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
;
163 const struct dst_entry
*dst
= __sk_dst_get(sk
);
165 if (sysctl_tcp_slow_start_after_idle
&&
166 (!tp
->packets_out
&& (s32
)(now
- tp
->lsndtime
) > icsk
->icsk_rto
))
167 tcp_cwnd_restart(sk
, __sk_dst_get(sk
));
171 /* If it is a reply for ato after last received
172 * packet, enter pingpong mode.
174 if ((u32
)(now
- icsk
->icsk_ack
.lrcvtime
) < icsk
->icsk_ack
.ato
&&
175 (!dst
|| !dst_metric(dst
, RTAX_QUICKACK
)))
176 icsk
->icsk_ack
.pingpong
= 1;
179 /* Account for an ACK we sent. */
180 static inline void tcp_event_ack_sent(struct sock
*sk
, unsigned int pkts
)
182 tcp_dec_quickack_mode(sk
, pkts
);
183 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_DACK
);
187 u32
tcp_default_init_rwnd(u32 mss
)
189 /* Initial receive window should be twice of TCP_INIT_CWND to
190 * enable proper sending of new unsent data during fast recovery
191 * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a
192 * limit when mss is larger than 1460.
194 u32 init_rwnd
= TCP_INIT_CWND
* 2;
197 init_rwnd
= max((1460 * init_rwnd
) / mss
, 2U);
201 /* Determine a window scaling and initial window to offer.
202 * Based on the assumption that the given amount of space
203 * will be offered. Store the results in the tp structure.
204 * NOTE: for smooth operation initial space offering should
205 * be a multiple of mss if possible. We assume here that mss >= 1.
206 * This MUST be enforced by all callers.
208 void tcp_select_initial_window(int __space
, __u32 mss
,
209 __u32
*rcv_wnd
, __u32
*window_clamp
,
210 int wscale_ok
, __u8
*rcv_wscale
,
213 unsigned int space
= (__space
< 0 ? 0 : __space
);
215 /* If no clamp set the clamp to the max possible scaled window */
216 if (*window_clamp
== 0)
217 (*window_clamp
) = (65535 << 14);
218 space
= min(*window_clamp
, space
);
220 /* Quantize space offering to a multiple of mss if possible. */
222 space
= (space
/ mss
) * mss
;
224 /* NOTE: offering an initial window larger than 32767
225 * will break some buggy TCP stacks. If the admin tells us
226 * it is likely we could be speaking with such a buggy stack
227 * we will truncate our initial window offering to 32K-1
228 * unless the remote has sent us a window scaling option,
229 * which we interpret as a sign the remote TCP is not
230 * misinterpreting the window field as a signed quantity.
232 if (sysctl_tcp_workaround_signed_windows
)
233 (*rcv_wnd
) = min(space
, MAX_TCP_WINDOW
);
239 /* Set window scaling on max possible window
240 * See RFC1323 for an explanation of the limit to 14
242 space
= max_t(u32
, sysctl_tcp_rmem
[2], sysctl_rmem_max
);
243 space
= min_t(u32
, space
, *window_clamp
);
244 while (space
> 65535 && (*rcv_wscale
) < 14) {
250 if (mss
> (1 << *rcv_wscale
)) {
251 if (!init_rcv_wnd
) /* Use default unless specified otherwise */
252 init_rcv_wnd
= tcp_default_init_rwnd(mss
);
253 *rcv_wnd
= min(*rcv_wnd
, init_rcv_wnd
* mss
);
256 /* Set the clamp no higher than max representable value */
257 (*window_clamp
) = min(65535U << (*rcv_wscale
), *window_clamp
);
259 EXPORT_SYMBOL(tcp_select_initial_window
);
261 /* Chose a new window to advertise, update state in tcp_sock for the
262 * socket, and return result with RFC1323 scaling applied. The return
263 * value can be stuffed directly into th->window for an outgoing
266 static u16
tcp_select_window(struct sock
*sk
)
268 struct tcp_sock
*tp
= tcp_sk(sk
);
269 u32 cur_win
= tcp_receive_window(tp
);
270 u32 new_win
= __tcp_select_window(sk
);
272 /* Never shrink the offered window */
273 if (new_win
< cur_win
) {
274 /* Danger Will Robinson!
275 * Don't update rcv_wup/rcv_wnd here or else
276 * we will not be able to advertise a zero
277 * window in time. --DaveM
279 * Relax Will Robinson.
281 new_win
= ALIGN(cur_win
, 1 << tp
->rx_opt
.rcv_wscale
);
283 tp
->rcv_wnd
= new_win
;
284 tp
->rcv_wup
= tp
->rcv_nxt
;
286 /* Make sure we do not exceed the maximum possible
289 if (!tp
->rx_opt
.rcv_wscale
&& sysctl_tcp_workaround_signed_windows
)
290 new_win
= min(new_win
, MAX_TCP_WINDOW
);
292 new_win
= min(new_win
, (65535U << tp
->rx_opt
.rcv_wscale
));
294 /* RFC1323 scaling applied */
295 new_win
>>= tp
->rx_opt
.rcv_wscale
;
297 /* If we advertise zero window, disable fast path. */
304 /* Packet ECN state for a SYN-ACK */
305 static inline void TCP_ECN_send_synack(const struct tcp_sock
*tp
, struct sk_buff
*skb
)
307 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_CWR
;
308 if (!(tp
->ecn_flags
& TCP_ECN_OK
))
309 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_ECE
;
312 /* Packet ECN state for a SYN. */
313 static inline void TCP_ECN_send_syn(struct sock
*sk
, struct sk_buff
*skb
)
315 struct tcp_sock
*tp
= tcp_sk(sk
);
318 if (sock_net(sk
)->ipv4
.sysctl_tcp_ecn
== 1) {
319 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_ECE
| TCPHDR_CWR
;
320 tp
->ecn_flags
= TCP_ECN_OK
;
324 static __inline__
void
325 TCP_ECN_make_synack(const struct request_sock
*req
, struct tcphdr
*th
)
327 if (inet_rsk(req
)->ecn_ok
)
331 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
334 static inline void TCP_ECN_send(struct sock
*sk
, struct sk_buff
*skb
,
337 struct tcp_sock
*tp
= tcp_sk(sk
);
339 if (tp
->ecn_flags
& TCP_ECN_OK
) {
340 /* Not-retransmitted data segment: set ECT and inject CWR. */
341 if (skb
->len
!= tcp_header_len
&&
342 !before(TCP_SKB_CB(skb
)->seq
, tp
->snd_nxt
)) {
344 if (tp
->ecn_flags
& TCP_ECN_QUEUE_CWR
) {
345 tp
->ecn_flags
&= ~TCP_ECN_QUEUE_CWR
;
346 tcp_hdr(skb
)->cwr
= 1;
347 skb_shinfo(skb
)->gso_type
|= SKB_GSO_TCP_ECN
;
350 /* ACK or retransmitted segment: clear ECT|CE */
351 INET_ECN_dontxmit(sk
);
353 if (tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)
354 tcp_hdr(skb
)->ece
= 1;
358 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
359 * auto increment end seqno.
361 static void tcp_init_nondata_skb(struct sk_buff
*skb
, u32 seq
, u8 flags
)
363 skb
->ip_summed
= CHECKSUM_PARTIAL
;
366 TCP_SKB_CB(skb
)->tcp_flags
= flags
;
367 TCP_SKB_CB(skb
)->sacked
= 0;
369 skb_shinfo(skb
)->gso_segs
= 1;
370 skb_shinfo(skb
)->gso_size
= 0;
371 skb_shinfo(skb
)->gso_type
= 0;
373 TCP_SKB_CB(skb
)->seq
= seq
;
374 if (flags
& (TCPHDR_SYN
| TCPHDR_FIN
))
376 TCP_SKB_CB(skb
)->end_seq
= seq
;
379 static inline bool tcp_urg_mode(const struct tcp_sock
*tp
)
381 return tp
->snd_una
!= tp
->snd_up
;
384 #define OPTION_SACK_ADVERTISE (1 << 0)
385 #define OPTION_TS (1 << 1)
386 #define OPTION_MD5 (1 << 2)
387 #define OPTION_WSCALE (1 << 3)
388 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
390 struct tcp_out_options
{
391 u16 options
; /* bit field of OPTION_* */
392 u16 mss
; /* 0 to disable */
393 u8 ws
; /* window scale, 0 to disable */
394 u8 num_sack_blocks
; /* number of SACK blocks to include */
395 u8 hash_size
; /* bytes in hash_location */
396 __u8
*hash_location
; /* temporary pointer, overloaded */
397 __u32 tsval
, tsecr
; /* need to include OPTION_TS */
398 struct tcp_fastopen_cookie
*fastopen_cookie
; /* Fast open cookie */
401 /* Write previously computed TCP options to the packet.
403 * Beware: Something in the Internet is very sensitive to the ordering of
404 * TCP options, we learned this through the hard way, so be careful here.
405 * Luckily we can at least blame others for their non-compliance but from
406 * inter-operatibility perspective it seems that we're somewhat stuck with
407 * the ordering which we have been using if we want to keep working with
408 * those broken things (not that it currently hurts anybody as there isn't
409 * particular reason why the ordering would need to be changed).
411 * At least SACK_PERM as the first option is known to lead to a disaster
412 * (but it may well be that other scenarios fail similarly).
414 static void tcp_options_write(__be32
*ptr
, struct tcp_sock
*tp
,
415 struct tcp_out_options
*opts
)
417 u16 options
= opts
->options
; /* mungable copy */
419 if (unlikely(OPTION_MD5
& options
)) {
420 *ptr
++ = htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16) |
421 (TCPOPT_MD5SIG
<< 8) | TCPOLEN_MD5SIG
);
422 /* overload cookie hash location */
423 opts
->hash_location
= (__u8
*)ptr
;
427 if (unlikely(opts
->mss
)) {
428 *ptr
++ = htonl((TCPOPT_MSS
<< 24) |
429 (TCPOLEN_MSS
<< 16) |
433 if (likely(OPTION_TS
& options
)) {
434 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
435 *ptr
++ = htonl((TCPOPT_SACK_PERM
<< 24) |
436 (TCPOLEN_SACK_PERM
<< 16) |
437 (TCPOPT_TIMESTAMP
<< 8) |
439 options
&= ~OPTION_SACK_ADVERTISE
;
441 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
443 (TCPOPT_TIMESTAMP
<< 8) |
446 *ptr
++ = htonl(opts
->tsval
);
447 *ptr
++ = htonl(opts
->tsecr
);
450 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
451 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
453 (TCPOPT_SACK_PERM
<< 8) |
457 if (unlikely(OPTION_WSCALE
& options
)) {
458 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
459 (TCPOPT_WINDOW
<< 16) |
460 (TCPOLEN_WINDOW
<< 8) |
464 if (unlikely(opts
->num_sack_blocks
)) {
465 struct tcp_sack_block
*sp
= tp
->rx_opt
.dsack
?
466 tp
->duplicate_sack
: tp
->selective_acks
;
469 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
472 (TCPOLEN_SACK_BASE
+ (opts
->num_sack_blocks
*
473 TCPOLEN_SACK_PERBLOCK
)));
475 for (this_sack
= 0; this_sack
< opts
->num_sack_blocks
;
477 *ptr
++ = htonl(sp
[this_sack
].start_seq
);
478 *ptr
++ = htonl(sp
[this_sack
].end_seq
);
481 tp
->rx_opt
.dsack
= 0;
484 if (unlikely(OPTION_FAST_OPEN_COOKIE
& options
)) {
485 struct tcp_fastopen_cookie
*foc
= opts
->fastopen_cookie
;
487 *ptr
++ = htonl((TCPOPT_EXP
<< 24) |
488 ((TCPOLEN_EXP_FASTOPEN_BASE
+ foc
->len
) << 16) |
489 TCPOPT_FASTOPEN_MAGIC
);
491 memcpy(ptr
, foc
->val
, foc
->len
);
492 if ((foc
->len
& 3) == 2) {
493 u8
*align
= ((u8
*)ptr
) + foc
->len
;
494 align
[0] = align
[1] = TCPOPT_NOP
;
496 ptr
+= (foc
->len
+ 3) >> 2;
500 /* Compute TCP options for SYN packets. This is not the final
501 * network wire format yet.
