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>
82 #include <net/mptcp.h>
84 int sysctl_tcp_max_orphans __read_mostly
= NR_FILE
;
86 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
87 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
88 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
89 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
90 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
91 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
92 #define FLAG_ECE 0x40 /* ECE in this ACK */
93 #define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */
94 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
95 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
96 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
97 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
98 #define FLAG_SET_XMIT_TIMER 0x1000 /* Set TLP or RTO timer */
99 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
100 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
101 #define FLAG_NO_CHALLENGE_ACK 0x8000 /* do not call tcp_send_challenge_ack() */
102 #define FLAG_ACK_MAYBE_DELAYED 0x10000 /* Likely a delayed ACK */
103 #define FLAG_DSACK_TLP 0x20000 /* DSACK for tail loss probe */
105 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
106 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
107 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE|FLAG_DSACKING_ACK)
108 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
110 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
111 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
113 #define REXMIT_NONE 0 /* no loss recovery to do */
114 #define REXMIT_LOST 1 /* retransmit packets marked lost */
115 #define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */
117 #if IS_ENABLED(CONFIG_TLS_DEVICE)
118 static DEFINE_STATIC_KEY_DEFERRED_FALSE(clean_acked_data_enabled
, HZ
);
120 void clean_acked_data_enable(struct inet_connection_sock
*icsk
,
121 void (*cad
)(struct sock
*sk
, u32 ack_seq
))
123 icsk
->icsk_clean_acked
= cad
;
124 static_branch_deferred_inc(&clean_acked_data_enabled
);
126 EXPORT_SYMBOL_GPL(clean_acked_data_enable
);
128 void clean_acked_data_disable(struct inet_connection_sock
*icsk
)
130 static_branch_slow_dec_deferred(&clean_acked_data_enabled
);
131 icsk
->icsk_clean_acked
= NULL
;
133 EXPORT_SYMBOL_GPL(clean_acked_data_disable
);
135 void clean_acked_data_flush(void)
137 static_key_deferred_flush(&clean_acked_data_enabled
);
139 EXPORT_SYMBOL_GPL(clean_acked_data_flush
);
142 #ifdef CONFIG_CGROUP_BPF
143 static void bpf_skops_parse_hdr(struct sock
*sk
, struct sk_buff
*skb
)
145 bool unknown_opt
= tcp_sk(sk
)->rx_opt
.saw_unknown
&&
146 BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk
),
147 BPF_SOCK_OPS_PARSE_UNKNOWN_HDR_OPT_CB_FLAG
);
148 bool parse_all_opt
= BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk
),
149 BPF_SOCK_OPS_PARSE_ALL_HDR_OPT_CB_FLAG
);
150 struct bpf_sock_ops_kern sock_ops
;
152 if (likely(!unknown_opt
&& !parse_all_opt
))
155 /* The skb will be handled in the
156 * bpf_skops_established() or
157 * bpf_skops_write_hdr_opt().
159 switch (sk
->sk_state
) {
166 sock_owned_by_me(sk
);
168 memset(&sock_ops
, 0, offsetof(struct bpf_sock_ops_kern
, temp
));
169 sock_ops
.op
= BPF_SOCK_OPS_PARSE_HDR_OPT_CB
;
170 sock_ops
.is_fullsock
= 1;
172 bpf_skops_init_skb(&sock_ops
, skb
, tcp_hdrlen(skb
));
174 BPF_CGROUP_RUN_PROG_SOCK_OPS(&sock_ops
);
177 static void bpf_skops_established(struct sock
*sk
, int bpf_op
,
180 struct bpf_sock_ops_kern sock_ops
;
182 sock_owned_by_me(sk
);
184 memset(&sock_ops
, 0, offsetof(struct bpf_sock_ops_kern
, temp
));
185 sock_ops
.op
= bpf_op
;
186 sock_ops
.is_fullsock
= 1;
188 /* sk with TCP_REPAIR_ON does not have skb in tcp_finish_connect */
190 bpf_skops_init_skb(&sock_ops
, skb
, tcp_hdrlen(skb
));
192 BPF_CGROUP_RUN_PROG_SOCK_OPS(&sock_ops
);
195 static void bpf_skops_parse_hdr(struct sock
*sk
, struct sk_buff
*skb
)
199 static void bpf_skops_established(struct sock
*sk
, int bpf_op
,
205 static void tcp_gro_dev_warn(struct sock
*sk
, const struct sk_buff
*skb
,
208 static bool __once __read_mostly
;
211 struct net_device
*dev
;
216 dev
= dev_get_by_index_rcu(sock_net(sk
), skb
->skb_iif
);
217 if (!dev
|| len
>= dev
->mtu
)
218 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n",
219 dev
? dev
->name
: "Unknown driver");
224 /* Adapt the MSS value used to make delayed ack decision to the
227 static void tcp_measure_rcv_mss(struct sock
*sk
, const struct sk_buff
*skb
)
229 struct inet_connection_sock
*icsk
= inet_csk(sk
);
230 const unsigned int lss
= icsk
->icsk_ack
.last_seg_size
;
233 icsk
->icsk_ack
.last_seg_size
= 0;
235 /* skb->len may jitter because of SACKs, even if peer
236 * sends good full-sized frames.
238 len
= skb_shinfo(skb
)->gso_size
? : skb
->len
;
239 if (len
>= icsk
->icsk_ack
.rcv_mss
) {
240 icsk
->icsk_ack
.rcv_mss
= min_t(unsigned int, len
,
242 /* Account for possibly-removed options */
243 if (unlikely(len
> icsk
->icsk_ack
.rcv_mss
+
244 MAX_TCP_OPTION_SPACE
))
245 tcp_gro_dev_warn(sk
, skb
, len
);
247 /* Otherwise, we make more careful check taking into account,
248 * that SACKs block is variable.
250 * "len" is invariant segment length, including TCP header.
252 len
+= skb
->data
- skb_transport_header(skb
);
253 if (len
>= TCP_MSS_DEFAULT
+ sizeof(struct tcphdr
) ||
254 /* If PSH is not set, packet should be
255 * full sized, provided peer TCP is not badly broken.
256 * This observation (if it is correct 8)) allows
257 * to handle super-low mtu links fairly.
259 (len
>= TCP_MIN_MSS
+ sizeof(struct tcphdr
) &&
260 !(tcp_flag_word(tcp_hdr(skb
)) & TCP_REMNANT
))) {
261 /* Subtract also invariant (if peer is RFC compliant),
262 * tcp header plus fixed timestamp option length.
263 * Resulting "len" is MSS free of SACK jitter.
265 len
-= tcp_sk(sk
)->tcp_header_len
;
266 icsk
->icsk_ack
.last_seg_size
= len
;
268 icsk
->icsk_ack
.rcv_mss
= len
;
272 if (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
)
273 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED2
;
274 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
278 static void tcp_incr_quickack(struct sock
*sk
, unsigned int max_quickacks
)
280 struct inet_connection_sock
*icsk
= inet_csk(sk
);
281 unsigned int quickacks
= tcp_sk(sk
)->rcv_wnd
/ (2 * icsk
->icsk_ack
.rcv_mss
);
285 quickacks
= min(quickacks
, max_quickacks
);
286 if (quickacks
> icsk
->icsk_ack
.quick
)
287 icsk
->icsk_ack
.quick
= quickacks
;
290 void tcp_enter_quickack_mode(struct sock
*sk
, unsigned int max_quickacks
)
292 struct inet_connection_sock
*icsk
= inet_csk(sk
);
294 tcp_incr_quickack(sk
, max_quickacks
);
295 inet_csk_exit_pingpong_mode(sk
);
296 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
298 EXPORT_SYMBOL(tcp_enter_quickack_mode
);
300 /* Send ACKs quickly, if "quick" count is not exhausted
301 * and the session is not interactive.
304 static bool tcp_in_quickack_mode(struct sock
*sk
)
306 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
307 const struct dst_entry
*dst
= __sk_dst_get(sk
);
309 return (dst
&& dst_metric(dst
, RTAX_QUICKACK
)) ||
310 (icsk
->icsk_ack
.quick
&& !inet_csk_in_pingpong_mode(sk
));
313 static void tcp_ecn_queue_cwr(struct tcp_sock
*tp
)
315 if (tp
->ecn_flags
& TCP_ECN_OK
)
316 tp
->ecn_flags
|= TCP_ECN_QUEUE_CWR
;
319 static void tcp_ecn_accept_cwr(struct sock
*sk
, const struct sk_buff
*skb
)
321 if (tcp_hdr(skb
)->cwr
) {
322 tcp_sk(sk
)->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
324 /* If the sender is telling us it has entered CWR, then its
325 * cwnd may be very low (even just 1 packet), so we should ACK
328 if (TCP_SKB_CB(skb
)->seq
!= TCP_SKB_CB(skb
)->end_seq
)
329 inet_csk(sk
)->icsk_ack
.pending
|= ICSK_ACK_NOW
;
333 static void tcp_ecn_withdraw_cwr(struct tcp_sock
*tp
)
335 tp
->ecn_flags
&= ~TCP_ECN_QUEUE_CWR
;
338 static void __tcp_ecn_check_ce(struct sock
*sk
, const struct sk_buff
*skb
)
340 struct tcp_sock
*tp
= tcp_sk(sk
);
342 switch (TCP_SKB_CB(skb
)->ip_dsfield
& INET_ECN_MASK
) {
343 case INET_ECN_NOT_ECT
:
344 /* Funny extension: if ECT is not set on a segment,
345 * and we already seen ECT on a previous segment,
346 * it is probably a retransmit.
348 if (tp
->ecn_flags
& TCP_ECN_SEEN
)
349 tcp_enter_quickack_mode(sk
, 2);
352 if (tcp_ca_needs_ecn(sk
))
353 tcp_ca_event(sk
, CA_EVENT_ECN_IS_CE
);
355 if (!(tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)) {
356 /* Better not delay acks, sender can have a very low cwnd */
357 tcp_enter_quickack_mode(sk
, 2);
358 tp
->ecn_flags
|= TCP_ECN_DEMAND_CWR
;
360 tp
->ecn_flags
|= TCP_ECN_SEEN
;
363 if (tcp_ca_needs_ecn(sk
))
364 tcp_ca_event(sk
, CA_EVENT_ECN_NO_CE
);
365 tp
->ecn_flags
|= TCP_ECN_SEEN
;
370 static void tcp_ecn_check_ce(struct sock
*sk
, const struct sk_buff
*skb
)
372 if (tcp_sk(sk
)->ecn_flags
& TCP_ECN_OK
)
373 __tcp_ecn_check_ce(sk
, skb
);
376 static void tcp_ecn_rcv_synack(struct tcp_sock
*tp
, const struct tcphdr
*th
)
378 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| th
->cwr
))
379 tp
->ecn_flags
&= ~TCP_ECN_OK
;
382 static void tcp_ecn_rcv_syn(struct tcp_sock
*tp
, const struct tcphdr
*th
)
384 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| !th
->cwr
))
385 tp
->ecn_flags
&= ~TCP_ECN_OK
;
388 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock
*tp
, const struct tcphdr
*th
)
390 if (th
->ece
&& !th
->syn
&& (tp
->ecn_flags
& TCP_ECN_OK
))
395 /* Buffer size and advertised window tuning.
397 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
400 static void tcp_sndbuf_expand(struct sock
*sk
)
402 const struct tcp_sock
*tp
= tcp_sk(sk
);
403 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
407 /* Worst case is non GSO/TSO : each frame consumes one skb
408 * and skb->head is kmalloced using power of two area of memory
410 per_mss
= max_t(u32
, tp
->rx_opt
.mss_clamp
, tp
->mss_cache
) +
412 SKB_DATA_ALIGN(sizeof(struct skb_shared_info
));
414 per_mss
= roundup_pow_of_two(per_mss
) +
415 SKB_DATA_ALIGN(sizeof(struct sk_buff
));
417 nr_segs
= max_t(u32
, TCP_INIT_CWND
, tcp_snd_cwnd(tp
));
418 nr_segs
= max_t(u32
, nr_segs
, tp
->reordering
+ 1);
420 /* Fast Recovery (RFC 5681 3.2) :
421 * Cubic needs 1.7 factor, rounded to 2 to include
422 * extra cushion (application might react slowly to EPOLLOUT)
424 sndmem
= ca_ops
->sndbuf_expand
? ca_ops
->sndbuf_expand(sk
) : 2;
425 sndmem
*= nr_segs
* per_mss
;
427 if (sk
->sk_sndbuf
< sndmem
)
428 WRITE_ONCE(sk
->sk_sndbuf
,
429 min(sndmem
, sock_net(sk
)->ipv4
.sysctl_tcp_wmem
[2]));
432 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
434 * All tcp_full_space() is split to two parts: "network" buffer, allocated
435 * forward and advertised in receiver window (tp->rcv_wnd) and
436 * "application buffer", required to isolate scheduling/application
437 * latencies from network.
438 * window_clamp is maximal advertised window. It can be less than
439 * tcp_full_space(), in this case tcp_full_space() - window_clamp
440 * is reserved for "application" buffer. The less window_clamp is
441 * the smoother our behaviour from viewpoint of network, but the lower
442 * throughput and the higher sensitivity of the connection to losses. 8)
444 * rcv_ssthresh is more strict window_clamp used at "slow start"
445 * phase to predict further behaviour of this connection.
446 * It is used for two goals:
447 * - to enforce header prediction at sender, even when application
448 * requires some significant "application buffer". It is check #1.
449 * - to prevent pruning of receive queue because of misprediction
450 * of receiver window. Check #2.
452 * The scheme does not work when sender sends good segments opening
453 * window and then starts to feed us spaghetti. But it should work
454 * in common situations. Otherwise, we have to rely on queue collapsing.
457 /* Slow part of check#2. */
458 static int __tcp_grow_window(const struct sock
*sk
, const struct sk_buff
*skb
,
459 unsigned int skbtruesize
)
461 struct tcp_sock
*tp
= tcp_sk(sk
);
463 int truesize
= tcp_win_from_space(sk
, skbtruesize
) >> 1;
464 int window
= tcp_win_from_space(sk
, sock_net(sk
)->ipv4
.sysctl_tcp_rmem
[2]) >> 1;
466 while (tp
->rcv_ssthresh
<= window
) {
467 if (truesize
<= skb
->len
)
468 return 2 * inet_csk(sk
)->icsk_ack
.rcv_mss
;
476 /* Even if skb appears to have a bad len/truesize ratio, TCP coalescing
477 * can play nice with us, as sk_buff and skb->head might be either
478 * freed or shared with up to MAX_SKB_FRAGS segments.
479 * Only give a boost to drivers using page frag(s) to hold the frame(s),
480 * and if no payload was pulled in skb->head before reaching us.
482 static u32
truesize_adjust(bool adjust
, const struct sk_buff
*skb
)
484 u32 truesize
= skb
->truesize
;
486 if (adjust
&& !skb_headlen(skb
)) {
487 truesize
-= SKB_TRUESIZE(skb_end_offset(skb
));
488 /* paranoid check, some drivers might be buggy */
489 if (unlikely((int)truesize
< (int)skb
->len
))
490 truesize
= skb
->truesize
;
495 static void tcp_grow_window(struct sock
*sk
, const struct sk_buff
*skb
,
498 struct tcp_sock
*tp
= tcp_sk(sk
);
501 room
= min_t(int, tp
->window_clamp
, tcp_space(sk
)) - tp
->rcv_ssthresh
;
504 if (room
> 0 && !tcp_under_memory_pressure(sk
)) {
505 unsigned int truesize
= truesize_adjust(adjust
, skb
);
508 /* Check #2. Increase window, if skb with such overhead
509 * will fit to rcvbuf in future.
511 if (tcp_win_from_space(sk
, truesize
) <= skb
->len
)
512 incr
= 2 * tp
->advmss
;
514 incr
= __tcp_grow_window(sk
, skb
, truesize
);
517 incr
= max_t(int, incr
, 2 * skb
->len
);
518 tp
->rcv_ssthresh
+= min(room
, incr
);
519 inet_csk(sk
)->icsk_ack
.quick
|= 1;
524 /* 3. Try to fixup all. It is made immediately after connection enters
527 static void tcp_init_buffer_space(struct sock
*sk
)
529 int tcp_app_win
= READ_ONCE(sock_net(sk
)->ipv4
.sysctl_tcp_app_win
);
530 struct tcp_sock
*tp
= tcp_sk(sk
);
533 if (!(sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
))
534 tcp_sndbuf_expand(sk
);
536 tcp_mstamp_refresh(tp
);
537 tp
->rcvq_space
.time
= tp
->tcp_mstamp
;
538 tp
->rcvq_space
.seq
= tp
->copied_seq
;
540 maxwin
= tcp_full_space(sk
);
542 if (tp
->window_clamp
>= maxwin
) {
543 tp
->window_clamp
= maxwin
;
545 if (tcp_app_win
&& maxwin
> 4 * tp
->advmss
)
546 tp
->window_clamp
= max(maxwin
-
547 (maxwin
>> tcp_app_win
),
551 /* Force reservation of one segment. */
553 tp
->window_clamp
> 2 * tp
->advmss
&&
554 tp
->window_clamp
+ tp
->advmss
> maxwin
)
555 tp
->window_clamp
= max(2 * tp
->advmss
, maxwin
- tp
->advmss
);
557 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, tp
->window_clamp
);
558 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
559 tp
->rcvq_space
.space
= min3(tp
->rcv_ssthresh
, tp
->rcv_wnd
,
560 (u32
)TCP_INIT_CWND
* tp
->advmss
);
563 /* 4. Recalculate window clamp after socket hit its memory bounds. */
564 static void tcp_clamp_window(struct sock
*sk
)
566 struct tcp_sock
*tp
= tcp_sk(sk
);
567 struct inet_connection_sock
*icsk
= inet_csk(sk
);
568 struct net
*net
= sock_net(sk
);
570 icsk
->icsk_ack
.quick
= 0;
572 if (sk
->sk_rcvbuf
< net
->ipv4
.sysctl_tcp_rmem
[2] &&
573 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
) &&
574 !tcp_under_memory_pressure(sk
) &&
575 sk_memory_allocated(sk
) < sk_prot_mem_limits(sk
, 0)) {
576 WRITE_ONCE(sk
->sk_rcvbuf
,
577 min(atomic_read(&sk
->sk_rmem_alloc
),
578 net
->ipv4
.sysctl_tcp_rmem
[2]));
580 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
)
581 tp
->rcv_ssthresh
= min(tp
->window_clamp
, 2U * tp
->advmss
);
584 /* Initialize RCV_MSS value.
585 * RCV_MSS is an our guess about MSS used by the peer.
586 * We haven't any direct information about the MSS.
587 * It's better to underestimate the RCV_MSS rather than overestimate.
588 * Overestimations make us ACKing less frequently than needed.
589 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
591 void tcp_initialize_rcv_mss(struct sock
*sk
)
593 const struct tcp_sock
*tp
= tcp_sk(sk
);
594 unsigned int hint
= min_t(unsigned int, tp
->advmss
, tp
->mss_cache
);
596 hint
= min(hint
, tp
->rcv_wnd
/ 2);
597 hint
= min(hint
, TCP_MSS_DEFAULT
);
598 hint
= max(hint
, TCP_MIN_MSS
);
600 inet_csk(sk
)->icsk_ack
.rcv_mss
= hint
;
602 EXPORT_SYMBOL(tcp_initialize_rcv_mss
);
604 /* Receiver "autotuning" code.
606 * The algorithm for RTT estimation w/o timestamps is based on
607 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
608 * <https://public.lanl.gov/radiant/pubs.html#DRS>
610 * More detail on this code can be found at
611 * <http://staff.psc.edu/jheffner/>,
612 * though this reference is out of date. A new paper
615 static void tcp_rcv_rtt_update(struct tcp_sock
*tp
, u32 sample
, int win_dep
)
617 u32 new_sample
= tp
->rcv_rtt_est
.rtt_us
;
620 if (new_sample
!= 0) {
621 /* If we sample in larger samples in the non-timestamp
622 * case, we could grossly overestimate the RTT especially
623 * with chatty applications or bulk transfer apps which
624 * are stalled on filesystem I/O.
626 * Also, since we are only going for a minimum in the
627 * non-timestamp case, we do not smooth things out
628 * else with timestamps disabled convergence takes too
632 m
-= (new_sample
>> 3);
640 /* No previous measure. */
644 tp
->rcv_rtt_est
.rtt_us
= new_sample
;
647 static inline void tcp_rcv_rtt_measure(struct tcp_sock
*tp
)
651 if (tp
->rcv_rtt_est
.time
== 0)
653 if (before(tp
->rcv_nxt
, tp
->rcv_rtt_est
.seq
))
655 delta_us
= tcp_stamp_us_delta(tp
->tcp_mstamp
, tp
->rcv_rtt_est
.time
);
658 tcp_rcv_rtt_update(tp
, delta_us
, 1);
661 tp
->rcv_rtt_est
.seq
= tp
->rcv_nxt
+ tp
->rcv_wnd
;
662 tp
->rcv_rtt_est
.time
= tp
->tcp_mstamp
;
665 static inline void tcp_rcv_rtt_measure_ts(struct sock
*sk
,
666 const struct sk_buff
*skb
)
668 struct tcp_sock
*tp
= tcp_sk(sk
);
670 if (tp
->rx_opt
.rcv_tsecr
== tp
->rcv_rtt_last_tsecr
)
672 tp
->rcv_rtt_last_tsecr
= tp
->rx_opt
.rcv_tsecr
;
674 if (TCP_SKB_CB(skb
)->end_seq
-
675 TCP_SKB_CB(skb
)->seq
>= inet_csk(sk
)->icsk_ack
.rcv_mss
) {
676 u32 delta
= tcp_time_stamp(tp
) - tp
->rx_opt
.rcv_tsecr
;
679 if (likely(delta
< INT_MAX
/ (USEC_PER_SEC
/ TCP_TS_HZ
))) {
682 delta_us
= delta
* (USEC_PER_SEC
/ TCP_TS_HZ
);
683 tcp_rcv_rtt_update(tp
, delta_us
, 0);
689 * This function should be called every time data is copied to user space.
690 * It calculates the appropriate TCP receive buffer space.
692 void tcp_rcv_space_adjust(struct sock
*sk
)
694 struct tcp_sock
*tp
= tcp_sk(sk
);
698 trace_tcp_rcv_space_adjust(sk
);
700 tcp_mstamp_refresh(tp
);
701 time
= tcp_stamp_us_delta(tp
->tcp_mstamp
, tp
->rcvq_space
.time
);
702 if (time
< (tp
->rcv_rtt_est
.rtt_us
>> 3) || tp
->rcv_rtt_est
.rtt_us
== 0)
705 /* Number of bytes copied to user in last RTT */
706 copied
= tp
->copied_seq
- tp
->rcvq_space
.seq
;
707 if (copied
<= tp
->rcvq_space
.space
)
711 * copied = bytes received in previous RTT, our base window
712 * To cope with packet losses, we need a 2x factor
713 * To cope with slow start, and sender growing its cwin by 100 %
714 * every RTT, we need a 4x factor, because the ACK we are sending
715 * now is for the next RTT, not the current one :
716 * <prev RTT . ><current RTT .. ><next RTT .... >
719 if (READ_ONCE(sock_net(sk
)->ipv4
.sysctl_tcp_moderate_rcvbuf
) &&
720 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
)) {
724 /* minimal window to cope with packet losses, assuming
725 * steady state. Add some cushion because of small variations.
727 rcvwin
= ((u64
)copied
<< 1) + 16 * tp
->advmss
;
729 /* Accommodate for sender rate increase (eg. slow start) */
730 grow
= rcvwin
* (copied
- tp
->rcvq_space
.space
);
731 do_div(grow
, tp
->rcvq_space
.space
);
732 rcvwin
+= (grow
<< 1);
734 rcvmem
= SKB_TRUESIZE(tp
->advmss
+ MAX_TCP_HEADER
);
735 while (tcp_win_from_space(sk
, rcvmem
) < tp
->advmss
)
738 do_div(rcvwin
, tp
->advmss
);
739 rcvbuf
= min_t(u64
, rcvwin
* rcvmem
,
740 sock_net(sk
)->ipv4
.sysctl_tcp_rmem
[2]);
741 if (rcvbuf
> sk
->sk_rcvbuf
) {
742 WRITE_ONCE(sk
->sk_rcvbuf
, rcvbuf
);
744 /* Make the window clamp follow along. */
745 tp
->window_clamp
= tcp_win_from_space(sk
, rcvbuf
);
748 tp
->rcvq_space
.space
= copied
;
751 tp
->rcvq_space
.seq
= tp
->copied_seq
;
752 tp
->rcvq_space
.time
= tp
->tcp_mstamp
;
755 /* There is something which you must keep in mind when you analyze the
756 * behavior of the tp->ato delayed ack timeout interval. When a
757 * connection starts up, we want to ack as quickly as possible. The
758 * problem is that "good" TCP's do slow start at the beginning of data
759 * transmission. The means that until we send the first few ACK's the
760 * sender will sit on his end and only queue most of his data, because
761 * he can only send snd_cwnd unacked packets at any given time. For
762 * each ACK we send, he increments snd_cwnd and transmits more of his
765 static void tcp_event_data_recv(struct sock
*sk
, struct sk_buff
*skb
)
767 struct tcp_sock
*tp
= tcp_sk(sk
);
768 struct inet_connection_sock
*icsk
= inet_csk(sk
);
771 inet_csk_schedule_ack(sk
);
773 tcp_measure_rcv_mss(sk
, skb
);
775 tcp_rcv_rtt_measure(tp
);
779 if (!icsk
->icsk_ack
.ato
) {
780 /* The _first_ data packet received, initialize
781 * delayed ACK engine.
783 tcp_incr_quickack(sk
, TCP_MAX_QUICKACKS
);
784 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
786 int m
= now
- icsk
->icsk_ack
.lrcvtime
;
788 if (m
<= TCP_ATO_MIN
/ 2) {
789 /* The fastest case is the first. */
790 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + TCP_ATO_MIN
/ 2;
791 } else if (m
< icsk
->icsk_ack
.ato
) {
792 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + m
;
793 if (icsk
->icsk_ack
.ato
> icsk
->icsk_rto
)
794 icsk
->icsk_ack
.ato
= icsk
->icsk_rto
;
795 } else if (m
> icsk
->icsk_rto
) {
796 /* Too long gap. Apparently sender failed to
797 * restart window, so that we send ACKs quickly.
799 tcp_incr_quickack(sk
, TCP_MAX_QUICKACKS
);
803 icsk
->icsk_ack
.lrcvtime
= now
;
805 tcp_ecn_check_ce(sk
, skb
);
808 tcp_grow_window(sk
, skb
, true);
811 /* Called to compute a smoothed rtt estimate. The data fed to this
812 * routine either comes from timestamps, or from segments that were
813 * known _not_ to have been retransmitted [see Karn/Partridge
814 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
815 * piece by Van Jacobson.
816 * NOTE: the next three routines used to be one big routine.
817 * To save cycles in the RFC 1323 implementation it was better to break
818 * it up into three procedures. -- erics
820 static void tcp_rtt_estimator(struct sock
*sk
, long mrtt_us
)
822 struct tcp_sock
*tp
= tcp_sk(sk
);
823 long m
= mrtt_us
; /* RTT */
824 u32 srtt
= tp
->srtt_us
;
826 /* The following amusing code comes from Jacobson's
827 * article in SIGCOMM '88. Note that rtt and mdev
828 * are scaled versions of rtt and mean deviation.
829 * This is designed to be as fast as possible
830 * m stands for "measurement".
832 * On a 1990 paper the rto value is changed to:
833 * RTO = rtt + 4 * mdev
835 * Funny. This algorithm seems to be very broken.
836 * These formulae increase RTO, when it should be decreased, increase
837 * too slowly, when it should be increased quickly, decrease too quickly
838 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
839 * does not matter how to _calculate_ it. Seems, it was trap
840 * that VJ failed to avoid. 8)
843 m
-= (srtt
>> 3); /* m is now error in rtt est */
844 srtt
+= m
; /* rtt = 7/8 rtt + 1/8 new */
846 m
= -m
; /* m is now abs(error) */
847 m
-= (tp
->mdev_us
>> 2); /* similar update on mdev */
848 /* This is similar to one of Eifel findings.
849 * Eifel blocks mdev updates when rtt decreases.
850 * This solution is a bit different: we use finer gain
851 * for mdev in this case (alpha*beta).
852 * Like Eifel it also prevents growth of rto,
853 * but also it limits too fast rto decreases,
854 * happening in pure Eifel.
