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 /* Calculate rto without backoff. This is the second half of Van Jacobson's
692 * routine referred to above.
694 void tcp_set_rto(struct sock
*sk
)
696 const struct tcp_sock
*tp
= tcp_sk(sk
);
697 /* Old crap is replaced with new one. 8)
700 * 1. If rtt variance happened to be less 50msec, it is hallucination.
701 * It cannot be less due to utterly erratic ACK generation made
702 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
703 * to do with delayed acks, because at cwnd>2 true delack timeout
704 * is invisible. Actually, Linux-2.4 also generates erratic
705 * ACKs in some circumstances.
707 inet_csk(sk
)->icsk_rto
= __tcp_set_rto(tp
);
709 /* 2. Fixups made earlier cannot be right.
710 * If we do not estimate RTO correctly without them,
711 * all the algo is pure shit and should be replaced
712 * with correct one. It is exactly, which we pretend to do.
715 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
716 * guarantees that rto is higher.
721 __u32
tcp_init_cwnd(const struct tcp_sock
*tp
, const struct dst_entry
*dst
)
723 __u32 cwnd
= (dst
? dst_metric(dst
, RTAX_INITCWND
) : 0);
726 cwnd
= TCP_INIT_CWND
;
727 return min_t(__u32
, cwnd
, tp
->snd_cwnd_clamp
);
731 * Packet counting of FACK is based on in-order assumptions, therefore TCP
732 * disables it when reordering is detected
734 void tcp_disable_fack(struct tcp_sock
*tp
)
736 /* RFC3517 uses different metric in lost marker => reset on change */
738 tp
->lost_skb_hint
= NULL
;
739 tp
->rx_opt
.sack_ok
&= ~TCP_FACK_ENABLED
;
742 /* Take a notice that peer is sending D-SACKs */
743 static void tcp_dsack_seen(struct tcp_sock
*tp
)
745 tp
->rx_opt
.sack_ok
|= TCP_DSACK_SEEN
;
748 static void tcp_update_reordering(struct sock
*sk
, const int metric
,
751 struct tcp_sock
*tp
= tcp_sk(sk
);
752 if (metric
> tp
->reordering
) {
755 tp
->reordering
= min(TCP_MAX_REORDERING
, metric
);
757 /* This exciting event is worth to be remembered. 8) */
759 mib_idx
= LINUX_MIB_TCPTSREORDER
;
760 else if (tcp_is_reno(tp
))
761 mib_idx
= LINUX_MIB_TCPRENOREORDER
;
762 else if (tcp_is_fack(tp
))
763 mib_idx
= LINUX_MIB_TCPFACKREORDER
;
765 mib_idx
= LINUX_MIB_TCPSACKREORDER
;
767 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
768 #if FASTRETRANS_DEBUG > 1
769 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
770 tp
->rx_opt
.sack_ok
, inet_csk(sk
)->icsk_ca_state
,
774 tp
->undo_marker
? tp
->undo_retrans
: 0);
776 tcp_disable_fack(tp
);
780 tcp_disable_early_retrans(tp
);
783 /* This must be called before lost_out is incremented */
784 static void tcp_verify_retransmit_hint(struct tcp_sock
*tp
, struct sk_buff
*skb
)
786 if ((tp
->retransmit_skb_hint
== NULL
) ||
787 before(TCP_SKB_CB(skb
)->seq
,
788 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
789 tp
->retransmit_skb_hint
= skb
;
792 after(TCP_SKB_CB(skb
)->end_seq
, tp
->retransmit_high
))
793 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
796 static void tcp_skb_mark_lost(struct tcp_sock
*tp
, struct sk_buff
*skb
)
798 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
799 tcp_verify_retransmit_hint(tp
, skb
);
801 tp
->lost_out
+= tcp_skb_pcount(skb
);
802 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
806 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock
*tp
,
809 tcp_verify_retransmit_hint(tp
, skb
);
811 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
812 tp
->lost_out
+= tcp_skb_pcount(skb
);
813 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
817 /* This procedure tags the retransmission queue when SACKs arrive.
819 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
820 * Packets in queue with these bits set are counted in variables
821 * sacked_out, retrans_out and lost_out, correspondingly.
823 * Valid combinations are:
824 * Tag InFlight Description
825 * 0 1 - orig segment is in flight.
826 * S 0 - nothing flies, orig reached receiver.
827 * L 0 - nothing flies, orig lost by net.
828 * R 2 - both orig and retransmit are in flight.
829 * L|R 1 - orig is lost, retransmit is in flight.
830 * S|R 1 - orig reached receiver, retrans is still in flight.
831 * (L|S|R is logically valid, it could occur when L|R is sacked,
832 * but it is equivalent to plain S and code short-curcuits it to S.
833 * L|S is logically invalid, it would mean -1 packet in flight 8))
835 * These 6 states form finite state machine, controlled by the following events:
836 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
837 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
838 * 3. Loss detection event of two flavors:
839 * A. Scoreboard estimator decided the packet is lost.
840 * A'. Reno "three dupacks" marks head of queue lost.
841 * A''. Its FACK modification, head until snd.fack is lost.
842 * B. SACK arrives sacking SND.NXT at the moment, when the
843 * segment was retransmitted.
844 * 4. D-SACK added new rule: D-SACK changes any tag to S.
846 * It is pleasant to note, that state diagram turns out to be commutative,
847 * so that we are allowed not to be bothered by order of our actions,
848 * when multiple events arrive simultaneously. (see the function below).
850 * Reordering detection.
851 * --------------------
852 * Reordering metric is maximal distance, which a packet can be displaced
853 * in packet stream. With SACKs we can estimate it:
855 * 1. SACK fills old hole and the corresponding segment was not
856 * ever retransmitted -> reordering. Alas, we cannot use it
857 * when segment was retransmitted.
858 * 2. The last flaw is solved with D-SACK. D-SACK arrives
859 * for retransmitted and already SACKed segment -> reordering..
860 * Both of these heuristics are not used in Loss state, when we cannot
861 * account for retransmits accurately.
863 * SACK block validation.
864 * ----------------------
866 * SACK block range validation checks that the received SACK block fits to
867 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
868 * Note that SND.UNA is not included to the range though being valid because
869 * it means that the receiver is rather inconsistent with itself reporting
870 * SACK reneging when it should advance SND.UNA. Such SACK block this is
871 * perfectly valid, however, in light of RFC2018 which explicitly states
872 * that "SACK block MUST reflect the newest segment. Even if the newest
873 * segment is going to be discarded ...", not that it looks very clever
874 * in case of head skb. Due to potentional receiver driven attacks, we
875 * choose to avoid immediate execution of a walk in write queue due to
876 * reneging and defer head skb's loss recovery to standard loss recovery
877 * procedure that will eventually trigger (nothing forbids us doing this).
879 * Implements also blockage to start_seq wrap-around. Problem lies in the
880 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
881 * there's no guarantee that it will be before snd_nxt (n). The problem
882 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
885 * <- outs wnd -> <- wrapzone ->
886 * u e n u_w e_w s n_w
888 * |<------------+------+----- TCP seqno space --------------+---------->|
889 * ...-- <2^31 ->| |<--------...
890 * ...---- >2^31 ------>| |<--------...
892 * Current code wouldn't be vulnerable but it's better still to discard such
893 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
894 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
895 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
896 * equal to the ideal case (infinite seqno space without wrap caused issues).
898 * With D-SACK the lower bound is extended to cover sequence space below
899 * SND.UNA down to undo_marker, which is the last point of interest. Yet
900 * again, D-SACK block must not to go across snd_una (for the same reason as
901 * for the normal SACK blocks, explained above). But there all simplicity
902 * ends, TCP might receive valid D-SACKs below that. As long as they reside
903 * fully below undo_marker they do not affect behavior in anyway and can
904 * therefore be safely ignored. In rare cases (which are more or less
905 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
906 * fragmentation and packet reordering past skb's retransmission. To consider
907 * them correctly, the acceptable range must be extended even more though
908 * the exact amount is rather hard to quantify. However, tp->max_window can
909 * be used as an exaggerated estimate.
911 static bool tcp_is_sackblock_valid(struct tcp_sock
*tp
, bool is_dsack
,
912 u32 start_seq
, u32 end_seq
)
914 /* Too far in future, or reversed (interpretation is ambiguous) */
915 if (after(end_seq
, tp
->snd_nxt
) || !before(start_seq
, end_seq
))
918 /* Nasty start_seq wrap-around check (see comments above) */
919 if (!before(start_seq
, tp
->snd_nxt
))
922 /* In outstanding window? ...This is valid exit for D-SACKs too.
923 * start_seq == snd_una is non-sensical (see comments above)
925 if (after(start_seq
, tp
->snd_una
))
928 if (!is_dsack
|| !tp
->undo_marker
)
931 /* ...Then it's D-SACK, and must reside below snd_una completely */
932 if (after(end_seq
, tp
->snd_una
))
935 if (!before(start_seq
, tp
->undo_marker
))
939 if (!after(end_seq
, tp
->undo_marker
))
942 /* Undo_marker boundary crossing (overestimates a lot). Known already:
943 * start_seq < undo_marker and end_seq >= undo_marker.
945 return !before(start_seq
, end_seq
- tp
->max_window
);
948 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
949 * Event "B". Later note: FACK people cheated me again 8), we have to account
950 * for reordering! Ugly, but should help.
952 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
953 * less than what is now known to be received by the other end (derived from
954 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
955 * retransmitted skbs to avoid some costly processing per ACKs.
957 static void tcp_mark_lost_retrans(struct sock
*sk
)
959 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
960 struct tcp_sock
*tp
= tcp_sk(sk
);
963 u32 new_low_seq
= tp
->snd_nxt
;
964 u32 received_upto
= tcp_highest_sack_seq(tp
);
966 if (!tcp_is_fack(tp
) || !tp
->retrans_out
||
967 !after(received_upto
, tp
->lost_retrans_low
) ||
968 icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
971 tcp_for_write_queue(skb
, sk
) {
972 u32 ack_seq
= TCP_SKB_CB(skb
)->ack_seq
;
974 if (skb
== tcp_send_head(sk
))
976 if (cnt
== tp
->retrans_out
)
978 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
981 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
))
984 /* TODO: We would like to get rid of tcp_is_fack(tp) only
985 * constraint here (see above) but figuring out that at
986 * least tp->reordering SACK blocks reside between ack_seq
987 * and received_upto is not easy task to do cheaply with
988 * the available datastructures.
990 * Whether FACK should check here for tp->reordering segs
991 * in-between one could argue for either way (it would be
992 * rather simple to implement as we could count fack_count
993 * during the walk and do tp->fackets_out - fack_count).
995 if (after(received_upto
, ack_seq
)) {
996 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
997 tp
->retrans_out
-= tcp_skb_pcount(skb
);
999 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
1000 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSTRETRANSMIT
);
1002 if (before(ack_seq
, new_low_seq
))
1003 new_low_seq
= ack_seq
;
1004 cnt
+= tcp_skb_pcount(skb
);
1008 if (tp
->retrans_out
)
1009 tp
->lost_retrans_low
= new_low_seq
;
1012 static bool tcp_check_dsack(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1013 struct tcp_sack_block_wire
*sp
, int num_sacks
,
1016 struct tcp_sock
*tp
= tcp_sk(sk
);
1017 u32 start_seq_0
= get_unaligned_be32(&sp
[0].start_seq
);
1018 u32 end_seq_0
= get_unaligned_be32(&sp
[0].end_seq
);
1019 bool dup_sack
= false;
1021 if (before(start_seq_0
, TCP_SKB_CB(ack_skb
)->ack_seq
)) {
1024 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPDSACKRECV
);
1025 } else if (num_sacks
> 1) {
1026 u32 end_seq_1
= get_unaligned_be32(&sp
[1].end_seq
);
1027 u32 start_seq_1
= get_unaligned_be32(&sp
[1].start_seq
);
1029 if (!after(end_seq_0
, end_seq_1
) &&
1030 !before(start_seq_0
, start_seq_1
)) {
1033 NET_INC_STATS_BH(sock_net(sk
),
1034 LINUX_MIB_TCPDSACKOFORECV
);
1038 /* D-SACK for already forgotten data... Do dumb counting. */
1039 if (dup_sack
&& tp
->undo_marker
&& tp
->undo_retrans
&&
1040 !after(end_seq_0
, prior_snd_una
) &&
1041 after(end_seq_0
, tp
->undo_marker
))
1047 struct tcp_sacktag_state
{
1053 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1054 * the incoming SACK may not exactly match but we can find smaller MSS
1055 * aligned portion of it that matches. Therefore we might need to fragment
1056 * which may fail and creates some hassle (caller must handle error case
1059 * FIXME: this could be merged to shift decision code
1061 static int tcp_match_skb_to_sack(struct sock
*sk
, struct sk_buff
*skb
,
1062 u32 start_seq
, u32 end_seq
)
1066 unsigned int pkt_len
;
1069 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1070 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1072 if (tcp_skb_pcount(skb
) > 1 && !in_sack
&&
1073 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
1074 mss
= tcp_skb_mss(skb
);
1075 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1078 pkt_len
= start_seq
- TCP_SKB_CB(skb
)->seq
;
1082 pkt_len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1087 /* Round if necessary so that SACKs cover only full MSSes
1088 * and/or the remaining small portion (if present)
1090 if (pkt_len
> mss
) {
1091 unsigned int new_len
= (pkt_len
/ mss
) * mss
;
1092 if (!in_sack
&& new_len
< pkt_len
) {
1094 if (new_len
> skb
->len
)
1099 err
= tcp_fragment(sk
, skb
, pkt_len
, mss
);
1107 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1108 static u8
tcp_sacktag_one(struct sock
*sk
,
1109 struct tcp_sacktag_state
*state
, u8 sacked
,
1110 u32 start_seq
, u32 end_seq
,
1111 bool dup_sack
, int pcount
)
1113 struct tcp_sock
*tp
= tcp_sk(sk
);
1114 int fack_count
= state
->fack_count
;
1116 /* Account D-SACK for retransmitted packet. */
1117 if (dup_sack
&& (sacked
& TCPCB_RETRANS
)) {
1118 if (tp
->undo_marker
&& tp
->undo_retrans
&&
1119 after(end_seq
, tp
->undo_marker
))
1121 if (sacked
& TCPCB_SACKED_ACKED
)
1122 state
->reord
= min(fack_count
, state
->reord
);
1125 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1126 if (!after(end_seq
, tp
->snd_una
))
1129 if (!(sacked
& TCPCB_SACKED_ACKED
)) {
1130 if (sacked
& TCPCB_SACKED_RETRANS
) {
1131 /* If the segment is not tagged as lost,
1132 * we do not clear RETRANS, believing
1133 * that retransmission is still in flight.
1135 if (sacked
& TCPCB_LOST
) {
1136 sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1137 tp
->lost_out
-= pcount
;
1138 tp
->retrans_out
-= pcount
;
1141 if (!(sacked
& TCPCB_RETRANS
)) {
1142 /* New sack for not retransmitted frame,
1143 * which was in hole. It is reordering.
1145 if (before(start_seq
,
1146 tcp_highest_sack_seq(tp
)))
1147 state
->reord
= min(fack_count
,
1149 if (!after(end_seq
, tp
->high_seq
))
1150 state
->flag
|= FLAG_ORIG_SACK_ACKED
;
1153 if (sacked
& TCPCB_LOST
) {
1154 sacked
&= ~TCPCB_LOST
;
1155 tp
->lost_out
-= pcount
;
1159 sacked
|= TCPCB_SACKED_ACKED
;
1160 state
->flag
|= FLAG_DATA_SACKED
;
1161 tp
->sacked_out
+= pcount
;
1163 fack_count
+= pcount
;
1165 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1166 if (!tcp_is_fack(tp
) && (tp
->lost_skb_hint
!= NULL
) &&
1167 before(start_seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
))
1168 tp
->lost_cnt_hint
+= pcount
;
1170 if (fack_count
> tp
->fackets_out
)
1171 tp
->fackets_out
= fack_count
;
1174 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1175 * frames and clear it. undo_retrans is decreased above, L|R frames
1176 * are accounted above as well.
1178 if (dup_sack
&& (sacked
& TCPCB_SACKED_RETRANS
)) {
1179 sacked
&= ~TCPCB_SACKED_RETRANS
;
1180 tp
->retrans_out
-= pcount
;
1186 /* Shift newly-SACKed bytes from this skb to the immediately previous
1187 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1189 static bool tcp_shifted_skb(struct sock
*sk
, struct sk_buff
*skb
,
1190 struct tcp_sacktag_state
*state
,
1191 unsigned int pcount
, int shifted
, int mss
,
1194 struct tcp_sock
*tp
= tcp_sk(sk
);
1195 struct sk_buff
*prev
= tcp_write_queue_prev(sk
, skb
);
1196 u32 start_seq
= TCP_SKB_CB(skb
)->seq
; /* start of newly-SACKed */
1197 u32 end_seq
= start_seq
+ shifted
; /* end of newly-SACKed */
1201 /* Adjust counters and hints for the newly sacked sequence
1202 * range but discard the return value since prev is already
1203 * marked. We must tag the range first because the seq
1204 * advancement below implicitly advances
1205 * tcp_highest_sack_seq() when skb is highest_sack.
1207 tcp_sacktag_one(sk
, state
, TCP_SKB_CB(skb
)->sacked
,
1208 start_seq
, end_seq
, dup_sack
, pcount
);
1210 if (skb
== tp
->lost_skb_hint
)
1211 tp
->lost_cnt_hint
+= pcount
;
1213 TCP_SKB_CB(prev
)->end_seq
+= shifted
;
1214 TCP_SKB_CB(skb
)->seq
+= shifted
;
1216 skb_shinfo(prev
)->gso_segs
+= pcount
;
1217 BUG_ON(skb_shinfo(skb
)->gso_segs
< pcount
);
1218 skb_shinfo(skb
)->gso_segs
-= pcount
;
1220 /* When we're adding to gso_segs == 1, gso_size will be zero,
1221 * in theory this shouldn't be necessary but as long as DSACK
1222 * code can come after this skb later on it's better to keep
1223 * setting gso_size to something.
