2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
23 * Pedro Roque : Fast Retransmit/Recovery.
25 * Retransmit queue handled by TCP.
26 * Better retransmit timer handling.
27 * New congestion avoidance.
31 * Eric : Fast Retransmit.
32 * Randy Scott : MSS option defines.
33 * Eric Schenk : Fixes to slow start algorithm.
34 * Eric Schenk : Yet another double ACK bug.
35 * Eric Schenk : Delayed ACK bug fixes.
36 * Eric Schenk : Floyd style fast retrans war avoidance.
37 * David S. Miller : Don't allow zero congestion window.
38 * Eric Schenk : Fix retransmitter so that it sends
39 * next packet on ack of previous packet.
40 * Andi Kleen : Moved open_request checking here
41 * and process RSTs for open_requests.
42 * Andi Kleen : Better prune_queue, and other fixes.
43 * Andrey Savochkin: Fix RTT measurements in the presence of
45 * Andrey Savochkin: Check sequence numbers correctly when
46 * removing SACKs due to in sequence incoming
48 * Andi Kleen: Make sure we never ack data there is not
49 * enough room for. Also make this condition
50 * a fatal error if it might still happen.
51 * Andi Kleen: Add tcp_measure_rcv_mss to make
52 * connections with MSS<min(MTU,ann. MSS)
53 * work without delayed acks.
54 * Andi Kleen: Process packets with PSH set in the
56 * J Hadi Salim: ECN support
59 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
60 * engine. Lots of bugs are found.
61 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
64 #define pr_fmt(fmt) "TCP: " fmt
67 #include <linux/slab.h>
68 #include <linux/module.h>
69 #include <linux/sysctl.h>
70 #include <linux/kernel.h>
71 #include <linux/reciprocal_div.h>
74 #include <net/inet_common.h>
75 #include <linux/ipsec.h>
76 #include <asm/unaligned.h>
77 #include <net/netdma.h>
79 int sysctl_tcp_timestamps __read_mostly
= 1;
80 int sysctl_tcp_window_scaling __read_mostly
= 1;
81 int sysctl_tcp_sack __read_mostly
= 1;
82 int sysctl_tcp_fack __read_mostly
= 1;
83 int sysctl_tcp_reordering __read_mostly
= TCP_FASTRETRANS_THRESH
;
84 EXPORT_SYMBOL(sysctl_tcp_reordering
);
85 int sysctl_tcp_dsack __read_mostly
= 1;
86 int sysctl_tcp_app_win __read_mostly
= 31;
87 int sysctl_tcp_adv_win_scale __read_mostly
= 1;
88 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale
);
90 /* rfc5961 challenge ack rate limiting */
91 int sysctl_tcp_challenge_ack_limit
= 1000;
93 int sysctl_tcp_stdurg __read_mostly
;
94 int sysctl_tcp_rfc1337 __read_mostly
;
95 int sysctl_tcp_max_orphans __read_mostly
= NR_FILE
;
96 int sysctl_tcp_frto __read_mostly
= 2;
98 int sysctl_tcp_thin_dupack __read_mostly
;
100 int sysctl_tcp_moderate_rcvbuf __read_mostly
= 1;
101 int sysctl_tcp_early_retrans __read_mostly
= 3;
102 int sysctl_tcp_default_init_rwnd __read_mostly
= TCP_DEFAULT_INIT_RCVWND
;
104 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
105 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
106 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
107 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
108 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
109 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
110 #define FLAG_ECE 0x40 /* ECE in this ACK */
111 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
112 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
113 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
114 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
115 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
116 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
118 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
119 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
120 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
121 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
123 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
124 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
126 /* Adapt the MSS value used to make delayed ack decision to the
129 static void tcp_measure_rcv_mss(struct sock
*sk
, const struct sk_buff
*skb
)
131 struct inet_connection_sock
*icsk
= inet_csk(sk
);
132 const unsigned int lss
= icsk
->icsk_ack
.last_seg_size
;
135 icsk
->icsk_ack
.last_seg_size
= 0;
137 /* skb->len may jitter because of SACKs, even if peer
138 * sends good full-sized frames.
140 len
= skb_shinfo(skb
)->gso_size
? : skb
->len
;
141 if (len
>= icsk
->icsk_ack
.rcv_mss
) {
142 icsk
->icsk_ack
.rcv_mss
= len
;
144 /* Otherwise, we make more careful check taking into account,
145 * that SACKs block is variable.
147 * "len" is invariant segment length, including TCP header.
149 len
+= skb
->data
- skb_transport_header(skb
);
150 if (len
>= TCP_MSS_DEFAULT
+ sizeof(struct tcphdr
) ||
151 /* If PSH is not set, packet should be
152 * full sized, provided peer TCP is not badly broken.
153 * This observation (if it is correct 8)) allows
154 * to handle super-low mtu links fairly.
156 (len
>= TCP_MIN_MSS
+ sizeof(struct tcphdr
) &&
157 !(tcp_flag_word(tcp_hdr(skb
)) & TCP_REMNANT
))) {
158 /* Subtract also invariant (if peer is RFC compliant),
159 * tcp header plus fixed timestamp option length.
160 * Resulting "len" is MSS free of SACK jitter.
162 len
-= tcp_sk(sk
)->tcp_header_len
;
163 icsk
->icsk_ack
.last_seg_size
= len
;
165 icsk
->icsk_ack
.rcv_mss
= len
;
169 if (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
)
170 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED2
;
171 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
175 static void tcp_incr_quickack(struct sock
*sk
)
177 struct inet_connection_sock
*icsk
= inet_csk(sk
);
178 unsigned int quickacks
= tcp_sk(sk
)->rcv_wnd
/ (2 * icsk
->icsk_ack
.rcv_mss
);
182 if (quickacks
> icsk
->icsk_ack
.quick
)
183 icsk
->icsk_ack
.quick
= min(quickacks
, TCP_MAX_QUICKACKS
);
186 static void tcp_enter_quickack_mode(struct sock
*sk
)
188 struct inet_connection_sock
*icsk
= inet_csk(sk
);
189 tcp_incr_quickack(sk
);
190 icsk
->icsk_ack
.pingpong
= 0;
191 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
194 /* Send ACKs quickly, if "quick" count is not exhausted
195 * and the session is not interactive.
198 static inline bool tcp_in_quickack_mode(const struct sock
*sk
)
200 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
202 return icsk
->icsk_ack
.quick
&& !icsk
->icsk_ack
.pingpong
;
205 static inline void TCP_ECN_queue_cwr(struct tcp_sock
*tp
)
207 if (tp
->ecn_flags
& TCP_ECN_OK
)
208 tp
->ecn_flags
|= TCP_ECN_QUEUE_CWR
;
211 static inline void TCP_ECN_accept_cwr(struct tcp_sock
*tp
, const struct sk_buff
*skb
)
213 if (tcp_hdr(skb
)->cwr
)
214 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
217 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock
*tp
)
219 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
222 static inline void TCP_ECN_check_ce(struct tcp_sock
*tp
, const struct sk_buff
*skb
)
224 if (!(tp
->ecn_flags
& TCP_ECN_OK
))
227 switch (TCP_SKB_CB(skb
)->ip_dsfield
& INET_ECN_MASK
) {
228 case INET_ECN_NOT_ECT
:
229 /* Funny extension: if ECT is not set on a segment,
230 * and we already seen ECT on a previous segment,
231 * it is probably a retransmit.
233 if (tp
->ecn_flags
& TCP_ECN_SEEN
)
234 tcp_enter_quickack_mode((struct sock
*)tp
);
237 if (!(tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)) {
238 /* Better not delay acks, sender can have a very low cwnd */
239 tcp_enter_quickack_mode((struct sock
*)tp
);
240 tp
->ecn_flags
|= TCP_ECN_DEMAND_CWR
;
244 tp
->ecn_flags
|= TCP_ECN_SEEN
;
248 static inline void TCP_ECN_rcv_synack(struct tcp_sock
*tp
, const struct tcphdr
*th
)
250 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| th
->cwr
))
251 tp
->ecn_flags
&= ~TCP_ECN_OK
;
254 static inline void TCP_ECN_rcv_syn(struct tcp_sock
*tp
, const struct tcphdr
*th
)
256 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| !th
->cwr
))
257 tp
->ecn_flags
&= ~TCP_ECN_OK
;
260 static bool TCP_ECN_rcv_ecn_echo(const struct tcp_sock
*tp
, const struct tcphdr
*th
)
262 if (th
->ece
&& !th
->syn
&& (tp
->ecn_flags
& TCP_ECN_OK
))
267 /* Buffer size and advertised window tuning.
269 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
272 static void tcp_fixup_sndbuf(struct sock
*sk
)
274 int sndmem
= SKB_TRUESIZE(tcp_sk(sk
)->rx_opt
.mss_clamp
+ MAX_TCP_HEADER
);
276 sndmem
*= TCP_INIT_CWND
;
277 if (sk
->sk_sndbuf
< sndmem
)
278 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
281 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
283 * All tcp_full_space() is split to two parts: "network" buffer, allocated
284 * forward and advertised in receiver window (tp->rcv_wnd) and
285 * "application buffer", required to isolate scheduling/application
286 * latencies from network.
287 * window_clamp is maximal advertised window. It can be less than
288 * tcp_full_space(), in this case tcp_full_space() - window_clamp
289 * is reserved for "application" buffer. The less window_clamp is
290 * the smoother our behaviour from viewpoint of network, but the lower
291 * throughput and the higher sensitivity of the connection to losses. 8)
293 * rcv_ssthresh is more strict window_clamp used at "slow start"
294 * phase to predict further behaviour of this connection.
295 * It is used for two goals:
296 * - to enforce header prediction at sender, even when application
297 * requires some significant "application buffer". It is check #1.
298 * - to prevent pruning of receive queue because of misprediction
299 * of receiver window. Check #2.
301 * The scheme does not work when sender sends good segments opening
302 * window and then starts to feed us spaghetti. But it should work
303 * in common situations. Otherwise, we have to rely on queue collapsing.
306 /* Slow part of check#2. */
307 static int __tcp_grow_window(const struct sock
*sk
, const struct sk_buff
*skb
)
309 struct tcp_sock
*tp
= tcp_sk(sk
);
311 int truesize
= tcp_win_from_space(skb
->truesize
) >> 1;
312 int window
= tcp_win_from_space(sysctl_tcp_rmem
[2]) >> 1;
314 while (tp
->rcv_ssthresh
<= window
) {
315 if (truesize
<= skb
->len
)
316 return 2 * inet_csk(sk
)->icsk_ack
.rcv_mss
;
324 static void tcp_grow_window(struct sock
*sk
, const struct sk_buff
*skb
)
326 struct tcp_sock
*tp
= tcp_sk(sk
);
329 if (tp
->rcv_ssthresh
< tp
->window_clamp
&&
330 (int)tp
->rcv_ssthresh
< tcp_space(sk
) &&
331 !sk_under_memory_pressure(sk
)) {
334 /* Check #2. Increase window, if skb with such overhead
335 * will fit to rcvbuf in future.
337 if (tcp_win_from_space(skb
->truesize
) <= skb
->len
)
338 incr
= 2 * tp
->advmss
;
340 incr
= __tcp_grow_window(sk
, skb
);
343 incr
= max_t(int, incr
, 2 * skb
->len
);
344 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
+ incr
,
346 inet_csk(sk
)->icsk_ack
.quick
|= 1;
351 /* 3. Tuning rcvbuf, when connection enters established state. */
353 static void tcp_fixup_rcvbuf(struct sock
*sk
)
355 u32 mss
= tcp_sk(sk
)->advmss
;
356 u32 icwnd
= sysctl_tcp_default_init_rwnd
;
359 /* Limit to 10 segments if mss <= 1460,
360 * or 14600/mss segments, with a minimum of two segments.
363 icwnd
= max_t(u32
, (1460 * icwnd
) / mss
, 2);
365 rcvmem
= SKB_TRUESIZE(mss
+ MAX_TCP_HEADER
);
366 while (tcp_win_from_space(rcvmem
) < mss
)
371 if (sk
->sk_rcvbuf
< rcvmem
)
372 sk
->sk_rcvbuf
= min(rcvmem
, sysctl_tcp_rmem
[2]);
375 /* 4. Try to fixup all. It is made immediately after connection enters
378 void tcp_init_buffer_space(struct sock
*sk
)
380 struct tcp_sock
*tp
= tcp_sk(sk
);
383 if (!(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
))
384 tcp_fixup_rcvbuf(sk
);
385 if (!(sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
))
386 tcp_fixup_sndbuf(sk
);
388 tp
->rcvq_space
.space
= tp
->rcv_wnd
;
390 maxwin
= tcp_full_space(sk
);
392 if (tp
->window_clamp
>= maxwin
) {
393 tp
->window_clamp
= maxwin
;
395 if (sysctl_tcp_app_win
&& maxwin
> 4 * tp
->advmss
)
396 tp
->window_clamp
= max(maxwin
-
397 (maxwin
>> sysctl_tcp_app_win
),
401 /* Force reservation of one segment. */
402 if (sysctl_tcp_app_win
&&
403 tp
->window_clamp
> 2 * tp
->advmss
&&
404 tp
->window_clamp
+ tp
->advmss
> maxwin
)
405 tp
->window_clamp
= max(2 * tp
->advmss
, maxwin
- tp
->advmss
);
407 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, tp
->window_clamp
);
408 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
411 /* 5. Recalculate window clamp after socket hit its memory bounds. */
412 static void tcp_clamp_window(struct sock
*sk
)
414 struct tcp_sock
*tp
= tcp_sk(sk
);
415 struct inet_connection_sock
*icsk
= inet_csk(sk
);
417 icsk
->icsk_ack
.quick
= 0;
419 if (sk
->sk_rcvbuf
< sysctl_tcp_rmem
[2] &&
420 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
) &&
421 !sk_under_memory_pressure(sk
) &&
422 sk_memory_allocated(sk
) < sk_prot_mem_limits(sk
, 0)) {
423 sk
->sk_rcvbuf
= min(atomic_read(&sk
->sk_rmem_alloc
),
426 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
)
427 tp
->rcv_ssthresh
= min(tp
->window_clamp
, 2U * tp
->advmss
);
430 /* Initialize RCV_MSS value.
431 * RCV_MSS is an our guess about MSS used by the peer.
432 * We haven't any direct information about the MSS.
433 * It's better to underestimate the RCV_MSS rather than overestimate.
434 * Overestimations make us ACKing less frequently than needed.
435 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
437 void tcp_initialize_rcv_mss(struct sock
*sk
)
439 const struct tcp_sock
*tp
= tcp_sk(sk
);
440 unsigned int hint
= min_t(unsigned int, tp
->advmss
, tp
->mss_cache
);
442 hint
= min(hint
, tp
->rcv_wnd
/ 2);
443 hint
= min(hint
, TCP_MSS_DEFAULT
);
444 hint
= max(hint
, TCP_MIN_MSS
);
446 inet_csk(sk
)->icsk_ack
.rcv_mss
= hint
;
448 EXPORT_SYMBOL(tcp_initialize_rcv_mss
);
450 /* Receiver "autotuning" code.
452 * The algorithm for RTT estimation w/o timestamps is based on
453 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
454 * <http://public.lanl.gov/radiant/pubs.html#DRS>
456 * More detail on this code can be found at
457 * <http://staff.psc.edu/jheffner/>,
458 * though this reference is out of date. A new paper
461 static void tcp_rcv_rtt_update(struct tcp_sock
*tp
, u32 sample
, int win_dep
)
463 u32 new_sample
= tp
->rcv_rtt_est
.rtt
;
469 if (new_sample
!= 0) {
470 /* If we sample in larger samples in the non-timestamp
471 * case, we could grossly overestimate the RTT especially
472 * with chatty applications or bulk transfer apps which
473 * are stalled on filesystem I/O.
475 * Also, since we are only going for a minimum in the
476 * non-timestamp case, we do not smooth things out
477 * else with timestamps disabled convergence takes too
481 m
-= (new_sample
>> 3);
489 /* No previous measure. */
493 if (tp
->rcv_rtt_est
.rtt
!= new_sample
)
494 tp
->rcv_rtt_est
.rtt
= new_sample
;
497 static inline void tcp_rcv_rtt_measure(struct tcp_sock
*tp
)
499 if (tp
->rcv_rtt_est
.time
== 0)
501 if (before(tp
->rcv_nxt
, tp
->rcv_rtt_est
.seq
))
503 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rcv_rtt_est
.time
, 1);
506 tp
->rcv_rtt_est
.seq
= tp
->rcv_nxt
+ tp
->rcv_wnd
;
507 tp
->rcv_rtt_est
.time
= tcp_time_stamp
;
510 static inline void tcp_rcv_rtt_measure_ts(struct sock
*sk
,
511 const struct sk_buff
*skb
)
513 struct tcp_sock
*tp
= tcp_sk(sk
);
514 if (tp
->rx_opt
.rcv_tsecr
&&
515 (TCP_SKB_CB(skb
)->end_seq
-
516 TCP_SKB_CB(skb
)->seq
>= inet_csk(sk
)->icsk_ack
.rcv_mss
))
517 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
, 0);
521 * This function should be called every time data is copied to user space.
522 * It calculates the appropriate TCP receive buffer space.
524 void tcp_rcv_space_adjust(struct sock
*sk
)
526 struct tcp_sock
*tp
= tcp_sk(sk
);
530 if (tp
->rcvq_space
.time
== 0)
533 time
= tcp_time_stamp
- tp
->rcvq_space
.time
;
534 if (time
< (tp
->rcv_rtt_est
.rtt
>> 3) || tp
->rcv_rtt_est
.rtt
== 0)
537 space
= 2 * (tp
->copied_seq
- tp
->rcvq_space
.seq
);
539 space
= max(tp
->rcvq_space
.space
, space
);
541 if (tp
->rcvq_space
.space
!= space
) {
544 tp
->rcvq_space
.space
= space
;
546 if (sysctl_tcp_moderate_rcvbuf
&&
547 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
)) {
548 int new_clamp
= space
;
550 /* Receive space grows, normalize in order to
551 * take into account packet headers and sk_buff
552 * structure overhead.
557 rcvmem
= SKB_TRUESIZE(tp
->advmss
+ MAX_TCP_HEADER
);
558 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
561 space
= min(space
, sysctl_tcp_rmem
[2]);
562 if (space
> sk
->sk_rcvbuf
) {
563 sk
->sk_rcvbuf
= space
;
565 /* Make the window clamp follow along. */
566 tp
->window_clamp
= new_clamp
;
572 tp
->rcvq_space
.seq
= tp
->copied_seq
;
573 tp
->rcvq_space
.time
= tcp_time_stamp
;
576 /* There is something which you must keep in mind when you analyze the
577 * behavior of the tp->ato delayed ack timeout interval. When a
578 * connection starts up, we want to ack as quickly as possible. The
579 * problem is that "good" TCP's do slow start at the beginning of data
580 * transmission. The means that until we send the first few ACK's the
581 * sender will sit on his end and only queue most of his data, because
582 * he can only send snd_cwnd unacked packets at any given time. For
583 * each ACK we send, he increments snd_cwnd and transmits more of his
586 static void tcp_event_data_recv(struct sock
*sk
, struct sk_buff
*skb
)
588 struct tcp_sock
*tp
= tcp_sk(sk
);
589 struct inet_connection_sock
*icsk
= inet_csk(sk
);
592 inet_csk_schedule_ack(sk
);
594 tcp_measure_rcv_mss(sk
, skb
);
596 tcp_rcv_rtt_measure(tp
);
598 now
= tcp_time_stamp
;
600 if (!icsk
->icsk_ack
.ato
) {
601 /* The _first_ data packet received, initialize
602 * delayed ACK engine.
604 tcp_incr_quickack(sk
);
605 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
607 int m
= now
- icsk
->icsk_ack
.lrcvtime
;
609 if (m
<= TCP_ATO_MIN
/ 2) {
610 /* The fastest case is the first. */
611 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + TCP_ATO_MIN
/ 2;
612 } else if (m
< icsk
->icsk_ack
.ato
) {
613 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + m
;
614 if (icsk
->icsk_ack
.ato
> icsk
->icsk_rto
)
615 icsk
->icsk_ack
.ato
= icsk
->icsk_rto
;
616 } else if (m
> icsk
->icsk_rto
) {
617 /* Too long gap. Apparently sender failed to
618 * restart window, so that we send ACKs quickly.
620 tcp_incr_quickack(sk
);
624 icsk
->icsk_ack
.lrcvtime
= now
;
626 TCP_ECN_check_ce(tp
, skb
);
629 tcp_grow_window(sk
, skb
);
632 /* Called to compute a smoothed rtt estimate. The data fed to this
633 * routine either comes from timestamps, or from segments that were
634 * known _not_ to have been retransmitted [see Karn/Partridge
635 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
636 * piece by Van Jacobson.
637 * NOTE: the next three routines used to be one big routine.
638 * To save cycles in the RFC 1323 implementation it was better to break
639 * it up into three procedures. -- erics
641 static void tcp_rtt_estimator(struct sock
*sk
, const __u32 mrtt
)
643 struct tcp_sock
*tp
= tcp_sk(sk
);
644 long m
= mrtt
; /* RTT */
646 /* The following amusing code comes from Jacobson's
647 * article in SIGCOMM '88. Note that rtt and mdev
648 * are scaled versions of rtt and mean deviation.
649 * This is designed to be as fast as possible
650 * m stands for "measurement".
652 * On a 1990 paper the rto value is changed to:
653 * RTO = rtt + 4 * mdev
655 * Funny. This algorithm seems to be very broken.
656 * These formulae increase RTO, when it should be decreased, increase
657 * too slowly, when it should be increased quickly, decrease too quickly
658 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
659 * does not matter how to _calculate_ it. Seems, it was trap
660 * that VJ failed to avoid. 8)
665 m
-= (tp
->srtt
>> 3); /* m is now error in rtt est */
666 tp
->srtt
+= m
; /* rtt = 7/8 rtt + 1/8 new */
668 m
= -m
; /* m is now abs(error) */
669 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
670 /* This is similar to one of Eifel findings.
671 * Eifel blocks mdev updates when rtt decreases.
672 * This solution is a bit different: we use finer gain
673 * for mdev in this case (alpha*beta).
674 * Like Eifel it also prevents growth of rto,
675 * but also it limits too fast rto decreases,
676 * happening in pure Eifel.
