5867 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5868 struct tcp_sock
*tp
= tcp_sk(sk
);
5869 struct tcp_fastopen_cookie foc
= { .len
= -1 };
5870 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
5873 tcp_parse_options(sock_net(sk
), skb
, &tp
->rx_opt
, 0, &foc
);
5874 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
5875 tp
->rx_opt
.rcv_tsecr
-= tp
->tsoffset
;
5879 * "If the state is SYN-SENT then
5880 * first check the ACK bit
5881 * If the ACK bit is set
5882 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5883 * a reset (unless the RST bit is set, if so drop
5884 * the segment and return)"
5886 if (!after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_una
) ||
5887 after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
))
5888 goto reset_and_undo
;
5890 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
5891 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
5892 tcp_time_stamp(tp
))) {
5893 NET_INC_STATS(sock_net(sk
),
5894 LINUX_MIB_PAWSACTIVEREJECTED
);
5895 goto reset_and_undo
;
5898 /* Now ACK is acceptable.
5900 * "If the RST bit is set
5901 * If the ACK was acceptable then signal the user "error:
5902 * connection reset", drop the segment, enter CLOSED state,
5903 * delete TCB, and return."
5912 * "fifth, if neither of the SYN or RST bits is set then
5913 * drop the segment and return."
5919 goto discard_and_undo
;
5922 * "If the SYN bit is on ...
5923 * are acceptable then ...
5924 * (our SYN has been ACKed), change the connection
5925 * state to ESTABLISHED..."
5928 tcp_ecn_rcv_synack(tp
, th
);
5930 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5931 tcp_try_undo_spurious_syn(sk
);
5932 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
5934 /* Ok.. it's good. Set up sequence numbers and
5935 * move to established.
5937 WRITE_ONCE(tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
+ 1);
5938 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5940 /* RFC1323: The window in SYN & SYN/ACK segments is
5943 tp
->snd_wnd
= ntohs(th
->window
);
5945 if (!tp
->rx_opt
.wscale_ok
) {
5946 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
5947 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
5950 if (tp
->rx_opt
.saw_tstamp
) {
5951 tp
->rx_opt
.tstamp_ok
= 1;
5952 tp
->tcp_header_len
=
5953 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5954 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5955 tcp_store_ts_recent(tp
);
5957 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5960 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5961 tcp_initialize_rcv_mss(sk
);
5963 /* Remember, tcp_poll() does not lock socket!
5964 * Change state from SYN-SENT only after copied_seq
5965 * is initialized. */
5966 WRITE_ONCE(tp
->copied_seq
, tp
->rcv_nxt
);
5968 smc_check_reset_syn(tp
);
5972 tcp_finish_connect(sk
, skb
);
5974 fastopen_fail
= (tp
->syn_fastopen
|| tp
->syn_data
) &&
5975 tcp_rcv_fastopen_synack(sk
, skb
, &foc
);
5977 if (!sock_flag(sk
, SOCK_DEAD
)) {
5978 sk
->sk_state_change(sk
);
5979 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5983 if (sk
->sk_write_pending
||
5984 icsk
->icsk_accept_queue
.rskq_defer_accept
||
5985 inet_csk_in_pingpong_mode(sk
)) {
5986 /* Save one ACK. Data will be ready after
5987 * several ticks, if write_pending is set.
5989 * It may be deleted, but with this feature tcpdumps
5990 * look so _wonderfully_ clever, that I was not able
5991 * to stand against the temptation 8) --ANK
5993 inet_csk_schedule_ack(sk
);
5994 tcp_enter_quickack_mode(sk
, TCP_MAX_QUICKACKS
);
5995 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
5996 TCP_DELACK_MAX
, TCP_RTO_MAX
);
6007 /* No ACK in the segment */
6011 * "If the RST bit is set
6013 * Otherwise (no ACK) drop the segment and return."
6016 goto discard_and_undo
;
6020 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&&
6021 tcp_paws_reject(&tp
->rx_opt
, 0))
6022 goto discard_and_undo
;
6025 /* We see SYN without ACK. It is attempt of
6026 * simultaneous connect with crossed SYNs.
6027 * Particularly, it can be connect to self.
