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>
22 * Changes: Pedro Roque : Retransmit queue handled by TCP.
23 * : Fragmentation on mtu decrease
24 * : Segment collapse on retransmit
27 * Linus Torvalds : send_delayed_ack
28 * David S. Miller : Charge memory using the right skb
29 * during syn/ack processing.
30 * David S. Miller : Output engine completely rewritten.
31 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
32 * Cacophonix Gaul : draft-minshall-nagle-01
33 * J Hadi Salim : ECN support
37 #define pr_fmt(fmt) "TCP: " fmt
41 #include <linux/compiler.h>
42 #include <linux/gfp.h>
43 #include <linux/module.h>
45 /* People can turn this off for buggy TCP's found in printers etc. */
46 int sysctl_tcp_retrans_collapse __read_mostly
= 1;
48 /* People can turn this on to work with those rare, broken TCPs that
49 * interpret the window field as a signed quantity.
51 int sysctl_tcp_workaround_signed_windows __read_mostly
= 0;
53 /* Default TSQ limit of two TSO segments */
54 int sysctl_tcp_limit_output_bytes __read_mostly
= 131072;
56 /* This limits the percentage of the congestion window which we
57 * will allow a single TSO frame to consume. Building TSO frames
58 * which are too large can cause TCP streams to be bursty.
60 int sysctl_tcp_tso_win_divisor __read_mostly
= 3;
62 int sysctl_tcp_mtu_probing __read_mostly
= 0;
63 int sysctl_tcp_base_mss __read_mostly
= TCP_BASE_MSS
;
65 /* By default, RFC2861 behavior. */
66 int sysctl_tcp_slow_start_after_idle __read_mostly
= 1;
68 static bool tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
,
69 int push_one
, gfp_t gfp
);
71 /* Account for new data that has been sent to the network. */
72 static void tcp_event_new_data_sent(struct sock
*sk
, const struct sk_buff
*skb
)
74 struct inet_connection_sock
*icsk
= inet_csk(sk
);
75 struct tcp_sock
*tp
= tcp_sk(sk
);
76 unsigned int prior_packets
= tp
->packets_out
;
78 tcp_advance_send_head(sk
, skb
);
79 tp
->snd_nxt
= TCP_SKB_CB(skb
)->end_seq
;
81 tp
->packets_out
+= tcp_skb_pcount(skb
);
82 if (!prior_packets
|| icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
||
83 icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
) {
88 /* SND.NXT, if window was not shrunk.
89 * If window has been shrunk, what should we make? It is not clear at all.
90 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
91 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
92 * invalid. OK, let's make this for now:
94 static inline __u32
tcp_acceptable_seq(const struct sock
*sk
)
96 const struct tcp_sock
*tp
= tcp_sk(sk
);
98 if (!before(tcp_wnd_end(tp
), tp
->snd_nxt
))
101 return tcp_wnd_end(tp
);
104 /* Calculate mss to advertise in SYN segment.
105 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
107 * 1. It is independent of path mtu.
108 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
109 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
110 * attached devices, because some buggy hosts are confused by
112 * 4. We do not make 3, we advertise MSS, calculated from first
113 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
114 * This may be overridden via information stored in routing table.
115 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
116 * probably even Jumbo".
118 static __u16
tcp_advertise_mss(struct sock
*sk
)
120 struct tcp_sock
*tp
= tcp_sk(sk
);
121 const struct dst_entry
*dst
= __sk_dst_get(sk
);
122 int mss
= tp
->advmss
;
125 unsigned int metric
= dst_metric_advmss(dst
);
136 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
137 * This is the first part of cwnd validation mechanism. */
138 static void tcp_cwnd_restart(struct sock
*sk
, const struct dst_entry
*dst
)
140 struct tcp_sock
*tp
= tcp_sk(sk
);
141 s32 delta
= tcp_time_stamp
- tp
->lsndtime
;
142 u32 restart_cwnd
= tcp_init_cwnd(tp
, dst
);
143 u32 cwnd
= tp
->snd_cwnd
;
145 tcp_ca_event(sk
, CA_EVENT_CWND_RESTART
);
147 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
148 restart_cwnd
= min(restart_cwnd
, cwnd
);
150 while ((delta
-= inet_csk(sk
)->icsk_rto
) > 0 && cwnd
> restart_cwnd
)
152 tp
->snd_cwnd
= max(cwnd
, restart_cwnd
);
153 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
154 tp
->snd_cwnd_used
= 0;
157 /* Congestion state accounting after a packet has been sent. */
158 static void tcp_event_data_sent(struct tcp_sock
*tp
,
161 struct inet_connection_sock
*icsk
= inet_csk(sk
);
162 const u32 now
= tcp_time_stamp
;
164 if (sysctl_tcp_slow_start_after_idle
&&
165 (!tp
->packets_out
&& (s32
)(now
- tp
->lsndtime
) > icsk
->icsk_rto
))
166 tcp_cwnd_restart(sk
, __sk_dst_get(sk
));
170 /* If it is a reply for ato after last received
171 * packet, enter pingpong mode.
173 if ((u32
)(now
- icsk
->icsk_ack
.lrcvtime
) < icsk
->icsk_ack
.ato
)
174 icsk
->icsk_ack
.pingpong
= 1;
177 /* Account for an ACK we sent. */
178 static inline void tcp_event_ack_sent(struct sock
*sk
, unsigned int pkts
)
180 tcp_dec_quickack_mode(sk
, pkts
);
181 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_DACK
);
185 u32
tcp_default_init_rwnd(u32 mss
)
187 /* Initial receive window should be twice of TCP_INIT_CWND to
188 * enable proper sending of new unset data during fast recovery
189 * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a
190 * limit when mss is larger than 1460.
192 u32 init_rwnd
= TCP_INIT_CWND
* 2;
195 init_rwnd
= max((1460 * init_rwnd
) / mss
, 2U);
199 /* Determine a window scaling and initial window to offer.
200 * Based on the assumption that the given amount of space
201 * will be offered. Store the results in the tp structure.
202 * NOTE: for smooth operation initial space offering should
203 * be a multiple of mss if possible. We assume here that mss >= 1.
204 * This MUST be enforced by all callers.
206 void tcp_select_initial_window(int __space
, __u32 mss
,
207 __u32
*rcv_wnd
, __u32
*window_clamp
,
208 int wscale_ok
, __u8
*rcv_wscale
,
211 unsigned int space
= (__space
< 0 ? 0 : __space
);
213 /* If no clamp set the clamp to the max possible scaled window */
214 if (*window_clamp
== 0)
215 (*window_clamp
) = (65535 << 14);
216 space
= min(*window_clamp
, space
);
218 /* Quantize space offering to a multiple of mss if possible. */
220 space
= (space
/ mss
) * mss
;
222 /* NOTE: offering an initial window larger than 32767
223 * will break some buggy TCP stacks. If the admin tells us
224 * it is likely we could be speaking with such a buggy stack
225 * we will truncate our initial window offering to 32K-1
226 * unless the remote has sent us a window scaling option,
227 * which we interpret as a sign the remote TCP is not
228 * misinterpreting the window field as a signed quantity.
230 if (sysctl_tcp_workaround_signed_windows
)
231 (*rcv_wnd
) = min(space
, MAX_TCP_WINDOW
);
237 /* Set window scaling on max possible window
238 * See RFC1323 for an explanation of the limit to 14
240 space
= max_t(u32
, sysctl_tcp_rmem
[2], sysctl_rmem_max
);
241 space
= min_t(u32
, space
, *window_clamp
);
242 while (space
> 65535 && (*rcv_wscale
) < 14) {
248 if (mss
> (1 << *rcv_wscale
)) {
249 if (!init_rcv_wnd
) /* Use default unless specified otherwise */
250 init_rcv_wnd
= tcp_default_init_rwnd(mss
);
251 *rcv_wnd
= min(*rcv_wnd
, init_rcv_wnd
* mss
);
254 /* Set the clamp no higher than max representable value */
255 (*window_clamp
) = min(65535U << (*rcv_wscale
), *window_clamp
);
257 EXPORT_SYMBOL(tcp_select_initial_window
);
259 /* Chose a new window to advertise, update state in tcp_sock for the
260 * socket, and return result with RFC1323 scaling applied. The return
261 * value can be stuffed directly into th->window for an outgoing
264 static u16
tcp_select_window(struct sock
*sk
)
266 struct tcp_sock
*tp
= tcp_sk(sk
);
267 u32 cur_win
= tcp_receive_window(tp
);
268 u32 new_win
= __tcp_select_window(sk
);
270 /* Never shrink the offered window */
271 if (new_win
< cur_win
) {
272 /* Danger Will Robinson!
273 * Don't update rcv_wup/rcv_wnd here or else
274 * we will not be able to advertise a zero
275 * window in time. --DaveM
277 * Relax Will Robinson.
279 new_win
= ALIGN(cur_win
, 1 << tp
->rx_opt
.rcv_wscale
);
281 tp
->rcv_wnd
= new_win
;
282 tp
->rcv_wup
= tp
->rcv_nxt
;
284 /* Make sure we do not exceed the maximum possible
287 if (!tp
->rx_opt
.rcv_wscale
&& sysctl_tcp_workaround_signed_windows
)
288 new_win
= min(new_win
, MAX_TCP_WINDOW
);
290 new_win
= min(new_win
, (65535U << tp
->rx_opt
.rcv_wscale
));
292 /* RFC1323 scaling applied */
293 new_win
>>= tp
->rx_opt
.rcv_wscale
;
295 /* If we advertise zero window, disable fast path. */
302 /* Packet ECN state for a SYN-ACK */
303 static inline void TCP_ECN_send_synack(const struct tcp_sock
*tp
, struct sk_buff
*skb
)
305 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_CWR
;
306 if (!(tp
->ecn_flags
& TCP_ECN_OK
))
307 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_ECE
;
310 /* Packet ECN state for a SYN. */
311 static inline void TCP_ECN_send_syn(struct sock
*sk
, struct sk_buff
*skb
)
313 struct tcp_sock
*tp
= tcp_sk(sk
);
316 if (sock_net(sk
)->ipv4
.sysctl_tcp_ecn
== 1) {
317 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_ECE
| TCPHDR_CWR
;
318 tp
->ecn_flags
= TCP_ECN_OK
;
322 static __inline__
void
323 TCP_ECN_make_synack(const struct request_sock
*req
, struct tcphdr
*th
)
325 if (inet_rsk(req
)->ecn_ok
)
329 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
332 static inline void TCP_ECN_send(struct sock
*sk
, struct sk_buff
*skb
,
335 struct tcp_sock
*tp
= tcp_sk(sk
);
337 if (tp
->ecn_flags
& TCP_ECN_OK
) {
338 /* Not-retransmitted data segment: set ECT and inject CWR. */
339 if (skb
->len
!= tcp_header_len
&&
340 !before(TCP_SKB_CB(skb
)->seq
, tp
->snd_nxt
)) {
342 if (tp
->ecn_flags
& TCP_ECN_QUEUE_CWR
) {
343 tp
->ecn_flags
&= ~TCP_ECN_QUEUE_CWR
;
344 tcp_hdr(skb
)->cwr
= 1;
345 skb_shinfo(skb
)->gso_type
|= SKB_GSO_TCP_ECN
;
348 /* ACK or retransmitted segment: clear ECT|CE */
349 INET_ECN_dontxmit(sk
);
351 if (tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)
352 tcp_hdr(skb
)->ece
= 1;
356 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
357 * auto increment end seqno.
359 static void tcp_init_nondata_skb(struct sk_buff
*skb
, u32 seq
, u8 flags
)
361 skb
->ip_summed
= CHECKSUM_PARTIAL
;
364 TCP_SKB_CB(skb
)->tcp_flags
= flags
;
365 TCP_SKB_CB(skb
)->sacked
= 0;
367 skb_shinfo(skb
)->gso_segs
= 1;
368 skb_shinfo(skb
)->gso_size
= 0;
369 skb_shinfo(skb
)->gso_type
= 0;
371 TCP_SKB_CB(skb
)->seq
= seq
;
372 if (flags
& (TCPHDR_SYN
| TCPHDR_FIN
))
374 TCP_SKB_CB(skb
)->end_seq
= seq
;
377 static inline bool tcp_urg_mode(const struct tcp_sock
*tp
)
379 return tp
->snd_una
!= tp
->snd_up
;
382 #define OPTION_SACK_ADVERTISE (1 << 0)
383 #define OPTION_TS (1 << 1)
384 #define OPTION_MD5 (1 << 2)
385 #define OPTION_WSCALE (1 << 3)
386 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
388 struct tcp_out_options
{
389 u16 options
; /* bit field of OPTION_* */
390 u16 mss
; /* 0 to disable */
391 u8 ws
; /* window scale, 0 to disable */
392 u8 num_sack_blocks
; /* number of SACK blocks to include */
393 u8 hash_size
; /* bytes in hash_location */
394 __u8
*hash_location
; /* temporary pointer, overloaded */
395 __u32 tsval
, tsecr
; /* need to include OPTION_TS */
396 struct tcp_fastopen_cookie
*fastopen_cookie
; /* Fast open cookie */
399 /* Write previously computed TCP options to the packet.