681 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
683 tp
->mdev
+= m
; /* mdev = 3/4 mdev + 1/4 new */
684 if (tp
->mdev
> tp
->mdev_max
) {
685 tp
->mdev_max
= tp
->mdev
;
686 if (tp
->mdev_max
> tp
->rttvar
)
687 tp
->rttvar
= tp
->mdev_max
;
689 if (after(tp
->snd_una
, tp
->rtt_seq
)) {
690 if (tp
->mdev_max
< tp
->rttvar
)
691 tp
->rttvar
-= (tp
->rttvar
- tp
->mdev_max
) >> 2;
692 tp
->rtt_seq
= tp
->snd_nxt
;
693 tp
->mdev_max
= tcp_rto_min(sk
);
696 /* no previous measure. */
697 tp
->srtt
= m
<< 3; /* take the measured time to be rtt */
698 tp
->mdev
= m
<< 1; /* make sure rto = 3*rtt */
699 tp
->mdev_max
= tp
->rttvar
= max(tp
->mdev
, tcp_rto_min(sk
));
700 tp
->rtt_seq
= tp
->snd_nxt
;
704 /* Set the sk_pacing_rate to allow proper sizing of TSO packets.
705 * Note: TCP stack does not yet implement pacing.
706 * FQ packet scheduler can be used to implement cheap but effective
707 * TCP pacing, to smooth the burst on large writes when packets
708 * in flight is significantly lower than cwnd (or rwin)
710 static void tcp_update_pacing_rate(struct sock
*sk
)
712 const struct tcp_sock
*tp
= tcp_sk(sk
);
715 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
716 rate
= (u64
)tp
->mss_cache
* 2 * (HZ
<< 3);
718 rate
*= max(tp
->snd_cwnd
, tp
->packets_out
);
720 /* Correction for small srtt : minimum srtt being 8 (1 jiffy << 3),
721 * be conservative and assume srtt = 1 (125 us instead of 1.25 ms)
722 * We probably need usec resolution in the future.
723 * Note: This also takes care of possible srtt=0 case,
724 * when tcp_rtt_estimator() was not yet called.
726 if (tp
->srtt
> 8 + 2)
727 do_div(rate
, tp
->srtt
);
729 sk
->sk_pacing_rate
= min_t(u64
, rate
, ~0U);
732 /* Calculate rto without backoff. This is the second half of Van Jacobson's
733 * routine referred to above.
735 void tcp_set_rto(struct sock
*sk
)
737 const struct tcp_sock
*tp
= tcp_sk(sk
);
738 /* Old crap is replaced with new one. 8)
741 * 1. If rtt variance happened to be less 50msec, it is hallucination.
742 * It cannot be less due to utterly erratic ACK generation made
743 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
744 * to do with delayed acks, because at cwnd>2 true delack timeout
745 * is invisible. Actually, Linux-2.4 also generates erratic
746 * ACKs in some circumstances.
748 inet_csk(sk
)->icsk_rto
= __tcp_set_rto(tp
);
750 /* 2. Fixups made earlier cannot be right.
751 * If we do not estimate RTO correctly without them,
752 * all the algo is pure shit and should be replaced
753 * with correct one. It is exactly, which we pretend to do.
756 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
757 * guarantees that rto is higher.
762 __u32
tcp_init_cwnd(const struct tcp_sock
*tp
, const struct dst_entry
*dst
)
764 __u32 cwnd
= (dst
? dst_metric(dst
, RTAX_INITCWND
) : 0);
767 cwnd
= TCP_INIT_CWND
;
768 return min_t(__u32
, cwnd
, tp
->snd_cwnd_clamp
);
772 * Packet counting of FACK is based on in-order assumptions, therefore TCP
773 * disables it when reordering is detected
775 void tcp_disable_fack(struct tcp_sock
*tp
)
777 /* RFC3517 uses different metric in lost marker => reset on change */
779 tp
->lost_skb_hint
= NULL
;
780 tp
->rx_opt
.sack_ok
&= ~TCP_FACK_ENABLED
;
783 /* Take a notice that peer is sending D-SACKs */
784 static void tcp_dsack_seen(struct tcp_sock
*tp
)
786 tp
->rx_opt
.sack_ok
|= TCP_DSACK_SEEN
;
789 static void tcp_update_reordering(struct sock
*sk
, const int metric
,
792 struct tcp_sock
*tp
= tcp_sk(sk
);
793 if (metric
> tp
->reordering
) {
796 tp
->reordering
= min(TCP_MAX_REORDERING
, metric
);
798 /* This exciting event is worth to be remembered. 8) */
800 mib_idx
= LINUX_MIB_TCPTSREORDER
;
801 else if (tcp_is_reno(tp
))
802 mib_idx
= LINUX_MIB_TCPRENOREORDER
;
803 else if (tcp_is_fack(tp
))
804 mib_idx
= LINUX_MIB_TCPFACKREORDER
;
806 mib_idx
= LINUX_MIB_TCPSACKREORDER
;
808 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
809 #if FASTRETRANS_DEBUG > 1
810 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
811 tp
->rx_opt
.sack_ok
, inet_csk(sk
)->icsk_ca_state
,
815 tp
->undo_marker
? tp
->undo_retrans
: 0);
817 tcp_disable_fack(tp
);
821 tcp_disable_early_retrans(tp
);
824 /* This must be called before lost_out is incremented */
825 static void tcp_verify_retransmit_hint(struct tcp_sock
*tp
, struct sk_buff
*skb
)
827 if ((tp
->retransmit_skb_hint
== NULL
) ||
828 before(TCP_SKB_CB(skb
)->seq
,
829 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
830 tp
->retransmit_skb_hint
= skb
;
833 after(TCP_SKB_CB(skb
)->end_seq
, tp
->retransmit_high
))
834 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
837 static void tcp_skb_mark_lost(struct tcp_sock
*tp
, struct sk_buff
*skb
)
839 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
840 tcp_verify_retransmit_hint(tp
, skb
);
842 tp
->lost_out
+= tcp_skb_pcount(skb
);
843 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
847 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock
*tp
,
850 tcp_verify_retransmit_hint(tp
, skb
);
852 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
853 tp
->lost_out
+= tcp_skb_pcount(skb
);
854 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
858 /* This procedure tags the retransmission queue when SACKs arrive.
860 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
861 * Packets in queue with these bits set are counted in variables
862 * sacked_out, retrans_out and lost_out, correspondingly.
864 * Valid combinations are:
865 * Tag InFlight Description
866 * 0 1 - orig segment is in flight.
867 * S 0 - nothing flies, orig reached receiver.
868 * L 0 - nothing flies, orig lost by net.
869 * R 2 - both orig and retransmit are in flight.
870 * L|R 1 - orig is lost, retransmit is in flight.
871 * S|R 1 - orig reached receiver, retrans is still in flight.
872 * (L|S|R is logically valid, it could occur when L|R is sacked,
873 * but it is equivalent to plain S and code short-curcuits it to S.
874 * L|S is logically invalid, it would mean -1 packet in flight 8))
876 * These 6 states form finite state machine, controlled by the following events:
877 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
878 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
879 * 3. Loss detection event of two flavors:
880 * A. Scoreboard estimator decided the packet is lost.
881 * A'. Reno "three dupacks" marks head of queue lost.
882 * A''. Its FACK modification, head until snd.fack is lost.
883 * B. SACK arrives sacking SND.NXT at the moment, when the
884 * segment was retransmitted.
885 * 4. D-SACK added new rule: D-SACK changes any tag to S.
887 * It is pleasant to note, that state diagram turns out to be commutative,
888 * so that we are allowed not to be bothered by order of our actions,
889 * when multiple events arrive simultaneously. (see the function below).
891 * Reordering detection.
892 * --------------------
893 * Reordering metric is maximal distance, which a packet can be displaced
894 * in packet stream. With SACKs we can estimate it:
896 * 1. SACK fills old hole and the corresponding segment was not
897 * ever retransmitted -> reordering. Alas, we cannot use it
898 * when segment was retransmitted.
899 * 2. The last flaw is solved with D-SACK. D-SACK arrives
900 * for retransmitted and already SACKed segment -> reordering..
901 * Both of these heuristics are not used in Loss state, when we cannot
902 * account for retransmits accurately.
904 * SACK block validation.
905 * ----------------------
907 * SACK block range validation checks that the received SACK block fits to
908 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
909 * Note that SND.UNA is not included to the range though being valid because
910 * it means that the receiver is rather inconsistent with itself reporting
911 * SACK reneging when it should advance SND.UNA. Such SACK block this is
912 * perfectly valid, however, in light of RFC2018 which explicitly states
913 * that "SACK block MUST reflect the newest segment. Even if the newest
914 * segment is going to be discarded ...", not that it looks very clever
915 * in case of head skb. Due to potentional receiver driven attacks, we
916 * choose to avoid immediate execution of a walk in write queue due to
917 * reneging and defer head skb's loss recovery to standard loss recovery
918 * procedure that will eventually trigger (nothing forbids us doing this).
920 * Implements also blockage to start_seq wrap-around. Problem lies in the
921 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
922 * there's no guarantee that it will be before snd_nxt (n). The problem
923 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
926 * <- outs wnd -> <- wrapzone ->
927 * u e n u_w e_w s n_w
929 * |<------------+------+----- TCP seqno space --------------+---------->|
930 * ...-- <2^31 ->| |<--------...
931 * ...---- >2^31 ------>| |<--------...
933 * Current code wouldn't be vulnerable but it's better still to discard such
934 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
935 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
936 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
937 * equal to the ideal case (infinite seqno space without wrap caused issues).
939 * With D-SACK the lower bound is extended to cover sequence space below
940 * SND.UNA down to undo_marker, which is the last point of interest. Yet
941 * again, D-SACK block must not to go across snd_una (for the same reason as
942 * for the normal SACK blocks, explained above). But there all simplicity
943 * ends, TCP might receive valid D-SACKs below that. As long as they reside
944 * fully below undo_marker they do not affect behavior in anyway and can
945 * therefore be safely ignored. In rare cases (which are more or less
946 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
947 * fragmentation and packet reordering past skb's retransmission. To consider
948 * them correctly, the acceptable range must be extended even more though
949 * the exact amount is rather hard to quantify. However, tp->max_window can
950 * be used as an exaggerated estimate.
952 static bool tcp_is_sackblock_valid(struct tcp_sock
*tp
, bool is_dsack
,
953 u32 start_seq
, u32 end_seq
)
955 /* Too far in future, or reversed (interpretation is ambiguous) */
956 if (after(end_seq
, tp
->snd_nxt
) || !before(start_seq
, end_seq
))
959 /* Nasty start_seq wrap-around check (see comments above) */
960 if (!before(start_seq
, tp
->snd_nxt
))
963 /* In outstanding window? ...This is valid exit for D-SACKs too.
964 * start_seq == snd_una is non-sensical (see comments above)
966 if (after(start_seq
, tp
->snd_una
))
969 if (!is_dsack
|| !tp
->undo_marker
)
972 /* ...Then it's D-SACK, and must reside below snd_una completely */
973 if (after(end_seq
, tp
->snd_una
))
976 if (!before(start_seq
, tp
->undo_marker
))
980 if (!after(end_seq
, tp
->undo_marker
))
983 /* Undo_marker boundary crossing (overestimates a lot). Known already:
984 * start_seq < undo_marker and end_seq >= undo_marker.
986 return !before(start_seq
, end_seq
- tp
->max_window
);
989 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
990 * Event "B". Later note: FACK people cheated me again 8), we have to account
991 * for reordering! Ugly, but should help.
993 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
994 * less than what is now known to be received by the other end (derived from
995 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
996 * retransmitted skbs to avoid some costly processing per ACKs.
998 static void tcp_mark_lost_retrans(struct sock
*sk
)
1000 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1001 struct tcp_sock
*tp
= tcp_sk(sk
);
1002 struct sk_buff
*skb
;
1004 u32 new_low_seq
= tp
->snd_nxt
;
1005 u32 received_upto
= tcp_highest_sack_seq(tp
);
1007 if (!tcp_is_fack(tp
) || !tp
->retrans_out
||
1008 !after(received_upto
, tp
->lost_retrans_low
) ||
1009 icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
1012 tcp_for_write_queue(skb
, sk
) {
1013 u32 ack_seq
= TCP_SKB_CB(skb
)->ack_seq
;
1015 if (skb
== tcp_send_head(sk
))
1017 if (cnt
== tp
->retrans_out
)
1019 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1022 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
))
1025 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1026 * constraint here (see above) but figuring out that at
1027 * least tp->reordering SACK blocks reside between ack_seq
1028 * and received_upto is not easy task to do cheaply with
1029 * the available datastructures.
1031 * Whether FACK should check here for tp->reordering segs
1032 * in-between one could argue for either way (it would be
1033 * rather simple to implement as we could count fack_count
1034 * during the walk and do tp->fackets_out - fack_count).
1036 if (after(received_upto
, ack_seq
)) {
1037 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1038 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1040 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
1041 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSTRETRANSMIT
);
1043 if (before(ack_seq
, new_low_seq
))
1044 new_low_seq
= ack_seq
;
1045 cnt
+= tcp_skb_pcount(skb
);
1049 if (tp
->retrans_out
)
1050 tp
->lost_retrans_low
= new_low_seq
;
1053 static bool tcp_check_dsack(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1054 struct tcp_sack_block_wire
*sp
, int num_sacks
,
1057 struct tcp_sock
*tp
= tcp_sk(sk
);
1058 u32 start_seq_0
= get_unaligned_be32(&sp
[0].start_seq
);
1059 u32 end_seq_0
= get_unaligned_be32(&sp
[0].end_seq
);
1060 bool dup_sack
= false;
1062 if (before(start_seq_0
, TCP_SKB_CB(ack_skb
)->ack_seq
)) {
1065 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPDSACKRECV
);
1066 } else if (num_sacks
> 1) {
1067 u32 end_seq_1
= get_unaligned_be32(&sp
[1].end_seq
);
1068 u32 start_seq_1
= get_unaligned_be32(&sp
[1].start_seq
);
1070 if (!after(end_seq_0
, end_seq_1
) &&
1071 !before(start_seq_0
, start_seq_1
)) {
1074 NET_INC_STATS_BH(sock_net(sk
),
1075 LINUX_MIB_TCPDSACKOFORECV
);
1079 /* D-SACK for already forgotten data... Do dumb counting. */
1080 if (dup_sack
&& tp
->undo_marker
&& tp
->undo_retrans
> 0 &&
1081 !after(end_seq_0
, prior_snd_una
) &&
1082 after(end_seq_0
, tp
->undo_marker
))
1088 struct tcp_sacktag_state
{
1094 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1095 * the incoming SACK may not exactly match but we can find smaller MSS
1096 * aligned portion of it that matches. Therefore we might need to fragment
1097 * which may fail and creates some hassle (caller must handle error case
1100 * FIXME: this could be merged to shift decision code
1102 static int tcp_match_skb_to_sack(struct sock
*sk
, struct sk_buff
*skb
,
1103 u32 start_seq
, u32 end_seq
)
1107 unsigned int pkt_len
;
1110 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1111 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1113 if (tcp_skb_pcount(skb
) > 1 && !in_sack
&&
1114 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
1115 mss
= tcp_skb_mss(skb
);
1116 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1119 pkt_len
= start_seq
- TCP_SKB_CB(skb
)->seq
;
1123 pkt_len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1128 /* Round if necessary so that SACKs cover only full MSSes
1129 * and/or the remaining small portion (if present)
1131 if (pkt_len
> mss
) {
1132 unsigned int new_len
= (pkt_len
/ mss
) * mss
;
1133 if (!in_sack
&& new_len
< pkt_len
) {
1135 if (new_len
>= skb
->len
)
1140 err
= tcp_fragment(sk
, skb
, pkt_len
, mss
);
1148 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1149 static u8
tcp_sacktag_one(struct sock
*sk
,
1150 struct tcp_sacktag_state
*state
, u8 sacked
,
1151 u32 start_seq
, u32 end_seq
,
1152 bool dup_sack
, int pcount
)
1154 struct tcp_sock
*tp
= tcp_sk(sk
);
1155 int fack_count
= state
->fack_count
;
1157 /* Account D-SACK for retransmitted packet. */
1158 if (dup_sack
&& (sacked
& TCPCB_RETRANS
)) {
1159 if (tp
->undo_marker
&& tp
->undo_retrans
> 0 &&
1160 after(end_seq
, tp
->undo_marker
))
1162 if (sacked
& TCPCB_SACKED_ACKED
)
1163 state
->reord
= min(fack_count
, state
->reord
);
1166 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1167 if (!after(end_seq
, tp
->snd_una
))
1170 if (!(sacked
& TCPCB_SACKED_ACKED
)) {
1171 if (sacked
& TCPCB_SACKED_RETRANS
) {
1172 /* If the segment is not tagged as lost,
1173 * we do not clear RETRANS, believing
1174 * that retransmission is still in flight.
1176 if (sacked
& TCPCB_LOST
) {
1177 sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1178 tp
->lost_out
-= pcount
;
1179 tp
->retrans_out
-= pcount
;
1182 if (!(sacked
& TCPCB_RETRANS
)) {
1183 /* New sack for not retransmitted frame,
1184 * which was in hole. It is reordering.
1186 if (before(start_seq
,
1187 tcp_highest_sack_seq(tp
)))
1188 state
->reord
= min(fack_count
,
1190 if (!after(end_seq
, tp
->high_seq
))
1191 state
->flag
|= FLAG_ORIG_SACK_ACKED
;
1194 if (sacked
& TCPCB_LOST
) {
1195 sacked
&= ~TCPCB_LOST
;
1196 tp
->lost_out
-= pcount
;
1200 sacked
|= TCPCB_SACKED_ACKED
;
1201 state
->flag
|= FLAG_DATA_SACKED
;
1202 tp
->sacked_out
+= pcount
;
1204 fack_count
+= pcount
;
1206 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1207 if (!tcp_is_fack(tp
) && (tp
->lost_skb_hint
!= NULL
) &&
1208 before(start_seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
))
1209 tp
->lost_cnt_hint
+= pcount
;
1211 if (fack_count
> tp
->fackets_out
)
1212 tp
->fackets_out
= fack_count
;
1215 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1216 * frames and clear it. undo_retrans is decreased above, L|R frames
1217 * are accounted above as well.
1219 if (dup_sack
&& (sacked
& TCPCB_SACKED_RETRANS
)) {
1220 sacked
&= ~TCPCB_SACKED_RETRANS
;
1221 tp
->retrans_out
-= pcount
;
1227 /* Shift newly-SACKed bytes from this skb to the immediately previous
1228 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1230 static bool tcp_shifted_skb(struct sock
*sk
, struct sk_buff
*skb
,
1231 struct tcp_sacktag_state
*state
,
1232 unsigned int pcount
, int shifted
, int mss
,
1235 struct tcp_sock
*tp
= tcp_sk(sk
);
1236 struct sk_buff
*prev
= tcp_write_queue_prev(sk
, skb
);
1237 u32 start_seq
= TCP_SKB_CB(skb
)->seq
; /* start of newly-SACKed */
1238 u32 end_seq
= start_seq
+ shifted
; /* end of newly-SACKed */
1242 /* Adjust counters and hints for the newly sacked sequence
1243 * range but discard the return value since prev is already
1244 * marked. We must tag the range first because the seq
1245 * advancement below implicitly advances
1246 * tcp_highest_sack_seq() when skb is highest_sack.
1248 tcp_sacktag_one(sk
, state
, TCP_SKB_CB(skb
)->sacked
,
1249 start_seq
, end_seq
, dup_sack
, pcount
);
1251 if (skb
== tp
->lost_skb_hint
)
1252 tp
->lost_cnt_hint
+= pcount
;
1254 TCP_SKB_CB(prev
)->end_seq
+= shifted
;
1255 TCP_SKB_CB(skb
)->seq
+= shifted
;
1257 skb_shinfo(prev
)->gso_segs
+= pcount
;
1258 BUG_ON(skb_shinfo(skb
)->gso_segs
< pcount
);
1259 skb_shinfo(skb
)->gso_segs
-= pcount
;
1261 /* When we're adding to gso_segs == 1, gso_size will be zero,
1262 * in theory this shouldn't be necessary but as long as DSACK
1263 * code can come after this skb later on it's better to keep
1264 * setting gso_size to something.
1266 if (!skb_shinfo(prev
)->gso_size
) {
1267 skb_shinfo(prev
)->gso_size
= mss
;
1268 skb_shinfo(prev
)->gso_type
= sk
->sk_gso_type
;
1271 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1272 if (skb_shinfo(skb
)->gso_segs
<= 1) {
1273 skb_shinfo(skb
)->gso_size
= 0;
1274 skb_shinfo(skb
)->gso_type
= 0;
1277 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1278 TCP_SKB_CB(prev
)->sacked
|= (TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
);
1281 BUG_ON(!tcp_skb_pcount(skb
));
1282 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKSHIFTED
);
1286 /* Whole SKB was eaten :-) */
1288 if (skb
== tp
->retransmit_skb_hint
)
1289 tp
->retransmit_skb_hint
= prev
;
1290 if (skb
== tp
->scoreboard_skb_hint
)
1291 tp
->scoreboard_skb_hint
= prev
;
1292 if (skb
== tp
->lost_skb_hint
) {
1293 tp
->lost_skb_hint
= prev
;
1294 tp
->lost_cnt_hint
-= tcp_skb_pcount(prev
);
1297 TCP_SKB_CB(prev
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
;
1298 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
1299 TCP_SKB_CB(prev
)->end_seq
++;
1301 if (skb
== tcp_highest_sack(sk
))
1302 tcp_advance_highest_sack(sk
, skb
);
1304 tcp_unlink_write_queue(skb
, sk
);
1305 sk_wmem_free_skb(sk
, skb
);
1307 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKMERGED
);
1312 /* I wish gso_size would have a bit more sane initialization than
1313 * something-or-zero which complicates things
1315 static int tcp_skb_seglen(const struct sk_buff
*skb
)
1317 return tcp_skb_pcount(skb
) == 1 ? skb
->len
: tcp_skb_mss(skb
);
1320 /* Shifting pages past head area doesn't work */
1321 static int skb_can_shift(const struct sk_buff
*skb
)
1323 return !skb_headlen(skb
) && skb_is_nonlinear(skb
);
1326 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1329 static struct sk_buff
*tcp_shift_skb_data(struct sock
*sk
, struct sk_buff
*skb
,
1330 struct tcp_sacktag_state
*state
,
1331 u32 start_seq
, u32 end_seq
,
1334 struct tcp_sock
*tp
= tcp_sk(sk
);
1335 struct sk_buff
*prev
;
1341 if (!sk_can_gso(sk
))
1344 /* Normally R but no L won't result in plain S */
1346 (TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_RETRANS
)) == TCPCB_SACKED_RETRANS
)
1348 if (!skb_can_shift(skb
))
1350 /* This frame is about to be dropped (was ACKed). */
1351 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1354 /* Can only happen with delayed DSACK + discard craziness */
1355 if (unlikely(skb
== tcp_write_queue_head(sk
)))
1357 prev
= tcp_write_queue_prev(sk
, skb
);
1359 if ((TCP_SKB_CB(prev
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
)
1362 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1363 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1367 pcount
= tcp_skb_pcount(skb
);
1368 mss
= tcp_skb_seglen(skb
);
1370 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1371 * drop this restriction as unnecessary
1373 if (mss
!= tcp_skb_seglen(prev
))
1376 if (!after(TCP_SKB_CB(skb
)->end_seq
, start_seq
))
1378 /* CHECKME: This is non-MSS split case only?, this will
1379 * cause skipped skbs due to advancing loop btw, original
1380 * has that feature too
1382 if (tcp_skb_pcount(skb
) <= 1)
1385 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1387 /* TODO: head merge to next could be attempted here
1388 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1389 * though it might not be worth of the additional hassle
1391 * ...we can probably just fallback to what was done
1392 * previously. We could try merging non-SACKed ones
1393 * as well but it probably isn't going to buy off
1394 * because later SACKs might again split them, and
1395 * it would make skb timestamp tracking considerably
1401 len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1403 BUG_ON(len
> skb
->len
);
1405 /* MSS boundaries should be honoured or else pcount will
1406 * severely break even though it makes things bit trickier.