1225 if (!skb_shinfo(prev
)->gso_size
) {
1226 skb_shinfo(prev
)->gso_size
= mss
;
1227 skb_shinfo(prev
)->gso_type
= sk
->sk_gso_type
;
1230 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1231 if (skb_shinfo(skb
)->gso_segs
<= 1) {
1232 skb_shinfo(skb
)->gso_size
= 0;
1233 skb_shinfo(skb
)->gso_type
= 0;
1236 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1237 TCP_SKB_CB(prev
)->sacked
|= (TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
);
1240 BUG_ON(!tcp_skb_pcount(skb
));
1241 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKSHIFTED
);
1245 /* Whole SKB was eaten :-) */
1247 if (skb
== tp
->retransmit_skb_hint
)
1248 tp
->retransmit_skb_hint
= prev
;
1249 if (skb
== tp
->lost_skb_hint
) {
1250 tp
->lost_skb_hint
= prev
;
1251 tp
->lost_cnt_hint
-= tcp_skb_pcount(prev
);
1254 TCP_SKB_CB(skb
)->tcp_flags
|= TCP_SKB_CB(prev
)->tcp_flags
;
1255 if (skb
== tcp_highest_sack(sk
))
1256 tcp_advance_highest_sack(sk
, skb
);
1258 tcp_unlink_write_queue(skb
, sk
);
1259 sk_wmem_free_skb(sk
, skb
);
1261 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKMERGED
);
1266 /* I wish gso_size would have a bit more sane initialization than
1267 * something-or-zero which complicates things
1269 static int tcp_skb_seglen(const struct sk_buff
*skb
)
1271 return tcp_skb_pcount(skb
) == 1 ? skb
->len
: tcp_skb_mss(skb
);
1274 /* Shifting pages past head area doesn't work */
1275 static int skb_can_shift(const struct sk_buff
*skb
)
1277 return !skb_headlen(skb
) && skb_is_nonlinear(skb
);
1280 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1283 static struct sk_buff
*tcp_shift_skb_data(struct sock
*sk
, struct sk_buff
*skb
,
1284 struct tcp_sacktag_state
*state
,
1285 u32 start_seq
, u32 end_seq
,
1288 struct tcp_sock
*tp
= tcp_sk(sk
);
1289 struct sk_buff
*prev
;
1295 if (!sk_can_gso(sk
))
1298 /* Normally R but no L won't result in plain S */
1300 (TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_RETRANS
)) == TCPCB_SACKED_RETRANS
)
1302 if (!skb_can_shift(skb
))
1304 /* This frame is about to be dropped (was ACKed). */
1305 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1308 /* Can only happen with delayed DSACK + discard craziness */
1309 if (unlikely(skb
== tcp_write_queue_head(sk
)))
1311 prev
= tcp_write_queue_prev(sk
, skb
);
1313 if ((TCP_SKB_CB(prev
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
)
1316 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1317 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1321 pcount
= tcp_skb_pcount(skb
);
1322 mss
= tcp_skb_seglen(skb
);
1324 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1325 * drop this restriction as unnecessary
1327 if (mss
!= tcp_skb_seglen(prev
))
1330 if (!after(TCP_SKB_CB(skb
)->end_seq
, start_seq
))
1332 /* CHECKME: This is non-MSS split case only?, this will
1333 * cause skipped skbs due to advancing loop btw, original
1334 * has that feature too
1336 if (tcp_skb_pcount(skb
) <= 1)
1339 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1341 /* TODO: head merge to next could be attempted here
1342 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1343 * though it might not be worth of the additional hassle
1345 * ...we can probably just fallback to what was done
1346 * previously. We could try merging non-SACKed ones
1347 * as well but it probably isn't going to buy off
1348 * because later SACKs might again split them, and
1349 * it would make skb timestamp tracking considerably
1355 len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1357 BUG_ON(len
> skb
->len
);
1359 /* MSS boundaries should be honoured or else pcount will
1360 * severely break even though it makes things bit trickier.
1361 * Optimize common case to avoid most of the divides
1363 mss
= tcp_skb_mss(skb
);
1365 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1366 * drop this restriction as unnecessary
1368 if (mss
!= tcp_skb_seglen(prev
))
1373 } else if (len
< mss
) {
1381 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1382 if (!after(TCP_SKB_CB(skb
)->seq
+ len
, tp
->snd_una
))
1385 if (!skb_shift(prev
, skb
, len
))
1387 if (!tcp_shifted_skb(sk
, skb
, state
, pcount
, len
, mss
, dup_sack
))
1390 /* Hole filled allows collapsing with the next as well, this is very
1391 * useful when hole on every nth skb pattern happens
1393 if (prev
== tcp_write_queue_tail(sk
))
1395 skb
= tcp_write_queue_next(sk
, prev
);
1397 if (!skb_can_shift(skb
) ||
1398 (skb
== tcp_send_head(sk
)) ||
1399 ((TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
) ||
1400 (mss
!= tcp_skb_seglen(skb
)))
1404 if (skb_shift(prev
, skb
, len
)) {
1405 pcount
+= tcp_skb_pcount(skb
);
1406 tcp_shifted_skb(sk
, skb
, state
, tcp_skb_pcount(skb
), len
, mss
, 0);
1410 state
->fack_count
+= pcount
;
1417 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKSHIFTFALLBACK
);
1421 static struct sk_buff
*tcp_sacktag_walk(struct sk_buff
*skb
, struct sock
*sk
,
1422 struct tcp_sack_block
*next_dup
,
1423 struct tcp_sacktag_state
*state
,
1424 u32 start_seq
, u32 end_seq
,
1427 struct tcp_sock
*tp
= tcp_sk(sk
);
1428 struct sk_buff
*tmp
;
1430 tcp_for_write_queue_from(skb
, sk
) {
1432 bool dup_sack
= dup_sack_in
;
1434 if (skb
== tcp_send_head(sk
))
1437 /* queue is in-order => we can short-circuit the walk early */
1438 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1441 if ((next_dup
!= NULL
) &&
1442 before(TCP_SKB_CB(skb
)->seq
, next_dup
->end_seq
)) {
1443 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1444 next_dup
->start_seq
,
1450 /* skb reference here is a bit tricky to get right, since
1451 * shifting can eat and free both this skb and the next,
1452 * so not even _safe variant of the loop is enough.
1455 tmp
= tcp_shift_skb_data(sk
, skb
, state
,
1456 start_seq
, end_seq
, dup_sack
);
1465 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1471 if (unlikely(in_sack
< 0))
1475 TCP_SKB_CB(skb
)->sacked
=
1478 TCP_SKB_CB(skb
)->sacked
,
1479 TCP_SKB_CB(skb
)->seq
,
1480 TCP_SKB_CB(skb
)->end_seq
,
1482 tcp_skb_pcount(skb
));
1484 if (!before(TCP_SKB_CB(skb
)->seq
,
1485 tcp_highest_sack_seq(tp
)))
1486 tcp_advance_highest_sack(sk
, skb
);
1489 state
->fack_count
+= tcp_skb_pcount(skb
);
1494 /* Avoid all extra work that is being done by sacktag while walking in
1497 static struct sk_buff
*tcp_sacktag_skip(struct sk_buff
*skb
, struct sock
*sk
,
1498 struct tcp_sacktag_state
*state
,
1501 tcp_for_write_queue_from(skb
, sk
) {
1502 if (skb
== tcp_send_head(sk
))
1505 if (after(TCP_SKB_CB(skb
)->end_seq
, skip_to_seq
))
1508 state
->fack_count
+= tcp_skb_pcount(skb
);
1513 static struct sk_buff
*tcp_maybe_skipping_dsack(struct sk_buff
*skb
,
1515 struct tcp_sack_block
*next_dup
,
1516 struct tcp_sacktag_state
*state
,
1519 if (next_dup
== NULL
)
1522 if (before(next_dup
->start_seq
, skip_to_seq
)) {
1523 skb
= tcp_sacktag_skip(skb
, sk
, state
, next_dup
->start_seq
);
1524 skb
= tcp_sacktag_walk(skb
, sk
, NULL
, state
,
1525 next_dup
->start_seq
, next_dup
->end_seq
,
1532 static int tcp_sack_cache_ok(const struct tcp_sock
*tp
, const struct tcp_sack_block
*cache
)
1534 return cache
< tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1538 tcp_sacktag_write_queue(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1541 struct tcp_sock
*tp
= tcp_sk(sk
);
1542 const unsigned char *ptr
= (skb_transport_header(ack_skb
) +
1543 TCP_SKB_CB(ack_skb
)->sacked
);
1544 struct tcp_sack_block_wire
*sp_wire
= (struct tcp_sack_block_wire
*)(ptr
+2);
1545 struct tcp_sack_block sp
[TCP_NUM_SACKS
];
1546 struct tcp_sack_block
*cache
;
1547 struct tcp_sacktag_state state
;
1548 struct sk_buff
*skb
;
1549 int num_sacks
= min(TCP_NUM_SACKS
, (ptr
[1] - TCPOLEN_SACK_BASE
) >> 3);
1551 bool found_dup_sack
= false;
1553 int first_sack_index
;
1556 state
.reord
= tp
->packets_out
;
1558 if (!tp
->sacked_out
) {
1559 if (WARN_ON(tp
->fackets_out
))
1560 tp
->fackets_out
= 0;
1561 tcp_highest_sack_reset(sk
);
1564 found_dup_sack
= tcp_check_dsack(sk
, ack_skb
, sp_wire
,
1565 num_sacks
, prior_snd_una
);
1567 state
.flag
|= FLAG_DSACKING_ACK
;
1569 /* Eliminate too old ACKs, but take into
1570 * account more or less fresh ones, they can
1571 * contain valid SACK info.
1573 if (before(TCP_SKB_CB(ack_skb
)->ack_seq
, prior_snd_una
- tp
->max_window
))
1576 if (!tp
->packets_out
)
1580 first_sack_index
= 0;
1581 for (i
= 0; i
< num_sacks
; i
++) {
1582 bool dup_sack
= !i
&& found_dup_sack
;
1584 sp
[used_sacks
].start_seq
= get_unaligned_be32(&sp_wire
[i
].start_seq
);
1585 sp
[used_sacks
].end_seq
= get_unaligned_be32(&sp_wire
[i
].end_seq
);
1587 if (!tcp_is_sackblock_valid(tp
, dup_sack
,
1588 sp
[used_sacks
].start_seq
,
1589 sp
[used_sacks
].end_seq
)) {
1593 if (!tp
->undo_marker
)
1594 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDNOUNDO
;
1596 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDOLD
;
1598 /* Don't count olds caused by ACK reordering */
1599 if ((TCP_SKB_CB(ack_skb
)->ack_seq
!= tp
->snd_una
) &&
1600 !after(sp
[used_sacks
].end_seq
, tp
->snd_una
))
1602 mib_idx
= LINUX_MIB_TCPSACKDISCARD
;
1605 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
1607 first_sack_index
= -1;
1611 /* Ignore very old stuff early */
1612 if (!after(sp
[used_sacks
].end_seq
, prior_snd_una
))
1618 /* order SACK blocks to allow in order walk of the retrans queue */
1619 for (i
= used_sacks
- 1; i
> 0; i
--) {
1620 for (j
= 0; j
< i
; j
++) {
1621 if (after(sp
[j
].start_seq
, sp
[j
+ 1].start_seq
)) {
1622 swap(sp
[j
], sp
[j
+ 1]);
1624 /* Track where the first SACK block goes to */
1625 if (j
== first_sack_index
)
1626 first_sack_index
= j
+ 1;
1631 skb
= tcp_write_queue_head(sk
);
1632 state
.fack_count
= 0;
1635 if (!tp
->sacked_out
) {
1636 /* It's already past, so skip checking against it */
1637 cache
= tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1639 cache
= tp
->recv_sack_cache
;
1640 /* Skip empty blocks in at head of the cache */
1641 while (tcp_sack_cache_ok(tp
, cache
) && !cache
->start_seq
&&
1646 while (i
< used_sacks
) {
1647 u32 start_seq
= sp
[i
].start_seq
;
1648 u32 end_seq
= sp
[i
].end_seq
;
1649 bool dup_sack
= (found_dup_sack
&& (i
== first_sack_index
));
1650 struct tcp_sack_block
*next_dup
= NULL
;
1652 if (found_dup_sack
&& ((i
+ 1) == first_sack_index
))
1653 next_dup
= &sp
[i
+ 1];
1655 /* Skip too early cached blocks */
1656 while (tcp_sack_cache_ok(tp
, cache
) &&
1657 !before(start_seq
, cache
->end_seq
))
1660 /* Can skip some work by looking recv_sack_cache? */
1661 if (tcp_sack_cache_ok(tp
, cache
) && !dup_sack
&&
1662 after(end_seq
, cache
->start_seq
)) {
1665 if (before(start_seq
, cache
->start_seq
)) {
1666 skb
= tcp_sacktag_skip(skb
, sk
, &state
,
1668 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
,
1675 /* Rest of the block already fully processed? */
1676 if (!after(end_seq
, cache
->end_seq
))
1679 skb
= tcp_maybe_skipping_dsack(skb
, sk
, next_dup
,
1683 /* ...tail remains todo... */
1684 if (tcp_highest_sack_seq(tp
) == cache
->end_seq
) {
1685 /* ...but better entrypoint exists! */
1686 skb
= tcp_highest_sack(sk
);
1689 state
.fack_count
= tp
->fackets_out
;
1694 skb
= tcp_sacktag_skip(skb
, sk
, &state
, cache
->end_seq
);
1695 /* Check overlap against next cached too (past this one already) */
1700 if (!before(start_seq
, tcp_highest_sack_seq(tp
))) {
1701 skb
= tcp_highest_sack(sk
);
1704 state
.fack_count
= tp
->fackets_out
;
1706 skb
= tcp_sacktag_skip(skb
, sk
, &state
, start_seq
);
1709 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
, &state
,
1710 start_seq
, end_seq
, dup_sack
);
1716 /* Clear the head of the cache sack blocks so we can skip it next time */
1717 for (i
= 0; i
< ARRAY_SIZE(tp
->recv_sack_cache
) - used_sacks
; i
++) {
1718 tp
->recv_sack_cache
[i
].start_seq
= 0;
1719 tp
->recv_sack_cache
[i
].end_seq
= 0;
1721 for (j
= 0; j
< used_sacks
; j
++)
1722 tp
->recv_sack_cache
[i
++] = sp
[j
];
1724 tcp_mark_lost_retrans(sk
);
1726 tcp_verify_left_out(tp
);
1728 if ((state
.reord
< tp
->fackets_out
) &&
1729 ((inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Loss
) || tp
->undo_marker
))
1730 tcp_update_reordering(sk
, tp
->fackets_out
- state
.reord
, 0);
1734 #if FASTRETRANS_DEBUG > 0
1735 WARN_ON((int)tp
->sacked_out
< 0);
1736 WARN_ON((int)tp
->lost_out
< 0);
1737 WARN_ON((int)tp
->retrans_out
< 0);
1738 WARN_ON((int)tcp_packets_in_flight(tp
) < 0);
1743 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1744 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1746 static bool tcp_limit_reno_sacked(struct tcp_sock
*tp
)
1750 holes
= max(tp
->lost_out
, 1U);
1751 holes
= min(holes
, tp
->packets_out
);
1753 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1754 tp
->sacked_out
= tp
->packets_out
- holes
;
1760 /* If we receive more dupacks than we expected counting segments
1761 * in assumption of absent reordering, interpret this as reordering.
1762 * The only another reason could be bug in receiver TCP.
1764 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1766 struct tcp_sock
*tp
= tcp_sk(sk
);
1767 if (tcp_limit_reno_sacked(tp
))
1768 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1771 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1773 static void tcp_add_reno_sack(struct sock
*sk
)
1775 struct tcp_sock
*tp
= tcp_sk(sk
);
1777 tcp_check_reno_reordering(sk
, 0);
1778 tcp_verify_left_out(tp
);
1781 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1783 static void tcp_remove_reno_sacks(struct sock
*sk
, int acked
)
1785 struct tcp_sock
*tp
= tcp_sk(sk
);
1788 /* One ACK acked hole. The rest eat duplicate ACKs. */
1789 if (acked
- 1 >= tp
->sacked_out
)
1792 tp
->sacked_out
-= acked
- 1;
1794 tcp_check_reno_reordering(sk
, acked
);
1795 tcp_verify_left_out(tp
);
1798 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1803 static void tcp_clear_retrans_partial(struct tcp_sock
*tp
)
1805 tp
->retrans_out
= 0;
1808 tp
->undo_marker
= 0;
1809 tp
->undo_retrans
= 0;
1812 void tcp_clear_retrans(struct tcp_sock
*tp
)
1814 tcp_clear_retrans_partial(tp
);
1816 tp
->fackets_out
= 0;
1820 /* Enter Loss state. If "how" is not zero, forget all SACK information
1821 * and reset tags completely, otherwise preserve SACKs. If receiver
1822 * dropped its ofo queue, we will know this due to reneging detection.
1824 void tcp_enter_loss(struct sock
*sk
, int how
)
1826 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1827 struct tcp_sock
*tp
= tcp_sk(sk
);
1828 struct sk_buff
*skb
;
1829 bool new_recovery
= false;
1831 /* Reduce ssthresh if it has not yet been made inside this window. */
1832 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
||
1833 !after(tp
->high_seq
, tp
->snd_una
) ||
1834 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
1835 new_recovery
= true;
1836 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1837 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1838 tcp_ca_event(sk
, CA_EVENT_LOSS
);
1841 tp
->snd_cwnd_cnt
= 0;
1842 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1844 tcp_clear_retrans_partial(tp
);
1846 if (tcp_is_reno(tp
))
1847 tcp_reset_reno_sack(tp
);
1849 tp
->undo_marker
= tp
->snd_una
;
1852 tp
->fackets_out
= 0;
1854 tcp_clear_all_retrans_hints(tp
);
1856 tcp_for_write_queue(skb
, sk
) {
1857 if (skb
== tcp_send_head(sk
))
1860 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
1861 tp
->undo_marker
= 0;
1862 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
1863 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) || how
) {
1864 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
1865 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1866 tp
->lost_out
+= tcp_skb_pcount(skb
);
1867 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
1870 tcp_verify_left_out(tp
);
1872 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1873 sysctl_tcp_reordering
);
1874 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1875 tp
->high_seq
= tp
->snd_nxt
;
1876 TCP_ECN_queue_cwr(tp
);
1878 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1879 * loss recovery is underway except recurring timeout(s) on
1880 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1882 tp
->frto
= sysctl_tcp_frto
&&
1883 (new_recovery
|| icsk
->icsk_retransmits
) &&
1884 !inet_csk(sk
)->icsk_mtup
.probe_size
;
1887 /* If ACK arrived pointing to a remembered SACK, it means that our
1888 * remembered SACKs do not reflect real state of receiver i.e.
1889 * receiver _host_ is heavily congested (or buggy).
1891 * Do processing similar to RTO timeout.