6029 tcp_set_state(sk
, TCP_SYN_RECV
);
6031 if (tp
->rx_opt
.saw_tstamp
) {
6032 tp
->rx_opt
.tstamp_ok
= 1;
6033 tcp_store_ts_recent(tp
);
6034 tp
->tcp_header_len
=
6035 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
6037 tp
->tcp_header_len
= sizeof(struct tcphdr
);
6040 WRITE_ONCE(tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
+ 1);
6041 WRITE_ONCE(tp
->copied_seq
, tp
->rcv_nxt
);
6042 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
6044 /* RFC1323: The window in SYN & SYN/ACK segments is
6047 tp
->snd_wnd
= ntohs(th
->window
);
6048 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
6049 tp
->max_window
= tp
->snd_wnd
;
6051 tcp_ecn_rcv_syn(tp
, th
);
6054 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
6055 tcp_initialize_rcv_mss(sk
);
6057 tcp_send_synack(sk
);
6059 /* Note, we could accept data and URG from this segment.
6060 * There are no obstacles to make this (except that we must
6061 * either change tcp_recvmsg() to prevent it from returning data
6062 * before 3WHS completes per RFC793, or employ TCP Fast Open).
6064 * However, if we ignore data in ACKless segments sometimes,
6065 * we have no reasons to accept it sometimes.
6066 * Also, seems the code doing it in step6 of tcp_rcv_state_process
6067 * is not flawless. So, discard packet for sanity.
6068 * Uncomment this return to process the data.
6075 /* "fifth, if neither of the SYN or RST bits is set then
6076 * drop the segment and return."
6080 tcp_clear_options(&tp
->rx_opt
);
6081 tp
->rx_opt
.mss_clamp
= saved_clamp
;
6085 tcp_clear_options(&tp
->rx_opt
);
6086 tp
->rx_opt
.mss_clamp
= saved_clamp
;
6090 static void tcp_rcv_synrecv_state_fastopen(struct sock
*sk
)
6092 struct request_sock
*req
;
6094 tcp_try_undo_loss(sk
, false);
6096 /* Reset rtx states to prevent spurious retransmits_timed_out() */
6097 tcp_sk(sk
)->retrans_stamp
= 0;
6098 inet_csk(sk
)->icsk_retransmits
= 0;
6100 /* Once we leave TCP_SYN_RECV or TCP_FIN_WAIT_1,
6101 * we no longer need req so release it.
6103 req
= rcu_dereference_protected(tcp_sk(sk
)->fastopen_rsk
,
6104 lockdep_sock_is_held(sk
));
6105 reqsk_fastopen_remove(sk
, req
, false);
6107 /* Re-arm the timer because data may have been sent out.
6108 * This is similar to the regular data transmission case
6109 * when new data has just been ack'ed.
6111 * (TFO) - we could try to be more aggressive and
6112 * retransmitting any data sooner based on when they
6119 * This function implements the receiving procedure of RFC 793 for
6120 * all states except ESTABLISHED and TIME_WAIT.
6121 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
6122 * address independent.
6125 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
)
6127 struct tcp_sock
*tp
= tcp_sk(sk
);
6128 struct inet_connection_sock
*icsk
= inet_csk(sk
);
6129 const struct tcphdr
*th
= tcp_hdr(skb
);
6130 struct request_sock
*req
;
6134 switch (sk
->sk_state
) {
6148 /* It is possible that we process SYN packets from backlog,
6149 * so we need to make sure to disable BH and RCU right there.