503 static unsigned int tcp_syn_options(struct sock
*sk
, struct sk_buff
*skb
,
504 struct tcp_out_options
*opts
,
505 struct tcp_md5sig_key
**md5
)
507 struct tcp_sock
*tp
= tcp_sk(sk
);
508 unsigned int remaining
= MAX_TCP_OPTION_SPACE
;
509 struct tcp_fastopen_request
*fastopen
= tp
->fastopen_req
;
511 #ifdef CONFIG_TCP_MD5SIG
512 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
514 opts
->options
|= OPTION_MD5
;
515 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
521 /* We always get an MSS option. The option bytes which will be seen in
522 * normal data packets should timestamps be used, must be in the MSS
523 * advertised. But we subtract them from tp->mss_cache so that
524 * calculations in tcp_sendmsg are simpler etc. So account for this
525 * fact here if necessary. If we don't do this correctly, as a
526 * receiver we won't recognize data packets as being full sized when we
527 * should, and thus we won't abide by the delayed ACK rules correctly.
528 * SACKs don't matter, we never delay an ACK when we have any of those
530 opts
->mss
= tcp_advertise_mss(sk
);
531 remaining
-= TCPOLEN_MSS_ALIGNED
;
533 if (likely(sysctl_tcp_timestamps
&& *md5
== NULL
)) {
534 opts
->options
|= OPTION_TS
;
535 opts
->tsval
= TCP_SKB_CB(skb
)->when
+ tp
->tsoffset
;
536 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
537 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
539 if (likely(sysctl_tcp_window_scaling
)) {
540 opts
->ws
= tp
->rx_opt
.rcv_wscale
;
541 opts
->options
|= OPTION_WSCALE
;
542 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
544 if (likely(sysctl_tcp_sack
)) {
545 opts
->options
|= OPTION_SACK_ADVERTISE
;
546 if (unlikely(!(OPTION_TS
& opts
->options
)))
547 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
550 if (fastopen
&& fastopen
->cookie
.len
>= 0) {
551 u32 need
= TCPOLEN_EXP_FASTOPEN_BASE
+ fastopen
->cookie
.len
;
552 need
= (need
+ 3) & ~3U; /* Align to 32 bits */
553 if (remaining
>= need
) {
554 opts
->options
|= OPTION_FAST_OPEN_COOKIE
;
555 opts
->fastopen_cookie
= &fastopen
->cookie
;
557 tp
->syn_fastopen
= 1;
561 return MAX_TCP_OPTION_SPACE
- remaining
;
564 /* Set up TCP options for SYN-ACKs. */
565 static unsigned int tcp_synack_options(struct sock
*sk
,
566 struct request_sock
*req
,
567 unsigned int mss
, struct sk_buff
*skb
,
568 struct tcp_out_options
*opts
,
569 struct tcp_md5sig_key
**md5
,
570 struct tcp_fastopen_cookie
*foc
)
572 struct inet_request_sock
*ireq
= inet_rsk(req
);
573 unsigned int remaining
= MAX_TCP_OPTION_SPACE
;
575 #ifdef CONFIG_TCP_MD5SIG
576 *md5
= tcp_rsk(req
)->af_specific
->md5_lookup(sk
, req
);
578 opts
->options
|= OPTION_MD5
;
579 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
581 /* We can't fit any SACK blocks in a packet with MD5 + TS
582 * options. There was discussion about disabling SACK
583 * rather than TS in order to fit in better with old,
584 * buggy kernels, but that was deemed to be unnecessary.
586 ireq
->tstamp_ok
&= !ireq
->sack_ok
;
592 /* We always send an MSS option. */
594 remaining
-= TCPOLEN_MSS_ALIGNED
;
596 if (likely(ireq
->wscale_ok
)) {
597 opts
->ws
= ireq
->rcv_wscale
;
598 opts
->options
|= OPTION_WSCALE
;
599 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
601 if (likely(ireq
->tstamp_ok
)) {
602 opts
->options
|= OPTION_TS
;
603 opts
->tsval
= TCP_SKB_CB(skb
)->when
;
604 opts
->tsecr
= req
->ts_recent
;
605 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
607 if (likely(ireq
->sack_ok
)) {
608 opts
->options
|= OPTION_SACK_ADVERTISE
;
609 if (unlikely(!ireq
->tstamp_ok
))
610 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
613 u32 need
= TCPOLEN_EXP_FASTOPEN_BASE
+ foc
->len
;
614 need
= (need
+ 3) & ~3U; /* Align to 32 bits */
615 if (remaining
>= need
) {
616 opts
->options
|= OPTION_FAST_OPEN_COOKIE
;
617 opts
->fastopen_cookie
= foc
;
622 return MAX_TCP_OPTION_SPACE
- remaining
;
625 /* Compute TCP options for ESTABLISHED sockets. This is not the
626 * final wire format yet.
628 static unsigned int tcp_established_options(struct sock
*sk
, struct sk_buff
*skb
,
629 struct tcp_out_options
*opts
,
630 struct tcp_md5sig_key
**md5
)
632 struct tcp_skb_cb
*tcb
= skb
? TCP_SKB_CB(skb
) : NULL
;
633 struct tcp_sock
*tp
= tcp_sk(sk
);
634 unsigned int size
= 0;
635 unsigned int eff_sacks
;
637 #ifdef CONFIG_TCP_MD5SIG
638 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
639 if (unlikely(*md5
)) {
640 opts
->options
|= OPTION_MD5
;
641 size
+= TCPOLEN_MD5SIG_ALIGNED
;
647 if (likely(tp
->rx_opt
.tstamp_ok
)) {
648 opts
->options
|= OPTION_TS
;
649 opts
->tsval
= tcb
? tcb
->when
+ tp
->tsoffset
: 0;
650 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
651 size
+= TCPOLEN_TSTAMP_ALIGNED
;
654 eff_sacks
= tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
;
655 if (unlikely(eff_sacks
)) {
656 const unsigned int remaining
= MAX_TCP_OPTION_SPACE
- size
;
657 opts
->num_sack_blocks
=
658 min_t(unsigned int, eff_sacks
,
659 (remaining
- TCPOLEN_SACK_BASE_ALIGNED
) /
660 TCPOLEN_SACK_PERBLOCK
);
661 size
+= TCPOLEN_SACK_BASE_ALIGNED
+
662 opts
->num_sack_blocks
* TCPOLEN_SACK_PERBLOCK
;
669 /* TCP SMALL QUEUES (TSQ)
671 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
672 * to reduce RTT and bufferbloat.
673 * We do this using a special skb destructor (tcp_wfree).
675 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
676 * needs to be reallocated in a driver.
677 * The invariant being skb->truesize substracted from sk->sk_wmem_alloc
679 * Since transmit from skb destructor is forbidden, we use a tasklet
680 * to process all sockets that eventually need to send more skbs.
681 * We use one tasklet per cpu, with its own queue of sockets.
684 struct tasklet_struct tasklet
;
685 struct list_head head
; /* queue of tcp sockets */
687 static DEFINE_PER_CPU(struct tsq_tasklet
, tsq_tasklet
);
689 static void tcp_tsq_handler(struct sock
*sk
)
691 if ((1 << sk
->sk_state
) &
692 (TCPF_ESTABLISHED
| TCPF_FIN_WAIT1
| TCPF_CLOSING
|
693 TCPF_CLOSE_WAIT
| TCPF_LAST_ACK
))
694 tcp_write_xmit(sk
, tcp_current_mss(sk
), 0, 0, GFP_ATOMIC
);
697 * One tasklest per cpu tries to send more skbs.
698 * We run in tasklet context but need to disable irqs when
699 * transfering tsq->head because tcp_wfree() might
700 * interrupt us (non NAPI drivers)
702 static void tcp_tasklet_func(unsigned long data
)
704 struct tsq_tasklet
*tsq
= (struct tsq_tasklet
*)data
;
707 struct list_head
*q
, *n
;
711 local_irq_save(flags
);
712 list_splice_init(&tsq
->head
, &list
);
713 local_irq_restore(flags
);
715 list_for_each_safe(q
, n
, &list
) {
716 tp
= list_entry(q
, struct tcp_sock
, tsq_node
);
717 list_del(&tp
->tsq_node
);
719 sk
= (struct sock
*)tp
;
722 if (!sock_owned_by_user(sk
)) {
725 /* defer the work to tcp_release_cb() */
726 set_bit(TCP_TSQ_DEFERRED
, &tp
->tsq_flags
);
730 clear_bit(TSQ_QUEUED
, &tp
->tsq_flags
);
735 #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \
736 (1UL << TCP_WRITE_TIMER_DEFERRED) | \
737 (1UL << TCP_DELACK_TIMER_DEFERRED) | \
738 (1UL << TCP_MTU_REDUCED_DEFERRED))
740 * tcp_release_cb - tcp release_sock() callback
743 * called from release_sock() to perform protocol dependent
744 * actions before socket release.
746 void tcp_release_cb(struct sock
*sk
)
748 struct tcp_sock
*tp
= tcp_sk(sk
);
749 unsigned long flags
, nflags
;
751 /* perform an atomic operation only if at least one flag is set */
753 flags
= tp
->tsq_flags
;
754 if (!(flags
& TCP_DEFERRED_ALL
))
756 nflags
= flags
& ~TCP_DEFERRED_ALL
;
757 } while (cmpxchg(&tp
->tsq_flags
, flags
, nflags
) != flags
);
759 if (flags
& (1UL << TCP_TSQ_DEFERRED
))
762 if (flags
& (1UL << TCP_WRITE_TIMER_DEFERRED
)) {
763 tcp_write_timer_handler(sk
);
766 if (flags
& (1UL << TCP_DELACK_TIMER_DEFERRED
)) {
767 tcp_delack_timer_handler(sk
);
770 if (flags
& (1UL << TCP_MTU_REDUCED_DEFERRED
)) {
771 sk
->sk_prot
->mtu_reduced(sk
);
775 EXPORT_SYMBOL(tcp_release_cb
);
777 void __init
tcp_tasklet_init(void)
781 for_each_possible_cpu(i
) {
782 struct tsq_tasklet
*tsq
= &per_cpu(tsq_tasklet
, i
);
784 INIT_LIST_HEAD(&tsq
->head
);
785 tasklet_init(&tsq
->tasklet
,
792 * Write buffer destructor automatically called from kfree_skb.
793 * We cant xmit new skbs from this context, as we might already
796 void tcp_wfree(struct sk_buff
*skb
)
798 struct sock
*sk
= skb
->sk
;
799 struct tcp_sock
*tp
= tcp_sk(sk
);
801 if (test_and_clear_bit(TSQ_THROTTLED
, &tp
->tsq_flags
) &&
802 !test_and_set_bit(TSQ_QUEUED
, &tp
->tsq_flags
)) {
804 struct tsq_tasklet
*tsq
;
806 /* Keep a ref on socket.
807 * This last ref will be released in tcp_tasklet_func()
809 atomic_sub(skb
->truesize
- 1, &sk
->sk_wmem_alloc
);
811 /* queue this socket to tasklet queue */
812 local_irq_save(flags
);
813 tsq
= &__get_cpu_var(tsq_tasklet
);
814 list_add(&tp
->tsq_node
, &tsq
->head
);
815 tasklet_schedule(&tsq
->tasklet
);
816 local_irq_restore(flags
);
822 /* This routine actually transmits TCP packets queued in by
823 * tcp_do_sendmsg(). This is used by both the initial
824 * transmission and possible later retransmissions.
825 * All SKB's seen here are completely headerless. It is our
826 * job to build the TCP header, and pass the packet down to
827 * IP so it can do the same plus pass the packet off to the
830 * We are working here with either a clone of the original
831 * SKB, or a fresh unique copy made by the retransmit engine.
833 static int tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int clone_it
,
836 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
837 struct inet_sock
*inet
;
839 struct tcp_skb_cb
*tcb
;
840 struct tcp_out_options opts
;
841 unsigned int tcp_options_size
, tcp_header_size
;
842 struct tcp_md5sig_key
*md5
;
846 BUG_ON(!skb
|| !tcp_skb_pcount(skb
));
848 /* If congestion control is doing timestamping, we must
849 * take such a timestamp before we potentially clone/copy.