1893 static bool tcp_check_sack_reneging(struct sock
*sk
, int flag
)
1895 if (flag
& FLAG_SACK_RENEGING
) {
1896 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1897 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSACKRENEGING
);
1899 tcp_enter_loss(sk
, 1);
1900 icsk
->icsk_retransmits
++;
1901 tcp_retransmit_skb(sk
, tcp_write_queue_head(sk
));
1902 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1903 icsk
->icsk_rto
, TCP_RTO_MAX
);
1909 static inline int tcp_fackets_out(const struct tcp_sock
*tp
)
1911 return tcp_is_reno(tp
) ? tp
->sacked_out
+ 1 : tp
->fackets_out
;
1914 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
1915 * counter when SACK is enabled (without SACK, sacked_out is used for
1918 * Instead, with FACK TCP uses fackets_out that includes both SACKed
1919 * segments up to the highest received SACK block so far and holes in
1922 * With reordering, holes may still be in flight, so RFC3517 recovery
1923 * uses pure sacked_out (total number of SACKed segments) even though
1924 * it violates the RFC that uses duplicate ACKs, often these are equal
1925 * but when e.g. out-of-window ACKs or packet duplication occurs,
1926 * they differ. Since neither occurs due to loss, TCP should really
1929 static inline int tcp_dupack_heuristics(const struct tcp_sock
*tp
)
1931 return tcp_is_fack(tp
) ? tp
->fackets_out
: tp
->sacked_out
+ 1;
1934 static bool tcp_pause_early_retransmit(struct sock
*sk
, int flag
)
1936 struct tcp_sock
*tp
= tcp_sk(sk
);
1937 unsigned long delay
;
1939 /* Delay early retransmit and entering fast recovery for
1940 * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
1941 * available, or RTO is scheduled to fire first.
1943 if (sysctl_tcp_early_retrans
< 2 || sysctl_tcp_early_retrans
> 3 ||
1944 (flag
& FLAG_ECE
) || !tp
->srtt
)
1947 delay
= max_t(unsigned long, (tp
->srtt
>> 5), msecs_to_jiffies(2));
1948 if (!time_after(inet_csk(sk
)->icsk_timeout
, (jiffies
+ delay
)))
1951 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_EARLY_RETRANS
, delay
,
1956 /* Linux NewReno/SACK/FACK/ECN state machine.
1957 * --------------------------------------
1959 * "Open" Normal state, no dubious events, fast path.
1960 * "Disorder" In all the respects it is "Open",
1961 * but requires a bit more attention. It is entered when
1962 * we see some SACKs or dupacks. It is split of "Open"
1963 * mainly to move some processing from fast path to slow one.
1964 * "CWR" CWND was reduced due to some Congestion Notification event.
1965 * It can be ECN, ICMP source quench, local device congestion.
1966 * "Recovery" CWND was reduced, we are fast-retransmitting.
1967 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
1969 * tcp_fastretrans_alert() is entered:
1970 * - each incoming ACK, if state is not "Open"
1971 * - when arrived ACK is unusual, namely:
1976 * Counting packets in flight is pretty simple.
1978 * in_flight = packets_out - left_out + retrans_out
1980 * packets_out is SND.NXT-SND.UNA counted in packets.
1982 * retrans_out is number of retransmitted segments.
1984 * left_out is number of segments left network, but not ACKed yet.
1986 * left_out = sacked_out + lost_out
1988 * sacked_out: Packets, which arrived to receiver out of order
1989 * and hence not ACKed. With SACKs this number is simply
1990 * amount of SACKed data. Even without SACKs
1991 * it is easy to give pretty reliable estimate of this number,
1992 * counting duplicate ACKs.
1994 * lost_out: Packets lost by network. TCP has no explicit
1995 * "loss notification" feedback from network (for now).
1996 * It means that this number can be only _guessed_.
1997 * Actually, it is the heuristics to predict lossage that
1998 * distinguishes different algorithms.
2000 * F.e. after RTO, when all the queue is considered as lost,
2001 * lost_out = packets_out and in_flight = retrans_out.
2003 * Essentially, we have now two algorithms counting
2006 * FACK: It is the simplest heuristics. As soon as we decided
2007 * that something is lost, we decide that _all_ not SACKed
2008 * packets until the most forward SACK are lost. I.e.
2009 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2010 * It is absolutely correct estimate, if network does not reorder
2011 * packets. And it loses any connection to reality when reordering
2012 * takes place. We use FACK by default until reordering
2013 * is suspected on the path to this destination.
2015 * NewReno: when Recovery is entered, we assume that one segment
2016 * is lost (classic Reno). While we are in Recovery and
2017 * a partial ACK arrives, we assume that one more packet
2018 * is lost (NewReno). This heuristics are the same in NewReno
2021 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2022 * deflation etc. CWND is real congestion window, never inflated, changes
2023 * only according to classic VJ rules.
2025 * Really tricky (and requiring careful tuning) part of algorithm
2026 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2027 * The first determines the moment _when_ we should reduce CWND and,
2028 * hence, slow down forward transmission. In fact, it determines the moment
2029 * when we decide that hole is caused by loss, rather than by a reorder.
2031 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2032 * holes, caused by lost packets.
2034 * And the most logically complicated part of algorithm is undo
2035 * heuristics. We detect false retransmits due to both too early
2036 * fast retransmit (reordering) and underestimated RTO, analyzing
2037 * timestamps and D-SACKs. When we detect that some segments were
2038 * retransmitted by mistake and CWND reduction was wrong, we undo
2039 * window reduction and abort recovery phase. This logic is hidden
2040 * inside several functions named tcp_try_undo_<something>.
2043 /* This function decides, when we should leave Disordered state
2044 * and enter Recovery phase, reducing congestion window.
2046 * Main question: may we further continue forward transmission
2047 * with the same cwnd?
2049 static bool tcp_time_to_recover(struct sock
*sk
, int flag
)
2051 struct tcp_sock
*tp
= tcp_sk(sk
);
2054 /* Trick#1: The loss is proven. */
2058 /* Not-A-Trick#2 : Classic rule... */
2059 if (tcp_dupack_heuristics(tp
) > tp
->reordering
)
2062 /* Trick#4: It is still not OK... But will it be useful to delay
2065 packets_out
= tp
->packets_out
;
2066 if (packets_out
<= tp
->reordering
&&
2067 tp
->sacked_out
>= max_t(__u32
, packets_out
/2, sysctl_tcp_reordering
) &&
2068 !tcp_may_send_now(sk
)) {
2069 /* We have nothing to send. This connection is limited
2070 * either by receiver window or by application.
2075 /* If a thin stream is detected, retransmit after first
2076 * received dupack. Employ only if SACK is supported in order
2077 * to avoid possible corner-case series of spurious retransmissions
2078 * Use only if there are no unsent data.
2080 if ((tp
->thin_dupack
|| sysctl_tcp_thin_dupack
) &&
2081 tcp_stream_is_thin(tp
) && tcp_dupack_heuristics(tp
) > 1 &&
2082 tcp_is_sack(tp
) && !tcp_send_head(sk
))
2085 /* Trick#6: TCP early retransmit, per RFC5827. To avoid spurious
2086 * retransmissions due to small network reorderings, we implement
2087 * Mitigation A.3 in the RFC and delay the retransmission for a short
2088 * interval if appropriate.
2090 if (tp
->do_early_retrans
&& !tp
->retrans_out
&& tp
->sacked_out
&&
2091 (tp
->packets_out
>= (tp
->sacked_out
+ 1) && tp
->packets_out
< 4) &&
2092 !tcp_may_send_now(sk
))
2093 return !tcp_pause_early_retransmit(sk
, flag
);
2098 /* Detect loss in event "A" above by marking head of queue up as lost.
2099 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2100 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2101 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2102 * the maximum SACKed segments to pass before reaching this limit.
2104 static void tcp_mark_head_lost(struct sock
*sk
, int packets
, int mark_head
)
2106 struct tcp_sock
*tp
= tcp_sk(sk
);
2107 struct sk_buff
*skb
;
2111 /* Use SACK to deduce losses of new sequences sent during recovery */
2112 const u32 loss_high
= tcp_is_sack(tp
) ? tp
->snd_nxt
: tp
->high_seq
;
2114 WARN_ON(packets
> tp
->packets_out
);
2115 if (tp
->lost_skb_hint
) {
2116 skb
= tp
->lost_skb_hint
;
2117 cnt
= tp
->lost_cnt_hint
;
2118 /* Head already handled? */
2119 if (mark_head
&& skb
!= tcp_write_queue_head(sk
))
2122 skb
= tcp_write_queue_head(sk
);
2126 tcp_for_write_queue_from(skb
, sk
) {
2127 if (skb
== tcp_send_head(sk
))
2129 /* TODO: do this better */
2130 /* this is not the most efficient way to do this... */
2131 tp
->lost_skb_hint
= skb
;
2132 tp
->lost_cnt_hint
= cnt
;
2134 if (after(TCP_SKB_CB(skb
)->end_seq
, loss_high
))
2138 if (tcp_is_fack(tp
) || tcp_is_reno(tp
) ||
2139 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2140 cnt
+= tcp_skb_pcount(skb
);
2142 if (cnt
> packets
) {
2143 if ((tcp_is_sack(tp
) && !tcp_is_fack(tp
)) ||
2144 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
) ||
2145 (oldcnt
>= packets
))
2148 mss
= skb_shinfo(skb
)->gso_size
;
2149 err
= tcp_fragment(sk
, skb
, (packets
- oldcnt
) * mss
, mss
);
2155 tcp_skb_mark_lost(tp
, skb
);
2160 tcp_verify_left_out(tp
);
2163 /* Account newly detected lost packet(s) */
2165 static void tcp_update_scoreboard(struct sock
*sk
, int fast_rexmit
)
2167 struct tcp_sock
*tp
= tcp_sk(sk
);
2169 if (tcp_is_reno(tp
)) {
2170 tcp_mark_head_lost(sk
, 1, 1);
2171 } else if (tcp_is_fack(tp
)) {
2172 int lost
= tp
->fackets_out
- tp
->reordering
;
2175 tcp_mark_head_lost(sk
, lost
, 0);
2177 int sacked_upto
= tp
->sacked_out
- tp
->reordering
;
2178 if (sacked_upto
>= 0)
2179 tcp_mark_head_lost(sk
, sacked_upto
, 0);
2180 else if (fast_rexmit
)
2181 tcp_mark_head_lost(sk
, 1, 1);
2185 /* CWND moderation, preventing bursts due to too big ACKs
2186 * in dubious situations.
2188 static inline void tcp_moderate_cwnd(struct tcp_sock
*tp
)
2190 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
2191 tcp_packets_in_flight(tp
) + tcp_max_burst(tp
));
2192 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2195 /* Nothing was retransmitted or returned timestamp is less
2196 * than timestamp of the first retransmission.
2198 static inline bool tcp_packet_delayed(const struct tcp_sock
*tp
)
2200 return !tp
->retrans_stamp
||
2201 (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2202 before(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
));
2205 /* Undo procedures. */
2207 #if FASTRETRANS_DEBUG > 1
2208 static void DBGUNDO(struct sock
*sk
, const char *msg
)
2210 struct tcp_sock
*tp
= tcp_sk(sk
);
2211 struct inet_sock
*inet
= inet_sk(sk
);
2213 if (sk
->sk_family
== AF_INET
) {
2214 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2216 &inet
->inet_daddr
, ntohs(inet
->inet_dport
),
2217 tp
->snd_cwnd
, tcp_left_out(tp
),
2218 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2221 #if IS_ENABLED(CONFIG_IPV6)
2222 else if (sk
->sk_family
== AF_INET6
) {
2223 struct ipv6_pinfo
*np
= inet6_sk(sk
);
2224 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2226 &np
->daddr
, ntohs(inet
->inet_dport
),
2227 tp
->snd_cwnd
, tcp_left_out(tp
),
2228 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2234 #define DBGUNDO(x...) do { } while (0)
2237 static void tcp_undo_cwnd_reduction(struct sock
*sk
, bool unmark_loss
)
2239 struct tcp_sock
*tp
= tcp_sk(sk
);
2242 struct sk_buff
*skb
;
2244 tcp_for_write_queue(skb
, sk
) {
2245 if (skb
== tcp_send_head(sk
))
2247 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2250 tcp_clear_all_retrans_hints(tp
);
2253 if (tp
->prior_ssthresh
) {
2254 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2256 if (icsk
->icsk_ca_ops
->undo_cwnd
)
2257 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
2259 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
<< 1);
2261 if (tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
2262 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
2263 TCP_ECN_withdraw_cwr(tp
);
2266 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2268 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2269 tp
->undo_marker
= 0;
2272 static inline bool tcp_may_undo(const struct tcp_sock
*tp
)
2274 return tp
->undo_marker
&& (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
2277 /* People celebrate: "We love our President!" */
2278 static bool tcp_try_undo_recovery(struct sock
*sk
)
2280 struct tcp_sock
*tp
= tcp_sk(sk
);
2282 if (tcp_may_undo(tp
)) {
2285 /* Happy end! We did not retransmit anything
2286 * or our original transmission succeeded.
2288 DBGUNDO(sk
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
2289 tcp_undo_cwnd_reduction(sk
, false);
2290 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
2291 mib_idx
= LINUX_MIB_TCPLOSSUNDO
;
2293 mib_idx
= LINUX_MIB_TCPFULLUNDO
;
2295 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2297 if (tp
->snd_una
== tp
->high_seq
&& tcp_is_reno(tp
)) {
2298 /* Hold old state until something *above* high_seq
2299 * is ACKed. For Reno it is MUST to prevent false
2300 * fast retransmits (RFC2582). SACK TCP is safe. */
2301 tcp_moderate_cwnd(tp
);
2304 tcp_set_ca_state(sk
, TCP_CA_Open
);
2308 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2309 static bool tcp_try_undo_dsack(struct sock
*sk
)
2311 struct tcp_sock
*tp
= tcp_sk(sk
);
2313 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
2314 DBGUNDO(sk
, "D-SACK");
2315 tcp_undo_cwnd_reduction(sk
, false);
2316 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPDSACKUNDO
);
2322 /* We can clear retrans_stamp when there are no retransmissions in the
2323 * window. It would seem that it is trivially available for us in
2324 * tp->retrans_out, however, that kind of assumptions doesn't consider
2325 * what will happen if errors occur when sending retransmission for the
2326 * second time. ...It could the that such segment has only
2327 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2328 * the head skb is enough except for some reneging corner cases that
2329 * are not worth the effort.
2331 * Main reason for all this complexity is the fact that connection dying
2332 * time now depends on the validity of the retrans_stamp, in particular,
2333 * that successive retransmissions of a segment must not advance
2334 * retrans_stamp under any conditions.
2336 static bool tcp_any_retrans_done(const struct sock
*sk
)
2338 const struct tcp_sock
*tp
= tcp_sk(sk
);
2339 struct sk_buff
*skb
;
2341 if (tp
->retrans_out
)
2344 skb
= tcp_write_queue_head(sk
);
2345 if (unlikely(skb
&& TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
))
2351 /* Undo during loss recovery after partial ACK or using F-RTO. */
2352 static bool tcp_try_undo_loss(struct sock
*sk
, bool frto_undo
)
2354 struct tcp_sock
*tp
= tcp_sk(sk
);
2356 if (frto_undo
|| tcp_may_undo(tp
)) {
2357 tcp_undo_cwnd_reduction(sk
, true);
2359 DBGUNDO(sk
, "partial loss");
2360 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSSUNDO
);
2362 NET_INC_STATS_BH(sock_net(sk
),
2363 LINUX_MIB_TCPSPURIOUSRTOS
);
2364 inet_csk(sk
)->icsk_retransmits
= 0;
2365 if (frto_undo
|| tcp_is_sack(tp
))
2366 tcp_set_ca_state(sk
, TCP_CA_Open
);
2372 /* The cwnd reduction in CWR and Recovery use the PRR algorithm
2373 * https://datatracker.ietf.org/doc/draft-ietf-tcpm-proportional-rate-reduction/
2374 * It computes the number of packets to send (sndcnt) based on packets newly
2376 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2377 * cwnd reductions across a full RTT.
2378 * 2) If packets in flight is lower than ssthresh (such as due to excess
2379 * losses and/or application stalls), do not perform any further cwnd
2380 * reductions, but instead slow start up to ssthresh.
2382 static void tcp_init_cwnd_reduction(struct sock
*sk
, const bool set_ssthresh
)
2384 struct tcp_sock
*tp
= tcp_sk(sk
);
2386 tp
->high_seq
= tp
->snd_nxt
;
2387 tp
->tlp_high_seq
= 0;
2388 tp
->snd_cwnd_cnt
= 0;
2389 tp
->prior_cwnd
= tp
->snd_cwnd
;
2390 tp
->prr_delivered
= 0;
2393 tp
->snd_ssthresh
= inet_csk(sk
)->icsk_ca_ops
->ssthresh(sk
);
2394 TCP_ECN_queue_cwr(tp
);
2397 static void tcp_cwnd_reduction(struct sock
*sk
, const int prior_unsacked
,
2400 struct tcp_sock
*tp
= tcp_sk(sk
);
2402 int delta
= tp
->snd_ssthresh
- tcp_packets_in_flight(tp
);
2403 int newly_acked_sacked
= prior_unsacked
-
2404 (tp
->packets_out
- tp
->sacked_out
);
2406 tp
->prr_delivered
+= newly_acked_sacked
;
2407 if (tcp_packets_in_flight(tp
) > tp
->snd_ssthresh
) {
2408 u64 dividend
= (u64
)tp
->snd_ssthresh
* tp
->prr_delivered
+
2410 sndcnt
= div_u64(dividend
, tp
->prior_cwnd
) - tp
->prr_out
;
2412 sndcnt
= min_t(int, delta
,
2413 max_t(int, tp
->prr_delivered
- tp
->prr_out
,
2414 newly_acked_sacked
) + 1);
2417 sndcnt
= max(sndcnt
, (fast_rexmit
? 1 : 0));
2418 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + sndcnt
;
2421 static inline void tcp_end_cwnd_reduction(struct sock
*sk
)
2423 struct tcp_sock
*tp
= tcp_sk(sk
);
2425 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2426 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_CWR
||
2427 (tp
->undo_marker
&& tp
->snd_ssthresh
< TCP_INFINITE_SSTHRESH
)) {
2428 tp
->snd_cwnd
= tp
->snd_ssthresh
;
2429 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2431 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
2434 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2435 void tcp_enter_cwr(struct sock
*sk
, const int set_ssthresh
)
2437 struct tcp_sock
*tp
= tcp_sk(sk
);
2439 tp
->prior_ssthresh
= 0;
2440 if (inet_csk(sk
)->icsk_ca_state
< TCP_CA_CWR
) {
2441 tp
->undo_marker
= 0;
2442 tcp_init_cwnd_reduction(sk
, set_ssthresh
);
2443 tcp_set_ca_state(sk
, TCP_CA_CWR
);
2447 static void tcp_try_keep_open(struct sock
*sk
)
2449 struct tcp_sock
*tp
= tcp_sk(sk
);
2450 int state
= TCP_CA_Open
;
2452 if (tcp_left_out(tp
) || tcp_any_retrans_done(sk
))
2453 state
= TCP_CA_Disorder
;
2455 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
2456 tcp_set_ca_state(sk
, state
);
2457 tp
->high_seq
= tp
->snd_nxt
;
2461 static void tcp_try_to_open(struct sock
*sk
, int flag
, const int prior_unsacked
)
2463 struct tcp_sock
*tp
= tcp_sk(sk
);
2465 tcp_verify_left_out(tp
);
2467 if (!tcp_any_retrans_done(sk
))
2468 tp
->retrans_stamp
= 0;
2470 if (flag
& FLAG_ECE
)
2471 tcp_enter_cwr(sk
, 1);
2473 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
2474 tcp_try_keep_open(sk
);
2475 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
)
2476 tcp_moderate_cwnd(tp
);
2478 tcp_cwnd_reduction(sk
, prior_unsacked
, 0);
2482 static void tcp_mtup_probe_failed(struct sock
*sk
)
2484 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2486 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
2487 icsk
->icsk_mtup
.probe_size
= 0;
2490 static void tcp_mtup_probe_success(struct sock
*sk
)
2492 struct tcp_sock
*tp
= tcp_sk(sk
);
2493 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2495 /* FIXME: breaks with very large cwnd */
2496 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2497 tp
->snd_cwnd
= tp
->snd_cwnd
*
2498 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
2499 icsk
->icsk_mtup
.probe_size
;
2500 tp
->snd_cwnd_cnt
= 0;
2501 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2502 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2504 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
2505 icsk
->icsk_mtup
.probe_size
= 0;
2506 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
2509 /* Do a simple retransmit without using the backoff mechanisms in
2510 * tcp_timer. This is used for path mtu discovery.