6153 acceptable
= icsk
->icsk_af_ops
->conn_request(sk
, skb
) >= 0;
6165 tp
->rx_opt
.saw_tstamp
= 0;
6166 tcp_mstamp_refresh(tp
);
6167 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
);
6171 /* Do step6 onward by hand. */
6172 tcp_urg(sk
, skb
, th
);
6174 tcp_data_snd_check(sk
);
6178 tcp_mstamp_refresh(tp
);
6179 tp
->rx_opt
.saw_tstamp
= 0;
6180 req
= rcu_dereference_protected(tp
->fastopen_rsk
,
6181 lockdep_sock_is_held(sk
));
6185 WARN_ON_ONCE(sk
->sk_state
!= TCP_SYN_RECV
&&
6186 sk
->sk_state
!= TCP_FIN_WAIT1
);
6188 if (!tcp_check_req(sk
, skb
, req
, true, &req_stolen
))
6192 if (!th
->ack
&& !th
->rst
&& !th
->syn
)
6195 if (!tcp_validate_incoming(sk
, skb
, th
, 0))
6198 /* step 5: check the ACK field */
6199 acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
|
6200 FLAG_UPDATE_TS_RECENT
|
6201 FLAG_NO_CHALLENGE_ACK
) > 0;
6204 if (sk
->sk_state
== TCP_SYN_RECV
)
6205 return 1; /* send one RST */
6206 tcp_send_challenge_ack(sk
, skb
);
6209 switch (sk
->sk_state
) {
6211 tp
->delivered
++; /* SYN-ACK delivery isn't tracked in tcp_ack */
6213 tcp_synack_rtt_meas(sk
, req
);
6216 tcp_rcv_synrecv_state_fastopen(sk
);
6218 tcp_try_undo_spurious_syn(sk
);
6219 tp
->retrans_stamp
= 0;
6220 tcp_init_transfer(sk
, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB
);
6221 WRITE_ONCE(tp
->copied_seq
, tp
->rcv_nxt
);
6224 tcp_set_state(sk
, TCP_ESTABLISHED
);
6225 sk
->sk_state_change(sk
);
6227 /* Note, that this wakeup is only for marginal crossed SYN case.
6228 * Passively open sockets are not waked up, because
6229 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
6232 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
6234 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
6235 tp
->snd_wnd
= ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
;
6236 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
6238 if (tp
->rx_opt
.tstamp_ok
)
6239 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
6241 if (!inet_csk(sk
)->icsk_ca_ops
->cong_control
)
6242 tcp_update_pacing_rate(sk
);
6244 /* Prevent spurious tcp_cwnd_restart() on first data packet */
6245 tp
->lsndtime
= tcp_jiffies32
;
6247 tcp_initialize_rcv_mss(sk
);
6248 tcp_fast_path_on(tp
);
6251 case TCP_FIN_WAIT1
: {
6255 tcp_rcv_synrecv_state_fastopen(sk
);
6257 if (tp
->snd_una
!= tp
->write_seq
)
6260 tcp_set_state(sk
, TCP_FIN_WAIT2
);
6261 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
6265 if (!sock_flag(sk
, SOCK_DEAD
)) {
6266 /* Wake up lingering close() */
6267 sk
->sk_state_change(sk
);
6271 if (tp
->linger2
< 0) {
6273 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
6276 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
6277 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
6278 /* Receive out of order FIN after close() */
6279 if (tp
->syn_fastopen
&& th
->fin
)
6280 tcp_fastopen_active_disable(sk
);
6282 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
6286 tmo
= tcp_fin_time(sk
);
6287 if (tmo
> TCP_TIMEWAIT_LEN
) {
6288 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
6289 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
6290 /* Bad case. We could lose such FIN otherwise.
6291 * It is not a big problem, but it looks confusing
6292 * and not so rare event. We still can lose it now,
6293 * if it spins in bh_lock_sock(), but it is really
6296 inet_csk_reset_keepalive_timer(sk
, tmo
);
6298 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
6305 if (tp
->snd_una
== tp
->write_seq
) {
6306 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
6312 if (tp
->snd_una
== tp
->write_seq
) {
6313 tcp_update_metrics(sk
);
6320 /* step 6: check the URG bit */
6321 tcp_urg(sk
, skb
, th
);
6323 /* step 7: process the segment text */
6324 switch (sk
->sk_state
) {
6325 case TCP_CLOSE_WAIT
:
6328 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
6333 /* RFC 793 says to queue data in these states,
6334 * RFC 1122 says we MUST send a reset.
6335 * BSD 4.4 also does reset.
6337 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
6338 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
6339 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
6340 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
6346 case TCP_ESTABLISHED
:
6347 tcp_data_queue(sk
, skb
);
6352 /* tcp_data could move socket to TIME-WAIT */
6353 if (sk
->sk_state
!= TCP_CLOSE
) {
6354 tcp_data_snd_check(sk
);
6355 tcp_ack_snd_check(sk
);
6364 EXPORT_SYMBOL(tcp_rcv_state_process
);
6366 static inline void pr_drop_req(struct request_sock
*req
, __u16 port
, int family
)
6368 struct inet_request_sock
*ireq
= inet_rsk(req
);
6370 if (family
== AF_INET
)
6371 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6372 &ireq
->ir_rmt_addr
, port
);
6373 #if IS_ENABLED(CONFIG_IPV6)
6374 else if (family
== AF_INET6
)
6375 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6376 &ireq
->ir_v6_rmt_addr
, port
);
6380 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6382 * If we receive a SYN packet with these bits set, it means a
6383 * network is playing bad games with TOS bits. In order to
6384 * avoid possible false congestion notifications, we disable
6385 * TCP ECN negotiation.
