2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
23 * Pedro Roque : Fast Retransmit/Recovery.
25 * Retransmit queue handled by TCP.
26 * Better retransmit timer handling.
27 * New congestion avoidance.
31 * Eric : Fast Retransmit.
32 * Randy Scott : MSS option defines.
33 * Eric Schenk : Fixes to slow start algorithm.
34 * Eric Schenk : Yet another double ACK bug.
35 * Eric Schenk : Delayed ACK bug fixes.
36 * Eric Schenk : Floyd style fast retrans war avoidance.
37 * David S. Miller : Don't allow zero congestion window.
38 * Eric Schenk : Fix retransmitter so that it sends
39 * next packet on ack of previous packet.
40 * Andi Kleen : Moved open_request checking here
41 * and process RSTs for open_requests.
42 * Andi Kleen : Better prune_queue, and other fixes.
43 * Andrey Savochkin: Fix RTT measurements in the presence of
45 * Andrey Savochkin: Check sequence numbers correctly when
46 * removing SACKs due to in sequence incoming
48 * Andi Kleen: Make sure we never ack data there is not
49 * enough room for. Also make this condition
50 * a fatal error if it might still happen.
51 * Andi Kleen: Add tcp_measure_rcv_mss to make
52 * connections with MSS<min(MTU,ann. MSS)
53 * work without delayed acks.
54 * Andi Kleen: Process packets with PSH set in the
56 * J Hadi Salim: ECN support
59 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
60 * engine. Lots of bugs are found.
61 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
64 #define pr_fmt(fmt) "TCP: " fmt
67 #include <linux/slab.h>
68 #include <linux/module.h>
69 #include <linux/sysctl.h>
70 #include <linux/kernel.h>
73 #include <net/inet_common.h>
74 #include <linux/ipsec.h>
75 #include <asm/unaligned.h>
76 #include <net/netdma.h>
78 int sysctl_tcp_timestamps __read_mostly
= 1;
79 int sysctl_tcp_window_scaling __read_mostly
= 1;
80 int sysctl_tcp_sack __read_mostly
= 1;
81 int sysctl_tcp_fack __read_mostly
= 1;
82 int sysctl_tcp_reordering __read_mostly
= TCP_FASTRETRANS_THRESH
;
83 EXPORT_SYMBOL(sysctl_tcp_reordering
);
84 int sysctl_tcp_dsack __read_mostly
= 1;
85 int sysctl_tcp_app_win __read_mostly
= 31;
86 int sysctl_tcp_adv_win_scale __read_mostly
= 1;
87 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale
);
89 /* rfc5961 challenge ack rate limiting */
90 int sysctl_tcp_challenge_ack_limit
= 100;
92 int sysctl_tcp_stdurg __read_mostly
;
93 int sysctl_tcp_rfc1337 __read_mostly
;
94 int sysctl_tcp_max_orphans __read_mostly
= NR_FILE
;
95 int sysctl_tcp_frto __read_mostly
= 2;
96 int sysctl_tcp_frto_response __read_mostly
;
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
= 2;
103 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
104 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
105 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
106 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
107 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
108 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
109 #define FLAG_ECE 0x40 /* ECE in this ACK */
110 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
111 #define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */
112 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
113 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
114 #define FLAG_NONHEAD_RETRANS_ACKED 0x1000 /* Non-head rexmitted data was ACKed */
115 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
117 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
118 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
119 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
120 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
121 #define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED)
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
= TCP_DEFAULT_INIT_RCVWND
;
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 * TCP_DEFAULT_INIT_RCVWND
) / 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 /* Calculate rto without backoff. This is the second half of Van Jacobson's
705 * routine referred to above.
707 void tcp_set_rto(struct sock
*sk
)
709 const struct tcp_sock
*tp
= tcp_sk(sk
);
710 /* Old crap is replaced with new one. 8)
713 * 1. If rtt variance happened to be less 50msec, it is hallucination.
714 * It cannot be less due to utterly erratic ACK generation made
715 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
716 * to do with delayed acks, because at cwnd>2 true delack timeout
717 * is invisible. Actually, Linux-2.4 also generates erratic
718 * ACKs in some circumstances.
720 inet_csk(sk
)->icsk_rto
= __tcp_set_rto(tp
);
722 /* 2. Fixups made earlier cannot be right.
723 * If we do not estimate RTO correctly without them,
724 * all the algo is pure shit and should be replaced
725 * with correct one. It is exactly, which we pretend to do.
728 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
729 * guarantees that rto is higher.
734 __u32
tcp_init_cwnd(const struct tcp_sock
*tp
, const struct dst_entry
*dst
)
736 __u32 cwnd
= (dst
? dst_metric(dst
, RTAX_INITCWND
) : 0);
739 cwnd
= TCP_INIT_CWND
;
740 return min_t(__u32
, cwnd
, tp
->snd_cwnd_clamp
);
744 * Packet counting of FACK is based on in-order assumptions, therefore TCP
745 * disables it when reordering is detected
747 void tcp_disable_fack(struct tcp_sock
*tp
)
749 /* RFC3517 uses different metric in lost marker => reset on change */
751 tp
->lost_skb_hint
= NULL
;
752 tp
->rx_opt
.sack_ok
&= ~TCP_FACK_ENABLED
;
755 /* Take a notice that peer is sending D-SACKs */
756 static void tcp_dsack_seen(struct tcp_sock
*tp
)
758 tp
->rx_opt
.sack_ok
|= TCP_DSACK_SEEN
;
761 static void tcp_update_reordering(struct sock
*sk
, const int metric
,
764 struct tcp_sock
*tp
= tcp_sk(sk
);
765 if (metric
> tp
->reordering
) {
768 tp
->reordering
= min(TCP_MAX_REORDERING
, metric
);
770 /* This exciting event is worth to be remembered. 8) */
772 mib_idx
= LINUX_MIB_TCPTSREORDER
;
773 else if (tcp_is_reno(tp
))
774 mib_idx
= LINUX_MIB_TCPRENOREORDER
;
775 else if (tcp_is_fack(tp
))
776 mib_idx
= LINUX_MIB_TCPFACKREORDER
;
778 mib_idx
= LINUX_MIB_TCPSACKREORDER
;
780 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
781 #if FASTRETRANS_DEBUG > 1
782 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
783 tp
->rx_opt
.sack_ok
, inet_csk(sk
)->icsk_ca_state
,
787 tp
->undo_marker
? tp
->undo_retrans
: 0);
789 tcp_disable_fack(tp
);
793 tcp_disable_early_retrans(tp
);
796 /* This must be called before lost_out is incremented */
797 static void tcp_verify_retransmit_hint(struct tcp_sock
*tp
, struct sk_buff
*skb
)
799 if ((tp
->retransmit_skb_hint
== NULL
) ||
800 before(TCP_SKB_CB(skb
)->seq
,
801 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
802 tp
->retransmit_skb_hint
= skb
;
805 after(TCP_SKB_CB(skb
)->end_seq
, tp
->retransmit_high
))
806 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
809 static void tcp_skb_mark_lost(struct tcp_sock
*tp
, struct sk_buff
*skb
)
811 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
812 tcp_verify_retransmit_hint(tp
, skb
);
814 tp
->lost_out
+= tcp_skb_pcount(skb
);
815 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
819 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock
*tp
,
822 tcp_verify_retransmit_hint(tp
, skb
);
824 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
825 tp
->lost_out
+= tcp_skb_pcount(skb
);
826 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
830 /* This procedure tags the retransmission queue when SACKs arrive.
832 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
833 * Packets in queue with these bits set are counted in variables
834 * sacked_out, retrans_out and lost_out, correspondingly.
836 * Valid combinations are:
837 * Tag InFlight Description
838 * 0 1 - orig segment is in flight.
839 * S 0 - nothing flies, orig reached receiver.
840 * L 0 - nothing flies, orig lost by net.
841 * R 2 - both orig and retransmit are in flight.
842 * L|R 1 - orig is lost, retransmit is in flight.
843 * S|R 1 - orig reached receiver, retrans is still in flight.
844 * (L|S|R is logically valid, it could occur when L|R is sacked,
845 * but it is equivalent to plain S and code short-curcuits it to S.
846 * L|S is logically invalid, it would mean -1 packet in flight 8))
848 * These 6 states form finite state machine, controlled by the following events:
849 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
850 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
851 * 3. Loss detection event of two flavors:
852 * A. Scoreboard estimator decided the packet is lost.
853 * A'. Reno "three dupacks" marks head of queue lost.
854 * A''. Its FACK modification, head until snd.fack is lost.
855 * B. SACK arrives sacking SND.NXT at the moment, when the
856 * segment was retransmitted.
857 * 4. D-SACK added new rule: D-SACK changes any tag to S.
859 * It is pleasant to note, that state diagram turns out to be commutative,
860 * so that we are allowed not to be bothered by order of our actions,
861 * when multiple events arrive simultaneously. (see the function below).
863 * Reordering detection.
864 * --------------------
865 * Reordering metric is maximal distance, which a packet can be displaced
866 * in packet stream. With SACKs we can estimate it:
868 * 1. SACK fills old hole and the corresponding segment was not
869 * ever retransmitted -> reordering. Alas, we cannot use it
870 * when segment was retransmitted.
871 * 2. The last flaw is solved with D-SACK. D-SACK arrives
872 * for retransmitted and already SACKed segment -> reordering..
873 * Both of these heuristics are not used in Loss state, when we cannot
874 * account for retransmits accurately.
876 * SACK block validation.
877 * ----------------------
879 * SACK block range validation checks that the received SACK block fits to
880 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
881 * Note that SND.UNA is not included to the range though being valid because
882 * it means that the receiver is rather inconsistent with itself reporting
883 * SACK reneging when it should advance SND.UNA. Such SACK block this is
884 * perfectly valid, however, in light of RFC2018 which explicitly states
885 * that "SACK block MUST reflect the newest segment. Even if the newest
886 * segment is going to be discarded ...", not that it looks very clever
887 * in case of head skb. Due to potentional receiver driven attacks, we
888 * choose to avoid immediate execution of a walk in write queue due to
889 * reneging and defer head skb's loss recovery to standard loss recovery
890 * procedure that will eventually trigger (nothing forbids us doing this).
892 * Implements also blockage to start_seq wrap-around. Problem lies in the
893 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
894 * there's no guarantee that it will be before snd_nxt (n). The problem
895 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
898 * <- outs wnd -> <- wrapzone ->
899 * u e n u_w e_w s n_w
901 * |<------------+------+----- TCP seqno space --------------+---------->|
902 * ...-- <2^31 ->| |<--------...
903 * ...---- >2^31 ------>| |<--------...
905 * Current code wouldn't be vulnerable but it's better still to discard such
906 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
907 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
908 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
909 * equal to the ideal case (infinite seqno space without wrap caused issues).
911 * With D-SACK the lower bound is extended to cover sequence space below
912 * SND.UNA down to undo_marker, which is the last point of interest. Yet
913 * again, D-SACK block must not to go across snd_una (for the same reason as
914 * for the normal SACK blocks, explained above). But there all simplicity
915 * ends, TCP might receive valid D-SACKs below that. As long as they reside
916 * fully below undo_marker they do not affect behavior in anyway and can
917 * therefore be safely ignored. In rare cases (which are more or less
918 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
919 * fragmentation and packet reordering past skb's retransmission. To consider
920 * them correctly, the acceptable range must be extended even more though
921 * the exact amount is rather hard to quantify. However, tp->max_window can
922 * be used as an exaggerated estimate.
924 static bool tcp_is_sackblock_valid(struct tcp_sock
*tp
, bool is_dsack
,
925 u32 start_seq
, u32 end_seq
)
927 /* Too far in future, or reversed (interpretation is ambiguous) */
928 if (after(end_seq
, tp
->snd_nxt
) || !before(start_seq
, end_seq
))
931 /* Nasty start_seq wrap-around check (see comments above) */
932 if (!before(start_seq
, tp
->snd_nxt
))
935 /* In outstanding window? ...This is valid exit for D-SACKs too.
936 * start_seq == snd_una is non-sensical (see comments above)
938 if (after(start_seq
, tp
->snd_una
))
941 if (!is_dsack
|| !tp
->undo_marker
)
944 /* ...Then it's D-SACK, and must reside below snd_una completely */
945 if (after(end_seq
, tp
->snd_una
))
948 if (!before(start_seq
, tp
->undo_marker
))
952 if (!after(end_seq
, tp
->undo_marker
))
955 /* Undo_marker boundary crossing (overestimates a lot). Known already:
956 * start_seq < undo_marker and end_seq >= undo_marker.
958 return !before(start_seq
, end_seq
- tp
->max_window
);
961 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
962 * Event "B". Later note: FACK people cheated me again 8), we have to account
963 * for reordering! Ugly, but should help.
965 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
966 * less than what is now known to be received by the other end (derived from
967 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
968 * retransmitted skbs to avoid some costly processing per ACKs.
970 static void tcp_mark_lost_retrans(struct sock
*sk
)
972 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
973 struct tcp_sock
*tp
= tcp_sk(sk
);
976 u32 new_low_seq
= tp
->snd_nxt
;
977 u32 received_upto
= tcp_highest_sack_seq(tp
);
979 if (!tcp_is_fack(tp
) || !tp
->retrans_out
||
980 !after(received_upto
, tp
->lost_retrans_low
) ||
981 icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
984 tcp_for_write_queue(skb
, sk
) {
985 u32 ack_seq
= TCP_SKB_CB(skb
)->ack_seq
;
987 if (skb
== tcp_send_head(sk
))
989 if (cnt
== tp
->retrans_out
)
991 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
994 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
))
997 /* TODO: We would like to get rid of tcp_is_fack(tp) only
998 * constraint here (see above) but figuring out that at
999 * least tp->reordering SACK blocks reside between ack_seq
1000 * and received_upto is not easy task to do cheaply with
1001 * the available datastructures.
1003 * Whether FACK should check here for tp->reordering segs
1004 * in-between one could argue for either way (it would be
1005 * rather simple to implement as we could count fack_count
1006 * during the walk and do tp->fackets_out - fack_count).
1008 if (after(received_upto
, ack_seq
)) {
1009 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1010 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1012 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
1013 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSTRETRANSMIT
);
1015 if (before(ack_seq
, new_low_seq
))
1016 new_low_seq
= ack_seq
;
1017 cnt
+= tcp_skb_pcount(skb
);
1021 if (tp
->retrans_out
)
1022 tp
->lost_retrans_low
= new_low_seq
;
1025 static bool tcp_check_dsack(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1026 struct tcp_sack_block_wire
*sp
, int num_sacks
,
1029 struct tcp_sock
*tp
= tcp_sk(sk
);
1030 u32 start_seq_0
= get_unaligned_be32(&sp
[0].start_seq
);
1031 u32 end_seq_0
= get_unaligned_be32(&sp
[0].end_seq
);
1032 bool dup_sack
= false;
1034 if (before(start_seq_0
, TCP_SKB_CB(ack_skb
)->ack_seq
)) {
1037 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPDSACKRECV
);
1038 } else if (num_sacks
> 1) {
1039 u32 end_seq_1
= get_unaligned_be32(&sp
[1].end_seq
);
1040 u32 start_seq_1
= get_unaligned_be32(&sp
[1].start_seq
);
1042 if (!after(end_seq_0
, end_seq_1
) &&
1043 !before(start_seq_0
, start_seq_1
)) {
1046 NET_INC_STATS_BH(sock_net(sk
),
1047 LINUX_MIB_TCPDSACKOFORECV
);
1051 /* D-SACK for already forgotten data... Do dumb counting. */
1052 if (dup_sack
&& tp
->undo_marker
&& tp
->undo_retrans
&&
1053 !after(end_seq_0
, prior_snd_una
) &&
1054 after(end_seq_0
, tp
->undo_marker
))
1060 struct tcp_sacktag_state
{
1066 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1067 * the incoming SACK may not exactly match but we can find smaller MSS
1068 * aligned portion of it that matches. Therefore we might need to fragment
1069 * which may fail and creates some hassle (caller must handle error case
1072 * FIXME: this could be merged to shift decision code
1074 static int tcp_match_skb_to_sack(struct sock
*sk
, struct sk_buff
*skb
,
1075 u32 start_seq
, u32 end_seq
)
1079 unsigned int pkt_len
;
1082 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1083 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1085 if (tcp_skb_pcount(skb
) > 1 && !in_sack
&&
1086 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
1087 mss
= tcp_skb_mss(skb
);
1088 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1091 pkt_len
= start_seq
- TCP_SKB_CB(skb
)->seq
;
1095 pkt_len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1100 /* Round if necessary so that SACKs cover only full MSSes
1101 * and/or the remaining small portion (if present)
1103 if (pkt_len
> mss
) {
1104 unsigned int new_len
= (pkt_len
/ mss
) * mss
;
1105 if (!in_sack
&& new_len
< pkt_len
) {
1107 if (new_len
> skb
->len
)
1112 err
= tcp_fragment(sk
, skb
, pkt_len
, mss
);
1120 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1121 static u8
tcp_sacktag_one(struct sock
*sk
,
1122 struct tcp_sacktag_state
*state
, u8 sacked
,
1123 u32 start_seq
, u32 end_seq
,
1124 bool dup_sack
, int pcount
)
1126 struct tcp_sock
*tp
= tcp_sk(sk
);
1127 int fack_count
= state
->fack_count
;
1129 /* Account D-SACK for retransmitted packet. */
1130 if (dup_sack
&& (sacked
& TCPCB_RETRANS
)) {
1131 if (tp
->undo_marker
&& tp
->undo_retrans
&&
1132 after(end_seq
, tp
->undo_marker
))
1134 if (sacked
& TCPCB_SACKED_ACKED
)
1135 state
->reord
= min(fack_count
, state
->reord
);
1138 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1139 if (!after(end_seq
, tp
->snd_una
))
1142 if (!(sacked
& TCPCB_SACKED_ACKED
)) {
1143 if (sacked
& TCPCB_SACKED_RETRANS
) {
1144 /* If the segment is not tagged as lost,
1145 * we do not clear RETRANS, believing
1146 * that retransmission is still in flight.
1148 if (sacked
& TCPCB_LOST
) {
1149 sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1150 tp
->lost_out
-= pcount
;
1151 tp
->retrans_out
-= pcount
;
1154 if (!(sacked
& TCPCB_RETRANS
)) {
1155 /* New sack for not retransmitted frame,
1156 * which was in hole. It is reordering.
1158 if (before(start_seq
,
1159 tcp_highest_sack_seq(tp
)))
1160 state
->reord
= min(fack_count
,
1163 /* SACK enhanced F-RTO (RFC4138; Appendix B) */
1164 if (!after(end_seq
, tp
->frto_highmark
))
1165 state
->flag
|= FLAG_ONLY_ORIG_SACKED
;
1168 if (sacked
& TCPCB_LOST
) {
1169 sacked
&= ~TCPCB_LOST
;
1170 tp
->lost_out
-= pcount
;
1174 sacked
|= TCPCB_SACKED_ACKED
;
1175 state
->flag
|= FLAG_DATA_SACKED
;
1176 tp
->sacked_out
+= pcount
;
1178 fack_count
+= pcount
;
1180 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1181 if (!tcp_is_fack(tp
) && (tp
->lost_skb_hint
!= NULL
) &&
1182 before(start_seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
))
1183 tp
->lost_cnt_hint
+= pcount
;
1185 if (fack_count
> tp
->fackets_out
)
1186 tp
->fackets_out
= fack_count
;
1189 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1190 * frames and clear it. undo_retrans is decreased above, L|R frames
1191 * are accounted above as well.
1193 if (dup_sack
&& (sacked
& TCPCB_SACKED_RETRANS
)) {
1194 sacked
&= ~TCPCB_SACKED_RETRANS
;
1195 tp
->retrans_out
-= pcount
;
1201 /* Shift newly-SACKed bytes from this skb to the immediately previous
1202 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1204 static bool tcp_shifted_skb(struct sock
*sk
, struct sk_buff
*skb
,
1205 struct tcp_sacktag_state
*state
,
1206 unsigned int pcount
, int shifted
, int mss
,
1209 struct tcp_sock
*tp
= tcp_sk(sk
);
1210 struct sk_buff
*prev
= tcp_write_queue_prev(sk
, skb
);
1211 u32 start_seq
= TCP_SKB_CB(skb
)->seq
; /* start of newly-SACKed */
1212 u32 end_seq
= start_seq
+ shifted
; /* end of newly-SACKed */
1216 /* Adjust counters and hints for the newly sacked sequence
1217 * range but discard the return value since prev is already
1218 * marked. We must tag the range first because the seq
1219 * advancement below implicitly advances
1220 * tcp_highest_sack_seq() when skb is highest_sack.
