2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
23 * Pedro Roque : Fast Retransmit/Recovery.
25 * Retransmit queue handled by TCP.
26 * Better retransmit timer handling.
27 * New congestion avoidance.
31 * Eric : Fast Retransmit.
32 * Randy Scott : MSS option defines.
33 * Eric Schenk : Fixes to slow start algorithm.
34 * Eric Schenk : Yet another double ACK bug.
35 * Eric Schenk : Delayed ACK bug fixes.
36 * Eric Schenk : Floyd style fast retrans war avoidance.
37 * David S. Miller : Don't allow zero congestion window.
38 * Eric Schenk : Fix retransmitter so that it sends
39 * next packet on ack of previous packet.
40 * Andi Kleen : Moved open_request checking here
41 * and process RSTs for open_requests.
42 * Andi Kleen : Better prune_queue, and other fixes.
43 * Andrey Savochkin: Fix RTT measurements in the presence of
45 * Andrey Savochkin: Check sequence numbers correctly when
46 * removing SACKs due to in sequence incoming
48 * Andi Kleen: Make sure we never ack data there is not
49 * enough room for. Also make this condition
50 * a fatal error if it might still happen.
51 * Andi Kleen: Add tcp_measure_rcv_mss to make
52 * connections with MSS<min(MTU,ann. MSS)
53 * work without delayed acks.
54 * Andi Kleen: Process packets with PSH set in the
56 * J Hadi Salim: ECN support
59 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
60 * engine. Lots of bugs are found.
61 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
64 #define pr_fmt(fmt) "TCP: " fmt
67 #include <linux/slab.h>
68 #include <linux/module.h>
69 #include <linux/sysctl.h>
70 #include <linux/kernel.h>
71 #include <linux/prefetch.h>
74 #include <net/inet_common.h>
75 #include <linux/ipsec.h>
76 #include <asm/unaligned.h>
77 #include <linux/errqueue.h>
79 int sysctl_tcp_timestamps __read_mostly
= 1;
80 int sysctl_tcp_window_scaling __read_mostly
= 1;
81 int sysctl_tcp_sack __read_mostly
= 1;
82 int sysctl_tcp_fack __read_mostly
= 1;
83 int sysctl_tcp_reordering __read_mostly
= TCP_FASTRETRANS_THRESH
;
84 int sysctl_tcp_max_reordering __read_mostly
= 300;
85 EXPORT_SYMBOL(sysctl_tcp_reordering
);
86 int sysctl_tcp_dsack __read_mostly
= 1;
87 int sysctl_tcp_app_win __read_mostly
= 31;
88 int sysctl_tcp_adv_win_scale __read_mostly
= 1;
89 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale
);
91 /* rfc5961 challenge ack rate limiting */
92 int sysctl_tcp_challenge_ack_limit
= 100;
94 int sysctl_tcp_stdurg __read_mostly
;
95 int sysctl_tcp_rfc1337 __read_mostly
;
96 int sysctl_tcp_max_orphans __read_mostly
= NR_FILE
;
97 int sysctl_tcp_frto __read_mostly
= 2;
98 int sysctl_tcp_min_rtt_wlen __read_mostly
= 300;
100 int sysctl_tcp_thin_dupack __read_mostly
;
102 int sysctl_tcp_moderate_rcvbuf __read_mostly
= 1;
103 int sysctl_tcp_early_retrans __read_mostly
= 3;
104 int sysctl_tcp_invalid_ratelimit __read_mostly
= HZ
/2;
106 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
107 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
108 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
109 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
110 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
111 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
112 #define FLAG_ECE 0x40 /* ECE in this ACK */
113 #define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */
114 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
115 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
116 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
117 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
118 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
119 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
121 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
122 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
123 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
124 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
126 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
127 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
129 /* Adapt the MSS value used to make delayed ack decision to the
132 static void tcp_measure_rcv_mss(struct sock
*sk
, const struct sk_buff
*skb
)
134 struct inet_connection_sock
*icsk
= inet_csk(sk
);
135 const unsigned int lss
= icsk
->icsk_ack
.last_seg_size
;
138 icsk
->icsk_ack
.last_seg_size
= 0;
140 /* skb->len may jitter because of SACKs, even if peer
141 * sends good full-sized frames.
143 len
= skb_shinfo(skb
)->gso_size
? : skb
->len
;
144 if (len
>= icsk
->icsk_ack
.rcv_mss
) {
145 icsk
->icsk_ack
.rcv_mss
= len
;
147 /* Otherwise, we make more careful check taking into account,
148 * that SACKs block is variable.
150 * "len" is invariant segment length, including TCP header.
152 len
+= skb
->data
- skb_transport_header(skb
);
153 if (len
>= TCP_MSS_DEFAULT
+ sizeof(struct tcphdr
) ||
154 /* If PSH is not set, packet should be
155 * full sized, provided peer TCP is not badly broken.
156 * This observation (if it is correct 8)) allows
157 * to handle super-low mtu links fairly.
159 (len
>= TCP_MIN_MSS
+ sizeof(struct tcphdr
) &&
160 !(tcp_flag_word(tcp_hdr(skb
)) & TCP_REMNANT
))) {
161 /* Subtract also invariant (if peer is RFC compliant),
162 * tcp header plus fixed timestamp option length.
163 * Resulting "len" is MSS free of SACK jitter.
165 len
-= tcp_sk(sk
)->tcp_header_len
;
166 icsk
->icsk_ack
.last_seg_size
= len
;
168 icsk
->icsk_ack
.rcv_mss
= len
;
172 if (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
)
173 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED2
;
174 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
178 static void tcp_incr_quickack(struct sock
*sk
)
180 struct inet_connection_sock
*icsk
= inet_csk(sk
);
181 unsigned int quickacks
= tcp_sk(sk
)->rcv_wnd
/ (2 * icsk
->icsk_ack
.rcv_mss
);
185 if (quickacks
> icsk
->icsk_ack
.quick
)
186 icsk
->icsk_ack
.quick
= min(quickacks
, TCP_MAX_QUICKACKS
);
189 static void tcp_enter_quickack_mode(struct sock
*sk
)
191 struct inet_connection_sock
*icsk
= inet_csk(sk
);
192 tcp_incr_quickack(sk
);
193 icsk
->icsk_ack
.pingpong
= 0;
194 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
197 /* Send ACKs quickly, if "quick" count is not exhausted
198 * and the session is not interactive.
201 static bool tcp_in_quickack_mode(struct sock
*sk
)
203 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
204 const struct dst_entry
*dst
= __sk_dst_get(sk
);
206 return (dst
&& dst_metric(dst
, RTAX_QUICKACK
)) ||
207 (icsk
->icsk_ack
.quick
&& !icsk
->icsk_ack
.pingpong
);
210 static void tcp_ecn_queue_cwr(struct tcp_sock
*tp
)
212 if (tp
->ecn_flags
& TCP_ECN_OK
)
213 tp
->ecn_flags
|= TCP_ECN_QUEUE_CWR
;
216 static void tcp_ecn_accept_cwr(struct tcp_sock
*tp
, const struct sk_buff
*skb
)
218 if (tcp_hdr(skb
)->cwr
)
219 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
222 static void tcp_ecn_withdraw_cwr(struct tcp_sock
*tp
)
224 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
227 static void __tcp_ecn_check_ce(struct tcp_sock
*tp
, const struct sk_buff
*skb
)
229 switch (TCP_SKB_CB(skb
)->ip_dsfield
& INET_ECN_MASK
) {
230 case INET_ECN_NOT_ECT
:
231 /* Funny extension: if ECT is not set on a segment,
232 * and we already seen ECT on a previous segment,
233 * it is probably a retransmit.
235 if (tp
->ecn_flags
& TCP_ECN_SEEN
)
236 tcp_enter_quickack_mode((struct sock
*)tp
);
239 if (tcp_ca_needs_ecn((struct sock
*)tp
))
240 tcp_ca_event((struct sock
*)tp
, CA_EVENT_ECN_IS_CE
);
242 if (!(tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)) {
243 /* Better not delay acks, sender can have a very low cwnd */
244 tcp_enter_quickack_mode((struct sock
*)tp
);
245 tp
->ecn_flags
|= TCP_ECN_DEMAND_CWR
;
247 tp
->ecn_flags
|= TCP_ECN_SEEN
;
250 if (tcp_ca_needs_ecn((struct sock
*)tp
))
251 tcp_ca_event((struct sock
*)tp
, CA_EVENT_ECN_NO_CE
);
252 tp
->ecn_flags
|= TCP_ECN_SEEN
;
257 static void tcp_ecn_check_ce(struct tcp_sock
*tp
, const struct sk_buff
*skb
)
259 if (tp
->ecn_flags
& TCP_ECN_OK
)
260 __tcp_ecn_check_ce(tp
, skb
);
263 static void tcp_ecn_rcv_synack(struct tcp_sock
*tp
, const struct tcphdr
*th
)
265 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| th
->cwr
))
266 tp
->ecn_flags
&= ~TCP_ECN_OK
;
269 static void tcp_ecn_rcv_syn(struct tcp_sock
*tp
, const struct tcphdr
*th
)
271 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| !th
->cwr
))
272 tp
->ecn_flags
&= ~TCP_ECN_OK
;
275 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock
*tp
, const struct tcphdr
*th
)
277 if (th
->ece
&& !th
->syn
&& (tp
->ecn_flags
& TCP_ECN_OK
))
282 /* Buffer size and advertised window tuning.
284 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
287 static void tcp_sndbuf_expand(struct sock
*sk
)
289 const struct tcp_sock
*tp
= tcp_sk(sk
);
293 /* Worst case is non GSO/TSO : each frame consumes one skb
294 * and skb->head is kmalloced using power of two area of memory
296 per_mss
= max_t(u32
, tp
->rx_opt
.mss_clamp
, tp
->mss_cache
) +
298 SKB_DATA_ALIGN(sizeof(struct skb_shared_info
));
300 per_mss
= roundup_pow_of_two(per_mss
) +
301 SKB_DATA_ALIGN(sizeof(struct sk_buff
));
303 nr_segs
= max_t(u32
, TCP_INIT_CWND
, tp
->snd_cwnd
);
304 nr_segs
= max_t(u32
, nr_segs
, tp
->reordering
+ 1);
306 /* Fast Recovery (RFC 5681 3.2) :
307 * Cubic needs 1.7 factor, rounded to 2 to include
308 * extra cushion (application might react slowly to POLLOUT)
310 sndmem
= 2 * nr_segs
* per_mss
;
312 if (sk
->sk_sndbuf
< sndmem
)
313 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
316 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
318 * All tcp_full_space() is split to two parts: "network" buffer, allocated
319 * forward and advertised in receiver window (tp->rcv_wnd) and
320 * "application buffer", required to isolate scheduling/application
321 * latencies from network.
322 * window_clamp is maximal advertised window. It can be less than
323 * tcp_full_space(), in this case tcp_full_space() - window_clamp
324 * is reserved for "application" buffer. The less window_clamp is
325 * the smoother our behaviour from viewpoint of network, but the lower
326 * throughput and the higher sensitivity of the connection to losses. 8)
328 * rcv_ssthresh is more strict window_clamp used at "slow start"
329 * phase to predict further behaviour of this connection.
330 * It is used for two goals:
331 * - to enforce header prediction at sender, even when application
332 * requires some significant "application buffer". It is check #1.
333 * - to prevent pruning of receive queue because of misprediction
334 * of receiver window. Check #2.
336 * The scheme does not work when sender sends good segments opening
337 * window and then starts to feed us spaghetti. But it should work
338 * in common situations. Otherwise, we have to rely on queue collapsing.
341 /* Slow part of check#2. */
342 static int __tcp_grow_window(const struct sock
*sk
, const struct sk_buff
*skb
)
344 struct tcp_sock
*tp
= tcp_sk(sk
);
346 int truesize
= tcp_win_from_space(skb
->truesize
) >> 1;
347 int window
= tcp_win_from_space(sysctl_tcp_rmem
[2]) >> 1;
349 while (tp
->rcv_ssthresh
<= window
) {
350 if (truesize
<= skb
->len
)
351 return 2 * inet_csk(sk
)->icsk_ack
.rcv_mss
;
359 static void tcp_grow_window(struct sock
*sk
, const struct sk_buff
*skb
)
361 struct tcp_sock
*tp
= tcp_sk(sk
);
364 if (tp
->rcv_ssthresh
< tp
->window_clamp
&&
365 (int)tp
->rcv_ssthresh
< tcp_space(sk
) &&
366 !tcp_under_memory_pressure(sk
)) {
369 /* Check #2. Increase window, if skb with such overhead
370 * will fit to rcvbuf in future.
372 if (tcp_win_from_space(skb
->truesize
) <= skb
->len
)
373 incr
= 2 * tp
->advmss
;
375 incr
= __tcp_grow_window(sk
, skb
);
378 incr
= max_t(int, incr
, 2 * skb
->len
);
379 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
+ incr
,
381 inet_csk(sk
)->icsk_ack
.quick
|= 1;
386 /* 3. Tuning rcvbuf, when connection enters established state. */
387 static void tcp_fixup_rcvbuf(struct sock
*sk
)
389 u32 mss
= tcp_sk(sk
)->advmss
;
392 rcvmem
= 2 * SKB_TRUESIZE(mss
+ MAX_TCP_HEADER
) *
393 tcp_default_init_rwnd(mss
);
395 /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
396 * Allow enough cushion so that sender is not limited by our window
398 if (sysctl_tcp_moderate_rcvbuf
)
401 if (sk
->sk_rcvbuf
< rcvmem
)
402 sk
->sk_rcvbuf
= min(rcvmem
, sysctl_tcp_rmem
[2]);
405 /* 4. Try to fixup all. It is made immediately after connection enters
408 void tcp_init_buffer_space(struct sock
*sk
)
410 struct tcp_sock
*tp
= tcp_sk(sk
);
413 if (!(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
))
414 tcp_fixup_rcvbuf(sk
);
415 if (!(sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
))
416 tcp_sndbuf_expand(sk
);
418 tp
->rcvq_space
.space
= tp
->rcv_wnd
;
419 tp
->rcvq_space
.time
= tcp_time_stamp
;
420 tp
->rcvq_space
.seq
= tp
->copied_seq
;
422 maxwin
= tcp_full_space(sk
);
424 if (tp
->window_clamp
>= maxwin
) {
425 tp
->window_clamp
= maxwin
;
427 if (sysctl_tcp_app_win
&& maxwin
> 4 * tp
->advmss
)
428 tp
->window_clamp
= max(maxwin
-
429 (maxwin
>> sysctl_tcp_app_win
),
433 /* Force reservation of one segment. */
434 if (sysctl_tcp_app_win
&&
435 tp
->window_clamp
> 2 * tp
->advmss
&&
436 tp
->window_clamp
+ tp
->advmss
> maxwin
)
437 tp
->window_clamp
= max(2 * tp
->advmss
, maxwin
- tp
->advmss
);
439 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, tp
->window_clamp
);
440 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
443 /* 5. Recalculate window clamp after socket hit its memory bounds. */
444 static void tcp_clamp_window(struct sock
*sk
)
446 struct tcp_sock
*tp
= tcp_sk(sk
);
447 struct inet_connection_sock
*icsk
= inet_csk(sk
);
449 icsk
->icsk_ack
.quick
= 0;
451 if (sk
->sk_rcvbuf
< sysctl_tcp_rmem
[2] &&
452 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
) &&
453 !tcp_under_memory_pressure(sk
) &&
454 sk_memory_allocated(sk
) < sk_prot_mem_limits(sk
, 0)) {
455 sk
->sk_rcvbuf
= min(atomic_read(&sk
->sk_rmem_alloc
),
458 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
)
459 tp
->rcv_ssthresh
= min(tp
->window_clamp
, 2U * tp
->advmss
);
462 /* Initialize RCV_MSS value.
463 * RCV_MSS is an our guess about MSS used by the peer.
464 * We haven't any direct information about the MSS.
465 * It's better to underestimate the RCV_MSS rather than overestimate.
466 * Overestimations make us ACKing less frequently than needed.
467 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
469 void tcp_initialize_rcv_mss(struct sock
*sk
)
471 const struct tcp_sock
*tp
= tcp_sk(sk
);
472 unsigned int hint
= min_t(unsigned int, tp
->advmss
, tp
->mss_cache
);
474 hint
= min(hint
, tp
->rcv_wnd
/ 2);
475 hint
= min(hint
, TCP_MSS_DEFAULT
);
476 hint
= max(hint
, TCP_MIN_MSS
);
478 inet_csk(sk
)->icsk_ack
.rcv_mss
= hint
;
480 EXPORT_SYMBOL(tcp_initialize_rcv_mss
);
482 /* Receiver "autotuning" code.
484 * The algorithm for RTT estimation w/o timestamps is based on
485 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
486 * <http://public.lanl.gov/radiant/pubs.html#DRS>
488 * More detail on this code can be found at
489 * <http://staff.psc.edu/jheffner/>,
490 * though this reference is out of date. A new paper
493 static void tcp_rcv_rtt_update(struct tcp_sock
*tp
, u32 sample
, int win_dep
)
495 u32 new_sample
= tp
->rcv_rtt_est
.rtt
;
501 if (new_sample
!= 0) {
502 /* If we sample in larger samples in the non-timestamp
503 * case, we could grossly overestimate the RTT especially
504 * with chatty applications or bulk transfer apps which
505 * are stalled on filesystem I/O.
507 * Also, since we are only going for a minimum in the
508 * non-timestamp case, we do not smooth things out
509 * else with timestamps disabled convergence takes too
513 m
-= (new_sample
>> 3);
521 /* No previous measure. */
525 if (tp
->rcv_rtt_est
.rtt
!= new_sample
)
526 tp
->rcv_rtt_est
.rtt
= new_sample
;
529 static inline void tcp_rcv_rtt_measure(struct tcp_sock
*tp
)
531 if (tp
->rcv_rtt_est
.time
== 0)
533 if (before(tp
->rcv_nxt
, tp
->rcv_rtt_est
.seq
))
535 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rcv_rtt_est
.time
, 1);
538 tp
->rcv_rtt_est
.seq
= tp
->rcv_nxt
+ tp
->rcv_wnd
;
539 tp
->rcv_rtt_est
.time
= tcp_time_stamp
;
542 static inline void tcp_rcv_rtt_measure_ts(struct sock
*sk
,
543 const struct sk_buff
*skb
)
545 struct tcp_sock
*tp
= tcp_sk(sk
);
546 if (tp
->rx_opt
.rcv_tsecr
&&
547 (TCP_SKB_CB(skb
)->end_seq
-
548 TCP_SKB_CB(skb
)->seq
>= inet_csk(sk
)->icsk_ack
.rcv_mss
))
549 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
, 0);
553 * This function should be called every time data is copied to user space.
554 * It calculates the appropriate TCP receive buffer space.
556 void tcp_rcv_space_adjust(struct sock
*sk
)
558 struct tcp_sock
*tp
= tcp_sk(sk
);
562 time
= tcp_time_stamp
- tp
->rcvq_space
.time
;
563 if (time
< (tp
->rcv_rtt_est
.rtt
>> 3) || tp
->rcv_rtt_est
.rtt
== 0)
566 /* Number of bytes copied to user in last RTT */
567 copied
= tp
->copied_seq
- tp
->rcvq_space
.seq
;
568 if (copied
<= tp
->rcvq_space
.space
)
572 * copied = bytes received in previous RTT, our base window
573 * To cope with packet losses, we need a 2x factor
574 * To cope with slow start, and sender growing its cwin by 100 %
575 * every RTT, we need a 4x factor, because the ACK we are sending
576 * now is for the next RTT, not the current one :
577 * <prev RTT . ><current RTT .. ><next RTT .... >
580 if (sysctl_tcp_moderate_rcvbuf
&&
581 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
)) {
582 int rcvwin
, rcvmem
, rcvbuf
;
584 /* minimal window to cope with packet losses, assuming
585 * steady state. Add some cushion because of small variations.
587 rcvwin
= (copied
<< 1) + 16 * tp
->advmss
;
589 /* If rate increased by 25%,
590 * assume slow start, rcvwin = 3 * copied
591 * If rate increased by 50%,
592 * assume sender can use 2x growth, rcvwin = 4 * copied
595 tp
->rcvq_space
.space
+ (tp
->rcvq_space
.space
>> 2)) {
597 tp
->rcvq_space
.space
+ (tp
->rcvq_space
.space
>> 1))
600 rcvwin
+= (rcvwin
>> 1);
603 rcvmem
= SKB_TRUESIZE(tp
->advmss
+ MAX_TCP_HEADER
);
604 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
607 rcvbuf
= min(rcvwin
/ tp
->advmss
* rcvmem
, sysctl_tcp_rmem
[2]);
608 if (rcvbuf
> sk
->sk_rcvbuf
) {
609 sk
->sk_rcvbuf
= rcvbuf
;
611 /* Make the window clamp follow along. */
612 tp
->window_clamp
= rcvwin
;
615 tp
->rcvq_space
.space
= copied
;
618 tp
->rcvq_space
.seq
= tp
->copied_seq
;
619 tp
->rcvq_space
.time
= tcp_time_stamp
;
622 /* There is something which you must keep in mind when you analyze the
623 * behavior of the tp->ato delayed ack timeout interval. When a
624 * connection starts up, we want to ack as quickly as possible. The
625 * problem is that "good" TCP's do slow start at the beginning of data
626 * transmission. The means that until we send the first few ACK's the
627 * sender will sit on his end and only queue most of his data, because
628 * he can only send snd_cwnd unacked packets at any given time. For
629 * each ACK we send, he increments snd_cwnd and transmits more of his
632 static void tcp_event_data_recv(struct sock
*sk
, struct sk_buff
*skb
)
634 struct tcp_sock
*tp
= tcp_sk(sk
);
635 struct inet_connection_sock
*icsk
= inet_csk(sk
);
638 inet_csk_schedule_ack(sk
);
640 tcp_measure_rcv_mss(sk
, skb
);
642 tcp_rcv_rtt_measure(tp
);
644 now
= tcp_time_stamp
;
646 if (!icsk
->icsk_ack
.ato
) {
647 /* The _first_ data packet received, initialize
648 * delayed ACK engine.
650 tcp_incr_quickack(sk
);
651 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
653 int m
= now
- icsk
->icsk_ack
.lrcvtime
;
655 if (m
<= TCP_ATO_MIN
/ 2) {
656 /* The fastest case is the first. */
657 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + TCP_ATO_MIN
/ 2;
658 } else if (m
< icsk
->icsk_ack
.ato
) {
659 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + m
;
660 if (icsk
->icsk_ack
.ato
> icsk
->icsk_rto
)
661 icsk
->icsk_ack
.ato
= icsk
->icsk_rto
;
662 } else if (m
> icsk
->icsk_rto
) {
663 /* Too long gap. Apparently sender failed to
664 * restart window, so that we send ACKs quickly.
666 tcp_incr_quickack(sk
);
670 icsk
->icsk_ack
.lrcvtime
= now
;
672 tcp_ecn_check_ce(tp
, skb
);
675 tcp_grow_window(sk
, skb
);
678 /* Called to compute a smoothed rtt estimate. The data fed to this
679 * routine either comes from timestamps, or from segments that were
680 * known _not_ to have been retransmitted [see Karn/Partridge
681 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
682 * piece by Van Jacobson.
683 * NOTE: the next three routines used to be one big routine.
684 * To save cycles in the RFC 1323 implementation it was better to break
685 * it up into three procedures. -- erics
687 static void tcp_rtt_estimator(struct sock
*sk
, long mrtt_us
)
689 struct tcp_sock
*tp
= tcp_sk(sk
);
690 long m
= mrtt_us
; /* RTT */
691 u32 srtt
= tp
->srtt_us
;
693 /* The following amusing code comes from Jacobson's
694 * article in SIGCOMM '88. Note that rtt and mdev
695 * are scaled versions of rtt and mean deviation.
696 * This is designed to be as fast as possible
697 * m stands for "measurement".
699 * On a 1990 paper the rto value is changed to:
700 * RTO = rtt + 4 * mdev
702 * Funny. This algorithm seems to be very broken.
703 * These formulae increase RTO, when it should be decreased, increase
704 * too slowly, when it should be increased quickly, decrease too quickly
705 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
706 * does not matter how to _calculate_ it. Seems, it was trap
707 * that VJ failed to avoid. 8)
710 m
-= (srtt
>> 3); /* m is now error in rtt est */
711 srtt
+= m
; /* rtt = 7/8 rtt + 1/8 new */
713 m
= -m
; /* m is now abs(error) */
714 m
-= (tp
->mdev_us
>> 2); /* similar update on mdev */
715 /* This is similar to one of Eifel findings.
