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 /* Don't override Nagle indefinitely with F-RTO */
82 if (tp
->frto_counter
== 2)
85 tp
->packets_out
+= tcp_skb_pcount(skb
);
86 if (!prior_packets
|| icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
||
87 icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
)
91 /* SND.NXT, if window was not shrunk.
92 * If window has been shrunk, what should we make? It is not clear at all.
93 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
94 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
95 * invalid. OK, let's make this for now:
97 static inline __u32
tcp_acceptable_seq(const struct sock
*sk
)
99 const struct tcp_sock
*tp
= tcp_sk(sk
);
101 if (!before(tcp_wnd_end(tp
), tp
->snd_nxt
))
104 return tcp_wnd_end(tp
);
107 /* Calculate mss to advertise in SYN segment.
108 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
110 * 1. It is independent of path mtu.
111 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
112 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
113 * attached devices, because some buggy hosts are confused by
115 * 4. We do not make 3, we advertise MSS, calculated from first
116 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
117 * This may be overridden via information stored in routing table.
118 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
119 * probably even Jumbo".
121 static __u16
tcp_advertise_mss(struct sock
*sk
)
123 struct tcp_sock
*tp
= tcp_sk(sk
);
124 const struct dst_entry
*dst
= __sk_dst_get(sk
);
125 int mss
= tp
->advmss
;
128 unsigned int metric
= dst_metric_advmss(dst
);
139 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
140 * This is the first part of cwnd validation mechanism. */
141 static void tcp_cwnd_restart(struct sock
*sk
, const struct dst_entry
*dst
)
143 struct tcp_sock
*tp
= tcp_sk(sk
);
144 s32 delta
= tcp_time_stamp
- tp
->lsndtime
;
145 u32 restart_cwnd
= tcp_init_cwnd(tp
, dst
);
146 u32 cwnd
= tp
->snd_cwnd
;
148 tcp_ca_event(sk
, CA_EVENT_CWND_RESTART
);
150 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
151 restart_cwnd
= min(restart_cwnd
, cwnd
);
153 while ((delta
-= inet_csk(sk
)->icsk_rto
) > 0 && cwnd
> restart_cwnd
)
155 tp
->snd_cwnd
= max(cwnd
, restart_cwnd
);
156 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
157 tp
->snd_cwnd_used
= 0;
160 /* Congestion state accounting after a packet has been sent. */
161 static void tcp_event_data_sent(struct tcp_sock
*tp
,
164 struct inet_connection_sock
*icsk
= inet_csk(sk
);
165 const u32 now
= tcp_time_stamp
;
167 if (sysctl_tcp_slow_start_after_idle
&&
168 (!tp
->packets_out
&& (s32
)(now
- tp
->lsndtime
) > icsk
->icsk_rto
))
169 tcp_cwnd_restart(sk
, __sk_dst_get(sk
));
173 /* If it is a reply for ato after last received
174 * packet, enter pingpong mode.
176 if ((u32
)(now
- icsk
->icsk_ack
.lrcvtime
) < icsk
->icsk_ack
.ato
)
177 icsk
->icsk_ack
.pingpong
= 1;
180 /* Account for an ACK we sent. */
181 static inline void tcp_event_ack_sent(struct sock
*sk
, unsigned int pkts
)
183 tcp_dec_quickack_mode(sk
, pkts
);
184 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_DACK
);
187 /* Determine a window scaling and initial window to offer.
188 * Based on the assumption that the given amount of space
189 * will be offered. Store the results in the tp structure.
190 * NOTE: for smooth operation initial space offering should
191 * be a multiple of mss if possible. We assume here that mss >= 1.
192 * This MUST be enforced by all callers.
194 void tcp_select_initial_window(int __space
, __u32 mss
,
195 __u32
*rcv_wnd
, __u32
*window_clamp
,
196 int wscale_ok
, __u8
*rcv_wscale
,
199 unsigned int space
= (__space
< 0 ? 0 : __space
);
201 /* If no clamp set the clamp to the max possible scaled window */
202 if (*window_clamp
== 0)
203 (*window_clamp
) = (65535 << 14);
204 space
= min(*window_clamp
, space
);
206 /* Quantize space offering to a multiple of mss if possible. */
208 space
= (space
/ mss
) * mss
;
210 /* NOTE: offering an initial window larger than 32767
211 * will break some buggy TCP stacks. If the admin tells us
212 * it is likely we could be speaking with such a buggy stack
213 * we will truncate our initial window offering to 32K-1
214 * unless the remote has sent us a window scaling option,
215 * which we interpret as a sign the remote TCP is not
216 * misinterpreting the window field as a signed quantity.
218 if (sysctl_tcp_workaround_signed_windows
)
219 (*rcv_wnd
) = min(space
, MAX_TCP_WINDOW
);
225 /* Set window scaling on max possible window
226 * See RFC1323 for an explanation of the limit to 14
228 space
= max_t(u32
, sysctl_tcp_rmem
[2], sysctl_rmem_max
);
229 space
= min_t(u32
, space
, *window_clamp
);
230 while (space
> 65535 && (*rcv_wscale
) < 14) {
236 /* Set initial window to a value enough for senders starting with
237 * initial congestion window of TCP_DEFAULT_INIT_RCVWND. Place
238 * a limit on the initial window when mss is larger than 1460.
240 if (mss
> (1 << *rcv_wscale
)) {
241 int init_cwnd
= TCP_DEFAULT_INIT_RCVWND
;
244 max_t(u32
, (1460 * TCP_DEFAULT_INIT_RCVWND
) / mss
, 2);
245 /* when initializing use the value from init_rcv_wnd
246 * rather than the default from above
249 *rcv_wnd
= min(*rcv_wnd
, init_rcv_wnd
* mss
);
251 *rcv_wnd
= min(*rcv_wnd
, init_cwnd
* mss
);
254 /* Set the clamp no higher than max representable value */
255 (*window_clamp
) = min(65535U << (*rcv_wscale
), *window_clamp
);
257 EXPORT_SYMBOL(tcp_select_initial_window
);
259 /* Chose a new window to advertise, update state in tcp_sock for the
260 * socket, and return result with RFC1323 scaling applied. The return
261 * value can be stuffed directly into th->window for an outgoing
264 static u16
tcp_select_window(struct sock
*sk
)
266 struct tcp_sock
*tp
= tcp_sk(sk
);
267 u32 cur_win
= tcp_receive_window(tp
);
268 u32 new_win
= __tcp_select_window(sk
);
270 /* Never shrink the offered window */
271 if (new_win
< cur_win
) {
272 /* Danger Will Robinson!
273 * Don't update rcv_wup/rcv_wnd here or else
274 * we will not be able to advertise a zero
275 * window in time. --DaveM
277 * Relax Will Robinson.
279 new_win
= ALIGN(cur_win
, 1 << tp
->rx_opt
.rcv_wscale
);
281 tp
->rcv_wnd
= new_win
;
282 tp
->rcv_wup
= tp
->rcv_nxt
;
284 /* Make sure we do not exceed the maximum possible
287 if (!tp
->rx_opt
.rcv_wscale
&& sysctl_tcp_workaround_signed_windows
)
288 new_win
= min(new_win
, MAX_TCP_WINDOW
);
290 new_win
= min(new_win
, (65535U << tp
->rx_opt
.rcv_wscale
));
292 /* RFC1323 scaling applied */
293 new_win
>>= tp
->rx_opt
.rcv_wscale
;
295 /* If we advertise zero window, disable fast path. */
302 /* Packet ECN state for a SYN-ACK */
303 static inline void TCP_ECN_send_synack(const struct tcp_sock
*tp
, struct sk_buff
*skb
)
305 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_CWR
;
306 if (!(tp
->ecn_flags
& TCP_ECN_OK
))
307 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_ECE
;
310 /* Packet ECN state for a SYN. */
311 static inline void TCP_ECN_send_syn(struct sock
*sk
, struct sk_buff
*skb
)
313 struct tcp_sock
*tp
= tcp_sk(sk
);
316 if (sock_net(sk
)->ipv4
.sysctl_tcp_ecn
== 1) {
317 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_ECE
| TCPHDR_CWR
;
318 tp
->ecn_flags
= TCP_ECN_OK
;
322 static __inline__
void
323 TCP_ECN_make_synack(const struct request_sock
*req
, struct tcphdr
*th
)
325 if (inet_rsk(req
)->ecn_ok
)
329 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
332 static inline void TCP_ECN_send(struct sock
*sk
, struct sk_buff
*skb
,
335 struct tcp_sock
*tp
= tcp_sk(sk
);
337 if (tp
->ecn_flags
& TCP_ECN_OK
) {
338 /* Not-retransmitted data segment: set ECT and inject CWR. */
339 if (skb
->len
!= tcp_header_len
&&
340 !before(TCP_SKB_CB(skb
)->seq
, tp
->snd_nxt
)) {
342 if (tp
->ecn_flags
& TCP_ECN_QUEUE_CWR
) {
343 tp
->ecn_flags
&= ~TCP_ECN_QUEUE_CWR
;
344 tcp_hdr(skb
)->cwr
= 1;
345 skb_shinfo(skb
)->gso_type
|= SKB_GSO_TCP_ECN
;
348 /* ACK or retransmitted segment: clear ECT|CE */
349 INET_ECN_dontxmit(sk
);
351 if (tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)
352 tcp_hdr(skb
)->ece
= 1;
356 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
357 * auto increment end seqno.
359 static void tcp_init_nondata_skb(struct sk_buff
*skb
, u32 seq
, u8 flags
)
361 skb
->ip_summed
= CHECKSUM_PARTIAL
;
364 TCP_SKB_CB(skb
)->tcp_flags
= flags
;
365 TCP_SKB_CB(skb
)->sacked
= 0;
367 skb_shinfo(skb
)->gso_segs
= 1;
368 skb_shinfo(skb
)->gso_size
= 0;
369 skb_shinfo(skb
)->gso_type
= 0;
371 TCP_SKB_CB(skb
)->seq
= seq
;
372 if (flags
& (TCPHDR_SYN
| TCPHDR_FIN
))
374 TCP_SKB_CB(skb
)->end_seq
= seq
;
377 static inline bool tcp_urg_mode(const struct tcp_sock
*tp
)
379 return tp
->snd_una
!= tp
->snd_up
;
382 #define OPTION_SACK_ADVERTISE (1 << 0)
383 #define OPTION_TS (1 << 1)
384 #define OPTION_MD5 (1 << 2)
385 #define OPTION_WSCALE (1 << 3)
386 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
388 struct tcp_out_options
{
389 u16 options
; /* bit field of OPTION_* */
390 u16 mss
; /* 0 to disable */
391 u8 ws
; /* window scale, 0 to disable */
392 u8 num_sack_blocks
; /* number of SACK blocks to include */
393 u8 hash_size
; /* bytes in hash_location */
394 __u8
*hash_location
; /* temporary pointer, overloaded */
395 __u32 tsval
, tsecr
; /* need to include OPTION_TS */
396 struct tcp_fastopen_cookie
*fastopen_cookie
; /* Fast open cookie */
399 /* Write previously computed TCP options to the packet.
401 * Beware: Something in the Internet is very sensitive to the ordering of
402 * TCP options, we learned this through the hard way, so be careful here.
403 * Luckily we can at least blame others for their non-compliance but from
404 * inter-operatibility perspective it seems that we're somewhat stuck with
405 * the ordering which we have been using if we want to keep working with
406 * those broken things (not that it currently hurts anybody as there isn't
407 * particular reason why the ordering would need to be changed).
409 * At least SACK_PERM as the first option is known to lead to a disaster
410 * (but it may well be that other scenarios fail similarly).