1407 * Optimize common case to avoid most of the divides
1409 mss
= tcp_skb_mss(skb
);
1411 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1412 * drop this restriction as unnecessary
1414 if (mss
!= tcp_skb_seglen(prev
))
1419 } else if (len
< mss
) {
1427 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1428 if (!after(TCP_SKB_CB(skb
)->seq
+ len
, tp
->snd_una
))
1431 if (!skb_shift(prev
, skb
, len
))
1433 if (!tcp_shifted_skb(sk
, skb
, state
, pcount
, len
, mss
, dup_sack
))
1436 /* Hole filled allows collapsing with the next as well, this is very
1437 * useful when hole on every nth skb pattern happens
1439 if (prev
== tcp_write_queue_tail(sk
))
1441 skb
= tcp_write_queue_next(sk
, prev
);
1443 if (!skb_can_shift(skb
) ||
1444 (skb
== tcp_send_head(sk
)) ||
1445 ((TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
) ||
1446 (mss
!= tcp_skb_seglen(skb
)))
1450 if (skb_shift(prev
, skb
, len
)) {
1451 pcount
+= tcp_skb_pcount(skb
);
1452 tcp_shifted_skb(sk
, skb
, state
, tcp_skb_pcount(skb
), len
, mss
, 0);
1456 state
->fack_count
+= pcount
;
1463 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKSHIFTFALLBACK
);
1467 static struct sk_buff
*tcp_sacktag_walk(struct sk_buff
*skb
, struct sock
*sk
,
1468 struct tcp_sack_block
*next_dup
,
1469 struct tcp_sacktag_state
*state
,
1470 u32 start_seq
, u32 end_seq
,
1473 struct tcp_sock
*tp
= tcp_sk(sk
);
1474 struct sk_buff
*tmp
;
1476 tcp_for_write_queue_from(skb
, sk
) {
1478 bool dup_sack
= dup_sack_in
;
1480 if (skb
== tcp_send_head(sk
))
1483 /* queue is in-order => we can short-circuit the walk early */
1484 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1487 if ((next_dup
!= NULL
) &&
1488 before(TCP_SKB_CB(skb
)->seq
, next_dup
->end_seq
)) {
1489 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1490 next_dup
->start_seq
,
1496 /* skb reference here is a bit tricky to get right, since
1497 * shifting can eat and free both this skb and the next,
1498 * so not even _safe variant of the loop is enough.
1501 tmp
= tcp_shift_skb_data(sk
, skb
, state
,
1502 start_seq
, end_seq
, dup_sack
);
1511 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1517 if (unlikely(in_sack
< 0))
1521 TCP_SKB_CB(skb
)->sacked
=
1524 TCP_SKB_CB(skb
)->sacked
,
1525 TCP_SKB_CB(skb
)->seq
,
1526 TCP_SKB_CB(skb
)->end_seq
,
1528 tcp_skb_pcount(skb
));
1530 if (!before(TCP_SKB_CB(skb
)->seq
,
1531 tcp_highest_sack_seq(tp
)))
1532 tcp_advance_highest_sack(sk
, skb
);
1535 state
->fack_count
+= tcp_skb_pcount(skb
);
1540 /* Avoid all extra work that is being done by sacktag while walking in
1543 static struct sk_buff
*tcp_sacktag_skip(struct sk_buff
*skb
, struct sock
*sk
,
1544 struct tcp_sacktag_state
*state
,
1547 tcp_for_write_queue_from(skb
, sk
) {
1548 if (skb
== tcp_send_head(sk
))
1551 if (after(TCP_SKB_CB(skb
)->end_seq
, skip_to_seq
))
1554 state
->fack_count
+= tcp_skb_pcount(skb
);
1559 static struct sk_buff
*tcp_maybe_skipping_dsack(struct sk_buff
*skb
,
1561 struct tcp_sack_block
*next_dup
,
1562 struct tcp_sacktag_state
*state
,
1565 if (next_dup
== NULL
)
1568 if (before(next_dup
->start_seq
, skip_to_seq
)) {
1569 skb
= tcp_sacktag_skip(skb
, sk
, state
, next_dup
->start_seq
);
1570 skb
= tcp_sacktag_walk(skb
, sk
, NULL
, state
,
1571 next_dup
->start_seq
, next_dup
->end_seq
,
1578 static int tcp_sack_cache_ok(const struct tcp_sock
*tp
, const struct tcp_sack_block
*cache
)
1580 return cache
< tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1584 tcp_sacktag_write_queue(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1587 struct tcp_sock
*tp
= tcp_sk(sk
);
1588 const unsigned char *ptr
= (skb_transport_header(ack_skb
) +
1589 TCP_SKB_CB(ack_skb
)->sacked
);
1590 struct tcp_sack_block_wire
*sp_wire
= (struct tcp_sack_block_wire
*)(ptr
+2);
1591 struct tcp_sack_block sp
[TCP_NUM_SACKS
];
1592 struct tcp_sack_block
*cache
;
1593 struct tcp_sacktag_state state
;
1594 struct sk_buff
*skb
;
1595 int num_sacks
= min(TCP_NUM_SACKS
, (ptr
[1] - TCPOLEN_SACK_BASE
) >> 3);
1597 bool found_dup_sack
= false;
1599 int first_sack_index
;
1602 state
.reord
= tp
->packets_out
;
1604 if (!tp
->sacked_out
) {
1605 if (WARN_ON(tp
->fackets_out
))
1606 tp
->fackets_out
= 0;
1607 tcp_highest_sack_reset(sk
);
1610 found_dup_sack
= tcp_check_dsack(sk
, ack_skb
, sp_wire
,
1611 num_sacks
, prior_snd_una
);
1613 state
.flag
|= FLAG_DSACKING_ACK
;
1615 /* Eliminate too old ACKs, but take into
1616 * account more or less fresh ones, they can
1617 * contain valid SACK info.
1619 if (before(TCP_SKB_CB(ack_skb
)->ack_seq
, prior_snd_una
- tp
->max_window
))
1622 if (!tp
->packets_out
)
1626 first_sack_index
= 0;
1627 for (i
= 0; i
< num_sacks
; i
++) {
1628 bool dup_sack
= !i
&& found_dup_sack
;
1630 sp
[used_sacks
].start_seq
= get_unaligned_be32(&sp_wire
[i
].start_seq
);
1631 sp
[used_sacks
].end_seq
= get_unaligned_be32(&sp_wire
[i
].end_seq
);
1633 if (!tcp_is_sackblock_valid(tp
, dup_sack
,
1634 sp
[used_sacks
].start_seq
,
1635 sp
[used_sacks
].end_seq
)) {
1639 if (!tp
->undo_marker
)
1640 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDNOUNDO
;
1642 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDOLD
;
1644 /* Don't count olds caused by ACK reordering */
1645 if ((TCP_SKB_CB(ack_skb
)->ack_seq
!= tp
->snd_una
) &&
1646 !after(sp
[used_sacks
].end_seq
, tp
->snd_una
))
1648 mib_idx
= LINUX_MIB_TCPSACKDISCARD
;
1651 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
1653 first_sack_index
= -1;
1657 /* Ignore very old stuff early */
1658 if (!after(sp
[used_sacks
].end_seq
, prior_snd_una
))
1664 /* order SACK blocks to allow in order walk of the retrans queue */
1665 for (i
= used_sacks
- 1; i
> 0; i
--) {
1666 for (j
= 0; j
< i
; j
++) {
1667 if (after(sp
[j
].start_seq
, sp
[j
+ 1].start_seq
)) {
1668 swap(sp
[j
], sp
[j
+ 1]);
1670 /* Track where the first SACK block goes to */
1671 if (j
== first_sack_index
)
1672 first_sack_index
= j
+ 1;
1677 skb
= tcp_write_queue_head(sk
);
1678 state
.fack_count
= 0;
1681 if (!tp
->sacked_out
) {
1682 /* It's already past, so skip checking against it */
1683 cache
= tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1685 cache
= tp
->recv_sack_cache
;
1686 /* Skip empty blocks in at head of the cache */
1687 while (tcp_sack_cache_ok(tp
, cache
) && !cache
->start_seq
&&
1692 while (i
< used_sacks
) {
1693 u32 start_seq
= sp
[i
].start_seq
;
1694 u32 end_seq
= sp
[i
].end_seq
;
1695 bool dup_sack
= (found_dup_sack
&& (i
== first_sack_index
));
1696 struct tcp_sack_block
*next_dup
= NULL
;
1698 if (found_dup_sack
&& ((i
+ 1) == first_sack_index
))
1699 next_dup
= &sp
[i
+ 1];
1701 /* Skip too early cached blocks */
1702 while (tcp_sack_cache_ok(tp
, cache
) &&
1703 !before(start_seq
, cache
->end_seq
))
1706 /* Can skip some work by looking recv_sack_cache? */
1707 if (tcp_sack_cache_ok(tp
, cache
) && !dup_sack
&&
1708 after(end_seq
, cache
->start_seq
)) {
1711 if (before(start_seq
, cache
->start_seq
)) {
1712 skb
= tcp_sacktag_skip(skb
, sk
, &state
,
1714 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
,
1721 /* Rest of the block already fully processed? */
1722 if (!after(end_seq
, cache
->end_seq
))
1725 skb
= tcp_maybe_skipping_dsack(skb
, sk
, next_dup
,
1729 /* ...tail remains todo... */
1730 if (tcp_highest_sack_seq(tp
) == cache
->end_seq
) {
1731 /* ...but better entrypoint exists! */
1732 skb
= tcp_highest_sack(sk
);
1735 state
.fack_count
= tp
->fackets_out
;
1740 skb
= tcp_sacktag_skip(skb
, sk
, &state
, cache
->end_seq
);
1741 /* Check overlap against next cached too (past this one already) */
1746 if (!before(start_seq
, tcp_highest_sack_seq(tp
))) {
1747 skb
= tcp_highest_sack(sk
);
1750 state
.fack_count
= tp
->fackets_out
;
1752 skb
= tcp_sacktag_skip(skb
, sk
, &state
, start_seq
);
1755 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
, &state
,
1756 start_seq
, end_seq
, dup_sack
);
1762 /* Clear the head of the cache sack blocks so we can skip it next time */
1763 for (i
= 0; i
< ARRAY_SIZE(tp
->recv_sack_cache
) - used_sacks
; i
++) {
1764 tp
->recv_sack_cache
[i
].start_seq
= 0;
1765 tp
->recv_sack_cache
[i
].end_seq
= 0;
1767 for (j
= 0; j
< used_sacks
; j
++)
1768 tp
->recv_sack_cache
[i
++] = sp
[j
];
1770 tcp_mark_lost_retrans(sk
);
1772 tcp_verify_left_out(tp
);
1774 if ((state
.reord
< tp
->fackets_out
) &&
1775 ((inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Loss
) || tp
->undo_marker
))
1776 tcp_update_reordering(sk
, tp
->fackets_out
- state
.reord
, 0);
1780 #if FASTRETRANS_DEBUG > 0
1781 WARN_ON((int)tp
->sacked_out
< 0);
1782 WARN_ON((int)tp
->lost_out
< 0);
1783 WARN_ON((int)tp
->retrans_out
< 0);
1784 WARN_ON((int)tcp_packets_in_flight(tp
) < 0);
1789 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1790 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1792 static bool tcp_limit_reno_sacked(struct tcp_sock
*tp
)
1796 holes
= max(tp
->lost_out
, 1U);
1797 holes
= min(holes
, tp
->packets_out
);
1799 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1800 tp
->sacked_out
= tp
->packets_out
- holes
;
1806 /* If we receive more dupacks than we expected counting segments
1807 * in assumption of absent reordering, interpret this as reordering.
1808 * The only another reason could be bug in receiver TCP.
1810 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1812 struct tcp_sock
*tp
= tcp_sk(sk
);
1813 if (tcp_limit_reno_sacked(tp
))
1814 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1817 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1819 static void tcp_add_reno_sack(struct sock
*sk
)
1821 struct tcp_sock
*tp
= tcp_sk(sk
);
1823 tcp_check_reno_reordering(sk
, 0);
1824 tcp_verify_left_out(tp
);
1827 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1829 static void tcp_remove_reno_sacks(struct sock
*sk
, int acked
)
1831 struct tcp_sock
*tp
= tcp_sk(sk
);
1834 /* One ACK acked hole. The rest eat duplicate ACKs. */
1835 if (acked
- 1 >= tp
->sacked_out
)
1838 tp
->sacked_out
-= acked
- 1;
1840 tcp_check_reno_reordering(sk
, acked
);
1841 tcp_verify_left_out(tp
);
1844 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1849 static void tcp_clear_retrans_partial(struct tcp_sock
*tp
)
1851 tp
->retrans_out
= 0;
1854 tp
->undo_marker
= 0;
1855 tp
->undo_retrans
= -1;
1858 void tcp_clear_retrans(struct tcp_sock
*tp
)
1860 tcp_clear_retrans_partial(tp
);
1862 tp
->fackets_out
= 0;
1866 /* Enter Loss state. If "how" is not zero, forget all SACK information
1867 * and reset tags completely, otherwise preserve SACKs. If receiver
1868 * dropped its ofo queue, we will know this due to reneging detection.
1870 void tcp_enter_loss(struct sock
*sk
, int how
)
1872 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1873 struct inet_connection_sock
*icsk1
= inet_csk(sk
);
1874 struct tcp_sock
*tp
= tcp_sk(sk
);
1875 struct sk_buff
*skb
;
1876 bool new_recovery
= false;
1878 /* Reduce ssthresh if it has not yet been made inside this window. */
1879 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
||
1880 !after(tp
->high_seq
, tp
->snd_una
) ||
1881 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
1882 new_recovery
= true;
1883 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1884 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1885 tcp_ca_event(sk
, CA_EVENT_LOSS
);
1887 if (icsk
->icsk_MMSRB
== 1)
1889 #ifdef CONFIG_MTK_NET_LOGGING
1890 printk("[mtk_net][mmspb] tcp_enter_loss snd_cwnd=%u, snd_cwnd_cnt=%u\n", tp
->snd_cwnd
, tp
->snd_cwnd_cnt
);
1892 if (tp
->mss_cache
!= 0)
1893 tp
->snd_cwnd
= (tp
->rcv_wnd
/ tp
->mss_cache
);
1896 tp
->snd_cwnd
= (tp
->rcv_wnd
/ tp
->advmss
);
1899 if (tp
->snd_ssthresh
> 16)
1901 tp
->snd_cwnd
= tp
->snd_ssthresh
/ 2;//set snd_cwnd is half of default snd_ssthresh
1905 tp
->snd_cwnd
= tp
->snd_ssthresh
/ 2 + 4;
1907 #ifdef CONFIG_MTK_NET_LOGGING
1908 printk("[mtk_net][mmspb] tcp_enter_loss update snd_cwnd=%u\n", tp
->snd_cwnd
);
1910 icsk1
->icsk_MMSRB
= 0;
1911 #ifdef CONFIG_MTK_NET_LOGGING
1912 printk("[mtk_net][mmspb] tcp_enter_loss set icsk_MMSRB=0\n");
1921 tp
->snd_cwnd_cnt
= 0;
1922 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1924 tcp_clear_retrans_partial(tp
);
1926 if (tcp_is_reno(tp
))
1927 tcp_reset_reno_sack(tp
);
1929 tp
->undo_marker
= tp
->snd_una
;
1932 tp
->fackets_out
= 0;
1934 tcp_clear_all_retrans_hints(tp
);
1936 tcp_for_write_queue(skb
, sk
) {
1937 if (skb
== tcp_send_head(sk
))
1940 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
1941 tp
->undo_marker
= 0;
1942 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
1943 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) || how
) {
1944 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
1945 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1946 tp
->lost_out
+= tcp_skb_pcount(skb
);
1947 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
1950 tcp_verify_left_out(tp
);
1952 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1953 sysctl_tcp_reordering
);
1954 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1955 tp
->high_seq
= tp
->snd_nxt
;
1956 TCP_ECN_queue_cwr(tp
);
1958 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1959 * loss recovery is underway except recurring timeout(s) on
1960 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1962 tp
->frto
= sysctl_tcp_frto
&&
1963 (new_recovery
|| icsk
->icsk_retransmits
) &&
1964 !inet_csk(sk
)->icsk_mtup
.probe_size
;
1967 /* If ACK arrived pointing to a remembered SACK, it means that our
1968 * remembered SACKs do not reflect real state of receiver i.e.
1969 * receiver _host_ is heavily congested (or buggy).
1971 * Do processing similar to RTO timeout.
1973 static bool tcp_check_sack_reneging(struct sock
*sk
, int flag
)
1975 if (flag
& FLAG_SACK_RENEGING
) {
1976 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1977 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSACKRENEGING
);
1979 tcp_enter_loss(sk
, 1);
1980 icsk
->icsk_retransmits
++;
1981 tcp_retransmit_skb(sk
, tcp_write_queue_head(sk
));
1982 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1983 icsk
->icsk_rto
, sysctl_tcp_rto_max
);
1989 static inline int tcp_fackets_out(const struct tcp_sock
*tp
)
1991 return tcp_is_reno(tp
) ? tp
->sacked_out
+ 1 : tp
->fackets_out
;
1994 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
1995 * counter when SACK is enabled (without SACK, sacked_out is used for
1998 * Instead, with FACK TCP uses fackets_out that includes both SACKed
1999 * segments up to the highest received SACK block so far and holes in
2002 * With reordering, holes may still be in flight, so RFC3517 recovery
2003 * uses pure sacked_out (total number of SACKed segments) even though
2004 * it violates the RFC that uses duplicate ACKs, often these are equal
2005 * but when e.g. out-of-window ACKs or packet duplication occurs,
2006 * they differ. Since neither occurs due to loss, TCP should really
2009 static inline int tcp_dupack_heuristics(const struct tcp_sock
*tp
)
2011 return tcp_is_fack(tp
) ? tp
->fackets_out
: tp
->sacked_out
+ 1;
2014 static bool tcp_pause_early_retransmit(struct sock
*sk
, int flag
)
2016 struct tcp_sock
*tp
= tcp_sk(sk
);
2017 unsigned long delay
;
2019 /* Delay early retransmit and entering fast recovery for
2020 * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
2021 * available, or RTO is scheduled to fire first.
2023 if (sysctl_tcp_early_retrans
< 2 || sysctl_tcp_early_retrans
> 3 ||
2024 (flag
& FLAG_ECE
) || !tp
->srtt
)
2027 delay
= max_t(unsigned long, (tp
->srtt
>> 5), msecs_to_jiffies(2));
2028 if (!time_after(inet_csk(sk
)->icsk_timeout
, (jiffies
+ delay
)))
2031 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_EARLY_RETRANS
, delay
,
2032 sysctl_tcp_rto_max
);
2036 static inline int tcp_skb_timedout(const struct sock
*sk
,
2037 const struct sk_buff
*skb
)
2039 return tcp_time_stamp
- TCP_SKB_CB(skb
)->when
> inet_csk(sk
)->icsk_rto
;
2042 static inline int tcp_head_timedout(const struct sock
*sk
)
2044 const struct tcp_sock
*tp
= tcp_sk(sk
);
2046 return tp
->packets_out
&&
2047 tcp_skb_timedout(sk
, tcp_write_queue_head(sk
));
2050 /* Linux NewReno/SACK/FACK/ECN state machine.
2051 * --------------------------------------
2053 * "Open" Normal state, no dubious events, fast path.
2054 * "Disorder" In all the respects it is "Open",
2055 * but requires a bit more attention. It is entered when
2056 * we see some SACKs or dupacks. It is split of "Open"
2057 * mainly to move some processing from fast path to slow one.
2058 * "CWR" CWND was reduced due to some Congestion Notification event.
2059 * It can be ECN, ICMP source quench, local device congestion.
2060 * "Recovery" CWND was reduced, we are fast-retransmitting.
2061 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2063 * tcp_fastretrans_alert() is entered:
2064 * - each incoming ACK, if state is not "Open"
2065 * - when arrived ACK is unusual, namely:
2070 * Counting packets in flight is pretty simple.
2072 * in_flight = packets_out - left_out + retrans_out
2074 * packets_out is SND.NXT-SND.UNA counted in packets.
2076 * retrans_out is number of retransmitted segments.
2078 * left_out is number of segments left network, but not ACKed yet.
2080 * left_out = sacked_out + lost_out
2082 * sacked_out: Packets, which arrived to receiver out of order
2083 * and hence not ACKed. With SACKs this number is simply
2084 * amount of SACKed data. Even without SACKs
2085 * it is easy to give pretty reliable estimate of this number,
2086 * counting duplicate ACKs.
