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>
73 #include <net/inet_common.h>
74 #include <linux/ipsec.h>
75 #include <asm/unaligned.h>
76 #include <net/netdma.h>
78 int sysctl_tcp_timestamps __read_mostly
= 1;
79 int sysctl_tcp_window_scaling __read_mostly
= 1;
80 int sysctl_tcp_sack __read_mostly
= 1;
81 int sysctl_tcp_fack __read_mostly
= 1;
82 int sysctl_tcp_reordering __read_mostly
= TCP_FASTRETRANS_THRESH
;
83 EXPORT_SYMBOL(sysctl_tcp_reordering
);
84 int sysctl_tcp_dsack __read_mostly
= 1;
85 int sysctl_tcp_app_win __read_mostly
= 31;
86 int sysctl_tcp_adv_win_scale __read_mostly
= 1;
87 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale
);
89 /* rfc5961 challenge ack rate limiting */
90 int sysctl_tcp_challenge_ack_limit
= 100;
92 int sysctl_tcp_stdurg __read_mostly
;
93 int sysctl_tcp_rfc1337 __read_mostly
;
94 int sysctl_tcp_max_orphans __read_mostly
= NR_FILE
;
95 int sysctl_tcp_frto __read_mostly
= 2;
97 int sysctl_tcp_thin_dupack __read_mostly
;
99 int sysctl_tcp_moderate_rcvbuf __read_mostly
= 1;
100 int sysctl_tcp_early_retrans __read_mostly
= 3;
102 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
103 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
104 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
105 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
106 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
107 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
108 #define FLAG_ECE 0x40 /* ECE in this ACK */
109 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
110 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
111 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
112 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
113 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
114 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
116 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
117 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
118 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
119 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
121 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
122 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
124 /* Adapt the MSS value used to make delayed ack decision to the
127 static void tcp_measure_rcv_mss(struct sock
*sk
, const struct sk_buff
*skb
)
129 struct inet_connection_sock
*icsk
= inet_csk(sk
);
130 const unsigned int lss
= icsk
->icsk_ack
.last_seg_size
;
133 icsk
->icsk_ack
.last_seg_size
= 0;
135 /* skb->len may jitter because of SACKs, even if peer
136 * sends good full-sized frames.
138 len
= skb_shinfo(skb
)->gso_size
? : skb
->len
;
139 if (len
>= icsk
->icsk_ack
.rcv_mss
) {
140 icsk
->icsk_ack
.rcv_mss
= len
;
142 /* Otherwise, we make more careful check taking into account,
143 * that SACKs block is variable.
145 * "len" is invariant segment length, including TCP header.
147 len
+= skb
->data
- skb_transport_header(skb
);
148 if (len
>= TCP_MSS_DEFAULT
+ sizeof(struct tcphdr
) ||
149 /* If PSH is not set, packet should be
150 * full sized, provided peer TCP is not badly broken.
151 * This observation (if it is correct 8)) allows
152 * to handle super-low mtu links fairly.
154 (len
>= TCP_MIN_MSS
+ sizeof(struct tcphdr
) &&
155 !(tcp_flag_word(tcp_hdr(skb
)) & TCP_REMNANT
))) {
156 /* Subtract also invariant (if peer is RFC compliant),
157 * tcp header plus fixed timestamp option length.
158 * Resulting "len" is MSS free of SACK jitter.
160 len
-= tcp_sk(sk
)->tcp_header_len
;
161 icsk
->icsk_ack
.last_seg_size
= len
;
163 icsk
->icsk_ack
.rcv_mss
= len
;
167 if (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
)
168 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED2
;
169 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
173 static void tcp_incr_quickack(struct sock
*sk
)
175 struct inet_connection_sock
*icsk
= inet_csk(sk
);
176 unsigned int quickacks
= tcp_sk(sk
)->rcv_wnd
/ (2 * icsk
->icsk_ack
.rcv_mss
);
180 if (quickacks
> icsk
->icsk_ack
.quick
)
181 icsk
->icsk_ack
.quick
= min(quickacks
, TCP_MAX_QUICKACKS
);
184 static void tcp_enter_quickack_mode(struct sock
*sk
)
186 struct inet_connection_sock
*icsk
= inet_csk(sk
);
187 tcp_incr_quickack(sk
);
188 icsk
->icsk_ack
.pingpong
= 0;
189 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
192 /* Send ACKs quickly, if "quick" count is not exhausted
193 * and the session is not interactive.
196 static inline bool tcp_in_quickack_mode(const struct sock
*sk
)
198 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
200 return icsk
->icsk_ack
.quick
&& !icsk
->icsk_ack
.pingpong
;
203 static inline void TCP_ECN_queue_cwr(struct tcp_sock
*tp
)
205 if (tp
->ecn_flags
& TCP_ECN_OK
)
206 tp
->ecn_flags
|= TCP_ECN_QUEUE_CWR
;
209 static inline void TCP_ECN_accept_cwr(struct tcp_sock
*tp
, const struct sk_buff
*skb
)
211 if (tcp_hdr(skb
)->cwr
)
212 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
215 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock
*tp
)
217 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
220 static inline void TCP_ECN_check_ce(struct tcp_sock
*tp
, const struct sk_buff
*skb
)
222 if (!(tp
->ecn_flags
& TCP_ECN_OK
))
225 switch (TCP_SKB_CB(skb
)->ip_dsfield
& INET_ECN_MASK
) {
226 case INET_ECN_NOT_ECT
:
227 /* Funny extension: if ECT is not set on a segment,
228 * and we already seen ECT on a previous segment,
229 * it is probably a retransmit.
231 if (tp
->ecn_flags
& TCP_ECN_SEEN
)
232 tcp_enter_quickack_mode((struct sock
*)tp
);
235 if (!(tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)) {
236 /* Better not delay acks, sender can have a very low cwnd */
237 tcp_enter_quickack_mode((struct sock
*)tp
);
238 tp
->ecn_flags
|= TCP_ECN_DEMAND_CWR
;
242 tp
->ecn_flags
|= TCP_ECN_SEEN
;
246 static inline void TCP_ECN_rcv_synack(struct tcp_sock
*tp
, const struct tcphdr
*th
)
248 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| th
->cwr
))
249 tp
->ecn_flags
&= ~TCP_ECN_OK
;
252 static inline void TCP_ECN_rcv_syn(struct tcp_sock
*tp
, const struct tcphdr
*th
)
254 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| !th
->cwr
))
255 tp
->ecn_flags
&= ~TCP_ECN_OK
;
258 static bool TCP_ECN_rcv_ecn_echo(const struct tcp_sock
*tp
, const struct tcphdr
*th
)
260 if (th
->ece
&& !th
->syn
&& (tp
->ecn_flags
& TCP_ECN_OK
))
265 /* Buffer size and advertised window tuning.
267 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
270 static void tcp_fixup_sndbuf(struct sock
*sk
)
272 int sndmem
= SKB_TRUESIZE(tcp_sk(sk
)->rx_opt
.mss_clamp
+ MAX_TCP_HEADER
);
274 sndmem
*= TCP_INIT_CWND
;
275 if (sk
->sk_sndbuf
< sndmem
)
276 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
279 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
281 * All tcp_full_space() is split to two parts: "network" buffer, allocated
282 * forward and advertised in receiver window (tp->rcv_wnd) and
283 * "application buffer", required to isolate scheduling/application
284 * latencies from network.
285 * window_clamp is maximal advertised window. It can be less than
286 * tcp_full_space(), in this case tcp_full_space() - window_clamp
287 * is reserved for "application" buffer. The less window_clamp is
288 * the smoother our behaviour from viewpoint of network, but the lower
289 * throughput and the higher sensitivity of the connection to losses. 8)
291 * rcv_ssthresh is more strict window_clamp used at "slow start"
292 * phase to predict further behaviour of this connection.
293 * It is used for two goals:
294 * - to enforce header prediction at sender, even when application
295 * requires some significant "application buffer". It is check #1.
296 * - to prevent pruning of receive queue because of misprediction
297 * of receiver window. Check #2.
299 * The scheme does not work when sender sends good segments opening
300 * window and then starts to feed us spaghetti. But it should work
301 * in common situations. Otherwise, we have to rely on queue collapsing.
304 /* Slow part of check#2. */
305 static int __tcp_grow_window(const struct sock
*sk
, const struct sk_buff
*skb
)
307 struct tcp_sock
*tp
= tcp_sk(sk
);
309 int truesize
= tcp_win_from_space(skb
->truesize
) >> 1;
310 int window
= tcp_win_from_space(sysctl_tcp_rmem
[2]) >> 1;
312 while (tp
->rcv_ssthresh
<= window
) {
313 if (truesize
<= skb
->len
)
314 return 2 * inet_csk(sk
)->icsk_ack
.rcv_mss
;
322 static void tcp_grow_window(struct sock
*sk
, const struct sk_buff
*skb
)
324 struct tcp_sock
*tp
= tcp_sk(sk
);
327 if (tp
->rcv_ssthresh
< tp
->window_clamp
&&
328 (int)tp
->rcv_ssthresh
< tcp_space(sk
) &&
329 !sk_under_memory_pressure(sk
)) {
332 /* Check #2. Increase window, if skb with such overhead
333 * will fit to rcvbuf in future.
335 if (tcp_win_from_space(skb
->truesize
) <= skb
->len
)
336 incr
= 2 * tp
->advmss
;
338 incr
= __tcp_grow_window(sk
, skb
);
341 incr
= max_t(int, incr
, 2 * skb
->len
);
342 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
+ incr
,
344 inet_csk(sk
)->icsk_ack
.quick
|= 1;
349 /* 3. Tuning rcvbuf, when connection enters established state. */
350 static void tcp_fixup_rcvbuf(struct sock
*sk
)
352 u32 mss
= tcp_sk(sk
)->advmss
;
355 rcvmem
= 2 * SKB_TRUESIZE(mss
+ MAX_TCP_HEADER
) *
356 tcp_default_init_rwnd(mss
);
358 if (sk
->sk_rcvbuf
< rcvmem
)
359 sk
->sk_rcvbuf
= min(rcvmem
, sysctl_tcp_rmem
[2]);
362 /* 4. Try to fixup all. It is made immediately after connection enters
365 void tcp_init_buffer_space(struct sock
*sk
)
367 struct tcp_sock
*tp
= tcp_sk(sk
);
370 if (!(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
))
371 tcp_fixup_rcvbuf(sk
);
372 if (!(sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
))
373 tcp_fixup_sndbuf(sk
);
375 tp
->rcvq_space
.space
= tp
->rcv_wnd
;
377 maxwin
= tcp_full_space(sk
);
379 if (tp
->window_clamp
>= maxwin
) {
380 tp
->window_clamp
= maxwin
;
382 if (sysctl_tcp_app_win
&& maxwin
> 4 * tp
->advmss
)
383 tp
->window_clamp
= max(maxwin
-
384 (maxwin
>> sysctl_tcp_app_win
),
388 /* Force reservation of one segment. */
389 if (sysctl_tcp_app_win
&&
390 tp
->window_clamp
> 2 * tp
->advmss
&&
391 tp
->window_clamp
+ tp
->advmss
> maxwin
)
392 tp
->window_clamp
= max(2 * tp
->advmss
, maxwin
- tp
->advmss
);
394 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, tp
->window_clamp
);
395 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
398 /* 5. Recalculate window clamp after socket hit its memory bounds. */
399 static void tcp_clamp_window(struct sock
*sk
)
401 struct tcp_sock
*tp
= tcp_sk(sk
);
402 struct inet_connection_sock
*icsk
= inet_csk(sk
);
404 icsk
->icsk_ack
.quick
= 0;
406 if (sk
->sk_rcvbuf
< sysctl_tcp_rmem
[2] &&
407 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
) &&
408 !sk_under_memory_pressure(sk
) &&
409 sk_memory_allocated(sk
) < sk_prot_mem_limits(sk
, 0)) {
410 sk
->sk_rcvbuf
= min(atomic_read(&sk
->sk_rmem_alloc
),
413 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
)
414 tp
->rcv_ssthresh
= min(tp
->window_clamp
, 2U * tp
->advmss
);
417 /* Initialize RCV_MSS value.
418 * RCV_MSS is an our guess about MSS used by the peer.
419 * We haven't any direct information about the MSS.
420 * It's better to underestimate the RCV_MSS rather than overestimate.
421 * Overestimations make us ACKing less frequently than needed.
422 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
424 void tcp_initialize_rcv_mss(struct sock
*sk
)
426 const struct tcp_sock
*tp
= tcp_sk(sk
);
427 unsigned int hint
= min_t(unsigned int, tp
->advmss
, tp
->mss_cache
);
429 hint
= min(hint
, tp
->rcv_wnd
/ 2);
430 hint
= min(hint
, TCP_MSS_DEFAULT
);
431 hint
= max(hint
, TCP_MIN_MSS
);
433 inet_csk(sk
)->icsk_ack
.rcv_mss
= hint
;
435 EXPORT_SYMBOL(tcp_initialize_rcv_mss
);
437 /* Receiver "autotuning" code.
439 * The algorithm for RTT estimation w/o timestamps is based on
440 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
441 * <http://public.lanl.gov/radiant/pubs.html#DRS>
443 * More detail on this code can be found at
444 * <http://staff.psc.edu/jheffner/>,
445 * though this reference is out of date. A new paper
448 static void tcp_rcv_rtt_update(struct tcp_sock
*tp
, u32 sample
, int win_dep
)
450 u32 new_sample
= tp
->rcv_rtt_est
.rtt
;
456 if (new_sample
!= 0) {
457 /* If we sample in larger samples in the non-timestamp
458 * case, we could grossly overestimate the RTT especially
459 * with chatty applications or bulk transfer apps which
460 * are stalled on filesystem I/O.
462 * Also, since we are only going for a minimum in the
463 * non-timestamp case, we do not smooth things out
464 * else with timestamps disabled convergence takes too
468 m
-= (new_sample
>> 3);
476 /* No previous measure. */
480 if (tp
->rcv_rtt_est
.rtt
!= new_sample
)
481 tp
->rcv_rtt_est
.rtt
= new_sample
;
484 static inline void tcp_rcv_rtt_measure(struct tcp_sock
*tp
)
486 if (tp
->rcv_rtt_est
.time
== 0)
488 if (before(tp
->rcv_nxt
, tp
->rcv_rtt_est
.seq
))
490 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rcv_rtt_est
.time
, 1);
493 tp
->rcv_rtt_est
.seq
= tp
->rcv_nxt
+ tp
->rcv_wnd
;
494 tp
->rcv_rtt_est
.time
= tcp_time_stamp
;
497 static inline void tcp_rcv_rtt_measure_ts(struct sock
*sk
,
498 const struct sk_buff
*skb
)
500 struct tcp_sock
*tp
= tcp_sk(sk
);
501 if (tp
->rx_opt
.rcv_tsecr
&&
502 (TCP_SKB_CB(skb
)->end_seq
-
503 TCP_SKB_CB(skb
)->seq
>= inet_csk(sk
)->icsk_ack
.rcv_mss
))
504 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
, 0);
508 * This function should be called every time data is copied to user space.
509 * It calculates the appropriate TCP receive buffer space.
511 void tcp_rcv_space_adjust(struct sock
*sk
)
513 struct tcp_sock
*tp
= tcp_sk(sk
);
517 if (tp
->rcvq_space
.time
== 0)
520 time
= tcp_time_stamp
- tp
->rcvq_space
.time
;
521 if (time
< (tp
->rcv_rtt_est
.rtt
>> 3) || tp
->rcv_rtt_est
.rtt
== 0)
524 space
= 2 * (tp
->copied_seq
- tp
->rcvq_space
.seq
);
526 space
= max(tp
->rcvq_space
.space
, space
);
528 if (tp
->rcvq_space
.space
!= space
) {
531 tp
->rcvq_space
.space
= space
;
533 if (sysctl_tcp_moderate_rcvbuf
&&
534 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
)) {
535 int new_clamp
= space
;
537 /* Receive space grows, normalize in order to
538 * take into account packet headers and sk_buff
539 * structure overhead.
544 rcvmem
= SKB_TRUESIZE(tp
->advmss
+ MAX_TCP_HEADER
);
545 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
548 space
= min(space
, sysctl_tcp_rmem
[2]);
549 if (space
> sk
->sk_rcvbuf
) {
550 sk
->sk_rcvbuf
= space
;
552 /* Make the window clamp follow along. */
553 tp
->window_clamp
= new_clamp
;
559 tp
->rcvq_space
.seq
= tp
->copied_seq
;
560 tp
->rcvq_space
.time
= tcp_time_stamp
;
563 /* There is something which you must keep in mind when you analyze the
564 * behavior of the tp->ato delayed ack timeout interval. When a
565 * connection starts up, we want to ack as quickly as possible. The
566 * problem is that "good" TCP's do slow start at the beginning of data
567 * transmission. The means that until we send the first few ACK's the
568 * sender will sit on his end and only queue most of his data, because
569 * he can only send snd_cwnd unacked packets at any given time. For
570 * each ACK we send, he increments snd_cwnd and transmits more of his
573 static void tcp_event_data_recv(struct sock
*sk
, struct sk_buff
*skb
)
575 struct tcp_sock
*tp
= tcp_sk(sk
);
576 struct inet_connection_sock
*icsk
= inet_csk(sk
);
579 inet_csk_schedule_ack(sk
);
581 tcp_measure_rcv_mss(sk
, skb
);
583 tcp_rcv_rtt_measure(tp
);
585 now
= tcp_time_stamp
;
587 if (!icsk
->icsk_ack
.ato
) {
588 /* The _first_ data packet received, initialize
589 * delayed ACK engine.
591 tcp_incr_quickack(sk
);
592 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
594 int m
= now
- icsk
->icsk_ack
.lrcvtime
;
596 if (m
<= TCP_ATO_MIN
/ 2) {
597 /* The fastest case is the first. */
598 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + TCP_ATO_MIN
/ 2;
599 } else if (m
< icsk
->icsk_ack
.ato
) {
600 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + m
;
601 if (icsk
->icsk_ack
.ato
> icsk
->icsk_rto
)
602 icsk
->icsk_ack
.ato
= icsk
->icsk_rto
;
603 } else if (m
> icsk
->icsk_rto
) {
604 /* Too long gap. Apparently sender failed to
605 * restart window, so that we send ACKs quickly.
607 tcp_incr_quickack(sk
);
611 icsk
->icsk_ack
.lrcvtime
= now
;
613 TCP_ECN_check_ce(tp
, skb
);
616 tcp_grow_window(sk
, skb
);
619 /* Called to compute a smoothed rtt estimate. The data fed to this
620 * routine either comes from timestamps, or from segments that were
621 * known _not_ to have been retransmitted [see Karn/Partridge
622 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
623 * piece by Van Jacobson.
624 * NOTE: the next three routines used to be one big routine.
625 * To save cycles in the RFC 1323 implementation it was better to break
626 * it up into three procedures. -- erics
628 static void tcp_rtt_estimator(struct sock
*sk
, const __u32 mrtt
)
630 struct tcp_sock
*tp
= tcp_sk(sk
);
631 long m
= mrtt
; /* RTT */
633 /* The following amusing code comes from Jacobson's
634 * article in SIGCOMM '88. Note that rtt and mdev
635 * are scaled versions of rtt and mean deviation.
636 * This is designed to be as fast as possible
637 * m stands for "measurement".
639 * On a 1990 paper the rto value is changed to:
640 * RTO = rtt + 4 * mdev
642 * Funny. This algorithm seems to be very broken.
643 * These formulae increase RTO, when it should be decreased, increase
644 * too slowly, when it should be increased quickly, decrease too quickly
645 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
646 * does not matter how to _calculate_ it. Seems, it was trap
647 * that VJ failed to avoid. 8)
652 m
-= (tp
->srtt
>> 3); /* m is now error in rtt est */
653 tp
->srtt
+= m
; /* rtt = 7/8 rtt + 1/8 new */
655 m
= -m
; /* m is now abs(error) */
656 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
657 /* This is similar to one of Eifel findings.
658 * Eifel blocks mdev updates when rtt decreases.
659 * This solution is a bit different: we use finer gain
660 * for mdev in this case (alpha*beta).
661 * Like Eifel it also prevents growth of rto,
662 * but also it limits too fast rto decreases,
663 * happening in pure Eifel.
668 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
670 tp
->mdev
+= m
; /* mdev = 3/4 mdev + 1/4 new */
671 if (tp
->mdev
> tp
->mdev_max
) {
672 tp
->mdev_max
= tp
->mdev
;
673 if (tp
->mdev_max
> tp
->rttvar
)
674 tp
->rttvar
= tp
->mdev_max
;
676 if (after(tp
->snd_una
, tp
->rtt_seq
)) {
677 if (tp
->mdev_max
< tp
->rttvar
)
678 tp
->rttvar
-= (tp
->rttvar
- tp
->mdev_max
) >> 2;
679 tp
->rtt_seq
= tp
->snd_nxt
;
680 tp
->mdev_max
= tcp_rto_min(sk
);
683 /* no previous measure. */
684 tp
->srtt
= m
<< 3; /* take the measured time to be rtt */
685 tp
->mdev
= m
<< 1; /* make sure rto = 3*rtt */
686 tp
->mdev_max
= tp
->rttvar
= max(tp
->mdev
, tcp_rto_min(sk
));
687 tp
->rtt_seq
= tp
->snd_nxt
;
691 /* Set the sk_pacing_rate to allow proper sizing of TSO packets.
692 * Note: TCP stack does not yet implement pacing.