1222 tcp_sacktag_one(sk
, state
, TCP_SKB_CB(skb
)->sacked
,
1223 start_seq
, end_seq
, dup_sack
, pcount
);
1225 if (skb
== tp
->lost_skb_hint
)
1226 tp
->lost_cnt_hint
+= pcount
;
1228 TCP_SKB_CB(prev
)->end_seq
+= shifted
;
1229 TCP_SKB_CB(skb
)->seq
+= shifted
;
1231 skb_shinfo(prev
)->gso_segs
+= pcount
;
1232 BUG_ON(skb_shinfo(skb
)->gso_segs
< pcount
);
1233 skb_shinfo(skb
)->gso_segs
-= pcount
;
1235 /* When we're adding to gso_segs == 1, gso_size will be zero,
1236 * in theory this shouldn't be necessary but as long as DSACK
1237 * code can come after this skb later on it's better to keep
1238 * setting gso_size to something.
1240 if (!skb_shinfo(prev
)->gso_size
) {
1241 skb_shinfo(prev
)->gso_size
= mss
;
1242 skb_shinfo(prev
)->gso_type
= sk
->sk_gso_type
;
1245 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1246 if (skb_shinfo(skb
)->gso_segs
<= 1) {
1247 skb_shinfo(skb
)->gso_size
= 0;
1248 skb_shinfo(skb
)->gso_type
= 0;
1251 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1252 TCP_SKB_CB(prev
)->sacked
|= (TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
);
1255 BUG_ON(!tcp_skb_pcount(skb
));
1256 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKSHIFTED
);
1260 /* Whole SKB was eaten :-) */
1262 if (skb
== tp
->retransmit_skb_hint
)
1263 tp
->retransmit_skb_hint
= prev
;
1264 if (skb
== tp
->scoreboard_skb_hint
)
1265 tp
->scoreboard_skb_hint
= prev
;
1266 if (skb
== tp
->lost_skb_hint
) {
1267 tp
->lost_skb_hint
= prev
;
1268 tp
->lost_cnt_hint
-= tcp_skb_pcount(prev
);
1271 TCP_SKB_CB(skb
)->tcp_flags
|= TCP_SKB_CB(prev
)->tcp_flags
;
1272 if (skb
== tcp_highest_sack(sk
))
1273 tcp_advance_highest_sack(sk
, skb
);
1275 tcp_unlink_write_queue(skb
, sk
);
1276 sk_wmem_free_skb(sk
, skb
);
1278 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKMERGED
);
1283 /* I wish gso_size would have a bit more sane initialization than
1284 * something-or-zero which complicates things
1286 static int tcp_skb_seglen(const struct sk_buff
*skb
)
1288 return tcp_skb_pcount(skb
) == 1 ? skb
->len
: tcp_skb_mss(skb
);
1291 /* Shifting pages past head area doesn't work */
1292 static int skb_can_shift(const struct sk_buff
*skb
)
1294 return !skb_headlen(skb
) && skb_is_nonlinear(skb
);
1297 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1300 static struct sk_buff
*tcp_shift_skb_data(struct sock
*sk
, struct sk_buff
*skb
,
1301 struct tcp_sacktag_state
*state
,
1302 u32 start_seq
, u32 end_seq
,
1305 struct tcp_sock
*tp
= tcp_sk(sk
);
1306 struct sk_buff
*prev
;
1312 if (!sk_can_gso(sk
))
1315 /* Normally R but no L won't result in plain S */
1317 (TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_RETRANS
)) == TCPCB_SACKED_RETRANS
)
1319 if (!skb_can_shift(skb
))
1321 /* This frame is about to be dropped (was ACKed). */
1322 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1325 /* Can only happen with delayed DSACK + discard craziness */
1326 if (unlikely(skb
== tcp_write_queue_head(sk
)))
1328 prev
= tcp_write_queue_prev(sk
, skb
);
1330 if ((TCP_SKB_CB(prev
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
)
1333 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1334 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1338 pcount
= tcp_skb_pcount(skb
);
1339 mss
= tcp_skb_seglen(skb
);
1341 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1342 * drop this restriction as unnecessary
1344 if (mss
!= tcp_skb_seglen(prev
))
1347 if (!after(TCP_SKB_CB(skb
)->end_seq
, start_seq
))
1349 /* CHECKME: This is non-MSS split case only?, this will
1350 * cause skipped skbs due to advancing loop btw, original
1351 * has that feature too
1353 if (tcp_skb_pcount(skb
) <= 1)
1356 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1358 /* TODO: head merge to next could be attempted here
1359 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1360 * though it might not be worth of the additional hassle
1362 * ...we can probably just fallback to what was done
1363 * previously. We could try merging non-SACKed ones
1364 * as well but it probably isn't going to buy off
1365 * because later SACKs might again split them, and
1366 * it would make skb timestamp tracking considerably
1372 len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1374 BUG_ON(len
> skb
->len
);
1376 /* MSS boundaries should be honoured or else pcount will
1377 * severely break even though it makes things bit trickier.
1378 * Optimize common case to avoid most of the divides
1380 mss
= tcp_skb_mss(skb
);
1382 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1383 * drop this restriction as unnecessary
1385 if (mss
!= tcp_skb_seglen(prev
))
1390 } else if (len
< mss
) {
1398 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1399 if (!after(TCP_SKB_CB(skb
)->seq
+ len
, tp
->snd_una
))
1402 if (!skb_shift(prev
, skb
, len
))
1404 if (!tcp_shifted_skb(sk
, skb
, state
, pcount
, len
, mss
, dup_sack
))
1407 /* Hole filled allows collapsing with the next as well, this is very
1408 * useful when hole on every nth skb pattern happens
1410 if (prev
== tcp_write_queue_tail(sk
))
1412 skb
= tcp_write_queue_next(sk
, prev
);
1414 if (!skb_can_shift(skb
) ||
1415 (skb
== tcp_send_head(sk
)) ||
1416 ((TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
) ||
1417 (mss
!= tcp_skb_seglen(skb
)))
1421 if (skb_shift(prev
, skb
, len
)) {
1422 pcount
+= tcp_skb_pcount(skb
);
1423 tcp_shifted_skb(sk
, skb
, state
, tcp_skb_pcount(skb
), len
, mss
, 0);
1427 state
->fack_count
+= pcount
;
1434 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKSHIFTFALLBACK
);
1438 static struct sk_buff
*tcp_sacktag_walk(struct sk_buff
*skb
, struct sock
*sk
,
1439 struct tcp_sack_block
*next_dup
,
1440 struct tcp_sacktag_state
*state
,
1441 u32 start_seq
, u32 end_seq
,
1444 struct tcp_sock
*tp
= tcp_sk(sk
);
1445 struct sk_buff
*tmp
;
1447 tcp_for_write_queue_from(skb
, sk
) {
1449 bool dup_sack
= dup_sack_in
;
1451 if (skb
== tcp_send_head(sk
))
1454 /* queue is in-order => we can short-circuit the walk early */
1455 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1458 if ((next_dup
!= NULL
) &&
1459 before(TCP_SKB_CB(skb
)->seq
, next_dup
->end_seq
)) {
1460 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1461 next_dup
->start_seq
,
1467 /* skb reference here is a bit tricky to get right, since
1468 * shifting can eat and free both this skb and the next,
1469 * so not even _safe variant of the loop is enough.
1472 tmp
= tcp_shift_skb_data(sk
, skb
, state
,
1473 start_seq
, end_seq
, dup_sack
);
1482 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1488 if (unlikely(in_sack
< 0))
1492 TCP_SKB_CB(skb
)->sacked
=
1495 TCP_SKB_CB(skb
)->sacked
,
1496 TCP_SKB_CB(skb
)->seq
,
1497 TCP_SKB_CB(skb
)->end_seq
,
1499 tcp_skb_pcount(skb
));
1501 if (!before(TCP_SKB_CB(skb
)->seq
,
1502 tcp_highest_sack_seq(tp
)))
1503 tcp_advance_highest_sack(sk
, skb
);
1506 state
->fack_count
+= tcp_skb_pcount(skb
);
1511 /* Avoid all extra work that is being done by sacktag while walking in
1514 static struct sk_buff
*tcp_sacktag_skip(struct sk_buff
*skb
, struct sock
*sk
,
1515 struct tcp_sacktag_state
*state
,
1518 tcp_for_write_queue_from(skb
, sk
) {
1519 if (skb
== tcp_send_head(sk
))
1522 if (after(TCP_SKB_CB(skb
)->end_seq
, skip_to_seq
))
1525 state
->fack_count
+= tcp_skb_pcount(skb
);
1530 static struct sk_buff
*tcp_maybe_skipping_dsack(struct sk_buff
*skb
,
1532 struct tcp_sack_block
*next_dup
,
1533 struct tcp_sacktag_state
*state
,
1536 if (next_dup
== NULL
)
1539 if (before(next_dup
->start_seq
, skip_to_seq
)) {
1540 skb
= tcp_sacktag_skip(skb
, sk
, state
, next_dup
->start_seq
);
1541 skb
= tcp_sacktag_walk(skb
, sk
, NULL
, state
,
1542 next_dup
->start_seq
, next_dup
->end_seq
,
1549 static int tcp_sack_cache_ok(const struct tcp_sock
*tp
, const struct tcp_sack_block
*cache
)
1551 return cache
< tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1555 tcp_sacktag_write_queue(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1558 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1559 struct tcp_sock
*tp
= tcp_sk(sk
);
1560 const unsigned char *ptr
= (skb_transport_header(ack_skb
) +
1561 TCP_SKB_CB(ack_skb
)->sacked
);
1562 struct tcp_sack_block_wire
*sp_wire
= (struct tcp_sack_block_wire
*)(ptr
+2);
1563 struct tcp_sack_block sp
[TCP_NUM_SACKS
];
1564 struct tcp_sack_block
*cache
;
1565 struct tcp_sacktag_state state
;
1566 struct sk_buff
*skb
;
1567 int num_sacks
= min(TCP_NUM_SACKS
, (ptr
[1] - TCPOLEN_SACK_BASE
) >> 3);
1569 bool found_dup_sack
= false;
1571 int first_sack_index
;
1574 state
.reord
= tp
->packets_out
;
1576 if (!tp
->sacked_out
) {
1577 if (WARN_ON(tp
->fackets_out
))
1578 tp
->fackets_out
= 0;
1579 tcp_highest_sack_reset(sk
);
1582 found_dup_sack
= tcp_check_dsack(sk
, ack_skb
, sp_wire
,
1583 num_sacks
, prior_snd_una
);
1585 state
.flag
|= FLAG_DSACKING_ACK
;
1587 /* Eliminate too old ACKs, but take into
1588 * account more or less fresh ones, they can
1589 * contain valid SACK info.
1591 if (before(TCP_SKB_CB(ack_skb
)->ack_seq
, prior_snd_una
- tp
->max_window
))
1594 if (!tp
->packets_out
)
1598 first_sack_index
= 0;
1599 for (i
= 0; i
< num_sacks
; i
++) {
1600 bool dup_sack
= !i
&& found_dup_sack
;
1602 sp
[used_sacks
].start_seq
= get_unaligned_be32(&sp_wire
[i
].start_seq
);
1603 sp
[used_sacks
].end_seq
= get_unaligned_be32(&sp_wire
[i
].end_seq
);
1605 if (!tcp_is_sackblock_valid(tp
, dup_sack
,
1606 sp
[used_sacks
].start_seq
,
1607 sp
[used_sacks
].end_seq
)) {
1611 if (!tp
->undo_marker
)
1612 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDNOUNDO
;
1614 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDOLD
;
1616 /* Don't count olds caused by ACK reordering */
1617 if ((TCP_SKB_CB(ack_skb
)->ack_seq
!= tp
->snd_una
) &&
1618 !after(sp
[used_sacks
].end_seq
, tp
->snd_una
))
1620 mib_idx
= LINUX_MIB_TCPSACKDISCARD
;
1623 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
1625 first_sack_index
= -1;
1629 /* Ignore very old stuff early */
1630 if (!after(sp
[used_sacks
].end_seq
, prior_snd_una
))
1636 /* order SACK blocks to allow in order walk of the retrans queue */
1637 for (i
= used_sacks
- 1; i
> 0; i
--) {
1638 for (j
= 0; j
< i
; j
++) {
1639 if (after(sp
[j
].start_seq
, sp
[j
+ 1].start_seq
)) {
1640 swap(sp
[j
], sp
[j
+ 1]);
1642 /* Track where the first SACK block goes to */
1643 if (j
== first_sack_index
)
1644 first_sack_index
= j
+ 1;
1649 skb
= tcp_write_queue_head(sk
);
1650 state
.fack_count
= 0;
1653 if (!tp
->sacked_out
) {
1654 /* It's already past, so skip checking against it */
1655 cache
= tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1657 cache
= tp
->recv_sack_cache
;
1658 /* Skip empty blocks in at head of the cache */
1659 while (tcp_sack_cache_ok(tp
, cache
) && !cache
->start_seq
&&
1664 while (i
< used_sacks
) {
1665 u32 start_seq
= sp
[i
].start_seq
;
1666 u32 end_seq
= sp
[i
].end_seq
;
1667 bool dup_sack
= (found_dup_sack
&& (i
== first_sack_index
));
1668 struct tcp_sack_block
*next_dup
= NULL
;
1670 if (found_dup_sack
&& ((i
+ 1) == first_sack_index
))
1671 next_dup
= &sp
[i
+ 1];
1673 /* Skip too early cached blocks */
1674 while (tcp_sack_cache_ok(tp
, cache
) &&
1675 !before(start_seq
, cache
->end_seq
))
1678 /* Can skip some work by looking recv_sack_cache? */
1679 if (tcp_sack_cache_ok(tp
, cache
) && !dup_sack
&&
1680 after(end_seq
, cache
->start_seq
)) {
1683 if (before(start_seq
, cache
->start_seq
)) {
1684 skb
= tcp_sacktag_skip(skb
, sk
, &state
,
1686 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
,
1693 /* Rest of the block already fully processed? */
1694 if (!after(end_seq
, cache
->end_seq
))
1697 skb
= tcp_maybe_skipping_dsack(skb
, sk
, next_dup
,
1701 /* ...tail remains todo... */
1702 if (tcp_highest_sack_seq(tp
) == cache
->end_seq
) {
1703 /* ...but better entrypoint exists! */
1704 skb
= tcp_highest_sack(sk
);
1707 state
.fack_count
= tp
->fackets_out
;
1712 skb
= tcp_sacktag_skip(skb
, sk
, &state
, cache
->end_seq
);
1713 /* Check overlap against next cached too (past this one already) */
1718 if (!before(start_seq
, tcp_highest_sack_seq(tp
))) {
1719 skb
= tcp_highest_sack(sk
);
1722 state
.fack_count
= tp
->fackets_out
;
1724 skb
= tcp_sacktag_skip(skb
, sk
, &state
, start_seq
);
1727 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
, &state
,
1728 start_seq
, end_seq
, dup_sack
);
1731 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1732 * due to in-order walk
1734 if (after(end_seq
, tp
->frto_highmark
))
1735 state
.flag
&= ~FLAG_ONLY_ORIG_SACKED
;
1740 /* Clear the head of the cache sack blocks so we can skip it next time */
1741 for (i
= 0; i
< ARRAY_SIZE(tp
->recv_sack_cache
) - used_sacks
; i
++) {
1742 tp
->recv_sack_cache
[i
].start_seq
= 0;
1743 tp
->recv_sack_cache
[i
].end_seq
= 0;
1745 for (j
= 0; j
< used_sacks
; j
++)
1746 tp
->recv_sack_cache
[i
++] = sp
[j
];
1748 tcp_mark_lost_retrans(sk
);
1750 tcp_verify_left_out(tp
);
1752 if ((state
.reord
< tp
->fackets_out
) &&
1753 ((icsk
->icsk_ca_state
!= TCP_CA_Loss
) || tp
->undo_marker
) &&
1754 (!tp
->frto_highmark
|| after(tp
->snd_una
, tp
->frto_highmark
)))
1755 tcp_update_reordering(sk
, tp
->fackets_out
- state
.reord
, 0);
1759 #if FASTRETRANS_DEBUG > 0
1760 WARN_ON((int)tp
->sacked_out
< 0);
1761 WARN_ON((int)tp
->lost_out
< 0);
1762 WARN_ON((int)tp
->retrans_out
< 0);
1763 WARN_ON((int)tcp_packets_in_flight(tp
) < 0);
1768 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1769 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1771 static bool tcp_limit_reno_sacked(struct tcp_sock
*tp
)
1775 holes
= max(tp
->lost_out
, 1U);
1776 holes
= min(holes
, tp
->packets_out
);
1778 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1779 tp
->sacked_out
= tp
->packets_out
- holes
;
1785 /* If we receive more dupacks than we expected counting segments
1786 * in assumption of absent reordering, interpret this as reordering.
1787 * The only another reason could be bug in receiver TCP.
1789 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1791 struct tcp_sock
*tp
= tcp_sk(sk
);
1792 if (tcp_limit_reno_sacked(tp
))
1793 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1796 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1798 static void tcp_add_reno_sack(struct sock
*sk
)
1800 struct tcp_sock
*tp
= tcp_sk(sk
);
1802 tcp_check_reno_reordering(sk
, 0);
1803 tcp_verify_left_out(tp
);
1806 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1808 static void tcp_remove_reno_sacks(struct sock
*sk
, int acked
)
1810 struct tcp_sock
*tp
= tcp_sk(sk
);
1813 /* One ACK acked hole. The rest eat duplicate ACKs. */
1814 if (acked
- 1 >= tp
->sacked_out
)
1817 tp
->sacked_out
-= acked
- 1;
1819 tcp_check_reno_reordering(sk
, acked
);
1820 tcp_verify_left_out(tp
);
1823 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1828 static int tcp_is_sackfrto(const struct tcp_sock
*tp
)
1830 return (sysctl_tcp_frto
== 0x2) && !tcp_is_reno(tp
);
1833 /* F-RTO can only be used if TCP has never retransmitted anything other than
1834 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
1836 bool tcp_use_frto(struct sock
*sk
)
1838 const struct tcp_sock
*tp
= tcp_sk(sk
);
1839 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1840 struct sk_buff
*skb
;
1842 if (!sysctl_tcp_frto
)
1845 /* MTU probe and F-RTO won't really play nicely along currently */
1846 if (icsk
->icsk_mtup
.probe_size
)
1849 if (tcp_is_sackfrto(tp
))
1852 /* Avoid expensive walking of rexmit queue if possible */
1853 if (tp
->retrans_out
> 1)
1856 skb
= tcp_write_queue_head(sk
);
1857 if (tcp_skb_is_last(sk
, skb
))
1859 skb
= tcp_write_queue_next(sk
, skb
); /* Skips head */
1860 tcp_for_write_queue_from(skb
, sk
) {
1861 if (skb
== tcp_send_head(sk
))
1863 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
1865 /* Short-circuit when first non-SACKed skb has been checked */
1866 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
1872 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
1873 * recovery a bit and use heuristics in tcp_process_frto() to detect if
1874 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
1875 * keep retrans_out counting accurate (with SACK F-RTO, other than head
1876 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
1877 * bits are handled if the Loss state is really to be entered (in
1878 * tcp_enter_frto_loss).
1880 * Do like tcp_enter_loss() would; when RTO expires the second time it
1882 * "Reduce ssthresh if it has not yet been made inside this window."
1884 void tcp_enter_frto(struct sock
*sk
)
1886 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1887 struct tcp_sock
*tp
= tcp_sk(sk
);
1888 struct sk_buff
*skb
;
1890 if ((!tp
->frto_counter
&& icsk
->icsk_ca_state
<= TCP_CA_Disorder
) ||
1891 tp
->snd_una
== tp
->high_seq
||
1892 ((icsk
->icsk_ca_state
== TCP_CA_Loss
|| tp
->frto_counter
) &&
1893 !icsk
->icsk_retransmits
)) {
1894 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1895 /* Our state is too optimistic in ssthresh() call because cwnd
1896 * is not reduced until tcp_enter_frto_loss() when previous F-RTO
1897 * recovery has not yet completed. Pattern would be this: RTO,
1898 * Cumulative ACK, RTO (2xRTO for the same segment does not end
1900 * RFC4138 should be more specific on what to do, even though
1901 * RTO is quite unlikely to occur after the first Cumulative ACK
1902 * due to back-off and complexity of triggering events ...