2511 * The socket is already locked here.
2513 void tcp_simple_retransmit(struct sock
*sk
)
2515 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2516 struct tcp_sock
*tp
= tcp_sk(sk
);
2517 struct sk_buff
*skb
;
2518 unsigned int mss
= tcp_current_mss(sk
);
2519 u32 prior_lost
= tp
->lost_out
;
2521 tcp_for_write_queue(skb
, sk
) {
2522 if (skb
== tcp_send_head(sk
))
2524 if (tcp_skb_seglen(skb
) > mss
&&
2525 !(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
2526 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2527 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
2528 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2530 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
2534 tcp_clear_retrans_hints_partial(tp
);
2536 if (prior_lost
== tp
->lost_out
)
2539 if (tcp_is_reno(tp
))
2540 tcp_limit_reno_sacked(tp
);
2542 tcp_verify_left_out(tp
);
2544 /* Don't muck with the congestion window here.
2545 * Reason is that we do not increase amount of _data_
2546 * in network, but units changed and effective
2547 * cwnd/ssthresh really reduced now.
2549 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
2550 tp
->high_seq
= tp
->snd_nxt
;
2551 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2552 tp
->prior_ssthresh
= 0;
2553 tp
->undo_marker
= 0;
2554 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2556 tcp_xmit_retransmit_queue(sk
);
2558 EXPORT_SYMBOL(tcp_simple_retransmit
);
2560 static void tcp_enter_recovery(struct sock
*sk
, bool ece_ack
)
2562 struct tcp_sock
*tp
= tcp_sk(sk
);
2565 if (tcp_is_reno(tp
))
2566 mib_idx
= LINUX_MIB_TCPRENORECOVERY
;
2568 mib_idx
= LINUX_MIB_TCPSACKRECOVERY
;
2570 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2572 tp
->prior_ssthresh
= 0;
2573 tp
->undo_marker
= tp
->snd_una
;
2574 tp
->undo_retrans
= tp
->retrans_out
;
2576 if (inet_csk(sk
)->icsk_ca_state
< TCP_CA_CWR
) {
2578 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2579 tcp_init_cwnd_reduction(sk
, true);
2581 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
2584 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2585 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2587 static void tcp_process_loss(struct sock
*sk
, int flag
, bool is_dupack
)
2589 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2590 struct tcp_sock
*tp
= tcp_sk(sk
);
2591 bool recovered
= !before(tp
->snd_una
, tp
->high_seq
);
2593 if (tp
->frto
) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2594 if (flag
& FLAG_ORIG_SACK_ACKED
) {
2595 /* Step 3.b. A timeout is spurious if not all data are
2596 * lost, i.e., never-retransmitted data are (s)acked.
2598 tcp_try_undo_loss(sk
, true);
2601 if (after(tp
->snd_nxt
, tp
->high_seq
) &&
2602 (flag
& FLAG_DATA_SACKED
|| is_dupack
)) {
2603 tp
->frto
= 0; /* Loss was real: 2nd part of step 3.a */
2604 } else if (flag
& FLAG_SND_UNA_ADVANCED
&& !recovered
) {
2605 tp
->high_seq
= tp
->snd_nxt
;
2606 __tcp_push_pending_frames(sk
, tcp_current_mss(sk
),
2608 if (after(tp
->snd_nxt
, tp
->high_seq
))
2609 return; /* Step 2.b */
2615 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2616 icsk
->icsk_retransmits
= 0;
2617 tcp_try_undo_recovery(sk
);
2620 if (flag
& FLAG_DATA_ACKED
)
2621 icsk
->icsk_retransmits
= 0;
2622 if (tcp_is_reno(tp
)) {
2623 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2624 * delivered. Lower inflight to clock out (re)tranmissions.
2626 if (after(tp
->snd_nxt
, tp
->high_seq
) && is_dupack
)
2627 tcp_add_reno_sack(sk
);
2628 else if (flag
& FLAG_SND_UNA_ADVANCED
)
2629 tcp_reset_reno_sack(tp
);
2631 if (tcp_try_undo_loss(sk
, false))
2633 tcp_xmit_retransmit_queue(sk
);
2636 /* Undo during fast recovery after partial ACK. */
2637 static bool tcp_try_undo_partial(struct sock
*sk
, const int acked
,
2638 const int prior_unsacked
)
2640 struct tcp_sock
*tp
= tcp_sk(sk
);
2642 if (tp
->undo_marker
&& tcp_packet_delayed(tp
)) {
2643 /* Plain luck! Hole if filled with delayed
2644 * packet, rather than with a retransmit.
2646 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
2648 /* We are getting evidence that the reordering degree is higher
2649 * than we realized. If there are no retransmits out then we
2650 * can undo. Otherwise we clock out new packets but do not
2651 * mark more packets lost or retransmit more.
2653 if (tp
->retrans_out
) {
2654 tcp_cwnd_reduction(sk
, prior_unsacked
, 0);
2658 if (!tcp_any_retrans_done(sk
))
2659 tp
->retrans_stamp
= 0;
2661 DBGUNDO(sk
, "partial recovery");
2662 tcp_undo_cwnd_reduction(sk
, true);
2663 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPARTIALUNDO
);
2664 tcp_try_keep_open(sk
);
2670 /* Process an event, which can update packets-in-flight not trivially.
2671 * Main goal of this function is to calculate new estimate for left_out,
2672 * taking into account both packets sitting in receiver's buffer and
2673 * packets lost by network.
2675 * Besides that it does CWND reduction, when packet loss is detected
2676 * and changes state of machine.
2678 * It does _not_ decide what to send, it is made in function
2679 * tcp_xmit_retransmit_queue().
2681 static void tcp_fastretrans_alert(struct sock
*sk
, const int acked
,
2682 const int prior_unsacked
,
2683 bool is_dupack
, int flag
)
2685 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2686 struct tcp_sock
*tp
= tcp_sk(sk
);
2687 bool do_lost
= is_dupack
|| ((flag
& FLAG_DATA_SACKED
) &&
2688 (tcp_fackets_out(tp
) > tp
->reordering
));
2689 int fast_rexmit
= 0;
2691 if (WARN_ON(!tp
->packets_out
&& tp
->sacked_out
))
2693 if (WARN_ON(!tp
->sacked_out
&& tp
->fackets_out
))
2694 tp
->fackets_out
= 0;
2696 /* Now state machine starts.
2697 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2698 if (flag
& FLAG_ECE
)
2699 tp
->prior_ssthresh
= 0;
2701 /* B. In all the states check for reneging SACKs. */
2702 if (tcp_check_sack_reneging(sk
, flag
))
2705 /* C. Check consistency of the current state. */
2706 tcp_verify_left_out(tp
);
2708 /* D. Check state exit conditions. State can be terminated
2709 * when high_seq is ACKed. */
2710 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
2711 WARN_ON(tp
->retrans_out
!= 0);
2712 tp
->retrans_stamp
= 0;
2713 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
2714 switch (icsk
->icsk_ca_state
) {
2716 /* CWR is to be held something *above* high_seq
2717 * is ACKed for CWR bit to reach receiver. */
2718 if (tp
->snd_una
!= tp
->high_seq
) {
2719 tcp_end_cwnd_reduction(sk
);
2720 tcp_set_ca_state(sk
, TCP_CA_Open
);
2724 case TCP_CA_Recovery
:
2725 if (tcp_is_reno(tp
))
2726 tcp_reset_reno_sack(tp
);
2727 if (tcp_try_undo_recovery(sk
))
2729 tcp_end_cwnd_reduction(sk
);
2734 /* E. Process state. */
2735 switch (icsk
->icsk_ca_state
) {
2736 case TCP_CA_Recovery
:
2737 if (!(flag
& FLAG_SND_UNA_ADVANCED
)) {
2738 if (tcp_is_reno(tp
) && is_dupack
)
2739 tcp_add_reno_sack(sk
);
2741 if (tcp_try_undo_partial(sk
, acked
, prior_unsacked
))
2743 /* Partial ACK arrived. Force fast retransmit. */
2744 do_lost
= tcp_is_reno(tp
) ||
2745 tcp_fackets_out(tp
) > tp
->reordering
;
2747 if (tcp_try_undo_dsack(sk
)) {
2748 tcp_try_keep_open(sk
);
2753 tcp_process_loss(sk
, flag
, is_dupack
);
2754 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
2756 /* Fall through to processing in Open state. */
2758 if (tcp_is_reno(tp
)) {
2759 if (flag
& FLAG_SND_UNA_ADVANCED
)
2760 tcp_reset_reno_sack(tp
);
2762 tcp_add_reno_sack(sk
);
2765 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
)
2766 tcp_try_undo_dsack(sk
);
2768 if (!tcp_time_to_recover(sk
, flag
)) {
2769 tcp_try_to_open(sk
, flag
, prior_unsacked
);
2773 /* MTU probe failure: don't reduce cwnd */
2774 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
2775 icsk
->icsk_mtup
.probe_size
&&
2776 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
2777 tcp_mtup_probe_failed(sk
);
2778 /* Restores the reduction we did in tcp_mtup_probe() */
2780 tcp_simple_retransmit(sk
);
2784 /* Otherwise enter Recovery state */
2785 tcp_enter_recovery(sk
, (flag
& FLAG_ECE
));
2790 tcp_update_scoreboard(sk
, fast_rexmit
);
2791 tcp_cwnd_reduction(sk
, prior_unsacked
, fast_rexmit
);
2792 tcp_xmit_retransmit_queue(sk
);
2795 void tcp_valid_rtt_meas(struct sock
*sk
, u32 seq_rtt
)
2797 tcp_rtt_estimator(sk
, seq_rtt
);
2799 inet_csk(sk
)->icsk_backoff
= 0;
2801 EXPORT_SYMBOL(tcp_valid_rtt_meas
);
2803 /* Read draft-ietf-tcplw-high-performance before mucking
2804 * with this code. (Supersedes RFC1323)
2806 static void tcp_ack_saw_tstamp(struct sock
*sk
, int flag
)
2808 /* RTTM Rule: A TSecr value received in a segment is used to
2809 * update the averaged RTT measurement only if the segment
2810 * acknowledges some new data, i.e., only if it advances the
2811 * left edge of the send window.
2813 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2814 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2816 * Changed: reset backoff as soon as we see the first valid sample.
2817 * If we do not, we get strongly overestimated rto. With timestamps
2818 * samples are accepted even from very old segments: f.e., when rtt=1
2819 * increases to 8, we retransmit 5 times and after 8 seconds delayed
2820 * answer arrives rto becomes 120 seconds! If at least one of segments
2821 * in window is lost... Voila. --ANK (010210)
2823 struct tcp_sock
*tp
= tcp_sk(sk
);
2825 tcp_valid_rtt_meas(sk
, tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
);
2828 static void tcp_ack_no_tstamp(struct sock
*sk
, u32 seq_rtt
, int flag
)
2830 /* We don't have a timestamp. Can only use
2831 * packets that are not retransmitted to determine
2832 * rtt estimates. Also, we must not reset the
2833 * backoff for rto until we get a non-retransmitted
2834 * packet. This allows us to deal with a situation
2835 * where the network delay has increased suddenly.
2836 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2839 if (flag
& FLAG_RETRANS_DATA_ACKED
)
2842 tcp_valid_rtt_meas(sk
, seq_rtt
);
2845 static inline void tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
2848 const struct tcp_sock
*tp
= tcp_sk(sk
);
2849 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2850 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
2851 tcp_ack_saw_tstamp(sk
, flag
);
2852 else if (seq_rtt
>= 0)
2853 tcp_ack_no_tstamp(sk
, seq_rtt
, flag
);
2856 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
, u32 in_flight
)
2858 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2859 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, in_flight
);
2860 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_time_stamp
;
2863 /* Restart timer after forward progress on connection.
2864 * RFC2988 recommends to restart timer to now+rto.
2866 void tcp_rearm_rto(struct sock
*sk
)
2868 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2869 struct tcp_sock
*tp
= tcp_sk(sk
);
2871 /* If the retrans timer is currently being used by Fast Open
2872 * for SYN-ACK retrans purpose, stay put.
2874 if (tp
->fastopen_rsk
)
2877 if (!tp
->packets_out
) {
2878 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
2880 u32 rto
= inet_csk(sk
)->icsk_rto
;
2881 /* Offset the time elapsed after installing regular RTO */
2882 if (icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
||
2883 icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
) {
2884 struct sk_buff
*skb
= tcp_write_queue_head(sk
);
2885 const u32 rto_time_stamp
= TCP_SKB_CB(skb
)->when
+ rto
;
2886 s32 delta
= (s32
)(rto_time_stamp
- tcp_time_stamp
);
2887 /* delta may not be positive if the socket is locked
2888 * when the retrans timer fires and is rescheduled.
2893 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
, rto
,
2898 /* This function is called when the delayed ER timer fires. TCP enters
2899 * fast recovery and performs fast-retransmit.
2901 void tcp_resume_early_retransmit(struct sock
*sk
)
2903 struct tcp_sock
*tp
= tcp_sk(sk
);
2907 /* Stop if ER is disabled after the delayed ER timer is scheduled */
2908 if (!tp
->do_early_retrans
)
2911 tcp_enter_recovery(sk
, false);
2912 tcp_update_scoreboard(sk
, 1);
2913 tcp_xmit_retransmit_queue(sk
);
2916 /* If we get here, the whole TSO packet has not been acked. */
2917 static u32
tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
)
2919 struct tcp_sock
*tp
= tcp_sk(sk
);
2922 BUG_ON(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
));
2924 packets_acked
= tcp_skb_pcount(skb
);
2925 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
2927 packets_acked
-= tcp_skb_pcount(skb
);
2929 if (packets_acked
) {
2930 BUG_ON(tcp_skb_pcount(skb
) == 0);
2931 BUG_ON(!before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
));
2934 return packets_acked
;
2937 /* Remove acknowledged frames from the retransmission queue. If our packet
2938 * is before the ack sequence we can discard it as it's confirmed to have
2939 * arrived at the other end.
2941 static int tcp_clean_rtx_queue(struct sock
*sk
, int prior_fackets
,
2944 struct tcp_sock
*tp
= tcp_sk(sk
);
2945 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2946 struct sk_buff
*skb
;
2947 u32 now
= tcp_time_stamp
;
2948 int fully_acked
= true;
2951 u32 reord
= tp
->packets_out
;
2952 u32 prior_sacked
= tp
->sacked_out
;
2954 s32 ca_seq_rtt
= -1;
2955 ktime_t last_ackt
= net_invalid_timestamp();
2957 while ((skb
= tcp_write_queue_head(sk
)) && skb
!= tcp_send_head(sk
)) {
2958 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
2960 u8 sacked
= scb
->sacked
;
2962 /* Determine how many packets and what bytes were acked, tso and else */
2963 if (after(scb
->end_seq
, tp
->snd_una
)) {
2964 if (tcp_skb_pcount(skb
) == 1 ||
2965 !after(tp
->snd_una
, scb
->seq
))
2968 acked_pcount
= tcp_tso_acked(sk
, skb
);
2972 fully_acked
= false;
2974 acked_pcount
= tcp_skb_pcount(skb
);
2977 if (sacked
& TCPCB_RETRANS
) {
2978 if (sacked
& TCPCB_SACKED_RETRANS
)
2979 tp
->retrans_out
-= acked_pcount
;
2980 flag
|= FLAG_RETRANS_DATA_ACKED
;
2984 ca_seq_rtt
= now
- scb
->when
;
2985 last_ackt
= skb
->tstamp
;
2987 seq_rtt
= ca_seq_rtt
;
2989 if (!(sacked
& TCPCB_SACKED_ACKED
))
2990 reord
= min(pkts_acked
, reord
);
2991 if (!after(scb
->end_seq
, tp
->high_seq
))
2992 flag
|= FLAG_ORIG_SACK_ACKED
;
2995 if (sacked
& TCPCB_SACKED_ACKED
)
2996 tp
->sacked_out
-= acked_pcount
;
2997 if (sacked
& TCPCB_LOST
)
2998 tp
->lost_out
-= acked_pcount
;
3000 tp
->packets_out
-= acked_pcount
;
3001 pkts_acked
+= acked_pcount
;
3003 /* Initial outgoing SYN's get put onto the write_queue
3004 * just like anything else we transmit. It is not
3005 * true data, and if we misinform our callers that
3006 * this ACK acks real data, we will erroneously exit
3007 * connection startup slow start one packet too
3008 * quickly. This is severely frowned upon behavior.