859 m
-= (tp
->mdev_us
>> 2); /* similar update on mdev */
861 tp
->mdev_us
+= m
; /* mdev = 3/4 mdev + 1/4 new */
862 if (tp
->mdev_us
> tp
->mdev_max_us
) {
863 tp
->mdev_max_us
= tp
->mdev_us
;
864 if (tp
->mdev_max_us
> tp
->rttvar_us
)
865 tp
->rttvar_us
= tp
->mdev_max_us
;
867 if (after(tp
->snd_una
, tp
->rtt_seq
)) {
868 if (tp
->mdev_max_us
< tp
->rttvar_us
)
869 tp
->rttvar_us
-= (tp
->rttvar_us
- tp
->mdev_max_us
) >> 2;
870 tp
->rtt_seq
= tp
->snd_nxt
;
871 tp
->mdev_max_us
= tcp_rto_min_us(sk
);
876 /* no previous measure. */
877 srtt
= m
<< 3; /* take the measured time to be rtt */
878 tp
->mdev_us
= m
<< 1; /* make sure rto = 3*rtt */
879 tp
->rttvar_us
= max(tp
->mdev_us
, tcp_rto_min_us(sk
));
880 tp
->mdev_max_us
= tp
->rttvar_us
;
881 tp
->rtt_seq
= tp
->snd_nxt
;
885 tp
->srtt_us
= max(1U, srtt
);
888 static void tcp_update_pacing_rate(struct sock
*sk
)
890 const struct tcp_sock
*tp
= tcp_sk(sk
);
893 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
894 rate
= (u64
)tp
->mss_cache
* ((USEC_PER_SEC
/ 100) << 3);
896 /* current rate is (cwnd * mss) / srtt
897 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
898 * In Congestion Avoidance phase, set it to 120 % the current rate.
900 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
901 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
902 * end of slow start and should slow down.
904 if (tcp_snd_cwnd(tp
) < tp
->snd_ssthresh
/ 2)
905 rate
*= sock_net(sk
)->ipv4
.sysctl_tcp_pacing_ss_ratio
;
907 rate
*= sock_net(sk
)->ipv4
.sysctl_tcp_pacing_ca_ratio
;
909 rate
*= max(tcp_snd_cwnd(tp
), tp
->packets_out
);
911 if (likely(tp
->srtt_us
))
912 do_div(rate
, tp
->srtt_us
);
914 /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate
915 * without any lock. We want to make sure compiler wont store
916 * intermediate values in this location.
918 WRITE_ONCE(sk
->sk_pacing_rate
, min_t(u64
, rate
,
919 sk
->sk_max_pacing_rate
));
922 /* Calculate rto without backoff. This is the second half of Van Jacobson's
923 * routine referred to above.
925 static void tcp_set_rto(struct sock
*sk
)
927 const struct tcp_sock
*tp
= tcp_sk(sk
);
928 /* Old crap is replaced with new one. 8)
931 * 1. If rtt variance happened to be less 50msec, it is hallucination.
932 * It cannot be less due to utterly erratic ACK generation made
933 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
934 * to do with delayed acks, because at cwnd>2 true delack timeout
935 * is invisible. Actually, Linux-2.4 also generates erratic
936 * ACKs in some circumstances.
938 inet_csk(sk
)->icsk_rto
= __tcp_set_rto(tp
);
940 /* 2. Fixups made earlier cannot be right.
941 * If we do not estimate RTO correctly without them,
942 * all the algo is pure shit and should be replaced
943 * with correct one. It is exactly, which we pretend to do.
946 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
947 * guarantees that rto is higher.
952 __u32
tcp_init_cwnd(const struct tcp_sock
*tp
, const struct dst_entry
*dst
)
954 __u32 cwnd
= (dst
? dst_metric(dst
, RTAX_INITCWND
) : 0);
957 cwnd
= TCP_INIT_CWND
;
958 return min_t(__u32
, cwnd
, tp
->snd_cwnd_clamp
);
961 struct tcp_sacktag_state
{
962 /* Timestamps for earliest and latest never-retransmitted segment
963 * that was SACKed. RTO needs the earliest RTT to stay conservative,
964 * but congestion control should still get an accurate delay signal.
971 unsigned int mss_now
;
972 struct rate_sample
*rate
;
975 /* Take a notice that peer is sending D-SACKs. Skip update of data delivery
976 * and spurious retransmission information if this DSACK is unlikely caused by
978 * - DSACKed sequence range is larger than maximum receiver's window.
979 * - Total no. of DSACKed segments exceed the total no. of retransmitted segs.
981 static u32
tcp_dsack_seen(struct tcp_sock
*tp
, u32 start_seq
,
982 u32 end_seq
, struct tcp_sacktag_state
*state
)
984 u32 seq_len
, dup_segs
= 1;
986 if (!before(start_seq
, end_seq
))
989 seq_len
= end_seq
- start_seq
;
990 /* Dubious DSACK: DSACKed range greater than maximum advertised rwnd */
991 if (seq_len
> tp
->max_window
)
993 if (seq_len
> tp
->mss_cache
)
994 dup_segs
= DIV_ROUND_UP(seq_len
, tp
->mss_cache
);
995 else if (tp
->tlp_high_seq
&& tp
->tlp_high_seq
== end_seq
)
996 state
->flag
|= FLAG_DSACK_TLP
;
998 tp
->dsack_dups
+= dup_segs
;
999 /* Skip the DSACK if dup segs weren't retransmitted by sender */
1000 if (tp
->dsack_dups
> tp
->total_retrans
)
1003 tp
->rx_opt
.sack_ok
|= TCP_DSACK_SEEN
;
1004 /* We increase the RACK ordering window in rounds where we receive
1005 * DSACKs that may have been due to reordering causing RACK to trigger
1006 * a spurious fast recovery. Thus RACK ignores DSACKs that happen
1007 * without having seen reordering, or that match TLP probes (TLP
1008 * is timer-driven, not triggered by RACK).
1010 if (tp
->reord_seen
&& !(state
->flag
& FLAG_DSACK_TLP
))
1011 tp
->rack
.dsack_seen
= 1;
1013 state
->flag
|= FLAG_DSACKING_ACK
;
1014 /* A spurious retransmission is delivered */
1015 state
->sack_delivered
+= dup_segs
;
1020 /* It's reordering when higher sequence was delivered (i.e. sacked) before
1021 * some lower never-retransmitted sequence ("low_seq"). The maximum reordering
1022 * distance is approximated in full-mss packet distance ("reordering").
1024 static void tcp_check_sack_reordering(struct sock
*sk
, const u32 low_seq
,
1027 struct tcp_sock
*tp
= tcp_sk(sk
);
1028 const u32 mss
= tp
->mss_cache
;
1031 fack
= tcp_highest_sack_seq(tp
);
1032 if (!before(low_seq
, fack
))
1035 metric
= fack
- low_seq
;
1036 if ((metric
> tp
->reordering
* mss
) && mss
) {
1037 #if FASTRETRANS_DEBUG > 1
1038 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
1039 tp
->rx_opt
.sack_ok
, inet_csk(sk
)->icsk_ca_state
,
1043 tp
->undo_marker
? tp
->undo_retrans
: 0);
1045 tp
->reordering
= min_t(u32
, (metric
+ mss
- 1) / mss
,
1046 READ_ONCE(sock_net(sk
)->ipv4
.sysctl_tcp_max_reordering
));
1049 /* This exciting event is worth to be remembered. 8) */
1051 NET_INC_STATS(sock_net(sk
),
1052 ts
? LINUX_MIB_TCPTSREORDER
: LINUX_MIB_TCPSACKREORDER
);
1055 /* This must be called before lost_out or retrans_out are updated
1056 * on a new loss, because we want to know if all skbs previously
1057 * known to be lost have already been retransmitted, indicating
1058 * that this newly lost skb is our next skb to retransmit.
1060 static void tcp_verify_retransmit_hint(struct tcp_sock
*tp
, struct sk_buff
*skb
)
1062 if ((!tp
->retransmit_skb_hint
&& tp
->retrans_out
>= tp
->lost_out
) ||
1063 (tp
->retransmit_skb_hint
&&
1064 before(TCP_SKB_CB(skb
)->seq
,
1065 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
)))
1066 tp
->retransmit_skb_hint
= skb
;
1069 /* Sum the number of packets on the wire we have marked as lost, and
1070 * notify the congestion control module that the given skb was marked lost.
1072 static void tcp_notify_skb_loss_event(struct tcp_sock
*tp
, const struct sk_buff
*skb
)
1074 tp
->lost
+= tcp_skb_pcount(skb
);
1077 void tcp_mark_skb_lost(struct sock
*sk
, struct sk_buff
*skb
)
1079 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
1080 struct tcp_sock
*tp
= tcp_sk(sk
);
1082 if (sacked
& TCPCB_SACKED_ACKED
)
1085 tcp_verify_retransmit_hint(tp
, skb
);
1086 if (sacked
& TCPCB_LOST
) {
1087 if (sacked
& TCPCB_SACKED_RETRANS
) {
1088 /* Account for retransmits that are lost again */
1089 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1090 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1091 NET_ADD_STATS(sock_net(sk
), LINUX_MIB_TCPLOSTRETRANSMIT
,
1092 tcp_skb_pcount(skb
));
1093 tcp_notify_skb_loss_event(tp
, skb
);
1096 tp
->lost_out
+= tcp_skb_pcount(skb
);
1097 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1098 tcp_notify_skb_loss_event(tp
, skb
);
1102 /* Updates the delivered and delivered_ce counts */
1103 static void tcp_count_delivered(struct tcp_sock
*tp
, u32 delivered
,
1106 tp
->delivered
+= delivered
;
1108 tp
->delivered_ce
+= delivered
;
1111 /* This procedure tags the retransmission queue when SACKs arrive.
1113 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1114 * Packets in queue with these bits set are counted in variables
1115 * sacked_out, retrans_out and lost_out, correspondingly.
1117 * Valid combinations are:
1118 * Tag InFlight Description
1119 * 0 1 - orig segment is in flight.
1120 * S 0 - nothing flies, orig reached receiver.
1121 * L 0 - nothing flies, orig lost by net.
1122 * R 2 - both orig and retransmit are in flight.
1123 * L|R 1 - orig is lost, retransmit is in flight.
1124 * S|R 1 - orig reached receiver, retrans is still in flight.
1125 * (L|S|R is logically valid, it could occur when L|R is sacked,
1126 * but it is equivalent to plain S and code short-curcuits it to S.
1127 * L|S is logically invalid, it would mean -1 packet in flight 8))
1129 * These 6 states form finite state machine, controlled by the following events:
1130 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1131 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1132 * 3. Loss detection event of two flavors:
1133 * A. Scoreboard estimator decided the packet is lost.
1134 * A'. Reno "three dupacks" marks head of queue lost.
1135 * B. SACK arrives sacking SND.NXT at the moment, when the
1136 * segment was retransmitted.
1137 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1139 * It is pleasant to note, that state diagram turns out to be commutative,
1140 * so that we are allowed not to be bothered by order of our actions,
1141 * when multiple events arrive simultaneously. (see the function below).
1143 * Reordering detection.
1144 * --------------------
1145 * Reordering metric is maximal distance, which a packet can be displaced
1146 * in packet stream. With SACKs we can estimate it:
1148 * 1. SACK fills old hole and the corresponding segment was not
1149 * ever retransmitted -> reordering. Alas, we cannot use it
1150 * when segment was retransmitted.
1151 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1152 * for retransmitted and already SACKed segment -> reordering..
1153 * Both of these heuristics are not used in Loss state, when we cannot
1154 * account for retransmits accurately.
1156 * SACK block validation.
1157 * ----------------------
1159 * SACK block range validation checks that the received SACK block fits to
1160 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1161 * Note that SND.UNA is not included to the range though being valid because
1162 * it means that the receiver is rather inconsistent with itself reporting
1163 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1164 * perfectly valid, however, in light of RFC2018 which explicitly states
1165 * that "SACK block MUST reflect the newest segment. Even if the newest
1166 * segment is going to be discarded ...", not that it looks very clever
1167 * in case of head skb. Due to potentional receiver driven attacks, we
1168 * choose to avoid immediate execution of a walk in write queue due to
1169 * reneging and defer head skb's loss recovery to standard loss recovery
1170 * procedure that will eventually trigger (nothing forbids us doing this).
1172 * Implements also blockage to start_seq wrap-around. Problem lies in the
1173 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1174 * there's no guarantee that it will be before snd_nxt (n). The problem
1175 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1178 * <- outs wnd -> <- wrapzone ->
1179 * u e n u_w e_w s n_w
1181 * |<------------+------+----- TCP seqno space --------------+---------->|
1182 * ...-- <2^31 ->| |<--------...
1183 * ...---- >2^31 ------>| |<--------...
1185 * Current code wouldn't be vulnerable but it's better still to discard such
1186 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1187 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1188 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1189 * equal to the ideal case (infinite seqno space without wrap caused issues).
1191 * With D-SACK the lower bound is extended to cover sequence space below
1192 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1193 * again, D-SACK block must not to go across snd_una (for the same reason as
1194 * for the normal SACK blocks, explained above). But there all simplicity
1195 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1196 * fully below undo_marker they do not affect behavior in anyway and can
1197 * therefore be safely ignored. In rare cases (which are more or less
1198 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1199 * fragmentation and packet reordering past skb's retransmission. To consider
1200 * them correctly, the acceptable range must be extended even more though
1201 * the exact amount is rather hard to quantify. However, tp->max_window can
1202 * be used as an exaggerated estimate.
1204 static bool tcp_is_sackblock_valid(struct tcp_sock
*tp
, bool is_dsack
,
1205 u32 start_seq
, u32 end_seq
)
1207 /* Too far in future, or reversed (interpretation is ambiguous) */
1208 if (after(end_seq
, tp
->snd_nxt
) || !before(start_seq
, end_seq
))
1211 /* Nasty start_seq wrap-around check (see comments above) */
1212 if (!before(start_seq
, tp
->snd_nxt
))
1215 /* In outstanding window? ...This is valid exit for D-SACKs too.
1216 * start_seq == snd_una is non-sensical (see comments above)
1218 if (after(start_seq
, tp
->snd_una
))
1221 if (!is_dsack
|| !tp
->undo_marker
)
1224 /* ...Then it's D-SACK, and must reside below snd_una completely */
1225 if (after(end_seq
, tp
->snd_una
))
1228 if (!before(start_seq
, tp
->undo_marker
))
1232 if (!after(end_seq
, tp
->undo_marker
))
1235 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1236 * start_seq < undo_marker and end_seq >= undo_marker.
1238 return !before(start_seq
, end_seq
- tp
->max_window
);
1241 static bool tcp_check_dsack(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1242 struct tcp_sack_block_wire
*sp
, int num_sacks
,
1243 u32 prior_snd_una
, struct tcp_sacktag_state
*state
)
1245 struct tcp_sock
*tp
= tcp_sk(sk
);
1246 u32 start_seq_0
= get_unaligned_be32(&sp
[0].start_seq
);
1247 u32 end_seq_0
= get_unaligned_be32(&sp
[0].end_seq
);
1250 if (before(start_seq_0
, TCP_SKB_CB(ack_skb
)->ack_seq
)) {
1251 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPDSACKRECV
);
1252 } else if (num_sacks
> 1) {
1253 u32 end_seq_1
= get_unaligned_be32(&sp
[1].end_seq
);
1254 u32 start_seq_1
= get_unaligned_be32(&sp
[1].start_seq
);
1256 if (after(end_seq_0
, end_seq_1
) || before(start_seq_0
, start_seq_1
))
1258 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPDSACKOFORECV
);
1263 dup_segs
= tcp_dsack_seen(tp
, start_seq_0
, end_seq_0
, state
);
1264 if (!dup_segs
) { /* Skip dubious DSACK */
1265 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPDSACKIGNOREDDUBIOUS
);
1269 NET_ADD_STATS(sock_net(sk
), LINUX_MIB_TCPDSACKRECVSEGS
, dup_segs
);
1271 /* D-SACK for already forgotten data... Do dumb counting. */
1272 if (tp
->undo_marker
&& tp
->undo_retrans
> 0 &&
1273 !after(end_seq_0
, prior_snd_una
) &&
1274 after(end_seq_0
, tp
->undo_marker
))
1275 tp
->undo_retrans
= max_t(int, 0, tp
->undo_retrans
- dup_segs
);
1280 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1281 * the incoming SACK may not exactly match but we can find smaller MSS
1282 * aligned portion of it that matches. Therefore we might need to fragment
1283 * which may fail and creates some hassle (caller must handle error case
1286 * FIXME: this could be merged to shift decision code
1288 static int tcp_match_skb_to_sack(struct sock
*sk
, struct sk_buff
*skb
,
1289 u32 start_seq
, u32 end_seq
)
1293 unsigned int pkt_len
;
1296 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1297 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1299 if (tcp_skb_pcount(skb
) > 1 && !in_sack
&&
1300 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
1301 mss
= tcp_skb_mss(skb
);
1302 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1305 pkt_len
= start_seq
- TCP_SKB_CB(skb
)->seq
;
1309 pkt_len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1314 /* Round if necessary so that SACKs cover only full MSSes
1315 * and/or the remaining small portion (if present)
1317 if (pkt_len
> mss
) {
1318 unsigned int new_len
= (pkt_len
/ mss
) * mss
;
1319 if (!in_sack
&& new_len
< pkt_len
)
1324 if (pkt_len
>= skb
->len
&& !in_sack
)
1327 err
= tcp_fragment(sk
, TCP_FRAG_IN_RTX_QUEUE
, skb
,
1328 pkt_len
, mss
, GFP_ATOMIC
);
1336 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1337 static u8
tcp_sacktag_one(struct sock
*sk
,
1338 struct tcp_sacktag_state
*state
, u8 sacked
,
1339 u32 start_seq
, u32 end_seq
,
1340 int dup_sack
, int pcount
,
1343 struct tcp_sock
*tp
= tcp_sk(sk
);
1345 /* Account D-SACK for retransmitted packet. */
1346 if (dup_sack
&& (sacked
& TCPCB_RETRANS
)) {
1347 if (tp
->undo_marker
&& tp
->undo_retrans
> 0 &&
1348 after(end_seq
, tp
->undo_marker
))
1349 tp
->undo_retrans
= max_t(int, 0, tp
->undo_retrans
- pcount
);
1350 if ((sacked
& TCPCB_SACKED_ACKED
) &&
1351 before(start_seq
, state
->reord
))
1352 state
->reord
= start_seq
;
1355 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1356 if (!after(end_seq
, tp
->snd_una
))
1359 if (!(sacked
& TCPCB_SACKED_ACKED
)) {
1360 tcp_rack_advance(tp
, sacked
, end_seq
, xmit_time
);
1362 if (sacked
& TCPCB_SACKED_RETRANS
) {
1363 /* If the segment is not tagged as lost,
1364 * we do not clear RETRANS, believing
1365 * that retransmission is still in flight.
1367 if (sacked
& TCPCB_LOST
) {
1368 sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1369 tp
->lost_out
-= pcount
;
1370 tp
->retrans_out
-= pcount
;
1373 if (!(sacked
& TCPCB_RETRANS
)) {
1374 /* New sack for not retransmitted frame,
1375 * which was in hole. It is reordering.
1377 if (before(start_seq
,
1378 tcp_highest_sack_seq(tp
)) &&
1379 before(start_seq
, state
->reord
))
1380 state
->reord
= start_seq
;
1382 if (!after(end_seq
, tp
->high_seq
))
1383 state
->flag
|= FLAG_ORIG_SACK_ACKED
;
1384 if (state
->first_sackt
== 0)
1385 state
->first_sackt
= xmit_time
;
1386 state
->last_sackt
= xmit_time
;
1389 if (sacked
& TCPCB_LOST
) {
1390 sacked
&= ~TCPCB_LOST
;
1391 tp
->lost_out
-= pcount
;
1395 sacked
|= TCPCB_SACKED_ACKED
;
1396 state
->flag
|= FLAG_DATA_SACKED
;
1397 tp
->sacked_out
+= pcount
;
1398 /* Out-of-order packets delivered */
1399 state
->sack_delivered
+= pcount
;
1401 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1402 if (tp
->lost_skb_hint
&&
1403 before(start_seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
))
1404 tp
->lost_cnt_hint
+= pcount
;
1407 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1408 * frames and clear it. undo_retrans is decreased above, L|R frames
1409 * are accounted above as well.
1411 if (dup_sack
&& (sacked
& TCPCB_SACKED_RETRANS
)) {
1412 sacked
&= ~TCPCB_SACKED_RETRANS
;
1413 tp
->retrans_out
-= pcount
;
1419 /* Shift newly-SACKed bytes from this skb to the immediately previous
1420 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1422 static bool tcp_shifted_skb(struct sock
*sk
, struct sk_buff
*prev
,
1423 struct sk_buff
*skb
,
1424 struct tcp_sacktag_state
*state
,
1425 unsigned int pcount
, int shifted
, int mss
,
1428 struct tcp_sock
*tp
= tcp_sk(sk
);
1429 u32 start_seq
= TCP_SKB_CB(skb
)->seq
; /* start of newly-SACKed */
1430 u32 end_seq
= start_seq
+ shifted
; /* end of newly-SACKed */
1434 /* Adjust counters and hints for the newly sacked sequence
1435 * range but discard the return value since prev is already
1436 * marked. We must tag the range first because the seq
1437 * advancement below implicitly advances
1438 * tcp_highest_sack_seq() when skb is highest_sack.
1440 tcp_sacktag_one(sk
, state
, TCP_SKB_CB(skb
)->sacked
,
1441 start_seq
, end_seq
, dup_sack
, pcount
,
1442 tcp_skb_timestamp_us(skb
));
1443 tcp_rate_skb_delivered(sk
, skb
, state
->rate
);
1445 if (skb
== tp
->lost_skb_hint
)
1446 tp
->lost_cnt_hint
+= pcount
;
1448 TCP_SKB_CB(prev
)->end_seq
+= shifted
;
1449 TCP_SKB_CB(skb
)->seq
+= shifted
;
1451 tcp_skb_pcount_add(prev
, pcount
);
1452 WARN_ON_ONCE(tcp_skb_pcount(skb
) < pcount
);
1453 tcp_skb_pcount_add(skb
, -pcount
);
1455 /* When we're adding to gso_segs == 1, gso_size will be zero,
1456 * in theory this shouldn't be necessary but as long as DSACK
1457 * code can come after this skb later on it's better to keep
1458 * setting gso_size to something.
1460 if (!TCP_SKB_CB(prev
)->tcp_gso_size
)
1461 TCP_SKB_CB(prev
)->tcp_gso_size
= mss
;
1463 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1464 if (tcp_skb_pcount(skb
) <= 1)
1465 TCP_SKB_CB(skb
)->tcp_gso_size
= 0;
1467 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1468 TCP_SKB_CB(prev
)->sacked
|= (TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
);
1471 BUG_ON(!tcp_skb_pcount(skb
));
1472 NET_INC_STATS(sock_net(sk
), LINUX_MIB_SACKSHIFTED
);
1476 /* Whole SKB was eaten :-) */
1478 if (skb
== tp
->retransmit_skb_hint
)
1479 tp
->retransmit_skb_hint
= prev
;
1480 if (skb
== tp
->lost_skb_hint
) {
1481 tp
->lost_skb_hint
= prev
;
1482 tp
->lost_cnt_hint
-= tcp_skb_pcount(prev
);
1485 TCP_SKB_CB(prev
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
;
1486 TCP_SKB_CB(prev
)->eor
= TCP_SKB_CB(skb
)->eor
;
1487 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
1488 TCP_SKB_CB(prev
)->end_seq
++;
1490 if (skb
== tcp_highest_sack(sk
))
1491 tcp_advance_highest_sack(sk
, skb
);
1493 tcp_skb_collapse_tstamp(prev
, skb
);
1494 if (unlikely(TCP_SKB_CB(prev
)->tx
.delivered_mstamp
))
1495 TCP_SKB_CB(prev
)->tx
.delivered_mstamp
= 0;
1497 tcp_rtx_queue_unlink_and_free(skb
, sk
);
1499 NET_INC_STATS(sock_net(sk
), LINUX_MIB_SACKMERGED
);
1504 /* I wish gso_size would have a bit more sane initialization than
1505 * something-or-zero which complicates things
1507 static int tcp_skb_seglen(const struct sk_buff
*skb
)
1509 return tcp_skb_pcount(skb
) == 1 ? skb
->len
: tcp_skb_mss(skb
);
1512 /* Shifting pages past head area doesn't work */
1513 static int skb_can_shift(const struct sk_buff
*skb
)
1515 return !skb_headlen(skb
) && skb_is_nonlinear(skb
);
1518 int tcp_skb_shift(struct sk_buff
*to
, struct sk_buff
*from
,
1519 int pcount
, int shiftlen
)
1521 /* TCP min gso_size is 8 bytes (TCP_MIN_GSO_SIZE)
1522 * Since TCP_SKB_CB(skb)->tcp_gso_segs is 16 bits, we need
1523 * to make sure not storing more than 65535 * 8 bytes per skb,
1524 * even if current MSS is bigger.
1526 if (unlikely(to
->len
+ shiftlen
>= 65535 * TCP_MIN_GSO_SIZE
))
1528 if (unlikely(tcp_skb_pcount(to
) + pcount
> 65535))
1530 return skb_shift(to
, from
, shiftlen
);
1533 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1536 static struct sk_buff
*tcp_shift_skb_data(struct sock
*sk
, struct sk_buff
*skb
,
1537 struct tcp_sacktag_state
*state
,
1538 u32 start_seq
, u32 end_seq
,
1541 struct tcp_sock
*tp
= tcp_sk(sk
);
1542 struct sk_buff
*prev
;
1548 /* Normally R but no L won't result in plain S */
1550 (TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_RETRANS
)) == TCPCB_SACKED_RETRANS
)
1552 if (!skb_can_shift(skb
))
1554 /* This frame is about to be dropped (was ACKed). */
1555 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1558 /* Can only happen with delayed DSACK + discard craziness */
1559 prev
= skb_rb_prev(skb
);
1563 if ((TCP_SKB_CB(prev
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
)
1566 if (!tcp_skb_can_collapse(prev
, skb
))
1569 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1570 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1574 pcount
= tcp_skb_pcount(skb
);
1575 mss
= tcp_skb_seglen(skb
);
1577 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1578 * drop this restriction as unnecessary
1580 if (mss
!= tcp_skb_seglen(prev
))
1583 if (!after(TCP_SKB_CB(skb
)->end_seq
, start_seq
))
1585 /* CHECKME: This is non-MSS split case only?, this will
1586 * cause skipped skbs due to advancing loop btw, original
1587 * has that feature too
1589 if (tcp_skb_pcount(skb
) <= 1)
1592 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1594 /* TODO: head merge to next could be attempted here
1595 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1596 * though it might not be worth of the additional hassle
1598 * ...we can probably just fallback to what was done
1599 * previously. We could try merging non-SACKed ones
1600 * as well but it probably isn't going to buy off
1601 * because later SACKs might again split them, and
1602 * it would make skb timestamp tracking considerably
1608 len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1610 BUG_ON(len
> skb
->len
);
1612 /* MSS boundaries should be honoured or else pcount will
1613 * severely break even though it makes things bit trickier.
1614 * Optimize common case to avoid most of the divides
1616 mss
= tcp_skb_mss(skb
);
1618 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1619 * drop this restriction as unnecessary
1621 if (mss
!= tcp_skb_seglen(prev
))
1626 } else if (len
< mss
) {
1634 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1635 if (!after(TCP_SKB_CB(skb
)->seq
+ len
, tp
->snd_una
))
1638 if (!tcp_skb_shift(prev
, skb
, pcount
, len
))
1640 if (!tcp_shifted_skb(sk
, prev
, skb
, state
, pcount
, len
, mss
, dup_sack
))
1643 /* Hole filled allows collapsing with the next as well, this is very
1644 * useful when hole on every nth skb pattern happens
1646 skb
= skb_rb_next(prev
);
1650 if (!skb_can_shift(skb
) ||
1651 ((TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
) ||
1652 (mss
!= tcp_skb_seglen(skb
)))
1655 if (!tcp_skb_can_collapse(prev
, skb
))
1658 pcount
= tcp_skb_pcount(skb
);
1659 if (tcp_skb_shift(prev
, skb
, pcount
, len
))
1660 tcp_shifted_skb(sk
, prev
, skb
, state
, pcount
,
1670 NET_INC_STATS(sock_net(sk
), LINUX_MIB_SACKSHIFTFALLBACK
);
1674 static struct sk_buff
*tcp_sacktag_walk(struct sk_buff
*skb
, struct sock
*sk
,
1675 struct tcp_sack_block
*next_dup
,
1676 struct tcp_sacktag_state
*state
,
1677 u32 start_seq
, u32 end_seq
,
1680 struct tcp_sock
*tp
= tcp_sk(sk
);
1681 struct sk_buff
*tmp
;
1683 skb_rbtree_walk_from(skb
) {
1685 bool dup_sack
= dup_sack_in
;
1687 /* queue is in-order => we can short-circuit the walk early */
1688 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1692 before(TCP_SKB_CB(skb
)->seq
, next_dup
->end_seq
)) {
1693 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1694 next_dup
->start_seq
,
1700 /* skb reference here is a bit tricky to get right, since
1701 * shifting can eat and free both this skb and the next,
1702 * so not even _safe variant of the loop is enough.