6387 * Exception: tcp_ca wants ECN. This is required for DCTCP
6388 * congestion control: Linux DCTCP asserts ECT on all packets,
6389 * including SYN, which is most optimal solution; however,
6390 * others, such as FreeBSD do not.
6392 * Exception: At least one of the reserved bits of the TCP header (th->res1) is
6393 * set, indicating the use of a future TCP extension (such as AccECN). See
6394 * RFC8311 §4.3 which updates RFC3168 to allow the development of such
6397 static void tcp_ecn_create_request(struct request_sock
*req
,
6398 const struct sk_buff
*skb
,
6399 const struct sock
*listen_sk
,
6400 const struct dst_entry
*dst
)
6402 const struct tcphdr
*th
= tcp_hdr(skb
);
6403 const struct net
*net
= sock_net(listen_sk
);
6404 bool th_ecn
= th
->ece
&& th
->cwr
;
6411 ect
= !INET_ECN_is_not_ect(TCP_SKB_CB(skb
)->ip_dsfield
);
6412 ecn_ok_dst
= dst_feature(dst
, DST_FEATURE_ECN_MASK
);
6413 ecn_ok
= net
->ipv4
.sysctl_tcp_ecn
|| ecn_ok_dst
;
6415 if (((!ect
|| th
->res1
) && ecn_ok
) || tcp_ca_needs_ecn(listen_sk
) ||
6416 (ecn_ok_dst
& DST_FEATURE_ECN_CA
) ||
6417 tcp_bpf_ca_needs_ecn((struct sock
*)req
))
6418 inet_rsk(req
)->ecn_ok
= 1;
6421 static void tcp_openreq_init(struct request_sock
*req
,
6422 const struct tcp_options_received
*rx_opt
,
6423 struct sk_buff
*skb
, const struct sock
*sk
)
6425 struct inet_request_sock
*ireq
= inet_rsk(req
);
6427 req
->rsk_rcv_wnd
= 0; /* So that tcp_send_synack() knows! */
6429 tcp_rsk(req
)->rcv_isn
= TCP_SKB_CB(skb
)->seq
;
6430 tcp_rsk(req
)->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
6431 tcp_rsk(req
)->snt_synack
= 0;
6432 tcp_rsk(req
)->last_oow_ack_time
= 0;
6433 req
->mss
= rx_opt
->mss_clamp
;
6434 req
->ts_recent
= rx_opt
->saw_tstamp
? rx_opt
->rcv_tsval
: 0;
6435 ireq
->tstamp_ok
= rx_opt
->tstamp_ok
;
6436 ireq
->sack_ok
= rx_opt
->sack_ok
;
6437 ireq
->snd_wscale
= rx_opt
->snd_wscale
;
6438 ireq
->wscale_ok
= rx_opt
->wscale_ok
;
6441 ireq
->ir_rmt_port
= tcp_hdr(skb
)->source
;
6442 ireq
->ir_num
= ntohs(tcp_hdr(skb
)->dest
);
6443 ireq
->ir_mark
= inet_request_mark(sk
, skb
);
6444 #if IS_ENABLED(CONFIG_SMC)
6445 ireq
->smc_ok
= rx_opt
->smc_ok
;
6449 struct request_sock
*inet_reqsk_alloc(const struct request_sock_ops
*ops
,
6450 struct sock
*sk_listener
,
6451 bool attach_listener
)
6453 struct request_sock
*req
= reqsk_alloc(ops
, sk_listener
,
6457 struct inet_request_sock
*ireq
= inet_rsk(req
);
6459 ireq
->ireq_opt
= NULL
;
6460 #if IS_ENABLED(CONFIG_IPV6)
6461 ireq
->pktopts
= NULL
;
6463 atomic64_set(&ireq
->ir_cookie
, 0);
6464 ireq
->ireq_state
= TCP_NEW_SYN_RECV
;
6465 write_pnet(&ireq
->ireq_net
, sock_net(sk_listener
));