2088 * lost_out: Packets lost by network. TCP has no explicit
2089 * "loss notification" feedback from network (for now).
2090 * It means that this number can be only _guessed_.
2091 * Actually, it is the heuristics to predict lossage that
2092 * distinguishes different algorithms.
2094 * F.e. after RTO, when all the queue is considered as lost,
2095 * lost_out = packets_out and in_flight = retrans_out.
2097 * Essentially, we have now two algorithms counting
2100 * FACK: It is the simplest heuristics. As soon as we decided
2101 * that something is lost, we decide that _all_ not SACKed
2102 * packets until the most forward SACK are lost. I.e.
2103 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2104 * It is absolutely correct estimate, if network does not reorder
2105 * packets. And it loses any connection to reality when reordering
2106 * takes place. We use FACK by default until reordering
2107 * is suspected on the path to this destination.
2109 * NewReno: when Recovery is entered, we assume that one segment
2110 * is lost (classic Reno). While we are in Recovery and
2111 * a partial ACK arrives, we assume that one more packet
2112 * is lost (NewReno). This heuristics are the same in NewReno
2115 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2116 * deflation etc. CWND is real congestion window, never inflated, changes
2117 * only according to classic VJ rules.
2119 * Really tricky (and requiring careful tuning) part of algorithm
2120 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2121 * The first determines the moment _when_ we should reduce CWND and,
2122 * hence, slow down forward transmission. In fact, it determines the moment
2123 * when we decide that hole is caused by loss, rather than by a reorder.
2125 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2126 * holes, caused by lost packets.
2128 * And the most logically complicated part of algorithm is undo
2129 * heuristics. We detect false retransmits due to both too early
2130 * fast retransmit (reordering) and underestimated RTO, analyzing
2131 * timestamps and D-SACKs. When we detect that some segments were
2132 * retransmitted by mistake and CWND reduction was wrong, we undo
2133 * window reduction and abort recovery phase. This logic is hidden
2134 * inside several functions named tcp_try_undo_<something>.
2137 /* This function decides, when we should leave Disordered state
2138 * and enter Recovery phase, reducing congestion window.
2140 * Main question: may we further continue forward transmission
2141 * with the same cwnd?
2143 static bool tcp_time_to_recover(struct sock
*sk
, int flag
)
2145 struct tcp_sock
*tp
= tcp_sk(sk
);
2148 /* Trick#1: The loss is proven. */
2152 /* Not-A-Trick#2 : Classic rule... */
2153 if (tcp_dupack_heuristics(tp
) > tp
->reordering
)
2156 /* Trick#3 : when we use RFC2988 timer restart, fast
2157 * retransmit can be triggered by timeout of queue head.
2159 if (tcp_is_fack(tp
) && tcp_head_timedout(sk
))
2162 /* Trick#4: It is still not OK... But will it be useful to delay
2165 packets_out
= tp
->packets_out
;
2166 if (packets_out
<= tp
->reordering
&&
2167 tp
->sacked_out
>= max_t(__u32
, packets_out
/2, sysctl_tcp_reordering
) &&
2168 !tcp_may_send_now(sk
)) {
2169 /* We have nothing to send. This connection is limited
2170 * either by receiver window or by application.
2175 /* If a thin stream is detected, retransmit after first
2176 * received dupack. Employ only if SACK is supported in order
2177 * to avoid possible corner-case series of spurious retransmissions
2178 * Use only if there are no unsent data.
2180 if ((tp
->thin_dupack
|| sysctl_tcp_thin_dupack
) &&
2181 tcp_stream_is_thin(tp
) && tcp_dupack_heuristics(tp
) > 1 &&
2182 tcp_is_sack(tp
) && !tcp_send_head(sk
))
2185 /* Trick#6: TCP early retransmit, per RFC5827. To avoid spurious
2186 * retransmissions due to small network reorderings, we implement
2187 * Mitigation A.3 in the RFC and delay the retransmission for a short
2188 * interval if appropriate.
2190 if (tp
->do_early_retrans
&& !tp
->retrans_out
&& tp
->sacked_out
&&
2191 (tp
->packets_out
>= (tp
->sacked_out
+ 1) && tp
->packets_out
< 4) &&
2192 !tcp_may_send_now(sk
))
2193 return !tcp_pause_early_retransmit(sk
, flag
);
2198 /* New heuristics: it is possible only after we switched to restart timer
2199 * each time when something is ACKed. Hence, we can detect timed out packets
2200 * during fast retransmit without falling to slow start.
2202 * Usefulness of this as is very questionable, since we should know which of
2203 * the segments is the next to timeout which is relatively expensive to find
2204 * in general case unless we add some data structure just for that. The
2205 * current approach certainly won't find the right one too often and when it
2206 * finally does find _something_ it usually marks large part of the window
2207 * right away (because a retransmission with a larger timestamp blocks the
2208 * loop from advancing). -ij
2210 static void tcp_timeout_skbs(struct sock
*sk
)
2212 struct tcp_sock
*tp
= tcp_sk(sk
);
2213 struct sk_buff
*skb
;
2215 if (!tcp_is_fack(tp
) || !tcp_head_timedout(sk
))
2218 skb
= tp
->scoreboard_skb_hint
;
2219 if (tp
->scoreboard_skb_hint
== NULL
)
2220 skb
= tcp_write_queue_head(sk
);
2222 tcp_for_write_queue_from(skb
, sk
) {
2223 if (skb
== tcp_send_head(sk
))
2225 if (!tcp_skb_timedout(sk
, skb
))
2228 tcp_skb_mark_lost(tp
, skb
);
2231 tp
->scoreboard_skb_hint
= skb
;
2233 tcp_verify_left_out(tp
);
2236 /* Detect loss in event "A" above by marking head of queue up as lost.
2237 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2238 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2239 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2240 * the maximum SACKed segments to pass before reaching this limit.
2242 static void tcp_mark_head_lost(struct sock
*sk
, int packets
, int mark_head
)
2244 struct tcp_sock
*tp
= tcp_sk(sk
);
2245 struct sk_buff
*skb
;
2249 /* Use SACK to deduce losses of new sequences sent during recovery */
2250 const u32 loss_high
= tcp_is_sack(tp
) ? tp
->snd_nxt
: tp
->high_seq
;
2252 WARN_ON(packets
> tp
->packets_out
);
2253 if (tp
->lost_skb_hint
) {
2254 skb
= tp
->lost_skb_hint
;
2255 cnt
= tp
->lost_cnt_hint
;
2256 /* Head already handled? */
2257 if (mark_head
&& skb
!= tcp_write_queue_head(sk
))
2260 skb
= tcp_write_queue_head(sk
);
2264 tcp_for_write_queue_from(skb
, sk
) {
2265 if (skb
== tcp_send_head(sk
))
2267 /* TODO: do this better */
2268 /* this is not the most efficient way to do this... */
2269 tp
->lost_skb_hint
= skb
;
2270 tp
->lost_cnt_hint
= cnt
;
2272 if (after(TCP_SKB_CB(skb
)->end_seq
, loss_high
))
2276 if (tcp_is_fack(tp
) || tcp_is_reno(tp
) ||
2277 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2278 cnt
+= tcp_skb_pcount(skb
);
2280 if (cnt
> packets
) {
2281 if ((tcp_is_sack(tp
) && !tcp_is_fack(tp
)) ||
2282 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
) ||
2283 (oldcnt
>= packets
))
2286 mss
= skb_shinfo(skb
)->gso_size
;
2287 err
= tcp_fragment(sk
, skb
, (packets
- oldcnt
) * mss
, mss
);
2293 tcp_skb_mark_lost(tp
, skb
);
2298 tcp_verify_left_out(tp
);
2301 /* Account newly detected lost packet(s) */
2303 static void tcp_update_scoreboard(struct sock
*sk
, int fast_rexmit
)
2305 struct tcp_sock
*tp
= tcp_sk(sk
);
2307 if (tcp_is_reno(tp
)) {
2308 tcp_mark_head_lost(sk
, 1, 1);
2309 } else if (tcp_is_fack(tp
)) {
2310 int lost
= tp
->fackets_out
- tp
->reordering
;
2313 tcp_mark_head_lost(sk
, lost
, 0);
2315 int sacked_upto
= tp
->sacked_out
- tp
->reordering
;
2316 if (sacked_upto
>= 0)
2317 tcp_mark_head_lost(sk
, sacked_upto
, 0);
2318 else if (fast_rexmit
)
2319 tcp_mark_head_lost(sk
, 1, 1);
2322 tcp_timeout_skbs(sk
);
2325 /* CWND moderation, preventing bursts due to too big ACKs
2326 * in dubious situations.
2328 static inline void tcp_moderate_cwnd(struct tcp_sock
*tp
)
2330 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
2331 tcp_packets_in_flight(tp
) + tcp_max_burst(tp
));
2332 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2335 /* Nothing was retransmitted or returned timestamp is less
2336 * than timestamp of the first retransmission.
2338 static inline bool tcp_packet_delayed(const struct tcp_sock
*tp
)
2340 return !tp
->retrans_stamp
||
2341 (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2342 before(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
));
2345 /* Undo procedures. */
2347 #if FASTRETRANS_DEBUG > 1
2348 static void DBGUNDO(struct sock
*sk
, const char *msg
)
2350 struct tcp_sock
*tp
= tcp_sk(sk
);
2351 struct inet_sock
*inet
= inet_sk(sk
);
2353 if (sk
->sk_family
== AF_INET
) {
2354 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2356 &inet
->inet_daddr
, ntohs(inet
->inet_dport
),
2357 tp
->snd_cwnd
, tcp_left_out(tp
),
2358 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2361 #if IS_ENABLED(CONFIG_IPV6)
2362 else if (sk
->sk_family
== AF_INET6
) {
2363 struct ipv6_pinfo
*np
= inet6_sk(sk
);
2364 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2366 &np
->daddr
, ntohs(inet
->inet_dport
),
2367 tp
->snd_cwnd
, tcp_left_out(tp
),
2368 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2374 #define DBGUNDO(x...) do { } while (0)
2377 static void tcp_undo_cwr(struct sock
*sk
, const bool undo_ssthresh
)
2379 struct tcp_sock
*tp
= tcp_sk(sk
);
2381 if (tp
->prior_ssthresh
) {
2382 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2384 if (icsk
->icsk_ca_ops
->undo_cwnd
)
2385 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
2387 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
<< 1);
2389 if (undo_ssthresh
&& tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
2390 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
2391 TCP_ECN_withdraw_cwr(tp
);
2394 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2396 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2399 static inline bool tcp_may_undo(const struct tcp_sock
*tp
)
2401 return tp
->undo_marker
&& (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
2404 /* People celebrate: "We love our President!" */
2405 static bool tcp_try_undo_recovery(struct sock
*sk
)
2407 struct tcp_sock
*tp
= tcp_sk(sk
);
2409 if (tcp_may_undo(tp
)) {
2412 /* Happy end! We did not retransmit anything
2413 * or our original transmission succeeded.
2415 DBGUNDO(sk
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
2416 tcp_undo_cwr(sk
, true);
2417 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
2418 mib_idx
= LINUX_MIB_TCPLOSSUNDO
;
2420 mib_idx
= LINUX_MIB_TCPFULLUNDO
;
2422 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2423 tp
->undo_marker
= 0;
2425 if (tp
->snd_una
== tp
->high_seq
&& tcp_is_reno(tp
)) {
2426 /* Hold old state until something *above* high_seq
2427 * is ACKed. For Reno it is MUST to prevent false
2428 * fast retransmits (RFC2582). SACK TCP is safe. */
2429 tcp_moderate_cwnd(tp
);
2432 tcp_set_ca_state(sk
, TCP_CA_Open
);
2436 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2437 static void tcp_try_undo_dsack(struct sock
*sk
)
2439 struct tcp_sock
*tp
= tcp_sk(sk
);
2441 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
2442 DBGUNDO(sk
, "D-SACK");
2443 tcp_undo_cwr(sk
, true);
2444 tp
->undo_marker
= 0;
2445 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPDSACKUNDO
);
2449 /* We can clear retrans_stamp when there are no retransmissions in the
2450 * window. It would seem that it is trivially available for us in
2451 * tp->retrans_out, however, that kind of assumptions doesn't consider
2452 * what will happen if errors occur when sending retransmission for the
2453 * second time. ...It could the that such segment has only
2454 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2455 * the head skb is enough except for some reneging corner cases that
2456 * are not worth the effort.
2458 * Main reason for all this complexity is the fact that connection dying
2459 * time now depends on the validity of the retrans_stamp, in particular,
2460 * that successive retransmissions of a segment must not advance
2461 * retrans_stamp under any conditions.
2463 static bool tcp_any_retrans_done(const struct sock
*sk
)
2465 const struct tcp_sock
*tp
= tcp_sk(sk
);
2466 struct sk_buff
*skb
;
2468 if (tp
->retrans_out
)
2471 skb
= tcp_write_queue_head(sk
);
2472 if (unlikely(skb
&& TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
))
2478 /* Undo during fast recovery after partial ACK. */
2480 static int tcp_try_undo_partial(struct sock
*sk
, int acked
)
2482 struct tcp_sock
*tp
= tcp_sk(sk
);
2483 /* Partial ACK arrived. Force Hoe's retransmit. */
2484 int failed
= tcp_is_reno(tp
) || (tcp_fackets_out(tp
) > tp
->reordering
);
2486 if (tcp_may_undo(tp
)) {
2487 /* Plain luck! Hole if filled with delayed
2488 * packet, rather than with a retransmit.
2490 if (!tcp_any_retrans_done(sk
))
2491 tp
->retrans_stamp
= 0;
2493 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
2496 tcp_undo_cwr(sk
, false);
2497 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPARTIALUNDO
);
2499 /* So... Do not make Hoe's retransmit yet.
2500 * If the first packet was delayed, the rest
2501 * ones are most probably delayed as well.
2508 /* Undo during loss recovery after partial ACK or using F-RTO. */
2509 static bool tcp_try_undo_loss(struct sock
*sk
, bool frto_undo
)
2511 struct tcp_sock
*tp
= tcp_sk(sk
);
2513 if (frto_undo
|| tcp_may_undo(tp
)) {
2514 struct sk_buff
*skb
;
2515 tcp_for_write_queue(skb
, sk
) {
2516 if (skb
== tcp_send_head(sk
))
2518 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2521 tcp_clear_all_retrans_hints(tp
);
2523 DBGUNDO(sk
, "partial loss");
2525 tcp_undo_cwr(sk
, true);
2526 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSSUNDO
);
2528 NET_INC_STATS_BH(sock_net(sk
),
2529 LINUX_MIB_TCPSPURIOUSRTOS
);
2530 inet_csk(sk
)->icsk_retransmits
= 0;
2531 tp
->undo_marker
= 0;
2532 if (frto_undo
|| tcp_is_sack(tp
))
2533 tcp_set_ca_state(sk
, TCP_CA_Open
);
2539 /* The cwnd reduction in CWR and Recovery use the PRR algorithm
2540 * https://datatracker.ietf.org/doc/draft-ietf-tcpm-proportional-rate-reduction/
2541 * It computes the number of packets to send (sndcnt) based on packets newly
2543 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2544 * cwnd reductions across a full RTT.
2545 * 2) If packets in flight is lower than ssthresh (such as due to excess
2546 * losses and/or application stalls), do not perform any further cwnd
2547 * reductions, but instead slow start up to ssthresh.
2549 static void tcp_init_cwnd_reduction(struct sock
*sk
, const bool set_ssthresh
)
2551 struct tcp_sock
*tp
= tcp_sk(sk
);
2553 tp
->high_seq
= tp
->snd_nxt
;
2554 tp
->tlp_high_seq
= 0;
2555 tp
->snd_cwnd_cnt
= 0;
2556 tp
->prior_cwnd
= tp
->snd_cwnd
;
2557 tp
->prr_delivered
= 0;
2560 tp
->snd_ssthresh
= inet_csk(sk
)->icsk_ca_ops
->ssthresh(sk
);
2561 TCP_ECN_queue_cwr(tp
);
2564 static void tcp_cwnd_reduction(struct sock
*sk
, int newly_acked_sacked
,
2567 struct tcp_sock
*tp
= tcp_sk(sk
);
2569 int delta
= tp
->snd_ssthresh
- tcp_packets_in_flight(tp
);
2571 tp
->prr_delivered
+= newly_acked_sacked
;
2572 if (tcp_packets_in_flight(tp
) > tp
->snd_ssthresh
) {
2573 u64 dividend
= (u64
)tp
->snd_ssthresh
* tp
->prr_delivered
+
2575 sndcnt
= div_u64(dividend
, tp
->prior_cwnd
) - tp
->prr_out
;
2577 sndcnt
= min_t(int, delta
,
2578 max_t(int, tp
->prr_delivered
- tp
->prr_out
,
2579 newly_acked_sacked
) + 1);
2582 sndcnt
= max(sndcnt
, (fast_rexmit
? 1 : 0));
2583 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + sndcnt
;
2586 static inline void tcp_end_cwnd_reduction(struct sock
*sk
)
2588 struct tcp_sock
*tp
= tcp_sk(sk
);
2590 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2591 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_CWR
||
2592 (tp
->undo_marker
&& tp
->snd_ssthresh
< TCP_INFINITE_SSTHRESH
)) {
2593 tp
->snd_cwnd
= tp
->snd_ssthresh
;
2594 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2596 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
2599 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2600 void tcp_enter_cwr(struct sock
*sk
, const int set_ssthresh
)
2602 struct tcp_sock
*tp
= tcp_sk(sk
);
2604 tp
->prior_ssthresh
= 0;
2605 if (inet_csk(sk
)->icsk_ca_state
< TCP_CA_CWR
) {
2606 tp
->undo_marker
= 0;
2607 tcp_init_cwnd_reduction(sk
, set_ssthresh
);
2608 tcp_set_ca_state(sk
, TCP_CA_CWR
);
2612 static void tcp_try_keep_open(struct sock
*sk
)
2614 struct tcp_sock
*tp
= tcp_sk(sk
);
2615 int state
= TCP_CA_Open
;
2617 if (tcp_left_out(tp
) || tcp_any_retrans_done(sk
))
2618 state
= TCP_CA_Disorder
;
2620 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
2621 tcp_set_ca_state(sk
, state
);
2622 tp
->high_seq
= tp
->snd_nxt
;
2626 static void tcp_try_to_open(struct sock
*sk
, int flag
, int newly_acked_sacked
)
2628 struct tcp_sock
*tp
= tcp_sk(sk
);
2630 tcp_verify_left_out(tp
);
2632 if (!tcp_any_retrans_done(sk
))
2633 tp
->retrans_stamp
= 0;
2635 if (flag
& FLAG_ECE
)
2636 tcp_enter_cwr(sk
, 1);
2638 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
2639 tcp_try_keep_open(sk
);
2640 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
)
2641 tcp_moderate_cwnd(tp
);
2643 tcp_cwnd_reduction(sk
, newly_acked_sacked
, 0);
2647 static void tcp_mtup_probe_failed(struct sock
*sk
)
2649 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2651 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
2652 icsk
->icsk_mtup
.probe_size
= 0;
2655 static void tcp_mtup_probe_success(struct sock
*sk
)
2657 struct tcp_sock
*tp
= tcp_sk(sk
);
2658 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2660 /* FIXME: breaks with very large cwnd */
2661 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2662 tp
->snd_cwnd
= tp
->snd_cwnd
*
2663 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
2664 icsk
->icsk_mtup
.probe_size
;
2665 tp
->snd_cwnd_cnt
= 0;
2666 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2667 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2669 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
2670 icsk
->icsk_mtup
.probe_size
= 0;
2671 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
2674 /* Do a simple retransmit without using the backoff mechanisms in
2675 * tcp_timer. This is used for path mtu discovery.
2676 * The socket is already locked here.
2678 void tcp_simple_retransmit(struct sock
*sk
)
2680 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2681 struct tcp_sock
*tp
= tcp_sk(sk
);
2682 struct sk_buff
*skb
;
2683 unsigned int mss
= tcp_current_mss(sk
);
2684 u32 prior_lost
= tp
->lost_out
;
2686 tcp_for_write_queue(skb
, sk
) {
2687 if (skb
== tcp_send_head(sk
))
2689 if (tcp_skb_seglen(skb
) > mss
&&
2690 !(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
2691 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2692 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
2693 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2695 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
2699 tcp_clear_retrans_hints_partial(tp
);
2701 if (prior_lost
== tp
->lost_out
)
2704 if (tcp_is_reno(tp
))
2705 tcp_limit_reno_sacked(tp
);
2707 tcp_verify_left_out(tp
);
2709 /* Don't muck with the congestion window here.
2710 * Reason is that we do not increase amount of _data_
2711 * in network, but units changed and effective
2712 * cwnd/ssthresh really reduced now.
2714 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
2715 tp
->high_seq
= tp
->snd_nxt
;
2716 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2717 tp
->prior_ssthresh
= 0;
2718 tp
->undo_marker
= 0;
2719 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2721 tcp_xmit_retransmit_queue(sk
);
2723 EXPORT_SYMBOL(tcp_simple_retransmit
);
2725 static void tcp_enter_recovery(struct sock
*sk
, bool ece_ack
)
2727 struct tcp_sock
*tp
= tcp_sk(sk
);
2730 if (tcp_is_reno(tp
))
2731 mib_idx
= LINUX_MIB_TCPRENORECOVERY
;
2733 mib_idx
= LINUX_MIB_TCPSACKRECOVERY
;
2735 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2737 tp
->prior_ssthresh
= 0;
2738 tp
->undo_marker
= tp
->snd_una
;
2739 tp
->undo_retrans
= tp
->retrans_out
? : -1;
2741 if (inet_csk(sk
)->icsk_ca_state
< TCP_CA_CWR
) {
2743 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2744 tcp_init_cwnd_reduction(sk
, true);
2746 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
2749 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2750 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2752 static void tcp_process_loss(struct sock
*sk
, int flag
, bool is_dupack
)
2754 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2755 struct tcp_sock
*tp
= tcp_sk(sk
);
2756 bool recovered
= !before(tp
->snd_una
, tp
->high_seq
);
2758 if (tp
->frto
) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2759 /* Step 3.b. A timeout is spurious if not all data are
2760 * lost, i.e., never-retransmitted data are (s)acked.