693 * FQ packet scheduler can be used to implement cheap but effective
694 * TCP pacing, to smooth the burst on large writes when packets
695 * in flight is significantly lower than cwnd (or rwin)
697 static void tcp_update_pacing_rate(struct sock
*sk
)
699 const struct tcp_sock
*tp
= tcp_sk(sk
);
702 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
703 rate
= (u64
)tp
->mss_cache
* 2 * (HZ
<< 3);
705 rate
*= max(tp
->snd_cwnd
, tp
->packets_out
);
707 /* Correction for small srtt : minimum srtt being 8 (1 jiffy << 3),
708 * be conservative and assume srtt = 1 (125 us instead of 1.25 ms)
709 * We probably need usec resolution in the future.
710 * Note: This also takes care of possible srtt=0 case,
711 * when tcp_rtt_estimator() was not yet called.
713 if (tp
->srtt
> 8 + 2)
714 do_div(rate
, tp
->srtt
);
716 sk
->sk_pacing_rate
= min_t(u64
, rate
, ~0U);
719 /* Calculate rto without backoff. This is the second half of Van Jacobson's
720 * routine referred to above.
722 void tcp_set_rto(struct sock
*sk
)
724 const struct tcp_sock
*tp
= tcp_sk(sk
);
725 /* Old crap is replaced with new one. 8)
728 * 1. If rtt variance happened to be less 50msec, it is hallucination.
729 * It cannot be less due to utterly erratic ACK generation made
730 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
731 * to do with delayed acks, because at cwnd>2 true delack timeout
732 * is invisible. Actually, Linux-2.4 also generates erratic
733 * ACKs in some circumstances.
735 inet_csk(sk
)->icsk_rto
= __tcp_set_rto(tp
);
737 /* 2. Fixups made earlier cannot be right.
738 * If we do not estimate RTO correctly without them,
739 * all the algo is pure shit and should be replaced
740 * with correct one. It is exactly, which we pretend to do.
743 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
744 * guarantees that rto is higher.
749 __u32
tcp_init_cwnd(const struct tcp_sock
*tp
, const struct dst_entry
*dst
)
751 __u32 cwnd
= (dst
? dst_metric(dst
, RTAX_INITCWND
) : 0);
754 cwnd
= TCP_INIT_CWND
;
755 return min_t(__u32
, cwnd
, tp
->snd_cwnd_clamp
);
759 * Packet counting of FACK is based on in-order assumptions, therefore TCP
760 * disables it when reordering is detected
762 void tcp_disable_fack(struct tcp_sock
*tp
)
764 /* RFC3517 uses different metric in lost marker => reset on change */
766 tp
->lost_skb_hint
= NULL
;
767 tp
->rx_opt
.sack_ok
&= ~TCP_FACK_ENABLED
;
770 /* Take a notice that peer is sending D-SACKs */
771 static void tcp_dsack_seen(struct tcp_sock
*tp
)
773 tp
->rx_opt
.sack_ok
|= TCP_DSACK_SEEN
;
776 static void tcp_update_reordering(struct sock
*sk
, const int metric
,
779 struct tcp_sock
*tp
= tcp_sk(sk
);
780 if (metric
> tp
->reordering
) {
783 tp
->reordering
= min(TCP_MAX_REORDERING
, metric
);
785 /* This exciting event is worth to be remembered. 8) */
787 mib_idx
= LINUX_MIB_TCPTSREORDER
;
788 else if (tcp_is_reno(tp
))
789 mib_idx
= LINUX_MIB_TCPRENOREORDER
;
790 else if (tcp_is_fack(tp
))
791 mib_idx
= LINUX_MIB_TCPFACKREORDER
;
793 mib_idx
= LINUX_MIB_TCPSACKREORDER
;
795 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
796 #if FASTRETRANS_DEBUG > 1
797 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
798 tp
->rx_opt
.sack_ok
, inet_csk(sk
)->icsk_ca_state
,
802 tp
->undo_marker
? tp
->undo_retrans
: 0);
804 tcp_disable_fack(tp
);
808 tcp_disable_early_retrans(tp
);
811 /* This must be called before lost_out is incremented */
812 static void tcp_verify_retransmit_hint(struct tcp_sock
*tp
, struct sk_buff
*skb
)
814 if ((tp
->retransmit_skb_hint
== NULL
) ||
815 before(TCP_SKB_CB(skb
)->seq
,
816 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
817 tp
->retransmit_skb_hint
= skb
;
820 after(TCP_SKB_CB(skb
)->end_seq
, tp
->retransmit_high
))
821 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
824 static void tcp_skb_mark_lost(struct tcp_sock
*tp
, struct sk_buff
*skb
)
826 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
827 tcp_verify_retransmit_hint(tp
, skb
);
829 tp
->lost_out
+= tcp_skb_pcount(skb
);
830 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
834 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock
*tp
,
837 tcp_verify_retransmit_hint(tp
, skb
);
839 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
840 tp
->lost_out
+= tcp_skb_pcount(skb
);
841 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
845 /* This procedure tags the retransmission queue when SACKs arrive.
847 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
848 * Packets in queue with these bits set are counted in variables
849 * sacked_out, retrans_out and lost_out, correspondingly.
851 * Valid combinations are:
852 * Tag InFlight Description
853 * 0 1 - orig segment is in flight.
854 * S 0 - nothing flies, orig reached receiver.
855 * L 0 - nothing flies, orig lost by net.
856 * R 2 - both orig and retransmit are in flight.
857 * L|R 1 - orig is lost, retransmit is in flight.
858 * S|R 1 - orig reached receiver, retrans is still in flight.
859 * (L|S|R is logically valid, it could occur when L|R is sacked,
860 * but it is equivalent to plain S and code short-curcuits it to S.
861 * L|S is logically invalid, it would mean -1 packet in flight 8))
863 * These 6 states form finite state machine, controlled by the following events:
864 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
865 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
866 * 3. Loss detection event of two flavors:
867 * A. Scoreboard estimator decided the packet is lost.
868 * A'. Reno "three dupacks" marks head of queue lost.
869 * A''. Its FACK modification, head until snd.fack is lost.
870 * B. SACK arrives sacking SND.NXT at the moment, when the
871 * segment was retransmitted.
872 * 4. D-SACK added new rule: D-SACK changes any tag to S.
874 * It is pleasant to note, that state diagram turns out to be commutative,
875 * so that we are allowed not to be bothered by order of our actions,
876 * when multiple events arrive simultaneously. (see the function below).
878 * Reordering detection.
879 * --------------------
880 * Reordering metric is maximal distance, which a packet can be displaced
881 * in packet stream. With SACKs we can estimate it:
883 * 1. SACK fills old hole and the corresponding segment was not
884 * ever retransmitted -> reordering. Alas, we cannot use it
885 * when segment was retransmitted.
886 * 2. The last flaw is solved with D-SACK. D-SACK arrives
887 * for retransmitted and already SACKed segment -> reordering..
888 * Both of these heuristics are not used in Loss state, when we cannot
889 * account for retransmits accurately.
891 * SACK block validation.
892 * ----------------------
894 * SACK block range validation checks that the received SACK block fits to
895 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
896 * Note that SND.UNA is not included to the range though being valid because
897 * it means that the receiver is rather inconsistent with itself reporting
898 * SACK reneging when it should advance SND.UNA. Such SACK block this is
899 * perfectly valid, however, in light of RFC2018 which explicitly states
900 * that "SACK block MUST reflect the newest segment. Even if the newest
901 * segment is going to be discarded ...", not that it looks very clever
902 * in case of head skb. Due to potentional receiver driven attacks, we
903 * choose to avoid immediate execution of a walk in write queue due to
904 * reneging and defer head skb's loss recovery to standard loss recovery
905 * procedure that will eventually trigger (nothing forbids us doing this).
907 * Implements also blockage to start_seq wrap-around. Problem lies in the
908 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
909 * there's no guarantee that it will be before snd_nxt (n). The problem
910 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
913 * <- outs wnd -> <- wrapzone ->
914 * u e n u_w e_w s n_w
916 * |<------------+------+----- TCP seqno space --------------+---------->|
917 * ...-- <2^31 ->| |<--------...
918 * ...---- >2^31 ------>| |<--------...
920 * Current code wouldn't be vulnerable but it's better still to discard such
921 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
922 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
923 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
924 * equal to the ideal case (infinite seqno space without wrap caused issues).
926 * With D-SACK the lower bound is extended to cover sequence space below
927 * SND.UNA down to undo_marker, which is the last point of interest. Yet
928 * again, D-SACK block must not to go across snd_una (for the same reason as
929 * for the normal SACK blocks, explained above). But there all simplicity
930 * ends, TCP might receive valid D-SACKs below that. As long as they reside
931 * fully below undo_marker they do not affect behavior in anyway and can
932 * therefore be safely ignored. In rare cases (which are more or less
933 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
934 * fragmentation and packet reordering past skb's retransmission. To consider
935 * them correctly, the acceptable range must be extended even more though
936 * the exact amount is rather hard to quantify. However, tp->max_window can
937 * be used as an exaggerated estimate.
939 static bool tcp_is_sackblock_valid(struct tcp_sock
*tp
, bool is_dsack
,
940 u32 start_seq
, u32 end_seq
)
942 /* Too far in future, or reversed (interpretation is ambiguous) */
943 if (after(end_seq
, tp
->snd_nxt
) || !before(start_seq
, end_seq
))
946 /* Nasty start_seq wrap-around check (see comments above) */
947 if (!before(start_seq
, tp
->snd_nxt
))
950 /* In outstanding window? ...This is valid exit for D-SACKs too.
951 * start_seq == snd_una is non-sensical (see comments above)
953 if (after(start_seq
, tp
->snd_una
))
956 if (!is_dsack
|| !tp
->undo_marker
)
959 /* ...Then it's D-SACK, and must reside below snd_una completely */
960 if (after(end_seq
, tp
->snd_una
))
963 if (!before(start_seq
, tp
->undo_marker
))
967 if (!after(end_seq
, tp
->undo_marker
))
970 /* Undo_marker boundary crossing (overestimates a lot). Known already:
971 * start_seq < undo_marker and end_seq >= undo_marker.
973 return !before(start_seq
, end_seq
- tp
->max_window
);
976 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
977 * Event "B". Later note: FACK people cheated me again 8), we have to account
978 * for reordering! Ugly, but should help.
980 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
981 * less than what is now known to be received by the other end (derived from
982 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
983 * retransmitted skbs to avoid some costly processing per ACKs.
985 static void tcp_mark_lost_retrans(struct sock
*sk
)
987 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
988 struct tcp_sock
*tp
= tcp_sk(sk
);
991 u32 new_low_seq
= tp
->snd_nxt
;
992 u32 received_upto
= tcp_highest_sack_seq(tp
);
994 if (!tcp_is_fack(tp
) || !tp
->retrans_out
||
995 !after(received_upto
, tp
->lost_retrans_low
) ||
996 icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
999 tcp_for_write_queue(skb
, sk
) {
1000 u32 ack_seq
= TCP_SKB_CB(skb
)->ack_seq
;
1002 if (skb
== tcp_send_head(sk
))
1004 if (cnt
== tp
->retrans_out
)
1006 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1009 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
))
1012 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1013 * constraint here (see above) but figuring out that at
1014 * least tp->reordering SACK blocks reside between ack_seq
1015 * and received_upto is not easy task to do cheaply with
1016 * the available datastructures.
1018 * Whether FACK should check here for tp->reordering segs
1019 * in-between one could argue for either way (it would be
1020 * rather simple to implement as we could count fack_count
1021 * during the walk and do tp->fackets_out - fack_count).
1023 if (after(received_upto
, ack_seq
)) {
1024 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1025 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1027 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
1028 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSTRETRANSMIT
);
1030 if (before(ack_seq
, new_low_seq
))
1031 new_low_seq
= ack_seq
;
1032 cnt
+= tcp_skb_pcount(skb
);
1036 if (tp
->retrans_out
)
1037 tp
->lost_retrans_low
= new_low_seq
;
1040 static bool tcp_check_dsack(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1041 struct tcp_sack_block_wire
*sp
, int num_sacks
,
1044 struct tcp_sock
*tp
= tcp_sk(sk
);
1045 u32 start_seq_0
= get_unaligned_be32(&sp
[0].start_seq
);
1046 u32 end_seq_0
= get_unaligned_be32(&sp
[0].end_seq
);
1047 bool dup_sack
= false;
1049 if (before(start_seq_0
, TCP_SKB_CB(ack_skb
)->ack_seq
)) {
1052 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPDSACKRECV
);
1053 } else if (num_sacks
> 1) {
1054 u32 end_seq_1
= get_unaligned_be32(&sp
[1].end_seq
);
1055 u32 start_seq_1
= get_unaligned_be32(&sp
[1].start_seq
);
1057 if (!after(end_seq_0
, end_seq_1
) &&
1058 !before(start_seq_0
, start_seq_1
)) {
1061 NET_INC_STATS_BH(sock_net(sk
),
1062 LINUX_MIB_TCPDSACKOFORECV
);
1066 /* D-SACK for already forgotten data... Do dumb counting. */
1067 if (dup_sack
&& tp
->undo_marker
&& tp
->undo_retrans
&&
1068 !after(end_seq_0
, prior_snd_una
) &&
1069 after(end_seq_0
, tp
->undo_marker
))
1075 struct tcp_sacktag_state
{
1079 s32 rtt
; /* RTT measured by SACKing never-retransmitted data */
1082 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1083 * the incoming SACK may not exactly match but we can find smaller MSS
1084 * aligned portion of it that matches. Therefore we might need to fragment
1085 * which may fail and creates some hassle (caller must handle error case
1088 * FIXME: this could be merged to shift decision code
1090 static int tcp_match_skb_to_sack(struct sock
*sk
, struct sk_buff
*skb
,
1091 u32 start_seq
, u32 end_seq
)
1095 unsigned int pkt_len
;
1098 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1099 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1101 if (tcp_skb_pcount(skb
) > 1 && !in_sack
&&
1102 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
1103 mss
= tcp_skb_mss(skb
);
1104 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1107 pkt_len
= start_seq
- TCP_SKB_CB(skb
)->seq
;
1111 pkt_len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1116 /* Round if necessary so that SACKs cover only full MSSes
1117 * and/or the remaining small portion (if present)
1119 if (pkt_len
> mss
) {
1120 unsigned int new_len
= (pkt_len
/ mss
) * mss
;
1121 if (!in_sack
&& new_len
< pkt_len
) {
1123 if (new_len
> skb
->len
)
1128 err
= tcp_fragment(sk
, skb
, pkt_len
, mss
);
1136 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1137 static u8
tcp_sacktag_one(struct sock
*sk
,
1138 struct tcp_sacktag_state
*state
, u8 sacked
,
1139 u32 start_seq
, u32 end_seq
,
1140 int dup_sack
, int pcount
, u32 xmit_time
)
1142 struct tcp_sock
*tp
= tcp_sk(sk
);
1143 int fack_count
= state
->fack_count
;
1145 /* Account D-SACK for retransmitted packet. */
1146 if (dup_sack
&& (sacked
& TCPCB_RETRANS
)) {
1147 if (tp
->undo_marker
&& tp
->undo_retrans
&&
1148 after(end_seq
, tp
->undo_marker
))
1150 if (sacked
& TCPCB_SACKED_ACKED
)
1151 state
->reord
= min(fack_count
, state
->reord
);
1154 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1155 if (!after(end_seq
, tp
->snd_una
))
1158 if (!(sacked
& TCPCB_SACKED_ACKED
)) {
1159 if (sacked
& TCPCB_SACKED_RETRANS
) {
1160 /* If the segment is not tagged as lost,
1161 * we do not clear RETRANS, believing
1162 * that retransmission is still in flight.
1164 if (sacked
& TCPCB_LOST
) {
1165 sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1166 tp
->lost_out
-= pcount
;
1167 tp
->retrans_out
-= pcount
;
1170 if (!(sacked
& TCPCB_RETRANS
)) {
1171 /* New sack for not retransmitted frame,
1172 * which was in hole. It is reordering.
1174 if (before(start_seq
,
1175 tcp_highest_sack_seq(tp
)))
1176 state
->reord
= min(fack_count
,
1178 if (!after(end_seq
, tp
->high_seq
))
1179 state
->flag
|= FLAG_ORIG_SACK_ACKED
;
1180 /* Pick the earliest sequence sacked for RTT */
1182 state
->rtt
= tcp_time_stamp
- xmit_time
;
1185 if (sacked
& TCPCB_LOST
) {
1186 sacked
&= ~TCPCB_LOST
;
1187 tp
->lost_out
-= pcount
;
1191 sacked
|= TCPCB_SACKED_ACKED
;
1192 state
->flag
|= FLAG_DATA_SACKED
;
1193 tp
->sacked_out
+= pcount
;
1195 fack_count
+= pcount
;
1197 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1198 if (!tcp_is_fack(tp
) && (tp
->lost_skb_hint
!= NULL
) &&
1199 before(start_seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
))
1200 tp
->lost_cnt_hint
+= pcount
;
1202 if (fack_count
> tp
->fackets_out
)
1203 tp
->fackets_out
= fack_count
;
1206 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1207 * frames and clear it. undo_retrans is decreased above, L|R frames
1208 * are accounted above as well.
1210 if (dup_sack
&& (sacked
& TCPCB_SACKED_RETRANS
)) {
1211 sacked
&= ~TCPCB_SACKED_RETRANS
;
1212 tp
->retrans_out
-= pcount
;
1218 /* Shift newly-SACKed bytes from this skb to the immediately previous
1219 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1221 static bool tcp_shifted_skb(struct sock
*sk
, struct sk_buff
*skb
,
1222 struct tcp_sacktag_state
*state
,
1223 unsigned int pcount
, int shifted
, int mss
,
1226 struct tcp_sock
*tp
= tcp_sk(sk
);
1227 struct sk_buff
*prev
= tcp_write_queue_prev(sk
, skb
);
1228 u32 start_seq
= TCP_SKB_CB(skb
)->seq
; /* start of newly-SACKed */
1229 u32 end_seq
= start_seq
+ shifted
; /* end of newly-SACKed */
1233 /* Adjust counters and hints for the newly sacked sequence
1234 * range but discard the return value since prev is already
1235 * marked. We must tag the range first because the seq
1236 * advancement below implicitly advances
1237 * tcp_highest_sack_seq() when skb is highest_sack.
1239 tcp_sacktag_one(sk
, state
, TCP_SKB_CB(skb
)->sacked
,
1240 start_seq
, end_seq
, dup_sack
, pcount
,
1241 TCP_SKB_CB(skb
)->when
);
1243 if (skb
== tp
->lost_skb_hint
)
1244 tp
->lost_cnt_hint
+= pcount
;
1246 TCP_SKB_CB(prev
)->end_seq
+= shifted
;
1247 TCP_SKB_CB(skb
)->seq
+= shifted
;
1249 skb_shinfo(prev
)->gso_segs
+= pcount
;
1250 BUG_ON(skb_shinfo(skb
)->gso_segs
< pcount
);
1251 skb_shinfo(skb
)->gso_segs
-= pcount
;
1253 /* When we're adding to gso_segs == 1, gso_size will be zero,
1254 * in theory this shouldn't be necessary but as long as DSACK
1255 * code can come after this skb later on it's better to keep
1256 * setting gso_size to something.
1258 if (!skb_shinfo(prev
)->gso_size
) {
1259 skb_shinfo(prev
)->gso_size
= mss
;
1260 skb_shinfo(prev
)->gso_type
= sk
->sk_gso_type
;
1263 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1264 if (skb_shinfo(skb
)->gso_segs
<= 1) {
1265 skb_shinfo(skb
)->gso_size
= 0;
1266 skb_shinfo(skb
)->gso_type
= 0;
1269 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1270 TCP_SKB_CB(prev
)->sacked
|= (TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
);
1273 BUG_ON(!tcp_skb_pcount(skb
));
1274 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKSHIFTED
);
1278 /* Whole SKB was eaten :-) */
1280 if (skb
== tp
->retransmit_skb_hint
)
1281 tp
->retransmit_skb_hint
= prev
;
1282 if (skb
== tp
->lost_skb_hint
) {
1283 tp
->lost_skb_hint
= prev
;
1284 tp
->lost_cnt_hint
-= tcp_skb_pcount(prev
);
1287 TCP_SKB_CB(skb
)->tcp_flags
|= TCP_SKB_CB(prev
)->tcp_flags
;
1288 if (skb
== tcp_highest_sack(sk
))
1289 tcp_advance_highest_sack(sk
, skb
);
1291 tcp_unlink_write_queue(skb
, sk
);
1292 sk_wmem_free_skb(sk
, skb
);
1294 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKMERGED
);
1299 /* I wish gso_size would have a bit more sane initialization than
1300 * something-or-zero which complicates things
1302 static int tcp_skb_seglen(const struct sk_buff
*skb
)
1304 return tcp_skb_pcount(skb
) == 1 ? skb
->len
: tcp_skb_mss(skb
);
1307 /* Shifting pages past head area doesn't work */
1308 static int skb_can_shift(const struct sk_buff
*skb
)
1310 return !skb_headlen(skb
) && skb_is_nonlinear(skb
);
1313 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1316 static struct sk_buff
*tcp_shift_skb_data(struct sock
*sk
, struct sk_buff
*skb
,
1317 struct tcp_sacktag_state
*state
,
1318 u32 start_seq
, u32 end_seq
,
1321 struct tcp_sock
*tp
= tcp_sk(sk
);
1322 struct sk_buff
*prev
;
1328 if (!sk_can_gso(sk
))
1331 /* Normally R but no L won't result in plain S */
1333 (TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_RETRANS
)) == TCPCB_SACKED_RETRANS
)
1335 if (!skb_can_shift(skb
))
1337 /* This frame is about to be dropped (was ACKed). */
1338 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1341 /* Can only happen with delayed DSACK + discard craziness */
1342 if (unlikely(skb
== tcp_write_queue_head(sk
)))
1344 prev
= tcp_write_queue_prev(sk
, skb
);
1346 if ((TCP_SKB_CB(prev
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
)
1349 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1350 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1354 pcount
= tcp_skb_pcount(skb
);
1355 mss
= tcp_skb_seglen(skb
);
1357 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1358 * drop this restriction as unnecessary
1360 if (mss
!= tcp_skb_seglen(prev
))
1363 if (!after(TCP_SKB_CB(skb
)->end_seq
, start_seq
))
1365 /* CHECKME: This is non-MSS split case only?, this will
1366 * cause skipped skbs due to advancing loop btw, original
1367 * has that feature too
1369 if (tcp_skb_pcount(skb
) <= 1)
1372 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1374 /* TODO: head merge to next could be attempted here
1375 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1376 * though it might not be worth of the additional hassle
1378 * ...we can probably just fallback to what was done
1379 * previously. We could try merging non-SACKed ones
1380 * as well but it probably isn't going to buy off
1381 * because later SACKs might again split them, and
1382 * it would make skb timestamp tracking considerably
1388 len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1390 BUG_ON(len
> skb
->len
);
1392 /* MSS boundaries should be honoured or else pcount will
1393 * severely break even though it makes things bit trickier.