401 * Beware: Something in the Internet is very sensitive to the ordering of
402 * TCP options, we learned this through the hard way, so be careful here.
403 * Luckily we can at least blame others for their non-compliance but from
404 * inter-operatibility perspective it seems that we're somewhat stuck with
405 * the ordering which we have been using if we want to keep working with
406 * those broken things (not that it currently hurts anybody as there isn't
407 * particular reason why the ordering would need to be changed).
409 * At least SACK_PERM as the first option is known to lead to a disaster
410 * (but it may well be that other scenarios fail similarly).
412 static void tcp_options_write(__be32
*ptr
, struct tcp_sock
*tp
,
413 struct tcp_out_options
*opts
)
415 u16 options
= opts
->options
; /* mungable copy */
417 if (unlikely(OPTION_MD5
& options
)) {
418 *ptr
++ = htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16) |
419 (TCPOPT_MD5SIG
<< 8) | TCPOLEN_MD5SIG
);
420 /* overload cookie hash location */
421 opts
->hash_location
= (__u8
*)ptr
;
425 if (unlikely(opts
->mss
)) {
426 *ptr
++ = htonl((TCPOPT_MSS
<< 24) |
427 (TCPOLEN_MSS
<< 16) |
431 if (likely(OPTION_TS
& options
)) {
432 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
433 *ptr
++ = htonl((TCPOPT_SACK_PERM
<< 24) |
434 (TCPOLEN_SACK_PERM
<< 16) |
435 (TCPOPT_TIMESTAMP
<< 8) |
437 options
&= ~OPTION_SACK_ADVERTISE
;
439 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
441 (TCPOPT_TIMESTAMP
<< 8) |
444 *ptr
++ = htonl(opts
->tsval
);
445 *ptr
++ = htonl(opts
->tsecr
);
448 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
449 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
451 (TCPOPT_SACK_PERM
<< 8) |
455 if (unlikely(OPTION_WSCALE
& options
)) {
456 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
457 (TCPOPT_WINDOW
<< 16) |
458 (TCPOLEN_WINDOW
<< 8) |
462 if (unlikely(opts
->num_sack_blocks
)) {
463 struct tcp_sack_block
*sp
= tp
->rx_opt
.dsack
?
464 tp
->duplicate_sack
: tp
->selective_acks
;
467 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
470 (TCPOLEN_SACK_BASE
+ (opts
->num_sack_blocks
*
471 TCPOLEN_SACK_PERBLOCK
)));
473 for (this_sack
= 0; this_sack
< opts
->num_sack_blocks
;
475 *ptr
++ = htonl(sp
[this_sack
].start_seq
);
476 *ptr
++ = htonl(sp
[this_sack
].end_seq
);
479 tp
->rx_opt
.dsack
= 0;
482 if (unlikely(OPTION_FAST_OPEN_COOKIE
& options
)) {
483 struct tcp_fastopen_cookie
*foc
= opts
->fastopen_cookie
;
485 *ptr
++ = htonl((TCPOPT_EXP
<< 24) |
486 ((TCPOLEN_EXP_FASTOPEN_BASE
+ foc
->len
) << 16) |
487 TCPOPT_FASTOPEN_MAGIC
);
489 memcpy(ptr
, foc
->val
, foc
->len
);
490 if ((foc
->len
& 3) == 2) {
491 u8
*align
= ((u8
*)ptr
) + foc
->len
;
492 align
[0] = align
[1] = TCPOPT_NOP
;
494 ptr
+= (foc
->len
+ 3) >> 2;
498 /* Compute TCP options for SYN packets. This is not the final
499 * network wire format yet.
501 static unsigned int tcp_syn_options(struct sock
*sk
, struct sk_buff
*skb
,
502 struct tcp_out_options
*opts
,
503 struct tcp_md5sig_key
**md5
)
505 struct tcp_sock
*tp
= tcp_sk(sk
);
506 unsigned int remaining
= MAX_TCP_OPTION_SPACE
;
507 struct tcp_fastopen_request
*fastopen
= tp
->fastopen_req
;
509 #ifdef CONFIG_TCP_MD5SIG
510 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
512 opts
->options
|= OPTION_MD5
;
513 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
519 /* We always get an MSS option. The option bytes which will be seen in
520 * normal data packets should timestamps be used, must be in the MSS
521 * advertised. But we subtract them from tp->mss_cache so that
522 * calculations in tcp_sendmsg are simpler etc. So account for this
523 * fact here if necessary. If we don't do this correctly, as a
524 * receiver we won't recognize data packets as being full sized when we
525 * should, and thus we won't abide by the delayed ACK rules correctly.
526 * SACKs don't matter, we never delay an ACK when we have any of those
528 opts
->mss
= tcp_advertise_mss(sk
);
529 remaining
-= TCPOLEN_MSS_ALIGNED
;
531 if (likely(sysctl_tcp_timestamps
&& *md5
== NULL
)) {
532 opts
->options
|= OPTION_TS
;
533 opts
->tsval
= TCP_SKB_CB(skb
)->when
+ tp
->tsoffset
;
534 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
535 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
537 if (likely(sysctl_tcp_window_scaling
)) {
538 opts
->ws
= tp
->rx_opt
.rcv_wscale
;
539 opts
->options
|= OPTION_WSCALE
;
540 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
542 if (likely(sysctl_tcp_sack
)) {
543 opts
->options
|= OPTION_SACK_ADVERTISE
;
544 if (unlikely(!(OPTION_TS
& opts
->options
)))
545 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
548 if (fastopen
&& fastopen
->cookie
.len
>= 0) {
549 u32 need
= TCPOLEN_EXP_FASTOPEN_BASE
+ fastopen
->cookie
.len
;
550 need
= (need
+ 3) & ~3U; /* Align to 32 bits */
551 if (remaining
>= need
) {
552 opts
->options
|= OPTION_FAST_OPEN_COOKIE
;
553 opts
->fastopen_cookie
= &fastopen
->cookie
;
555 tp
->syn_fastopen
= 1;
559 return MAX_TCP_OPTION_SPACE
- remaining
;
562 /* Set up TCP options for SYN-ACKs. */
563 static unsigned int tcp_synack_options(struct sock
*sk
,
564 struct request_sock
*req
,
565 unsigned int mss
, struct sk_buff
*skb
,
566 struct tcp_out_options
*opts
,
567 struct tcp_md5sig_key
**md5
,
568 struct tcp_fastopen_cookie
*foc
)
570 struct inet_request_sock
*ireq
= inet_rsk(req
);
571 unsigned int remaining
= MAX_TCP_OPTION_SPACE
;
573 #ifdef CONFIG_TCP_MD5SIG
574 *md5
= tcp_rsk(req
)->af_specific
->md5_lookup(sk
, req
);
576 opts
->options
|= OPTION_MD5
;
577 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
579 /* We can't fit any SACK blocks in a packet with MD5 + TS
580 * options. There was discussion about disabling SACK
581 * rather than TS in order to fit in better with old,
582 * buggy kernels, but that was deemed to be unnecessary.
584 ireq
->tstamp_ok
&= !ireq
->sack_ok
;
590 /* We always send an MSS option. */
592 remaining
-= TCPOLEN_MSS_ALIGNED
;
594 if (likely(ireq
->wscale_ok
)) {
595 opts
->ws
= ireq
->rcv_wscale
;
596 opts
->options
|= OPTION_WSCALE
;
597 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
599 if (likely(ireq
->tstamp_ok
)) {
600 opts
->options
|= OPTION_TS
;
601 opts
->tsval
= TCP_SKB_CB(skb
)->when
;
602 opts
->tsecr
= req
->ts_recent
;
603 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
605 if (likely(ireq
->sack_ok
)) {
606 opts
->options
|= OPTION_SACK_ADVERTISE
;
607 if (unlikely(!ireq
->tstamp_ok
))
608 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
611 u32 need
= TCPOLEN_EXP_FASTOPEN_BASE
+ foc
->len
;
612 need
= (need
+ 3) & ~3U; /* Align to 32 bits */
613 if (remaining
>= need
) {
614 opts
->options
|= OPTION_FAST_OPEN_COOKIE
;
615 opts
->fastopen_cookie
= foc
;
620 return MAX_TCP_OPTION_SPACE
- remaining
;
623 /* Compute TCP options for ESTABLISHED sockets. This is not the
624 * final wire format yet.
626 static unsigned int tcp_established_options(struct sock
*sk
, struct sk_buff
*skb
,
627 struct tcp_out_options
*opts
,
628 struct tcp_md5sig_key
**md5
)
630 struct tcp_skb_cb
*tcb
= skb
? TCP_SKB_CB(skb
) : NULL
;
631 struct tcp_sock
*tp
= tcp_sk(sk
);
632 unsigned int size
= 0;
633 unsigned int eff_sacks
;
635 #ifdef CONFIG_TCP_MD5SIG
636 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
637 if (unlikely(*md5
)) {
638 opts
->options
|= OPTION_MD5
;
639 size
+= TCPOLEN_MD5SIG_ALIGNED
;
645 if (likely(tp
->rx_opt
.tstamp_ok
)) {
646 opts
->options
|= OPTION_TS
;
647 opts
->tsval
= tcb
? tcb
->when
+ tp
->tsoffset
: 0;
648 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
649 size
+= TCPOLEN_TSTAMP_ALIGNED
;
652 eff_sacks
= tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
;
653 if (unlikely(eff_sacks
)) {
654 const unsigned int remaining
= MAX_TCP_OPTION_SPACE
- size
;
655 opts
->num_sack_blocks
=
656 min_t(unsigned int, eff_sacks
,
657 (remaining
- TCPOLEN_SACK_BASE_ALIGNED
) /
658 TCPOLEN_SACK_PERBLOCK
);
659 size
+= TCPOLEN_SACK_BASE_ALIGNED
+
660 opts
->num_sack_blocks
* TCPOLEN_SACK_PERBLOCK
;
667 /* TCP SMALL QUEUES (TSQ)
669 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
670 * to reduce RTT and bufferbloat.
671 * We do this using a special skb destructor (tcp_wfree).
673 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
674 * needs to be reallocated in a driver.
675 * The invariant being skb->truesize substracted from sk->sk_wmem_alloc
677 * Since transmit from skb destructor is forbidden, we use a tasklet
678 * to process all sockets that eventually need to send more skbs.
679 * We use one tasklet per cpu, with its own queue of sockets.
682 struct tasklet_struct tasklet
;
683 struct list_head head
; /* queue of tcp sockets */
685 static DEFINE_PER_CPU(struct tsq_tasklet
, tsq_tasklet
);
687 static void tcp_tsq_handler(struct sock
*sk
)
689 if ((1 << sk
->sk_state
) &
690 (TCPF_ESTABLISHED
| TCPF_FIN_WAIT1
| TCPF_CLOSING
|
691 TCPF_CLOSE_WAIT
| TCPF_LAST_ACK
))
692 tcp_write_xmit(sk
, tcp_current_mss(sk
), 0, 0, GFP_ATOMIC
);
695 * One tasklest per cpu tries to send more skbs.
696 * We run in tasklet context but need to disable irqs when
697 * transfering tsq->head because tcp_wfree() might
698 * interrupt us (non NAPI drivers)
700 static void tcp_tasklet_func(unsigned long data
)
702 struct tsq_tasklet
*tsq
= (struct tsq_tasklet
*)data
;
705 struct list_head
*q
, *n
;
709 local_irq_save(flags
);
710 list_splice_init(&tsq
->head
, &list
);
711 local_irq_restore(flags
);
713 list_for_each_safe(q
, n
, &list
) {
714 tp
= list_entry(q
, struct tcp_sock
, tsq_node
);
715 list_del(&tp
->tsq_node
);
717 sk
= (struct sock
*)tp
;
720 if (!sock_owned_by_user(sk
)) {
723 /* defer the work to tcp_release_cb() */
724 set_bit(TCP_TSQ_DEFERRED
, &tp
->tsq_flags
);
728 clear_bit(TSQ_QUEUED
, &tp
->tsq_flags
);
733 #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \
734 (1UL << TCP_WRITE_TIMER_DEFERRED) | \
735 (1UL << TCP_DELACK_TIMER_DEFERRED) | \
736 (1UL << TCP_MTU_REDUCED_DEFERRED))
738 * tcp_release_cb - tcp release_sock() callback
741 * called from release_sock() to perform protocol dependent
742 * actions before socket release.
744 void tcp_release_cb(struct sock
*sk
)
746 struct tcp_sock
*tp
= tcp_sk(sk
);
747 unsigned long flags
, nflags
;
749 /* perform an atomic operation only if at least one flag is set */
751 flags
= tp
->tsq_flags
;
752 if (!(flags
& TCP_DEFERRED_ALL
))
754 nflags
= flags
& ~TCP_DEFERRED_ALL
;
755 } while (cmpxchg(&tp
->tsq_flags
, flags
, nflags
) != flags
);
757 if (flags
& (1UL << TCP_TSQ_DEFERRED
))
760 if (flags
& (1UL << TCP_WRITE_TIMER_DEFERRED
)) {
761 tcp_write_timer_handler(sk
);
764 if (flags
& (1UL << TCP_DELACK_TIMER_DEFERRED
)) {
765 tcp_delack_timer_handler(sk
);
768 if (flags
& (1UL << TCP_MTU_REDUCED_DEFERRED
)) {
769 sk
->sk_prot
->mtu_reduced(sk
);
773 EXPORT_SYMBOL(tcp_release_cb
);
775 void __init
tcp_tasklet_init(void)
779 for_each_possible_cpu(i
) {
780 struct tsq_tasklet
*tsq
= &per_cpu(tsq_tasklet
, i
);
782 INIT_LIST_HEAD(&tsq
->head
);
783 tasklet_init(&tsq
->tasklet
,
790 * Write buffer destructor automatically called from kfree_skb.