3010 if (!(scb
->tcp_flags
& TCPHDR_SYN
)) {
3011 flag
|= FLAG_DATA_ACKED
;
3013 flag
|= FLAG_SYN_ACKED
;
3014 tp
->retrans_stamp
= 0;
3020 tcp_unlink_write_queue(skb
, sk
);
3021 sk_wmem_free_skb(sk
, skb
);
3022 if (skb
== tp
->retransmit_skb_hint
)
3023 tp
->retransmit_skb_hint
= NULL
;
3024 if (skb
== tp
->lost_skb_hint
)
3025 tp
->lost_skb_hint
= NULL
;
3028 if (likely(between(tp
->snd_up
, prior_snd_una
, tp
->snd_una
)))
3029 tp
->snd_up
= tp
->snd_una
;
3031 if (skb
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
3032 flag
|= FLAG_SACK_RENEGING
;
3034 if (flag
& FLAG_ACKED
) {
3035 const struct tcp_congestion_ops
*ca_ops
3036 = inet_csk(sk
)->icsk_ca_ops
;
3038 if (unlikely(icsk
->icsk_mtup
.probe_size
&&
3039 !after(tp
->mtu_probe
.probe_seq_end
, tp
->snd_una
))) {
3040 tcp_mtup_probe_success(sk
);
3043 tcp_ack_update_rtt(sk
, flag
, seq_rtt
);
3046 if (tcp_is_reno(tp
)) {
3047 tcp_remove_reno_sacks(sk
, pkts_acked
);
3051 /* Non-retransmitted hole got filled? That's reordering */
3052 if (reord
< prior_fackets
)
3053 tcp_update_reordering(sk
, tp
->fackets_out
- reord
, 0);
3055 delta
= tcp_is_fack(tp
) ? pkts_acked
:
3056 prior_sacked
- tp
->sacked_out
;
3057 tp
->lost_cnt_hint
-= min(tp
->lost_cnt_hint
, delta
);
3060 tp
->fackets_out
-= min(pkts_acked
, tp
->fackets_out
);
3062 if (ca_ops
->pkts_acked
) {
3065 /* Is the ACK triggering packet unambiguous? */
3066 if (!(flag
& FLAG_RETRANS_DATA_ACKED
)) {
3067 /* High resolution needed and available? */
3068 if (ca_ops
->flags
& TCP_CONG_RTT_STAMP
&&
3069 !ktime_equal(last_ackt
,
3070 net_invalid_timestamp()))
3071 rtt_us
= ktime_us_delta(ktime_get_real(),
3073 else if (ca_seq_rtt
>= 0)
3074 rtt_us
= jiffies_to_usecs(ca_seq_rtt
);
3077 ca_ops
->pkts_acked(sk
, pkts_acked
, rtt_us
);
3081 #if FASTRETRANS_DEBUG > 0
3082 WARN_ON((int)tp
->sacked_out
< 0);
3083 WARN_ON((int)tp
->lost_out
< 0);
3084 WARN_ON((int)tp
->retrans_out
< 0);
3085 if (!tp
->packets_out
&& tcp_is_sack(tp
)) {
3086 icsk
= inet_csk(sk
);
3088 pr_debug("Leak l=%u %d\n",
3089 tp
->lost_out
, icsk
->icsk_ca_state
);
3092 if (tp
->sacked_out
) {
3093 pr_debug("Leak s=%u %d\n",
3094 tp
->sacked_out
, icsk
->icsk_ca_state
);
3097 if (tp
->retrans_out
) {
3098 pr_debug("Leak r=%u %d\n",
3099 tp
->retrans_out
, icsk
->icsk_ca_state
);
3100 tp
->retrans_out
= 0;
3107 static void tcp_ack_probe(struct sock
*sk
)
3109 const struct tcp_sock
*tp
= tcp_sk(sk
);
3110 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3112 /* Was it a usable window open? */
3114 if (!after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
, tcp_wnd_end(tp
))) {
3115 icsk
->icsk_backoff
= 0;
3116 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
3117 /* Socket must be waked up by subsequent tcp_data_snd_check().
3118 * This function is not for random using!
3121 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3122 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
3127 static inline bool tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
3129 return !(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
3130 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
;
3133 static inline bool tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
3135 const struct tcp_sock
*tp
= tcp_sk(sk
);
3136 return (!(flag
& FLAG_ECE
) || tp
->snd_cwnd
< tp
->snd_ssthresh
) &&
3137 !tcp_in_cwnd_reduction(sk
);
3140 /* Check that window update is acceptable.
3141 * The function assumes that snd_una<=ack<=snd_next.
3143 static inline bool tcp_may_update_window(const struct tcp_sock
*tp
,
3144 const u32 ack
, const u32 ack_seq
,
3147 return after(ack
, tp
->snd_una
) ||
3148 after(ack_seq
, tp
->snd_wl1
) ||
3149 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
);
3152 /* Update our send window.
3154 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3155 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3157 static int tcp_ack_update_window(struct sock
*sk
, const struct sk_buff
*skb
, u32 ack
,
3160 struct tcp_sock
*tp
= tcp_sk(sk
);
3162 u32 nwin
= ntohs(tcp_hdr(skb
)->window
);
3164 if (likely(!tcp_hdr(skb
)->syn
))
3165 nwin
<<= tp
->rx_opt
.snd_wscale
;
3167 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
3168 flag
|= FLAG_WIN_UPDATE
;
3169 tcp_update_wl(tp
, ack_seq
);
3171 if (tp
->snd_wnd
!= nwin
) {
3174 /* Note, it is the only place, where
3175 * fast path is recovered for sending TCP.
3178 tcp_fast_path_check(sk
);
3180 if (nwin
> tp
->max_window
) {
3181 tp
->max_window
= nwin
;
3182 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
3192 /* RFC 5961 7 [ACK Throttling] */
3193 static void tcp_send_challenge_ack(struct sock
*sk
)
3195 /* unprotected vars, we dont care of overwrites */
3196 static u32 challenge_timestamp
;
3197 static unsigned int challenge_count
;
3198 u32 now
= jiffies
/ HZ
;
3200 if (now
!= challenge_timestamp
) {
3201 challenge_timestamp
= now
;
3202 challenge_count
= 0;
3204 if (++challenge_count
<= sysctl_tcp_challenge_ack_limit
) {
3205 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPCHALLENGEACK
);
3210 static void tcp_store_ts_recent(struct tcp_sock
*tp
)
3212 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
3213 tp
->rx_opt
.ts_recent_stamp
= get_seconds();
3216 static void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
3218 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
3219 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3220 * extra check below makes sure this can only happen
3221 * for pure ACK frames. -DaveM
3223 * Not only, also it occurs for expired timestamps.
3226 if (tcp_paws_check(&tp
->rx_opt
, 0))
3227 tcp_store_ts_recent(tp
);
3231 /* This routine deals with acks during a TLP episode.
3232 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3234 static void tcp_process_tlp_ack(struct sock
*sk
, u32 ack
, int flag
)
3236 struct tcp_sock
*tp
= tcp_sk(sk
);
3237 bool is_tlp_dupack
= (ack
== tp
->tlp_high_seq
) &&
3238 !(flag
& (FLAG_SND_UNA_ADVANCED
|
3239 FLAG_NOT_DUP
| FLAG_DATA_SACKED
));
3241 /* Mark the end of TLP episode on receiving TLP dupack or when
3242 * ack is after tlp_high_seq.
3244 if (is_tlp_dupack
) {
3245 tp
->tlp_high_seq
= 0;
3249 if (after(ack
, tp
->tlp_high_seq
)) {
3250 tp
->tlp_high_seq
= 0;
3251 /* Don't reduce cwnd if DSACK arrives for TLP retrans. */
3252 if (!(flag
& FLAG_DSACKING_ACK
)) {
3253 tcp_init_cwnd_reduction(sk
, true);
3254 tcp_set_ca_state(sk
, TCP_CA_CWR
);
3255 tcp_end_cwnd_reduction(sk
);
3256 tcp_set_ca_state(sk
, TCP_CA_Open
);
3257 NET_INC_STATS_BH(sock_net(sk
),
3258 LINUX_MIB_TCPLOSSPROBERECOVERY
);
3263 /* This routine deals with incoming acks, but not outgoing ones. */
3264 static int tcp_ack(struct sock
*sk
, const struct sk_buff
*skb
, int flag
)
3266 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3267 struct tcp_sock
*tp
= tcp_sk(sk
);
3268 u32 prior_snd_una
= tp
->snd_una
;
3269 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
3270 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3271 bool is_dupack
= false;
3272 u32 prior_in_flight
;
3274 int prior_packets
= tp
->packets_out
;
3275 const int prior_unsacked
= tp
->packets_out
- tp
->sacked_out
;
3276 int acked
= 0; /* Number of packets newly acked */
3278 /* If the ack is older than previous acks
3279 * then we can probably ignore it.
3281 if (before(ack
, prior_snd_una
)) {
3282 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3283 if (before(ack
, prior_snd_una
- tp
->max_window
)) {
3284 tcp_send_challenge_ack(sk
);
3290 /* If the ack includes data we haven't sent yet, discard
3291 * this segment (RFC793 Section 3.9).
3293 if (after(ack
, tp
->snd_nxt
))
3296 if (icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
||
3297 icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
)
3300 if (after(ack
, prior_snd_una
))
3301 flag
|= FLAG_SND_UNA_ADVANCED
;
3303 prior_fackets
= tp
->fackets_out
;
3304 prior_in_flight
= tcp_packets_in_flight(tp
);
3306 /* ts_recent update must be made after we are sure that the packet
3309 if (flag
& FLAG_UPDATE_TS_RECENT
)
3310 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
3312 if (!(flag
& FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
3313 /* Window is constant, pure forward advance.
3314 * No more checks are required.
3315 * Note, we use the fact that SND.UNA>=SND.WL2.
3317 tcp_update_wl(tp
, ack_seq
);
3319 flag
|= FLAG_WIN_UPDATE
;
3321 tcp_ca_event(sk
, CA_EVENT_FAST_ACK
);
3323 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPACKS
);
3325 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
3328 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPUREACKS
);
3330 flag
|= tcp_ack_update_window(sk
, skb
, ack
, ack_seq
);
3332 if (TCP_SKB_CB(skb
)->sacked
)
3333 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
3335 if (TCP_ECN_rcv_ecn_echo(tp
, tcp_hdr(skb
)))
3338 tcp_ca_event(sk
, CA_EVENT_SLOW_ACK
);
3341 /* We passed data and got it acked, remove any soft error
3342 * log. Something worked...
3344 sk
->sk_err_soft
= 0;
3345 icsk
->icsk_probes_out
= 0;
3346 tp
->rcv_tstamp
= tcp_time_stamp
;
3350 /* See if we can take anything off of the retransmit queue. */
3351 acked
= tp
->packets_out
;
3352 flag
|= tcp_clean_rtx_queue(sk
, prior_fackets
, prior_snd_una
);
3353 acked
-= tp
->packets_out
;
3355 if (tcp_ack_is_dubious(sk
, flag
)) {
3356 /* Advance CWND, if state allows this. */
3357 if ((flag
& FLAG_DATA_ACKED
) && tcp_may_raise_cwnd(sk
, flag
))
3358 tcp_cong_avoid(sk
, ack
, prior_in_flight
);
3359 is_dupack
= !(flag
& (FLAG_SND_UNA_ADVANCED
| FLAG_NOT_DUP
));
3360 tcp_fastretrans_alert(sk
, acked
, prior_unsacked
,
3363 if (flag
& FLAG_DATA_ACKED
)
3364 tcp_cong_avoid(sk
, ack
, prior_in_flight
);
3367 if (tp
->tlp_high_seq
)
3368 tcp_process_tlp_ack(sk
, ack
, flag
);
3370 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
& FLAG_NOT_DUP
)) {
3371 struct dst_entry
*dst
= __sk_dst_get(sk
);
3376 if (icsk
->icsk_pending
== ICSK_TIME_RETRANS
)
3377 tcp_schedule_loss_probe(sk
);
3381 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3382 if (flag
& FLAG_DSACKING_ACK
)
3383 tcp_fastretrans_alert(sk
, acked
, prior_unsacked
,
3385 /* If this ack opens up a zero window, clear backoff. It was
3386 * being used to time the probes, and is probably far higher than
3387 * it needs to be for normal retransmission.
3389 if (tcp_send_head(sk
))
3392 if (tp
->tlp_high_seq
)
3393 tcp_process_tlp_ack(sk
, ack
, flag
);
3397 SOCK_DEBUG(sk
, "Ack %u after %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3401 /* If data was SACKed, tag it and see if we should send more data.
3402 * If data was DSACKed, see if we can undo a cwnd reduction.
3404 if (TCP_SKB_CB(skb
)->sacked
) {
3405 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
3406 tcp_fastretrans_alert(sk
, acked
, prior_unsacked
,
3410 SOCK_DEBUG(sk
, "Ack %u before %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3414 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3415 * But, this can also be called on packets in the established flow when
3416 * the fast version below fails.
3418 void tcp_parse_options(const struct sk_buff
*skb
,
3419 struct tcp_options_received
*opt_rx
, int estab
,
3420 struct tcp_fastopen_cookie
*foc
)
3422 const unsigned char *ptr
;
3423 const struct tcphdr
*th
= tcp_hdr(skb
);
3424 int length
= (th
->doff
* 4) - sizeof(struct tcphdr
);
3426 ptr
= (const unsigned char *)(th
+ 1);
3427 opt_rx
->saw_tstamp
= 0;
3429 while (length
> 0) {
3430 int opcode
= *ptr
++;
3436 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
3441 if (opsize
< 2) /* "silly options" */
3443 if (opsize
> length
)
3444 return; /* don't parse partial options */
3447 if (opsize
== TCPOLEN_MSS
&& th
->syn
&& !estab
) {
3448 u16 in_mss
= get_unaligned_be16(ptr
);
3450 if (opt_rx
->user_mss
&&
3451 opt_rx
->user_mss
< in_mss
)
3452 in_mss
= opt_rx
->user_mss
;
3453 opt_rx
->mss_clamp
= in_mss
;
3458 if (opsize
== TCPOLEN_WINDOW
&& th
->syn
&&
3459 !estab
&& sysctl_tcp_window_scaling
) {
3460 __u8 snd_wscale
= *(__u8
*)ptr
;
3461 opt_rx
->wscale_ok
= 1;
3462 if (snd_wscale
> 14) {
3463 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3468 opt_rx
->snd_wscale
= snd_wscale
;
3471 case TCPOPT_TIMESTAMP
:
3472 if ((opsize
== TCPOLEN_TIMESTAMP
) &&
3473 ((estab
&& opt_rx
->tstamp_ok
) ||
3474 (!estab
&& sysctl_tcp_timestamps
))) {
3475 opt_rx
->saw_tstamp
= 1;
3476 opt_rx
->rcv_tsval
= get_unaligned_be32(ptr
);
3477 opt_rx
->rcv_tsecr
= get_unaligned_be32(ptr
+ 4);
3480 case TCPOPT_SACK_PERM
:
3481 if (opsize
== TCPOLEN_SACK_PERM
&& th
->syn
&&
3482 !estab
&& sysctl_tcp_sack
) {
3483 opt_rx
->sack_ok
= TCP_SACK_SEEN
;
3484 tcp_sack_reset(opt_rx
);
3489 if ((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
3490 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
3492 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
3495 #ifdef CONFIG_TCP_MD5SIG
3498 * The MD5 Hash has already been
3499 * checked (see tcp_v{4,6}_do_rcv()).
3504 /* Fast Open option shares code 254 using a
3505 * 16 bits magic number. It's valid only in
3506 * SYN or SYN-ACK with an even size.
3508 if (opsize
< TCPOLEN_EXP_FASTOPEN_BASE
||
3509 get_unaligned_be16(ptr
) != TCPOPT_FASTOPEN_MAGIC
||
3510 foc
== NULL
|| !th
->syn
|| (opsize
& 1))
3512 foc
->len
= opsize
- TCPOLEN_EXP_FASTOPEN_BASE
;
3513 if (foc
->len
>= TCP_FASTOPEN_COOKIE_MIN
&&
3514 foc
->len
<= TCP_FASTOPEN_COOKIE_MAX
)
3515 memcpy(foc
->val
, ptr
+ 2, foc
->len
);
3516 else if (foc
->len
!= 0)
3526 EXPORT_SYMBOL(tcp_parse_options
);
3528 static bool tcp_parse_aligned_timestamp(struct tcp_sock
*tp
, const struct tcphdr
*th
)
3530 const __be32
*ptr
= (const __be32
*)(th
+ 1);
3532 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3533 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
3534 tp
->rx_opt
.saw_tstamp
= 1;
3536 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3538 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
) - tp
->tsoffset
;
3544 /* Fast parse options. This hopes to only see timestamps.
3545 * If it is wrong it falls back on tcp_parse_options().
3547 static bool tcp_fast_parse_options(const struct sk_buff
*skb
,
3548 const struct tcphdr
*th
, struct tcp_sock
*tp
)
3550 /* In the spirit of fast parsing, compare doff directly to constant
3551 * values. Because equality is used, short doff can be ignored here.
3553 if (th
->doff
== (sizeof(*th
) / 4)) {
3554 tp
->rx_opt
.saw_tstamp
= 0;
3556 } else if (tp
->rx_opt
.tstamp_ok
&&
3557 th
->doff
== ((sizeof(*th
) + TCPOLEN_TSTAMP_ALIGNED
) / 4)) {
3558 if (tcp_parse_aligned_timestamp(tp
, th
))
3562 tcp_parse_options(skb
, &tp
->rx_opt
, 1, NULL
);
3563 if (tp
->rx_opt
.saw_tstamp
)
3564 tp
->rx_opt
.rcv_tsecr
-= tp
->tsoffset
;
3569 #ifdef CONFIG_TCP_MD5SIG
3571 * Parse MD5 Signature option
3573 const u8
*tcp_parse_md5sig_option(const struct tcphdr
*th
)
3575 int length
= (th
->doff
<< 2) - sizeof(*th
);
3576 const u8
*ptr
= (const u8
*)(th
+ 1);
3578 /* If the TCP option is too short, we can short cut */
3579 if (length
< TCPOLEN_MD5SIG
)
3582 while (length
> 0) {
3583 int opcode
= *ptr
++;
3594 if (opsize
< 2 || opsize
> length
)
3596 if (opcode
== TCPOPT_MD5SIG
)
3597 return opsize
== TCPOLEN_MD5SIG
? ptr
: NULL
;
3604 EXPORT_SYMBOL(tcp_parse_md5sig_option
);
3607 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3609 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3610 * it can pass through stack. So, the following predicate verifies that
3611 * this segment is not used for anything but congestion avoidance or
3612 * fast retransmit. Moreover, we even are able to eliminate most of such
3613 * second order effects, if we apply some small "replay" window (~RTO)
3614 * to timestamp space.
3616 * All these measures still do not guarantee that we reject wrapped ACKs
3617 * on networks with high bandwidth, when sequence space is recycled fastly,
3618 * but it guarantees that such events will be very rare and do not affect
3619 * connection seriously. This doesn't look nice, but alas, PAWS is really
3622 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3623 * states that events when retransmit arrives after original data are rare.
3624 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3625 * the biggest problem on large power networks even with minor reordering.