716 * Eifel blocks mdev updates when rtt decreases.
717 * This solution is a bit different: we use finer gain
718 * for mdev in this case (alpha*beta).
719 * Like Eifel it also prevents growth of rto,
720 * but also it limits too fast rto decreases,
721 * happening in pure Eifel.
726 m
-= (tp
->mdev_us
>> 2); /* similar update on mdev */
728 tp
->mdev_us
+= m
; /* mdev = 3/4 mdev + 1/4 new */
729 if (tp
->mdev_us
> tp
->mdev_max_us
) {
730 tp
->mdev_max_us
= tp
->mdev_us
;
731 if (tp
->mdev_max_us
> tp
->rttvar_us
)
732 tp
->rttvar_us
= tp
->mdev_max_us
;
734 if (after(tp
->snd_una
, tp
->rtt_seq
)) {
735 if (tp
->mdev_max_us
< tp
->rttvar_us
)
736 tp
->rttvar_us
-= (tp
->rttvar_us
- tp
->mdev_max_us
) >> 2;
737 tp
->rtt_seq
= tp
->snd_nxt
;
738 tp
->mdev_max_us
= tcp_rto_min_us(sk
);
741 /* no previous measure. */
742 srtt
= m
<< 3; /* take the measured time to be rtt */
743 tp
->mdev_us
= m
<< 1; /* make sure rto = 3*rtt */
744 tp
->rttvar_us
= max(tp
->mdev_us
, tcp_rto_min_us(sk
));
745 tp
->mdev_max_us
= tp
->rttvar_us
;
746 tp
->rtt_seq
= tp
->snd_nxt
;
748 tp
->srtt_us
= max(1U, srtt
);
751 /* Set the sk_pacing_rate to allow proper sizing of TSO packets.
752 * Note: TCP stack does not yet implement pacing.
753 * FQ packet scheduler can be used to implement cheap but effective
754 * TCP pacing, to smooth the burst on large writes when packets
755 * in flight is significantly lower than cwnd (or rwin)
757 int sysctl_tcp_pacing_ss_ratio __read_mostly
= 200;
758 int sysctl_tcp_pacing_ca_ratio __read_mostly
= 120;
760 static void tcp_update_pacing_rate(struct sock
*sk
)
762 const struct tcp_sock
*tp
= tcp_sk(sk
);
765 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
766 rate
= (u64
)tp
->mss_cache
* ((USEC_PER_SEC
/ 100) << 3);
768 /* current rate is (cwnd * mss) / srtt
769 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
770 * In Congestion Avoidance phase, set it to 120 % the current rate.
772 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
773 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
774 * end of slow start and should slow down.
776 if (tp
->snd_cwnd
< tp
->snd_ssthresh
/ 2)
777 rate
*= sysctl_tcp_pacing_ss_ratio
;
779 rate
*= sysctl_tcp_pacing_ca_ratio
;
781 rate
*= max(tp
->snd_cwnd
, tp
->packets_out
);
783 if (likely(tp
->srtt_us
))
784 do_div(rate
, tp
->srtt_us
);
786 /* ACCESS_ONCE() is needed because sch_fq fetches sk_pacing_rate
787 * without any lock. We want to make sure compiler wont store
788 * intermediate values in this location.
790 ACCESS_ONCE(sk
->sk_pacing_rate
) = min_t(u64
, rate
,
791 sk
->sk_max_pacing_rate
);
794 /* Calculate rto without backoff. This is the second half of Van Jacobson's
795 * routine referred to above.
797 static void tcp_set_rto(struct sock
*sk
)
799 const struct tcp_sock
*tp
= tcp_sk(sk
);
800 /* Old crap is replaced with new one. 8)
803 * 1. If rtt variance happened to be less 50msec, it is hallucination.
804 * It cannot be less due to utterly erratic ACK generation made
805 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
806 * to do with delayed acks, because at cwnd>2 true delack timeout
807 * is invisible. Actually, Linux-2.4 also generates erratic
808 * ACKs in some circumstances.
810 inet_csk(sk
)->icsk_rto
= __tcp_set_rto(tp
);
812 /* 2. Fixups made earlier cannot be right.
813 * If we do not estimate RTO correctly without them,
814 * all the algo is pure shit and should be replaced
815 * with correct one. It is exactly, which we pretend to do.
818 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
819 * guarantees that rto is higher.
824 __u32
tcp_init_cwnd(const struct tcp_sock
*tp
, const struct dst_entry
*dst
)
826 __u32 cwnd
= (dst
? dst_metric(dst
, RTAX_INITCWND
) : 0);
829 cwnd
= TCP_INIT_CWND
;
830 return min_t(__u32
, cwnd
, tp
->snd_cwnd_clamp
);
834 * Packet counting of FACK is based on in-order assumptions, therefore TCP
835 * disables it when reordering is detected
837 void tcp_disable_fack(struct tcp_sock
*tp
)
839 /* RFC3517 uses different metric in lost marker => reset on change */
841 tp
->lost_skb_hint
= NULL
;
842 tp
->rx_opt
.sack_ok
&= ~TCP_FACK_ENABLED
;
845 /* Take a notice that peer is sending D-SACKs */
846 static void tcp_dsack_seen(struct tcp_sock
*tp
)
848 tp
->rx_opt
.sack_ok
|= TCP_DSACK_SEEN
;
851 static void tcp_update_reordering(struct sock
*sk
, const int metric
,
854 struct tcp_sock
*tp
= tcp_sk(sk
);
855 if (metric
> tp
->reordering
) {
858 tp
->reordering
= min(sysctl_tcp_max_reordering
, metric
);
860 /* This exciting event is worth to be remembered. 8) */
862 mib_idx
= LINUX_MIB_TCPTSREORDER
;
863 else if (tcp_is_reno(tp
))
864 mib_idx
= LINUX_MIB_TCPRENOREORDER
;
865 else if (tcp_is_fack(tp
))
866 mib_idx
= LINUX_MIB_TCPFACKREORDER
;
868 mib_idx
= LINUX_MIB_TCPSACKREORDER
;
870 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
871 #if FASTRETRANS_DEBUG > 1
872 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
873 tp
->rx_opt
.sack_ok
, inet_csk(sk
)->icsk_ca_state
,
877 tp
->undo_marker
? tp
->undo_retrans
: 0);
879 tcp_disable_fack(tp
);
883 tcp_disable_early_retrans(tp
);
887 /* This must be called before lost_out is incremented */
888 static void tcp_verify_retransmit_hint(struct tcp_sock
*tp
, struct sk_buff
*skb
)
890 if (!tp
->retransmit_skb_hint
||
891 before(TCP_SKB_CB(skb
)->seq
,
892 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
893 tp
->retransmit_skb_hint
= skb
;
896 after(TCP_SKB_CB(skb
)->end_seq
, tp
->retransmit_high
))
897 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
900 static void tcp_skb_mark_lost(struct tcp_sock
*tp
, struct sk_buff
*skb
)
902 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
903 tcp_verify_retransmit_hint(tp
, skb
);
905 tp
->lost_out
+= tcp_skb_pcount(skb
);
906 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
910 void tcp_skb_mark_lost_uncond_verify(struct tcp_sock
*tp
, struct sk_buff
*skb
)
912 tcp_verify_retransmit_hint(tp
, skb
);
914 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
915 tp
->lost_out
+= tcp_skb_pcount(skb
);
916 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
920 /* This procedure tags the retransmission queue when SACKs arrive.
922 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
923 * Packets in queue with these bits set are counted in variables
924 * sacked_out, retrans_out and lost_out, correspondingly.
926 * Valid combinations are:
927 * Tag InFlight Description
928 * 0 1 - orig segment is in flight.
929 * S 0 - nothing flies, orig reached receiver.
930 * L 0 - nothing flies, orig lost by net.
931 * R 2 - both orig and retransmit are in flight.
932 * L|R 1 - orig is lost, retransmit is in flight.
933 * S|R 1 - orig reached receiver, retrans is still in flight.
934 * (L|S|R is logically valid, it could occur when L|R is sacked,
935 * but it is equivalent to plain S and code short-curcuits it to S.
936 * L|S is logically invalid, it would mean -1 packet in flight 8))
938 * These 6 states form finite state machine, controlled by the following events:
939 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
940 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
941 * 3. Loss detection event of two flavors:
942 * A. Scoreboard estimator decided the packet is lost.
943 * A'. Reno "three dupacks" marks head of queue lost.
944 * A''. Its FACK modification, head until snd.fack is lost.
945 * B. SACK arrives sacking SND.NXT at the moment, when the
946 * segment was retransmitted.
947 * 4. D-SACK added new rule: D-SACK changes any tag to S.
949 * It is pleasant to note, that state diagram turns out to be commutative,
950 * so that we are allowed not to be bothered by order of our actions,
951 * when multiple events arrive simultaneously. (see the function below).
953 * Reordering detection.
954 * --------------------
955 * Reordering metric is maximal distance, which a packet can be displaced
956 * in packet stream. With SACKs we can estimate it:
958 * 1. SACK fills old hole and the corresponding segment was not
959 * ever retransmitted -> reordering. Alas, we cannot use it
960 * when segment was retransmitted.
961 * 2. The last flaw is solved with D-SACK. D-SACK arrives
962 * for retransmitted and already SACKed segment -> reordering..
963 * Both of these heuristics are not used in Loss state, when we cannot
964 * account for retransmits accurately.
966 * SACK block validation.
967 * ----------------------
969 * SACK block range validation checks that the received SACK block fits to
970 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
971 * Note that SND.UNA is not included to the range though being valid because
972 * it means that the receiver is rather inconsistent with itself reporting
973 * SACK reneging when it should advance SND.UNA. Such SACK block this is
974 * perfectly valid, however, in light of RFC2018 which explicitly states
975 * that "SACK block MUST reflect the newest segment. Even if the newest
976 * segment is going to be discarded ...", not that it looks very clever
977 * in case of head skb. Due to potentional receiver driven attacks, we
978 * choose to avoid immediate execution of a walk in write queue due to
979 * reneging and defer head skb's loss recovery to standard loss recovery
980 * procedure that will eventually trigger (nothing forbids us doing this).
982 * Implements also blockage to start_seq wrap-around. Problem lies in the
983 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
984 * there's no guarantee that it will be before snd_nxt (n). The problem
985 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
988 * <- outs wnd -> <- wrapzone ->
989 * u e n u_w e_w s n_w
991 * |<------------+------+----- TCP seqno space --------------+---------->|
992 * ...-- <2^31 ->| |<--------...
993 * ...---- >2^31 ------>| |<--------...
995 * Current code wouldn't be vulnerable but it's better still to discard such
996 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
997 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
998 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
999 * equal to the ideal case (infinite seqno space without wrap caused issues).
1001 * With D-SACK the lower bound is extended to cover sequence space below
1002 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1003 * again, D-SACK block must not to go across snd_una (for the same reason as
1004 * for the normal SACK blocks, explained above). But there all simplicity
1005 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1006 * fully below undo_marker they do not affect behavior in anyway and can
1007 * therefore be safely ignored. In rare cases (which are more or less
1008 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1009 * fragmentation and packet reordering past skb's retransmission. To consider
1010 * them correctly, the acceptable range must be extended even more though
1011 * the exact amount is rather hard to quantify. However, tp->max_window can
1012 * be used as an exaggerated estimate.
1014 static bool tcp_is_sackblock_valid(struct tcp_sock
*tp
, bool is_dsack
,
1015 u32 start_seq
, u32 end_seq
)
1017 /* Too far in future, or reversed (interpretation is ambiguous) */
1018 if (after(end_seq
, tp
->snd_nxt
) || !before(start_seq
, end_seq
))
1021 /* Nasty start_seq wrap-around check (see comments above) */
1022 if (!before(start_seq
, tp
->snd_nxt
))
1025 /* In outstanding window? ...This is valid exit for D-SACKs too.
1026 * start_seq == snd_una is non-sensical (see comments above)
1028 if (after(start_seq
, tp
->snd_una
))
1031 if (!is_dsack
|| !tp
->undo_marker
)
1034 /* ...Then it's D-SACK, and must reside below snd_una completely */
1035 if (after(end_seq
, tp
->snd_una
))
1038 if (!before(start_seq
, tp
->undo_marker
))
1042 if (!after(end_seq
, tp
->undo_marker
))
1045 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1046 * start_seq < undo_marker and end_seq >= undo_marker.
1048 return !before(start_seq
, end_seq
- tp
->max_window
);
1051 static bool tcp_check_dsack(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1052 struct tcp_sack_block_wire
*sp
, int num_sacks
,
1055 struct tcp_sock
*tp
= tcp_sk(sk
);
1056 u32 start_seq_0
= get_unaligned_be32(&sp
[0].start_seq
);
1057 u32 end_seq_0
= get_unaligned_be32(&sp
[0].end_seq
);
1058 bool dup_sack
= false;
1060 if (before(start_seq_0
, TCP_SKB_CB(ack_skb
)->ack_seq
)) {
1063 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPDSACKRECV
);
1064 } else if (num_sacks
> 1) {
1065 u32 end_seq_1
= get_unaligned_be32(&sp
[1].end_seq
);
1066 u32 start_seq_1
= get_unaligned_be32(&sp
[1].start_seq
);
1068 if (!after(end_seq_0
, end_seq_1
) &&
1069 !before(start_seq_0
, start_seq_1
)) {
1072 NET_INC_STATS_BH(sock_net(sk
),
1073 LINUX_MIB_TCPDSACKOFORECV
);
1077 /* D-SACK for already forgotten data... Do dumb counting. */
1078 if (dup_sack
&& tp
->undo_marker
&& tp
->undo_retrans
> 0 &&
1079 !after(end_seq_0
, prior_snd_una
) &&
1080 after(end_seq_0
, tp
->undo_marker
))
1086 struct tcp_sacktag_state
{
1089 /* Timestamps for earliest and latest never-retransmitted segment
1090 * that was SACKed. RTO needs the earliest RTT to stay conservative,
1091 * but congestion control should still get an accurate delay signal.
1093 struct skb_mstamp first_sackt
;
1094 struct skb_mstamp last_sackt
;
1098 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1099 * the incoming SACK may not exactly match but we can find smaller MSS
1100 * aligned portion of it that matches. Therefore we might need to fragment
1101 * which may fail and creates some hassle (caller must handle error case
1104 * FIXME: this could be merged to shift decision code
1106 static int tcp_match_skb_to_sack(struct sock
*sk
, struct sk_buff
*skb
,
1107 u32 start_seq
, u32 end_seq
)
1111 unsigned int pkt_len
;
1114 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1115 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1117 if (tcp_skb_pcount(skb
) > 1 && !in_sack
&&
1118 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
1119 mss
= tcp_skb_mss(skb
);
1120 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1123 pkt_len
= start_seq
- TCP_SKB_CB(skb
)->seq
;
1127 pkt_len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1132 /* Round if necessary so that SACKs cover only full MSSes
1133 * and/or the remaining small portion (if present)
1135 if (pkt_len
> mss
) {
1136 unsigned int new_len
= (pkt_len
/ mss
) * mss
;
1137 if (!in_sack
&& new_len
< pkt_len
) {
1139 if (new_len
>= skb
->len
)
1144 err
= tcp_fragment(sk
, skb
, pkt_len
, mss
, GFP_ATOMIC
);
1152 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1153 static u8
tcp_sacktag_one(struct sock
*sk
,
1154 struct tcp_sacktag_state
*state
, u8 sacked
,
1155 u32 start_seq
, u32 end_seq
,
1156 int dup_sack
, int pcount
,
1157 const struct skb_mstamp
*xmit_time
)
1159 struct tcp_sock
*tp
= tcp_sk(sk
);
1160 int fack_count
= state
->fack_count
;
1162 /* Account D-SACK for retransmitted packet. */
1163 if (dup_sack
&& (sacked
& TCPCB_RETRANS
)) {
1164 if (tp
->undo_marker
&& tp
->undo_retrans
> 0 &&
1165 after(end_seq
, tp
->undo_marker
))
1167 if (sacked
& TCPCB_SACKED_ACKED
)
1168 state
->reord
= min(fack_count
, state
->reord
);
1171 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1172 if (!after(end_seq
, tp
->snd_una
))
1175 if (!(sacked
& TCPCB_SACKED_ACKED
)) {
1176 tcp_rack_advance(tp
, xmit_time
, sacked
);
1178 if (sacked
& TCPCB_SACKED_RETRANS
) {
1179 /* If the segment is not tagged as lost,
1180 * we do not clear RETRANS, believing
1181 * that retransmission is still in flight.
1183 if (sacked
& TCPCB_LOST
) {
1184 sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1185 tp
->lost_out
-= pcount
;
1186 tp
->retrans_out
-= pcount
;
1189 if (!(sacked
& TCPCB_RETRANS
)) {
1190 /* New sack for not retransmitted frame,
1191 * which was in hole. It is reordering.
1193 if (before(start_seq
,
1194 tcp_highest_sack_seq(tp
)))
1195 state
->reord
= min(fack_count
,
1197 if (!after(end_seq
, tp
->high_seq
))
1198 state
->flag
|= FLAG_ORIG_SACK_ACKED
;
1199 if (state
->first_sackt
.v64
== 0)
1200 state
->first_sackt
= *xmit_time
;
1201 state
->last_sackt
= *xmit_time
;
1204 if (sacked
& TCPCB_LOST
) {
1205 sacked
&= ~TCPCB_LOST
;
1206 tp
->lost_out
-= pcount
;
1210 sacked
|= TCPCB_SACKED_ACKED
;
1211 state
->flag
|= FLAG_DATA_SACKED
;
1212 tp
->sacked_out
+= pcount
;
1214 fack_count
+= pcount
;
1216 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1217 if (!tcp_is_fack(tp
) && tp
->lost_skb_hint
&&
1218 before(start_seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
))
1219 tp
->lost_cnt_hint
+= pcount
;
1221 if (fack_count
> tp
->fackets_out
)
1222 tp
->fackets_out
= fack_count
;
1225 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1226 * frames and clear it. undo_retrans is decreased above, L|R frames
1227 * are accounted above as well.
1229 if (dup_sack
&& (sacked
& TCPCB_SACKED_RETRANS
)) {
1230 sacked
&= ~TCPCB_SACKED_RETRANS
;
1231 tp
->retrans_out
-= pcount
;
1237 /* Shift newly-SACKed bytes from this skb to the immediately previous
1238 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1240 static bool tcp_shifted_skb(struct sock
*sk
, struct sk_buff
*skb
,
1241 struct tcp_sacktag_state
*state
,
1242 unsigned int pcount
, int shifted
, int mss
,
1245 struct tcp_sock
*tp
= tcp_sk(sk
);
1246 struct sk_buff
*prev
= tcp_write_queue_prev(sk
, skb
);
1247 u32 start_seq
= TCP_SKB_CB(skb
)->seq
; /* start of newly-SACKed */
1248 u32 end_seq
= start_seq
+ shifted
; /* end of newly-SACKed */
1252 /* Adjust counters and hints for the newly sacked sequence
1253 * range but discard the return value since prev is already
1254 * marked. We must tag the range first because the seq
1255 * advancement below implicitly advances
1256 * tcp_highest_sack_seq() when skb is highest_sack.
1258 tcp_sacktag_one(sk
, state
, TCP_SKB_CB(skb
)->sacked
,
1259 start_seq
, end_seq
, dup_sack
, pcount
,
1262 if (skb
== tp
->lost_skb_hint
)
1263 tp
->lost_cnt_hint
+= pcount
;
1265 TCP_SKB_CB(prev
)->end_seq
+= shifted
;
1266 TCP_SKB_CB(skb
)->seq
+= shifted
;
1268 tcp_skb_pcount_add(prev
, pcount
);
1269 BUG_ON(tcp_skb_pcount(skb
) < pcount
);
1270 tcp_skb_pcount_add(skb
, -pcount
);
1272 /* When we're adding to gso_segs == 1, gso_size will be zero,
1273 * in theory this shouldn't be necessary but as long as DSACK
1274 * code can come after this skb later on it's better to keep
1275 * setting gso_size to something.
1277 if (!TCP_SKB_CB(prev
)->tcp_gso_size
)
1278 TCP_SKB_CB(prev
)->tcp_gso_size
= mss
;
1280 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1281 if (tcp_skb_pcount(skb
) <= 1)
1282 TCP_SKB_CB(skb
)->tcp_gso_size
= 0;
1284 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1285 TCP_SKB_CB(prev
)->sacked
|= (TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
);
1288 BUG_ON(!tcp_skb_pcount(skb
));
1289 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKSHIFTED
);
1293 /* Whole SKB was eaten :-) */
1295 if (skb
== tp
->retransmit_skb_hint
)
1296 tp
->retransmit_skb_hint
= prev
;
1297 if (skb
== tp
->lost_skb_hint
) {
1298 tp
->lost_skb_hint
= prev
;
1299 tp
->lost_cnt_hint
-= tcp_skb_pcount(prev
);
1302 TCP_SKB_CB(prev
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
;
1303 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
1304 TCP_SKB_CB(prev
)->end_seq
++;
1306 if (skb
== tcp_highest_sack(sk
))
1307 tcp_advance_highest_sack(sk
, skb
);
1309 tcp_unlink_write_queue(skb
, sk
);
1310 sk_wmem_free_skb(sk
, skb
);
1312 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKMERGED
);
1317 /* I wish gso_size would have a bit more sane initialization than
1318 * something-or-zero which complicates things
1320 static int tcp_skb_seglen(const struct sk_buff
*skb
)
1322 return tcp_skb_pcount(skb
) == 1 ? skb
->len
: tcp_skb_mss(skb
);
1325 /* Shifting pages past head area doesn't work */
1326 static int skb_can_shift(const struct sk_buff
*skb
)
1328 return !skb_headlen(skb
) && skb_is_nonlinear(skb
);
1331 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1334 static struct sk_buff
*tcp_shift_skb_data(struct sock
*sk
, struct sk_buff
*skb
,
1335 struct tcp_sacktag_state
*state
,
1336 u32 start_seq
, u32 end_seq
,
1339 struct tcp_sock
*tp
= tcp_sk(sk
);
1340 struct sk_buff
*prev
;
1346 if (!sk_can_gso(sk
))
1349 /* Normally R but no L won't result in plain S */
1351 (TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_RETRANS
)) == TCPCB_SACKED_RETRANS
)
1353 if (!skb_can_shift(skb
))
1355 /* This frame is about to be dropped (was ACKed). */
1356 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1359 /* Can only happen with delayed DSACK + discard craziness */
1360 if (unlikely(skb
== tcp_write_queue_head(sk
)))
1362 prev
= tcp_write_queue_prev(sk
, skb
);
1364 if ((TCP_SKB_CB(prev
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
)
1367 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1368 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1372 pcount
= tcp_skb_pcount(skb
);
1373 mss
= tcp_skb_seglen(skb
);
1375 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1376 * drop this restriction as unnecessary
1378 if (mss
!= tcp_skb_seglen(prev
))
1381 if (!after(TCP_SKB_CB(skb
)->end_seq
, start_seq
))
1383 /* CHECKME: This is non-MSS split case only?, this will
1384 * cause skipped skbs due to advancing loop btw, original
1385 * has that feature too
1387 if (tcp_skb_pcount(skb
) <= 1)
1390 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1392 /* TODO: head merge to next could be attempted here
1393 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1394 * though it might not be worth of the additional hassle
1396 * ...we can probably just fallback to what was done
1397 * previously. We could try merging non-SACKed ones
1398 * as well but it probably isn't going to buy off
1399 * because later SACKs might again split them, and
1400 * it would make skb timestamp tracking considerably
1406 len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1408 BUG_ON(len
> skb
->len
);
1410 /* MSS boundaries should be honoured or else pcount will
1411 * severely break even though it makes things bit trickier.