791 * We cant xmit new skbs from this context, as we might already
794 void tcp_wfree(struct sk_buff
*skb
)
796 struct sock
*sk
= skb
->sk
;
797 struct tcp_sock
*tp
= tcp_sk(sk
);
799 if (test_and_clear_bit(TSQ_THROTTLED
, &tp
->tsq_flags
) &&
800 !test_and_set_bit(TSQ_QUEUED
, &tp
->tsq_flags
)) {
802 struct tsq_tasklet
*tsq
;
804 /* Keep a ref on socket.
805 * This last ref will be released in tcp_tasklet_func()
807 atomic_sub(skb
->truesize
- 1, &sk
->sk_wmem_alloc
);
809 /* queue this socket to tasklet queue */
810 local_irq_save(flags
);
811 tsq
= &__get_cpu_var(tsq_tasklet
);
812 list_add(&tp
->tsq_node
, &tsq
->head
);
813 tasklet_schedule(&tsq
->tasklet
);
814 local_irq_restore(flags
);
820 /* This routine actually transmits TCP packets queued in by
821 * tcp_do_sendmsg(). This is used by both the initial
822 * transmission and possible later retransmissions.
823 * All SKB's seen here are completely headerless. It is our
824 * job to build the TCP header, and pass the packet down to
825 * IP so it can do the same plus pass the packet off to the
828 * We are working here with either a clone of the original
829 * SKB, or a fresh unique copy made by the retransmit engine.
831 static int tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int clone_it
,
834 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
835 struct inet_sock
*inet
;
837 struct tcp_skb_cb
*tcb
;
838 struct tcp_out_options opts
;
839 unsigned int tcp_options_size
, tcp_header_size
;
840 struct tcp_md5sig_key
*md5
;
844 BUG_ON(!skb
|| !tcp_skb_pcount(skb
));
846 /* If congestion control is doing timestamping, we must
847 * take such a timestamp before we potentially clone/copy.
849 if (icsk
->icsk_ca_ops
->flags
& TCP_CONG_RTT_STAMP
)
850 __net_timestamp(skb
);
852 if (likely(clone_it
)) {
853 const struct sk_buff
*fclone
= skb
+ 1;
855 if (unlikely(skb
->fclone
== SKB_FCLONE_ORIG
&&
856 fclone
->fclone
== SKB_FCLONE_CLONE
))
857 NET_INC_STATS_BH(sock_net(sk
),
858 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES
);
860 if (unlikely(skb_cloned(skb
)))
861 skb
= pskb_copy(skb
, gfp_mask
);
863 skb
= skb_clone(skb
, gfp_mask
);
870 tcb
= TCP_SKB_CB(skb
);
871 memset(&opts
, 0, sizeof(opts
));
873 if (unlikely(tcb
->tcp_flags
& TCPHDR_SYN
))
874 tcp_options_size
= tcp_syn_options(sk
, skb
, &opts
, &md5
);
876 tcp_options_size
= tcp_established_options(sk
, skb
, &opts
,
878 tcp_header_size
= tcp_options_size
+ sizeof(struct tcphdr
);
880 if (tcp_packets_in_flight(tp
) == 0)
881 tcp_ca_event(sk
, CA_EVENT_TX_START
);
883 /* if no packet is in qdisc/device queue, then allow XPS to select
886 skb
->ooo_okay
= sk_wmem_alloc_get(sk
) == 0;
888 skb_push(skb
, tcp_header_size
);
889 skb_reset_transport_header(skb
);
893 skb
->destructor
= (sysctl_tcp_limit_output_bytes
> 0) ?
894 tcp_wfree
: sock_wfree
;
895 atomic_add(skb
->truesize
, &sk
->sk_wmem_alloc
);
897 /* Build TCP header and checksum it. */
899 th
->source
= inet
->inet_sport
;
900 th
->dest
= inet
->inet_dport
;
901 th
->seq
= htonl(tcb
->seq
);
902 th
->ack_seq
= htonl(tp
->rcv_nxt
);
903 *(((__be16
*)th
) + 6) = htons(((tcp_header_size
>> 2) << 12) |
906 if (unlikely(tcb
->tcp_flags
& TCPHDR_SYN
)) {
907 /* RFC1323: The window in SYN & SYN/ACK segments
910 th
->window
= htons(min(tp
->rcv_wnd
, 65535U));
912 th
->window
= htons(tcp_select_window(sk
));
917 /* The urg_mode check is necessary during a below snd_una win probe */
918 if (unlikely(tcp_urg_mode(tp
) && before(tcb
->seq
, tp
->snd_up
))) {
919 if (before(tp
->snd_up
, tcb
->seq
+ 0x10000)) {
920 th
->urg_ptr
= htons(tp
->snd_up
- tcb
->seq
);
922 } else if (after(tcb
->seq
+ 0xFFFF, tp
->snd_nxt
)) {
923 th
->urg_ptr
= htons(0xFFFF);
928 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
);
929 if (likely((tcb
->tcp_flags
& TCPHDR_SYN
) == 0))
930 TCP_ECN_send(sk
, skb
, tcp_header_size
);
932 #ifdef CONFIG_TCP_MD5SIG
933 /* Calculate the MD5 hash, as we have all we need now */
935 sk_nocaps_add(sk
, NETIF_F_GSO_MASK
);
936 tp
->af_specific
->calc_md5_hash(opts
.hash_location
,
941 icsk
->icsk_af_ops
->send_check(sk
, skb
);
943 if (likely(tcb
->tcp_flags
& TCPHDR_ACK
))
944 tcp_event_ack_sent(sk
, tcp_skb_pcount(skb
));
946 if (skb
->len
!= tcp_header_size
)
947 tcp_event_data_sent(tp
, sk
);
949 if (after(tcb
->end_seq
, tp
->snd_nxt
) || tcb
->seq
== tcb
->end_seq
)
950 TCP_ADD_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
,
951 tcp_skb_pcount(skb
));
953 err
= icsk
->icsk_af_ops
->queue_xmit(skb
, &inet
->cork
.fl
);
954 if (likely(err
<= 0))
957 tcp_enter_cwr(sk
, 1);
959 return net_xmit_eval(err
);
962 /* This routine just queues the buffer for sending.
964 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
965 * otherwise socket can stall.
967 static void tcp_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
969 struct tcp_sock
*tp
= tcp_sk(sk
);
971 /* Advance write_seq and place onto the write_queue. */
972 tp
->write_seq
= TCP_SKB_CB(skb
)->end_seq
;
973 skb_header_release(skb
);
974 tcp_add_write_queue_tail(sk
, skb
);
975 sk
->sk_wmem_queued
+= skb
->truesize
;
976 sk_mem_charge(sk
, skb
->truesize
);
979 /* Initialize TSO segments for a packet. */
980 static void tcp_set_skb_tso_segs(const struct sock
*sk
, struct sk_buff
*skb
,
981 unsigned int mss_now
)
983 if (skb
->len
<= mss_now
|| !sk_can_gso(sk
) ||
984 skb
->ip_summed
== CHECKSUM_NONE
) {
985 /* Avoid the costly divide in the normal
988 skb_shinfo(skb
)->gso_segs
= 1;
989 skb_shinfo(skb
)->gso_size
= 0;
990 skb_shinfo(skb
)->gso_type
= 0;
992 skb_shinfo(skb
)->gso_segs
= DIV_ROUND_UP(skb
->len
, mss_now
);
993 skb_shinfo(skb
)->gso_size
= mss_now
;
994 skb_shinfo(skb
)->gso_type
= sk
->sk_gso_type
;
998 /* When a modification to fackets out becomes necessary, we need to check
999 * skb is counted to fackets_out or not.
1001 static void tcp_adjust_fackets_out(struct sock
*sk
, const struct sk_buff
*skb
,
1004 struct tcp_sock
*tp
= tcp_sk(sk
);
1006 if (!tp
->sacked_out
|| tcp_is_reno(tp
))
1009 if (after(tcp_highest_sack_seq(tp
), TCP_SKB_CB(skb
)->seq
))
1010 tp
->fackets_out
-= decr
;
1013 /* Pcount in the middle of the write queue got changed, we need to do various
1014 * tweaks to fix counters
1016 static void tcp_adjust_pcount(struct sock
*sk
, const struct sk_buff
*skb
, int decr
)
1018 struct tcp_sock
*tp
= tcp_sk(sk
);
1020 tp
->packets_out
-= decr
;
1022 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
1023 tp
->sacked_out
-= decr
;
1024 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
1025 tp
->retrans_out
-= decr
;
1026 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
)
1027 tp
->lost_out
-= decr
;
1029 /* Reno case is special. Sigh... */
1030 if (tcp_is_reno(tp
) && decr
> 0)
1031 tp
->sacked_out
-= min_t(u32
, tp
->sacked_out
, decr
);
1033 tcp_adjust_fackets_out(sk
, skb
, decr
);
1035 if (tp
->lost_skb_hint
&&
1036 before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
) &&
1037 (tcp_is_fack(tp
) || (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)))
1038 tp
->lost_cnt_hint
-= decr
;
1040 tcp_verify_left_out(tp
);
1043 /* Function to create two new TCP segments. Shrinks the given segment
1044 * to the specified size and appends a new segment with the rest of the
1045 * packet to the list. This won't be called frequently, I hope.
1046 * Remember, these are still headerless SKBs at this point.
1048 int tcp_fragment(struct sock
*sk
, struct sk_buff
*skb
, u32 len
,
1049 unsigned int mss_now
)
1051 struct tcp_sock
*tp
= tcp_sk(sk
);
1052 struct sk_buff
*buff
;
1053 int nsize
, old_factor
;
1057 if (WARN_ON(len
> skb
->len
))
1060 nsize
= skb_headlen(skb
) - len
;
1064 if (skb_cloned(skb
) &&
1065 skb_is_nonlinear(skb
) &&
1066 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
1069 /* Get a new skb... force flag on. */
1070 buff
= sk_stream_alloc_skb(sk
, nsize
, GFP_ATOMIC
);
1072 return -ENOMEM
; /* We'll just try again later. */
1074 sk
->sk_wmem_queued
+= buff
->truesize
;
1075 sk_mem_charge(sk
, buff
->truesize
);
1076 nlen
= skb
->len
- len
- nsize
;
1077 buff
->truesize
+= nlen
;
1078 skb
->truesize
-= nlen
;
1080 /* Correct the sequence numbers. */
1081 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1082 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1083 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1085 /* PSH and FIN should only be set in the second packet. */
1086 flags
= TCP_SKB_CB(skb
)->tcp_flags
;
1087 TCP_SKB_CB(skb
)->tcp_flags
= flags
& ~(TCPHDR_FIN
| TCPHDR_PSH
);
1088 TCP_SKB_CB(buff
)->tcp_flags
= flags
;
1089 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
1091 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_PARTIAL
) {
1092 /* Copy and checksum data tail into the new buffer. */
1093 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
,
1094 skb_put(buff
, nsize
),
1099 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
1101 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1102 skb_split(skb
, buff
, len
);
1105 buff
->ip_summed
= skb
->ip_summed
;
1107 /* Looks stupid, but our code really uses when of
1108 * skbs, which it never sent before. --ANK
1110 TCP_SKB_CB(buff
)->when
= TCP_SKB_CB(skb
)->when
;
1111 buff
->tstamp
= skb
->tstamp
;
1113 old_factor
= tcp_skb_pcount(skb
);
1115 /* Fix up tso_factor for both original and new SKB. */
1116 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1117 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1119 /* If this packet has been sent out already, we must
1120 * adjust the various packet counters.
1122 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
1123 int diff
= old_factor
- tcp_skb_pcount(skb
) -
1124 tcp_skb_pcount(buff
);
1127 tcp_adjust_pcount(sk
, skb
, diff
);
1130 /* Link BUFF into the send queue. */
1131 skb_header_release(buff
);
1132 tcp_insert_write_queue_after(skb
, buff
, sk
);
1137 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1138 * eventually). The difference is that pulled data not copied, but
1139 * immediately discarded.