3626 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3627 * up to bandwidth of 18Gigabit/sec. 8) ]
3630 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
3632 const struct tcp_sock
*tp
= tcp_sk(sk
);
3633 const struct tcphdr
*th
= tcp_hdr(skb
);
3634 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3635 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3637 return (/* 1. Pure ACK with correct sequence number. */
3638 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
3640 /* 2. ... and duplicate ACK. */
3641 ack
== tp
->snd_una
&&
3643 /* 3. ... and does not update window. */
3644 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
3646 /* 4. ... and sits in replay window. */
3647 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
3650 static inline bool tcp_paws_discard(const struct sock
*sk
,
3651 const struct sk_buff
*skb
)
3653 const struct tcp_sock
*tp
= tcp_sk(sk
);
3655 return !tcp_paws_check(&tp
->rx_opt
, TCP_PAWS_WINDOW
) &&
3656 !tcp_disordered_ack(sk
, skb
);
3659 /* Check segment sequence number for validity.
3661 * Segment controls are considered valid, if the segment
3662 * fits to the window after truncation to the window. Acceptability
3663 * of data (and SYN, FIN, of course) is checked separately.
3664 * See tcp_data_queue(), for example.
3666 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3667 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3668 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3669 * (borrowed from freebsd)
3672 static inline bool tcp_sequence(const struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3674 return !before(end_seq
, tp
->rcv_wup
) &&
3675 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
3678 /* When we get a reset we do this. */
3679 void tcp_reset(struct sock
*sk
)
3681 /* We want the right error as BSD sees it (and indeed as we do). */
3682 switch (sk
->sk_state
) {
3684 sk
->sk_err
= ECONNREFUSED
;
3686 case TCP_CLOSE_WAIT
:
3692 sk
->sk_err
= ECONNRESET
;
3694 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3697 if (!sock_flag(sk
, SOCK_DEAD
))
3698 sk
->sk_error_report(sk
);
3704 * Process the FIN bit. This now behaves as it is supposed to work
3705 * and the FIN takes effect when it is validly part of sequence
3706 * space. Not before when we get holes.
3708 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3709 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3712 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3713 * close and we go into CLOSING (and later onto TIME-WAIT)
3715 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3717 static void tcp_fin(struct sock
*sk
)
3719 struct tcp_sock
*tp
= tcp_sk(sk
);
3720 const struct dst_entry
*dst
;
3722 inet_csk_schedule_ack(sk
);
3724 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
3725 sock_set_flag(sk
, SOCK_DONE
);
3727 switch (sk
->sk_state
) {
3729 case TCP_ESTABLISHED
:
3730 /* Move to CLOSE_WAIT */
3731 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
3732 dst
= __sk_dst_get(sk
);
3733 if (!dst
|| !dst_metric(dst
, RTAX_QUICKACK
))
3734 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
3737 case TCP_CLOSE_WAIT
:
3739 /* Received a retransmission of the FIN, do
3744 /* RFC793: Remain in the LAST-ACK state. */
3748 /* This case occurs when a simultaneous close
3749 * happens, we must ack the received FIN and
3750 * enter the CLOSING state.
3753 tcp_set_state(sk
, TCP_CLOSING
);
3756 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3758 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
3761 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3762 * cases we should never reach this piece of code.
3764 pr_err("%s: Impossible, sk->sk_state=%d\n",
3765 __func__
, sk
->sk_state
);
3769 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3770 * Probably, we should reset in this case. For now drop them.
3772 __skb_queue_purge(&tp
->out_of_order_queue
);
3773 if (tcp_is_sack(tp
))
3774 tcp_sack_reset(&tp
->rx_opt
);
3777 if (!sock_flag(sk
, SOCK_DEAD
)) {
3778 sk
->sk_state_change(sk
);
3780 /* Do not send POLL_HUP for half duplex close. */
3781 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
3782 sk
->sk_state
== TCP_CLOSE
)
3783 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_HUP
);
3785 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_IN
);
3789 static inline bool tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
,
3792 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
3793 if (before(seq
, sp
->start_seq
))
3794 sp
->start_seq
= seq
;
3795 if (after(end_seq
, sp
->end_seq
))
3796 sp
->end_seq
= end_seq
;
3802 static void tcp_dsack_set(struct sock
*sk
, u32 seq
, u32 end_seq
)
3804 struct tcp_sock
*tp
= tcp_sk(sk
);
3806 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
3809 if (before(seq
, tp
->rcv_nxt
))
3810 mib_idx
= LINUX_MIB_TCPDSACKOLDSENT
;
3812 mib_idx
= LINUX_MIB_TCPDSACKOFOSENT
;
3814 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
3816 tp
->rx_opt
.dsack
= 1;
3817 tp
->duplicate_sack
[0].start_seq
= seq
;
3818 tp
->duplicate_sack
[0].end_seq
= end_seq
;
3822 static void tcp_dsack_extend(struct sock
*sk
, u32 seq
, u32 end_seq
)
3824 struct tcp_sock
*tp
= tcp_sk(sk
);
3826 if (!tp
->rx_opt
.dsack
)
3827 tcp_dsack_set(sk
, seq
, end_seq
);
3829 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
3832 static void tcp_send_dupack(struct sock
*sk
, const struct sk_buff
*skb
)
3834 struct tcp_sock
*tp
= tcp_sk(sk
);
3836 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
3837 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
3838 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
3839 tcp_enter_quickack_mode(sk
);
3841 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
3842 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
3844 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
3845 end_seq
= tp
->rcv_nxt
;
3846 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, end_seq
);
3853 /* These routines update the SACK block as out-of-order packets arrive or
3854 * in-order packets close up the sequence space.
3856 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
3859 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3860 struct tcp_sack_block
*swalk
= sp
+ 1;
3862 /* See if the recent change to the first SACK eats into
3863 * or hits the sequence space of other SACK blocks, if so coalesce.
3865 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
;) {
3866 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
3869 /* Zap SWALK, by moving every further SACK up by one slot.
3870 * Decrease num_sacks.
3872 tp
->rx_opt
.num_sacks
--;
3873 for (i
= this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
3877 this_sack
++, swalk
++;
3881 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
3883 struct tcp_sock
*tp
= tcp_sk(sk
);
3884 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3885 int cur_sacks
= tp
->rx_opt
.num_sacks
;
3891 for (this_sack
= 0; this_sack
< cur_sacks
; this_sack
++, sp
++) {
3892 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
3893 /* Rotate this_sack to the first one. */
3894 for (; this_sack
> 0; this_sack
--, sp
--)
3895 swap(*sp
, *(sp
- 1));
3897 tcp_sack_maybe_coalesce(tp
);
3902 /* Could not find an adjacent existing SACK, build a new one,
3903 * put it at the front, and shift everyone else down. We
3904 * always know there is at least one SACK present already here.
3906 * If the sack array is full, forget about the last one.
3908 if (this_sack
>= TCP_NUM_SACKS
) {
3910 tp
->rx_opt
.num_sacks
--;
3913 for (; this_sack
> 0; this_sack
--, sp
--)
3917 /* Build the new head SACK, and we're done. */
3918 sp
->start_seq
= seq
;
3919 sp
->end_seq
= end_seq
;
3920 tp
->rx_opt
.num_sacks
++;
3923 /* RCV.NXT advances, some SACKs should be eaten. */
3925 static void tcp_sack_remove(struct tcp_sock
*tp
)
3927 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3928 int num_sacks
= tp
->rx_opt
.num_sacks
;
3931 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3932 if (skb_queue_empty(&tp
->out_of_order_queue
)) {
3933 tp
->rx_opt
.num_sacks
= 0;
3937 for (this_sack
= 0; this_sack
< num_sacks
;) {
3938 /* Check if the start of the sack is covered by RCV.NXT. */
3939 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
3942 /* RCV.NXT must cover all the block! */
3943 WARN_ON(before(tp
->rcv_nxt
, sp
->end_seq
));
3945 /* Zap this SACK, by moving forward any other SACKS. */
3946 for (i
=this_sack
+1; i
< num_sacks
; i
++)
3947 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
3954 tp
->rx_opt
.num_sacks
= num_sacks
;
3957 /* This one checks to see if we can put data from the
3958 * out_of_order queue into the receive_queue.
3960 static void tcp_ofo_queue(struct sock
*sk
)
3962 struct tcp_sock
*tp
= tcp_sk(sk
);
3963 __u32 dsack_high
= tp
->rcv_nxt
;
3964 struct sk_buff
*skb
;
3966 while ((skb
= skb_peek(&tp
->out_of_order_queue
)) != NULL
) {
3967 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
3970 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
3971 __u32 dsack
= dsack_high
;
3972 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
3973 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
3974 tcp_dsack_extend(sk
, TCP_SKB_CB(skb
)->seq
, dsack
);
3977 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
3978 SOCK_DEBUG(sk
, "ofo packet was already received\n");
3979 __skb_unlink(skb
, &tp
->out_of_order_queue
);
3983 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
3984 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
3985 TCP_SKB_CB(skb
)->end_seq
);
3987 __skb_unlink(skb
, &tp
->out_of_order_queue
);
3988 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
3989 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
3990 if (tcp_hdr(skb
)->fin
)
3995 static bool tcp_prune_ofo_queue(struct sock
*sk
);
3996 static int tcp_prune_queue(struct sock
*sk
);
3998 static int tcp_try_rmem_schedule(struct sock
*sk
, struct sk_buff
*skb
,
4001 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
4002 !sk_rmem_schedule(sk
, skb
, size
)) {
4004 if (tcp_prune_queue(sk
) < 0)
4007 if (!sk_rmem_schedule(sk
, skb
, size
)) {
4008 if (!tcp_prune_ofo_queue(sk
))
4011 if (!sk_rmem_schedule(sk
, skb
, size
))
4019 * tcp_try_coalesce - try to merge skb to prior one
4022 * @from: buffer to add in queue
4023 * @fragstolen: pointer to boolean
4025 * Before queueing skb @from after @to, try to merge them
4026 * to reduce overall memory use and queue lengths, if cost is small.
4027 * Packets in ofo or receive queues can stay a long time.
4028 * Better try to coalesce them right now to avoid future collapses.
4029 * Returns true if caller should free @from instead of queueing it
4031 static bool tcp_try_coalesce(struct sock
*sk
,
4033 struct sk_buff
*from
,
4038 *fragstolen
= false;
4040 if (tcp_hdr(from
)->fin
)
4043 /* Its possible this segment overlaps with prior segment in queue */
4044 if (TCP_SKB_CB(from
)->seq
!= TCP_SKB_CB(to
)->end_seq
)
4047 if (!skb_try_coalesce(to
, from
, fragstolen
, &delta
))
4050 atomic_add(delta
, &sk
->sk_rmem_alloc
);
4051 sk_mem_charge(sk
, delta
);
4052 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRCVCOALESCE
);
4053 TCP_SKB_CB(to
)->end_seq
= TCP_SKB_CB(from
)->end_seq
;
4054 TCP_SKB_CB(to
)->ack_seq
= TCP_SKB_CB(from
)->ack_seq
;
4058 static void tcp_data_queue_ofo(struct sock
*sk
, struct sk_buff
*skb
)
4060 struct tcp_sock
*tp
= tcp_sk(sk
);
4061 struct sk_buff
*skb1
;
4064 TCP_ECN_check_ce(tp
, skb
);
4066 if (unlikely(tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))) {
4067 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFODROP
);
4072 /* Disable header prediction. */
4074 inet_csk_schedule_ack(sk
);
4076 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFOQUEUE
);
4077 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
4078 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4080 skb1
= skb_peek_tail(&tp
->out_of_order_queue
);
4082 /* Initial out of order segment, build 1 SACK. */
4083 if (tcp_is_sack(tp
)) {
4084 tp
->rx_opt
.num_sacks
= 1;
4085 tp
->selective_acks
[0].start_seq
= TCP_SKB_CB(skb
)->seq
;
4086 tp
->selective_acks
[0].end_seq
=
4087 TCP_SKB_CB(skb
)->end_seq
;
4089 __skb_queue_head(&tp
->out_of_order_queue
, skb
);
4093 seq
= TCP_SKB_CB(skb
)->seq
;
4094 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4096 if (seq
== TCP_SKB_CB(skb1
)->end_seq
) {
4099 if (!tcp_try_coalesce(sk
, skb1
, skb
, &fragstolen
)) {
4100 __skb_queue_after(&tp
->out_of_order_queue
, skb1
, skb
);
4102 kfree_skb_partial(skb
, fragstolen
);
4106 if (!tp
->rx_opt
.num_sacks
||
4107 tp
->selective_acks
[0].end_seq
!= seq
)
4110 /* Common case: data arrive in order after hole. */
4111 tp
->selective_acks
[0].end_seq
= end_seq
;
4115 /* Find place to insert this segment. */
4117 if (!after(TCP_SKB_CB(skb1
)->seq
, seq
))
4119 if (skb_queue_is_first(&tp
->out_of_order_queue
, skb1
)) {
4123 skb1
= skb_queue_prev(&tp
->out_of_order_queue
, skb1
);
4126 /* Do skb overlap to previous one? */
4127 if (skb1
&& before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4128 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4129 /* All the bits are present. Drop. */
4130 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFOMERGE
);
4133 tcp_dsack_set(sk
, seq
, end_seq
);
4136 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
4137 /* Partial overlap. */
4138 tcp_dsack_set(sk
, seq
,
4139 TCP_SKB_CB(skb1
)->end_seq
);
4141 if (skb_queue_is_first(&tp
->out_of_order_queue
,
4145 skb1
= skb_queue_prev(
4146 &tp
->out_of_order_queue
,
4151 __skb_queue_head(&tp
->out_of_order_queue
, skb
);
4153 __skb_queue_after(&tp
->out_of_order_queue
, skb1
, skb
);
4155 /* And clean segments covered by new one as whole. */
4156 while (!skb_queue_is_last(&tp
->out_of_order_queue
, skb
)) {
4157 skb1
= skb_queue_next(&tp
->out_of_order_queue
, skb
);
4159 if (!after(end_seq
, TCP_SKB_CB(skb1
)->seq
))
4161 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4162 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4166 __skb_unlink(skb1
, &tp
->out_of_order_queue
);
4167 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4168 TCP_SKB_CB(skb1
)->end_seq
);
4169 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFOMERGE
);
4174 if (tcp_is_sack(tp
))
4175 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
4178 skb_set_owner_r(skb
, sk
);
4181 static int __must_check
tcp_queue_rcv(struct sock
*sk
, struct sk_buff
*skb
, int hdrlen
,
4185 struct sk_buff
*tail
= skb_peek_tail(&sk
->sk_receive_queue
);
4187 __skb_pull(skb
, hdrlen
);
4189 tcp_try_coalesce(sk
, tail
, skb
, fragstolen
)) ? 1 : 0;
4190 tcp_sk(sk
)->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4192 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4193 skb_set_owner_r(skb
, sk
);
4198 int tcp_send_rcvq(struct sock
*sk
, struct msghdr
*msg
, size_t size
)
4200 struct sk_buff
*skb
= NULL
;
4207 skb
= alloc_skb(size
+ sizeof(*th
), sk
->sk_allocation
);
4211 if (tcp_try_rmem_schedule(sk
, skb
, size
+ sizeof(*th
)))
4214 th
= (struct tcphdr
*)skb_put(skb
, sizeof(*th
));
4215 skb_reset_transport_header(skb
);
4216 memset(th
, 0, sizeof(*th
));
4218 if (memcpy_fromiovec(skb_put(skb
, size
), msg
->msg_iov
, size
))
4221 TCP_SKB_CB(skb
)->seq
= tcp_sk(sk
)->rcv_nxt
;
4222 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ size
;
4223 TCP_SKB_CB(skb
)->ack_seq
= tcp_sk(sk
)->snd_una
- 1;
4225 if (tcp_queue_rcv(sk
, skb
, sizeof(*th
), &fragstolen
)) {
4226 WARN_ON_ONCE(fragstolen
); /* should not happen */
4237 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
4239 const struct tcphdr
*th
= tcp_hdr(skb
);
4240 struct tcp_sock
*tp
= tcp_sk(sk
);
4242 bool fragstolen
= false;
4244 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
)
4248 __skb_pull(skb
, th
->doff
* 4);
4250 TCP_ECN_accept_cwr(tp
, skb
);
4252 tp
->rx_opt
.dsack
= 0;
4254 /* Queue data for delivery to the user.
4255 * Packets in sequence go to the receive queue.
4256 * Out of sequence packets to the out_of_order_queue.
4258 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
4259 if (tcp_receive_window(tp
) == 0)
4262 /* Ok. In sequence. In window. */
4263 if (tp
->ucopy
.task
== current
&&
4264 tp
->copied_seq
== tp
->rcv_nxt
&& tp
->ucopy
.len
&&
4265 sock_owned_by_user(sk
) && !tp
->urg_data
) {
4266 int chunk
= min_t(unsigned int, skb
->len
,
4269 __set_current_state(TASK_RUNNING
);
4272 if (!skb_copy_datagram_iovec(skb
, 0, tp
->ucopy
.iov
, chunk
)) {
4273 tp
->ucopy
.len
-= chunk
;
4274 tp
->copied_seq
+= chunk
;
4275 eaten
= (chunk
== skb
->len
);
4276 tcp_rcv_space_adjust(sk
);
4284 tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))
4287 eaten
= tcp_queue_rcv(sk
, skb
, 0, &fragstolen
);
4289 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4291 tcp_event_data_recv(sk
, skb
);
4295 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4298 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4299 * gap in queue is filled.
4301 if (skb_queue_empty(&tp
->out_of_order_queue
))
4302 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
4305 if (tp
->rx_opt
.num_sacks
)
4306 tcp_sack_remove(tp
);
4308 tcp_fast_path_check(sk
);
4311 kfree_skb_partial(skb
, fragstolen
);
4312 if (!sock_flag(sk
, SOCK_DEAD
))
4313 sk
->sk_data_ready(sk
, 0);
4317 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4318 /* A retransmit, 2nd most common case. Force an immediate ack. */
4319 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4320 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4323 tcp_enter_quickack_mode(sk
);
4324 inet_csk_schedule_ack(sk
);
4330 /* Out of window. F.e. zero window probe. */
4331 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
4334 tcp_enter_quickack_mode(sk
);
4336 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4337 /* Partial packet, seq < rcv_next < end_seq */
4338 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
4339 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4340 TCP_SKB_CB(skb
)->end_seq
);
4342 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
4344 /* If window is closed, drop tail of packet. But after
4345 * remembering D-SACK for its head made in previous line.
4347 if (!tcp_receive_window(tp
))
4352 tcp_data_queue_ofo(sk
, skb
);
4355 static struct sk_buff
*tcp_collapse_one(struct sock
*sk
, struct sk_buff
*skb
,
4356 struct sk_buff_head
*list
)
4358 struct sk_buff
*next
= NULL
;
4360 if (!skb_queue_is_last(list
, skb
))
4361 next
= skb_queue_next(list
, skb
);
4363 __skb_unlink(skb
, list
);
4365 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRCVCOLLAPSED
);
4370 /* Collapse contiguous sequence of skbs head..tail with
4371 * sequence numbers start..end.