851 if (icsk
->icsk_ca_ops
->flags
& TCP_CONG_RTT_STAMP
)
852 __net_timestamp(skb
);
854 if (likely(clone_it
)) {
855 const struct sk_buff
*fclone
= skb
+ 1;
857 if (unlikely(skb
->fclone
== SKB_FCLONE_ORIG
&&
858 fclone
->fclone
== SKB_FCLONE_CLONE
))
859 NET_INC_STATS_BH(sock_net(sk
),
860 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES
);
862 if (unlikely(skb_cloned(skb
)))
863 skb
= pskb_copy(skb
, gfp_mask
);
865 skb
= skb_clone(skb
, gfp_mask
);
872 tcb
= TCP_SKB_CB(skb
);
873 memset(&opts
, 0, sizeof(opts
));
875 if (unlikely(tcb
->tcp_flags
& TCPHDR_SYN
))
876 tcp_options_size
= tcp_syn_options(sk
, skb
, &opts
, &md5
);
878 tcp_options_size
= tcp_established_options(sk
, skb
, &opts
,
880 tcp_header_size
= tcp_options_size
+ sizeof(struct tcphdr
);
882 if (tcp_packets_in_flight(tp
) == 0)
883 tcp_ca_event(sk
, CA_EVENT_TX_START
);
885 /* if no packet is in qdisc/device queue, then allow XPS to select
888 skb
->ooo_okay
= sk_wmem_alloc_get(sk
) == 0;
890 skb_push(skb
, tcp_header_size
);
891 skb_reset_transport_header(skb
);
895 skb
->destructor
= (sysctl_tcp_limit_output_bytes
> 0) ?
896 tcp_wfree
: sock_wfree
;
897 atomic_add(skb
->truesize
, &sk
->sk_wmem_alloc
);
899 /* Build TCP header and checksum it. */
901 th
->source
= inet
->inet_sport
;
902 th
->dest
= inet
->inet_dport
;
903 th
->seq
= htonl(tcb
->seq
);
904 th
->ack_seq
= htonl(tp
->rcv_nxt
);
905 *(((__be16
*)th
) + 6) = htons(((tcp_header_size
>> 2) << 12) |
908 if (unlikely(tcb
->tcp_flags
& TCPHDR_SYN
)) {
909 /* RFC1323: The window in SYN & SYN/ACK segments
912 th
->window
= htons(min(tp
->rcv_wnd
, 65535U));
914 th
->window
= htons(tcp_select_window(sk
));
919 /* The urg_mode check is necessary during a below snd_una win probe */
920 if (unlikely(tcp_urg_mode(tp
) && before(tcb
->seq
, tp
->snd_up
))) {
921 if (before(tp
->snd_up
, tcb
->seq
+ 0x10000)) {
922 th
->urg_ptr
= htons(tp
->snd_up
- tcb
->seq
);
924 } else if (after(tcb
->seq
+ 0xFFFF, tp
->snd_nxt
)) {
925 th
->urg_ptr
= htons(0xFFFF);
930 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
);
931 if (likely((tcb
->tcp_flags
& TCPHDR_SYN
) == 0))
932 TCP_ECN_send(sk
, skb
, tcp_header_size
);
934 #ifdef CONFIG_TCP_MD5SIG
935 /* Calculate the MD5 hash, as we have all we need now */
937 sk_nocaps_add(sk
, NETIF_F_GSO_MASK
);
938 tp
->af_specific
->calc_md5_hash(opts
.hash_location
,
943 icsk
->icsk_af_ops
->send_check(sk
, skb
);
945 if (likely(tcb
->tcp_flags
& TCPHDR_ACK
))
946 tcp_event_ack_sent(sk
, tcp_skb_pcount(skb
));
948 if (skb
->len
!= tcp_header_size
)
949 tcp_event_data_sent(tp
, sk
);
951 if (after(tcb
->end_seq
, tp
->snd_nxt
) || tcb
->seq
== tcb
->end_seq
)
952 TCP_ADD_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
,
953 tcp_skb_pcount(skb
));
955 err
= icsk
->icsk_af_ops
->queue_xmit(skb
, &inet
->cork
.fl
);
956 if (likely(err
<= 0))
959 tcp_enter_cwr(sk
, 1);
961 return net_xmit_eval(err
);
964 /* This routine just queues the buffer for sending.
966 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
967 * otherwise socket can stall.
969 static void tcp_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
971 struct tcp_sock
*tp
= tcp_sk(sk
);
973 /* Advance write_seq and place onto the write_queue. */
974 tp
->write_seq
= TCP_SKB_CB(skb
)->end_seq
;
975 skb_header_release(skb
);
976 tcp_add_write_queue_tail(sk
, skb
);
977 sk
->sk_wmem_queued
+= skb
->truesize
;
978 sk_mem_charge(sk
, skb
->truesize
);
981 /* Initialize TSO segments for a packet. */
982 static void tcp_set_skb_tso_segs(const struct sock
*sk
, struct sk_buff
*skb
,
983 unsigned int mss_now
)
985 if (skb
->len
<= mss_now
|| !sk_can_gso(sk
) ||
986 skb
->ip_summed
== CHECKSUM_NONE
) {
987 /* Avoid the costly divide in the normal
990 skb_shinfo(skb
)->gso_segs
= 1;
991 skb_shinfo(skb
)->gso_size
= 0;
992 skb_shinfo(skb
)->gso_type
= 0;
994 skb_shinfo(skb
)->gso_segs
= DIV_ROUND_UP(skb
->len
, mss_now
);
995 skb_shinfo(skb
)->gso_size
= mss_now
;
996 skb_shinfo(skb
)->gso_type
= sk
->sk_gso_type
;
1000 /* When a modification to fackets out becomes necessary, we need to check
1001 * skb is counted to fackets_out or not.
1003 static void tcp_adjust_fackets_out(struct sock
*sk
, const struct sk_buff
*skb
,
1006 struct tcp_sock
*tp
= tcp_sk(sk
);
1008 if (!tp
->sacked_out
|| tcp_is_reno(tp
))
1011 if (after(tcp_highest_sack_seq(tp
), TCP_SKB_CB(skb
)->seq
))
1012 tp
->fackets_out
-= decr
;
1015 /* Pcount in the middle of the write queue got changed, we need to do various
1016 * tweaks to fix counters
1018 static void tcp_adjust_pcount(struct sock
*sk
, const struct sk_buff
*skb
, int decr
)
1020 struct tcp_sock
*tp
= tcp_sk(sk
);
1022 tp
->packets_out
-= decr
;
1024 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
1025 tp
->sacked_out
-= decr
;
1026 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
1027 tp
->retrans_out
-= decr
;
1028 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
)
1029 tp
->lost_out
-= decr
;
1031 /* Reno case is special. Sigh... */
1032 if (tcp_is_reno(tp
) && decr
> 0)
1033 tp
->sacked_out
-= min_t(u32
, tp
->sacked_out
, decr
);
1035 tcp_adjust_fackets_out(sk
, skb
, decr
);
1037 if (tp
->lost_skb_hint
&&
1038 before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
) &&
1039 (tcp_is_fack(tp
) || (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)))
1040 tp
->lost_cnt_hint
-= decr
;
1042 tcp_verify_left_out(tp
);
1045 /* Function to create two new TCP segments. Shrinks the given segment
1046 * to the specified size and appends a new segment with the rest of the
1047 * packet to the list. This won't be called frequently, I hope.
1048 * Remember, these are still headerless SKBs at this point.
1050 int tcp_fragment(struct sock
*sk
, struct sk_buff
*skb
, u32 len
,
1051 unsigned int mss_now
)
1053 struct tcp_sock
*tp
= tcp_sk(sk
);
1054 struct sk_buff
*buff
;
1055 int nsize
, old_factor
;
1059 if (WARN_ON(len
> skb
->len
))
1062 nsize
= skb_headlen(skb
) - len
;
1066 if (skb_cloned(skb
) &&
1067 skb_is_nonlinear(skb
) &&
1068 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
1071 /* Get a new skb... force flag on. */
1072 buff
= sk_stream_alloc_skb(sk
, nsize
, GFP_ATOMIC
);
1074 return -ENOMEM
; /* We'll just try again later. */
1076 sk
->sk_wmem_queued
+= buff
->truesize
;
1077 sk_mem_charge(sk
, buff
->truesize
);
1078 nlen
= skb
->len
- len
- nsize
;
1079 buff
->truesize
+= nlen
;
1080 skb
->truesize
-= nlen
;
1082 /* Correct the sequence numbers. */
1083 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1084 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1085 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1087 /* PSH and FIN should only be set in the second packet. */
1088 flags
= TCP_SKB_CB(skb
)->tcp_flags
;
1089 TCP_SKB_CB(skb
)->tcp_flags
= flags
& ~(TCPHDR_FIN
| TCPHDR_PSH
);
1090 TCP_SKB_CB(buff
)->tcp_flags
= flags
;
1091 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
1093 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_PARTIAL
) {
1094 /* Copy and checksum data tail into the new buffer. */
1095 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
,
1096 skb_put(buff
, nsize
),
1101 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
1103 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1104 skb_split(skb
, buff
, len
);
1107 buff
->ip_summed
= skb
->ip_summed
;
1109 /* Looks stupid, but our code really uses when of
1110 * skbs, which it never sent before. --ANK
1112 TCP_SKB_CB(buff
)->when
= TCP_SKB_CB(skb
)->when
;
1113 buff
->tstamp
= skb
->tstamp
;
1115 old_factor
= tcp_skb_pcount(skb
);
1117 /* Fix up tso_factor for both original and new SKB. */
1118 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1119 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1121 /* If this packet has been sent out already, we must
1122 * adjust the various packet counters.
1124 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
1125 int diff
= old_factor
- tcp_skb_pcount(skb
) -
1126 tcp_skb_pcount(buff
);
1129 tcp_adjust_pcount(sk
, skb
, diff
);
1132 /* Link BUFF into the send queue. */
1133 skb_header_release(buff
);
1134 tcp_insert_write_queue_after(skb
, buff
, sk
);
1139 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1140 * eventually). The difference is that pulled data not copied, but
1141 * immediately discarded.
1143 static void __pskb_trim_head(struct sk_buff
*skb
, int len
)
1147 eat
= min_t(int, len
, skb_headlen(skb
));
1149 __skb_pull(skb
, eat
);
1156 for (i
= 0; i
< skb_shinfo(skb
)->nr_frags
; i
++) {
1157 int size
= skb_frag_size(&skb_shinfo(skb
)->frags
[i
]);
1160 skb_frag_unref(skb
, i
);
1163 skb_shinfo(skb
)->frags
[k
] = skb_shinfo(skb
)->frags
[i
];
1165 skb_shinfo(skb
)->frags
[k
].page_offset
+= eat
;
1166 skb_frag_size_sub(&skb_shinfo(skb
)->frags
[k
], eat
);
1172 skb_shinfo(skb
)->nr_frags
= k
;
1174 skb_reset_tail_pointer(skb
);
1175 skb
->data_len
-= len
;
1176 skb
->len
= skb
->data_len
;
1179 /* Remove acked data from a packet in the transmit queue. */
1180 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
1182 if (skb_unclone(skb
, GFP_ATOMIC
))
1185 __pskb_trim_head(skb
, len
);
1187 TCP_SKB_CB(skb
)->seq
+= len
;
1188 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1190 skb
->truesize
-= len
;
1191 sk
->sk_wmem_queued
-= len
;
1192 sk_mem_uncharge(sk
, len
);
1193 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
1195 /* Any change of skb->len requires recalculation of tso factor. */
1196 if (tcp_skb_pcount(skb
) > 1)
1197 tcp_set_skb_tso_segs(sk
, skb
, tcp_skb_mss(skb
));
1202 /* Calculate MSS not accounting any TCP options. */
1203 static inline int __tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1205 const struct tcp_sock
*tp
= tcp_sk(sk
);
1206 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1209 /* Calculate base mss without TCP options:
1210 It is MMS_S - sizeof(tcphdr) of rfc1122
1212 mss_now
= pmtu
- icsk
->icsk_af_ops
->net_header_len
- sizeof(struct tcphdr
);
1214 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1215 if (icsk
->icsk_af_ops
->net_frag_header_len
) {
1216 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1218 if (dst
&& dst_allfrag(dst
))
1219 mss_now
-= icsk
->icsk_af_ops
->net_frag_header_len
;
1222 /* Clamp it (mss_clamp does not include tcp options) */
1223 if (mss_now
> tp
->rx_opt
.mss_clamp
)
1224 mss_now
= tp
->rx_opt
.mss_clamp
;
1226 /* Now subtract optional transport overhead */
1227 mss_now
-= icsk
->icsk_ext_hdr_len
;
1229 /* Then reserve room for full set of TCP options and 8 bytes of data */
1235 /* Calculate MSS. Not accounting for SACKs here. */
1236 int tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1238 /* Subtract TCP options size, not including SACKs */
1239 return __tcp_mtu_to_mss(sk
, pmtu
) -
1240 (tcp_sk(sk
)->tcp_header_len
- sizeof(struct tcphdr
));
1243 /* Inverse of above */
1244 int tcp_mss_to_mtu(struct sock
*sk
, int mss
)
1246 const struct tcp_sock
*tp
= tcp_sk(sk
);
1247 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1251 tp
->tcp_header_len
+
1252 icsk
->icsk_ext_hdr_len
+
1253 icsk
->icsk_af_ops
->net_header_len
;
1255 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1256 if (icsk
->icsk_af_ops
->net_frag_header_len
) {
1257 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1259 if (dst
&& dst_allfrag(dst
))
1260 mtu
+= icsk
->icsk_af_ops
->net_frag_header_len
;
1265 /* MTU probing init per socket */
1266 void tcp_mtup_init(struct sock
*sk
)
1268 struct tcp_sock
*tp
= tcp_sk(sk
);
1269 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1271 icsk
->icsk_mtup
.enabled
= sysctl_tcp_mtu_probing
> 1;
1272 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+ sizeof(struct tcphdr
) +
1273 icsk
->icsk_af_ops
->net_header_len
;
1274 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, sysctl_tcp_base_mss
);
1275 icsk
->icsk_mtup
.probe_size
= 0;
1277 EXPORT_SYMBOL(tcp_mtup_init
);
1279 /* This function synchronize snd mss to current pmtu/exthdr set.