1141 static void __pskb_trim_head(struct sk_buff
*skb
, int len
)
1145 eat
= min_t(int, len
, skb_headlen(skb
));
1147 __skb_pull(skb
, eat
);
1154 for (i
= 0; i
< skb_shinfo(skb
)->nr_frags
; i
++) {
1155 int size
= skb_frag_size(&skb_shinfo(skb
)->frags
[i
]);
1158 skb_frag_unref(skb
, i
);
1161 skb_shinfo(skb
)->frags
[k
] = skb_shinfo(skb
)->frags
[i
];
1163 skb_shinfo(skb
)->frags
[k
].page_offset
+= eat
;
1164 skb_frag_size_sub(&skb_shinfo(skb
)->frags
[k
], eat
);
1170 skb_shinfo(skb
)->nr_frags
= k
;
1172 skb_reset_tail_pointer(skb
);
1173 skb
->data_len
-= len
;
1174 skb
->len
= skb
->data_len
;
1177 /* Remove acked data from a packet in the transmit queue. */
1178 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
1180 if (skb_unclone(skb
, GFP_ATOMIC
))
1183 __pskb_trim_head(skb
, len
);
1185 TCP_SKB_CB(skb
)->seq
+= len
;
1186 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1188 skb
->truesize
-= len
;
1189 sk
->sk_wmem_queued
-= len
;
1190 sk_mem_uncharge(sk
, len
);
1191 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
1193 /* Any change of skb->len requires recalculation of tso factor. */
1194 if (tcp_skb_pcount(skb
) > 1)
1195 tcp_set_skb_tso_segs(sk
, skb
, tcp_skb_mss(skb
));
1200 /* Calculate MSS not accounting any TCP options. */
1201 static inline int __tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1203 const struct tcp_sock
*tp
= tcp_sk(sk
);
1204 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1207 /* Calculate base mss without TCP options:
1208 It is MMS_S - sizeof(tcphdr) of rfc1122
1210 mss_now
= pmtu
- icsk
->icsk_af_ops
->net_header_len
- sizeof(struct tcphdr
);
1212 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1213 if (icsk
->icsk_af_ops
->net_frag_header_len
) {
1214 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1216 if (dst
&& dst_allfrag(dst
))
1217 mss_now
-= icsk
->icsk_af_ops
->net_frag_header_len
;
1220 /* Clamp it (mss_clamp does not include tcp options) */
1221 if (mss_now
> tp
->rx_opt
.mss_clamp
)
1222 mss_now
= tp
->rx_opt
.mss_clamp
;
1224 /* Now subtract optional transport overhead */
1225 mss_now
-= icsk
->icsk_ext_hdr_len
;
1227 /* Then reserve room for full set of TCP options and 8 bytes of data */
1233 /* Calculate MSS. Not accounting for SACKs here. */
1234 int tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1236 /* Subtract TCP options size, not including SACKs */
1237 return __tcp_mtu_to_mss(sk
, pmtu
) -
1238 (tcp_sk(sk
)->tcp_header_len
- sizeof(struct tcphdr
));
1241 /* Inverse of above */
1242 int tcp_mss_to_mtu(struct sock
*sk
, int mss
)
1244 const struct tcp_sock
*tp
= tcp_sk(sk
);
1245 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1249 tp
->tcp_header_len
+
1250 icsk
->icsk_ext_hdr_len
+
1251 icsk
->icsk_af_ops
->net_header_len
;
1253 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1254 if (icsk
->icsk_af_ops
->net_frag_header_len
) {
1255 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1257 if (dst
&& dst_allfrag(dst
))
1258 mtu
+= icsk
->icsk_af_ops
->net_frag_header_len
;
1263 /* MTU probing init per socket */
1264 void tcp_mtup_init(struct sock
*sk
)
1266 struct tcp_sock
*tp
= tcp_sk(sk
);
1267 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1269 icsk
->icsk_mtup
.enabled
= sysctl_tcp_mtu_probing
> 1;
1270 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+ sizeof(struct tcphdr
) +
1271 icsk
->icsk_af_ops
->net_header_len
;
1272 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, sysctl_tcp_base_mss
);
1273 icsk
->icsk_mtup
.probe_size
= 0;
1275 EXPORT_SYMBOL(tcp_mtup_init
);
1277 /* This function synchronize snd mss to current pmtu/exthdr set.
1279 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1280 for TCP options, but includes only bare TCP header.
1282 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1283 It is minimum of user_mss and mss received with SYN.
1284 It also does not include TCP options.
1286 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1288 tp->mss_cache is current effective sending mss, including
1289 all tcp options except for SACKs. It is evaluated,
1290 taking into account current pmtu, but never exceeds
1291 tp->rx_opt.mss_clamp.
1293 NOTE1. rfc1122 clearly states that advertised MSS
1294 DOES NOT include either tcp or ip options.
1296 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1297 are READ ONLY outside this function. --ANK (980731)
1299 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
1301 struct tcp_sock
*tp
= tcp_sk(sk
);
1302 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1305 if (icsk
->icsk_mtup
.search_high
> pmtu
)
1306 icsk
->icsk_mtup
.search_high
= pmtu
;
1308 mss_now
= tcp_mtu_to_mss(sk
, pmtu
);
1309 mss_now
= tcp_bound_to_half_wnd(tp
, mss_now
);
1311 /* And store cached results */
1312 icsk
->icsk_pmtu_cookie
= pmtu
;
1313 if (icsk
->icsk_mtup
.enabled
)
1314 mss_now
= min(mss_now
, tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
));
1315 tp
->mss_cache
= mss_now
;
1319 EXPORT_SYMBOL(tcp_sync_mss
);
1321 /* Compute the current effective MSS, taking SACKs and IP options,
1322 * and even PMTU discovery events into account.
1324 unsigned int tcp_current_mss(struct sock
*sk
)
1326 const struct tcp_sock
*tp
= tcp_sk(sk
);
1327 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1329 unsigned int header_len
;
1330 struct tcp_out_options opts
;
1331 struct tcp_md5sig_key
*md5
;
1333 mss_now
= tp
->mss_cache
;
1336 u32 mtu
= dst_mtu(dst
);
1337 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
1338 mss_now
= tcp_sync_mss(sk
, mtu
);
1341 header_len
= tcp_established_options(sk
, NULL
, &opts
, &md5
) +
1342 sizeof(struct tcphdr
);
1343 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1344 * some common options. If this is an odd packet (because we have SACK
1345 * blocks etc) then our calculated header_len will be different, and
1346 * we have to adjust mss_now correspondingly */
1347 if (header_len
!= tp
->tcp_header_len
) {
1348 int delta
= (int) header_len
- tp
->tcp_header_len
;
1355 /* Congestion window validation. (RFC2861) */
1356 static void tcp_cwnd_validate(struct sock
*sk
)
1358 struct tcp_sock
*tp
= tcp_sk(sk
);
1360 if (tp
->packets_out
>= tp
->snd_cwnd
) {
1361 /* Network is feed fully. */
1362 tp
->snd_cwnd_used
= 0;
1363 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1365 /* Network starves. */
1366 if (tp
->packets_out
> tp
->snd_cwnd_used
)
1367 tp
->snd_cwnd_used
= tp
->packets_out
;
1369 if (sysctl_tcp_slow_start_after_idle
&&
1370 (s32
)(tcp_time_stamp
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
)
1371 tcp_cwnd_application_limited(sk
);
1375 /* Returns the portion of skb which can be sent right away without
1376 * introducing MSS oddities to segment boundaries. In rare cases where
1377 * mss_now != mss_cache, we will request caller to create a small skb
1378 * per input skb which could be mostly avoided here (if desired).
1380 * We explicitly want to create a request for splitting write queue tail
1381 * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1382 * thus all the complexity (cwnd_len is always MSS multiple which we
1383 * return whenever allowed by the other factors). Basically we need the
1384 * modulo only when the receiver window alone is the limiting factor or
1385 * when we would be allowed to send the split-due-to-Nagle skb fully.
1387 static unsigned int tcp_mss_split_point(const struct sock
*sk
, const struct sk_buff
*skb
,
1388 unsigned int mss_now
, unsigned int max_segs
)
1390 const struct tcp_sock
*tp
= tcp_sk(sk
);
1391 u32 needed
, window
, max_len
;
1393 window
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1394 max_len
= mss_now
* max_segs
;
1396 if (likely(max_len
<= window
&& skb
!= tcp_write_queue_tail(sk
)))
1399 needed
= min(skb
->len
, window
);
1401 if (max_len
<= needed
)
1404 return needed
- needed
% mss_now
;
1407 /* Can at least one segment of SKB be sent right now, according to the
1408 * congestion window rules? If so, return how many segments are allowed.
1410 static inline unsigned int tcp_cwnd_test(const struct tcp_sock
*tp
,
1411 const struct sk_buff
*skb
)
1413 u32 in_flight
, cwnd
;
1415 /* Don't be strict about the congestion window for the final FIN. */
1416 if ((TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
) &&
1417 tcp_skb_pcount(skb
) == 1)
1420 in_flight
= tcp_packets_in_flight(tp
);
1421 cwnd
= tp
->snd_cwnd
;
1422 if (in_flight
< cwnd
)
1423 return (cwnd
- in_flight
);
1428 /* Initialize TSO state of a skb.
1429 * This must be invoked the first time we consider transmitting
1430 * SKB onto the wire.
1432 static int tcp_init_tso_segs(const struct sock
*sk
, struct sk_buff
*skb
,
1433 unsigned int mss_now
)
1435 int tso_segs
= tcp_skb_pcount(skb
);
1437 if (!tso_segs
|| (tso_segs
> 1 && tcp_skb_mss(skb
) != mss_now
)) {
1438 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1439 tso_segs
= tcp_skb_pcount(skb
);
1444 /* Minshall's variant of the Nagle send check. */
1445 static inline bool tcp_minshall_check(const struct tcp_sock
*tp
)
1447 return after(tp
->snd_sml
, tp
->snd_una
) &&
1448 !after(tp
->snd_sml
, tp
->snd_nxt
);
1451 /* Return false, if packet can be sent now without violation Nagle's rules:
1452 * 1. It is full sized.
1453 * 2. Or it contains FIN. (already checked by caller)
1454 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1455 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1456 * With Minshall's modification: all sent small packets are ACKed.
1458 static inline bool tcp_nagle_check(const struct tcp_sock
*tp
,
1459 const struct sk_buff
*skb
,
1460 unsigned int mss_now
, int nonagle
)
1462 return skb
->len
< mss_now
&&
1463 ((nonagle
& TCP_NAGLE_CORK
) ||
1464 (!nonagle
&& tp
->packets_out
&& tcp_minshall_check(tp
)));
1467 /* Return true if the Nagle test allows this packet to be
1470 static inline bool tcp_nagle_test(const struct tcp_sock
*tp
, const struct sk_buff
*skb
,
1471 unsigned int cur_mss
, int nonagle
)
1473 /* Nagle rule does not apply to frames, which sit in the middle of the
1474 * write_queue (they have no chances to get new data).
1476 * This is implemented in the callers, where they modify the 'nonagle'
1477 * argument based upon the location of SKB in the send queue.
1479 if (nonagle
& TCP_NAGLE_PUSH
)
1482 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1483 if (tcp_urg_mode(tp
) || (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
))
1486 if (!tcp_nagle_check(tp
, skb
, cur_mss
, nonagle
))
1492 /* Does at least the first segment of SKB fit into the send window? */
1493 static bool tcp_snd_wnd_test(const struct tcp_sock
*tp
,
1494 const struct sk_buff
*skb
,
1495 unsigned int cur_mss
)
1497 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1499 if (skb
->len
> cur_mss
)
1500 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
1502 return !after(end_seq
, tcp_wnd_end(tp
));
1505 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1506 * should be put on the wire right now. If so, it returns the number of
1507 * packets allowed by the congestion window.
1509 static unsigned int tcp_snd_test(const struct sock
*sk
, struct sk_buff
*skb
,
1510 unsigned int cur_mss
, int nonagle
)
1512 const struct tcp_sock
*tp
= tcp_sk(sk
);
1513 unsigned int cwnd_quota
;
1515 tcp_init_tso_segs(sk
, skb
, cur_mss
);
1517 if (!tcp_nagle_test(tp
, skb
, cur_mss
, nonagle
))
1520 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1521 if (cwnd_quota
&& !tcp_snd_wnd_test(tp
, skb
, cur_mss
))
1527 /* Test if sending is allowed right now. */
1528 bool tcp_may_send_now(struct sock
*sk
)
1530 const struct tcp_sock
*tp
= tcp_sk(sk
);
1531 struct sk_buff
*skb
= tcp_send_head(sk
);
1534 tcp_snd_test(sk
, skb
, tcp_current_mss(sk
),
1535 (tcp_skb_is_last(sk
, skb
) ?
1536 tp
->nonagle
: TCP_NAGLE_PUSH
));
1539 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1540 * which is put after SKB on the list. It is very much like
1541 * tcp_fragment() except that it may make several kinds of assumptions
1542 * in order to speed up the splitting operation. In particular, we
1543 * know that all the data is in scatter-gather pages, and that the
1544 * packet has never been sent out before (and thus is not cloned).