1412 * Optimize common case to avoid most of the divides
1414 mss
= tcp_skb_mss(skb
);
1416 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1417 * drop this restriction as unnecessary
1419 if (mss
!= tcp_skb_seglen(prev
))
1424 } else if (len
< mss
) {
1432 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1433 if (!after(TCP_SKB_CB(skb
)->seq
+ len
, tp
->snd_una
))
1436 if (!skb_shift(prev
, skb
, len
))
1438 if (!tcp_shifted_skb(sk
, skb
, state
, pcount
, len
, mss
, dup_sack
))
1441 /* Hole filled allows collapsing with the next as well, this is very
1442 * useful when hole on every nth skb pattern happens
1444 if (prev
== tcp_write_queue_tail(sk
))
1446 skb
= tcp_write_queue_next(sk
, prev
);
1448 if (!skb_can_shift(skb
) ||
1449 (skb
== tcp_send_head(sk
)) ||
1450 ((TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
) ||
1451 (mss
!= tcp_skb_seglen(skb
)))
1455 if (skb_shift(prev
, skb
, len
)) {
1456 pcount
+= tcp_skb_pcount(skb
);
1457 tcp_shifted_skb(sk
, skb
, state
, tcp_skb_pcount(skb
), len
, mss
, 0);
1461 state
->fack_count
+= pcount
;
1468 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKSHIFTFALLBACK
);
1472 static struct sk_buff
*tcp_sacktag_walk(struct sk_buff
*skb
, struct sock
*sk
,
1473 struct tcp_sack_block
*next_dup
,
1474 struct tcp_sacktag_state
*state
,
1475 u32 start_seq
, u32 end_seq
,
1478 struct tcp_sock
*tp
= tcp_sk(sk
);
1479 struct sk_buff
*tmp
;
1481 tcp_for_write_queue_from(skb
, sk
) {
1483 bool dup_sack
= dup_sack_in
;
1485 if (skb
== tcp_send_head(sk
))
1488 /* queue is in-order => we can short-circuit the walk early */
1489 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1493 before(TCP_SKB_CB(skb
)->seq
, next_dup
->end_seq
)) {
1494 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1495 next_dup
->start_seq
,
1501 /* skb reference here is a bit tricky to get right, since
1502 * shifting can eat and free both this skb and the next,
1503 * so not even _safe variant of the loop is enough.
1506 tmp
= tcp_shift_skb_data(sk
, skb
, state
,
1507 start_seq
, end_seq
, dup_sack
);
1516 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1522 if (unlikely(in_sack
< 0))
1526 TCP_SKB_CB(skb
)->sacked
=
1529 TCP_SKB_CB(skb
)->sacked
,
1530 TCP_SKB_CB(skb
)->seq
,
1531 TCP_SKB_CB(skb
)->end_seq
,
1533 tcp_skb_pcount(skb
),
1536 if (!before(TCP_SKB_CB(skb
)->seq
,
1537 tcp_highest_sack_seq(tp
)))
1538 tcp_advance_highest_sack(sk
, skb
);
1541 state
->fack_count
+= tcp_skb_pcount(skb
);
1546 /* Avoid all extra work that is being done by sacktag while walking in
1549 static struct sk_buff
*tcp_sacktag_skip(struct sk_buff
*skb
, struct sock
*sk
,
1550 struct tcp_sacktag_state
*state
,
1553 tcp_for_write_queue_from(skb
, sk
) {
1554 if (skb
== tcp_send_head(sk
))
1557 if (after(TCP_SKB_CB(skb
)->end_seq
, skip_to_seq
))
1560 state
->fack_count
+= tcp_skb_pcount(skb
);
1565 static struct sk_buff
*tcp_maybe_skipping_dsack(struct sk_buff
*skb
,
1567 struct tcp_sack_block
*next_dup
,
1568 struct tcp_sacktag_state
*state
,
1574 if (before(next_dup
->start_seq
, skip_to_seq
)) {
1575 skb
= tcp_sacktag_skip(skb
, sk
, state
, next_dup
->start_seq
);
1576 skb
= tcp_sacktag_walk(skb
, sk
, NULL
, state
,
1577 next_dup
->start_seq
, next_dup
->end_seq
,
1584 static int tcp_sack_cache_ok(const struct tcp_sock
*tp
, const struct tcp_sack_block
*cache
)
1586 return cache
< tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1590 tcp_sacktag_write_queue(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1591 u32 prior_snd_una
, struct tcp_sacktag_state
*state
)
1593 struct tcp_sock
*tp
= tcp_sk(sk
);
1594 const unsigned char *ptr
= (skb_transport_header(ack_skb
) +
1595 TCP_SKB_CB(ack_skb
)->sacked
);
1596 struct tcp_sack_block_wire
*sp_wire
= (struct tcp_sack_block_wire
*)(ptr
+2);
1597 struct tcp_sack_block sp
[TCP_NUM_SACKS
];
1598 struct tcp_sack_block
*cache
;
1599 struct sk_buff
*skb
;
1600 int num_sacks
= min(TCP_NUM_SACKS
, (ptr
[1] - TCPOLEN_SACK_BASE
) >> 3);
1602 bool found_dup_sack
= false;
1604 int first_sack_index
;
1607 state
->reord
= tp
->packets_out
;
1609 if (!tp
->sacked_out
) {
1610 if (WARN_ON(tp
->fackets_out
))
1611 tp
->fackets_out
= 0;
1612 tcp_highest_sack_reset(sk
);
1615 found_dup_sack
= tcp_check_dsack(sk
, ack_skb
, sp_wire
,
1616 num_sacks
, prior_snd_una
);
1618 state
->flag
|= FLAG_DSACKING_ACK
;
1620 /* Eliminate too old ACKs, but take into
1621 * account more or less fresh ones, they can
1622 * contain valid SACK info.
1624 if (before(TCP_SKB_CB(ack_skb
)->ack_seq
, prior_snd_una
- tp
->max_window
))
1627 if (!tp
->packets_out
)
1631 first_sack_index
= 0;
1632 for (i
= 0; i
< num_sacks
; i
++) {
1633 bool dup_sack
= !i
&& found_dup_sack
;
1635 sp
[used_sacks
].start_seq
= get_unaligned_be32(&sp_wire
[i
].start_seq
);
1636 sp
[used_sacks
].end_seq
= get_unaligned_be32(&sp_wire
[i
].end_seq
);
1638 if (!tcp_is_sackblock_valid(tp
, dup_sack
,
1639 sp
[used_sacks
].start_seq
,
1640 sp
[used_sacks
].end_seq
)) {
1644 if (!tp
->undo_marker
)
1645 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDNOUNDO
;
1647 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDOLD
;
1649 /* Don't count olds caused by ACK reordering */
1650 if ((TCP_SKB_CB(ack_skb
)->ack_seq
!= tp
->snd_una
) &&
1651 !after(sp
[used_sacks
].end_seq
, tp
->snd_una
))
1653 mib_idx
= LINUX_MIB_TCPSACKDISCARD
;
1656 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
1658 first_sack_index
= -1;
1662 /* Ignore very old stuff early */
1663 if (!after(sp
[used_sacks
].end_seq
, prior_snd_una
))
1669 /* order SACK blocks to allow in order walk of the retrans queue */
1670 for (i
= used_sacks
- 1; i
> 0; i
--) {
1671 for (j
= 0; j
< i
; j
++) {
1672 if (after(sp
[j
].start_seq
, sp
[j
+ 1].start_seq
)) {
1673 swap(sp
[j
], sp
[j
+ 1]);
1675 /* Track where the first SACK block goes to */
1676 if (j
== first_sack_index
)
1677 first_sack_index
= j
+ 1;
1682 skb
= tcp_write_queue_head(sk
);
1683 state
->fack_count
= 0;
1686 if (!tp
->sacked_out
) {
1687 /* It's already past, so skip checking against it */
1688 cache
= tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1690 cache
= tp
->recv_sack_cache
;
1691 /* Skip empty blocks in at head of the cache */
1692 while (tcp_sack_cache_ok(tp
, cache
) && !cache
->start_seq
&&
1697 while (i
< used_sacks
) {
1698 u32 start_seq
= sp
[i
].start_seq
;
1699 u32 end_seq
= sp
[i
].end_seq
;
1700 bool dup_sack
= (found_dup_sack
&& (i
== first_sack_index
));
1701 struct tcp_sack_block
*next_dup
= NULL
;
1703 if (found_dup_sack
&& ((i
+ 1) == first_sack_index
))
1704 next_dup
= &sp
[i
+ 1];
1706 /* Skip too early cached blocks */
1707 while (tcp_sack_cache_ok(tp
, cache
) &&
1708 !before(start_seq
, cache
->end_seq
))
1711 /* Can skip some work by looking recv_sack_cache? */
1712 if (tcp_sack_cache_ok(tp
, cache
) && !dup_sack
&&
1713 after(end_seq
, cache
->start_seq
)) {
1716 if (before(start_seq
, cache
->start_seq
)) {
1717 skb
= tcp_sacktag_skip(skb
, sk
, state
,
1719 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
,
1726 /* Rest of the block already fully processed? */
1727 if (!after(end_seq
, cache
->end_seq
))
1730 skb
= tcp_maybe_skipping_dsack(skb
, sk
, next_dup
,
1734 /* ...tail remains todo... */
1735 if (tcp_highest_sack_seq(tp
) == cache
->end_seq
) {
1736 /* ...but better entrypoint exists! */
1737 skb
= tcp_highest_sack(sk
);
1740 state
->fack_count
= tp
->fackets_out
;
1745 skb
= tcp_sacktag_skip(skb
, sk
, state
, cache
->end_seq
);
1746 /* Check overlap against next cached too (past this one already) */
1751 if (!before(start_seq
, tcp_highest_sack_seq(tp
))) {
1752 skb
= tcp_highest_sack(sk
);
1755 state
->fack_count
= tp
->fackets_out
;
1757 skb
= tcp_sacktag_skip(skb
, sk
, state
, start_seq
);
1760 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
, state
,
1761 start_seq
, end_seq
, dup_sack
);
1767 /* Clear the head of the cache sack blocks so we can skip it next time */
1768 for (i
= 0; i
< ARRAY_SIZE(tp
->recv_sack_cache
) - used_sacks
; i
++) {
1769 tp
->recv_sack_cache
[i
].start_seq
= 0;
1770 tp
->recv_sack_cache
[i
].end_seq
= 0;
1772 for (j
= 0; j
< used_sacks
; j
++)
1773 tp
->recv_sack_cache
[i
++] = sp
[j
];
1775 if ((state
->reord
< tp
->fackets_out
) &&
1776 ((inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Loss
) || tp
->undo_marker
))
1777 tcp_update_reordering(sk
, tp
->fackets_out
- state
->reord
, 0);
1779 tcp_verify_left_out(tp
);
1782 #if FASTRETRANS_DEBUG > 0
1783 WARN_ON((int)tp
->sacked_out
< 0);
1784 WARN_ON((int)tp
->lost_out
< 0);
1785 WARN_ON((int)tp
->retrans_out
< 0);
1786 WARN_ON((int)tcp_packets_in_flight(tp
) < 0);
1791 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1792 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1794 static bool tcp_limit_reno_sacked(struct tcp_sock
*tp
)
1798 holes
= max(tp
->lost_out
, 1U);
1799 holes
= min(holes
, tp
->packets_out
);
1801 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1802 tp
->sacked_out
= tp
->packets_out
- holes
;
1808 /* If we receive more dupacks than we expected counting segments
1809 * in assumption of absent reordering, interpret this as reordering.
1810 * The only another reason could be bug in receiver TCP.
1812 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1814 struct tcp_sock
*tp
= tcp_sk(sk
);
1815 if (tcp_limit_reno_sacked(tp
))
1816 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1819 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1821 static void tcp_add_reno_sack(struct sock
*sk
)
1823 struct tcp_sock
*tp
= tcp_sk(sk
);
1825 tcp_check_reno_reordering(sk
, 0);
1826 tcp_verify_left_out(tp
);
1829 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1831 static void tcp_remove_reno_sacks(struct sock
*sk
, int acked
)
1833 struct tcp_sock
*tp
= tcp_sk(sk
);
1836 /* One ACK acked hole. The rest eat duplicate ACKs. */
1837 if (acked
- 1 >= tp
->sacked_out
)
1840 tp
->sacked_out
-= acked
- 1;
1842 tcp_check_reno_reordering(sk
, acked
);
1843 tcp_verify_left_out(tp
);
1846 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1851 void tcp_clear_retrans(struct tcp_sock
*tp
)
1853 tp
->retrans_out
= 0;
1855 tp
->undo_marker
= 0;
1856 tp
->undo_retrans
= -1;
1857 tp
->fackets_out
= 0;
1861 static inline void tcp_init_undo(struct tcp_sock
*tp
)
1863 tp
->undo_marker
= tp
->snd_una
;
1864 /* Retransmission still in flight may cause DSACKs later. */
1865 tp
->undo_retrans
= tp
->retrans_out
? : -1;
1868 /* Enter Loss state. If we detect SACK reneging, forget all SACK information
1869 * and reset tags completely, otherwise preserve SACKs. If receiver
1870 * dropped its ofo queue, we will know this due to reneging detection.
1872 void tcp_enter_loss(struct sock
*sk
)
1874 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1875 struct tcp_sock
*tp
= tcp_sk(sk
);
1876 struct sk_buff
*skb
;
1877 bool new_recovery
= icsk
->icsk_ca_state
< TCP_CA_Recovery
;
1878 bool is_reneg
; /* is receiver reneging on SACKs? */
1880 /* Reduce ssthresh if it has not yet been made inside this window. */
1881 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
||
1882 !after(tp
->high_seq
, tp
->snd_una
) ||
1883 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
1884 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1885 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1886 tcp_ca_event(sk
, CA_EVENT_LOSS
);
1890 tp
->snd_cwnd_cnt
= 0;
1891 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1893 tp
->retrans_out
= 0;
1896 if (tcp_is_reno(tp
))
1897 tcp_reset_reno_sack(tp
);
1899 skb
= tcp_write_queue_head(sk
);
1900 is_reneg
= skb
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
);
1902 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSACKRENEGING
);
1904 tp
->fackets_out
= 0;
1906 tcp_clear_all_retrans_hints(tp
);
1908 tcp_for_write_queue(skb
, sk
) {
1909 if (skb
== tcp_send_head(sk
))
1912 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
1913 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) || is_reneg
) {
1914 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
1915 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1916 tp
->lost_out
+= tcp_skb_pcount(skb
);
1917 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
1920 tcp_verify_left_out(tp
);
1922 /* Timeout in disordered state after receiving substantial DUPACKs
1923 * suggests that the degree of reordering is over-estimated.
1925 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
&&
1926 tp
->sacked_out
>= sysctl_tcp_reordering
)
1927 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1928 sysctl_tcp_reordering
);
1929 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1930 tp
->high_seq
= tp
->snd_nxt
;
1931 tcp_ecn_queue_cwr(tp
);
1933 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1934 * loss recovery is underway except recurring timeout(s) on
1935 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1937 tp
->frto
= sysctl_tcp_frto
&&
1938 (new_recovery
|| icsk
->icsk_retransmits
) &&
1939 !inet_csk(sk
)->icsk_mtup
.probe_size
;
1942 /* If ACK arrived pointing to a remembered SACK, it means that our
1943 * remembered SACKs do not reflect real state of receiver i.e.
1944 * receiver _host_ is heavily congested (or buggy).
1946 * To avoid big spurious retransmission bursts due to transient SACK
1947 * scoreboard oddities that look like reneging, we give the receiver a
1948 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
1949 * restore sanity to the SACK scoreboard. If the apparent reneging
1950 * persists until this RTO then we'll clear the SACK scoreboard.
1952 static bool tcp_check_sack_reneging(struct sock
*sk
, int flag
)
1954 if (flag
& FLAG_SACK_RENEGING
) {
1955 struct tcp_sock
*tp
= tcp_sk(sk
);
1956 unsigned long delay
= max(usecs_to_jiffies(tp
->srtt_us
>> 4),
1957 msecs_to_jiffies(10));
1959 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1960 delay
, TCP_RTO_MAX
);
1966 static inline int tcp_fackets_out(const struct tcp_sock
*tp
)
1968 return tcp_is_reno(tp
) ? tp
->sacked_out
+ 1 : tp
->fackets_out
;
1971 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
1972 * counter when SACK is enabled (without SACK, sacked_out is used for
1975 * Instead, with FACK TCP uses fackets_out that includes both SACKed
1976 * segments up to the highest received SACK block so far and holes in
1979 * With reordering, holes may still be in flight, so RFC3517 recovery
1980 * uses pure sacked_out (total number of SACKed segments) even though
1981 * it violates the RFC that uses duplicate ACKs, often these are equal
1982 * but when e.g. out-of-window ACKs or packet duplication occurs,
1983 * they differ. Since neither occurs due to loss, TCP should really
1986 static inline int tcp_dupack_heuristics(const struct tcp_sock
*tp
)
1988 return tcp_is_fack(tp
) ? tp
->fackets_out
: tp
->sacked_out
+ 1;
1991 static bool tcp_pause_early_retransmit(struct sock
*sk
, int flag
)
1993 struct tcp_sock
*tp
= tcp_sk(sk
);
1994 unsigned long delay
;
1996 /* Delay early retransmit and entering fast recovery for
1997 * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
1998 * available, or RTO is scheduled to fire first.
2000 if (sysctl_tcp_early_retrans
< 2 || sysctl_tcp_early_retrans
> 3 ||
2001 (flag
& FLAG_ECE
) || !tp
->srtt_us
)
2004 delay
= max(usecs_to_jiffies(tp
->srtt_us
>> 5),
2005 msecs_to_jiffies(2));
2007 if (!time_after(inet_csk(sk
)->icsk_timeout
, (jiffies
+ delay
)))
2010 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_EARLY_RETRANS
, delay
,
2015 /* Linux NewReno/SACK/FACK/ECN state machine.
2016 * --------------------------------------
2018 * "Open" Normal state, no dubious events, fast path.
2019 * "Disorder" In all the respects it is "Open",
2020 * but requires a bit more attention. It is entered when
2021 * we see some SACKs or dupacks. It is split of "Open"
2022 * mainly to move some processing from fast path to slow one.
2023 * "CWR" CWND was reduced due to some Congestion Notification event.
2024 * It can be ECN, ICMP source quench, local device congestion.
2025 * "Recovery" CWND was reduced, we are fast-retransmitting.
2026 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2028 * tcp_fastretrans_alert() is entered:
2029 * - each incoming ACK, if state is not "Open"
2030 * - when arrived ACK is unusual, namely:
2035 * Counting packets in flight is pretty simple.
2037 * in_flight = packets_out - left_out + retrans_out
2039 * packets_out is SND.NXT-SND.UNA counted in packets.
2041 * retrans_out is number of retransmitted segments.
2043 * left_out is number of segments left network, but not ACKed yet.
2045 * left_out = sacked_out + lost_out
2047 * sacked_out: Packets, which arrived to receiver out of order
2048 * and hence not ACKed. With SACKs this number is simply
2049 * amount of SACKed data. Even without SACKs
2050 * it is easy to give pretty reliable estimate of this number,
2051 * counting duplicate ACKs.
2053 * lost_out: Packets lost by network. TCP has no explicit
2054 * "loss notification" feedback from network (for now).
2055 * It means that this number can be only _guessed_.
2056 * Actually, it is the heuristics to predict lossage that
2057 * distinguishes different algorithms.
2059 * F.e. after RTO, when all the queue is considered as lost,
2060 * lost_out = packets_out and in_flight = retrans_out.
2062 * Essentially, we have now two algorithms counting
2065 * FACK: It is the simplest heuristics. As soon as we decided
2066 * that something is lost, we decide that _all_ not SACKed
2067 * packets until the most forward SACK are lost. I.e.
2068 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2069 * It is absolutely correct estimate, if network does not reorder
2070 * packets. And it loses any connection to reality when reordering
2071 * takes place. We use FACK by default until reordering
2072 * is suspected on the path to this destination.
2074 * NewReno: when Recovery is entered, we assume that one segment
2075 * is lost (classic Reno). While we are in Recovery and
2076 * a partial ACK arrives, we assume that one more packet
2077 * is lost (NewReno). This heuristics are the same in NewReno
2080 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2081 * deflation etc. CWND is real congestion window, never inflated, changes
2082 * only according to classic VJ rules.
2084 * Really tricky (and requiring careful tuning) part of algorithm
2085 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2086 * The first determines the moment _when_ we should reduce CWND and,
2087 * hence, slow down forward transmission. In fact, it determines the moment
2088 * when we decide that hole is caused by loss, rather than by a reorder.
2090 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2091 * holes, caused by lost packets.
2093 * And the most logically complicated part of algorithm is undo
2094 * heuristics. We detect false retransmits due to both too early
2095 * fast retransmit (reordering) and underestimated RTO, analyzing
2096 * timestamps and D-SACKs. When we detect that some segments were
2097 * retransmitted by mistake and CWND reduction was wrong, we undo
2098 * window reduction and abort recovery phase. This logic is hidden
2099 * inside several functions named tcp_try_undo_<something>.
2102 /* This function decides, when we should leave Disordered state
2103 * and enter Recovery phase, reducing congestion window.
2105 * Main question: may we further continue forward transmission
2106 * with the same cwnd?
2108 static bool tcp_time_to_recover(struct sock
*sk
, int flag
)
2110 struct tcp_sock
*tp
= tcp_sk(sk
);
2113 /* Trick#1: The loss is proven. */
2117 /* Not-A-Trick#2 : Classic rule... */
2118 if (tcp_dupack_heuristics(tp
) > tp
->reordering
)
2121 /* Trick#4: It is still not OK... But will it be useful to delay
2124 packets_out
= tp
->packets_out
;
2125 if (packets_out
<= tp
->reordering
&&
2126 tp
->sacked_out
>= max_t(__u32
, packets_out
/2, sysctl_tcp_reordering
) &&
2127 !tcp_may_send_now(sk
)) {
2128 /* We have nothing to send. This connection is limited
2129 * either by receiver window or by application.
2134 /* If a thin stream is detected, retransmit after first
2135 * received dupack. Employ only if SACK is supported in order
2136 * to avoid possible corner-case series of spurious retransmissions
2137 * Use only if there are no unsent data.
2139 if ((tp
->thin_dupack
|| sysctl_tcp_thin_dupack
) &&
2140 tcp_stream_is_thin(tp
) && tcp_dupack_heuristics(tp
) > 1 &&
2141 tcp_is_sack(tp
) && !tcp_send_head(sk
))
2144 /* Trick#6: TCP early retransmit, per RFC5827. To avoid spurious
2145 * retransmissions due to small network reorderings, we implement
2146 * Mitigation A.3 in the RFC and delay the retransmission for a short
2147 * interval if appropriate.
2149 if (tp
->do_early_retrans
&& !tp
->retrans_out
&& tp
->sacked_out
&&
2150 (tp
->packets_out
>= (tp
->sacked_out
+ 1) && tp
->packets_out
< 4) &&
2151 !tcp_may_send_now(sk
))
2152 return !tcp_pause_early_retransmit(sk
, flag
);
2157 /* Detect loss in event "A" above by marking head of queue up as lost.
2158 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2159 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2160 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2161 * the maximum SACKed segments to pass before reaching this limit.
2163 static void tcp_mark_head_lost(struct sock
*sk
, int packets
, int mark_head
)
2165 struct tcp_sock
*tp
= tcp_sk(sk
);
2166 struct sk_buff
*skb
;
2167 int cnt
, oldcnt
, lost
;
2169 /* Use SACK to deduce losses of new sequences sent during recovery */
2170 const u32 loss_high
= tcp_is_sack(tp
) ? tp
->snd_nxt
: tp
->high_seq
;
2172 WARN_ON(packets
> tp
->packets_out
);
2173 if (tp
->lost_skb_hint
) {
2174 skb
= tp
->lost_skb_hint
;
2175 cnt
= tp
->lost_cnt_hint
;
2176 /* Head already handled? */
2177 if (mark_head
&& skb
!= tcp_write_queue_head(sk
))
2180 skb
= tcp_write_queue_head(sk
);
2184 tcp_for_write_queue_from(skb
, sk
) {
2185 if (skb
== tcp_send_head(sk
))
2187 /* TODO: do this better */
2188 /* this is not the most efficient way to do this... */
2189 tp
->lost_skb_hint
= skb
;
2190 tp
->lost_cnt_hint
= cnt
;
2192 if (after(TCP_SKB_CB(skb
)->end_seq
, loss_high
))
2196 if (tcp_is_fack(tp
) || tcp_is_reno(tp
) ||
2197 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2198 cnt
+= tcp_skb_pcount(skb
);
2200 if (cnt
> packets
) {
2201 if ((tcp_is_sack(tp
) && !tcp_is_fack(tp
)) ||
2202 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
) ||
2203 (oldcnt
>= packets
))
2206 mss
= tcp_skb_mss(skb
);
2207 /* If needed, chop off the prefix to mark as lost. */
2208 lost
= (packets
- oldcnt
) * mss
;
2209 if (lost
< skb
->len
&&
2210 tcp_fragment(sk
, skb
, lost
, mss
, GFP_ATOMIC
) < 0)
2215 tcp_skb_mark_lost(tp
, skb
);
2220 tcp_verify_left_out(tp
);
2223 /* Account newly detected lost packet(s) */
2225 static void tcp_update_scoreboard(struct sock
*sk
, int fast_rexmit
)
2227 struct tcp_sock
*tp
= tcp_sk(sk
);
2229 if (tcp_is_reno(tp
)) {
2230 tcp_mark_head_lost(sk
, 1, 1);
2231 } else if (tcp_is_fack(tp
)) {
2232 int lost
= tp
->fackets_out
- tp
->reordering
;
2235 tcp_mark_head_lost(sk
, lost
, 0);
2237 int sacked_upto
= tp
->sacked_out
- tp
->reordering
;
2238 if (sacked_upto
>= 0)
2239 tcp_mark_head_lost(sk
, sacked_upto
, 0);
2240 else if (fast_rexmit
)
2241 tcp_mark_head_lost(sk
, 1, 1);
2245 /* CWND moderation, preventing bursts due to too big ACKs
2246 * in dubious situations.