1394 * Optimize common case to avoid most of the divides
1396 mss
= tcp_skb_mss(skb
);
1398 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1399 * drop this restriction as unnecessary
1401 if (mss
!= tcp_skb_seglen(prev
))
1406 } else if (len
< mss
) {
1414 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1415 if (!after(TCP_SKB_CB(skb
)->seq
+ len
, tp
->snd_una
))
1418 if (!skb_shift(prev
, skb
, len
))
1420 if (!tcp_shifted_skb(sk
, skb
, state
, pcount
, len
, mss
, dup_sack
))
1423 /* Hole filled allows collapsing with the next as well, this is very
1424 * useful when hole on every nth skb pattern happens
1426 if (prev
== tcp_write_queue_tail(sk
))
1428 skb
= tcp_write_queue_next(sk
, prev
);
1430 if (!skb_can_shift(skb
) ||
1431 (skb
== tcp_send_head(sk
)) ||
1432 ((TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
) ||
1433 (mss
!= tcp_skb_seglen(skb
)))
1437 if (skb_shift(prev
, skb
, len
)) {
1438 pcount
+= tcp_skb_pcount(skb
);
1439 tcp_shifted_skb(sk
, skb
, state
, tcp_skb_pcount(skb
), len
, mss
, 0);
1443 state
->fack_count
+= pcount
;
1450 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKSHIFTFALLBACK
);
1454 static struct sk_buff
*tcp_sacktag_walk(struct sk_buff
*skb
, struct sock
*sk
,
1455 struct tcp_sack_block
*next_dup
,
1456 struct tcp_sacktag_state
*state
,
1457 u32 start_seq
, u32 end_seq
,
1460 struct tcp_sock
*tp
= tcp_sk(sk
);
1461 struct sk_buff
*tmp
;
1463 tcp_for_write_queue_from(skb
, sk
) {
1465 bool dup_sack
= dup_sack_in
;
1467 if (skb
== tcp_send_head(sk
))
1470 /* queue is in-order => we can short-circuit the walk early */
1471 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1474 if ((next_dup
!= NULL
) &&
1475 before(TCP_SKB_CB(skb
)->seq
, next_dup
->end_seq
)) {
1476 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1477 next_dup
->start_seq
,
1483 /* skb reference here is a bit tricky to get right, since
1484 * shifting can eat and free both this skb and the next,
1485 * so not even _safe variant of the loop is enough.
1488 tmp
= tcp_shift_skb_data(sk
, skb
, state
,
1489 start_seq
, end_seq
, dup_sack
);
1498 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1504 if (unlikely(in_sack
< 0))
1508 TCP_SKB_CB(skb
)->sacked
=
1511 TCP_SKB_CB(skb
)->sacked
,
1512 TCP_SKB_CB(skb
)->seq
,
1513 TCP_SKB_CB(skb
)->end_seq
,
1515 tcp_skb_pcount(skb
),
1516 TCP_SKB_CB(skb
)->when
);
1518 if (!before(TCP_SKB_CB(skb
)->seq
,
1519 tcp_highest_sack_seq(tp
)))
1520 tcp_advance_highest_sack(sk
, skb
);
1523 state
->fack_count
+= tcp_skb_pcount(skb
);
1528 /* Avoid all extra work that is being done by sacktag while walking in
1531 static struct sk_buff
*tcp_sacktag_skip(struct sk_buff
*skb
, struct sock
*sk
,
1532 struct tcp_sacktag_state
*state
,
1535 tcp_for_write_queue_from(skb
, sk
) {
1536 if (skb
== tcp_send_head(sk
))
1539 if (after(TCP_SKB_CB(skb
)->end_seq
, skip_to_seq
))
1542 state
->fack_count
+= tcp_skb_pcount(skb
);
1547 static struct sk_buff
*tcp_maybe_skipping_dsack(struct sk_buff
*skb
,
1549 struct tcp_sack_block
*next_dup
,
1550 struct tcp_sacktag_state
*state
,
1553 if (next_dup
== NULL
)
1556 if (before(next_dup
->start_seq
, skip_to_seq
)) {
1557 skb
= tcp_sacktag_skip(skb
, sk
, state
, next_dup
->start_seq
);
1558 skb
= tcp_sacktag_walk(skb
, sk
, NULL
, state
,
1559 next_dup
->start_seq
, next_dup
->end_seq
,
1566 static int tcp_sack_cache_ok(const struct tcp_sock
*tp
, const struct tcp_sack_block
*cache
)
1568 return cache
< tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1572 tcp_sacktag_write_queue(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1573 u32 prior_snd_una
, s32
*sack_rtt
)
1575 struct tcp_sock
*tp
= tcp_sk(sk
);
1576 const unsigned char *ptr
= (skb_transport_header(ack_skb
) +
1577 TCP_SKB_CB(ack_skb
)->sacked
);
1578 struct tcp_sack_block_wire
*sp_wire
= (struct tcp_sack_block_wire
*)(ptr
+2);
1579 struct tcp_sack_block sp
[TCP_NUM_SACKS
];
1580 struct tcp_sack_block
*cache
;
1581 struct tcp_sacktag_state state
;
1582 struct sk_buff
*skb
;
1583 int num_sacks
= min(TCP_NUM_SACKS
, (ptr
[1] - TCPOLEN_SACK_BASE
) >> 3);
1585 bool found_dup_sack
= false;
1587 int first_sack_index
;
1590 state
.reord
= tp
->packets_out
;
1593 if (!tp
->sacked_out
) {
1594 if (WARN_ON(tp
->fackets_out
))
1595 tp
->fackets_out
= 0;
1596 tcp_highest_sack_reset(sk
);
1599 found_dup_sack
= tcp_check_dsack(sk
, ack_skb
, sp_wire
,
1600 num_sacks
, prior_snd_una
);
1602 state
.flag
|= FLAG_DSACKING_ACK
;
1604 /* Eliminate too old ACKs, but take into
1605 * account more or less fresh ones, they can
1606 * contain valid SACK info.
1608 if (before(TCP_SKB_CB(ack_skb
)->ack_seq
, prior_snd_una
- tp
->max_window
))
1611 if (!tp
->packets_out
)
1615 first_sack_index
= 0;
1616 for (i
= 0; i
< num_sacks
; i
++) {
1617 bool dup_sack
= !i
&& found_dup_sack
;
1619 sp
[used_sacks
].start_seq
= get_unaligned_be32(&sp_wire
[i
].start_seq
);
1620 sp
[used_sacks
].end_seq
= get_unaligned_be32(&sp_wire
[i
].end_seq
);
1622 if (!tcp_is_sackblock_valid(tp
, dup_sack
,
1623 sp
[used_sacks
].start_seq
,
1624 sp
[used_sacks
].end_seq
)) {
1628 if (!tp
->undo_marker
)
1629 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDNOUNDO
;
1631 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDOLD
;
1633 /* Don't count olds caused by ACK reordering */
1634 if ((TCP_SKB_CB(ack_skb
)->ack_seq
!= tp
->snd_una
) &&
1635 !after(sp
[used_sacks
].end_seq
, tp
->snd_una
))
1637 mib_idx
= LINUX_MIB_TCPSACKDISCARD
;
1640 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
1642 first_sack_index
= -1;
1646 /* Ignore very old stuff early */
1647 if (!after(sp
[used_sacks
].end_seq
, prior_snd_una
))
1653 /* order SACK blocks to allow in order walk of the retrans queue */
1654 for (i
= used_sacks
- 1; i
> 0; i
--) {
1655 for (j
= 0; j
< i
; j
++) {
1656 if (after(sp
[j
].start_seq
, sp
[j
+ 1].start_seq
)) {
1657 swap(sp
[j
], sp
[j
+ 1]);
1659 /* Track where the first SACK block goes to */
1660 if (j
== first_sack_index
)
1661 first_sack_index
= j
+ 1;
1666 skb
= tcp_write_queue_head(sk
);
1667 state
.fack_count
= 0;
1670 if (!tp
->sacked_out
) {
1671 /* It's already past, so skip checking against it */
1672 cache
= tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1674 cache
= tp
->recv_sack_cache
;
1675 /* Skip empty blocks in at head of the cache */
1676 while (tcp_sack_cache_ok(tp
, cache
) && !cache
->start_seq
&&
1681 while (i
< used_sacks
) {
1682 u32 start_seq
= sp
[i
].start_seq
;
1683 u32 end_seq
= sp
[i
].end_seq
;
1684 bool dup_sack
= (found_dup_sack
&& (i
== first_sack_index
));
1685 struct tcp_sack_block
*next_dup
= NULL
;
1687 if (found_dup_sack
&& ((i
+ 1) == first_sack_index
))
1688 next_dup
= &sp
[i
+ 1];
1690 /* Skip too early cached blocks */
1691 while (tcp_sack_cache_ok(tp
, cache
) &&
1692 !before(start_seq
, cache
->end_seq
))
1695 /* Can skip some work by looking recv_sack_cache? */
1696 if (tcp_sack_cache_ok(tp
, cache
) && !dup_sack
&&
1697 after(end_seq
, cache
->start_seq
)) {
1700 if (before(start_seq
, cache
->start_seq
)) {
1701 skb
= tcp_sacktag_skip(skb
, sk
, &state
,
1703 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
,
1710 /* Rest of the block already fully processed? */
1711 if (!after(end_seq
, cache
->end_seq
))
1714 skb
= tcp_maybe_skipping_dsack(skb
, sk
, next_dup
,
1718 /* ...tail remains todo... */
1719 if (tcp_highest_sack_seq(tp
) == cache
->end_seq
) {
1720 /* ...but better entrypoint exists! */
1721 skb
= tcp_highest_sack(sk
);
1724 state
.fack_count
= tp
->fackets_out
;
1729 skb
= tcp_sacktag_skip(skb
, sk
, &state
, cache
->end_seq
);
1730 /* Check overlap against next cached too (past this one already) */
1735 if (!before(start_seq
, tcp_highest_sack_seq(tp
))) {
1736 skb
= tcp_highest_sack(sk
);
1739 state
.fack_count
= tp
->fackets_out
;
1741 skb
= tcp_sacktag_skip(skb
, sk
, &state
, start_seq
);
1744 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
, &state
,
1745 start_seq
, end_seq
, dup_sack
);
1751 /* Clear the head of the cache sack blocks so we can skip it next time */
1752 for (i
= 0; i
< ARRAY_SIZE(tp
->recv_sack_cache
) - used_sacks
; i
++) {
1753 tp
->recv_sack_cache
[i
].start_seq
= 0;
1754 tp
->recv_sack_cache
[i
].end_seq
= 0;
1756 for (j
= 0; j
< used_sacks
; j
++)
1757 tp
->recv_sack_cache
[i
++] = sp
[j
];
1759 tcp_mark_lost_retrans(sk
);
1761 tcp_verify_left_out(tp
);
1763 if ((state
.reord
< tp
->fackets_out
) &&
1764 ((inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Loss
) || tp
->undo_marker
))
1765 tcp_update_reordering(sk
, tp
->fackets_out
- state
.reord
, 0);
1769 #if FASTRETRANS_DEBUG > 0
1770 WARN_ON((int)tp
->sacked_out
< 0);
1771 WARN_ON((int)tp
->lost_out
< 0);
1772 WARN_ON((int)tp
->retrans_out
< 0);
1773 WARN_ON((int)tcp_packets_in_flight(tp
) < 0);
1775 *sack_rtt
= state
.rtt
;
1779 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1780 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1782 static bool tcp_limit_reno_sacked(struct tcp_sock
*tp
)
1786 holes
= max(tp
->lost_out
, 1U);
1787 holes
= min(holes
, tp
->packets_out
);
1789 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1790 tp
->sacked_out
= tp
->packets_out
- holes
;
1796 /* If we receive more dupacks than we expected counting segments
1797 * in assumption of absent reordering, interpret this as reordering.
1798 * The only another reason could be bug in receiver TCP.
1800 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1802 struct tcp_sock
*tp
= tcp_sk(sk
);
1803 if (tcp_limit_reno_sacked(tp
))
1804 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1807 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1809 static void tcp_add_reno_sack(struct sock
*sk
)
1811 struct tcp_sock
*tp
= tcp_sk(sk
);
1813 tcp_check_reno_reordering(sk
, 0);
1814 tcp_verify_left_out(tp
);
1817 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1819 static void tcp_remove_reno_sacks(struct sock
*sk
, int acked
)
1821 struct tcp_sock
*tp
= tcp_sk(sk
);
1824 /* One ACK acked hole. The rest eat duplicate ACKs. */
1825 if (acked
- 1 >= tp
->sacked_out
)
1828 tp
->sacked_out
-= acked
- 1;
1830 tcp_check_reno_reordering(sk
, acked
);
1831 tcp_verify_left_out(tp
);
1834 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1839 static void tcp_clear_retrans_partial(struct tcp_sock
*tp
)
1841 tp
->retrans_out
= 0;
1844 tp
->undo_marker
= 0;
1845 tp
->undo_retrans
= 0;
1848 void tcp_clear_retrans(struct tcp_sock
*tp
)
1850 tcp_clear_retrans_partial(tp
);
1852 tp
->fackets_out
= 0;
1856 /* Enter Loss state. If "how" is not zero, forget all SACK information
1857 * and reset tags completely, otherwise preserve SACKs. If receiver
1858 * dropped its ofo queue, we will know this due to reneging detection.
1860 void tcp_enter_loss(struct sock
*sk
, int how
)
1862 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1863 struct tcp_sock
*tp
= tcp_sk(sk
);
1864 struct sk_buff
*skb
;
1865 bool new_recovery
= false;
1867 /* Reduce ssthresh if it has not yet been made inside this window. */
1868 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
||
1869 !after(tp
->high_seq
, tp
->snd_una
) ||
1870 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
1871 new_recovery
= true;
1872 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1873 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1874 tcp_ca_event(sk
, CA_EVENT_LOSS
);
1877 tp
->snd_cwnd_cnt
= 0;
1878 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1880 tcp_clear_retrans_partial(tp
);
1882 if (tcp_is_reno(tp
))
1883 tcp_reset_reno_sack(tp
);
1885 tp
->undo_marker
= tp
->snd_una
;
1888 tp
->fackets_out
= 0;
1890 tcp_clear_all_retrans_hints(tp
);
1892 tcp_for_write_queue(skb
, sk
) {
1893 if (skb
== tcp_send_head(sk
))
1896 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
1897 tp
->undo_marker
= 0;
1898 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
1899 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) || how
) {
1900 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
1901 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1902 tp
->lost_out
+= tcp_skb_pcount(skb
);
1903 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
1906 tcp_verify_left_out(tp
);
1908 /* Timeout in disordered state after receiving substantial DUPACKs
1909 * suggests that the degree of reordering is over-estimated.
1911 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
&&
1912 tp
->sacked_out
>= sysctl_tcp_reordering
)
1913 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1914 sysctl_tcp_reordering
);
1915 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1916 tp
->high_seq
= tp
->snd_nxt
;
1917 TCP_ECN_queue_cwr(tp
);
1919 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1920 * loss recovery is underway except recurring timeout(s) on
1921 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1923 tp
->frto
= sysctl_tcp_frto
&&
1924 (new_recovery
|| icsk
->icsk_retransmits
) &&
1925 !inet_csk(sk
)->icsk_mtup
.probe_size
;
1928 /* If ACK arrived pointing to a remembered SACK, it means that our
1929 * remembered SACKs do not reflect real state of receiver i.e.
1930 * receiver _host_ is heavily congested (or buggy).
1932 * Do processing similar to RTO timeout.
1934 static bool tcp_check_sack_reneging(struct sock
*sk
, int flag
)
1936 if (flag
& FLAG_SACK_RENEGING
) {
1937 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1938 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSACKRENEGING
);
1940 tcp_enter_loss(sk
, 1);
1941 icsk
->icsk_retransmits
++;
1942 tcp_retransmit_skb(sk
, tcp_write_queue_head(sk
));
1943 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1944 icsk
->icsk_rto
, TCP_RTO_MAX
);
1950 static inline int tcp_fackets_out(const struct tcp_sock
*tp
)
1952 return tcp_is_reno(tp
) ? tp
->sacked_out
+ 1 : tp
->fackets_out
;
1955 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
1956 * counter when SACK is enabled (without SACK, sacked_out is used for
1959 * Instead, with FACK TCP uses fackets_out that includes both SACKed
1960 * segments up to the highest received SACK block so far and holes in
1963 * With reordering, holes may still be in flight, so RFC3517 recovery
1964 * uses pure sacked_out (total number of SACKed segments) even though
1965 * it violates the RFC that uses duplicate ACKs, often these are equal
1966 * but when e.g. out-of-window ACKs or packet duplication occurs,
1967 * they differ. Since neither occurs due to loss, TCP should really
1970 static inline int tcp_dupack_heuristics(const struct tcp_sock
*tp
)
1972 return tcp_is_fack(tp
) ? tp
->fackets_out
: tp
->sacked_out
+ 1;
1975 static bool tcp_pause_early_retransmit(struct sock
*sk
, int flag
)
1977 struct tcp_sock
*tp
= tcp_sk(sk
);
1978 unsigned long delay
;
1980 /* Delay early retransmit and entering fast recovery for
1981 * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
1982 * available, or RTO is scheduled to fire first.
1984 if (sysctl_tcp_early_retrans
< 2 || sysctl_tcp_early_retrans
> 3 ||
1985 (flag
& FLAG_ECE
) || !tp
->srtt
)
1988 delay
= max_t(unsigned long, (tp
->srtt
>> 5), msecs_to_jiffies(2));
1989 if (!time_after(inet_csk(sk
)->icsk_timeout
, (jiffies
+ delay
)))
1992 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_EARLY_RETRANS
, delay
,
1997 /* Linux NewReno/SACK/FACK/ECN state machine.
1998 * --------------------------------------
2000 * "Open" Normal state, no dubious events, fast path.
2001 * "Disorder" In all the respects it is "Open",
2002 * but requires a bit more attention. It is entered when
2003 * we see some SACKs or dupacks. It is split of "Open"
2004 * mainly to move some processing from fast path to slow one.
2005 * "CWR" CWND was reduced due to some Congestion Notification event.
2006 * It can be ECN, ICMP source quench, local device congestion.
2007 * "Recovery" CWND was reduced, we are fast-retransmitting.
2008 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2010 * tcp_fastretrans_alert() is entered:
2011 * - each incoming ACK, if state is not "Open"
2012 * - when arrived ACK is unusual, namely:
2017 * Counting packets in flight is pretty simple.
2019 * in_flight = packets_out - left_out + retrans_out
2021 * packets_out is SND.NXT-SND.UNA counted in packets.
2023 * retrans_out is number of retransmitted segments.
2025 * left_out is number of segments left network, but not ACKed yet.
2027 * left_out = sacked_out + lost_out
2029 * sacked_out: Packets, which arrived to receiver out of order
2030 * and hence not ACKed. With SACKs this number is simply
2031 * amount of SACKed data. Even without SACKs
2032 * it is easy to give pretty reliable estimate of this number,
2033 * counting duplicate ACKs.
2035 * lost_out: Packets lost by network. TCP has no explicit
2036 * "loss notification" feedback from network (for now).
2037 * It means that this number can be only _guessed_.
2038 * Actually, it is the heuristics to predict lossage that
2039 * distinguishes different algorithms.
2041 * F.e. after RTO, when all the queue is considered as lost,
2042 * lost_out = packets_out and in_flight = retrans_out.
2044 * Essentially, we have now two algorithms counting
2047 * FACK: It is the simplest heuristics. As soon as we decided
2048 * that something is lost, we decide that _all_ not SACKed
2049 * packets until the most forward SACK are lost. I.e.
2050 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2051 * It is absolutely correct estimate, if network does not reorder
2052 * packets. And it loses any connection to reality when reordering
2053 * takes place. We use FACK by default until reordering
2054 * is suspected on the path to this destination.
2056 * NewReno: when Recovery is entered, we assume that one segment
2057 * is lost (classic Reno). While we are in Recovery and
2058 * a partial ACK arrives, we assume that one more packet
2059 * is lost (NewReno). This heuristics are the same in NewReno
2062 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2063 * deflation etc. CWND is real congestion window, never inflated, changes
2064 * only according to classic VJ rules.
