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. */
351 static void tcp_fixup_rcvbuf(struct sock
*sk
)
353 u32 mss
= tcp_sk(sk
)->advmss
;
354 u32 icwnd
= TCP_DEFAULT_INIT_RCVWND
;
357 /* Limit to 10 segments if mss <= 1460,
358 * or 14600/mss segments, with a minimum of two segments.
361 icwnd
= max_t(u32
, (1460 * TCP_DEFAULT_INIT_RCVWND
) / mss
, 2);
363 rcvmem
= SKB_TRUESIZE(mss
+ MAX_TCP_HEADER
);
364 while (tcp_win_from_space(rcvmem
) < mss
)
369 if (sk
->sk_rcvbuf
< rcvmem
)
370 sk
->sk_rcvbuf
= min(rcvmem
, sysctl_tcp_rmem
[2]);
373 /* 4. Try to fixup all. It is made immediately after connection enters
376 void tcp_init_buffer_space(struct sock
*sk
)
378 struct tcp_sock
*tp
= tcp_sk(sk
);
381 if (!(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
))
382 tcp_fixup_rcvbuf(sk
);
383 if (!(sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
))
384 tcp_fixup_sndbuf(sk
);
386 tp
->rcvq_space
.space
= tp
->rcv_wnd
;
388 maxwin
= tcp_full_space(sk
);
390 if (tp
->window_clamp
>= maxwin
) {
391 tp
->window_clamp
= maxwin
;
393 if (sysctl_tcp_app_win
&& maxwin
> 4 * tp
->advmss
)
394 tp
->window_clamp
= max(maxwin
-
395 (maxwin
>> sysctl_tcp_app_win
),
399 /* Force reservation of one segment. */
400 if (sysctl_tcp_app_win
&&
401 tp
->window_clamp
> 2 * tp
->advmss
&&
402 tp
->window_clamp
+ tp
->advmss
> maxwin
)
403 tp
->window_clamp
= max(2 * tp
->advmss
, maxwin
- tp
->advmss
);
405 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, tp
->window_clamp
);
406 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
409 /* 5. Recalculate window clamp after socket hit its memory bounds. */
410 static void tcp_clamp_window(struct sock
*sk
)
412 struct tcp_sock
*tp
= tcp_sk(sk
);
413 struct inet_connection_sock
*icsk
= inet_csk(sk
);
415 icsk
->icsk_ack
.quick
= 0;
417 if (sk
->sk_rcvbuf
< sysctl_tcp_rmem
[2] &&
418 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
) &&
419 !sk_under_memory_pressure(sk
) &&
420 sk_memory_allocated(sk
) < sk_prot_mem_limits(sk
, 0)) {
421 sk
->sk_rcvbuf
= min(atomic_read(&sk
->sk_rmem_alloc
),
424 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
)
425 tp
->rcv_ssthresh
= min(tp
->window_clamp
, 2U * tp
->advmss
);
428 /* Initialize RCV_MSS value.
429 * RCV_MSS is an our guess about MSS used by the peer.
430 * We haven't any direct information about the MSS.
431 * It's better to underestimate the RCV_MSS rather than overestimate.
432 * Overestimations make us ACKing less frequently than needed.
433 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
435 void tcp_initialize_rcv_mss(struct sock
*sk
)
437 const struct tcp_sock
*tp
= tcp_sk(sk
);
438 unsigned int hint
= min_t(unsigned int, tp
->advmss
, tp
->mss_cache
);
440 hint
= min(hint
, tp
->rcv_wnd
/ 2);
441 hint
= min(hint
, TCP_MSS_DEFAULT
);
442 hint
= max(hint
, TCP_MIN_MSS
);
444 inet_csk(sk
)->icsk_ack
.rcv_mss
= hint
;
446 EXPORT_SYMBOL(tcp_initialize_rcv_mss
);
448 /* Receiver "autotuning" code.
450 * The algorithm for RTT estimation w/o timestamps is based on
451 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
452 * <http://public.lanl.gov/radiant/pubs.html#DRS>
454 * More detail on this code can be found at
455 * <http://staff.psc.edu/jheffner/>,
456 * though this reference is out of date. A new paper
459 static void tcp_rcv_rtt_update(struct tcp_sock
*tp
, u32 sample
, int win_dep
)
461 u32 new_sample
= tp
->rcv_rtt_est
.rtt
;
467 if (new_sample
!= 0) {
468 /* If we sample in larger samples in the non-timestamp
469 * case, we could grossly overestimate the RTT especially
470 * with chatty applications or bulk transfer apps which
471 * are stalled on filesystem I/O.
473 * Also, since we are only going for a minimum in the
474 * non-timestamp case, we do not smooth things out
475 * else with timestamps disabled convergence takes too
479 m
-= (new_sample
>> 3);
487 /* No previous measure. */
491 if (tp
->rcv_rtt_est
.rtt
!= new_sample
)
492 tp
->rcv_rtt_est
.rtt
= new_sample
;
495 static inline void tcp_rcv_rtt_measure(struct tcp_sock
*tp
)
497 if (tp
->rcv_rtt_est
.time
== 0)
499 if (before(tp
->rcv_nxt
, tp
->rcv_rtt_est
.seq
))
501 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rcv_rtt_est
.time
, 1);
504 tp
->rcv_rtt_est
.seq
= tp
->rcv_nxt
+ tp
->rcv_wnd
;
505 tp
->rcv_rtt_est
.time
= tcp_time_stamp
;
508 static inline void tcp_rcv_rtt_measure_ts(struct sock
*sk
,
509 const struct sk_buff
*skb
)
511 struct tcp_sock
*tp
= tcp_sk(sk
);
512 if (tp
->rx_opt
.rcv_tsecr
&&
513 (TCP_SKB_CB(skb
)->end_seq
-
514 TCP_SKB_CB(skb
)->seq
>= inet_csk(sk
)->icsk_ack
.rcv_mss
))
515 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
, 0);
519 * This function should be called every time data is copied to user space.
520 * It calculates the appropriate TCP receive buffer space.
522 void tcp_rcv_space_adjust(struct sock
*sk
)
524 struct tcp_sock
*tp
= tcp_sk(sk
);
528 if (tp
->rcvq_space
.time
== 0)
531 time
= tcp_time_stamp
- tp
->rcvq_space
.time
;
532 if (time
< (tp
->rcv_rtt_est
.rtt
>> 3) || tp
->rcv_rtt_est
.rtt
== 0)
535 space
= 2 * (tp
->copied_seq
- tp
->rcvq_space
.seq
);
537 space
= max(tp
->rcvq_space
.space
, space
);
539 if (tp
->rcvq_space
.space
!= space
) {
542 tp
->rcvq_space
.space
= space
;
544 if (sysctl_tcp_moderate_rcvbuf
&&
545 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
)) {
546 int new_clamp
= space
;
548 /* Receive space grows, normalize in order to
549 * take into account packet headers and sk_buff
550 * structure overhead.
555 rcvmem
= SKB_TRUESIZE(tp
->advmss
+ MAX_TCP_HEADER
);
556 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
559 space
= min(space
, sysctl_tcp_rmem
[2]);
560 if (space
> sk
->sk_rcvbuf
) {
561 sk
->sk_rcvbuf
= space
;
563 /* Make the window clamp follow along. */
564 tp
->window_clamp
= new_clamp
;
570 tp
->rcvq_space
.seq
= tp
->copied_seq
;
571 tp
->rcvq_space
.time
= tcp_time_stamp
;
574 /* There is something which you must keep in mind when you analyze the
575 * behavior of the tp->ato delayed ack timeout interval. When a
576 * connection starts up, we want to ack as quickly as possible. The
577 * problem is that "good" TCP's do slow start at the beginning of data
578 * transmission. The means that until we send the first few ACK's the
579 * sender will sit on his end and only queue most of his data, because
580 * he can only send snd_cwnd unacked packets at any given time. For
581 * each ACK we send, he increments snd_cwnd and transmits more of his
584 static void tcp_event_data_recv(struct sock
*sk
, struct sk_buff
*skb
)
586 struct tcp_sock
*tp
= tcp_sk(sk
);
587 struct inet_connection_sock
*icsk
= inet_csk(sk
);
590 inet_csk_schedule_ack(sk
);
592 tcp_measure_rcv_mss(sk
, skb
);
594 tcp_rcv_rtt_measure(tp
);
596 now
= tcp_time_stamp
;
598 if (!icsk
->icsk_ack
.ato
) {
599 /* The _first_ data packet received, initialize
600 * delayed ACK engine.
602 tcp_incr_quickack(sk
);
603 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
605 int m
= now
- icsk
->icsk_ack
.lrcvtime
;
607 if (m
<= TCP_ATO_MIN
/ 2) {
608 /* The fastest case is the first. */
609 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + TCP_ATO_MIN
/ 2;
610 } else if (m
< icsk
->icsk_ack
.ato
) {
611 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + m
;
612 if (icsk
->icsk_ack
.ato
> icsk
->icsk_rto
)
613 icsk
->icsk_ack
.ato
= icsk
->icsk_rto
;
614 } else if (m
> icsk
->icsk_rto
) {
615 /* Too long gap. Apparently sender failed to
616 * restart window, so that we send ACKs quickly.
618 tcp_incr_quickack(sk
);
622 icsk
->icsk_ack
.lrcvtime
= now
;
624 TCP_ECN_check_ce(tp
, skb
);
627 tcp_grow_window(sk
, skb
);
630 /* Called to compute a smoothed rtt estimate. The data fed to this
631 * routine either comes from timestamps, or from segments that were
632 * known _not_ to have been retransmitted [see Karn/Partridge
633 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
634 * piece by Van Jacobson.
635 * NOTE: the next three routines used to be one big routine.
636 * To save cycles in the RFC 1323 implementation it was better to break
637 * it up into three procedures. -- erics
639 static void tcp_rtt_estimator(struct sock
*sk
, const __u32 mrtt
)
641 struct tcp_sock
*tp
= tcp_sk(sk
);
642 long m
= mrtt
; /* RTT */
644 /* The following amusing code comes from Jacobson's
645 * article in SIGCOMM '88. Note that rtt and mdev
646 * are scaled versions of rtt and mean deviation.
647 * This is designed to be as fast as possible
648 * m stands for "measurement".
650 * On a 1990 paper the rto value is changed to:
651 * RTO = rtt + 4 * mdev
653 * Funny. This algorithm seems to be very broken.
654 * These formulae increase RTO, when it should be decreased, increase
655 * too slowly, when it should be increased quickly, decrease too quickly
656 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
657 * does not matter how to _calculate_ it. Seems, it was trap
658 * that VJ failed to avoid. 8)
663 m
-= (tp
->srtt
>> 3); /* m is now error in rtt est */
664 tp
->srtt
+= m
; /* rtt = 7/8 rtt + 1/8 new */
666 m
= -m
; /* m is now abs(error) */
667 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
668 /* This is similar to one of Eifel findings.
669 * Eifel blocks mdev updates when rtt decreases.
670 * This solution is a bit different: we use finer gain
671 * for mdev in this case (alpha*beta).
672 * Like Eifel it also prevents growth of rto,
673 * but also it limits too fast rto decreases,
674 * happening in pure Eifel.
679 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
681 tp
->mdev
+= m
; /* mdev = 3/4 mdev + 1/4 new */
682 if (tp
->mdev
> tp
->mdev_max
) {
683 tp
->mdev_max
= tp
->mdev
;
684 if (tp
->mdev_max
> tp
->rttvar
)
685 tp
->rttvar
= tp
->mdev_max
;
687 if (after(tp
->snd_una
, tp
->rtt_seq
)) {
688 if (tp
->mdev_max
< tp
->rttvar
)
689 tp
->rttvar
-= (tp
->rttvar
- tp
->mdev_max
) >> 2;
690 tp
->rtt_seq
= tp
->snd_nxt
;
691 tp
->mdev_max
= tcp_rto_min(sk
);
694 /* no previous measure. */
695 tp
->srtt
= m
<< 3; /* take the measured time to be rtt */
696 tp
->mdev
= m
<< 1; /* make sure rto = 3*rtt */
697 tp
->mdev_max
= tp
->rttvar
= max(tp
->mdev
, tcp_rto_min(sk
));
698 tp
->rtt_seq
= tp
->snd_nxt
;
702 /* Calculate rto without backoff. This is the second half of Van Jacobson's
703 * routine referred to above.
705 void tcp_set_rto(struct sock
*sk
)
707 const struct tcp_sock
*tp
= tcp_sk(sk
);
708 /* Old crap is replaced with new one. 8)
711 * 1. If rtt variance happened to be less 50msec, it is hallucination.
712 * It cannot be less due to utterly erratic ACK generation made
713 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
714 * to do with delayed acks, because at cwnd>2 true delack timeout
715 * is invisible. Actually, Linux-2.4 also generates erratic
716 * ACKs in some circumstances.
718 inet_csk(sk
)->icsk_rto
= __tcp_set_rto(tp
);
720 /* 2. Fixups made earlier cannot be right.
721 * If we do not estimate RTO correctly without them,
722 * all the algo is pure shit and should be replaced
723 * with correct one. It is exactly, which we pretend to do.
726 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
727 * guarantees that rto is higher.
732 __u32
tcp_init_cwnd(const struct tcp_sock
*tp
, const struct dst_entry
*dst
)
734 __u32 cwnd
= (dst
? dst_metric(dst
, RTAX_INITCWND
) : 0);
737 cwnd
= TCP_INIT_CWND
;
738 return min_t(__u32
, cwnd
, tp
->snd_cwnd_clamp
);
742 * Packet counting of FACK is based on in-order assumptions, therefore TCP
743 * disables it when reordering is detected
745 void tcp_disable_fack(struct tcp_sock
*tp
)
747 /* RFC3517 uses different metric in lost marker => reset on change */
749 tp
->lost_skb_hint
= NULL
;
750 tp
->rx_opt
.sack_ok
&= ~TCP_FACK_ENABLED
;
753 /* Take a notice that peer is sending D-SACKs */
754 static void tcp_dsack_seen(struct tcp_sock
*tp
)
756 tp
->rx_opt
.sack_ok
|= TCP_DSACK_SEEN
;
759 static void tcp_update_reordering(struct sock
*sk
, const int metric
,
762 struct tcp_sock
*tp
= tcp_sk(sk
);
763 if (metric
> tp
->reordering
) {
766 tp
->reordering
= min(TCP_MAX_REORDERING
, metric
);
768 /* This exciting event is worth to be remembered. 8) */
770 mib_idx
= LINUX_MIB_TCPTSREORDER
;
771 else if (tcp_is_reno(tp
))
772 mib_idx
= LINUX_MIB_TCPRENOREORDER
;
773 else if (tcp_is_fack(tp
))
774 mib_idx
= LINUX_MIB_TCPFACKREORDER
;
776 mib_idx
= LINUX_MIB_TCPSACKREORDER
;
778 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
779 #if FASTRETRANS_DEBUG > 1
780 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
781 tp
->rx_opt
.sack_ok
, inet_csk(sk
)->icsk_ca_state
,
785 tp
->undo_marker
? tp
->undo_retrans
: 0);
787 tcp_disable_fack(tp
);
791 tcp_disable_early_retrans(tp
);
794 /* This must be called before lost_out is incremented */
795 static void tcp_verify_retransmit_hint(struct tcp_sock
*tp
, struct sk_buff
*skb
)
797 if ((tp
->retransmit_skb_hint
== NULL
) ||
798 before(TCP_SKB_CB(skb
)->seq
,
799 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
800 tp
->retransmit_skb_hint
= skb
;
803 after(TCP_SKB_CB(skb
)->end_seq
, tp
->retransmit_high
))
804 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
807 static void tcp_skb_mark_lost(struct tcp_sock
*tp
, struct sk_buff
*skb
)
809 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
810 tcp_verify_retransmit_hint(tp
, skb
);
812 tp
->lost_out
+= tcp_skb_pcount(skb
);
813 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
817 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock
*tp
,
820 tcp_verify_retransmit_hint(tp
, skb
);
822 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
823 tp
->lost_out
+= tcp_skb_pcount(skb
);
824 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
828 /* This procedure tags the retransmission queue when SACKs arrive.
830 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
831 * Packets in queue with these bits set are counted in variables
832 * sacked_out, retrans_out and lost_out, correspondingly.
834 * Valid combinations are:
835 * Tag InFlight Description
836 * 0 1 - orig segment is in flight.
837 * S 0 - nothing flies, orig reached receiver.
838 * L 0 - nothing flies, orig lost by net.
839 * R 2 - both orig and retransmit are in flight.
840 * L|R 1 - orig is lost, retransmit is in flight.
841 * S|R 1 - orig reached receiver, retrans is still in flight.
842 * (L|S|R is logically valid, it could occur when L|R is sacked,
843 * but it is equivalent to plain S and code short-curcuits it to S.
844 * L|S is logically invalid, it would mean -1 packet in flight 8))
846 * These 6 states form finite state machine, controlled by the following events:
847 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
848 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
849 * 3. Loss detection event of two flavors:
850 * A. Scoreboard estimator decided the packet is lost.
851 * A'. Reno "three dupacks" marks head of queue lost.
852 * A''. Its FACK modification, head until snd.fack is lost.
853 * B. SACK arrives sacking SND.NXT at the moment, when the
854 * segment was retransmitted.
855 * 4. D-SACK added new rule: D-SACK changes any tag to S.
857 * It is pleasant to note, that state diagram turns out to be commutative,
858 * so that we are allowed not to be bothered by order of our actions,
859 * when multiple events arrive simultaneously. (see the function below).
861 * Reordering detection.
862 * --------------------
863 * Reordering metric is maximal distance, which a packet can be displaced
864 * in packet stream. With SACKs we can estimate it:
866 * 1. SACK fills old hole and the corresponding segment was not
867 * ever retransmitted -> reordering. Alas, we cannot use it
868 * when segment was retransmitted.
869 * 2. The last flaw is solved with D-SACK. D-SACK arrives
870 * for retransmitted and already SACKed segment -> reordering..
871 * Both of these heuristics are not used in Loss state, when we cannot
872 * account for retransmits accurately.
874 * SACK block validation.
875 * ----------------------
877 * SACK block range validation checks that the received SACK block fits to
878 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
879 * Note that SND.UNA is not included to the range though being valid because
880 * it means that the receiver is rather inconsistent with itself reporting
881 * SACK reneging when it should advance SND.UNA. Such SACK block this is
882 * perfectly valid, however, in light of RFC2018 which explicitly states
883 * that "SACK block MUST reflect the newest segment. Even if the newest
884 * segment is going to be discarded ...", not that it looks very clever
885 * in case of head skb. Due to potentional receiver driven attacks, we
886 * choose to avoid immediate execution of a walk in write queue due to
887 * reneging and defer head skb's loss recovery to standard loss recovery
888 * procedure that will eventually trigger (nothing forbids us doing this).
890 * Implements also blockage to start_seq wrap-around. Problem lies in the
891 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
892 * there's no guarantee that it will be before snd_nxt (n). The problem
893 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
896 * <- outs wnd -> <- wrapzone ->
897 * u e n u_w e_w s n_w
899 * |<------------+------+----- TCP seqno space --------------+---------->|
900 * ...-- <2^31 ->| |<--------...
901 * ...---- >2^31 ------>| |<--------...
903 * Current code wouldn't be vulnerable but it's better still to discard such
904 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
905 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
906 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
907 * equal to the ideal case (infinite seqno space without wrap caused issues).
909 * With D-SACK the lower bound is extended to cover sequence space below
910 * SND.UNA down to undo_marker, which is the last point of interest. Yet
911 * again, D-SACK block must not to go across snd_una (for the same reason as
912 * for the normal SACK blocks, explained above). But there all simplicity
913 * ends, TCP might receive valid D-SACKs below that. As long as they reside
914 * fully below undo_marker they do not affect behavior in anyway and can
915 * therefore be safely ignored. In rare cases (which are more or less
916 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
917 * fragmentation and packet reordering past skb's retransmission. To consider
918 * them correctly, the acceptable range must be extended even more though
919 * the exact amount is rather hard to quantify. However, tp->max_window can
920 * be used as an exaggerated estimate.
922 static bool tcp_is_sackblock_valid(struct tcp_sock
*tp
, bool is_dsack
,
923 u32 start_seq
, u32 end_seq
)
925 /* Too far in future, or reversed (interpretation is ambiguous) */
926 if (after(end_seq
, tp
->snd_nxt
) || !before(start_seq
, end_seq
))
929 /* Nasty start_seq wrap-around check (see comments above) */
930 if (!before(start_seq
, tp
->snd_nxt
))
933 /* In outstanding window? ...This is valid exit for D-SACKs too.