6466 ireq
->ireq_family
= sk_listener
->sk_family
;
6471 EXPORT_SYMBOL(inet_reqsk_alloc
);
6474 * Return true if a syncookie should be sent
6476 static bool tcp_syn_flood_action(const struct sock
*sk
, const char *proto
)
6478 struct request_sock_queue
*queue
= &inet_csk(sk
)->icsk_accept_queue
;
6479 const char *msg
= "Dropping request";
6480 bool want_cookie
= false;
6481 struct net
*net
= sock_net(sk
);
6483 #ifdef CONFIG_SYN_COOKIES
6484 if (net
->ipv4
.sysctl_tcp_syncookies
) {
6485 msg
= "Sending cookies";
6487 __NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPREQQFULLDOCOOKIES
);
6490 __NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPREQQFULLDROP
);
6492 if (!queue
->synflood_warned
&&
6493 net
->ipv4
.sysctl_tcp_syncookies
!= 2 &&
6494 xchg(&queue
->synflood_warned
, 1) == 0)
6495 net_info_ratelimited("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6496 proto
, sk
->sk_num
, msg
);
6501 static void tcp_reqsk_record_syn(const struct sock
*sk
,
6502 struct request_sock
*req
,
6503 const struct sk_buff
*skb
)
6505 if (tcp_sk(sk
)->save_syn
) {
6506 u32 len
= skb_network_header_len(skb
) + tcp_hdrlen(skb
);
6509 copy
= kmalloc(len
+ sizeof(u32
), GFP_ATOMIC
);
6512 memcpy(©
[1], skb_network_header(skb
), len
);
6513 req
->saved_syn
= copy
;
6518 /* If a SYN cookie is required and supported, returns a clamped MSS value to be
6519 * used for SYN cookie generation.
6521 u16
tcp_get_syncookie_mss(struct request_sock_ops
*rsk_ops
,
6522 const struct tcp_request_sock_ops
*af_ops
,
6523 struct sock
*sk
, struct tcphdr
*th
)
6525 struct tcp_sock
*tp
= tcp_sk(sk
);
6528 if (sock_net(sk
)->ipv4
.sysctl_tcp_syncookies
!= 2 &&
6529 !inet_csk_reqsk_queue_is_full(sk
))
6532 if (!tcp_syn_flood_action(sk
, rsk_ops
->slab_name
))
6535 if (sk_acceptq_is_full(sk
)) {
6536 NET_INC_STATS(sock_net(sk
), LINUX_MIB_LISTENOVERFLOWS
);
6540 mss
= tcp_parse_mss_option(th
, tp
->rx_opt
.user_mss
);
6542 mss
= af_ops
->mss_clamp
;
6546 EXPORT_SYMBOL_GPL(tcp_get_syncookie_mss
);
6548 int tcp_conn_request(struct request_sock_ops
*rsk_ops
,
6549 const struct tcp_request_sock_ops
*af_ops
,
6550 struct sock
*sk
, struct sk_buff
*skb
)
6552 struct tcp_fastopen_cookie foc
= { .len
= -1 };
6553 __u32 isn
= TCP_SKB_CB(skb
)->tcp_tw_isn
;
6554 struct tcp_options_received tmp_opt
;
6555 struct tcp_sock
*tp
= tcp_sk(sk
);
6556 struct net
*net
= sock_net(sk
);
6557 struct sock
*fastopen_sk
= NULL
;
6558 struct request_sock
*req
;
6559 bool want_cookie
= false;
6560 struct dst_entry
*dst
;
6563 /* TW buckets are converted to open requests without
6564 * limitations, they conserve resources and peer is
6565 * evidently real one.