2066 * Really tricky (and requiring careful tuning) part of algorithm
2067 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2068 * The first determines the moment _when_ we should reduce CWND and,
2069 * hence, slow down forward transmission. In fact, it determines the moment
2070 * when we decide that hole is caused by loss, rather than by a reorder.
2072 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2073 * holes, caused by lost packets.
2075 * And the most logically complicated part of algorithm is undo
2076 * heuristics. We detect false retransmits due to both too early
2077 * fast retransmit (reordering) and underestimated RTO, analyzing
2078 * timestamps and D-SACKs. When we detect that some segments were
2079 * retransmitted by mistake and CWND reduction was wrong, we undo
2080 * window reduction and abort recovery phase. This logic is hidden
2081 * inside several functions named tcp_try_undo_<something>.
2084 /* This function decides, when we should leave Disordered state
2085 * and enter Recovery phase, reducing congestion window.
2087 * Main question: may we further continue forward transmission
2088 * with the same cwnd?
2090 static bool tcp_time_to_recover(struct sock
*sk
, int flag
)
2092 struct tcp_sock
*tp
= tcp_sk(sk
);
2095 /* Trick#1: The loss is proven. */
2099 /* Not-A-Trick#2 : Classic rule... */
2100 if (tcp_dupack_heuristics(tp
) > tp
->reordering
)
2103 /* Trick#4: It is still not OK... But will it be useful to delay
2106 packets_out
= tp
->packets_out
;
2107 if (packets_out
<= tp
->reordering
&&
2108 tp
->sacked_out
>= max_t(__u32
, packets_out
/2, sysctl_tcp_reordering
) &&
2109 !tcp_may_send_now(sk
)) {
2110 /* We have nothing to send. This connection is limited
2111 * either by receiver window or by application.
2116 /* If a thin stream is detected, retransmit after first
2117 * received dupack. Employ only if SACK is supported in order
2118 * to avoid possible corner-case series of spurious retransmissions
2119 * Use only if there are no unsent data.
2121 if ((tp
->thin_dupack
|| sysctl_tcp_thin_dupack
) &&
2122 tcp_stream_is_thin(tp
) && tcp_dupack_heuristics(tp
) > 1 &&
2123 tcp_is_sack(tp
) && !tcp_send_head(sk
))
2126 /* Trick#6: TCP early retransmit, per RFC5827. To avoid spurious
2127 * retransmissions due to small network reorderings, we implement
2128 * Mitigation A.3 in the RFC and delay the retransmission for a short
2129 * interval if appropriate.
2131 if (tp
->do_early_retrans
&& !tp
->retrans_out
&& tp
->sacked_out
&&
2132 (tp
->packets_out
>= (tp
->sacked_out
+ 1) && tp
->packets_out
< 4) &&
2133 !tcp_may_send_now(sk
))
2134 return !tcp_pause_early_retransmit(sk
, flag
);
2139 /* Detect loss in event "A" above by marking head of queue up as lost.
2140 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2141 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2142 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2143 * the maximum SACKed segments to pass before reaching this limit.
2145 static void tcp_mark_head_lost(struct sock
*sk
, int packets
, int mark_head
)
2147 struct tcp_sock
*tp
= tcp_sk(sk
);
2148 struct sk_buff
*skb
;
2152 /* Use SACK to deduce losses of new sequences sent during recovery */
2153 const u32 loss_high
= tcp_is_sack(tp
) ? tp
->snd_nxt
: tp
->high_seq
;
2155 WARN_ON(packets
> tp
->packets_out
);
2156 if (tp
->lost_skb_hint
) {
2157 skb
= tp
->lost_skb_hint
;
2158 cnt
= tp
->lost_cnt_hint
;
2159 /* Head already handled? */
2160 if (mark_head
&& skb
!= tcp_write_queue_head(sk
))
2163 skb
= tcp_write_queue_head(sk
);
2167 tcp_for_write_queue_from(skb
, sk
) {
2168 if (skb
== tcp_send_head(sk
))
2170 /* TODO: do this better */
2171 /* this is not the most efficient way to do this... */
2172 tp
->lost_skb_hint
= skb
;
2173 tp
->lost_cnt_hint
= cnt
;
2175 if (after(TCP_SKB_CB(skb
)->end_seq
, loss_high
))
2179 if (tcp_is_fack(tp
) || tcp_is_reno(tp
) ||
2180 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2181 cnt
+= tcp_skb_pcount(skb
);
2183 if (cnt
> packets
) {
2184 if ((tcp_is_sack(tp
) && !tcp_is_fack(tp
)) ||
2185 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
) ||
2186 (oldcnt
>= packets
))
2189 mss
= skb_shinfo(skb
)->gso_size
;
2190 err
= tcp_fragment(sk
, skb
, (packets
- oldcnt
) * mss
, mss
);
2196 tcp_skb_mark_lost(tp
, skb
);
2201 tcp_verify_left_out(tp
);
2204 /* Account newly detected lost packet(s) */
2206 static void tcp_update_scoreboard(struct sock
*sk
, int fast_rexmit
)
2208 struct tcp_sock
*tp
= tcp_sk(sk
);
2210 if (tcp_is_reno(tp
)) {
2211 tcp_mark_head_lost(sk
, 1, 1);
2212 } else if (tcp_is_fack(tp
)) {
2213 int lost
= tp
->fackets_out
- tp
->reordering
;
2216 tcp_mark_head_lost(sk
, lost
, 0);
2218 int sacked_upto
= tp
->sacked_out
- tp
->reordering
;
2219 if (sacked_upto
>= 0)
2220 tcp_mark_head_lost(sk
, sacked_upto
, 0);
2221 else if (fast_rexmit
)
2222 tcp_mark_head_lost(sk
, 1, 1);
2226 /* CWND moderation, preventing bursts due to too big ACKs
2227 * in dubious situations.
2229 static inline void tcp_moderate_cwnd(struct tcp_sock
*tp
)
2231 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
2232 tcp_packets_in_flight(tp
) + tcp_max_burst(tp
));
2233 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2236 /* Nothing was retransmitted or returned timestamp is less
2237 * than timestamp of the first retransmission.
2239 static inline bool tcp_packet_delayed(const struct tcp_sock
*tp
)
2241 return !tp
->retrans_stamp
||
2242 (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2243 before(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
));
2246 /* Undo procedures. */
2248 #if FASTRETRANS_DEBUG > 1
2249 static void DBGUNDO(struct sock
*sk
, const char *msg
)
2251 struct tcp_sock
*tp
= tcp_sk(sk
);
2252 struct inet_sock
*inet
= inet_sk(sk
);
2254 if (sk
->sk_family
== AF_INET
) {
2255 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2257 &inet
->inet_daddr
, ntohs(inet
->inet_dport
),
2258 tp
->snd_cwnd
, tcp_left_out(tp
),
2259 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2262 #if IS_ENABLED(CONFIG_IPV6)
2263 else if (sk
->sk_family
== AF_INET6
) {
2264 struct ipv6_pinfo
*np
= inet6_sk(sk
);
2265 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2267 &np
->daddr
, ntohs(inet
->inet_dport
),
2268 tp
->snd_cwnd
, tcp_left_out(tp
),
2269 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2275 #define DBGUNDO(x...) do { } while (0)
2278 static void tcp_undo_cwnd_reduction(struct sock
*sk
, bool unmark_loss
)
2280 struct tcp_sock
*tp
= tcp_sk(sk
);
2283 struct sk_buff
*skb
;
2285 tcp_for_write_queue(skb
, sk
) {
2286 if (skb
== tcp_send_head(sk
))
2288 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2291 tcp_clear_all_retrans_hints(tp
);
2294 if (tp
->prior_ssthresh
) {
2295 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2297 if (icsk
->icsk_ca_ops
->undo_cwnd
)
2298 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
2300 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
<< 1);
2302 if (tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
2303 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
2304 TCP_ECN_withdraw_cwr(tp
);
2307 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2309 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2310 tp
->undo_marker
= 0;
2313 static inline bool tcp_may_undo(const struct tcp_sock
*tp
)
2315 return tp
->undo_marker
&& (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
2318 /* People celebrate: "We love our President!" */
2319 static bool tcp_try_undo_recovery(struct sock
*sk
)
2321 struct tcp_sock
*tp
= tcp_sk(sk
);
2323 if (tcp_may_undo(tp
)) {
2326 /* Happy end! We did not retransmit anything
2327 * or our original transmission succeeded.
2329 DBGUNDO(sk
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
2330 tcp_undo_cwnd_reduction(sk
, false);
2331 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
2332 mib_idx
= LINUX_MIB_TCPLOSSUNDO
;
2334 mib_idx
= LINUX_MIB_TCPFULLUNDO
;
2336 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2338 if (tp
->snd_una
== tp
->high_seq
&& tcp_is_reno(tp
)) {
2339 /* Hold old state until something *above* high_seq
2340 * is ACKed. For Reno it is MUST to prevent false
2341 * fast retransmits (RFC2582). SACK TCP is safe. */
2342 tcp_moderate_cwnd(tp
);
2345 tcp_set_ca_state(sk
, TCP_CA_Open
);
2349 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2350 static bool tcp_try_undo_dsack(struct sock
*sk
)
2352 struct tcp_sock
*tp
= tcp_sk(sk
);
2354 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
2355 DBGUNDO(sk
, "D-SACK");
2356 tcp_undo_cwnd_reduction(sk
, false);
2357 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPDSACKUNDO
);
2363 /* We can clear retrans_stamp when there are no retransmissions in the
2364 * window. It would seem that it is trivially available for us in
2365 * tp->retrans_out, however, that kind of assumptions doesn't consider
2366 * what will happen if errors occur when sending retransmission for the
2367 * second time. ...It could the that such segment has only
2368 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2369 * the head skb is enough except for some reneging corner cases that
2370 * are not worth the effort.
2372 * Main reason for all this complexity is the fact that connection dying
2373 * time now depends on the validity of the retrans_stamp, in particular,
2374 * that successive retransmissions of a segment must not advance
2375 * retrans_stamp under any conditions.
2377 static bool tcp_any_retrans_done(const struct sock
*sk
)
2379 const struct tcp_sock
*tp
= tcp_sk(sk
);
2380 struct sk_buff
*skb
;
2382 if (tp
->retrans_out
)
2385 skb
= tcp_write_queue_head(sk
);
2386 if (unlikely(skb
&& TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
))
2392 /* Undo during loss recovery after partial ACK or using F-RTO. */
2393 static bool tcp_try_undo_loss(struct sock
*sk
, bool frto_undo
)
2395 struct tcp_sock
*tp
= tcp_sk(sk
);
2397 if (frto_undo
|| tcp_may_undo(tp
)) {
2398 tcp_undo_cwnd_reduction(sk
, true);
2400 DBGUNDO(sk
, "partial loss");
2401 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSSUNDO
);
2403 NET_INC_STATS_BH(sock_net(sk
),
2404 LINUX_MIB_TCPSPURIOUSRTOS
);
2405 inet_csk(sk
)->icsk_retransmits
= 0;
2406 if (frto_undo
|| tcp_is_sack(tp
))
2407 tcp_set_ca_state(sk
, TCP_CA_Open
);
2413 /* The cwnd reduction in CWR and Recovery use the PRR algorithm
2414 * https://datatracker.ietf.org/doc/draft-ietf-tcpm-proportional-rate-reduction/
2415 * It computes the number of packets to send (sndcnt) based on packets newly
2417 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2418 * cwnd reductions across a full RTT.
2419 * 2) If packets in flight is lower than ssthresh (such as due to excess
2420 * losses and/or application stalls), do not perform any further cwnd
2421 * reductions, but instead slow start up to ssthresh.
2423 static void tcp_init_cwnd_reduction(struct sock
*sk
, const bool set_ssthresh
)
2425 struct tcp_sock
*tp
= tcp_sk(sk
);
2427 tp
->high_seq
= tp
->snd_nxt
;
2428 tp
->tlp_high_seq
= 0;
2429 tp
->snd_cwnd_cnt
= 0;
2430 tp
->prior_cwnd
= tp
->snd_cwnd
;
2431 tp
->prr_delivered
= 0;
2434 tp
->snd_ssthresh
= inet_csk(sk
)->icsk_ca_ops
->ssthresh(sk
);
2435 TCP_ECN_queue_cwr(tp
);
2438 static void tcp_cwnd_reduction(struct sock
*sk
, const int prior_unsacked
,
2441 struct tcp_sock
*tp
= tcp_sk(sk
);
2443 int delta
= tp
->snd_ssthresh
- tcp_packets_in_flight(tp
);
2444 int newly_acked_sacked
= prior_unsacked
-
2445 (tp
->packets_out
- tp
->sacked_out
);
2447 tp
->prr_delivered
+= newly_acked_sacked
;
2448 if (tcp_packets_in_flight(tp
) > tp
->snd_ssthresh
) {
2449 u64 dividend
= (u64
)tp
->snd_ssthresh
* tp
->prr_delivered
+
2451 sndcnt
= div_u64(dividend
, tp
->prior_cwnd
) - tp
->prr_out
;
2453 sndcnt
= min_t(int, delta
,
2454 max_t(int, tp
->prr_delivered
- tp
->prr_out
,
2455 newly_acked_sacked
) + 1);
2458 sndcnt
= max(sndcnt
, (fast_rexmit
? 1 : 0));
2459 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + sndcnt
;
2462 static inline void tcp_end_cwnd_reduction(struct sock
*sk
)
2464 struct tcp_sock
*tp
= tcp_sk(sk
);
2466 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2467 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_CWR
||
2468 (tp
->undo_marker
&& tp
->snd_ssthresh
< TCP_INFINITE_SSTHRESH
)) {
2469 tp
->snd_cwnd
= tp
->snd_ssthresh
;
2470 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2472 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
2475 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2476 void tcp_enter_cwr(struct sock
*sk
, const int set_ssthresh
)
2478 struct tcp_sock
*tp
= tcp_sk(sk
);
2480 tp
->prior_ssthresh
= 0;
2481 if (inet_csk(sk
)->icsk_ca_state
< TCP_CA_CWR
) {
2482 tp
->undo_marker
= 0;
2483 tcp_init_cwnd_reduction(sk
, set_ssthresh
);
2484 tcp_set_ca_state(sk
, TCP_CA_CWR
);
2488 static void tcp_try_keep_open(struct sock
*sk
)
2490 struct tcp_sock
*tp
= tcp_sk(sk
);
2491 int state
= TCP_CA_Open
;
2493 if (tcp_left_out(tp
) || tcp_any_retrans_done(sk
))
2494 state
= TCP_CA_Disorder
;
2496 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
2497 tcp_set_ca_state(sk
, state
);
2498 tp
->high_seq
= tp
->snd_nxt
;
2502 static void tcp_try_to_open(struct sock
*sk
, int flag
, const int prior_unsacked
)
2504 struct tcp_sock
*tp
= tcp_sk(sk
);
2506 tcp_verify_left_out(tp
);
2508 if (!tcp_any_retrans_done(sk
))
2509 tp
->retrans_stamp
= 0;
2511 if (flag
& FLAG_ECE
)
2512 tcp_enter_cwr(sk
, 1);
2514 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
2515 tcp_try_keep_open(sk
);
2517 tcp_cwnd_reduction(sk
, prior_unsacked
, 0);
2521 static void tcp_mtup_probe_failed(struct sock
*sk
)
2523 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2525 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
2526 icsk
->icsk_mtup
.probe_size
= 0;
2529 static void tcp_mtup_probe_success(struct sock
*sk
)
2531 struct tcp_sock
*tp
= tcp_sk(sk
);
2532 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2534 /* FIXME: breaks with very large cwnd */
2535 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2536 tp
->snd_cwnd
= tp
->snd_cwnd
*
2537 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
2538 icsk
->icsk_mtup
.probe_size
;
2539 tp
->snd_cwnd_cnt
= 0;
2540 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2541 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2543 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
2544 icsk
->icsk_mtup
.probe_size
= 0;
2545 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
2548 /* Do a simple retransmit without using the backoff mechanisms in
2549 * tcp_timer. This is used for path mtu discovery.
2550 * The socket is already locked here.
2552 void tcp_simple_retransmit(struct sock
*sk
)
2554 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2555 struct tcp_sock
*tp
= tcp_sk(sk
);
2556 struct sk_buff
*skb
;
2557 unsigned int mss
= tcp_current_mss(sk
);
2558 u32 prior_lost
= tp
->lost_out
;
2560 tcp_for_write_queue(skb
, sk
) {
2561 if (skb
== tcp_send_head(sk
))
2563 if (tcp_skb_seglen(skb
) > mss
&&
2564 !(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
2565 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2566 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
2567 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2569 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
2573 tcp_clear_retrans_hints_partial(tp
);
2575 if (prior_lost
== tp
->lost_out
)
2578 if (tcp_is_reno(tp
))
2579 tcp_limit_reno_sacked(tp
);
2581 tcp_verify_left_out(tp
);
2583 /* Don't muck with the congestion window here.
2584 * Reason is that we do not increase amount of _data_
2585 * in network, but units changed and effective
2586 * cwnd/ssthresh really reduced now.
2588 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
2589 tp
->high_seq
= tp
->snd_nxt
;
2590 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2591 tp
->prior_ssthresh
= 0;
2592 tp
->undo_marker
= 0;
2593 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2595 tcp_xmit_retransmit_queue(sk
);
2597 EXPORT_SYMBOL(tcp_simple_retransmit
);
2599 static void tcp_enter_recovery(struct sock
*sk
, bool ece_ack
)
2601 struct tcp_sock
*tp
= tcp_sk(sk
);
2604 if (tcp_is_reno(tp
))
2605 mib_idx
= LINUX_MIB_TCPRENORECOVERY
;
2607 mib_idx
= LINUX_MIB_TCPSACKRECOVERY
;
2609 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2611 tp
->prior_ssthresh
= 0;
2612 tp
->undo_marker
= tp
->snd_una
;
2613 tp
->undo_retrans
= tp
->retrans_out
;
2615 if (inet_csk(sk
)->icsk_ca_state
< TCP_CA_CWR
) {
2617 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2618 tcp_init_cwnd_reduction(sk
, true);
2620 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
2623 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2624 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2626 static void tcp_process_loss(struct sock
*sk
, int flag
, bool is_dupack
)
2628 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2629 struct tcp_sock
*tp
= tcp_sk(sk
);
2630 bool recovered
= !before(tp
->snd_una
, tp
->high_seq
);
2632 if (tp
->frto
) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2633 if (flag
& FLAG_ORIG_SACK_ACKED
) {
2634 /* Step 3.b. A timeout is spurious if not all data are
2635 * lost, i.e., never-retransmitted data are (s)acked.
2637 tcp_try_undo_loss(sk
, true);
2640 if (after(tp
->snd_nxt
, tp
->high_seq
) &&
2641 (flag
& FLAG_DATA_SACKED
|| is_dupack
)) {
2642 tp
->frto
= 0; /* Loss was real: 2nd part of step 3.a */
2643 } else if (flag
& FLAG_SND_UNA_ADVANCED
&& !recovered
) {
2644 tp
->high_seq
= tp
->snd_nxt
;
2645 __tcp_push_pending_frames(sk
, tcp_current_mss(sk
),
2647 if (after(tp
->snd_nxt
, tp
->high_seq
))
2648 return; /* Step 2.b */
2654 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2655 icsk
->icsk_retransmits
= 0;
2656 tcp_try_undo_recovery(sk
);
2659 if (flag
& FLAG_DATA_ACKED
)
2660 icsk
->icsk_retransmits
= 0;
2661 if (tcp_is_reno(tp
)) {
2662 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2663 * delivered. Lower inflight to clock out (re)tranmissions.
2665 if (after(tp
->snd_nxt
, tp
->high_seq
) && is_dupack
)
2666 tcp_add_reno_sack(sk
);
2667 else if (flag
& FLAG_SND_UNA_ADVANCED
)
2668 tcp_reset_reno_sack(tp
);
2670 if (tcp_try_undo_loss(sk
, false))
2672 tcp_xmit_retransmit_queue(sk
);
2675 /* Undo during fast recovery after partial ACK. */
2676 static bool tcp_try_undo_partial(struct sock
*sk
, const int acked
,
2677 const int prior_unsacked
)
2679 struct tcp_sock
*tp
= tcp_sk(sk
);
2681 if (tp
->undo_marker
&& tcp_packet_delayed(tp
)) {
2682 /* Plain luck! Hole if filled with delayed
2683 * packet, rather than with a retransmit.
2685 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
2687 /* We are getting evidence that the reordering degree is higher
2688 * than we realized. If there are no retransmits out then we
2689 * can undo. Otherwise we clock out new packets but do not
2690 * mark more packets lost or retransmit more.
2692 if (tp
->retrans_out
) {
2693 tcp_cwnd_reduction(sk
, prior_unsacked
, 0);
2697 if (!tcp_any_retrans_done(sk
))
2698 tp
->retrans_stamp
= 0;
2700 DBGUNDO(sk
, "partial recovery");
2701 tcp_undo_cwnd_reduction(sk
, true);
2702 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPARTIALUNDO
);
2703 tcp_try_keep_open(sk
);
2709 /* Process an event, which can update packets-in-flight not trivially.
2710 * Main goal of this function is to calculate new estimate for left_out,
2711 * taking into account both packets sitting in receiver's buffer and
2712 * packets lost by network.
2714 * Besides that it does CWND reduction, when packet loss is detected
2715 * and changes state of machine.
2717 * It does _not_ decide what to send, it is made in function
2718 * tcp_xmit_retransmit_queue().