2248 static inline void tcp_moderate_cwnd(struct tcp_sock
*tp
)
2250 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
2251 tcp_packets_in_flight(tp
) + tcp_max_burst(tp
));
2252 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2255 static bool tcp_tsopt_ecr_before(const struct tcp_sock
*tp
, u32 when
)
2257 return tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2258 before(tp
->rx_opt
.rcv_tsecr
, when
);
2261 /* skb is spurious retransmitted if the returned timestamp echo
2262 * reply is prior to the skb transmission time
2264 static bool tcp_skb_spurious_retrans(const struct tcp_sock
*tp
,
2265 const struct sk_buff
*skb
)
2267 return (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
) &&
2268 tcp_tsopt_ecr_before(tp
, tcp_skb_timestamp(skb
));
2271 /* Nothing was retransmitted or returned timestamp is less
2272 * than timestamp of the first retransmission.
2274 static inline bool tcp_packet_delayed(const struct tcp_sock
*tp
)
2276 return !tp
->retrans_stamp
||
2277 tcp_tsopt_ecr_before(tp
, tp
->retrans_stamp
);
2280 /* Undo procedures. */
2282 /* We can clear retrans_stamp when there are no retransmissions in the
2283 * window. It would seem that it is trivially available for us in
2284 * tp->retrans_out, however, that kind of assumptions doesn't consider
2285 * what will happen if errors occur when sending retransmission for the
2286 * second time. ...It could the that such segment has only
2287 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2288 * the head skb is enough except for some reneging corner cases that
2289 * are not worth the effort.
2291 * Main reason for all this complexity is the fact that connection dying
2292 * time now depends on the validity of the retrans_stamp, in particular,
2293 * that successive retransmissions of a segment must not advance
2294 * retrans_stamp under any conditions.
2296 static bool tcp_any_retrans_done(const struct sock
*sk
)
2298 const struct tcp_sock
*tp
= tcp_sk(sk
);
2299 struct sk_buff
*skb
;
2301 if (tp
->retrans_out
)
2304 skb
= tcp_write_queue_head(sk
);
2305 if (unlikely(skb
&& TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
))
2311 #if FASTRETRANS_DEBUG > 1
2312 static void DBGUNDO(struct sock
*sk
, const char *msg
)
2314 struct tcp_sock
*tp
= tcp_sk(sk
);
2315 struct inet_sock
*inet
= inet_sk(sk
);
2317 if (sk
->sk_family
== AF_INET
) {
2318 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2320 &inet
->inet_daddr
, ntohs(inet
->inet_dport
),
2321 tp
->snd_cwnd
, tcp_left_out(tp
),
2322 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2325 #if IS_ENABLED(CONFIG_IPV6)
2326 else if (sk
->sk_family
== AF_INET6
) {
2327 struct ipv6_pinfo
*np
= inet6_sk(sk
);
2328 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2330 &np
->daddr
, ntohs(inet
->inet_dport
),
2331 tp
->snd_cwnd
, tcp_left_out(tp
),
2332 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2338 #define DBGUNDO(x...) do { } while (0)
2341 static void tcp_undo_cwnd_reduction(struct sock
*sk
, bool unmark_loss
)
2343 struct tcp_sock
*tp
= tcp_sk(sk
);
2346 struct sk_buff
*skb
;
2348 tcp_for_write_queue(skb
, sk
) {
2349 if (skb
== tcp_send_head(sk
))
2351 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2354 tcp_clear_all_retrans_hints(tp
);
2357 if (tp
->prior_ssthresh
) {
2358 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2360 if (icsk
->icsk_ca_ops
->undo_cwnd
)
2361 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
2363 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
<< 1);
2365 if (tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
2366 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
2367 tcp_ecn_withdraw_cwr(tp
);
2370 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2371 tp
->undo_marker
= 0;
2374 static inline bool tcp_may_undo(const struct tcp_sock
*tp
)
2376 return tp
->undo_marker
&& (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
2379 /* People celebrate: "We love our President!" */
2380 static bool tcp_try_undo_recovery(struct sock
*sk
)
2382 struct tcp_sock
*tp
= tcp_sk(sk
);
2384 if (tcp_may_undo(tp
)) {
2387 /* Happy end! We did not retransmit anything
2388 * or our original transmission succeeded.
2390 DBGUNDO(sk
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
2391 tcp_undo_cwnd_reduction(sk
, false);
2392 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
2393 mib_idx
= LINUX_MIB_TCPLOSSUNDO
;
2395 mib_idx
= LINUX_MIB_TCPFULLUNDO
;
2397 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2399 if (tp
->snd_una
== tp
->high_seq
&& tcp_is_reno(tp
)) {
2400 /* Hold old state until something *above* high_seq
2401 * is ACKed. For Reno it is MUST to prevent false
2402 * fast retransmits (RFC2582). SACK TCP is safe. */
2403 tcp_moderate_cwnd(tp
);
2404 if (!tcp_any_retrans_done(sk
))
2405 tp
->retrans_stamp
= 0;
2408 tcp_set_ca_state(sk
, TCP_CA_Open
);
2412 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2413 static bool tcp_try_undo_dsack(struct sock
*sk
)
2415 struct tcp_sock
*tp
= tcp_sk(sk
);
2417 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
2418 DBGUNDO(sk
, "D-SACK");
2419 tcp_undo_cwnd_reduction(sk
, false);
2420 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPDSACKUNDO
);
2426 /* Undo during loss recovery after partial ACK or using F-RTO. */
2427 static bool tcp_try_undo_loss(struct sock
*sk
, bool frto_undo
)
2429 struct tcp_sock
*tp
= tcp_sk(sk
);
2431 if (frto_undo
|| tcp_may_undo(tp
)) {
2432 tcp_undo_cwnd_reduction(sk
, true);
2434 DBGUNDO(sk
, "partial loss");
2435 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSSUNDO
);
2437 NET_INC_STATS_BH(sock_net(sk
),
2438 LINUX_MIB_TCPSPURIOUSRTOS
);
2439 inet_csk(sk
)->icsk_retransmits
= 0;
2440 if (frto_undo
|| tcp_is_sack(tp
))
2441 tcp_set_ca_state(sk
, TCP_CA_Open
);
2447 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2448 * It computes the number of packets to send (sndcnt) based on packets newly
2450 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2451 * cwnd reductions across a full RTT.
2452 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2453 * But when the retransmits are acked without further losses, PRR
2454 * slow starts cwnd up to ssthresh to speed up the recovery.
2456 static void tcp_init_cwnd_reduction(struct sock
*sk
)
2458 struct tcp_sock
*tp
= tcp_sk(sk
);
2460 tp
->high_seq
= tp
->snd_nxt
;
2461 tp
->tlp_high_seq
= 0;
2462 tp
->snd_cwnd_cnt
= 0;
2463 tp
->prior_cwnd
= tp
->snd_cwnd
;
2464 tp
->prr_delivered
= 0;
2466 tp
->snd_ssthresh
= inet_csk(sk
)->icsk_ca_ops
->ssthresh(sk
);
2467 tcp_ecn_queue_cwr(tp
);
2470 static void tcp_cwnd_reduction(struct sock
*sk
, const int prior_unsacked
,
2471 int fast_rexmit
, int flag
)
2473 struct tcp_sock
*tp
= tcp_sk(sk
);
2475 int delta
= tp
->snd_ssthresh
- tcp_packets_in_flight(tp
);
2476 int newly_acked_sacked
= prior_unsacked
-
2477 (tp
->packets_out
- tp
->sacked_out
);
2479 if (newly_acked_sacked
<= 0 || WARN_ON_ONCE(!tp
->prior_cwnd
))
2482 tp
->prr_delivered
+= newly_acked_sacked
;
2484 u64 dividend
= (u64
)tp
->snd_ssthresh
* tp
->prr_delivered
+
2486 sndcnt
= div_u64(dividend
, tp
->prior_cwnd
) - tp
->prr_out
;
2487 } else if ((flag
& FLAG_RETRANS_DATA_ACKED
) &&
2488 !(flag
& FLAG_LOST_RETRANS
)) {
2489 sndcnt
= min_t(int, delta
,
2490 max_t(int, tp
->prr_delivered
- tp
->prr_out
,
2491 newly_acked_sacked
) + 1);
2493 sndcnt
= min(delta
, newly_acked_sacked
);
2495 sndcnt
= max(sndcnt
, (fast_rexmit
? 1 : 0));
2496 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + sndcnt
;
2499 static inline void tcp_end_cwnd_reduction(struct sock
*sk
)
2501 struct tcp_sock
*tp
= tcp_sk(sk
);
2503 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2504 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_CWR
||
2505 (tp
->undo_marker
&& tp
->snd_ssthresh
< TCP_INFINITE_SSTHRESH
)) {
2506 tp
->snd_cwnd
= tp
->snd_ssthresh
;
2507 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2509 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
2512 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2513 void tcp_enter_cwr(struct sock
*sk
)
2515 struct tcp_sock
*tp
= tcp_sk(sk
);
2517 tp
->prior_ssthresh
= 0;
2518 if (inet_csk(sk
)->icsk_ca_state
< TCP_CA_CWR
) {
2519 tp
->undo_marker
= 0;
2520 tcp_init_cwnd_reduction(sk
);
2521 tcp_set_ca_state(sk
, TCP_CA_CWR
);
2524 EXPORT_SYMBOL(tcp_enter_cwr
);
2526 static void tcp_try_keep_open(struct sock
*sk
)
2528 struct tcp_sock
*tp
= tcp_sk(sk
);
2529 int state
= TCP_CA_Open
;
2531 if (tcp_left_out(tp
) || tcp_any_retrans_done(sk
))
2532 state
= TCP_CA_Disorder
;
2534 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
2535 tcp_set_ca_state(sk
, state
);
2536 tp
->high_seq
= tp
->snd_nxt
;
2540 static void tcp_try_to_open(struct sock
*sk
, int flag
, const int prior_unsacked
)
2542 struct tcp_sock
*tp
= tcp_sk(sk
);
2544 tcp_verify_left_out(tp
);
2546 if (!tcp_any_retrans_done(sk
))
2547 tp
->retrans_stamp
= 0;
2549 if (flag
& FLAG_ECE
)
2552 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
2553 tcp_try_keep_open(sk
);
2555 tcp_cwnd_reduction(sk
, prior_unsacked
, 0, flag
);
2559 static void tcp_mtup_probe_failed(struct sock
*sk
)
2561 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2563 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
2564 icsk
->icsk_mtup
.probe_size
= 0;
2565 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPMTUPFAIL
);
2568 static void tcp_mtup_probe_success(struct sock
*sk
)
2570 struct tcp_sock
*tp
= tcp_sk(sk
);
2571 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2573 /* FIXME: breaks with very large cwnd */
2574 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2575 tp
->snd_cwnd
= tp
->snd_cwnd
*
2576 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
2577 icsk
->icsk_mtup
.probe_size
;
2578 tp
->snd_cwnd_cnt
= 0;
2579 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2580 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2582 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
2583 icsk
->icsk_mtup
.probe_size
= 0;
2584 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
2585 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPMTUPSUCCESS
);
2588 /* Do a simple retransmit without using the backoff mechanisms in
2589 * tcp_timer. This is used for path mtu discovery.
2590 * The socket is already locked here.
2592 void tcp_simple_retransmit(struct sock
*sk
)
2594 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2595 struct tcp_sock
*tp
= tcp_sk(sk
);
2596 struct sk_buff
*skb
;
2597 unsigned int mss
= tcp_current_mss(sk
);
2598 u32 prior_lost
= tp
->lost_out
;
2600 tcp_for_write_queue(skb
, sk
) {
2601 if (skb
== tcp_send_head(sk
))
2603 if (tcp_skb_seglen(skb
) > mss
&&
2604 !(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
2605 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2606 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
2607 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2609 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
2613 tcp_clear_retrans_hints_partial(tp
);
2615 if (prior_lost
== tp
->lost_out
)
2618 if (tcp_is_reno(tp
))
2619 tcp_limit_reno_sacked(tp
);
2621 tcp_verify_left_out(tp
);
2623 /* Don't muck with the congestion window here.
2624 * Reason is that we do not increase amount of _data_
2625 * in network, but units changed and effective
2626 * cwnd/ssthresh really reduced now.
2628 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
2629 tp
->high_seq
= tp
->snd_nxt
;
2630 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2631 tp
->prior_ssthresh
= 0;
2632 tp
->undo_marker
= 0;
2633 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2635 tcp_xmit_retransmit_queue(sk
);
2637 EXPORT_SYMBOL(tcp_simple_retransmit
);
2639 static void tcp_enter_recovery(struct sock
*sk
, bool ece_ack
)
2641 struct tcp_sock
*tp
= tcp_sk(sk
);
2644 if (tcp_is_reno(tp
))
2645 mib_idx
= LINUX_MIB_TCPRENORECOVERY
;
2647 mib_idx
= LINUX_MIB_TCPSACKRECOVERY
;
2649 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2651 tp
->prior_ssthresh
= 0;
2654 if (!tcp_in_cwnd_reduction(sk
)) {
2656 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2657 tcp_init_cwnd_reduction(sk
);
2659 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
2662 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2663 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2665 static void tcp_process_loss(struct sock
*sk
, int flag
, bool is_dupack
)
2667 struct tcp_sock
*tp
= tcp_sk(sk
);
2668 bool recovered
= !before(tp
->snd_una
, tp
->high_seq
);
2670 if ((flag
& FLAG_SND_UNA_ADVANCED
) &&
2671 tcp_try_undo_loss(sk
, false))
2674 if (tp
->frto
) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2675 /* Step 3.b. A timeout is spurious if not all data are
2676 * lost, i.e., never-retransmitted data are (s)acked.
2678 if ((flag
& FLAG_ORIG_SACK_ACKED
) &&
2679 tcp_try_undo_loss(sk
, true))
2682 if (after(tp
->snd_nxt
, tp
->high_seq
)) {
2683 if (flag
& FLAG_DATA_SACKED
|| is_dupack
)
2684 tp
->frto
= 0; /* Step 3.a. loss was real */
2685 } else if (flag
& FLAG_SND_UNA_ADVANCED
&& !recovered
) {
2686 tp
->high_seq
= tp
->snd_nxt
;
2687 __tcp_push_pending_frames(sk
, tcp_current_mss(sk
),
2689 if (after(tp
->snd_nxt
, tp
->high_seq
))
2690 return; /* Step 2.b */
2696 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2697 tcp_try_undo_recovery(sk
);
2700 if (tcp_is_reno(tp
)) {
2701 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2702 * delivered. Lower inflight to clock out (re)tranmissions.
2704 if (after(tp
->snd_nxt
, tp
->high_seq
) && is_dupack
)
2705 tcp_add_reno_sack(sk
);
2706 else if (flag
& FLAG_SND_UNA_ADVANCED
)
2707 tcp_reset_reno_sack(tp
);
2709 tcp_xmit_retransmit_queue(sk
);
2712 /* Undo during fast recovery after partial ACK. */
2713 static bool tcp_try_undo_partial(struct sock
*sk
, const int acked
,
2714 const int prior_unsacked
, int flag
)
2716 struct tcp_sock
*tp
= tcp_sk(sk
);
2718 if (tp
->undo_marker
&& tcp_packet_delayed(tp
)) {
2719 /* Plain luck! Hole if filled with delayed
2720 * packet, rather than with a retransmit.
2722 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
2724 /* We are getting evidence that the reordering degree is higher
2725 * than we realized. If there are no retransmits out then we
2726 * can undo. Otherwise we clock out new packets but do not
2727 * mark more packets lost or retransmit more.
2729 if (tp
->retrans_out
) {
2730 tcp_cwnd_reduction(sk
, prior_unsacked
, 0, flag
);
2734 if (!tcp_any_retrans_done(sk
))
2735 tp
->retrans_stamp
= 0;
2737 DBGUNDO(sk
, "partial recovery");
2738 tcp_undo_cwnd_reduction(sk
, true);
2739 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPARTIALUNDO
);
2740 tcp_try_keep_open(sk
);
2746 /* Process an event, which can update packets-in-flight not trivially.
2747 * Main goal of this function is to calculate new estimate for left_out,
2748 * taking into account both packets sitting in receiver's buffer and
2749 * packets lost by network.
2751 * Besides that it does CWND reduction, when packet loss is detected
2752 * and changes state of machine.
2754 * It does _not_ decide what to send, it is made in function
2755 * tcp_xmit_retransmit_queue().
2757 static void tcp_fastretrans_alert(struct sock
*sk
, const int acked
,
2758 const int prior_unsacked
,
2759 bool is_dupack
, int flag
)
2761 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2762 struct tcp_sock
*tp
= tcp_sk(sk
);
2763 bool do_lost
= is_dupack
|| ((flag
& FLAG_DATA_SACKED
) &&
2764 (tcp_fackets_out(tp
) > tp
->reordering
));
2765 int fast_rexmit
= 0;
2767 if (WARN_ON(!tp
->packets_out
&& tp
->sacked_out
))
2769 if (WARN_ON(!tp
->sacked_out
&& tp
->fackets_out
))
2770 tp
->fackets_out
= 0;
2772 /* Now state machine starts.
2773 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2774 if (flag
& FLAG_ECE
)
2775 tp
->prior_ssthresh
= 0;
2777 /* B. In all the states check for reneging SACKs. */
2778 if (tcp_check_sack_reneging(sk
, flag
))
2781 /* C. Check consistency of the current state. */
2782 tcp_verify_left_out(tp
);
2784 /* D. Check state exit conditions. State can be terminated
2785 * when high_seq is ACKed. */
2786 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
2787 WARN_ON(tp
->retrans_out
!= 0);
2788 tp
->retrans_stamp
= 0;
2789 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
2790 switch (icsk
->icsk_ca_state
) {
2792 /* CWR is to be held something *above* high_seq
2793 * is ACKed for CWR bit to reach receiver. */
2794 if (tp
->snd_una
!= tp
->high_seq
) {
2795 tcp_end_cwnd_reduction(sk
);
2796 tcp_set_ca_state(sk
, TCP_CA_Open
);
2800 case TCP_CA_Recovery
:
2801 if (tcp_is_reno(tp
))
2802 tcp_reset_reno_sack(tp
);
2803 if (tcp_try_undo_recovery(sk
))
2805 tcp_end_cwnd_reduction(sk
);
2810 /* Use RACK to detect loss */
2811 if (sysctl_tcp_recovery
& TCP_RACK_LOST_RETRANS
&&
2812 tcp_rack_mark_lost(sk
))
2813 flag
|= FLAG_LOST_RETRANS
;
2815 /* E. Process state. */
2816 switch (icsk
->icsk_ca_state
) {
2817 case TCP_CA_Recovery
:
2818 if (!(flag
& FLAG_SND_UNA_ADVANCED
)) {
2819 if (tcp_is_reno(tp
) && is_dupack
)
2820 tcp_add_reno_sack(sk
);
2822 if (tcp_try_undo_partial(sk
, acked
, prior_unsacked
, flag
))
2824 /* Partial ACK arrived. Force fast retransmit. */
2825 do_lost
= tcp_is_reno(tp
) ||
2826 tcp_fackets_out(tp
) > tp
->reordering
;
2828 if (tcp_try_undo_dsack(sk
)) {
2829 tcp_try_keep_open(sk
);
2834 tcp_process_loss(sk
, flag
, is_dupack
);
2835 if (icsk
->icsk_ca_state
!= TCP_CA_Open
&&
2836 !(flag
& FLAG_LOST_RETRANS
))
2838 /* Change state if cwnd is undone or retransmits are lost */
2840 if (tcp_is_reno(tp
)) {
2841 if (flag
& FLAG_SND_UNA_ADVANCED
)
2842 tcp_reset_reno_sack(tp
);
2844 tcp_add_reno_sack(sk
);
2847 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
)
2848 tcp_try_undo_dsack(sk
);
2850 if (!tcp_time_to_recover(sk
, flag
)) {
2851 tcp_try_to_open(sk
, flag
, prior_unsacked
);
2855 /* MTU probe failure: don't reduce cwnd */
2856 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
2857 icsk
->icsk_mtup
.probe_size
&&
2858 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
2859 tcp_mtup_probe_failed(sk
);
2860 /* Restores the reduction we did in tcp_mtup_probe() */
2862 tcp_simple_retransmit(sk
);
2866 /* Otherwise enter Recovery state */
2867 tcp_enter_recovery(sk
, (flag
& FLAG_ECE
));
2872 tcp_update_scoreboard(sk
, fast_rexmit
);
2873 tcp_cwnd_reduction(sk
, prior_unsacked
, fast_rexmit
, flag
);
2874 tcp_xmit_retransmit_queue(sk
);
2877 /* Kathleen Nichols' algorithm for tracking the minimum value of
2878 * a data stream over some fixed time interval. (E.g., the minimum
2879 * RTT over the past five minutes.) It uses constant space and constant
2880 * time per update yet almost always delivers the same minimum as an
2881 * implementation that has to keep all the data in the window.
2883 * The algorithm keeps track of the best, 2nd best & 3rd best min
2884 * values, maintaining an invariant that the measurement time of the
2885 * n'th best >= n-1'th best. It also makes sure that the three values
2886 * are widely separated in the time window since that bounds the worse
2887 * case error when that data is monotonically increasing over the window.
2889 * Upon getting a new min, we can forget everything earlier because it
2890 * has no value - the new min is <= everything else in the window by
2891 * definition and it's the most recent. So we restart fresh on every new min
2892 * and overwrites 2nd & 3rd choices. The same property holds for 2nd & 3rd
2895 static void tcp_update_rtt_min(struct sock
*sk
, u32 rtt_us
)
2897 const u32 now
= tcp_time_stamp
, wlen
= sysctl_tcp_min_rtt_wlen
* HZ
;
2898 struct rtt_meas
*m
= tcp_sk(sk
)->rtt_min
;
2899 struct rtt_meas rttm
= {
2900 .rtt
= likely(rtt_us
) ? rtt_us
: jiffies_to_usecs(1),
2905 /* Check if the new measurement updates the 1st, 2nd, or 3rd choices */
2906 if (unlikely(rttm
.rtt
<= m
[0].rtt
))
2907 m
[0] = m
[1] = m
[2] = rttm
;
2908 else if (rttm
.rtt
<= m
[1].rtt
)
2910 else if (rttm
.rtt
<= m
[2].rtt
)
2913 elapsed
= now
- m
[0].ts
;
2914 if (unlikely(elapsed
> wlen
)) {
2915 /* Passed entire window without a new min so make 2nd choice
2916 * the new min & 3rd choice the new 2nd. So forth and so on.
2921 if (now
- m
[0].ts
> wlen
) {
2924 if (now
- m
[0].ts
> wlen
)
2927 } else if (m
[1].ts
== m
[0].ts
&& elapsed
> wlen
/ 4) {
2928 /* Passed a quarter of the window without a new min so
2929 * take 2nd choice from the 2nd quarter of the window.
2932 } else if (m
[2].ts
== m
[1].ts
&& elapsed
> wlen
/ 2) {
2933 /* Passed half the window without a new min so take the 3rd
2934 * choice from the last half of the window.