1904 if (tp
->frto_counter
) {
1906 stored_cwnd
= tp
->snd_cwnd
;
1908 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1909 tp
->snd_cwnd
= stored_cwnd
;
1911 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1913 /* ... in theory, cong.control module could do "any tricks" in
1914 * ssthresh(), which means that ca_state, lost bits and lost_out
1915 * counter would have to be faked before the call occurs. We
1916 * consider that too expensive, unlikely and hacky, so modules
1917 * using these in ssthresh() must deal these incompatibility
1918 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
1920 tcp_ca_event(sk
, CA_EVENT_FRTO
);
1923 tp
->undo_marker
= tp
->snd_una
;
1924 tp
->undo_retrans
= 0;
1926 skb
= tcp_write_queue_head(sk
);
1927 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
1928 tp
->undo_marker
= 0;
1929 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
1930 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1931 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1933 tcp_verify_left_out(tp
);
1935 /* Too bad if TCP was application limited */
1936 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tcp_packets_in_flight(tp
) + 1);
1938 /* Earlier loss recovery underway (see RFC4138; Appendix B).
1939 * The last condition is necessary at least in tp->frto_counter case.
1941 if (tcp_is_sackfrto(tp
) && (tp
->frto_counter
||
1942 ((1 << icsk
->icsk_ca_state
) & (TCPF_CA_Recovery
|TCPF_CA_Loss
))) &&
1943 after(tp
->high_seq
, tp
->snd_una
)) {
1944 tp
->frto_highmark
= tp
->high_seq
;
1946 tp
->frto_highmark
= tp
->snd_nxt
;
1948 tcp_set_ca_state(sk
, TCP_CA_Disorder
);
1949 tp
->high_seq
= tp
->snd_nxt
;
1950 tp
->frto_counter
= 1;
1953 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1954 * which indicates that we should follow the traditional RTO recovery,
1955 * i.e. mark everything lost and do go-back-N retransmission.
1957 static void tcp_enter_frto_loss(struct sock
*sk
, int allowed_segments
, int flag
)
1959 struct tcp_sock
*tp
= tcp_sk(sk
);
1960 struct sk_buff
*skb
;
1963 tp
->retrans_out
= 0;
1964 if (tcp_is_reno(tp
))
1965 tcp_reset_reno_sack(tp
);
1967 tcp_for_write_queue(skb
, sk
) {
1968 if (skb
== tcp_send_head(sk
))
1971 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
1973 * Count the retransmission made on RTO correctly (only when
1974 * waiting for the first ACK and did not get it)...
1976 if ((tp
->frto_counter
== 1) && !(flag
& FLAG_DATA_ACKED
)) {
1977 /* For some reason this R-bit might get cleared? */
1978 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
1979 tp
->retrans_out
+= tcp_skb_pcount(skb
);
1980 /* ...enter this if branch just for the first segment */
1981 flag
|= FLAG_DATA_ACKED
;
1983 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
1984 tp
->undo_marker
= 0;
1985 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1988 /* Marking forward transmissions that were made after RTO lost
1989 * can cause unnecessary retransmissions in some scenarios,
1990 * SACK blocks will mitigate that in some but not in all cases.
1991 * We used to not mark them but it was causing break-ups with
1992 * receivers that do only in-order receival.
1994 * TODO: we could detect presence of such receiver and select
1995 * different behavior per flow.
1997 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
1998 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1999 tp
->lost_out
+= tcp_skb_pcount(skb
);
2000 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
2003 tcp_verify_left_out(tp
);
2005 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + allowed_segments
;
2006 tp
->snd_cwnd_cnt
= 0;
2007 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2008 tp
->frto_counter
= 0;
2010 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
2011 sysctl_tcp_reordering
);
2012 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2013 tp
->high_seq
= tp
->snd_nxt
;
2014 TCP_ECN_queue_cwr(tp
);
2016 tcp_clear_all_retrans_hints(tp
);
2019 static void tcp_clear_retrans_partial(struct tcp_sock
*tp
)
2021 tp
->retrans_out
= 0;
2024 tp
->undo_marker
= 0;
2025 tp
->undo_retrans
= 0;
2028 void tcp_clear_retrans(struct tcp_sock
*tp
)
2030 tcp_clear_retrans_partial(tp
);
2032 tp
->fackets_out
= 0;
2036 /* Enter Loss state. If "how" is not zero, forget all SACK information
2037 * and reset tags completely, otherwise preserve SACKs. If receiver
2038 * dropped its ofo queue, we will know this due to reneging detection.
2040 void tcp_enter_loss(struct sock
*sk
, int how
)
2042 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2043 struct tcp_sock
*tp
= tcp_sk(sk
);
2044 struct sk_buff
*skb
;
2046 /* Reduce ssthresh if it has not yet been made inside this window. */
2047 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
|| tp
->snd_una
== tp
->high_seq
||
2048 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
2049 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2050 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
2051 tcp_ca_event(sk
, CA_EVENT_LOSS
);
2054 tp
->snd_cwnd_cnt
= 0;
2055 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2057 tcp_clear_retrans_partial(tp
);
2059 if (tcp_is_reno(tp
))
2060 tcp_reset_reno_sack(tp
);
2062 tp
->undo_marker
= tp
->snd_una
;
2065 tp
->fackets_out
= 0;
2067 tcp_clear_all_retrans_hints(tp
);
2069 tcp_for_write_queue(skb
, sk
) {
2070 if (skb
== tcp_send_head(sk
))
2073 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
2074 tp
->undo_marker
= 0;
2075 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
2076 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) || how
) {
2077 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
2078 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
2079 tp
->lost_out
+= tcp_skb_pcount(skb
);
2080 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
2083 tcp_verify_left_out(tp
);
2085 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
2086 sysctl_tcp_reordering
);
2087 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2088 tp
->high_seq
= tp
->snd_nxt
;
2089 TCP_ECN_queue_cwr(tp
);
2090 /* Abort F-RTO algorithm if one is in progress */
2091 tp
->frto_counter
= 0;
2094 /* If ACK arrived pointing to a remembered SACK, it means that our
2095 * remembered SACKs do not reflect real state of receiver i.e.
2096 * receiver _host_ is heavily congested (or buggy).
2098 * Do processing similar to RTO timeout.
2100 static bool tcp_check_sack_reneging(struct sock
*sk
, int flag
)
2102 if (flag
& FLAG_SACK_RENEGING
) {
2103 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2104 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSACKRENEGING
);
2106 tcp_enter_loss(sk
, 1);
2107 icsk
->icsk_retransmits
++;
2108 tcp_retransmit_skb(sk
, tcp_write_queue_head(sk
));
2109 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2110 icsk
->icsk_rto
, TCP_RTO_MAX
);
2116 static inline int tcp_fackets_out(const struct tcp_sock
*tp
)
2118 return tcp_is_reno(tp
) ? tp
->sacked_out
+ 1 : tp
->fackets_out
;
2121 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2122 * counter when SACK is enabled (without SACK, sacked_out is used for
2125 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2126 * segments up to the highest received SACK block so far and holes in
2129 * With reordering, holes may still be in flight, so RFC3517 recovery
2130 * uses pure sacked_out (total number of SACKed segments) even though
2131 * it violates the RFC that uses duplicate ACKs, often these are equal
2132 * but when e.g. out-of-window ACKs or packet duplication occurs,
2133 * they differ. Since neither occurs due to loss, TCP should really
2136 static inline int tcp_dupack_heuristics(const struct tcp_sock
*tp
)
2138 return tcp_is_fack(tp
) ? tp
->fackets_out
: tp
->sacked_out
+ 1;
2141 static bool tcp_pause_early_retransmit(struct sock
*sk
, int flag
)
2143 struct tcp_sock
*tp
= tcp_sk(sk
);
2144 unsigned long delay
;
2146 /* Delay early retransmit and entering fast recovery for
2147 * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
2148 * available, or RTO is scheduled to fire first.
2150 if (sysctl_tcp_early_retrans
< 2 || (flag
& FLAG_ECE
) || !tp
->srtt
)
2153 delay
= max_t(unsigned long, (tp
->srtt
>> 5), msecs_to_jiffies(2));
2154 if (!time_after(inet_csk(sk
)->icsk_timeout
, (jiffies
+ delay
)))
2157 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
, delay
, TCP_RTO_MAX
);
2158 tp
->early_retrans_delayed
= 1;
2162 static inline int tcp_skb_timedout(const struct sock
*sk
,
2163 const struct sk_buff
*skb
)
2165 return tcp_time_stamp
- TCP_SKB_CB(skb
)->when
> inet_csk(sk
)->icsk_rto
;
2168 static inline int tcp_head_timedout(const struct sock
*sk
)
2170 const struct tcp_sock
*tp
= tcp_sk(sk
);
2172 return tp
->packets_out
&&
2173 tcp_skb_timedout(sk
, tcp_write_queue_head(sk
));
2176 /* Linux NewReno/SACK/FACK/ECN state machine.
2177 * --------------------------------------
2179 * "Open" Normal state, no dubious events, fast path.
2180 * "Disorder" In all the respects it is "Open",
2181 * but requires a bit more attention. It is entered when
2182 * we see some SACKs or dupacks. It is split of "Open"
2183 * mainly to move some processing from fast path to slow one.
2184 * "CWR" CWND was reduced due to some Congestion Notification event.
2185 * It can be ECN, ICMP source quench, local device congestion.
2186 * "Recovery" CWND was reduced, we are fast-retransmitting.
2187 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2189 * tcp_fastretrans_alert() is entered:
2190 * - each incoming ACK, if state is not "Open"
2191 * - when arrived ACK is unusual, namely:
2196 * Counting packets in flight is pretty simple.
2198 * in_flight = packets_out - left_out + retrans_out
2200 * packets_out is SND.NXT-SND.UNA counted in packets.
2202 * retrans_out is number of retransmitted segments.
2204 * left_out is number of segments left network, but not ACKed yet.
2206 * left_out = sacked_out + lost_out
2208 * sacked_out: Packets, which arrived to receiver out of order
2209 * and hence not ACKed. With SACKs this number is simply
2210 * amount of SACKed data. Even without SACKs
2211 * it is easy to give pretty reliable estimate of this number,
2212 * counting duplicate ACKs.
2214 * lost_out: Packets lost by network. TCP has no explicit
2215 * "loss notification" feedback from network (for now).
2216 * It means that this number can be only _guessed_.
2217 * Actually, it is the heuristics to predict lossage that
2218 * distinguishes different algorithms.
2220 * F.e. after RTO, when all the queue is considered as lost,
2221 * lost_out = packets_out and in_flight = retrans_out.
2223 * Essentially, we have now two algorithms counting
2226 * FACK: It is the simplest heuristics. As soon as we decided
2227 * that something is lost, we decide that _all_ not SACKed
2228 * packets until the most forward SACK are lost. I.e.
2229 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2230 * It is absolutely correct estimate, if network does not reorder
2231 * packets. And it loses any connection to reality when reordering
2232 * takes place. We use FACK by default until reordering
2233 * is suspected on the path to this destination.
2235 * NewReno: when Recovery is entered, we assume that one segment
2236 * is lost (classic Reno). While we are in Recovery and
2237 * a partial ACK arrives, we assume that one more packet
2238 * is lost (NewReno). This heuristics are the same in NewReno
2241 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2242 * deflation etc. CWND is real congestion window, never inflated, changes
2243 * only according to classic VJ rules.
2245 * Really tricky (and requiring careful tuning) part of algorithm
2246 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2247 * The first determines the moment _when_ we should reduce CWND and,
2248 * hence, slow down forward transmission. In fact, it determines the moment
2249 * when we decide that hole is caused by loss, rather than by a reorder.
2251 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2252 * holes, caused by lost packets.
2254 * And the most logically complicated part of algorithm is undo
2255 * heuristics. We detect false retransmits due to both too early
2256 * fast retransmit (reordering) and underestimated RTO, analyzing
2257 * timestamps and D-SACKs. When we detect that some segments were
2258 * retransmitted by mistake and CWND reduction was wrong, we undo
2259 * window reduction and abort recovery phase. This logic is hidden
2260 * inside several functions named tcp_try_undo_<something>.
2263 /* This function decides, when we should leave Disordered state
2264 * and enter Recovery phase, reducing congestion window.
2266 * Main question: may we further continue forward transmission
2267 * with the same cwnd?
2269 static bool tcp_time_to_recover(struct sock
*sk
, int flag
)
2271 struct tcp_sock
*tp
= tcp_sk(sk
);
2274 /* Do not perform any recovery during F-RTO algorithm */
2275 if (tp
->frto_counter
)
2278 /* Trick#1: The loss is proven. */
2282 /* Not-A-Trick#2 : Classic rule... */
2283 if (tcp_dupack_heuristics(tp
) > tp
->reordering
)
2286 /* Trick#3 : when we use RFC2988 timer restart, fast
2287 * retransmit can be triggered by timeout of queue head.
2289 if (tcp_is_fack(tp
) && tcp_head_timedout(sk
))
2292 /* Trick#4: It is still not OK... But will it be useful to delay
2295 packets_out
= tp
->packets_out
;
2296 if (packets_out
<= tp
->reordering
&&
2297 tp
->sacked_out
>= max_t(__u32
, packets_out
/2, sysctl_tcp_reordering
) &&
2298 !tcp_may_send_now(sk
)) {
2299 /* We have nothing to send. This connection is limited
2300 * either by receiver window or by application.
2305 /* If a thin stream is detected, retransmit after first
2306 * received dupack. Employ only if SACK is supported in order
2307 * to avoid possible corner-case series of spurious retransmissions
2308 * Use only if there are no unsent data.
2310 if ((tp
->thin_dupack
|| sysctl_tcp_thin_dupack
) &&
2311 tcp_stream_is_thin(tp
) && tcp_dupack_heuristics(tp
) > 1 &&
2312 tcp_is_sack(tp
) && !tcp_send_head(sk
))
2315 /* Trick#6: TCP early retransmit, per RFC5827. To avoid spurious
2316 * retransmissions due to small network reorderings, we implement
2317 * Mitigation A.3 in the RFC and delay the retransmission for a short
2318 * interval if appropriate.
2320 if (tp
->do_early_retrans
&& !tp
->retrans_out
&& tp
->sacked_out
&&
2321 (tp
->packets_out
== (tp
->sacked_out
+ 1) && tp
->packets_out
< 4) &&
2322 !tcp_may_send_now(sk
))
2323 return !tcp_pause_early_retransmit(sk
, flag
);
2328 /* New heuristics: it is possible only after we switched to restart timer
2329 * each time when something is ACKed. Hence, we can detect timed out packets
2330 * during fast retransmit without falling to slow start.
2332 * Usefulness of this as is very questionable, since we should know which of
2333 * the segments is the next to timeout which is relatively expensive to find
2334 * in general case unless we add some data structure just for that. The
2335 * current approach certainly won't find the right one too often and when it
2336 * finally does find _something_ it usually marks large part of the window
2337 * right away (because a retransmission with a larger timestamp blocks the
2338 * loop from advancing). -ij
2340 static void tcp_timeout_skbs(struct sock
*sk
)
2342 struct tcp_sock
*tp
= tcp_sk(sk
);
2343 struct sk_buff
*skb
;
2345 if (!tcp_is_fack(tp
) || !tcp_head_timedout(sk
))
2348 skb
= tp
->scoreboard_skb_hint
;
2349 if (tp
->scoreboard_skb_hint
== NULL
)
2350 skb
= tcp_write_queue_head(sk
);
2352 tcp_for_write_queue_from(skb
, sk
) {
2353 if (skb
== tcp_send_head(sk
))
2355 if (!tcp_skb_timedout(sk
, skb
))
2358 tcp_skb_mark_lost(tp
, skb
);
2361 tp
->scoreboard_skb_hint
= skb
;
2363 tcp_verify_left_out(tp
);
2366 /* Detect loss in event "A" above by marking head of queue up as lost.
2367 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2368 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2369 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2370 * the maximum SACKed segments to pass before reaching this limit.
2372 static void tcp_mark_head_lost(struct sock
*sk
, int packets
, int mark_head
)
2374 struct tcp_sock
*tp
= tcp_sk(sk
);
2375 struct sk_buff
*skb
;
2379 /* Use SACK to deduce losses of new sequences sent during recovery */
2380 const u32 loss_high
= tcp_is_sack(tp
) ? tp
->snd_nxt
: tp
->high_seq
;
2382 WARN_ON(packets
> tp
->packets_out
);
2383 if (tp
->lost_skb_hint
) {
2384 skb
= tp
->lost_skb_hint
;
2385 cnt
= tp
->lost_cnt_hint
;
2386 /* Head already handled? */
2387 if (mark_head
&& skb
!= tcp_write_queue_head(sk
))
2390 skb
= tcp_write_queue_head(sk
);
2394 tcp_for_write_queue_from(skb
, sk
) {
2395 if (skb
== tcp_send_head(sk
))
2397 /* TODO: do this better */
2398 /* this is not the most efficient way to do this... */
2399 tp
->lost_skb_hint
= skb
;
2400 tp
->lost_cnt_hint
= cnt
;
2402 if (after(TCP_SKB_CB(skb
)->end_seq
, loss_high
))
2406 if (tcp_is_fack(tp
) || tcp_is_reno(tp
) ||
2407 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2408 cnt
+= tcp_skb_pcount(skb
);
2410 if (cnt
> packets
) {
2411 if ((tcp_is_sack(tp
) && !tcp_is_fack(tp
)) ||
2412 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
) ||
2413 (oldcnt
>= packets
))
2416 mss
= skb_shinfo(skb
)->gso_size
;
2417 err
= tcp_fragment(sk
, skb
, (packets
- oldcnt
) * mss
, mss
);
2423 tcp_skb_mark_lost(tp
, skb
);
2428 tcp_verify_left_out(tp
);
2431 /* Account newly detected lost packet(s) */
2433 static void tcp_update_scoreboard(struct sock
*sk
, int fast_rexmit
)
2435 struct tcp_sock
*tp
= tcp_sk(sk
);
2437 if (tcp_is_reno(tp
)) {
2438 tcp_mark_head_lost(sk
, 1, 1);
2439 } else if (tcp_is_fack(tp
)) {
2440 int lost
= tp
->fackets_out
- tp
->reordering
;
2443 tcp_mark_head_lost(sk
, lost
, 0);
2445 int sacked_upto
= tp
->sacked_out
- tp
->reordering
;
2446 if (sacked_upto
>= 0)
2447 tcp_mark_head_lost(sk
, sacked_upto
, 0);
2448 else if (fast_rexmit
)
2449 tcp_mark_head_lost(sk
, 1, 1);
2452 tcp_timeout_skbs(sk
);
2455 /* CWND moderation, preventing bursts due to too big ACKs
2456 * in dubious situations.
2458 static inline void tcp_moderate_cwnd(struct tcp_sock
*tp
)
2460 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
2461 tcp_packets_in_flight(tp
) + tcp_max_burst(tp
));
2462 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2465 /* Nothing was retransmitted or returned timestamp is less
2466 * than timestamp of the first retransmission.
2468 static inline bool tcp_packet_delayed(const struct tcp_sock
*tp
)
2470 return !tp
->retrans_stamp
||
2471 (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2472 before(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
));
2475 /* Undo procedures. */
2477 #if FASTRETRANS_DEBUG > 1
2478 static void DBGUNDO(struct sock
*sk
, const char *msg
)
2480 struct tcp_sock
*tp
= tcp_sk(sk
);
2481 struct inet_sock
*inet
= inet_sk(sk
);
2483 if (sk
->sk_family
== AF_INET
) {
2484 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2486 &inet
->inet_daddr
, ntohs(inet
->inet_dport
),
2487 tp
->snd_cwnd
, tcp_left_out(tp
),
2488 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2491 #if IS_ENABLED(CONFIG_IPV6)
2492 else if (sk
->sk_family
== AF_INET6
) {
2493 struct ipv6_pinfo
*np
= inet6_sk(sk
);
2494 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2496 &np
->daddr
, ntohs(inet
->inet_dport
),
2497 tp
->snd_cwnd
, tcp_left_out(tp
),
2498 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2504 #define DBGUNDO(x...) do { } while (0)
2507 static void tcp_undo_cwr(struct sock
*sk
, const bool undo_ssthresh
)
2509 struct tcp_sock
*tp
= tcp_sk(sk
);
2511 if (tp
->prior_ssthresh
) {
2512 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2514 if (icsk
->icsk_ca_ops
->undo_cwnd
)
2515 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
2517 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
<< 1);
2519 if (undo_ssthresh
&& tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
2520 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
2521 TCP_ECN_withdraw_cwr(tp
);
2524 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2526 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2529 static inline bool tcp_may_undo(const struct tcp_sock
*tp
)
2531 return tp
->undo_marker
&& (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
2534 /* People celebrate: "We love our President!" */
2535 static bool tcp_try_undo_recovery(struct sock
*sk
)
2537 struct tcp_sock
*tp
= tcp_sk(sk
);
2539 if (tcp_may_undo(tp
)) {
2542 /* Happy end! We did not retransmit anything
2543 * or our original transmission succeeded.