2720 static void tcp_fastretrans_alert(struct sock
*sk
, const int acked
,
2721 const int prior_unsacked
,
2722 bool is_dupack
, int flag
)
2724 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2725 struct tcp_sock
*tp
= tcp_sk(sk
);
2726 bool do_lost
= is_dupack
|| ((flag
& FLAG_DATA_SACKED
) &&
2727 (tcp_fackets_out(tp
) > tp
->reordering
));
2728 int fast_rexmit
= 0;
2730 if (WARN_ON(!tp
->packets_out
&& tp
->sacked_out
))
2732 if (WARN_ON(!tp
->sacked_out
&& tp
->fackets_out
))
2733 tp
->fackets_out
= 0;
2735 /* Now state machine starts.
2736 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2737 if (flag
& FLAG_ECE
)
2738 tp
->prior_ssthresh
= 0;
2740 /* B. In all the states check for reneging SACKs. */
2741 if (tcp_check_sack_reneging(sk
, flag
))
2744 /* C. Check consistency of the current state. */
2745 tcp_verify_left_out(tp
);
2747 /* D. Check state exit conditions. State can be terminated
2748 * when high_seq is ACKed. */
2749 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
2750 WARN_ON(tp
->retrans_out
!= 0);
2751 tp
->retrans_stamp
= 0;
2752 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
2753 switch (icsk
->icsk_ca_state
) {
2755 /* CWR is to be held something *above* high_seq
2756 * is ACKed for CWR bit to reach receiver. */
2757 if (tp
->snd_una
!= tp
->high_seq
) {
2758 tcp_end_cwnd_reduction(sk
);
2759 tcp_set_ca_state(sk
, TCP_CA_Open
);
2763 case TCP_CA_Recovery
:
2764 if (tcp_is_reno(tp
))
2765 tcp_reset_reno_sack(tp
);
2766 if (tcp_try_undo_recovery(sk
))
2768 tcp_end_cwnd_reduction(sk
);
2773 /* E. Process state. */
2774 switch (icsk
->icsk_ca_state
) {
2775 case TCP_CA_Recovery
:
2776 if (!(flag
& FLAG_SND_UNA_ADVANCED
)) {
2777 if (tcp_is_reno(tp
) && is_dupack
)
2778 tcp_add_reno_sack(sk
);
2780 if (tcp_try_undo_partial(sk
, acked
, prior_unsacked
))
2782 /* Partial ACK arrived. Force fast retransmit. */
2783 do_lost
= tcp_is_reno(tp
) ||
2784 tcp_fackets_out(tp
) > tp
->reordering
;
2786 if (tcp_try_undo_dsack(sk
)) {
2787 tcp_try_keep_open(sk
);
2792 tcp_process_loss(sk
, flag
, is_dupack
);
2793 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
2795 /* Fall through to processing in Open state. */
2797 if (tcp_is_reno(tp
)) {
2798 if (flag
& FLAG_SND_UNA_ADVANCED
)
2799 tcp_reset_reno_sack(tp
);
2801 tcp_add_reno_sack(sk
);
2804 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
)
2805 tcp_try_undo_dsack(sk
);
2807 if (!tcp_time_to_recover(sk
, flag
)) {
2808 tcp_try_to_open(sk
, flag
, prior_unsacked
);
2812 /* MTU probe failure: don't reduce cwnd */
2813 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
2814 icsk
->icsk_mtup
.probe_size
&&
2815 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
2816 tcp_mtup_probe_failed(sk
);
2817 /* Restores the reduction we did in tcp_mtup_probe() */
2819 tcp_simple_retransmit(sk
);
2823 /* Otherwise enter Recovery state */
2824 tcp_enter_recovery(sk
, (flag
& FLAG_ECE
));
2829 tcp_update_scoreboard(sk
, fast_rexmit
);
2830 tcp_cwnd_reduction(sk
, prior_unsacked
, fast_rexmit
);
2831 tcp_xmit_retransmit_queue(sk
);
2834 static inline bool tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
2835 s32 seq_rtt
, s32 sack_rtt
)
2837 const struct tcp_sock
*tp
= tcp_sk(sk
);
2839 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2840 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2841 * Karn's algorithm forbids taking RTT if some retransmitted data
2842 * is acked (RFC6298).
2844 if (flag
& FLAG_RETRANS_DATA_ACKED
)
2850 /* RTTM Rule: A TSecr value received in a segment is used to
2851 * update the averaged RTT measurement only if the segment
2852 * acknowledges some new data, i.e., only if it advances the
2853 * left edge of the send window.
2854 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2856 if (seq_rtt
< 0 && tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
2857 seq_rtt
= tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
;
2862 tcp_rtt_estimator(sk
, seq_rtt
);
2865 /* RFC6298: only reset backoff on valid RTT measurement. */
2866 inet_csk(sk
)->icsk_backoff
= 0;
2870 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2871 static void tcp_synack_rtt_meas(struct sock
*sk
, struct request_sock
*req
)
2873 struct tcp_sock
*tp
= tcp_sk(sk
);
2876 if (tp
->lsndtime
&& !tp
->total_retrans
)
2877 seq_rtt
= tcp_time_stamp
- tp
->lsndtime
;
2878 tcp_ack_update_rtt(sk
, FLAG_SYN_ACKED
, seq_rtt
, -1);
2881 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
, u32 in_flight
)
2883 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2884 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, in_flight
);
2885 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_time_stamp
;
2888 /* Restart timer after forward progress on connection.
2889 * RFC2988 recommends to restart timer to now+rto.
2891 void tcp_rearm_rto(struct sock
*sk
)
2893 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2894 struct tcp_sock
*tp
= tcp_sk(sk
);
2896 /* If the retrans timer is currently being used by Fast Open
2897 * for SYN-ACK retrans purpose, stay put.
2899 if (tp
->fastopen_rsk
)
2902 if (!tp
->packets_out
) {
2903 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
2905 u32 rto
= inet_csk(sk
)->icsk_rto
;
2906 /* Offset the time elapsed after installing regular RTO */
2907 if (icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
||
2908 icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
) {
2909 struct sk_buff
*skb
= tcp_write_queue_head(sk
);
2910 const u32 rto_time_stamp
= TCP_SKB_CB(skb
)->when
+ rto
;
2911 s32 delta
= (s32
)(rto_time_stamp
- tcp_time_stamp
);
2912 /* delta may not be positive if the socket is locked
2913 * when the retrans timer fires and is rescheduled.
2918 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
, rto
,
2923 /* This function is called when the delayed ER timer fires. TCP enters
2924 * fast recovery and performs fast-retransmit.
2926 void tcp_resume_early_retransmit(struct sock
*sk
)
2928 struct tcp_sock
*tp
= tcp_sk(sk
);
2932 /* Stop if ER is disabled after the delayed ER timer is scheduled */
2933 if (!tp
->do_early_retrans
)
2936 tcp_enter_recovery(sk
, false);
2937 tcp_update_scoreboard(sk
, 1);
2938 tcp_xmit_retransmit_queue(sk
);
2941 /* If we get here, the whole TSO packet has not been acked. */
2942 static u32
tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
)
2944 struct tcp_sock
*tp
= tcp_sk(sk
);
2947 BUG_ON(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
));
2949 packets_acked
= tcp_skb_pcount(skb
);
2950 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
2952 packets_acked
-= tcp_skb_pcount(skb
);
2954 if (packets_acked
) {
2955 BUG_ON(tcp_skb_pcount(skb
) == 0);
2956 BUG_ON(!before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
));
2959 return packets_acked
;
2962 /* Remove acknowledged frames from the retransmission queue. If our packet
2963 * is before the ack sequence we can discard it as it's confirmed to have
2964 * arrived at the other end.
2966 static int tcp_clean_rtx_queue(struct sock
*sk
, int prior_fackets
,
2967 u32 prior_snd_una
, s32 sack_rtt
)
2969 struct tcp_sock
*tp
= tcp_sk(sk
);
2970 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2971 struct sk_buff
*skb
;
2972 u32 now
= tcp_time_stamp
;
2973 int fully_acked
= true;
2976 u32 reord
= tp
->packets_out
;
2977 u32 prior_sacked
= tp
->sacked_out
;
2979 s32 ca_seq_rtt
= -1;
2980 ktime_t last_ackt
= net_invalid_timestamp();
2982 while ((skb
= tcp_write_queue_head(sk
)) && skb
!= tcp_send_head(sk
)) {
2983 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
2985 u8 sacked
= scb
->sacked
;
2987 /* Determine how many packets and what bytes were acked, tso and else */
2988 if (after(scb
->end_seq
, tp
->snd_una
)) {
2989 if (tcp_skb_pcount(skb
) == 1 ||
2990 !after(tp
->snd_una
, scb
->seq
))
2993 acked_pcount
= tcp_tso_acked(sk
, skb
);
2997 fully_acked
= false;
2999 acked_pcount
= tcp_skb_pcount(skb
);
3002 if (sacked
& TCPCB_RETRANS
) {
3003 if (sacked
& TCPCB_SACKED_RETRANS
)
3004 tp
->retrans_out
-= acked_pcount
;
3005 flag
|= FLAG_RETRANS_DATA_ACKED
;
3007 ca_seq_rtt
= now
- scb
->when
;
3008 last_ackt
= skb
->tstamp
;
3010 seq_rtt
= ca_seq_rtt
;
3012 if (!(sacked
& TCPCB_SACKED_ACKED
))
3013 reord
= min(pkts_acked
, reord
);
3014 if (!after(scb
->end_seq
, tp
->high_seq
))
3015 flag
|= FLAG_ORIG_SACK_ACKED
;
3018 if (sacked
& TCPCB_SACKED_ACKED
)
3019 tp
->sacked_out
-= acked_pcount
;
3020 if (sacked
& TCPCB_LOST
)
3021 tp
->lost_out
-= acked_pcount
;
3023 tp
->packets_out
-= acked_pcount
;
3024 pkts_acked
+= acked_pcount
;
3026 /* Initial outgoing SYN's get put onto the write_queue
3027 * just like anything else we transmit. It is not
3028 * true data, and if we misinform our callers that
3029 * this ACK acks real data, we will erroneously exit
3030 * connection startup slow start one packet too
3031 * quickly. This is severely frowned upon behavior.
3033 if (!(scb
->tcp_flags
& TCPHDR_SYN
)) {
3034 flag
|= FLAG_DATA_ACKED
;
3036 flag
|= FLAG_SYN_ACKED
;
3037 tp
->retrans_stamp
= 0;
3043 tcp_unlink_write_queue(skb
, sk
);
3044 sk_wmem_free_skb(sk
, skb
);
3045 if (skb
== tp
->retransmit_skb_hint
)
3046 tp
->retransmit_skb_hint
= NULL
;
3047 if (skb
== tp
->lost_skb_hint
)
3048 tp
->lost_skb_hint
= NULL
;
3051 if (likely(between(tp
->snd_up
, prior_snd_una
, tp
->snd_una
)))
3052 tp
->snd_up
= tp
->snd_una
;
3054 if (skb
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
3055 flag
|= FLAG_SACK_RENEGING
;
3057 if (tcp_ack_update_rtt(sk
, flag
, seq_rtt
, sack_rtt
) ||
3058 (flag
& FLAG_ACKED
))
3061 if (flag
& FLAG_ACKED
) {
3062 const struct tcp_congestion_ops
*ca_ops
3063 = inet_csk(sk
)->icsk_ca_ops
;
3065 if (unlikely(icsk
->icsk_mtup
.probe_size
&&
3066 !after(tp
->mtu_probe
.probe_seq_end
, tp
->snd_una
))) {
3067 tcp_mtup_probe_success(sk
);
3070 if (tcp_is_reno(tp
)) {
3071 tcp_remove_reno_sacks(sk
, pkts_acked
);
3075 /* Non-retransmitted hole got filled? That's reordering */
3076 if (reord
< prior_fackets
)
3077 tcp_update_reordering(sk
, tp
->fackets_out
- reord
, 0);
3079 delta
= tcp_is_fack(tp
) ? pkts_acked
:
3080 prior_sacked
- tp
->sacked_out
;
3081 tp
->lost_cnt_hint
-= min(tp
->lost_cnt_hint
, delta
);
3084 tp
->fackets_out
-= min(pkts_acked
, tp
->fackets_out
);
3086 if (ca_ops
->pkts_acked
) {
3089 /* Is the ACK triggering packet unambiguous? */
3090 if (!(flag
& FLAG_RETRANS_DATA_ACKED
)) {
3091 /* High resolution needed and available? */
3092 if (ca_ops
->flags
& TCP_CONG_RTT_STAMP
&&
3093 !ktime_equal(last_ackt
,
3094 net_invalid_timestamp()))
3095 rtt_us
= ktime_us_delta(ktime_get_real(),
3097 else if (ca_seq_rtt
>= 0)
3098 rtt_us
= jiffies_to_usecs(ca_seq_rtt
);
3101 ca_ops
->pkts_acked(sk
, pkts_acked
, rtt_us
);
3105 #if FASTRETRANS_DEBUG > 0
3106 WARN_ON((int)tp
->sacked_out
< 0);
3107 WARN_ON((int)tp
->lost_out
< 0);
3108 WARN_ON((int)tp
->retrans_out
< 0);
3109 if (!tp
->packets_out
&& tcp_is_sack(tp
)) {
3110 icsk
= inet_csk(sk
);
3112 pr_debug("Leak l=%u %d\n",
3113 tp
->lost_out
, icsk
->icsk_ca_state
);
3116 if (tp
->sacked_out
) {
3117 pr_debug("Leak s=%u %d\n",
3118 tp
->sacked_out
, icsk
->icsk_ca_state
);
3121 if (tp
->retrans_out
) {
3122 pr_debug("Leak r=%u %d\n",
3123 tp
->retrans_out
, icsk
->icsk_ca_state
);
3124 tp
->retrans_out
= 0;
3131 static void tcp_ack_probe(struct sock
*sk
)
3133 const struct tcp_sock
*tp
= tcp_sk(sk
);
3134 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3136 /* Was it a usable window open? */
3138 if (!after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
, tcp_wnd_end(tp
))) {
3139 icsk
->icsk_backoff
= 0;
3140 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
3141 /* Socket must be waked up by subsequent tcp_data_snd_check().
3142 * This function is not for random using!
3145 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3146 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
3151 static inline bool tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
3153 return !(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
3154 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
;
3157 /* Decide wheather to run the increase function of congestion control. */
3158 static inline bool tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
3160 if (tcp_in_cwnd_reduction(sk
))
3163 /* If reordering is high then always grow cwnd whenever data is
3164 * delivered regardless of its ordering. Otherwise stay conservative
3165 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3166 * new SACK or ECE mark may first advance cwnd here and later reduce
3167 * cwnd in tcp_fastretrans_alert() based on more states.
3169 if (tcp_sk(sk
)->reordering
> sysctl_tcp_reordering
)
3170 return flag
& FLAG_FORWARD_PROGRESS
;
3172 return flag
& FLAG_DATA_ACKED
;
3175 /* Check that window update is acceptable.
3176 * The function assumes that snd_una<=ack<=snd_next.
3178 static inline bool tcp_may_update_window(const struct tcp_sock
*tp
,
3179 const u32 ack
, const u32 ack_seq
,
3182 return after(ack
, tp
->snd_una
) ||
3183 after(ack_seq
, tp
->snd_wl1
) ||
3184 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
);
3187 /* Update our send window.
3189 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3190 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3192 static int tcp_ack_update_window(struct sock
*sk
, const struct sk_buff
*skb
, u32 ack
,
3195 struct tcp_sock
*tp
= tcp_sk(sk
);
3197 u32 nwin
= ntohs(tcp_hdr(skb
)->window
);
3199 if (likely(!tcp_hdr(skb
)->syn
))
3200 nwin
<<= tp
->rx_opt
.snd_wscale
;
3202 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
3203 flag
|= FLAG_WIN_UPDATE
;
3204 tcp_update_wl(tp
, ack_seq
);
3206 if (tp
->snd_wnd
!= nwin
) {
3209 /* Note, it is the only place, where
3210 * fast path is recovered for sending TCP.
3213 tcp_fast_path_check(sk
);
3215 if (nwin
> tp
->max_window
) {
3216 tp
->max_window
= nwin
;
3217 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
3227 /* RFC 5961 7 [ACK Throttling] */
3228 static void tcp_send_challenge_ack(struct sock
*sk
)
3230 /* unprotected vars, we dont care of overwrites */
3231 static u32 challenge_timestamp
;
3232 static unsigned int challenge_count
;
3233 u32 now
= jiffies
/ HZ
;
3235 if (now
!= challenge_timestamp
) {
3236 challenge_timestamp
= now
;
3237 challenge_count
= 0;
3239 if (++challenge_count
<= sysctl_tcp_challenge_ack_limit
) {
3240 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPCHALLENGEACK
);
3245 static void tcp_store_ts_recent(struct tcp_sock
*tp
)
3247 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
3248 tp
->rx_opt
.ts_recent_stamp
= get_seconds();
3251 static void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
3253 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
3254 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3255 * extra check below makes sure this can only happen
3256 * for pure ACK frames. -DaveM
3258 * Not only, also it occurs for expired timestamps.
3261 if (tcp_paws_check(&tp
->rx_opt
, 0))
3262 tcp_store_ts_recent(tp
);
3266 /* This routine deals with acks during a TLP episode.
3267 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3269 static void tcp_process_tlp_ack(struct sock
*sk
, u32 ack
, int flag
)
3271 struct tcp_sock
*tp
= tcp_sk(sk
);
3272 bool is_tlp_dupack
= (ack
== tp
->tlp_high_seq
) &&
3273 !(flag
& (FLAG_SND_UNA_ADVANCED
|
3274 FLAG_NOT_DUP
| FLAG_DATA_SACKED
));
3276 /* Mark the end of TLP episode on receiving TLP dupack or when
3277 * ack is after tlp_high_seq.
3279 if (is_tlp_dupack
) {
3280 tp
->tlp_high_seq
= 0;
3284 if (after(ack
, tp
->tlp_high_seq
)) {
3285 tp
->tlp_high_seq
= 0;
3286 /* Don't reduce cwnd if DSACK arrives for TLP retrans. */
3287 if (!(flag
& FLAG_DSACKING_ACK
)) {
3288 tcp_init_cwnd_reduction(sk
, true);
3289 tcp_set_ca_state(sk
, TCP_CA_CWR
);
3290 tcp_end_cwnd_reduction(sk
);
3291 tcp_set_ca_state(sk
, TCP_CA_Open
);
3292 NET_INC_STATS_BH(sock_net(sk
),
3293 LINUX_MIB_TCPLOSSPROBERECOVERY
);
3298 /* This routine deals with incoming acks, but not outgoing ones. */
3299 static int tcp_ack(struct sock
*sk
, const struct sk_buff
*skb
, int flag
)
3301 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3302 struct tcp_sock
*tp
= tcp_sk(sk
);
3303 u32 prior_snd_una
= tp
->snd_una
;
3304 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
3305 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3306 bool is_dupack
= false;
3307 u32 prior_in_flight
, prior_cwnd
= tp
->snd_cwnd
, prior_rtt
= tp
->srtt
;
3309 int prior_packets
= tp
->packets_out
;
3310 const int prior_unsacked
= tp
->packets_out
- tp
->sacked_out
;
3311 int acked
= 0; /* Number of packets newly acked */
3314 /* If the ack is older than previous acks
3315 * then we can probably ignore it.
3317 if (before(ack
, prior_snd_una
)) {
3318 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3319 if (before(ack
, prior_snd_una
- tp
->max_window
)) {
3320 tcp_send_challenge_ack(sk
);
3326 /* If the ack includes data we haven't sent yet, discard
3327 * this segment (RFC793 Section 3.9).
3329 if (after(ack
, tp
->snd_nxt
))
3332 if (icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
||
3333 icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
)
3336 if (after(ack
, prior_snd_una
))
3337 flag
|= FLAG_SND_UNA_ADVANCED
;
3339 prior_fackets
= tp
->fackets_out
;
3340 prior_in_flight
= tcp_packets_in_flight(tp
);
3342 /* ts_recent update must be made after we are sure that the packet
3345 if (flag
& FLAG_UPDATE_TS_RECENT
)
3346 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
3348 if (!(flag
& FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
3349 /* Window is constant, pure forward advance.
3350 * No more checks are required.
3351 * Note, we use the fact that SND.UNA>=SND.WL2.
3353 tcp_update_wl(tp
, ack_seq
);
3355 flag
|= FLAG_WIN_UPDATE
;
3357 tcp_ca_event(sk
, CA_EVENT_FAST_ACK
);
3359 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPACKS
);
3361 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
3364 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPUREACKS
);
3366 flag
|= tcp_ack_update_window(sk
, skb
, ack
, ack_seq
);
3368 if (TCP_SKB_CB(skb
)->sacked
)
3369 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
,
3372 if (TCP_ECN_rcv_ecn_echo(tp
, tcp_hdr(skb
)))
3375 tcp_ca_event(sk
, CA_EVENT_SLOW_ACK
);
3378 /* We passed data and got it acked, remove any soft error
3379 * log. Something worked...