412 static void tcp_options_write(__be32
*ptr
, struct tcp_sock
*tp
,
413 struct tcp_out_options
*opts
)
415 u16 options
= opts
->options
; /* mungable copy */
417 if (unlikely(OPTION_MD5
& options
)) {
418 *ptr
++ = htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16) |
419 (TCPOPT_MD5SIG
<< 8) | TCPOLEN_MD5SIG
);
420 /* overload cookie hash location */
421 opts
->hash_location
= (__u8
*)ptr
;
425 if (unlikely(opts
->mss
)) {
426 *ptr
++ = htonl((TCPOPT_MSS
<< 24) |
427 (TCPOLEN_MSS
<< 16) |
431 if (likely(OPTION_TS
& options
)) {
432 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
433 *ptr
++ = htonl((TCPOPT_SACK_PERM
<< 24) |
434 (TCPOLEN_SACK_PERM
<< 16) |
435 (TCPOPT_TIMESTAMP
<< 8) |
437 options
&= ~OPTION_SACK_ADVERTISE
;
439 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
441 (TCPOPT_TIMESTAMP
<< 8) |
444 *ptr
++ = htonl(opts
->tsval
);
445 *ptr
++ = htonl(opts
->tsecr
);
448 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
449 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
451 (TCPOPT_SACK_PERM
<< 8) |
455 if (unlikely(OPTION_WSCALE
& options
)) {
456 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
457 (TCPOPT_WINDOW
<< 16) |
458 (TCPOLEN_WINDOW
<< 8) |
462 if (unlikely(opts
->num_sack_blocks
)) {
463 struct tcp_sack_block
*sp
= tp
->rx_opt
.dsack
?
464 tp
->duplicate_sack
: tp
->selective_acks
;
467 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
470 (TCPOLEN_SACK_BASE
+ (opts
->num_sack_blocks
*
471 TCPOLEN_SACK_PERBLOCK
)));
473 for (this_sack
= 0; this_sack
< opts
->num_sack_blocks
;
475 *ptr
++ = htonl(sp
[this_sack
].start_seq
);
476 *ptr
++ = htonl(sp
[this_sack
].end_seq
);
479 tp
->rx_opt
.dsack
= 0;
482 if (unlikely(OPTION_FAST_OPEN_COOKIE
& options
)) {
483 struct tcp_fastopen_cookie
*foc
= opts
->fastopen_cookie
;
485 *ptr
++ = htonl((TCPOPT_EXP
<< 24) |
486 ((TCPOLEN_EXP_FASTOPEN_BASE
+ foc
->len
) << 16) |
487 TCPOPT_FASTOPEN_MAGIC
);
489 memcpy(ptr
, foc
->val
, foc
->len
);
490 if ((foc
->len
& 3) == 2) {
491 u8
*align
= ((u8
*)ptr
) + foc
->len
;
492 align
[0] = align
[1] = TCPOPT_NOP
;
494 ptr
+= (foc
->len
+ 3) >> 2;
498 /* Compute TCP options for SYN packets. This is not the final
499 * network wire format yet.
501 static unsigned int tcp_syn_options(struct sock
*sk
, struct sk_buff
*skb
,
502 struct tcp_out_options
*opts
,
503 struct tcp_md5sig_key
**md5
)
505 struct tcp_sock
*tp
= tcp_sk(sk
);
506 unsigned int remaining
= MAX_TCP_OPTION_SPACE
;
507 struct tcp_fastopen_request
*fastopen
= tp
->fastopen_req
;
509 #ifdef CONFIG_TCP_MD5SIG
510 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
512 opts
->options
|= OPTION_MD5
;
513 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
519 /* We always get an MSS option. The option bytes which will be seen in
520 * normal data packets should timestamps be used, must be in the MSS
521 * advertised. But we subtract them from tp->mss_cache so that
522 * calculations in tcp_sendmsg are simpler etc. So account for this
523 * fact here if necessary. If we don't do this correctly, as a
524 * receiver we won't recognize data packets as being full sized when we
525 * should, and thus we won't abide by the delayed ACK rules correctly.
526 * SACKs don't matter, we never delay an ACK when we have any of those
528 opts
->mss
= tcp_advertise_mss(sk
);
529 remaining
-= TCPOLEN_MSS_ALIGNED
;
531 if (likely(sysctl_tcp_timestamps
&& *md5
== NULL
)) {
532 opts
->options
|= OPTION_TS
;
533 opts
->tsval
= TCP_SKB_CB(skb
)->when
+ tp
->tsoffset
;
534 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
535 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
537 if (likely(sysctl_tcp_window_scaling
)) {
538 opts
->ws
= tp
->rx_opt
.rcv_wscale
;
539 opts
->options
|= OPTION_WSCALE
;
540 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
542 if (likely(sysctl_tcp_sack
)) {
543 opts
->options
|= OPTION_SACK_ADVERTISE
;
544 if (unlikely(!(OPTION_TS
& opts
->options
)))
545 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
548 if (fastopen
&& fastopen
->cookie
.len
>= 0) {
549 u32 need
= TCPOLEN_EXP_FASTOPEN_BASE
+ fastopen
->cookie
.len
;
550 need
= (need
+ 3) & ~3U; /* Align to 32 bits */
551 if (remaining
>= need
) {
552 opts
->options
|= OPTION_FAST_OPEN_COOKIE
;
553 opts
->fastopen_cookie
= &fastopen
->cookie
;
555 tp
->syn_fastopen
= 1;
559 return MAX_TCP_OPTION_SPACE
- remaining
;
562 /* Set up TCP options for SYN-ACKs. */
563 static unsigned int tcp_synack_options(struct sock
*sk
,
564 struct request_sock
*req
,
565 unsigned int mss
, struct sk_buff
*skb
,
566 struct tcp_out_options
*opts
,
567 struct tcp_md5sig_key
**md5
,
568 struct tcp_fastopen_cookie
*foc
)
570 struct inet_request_sock
*ireq
= inet_rsk(req
);
571 unsigned int remaining
= MAX_TCP_OPTION_SPACE
;
573 #ifdef CONFIG_TCP_MD5SIG
574 *md5
= tcp_rsk(req
)->af_specific
->md5_lookup(sk
, req
);
576 opts
->options
|= OPTION_MD5
;
577 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
579 /* We can't fit any SACK blocks in a packet with MD5 + TS
580 * options. There was discussion about disabling SACK
581 * rather than TS in order to fit in better with old,
582 * buggy kernels, but that was deemed to be unnecessary.
584 ireq
->tstamp_ok
&= !ireq
->sack_ok
;
590 /* We always send an MSS option. */
592 remaining
-= TCPOLEN_MSS_ALIGNED
;
594 if (likely(ireq
->wscale_ok
)) {
595 opts
->ws
= ireq
->rcv_wscale
;
596 opts
->options
|= OPTION_WSCALE
;
597 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
599 if (likely(ireq
->tstamp_ok
)) {
600 opts
->options
|= OPTION_TS
;
601 opts
->tsval
= TCP_SKB_CB(skb
)->when
;
602 opts
->tsecr
= req
->ts_recent
;
603 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
605 if (likely(ireq
->sack_ok
)) {
606 opts
->options
|= OPTION_SACK_ADVERTISE
;
607 if (unlikely(!ireq
->tstamp_ok
))
608 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
611 u32 need
= TCPOLEN_EXP_FASTOPEN_BASE
+ foc
->len
;
612 need
= (need
+ 3) & ~3U; /* Align to 32 bits */
613 if (remaining
>= need
) {
614 opts
->options
|= OPTION_FAST_OPEN_COOKIE
;
615 opts
->fastopen_cookie
= foc
;
620 return MAX_TCP_OPTION_SPACE
- remaining
;
623 /* Compute TCP options for ESTABLISHED sockets. This is not the
624 * final wire format yet.
626 static unsigned int tcp_established_options(struct sock
*sk
, struct sk_buff
*skb
,
627 struct tcp_out_options
*opts
,
628 struct tcp_md5sig_key
**md5
)
630 struct tcp_skb_cb
*tcb
= skb
? TCP_SKB_CB(skb
) : NULL
;
631 struct tcp_sock
*tp
= tcp_sk(sk
);
632 unsigned int size
= 0;
633 unsigned int eff_sacks
;
635 #ifdef CONFIG_TCP_MD5SIG
636 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
637 if (unlikely(*md5
)) {
638 opts
->options
|= OPTION_MD5
;
639 size
+= TCPOLEN_MD5SIG_ALIGNED
;
645 if (likely(tp
->rx_opt
.tstamp_ok
)) {
646 opts
->options
|= OPTION_TS
;
647 opts
->tsval
= tcb
? tcb
->when
+ tp
->tsoffset
: 0;
648 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
649 size
+= TCPOLEN_TSTAMP_ALIGNED
;
652 eff_sacks
= tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
;
653 if (unlikely(eff_sacks
)) {
654 const unsigned int remaining
= MAX_TCP_OPTION_SPACE
- size
;
655 opts
->num_sack_blocks
=
656 min_t(unsigned int, eff_sacks
,
657 (remaining
- TCPOLEN_SACK_BASE_ALIGNED
) /
658 TCPOLEN_SACK_PERBLOCK
);
659 size
+= TCPOLEN_SACK_BASE_ALIGNED
+
660 opts
->num_sack_blocks
* TCPOLEN_SACK_PERBLOCK
;
667 /* TCP SMALL QUEUES (TSQ)
669 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
670 * to reduce RTT and bufferbloat.
671 * We do this using a special skb destructor (tcp_wfree).
673 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
674 * needs to be reallocated in a driver.
675 * The invariant being skb->truesize substracted from sk->sk_wmem_alloc
677 * Since transmit from skb destructor is forbidden, we use a tasklet
678 * to process all sockets that eventually need to send more skbs.
679 * We use one tasklet per cpu, with its own queue of sockets.
682 struct tasklet_struct tasklet
;
683 struct list_head head
; /* queue of tcp sockets */
685 static DEFINE_PER_CPU(struct tsq_tasklet
, tsq_tasklet
);
687 static void tcp_tsq_handler(struct sock
*sk
)
689 if ((1 << sk
->sk_state
) &
690 (TCPF_ESTABLISHED
| TCPF_FIN_WAIT1
| TCPF_CLOSING
|
691 TCPF_CLOSE_WAIT
| TCPF_LAST_ACK
))
692 tcp_write_xmit(sk
, tcp_current_mss(sk
), 0, 0, GFP_ATOMIC
);
695 * One tasklest per cpu tries to send more skbs.
696 * We run in tasklet context but need to disable irqs when
697 * transfering tsq->head because tcp_wfree() might
698 * interrupt us (non NAPI drivers)
700 static void tcp_tasklet_func(unsigned long data
)
702 struct tsq_tasklet
*tsq
= (struct tsq_tasklet
*)data
;
705 struct list_head
*q
, *n
;
709 local_irq_save(flags
);
710 list_splice_init(&tsq
->head
, &list
);
711 local_irq_restore(flags
);
713 list_for_each_safe(q
, n
, &list
) {
714 tp
= list_entry(q
, struct tcp_sock
, tsq_node
);
715 list_del(&tp
->tsq_node
);
717 sk
= (struct sock
*)tp
;
720 if (!sock_owned_by_user(sk
)) {
723 /* defer the work to tcp_release_cb() */
724 set_bit(TCP_TSQ_DEFERRED
, &tp
->tsq_flags
);
728 clear_bit(TSQ_QUEUED
, &tp
->tsq_flags
);
733 #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \
734 (1UL << TCP_WRITE_TIMER_DEFERRED) | \
735 (1UL << TCP_DELACK_TIMER_DEFERRED) | \
736 (1UL << TCP_MTU_REDUCED_DEFERRED))
738 * tcp_release_cb - tcp release_sock() callback
741 * called from release_sock() to perform protocol dependent
742 * actions before socket release.