934 * start_seq == snd_una is non-sensical (see comments above)
936 if (after(start_seq
, tp
->snd_una
))
939 if (!is_dsack
|| !tp
->undo_marker
)
942 /* ...Then it's D-SACK, and must reside below snd_una completely */
943 if (after(end_seq
, tp
->snd_una
))
946 if (!before(start_seq
, tp
->undo_marker
))
950 if (!after(end_seq
, tp
->undo_marker
))
953 /* Undo_marker boundary crossing (overestimates a lot). Known already:
954 * start_seq < undo_marker and end_seq >= undo_marker.
956 return !before(start_seq
, end_seq
- tp
->max_window
);
959 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
960 * Event "B". Later note: FACK people cheated me again 8), we have to account
961 * for reordering! Ugly, but should help.
963 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
964 * less than what is now known to be received by the other end (derived from
965 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
966 * retransmitted skbs to avoid some costly processing per ACKs.
968 static void tcp_mark_lost_retrans(struct sock
*sk
)
970 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
971 struct tcp_sock
*tp
= tcp_sk(sk
);
974 u32 new_low_seq
= tp
->snd_nxt
;
975 u32 received_upto
= tcp_highest_sack_seq(tp
);
977 if (!tcp_is_fack(tp
) || !tp
->retrans_out
||
978 !after(received_upto
, tp
->lost_retrans_low
) ||
979 icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
982 tcp_for_write_queue(skb
, sk
) {
983 u32 ack_seq
= TCP_SKB_CB(skb
)->ack_seq
;
985 if (skb
== tcp_send_head(sk
))
987 if (cnt
== tp
->retrans_out
)
989 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
992 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
))
995 /* TODO: We would like to get rid of tcp_is_fack(tp) only
996 * constraint here (see above) but figuring out that at
997 * least tp->reordering SACK blocks reside between ack_seq
998 * and received_upto is not easy task to do cheaply with
999 * the available datastructures.
1001 * Whether FACK should check here for tp->reordering segs
1002 * in-between one could argue for either way (it would be
1003 * rather simple to implement as we could count fack_count
1004 * during the walk and do tp->fackets_out - fack_count).
1006 if (after(received_upto
, ack_seq
)) {
1007 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1008 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1010 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
1011 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSTRETRANSMIT
);
1013 if (before(ack_seq
, new_low_seq
))
1014 new_low_seq
= ack_seq
;
1015 cnt
+= tcp_skb_pcount(skb
);
1019 if (tp
->retrans_out
)
1020 tp
->lost_retrans_low
= new_low_seq
;
1023 static bool tcp_check_dsack(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1024 struct tcp_sack_block_wire
*sp
, int num_sacks
,
1027 struct tcp_sock
*tp
= tcp_sk(sk
);
1028 u32 start_seq_0
= get_unaligned_be32(&sp
[0].start_seq
);
1029 u32 end_seq_0
= get_unaligned_be32(&sp
[0].end_seq
);
1030 bool dup_sack
= false;
1032 if (before(start_seq_0
, TCP_SKB_CB(ack_skb
)->ack_seq
)) {
1035 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPDSACKRECV
);
1036 } else if (num_sacks
> 1) {
1037 u32 end_seq_1
= get_unaligned_be32(&sp
[1].end_seq
);
1038 u32 start_seq_1
= get_unaligned_be32(&sp
[1].start_seq
);
1040 if (!after(end_seq_0
, end_seq_1
) &&
1041 !before(start_seq_0
, start_seq_1
)) {
1044 NET_INC_STATS_BH(sock_net(sk
),
1045 LINUX_MIB_TCPDSACKOFORECV
);
1049 /* D-SACK for already forgotten data... Do dumb counting. */
1050 if (dup_sack
&& tp
->undo_marker
&& tp
->undo_retrans
&&
1051 !after(end_seq_0
, prior_snd_una
) &&
1052 after(end_seq_0
, tp
->undo_marker
))
1058 struct tcp_sacktag_state
{
1064 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1065 * the incoming SACK may not exactly match but we can find smaller MSS
1066 * aligned portion of it that matches. Therefore we might need to fragment
1067 * which may fail and creates some hassle (caller must handle error case
1070 * FIXME: this could be merged to shift decision code
1072 static int tcp_match_skb_to_sack(struct sock
*sk
, struct sk_buff
*skb
,
1073 u32 start_seq
, u32 end_seq
)
1077 unsigned int pkt_len
;
1080 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1081 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1083 if (tcp_skb_pcount(skb
) > 1 && !in_sack
&&
1084 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
1085 mss
= tcp_skb_mss(skb
);
1086 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1089 pkt_len
= start_seq
- TCP_SKB_CB(skb
)->seq
;
1093 pkt_len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1098 /* Round if necessary so that SACKs cover only full MSSes
1099 * and/or the remaining small portion (if present)
1101 if (pkt_len
> mss
) {
1102 unsigned int new_len
= (pkt_len
/ mss
) * mss
;
1103 if (!in_sack
&& new_len
< pkt_len
) {
1105 if (new_len
> skb
->len
)
1110 err
= tcp_fragment(sk
, skb
, pkt_len
, mss
);
1118 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1119 static u8
tcp_sacktag_one(struct sock
*sk
,
1120 struct tcp_sacktag_state
*state
, u8 sacked
,
1121 u32 start_seq
, u32 end_seq
,
1122 bool dup_sack
, int pcount
)
1124 struct tcp_sock
*tp
= tcp_sk(sk
);
1125 int fack_count
= state
->fack_count
;
1127 /* Account D-SACK for retransmitted packet. */
1128 if (dup_sack
&& (sacked
& TCPCB_RETRANS
)) {
1129 if (tp
->undo_marker
&& tp
->undo_retrans
&&
1130 after(end_seq
, tp
->undo_marker
))
1132 if (sacked
& TCPCB_SACKED_ACKED
)
1133 state
->reord
= min(fack_count
, state
->reord
);
1136 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1137 if (!after(end_seq
, tp
->snd_una
))
1140 if (!(sacked
& TCPCB_SACKED_ACKED
)) {
1141 if (sacked
& TCPCB_SACKED_RETRANS
) {
1142 /* If the segment is not tagged as lost,
1143 * we do not clear RETRANS, believing
1144 * that retransmission is still in flight.
1146 if (sacked
& TCPCB_LOST
) {
1147 sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1148 tp
->lost_out
-= pcount
;
1149 tp
->retrans_out
-= pcount
;
1152 if (!(sacked
& TCPCB_RETRANS
)) {
1153 /* New sack for not retransmitted frame,
1154 * which was in hole. It is reordering.
1156 if (before(start_seq
,
1157 tcp_highest_sack_seq(tp
)))
1158 state
->reord
= min(fack_count
,
1160 if (!after(end_seq
, tp
->high_seq
))
1161 state
->flag
|= FLAG_ORIG_SACK_ACKED
;
1164 if (sacked
& TCPCB_LOST
) {
1165 sacked
&= ~TCPCB_LOST
;
1166 tp
->lost_out
-= pcount
;
1170 sacked
|= TCPCB_SACKED_ACKED
;
1171 state
->flag
|= FLAG_DATA_SACKED
;
1172 tp
->sacked_out
+= pcount
;
1174 fack_count
+= pcount
;
1176 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1177 if (!tcp_is_fack(tp
) && (tp
->lost_skb_hint
!= NULL
) &&
1178 before(start_seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
))
1179 tp
->lost_cnt_hint
+= pcount
;
1181 if (fack_count
> tp
->fackets_out
)
1182 tp
->fackets_out
= fack_count
;
1185 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1186 * frames and clear it. undo_retrans is decreased above, L|R frames
1187 * are accounted above as well.
1189 if (dup_sack
&& (sacked
& TCPCB_SACKED_RETRANS
)) {
1190 sacked
&= ~TCPCB_SACKED_RETRANS
;
1191 tp
->retrans_out
-= pcount
;
1197 /* Shift newly-SACKed bytes from this skb to the immediately previous
1198 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1200 static bool tcp_shifted_skb(struct sock
*sk
, struct sk_buff
*skb
,
1201 struct tcp_sacktag_state
*state
,
1202 unsigned int pcount
, int shifted
, int mss
,
1205 struct tcp_sock
*tp
= tcp_sk(sk
);
1206 struct sk_buff
*prev
= tcp_write_queue_prev(sk
, skb
);
1207 u32 start_seq
= TCP_SKB_CB(skb
)->seq
; /* start of newly-SACKed */
1208 u32 end_seq
= start_seq
+ shifted
; /* end of newly-SACKed */
1212 /* Adjust counters and hints for the newly sacked sequence
1213 * range but discard the return value since prev is already
1214 * marked. We must tag the range first because the seq
1215 * advancement below implicitly advances
1216 * tcp_highest_sack_seq() when skb is highest_sack.
1218 tcp_sacktag_one(sk
, state
, TCP_SKB_CB(skb
)->sacked
,
1219 start_seq
, end_seq
, dup_sack
, pcount
);
1221 if (skb
== tp
->lost_skb_hint
)
1222 tp
->lost_cnt_hint
+= pcount
;
1224 TCP_SKB_CB(prev
)->end_seq
+= shifted
;
1225 TCP_SKB_CB(skb
)->seq
+= shifted
;
1227 skb_shinfo(prev
)->gso_segs
+= pcount
;
1228 BUG_ON(skb_shinfo(skb
)->gso_segs
< pcount
);
1229 skb_shinfo(skb
)->gso_segs
-= pcount
;
1231 /* When we're adding to gso_segs == 1, gso_size will be zero,
1232 * in theory this shouldn't be necessary but as long as DSACK
1233 * code can come after this skb later on it's better to keep
1234 * setting gso_size to something.
1236 if (!skb_shinfo(prev
)->gso_size
) {
1237 skb_shinfo(prev
)->gso_size
= mss
;
1238 skb_shinfo(prev
)->gso_type
= sk
->sk_gso_type
;
1241 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1242 if (skb_shinfo(skb
)->gso_segs
<= 1) {
1243 skb_shinfo(skb
)->gso_size
= 0;
1244 skb_shinfo(skb
)->gso_type
= 0;
1247 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1248 TCP_SKB_CB(prev
)->sacked
|= (TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
);
1251 BUG_ON(!tcp_skb_pcount(skb
));
1252 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKSHIFTED
);
1256 /* Whole SKB was eaten :-) */
1258 if (skb
== tp
->retransmit_skb_hint
)
1259 tp
->retransmit_skb_hint
= prev
;
1260 if (skb
== tp
->scoreboard_skb_hint
)
1261 tp
->scoreboard_skb_hint
= prev
;
1262 if (skb
== tp
->lost_skb_hint
) {
1263 tp
->lost_skb_hint
= prev
;
1264 tp
->lost_cnt_hint
-= tcp_skb_pcount(prev
);
1267 TCP_SKB_CB(skb
)->tcp_flags
|= TCP_SKB_CB(prev
)->tcp_flags
;
1268 if (skb
== tcp_highest_sack(sk
))
1269 tcp_advance_highest_sack(sk
, skb
);
1271 tcp_unlink_write_queue(skb
, sk
);
1272 sk_wmem_free_skb(sk
, skb
);
1274 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKMERGED
);
1279 /* I wish gso_size would have a bit more sane initialization than
1280 * something-or-zero which complicates things
1282 static int tcp_skb_seglen(const struct sk_buff
*skb
)
1284 return tcp_skb_pcount(skb
) == 1 ? skb
->len
: tcp_skb_mss(skb
);
1287 /* Shifting pages past head area doesn't work */
1288 static int skb_can_shift(const struct sk_buff
*skb
)
1290 return !skb_headlen(skb
) && skb_is_nonlinear(skb
);
1293 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1296 static struct sk_buff
*tcp_shift_skb_data(struct sock
*sk
, struct sk_buff
*skb
,
1297 struct tcp_sacktag_state
*state
,
1298 u32 start_seq
, u32 end_seq
,
1301 struct tcp_sock
*tp
= tcp_sk(sk
);
1302 struct sk_buff
*prev
;
1308 if (!sk_can_gso(sk
))
1311 /* Normally R but no L won't result in plain S */
1313 (TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_RETRANS
)) == TCPCB_SACKED_RETRANS
)
1315 if (!skb_can_shift(skb
))
1317 /* This frame is about to be dropped (was ACKed). */
1318 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1321 /* Can only happen with delayed DSACK + discard craziness */
1322 if (unlikely(skb
== tcp_write_queue_head(sk
)))
1324 prev
= tcp_write_queue_prev(sk
, skb
);
1326 if ((TCP_SKB_CB(prev
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
)
1329 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1330 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1334 pcount
= tcp_skb_pcount(skb
);
1335 mss
= tcp_skb_seglen(skb
);
1337 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1338 * drop this restriction as unnecessary
1340 if (mss
!= tcp_skb_seglen(prev
))
1343 if (!after(TCP_SKB_CB(skb
)->end_seq
, start_seq
))
1345 /* CHECKME: This is non-MSS split case only?, this will
1346 * cause skipped skbs due to advancing loop btw, original
1347 * has that feature too
1349 if (tcp_skb_pcount(skb
) <= 1)
1352 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1354 /* TODO: head merge to next could be attempted here
1355 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1356 * though it might not be worth of the additional hassle
1358 * ...we can probably just fallback to what was done
1359 * previously. We could try merging non-SACKed ones
1360 * as well but it probably isn't going to buy off
1361 * because later SACKs might again split them, and
1362 * it would make skb timestamp tracking considerably
1368 len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1370 BUG_ON(len
> skb
->len
);
1372 /* MSS boundaries should be honoured or else pcount will
1373 * severely break even though it makes things bit trickier.
1374 * Optimize common case to avoid most of the divides
1376 mss
= tcp_skb_mss(skb
);
1378 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1379 * drop this restriction as unnecessary
1381 if (mss
!= tcp_skb_seglen(prev
))
1386 } else if (len
< mss
) {
1394 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1395 if (!after(TCP_SKB_CB(skb
)->seq
+ len
, tp
->snd_una
))
1398 if (!skb_shift(prev
, skb
, len
))
1400 if (!tcp_shifted_skb(sk
, skb
, state
, pcount
, len
, mss
, dup_sack
))
1403 /* Hole filled allows collapsing with the next as well, this is very
1404 * useful when hole on every nth skb pattern happens
1406 if (prev
== tcp_write_queue_tail(sk
))
1408 skb
= tcp_write_queue_next(sk
, prev
);
1410 if (!skb_can_shift(skb
) ||
1411 (skb
== tcp_send_head(sk
)) ||
1412 ((TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
) ||
1413 (mss
!= tcp_skb_seglen(skb
)))
1417 if (skb_shift(prev
, skb
, len
)) {
1418 pcount
+= tcp_skb_pcount(skb
);
1419 tcp_shifted_skb(sk
, skb
, state
, tcp_skb_pcount(skb
), len
, mss
, 0);
1423 state
->fack_count
+= pcount
;
1430 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKSHIFTFALLBACK
);
1434 static struct sk_buff
*tcp_sacktag_walk(struct sk_buff
*skb
, struct sock
*sk
,
1435 struct tcp_sack_block
*next_dup
,
1436 struct tcp_sacktag_state
*state
,
1437 u32 start_seq
, u32 end_seq
,
1440 struct tcp_sock
*tp
= tcp_sk(sk
);
1441 struct sk_buff
*tmp
;
1443 tcp_for_write_queue_from(skb
, sk
) {
1445 bool dup_sack
= dup_sack_in
;
1447 if (skb
== tcp_send_head(sk
))
1450 /* queue is in-order => we can short-circuit the walk early */
1451 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1454 if ((next_dup
!= NULL
) &&
1455 before(TCP_SKB_CB(skb
)->seq
, next_dup
->end_seq
)) {
1456 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1457 next_dup
->start_seq
,
1463 /* skb reference here is a bit tricky to get right, since
1464 * shifting can eat and free both this skb and the next,
1465 * so not even _safe variant of the loop is enough.
1468 tmp
= tcp_shift_skb_data(sk
, skb
, state
,
1469 start_seq
, end_seq
, dup_sack
);
1478 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1484 if (unlikely(in_sack
< 0))
1488 TCP_SKB_CB(skb
)->sacked
=
1491 TCP_SKB_CB(skb
)->sacked
,
1492 TCP_SKB_CB(skb
)->seq
,
1493 TCP_SKB_CB(skb
)->end_seq
,
1495 tcp_skb_pcount(skb
));
1497 if (!before(TCP_SKB_CB(skb
)->seq
,
1498 tcp_highest_sack_seq(tp
)))
1499 tcp_advance_highest_sack(sk
, skb
);
1502 state
->fack_count
+= tcp_skb_pcount(skb
);
1507 /* Avoid all extra work that is being done by sacktag while walking in
1510 static struct sk_buff
*tcp_sacktag_skip(struct sk_buff
*skb
, struct sock
*sk
,
1511 struct tcp_sacktag_state
*state
,
1514 tcp_for_write_queue_from(skb
, sk
) {
1515 if (skb
== tcp_send_head(sk
))
1518 if (after(TCP_SKB_CB(skb
)->end_seq
, skip_to_seq
))
1521 state
->fack_count
+= tcp_skb_pcount(skb
);
1526 static struct sk_buff
*tcp_maybe_skipping_dsack(struct sk_buff
*skb
,
1528 struct tcp_sack_block
*next_dup
,
1529 struct tcp_sacktag_state
*state
,
1532 if (next_dup
== NULL
)
1535 if (before(next_dup
->start_seq
, skip_to_seq
)) {
1536 skb
= tcp_sacktag_skip(skb
, sk
, state
, next_dup
->start_seq
);
1537 skb
= tcp_sacktag_walk(skb
, sk
, NULL
, state
,
1538 next_dup
->start_seq
, next_dup
->end_seq
,
1545 static int tcp_sack_cache_ok(const struct tcp_sock
*tp
, const struct tcp_sack_block
*cache
)
1547 return cache
< tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1551 tcp_sacktag_write_queue(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1554 struct tcp_sock
*tp
= tcp_sk(sk
);
1555 const unsigned char *ptr
= (skb_transport_header(ack_skb
) +
1556 TCP_SKB_CB(ack_skb
)->sacked
);
1557 struct tcp_sack_block_wire
*sp_wire
= (struct tcp_sack_block_wire
*)(ptr
+2);
1558 struct tcp_sack_block sp
[TCP_NUM_SACKS
];
1559 struct tcp_sack_block
*cache
;
1560 struct tcp_sacktag_state state
;
1561 struct sk_buff
*skb
;
1562 int num_sacks
= min(TCP_NUM_SACKS
, (ptr
[1] - TCPOLEN_SACK_BASE
) >> 3);
1564 bool found_dup_sack
= false;
1566 int first_sack_index
;
1569 state
.reord
= tp
->packets_out
;
1571 if (!tp
->sacked_out
) {
1572 if (WARN_ON(tp
->fackets_out
))
1573 tp
->fackets_out
= 0;
1574 tcp_highest_sack_reset(sk
);
1577 found_dup_sack
= tcp_check_dsack(sk
, ack_skb
, sp_wire
,
1578 num_sacks
, prior_snd_una
);
1580 state
.flag
|= FLAG_DSACKING_ACK
;
1582 /* Eliminate too old ACKs, but take into
1583 * account more or less fresh ones, they can
1584 * contain valid SACK info.