3381 sk
->sk_err_soft
= 0;
3382 icsk
->icsk_probes_out
= 0;
3383 tp
->rcv_tstamp
= tcp_time_stamp
;
3387 /* See if we can take anything off of the retransmit queue. */
3388 acked
= tp
->packets_out
;
3389 flag
|= tcp_clean_rtx_queue(sk
, prior_fackets
, prior_snd_una
, sack_rtt
);
3390 acked
-= tp
->packets_out
;
3392 /* Advance cwnd if state allows */
3393 if (tcp_may_raise_cwnd(sk
, flag
))
3394 tcp_cong_avoid(sk
, ack
, prior_in_flight
);
3396 if (tcp_ack_is_dubious(sk
, flag
)) {
3397 is_dupack
= !(flag
& (FLAG_SND_UNA_ADVANCED
| FLAG_NOT_DUP
));
3398 tcp_fastretrans_alert(sk
, acked
, prior_unsacked
,
3401 if (tp
->tlp_high_seq
)
3402 tcp_process_tlp_ack(sk
, ack
, flag
);
3404 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
& FLAG_NOT_DUP
)) {
3405 struct dst_entry
*dst
= __sk_dst_get(sk
);
3410 if (icsk
->icsk_pending
== ICSK_TIME_RETRANS
)
3411 tcp_schedule_loss_probe(sk
);
3412 if (tp
->srtt
!= prior_rtt
|| tp
->snd_cwnd
!= prior_cwnd
)
3413 tcp_update_pacing_rate(sk
);
3417 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3418 if (flag
& FLAG_DSACKING_ACK
)
3419 tcp_fastretrans_alert(sk
, acked
, prior_unsacked
,
3421 /* If this ack opens up a zero window, clear backoff. It was
3422 * being used to time the probes, and is probably far higher than
3423 * it needs to be for normal retransmission.
3425 if (tcp_send_head(sk
))
3428 if (tp
->tlp_high_seq
)
3429 tcp_process_tlp_ack(sk
, ack
, flag
);
3433 SOCK_DEBUG(sk
, "Ack %u after %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3437 /* If data was SACKed, tag it and see if we should send more data.
3438 * If data was DSACKed, see if we can undo a cwnd reduction.
3440 if (TCP_SKB_CB(skb
)->sacked
) {
3441 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
,
3443 tcp_fastretrans_alert(sk
, acked
, prior_unsacked
,
3447 SOCK_DEBUG(sk
, "Ack %u before %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3451 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3452 * But, this can also be called on packets in the established flow when
3453 * the fast version below fails.
3455 void tcp_parse_options(const struct sk_buff
*skb
,
3456 struct tcp_options_received
*opt_rx
, int estab
,
3457 struct tcp_fastopen_cookie
*foc
)
3459 const unsigned char *ptr
;
3460 const struct tcphdr
*th
= tcp_hdr(skb
);
3461 int length
= (th
->doff
* 4) - sizeof(struct tcphdr
);
3463 ptr
= (const unsigned char *)(th
+ 1);
3464 opt_rx
->saw_tstamp
= 0;
3466 while (length
> 0) {
3467 int opcode
= *ptr
++;
3473 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
3478 if (opsize
< 2) /* "silly options" */
3480 if (opsize
> length
)
3481 return; /* don't parse partial options */
3484 if (opsize
== TCPOLEN_MSS
&& th
->syn
&& !estab
) {
3485 u16 in_mss
= get_unaligned_be16(ptr
);
3487 if (opt_rx
->user_mss
&&
3488 opt_rx
->user_mss
< in_mss
)
3489 in_mss
= opt_rx
->user_mss
;
3490 opt_rx
->mss_clamp
= in_mss
;
3495 if (opsize
== TCPOLEN_WINDOW
&& th
->syn
&&
3496 !estab
&& sysctl_tcp_window_scaling
) {
3497 __u8 snd_wscale
= *(__u8
*)ptr
;
3498 opt_rx
->wscale_ok
= 1;
3499 if (snd_wscale
> 14) {
3500 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3505 opt_rx
->snd_wscale
= snd_wscale
;
3508 case TCPOPT_TIMESTAMP
:
3509 if ((opsize
== TCPOLEN_TIMESTAMP
) &&
3510 ((estab
&& opt_rx
->tstamp_ok
) ||
3511 (!estab
&& sysctl_tcp_timestamps
))) {
3512 opt_rx
->saw_tstamp
= 1;
3513 opt_rx
->rcv_tsval
= get_unaligned_be32(ptr
);
3514 opt_rx
->rcv_tsecr
= get_unaligned_be32(ptr
+ 4);
3517 case TCPOPT_SACK_PERM
:
3518 if (opsize
== TCPOLEN_SACK_PERM
&& th
->syn
&&
3519 !estab
&& sysctl_tcp_sack
) {
3520 opt_rx
->sack_ok
= TCP_SACK_SEEN
;
3521 tcp_sack_reset(opt_rx
);
3526 if ((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
3527 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
3529 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
3532 #ifdef CONFIG_TCP_MD5SIG
3535 * The MD5 Hash has already been
3536 * checked (see tcp_v{4,6}_do_rcv()).
3541 /* Fast Open option shares code 254 using a
3542 * 16 bits magic number. It's valid only in
3543 * SYN or SYN-ACK with an even size.
3545 if (opsize
< TCPOLEN_EXP_FASTOPEN_BASE
||
3546 get_unaligned_be16(ptr
) != TCPOPT_FASTOPEN_MAGIC
||
3547 foc
== NULL
|| !th
->syn
|| (opsize
& 1))
3549 foc
->len
= opsize
- TCPOLEN_EXP_FASTOPEN_BASE
;
3550 if (foc
->len
>= TCP_FASTOPEN_COOKIE_MIN
&&
3551 foc
->len
<= TCP_FASTOPEN_COOKIE_MAX
)
3552 memcpy(foc
->val
, ptr
+ 2, foc
->len
);
3553 else if (foc
->len
!= 0)
3563 EXPORT_SYMBOL(tcp_parse_options
);
3565 static bool tcp_parse_aligned_timestamp(struct tcp_sock
*tp
, const struct tcphdr
*th
)
3567 const __be32
*ptr
= (const __be32
*)(th
+ 1);
3569 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3570 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
3571 tp
->rx_opt
.saw_tstamp
= 1;
3573 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3576 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
) - tp
->tsoffset
;
3578 tp
->rx_opt
.rcv_tsecr
= 0;
3584 /* Fast parse options. This hopes to only see timestamps.
3585 * If it is wrong it falls back on tcp_parse_options().
3587 static bool tcp_fast_parse_options(const struct sk_buff
*skb
,
3588 const struct tcphdr
*th
, struct tcp_sock
*tp
)
3590 /* In the spirit of fast parsing, compare doff directly to constant
3591 * values. Because equality is used, short doff can be ignored here.
3593 if (th
->doff
== (sizeof(*th
) / 4)) {
3594 tp
->rx_opt
.saw_tstamp
= 0;
3596 } else if (tp
->rx_opt
.tstamp_ok
&&
3597 th
->doff
== ((sizeof(*th
) + TCPOLEN_TSTAMP_ALIGNED
) / 4)) {
3598 if (tcp_parse_aligned_timestamp(tp
, th
))
3602 tcp_parse_options(skb
, &tp
->rx_opt
, 1, NULL
);
3603 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
3604 tp
->rx_opt
.rcv_tsecr
-= tp
->tsoffset
;
3609 #ifdef CONFIG_TCP_MD5SIG
3611 * Parse MD5 Signature option
3613 const u8
*tcp_parse_md5sig_option(const struct tcphdr
*th
)
3615 int length
= (th
->doff
<< 2) - sizeof(*th
);
3616 const u8
*ptr
= (const u8
*)(th
+ 1);
3618 /* If the TCP option is too short, we can short cut */
3619 if (length
< TCPOLEN_MD5SIG
)
3622 while (length
> 0) {
3623 int opcode
= *ptr
++;
3634 if (opsize
< 2 || opsize
> length
)
3636 if (opcode
== TCPOPT_MD5SIG
)
3637 return opsize
== TCPOLEN_MD5SIG
? ptr
: NULL
;
3644 EXPORT_SYMBOL(tcp_parse_md5sig_option
);
3647 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3649 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3650 * it can pass through stack. So, the following predicate verifies that
3651 * this segment is not used for anything but congestion avoidance or
3652 * fast retransmit. Moreover, we even are able to eliminate most of such
3653 * second order effects, if we apply some small "replay" window (~RTO)
3654 * to timestamp space.
3656 * All these measures still do not guarantee that we reject wrapped ACKs
3657 * on networks with high bandwidth, when sequence space is recycled fastly,
3658 * but it guarantees that such events will be very rare and do not affect
3659 * connection seriously. This doesn't look nice, but alas, PAWS is really
3662 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3663 * states that events when retransmit arrives after original data are rare.
3664 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3665 * the biggest problem on large power networks even with minor reordering.
3666 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3667 * up to bandwidth of 18Gigabit/sec. 8) ]
3670 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
3672 const struct tcp_sock
*tp
= tcp_sk(sk
);
3673 const struct tcphdr
*th
= tcp_hdr(skb
);
3674 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3675 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3677 return (/* 1. Pure ACK with correct sequence number. */
3678 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
3680 /* 2. ... and duplicate ACK. */
3681 ack
== tp
->snd_una
&&
3683 /* 3. ... and does not update window. */
3684 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
3686 /* 4. ... and sits in replay window. */
3687 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
3690 static inline bool tcp_paws_discard(const struct sock
*sk
,
3691 const struct sk_buff
*skb
)
3693 const struct tcp_sock
*tp
= tcp_sk(sk
);
3695 return !tcp_paws_check(&tp
->rx_opt
, TCP_PAWS_WINDOW
) &&
3696 !tcp_disordered_ack(sk
, skb
);
3699 /* Check segment sequence number for validity.
3701 * Segment controls are considered valid, if the segment
3702 * fits to the window after truncation to the window. Acceptability
3703 * of data (and SYN, FIN, of course) is checked separately.
3704 * See tcp_data_queue(), for example.
3706 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3707 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3708 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3709 * (borrowed from freebsd)
3712 static inline bool tcp_sequence(const struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3714 return !before(end_seq
, tp
->rcv_wup
) &&
3715 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
3718 /* When we get a reset we do this. */
3719 void tcp_reset(struct sock
*sk
)
3721 /* We want the right error as BSD sees it (and indeed as we do). */
3722 switch (sk
->sk_state
) {
3724 sk
->sk_err
= ECONNREFUSED
;
3726 case TCP_CLOSE_WAIT
:
3732 sk
->sk_err
= ECONNRESET
;
3734 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3737 if (!sock_flag(sk
, SOCK_DEAD
))
3738 sk
->sk_error_report(sk
);
3744 * Process the FIN bit. This now behaves as it is supposed to work
3745 * and the FIN takes effect when it is validly part of sequence
3746 * space. Not before when we get holes.
3748 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3749 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3752 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3753 * close and we go into CLOSING (and later onto TIME-WAIT)
3755 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3757 static void tcp_fin(struct sock
*sk
)
3759 struct tcp_sock
*tp
= tcp_sk(sk
);
3760 const struct dst_entry
*dst
;
3762 inet_csk_schedule_ack(sk
);
3764 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
3765 sock_set_flag(sk
, SOCK_DONE
);
3767 switch (sk
->sk_state
) {
3769 case TCP_ESTABLISHED
:
3770 /* Move to CLOSE_WAIT */
3771 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
3772 dst
= __sk_dst_get(sk
);
3773 if (!dst
|| !dst_metric(dst
, RTAX_QUICKACK
))
3774 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
3777 case TCP_CLOSE_WAIT
:
3779 /* Received a retransmission of the FIN, do
3784 /* RFC793: Remain in the LAST-ACK state. */
3788 /* This case occurs when a simultaneous close
3789 * happens, we must ack the received FIN and
3790 * enter the CLOSING state.
3793 tcp_set_state(sk
, TCP_CLOSING
);
3796 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3798 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
3801 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3802 * cases we should never reach this piece of code.
3804 pr_err("%s: Impossible, sk->sk_state=%d\n",
3805 __func__
, sk
->sk_state
);
3809 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3810 * Probably, we should reset in this case. For now drop them.
3812 __skb_queue_purge(&tp
->out_of_order_queue
);
3813 if (tcp_is_sack(tp
))
3814 tcp_sack_reset(&tp
->rx_opt
);
3817 if (!sock_flag(sk
, SOCK_DEAD
)) {
3818 sk
->sk_state_change(sk
);
3820 /* Do not send POLL_HUP for half duplex close. */
3821 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
3822 sk
->sk_state
== TCP_CLOSE
)
3823 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_HUP
);
3825 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_IN
);
3829 static inline bool tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
,
3832 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
3833 if (before(seq
, sp
->start_seq
))
3834 sp
->start_seq
= seq
;
3835 if (after(end_seq
, sp
->end_seq
))
3836 sp
->end_seq
= end_seq
;
3842 static void tcp_dsack_set(struct sock
*sk
, u32 seq
, u32 end_seq
)
3844 struct tcp_sock
*tp
= tcp_sk(sk
);
3846 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
3849 if (before(seq
, tp
->rcv_nxt
))
3850 mib_idx
= LINUX_MIB_TCPDSACKOLDSENT
;
3852 mib_idx
= LINUX_MIB_TCPDSACKOFOSENT
;
3854 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
3856 tp
->rx_opt
.dsack
= 1;
3857 tp
->duplicate_sack
[0].start_seq
= seq
;
3858 tp
->duplicate_sack
[0].end_seq
= end_seq
;
3862 static void tcp_dsack_extend(struct sock
*sk
, u32 seq
, u32 end_seq
)
3864 struct tcp_sock
*tp
= tcp_sk(sk
);
3866 if (!tp
->rx_opt
.dsack
)
3867 tcp_dsack_set(sk
, seq
, end_seq
);
3869 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
3872 static void tcp_send_dupack(struct sock
*sk
, const struct sk_buff
*skb
)
3874 struct tcp_sock
*tp
= tcp_sk(sk
);
3876 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
3877 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
3878 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
3879 tcp_enter_quickack_mode(sk
);
3881 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
3882 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
3884 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
3885 end_seq
= tp
->rcv_nxt
;
3886 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, end_seq
);
3893 /* These routines update the SACK block as out-of-order packets arrive or
3894 * in-order packets close up the sequence space.
3896 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
3899 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3900 struct tcp_sack_block
*swalk
= sp
+ 1;
3902 /* See if the recent change to the first SACK eats into
3903 * or hits the sequence space of other SACK blocks, if so coalesce.
3905 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
;) {
3906 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
3909 /* Zap SWALK, by moving every further SACK up by one slot.
3910 * Decrease num_sacks.
3912 tp
->rx_opt
.num_sacks
--;
3913 for (i
= this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
3917 this_sack
++, swalk
++;
3921 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
3923 struct tcp_sock
*tp
= tcp_sk(sk
);
3924 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3925 int cur_sacks
= tp
->rx_opt
.num_sacks
;
3931 for (this_sack
= 0; this_sack
< cur_sacks
; this_sack
++, sp
++) {
3932 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
3933 /* Rotate this_sack to the first one. */
3934 for (; this_sack
> 0; this_sack
--, sp
--)
3935 swap(*sp
, *(sp
- 1));
3937 tcp_sack_maybe_coalesce(tp
);
3942 /* Could not find an adjacent existing SACK, build a new one,
3943 * put it at the front, and shift everyone else down. We
3944 * always know there is at least one SACK present already here.
3946 * If the sack array is full, forget about the last one.
3948 if (this_sack
>= TCP_NUM_SACKS
) {
3950 tp
->rx_opt
.num_sacks
--;
3953 for (; this_sack
> 0; this_sack
--, sp
--)
3957 /* Build the new head SACK, and we're done. */
3958 sp
->start_seq
= seq
;
3959 sp
->end_seq
= end_seq
;
3960 tp
->rx_opt
.num_sacks
++;
3963 /* RCV.NXT advances, some SACKs should be eaten. */
3965 static void tcp_sack_remove(struct tcp_sock
*tp
)
3967 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3968 int num_sacks
= tp
->rx_opt
.num_sacks
;
3971 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3972 if (skb_queue_empty(&tp
->out_of_order_queue
)) {
3973 tp
->rx_opt
.num_sacks
= 0;
3977 for (this_sack
= 0; this_sack
< num_sacks
;) {
3978 /* Check if the start of the sack is covered by RCV.NXT. */
3979 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
3982 /* RCV.NXT must cover all the block! */
3983 WARN_ON(before(tp
->rcv_nxt
, sp
->end_seq
));
3985 /* Zap this SACK, by moving forward any other SACKS. */
3986 for (i
=this_sack
+1; i
< num_sacks
; i
++)
3987 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
3994 tp
->rx_opt
.num_sacks
= num_sacks
;
3997 /* This one checks to see if we can put data from the
3998 * out_of_order queue into the receive_queue.
4000 static void tcp_ofo_queue(struct sock
*sk
)
4002 struct tcp_sock
*tp
= tcp_sk(sk
);
4003 __u32 dsack_high
= tp
->rcv_nxt
;
4004 struct sk_buff
*skb
;
4006 while ((skb
= skb_peek(&tp
->out_of_order_queue
)) != NULL
) {
4007 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
4010 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
4011 __u32 dsack
= dsack_high
;
4012 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
4013 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
4014 tcp_dsack_extend(sk
, TCP_SKB_CB(skb
)->seq
, dsack
);
4017 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4018 SOCK_DEBUG(sk
, "ofo packet was already received\n");
4019 __skb_unlink(skb
, &tp
->out_of_order_queue
);
4023 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
4024 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4025 TCP_SKB_CB(skb
)->end_seq
);
4027 __skb_unlink(skb
, &tp
->out_of_order_queue
);
4028 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4029 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4030 if (tcp_hdr(skb
)->fin
)
4035 static bool tcp_prune_ofo_queue(struct sock
*sk
);
4036 static int tcp_prune_queue(struct sock
*sk
);
4038 static int tcp_try_rmem_schedule(struct sock
*sk
, struct sk_buff
*skb
,
4041 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
4042 !sk_rmem_schedule(sk
, skb
, size
)) {
4044 if (tcp_prune_queue(sk
) < 0)
4047 if (!sk_rmem_schedule(sk
, skb
, size
)) {
4048 if (!tcp_prune_ofo_queue(sk
))
4051 if (!sk_rmem_schedule(sk
, skb
, size
))
4059 * tcp_try_coalesce - try to merge skb to prior one
4062 * @from: buffer to add in queue
4063 * @fragstolen: pointer to boolean
4065 * Before queueing skb @from after @to, try to merge them
4066 * to reduce overall memory use and queue lengths, if cost is small.
4067 * Packets in ofo or receive queues can stay a long time.
4068 * Better try to coalesce them right now to avoid future collapses.
4069 * Returns true if caller should free @from instead of queueing it
4071 static bool tcp_try_coalesce(struct sock
*sk
,
4073 struct sk_buff
*from
,
4078 *fragstolen
= false;
4080 if (tcp_hdr(from
)->fin
)
4083 /* Its possible this segment overlaps with prior segment in queue */
4084 if (TCP_SKB_CB(from
)->seq
!= TCP_SKB_CB(to
)->end_seq
)
4087 if (!skb_try_coalesce(to
, from
, fragstolen
, &delta
))
4090 atomic_add(delta
, &sk
->sk_rmem_alloc
);
4091 sk_mem_charge(sk
, delta
);
4092 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRCVCOALESCE
);
4093 TCP_SKB_CB(to
)->end_seq
= TCP_SKB_CB(from
)->end_seq
;
4094 TCP_SKB_CB(to
)->ack_seq
= TCP_SKB_CB(from
)->ack_seq
;
4098 static void tcp_data_queue_ofo(struct sock
*sk
, struct sk_buff
*skb
)
4100 struct tcp_sock
*tp
= tcp_sk(sk
);
4101 struct sk_buff
*skb1
;
4104 TCP_ECN_check_ce(tp
, skb
);
4106 if (unlikely(tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))) {
4107 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFODROP
);
4112 /* Disable header prediction. */
4114 inet_csk_schedule_ack(sk
);
4116 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFOQUEUE
);
4117 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
4118 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4120 skb1
= skb_peek_tail(&tp
->out_of_order_queue
);
4122 /* Initial out of order segment, build 1 SACK. */
4123 if (tcp_is_sack(tp
)) {
4124 tp
->rx_opt
.num_sacks
= 1;
4125 tp
->selective_acks
[0].start_seq
= TCP_SKB_CB(skb
)->seq
;
4126 tp
->selective_acks
[0].end_seq
=
4127 TCP_SKB_CB(skb
)->end_seq
;
4129 __skb_queue_head(&tp
->out_of_order_queue
, skb
);
4133 seq
= TCP_SKB_CB(skb
)->seq
;
4134 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4136 if (seq
== TCP_SKB_CB(skb1
)->end_seq
) {
4139 if (!tcp_try_coalesce(sk
, skb1
, skb
, &fragstolen
)) {
4140 __skb_queue_after(&tp
->out_of_order_queue
, skb1
, skb
);
4142 tcp_grow_window(sk
, skb
);
4143 kfree_skb_partial(skb
, fragstolen
);
4147 if (!tp
->rx_opt
.num_sacks
||
4148 tp
->selective_acks
[0].end_seq
!= seq
)
4151 /* Common case: data arrive in order after hole. */
4152 tp
->selective_acks
[0].end_seq
= end_seq
;
4156 /* Find place to insert this segment. */
4158 if (!after(TCP_SKB_CB(skb1
)->seq
, seq
))
4160 if (skb_queue_is_first(&tp
->out_of_order_queue
, skb1
)) {
4164 skb1
= skb_queue_prev(&tp
->out_of_order_queue
, skb1
);
4167 /* Do skb overlap to previous one? */
4168 if (skb1
&& before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4169 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4170 /* All the bits are present. Drop. */
4171 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFOMERGE
);
4174 tcp_dsack_set(sk
, seq
, end_seq
);
4177 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
4178 /* Partial overlap. */
4179 tcp_dsack_set(sk
, seq
,
4180 TCP_SKB_CB(skb1
)->end_seq
);
4182 if (skb_queue_is_first(&tp
->out_of_order_queue
,
4186 skb1
= skb_queue_prev(
4187 &tp
->out_of_order_queue
,
4192 __skb_queue_head(&tp
->out_of_order_queue
, skb
);
4194 __skb_queue_after(&tp
->out_of_order_queue
, skb1
, skb
);
4196 /* And clean segments covered by new one as whole. */
4197 while (!skb_queue_is_last(&tp
->out_of_order_queue
, skb
)) {
4198 skb1
= skb_queue_next(&tp
->out_of_order_queue
, skb
);
4200 if (!after(end_seq
, TCP_SKB_CB(skb1
)->seq
))
4202 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4203 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4207 __skb_unlink(skb1
, &tp
->out_of_order_queue
);
4208 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4209 TCP_SKB_CB(skb1
)->end_seq
);
4210 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFOMERGE
);
4215 if (tcp_is_sack(tp
))
4216 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
4219 tcp_grow_window(sk
, skb
);
4220 skb_set_owner_r(skb
, sk
);
4224 static int __must_check
tcp_queue_rcv(struct sock
*sk
, struct sk_buff
*skb
, int hdrlen
,
4228 struct sk_buff
*tail
= skb_peek_tail(&sk
->sk_receive_queue
);
4230 __skb_pull(skb
, hdrlen
);
4232 tcp_try_coalesce(sk
, tail
, skb
, fragstolen
)) ? 1 : 0;
4233 tcp_sk(sk
)->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4235 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4236 skb_set_owner_r(skb
, sk
);
4241 int tcp_send_rcvq(struct sock
*sk
, struct msghdr
*msg
, size_t size
)
4243 struct sk_buff
*skb
= NULL
;
4250 skb
= alloc_skb(size
+ sizeof(*th
), sk
->sk_allocation
);
4254 if (tcp_try_rmem_schedule(sk
, skb
, size
+ sizeof(*th
)))
4257 th
= (struct tcphdr
*)skb_put(skb
, sizeof(*th
));
4258 skb_reset_transport_header(skb
);
4259 memset(th
, 0, sizeof(*th
));
4261 if (memcpy_fromiovec(skb_put(skb
, size
), msg
->msg_iov
, size
))
4264 TCP_SKB_CB(skb
)->seq
= tcp_sk(sk
)->rcv_nxt
;
4265 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ size
;
4266 TCP_SKB_CB(skb
)->ack_seq
= tcp_sk(sk
)->snd_una
- 1;
4268 if (tcp_queue_rcv(sk
, skb
, sizeof(*th
), &fragstolen
)) {
4269 WARN_ON_ONCE(fragstolen
); /* should not happen */
4280 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
4282 const struct tcphdr
*th
= tcp_hdr(skb
);
4283 struct tcp_sock
*tp
= tcp_sk(sk
);
4285 bool fragstolen
= false;
4287 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
)
4291 __skb_pull(skb
, th
->doff
* 4);
4293 TCP_ECN_accept_cwr(tp
, skb
);
4295 tp
->rx_opt
.dsack
= 0;
4297 /* Queue data for delivery to the user.