1546 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
,
1547 unsigned int mss_now
, gfp_t gfp
)
1549 struct sk_buff
*buff
;
1550 int nlen
= skb
->len
- len
;
1553 /* All of a TSO frame must be composed of paged data. */
1554 if (skb
->len
!= skb
->data_len
)
1555 return tcp_fragment(sk
, skb
, len
, mss_now
);
1557 buff
= sk_stream_alloc_skb(sk
, 0, gfp
);
1558 if (unlikely(buff
== NULL
))
1561 sk
->sk_wmem_queued
+= buff
->truesize
;
1562 sk_mem_charge(sk
, buff
->truesize
);
1563 buff
->truesize
+= nlen
;
1564 skb
->truesize
-= nlen
;
1566 /* Correct the sequence numbers. */
1567 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1568 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1569 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1571 /* PSH and FIN should only be set in the second packet. */
1572 flags
= TCP_SKB_CB(skb
)->tcp_flags
;
1573 TCP_SKB_CB(skb
)->tcp_flags
= flags
& ~(TCPHDR_FIN
| TCPHDR_PSH
);
1574 TCP_SKB_CB(buff
)->tcp_flags
= flags
;
1576 /* This packet was never sent out yet, so no SACK bits. */
1577 TCP_SKB_CB(buff
)->sacked
= 0;
1579 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_PARTIAL
;
1580 skb_split(skb
, buff
, len
);
1582 /* Fix up tso_factor for both original and new SKB. */
1583 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1584 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1586 /* Link BUFF into the send queue. */
1587 skb_header_release(buff
);
1588 tcp_insert_write_queue_after(skb
, buff
, sk
);
1593 /* Try to defer sending, if possible, in order to minimize the amount
1594 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1596 * This algorithm is from John Heffner.
1598 static bool tcp_tso_should_defer(struct sock
*sk
, struct sk_buff
*skb
)
1600 struct tcp_sock
*tp
= tcp_sk(sk
);
1601 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1602 u32 send_win
, cong_win
, limit
, in_flight
;
1605 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
1608 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
1611 /* Defer for less than two clock ticks. */
1612 if (tp
->tso_deferred
&&
1613 (((u32
)jiffies
<< 1) >> 1) - (tp
->tso_deferred
>> 1) > 1)
1616 in_flight
= tcp_packets_in_flight(tp
);
1618 BUG_ON(tcp_skb_pcount(skb
) <= 1 || (tp
->snd_cwnd
<= in_flight
));
1620 send_win
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1622 /* From in_flight test above, we know that cwnd > in_flight. */
1623 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
1625 limit
= min(send_win
, cong_win
);
1627 /* If a full-sized TSO skb can be sent, do it. */
1628 if (limit
>= min_t(unsigned int, sk
->sk_gso_max_size
,
1629 sk
->sk_gso_max_segs
* tp
->mss_cache
))
1632 /* Middle in queue won't get any more data, full sendable already? */
1633 if ((skb
!= tcp_write_queue_tail(sk
)) && (limit
>= skb
->len
))
1636 win_divisor
= ACCESS_ONCE(sysctl_tcp_tso_win_divisor
);
1638 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
1640 /* If at least some fraction of a window is available,
1643 chunk
/= win_divisor
;
1647 /* Different approach, try not to defer past a single
1648 * ACK. Receiver should ACK every other full sized
1649 * frame, so if we have space for more than 3 frames
1652 if (limit
> tcp_max_tso_deferred_mss(tp
) * tp
->mss_cache
)
1656 /* Ok, it looks like it is advisable to defer.
1657 * Do not rearm the timer if already set to not break TCP ACK clocking.
1659 if (!tp
->tso_deferred
)
1660 tp
->tso_deferred
= 1 | (jiffies
<< 1);
1665 tp
->tso_deferred
= 0;
1669 /* Create a new MTU probe if we are ready.
1670 * MTU probe is regularly attempting to increase the path MTU by
1671 * deliberately sending larger packets. This discovers routing
1672 * changes resulting in larger path MTUs.
1674 * Returns 0 if we should wait to probe (no cwnd available),
1675 * 1 if a probe was sent,
1678 static int tcp_mtu_probe(struct sock
*sk
)
1680 struct tcp_sock
*tp
= tcp_sk(sk
);
1681 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1682 struct sk_buff
*skb
, *nskb
, *next
;
1689 /* Not currently probing/verifying,
1691 * have enough cwnd, and
1692 * not SACKing (the variable headers throw things off) */
1693 if (!icsk
->icsk_mtup
.enabled
||
1694 icsk
->icsk_mtup
.probe_size
||
1695 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
||
1696 tp
->snd_cwnd
< 11 ||
1697 tp
->rx_opt
.num_sacks
|| tp
->rx_opt
.dsack
)
1700 /* Very simple search strategy: just double the MSS. */
1701 mss_now
= tcp_current_mss(sk
);
1702 probe_size
= 2 * tp
->mss_cache
;
1703 size_needed
= probe_size
+ (tp
->reordering
+ 1) * tp
->mss_cache
;
1704 if (probe_size
> tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_high
)) {
1705 /* TODO: set timer for probe_converge_event */
1709 /* Have enough data in the send queue to probe? */
1710 if (tp
->write_seq
- tp
->snd_nxt
< size_needed
)
1713 if (tp
->snd_wnd
< size_needed
)
1715 if (after(tp
->snd_nxt
+ size_needed
, tcp_wnd_end(tp
)))
1718 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1719 if (tcp_packets_in_flight(tp
) + 2 > tp
->snd_cwnd
) {
1720 if (!tcp_packets_in_flight(tp
))
1726 /* We're allowed to probe. Build it now. */
1727 if ((nskb
= sk_stream_alloc_skb(sk
, probe_size
, GFP_ATOMIC
)) == NULL
)
1729 sk
->sk_wmem_queued
+= nskb
->truesize
;
1730 sk_mem_charge(sk
, nskb
->truesize
);
1732 skb
= tcp_send_head(sk
);
1734 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(skb
)->seq
;
1735 TCP_SKB_CB(nskb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ probe_size
;
1736 TCP_SKB_CB(nskb
)->tcp_flags
= TCPHDR_ACK
;
1737 TCP_SKB_CB(nskb
)->sacked
= 0;
1739 nskb
->ip_summed
= skb
->ip_summed
;
1741 tcp_insert_write_queue_before(nskb
, skb
, sk
);
1744 tcp_for_write_queue_from_safe(skb
, next
, sk
) {
1745 copy
= min_t(int, skb
->len
, probe_size
- len
);
1746 if (nskb
->ip_summed
)
1747 skb_copy_bits(skb
, 0, skb_put(nskb
, copy
), copy
);
1749 nskb
->csum
= skb_copy_and_csum_bits(skb
, 0,
1750 skb_put(nskb
, copy
),
1753 if (skb
->len
<= copy
) {
1754 /* We've eaten all the data from this skb.
1756 TCP_SKB_CB(nskb
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
;
1757 tcp_unlink_write_queue(skb
, sk
);
1758 sk_wmem_free_skb(sk
, skb
);
1760 TCP_SKB_CB(nskb
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
&
1761 ~(TCPHDR_FIN
|TCPHDR_PSH
);
1762 if (!skb_shinfo(skb
)->nr_frags
) {
1763 skb_pull(skb
, copy
);
1764 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1765 skb
->csum
= csum_partial(skb
->data
,
1768 __pskb_trim_head(skb
, copy
);
1769 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1771 TCP_SKB_CB(skb
)->seq
+= copy
;
1776 if (len
>= probe_size
)
1779 tcp_init_tso_segs(sk
, nskb
, nskb
->len
);
1781 /* We're ready to send. If this fails, the probe will
1782 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1783 TCP_SKB_CB(nskb
)->when
= tcp_time_stamp
;
1784 if (!tcp_transmit_skb(sk
, nskb
, 1, GFP_ATOMIC
)) {
1785 /* Decrement cwnd here because we are sending
1786 * effectively two packets. */
1788 tcp_event_new_data_sent(sk
, nskb
);
1790 icsk
->icsk_mtup
.probe_size
= tcp_mss_to_mtu(sk
, nskb
->len
);
1791 tp
->mtu_probe
.probe_seq_start
= TCP_SKB_CB(nskb
)->seq
;
1792 tp
->mtu_probe
.probe_seq_end
= TCP_SKB_CB(nskb
)->end_seq
;
1800 /* This routine writes packets to the network. It advances the
1801 * send_head. This happens as incoming acks open up the remote
1804 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1805 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1806 * account rare use of URG, this is not a big flaw.
1808 * Send at most one packet when push_one > 0. Temporarily ignore
1809 * cwnd limit to force at most one packet out when push_one == 2.
1811 * Returns true, if no segments are in flight and we have queued segments,
1812 * but cannot send anything now because of SWS or another problem.
1814 static bool tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
,
1815 int push_one
, gfp_t gfp
)
1817 struct tcp_sock
*tp
= tcp_sk(sk
);
1818 struct sk_buff
*skb
;
1819 unsigned int tso_segs
, sent_pkts
;
1826 /* Do MTU probing. */
1827 result
= tcp_mtu_probe(sk
);
1830 } else if (result
> 0) {
1835 while ((skb
= tcp_send_head(sk
))) {
1839 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1842 if (unlikely(tp
->repair
) && tp
->repair_queue
== TCP_SEND_QUEUE
)
1843 goto repair
; /* Skip network transmission */
1845 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1848 /* Force out a loss probe pkt. */
1854 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
)))
1857 if (tso_segs
== 1) {
1858 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
1859 (tcp_skb_is_last(sk
, skb
) ?
1860 nonagle
: TCP_NAGLE_PUSH
))))
1863 if (!push_one
&& tcp_tso_should_defer(sk
, skb
))
1867 /* TSQ : sk_wmem_alloc accounts skb truesize,
1868 * including skb overhead. But thats OK.
1870 if (atomic_read(&sk
->sk_wmem_alloc
) >= sysctl_tcp_limit_output_bytes
) {
1871 set_bit(TSQ_THROTTLED
, &tp
->tsq_flags
);
1875 if (tso_segs
> 1 && !tcp_urg_mode(tp
))
1876 limit
= tcp_mss_split_point(sk
, skb
, mss_now
,
1879 sk
->sk_gso_max_segs
));
1881 if (skb
->len
> limit
&&
1882 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
, gfp
)))
1885 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1887 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, gfp
)))
1891 /* Advance the send_head. This one is sent out.
1892 * This call will increment packets_out.
1894 tcp_event_new_data_sent(sk
, skb
);
1896 tcp_minshall_update(tp
, mss_now
, skb
);
1897 sent_pkts
+= tcp_skb_pcount(skb
);
1903 if (likely(sent_pkts
)) {
1904 if (tcp_in_cwnd_reduction(sk
))
1905 tp
->prr_out
+= sent_pkts
;
1907 /* Send one loss probe per tail loss episode. */
1909 tcp_schedule_loss_probe(sk
);
1910 tcp_cwnd_validate(sk
);
1913 return (push_one
== 2) || (!tp
->packets_out
&& tcp_send_head(sk
));
1916 bool tcp_schedule_loss_probe(struct sock
*sk
)
1918 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1919 struct tcp_sock
*tp
= tcp_sk(sk
);
1920 u32 timeout
, tlp_time_stamp
, rto_time_stamp
;
1921 u32 rtt
= tp
->srtt
>> 3;
1923 if (WARN_ON(icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
))
1925 /* No consecutive loss probes. */
1926 if (WARN_ON(icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
)) {
1930 /* Don't do any loss probe on a Fast Open connection before 3WHS
1933 if (sk
->sk_state
== TCP_SYN_RECV
)
1936 /* TLP is only scheduled when next timer event is RTO. */
1937 if (icsk
->icsk_pending
!= ICSK_TIME_RETRANS
)
1940 /* Schedule a loss probe in 2*RTT for SACK capable connections
1941 * in Open state, that are either limited by cwnd or application.
1943 if (sysctl_tcp_early_retrans
< 3 || !rtt
|| !tp
->packets_out
||
1944 !tcp_is_sack(tp
) || inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
)
1947 if ((tp
->snd_cwnd
> tcp_packets_in_flight(tp
)) &&
1951 /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
1952 * for delayed ack when there's one outstanding packet.
1955 if (tp
->packets_out
== 1)
1956 timeout
= max_t(u32
, timeout
,
1957 (rtt
+ (rtt
>> 1) + TCP_DELACK_MAX
));
1958 timeout
= max_t(u32
, timeout
, msecs_to_jiffies(10));
1960 /* If RTO is shorter, just schedule TLP in its place. */
1961 tlp_time_stamp
= tcp_time_stamp
+ timeout
;
1962 rto_time_stamp
= (u32
)inet_csk(sk
)->icsk_timeout
;
1963 if ((s32
)(tlp_time_stamp
- rto_time_stamp
) > 0) {
1964 s32 delta
= rto_time_stamp
- tcp_time_stamp
;
1969 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_LOSS_PROBE
, timeout
,
1974 /* When probe timeout (PTO) fires, send a new segment if one exists, else
1975 * retransmit the last segment.