1586 if (before(TCP_SKB_CB(ack_skb
)->ack_seq
, prior_snd_una
- tp
->max_window
))
1589 if (!tp
->packets_out
)
1593 first_sack_index
= 0;
1594 for (i
= 0; i
< num_sacks
; i
++) {
1595 bool dup_sack
= !i
&& found_dup_sack
;
1597 sp
[used_sacks
].start_seq
= get_unaligned_be32(&sp_wire
[i
].start_seq
);
1598 sp
[used_sacks
].end_seq
= get_unaligned_be32(&sp_wire
[i
].end_seq
);
1600 if (!tcp_is_sackblock_valid(tp
, dup_sack
,
1601 sp
[used_sacks
].start_seq
,
1602 sp
[used_sacks
].end_seq
)) {
1606 if (!tp
->undo_marker
)
1607 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDNOUNDO
;
1609 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDOLD
;
1611 /* Don't count olds caused by ACK reordering */
1612 if ((TCP_SKB_CB(ack_skb
)->ack_seq
!= tp
->snd_una
) &&
1613 !after(sp
[used_sacks
].end_seq
, tp
->snd_una
))
1615 mib_idx
= LINUX_MIB_TCPSACKDISCARD
;
1618 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
1620 first_sack_index
= -1;
1624 /* Ignore very old stuff early */
1625 if (!after(sp
[used_sacks
].end_seq
, prior_snd_una
))
1631 /* order SACK blocks to allow in order walk of the retrans queue */
1632 for (i
= used_sacks
- 1; i
> 0; i
--) {
1633 for (j
= 0; j
< i
; j
++) {
1634 if (after(sp
[j
].start_seq
, sp
[j
+ 1].start_seq
)) {
1635 swap(sp
[j
], sp
[j
+ 1]);
1637 /* Track where the first SACK block goes to */
1638 if (j
== first_sack_index
)
1639 first_sack_index
= j
+ 1;
1644 skb
= tcp_write_queue_head(sk
);
1645 state
.fack_count
= 0;
1648 if (!tp
->sacked_out
) {
1649 /* It's already past, so skip checking against it */
1650 cache
= tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1652 cache
= tp
->recv_sack_cache
;
1653 /* Skip empty blocks in at head of the cache */
1654 while (tcp_sack_cache_ok(tp
, cache
) && !cache
->start_seq
&&
1659 while (i
< used_sacks
) {
1660 u32 start_seq
= sp
[i
].start_seq
;
1661 u32 end_seq
= sp
[i
].end_seq
;
1662 bool dup_sack
= (found_dup_sack
&& (i
== first_sack_index
));
1663 struct tcp_sack_block
*next_dup
= NULL
;
1665 if (found_dup_sack
&& ((i
+ 1) == first_sack_index
))
1666 next_dup
= &sp
[i
+ 1];
1668 /* Skip too early cached blocks */
1669 while (tcp_sack_cache_ok(tp
, cache
) &&
1670 !before(start_seq
, cache
->end_seq
))
1673 /* Can skip some work by looking recv_sack_cache? */
1674 if (tcp_sack_cache_ok(tp
, cache
) && !dup_sack
&&
1675 after(end_seq
, cache
->start_seq
)) {
1678 if (before(start_seq
, cache
->start_seq
)) {
1679 skb
= tcp_sacktag_skip(skb
, sk
, &state
,
1681 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
,
1688 /* Rest of the block already fully processed? */
1689 if (!after(end_seq
, cache
->end_seq
))
1692 skb
= tcp_maybe_skipping_dsack(skb
, sk
, next_dup
,
1696 /* ...tail remains todo... */
1697 if (tcp_highest_sack_seq(tp
) == cache
->end_seq
) {
1698 /* ...but better entrypoint exists! */
1699 skb
= tcp_highest_sack(sk
);
1702 state
.fack_count
= tp
->fackets_out
;
1707 skb
= tcp_sacktag_skip(skb
, sk
, &state
, cache
->end_seq
);
1708 /* Check overlap against next cached too (past this one already) */
1713 if (!before(start_seq
, tcp_highest_sack_seq(tp
))) {
1714 skb
= tcp_highest_sack(sk
);
1717 state
.fack_count
= tp
->fackets_out
;
1719 skb
= tcp_sacktag_skip(skb
, sk
, &state
, start_seq
);
1722 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
, &state
,
1723 start_seq
, end_seq
, dup_sack
);
1729 /* Clear the head of the cache sack blocks so we can skip it next time */
1730 for (i
= 0; i
< ARRAY_SIZE(tp
->recv_sack_cache
) - used_sacks
; i
++) {
1731 tp
->recv_sack_cache
[i
].start_seq
= 0;
1732 tp
->recv_sack_cache
[i
].end_seq
= 0;
1734 for (j
= 0; j
< used_sacks
; j
++)
1735 tp
->recv_sack_cache
[i
++] = sp
[j
];
1737 tcp_mark_lost_retrans(sk
);
1739 tcp_verify_left_out(tp
);
1741 if ((state
.reord
< tp
->fackets_out
) &&
1742 ((inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Loss
) || tp
->undo_marker
))
1743 tcp_update_reordering(sk
, tp
->fackets_out
- state
.reord
, 0);
1747 #if FASTRETRANS_DEBUG > 0
1748 WARN_ON((int)tp
->sacked_out
< 0);
1749 WARN_ON((int)tp
->lost_out
< 0);
1750 WARN_ON((int)tp
->retrans_out
< 0);
1751 WARN_ON((int)tcp_packets_in_flight(tp
) < 0);
1756 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1757 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1759 static bool tcp_limit_reno_sacked(struct tcp_sock
*tp
)
1763 holes
= max(tp
->lost_out
, 1U);
1764 holes
= min(holes
, tp
->packets_out
);
1766 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1767 tp
->sacked_out
= tp
->packets_out
- holes
;
1773 /* If we receive more dupacks than we expected counting segments
1774 * in assumption of absent reordering, interpret this as reordering.
1775 * The only another reason could be bug in receiver TCP.
1777 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1779 struct tcp_sock
*tp
= tcp_sk(sk
);
1780 if (tcp_limit_reno_sacked(tp
))
1781 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1784 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1786 static void tcp_add_reno_sack(struct sock
*sk
)
1788 struct tcp_sock
*tp
= tcp_sk(sk
);
1790 tcp_check_reno_reordering(sk
, 0);
1791 tcp_verify_left_out(tp
);
1794 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1796 static void tcp_remove_reno_sacks(struct sock
*sk
, int acked
)
1798 struct tcp_sock
*tp
= tcp_sk(sk
);
1801 /* One ACK acked hole. The rest eat duplicate ACKs. */
1802 if (acked
- 1 >= tp
->sacked_out
)
1805 tp
->sacked_out
-= acked
- 1;
1807 tcp_check_reno_reordering(sk
, acked
);
1808 tcp_verify_left_out(tp
);
1811 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1816 static void tcp_clear_retrans_partial(struct tcp_sock
*tp
)
1818 tp
->retrans_out
= 0;
1821 tp
->undo_marker
= 0;
1822 tp
->undo_retrans
= 0;
1825 void tcp_clear_retrans(struct tcp_sock
*tp
)
1827 tcp_clear_retrans_partial(tp
);
1829 tp
->fackets_out
= 0;
1833 /* Enter Loss state. If "how" is not zero, forget all SACK information
1834 * and reset tags completely, otherwise preserve SACKs. If receiver
1835 * dropped its ofo queue, we will know this due to reneging detection.
1837 void tcp_enter_loss(struct sock
*sk
, int how
)
1839 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1840 struct tcp_sock
*tp
= tcp_sk(sk
);
1841 struct sk_buff
*skb
;
1842 bool new_recovery
= false;
1844 /* Reduce ssthresh if it has not yet been made inside this window. */
1845 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
||
1846 !after(tp
->high_seq
, tp
->snd_una
) ||
1847 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
1848 new_recovery
= true;
1849 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1850 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1851 tcp_ca_event(sk
, CA_EVENT_LOSS
);
1854 tp
->snd_cwnd_cnt
= 0;
1855 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1857 tcp_clear_retrans_partial(tp
);
1859 if (tcp_is_reno(tp
))
1860 tcp_reset_reno_sack(tp
);
1862 tp
->undo_marker
= tp
->snd_una
;
1865 tp
->fackets_out
= 0;
1867 tcp_clear_all_retrans_hints(tp
);
1869 tcp_for_write_queue(skb
, sk
) {
1870 if (skb
== tcp_send_head(sk
))
1873 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
1874 tp
->undo_marker
= 0;
1875 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
1876 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) || how
) {
1877 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
1878 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1879 tp
->lost_out
+= tcp_skb_pcount(skb
);
1880 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
1883 tcp_verify_left_out(tp
);
1885 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1886 sysctl_tcp_reordering
);
1887 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1888 tp
->high_seq
= tp
->snd_nxt
;
1889 TCP_ECN_queue_cwr(tp
);
1891 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1892 * loss recovery is underway except recurring timeout(s) on
1893 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1895 tp
->frto
= sysctl_tcp_frto
&&
1896 (new_recovery
|| icsk
->icsk_retransmits
) &&
1897 !inet_csk(sk
)->icsk_mtup
.probe_size
;
1900 /* If ACK arrived pointing to a remembered SACK, it means that our
1901 * remembered SACKs do not reflect real state of receiver i.e.
1902 * receiver _host_ is heavily congested (or buggy).
1904 * Do processing similar to RTO timeout.
1906 static bool tcp_check_sack_reneging(struct sock
*sk
, int flag
)
1908 if (flag
& FLAG_SACK_RENEGING
) {
1909 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1910 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSACKRENEGING
);
1912 tcp_enter_loss(sk
, 1);
1913 icsk
->icsk_retransmits
++;
1914 tcp_retransmit_skb(sk
, tcp_write_queue_head(sk
));
1915 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1916 icsk
->icsk_rto
, TCP_RTO_MAX
);
1922 static inline int tcp_fackets_out(const struct tcp_sock
*tp
)
1924 return tcp_is_reno(tp
) ? tp
->sacked_out
+ 1 : tp
->fackets_out
;
1927 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
1928 * counter when SACK is enabled (without SACK, sacked_out is used for
1931 * Instead, with FACK TCP uses fackets_out that includes both SACKed
1932 * segments up to the highest received SACK block so far and holes in
1935 * With reordering, holes may still be in flight, so RFC3517 recovery
1936 * uses pure sacked_out (total number of SACKed segments) even though
1937 * it violates the RFC that uses duplicate ACKs, often these are equal
1938 * but when e.g. out-of-window ACKs or packet duplication occurs,
1939 * they differ. Since neither occurs due to loss, TCP should really
1942 static inline int tcp_dupack_heuristics(const struct tcp_sock
*tp
)
1944 return tcp_is_fack(tp
) ? tp
->fackets_out
: tp
->sacked_out
+ 1;
1947 static bool tcp_pause_early_retransmit(struct sock
*sk
, int flag
)
1949 struct tcp_sock
*tp
= tcp_sk(sk
);
1950 unsigned long delay
;
1952 /* Delay early retransmit and entering fast recovery for
1953 * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
1954 * available, or RTO is scheduled to fire first.
1956 if (sysctl_tcp_early_retrans
< 2 || sysctl_tcp_early_retrans
> 3 ||
1957 (flag
& FLAG_ECE
) || !tp
->srtt
)
1960 delay
= max_t(unsigned long, (tp
->srtt
>> 5), msecs_to_jiffies(2));
1961 if (!time_after(inet_csk(sk
)->icsk_timeout
, (jiffies
+ delay
)))
1964 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_EARLY_RETRANS
, delay
,
1969 static inline int tcp_skb_timedout(const struct sock
*sk
,
1970 const struct sk_buff
*skb
)
1972 return tcp_time_stamp
- TCP_SKB_CB(skb
)->when
> inet_csk(sk
)->icsk_rto
;
1975 static inline int tcp_head_timedout(const struct sock
*sk
)
1977 const struct tcp_sock
*tp
= tcp_sk(sk
);
1979 return tp
->packets_out
&&
1980 tcp_skb_timedout(sk
, tcp_write_queue_head(sk
));
1983 /* Linux NewReno/SACK/FACK/ECN state machine.
1984 * --------------------------------------
1986 * "Open" Normal state, no dubious events, fast path.
1987 * "Disorder" In all the respects it is "Open",
1988 * but requires a bit more attention. It is entered when
1989 * we see some SACKs or dupacks. It is split of "Open"
1990 * mainly to move some processing from fast path to slow one.
1991 * "CWR" CWND was reduced due to some Congestion Notification event.
1992 * It can be ECN, ICMP source quench, local device congestion.
1993 * "Recovery" CWND was reduced, we are fast-retransmitting.
1994 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
1996 * tcp_fastretrans_alert() is entered:
1997 * - each incoming ACK, if state is not "Open"
1998 * - when arrived ACK is unusual, namely:
2003 * Counting packets in flight is pretty simple.
2005 * in_flight = packets_out - left_out + retrans_out
2007 * packets_out is SND.NXT-SND.UNA counted in packets.
2009 * retrans_out is number of retransmitted segments.
2011 * left_out is number of segments left network, but not ACKed yet.
2013 * left_out = sacked_out + lost_out
2015 * sacked_out: Packets, which arrived to receiver out of order
2016 * and hence not ACKed. With SACKs this number is simply
2017 * amount of SACKed data. Even without SACKs
2018 * it is easy to give pretty reliable estimate of this number,
2019 * counting duplicate ACKs.
2021 * lost_out: Packets lost by network. TCP has no explicit
2022 * "loss notification" feedback from network (for now).
2023 * It means that this number can be only _guessed_.
2024 * Actually, it is the heuristics to predict lossage that
2025 * distinguishes different algorithms.
2027 * F.e. after RTO, when all the queue is considered as lost,
2028 * lost_out = packets_out and in_flight = retrans_out.
2030 * Essentially, we have now two algorithms counting
2033 * FACK: It is the simplest heuristics. As soon as we decided
2034 * that something is lost, we decide that _all_ not SACKed
2035 * packets until the most forward SACK are lost. I.e.
2036 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2037 * It is absolutely correct estimate, if network does not reorder
2038 * packets. And it loses any connection to reality when reordering
2039 * takes place. We use FACK by default until reordering
2040 * is suspected on the path to this destination.
2042 * NewReno: when Recovery is entered, we assume that one segment
2043 * is lost (classic Reno). While we are in Recovery and
2044 * a partial ACK arrives, we assume that one more packet
2045 * is lost (NewReno). This heuristics are the same in NewReno
2048 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2049 * deflation etc. CWND is real congestion window, never inflated, changes
2050 * only according to classic VJ rules.
2052 * Really tricky (and requiring careful tuning) part of algorithm
2053 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2054 * The first determines the moment _when_ we should reduce CWND and,
2055 * hence, slow down forward transmission. In fact, it determines the moment
2056 * when we decide that hole is caused by loss, rather than by a reorder.
2058 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2059 * holes, caused by lost packets.
2061 * And the most logically complicated part of algorithm is undo
2062 * heuristics. We detect false retransmits due to both too early
2063 * fast retransmit (reordering) and underestimated RTO, analyzing
2064 * timestamps and D-SACKs. When we detect that some segments were
2065 * retransmitted by mistake and CWND reduction was wrong, we undo
2066 * window reduction and abort recovery phase. This logic is hidden
2067 * inside several functions named tcp_try_undo_<something>.
2070 /* This function decides, when we should leave Disordered state
2071 * and enter Recovery phase, reducing congestion window.
2073 * Main question: may we further continue forward transmission
2074 * with the same cwnd?
2076 static bool tcp_time_to_recover(struct sock
*sk
, int flag
)
2078 struct tcp_sock
*tp
= tcp_sk(sk
);
2081 /* Trick#1: The loss is proven. */
2085 /* Not-A-Trick#2 : Classic rule... */
2086 if (tcp_dupack_heuristics(tp
) > tp
->reordering
)
2089 /* Trick#3 : when we use RFC2988 timer restart, fast
2090 * retransmit can be triggered by timeout of queue head.
2092 if (tcp_is_fack(tp
) && tcp_head_timedout(sk
))
2095 /* Trick#4: It is still not OK... But will it be useful to delay
2098 packets_out
= tp
->packets_out
;
2099 if (packets_out
<= tp
->reordering
&&
2100 tp
->sacked_out
>= max_t(__u32
, packets_out
/2, sysctl_tcp_reordering
) &&
2101 !tcp_may_send_now(sk
)) {
2102 /* We have nothing to send. This connection is limited
2103 * either by receiver window or by application.
2108 /* If a thin stream is detected, retransmit after first
2109 * received dupack. Employ only if SACK is supported in order
2110 * to avoid possible corner-case series of spurious retransmissions
2111 * Use only if there are no unsent data.
2113 if ((tp
->thin_dupack
|| sysctl_tcp_thin_dupack
) &&
2114 tcp_stream_is_thin(tp
) && tcp_dupack_heuristics(tp
) > 1 &&
2115 tcp_is_sack(tp
) && !tcp_send_head(sk
))
2118 /* Trick#6: TCP early retransmit, per RFC5827. To avoid spurious
2119 * retransmissions due to small network reorderings, we implement
2120 * Mitigation A.3 in the RFC and delay the retransmission for a short
2121 * interval if appropriate.
2123 if (tp
->do_early_retrans
&& !tp
->retrans_out
&& tp
->sacked_out
&&
2124 (tp
->packets_out
>= (tp
->sacked_out
+ 1) && tp
->packets_out
< 4) &&
2125 !tcp_may_send_now(sk
))
2126 return !tcp_pause_early_retransmit(sk
, flag
);
2131 /* New heuristics: it is possible only after we switched to restart timer
2132 * each time when something is ACKed. Hence, we can detect timed out packets
2133 * during fast retransmit without falling to slow start.
2135 * Usefulness of this as is very questionable, since we should know which of
2136 * the segments is the next to timeout which is relatively expensive to find
2137 * in general case unless we add some data structure just for that. The
2138 * current approach certainly won't find the right one too often and when it
2139 * finally does find _something_ it usually marks large part of the window
2140 * right away (because a retransmission with a larger timestamp blocks the
2141 * loop from advancing). -ij
2143 static void tcp_timeout_skbs(struct sock
*sk
)
2145 struct tcp_sock
*tp
= tcp_sk(sk
);
2146 struct sk_buff
*skb
;
2148 if (!tcp_is_fack(tp
) || !tcp_head_timedout(sk
))
2151 skb
= tp
->scoreboard_skb_hint
;
2152 if (tp
->scoreboard_skb_hint
== NULL
)
2153 skb
= tcp_write_queue_head(sk
);
2155 tcp_for_write_queue_from(skb
, sk
) {
2156 if (skb
== tcp_send_head(sk
))
2158 if (!tcp_skb_timedout(sk
, skb
))
2161 tcp_skb_mark_lost(tp
, skb
);
2164 tp
->scoreboard_skb_hint
= skb
;
2166 tcp_verify_left_out(tp
);
2169 /* Detect loss in event "A" above by marking head of queue up as lost.
2170 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2171 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2172 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2173 * the maximum SACKed segments to pass before reaching this limit.
2175 static void tcp_mark_head_lost(struct sock
*sk
, int packets
, int mark_head
)
2177 struct tcp_sock
*tp
= tcp_sk(sk
);
2178 struct sk_buff
*skb
;
2182 /* Use SACK to deduce losses of new sequences sent during recovery */
2183 const u32 loss_high
= tcp_is_sack(tp
) ? tp
->snd_nxt
: tp
->high_seq
;
2185 WARN_ON(packets
> tp
->packets_out
);
2186 if (tp
->lost_skb_hint
) {
2187 skb
= tp
->lost_skb_hint
;
2188 cnt
= tp
->lost_cnt_hint
;
2189 /* Head already handled? */
2190 if (mark_head
&& skb
!= tcp_write_queue_head(sk
))
2193 skb
= tcp_write_queue_head(sk
);
2197 tcp_for_write_queue_from(skb
, sk
) {
2198 if (skb
== tcp_send_head(sk
))
2200 /* TODO: do this better */
2201 /* this is not the most efficient way to do this... */
2202 tp
->lost_skb_hint
= skb
;
2203 tp
->lost_cnt_hint
= cnt
;
2205 if (after(TCP_SKB_CB(skb
)->end_seq
, loss_high
))
2209 if (tcp_is_fack(tp
) || tcp_is_reno(tp
) ||
2210 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2211 cnt
+= tcp_skb_pcount(skb
);
2213 if (cnt
> packets
) {
2214 if ((tcp_is_sack(tp
) && !tcp_is_fack(tp
)) ||
2215 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
) ||
2216 (oldcnt
>= packets
))
2219 mss
= skb_shinfo(skb
)->gso_size
;
2220 err
= tcp_fragment(sk
, skb
, (packets
- oldcnt
) * mss
, mss
);
2226 tcp_skb_mark_lost(tp
, skb
);
2231 tcp_verify_left_out(tp
);
2234 /* Account newly detected lost packet(s) */
2236 static void tcp_update_scoreboard(struct sock
*sk
, int fast_rexmit
)
2238 struct tcp_sock
*tp
= tcp_sk(sk
);
2240 if (tcp_is_reno(tp
)) {
2241 tcp_mark_head_lost(sk
, 1, 1);
2242 } else if (tcp_is_fack(tp
)) {
2243 int lost
= tp
->fackets_out
- tp
->reordering
;
2246 tcp_mark_head_lost(sk
, lost
, 0);
2248 int sacked_upto
= tp
->sacked_out
- tp
->reordering
;
2249 if (sacked_upto
>= 0)
2250 tcp_mark_head_lost(sk
, sacked_upto
, 0);
2251 else if (fast_rexmit
)
2252 tcp_mark_head_lost(sk
, 1, 1);
2255 tcp_timeout_skbs(sk
);
2258 /* CWND moderation, preventing bursts due to too big ACKs
2259 * in dubious situations.