1281 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1282 for TCP options, but includes only bare TCP header.
1284 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1285 It is minimum of user_mss and mss received with SYN.
1286 It also does not include TCP options.
1288 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1290 tp->mss_cache is current effective sending mss, including
1291 all tcp options except for SACKs. It is evaluated,
1292 taking into account current pmtu, but never exceeds
1293 tp->rx_opt.mss_clamp.
1295 NOTE1. rfc1122 clearly states that advertised MSS
1296 DOES NOT include either tcp or ip options.
1298 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1299 are READ ONLY outside this function. --ANK (980731)
1301 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
1303 struct tcp_sock
*tp
= tcp_sk(sk
);
1304 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1307 if (icsk
->icsk_mtup
.search_high
> pmtu
)
1308 icsk
->icsk_mtup
.search_high
= pmtu
;
1310 mss_now
= tcp_mtu_to_mss(sk
, pmtu
);
1311 mss_now
= tcp_bound_to_half_wnd(tp
, mss_now
);
1313 /* And store cached results */
1314 icsk
->icsk_pmtu_cookie
= pmtu
;
1315 if (icsk
->icsk_mtup
.enabled
)
1316 mss_now
= min(mss_now
, tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
));
1317 tp
->mss_cache
= mss_now
;
1321 EXPORT_SYMBOL(tcp_sync_mss
);
1323 /* Compute the current effective MSS, taking SACKs and IP options,
1324 * and even PMTU discovery events into account.
1326 unsigned int tcp_current_mss(struct sock
*sk
)
1328 const struct tcp_sock
*tp
= tcp_sk(sk
);
1329 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1331 unsigned int header_len
;
1332 struct tcp_out_options opts
;
1333 struct tcp_md5sig_key
*md5
;
1335 mss_now
= tp
->mss_cache
;
1338 u32 mtu
= dst_mtu(dst
);
1339 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
1340 mss_now
= tcp_sync_mss(sk
, mtu
);
1343 header_len
= tcp_established_options(sk
, NULL
, &opts
, &md5
) +
1344 sizeof(struct tcphdr
);
1345 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1346 * some common options. If this is an odd packet (because we have SACK
1347 * blocks etc) then our calculated header_len will be different, and
1348 * we have to adjust mss_now correspondingly */
1349 if (header_len
!= tp
->tcp_header_len
) {
1350 int delta
= (int) header_len
- tp
->tcp_header_len
;
1357 /* Congestion window validation. (RFC2861) */
1358 static void tcp_cwnd_validate(struct sock
*sk
)
1360 struct tcp_sock
*tp
= tcp_sk(sk
);
1362 if (tp
->packets_out
>= tp
->snd_cwnd
) {
1363 /* Network is feed fully. */
1364 tp
->snd_cwnd_used
= 0;
1365 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1367 /* Network starves. */
1368 if (tp
->packets_out
> tp
->snd_cwnd_used
)
1369 tp
->snd_cwnd_used
= tp
->packets_out
;
1371 if (sysctl_tcp_slow_start_after_idle
&&
1372 (s32
)(tcp_time_stamp
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
)
1373 tcp_cwnd_application_limited(sk
);
1377 /* Returns the portion of skb which can be sent right away without
1378 * introducing MSS oddities to segment boundaries. In rare cases where
1379 * mss_now != mss_cache, we will request caller to create a small skb
1380 * per input skb which could be mostly avoided here (if desired).
1382 * We explicitly want to create a request for splitting write queue tail
1383 * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1384 * thus all the complexity (cwnd_len is always MSS multiple which we
1385 * return whenever allowed by the other factors). Basically we need the
1386 * modulo only when the receiver window alone is the limiting factor or
1387 * when we would be allowed to send the split-due-to-Nagle skb fully.
1389 static unsigned int tcp_mss_split_point(const struct sock
*sk
, const struct sk_buff
*skb
,
1390 unsigned int mss_now
, unsigned int max_segs
)
1392 const struct tcp_sock
*tp
= tcp_sk(sk
);
1393 u32 needed
, window
, max_len
;
1395 window
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1396 max_len
= mss_now
* max_segs
;
1398 if (likely(max_len
<= window
&& skb
!= tcp_write_queue_tail(sk
)))
1401 needed
= min(skb
->len
, window
);
1403 if (max_len
<= needed
)
1406 return needed
- needed
% mss_now
;
1409 /* Can at least one segment of SKB be sent right now, according to the
1410 * congestion window rules? If so, return how many segments are allowed.
1412 static inline unsigned int tcp_cwnd_test(const struct tcp_sock
*tp
,
1413 const struct sk_buff
*skb
)
1415 u32 in_flight
, cwnd
;
1417 /* Don't be strict about the congestion window for the final FIN. */
1418 if ((TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
) &&
1419 tcp_skb_pcount(skb
) == 1)
1422 in_flight
= tcp_packets_in_flight(tp
);
1423 cwnd
= tp
->snd_cwnd
;
1424 if (in_flight
< cwnd
)
1425 return (cwnd
- in_flight
);
1430 /* Initialize TSO state of a skb.
1431 * This must be invoked the first time we consider transmitting
1432 * SKB onto the wire.
1434 static int tcp_init_tso_segs(const struct sock
*sk
, struct sk_buff
*skb
,
1435 unsigned int mss_now
)
1437 int tso_segs
= tcp_skb_pcount(skb
);
1439 if (!tso_segs
|| (tso_segs
> 1 && tcp_skb_mss(skb
) != mss_now
)) {
1440 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1441 tso_segs
= tcp_skb_pcount(skb
);
1446 /* Minshall's variant of the Nagle send check. */
1447 static inline bool tcp_minshall_check(const struct tcp_sock
*tp
)
1449 return after(tp
->snd_sml
, tp
->snd_una
) &&
1450 !after(tp
->snd_sml
, tp
->snd_nxt
);
1453 /* Return false, if packet can be sent now without violation Nagle's rules:
1454 * 1. It is full sized.
1455 * 2. Or it contains FIN. (already checked by caller)
1456 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1457 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1458 * With Minshall's modification: all sent small packets are ACKed.
1460 static inline bool tcp_nagle_check(const struct tcp_sock
*tp
,
1461 const struct sk_buff
*skb
,
1462 unsigned int mss_now
, int nonagle
)
1464 return skb
->len
< mss_now
&&
1465 ((nonagle
& TCP_NAGLE_CORK
) ||
1466 (!nonagle
&& tp
->packets_out
&& tcp_minshall_check(tp
)));
1469 /* Return true if the Nagle test allows this packet to be
1472 static inline bool tcp_nagle_test(const struct tcp_sock
*tp
, const struct sk_buff
*skb
,
1473 unsigned int cur_mss
, int nonagle
)
1475 /* Nagle rule does not apply to frames, which sit in the middle of the
1476 * write_queue (they have no chances to get new data).
1478 * This is implemented in the callers, where they modify the 'nonagle'
1479 * argument based upon the location of SKB in the send queue.
1481 if (nonagle
& TCP_NAGLE_PUSH
)
1484 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1485 if (tcp_urg_mode(tp
) || (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
))
1488 if (!tcp_nagle_check(tp
, skb
, cur_mss
, nonagle
))
1494 /* Does at least the first segment of SKB fit into the send window? */
1495 static bool tcp_snd_wnd_test(const struct tcp_sock
*tp
,
1496 const struct sk_buff
*skb
,
1497 unsigned int cur_mss
)
1499 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1501 if (skb
->len
> cur_mss
)
1502 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
1504 return !after(end_seq
, tcp_wnd_end(tp
));
1507 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1508 * should be put on the wire right now. If so, it returns the number of
1509 * packets allowed by the congestion window.
1511 static unsigned int tcp_snd_test(const struct sock
*sk
, struct sk_buff
*skb
,
1512 unsigned int cur_mss
, int nonagle
)
1514 const struct tcp_sock
*tp
= tcp_sk(sk
);
1515 unsigned int cwnd_quota
;
1517 tcp_init_tso_segs(sk
, skb
, cur_mss
);
1519 if (!tcp_nagle_test(tp
, skb
, cur_mss
, nonagle
))
1522 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1523 if (cwnd_quota
&& !tcp_snd_wnd_test(tp
, skb
, cur_mss
))
1529 /* Test if sending is allowed right now. */
1530 bool tcp_may_send_now(struct sock
*sk
)
1532 const struct tcp_sock
*tp
= tcp_sk(sk
);
1533 struct sk_buff
*skb
= tcp_send_head(sk
);
1536 tcp_snd_test(sk
, skb
, tcp_current_mss(sk
),
1537 (tcp_skb_is_last(sk
, skb
) ?
1538 tp
->nonagle
: TCP_NAGLE_PUSH
));
1541 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1542 * which is put after SKB on the list. It is very much like
1543 * tcp_fragment() except that it may make several kinds of assumptions
1544 * in order to speed up the splitting operation. In particular, we
1545 * know that all the data is in scatter-gather pages, and that the
1546 * packet has never been sent out before (and thus is not cloned).
1548 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
,
1549 unsigned int mss_now
, gfp_t gfp
)
1551 struct sk_buff
*buff
;
1552 int nlen
= skb
->len
- len
;
1555 /* All of a TSO frame must be composed of paged data. */
1556 if (skb
->len
!= skb
->data_len
)
1557 return tcp_fragment(sk
, skb
, len
, mss_now
);
1559 buff
= sk_stream_alloc_skb(sk
, 0, gfp
);
1560 if (unlikely(buff
== NULL
))
1563 sk
->sk_wmem_queued
+= buff
->truesize
;
1564 sk_mem_charge(sk
, buff
->truesize
);
1565 buff
->truesize
+= nlen
;
1566 skb
->truesize
-= nlen
;
1568 /* Correct the sequence numbers. */
1569 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1570 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1571 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1573 /* PSH and FIN should only be set in the second packet. */
1574 flags
= TCP_SKB_CB(skb
)->tcp_flags
;
1575 TCP_SKB_CB(skb
)->tcp_flags
= flags
& ~(TCPHDR_FIN
| TCPHDR_PSH
);
1576 TCP_SKB_CB(buff
)->tcp_flags
= flags
;
1578 /* This packet was never sent out yet, so no SACK bits. */
1579 TCP_SKB_CB(buff
)->sacked
= 0;
1581 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_PARTIAL
;
1582 skb_split(skb
, buff
, len
);
1584 /* Fix up tso_factor for both original and new SKB. */
1585 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1586 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1588 /* Link BUFF into the send queue. */
1589 skb_header_release(buff
);
1590 tcp_insert_write_queue_after(skb
, buff
, sk
);
1595 /* Try to defer sending, if possible, in order to minimize the amount
1596 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1598 * This algorithm is from John Heffner.