2545 DBGUNDO(sk
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
2546 tcp_undo_cwr(sk
, true);
2547 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
2548 mib_idx
= LINUX_MIB_TCPLOSSUNDO
;
2550 mib_idx
= LINUX_MIB_TCPFULLUNDO
;
2552 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2553 tp
->undo_marker
= 0;
2555 if (tp
->snd_una
== tp
->high_seq
&& tcp_is_reno(tp
)) {
2556 /* Hold old state until something *above* high_seq
2557 * is ACKed. For Reno it is MUST to prevent false
2558 * fast retransmits (RFC2582). SACK TCP is safe. */
2559 tcp_moderate_cwnd(tp
);
2562 tcp_set_ca_state(sk
, TCP_CA_Open
);
2566 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2567 static void tcp_try_undo_dsack(struct sock
*sk
)
2569 struct tcp_sock
*tp
= tcp_sk(sk
);
2571 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
2572 DBGUNDO(sk
, "D-SACK");
2573 tcp_undo_cwr(sk
, true);
2574 tp
->undo_marker
= 0;
2575 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPDSACKUNDO
);
2579 /* We can clear retrans_stamp when there are no retransmissions in the
2580 * window. It would seem that it is trivially available for us in
2581 * tp->retrans_out, however, that kind of assumptions doesn't consider
2582 * what will happen if errors occur when sending retransmission for the
2583 * second time. ...It could the that such segment has only
2584 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2585 * the head skb is enough except for some reneging corner cases that
2586 * are not worth the effort.
2588 * Main reason for all this complexity is the fact that connection dying
2589 * time now depends on the validity of the retrans_stamp, in particular,
2590 * that successive retransmissions of a segment must not advance
2591 * retrans_stamp under any conditions.
2593 static bool tcp_any_retrans_done(const struct sock
*sk
)
2595 const struct tcp_sock
*tp
= tcp_sk(sk
);
2596 struct sk_buff
*skb
;
2598 if (tp
->retrans_out
)
2601 skb
= tcp_write_queue_head(sk
);
2602 if (unlikely(skb
&& TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
))
2608 /* Undo during fast recovery after partial ACK. */
2610 static int tcp_try_undo_partial(struct sock
*sk
, int acked
)
2612 struct tcp_sock
*tp
= tcp_sk(sk
);
2613 /* Partial ACK arrived. Force Hoe's retransmit. */
2614 int failed
= tcp_is_reno(tp
) || (tcp_fackets_out(tp
) > tp
->reordering
);
2616 if (tcp_may_undo(tp
)) {
2617 /* Plain luck! Hole if filled with delayed
2618 * packet, rather than with a retransmit.
2620 if (!tcp_any_retrans_done(sk
))
2621 tp
->retrans_stamp
= 0;
2623 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
2626 tcp_undo_cwr(sk
, false);
2627 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPARTIALUNDO
);
2629 /* So... Do not make Hoe's retransmit yet.
2630 * If the first packet was delayed, the rest
2631 * ones are most probably delayed as well.
2638 /* Undo during loss recovery after partial ACK. */
2639 static bool tcp_try_undo_loss(struct sock
*sk
)
2641 struct tcp_sock
*tp
= tcp_sk(sk
);
2643 if (tcp_may_undo(tp
)) {
2644 struct sk_buff
*skb
;
2645 tcp_for_write_queue(skb
, sk
) {
2646 if (skb
== tcp_send_head(sk
))
2648 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2651 tcp_clear_all_retrans_hints(tp
);
2653 DBGUNDO(sk
, "partial loss");
2655 tcp_undo_cwr(sk
, true);
2656 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSSUNDO
);
2657 inet_csk(sk
)->icsk_retransmits
= 0;
2658 tp
->undo_marker
= 0;
2659 if (tcp_is_sack(tp
))
2660 tcp_set_ca_state(sk
, TCP_CA_Open
);
2666 /* The cwnd reduction in CWR and Recovery use the PRR algorithm
2667 * https://datatracker.ietf.org/doc/draft-ietf-tcpm-proportional-rate-reduction/
2668 * It computes the number of packets to send (sndcnt) based on packets newly
2670 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2671 * cwnd reductions across a full RTT.
2672 * 2) If packets in flight is lower than ssthresh (such as due to excess
2673 * losses and/or application stalls), do not perform any further cwnd
2674 * reductions, but instead slow start up to ssthresh.
2676 static void tcp_init_cwnd_reduction(struct sock
*sk
, const bool set_ssthresh
)
2678 struct tcp_sock
*tp
= tcp_sk(sk
);
2680 tp
->high_seq
= tp
->snd_nxt
;
2681 tp
->snd_cwnd_cnt
= 0;
2682 tp
->prior_cwnd
= tp
->snd_cwnd
;
2683 tp
->prr_delivered
= 0;
2686 tp
->snd_ssthresh
= inet_csk(sk
)->icsk_ca_ops
->ssthresh(sk
);
2687 TCP_ECN_queue_cwr(tp
);
2690 static void tcp_cwnd_reduction(struct sock
*sk
, int newly_acked_sacked
,
2693 struct tcp_sock
*tp
= tcp_sk(sk
);
2695 int delta
= tp
->snd_ssthresh
- tcp_packets_in_flight(tp
);
2697 tp
->prr_delivered
+= newly_acked_sacked
;
2698 if (tcp_packets_in_flight(tp
) > tp
->snd_ssthresh
) {
2699 u64 dividend
= (u64
)tp
->snd_ssthresh
* tp
->prr_delivered
+
2701 sndcnt
= div_u64(dividend
, tp
->prior_cwnd
) - tp
->prr_out
;
2703 sndcnt
= min_t(int, delta
,
2704 max_t(int, tp
->prr_delivered
- tp
->prr_out
,
2705 newly_acked_sacked
) + 1);
2708 sndcnt
= max(sndcnt
, (fast_rexmit
? 1 : 0));
2709 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + sndcnt
;
2712 static inline void tcp_end_cwnd_reduction(struct sock
*sk
)
2714 struct tcp_sock
*tp
= tcp_sk(sk
);
2716 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2717 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_CWR
||
2718 (tp
->undo_marker
&& tp
->snd_ssthresh
< TCP_INFINITE_SSTHRESH
)) {
2719 tp
->snd_cwnd
= tp
->snd_ssthresh
;
2720 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2722 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
2725 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2726 void tcp_enter_cwr(struct sock
*sk
, const int set_ssthresh
)
2728 struct tcp_sock
*tp
= tcp_sk(sk
);
2730 tp
->prior_ssthresh
= 0;
2731 if (inet_csk(sk
)->icsk_ca_state
< TCP_CA_CWR
) {
2732 tp
->undo_marker
= 0;
2733 tcp_init_cwnd_reduction(sk
, set_ssthresh
);
2734 tcp_set_ca_state(sk
, TCP_CA_CWR
);
2738 static void tcp_try_keep_open(struct sock
*sk
)
2740 struct tcp_sock
*tp
= tcp_sk(sk
);
2741 int state
= TCP_CA_Open
;
2743 if (tcp_left_out(tp
) || tcp_any_retrans_done(sk
))
2744 state
= TCP_CA_Disorder
;
2746 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
2747 tcp_set_ca_state(sk
, state
);
2748 tp
->high_seq
= tp
->snd_nxt
;
2752 static void tcp_try_to_open(struct sock
*sk
, int flag
, int newly_acked_sacked
)
2754 struct tcp_sock
*tp
= tcp_sk(sk
);
2756 tcp_verify_left_out(tp
);
2758 if (!tp
->frto_counter
&& !tcp_any_retrans_done(sk
))
2759 tp
->retrans_stamp
= 0;
2761 if (flag
& FLAG_ECE
)
2762 tcp_enter_cwr(sk
, 1);
2764 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
2765 tcp_try_keep_open(sk
);
2766 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
)
2767 tcp_moderate_cwnd(tp
);
2769 tcp_cwnd_reduction(sk
, newly_acked_sacked
, 0);
2773 static void tcp_mtup_probe_failed(struct sock
*sk
)
2775 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2777 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
2778 icsk
->icsk_mtup
.probe_size
= 0;
2781 static void tcp_mtup_probe_success(struct sock
*sk
)
2783 struct tcp_sock
*tp
= tcp_sk(sk
);
2784 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2786 /* FIXME: breaks with very large cwnd */
2787 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2788 tp
->snd_cwnd
= tp
->snd_cwnd
*
2789 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
2790 icsk
->icsk_mtup
.probe_size
;
2791 tp
->snd_cwnd_cnt
= 0;
2792 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2793 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2795 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
2796 icsk
->icsk_mtup
.probe_size
= 0;
2797 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
2800 /* Do a simple retransmit without using the backoff mechanisms in
2801 * tcp_timer. This is used for path mtu discovery.
2802 * The socket is already locked here.
2804 void tcp_simple_retransmit(struct sock
*sk
)
2806 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2807 struct tcp_sock
*tp
= tcp_sk(sk
);
2808 struct sk_buff
*skb
;
2809 unsigned int mss
= tcp_current_mss(sk
);
2810 u32 prior_lost
= tp
->lost_out
;
2812 tcp_for_write_queue(skb
, sk
) {
2813 if (skb
== tcp_send_head(sk
))
2815 if (tcp_skb_seglen(skb
) > mss
&&
2816 !(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
2817 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2818 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
2819 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2821 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
2825 tcp_clear_retrans_hints_partial(tp
);
2827 if (prior_lost
== tp
->lost_out
)
2830 if (tcp_is_reno(tp
))
2831 tcp_limit_reno_sacked(tp
);
2833 tcp_verify_left_out(tp
);
2835 /* Don't muck with the congestion window here.
2836 * Reason is that we do not increase amount of _data_
2837 * in network, but units changed and effective
2838 * cwnd/ssthresh really reduced now.
2840 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
2841 tp
->high_seq
= tp
->snd_nxt
;
2842 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2843 tp
->prior_ssthresh
= 0;
2844 tp
->undo_marker
= 0;
2845 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2847 tcp_xmit_retransmit_queue(sk
);
2849 EXPORT_SYMBOL(tcp_simple_retransmit
);
2851 static void tcp_enter_recovery(struct sock
*sk
, bool ece_ack
)
2853 struct tcp_sock
*tp
= tcp_sk(sk
);
2856 if (tcp_is_reno(tp
))
2857 mib_idx
= LINUX_MIB_TCPRENORECOVERY
;
2859 mib_idx
= LINUX_MIB_TCPSACKRECOVERY
;
2861 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2863 tp
->prior_ssthresh
= 0;
2864 tp
->undo_marker
= tp
->snd_una
;
2865 tp
->undo_retrans
= tp
->retrans_out
;
2867 if (inet_csk(sk
)->icsk_ca_state
< TCP_CA_CWR
) {
2869 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2870 tcp_init_cwnd_reduction(sk
, true);
2872 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
2875 /* Process an event, which can update packets-in-flight not trivially.
2876 * Main goal of this function is to calculate new estimate for left_out,
2877 * taking into account both packets sitting in receiver's buffer and
2878 * packets lost by network.
2880 * Besides that it does CWND reduction, when packet loss is detected
2881 * and changes state of machine.
2883 * It does _not_ decide what to send, it is made in function
2884 * tcp_xmit_retransmit_queue().
2886 static void tcp_fastretrans_alert(struct sock
*sk
, int pkts_acked
,
2887 int prior_sacked
, bool is_dupack
,
2890 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2891 struct tcp_sock
*tp
= tcp_sk(sk
);
2892 int do_lost
= is_dupack
|| ((flag
& FLAG_DATA_SACKED
) &&
2893 (tcp_fackets_out(tp
) > tp
->reordering
));
2894 int newly_acked_sacked
= 0;
2895 int fast_rexmit
= 0;
2897 if (WARN_ON(!tp
->packets_out
&& tp
->sacked_out
))
2899 if (WARN_ON(!tp
->sacked_out
&& tp
->fackets_out
))
2900 tp
->fackets_out
= 0;
2902 /* Now state machine starts.
2903 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2904 if (flag
& FLAG_ECE
)
2905 tp
->prior_ssthresh
= 0;
2907 /* B. In all the states check for reneging SACKs. */
2908 if (tcp_check_sack_reneging(sk
, flag
))
2911 /* C. Check consistency of the current state. */
2912 tcp_verify_left_out(tp
);
2914 /* D. Check state exit conditions. State can be terminated
2915 * when high_seq is ACKed. */
2916 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
2917 WARN_ON(tp
->retrans_out
!= 0);
2918 tp
->retrans_stamp
= 0;
2919 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
2920 switch (icsk
->icsk_ca_state
) {
2922 icsk
->icsk_retransmits
= 0;
2923 if (tcp_try_undo_recovery(sk
))
2928 /* CWR is to be held something *above* high_seq
2929 * is ACKed for CWR bit to reach receiver. */
2930 if (tp
->snd_una
!= tp
->high_seq
) {
2931 tcp_end_cwnd_reduction(sk
);
2932 tcp_set_ca_state(sk
, TCP_CA_Open
);
2936 case TCP_CA_Recovery
:
2937 if (tcp_is_reno(tp
))
2938 tcp_reset_reno_sack(tp
);
2939 if (tcp_try_undo_recovery(sk
))
2941 tcp_end_cwnd_reduction(sk
);
2946 /* E. Process state. */
2947 switch (icsk
->icsk_ca_state
) {
2948 case TCP_CA_Recovery
:
2949 if (!(flag
& FLAG_SND_UNA_ADVANCED
)) {
2950 if (tcp_is_reno(tp
) && is_dupack
)
2951 tcp_add_reno_sack(sk
);
2953 do_lost
= tcp_try_undo_partial(sk
, pkts_acked
);
2954 newly_acked_sacked
= pkts_acked
+ tp
->sacked_out
- prior_sacked
;
2957 if (flag
& FLAG_DATA_ACKED
)
2958 icsk
->icsk_retransmits
= 0;
2959 if (tcp_is_reno(tp
) && flag
& FLAG_SND_UNA_ADVANCED
)
2960 tcp_reset_reno_sack(tp
);
2961 if (!tcp_try_undo_loss(sk
)) {
2962 tcp_moderate_cwnd(tp
);
2963 tcp_xmit_retransmit_queue(sk
);
2966 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
2968 /* Loss is undone; fall through to processing in Open state. */
2970 if (tcp_is_reno(tp
)) {
2971 if (flag
& FLAG_SND_UNA_ADVANCED
)
2972 tcp_reset_reno_sack(tp
);
2974 tcp_add_reno_sack(sk
);
2976 newly_acked_sacked
= pkts_acked
+ tp
->sacked_out
- prior_sacked
;
2978 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
)
2979 tcp_try_undo_dsack(sk
);
2981 if (!tcp_time_to_recover(sk
, flag
)) {
2982 tcp_try_to_open(sk
, flag
, newly_acked_sacked
);
2986 /* MTU probe failure: don't reduce cwnd */
2987 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
2988 icsk
->icsk_mtup
.probe_size
&&
2989 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
2990 tcp_mtup_probe_failed(sk
);
2991 /* Restores the reduction we did in tcp_mtup_probe() */
2993 tcp_simple_retransmit(sk
);
2997 /* Otherwise enter Recovery state */
2998 tcp_enter_recovery(sk
, (flag
& FLAG_ECE
));
3002 if (do_lost
|| (tcp_is_fack(tp
) && tcp_head_timedout(sk
)))
3003 tcp_update_scoreboard(sk
, fast_rexmit
);
3004 tcp_cwnd_reduction(sk
, newly_acked_sacked
, fast_rexmit
);
3005 tcp_xmit_retransmit_queue(sk
);
3008 void tcp_valid_rtt_meas(struct sock
*sk
, u32 seq_rtt
)
3010 tcp_rtt_estimator(sk
, seq_rtt
);
3012 inet_csk(sk
)->icsk_backoff
= 0;
3014 EXPORT_SYMBOL(tcp_valid_rtt_meas
);
3016 /* Read draft-ietf-tcplw-high-performance before mucking
3017 * with this code. (Supersedes RFC1323)
3019 static void tcp_ack_saw_tstamp(struct sock
*sk
, int flag
)
3021 /* RTTM Rule: A TSecr value received in a segment is used to
3022 * update the averaged RTT measurement only if the segment
3023 * acknowledges some new data, i.e., only if it advances the
3024 * left edge of the send window.
3026 * See draft-ietf-tcplw-high-performance-00, section 3.3.
3027 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
3029 * Changed: reset backoff as soon as we see the first valid sample.
3030 * If we do not, we get strongly overestimated rto. With timestamps
3031 * samples are accepted even from very old segments: f.e., when rtt=1
3032 * increases to 8, we retransmit 5 times and after 8 seconds delayed
3033 * answer arrives rto becomes 120 seconds! If at least one of segments
3034 * in window is lost... Voila. --ANK (010210)
3036 struct tcp_sock
*tp
= tcp_sk(sk
);
3038 tcp_valid_rtt_meas(sk
, tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
);
3041 static void tcp_ack_no_tstamp(struct sock
*sk
, u32 seq_rtt
, int flag
)
3043 /* We don't have a timestamp. Can only use
3044 * packets that are not retransmitted to determine
3045 * rtt estimates. Also, we must not reset the
3046 * backoff for rto until we get a non-retransmitted
3047 * packet. This allows us to deal with a situation
3048 * where the network delay has increased suddenly.
3049 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
3052 if (flag
& FLAG_RETRANS_DATA_ACKED
)
3055 tcp_valid_rtt_meas(sk
, seq_rtt
);
3058 static inline void tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
3061 const struct tcp_sock
*tp
= tcp_sk(sk
);
3062 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
3063 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
3064 tcp_ack_saw_tstamp(sk
, flag
);
3065 else if (seq_rtt
>= 0)
3066 tcp_ack_no_tstamp(sk
, seq_rtt
, flag
);
3069 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
, u32 in_flight
)
3071 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3072 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, in_flight
);
3073 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_time_stamp
;
3076 /* Restart timer after forward progress on connection.
3077 * RFC2988 recommends to restart timer to now+rto.
3079 void tcp_rearm_rto(struct sock
*sk
)
3081 struct tcp_sock
*tp
= tcp_sk(sk
);
3083 /* If the retrans timer is currently being used by Fast Open
3084 * for SYN-ACK retrans purpose, stay put.
3086 if (tp
->fastopen_rsk
)
3089 if (!tp
->packets_out
) {
3090 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
3092 u32 rto
= inet_csk(sk
)->icsk_rto
;
3093 /* Offset the time elapsed after installing regular RTO */
3094 if (tp
->early_retrans_delayed
) {
3095 struct sk_buff
*skb
= tcp_write_queue_head(sk
);
3096 const u32 rto_time_stamp
= TCP_SKB_CB(skb
)->when
+ rto
;
3097 s32 delta
= (s32
)(rto_time_stamp
- tcp_time_stamp
);
3098 /* delta may not be positive if the socket is locked
3099 * when the delayed ER timer fires and is rescheduled.
3104 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
, rto
,
3107 tp
->early_retrans_delayed
= 0;
3110 /* This function is called when the delayed ER timer fires. TCP enters
3111 * fast recovery and performs fast-retransmit.
3113 void tcp_resume_early_retransmit(struct sock
*sk
)
3115 struct tcp_sock
*tp
= tcp_sk(sk
);
3119 /* Stop if ER is disabled after the delayed ER timer is scheduled */
3120 if (!tp
->do_early_retrans
)
3123 tcp_enter_recovery(sk
, false);
3124 tcp_update_scoreboard(sk
, 1);
3125 tcp_xmit_retransmit_queue(sk
);
3128 /* If we get here, the whole TSO packet has not been acked. */
3129 static u32
tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
)
3131 struct tcp_sock
*tp
= tcp_sk(sk
);
3134 BUG_ON(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
));
3136 packets_acked
= tcp_skb_pcount(skb
);
3137 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
3139 packets_acked
-= tcp_skb_pcount(skb
);
3141 if (packets_acked
) {
3142 BUG_ON(tcp_skb_pcount(skb
) == 0);
3143 BUG_ON(!before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
));
3146 return packets_acked
;
3149 /* Remove acknowledged frames from the retransmission queue. If our packet
3150 * is before the ack sequence we can discard it as it's confirmed to have
3151 * arrived at the other end.