4373 * If tail is NULL, this means until the end of the list.
4375 * Segments with FIN/SYN are not collapsed (only because this
4379 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
,
4380 struct sk_buff
*head
, struct sk_buff
*tail
,
4383 struct sk_buff
*skb
, *n
;
4386 /* First, check that queue is collapsible and find
4387 * the point where collapsing can be useful. */
4391 skb_queue_walk_from_safe(list
, skb
, n
) {
4394 /* No new bits? It is possible on ofo queue. */
4395 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4396 skb
= tcp_collapse_one(sk
, skb
, list
);
4402 /* The first skb to collapse is:
4404 * - bloated or contains data before "start" or
4405 * overlaps to the next one.
4407 if (!tcp_hdr(skb
)->syn
&& !tcp_hdr(skb
)->fin
&&
4408 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
4409 before(TCP_SKB_CB(skb
)->seq
, start
))) {
4410 end_of_skbs
= false;
4414 if (!skb_queue_is_last(list
, skb
)) {
4415 struct sk_buff
*next
= skb_queue_next(list
, skb
);
4417 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(next
)->seq
) {
4418 end_of_skbs
= false;
4423 /* Decided to skip this, advance start seq. */
4424 start
= TCP_SKB_CB(skb
)->end_seq
;
4426 if (end_of_skbs
|| tcp_hdr(skb
)->syn
|| tcp_hdr(skb
)->fin
)
4429 while (before(start
, end
)) {
4430 struct sk_buff
*nskb
;
4431 unsigned int header
= skb_headroom(skb
);
4432 int copy
= SKB_MAX_ORDER(header
, 0);
4434 /* Too big header? This can happen with IPv6. */
4437 if (end
- start
< copy
)
4439 nskb
= alloc_skb(copy
+ header
, GFP_ATOMIC
);
4443 skb_set_mac_header(nskb
, skb_mac_header(skb
) - skb
->head
);
4444 skb_set_network_header(nskb
, (skb_network_header(skb
) -
4446 skb_set_transport_header(nskb
, (skb_transport_header(skb
) -
4448 skb_reserve(nskb
, header
);
4449 memcpy(nskb
->head
, skb
->head
, header
);
4450 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
4451 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
4452 __skb_queue_before(list
, skb
, nskb
);
4453 skb_set_owner_r(nskb
, sk
);
4455 /* Copy data, releasing collapsed skbs. */
4457 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
4458 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
4462 size
= min(copy
, size
);
4463 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
4465 TCP_SKB_CB(nskb
)->end_seq
+= size
;
4469 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4470 skb
= tcp_collapse_one(sk
, skb
, list
);
4473 tcp_hdr(skb
)->syn
||
4481 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4482 * and tcp_collapse() them until all the queue is collapsed.
4484 static void tcp_collapse_ofo_queue(struct sock
*sk
)
4486 struct tcp_sock
*tp
= tcp_sk(sk
);
4487 struct sk_buff
*skb
= skb_peek(&tp
->out_of_order_queue
);
4488 struct sk_buff
*head
;
4494 start
= TCP_SKB_CB(skb
)->seq
;
4495 end
= TCP_SKB_CB(skb
)->end_seq
;
4499 struct sk_buff
*next
= NULL
;
4501 if (!skb_queue_is_last(&tp
->out_of_order_queue
, skb
))
4502 next
= skb_queue_next(&tp
->out_of_order_queue
, skb
);
4505 /* Segment is terminated when we see gap or when
4506 * we are at the end of all the queue. */
4508 after(TCP_SKB_CB(skb
)->seq
, end
) ||
4509 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
4510 tcp_collapse(sk
, &tp
->out_of_order_queue
,
4511 head
, skb
, start
, end
);
4515 /* Start new segment */
4516 start
= TCP_SKB_CB(skb
)->seq
;
4517 end
= TCP_SKB_CB(skb
)->end_seq
;
4519 if (before(TCP_SKB_CB(skb
)->seq
, start
))
4520 start
= TCP_SKB_CB(skb
)->seq
;
4521 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
4522 end
= TCP_SKB_CB(skb
)->end_seq
;
4528 * Purge the out-of-order queue.
4529 * Return true if queue was pruned.
4531 static bool tcp_prune_ofo_queue(struct sock
*sk
)
4533 struct tcp_sock
*tp
= tcp_sk(sk
);
4536 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4537 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_OFOPRUNED
);
4538 __skb_queue_purge(&tp
->out_of_order_queue
);
4540 /* Reset SACK state. A conforming SACK implementation will
4541 * do the same at a timeout based retransmit. When a connection
4542 * is in a sad state like this, we care only about integrity
4543 * of the connection not performance.
4545 if (tp
->rx_opt
.sack_ok
)
4546 tcp_sack_reset(&tp
->rx_opt
);
4553 /* Reduce allocated memory if we can, trying to get
4554 * the socket within its memory limits again.
4556 * Return less than zero if we should start dropping frames
4557 * until the socket owning process reads some of the data
4558 * to stabilize the situation.
4560 static int tcp_prune_queue(struct sock
*sk
)
4562 struct tcp_sock
*tp
= tcp_sk(sk
);
4564 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
4566 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PRUNECALLED
);
4568 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
4569 tcp_clamp_window(sk
);
4570 else if (sk_under_memory_pressure(sk
))
4571 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
4573 tcp_collapse_ofo_queue(sk
);
4574 if (!skb_queue_empty(&sk
->sk_receive_queue
))
4575 tcp_collapse(sk
, &sk
->sk_receive_queue
,
4576 skb_peek(&sk
->sk_receive_queue
),
4578 tp
->copied_seq
, tp
->rcv_nxt
);
4581 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4584 /* Collapsing did not help, destructive actions follow.
4585 * This must not ever occur. */
4587 tcp_prune_ofo_queue(sk
);
4589 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4592 /* If we are really being abused, tell the caller to silently
4593 * drop receive data on the floor. It will get retransmitted
4594 * and hopefully then we'll have sufficient space.
4596 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_RCVPRUNED
);
4598 /* Massive buffer overcommit. */
4603 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4604 * As additional protections, we do not touch cwnd in retransmission phases,
4605 * and if application hit its sndbuf limit recently.
4607 void tcp_cwnd_application_limited(struct sock
*sk
)
4609 struct tcp_sock
*tp
= tcp_sk(sk
);
4611 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
4612 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
4613 /* Limited by application or receiver window. */
4614 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
4615 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
4616 if (win_used
< tp
->snd_cwnd
) {
4617 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
4618 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
4620 tp
->snd_cwnd_used
= 0;
4622 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4625 static bool tcp_should_expand_sndbuf(const struct sock
*sk
)
4627 const struct tcp_sock
*tp
= tcp_sk(sk
);
4629 /* If the user specified a specific send buffer setting, do
4632 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
4635 /* If we are under global TCP memory pressure, do not expand. */
4636 if (sk_under_memory_pressure(sk
))
4639 /* If we are under soft global TCP memory pressure, do not expand. */
4640 if (sk_memory_allocated(sk
) >= sk_prot_mem_limits(sk
, 0))
4643 /* If we filled the congestion window, do not expand. */
4644 if (tp
->packets_out
>= tp
->snd_cwnd
)
4650 /* When incoming ACK allowed to free some skb from write_queue,
4651 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4652 * on the exit from tcp input handler.
4654 * PROBLEM: sndbuf expansion does not work well with largesend.
4656 static void tcp_new_space(struct sock
*sk
)
4658 struct tcp_sock
*tp
= tcp_sk(sk
);
4660 if (tcp_should_expand_sndbuf(sk
)) {
4661 int sndmem
= SKB_TRUESIZE(max_t(u32
,
4662 tp
->rx_opt
.mss_clamp
,
4665 int demanded
= max_t(unsigned int, tp
->snd_cwnd
,
4666 tp
->reordering
+ 1);
4667 sndmem
*= 2 * demanded
;
4668 if (sndmem
> sk
->sk_sndbuf
)
4669 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
4670 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4673 sk
->sk_write_space(sk
);
4676 static void tcp_check_space(struct sock
*sk
)
4678 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
4679 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
4680 if (sk
->sk_socket
&&
4681 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
4686 static inline void tcp_data_snd_check(struct sock
*sk
)
4688 tcp_push_pending_frames(sk
);
4689 tcp_check_space(sk
);
4693 * Check if sending an ack is needed.
4695 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
4697 struct tcp_sock
*tp
= tcp_sk(sk
);
4699 /* More than one full frame received... */
4700 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
&&
4701 /* ... and right edge of window advances far enough.
4702 * (tcp_recvmsg() will send ACK otherwise). Or...
4704 __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
4705 /* We ACK each frame or... */
4706 tcp_in_quickack_mode(sk
) ||
4707 /* We have out of order data. */
4708 (ofo_possible
&& skb_peek(&tp
->out_of_order_queue
))) {
4709 /* Then ack it now */
4712 /* Else, send delayed ack. */
4713 tcp_send_delayed_ack(sk
);
4717 static inline void tcp_ack_snd_check(struct sock
*sk
)
4719 if (!inet_csk_ack_scheduled(sk
)) {
4720 /* We sent a data segment already. */
4723 __tcp_ack_snd_check(sk
, 1);
4727 * This routine is only called when we have urgent data
4728 * signaled. Its the 'slow' part of tcp_urg. It could be
4729 * moved inline now as tcp_urg is only called from one
4730 * place. We handle URGent data wrong. We have to - as
4731 * BSD still doesn't use the correction from RFC961.
4732 * For 1003.1g we should support a new option TCP_STDURG to permit
4733 * either form (or just set the sysctl tcp_stdurg).
4736 static void tcp_check_urg(struct sock
*sk
, const struct tcphdr
*th
)
4738 struct tcp_sock
*tp
= tcp_sk(sk
);
4739 u32 ptr
= ntohs(th
->urg_ptr
);
4741 if (ptr
&& !sysctl_tcp_stdurg
)
4743 ptr
+= ntohl(th
->seq
);
4745 /* Ignore urgent data that we've already seen and read. */
4746 if (after(tp
->copied_seq
, ptr
))
4749 /* Do not replay urg ptr.
4751 * NOTE: interesting situation not covered by specs.
4752 * Misbehaving sender may send urg ptr, pointing to segment,
4753 * which we already have in ofo queue. We are not able to fetch
4754 * such data and will stay in TCP_URG_NOTYET until will be eaten
4755 * by recvmsg(). Seems, we are not obliged to handle such wicked
4756 * situations. But it is worth to think about possibility of some
4757 * DoSes using some hypothetical application level deadlock.
4759 if (before(ptr
, tp
->rcv_nxt
))
4762 /* Do we already have a newer (or duplicate) urgent pointer? */
4763 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
4766 /* Tell the world about our new urgent pointer. */
4769 /* We may be adding urgent data when the last byte read was
4770 * urgent. To do this requires some care. We cannot just ignore
4771 * tp->copied_seq since we would read the last urgent byte again
4772 * as data, nor can we alter copied_seq until this data arrives
4773 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4775 * NOTE. Double Dutch. Rendering to plain English: author of comment
4776 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4777 * and expect that both A and B disappear from stream. This is _wrong_.
4778 * Though this happens in BSD with high probability, this is occasional.
4779 * Any application relying on this is buggy. Note also, that fix "works"
4780 * only in this artificial test. Insert some normal data between A and B and we will
4781 * decline of BSD again. Verdict: it is better to remove to trap
4784 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
4785 !sock_flag(sk
, SOCK_URGINLINE
) && tp
->copied_seq
!= tp
->rcv_nxt
) {
4786 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
4788 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4789 __skb_unlink(skb
, &sk
->sk_receive_queue
);
4794 tp
->urg_data
= TCP_URG_NOTYET
;
4797 /* Disable header prediction. */
4801 /* This is the 'fast' part of urgent handling. */
4802 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, const struct tcphdr
*th
)
4804 struct tcp_sock
*tp
= tcp_sk(sk
);
4806 /* Check if we get a new urgent pointer - normally not. */
4808 tcp_check_urg(sk
, th
);
4810 /* Do we wait for any urgent data? - normally not... */
4811 if (tp
->urg_data
== TCP_URG_NOTYET
) {
4812 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
4815 /* Is the urgent pointer pointing into this packet? */
4816 if (ptr
< skb
->len
) {
4818 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
4820 tp
->urg_data
= TCP_URG_VALID
| tmp
;
4821 if (!sock_flag(sk
, SOCK_DEAD
))
4822 sk
->sk_data_ready(sk
, 0);
4827 static int tcp_copy_to_iovec(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
4829 struct tcp_sock
*tp
= tcp_sk(sk
);
4830 int chunk
= skb
->len
- hlen
;
4834 if (skb_csum_unnecessary(skb
))
4835 err
= skb_copy_datagram_iovec(skb
, hlen
, tp
->ucopy
.iov
, chunk
);
4837 err
= skb_copy_and_csum_datagram_iovec(skb
, hlen
,
4841 tp
->ucopy
.len
-= chunk
;
4842 tp
->copied_seq
+= chunk
;
4843 tcp_rcv_space_adjust(sk
);
4850 static __sum16
__tcp_checksum_complete_user(struct sock
*sk
,
4851 struct sk_buff
*skb
)
4855 if (sock_owned_by_user(sk
)) {
4857 result
= __tcp_checksum_complete(skb
);
4860 result
= __tcp_checksum_complete(skb
);
4865 static inline bool tcp_checksum_complete_user(struct sock
*sk
,
4866 struct sk_buff
*skb
)
4868 return !skb_csum_unnecessary(skb
) &&
4869 __tcp_checksum_complete_user(sk
, skb
);
4872 #ifdef CONFIG_NET_DMA
4873 static bool tcp_dma_try_early_copy(struct sock
*sk
, struct sk_buff
*skb
,
4876 struct tcp_sock
*tp
= tcp_sk(sk
);
4877 int chunk
= skb
->len
- hlen
;
4879 bool copied_early
= false;
4881 if (tp
->ucopy
.wakeup
)
4884 if (!tp
->ucopy
.dma_chan
&& tp
->ucopy
.pinned_list
)
4885 tp
->ucopy
.dma_chan
= net_dma_find_channel();
4887 if (tp
->ucopy
.dma_chan
&& skb_csum_unnecessary(skb
)) {
4889 dma_cookie
= dma_skb_copy_datagram_iovec(tp
->ucopy
.dma_chan
,
4891 tp
->ucopy
.iov
, chunk
,
4892 tp
->ucopy
.pinned_list
);
4897 tp
->ucopy
.dma_cookie
= dma_cookie
;
4898 copied_early
= true;
4900 tp
->ucopy
.len
-= chunk
;
4901 tp
->copied_seq
+= chunk
;
4902 tcp_rcv_space_adjust(sk
);
4904 if ((tp
->ucopy
.len
== 0) ||
4905 (tcp_flag_word(tcp_hdr(skb
)) & TCP_FLAG_PSH
) ||
4906 (atomic_read(&sk
->sk_rmem_alloc
) > (sk
->sk_rcvbuf
>> 1))) {
4907 tp
->ucopy
.wakeup
= 1;
4908 sk
->sk_data_ready(sk
, 0);
4910 } else if (chunk
> 0) {
4911 tp
->ucopy
.wakeup
= 1;
4912 sk
->sk_data_ready(sk
, 0);
4915 return copied_early
;
4917 #endif /* CONFIG_NET_DMA */
4919 /* Does PAWS and seqno based validation of an incoming segment, flags will
4920 * play significant role here.
4922 static bool tcp_validate_incoming(struct sock
*sk
, struct sk_buff
*skb
,
4923 const struct tcphdr
*th
, int syn_inerr
)
4925 struct tcp_sock
*tp
= tcp_sk(sk
);
4927 /* RFC1323: H1. Apply PAWS check first. */
4928 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
4929 tcp_paws_discard(sk
, skb
)) {
4931 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSESTABREJECTED
);
4932 tcp_send_dupack(sk
, skb
);
4935 /* Reset is accepted even if it did not pass PAWS. */
4938 /* Step 1: check sequence number */
4939 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4940 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4941 * (RST) segments are validated by checking their SEQ-fields."
4942 * And page 69: "If an incoming segment is not acceptable,
4943 * an acknowledgment should be sent in reply (unless the RST
4944 * bit is set, if so drop the segment and return)".
4949 tcp_send_dupack(sk
, skb
);
4954 /* Step 2: check RST bit */
4957 * If sequence number exactly matches RCV.NXT, then
4958 * RESET the connection
4960 * Send a challenge ACK
4962 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
)
4965 tcp_send_challenge_ack(sk
);
4969 /* step 3: check security and precedence [ignored] */
4971 /* step 4: Check for a SYN
4972 * RFC 5691 4.2 : Send a challenge ack
4977 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
4978 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSYNCHALLENGE
);
4979 tcp_send_challenge_ack(sk
);
4991 * TCP receive function for the ESTABLISHED state.
4993 * It is split into a fast path and a slow path. The fast path is
4995 * - A zero window was announced from us - zero window probing
4996 * is only handled properly in the slow path.
4997 * - Out of order segments arrived.
4998 * - Urgent data is expected.
4999 * - There is no buffer space left
5000 * - Unexpected TCP flags/window values/header lengths are received
5001 * (detected by checking the TCP header against pred_flags)
5002 * - Data is sent in both directions. Fast path only supports pure senders
5003 * or pure receivers (this means either the sequence number or the ack
5004 * value must stay constant)
5005 * - Unexpected TCP option.
5007 * When these conditions are not satisfied it drops into a standard
5008 * receive procedure patterned after RFC793 to handle all cases.
5009 * The first three cases are guaranteed by proper pred_flags setting,
5010 * the rest is checked inline. Fast processing is turned on in
5011 * tcp_data_queue when everything is OK.
5013 int tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
5014 const struct tcphdr
*th
, unsigned int len
)
5016 struct tcp_sock
*tp
= tcp_sk(sk
);
5018 if (unlikely(sk
->sk_rx_dst
== NULL
))
5019 inet_csk(sk
)->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
5021 * Header prediction.
5022 * The code loosely follows the one in the famous
5023 * "30 instruction TCP receive" Van Jacobson mail.
5025 * Van's trick is to deposit buffers into socket queue
5026 * on a device interrupt, to call tcp_recv function
5027 * on the receive process context and checksum and copy
5028 * the buffer to user space. smart...
5030 * Our current scheme is not silly either but we take the
5031 * extra cost of the net_bh soft interrupt processing...
5032 * We do checksum and copy also but from device to kernel.