744 void tcp_release_cb(struct sock
*sk
)
746 struct tcp_sock
*tp
= tcp_sk(sk
);
747 unsigned long flags
, nflags
;
749 /* perform an atomic operation only if at least one flag is set */
751 flags
= tp
->tsq_flags
;
752 if (!(flags
& TCP_DEFERRED_ALL
))
754 nflags
= flags
& ~TCP_DEFERRED_ALL
;
755 } while (cmpxchg(&tp
->tsq_flags
, flags
, nflags
) != flags
);
757 if (flags
& (1UL << TCP_TSQ_DEFERRED
))
760 if (flags
& (1UL << TCP_WRITE_TIMER_DEFERRED
)) {
761 tcp_write_timer_handler(sk
);
764 if (flags
& (1UL << TCP_DELACK_TIMER_DEFERRED
)) {
765 tcp_delack_timer_handler(sk
);
768 if (flags
& (1UL << TCP_MTU_REDUCED_DEFERRED
)) {
769 sk
->sk_prot
->mtu_reduced(sk
);
773 EXPORT_SYMBOL(tcp_release_cb
);
775 void __init
tcp_tasklet_init(void)
779 for_each_possible_cpu(i
) {
780 struct tsq_tasklet
*tsq
= &per_cpu(tsq_tasklet
, i
);
782 INIT_LIST_HEAD(&tsq
->head
);
783 tasklet_init(&tsq
->tasklet
,
790 * Write buffer destructor automatically called from kfree_skb.
791 * We cant xmit new skbs from this context, as we might already
794 static void tcp_wfree(struct sk_buff
*skb
)
796 struct sock
*sk
= skb
->sk
;
797 struct tcp_sock
*tp
= tcp_sk(sk
);
799 if (test_and_clear_bit(TSQ_THROTTLED
, &tp
->tsq_flags
) &&
800 !test_and_set_bit(TSQ_QUEUED
, &tp
->tsq_flags
)) {
802 struct tsq_tasklet
*tsq
;
804 /* Keep a ref on socket.
805 * This last ref will be released in tcp_tasklet_func()
807 atomic_sub(skb
->truesize
- 1, &sk
->sk_wmem_alloc
);
809 /* queue this socket to tasklet queue */
810 local_irq_save(flags
);
811 tsq
= &__get_cpu_var(tsq_tasklet
);
812 list_add(&tp
->tsq_node
, &tsq
->head
);
813 tasklet_schedule(&tsq
->tasklet
);
814 local_irq_restore(flags
);
820 /* This routine actually transmits TCP packets queued in by
821 * tcp_do_sendmsg(). This is used by both the initial
822 * transmission and possible later retransmissions.
823 * All SKB's seen here are completely headerless. It is our
824 * job to build the TCP header, and pass the packet down to
825 * IP so it can do the same plus pass the packet off to the
828 * We are working here with either a clone of the original
829 * SKB, or a fresh unique copy made by the retransmit engine.
831 static int tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int clone_it
,
834 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
835 struct inet_sock
*inet
;
837 struct tcp_skb_cb
*tcb
;
838 struct tcp_out_options opts
;
839 unsigned int tcp_options_size
, tcp_header_size
;
840 struct tcp_md5sig_key
*md5
;
844 BUG_ON(!skb
|| !tcp_skb_pcount(skb
));
846 /* If congestion control is doing timestamping, we must
847 * take such a timestamp before we potentially clone/copy.
849 if (icsk
->icsk_ca_ops
->flags
& TCP_CONG_RTT_STAMP
)
850 __net_timestamp(skb
);
852 if (likely(clone_it
)) {
853 if (unlikely(skb_cloned(skb
)))
854 skb
= pskb_copy(skb
, gfp_mask
);
856 skb
= skb_clone(skb
, gfp_mask
);
863 tcb
= TCP_SKB_CB(skb
);
864 memset(&opts
, 0, sizeof(opts
));
866 if (unlikely(tcb
->tcp_flags
& TCPHDR_SYN
))
867 tcp_options_size
= tcp_syn_options(sk
, skb
, &opts
, &md5
);
869 tcp_options_size
= tcp_established_options(sk
, skb
, &opts
,
871 tcp_header_size
= tcp_options_size
+ sizeof(struct tcphdr
);
873 if (tcp_packets_in_flight(tp
) == 0) {
874 tcp_ca_event(sk
, CA_EVENT_TX_START
);
879 skb_push(skb
, tcp_header_size
);
880 skb_reset_transport_header(skb
);
884 skb
->destructor
= (sysctl_tcp_limit_output_bytes
> 0) ?
885 tcp_wfree
: sock_wfree
;
886 atomic_add(skb
->truesize
, &sk
->sk_wmem_alloc
);
888 /* Build TCP header and checksum it. */
890 th
->source
= inet
->inet_sport
;
891 th
->dest
= inet
->inet_dport
;
892 th
->seq
= htonl(tcb
->seq
);
893 th
->ack_seq
= htonl(tp
->rcv_nxt
);
894 *(((__be16
*)th
) + 6) = htons(((tcp_header_size
>> 2) << 12) |
897 if (unlikely(tcb
->tcp_flags
& TCPHDR_SYN
)) {
898 /* RFC1323: The window in SYN & SYN/ACK segments
901 th
->window
= htons(min(tp
->rcv_wnd
, 65535U));
903 th
->window
= htons(tcp_select_window(sk
));
908 /* The urg_mode check is necessary during a below snd_una win probe */
909 if (unlikely(tcp_urg_mode(tp
) && before(tcb
->seq
, tp
->snd_up
))) {
910 if (before(tp
->snd_up
, tcb
->seq
+ 0x10000)) {
911 th
->urg_ptr
= htons(tp
->snd_up
- tcb
->seq
);
913 } else if (after(tcb
->seq
+ 0xFFFF, tp
->snd_nxt
)) {
914 th
->urg_ptr
= htons(0xFFFF);
919 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
);
920 if (likely((tcb
->tcp_flags
& TCPHDR_SYN
) == 0))
921 TCP_ECN_send(sk
, skb
, tcp_header_size
);
923 #ifdef CONFIG_TCP_MD5SIG
924 /* Calculate the MD5 hash, as we have all we need now */
926 sk_nocaps_add(sk
, NETIF_F_GSO_MASK
);
927 tp
->af_specific
->calc_md5_hash(opts
.hash_location
,
932 icsk
->icsk_af_ops
->send_check(sk
, skb
);
934 if (likely(tcb
->tcp_flags
& TCPHDR_ACK
))
935 tcp_event_ack_sent(sk
, tcp_skb_pcount(skb
));
937 if (skb
->len
!= tcp_header_size
)
938 tcp_event_data_sent(tp
, sk
);
940 if (after(tcb
->end_seq
, tp
->snd_nxt
) || tcb
->seq
== tcb
->end_seq
)
941 TCP_ADD_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
,
942 tcp_skb_pcount(skb
));
944 err
= icsk
->icsk_af_ops
->queue_xmit(skb
, &inet
->cork
.fl
);
945 if (likely(err
<= 0))
948 tcp_enter_cwr(sk
, 1);
950 return net_xmit_eval(err
);
953 /* This routine just queues the buffer for sending.
955 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
956 * otherwise socket can stall.
958 static void tcp_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
960 struct tcp_sock
*tp
= tcp_sk(sk
);
962 /* Advance write_seq and place onto the write_queue. */
963 tp
->write_seq
= TCP_SKB_CB(skb
)->end_seq
;
964 skb_header_release(skb
);
965 tcp_add_write_queue_tail(sk
, skb
);
966 sk
->sk_wmem_queued
+= skb
->truesize
;
967 sk_mem_charge(sk
, skb
->truesize
);
970 /* Initialize TSO segments for a packet. */
971 static void tcp_set_skb_tso_segs(const struct sock
*sk
, struct sk_buff
*skb
,
972 unsigned int mss_now
)
974 if (skb
->len
<= mss_now
|| !sk_can_gso(sk
) ||
975 skb
->ip_summed
== CHECKSUM_NONE
) {
976 /* Avoid the costly divide in the normal
979 skb_shinfo(skb
)->gso_segs
= 1;
980 skb_shinfo(skb
)->gso_size
= 0;
981 skb_shinfo(skb
)->gso_type
= 0;
983 skb_shinfo(skb
)->gso_segs
= DIV_ROUND_UP(skb
->len
, mss_now
);
984 skb_shinfo(skb
)->gso_size
= mss_now
;
985 skb_shinfo(skb
)->gso_type
= sk
->sk_gso_type
;
989 /* When a modification to fackets out becomes necessary, we need to check
990 * skb is counted to fackets_out or not.
992 static void tcp_adjust_fackets_out(struct sock
*sk
, const struct sk_buff
*skb
,
995 struct tcp_sock
*tp
= tcp_sk(sk
);
997 if (!tp
->sacked_out
|| tcp_is_reno(tp
))
1000 if (after(tcp_highest_sack_seq(tp
), TCP_SKB_CB(skb
)->seq
))
1001 tp
->fackets_out
-= decr
;
1004 /* Pcount in the middle of the write queue got changed, we need to do various
1005 * tweaks to fix counters
1007 static void tcp_adjust_pcount(struct sock
*sk
, const struct sk_buff
*skb
, int decr
)
1009 struct tcp_sock
*tp
= tcp_sk(sk
);
1011 tp
->packets_out
-= decr
;
1013 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
1014 tp
->sacked_out
-= decr
;
1015 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
1016 tp
->retrans_out
-= decr
;
1017 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
)
1018 tp
->lost_out
-= decr
;
1020 /* Reno case is special. Sigh... */
1021 if (tcp_is_reno(tp
) && decr
> 0)
1022 tp
->sacked_out
-= min_t(u32
, tp
->sacked_out
, decr
);
1024 tcp_adjust_fackets_out(sk
, skb
, decr
);
1026 if (tp
->lost_skb_hint
&&
1027 before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
) &&
1028 (tcp_is_fack(tp
) || (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)))
1029 tp
->lost_cnt_hint
-= decr
;
1031 tcp_verify_left_out(tp
);
1034 /* Function to create two new TCP segments. Shrinks the given segment
1035 * to the specified size and appends a new segment with the rest of the
1036 * packet to the list. This won't be called frequently, I hope.
1037 * Remember, these are still headerless SKBs at this point.
1039 int tcp_fragment(struct sock
*sk
, struct sk_buff
*skb
, u32 len
,
1040 unsigned int mss_now
)
1042 struct tcp_sock
*tp
= tcp_sk(sk
);
1043 struct sk_buff
*buff
;
1044 int nsize
, old_factor
;
1048 if (WARN_ON(len
> skb
->len
))
1051 nsize
= skb_headlen(skb
) - len
;
1055 if (skb_cloned(skb
) &&
1056 skb_is_nonlinear(skb
) &&
1057 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
1060 /* Get a new skb... force flag on. */
1061 buff
= sk_stream_alloc_skb(sk
, nsize
, GFP_ATOMIC
);
1063 return -ENOMEM
; /* We'll just try again later. */
1065 sk
->sk_wmem_queued
+= buff
->truesize
;
1066 sk_mem_charge(sk
, buff
->truesize
);
1067 nlen
= skb
->len
- len
- nsize
;
1068 buff
->truesize
+= nlen
;
1069 skb
->truesize
-= nlen
;
1071 /* Correct the sequence numbers. */
1072 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1073 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1074 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1076 /* PSH and FIN should only be set in the second packet. */
1077 flags
= TCP_SKB_CB(skb
)->tcp_flags
;
1078 TCP_SKB_CB(skb
)->tcp_flags
= flags
& ~(TCPHDR_FIN
| TCPHDR_PSH
);
1079 TCP_SKB_CB(buff
)->tcp_flags
= flags
;
1080 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
1082 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_PARTIAL
) {
1083 /* Copy and checksum data tail into the new buffer. */
1084 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
,
1085 skb_put(buff
, nsize
),
1090 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
1092 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1093 skb_split(skb
, buff
, len
);
1096 buff
->ip_summed
= skb
->ip_summed
;
1098 /* Looks stupid, but our code really uses when of
1099 * skbs, which it never sent before. --ANK
1101 TCP_SKB_CB(buff
)->when
= TCP_SKB_CB(skb
)->when
;
1102 buff
->tstamp
= skb
->tstamp
;
1104 old_factor
= tcp_skb_pcount(skb
);
1106 /* Fix up tso_factor for both original and new SKB. */
1107 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1108 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1110 /* If this packet has been sent out already, we must
1111 * adjust the various packet counters.