1705 tmp
= tcp_shift_skb_data(sk
, skb
, state
,
1706 start_seq
, end_seq
, dup_sack
);
1715 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1721 if (unlikely(in_sack
< 0))
1725 TCP_SKB_CB(skb
)->sacked
=
1728 TCP_SKB_CB(skb
)->sacked
,
1729 TCP_SKB_CB(skb
)->seq
,
1730 TCP_SKB_CB(skb
)->end_seq
,
1732 tcp_skb_pcount(skb
),
1733 tcp_skb_timestamp_us(skb
));
1734 tcp_rate_skb_delivered(sk
, skb
, state
->rate
);
1735 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
1736 list_del_init(&skb
->tcp_tsorted_anchor
);
1738 if (!before(TCP_SKB_CB(skb
)->seq
,
1739 tcp_highest_sack_seq(tp
)))
1740 tcp_advance_highest_sack(sk
, skb
);
1746 static struct sk_buff
*tcp_sacktag_bsearch(struct sock
*sk
, u32 seq
)
1748 struct rb_node
*parent
, **p
= &sk
->tcp_rtx_queue
.rb_node
;
1749 struct sk_buff
*skb
;
1753 skb
= rb_to_skb(parent
);
1754 if (before(seq
, TCP_SKB_CB(skb
)->seq
)) {
1755 p
= &parent
->rb_left
;
1758 if (!before(seq
, TCP_SKB_CB(skb
)->end_seq
)) {
1759 p
= &parent
->rb_right
;
1767 static struct sk_buff
*tcp_sacktag_skip(struct sk_buff
*skb
, struct sock
*sk
,
1770 if (skb
&& after(TCP_SKB_CB(skb
)->seq
, skip_to_seq
))
1773 return tcp_sacktag_bsearch(sk
, skip_to_seq
);
1776 static struct sk_buff
*tcp_maybe_skipping_dsack(struct sk_buff
*skb
,
1778 struct tcp_sack_block
*next_dup
,
1779 struct tcp_sacktag_state
*state
,
1785 if (before(next_dup
->start_seq
, skip_to_seq
)) {
1786 skb
= tcp_sacktag_skip(skb
, sk
, next_dup
->start_seq
);
1787 skb
= tcp_sacktag_walk(skb
, sk
, NULL
, state
,
1788 next_dup
->start_seq
, next_dup
->end_seq
,
1795 static int tcp_sack_cache_ok(const struct tcp_sock
*tp
, const struct tcp_sack_block
*cache
)
1797 return cache
< tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1801 tcp_sacktag_write_queue(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1802 u32 prior_snd_una
, struct tcp_sacktag_state
*state
)
1804 struct tcp_sock
*tp
= tcp_sk(sk
);
1805 const unsigned char *ptr
= (skb_transport_header(ack_skb
) +
1806 TCP_SKB_CB(ack_skb
)->sacked
);
1807 struct tcp_sack_block_wire
*sp_wire
= (struct tcp_sack_block_wire
*)(ptr
+2);
1808 struct tcp_sack_block sp
[TCP_NUM_SACKS
];
1809 struct tcp_sack_block
*cache
;
1810 struct sk_buff
*skb
;
1811 int num_sacks
= min(TCP_NUM_SACKS
, (ptr
[1] - TCPOLEN_SACK_BASE
) >> 3);
1813 bool found_dup_sack
= false;
1815 int first_sack_index
;
1818 state
->reord
= tp
->snd_nxt
;
1820 if (!tp
->sacked_out
)
1821 tcp_highest_sack_reset(sk
);
1823 found_dup_sack
= tcp_check_dsack(sk
, ack_skb
, sp_wire
,
1824 num_sacks
, prior_snd_una
, state
);
1826 /* Eliminate too old ACKs, but take into
1827 * account more or less fresh ones, they can
1828 * contain valid SACK info.
1830 if (before(TCP_SKB_CB(ack_skb
)->ack_seq
, prior_snd_una
- tp
->max_window
))
1833 if (!tp
->packets_out
)
1837 first_sack_index
= 0;
1838 for (i
= 0; i
< num_sacks
; i
++) {
1839 bool dup_sack
= !i
&& found_dup_sack
;
1841 sp
[used_sacks
].start_seq
= get_unaligned_be32(&sp_wire
[i
].start_seq
);
1842 sp
[used_sacks
].end_seq
= get_unaligned_be32(&sp_wire
[i
].end_seq
);
1844 if (!tcp_is_sackblock_valid(tp
, dup_sack
,
1845 sp
[used_sacks
].start_seq
,
1846 sp
[used_sacks
].end_seq
)) {
1850 if (!tp
->undo_marker
)
1851 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDNOUNDO
;
1853 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDOLD
;
1855 /* Don't count olds caused by ACK reordering */
1856 if ((TCP_SKB_CB(ack_skb
)->ack_seq
!= tp
->snd_una
) &&
1857 !after(sp
[used_sacks
].end_seq
, tp
->snd_una
))
1859 mib_idx
= LINUX_MIB_TCPSACKDISCARD
;
1862 NET_INC_STATS(sock_net(sk
), mib_idx
);
1864 first_sack_index
= -1;
1868 /* Ignore very old stuff early */
1869 if (!after(sp
[used_sacks
].end_seq
, prior_snd_una
)) {
1871 first_sack_index
= -1;
1878 /* order SACK blocks to allow in order walk of the retrans queue */
1879 for (i
= used_sacks
- 1; i
> 0; i
--) {
1880 for (j
= 0; j
< i
; j
++) {
1881 if (after(sp
[j
].start_seq
, sp
[j
+ 1].start_seq
)) {
1882 swap(sp
[j
], sp
[j
+ 1]);
1884 /* Track where the first SACK block goes to */
1885 if (j
== first_sack_index
)
1886 first_sack_index
= j
+ 1;
1891 state
->mss_now
= tcp_current_mss(sk
);
1895 if (!tp
->sacked_out
) {
1896 /* It's already past, so skip checking against it */
1897 cache
= tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1899 cache
= tp
->recv_sack_cache
;
1900 /* Skip empty blocks in at head of the cache */
1901 while (tcp_sack_cache_ok(tp
, cache
) && !cache
->start_seq
&&
1906 while (i
< used_sacks
) {
1907 u32 start_seq
= sp
[i
].start_seq
;
1908 u32 end_seq
= sp
[i
].end_seq
;
1909 bool dup_sack
= (found_dup_sack
&& (i
== first_sack_index
));
1910 struct tcp_sack_block
*next_dup
= NULL
;
1912 if (found_dup_sack
&& ((i
+ 1) == first_sack_index
))
1913 next_dup
= &sp
[i
+ 1];
1915 /* Skip too early cached blocks */
1916 while (tcp_sack_cache_ok(tp
, cache
) &&
1917 !before(start_seq
, cache
->end_seq
))
1920 /* Can skip some work by looking recv_sack_cache? */
1921 if (tcp_sack_cache_ok(tp
, cache
) && !dup_sack
&&
1922 after(end_seq
, cache
->start_seq
)) {
1925 if (before(start_seq
, cache
->start_seq
)) {
1926 skb
= tcp_sacktag_skip(skb
, sk
, start_seq
);
1927 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
,
1934 /* Rest of the block already fully processed? */
1935 if (!after(end_seq
, cache
->end_seq
))
1938 skb
= tcp_maybe_skipping_dsack(skb
, sk
, next_dup
,
1942 /* ...tail remains todo... */
1943 if (tcp_highest_sack_seq(tp
) == cache
->end_seq
) {
1944 /* ...but better entrypoint exists! */
1945 skb
= tcp_highest_sack(sk
);
1952 skb
= tcp_sacktag_skip(skb
, sk
, cache
->end_seq
);
1953 /* Check overlap against next cached too (past this one already) */
1958 if (!before(start_seq
, tcp_highest_sack_seq(tp
))) {
1959 skb
= tcp_highest_sack(sk
);
1963 skb
= tcp_sacktag_skip(skb
, sk
, start_seq
);
1966 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
, state
,
1967 start_seq
, end_seq
, dup_sack
);
1973 /* Clear the head of the cache sack blocks so we can skip it next time */
1974 for (i
= 0; i
< ARRAY_SIZE(tp
->recv_sack_cache
) - used_sacks
; i
++) {
1975 tp
->recv_sack_cache
[i
].start_seq
= 0;
1976 tp
->recv_sack_cache
[i
].end_seq
= 0;
1978 for (j
= 0; j
< used_sacks
; j
++)
1979 tp
->recv_sack_cache
[i
++] = sp
[j
];
1981 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Loss
|| tp
->undo_marker
)
1982 tcp_check_sack_reordering(sk
, state
->reord
, 0);
1984 tcp_verify_left_out(tp
);
1987 #if FASTRETRANS_DEBUG > 0
1988 WARN_ON((int)tp
->sacked_out
< 0);
1989 WARN_ON((int)tp
->lost_out
< 0);
1990 WARN_ON((int)tp
->retrans_out
< 0);
1991 WARN_ON((int)tcp_packets_in_flight(tp
) < 0);
1996 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1997 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1999 static bool tcp_limit_reno_sacked(struct tcp_sock
*tp
)
2003 holes
= max(tp
->lost_out
, 1U);
2004 holes
= min(holes
, tp
->packets_out
);
2006 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
2007 tp
->sacked_out
= tp
->packets_out
- holes
;
2013 /* If we receive more dupacks than we expected counting segments
2014 * in assumption of absent reordering, interpret this as reordering.
2015 * The only another reason could be bug in receiver TCP.
2017 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
2019 struct tcp_sock
*tp
= tcp_sk(sk
);
2021 if (!tcp_limit_reno_sacked(tp
))
2024 tp
->reordering
= min_t(u32
, tp
->packets_out
+ addend
,
2025 READ_ONCE(sock_net(sk
)->ipv4
.sysctl_tcp_max_reordering
));
2027 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPRENOREORDER
);
2030 /* Emulate SACKs for SACKless connection: account for a new dupack. */
2032 static void tcp_add_reno_sack(struct sock
*sk
, int num_dupack
, bool ece_ack
)
2035 struct tcp_sock
*tp
= tcp_sk(sk
);
2036 u32 prior_sacked
= tp
->sacked_out
;
2039 tp
->sacked_out
+= num_dupack
;
2040 tcp_check_reno_reordering(sk
, 0);
2041 delivered
= tp
->sacked_out
- prior_sacked
;
2043 tcp_count_delivered(tp
, delivered
, ece_ack
);
2044 tcp_verify_left_out(tp
);
2048 /* Account for ACK, ACKing some data in Reno Recovery phase. */
2050 static void tcp_remove_reno_sacks(struct sock
*sk
, int acked
, bool ece_ack
)
2052 struct tcp_sock
*tp
= tcp_sk(sk
);
2055 /* One ACK acked hole. The rest eat duplicate ACKs. */
2056 tcp_count_delivered(tp
, max_t(int, acked
- tp
->sacked_out
, 1),
2058 if (acked
- 1 >= tp
->sacked_out
)
2061 tp
->sacked_out
-= acked
- 1;
2063 tcp_check_reno_reordering(sk
, acked
);
2064 tcp_verify_left_out(tp
);
2067 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
2072 void tcp_clear_retrans(struct tcp_sock
*tp
)
2074 tp
->retrans_out
= 0;
2076 tp
->undo_marker
= 0;
2077 tp
->undo_retrans
= -1;
2081 static inline void tcp_init_undo(struct tcp_sock
*tp
)
2083 tp
->undo_marker
= tp
->snd_una
;
2084 /* Retransmission still in flight may cause DSACKs later. */
2085 tp
->undo_retrans
= tp
->retrans_out
? : -1;
2088 static bool tcp_is_rack(const struct sock
*sk
)
2090 return READ_ONCE(sock_net(sk
)->ipv4
.sysctl_tcp_recovery
) &
2091 TCP_RACK_LOSS_DETECTION
;
2094 /* If we detect SACK reneging, forget all SACK information
2095 * and reset tags completely, otherwise preserve SACKs. If receiver
2096 * dropped its ofo queue, we will know this due to reneging detection.
2098 static void tcp_timeout_mark_lost(struct sock
*sk
)
2100 struct tcp_sock
*tp
= tcp_sk(sk
);
2101 struct sk_buff
*skb
, *head
;
2102 bool is_reneg
; /* is receiver reneging on SACKs? */
2104 head
= tcp_rtx_queue_head(sk
);
2105 is_reneg
= head
&& (TCP_SKB_CB(head
)->sacked
& TCPCB_SACKED_ACKED
);
2107 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPSACKRENEGING
);
2109 /* Mark SACK reneging until we recover from this loss event. */
2110 tp
->is_sack_reneg
= 1;
2111 } else if (tcp_is_reno(tp
)) {
2112 tcp_reset_reno_sack(tp
);
2116 skb_rbtree_walk_from(skb
) {
2118 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
2119 else if (tcp_is_rack(sk
) && skb
!= head
&&
2120 tcp_rack_skb_timeout(tp
, skb
, 0) > 0)
2121 continue; /* Don't mark recently sent ones lost yet */
2122 tcp_mark_skb_lost(sk
, skb
);
2124 tcp_verify_left_out(tp
);
2125 tcp_clear_all_retrans_hints(tp
);
2128 /* Enter Loss state. */
2129 void tcp_enter_loss(struct sock
*sk
)
2131 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2132 struct tcp_sock
*tp
= tcp_sk(sk
);
2133 struct net
*net
= sock_net(sk
);
2134 bool new_recovery
= icsk
->icsk_ca_state
< TCP_CA_Recovery
;
2137 tcp_timeout_mark_lost(sk
);
2139 /* Reduce ssthresh if it has not yet been made inside this window. */
2140 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
||
2141 !after(tp
->high_seq
, tp
->snd_una
) ||
2142 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
2143 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2144 tp
->prior_cwnd
= tcp_snd_cwnd(tp
);
2145 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
2146 tcp_ca_event(sk
, CA_EVENT_LOSS
);
2149 tcp_snd_cwnd_set(tp
, tcp_packets_in_flight(tp
) + 1);
2150 tp
->snd_cwnd_cnt
= 0;
2151 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
2153 /* Timeout in disordered state after receiving substantial DUPACKs
2154 * suggests that the degree of reordering is over-estimated.
2156 reordering
= READ_ONCE(net
->ipv4
.sysctl_tcp_reordering
);
2157 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
&&
2158 tp
->sacked_out
>= reordering
)
2159 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
2162 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2163 tp
->high_seq
= tp
->snd_nxt
;
2164 tcp_ecn_queue_cwr(tp
);
2166 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
2167 * loss recovery is underway except recurring timeout(s) on
2168 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
2170 tp
->frto
= READ_ONCE(net
->ipv4
.sysctl_tcp_frto
) &&
2171 (new_recovery
|| icsk
->icsk_retransmits
) &&
2172 !inet_csk(sk
)->icsk_mtup
.probe_size
;
2175 /* If ACK arrived pointing to a remembered SACK, it means that our
2176 * remembered SACKs do not reflect real state of receiver i.e.
2177 * receiver _host_ is heavily congested (or buggy).
2179 * To avoid big spurious retransmission bursts due to transient SACK
2180 * scoreboard oddities that look like reneging, we give the receiver a
2181 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2182 * restore sanity to the SACK scoreboard. If the apparent reneging
2183 * persists until this RTO then we'll clear the SACK scoreboard.
2185 static bool tcp_check_sack_reneging(struct sock
*sk
, int flag
)
2187 if (flag
& FLAG_SACK_RENEGING
) {
2188 struct tcp_sock
*tp
= tcp_sk(sk
);
2189 unsigned long delay
= max(usecs_to_jiffies(tp
->srtt_us
>> 4),
2190 msecs_to_jiffies(10));
2192 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2193 delay
, TCP_RTO_MAX
);
2199 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2200 * counter when SACK is enabled (without SACK, sacked_out is used for
2203 * With reordering, holes may still be in flight, so RFC3517 recovery
2204 * uses pure sacked_out (total number of SACKed segments) even though
2205 * it violates the RFC that uses duplicate ACKs, often these are equal
2206 * but when e.g. out-of-window ACKs or packet duplication occurs,
2207 * they differ. Since neither occurs due to loss, TCP should really
2210 static inline int tcp_dupack_heuristics(const struct tcp_sock
*tp
)
2212 return tp
->sacked_out
+ 1;
2215 /* Linux NewReno/SACK/ECN state machine.
2216 * --------------------------------------
2218 * "Open" Normal state, no dubious events, fast path.
2219 * "Disorder" In all the respects it is "Open",
2220 * but requires a bit more attention. It is entered when
2221 * we see some SACKs or dupacks. It is split of "Open"
2222 * mainly to move some processing from fast path to slow one.
2223 * "CWR" CWND was reduced due to some Congestion Notification event.
2224 * It can be ECN, ICMP source quench, local device congestion.
2225 * "Recovery" CWND was reduced, we are fast-retransmitting.
2226 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2228 * tcp_fastretrans_alert() is entered:
2229 * - each incoming ACK, if state is not "Open"
2230 * - when arrived ACK is unusual, namely:
2235 * Counting packets in flight is pretty simple.
2237 * in_flight = packets_out - left_out + retrans_out
2239 * packets_out is SND.NXT-SND.UNA counted in packets.
2241 * retrans_out is number of retransmitted segments.
2243 * left_out is number of segments left network, but not ACKed yet.
2245 * left_out = sacked_out + lost_out
2247 * sacked_out: Packets, which arrived to receiver out of order
2248 * and hence not ACKed. With SACKs this number is simply
2249 * amount of SACKed data. Even without SACKs
2250 * it is easy to give pretty reliable estimate of this number,
2251 * counting duplicate ACKs.
2253 * lost_out: Packets lost by network. TCP has no explicit
2254 * "loss notification" feedback from network (for now).
2255 * It means that this number can be only _guessed_.
2256 * Actually, it is the heuristics to predict lossage that
2257 * distinguishes different algorithms.
2259 * F.e. after RTO, when all the queue is considered as lost,
2260 * lost_out = packets_out and in_flight = retrans_out.
2262 * Essentially, we have now a few algorithms detecting
2265 * If the receiver supports SACK:
2267 * RFC6675/3517: It is the conventional algorithm. A packet is
2268 * considered lost if the number of higher sequence packets
2269 * SACKed is greater than or equal the DUPACK thoreshold
2270 * (reordering). This is implemented in tcp_mark_head_lost and
2271 * tcp_update_scoreboard.
2273 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm
2274 * (2017-) that checks timing instead of counting DUPACKs.
2275 * Essentially a packet is considered lost if it's not S/ACKed
2276 * after RTT + reordering_window, where both metrics are
2277 * dynamically measured and adjusted. This is implemented in
2278 * tcp_rack_mark_lost.
2280 * If the receiver does not support SACK:
2282 * NewReno (RFC6582): in Recovery we assume that one segment
2283 * is lost (classic Reno). While we are in Recovery and
2284 * a partial ACK arrives, we assume that one more packet
2285 * is lost (NewReno). This heuristics are the same in NewReno
2288 * Really tricky (and requiring careful tuning) part of algorithm
2289 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2290 * The first determines the moment _when_ we should reduce CWND and,
2291 * hence, slow down forward transmission. In fact, it determines the moment
2292 * when we decide that hole is caused by loss, rather than by a reorder.
2294 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2295 * holes, caused by lost packets.
2297 * And the most logically complicated part of algorithm is undo
2298 * heuristics. We detect false retransmits due to both too early
2299 * fast retransmit (reordering) and underestimated RTO, analyzing
2300 * timestamps and D-SACKs. When we detect that some segments were
2301 * retransmitted by mistake and CWND reduction was wrong, we undo
2302 * window reduction and abort recovery phase. This logic is hidden
2303 * inside several functions named tcp_try_undo_<something>.
2306 /* This function decides, when we should leave Disordered state
2307 * and enter Recovery phase, reducing congestion window.
2309 * Main question: may we further continue forward transmission
2310 * with the same cwnd?
2312 static bool tcp_time_to_recover(struct sock
*sk
, int flag
)
2314 struct tcp_sock
*tp
= tcp_sk(sk
);
2316 /* Trick#1: The loss is proven. */
2320 /* Not-A-Trick#2 : Classic rule... */
2321 if (!tcp_is_rack(sk
) && tcp_dupack_heuristics(tp
) > tp
->reordering
)
2327 /* Detect loss in event "A" above by marking head of queue up as lost.
2328 * For RFC3517 SACK, a segment is considered lost if it
2329 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2330 * the maximum SACKed segments to pass before reaching this limit.
2332 static void tcp_mark_head_lost(struct sock
*sk
, int packets
, int mark_head
)
2334 struct tcp_sock
*tp
= tcp_sk(sk
);
2335 struct sk_buff
*skb
;
2337 /* Use SACK to deduce losses of new sequences sent during recovery */
2338 const u32 loss_high
= tp
->snd_nxt
;
2340 WARN_ON(packets
> tp
->packets_out
);
2341 skb
= tp
->lost_skb_hint
;
2343 /* Head already handled? */
2344 if (mark_head
&& after(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
))
2346 cnt
= tp
->lost_cnt_hint
;
2348 skb
= tcp_rtx_queue_head(sk
);
2352 skb_rbtree_walk_from(skb
) {
2353 /* TODO: do this better */
2354 /* this is not the most efficient way to do this... */
2355 tp
->lost_skb_hint
= skb
;
2356 tp
->lost_cnt_hint
= cnt
;
2358 if (after(TCP_SKB_CB(skb
)->end_seq
, loss_high
))
2361 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
2362 cnt
+= tcp_skb_pcount(skb
);
2367 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
))
2368 tcp_mark_skb_lost(sk
, skb
);
2373 tcp_verify_left_out(tp
);
2376 /* Account newly detected lost packet(s) */
2378 static void tcp_update_scoreboard(struct sock
*sk
, int fast_rexmit
)
2380 struct tcp_sock
*tp
= tcp_sk(sk
);
2382 if (tcp_is_sack(tp
)) {
2383 int sacked_upto
= tp
->sacked_out
- tp
->reordering
;
2384 if (sacked_upto
>= 0)
2385 tcp_mark_head_lost(sk
, sacked_upto
, 0);
2386 else if (fast_rexmit
)
2387 tcp_mark_head_lost(sk
, 1, 1);
2391 static bool tcp_tsopt_ecr_before(const struct tcp_sock
*tp
, u32 when
)
2393 return tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2394 before(tp
->rx_opt
.rcv_tsecr
, when
);
2397 /* skb is spurious retransmitted if the returned timestamp echo
2398 * reply is prior to the skb transmission time
2400 static bool tcp_skb_spurious_retrans(const struct tcp_sock
*tp
,
2401 const struct sk_buff
*skb
)
2403 return (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
) &&
2404 tcp_tsopt_ecr_before(tp
, tcp_skb_timestamp(skb
));
2407 /* Nothing was retransmitted or returned timestamp is less
2408 * than timestamp of the first retransmission.
2410 static inline bool tcp_packet_delayed(const struct tcp_sock
*tp
)
2412 return tp
->retrans_stamp
&&
2413 tcp_tsopt_ecr_before(tp
, tp
->retrans_stamp
);
2416 /* Undo procedures. */
2418 /* We can clear retrans_stamp when there are no retransmissions in the
2419 * window. It would seem that it is trivially available for us in
2420 * tp->retrans_out, however, that kind of assumptions doesn't consider
2421 * what will happen if errors occur when sending retransmission for the
2422 * second time. ...It could the that such segment has only
2423 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2424 * the head skb is enough except for some reneging corner cases that
2425 * are not worth the effort.
2427 * Main reason for all this complexity is the fact that connection dying
2428 * time now depends on the validity of the retrans_stamp, in particular,
2429 * that successive retransmissions of a segment must not advance
2430 * retrans_stamp under any conditions.
2432 static bool tcp_any_retrans_done(const struct sock
*sk
)
2434 const struct tcp_sock
*tp
= tcp_sk(sk
);
2435 struct sk_buff
*skb
;
2437 if (tp
->retrans_out
)
2440 skb
= tcp_rtx_queue_head(sk
);
2441 if (unlikely(skb
&& TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
))
2447 static void DBGUNDO(struct sock
*sk
, const char *msg
)
2449 #if FASTRETRANS_DEBUG > 1
2450 struct tcp_sock
*tp
= tcp_sk(sk
);
2451 struct inet_sock
*inet
= inet_sk(sk
);
2453 if (sk
->sk_family
== AF_INET
) {
2454 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2456 &inet
->inet_daddr
, ntohs(inet
->inet_dport
),
2457 tcp_snd_cwnd(tp
), tcp_left_out(tp
),
2458 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2461 #if IS_ENABLED(CONFIG_IPV6)
2462 else if (sk
->sk_family
== AF_INET6
) {
2463 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2465 &sk
->sk_v6_daddr
, ntohs(inet
->inet_dport
),
2466 tcp_snd_cwnd(tp
), tcp_left_out(tp
),
2467 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2474 static void tcp_undo_cwnd_reduction(struct sock
*sk
, bool unmark_loss
)
2476 struct tcp_sock
*tp
= tcp_sk(sk
);
2479 struct sk_buff
*skb
;
2481 skb_rbtree_walk(skb
, &sk
->tcp_rtx_queue
) {
2482 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2485 tcp_clear_all_retrans_hints(tp
);
2488 if (tp
->prior_ssthresh
) {
2489 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2491 tcp_snd_cwnd_set(tp
, icsk
->icsk_ca_ops
->undo_cwnd(sk
));
2493 if (tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
2494 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
2495 tcp_ecn_withdraw_cwr(tp
);
2498 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
2499 tp
->undo_marker
= 0;
2500 tp
->rack
.advanced
= 1; /* Force RACK to re-exam losses */
2503 static inline bool tcp_may_undo(const struct tcp_sock
*tp
)
2505 return tp
->undo_marker
&& (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
2508 /* People celebrate: "We love our President!" */
2509 static bool tcp_try_undo_recovery(struct sock
*sk
)
2511 struct tcp_sock
*tp
= tcp_sk(sk
);
2513 if (tcp_may_undo(tp
)) {
2516 /* Happy end! We did not retransmit anything
2517 * or our original transmission succeeded.
2519 DBGUNDO(sk
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
2520 tcp_undo_cwnd_reduction(sk
, false);
2521 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
2522 mib_idx
= LINUX_MIB_TCPLOSSUNDO
;
2524 mib_idx
= LINUX_MIB_TCPFULLUNDO
;
2526 NET_INC_STATS(sock_net(sk
), mib_idx
);
2527 } else if (tp
->rack
.reo_wnd_persist
) {
2528 tp
->rack
.reo_wnd_persist
--;
2530 if (tp
->snd_una
== tp
->high_seq
&& tcp_is_reno(tp
)) {
2531 /* Hold old state until something *above* high_seq
2532 * is ACKed. For Reno it is MUST to prevent false
2533 * fast retransmits (RFC2582). SACK TCP is safe. */
2534 if (!tcp_any_retrans_done(sk
))
2535 tp
->retrans_stamp
= 0;
2538 tcp_set_ca_state(sk
, TCP_CA_Open
);
2539 tp
->is_sack_reneg
= 0;
2543 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2544 static bool tcp_try_undo_dsack(struct sock
*sk
)
2546 struct tcp_sock
*tp
= tcp_sk(sk
);
2548 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
2549 tp
->rack
.reo_wnd_persist
= min(TCP_RACK_RECOVERY_THRESH
,
2550 tp
->rack
.reo_wnd_persist
+ 1);
2551 DBGUNDO(sk
, "D-SACK");
2552 tcp_undo_cwnd_reduction(sk
, false);
2553 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPDSACKUNDO
);
2559 /* Undo during loss recovery after partial ACK or using F-RTO. */
2560 static bool tcp_try_undo_loss(struct sock
*sk
, bool frto_undo
)
2562 struct tcp_sock
*tp
= tcp_sk(sk
);
2564 if (frto_undo
|| tcp_may_undo(tp
)) {
2565 tcp_undo_cwnd_reduction(sk
, true);
2567 DBGUNDO(sk
, "partial loss");
2568 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPLOSSUNDO
);
2570 NET_INC_STATS(sock_net(sk
),
2571 LINUX_MIB_TCPSPURIOUSRTOS
);
2572 inet_csk(sk
)->icsk_retransmits
= 0;
2573 if (frto_undo
|| tcp_is_sack(tp
)) {
2574 tcp_set_ca_state(sk
, TCP_CA_Open
);
2575 tp
->is_sack_reneg
= 0;
2582 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2583 * It computes the number of packets to send (sndcnt) based on packets newly
2585 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2586 * cwnd reductions across a full RTT.