3153 static int tcp_clean_rtx_queue(struct sock
*sk
, int prior_fackets
,
3156 struct tcp_sock
*tp
= tcp_sk(sk
);
3157 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3158 struct sk_buff
*skb
;
3159 u32 now
= tcp_time_stamp
;
3160 int fully_acked
= true;
3163 u32 reord
= tp
->packets_out
;
3164 u32 prior_sacked
= tp
->sacked_out
;
3166 s32 ca_seq_rtt
= -1;
3167 ktime_t last_ackt
= net_invalid_timestamp();
3169 while ((skb
= tcp_write_queue_head(sk
)) && skb
!= tcp_send_head(sk
)) {
3170 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
3172 u8 sacked
= scb
->sacked
;
3174 /* Determine how many packets and what bytes were acked, tso and else */
3175 if (after(scb
->end_seq
, tp
->snd_una
)) {
3176 if (tcp_skb_pcount(skb
) == 1 ||
3177 !after(tp
->snd_una
, scb
->seq
))
3180 acked_pcount
= tcp_tso_acked(sk
, skb
);
3184 fully_acked
= false;
3186 acked_pcount
= tcp_skb_pcount(skb
);
3189 if (sacked
& TCPCB_RETRANS
) {
3190 if (sacked
& TCPCB_SACKED_RETRANS
)
3191 tp
->retrans_out
-= acked_pcount
;
3192 flag
|= FLAG_RETRANS_DATA_ACKED
;
3195 if ((flag
& FLAG_DATA_ACKED
) || (acked_pcount
> 1))
3196 flag
|= FLAG_NONHEAD_RETRANS_ACKED
;
3198 ca_seq_rtt
= now
- scb
->when
;
3199 last_ackt
= skb
->tstamp
;
3201 seq_rtt
= ca_seq_rtt
;
3203 if (!(sacked
& TCPCB_SACKED_ACKED
))
3204 reord
= min(pkts_acked
, reord
);
3207 if (sacked
& TCPCB_SACKED_ACKED
)
3208 tp
->sacked_out
-= acked_pcount
;
3209 if (sacked
& TCPCB_LOST
)
3210 tp
->lost_out
-= acked_pcount
;
3212 tp
->packets_out
-= acked_pcount
;
3213 pkts_acked
+= acked_pcount
;
3215 /* Initial outgoing SYN's get put onto the write_queue
3216 * just like anything else we transmit. It is not
3217 * true data, and if we misinform our callers that
3218 * this ACK acks real data, we will erroneously exit
3219 * connection startup slow start one packet too
3220 * quickly. This is severely frowned upon behavior.
3222 if (!(scb
->tcp_flags
& TCPHDR_SYN
)) {
3223 flag
|= FLAG_DATA_ACKED
;
3225 flag
|= FLAG_SYN_ACKED
;
3226 tp
->retrans_stamp
= 0;
3232 tcp_unlink_write_queue(skb
, sk
);
3233 sk_wmem_free_skb(sk
, skb
);
3234 tp
->scoreboard_skb_hint
= NULL
;
3235 if (skb
== tp
->retransmit_skb_hint
)
3236 tp
->retransmit_skb_hint
= NULL
;
3237 if (skb
== tp
->lost_skb_hint
)
3238 tp
->lost_skb_hint
= NULL
;
3241 if (likely(between(tp
->snd_up
, prior_snd_una
, tp
->snd_una
)))
3242 tp
->snd_up
= tp
->snd_una
;
3244 if (skb
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
3245 flag
|= FLAG_SACK_RENEGING
;
3247 if (flag
& FLAG_ACKED
) {
3248 const struct tcp_congestion_ops
*ca_ops
3249 = inet_csk(sk
)->icsk_ca_ops
;
3251 if (unlikely(icsk
->icsk_mtup
.probe_size
&&
3252 !after(tp
->mtu_probe
.probe_seq_end
, tp
->snd_una
))) {
3253 tcp_mtup_probe_success(sk
);
3256 tcp_ack_update_rtt(sk
, flag
, seq_rtt
);
3259 if (tcp_is_reno(tp
)) {
3260 tcp_remove_reno_sacks(sk
, pkts_acked
);
3264 /* Non-retransmitted hole got filled? That's reordering */
3265 if (reord
< prior_fackets
)
3266 tcp_update_reordering(sk
, tp
->fackets_out
- reord
, 0);
3268 delta
= tcp_is_fack(tp
) ? pkts_acked
:
3269 prior_sacked
- tp
->sacked_out
;
3270 tp
->lost_cnt_hint
-= min(tp
->lost_cnt_hint
, delta
);
3273 tp
->fackets_out
-= min(pkts_acked
, tp
->fackets_out
);
3275 if (ca_ops
->pkts_acked
) {
3278 /* Is the ACK triggering packet unambiguous? */
3279 if (!(flag
& FLAG_RETRANS_DATA_ACKED
)) {
3280 /* High resolution needed and available? */
3281 if (ca_ops
->flags
& TCP_CONG_RTT_STAMP
&&
3282 !ktime_equal(last_ackt
,
3283 net_invalid_timestamp()))
3284 rtt_us
= ktime_us_delta(ktime_get_real(),
3286 else if (ca_seq_rtt
>= 0)
3287 rtt_us
= jiffies_to_usecs(ca_seq_rtt
);
3290 ca_ops
->pkts_acked(sk
, pkts_acked
, rtt_us
);
3294 #if FASTRETRANS_DEBUG > 0
3295 WARN_ON((int)tp
->sacked_out
< 0);
3296 WARN_ON((int)tp
->lost_out
< 0);
3297 WARN_ON((int)tp
->retrans_out
< 0);
3298 if (!tp
->packets_out
&& tcp_is_sack(tp
)) {
3299 icsk
= inet_csk(sk
);
3301 pr_debug("Leak l=%u %d\n",
3302 tp
->lost_out
, icsk
->icsk_ca_state
);
3305 if (tp
->sacked_out
) {
3306 pr_debug("Leak s=%u %d\n",
3307 tp
->sacked_out
, icsk
->icsk_ca_state
);
3310 if (tp
->retrans_out
) {
3311 pr_debug("Leak r=%u %d\n",
3312 tp
->retrans_out
, icsk
->icsk_ca_state
);
3313 tp
->retrans_out
= 0;
3320 static void tcp_ack_probe(struct sock
*sk
)
3322 const struct tcp_sock
*tp
= tcp_sk(sk
);
3323 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3325 /* Was it a usable window open? */
3327 if (!after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
, tcp_wnd_end(tp
))) {
3328 icsk
->icsk_backoff
= 0;
3329 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
3330 /* Socket must be waked up by subsequent tcp_data_snd_check().
3331 * This function is not for random using!
3334 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3335 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
3340 static inline bool tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
3342 return !(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
3343 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
;
3346 static inline bool tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
3348 const struct tcp_sock
*tp
= tcp_sk(sk
);
3349 return (!(flag
& FLAG_ECE
) || tp
->snd_cwnd
< tp
->snd_ssthresh
) &&
3350 !tcp_in_cwnd_reduction(sk
);
3353 /* Check that window update is acceptable.
3354 * The function assumes that snd_una<=ack<=snd_next.
3356 static inline bool tcp_may_update_window(const struct tcp_sock
*tp
,
3357 const u32 ack
, const u32 ack_seq
,
3360 return after(ack
, tp
->snd_una
) ||
3361 after(ack_seq
, tp
->snd_wl1
) ||
3362 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
);
3365 /* Update our send window.
3367 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3368 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3370 static int tcp_ack_update_window(struct sock
*sk
, const struct sk_buff
*skb
, u32 ack
,
3373 struct tcp_sock
*tp
= tcp_sk(sk
);
3375 u32 nwin
= ntohs(tcp_hdr(skb
)->window
);
3377 if (likely(!tcp_hdr(skb
)->syn
))
3378 nwin
<<= tp
->rx_opt
.snd_wscale
;
3380 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
3381 flag
|= FLAG_WIN_UPDATE
;
3382 tcp_update_wl(tp
, ack_seq
);
3384 if (tp
->snd_wnd
!= nwin
) {
3387 /* Note, it is the only place, where
3388 * fast path is recovered for sending TCP.
3391 tcp_fast_path_check(sk
);
3393 if (nwin
> tp
->max_window
) {
3394 tp
->max_window
= nwin
;
3395 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
3405 /* A very conservative spurious RTO response algorithm: reduce cwnd and
3406 * continue in congestion avoidance.
3408 static void tcp_conservative_spur_to_response(struct tcp_sock
*tp
)
3410 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
3411 tp
->snd_cwnd_cnt
= 0;
3412 TCP_ECN_queue_cwr(tp
);
3413 tcp_moderate_cwnd(tp
);
3416 /* A conservative spurious RTO response algorithm: reduce cwnd using
3417 * PRR and continue in congestion avoidance.
3419 static void tcp_cwr_spur_to_response(struct sock
*sk
)
3421 tcp_enter_cwr(sk
, 0);
3424 static void tcp_undo_spur_to_response(struct sock
*sk
, int flag
)
3426 if (flag
& FLAG_ECE
)
3427 tcp_cwr_spur_to_response(sk
);
3429 tcp_undo_cwr(sk
, true);
3432 /* F-RTO spurious RTO detection algorithm (RFC4138)
3434 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3435 * comments). State (ACK number) is kept in frto_counter. When ACK advances
3436 * window (but not to or beyond highest sequence sent before RTO):
3437 * On First ACK, send two new segments out.
3438 * On Second ACK, RTO was likely spurious. Do spurious response (response
3439 * algorithm is not part of the F-RTO detection algorithm
3440 * given in RFC4138 but can be selected separately).
3441 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
3442 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3443 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3444 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
3446 * Rationale: if the RTO was spurious, new ACKs should arrive from the
3447 * original window even after we transmit two new data segments.
3450 * on first step, wait until first cumulative ACK arrives, then move to
3451 * the second step. In second step, the next ACK decides.
3453 * F-RTO is implemented (mainly) in four functions:
3454 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
3455 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3456 * called when tcp_use_frto() showed green light
3457 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3458 * - tcp_enter_frto_loss() is called if there is not enough evidence
3459 * to prove that the RTO is indeed spurious. It transfers the control
3460 * from F-RTO to the conventional RTO recovery
3462 static bool tcp_process_frto(struct sock
*sk
, int flag
)
3464 struct tcp_sock
*tp
= tcp_sk(sk
);
3466 tcp_verify_left_out(tp
);
3468 /* Duplicate the behavior from Loss state (fastretrans_alert) */
3469 if (flag
& FLAG_DATA_ACKED
)
3470 inet_csk(sk
)->icsk_retransmits
= 0;
3472 if ((flag
& FLAG_NONHEAD_RETRANS_ACKED
) ||
3473 ((tp
->frto_counter
>= 2) && (flag
& FLAG_RETRANS_DATA_ACKED
)))
3474 tp
->undo_marker
= 0;
3476 if (!before(tp
->snd_una
, tp
->frto_highmark
)) {
3477 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 2 : 3), flag
);
3481 if (!tcp_is_sackfrto(tp
)) {
3482 /* RFC4138 shortcoming in step 2; should also have case c):
3483 * ACK isn't duplicate nor advances window, e.g., opposite dir
3486 if (!(flag
& FLAG_ANY_PROGRESS
) && (flag
& FLAG_NOT_DUP
))
3489 if (!(flag
& FLAG_DATA_ACKED
)) {
3490 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 0 : 3),
3495 if (!(flag
& FLAG_DATA_ACKED
) && (tp
->frto_counter
== 1)) {
3496 if (!tcp_packets_in_flight(tp
)) {
3497 tcp_enter_frto_loss(sk
, 2, flag
);
3501 /* Prevent sending of new data. */
3502 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
3503 tcp_packets_in_flight(tp
));
3507 if ((tp
->frto_counter
>= 2) &&
3508 (!(flag
& FLAG_FORWARD_PROGRESS
) ||
3509 ((flag
& FLAG_DATA_SACKED
) &&
3510 !(flag
& FLAG_ONLY_ORIG_SACKED
)))) {
3511 /* RFC4138 shortcoming (see comment above) */
3512 if (!(flag
& FLAG_FORWARD_PROGRESS
) &&
3513 (flag
& FLAG_NOT_DUP
))
3516 tcp_enter_frto_loss(sk
, 3, flag
);
3521 if (tp
->frto_counter
== 1) {
3522 /* tcp_may_send_now needs to see updated state */
3523 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + 2;
3524 tp
->frto_counter
= 2;
3526 if (!tcp_may_send_now(sk
))
3527 tcp_enter_frto_loss(sk
, 2, flag
);
3531 switch (sysctl_tcp_frto_response
) {
3533 tcp_undo_spur_to_response(sk
, flag
);
3536 tcp_conservative_spur_to_response(tp
);
3539 tcp_cwr_spur_to_response(sk
);
3542 tp
->frto_counter
= 0;
3543 tp
->undo_marker
= 0;
3544 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSPURIOUSRTOS
);
3549 /* RFC 5961 7 [ACK Throttling] */
3550 static void tcp_send_challenge_ack(struct sock
*sk
)
3552 /* unprotected vars, we dont care of overwrites */
3553 static u32 challenge_timestamp
;
3554 static unsigned int challenge_count
;
3555 u32 now
= jiffies
/ HZ
;
3557 if (now
!= challenge_timestamp
) {
3558 challenge_timestamp
= now
;
3559 challenge_count
= 0;
3561 if (++challenge_count
<= sysctl_tcp_challenge_ack_limit
) {
3562 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPCHALLENGEACK
);
3567 /* This routine deals with incoming acks, but not outgoing ones. */
3568 static int tcp_ack(struct sock
*sk
, const struct sk_buff
*skb
, int flag
)
3570 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3571 struct tcp_sock
*tp
= tcp_sk(sk
);
3572 u32 prior_snd_una
= tp
->snd_una
;
3573 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
3574 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3575 bool is_dupack
= false;
3576 u32 prior_in_flight
;
3579 int prior_sacked
= tp
->sacked_out
;
3581 bool frto_cwnd
= false;
3583 /* If the ack is older than previous acks
3584 * then we can probably ignore it.
3586 if (before(ack
, prior_snd_una
)) {
3587 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3588 if (before(ack
, prior_snd_una
- tp
->max_window
)) {
3589 tcp_send_challenge_ack(sk
);
3595 /* If the ack includes data we haven't sent yet, discard
3596 * this segment (RFC793 Section 3.9).
3598 if (after(ack
, tp
->snd_nxt
))
3601 if (tp
->early_retrans_delayed
)
3604 if (after(ack
, prior_snd_una
))
3605 flag
|= FLAG_SND_UNA_ADVANCED
;
3607 prior_fackets
= tp
->fackets_out
;
3608 prior_in_flight
= tcp_packets_in_flight(tp
);
3610 if (!(flag
& FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
3611 /* Window is constant, pure forward advance.
3612 * No more checks are required.
3613 * Note, we use the fact that SND.UNA>=SND.WL2.
3615 tcp_update_wl(tp
, ack_seq
);
3617 flag
|= FLAG_WIN_UPDATE
;
3619 tcp_ca_event(sk
, CA_EVENT_FAST_ACK
);
3621 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPACKS
);
3623 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
3626 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPUREACKS
);
3628 flag
|= tcp_ack_update_window(sk
, skb
, ack
, ack_seq
);
3630 if (TCP_SKB_CB(skb
)->sacked
)
3631 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
3633 if (TCP_ECN_rcv_ecn_echo(tp
, tcp_hdr(skb
)))
3636 tcp_ca_event(sk
, CA_EVENT_SLOW_ACK
);
3639 /* We passed data and got it acked, remove any soft error
3640 * log. Something worked...
3642 sk
->sk_err_soft
= 0;
3643 icsk
->icsk_probes_out
= 0;
3644 tp
->rcv_tstamp
= tcp_time_stamp
;
3645 prior_packets
= tp
->packets_out
;
3649 /* See if we can take anything off of the retransmit queue. */
3650 flag
|= tcp_clean_rtx_queue(sk
, prior_fackets
, prior_snd_una
);
3652 pkts_acked
= prior_packets
- tp
->packets_out
;
3654 if (tp
->frto_counter
)
3655 frto_cwnd
= tcp_process_frto(sk
, flag
);
3656 /* Guarantee sacktag reordering detection against wrap-arounds */
3657 if (before(tp
->frto_highmark
, tp
->snd_una
))
3658 tp
->frto_highmark
= 0;
3660 if (tcp_ack_is_dubious(sk
, flag
)) {
3661 /* Advance CWND, if state allows this. */
3662 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
&&
3663 tcp_may_raise_cwnd(sk
, flag
))
3664 tcp_cong_avoid(sk
, ack
, prior_in_flight
);
3665 is_dupack
= !(flag
& (FLAG_SND_UNA_ADVANCED
| FLAG_NOT_DUP
));
3666 tcp_fastretrans_alert(sk
, pkts_acked
, prior_sacked
,
3669 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
)
3670 tcp_cong_avoid(sk
, ack
, prior_in_flight
);
3673 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
& FLAG_NOT_DUP
)) {
3674 struct dst_entry
*dst
= __sk_dst_get(sk
);
3681 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3682 if (flag
& FLAG_DSACKING_ACK
)
3683 tcp_fastretrans_alert(sk
, pkts_acked
, prior_sacked
,
3685 /* If this ack opens up a zero window, clear backoff. It was
3686 * being used to time the probes, and is probably far higher than
3687 * it needs to be for normal retransmission.
3689 if (tcp_send_head(sk
))
3694 SOCK_DEBUG(sk
, "Ack %u after %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3698 /* If data was SACKed, tag it and see if we should send more data.
3699 * If data was DSACKed, see if we can undo a cwnd reduction.
3701 if (TCP_SKB_CB(skb
)->sacked
) {
3702 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
3703 tcp_fastretrans_alert(sk
, pkts_acked
, prior_sacked
,
3707 SOCK_DEBUG(sk
, "Ack %u before %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3711 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3712 * But, this can also be called on packets in the established flow when
3713 * the fast version below fails.
3715 void tcp_parse_options(const struct sk_buff
*skb
, struct tcp_options_received
*opt_rx
,
3716 const u8
**hvpp
, int estab
,
3717 struct tcp_fastopen_cookie
*foc
)
3719 const unsigned char *ptr
;
3720 const struct tcphdr
*th
= tcp_hdr(skb
);
3721 int length
= (th
->doff
* 4) - sizeof(struct tcphdr
);
3723 ptr
= (const unsigned char *)(th
+ 1);
3724 opt_rx
->saw_tstamp
= 0;
3726 while (length
> 0) {
3727 int opcode
= *ptr
++;
3733 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
3738 if (opsize
< 2) /* "silly options" */
3740 if (opsize
> length
)
3741 return; /* don't parse partial options */
3744 if (opsize
== TCPOLEN_MSS
&& th
->syn
&& !estab
) {
3745 u16 in_mss
= get_unaligned_be16(ptr
);
3747 if (opt_rx
->user_mss
&&
3748 opt_rx
->user_mss
< in_mss
)
3749 in_mss
= opt_rx
->user_mss
;
3750 opt_rx
->mss_clamp
= in_mss
;
3755 if (opsize
== TCPOLEN_WINDOW
&& th
->syn
&&
3756 !estab
&& sysctl_tcp_window_scaling
) {
3757 __u8 snd_wscale
= *(__u8
*)ptr
;
3758 opt_rx
->wscale_ok
= 1;
3759 if (snd_wscale
> 14) {
3760 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3765 opt_rx
->snd_wscale
= snd_wscale
;
3768 case TCPOPT_TIMESTAMP
:
3769 if ((opsize
== TCPOLEN_TIMESTAMP
) &&
3770 ((estab
&& opt_rx
->tstamp_ok
) ||
3771 (!estab
&& sysctl_tcp_timestamps
))) {
3772 opt_rx
->saw_tstamp
= 1;
3773 opt_rx
->rcv_tsval
= get_unaligned_be32(ptr
);
3774 opt_rx
->rcv_tsecr
= get_unaligned_be32(ptr
+ 4);
3777 case TCPOPT_SACK_PERM
:
3778 if (opsize
== TCPOLEN_SACK_PERM
&& th
->syn
&&
3779 !estab
&& sysctl_tcp_sack
) {
3780 opt_rx
->sack_ok
= TCP_SACK_SEEN
;
3781 tcp_sack_reset(opt_rx
);
3786 if ((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
3787 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
3789 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
3792 #ifdef CONFIG_TCP_MD5SIG
3795 * The MD5 Hash has already been
3796 * checked (see tcp_v{4,6}_do_rcv()).
3801 /* This option is variable length.
3804 case TCPOLEN_COOKIE_BASE
:
3805 /* not yet implemented */
3807 case TCPOLEN_COOKIE_PAIR
:
3808 /* not yet implemented */
3810 case TCPOLEN_COOKIE_MIN
+0:
3811 case TCPOLEN_COOKIE_MIN
+2:
3812 case TCPOLEN_COOKIE_MIN
+4:
3813 case TCPOLEN_COOKIE_MIN
+6:
3814 case TCPOLEN_COOKIE_MAX
:
3815 /* 16-bit multiple */
3816 opt_rx
->cookie_plus
= opsize
;
3826 /* Fast Open option shares code 254 using a
3827 * 16 bits magic number. It's valid only in
3828 * SYN or SYN-ACK with an even size.