2762 if (tcp_try_undo_loss(sk
, flag
& FLAG_ORIG_SACK_ACKED
))
2765 if (after(tp
->snd_nxt
, tp
->high_seq
) &&
2766 (flag
& FLAG_DATA_SACKED
|| is_dupack
)) {
2767 tp
->frto
= 0; /* Loss was real: 2nd part of step 3.a */
2768 } else if (flag
& FLAG_SND_UNA_ADVANCED
&& !recovered
) {
2769 tp
->high_seq
= tp
->snd_nxt
;
2770 __tcp_push_pending_frames(sk
, tcp_current_mss(sk
),
2772 if (after(tp
->snd_nxt
, tp
->high_seq
))
2773 return; /* Step 2.b */
2779 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2780 icsk
->icsk_retransmits
= 0;
2781 tcp_try_undo_recovery(sk
);
2784 if (flag
& FLAG_DATA_ACKED
)
2785 icsk
->icsk_retransmits
= 0;
2786 if (tcp_is_reno(tp
)) {
2787 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2788 * delivered. Lower inflight to clock out (re)tranmissions.
2790 if (after(tp
->snd_nxt
, tp
->high_seq
) && is_dupack
)
2791 tcp_add_reno_sack(sk
);
2792 else if (flag
& FLAG_SND_UNA_ADVANCED
)
2793 tcp_reset_reno_sack(tp
);
2795 if (tcp_try_undo_loss(sk
, false))
2797 tcp_xmit_retransmit_queue(sk
);
2800 /* Process an event, which can update packets-in-flight not trivially.
2801 * Main goal of this function is to calculate new estimate for left_out,
2802 * taking into account both packets sitting in receiver's buffer and
2803 * packets lost by network.
2805 * Besides that it does CWND reduction, when packet loss is detected
2806 * and changes state of machine.
2808 * It does _not_ decide what to send, it is made in function
2809 * tcp_xmit_retransmit_queue().
2811 static void tcp_fastretrans_alert(struct sock
*sk
, int pkts_acked
,
2812 int prior_sacked
, int prior_packets
,
2813 bool is_dupack
, int flag
)
2815 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2816 struct tcp_sock
*tp
= tcp_sk(sk
);
2817 int do_lost
= is_dupack
|| ((flag
& FLAG_DATA_SACKED
) &&
2818 (tcp_fackets_out(tp
) > tp
->reordering
));
2819 int newly_acked_sacked
= 0;
2820 int fast_rexmit
= 0;
2822 if (WARN_ON(!tp
->packets_out
&& tp
->sacked_out
))
2824 if (WARN_ON(!tp
->sacked_out
&& tp
->fackets_out
))
2825 tp
->fackets_out
= 0;
2827 /* Now state machine starts.
2828 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2829 if (flag
& FLAG_ECE
)
2830 tp
->prior_ssthresh
= 0;
2832 /* B. In all the states check for reneging SACKs. */
2833 if (tcp_check_sack_reneging(sk
, flag
))
2836 /* C. Check consistency of the current state. */
2837 tcp_verify_left_out(tp
);
2839 /* D. Check state exit conditions. State can be terminated
2840 * when high_seq is ACKed. */
2841 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
2842 WARN_ON(tp
->retrans_out
!= 0);
2843 tp
->retrans_stamp
= 0;
2844 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
2845 switch (icsk
->icsk_ca_state
) {
2847 /* CWR is to be held something *above* high_seq
2848 * is ACKed for CWR bit to reach receiver. */
2849 if (tp
->snd_una
!= tp
->high_seq
) {
2850 tcp_end_cwnd_reduction(sk
);
2851 tcp_set_ca_state(sk
, TCP_CA_Open
);
2855 case TCP_CA_Recovery
:
2856 if (tcp_is_reno(tp
))
2857 tcp_reset_reno_sack(tp
);
2858 if (tcp_try_undo_recovery(sk
))
2860 tcp_end_cwnd_reduction(sk
);
2865 /* E. Process state. */
2866 switch (icsk
->icsk_ca_state
) {
2867 case TCP_CA_Recovery
:
2868 if (!(flag
& FLAG_SND_UNA_ADVANCED
)) {
2869 if (tcp_is_reno(tp
) && is_dupack
)
2870 tcp_add_reno_sack(sk
);
2872 do_lost
= tcp_try_undo_partial(sk
, pkts_acked
);
2873 newly_acked_sacked
= prior_packets
- tp
->packets_out
+
2874 tp
->sacked_out
- prior_sacked
;
2877 tcp_process_loss(sk
, flag
, is_dupack
);
2878 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
2880 /* Fall through to processing in Open state. */
2882 if (tcp_is_reno(tp
)) {
2883 if (flag
& FLAG_SND_UNA_ADVANCED
)
2884 tcp_reset_reno_sack(tp
);
2886 tcp_add_reno_sack(sk
);
2888 newly_acked_sacked
= prior_packets
- tp
->packets_out
+
2889 tp
->sacked_out
- prior_sacked
;
2891 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
)
2892 tcp_try_undo_dsack(sk
);
2894 if (!tcp_time_to_recover(sk
, flag
)) {
2895 tcp_try_to_open(sk
, flag
, newly_acked_sacked
);
2899 /* MTU probe failure: don't reduce cwnd */
2900 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
2901 icsk
->icsk_mtup
.probe_size
&&
2902 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
2903 tcp_mtup_probe_failed(sk
);
2904 /* Restores the reduction we did in tcp_mtup_probe() */
2906 tcp_simple_retransmit(sk
);
2910 /* Otherwise enter Recovery state */
2911 tcp_enter_recovery(sk
, (flag
& FLAG_ECE
));
2915 if (do_lost
|| (tcp_is_fack(tp
) && tcp_head_timedout(sk
)))
2916 tcp_update_scoreboard(sk
, fast_rexmit
);
2917 tcp_cwnd_reduction(sk
, newly_acked_sacked
, fast_rexmit
);
2918 tcp_xmit_retransmit_queue(sk
);
2921 void tcp_valid_rtt_meas(struct sock
*sk
, u32 seq_rtt
)
2923 tcp_rtt_estimator(sk
, seq_rtt
);
2925 inet_csk(sk
)->icsk_backoff
= 0;
2927 EXPORT_SYMBOL(tcp_valid_rtt_meas
);
2929 /* Read draft-ietf-tcplw-high-performance before mucking
2930 * with this code. (Supersedes RFC1323)
2932 static void tcp_ack_saw_tstamp(struct sock
*sk
, int flag
)
2934 /* RTTM Rule: A TSecr value received in a segment is used to
2935 * update the averaged RTT measurement only if the segment
2936 * acknowledges some new data, i.e., only if it advances the
2937 * left edge of the send window.
2939 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2940 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2942 * Changed: reset backoff as soon as we see the first valid sample.
2943 * If we do not, we get strongly overestimated rto. With timestamps
2944 * samples are accepted even from very old segments: f.e., when rtt=1
2945 * increases to 8, we retransmit 5 times and after 8 seconds delayed
2946 * answer arrives rto becomes 120 seconds! If at least one of segments
2947 * in window is lost... Voila. --ANK (010210)
2949 struct tcp_sock
*tp
= tcp_sk(sk
);
2951 tcp_valid_rtt_meas(sk
, tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
);
2954 static void tcp_ack_no_tstamp(struct sock
*sk
, u32 seq_rtt
, int flag
)
2956 /* We don't have a timestamp. Can only use
2957 * packets that are not retransmitted to determine
2958 * rtt estimates. Also, we must not reset the
2959 * backoff for rto until we get a non-retransmitted
2960 * packet. This allows us to deal with a situation
2961 * where the network delay has increased suddenly.
2962 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2965 if (flag
& FLAG_RETRANS_DATA_ACKED
)
2968 tcp_valid_rtt_meas(sk
, seq_rtt
);
2971 static inline void tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
2974 const struct tcp_sock
*tp
= tcp_sk(sk
);
2975 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2976 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
2977 tcp_ack_saw_tstamp(sk
, flag
);
2978 else if (seq_rtt
>= 0)
2979 tcp_ack_no_tstamp(sk
, seq_rtt
, flag
);
2982 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
, u32 in_flight
)
2984 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2985 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, in_flight
);
2986 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_time_stamp
;
2989 /* Restart timer after forward progress on connection.
2990 * RFC2988 recommends to restart timer to now+rto.
2992 void tcp_rearm_rto(struct sock
*sk
)
2994 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2995 struct tcp_sock
*tp
= tcp_sk(sk
);
2997 /* If the retrans timer is currently being used by Fast Open
2998 * for SYN-ACK retrans purpose, stay put.
3000 if (tp
->fastopen_rsk
)
3003 if (!tp
->packets_out
) {
3004 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
3006 u32 rto
= inet_csk(sk
)->icsk_rto
;
3007 /* Offset the time elapsed after installing regular RTO */
3008 if (icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
||
3009 icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
) {
3010 struct sk_buff
*skb
= tcp_write_queue_head(sk
);
3011 const u32 rto_time_stamp
= TCP_SKB_CB(skb
)->when
+ rto
;
3012 s32 delta
= (s32
)(rto_time_stamp
- tcp_time_stamp
);
3013 /* delta may not be positive if the socket is locked
3014 * when the retrans timer fires and is rescheduled.
3019 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
, rto
,
3020 sysctl_tcp_rto_max
);
3024 /* This function is called when the delayed ER timer fires. TCP enters
3025 * fast recovery and performs fast-retransmit.
3027 void tcp_resume_early_retransmit(struct sock
*sk
)
3029 struct tcp_sock
*tp
= tcp_sk(sk
);
3033 /* Stop if ER is disabled after the delayed ER timer is scheduled */
3034 if (!tp
->do_early_retrans
)
3037 tcp_enter_recovery(sk
, false);
3038 tcp_update_scoreboard(sk
, 1);
3039 tcp_xmit_retransmit_queue(sk
);
3042 /* If we get here, the whole TSO packet has not been acked. */
3043 static u32
tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
)
3045 struct tcp_sock
*tp
= tcp_sk(sk
);
3048 BUG_ON(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
));
3050 packets_acked
= tcp_skb_pcount(skb
);
3051 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
3053 packets_acked
-= tcp_skb_pcount(skb
);
3055 if (packets_acked
) {
3056 BUG_ON(tcp_skb_pcount(skb
) == 0);
3057 BUG_ON(!before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
));
3060 return packets_acked
;
3063 /* Remove acknowledged frames from the retransmission queue. If our packet
3064 * is before the ack sequence we can discard it as it's confirmed to have
3065 * arrived at the other end.
3067 static int tcp_clean_rtx_queue(struct sock
*sk
, int prior_fackets
,
3070 struct tcp_sock
*tp
= tcp_sk(sk
);
3071 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3072 struct sk_buff
*skb
;
3073 u32 now
= tcp_time_stamp
;
3074 int fully_acked
= true;
3077 u32 reord
= tp
->packets_out
;
3078 u32 prior_sacked
= tp
->sacked_out
;
3080 s32 ca_seq_rtt
= -1;
3081 ktime_t last_ackt
= net_invalid_timestamp();
3083 while ((skb
= tcp_write_queue_head(sk
)) && skb
!= tcp_send_head(sk
)) {
3084 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
3086 u8 sacked
= scb
->sacked
;
3088 /* Determine how many packets and what bytes were acked, tso and else */
3089 if (after(scb
->end_seq
, tp
->snd_una
)) {
3090 if (tcp_skb_pcount(skb
) == 1 ||
3091 !after(tp
->snd_una
, scb
->seq
))
3094 acked_pcount
= tcp_tso_acked(sk
, skb
);
3098 fully_acked
= false;
3100 acked_pcount
= tcp_skb_pcount(skb
);
3103 if (sacked
& TCPCB_RETRANS
) {
3104 if (sacked
& TCPCB_SACKED_RETRANS
)
3105 tp
->retrans_out
-= acked_pcount
;
3106 flag
|= FLAG_RETRANS_DATA_ACKED
;
3110 ca_seq_rtt
= now
- scb
->when
;
3111 last_ackt
= skb
->tstamp
;
3113 seq_rtt
= ca_seq_rtt
;
3115 if (!(sacked
& TCPCB_SACKED_ACKED
)) {
3116 reord
= min(pkts_acked
, reord
);
3117 if (!after(scb
->end_seq
, tp
->high_seq
))
3118 flag
|= FLAG_ORIG_SACK_ACKED
;
3122 if (sacked
& TCPCB_SACKED_ACKED
)
3123 tp
->sacked_out
-= acked_pcount
;
3124 if (sacked
& TCPCB_LOST
)
3125 tp
->lost_out
-= acked_pcount
;
3127 tp
->packets_out
-= acked_pcount
;
3128 pkts_acked
+= acked_pcount
;
3130 /* Initial outgoing SYN's get put onto the write_queue
3131 * just like anything else we transmit. It is not
3132 * true data, and if we misinform our callers that
3133 * this ACK acks real data, we will erroneously exit
3134 * connection startup slow start one packet too
3135 * quickly. This is severely frowned upon behavior.
3137 if (!(scb
->tcp_flags
& TCPHDR_SYN
)) {
3138 flag
|= FLAG_DATA_ACKED
;
3140 flag
|= FLAG_SYN_ACKED
;
3141 tp
->retrans_stamp
= 0;
3147 tcp_unlink_write_queue(skb
, sk
);
3148 sk_wmem_free_skb(sk
, skb
);
3149 tp
->scoreboard_skb_hint
= NULL
;
3150 if (skb
== tp
->retransmit_skb_hint
)
3151 tp
->retransmit_skb_hint
= NULL
;
3152 if (skb
== tp
->lost_skb_hint
)
3153 tp
->lost_skb_hint
= NULL
;
3156 if (likely(between(tp
->snd_up
, prior_snd_una
, tp
->snd_una
)))
3157 tp
->snd_up
= tp
->snd_una
;
3159 if (skb
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
3160 flag
|= FLAG_SACK_RENEGING
;
3162 if (flag
& FLAG_ACKED
) {
3163 const struct tcp_congestion_ops
*ca_ops
3164 = inet_csk(sk
)->icsk_ca_ops
;
3166 if (unlikely(icsk
->icsk_mtup
.probe_size
&&
3167 !after(tp
->mtu_probe
.probe_seq_end
, tp
->snd_una
))) {
3168 tcp_mtup_probe_success(sk
);
3171 tcp_ack_update_rtt(sk
, flag
, seq_rtt
);
3174 if (tcp_is_reno(tp
)) {
3175 tcp_remove_reno_sacks(sk
, pkts_acked
);
3179 /* Non-retransmitted hole got filled? That's reordering */
3180 if (reord
< prior_fackets
)
3181 tcp_update_reordering(sk
, tp
->fackets_out
- reord
, 0);
3183 delta
= tcp_is_fack(tp
) ? pkts_acked
:
3184 prior_sacked
- tp
->sacked_out
;
3185 tp
->lost_cnt_hint
-= min(tp
->lost_cnt_hint
, delta
);
3188 tp
->fackets_out
-= min(pkts_acked
, tp
->fackets_out
);
3190 if (ca_ops
->pkts_acked
) {
3193 /* Is the ACK triggering packet unambiguous? */
3194 if (!(flag
& FLAG_RETRANS_DATA_ACKED
)) {
3195 /* High resolution needed and available? */
3196 if (ca_ops
->flags
& TCP_CONG_RTT_STAMP
&&
3197 !ktime_equal(last_ackt
,
3198 net_invalid_timestamp()))
3199 rtt_us
= ktime_us_delta(ktime_get_real(),
3201 else if (ca_seq_rtt
>= 0)
3202 rtt_us
= jiffies_to_usecs(ca_seq_rtt
);
3205 ca_ops
->pkts_acked(sk
, pkts_acked
, rtt_us
);
3209 #if FASTRETRANS_DEBUG > 0
3210 WARN_ON((int)tp
->sacked_out
< 0);
3211 WARN_ON((int)tp
->lost_out
< 0);
3212 WARN_ON((int)tp
->retrans_out
< 0);
3213 if (!tp
->packets_out
&& tcp_is_sack(tp
)) {
3214 icsk
= inet_csk(sk
);
3216 pr_debug("Leak l=%u %d\n",
3217 tp
->lost_out
, icsk
->icsk_ca_state
);
3220 if (tp
->sacked_out
) {
3221 pr_debug("Leak s=%u %d\n",
3222 tp
->sacked_out
, icsk
->icsk_ca_state
);
3225 if (tp
->retrans_out
) {
3226 pr_debug("Leak r=%u %d\n",
3227 tp
->retrans_out
, icsk
->icsk_ca_state
);
3228 tp
->retrans_out
= 0;
3235 static void tcp_ack_probe(struct sock
*sk
)
3237 const struct tcp_sock
*tp
= tcp_sk(sk
);
3238 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3240 /* Was it a usable window open? */
3242 if (!after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
, tcp_wnd_end(tp
))) {
3243 icsk
->icsk_backoff
= 0;
3244 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
3245 /* Socket must be waked up by subsequent tcp_data_snd_check().
3246 * This function is not for random using!
3249 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3250 min_t(unsigned int, icsk
->icsk_rto
<< icsk
->icsk_backoff
, sysctl_tcp_rto_max
),
3251 sysctl_tcp_rto_max
);
3255 static inline bool tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
3257 return !(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
3258 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
;
3261 static inline bool tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
3263 const struct tcp_sock
*tp
= tcp_sk(sk
);
3264 return (!(flag
& FLAG_ECE
) || tp
->snd_cwnd
< tp
->snd_ssthresh
) &&
3265 !tcp_in_cwnd_reduction(sk
);
3268 /* Check that window update is acceptable.
3269 * The function assumes that snd_una<=ack<=snd_next.
3271 static inline bool tcp_may_update_window(const struct tcp_sock
*tp
,
3272 const u32 ack
, const u32 ack_seq
,
3275 return after(ack
, tp
->snd_una
) ||
3276 after(ack_seq
, tp
->snd_wl1
) ||
3277 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
);
3280 /* Update our send window.
3282 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3283 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3285 static int tcp_ack_update_window(struct sock
*sk
, const struct sk_buff
*skb
, u32 ack
,
3288 struct tcp_sock
*tp
= tcp_sk(sk
);
3290 u32 nwin
= ntohs(tcp_hdr(skb
)->window
);
3292 if (likely(!tcp_hdr(skb
)->syn
))
3293 nwin
<<= tp
->rx_opt
.snd_wscale
;
3295 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
3296 flag
|= FLAG_WIN_UPDATE
;
3297 tcp_update_wl(tp
, ack_seq
);
3299 if (tp
->snd_wnd
!= nwin
) {
3302 /* Note, it is the only place, where
3303 * fast path is recovered for sending TCP.
3306 tcp_fast_path_check(sk
);
3308 if (nwin
> tp
->max_window
) {
3309 tp
->max_window
= nwin
;
3310 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
3320 /* RFC 5961 7 [ACK Throttling] */
3321 static void tcp_send_challenge_ack(struct sock
*sk
)
3323 /* unprotected vars, we dont care of overwrites */
3324 static u32 challenge_timestamp
;
3325 static unsigned int challenge_count
;
3326 u32 now
= jiffies
/ HZ
;
3329 if (now
!= challenge_timestamp
) {
3330 u32 half
= (sysctl_tcp_challenge_ack_limit
+ 1) >> 1;
3332 challenge_timestamp
= now
;
3333 ACCESS_ONCE(challenge_count
) = half
+
3334 reciprocal_divide(prandom_u32(),
3335 sysctl_tcp_challenge_ack_limit
);
3337 count
= ACCESS_ONCE(challenge_count
);
3339 ACCESS_ONCE(challenge_count
) = count
- 1;
3340 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPCHALLENGEACK
);
3345 static void tcp_store_ts_recent(struct tcp_sock
*tp
)
3347 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
3348 tp
->rx_opt
.ts_recent_stamp
= get_seconds();
3351 static void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
3353 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
3354 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3355 * extra check below makes sure this can only happen
3356 * for pure ACK frames. -DaveM
3358 * Not only, also it occurs for expired timestamps.
3361 if (tcp_paws_check(&tp
->rx_opt
, 0))
3362 tcp_store_ts_recent(tp
);
3366 /* This routine deals with acks during a TLP episode.
3367 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3369 static void tcp_process_tlp_ack(struct sock
*sk
, u32 ack
, int flag
)
3371 struct tcp_sock
*tp
= tcp_sk(sk
);
3372 bool is_tlp_dupack
= (ack
== tp
->tlp_high_seq
) &&
3373 !(flag
& (FLAG_SND_UNA_ADVANCED
|
3374 FLAG_NOT_DUP
| FLAG_DATA_SACKED
));
3376 /* Mark the end of TLP episode on receiving TLP dupack or when
3377 * ack is after tlp_high_seq.
3379 if (is_tlp_dupack
) {
3380 tp
->tlp_high_seq
= 0;
3384 if (after(ack
, tp
->tlp_high_seq
)) {
3385 tp
->tlp_high_seq
= 0;
3386 /* Don't reduce cwnd if DSACK arrives for TLP retrans. */
3387 if (!(flag
& FLAG_DSACKING_ACK
)) {
3388 tcp_init_cwnd_reduction(sk
, true);
3389 tcp_set_ca_state(sk
, TCP_CA_CWR
);
3390 tcp_end_cwnd_reduction(sk
);
3391 tcp_try_keep_open(sk
);
3392 NET_INC_STATS_BH(sock_net(sk
),
3393 LINUX_MIB_TCPLOSSPROBERECOVERY
);
3398 /* This routine deals with incoming acks, but not outgoing ones. */
3399 static int tcp_ack(struct sock
*sk
, const struct sk_buff
*skb
, int flag
)
3401 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3402 struct tcp_sock
*tp
= tcp_sk(sk
);
3403 u32 prior_snd_una
= tp
->snd_una
;
3404 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
3405 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3406 bool is_dupack
= false;
3407 u32 prior_in_flight
, prior_cwnd
= tp
->snd_cwnd
, prior_rtt
= tp
->srtt
;
3409 int prior_packets
= tp
->packets_out
;
3410 int prior_sacked
= tp
->sacked_out
;
3412 int previous_packets_out
= 0;
3414 /* If the ack is older than previous acks
3415 * then we can probably ignore it.
3417 if (before(ack
, prior_snd_una
)) {
3418 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3419 if (before(ack
, prior_snd_una
- tp
->max_window
)) {
3420 tcp_send_challenge_ack(sk
);
3426 /* If the ack includes data we haven't sent yet, discard
3427 * this segment (RFC793 Section 3.9).