2587 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2588 * But when SND_UNA is acked without further losses,
2589 * slow starts cwnd up to ssthresh to speed up the recovery.
2591 static void tcp_init_cwnd_reduction(struct sock
*sk
)
2593 struct tcp_sock
*tp
= tcp_sk(sk
);
2595 tp
->high_seq
= tp
->snd_nxt
;
2596 tp
->tlp_high_seq
= 0;
2597 tp
->snd_cwnd_cnt
= 0;
2598 tp
->prior_cwnd
= tcp_snd_cwnd(tp
);
2599 tp
->prr_delivered
= 0;
2601 tp
->snd_ssthresh
= inet_csk(sk
)->icsk_ca_ops
->ssthresh(sk
);
2602 tcp_ecn_queue_cwr(tp
);
2605 void tcp_cwnd_reduction(struct sock
*sk
, int newly_acked_sacked
, int newly_lost
, int flag
)
2607 struct tcp_sock
*tp
= tcp_sk(sk
);
2609 int delta
= tp
->snd_ssthresh
- tcp_packets_in_flight(tp
);
2611 if (newly_acked_sacked
<= 0 || WARN_ON_ONCE(!tp
->prior_cwnd
))
2614 tp
->prr_delivered
+= newly_acked_sacked
;
2616 u64 dividend
= (u64
)tp
->snd_ssthresh
* tp
->prr_delivered
+
2618 sndcnt
= div_u64(dividend
, tp
->prior_cwnd
) - tp
->prr_out
;
2619 } else if (flag
& FLAG_SND_UNA_ADVANCED
&& !newly_lost
) {
2620 sndcnt
= min_t(int, delta
,
2621 max_t(int, tp
->prr_delivered
- tp
->prr_out
,
2622 newly_acked_sacked
) + 1);
2624 sndcnt
= min(delta
, newly_acked_sacked
);
2626 /* Force a fast retransmit upon entering fast recovery */
2627 sndcnt
= max(sndcnt
, (tp
->prr_out
? 0 : 1));
2628 tcp_snd_cwnd_set(tp
, tcp_packets_in_flight(tp
) + sndcnt
);
2631 static inline void tcp_end_cwnd_reduction(struct sock
*sk
)
2633 struct tcp_sock
*tp
= tcp_sk(sk
);
2635 if (inet_csk(sk
)->icsk_ca_ops
->cong_control
)
2638 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2639 if (tp
->snd_ssthresh
< TCP_INFINITE_SSTHRESH
&&
2640 (inet_csk(sk
)->icsk_ca_state
== TCP_CA_CWR
|| tp
->undo_marker
)) {
2641 tcp_snd_cwnd_set(tp
, tp
->snd_ssthresh
);
2642 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
2644 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
2647 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2648 void tcp_enter_cwr(struct sock
*sk
)
2650 struct tcp_sock
*tp
= tcp_sk(sk
);
2652 tp
->prior_ssthresh
= 0;
2653 if (inet_csk(sk
)->icsk_ca_state
< TCP_CA_CWR
) {
2654 tp
->undo_marker
= 0;
2655 tcp_init_cwnd_reduction(sk
);
2656 tcp_set_ca_state(sk
, TCP_CA_CWR
);
2659 EXPORT_SYMBOL(tcp_enter_cwr
);
2661 static void tcp_try_keep_open(struct sock
*sk
)
2663 struct tcp_sock
*tp
= tcp_sk(sk
);
2664 int state
= TCP_CA_Open
;
2666 if (tcp_left_out(tp
) || tcp_any_retrans_done(sk
))
2667 state
= TCP_CA_Disorder
;
2669 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
2670 tcp_set_ca_state(sk
, state
);
2671 tp
->high_seq
= tp
->snd_nxt
;
2675 static void tcp_try_to_open(struct sock
*sk
, int flag
)
2677 struct tcp_sock
*tp
= tcp_sk(sk
);
2679 tcp_verify_left_out(tp
);
2681 if (!tcp_any_retrans_done(sk
))
2682 tp
->retrans_stamp
= 0;
2684 if (flag
& FLAG_ECE
)
2687 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
2688 tcp_try_keep_open(sk
);
2692 static void tcp_mtup_probe_failed(struct sock
*sk
)
2694 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2696 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
2697 icsk
->icsk_mtup
.probe_size
= 0;
2698 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPMTUPFAIL
);
2701 static void tcp_mtup_probe_success(struct sock
*sk
)
2703 struct tcp_sock
*tp
= tcp_sk(sk
);
2704 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2707 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2709 val
= (u64
)tcp_snd_cwnd(tp
) * tcp_mss_to_mtu(sk
, tp
->mss_cache
);
2710 do_div(val
, icsk
->icsk_mtup
.probe_size
);
2711 WARN_ON_ONCE((u32
)val
!= val
);
2712 tcp_snd_cwnd_set(tp
, max_t(u32
, 1U, val
));
2714 tp
->snd_cwnd_cnt
= 0;
2715 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
2716 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2718 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
2719 icsk
->icsk_mtup
.probe_size
= 0;
2720 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
2721 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPMTUPSUCCESS
);
2724 /* Do a simple retransmit without using the backoff mechanisms in
2725 * tcp_timer. This is used for path mtu discovery.
2726 * The socket is already locked here.
2728 void tcp_simple_retransmit(struct sock
*sk
)
2730 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2731 struct tcp_sock
*tp
= tcp_sk(sk
);
2732 struct sk_buff
*skb
;
2735 /* A fastopen SYN request is stored as two separate packets within
2736 * the retransmit queue, this is done by tcp_send_syn_data().
2737 * As a result simply checking the MSS of the frames in the queue
2738 * will not work for the SYN packet.
2740 * Us being here is an indication of a path MTU issue so we can
2741 * assume that the fastopen SYN was lost and just mark all the
2742 * frames in the retransmit queue as lost. We will use an MSS of
2743 * -1 to mark all frames as lost, otherwise compute the current MSS.
2745 if (tp
->syn_data
&& sk
->sk_state
== TCP_SYN_SENT
)
2748 mss
= tcp_current_mss(sk
);
2750 skb_rbtree_walk(skb
, &sk
->tcp_rtx_queue
) {
2751 if (tcp_skb_seglen(skb
) > mss
)
2752 tcp_mark_skb_lost(sk
, skb
);
2755 tcp_clear_retrans_hints_partial(tp
);
2760 if (tcp_is_reno(tp
))
2761 tcp_limit_reno_sacked(tp
);
2763 tcp_verify_left_out(tp
);
2765 /* Don't muck with the congestion window here.
2766 * Reason is that we do not increase amount of _data_
2767 * in network, but units changed and effective
2768 * cwnd/ssthresh really reduced now.
2770 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
2771 tp
->high_seq
= tp
->snd_nxt
;
2772 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2773 tp
->prior_ssthresh
= 0;
2774 tp
->undo_marker
= 0;
2775 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2777 tcp_xmit_retransmit_queue(sk
);
2779 EXPORT_SYMBOL(tcp_simple_retransmit
);
2781 void tcp_enter_recovery(struct sock
*sk
, bool ece_ack
)
2783 struct tcp_sock
*tp
= tcp_sk(sk
);
2786 if (tcp_is_reno(tp
))
2787 mib_idx
= LINUX_MIB_TCPRENORECOVERY
;
2789 mib_idx
= LINUX_MIB_TCPSACKRECOVERY
;
2791 NET_INC_STATS(sock_net(sk
), mib_idx
);
2793 tp
->prior_ssthresh
= 0;
2796 if (!tcp_in_cwnd_reduction(sk
)) {
2798 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2799 tcp_init_cwnd_reduction(sk
);
2801 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
2804 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2805 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2807 static void tcp_process_loss(struct sock
*sk
, int flag
, int num_dupack
,
2810 struct tcp_sock
*tp
= tcp_sk(sk
);
2811 bool recovered
= !before(tp
->snd_una
, tp
->high_seq
);
2813 if ((flag
& FLAG_SND_UNA_ADVANCED
|| rcu_access_pointer(tp
->fastopen_rsk
)) &&
2814 tcp_try_undo_loss(sk
, false))
2817 if (tp
->frto
) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2818 /* Step 3.b. A timeout is spurious if not all data are
2819 * lost, i.e., never-retransmitted data are (s)acked.
2821 if ((flag
& FLAG_ORIG_SACK_ACKED
) &&
2822 tcp_try_undo_loss(sk
, true))
2825 if (after(tp
->snd_nxt
, tp
->high_seq
)) {
2826 if (flag
& FLAG_DATA_SACKED
|| num_dupack
)
2827 tp
->frto
= 0; /* Step 3.a. loss was real */
2828 } else if (flag
& FLAG_SND_UNA_ADVANCED
&& !recovered
) {
2829 tp
->high_seq
= tp
->snd_nxt
;
2830 /* Step 2.b. Try send new data (but deferred until cwnd
2831 * is updated in tcp_ack()). Otherwise fall back to
2832 * the conventional recovery.
2834 if (!tcp_write_queue_empty(sk
) &&
2835 after(tcp_wnd_end(tp
), tp
->snd_nxt
)) {
2836 *rexmit
= REXMIT_NEW
;
2844 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2845 tcp_try_undo_recovery(sk
);
2848 if (tcp_is_reno(tp
)) {
2849 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2850 * delivered. Lower inflight to clock out (re)tranmissions.
2852 if (after(tp
->snd_nxt
, tp
->high_seq
) && num_dupack
)
2853 tcp_add_reno_sack(sk
, num_dupack
, flag
& FLAG_ECE
);
2854 else if (flag
& FLAG_SND_UNA_ADVANCED
)
2855 tcp_reset_reno_sack(tp
);
2857 *rexmit
= REXMIT_LOST
;
2860 static bool tcp_force_fast_retransmit(struct sock
*sk
)
2862 struct tcp_sock
*tp
= tcp_sk(sk
);
2864 return after(tcp_highest_sack_seq(tp
),
2865 tp
->snd_una
+ tp
->reordering
* tp
->mss_cache
);
2868 /* Undo during fast recovery after partial ACK. */
2869 static bool tcp_try_undo_partial(struct sock
*sk
, u32 prior_snd_una
,
2872 struct tcp_sock
*tp
= tcp_sk(sk
);
2874 if (tp
->undo_marker
&& tcp_packet_delayed(tp
)) {
2875 /* Plain luck! Hole if filled with delayed
2876 * packet, rather than with a retransmit. Check reordering.
2878 tcp_check_sack_reordering(sk
, prior_snd_una
, 1);
2880 /* We are getting evidence that the reordering degree is higher
2881 * than we realized. If there are no retransmits out then we
2882 * can undo. Otherwise we clock out new packets but do not
2883 * mark more packets lost or retransmit more.
2885 if (tp
->retrans_out
)
2888 if (!tcp_any_retrans_done(sk
))
2889 tp
->retrans_stamp
= 0;
2891 DBGUNDO(sk
, "partial recovery");
2892 tcp_undo_cwnd_reduction(sk
, true);
2893 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPPARTIALUNDO
);
2894 tcp_try_keep_open(sk
);
2896 /* Partial ACK arrived. Force fast retransmit. */
2897 *do_lost
= tcp_force_fast_retransmit(sk
);
2902 static void tcp_identify_packet_loss(struct sock
*sk
, int *ack_flag
)
2904 struct tcp_sock
*tp
= tcp_sk(sk
);
2906 if (tcp_rtx_queue_empty(sk
))
2909 if (unlikely(tcp_is_reno(tp
))) {
2910 tcp_newreno_mark_lost(sk
, *ack_flag
& FLAG_SND_UNA_ADVANCED
);
2911 } else if (tcp_is_rack(sk
)) {
2912 u32 prior_retrans
= tp
->retrans_out
;
2914 if (tcp_rack_mark_lost(sk
))
2915 *ack_flag
&= ~FLAG_SET_XMIT_TIMER
;
2916 if (prior_retrans
> tp
->retrans_out
)
2917 *ack_flag
|= FLAG_LOST_RETRANS
;
2921 /* Process an event, which can update packets-in-flight not trivially.
2922 * Main goal of this function is to calculate new estimate for left_out,
2923 * taking into account both packets sitting in receiver's buffer and
2924 * packets lost by network.
2926 * Besides that it updates the congestion state when packet loss or ECN
2927 * is detected. But it does not reduce the cwnd, it is done by the
2928 * congestion control later.
2930 * It does _not_ decide what to send, it is made in function
2931 * tcp_xmit_retransmit_queue().
2933 static void tcp_fastretrans_alert(struct sock
*sk
, const u32 prior_snd_una
,
2934 int num_dupack
, int *ack_flag
, int *rexmit
)
2936 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2937 struct tcp_sock
*tp
= tcp_sk(sk
);
2938 int fast_rexmit
= 0, flag
= *ack_flag
;
2939 bool ece_ack
= flag
& FLAG_ECE
;
2940 bool do_lost
= num_dupack
|| ((flag
& FLAG_DATA_SACKED
) &&
2941 tcp_force_fast_retransmit(sk
));
2943 if (!tp
->packets_out
&& tp
->sacked_out
)
2946 /* Now state machine starts.
2947 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2949 tp
->prior_ssthresh
= 0;
2951 /* B. In all the states check for reneging SACKs. */
2952 if (tcp_check_sack_reneging(sk
, flag
))
2955 /* C. Check consistency of the current state. */
2956 tcp_verify_left_out(tp
);
2958 /* D. Check state exit conditions. State can be terminated
2959 * when high_seq is ACKed. */
2960 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
2961 WARN_ON(tp
->retrans_out
!= 0 && !tp
->syn_data
);
2962 tp
->retrans_stamp
= 0;
2963 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
2964 switch (icsk
->icsk_ca_state
) {
2966 /* CWR is to be held something *above* high_seq
2967 * is ACKed for CWR bit to reach receiver. */
2968 if (tp
->snd_una
!= tp
->high_seq
) {
2969 tcp_end_cwnd_reduction(sk
);
2970 tcp_set_ca_state(sk
, TCP_CA_Open
);
2974 case TCP_CA_Recovery
:
2975 if (tcp_is_reno(tp
))
2976 tcp_reset_reno_sack(tp
);
2977 if (tcp_try_undo_recovery(sk
))
2979 tcp_end_cwnd_reduction(sk
);
2984 /* E. Process state. */
2985 switch (icsk
->icsk_ca_state
) {
2986 case TCP_CA_Recovery
:
2987 if (!(flag
& FLAG_SND_UNA_ADVANCED
)) {
2988 if (tcp_is_reno(tp
))
2989 tcp_add_reno_sack(sk
, num_dupack
, ece_ack
);
2990 } else if (tcp_try_undo_partial(sk
, prior_snd_una
, &do_lost
))
2993 if (tcp_try_undo_dsack(sk
))
2994 tcp_try_keep_open(sk
);
2996 tcp_identify_packet_loss(sk
, ack_flag
);
2997 if (icsk
->icsk_ca_state
!= TCP_CA_Recovery
) {
2998 if (!tcp_time_to_recover(sk
, flag
))
3000 /* Undo reverts the recovery state. If loss is evident,
3001 * starts a new recovery (e.g. reordering then loss);
3003 tcp_enter_recovery(sk
, ece_ack
);
3007 tcp_process_loss(sk
, flag
, num_dupack
, rexmit
);
3008 tcp_identify_packet_loss(sk
, ack_flag
);
3009 if (!(icsk
->icsk_ca_state
== TCP_CA_Open
||
3010 (*ack_flag
& FLAG_LOST_RETRANS
)))
3012 /* Change state if cwnd is undone or retransmits are lost */
3015 if (tcp_is_reno(tp
)) {
3016 if (flag
& FLAG_SND_UNA_ADVANCED
)
3017 tcp_reset_reno_sack(tp
);
3018 tcp_add_reno_sack(sk
, num_dupack
, ece_ack
);
3021 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
)
3022 tcp_try_undo_dsack(sk
);
3024 tcp_identify_packet_loss(sk
, ack_flag
);
3025 if (!tcp_time_to_recover(sk
, flag
)) {
3026 tcp_try_to_open(sk
, flag
);
3030 /* MTU probe failure: don't reduce cwnd */
3031 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
3032 icsk
->icsk_mtup
.probe_size
&&
3033 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
3034 tcp_mtup_probe_failed(sk
);
3035 /* Restores the reduction we did in tcp_mtup_probe() */
3036 tcp_snd_cwnd_set(tp
, tcp_snd_cwnd(tp
) + 1);
3037 tcp_simple_retransmit(sk
);
3041 /* Otherwise enter Recovery state */
3042 tcp_enter_recovery(sk
, ece_ack
);
3046 if (!tcp_is_rack(sk
) && do_lost
)
3047 tcp_update_scoreboard(sk
, fast_rexmit
);
3048 *rexmit
= REXMIT_LOST
;
3051 static void tcp_update_rtt_min(struct sock
*sk
, u32 rtt_us
, const int flag
)
3053 u32 wlen
= READ_ONCE(sock_net(sk
)->ipv4
.sysctl_tcp_min_rtt_wlen
) * HZ
;
3054 struct tcp_sock
*tp
= tcp_sk(sk
);
3056 if ((flag
& FLAG_ACK_MAYBE_DELAYED
) && rtt_us
> tcp_min_rtt(tp
)) {
3057 /* If the remote keeps returning delayed ACKs, eventually
3058 * the min filter would pick it up and overestimate the
3059 * prop. delay when it expires. Skip suspected delayed ACKs.
3063 minmax_running_min(&tp
->rtt_min
, wlen
, tcp_jiffies32
,
3064 rtt_us
? : jiffies_to_usecs(1));
3067 static bool tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
3068 long seq_rtt_us
, long sack_rtt_us
,
3069 long ca_rtt_us
, struct rate_sample
*rs
)
3071 const struct tcp_sock
*tp
= tcp_sk(sk
);
3073 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
3074 * broken middle-boxes or peers may corrupt TS-ECR fields. But
3075 * Karn's algorithm forbids taking RTT if some retransmitted data
3076 * is acked (RFC6298).
3079 seq_rtt_us
= sack_rtt_us
;
3081 /* RTTM Rule: A TSecr value received in a segment is used to
3082 * update the averaged RTT measurement only if the segment
3083 * acknowledges some new data, i.e., only if it advances the
3084 * left edge of the send window.
3085 * See draft-ietf-tcplw-high-performance-00, section 3.3.
3087 if (seq_rtt_us
< 0 && tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
3088 flag
& FLAG_ACKED
) {
3089 u32 delta
= tcp_time_stamp(tp
) - tp
->rx_opt
.rcv_tsecr
;
3091 if (likely(delta
< INT_MAX
/ (USEC_PER_SEC
/ TCP_TS_HZ
))) {
3094 seq_rtt_us
= delta
* (USEC_PER_SEC
/ TCP_TS_HZ
);
3095 ca_rtt_us
= seq_rtt_us
;
3098 rs
->rtt_us
= ca_rtt_us
; /* RTT of last (S)ACKed packet (or -1) */
3102 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
3103 * always taken together with ACK, SACK, or TS-opts. Any negative
3104 * values will be skipped with the seq_rtt_us < 0 check above.
3106 tcp_update_rtt_min(sk
, ca_rtt_us
, flag
);
3107 tcp_rtt_estimator(sk
, seq_rtt_us
);
3110 /* RFC6298: only reset backoff on valid RTT measurement. */
3111 inet_csk(sk
)->icsk_backoff
= 0;
3115 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
3116 void tcp_synack_rtt_meas(struct sock
*sk
, struct request_sock
*req
)
3118 struct rate_sample rs
;
3121 if (req
&& !req
->num_retrans
&& tcp_rsk(req
)->snt_synack
)
3122 rtt_us
= tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req
)->snt_synack
);
3124 tcp_ack_update_rtt(sk
, FLAG_SYN_ACKED
, rtt_us
, -1L, rtt_us
, &rs
);
3128 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
, u32 acked
)
3130 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3132 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, acked
);
3133 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_jiffies32
;
3136 /* Restart timer after forward progress on connection.
3137 * RFC2988 recommends to restart timer to now+rto.
3139 void tcp_rearm_rto(struct sock
*sk
)
3141 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3142 struct tcp_sock
*tp
= tcp_sk(sk
);
3144 /* If the retrans timer is currently being used by Fast Open
3145 * for SYN-ACK retrans purpose, stay put.
3147 if (rcu_access_pointer(tp
->fastopen_rsk
))
3150 if (!tp
->packets_out
) {
3151 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
3153 u32 rto
= inet_csk(sk
)->icsk_rto
;
3154 /* Offset the time elapsed after installing regular RTO */
3155 if (icsk
->icsk_pending
== ICSK_TIME_REO_TIMEOUT
||
3156 icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
) {
3157 s64 delta_us
= tcp_rto_delta_us(sk
);
3158 /* delta_us may not be positive if the socket is locked
3159 * when the retrans timer fires and is rescheduled.
3161 rto
= usecs_to_jiffies(max_t(int, delta_us
, 1));
3163 tcp_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
, rto
,
3168 /* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */
3169 static void tcp_set_xmit_timer(struct sock
*sk
)
3171 if (!tcp_schedule_loss_probe(sk
, true))
3175 /* If we get here, the whole TSO packet has not been acked. */
3176 static u32
tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
)
3178 struct tcp_sock
*tp
= tcp_sk(sk
);
3181 BUG_ON(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
));
3183 packets_acked
= tcp_skb_pcount(skb
);
3184 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
3186 packets_acked
-= tcp_skb_pcount(skb
);
3188 if (packets_acked
) {
3189 BUG_ON(tcp_skb_pcount(skb
) == 0);
3190 BUG_ON(!before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
));
3193 return packets_acked
;
3196 static void tcp_ack_tstamp(struct sock
*sk
, struct sk_buff
*skb
,
3197 const struct sk_buff
*ack_skb
, u32 prior_snd_una
)
3199 const struct skb_shared_info
*shinfo
;
3201 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3202 if (likely(!TCP_SKB_CB(skb
)->txstamp_ack
))
3205 shinfo
= skb_shinfo(skb
);
3206 if (!before(shinfo
->tskey
, prior_snd_una
) &&
3207 before(shinfo
->tskey
, tcp_sk(sk
)->snd_una
)) {
3208 tcp_skb_tsorted_save(skb
) {
3209 __skb_tstamp_tx(skb
, ack_skb
, NULL
, sk
, SCM_TSTAMP_ACK
);
3210 } tcp_skb_tsorted_restore(skb
);
3214 /* Remove acknowledged frames from the retransmission queue. If our packet
3215 * is before the ack sequence we can discard it as it's confirmed to have
3216 * arrived at the other end.
3218 static int tcp_clean_rtx_queue(struct sock
*sk
, const struct sk_buff
*ack_skb
,
3219 u32 prior_fack
, u32 prior_snd_una
,
3220 struct tcp_sacktag_state
*sack
, bool ece_ack
)
3222 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3223 u64 first_ackt
, last_ackt
;
3224 struct tcp_sock
*tp
= tcp_sk(sk
);
3225 u32 prior_sacked
= tp
->sacked_out
;
3226 u32 reord
= tp
->snd_nxt
; /* lowest acked un-retx un-sacked seq */
3227 struct sk_buff
*skb
, *next
;
3228 bool fully_acked
= true;
3229 long sack_rtt_us
= -1L;
3230 long seq_rtt_us
= -1L;
3231 long ca_rtt_us
= -1L;
3233 u32 last_in_flight
= 0;
3239 for (skb
= skb_rb_first(&sk
->tcp_rtx_queue
); skb
; skb
= next
) {
3240 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
3241 const u32 start_seq
= scb
->seq
;
3242 u8 sacked
= scb
->sacked
;
3245 /* Determine how many packets and what bytes were acked, tso and else */
3246 if (after(scb
->end_seq
, tp
->snd_una
)) {
3247 if (tcp_skb_pcount(skb
) == 1 ||
3248 !after(tp
->snd_una
, scb
->seq
))
3251 acked_pcount
= tcp_tso_acked(sk
, skb
);
3254 fully_acked
= false;
3256 acked_pcount
= tcp_skb_pcount(skb
);
3259 if (unlikely(sacked
& TCPCB_RETRANS
)) {
3260 if (sacked
& TCPCB_SACKED_RETRANS
)
3261 tp
->retrans_out
-= acked_pcount
;
3262 flag
|= FLAG_RETRANS_DATA_ACKED
;
3263 } else if (!(sacked
& TCPCB_SACKED_ACKED
)) {
3264 last_ackt
= tcp_skb_timestamp_us(skb
);
3265 WARN_ON_ONCE(last_ackt
== 0);
3267 first_ackt
= last_ackt
;
3269 last_in_flight
= TCP_SKB_CB(skb
)->tx
.in_flight
;
3270 if (before(start_seq
, reord
))
3272 if (!after(scb
->end_seq
, tp
->high_seq
))
3273 flag
|= FLAG_ORIG_SACK_ACKED
;
3276 if (sacked
& TCPCB_SACKED_ACKED
) {
3277 tp
->sacked_out
-= acked_pcount
;
3278 } else if (tcp_is_sack(tp
)) {
3279 tcp_count_delivered(tp
, acked_pcount
, ece_ack
);
3280 if (!tcp_skb_spurious_retrans(tp
, skb
))
3281 tcp_rack_advance(tp
, sacked
, scb
->end_seq
,
3282 tcp_skb_timestamp_us(skb
));
3284 if (sacked
& TCPCB_LOST
)
3285 tp
->lost_out
-= acked_pcount
;
3287 tp
->packets_out
-= acked_pcount
;
3288 pkts_acked
+= acked_pcount
;
3289 tcp_rate_skb_delivered(sk
, skb
, sack
->rate
);
3291 /* Initial outgoing SYN's get put onto the write_queue
3292 * just like anything else we transmit. It is not
3293 * true data, and if we misinform our callers that
3294 * this ACK acks real data, we will erroneously exit
3295 * connection startup slow start one packet too
3296 * quickly. This is severely frowned upon behavior.
3298 if (likely(!(scb
->tcp_flags
& TCPHDR_SYN
))) {
3299 flag
|= FLAG_DATA_ACKED
;
3301 flag
|= FLAG_SYN_ACKED
;
3302 tp
->retrans_stamp
= 0;
3308 tcp_ack_tstamp(sk
, skb
, ack_skb
, prior_snd_una
);
3310 next
= skb_rb_next(skb
);
3311 if (unlikely(skb
== tp
->retransmit_skb_hint
))
3312 tp
->retransmit_skb_hint
= NULL
;
3313 if (unlikely(skb
== tp
->lost_skb_hint
))
3314 tp
->lost_skb_hint
= NULL
;
3315 tcp_highest_sack_replace(sk
, skb
, next
);
3316 tcp_rtx_queue_unlink_and_free(skb
, sk
);
3320 tcp_chrono_stop(sk
, TCP_CHRONO_BUSY
);
3322 if (likely(between(tp
->snd_up
, prior_snd_una
, tp
->snd_una
)))
3323 tp
->snd_up
= tp
->snd_una
;
3326 tcp_ack_tstamp(sk
, skb
, ack_skb
, prior_snd_una
);
3327 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
3328 flag
|= FLAG_SACK_RENEGING
;
3331 if (likely(first_ackt
) && !(flag
& FLAG_RETRANS_DATA_ACKED
)) {
3332 seq_rtt_us
= tcp_stamp_us_delta(tp
->tcp_mstamp
, first_ackt
);
3333 ca_rtt_us
= tcp_stamp_us_delta(tp
->tcp_mstamp
, last_ackt
);
3335 if (pkts_acked
== 1 && last_in_flight
< tp
->mss_cache
&&
3336 last_in_flight
&& !prior_sacked
&& fully_acked
&&
3337 sack
->rate
->prior_delivered
+ 1 == tp
->delivered
&&
3338 !(flag
& (FLAG_CA_ALERT
| FLAG_SYN_ACKED
))) {
3339 /* Conservatively mark a delayed ACK. It's typically
3340 * from a lone runt packet over the round trip to
3341 * a receiver w/o out-of-order or CE events.