1113 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
1114 int diff
= old_factor
- tcp_skb_pcount(skb
) -
1115 tcp_skb_pcount(buff
);
1118 tcp_adjust_pcount(sk
, skb
, diff
);
1121 /* Link BUFF into the send queue. */
1122 skb_header_release(buff
);
1123 tcp_insert_write_queue_after(skb
, buff
, sk
);
1128 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1129 * eventually). The difference is that pulled data not copied, but
1130 * immediately discarded.
1132 static void __pskb_trim_head(struct sk_buff
*skb
, int len
)
1136 eat
= min_t(int, len
, skb_headlen(skb
));
1138 __skb_pull(skb
, eat
);
1139 skb
->avail_size
-= eat
;
1146 for (i
= 0; i
< skb_shinfo(skb
)->nr_frags
; i
++) {
1147 int size
= skb_frag_size(&skb_shinfo(skb
)->frags
[i
]);
1150 skb_frag_unref(skb
, i
);
1153 skb_shinfo(skb
)->frags
[k
] = skb_shinfo(skb
)->frags
[i
];
1155 skb_shinfo(skb
)->frags
[k
].page_offset
+= eat
;
1156 skb_frag_size_sub(&skb_shinfo(skb
)->frags
[k
], eat
);
1162 skb_shinfo(skb
)->nr_frags
= k
;
1164 skb_reset_tail_pointer(skb
);
1165 skb
->data_len
-= len
;
1166 skb
->len
= skb
->data_len
;
1169 /* Remove acked data from a packet in the transmit queue. */
1170 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
1172 if (skb_unclone(skb
, GFP_ATOMIC
))
1175 __pskb_trim_head(skb
, len
);
1177 TCP_SKB_CB(skb
)->seq
+= len
;
1178 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1180 skb
->truesize
-= len
;
1181 sk
->sk_wmem_queued
-= len
;
1182 sk_mem_uncharge(sk
, len
);
1183 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
1185 /* Any change of skb->len requires recalculation of tso factor. */
1186 if (tcp_skb_pcount(skb
) > 1)
1187 tcp_set_skb_tso_segs(sk
, skb
, tcp_skb_mss(skb
));
1192 /* Calculate MSS not accounting any TCP options. */
1193 static inline int __tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1195 const struct tcp_sock
*tp
= tcp_sk(sk
);
1196 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1199 /* Calculate base mss without TCP options:
1200 It is MMS_S - sizeof(tcphdr) of rfc1122
1202 mss_now
= pmtu
- icsk
->icsk_af_ops
->net_header_len
- sizeof(struct tcphdr
);
1204 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1205 if (icsk
->icsk_af_ops
->net_frag_header_len
) {
1206 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1208 if (dst
&& dst_allfrag(dst
))
1209 mss_now
-= icsk
->icsk_af_ops
->net_frag_header_len
;
1212 /* Clamp it (mss_clamp does not include tcp options) */
1213 if (mss_now
> tp
->rx_opt
.mss_clamp
)
1214 mss_now
= tp
->rx_opt
.mss_clamp
;
1216 /* Now subtract optional transport overhead */
1217 mss_now
-= icsk
->icsk_ext_hdr_len
;
1219 /* Then reserve room for full set of TCP options and 8 bytes of data */
1225 /* Calculate MSS. Not accounting for SACKs here. */
1226 int tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1228 /* Subtract TCP options size, not including SACKs */
1229 return __tcp_mtu_to_mss(sk
, pmtu
) -
1230 (tcp_sk(sk
)->tcp_header_len
- sizeof(struct tcphdr
));
1233 /* Inverse of above */
1234 int tcp_mss_to_mtu(struct sock
*sk
, int mss
)
1236 const struct tcp_sock
*tp
= tcp_sk(sk
);
1237 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1241 tp
->tcp_header_len
+
1242 icsk
->icsk_ext_hdr_len
+
1243 icsk
->icsk_af_ops
->net_header_len
;
1245 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1246 if (icsk
->icsk_af_ops
->net_frag_header_len
) {
1247 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1249 if (dst
&& dst_allfrag(dst
))
1250 mtu
+= icsk
->icsk_af_ops
->net_frag_header_len
;
1255 /* MTU probing init per socket */
1256 void tcp_mtup_init(struct sock
*sk
)
1258 struct tcp_sock
*tp
= tcp_sk(sk
);
1259 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1261 icsk
->icsk_mtup
.enabled
= sysctl_tcp_mtu_probing
> 1;
1262 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+ sizeof(struct tcphdr
) +
1263 icsk
->icsk_af_ops
->net_header_len
;
1264 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, sysctl_tcp_base_mss
);
1265 icsk
->icsk_mtup
.probe_size
= 0;
1267 EXPORT_SYMBOL(tcp_mtup_init
);
1269 /* This function synchronize snd mss to current pmtu/exthdr set.
1271 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1272 for TCP options, but includes only bare TCP header.
1274 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1275 It is minimum of user_mss and mss received with SYN.
1276 It also does not include TCP options.
1278 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1280 tp->mss_cache is current effective sending mss, including
1281 all tcp options except for SACKs. It is evaluated,
1282 taking into account current pmtu, but never exceeds
1283 tp->rx_opt.mss_clamp.
1285 NOTE1. rfc1122 clearly states that advertised MSS
1286 DOES NOT include either tcp or ip options.
1288 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1289 are READ ONLY outside this function. --ANK (980731)
1291 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
1293 struct tcp_sock
*tp
= tcp_sk(sk
);
1294 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1297 if (icsk
->icsk_mtup
.search_high
> pmtu
)
1298 icsk
->icsk_mtup
.search_high
= pmtu
;
1300 mss_now
= tcp_mtu_to_mss(sk
, pmtu
);
1301 mss_now
= tcp_bound_to_half_wnd(tp
, mss_now
);
1303 /* And store cached results */
1304 icsk
->icsk_pmtu_cookie
= pmtu
;
1305 if (icsk
->icsk_mtup
.enabled
)
1306 mss_now
= min(mss_now
, tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
));
1307 tp
->mss_cache
= mss_now
;
1311 EXPORT_SYMBOL(tcp_sync_mss
);
1313 /* Compute the current effective MSS, taking SACKs and IP options,
1314 * and even PMTU discovery events into account.
1316 unsigned int tcp_current_mss(struct sock
*sk
)
1318 const struct tcp_sock
*tp
= tcp_sk(sk
);
1319 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1321 unsigned int header_len
;
1322 struct tcp_out_options opts
;
1323 struct tcp_md5sig_key
*md5
;
1325 mss_now
= tp
->mss_cache
;
1328 u32 mtu
= dst_mtu(dst
);
1329 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
1330 mss_now
= tcp_sync_mss(sk
, mtu
);
1333 header_len
= tcp_established_options(sk
, NULL
, &opts
, &md5
) +
1334 sizeof(struct tcphdr
);
1335 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1336 * some common options. If this is an odd packet (because we have SACK
1337 * blocks etc) then our calculated header_len will be different, and
1338 * we have to adjust mss_now correspondingly */
1339 if (header_len
!= tp
->tcp_header_len
) {
1340 int delta
= (int) header_len
- tp
->tcp_header_len
;
1347 /* Congestion window validation. (RFC2861) */
1348 static void tcp_cwnd_validate(struct sock
*sk
)
1350 struct tcp_sock
*tp
= tcp_sk(sk
);
1352 if (tp
->packets_out
>= tp
->snd_cwnd
) {
1353 /* Network is feed fully. */
1354 tp
->snd_cwnd_used
= 0;
1355 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1357 /* Network starves. */
1358 if (tp
->packets_out
> tp
->snd_cwnd_used
)
1359 tp
->snd_cwnd_used
= tp
->packets_out
;
1361 if (sysctl_tcp_slow_start_after_idle
&&
1362 (s32
)(tcp_time_stamp
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
)
1363 tcp_cwnd_application_limited(sk
);
1367 /* Returns the portion of skb which can be sent right away without
1368 * introducing MSS oddities to segment boundaries. In rare cases where
1369 * mss_now != mss_cache, we will request caller to create a small skb
1370 * per input skb which could be mostly avoided here (if desired).
1372 * We explicitly want to create a request for splitting write queue tail
1373 * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1374 * thus all the complexity (cwnd_len is always MSS multiple which we
1375 * return whenever allowed by the other factors). Basically we need the
1376 * modulo only when the receiver window alone is the limiting factor or
1377 * when we would be allowed to send the split-due-to-Nagle skb fully.
1379 static unsigned int tcp_mss_split_point(const struct sock
*sk
, const struct sk_buff
*skb
,
1380 unsigned int mss_now
, unsigned int max_segs
)
1382 const struct tcp_sock
*tp
= tcp_sk(sk
);
1383 u32 needed
, window
, max_len
;
1385 window
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1386 max_len
= mss_now
* max_segs
;
1388 if (likely(max_len
<= window
&& skb
!= tcp_write_queue_tail(sk
)))
1391 needed
= min(skb
->len
, window
);
1393 if (max_len
<= needed
)
1396 return needed
- needed
% mss_now
;
1399 /* Can at least one segment of SKB be sent right now, according to the
1400 * congestion window rules? If so, return how many segments are allowed.
1402 static inline unsigned int tcp_cwnd_test(const struct tcp_sock
*tp
,
1403 const struct sk_buff
*skb
)
1405 u32 in_flight
, cwnd
;
1407 /* Don't be strict about the congestion window for the final FIN. */
1408 if ((TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
) &&
1409 tcp_skb_pcount(skb
) == 1)
1412 in_flight
= tcp_packets_in_flight(tp
);
1413 cwnd
= tp
->snd_cwnd
;
1414 if (in_flight
< cwnd
)
1415 return (cwnd
- in_flight
);
1420 /* Initialize TSO state of a skb.
1421 * This must be invoked the first time we consider transmitting
1422 * SKB onto the wire.
1424 static int tcp_init_tso_segs(const struct sock
*sk
, struct sk_buff
*skb
,
1425 unsigned int mss_now
)
1427 int tso_segs
= tcp_skb_pcount(skb
);
1429 if (!tso_segs
|| (tso_segs
> 1 && tcp_skb_mss(skb
) != mss_now
)) {
1430 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1431 tso_segs
= tcp_skb_pcount(skb
);
1436 /* Minshall's variant of the Nagle send check. */
1437 static inline bool tcp_minshall_check(const struct tcp_sock
*tp
)
1439 return after(tp
->snd_sml
, tp
->snd_una
) &&
1440 !after(tp
->snd_sml
, tp
->snd_nxt
);
1443 /* Return false, if packet can be sent now without violation Nagle's rules:
1444 * 1. It is full sized.
1445 * 2. Or it contains FIN. (already checked by caller)
1446 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1447 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1448 * With Minshall's modification: all sent small packets are ACKed.
1450 static inline bool tcp_nagle_check(const struct tcp_sock
*tp
,
1451 const struct sk_buff
*skb
,
1452 unsigned int mss_now
, int nonagle
)
1454 return skb
->len
< mss_now
&&
1455 ((nonagle
& TCP_NAGLE_CORK
) ||
1456 (!nonagle
&& tp
->packets_out
&& tcp_minshall_check(tp
)));
1459 /* Return true if the Nagle test allows this packet to be
1462 static inline bool tcp_nagle_test(const struct tcp_sock
*tp
, const struct sk_buff
*skb
,
1463 unsigned int cur_mss
, int nonagle
)
1465 /* Nagle rule does not apply to frames, which sit in the middle of the
1466 * write_queue (they have no chances to get new data).
1468 * This is implemented in the callers, where they modify the 'nonagle'
1469 * argument based upon the location of SKB in the send queue.
1471 if (nonagle
& TCP_NAGLE_PUSH
)
1474 /* Don't use the nagle rule for urgent data (or for the final FIN).