6567 if ((net
->ipv4
.sysctl_tcp_syncookies
== 2 ||
6568 inet_csk_reqsk_queue_is_full(sk
)) && !isn
) {
6569 want_cookie
= tcp_syn_flood_action(sk
, rsk_ops
->slab_name
);
6574 if (sk_acceptq_is_full(sk
)) {
6575 NET_INC_STATS(sock_net(sk
), LINUX_MIB_LISTENOVERFLOWS
);
6579 req
= inet_reqsk_alloc(rsk_ops
, sk
, !want_cookie
);
6583 tcp_rsk(req
)->af_specific
= af_ops
;
6584 tcp_rsk(req
)->ts_off
= 0;
6586 tcp_clear_options(&tmp_opt
);
6587 tmp_opt
.mss_clamp
= af_ops
->mss_clamp
;
6588 tmp_opt
.user_mss
= tp
->rx_opt
.user_mss
;
6589 tcp_parse_options(sock_net(sk
), skb
, &tmp_opt
, 0,
6590 want_cookie
? NULL
: &foc
);
6592 if (want_cookie
&& !tmp_opt
.saw_tstamp
)
6593 tcp_clear_options(&tmp_opt
);
6595 if (IS_ENABLED(CONFIG_SMC
) && want_cookie
)
6598 tmp_opt
.tstamp_ok
= tmp_opt
.saw_tstamp
;
6599 tcp_openreq_init(req
, &tmp_opt
, skb
, sk
);
6600 inet_rsk(req
)->no_srccheck
= inet_sk(sk
)->transparent
;
6602 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6603 inet_rsk(req
)->ir_iif
= inet_request_bound_dev_if(sk
, skb
);
6605 af_ops
->init_req(req
, sk
, skb
);
6607 if (security_inet_conn_request(sk
, skb
, req
))
6610 if (tmp_opt
.tstamp_ok
)
6611 tcp_rsk(req
)->ts_off
= af_ops
->init_ts_off(net
, skb
);
6613 dst
= af_ops
->route_req(sk
, &fl
, req
);
6617 if (!want_cookie
&& !isn
) {
6618 /* Kill the following clause, if you dislike this way. */
6619 if (!net
->ipv4
.sysctl_tcp_syncookies
&&
6620 (net
->ipv4
.sysctl_max_syn_backlog
- inet_csk_reqsk_queue_len(sk
) <
6621 (net
->ipv4
.sysctl_max_syn_backlog
>> 2)) &&
6622 !tcp_peer_is_proven(req
, dst
)) {
6623 /* Without syncookies last quarter of
6624 * backlog is filled with destinations,
6625 * proven to be alive.
6626 * It means that we continue to communicate
6627 * to destinations, already remembered
6628 * to the moment of synflood.
6630 pr_drop_req(req
, ntohs(tcp_hdr(skb
)->source
),
6632 goto drop_and_release
;
6635 isn
= af_ops
->init_seq(skb
);
6638 tcp_ecn_create_request(req
, skb
, sk
, dst
);
6641 isn
= cookie_init_sequence(af_ops
, sk
, skb
, &req
->mss
);
6642 req
->cookie_ts
= tmp_opt
.tstamp_ok
;
6643 if (!tmp_opt
.tstamp_ok
)
6644 inet_rsk(req
)->ecn_ok
= 0;
6647 tcp_rsk(req
)->snt_isn
= isn
;
6648 tcp_rsk(req
)->txhash
= net_tx_rndhash();
6649 tcp_openreq_init_rwin(req
, sk
, dst
);
6650 sk_rx_queue_set(req_to_sk(req
), skb
);
6652 tcp_reqsk_record_syn(sk
, req
, skb
);
6653 fastopen_sk
= tcp_try_fastopen(sk
, skb
, req
, &foc
, dst
);
6656 af_ops
->send_synack(fastopen_sk
, dst
, &fl
, req
,
6657 &foc
, TCP_SYNACK_FASTOPEN
);
6658 /* Add the child socket directly into the accept queue */
6659 if (!inet_csk_reqsk_queue_add(sk
, req
, fastopen_sk
)) {
6660 reqsk_fastopen_remove(fastopen_sk
, req
, false);
6661 bh_unlock_sock(fastopen_sk
);
6662 sock_put(fastopen_sk
);
6665 sk
->sk_data_ready(sk
);
6666 bh_unlock_sock(fastopen_sk
);
6667 sock_put(fastopen_sk
);
6669 tcp_rsk(req
)->tfo_listener
= false;
6671 inet_csk_reqsk_queue_hash_add(sk
, req
,
6672 tcp_timeout_init((struct sock
*)req
));
6673 af_ops
->send_synack(sk
, dst
, &fl
, req
, &foc
,
6674 !want_cookie
? TCP_SYNACK_NORMAL
:
6692 EXPORT_SYMBOL(tcp_conn_request
);