4298 * Packets in sequence go to the receive queue.
4299 * Out of sequence packets to the out_of_order_queue.
4301 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
4302 if (tcp_receive_window(tp
) == 0)
4305 /* Ok. In sequence. In window. */
4306 if (tp
->ucopy
.task
== current
&&
4307 tp
->copied_seq
== tp
->rcv_nxt
&& tp
->ucopy
.len
&&
4308 sock_owned_by_user(sk
) && !tp
->urg_data
) {
4309 int chunk
= min_t(unsigned int, skb
->len
,
4312 __set_current_state(TASK_RUNNING
);
4315 if (!skb_copy_datagram_iovec(skb
, 0, tp
->ucopy
.iov
, chunk
)) {
4316 tp
->ucopy
.len
-= chunk
;
4317 tp
->copied_seq
+= chunk
;
4318 eaten
= (chunk
== skb
->len
);
4319 tcp_rcv_space_adjust(sk
);
4327 tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))
4330 eaten
= tcp_queue_rcv(sk
, skb
, 0, &fragstolen
);
4332 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4334 tcp_event_data_recv(sk
, skb
);
4338 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4341 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4342 * gap in queue is filled.
4344 if (skb_queue_empty(&tp
->out_of_order_queue
))
4345 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
4348 if (tp
->rx_opt
.num_sacks
)
4349 tcp_sack_remove(tp
);
4351 tcp_fast_path_check(sk
);
4354 kfree_skb_partial(skb
, fragstolen
);
4355 if (!sock_flag(sk
, SOCK_DEAD
))
4356 sk
->sk_data_ready(sk
, 0);
4360 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4361 /* A retransmit, 2nd most common case. Force an immediate ack. */
4362 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4363 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4366 tcp_enter_quickack_mode(sk
);
4367 inet_csk_schedule_ack(sk
);
4373 /* Out of window. F.e. zero window probe. */
4374 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
4377 tcp_enter_quickack_mode(sk
);
4379 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4380 /* Partial packet, seq < rcv_next < end_seq */
4381 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
4382 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4383 TCP_SKB_CB(skb
)->end_seq
);
4385 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
4387 /* If window is closed, drop tail of packet. But after
4388 * remembering D-SACK for its head made in previous line.
4390 if (!tcp_receive_window(tp
))
4395 tcp_data_queue_ofo(sk
, skb
);
4398 static struct sk_buff
*tcp_collapse_one(struct sock
*sk
, struct sk_buff
*skb
,
4399 struct sk_buff_head
*list
)
4401 struct sk_buff
*next
= NULL
;
4403 if (!skb_queue_is_last(list
, skb
))
4404 next
= skb_queue_next(list
, skb
);
4406 __skb_unlink(skb
, list
);
4408 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRCVCOLLAPSED
);
4413 /* Collapse contiguous sequence of skbs head..tail with
4414 * sequence numbers start..end.
4416 * If tail is NULL, this means until the end of the list.
4418 * Segments with FIN/SYN are not collapsed (only because this
4422 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
,
4423 struct sk_buff
*head
, struct sk_buff
*tail
,
4426 struct sk_buff
*skb
, *n
;
4429 /* First, check that queue is collapsible and find
4430 * the point where collapsing can be useful. */
4434 skb_queue_walk_from_safe(list
, skb
, n
) {
4437 /* No new bits? It is possible on ofo queue. */
4438 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4439 skb
= tcp_collapse_one(sk
, skb
, list
);
4445 /* The first skb to collapse is:
4447 * - bloated or contains data before "start" or
4448 * overlaps to the next one.
4450 if (!tcp_hdr(skb
)->syn
&& !tcp_hdr(skb
)->fin
&&
4451 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
4452 before(TCP_SKB_CB(skb
)->seq
, start
))) {
4453 end_of_skbs
= false;
4457 if (!skb_queue_is_last(list
, skb
)) {
4458 struct sk_buff
*next
= skb_queue_next(list
, skb
);
4460 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(next
)->seq
) {
4461 end_of_skbs
= false;
4466 /* Decided to skip this, advance start seq. */
4467 start
= TCP_SKB_CB(skb
)->end_seq
;
4469 if (end_of_skbs
|| tcp_hdr(skb
)->syn
|| tcp_hdr(skb
)->fin
)
4472 while (before(start
, end
)) {
4473 struct sk_buff
*nskb
;
4474 unsigned int header
= skb_headroom(skb
);
4475 int copy
= SKB_MAX_ORDER(header
, 0);
4477 /* Too big header? This can happen with IPv6. */
4480 if (end
- start
< copy
)
4482 nskb
= alloc_skb(copy
+ header
, GFP_ATOMIC
);
4486 skb_set_mac_header(nskb
, skb_mac_header(skb
) - skb
->head
);
4487 skb_set_network_header(nskb
, (skb_network_header(skb
) -
4489 skb_set_transport_header(nskb
, (skb_transport_header(skb
) -
4491 skb_reserve(nskb
, header
);
4492 memcpy(nskb
->head
, skb
->head
, header
);
4493 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
4494 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
4495 __skb_queue_before(list
, skb
, nskb
);
4496 skb_set_owner_r(nskb
, sk
);
4498 /* Copy data, releasing collapsed skbs. */
4500 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
4501 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
4505 size
= min(copy
, size
);
4506 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
4508 TCP_SKB_CB(nskb
)->end_seq
+= size
;
4512 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4513 skb
= tcp_collapse_one(sk
, skb
, list
);
4516 tcp_hdr(skb
)->syn
||
4524 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4525 * and tcp_collapse() them until all the queue is collapsed.
4527 static void tcp_collapse_ofo_queue(struct sock
*sk
)
4529 struct tcp_sock
*tp
= tcp_sk(sk
);
4530 struct sk_buff
*skb
= skb_peek(&tp
->out_of_order_queue
);
4531 struct sk_buff
*head
;
4537 start
= TCP_SKB_CB(skb
)->seq
;
4538 end
= TCP_SKB_CB(skb
)->end_seq
;
4542 struct sk_buff
*next
= NULL
;
4544 if (!skb_queue_is_last(&tp
->out_of_order_queue
, skb
))
4545 next
= skb_queue_next(&tp
->out_of_order_queue
, skb
);
4548 /* Segment is terminated when we see gap or when
4549 * we are at the end of all the queue. */
4551 after(TCP_SKB_CB(skb
)->seq
, end
) ||
4552 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
4553 tcp_collapse(sk
, &tp
->out_of_order_queue
,
4554 head
, skb
, start
, end
);
4558 /* Start new segment */
4559 start
= TCP_SKB_CB(skb
)->seq
;
4560 end
= TCP_SKB_CB(skb
)->end_seq
;
4562 if (before(TCP_SKB_CB(skb
)->seq
, start
))
4563 start
= TCP_SKB_CB(skb
)->seq
;
4564 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
4565 end
= TCP_SKB_CB(skb
)->end_seq
;
4571 * Purge the out-of-order queue.
4572 * Return true if queue was pruned.
4574 static bool tcp_prune_ofo_queue(struct sock
*sk
)
4576 struct tcp_sock
*tp
= tcp_sk(sk
);
4579 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4580 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_OFOPRUNED
);
4581 __skb_queue_purge(&tp
->out_of_order_queue
);
4583 /* Reset SACK state. A conforming SACK implementation will
4584 * do the same at a timeout based retransmit. When a connection
4585 * is in a sad state like this, we care only about integrity
4586 * of the connection not performance.
4588 if (tp
->rx_opt
.sack_ok
)
4589 tcp_sack_reset(&tp
->rx_opt
);
4596 /* Reduce allocated memory if we can, trying to get
4597 * the socket within its memory limits again.
4599 * Return less than zero if we should start dropping frames
4600 * until the socket owning process reads some of the data
4601 * to stabilize the situation.
4603 static int tcp_prune_queue(struct sock
*sk
)
4605 struct tcp_sock
*tp
= tcp_sk(sk
);
4607 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
4609 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PRUNECALLED
);
4611 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
4612 tcp_clamp_window(sk
);
4613 else if (sk_under_memory_pressure(sk
))
4614 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
4616 tcp_collapse_ofo_queue(sk
);
4617 if (!skb_queue_empty(&sk
->sk_receive_queue
))
4618 tcp_collapse(sk
, &sk
->sk_receive_queue
,
4619 skb_peek(&sk
->sk_receive_queue
),
4621 tp
->copied_seq
, tp
->rcv_nxt
);
4624 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4627 /* Collapsing did not help, destructive actions follow.
4628 * This must not ever occur. */
4630 tcp_prune_ofo_queue(sk
);
4632 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4635 /* If we are really being abused, tell the caller to silently
4636 * drop receive data on the floor. It will get retransmitted
4637 * and hopefully then we'll have sufficient space.
4639 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_RCVPRUNED
);
4641 /* Massive buffer overcommit. */
4646 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4647 * As additional protections, we do not touch cwnd in retransmission phases,
4648 * and if application hit its sndbuf limit recently.
4650 void tcp_cwnd_application_limited(struct sock
*sk
)
4652 struct tcp_sock
*tp
= tcp_sk(sk
);
4654 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
4655 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
4656 /* Limited by application or receiver window. */
4657 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
4658 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
4659 if (win_used
< tp
->snd_cwnd
) {
4660 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
4661 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
4663 tp
->snd_cwnd_used
= 0;
4665 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4668 static bool tcp_should_expand_sndbuf(const struct sock
*sk
)
4670 const struct tcp_sock
*tp
= tcp_sk(sk
);
4672 /* If the user specified a specific send buffer setting, do
4675 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
4678 /* If we are under global TCP memory pressure, do not expand. */
4679 if (sk_under_memory_pressure(sk
))
4682 /* If we are under soft global TCP memory pressure, do not expand. */
4683 if (sk_memory_allocated(sk
) >= sk_prot_mem_limits(sk
, 0))
4686 /* If we filled the congestion window, do not expand. */
4687 if (tp
->packets_out
>= tp
->snd_cwnd
)
4693 /* When incoming ACK allowed to free some skb from write_queue,
4694 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4695 * on the exit from tcp input handler.
4697 * PROBLEM: sndbuf expansion does not work well with largesend.
4699 static void tcp_new_space(struct sock
*sk
)
4701 struct tcp_sock
*tp
= tcp_sk(sk
);
4703 if (tcp_should_expand_sndbuf(sk
)) {
4704 int sndmem
= SKB_TRUESIZE(max_t(u32
,
4705 tp
->rx_opt
.mss_clamp
,
4708 int demanded
= max_t(unsigned int, tp
->snd_cwnd
,
4709 tp
->reordering
+ 1);
4710 sndmem
*= 2 * demanded
;
4711 if (sndmem
> sk
->sk_sndbuf
)
4712 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
4713 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4716 sk
->sk_write_space(sk
);
4719 static void tcp_check_space(struct sock
*sk
)
4721 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
4722 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
4723 if (sk
->sk_socket
&&
4724 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
4729 static inline void tcp_data_snd_check(struct sock
*sk
)
4731 tcp_push_pending_frames(sk
);
4732 tcp_check_space(sk
);
4736 * Check if sending an ack is needed.
4738 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
4740 struct tcp_sock
*tp
= tcp_sk(sk
);
4742 /* More than one full frame received... */
4743 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
&&
4744 /* ... and right edge of window advances far enough.
4745 * (tcp_recvmsg() will send ACK otherwise). Or...
4747 __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
4748 /* We ACK each frame or... */
4749 tcp_in_quickack_mode(sk
) ||
4750 /* We have out of order data. */
4751 (ofo_possible
&& skb_peek(&tp
->out_of_order_queue
))) {
4752 /* Then ack it now */
4755 /* Else, send delayed ack. */
4756 tcp_send_delayed_ack(sk
);
4760 static inline void tcp_ack_snd_check(struct sock
*sk
)
4762 if (!inet_csk_ack_scheduled(sk
)) {
4763 /* We sent a data segment already. */
4766 __tcp_ack_snd_check(sk
, 1);
4770 * This routine is only called when we have urgent data
4771 * signaled. Its the 'slow' part of tcp_urg. It could be
4772 * moved inline now as tcp_urg is only called from one
4773 * place. We handle URGent data wrong. We have to - as
4774 * BSD still doesn't use the correction from RFC961.
4775 * For 1003.1g we should support a new option TCP_STDURG to permit
4776 * either form (or just set the sysctl tcp_stdurg).
4779 static void tcp_check_urg(struct sock
*sk
, const struct tcphdr
*th
)
4781 struct tcp_sock
*tp
= tcp_sk(sk
);
4782 u32 ptr
= ntohs(th
->urg_ptr
);
4784 if (ptr
&& !sysctl_tcp_stdurg
)
4786 ptr
+= ntohl(th
->seq
);
4788 /* Ignore urgent data that we've already seen and read. */
4789 if (after(tp
->copied_seq
, ptr
))
4792 /* Do not replay urg ptr.
4794 * NOTE: interesting situation not covered by specs.
4795 * Misbehaving sender may send urg ptr, pointing to segment,
4796 * which we already have in ofo queue. We are not able to fetch
4797 * such data and will stay in TCP_URG_NOTYET until will be eaten
4798 * by recvmsg(). Seems, we are not obliged to handle such wicked
4799 * situations. But it is worth to think about possibility of some
4800 * DoSes using some hypothetical application level deadlock.
4802 if (before(ptr
, tp
->rcv_nxt
))
4805 /* Do we already have a newer (or duplicate) urgent pointer? */
4806 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
4809 /* Tell the world about our new urgent pointer. */
4812 /* We may be adding urgent data when the last byte read was
4813 * urgent. To do this requires some care. We cannot just ignore
4814 * tp->copied_seq since we would read the last urgent byte again
4815 * as data, nor can we alter copied_seq until this data arrives
4816 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4818 * NOTE. Double Dutch. Rendering to plain English: author of comment
4819 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4820 * and expect that both A and B disappear from stream. This is _wrong_.
4821 * Though this happens in BSD with high probability, this is occasional.
4822 * Any application relying on this is buggy. Note also, that fix "works"
4823 * only in this artificial test. Insert some normal data between A and B and we will
4824 * decline of BSD again. Verdict: it is better to remove to trap
4827 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
4828 !sock_flag(sk
, SOCK_URGINLINE
) && tp
->copied_seq
!= tp
->rcv_nxt
) {
4829 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
4831 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4832 __skb_unlink(skb
, &sk
->sk_receive_queue
);
4837 tp
->urg_data
= TCP_URG_NOTYET
;
4840 /* Disable header prediction. */
4844 /* This is the 'fast' part of urgent handling. */
4845 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, const struct tcphdr
*th
)
4847 struct tcp_sock
*tp
= tcp_sk(sk
);
4849 /* Check if we get a new urgent pointer - normally not. */
4851 tcp_check_urg(sk
, th
);
4853 /* Do we wait for any urgent data? - normally not... */
4854 if (tp
->urg_data
== TCP_URG_NOTYET
) {
4855 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
4858 /* Is the urgent pointer pointing into this packet? */
4859 if (ptr
< skb
->len
) {
4861 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
4863 tp
->urg_data
= TCP_URG_VALID
| tmp
;
4864 if (!sock_flag(sk
, SOCK_DEAD
))
4865 sk
->sk_data_ready(sk
, 0);
4870 static int tcp_copy_to_iovec(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
4872 struct tcp_sock
*tp
= tcp_sk(sk
);
4873 int chunk
= skb
->len
- hlen
;
4877 if (skb_csum_unnecessary(skb
))
4878 err
= skb_copy_datagram_iovec(skb
, hlen
, tp
->ucopy
.iov
, chunk
);
4880 err
= skb_copy_and_csum_datagram_iovec(skb
, hlen
,
4884 tp
->ucopy
.len
-= chunk
;
4885 tp
->copied_seq
+= chunk
;
4886 tcp_rcv_space_adjust(sk
);
4893 static __sum16
__tcp_checksum_complete_user(struct sock
*sk
,
4894 struct sk_buff
*skb
)
4898 if (sock_owned_by_user(sk
)) {
4900 result
= __tcp_checksum_complete(skb
);
4903 result
= __tcp_checksum_complete(skb
);
4908 static inline bool tcp_checksum_complete_user(struct sock
*sk
,
4909 struct sk_buff
*skb
)
4911 return !skb_csum_unnecessary(skb
) &&
4912 __tcp_checksum_complete_user(sk
, skb
);
4915 #ifdef CONFIG_NET_DMA
4916 static bool tcp_dma_try_early_copy(struct sock
*sk
, struct sk_buff
*skb
,
4919 struct tcp_sock
*tp
= tcp_sk(sk
);
4920 int chunk
= skb
->len
- hlen
;
4922 bool copied_early
= false;
4924 if (tp
->ucopy
.wakeup
)
4927 if (!tp
->ucopy
.dma_chan
&& tp
->ucopy
.pinned_list
)
4928 tp
->ucopy
.dma_chan
= net_dma_find_channel();
4930 if (tp
->ucopy
.dma_chan
&& skb_csum_unnecessary(skb
)) {
4932 dma_cookie
= dma_skb_copy_datagram_iovec(tp
->ucopy
.dma_chan
,
4934 tp
->ucopy
.iov
, chunk
,
4935 tp
->ucopy
.pinned_list
);
4940 tp
->ucopy
.dma_cookie
= dma_cookie
;
4941 copied_early
= true;
4943 tp
->ucopy
.len
-= chunk
;
4944 tp
->copied_seq
+= chunk
;
4945 tcp_rcv_space_adjust(sk
);
4947 if ((tp
->ucopy
.len
== 0) ||
4948 (tcp_flag_word(tcp_hdr(skb
)) & TCP_FLAG_PSH
) ||
4949 (atomic_read(&sk
->sk_rmem_alloc
) > (sk
->sk_rcvbuf
>> 1))) {
4950 tp
->ucopy
.wakeup
= 1;
4951 sk
->sk_data_ready(sk
, 0);
4953 } else if (chunk
> 0) {
4954 tp
->ucopy
.wakeup
= 1;
4955 sk
->sk_data_ready(sk
, 0);
4958 return copied_early
;
4960 #endif /* CONFIG_NET_DMA */
4962 /* Does PAWS and seqno based validation of an incoming segment, flags will
4963 * play significant role here.
4965 static bool tcp_validate_incoming(struct sock
*sk
, struct sk_buff
*skb
,
4966 const struct tcphdr
*th
, int syn_inerr
)
4968 struct tcp_sock
*tp
= tcp_sk(sk
);
4970 /* RFC1323: H1. Apply PAWS check first. */
4971 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
4972 tcp_paws_discard(sk
, skb
)) {
4974 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSESTABREJECTED
);
4975 tcp_send_dupack(sk
, skb
);
4978 /* Reset is accepted even if it did not pass PAWS. */
4981 /* Step 1: check sequence number */
4982 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4983 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4984 * (RST) segments are validated by checking their SEQ-fields."