2261 static inline void tcp_moderate_cwnd(struct tcp_sock
*tp
)
2263 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
2264 tcp_packets_in_flight(tp
) + tcp_max_burst(tp
));
2265 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2268 /* Nothing was retransmitted or returned timestamp is less
2269 * than timestamp of the first retransmission.
2271 static inline bool tcp_packet_delayed(const struct tcp_sock
*tp
)
2273 return !tp
->retrans_stamp
||
2274 (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2275 before(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
));
2278 /* Undo procedures. */
2280 #if FASTRETRANS_DEBUG > 1
2281 static void DBGUNDO(struct sock
*sk
, const char *msg
)
2283 struct tcp_sock
*tp
= tcp_sk(sk
);
2284 struct inet_sock
*inet
= inet_sk(sk
);
2286 if (sk
->sk_family
== AF_INET
) {
2287 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2289 &inet
->inet_daddr
, ntohs(inet
->inet_dport
),
2290 tp
->snd_cwnd
, tcp_left_out(tp
),
2291 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2294 #if IS_ENABLED(CONFIG_IPV6)
2295 else if (sk
->sk_family
== AF_INET6
) {
2296 struct ipv6_pinfo
*np
= inet6_sk(sk
);
2297 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2299 &np
->daddr
, ntohs(inet
->inet_dport
),
2300 tp
->snd_cwnd
, tcp_left_out(tp
),
2301 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2307 #define DBGUNDO(x...) do { } while (0)
2310 static void tcp_undo_cwr(struct sock
*sk
, const bool undo_ssthresh
)
2312 struct tcp_sock
*tp
= tcp_sk(sk
);
2314 if (tp
->prior_ssthresh
) {
2315 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2317 if (icsk
->icsk_ca_ops
->undo_cwnd
)
2318 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
2320 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
<< 1);
2322 if (undo_ssthresh
&& tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
2323 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
2324 TCP_ECN_withdraw_cwr(tp
);
2327 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2329 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2332 static inline bool tcp_may_undo(const struct tcp_sock
*tp
)
2334 return tp
->undo_marker
&& (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
2337 /* People celebrate: "We love our President!" */
2338 static bool tcp_try_undo_recovery(struct sock
*sk
)
2340 struct tcp_sock
*tp
= tcp_sk(sk
);
2342 if (tcp_may_undo(tp
)) {
2345 /* Happy end! We did not retransmit anything
2346 * or our original transmission succeeded.
2348 DBGUNDO(sk
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
2349 tcp_undo_cwr(sk
, true);
2350 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
2351 mib_idx
= LINUX_MIB_TCPLOSSUNDO
;
2353 mib_idx
= LINUX_MIB_TCPFULLUNDO
;
2355 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2356 tp
->undo_marker
= 0;
2358 if (tp
->snd_una
== tp
->high_seq
&& tcp_is_reno(tp
)) {
2359 /* Hold old state until something *above* high_seq
2360 * is ACKed. For Reno it is MUST to prevent false
2361 * fast retransmits (RFC2582). SACK TCP is safe. */
2362 tcp_moderate_cwnd(tp
);
2365 tcp_set_ca_state(sk
, TCP_CA_Open
);
2369 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2370 static void tcp_try_undo_dsack(struct sock
*sk
)
2372 struct tcp_sock
*tp
= tcp_sk(sk
);
2374 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
2375 DBGUNDO(sk
, "D-SACK");
2376 tcp_undo_cwr(sk
, true);
2377 tp
->undo_marker
= 0;
2378 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPDSACKUNDO
);
2382 /* We can clear retrans_stamp when there are no retransmissions in the
2383 * window. It would seem that it is trivially available for us in
2384 * tp->retrans_out, however, that kind of assumptions doesn't consider
2385 * what will happen if errors occur when sending retransmission for the
2386 * second time. ...It could the that such segment has only
2387 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2388 * the head skb is enough except for some reneging corner cases that
2389 * are not worth the effort.
2391 * Main reason for all this complexity is the fact that connection dying
2392 * time now depends on the validity of the retrans_stamp, in particular,
2393 * that successive retransmissions of a segment must not advance
2394 * retrans_stamp under any conditions.
2396 static bool tcp_any_retrans_done(const struct sock
*sk
)
2398 const struct tcp_sock
*tp
= tcp_sk(sk
);
2399 struct sk_buff
*skb
;
2401 if (tp
->retrans_out
)
2404 skb
= tcp_write_queue_head(sk
);
2405 if (unlikely(skb
&& TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
))
2411 /* Undo during fast recovery after partial ACK. */
2413 static int tcp_try_undo_partial(struct sock
*sk
, int acked
)
2415 struct tcp_sock
*tp
= tcp_sk(sk
);
2416 /* Partial ACK arrived. Force Hoe's retransmit. */
2417 int failed
= tcp_is_reno(tp
) || (tcp_fackets_out(tp
) > tp
->reordering
);
2419 if (tcp_may_undo(tp
)) {
2420 /* Plain luck! Hole if filled with delayed
2421 * packet, rather than with a retransmit.
2423 if (!tcp_any_retrans_done(sk
))
2424 tp
->retrans_stamp
= 0;
2426 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
2429 tcp_undo_cwr(sk
, false);
2430 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPARTIALUNDO
);
2432 /* So... Do not make Hoe's retransmit yet.
2433 * If the first packet was delayed, the rest
2434 * ones are most probably delayed as well.
2441 /* Undo during loss recovery after partial ACK or using F-RTO. */
2442 static bool tcp_try_undo_loss(struct sock
*sk
, bool frto_undo
)
2444 struct tcp_sock
*tp
= tcp_sk(sk
);
2446 if (frto_undo
|| tcp_may_undo(tp
)) {
2447 struct sk_buff
*skb
;
2448 tcp_for_write_queue(skb
, sk
) {
2449 if (skb
== tcp_send_head(sk
))
2451 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2454 tcp_clear_all_retrans_hints(tp
);
2456 DBGUNDO(sk
, "partial loss");
2458 tcp_undo_cwr(sk
, true);
2459 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSSUNDO
);
2461 NET_INC_STATS_BH(sock_net(sk
),
2462 LINUX_MIB_TCPSPURIOUSRTOS
);
2463 inet_csk(sk
)->icsk_retransmits
= 0;
2464 tp
->undo_marker
= 0;
2465 if (frto_undo
|| tcp_is_sack(tp
))
2466 tcp_set_ca_state(sk
, TCP_CA_Open
);
2472 /* The cwnd reduction in CWR and Recovery use the PRR algorithm
2473 * https://datatracker.ietf.org/doc/draft-ietf-tcpm-proportional-rate-reduction/
2474 * It computes the number of packets to send (sndcnt) based on packets newly
2476 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2477 * cwnd reductions across a full RTT.
2478 * 2) If packets in flight is lower than ssthresh (such as due to excess
2479 * losses and/or application stalls), do not perform any further cwnd
2480 * reductions, but instead slow start up to ssthresh.
2482 static void tcp_init_cwnd_reduction(struct sock
*sk
, const bool set_ssthresh
)
2484 struct tcp_sock
*tp
= tcp_sk(sk
);
2486 tp
->high_seq
= tp
->snd_nxt
;
2487 tp
->tlp_high_seq
= 0;
2488 tp
->snd_cwnd_cnt
= 0;
2489 tp
->prior_cwnd
= tp
->snd_cwnd
;
2490 tp
->prr_delivered
= 0;
2493 tp
->snd_ssthresh
= inet_csk(sk
)->icsk_ca_ops
->ssthresh(sk
);
2494 TCP_ECN_queue_cwr(tp
);
2497 static void tcp_cwnd_reduction(struct sock
*sk
, int newly_acked_sacked
,
2500 struct tcp_sock
*tp
= tcp_sk(sk
);
2502 int delta
= tp
->snd_ssthresh
- tcp_packets_in_flight(tp
);
2504 tp
->prr_delivered
+= newly_acked_sacked
;
2505 if (tcp_packets_in_flight(tp
) > tp
->snd_ssthresh
) {
2506 u64 dividend
= (u64
)tp
->snd_ssthresh
* tp
->prr_delivered
+
2508 sndcnt
= div_u64(dividend
, tp
->prior_cwnd
) - tp
->prr_out
;
2510 sndcnt
= min_t(int, delta
,
2511 max_t(int, tp
->prr_delivered
- tp
->prr_out
,
2512 newly_acked_sacked
) + 1);
2515 sndcnt
= max(sndcnt
, (fast_rexmit
? 1 : 0));
2516 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + sndcnt
;
2519 static inline void tcp_end_cwnd_reduction(struct sock
*sk
)
2521 struct tcp_sock
*tp
= tcp_sk(sk
);
2523 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2524 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_CWR
||
2525 (tp
->undo_marker
&& tp
->snd_ssthresh
< TCP_INFINITE_SSTHRESH
)) {
2526 tp
->snd_cwnd
= tp
->snd_ssthresh
;
2527 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2529 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
2532 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2533 void tcp_enter_cwr(struct sock
*sk
, const int set_ssthresh
)
2535 struct tcp_sock
*tp
= tcp_sk(sk
);
2537 tp
->prior_ssthresh
= 0;
2538 if (inet_csk(sk
)->icsk_ca_state
< TCP_CA_CWR
) {
2539 tp
->undo_marker
= 0;
2540 tcp_init_cwnd_reduction(sk
, set_ssthresh
);
2541 tcp_set_ca_state(sk
, TCP_CA_CWR
);
2545 static void tcp_try_keep_open(struct sock
*sk
)
2547 struct tcp_sock
*tp
= tcp_sk(sk
);
2548 int state
= TCP_CA_Open
;
2550 if (tcp_left_out(tp
) || tcp_any_retrans_done(sk
))
2551 state
= TCP_CA_Disorder
;
2553 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
2554 tcp_set_ca_state(sk
, state
);
2555 tp
->high_seq
= tp
->snd_nxt
;
2559 static void tcp_try_to_open(struct sock
*sk
, int flag
, int newly_acked_sacked
)
2561 struct tcp_sock
*tp
= tcp_sk(sk
);
2563 tcp_verify_left_out(tp
);
2565 if (!tcp_any_retrans_done(sk
))
2566 tp
->retrans_stamp
= 0;
2568 if (flag
& FLAG_ECE
)
2569 tcp_enter_cwr(sk
, 1);
2571 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
2572 tcp_try_keep_open(sk
);
2573 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
)
2574 tcp_moderate_cwnd(tp
);
2576 tcp_cwnd_reduction(sk
, newly_acked_sacked
, 0);
2580 static void tcp_mtup_probe_failed(struct sock
*sk
)
2582 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2584 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
2585 icsk
->icsk_mtup
.probe_size
= 0;
2588 static void tcp_mtup_probe_success(struct sock
*sk
)
2590 struct tcp_sock
*tp
= tcp_sk(sk
);
2591 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2593 /* FIXME: breaks with very large cwnd */
2594 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2595 tp
->snd_cwnd
= tp
->snd_cwnd
*
2596 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
2597 icsk
->icsk_mtup
.probe_size
;
2598 tp
->snd_cwnd_cnt
= 0;
2599 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2600 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2602 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
2603 icsk
->icsk_mtup
.probe_size
= 0;
2604 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
2607 /* Do a simple retransmit without using the backoff mechanisms in
2608 * tcp_timer. This is used for path mtu discovery.
2609 * The socket is already locked here.
2611 void tcp_simple_retransmit(struct sock
*sk
)
2613 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2614 struct tcp_sock
*tp
= tcp_sk(sk
);
2615 struct sk_buff
*skb
;
2616 unsigned int mss
= tcp_current_mss(sk
);
2617 u32 prior_lost
= tp
->lost_out
;
2619 tcp_for_write_queue(skb
, sk
) {
2620 if (skb
== tcp_send_head(sk
))
2622 if (tcp_skb_seglen(skb
) > mss
&&
2623 !(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
2624 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2625 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
2626 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2628 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
2632 tcp_clear_retrans_hints_partial(tp
);
2634 if (prior_lost
== tp
->lost_out
)
2637 if (tcp_is_reno(tp
))
2638 tcp_limit_reno_sacked(tp
);
2640 tcp_verify_left_out(tp
);
2642 /* Don't muck with the congestion window here.
2643 * Reason is that we do not increase amount of _data_
2644 * in network, but units changed and effective
2645 * cwnd/ssthresh really reduced now.
2647 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
2648 tp
->high_seq
= tp
->snd_nxt
;
2649 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2650 tp
->prior_ssthresh
= 0;
2651 tp
->undo_marker
= 0;
2652 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2654 tcp_xmit_retransmit_queue(sk
);
2656 EXPORT_SYMBOL(tcp_simple_retransmit
);
2658 static void tcp_enter_recovery(struct sock
*sk
, bool ece_ack
)
2660 struct tcp_sock
*tp
= tcp_sk(sk
);
2663 if (tcp_is_reno(tp
))
2664 mib_idx
= LINUX_MIB_TCPRENORECOVERY
;
2666 mib_idx
= LINUX_MIB_TCPSACKRECOVERY
;
2668 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2670 tp
->prior_ssthresh
= 0;
2671 tp
->undo_marker
= tp
->snd_una
;
2672 tp
->undo_retrans
= tp
->retrans_out
;
2674 if (inet_csk(sk
)->icsk_ca_state
< TCP_CA_CWR
) {
2676 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2677 tcp_init_cwnd_reduction(sk
, true);
2679 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
2682 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2683 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2685 static void tcp_process_loss(struct sock
*sk
, int flag
, bool is_dupack
)
2687 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2688 struct tcp_sock
*tp
= tcp_sk(sk
);
2689 bool recovered
= !before(tp
->snd_una
, tp
->high_seq
);
2691 if (tp
->frto
) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2692 if (flag
& FLAG_ORIG_SACK_ACKED
) {
2693 /* Step 3.b. A timeout is spurious if not all data are
2694 * lost, i.e., never-retransmitted data are (s)acked.
2696 tcp_try_undo_loss(sk
, true);
2699 if (after(tp
->snd_nxt
, tp
->high_seq
) &&
2700 (flag
& FLAG_DATA_SACKED
|| is_dupack
)) {
2701 tp
->frto
= 0; /* Loss was real: 2nd part of step 3.a */
2702 } else if (flag
& FLAG_SND_UNA_ADVANCED
&& !recovered
) {
2703 tp
->high_seq
= tp
->snd_nxt
;
2704 __tcp_push_pending_frames(sk
, tcp_current_mss(sk
),
2706 if (after(tp
->snd_nxt
, tp
->high_seq
))
2707 return; /* Step 2.b */
2713 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2714 icsk
->icsk_retransmits
= 0;
2715 tcp_try_undo_recovery(sk
);
2718 if (flag
& FLAG_DATA_ACKED
)
2719 icsk
->icsk_retransmits
= 0;
2720 if (tcp_is_reno(tp
)) {
2721 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2722 * delivered. Lower inflight to clock out (re)tranmissions.
2724 if (after(tp
->snd_nxt
, tp
->high_seq
) && is_dupack
)
2725 tcp_add_reno_sack(sk
);
2726 else if (flag
& FLAG_SND_UNA_ADVANCED
)
2727 tcp_reset_reno_sack(tp
);
2729 if (tcp_try_undo_loss(sk
, false))
2731 tcp_xmit_retransmit_queue(sk
);
2734 /* Process an event, which can update packets-in-flight not trivially.
2735 * Main goal of this function is to calculate new estimate for left_out,
2736 * taking into account both packets sitting in receiver's buffer and
2737 * packets lost by network.
2739 * Besides that it does CWND reduction, when packet loss is detected
2740 * and changes state of machine.
2742 * It does _not_ decide what to send, it is made in function
2743 * tcp_xmit_retransmit_queue().
2745 static void tcp_fastretrans_alert(struct sock
*sk
, int pkts_acked
,
2746 int prior_sacked
, bool is_dupack
,
2749 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2750 struct tcp_sock
*tp
= tcp_sk(sk
);
2751 int do_lost
= is_dupack
|| ((flag
& FLAG_DATA_SACKED
) &&
2752 (tcp_fackets_out(tp
) > tp
->reordering
));
2753 int newly_acked_sacked
= 0;
2754 int fast_rexmit
= 0;
2756 if (WARN_ON(!tp
->packets_out
&& tp
->sacked_out
))
2758 if (WARN_ON(!tp
->sacked_out
&& tp
->fackets_out
))
2759 tp
->fackets_out
= 0;
2761 /* Now state machine starts.
2762 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2763 if (flag
& FLAG_ECE
)
2764 tp
->prior_ssthresh
= 0;
2766 /* B. In all the states check for reneging SACKs. */
2767 if (tcp_check_sack_reneging(sk
, flag
))
2770 /* C. Check consistency of the current state. */
2771 tcp_verify_left_out(tp
);
2773 /* D. Check state exit conditions. State can be terminated
2774 * when high_seq is ACKed. */
2775 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
2776 WARN_ON(tp
->retrans_out
!= 0);
2777 tp
->retrans_stamp
= 0;
2778 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
2779 switch (icsk
->icsk_ca_state
) {
2781 /* CWR is to be held something *above* high_seq
2782 * is ACKed for CWR bit to reach receiver. */
2783 if (tp
->snd_una
!= tp
->high_seq
) {
2784 tcp_end_cwnd_reduction(sk
);
2785 tcp_set_ca_state(sk
, TCP_CA_Open
);
2789 case TCP_CA_Recovery
:
2790 if (tcp_is_reno(tp
))
2791 tcp_reset_reno_sack(tp
);
2792 if (tcp_try_undo_recovery(sk
))
2794 tcp_end_cwnd_reduction(sk
);
2799 /* E. Process state. */
2800 switch (icsk
->icsk_ca_state
) {
2801 case TCP_CA_Recovery
:
2802 if (!(flag
& FLAG_SND_UNA_ADVANCED
)) {
2803 if (tcp_is_reno(tp
) && is_dupack
)
2804 tcp_add_reno_sack(sk
);
2806 do_lost
= tcp_try_undo_partial(sk
, pkts_acked
);
2807 newly_acked_sacked
= pkts_acked
+ tp
->sacked_out
- prior_sacked
;
2810 tcp_process_loss(sk
, flag
, is_dupack
);
2811 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
2813 /* Fall through to processing in Open state. */
2815 if (tcp_is_reno(tp
)) {
2816 if (flag
& FLAG_SND_UNA_ADVANCED
)
2817 tcp_reset_reno_sack(tp
);
2819 tcp_add_reno_sack(sk
);
2821 newly_acked_sacked
= pkts_acked
+ tp
->sacked_out
- prior_sacked
;
2823 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
)
2824 tcp_try_undo_dsack(sk
);
2826 if (!tcp_time_to_recover(sk
, flag
)) {
2827 tcp_try_to_open(sk
, flag
, newly_acked_sacked
);
2831 /* MTU probe failure: don't reduce cwnd */
2832 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
2833 icsk
->icsk_mtup
.probe_size
&&
2834 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
2835 tcp_mtup_probe_failed(sk
);
2836 /* Restores the reduction we did in tcp_mtup_probe() */
2838 tcp_simple_retransmit(sk
);
2842 /* Otherwise enter Recovery state */
2843 tcp_enter_recovery(sk
, (flag
& FLAG_ECE
));
2847 if (do_lost
|| (tcp_is_fack(tp
) && tcp_head_timedout(sk
)))
2848 tcp_update_scoreboard(sk
, fast_rexmit
);
2849 tcp_cwnd_reduction(sk
, newly_acked_sacked
, fast_rexmit
);
2850 tcp_xmit_retransmit_queue(sk
);
2853 void tcp_valid_rtt_meas(struct sock
*sk
, u32 seq_rtt
)
2855 tcp_rtt_estimator(sk
, seq_rtt
);
2857 inet_csk(sk
)->icsk_backoff
= 0;
2859 EXPORT_SYMBOL(tcp_valid_rtt_meas
);
2861 /* Read draft-ietf-tcplw-high-performance before mucking
2862 * with this code. (Supersedes RFC1323)
2864 static void tcp_ack_saw_tstamp(struct sock
*sk
, int flag
)
2866 /* RTTM Rule: A TSecr value received in a segment is used to
2867 * update the averaged RTT measurement only if the segment
2868 * acknowledges some new data, i.e., only if it advances the
2869 * left edge of the send window.