1600 static bool tcp_tso_should_defer(struct sock
*sk
, struct sk_buff
*skb
)
1602 struct tcp_sock
*tp
= tcp_sk(sk
);
1603 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1604 u32 send_win
, cong_win
, limit
, in_flight
;
1607 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
1610 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
1613 /* Defer for less than two clock ticks. */
1614 if (tp
->tso_deferred
&&
1615 (((u32
)jiffies
<< 1) >> 1) - (tp
->tso_deferred
>> 1) > 1)
1618 in_flight
= tcp_packets_in_flight(tp
);
1620 BUG_ON(tcp_skb_pcount(skb
) <= 1 || (tp
->snd_cwnd
<= in_flight
));
1622 send_win
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1624 /* From in_flight test above, we know that cwnd > in_flight. */
1625 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
1627 limit
= min(send_win
, cong_win
);
1629 /* If a full-sized TSO skb can be sent, do it. */
1630 if (limit
>= min_t(unsigned int, sk
->sk_gso_max_size
,
1631 sk
->sk_gso_max_segs
* tp
->mss_cache
))
1634 /* Middle in queue won't get any more data, full sendable already? */
1635 if ((skb
!= tcp_write_queue_tail(sk
)) && (limit
>= skb
->len
))
1638 win_divisor
= ACCESS_ONCE(sysctl_tcp_tso_win_divisor
);
1640 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
1642 /* If at least some fraction of a window is available,
1645 chunk
/= win_divisor
;
1649 /* Different approach, try not to defer past a single
1650 * ACK. Receiver should ACK every other full sized
1651 * frame, so if we have space for more than 3 frames
1654 if (limit
> tcp_max_tso_deferred_mss(tp
) * tp
->mss_cache
)
1658 /* Ok, it looks like it is advisable to defer.
1659 * Do not rearm the timer if already set to not break TCP ACK clocking.
1661 if (!tp
->tso_deferred
)
1662 tp
->tso_deferred
= 1 | (jiffies
<< 1);
1667 tp
->tso_deferred
= 0;
1671 /* Create a new MTU probe if we are ready.
1672 * MTU probe is regularly attempting to increase the path MTU by
1673 * deliberately sending larger packets. This discovers routing
1674 * changes resulting in larger path MTUs.
1676 * Returns 0 if we should wait to probe (no cwnd available),
1677 * 1 if a probe was sent,
1680 static int tcp_mtu_probe(struct sock
*sk
)
1682 struct tcp_sock
*tp
= tcp_sk(sk
);
1683 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1684 struct sk_buff
*skb
, *nskb
, *next
;
1691 /* Not currently probing/verifying,
1693 * have enough cwnd, and
1694 * not SACKing (the variable headers throw things off) */
1695 if (!icsk
->icsk_mtup
.enabled
||
1696 icsk
->icsk_mtup
.probe_size
||
1697 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
||
1698 tp
->snd_cwnd
< 11 ||
1699 tp
->rx_opt
.num_sacks
|| tp
->rx_opt
.dsack
)
1702 /* Very simple search strategy: just double the MSS. */
1703 mss_now
= tcp_current_mss(sk
);
1704 probe_size
= 2 * tp
->mss_cache
;
1705 size_needed
= probe_size
+ (tp
->reordering
+ 1) * tp
->mss_cache
;
1706 if (probe_size
> tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_high
)) {
1707 /* TODO: set timer for probe_converge_event */
1711 /* Have enough data in the send queue to probe? */
1712 if (tp
->write_seq
- tp
->snd_nxt
< size_needed
)
1715 if (tp
->snd_wnd
< size_needed
)
1717 if (after(tp
->snd_nxt
+ size_needed
, tcp_wnd_end(tp
)))
1720 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1721 if (tcp_packets_in_flight(tp
) + 2 > tp
->snd_cwnd
) {
1722 if (!tcp_packets_in_flight(tp
))
1728 /* We're allowed to probe. Build it now. */
1729 if ((nskb
= sk_stream_alloc_skb(sk
, probe_size
, GFP_ATOMIC
)) == NULL
)
1731 sk
->sk_wmem_queued
+= nskb
->truesize
;
1732 sk_mem_charge(sk
, nskb
->truesize
);
1734 skb
= tcp_send_head(sk
);
1736 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(skb
)->seq
;
1737 TCP_SKB_CB(nskb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ probe_size
;
1738 TCP_SKB_CB(nskb
)->tcp_flags
= TCPHDR_ACK
;
1739 TCP_SKB_CB(nskb
)->sacked
= 0;
1741 nskb
->ip_summed
= skb
->ip_summed
;
1743 tcp_insert_write_queue_before(nskb
, skb
, sk
);
1746 tcp_for_write_queue_from_safe(skb
, next
, sk
) {
1747 copy
= min_t(int, skb
->len
, probe_size
- len
);
1748 if (nskb
->ip_summed
)
1749 skb_copy_bits(skb
, 0, skb_put(nskb
, copy
), copy
);
1751 nskb
->csum
= skb_copy_and_csum_bits(skb
, 0,
1752 skb_put(nskb
, copy
),
1755 if (skb
->len
<= copy
) {
1756 /* We've eaten all the data from this skb.
1758 TCP_SKB_CB(nskb
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
;
1759 tcp_unlink_write_queue(skb
, sk
);
1760 sk_wmem_free_skb(sk
, skb
);
1762 TCP_SKB_CB(nskb
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
&
1763 ~(TCPHDR_FIN
|TCPHDR_PSH
);
1764 if (!skb_shinfo(skb
)->nr_frags
) {
1765 skb_pull(skb
, copy
);
1766 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1767 skb
->csum
= csum_partial(skb
->data
,
1770 __pskb_trim_head(skb
, copy
);
1771 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1773 TCP_SKB_CB(skb
)->seq
+= copy
;
1778 if (len
>= probe_size
)
1781 tcp_init_tso_segs(sk
, nskb
, nskb
->len
);
1783 /* We're ready to send. If this fails, the probe will
1784 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1785 TCP_SKB_CB(nskb
)->when
= tcp_time_stamp
;
1786 if (!tcp_transmit_skb(sk
, nskb
, 1, GFP_ATOMIC
)) {
1787 /* Decrement cwnd here because we are sending
1788 * effectively two packets. */
1790 tcp_event_new_data_sent(sk
, nskb
);
1792 icsk
->icsk_mtup
.probe_size
= tcp_mss_to_mtu(sk
, nskb
->len
);
1793 tp
->mtu_probe
.probe_seq_start
= TCP_SKB_CB(nskb
)->seq
;
1794 tp
->mtu_probe
.probe_seq_end
= TCP_SKB_CB(nskb
)->end_seq
;
1802 /* This routine writes packets to the network. It advances the
1803 * send_head. This happens as incoming acks open up the remote
1806 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1807 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1808 * account rare use of URG, this is not a big flaw.
1810 * Send at most one packet when push_one > 0. Temporarily ignore
1811 * cwnd limit to force at most one packet out when push_one == 2.
1813 * Returns true, if no segments are in flight and we have queued segments,
1814 * but cannot send anything now because of SWS or another problem.
1816 static bool tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
,
1817 int push_one
, gfp_t gfp
)
1819 struct tcp_sock
*tp
= tcp_sk(sk
);
1820 struct sk_buff
*skb
;
1821 unsigned int tso_segs
, sent_pkts
;
1828 /* Do MTU probing. */
1829 result
= tcp_mtu_probe(sk
);
1832 } else if (result
> 0) {
1837 while ((skb
= tcp_send_head(sk
))) {
1841 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1844 if (unlikely(tp
->repair
) && tp
->repair_queue
== TCP_SEND_QUEUE
)
1845 goto repair
; /* Skip network transmission */
1847 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1850 /* Force out a loss probe pkt. */
1856 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
)))
1859 if (tso_segs
== 1) {
1860 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
1861 (tcp_skb_is_last(sk
, skb
) ?
1862 nonagle
: TCP_NAGLE_PUSH
))))
1865 if (!push_one
&& tcp_tso_should_defer(sk
, skb
))
1869 /* TSQ : sk_wmem_alloc accounts skb truesize,
1870 * including skb overhead. But thats OK.
1872 if (atomic_read(&sk
->sk_wmem_alloc
) >= sysctl_tcp_limit_output_bytes
) {
1873 set_bit(TSQ_THROTTLED
, &tp
->tsq_flags
);
1877 if (tso_segs
> 1 && !tcp_urg_mode(tp
))
1878 limit
= tcp_mss_split_point(sk
, skb
, mss_now
,
1881 sk
->sk_gso_max_segs
));
1883 if (skb
->len
> limit
&&
1884 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
, gfp
)))
1887 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1889 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, gfp
)))
1893 /* Advance the send_head. This one is sent out.
1894 * This call will increment packets_out.
1896 tcp_event_new_data_sent(sk
, skb
);
1898 tcp_minshall_update(tp
, mss_now
, skb
);
1899 sent_pkts
+= tcp_skb_pcount(skb
);
1905 if (likely(sent_pkts
)) {
1906 if (tcp_in_cwnd_reduction(sk
))
1907 tp
->prr_out
+= sent_pkts
;
1909 /* Send one loss probe per tail loss episode. */
1911 tcp_schedule_loss_probe(sk
);
1912 tcp_cwnd_validate(sk
);
1915 return (push_one
== 2) || (!tp
->packets_out
&& tcp_send_head(sk
));
1918 bool tcp_schedule_loss_probe(struct sock
*sk
)
1920 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1921 struct tcp_sock
*tp
= tcp_sk(sk
);
1922 u32 timeout
, tlp_time_stamp
, rto_time_stamp
;
1923 u32 rtt
= tp
->srtt
>> 3;
1925 if (WARN_ON(icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
))
1927 /* No consecutive loss probes. */
1928 if (WARN_ON(icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
)) {
1932 /* Don't do any loss probe on a Fast Open connection before 3WHS
1935 if (sk
->sk_state
== TCP_SYN_RECV
)
1938 /* TLP is only scheduled when next timer event is RTO. */
1939 if (icsk
->icsk_pending
!= ICSK_TIME_RETRANS
)
1942 /* Schedule a loss probe in 2*RTT for SACK capable connections
1943 * in Open state, that are either limited by cwnd or application.
1945 if (sysctl_tcp_early_retrans
< 3 || !rtt
|| !tp
->packets_out
||
1946 !tcp_is_sack(tp
) || inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
)
1949 if ((tp
->snd_cwnd
> tcp_packets_in_flight(tp
)) &&
1953 /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
1954 * for delayed ack when there's one outstanding packet.
1957 if (tp
->packets_out
== 1)
1958 timeout
= max_t(u32
, timeout
,
1959 (rtt
+ (rtt
>> 1) + TCP_DELACK_MAX
));
1960 timeout
= max_t(u32
, timeout
, msecs_to_jiffies(10));
1962 /* If RTO is shorter, just schedule TLP in its place. */
1963 tlp_time_stamp
= tcp_time_stamp
+ timeout
;
1964 rto_time_stamp
= (u32
)inet_csk(sk
)->icsk_timeout
;
1965 if ((s32
)(tlp_time_stamp
- rto_time_stamp
) > 0) {
1966 s32 delta
= rto_time_stamp
- tcp_time_stamp
;
1971 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_LOSS_PROBE
, timeout
,
1976 /* When probe timeout (PTO) fires, send a new segment if one exists, else
1977 * retransmit the last segment.
1979 void tcp_send_loss_probe(struct sock
*sk
)
1981 struct tcp_sock
*tp
= tcp_sk(sk
);
1982 struct sk_buff
*skb
;
1984 int mss
= tcp_current_mss(sk
);
1987 if (tcp_send_head(sk
) != NULL
) {
1988 err
= tcp_write_xmit(sk
, mss
, TCP_NAGLE_OFF
, 2, GFP_ATOMIC
);
1992 /* At most one outstanding TLP retransmission. */
1993 if (tp
->tlp_high_seq
)
1996 /* Retransmit last segment. */
1997 skb
= tcp_write_queue_tail(sk
);
2001 pcount
= tcp_skb_pcount(skb
);
2002 if (WARN_ON(!pcount
))
2005 if ((pcount
> 1) && (skb
->len
> (pcount
- 1) * mss
)) {
2006 if (unlikely(tcp_fragment(sk
, skb
, (pcount
- 1) * mss
, mss
)))
2008 skb
= tcp_write_queue_tail(sk
);
2011 if (WARN_ON(!skb
|| !tcp_skb_pcount(skb
)))
2014 /* Probe with zero data doesn't trigger fast recovery. */
2016 err
= __tcp_retransmit_skb(sk
, skb
);
2018 /* Record snd_nxt for loss detection. */
2020 tp
->tlp_high_seq
= tp
->snd_nxt
;
2023 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2024 inet_csk(sk
)->icsk_rto
,
2028 NET_INC_STATS_BH(sock_net(sk
),
2029 LINUX_MIB_TCPLOSSPROBES
);
2033 /* Push out any pending frames which were held back due to
2034 * TCP_CORK or attempt at coalescing tiny packets.
2035 * The socket must be locked by the caller.
2037 void __tcp_push_pending_frames(struct sock
*sk
, unsigned int cur_mss
,
2040 /* If we are closed, the bytes will have to remain here.
2041 * In time closedown will finish, we empty the write queue and
2042 * all will be happy.