3343 flag
|= FLAG_ACK_MAYBE_DELAYED
;
3346 if (sack
->first_sackt
) {
3347 sack_rtt_us
= tcp_stamp_us_delta(tp
->tcp_mstamp
, sack
->first_sackt
);
3348 ca_rtt_us
= tcp_stamp_us_delta(tp
->tcp_mstamp
, sack
->last_sackt
);
3350 rtt_update
= tcp_ack_update_rtt(sk
, flag
, seq_rtt_us
, sack_rtt_us
,
3351 ca_rtt_us
, sack
->rate
);
3353 if (flag
& FLAG_ACKED
) {
3354 flag
|= FLAG_SET_XMIT_TIMER
; /* set TLP or RTO timer */
3355 if (unlikely(icsk
->icsk_mtup
.probe_size
&&
3356 !after(tp
->mtu_probe
.probe_seq_end
, tp
->snd_una
))) {
3357 tcp_mtup_probe_success(sk
);
3360 if (tcp_is_reno(tp
)) {
3361 tcp_remove_reno_sacks(sk
, pkts_acked
, ece_ack
);
3363 /* If any of the cumulatively ACKed segments was
3364 * retransmitted, non-SACK case cannot confirm that
3365 * progress was due to original transmission due to
3366 * lack of TCPCB_SACKED_ACKED bits even if some of
3367 * the packets may have been never retransmitted.
3369 if (flag
& FLAG_RETRANS_DATA_ACKED
)
3370 flag
&= ~FLAG_ORIG_SACK_ACKED
;
3374 /* Non-retransmitted hole got filled? That's reordering */
3375 if (before(reord
, prior_fack
))
3376 tcp_check_sack_reordering(sk
, reord
, 0);
3378 delta
= prior_sacked
- tp
->sacked_out
;
3379 tp
->lost_cnt_hint
-= min(tp
->lost_cnt_hint
, delta
);
3381 } else if (skb
&& rtt_update
&& sack_rtt_us
>= 0 &&
3382 sack_rtt_us
> tcp_stamp_us_delta(tp
->tcp_mstamp
,
3383 tcp_skb_timestamp_us(skb
))) {
3384 /* Do not re-arm RTO if the sack RTT is measured from data sent
3385 * after when the head was last (re)transmitted. Otherwise the
3386 * timeout may continue to extend in loss recovery.
3388 flag
|= FLAG_SET_XMIT_TIMER
; /* set TLP or RTO timer */
3391 if (icsk
->icsk_ca_ops
->pkts_acked
) {
3392 struct ack_sample sample
= { .pkts_acked
= pkts_acked
,
3393 .rtt_us
= sack
->rate
->rtt_us
,
3394 .in_flight
= last_in_flight
};
3396 icsk
->icsk_ca_ops
->pkts_acked(sk
, &sample
);
3399 #if FASTRETRANS_DEBUG > 0
3400 WARN_ON((int)tp
->sacked_out
< 0);
3401 WARN_ON((int)tp
->lost_out
< 0);
3402 WARN_ON((int)tp
->retrans_out
< 0);
3403 if (!tp
->packets_out
&& tcp_is_sack(tp
)) {
3404 icsk
= inet_csk(sk
);
3406 pr_debug("Leak l=%u %d\n",
3407 tp
->lost_out
, icsk
->icsk_ca_state
);
3410 if (tp
->sacked_out
) {
3411 pr_debug("Leak s=%u %d\n",
3412 tp
->sacked_out
, icsk
->icsk_ca_state
);
3415 if (tp
->retrans_out
) {
3416 pr_debug("Leak r=%u %d\n",
3417 tp
->retrans_out
, icsk
->icsk_ca_state
);
3418 tp
->retrans_out
= 0;
3425 static void tcp_ack_probe(struct sock
*sk
)
3427 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3428 struct sk_buff
*head
= tcp_send_head(sk
);
3429 const struct tcp_sock
*tp
= tcp_sk(sk
);
3431 /* Was it a usable window open? */
3434 if (!after(TCP_SKB_CB(head
)->end_seq
, tcp_wnd_end(tp
))) {
3435 icsk
->icsk_backoff
= 0;
3436 icsk
->icsk_probes_tstamp
= 0;
3437 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
3438 /* Socket must be waked up by subsequent tcp_data_snd_check().
3439 * This function is not for random using!
3442 unsigned long when
= tcp_probe0_when(sk
, TCP_RTO_MAX
);
3444 when
= tcp_clamp_probe0_to_user_timeout(sk
, when
);
3445 tcp_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
, when
, TCP_RTO_MAX
);
3449 static inline bool tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
3451 return !(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
3452 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
;
3455 /* Decide wheather to run the increase function of congestion control. */
3456 static inline bool tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
3458 /* If reordering is high then always grow cwnd whenever data is
3459 * delivered regardless of its ordering. Otherwise stay conservative
3460 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3461 * new SACK or ECE mark may first advance cwnd here and later reduce
3462 * cwnd in tcp_fastretrans_alert() based on more states.
3464 if (tcp_sk(sk
)->reordering
>
3465 READ_ONCE(sock_net(sk
)->ipv4
.sysctl_tcp_reordering
))
3466 return flag
& FLAG_FORWARD_PROGRESS
;
3468 return flag
& FLAG_DATA_ACKED
;
3471 /* The "ultimate" congestion control function that aims to replace the rigid
3472 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3473 * It's called toward the end of processing an ACK with precise rate
3474 * information. All transmission or retransmission are delayed afterwards.
3476 static void tcp_cong_control(struct sock
*sk
, u32 ack
, u32 acked_sacked
,
3477 int flag
, const struct rate_sample
*rs
)
3479 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3481 if (icsk
->icsk_ca_ops
->cong_control
) {
3482 icsk
->icsk_ca_ops
->cong_control(sk
, rs
);
3486 if (tcp_in_cwnd_reduction(sk
)) {
3487 /* Reduce cwnd if state mandates */
3488 tcp_cwnd_reduction(sk
, acked_sacked
, rs
->losses
, flag
);
3489 } else if (tcp_may_raise_cwnd(sk
, flag
)) {
3490 /* Advance cwnd if state allows */
3491 tcp_cong_avoid(sk
, ack
, acked_sacked
);
3493 tcp_update_pacing_rate(sk
);
3496 /* Check that window update is acceptable.
3497 * The function assumes that snd_una<=ack<=snd_next.
3499 static inline bool tcp_may_update_window(const struct tcp_sock
*tp
,
3500 const u32 ack
, const u32 ack_seq
,
3503 return after(ack
, tp
->snd_una
) ||
3504 after(ack_seq
, tp
->snd_wl1
) ||
3505 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
);
3508 /* If we update tp->snd_una, also update tp->bytes_acked */
3509 static void tcp_snd_una_update(struct tcp_sock
*tp
, u32 ack
)
3511 u32 delta
= ack
- tp
->snd_una
;
3513 sock_owned_by_me((struct sock
*)tp
);
3514 tp
->bytes_acked
+= delta
;
3518 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3519 static void tcp_rcv_nxt_update(struct tcp_sock
*tp
, u32 seq
)
3521 u32 delta
= seq
- tp
->rcv_nxt
;
3523 sock_owned_by_me((struct sock
*)tp
);
3524 tp
->bytes_received
+= delta
;
3525 WRITE_ONCE(tp
->rcv_nxt
, seq
);
3528 /* Update our send window.
3530 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3531 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3533 static int tcp_ack_update_window(struct sock
*sk
, const struct sk_buff
*skb
, u32 ack
,
3536 struct tcp_sock
*tp
= tcp_sk(sk
);
3538 u32 nwin
= ntohs(tcp_hdr(skb
)->window
);
3540 if (likely(!tcp_hdr(skb
)->syn
))
3541 nwin
<<= tp
->rx_opt
.snd_wscale
;
3543 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
3544 flag
|= FLAG_WIN_UPDATE
;
3545 tcp_update_wl(tp
, ack_seq
);
3547 if (tp
->snd_wnd
!= nwin
) {
3550 /* Note, it is the only place, where
3551 * fast path is recovered for sending TCP.
3554 tcp_fast_path_check(sk
);
3556 if (!tcp_write_queue_empty(sk
))
3557 tcp_slow_start_after_idle_check(sk
);
3559 if (nwin
> tp
->max_window
) {
3560 tp
->max_window
= nwin
;
3561 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
3566 tcp_snd_una_update(tp
, ack
);
3571 static bool __tcp_oow_rate_limited(struct net
*net
, int mib_idx
,
3572 u32
*last_oow_ack_time
)
3574 if (*last_oow_ack_time
) {
3575 s32 elapsed
= (s32
)(tcp_jiffies32
- *last_oow_ack_time
);
3578 elapsed
< READ_ONCE(net
->ipv4
.sysctl_tcp_invalid_ratelimit
)) {
3579 NET_INC_STATS(net
, mib_idx
);
3580 return true; /* rate-limited: don't send yet! */
3584 *last_oow_ack_time
= tcp_jiffies32
;
3586 return false; /* not rate-limited: go ahead, send dupack now! */
3589 /* Return true if we're currently rate-limiting out-of-window ACKs and
3590 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3591 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3592 * attacks that send repeated SYNs or ACKs for the same connection. To
3593 * do this, we do not send a duplicate SYNACK or ACK if the remote
3594 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3596 bool tcp_oow_rate_limited(struct net
*net
, const struct sk_buff
*skb
,
3597 int mib_idx
, u32
*last_oow_ack_time
)
3599 /* Data packets without SYNs are not likely part of an ACK loop. */
3600 if ((TCP_SKB_CB(skb
)->seq
!= TCP_SKB_CB(skb
)->end_seq
) &&
3604 return __tcp_oow_rate_limited(net
, mib_idx
, last_oow_ack_time
);
3607 /* RFC 5961 7 [ACK Throttling] */
3608 static void tcp_send_challenge_ack(struct sock
*sk
, const struct sk_buff
*skb
)
3610 /* unprotected vars, we dont care of overwrites */
3611 static u32 challenge_timestamp
;
3612 static unsigned int challenge_count
;
3613 struct tcp_sock
*tp
= tcp_sk(sk
);
3614 struct net
*net
= sock_net(sk
);
3617 /* First check our per-socket dupack rate limit. */
3618 if (__tcp_oow_rate_limited(net
,
3619 LINUX_MIB_TCPACKSKIPPEDCHALLENGE
,
3620 &tp
->last_oow_ack_time
))
3623 /* Then check host-wide RFC 5961 rate limit. */
3625 if (now
!= challenge_timestamp
) {
3626 u32 ack_limit
= READ_ONCE(net
->ipv4
.sysctl_tcp_challenge_ack_limit
);
3627 u32 half
= (ack_limit
+ 1) >> 1;
3629 challenge_timestamp
= now
;
3630 WRITE_ONCE(challenge_count
, half
+ prandom_u32_max(ack_limit
));
3632 count
= READ_ONCE(challenge_count
);
3634 WRITE_ONCE(challenge_count
, count
- 1);
3635 NET_INC_STATS(net
, LINUX_MIB_TCPCHALLENGEACK
);
3640 static void tcp_store_ts_recent(struct tcp_sock
*tp
)
3642 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
3643 tp
->rx_opt
.ts_recent_stamp
= ktime_get_seconds();
3646 static void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
3648 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
3649 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3650 * extra check below makes sure this can only happen
3651 * for pure ACK frames. -DaveM
3653 * Not only, also it occurs for expired timestamps.
3656 if (tcp_paws_check(&tp
->rx_opt
, 0))
3657 tcp_store_ts_recent(tp
);
3661 /* This routine deals with acks during a TLP episode and ends an episode by
3662 * resetting tlp_high_seq. Ref: TLP algorithm in draft-ietf-tcpm-rack
3664 static void tcp_process_tlp_ack(struct sock
*sk
, u32 ack
, int flag
)
3666 struct tcp_sock
*tp
= tcp_sk(sk
);
3668 if (before(ack
, tp
->tlp_high_seq
))
3671 if (!tp
->tlp_retrans
) {
3672 /* TLP of new data has been acknowledged */
3673 tp
->tlp_high_seq
= 0;
3674 } else if (flag
& FLAG_DSACK_TLP
) {
3675 /* This DSACK means original and TLP probe arrived; no loss */
3676 tp
->tlp_high_seq
= 0;
3677 } else if (after(ack
, tp
->tlp_high_seq
)) {
3678 /* ACK advances: there was a loss, so reduce cwnd. Reset
3679 * tlp_high_seq in tcp_init_cwnd_reduction()
3681 tcp_init_cwnd_reduction(sk
);
3682 tcp_set_ca_state(sk
, TCP_CA_CWR
);
3683 tcp_end_cwnd_reduction(sk
);
3684 tcp_try_keep_open(sk
);
3685 NET_INC_STATS(sock_net(sk
),
3686 LINUX_MIB_TCPLOSSPROBERECOVERY
);
3687 } else if (!(flag
& (FLAG_SND_UNA_ADVANCED
|
3688 FLAG_NOT_DUP
| FLAG_DATA_SACKED
))) {
3689 /* Pure dupack: original and TLP probe arrived; no loss */
3690 tp
->tlp_high_seq
= 0;
3694 static inline void tcp_in_ack_event(struct sock
*sk
, u32 flags
)
3696 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3698 if (icsk
->icsk_ca_ops
->in_ack_event
)
3699 icsk
->icsk_ca_ops
->in_ack_event(sk
, flags
);
3702 /* Congestion control has updated the cwnd already. So if we're in
3703 * loss recovery then now we do any new sends (for FRTO) or
3704 * retransmits (for CA_Loss or CA_recovery) that make sense.
3706 static void tcp_xmit_recovery(struct sock
*sk
, int rexmit
)
3708 struct tcp_sock
*tp
= tcp_sk(sk
);
3710 if (rexmit
== REXMIT_NONE
|| sk
->sk_state
== TCP_SYN_SENT
)
3713 if (unlikely(rexmit
== REXMIT_NEW
)) {
3714 __tcp_push_pending_frames(sk
, tcp_current_mss(sk
),
3716 if (after(tp
->snd_nxt
, tp
->high_seq
))
3720 tcp_xmit_retransmit_queue(sk
);
3723 /* Returns the number of packets newly acked or sacked by the current ACK */
3724 static u32
tcp_newly_delivered(struct sock
*sk
, u32 prior_delivered
, int flag
)
3726 const struct net
*net
= sock_net(sk
);
3727 struct tcp_sock
*tp
= tcp_sk(sk
);
3730 delivered
= tp
->delivered
- prior_delivered
;
3731 NET_ADD_STATS(net
, LINUX_MIB_TCPDELIVERED
, delivered
);
3732 if (flag
& FLAG_ECE
)
3733 NET_ADD_STATS(net
, LINUX_MIB_TCPDELIVEREDCE
, delivered
);
3738 /* This routine deals with incoming acks, but not outgoing ones. */
3739 static int tcp_ack(struct sock
*sk
, const struct sk_buff
*skb
, int flag
)
3741 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3742 struct tcp_sock
*tp
= tcp_sk(sk
);
3743 struct tcp_sacktag_state sack_state
;
3744 struct rate_sample rs
= { .prior_delivered
= 0 };
3745 u32 prior_snd_una
= tp
->snd_una
;
3746 bool is_sack_reneg
= tp
->is_sack_reneg
;
3747 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
3748 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3750 int prior_packets
= tp
->packets_out
;
3751 u32 delivered
= tp
->delivered
;
3752 u32 lost
= tp
->lost
;
3753 int rexmit
= REXMIT_NONE
; /* Flag to (re)transmit to recover losses */
3756 sack_state
.first_sackt
= 0;
3757 sack_state
.rate
= &rs
;
3758 sack_state
.sack_delivered
= 0;
3760 /* We very likely will need to access rtx queue. */
3761 prefetch(sk
->tcp_rtx_queue
.rb_node
);
3763 /* If the ack is older than previous acks
3764 * then we can probably ignore it.
3766 if (before(ack
, prior_snd_una
)) {
3767 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3768 if (before(ack
, prior_snd_una
- tp
->max_window
)) {
3769 if (!(flag
& FLAG_NO_CHALLENGE_ACK
))
3770 tcp_send_challenge_ack(sk
, skb
);
3776 /* If the ack includes data we haven't sent yet, discard
3777 * this segment (RFC793 Section 3.9).
3779 if (after(ack
, tp
->snd_nxt
))
3782 if (after(ack
, prior_snd_una
)) {
3783 flag
|= FLAG_SND_UNA_ADVANCED
;
3784 icsk
->icsk_retransmits
= 0;
3786 #if IS_ENABLED(CONFIG_TLS_DEVICE)
3787 if (static_branch_unlikely(&clean_acked_data_enabled
.key
))
3788 if (icsk
->icsk_clean_acked
)
3789 icsk
->icsk_clean_acked(sk
, ack
);
3793 prior_fack
= tcp_is_sack(tp
) ? tcp_highest_sack_seq(tp
) : tp
->snd_una
;
3794 rs
.prior_in_flight
= tcp_packets_in_flight(tp
);
3796 /* ts_recent update must be made after we are sure that the packet
3799 if (flag
& FLAG_UPDATE_TS_RECENT
)
3800 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
3802 if ((flag
& (FLAG_SLOWPATH
| FLAG_SND_UNA_ADVANCED
)) ==
3803 FLAG_SND_UNA_ADVANCED
) {
3804 /* Window is constant, pure forward advance.
3805 * No more checks are required.
3806 * Note, we use the fact that SND.UNA>=SND.WL2.
3808 tcp_update_wl(tp
, ack_seq
);
3809 tcp_snd_una_update(tp
, ack
);
3810 flag
|= FLAG_WIN_UPDATE
;
3812 tcp_in_ack_event(sk
, CA_ACK_WIN_UPDATE
);
3814 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPHPACKS
);
3816 u32 ack_ev_flags
= CA_ACK_SLOWPATH
;
3818 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
3821 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPPUREACKS
);
3823 flag
|= tcp_ack_update_window(sk
, skb
, ack
, ack_seq
);
3825 if (TCP_SKB_CB(skb
)->sacked
)
3826 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
,
3829 if (tcp_ecn_rcv_ecn_echo(tp
, tcp_hdr(skb
))) {
3831 ack_ev_flags
|= CA_ACK_ECE
;
3834 if (sack_state
.sack_delivered
)
3835 tcp_count_delivered(tp
, sack_state
.sack_delivered
,
3838 if (flag
& FLAG_WIN_UPDATE
)
3839 ack_ev_flags
|= CA_ACK_WIN_UPDATE
;
3841 tcp_in_ack_event(sk
, ack_ev_flags
);
3844 /* This is a deviation from RFC3168 since it states that:
3845 * "When the TCP data sender is ready to set the CWR bit after reducing
3846 * the congestion window, it SHOULD set the CWR bit only on the first
3847 * new data packet that it transmits."
3848 * We accept CWR on pure ACKs to be more robust
3849 * with widely-deployed TCP implementations that do this.
3851 tcp_ecn_accept_cwr(sk
, skb
);
3853 /* We passed data and got it acked, remove any soft error
3854 * log. Something worked...
3856 sk
->sk_err_soft
= 0;
3857 icsk
->icsk_probes_out
= 0;
3858 tp
->rcv_tstamp
= tcp_jiffies32
;
3862 /* See if we can take anything off of the retransmit queue. */
3863 flag
|= tcp_clean_rtx_queue(sk
, skb
, prior_fack
, prior_snd_una
,
3864 &sack_state
, flag
& FLAG_ECE
);
3866 tcp_rack_update_reo_wnd(sk
, &rs
);
3868 if (tp
->tlp_high_seq
)
3869 tcp_process_tlp_ack(sk
, ack
, flag
);
3871 if (tcp_ack_is_dubious(sk
, flag
)) {
3872 if (!(flag
& (FLAG_SND_UNA_ADVANCED
|
3873 FLAG_NOT_DUP
| FLAG_DSACKING_ACK
))) {
3875 /* Consider if pure acks were aggregated in tcp_add_backlog() */
3876 if (!(flag
& FLAG_DATA
))
3877 num_dupack
= max_t(u16
, 1, skb_shinfo(skb
)->gso_segs
);
3879 tcp_fastretrans_alert(sk
, prior_snd_una
, num_dupack
, &flag
,
3883 /* If needed, reset TLP/RTO timer when RACK doesn't set. */
3884 if (flag
& FLAG_SET_XMIT_TIMER
)
3885 tcp_set_xmit_timer(sk
);
3887 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
& FLAG_NOT_DUP
))
3890 delivered
= tcp_newly_delivered(sk
, delivered
, flag
);
3891 lost
= tp
->lost
- lost
; /* freshly marked lost */
3892 rs
.is_ack_delayed
= !!(flag
& FLAG_ACK_MAYBE_DELAYED
);
3893 tcp_rate_gen(sk
, delivered
, lost
, is_sack_reneg
, sack_state
.rate
);
3894 tcp_cong_control(sk
, ack
, delivered
, flag
, sack_state
.rate
);
3895 tcp_xmit_recovery(sk
, rexmit
);
3899 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3900 if (flag
& FLAG_DSACKING_ACK
) {
3901 tcp_fastretrans_alert(sk
, prior_snd_una
, num_dupack
, &flag
,
3903 tcp_newly_delivered(sk
, delivered
, flag
);
3905 /* If this ack opens up a zero window, clear backoff. It was
3906 * being used to time the probes, and is probably far higher than
3907 * it needs to be for normal retransmission.
3911 if (tp
->tlp_high_seq
)
3912 tcp_process_tlp_ack(sk
, ack
, flag
);
3916 /* If data was SACKed, tag it and see if we should send more data.
3917 * If data was DSACKed, see if we can undo a cwnd reduction.
3919 if (TCP_SKB_CB(skb
)->sacked
) {
3920 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
,
3922 tcp_fastretrans_alert(sk
, prior_snd_una
, num_dupack
, &flag
,
3924 tcp_newly_delivered(sk
, delivered
, flag
);
3925 tcp_xmit_recovery(sk
, rexmit
);
3931 static void tcp_parse_fastopen_option(int len
, const unsigned char *cookie
,
3932 bool syn
, struct tcp_fastopen_cookie
*foc
,
3935 /* Valid only in SYN or SYN-ACK with an even length. */
3936 if (!foc
|| !syn
|| len
< 0 || (len
& 1))
3939 if (len
>= TCP_FASTOPEN_COOKIE_MIN
&&
3940 len
<= TCP_FASTOPEN_COOKIE_MAX
)
3941 memcpy(foc
->val
, cookie
, len
);
3948 static bool smc_parse_options(const struct tcphdr
*th
,
3949 struct tcp_options_received
*opt_rx
,
3950 const unsigned char *ptr
,
3953 #if IS_ENABLED(CONFIG_SMC)
3954 if (static_branch_unlikely(&tcp_have_smc
)) {
3955 if (th
->syn
&& !(opsize
& 1) &&
3956 opsize
>= TCPOLEN_EXP_SMC_BASE
&&
3957 get_unaligned_be32(ptr
) == TCPOPT_SMC_MAGIC
) {
3966 /* Try to parse the MSS option from the TCP header. Return 0 on failure, clamped
3969 static u16
tcp_parse_mss_option(const struct tcphdr
*th
, u16 user_mss
)
3971 const unsigned char *ptr
= (const unsigned char *)(th
+ 1);
3972 int length
= (th
->doff
* 4) - sizeof(struct tcphdr
);
3975 while (length
> 0) {
3976 int opcode
= *ptr
++;
3982 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
3989 if (opsize
< 2) /* "silly options" */
3991 if (opsize
> length
)
3992 return mss
; /* fail on partial options */
3993 if (opcode
== TCPOPT_MSS
&& opsize
== TCPOLEN_MSS
) {
3994 u16 in_mss
= get_unaligned_be16(ptr
);
3997 if (user_mss
&& user_mss
< in_mss
)
4009 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
4010 * But, this can also be called on packets in the established flow when
4011 * the fast version below fails.
4013 void tcp_parse_options(const struct net
*net
,
4014 const struct sk_buff
*skb
,
4015 struct tcp_options_received
*opt_rx
, int estab
,
4016 struct tcp_fastopen_cookie
*foc
)
4018 const unsigned char *ptr
;
4019 const struct tcphdr
*th
= tcp_hdr(skb
);
4020 int length
= (th
->doff
* 4) - sizeof(struct tcphdr
);
4022 ptr
= (const unsigned char *)(th
+ 1);
4023 opt_rx
->saw_tstamp
= 0;
4024 opt_rx
->saw_unknown
= 0;
4026 while (length
> 0) {
4027 int opcode
= *ptr
++;
4033 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
4040 if (opsize
< 2) /* "silly options" */
4042 if (opsize
> length
)
4043 return; /* don't parse partial options */
4046 if (opsize
== TCPOLEN_MSS
&& th
->syn
&& !estab
) {
4047 u16 in_mss
= get_unaligned_be16(ptr
);
4049 if (opt_rx
->user_mss
&&
4050 opt_rx
->user_mss
< in_mss
)
4051 in_mss
= opt_rx
->user_mss
;
4052 opt_rx
->mss_clamp
= in_mss
;
4057 if (opsize
== TCPOLEN_WINDOW
&& th
->syn
&&
4058 !estab
&& READ_ONCE(net
->ipv4
.sysctl_tcp_window_scaling
)) {
4059 __u8 snd_wscale
= *(__u8
*)ptr
;
4060 opt_rx
->wscale_ok
= 1;
4061 if (snd_wscale
> TCP_MAX_WSCALE
) {
4062 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n",
4066 snd_wscale
= TCP_MAX_WSCALE
;
4068 opt_rx
->snd_wscale
= snd_wscale
;
4071 case TCPOPT_TIMESTAMP
:
4072 if ((opsize
== TCPOLEN_TIMESTAMP
) &&
4073 ((estab
&& opt_rx
->tstamp_ok
) ||
4074 (!estab
&& READ_ONCE(net
->ipv4
.sysctl_tcp_timestamps
)))) {
4075 opt_rx
->saw_tstamp
= 1;
4076 opt_rx
->rcv_tsval
= get_unaligned_be32(ptr
);
4077 opt_rx
->rcv_tsecr
= get_unaligned_be32(ptr
+ 4);
4080 case TCPOPT_SACK_PERM
:
4081 if (opsize
== TCPOLEN_SACK_PERM
&& th
->syn
&&
4082 !estab
&& READ_ONCE(net
->ipv4
.sysctl_tcp_sack
)) {
4083 opt_rx
->sack_ok
= TCP_SACK_SEEN
;
4084 tcp_sack_reset(opt_rx
);
4089 if ((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
4090 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
4092 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
4095 #ifdef CONFIG_TCP_MD5SIG
4098 * The MD5 Hash has already been
4099 * checked (see tcp_v{4,6}_do_rcv()).
4103 case TCPOPT_FASTOPEN
:
4104 tcp_parse_fastopen_option(
4105 opsize
- TCPOLEN_FASTOPEN_BASE
,
4106 ptr
, th
->syn
, foc
, false);
4110 /* Fast Open option shares code 254 using a
4111 * 16 bits magic number.
4113 if (opsize
>= TCPOLEN_EXP_FASTOPEN_BASE
&&
4114 get_unaligned_be16(ptr
) ==
4115 TCPOPT_FASTOPEN_MAGIC
) {
4116 tcp_parse_fastopen_option(opsize
-
4117 TCPOLEN_EXP_FASTOPEN_BASE
,
4118 ptr
+ 2, th
->syn
, foc
, true);
4122 if (smc_parse_options(th
, opt_rx
, ptr
, opsize
))
4125 opt_rx
->saw_unknown
= 1;
4129 opt_rx
->saw_unknown
= 1;
4136 EXPORT_SYMBOL(tcp_parse_options
);
4138 static bool tcp_parse_aligned_timestamp(struct tcp_sock
*tp
, const struct tcphdr
*th
)
4140 const __be32
*ptr
= (const __be32
*)(th
+ 1);
4142 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
4143 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
4144 tp
->rx_opt
.saw_tstamp
= 1;
4146 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
4149 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
) - tp
->tsoffset
;
4151 tp
->rx_opt
.rcv_tsecr
= 0;
4157 /* Fast parse options. This hopes to only see timestamps.
4158 * If it is wrong it falls back on tcp_parse_options().
4160 static bool tcp_fast_parse_options(const struct net
*net
,
4161 const struct sk_buff
*skb
,
4162 const struct tcphdr
*th
, struct tcp_sock
*tp
)
4164 /* In the spirit of fast parsing, compare doff directly to constant
4165 * values. Because equality is used, short doff can be ignored here.