3429 if (after(ack
, tp
->snd_nxt
))
3432 if (icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
||
3433 icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
)
3436 if (after(ack
, prior_snd_una
))
3437 flag
|= FLAG_SND_UNA_ADVANCED
;
3439 prior_fackets
= tp
->fackets_out
;
3440 prior_in_flight
= tcp_packets_in_flight(tp
);
3442 /* ts_recent update must be made after we are sure that the packet
3445 if (flag
& FLAG_UPDATE_TS_RECENT
)
3446 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
3448 if (!(flag
& FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
3449 /* Window is constant, pure forward advance.
3450 * No more checks are required.
3451 * Note, we use the fact that SND.UNA>=SND.WL2.
3453 tcp_update_wl(tp
, ack_seq
);
3455 flag
|= FLAG_WIN_UPDATE
;
3457 tcp_ca_event(sk
, CA_EVENT_FAST_ACK
);
3459 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPACKS
);
3461 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
3464 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPUREACKS
);
3466 flag
|= tcp_ack_update_window(sk
, skb
, ack
, ack_seq
);
3468 if (TCP_SKB_CB(skb
)->sacked
)
3469 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
3471 if (TCP_ECN_rcv_ecn_echo(tp
, tcp_hdr(skb
)))
3474 tcp_ca_event(sk
, CA_EVENT_SLOW_ACK
);
3477 /* We passed data and got it acked, remove any soft error
3478 * log. Something worked...
3480 sk
->sk_err_soft
= 0;
3481 icsk
->icsk_probes_out
= 0;
3482 tp
->rcv_tstamp
= tcp_time_stamp
;
3486 /* See if we can take anything off of the retransmit queue. */
3487 previous_packets_out
= tp
->packets_out
;
3488 flag
|= tcp_clean_rtx_queue(sk
, prior_fackets
, prior_snd_una
);
3490 pkts_acked
= previous_packets_out
- tp
->packets_out
;
3492 if (tcp_ack_is_dubious(sk
, flag
)) {
3493 /* Advance CWND, if state allows this. */
3494 if ((flag
& FLAG_DATA_ACKED
) && tcp_may_raise_cwnd(sk
, flag
))
3495 tcp_cong_avoid(sk
, ack
, prior_in_flight
);
3496 is_dupack
= !(flag
& (FLAG_SND_UNA_ADVANCED
| FLAG_NOT_DUP
));
3497 tcp_fastretrans_alert(sk
, pkts_acked
, prior_sacked
,
3498 prior_packets
, is_dupack
, flag
);
3500 if (flag
& FLAG_DATA_ACKED
)
3501 tcp_cong_avoid(sk
, ack
, prior_in_flight
);
3504 if (tp
->tlp_high_seq
)
3505 tcp_process_tlp_ack(sk
, ack
, flag
);
3507 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
& FLAG_NOT_DUP
)) {
3508 struct dst_entry
*dst
= __sk_dst_get(sk
);
3513 if (icsk
->icsk_pending
== ICSK_TIME_RETRANS
)
3514 tcp_schedule_loss_probe(sk
);
3515 if (tp
->srtt
!= prior_rtt
|| tp
->snd_cwnd
!= prior_cwnd
)
3516 tcp_update_pacing_rate(sk
);
3520 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3521 if (flag
& FLAG_DSACKING_ACK
)
3522 tcp_fastretrans_alert(sk
, pkts_acked
, prior_sacked
,
3523 prior_packets
, is_dupack
, flag
);
3524 /* If this ack opens up a zero window, clear backoff. It was
3525 * being used to time the probes, and is probably far higher than
3526 * it needs to be for normal retransmission.
3528 if (tcp_send_head(sk
))
3531 if (tp
->tlp_high_seq
)
3532 tcp_process_tlp_ack(sk
, ack
, flag
);
3536 SOCK_DEBUG(sk
, "Ack %u after %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3540 /* If data was SACKed, tag it and see if we should send more data.
3541 * If data was DSACKed, see if we can undo a cwnd reduction.
3543 if (TCP_SKB_CB(skb
)->sacked
) {
3544 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
3545 tcp_fastretrans_alert(sk
, pkts_acked
, prior_sacked
,
3546 prior_packets
, is_dupack
, flag
);
3549 SOCK_DEBUG(sk
, "Ack %u before %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3553 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3554 * But, this can also be called on packets in the established flow when
3555 * the fast version below fails.
3557 void tcp_parse_options(const struct sk_buff
*skb
,
3558 struct tcp_options_received
*opt_rx
, int estab
,
3559 struct tcp_fastopen_cookie
*foc
)
3561 const unsigned char *ptr
;
3562 const struct tcphdr
*th
= tcp_hdr(skb
);
3563 int length
= (th
->doff
* 4) - sizeof(struct tcphdr
);
3565 ptr
= (const unsigned char *)(th
+ 1);
3566 opt_rx
->saw_tstamp
= 0;
3568 while (length
> 0) {
3569 int opcode
= *ptr
++;
3575 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
3580 if (opsize
< 2) /* "silly options" */
3582 if (opsize
> length
)
3583 return; /* don't parse partial options */
3586 if (opsize
== TCPOLEN_MSS
&& th
->syn
&& !estab
) {
3587 u16 in_mss
= get_unaligned_be16(ptr
);
3589 if (opt_rx
->user_mss
&&
3590 opt_rx
->user_mss
< in_mss
)
3591 in_mss
= opt_rx
->user_mss
;
3592 opt_rx
->mss_clamp
= in_mss
;
3597 if (opsize
== TCPOLEN_WINDOW
&& th
->syn
&&
3598 !estab
&& sysctl_tcp_window_scaling
) {
3599 __u8 snd_wscale
= *(__u8
*)ptr
;
3600 opt_rx
->wscale_ok
= 1;
3601 if (snd_wscale
> 14) {
3602 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3607 opt_rx
->snd_wscale
= snd_wscale
;
3610 case TCPOPT_TIMESTAMP
:
3611 if ((opsize
== TCPOLEN_TIMESTAMP
) &&
3612 ((estab
&& opt_rx
->tstamp_ok
) ||
3613 (!estab
&& sysctl_tcp_timestamps
))) {
3614 opt_rx
->saw_tstamp
= 1;
3615 opt_rx
->rcv_tsval
= get_unaligned_be32(ptr
);
3616 opt_rx
->rcv_tsecr
= get_unaligned_be32(ptr
+ 4);
3619 case TCPOPT_SACK_PERM
:
3620 if (opsize
== TCPOLEN_SACK_PERM
&& th
->syn
&&
3621 !estab
&& sysctl_tcp_sack
) {
3622 opt_rx
->sack_ok
= TCP_SACK_SEEN
;
3623 tcp_sack_reset(opt_rx
);
3628 if ((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
3629 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
3631 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
3634 #ifdef CONFIG_TCP_MD5SIG
3637 * The MD5 Hash has already been
3638 * checked (see tcp_v{4,6}_do_rcv()).
3643 /* Fast Open option shares code 254 using a
3644 * 16 bits magic number. It's valid only in
3645 * SYN or SYN-ACK with an even size.
3647 if (opsize
< TCPOLEN_EXP_FASTOPEN_BASE
||
3648 get_unaligned_be16(ptr
) != TCPOPT_FASTOPEN_MAGIC
||
3649 foc
== NULL
|| !th
->syn
|| (opsize
& 1))
3651 foc
->len
= opsize
- TCPOLEN_EXP_FASTOPEN_BASE
;
3652 if (foc
->len
>= TCP_FASTOPEN_COOKIE_MIN
&&
3653 foc
->len
<= TCP_FASTOPEN_COOKIE_MAX
)
3654 memcpy(foc
->val
, ptr
+ 2, foc
->len
);
3655 else if (foc
->len
!= 0)
3665 EXPORT_SYMBOL(tcp_parse_options
);
3667 static bool tcp_parse_aligned_timestamp(struct tcp_sock
*tp
, const struct tcphdr
*th
)
3669 const __be32
*ptr
= (const __be32
*)(th
+ 1);
3671 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3672 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
3673 tp
->rx_opt
.saw_tstamp
= 1;
3675 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3678 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
) - tp
->tsoffset
;
3680 tp
->rx_opt
.rcv_tsecr
= 0;
3686 /* Fast parse options. This hopes to only see timestamps.
3687 * If it is wrong it falls back on tcp_parse_options().
3689 static bool tcp_fast_parse_options(const struct sk_buff
*skb
,
3690 const struct tcphdr
*th
, struct tcp_sock
*tp
)
3692 /* In the spirit of fast parsing, compare doff directly to constant
3693 * values. Because equality is used, short doff can be ignored here.
3695 if (th
->doff
== (sizeof(*th
) / 4)) {
3696 tp
->rx_opt
.saw_tstamp
= 0;
3698 } else if (tp
->rx_opt
.tstamp_ok
&&
3699 th
->doff
== ((sizeof(*th
) + TCPOLEN_TSTAMP_ALIGNED
) / 4)) {
3700 if (tcp_parse_aligned_timestamp(tp
, th
))
3704 tcp_parse_options(skb
, &tp
->rx_opt
, 1, NULL
);
3705 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
3706 tp
->rx_opt
.rcv_tsecr
-= tp
->tsoffset
;
3711 #ifdef CONFIG_TCP_MD5SIG
3713 * Parse MD5 Signature option
3715 const u8
*tcp_parse_md5sig_option(const struct tcphdr
*th
)
3717 int length
= (th
->doff
<< 2) - sizeof(*th
);
3718 const u8
*ptr
= (const u8
*)(th
+ 1);
3720 /* If the TCP option is too short, we can short cut */
3721 if (length
< TCPOLEN_MD5SIG
)
3724 while (length
> 0) {
3725 int opcode
= *ptr
++;
3736 if (opsize
< 2 || opsize
> length
)
3738 if (opcode
== TCPOPT_MD5SIG
)
3739 return opsize
== TCPOLEN_MD5SIG
? ptr
: NULL
;
3746 EXPORT_SYMBOL(tcp_parse_md5sig_option
);
3749 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3751 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3752 * it can pass through stack. So, the following predicate verifies that
3753 * this segment is not used for anything but congestion avoidance or
3754 * fast retransmit. Moreover, we even are able to eliminate most of such
3755 * second order effects, if we apply some small "replay" window (~RTO)
3756 * to timestamp space.
3758 * All these measures still do not guarantee that we reject wrapped ACKs
3759 * on networks with high bandwidth, when sequence space is recycled fastly,
3760 * but it guarantees that such events will be very rare and do not affect
3761 * connection seriously. This doesn't look nice, but alas, PAWS is really
3764 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3765 * states that events when retransmit arrives after original data are rare.
3766 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3767 * the biggest problem on large power networks even with minor reordering.
3768 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3769 * up to bandwidth of 18Gigabit/sec. 8) ]
3772 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
3774 const struct tcp_sock
*tp
= tcp_sk(sk
);
3775 const struct tcphdr
*th
= tcp_hdr(skb
);
3776 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3777 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3779 return (/* 1. Pure ACK with correct sequence number. */
3780 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
3782 /* 2. ... and duplicate ACK. */
3783 ack
== tp
->snd_una
&&
3785 /* 3. ... and does not update window. */
3786 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
3788 /* 4. ... and sits in replay window. */
3789 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
3792 static inline bool tcp_paws_discard(const struct sock
*sk
,
3793 const struct sk_buff
*skb
)
3795 const struct tcp_sock
*tp
= tcp_sk(sk
);
3797 return !tcp_paws_check(&tp
->rx_opt
, TCP_PAWS_WINDOW
) &&
3798 !tcp_disordered_ack(sk
, skb
);
3801 /* Check segment sequence number for validity.
3803 * Segment controls are considered valid, if the segment
3804 * fits to the window after truncation to the window. Acceptability
3805 * of data (and SYN, FIN, of course) is checked separately.
3806 * See tcp_data_queue(), for example.
3808 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3809 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3810 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3811 * (borrowed from freebsd)
3814 static inline bool tcp_sequence(const struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3816 return !before(end_seq
, tp
->rcv_wup
) &&
3817 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
3820 /* When we get a reset we do this. */
3821 void tcp_reset(struct sock
*sk
)
3823 /* We want the right error as BSD sees it (and indeed as we do). */
3824 switch (sk
->sk_state
) {
3826 sk
->sk_err
= ECONNREFUSED
;
3828 case TCP_CLOSE_WAIT
:
3834 sk
->sk_err
= ECONNRESET
;
3836 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3839 if (!sock_flag(sk
, SOCK_DEAD
))
3840 sk
->sk_error_report(sk
);
3846 * Process the FIN bit. This now behaves as it is supposed to work
3847 * and the FIN takes effect when it is validly part of sequence
3848 * space. Not before when we get holes.
3850 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3851 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3854 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3855 * close and we go into CLOSING (and later onto TIME-WAIT)
3857 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3859 static void tcp_fin(struct sock
*sk
)
3861 struct tcp_sock
*tp
= tcp_sk(sk
);
3863 inet_csk_schedule_ack(sk
);
3865 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
3866 sock_set_flag(sk
, SOCK_DONE
);
3868 switch (sk
->sk_state
) {
3870 case TCP_ESTABLISHED
:
3871 /* Move to CLOSE_WAIT */
3872 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
3873 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
3876 case TCP_CLOSE_WAIT
:
3878 /* Received a retransmission of the FIN, do
3883 /* RFC793: Remain in the LAST-ACK state. */
3887 /* This case occurs when a simultaneous close
3888 * happens, we must ack the received FIN and
3889 * enter the CLOSING state.
3892 tcp_set_state(sk
, TCP_CLOSING
);
3895 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3897 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
3900 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3901 * cases we should never reach this piece of code.
3903 pr_err("%s: Impossible, sk->sk_state=%d\n",
3904 __func__
, sk
->sk_state
);
3908 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3909 * Probably, we should reset in this case. For now drop them.
3911 __skb_queue_purge(&tp
->out_of_order_queue
);
3912 if (tcp_is_sack(tp
))
3913 tcp_sack_reset(&tp
->rx_opt
);
3916 if (!sock_flag(sk
, SOCK_DEAD
)) {
3917 sk
->sk_state_change(sk
);
3919 /* Do not send POLL_HUP for half duplex close. */
3920 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
3921 sk
->sk_state
== TCP_CLOSE
)
3922 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_HUP
);
3924 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_IN
);
3928 static inline bool tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
,
3931 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
3932 if (before(seq
, sp
->start_seq
))
3933 sp
->start_seq
= seq
;
3934 if (after(end_seq
, sp
->end_seq
))
3935 sp
->end_seq
= end_seq
;
3941 static void tcp_dsack_set(struct sock
*sk
, u32 seq
, u32 end_seq
)
3943 struct tcp_sock
*tp
= tcp_sk(sk
);
3945 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
3948 if (before(seq
, tp
->rcv_nxt
))
3949 mib_idx
= LINUX_MIB_TCPDSACKOLDSENT
;
3951 mib_idx
= LINUX_MIB_TCPDSACKOFOSENT
;
3953 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
3955 tp
->rx_opt
.dsack
= 1;
3956 tp
->duplicate_sack
[0].start_seq
= seq
;
3957 tp
->duplicate_sack
[0].end_seq
= end_seq
;
3961 static void tcp_dsack_extend(struct sock
*sk
, u32 seq
, u32 end_seq
)
3963 struct tcp_sock
*tp
= tcp_sk(sk
);
3965 if (!tp
->rx_opt
.dsack
)
3966 tcp_dsack_set(sk
, seq
, end_seq
);
3968 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
3971 static void tcp_send_dupack(struct sock
*sk
, const struct sk_buff
*skb
)
3973 struct tcp_sock
*tp
= tcp_sk(sk
);
3975 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
3976 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
3977 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
3978 tcp_enter_quickack_mode(sk
);
3980 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
3981 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
3983 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
3984 end_seq
= tp
->rcv_nxt
;
3985 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, end_seq
);
3992 /* These routines update the SACK block as out-of-order packets arrive or
3993 * in-order packets close up the sequence space.
3995 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
3998 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3999 struct tcp_sack_block
*swalk
= sp
+ 1;
4001 /* See if the recent change to the first SACK eats into
4002 * or hits the sequence space of other SACK blocks, if so coalesce.
4004 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
;) {
4005 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
4008 /* Zap SWALK, by moving every further SACK up by one slot.
4009 * Decrease num_sacks.
4011 tp
->rx_opt
.num_sacks
--;
4012 for (i
= this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
4016 this_sack
++, swalk
++;
4020 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
4022 struct tcp_sock
*tp
= tcp_sk(sk
);
4023 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4024 int cur_sacks
= tp
->rx_opt
.num_sacks
;
4030 for (this_sack
= 0; this_sack
< cur_sacks
; this_sack
++, sp
++) {
4031 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
4032 /* Rotate this_sack to the first one. */
4033 for (; this_sack
> 0; this_sack
--, sp
--)
4034 swap(*sp
, *(sp
- 1));
4036 tcp_sack_maybe_coalesce(tp
);
4041 /* Could not find an adjacent existing SACK, build a new one,
4042 * put it at the front, and shift everyone else down. We
4043 * always know there is at least one SACK present already here.
4045 * If the sack array is full, forget about the last one.
4047 if (this_sack
>= TCP_NUM_SACKS
) {
4049 tp
->rx_opt
.num_sacks
--;
4052 for (; this_sack
> 0; this_sack
--, sp
--)
4056 /* Build the new head SACK, and we're done. */
4057 sp
->start_seq
= seq
;
4058 sp
->end_seq
= end_seq
;
4059 tp
->rx_opt
.num_sacks
++;
4062 /* RCV.NXT advances, some SACKs should be eaten. */
4064 static void tcp_sack_remove(struct tcp_sock
*tp
)
4066 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4067 int num_sacks
= tp
->rx_opt
.num_sacks
;
4070 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4071 if (skb_queue_empty(&tp
->out_of_order_queue
)) {
4072 tp
->rx_opt
.num_sacks
= 0;
4076 for (this_sack
= 0; this_sack
< num_sacks
;) {
4077 /* Check if the start of the sack is covered by RCV.NXT. */
4078 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
4081 /* RCV.NXT must cover all the block! */
4082 WARN_ON(before(tp
->rcv_nxt
, sp
->end_seq
));
4084 /* Zap this SACK, by moving forward any other SACKS. */
4085 for (i
=this_sack
+1; i
< num_sacks
; i
++)
4086 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
4093 tp
->rx_opt
.num_sacks
= num_sacks
;
4096 /* This one checks to see if we can put data from the
4097 * out_of_order queue into the receive_queue.
4099 static void tcp_ofo_queue(struct sock
*sk
)
4101 struct tcp_sock
*tp
= tcp_sk(sk
);
4102 __u32 dsack_high
= tp
->rcv_nxt
;
4103 struct sk_buff
*skb
;
4105 while ((skb
= skb_peek(&tp
->out_of_order_queue
)) != NULL
) {
4106 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
4109 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
4110 __u32 dsack
= dsack_high
;
4111 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
4112 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
4113 tcp_dsack_extend(sk
, TCP_SKB_CB(skb
)->seq
, dsack
);
4116 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4117 SOCK_DEBUG(sk
, "ofo packet was already received\n");
4118 __skb_unlink(skb
, &tp
->out_of_order_queue
);
4122 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
4123 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4124 TCP_SKB_CB(skb
)->end_seq
);
4126 __skb_unlink(skb
, &tp
->out_of_order_queue
);
4127 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4128 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4129 if (tcp_hdr(skb
)->fin
)
4134 static bool tcp_prune_ofo_queue(struct sock
*sk
);
4135 static int tcp_prune_queue(struct sock
*sk
);
4137 static int tcp_try_rmem_schedule(struct sock
*sk
, struct sk_buff
*skb
,
4140 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
4141 !sk_rmem_schedule(sk
, skb
, size
)) {
4143 if (tcp_prune_queue(sk
) < 0)
4146 if (!sk_rmem_schedule(sk
, skb
, size
)) {
4147 if (!tcp_prune_ofo_queue(sk
))
4150 if (!sk_rmem_schedule(sk
, skb
, size
))
4158 * tcp_try_coalesce - try to merge skb to prior one
4161 * @from: buffer to add in queue
4162 * @fragstolen: pointer to boolean
4164 * Before queueing skb @from after @to, try to merge them
4165 * to reduce overall memory use and queue lengths, if cost is small.
4166 * Packets in ofo or receive queues can stay a long time.
4167 * Better try to coalesce them right now to avoid future collapses.