3830 if (opsize
< TCPOLEN_EXP_FASTOPEN_BASE
||
3831 get_unaligned_be16(ptr
) != TCPOPT_FASTOPEN_MAGIC
||
3832 foc
== NULL
|| !th
->syn
|| (opsize
& 1))
3834 foc
->len
= opsize
- TCPOLEN_EXP_FASTOPEN_BASE
;
3835 if (foc
->len
>= TCP_FASTOPEN_COOKIE_MIN
&&
3836 foc
->len
<= TCP_FASTOPEN_COOKIE_MAX
)
3837 memcpy(foc
->val
, ptr
+ 2, foc
->len
);
3838 else if (foc
->len
!= 0)
3848 EXPORT_SYMBOL(tcp_parse_options
);
3850 static bool tcp_parse_aligned_timestamp(struct tcp_sock
*tp
, const struct tcphdr
*th
)
3852 const __be32
*ptr
= (const __be32
*)(th
+ 1);
3854 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3855 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
3856 tp
->rx_opt
.saw_tstamp
= 1;
3858 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3860 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
) - tp
->tsoffset
;
3866 /* Fast parse options. This hopes to only see timestamps.
3867 * If it is wrong it falls back on tcp_parse_options().
3869 static bool tcp_fast_parse_options(const struct sk_buff
*skb
,
3870 const struct tcphdr
*th
,
3871 struct tcp_sock
*tp
, const u8
**hvpp
)
3873 /* In the spirit of fast parsing, compare doff directly to constant
3874 * values. Because equality is used, short doff can be ignored here.
3876 if (th
->doff
== (sizeof(*th
) / 4)) {
3877 tp
->rx_opt
.saw_tstamp
= 0;
3879 } else if (tp
->rx_opt
.tstamp_ok
&&
3880 th
->doff
== ((sizeof(*th
) + TCPOLEN_TSTAMP_ALIGNED
) / 4)) {
3881 if (tcp_parse_aligned_timestamp(tp
, th
))
3885 tcp_parse_options(skb
, &tp
->rx_opt
, hvpp
, 1, NULL
);
3886 if (tp
->rx_opt
.saw_tstamp
)
3887 tp
->rx_opt
.rcv_tsecr
-= tp
->tsoffset
;
3892 #ifdef CONFIG_TCP_MD5SIG
3894 * Parse MD5 Signature option
3896 const u8
*tcp_parse_md5sig_option(const struct tcphdr
*th
)
3898 int length
= (th
->doff
<< 2) - sizeof(*th
);
3899 const u8
*ptr
= (const u8
*)(th
+ 1);
3901 /* If the TCP option is too short, we can short cut */
3902 if (length
< TCPOLEN_MD5SIG
)
3905 while (length
> 0) {
3906 int opcode
= *ptr
++;
3917 if (opsize
< 2 || opsize
> length
)
3919 if (opcode
== TCPOPT_MD5SIG
)
3920 return opsize
== TCPOLEN_MD5SIG
? ptr
: NULL
;
3927 EXPORT_SYMBOL(tcp_parse_md5sig_option
);
3930 static inline void tcp_store_ts_recent(struct tcp_sock
*tp
)
3932 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
3933 tp
->rx_opt
.ts_recent_stamp
= get_seconds();
3936 static inline void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
3938 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
3939 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3940 * extra check below makes sure this can only happen
3941 * for pure ACK frames. -DaveM
3943 * Not only, also it occurs for expired timestamps.
3946 if (tcp_paws_check(&tp
->rx_opt
, 0))
3947 tcp_store_ts_recent(tp
);
3951 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3953 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3954 * it can pass through stack. So, the following predicate verifies that
3955 * this segment is not used for anything but congestion avoidance or
3956 * fast retransmit. Moreover, we even are able to eliminate most of such
3957 * second order effects, if we apply some small "replay" window (~RTO)
3958 * to timestamp space.
3960 * All these measures still do not guarantee that we reject wrapped ACKs
3961 * on networks with high bandwidth, when sequence space is recycled fastly,
3962 * but it guarantees that such events will be very rare and do not affect
3963 * connection seriously. This doesn't look nice, but alas, PAWS is really
3966 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3967 * states that events when retransmit arrives after original data are rare.
3968 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3969 * the biggest problem on large power networks even with minor reordering.
3970 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3971 * up to bandwidth of 18Gigabit/sec. 8) ]
3974 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
3976 const struct tcp_sock
*tp
= tcp_sk(sk
);
3977 const struct tcphdr
*th
= tcp_hdr(skb
);
3978 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3979 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3981 return (/* 1. Pure ACK with correct sequence number. */
3982 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
3984 /* 2. ... and duplicate ACK. */
3985 ack
== tp
->snd_una
&&
3987 /* 3. ... and does not update window. */
3988 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
3990 /* 4. ... and sits in replay window. */
3991 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
3994 static inline bool tcp_paws_discard(const struct sock
*sk
,
3995 const struct sk_buff
*skb
)
3997 const struct tcp_sock
*tp
= tcp_sk(sk
);
3999 return !tcp_paws_check(&tp
->rx_opt
, TCP_PAWS_WINDOW
) &&
4000 !tcp_disordered_ack(sk
, skb
);
4003 /* Check segment sequence number for validity.
4005 * Segment controls are considered valid, if the segment
4006 * fits to the window after truncation to the window. Acceptability
4007 * of data (and SYN, FIN, of course) is checked separately.
4008 * See tcp_data_queue(), for example.
4010 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4011 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4012 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4013 * (borrowed from freebsd)
4016 static inline bool tcp_sequence(const struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
4018 return !before(end_seq
, tp
->rcv_wup
) &&
4019 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
4022 /* When we get a reset we do this. */
4023 void tcp_reset(struct sock
*sk
)
4025 /* We want the right error as BSD sees it (and indeed as we do). */
4026 switch (sk
->sk_state
) {
4028 sk
->sk_err
= ECONNREFUSED
;
4030 case TCP_CLOSE_WAIT
:
4036 sk
->sk_err
= ECONNRESET
;
4038 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4041 if (!sock_flag(sk
, SOCK_DEAD
))
4042 sk
->sk_error_report(sk
);
4048 * Process the FIN bit. This now behaves as it is supposed to work
4049 * and the FIN takes effect when it is validly part of sequence
4050 * space. Not before when we get holes.
4052 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4053 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4056 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4057 * close and we go into CLOSING (and later onto TIME-WAIT)
4059 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4061 static void tcp_fin(struct sock
*sk
)
4063 struct tcp_sock
*tp
= tcp_sk(sk
);
4065 inet_csk_schedule_ack(sk
);
4067 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
4068 sock_set_flag(sk
, SOCK_DONE
);
4070 switch (sk
->sk_state
) {
4072 case TCP_ESTABLISHED
:
4073 /* Move to CLOSE_WAIT */
4074 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
4075 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
4078 case TCP_CLOSE_WAIT
:
4080 /* Received a retransmission of the FIN, do
4085 /* RFC793: Remain in the LAST-ACK state. */
4089 /* This case occurs when a simultaneous close
4090 * happens, we must ack the received FIN and
4091 * enter the CLOSING state.
4094 tcp_set_state(sk
, TCP_CLOSING
);
4097 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4099 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
4102 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4103 * cases we should never reach this piece of code.
4105 pr_err("%s: Impossible, sk->sk_state=%d\n",
4106 __func__
, sk
->sk_state
);
4110 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4111 * Probably, we should reset in this case. For now drop them.
4113 __skb_queue_purge(&tp
->out_of_order_queue
);
4114 if (tcp_is_sack(tp
))
4115 tcp_sack_reset(&tp
->rx_opt
);
4118 if (!sock_flag(sk
, SOCK_DEAD
)) {
4119 sk
->sk_state_change(sk
);
4121 /* Do not send POLL_HUP for half duplex close. */
4122 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
4123 sk
->sk_state
== TCP_CLOSE
)
4124 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_HUP
);
4126 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_IN
);
4130 static inline bool tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
,
4133 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
4134 if (before(seq
, sp
->start_seq
))
4135 sp
->start_seq
= seq
;
4136 if (after(end_seq
, sp
->end_seq
))
4137 sp
->end_seq
= end_seq
;
4143 static void tcp_dsack_set(struct sock
*sk
, u32 seq
, u32 end_seq
)
4145 struct tcp_sock
*tp
= tcp_sk(sk
);
4147 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
4150 if (before(seq
, tp
->rcv_nxt
))
4151 mib_idx
= LINUX_MIB_TCPDSACKOLDSENT
;
4153 mib_idx
= LINUX_MIB_TCPDSACKOFOSENT
;
4155 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
4157 tp
->rx_opt
.dsack
= 1;
4158 tp
->duplicate_sack
[0].start_seq
= seq
;
4159 tp
->duplicate_sack
[0].end_seq
= end_seq
;
4163 static void tcp_dsack_extend(struct sock
*sk
, u32 seq
, u32 end_seq
)
4165 struct tcp_sock
*tp
= tcp_sk(sk
);
4167 if (!tp
->rx_opt
.dsack
)
4168 tcp_dsack_set(sk
, seq
, end_seq
);
4170 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
4173 static void tcp_send_dupack(struct sock
*sk
, const struct sk_buff
*skb
)
4175 struct tcp_sock
*tp
= tcp_sk(sk
);
4177 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
4178 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4179 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4180 tcp_enter_quickack_mode(sk
);
4182 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
4183 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4185 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
4186 end_seq
= tp
->rcv_nxt
;
4187 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, end_seq
);
4194 /* These routines update the SACK block as out-of-order packets arrive or
4195 * in-order packets close up the sequence space.
4197 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
4200 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4201 struct tcp_sack_block
*swalk
= sp
+ 1;
4203 /* See if the recent change to the first SACK eats into
4204 * or hits the sequence space of other SACK blocks, if so coalesce.
4206 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
;) {
4207 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
4210 /* Zap SWALK, by moving every further SACK up by one slot.
4211 * Decrease num_sacks.
4213 tp
->rx_opt
.num_sacks
--;
4214 for (i
= this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
4218 this_sack
++, swalk
++;
4222 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
4224 struct tcp_sock
*tp
= tcp_sk(sk
);
4225 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4226 int cur_sacks
= tp
->rx_opt
.num_sacks
;
4232 for (this_sack
= 0; this_sack
< cur_sacks
; this_sack
++, sp
++) {
4233 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
4234 /* Rotate this_sack to the first one. */
4235 for (; this_sack
> 0; this_sack
--, sp
--)
4236 swap(*sp
, *(sp
- 1));
4238 tcp_sack_maybe_coalesce(tp
);
4243 /* Could not find an adjacent existing SACK, build a new one,
4244 * put it at the front, and shift everyone else down. We
4245 * always know there is at least one SACK present already here.
4247 * If the sack array is full, forget about the last one.
4249 if (this_sack
>= TCP_NUM_SACKS
) {
4251 tp
->rx_opt
.num_sacks
--;
4254 for (; this_sack
> 0; this_sack
--, sp
--)
4258 /* Build the new head SACK, and we're done. */
4259 sp
->start_seq
= seq
;
4260 sp
->end_seq
= end_seq
;
4261 tp
->rx_opt
.num_sacks
++;
4264 /* RCV.NXT advances, some SACKs should be eaten. */
4266 static void tcp_sack_remove(struct tcp_sock
*tp
)
4268 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4269 int num_sacks
= tp
->rx_opt
.num_sacks
;
4272 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4273 if (skb_queue_empty(&tp
->out_of_order_queue
)) {
4274 tp
->rx_opt
.num_sacks
= 0;
4278 for (this_sack
= 0; this_sack
< num_sacks
;) {
4279 /* Check if the start of the sack is covered by RCV.NXT. */
4280 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
4283 /* RCV.NXT must cover all the block! */
4284 WARN_ON(before(tp
->rcv_nxt
, sp
->end_seq
));
4286 /* Zap this SACK, by moving forward any other SACKS. */
4287 for (i
=this_sack
+1; i
< num_sacks
; i
++)
4288 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
4295 tp
->rx_opt
.num_sacks
= num_sacks
;
4298 /* This one checks to see if we can put data from the
4299 * out_of_order queue into the receive_queue.
4301 static void tcp_ofo_queue(struct sock
*sk
)
4303 struct tcp_sock
*tp
= tcp_sk(sk
);
4304 __u32 dsack_high
= tp
->rcv_nxt
;
4305 struct sk_buff
*skb
;
4307 while ((skb
= skb_peek(&tp
->out_of_order_queue
)) != NULL
) {
4308 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
4311 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
4312 __u32 dsack
= dsack_high
;
4313 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
4314 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
4315 tcp_dsack_extend(sk
, TCP_SKB_CB(skb
)->seq
, dsack
);
4318 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4319 SOCK_DEBUG(sk
, "ofo packet was already received\n");
4320 __skb_unlink(skb
, &tp
->out_of_order_queue
);
4324 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
4325 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4326 TCP_SKB_CB(skb
)->end_seq
);
4328 __skb_unlink(skb
, &tp
->out_of_order_queue
);
4329 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4330 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4331 if (tcp_hdr(skb
)->fin
)
4336 static bool tcp_prune_ofo_queue(struct sock
*sk
);
4337 static int tcp_prune_queue(struct sock
*sk
);
4339 static int tcp_try_rmem_schedule(struct sock
*sk
, struct sk_buff
*skb
,
4342 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
4343 !sk_rmem_schedule(sk
, skb
, size
)) {
4345 if (tcp_prune_queue(sk
) < 0)
4348 if (!sk_rmem_schedule(sk
, skb
, size
)) {
4349 if (!tcp_prune_ofo_queue(sk
))
4352 if (!sk_rmem_schedule(sk
, skb
, size
))
4360 * tcp_try_coalesce - try to merge skb to prior one
4363 * @from: buffer to add in queue
4364 * @fragstolen: pointer to boolean
4366 * Before queueing skb @from after @to, try to merge them
4367 * to reduce overall memory use and queue lengths, if cost is small.
4368 * Packets in ofo or receive queues can stay a long time.
4369 * Better try to coalesce them right now to avoid future collapses.
4370 * Returns true if caller should free @from instead of queueing it
4372 static bool tcp_try_coalesce(struct sock
*sk
,
4374 struct sk_buff
*from
,
4379 *fragstolen
= false;
4381 if (tcp_hdr(from
)->fin
)
4384 /* Its possible this segment overlaps with prior segment in queue */
4385 if (TCP_SKB_CB(from
)->seq
!= TCP_SKB_CB(to
)->end_seq
)
4388 if (!skb_try_coalesce(to
, from
, fragstolen
, &delta
))
4391 atomic_add(delta
, &sk
->sk_rmem_alloc
);
4392 sk_mem_charge(sk
, delta
);
4393 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRCVCOALESCE
);
4394 TCP_SKB_CB(to
)->end_seq
= TCP_SKB_CB(from
)->end_seq
;
4395 TCP_SKB_CB(to
)->ack_seq
= TCP_SKB_CB(from
)->ack_seq
;
4399 static void tcp_data_queue_ofo(struct sock
*sk
, struct sk_buff
*skb
)
4401 struct tcp_sock
*tp
= tcp_sk(sk
);
4402 struct sk_buff
*skb1
;
4405 TCP_ECN_check_ce(tp
, skb
);
4407 if (unlikely(tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))) {
4408 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFODROP
);
4413 /* Disable header prediction. */
4415 inet_csk_schedule_ack(sk
);
4417 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFOQUEUE
);
4418 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
4419 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4421 skb1
= skb_peek_tail(&tp
->out_of_order_queue
);
4423 /* Initial out of order segment, build 1 SACK. */
4424 if (tcp_is_sack(tp
)) {
4425 tp
->rx_opt
.num_sacks
= 1;
4426 tp
->selective_acks
[0].start_seq
= TCP_SKB_CB(skb
)->seq
;
4427 tp
->selective_acks
[0].end_seq
=
4428 TCP_SKB_CB(skb
)->end_seq
;
4430 __skb_queue_head(&tp
->out_of_order_queue
, skb
);
4434 seq
= TCP_SKB_CB(skb
)->seq
;
4435 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4437 if (seq
== TCP_SKB_CB(skb1
)->end_seq
) {
4440 if (!tcp_try_coalesce(sk
, skb1
, skb
, &fragstolen
)) {
4441 __skb_queue_after(&tp
->out_of_order_queue
, skb1
, skb
);
4443 kfree_skb_partial(skb
, fragstolen
);
4447 if (!tp
->rx_opt
.num_sacks
||
4448 tp
->selective_acks
[0].end_seq
!= seq
)
4451 /* Common case: data arrive in order after hole. */
4452 tp
->selective_acks
[0].end_seq
= end_seq
;
4456 /* Find place to insert this segment. */
4458 if (!after(TCP_SKB_CB(skb1
)->seq
, seq
))
4460 if (skb_queue_is_first(&tp
->out_of_order_queue
, skb1
)) {
4464 skb1
= skb_queue_prev(&tp
->out_of_order_queue
, skb1
);
4467 /* Do skb overlap to previous one? */
4468 if (skb1
&& before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4469 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4470 /* All the bits are present. Drop. */
4471 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFOMERGE
);
4474 tcp_dsack_set(sk
, seq
, end_seq
);
4477 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
4478 /* Partial overlap. */
4479 tcp_dsack_set(sk
, seq
,
4480 TCP_SKB_CB(skb1
)->end_seq
);
4482 if (skb_queue_is_first(&tp
->out_of_order_queue
,
4486 skb1
= skb_queue_prev(
4487 &tp
->out_of_order_queue
,
4492 __skb_queue_head(&tp
->out_of_order_queue
, skb
);
4494 __skb_queue_after(&tp
->out_of_order_queue
, skb1
, skb
);
4496 /* And clean segments covered by new one as whole. */
4497 while (!skb_queue_is_last(&tp
->out_of_order_queue
, skb
)) {
4498 skb1
= skb_queue_next(&tp
->out_of_order_queue
, skb
);
4500 if (!after(end_seq
, TCP_SKB_CB(skb1
)->seq
))
4502 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4503 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4507 __skb_unlink(skb1
, &tp
->out_of_order_queue
);
4508 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4509 TCP_SKB_CB(skb1
)->end_seq
);
4510 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFOMERGE
);
4515 if (tcp_is_sack(tp
))
4516 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
4519 skb_set_owner_r(skb
, sk
);
4522 static int __must_check
tcp_queue_rcv(struct sock
*sk
, struct sk_buff
*skb
, int hdrlen
,
4526 struct sk_buff
*tail
= skb_peek_tail(&sk
->sk_receive_queue
);
4528 __skb_pull(skb
, hdrlen
);
4530 tcp_try_coalesce(sk
, tail
, skb
, fragstolen
)) ? 1 : 0;
4531 tcp_sk(sk
)->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4533 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4534 skb_set_owner_r(skb
, sk
);
4539 int tcp_send_rcvq(struct sock
*sk
, struct msghdr
*msg
, size_t size
)
4541 struct sk_buff
*skb
= NULL
;
4548 skb
= alloc_skb(size
+ sizeof(*th
), sk
->sk_allocation
);
4552 if (tcp_try_rmem_schedule(sk
, skb
, size
+ sizeof(*th
)))
4555 th
= (struct tcphdr
*)skb_put(skb
, sizeof(*th
));
4556 skb_reset_transport_header(skb
);
4557 memset(th
, 0, sizeof(*th
));
4559 if (memcpy_fromiovec(skb_put(skb
, size
), msg
->msg_iov
, size
))
4562 TCP_SKB_CB(skb
)->seq
= tcp_sk(sk
)->rcv_nxt
;
4563 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ size
;
4564 TCP_SKB_CB(skb
)->ack_seq
= tcp_sk(sk
)->snd_una
- 1;
4566 if (tcp_queue_rcv(sk
, skb
, sizeof(*th
), &fragstolen
)) {
4567 WARN_ON_ONCE(fragstolen
); /* should not happen */
4578 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
4580 const struct tcphdr
*th
= tcp_hdr(skb
);
4581 struct tcp_sock
*tp
= tcp_sk(sk
);
4583 bool fragstolen
= false;
4585 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
)
4589 __skb_pull(skb
, th
->doff
* 4);
4591 TCP_ECN_accept_cwr(tp
, skb
);
4593 tp
->rx_opt
.dsack
= 0;
4595 /* Queue data for delivery to the user.
4596 * Packets in sequence go to the receive queue.
4597 * Out of sequence packets to the out_of_order_queue.