2044 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
2047 if (tcp_write_xmit(sk
, cur_mss
, nonagle
, 0,
2048 sk_gfp_atomic(sk
, GFP_ATOMIC
)))
2049 tcp_check_probe_timer(sk
);
2052 /* Send _single_ skb sitting at the send head. This function requires
2053 * true push pending frames to setup probe timer etc.
2055 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
2057 struct sk_buff
*skb
= tcp_send_head(sk
);
2059 BUG_ON(!skb
|| skb
->len
< mss_now
);
2061 tcp_write_xmit(sk
, mss_now
, TCP_NAGLE_PUSH
, 1, sk
->sk_allocation
);
2064 /* This function returns the amount that we can raise the
2065 * usable window based on the following constraints
2067 * 1. The window can never be shrunk once it is offered (RFC 793)
2068 * 2. We limit memory per socket
2071 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2072 * RECV.NEXT + RCV.WIN fixed until:
2073 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2075 * i.e. don't raise the right edge of the window until you can raise
2076 * it at least MSS bytes.
2078 * Unfortunately, the recommended algorithm breaks header prediction,
2079 * since header prediction assumes th->window stays fixed.
2081 * Strictly speaking, keeping th->window fixed violates the receiver
2082 * side SWS prevention criteria. The problem is that under this rule
2083 * a stream of single byte packets will cause the right side of the
2084 * window to always advance by a single byte.
2086 * Of course, if the sender implements sender side SWS prevention
2087 * then this will not be a problem.
2089 * BSD seems to make the following compromise:
2091 * If the free space is less than the 1/4 of the maximum
2092 * space available and the free space is less than 1/2 mss,
2093 * then set the window to 0.
2094 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2095 * Otherwise, just prevent the window from shrinking
2096 * and from being larger than the largest representable value.
2098 * This prevents incremental opening of the window in the regime
2099 * where TCP is limited by the speed of the reader side taking
2100 * data out of the TCP receive queue. It does nothing about
2101 * those cases where the window is constrained on the sender side
2102 * because the pipeline is full.
2104 * BSD also seems to "accidentally" limit itself to windows that are a
2105 * multiple of MSS, at least until the free space gets quite small.
2106 * This would appear to be a side effect of the mbuf implementation.
2107 * Combining these two algorithms results in the observed behavior
2108 * of having a fixed window size at almost all times.
2110 * Below we obtain similar behavior by forcing the offered window to
2111 * a multiple of the mss when it is feasible to do so.
2113 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2114 * Regular options like TIMESTAMP are taken into account.
2116 u32
__tcp_select_window(struct sock
*sk
)
2118 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2119 struct tcp_sock
*tp
= tcp_sk(sk
);
2120 /* MSS for the peer's data. Previous versions used mss_clamp
2121 * here. I don't know if the value based on our guesses
2122 * of peer's MSS is better for the performance. It's more correct
2123 * but may be worse for the performance because of rcv_mss
2124 * fluctuations. --SAW 1998/11/1
2126 int mss
= icsk
->icsk_ack
.rcv_mss
;
2127 int free_space
= tcp_space(sk
);
2128 int full_space
= min_t(int, tp
->window_clamp
, tcp_full_space(sk
));
2131 if (mss
> full_space
)
2134 if (free_space
< (full_space
>> 1)) {
2135 icsk
->icsk_ack
.quick
= 0;
2137 if (sk_under_memory_pressure(sk
))
2138 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
,
2141 if (free_space
< mss
)
2145 if (free_space
> tp
->rcv_ssthresh
)
2146 free_space
= tp
->rcv_ssthresh
;
2148 /* Don't do rounding if we are using window scaling, since the
2149 * scaled window will not line up with the MSS boundary anyway.
2151 window
= tp
->rcv_wnd
;
2152 if (tp
->rx_opt
.rcv_wscale
) {
2153 window
= free_space
;
2155 /* Advertise enough space so that it won't get scaled away.
2156 * Import case: prevent zero window announcement if
2157 * 1<<rcv_wscale > mss.
2159 if (((window
>> tp
->rx_opt
.rcv_wscale
) << tp
->rx_opt
.rcv_wscale
) != window
)
2160 window
= (((window
>> tp
->rx_opt
.rcv_wscale
) + 1)
2161 << tp
->rx_opt
.rcv_wscale
);
2163 /* Get the largest window that is a nice multiple of mss.
2164 * Window clamp already applied above.
2165 * If our current window offering is within 1 mss of the
2166 * free space we just keep it. This prevents the divide
2167 * and multiply from happening most of the time.
2168 * We also don't do any window rounding when the free space
2171 if (window
<= free_space
- mss
|| window
> free_space
)
2172 window
= (free_space
/ mss
) * mss
;
2173 else if (mss
== full_space
&&
2174 free_space
> window
+ (full_space
>> 1))
2175 window
= free_space
;
2181 /* Collapses two adjacent SKB's during retransmission. */
2182 static void tcp_collapse_retrans(struct sock
*sk
, struct sk_buff
*skb
)
2184 struct tcp_sock
*tp
= tcp_sk(sk
);
2185 struct sk_buff
*next_skb
= tcp_write_queue_next(sk
, skb
);
2186 int skb_size
, next_skb_size
;
2188 skb_size
= skb
->len
;
2189 next_skb_size
= next_skb
->len
;
2191 BUG_ON(tcp_skb_pcount(skb
) != 1 || tcp_skb_pcount(next_skb
) != 1);
2193 tcp_highest_sack_combine(sk
, next_skb
, skb
);
2195 tcp_unlink_write_queue(next_skb
, sk
);
2197 skb_copy_from_linear_data(next_skb
, skb_put(skb
, next_skb_size
),
2200 if (next_skb
->ip_summed
== CHECKSUM_PARTIAL
)
2201 skb
->ip_summed
= CHECKSUM_PARTIAL
;
2203 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
2204 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
2206 /* Update sequence range on original skb. */
2207 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
2209 /* Merge over control information. This moves PSH/FIN etc. over */
2210 TCP_SKB_CB(skb
)->tcp_flags
|= TCP_SKB_CB(next_skb
)->tcp_flags
;
2212 /* All done, get rid of second SKB and account for it so
2213 * packet counting does not break.
2215 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
& TCPCB_EVER_RETRANS
;
2217 /* changed transmit queue under us so clear hints */
2218 tcp_clear_retrans_hints_partial(tp
);
2219 if (next_skb
== tp
->retransmit_skb_hint
)
2220 tp
->retransmit_skb_hint
= skb
;
2222 tcp_adjust_pcount(sk
, next_skb
, tcp_skb_pcount(next_skb
));
2224 sk_wmem_free_skb(sk
, next_skb
);
2227 /* Check if coalescing SKBs is legal. */
2228 static bool tcp_can_collapse(const struct sock
*sk
, const struct sk_buff
*skb
)
2230 if (tcp_skb_pcount(skb
) > 1)
2232 /* TODO: SACK collapsing could be used to remove this condition */
2233 if (skb_shinfo(skb
)->nr_frags
!= 0)
2235 if (skb_cloned(skb
))
2237 if (skb
== tcp_send_head(sk
))
2239 /* Some heurestics for collapsing over SACK'd could be invented */
2240 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
2246 /* Collapse packets in the retransmit queue to make to create
2247 * less packets on the wire. This is only done on retransmission.
2249 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*to
,
2252 struct tcp_sock
*tp
= tcp_sk(sk
);
2253 struct sk_buff
*skb
= to
, *tmp
;
2256 if (!sysctl_tcp_retrans_collapse
)
2258 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)
2261 tcp_for_write_queue_from_safe(skb
, tmp
, sk
) {
2262 if (!tcp_can_collapse(sk
, skb
))
2274 /* Punt if not enough space exists in the first SKB for
2275 * the data in the second
2277 if (skb
->len
> skb_availroom(to
))
2280 if (after(TCP_SKB_CB(skb
)->end_seq
, tcp_wnd_end(tp
)))
2283 tcp_collapse_retrans(sk
, to
);
2287 /* This retransmits one SKB. Policy decisions and retransmit queue
2288 * state updates are done by the caller. Returns non-zero if an
2289 * error occurred which prevented the send.
2291 int __tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
2293 struct tcp_sock
*tp
= tcp_sk(sk
);
2294 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2295 unsigned int cur_mss
;
2297 /* Inconslusive MTU probe */
2298 if (icsk
->icsk_mtup
.probe_size
) {
2299 icsk
->icsk_mtup
.probe_size
= 0;
2302 /* Do not sent more than we queued. 1/4 is reserved for possible
2303 * copying overhead: fragmentation, tunneling, mangling etc.
2305 if (atomic_read(&sk
->sk_wmem_alloc
) >
2306 min(sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2), sk
->sk_sndbuf
))
2309 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
2310 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
2312 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
2316 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
2317 return -EHOSTUNREACH
; /* Routing failure or similar. */
2319 cur_mss
= tcp_current_mss(sk
);
2321 /* If receiver has shrunk his window, and skb is out of
2322 * new window, do not retransmit it. The exception is the
2323 * case, when window is shrunk to zero. In this case
2324 * our retransmit serves as a zero window probe.
2326 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
)) &&
2327 TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
2330 if (skb
->len
> cur_mss
) {
2331 if (tcp_fragment(sk
, skb
, cur_mss
, cur_mss
))
2332 return -ENOMEM
; /* We'll try again later. */
2334 int oldpcount
= tcp_skb_pcount(skb
);
2336 if (unlikely(oldpcount
> 1)) {
2337 tcp_init_tso_segs(sk
, skb
, cur_mss
);
2338 tcp_adjust_pcount(sk
, skb
, oldpcount
- tcp_skb_pcount(skb
));
2342 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
2344 /* Some Solaris stacks overoptimize and ignore the FIN on a
2345 * retransmit when old data is attached. So strip it off
2346 * since it is cheap to do so and saves bytes on the network.
2349 (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
) &&
2350 tp
->snd_una
== (TCP_SKB_CB(skb
)->end_seq
- 1)) {
2351 if (!pskb_trim(skb
, 0)) {
2352 /* Reuse, even though it does some unnecessary work */
2353 tcp_init_nondata_skb(skb
, TCP_SKB_CB(skb
)->end_seq
- 1,
2354 TCP_SKB_CB(skb
)->tcp_flags
);
2355 skb
->ip_summed
= CHECKSUM_NONE
;
2359 /* Make a copy, if the first transmission SKB clone we made
2360 * is still in somebody's hands, else make a clone.
2362 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2364 /* make sure skb->data is aligned on arches that require it
2365 * and check if ack-trimming & collapsing extended the headroom
2366 * beyond what csum_start can cover.
2368 if (unlikely((NET_IP_ALIGN
&& ((unsigned long)skb
->data
& 3)) ||
2369 skb_headroom(skb
) >= 0xFFFF)) {
2370 struct sk_buff
*nskb
= __pskb_copy(skb
, MAX_TCP_HEADER
,
2372 return nskb
? tcp_transmit_skb(sk
, nskb
, 0, GFP_ATOMIC
) :
2375 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2379 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
2381 struct tcp_sock
*tp
= tcp_sk(sk
);
2382 int err
= __tcp_retransmit_skb(sk
, skb
);
2385 /* Update global TCP statistics. */
2386 TCP_INC_STATS(sock_net(sk
), TCP_MIB_RETRANSSEGS
);
2388 tp
->total_retrans
++;
2390 #if FASTRETRANS_DEBUG > 0
2391 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2392 net_dbg_ratelimited("retrans_out leaked\n");
2395 if (!tp
->retrans_out
)
2396 tp
->lost_retrans_low
= tp
->snd_nxt
;
2397 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
2398 tp
->retrans_out
+= tcp_skb_pcount(skb
);
2400 /* Save stamp of the first retransmit. */
2401 if (!tp
->retrans_stamp
)
2402 tp
->retrans_stamp
= TCP_SKB_CB(skb
)->when
;
2404 tp
->undo_retrans
+= tcp_skb_pcount(skb
);
2406 /* snd_nxt is stored to detect loss of retransmitted segment,
2407 * see tcp_input.c tcp_sacktag_write_queue().
2409 TCP_SKB_CB(skb
)->ack_seq
= tp
->snd_nxt
;
2411 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRETRANSFAIL
);
2416 /* Check if we forward retransmits are possible in the current
2417 * window/congestion state.
2419 static bool tcp_can_forward_retransmit(struct sock
*sk
)
2421 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2422 const struct tcp_sock
*tp
= tcp_sk(sk
);
2424 /* Forward retransmissions are possible only during Recovery. */
2425 if (icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
2428 /* No forward retransmissions in Reno are possible. */
2429 if (tcp_is_reno(tp
))
2432 /* Yeah, we have to make difficult choice between forward transmission
2433 * and retransmission... Both ways have their merits...