1977 void tcp_send_loss_probe(struct sock
*sk
)
1979 struct tcp_sock
*tp
= tcp_sk(sk
);
1980 struct sk_buff
*skb
;
1982 int mss
= tcp_current_mss(sk
);
1985 if (tcp_send_head(sk
) != NULL
) {
1986 err
= tcp_write_xmit(sk
, mss
, TCP_NAGLE_OFF
, 2, GFP_ATOMIC
);
1990 /* At most one outstanding TLP retransmission. */
1991 if (tp
->tlp_high_seq
)
1994 /* Retransmit last segment. */
1995 skb
= tcp_write_queue_tail(sk
);
1999 pcount
= tcp_skb_pcount(skb
);
2000 if (WARN_ON(!pcount
))
2003 if ((pcount
> 1) && (skb
->len
> (pcount
- 1) * mss
)) {
2004 if (unlikely(tcp_fragment(sk
, skb
, (pcount
- 1) * mss
, mss
)))
2006 skb
= tcp_write_queue_tail(sk
);
2009 if (WARN_ON(!skb
|| !tcp_skb_pcount(skb
)))
2012 /* Probe with zero data doesn't trigger fast recovery. */
2014 err
= __tcp_retransmit_skb(sk
, skb
);
2016 /* Record snd_nxt for loss detection. */
2018 tp
->tlp_high_seq
= tp
->snd_nxt
;
2021 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2022 inet_csk(sk
)->icsk_rto
,
2026 NET_INC_STATS_BH(sock_net(sk
),
2027 LINUX_MIB_TCPLOSSPROBES
);
2031 /* Push out any pending frames which were held back due to
2032 * TCP_CORK or attempt at coalescing tiny packets.
2033 * The socket must be locked by the caller.
2035 void __tcp_push_pending_frames(struct sock
*sk
, unsigned int cur_mss
,
2038 /* If we are closed, the bytes will have to remain here.
2039 * In time closedown will finish, we empty the write queue and
2040 * all will be happy.
2042 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
2045 if (tcp_write_xmit(sk
, cur_mss
, nonagle
, 0,
2046 sk_gfp_atomic(sk
, GFP_ATOMIC
)))
2047 tcp_check_probe_timer(sk
);
2050 /* Send _single_ skb sitting at the send head. This function requires
2051 * true push pending frames to setup probe timer etc.
2053 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
2055 struct sk_buff
*skb
= tcp_send_head(sk
);
2057 BUG_ON(!skb
|| skb
->len
< mss_now
);
2059 tcp_write_xmit(sk
, mss_now
, TCP_NAGLE_PUSH
, 1, sk
->sk_allocation
);
2062 /* This function returns the amount that we can raise the
2063 * usable window based on the following constraints
2065 * 1. The window can never be shrunk once it is offered (RFC 793)
2066 * 2. We limit memory per socket
2069 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2070 * RECV.NEXT + RCV.WIN fixed until:
2071 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2073 * i.e. don't raise the right edge of the window until you can raise
2074 * it at least MSS bytes.
2076 * Unfortunately, the recommended algorithm breaks header prediction,
2077 * since header prediction assumes th->window stays fixed.
2079 * Strictly speaking, keeping th->window fixed violates the receiver
2080 * side SWS prevention criteria. The problem is that under this rule
2081 * a stream of single byte packets will cause the right side of the
2082 * window to always advance by a single byte.
2084 * Of course, if the sender implements sender side SWS prevention
2085 * then this will not be a problem.
2087 * BSD seems to make the following compromise:
2089 * If the free space is less than the 1/4 of the maximum
2090 * space available and the free space is less than 1/2 mss,
2091 * then set the window to 0.
2092 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2093 * Otherwise, just prevent the window from shrinking
2094 * and from being larger than the largest representable value.
2096 * This prevents incremental opening of the window in the regime
2097 * where TCP is limited by the speed of the reader side taking
2098 * data out of the TCP receive queue. It does nothing about
2099 * those cases where the window is constrained on the sender side
2100 * because the pipeline is full.
2102 * BSD also seems to "accidentally" limit itself to windows that are a
2103 * multiple of MSS, at least until the free space gets quite small.
2104 * This would appear to be a side effect of the mbuf implementation.
2105 * Combining these two algorithms results in the observed behavior
2106 * of having a fixed window size at almost all times.
2108 * Below we obtain similar behavior by forcing the offered window to
2109 * a multiple of the mss when it is feasible to do so.
2111 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2112 * Regular options like TIMESTAMP are taken into account.
2114 u32
__tcp_select_window(struct sock
*sk
)
2116 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2117 struct tcp_sock
*tp
= tcp_sk(sk
);
2118 /* MSS for the peer's data. Previous versions used mss_clamp
2119 * here. I don't know if the value based on our guesses
2120 * of peer's MSS is better for the performance. It's more correct
2121 * but may be worse for the performance because of rcv_mss
2122 * fluctuations. --SAW 1998/11/1
2124 int mss
= icsk
->icsk_ack
.rcv_mss
;
2125 int free_space
= tcp_space(sk
);
2126 int full_space
= min_t(int, tp
->window_clamp
, tcp_full_space(sk
));
2129 if (mss
> full_space
)
2132 if (free_space
< (full_space
>> 1)) {
2133 icsk
->icsk_ack
.quick
= 0;
2135 if (sk_under_memory_pressure(sk
))
2136 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
,
2139 if (free_space
< mss
)
2143 if (free_space
> tp
->rcv_ssthresh
)
2144 free_space
= tp
->rcv_ssthresh
;
2146 /* Don't do rounding if we are using window scaling, since the
2147 * scaled window will not line up with the MSS boundary anyway.
2149 window
= tp
->rcv_wnd
;
2150 if (tp
->rx_opt
.rcv_wscale
) {
2151 window
= free_space
;
2153 /* Advertise enough space so that it won't get scaled away.
2154 * Import case: prevent zero window announcement if
2155 * 1<<rcv_wscale > mss.
2157 if (((window
>> tp
->rx_opt
.rcv_wscale
) << tp
->rx_opt
.rcv_wscale
) != window
)
2158 window
= (((window
>> tp
->rx_opt
.rcv_wscale
) + 1)
2159 << tp
->rx_opt
.rcv_wscale
);
2161 /* Get the largest window that is a nice multiple of mss.
2162 * Window clamp already applied above.
2163 * If our current window offering is within 1 mss of the
2164 * free space we just keep it. This prevents the divide
2165 * and multiply from happening most of the time.
2166 * We also don't do any window rounding when the free space
2169 if (window
<= free_space
- mss
|| window
> free_space
)
2170 window
= (free_space
/ mss
) * mss
;
2171 else if (mss
== full_space
&&
2172 free_space
> window
+ (full_space
>> 1))
2173 window
= free_space
;
2179 /* Collapses two adjacent SKB's during retransmission. */
2180 static void tcp_collapse_retrans(struct sock
*sk
, struct sk_buff
*skb
)
2182 struct tcp_sock
*tp
= tcp_sk(sk
);
2183 struct sk_buff
*next_skb
= tcp_write_queue_next(sk
, skb
);
2184 int skb_size
, next_skb_size
;
2186 skb_size
= skb
->len
;
2187 next_skb_size
= next_skb
->len
;
2189 BUG_ON(tcp_skb_pcount(skb
) != 1 || tcp_skb_pcount(next_skb
) != 1);
2191 tcp_highest_sack_combine(sk
, next_skb
, skb
);
2193 tcp_unlink_write_queue(next_skb
, sk
);
2195 skb_copy_from_linear_data(next_skb
, skb_put(skb
, next_skb_size
),
2198 if (next_skb
->ip_summed
== CHECKSUM_PARTIAL
)
2199 skb
->ip_summed
= CHECKSUM_PARTIAL
;
2201 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
2202 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
2204 /* Update sequence range on original skb. */
2205 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
2207 /* Merge over control information. This moves PSH/FIN etc. over */
2208 TCP_SKB_CB(skb
)->tcp_flags
|= TCP_SKB_CB(next_skb
)->tcp_flags
;
2210 /* All done, get rid of second SKB and account for it so
2211 * packet counting does not break.
2213 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
& TCPCB_EVER_RETRANS
;
2215 /* changed transmit queue under us so clear hints */
2216 tcp_clear_retrans_hints_partial(tp
);
2217 if (next_skb
== tp
->retransmit_skb_hint
)
2218 tp
->retransmit_skb_hint
= skb
;
2220 tcp_adjust_pcount(sk
, next_skb
, tcp_skb_pcount(next_skb
));
2222 sk_wmem_free_skb(sk
, next_skb
);
2225 /* Check if coalescing SKBs is legal. */
2226 static bool tcp_can_collapse(const struct sock
*sk
, const struct sk_buff
*skb
)
2228 if (tcp_skb_pcount(skb
) > 1)
2230 /* TODO: SACK collapsing could be used to remove this condition */
2231 if (skb_shinfo(skb
)->nr_frags
!= 0)
2233 if (skb_cloned(skb
))
2235 if (skb
== tcp_send_head(sk
))
2237 /* Some heurestics for collapsing over SACK'd could be invented */
2238 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
2244 /* Collapse packets in the retransmit queue to make to create
2245 * less packets on the wire. This is only done on retransmission.
2247 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*to
,
2250 struct tcp_sock
*tp
= tcp_sk(sk
);
2251 struct sk_buff
*skb
= to
, *tmp
;
2254 if (!sysctl_tcp_retrans_collapse
)
2256 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)
2259 tcp_for_write_queue_from_safe(skb
, tmp
, sk
) {
2260 if (!tcp_can_collapse(sk
, skb
))
2272 /* Punt if not enough space exists in the first SKB for
2273 * the data in the second
2275 if (skb
->len
> skb_availroom(to
))
2278 if (after(TCP_SKB_CB(skb
)->end_seq
, tcp_wnd_end(tp
)))
2281 tcp_collapse_retrans(sk
, to
);
2285 /* This retransmits one SKB. Policy decisions and retransmit queue
2286 * state updates are done by the caller. Returns non-zero if an
2287 * error occurred which prevented the send.
2289 int __tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
2291 struct tcp_sock
*tp
= tcp_sk(sk
);
2292 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2293 unsigned int cur_mss
;
2295 /* Inconslusive MTU probe */
2296 if (icsk
->icsk_mtup
.probe_size
) {
2297 icsk
->icsk_mtup
.probe_size
= 0;
2300 /* Do not sent more than we queued. 1/4 is reserved for possible
2301 * copying overhead: fragmentation, tunneling, mangling etc.
2303 if (atomic_read(&sk
->sk_wmem_alloc
) >
2304 min(sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2), sk
->sk_sndbuf
))
2307 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
2308 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
2310 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
2314 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
2315 return -EHOSTUNREACH
; /* Routing failure or similar. */
2317 cur_mss
= tcp_current_mss(sk
);
2319 /* If receiver has shrunk his window, and skb is out of
2320 * new window, do not retransmit it. The exception is the
2321 * case, when window is shrunk to zero. In this case
2322 * our retransmit serves as a zero window probe.
2324 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
)) &&
2325 TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
2328 if (skb
->len
> cur_mss
) {
2329 if (tcp_fragment(sk
, skb
, cur_mss
, cur_mss
))
2330 return -ENOMEM
; /* We'll try again later. */
2332 int oldpcount
= tcp_skb_pcount(skb
);
2334 if (unlikely(oldpcount
> 1)) {
2335 tcp_init_tso_segs(sk
, skb
, cur_mss
);
2336 tcp_adjust_pcount(sk
, skb
, oldpcount
- tcp_skb_pcount(skb
));
2340 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
2342 /* Some Solaris stacks overoptimize and ignore the FIN on a
2343 * retransmit when old data is attached. So strip it off
2344 * since it is cheap to do so and saves bytes on the network.
2347 (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
) &&
2348 tp
->snd_una
== (TCP_SKB_CB(skb
)->end_seq
- 1)) {
2349 if (!pskb_trim(skb
, 0)) {
2350 /* Reuse, even though it does some unnecessary work */
2351 tcp_init_nondata_skb(skb
, TCP_SKB_CB(skb
)->end_seq
- 1,
2352 TCP_SKB_CB(skb
)->tcp_flags
);
2353 skb
->ip_summed
= CHECKSUM_NONE
;
2357 /* Make a copy, if the first transmission SKB clone we made
2358 * is still in somebody's hands, else make a clone.
2360 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2362 /* make sure skb->data is aligned on arches that require it
2363 * and check if ack-trimming & collapsing extended the headroom
2364 * beyond what csum_start can cover.
2366 if (unlikely((NET_IP_ALIGN
&& ((unsigned long)skb
->data
& 3)) ||
2367 skb_headroom(skb
) >= 0xFFFF)) {
2368 struct sk_buff
*nskb
= __pskb_copy(skb
, MAX_TCP_HEADER
,
2370 return nskb
? tcp_transmit_skb(sk
, nskb
, 0, GFP_ATOMIC
) :
2373 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2377 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
2379 struct tcp_sock
*tp
= tcp_sk(sk
);
2380 int err
= __tcp_retransmit_skb(sk
, skb
);
2383 /* Update global TCP statistics. */
2384 TCP_INC_STATS(sock_net(sk
), TCP_MIB_RETRANSSEGS
);
2386 tp
->total_retrans
++;
2388 #if FASTRETRANS_DEBUG > 0
2389 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2390 net_dbg_ratelimited("retrans_out leaked\n");
2393 if (!tp
->retrans_out
)
2394 tp
->lost_retrans_low
= tp
->snd_nxt
;
2395 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
2396 tp
->retrans_out
+= tcp_skb_pcount(skb
);
2398 /* Save stamp of the first retransmit. */
2399 if (!tp
->retrans_stamp
)
2400 tp
->retrans_stamp
= TCP_SKB_CB(skb
)->when
;
2402 tp
->undo_retrans
+= tcp_skb_pcount(skb
);
2404 /* snd_nxt is stored to detect loss of retransmitted segment,
2405 * see tcp_input.c tcp_sacktag_write_queue().