2871 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2872 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2874 * Changed: reset backoff as soon as we see the first valid sample.
2875 * If we do not, we get strongly overestimated rto. With timestamps
2876 * samples are accepted even from very old segments: f.e., when rtt=1
2877 * increases to 8, we retransmit 5 times and after 8 seconds delayed
2878 * answer arrives rto becomes 120 seconds! If at least one of segments
2879 * in window is lost... Voila. --ANK (010210)
2881 struct tcp_sock
*tp
= tcp_sk(sk
);
2883 tcp_valid_rtt_meas(sk
, tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
);
2886 static void tcp_ack_no_tstamp(struct sock
*sk
, u32 seq_rtt
, int flag
)
2888 /* We don't have a timestamp. Can only use
2889 * packets that are not retransmitted to determine
2890 * rtt estimates. Also, we must not reset the
2891 * backoff for rto until we get a non-retransmitted
2892 * packet. This allows us to deal with a situation
2893 * where the network delay has increased suddenly.
2894 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2897 if (flag
& FLAG_RETRANS_DATA_ACKED
)
2900 tcp_valid_rtt_meas(sk
, seq_rtt
);
2903 static inline void tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
2906 const struct tcp_sock
*tp
= tcp_sk(sk
);
2907 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2908 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
2909 tcp_ack_saw_tstamp(sk
, flag
);
2910 else if (seq_rtt
>= 0)
2911 tcp_ack_no_tstamp(sk
, seq_rtt
, flag
);
2914 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
, u32 in_flight
)
2916 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2917 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, in_flight
);
2918 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_time_stamp
;
2921 /* Restart timer after forward progress on connection.
2922 * RFC2988 recommends to restart timer to now+rto.
2924 void tcp_rearm_rto(struct sock
*sk
)
2926 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2927 struct tcp_sock
*tp
= tcp_sk(sk
);
2929 /* If the retrans timer is currently being used by Fast Open
2930 * for SYN-ACK retrans purpose, stay put.
2932 if (tp
->fastopen_rsk
)
2935 if (!tp
->packets_out
) {
2936 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
2938 u32 rto
= inet_csk(sk
)->icsk_rto
;
2939 /* Offset the time elapsed after installing regular RTO */
2940 if (icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
||
2941 icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
) {
2942 struct sk_buff
*skb
= tcp_write_queue_head(sk
);
2943 const u32 rto_time_stamp
= TCP_SKB_CB(skb
)->when
+ rto
;
2944 s32 delta
= (s32
)(rto_time_stamp
- tcp_time_stamp
);
2945 /* delta may not be positive if the socket is locked
2946 * when the retrans timer fires and is rescheduled.
2951 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
, rto
,
2956 /* This function is called when the delayed ER timer fires. TCP enters
2957 * fast recovery and performs fast-retransmit.
2959 void tcp_resume_early_retransmit(struct sock
*sk
)
2961 struct tcp_sock
*tp
= tcp_sk(sk
);
2965 /* Stop if ER is disabled after the delayed ER timer is scheduled */
2966 if (!tp
->do_early_retrans
)
2969 tcp_enter_recovery(sk
, false);
2970 tcp_update_scoreboard(sk
, 1);
2971 tcp_xmit_retransmit_queue(sk
);
2974 /* If we get here, the whole TSO packet has not been acked. */
2975 static u32
tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
)
2977 struct tcp_sock
*tp
= tcp_sk(sk
);
2980 BUG_ON(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
));
2982 packets_acked
= tcp_skb_pcount(skb
);
2983 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
2985 packets_acked
-= tcp_skb_pcount(skb
);
2987 if (packets_acked
) {
2988 BUG_ON(tcp_skb_pcount(skb
) == 0);
2989 BUG_ON(!before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
));
2992 return packets_acked
;
2995 /* Remove acknowledged frames from the retransmission queue. If our packet
2996 * is before the ack sequence we can discard it as it's confirmed to have
2997 * arrived at the other end.
2999 static int tcp_clean_rtx_queue(struct sock
*sk
, int prior_fackets
,
3002 struct tcp_sock
*tp
= tcp_sk(sk
);
3003 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3004 struct sk_buff
*skb
;
3005 u32 now
= tcp_time_stamp
;
3006 int fully_acked
= true;
3009 u32 reord
= tp
->packets_out
;
3010 u32 prior_sacked
= tp
->sacked_out
;
3012 s32 ca_seq_rtt
= -1;
3013 ktime_t last_ackt
= net_invalid_timestamp();
3015 while ((skb
= tcp_write_queue_head(sk
)) && skb
!= tcp_send_head(sk
)) {
3016 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
3018 u8 sacked
= scb
->sacked
;
3020 /* Determine how many packets and what bytes were acked, tso and else */
3021 if (after(scb
->end_seq
, tp
->snd_una
)) {
3022 if (tcp_skb_pcount(skb
) == 1 ||
3023 !after(tp
->snd_una
, scb
->seq
))
3026 acked_pcount
= tcp_tso_acked(sk
, skb
);
3030 fully_acked
= false;
3032 acked_pcount
= tcp_skb_pcount(skb
);
3035 if (sacked
& TCPCB_RETRANS
) {
3036 if (sacked
& TCPCB_SACKED_RETRANS
)
3037 tp
->retrans_out
-= acked_pcount
;
3038 flag
|= FLAG_RETRANS_DATA_ACKED
;
3042 ca_seq_rtt
= now
- scb
->when
;
3043 last_ackt
= skb
->tstamp
;
3045 seq_rtt
= ca_seq_rtt
;
3047 if (!(sacked
& TCPCB_SACKED_ACKED
))
3048 reord
= min(pkts_acked
, reord
);
3049 if (!after(scb
->end_seq
, tp
->high_seq
))
3050 flag
|= FLAG_ORIG_SACK_ACKED
;
3053 if (sacked
& TCPCB_SACKED_ACKED
)
3054 tp
->sacked_out
-= acked_pcount
;
3055 if (sacked
& TCPCB_LOST
)
3056 tp
->lost_out
-= acked_pcount
;
3058 tp
->packets_out
-= acked_pcount
;
3059 pkts_acked
+= acked_pcount
;
3061 /* Initial outgoing SYN's get put onto the write_queue
3062 * just like anything else we transmit. It is not
3063 * true data, and if we misinform our callers that
3064 * this ACK acks real data, we will erroneously exit
3065 * connection startup slow start one packet too
3066 * quickly. This is severely frowned upon behavior.
3068 if (!(scb
->tcp_flags
& TCPHDR_SYN
)) {
3069 flag
|= FLAG_DATA_ACKED
;
3071 flag
|= FLAG_SYN_ACKED
;
3072 tp
->retrans_stamp
= 0;
3078 tcp_unlink_write_queue(skb
, sk
);
3079 sk_wmem_free_skb(sk
, skb
);
3080 tp
->scoreboard_skb_hint
= NULL
;
3081 if (skb
== tp
->retransmit_skb_hint
)
3082 tp
->retransmit_skb_hint
= NULL
;
3083 if (skb
== tp
->lost_skb_hint
)
3084 tp
->lost_skb_hint
= NULL
;
3087 if (likely(between(tp
->snd_up
, prior_snd_una
, tp
->snd_una
)))
3088 tp
->snd_up
= tp
->snd_una
;
3090 if (skb
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
3091 flag
|= FLAG_SACK_RENEGING
;
3093 if (flag
& FLAG_ACKED
) {
3094 const struct tcp_congestion_ops
*ca_ops
3095 = inet_csk(sk
)->icsk_ca_ops
;
3097 if (unlikely(icsk
->icsk_mtup
.probe_size
&&
3098 !after(tp
->mtu_probe
.probe_seq_end
, tp
->snd_una
))) {
3099 tcp_mtup_probe_success(sk
);
3102 tcp_ack_update_rtt(sk
, flag
, seq_rtt
);
3105 if (tcp_is_reno(tp
)) {
3106 tcp_remove_reno_sacks(sk
, pkts_acked
);
3110 /* Non-retransmitted hole got filled? That's reordering */
3111 if (reord
< prior_fackets
)
3112 tcp_update_reordering(sk
, tp
->fackets_out
- reord
, 0);
3114 delta
= tcp_is_fack(tp
) ? pkts_acked
:
3115 prior_sacked
- tp
->sacked_out
;
3116 tp
->lost_cnt_hint
-= min(tp
->lost_cnt_hint
, delta
);
3119 tp
->fackets_out
-= min(pkts_acked
, tp
->fackets_out
);
3121 if (ca_ops
->pkts_acked
) {
3124 /* Is the ACK triggering packet unambiguous? */
3125 if (!(flag
& FLAG_RETRANS_DATA_ACKED
)) {
3126 /* High resolution needed and available? */
3127 if (ca_ops
->flags
& TCP_CONG_RTT_STAMP
&&
3128 !ktime_equal(last_ackt
,
3129 net_invalid_timestamp()))
3130 rtt_us
= ktime_us_delta(ktime_get_real(),
3132 else if (ca_seq_rtt
>= 0)
3133 rtt_us
= jiffies_to_usecs(ca_seq_rtt
);
3136 ca_ops
->pkts_acked(sk
, pkts_acked
, rtt_us
);
3140 #if FASTRETRANS_DEBUG > 0
3141 WARN_ON((int)tp
->sacked_out
< 0);
3142 WARN_ON((int)tp
->lost_out
< 0);
3143 WARN_ON((int)tp
->retrans_out
< 0);
3144 if (!tp
->packets_out
&& tcp_is_sack(tp
)) {
3145 icsk
= inet_csk(sk
);
3147 pr_debug("Leak l=%u %d\n",
3148 tp
->lost_out
, icsk
->icsk_ca_state
);
3151 if (tp
->sacked_out
) {
3152 pr_debug("Leak s=%u %d\n",
3153 tp
->sacked_out
, icsk
->icsk_ca_state
);
3156 if (tp
->retrans_out
) {
3157 pr_debug("Leak r=%u %d\n",
3158 tp
->retrans_out
, icsk
->icsk_ca_state
);
3159 tp
->retrans_out
= 0;
3166 static void tcp_ack_probe(struct sock
*sk
)
3168 const struct tcp_sock
*tp
= tcp_sk(sk
);
3169 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3171 /* Was it a usable window open? */
3173 if (!after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
, tcp_wnd_end(tp
))) {
3174 icsk
->icsk_backoff
= 0;
3175 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
3176 /* Socket must be waked up by subsequent tcp_data_snd_check().
3177 * This function is not for random using!
3180 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3181 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
3186 static inline bool tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
3188 return !(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
3189 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
;
3192 static inline bool tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
3194 const struct tcp_sock
*tp
= tcp_sk(sk
);
3195 return (!(flag
& FLAG_ECE
) || tp
->snd_cwnd
< tp
->snd_ssthresh
) &&
3196 !tcp_in_cwnd_reduction(sk
);
3199 /* Check that window update is acceptable.
3200 * The function assumes that snd_una<=ack<=snd_next.
3202 static inline bool tcp_may_update_window(const struct tcp_sock
*tp
,
3203 const u32 ack
, const u32 ack_seq
,
3206 return after(ack
, tp
->snd_una
) ||
3207 after(ack_seq
, tp
->snd_wl1
) ||
3208 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
);
3211 /* Update our send window.
3213 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3214 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3216 static int tcp_ack_update_window(struct sock
*sk
, const struct sk_buff
*skb
, u32 ack
,
3219 struct tcp_sock
*tp
= tcp_sk(sk
);
3221 u32 nwin
= ntohs(tcp_hdr(skb
)->window
);
3223 if (likely(!tcp_hdr(skb
)->syn
))
3224 nwin
<<= tp
->rx_opt
.snd_wscale
;
3226 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
3227 flag
|= FLAG_WIN_UPDATE
;
3228 tcp_update_wl(tp
, ack_seq
);
3230 if (tp
->snd_wnd
!= nwin
) {
3233 /* Note, it is the only place, where
3234 * fast path is recovered for sending TCP.
3237 tcp_fast_path_check(sk
);
3239 if (nwin
> tp
->max_window
) {
3240 tp
->max_window
= nwin
;
3241 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
3251 /* RFC 5961 7 [ACK Throttling] */
3252 static void tcp_send_challenge_ack(struct sock
*sk
)
3254 /* unprotected vars, we dont care of overwrites */
3255 static u32 challenge_timestamp
;
3256 static unsigned int challenge_count
;
3257 u32 now
= jiffies
/ HZ
;
3259 if (now
!= challenge_timestamp
) {
3260 challenge_timestamp
= now
;
3261 challenge_count
= 0;
3263 if (++challenge_count
<= sysctl_tcp_challenge_ack_limit
) {
3264 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPCHALLENGEACK
);
3269 static void tcp_store_ts_recent(struct tcp_sock
*tp
)
3271 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
3272 tp
->rx_opt
.ts_recent_stamp
= get_seconds();
3275 static void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
3277 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
3278 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3279 * extra check below makes sure this can only happen
3280 * for pure ACK frames. -DaveM
3282 * Not only, also it occurs for expired timestamps.
3285 if (tcp_paws_check(&tp
->rx_opt
, 0))
3286 tcp_store_ts_recent(tp
);
3290 /* This routine deals with acks during a TLP episode.
3291 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3293 static void tcp_process_tlp_ack(struct sock
*sk
, u32 ack
, int flag
)
3295 struct tcp_sock
*tp
= tcp_sk(sk
);
3296 bool is_tlp_dupack
= (ack
== tp
->tlp_high_seq
) &&
3297 !(flag
& (FLAG_SND_UNA_ADVANCED
|
3298 FLAG_NOT_DUP
| FLAG_DATA_SACKED
));
3300 /* Mark the end of TLP episode on receiving TLP dupack or when
3301 * ack is after tlp_high_seq.
3303 if (is_tlp_dupack
) {
3304 tp
->tlp_high_seq
= 0;
3308 if (after(ack
, tp
->tlp_high_seq
)) {
3309 tp
->tlp_high_seq
= 0;
3310 /* Don't reduce cwnd if DSACK arrives for TLP retrans. */
3311 if (!(flag
& FLAG_DSACKING_ACK
)) {
3312 tcp_init_cwnd_reduction(sk
, true);
3313 tcp_set_ca_state(sk
, TCP_CA_CWR
);
3314 tcp_end_cwnd_reduction(sk
);
3315 tcp_set_ca_state(sk
, TCP_CA_Open
);
3316 NET_INC_STATS_BH(sock_net(sk
),
3317 LINUX_MIB_TCPLOSSPROBERECOVERY
);
3322 /* This routine deals with incoming acks, but not outgoing ones. */
3323 static int tcp_ack(struct sock
*sk
, const struct sk_buff
*skb
, int flag
)
3325 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3326 struct tcp_sock
*tp
= tcp_sk(sk
);
3327 u32 prior_snd_una
= tp
->snd_una
;
3328 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
3329 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3330 bool is_dupack
= false;
3331 u32 prior_in_flight
;
3334 int prior_sacked
= tp
->sacked_out
;
3337 /* If the ack is older than previous acks
3338 * then we can probably ignore it.
3340 if (before(ack
, prior_snd_una
)) {
3341 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3342 if (before(ack
, prior_snd_una
- tp
->max_window
)) {
3343 tcp_send_challenge_ack(sk
);
3349 /* If the ack includes data we haven't sent yet, discard
3350 * this segment (RFC793 Section 3.9).
3352 if (after(ack
, tp
->snd_nxt
))
3355 if (icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
||
3356 icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
)
3359 if (after(ack
, prior_snd_una
))
3360 flag
|= FLAG_SND_UNA_ADVANCED
;
3362 prior_fackets
= tp
->fackets_out
;
3363 prior_in_flight
= tcp_packets_in_flight(tp
);
3365 /* ts_recent update must be made after we are sure that the packet
3368 if (flag
& FLAG_UPDATE_TS_RECENT
)
3369 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
3371 if (!(flag
& FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
3372 /* Window is constant, pure forward advance.
3373 * No more checks are required.
3374 * Note, we use the fact that SND.UNA>=SND.WL2.
3376 tcp_update_wl(tp
, ack_seq
);
3378 flag
|= FLAG_WIN_UPDATE
;
3380 tcp_ca_event(sk
, CA_EVENT_FAST_ACK
);
3382 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPACKS
);
3384 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
3387 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPUREACKS
);
3389 flag
|= tcp_ack_update_window(sk
, skb
, ack
, ack_seq
);
3391 if (TCP_SKB_CB(skb
)->sacked
)
3392 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
3394 if (TCP_ECN_rcv_ecn_echo(tp
, tcp_hdr(skb
)))
3397 tcp_ca_event(sk
, CA_EVENT_SLOW_ACK
);
3400 /* We passed data and got it acked, remove any soft error
3401 * log. Something worked...
3403 sk
->sk_err_soft
= 0;
3404 icsk
->icsk_probes_out
= 0;
3405 tp
->rcv_tstamp
= tcp_time_stamp
;
3406 prior_packets
= tp
->packets_out
;
3410 /* See if we can take anything off of the retransmit queue. */
3411 flag
|= tcp_clean_rtx_queue(sk
, prior_fackets
, prior_snd_una
);
3413 pkts_acked
= prior_packets
- tp
->packets_out
;
3415 if (tcp_ack_is_dubious(sk
, flag
)) {
3416 /* Advance CWND, if state allows this. */
3417 if ((flag
& FLAG_DATA_ACKED
) && tcp_may_raise_cwnd(sk
, flag
))
3418 tcp_cong_avoid(sk
, ack
, prior_in_flight
);
3419 is_dupack
= !(flag
& (FLAG_SND_UNA_ADVANCED
| FLAG_NOT_DUP
));
3420 tcp_fastretrans_alert(sk
, pkts_acked
, prior_sacked
,
3423 if (flag
& FLAG_DATA_ACKED
)
3424 tcp_cong_avoid(sk
, ack
, prior_in_flight
);
3427 if (tp
->tlp_high_seq
)
3428 tcp_process_tlp_ack(sk
, ack
, flag
);
3430 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
& FLAG_NOT_DUP
)) {
3431 struct dst_entry
*dst
= __sk_dst_get(sk
);
3436 if (icsk
->icsk_pending
== ICSK_TIME_RETRANS
)
3437 tcp_schedule_loss_probe(sk
);
3441 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3442 if (flag
& FLAG_DSACKING_ACK
)
3443 tcp_fastretrans_alert(sk
, pkts_acked
, prior_sacked
,
3445 /* If this ack opens up a zero window, clear backoff. It was
3446 * being used to time the probes, and is probably far higher than
3447 * it needs to be for normal retransmission.
3449 if (tcp_send_head(sk
))
3452 if (tp
->tlp_high_seq
)
3453 tcp_process_tlp_ack(sk
, ack
, flag
);
3457 SOCK_DEBUG(sk
, "Ack %u after %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3461 /* If data was SACKed, tag it and see if we should send more data.
3462 * If data was DSACKed, see if we can undo a cwnd reduction.
3464 if (TCP_SKB_CB(skb
)->sacked
) {
3465 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
3466 tcp_fastretrans_alert(sk
, pkts_acked
, prior_sacked
,
3470 SOCK_DEBUG(sk
, "Ack %u before %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3474 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3475 * But, this can also be called on packets in the established flow when
3476 * the fast version below fails.
3478 void tcp_parse_options(const struct sk_buff
*skb
,
3479 struct tcp_options_received
*opt_rx
, int estab
,
3480 struct tcp_fastopen_cookie
*foc
)
3482 const unsigned char *ptr
;
3483 const struct tcphdr
*th
= tcp_hdr(skb
);
3484 int length
= (th
->doff
* 4) - sizeof(struct tcphdr
);
3486 ptr
= (const unsigned char *)(th
+ 1);
3487 opt_rx
->saw_tstamp
= 0;
3489 while (length
> 0) {
3490 int opcode
= *ptr
++;
3496 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
3501 if (opsize
< 2) /* "silly options" */
3503 if (opsize
> length
)
3504 return; /* don't parse partial options */
3507 if (opsize
== TCPOLEN_MSS
&& th
->syn
&& !estab
) {
3508 u16 in_mss
= get_unaligned_be16(ptr
);
3510 if (opt_rx
->user_mss
&&
3511 opt_rx
->user_mss
< in_mss
)
3512 in_mss
= opt_rx
->user_mss
;
3513 opt_rx
->mss_clamp
= in_mss
;
3518 if (opsize
== TCPOLEN_WINDOW
&& th
->syn
&&
3519 !estab
&& sysctl_tcp_window_scaling
) {
3520 __u8 snd_wscale
= *(__u8
*)ptr
;
3521 opt_rx
->wscale_ok
= 1;
3522 if (snd_wscale
> 14) {
3523 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3528 opt_rx
->snd_wscale
= snd_wscale
;
3531 case TCPOPT_TIMESTAMP
:
3532 if ((opsize
== TCPOLEN_TIMESTAMP
) &&
3533 ((estab
&& opt_rx
->tstamp_ok
) ||
3534 (!estab
&& sysctl_tcp_timestamps
))) {
3535 opt_rx
->saw_tstamp
= 1;
3536 opt_rx
->rcv_tsval
= get_unaligned_be32(ptr
);
3537 opt_rx
->rcv_tsecr
= get_unaligned_be32(ptr
+ 4);
3540 case TCPOPT_SACK_PERM
:
3541 if (opsize
== TCPOLEN_SACK_PERM
&& th
->syn
&&
3542 !estab
&& sysctl_tcp_sack
) {
3543 opt_rx
->sack_ok
= TCP_SACK_SEEN
;
3544 tcp_sack_reset(opt_rx
);
3549 if ((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
3550 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
3552 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
3555 #ifdef CONFIG_TCP_MD5SIG
3558 * The MD5 Hash has already been
3559 * checked (see tcp_v{4,6}_do_rcv()).