1475 * Nagle can be ignored during F-RTO too (see RFC4138).
1477 if (tcp_urg_mode(tp
) || (tp
->frto_counter
== 2) ||
1478 (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
))
1481 if (!tcp_nagle_check(tp
, skb
, cur_mss
, nonagle
))
1487 /* Does at least the first segment of SKB fit into the send window? */
1488 static bool tcp_snd_wnd_test(const struct tcp_sock
*tp
,
1489 const struct sk_buff
*skb
,
1490 unsigned int cur_mss
)
1492 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1494 if (skb
->len
> cur_mss
)
1495 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
1497 return !after(end_seq
, tcp_wnd_end(tp
));
1500 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1501 * should be put on the wire right now. If so, it returns the number of
1502 * packets allowed by the congestion window.
1504 static unsigned int tcp_snd_test(const struct sock
*sk
, struct sk_buff
*skb
,
1505 unsigned int cur_mss
, int nonagle
)
1507 const struct tcp_sock
*tp
= tcp_sk(sk
);
1508 unsigned int cwnd_quota
;
1510 tcp_init_tso_segs(sk
, skb
, cur_mss
);
1512 if (!tcp_nagle_test(tp
, skb
, cur_mss
, nonagle
))
1515 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1516 if (cwnd_quota
&& !tcp_snd_wnd_test(tp
, skb
, cur_mss
))
1522 /* Test if sending is allowed right now. */
1523 bool tcp_may_send_now(struct sock
*sk
)
1525 const struct tcp_sock
*tp
= tcp_sk(sk
);
1526 struct sk_buff
*skb
= tcp_send_head(sk
);
1529 tcp_snd_test(sk
, skb
, tcp_current_mss(sk
),
1530 (tcp_skb_is_last(sk
, skb
) ?
1531 tp
->nonagle
: TCP_NAGLE_PUSH
));
1534 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1535 * which is put after SKB on the list. It is very much like
1536 * tcp_fragment() except that it may make several kinds of assumptions
1537 * in order to speed up the splitting operation. In particular, we
1538 * know that all the data is in scatter-gather pages, and that the
1539 * packet has never been sent out before (and thus is not cloned).
1541 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
,
1542 unsigned int mss_now
, gfp_t gfp
)
1544 struct sk_buff
*buff
;
1545 int nlen
= skb
->len
- len
;
1548 /* All of a TSO frame must be composed of paged data. */
1549 if (skb
->len
!= skb
->data_len
)
1550 return tcp_fragment(sk
, skb
, len
, mss_now
);
1552 buff
= sk_stream_alloc_skb(sk
, 0, gfp
);
1553 if (unlikely(buff
== NULL
))
1556 sk
->sk_wmem_queued
+= buff
->truesize
;
1557 sk_mem_charge(sk
, buff
->truesize
);
1558 buff
->truesize
+= nlen
;
1559 skb
->truesize
-= nlen
;
1561 /* Correct the sequence numbers. */
1562 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1563 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1564 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1566 /* PSH and FIN should only be set in the second packet. */
1567 flags
= TCP_SKB_CB(skb
)->tcp_flags
;
1568 TCP_SKB_CB(skb
)->tcp_flags
= flags
& ~(TCPHDR_FIN
| TCPHDR_PSH
);
1569 TCP_SKB_CB(buff
)->tcp_flags
= flags
;
1571 /* This packet was never sent out yet, so no SACK bits. */
1572 TCP_SKB_CB(buff
)->sacked
= 0;
1574 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_PARTIAL
;
1575 skb_split(skb
, buff
, len
);
1577 /* Fix up tso_factor for both original and new SKB. */
1578 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1579 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1581 /* Link BUFF into the send queue. */
1582 skb_header_release(buff
);
1583 tcp_insert_write_queue_after(skb
, buff
, sk
);
1588 /* Try to defer sending, if possible, in order to minimize the amount
1589 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1591 * This algorithm is from John Heffner.
1593 static bool tcp_tso_should_defer(struct sock
*sk
, struct sk_buff
*skb
)
1595 struct tcp_sock
*tp
= tcp_sk(sk
);
1596 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1597 u32 send_win
, cong_win
, limit
, in_flight
;
1600 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
1603 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
1606 /* Defer for less than two clock ticks. */
1607 if (tp
->tso_deferred
&&
1608 (((u32
)jiffies
<< 1) >> 1) - (tp
->tso_deferred
>> 1) > 1)
1611 in_flight
= tcp_packets_in_flight(tp
);
1613 BUG_ON(tcp_skb_pcount(skb
) <= 1 || (tp
->snd_cwnd
<= in_flight
));
1615 send_win
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1617 /* From in_flight test above, we know that cwnd > in_flight. */
1618 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
1620 limit
= min(send_win
, cong_win
);
1622 /* If a full-sized TSO skb can be sent, do it. */
1623 if (limit
>= min_t(unsigned int, sk
->sk_gso_max_size
,
1624 sk
->sk_gso_max_segs
* tp
->mss_cache
))
1627 /* Middle in queue won't get any more data, full sendable already? */
1628 if ((skb
!= tcp_write_queue_tail(sk
)) && (limit
>= skb
->len
))
1631 win_divisor
= ACCESS_ONCE(sysctl_tcp_tso_win_divisor
);
1633 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
1635 /* If at least some fraction of a window is available,
1638 chunk
/= win_divisor
;
1642 /* Different approach, try not to defer past a single
1643 * ACK. Receiver should ACK every other full sized
1644 * frame, so if we have space for more than 3 frames
1647 if (limit
> tcp_max_tso_deferred_mss(tp
) * tp
->mss_cache
)
1651 /* Ok, it looks like it is advisable to defer. */
1652 tp
->tso_deferred
= 1 | (jiffies
<< 1);
1657 tp
->tso_deferred
= 0;
1661 /* Create a new MTU probe if we are ready.
1662 * MTU probe is regularly attempting to increase the path MTU by
1663 * deliberately sending larger packets. This discovers routing
1664 * changes resulting in larger path MTUs.
1666 * Returns 0 if we should wait to probe (no cwnd available),
1667 * 1 if a probe was sent,
1670 static int tcp_mtu_probe(struct sock
*sk
)
1672 struct tcp_sock
*tp
= tcp_sk(sk
);
1673 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1674 struct sk_buff
*skb
, *nskb
, *next
;
1681 /* Not currently probing/verifying,
1683 * have enough cwnd, and
1684 * not SACKing (the variable headers throw things off) */
1685 if (!icsk
->icsk_mtup
.enabled
||
1686 icsk
->icsk_mtup
.probe_size
||
1687 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
||
1688 tp
->snd_cwnd
< 11 ||
1689 tp
->rx_opt
.num_sacks
|| tp
->rx_opt
.dsack
)
1692 /* Very simple search strategy: just double the MSS. */
1693 mss_now
= tcp_current_mss(sk
);
1694 probe_size
= 2 * tp
->mss_cache
;
1695 size_needed
= probe_size
+ (tp
->reordering
+ 1) * tp
->mss_cache
;
1696 if (probe_size
> tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_high
)) {
1697 /* TODO: set timer for probe_converge_event */
1701 /* Have enough data in the send queue to probe? */
1702 if (tp
->write_seq
- tp
->snd_nxt
< size_needed
)
1705 if (tp
->snd_wnd
< size_needed
)
1707 if (after(tp
->snd_nxt
+ size_needed
, tcp_wnd_end(tp
)))
1710 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1711 if (tcp_packets_in_flight(tp
) + 2 > tp
->snd_cwnd
) {
1712 if (!tcp_packets_in_flight(tp
))
1718 /* We're allowed to probe. Build it now. */
1719 if ((nskb
= sk_stream_alloc_skb(sk
, probe_size
, GFP_ATOMIC
)) == NULL
)
1721 sk
->sk_wmem_queued
+= nskb
->truesize
;
1722 sk_mem_charge(sk
, nskb
->truesize
);
1724 skb
= tcp_send_head(sk
);
1726 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(skb
)->seq
;
1727 TCP_SKB_CB(nskb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ probe_size
;
1728 TCP_SKB_CB(nskb
)->tcp_flags
= TCPHDR_ACK
;
1729 TCP_SKB_CB(nskb
)->sacked
= 0;
1731 nskb
->ip_summed
= skb
->ip_summed
;
1733 tcp_insert_write_queue_before(nskb
, skb
, sk
);
1736 tcp_for_write_queue_from_safe(skb
, next
, sk
) {
1737 copy
= min_t(int, skb
->len
, probe_size
- len
);
1738 if (nskb
->ip_summed
)
1739 skb_copy_bits(skb
, 0, skb_put(nskb
, copy
), copy
);
1741 nskb
->csum
= skb_copy_and_csum_bits(skb
, 0,
1742 skb_put(nskb
, copy
),
1745 if (skb
->len
<= copy
) {
1746 /* We've eaten all the data from this skb.
1748 TCP_SKB_CB(nskb
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
;
1749 tcp_unlink_write_queue(skb
, sk
);
1750 sk_wmem_free_skb(sk
, skb
);
1752 TCP_SKB_CB(nskb
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
&
1753 ~(TCPHDR_FIN
|TCPHDR_PSH
);
1754 if (!skb_shinfo(skb
)->nr_frags
) {
1755 skb_pull(skb
, copy
);
1756 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1757 skb
->csum
= csum_partial(skb
->data
,
1760 __pskb_trim_head(skb
, copy
);
1761 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1763 TCP_SKB_CB(skb
)->seq
+= copy
;
1768 if (len
>= probe_size
)
1771 tcp_init_tso_segs(sk
, nskb
, nskb
->len
);
1773 /* We're ready to send. If this fails, the probe will
1774 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1775 TCP_SKB_CB(nskb
)->when
= tcp_time_stamp
;
1776 if (!tcp_transmit_skb(sk
, nskb
, 1, GFP_ATOMIC
)) {
1777 /* Decrement cwnd here because we are sending
1778 * effectively two packets. */
1780 tcp_event_new_data_sent(sk
, nskb
);
1782 icsk
->icsk_mtup
.probe_size
= tcp_mss_to_mtu(sk
, nskb
->len
);
1783 tp
->mtu_probe
.probe_seq_start
= TCP_SKB_CB(nskb
)->seq
;
1784 tp
->mtu_probe
.probe_seq_end
= TCP_SKB_CB(nskb
)->end_seq
;
1792 /* This routine writes packets to the network. It advances the
1793 * send_head. This happens as incoming acks open up the remote
1796 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1797 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1798 * account rare use of URG, this is not a big flaw.
1800 * Send at most one packet when push_one > 0. Temporarily ignore
1801 * cwnd limit to force at most one packet out when push_one == 2.
1803 * Returns true, if no segments are in flight and we have queued segments,
1804 * but cannot send anything now because of SWS or another problem.
1806 static bool tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
,
1807 int push_one
, gfp_t gfp
)
1809 struct tcp_sock
*tp
= tcp_sk(sk
);
1810 struct sk_buff
*skb
;
1811 unsigned int tso_segs
, sent_pkts
;
1818 /* Do MTU probing. */
1819 result
= tcp_mtu_probe(sk
);
1822 } else if (result
> 0) {
1827 while ((skb
= tcp_send_head(sk
))) {
1831 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1834 if (unlikely(tp
->repair
) && tp
->repair_queue
== TCP_SEND_QUEUE
)
1835 goto repair
; /* Skip network transmission */
1837 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1840 /* Force out a loss probe pkt. */
1846 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
)))
1849 if (tso_segs
== 1) {
1850 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
1851 (tcp_skb_is_last(sk
, skb
) ?
1852 nonagle
: TCP_NAGLE_PUSH
))))
1855 if (!push_one
&& tcp_tso_should_defer(sk
, skb
))
1859 /* TSQ : sk_wmem_alloc accounts skb truesize,
1860 * including skb overhead. But thats OK.