2407 TCP_SKB_CB(skb
)->ack_seq
= tp
->snd_nxt
;
2412 /* Check if we forward retransmits are possible in the current
2413 * window/congestion state.
2415 static bool tcp_can_forward_retransmit(struct sock
*sk
)
2417 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2418 const struct tcp_sock
*tp
= tcp_sk(sk
);
2420 /* Forward retransmissions are possible only during Recovery. */
2421 if (icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
2424 /* No forward retransmissions in Reno are possible. */
2425 if (tcp_is_reno(tp
))
2428 /* Yeah, we have to make difficult choice between forward transmission
2429 * and retransmission... Both ways have their merits...
2431 * For now we do not retransmit anything, while we have some new
2432 * segments to send. In the other cases, follow rule 3 for
2433 * NextSeg() specified in RFC3517.
2436 if (tcp_may_send_now(sk
))
2442 /* This gets called after a retransmit timeout, and the initially
2443 * retransmitted data is acknowledged. It tries to continue
2444 * resending the rest of the retransmit queue, until either
2445 * we've sent it all or the congestion window limit is reached.
2446 * If doing SACK, the first ACK which comes back for a timeout
2447 * based retransmit packet might feed us FACK information again.
2448 * If so, we use it to avoid unnecessarily retransmissions.
2450 void tcp_xmit_retransmit_queue(struct sock
*sk
)
2452 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2453 struct tcp_sock
*tp
= tcp_sk(sk
);
2454 struct sk_buff
*skb
;
2455 struct sk_buff
*hole
= NULL
;
2458 int fwd_rexmitting
= 0;
2460 if (!tp
->packets_out
)
2464 tp
->retransmit_high
= tp
->snd_una
;
2466 if (tp
->retransmit_skb_hint
) {
2467 skb
= tp
->retransmit_skb_hint
;
2468 last_lost
= TCP_SKB_CB(skb
)->end_seq
;
2469 if (after(last_lost
, tp
->retransmit_high
))
2470 last_lost
= tp
->retransmit_high
;
2472 skb
= tcp_write_queue_head(sk
);
2473 last_lost
= tp
->snd_una
;
2476 tcp_for_write_queue_from(skb
, sk
) {
2477 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
2479 if (skb
== tcp_send_head(sk
))
2481 /* we could do better than to assign each time */
2483 tp
->retransmit_skb_hint
= skb
;
2485 /* Assume this retransmit will generate
2486 * only one packet for congestion window
2487 * calculation purposes. This works because
2488 * tcp_retransmit_skb() will chop up the
2489 * packet to be MSS sized and all the
2490 * packet counting works out.
2492 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
2495 if (fwd_rexmitting
) {
2497 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_highest_sack_seq(tp
)))
2499 mib_idx
= LINUX_MIB_TCPFORWARDRETRANS
;
2501 } else if (!before(TCP_SKB_CB(skb
)->seq
, tp
->retransmit_high
)) {
2502 tp
->retransmit_high
= last_lost
;
2503 if (!tcp_can_forward_retransmit(sk
))
2505 /* Backtrack if necessary to non-L'ed skb */
2513 } else if (!(sacked
& TCPCB_LOST
)) {
2514 if (hole
== NULL
&& !(sacked
& (TCPCB_SACKED_RETRANS
|TCPCB_SACKED_ACKED
)))
2519 last_lost
= TCP_SKB_CB(skb
)->end_seq
;
2520 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
2521 mib_idx
= LINUX_MIB_TCPFASTRETRANS
;
2523 mib_idx
= LINUX_MIB_TCPSLOWSTARTRETRANS
;
2526 if (sacked
& (TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))
2529 if (tcp_retransmit_skb(sk
, skb
)) {
2530 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRETRANSFAIL
);
2533 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2535 if (tcp_in_cwnd_reduction(sk
))
2536 tp
->prr_out
+= tcp_skb_pcount(skb
);
2538 if (skb
== tcp_write_queue_head(sk
))
2539 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2540 inet_csk(sk
)->icsk_rto
,
2545 /* Send a fin. The caller locks the socket for us. This cannot be
2546 * allowed to fail queueing a FIN frame under any circumstances.
2548 void tcp_send_fin(struct sock
*sk
)
2550 struct tcp_sock
*tp
= tcp_sk(sk
);
2551 struct sk_buff
*skb
= tcp_write_queue_tail(sk
);
2554 /* Optimization, tack on the FIN if we have a queue of
2555 * unsent frames. But be careful about outgoing SACKS
2558 mss_now
= tcp_current_mss(sk
);
2560 if (tcp_send_head(sk
) != NULL
) {
2561 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_FIN
;
2562 TCP_SKB_CB(skb
)->end_seq
++;
2565 /* Socket is locked, keep trying until memory is available. */
2567 skb
= alloc_skb_fclone(MAX_TCP_HEADER
,
2574 /* Reserve space for headers and prepare control bits. */
2575 skb_reserve(skb
, MAX_TCP_HEADER
);
2576 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2577 tcp_init_nondata_skb(skb
, tp
->write_seq
,
2578 TCPHDR_ACK
| TCPHDR_FIN
);
2579 tcp_queue_skb(sk
, skb
);
2581 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_OFF
);
2584 /* We get here when a process closes a file descriptor (either due to
2585 * an explicit close() or as a byproduct of exit()'ing) and there
2586 * was unread data in the receive queue. This behavior is recommended
2587 * by RFC 2525, section 2.17. -DaveM
2589 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
2591 struct sk_buff
*skb
;
2593 /* NOTE: No TCP options attached and we never retransmit this. */
2594 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
2596 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
2600 /* Reserve space for headers and prepare control bits. */
2601 skb_reserve(skb
, MAX_TCP_HEADER
);
2602 tcp_init_nondata_skb(skb
, tcp_acceptable_seq(sk
),
2603 TCPHDR_ACK
| TCPHDR_RST
);
2605 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2606 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
2607 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
2609 TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTRSTS
);
2612 /* Send a crossed SYN-ACK during socket establishment.
2613 * WARNING: This routine must only be called when we have already sent
2614 * a SYN packet that crossed the incoming SYN that caused this routine
2615 * to get called. If this assumption fails then the initial rcv_wnd
2616 * and rcv_wscale values will not be correct.
2618 int tcp_send_synack(struct sock
*sk
)
2620 struct sk_buff
*skb
;
2622 skb
= tcp_write_queue_head(sk
);
2623 if (skb
== NULL
|| !(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)) {
2624 pr_debug("%s: wrong queue state\n", __func__
);
2627 if (!(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_ACK
)) {
2628 if (skb_cloned(skb
)) {
2629 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
2632 tcp_unlink_write_queue(skb
, sk
);
2633 skb_header_release(nskb
);
2634 __tcp_add_write_queue_head(sk
, nskb
);
2635 sk_wmem_free_skb(sk
, skb
);
2636 sk
->sk_wmem_queued
+= nskb
->truesize
;
2637 sk_mem_charge(sk
, nskb
->truesize
);
2641 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_ACK
;
2642 TCP_ECN_send_synack(tcp_sk(sk
), skb
);
2644 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2645 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2649 * tcp_make_synack - Prepare a SYN-ACK.
2650 * sk: listener socket
2651 * dst: dst entry attached to the SYNACK
2652 * req: request_sock pointer
2654 * Allocate one skb and build a SYNACK packet.
2655 * @dst is consumed : Caller should not use it again.
2657 struct sk_buff
*tcp_make_synack(struct sock
*sk
, struct dst_entry
*dst
,
2658 struct request_sock
*req
,
2659 struct tcp_fastopen_cookie
*foc
)
2661 struct tcp_out_options opts
;
2662 struct inet_request_sock
*ireq
= inet_rsk(req
);
2663 struct tcp_sock
*tp
= tcp_sk(sk
);
2665 struct sk_buff
*skb
;
2666 struct tcp_md5sig_key
*md5
;
2667 int tcp_header_size
;
2670 skb
= alloc_skb(MAX_TCP_HEADER
+ 15, sk_gfp_atomic(sk
, GFP_ATOMIC
));
2671 if (unlikely(!skb
)) {
2675 /* Reserve space for headers. */
2676 skb_reserve(skb
, MAX_TCP_HEADER
);
2678 skb_dst_set(skb
, dst
);
2679 security_skb_owned_by(skb
, sk
);
2681 mss
= dst_metric_advmss(dst
);
2682 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< mss
)
2683 mss
= tp
->rx_opt
.user_mss
;
2685 if (req
->rcv_wnd
== 0) { /* ignored for retransmitted syns */
2687 /* Set this up on the first call only */
2688 req
->window_clamp
= tp
->window_clamp
? : dst_metric(dst
, RTAX_WINDOW
);
2690 /* limit the window selection if the user enforce a smaller rx buffer */
2691 if (sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
&&
2692 (req
->window_clamp
> tcp_full_space(sk
) || req
->window_clamp
== 0))
2693 req
->window_clamp
= tcp_full_space(sk
);
2695 /* tcp_full_space because it is guaranteed to be the first packet */
2696 tcp_select_initial_window(tcp_full_space(sk
),
2697 mss
- (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0),
2702 dst_metric(dst
, RTAX_INITRWND
));
2703 ireq
->rcv_wscale
= rcv_wscale
;
2706 memset(&opts
, 0, sizeof(opts
));
2707 #ifdef CONFIG_SYN_COOKIES
2708 if (unlikely(req
->cookie_ts
))
2709 TCP_SKB_CB(skb
)->when
= cookie_init_timestamp(req
);
2712 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2713 tcp_header_size
= tcp_synack_options(sk
, req
, mss
, skb
, &opts
, &md5
,
2716 skb_push(skb
, tcp_header_size
);
2717 skb_reset_transport_header(skb
);
2720 memset(th
, 0, sizeof(struct tcphdr
));
2723 TCP_ECN_make_synack(req
, th
);
2724 th
->source
= ireq
->loc_port
;
2725 th
->dest
= ireq
->rmt_port
;
2726 /* Setting of flags are superfluous here for callers (and ECE is
2727 * not even correctly set)
2729 tcp_init_nondata_skb(skb
, tcp_rsk(req
)->snt_isn
,
2730 TCPHDR_SYN
| TCPHDR_ACK
);
2732 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
2733 /* XXX data is queued and acked as is. No buffer/window check */
2734 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_nxt
);
2736 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2737 th
->window
= htons(min(req
->rcv_wnd
, 65535U));
2738 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
);
2739 th
->doff
= (tcp_header_size
>> 2);
2740 TCP_ADD_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
, tcp_skb_pcount(skb
));
2742 #ifdef CONFIG_TCP_MD5SIG
2743 /* Okay, we have all we need - do the md5 hash if needed */
2745 tcp_rsk(req
)->af_specific
->calc_md5_hash(opts
.hash_location
,
2746 md5
, NULL
, req
, skb
);
2752 EXPORT_SYMBOL(tcp_make_synack
);
2754 /* Do all connect socket setups that can be done AF independent. */
2755 void tcp_connect_init(struct sock
*sk
)
2757 const struct dst_entry
*dst
= __sk_dst_get(sk
);
2758 struct tcp_sock
*tp
= tcp_sk(sk
);
2761 /* We'll fix this up when we get a response from the other end.
2762 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2764 tp
->tcp_header_len
= sizeof(struct tcphdr
) +
2765 (sysctl_tcp_timestamps
? TCPOLEN_TSTAMP_ALIGNED
: 0);
2767 #ifdef CONFIG_TCP_MD5SIG
2768 if (tp
->af_specific
->md5_lookup(sk
, sk
) != NULL
)
2769 tp
->tcp_header_len
+= TCPOLEN_MD5SIG_ALIGNED
;
2772 /* If user gave his TCP_MAXSEG, record it to clamp */
2773 if (tp
->rx_opt
.user_mss
)
2774 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
2777 tcp_sync_mss(sk
, dst_mtu(dst
));
2779 if (!tp
->window_clamp
)
2780 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
2781 tp
->advmss
= dst_metric_advmss(dst
);
2782 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< tp
->advmss
)
2783 tp
->advmss
= tp
->rx_opt
.user_mss
;
2785 tcp_initialize_rcv_mss(sk
);
2787 /* limit the window selection if the user enforce a smaller rx buffer */
2788 if (sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
&&
2789 (tp
->window_clamp
> tcp_full_space(sk
) || tp
->window_clamp
== 0))
2790 tp
->window_clamp
= tcp_full_space(sk
);
2792 tcp_select_initial_window(tcp_full_space(sk
),
2793 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
2796 sysctl_tcp_window_scaling
,
2798 dst_metric(dst
, RTAX_INITRWND
));
2800 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
2801 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
2804 sock_reset_flag(sk
, SOCK_DONE
);
2807 tp
->snd_una
= tp
->write_seq
;
2808 tp
->snd_sml
= tp
->write_seq
;
2809 tp
->snd_up
= tp
->write_seq
;
2810 tp
->snd_nxt
= tp
->write_seq
;
2812 if (likely(!tp
->repair
))
2814 tp
->rcv_wup
= tp
->rcv_nxt
;
2815 tp
->copied_seq
= tp
->rcv_nxt
;
2817 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
2818 inet_csk(sk
)->icsk_retransmits
= 0;
2819 tcp_clear_retrans(tp
);
2822 static void tcp_connect_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
2824 struct tcp_sock
*tp
= tcp_sk(sk
);
2825 struct tcp_skb_cb
*tcb
= TCP_SKB_CB(skb
);
2827 tcb
->end_seq
+= skb
->len
;
2828 skb_header_release(skb
);
2829 __tcp_add_write_queue_tail(sk
, skb
);
2830 sk
->sk_wmem_queued
+= skb
->truesize
;
2831 sk_mem_charge(sk
, skb
->truesize
);
2832 tp
->write_seq
= tcb
->end_seq
;
2833 tp
->packets_out
+= tcp_skb_pcount(skb
);
2836 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
2837 * queue a data-only packet after the regular SYN, such that regular SYNs
2838 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
2839 * only the SYN sequence, the data are retransmitted in the first ACK.