3564 /* Fast Open option shares code 254 using a
3565 * 16 bits magic number. It's valid only in
3566 * SYN or SYN-ACK with an even size.
3568 if (opsize
< TCPOLEN_EXP_FASTOPEN_BASE
||
3569 get_unaligned_be16(ptr
) != TCPOPT_FASTOPEN_MAGIC
||
3570 foc
== NULL
|| !th
->syn
|| (opsize
& 1))
3572 foc
->len
= opsize
- TCPOLEN_EXP_FASTOPEN_BASE
;
3573 if (foc
->len
>= TCP_FASTOPEN_COOKIE_MIN
&&
3574 foc
->len
<= TCP_FASTOPEN_COOKIE_MAX
)
3575 memcpy(foc
->val
, ptr
+ 2, foc
->len
);
3576 else if (foc
->len
!= 0)
3586 EXPORT_SYMBOL(tcp_parse_options
);
3588 static bool tcp_parse_aligned_timestamp(struct tcp_sock
*tp
, const struct tcphdr
*th
)
3590 const __be32
*ptr
= (const __be32
*)(th
+ 1);
3592 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3593 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
3594 tp
->rx_opt
.saw_tstamp
= 1;
3596 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3598 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
) - tp
->tsoffset
;
3604 /* Fast parse options. This hopes to only see timestamps.
3605 * If it is wrong it falls back on tcp_parse_options().
3607 static bool tcp_fast_parse_options(const struct sk_buff
*skb
,
3608 const struct tcphdr
*th
, struct tcp_sock
*tp
)
3610 /* In the spirit of fast parsing, compare doff directly to constant
3611 * values. Because equality is used, short doff can be ignored here.
3613 if (th
->doff
== (sizeof(*th
) / 4)) {
3614 tp
->rx_opt
.saw_tstamp
= 0;
3616 } else if (tp
->rx_opt
.tstamp_ok
&&
3617 th
->doff
== ((sizeof(*th
) + TCPOLEN_TSTAMP_ALIGNED
) / 4)) {
3618 if (tcp_parse_aligned_timestamp(tp
, th
))
3622 tcp_parse_options(skb
, &tp
->rx_opt
, 1, NULL
);
3623 if (tp
->rx_opt
.saw_tstamp
)
3624 tp
->rx_opt
.rcv_tsecr
-= tp
->tsoffset
;
3629 #ifdef CONFIG_TCP_MD5SIG
3631 * Parse MD5 Signature option
3633 const u8
*tcp_parse_md5sig_option(const struct tcphdr
*th
)
3635 int length
= (th
->doff
<< 2) - sizeof(*th
);
3636 const u8
*ptr
= (const u8
*)(th
+ 1);
3638 /* If the TCP option is too short, we can short cut */
3639 if (length
< TCPOLEN_MD5SIG
)
3642 while (length
> 0) {
3643 int opcode
= *ptr
++;
3654 if (opsize
< 2 || opsize
> length
)
3656 if (opcode
== TCPOPT_MD5SIG
)
3657 return opsize
== TCPOLEN_MD5SIG
? ptr
: NULL
;
3664 EXPORT_SYMBOL(tcp_parse_md5sig_option
);
3667 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3669 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3670 * it can pass through stack. So, the following predicate verifies that
3671 * this segment is not used for anything but congestion avoidance or
3672 * fast retransmit. Moreover, we even are able to eliminate most of such
3673 * second order effects, if we apply some small "replay" window (~RTO)
3674 * to timestamp space.
3676 * All these measures still do not guarantee that we reject wrapped ACKs
3677 * on networks with high bandwidth, when sequence space is recycled fastly,
3678 * but it guarantees that such events will be very rare and do not affect
3679 * connection seriously. This doesn't look nice, but alas, PAWS is really
3682 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3683 * states that events when retransmit arrives after original data are rare.
3684 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3685 * the biggest problem on large power networks even with minor reordering.
3686 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3687 * up to bandwidth of 18Gigabit/sec. 8) ]
3690 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
3692 const struct tcp_sock
*tp
= tcp_sk(sk
);
3693 const struct tcphdr
*th
= tcp_hdr(skb
);
3694 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3695 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3697 return (/* 1. Pure ACK with correct sequence number. */
3698 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
3700 /* 2. ... and duplicate ACK. */
3701 ack
== tp
->snd_una
&&
3703 /* 3. ... and does not update window. */
3704 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
3706 /* 4. ... and sits in replay window. */
3707 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
3710 static inline bool tcp_paws_discard(const struct sock
*sk
,
3711 const struct sk_buff
*skb
)
3713 const struct tcp_sock
*tp
= tcp_sk(sk
);
3715 return !tcp_paws_check(&tp
->rx_opt
, TCP_PAWS_WINDOW
) &&
3716 !tcp_disordered_ack(sk
, skb
);
3719 /* Check segment sequence number for validity.
3721 * Segment controls are considered valid, if the segment
3722 * fits to the window after truncation to the window. Acceptability
3723 * of data (and SYN, FIN, of course) is checked separately.
3724 * See tcp_data_queue(), for example.
3726 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3727 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3728 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3729 * (borrowed from freebsd)
3732 static inline bool tcp_sequence(const struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3734 return !before(end_seq
, tp
->rcv_wup
) &&
3735 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
3738 /* When we get a reset we do this. */
3739 void tcp_reset(struct sock
*sk
)
3741 /* We want the right error as BSD sees it (and indeed as we do). */
3742 switch (sk
->sk_state
) {
3744 sk
->sk_err
= ECONNREFUSED
;
3746 case TCP_CLOSE_WAIT
:
3752 sk
->sk_err
= ECONNRESET
;
3754 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3757 if (!sock_flag(sk
, SOCK_DEAD
))
3758 sk
->sk_error_report(sk
);
3764 * Process the FIN bit. This now behaves as it is supposed to work
3765 * and the FIN takes effect when it is validly part of sequence
3766 * space. Not before when we get holes.
3768 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3769 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3772 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3773 * close and we go into CLOSING (and later onto TIME-WAIT)
3775 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3777 static void tcp_fin(struct sock
*sk
)
3779 struct tcp_sock
*tp
= tcp_sk(sk
);
3781 inet_csk_schedule_ack(sk
);
3783 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
3784 sock_set_flag(sk
, SOCK_DONE
);
3786 switch (sk
->sk_state
) {
3788 case TCP_ESTABLISHED
:
3789 /* Move to CLOSE_WAIT */
3790 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
3791 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
3794 case TCP_CLOSE_WAIT
:
3796 /* Received a retransmission of the FIN, do
3801 /* RFC793: Remain in the LAST-ACK state. */
3805 /* This case occurs when a simultaneous close
3806 * happens, we must ack the received FIN and
3807 * enter the CLOSING state.
3810 tcp_set_state(sk
, TCP_CLOSING
);
3813 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3815 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
3818 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3819 * cases we should never reach this piece of code.
3821 pr_err("%s: Impossible, sk->sk_state=%d\n",
3822 __func__
, sk
->sk_state
);
3826 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3827 * Probably, we should reset in this case. For now drop them.
3829 __skb_queue_purge(&tp
->out_of_order_queue
);
3830 if (tcp_is_sack(tp
))
3831 tcp_sack_reset(&tp
->rx_opt
);
3834 if (!sock_flag(sk
, SOCK_DEAD
)) {
3835 sk
->sk_state_change(sk
);
3837 /* Do not send POLL_HUP for half duplex close. */
3838 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
3839 sk
->sk_state
== TCP_CLOSE
)
3840 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_HUP
);
3842 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_IN
);
3846 static inline bool tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
,
3849 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
3850 if (before(seq
, sp
->start_seq
))
3851 sp
->start_seq
= seq
;
3852 if (after(end_seq
, sp
->end_seq
))
3853 sp
->end_seq
= end_seq
;
3859 static void tcp_dsack_set(struct sock
*sk
, u32 seq
, u32 end_seq
)
3861 struct tcp_sock
*tp
= tcp_sk(sk
);
3863 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
3866 if (before(seq
, tp
->rcv_nxt
))
3867 mib_idx
= LINUX_MIB_TCPDSACKOLDSENT
;
3869 mib_idx
= LINUX_MIB_TCPDSACKOFOSENT
;
3871 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
3873 tp
->rx_opt
.dsack
= 1;
3874 tp
->duplicate_sack
[0].start_seq
= seq
;
3875 tp
->duplicate_sack
[0].end_seq
= end_seq
;
3879 static void tcp_dsack_extend(struct sock
*sk
, u32 seq
, u32 end_seq
)
3881 struct tcp_sock
*tp
= tcp_sk(sk
);
3883 if (!tp
->rx_opt
.dsack
)
3884 tcp_dsack_set(sk
, seq
, end_seq
);
3886 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
3889 static void tcp_send_dupack(struct sock
*sk
, const struct sk_buff
*skb
)
3891 struct tcp_sock
*tp
= tcp_sk(sk
);
3893 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
3894 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
3895 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
3896 tcp_enter_quickack_mode(sk
);
3898 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
3899 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
3901 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
3902 end_seq
= tp
->rcv_nxt
;
3903 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, end_seq
);
3910 /* These routines update the SACK block as out-of-order packets arrive or
3911 * in-order packets close up the sequence space.
3913 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
3916 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3917 struct tcp_sack_block
*swalk
= sp
+ 1;
3919 /* See if the recent change to the first SACK eats into
3920 * or hits the sequence space of other SACK blocks, if so coalesce.
3922 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
;) {
3923 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
3926 /* Zap SWALK, by moving every further SACK up by one slot.
3927 * Decrease num_sacks.
3929 tp
->rx_opt
.num_sacks
--;
3930 for (i
= this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
3934 this_sack
++, swalk
++;
3938 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
3940 struct tcp_sock
*tp
= tcp_sk(sk
);
3941 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3942 int cur_sacks
= tp
->rx_opt
.num_sacks
;
3948 for (this_sack
= 0; this_sack
< cur_sacks
; this_sack
++, sp
++) {
3949 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
3950 /* Rotate this_sack to the first one. */
3951 for (; this_sack
> 0; this_sack
--, sp
--)
3952 swap(*sp
, *(sp
- 1));
3954 tcp_sack_maybe_coalesce(tp
);
3959 /* Could not find an adjacent existing SACK, build a new one,
3960 * put it at the front, and shift everyone else down. We
3961 * always know there is at least one SACK present already here.
3963 * If the sack array is full, forget about the last one.
3965 if (this_sack
>= TCP_NUM_SACKS
) {
3967 tp
->rx_opt
.num_sacks
--;
3970 for (; this_sack
> 0; this_sack
--, sp
--)
3974 /* Build the new head SACK, and we're done. */
3975 sp
->start_seq
= seq
;
3976 sp
->end_seq
= end_seq
;
3977 tp
->rx_opt
.num_sacks
++;
3980 /* RCV.NXT advances, some SACKs should be eaten. */
3982 static void tcp_sack_remove(struct tcp_sock
*tp
)
3984 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3985 int num_sacks
= tp
->rx_opt
.num_sacks
;
3988 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3989 if (skb_queue_empty(&tp
->out_of_order_queue
)) {
3990 tp
->rx_opt
.num_sacks
= 0;
3994 for (this_sack
= 0; this_sack
< num_sacks
;) {
3995 /* Check if the start of the sack is covered by RCV.NXT. */
3996 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
3999 /* RCV.NXT must cover all the block! */
4000 WARN_ON(before(tp
->rcv_nxt
, sp
->end_seq
));
4002 /* Zap this SACK, by moving forward any other SACKS. */
4003 for (i
=this_sack
+1; i
< num_sacks
; i
++)
4004 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
4011 tp
->rx_opt
.num_sacks
= num_sacks
;
4014 /* This one checks to see if we can put data from the
4015 * out_of_order queue into the receive_queue.
4017 static void tcp_ofo_queue(struct sock
*sk
)
4019 struct tcp_sock
*tp
= tcp_sk(sk
);
4020 __u32 dsack_high
= tp
->rcv_nxt
;
4021 struct sk_buff
*skb
;
4023 while ((skb
= skb_peek(&tp
->out_of_order_queue
)) != NULL
) {
4024 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
4027 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
4028 __u32 dsack
= dsack_high
;
4029 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
4030 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
4031 tcp_dsack_extend(sk
, TCP_SKB_CB(skb
)->seq
, dsack
);
4034 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4035 SOCK_DEBUG(sk
, "ofo packet was already received\n");
4036 __skb_unlink(skb
, &tp
->out_of_order_queue
);
4040 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
4041 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4042 TCP_SKB_CB(skb
)->end_seq
);
4044 __skb_unlink(skb
, &tp
->out_of_order_queue
);
4045 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4046 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4047 if (tcp_hdr(skb
)->fin
)
4052 static bool tcp_prune_ofo_queue(struct sock
*sk
);
4053 static int tcp_prune_queue(struct sock
*sk
);
4055 static int tcp_try_rmem_schedule(struct sock
*sk
, struct sk_buff
*skb
,
4058 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
4059 !sk_rmem_schedule(sk
, skb
, size
)) {
4061 if (tcp_prune_queue(sk
) < 0)
4064 if (!sk_rmem_schedule(sk
, skb
, size
)) {
4065 if (!tcp_prune_ofo_queue(sk
))
4068 if (!sk_rmem_schedule(sk
, skb
, size
))
4076 * tcp_try_coalesce - try to merge skb to prior one
4079 * @from: buffer to add in queue
4080 * @fragstolen: pointer to boolean
4082 * Before queueing skb @from after @to, try to merge them
4083 * to reduce overall memory use and queue lengths, if cost is small.
4084 * Packets in ofo or receive queues can stay a long time.
4085 * Better try to coalesce them right now to avoid future collapses.
4086 * Returns true if caller should free @from instead of queueing it
4088 static bool tcp_try_coalesce(struct sock
*sk
,
4090 struct sk_buff
*from
,
4095 *fragstolen
= false;
4097 if (tcp_hdr(from
)->fin
)
4100 /* Its possible this segment overlaps with prior segment in queue */
4101 if (TCP_SKB_CB(from
)->seq
!= TCP_SKB_CB(to
)->end_seq
)
4104 if (!skb_try_coalesce(to
, from
, fragstolen
, &delta
))
4107 atomic_add(delta
, &sk
->sk_rmem_alloc
);
4108 sk_mem_charge(sk
, delta
);
4109 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRCVCOALESCE
);
4110 TCP_SKB_CB(to
)->end_seq
= TCP_SKB_CB(from
)->end_seq
;
4111 TCP_SKB_CB(to
)->ack_seq
= TCP_SKB_CB(from
)->ack_seq
;
4115 static void tcp_data_queue_ofo(struct sock
*sk
, struct sk_buff
*skb
)
4117 struct tcp_sock
*tp
= tcp_sk(sk
);
4118 struct sk_buff
*skb1
;
4121 TCP_ECN_check_ce(tp
, skb
);
4123 if (unlikely(tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))) {
4124 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFODROP
);
4129 /* Disable header prediction. */
4131 inet_csk_schedule_ack(sk
);
4133 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFOQUEUE
);
4134 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
4135 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4137 skb1
= skb_peek_tail(&tp
->out_of_order_queue
);
4139 /* Initial out of order segment, build 1 SACK. */
4140 if (tcp_is_sack(tp
)) {
4141 tp
->rx_opt
.num_sacks
= 1;
4142 tp
->selective_acks
[0].start_seq
= TCP_SKB_CB(skb
)->seq
;
4143 tp
->selective_acks
[0].end_seq
=
4144 TCP_SKB_CB(skb
)->end_seq
;
4146 __skb_queue_head(&tp
->out_of_order_queue
, skb
);
4150 seq
= TCP_SKB_CB(skb
)->seq
;
4151 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4153 if (seq
== TCP_SKB_CB(skb1
)->end_seq
) {
4156 if (!tcp_try_coalesce(sk
, skb1
, skb
, &fragstolen
)) {
4157 __skb_queue_after(&tp
->out_of_order_queue
, skb1
, skb
);
4159 kfree_skb_partial(skb
, fragstolen
);
4163 if (!tp
->rx_opt
.num_sacks
||
4164 tp
->selective_acks
[0].end_seq
!= seq
)
4167 /* Common case: data arrive in order after hole. */
4168 tp
->selective_acks
[0].end_seq
= end_seq
;
4172 /* Find place to insert this segment. */
4174 if (!after(TCP_SKB_CB(skb1
)->seq
, seq
))
4176 if (skb_queue_is_first(&tp
->out_of_order_queue
, skb1
)) {
4180 skb1
= skb_queue_prev(&tp
->out_of_order_queue
, skb1
);
4183 /* Do skb overlap to previous one? */
4184 if (skb1
&& before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4185 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4186 /* All the bits are present. Drop. */
4187 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFOMERGE
);
4190 tcp_dsack_set(sk
, seq
, end_seq
);
4193 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
4194 /* Partial overlap. */
4195 tcp_dsack_set(sk
, seq
,
4196 TCP_SKB_CB(skb1
)->end_seq
);
4198 if (skb_queue_is_first(&tp
->out_of_order_queue
,
4202 skb1
= skb_queue_prev(
4203 &tp
->out_of_order_queue
,
4208 __skb_queue_head(&tp
->out_of_order_queue
, skb
);
4210 __skb_queue_after(&tp
->out_of_order_queue
, skb1
, skb
);
4212 /* And clean segments covered by new one as whole. */
4213 while (!skb_queue_is_last(&tp
->out_of_order_queue
, skb
)) {
4214 skb1
= skb_queue_next(&tp
->out_of_order_queue
, skb
);
4216 if (!after(end_seq
, TCP_SKB_CB(skb1
)->seq
))
4218 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4219 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4223 __skb_unlink(skb1
, &tp
->out_of_order_queue
);
4224 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4225 TCP_SKB_CB(skb1
)->end_seq
);
4226 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFOMERGE
);
4231 if (tcp_is_sack(tp
))
4232 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
4235 skb_set_owner_r(skb
, sk
);
4238 static int __must_check
tcp_queue_rcv(struct sock
*sk
, struct sk_buff
*skb
, int hdrlen
,
4242 struct sk_buff
*tail
= skb_peek_tail(&sk
->sk_receive_queue
);
4244 __skb_pull(skb
, hdrlen
);
4246 tcp_try_coalesce(sk
, tail
, skb
, fragstolen
)) ? 1 : 0;
4247 tcp_sk(sk
)->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4249 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4250 skb_set_owner_r(skb
, sk
);
4255 int tcp_send_rcvq(struct sock
*sk
, struct msghdr
*msg
, size_t size
)
4257 struct sk_buff
*skb
= NULL
;
4264 skb
= alloc_skb(size
+ sizeof(*th
), sk
->sk_allocation
);
4268 if (tcp_try_rmem_schedule(sk
, skb
, size
+ sizeof(*th
)))
4271 th
= (struct tcphdr
*)skb_put(skb
, sizeof(*th
));
4272 skb_reset_transport_header(skb
);
4273 memset(th
, 0, sizeof(*th
));
4275 if (memcpy_fromiovec(skb_put(skb
, size
), msg
->msg_iov
, size
))
4278 TCP_SKB_CB(skb
)->seq
= tcp_sk(sk
)->rcv_nxt
;
4279 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ size
;
4280 TCP_SKB_CB(skb
)->ack_seq
= tcp_sk(sk
)->snd_una
- 1;
4282 if (tcp_queue_rcv(sk
, skb
, sizeof(*th
), &fragstolen
)) {
4283 WARN_ON_ONCE(fragstolen
); /* should not happen */
4294 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
4296 const struct tcphdr
*th
= tcp_hdr(skb
);
4297 struct tcp_sock
*tp
= tcp_sk(sk
);
4299 bool fragstolen
= false;
4301 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
)
4305 __skb_pull(skb
, th
->doff
* 4);
4307 TCP_ECN_accept_cwr(tp
, skb
);
4309 tp
->rx_opt
.dsack
= 0;
4311 /* Queue data for delivery to the user.