2940 static inline bool tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
2941 long seq_rtt_us
, long sack_rtt_us
,
2944 const struct tcp_sock
*tp
= tcp_sk(sk
);
2946 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2947 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2948 * Karn's algorithm forbids taking RTT if some retransmitted data
2949 * is acked (RFC6298).
2952 seq_rtt_us
= sack_rtt_us
;
2954 /* RTTM Rule: A TSecr value received in a segment is used to
2955 * update the averaged RTT measurement only if the segment
2956 * acknowledges some new data, i.e., only if it advances the
2957 * left edge of the send window.
2958 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2960 if (seq_rtt_us
< 0 && tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2962 seq_rtt_us
= ca_rtt_us
= jiffies_to_usecs(tcp_time_stamp
-
2963 tp
->rx_opt
.rcv_tsecr
);
2967 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
2968 * always taken together with ACK, SACK, or TS-opts. Any negative
2969 * values will be skipped with the seq_rtt_us < 0 check above.
2971 tcp_update_rtt_min(sk
, ca_rtt_us
);
2972 tcp_rtt_estimator(sk
, seq_rtt_us
);
2975 /* RFC6298: only reset backoff on valid RTT measurement. */
2976 inet_csk(sk
)->icsk_backoff
= 0;
2980 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2981 void tcp_synack_rtt_meas(struct sock
*sk
, struct request_sock
*req
)
2985 if (req
&& !req
->num_retrans
&& tcp_rsk(req
)->snt_synack
.v64
) {
2986 struct skb_mstamp now
;
2988 skb_mstamp_get(&now
);
2989 rtt_us
= skb_mstamp_us_delta(&now
, &tcp_rsk(req
)->snt_synack
);
2992 tcp_ack_update_rtt(sk
, FLAG_SYN_ACKED
, rtt_us
, -1L, rtt_us
);
2996 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
, u32 acked
)
2998 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3000 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, acked
);
3001 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_time_stamp
;
3004 /* Restart timer after forward progress on connection.
3005 * RFC2988 recommends to restart timer to now+rto.
3007 void tcp_rearm_rto(struct sock
*sk
)
3009 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3010 struct tcp_sock
*tp
= tcp_sk(sk
);
3012 /* If the retrans timer is currently being used by Fast Open
3013 * for SYN-ACK retrans purpose, stay put.
3015 if (tp
->fastopen_rsk
)
3018 if (!tp
->packets_out
) {
3019 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
3021 u32 rto
= inet_csk(sk
)->icsk_rto
;
3022 /* Offset the time elapsed after installing regular RTO */
3023 if (icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
||
3024 icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
) {
3025 struct sk_buff
*skb
= tcp_write_queue_head(sk
);
3026 const u32 rto_time_stamp
=
3027 tcp_skb_timestamp(skb
) + rto
;
3028 s32 delta
= (s32
)(rto_time_stamp
- tcp_time_stamp
);
3029 /* delta may not be positive if the socket is locked
3030 * when the retrans timer fires and is rescheduled.
3035 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
, rto
,
3040 /* This function is called when the delayed ER timer fires. TCP enters
3041 * fast recovery and performs fast-retransmit.
3043 void tcp_resume_early_retransmit(struct sock
*sk
)
3045 struct tcp_sock
*tp
= tcp_sk(sk
);
3049 /* Stop if ER is disabled after the delayed ER timer is scheduled */
3050 if (!tp
->do_early_retrans
)
3053 tcp_enter_recovery(sk
, false);
3054 tcp_update_scoreboard(sk
, 1);
3055 tcp_xmit_retransmit_queue(sk
);
3058 /* If we get here, the whole TSO packet has not been acked. */
3059 static u32
tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
)
3061 struct tcp_sock
*tp
= tcp_sk(sk
);
3064 BUG_ON(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
));
3066 packets_acked
= tcp_skb_pcount(skb
);
3067 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
3069 packets_acked
-= tcp_skb_pcount(skb
);
3071 if (packets_acked
) {
3072 BUG_ON(tcp_skb_pcount(skb
) == 0);
3073 BUG_ON(!before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
));
3076 return packets_acked
;
3079 static void tcp_ack_tstamp(struct sock
*sk
, struct sk_buff
*skb
,
3082 const struct skb_shared_info
*shinfo
;
3084 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3085 if (likely(!(sk
->sk_tsflags
& SOF_TIMESTAMPING_TX_ACK
)))
3088 shinfo
= skb_shinfo(skb
);
3089 if ((shinfo
->tx_flags
& SKBTX_ACK_TSTAMP
) &&
3090 between(shinfo
->tskey
, prior_snd_una
, tcp_sk(sk
)->snd_una
- 1))
3091 __skb_tstamp_tx(skb
, NULL
, sk
, SCM_TSTAMP_ACK
);
3094 /* Remove acknowledged frames from the retransmission queue. If our packet
3095 * is before the ack sequence we can discard it as it's confirmed to have
3096 * arrived at the other end.
3098 static int tcp_clean_rtx_queue(struct sock
*sk
, int prior_fackets
,
3100 struct tcp_sacktag_state
*sack
)
3102 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3103 struct skb_mstamp first_ackt
, last_ackt
, now
;
3104 struct tcp_sock
*tp
= tcp_sk(sk
);
3105 u32 prior_sacked
= tp
->sacked_out
;
3106 u32 reord
= tp
->packets_out
;
3107 bool fully_acked
= true;
3108 long sack_rtt_us
= -1L;
3109 long seq_rtt_us
= -1L;
3110 long ca_rtt_us
= -1L;
3111 struct sk_buff
*skb
;
3118 while ((skb
= tcp_write_queue_head(sk
)) && skb
!= tcp_send_head(sk
)) {
3119 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
3120 u8 sacked
= scb
->sacked
;
3123 tcp_ack_tstamp(sk
, skb
, prior_snd_una
);
3125 /* Determine how many packets and what bytes were acked, tso and else */
3126 if (after(scb
->end_seq
, tp
->snd_una
)) {
3127 if (tcp_skb_pcount(skb
) == 1 ||
3128 !after(tp
->snd_una
, scb
->seq
))
3131 acked_pcount
= tcp_tso_acked(sk
, skb
);
3135 fully_acked
= false;
3137 /* Speedup tcp_unlink_write_queue() and next loop */
3138 prefetchw(skb
->next
);
3139 acked_pcount
= tcp_skb_pcount(skb
);
3142 if (unlikely(sacked
& TCPCB_RETRANS
)) {
3143 if (sacked
& TCPCB_SACKED_RETRANS
)
3144 tp
->retrans_out
-= acked_pcount
;
3145 flag
|= FLAG_RETRANS_DATA_ACKED
;
3146 } else if (!(sacked
& TCPCB_SACKED_ACKED
)) {
3147 last_ackt
= skb
->skb_mstamp
;
3148 WARN_ON_ONCE(last_ackt
.v64
== 0);
3149 if (!first_ackt
.v64
)
3150 first_ackt
= last_ackt
;
3152 reord
= min(pkts_acked
, reord
);
3153 if (!after(scb
->end_seq
, tp
->high_seq
))
3154 flag
|= FLAG_ORIG_SACK_ACKED
;
3157 if (sacked
& TCPCB_SACKED_ACKED
)
3158 tp
->sacked_out
-= acked_pcount
;
3159 else if (tcp_is_sack(tp
) && !tcp_skb_spurious_retrans(tp
, skb
))
3160 tcp_rack_advance(tp
, &skb
->skb_mstamp
, sacked
);
3161 if (sacked
& TCPCB_LOST
)
3162 tp
->lost_out
-= acked_pcount
;
3164 tp
->packets_out
-= acked_pcount
;
3165 pkts_acked
+= acked_pcount
;
3167 /* Initial outgoing SYN's get put onto the write_queue
3168 * just like anything else we transmit. It is not
3169 * true data, and if we misinform our callers that
3170 * this ACK acks real data, we will erroneously exit
3171 * connection startup slow start one packet too
3172 * quickly. This is severely frowned upon behavior.
3174 if (likely(!(scb
->tcp_flags
& TCPHDR_SYN
))) {
3175 flag
|= FLAG_DATA_ACKED
;
3177 flag
|= FLAG_SYN_ACKED
;
3178 tp
->retrans_stamp
= 0;
3184 tcp_unlink_write_queue(skb
, sk
);
3185 sk_wmem_free_skb(sk
, skb
);
3186 if (unlikely(skb
== tp
->retransmit_skb_hint
))
3187 tp
->retransmit_skb_hint
= NULL
;
3188 if (unlikely(skb
== tp
->lost_skb_hint
))
3189 tp
->lost_skb_hint
= NULL
;
3192 if (likely(between(tp
->snd_up
, prior_snd_una
, tp
->snd_una
)))
3193 tp
->snd_up
= tp
->snd_una
;
3195 if (skb
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
3196 flag
|= FLAG_SACK_RENEGING
;
3198 skb_mstamp_get(&now
);
3199 if (likely(first_ackt
.v64
) && !(flag
& FLAG_RETRANS_DATA_ACKED
)) {
3200 seq_rtt_us
= skb_mstamp_us_delta(&now
, &first_ackt
);
3201 ca_rtt_us
= skb_mstamp_us_delta(&now
, &last_ackt
);
3203 if (sack
->first_sackt
.v64
) {
3204 sack_rtt_us
= skb_mstamp_us_delta(&now
, &sack
->first_sackt
);
3205 ca_rtt_us
= skb_mstamp_us_delta(&now
, &sack
->last_sackt
);
3208 rtt_update
= tcp_ack_update_rtt(sk
, flag
, seq_rtt_us
, sack_rtt_us
,
3211 if (flag
& FLAG_ACKED
) {
3213 if (unlikely(icsk
->icsk_mtup
.probe_size
&&
3214 !after(tp
->mtu_probe
.probe_seq_end
, tp
->snd_una
))) {
3215 tcp_mtup_probe_success(sk
);
3218 if (tcp_is_reno(tp
)) {
3219 tcp_remove_reno_sacks(sk
, pkts_acked
);
3223 /* Non-retransmitted hole got filled? That's reordering */
3224 if (reord
< prior_fackets
)
3225 tcp_update_reordering(sk
, tp
->fackets_out
- reord
, 0);
3227 delta
= tcp_is_fack(tp
) ? pkts_acked
:
3228 prior_sacked
- tp
->sacked_out
;
3229 tp
->lost_cnt_hint
-= min(tp
->lost_cnt_hint
, delta
);
3232 tp
->fackets_out
-= min(pkts_acked
, tp
->fackets_out
);
3234 } else if (skb
&& rtt_update
&& sack_rtt_us
>= 0 &&
3235 sack_rtt_us
> skb_mstamp_us_delta(&now
, &skb
->skb_mstamp
)) {
3236 /* Do not re-arm RTO if the sack RTT is measured from data sent
3237 * after when the head was last (re)transmitted. Otherwise the
3238 * timeout may continue to extend in loss recovery.
3243 if (icsk
->icsk_ca_ops
->pkts_acked
)
3244 icsk
->icsk_ca_ops
->pkts_acked(sk
, pkts_acked
, ca_rtt_us
);
3246 #if FASTRETRANS_DEBUG > 0
3247 WARN_ON((int)tp
->sacked_out
< 0);
3248 WARN_ON((int)tp
->lost_out
< 0);
3249 WARN_ON((int)tp
->retrans_out
< 0);
3250 if (!tp
->packets_out
&& tcp_is_sack(tp
)) {
3251 icsk
= inet_csk(sk
);
3253 pr_debug("Leak l=%u %d\n",
3254 tp
->lost_out
, icsk
->icsk_ca_state
);
3257 if (tp
->sacked_out
) {
3258 pr_debug("Leak s=%u %d\n",
3259 tp
->sacked_out
, icsk
->icsk_ca_state
);
3262 if (tp
->retrans_out
) {
3263 pr_debug("Leak r=%u %d\n",
3264 tp
->retrans_out
, icsk
->icsk_ca_state
);
3265 tp
->retrans_out
= 0;
3272 static void tcp_ack_probe(struct sock
*sk
)
3274 const struct tcp_sock
*tp
= tcp_sk(sk
);
3275 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3277 /* Was it a usable window open? */
3279 if (!after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
, tcp_wnd_end(tp
))) {
3280 icsk
->icsk_backoff
= 0;
3281 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
3282 /* Socket must be waked up by subsequent tcp_data_snd_check().
3283 * This function is not for random using!
3286 unsigned long when
= tcp_probe0_when(sk
, TCP_RTO_MAX
);
3288 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3293 static inline bool tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
3295 return !(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
3296 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
;
3299 /* Decide wheather to run the increase function of congestion control. */
3300 static inline bool tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
3302 if (tcp_in_cwnd_reduction(sk
))
3305 /* If reordering is high then always grow cwnd whenever data is
3306 * delivered regardless of its ordering. Otherwise stay conservative
3307 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3308 * new SACK or ECE mark may first advance cwnd here and later reduce
3309 * cwnd in tcp_fastretrans_alert() based on more states.
3311 if (tcp_sk(sk
)->reordering
> sysctl_tcp_reordering
)
3312 return flag
& FLAG_FORWARD_PROGRESS
;
3314 return flag
& FLAG_DATA_ACKED
;
3317 /* Check that window update is acceptable.
3318 * The function assumes that snd_una<=ack<=snd_next.
3320 static inline bool tcp_may_update_window(const struct tcp_sock
*tp
,
3321 const u32 ack
, const u32 ack_seq
,
3324 return after(ack
, tp
->snd_una
) ||
3325 after(ack_seq
, tp
->snd_wl1
) ||
3326 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
);
3329 /* If we update tp->snd_una, also update tp->bytes_acked */
3330 static void tcp_snd_una_update(struct tcp_sock
*tp
, u32 ack
)
3332 u32 delta
= ack
- tp
->snd_una
;
3334 u64_stats_update_begin(&tp
->syncp
);
3335 tp
->bytes_acked
+= delta
;
3336 u64_stats_update_end(&tp
->syncp
);
3340 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3341 static void tcp_rcv_nxt_update(struct tcp_sock
*tp
, u32 seq
)
3343 u32 delta
= seq
- tp
->rcv_nxt
;
3345 u64_stats_update_begin(&tp
->syncp
);
3346 tp
->bytes_received
+= delta
;
3347 u64_stats_update_end(&tp
->syncp
);
3351 /* Update our send window.
3353 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3354 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3356 static int tcp_ack_update_window(struct sock
*sk
, const struct sk_buff
*skb
, u32 ack
,
3359 struct tcp_sock
*tp
= tcp_sk(sk
);
3361 u32 nwin
= ntohs(tcp_hdr(skb
)->window
);
3363 if (likely(!tcp_hdr(skb
)->syn
))
3364 nwin
<<= tp
->rx_opt
.snd_wscale
;
3366 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
3367 flag
|= FLAG_WIN_UPDATE
;
3368 tcp_update_wl(tp
, ack_seq
);
3370 if (tp
->snd_wnd
!= nwin
) {
3373 /* Note, it is the only place, where
3374 * fast path is recovered for sending TCP.
3377 tcp_fast_path_check(sk
);
3379 if (tcp_send_head(sk
))
3380 tcp_slow_start_after_idle_check(sk
);
3382 if (nwin
> tp
->max_window
) {
3383 tp
->max_window
= nwin
;
3384 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
3389 tcp_snd_una_update(tp
, ack
);
3394 /* Return true if we're currently rate-limiting out-of-window ACKs and
3395 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3396 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3397 * attacks that send repeated SYNs or ACKs for the same connection. To
3398 * do this, we do not send a duplicate SYNACK or ACK if the remote
3399 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3401 bool tcp_oow_rate_limited(struct net
*net
, const struct sk_buff
*skb
,
3402 int mib_idx
, u32
*last_oow_ack_time
)
3404 /* Data packets without SYNs are not likely part of an ACK loop. */
3405 if ((TCP_SKB_CB(skb
)->seq
!= TCP_SKB_CB(skb
)->end_seq
) &&
3407 goto not_rate_limited
;
3409 if (*last_oow_ack_time
) {
3410 s32 elapsed
= (s32
)(tcp_time_stamp
- *last_oow_ack_time
);
3412 if (0 <= elapsed
&& elapsed
< sysctl_tcp_invalid_ratelimit
) {
3413 NET_INC_STATS_BH(net
, mib_idx
);
3414 return true; /* rate-limited: don't send yet! */
3418 *last_oow_ack_time
= tcp_time_stamp
;
3421 return false; /* not rate-limited: go ahead, send dupack now! */
3424 /* RFC 5961 7 [ACK Throttling] */
3425 static void tcp_send_challenge_ack(struct sock
*sk
, const struct sk_buff
*skb
)
3427 /* unprotected vars, we dont care of overwrites */
3428 static u32 challenge_timestamp
;
3429 static unsigned int challenge_count
;
3430 struct tcp_sock
*tp
= tcp_sk(sk
);
3433 /* First check our per-socket dupack rate limit. */
3434 if (tcp_oow_rate_limited(sock_net(sk
), skb
,
3435 LINUX_MIB_TCPACKSKIPPEDCHALLENGE
,
3436 &tp
->last_oow_ack_time
))
3439 /* Then check the check host-wide RFC 5961 rate limit. */
3441 if (now
!= challenge_timestamp
) {
3442 challenge_timestamp
= now
;
3443 challenge_count
= 0;
3445 if (++challenge_count
<= sysctl_tcp_challenge_ack_limit
) {
3446 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPCHALLENGEACK
);
3451 static void tcp_store_ts_recent(struct tcp_sock
*tp
)
3453 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
3454 tp
->rx_opt
.ts_recent_stamp
= get_seconds();
3457 static void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
3459 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
3460 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3461 * extra check below makes sure this can only happen
3462 * for pure ACK frames. -DaveM
3464 * Not only, also it occurs for expired timestamps.
3467 if (tcp_paws_check(&tp
->rx_opt
, 0))
3468 tcp_store_ts_recent(tp
);
3472 /* This routine deals with acks during a TLP episode.
3473 * We mark the end of a TLP episode on receiving TLP dupack or when
3474 * ack is after tlp_high_seq.
3475 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3477 static void tcp_process_tlp_ack(struct sock
*sk
, u32 ack
, int flag
)
3479 struct tcp_sock
*tp
= tcp_sk(sk
);
3481 if (before(ack
, tp
->tlp_high_seq
))
3484 if (flag
& FLAG_DSACKING_ACK
) {
3485 /* This DSACK means original and TLP probe arrived; no loss */
3486 tp
->tlp_high_seq
= 0;
3487 } else if (after(ack
, tp
->tlp_high_seq
)) {
3488 /* ACK advances: there was a loss, so reduce cwnd. Reset
3489 * tlp_high_seq in tcp_init_cwnd_reduction()
3491 tcp_init_cwnd_reduction(sk
);
3492 tcp_set_ca_state(sk
, TCP_CA_CWR
);
3493 tcp_end_cwnd_reduction(sk
);
3494 tcp_try_keep_open(sk
);
3495 NET_INC_STATS_BH(sock_net(sk
),
3496 LINUX_MIB_TCPLOSSPROBERECOVERY
);
3497 } else if (!(flag
& (FLAG_SND_UNA_ADVANCED
|
3498 FLAG_NOT_DUP
| FLAG_DATA_SACKED
))) {
3499 /* Pure dupack: original and TLP probe arrived; no loss */
3500 tp
->tlp_high_seq
= 0;
3504 static inline void tcp_in_ack_event(struct sock
*sk
, u32 flags
)
3506 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3508 if (icsk
->icsk_ca_ops
->in_ack_event
)
3509 icsk
->icsk_ca_ops
->in_ack_event(sk
, flags
);
3512 /* This routine deals with incoming acks, but not outgoing ones. */
3513 static int tcp_ack(struct sock
*sk
, const struct sk_buff
*skb
, int flag
)
3515 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3516 struct tcp_sock
*tp
= tcp_sk(sk
);
3517 struct tcp_sacktag_state sack_state
;
3518 u32 prior_snd_una
= tp
->snd_una
;
3519 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
3520 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3521 bool is_dupack
= false;
3523 int prior_packets
= tp
->packets_out
;
3524 const int prior_unsacked
= tp
->packets_out
- tp
->sacked_out
;
3525 int acked
= 0; /* Number of packets newly acked */
3527 sack_state
.first_sackt
.v64
= 0;
3529 /* We very likely will need to access write queue head. */
3530 prefetchw(sk
->sk_write_queue
.next
);
3532 /* If the ack is older than previous acks
3533 * then we can probably ignore it.
3535 if (before(ack
, prior_snd_una
)) {
3536 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3537 if (before(ack
, prior_snd_una
- tp
->max_window
)) {
3538 tcp_send_challenge_ack(sk
, skb
);
3544 /* If the ack includes data we haven't sent yet, discard
3545 * this segment (RFC793 Section 3.9).
3547 if (after(ack
, tp
->snd_nxt
))
3550 if (icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
||
3551 icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
)
3554 if (after(ack
, prior_snd_una
)) {
3555 flag
|= FLAG_SND_UNA_ADVANCED
;
3556 icsk
->icsk_retransmits
= 0;
3559 prior_fackets
= tp
->fackets_out
;
3561 /* ts_recent update must be made after we are sure that the packet
3564 if (flag
& FLAG_UPDATE_TS_RECENT
)
3565 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
3567 if (!(flag
& FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
3568 /* Window is constant, pure forward advance.
3569 * No more checks are required.
3570 * Note, we use the fact that SND.UNA>=SND.WL2.
3572 tcp_update_wl(tp
, ack_seq
);
3573 tcp_snd_una_update(tp
, ack
);
3574 flag
|= FLAG_WIN_UPDATE
;
3576 tcp_in_ack_event(sk
, CA_ACK_WIN_UPDATE
);
3578 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPACKS
);
3580 u32 ack_ev_flags
= CA_ACK_SLOWPATH
;
3582 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
3585 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPUREACKS
);
3587 flag
|= tcp_ack_update_window(sk
, skb
, ack
, ack_seq
);
3589 if (TCP_SKB_CB(skb
)->sacked
)
3590 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
,
3593 if (tcp_ecn_rcv_ecn_echo(tp
, tcp_hdr(skb
))) {
3595 ack_ev_flags
|= CA_ACK_ECE
;
3598 if (flag
& FLAG_WIN_UPDATE
)
3599 ack_ev_flags
|= CA_ACK_WIN_UPDATE
;
3601 tcp_in_ack_event(sk
, ack_ev_flags
);
3604 /* We passed data and got it acked, remove any soft error
3605 * log. Something worked...
3607 sk
->sk_err_soft
= 0;
3608 icsk
->icsk_probes_out
= 0;
3609 tp
->rcv_tstamp
= tcp_time_stamp
;
3613 /* See if we can take anything off of the retransmit queue. */
3614 acked
= tp
->packets_out
;
3615 flag
|= tcp_clean_rtx_queue(sk
, prior_fackets
, prior_snd_una
,
3617 acked
-= tp
->packets_out
;
3619 if (tcp_ack_is_dubious(sk
, flag
)) {
3620 is_dupack
= !(flag
& (FLAG_SND_UNA_ADVANCED
| FLAG_NOT_DUP
));
3621 tcp_fastretrans_alert(sk
, acked
, prior_unsacked
,
3624 if (tp
->tlp_high_seq
)
3625 tcp_process_tlp_ack(sk
, ack
, flag
);
3627 /* Advance cwnd if state allows */
3628 if (tcp_may_raise_cwnd(sk
, flag
))
3629 tcp_cong_avoid(sk
, ack
, acked
);
3631 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
& FLAG_NOT_DUP
)) {
3632 struct dst_entry
*dst
= __sk_dst_get(sk
);
3637 if (icsk
->icsk_pending
== ICSK_TIME_RETRANS
)
3638 tcp_schedule_loss_probe(sk
);
3639 tcp_update_pacing_rate(sk
);
3643 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3644 if (flag
& FLAG_DSACKING_ACK
)
3645 tcp_fastretrans_alert(sk
, acked
, prior_unsacked
,
3647 /* If this ack opens up a zero window, clear backoff. It was
3648 * being used to time the probes, and is probably far higher than
3649 * it needs to be for normal retransmission.
3651 if (tcp_send_head(sk
))
3654 if (tp
->tlp_high_seq
)
3655 tcp_process_tlp_ack(sk
, ack
, flag
);
3659 SOCK_DEBUG(sk
, "Ack %u after %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3663 /* If data was SACKed, tag it and see if we should send more data.