2435 * For now we do not retransmit anything, while we have some new
2436 * segments to send. In the other cases, follow rule 3 for
2437 * NextSeg() specified in RFC3517.
2440 if (tcp_may_send_now(sk
))
2446 /* This gets called after a retransmit timeout, and the initially
2447 * retransmitted data is acknowledged. It tries to continue
2448 * resending the rest of the retransmit queue, until either
2449 * we've sent it all or the congestion window limit is reached.
2450 * If doing SACK, the first ACK which comes back for a timeout
2451 * based retransmit packet might feed us FACK information again.
2452 * If so, we use it to avoid unnecessarily retransmissions.
2454 void tcp_xmit_retransmit_queue(struct sock
*sk
)
2456 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2457 struct tcp_sock
*tp
= tcp_sk(sk
);
2458 struct sk_buff
*skb
;
2459 struct sk_buff
*hole
= NULL
;
2462 int fwd_rexmitting
= 0;
2464 if (!tp
->packets_out
)
2468 tp
->retransmit_high
= tp
->snd_una
;
2470 if (tp
->retransmit_skb_hint
) {
2471 skb
= tp
->retransmit_skb_hint
;
2472 last_lost
= TCP_SKB_CB(skb
)->end_seq
;
2473 if (after(last_lost
, tp
->retransmit_high
))
2474 last_lost
= tp
->retransmit_high
;
2476 skb
= tcp_write_queue_head(sk
);
2477 last_lost
= tp
->snd_una
;
2480 tcp_for_write_queue_from(skb
, sk
) {
2481 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
2483 if (skb
== tcp_send_head(sk
))
2485 /* we could do better than to assign each time */
2487 tp
->retransmit_skb_hint
= skb
;
2489 /* Assume this retransmit will generate
2490 * only one packet for congestion window
2491 * calculation purposes. This works because
2492 * tcp_retransmit_skb() will chop up the
2493 * packet to be MSS sized and all the
2494 * packet counting works out.
2496 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
2499 if (fwd_rexmitting
) {
2501 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_highest_sack_seq(tp
)))
2503 mib_idx
= LINUX_MIB_TCPFORWARDRETRANS
;
2505 } else if (!before(TCP_SKB_CB(skb
)->seq
, tp
->retransmit_high
)) {
2506 tp
->retransmit_high
= last_lost
;
2507 if (!tcp_can_forward_retransmit(sk
))
2509 /* Backtrack if necessary to non-L'ed skb */
2517 } else if (!(sacked
& TCPCB_LOST
)) {
2518 if (hole
== NULL
&& !(sacked
& (TCPCB_SACKED_RETRANS
|TCPCB_SACKED_ACKED
)))
2523 last_lost
= TCP_SKB_CB(skb
)->end_seq
;
2524 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
2525 mib_idx
= LINUX_MIB_TCPFASTRETRANS
;
2527 mib_idx
= LINUX_MIB_TCPSLOWSTARTRETRANS
;
2530 if (sacked
& (TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))
2533 if (tcp_retransmit_skb(sk
, skb
))
2536 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2538 if (tcp_in_cwnd_reduction(sk
))
2539 tp
->prr_out
+= tcp_skb_pcount(skb
);
2541 if (skb
== tcp_write_queue_head(sk
))
2542 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2543 inet_csk(sk
)->icsk_rto
,
2548 /* Send a fin. The caller locks the socket for us. This cannot be
2549 * allowed to fail queueing a FIN frame under any circumstances.
2551 void tcp_send_fin(struct sock
*sk
)
2553 struct tcp_sock
*tp
= tcp_sk(sk
);
2554 struct sk_buff
*skb
= tcp_write_queue_tail(sk
);
2557 /* Optimization, tack on the FIN if we have a queue of
2558 * unsent frames. But be careful about outgoing SACKS
2561 mss_now
= tcp_current_mss(sk
);
2563 if (tcp_send_head(sk
) != NULL
) {
2564 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_FIN
;
2565 TCP_SKB_CB(skb
)->end_seq
++;
2568 /* Socket is locked, keep trying until memory is available. */
2570 skb
= alloc_skb_fclone(MAX_TCP_HEADER
,
2577 /* Reserve space for headers and prepare control bits. */
2578 skb_reserve(skb
, MAX_TCP_HEADER
);
2579 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2580 tcp_init_nondata_skb(skb
, tp
->write_seq
,
2581 TCPHDR_ACK
| TCPHDR_FIN
);
2582 tcp_queue_skb(sk
, skb
);
2584 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_OFF
);
2587 /* We get here when a process closes a file descriptor (either due to
2588 * an explicit close() or as a byproduct of exit()'ing) and there
2589 * was unread data in the receive queue. This behavior is recommended
2590 * by RFC 2525, section 2.17. -DaveM
2592 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
2594 struct sk_buff
*skb
;
2596 /* NOTE: No TCP options attached and we never retransmit this. */
2597 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
2599 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
2603 /* Reserve space for headers and prepare control bits. */
2604 skb_reserve(skb
, MAX_TCP_HEADER
);
2605 tcp_init_nondata_skb(skb
, tcp_acceptable_seq(sk
),
2606 TCPHDR_ACK
| TCPHDR_RST
);
2608 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2609 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
2610 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
2612 TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTRSTS
);
2615 /* Send a crossed SYN-ACK during socket establishment.
2616 * WARNING: This routine must only be called when we have already sent
2617 * a SYN packet that crossed the incoming SYN that caused this routine
2618 * to get called. If this assumption fails then the initial rcv_wnd
2619 * and rcv_wscale values will not be correct.
2621 int tcp_send_synack(struct sock
*sk
)
2623 struct sk_buff
*skb
;
2625 skb
= tcp_write_queue_head(sk
);
2626 if (skb
== NULL
|| !(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)) {
2627 pr_debug("%s: wrong queue state\n", __func__
);
2630 if (!(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_ACK
)) {
2631 if (skb_cloned(skb
)) {
2632 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
2635 tcp_unlink_write_queue(skb
, sk
);
2636 skb_header_release(nskb
);
2637 __tcp_add_write_queue_head(sk
, nskb
);
2638 sk_wmem_free_skb(sk
, skb
);
2639 sk
->sk_wmem_queued
+= nskb
->truesize
;
2640 sk_mem_charge(sk
, nskb
->truesize
);
2644 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_ACK
;
2645 TCP_ECN_send_synack(tcp_sk(sk
), skb
);
2647 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2648 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2652 * tcp_make_synack - Prepare a SYN-ACK.
2653 * sk: listener socket
2654 * dst: dst entry attached to the SYNACK
2655 * req: request_sock pointer
2657 * Allocate one skb and build a SYNACK packet.
2658 * @dst is consumed : Caller should not use it again.
2660 struct sk_buff
*tcp_make_synack(struct sock
*sk
, struct dst_entry
*dst
,
2661 struct request_sock
*req
,
2662 struct tcp_fastopen_cookie
*foc
)
2664 struct tcp_out_options opts
;
2665 struct inet_request_sock
*ireq
= inet_rsk(req
);
2666 struct tcp_sock
*tp
= tcp_sk(sk
);
2668 struct sk_buff
*skb
;
2669 struct tcp_md5sig_key
*md5
;
2670 int tcp_header_size
;
2673 skb
= alloc_skb(MAX_TCP_HEADER
+ 15, sk_gfp_atomic(sk
, GFP_ATOMIC
));
2674 if (unlikely(!skb
)) {
2678 /* Reserve space for headers. */
2679 skb_reserve(skb
, MAX_TCP_HEADER
);
2681 skb_dst_set(skb
, dst
);
2682 security_skb_owned_by(skb
, sk
);
2684 mss
= dst_metric_advmss(dst
);
2685 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< mss
)
2686 mss
= tp
->rx_opt
.user_mss
;
2688 if (req
->rcv_wnd
== 0) { /* ignored for retransmitted syns */
2690 /* Set this up on the first call only */
2691 req
->window_clamp
= tp
->window_clamp
? : dst_metric(dst
, RTAX_WINDOW
);
2693 /* limit the window selection if the user enforce a smaller rx buffer */
2694 if (sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
&&
2695 (req
->window_clamp
> tcp_full_space(sk
) || req
->window_clamp
== 0))
2696 req
->window_clamp
= tcp_full_space(sk
);
2698 /* tcp_full_space because it is guaranteed to be the first packet */
2699 tcp_select_initial_window(tcp_full_space(sk
),
2700 mss
- (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0),
2705 dst_metric(dst
, RTAX_INITRWND
));
2706 ireq
->rcv_wscale
= rcv_wscale
;
2709 memset(&opts
, 0, sizeof(opts
));
2710 #ifdef CONFIG_SYN_COOKIES
2711 if (unlikely(req
->cookie_ts
))
2712 TCP_SKB_CB(skb
)->when
= cookie_init_timestamp(req
);
2715 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2716 tcp_header_size
= tcp_synack_options(sk
, req
, mss
, skb
, &opts
, &md5
,
2719 skb_push(skb
, tcp_header_size
);
2720 skb_reset_transport_header(skb
);
2723 memset(th
, 0, sizeof(struct tcphdr
));
2726 TCP_ECN_make_synack(req
, th
);
2727 th
->source
= ireq
->loc_port
;
2728 th
->dest
= ireq
->rmt_port
;
2729 /* Setting of flags are superfluous here for callers (and ECE is
2730 * not even correctly set)
2732 tcp_init_nondata_skb(skb
, tcp_rsk(req
)->snt_isn
,
2733 TCPHDR_SYN
| TCPHDR_ACK
);
2735 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
2736 /* XXX data is queued and acked as is. No buffer/window check */
2737 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_nxt
);
2739 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2740 th
->window
= htons(min(req
->rcv_wnd
, 65535U));
2741 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
);
2742 th
->doff
= (tcp_header_size
>> 2);
2743 TCP_ADD_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
, tcp_skb_pcount(skb
));
2745 #ifdef CONFIG_TCP_MD5SIG
2746 /* Okay, we have all we need - do the md5 hash if needed */
2748 tcp_rsk(req
)->af_specific
->calc_md5_hash(opts
.hash_location
,
2749 md5
, NULL
, req
, skb
);
2755 EXPORT_SYMBOL(tcp_make_synack
);
2757 /* Do all connect socket setups that can be done AF independent. */
2758 void tcp_connect_init(struct sock
*sk
)
2760 const struct dst_entry
*dst
= __sk_dst_get(sk
);
2761 struct tcp_sock
*tp
= tcp_sk(sk
);
2764 /* We'll fix this up when we get a response from the other end.
2765 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2767 tp
->tcp_header_len
= sizeof(struct tcphdr
) +
2768 (sysctl_tcp_timestamps
? TCPOLEN_TSTAMP_ALIGNED
: 0);
2770 #ifdef CONFIG_TCP_MD5SIG
2771 if (tp
->af_specific
->md5_lookup(sk
, sk
) != NULL
)
2772 tp
->tcp_header_len
+= TCPOLEN_MD5SIG_ALIGNED
;
2775 /* If user gave his TCP_MAXSEG, record it to clamp */
2776 if (tp
->rx_opt
.user_mss
)
2777 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
2780 tcp_sync_mss(sk
, dst_mtu(dst
));
2782 if (!tp
->window_clamp
)
2783 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
2784 tp
->advmss
= dst_metric_advmss(dst
);
2785 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< tp
->advmss
)
2786 tp
->advmss
= tp
->rx_opt
.user_mss
;
2788 tcp_initialize_rcv_mss(sk
);
2790 /* limit the window selection if the user enforce a smaller rx buffer */
2791 if (sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
&&
2792 (tp
->window_clamp
> tcp_full_space(sk
) || tp
->window_clamp
== 0))
2793 tp
->window_clamp
= tcp_full_space(sk
);
2795 tcp_select_initial_window(tcp_full_space(sk
),
2796 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
2799 sysctl_tcp_window_scaling
,
2801 dst_metric(dst
, RTAX_INITRWND
));
2803 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
2804 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
2807 sock_reset_flag(sk
, SOCK_DONE
);
2810 tp
->snd_una
= tp
->write_seq
;
2811 tp
->snd_sml
= tp
->write_seq
;
2812 tp
->snd_up
= tp
->write_seq
;
2813 tp
->snd_nxt
= tp
->write_seq
;
2815 if (likely(!tp
->repair
))
2817 tp
->rcv_wup
= tp
->rcv_nxt
;
2818 tp
->copied_seq
= tp
->rcv_nxt
;
2820 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
2821 inet_csk(sk
)->icsk_retransmits
= 0;
2822 tcp_clear_retrans(tp
);
2825 static void tcp_connect_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
2827 struct tcp_sock
*tp
= tcp_sk(sk
);
2828 struct tcp_skb_cb
*tcb
= TCP_SKB_CB(skb
);
2830 tcb
->end_seq
+= skb
->len
;
2831 skb_header_release(skb
);
2832 __tcp_add_write_queue_tail(sk
, skb
);
2833 sk
->sk_wmem_queued
+= skb
->truesize
;
2834 sk_mem_charge(sk
, skb
->truesize
);
2835 tp
->write_seq
= tcb
->end_seq
;
2836 tp
->packets_out
+= tcp_skb_pcount(skb
);
2839 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
2840 * queue a data-only packet after the regular SYN, such that regular SYNs
2841 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
2842 * only the SYN sequence, the data are retransmitted in the first ACK.