4312 * Packets in sequence go to the receive queue.
4313 * Out of sequence packets to the out_of_order_queue.
4315 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
4316 if (tcp_receive_window(tp
) == 0)
4319 /* Ok. In sequence. In window. */
4320 if (tp
->ucopy
.task
== current
&&
4321 tp
->copied_seq
== tp
->rcv_nxt
&& tp
->ucopy
.len
&&
4322 sock_owned_by_user(sk
) && !tp
->urg_data
) {
4323 int chunk
= min_t(unsigned int, skb
->len
,
4326 __set_current_state(TASK_RUNNING
);
4329 if (!skb_copy_datagram_iovec(skb
, 0, tp
->ucopy
.iov
, chunk
)) {
4330 tp
->ucopy
.len
-= chunk
;
4331 tp
->copied_seq
+= chunk
;
4332 eaten
= (chunk
== skb
->len
);
4333 tcp_rcv_space_adjust(sk
);
4341 tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))
4344 eaten
= tcp_queue_rcv(sk
, skb
, 0, &fragstolen
);
4346 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4348 tcp_event_data_recv(sk
, skb
);
4352 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4355 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4356 * gap in queue is filled.
4358 if (skb_queue_empty(&tp
->out_of_order_queue
))
4359 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
4362 if (tp
->rx_opt
.num_sacks
)
4363 tcp_sack_remove(tp
);
4365 tcp_fast_path_check(sk
);
4368 kfree_skb_partial(skb
, fragstolen
);
4369 if (!sock_flag(sk
, SOCK_DEAD
))
4370 sk
->sk_data_ready(sk
, 0);
4374 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4375 /* A retransmit, 2nd most common case. Force an immediate ack. */
4376 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4377 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4380 tcp_enter_quickack_mode(sk
);
4381 inet_csk_schedule_ack(sk
);
4387 /* Out of window. F.e. zero window probe. */
4388 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
4391 tcp_enter_quickack_mode(sk
);
4393 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4394 /* Partial packet, seq < rcv_next < end_seq */
4395 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
4396 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4397 TCP_SKB_CB(skb
)->end_seq
);
4399 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
4401 /* If window is closed, drop tail of packet. But after
4402 * remembering D-SACK for its head made in previous line.
4404 if (!tcp_receive_window(tp
))
4409 tcp_data_queue_ofo(sk
, skb
);
4412 static struct sk_buff
*tcp_collapse_one(struct sock
*sk
, struct sk_buff
*skb
,
4413 struct sk_buff_head
*list
)
4415 struct sk_buff
*next
= NULL
;
4417 if (!skb_queue_is_last(list
, skb
))
4418 next
= skb_queue_next(list
, skb
);
4420 __skb_unlink(skb
, list
);
4422 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRCVCOLLAPSED
);
4427 /* Collapse contiguous sequence of skbs head..tail with
4428 * sequence numbers start..end.
4430 * If tail is NULL, this means until the end of the list.
4432 * Segments with FIN/SYN are not collapsed (only because this
4436 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
,
4437 struct sk_buff
*head
, struct sk_buff
*tail
,
4440 struct sk_buff
*skb
, *n
;
4443 /* First, check that queue is collapsible and find
4444 * the point where collapsing can be useful. */
4448 skb_queue_walk_from_safe(list
, skb
, n
) {
4451 /* No new bits? It is possible on ofo queue. */
4452 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4453 skb
= tcp_collapse_one(sk
, skb
, list
);
4459 /* The first skb to collapse is:
4461 * - bloated or contains data before "start" or
4462 * overlaps to the next one.
4464 if (!tcp_hdr(skb
)->syn
&& !tcp_hdr(skb
)->fin
&&
4465 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
4466 before(TCP_SKB_CB(skb
)->seq
, start
))) {
4467 end_of_skbs
= false;
4471 if (!skb_queue_is_last(list
, skb
)) {
4472 struct sk_buff
*next
= skb_queue_next(list
, skb
);
4474 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(next
)->seq
) {
4475 end_of_skbs
= false;
4480 /* Decided to skip this, advance start seq. */
4481 start
= TCP_SKB_CB(skb
)->end_seq
;
4483 if (end_of_skbs
|| tcp_hdr(skb
)->syn
|| tcp_hdr(skb
)->fin
)
4486 while (before(start
, end
)) {
4487 struct sk_buff
*nskb
;
4488 unsigned int header
= skb_headroom(skb
);
4489 int copy
= SKB_MAX_ORDER(header
, 0);
4491 /* Too big header? This can happen with IPv6. */
4494 if (end
- start
< copy
)
4496 nskb
= alloc_skb(copy
+ header
, GFP_ATOMIC
);
4500 skb_set_mac_header(nskb
, skb_mac_header(skb
) - skb
->head
);
4501 skb_set_network_header(nskb
, (skb_network_header(skb
) -
4503 skb_set_transport_header(nskb
, (skb_transport_header(skb
) -
4505 skb_reserve(nskb
, header
);
4506 memcpy(nskb
->head
, skb
->head
, header
);
4507 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
4508 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
4509 __skb_queue_before(list
, skb
, nskb
);
4510 skb_set_owner_r(nskb
, sk
);
4512 /* Copy data, releasing collapsed skbs. */
4514 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
4515 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
4519 size
= min(copy
, size
);
4520 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
4522 TCP_SKB_CB(nskb
)->end_seq
+= size
;
4526 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4527 skb
= tcp_collapse_one(sk
, skb
, list
);
4530 tcp_hdr(skb
)->syn
||
4538 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4539 * and tcp_collapse() them until all the queue is collapsed.
4541 static void tcp_collapse_ofo_queue(struct sock
*sk
)
4543 struct tcp_sock
*tp
= tcp_sk(sk
);
4544 struct sk_buff
*skb
= skb_peek(&tp
->out_of_order_queue
);
4545 struct sk_buff
*head
;
4551 start
= TCP_SKB_CB(skb
)->seq
;
4552 end
= TCP_SKB_CB(skb
)->end_seq
;
4556 struct sk_buff
*next
= NULL
;
4558 if (!skb_queue_is_last(&tp
->out_of_order_queue
, skb
))
4559 next
= skb_queue_next(&tp
->out_of_order_queue
, skb
);
4562 /* Segment is terminated when we see gap or when
4563 * we are at the end of all the queue. */
4565 after(TCP_SKB_CB(skb
)->seq
, end
) ||
4566 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
4567 tcp_collapse(sk
, &tp
->out_of_order_queue
,
4568 head
, skb
, start
, end
);
4572 /* Start new segment */
4573 start
= TCP_SKB_CB(skb
)->seq
;
4574 end
= TCP_SKB_CB(skb
)->end_seq
;
4576 if (before(TCP_SKB_CB(skb
)->seq
, start
))
4577 start
= TCP_SKB_CB(skb
)->seq
;
4578 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
4579 end
= TCP_SKB_CB(skb
)->end_seq
;
4585 * Purge the out-of-order queue.
4586 * Return true if queue was pruned.
4588 static bool tcp_prune_ofo_queue(struct sock
*sk
)
4590 struct tcp_sock
*tp
= tcp_sk(sk
);
4593 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4594 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_OFOPRUNED
);
4595 __skb_queue_purge(&tp
->out_of_order_queue
);
4597 /* Reset SACK state. A conforming SACK implementation will
4598 * do the same at a timeout based retransmit. When a connection
4599 * is in a sad state like this, we care only about integrity
4600 * of the connection not performance.
4602 if (tp
->rx_opt
.sack_ok
)
4603 tcp_sack_reset(&tp
->rx_opt
);
4610 /* Reduce allocated memory if we can, trying to get
4611 * the socket within its memory limits again.
4613 * Return less than zero if we should start dropping frames
4614 * until the socket owning process reads some of the data
4615 * to stabilize the situation.
4617 static int tcp_prune_queue(struct sock
*sk
)
4619 struct tcp_sock
*tp
= tcp_sk(sk
);
4621 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
4623 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PRUNECALLED
);
4625 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
4626 tcp_clamp_window(sk
);
4627 else if (sk_under_memory_pressure(sk
))
4628 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
4630 tcp_collapse_ofo_queue(sk
);
4631 if (!skb_queue_empty(&sk
->sk_receive_queue
))
4632 tcp_collapse(sk
, &sk
->sk_receive_queue
,
4633 skb_peek(&sk
->sk_receive_queue
),
4635 tp
->copied_seq
, tp
->rcv_nxt
);
4638 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4641 /* Collapsing did not help, destructive actions follow.
4642 * This must not ever occur. */
4644 tcp_prune_ofo_queue(sk
);
4646 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4649 /* If we are really being abused, tell the caller to silently
4650 * drop receive data on the floor. It will get retransmitted
4651 * and hopefully then we'll have sufficient space.
4653 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_RCVPRUNED
);
4655 /* Massive buffer overcommit. */
4660 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4661 * As additional protections, we do not touch cwnd in retransmission phases,
4662 * and if application hit its sndbuf limit recently.
4664 void tcp_cwnd_application_limited(struct sock
*sk
)
4666 struct tcp_sock
*tp
= tcp_sk(sk
);
4668 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
4669 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
4670 /* Limited by application or receiver window. */
4671 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
4672 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
4673 if (win_used
< tp
->snd_cwnd
) {
4674 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
4675 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
4677 tp
->snd_cwnd_used
= 0;
4679 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4682 static bool tcp_should_expand_sndbuf(const struct sock
*sk
)
4684 const struct tcp_sock
*tp
= tcp_sk(sk
);
4686 /* If the user specified a specific send buffer setting, do
4689 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
4692 /* If we are under global TCP memory pressure, do not expand. */
4693 if (sk_under_memory_pressure(sk
))
4696 /* If we are under soft global TCP memory pressure, do not expand. */
4697 if (sk_memory_allocated(sk
) >= sk_prot_mem_limits(sk
, 0))
4700 /* If we filled the congestion window, do not expand. */
4701 if (tp
->packets_out
>= tp
->snd_cwnd
)
4707 /* When incoming ACK allowed to free some skb from write_queue,
4708 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4709 * on the exit from tcp input handler.
4711 * PROBLEM: sndbuf expansion does not work well with largesend.
4713 static void tcp_new_space(struct sock
*sk
)
4715 struct tcp_sock
*tp
= tcp_sk(sk
);
4717 if (tcp_should_expand_sndbuf(sk
)) {
4718 int sndmem
= SKB_TRUESIZE(max_t(u32
,
4719 tp
->rx_opt
.mss_clamp
,
4722 int demanded
= max_t(unsigned int, tp
->snd_cwnd
,
4723 tp
->reordering
+ 1);
4724 sndmem
*= 2 * demanded
;
4725 if (sndmem
> sk
->sk_sndbuf
)
4726 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
4727 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4730 sk
->sk_write_space(sk
);
4733 static void tcp_check_space(struct sock
*sk
)
4735 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
4736 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
4737 if (sk
->sk_socket
&&
4738 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
4743 static inline void tcp_data_snd_check(struct sock
*sk
)
4745 tcp_push_pending_frames(sk
);
4746 tcp_check_space(sk
);
4750 * Check if sending an ack is needed.
4752 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
4754 struct tcp_sock
*tp
= tcp_sk(sk
);
4756 /* More than one full frame received... */
4757 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
&&
4758 /* ... and right edge of window advances far enough.
4759 * (tcp_recvmsg() will send ACK otherwise). Or...
4761 __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
4762 /* We ACK each frame or... */
4763 tcp_in_quickack_mode(sk
) ||
4764 /* We have out of order data. */
4765 (ofo_possible
&& skb_peek(&tp
->out_of_order_queue
))) {
4766 /* Then ack it now */
4769 /* Else, send delayed ack. */
4770 tcp_send_delayed_ack(sk
);
4774 static inline void tcp_ack_snd_check(struct sock
*sk
)
4776 if (!inet_csk_ack_scheduled(sk
)) {
4777 /* We sent a data segment already. */
4780 __tcp_ack_snd_check(sk
, 1);
4784 * This routine is only called when we have urgent data
4785 * signaled. Its the 'slow' part of tcp_urg. It could be
4786 * moved inline now as tcp_urg is only called from one
4787 * place. We handle URGent data wrong. We have to - as
4788 * BSD still doesn't use the correction from RFC961.
4789 * For 1003.1g we should support a new option TCP_STDURG to permit
4790 * either form (or just set the sysctl tcp_stdurg).
4793 static void tcp_check_urg(struct sock
*sk
, const struct tcphdr
*th
)
4795 struct tcp_sock
*tp
= tcp_sk(sk
);
4796 u32 ptr
= ntohs(th
->urg_ptr
);
4798 if (ptr
&& !sysctl_tcp_stdurg
)
4800 ptr
+= ntohl(th
->seq
);
4802 /* Ignore urgent data that we've already seen and read. */
4803 if (after(tp
->copied_seq
, ptr
))
4806 /* Do not replay urg ptr.
4808 * NOTE: interesting situation not covered by specs.
4809 * Misbehaving sender may send urg ptr, pointing to segment,
4810 * which we already have in ofo queue. We are not able to fetch
4811 * such data and will stay in TCP_URG_NOTYET until will be eaten
4812 * by recvmsg(). Seems, we are not obliged to handle such wicked
4813 * situations. But it is worth to think about possibility of some
4814 * DoSes using some hypothetical application level deadlock.
4816 if (before(ptr
, tp
->rcv_nxt
))
4819 /* Do we already have a newer (or duplicate) urgent pointer? */
4820 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
4823 /* Tell the world about our new urgent pointer. */
4826 /* We may be adding urgent data when the last byte read was
4827 * urgent. To do this requires some care. We cannot just ignore
4828 * tp->copied_seq since we would read the last urgent byte again
4829 * as data, nor can we alter copied_seq until this data arrives
4830 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4832 * NOTE. Double Dutch. Rendering to plain English: author of comment
4833 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4834 * and expect that both A and B disappear from stream. This is _wrong_.
4835 * Though this happens in BSD with high probability, this is occasional.
4836 * Any application relying on this is buggy. Note also, that fix "works"
4837 * only in this artificial test. Insert some normal data between A and B and we will
4838 * decline of BSD again. Verdict: it is better to remove to trap
4841 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
4842 !sock_flag(sk
, SOCK_URGINLINE
) && tp
->copied_seq
!= tp
->rcv_nxt
) {
4843 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
4845 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4846 __skb_unlink(skb
, &sk
->sk_receive_queue
);
4851 tp
->urg_data
= TCP_URG_NOTYET
;
4854 /* Disable header prediction. */
4858 /* This is the 'fast' part of urgent handling. */
4859 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, const struct tcphdr
*th
)
4861 struct tcp_sock
*tp
= tcp_sk(sk
);
4863 /* Check if we get a new urgent pointer - normally not. */
4865 tcp_check_urg(sk
, th
);
4867 /* Do we wait for any urgent data? - normally not... */
4868 if (tp
->urg_data
== TCP_URG_NOTYET
) {
4869 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
4872 /* Is the urgent pointer pointing into this packet? */
4873 if (ptr
< skb
->len
) {
4875 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
4877 tp
->urg_data
= TCP_URG_VALID
| tmp
;
4878 if (!sock_flag(sk
, SOCK_DEAD
))
4879 sk
->sk_data_ready(sk
, 0);
4884 static int tcp_copy_to_iovec(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
4886 struct tcp_sock
*tp
= tcp_sk(sk
);
4887 int chunk
= skb
->len
- hlen
;
4891 if (skb_csum_unnecessary(skb
))
4892 err
= skb_copy_datagram_iovec(skb
, hlen
, tp
->ucopy
.iov
, chunk
);
4894 err
= skb_copy_and_csum_datagram_iovec(skb
, hlen
,
4898 tp
->ucopy
.len
-= chunk
;
4899 tp
->copied_seq
+= chunk
;
4900 tcp_rcv_space_adjust(sk
);
4907 static __sum16
__tcp_checksum_complete_user(struct sock
*sk
,
4908 struct sk_buff
*skb
)
4912 if (sock_owned_by_user(sk
)) {
4914 result
= __tcp_checksum_complete(skb
);
4917 result
= __tcp_checksum_complete(skb
);
4922 static inline bool tcp_checksum_complete_user(struct sock
*sk
,
4923 struct sk_buff
*skb
)
4925 return !skb_csum_unnecessary(skb
) &&
4926 __tcp_checksum_complete_user(sk
, skb
);
4929 #ifdef CONFIG_NET_DMA
4930 static bool tcp_dma_try_early_copy(struct sock
*sk
, struct sk_buff
*skb
,
4933 struct tcp_sock
*tp
= tcp_sk(sk
);
4934 int chunk
= skb
->len
- hlen
;
4936 bool copied_early
= false;
4938 if (tp
->ucopy
.wakeup
)
4941 if (!tp
->ucopy
.dma_chan
&& tp
->ucopy
.pinned_list
)
4942 tp
->ucopy
.dma_chan
= net_dma_find_channel();
4944 if (tp
->ucopy
.dma_chan
&& skb_csum_unnecessary(skb
)) {
4946 dma_cookie
= dma_skb_copy_datagram_iovec(tp
->ucopy
.dma_chan
,
4948 tp
->ucopy
.iov
, chunk
,
4949 tp
->ucopy
.pinned_list
);
4954 tp
->ucopy
.dma_cookie
= dma_cookie
;
4955 copied_early
= true;
4957 tp
->ucopy
.len
-= chunk
;
4958 tp
->copied_seq
+= chunk
;
4959 tcp_rcv_space_adjust(sk
);
4961 if ((tp
->ucopy
.len
== 0) ||
4962 (tcp_flag_word(tcp_hdr(skb
)) & TCP_FLAG_PSH
) ||
4963 (atomic_read(&sk
->sk_rmem_alloc
) > (sk
->sk_rcvbuf
>> 1))) {
4964 tp
->ucopy
.wakeup
= 1;
4965 sk
->sk_data_ready(sk
, 0);
4967 } else if (chunk
> 0) {
4968 tp
->ucopy
.wakeup
= 1;
4969 sk
->sk_data_ready(sk
, 0);
4972 return copied_early
;
4974 #endif /* CONFIG_NET_DMA */
4976 /* Does PAWS and seqno based validation of an incoming segment, flags will
4977 * play significant role here.
4979 static bool tcp_validate_incoming(struct sock
*sk
, struct sk_buff
*skb
,
4980 const struct tcphdr
*th
, int syn_inerr
)
4982 struct tcp_sock
*tp
= tcp_sk(sk
);
4984 /* RFC1323: H1. Apply PAWS check first. */
4985 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
4986 tcp_paws_discard(sk
, skb
)) {
4988 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSESTABREJECTED
);
4989 tcp_send_dupack(sk
, skb
);
4992 /* Reset is accepted even if it did not pass PAWS. */
4995 /* Step 1: check sequence number */
4996 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4997 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4998 * (RST) segments are validated by checking their SEQ-fields."
4999 * And page 69: "If an incoming segment is not acceptable,
5000 * an acknowledgment should be sent in reply (unless the RST
5001 * bit is set, if so drop the segment and return)".
5006 tcp_send_dupack(sk
, skb
);
5011 /* Step 2: check RST bit */
5014 * If sequence number exactly matches RCV.NXT, then
5015 * RESET the connection
5017 * Send a challenge ACK
5019 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
)
5022 tcp_send_challenge_ack(sk
);
5026 /* step 3: check security and precedence [ignored] */
5028 /* step 4: Check for a SYN
5029 * RFC 5691 4.2 : Send a challenge ack
5034 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5035 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSYNCHALLENGE
);
5036 tcp_send_challenge_ack(sk
);
5048 * TCP receive function for the ESTABLISHED state.
5050 * It is split into a fast path and a slow path. The fast path is
5052 * - A zero window was announced from us - zero window probing
5053 * is only handled properly in the slow path.
5054 * - Out of order segments arrived.
5055 * - Urgent data is expected.