4167 if (th
->doff
== (sizeof(*th
) / 4)) {
4168 tp
->rx_opt
.saw_tstamp
= 0;
4170 } else if (tp
->rx_opt
.tstamp_ok
&&
4171 th
->doff
== ((sizeof(*th
) + TCPOLEN_TSTAMP_ALIGNED
) / 4)) {
4172 if (tcp_parse_aligned_timestamp(tp
, th
))
4176 tcp_parse_options(net
, skb
, &tp
->rx_opt
, 1, NULL
);
4177 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
4178 tp
->rx_opt
.rcv_tsecr
-= tp
->tsoffset
;
4183 #ifdef CONFIG_TCP_MD5SIG
4185 * Parse MD5 Signature option
4187 const u8
*tcp_parse_md5sig_option(const struct tcphdr
*th
)
4189 int length
= (th
->doff
<< 2) - sizeof(*th
);
4190 const u8
*ptr
= (const u8
*)(th
+ 1);
4192 /* If not enough data remaining, we can short cut */
4193 while (length
>= TCPOLEN_MD5SIG
) {
4194 int opcode
= *ptr
++;
4205 if (opsize
< 2 || opsize
> length
)
4207 if (opcode
== TCPOPT_MD5SIG
)
4208 return opsize
== TCPOLEN_MD5SIG
? ptr
: NULL
;
4215 EXPORT_SYMBOL(tcp_parse_md5sig_option
);
4218 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
4220 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
4221 * it can pass through stack. So, the following predicate verifies that
4222 * this segment is not used for anything but congestion avoidance or
4223 * fast retransmit. Moreover, we even are able to eliminate most of such
4224 * second order effects, if we apply some small "replay" window (~RTO)
4225 * to timestamp space.
4227 * All these measures still do not guarantee that we reject wrapped ACKs
4228 * on networks with high bandwidth, when sequence space is recycled fastly,
4229 * but it guarantees that such events will be very rare and do not affect
4230 * connection seriously. This doesn't look nice, but alas, PAWS is really
4233 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
4234 * states that events when retransmit arrives after original data are rare.
4235 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
4236 * the biggest problem on large power networks even with minor reordering.
4237 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
4238 * up to bandwidth of 18Gigabit/sec. 8) ]
4241 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
4243 const struct tcp_sock
*tp
= tcp_sk(sk
);
4244 const struct tcphdr
*th
= tcp_hdr(skb
);
4245 u32 seq
= TCP_SKB_CB(skb
)->seq
;
4246 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
4248 return (/* 1. Pure ACK with correct sequence number. */
4249 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
4251 /* 2. ... and duplicate ACK. */
4252 ack
== tp
->snd_una
&&
4254 /* 3. ... and does not update window. */
4255 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
4257 /* 4. ... and sits in replay window. */
4258 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
4261 static inline bool tcp_paws_discard(const struct sock
*sk
,
4262 const struct sk_buff
*skb
)
4264 const struct tcp_sock
*tp
= tcp_sk(sk
);
4266 return !tcp_paws_check(&tp
->rx_opt
, TCP_PAWS_WINDOW
) &&
4267 !tcp_disordered_ack(sk
, skb
);
4270 /* Check segment sequence number for validity.
4272 * Segment controls are considered valid, if the segment
4273 * fits to the window after truncation to the window. Acceptability
4274 * of data (and SYN, FIN, of course) is checked separately.
4275 * See tcp_data_queue(), for example.
4277 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4278 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4279 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4280 * (borrowed from freebsd)
4283 static inline bool tcp_sequence(const struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
4285 return !before(end_seq
, tp
->rcv_wup
) &&
4286 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
4289 /* When we get a reset we do this. */
4290 void tcp_reset(struct sock
*sk
, struct sk_buff
*skb
)
4292 trace_tcp_receive_reset(sk
);
4294 /* mptcp can't tell us to ignore reset pkts,
4295 * so just ignore the return value of mptcp_incoming_options().
4297 if (sk_is_mptcp(sk
))
4298 mptcp_incoming_options(sk
, skb
);
4300 /* We want the right error as BSD sees it (and indeed as we do). */
4301 switch (sk
->sk_state
) {
4303 sk
->sk_err
= ECONNREFUSED
;
4305 case TCP_CLOSE_WAIT
:
4311 sk
->sk_err
= ECONNRESET
;
4313 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4316 tcp_write_queue_purge(sk
);
4319 if (!sock_flag(sk
, SOCK_DEAD
))
4320 sk_error_report(sk
);
4324 * Process the FIN bit. This now behaves as it is supposed to work
4325 * and the FIN takes effect when it is validly part of sequence
4326 * space. Not before when we get holes.
4328 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4329 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4332 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4333 * close and we go into CLOSING (and later onto TIME-WAIT)
4335 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4337 void tcp_fin(struct sock
*sk
)
4339 struct tcp_sock
*tp
= tcp_sk(sk
);
4341 inet_csk_schedule_ack(sk
);
4343 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
4344 sock_set_flag(sk
, SOCK_DONE
);
4346 switch (sk
->sk_state
) {
4348 case TCP_ESTABLISHED
:
4349 /* Move to CLOSE_WAIT */
4350 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
4351 inet_csk_enter_pingpong_mode(sk
);
4354 case TCP_CLOSE_WAIT
:
4356 /* Received a retransmission of the FIN, do
4361 /* RFC793: Remain in the LAST-ACK state. */
4365 /* This case occurs when a simultaneous close
4366 * happens, we must ack the received FIN and
4367 * enter the CLOSING state.
4370 tcp_set_state(sk
, TCP_CLOSING
);
4373 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4375 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
4378 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4379 * cases we should never reach this piece of code.
4381 pr_err("%s: Impossible, sk->sk_state=%d\n",
4382 __func__
, sk
->sk_state
);
4386 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4387 * Probably, we should reset in this case. For now drop them.
4389 skb_rbtree_purge(&tp
->out_of_order_queue
);
4390 if (tcp_is_sack(tp
))
4391 tcp_sack_reset(&tp
->rx_opt
);
4394 if (!sock_flag(sk
, SOCK_DEAD
)) {
4395 sk
->sk_state_change(sk
);
4397 /* Do not send POLL_HUP for half duplex close. */
4398 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
4399 sk
->sk_state
== TCP_CLOSE
)
4400 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_HUP
);
4402 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_IN
);
4406 static inline bool tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
,
4409 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
4410 if (before(seq
, sp
->start_seq
))
4411 sp
->start_seq
= seq
;
4412 if (after(end_seq
, sp
->end_seq
))
4413 sp
->end_seq
= end_seq
;
4419 static void tcp_dsack_set(struct sock
*sk
, u32 seq
, u32 end_seq
)
4421 struct tcp_sock
*tp
= tcp_sk(sk
);
4423 if (tcp_is_sack(tp
) && READ_ONCE(sock_net(sk
)->ipv4
.sysctl_tcp_dsack
)) {
4426 if (before(seq
, tp
->rcv_nxt
))
4427 mib_idx
= LINUX_MIB_TCPDSACKOLDSENT
;
4429 mib_idx
= LINUX_MIB_TCPDSACKOFOSENT
;
4431 NET_INC_STATS(sock_net(sk
), mib_idx
);
4433 tp
->rx_opt
.dsack
= 1;
4434 tp
->duplicate_sack
[0].start_seq
= seq
;
4435 tp
->duplicate_sack
[0].end_seq
= end_seq
;
4439 static void tcp_dsack_extend(struct sock
*sk
, u32 seq
, u32 end_seq
)
4441 struct tcp_sock
*tp
= tcp_sk(sk
);
4443 if (!tp
->rx_opt
.dsack
)
4444 tcp_dsack_set(sk
, seq
, end_seq
);
4446 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
4449 static void tcp_rcv_spurious_retrans(struct sock
*sk
, const struct sk_buff
*skb
)
4451 /* When the ACK path fails or drops most ACKs, the sender would
4452 * timeout and spuriously retransmit the same segment repeatedly.
4453 * The receiver remembers and reflects via DSACKs. Leverage the
4454 * DSACK state and change the txhash to re-route speculatively.
4456 if (TCP_SKB_CB(skb
)->seq
== tcp_sk(sk
)->duplicate_sack
[0].start_seq
&&
4457 sk_rethink_txhash(sk
))
4458 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPDUPLICATEDATAREHASH
);
4461 static void tcp_send_dupack(struct sock
*sk
, const struct sk_buff
*skb
)
4463 struct tcp_sock
*tp
= tcp_sk(sk
);
4465 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
4466 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4467 NET_INC_STATS(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4468 tcp_enter_quickack_mode(sk
, TCP_MAX_QUICKACKS
);
4470 if (tcp_is_sack(tp
) && READ_ONCE(sock_net(sk
)->ipv4
.sysctl_tcp_dsack
)) {
4471 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4473 tcp_rcv_spurious_retrans(sk
, skb
);
4474 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
4475 end_seq
= tp
->rcv_nxt
;
4476 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, end_seq
);
4483 /* These routines update the SACK block as out-of-order packets arrive or
4484 * in-order packets close up the sequence space.
4486 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
4489 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4490 struct tcp_sack_block
*swalk
= sp
+ 1;
4492 /* See if the recent change to the first SACK eats into
4493 * or hits the sequence space of other SACK blocks, if so coalesce.
4495 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
;) {
4496 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
4499 /* Zap SWALK, by moving every further SACK up by one slot.
4500 * Decrease num_sacks.
4502 tp
->rx_opt
.num_sacks
--;
4503 for (i
= this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
4512 static void tcp_sack_compress_send_ack(struct sock
*sk
)
4514 struct tcp_sock
*tp
= tcp_sk(sk
);
4516 if (!tp
->compressed_ack
)
4519 if (hrtimer_try_to_cancel(&tp
->compressed_ack_timer
) == 1)
4522 /* Since we have to send one ack finally,
4523 * substract one from tp->compressed_ack to keep
4524 * LINUX_MIB_TCPACKCOMPRESSED accurate.
4526 NET_ADD_STATS(sock_net(sk
), LINUX_MIB_TCPACKCOMPRESSED
,
4527 tp
->compressed_ack
- 1);
4529 tp
->compressed_ack
= 0;
4533 /* Reasonable amount of sack blocks included in TCP SACK option
4534 * The max is 4, but this becomes 3 if TCP timestamps are there.
4535 * Given that SACK packets might be lost, be conservative and use 2.
4537 #define TCP_SACK_BLOCKS_EXPECTED 2
4539 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
4541 struct tcp_sock
*tp
= tcp_sk(sk
);
4542 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4543 int cur_sacks
= tp
->rx_opt
.num_sacks
;
4549 for (this_sack
= 0; this_sack
< cur_sacks
; this_sack
++, sp
++) {
4550 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
4551 if (this_sack
>= TCP_SACK_BLOCKS_EXPECTED
)
4552 tcp_sack_compress_send_ack(sk
);
4553 /* Rotate this_sack to the first one. */
4554 for (; this_sack
> 0; this_sack
--, sp
--)
4555 swap(*sp
, *(sp
- 1));
4557 tcp_sack_maybe_coalesce(tp
);
4562 if (this_sack
>= TCP_SACK_BLOCKS_EXPECTED
)
4563 tcp_sack_compress_send_ack(sk
);
4565 /* Could not find an adjacent existing SACK, build a new one,
4566 * put it at the front, and shift everyone else down. We
4567 * always know there is at least one SACK present already here.
4569 * If the sack array is full, forget about the last one.
4571 if (this_sack
>= TCP_NUM_SACKS
) {
4573 tp
->rx_opt
.num_sacks
--;
4576 for (; this_sack
> 0; this_sack
--, sp
--)
4580 /* Build the new head SACK, and we're done. */
4581 sp
->start_seq
= seq
;
4582 sp
->end_seq
= end_seq
;
4583 tp
->rx_opt
.num_sacks
++;
4586 /* RCV.NXT advances, some SACKs should be eaten. */
4588 static void tcp_sack_remove(struct tcp_sock
*tp
)
4590 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4591 int num_sacks
= tp
->rx_opt
.num_sacks
;
4594 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4595 if (RB_EMPTY_ROOT(&tp
->out_of_order_queue
)) {
4596 tp
->rx_opt
.num_sacks
= 0;
4600 for (this_sack
= 0; this_sack
< num_sacks
;) {
4601 /* Check if the start of the sack is covered by RCV.NXT. */
4602 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
4605 /* RCV.NXT must cover all the block! */
4606 WARN_ON(before(tp
->rcv_nxt
, sp
->end_seq
));
4608 /* Zap this SACK, by moving forward any other SACKS. */
4609 for (i
= this_sack
+1; i
< num_sacks
; i
++)
4610 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
4617 tp
->rx_opt
.num_sacks
= num_sacks
;
4621 * tcp_try_coalesce - try to merge skb to prior one
4624 * @from: buffer to add in queue
4625 * @fragstolen: pointer to boolean
4627 * Before queueing skb @from after @to, try to merge them
4628 * to reduce overall memory use and queue lengths, if cost is small.
4629 * Packets in ofo or receive queues can stay a long time.
4630 * Better try to coalesce them right now to avoid future collapses.
4631 * Returns true if caller should free @from instead of queueing it
4633 static bool tcp_try_coalesce(struct sock
*sk
,
4635 struct sk_buff
*from
,
4640 *fragstolen
= false;
4642 /* Its possible this segment overlaps with prior segment in queue */
4643 if (TCP_SKB_CB(from
)->seq
!= TCP_SKB_CB(to
)->end_seq
)
4646 if (!mptcp_skb_can_collapse(to
, from
))
4649 #ifdef CONFIG_TLS_DEVICE
4650 if (from
->decrypted
!= to
->decrypted
)
4654 if (!skb_try_coalesce(to
, from
, fragstolen
, &delta
))
4657 atomic_add(delta
, &sk
->sk_rmem_alloc
);
4658 sk_mem_charge(sk
, delta
);
4659 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPRCVCOALESCE
);
4660 TCP_SKB_CB(to
)->end_seq
= TCP_SKB_CB(from
)->end_seq
;
4661 TCP_SKB_CB(to
)->ack_seq
= TCP_SKB_CB(from
)->ack_seq
;
4662 TCP_SKB_CB(to
)->tcp_flags
|= TCP_SKB_CB(from
)->tcp_flags
;
4664 if (TCP_SKB_CB(from
)->has_rxtstamp
) {
4665 TCP_SKB_CB(to
)->has_rxtstamp
= true;
4666 to
->tstamp
= from
->tstamp
;
4667 skb_hwtstamps(to
)->hwtstamp
= skb_hwtstamps(from
)->hwtstamp
;
4673 static bool tcp_ooo_try_coalesce(struct sock
*sk
,
4675 struct sk_buff
*from
,
4678 bool res
= tcp_try_coalesce(sk
, to
, from
, fragstolen
);
4680 /* In case tcp_drop() is called later, update to->gso_segs */
4682 u32 gso_segs
= max_t(u16
, 1, skb_shinfo(to
)->gso_segs
) +
4683 max_t(u16
, 1, skb_shinfo(from
)->gso_segs
);
4685 skb_shinfo(to
)->gso_segs
= min_t(u32
, gso_segs
, 0xFFFF);
4690 static void tcp_drop(struct sock
*sk
, struct sk_buff
*skb
)
4692 sk_drops_add(sk
, skb
);
4696 /* This one checks to see if we can put data from the
4697 * out_of_order queue into the receive_queue.
4699 static void tcp_ofo_queue(struct sock
*sk
)
4701 struct tcp_sock
*tp
= tcp_sk(sk
);
4702 __u32 dsack_high
= tp
->rcv_nxt
;
4703 bool fin
, fragstolen
, eaten
;
4704 struct sk_buff
*skb
, *tail
;
4707 p
= rb_first(&tp
->out_of_order_queue
);
4710 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
4713 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
4714 __u32 dsack
= dsack_high
;
4715 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
4716 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
4717 tcp_dsack_extend(sk
, TCP_SKB_CB(skb
)->seq
, dsack
);
4720 rb_erase(&skb
->rbnode
, &tp
->out_of_order_queue
);
4722 if (unlikely(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))) {
4727 tail
= skb_peek_tail(&sk
->sk_receive_queue
);
4728 eaten
= tail
&& tcp_try_coalesce(sk
, tail
, skb
, &fragstolen
);
4729 tcp_rcv_nxt_update(tp
, TCP_SKB_CB(skb
)->end_seq
);
4730 fin
= TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
;
4732 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4734 kfree_skb_partial(skb
, fragstolen
);
4736 if (unlikely(fin
)) {
4738 /* tcp_fin() purges tp->out_of_order_queue,
4739 * so we must end this loop right now.
4746 static bool tcp_prune_ofo_queue(struct sock
*sk
);
4747 static int tcp_prune_queue(struct sock
*sk
);
4749 static int tcp_try_rmem_schedule(struct sock
*sk
, struct sk_buff
*skb
,
4752 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
4753 !sk_rmem_schedule(sk
, skb
, size
)) {
4755 if (tcp_prune_queue(sk
) < 0)
4758 while (!sk_rmem_schedule(sk
, skb
, size
)) {
4759 if (!tcp_prune_ofo_queue(sk
))
4766 static void tcp_data_queue_ofo(struct sock
*sk
, struct sk_buff
*skb
)
4768 struct tcp_sock
*tp
= tcp_sk(sk
);
4769 struct rb_node
**p
, *parent
;
4770 struct sk_buff
*skb1
;
4774 tcp_ecn_check_ce(sk
, skb
);
4776 if (unlikely(tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))) {
4777 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPOFODROP
);
4778 sk
->sk_data_ready(sk
);
4783 /* Disable header prediction. */
4785 inet_csk_schedule_ack(sk
);
4787 tp
->rcv_ooopack
+= max_t(u16
, 1, skb_shinfo(skb
)->gso_segs
);
4788 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPOFOQUEUE
);
4789 seq
= TCP_SKB_CB(skb
)->seq
;
4790 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4792 p
= &tp
->out_of_order_queue
.rb_node
;
4793 if (RB_EMPTY_ROOT(&tp
->out_of_order_queue
)) {
4794 /* Initial out of order segment, build 1 SACK. */
4795 if (tcp_is_sack(tp
)) {
4796 tp
->rx_opt
.num_sacks
= 1;
4797 tp
->selective_acks
[0].start_seq
= seq
;
4798 tp
->selective_acks
[0].end_seq
= end_seq
;
4800 rb_link_node(&skb
->rbnode
, NULL
, p
);
4801 rb_insert_color(&skb
->rbnode
, &tp
->out_of_order_queue
);
4802 tp
->ooo_last_skb
= skb
;
4806 /* In the typical case, we are adding an skb to the end of the list.
4807 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4809 if (tcp_ooo_try_coalesce(sk
, tp
->ooo_last_skb
,
4810 skb
, &fragstolen
)) {
4812 /* For non sack flows, do not grow window to force DUPACK
4813 * and trigger fast retransmit.
4815 if (tcp_is_sack(tp
))
4816 tcp_grow_window(sk
, skb
, true);
4817 kfree_skb_partial(skb
, fragstolen
);
4821 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4822 if (!before(seq
, TCP_SKB_CB(tp
->ooo_last_skb
)->end_seq
)) {
4823 parent
= &tp
->ooo_last_skb
->rbnode
;
4824 p
= &parent
->rb_right
;
4828 /* Find place to insert this segment. Handle overlaps on the way. */
4832 skb1
= rb_to_skb(parent
);
4833 if (before(seq
, TCP_SKB_CB(skb1
)->seq
)) {
4834 p
= &parent
->rb_left
;
4837 if (before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4838 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4839 /* All the bits are present. Drop. */
4840 NET_INC_STATS(sock_net(sk
),
4841 LINUX_MIB_TCPOFOMERGE
);
4844 tcp_dsack_set(sk
, seq
, end_seq
);
4847 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
4848 /* Partial overlap. */
4849 tcp_dsack_set(sk
, seq
, TCP_SKB_CB(skb1
)->end_seq
);
4851 /* skb's seq == skb1's seq and skb covers skb1.
4852 * Replace skb1 with skb.
4854 rb_replace_node(&skb1
->rbnode
, &skb
->rbnode
,
4855 &tp
->out_of_order_queue
);
4856 tcp_dsack_extend(sk
,
4857 TCP_SKB_CB(skb1
)->seq
,
4858 TCP_SKB_CB(skb1
)->end_seq
);
4859 NET_INC_STATS(sock_net(sk
),
4860 LINUX_MIB_TCPOFOMERGE
);
4864 } else if (tcp_ooo_try_coalesce(sk
, skb1
,
4865 skb
, &fragstolen
)) {
4868 p
= &parent
->rb_right
;
4871 /* Insert segment into RB tree. */
4872 rb_link_node(&skb
->rbnode
, parent
, p
);
4873 rb_insert_color(&skb
->rbnode
, &tp
->out_of_order_queue
);
4876 /* Remove other segments covered by skb. */
4877 while ((skb1
= skb_rb_next(skb
)) != NULL
) {
4878 if (!after(end_seq
, TCP_SKB_CB(skb1
)->seq
))
4880 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4881 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4885 rb_erase(&skb1
->rbnode
, &tp
->out_of_order_queue
);
4886 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4887 TCP_SKB_CB(skb1
)->end_seq
);
4888 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPOFOMERGE
);
4891 /* If there is no skb after us, we are the last_skb ! */
4893 tp
->ooo_last_skb
= skb
;
4896 if (tcp_is_sack(tp
))
4897 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
4900 /* For non sack flows, do not grow window to force DUPACK
4901 * and trigger fast retransmit.
4903 if (tcp_is_sack(tp
))
4904 tcp_grow_window(sk
, skb
, false);
4906 skb_set_owner_r(skb
, sk
);
4910 static int __must_check
tcp_queue_rcv(struct sock
*sk
, struct sk_buff
*skb
,
4914 struct sk_buff
*tail
= skb_peek_tail(&sk
->sk_receive_queue
);
4917 tcp_try_coalesce(sk
, tail
,
4918 skb
, fragstolen
)) ? 1 : 0;
4919 tcp_rcv_nxt_update(tcp_sk(sk
), TCP_SKB_CB(skb
)->end_seq
);
4921 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4922 skb_set_owner_r(skb
, sk
);
4927 int tcp_send_rcvq(struct sock
*sk
, struct msghdr
*msg
, size_t size
)
4929 struct sk_buff
*skb
;
4937 if (size
> PAGE_SIZE
) {
4938 int npages
= min_t(size_t, size
>> PAGE_SHIFT
, MAX_SKB_FRAGS
);
4940 data_len
= npages
<< PAGE_SHIFT
;
4941 size
= data_len
+ (size
& ~PAGE_MASK
);
4943 skb
= alloc_skb_with_frags(size
- data_len
, data_len
,
4944 PAGE_ALLOC_COSTLY_ORDER
,
4945 &err
, sk
->sk_allocation
);
4949 skb_put(skb
, size
- data_len
);
4950 skb
->data_len
= data_len
;
4953 if (tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
)) {
4954 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPRCVQDROP
);
4958 err
= skb_copy_datagram_from_iter(skb
, 0, &msg
->msg_iter
, size
);
4962 TCP_SKB_CB(skb
)->seq
= tcp_sk(sk
)->rcv_nxt
;
4963 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ size
;
4964 TCP_SKB_CB(skb
)->ack_seq
= tcp_sk(sk
)->snd_una
- 1;
4966 if (tcp_queue_rcv(sk
, skb
, &fragstolen
)) {
4967 WARN_ON_ONCE(fragstolen
); /* should not happen */
4979 void tcp_data_ready(struct sock
*sk
)
4981 if (tcp_epollin_ready(sk
, sk
->sk_rcvlowat
) || sock_flag(sk
, SOCK_DONE
))
4982 sk
->sk_data_ready(sk
);
4985 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
4987 struct tcp_sock
*tp
= tcp_sk(sk
);
4991 /* If a subflow has been reset, the packet should not continue
4992 * to be processed, drop the packet.
4994 if (sk_is_mptcp(sk
) && !mptcp_incoming_options(sk
, skb
)) {
4999 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
) {
5004 __skb_pull(skb
, tcp_hdr(skb
)->doff
* 4);
5006 tp
->rx_opt
.dsack
= 0;
5008 /* Queue data for delivery to the user.
5009 * Packets in sequence go to the receive queue.
5010 * Out of sequence packets to the out_of_order_queue.
5012 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
5013 if (tcp_receive_window(tp
) == 0) {
5014 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPZEROWINDOWDROP
);
5018 /* Ok. In sequence. In window. */
5020 if (skb_queue_len(&sk
->sk_receive_queue
) == 0)
5021 sk_forced_mem_schedule(sk
, skb
->truesize
);
5022 else if (tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
)) {
5023 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPRCVQDROP
);
5024 sk
->sk_data_ready(sk
);
5028 eaten
= tcp_queue_rcv(sk
, skb
, &fragstolen
);
5030 tcp_event_data_recv(sk
, skb
);
5031 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
5034 if (!RB_EMPTY_ROOT(&tp
->out_of_order_queue
)) {
5037 /* RFC5681. 4.2. SHOULD send immediate ACK, when
5038 * gap in queue is filled.
5040 if (RB_EMPTY_ROOT(&tp
->out_of_order_queue
))
5041 inet_csk(sk
)->icsk_ack
.pending
|= ICSK_ACK_NOW
;
5044 if (tp
->rx_opt
.num_sacks
)
5045 tcp_sack_remove(tp
);
5047 tcp_fast_path_check(sk
);
5050 kfree_skb_partial(skb
, fragstolen
);
5051 if (!sock_flag(sk
, SOCK_DEAD
))
5056 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
5057 tcp_rcv_spurious_retrans(sk
, skb
);
5058 /* A retransmit, 2nd most common case. Force an immediate ack. */
5059 NET_INC_STATS(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
5060 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
5063 tcp_enter_quickack_mode(sk
, TCP_MAX_QUICKACKS
);
5064 inet_csk_schedule_ack(sk
);
5070 /* Out of window. F.e. zero window probe. */
5071 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
5074 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
5075 /* Partial packet, seq < rcv_next < end_seq */
5076 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
5078 /* If window is closed, drop tail of packet. But after
5079 * remembering D-SACK for its head made in previous line.
5081 if (!tcp_receive_window(tp
)) {
5082 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPZEROWINDOWDROP
);
5088 tcp_data_queue_ofo(sk
, skb
);
5091 static struct sk_buff
*tcp_skb_next(struct sk_buff
*skb
, struct sk_buff_head
*list
)
5094 return !skb_queue_is_last(list
, skb
) ? skb
->next
: NULL
;
5096 return skb_rb_next(skb
);
5099 static struct sk_buff
*tcp_collapse_one(struct sock
*sk
, struct sk_buff
*skb
,
5100 struct sk_buff_head
*list
,
5101 struct rb_root
*root
)
5103 struct sk_buff
*next
= tcp_skb_next(skb
, list
);
5106 __skb_unlink(skb
, list
);
5108 rb_erase(&skb
->rbnode
, root
);
5111 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPRCVCOLLAPSED
);
5116 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
5117 void tcp_rbtree_insert(struct rb_root
*root
, struct sk_buff
*skb
)
5119 struct rb_node
**p
= &root
->rb_node
;
5120 struct rb_node
*parent
= NULL
;
5121 struct sk_buff
*skb1
;
5125 skb1
= rb_to_skb(parent
);
5126 if (before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb1
)->seq
))
5127 p
= &parent
->rb_left
;
5129 p
= &parent
->rb_right
;
5131 rb_link_node(&skb
->rbnode
, parent
, p
);
5132 rb_insert_color(&skb
->rbnode
, root
);
5135 /* Collapse contiguous sequence of skbs head..tail with
5136 * sequence numbers start..end.
5138 * If tail is NULL, this means until the end of the queue.
5140 * Segments with FIN/SYN are not collapsed (only because this
5144 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
, struct rb_root
*root
,
5145 struct sk_buff
*head
, struct sk_buff
*tail
, u32 start
, u32 end
)
5147 struct sk_buff
*skb
= head
, *n
;
5148 struct sk_buff_head tmp
;
5151 /* First, check that queue is collapsible and find
5152 * the point where collapsing can be useful.
5155 for (end_of_skbs
= true; skb
!= NULL
&& skb
!= tail
; skb
= n
) {
5156 n
= tcp_skb_next(skb
, list
);
5158 /* No new bits? It is possible on ofo queue. */
5159 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
5160 skb
= tcp_collapse_one(sk
, skb
, list
, root
);
5166 /* The first skb to collapse is:
5168 * - bloated or contains data before "start" or
5169 * overlaps to the next one and mptcp allow collapsing.