4168 * Returns true if caller should free @from instead of queueing it
4170 static bool tcp_try_coalesce(struct sock
*sk
,
4172 struct sk_buff
*from
,
4177 *fragstolen
= false;
4179 if (tcp_hdr(from
)->fin
)
4182 /* Its possible this segment overlaps with prior segment in queue */
4183 if (TCP_SKB_CB(from
)->seq
!= TCP_SKB_CB(to
)->end_seq
)
4186 if (!skb_try_coalesce(to
, from
, fragstolen
, &delta
))
4189 atomic_add(delta
, &sk
->sk_rmem_alloc
);
4190 sk_mem_charge(sk
, delta
);
4191 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRCVCOALESCE
);
4192 TCP_SKB_CB(to
)->end_seq
= TCP_SKB_CB(from
)->end_seq
;
4193 TCP_SKB_CB(to
)->ack_seq
= TCP_SKB_CB(from
)->ack_seq
;
4197 static void tcp_data_queue_ofo(struct sock
*sk
, struct sk_buff
*skb
)
4199 struct tcp_sock
*tp
= tcp_sk(sk
);
4200 struct sk_buff
*skb1
;
4203 TCP_ECN_check_ce(tp
, skb
);
4205 if (unlikely(tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))) {
4206 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFODROP
);
4211 /* Disable header prediction. */
4213 inet_csk_schedule_ack(sk
);
4215 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFOQUEUE
);
4216 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
4217 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4219 skb1
= skb_peek_tail(&tp
->out_of_order_queue
);
4221 /* Initial out of order segment, build 1 SACK. */
4222 if (tcp_is_sack(tp
)) {
4223 tp
->rx_opt
.num_sacks
= 1;
4224 tp
->selective_acks
[0].start_seq
= TCP_SKB_CB(skb
)->seq
;
4225 tp
->selective_acks
[0].end_seq
=
4226 TCP_SKB_CB(skb
)->end_seq
;
4228 __skb_queue_head(&tp
->out_of_order_queue
, skb
);
4232 seq
= TCP_SKB_CB(skb
)->seq
;
4233 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4235 if (seq
== TCP_SKB_CB(skb1
)->end_seq
) {
4238 if (!tcp_try_coalesce(sk
, skb1
, skb
, &fragstolen
)) {
4239 __skb_queue_after(&tp
->out_of_order_queue
, skb1
, skb
);
4241 kfree_skb_partial(skb
, fragstolen
);
4245 if (!tp
->rx_opt
.num_sacks
||
4246 tp
->selective_acks
[0].end_seq
!= seq
)
4249 /* Common case: data arrive in order after hole. */
4250 tp
->selective_acks
[0].end_seq
= end_seq
;
4254 /* Find place to insert this segment. */
4256 if (!after(TCP_SKB_CB(skb1
)->seq
, seq
))
4258 if (skb_queue_is_first(&tp
->out_of_order_queue
, skb1
)) {
4262 skb1
= skb_queue_prev(&tp
->out_of_order_queue
, skb1
);
4265 /* Do skb overlap to previous one? */
4266 if (skb1
&& before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4267 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4268 /* All the bits are present. Drop. */
4269 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFOMERGE
);
4272 tcp_dsack_set(sk
, seq
, end_seq
);
4275 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
4276 /* Partial overlap. */
4277 tcp_dsack_set(sk
, seq
,
4278 TCP_SKB_CB(skb1
)->end_seq
);
4280 if (skb_queue_is_first(&tp
->out_of_order_queue
,
4284 skb1
= skb_queue_prev(
4285 &tp
->out_of_order_queue
,
4290 __skb_queue_head(&tp
->out_of_order_queue
, skb
);
4292 __skb_queue_after(&tp
->out_of_order_queue
, skb1
, skb
);
4294 /* And clean segments covered by new one as whole. */
4295 while (!skb_queue_is_last(&tp
->out_of_order_queue
, skb
)) {
4296 skb1
= skb_queue_next(&tp
->out_of_order_queue
, skb
);
4298 if (!after(end_seq
, TCP_SKB_CB(skb1
)->seq
))
4300 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4301 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4305 __skb_unlink(skb1
, &tp
->out_of_order_queue
);
4306 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4307 TCP_SKB_CB(skb1
)->end_seq
);
4308 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFOMERGE
);
4313 if (tcp_is_sack(tp
))
4314 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
4317 skb_set_owner_r(skb
, sk
);
4320 static int __must_check
tcp_queue_rcv(struct sock
*sk
, struct sk_buff
*skb
, int hdrlen
,
4324 struct sk_buff
*tail
= skb_peek_tail(&sk
->sk_receive_queue
);
4326 __skb_pull(skb
, hdrlen
);
4328 tcp_try_coalesce(sk
, tail
, skb
, fragstolen
)) ? 1 : 0;
4329 tcp_sk(sk
)->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4331 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4332 skb_set_owner_r(skb
, sk
);
4337 int tcp_send_rcvq(struct sock
*sk
, struct msghdr
*msg
, size_t size
)
4339 struct sk_buff
*skb
= NULL
;
4346 skb
= alloc_skb(size
+ sizeof(*th
), sk
->sk_allocation
);
4350 if (tcp_try_rmem_schedule(sk
, skb
, size
+ sizeof(*th
)))
4353 th
= (struct tcphdr
*)skb_put(skb
, sizeof(*th
));
4354 skb_reset_transport_header(skb
);
4355 memset(th
, 0, sizeof(*th
));
4357 if (memcpy_fromiovec(skb_put(skb
, size
), msg
->msg_iov
, size
))
4360 TCP_SKB_CB(skb
)->seq
= tcp_sk(sk
)->rcv_nxt
;
4361 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ size
;
4362 TCP_SKB_CB(skb
)->ack_seq
= tcp_sk(sk
)->snd_una
- 1;
4364 if (tcp_queue_rcv(sk
, skb
, sizeof(*th
), &fragstolen
)) {
4365 WARN_ON_ONCE(fragstolen
); /* should not happen */
4376 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
4378 const struct tcphdr
*th
= tcp_hdr(skb
);
4379 struct tcp_sock
*tp
= tcp_sk(sk
);
4381 bool fragstolen
= false;
4383 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
)
4387 __skb_pull(skb
, th
->doff
* 4);
4389 TCP_ECN_accept_cwr(tp
, skb
);
4391 tp
->rx_opt
.dsack
= 0;
4393 /* Queue data for delivery to the user.
4394 * Packets in sequence go to the receive queue.
4395 * Out of sequence packets to the out_of_order_queue.
4397 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
4398 if (tcp_receive_window(tp
) == 0)
4401 /* Ok. In sequence. In window. */
4402 if (tp
->ucopy
.task
== current
&&
4403 tp
->copied_seq
== tp
->rcv_nxt
&& tp
->ucopy
.len
&&
4404 sock_owned_by_user(sk
) && !tp
->urg_data
) {
4405 int chunk
= min_t(unsigned int, skb
->len
,
4408 __set_current_state(TASK_RUNNING
);
4411 if (!skb_copy_datagram_iovec(skb
, 0, tp
->ucopy
.iov
, chunk
)) {
4412 tp
->ucopy
.len
-= chunk
;
4413 tp
->copied_seq
+= chunk
;
4414 eaten
= (chunk
== skb
->len
);
4415 tcp_rcv_space_adjust(sk
);
4423 tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))
4426 eaten
= tcp_queue_rcv(sk
, skb
, 0, &fragstolen
);
4428 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4430 tcp_event_data_recv(sk
, skb
);
4434 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4437 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4438 * gap in queue is filled.
4440 if (skb_queue_empty(&tp
->out_of_order_queue
))
4441 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
4444 if (tp
->rx_opt
.num_sacks
)
4445 tcp_sack_remove(tp
);
4447 tcp_fast_path_check(sk
);
4450 kfree_skb_partial(skb
, fragstolen
);
4451 if (!sock_flag(sk
, SOCK_DEAD
))
4452 sk
->sk_data_ready(sk
, 0);
4456 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4457 /* A retransmit, 2nd most common case. Force an immediate ack. */
4458 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4459 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4462 tcp_enter_quickack_mode(sk
);
4463 inet_csk_schedule_ack(sk
);
4469 /* Out of window. F.e. zero window probe. */
4470 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
4473 tcp_enter_quickack_mode(sk
);
4475 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4476 /* Partial packet, seq < rcv_next < end_seq */
4477 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
4478 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4479 TCP_SKB_CB(skb
)->end_seq
);
4481 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
4483 /* If window is closed, drop tail of packet. But after
4484 * remembering D-SACK for its head made in previous line.
4486 if (!tcp_receive_window(tp
))
4491 tcp_data_queue_ofo(sk
, skb
);
4494 static struct sk_buff
*tcp_collapse_one(struct sock
*sk
, struct sk_buff
*skb
,
4495 struct sk_buff_head
*list
)
4497 struct sk_buff
*next
= NULL
;
4499 if (!skb_queue_is_last(list
, skb
))
4500 next
= skb_queue_next(list
, skb
);
4502 __skb_unlink(skb
, list
);
4504 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRCVCOLLAPSED
);
4509 /* Collapse contiguous sequence of skbs head..tail with
4510 * sequence numbers start..end.
4512 * If tail is NULL, this means until the end of the list.
4514 * Segments with FIN/SYN are not collapsed (only because this
4518 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
,
4519 struct sk_buff
*head
, struct sk_buff
*tail
,
4522 struct sk_buff
*skb
, *n
;
4525 /* First, check that queue is collapsible and find
4526 * the point where collapsing can be useful. */
4530 skb_queue_walk_from_safe(list
, skb
, n
) {
4533 /* No new bits? It is possible on ofo queue. */
4534 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4535 skb
= tcp_collapse_one(sk
, skb
, list
);
4541 /* The first skb to collapse is:
4543 * - bloated or contains data before "start" or
4544 * overlaps to the next one.
4546 if (!tcp_hdr(skb
)->syn
&& !tcp_hdr(skb
)->fin
&&
4547 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
4548 before(TCP_SKB_CB(skb
)->seq
, start
))) {
4549 end_of_skbs
= false;
4553 if (!skb_queue_is_last(list
, skb
)) {
4554 struct sk_buff
*next
= skb_queue_next(list
, skb
);
4556 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(next
)->seq
) {
4557 end_of_skbs
= false;
4562 /* Decided to skip this, advance start seq. */
4563 start
= TCP_SKB_CB(skb
)->end_seq
;
4565 if (end_of_skbs
|| tcp_hdr(skb
)->syn
|| tcp_hdr(skb
)->fin
)
4568 while (before(start
, end
)) {
4569 struct sk_buff
*nskb
;
4570 unsigned int header
= skb_headroom(skb
);
4571 int copy
= SKB_MAX_ORDER(header
, 0);
4573 /* Too big header? This can happen with IPv6. */
4576 if (end
- start
< copy
)
4578 nskb
= alloc_skb(copy
+ header
, GFP_ATOMIC
);
4582 skb_set_mac_header(nskb
, skb_mac_header(skb
) - skb
->head
);
4583 skb_set_network_header(nskb
, (skb_network_header(skb
) -
4585 skb_set_transport_header(nskb
, (skb_transport_header(skb
) -
4587 skb_reserve(nskb
, header
);
4588 memcpy(nskb
->head
, skb
->head
, header
);
4589 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
4590 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
4591 __skb_queue_before(list
, skb
, nskb
);
4592 skb_set_owner_r(nskb
, sk
);
4594 /* Copy data, releasing collapsed skbs. */
4596 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
4597 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
4601 size
= min(copy
, size
);
4602 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
4604 TCP_SKB_CB(nskb
)->end_seq
+= size
;
4608 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4609 skb
= tcp_collapse_one(sk
, skb
, list
);
4612 tcp_hdr(skb
)->syn
||
4620 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4621 * and tcp_collapse() them until all the queue is collapsed.
4623 static void tcp_collapse_ofo_queue(struct sock
*sk
)
4625 struct tcp_sock
*tp
= tcp_sk(sk
);
4626 struct sk_buff
*skb
= skb_peek(&tp
->out_of_order_queue
);
4627 struct sk_buff
*head
;
4633 start
= TCP_SKB_CB(skb
)->seq
;
4634 end
= TCP_SKB_CB(skb
)->end_seq
;
4638 struct sk_buff
*next
= NULL
;
4640 if (!skb_queue_is_last(&tp
->out_of_order_queue
, skb
))
4641 next
= skb_queue_next(&tp
->out_of_order_queue
, skb
);
4644 /* Segment is terminated when we see gap or when
4645 * we are at the end of all the queue. */
4647 after(TCP_SKB_CB(skb
)->seq
, end
) ||
4648 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
4649 tcp_collapse(sk
, &tp
->out_of_order_queue
,
4650 head
, skb
, start
, end
);
4654 /* Start new segment */
4655 start
= TCP_SKB_CB(skb
)->seq
;
4656 end
= TCP_SKB_CB(skb
)->end_seq
;
4658 if (before(TCP_SKB_CB(skb
)->seq
, start
))
4659 start
= TCP_SKB_CB(skb
)->seq
;
4660 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
4661 end
= TCP_SKB_CB(skb
)->end_seq
;
4667 * Purge the out-of-order queue.
4668 * Return true if queue was pruned.
4670 static bool tcp_prune_ofo_queue(struct sock
*sk
)
4672 struct tcp_sock
*tp
= tcp_sk(sk
);
4675 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4676 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_OFOPRUNED
);
4677 __skb_queue_purge(&tp
->out_of_order_queue
);
4679 /* Reset SACK state. A conforming SACK implementation will
4680 * do the same at a timeout based retransmit. When a connection
4681 * is in a sad state like this, we care only about integrity
4682 * of the connection not performance.
4684 if (tp
->rx_opt
.sack_ok
)
4685 tcp_sack_reset(&tp
->rx_opt
);
4692 /* Reduce allocated memory if we can, trying to get
4693 * the socket within its memory limits again.
4695 * Return less than zero if we should start dropping frames
4696 * until the socket owning process reads some of the data
4697 * to stabilize the situation.
4699 static int tcp_prune_queue(struct sock
*sk
)
4701 struct tcp_sock
*tp
= tcp_sk(sk
);
4703 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
4705 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PRUNECALLED
);
4707 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
4708 tcp_clamp_window(sk
);
4709 else if (sk_under_memory_pressure(sk
))
4710 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
4712 tcp_collapse_ofo_queue(sk
);
4713 if (!skb_queue_empty(&sk
->sk_receive_queue
))
4714 tcp_collapse(sk
, &sk
->sk_receive_queue
,
4715 skb_peek(&sk
->sk_receive_queue
),
4717 tp
->copied_seq
, tp
->rcv_nxt
);
4720 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4723 /* Collapsing did not help, destructive actions follow.
4724 * This must not ever occur. */
4726 tcp_prune_ofo_queue(sk
);
4728 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4731 /* If we are really being abused, tell the caller to silently
4732 * drop receive data on the floor. It will get retransmitted
4733 * and hopefully then we'll have sufficient space.
4735 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_RCVPRUNED
);
4737 /* Massive buffer overcommit. */
4742 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4743 * As additional protections, we do not touch cwnd in retransmission phases,
4744 * and if application hit its sndbuf limit recently.
4746 void tcp_cwnd_application_limited(struct sock
*sk
)
4748 struct tcp_sock
*tp
= tcp_sk(sk
);
4750 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
4751 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
4752 /* Limited by application or receiver window. */
4753 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
4754 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
4755 if (win_used
< tp
->snd_cwnd
) {
4756 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
4757 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
4759 tp
->snd_cwnd_used
= 0;
4761 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4764 static bool tcp_should_expand_sndbuf(const struct sock
*sk
)
4766 const struct tcp_sock
*tp
= tcp_sk(sk
);
4768 /* If the user specified a specific send buffer setting, do
4771 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
4774 /* If we are under global TCP memory pressure, do not expand. */
4775 if (sk_under_memory_pressure(sk
))
4778 /* If we are under soft global TCP memory pressure, do not expand. */
4779 if (sk_memory_allocated(sk
) >= sk_prot_mem_limits(sk
, 0))
4782 /* If we filled the congestion window, do not expand. */
4783 if (tp
->packets_out
>= tp
->snd_cwnd
)
4789 /* When incoming ACK allowed to free some skb from write_queue,
4790 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4791 * on the exit from tcp input handler.
4793 * PROBLEM: sndbuf expansion does not work well with largesend.
4795 static void tcp_new_space(struct sock
*sk
)
4797 struct tcp_sock
*tp
= tcp_sk(sk
);
4799 if (tcp_should_expand_sndbuf(sk
)) {
4800 int sndmem
= SKB_TRUESIZE(max_t(u32
,
4801 tp
->rx_opt
.mss_clamp
,
4804 int demanded
= max_t(unsigned int, tp
->snd_cwnd
,
4805 tp
->reordering
+ 1);
4806 sndmem
*= 2 * demanded
;
4807 if (sndmem
> sk
->sk_sndbuf
)
4808 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
4809 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4812 sk
->sk_write_space(sk
);
4815 static void tcp_check_space(struct sock
*sk
)
4817 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
4818 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
4819 if (sk
->sk_socket
&&
4820 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
4825 static inline void tcp_data_snd_check(struct sock
*sk
)
4827 tcp_push_pending_frames(sk
);
4828 tcp_check_space(sk
);
4832 * Check if sending an ack is needed.
4834 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
4836 struct tcp_sock
*tp
= tcp_sk(sk
);
4838 /* More than one full frame received... */
4839 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
&&
4840 /* ... and right edge of window advances far enough.
4841 * (tcp_recvmsg() will send ACK otherwise). Or...
4843 __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
4844 /* We ACK each frame or... */
4845 tcp_in_quickack_mode(sk
) ||
4846 /* We have out of order data. */
4847 (ofo_possible
&& skb_peek(&tp
->out_of_order_queue
))) {
4848 /* Then ack it now */
4851 /* Else, send delayed ack. */
4852 tcp_send_delayed_ack(sk
);
4856 static inline void tcp_ack_snd_check(struct sock
*sk
)
4858 if (!inet_csk_ack_scheduled(sk
)) {
4859 /* We sent a data segment already. */
4862 __tcp_ack_snd_check(sk
, 1);
4866 * This routine is only called when we have urgent data
4867 * signaled. Its the 'slow' part of tcp_urg. It could be
4868 * moved inline now as tcp_urg is only called from one
4869 * place. We handle URGent data wrong. We have to - as
4870 * BSD still doesn't use the correction from RFC961.
4871 * For 1003.1g we should support a new option TCP_STDURG to permit
4872 * either form (or just set the sysctl tcp_stdurg).
4875 static void tcp_check_urg(struct sock
*sk
, const struct tcphdr
*th
)
4877 struct tcp_sock
*tp
= tcp_sk(sk
);
4878 u32 ptr
= ntohs(th
->urg_ptr
);
4880 if (ptr
&& !sysctl_tcp_stdurg
)
4882 ptr
+= ntohl(th
->seq
);
4884 /* Ignore urgent data that we've already seen and read. */
4885 if (after(tp
->copied_seq
, ptr
))
4888 /* Do not replay urg ptr.
4890 * NOTE: interesting situation not covered by specs.
4891 * Misbehaving sender may send urg ptr, pointing to segment,
4892 * which we already have in ofo queue. We are not able to fetch
4893 * such data and will stay in TCP_URG_NOTYET until will be eaten
4894 * by recvmsg(). Seems, we are not obliged to handle such wicked
4895 * situations. But it is worth to think about possibility of some
4896 * DoSes using some hypothetical application level deadlock.
4898 if (before(ptr
, tp
->rcv_nxt
))
4901 /* Do we already have a newer (or duplicate) urgent pointer? */
4902 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
4905 /* Tell the world about our new urgent pointer. */
4908 /* We may be adding urgent data when the last byte read was
4909 * urgent. To do this requires some care. We cannot just ignore
4910 * tp->copied_seq since we would read the last urgent byte again
4911 * as data, nor can we alter copied_seq until this data arrives
4912 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4914 * NOTE. Double Dutch. Rendering to plain English: author of comment
4915 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4916 * and expect that both A and B disappear from stream. This is _wrong_.
4917 * Though this happens in BSD with high probability, this is occasional.
4918 * Any application relying on this is buggy. Note also, that fix "works"
4919 * only in this artificial test. Insert some normal data between A and B and we will
4920 * decline of BSD again. Verdict: it is better to remove to trap
4923 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
4924 !sock_flag(sk
, SOCK_URGINLINE
) && tp
->copied_seq
!= tp
->rcv_nxt
) {
4925 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
4927 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4928 __skb_unlink(skb
, &sk
->sk_receive_queue
);
4933 tp
->urg_data
= TCP_URG_NOTYET
;
4936 /* Disable header prediction. */
4940 /* This is the 'fast' part of urgent handling. */
4941 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, const struct tcphdr
*th
)
4943 struct tcp_sock
*tp
= tcp_sk(sk
);
4945 /* Check if we get a new urgent pointer - normally not. */
4947 tcp_check_urg(sk
, th
);
4949 /* Do we wait for any urgent data? - normally not... */
4950 if (tp
->urg_data
== TCP_URG_NOTYET
) {
4951 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
4954 /* Is the urgent pointer pointing into this packet? */
4955 if (ptr
< skb
->len
) {
4957 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
4959 tp
->urg_data
= TCP_URG_VALID
| tmp
;
4960 if (!sock_flag(sk
, SOCK_DEAD
))
4961 sk
->sk_data_ready(sk
, 0);
4966 static int tcp_copy_to_iovec(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
4968 struct tcp_sock
*tp
= tcp_sk(sk
);
4969 int chunk
= skb
->len
- hlen
;
4973 if (skb_csum_unnecessary(skb
))
4974 err
= skb_copy_datagram_iovec(skb
, hlen
, tp
->ucopy
.iov
, chunk
);
4976 err
= skb_copy_and_csum_datagram_iovec(skb
, hlen
,
4980 tp
->ucopy
.len
-= chunk
;
4981 tp
->copied_seq
+= chunk
;
4982 tcp_rcv_space_adjust(sk
);
4989 static __sum16
__tcp_checksum_complete_user(struct sock
*sk
,
4990 struct sk_buff
*skb
)
4994 if (sock_owned_by_user(sk
)) {
4996 result
= __tcp_checksum_complete(skb
);
4999 result
= __tcp_checksum_complete(skb
);
5004 static inline bool tcp_checksum_complete_user(struct sock
*sk
,
5005 struct sk_buff
*skb
)
5007 return !skb_csum_unnecessary(skb
) &&
5008 __tcp_checksum_complete_user(sk
, skb
);
5011 #ifdef CONFIG_NET_DMA
5012 static bool tcp_dma_try_early_copy(struct sock
*sk
, struct sk_buff
*skb
,
5015 struct tcp_sock
*tp
= tcp_sk(sk
);
5016 int chunk
= skb
->len
- hlen
;
5018 bool copied_early
= false;
5020 if (tp
->ucopy
.wakeup
)
5023 if (!tp
->ucopy
.dma_chan
&& tp
->ucopy
.pinned_list
)
5024 tp
->ucopy
.dma_chan
= net_dma_find_channel();
5026 if (tp
->ucopy
.dma_chan
&& skb_csum_unnecessary(skb
)) {
5028 dma_cookie
= dma_skb_copy_datagram_iovec(tp
->ucopy
.dma_chan
,
5030 tp
->ucopy
.iov
, chunk
,
5031 tp
->ucopy
.pinned_list
);
5036 tp
->ucopy
.dma_cookie
= dma_cookie
;
5037 copied_early
= true;
5039 tp
->ucopy
.len
-= chunk
;
5040 tp
->copied_seq
+= chunk
;
5041 tcp_rcv_space_adjust(sk
);
5043 if ((tp
->ucopy
.len
== 0) ||
5044 (tcp_flag_word(tcp_hdr(skb
)) & TCP_FLAG_PSH
) ||
5045 (atomic_read(&sk
->sk_rmem_alloc
) > (sk
->sk_rcvbuf
>> 1))) {
5046 tp
->ucopy
.wakeup
= 1;
5047 sk
->sk_data_ready(sk
, 0);
5049 } else if (chunk
> 0) {
5050 tp
->ucopy
.wakeup
= 1;
5051 sk
->sk_data_ready(sk
, 0);
5054 return copied_early
;
5056 #endif /* CONFIG_NET_DMA */
5058 /* Does PAWS and seqno based validation of an incoming segment, flags will
5059 * play significant role here.