5056 * - There is no buffer space left
5057 * - Unexpected TCP flags/window values/header lengths are received
5058 * (detected by checking the TCP header against pred_flags)
5059 * - Data is sent in both directions. Fast path only supports pure senders
5060 * or pure receivers (this means either the sequence number or the ack
5061 * value must stay constant)
5062 * - Unexpected TCP option.
5064 * When these conditions are not satisfied it drops into a standard
5065 * receive procedure patterned after RFC793 to handle all cases.
5066 * The first three cases are guaranteed by proper pred_flags setting,
5067 * the rest is checked inline. Fast processing is turned on in
5068 * tcp_data_queue when everything is OK.
5070 int tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
5071 const struct tcphdr
*th
, unsigned int len
)
5073 struct tcp_sock
*tp
= tcp_sk(sk
);
5075 if (unlikely(sk
->sk_rx_dst
== NULL
))
5076 inet_csk(sk
)->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
5078 * Header prediction.
5079 * The code loosely follows the one in the famous
5080 * "30 instruction TCP receive" Van Jacobson mail.
5082 * Van's trick is to deposit buffers into socket queue
5083 * on a device interrupt, to call tcp_recv function
5084 * on the receive process context and checksum and copy
5085 * the buffer to user space. smart...
5087 * Our current scheme is not silly either but we take the
5088 * extra cost of the net_bh soft interrupt processing...
5089 * We do checksum and copy also but from device to kernel.
5092 tp
->rx_opt
.saw_tstamp
= 0;
5094 /* pred_flags is 0xS?10 << 16 + snd_wnd
5095 * if header_prediction is to be made
5096 * 'S' will always be tp->tcp_header_len >> 2
5097 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5098 * turn it off (when there are holes in the receive
5099 * space for instance)
5100 * PSH flag is ignored.
5103 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
5104 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
&&
5105 !after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
)) {
5106 int tcp_header_len
= tp
->tcp_header_len
;
5108 /* Timestamp header prediction: tcp_header_len
5109 * is automatically equal to th->doff*4 due to pred_flags
5113 /* Check timestamp */
5114 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
5115 /* No? Slow path! */
5116 if (!tcp_parse_aligned_timestamp(tp
, th
))
5119 /* If PAWS failed, check it more carefully in slow path */
5120 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
5123 /* DO NOT update ts_recent here, if checksum fails
5124 * and timestamp was corrupted part, it will result
5125 * in a hung connection since we will drop all
5126 * future packets due to the PAWS test.
5130 if (len
<= tcp_header_len
) {
5131 /* Bulk data transfer: sender */
5132 if (len
== tcp_header_len
) {
5133 /* Predicted packet is in window by definition.
5134 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5135 * Hence, check seq<=rcv_wup reduces to:
5137 if (tcp_header_len
==
5138 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5139 tp
->rcv_nxt
== tp
->rcv_wup
)
5140 tcp_store_ts_recent(tp
);
5142 /* We know that such packets are checksummed
5145 tcp_ack(sk
, skb
, 0);
5147 tcp_data_snd_check(sk
);
5149 } else { /* Header too small */
5150 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5155 int copied_early
= 0;
5156 bool fragstolen
= false;
5158 if (tp
->copied_seq
== tp
->rcv_nxt
&&
5159 len
- tcp_header_len
<= tp
->ucopy
.len
) {
5160 #ifdef CONFIG_NET_DMA
5161 if (tp
->ucopy
.task
== current
&&
5162 sock_owned_by_user(sk
) &&
5163 tcp_dma_try_early_copy(sk
, skb
, tcp_header_len
)) {
5168 if (tp
->ucopy
.task
== current
&&
5169 sock_owned_by_user(sk
) && !copied_early
) {
5170 __set_current_state(TASK_RUNNING
);
5172 if (!tcp_copy_to_iovec(sk
, skb
, tcp_header_len
))
5176 /* Predicted packet is in window by definition.
5177 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5178 * Hence, check seq<=rcv_wup reduces to:
5180 if (tcp_header_len
==
5181 (sizeof(struct tcphdr
) +
5182 TCPOLEN_TSTAMP_ALIGNED
) &&
5183 tp
->rcv_nxt
== tp
->rcv_wup
)
5184 tcp_store_ts_recent(tp
);
5186 tcp_rcv_rtt_measure_ts(sk
, skb
);
5188 __skb_pull(skb
, tcp_header_len
);
5189 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
5190 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITSTOUSER
);
5193 tcp_cleanup_rbuf(sk
, skb
->len
);
5196 if (tcp_checksum_complete_user(sk
, skb
))
5199 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
5202 /* Predicted packet is in window by definition.
5203 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5204 * Hence, check seq<=rcv_wup reduces to:
5206 if (tcp_header_len
==
5207 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5208 tp
->rcv_nxt
== tp
->rcv_wup
)
5209 tcp_store_ts_recent(tp
);
5211 tcp_rcv_rtt_measure_ts(sk
, skb
);
5213 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITS
);
5215 /* Bulk data transfer: receiver */
5216 eaten
= tcp_queue_rcv(sk
, skb
, tcp_header_len
,
5220 tcp_event_data_recv(sk
, skb
);
5222 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
5223 /* Well, only one small jumplet in fast path... */
5224 tcp_ack(sk
, skb
, FLAG_DATA
);
5225 tcp_data_snd_check(sk
);
5226 if (!inet_csk_ack_scheduled(sk
))
5230 if (!copied_early
|| tp
->rcv_nxt
!= tp
->rcv_wup
)
5231 __tcp_ack_snd_check(sk
, 0);
5233 #ifdef CONFIG_NET_DMA
5235 __skb_queue_tail(&sk
->sk_async_wait_queue
, skb
);
5239 kfree_skb_partial(skb
, fragstolen
);
5240 sk
->sk_data_ready(sk
, 0);
5246 if (len
< (th
->doff
<< 2) || tcp_checksum_complete_user(sk
, skb
))
5249 if (!th
->ack
&& !th
->rst
)
5253 * Standard slow path.
5256 if (!tcp_validate_incoming(sk
, skb
, th
, 1))
5260 if (tcp_ack(sk
, skb
, FLAG_SLOWPATH
| FLAG_UPDATE_TS_RECENT
) < 0)
5263 tcp_rcv_rtt_measure_ts(sk
, skb
);
5265 /* Process urgent data. */
5266 tcp_urg(sk
, skb
, th
);
5268 /* step 7: process the segment text */
5269 tcp_data_queue(sk
, skb
);
5271 tcp_data_snd_check(sk
);
5272 tcp_ack_snd_check(sk
);
5276 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_CSUMERRORS
);
5277 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5283 EXPORT_SYMBOL(tcp_rcv_established
);
5285 void tcp_finish_connect(struct sock
*sk
, struct sk_buff
*skb
)
5287 struct tcp_sock
*tp
= tcp_sk(sk
);
5288 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5290 tcp_set_state(sk
, TCP_ESTABLISHED
);
5293 icsk
->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
5294 security_inet_conn_established(sk
, skb
);
5297 /* Make sure socket is routed, for correct metrics. */
5298 icsk
->icsk_af_ops
->rebuild_header(sk
);
5300 tcp_init_metrics(sk
);
5302 tcp_init_congestion_control(sk
);
5304 /* Prevent spurious tcp_cwnd_restart() on first data
5307 tp
->lsndtime
= tcp_time_stamp
;
5309 tcp_init_buffer_space(sk
);
5311 if (sock_flag(sk
, SOCK_KEEPOPEN
))
5312 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
5314 if (!tp
->rx_opt
.snd_wscale
)
5315 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
5319 if (!sock_flag(sk
, SOCK_DEAD
)) {
5320 sk
->sk_state_change(sk
);
5321 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5325 static bool tcp_rcv_fastopen_synack(struct sock
*sk
, struct sk_buff
*synack
,
5326 struct tcp_fastopen_cookie
*cookie
)
5328 struct tcp_sock
*tp
= tcp_sk(sk
);
5329 struct sk_buff
*data
= tp
->syn_data
? tcp_write_queue_head(sk
) : NULL
;
5330 u16 mss
= tp
->rx_opt
.mss_clamp
;
5333 if (mss
== tp
->rx_opt
.user_mss
) {
5334 struct tcp_options_received opt
;
5336 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5337 tcp_clear_options(&opt
);
5338 opt
.user_mss
= opt
.mss_clamp
= 0;
5339 tcp_parse_options(synack
, &opt
, 0, NULL
);
5340 mss
= opt
.mss_clamp
;
5343 if (!tp
->syn_fastopen
) /* Ignore an unsolicited cookie */
5346 /* The SYN-ACK neither has cookie nor acknowledges the data. Presumably
5347 * the remote receives only the retransmitted (regular) SYNs: either
5348 * the original SYN-data or the corresponding SYN-ACK is lost.
5350 syn_drop
= (cookie
->len
<= 0 && data
&& tp
->total_retrans
);
5352 tcp_fastopen_cache_set(sk
, mss
, cookie
, syn_drop
);
5354 if (data
) { /* Retransmit unacked data in SYN */
5355 tcp_for_write_queue_from(data
, sk
) {
5356 if (data
== tcp_send_head(sk
) ||
5357 __tcp_retransmit_skb(sk
, data
))
5363 tp
->syn_data_acked
= tp
->syn_data
;
5367 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5368 const struct tcphdr
*th
, unsigned int len
)
5370 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5371 struct tcp_sock
*tp
= tcp_sk(sk
);
5372 struct tcp_fastopen_cookie foc
= { .len
= -1 };
5373 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
5375 tcp_parse_options(skb
, &tp
->rx_opt
, 0, &foc
);
5376 if (tp
->rx_opt
.saw_tstamp
)
5377 tp
->rx_opt
.rcv_tsecr
-= tp
->tsoffset
;
5381 * "If the state is SYN-SENT then
5382 * first check the ACK bit
5383 * If the ACK bit is set
5384 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5385 * a reset (unless the RST bit is set, if so drop
5386 * the segment and return)"
5388 if (!after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_una
) ||
5389 after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
))
5390 goto reset_and_undo
;
5392 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
5393 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
5395 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSACTIVEREJECTED
);
5396 goto reset_and_undo
;
5399 /* Now ACK is acceptable.
5401 * "If the RST bit is set
5402 * If the ACK was acceptable then signal the user "error:
5403 * connection reset", drop the segment, enter CLOSED state,
5404 * delete TCB, and return."
5413 * "fifth, if neither of the SYN or RST bits is set then
5414 * drop the segment and return."
5420 goto discard_and_undo
;
5423 * "If the SYN bit is on ...
5424 * are acceptable then ...
5425 * (our SYN has been ACKed), change the connection
5426 * state to ESTABLISHED..."
5429 TCP_ECN_rcv_synack(tp
, th
);
5431 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5432 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
5434 /* Ok.. it's good. Set up sequence numbers and
5435 * move to established.
5437 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5438 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5440 /* RFC1323: The window in SYN & SYN/ACK segments is
5443 tp
->snd_wnd
= ntohs(th
->window
);
5445 if (!tp
->rx_opt
.wscale_ok
) {
5446 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
5447 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
5450 if (tp
->rx_opt
.saw_tstamp
) {
5451 tp
->rx_opt
.tstamp_ok
= 1;
5452 tp
->tcp_header_len
=
5453 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5454 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5455 tcp_store_ts_recent(tp
);
5457 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5460 if (tcp_is_sack(tp
) && sysctl_tcp_fack
)
5461 tcp_enable_fack(tp
);
5464 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5465 tcp_initialize_rcv_mss(sk
);
5467 /* Remember, tcp_poll() does not lock socket!
5468 * Change state from SYN-SENT only after copied_seq
5469 * is initialized. */
5470 tp
->copied_seq
= tp
->rcv_nxt
;
5474 tcp_finish_connect(sk
, skb
);
5476 if ((tp
->syn_fastopen
|| tp
->syn_data
) &&
5477 tcp_rcv_fastopen_synack(sk
, skb
, &foc
))
5480 if (sk
->sk_write_pending
||
5481 icsk
->icsk_accept_queue
.rskq_defer_accept
||
5482 icsk
->icsk_ack
.pingpong
) {
5483 /* Save one ACK. Data will be ready after
5484 * several ticks, if write_pending is set.
5486 * It may be deleted, but with this feature tcpdumps
5487 * look so _wonderfully_ clever, that I was not able
5488 * to stand against the temptation 8) --ANK
5490 inet_csk_schedule_ack(sk
);
5491 icsk
->icsk_ack
.lrcvtime
= tcp_time_stamp
;
5492 tcp_enter_quickack_mode(sk
);
5493 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
5494 TCP_DELACK_MAX
, TCP_RTO_MAX
);
5505 /* No ACK in the segment */
5509 * "If the RST bit is set
5511 * Otherwise (no ACK) drop the segment and return."
5514 goto discard_and_undo
;
5518 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&&
5519 tcp_paws_reject(&tp
->rx_opt
, 0))
5520 goto discard_and_undo
;
5523 /* We see SYN without ACK. It is attempt of
5524 * simultaneous connect with crossed SYNs.
5525 * Particularly, it can be connect to self.
5527 tcp_set_state(sk
, TCP_SYN_RECV
);
5529 if (tp
->rx_opt
.saw_tstamp
) {
5530 tp
->rx_opt
.tstamp_ok
= 1;
5531 tcp_store_ts_recent(tp
);
5532 tp
->tcp_header_len
=
5533 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5535 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5538 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5539 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5541 /* RFC1323: The window in SYN & SYN/ACK segments is
5544 tp
->snd_wnd
= ntohs(th
->window
);
5545 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
5546 tp
->max_window
= tp
->snd_wnd
;
5548 TCP_ECN_rcv_syn(tp
, th
);
5551 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5552 tcp_initialize_rcv_mss(sk
);
5554 tcp_send_synack(sk
);
5556 /* Note, we could accept data and URG from this segment.
5557 * There are no obstacles to make this (except that we must
5558 * either change tcp_recvmsg() to prevent it from returning data
5559 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5561 * However, if we ignore data in ACKless segments sometimes,
5562 * we have no reasons to accept it sometimes.
5563 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5564 * is not flawless. So, discard packet for sanity.
5565 * Uncomment this return to process the data.
5572 /* "fifth, if neither of the SYN or RST bits is set then
5573 * drop the segment and return."
5577 tcp_clear_options(&tp
->rx_opt
);
5578 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5582 tcp_clear_options(&tp
->rx_opt
);
5583 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5588 * This function implements the receiving procedure of RFC 793 for
5589 * all states except ESTABLISHED and TIME_WAIT.
5590 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5591 * address independent.
5594 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5595 const struct tcphdr
*th
, unsigned int len
)
5597 struct tcp_sock
*tp
= tcp_sk(sk
);
5598 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5599 struct request_sock
*req
;
5602 tp
->rx_opt
.saw_tstamp
= 0;
5604 switch (sk
->sk_state
) {
5618 if (icsk
->icsk_af_ops
->conn_request(sk
, skb
) < 0)
5621 /* Now we have several options: In theory there is
5622 * nothing else in the frame. KA9Q has an option to
5623 * send data with the syn, BSD accepts data with the
5624 * syn up to the [to be] advertised window and
5625 * Solaris 2.1 gives you a protocol error. For now
5626 * we just ignore it, that fits the spec precisely
5627 * and avoids incompatibilities. It would be nice in
5628 * future to drop through and process the data.
5630 * Now that TTCP is starting to be used we ought to
5632 * But, this leaves one open to an easy denial of
5633 * service attack, and SYN cookies can't defend
5634 * against this problem. So, we drop the data
5635 * in the interest of security over speed unless
5636 * it's still in use.
5644 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
, len
);
5648 /* Do step6 onward by hand. */
5649 tcp_urg(sk
, skb
, th
);
5651 tcp_data_snd_check(sk
);
5655 req
= tp
->fastopen_rsk
;
5657 WARN_ON_ONCE(sk
->sk_state
!= TCP_SYN_RECV
&&
5658 sk
->sk_state
!= TCP_FIN_WAIT1
);
5660 if (tcp_check_req(sk
, skb
, req
, NULL
, true) == NULL
)
5664 if (!th
->ack
&& !th
->rst
)
5667 if (!tcp_validate_incoming(sk
, skb
, th
, 0))
5670 /* step 5: check the ACK field */
5672 int acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
|
5673 FLAG_UPDATE_TS_RECENT
) > 0;
5675 switch (sk
->sk_state
) {
5678 /* Once we leave TCP_SYN_RECV, we no longer
5679 * need req so release it.
5682 tcp_synack_rtt_meas(sk
, req
);
5683 tp
->total_retrans
= req
->num_retrans
;
5685 reqsk_fastopen_remove(sk
, req
, false);
5687 /* Make sure socket is routed, for
5690 icsk
->icsk_af_ops
->rebuild_header(sk
);
5691 tcp_init_congestion_control(sk
);
5694 tcp_init_buffer_space(sk
);
5695 tp
->copied_seq
= tp
->rcv_nxt
;
5698 tcp_set_state(sk
, TCP_ESTABLISHED
);
5699 sk
->sk_state_change(sk
);
5701 /* Note, that this wakeup is only for marginal
5702 * crossed SYN case. Passively open sockets
5703 * are not waked up, because sk->sk_sleep ==
5704 * NULL and sk->sk_socket == NULL.
5708 SOCK_WAKE_IO
, POLL_OUT
);
5710 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
5711 tp
->snd_wnd
= ntohs(th
->window
) <<
5712 tp
->rx_opt
.snd_wscale
;
5713 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5715 if (tp
->rx_opt
.tstamp_ok
)
5716 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5719 /* Re-arm the timer because data may
5720 * have been sent out. This is similar
5721 * to the regular data transmission case
5722 * when new data has just been ack'ed.
5724 * (TFO) - we could try to be more
5725 * aggressive and retranmitting any data
5726 * sooner based on when they were sent
5731 tcp_init_metrics(sk
);
5733 /* Prevent spurious tcp_cwnd_restart() on
5734 * first data packet.
5736 tp
->lsndtime
= tcp_time_stamp
;
5738 tcp_initialize_rcv_mss(sk
);
5739 tcp_fast_path_on(tp
);
5746 /* If we enter the TCP_FIN_WAIT1 state and we are a
5747 * Fast Open socket and this is the first acceptable
5748 * ACK we have received, this would have acknowledged
5749 * our SYNACK so stop the SYNACK timer.
5752 /* Return RST if ack_seq is invalid.
5753 * Note that RFC793 only says to generate a
5754 * DUPACK for it but for TCP Fast Open it seems
5755 * better to treat this case like TCP_SYN_RECV
5760 /* We no longer need the request sock. */
5761 reqsk_fastopen_remove(sk
, req
, false);
5764 if (tp
->snd_una
== tp
->write_seq
) {
5765 struct dst_entry
*dst
;
5767 tcp_set_state(sk
, TCP_FIN_WAIT2
);
5768 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
5770 dst
= __sk_dst_get(sk
);
5774 if (!sock_flag(sk
, SOCK_DEAD
))
5775 /* Wake up lingering close() */
5776 sk
->sk_state_change(sk
);
5780 if (tp
->linger2
< 0 ||
5781 (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5782 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
))) {
5784 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
5788 tmo
= tcp_fin_time(sk
);
5789 if (tmo
> TCP_TIMEWAIT_LEN
) {
5790 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
5791 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
5792 /* Bad case. We could lose such FIN otherwise.
5793 * It is not a big problem, but it looks confusing
5794 * and not so rare event. We still can lose it now,
5795 * if it spins in bh_lock_sock(), but it is really
5798 inet_csk_reset_keepalive_timer(sk
, tmo
);
5800 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
5808 if (tp
->snd_una
== tp
->write_seq
) {
5809 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
5815 if (tp
->snd_una
== tp
->write_seq
) {
5816 tcp_update_metrics(sk
);
5824 /* step 6: check the URG bit */
5825 tcp_urg(sk
, skb
, th
);
5827 /* step 7: process the segment text */
5828 switch (sk
->sk_state
) {
5829 case TCP_CLOSE_WAIT
:
5832 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
5836 /* RFC 793 says to queue data in these states,
5837 * RFC 1122 says we MUST send a reset.
5838 * BSD 4.4 also does reset.
5840 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
5841 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5842 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
5843 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
5849 case TCP_ESTABLISHED
:
5850 tcp_data_queue(sk
, skb
);
5855 /* tcp_data could move socket to TIME-WAIT */
5856 if (sk
->sk_state
!= TCP_CLOSE
) {
5857 tcp_data_snd_check(sk
);
5858 tcp_ack_snd_check(sk
);
5867 EXPORT_SYMBOL(tcp_rcv_state_process
);