2843 * If cookie is not cached or other error occurs, falls back to send a
2844 * regular SYN with Fast Open cookie request option.
2846 static int tcp_send_syn_data(struct sock
*sk
, struct sk_buff
*syn
)
2848 struct tcp_sock
*tp
= tcp_sk(sk
);
2849 struct tcp_fastopen_request
*fo
= tp
->fastopen_req
;
2850 int syn_loss
= 0, space
, i
, err
= 0, iovlen
= fo
->data
->msg_iovlen
;
2851 struct sk_buff
*syn_data
= NULL
, *data
;
2852 unsigned long last_syn_loss
= 0;
2854 tp
->rx_opt
.mss_clamp
= tp
->advmss
; /* If MSS is not cached */
2855 tcp_fastopen_cache_get(sk
, &tp
->rx_opt
.mss_clamp
, &fo
->cookie
,
2856 &syn_loss
, &last_syn_loss
);
2857 /* Recurring FO SYN losses: revert to regular handshake temporarily */
2859 time_before(jiffies
, last_syn_loss
+ (60*HZ
<< syn_loss
))) {
2860 fo
->cookie
.len
= -1;
2864 if (sysctl_tcp_fastopen
& TFO_CLIENT_NO_COOKIE
)
2865 fo
->cookie
.len
= -1;
2866 else if (fo
->cookie
.len
<= 0)
2869 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
2870 * user-MSS. Reserve maximum option space for middleboxes that add
2871 * private TCP options. The cost is reduced data space in SYN :(
2873 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< tp
->rx_opt
.mss_clamp
)
2874 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
2875 space
= __tcp_mtu_to_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
) -
2876 MAX_TCP_OPTION_SPACE
;
2878 syn_data
= skb_copy_expand(syn
, skb_headroom(syn
), space
,
2880 if (syn_data
== NULL
)
2883 for (i
= 0; i
< iovlen
&& syn_data
->len
< space
; ++i
) {
2884 struct iovec
*iov
= &fo
->data
->msg_iov
[i
];
2885 unsigned char __user
*from
= iov
->iov_base
;
2886 int len
= iov
->iov_len
;
2888 if (syn_data
->len
+ len
> space
)
2889 len
= space
- syn_data
->len
;
2890 else if (i
+ 1 == iovlen
)
2891 /* No more data pending in inet_wait_for_connect() */
2894 if (skb_add_data(syn_data
, from
, len
))
2898 /* Queue a data-only packet after the regular SYN for retransmission */
2899 data
= pskb_copy(syn_data
, sk
->sk_allocation
);
2902 TCP_SKB_CB(data
)->seq
++;
2903 TCP_SKB_CB(data
)->tcp_flags
&= ~TCPHDR_SYN
;
2904 TCP_SKB_CB(data
)->tcp_flags
= (TCPHDR_ACK
|TCPHDR_PSH
);
2905 tcp_connect_queue_skb(sk
, data
);
2906 fo
->copied
= data
->len
;
2908 if (tcp_transmit_skb(sk
, syn_data
, 0, sk
->sk_allocation
) == 0) {
2909 tp
->syn_data
= (fo
->copied
> 0);
2910 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPFASTOPENACTIVE
);
2916 /* Send a regular SYN with Fast Open cookie request option */
2917 if (fo
->cookie
.len
> 0)
2919 err
= tcp_transmit_skb(sk
, syn
, 1, sk
->sk_allocation
);
2921 tp
->syn_fastopen
= 0;
2922 kfree_skb(syn_data
);
2924 fo
->cookie
.len
= -1; /* Exclude Fast Open option for SYN retries */
2928 /* Build a SYN and send it off. */
2929 int tcp_connect(struct sock
*sk
)
2931 struct tcp_sock
*tp
= tcp_sk(sk
);
2932 struct sk_buff
*buff
;
2935 tcp_connect_init(sk
);
2937 if (unlikely(tp
->repair
)) {
2938 tcp_finish_connect(sk
, NULL
);
2942 buff
= alloc_skb_fclone(MAX_TCP_HEADER
+ 15, sk
->sk_allocation
);
2943 if (unlikely(buff
== NULL
))
2946 /* Reserve space for headers. */
2947 skb_reserve(buff
, MAX_TCP_HEADER
);
2949 tcp_init_nondata_skb(buff
, tp
->write_seq
++, TCPHDR_SYN
);
2950 tp
->retrans_stamp
= TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2951 tcp_connect_queue_skb(sk
, buff
);
2952 TCP_ECN_send_syn(sk
, buff
);
2954 /* Send off SYN; include data in Fast Open. */
2955 err
= tp
->fastopen_req
? tcp_send_syn_data(sk
, buff
) :
2956 tcp_transmit_skb(sk
, buff
, 1, sk
->sk_allocation
);
2957 if (err
== -ECONNREFUSED
)
2960 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2961 * in order to make this packet get counted in tcpOutSegs.
2963 tp
->snd_nxt
= tp
->write_seq
;
2964 tp
->pushed_seq
= tp
->write_seq
;
2965 TCP_INC_STATS(sock_net(sk
), TCP_MIB_ACTIVEOPENS
);
2967 /* Timer for repeating the SYN until an answer. */
2968 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2969 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2972 EXPORT_SYMBOL(tcp_connect
);
2974 /* Send out a delayed ack, the caller does the policy checking
2975 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2978 void tcp_send_delayed_ack(struct sock
*sk
)
2980 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2981 int ato
= icsk
->icsk_ack
.ato
;
2982 unsigned long timeout
;
2984 if (ato
> TCP_DELACK_MIN
) {
2985 const struct tcp_sock
*tp
= tcp_sk(sk
);
2986 int max_ato
= HZ
/ 2;
2988 if (icsk
->icsk_ack
.pingpong
||
2989 (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
2990 max_ato
= TCP_DELACK_MAX
;
2992 /* Slow path, intersegment interval is "high". */
2994 /* If some rtt estimate is known, use it to bound delayed ack.
2995 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2999 int rtt
= max(tp
->srtt
>> 3, TCP_DELACK_MIN
);
3005 ato
= min(ato
, max_ato
);
3008 /* Stay within the limit we were given */
3009 timeout
= jiffies
+ ato
;
3011 /* Use new timeout only if there wasn't a older one earlier. */
3012 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
3013 /* If delack timer was blocked or is about to expire,
3016 if (icsk
->icsk_ack
.blocked
||
3017 time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
3022 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
3023 timeout
= icsk
->icsk_ack
.timeout
;
3025 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
3026 icsk
->icsk_ack
.timeout
= timeout
;
3027 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
3030 /* This routine sends an ack and also updates the window. */
3031 void tcp_send_ack(struct sock
*sk
)
3033 struct sk_buff
*buff
;
3035 /* If we have been reset, we may not send again. */
3036 if (sk
->sk_state
== TCP_CLOSE
)
3039 /* We are not putting this on the write queue, so
3040 * tcp_transmit_skb() will set the ownership to this
3043 buff
= alloc_skb(MAX_TCP_HEADER
, sk_gfp_atomic(sk
, GFP_ATOMIC
));
3045 inet_csk_schedule_ack(sk
);
3046 inet_csk(sk
)->icsk_ack
.ato
= TCP_ATO_MIN
;
3047 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
3048 TCP_DELACK_MAX
, TCP_RTO_MAX
);
3052 /* Reserve space for headers and prepare control bits. */
3053 skb_reserve(buff
, MAX_TCP_HEADER
);
3054 tcp_init_nondata_skb(buff
, tcp_acceptable_seq(sk
), TCPHDR_ACK
);
3056 /* Send it off, this clears delayed acks for us. */
3057 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
3058 tcp_transmit_skb(sk
, buff
, 0, sk_gfp_atomic(sk
, GFP_ATOMIC
));
3061 /* This routine sends a packet with an out of date sequence
3062 * number. It assumes the other end will try to ack it.
3064 * Question: what should we make while urgent mode?
3065 * 4.4BSD forces sending single byte of data. We cannot send
3066 * out of window data, because we have SND.NXT==SND.MAX...
3068 * Current solution: to send TWO zero-length segments in urgent mode:
3069 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3070 * out-of-date with SND.UNA-1 to probe window.
3072 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
)
3074 struct tcp_sock
*tp
= tcp_sk(sk
);
3075 struct sk_buff
*skb
;
3077 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3078 skb
= alloc_skb(MAX_TCP_HEADER
, sk_gfp_atomic(sk
, GFP_ATOMIC
));
3082 /* Reserve space for headers and set control bits. */
3083 skb_reserve(skb
, MAX_TCP_HEADER
);
3084 /* Use a previous sequence. This should cause the other
3085 * end to send an ack. Don't queue or clone SKB, just
3088 tcp_init_nondata_skb(skb
, tp
->snd_una
- !urgent
, TCPHDR_ACK
);
3089 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
3090 return tcp_transmit_skb(sk
, skb
, 0, GFP_ATOMIC
);
3093 void tcp_send_window_probe(struct sock
*sk
)
3095 if (sk
->sk_state
== TCP_ESTABLISHED
) {
3096 tcp_sk(sk
)->snd_wl1
= tcp_sk(sk
)->rcv_nxt
- 1;
3097 tcp_sk(sk
)->snd_nxt
= tcp_sk(sk
)->write_seq
;
3098 tcp_xmit_probe_skb(sk
, 0);
3102 /* Initiate keepalive or window probe from timer. */
3103 int tcp_write_wakeup(struct sock
*sk
)
3105 struct tcp_sock
*tp
= tcp_sk(sk
);
3106 struct sk_buff
*skb
;
3108 if (sk
->sk_state
== TCP_CLOSE
)
3111 if ((skb
= tcp_send_head(sk
)) != NULL
&&
3112 before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
))) {
3114 unsigned int mss
= tcp_current_mss(sk
);
3115 unsigned int seg_size
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
3117 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
3118 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
3120 /* We are probing the opening of a window
3121 * but the window size is != 0
3122 * must have been a result SWS avoidance ( sender )
3124 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
3126 seg_size
= min(seg_size
, mss
);
3127 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
3128 if (tcp_fragment(sk
, skb
, seg_size
, mss
))
3130 } else if (!tcp_skb_pcount(skb
))
3131 tcp_set_skb_tso_segs(sk
, skb
, mss
);
3133 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
3134 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
3135 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
3137 tcp_event_new_data_sent(sk
, skb
);
3140 if (between(tp
->snd_up
, tp
->snd_una
+ 1, tp
->snd_una
+ 0xFFFF))
3141 tcp_xmit_probe_skb(sk
, 1);
3142 return tcp_xmit_probe_skb(sk
, 0);
3146 /* A window probe timeout has occurred. If window is not closed send
3147 * a partial packet else a zero probe.
3149 void tcp_send_probe0(struct sock
*sk
)
3151 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3152 struct tcp_sock
*tp
= tcp_sk(sk
);
3155 err
= tcp_write_wakeup(sk
);
3157 if (tp
->packets_out
|| !tcp_send_head(sk
)) {
3158 /* Cancel probe timer, if it is not required. */
3159 icsk
->icsk_probes_out
= 0;
3160 icsk
->icsk_backoff
= 0;
3165 if (icsk
->icsk_backoff
< sysctl_tcp_retries2
)
3166 icsk
->icsk_backoff
++;
3167 icsk
->icsk_probes_out
++;
3168 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3169 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
3172 /* If packet was not sent due to local congestion,
3173 * do not backoff and do not remember icsk_probes_out.
3174 * Let local senders to fight for local resources.
3176 * Use accumulated backoff yet.
3178 if (!icsk
->icsk_probes_out
)
3179 icsk
->icsk_probes_out
= 1;
3180 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3181 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
,
3182 TCP_RESOURCE_PROBE_INTERVAL
),