4599 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
4600 if (tcp_receive_window(tp
) == 0)
4603 /* Ok. In sequence. In window. */
4604 if (tp
->ucopy
.task
== current
&&
4605 tp
->copied_seq
== tp
->rcv_nxt
&& tp
->ucopy
.len
&&
4606 sock_owned_by_user(sk
) && !tp
->urg_data
) {
4607 int chunk
= min_t(unsigned int, skb
->len
,
4610 __set_current_state(TASK_RUNNING
);
4613 if (!skb_copy_datagram_iovec(skb
, 0, tp
->ucopy
.iov
, chunk
)) {
4614 tp
->ucopy
.len
-= chunk
;
4615 tp
->copied_seq
+= chunk
;
4616 eaten
= (chunk
== skb
->len
);
4617 tcp_rcv_space_adjust(sk
);
4625 tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))
4628 eaten
= tcp_queue_rcv(sk
, skb
, 0, &fragstolen
);
4630 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4632 tcp_event_data_recv(sk
, skb
);
4636 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4639 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4640 * gap in queue is filled.
4642 if (skb_queue_empty(&tp
->out_of_order_queue
))
4643 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
4646 if (tp
->rx_opt
.num_sacks
)
4647 tcp_sack_remove(tp
);
4649 tcp_fast_path_check(sk
);
4652 kfree_skb_partial(skb
, fragstolen
);
4653 if (!sock_flag(sk
, SOCK_DEAD
))
4654 sk
->sk_data_ready(sk
, 0);
4658 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4659 /* A retransmit, 2nd most common case. Force an immediate ack. */
4660 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4661 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4664 tcp_enter_quickack_mode(sk
);
4665 inet_csk_schedule_ack(sk
);
4671 /* Out of window. F.e. zero window probe. */
4672 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
4675 tcp_enter_quickack_mode(sk
);
4677 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4678 /* Partial packet, seq < rcv_next < end_seq */
4679 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
4680 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4681 TCP_SKB_CB(skb
)->end_seq
);
4683 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
4685 /* If window is closed, drop tail of packet. But after
4686 * remembering D-SACK for its head made in previous line.
4688 if (!tcp_receive_window(tp
))
4693 tcp_data_queue_ofo(sk
, skb
);
4696 static struct sk_buff
*tcp_collapse_one(struct sock
*sk
, struct sk_buff
*skb
,
4697 struct sk_buff_head
*list
)
4699 struct sk_buff
*next
= NULL
;
4701 if (!skb_queue_is_last(list
, skb
))
4702 next
= skb_queue_next(list
, skb
);
4704 __skb_unlink(skb
, list
);
4706 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRCVCOLLAPSED
);
4711 /* Collapse contiguous sequence of skbs head..tail with
4712 * sequence numbers start..end.
4714 * If tail is NULL, this means until the end of the list.
4716 * Segments with FIN/SYN are not collapsed (only because this
4720 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
,
4721 struct sk_buff
*head
, struct sk_buff
*tail
,
4724 struct sk_buff
*skb
, *n
;
4727 /* First, check that queue is collapsible and find
4728 * the point where collapsing can be useful. */
4732 skb_queue_walk_from_safe(list
, skb
, n
) {
4735 /* No new bits? It is possible on ofo queue. */
4736 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4737 skb
= tcp_collapse_one(sk
, skb
, list
);
4743 /* The first skb to collapse is:
4745 * - bloated or contains data before "start" or
4746 * overlaps to the next one.
4748 if (!tcp_hdr(skb
)->syn
&& !tcp_hdr(skb
)->fin
&&
4749 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
4750 before(TCP_SKB_CB(skb
)->seq
, start
))) {
4751 end_of_skbs
= false;
4755 if (!skb_queue_is_last(list
, skb
)) {
4756 struct sk_buff
*next
= skb_queue_next(list
, skb
);
4758 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(next
)->seq
) {
4759 end_of_skbs
= false;
4764 /* Decided to skip this, advance start seq. */
4765 start
= TCP_SKB_CB(skb
)->end_seq
;
4767 if (end_of_skbs
|| tcp_hdr(skb
)->syn
|| tcp_hdr(skb
)->fin
)
4770 while (before(start
, end
)) {
4771 struct sk_buff
*nskb
;
4772 unsigned int header
= skb_headroom(skb
);
4773 int copy
= SKB_MAX_ORDER(header
, 0);
4775 /* Too big header? This can happen with IPv6. */
4778 if (end
- start
< copy
)
4780 nskb
= alloc_skb(copy
+ header
, GFP_ATOMIC
);
4784 skb_set_mac_header(nskb
, skb_mac_header(skb
) - skb
->head
);
4785 skb_set_network_header(nskb
, (skb_network_header(skb
) -
4787 skb_set_transport_header(nskb
, (skb_transport_header(skb
) -
4789 skb_reserve(nskb
, header
);
4790 memcpy(nskb
->head
, skb
->head
, header
);
4791 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
4792 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
4793 __skb_queue_before(list
, skb
, nskb
);
4794 skb_set_owner_r(nskb
, sk
);
4796 /* Copy data, releasing collapsed skbs. */
4798 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
4799 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
4803 size
= min(copy
, size
);
4804 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
4806 TCP_SKB_CB(nskb
)->end_seq
+= size
;
4810 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4811 skb
= tcp_collapse_one(sk
, skb
, list
);
4814 tcp_hdr(skb
)->syn
||
4822 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4823 * and tcp_collapse() them until all the queue is collapsed.
4825 static void tcp_collapse_ofo_queue(struct sock
*sk
)
4827 struct tcp_sock
*tp
= tcp_sk(sk
);
4828 struct sk_buff
*skb
= skb_peek(&tp
->out_of_order_queue
);
4829 struct sk_buff
*head
;
4835 start
= TCP_SKB_CB(skb
)->seq
;
4836 end
= TCP_SKB_CB(skb
)->end_seq
;
4840 struct sk_buff
*next
= NULL
;
4842 if (!skb_queue_is_last(&tp
->out_of_order_queue
, skb
))
4843 next
= skb_queue_next(&tp
->out_of_order_queue
, skb
);
4846 /* Segment is terminated when we see gap or when
4847 * we are at the end of all the queue. */
4849 after(TCP_SKB_CB(skb
)->seq
, end
) ||
4850 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
4851 tcp_collapse(sk
, &tp
->out_of_order_queue
,
4852 head
, skb
, start
, end
);
4856 /* Start new segment */
4857 start
= TCP_SKB_CB(skb
)->seq
;
4858 end
= TCP_SKB_CB(skb
)->end_seq
;
4860 if (before(TCP_SKB_CB(skb
)->seq
, start
))
4861 start
= TCP_SKB_CB(skb
)->seq
;
4862 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
4863 end
= TCP_SKB_CB(skb
)->end_seq
;
4869 * Purge the out-of-order queue.
4870 * Return true if queue was pruned.
4872 static bool tcp_prune_ofo_queue(struct sock
*sk
)
4874 struct tcp_sock
*tp
= tcp_sk(sk
);
4877 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4878 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_OFOPRUNED
);
4879 __skb_queue_purge(&tp
->out_of_order_queue
);
4881 /* Reset SACK state. A conforming SACK implementation will
4882 * do the same at a timeout based retransmit. When a connection
4883 * is in a sad state like this, we care only about integrity
4884 * of the connection not performance.
4886 if (tp
->rx_opt
.sack_ok
)
4887 tcp_sack_reset(&tp
->rx_opt
);
4894 /* Reduce allocated memory if we can, trying to get
4895 * the socket within its memory limits again.
4897 * Return less than zero if we should start dropping frames
4898 * until the socket owning process reads some of the data
4899 * to stabilize the situation.
4901 static int tcp_prune_queue(struct sock
*sk
)
4903 struct tcp_sock
*tp
= tcp_sk(sk
);
4905 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
4907 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PRUNECALLED
);
4909 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
4910 tcp_clamp_window(sk
);
4911 else if (sk_under_memory_pressure(sk
))
4912 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
4914 tcp_collapse_ofo_queue(sk
);
4915 if (!skb_queue_empty(&sk
->sk_receive_queue
))
4916 tcp_collapse(sk
, &sk
->sk_receive_queue
,
4917 skb_peek(&sk
->sk_receive_queue
),
4919 tp
->copied_seq
, tp
->rcv_nxt
);
4922 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4925 /* Collapsing did not help, destructive actions follow.
4926 * This must not ever occur. */
4928 tcp_prune_ofo_queue(sk
);
4930 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4933 /* If we are really being abused, tell the caller to silently
4934 * drop receive data on the floor. It will get retransmitted
4935 * and hopefully then we'll have sufficient space.
4937 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_RCVPRUNED
);
4939 /* Massive buffer overcommit. */
4944 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4945 * As additional protections, we do not touch cwnd in retransmission phases,
4946 * and if application hit its sndbuf limit recently.
4948 void tcp_cwnd_application_limited(struct sock
*sk
)
4950 struct tcp_sock
*tp
= tcp_sk(sk
);
4952 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
4953 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
4954 /* Limited by application or receiver window. */
4955 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
4956 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
4957 if (win_used
< tp
->snd_cwnd
) {
4958 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
4959 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
4961 tp
->snd_cwnd_used
= 0;
4963 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4966 static bool tcp_should_expand_sndbuf(const struct sock
*sk
)
4968 const struct tcp_sock
*tp
= tcp_sk(sk
);
4970 /* If the user specified a specific send buffer setting, do
4973 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
4976 /* If we are under global TCP memory pressure, do not expand. */
4977 if (sk_under_memory_pressure(sk
))
4980 /* If we are under soft global TCP memory pressure, do not expand. */
4981 if (sk_memory_allocated(sk
) >= sk_prot_mem_limits(sk
, 0))
4984 /* If we filled the congestion window, do not expand. */
4985 if (tp
->packets_out
>= tp
->snd_cwnd
)
4991 /* When incoming ACK allowed to free some skb from write_queue,
4992 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4993 * on the exit from tcp input handler.
4995 * PROBLEM: sndbuf expansion does not work well with largesend.
4997 static void tcp_new_space(struct sock
*sk
)
4999 struct tcp_sock
*tp
= tcp_sk(sk
);
5001 if (tcp_should_expand_sndbuf(sk
)) {
5002 int sndmem
= SKB_TRUESIZE(max_t(u32
,
5003 tp
->rx_opt
.mss_clamp
,
5006 int demanded
= max_t(unsigned int, tp
->snd_cwnd
,
5007 tp
->reordering
+ 1);
5008 sndmem
*= 2 * demanded
;
5009 if (sndmem
> sk
->sk_sndbuf
)
5010 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
5011 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
5014 sk
->sk_write_space(sk
);
5017 static void tcp_check_space(struct sock
*sk
)
5019 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
5020 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
5021 if (sk
->sk_socket
&&
5022 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
5027 static inline void tcp_data_snd_check(struct sock
*sk
)
5029 tcp_push_pending_frames(sk
);
5030 tcp_check_space(sk
);
5034 * Check if sending an ack is needed.
5036 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
5038 struct tcp_sock
*tp
= tcp_sk(sk
);
5040 /* More than one full frame received... */
5041 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
&&
5042 /* ... and right edge of window advances far enough.
5043 * (tcp_recvmsg() will send ACK otherwise). Or...
5045 __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
5046 /* We ACK each frame or... */
5047 tcp_in_quickack_mode(sk
) ||
5048 /* We have out of order data. */
5049 (ofo_possible
&& skb_peek(&tp
->out_of_order_queue
))) {
5050 /* Then ack it now */
5053 /* Else, send delayed ack. */
5054 tcp_send_delayed_ack(sk
);
5058 static inline void tcp_ack_snd_check(struct sock
*sk
)
5060 if (!inet_csk_ack_scheduled(sk
)) {
5061 /* We sent a data segment already. */
5064 __tcp_ack_snd_check(sk
, 1);
5068 * This routine is only called when we have urgent data
5069 * signaled. Its the 'slow' part of tcp_urg. It could be
5070 * moved inline now as tcp_urg is only called from one
5071 * place. We handle URGent data wrong. We have to - as
5072 * BSD still doesn't use the correction from RFC961.
5073 * For 1003.1g we should support a new option TCP_STDURG to permit
5074 * either form (or just set the sysctl tcp_stdurg).
5077 static void tcp_check_urg(struct sock
*sk
, const struct tcphdr
*th
)
5079 struct tcp_sock
*tp
= tcp_sk(sk
);
5080 u32 ptr
= ntohs(th
->urg_ptr
);
5082 if (ptr
&& !sysctl_tcp_stdurg
)
5084 ptr
+= ntohl(th
->seq
);
5086 /* Ignore urgent data that we've already seen and read. */
5087 if (after(tp
->copied_seq
, ptr
))
5090 /* Do not replay urg ptr.
5092 * NOTE: interesting situation not covered by specs.
5093 * Misbehaving sender may send urg ptr, pointing to segment,
5094 * which we already have in ofo queue. We are not able to fetch
5095 * such data and will stay in TCP_URG_NOTYET until will be eaten
5096 * by recvmsg(). Seems, we are not obliged to handle such wicked
5097 * situations. But it is worth to think about possibility of some
5098 * DoSes using some hypothetical application level deadlock.
5100 if (before(ptr
, tp
->rcv_nxt
))
5103 /* Do we already have a newer (or duplicate) urgent pointer? */
5104 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
5107 /* Tell the world about our new urgent pointer. */
5110 /* We may be adding urgent data when the last byte read was
5111 * urgent. To do this requires some care. We cannot just ignore
5112 * tp->copied_seq since we would read the last urgent byte again
5113 * as data, nor can we alter copied_seq until this data arrives
5114 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5116 * NOTE. Double Dutch. Rendering to plain English: author of comment
5117 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5118 * and expect that both A and B disappear from stream. This is _wrong_.
5119 * Though this happens in BSD with high probability, this is occasional.
5120 * Any application relying on this is buggy. Note also, that fix "works"
5121 * only in this artificial test. Insert some normal data between A and B and we will
5122 * decline of BSD again. Verdict: it is better to remove to trap
5125 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
5126 !sock_flag(sk
, SOCK_URGINLINE
) && tp
->copied_seq
!= tp
->rcv_nxt
) {
5127 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
5129 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5130 __skb_unlink(skb
, &sk
->sk_receive_queue
);
5135 tp
->urg_data
= TCP_URG_NOTYET
;
5138 /* Disable header prediction. */
5142 /* This is the 'fast' part of urgent handling. */
5143 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, const struct tcphdr
*th
)
5145 struct tcp_sock
*tp
= tcp_sk(sk
);
5147 /* Check if we get a new urgent pointer - normally not. */
5149 tcp_check_urg(sk
, th
);
5151 /* Do we wait for any urgent data? - normally not... */
5152 if (tp
->urg_data
== TCP_URG_NOTYET
) {
5153 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
5156 /* Is the urgent pointer pointing into this packet? */
5157 if (ptr
< skb
->len
) {
5159 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
5161 tp
->urg_data
= TCP_URG_VALID
| tmp
;
5162 if (!sock_flag(sk
, SOCK_DEAD
))
5163 sk
->sk_data_ready(sk
, 0);
5168 static int tcp_copy_to_iovec(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
5170 struct tcp_sock
*tp
= tcp_sk(sk
);
5171 int chunk
= skb
->len
- hlen
;
5175 if (skb_csum_unnecessary(skb
))
5176 err
= skb_copy_datagram_iovec(skb
, hlen
, tp
->ucopy
.iov
, chunk
);
5178 err
= skb_copy_and_csum_datagram_iovec(skb
, hlen
,
5182 tp
->ucopy
.len
-= chunk
;
5183 tp
->copied_seq
+= chunk
;
5184 tcp_rcv_space_adjust(sk
);
5191 static __sum16
__tcp_checksum_complete_user(struct sock
*sk
,
5192 struct sk_buff
*skb
)
5196 if (sock_owned_by_user(sk
)) {
5198 result
= __tcp_checksum_complete(skb
);
5201 result
= __tcp_checksum_complete(skb
);
5206 static inline bool tcp_checksum_complete_user(struct sock
*sk
,
5207 struct sk_buff
*skb
)
5209 return !skb_csum_unnecessary(skb
) &&
5210 __tcp_checksum_complete_user(sk
, skb
);
5213 #ifdef CONFIG_NET_DMA
5214 static bool tcp_dma_try_early_copy(struct sock
*sk
, struct sk_buff
*skb
,
5217 struct tcp_sock
*tp
= tcp_sk(sk
);
5218 int chunk
= skb
->len
- hlen
;
5220 bool copied_early
= false;
5222 if (tp
->ucopy
.wakeup
)
5225 if (!tp
->ucopy
.dma_chan
&& tp
->ucopy
.pinned_list
)
5226 tp
->ucopy
.dma_chan
= net_dma_find_channel();
5228 if (tp
->ucopy
.dma_chan
&& skb_csum_unnecessary(skb
)) {
5230 dma_cookie
= dma_skb_copy_datagram_iovec(tp
->ucopy
.dma_chan
,
5232 tp
->ucopy
.iov
, chunk
,
5233 tp
->ucopy
.pinned_list
);
5238 tp
->ucopy
.dma_cookie
= dma_cookie
;
5239 copied_early
= true;
5241 tp
->ucopy
.len
-= chunk
;
5242 tp
->copied_seq
+= chunk
;
5243 tcp_rcv_space_adjust(sk
);
5245 if ((tp
->ucopy
.len
== 0) ||
5246 (tcp_flag_word(tcp_hdr(skb
)) & TCP_FLAG_PSH
) ||
5247 (atomic_read(&sk
->sk_rmem_alloc
) > (sk
->sk_rcvbuf
>> 1))) {
5248 tp
->ucopy
.wakeup
= 1;
5249 sk
->sk_data_ready(sk
, 0);
5251 } else if (chunk
> 0) {
5252 tp
->ucopy
.wakeup
= 1;
5253 sk
->sk_data_ready(sk
, 0);
5256 return copied_early
;
5258 #endif /* CONFIG_NET_DMA */
5260 /* Does PAWS and seqno based validation of an incoming segment, flags will
5261 * play significant role here.
5263 static bool tcp_validate_incoming(struct sock
*sk
, struct sk_buff
*skb
,
5264 const struct tcphdr
*th
, int syn_inerr
)
5266 const u8
*hash_location
;
5267 struct tcp_sock
*tp
= tcp_sk(sk
);
5269 /* RFC1323: H1. Apply PAWS check first. */
5270 if (tcp_fast_parse_options(skb
, th
, tp
, &hash_location
) &&
5271 tp
->rx_opt
.saw_tstamp
&&
5272 tcp_paws_discard(sk
, skb
)) {
5274 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSESTABREJECTED
);
5275 tcp_send_dupack(sk
, skb
);
5278 /* Reset is accepted even if it did not pass PAWS. */
5281 /* Step 1: check sequence number */
5282 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5283 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5284 * (RST) segments are validated by checking their SEQ-fields."
5285 * And page 69: "If an incoming segment is not acceptable,
5286 * an acknowledgment should be sent in reply (unless the RST
5287 * bit is set, if so drop the segment and return)".
5292 tcp_send_dupack(sk
, skb
);
5297 /* Step 2: check RST bit */
5300 * If sequence number exactly matches RCV.NXT, then
5301 * RESET the connection
5303 * Send a challenge ACK
5305 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
)
5308 tcp_send_challenge_ack(sk
);
5312 /* step 3: check security and precedence [ignored] */
5314 /* step 4: Check for a SYN
5315 * RFC 5691 4.2 : Send a challenge ack
5320 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5321 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSYNCHALLENGE
);
5322 tcp_send_challenge_ack(sk
);
5334 * TCP receive function for the ESTABLISHED state.
5336 * It is split into a fast path and a slow path. The fast path is
5338 * - A zero window was announced from us - zero window probing
5339 * is only handled properly in the slow path.
5340 * - Out of order segments arrived.
5341 * - Urgent data is expected.
5342 * - There is no buffer space left
5343 * - Unexpected TCP flags/window values/header lengths are received
5344 * (detected by checking the TCP header against pred_flags)
5345 * - Data is sent in both directions. Fast path only supports pure senders
5346 * or pure receivers (this means either the sequence number or the ack
5347 * value must stay constant)
5348 * - Unexpected TCP option.
5350 * When these conditions are not satisfied it drops into a standard
5351 * receive procedure patterned after RFC793 to handle all cases.
5352 * The first three cases are guaranteed by proper pred_flags setting,
5353 * the rest is checked inline. Fast processing is turned on in
5354 * tcp_data_queue when everything is OK.
5356 int tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
5357 const struct tcphdr
*th
, unsigned int len
)
5359 struct tcp_sock
*tp
= tcp_sk(sk
);
5361 if (unlikely(sk
->sk_rx_dst
== NULL
))
5362 inet_csk(sk
)->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
5364 * Header prediction.
5365 * The code loosely follows the one in the famous
5366 * "30 instruction TCP receive" Van Jacobson mail.