1862 if (atomic_read(&sk
->sk_wmem_alloc
) >= sysctl_tcp_limit_output_bytes
) {
1863 set_bit(TSQ_THROTTLED
, &tp
->tsq_flags
);
1867 if (tso_segs
> 1 && !tcp_urg_mode(tp
))
1868 limit
= tcp_mss_split_point(sk
, skb
, mss_now
,
1871 sk
->sk_gso_max_segs
));
1873 if (skb
->len
> limit
&&
1874 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
, gfp
)))
1877 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1879 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, gfp
)))
1883 /* Advance the send_head. This one is sent out.
1884 * This call will increment packets_out.
1886 tcp_event_new_data_sent(sk
, skb
);
1888 tcp_minshall_update(tp
, mss_now
, skb
);
1889 sent_pkts
+= tcp_skb_pcount(skb
);
1895 if (likely(sent_pkts
)) {
1896 if (tcp_in_cwnd_reduction(sk
))
1897 tp
->prr_out
+= sent_pkts
;
1899 /* Send one loss probe per tail loss episode. */
1901 tcp_schedule_loss_probe(sk
);
1902 tcp_cwnd_validate(sk
);
1905 return (push_one
== 2) || (!tp
->packets_out
&& tcp_send_head(sk
));
1908 bool tcp_schedule_loss_probe(struct sock
*sk
)
1910 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1911 struct tcp_sock
*tp
= tcp_sk(sk
);
1912 u32 timeout
, tlp_time_stamp
, rto_time_stamp
;
1913 u32 rtt
= tp
->srtt
>> 3;
1915 if (WARN_ON(icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
))
1917 /* No consecutive loss probes. */
1918 if (WARN_ON(icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
)) {
1922 /* Don't do any loss probe on a Fast Open connection before 3WHS
1925 if (sk
->sk_state
== TCP_SYN_RECV
)
1928 /* TLP is only scheduled when next timer event is RTO. */
1929 if (icsk
->icsk_pending
!= ICSK_TIME_RETRANS
)
1932 /* Schedule a loss probe in 2*RTT for SACK capable connections
1933 * in Open state, that are either limited by cwnd or application.
1935 if (sysctl_tcp_early_retrans
< 3 || !rtt
|| !tp
->packets_out
||
1936 !tcp_is_sack(tp
) || inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
)
1939 if ((tp
->snd_cwnd
> tcp_packets_in_flight(tp
)) &&
1943 /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
1944 * for delayed ack when there's one outstanding packet.
1947 if (tp
->packets_out
== 1)
1948 timeout
= max_t(u32
, timeout
,
1949 (rtt
+ (rtt
>> 1) + TCP_DELACK_MAX
));
1950 timeout
= max_t(u32
, timeout
, msecs_to_jiffies(10));
1952 /* If RTO is shorter, just schedule TLP in its place. */
1953 tlp_time_stamp
= tcp_time_stamp
+ timeout
;
1954 rto_time_stamp
= (u32
)inet_csk(sk
)->icsk_timeout
;
1955 if ((s32
)(tlp_time_stamp
- rto_time_stamp
) > 0) {
1956 s32 delta
= rto_time_stamp
- tcp_time_stamp
;
1961 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_LOSS_PROBE
, timeout
,
1966 /* When probe timeout (PTO) fires, send a new segment if one exists, else
1967 * retransmit the last segment.
1969 void tcp_send_loss_probe(struct sock
*sk
)
1971 struct tcp_sock
*tp
= tcp_sk(sk
);
1972 struct sk_buff
*skb
;
1974 int mss
= tcp_current_mss(sk
);
1977 if (tcp_send_head(sk
) != NULL
) {
1978 err
= tcp_write_xmit(sk
, mss
, TCP_NAGLE_OFF
, 2, GFP_ATOMIC
);
1982 /* At most one outstanding TLP retransmission. */
1983 if (tp
->tlp_high_seq
)
1986 /* Retransmit last segment. */
1987 skb
= tcp_write_queue_tail(sk
);
1991 pcount
= tcp_skb_pcount(skb
);
1992 if (WARN_ON(!pcount
))
1995 if ((pcount
> 1) && (skb
->len
> (pcount
- 1) * mss
)) {
1996 if (unlikely(tcp_fragment(sk
, skb
, (pcount
- 1) * mss
, mss
)))
1998 skb
= tcp_write_queue_tail(sk
);
2001 if (WARN_ON(!skb
|| !tcp_skb_pcount(skb
)))
2004 /* Probe with zero data doesn't trigger fast recovery. */
2006 err
= __tcp_retransmit_skb(sk
, skb
);
2008 /* Record snd_nxt for loss detection. */
2010 tp
->tlp_high_seq
= tp
->snd_nxt
;
2013 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2014 inet_csk(sk
)->icsk_rto
,
2018 NET_INC_STATS_BH(sock_net(sk
),
2019 LINUX_MIB_TCPLOSSPROBES
);
2023 /* Push out any pending frames which were held back due to
2024 * TCP_CORK or attempt at coalescing tiny packets.
2025 * The socket must be locked by the caller.
2027 void __tcp_push_pending_frames(struct sock
*sk
, unsigned int cur_mss
,
2030 /* If we are closed, the bytes will have to remain here.
2031 * In time closedown will finish, we empty the write queue and
2032 * all will be happy.
2034 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
2037 if (tcp_write_xmit(sk
, cur_mss
, nonagle
, 0,
2038 sk_gfp_atomic(sk
, GFP_ATOMIC
)))
2039 tcp_check_probe_timer(sk
);
2042 /* Send _single_ skb sitting at the send head. This function requires
2043 * true push pending frames to setup probe timer etc.
2045 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
2047 struct sk_buff
*skb
= tcp_send_head(sk
);
2049 BUG_ON(!skb
|| skb
->len
< mss_now
);
2051 tcp_write_xmit(sk
, mss_now
, TCP_NAGLE_PUSH
, 1, sk
->sk_allocation
);
2054 /* This function returns the amount that we can raise the
2055 * usable window based on the following constraints
2057 * 1. The window can never be shrunk once it is offered (RFC 793)
2058 * 2. We limit memory per socket
2061 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2062 * RECV.NEXT + RCV.WIN fixed until:
2063 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2065 * i.e. don't raise the right edge of the window until you can raise
2066 * it at least MSS bytes.
2068 * Unfortunately, the recommended algorithm breaks header prediction,
2069 * since header prediction assumes th->window stays fixed.
2071 * Strictly speaking, keeping th->window fixed violates the receiver
2072 * side SWS prevention criteria. The problem is that under this rule
2073 * a stream of single byte packets will cause the right side of the
2074 * window to always advance by a single byte.
2076 * Of course, if the sender implements sender side SWS prevention
2077 * then this will not be a problem.
2079 * BSD seems to make the following compromise:
2081 * If the free space is less than the 1/4 of the maximum
2082 * space available and the free space is less than 1/2 mss,
2083 * then set the window to 0.
2084 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2085 * Otherwise, just prevent the window from shrinking
2086 * and from being larger than the largest representable value.
2088 * This prevents incremental opening of the window in the regime
2089 * where TCP is limited by the speed of the reader side taking
2090 * data out of the TCP receive queue. It does nothing about
2091 * those cases where the window is constrained on the sender side
2092 * because the pipeline is full.
2094 * BSD also seems to "accidentally" limit itself to windows that are a
2095 * multiple of MSS, at least until the free space gets quite small.
2096 * This would appear to be a side effect of the mbuf implementation.
2097 * Combining these two algorithms results in the observed behavior
2098 * of having a fixed window size at almost all times.
2100 * Below we obtain similar behavior by forcing the offered window to
2101 * a multiple of the mss when it is feasible to do so.
2103 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2104 * Regular options like TIMESTAMP are taken into account.
2106 u32
__tcp_select_window(struct sock
*sk
)
2108 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2109 struct tcp_sock
*tp
= tcp_sk(sk
);
2110 /* MSS for the peer's data. Previous versions used mss_clamp
2111 * here. I don't know if the value based on our guesses
2112 * of peer's MSS is better for the performance. It's more correct
2113 * but may be worse for the performance because of rcv_mss
2114 * fluctuations. --SAW 1998/11/1
2116 int mss
= icsk
->icsk_ack
.rcv_mss
;
2117 int free_space
= tcp_space(sk
);
2118 int full_space
= min_t(int, tp
->window_clamp
, tcp_full_space(sk
));
2121 if (mss
> full_space
)
2124 if (free_space
< (full_space
>> 1)) {
2125 icsk
->icsk_ack
.quick
= 0;
2127 if (sk_under_memory_pressure(sk
))
2128 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
,
2131 if (free_space
< mss
)
2135 if (free_space
> tp
->rcv_ssthresh
)
2136 free_space
= tp
->rcv_ssthresh
;
2138 /* Don't do rounding if we are using window scaling, since the
2139 * scaled window will not line up with the MSS boundary anyway.
2141 window
= tp
->rcv_wnd
;
2142 if (tp
->rx_opt
.rcv_wscale
) {
2143 window
= free_space
;
2145 /* Advertise enough space so that it won't get scaled away.
2146 * Import case: prevent zero window announcement if
2147 * 1<<rcv_wscale > mss.
2149 if (((window
>> tp
->rx_opt
.rcv_wscale
) << tp
->rx_opt
.rcv_wscale
) != window
)
2150 window
= (((window
>> tp
->rx_opt
.rcv_wscale
) + 1)
2151 << tp
->rx_opt
.rcv_wscale
);
2153 /* Get the largest window that is a nice multiple of mss.
2154 * Window clamp already applied above.
2155 * If our current window offering is within 1 mss of the
2156 * free space we just keep it. This prevents the divide
2157 * and multiply from happening most of the time.
2158 * We also don't do any window rounding when the free space
2161 if (window
<= free_space
- mss
|| window
> free_space
)
2162 window
= (free_space
/ mss
) * mss
;
2163 else if (mss
== full_space
&&
2164 free_space
> window
+ (full_space
>> 1))
2165 window
= free_space
;
2171 /* Collapses two adjacent SKB's during retransmission. */
2172 static void tcp_collapse_retrans(struct sock
*sk
, struct sk_buff
*skb
)
2174 struct tcp_sock
*tp
= tcp_sk(sk
);
2175 struct sk_buff
*next_skb
= tcp_write_queue_next(sk
, skb
);
2176 int skb_size
, next_skb_size
;
2178 skb_size
= skb
->len
;
2179 next_skb_size
= next_skb
->len
;
2181 BUG_ON(tcp_skb_pcount(skb
) != 1 || tcp_skb_pcount(next_skb
) != 1);
2183 tcp_highest_sack_combine(sk
, next_skb
, skb
);
2185 tcp_unlink_write_queue(next_skb
, sk
);
2187 skb_copy_from_linear_data(next_skb
, skb_put(skb
, next_skb_size
),
2190 if (next_skb
->ip_summed
== CHECKSUM_PARTIAL
)
2191 skb
->ip_summed
= CHECKSUM_PARTIAL
;
2193 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
2194 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
2196 /* Update sequence range on original skb. */
2197 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
2199 /* Merge over control information. This moves PSH/FIN etc. over */
2200 TCP_SKB_CB(skb
)->tcp_flags
|= TCP_SKB_CB(next_skb
)->tcp_flags
;
2202 /* All done, get rid of second SKB and account for it so
2203 * packet counting does not break.
2205 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
& TCPCB_EVER_RETRANS
;
2207 /* changed transmit queue under us so clear hints */
2208 tcp_clear_retrans_hints_partial(tp
);
2209 if (next_skb
== tp
->retransmit_skb_hint
)
2210 tp
->retransmit_skb_hint
= skb
;
2212 tcp_adjust_pcount(sk
, next_skb
, tcp_skb_pcount(next_skb
));
2214 sk_wmem_free_skb(sk
, next_skb
);
2217 /* Check if coalescing SKBs is legal. */
2218 static bool tcp_can_collapse(const struct sock
*sk
, const struct sk_buff
*skb
)
2220 if (tcp_skb_pcount(skb
) > 1)
2222 /* TODO: SACK collapsing could be used to remove this condition */
2223 if (skb_shinfo(skb
)->nr_frags
!= 0)
2225 if (skb_cloned(skb
))
2227 if (skb
== tcp_send_head(sk
))
2229 /* Some heurestics for collapsing over SACK'd could be invented */
2230 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
2236 /* Collapse packets in the retransmit queue to make to create
2237 * less packets on the wire. This is only done on retransmission.