5035 tp
->rx_opt
.saw_tstamp
= 0;
5037 /* pred_flags is 0xS?10 << 16 + snd_wnd
5038 * if header_prediction is to be made
5039 * 'S' will always be tp->tcp_header_len >> 2
5040 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5041 * turn it off (when there are holes in the receive
5042 * space for instance)
5043 * PSH flag is ignored.
5046 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
5047 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
&&
5048 !after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
)) {
5049 int tcp_header_len
= tp
->tcp_header_len
;
5051 /* Timestamp header prediction: tcp_header_len
5052 * is automatically equal to th->doff*4 due to pred_flags
5056 /* Check timestamp */
5057 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
5058 /* No? Slow path! */
5059 if (!tcp_parse_aligned_timestamp(tp
, th
))
5062 /* If PAWS failed, check it more carefully in slow path */
5063 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
5066 /* DO NOT update ts_recent here, if checksum fails
5067 * and timestamp was corrupted part, it will result
5068 * in a hung connection since we will drop all
5069 * future packets due to the PAWS test.
5073 if (len
<= tcp_header_len
) {
5074 /* Bulk data transfer: sender */
5075 if (len
== tcp_header_len
) {
5076 /* Predicted packet is in window by definition.
5077 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5078 * Hence, check seq<=rcv_wup reduces to:
5080 if (tcp_header_len
==
5081 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5082 tp
->rcv_nxt
== tp
->rcv_wup
)
5083 tcp_store_ts_recent(tp
);
5085 /* We know that such packets are checksummed
5088 tcp_ack(sk
, skb
, 0);
5090 tcp_data_snd_check(sk
);
5092 } else { /* Header too small */
5093 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5098 int copied_early
= 0;
5099 bool fragstolen
= false;
5101 if (tp
->copied_seq
== tp
->rcv_nxt
&&
5102 len
- tcp_header_len
<= tp
->ucopy
.len
) {
5103 #ifdef CONFIG_NET_DMA
5104 if (tp
->ucopy
.task
== current
&&
5105 sock_owned_by_user(sk
) &&
5106 tcp_dma_try_early_copy(sk
, skb
, tcp_header_len
)) {
5111 if (tp
->ucopy
.task
== current
&&
5112 sock_owned_by_user(sk
) && !copied_early
) {
5113 __set_current_state(TASK_RUNNING
);
5115 if (!tcp_copy_to_iovec(sk
, skb
, 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
) +
5125 TCPOLEN_TSTAMP_ALIGNED
) &&
5126 tp
->rcv_nxt
== tp
->rcv_wup
)
5127 tcp_store_ts_recent(tp
);
5129 tcp_rcv_rtt_measure_ts(sk
, skb
);
5131 __skb_pull(skb
, tcp_header_len
);
5132 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
5133 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITSTOUSER
);
5136 tcp_cleanup_rbuf(sk
, skb
->len
);
5139 if (tcp_checksum_complete_user(sk
, skb
))
5142 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
5145 /* Predicted packet is in window by definition.
5146 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5147 * Hence, check seq<=rcv_wup reduces to:
5149 if (tcp_header_len
==
5150 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5151 tp
->rcv_nxt
== tp
->rcv_wup
)
5152 tcp_store_ts_recent(tp
);
5154 tcp_rcv_rtt_measure_ts(sk
, skb
);
5156 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITS
);
5158 /* Bulk data transfer: receiver */
5159 eaten
= tcp_queue_rcv(sk
, skb
, tcp_header_len
,
5163 tcp_event_data_recv(sk
, skb
);
5165 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
5166 /* Well, only one small jumplet in fast path... */
5167 tcp_ack(sk
, skb
, FLAG_DATA
);
5168 tcp_data_snd_check(sk
);
5169 if (!inet_csk_ack_scheduled(sk
))
5173 if (!copied_early
|| tp
->rcv_nxt
!= tp
->rcv_wup
)
5174 __tcp_ack_snd_check(sk
, 0);
5176 #ifdef CONFIG_NET_DMA
5178 __skb_queue_tail(&sk
->sk_async_wait_queue
, skb
);
5182 kfree_skb_partial(skb
, fragstolen
);
5183 sk
->sk_data_ready(sk
, 0);
5189 if (len
< (th
->doff
<< 2) || tcp_checksum_complete_user(sk
, skb
))
5192 if (!th
->ack
&& !th
->rst
)
5196 * Standard slow path.
5199 if (!tcp_validate_incoming(sk
, skb
, th
, 1))
5203 if (tcp_ack(sk
, skb
, FLAG_SLOWPATH
| FLAG_UPDATE_TS_RECENT
) < 0)
5206 tcp_rcv_rtt_measure_ts(sk
, skb
);
5208 /* Process urgent data. */
5209 tcp_urg(sk
, skb
, th
);
5211 /* step 7: process the segment text */
5212 tcp_data_queue(sk
, skb
);
5214 tcp_data_snd_check(sk
);
5215 tcp_ack_snd_check(sk
);
5219 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_CSUMERRORS
);
5220 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5226 EXPORT_SYMBOL(tcp_rcv_established
);
5228 void tcp_finish_connect(struct sock
*sk
, struct sk_buff
*skb
)
5230 struct tcp_sock
*tp
= tcp_sk(sk
);
5231 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5233 tcp_set_state(sk
, TCP_ESTABLISHED
);
5236 icsk
->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
5237 security_inet_conn_established(sk
, skb
);
5240 /* Make sure socket is routed, for correct metrics. */
5241 icsk
->icsk_af_ops
->rebuild_header(sk
);
5243 tcp_init_metrics(sk
);
5245 tcp_init_congestion_control(sk
);
5247 /* Prevent spurious tcp_cwnd_restart() on first data
5250 tp
->lsndtime
= tcp_time_stamp
;
5252 tcp_init_buffer_space(sk
);
5254 if (sock_flag(sk
, SOCK_KEEPOPEN
))
5255 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
5257 if (!tp
->rx_opt
.snd_wscale
)
5258 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
5262 if (!sock_flag(sk
, SOCK_DEAD
)) {
5263 sk
->sk_state_change(sk
);
5264 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5268 static bool tcp_rcv_fastopen_synack(struct sock
*sk
, struct sk_buff
*synack
,
5269 struct tcp_fastopen_cookie
*cookie
)
5271 struct tcp_sock
*tp
= tcp_sk(sk
);
5272 struct sk_buff
*data
= tp
->syn_data
? tcp_write_queue_head(sk
) : NULL
;
5273 u16 mss
= tp
->rx_opt
.mss_clamp
;
5276 if (mss
== tp
->rx_opt
.user_mss
) {
5277 struct tcp_options_received opt
;
5279 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5280 tcp_clear_options(&opt
);
5281 opt
.user_mss
= opt
.mss_clamp
= 0;
5282 tcp_parse_options(synack
, &opt
, 0, NULL
);
5283 mss
= opt
.mss_clamp
;
5286 if (!tp
->syn_fastopen
) /* Ignore an unsolicited cookie */
5289 /* The SYN-ACK neither has cookie nor acknowledges the data. Presumably
5290 * the remote receives only the retransmitted (regular) SYNs: either
5291 * the original SYN-data or the corresponding SYN-ACK is lost.
5293 syn_drop
= (cookie
->len
<= 0 && data
&& tp
->total_retrans
);
5295 tcp_fastopen_cache_set(sk
, mss
, cookie
, syn_drop
);
5297 if (data
) { /* Retransmit unacked data in SYN */
5298 tcp_for_write_queue_from(data
, sk
) {
5299 if (data
== tcp_send_head(sk
) ||
5300 __tcp_retransmit_skb(sk
, data
))
5306 tp
->syn_data_acked
= tp
->syn_data
;
5310 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5311 const struct tcphdr
*th
, unsigned int len
)
5313 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5314 struct tcp_sock
*tp
= tcp_sk(sk
);
5315 struct tcp_fastopen_cookie foc
= { .len
= -1 };
5316 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
5318 tcp_parse_options(skb
, &tp
->rx_opt
, 0, &foc
);
5319 if (tp
->rx_opt
.saw_tstamp
)
5320 tp
->rx_opt
.rcv_tsecr
-= tp
->tsoffset
;
5324 * "If the state is SYN-SENT then
5325 * first check the ACK bit
5326 * If the ACK bit is set
5327 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5328 * a reset (unless the RST bit is set, if so drop
5329 * the segment and return)"
5331 if (!after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_una
) ||
5332 after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
))
5333 goto reset_and_undo
;
5335 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
5336 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
5338 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSACTIVEREJECTED
);
5339 goto reset_and_undo
;
5342 /* Now ACK is acceptable.
5344 * "If the RST bit is set
5345 * If the ACK was acceptable then signal the user "error:
5346 * connection reset", drop the segment, enter CLOSED state,
5347 * delete TCB, and return."
5356 * "fifth, if neither of the SYN or RST bits is set then
5357 * drop the segment and return."
5363 goto discard_and_undo
;
5366 * "If the SYN bit is on ...
5367 * are acceptable then ...
5368 * (our SYN has been ACKed), change the connection
5369 * state to ESTABLISHED..."
5372 TCP_ECN_rcv_synack(tp
, th
);
5374 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5375 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
5377 /* Ok.. it's good. Set up sequence numbers and
5378 * move to established.
5380 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5381 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5383 /* RFC1323: The window in SYN & SYN/ACK segments is
5386 tp
->snd_wnd
= ntohs(th
->window
);
5388 if (!tp
->rx_opt
.wscale_ok
) {
5389 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
5390 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
5393 if (tp
->rx_opt
.saw_tstamp
) {
5394 tp
->rx_opt
.tstamp_ok
= 1;
5395 tp
->tcp_header_len
=
5396 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5397 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5398 tcp_store_ts_recent(tp
);
5400 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5403 if (tcp_is_sack(tp
) && sysctl_tcp_fack
)
5404 tcp_enable_fack(tp
);
5407 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5408 tcp_initialize_rcv_mss(sk
);
5410 /* Remember, tcp_poll() does not lock socket!
5411 * Change state from SYN-SENT only after copied_seq
5412 * is initialized. */
5413 tp
->copied_seq
= tp
->rcv_nxt
;
5417 tcp_finish_connect(sk
, skb
);
5419 if ((tp
->syn_fastopen
|| tp
->syn_data
) &&
5420 tcp_rcv_fastopen_synack(sk
, skb
, &foc
))
5423 if (sk
->sk_write_pending
||
5424 icsk
->icsk_accept_queue
.rskq_defer_accept
||
5425 icsk
->icsk_ack
.pingpong
) {
5426 /* Save one ACK. Data will be ready after
5427 * several ticks, if write_pending is set.
5429 * It may be deleted, but with this feature tcpdumps
5430 * look so _wonderfully_ clever, that I was not able
5431 * to stand against the temptation 8) --ANK
5433 inet_csk_schedule_ack(sk
);
5434 icsk
->icsk_ack
.lrcvtime
= tcp_time_stamp
;
5435 tcp_enter_quickack_mode(sk
);
5436 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
5437 TCP_DELACK_MAX
, TCP_RTO_MAX
);
5448 /* No ACK in the segment */
5452 * "If the RST bit is set
5454 * Otherwise (no ACK) drop the segment and return."
5457 goto discard_and_undo
;
5461 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&&
5462 tcp_paws_reject(&tp
->rx_opt
, 0))
5463 goto discard_and_undo
;
5466 /* We see SYN without ACK. It is attempt of
5467 * simultaneous connect with crossed SYNs.
5468 * Particularly, it can be connect to self.
5470 tcp_set_state(sk
, TCP_SYN_RECV
);
5472 if (tp
->rx_opt
.saw_tstamp
) {
5473 tp
->rx_opt
.tstamp_ok
= 1;
5474 tcp_store_ts_recent(tp
);
5475 tp
->tcp_header_len
=
5476 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5478 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5481 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5482 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5484 /* RFC1323: The window in SYN & SYN/ACK segments is
5487 tp
->snd_wnd
= ntohs(th
->window
);
5488 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
5489 tp
->max_window
= tp
->snd_wnd
;
5491 TCP_ECN_rcv_syn(tp
, th
);
5494 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5495 tcp_initialize_rcv_mss(sk
);
5497 tcp_send_synack(sk
);
5499 /* Note, we could accept data and URG from this segment.
5500 * There are no obstacles to make this (except that we must
5501 * either change tcp_recvmsg() to prevent it from returning data
5502 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5504 * However, if we ignore data in ACKless segments sometimes,
5505 * we have no reasons to accept it sometimes.
5506 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5507 * is not flawless. So, discard packet for sanity.
5508 * Uncomment this return to process the data.
5515 /* "fifth, if neither of the SYN or RST bits is set then
5516 * drop the segment and return."
5520 tcp_clear_options(&tp
->rx_opt
);
5521 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5525 tcp_clear_options(&tp
->rx_opt
);
5526 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5531 * This function implements the receiving procedure of RFC 793 for
5532 * all states except ESTABLISHED and TIME_WAIT.
5533 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5534 * address independent.
5537 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5538 const struct tcphdr
*th
, unsigned int len
)
5540 struct tcp_sock
*tp
= tcp_sk(sk
);
5541 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5542 struct request_sock
*req
;
5546 tp
->rx_opt
.saw_tstamp
= 0;
5548 switch (sk
->sk_state
) {
5562 if (icsk
->icsk_af_ops
->conn_request(sk
, skb
) < 0)
5565 /* Now we have several options: In theory there is
5566 * nothing else in the frame. KA9Q has an option to
5567 * send data with the syn, BSD accepts data with the
5568 * syn up to the [to be] advertised window and
5569 * Solaris 2.1 gives you a protocol error. For now
5570 * we just ignore it, that fits the spec precisely
5571 * and avoids incompatibilities. It would be nice in
5572 * future to drop through and process the data.
5574 * Now that TTCP is starting to be used we ought to
5576 * But, this leaves one open to an easy denial of
5577 * service attack, and SYN cookies can't defend
5578 * against this problem. So, we drop the data
5579 * in the interest of security over speed unless
5580 * it's still in use.
5588 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
, len
);
5592 /* Do step6 onward by hand. */
5593 tcp_urg(sk
, skb
, th
);
5595 tcp_data_snd_check(sk
);
5599 req
= tp
->fastopen_rsk
;
5601 WARN_ON_ONCE(sk
->sk_state
!= TCP_SYN_RECV
&&
5602 sk
->sk_state
!= TCP_FIN_WAIT1
);
5604 if (tcp_check_req(sk
, skb
, req
, NULL
, true) == NULL
)
5608 if (!th
->ack
&& !th
->rst
)
5611 if (!tcp_validate_incoming(sk
, skb
, th
, 0))
5614 /* step 5: check the ACK field */
5615 acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
|
5616 FLAG_UPDATE_TS_RECENT
) > 0;
5618 switch (sk
->sk_state
) {
5623 /* Once we leave TCP_SYN_RECV, we no longer need req
5627 tcp_synack_rtt_meas(sk
, req
);
5628 tp
->total_retrans
= req
->num_retrans
;
5630 reqsk_fastopen_remove(sk
, req
, false);
5632 /* Make sure socket is routed, for correct metrics. */
5633 icsk
->icsk_af_ops
->rebuild_header(sk
);
5634 tcp_init_congestion_control(sk
);
5637 tcp_init_buffer_space(sk
);
5638 tp
->copied_seq
= tp
->rcv_nxt
;
5641 tcp_set_state(sk
, TCP_ESTABLISHED
);
5642 sk
->sk_state_change(sk
);
5644 /* Note, that this wakeup is only for marginal crossed SYN case.
5645 * Passively open sockets are not waked up, because
5646 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5649 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5651 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
5652 tp
->snd_wnd
= ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
;
5653 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5655 if (tp
->rx_opt
.tstamp_ok
)
5656 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5659 /* Re-arm the timer because data may have been sent out.
5660 * This is similar to the regular data transmission case
5661 * when new data has just been ack'ed.
5663 * (TFO) - we could try to be more aggressive and
5664 * retransmitting any data sooner based on when they
5669 tcp_init_metrics(sk
);
5671 /* Prevent spurious tcp_cwnd_restart() on first data packet */
5672 tp
->lsndtime
= tcp_time_stamp
;
5674 tcp_initialize_rcv_mss(sk
);
5675 tcp_fast_path_on(tp
);
5678 case TCP_FIN_WAIT1
: {
5679 struct dst_entry
*dst
;
5682 /* If we enter the TCP_FIN_WAIT1 state and we are a
5683 * Fast Open socket and this is the first acceptable
5684 * ACK we have received, this would have acknowledged
5685 * our SYNACK so stop the SYNACK timer.
5688 /* Return RST if ack_seq is invalid.
5689 * Note that RFC793 only says to generate a
5690 * DUPACK for it but for TCP Fast Open it seems
5691 * better to treat this case like TCP_SYN_RECV
5696 /* We no longer need the request sock. */
5697 reqsk_fastopen_remove(sk
, req
, false);
5700 if (tp
->snd_una
!= tp
->write_seq
)
5703 tcp_set_state(sk
, TCP_FIN_WAIT2
);
5704 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
5706 dst
= __sk_dst_get(sk
);
5710 if (!sock_flag(sk
, SOCK_DEAD
)) {
5711 /* Wake up lingering close() */
5712 sk
->sk_state_change(sk
);
5716 if (tp
->linger2
< 0 ||
5717 (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5718 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
))) {
5720 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
5724 tmo
= tcp_fin_time(sk
);
5725 if (tmo
> TCP_TIMEWAIT_LEN
) {
5726 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
5727 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
5728 /* Bad case. We could lose such FIN otherwise.
5729 * It is not a big problem, but it looks confusing
5730 * and not so rare event. We still can lose it now,
5731 * if it spins in bh_lock_sock(), but it is really
5734 inet_csk_reset_keepalive_timer(sk
, tmo
);
5736 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
5743 if (tp
->snd_una
== tp
->write_seq
) {
5744 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
5750 if (tp
->snd_una
== tp
->write_seq
) {
5751 tcp_update_metrics(sk
);
5758 /* step 6: check the URG bit */
5759 tcp_urg(sk
, skb
, th
);
5761 /* step 7: process the segment text */
5762 switch (sk
->sk_state
) {
5763 case TCP_CLOSE_WAIT
:
5766 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
5770 /* RFC 793 says to queue data in these states,
5771 * RFC 1122 says we MUST send a reset.
5772 * BSD 4.4 also does reset.
5774 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
5775 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5776 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
5777 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
5783 case TCP_ESTABLISHED
:
5784 tcp_data_queue(sk
, skb
);
5789 /* tcp_data could move socket to TIME-WAIT */
5790 if (sk
->sk_state
!= TCP_CLOSE
) {
5791 tcp_data_snd_check(sk
);
5792 tcp_ack_snd_check(sk
);
5801 EXPORT_SYMBOL(tcp_rcv_state_process
);