4985 * And page 69: "If an incoming segment is not acceptable,
4986 * an acknowledgment should be sent in reply (unless the RST
4987 * bit is set, if so drop the segment and return)".
4992 tcp_send_dupack(sk
, skb
);
4997 /* Step 2: check RST bit */
5000 * If sequence number exactly matches RCV.NXT, then
5001 * RESET the connection
5003 * Send a challenge ACK
5005 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
)
5008 tcp_send_challenge_ack(sk
);
5012 /* step 3: check security and precedence [ignored] */
5014 /* step 4: Check for a SYN
5015 * RFC 5691 4.2 : Send a challenge ack
5020 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5021 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSYNCHALLENGE
);
5022 tcp_send_challenge_ack(sk
);
5034 * TCP receive function for the ESTABLISHED state.
5036 * It is split into a fast path and a slow path. The fast path is
5038 * - A zero window was announced from us - zero window probing
5039 * is only handled properly in the slow path.
5040 * - Out of order segments arrived.
5041 * - Urgent data is expected.
5042 * - There is no buffer space left
5043 * - Unexpected TCP flags/window values/header lengths are received
5044 * (detected by checking the TCP header against pred_flags)
5045 * - Data is sent in both directions. Fast path only supports pure senders
5046 * or pure receivers (this means either the sequence number or the ack
5047 * value must stay constant)
5048 * - Unexpected TCP option.
5050 * When these conditions are not satisfied it drops into a standard
5051 * receive procedure patterned after RFC793 to handle all cases.
5052 * The first three cases are guaranteed by proper pred_flags setting,
5053 * the rest is checked inline. Fast processing is turned on in
5054 * tcp_data_queue when everything is OK.
5056 void tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
5057 const struct tcphdr
*th
, unsigned int len
)
5059 struct tcp_sock
*tp
= tcp_sk(sk
);
5061 if (unlikely(sk
->sk_rx_dst
== NULL
))
5062 inet_csk(sk
)->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
5064 * Header prediction.
5065 * The code loosely follows the one in the famous
5066 * "30 instruction TCP receive" Van Jacobson mail.
5068 * Van's trick is to deposit buffers into socket queue
5069 * on a device interrupt, to call tcp_recv function
5070 * on the receive process context and checksum and copy
5071 * the buffer to user space. smart...
5073 * Our current scheme is not silly either but we take the
5074 * extra cost of the net_bh soft interrupt processing...
5075 * We do checksum and copy also but from device to kernel.
5078 tp
->rx_opt
.saw_tstamp
= 0;
5080 /* pred_flags is 0xS?10 << 16 + snd_wnd
5081 * if header_prediction is to be made
5082 * 'S' will always be tp->tcp_header_len >> 2
5083 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5084 * turn it off (when there are holes in the receive
5085 * space for instance)
5086 * PSH flag is ignored.
5089 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
5090 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
&&
5091 !after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
)) {
5092 int tcp_header_len
= tp
->tcp_header_len
;
5094 /* Timestamp header prediction: tcp_header_len
5095 * is automatically equal to th->doff*4 due to pred_flags
5099 /* Check timestamp */
5100 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
5101 /* No? Slow path! */
5102 if (!tcp_parse_aligned_timestamp(tp
, th
))
5105 /* If PAWS failed, check it more carefully in slow path */
5106 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
5109 /* DO NOT update ts_recent here, if checksum fails
5110 * and timestamp was corrupted part, it will result
5111 * in a hung connection since we will drop all
5112 * future packets due to the PAWS test.
5116 if (len
<= tcp_header_len
) {
5117 /* Bulk data transfer: sender */
5118 if (len
== tcp_header_len
) {
5119 /* Predicted packet is in window by definition.
5120 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5121 * Hence, check seq<=rcv_wup reduces to:
5123 if (tcp_header_len
==
5124 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5125 tp
->rcv_nxt
== tp
->rcv_wup
)
5126 tcp_store_ts_recent(tp
);
5128 /* We know that such packets are checksummed
5131 tcp_ack(sk
, skb
, 0);
5133 tcp_data_snd_check(sk
);
5135 } else { /* Header too small */
5136 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5141 int copied_early
= 0;
5142 bool fragstolen
= false;
5144 if (tp
->copied_seq
== tp
->rcv_nxt
&&
5145 len
- tcp_header_len
<= tp
->ucopy
.len
) {
5146 #ifdef CONFIG_NET_DMA
5147 if (tp
->ucopy
.task
== current
&&
5148 sock_owned_by_user(sk
) &&
5149 tcp_dma_try_early_copy(sk
, skb
, tcp_header_len
)) {
5154 if (tp
->ucopy
.task
== current
&&
5155 sock_owned_by_user(sk
) && !copied_early
) {
5156 __set_current_state(TASK_RUNNING
);
5158 if (!tcp_copy_to_iovec(sk
, skb
, tcp_header_len
))
5162 /* Predicted packet is in window by definition.
5163 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5164 * Hence, check seq<=rcv_wup reduces to:
5166 if (tcp_header_len
==
5167 (sizeof(struct tcphdr
) +
5168 TCPOLEN_TSTAMP_ALIGNED
) &&
5169 tp
->rcv_nxt
== tp
->rcv_wup
)
5170 tcp_store_ts_recent(tp
);
5172 tcp_rcv_rtt_measure_ts(sk
, skb
);
5174 __skb_pull(skb
, tcp_header_len
);
5175 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
5176 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITSTOUSER
);
5179 tcp_cleanup_rbuf(sk
, skb
->len
);
5182 if (tcp_checksum_complete_user(sk
, skb
))
5185 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
5188 /* Predicted packet is in window by definition.
5189 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5190 * Hence, check seq<=rcv_wup reduces to:
5192 if (tcp_header_len
==
5193 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5194 tp
->rcv_nxt
== tp
->rcv_wup
)
5195 tcp_store_ts_recent(tp
);
5197 tcp_rcv_rtt_measure_ts(sk
, skb
);
5199 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITS
);
5201 /* Bulk data transfer: receiver */
5202 eaten
= tcp_queue_rcv(sk
, skb
, tcp_header_len
,
5206 tcp_event_data_recv(sk
, skb
);
5208 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
5209 /* Well, only one small jumplet in fast path... */
5210 tcp_ack(sk
, skb
, FLAG_DATA
);
5211 tcp_data_snd_check(sk
);
5212 if (!inet_csk_ack_scheduled(sk
))
5216 if (!copied_early
|| tp
->rcv_nxt
!= tp
->rcv_wup
)
5217 __tcp_ack_snd_check(sk
, 0);
5219 #ifdef CONFIG_NET_DMA
5221 __skb_queue_tail(&sk
->sk_async_wait_queue
, skb
);
5225 kfree_skb_partial(skb
, fragstolen
);
5226 sk
->sk_data_ready(sk
, 0);
5232 if (len
< (th
->doff
<< 2) || tcp_checksum_complete_user(sk
, skb
))
5235 if (!th
->ack
&& !th
->rst
)
5239 * Standard slow path.
5242 if (!tcp_validate_incoming(sk
, skb
, th
, 1))
5246 if (tcp_ack(sk
, skb
, FLAG_SLOWPATH
| FLAG_UPDATE_TS_RECENT
) < 0)
5249 tcp_rcv_rtt_measure_ts(sk
, skb
);
5251 /* Process urgent data. */
5252 tcp_urg(sk
, skb
, th
);
5254 /* step 7: process the segment text */
5255 tcp_data_queue(sk
, skb
);
5257 tcp_data_snd_check(sk
);
5258 tcp_ack_snd_check(sk
);
5262 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_CSUMERRORS
);
5263 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5268 EXPORT_SYMBOL(tcp_rcv_established
);
5270 void tcp_finish_connect(struct sock
*sk
, struct sk_buff
*skb
)
5272 struct tcp_sock
*tp
= tcp_sk(sk
);
5273 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5275 tcp_set_state(sk
, TCP_ESTABLISHED
);
5278 icsk
->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
5279 security_inet_conn_established(sk
, skb
);
5282 /* Make sure socket is routed, for correct metrics. */
5283 icsk
->icsk_af_ops
->rebuild_header(sk
);
5285 tcp_init_metrics(sk
);
5287 tcp_init_congestion_control(sk
);
5289 /* Prevent spurious tcp_cwnd_restart() on first data
5292 tp
->lsndtime
= tcp_time_stamp
;
5294 tcp_init_buffer_space(sk
);
5296 if (sock_flag(sk
, SOCK_KEEPOPEN
))
5297 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
5299 if (!tp
->rx_opt
.snd_wscale
)
5300 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
5304 if (!sock_flag(sk
, SOCK_DEAD
)) {
5305 sk
->sk_state_change(sk
);
5306 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5310 static bool tcp_rcv_fastopen_synack(struct sock
*sk
, struct sk_buff
*synack
,
5311 struct tcp_fastopen_cookie
*cookie
)
5313 struct tcp_sock
*tp
= tcp_sk(sk
);
5314 struct sk_buff
*data
= tp
->syn_data
? tcp_write_queue_head(sk
) : NULL
;
5315 u16 mss
= tp
->rx_opt
.mss_clamp
;
5318 if (mss
== tp
->rx_opt
.user_mss
) {
5319 struct tcp_options_received opt
;
5321 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5322 tcp_clear_options(&opt
);
5323 opt
.user_mss
= opt
.mss_clamp
= 0;
5324 tcp_parse_options(synack
, &opt
, 0, NULL
);
5325 mss
= opt
.mss_clamp
;
5328 if (!tp
->syn_fastopen
) /* Ignore an unsolicited cookie */
5331 /* The SYN-ACK neither has cookie nor acknowledges the data. Presumably
5332 * the remote receives only the retransmitted (regular) SYNs: either
5333 * the original SYN-data or the corresponding SYN-ACK is lost.
5335 syn_drop
= (cookie
->len
<= 0 && data
&& tp
->total_retrans
);
5337 tcp_fastopen_cache_set(sk
, mss
, cookie
, syn_drop
);
5339 if (data
) { /* Retransmit unacked data in SYN */
5340 tcp_for_write_queue_from(data
, sk
) {
5341 if (data
== tcp_send_head(sk
) ||
5342 __tcp_retransmit_skb(sk
, data
))
5348 tp
->syn_data_acked
= tp
->syn_data
;
5352 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5353 const struct tcphdr
*th
, unsigned int len
)
5355 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5356 struct tcp_sock
*tp
= tcp_sk(sk
);
5357 struct tcp_fastopen_cookie foc
= { .len
= -1 };
5358 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
5360 tcp_parse_options(skb
, &tp
->rx_opt
, 0, &foc
);
5361 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
5362 tp
->rx_opt
.rcv_tsecr
-= tp
->tsoffset
;
5366 * "If the state is SYN-SENT then
5367 * first check the ACK bit
5368 * If the ACK bit is set
5369 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5370 * a reset (unless the RST bit is set, if so drop
5371 * the segment and return)"
5373 if (!after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_una
) ||
5374 after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
))
5375 goto reset_and_undo
;
5377 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
5378 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
5380 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSACTIVEREJECTED
);
5381 goto reset_and_undo
;
5384 /* Now ACK is acceptable.
5386 * "If the RST bit is set
5387 * If the ACK was acceptable then signal the user "error:
5388 * connection reset", drop the segment, enter CLOSED state,
5389 * delete TCB, and return."
5398 * "fifth, if neither of the SYN or RST bits is set then
5399 * drop the segment and return."
5405 goto discard_and_undo
;
5408 * "If the SYN bit is on ...
5409 * are acceptable then ...
5410 * (our SYN has been ACKed), change the connection
5411 * state to ESTABLISHED..."
5414 TCP_ECN_rcv_synack(tp
, th
);
5416 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5417 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
5419 /* Ok.. it's good. Set up sequence numbers and
5420 * move to established.
5422 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5423 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5425 /* RFC1323: The window in SYN & SYN/ACK segments is
5428 tp
->snd_wnd
= ntohs(th
->window
);
5430 if (!tp
->rx_opt
.wscale_ok
) {
5431 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
5432 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
5435 if (tp
->rx_opt
.saw_tstamp
) {
5436 tp
->rx_opt
.tstamp_ok
= 1;
5437 tp
->tcp_header_len
=
5438 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5439 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5440 tcp_store_ts_recent(tp
);
5442 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5445 if (tcp_is_sack(tp
) && sysctl_tcp_fack
)
5446 tcp_enable_fack(tp
);
5449 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5450 tcp_initialize_rcv_mss(sk
);
5452 /* Remember, tcp_poll() does not lock socket!
5453 * Change state from SYN-SENT only after copied_seq
5454 * is initialized. */
5455 tp
->copied_seq
= tp
->rcv_nxt
;
5459 tcp_finish_connect(sk
, skb
);
5461 if ((tp
->syn_fastopen
|| tp
->syn_data
) &&
5462 tcp_rcv_fastopen_synack(sk
, skb
, &foc
))
5465 if (sk
->sk_write_pending
||
5466 icsk
->icsk_accept_queue
.rskq_defer_accept
||
5467 icsk
->icsk_ack
.pingpong
) {
5468 /* Save one ACK. Data will be ready after
5469 * several ticks, if write_pending is set.
5471 * It may be deleted, but with this feature tcpdumps
5472 * look so _wonderfully_ clever, that I was not able
5473 * to stand against the temptation 8) --ANK
5475 inet_csk_schedule_ack(sk
);
5476 icsk
->icsk_ack
.lrcvtime
= tcp_time_stamp
;
5477 tcp_enter_quickack_mode(sk
);
5478 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
5479 TCP_DELACK_MAX
, TCP_RTO_MAX
);
5490 /* No ACK in the segment */
5494 * "If the RST bit is set
5496 * Otherwise (no ACK) drop the segment and return."
5499 goto discard_and_undo
;
5503 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&&
5504 tcp_paws_reject(&tp
->rx_opt
, 0))
5505 goto discard_and_undo
;
5508 /* We see SYN without ACK. It is attempt of
5509 * simultaneous connect with crossed SYNs.
5510 * Particularly, it can be connect to self.
5512 tcp_set_state(sk
, TCP_SYN_RECV
);
5514 if (tp
->rx_opt
.saw_tstamp
) {
5515 tp
->rx_opt
.tstamp_ok
= 1;
5516 tcp_store_ts_recent(tp
);
5517 tp
->tcp_header_len
=
5518 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5520 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5523 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5524 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5526 /* RFC1323: The window in SYN & SYN/ACK segments is
5529 tp
->snd_wnd
= ntohs(th
->window
);
5530 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
5531 tp
->max_window
= tp
->snd_wnd
;
5533 TCP_ECN_rcv_syn(tp
, th
);
5536 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5537 tcp_initialize_rcv_mss(sk
);
5539 tcp_send_synack(sk
);
5541 /* Note, we could accept data and URG from this segment.
5542 * There are no obstacles to make this (except that we must
5543 * either change tcp_recvmsg() to prevent it from returning data
5544 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5546 * However, if we ignore data in ACKless segments sometimes,
5547 * we have no reasons to accept it sometimes.
5548 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5549 * is not flawless. So, discard packet for sanity.
5550 * Uncomment this return to process the data.
5557 /* "fifth, if neither of the SYN or RST bits is set then
5558 * drop the segment and return."
5562 tcp_clear_options(&tp
->rx_opt
);
5563 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5567 tcp_clear_options(&tp
->rx_opt
);
5568 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5573 * This function implements the receiving procedure of RFC 793 for
5574 * all states except ESTABLISHED and TIME_WAIT.
5575 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5576 * address independent.
5579 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5580 const struct tcphdr
*th
, unsigned int len
)
5582 struct tcp_sock
*tp
= tcp_sk(sk
);
5583 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5584 struct request_sock
*req
;
5588 tp
->rx_opt
.saw_tstamp
= 0;
5590 switch (sk
->sk_state
) {
5604 if (icsk
->icsk_af_ops
->conn_request(sk
, skb
) < 0)
5607 /* Now we have several options: In theory there is
5608 * nothing else in the frame. KA9Q has an option to
5609 * send data with the syn, BSD accepts data with the
5610 * syn up to the [to be] advertised window and
5611 * Solaris 2.1 gives you a protocol error. For now
5612 * we just ignore it, that fits the spec precisely
5613 * and avoids incompatibilities. It would be nice in
5614 * future to drop through and process the data.
5616 * Now that TTCP is starting to be used we ought to
5618 * But, this leaves one open to an easy denial of
5619 * service attack, and SYN cookies can't defend
5620 * against this problem. So, we drop the data
5621 * in the interest of security over speed unless
5622 * it's still in use.
5630 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
, len
);
5634 /* Do step6 onward by hand. */
5635 tcp_urg(sk
, skb
, th
);
5637 tcp_data_snd_check(sk
);
5641 req
= tp
->fastopen_rsk
;
5643 WARN_ON_ONCE(sk
->sk_state
!= TCP_SYN_RECV
&&
5644 sk
->sk_state
!= TCP_FIN_WAIT1
);
5646 if (tcp_check_req(sk
, skb
, req
, NULL
, true) == NULL
)
5650 if (!th
->ack
&& !th
->rst
)
5653 if (!tcp_validate_incoming(sk
, skb
, th
, 0))
5656 /* step 5: check the ACK field */
5657 acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
|
5658 FLAG_UPDATE_TS_RECENT
) > 0;
5660 switch (sk
->sk_state
) {
5665 /* Once we leave TCP_SYN_RECV, we no longer need req
5669 tp
->total_retrans
= req
->num_retrans
;
5670 reqsk_fastopen_remove(sk
, req
, false);
5672 /* Make sure socket is routed, for correct metrics. */
5673 icsk
->icsk_af_ops
->rebuild_header(sk
);
5674 tcp_init_congestion_control(sk
);
5677 tcp_init_buffer_space(sk
);
5678 tp
->copied_seq
= tp
->rcv_nxt
;
5681 tcp_set_state(sk
, TCP_ESTABLISHED
);
5682 sk
->sk_state_change(sk
);
5684 /* Note, that this wakeup is only for marginal crossed SYN case.
5685 * Passively open sockets are not waked up, because
5686 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5689 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5691 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
5692 tp
->snd_wnd
= ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
;
5693 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5694 tcp_synack_rtt_meas(sk
, req
);
5696 if (tp
->rx_opt
.tstamp_ok
)
5697 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5700 /* Re-arm the timer because data may have been sent out.
5701 * This is similar to the regular data transmission case
5702 * when new data has just been ack'ed.
5704 * (TFO) - we could try to be more aggressive and
5705 * retransmitting any data sooner based on when they
5710 tcp_init_metrics(sk
);
5712 /* Prevent spurious tcp_cwnd_restart() on first data packet */
5713 tp
->lsndtime
= tcp_time_stamp
;
5715 tcp_initialize_rcv_mss(sk
);
5716 tcp_fast_path_on(tp
);
5719 case TCP_FIN_WAIT1
: {
5720 struct dst_entry
*dst
;
5723 /* If we enter the TCP_FIN_WAIT1 state and we are a
5724 * Fast Open socket and this is the first acceptable
5725 * ACK we have received, this would have acknowledged
5726 * our SYNACK so stop the SYNACK timer.
5729 /* Return RST if ack_seq is invalid.
5730 * Note that RFC793 only says to generate a
5731 * DUPACK for it but for TCP Fast Open it seems
5732 * better to treat this case like TCP_SYN_RECV
5737 /* We no longer need the request sock. */
5738 reqsk_fastopen_remove(sk
, req
, false);
5741 if (tp
->snd_una
!= tp
->write_seq
)
5744 tcp_set_state(sk
, TCP_FIN_WAIT2
);
5745 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
5747 dst
= __sk_dst_get(sk
);
5751 if (!sock_flag(sk
, SOCK_DEAD
)) {
5752 /* Wake up lingering close() */
5753 sk
->sk_state_change(sk
);
5757 if (tp
->linger2
< 0 ||
5758 (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5759 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
))) {
5761 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
5765 tmo
= tcp_fin_time(sk
);
5766 if (tmo
> TCP_TIMEWAIT_LEN
) {
5767 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
5768 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
5769 /* Bad case. We could lose such FIN otherwise.
5770 * It is not a big problem, but it looks confusing
5771 * and not so rare event. We still can lose it now,
5772 * if it spins in bh_lock_sock(), but it is really
5775 inet_csk_reset_keepalive_timer(sk
, tmo
);
5777 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
5784 if (tp
->snd_una
== tp
->write_seq
) {
5785 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
5791 if (tp
->snd_una
== tp
->write_seq
) {
5792 tcp_update_metrics(sk
);
5799 /* step 6: check the URG bit */
5800 tcp_urg(sk
, skb
, th
);
5802 /* step 7: process the segment text */
5803 switch (sk
->sk_state
) {
5804 case TCP_CLOSE_WAIT
:
5807 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
5811 /* RFC 793 says to queue data in these states,
5812 * RFC 1122 says we MUST send a reset.
5813 * BSD 4.4 also does reset.
5815 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
5816 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5817 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
5818 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
5824 case TCP_ESTABLISHED
:
5825 tcp_data_queue(sk
, skb
);
5830 /* tcp_data could move socket to TIME-WAIT */
5831 if (sk
->sk_state
!= TCP_CLOSE
) {
5832 tcp_data_snd_check(sk
);
5833 tcp_ack_snd_check(sk
);
5842 EXPORT_SYMBOL(tcp_rcv_state_process
);