5061 static bool tcp_validate_incoming(struct sock
*sk
, struct sk_buff
*skb
,
5062 const struct tcphdr
*th
, int syn_inerr
)
5064 struct tcp_sock
*tp
= tcp_sk(sk
);
5066 /* RFC1323: H1. Apply PAWS check first. */
5067 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
5068 tcp_paws_discard(sk
, skb
)) {
5070 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSESTABREJECTED
);
5071 tcp_send_dupack(sk
, skb
);
5074 /* Reset is accepted even if it did not pass PAWS. */
5077 /* Step 1: check sequence number */
5078 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5079 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5080 * (RST) segments are validated by checking their SEQ-fields."
5081 * And page 69: "If an incoming segment is not acceptable,
5082 * an acknowledgment should be sent in reply (unless the RST
5083 * bit is set, if so drop the segment and return)".
5088 tcp_send_dupack(sk
, skb
);
5093 /* Step 2: check RST bit */
5096 * If sequence number exactly matches RCV.NXT, then
5097 * RESET the connection
5099 * Send a challenge ACK
5101 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
)
5104 tcp_send_challenge_ack(sk
);
5108 /* step 3: check security and precedence [ignored] */
5110 /* step 4: Check for a SYN
5111 * RFC 5691 4.2 : Send a challenge ack
5116 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5117 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSYNCHALLENGE
);
5118 tcp_send_challenge_ack(sk
);
5130 * TCP receive function for the ESTABLISHED state.
5132 * It is split into a fast path and a slow path. The fast path is
5134 * - A zero window was announced from us - zero window probing
5135 * is only handled properly in the slow path.
5136 * - Out of order segments arrived.
5137 * - Urgent data is expected.
5138 * - There is no buffer space left
5139 * - Unexpected TCP flags/window values/header lengths are received
5140 * (detected by checking the TCP header against pred_flags)
5141 * - Data is sent in both directions. Fast path only supports pure senders
5142 * or pure receivers (this means either the sequence number or the ack
5143 * value must stay constant)
5144 * - Unexpected TCP option.
5146 * When these conditions are not satisfied it drops into a standard
5147 * receive procedure patterned after RFC793 to handle all cases.
5148 * The first three cases are guaranteed by proper pred_flags setting,
5149 * the rest is checked inline. Fast processing is turned on in
5150 * tcp_data_queue when everything is OK.
5152 int tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
5153 const struct tcphdr
*th
, unsigned int len
)
5155 struct tcp_sock
*tp
= tcp_sk(sk
);
5157 if (unlikely(sk
->sk_rx_dst
== NULL
))
5158 inet_csk(sk
)->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
5160 * Header prediction.
5161 * The code loosely follows the one in the famous
5162 * "30 instruction TCP receive" Van Jacobson mail.
5164 * Van's trick is to deposit buffers into socket queue
5165 * on a device interrupt, to call tcp_recv function
5166 * on the receive process context and checksum and copy
5167 * the buffer to user space. smart...
5169 * Our current scheme is not silly either but we take the
5170 * extra cost of the net_bh soft interrupt processing...
5171 * We do checksum and copy also but from device to kernel.
5174 tp
->rx_opt
.saw_tstamp
= 0;
5176 /* pred_flags is 0xS?10 << 16 + snd_wnd
5177 * if header_prediction is to be made
5178 * 'S' will always be tp->tcp_header_len >> 2
5179 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5180 * turn it off (when there are holes in the receive
5181 * space for instance)
5182 * PSH flag is ignored.
5185 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
5186 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
&&
5187 !after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
)) {
5188 int tcp_header_len
= tp
->tcp_header_len
;
5190 /* Timestamp header prediction: tcp_header_len
5191 * is automatically equal to th->doff*4 due to pred_flags
5195 /* Check timestamp */
5196 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
5197 /* No? Slow path! */
5198 if (!tcp_parse_aligned_timestamp(tp
, th
))
5201 /* If PAWS failed, check it more carefully in slow path */
5202 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
5205 /* DO NOT update ts_recent here, if checksum fails
5206 * and timestamp was corrupted part, it will result
5207 * in a hung connection since we will drop all
5208 * future packets due to the PAWS test.
5212 if (len
<= tcp_header_len
) {
5213 /* Bulk data transfer: sender */
5214 if (len
== tcp_header_len
) {
5215 /* Predicted packet is in window by definition.
5216 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5217 * Hence, check seq<=rcv_wup reduces to:
5219 if (tcp_header_len
==
5220 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5221 tp
->rcv_nxt
== tp
->rcv_wup
)
5222 tcp_store_ts_recent(tp
);
5224 /* We know that such packets are checksummed
5227 tcp_ack(sk
, skb
, 0);
5229 tcp_data_snd_check(sk
);
5231 } else { /* Header too small */
5232 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5237 int copied_early
= 0;
5238 bool fragstolen
= false;
5240 if (tp
->copied_seq
== tp
->rcv_nxt
&&
5241 len
- tcp_header_len
<= tp
->ucopy
.len
) {
5242 #ifdef CONFIG_NET_DMA
5243 if (tp
->ucopy
.task
== current
&&
5244 sock_owned_by_user(sk
) &&
5245 tcp_dma_try_early_copy(sk
, skb
, tcp_header_len
)) {
5250 if (tp
->ucopy
.task
== current
&&
5251 sock_owned_by_user(sk
) && !copied_early
) {
5252 __set_current_state(TASK_RUNNING
);
5254 if (!tcp_copy_to_iovec(sk
, skb
, tcp_header_len
))
5258 /* Predicted packet is in window by definition.
5259 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5260 * Hence, check seq<=rcv_wup reduces to:
5262 if (tcp_header_len
==
5263 (sizeof(struct tcphdr
) +
5264 TCPOLEN_TSTAMP_ALIGNED
) &&
5265 tp
->rcv_nxt
== tp
->rcv_wup
)
5266 tcp_store_ts_recent(tp
);
5268 tcp_rcv_rtt_measure_ts(sk
, skb
);
5270 __skb_pull(skb
, tcp_header_len
);
5271 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
5272 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITSTOUSER
);
5275 tcp_cleanup_rbuf(sk
, skb
->len
);
5278 if (tcp_checksum_complete_user(sk
, skb
))
5281 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
5284 /* Predicted packet is in window by definition.
5285 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5286 * Hence, check seq<=rcv_wup reduces to:
5288 if (tcp_header_len
==
5289 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5290 tp
->rcv_nxt
== tp
->rcv_wup
)
5291 tcp_store_ts_recent(tp
);
5293 tcp_rcv_rtt_measure_ts(sk
, skb
);
5295 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITS
);
5297 /* Bulk data transfer: receiver */
5298 eaten
= tcp_queue_rcv(sk
, skb
, tcp_header_len
,
5302 tcp_event_data_recv(sk
, skb
);
5304 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
5305 /* Well, only one small jumplet in fast path... */
5306 tcp_ack(sk
, skb
, FLAG_DATA
);
5307 tcp_data_snd_check(sk
);
5308 if (!inet_csk_ack_scheduled(sk
))
5312 if (!copied_early
|| tp
->rcv_nxt
!= tp
->rcv_wup
)
5313 __tcp_ack_snd_check(sk
, 0);
5315 #ifdef CONFIG_NET_DMA
5317 __skb_queue_tail(&sk
->sk_async_wait_queue
, skb
);
5321 kfree_skb_partial(skb
, fragstolen
);
5322 sk
->sk_data_ready(sk
, 0);
5328 if (len
< (th
->doff
<< 2) || tcp_checksum_complete_user(sk
, skb
))
5331 if (!th
->ack
&& !th
->rst
)
5335 * Standard slow path.
5338 if (!tcp_validate_incoming(sk
, skb
, th
, 1))
5342 if (tcp_ack(sk
, skb
, FLAG_SLOWPATH
| FLAG_UPDATE_TS_RECENT
) < 0)
5345 tcp_rcv_rtt_measure_ts(sk
, skb
);
5347 /* Process urgent data. */
5348 tcp_urg(sk
, skb
, th
);
5350 /* step 7: process the segment text */
5351 tcp_data_queue(sk
, skb
);
5353 tcp_data_snd_check(sk
);
5354 tcp_ack_snd_check(sk
);
5358 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_CSUMERRORS
);
5359 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5365 EXPORT_SYMBOL(tcp_rcv_established
);
5367 void tcp_finish_connect(struct sock
*sk
, struct sk_buff
*skb
)
5369 struct tcp_sock
*tp
= tcp_sk(sk
);
5370 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5372 tcp_set_state(sk
, TCP_ESTABLISHED
);
5375 icsk
->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
5376 security_inet_conn_established(sk
, skb
);
5379 /* Make sure socket is routed, for correct metrics. */
5380 icsk
->icsk_af_ops
->rebuild_header(sk
);
5382 tcp_init_metrics(sk
);
5384 tcp_init_congestion_control(sk
);
5386 /* Prevent spurious tcp_cwnd_restart() on first data
5389 tp
->lsndtime
= tcp_time_stamp
;
5391 tcp_init_buffer_space(sk
);
5393 if (sock_flag(sk
, SOCK_KEEPOPEN
))
5394 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
5396 if (!tp
->rx_opt
.snd_wscale
)
5397 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
5401 if (!sock_flag(sk
, SOCK_DEAD
)) {
5402 sk
->sk_state_change(sk
);
5403 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5407 static bool tcp_rcv_fastopen_synack(struct sock
*sk
, struct sk_buff
*synack
,
5408 struct tcp_fastopen_cookie
*cookie
)
5410 struct tcp_sock
*tp
= tcp_sk(sk
);
5411 struct sk_buff
*data
= tp
->syn_data
? tcp_write_queue_head(sk
) : NULL
;
5412 u16 mss
= tp
->rx_opt
.mss_clamp
;
5415 if (mss
== tp
->rx_opt
.user_mss
) {
5416 struct tcp_options_received opt
;
5418 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5419 tcp_clear_options(&opt
);
5420 opt
.user_mss
= opt
.mss_clamp
= 0;
5421 tcp_parse_options(synack
, &opt
, 0, NULL
);
5422 mss
= opt
.mss_clamp
;
5425 if (!tp
->syn_fastopen
) /* Ignore an unsolicited cookie */
5428 /* The SYN-ACK neither has cookie nor acknowledges the data. Presumably
5429 * the remote receives only the retransmitted (regular) SYNs: either
5430 * the original SYN-data or the corresponding SYN-ACK is lost.
5432 syn_drop
= (cookie
->len
<= 0 && data
&& tp
->total_retrans
);
5434 tcp_fastopen_cache_set(sk
, mss
, cookie
, syn_drop
);
5436 if (data
) { /* Retransmit unacked data in SYN */
5437 tcp_for_write_queue_from(data
, sk
) {
5438 if (data
== tcp_send_head(sk
) ||
5439 __tcp_retransmit_skb(sk
, data
))
5445 tp
->syn_data_acked
= tp
->syn_data
;
5449 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5450 const struct tcphdr
*th
, unsigned int len
)
5452 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5453 struct tcp_sock
*tp
= tcp_sk(sk
);
5454 struct tcp_fastopen_cookie foc
= { .len
= -1 };
5455 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
5457 tcp_parse_options(skb
, &tp
->rx_opt
, 0, &foc
);
5458 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
5459 tp
->rx_opt
.rcv_tsecr
-= tp
->tsoffset
;
5463 * "If the state is SYN-SENT then
5464 * first check the ACK bit
5465 * If the ACK bit is set
5466 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5467 * a reset (unless the RST bit is set, if so drop
5468 * the segment and return)"
5470 if (!after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_una
) ||
5471 after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
))
5472 goto reset_and_undo
;
5474 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
5475 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
5477 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSACTIVEREJECTED
);
5478 goto reset_and_undo
;
5481 /* Now ACK is acceptable.
5483 * "If the RST bit is set
5484 * If the ACK was acceptable then signal the user "error:
5485 * connection reset", drop the segment, enter CLOSED state,
5486 * delete TCB, and return."
5495 * "fifth, if neither of the SYN or RST bits is set then
5496 * drop the segment and return."
5502 goto discard_and_undo
;
5505 * "If the SYN bit is on ...
5506 * are acceptable then ...
5507 * (our SYN has been ACKed), change the connection
5508 * state to ESTABLISHED..."
5511 TCP_ECN_rcv_synack(tp
, th
);
5513 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5514 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
5516 /* Ok.. it's good. Set up sequence numbers and
5517 * move to established.
5519 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5520 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5522 /* RFC1323: The window in SYN & SYN/ACK segments is
5525 tp
->snd_wnd
= ntohs(th
->window
);
5527 if (!tp
->rx_opt
.wscale_ok
) {
5528 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
5529 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
5532 if (tp
->rx_opt
.saw_tstamp
) {
5533 tp
->rx_opt
.tstamp_ok
= 1;
5534 tp
->tcp_header_len
=
5535 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5536 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5537 tcp_store_ts_recent(tp
);
5539 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5542 if (tcp_is_sack(tp
) && sysctl_tcp_fack
)
5543 tcp_enable_fack(tp
);
5546 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5547 tcp_initialize_rcv_mss(sk
);
5549 /* Remember, tcp_poll() does not lock socket!
5550 * Change state from SYN-SENT only after copied_seq
5551 * is initialized. */
5552 tp
->copied_seq
= tp
->rcv_nxt
;
5556 tcp_finish_connect(sk
, skb
);
5558 if ((tp
->syn_fastopen
|| tp
->syn_data
) &&
5559 tcp_rcv_fastopen_synack(sk
, skb
, &foc
))
5562 if (sk
->sk_write_pending
||
5563 icsk
->icsk_accept_queue
.rskq_defer_accept
||
5564 icsk
->icsk_ack
.pingpong
) {
5565 /* Save one ACK. Data will be ready after
5566 * several ticks, if write_pending is set.
5568 * It may be deleted, but with this feature tcpdumps
5569 * look so _wonderfully_ clever, that I was not able
5570 * to stand against the temptation 8) --ANK
5572 inet_csk_schedule_ack(sk
);
5573 icsk
->icsk_ack
.lrcvtime
= tcp_time_stamp
;
5574 tcp_enter_quickack_mode(sk
);
5575 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
5576 TCP_DELACK_MAX
, sysctl_tcp_rto_max
);
5587 /* No ACK in the segment */
5591 * "If the RST bit is set
5593 * Otherwise (no ACK) drop the segment and return."
5596 goto discard_and_undo
;
5600 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&&
5601 tcp_paws_reject(&tp
->rx_opt
, 0))
5602 goto discard_and_undo
;
5605 /* We see SYN without ACK. It is attempt of
5606 * simultaneous connect with crossed SYNs.
5607 * Particularly, it can be connect to self.
5609 tcp_set_state(sk
, TCP_SYN_RECV
);
5611 if (tp
->rx_opt
.saw_tstamp
) {
5612 tp
->rx_opt
.tstamp_ok
= 1;
5613 tcp_store_ts_recent(tp
);
5614 tp
->tcp_header_len
=
5615 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5617 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5620 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5621 tp
->copied_seq
= tp
->rcv_nxt
;
5622 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5624 /* RFC1323: The window in SYN & SYN/ACK segments is
5627 tp
->snd_wnd
= ntohs(th
->window
);
5628 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
5629 tp
->max_window
= tp
->snd_wnd
;
5631 TCP_ECN_rcv_syn(tp
, th
);
5634 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5635 tcp_initialize_rcv_mss(sk
);
5637 tcp_send_synack(sk
);
5639 /* Note, we could accept data and URG from this segment.
5640 * There are no obstacles to make this (except that we must
5641 * either change tcp_recvmsg() to prevent it from returning data
5642 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5644 * However, if we ignore data in ACKless segments sometimes,
5645 * we have no reasons to accept it sometimes.
5646 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5647 * is not flawless. So, discard packet for sanity.
5648 * Uncomment this return to process the data.
5655 /* "fifth, if neither of the SYN or RST bits is set then
5656 * drop the segment and return."
5660 tcp_clear_options(&tp
->rx_opt
);
5661 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5665 tcp_clear_options(&tp
->rx_opt
);
5666 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5671 * This function implements the receiving procedure of RFC 793 for
5672 * all states except ESTABLISHED and TIME_WAIT.
5673 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5674 * address independent.
5677 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5678 const struct tcphdr
*th
, unsigned int len
)
5680 struct tcp_sock
*tp
= tcp_sk(sk
);
5681 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5682 struct request_sock
*req
;
5685 tp
->rx_opt
.saw_tstamp
= 0;
5687 switch (sk
->sk_state
) {
5701 if (icsk
->icsk_af_ops
->conn_request(sk
, skb
) < 0)
5704 /* Now we have several options: In theory there is
5705 * nothing else in the frame. KA9Q has an option to
5706 * send data with the syn, BSD accepts data with the
5707 * syn up to the [to be] advertised window and
5708 * Solaris 2.1 gives you a protocol error. For now
5709 * we just ignore it, that fits the spec precisely
5710 * and avoids incompatibilities. It would be nice in
5711 * future to drop through and process the data.
5713 * Now that TTCP is starting to be used we ought to
5715 * But, this leaves one open to an easy denial of
5716 * service attack, and SYN cookies can't defend
5717 * against this problem. So, we drop the data
5718 * in the interest of security over speed unless
5719 * it's still in use.
5727 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
, len
);
5731 /* Do step6 onward by hand. */
5732 tcp_urg(sk
, skb
, th
);
5734 tcp_data_snd_check(sk
);
5738 req
= tp
->fastopen_rsk
;
5740 WARN_ON_ONCE(sk
->sk_state
!= TCP_SYN_RECV
&&
5741 sk
->sk_state
!= TCP_FIN_WAIT1
);
5743 if (tcp_check_req(sk
, skb
, req
, NULL
, true) == NULL
)
5747 if (!th
->ack
&& !th
->rst
)
5750 if (!tcp_validate_incoming(sk
, skb
, th
, 0))
5753 /* step 5: check the ACK field */
5755 int acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
|
5756 FLAG_UPDATE_TS_RECENT
) > 0;
5758 switch (sk
->sk_state
) {
5761 /* Once we leave TCP_SYN_RECV, we no longer
5762 * need req so release it.
5765 tcp_synack_rtt_meas(sk
, req
);
5766 tp
->total_retrans
= req
->num_retrans
;
5768 reqsk_fastopen_remove(sk
, req
, false);
5770 /* Make sure socket is routed, for
5773 icsk
->icsk_af_ops
->rebuild_header(sk
);
5774 tcp_init_congestion_control(sk
);
5777 tcp_init_buffer_space(sk
);
5778 tp
->copied_seq
= tp
->rcv_nxt
;
5781 tcp_set_state(sk
, TCP_ESTABLISHED
);
5782 sk
->sk_state_change(sk
);
5784 /* Note, that this wakeup is only for marginal
5785 * crossed SYN case. Passively open sockets
5786 * are not waked up, because sk->sk_sleep ==
5787 * NULL and sk->sk_socket == NULL.
5791 SOCK_WAKE_IO
, POLL_OUT
);
5793 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
5794 tp
->snd_wnd
= ntohs(th
->window
) <<
5795 tp
->rx_opt
.snd_wscale
;
5796 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5798 if (tp
->rx_opt
.tstamp_ok
)
5799 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5802 /* Re-arm the timer because data may
5803 * have been sent out. This is similar
5804 * to the regular data transmission case
5805 * when new data has just been ack'ed.
5807 * (TFO) - we could try to be more
5808 * aggressive and retranmitting any data
5809 * sooner based on when they were sent
5814 tcp_init_metrics(sk
);
5816 tcp_update_pacing_rate(sk
);
5818 /* Prevent spurious tcp_cwnd_restart() on
5819 * first data packet.
5821 tp
->lsndtime
= tcp_time_stamp
;
5823 tcp_initialize_rcv_mss(sk
);
5824 tcp_fast_path_on(tp
);
5831 /* If we enter the TCP_FIN_WAIT1 state and we are a
5832 * Fast Open socket and this is the first acceptable
5833 * ACK we have received, this would have acknowledged
5834 * our SYNACK so stop the SYNACK timer.
5837 /* Return RST if ack_seq is invalid.
5838 * Note that RFC793 only says to generate a
5839 * DUPACK for it but for TCP Fast Open it seems
5840 * better to treat this case like TCP_SYN_RECV
5845 /* We no longer need the request sock. */
5846 reqsk_fastopen_remove(sk
, req
, false);
5849 if (tp
->snd_una
== tp
->write_seq
) {
5850 struct dst_entry
*dst
;
5852 tcp_set_state(sk
, TCP_FIN_WAIT2
);
5853 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
5855 dst
= __sk_dst_get(sk
);
5859 if (!sock_flag(sk
, SOCK_DEAD
))
5860 /* Wake up lingering close() */
5861 sk
->sk_state_change(sk
);
5865 if (tp
->linger2
< 0 ||
5866 (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5867 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
))) {
5869 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
5873 tmo
= tcp_fin_time(sk
);
5874 if (tmo
> TCP_TIMEWAIT_LEN
) {
5875 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
5876 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
5877 /* Bad case. We could lose such FIN otherwise.
5878 * It is not a big problem, but it looks confusing
5879 * and not so rare event. We still can lose it now,
5880 * if it spins in bh_lock_sock(), but it is really
5883 inet_csk_reset_keepalive_timer(sk
, tmo
);
5885 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
5893 if (tp
->snd_una
== tp
->write_seq
) {
5894 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
5900 if (tp
->snd_una
== tp
->write_seq
) {
5901 tcp_update_metrics(sk
);
5909 /* step 6: check the URG bit */
5910 tcp_urg(sk
, skb
, th
);
5912 /* step 7: process the segment text */
5913 switch (sk
->sk_state
) {
5914 case TCP_CLOSE_WAIT
:
5917 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
5921 /* RFC 793 says to queue data in these states,
5922 * RFC 1122 says we MUST send a reset.
5923 * BSD 4.4 also does reset.
5925 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
5926 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5927 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
5928 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
5934 case TCP_ESTABLISHED
:
5935 tcp_data_queue(sk
, skb
);
5940 /* tcp_data could move socket to TIME-WAIT */
5941 if (sk
->sk_state
!= TCP_CLOSE
) {
5942 tcp_data_snd_check(sk
);
5943 tcp_ack_snd_check(sk
);
5952 EXPORT_SYMBOL(tcp_rcv_state_process
);