3664 * If data was DSACKed, see if we can undo a cwnd reduction.
3666 if (TCP_SKB_CB(skb
)->sacked
) {
3667 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
,
3669 tcp_fastretrans_alert(sk
, acked
, prior_unsacked
,
3673 SOCK_DEBUG(sk
, "Ack %u before %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3677 static void tcp_parse_fastopen_option(int len
, const unsigned char *cookie
,
3678 bool syn
, struct tcp_fastopen_cookie
*foc
,
3681 /* Valid only in SYN or SYN-ACK with an even length. */
3682 if (!foc
|| !syn
|| len
< 0 || (len
& 1))
3685 if (len
>= TCP_FASTOPEN_COOKIE_MIN
&&
3686 len
<= TCP_FASTOPEN_COOKIE_MAX
)
3687 memcpy(foc
->val
, cookie
, len
);
3694 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3695 * But, this can also be called on packets in the established flow when
3696 * the fast version below fails.
3698 void tcp_parse_options(const struct sk_buff
*skb
,
3699 struct tcp_options_received
*opt_rx
, int estab
,
3700 struct tcp_fastopen_cookie
*foc
)
3702 const unsigned char *ptr
;
3703 const struct tcphdr
*th
= tcp_hdr(skb
);
3704 int length
= (th
->doff
* 4) - sizeof(struct tcphdr
);
3706 ptr
= (const unsigned char *)(th
+ 1);
3707 opt_rx
->saw_tstamp
= 0;
3709 while (length
> 0) {
3710 int opcode
= *ptr
++;
3716 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
3721 if (opsize
< 2) /* "silly options" */
3723 if (opsize
> length
)
3724 return; /* don't parse partial options */
3727 if (opsize
== TCPOLEN_MSS
&& th
->syn
&& !estab
) {
3728 u16 in_mss
= get_unaligned_be16(ptr
);
3730 if (opt_rx
->user_mss
&&
3731 opt_rx
->user_mss
< in_mss
)
3732 in_mss
= opt_rx
->user_mss
;
3733 opt_rx
->mss_clamp
= in_mss
;
3738 if (opsize
== TCPOLEN_WINDOW
&& th
->syn
&&
3739 !estab
&& sysctl_tcp_window_scaling
) {
3740 __u8 snd_wscale
= *(__u8
*)ptr
;
3741 opt_rx
->wscale_ok
= 1;
3742 if (snd_wscale
> 14) {
3743 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3748 opt_rx
->snd_wscale
= snd_wscale
;
3751 case TCPOPT_TIMESTAMP
:
3752 if ((opsize
== TCPOLEN_TIMESTAMP
) &&
3753 ((estab
&& opt_rx
->tstamp_ok
) ||
3754 (!estab
&& sysctl_tcp_timestamps
))) {
3755 opt_rx
->saw_tstamp
= 1;
3756 opt_rx
->rcv_tsval
= get_unaligned_be32(ptr
);
3757 opt_rx
->rcv_tsecr
= get_unaligned_be32(ptr
+ 4);
3760 case TCPOPT_SACK_PERM
:
3761 if (opsize
== TCPOLEN_SACK_PERM
&& th
->syn
&&
3762 !estab
&& sysctl_tcp_sack
) {
3763 opt_rx
->sack_ok
= TCP_SACK_SEEN
;
3764 tcp_sack_reset(opt_rx
);
3769 if ((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
3770 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
3772 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
3775 #ifdef CONFIG_TCP_MD5SIG
3778 * The MD5 Hash has already been
3779 * checked (see tcp_v{4,6}_do_rcv()).
3783 case TCPOPT_FASTOPEN
:
3784 tcp_parse_fastopen_option(
3785 opsize
- TCPOLEN_FASTOPEN_BASE
,
3786 ptr
, th
->syn
, foc
, false);
3790 /* Fast Open option shares code 254 using a
3791 * 16 bits magic number.
3793 if (opsize
>= TCPOLEN_EXP_FASTOPEN_BASE
&&
3794 get_unaligned_be16(ptr
) ==
3795 TCPOPT_FASTOPEN_MAGIC
)
3796 tcp_parse_fastopen_option(opsize
-
3797 TCPOLEN_EXP_FASTOPEN_BASE
,
3798 ptr
+ 2, th
->syn
, foc
, true);
3807 EXPORT_SYMBOL(tcp_parse_options
);
3809 static bool tcp_parse_aligned_timestamp(struct tcp_sock
*tp
, const struct tcphdr
*th
)
3811 const __be32
*ptr
= (const __be32
*)(th
+ 1);
3813 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3814 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
3815 tp
->rx_opt
.saw_tstamp
= 1;
3817 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3820 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
) - tp
->tsoffset
;
3822 tp
->rx_opt
.rcv_tsecr
= 0;
3828 /* Fast parse options. This hopes to only see timestamps.
3829 * If it is wrong it falls back on tcp_parse_options().
3831 static bool tcp_fast_parse_options(const struct sk_buff
*skb
,
3832 const struct tcphdr
*th
, struct tcp_sock
*tp
)
3834 /* In the spirit of fast parsing, compare doff directly to constant
3835 * values. Because equality is used, short doff can be ignored here.
3837 if (th
->doff
== (sizeof(*th
) / 4)) {
3838 tp
->rx_opt
.saw_tstamp
= 0;
3840 } else if (tp
->rx_opt
.tstamp_ok
&&
3841 th
->doff
== ((sizeof(*th
) + TCPOLEN_TSTAMP_ALIGNED
) / 4)) {
3842 if (tcp_parse_aligned_timestamp(tp
, th
))
3846 tcp_parse_options(skb
, &tp
->rx_opt
, 1, NULL
);
3847 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
3848 tp
->rx_opt
.rcv_tsecr
-= tp
->tsoffset
;
3853 #ifdef CONFIG_TCP_MD5SIG
3855 * Parse MD5 Signature option
3857 const u8
*tcp_parse_md5sig_option(const struct tcphdr
*th
)
3859 int length
= (th
->doff
<< 2) - sizeof(*th
);
3860 const u8
*ptr
= (const u8
*)(th
+ 1);
3862 /* If the TCP option is too short, we can short cut */
3863 if (length
< TCPOLEN_MD5SIG
)
3866 while (length
> 0) {
3867 int opcode
= *ptr
++;
3878 if (opsize
< 2 || opsize
> length
)
3880 if (opcode
== TCPOPT_MD5SIG
)
3881 return opsize
== TCPOLEN_MD5SIG
? ptr
: NULL
;
3888 EXPORT_SYMBOL(tcp_parse_md5sig_option
);
3891 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3893 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3894 * it can pass through stack. So, the following predicate verifies that
3895 * this segment is not used for anything but congestion avoidance or
3896 * fast retransmit. Moreover, we even are able to eliminate most of such
3897 * second order effects, if we apply some small "replay" window (~RTO)
3898 * to timestamp space.
3900 * All these measures still do not guarantee that we reject wrapped ACKs
3901 * on networks with high bandwidth, when sequence space is recycled fastly,
3902 * but it guarantees that such events will be very rare and do not affect
3903 * connection seriously. This doesn't look nice, but alas, PAWS is really
3906 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3907 * states that events when retransmit arrives after original data are rare.
3908 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3909 * the biggest problem on large power networks even with minor reordering.
3910 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3911 * up to bandwidth of 18Gigabit/sec. 8) ]
3914 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
3916 const struct tcp_sock
*tp
= tcp_sk(sk
);
3917 const struct tcphdr
*th
= tcp_hdr(skb
);
3918 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3919 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3921 return (/* 1. Pure ACK with correct sequence number. */
3922 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
3924 /* 2. ... and duplicate ACK. */
3925 ack
== tp
->snd_una
&&
3927 /* 3. ... and does not update window. */
3928 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
3930 /* 4. ... and sits in replay window. */
3931 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
3934 static inline bool tcp_paws_discard(const struct sock
*sk
,
3935 const struct sk_buff
*skb
)
3937 const struct tcp_sock
*tp
= tcp_sk(sk
);
3939 return !tcp_paws_check(&tp
->rx_opt
, TCP_PAWS_WINDOW
) &&
3940 !tcp_disordered_ack(sk
, skb
);
3943 /* Check segment sequence number for validity.
3945 * Segment controls are considered valid, if the segment
3946 * fits to the window after truncation to the window. Acceptability
3947 * of data (and SYN, FIN, of course) is checked separately.
3948 * See tcp_data_queue(), for example.
3950 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3951 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3952 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3953 * (borrowed from freebsd)
3956 static inline bool tcp_sequence(const struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3958 return !before(end_seq
, tp
->rcv_wup
) &&
3959 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
3962 /* When we get a reset we do this. */
3963 void tcp_reset(struct sock
*sk
)
3965 /* We want the right error as BSD sees it (and indeed as we do). */
3966 switch (sk
->sk_state
) {
3968 sk
->sk_err
= ECONNREFUSED
;
3970 case TCP_CLOSE_WAIT
:
3976 sk
->sk_err
= ECONNRESET
;
3978 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3981 if (!sock_flag(sk
, SOCK_DEAD
))
3982 sk
->sk_error_report(sk
);
3988 * Process the FIN bit. This now behaves as it is supposed to work
3989 * and the FIN takes effect when it is validly part of sequence
3990 * space. Not before when we get holes.
3992 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3993 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3996 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3997 * close and we go into CLOSING (and later onto TIME-WAIT)
3999 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4001 static void tcp_fin(struct sock
*sk
)
4003 struct tcp_sock
*tp
= tcp_sk(sk
);
4005 inet_csk_schedule_ack(sk
);
4007 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
4008 sock_set_flag(sk
, SOCK_DONE
);
4010 switch (sk
->sk_state
) {
4012 case TCP_ESTABLISHED
:
4013 /* Move to CLOSE_WAIT */
4014 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
4015 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
4018 case TCP_CLOSE_WAIT
:
4020 /* Received a retransmission of the FIN, do
4025 /* RFC793: Remain in the LAST-ACK state. */
4029 /* This case occurs when a simultaneous close
4030 * happens, we must ack the received FIN and
4031 * enter the CLOSING state.
4034 tcp_set_state(sk
, TCP_CLOSING
);
4037 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4039 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
4042 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4043 * cases we should never reach this piece of code.
4045 pr_err("%s: Impossible, sk->sk_state=%d\n",
4046 __func__
, sk
->sk_state
);
4050 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4051 * Probably, we should reset in this case. For now drop them.
4053 __skb_queue_purge(&tp
->out_of_order_queue
);
4054 if (tcp_is_sack(tp
))
4055 tcp_sack_reset(&tp
->rx_opt
);
4058 if (!sock_flag(sk
, SOCK_DEAD
)) {
4059 sk
->sk_state_change(sk
);
4061 /* Do not send POLL_HUP for half duplex close. */
4062 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
4063 sk
->sk_state
== TCP_CLOSE
)
4064 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_HUP
);
4066 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_IN
);
4070 static inline bool tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
,
4073 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
4074 if (before(seq
, sp
->start_seq
))
4075 sp
->start_seq
= seq
;
4076 if (after(end_seq
, sp
->end_seq
))
4077 sp
->end_seq
= end_seq
;
4083 static void tcp_dsack_set(struct sock
*sk
, u32 seq
, u32 end_seq
)
4085 struct tcp_sock
*tp
= tcp_sk(sk
);
4087 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
4090 if (before(seq
, tp
->rcv_nxt
))
4091 mib_idx
= LINUX_MIB_TCPDSACKOLDSENT
;
4093 mib_idx
= LINUX_MIB_TCPDSACKOFOSENT
;
4095 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
4097 tp
->rx_opt
.dsack
= 1;
4098 tp
->duplicate_sack
[0].start_seq
= seq
;
4099 tp
->duplicate_sack
[0].end_seq
= end_seq
;
4103 static void tcp_dsack_extend(struct sock
*sk
, u32 seq
, u32 end_seq
)
4105 struct tcp_sock
*tp
= tcp_sk(sk
);
4107 if (!tp
->rx_opt
.dsack
)
4108 tcp_dsack_set(sk
, seq
, end_seq
);
4110 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
4113 static void tcp_send_dupack(struct sock
*sk
, const struct sk_buff
*skb
)
4115 struct tcp_sock
*tp
= tcp_sk(sk
);
4117 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
4118 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4119 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4120 tcp_enter_quickack_mode(sk
);
4122 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
4123 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4125 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
4126 end_seq
= tp
->rcv_nxt
;
4127 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, end_seq
);
4134 /* These routines update the SACK block as out-of-order packets arrive or
4135 * in-order packets close up the sequence space.
4137 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
4140 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4141 struct tcp_sack_block
*swalk
= sp
+ 1;
4143 /* See if the recent change to the first SACK eats into
4144 * or hits the sequence space of other SACK blocks, if so coalesce.
4146 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
;) {
4147 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
4150 /* Zap SWALK, by moving every further SACK up by one slot.
4151 * Decrease num_sacks.
4153 tp
->rx_opt
.num_sacks
--;
4154 for (i
= this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
4158 this_sack
++, swalk
++;
4162 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
4164 struct tcp_sock
*tp
= tcp_sk(sk
);
4165 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4166 int cur_sacks
= tp
->rx_opt
.num_sacks
;
4172 for (this_sack
= 0; this_sack
< cur_sacks
; this_sack
++, sp
++) {
4173 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
4174 /* Rotate this_sack to the first one. */
4175 for (; this_sack
> 0; this_sack
--, sp
--)
4176 swap(*sp
, *(sp
- 1));
4178 tcp_sack_maybe_coalesce(tp
);
4183 /* Could not find an adjacent existing SACK, build a new one,
4184 * put it at the front, and shift everyone else down. We
4185 * always know there is at least one SACK present already here.
4187 * If the sack array is full, forget about the last one.
4189 if (this_sack
>= TCP_NUM_SACKS
) {
4191 tp
->rx_opt
.num_sacks
--;
4194 for (; this_sack
> 0; this_sack
--, sp
--)
4198 /* Build the new head SACK, and we're done. */
4199 sp
->start_seq
= seq
;
4200 sp
->end_seq
= end_seq
;
4201 tp
->rx_opt
.num_sacks
++;
4204 /* RCV.NXT advances, some SACKs should be eaten. */
4206 static void tcp_sack_remove(struct tcp_sock
*tp
)
4208 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4209 int num_sacks
= tp
->rx_opt
.num_sacks
;
4212 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4213 if (skb_queue_empty(&tp
->out_of_order_queue
)) {
4214 tp
->rx_opt
.num_sacks
= 0;
4218 for (this_sack
= 0; this_sack
< num_sacks
;) {
4219 /* Check if the start of the sack is covered by RCV.NXT. */
4220 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
4223 /* RCV.NXT must cover all the block! */
4224 WARN_ON(before(tp
->rcv_nxt
, sp
->end_seq
));
4226 /* Zap this SACK, by moving forward any other SACKS. */
4227 for (i
= this_sack
+1; i
< num_sacks
; i
++)
4228 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
4235 tp
->rx_opt
.num_sacks
= num_sacks
;
4239 * tcp_try_coalesce - try to merge skb to prior one
4242 * @from: buffer to add in queue
4243 * @fragstolen: pointer to boolean
4245 * Before queueing skb @from after @to, try to merge them
4246 * to reduce overall memory use and queue lengths, if cost is small.
4247 * Packets in ofo or receive queues can stay a long time.
4248 * Better try to coalesce them right now to avoid future collapses.
4249 * Returns true if caller should free @from instead of queueing it
4251 static bool tcp_try_coalesce(struct sock
*sk
,
4253 struct sk_buff
*from
,
4258 *fragstolen
= false;
4260 /* Its possible this segment overlaps with prior segment in queue */
4261 if (TCP_SKB_CB(from
)->seq
!= TCP_SKB_CB(to
)->end_seq
)
4264 if (!skb_try_coalesce(to
, from
, fragstolen
, &delta
))
4267 atomic_add(delta
, &sk
->sk_rmem_alloc
);
4268 sk_mem_charge(sk
, delta
);
4269 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRCVCOALESCE
);
4270 TCP_SKB_CB(to
)->end_seq
= TCP_SKB_CB(from
)->end_seq
;
4271 TCP_SKB_CB(to
)->ack_seq
= TCP_SKB_CB(from
)->ack_seq
;
4272 TCP_SKB_CB(to
)->tcp_flags
|= TCP_SKB_CB(from
)->tcp_flags
;
4276 /* This one checks to see if we can put data from the
4277 * out_of_order queue into the receive_queue.
4279 static void tcp_ofo_queue(struct sock
*sk
)
4281 struct tcp_sock
*tp
= tcp_sk(sk
);
4282 __u32 dsack_high
= tp
->rcv_nxt
;
4283 struct sk_buff
*skb
, *tail
;
4284 bool fragstolen
, eaten
;
4286 while ((skb
= skb_peek(&tp
->out_of_order_queue
)) != NULL
) {
4287 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
4290 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
4291 __u32 dsack
= dsack_high
;
4292 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
4293 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
4294 tcp_dsack_extend(sk
, TCP_SKB_CB(skb
)->seq
, dsack
);
4297 __skb_unlink(skb
, &tp
->out_of_order_queue
);
4298 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4299 SOCK_DEBUG(sk
, "ofo packet was already received\n");
4303 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
4304 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4305 TCP_SKB_CB(skb
)->end_seq
);
4307 tail
= skb_peek_tail(&sk
->sk_receive_queue
);
4308 eaten
= tail
&& tcp_try_coalesce(sk
, tail
, skb
, &fragstolen
);
4309 tcp_rcv_nxt_update(tp
, TCP_SKB_CB(skb
)->end_seq
);
4311 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4312 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
4315 kfree_skb_partial(skb
, fragstolen
);
4319 static bool tcp_prune_ofo_queue(struct sock
*sk
);
4320 static int tcp_prune_queue(struct sock
*sk
);
4322 static int tcp_try_rmem_schedule(struct sock
*sk
, struct sk_buff
*skb
,
4325 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
4326 !sk_rmem_schedule(sk
, skb
, size
)) {
4328 if (tcp_prune_queue(sk
) < 0)
4331 if (!sk_rmem_schedule(sk
, skb
, size
)) {
4332 if (!tcp_prune_ofo_queue(sk
))
4335 if (!sk_rmem_schedule(sk
, skb
, size
))
4342 static void tcp_data_queue_ofo(struct sock
*sk
, struct sk_buff
*skb
)
4344 struct tcp_sock
*tp
= tcp_sk(sk
);
4345 struct sk_buff
*skb1
;
4348 tcp_ecn_check_ce(tp
, skb
);
4350 if (unlikely(tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))) {
4351 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFODROP
);
4356 /* Disable header prediction. */
4358 inet_csk_schedule_ack(sk
);
4360 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFOQUEUE
);
4361 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
4362 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4364 skb1
= skb_peek_tail(&tp
->out_of_order_queue
);
4366 /* Initial out of order segment, build 1 SACK. */
4367 if (tcp_is_sack(tp
)) {
4368 tp
->rx_opt
.num_sacks
= 1;
4369 tp
->selective_acks
[0].start_seq
= TCP_SKB_CB(skb
)->seq
;
4370 tp
->selective_acks
[0].end_seq
=
4371 TCP_SKB_CB(skb
)->end_seq
;
4373 __skb_queue_head(&tp
->out_of_order_queue
, skb
);
4377 seq
= TCP_SKB_CB(skb
)->seq
;
4378 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4380 if (seq
== TCP_SKB_CB(skb1
)->end_seq
) {
4383 if (!tcp_try_coalesce(sk
, skb1
, skb
, &fragstolen
)) {
4384 __skb_queue_after(&tp
->out_of_order_queue
, skb1
, skb
);
4386 tcp_grow_window(sk
, skb
);
4387 kfree_skb_partial(skb
, fragstolen
);
4391 if (!tp
->rx_opt
.num_sacks
||
4392 tp
->selective_acks
[0].end_seq
!= seq
)
4395 /* Common case: data arrive in order after hole. */
4396 tp
->selective_acks
[0].end_seq
= end_seq
;
4400 /* Find place to insert this segment. */
4402 if (!after(TCP_SKB_CB(skb1
)->seq
, seq
))
4404 if (skb_queue_is_first(&tp
->out_of_order_queue
, skb1
)) {
4408 skb1
= skb_queue_prev(&tp
->out_of_order_queue
, skb1
);
4411 /* Do skb overlap to previous one? */
4412 if (skb1
&& before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4413 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4414 /* All the bits are present. Drop. */
4415 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFOMERGE
);
4418 tcp_dsack_set(sk
, seq
, end_seq
);
4421 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
4422 /* Partial overlap. */
4423 tcp_dsack_set(sk
, seq
,
4424 TCP_SKB_CB(skb1
)->end_seq
);
4426 if (skb_queue_is_first(&tp
->out_of_order_queue
,
4430 skb1
= skb_queue_prev(
4431 &tp
->out_of_order_queue
,
4436 __skb_queue_head(&tp
->out_of_order_queue
, skb
);
4438 __skb_queue_after(&tp
->out_of_order_queue
, skb1
, skb
);
4440 /* And clean segments covered by new one as whole. */
4441 while (!skb_queue_is_last(&tp
->out_of_order_queue
, skb
)) {
4442 skb1
= skb_queue_next(&tp
->out_of_order_queue
, skb
);
4444 if (!after(end_seq
, TCP_SKB_CB(skb1
)->seq
))
4446 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4447 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4451 __skb_unlink(skb1
, &tp
->out_of_order_queue
);
4452 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4453 TCP_SKB_CB(skb1
)->end_seq
);
4454 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFOMERGE
);
4459 if (tcp_is_sack(tp
))
4460 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
4463 tcp_grow_window(sk
, skb
);
4464 skb_set_owner_r(skb
, sk
);
4468 static int __must_check
tcp_queue_rcv(struct sock
*sk
, struct sk_buff
*skb
, int hdrlen
,
4472 struct sk_buff
*tail
= skb_peek_tail(&sk
->sk_receive_queue
);
4474 __skb_pull(skb
, hdrlen
);
4476 tcp_try_coalesce(sk
, tail
, skb
, fragstolen
)) ? 1 : 0;
4477 tcp_rcv_nxt_update(tcp_sk(sk
), TCP_SKB_CB(skb
)->end_seq
);
4479 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4480 skb_set_owner_r(skb
, sk
);
4485 int tcp_send_rcvq(struct sock
*sk
, struct msghdr
*msg
, size_t size
)
4487 struct sk_buff
*skb
;
4495 if (size
> PAGE_SIZE
) {
4496 int npages
= min_t(size_t, size
>> PAGE_SHIFT
, MAX_SKB_FRAGS
);
4498 data_len
= npages
<< PAGE_SHIFT
;
4499 size
= data_len
+ (size
& ~PAGE_MASK
);
4501 skb
= alloc_skb_with_frags(size
- data_len
, data_len
,
4502 PAGE_ALLOC_COSTLY_ORDER
,
4503 &err
, sk
->sk_allocation
);
4507 skb_put(skb
, size
- data_len
);
4508 skb
->data_len
= data_len
;
4511 if (tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))
4514 err
= skb_copy_datagram_from_iter(skb
, 0, &msg
->msg_iter
, size
);
4518 TCP_SKB_CB(skb
)->seq
= tcp_sk(sk
)->rcv_nxt
;
4519 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ size
;
4520 TCP_SKB_CB(skb
)->ack_seq
= tcp_sk(sk
)->snd_una
- 1;
4522 if (tcp_queue_rcv(sk
, skb
, 0, &fragstolen
)) {
4523 WARN_ON_ONCE(fragstolen
); /* should not happen */
4535 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
4537 struct tcp_sock
*tp
= tcp_sk(sk
);
4539 bool fragstolen
= false;
4541 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
)
4545 __skb_pull(skb
, tcp_hdr(skb
)->doff
* 4);
4547 tcp_ecn_accept_cwr(tp
, skb
);
4549 tp
->rx_opt
.dsack
= 0;
4551 /* Queue data for delivery to the user.
4552 * Packets in sequence go to the receive queue.
4553 * Out of sequence packets to the out_of_order_queue.