2239 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*to
,
2242 struct tcp_sock
*tp
= tcp_sk(sk
);
2243 struct sk_buff
*skb
= to
, *tmp
;
2246 if (!sysctl_tcp_retrans_collapse
)
2248 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)
2251 tcp_for_write_queue_from_safe(skb
, tmp
, sk
) {
2252 if (!tcp_can_collapse(sk
, skb
))
2264 /* Punt if not enough space exists in the first SKB for
2265 * the data in the second
2267 if (skb
->len
> skb_availroom(to
))
2270 if (after(TCP_SKB_CB(skb
)->end_seq
, tcp_wnd_end(tp
)))
2273 tcp_collapse_retrans(sk
, to
);
2277 /* This retransmits one SKB. Policy decisions and retransmit queue
2278 * state updates are done by the caller. Returns non-zero if an
2279 * error occurred which prevented the send.
2281 int __tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
2283 struct tcp_sock
*tp
= tcp_sk(sk
);
2284 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2285 unsigned int cur_mss
;
2287 /* Inconslusive MTU probe */
2288 if (icsk
->icsk_mtup
.probe_size
) {
2289 icsk
->icsk_mtup
.probe_size
= 0;
2292 /* Do not sent more than we queued. 1/4 is reserved for possible
2293 * copying overhead: fragmentation, tunneling, mangling etc.
2295 if (atomic_read(&sk
->sk_wmem_alloc
) >
2296 min(sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2), sk
->sk_sndbuf
))
2299 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
2300 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
2302 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
2306 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
2307 return -EHOSTUNREACH
; /* Routing failure or similar. */
2309 cur_mss
= tcp_current_mss(sk
);
2311 /* If receiver has shrunk his window, and skb is out of
2312 * new window, do not retransmit it. The exception is the
2313 * case, when window is shrunk to zero. In this case
2314 * our retransmit serves as a zero window probe.
2316 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
)) &&
2317 TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
2320 if (skb
->len
> cur_mss
) {
2321 if (tcp_fragment(sk
, skb
, cur_mss
, cur_mss
))
2322 return -ENOMEM
; /* We'll try again later. */
2324 int oldpcount
= tcp_skb_pcount(skb
);
2326 if (unlikely(oldpcount
> 1)) {
2327 tcp_init_tso_segs(sk
, skb
, cur_mss
);
2328 tcp_adjust_pcount(sk
, skb
, oldpcount
- tcp_skb_pcount(skb
));
2332 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
2334 /* Some Solaris stacks overoptimize and ignore the FIN on a
2335 * retransmit when old data is attached. So strip it off
2336 * since it is cheap to do so and saves bytes on the network.
2339 (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
) &&
2340 tp
->snd_una
== (TCP_SKB_CB(skb
)->end_seq
- 1)) {
2341 if (!pskb_trim(skb
, 0)) {
2342 /* Reuse, even though it does some unnecessary work */
2343 tcp_init_nondata_skb(skb
, TCP_SKB_CB(skb
)->end_seq
- 1,
2344 TCP_SKB_CB(skb
)->tcp_flags
);
2345 skb
->ip_summed
= CHECKSUM_NONE
;
2349 /* Make a copy, if the first transmission SKB clone we made
2350 * is still in somebody's hands, else make a clone.
2352 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2354 /* make sure skb->data is aligned on arches that require it */
2355 if (unlikely(NET_IP_ALIGN
&& ((unsigned long)skb
->data
& 3))) {
2356 struct sk_buff
*nskb
= __pskb_copy(skb
, MAX_TCP_HEADER
,
2358 return nskb
? tcp_transmit_skb(sk
, nskb
, 0, GFP_ATOMIC
) :
2361 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2365 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
2367 struct tcp_sock
*tp
= tcp_sk(sk
);
2368 int err
= __tcp_retransmit_skb(sk
, skb
);
2371 /* Update global TCP statistics. */
2372 TCP_INC_STATS(sock_net(sk
), TCP_MIB_RETRANSSEGS
);
2374 tp
->total_retrans
++;
2376 #if FASTRETRANS_DEBUG > 0
2377 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2378 net_dbg_ratelimited("retrans_out leaked\n");
2381 if (!tp
->retrans_out
)
2382 tp
->lost_retrans_low
= tp
->snd_nxt
;
2383 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
2384 tp
->retrans_out
+= tcp_skb_pcount(skb
);
2386 /* Save stamp of the first retransmit. */
2387 if (!tp
->retrans_stamp
)
2388 tp
->retrans_stamp
= TCP_SKB_CB(skb
)->when
;
2390 tp
->undo_retrans
+= tcp_skb_pcount(skb
);
2392 /* snd_nxt is stored to detect loss of retransmitted segment,
2393 * see tcp_input.c tcp_sacktag_write_queue().
2395 TCP_SKB_CB(skb
)->ack_seq
= tp
->snd_nxt
;
2400 /* Check if we forward retransmits are possible in the current
2401 * window/congestion state.
2403 static bool tcp_can_forward_retransmit(struct sock
*sk
)
2405 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2406 const struct tcp_sock
*tp
= tcp_sk(sk
);
2408 /* Forward retransmissions are possible only during Recovery. */
2409 if (icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
2412 /* No forward retransmissions in Reno are possible. */
2413 if (tcp_is_reno(tp
))
2416 /* Yeah, we have to make difficult choice between forward transmission
2417 * and retransmission... Both ways have their merits...
2419 * For now we do not retransmit anything, while we have some new
2420 * segments to send. In the other cases, follow rule 3 for
2421 * NextSeg() specified in RFC3517.
2424 if (tcp_may_send_now(sk
))
2430 /* This gets called after a retransmit timeout, and the initially
2431 * retransmitted data is acknowledged. It tries to continue
2432 * resending the rest of the retransmit queue, until either
2433 * we've sent it all or the congestion window limit is reached.
2434 * If doing SACK, the first ACK which comes back for a timeout
2435 * based retransmit packet might feed us FACK information again.
2436 * If so, we use it to avoid unnecessarily retransmissions.
2438 void tcp_xmit_retransmit_queue(struct sock
*sk
)
2440 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2441 struct tcp_sock
*tp
= tcp_sk(sk
);
2442 struct sk_buff
*skb
;
2443 struct sk_buff
*hole
= NULL
;
2446 int fwd_rexmitting
= 0;
2448 if (!tp
->packets_out
)
2452 tp
->retransmit_high
= tp
->snd_una
;
2454 if (tp
->retransmit_skb_hint
) {
2455 skb
= tp
->retransmit_skb_hint
;
2456 last_lost
= TCP_SKB_CB(skb
)->end_seq
;
2457 if (after(last_lost
, tp
->retransmit_high
))
2458 last_lost
= tp
->retransmit_high
;
2460 skb
= tcp_write_queue_head(sk
);
2461 last_lost
= tp
->snd_una
;
2464 tcp_for_write_queue_from(skb
, sk
) {
2465 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
2467 if (skb
== tcp_send_head(sk
))
2469 /* we could do better than to assign each time */
2471 tp
->retransmit_skb_hint
= skb
;
2473 /* Assume this retransmit will generate
2474 * only one packet for congestion window
2475 * calculation purposes. This works because
2476 * tcp_retransmit_skb() will chop up the
2477 * packet to be MSS sized and all the
2478 * packet counting works out.
2480 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
2483 if (fwd_rexmitting
) {
2485 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_highest_sack_seq(tp
)))
2487 mib_idx
= LINUX_MIB_TCPFORWARDRETRANS
;
2489 } else if (!before(TCP_SKB_CB(skb
)->seq
, tp
->retransmit_high
)) {
2490 tp
->retransmit_high
= last_lost
;
2491 if (!tcp_can_forward_retransmit(sk
))
2493 /* Backtrack if necessary to non-L'ed skb */
2501 } else if (!(sacked
& TCPCB_LOST
)) {
2502 if (hole
== NULL
&& !(sacked
& (TCPCB_SACKED_RETRANS
|TCPCB_SACKED_ACKED
)))
2507 last_lost
= TCP_SKB_CB(skb
)->end_seq
;
2508 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
2509 mib_idx
= LINUX_MIB_TCPFASTRETRANS
;
2511 mib_idx
= LINUX_MIB_TCPSLOWSTARTRETRANS
;
2514 if (sacked
& (TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))
2517 if (tcp_retransmit_skb(sk
, skb
)) {
2518 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRETRANSFAIL
);
2521 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2523 if (tcp_in_cwnd_reduction(sk
))
2524 tp
->prr_out
+= tcp_skb_pcount(skb
);
2526 if (skb
== tcp_write_queue_head(sk
))
2527 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2528 inet_csk(sk
)->icsk_rto
,
2533 /* Send a fin. The caller locks the socket for us. This cannot be
2534 * allowed to fail queueing a FIN frame under any circumstances.
2536 void tcp_send_fin(struct sock
*sk
)
2538 struct tcp_sock
*tp
= tcp_sk(sk
);
2539 struct sk_buff
*skb
= tcp_write_queue_tail(sk
);
2542 /* Optimization, tack on the FIN if we have a queue of
2543 * unsent frames. But be careful about outgoing SACKS
2546 mss_now
= tcp_current_mss(sk
);
2548 if (tcp_send_head(sk
) != NULL
) {
2549 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_FIN
;
2550 TCP_SKB_CB(skb
)->end_seq
++;
2553 /* Socket is locked, keep trying until memory is available. */
2555 skb
= alloc_skb_fclone(MAX_TCP_HEADER
,
2562 /* Reserve space for headers and prepare control bits. */
2563 skb_reserve(skb
, MAX_TCP_HEADER
);
2564 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2565 tcp_init_nondata_skb(skb
, tp
->write_seq
,
2566 TCPHDR_ACK
| TCPHDR_FIN
);
2567 tcp_queue_skb(sk
, skb
);
2569 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_OFF
);
2572 /* We get here when a process closes a file descriptor (either due to
2573 * an explicit close() or as a byproduct of exit()'ing) and there
2574 * was unread data in the receive queue. This behavior is recommended
2575 * by RFC 2525, section 2.17. -DaveM
2577 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
2579 struct sk_buff
*skb
;
2581 /* NOTE: No TCP options attached and we never retransmit this. */
2582 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
2584 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
2588 /* Reserve space for headers and prepare control bits. */
2589 skb_reserve(skb
, MAX_TCP_HEADER
);
2590 tcp_init_nondata_skb(skb
, tcp_acceptable_seq(sk
),
2591 TCPHDR_ACK
| TCPHDR_RST
);
2593 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2594 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
2595 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
2597 TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTRSTS
);
2600 /* Send a crossed SYN-ACK during socket establishment.
2601 * WARNING: This routine must only be called when we have already sent
2602 * a SYN packet that crossed the incoming SYN that caused this routine
2603 * to get called. If this assumption fails then the initial rcv_wnd
2604 * and rcv_wscale values will not be correct.
2606 int tcp_send_synack(struct sock
*sk
)
2608 struct sk_buff
*skb
;
2610 skb
= tcp_write_queue_head(sk
);
2611 if (skb
== NULL
|| !(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)) {
2612 pr_debug("%s: wrong queue state\n", __func__
);
2615 if (!(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_ACK
)) {
2616 if (skb_cloned(skb
)) {
2617 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
2620 tcp_unlink_write_queue(skb
, sk
);
2621 skb_header_release(nskb
);
2622 __tcp_add_write_queue_head(sk
, nskb
);
2623 sk_wmem_free_skb(sk
, skb
);
2624 sk
->sk_wmem_queued
+= nskb
->truesize
;
2625 sk_mem_charge(sk
, nskb
->truesize
);
2629 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_ACK
;
2630 TCP_ECN_send_synack(tcp_sk(sk
), skb
);
2632 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2633 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2637 * tcp_make_synack - Prepare a SYN-ACK.