5171 if (!(TCP_SKB_CB(skb
)->tcp_flags
& (TCPHDR_SYN
| TCPHDR_FIN
)) &&
5172 (tcp_win_from_space(sk
, skb
->truesize
) > skb
->len
||
5173 before(TCP_SKB_CB(skb
)->seq
, start
))) {
5174 end_of_skbs
= false;
5178 if (n
&& n
!= tail
&& mptcp_skb_can_collapse(skb
, n
) &&
5179 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(n
)->seq
) {
5180 end_of_skbs
= false;
5184 /* Decided to skip this, advance start seq. */
5185 start
= TCP_SKB_CB(skb
)->end_seq
;
5188 (TCP_SKB_CB(skb
)->tcp_flags
& (TCPHDR_SYN
| TCPHDR_FIN
)))
5191 __skb_queue_head_init(&tmp
);
5193 while (before(start
, end
)) {
5194 int copy
= min_t(int, SKB_MAX_ORDER(0, 0), end
- start
);
5195 struct sk_buff
*nskb
;
5197 nskb
= alloc_skb(copy
, GFP_ATOMIC
);
5201 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
5202 #ifdef CONFIG_TLS_DEVICE
5203 nskb
->decrypted
= skb
->decrypted
;
5205 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
5207 __skb_queue_before(list
, skb
, nskb
);
5209 __skb_queue_tail(&tmp
, nskb
); /* defer rbtree insertion */
5210 skb_set_owner_r(nskb
, sk
);
5211 mptcp_skb_ext_move(nskb
, skb
);
5213 /* Copy data, releasing collapsed skbs. */
5215 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
5216 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
5220 size
= min(copy
, size
);
5221 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
5223 TCP_SKB_CB(nskb
)->end_seq
+= size
;
5227 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
5228 skb
= tcp_collapse_one(sk
, skb
, list
, root
);
5231 !mptcp_skb_can_collapse(nskb
, skb
) ||
5232 (TCP_SKB_CB(skb
)->tcp_flags
& (TCPHDR_SYN
| TCPHDR_FIN
)))
5234 #ifdef CONFIG_TLS_DEVICE
5235 if (skb
->decrypted
!= nskb
->decrypted
)
5242 skb_queue_walk_safe(&tmp
, skb
, n
)
5243 tcp_rbtree_insert(root
, skb
);
5246 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
5247 * and tcp_collapse() them until all the queue is collapsed.
5249 static void tcp_collapse_ofo_queue(struct sock
*sk
)
5251 struct tcp_sock
*tp
= tcp_sk(sk
);
5252 u32 range_truesize
, sum_tiny
= 0;
5253 struct sk_buff
*skb
, *head
;
5256 skb
= skb_rb_first(&tp
->out_of_order_queue
);
5259 tp
->ooo_last_skb
= skb_rb_last(&tp
->out_of_order_queue
);
5262 start
= TCP_SKB_CB(skb
)->seq
;
5263 end
= TCP_SKB_CB(skb
)->end_seq
;
5264 range_truesize
= skb
->truesize
;
5266 for (head
= skb
;;) {
5267 skb
= skb_rb_next(skb
);
5269 /* Range is terminated when we see a gap or when
5270 * we are at the queue end.
5273 after(TCP_SKB_CB(skb
)->seq
, end
) ||
5274 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
5275 /* Do not attempt collapsing tiny skbs */
5276 if (range_truesize
!= head
->truesize
||
5277 end
- start
>= SKB_WITH_OVERHEAD(SK_MEM_QUANTUM
)) {
5278 tcp_collapse(sk
, NULL
, &tp
->out_of_order_queue
,
5279 head
, skb
, start
, end
);
5281 sum_tiny
+= range_truesize
;
5282 if (sum_tiny
> sk
->sk_rcvbuf
>> 3)
5288 range_truesize
+= skb
->truesize
;
5289 if (unlikely(before(TCP_SKB_CB(skb
)->seq
, start
)))
5290 start
= TCP_SKB_CB(skb
)->seq
;
5291 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
5292 end
= TCP_SKB_CB(skb
)->end_seq
;
5297 * Clean the out-of-order queue to make room.
5298 * We drop high sequences packets to :
5299 * 1) Let a chance for holes to be filled.
5300 * 2) not add too big latencies if thousands of packets sit there.
5301 * (But if application shrinks SO_RCVBUF, we could still end up
5302 * freeing whole queue here)
5303 * 3) Drop at least 12.5 % of sk_rcvbuf to avoid malicious attacks.
5305 * Return true if queue has shrunk.
5307 static bool tcp_prune_ofo_queue(struct sock
*sk
)
5309 struct tcp_sock
*tp
= tcp_sk(sk
);
5310 struct rb_node
*node
, *prev
;
5313 if (RB_EMPTY_ROOT(&tp
->out_of_order_queue
))
5316 NET_INC_STATS(sock_net(sk
), LINUX_MIB_OFOPRUNED
);
5317 goal
= sk
->sk_rcvbuf
>> 3;
5318 node
= &tp
->ooo_last_skb
->rbnode
;
5320 prev
= rb_prev(node
);
5321 rb_erase(node
, &tp
->out_of_order_queue
);
5322 goal
-= rb_to_skb(node
)->truesize
;
5323 tcp_drop(sk
, rb_to_skb(node
));
5324 if (!prev
|| goal
<= 0) {
5326 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
&&
5327 !tcp_under_memory_pressure(sk
))
5329 goal
= sk
->sk_rcvbuf
>> 3;
5333 tp
->ooo_last_skb
= rb_to_skb(prev
);
5335 /* Reset SACK state. A conforming SACK implementation will
5336 * do the same at a timeout based retransmit. When a connection
5337 * is in a sad state like this, we care only about integrity
5338 * of the connection not performance.
5340 if (tp
->rx_opt
.sack_ok
)
5341 tcp_sack_reset(&tp
->rx_opt
);
5345 /* Reduce allocated memory if we can, trying to get
5346 * the socket within its memory limits again.
5348 * Return less than zero if we should start dropping frames
5349 * until the socket owning process reads some of the data
5350 * to stabilize the situation.
5352 static int tcp_prune_queue(struct sock
*sk
)
5354 struct tcp_sock
*tp
= tcp_sk(sk
);
5356 NET_INC_STATS(sock_net(sk
), LINUX_MIB_PRUNECALLED
);
5358 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
5359 tcp_clamp_window(sk
);
5360 else if (tcp_under_memory_pressure(sk
))
5361 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
5363 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
5366 tcp_collapse_ofo_queue(sk
);
5367 if (!skb_queue_empty(&sk
->sk_receive_queue
))
5368 tcp_collapse(sk
, &sk
->sk_receive_queue
, NULL
,
5369 skb_peek(&sk
->sk_receive_queue
),
5371 tp
->copied_seq
, tp
->rcv_nxt
);
5374 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
5377 /* Collapsing did not help, destructive actions follow.
5378 * This must not ever occur. */
5380 tcp_prune_ofo_queue(sk
);
5382 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
5385 /* If we are really being abused, tell the caller to silently
5386 * drop receive data on the floor. It will get retransmitted
5387 * and hopefully then we'll have sufficient space.
5389 NET_INC_STATS(sock_net(sk
), LINUX_MIB_RCVPRUNED
);
5391 /* Massive buffer overcommit. */
5396 static bool tcp_should_expand_sndbuf(const struct sock
*sk
)
5398 const struct tcp_sock
*tp
= tcp_sk(sk
);
5400 /* If the user specified a specific send buffer setting, do
5403 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
5406 /* If we are under global TCP memory pressure, do not expand. */
5407 if (tcp_under_memory_pressure(sk
))
5410 /* If we are under soft global TCP memory pressure, do not expand. */
5411 if (sk_memory_allocated(sk
) >= sk_prot_mem_limits(sk
, 0))
5414 /* If we filled the congestion window, do not expand. */
5415 if (tcp_packets_in_flight(tp
) >= tcp_snd_cwnd(tp
))
5421 static void tcp_new_space(struct sock
*sk
)
5423 struct tcp_sock
*tp
= tcp_sk(sk
);
5425 if (tcp_should_expand_sndbuf(sk
)) {
5426 tcp_sndbuf_expand(sk
);
5427 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
5430 INDIRECT_CALL_1(sk
->sk_write_space
, sk_stream_write_space
, sk
);
5433 /* Caller made space either from:
5434 * 1) Freeing skbs in rtx queues (after tp->snd_una has advanced)
5435 * 2) Sent skbs from output queue (and thus advancing tp->snd_nxt)
5437 * We might be able to generate EPOLLOUT to the application if:
5438 * 1) Space consumed in output/rtx queues is below sk->sk_sndbuf/2
5439 * 2) notsent amount (tp->write_seq - tp->snd_nxt) became
5440 * small enough that tcp_stream_memory_free() decides it
5441 * is time to generate EPOLLOUT.
5443 void tcp_check_space(struct sock
*sk
)
5445 /* pairs with tcp_poll() */
5447 if (sk
->sk_socket
&&
5448 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
5450 if (!test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
5451 tcp_chrono_stop(sk
, TCP_CHRONO_SNDBUF_LIMITED
);
5455 static inline void tcp_data_snd_check(struct sock
*sk
)
5457 tcp_push_pending_frames(sk
);
5458 tcp_check_space(sk
);
5462 * Check if sending an ack is needed.
5464 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
5466 struct tcp_sock
*tp
= tcp_sk(sk
);
5467 unsigned long rtt
, delay
;
5469 /* More than one full frame received... */
5470 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
&&
5471 /* ... and right edge of window advances far enough.
5472 * (tcp_recvmsg() will send ACK otherwise).
5473 * If application uses SO_RCVLOWAT, we want send ack now if
5474 * we have not received enough bytes to satisfy the condition.
5476 (tp
->rcv_nxt
- tp
->copied_seq
< sk
->sk_rcvlowat
||
5477 __tcp_select_window(sk
) >= tp
->rcv_wnd
)) ||
5478 /* We ACK each frame or... */
5479 tcp_in_quickack_mode(sk
) ||
5480 /* Protocol state mandates a one-time immediate ACK */
5481 inet_csk(sk
)->icsk_ack
.pending
& ICSK_ACK_NOW
) {
5487 if (!ofo_possible
|| RB_EMPTY_ROOT(&tp
->out_of_order_queue
)) {
5488 tcp_send_delayed_ack(sk
);
5492 if (!tcp_is_sack(tp
) ||
5493 tp
->compressed_ack
>= sock_net(sk
)->ipv4
.sysctl_tcp_comp_sack_nr
)
5496 if (tp
->compressed_ack_rcv_nxt
!= tp
->rcv_nxt
) {
5497 tp
->compressed_ack_rcv_nxt
= tp
->rcv_nxt
;
5498 tp
->dup_ack_counter
= 0;
5500 if (tp
->dup_ack_counter
< TCP_FASTRETRANS_THRESH
) {
5501 tp
->dup_ack_counter
++;
5504 tp
->compressed_ack
++;
5505 if (hrtimer_is_queued(&tp
->compressed_ack_timer
))
5508 /* compress ack timer : 5 % of rtt, but no more than tcp_comp_sack_delay_ns */
5510 rtt
= tp
->rcv_rtt_est
.rtt_us
;
5511 if (tp
->srtt_us
&& tp
->srtt_us
< rtt
)
5514 delay
= min_t(unsigned long, sock_net(sk
)->ipv4
.sysctl_tcp_comp_sack_delay_ns
,
5515 rtt
* (NSEC_PER_USEC
>> 3)/20);
5517 hrtimer_start_range_ns(&tp
->compressed_ack_timer
, ns_to_ktime(delay
),
5518 sock_net(sk
)->ipv4
.sysctl_tcp_comp_sack_slack_ns
,
5519 HRTIMER_MODE_REL_PINNED_SOFT
);
5522 static inline void tcp_ack_snd_check(struct sock
*sk
)
5524 if (!inet_csk_ack_scheduled(sk
)) {
5525 /* We sent a data segment already. */
5528 __tcp_ack_snd_check(sk
, 1);
5532 * This routine is only called when we have urgent data
5533 * signaled. Its the 'slow' part of tcp_urg. It could be
5534 * moved inline now as tcp_urg is only called from one
5535 * place. We handle URGent data wrong. We have to - as
5536 * BSD still doesn't use the correction from RFC961.
5537 * For 1003.1g we should support a new option TCP_STDURG to permit
5538 * either form (or just set the sysctl tcp_stdurg).
5541 static void tcp_check_urg(struct sock
*sk
, const struct tcphdr
*th
)
5543 struct tcp_sock
*tp
= tcp_sk(sk
);
5544 u32 ptr
= ntohs(th
->urg_ptr
);
5546 if (ptr
&& !READ_ONCE(sock_net(sk
)->ipv4
.sysctl_tcp_stdurg
))
5548 ptr
+= ntohl(th
->seq
);
5550 /* Ignore urgent data that we've already seen and read. */
5551 if (after(tp
->copied_seq
, ptr
))
5554 /* Do not replay urg ptr.
5556 * NOTE: interesting situation not covered by specs.
5557 * Misbehaving sender may send urg ptr, pointing to segment,
5558 * which we already have in ofo queue. We are not able to fetch
5559 * such data and will stay in TCP_URG_NOTYET until will be eaten
5560 * by recvmsg(). Seems, we are not obliged to handle such wicked
5561 * situations. But it is worth to think about possibility of some
5562 * DoSes using some hypothetical application level deadlock.
5564 if (before(ptr
, tp
->rcv_nxt
))
5567 /* Do we already have a newer (or duplicate) urgent pointer? */
5568 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
5571 /* Tell the world about our new urgent pointer. */
5574 /* We may be adding urgent data when the last byte read was
5575 * urgent. To do this requires some care. We cannot just ignore
5576 * tp->copied_seq since we would read the last urgent byte again
5577 * as data, nor can we alter copied_seq until this data arrives
5578 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5580 * NOTE. Double Dutch. Rendering to plain English: author of comment
5581 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5582 * and expect that both A and B disappear from stream. This is _wrong_.
5583 * Though this happens in BSD with high probability, this is occasional.
5584 * Any application relying on this is buggy. Note also, that fix "works"
5585 * only in this artificial test. Insert some normal data between A and B and we will
5586 * decline of BSD again. Verdict: it is better to remove to trap
5589 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
5590 !sock_flag(sk
, SOCK_URGINLINE
) && tp
->copied_seq
!= tp
->rcv_nxt
) {
5591 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
5593 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5594 __skb_unlink(skb
, &sk
->sk_receive_queue
);
5599 tp
->urg_data
= TCP_URG_NOTYET
;
5600 WRITE_ONCE(tp
->urg_seq
, ptr
);
5602 /* Disable header prediction. */
5606 /* This is the 'fast' part of urgent handling. */
5607 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, const struct tcphdr
*th
)
5609 struct tcp_sock
*tp
= tcp_sk(sk
);
5611 /* Check if we get a new urgent pointer - normally not. */
5613 tcp_check_urg(sk
, th
);
5615 /* Do we wait for any urgent data? - normally not... */
5616 if (tp
->urg_data
== TCP_URG_NOTYET
) {
5617 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
5620 /* Is the urgent pointer pointing into this packet? */
5621 if (ptr
< skb
->len
) {
5623 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
5625 tp
->urg_data
= TCP_URG_VALID
| tmp
;
5626 if (!sock_flag(sk
, SOCK_DEAD
))
5627 sk
->sk_data_ready(sk
);
5632 /* Accept RST for rcv_nxt - 1 after a FIN.
5633 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a
5634 * FIN is sent followed by a RST packet. The RST is sent with the same
5635 * sequence number as the FIN, and thus according to RFC 5961 a challenge
5636 * ACK should be sent. However, Mac OSX rate limits replies to challenge
5637 * ACKs on the closed socket. In addition middleboxes can drop either the
5638 * challenge ACK or a subsequent RST.
5640 static bool tcp_reset_check(const struct sock
*sk
, const struct sk_buff
*skb
)
5642 struct tcp_sock
*tp
= tcp_sk(sk
);
5644 return unlikely(TCP_SKB_CB(skb
)->seq
== (tp
->rcv_nxt
- 1) &&
5645 (1 << sk
->sk_state
) & (TCPF_CLOSE_WAIT
| TCPF_LAST_ACK
|
5649 /* Does PAWS and seqno based validation of an incoming segment, flags will
5650 * play significant role here.
5652 static bool tcp_validate_incoming(struct sock
*sk
, struct sk_buff
*skb
,
5653 const struct tcphdr
*th
, int syn_inerr
)
5655 struct tcp_sock
*tp
= tcp_sk(sk
);
5656 bool rst_seq_match
= false;
5658 /* RFC1323: H1. Apply PAWS check first. */
5659 if (tcp_fast_parse_options(sock_net(sk
), skb
, th
, tp
) &&
5660 tp
->rx_opt
.saw_tstamp
&&
5661 tcp_paws_discard(sk
, skb
)) {
5663 NET_INC_STATS(sock_net(sk
), LINUX_MIB_PAWSESTABREJECTED
);
5664 if (!tcp_oow_rate_limited(sock_net(sk
), skb
,
5665 LINUX_MIB_TCPACKSKIPPEDPAWS
,
5666 &tp
->last_oow_ack_time
))
5667 tcp_send_dupack(sk
, skb
);
5670 /* Reset is accepted even if it did not pass PAWS. */
5673 /* Step 1: check sequence number */
5674 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5675 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5676 * (RST) segments are validated by checking their SEQ-fields."
5677 * And page 69: "If an incoming segment is not acceptable,
5678 * an acknowledgment should be sent in reply (unless the RST
5679 * bit is set, if so drop the segment and return)".
5684 if (!tcp_oow_rate_limited(sock_net(sk
), skb
,
5685 LINUX_MIB_TCPACKSKIPPEDSEQ
,
5686 &tp
->last_oow_ack_time
))
5687 tcp_send_dupack(sk
, skb
);
5688 } else if (tcp_reset_check(sk
, skb
)) {
5694 /* Step 2: check RST bit */
5696 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a
5697 * FIN and SACK too if available):
5698 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or
5699 * the right-most SACK block,
5701 * RESET the connection
5703 * Send a challenge ACK
5705 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
||
5706 tcp_reset_check(sk
, skb
)) {
5707 rst_seq_match
= true;
5708 } else if (tcp_is_sack(tp
) && tp
->rx_opt
.num_sacks
> 0) {
5709 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
5710 int max_sack
= sp
[0].end_seq
;
5713 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
;
5715 max_sack
= after(sp
[this_sack
].end_seq
,
5717 sp
[this_sack
].end_seq
: max_sack
;
5720 if (TCP_SKB_CB(skb
)->seq
== max_sack
)
5721 rst_seq_match
= true;
5727 /* Disable TFO if RST is out-of-order
5728 * and no data has been received
5729 * for current active TFO socket
5731 if (tp
->syn_fastopen
&& !tp
->data_segs_in
&&
5732 sk
->sk_state
== TCP_ESTABLISHED
)
5733 tcp_fastopen_active_disable(sk
);
5734 tcp_send_challenge_ack(sk
, skb
);
5739 /* step 3: check security and precedence [ignored] */
5741 /* step 4: Check for a SYN
5742 * RFC 5961 4.2 : Send a challenge ack
5747 TCP_INC_STATS(sock_net(sk
), TCP_MIB_INERRS
);
5748 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPSYNCHALLENGE
);
5749 tcp_send_challenge_ack(sk
, skb
);
5753 bpf_skops_parse_hdr(sk
, skb
);
5763 * TCP receive function for the ESTABLISHED state.
5765 * It is split into a fast path and a slow path. The fast path is
5767 * - A zero window was announced from us - zero window probing
5768 * is only handled properly in the slow path.
5769 * - Out of order segments arrived.
5770 * - Urgent data is expected.
5771 * - There is no buffer space left
5772 * - Unexpected TCP flags/window values/header lengths are received
5773 * (detected by checking the TCP header against pred_flags)
5774 * - Data is sent in both directions. Fast path only supports pure senders
5775 * or pure receivers (this means either the sequence number or the ack
5776 * value must stay constant)
5777 * - Unexpected TCP option.
5779 * When these conditions are not satisfied it drops into a standard
5780 * receive procedure patterned after RFC793 to handle all cases.
5781 * The first three cases are guaranteed by proper pred_flags setting,
5782 * the rest is checked inline. Fast processing is turned on in
5783 * tcp_data_queue when everything is OK.
5785 void tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
)
5787 const struct tcphdr
*th
= (const struct tcphdr
*)skb
->data
;
5788 struct tcp_sock
*tp
= tcp_sk(sk
);
5789 unsigned int len
= skb
->len
;
5791 /* TCP congestion window tracking */
5792 trace_tcp_probe(sk
, skb
);
5794 tcp_mstamp_refresh(tp
);
5795 if (unlikely(!rcu_access_pointer(sk
->sk_rx_dst
)))
5796 inet_csk(sk
)->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
5798 * Header prediction.
5799 * The code loosely follows the one in the famous
5800 * "30 instruction TCP receive" Van Jacobson mail.
5802 * Van's trick is to deposit buffers into socket queue
5803 * on a device interrupt, to call tcp_recv function
5804 * on the receive process context and checksum and copy
5805 * the buffer to user space. smart...
5807 * Our current scheme is not silly either but we take the
5808 * extra cost of the net_bh soft interrupt processing...
5809 * We do checksum and copy also but from device to kernel.
5812 tp
->rx_opt
.saw_tstamp
= 0;
5814 /* pred_flags is 0xS?10 << 16 + snd_wnd
5815 * if header_prediction is to be made
5816 * 'S' will always be tp->tcp_header_len >> 2
5817 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5818 * turn it off (when there are holes in the receive
5819 * space for instance)
5820 * PSH flag is ignored.
5823 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
5824 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
&&
5825 !after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
)) {
5826 int tcp_header_len
= tp
->tcp_header_len
;
5828 /* Timestamp header prediction: tcp_header_len
5829 * is automatically equal to th->doff*4 due to pred_flags
5833 /* Check timestamp */
5834 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
5835 /* No? Slow path! */
5836 if (!tcp_parse_aligned_timestamp(tp
, th
))
5839 /* If PAWS failed, check it more carefully in slow path */
5840 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
5843 /* DO NOT update ts_recent here, if checksum fails
5844 * and timestamp was corrupted part, it will result
5845 * in a hung connection since we will drop all
5846 * future packets due to the PAWS test.
5850 if (len
<= tcp_header_len
) {
5851 /* Bulk data transfer: sender */
5852 if (len
== tcp_header_len
) {
5853 /* Predicted packet is in window by definition.
5854 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5855 * Hence, check seq<=rcv_wup reduces to:
5857 if (tcp_header_len
==
5858 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5859 tp
->rcv_nxt
== tp
->rcv_wup
)
5860 tcp_store_ts_recent(tp
);
5862 /* We know that such packets are checksummed
5865 tcp_ack(sk
, skb
, 0);
5867 tcp_data_snd_check(sk
);
5868 /* When receiving pure ack in fast path, update
5869 * last ts ecr directly instead of calling
5870 * tcp_rcv_rtt_measure_ts()
5872 tp
->rcv_rtt_last_tsecr
= tp
->rx_opt
.rcv_tsecr
;
5874 } else { /* Header too small */
5875 TCP_INC_STATS(sock_net(sk
), TCP_MIB_INERRS
);
5880 bool fragstolen
= false;
5882 if (tcp_checksum_complete(skb
))
5885 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
5888 /* Predicted packet is in window by definition.
5889 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5890 * Hence, check seq<=rcv_wup reduces to:
5892 if (tcp_header_len
==
5893 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5894 tp
->rcv_nxt
== tp
->rcv_wup
)
5895 tcp_store_ts_recent(tp
);
5897 tcp_rcv_rtt_measure_ts(sk
, skb
);
5899 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPHPHITS
);
5901 /* Bulk data transfer: receiver */
5902 __skb_pull(skb
, tcp_header_len
);
5903 eaten
= tcp_queue_rcv(sk
, skb
, &fragstolen
);
5905 tcp_event_data_recv(sk
, skb
);
5907 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
5908 /* Well, only one small jumplet in fast path... */
5909 tcp_ack(sk
, skb
, FLAG_DATA
);
5910 tcp_data_snd_check(sk
);
5911 if (!inet_csk_ack_scheduled(sk
))
5914 tcp_update_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5917 __tcp_ack_snd_check(sk
, 0);
5920 kfree_skb_partial(skb
, fragstolen
);
5927 if (len
< (th
->doff
<< 2) || tcp_checksum_complete(skb
))
5930 if (!th
->ack
&& !th
->rst
&& !th
->syn
)
5934 * Standard slow path.
5937 if (!tcp_validate_incoming(sk
, skb
, th
, 1))
5941 if (tcp_ack(sk
, skb
, FLAG_SLOWPATH
| FLAG_UPDATE_TS_RECENT
) < 0)
5944 tcp_rcv_rtt_measure_ts(sk
, skb
);
5946 /* Process urgent data. */
5947 tcp_urg(sk
, skb
, th
);
5949 /* step 7: process the segment text */
5950 tcp_data_queue(sk
, skb
);
5952 tcp_data_snd_check(sk
);
5953 tcp_ack_snd_check(sk
);
5957 trace_tcp_bad_csum(skb
);
5958 TCP_INC_STATS(sock_net(sk
), TCP_MIB_CSUMERRORS
);
5959 TCP_INC_STATS(sock_net(sk
), TCP_MIB_INERRS
);
5964 EXPORT_SYMBOL(tcp_rcv_established
);
5966 void tcp_init_transfer(struct sock
*sk
, int bpf_op
, struct sk_buff
*skb
)
5968 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5969 struct tcp_sock
*tp
= tcp_sk(sk
);
5972 icsk
->icsk_af_ops
->rebuild_header(sk
);
5973 tcp_init_metrics(sk
);
5975 /* Initialize the congestion window to start the transfer.
5976 * Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
5977 * retransmitted. In light of RFC6298 more aggressive 1sec
5978 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
5979 * retransmission has occurred.
5981 if (tp
->total_retrans
> 1 && tp
->undo_marker
)
5982 tcp_snd_cwnd_set(tp
, 1);
5984 tcp_snd_cwnd_set(tp
, tcp_init_cwnd(tp
, __sk_dst_get(sk
)));
5985 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
5987 bpf_skops_established(sk
, bpf_op
, skb
);
5988 /* Initialize congestion control unless BPF initialized it already: */
5989 if (!icsk
->icsk_ca_initialized
)
5990 tcp_init_congestion_control(sk
);
5991 tcp_init_buffer_space(sk
);
5994 void tcp_finish_connect(struct sock
*sk
, struct sk_buff
*skb
)
5996 struct tcp_sock
*tp
= tcp_sk(sk
);
5997 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5999 tcp_set_state(sk
, TCP_ESTABLISHED
);
6000 icsk
->icsk_ack
.lrcvtime
= tcp_jiffies32
;
6003 icsk
->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
6004 security_inet_conn_established(sk
, skb
);
6005 sk_mark_napi_id(sk
, skb
);
6008 tcp_init_transfer(sk
, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB
, skb
);
6010 /* Prevent spurious tcp_cwnd_restart() on first data
6013 tp
->lsndtime
= tcp_jiffies32
;
6015 if (sock_flag(sk
, SOCK_KEEPOPEN
))
6016 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
6018 if (!tp
->rx_opt
.snd_wscale
)
6019 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
6024 static bool tcp_rcv_fastopen_synack(struct sock
*sk
, struct sk_buff
*synack
,
6025 struct tcp_fastopen_cookie
*cookie
)
6027 struct tcp_sock
*tp
= tcp_sk(sk
);
6028 struct sk_buff
*data
= tp
->syn_data
? tcp_rtx_queue_head(sk
) : NULL
;
6029 u16 mss
= tp
->rx_opt
.mss_clamp
, try_exp
= 0;
6030 bool syn_drop
= false;
6032 if (mss
== tp
->rx_opt
.user_mss
) {
6033 struct tcp_options_received opt
;
6035 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
6036 tcp_clear_options(&opt
);
6037 opt
.user_mss
= opt
.mss_clamp
= 0;
6038 tcp_parse_options(sock_net(sk
), synack
, &opt
, 0, NULL
);
6039 mss
= opt
.mss_clamp
;
6042 if (!tp
->syn_fastopen
) {
6043 /* Ignore an unsolicited cookie */
6045 } else if (tp
->total_retrans
) {
6046 /* SYN timed out and the SYN-ACK neither has a cookie nor
6047 * acknowledges data. Presumably the remote received only
6048 * the retransmitted (regular) SYNs: either the original
6049 * SYN-data or the corresponding SYN-ACK was dropped.