2840 * If cookie is not cached or other error occurs, falls back to send a
2841 * regular SYN with Fast Open cookie request option.
2843 static int tcp_send_syn_data(struct sock
*sk
, struct sk_buff
*syn
)
2845 struct tcp_sock
*tp
= tcp_sk(sk
);
2846 struct tcp_fastopen_request
*fo
= tp
->fastopen_req
;
2847 int syn_loss
= 0, space
, i
, err
= 0, iovlen
= fo
->data
->msg_iovlen
;
2848 struct sk_buff
*syn_data
= NULL
, *data
;
2849 unsigned long last_syn_loss
= 0;
2851 tp
->rx_opt
.mss_clamp
= tp
->advmss
; /* If MSS is not cached */
2852 tcp_fastopen_cache_get(sk
, &tp
->rx_opt
.mss_clamp
, &fo
->cookie
,
2853 &syn_loss
, &last_syn_loss
);
2854 /* Recurring FO SYN losses: revert to regular handshake temporarily */
2856 time_before(jiffies
, last_syn_loss
+ (60*HZ
<< syn_loss
))) {
2857 fo
->cookie
.len
= -1;
2861 if (sysctl_tcp_fastopen
& TFO_CLIENT_NO_COOKIE
)
2862 fo
->cookie
.len
= -1;
2863 else if (fo
->cookie
.len
<= 0)
2866 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
2867 * user-MSS. Reserve maximum option space for middleboxes that add
2868 * private TCP options. The cost is reduced data space in SYN :(
2870 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< tp
->rx_opt
.mss_clamp
)
2871 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
2872 space
= __tcp_mtu_to_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
) -
2873 MAX_TCP_OPTION_SPACE
;
2875 syn_data
= skb_copy_expand(syn
, skb_headroom(syn
), space
,
2877 if (syn_data
== NULL
)
2880 for (i
= 0; i
< iovlen
&& syn_data
->len
< space
; ++i
) {
2881 struct iovec
*iov
= &fo
->data
->msg_iov
[i
];
2882 unsigned char __user
*from
= iov
->iov_base
;
2883 int len
= iov
->iov_len
;
2885 if (syn_data
->len
+ len
> space
)
2886 len
= space
- syn_data
->len
;
2887 else if (i
+ 1 == iovlen
)
2888 /* No more data pending in inet_wait_for_connect() */
2891 if (skb_add_data(syn_data
, from
, len
))
2895 /* Queue a data-only packet after the regular SYN for retransmission */
2896 data
= pskb_copy(syn_data
, sk
->sk_allocation
);
2899 TCP_SKB_CB(data
)->seq
++;
2900 TCP_SKB_CB(data
)->tcp_flags
&= ~TCPHDR_SYN
;
2901 TCP_SKB_CB(data
)->tcp_flags
= (TCPHDR_ACK
|TCPHDR_PSH
);
2902 tcp_connect_queue_skb(sk
, data
);
2903 fo
->copied
= data
->len
;
2905 if (tcp_transmit_skb(sk
, syn_data
, 0, sk
->sk_allocation
) == 0) {
2906 tp
->syn_data
= (fo
->copied
> 0);
2907 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPFASTOPENACTIVE
);
2913 /* Send a regular SYN with Fast Open cookie request option */
2914 if (fo
->cookie
.len
> 0)
2916 err
= tcp_transmit_skb(sk
, syn
, 1, sk
->sk_allocation
);
2918 tp
->syn_fastopen
= 0;
2919 kfree_skb(syn_data
);
2921 fo
->cookie
.len
= -1; /* Exclude Fast Open option for SYN retries */
2925 /* Build a SYN and send it off. */
2926 int tcp_connect(struct sock
*sk
)
2928 struct tcp_sock
*tp
= tcp_sk(sk
);
2929 struct sk_buff
*buff
;
2932 tcp_connect_init(sk
);
2934 if (unlikely(tp
->repair
)) {
2935 tcp_finish_connect(sk
, NULL
);
2939 buff
= alloc_skb_fclone(MAX_TCP_HEADER
+ 15, sk
->sk_allocation
);
2940 if (unlikely(buff
== NULL
))
2943 /* Reserve space for headers. */
2944 skb_reserve(buff
, MAX_TCP_HEADER
);
2946 tcp_init_nondata_skb(buff
, tp
->write_seq
++, TCPHDR_SYN
);
2947 tp
->retrans_stamp
= TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2948 tcp_connect_queue_skb(sk
, buff
);
2949 TCP_ECN_send_syn(sk
, buff
);
2951 /* Send off SYN; include data in Fast Open. */
2952 err
= tp
->fastopen_req
? tcp_send_syn_data(sk
, buff
) :
2953 tcp_transmit_skb(sk
, buff
, 1, sk
->sk_allocation
);
2954 if (err
== -ECONNREFUSED
)
2957 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2958 * in order to make this packet get counted in tcpOutSegs.
2960 tp
->snd_nxt
= tp
->write_seq
;
2961 tp
->pushed_seq
= tp
->write_seq
;
2962 TCP_INC_STATS(sock_net(sk
), TCP_MIB_ACTIVEOPENS
);
2964 /* Timer for repeating the SYN until an answer. */
2965 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2966 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2969 EXPORT_SYMBOL(tcp_connect
);
2971 /* Send out a delayed ack, the caller does the policy checking
2972 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2975 void tcp_send_delayed_ack(struct sock
*sk
)
2977 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2978 int ato
= icsk
->icsk_ack
.ato
;
2979 unsigned long timeout
;
2981 if (ato
> TCP_DELACK_MIN
) {
2982 const struct tcp_sock
*tp
= tcp_sk(sk
);
2983 int max_ato
= HZ
/ 2;
2985 if (icsk
->icsk_ack
.pingpong
||
2986 (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
2987 max_ato
= TCP_DELACK_MAX
;
2989 /* Slow path, intersegment interval is "high". */
2991 /* If some rtt estimate is known, use it to bound delayed ack.
2992 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2996 int rtt
= max(tp
->srtt
>> 3, TCP_DELACK_MIN
);
3002 ato
= min(ato
, max_ato
);
3005 /* Stay within the limit we were given */
3006 timeout
= jiffies
+ ato
;
3008 /* Use new timeout only if there wasn't a older one earlier. */
3009 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
3010 /* If delack timer was blocked or is about to expire,
3013 if (icsk
->icsk_ack
.blocked
||
3014 time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
3019 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
3020 timeout
= icsk
->icsk_ack
.timeout
;
3022 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
3023 icsk
->icsk_ack
.timeout
= timeout
;
3024 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
3027 /* This routine sends an ack and also updates the window. */
3028 void tcp_send_ack(struct sock
*sk
)
3030 struct sk_buff
*buff
;
3032 /* If we have been reset, we may not send again. */
3033 if (sk
->sk_state
== TCP_CLOSE
)
3036 /* We are not putting this on the write queue, so
3037 * tcp_transmit_skb() will set the ownership to this
3040 buff
= alloc_skb(MAX_TCP_HEADER
, sk_gfp_atomic(sk
, GFP_ATOMIC
));
3042 inet_csk_schedule_ack(sk
);
3043 inet_csk(sk
)->icsk_ack
.ato
= TCP_ATO_MIN
;
3044 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
3045 TCP_DELACK_MAX
, TCP_RTO_MAX
);
3049 /* Reserve space for headers and prepare control bits. */
3050 skb_reserve(buff
, MAX_TCP_HEADER
);
3051 tcp_init_nondata_skb(buff
, tcp_acceptable_seq(sk
), TCPHDR_ACK
);
3053 /* Send it off, this clears delayed acks for us. */
3054 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
3055 tcp_transmit_skb(sk
, buff
, 0, sk_gfp_atomic(sk
, GFP_ATOMIC
));
3058 /* This routine sends a packet with an out of date sequence
3059 * number. It assumes the other end will try to ack it.
3061 * Question: what should we make while urgent mode?
3062 * 4.4BSD forces sending single byte of data. We cannot send
3063 * out of window data, because we have SND.NXT==SND.MAX...
3065 * Current solution: to send TWO zero-length segments in urgent mode:
3066 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3067 * out-of-date with SND.UNA-1 to probe window.
3069 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
)
3071 struct tcp_sock
*tp
= tcp_sk(sk
);
3072 struct sk_buff
*skb
;
3074 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3075 skb
= alloc_skb(MAX_TCP_HEADER
, sk_gfp_atomic(sk
, GFP_ATOMIC
));
3079 /* Reserve space for headers and set control bits. */
3080 skb_reserve(skb
, MAX_TCP_HEADER
);
3081 /* Use a previous sequence. This should cause the other
3082 * end to send an ack. Don't queue or clone SKB, just
3085 tcp_init_nondata_skb(skb
, tp
->snd_una
- !urgent
, TCPHDR_ACK
);
3086 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
3087 return tcp_transmit_skb(sk
, skb
, 0, GFP_ATOMIC
);
3090 void tcp_send_window_probe(struct sock
*sk
)
3092 if (sk
->sk_state
== TCP_ESTABLISHED
) {
3093 tcp_sk(sk
)->snd_wl1
= tcp_sk(sk
)->rcv_nxt
- 1;
3094 tcp_sk(sk
)->snd_nxt
= tcp_sk(sk
)->write_seq
;
3095 tcp_xmit_probe_skb(sk
, 0);
3099 /* Initiate keepalive or window probe from timer. */
3100 int tcp_write_wakeup(struct sock
*sk
)
3102 struct tcp_sock
*tp
= tcp_sk(sk
);
3103 struct sk_buff
*skb
;
3105 if (sk
->sk_state
== TCP_CLOSE
)
3108 if ((skb
= tcp_send_head(sk
)) != NULL
&&
3109 before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
))) {
3111 unsigned int mss
= tcp_current_mss(sk
);
3112 unsigned int seg_size
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
3114 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
3115 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
3117 /* We are probing the opening of a window
3118 * but the window size is != 0
3119 * must have been a result SWS avoidance ( sender )
3121 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
3123 seg_size
= min(seg_size
, mss
);
3124 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
3125 if (tcp_fragment(sk
, skb
, seg_size
, mss
))
3127 } else if (!tcp_skb_pcount(skb
))
3128 tcp_set_skb_tso_segs(sk
, skb
, mss
);
3130 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
3131 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
3132 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
3134 tcp_event_new_data_sent(sk
, skb
);
3137 if (between(tp
->snd_up
, tp
->snd_una
+ 1, tp
->snd_una
+ 0xFFFF))
3138 tcp_xmit_probe_skb(sk
, 1);
3139 return tcp_xmit_probe_skb(sk
, 0);
3143 /* A window probe timeout has occurred. If window is not closed send
3144 * a partial packet else a zero probe.
3146 void tcp_send_probe0(struct sock
*sk
)
3148 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3149 struct tcp_sock
*tp
= tcp_sk(sk
);
3152 err
= tcp_write_wakeup(sk
);
3154 if (tp
->packets_out
|| !tcp_send_head(sk
)) {
3155 /* Cancel probe timer, if it is not required. */
3156 icsk
->icsk_probes_out
= 0;
3157 icsk
->icsk_backoff
= 0;
3162 if (icsk
->icsk_backoff
< sysctl_tcp_retries2
)
3163 icsk
->icsk_backoff
++;
3164 icsk
->icsk_probes_out
++;
3165 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3166 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
3169 /* If packet was not sent due to local congestion,
3170 * do not backoff and do not remember icsk_probes_out.
3171 * Let local senders to fight for local resources.
3173 * Use accumulated backoff yet.
3175 if (!icsk
->icsk_probes_out
)
3176 icsk
->icsk_probes_out
= 1;
3177 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3178 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
,
3179 TCP_RESOURCE_PROBE_INTERVAL
),