4555 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
4556 if (tcp_receive_window(tp
) == 0)
4559 /* Ok. In sequence. In window. */
4560 if (tp
->ucopy
.task
== current
&&
4561 tp
->copied_seq
== tp
->rcv_nxt
&& tp
->ucopy
.len
&&
4562 sock_owned_by_user(sk
) && !tp
->urg_data
) {
4563 int chunk
= min_t(unsigned int, skb
->len
,
4566 __set_current_state(TASK_RUNNING
);
4569 if (!skb_copy_datagram_msg(skb
, 0, tp
->ucopy
.msg
, chunk
)) {
4570 tp
->ucopy
.len
-= chunk
;
4571 tp
->copied_seq
+= chunk
;
4572 eaten
= (chunk
== skb
->len
);
4573 tcp_rcv_space_adjust(sk
);
4581 if (skb_queue_len(&sk
->sk_receive_queue
) == 0)
4582 sk_forced_mem_schedule(sk
, skb
->truesize
);
4583 else if (tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))
4586 eaten
= tcp_queue_rcv(sk
, skb
, 0, &fragstolen
);
4588 tcp_rcv_nxt_update(tp
, TCP_SKB_CB(skb
)->end_seq
);
4590 tcp_event_data_recv(sk
, skb
);
4591 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
4594 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4597 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4598 * gap in queue is filled.
4600 if (skb_queue_empty(&tp
->out_of_order_queue
))
4601 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
4604 if (tp
->rx_opt
.num_sacks
)
4605 tcp_sack_remove(tp
);
4607 tcp_fast_path_check(sk
);
4610 kfree_skb_partial(skb
, fragstolen
);
4611 if (!sock_flag(sk
, SOCK_DEAD
))
4612 sk
->sk_data_ready(sk
);
4616 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4617 /* A retransmit, 2nd most common case. Force an immediate ack. */
4618 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4619 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4622 tcp_enter_quickack_mode(sk
);
4623 inet_csk_schedule_ack(sk
);
4629 /* Out of window. F.e. zero window probe. */
4630 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
4633 tcp_enter_quickack_mode(sk
);
4635 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4636 /* Partial packet, seq < rcv_next < end_seq */
4637 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
4638 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4639 TCP_SKB_CB(skb
)->end_seq
);
4641 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
4643 /* If window is closed, drop tail of packet. But after
4644 * remembering D-SACK for its head made in previous line.
4646 if (!tcp_receive_window(tp
))
4651 tcp_data_queue_ofo(sk
, skb
);
4654 static struct sk_buff
*tcp_collapse_one(struct sock
*sk
, struct sk_buff
*skb
,
4655 struct sk_buff_head
*list
)
4657 struct sk_buff
*next
= NULL
;
4659 if (!skb_queue_is_last(list
, skb
))
4660 next
= skb_queue_next(list
, skb
);
4662 __skb_unlink(skb
, list
);
4664 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRCVCOLLAPSED
);
4669 /* Collapse contiguous sequence of skbs head..tail with
4670 * sequence numbers start..end.
4672 * If tail is NULL, this means until the end of the list.
4674 * Segments with FIN/SYN are not collapsed (only because this
4678 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
,
4679 struct sk_buff
*head
, struct sk_buff
*tail
,
4682 struct sk_buff
*skb
, *n
;
4685 /* First, check that queue is collapsible and find
4686 * the point where collapsing can be useful. */
4690 skb_queue_walk_from_safe(list
, skb
, n
) {
4693 /* No new bits? It is possible on ofo queue. */
4694 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4695 skb
= tcp_collapse_one(sk
, skb
, list
);
4701 /* The first skb to collapse is:
4703 * - bloated or contains data before "start" or
4704 * overlaps to the next one.
4706 if (!(TCP_SKB_CB(skb
)->tcp_flags
& (TCPHDR_SYN
| TCPHDR_FIN
)) &&
4707 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
4708 before(TCP_SKB_CB(skb
)->seq
, start
))) {
4709 end_of_skbs
= false;
4713 if (!skb_queue_is_last(list
, skb
)) {
4714 struct sk_buff
*next
= skb_queue_next(list
, skb
);
4716 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(next
)->seq
) {
4717 end_of_skbs
= false;
4722 /* Decided to skip this, advance start seq. */
4723 start
= TCP_SKB_CB(skb
)->end_seq
;
4726 (TCP_SKB_CB(skb
)->tcp_flags
& (TCPHDR_SYN
| TCPHDR_FIN
)))
4729 while (before(start
, end
)) {
4730 int copy
= min_t(int, SKB_MAX_ORDER(0, 0), end
- start
);
4731 struct sk_buff
*nskb
;
4733 nskb
= alloc_skb(copy
, GFP_ATOMIC
);
4737 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
4738 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
4739 __skb_queue_before(list
, skb
, nskb
);
4740 skb_set_owner_r(nskb
, sk
);
4742 /* Copy data, releasing collapsed skbs. */
4744 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
4745 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
4749 size
= min(copy
, size
);
4750 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
4752 TCP_SKB_CB(nskb
)->end_seq
+= size
;
4756 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4757 skb
= tcp_collapse_one(sk
, skb
, list
);
4760 (TCP_SKB_CB(skb
)->tcp_flags
& (TCPHDR_SYN
| TCPHDR_FIN
)))
4767 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4768 * and tcp_collapse() them until all the queue is collapsed.
4770 static void tcp_collapse_ofo_queue(struct sock
*sk
)
4772 struct tcp_sock
*tp
= tcp_sk(sk
);
4773 struct sk_buff
*skb
= skb_peek(&tp
->out_of_order_queue
);
4774 struct sk_buff
*head
;
4780 start
= TCP_SKB_CB(skb
)->seq
;
4781 end
= TCP_SKB_CB(skb
)->end_seq
;
4785 struct sk_buff
*next
= NULL
;
4787 if (!skb_queue_is_last(&tp
->out_of_order_queue
, skb
))
4788 next
= skb_queue_next(&tp
->out_of_order_queue
, skb
);
4791 /* Segment is terminated when we see gap or when
4792 * we are at the end of all the queue. */
4794 after(TCP_SKB_CB(skb
)->seq
, end
) ||
4795 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
4796 tcp_collapse(sk
, &tp
->out_of_order_queue
,
4797 head
, skb
, start
, end
);
4801 /* Start new segment */
4802 start
= TCP_SKB_CB(skb
)->seq
;
4803 end
= TCP_SKB_CB(skb
)->end_seq
;
4805 if (before(TCP_SKB_CB(skb
)->seq
, start
))
4806 start
= TCP_SKB_CB(skb
)->seq
;
4807 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
4808 end
= TCP_SKB_CB(skb
)->end_seq
;
4814 * Purge the out-of-order queue.
4815 * Return true if queue was pruned.
4817 static bool tcp_prune_ofo_queue(struct sock
*sk
)
4819 struct tcp_sock
*tp
= tcp_sk(sk
);
4822 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4823 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_OFOPRUNED
);
4824 __skb_queue_purge(&tp
->out_of_order_queue
);
4826 /* Reset SACK state. A conforming SACK implementation will
4827 * do the same at a timeout based retransmit. When a connection
4828 * is in a sad state like this, we care only about integrity
4829 * of the connection not performance.
4831 if (tp
->rx_opt
.sack_ok
)
4832 tcp_sack_reset(&tp
->rx_opt
);
4839 /* Reduce allocated memory if we can, trying to get
4840 * the socket within its memory limits again.
4842 * Return less than zero if we should start dropping frames
4843 * until the socket owning process reads some of the data
4844 * to stabilize the situation.
4846 static int tcp_prune_queue(struct sock
*sk
)
4848 struct tcp_sock
*tp
= tcp_sk(sk
);
4850 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
4852 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PRUNECALLED
);
4854 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
4855 tcp_clamp_window(sk
);
4856 else if (tcp_under_memory_pressure(sk
))
4857 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
4859 tcp_collapse_ofo_queue(sk
);
4860 if (!skb_queue_empty(&sk
->sk_receive_queue
))
4861 tcp_collapse(sk
, &sk
->sk_receive_queue
,
4862 skb_peek(&sk
->sk_receive_queue
),
4864 tp
->copied_seq
, tp
->rcv_nxt
);
4867 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4870 /* Collapsing did not help, destructive actions follow.
4871 * This must not ever occur. */
4873 tcp_prune_ofo_queue(sk
);
4875 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4878 /* If we are really being abused, tell the caller to silently
4879 * drop receive data on the floor. It will get retransmitted
4880 * and hopefully then we'll have sufficient space.
4882 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_RCVPRUNED
);
4884 /* Massive buffer overcommit. */
4889 static bool tcp_should_expand_sndbuf(const struct sock
*sk
)
4891 const struct tcp_sock
*tp
= tcp_sk(sk
);
4893 /* If the user specified a specific send buffer setting, do
4896 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
4899 /* If we are under global TCP memory pressure, do not expand. */
4900 if (tcp_under_memory_pressure(sk
))
4903 /* If we are under soft global TCP memory pressure, do not expand. */
4904 if (sk_memory_allocated(sk
) >= sk_prot_mem_limits(sk
, 0))
4907 /* If we filled the congestion window, do not expand. */
4908 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
4914 /* When incoming ACK allowed to free some skb from write_queue,
4915 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4916 * on the exit from tcp input handler.
4918 * PROBLEM: sndbuf expansion does not work well with largesend.
4920 static void tcp_new_space(struct sock
*sk
)
4922 struct tcp_sock
*tp
= tcp_sk(sk
);
4924 if (tcp_should_expand_sndbuf(sk
)) {
4925 tcp_sndbuf_expand(sk
);
4926 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4929 sk
->sk_write_space(sk
);
4932 static void tcp_check_space(struct sock
*sk
)
4934 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
4935 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
4936 /* pairs with tcp_poll() */
4937 smp_mb__after_atomic();
4938 if (sk
->sk_socket
&&
4939 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
4944 static inline void tcp_data_snd_check(struct sock
*sk
)
4946 tcp_push_pending_frames(sk
);
4947 tcp_check_space(sk
);
4951 * Check if sending an ack is needed.
4953 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
4955 struct tcp_sock
*tp
= tcp_sk(sk
);
4957 /* More than one full frame received... */
4958 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
&&
4959 /* ... and right edge of window advances far enough.
4960 * (tcp_recvmsg() will send ACK otherwise). Or...
4962 __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
4963 /* We ACK each frame or... */
4964 tcp_in_quickack_mode(sk
) ||
4965 /* We have out of order data. */
4966 (ofo_possible
&& skb_peek(&tp
->out_of_order_queue
))) {
4967 /* Then ack it now */
4970 /* Else, send delayed ack. */
4971 tcp_send_delayed_ack(sk
);
4975 static inline void tcp_ack_snd_check(struct sock
*sk
)
4977 if (!inet_csk_ack_scheduled(sk
)) {
4978 /* We sent a data segment already. */
4981 __tcp_ack_snd_check(sk
, 1);
4985 * This routine is only called when we have urgent data
4986 * signaled. Its the 'slow' part of tcp_urg. It could be
4987 * moved inline now as tcp_urg is only called from one
4988 * place. We handle URGent data wrong. We have to - as
4989 * BSD still doesn't use the correction from RFC961.
4990 * For 1003.1g we should support a new option TCP_STDURG to permit
4991 * either form (or just set the sysctl tcp_stdurg).
4994 static void tcp_check_urg(struct sock
*sk
, const struct tcphdr
*th
)
4996 struct tcp_sock
*tp
= tcp_sk(sk
);
4997 u32 ptr
= ntohs(th
->urg_ptr
);
4999 if (ptr
&& !sysctl_tcp_stdurg
)
5001 ptr
+= ntohl(th
->seq
);
5003 /* Ignore urgent data that we've already seen and read. */
5004 if (after(tp
->copied_seq
, ptr
))
5007 /* Do not replay urg ptr.
5009 * NOTE: interesting situation not covered by specs.
5010 * Misbehaving sender may send urg ptr, pointing to segment,
5011 * which we already have in ofo queue. We are not able to fetch
5012 * such data and will stay in TCP_URG_NOTYET until will be eaten
5013 * by recvmsg(). Seems, we are not obliged to handle such wicked
5014 * situations. But it is worth to think about possibility of some
5015 * DoSes using some hypothetical application level deadlock.
5017 if (before(ptr
, tp
->rcv_nxt
))
5020 /* Do we already have a newer (or duplicate) urgent pointer? */
5021 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
5024 /* Tell the world about our new urgent pointer. */
5027 /* We may be adding urgent data when the last byte read was
5028 * urgent. To do this requires some care. We cannot just ignore
5029 * tp->copied_seq since we would read the last urgent byte again
5030 * as data, nor can we alter copied_seq until this data arrives
5031 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5033 * NOTE. Double Dutch. Rendering to plain English: author of comment
5034 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5035 * and expect that both A and B disappear from stream. This is _wrong_.
5036 * Though this happens in BSD with high probability, this is occasional.
5037 * Any application relying on this is buggy. Note also, that fix "works"
5038 * only in this artificial test. Insert some normal data between A and B and we will
5039 * decline of BSD again. Verdict: it is better to remove to trap
5042 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
5043 !sock_flag(sk
, SOCK_URGINLINE
) && tp
->copied_seq
!= tp
->rcv_nxt
) {
5044 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
5046 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5047 __skb_unlink(skb
, &sk
->sk_receive_queue
);
5052 tp
->urg_data
= TCP_URG_NOTYET
;
5055 /* Disable header prediction. */
5059 /* This is the 'fast' part of urgent handling. */
5060 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, const struct tcphdr
*th
)
5062 struct tcp_sock
*tp
= tcp_sk(sk
);
5064 /* Check if we get a new urgent pointer - normally not. */
5066 tcp_check_urg(sk
, th
);
5068 /* Do we wait for any urgent data? - normally not... */
5069 if (tp
->urg_data
== TCP_URG_NOTYET
) {
5070 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
5073 /* Is the urgent pointer pointing into this packet? */
5074 if (ptr
< skb
->len
) {
5076 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
5078 tp
->urg_data
= TCP_URG_VALID
| tmp
;
5079 if (!sock_flag(sk
, SOCK_DEAD
))
5080 sk
->sk_data_ready(sk
);
5085 static int tcp_copy_to_iovec(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
5087 struct tcp_sock
*tp
= tcp_sk(sk
);
5088 int chunk
= skb
->len
- hlen
;
5092 if (skb_csum_unnecessary(skb
))
5093 err
= skb_copy_datagram_msg(skb
, hlen
, tp
->ucopy
.msg
, chunk
);
5095 err
= skb_copy_and_csum_datagram_msg(skb
, hlen
, tp
->ucopy
.msg
);
5098 tp
->ucopy
.len
-= chunk
;
5099 tp
->copied_seq
+= chunk
;
5100 tcp_rcv_space_adjust(sk
);
5107 static __sum16
__tcp_checksum_complete_user(struct sock
*sk
,
5108 struct sk_buff
*skb
)
5112 if (sock_owned_by_user(sk
)) {
5114 result
= __tcp_checksum_complete(skb
);
5117 result
= __tcp_checksum_complete(skb
);
5122 static inline bool tcp_checksum_complete_user(struct sock
*sk
,
5123 struct sk_buff
*skb
)
5125 return !skb_csum_unnecessary(skb
) &&
5126 __tcp_checksum_complete_user(sk
, skb
);
5129 /* Does PAWS and seqno based validation of an incoming segment, flags will
5130 * play significant role here.
5132 static bool tcp_validate_incoming(struct sock
*sk
, struct sk_buff
*skb
,
5133 const struct tcphdr
*th
, int syn_inerr
)
5135 struct tcp_sock
*tp
= tcp_sk(sk
);
5137 /* RFC1323: H1. Apply PAWS check first. */
5138 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
5139 tcp_paws_discard(sk
, skb
)) {
5141 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSESTABREJECTED
);
5142 if (!tcp_oow_rate_limited(sock_net(sk
), skb
,
5143 LINUX_MIB_TCPACKSKIPPEDPAWS
,
5144 &tp
->last_oow_ack_time
))
5145 tcp_send_dupack(sk
, skb
);
5148 /* Reset is accepted even if it did not pass PAWS. */
5151 /* Step 1: check sequence number */
5152 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5153 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5154 * (RST) segments are validated by checking their SEQ-fields."
5155 * And page 69: "If an incoming segment is not acceptable,
5156 * an acknowledgment should be sent in reply (unless the RST
5157 * bit is set, if so drop the segment and return)".
5162 if (!tcp_oow_rate_limited(sock_net(sk
), skb
,
5163 LINUX_MIB_TCPACKSKIPPEDSEQ
,
5164 &tp
->last_oow_ack_time
))
5165 tcp_send_dupack(sk
, skb
);
5170 /* Step 2: check RST bit */
5173 * If sequence number exactly matches RCV.NXT, then
5174 * RESET the connection
5176 * Send a challenge ACK
5178 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
)
5181 tcp_send_challenge_ack(sk
, skb
);
5185 /* step 3: check security and precedence [ignored] */
5187 /* step 4: Check for a SYN
5188 * RFC 5961 4.2 : Send a challenge ack
5193 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5194 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSYNCHALLENGE
);
5195 tcp_send_challenge_ack(sk
, skb
);
5207 * TCP receive function for the ESTABLISHED state.
5209 * It is split into a fast path and a slow path. The fast path is
5211 * - A zero window was announced from us - zero window probing
5212 * is only handled properly in the slow path.
5213 * - Out of order segments arrived.
5214 * - Urgent data is expected.
5215 * - There is no buffer space left
5216 * - Unexpected TCP flags/window values/header lengths are received
5217 * (detected by checking the TCP header against pred_flags)
5218 * - Data is sent in both directions. Fast path only supports pure senders
5219 * or pure receivers (this means either the sequence number or the ack
5220 * value must stay constant)
5221 * - Unexpected TCP option.
5223 * When these conditions are not satisfied it drops into a standard
5224 * receive procedure patterned after RFC793 to handle all cases.
5225 * The first three cases are guaranteed by proper pred_flags setting,
5226 * the rest is checked inline. Fast processing is turned on in
5227 * tcp_data_queue when everything is OK.
5229 void tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
5230 const struct tcphdr
*th
, unsigned int len
)
5232 struct tcp_sock
*tp
= tcp_sk(sk
);
5234 if (unlikely(!sk
->sk_rx_dst
))
5235 inet_csk(sk
)->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
5237 * Header prediction.
5238 * The code loosely follows the one in the famous
5239 * "30 instruction TCP receive" Van Jacobson mail.
5241 * Van's trick is to deposit buffers into socket queue
5242 * on a device interrupt, to call tcp_recv function
5243 * on the receive process context and checksum and copy
5244 * the buffer to user space. smart...
5246 * Our current scheme is not silly either but we take the
5247 * extra cost of the net_bh soft interrupt processing...
5248 * We do checksum and copy also but from device to kernel.
5251 tp
->rx_opt
.saw_tstamp
= 0;
5253 /* pred_flags is 0xS?10 << 16 + snd_wnd
5254 * if header_prediction is to be made
5255 * 'S' will always be tp->tcp_header_len >> 2
5256 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5257 * turn it off (when there are holes in the receive
5258 * space for instance)
5259 * PSH flag is ignored.
5262 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
5263 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
&&
5264 !after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
)) {
5265 int tcp_header_len
= tp
->tcp_header_len
;
5267 /* Timestamp header prediction: tcp_header_len
5268 * is automatically equal to th->doff*4 due to pred_flags
5272 /* Check timestamp */
5273 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
5274 /* No? Slow path! */
5275 if (!tcp_parse_aligned_timestamp(tp
, th
))
5278 /* If PAWS failed, check it more carefully in slow path */
5279 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
5282 /* DO NOT update ts_recent here, if checksum fails
5283 * and timestamp was corrupted part, it will result
5284 * in a hung connection since we will drop all
5285 * future packets due to the PAWS test.
5289 if (len
<= tcp_header_len
) {
5290 /* Bulk data transfer: sender */
5291 if (len
== tcp_header_len
) {
5292 /* Predicted packet is in window by definition.
5293 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5294 * Hence, check seq<=rcv_wup reduces to:
5296 if (tcp_header_len
==
5297 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5298 tp
->rcv_nxt
== tp
->rcv_wup
)
5299 tcp_store_ts_recent(tp
);
5301 /* We know that such packets are checksummed
5304 tcp_ack(sk
, skb
, 0);
5306 tcp_data_snd_check(sk
);
5308 } else { /* Header too small */
5309 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5314 bool fragstolen
= false;
5316 if (tp
->ucopy
.task
== current
&&
5317 tp
->copied_seq
== tp
->rcv_nxt
&&
5318 len
- tcp_header_len
<= tp
->ucopy
.len
&&
5319 sock_owned_by_user(sk
)) {
5320 __set_current_state(TASK_RUNNING
);
5322 if (!tcp_copy_to_iovec(sk
, skb
, tcp_header_len
)) {
5323 /* Predicted packet is in window by definition.
5324 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5325 * Hence, check seq<=rcv_wup reduces to:
5327 if (tcp_header_len
==
5328 (sizeof(struct tcphdr
) +
5329 TCPOLEN_TSTAMP_ALIGNED
) &&
5330 tp
->rcv_nxt
== tp
->rcv_wup
)
5331 tcp_store_ts_recent(tp
);
5333 tcp_rcv_rtt_measure_ts(sk
, skb
);
5335 __skb_pull(skb
, tcp_header_len
);
5336 tcp_rcv_nxt_update(tp
, TCP_SKB_CB(skb
)->end_seq
);
5337 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITSTOUSER
);
5342 if (tcp_checksum_complete_user(sk
, skb
))
5345 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
5348 /* Predicted packet is in window by definition.
5349 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5350 * Hence, check seq<=rcv_wup reduces to:
5352 if (tcp_header_len
==
5353 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5354 tp
->rcv_nxt
== tp
->rcv_wup
)
5355 tcp_store_ts_recent(tp
);
5357 tcp_rcv_rtt_measure_ts(sk
, skb
);
5359 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITS
);
5361 /* Bulk data transfer: receiver */
5362 eaten
= tcp_queue_rcv(sk
, skb
, tcp_header_len
,
5366 tcp_event_data_recv(sk
, skb
);
5368 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
5369 /* Well, only one small jumplet in fast path... */
5370 tcp_ack(sk
, skb
, FLAG_DATA
);
5371 tcp_data_snd_check(sk
);
5372 if (!inet_csk_ack_scheduled(sk
))
5376 __tcp_ack_snd_check(sk
, 0);
5379 kfree_skb_partial(skb
, fragstolen
);
5380 sk
->sk_data_ready(sk
);
5386 if (len
< (th
->doff
<< 2) || tcp_checksum_complete_user(sk
, skb
))
5389 if (!th
->ack
&& !th
->rst
&& !th
->syn
)
5393 * Standard slow path.
5396 if (!tcp_validate_incoming(sk
, skb
, th
, 1))
5400 if (tcp_ack(sk
, skb
, FLAG_SLOWPATH
| FLAG_UPDATE_TS_RECENT
) < 0)
5403 tcp_rcv_rtt_measure_ts(sk
, skb
);
5405 /* Process urgent data. */
5406 tcp_urg(sk
, skb
, th
);
5408 /* step 7: process the segment text */
5409 tcp_data_queue(sk
, skb
);
5411 tcp_data_snd_check(sk
);
5412 tcp_ack_snd_check(sk
);
5416 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_CSUMERRORS
);
5417 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5422 EXPORT_SYMBOL(tcp_rcv_established
);
5424 void tcp_finish_connect(struct sock
*sk
, struct sk_buff
*skb
)
5426 struct tcp_sock
*tp
= tcp_sk(sk
);
5427 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5429 tcp_set_state(sk
, TCP_ESTABLISHED
);
5432 icsk
->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
5433 security_inet_conn_established(sk
, skb
);
5436 /* Make sure socket is routed, for correct metrics. */
5437 icsk
->icsk_af_ops
->rebuild_header(sk
);
5439 tcp_init_metrics(sk
);
5441 tcp_init_congestion_control(sk
);
5443 /* Prevent spurious tcp_cwnd_restart() on first data
5446 tp
->lsndtime
= tcp_time_stamp
;
5448 tcp_init_buffer_space(sk
);
5450 if (sock_flag(sk
, SOCK_KEEPOPEN
))
5451 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
5453 if (!tp
->rx_opt
.snd_wscale
)
5454 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
5458 if (!sock_flag(sk
, SOCK_DEAD
)) {
5459 sk
->sk_state_change(sk
);
5460 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5464 static bool tcp_rcv_fastopen_synack(struct sock
*sk
, struct sk_buff
*synack
,
5465 struct tcp_fastopen_cookie
*cookie
)
5467 struct tcp_sock
*tp
= tcp_sk(sk
);
5468 struct sk_buff
*data
= tp
->syn_data
? tcp_write_queue_head(sk
) : NULL
;
5469 u16 mss
= tp
->rx_opt
.mss_clamp
, try_exp
= 0;
5470 bool syn_drop
= false;
5472 if (mss
== tp
->rx_opt
.user_mss
) {
5473 struct tcp_options_received opt
;
5475 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5476 tcp_clear_options(&opt
);
5477 opt
.user_mss
= opt
.mss_clamp
= 0;
5478 tcp_parse_options(synack
, &opt
, 0, NULL
);
5479 mss
= opt
.mss_clamp
;
5482 if (!tp
->syn_fastopen
) {
5483 /* Ignore an unsolicited cookie */
5485 } else if (tp
->total_retrans
) {
5486 /* SYN timed out and the SYN-ACK neither has a cookie nor
5487 * acknowledges data. Presumably the remote received only
5488 * the retransmitted (regular) SYNs: either the original
5489 * SYN-data or the corresponding SYN-ACK was dropped.