5368 * Van's trick is to deposit buffers into socket queue
5369 * on a device interrupt, to call tcp_recv function
5370 * on the receive process context and checksum and copy
5371 * the buffer to user space. smart...
5373 * Our current scheme is not silly either but we take the
5374 * extra cost of the net_bh soft interrupt processing...
5375 * We do checksum and copy also but from device to kernel.
5378 tp
->rx_opt
.saw_tstamp
= 0;
5380 /* pred_flags is 0xS?10 << 16 + snd_wnd
5381 * if header_prediction is to be made
5382 * 'S' will always be tp->tcp_header_len >> 2
5383 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5384 * turn it off (when there are holes in the receive
5385 * space for instance)
5386 * PSH flag is ignored.
5389 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
5390 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
&&
5391 !after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
)) {
5392 int tcp_header_len
= tp
->tcp_header_len
;
5394 /* Timestamp header prediction: tcp_header_len
5395 * is automatically equal to th->doff*4 due to pred_flags
5399 /* Check timestamp */
5400 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
5401 /* No? Slow path! */
5402 if (!tcp_parse_aligned_timestamp(tp
, th
))
5405 /* If PAWS failed, check it more carefully in slow path */
5406 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
5409 /* DO NOT update ts_recent here, if checksum fails
5410 * and timestamp was corrupted part, it will result
5411 * in a hung connection since we will drop all
5412 * future packets due to the PAWS test.
5416 if (len
<= tcp_header_len
) {
5417 /* Bulk data transfer: sender */
5418 if (len
== tcp_header_len
) {
5419 /* Predicted packet is in window by definition.
5420 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5421 * Hence, check seq<=rcv_wup reduces to:
5423 if (tcp_header_len
==
5424 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5425 tp
->rcv_nxt
== tp
->rcv_wup
)
5426 tcp_store_ts_recent(tp
);
5428 /* We know that such packets are checksummed
5431 tcp_ack(sk
, skb
, 0);
5433 tcp_data_snd_check(sk
);
5435 } else { /* Header too small */
5436 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5441 int copied_early
= 0;
5442 bool fragstolen
= false;
5444 if (tp
->copied_seq
== tp
->rcv_nxt
&&
5445 len
- tcp_header_len
<= tp
->ucopy
.len
) {
5446 #ifdef CONFIG_NET_DMA
5447 if (tp
->ucopy
.task
== current
&&
5448 sock_owned_by_user(sk
) &&
5449 tcp_dma_try_early_copy(sk
, skb
, tcp_header_len
)) {
5454 if (tp
->ucopy
.task
== current
&&
5455 sock_owned_by_user(sk
) && !copied_early
) {
5456 __set_current_state(TASK_RUNNING
);
5458 if (!tcp_copy_to_iovec(sk
, skb
, tcp_header_len
))
5462 /* Predicted packet is in window by definition.
5463 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5464 * Hence, check seq<=rcv_wup reduces to:
5466 if (tcp_header_len
==
5467 (sizeof(struct tcphdr
) +
5468 TCPOLEN_TSTAMP_ALIGNED
) &&
5469 tp
->rcv_nxt
== tp
->rcv_wup
)
5470 tcp_store_ts_recent(tp
);
5472 tcp_rcv_rtt_measure_ts(sk
, skb
);
5474 __skb_pull(skb
, tcp_header_len
);
5475 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
5476 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITSTOUSER
);
5479 tcp_cleanup_rbuf(sk
, skb
->len
);
5482 if (tcp_checksum_complete_user(sk
, skb
))
5485 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
5488 /* Predicted packet is in window by definition.
5489 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5490 * Hence, check seq<=rcv_wup reduces to:
5492 if (tcp_header_len
==
5493 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5494 tp
->rcv_nxt
== tp
->rcv_wup
)
5495 tcp_store_ts_recent(tp
);
5497 tcp_rcv_rtt_measure_ts(sk
, skb
);
5499 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITS
);
5501 /* Bulk data transfer: receiver */
5502 eaten
= tcp_queue_rcv(sk
, skb
, tcp_header_len
,
5506 tcp_event_data_recv(sk
, skb
);
5508 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
5509 /* Well, only one small jumplet in fast path... */
5510 tcp_ack(sk
, skb
, FLAG_DATA
);
5511 tcp_data_snd_check(sk
);
5512 if (!inet_csk_ack_scheduled(sk
))
5516 if (!copied_early
|| tp
->rcv_nxt
!= tp
->rcv_wup
)
5517 __tcp_ack_snd_check(sk
, 0);
5519 #ifdef CONFIG_NET_DMA
5521 __skb_queue_tail(&sk
->sk_async_wait_queue
, skb
);
5525 kfree_skb_partial(skb
, fragstolen
);
5526 sk
->sk_data_ready(sk
, 0);
5532 if (len
< (th
->doff
<< 2) || tcp_checksum_complete_user(sk
, skb
))
5535 if (!th
->ack
&& !th
->rst
)
5539 * Standard slow path.
5542 if (!tcp_validate_incoming(sk
, skb
, th
, 1))
5546 if (tcp_ack(sk
, skb
, FLAG_SLOWPATH
) < 0)
5549 /* ts_recent update must be made after we are sure that the packet
5552 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
5554 tcp_rcv_rtt_measure_ts(sk
, skb
);
5556 /* Process urgent data. */
5557 tcp_urg(sk
, skb
, th
);
5559 /* step 7: process the segment text */
5560 tcp_data_queue(sk
, skb
);
5562 tcp_data_snd_check(sk
);
5563 tcp_ack_snd_check(sk
);
5567 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5573 EXPORT_SYMBOL(tcp_rcv_established
);
5575 void tcp_finish_connect(struct sock
*sk
, struct sk_buff
*skb
)
5577 struct tcp_sock
*tp
= tcp_sk(sk
);
5578 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5580 tcp_set_state(sk
, TCP_ESTABLISHED
);
5583 icsk
->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
5584 security_inet_conn_established(sk
, skb
);
5587 /* Make sure socket is routed, for correct metrics. */
5588 icsk
->icsk_af_ops
->rebuild_header(sk
);
5590 tcp_init_metrics(sk
);
5592 tcp_init_congestion_control(sk
);
5594 /* Prevent spurious tcp_cwnd_restart() on first data
5597 tp
->lsndtime
= tcp_time_stamp
;
5599 tcp_init_buffer_space(sk
);
5601 if (sock_flag(sk
, SOCK_KEEPOPEN
))
5602 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
5604 if (!tp
->rx_opt
.snd_wscale
)
5605 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
5609 if (!sock_flag(sk
, SOCK_DEAD
)) {
5610 sk
->sk_state_change(sk
);
5611 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5615 static bool tcp_rcv_fastopen_synack(struct sock
*sk
, struct sk_buff
*synack
,
5616 struct tcp_fastopen_cookie
*cookie
)
5618 struct tcp_sock
*tp
= tcp_sk(sk
);
5619 struct sk_buff
*data
= tp
->syn_data
? tcp_write_queue_head(sk
) : NULL
;
5620 u16 mss
= tp
->rx_opt
.mss_clamp
;
5623 if (mss
== tp
->rx_opt
.user_mss
) {
5624 struct tcp_options_received opt
;
5625 const u8
*hash_location
;
5627 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5628 tcp_clear_options(&opt
);
5629 opt
.user_mss
= opt
.mss_clamp
= 0;
5630 tcp_parse_options(synack
, &opt
, &hash_location
, 0, NULL
);
5631 mss
= opt
.mss_clamp
;
5634 if (!tp
->syn_fastopen
) /* Ignore an unsolicited cookie */
5637 /* The SYN-ACK neither has cookie nor acknowledges the data. Presumably
5638 * the remote receives only the retransmitted (regular) SYNs: either
5639 * the original SYN-data or the corresponding SYN-ACK is lost.
5641 syn_drop
= (cookie
->len
<= 0 && data
&& tp
->total_retrans
);
5643 tcp_fastopen_cache_set(sk
, mss
, cookie
, syn_drop
);
5645 if (data
) { /* Retransmit unacked data in SYN */
5646 tcp_for_write_queue_from(data
, sk
) {
5647 if (data
== tcp_send_head(sk
) ||
5648 __tcp_retransmit_skb(sk
, data
))
5654 tp
->syn_data_acked
= tp
->syn_data
;
5658 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5659 const struct tcphdr
*th
, unsigned int len
)
5661 const u8
*hash_location
;
5662 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5663 struct tcp_sock
*tp
= tcp_sk(sk
);
5664 struct tcp_cookie_values
*cvp
= tp
->cookie_values
;
5665 struct tcp_fastopen_cookie foc
= { .len
= -1 };
5666 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
5668 tcp_parse_options(skb
, &tp
->rx_opt
, &hash_location
, 0, &foc
);
5669 if (tp
->rx_opt
.saw_tstamp
)
5670 tp
->rx_opt
.rcv_tsecr
-= tp
->tsoffset
;
5674 * "If the state is SYN-SENT then
5675 * first check the ACK bit
5676 * If the ACK bit is set
5677 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5678 * a reset (unless the RST bit is set, if so drop
5679 * the segment and return)"
5681 if (!after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_una
) ||
5682 after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
))
5683 goto reset_and_undo
;
5685 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
5686 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
5688 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSACTIVEREJECTED
);
5689 goto reset_and_undo
;
5692 /* Now ACK is acceptable.
5694 * "If the RST bit is set
5695 * If the ACK was acceptable then signal the user "error:
5696 * connection reset", drop the segment, enter CLOSED state,
5697 * delete TCB, and return."
5706 * "fifth, if neither of the SYN or RST bits is set then
5707 * drop the segment and return."
5713 goto discard_and_undo
;
5716 * "If the SYN bit is on ...
5717 * are acceptable then ...
5718 * (our SYN has been ACKed), change the connection
5719 * state to ESTABLISHED..."
5722 TCP_ECN_rcv_synack(tp
, th
);
5724 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5725 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
5727 /* Ok.. it's good. Set up sequence numbers and
5728 * move to established.
5730 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5731 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5733 /* RFC1323: The window in SYN & SYN/ACK segments is
5736 tp
->snd_wnd
= ntohs(th
->window
);
5738 if (!tp
->rx_opt
.wscale_ok
) {
5739 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
5740 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
5743 if (tp
->rx_opt
.saw_tstamp
) {
5744 tp
->rx_opt
.tstamp_ok
= 1;
5745 tp
->tcp_header_len
=
5746 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5747 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5748 tcp_store_ts_recent(tp
);
5750 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5753 if (tcp_is_sack(tp
) && sysctl_tcp_fack
)
5754 tcp_enable_fack(tp
);
5757 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5758 tcp_initialize_rcv_mss(sk
);
5760 /* Remember, tcp_poll() does not lock socket!
5761 * Change state from SYN-SENT only after copied_seq
5762 * is initialized. */
5763 tp
->copied_seq
= tp
->rcv_nxt
;
5766 cvp
->cookie_pair_size
> 0 &&
5767 tp
->rx_opt
.cookie_plus
> 0) {
5768 int cookie_size
= tp
->rx_opt
.cookie_plus
5769 - TCPOLEN_COOKIE_BASE
;
5770 int cookie_pair_size
= cookie_size
5771 + cvp
->cookie_desired
;
5773 /* A cookie extension option was sent and returned.
5774 * Note that each incoming SYNACK replaces the
5775 * Responder cookie. The initial exchange is most
5776 * fragile, as protection against spoofing relies
5777 * entirely upon the sequence and timestamp (above).
5778 * This replacement strategy allows the correct pair to
5779 * pass through, while any others will be filtered via
5780 * Responder verification later.
5782 if (sizeof(cvp
->cookie_pair
) >= cookie_pair_size
) {
5783 memcpy(&cvp
->cookie_pair
[cvp
->cookie_desired
],
5784 hash_location
, cookie_size
);
5785 cvp
->cookie_pair_size
= cookie_pair_size
;
5791 tcp_finish_connect(sk
, skb
);
5793 if ((tp
->syn_fastopen
|| tp
->syn_data
) &&
5794 tcp_rcv_fastopen_synack(sk
, skb
, &foc
))
5797 if (sk
->sk_write_pending
||
5798 icsk
->icsk_accept_queue
.rskq_defer_accept
||
5799 icsk
->icsk_ack
.pingpong
) {
5800 /* Save one ACK. Data will be ready after
5801 * several ticks, if write_pending is set.
5803 * It may be deleted, but with this feature tcpdumps
5804 * look so _wonderfully_ clever, that I was not able
5805 * to stand against the temptation 8) --ANK
5807 inet_csk_schedule_ack(sk
);
5808 icsk
->icsk_ack
.lrcvtime
= tcp_time_stamp
;
5809 tcp_enter_quickack_mode(sk
);
5810 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
5811 TCP_DELACK_MAX
, TCP_RTO_MAX
);
5822 /* No ACK in the segment */
5826 * "If the RST bit is set
5828 * Otherwise (no ACK) drop the segment and return."
5831 goto discard_and_undo
;
5835 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&&
5836 tcp_paws_reject(&tp
->rx_opt
, 0))
5837 goto discard_and_undo
;
5840 /* We see SYN without ACK. It is attempt of
5841 * simultaneous connect with crossed SYNs.
5842 * Particularly, it can be connect to self.
5844 tcp_set_state(sk
, TCP_SYN_RECV
);
5846 if (tp
->rx_opt
.saw_tstamp
) {
5847 tp
->rx_opt
.tstamp_ok
= 1;
5848 tcp_store_ts_recent(tp
);
5849 tp
->tcp_header_len
=
5850 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5852 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5855 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5856 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5858 /* RFC1323: The window in SYN & SYN/ACK segments is
5861 tp
->snd_wnd
= ntohs(th
->window
);
5862 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
5863 tp
->max_window
= tp
->snd_wnd
;
5865 TCP_ECN_rcv_syn(tp
, th
);
5868 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5869 tcp_initialize_rcv_mss(sk
);
5871 tcp_send_synack(sk
);
5873 /* Note, we could accept data and URG from this segment.
5874 * There are no obstacles to make this (except that we must
5875 * either change tcp_recvmsg() to prevent it from returning data
5876 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5878 * However, if we ignore data in ACKless segments sometimes,
5879 * we have no reasons to accept it sometimes.
5880 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5881 * is not flawless. So, discard packet for sanity.
5882 * Uncomment this return to process the data.
5889 /* "fifth, if neither of the SYN or RST bits is set then
5890 * drop the segment and return."
5894 tcp_clear_options(&tp
->rx_opt
);
5895 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5899 tcp_clear_options(&tp
->rx_opt
);
5900 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5905 * This function implements the receiving procedure of RFC 793 for
5906 * all states except ESTABLISHED and TIME_WAIT.
5907 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5908 * address independent.
5911 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5912 const struct tcphdr
*th
, unsigned int len
)
5914 struct tcp_sock
*tp
= tcp_sk(sk
);
5915 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5916 struct request_sock
*req
;
5919 tp
->rx_opt
.saw_tstamp
= 0;
5921 switch (sk
->sk_state
) {
5935 if (icsk
->icsk_af_ops
->conn_request(sk
, skb
) < 0)
5938 /* Now we have several options: In theory there is
5939 * nothing else in the frame. KA9Q has an option to
5940 * send data with the syn, BSD accepts data with the
5941 * syn up to the [to be] advertised window and
5942 * Solaris 2.1 gives you a protocol error. For now
5943 * we just ignore it, that fits the spec precisely
5944 * and avoids incompatibilities. It would be nice in
5945 * future to drop through and process the data.
5947 * Now that TTCP is starting to be used we ought to
5949 * But, this leaves one open to an easy denial of
5950 * service attack, and SYN cookies can't defend
5951 * against this problem. So, we drop the data
5952 * in the interest of security over speed unless
5953 * it's still in use.
5961 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
, len
);
5965 /* Do step6 onward by hand. */
5966 tcp_urg(sk
, skb
, th
);
5968 tcp_data_snd_check(sk
);
5972 req
= tp
->fastopen_rsk
;
5974 WARN_ON_ONCE(sk
->sk_state
!= TCP_SYN_RECV
&&
5975 sk
->sk_state
!= TCP_FIN_WAIT1
);
5977 if (tcp_check_req(sk
, skb
, req
, NULL
, true) == NULL
)
5981 if (!th
->ack
&& !th
->rst
)
5984 if (!tcp_validate_incoming(sk
, skb
, th
, 0))
5987 /* step 5: check the ACK field */
5989 int acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
) > 0;
5991 switch (sk
->sk_state
) {
5994 /* Once we leave TCP_SYN_RECV, we no longer
5995 * need req so release it.
5998 tcp_synack_rtt_meas(sk
, req
);
5999 tp
->total_retrans
= req
->num_retrans
;
6001 reqsk_fastopen_remove(sk
, req
, false);
6003 /* Make sure socket is routed, for
6006 icsk
->icsk_af_ops
->rebuild_header(sk
);
6007 tcp_init_congestion_control(sk
);
6010 tcp_init_buffer_space(sk
);
6011 tp
->copied_seq
= tp
->rcv_nxt
;
6014 tcp_set_state(sk
, TCP_ESTABLISHED
);
6015 sk
->sk_state_change(sk
);
6017 /* Note, that this wakeup is only for marginal
6018 * crossed SYN case. Passively open sockets
6019 * are not waked up, because sk->sk_sleep ==
6020 * NULL and sk->sk_socket == NULL.
6024 SOCK_WAKE_IO
, POLL_OUT
);
6026 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
6027 tp
->snd_wnd
= ntohs(th
->window
) <<
6028 tp
->rx_opt
.snd_wscale
;
6029 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
6031 if (tp
->rx_opt
.tstamp_ok
)
6032 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
6035 /* Re-arm the timer because data may
6036 * have been sent out. This is similar
6037 * to the regular data transmission case
6038 * when new data has just been ack'ed.
6040 * (TFO) - we could try to be more
6041 * aggressive and retranmitting any data
6042 * sooner based on when they were sent
6047 tcp_init_metrics(sk
);
6049 /* Prevent spurious tcp_cwnd_restart() on
6050 * first data packet.
6052 tp
->lsndtime
= tcp_time_stamp
;
6054 tcp_initialize_rcv_mss(sk
);
6055 tcp_fast_path_on(tp
);
6062 /* If we enter the TCP_FIN_WAIT1 state and we are a
6063 * Fast Open socket and this is the first acceptable
6064 * ACK we have received, this would have acknowledged
6065 * our SYNACK so stop the SYNACK timer.
6068 /* Return RST if ack_seq is invalid.
6069 * Note that RFC793 only says to generate a
6070 * DUPACK for it but for TCP Fast Open it seems
6071 * better to treat this case like TCP_SYN_RECV
6076 /* We no longer need the request sock. */
6077 reqsk_fastopen_remove(sk
, req
, false);
6080 if (tp
->snd_una
== tp
->write_seq
) {
6081 struct dst_entry
*dst
;
6083 tcp_set_state(sk
, TCP_FIN_WAIT2
);
6084 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
6086 dst
= __sk_dst_get(sk
);
6090 if (!sock_flag(sk
, SOCK_DEAD
))
6091 /* Wake up lingering close() */
6092 sk
->sk_state_change(sk
);
6096 if (tp
->linger2
< 0 ||
6097 (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
6098 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
))) {
6100 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
6104 tmo
= tcp_fin_time(sk
);
6105 if (tmo
> TCP_TIMEWAIT_LEN
) {
6106 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
6107 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
6108 /* Bad case. We could lose such FIN otherwise.
6109 * It is not a big problem, but it looks confusing
6110 * and not so rare event. We still can lose it now,
6111 * if it spins in bh_lock_sock(), but it is really
6114 inet_csk_reset_keepalive_timer(sk
, tmo
);
6116 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
6124 if (tp
->snd_una
== tp
->write_seq
) {
6125 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
6131 if (tp
->snd_una
== tp
->write_seq
) {
6132 tcp_update_metrics(sk
);
6140 /* ts_recent update must be made after we are sure that the packet
6143 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
6145 /* step 6: check the URG bit */
6146 tcp_urg(sk
, skb
, th
);
6148 /* step 7: process the segment text */
6149 switch (sk
->sk_state
) {
6150 case TCP_CLOSE_WAIT
:
6153 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
6157 /* RFC 793 says to queue data in these states,
6158 * RFC 1122 says we MUST send a reset.
6159 * BSD 4.4 also does reset.
6161 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
6162 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
6163 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
6164 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
6170 case TCP_ESTABLISHED
:
6171 tcp_data_queue(sk
, skb
);
6176 /* tcp_data could move socket to TIME-WAIT */
6177 if (sk
->sk_state
!= TCP_CLOSE
) {
6178 tcp_data_snd_check(sk
);
6179 tcp_ack_snd_check(sk
);
6188 EXPORT_SYMBOL(tcp_rcv_state_process
);