6051 syn_drop
= (cookie
->len
< 0 && data
);
6052 } else if (cookie
->len
< 0 && !tp
->syn_data
) {
6053 /* We requested a cookie but didn't get it. If we did not use
6054 * the (old) exp opt format then try so next time (try_exp=1).
6055 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
6057 try_exp
= tp
->syn_fastopen_exp
? 2 : 1;
6060 tcp_fastopen_cache_set(sk
, mss
, cookie
, syn_drop
, try_exp
);
6062 if (data
) { /* Retransmit unacked data in SYN */
6063 if (tp
->total_retrans
)
6064 tp
->fastopen_client_fail
= TFO_SYN_RETRANSMITTED
;
6066 tp
->fastopen_client_fail
= TFO_DATA_NOT_ACKED
;
6067 skb_rbtree_walk_from(data
)
6068 tcp_mark_skb_lost(sk
, data
);
6069 tcp_xmit_retransmit_queue(sk
);
6070 NET_INC_STATS(sock_net(sk
),
6071 LINUX_MIB_TCPFASTOPENACTIVEFAIL
);
6074 tp
->syn_data_acked
= tp
->syn_data
;
6075 if (tp
->syn_data_acked
) {
6076 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPFASTOPENACTIVE
);
6077 /* SYN-data is counted as two separate packets in tcp_ack() */
6078 if (tp
->delivered
> 1)
6082 tcp_fastopen_add_skb(sk
, synack
);
6087 static void smc_check_reset_syn(struct tcp_sock
*tp
)
6089 #if IS_ENABLED(CONFIG_SMC)
6090 if (static_branch_unlikely(&tcp_have_smc
)) {
6091 if (tp
->syn_smc
&& !tp
->rx_opt
.smc_ok
)
6097 static void tcp_try_undo_spurious_syn(struct sock
*sk
)
6099 struct tcp_sock
*tp
= tcp_sk(sk
);
6102 /* undo_marker is set when SYN or SYNACK times out. The timeout is
6103 * spurious if the ACK's timestamp option echo value matches the
6104 * original SYN timestamp.
6106 syn_stamp
= tp
->retrans_stamp
;
6107 if (tp
->undo_marker
&& syn_stamp
&& tp
->rx_opt
.saw_tstamp
&&
6108 syn_stamp
== tp
->rx_opt
.rcv_tsecr
)
6109 tp
->undo_marker
= 0;
6112 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
6113 const struct tcphdr
*th
)
6115 struct inet_connection_sock
*icsk
= inet_csk(sk
);
6116 struct tcp_sock
*tp
= tcp_sk(sk
);
6117 struct tcp_fastopen_cookie foc
= { .len
= -1 };
6118 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
6121 tcp_parse_options(sock_net(sk
), skb
, &tp
->rx_opt
, 0, &foc
);
6122 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
6123 tp
->rx_opt
.rcv_tsecr
-= tp
->tsoffset
;
6127 * "If the state is SYN-SENT then
6128 * first check the ACK bit
6129 * If the ACK bit is set
6130 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
6131 * a reset (unless the RST bit is set, if so drop
6132 * the segment and return)"
6134 if (!after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_una
) ||
6135 after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
)) {
6136 /* Previous FIN/ACK or RST/ACK might be ignored. */
6137 if (icsk
->icsk_retransmits
== 0)
6138 inet_csk_reset_xmit_timer(sk
,
6140 TCP_TIMEOUT_MIN
, TCP_RTO_MAX
);
6141 goto reset_and_undo
;
6144 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
6145 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
6146 tcp_time_stamp(tp
))) {
6147 NET_INC_STATS(sock_net(sk
),
6148 LINUX_MIB_PAWSACTIVEREJECTED
);
6149 goto reset_and_undo
;
6152 /* Now ACK is acceptable.
6154 * "If the RST bit is set
6155 * If the ACK was acceptable then signal the user "error:
6156 * connection reset", drop the segment, enter CLOSED state,
6157 * delete TCB, and return."
6166 * "fifth, if neither of the SYN or RST bits is set then
6167 * drop the segment and return."
6173 goto discard_and_undo
;
6176 * "If the SYN bit is on ...
6177 * are acceptable then ...
6178 * (our SYN has been ACKed), change the connection
6179 * state to ESTABLISHED..."
6182 tcp_ecn_rcv_synack(tp
, th
);
6184 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
6185 tcp_try_undo_spurious_syn(sk
);
6186 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
6188 /* Ok.. it's good. Set up sequence numbers and
6189 * move to established.
6191 WRITE_ONCE(tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
+ 1);
6192 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
6194 /* RFC1323: The window in SYN & SYN/ACK segments is
6197 tp
->snd_wnd
= ntohs(th
->window
);
6199 if (!tp
->rx_opt
.wscale_ok
) {
6200 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
6201 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
6204 if (tp
->rx_opt
.saw_tstamp
) {
6205 tp
->rx_opt
.tstamp_ok
= 1;
6206 tp
->tcp_header_len
=
6207 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
6208 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
6209 tcp_store_ts_recent(tp
);
6211 tp
->tcp_header_len
= sizeof(struct tcphdr
);
6214 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
6215 tcp_initialize_rcv_mss(sk
);
6217 /* Remember, tcp_poll() does not lock socket!
6218 * Change state from SYN-SENT only after copied_seq
6219 * is initialized. */
6220 WRITE_ONCE(tp
->copied_seq
, tp
->rcv_nxt
);
6222 smc_check_reset_syn(tp
);
6226 tcp_finish_connect(sk
, skb
);
6228 fastopen_fail
= (tp
->syn_fastopen
|| tp
->syn_data
) &&
6229 tcp_rcv_fastopen_synack(sk
, skb
, &foc
);
6231 if (!sock_flag(sk
, SOCK_DEAD
)) {
6232 sk
->sk_state_change(sk
);
6233 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
6237 if (sk
->sk_write_pending
||
6238 icsk
->icsk_accept_queue
.rskq_defer_accept
||
6239 inet_csk_in_pingpong_mode(sk
)) {
6240 /* Save one ACK. Data will be ready after
6241 * several ticks, if write_pending is set.
6243 * It may be deleted, but with this feature tcpdumps
6244 * look so _wonderfully_ clever, that I was not able
6245 * to stand against the temptation 8) --ANK
6247 inet_csk_schedule_ack(sk
);
6248 tcp_enter_quickack_mode(sk
, TCP_MAX_QUICKACKS
);
6249 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
6250 TCP_DELACK_MAX
, TCP_RTO_MAX
);
6261 /* No ACK in the segment */
6265 * "If the RST bit is set
6267 * Otherwise (no ACK) drop the segment and return."
6270 goto discard_and_undo
;
6274 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&&
6275 tcp_paws_reject(&tp
->rx_opt
, 0))
6276 goto discard_and_undo
;
6279 /* We see SYN without ACK. It is attempt of
6280 * simultaneous connect with crossed SYNs.
6281 * Particularly, it can be connect to self.
6283 tcp_set_state(sk
, TCP_SYN_RECV
);
6285 if (tp
->rx_opt
.saw_tstamp
) {
6286 tp
->rx_opt
.tstamp_ok
= 1;
6287 tcp_store_ts_recent(tp
);
6288 tp
->tcp_header_len
=
6289 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
6291 tp
->tcp_header_len
= sizeof(struct tcphdr
);
6294 WRITE_ONCE(tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
+ 1);
6295 WRITE_ONCE(tp
->copied_seq
, tp
->rcv_nxt
);
6296 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
6298 /* RFC1323: The window in SYN & SYN/ACK segments is
6301 tp
->snd_wnd
= ntohs(th
->window
);
6302 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
6303 tp
->max_window
= tp
->snd_wnd
;
6305 tcp_ecn_rcv_syn(tp
, th
);
6308 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
6309 tcp_initialize_rcv_mss(sk
);
6311 tcp_send_synack(sk
);
6313 /* Note, we could accept data and URG from this segment.
6314 * There are no obstacles to make this (except that we must
6315 * either change tcp_recvmsg() to prevent it from returning data
6316 * before 3WHS completes per RFC793, or employ TCP Fast Open).
6318 * However, if we ignore data in ACKless segments sometimes,
6319 * we have no reasons to accept it sometimes.
6320 * Also, seems the code doing it in step6 of tcp_rcv_state_process
6321 * is not flawless. So, discard packet for sanity.
6322 * Uncomment this return to process the data.
6329 /* "fifth, if neither of the SYN or RST bits is set then
6330 * drop the segment and return."
6334 tcp_clear_options(&tp
->rx_opt
);
6335 tp
->rx_opt
.mss_clamp
= saved_clamp
;
6339 tcp_clear_options(&tp
->rx_opt
);
6340 tp
->rx_opt
.mss_clamp
= saved_clamp
;
6344 static void tcp_rcv_synrecv_state_fastopen(struct sock
*sk
)
6346 struct request_sock
*req
;
6348 /* If we are still handling the SYNACK RTO, see if timestamp ECR allows
6349 * undo. If peer SACKs triggered fast recovery, we can't undo here.
6351 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
6352 tcp_try_undo_loss(sk
, false);
6354 /* Reset rtx states to prevent spurious retransmits_timed_out() */
6355 tcp_sk(sk
)->retrans_stamp
= 0;
6356 inet_csk(sk
)->icsk_retransmits
= 0;
6358 /* Once we leave TCP_SYN_RECV or TCP_FIN_WAIT_1,
6359 * we no longer need req so release it.
6361 req
= rcu_dereference_protected(tcp_sk(sk
)->fastopen_rsk
,
6362 lockdep_sock_is_held(sk
));
6363 reqsk_fastopen_remove(sk
, req
, false);
6365 /* Re-arm the timer because data may have been sent out.
6366 * This is similar to the regular data transmission case
6367 * when new data has just been ack'ed.
6369 * (TFO) - we could try to be more aggressive and
6370 * retransmitting any data sooner based on when they
6377 * This function implements the receiving procedure of RFC 793 for
6378 * all states except ESTABLISHED and TIME_WAIT.
6379 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
6380 * address independent.
6383 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
)
6385 struct tcp_sock
*tp
= tcp_sk(sk
);
6386 struct inet_connection_sock
*icsk
= inet_csk(sk
);
6387 const struct tcphdr
*th
= tcp_hdr(skb
);
6388 struct request_sock
*req
;
6392 switch (sk
->sk_state
) {
6406 /* It is possible that we process SYN packets from backlog,
6407 * so we need to make sure to disable BH and RCU right there.
6411 acceptable
= icsk
->icsk_af_ops
->conn_request(sk
, skb
) >= 0;
6423 tp
->rx_opt
.saw_tstamp
= 0;
6424 tcp_mstamp_refresh(tp
);
6425 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
);
6429 /* Do step6 onward by hand. */
6430 tcp_urg(sk
, skb
, th
);
6432 tcp_data_snd_check(sk
);
6436 tcp_mstamp_refresh(tp
);
6437 tp
->rx_opt
.saw_tstamp
= 0;
6438 req
= rcu_dereference_protected(tp
->fastopen_rsk
,
6439 lockdep_sock_is_held(sk
));
6443 WARN_ON_ONCE(sk
->sk_state
!= TCP_SYN_RECV
&&
6444 sk
->sk_state
!= TCP_FIN_WAIT1
);
6446 if (!tcp_check_req(sk
, skb
, req
, true, &req_stolen
))
6450 if (!th
->ack
&& !th
->rst
&& !th
->syn
)
6453 if (!tcp_validate_incoming(sk
, skb
, th
, 0))
6456 /* step 5: check the ACK field */
6457 acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
|
6458 FLAG_UPDATE_TS_RECENT
|
6459 FLAG_NO_CHALLENGE_ACK
) > 0;
6462 if (sk
->sk_state
== TCP_SYN_RECV
)
6463 return 1; /* send one RST */
6464 tcp_send_challenge_ack(sk
, skb
);
6467 switch (sk
->sk_state
) {
6469 tp
->delivered
++; /* SYN-ACK delivery isn't tracked in tcp_ack */
6471 tcp_synack_rtt_meas(sk
, req
);
6474 tcp_rcv_synrecv_state_fastopen(sk
);
6476 tcp_try_undo_spurious_syn(sk
);
6477 tp
->retrans_stamp
= 0;
6478 tcp_init_transfer(sk
, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB
,
6480 WRITE_ONCE(tp
->copied_seq
, tp
->rcv_nxt
);
6483 tcp_set_state(sk
, TCP_ESTABLISHED
);
6484 sk
->sk_state_change(sk
);
6486 /* Note, that this wakeup is only for marginal crossed SYN case.
6487 * Passively open sockets are not waked up, because
6488 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
6491 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
6493 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
6494 tp
->snd_wnd
= ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
;
6495 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
6497 if (tp
->rx_opt
.tstamp_ok
)
6498 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
6500 if (!inet_csk(sk
)->icsk_ca_ops
->cong_control
)
6501 tcp_update_pacing_rate(sk
);
6503 /* Prevent spurious tcp_cwnd_restart() on first data packet */
6504 tp
->lsndtime
= tcp_jiffies32
;
6506 tcp_initialize_rcv_mss(sk
);
6507 tcp_fast_path_on(tp
);
6510 case TCP_FIN_WAIT1
: {
6514 tcp_rcv_synrecv_state_fastopen(sk
);
6516 if (tp
->snd_una
!= tp
->write_seq
)
6519 tcp_set_state(sk
, TCP_FIN_WAIT2
);
6520 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
6524 if (!sock_flag(sk
, SOCK_DEAD
)) {
6525 /* Wake up lingering close() */
6526 sk
->sk_state_change(sk
);
6530 if (tp
->linger2
< 0) {
6532 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
6535 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
6536 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
6537 /* Receive out of order FIN after close() */
6538 if (tp
->syn_fastopen
&& th
->fin
)
6539 tcp_fastopen_active_disable(sk
);
6541 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
6545 tmo
= tcp_fin_time(sk
);
6546 if (tmo
> TCP_TIMEWAIT_LEN
) {
6547 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
6548 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
6549 /* Bad case. We could lose such FIN otherwise.
6550 * It is not a big problem, but it looks confusing
6551 * and not so rare event. We still can lose it now,
6552 * if it spins in bh_lock_sock(), but it is really
6555 inet_csk_reset_keepalive_timer(sk
, tmo
);
6557 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
6564 if (tp
->snd_una
== tp
->write_seq
) {
6565 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
6571 if (tp
->snd_una
== tp
->write_seq
) {
6572 tcp_update_metrics(sk
);
6579 /* step 6: check the URG bit */
6580 tcp_urg(sk
, skb
, th
);
6582 /* step 7: process the segment text */
6583 switch (sk
->sk_state
) {
6584 case TCP_CLOSE_WAIT
:
6587 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
6588 /* If a subflow has been reset, the packet should not
6589 * continue to be processed, drop the packet.
6591 if (sk_is_mptcp(sk
) && !mptcp_incoming_options(sk
, skb
))
6598 /* RFC 793 says to queue data in these states,
6599 * RFC 1122 says we MUST send a reset.
6600 * BSD 4.4 also does reset.
6602 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
6603 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
6604 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
6605 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
6611 case TCP_ESTABLISHED
:
6612 tcp_data_queue(sk
, skb
);
6617 /* tcp_data could move socket to TIME-WAIT */
6618 if (sk
->sk_state
!= TCP_CLOSE
) {
6619 tcp_data_snd_check(sk
);
6620 tcp_ack_snd_check(sk
);
6629 EXPORT_SYMBOL(tcp_rcv_state_process
);
6631 static inline void pr_drop_req(struct request_sock
*req
, __u16 port
, int family
)
6633 struct inet_request_sock
*ireq
= inet_rsk(req
);
6635 if (family
== AF_INET
)
6636 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6637 &ireq
->ir_rmt_addr
, port
);
6638 #if IS_ENABLED(CONFIG_IPV6)
6639 else if (family
== AF_INET6
)
6640 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6641 &ireq
->ir_v6_rmt_addr
, port
);
6645 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6647 * If we receive a SYN packet with these bits set, it means a
6648 * network is playing bad games with TOS bits. In order to
6649 * avoid possible false congestion notifications, we disable
6650 * TCP ECN negotiation.
6652 * Exception: tcp_ca wants ECN. This is required for DCTCP
6653 * congestion control: Linux DCTCP asserts ECT on all packets,
6654 * including SYN, which is most optimal solution; however,
6655 * others, such as FreeBSD do not.
6657 * Exception: At least one of the reserved bits of the TCP header (th->res1) is
6658 * set, indicating the use of a future TCP extension (such as AccECN). See
6659 * RFC8311 §4.3 which updates RFC3168 to allow the development of such
6662 static void tcp_ecn_create_request(struct request_sock
*req
,
6663 const struct sk_buff
*skb
,
6664 const struct sock
*listen_sk
,
6665 const struct dst_entry
*dst
)
6667 const struct tcphdr
*th
= tcp_hdr(skb
);
6668 const struct net
*net
= sock_net(listen_sk
);
6669 bool th_ecn
= th
->ece
&& th
->cwr
;
6676 ect
= !INET_ECN_is_not_ect(TCP_SKB_CB(skb
)->ip_dsfield
);
6677 ecn_ok_dst
= dst_feature(dst
, DST_FEATURE_ECN_MASK
);
6678 ecn_ok
= READ_ONCE(net
->ipv4
.sysctl_tcp_ecn
) || ecn_ok_dst
;
6680 if (((!ect
|| th
->res1
) && ecn_ok
) || tcp_ca_needs_ecn(listen_sk
) ||
6681 (ecn_ok_dst
& DST_FEATURE_ECN_CA
) ||
6682 tcp_bpf_ca_needs_ecn((struct sock
*)req
))
6683 inet_rsk(req
)->ecn_ok
= 1;
6686 static void tcp_openreq_init(struct request_sock
*req
,
6687 const struct tcp_options_received
*rx_opt
,
6688 struct sk_buff
*skb
, const struct sock
*sk
)
6690 struct inet_request_sock
*ireq
= inet_rsk(req
);
6692 req
->rsk_rcv_wnd
= 0; /* So that tcp_send_synack() knows! */
6693 tcp_rsk(req
)->rcv_isn
= TCP_SKB_CB(skb
)->seq
;
6694 tcp_rsk(req
)->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
6695 tcp_rsk(req
)->snt_synack
= 0;
6696 tcp_rsk(req
)->last_oow_ack_time
= 0;
6697 req
->mss
= rx_opt
->mss_clamp
;
6698 req
->ts_recent
= rx_opt
->saw_tstamp
? rx_opt
->rcv_tsval
: 0;
6699 ireq
->tstamp_ok
= rx_opt
->tstamp_ok
;
6700 ireq
->sack_ok
= rx_opt
->sack_ok
;
6701 ireq
->snd_wscale
= rx_opt
->snd_wscale
;
6702 ireq
->wscale_ok
= rx_opt
->wscale_ok
;
6705 ireq
->ir_rmt_port
= tcp_hdr(skb
)->source
;
6706 ireq
->ir_num
= ntohs(tcp_hdr(skb
)->dest
);
6707 ireq
->ir_mark
= inet_request_mark(sk
, skb
);
6708 #if IS_ENABLED(CONFIG_SMC)
6709 ireq
->smc_ok
= rx_opt
->smc_ok
;
6713 struct request_sock
*inet_reqsk_alloc(const struct request_sock_ops
*ops
,
6714 struct sock
*sk_listener
,
6715 bool attach_listener
)
6717 struct request_sock
*req
= reqsk_alloc(ops
, sk_listener
,
6721 struct inet_request_sock
*ireq
= inet_rsk(req
);
6723 ireq
->ireq_opt
= NULL
;
6724 #if IS_ENABLED(CONFIG_IPV6)
6725 ireq
->pktopts
= NULL
;
6727 atomic64_set(&ireq
->ir_cookie
, 0);
6728 ireq
->ireq_state
= TCP_NEW_SYN_RECV
;
6729 write_pnet(&ireq
->ireq_net
, sock_net(sk_listener
));
6730 ireq
->ireq_family
= sk_listener
->sk_family
;
6735 EXPORT_SYMBOL(inet_reqsk_alloc
);
6738 * Return true if a syncookie should be sent
6740 static bool tcp_syn_flood_action(const struct sock
*sk
, const char *proto
)
6742 struct request_sock_queue
*queue
= &inet_csk(sk
)->icsk_accept_queue
;
6743 const char *msg
= "Dropping request";
6744 struct net
*net
= sock_net(sk
);
6745 bool want_cookie
= false;
6748 syncookies
= READ_ONCE(net
->ipv4
.sysctl_tcp_syncookies
);
6750 #ifdef CONFIG_SYN_COOKIES
6752 msg
= "Sending cookies";
6754 __NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPREQQFULLDOCOOKIES
);
6757 __NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPREQQFULLDROP
);
6759 if (!queue
->synflood_warned
&& syncookies
!= 2 &&
6760 xchg(&queue
->synflood_warned
, 1) == 0)
6761 net_info_ratelimited("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6762 proto
, sk
->sk_num
, msg
);
6767 static void tcp_reqsk_record_syn(const struct sock
*sk
,
6768 struct request_sock
*req
,
6769 const struct sk_buff
*skb
)
6771 if (tcp_sk(sk
)->save_syn
) {
6772 u32 len
= skb_network_header_len(skb
) + tcp_hdrlen(skb
);
6773 struct saved_syn
*saved_syn
;
6777 if (tcp_sk(sk
)->save_syn
== 2) { /* Save full header. */
6778 base
= skb_mac_header(skb
);
6779 mac_hdrlen
= skb_mac_header_len(skb
);
6782 base
= skb_network_header(skb
);
6786 saved_syn
= kmalloc(struct_size(saved_syn
, data
, len
),
6789 saved_syn
->mac_hdrlen
= mac_hdrlen
;
6790 saved_syn
->network_hdrlen
= skb_network_header_len(skb
);
6791 saved_syn
->tcp_hdrlen
= tcp_hdrlen(skb
);
6792 memcpy(saved_syn
->data
, base
, len
);
6793 req
->saved_syn
= saved_syn
;
6798 /* If a SYN cookie is required and supported, returns a clamped MSS value to be
6799 * used for SYN cookie generation.
6801 u16
tcp_get_syncookie_mss(struct request_sock_ops
*rsk_ops
,
6802 const struct tcp_request_sock_ops
*af_ops
,
6803 struct sock
*sk
, struct tcphdr
*th
)
6805 struct tcp_sock
*tp
= tcp_sk(sk
);
6808 if (READ_ONCE(sock_net(sk
)->ipv4
.sysctl_tcp_syncookies
) != 2 &&
6809 !inet_csk_reqsk_queue_is_full(sk
))
6812 if (!tcp_syn_flood_action(sk
, rsk_ops
->slab_name
))
6815 if (sk_acceptq_is_full(sk
)) {
6816 NET_INC_STATS(sock_net(sk
), LINUX_MIB_LISTENOVERFLOWS
);
6820 mss
= tcp_parse_mss_option(th
, tp
->rx_opt
.user_mss
);
6822 mss
= af_ops
->mss_clamp
;
6826 EXPORT_SYMBOL_GPL(tcp_get_syncookie_mss
);
6828 int tcp_conn_request(struct request_sock_ops
*rsk_ops
,
6829 const struct tcp_request_sock_ops
*af_ops
,
6830 struct sock
*sk
, struct sk_buff
*skb
)
6832 struct tcp_fastopen_cookie foc
= { .len
= -1 };
6833 __u32 isn
= TCP_SKB_CB(skb
)->tcp_tw_isn
;
6834 struct tcp_options_received tmp_opt
;
6835 struct tcp_sock
*tp
= tcp_sk(sk
);
6836 struct net
*net
= sock_net(sk
);
6837 struct sock
*fastopen_sk
= NULL
;
6838 struct request_sock
*req
;
6839 bool want_cookie
= false;
6840 struct dst_entry
*dst
;
6844 syncookies
= READ_ONCE(net
->ipv4
.sysctl_tcp_syncookies
);
6846 /* TW buckets are converted to open requests without
6847 * limitations, they conserve resources and peer is
6848 * evidently real one.
6850 if ((syncookies
== 2 || inet_csk_reqsk_queue_is_full(sk
)) && !isn
) {
6851 want_cookie
= tcp_syn_flood_action(sk
, rsk_ops
->slab_name
);
6856 if (sk_acceptq_is_full(sk
)) {
6857 NET_INC_STATS(sock_net(sk
), LINUX_MIB_LISTENOVERFLOWS
);
6861 req
= inet_reqsk_alloc(rsk_ops
, sk
, !want_cookie
);
6865 req
->syncookie
= want_cookie
;
6866 tcp_rsk(req
)->af_specific
= af_ops
;
6867 tcp_rsk(req
)->ts_off
= 0;
6868 #if IS_ENABLED(CONFIG_MPTCP)
6869 tcp_rsk(req
)->is_mptcp
= 0;
6872 tcp_clear_options(&tmp_opt
);
6873 tmp_opt
.mss_clamp
= af_ops
->mss_clamp
;
6874 tmp_opt
.user_mss
= tp
->rx_opt
.user_mss
;
6875 tcp_parse_options(sock_net(sk
), skb
, &tmp_opt
, 0,
6876 want_cookie
? NULL
: &foc
);
6878 if (want_cookie
&& !tmp_opt
.saw_tstamp
)
6879 tcp_clear_options(&tmp_opt
);
6881 if (IS_ENABLED(CONFIG_SMC
) && want_cookie
)
6884 tmp_opt
.tstamp_ok
= tmp_opt
.saw_tstamp
;
6885 tcp_openreq_init(req
, &tmp_opt
, skb
, sk
);
6886 inet_rsk(req
)->no_srccheck
= inet_sk(sk
)->transparent
;
6888 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6889 inet_rsk(req
)->ir_iif
= inet_request_bound_dev_if(sk
, skb
);
6891 dst
= af_ops
->route_req(sk
, skb
, &fl
, req
);
6895 if (tmp_opt
.tstamp_ok
)
6896 tcp_rsk(req
)->ts_off
= af_ops
->init_ts_off(net
, skb
);
6898 if (!want_cookie
&& !isn
) {
6899 int max_syn_backlog
= READ_ONCE(net
->ipv4
.sysctl_max_syn_backlog
);
6901 /* Kill the following clause, if you dislike this way. */
6903 (max_syn_backlog
- inet_csk_reqsk_queue_len(sk
) <
6904 (max_syn_backlog
>> 2)) &&
6905 !tcp_peer_is_proven(req
, dst
)) {
6906 /* Without syncookies last quarter of
6907 * backlog is filled with destinations,
6908 * proven to be alive.
6909 * It means that we continue to communicate
6910 * to destinations, already remembered
6911 * to the moment of synflood.
6913 pr_drop_req(req
, ntohs(tcp_hdr(skb
)->source
),
6915 goto drop_and_release
;
6918 isn
= af_ops
->init_seq(skb
);
6921 tcp_ecn_create_request(req
, skb
, sk
, dst
);
6924 isn
= cookie_init_sequence(af_ops
, sk
, skb
, &req
->mss
);
6925 if (!tmp_opt
.tstamp_ok
)
6926 inet_rsk(req
)->ecn_ok
= 0;
6929 tcp_rsk(req
)->snt_isn
= isn
;
6930 tcp_rsk(req
)->txhash
= net_tx_rndhash();
6931 tcp_rsk(req
)->syn_tos
= TCP_SKB_CB(skb
)->ip_dsfield
;
6932 tcp_openreq_init_rwin(req
, sk
, dst
);
6933 sk_rx_queue_set(req_to_sk(req
), skb
);
6935 tcp_reqsk_record_syn(sk
, req
, skb
);
6936 fastopen_sk
= tcp_try_fastopen(sk
, skb
, req
, &foc
, dst
);
6939 af_ops
->send_synack(fastopen_sk
, dst
, &fl
, req
,
6940 &foc
, TCP_SYNACK_FASTOPEN
, skb
);
6941 /* Add the child socket directly into the accept queue */
6942 if (!inet_csk_reqsk_queue_add(sk
, req
, fastopen_sk
)) {
6943 reqsk_fastopen_remove(fastopen_sk
, req
, false);
6944 bh_unlock_sock(fastopen_sk
);
6945 sock_put(fastopen_sk
);
6948 sk
->sk_data_ready(sk
);
6949 bh_unlock_sock(fastopen_sk
);
6950 sock_put(fastopen_sk
);
6952 tcp_rsk(req
)->tfo_listener
= false;
6954 inet_csk_reqsk_queue_hash_add(sk
, req
,
6955 tcp_timeout_init((struct sock
*)req
));
6956 af_ops
->send_synack(sk
, dst
, &fl
, req
, &foc
,
6957 !want_cookie
? TCP_SYNACK_NORMAL
:
6976 EXPORT_SYMBOL(tcp_conn_request
);