5491 syn_drop
= (cookie
->len
< 0 && data
);
5492 } else if (cookie
->len
< 0 && !tp
->syn_data
) {
5493 /* We requested a cookie but didn't get it. If we did not use
5494 * the (old) exp opt format then try so next time (try_exp=1).
5495 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5497 try_exp
= tp
->syn_fastopen_exp
? 2 : 1;
5500 tcp_fastopen_cache_set(sk
, mss
, cookie
, syn_drop
, try_exp
);
5502 if (data
) { /* Retransmit unacked data in SYN */
5503 tcp_for_write_queue_from(data
, sk
) {
5504 if (data
== tcp_send_head(sk
) ||
5505 __tcp_retransmit_skb(sk
, data
))
5509 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPFASTOPENACTIVEFAIL
);
5512 tp
->syn_data_acked
= tp
->syn_data
;
5513 if (tp
->syn_data_acked
)
5514 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPFASTOPENACTIVE
);
5518 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5519 const struct tcphdr
*th
)
5521 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5522 struct tcp_sock
*tp
= tcp_sk(sk
);
5523 struct tcp_fastopen_cookie foc
= { .len
= -1 };
5524 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
5526 tcp_parse_options(skb
, &tp
->rx_opt
, 0, &foc
);
5527 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
5528 tp
->rx_opt
.rcv_tsecr
-= tp
->tsoffset
;
5532 * "If the state is SYN-SENT then
5533 * first check the ACK bit
5534 * If the ACK bit is set
5535 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5536 * a reset (unless the RST bit is set, if so drop
5537 * the segment and return)"
5539 if (!after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_una
) ||
5540 after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
))
5541 goto reset_and_undo
;
5543 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
5544 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
5546 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSACTIVEREJECTED
);
5547 goto reset_and_undo
;
5550 /* Now ACK is acceptable.
5552 * "If the RST bit is set
5553 * If the ACK was acceptable then signal the user "error:
5554 * connection reset", drop the segment, enter CLOSED state,
5555 * delete TCB, and return."
5564 * "fifth, if neither of the SYN or RST bits is set then
5565 * drop the segment and return."
5571 goto discard_and_undo
;
5574 * "If the SYN bit is on ...
5575 * are acceptable then ...
5576 * (our SYN has been ACKed), change the connection
5577 * state to ESTABLISHED..."
5580 tcp_ecn_rcv_synack(tp
, th
);
5582 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5583 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
5585 /* Ok.. it's good. Set up sequence numbers and
5586 * move to established.
5588 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5589 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5591 /* RFC1323: The window in SYN & SYN/ACK segments is
5594 tp
->snd_wnd
= ntohs(th
->window
);
5596 if (!tp
->rx_opt
.wscale_ok
) {
5597 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
5598 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
5601 if (tp
->rx_opt
.saw_tstamp
) {
5602 tp
->rx_opt
.tstamp_ok
= 1;
5603 tp
->tcp_header_len
=
5604 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5605 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5606 tcp_store_ts_recent(tp
);
5608 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5611 if (tcp_is_sack(tp
) && sysctl_tcp_fack
)
5612 tcp_enable_fack(tp
);
5615 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5616 tcp_initialize_rcv_mss(sk
);
5618 /* Remember, tcp_poll() does not lock socket!
5619 * Change state from SYN-SENT only after copied_seq
5620 * is initialized. */
5621 tp
->copied_seq
= tp
->rcv_nxt
;
5625 tcp_finish_connect(sk
, skb
);
5627 if ((tp
->syn_fastopen
|| tp
->syn_data
) &&
5628 tcp_rcv_fastopen_synack(sk
, skb
, &foc
))
5631 if (sk
->sk_write_pending
||
5632 icsk
->icsk_accept_queue
.rskq_defer_accept
||
5633 icsk
->icsk_ack
.pingpong
) {
5634 /* Save one ACK. Data will be ready after
5635 * several ticks, if write_pending is set.
5637 * It may be deleted, but with this feature tcpdumps
5638 * look so _wonderfully_ clever, that I was not able
5639 * to stand against the temptation 8) --ANK
5641 inet_csk_schedule_ack(sk
);
5642 icsk
->icsk_ack
.lrcvtime
= tcp_time_stamp
;
5643 tcp_enter_quickack_mode(sk
);
5644 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
5645 TCP_DELACK_MAX
, TCP_RTO_MAX
);
5656 /* No ACK in the segment */
5660 * "If the RST bit is set
5662 * Otherwise (no ACK) drop the segment and return."
5665 goto discard_and_undo
;
5669 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&&
5670 tcp_paws_reject(&tp
->rx_opt
, 0))
5671 goto discard_and_undo
;
5674 /* We see SYN without ACK. It is attempt of
5675 * simultaneous connect with crossed SYNs.
5676 * Particularly, it can be connect to self.
5678 tcp_set_state(sk
, TCP_SYN_RECV
);
5680 if (tp
->rx_opt
.saw_tstamp
) {
5681 tp
->rx_opt
.tstamp_ok
= 1;
5682 tcp_store_ts_recent(tp
);
5683 tp
->tcp_header_len
=
5684 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5686 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5689 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5690 tp
->copied_seq
= tp
->rcv_nxt
;
5691 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5693 /* RFC1323: The window in SYN & SYN/ACK segments is
5696 tp
->snd_wnd
= ntohs(th
->window
);
5697 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
5698 tp
->max_window
= tp
->snd_wnd
;
5700 tcp_ecn_rcv_syn(tp
, th
);
5703 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5704 tcp_initialize_rcv_mss(sk
);
5706 tcp_send_synack(sk
);
5708 /* Note, we could accept data and URG from this segment.
5709 * There are no obstacles to make this (except that we must
5710 * either change tcp_recvmsg() to prevent it from returning data
5711 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5713 * However, if we ignore data in ACKless segments sometimes,
5714 * we have no reasons to accept it sometimes.
5715 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5716 * is not flawless. So, discard packet for sanity.
5717 * Uncomment this return to process the data.
5724 /* "fifth, if neither of the SYN or RST bits is set then
5725 * drop the segment and return."
5729 tcp_clear_options(&tp
->rx_opt
);
5730 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5734 tcp_clear_options(&tp
->rx_opt
);
5735 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5740 * This function implements the receiving procedure of RFC 793 for
5741 * all states except ESTABLISHED and TIME_WAIT.
5742 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5743 * address independent.
5746 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
)
5748 struct tcp_sock
*tp
= tcp_sk(sk
);
5749 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5750 const struct tcphdr
*th
= tcp_hdr(skb
);
5751 struct request_sock
*req
;
5755 tp
->rx_opt
.saw_tstamp
= 0;
5757 switch (sk
->sk_state
) {
5771 if (icsk
->icsk_af_ops
->conn_request(sk
, skb
) < 0)
5774 /* Now we have several options: In theory there is
5775 * nothing else in the frame. KA9Q has an option to
5776 * send data with the syn, BSD accepts data with the
5777 * syn up to the [to be] advertised window and
5778 * Solaris 2.1 gives you a protocol error. For now
5779 * we just ignore it, that fits the spec precisely
5780 * and avoids incompatibilities. It would be nice in
5781 * future to drop through and process the data.
5783 * Now that TTCP is starting to be used we ought to
5785 * But, this leaves one open to an easy denial of
5786 * service attack, and SYN cookies can't defend
5787 * against this problem. So, we drop the data
5788 * in the interest of security over speed unless
5789 * it's still in use.
5797 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
);
5801 /* Do step6 onward by hand. */
5802 tcp_urg(sk
, skb
, th
);
5804 tcp_data_snd_check(sk
);
5808 req
= tp
->fastopen_rsk
;
5810 WARN_ON_ONCE(sk
->sk_state
!= TCP_SYN_RECV
&&
5811 sk
->sk_state
!= TCP_FIN_WAIT1
);
5813 if (!tcp_check_req(sk
, skb
, req
, true))
5817 if (!th
->ack
&& !th
->rst
&& !th
->syn
)
5820 if (!tcp_validate_incoming(sk
, skb
, th
, 0))
5823 /* step 5: check the ACK field */
5824 acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
|
5825 FLAG_UPDATE_TS_RECENT
) > 0;
5827 switch (sk
->sk_state
) {
5833 tcp_synack_rtt_meas(sk
, req
);
5835 /* Once we leave TCP_SYN_RECV, we no longer need req
5839 tp
->total_retrans
= req
->num_retrans
;
5840 reqsk_fastopen_remove(sk
, req
, false);
5842 /* Make sure socket is routed, for correct metrics. */
5843 icsk
->icsk_af_ops
->rebuild_header(sk
);
5844 tcp_init_congestion_control(sk
);
5847 tp
->copied_seq
= tp
->rcv_nxt
;
5848 tcp_init_buffer_space(sk
);
5851 tcp_set_state(sk
, TCP_ESTABLISHED
);
5852 sk
->sk_state_change(sk
);
5854 /* Note, that this wakeup is only for marginal crossed SYN case.
5855 * Passively open sockets are not waked up, because
5856 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5859 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5861 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
5862 tp
->snd_wnd
= ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
;
5863 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5865 if (tp
->rx_opt
.tstamp_ok
)
5866 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5869 /* Re-arm the timer because data may have been sent out.
5870 * This is similar to the regular data transmission case
5871 * when new data has just been ack'ed.
5873 * (TFO) - we could try to be more aggressive and
5874 * retransmitting any data sooner based on when they
5879 tcp_init_metrics(sk
);
5881 tcp_update_pacing_rate(sk
);
5883 /* Prevent spurious tcp_cwnd_restart() on first data packet */
5884 tp
->lsndtime
= tcp_time_stamp
;
5886 tcp_initialize_rcv_mss(sk
);
5887 tcp_fast_path_on(tp
);
5890 case TCP_FIN_WAIT1
: {
5891 struct dst_entry
*dst
;
5894 /* If we enter the TCP_FIN_WAIT1 state and we are a
5895 * Fast Open socket and this is the first acceptable
5896 * ACK we have received, this would have acknowledged
5897 * our SYNACK so stop the SYNACK timer.
5900 /* Return RST if ack_seq is invalid.
5901 * Note that RFC793 only says to generate a
5902 * DUPACK for it but for TCP Fast Open it seems
5903 * better to treat this case like TCP_SYN_RECV
5908 /* We no longer need the request sock. */
5909 reqsk_fastopen_remove(sk
, req
, false);
5912 if (tp
->snd_una
!= tp
->write_seq
)
5915 tcp_set_state(sk
, TCP_FIN_WAIT2
);
5916 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
5918 dst
= __sk_dst_get(sk
);
5922 if (!sock_flag(sk
, SOCK_DEAD
)) {
5923 /* Wake up lingering close() */
5924 sk
->sk_state_change(sk
);
5928 if (tp
->linger2
< 0 ||
5929 (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5930 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
))) {
5932 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
5936 tmo
= tcp_fin_time(sk
);
5937 if (tmo
> TCP_TIMEWAIT_LEN
) {
5938 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
5939 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
5940 /* Bad case. We could lose such FIN otherwise.
5941 * It is not a big problem, but it looks confusing
5942 * and not so rare event. We still can lose it now,
5943 * if it spins in bh_lock_sock(), but it is really
5946 inet_csk_reset_keepalive_timer(sk
, tmo
);
5948 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
5955 if (tp
->snd_una
== tp
->write_seq
) {
5956 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
5962 if (tp
->snd_una
== tp
->write_seq
) {
5963 tcp_update_metrics(sk
);
5970 /* step 6: check the URG bit */
5971 tcp_urg(sk
, skb
, th
);
5973 /* step 7: process the segment text */
5974 switch (sk
->sk_state
) {
5975 case TCP_CLOSE_WAIT
:
5978 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
5982 /* RFC 793 says to queue data in these states,
5983 * RFC 1122 says we MUST send a reset.
5984 * BSD 4.4 also does reset.
5986 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
5987 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5988 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
5989 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
5995 case TCP_ESTABLISHED
:
5996 tcp_data_queue(sk
, skb
);
6001 /* tcp_data could move socket to TIME-WAIT */
6002 if (sk
->sk_state
!= TCP_CLOSE
) {
6003 tcp_data_snd_check(sk
);
6004 tcp_ack_snd_check(sk
);
6013 EXPORT_SYMBOL(tcp_rcv_state_process
);
6015 static inline void pr_drop_req(struct request_sock
*req
, __u16 port
, int family
)
6017 struct inet_request_sock
*ireq
= inet_rsk(req
);
6019 if (family
== AF_INET
)
6020 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6021 &ireq
->ir_rmt_addr
, port
);
6022 #if IS_ENABLED(CONFIG_IPV6)
6023 else if (family
== AF_INET6
)
6024 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6025 &ireq
->ir_v6_rmt_addr
, port
);
6029 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6031 * If we receive a SYN packet with these bits set, it means a
6032 * network is playing bad games with TOS bits. In order to
6033 * avoid possible false congestion notifications, we disable
6034 * TCP ECN negotiation.
6036 * Exception: tcp_ca wants ECN. This is required for DCTCP
6037 * congestion control: Linux DCTCP asserts ECT on all packets,
6038 * including SYN, which is most optimal solution; however,
6039 * others, such as FreeBSD do not.
6041 static void tcp_ecn_create_request(struct request_sock
*req
,
6042 const struct sk_buff
*skb
,
6043 const struct sock
*listen_sk
,
6044 const struct dst_entry
*dst
)
6046 const struct tcphdr
*th
= tcp_hdr(skb
);
6047 const struct net
*net
= sock_net(listen_sk
);
6048 bool th_ecn
= th
->ece
&& th
->cwr
;
6055 ect
= !INET_ECN_is_not_ect(TCP_SKB_CB(skb
)->ip_dsfield
);
6056 ecn_ok_dst
= dst_feature(dst
, DST_FEATURE_ECN_MASK
);
6057 ecn_ok
= net
->ipv4
.sysctl_tcp_ecn
|| ecn_ok_dst
;
6059 if ((!ect
&& ecn_ok
) || tcp_ca_needs_ecn(listen_sk
) ||
6060 (ecn_ok_dst
& DST_FEATURE_ECN_CA
))
6061 inet_rsk(req
)->ecn_ok
= 1;
6064 static void tcp_openreq_init(struct request_sock
*req
,
6065 const struct tcp_options_received
*rx_opt
,
6066 struct sk_buff
*skb
, const struct sock
*sk
)
6068 struct inet_request_sock
*ireq
= inet_rsk(req
);
6070 req
->rsk_rcv_wnd
= 0; /* So that tcp_send_synack() knows! */
6072 tcp_rsk(req
)->rcv_isn
= TCP_SKB_CB(skb
)->seq
;
6073 tcp_rsk(req
)->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
6074 skb_mstamp_get(&tcp_rsk(req
)->snt_synack
);
6075 tcp_rsk(req
)->last_oow_ack_time
= 0;
6076 req
->mss
= rx_opt
->mss_clamp
;
6077 req
->ts_recent
= rx_opt
->saw_tstamp
? rx_opt
->rcv_tsval
: 0;
6078 ireq
->tstamp_ok
= rx_opt
->tstamp_ok
;
6079 ireq
->sack_ok
= rx_opt
->sack_ok
;
6080 ireq
->snd_wscale
= rx_opt
->snd_wscale
;
6081 ireq
->wscale_ok
= rx_opt
->wscale_ok
;
6084 ireq
->ir_rmt_port
= tcp_hdr(skb
)->source
;
6085 ireq
->ir_num
= ntohs(tcp_hdr(skb
)->dest
);
6086 ireq
->ir_mark
= inet_request_mark(sk
, skb
);
6089 struct request_sock
*inet_reqsk_alloc(const struct request_sock_ops
*ops
,
6090 struct sock
*sk_listener
,
6091 bool attach_listener
)
6093 struct request_sock
*req
= reqsk_alloc(ops
, sk_listener
,
6097 struct inet_request_sock
*ireq
= inet_rsk(req
);
6099 kmemcheck_annotate_bitfield(ireq
, flags
);
6101 atomic64_set(&ireq
->ir_cookie
, 0);
6102 ireq
->ireq_state
= TCP_NEW_SYN_RECV
;
6103 write_pnet(&ireq
->ireq_net
, sock_net(sk_listener
));
6104 ireq
->ireq_family
= sk_listener
->sk_family
;
6109 EXPORT_SYMBOL(inet_reqsk_alloc
);
6112 * Return true if a syncookie should be sent
6114 static bool tcp_syn_flood_action(const struct sock
*sk
,
6115 const struct sk_buff
*skb
,
6118 struct request_sock_queue
*queue
= &inet_csk(sk
)->icsk_accept_queue
;
6119 const char *msg
= "Dropping request";
6120 bool want_cookie
= false;
6122 #ifdef CONFIG_SYN_COOKIES
6123 if (sysctl_tcp_syncookies
) {
6124 msg
= "Sending cookies";
6126 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPREQQFULLDOCOOKIES
);
6129 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPREQQFULLDROP
);
6131 if (!queue
->synflood_warned
&&
6132 sysctl_tcp_syncookies
!= 2 &&
6133 xchg(&queue
->synflood_warned
, 1) == 0)
6134 pr_info("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6135 proto
, ntohs(tcp_hdr(skb
)->dest
), msg
);
6140 static void tcp_reqsk_record_syn(const struct sock
*sk
,
6141 struct request_sock
*req
,
6142 const struct sk_buff
*skb
)
6144 if (tcp_sk(sk
)->save_syn
) {
6145 u32 len
= skb_network_header_len(skb
) + tcp_hdrlen(skb
);
6148 copy
= kmalloc(len
+ sizeof(u32
), GFP_ATOMIC
);
6151 memcpy(©
[1], skb_network_header(skb
), len
);
6152 req
->saved_syn
= copy
;
6157 int tcp_conn_request(struct request_sock_ops
*rsk_ops
,
6158 const struct tcp_request_sock_ops
*af_ops
,
6159 struct sock
*sk
, struct sk_buff
*skb
)
6161 struct tcp_fastopen_cookie foc
= { .len
= -1 };
6162 __u32 isn
= TCP_SKB_CB(skb
)->tcp_tw_isn
;
6163 struct tcp_options_received tmp_opt
;
6164 struct tcp_sock
*tp
= tcp_sk(sk
);
6165 struct sock
*fastopen_sk
= NULL
;
6166 struct dst_entry
*dst
= NULL
;
6167 struct request_sock
*req
;
6168 bool want_cookie
= false;
6171 /* TW buckets are converted to open requests without
6172 * limitations, they conserve resources and peer is
6173 * evidently real one.
6175 if ((sysctl_tcp_syncookies
== 2 ||
6176 inet_csk_reqsk_queue_is_full(sk
)) && !isn
) {
6177 want_cookie
= tcp_syn_flood_action(sk
, skb
, rsk_ops
->slab_name
);
6183 /* Accept backlog is full. If we have already queued enough
6184 * of warm entries in syn queue, drop request. It is better than
6185 * clogging syn queue with openreqs with exponentially increasing
6188 if (sk_acceptq_is_full(sk
) && inet_csk_reqsk_queue_young(sk
) > 1) {
6189 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_LISTENOVERFLOWS
);
6193 req
= inet_reqsk_alloc(rsk_ops
, sk
, !want_cookie
);
6197 tcp_rsk(req
)->af_specific
= af_ops
;
6199 tcp_clear_options(&tmp_opt
);
6200 tmp_opt
.mss_clamp
= af_ops
->mss_clamp
;
6201 tmp_opt
.user_mss
= tp
->rx_opt
.user_mss
;
6202 tcp_parse_options(skb
, &tmp_opt
, 0, want_cookie
? NULL
: &foc
);
6204 if (want_cookie
&& !tmp_opt
.saw_tstamp
)
6205 tcp_clear_options(&tmp_opt
);
6207 tmp_opt
.tstamp_ok
= tmp_opt
.saw_tstamp
;
6208 tcp_openreq_init(req
, &tmp_opt
, skb
, sk
);
6210 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6211 inet_rsk(req
)->ir_iif
= inet_request_bound_dev_if(sk
, skb
);
6213 af_ops
->init_req(req
, sk
, skb
);
6215 if (security_inet_conn_request(sk
, skb
, req
))
6218 if (!want_cookie
&& !isn
) {
6219 /* VJ's idea. We save last timestamp seen
6220 * from the destination in peer table, when entering
6221 * state TIME-WAIT, and check against it before
6222 * accepting new connection request.
6224 * If "isn" is not zero, this request hit alive
6225 * timewait bucket, so that all the necessary checks
6226 * are made in the function processing timewait state.
6228 if (tcp_death_row
.sysctl_tw_recycle
) {
6231 dst
= af_ops
->route_req(sk
, &fl
, req
, &strict
);
6233 if (dst
&& strict
&&
6234 !tcp_peer_is_proven(req
, dst
, true,
6235 tmp_opt
.saw_tstamp
)) {
6236 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSPASSIVEREJECTED
);
6237 goto drop_and_release
;
6240 /* Kill the following clause, if you dislike this way. */
6241 else if (!sysctl_tcp_syncookies
&&
6242 (sysctl_max_syn_backlog
- inet_csk_reqsk_queue_len(sk
) <
6243 (sysctl_max_syn_backlog
>> 2)) &&
6244 !tcp_peer_is_proven(req
, dst
, false,
6245 tmp_opt
.saw_tstamp
)) {
6246 /* Without syncookies last quarter of
6247 * backlog is filled with destinations,
6248 * proven to be alive.
6249 * It means that we continue to communicate
6250 * to destinations, already remembered
6251 * to the moment of synflood.
6253 pr_drop_req(req
, ntohs(tcp_hdr(skb
)->source
),
6255 goto drop_and_release
;
6258 isn
= af_ops
->init_seq(skb
);
6261 dst
= af_ops
->route_req(sk
, &fl
, req
, NULL
);
6266 tcp_ecn_create_request(req
, skb
, sk
, dst
);
6269 isn
= cookie_init_sequence(af_ops
, sk
, skb
, &req
->mss
);
6270 req
->cookie_ts
= tmp_opt
.tstamp_ok
;
6271 if (!tmp_opt
.tstamp_ok
)
6272 inet_rsk(req
)->ecn_ok
= 0;
6275 tcp_rsk(req
)->snt_isn
= isn
;
6276 tcp_rsk(req
)->txhash
= net_tx_rndhash();
6277 tcp_openreq_init_rwin(req
, sk
, dst
);
6279 tcp_reqsk_record_syn(sk
, req
, skb
);
6280 fastopen_sk
= tcp_try_fastopen(sk
, skb
, req
, &foc
, dst
);
6283 af_ops
->send_synack(fastopen_sk
, dst
, &fl
, req
,
6285 /* Add the child socket directly into the accept queue */
6286 inet_csk_reqsk_queue_add(sk
, req
, fastopen_sk
);
6287 sk
->sk_data_ready(sk
);
6288 bh_unlock_sock(fastopen_sk
);
6289 sock_put(fastopen_sk
);
6291 tcp_rsk(req
)->tfo_listener
= false;
6293 inet_csk_reqsk_queue_hash_add(sk
, req
, TCP_TIMEOUT_INIT
);
6294 af_ops
->send_synack(sk
, dst
, &fl
, req
,
6295 &foc
, !want_cookie
);
6307 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_LISTENDROPS
);
6310 EXPORT_SYMBOL(tcp_conn_request
);