2638 * sk: listener socket
2639 * dst: dst entry attached to the SYNACK
2640 * req: request_sock pointer
2642 * Allocate one skb and build a SYNACK packet.
2643 * @dst is consumed : Caller should not use it again.
2645 struct sk_buff
*tcp_make_synack(struct sock
*sk
, struct dst_entry
*dst
,
2646 struct request_sock
*req
,
2647 struct tcp_fastopen_cookie
*foc
)
2649 struct tcp_out_options opts
;
2650 struct inet_request_sock
*ireq
= inet_rsk(req
);
2651 struct tcp_sock
*tp
= tcp_sk(sk
);
2653 struct sk_buff
*skb
;
2654 struct tcp_md5sig_key
*md5
;
2655 int tcp_header_size
;
2658 skb
= alloc_skb(MAX_TCP_HEADER
+ 15, sk_gfp_atomic(sk
, GFP_ATOMIC
));
2659 if (unlikely(!skb
)) {
2663 /* Reserve space for headers. */
2664 skb_reserve(skb
, MAX_TCP_HEADER
);
2666 skb_dst_set(skb
, dst
);
2668 mss
= dst_metric_advmss(dst
);
2669 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< mss
)
2670 mss
= tp
->rx_opt
.user_mss
;
2672 if (req
->rcv_wnd
== 0) { /* ignored for retransmitted syns */
2674 /* Set this up on the first call only */
2675 req
->window_clamp
= tp
->window_clamp
? : dst_metric(dst
, RTAX_WINDOW
);
2677 /* limit the window selection if the user enforce a smaller rx buffer */
2678 if (sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
&&
2679 (req
->window_clamp
> tcp_full_space(sk
) || req
->window_clamp
== 0))
2680 req
->window_clamp
= tcp_full_space(sk
);
2682 /* tcp_full_space because it is guaranteed to be the first packet */
2683 tcp_select_initial_window(tcp_full_space(sk
),
2684 mss
- (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0),
2689 dst_metric(dst
, RTAX_INITRWND
));
2690 ireq
->rcv_wscale
= rcv_wscale
;
2693 memset(&opts
, 0, sizeof(opts
));
2694 #ifdef CONFIG_SYN_COOKIES
2695 if (unlikely(req
->cookie_ts
))
2696 TCP_SKB_CB(skb
)->when
= cookie_init_timestamp(req
);
2699 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2700 tcp_header_size
= tcp_synack_options(sk
, req
, mss
, skb
, &opts
, &md5
,
2703 skb_push(skb
, tcp_header_size
);
2704 skb_reset_transport_header(skb
);
2707 memset(th
, 0, sizeof(struct tcphdr
));
2710 TCP_ECN_make_synack(req
, th
);
2711 th
->source
= ireq
->loc_port
;
2712 th
->dest
= ireq
->rmt_port
;
2713 /* Setting of flags are superfluous here for callers (and ECE is
2714 * not even correctly set)
2716 tcp_init_nondata_skb(skb
, tcp_rsk(req
)->snt_isn
,
2717 TCPHDR_SYN
| TCPHDR_ACK
);
2719 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
2720 /* XXX data is queued and acked as is. No buffer/window check */
2721 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_nxt
);
2723 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2724 th
->window
= htons(min(req
->rcv_wnd
, 65535U));
2725 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
);
2726 th
->doff
= (tcp_header_size
>> 2);
2727 TCP_ADD_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
, tcp_skb_pcount(skb
));
2729 #ifdef CONFIG_TCP_MD5SIG
2730 /* Okay, we have all we need - do the md5 hash if needed */
2732 tcp_rsk(req
)->af_specific
->calc_md5_hash(opts
.hash_location
,
2733 md5
, NULL
, req
, skb
);
2739 EXPORT_SYMBOL(tcp_make_synack
);
2741 /* Do all connect socket setups that can be done AF independent. */
2742 void tcp_connect_init(struct sock
*sk
)
2744 const struct dst_entry
*dst
= __sk_dst_get(sk
);
2745 struct tcp_sock
*tp
= tcp_sk(sk
);
2748 /* We'll fix this up when we get a response from the other end.
2749 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2751 tp
->tcp_header_len
= sizeof(struct tcphdr
) +
2752 (sysctl_tcp_timestamps
? TCPOLEN_TSTAMP_ALIGNED
: 0);
2754 #ifdef CONFIG_TCP_MD5SIG
2755 if (tp
->af_specific
->md5_lookup(sk
, sk
) != NULL
)
2756 tp
->tcp_header_len
+= TCPOLEN_MD5SIG_ALIGNED
;
2759 /* If user gave his TCP_MAXSEG, record it to clamp */
2760 if (tp
->rx_opt
.user_mss
)
2761 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
2764 tcp_sync_mss(sk
, dst_mtu(dst
));
2766 if (!tp
->window_clamp
)
2767 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
2768 tp
->advmss
= dst_metric_advmss(dst
);
2769 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< tp
->advmss
)
2770 tp
->advmss
= tp
->rx_opt
.user_mss
;
2772 tcp_initialize_rcv_mss(sk
);
2774 /* limit the window selection if the user enforce a smaller rx buffer */
2775 if (sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
&&
2776 (tp
->window_clamp
> tcp_full_space(sk
) || tp
->window_clamp
== 0))
2777 tp
->window_clamp
= tcp_full_space(sk
);
2779 tcp_select_initial_window(tcp_full_space(sk
),
2780 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
2783 sysctl_tcp_window_scaling
,
2785 dst_metric(dst
, RTAX_INITRWND
));
2787 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
2788 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
2791 sock_reset_flag(sk
, SOCK_DONE
);
2794 tp
->snd_una
= tp
->write_seq
;
2795 tp
->snd_sml
= tp
->write_seq
;
2796 tp
->snd_up
= tp
->write_seq
;
2797 tp
->snd_nxt
= tp
->write_seq
;
2799 if (likely(!tp
->repair
))
2801 tp
->rcv_wup
= tp
->rcv_nxt
;
2802 tp
->copied_seq
= tp
->rcv_nxt
;
2804 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
2805 inet_csk(sk
)->icsk_retransmits
= 0;
2806 tcp_clear_retrans(tp
);
2809 static void tcp_connect_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
2811 struct tcp_sock
*tp
= tcp_sk(sk
);
2812 struct tcp_skb_cb
*tcb
= TCP_SKB_CB(skb
);
2814 tcb
->end_seq
+= skb
->len
;
2815 skb_header_release(skb
);
2816 __tcp_add_write_queue_tail(sk
, skb
);
2817 sk
->sk_wmem_queued
+= skb
->truesize
;
2818 sk_mem_charge(sk
, skb
->truesize
);
2819 tp
->write_seq
= tcb
->end_seq
;
2820 tp
->packets_out
+= tcp_skb_pcount(skb
);
2823 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
2824 * queue a data-only packet after the regular SYN, such that regular SYNs
2825 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
2826 * only the SYN sequence, the data are retransmitted in the first ACK.
2827 * If cookie is not cached or other error occurs, falls back to send a
2828 * regular SYN with Fast Open cookie request option.
2830 static int tcp_send_syn_data(struct sock
*sk
, struct sk_buff
*syn
)
2832 struct tcp_sock
*tp
= tcp_sk(sk
);
2833 struct tcp_fastopen_request
*fo
= tp
->fastopen_req
;
2834 int syn_loss
= 0, space
, i
, err
= 0, iovlen
= fo
->data
->msg_iovlen
;
2835 struct sk_buff
*syn_data
= NULL
, *data
;
2836 unsigned long last_syn_loss
= 0;
2838 tp
->rx_opt
.mss_clamp
= tp
->advmss
; /* If MSS is not cached */
2839 tcp_fastopen_cache_get(sk
, &tp
->rx_opt
.mss_clamp
, &fo
->cookie
,
2840 &syn_loss
, &last_syn_loss
);
2841 /* Recurring FO SYN losses: revert to regular handshake temporarily */
2843 time_before(jiffies
, last_syn_loss
+ (60*HZ
<< syn_loss
))) {
2844 fo
->cookie
.len
= -1;
2848 if (sysctl_tcp_fastopen
& TFO_CLIENT_NO_COOKIE
)
2849 fo
->cookie
.len
= -1;
2850 else if (fo
->cookie
.len
<= 0)
2853 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
2854 * user-MSS. Reserve maximum option space for middleboxes that add
2855 * private TCP options. The cost is reduced data space in SYN :(
2857 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< tp
->rx_opt
.mss_clamp
)
2858 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
2859 space
= __tcp_mtu_to_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
) -
2860 MAX_TCP_OPTION_SPACE
;
2862 syn_data
= skb_copy_expand(syn
, skb_headroom(syn
), space
,
2864 if (syn_data
== NULL
)
2867 for (i
= 0; i
< iovlen
&& syn_data
->len
< space
; ++i
) {
2868 struct iovec
*iov
= &fo
->data
->msg_iov
[i
];
2869 unsigned char __user
*from
= iov
->iov_base
;
2870 int len
= iov
->iov_len
;
2872 if (syn_data
->len
+ len
> space
)
2873 len
= space
- syn_data
->len
;
2874 else if (i
+ 1 == iovlen
)
2875 /* No more data pending in inet_wait_for_connect() */
2878 if (skb_add_data(syn_data
, from
, len
))
2882 /* Queue a data-only packet after the regular SYN for retransmission */
2883 data
= pskb_copy(syn_data
, sk
->sk_allocation
);
2886 TCP_SKB_CB(data
)->seq
++;
2887 TCP_SKB_CB(data
)->tcp_flags
&= ~TCPHDR_SYN
;
2888 TCP_SKB_CB(data
)->tcp_flags
= (TCPHDR_ACK
|TCPHDR_PSH
);
2889 tcp_connect_queue_skb(sk
, data
);
2890 fo
->copied
= data
->len
;
2892 if (tcp_transmit_skb(sk
, syn_data
, 0, sk
->sk_allocation
) == 0) {
2893 tp
->syn_data
= (fo
->copied
> 0);
2894 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPFASTOPENACTIVE
);
2900 /* Send a regular SYN with Fast Open cookie request option */
2901 if (fo
->cookie
.len
> 0)
2903 err
= tcp_transmit_skb(sk
, syn
, 1, sk
->sk_allocation
);
2905 tp
->syn_fastopen
= 0;
2906 kfree_skb(syn_data
);
2908 fo
->cookie
.len
= -1; /* Exclude Fast Open option for SYN retries */
2912 /* Build a SYN and send it off. */
2913 int tcp_connect(struct sock
*sk
)
2915 struct tcp_sock
*tp
= tcp_sk(sk
);
2916 struct sk_buff
*buff
;
2919 tcp_connect_init(sk
);
2921 if (unlikely(tp
->repair
)) {
2922 tcp_finish_connect(sk
, NULL
);
2926 buff
= alloc_skb_fclone(MAX_TCP_HEADER
+ 15, sk
->sk_allocation
);
2927 if (unlikely(buff
== NULL
))
2930 /* Reserve space for headers. */
2931 skb_reserve(buff
, MAX_TCP_HEADER
);
2933 tcp_init_nondata_skb(buff
, tp
->write_seq
++, TCPHDR_SYN
);
2934 tp
->retrans_stamp
= TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2935 tcp_connect_queue_skb(sk
, buff
);
2936 TCP_ECN_send_syn(sk
, buff
);
2938 /* Send off SYN; include data in Fast Open. */
2939 err
= tp
->fastopen_req
? tcp_send_syn_data(sk
, buff
) :
2940 tcp_transmit_skb(sk
, buff
, 1, sk
->sk_allocation
);
2941 if (err
== -ECONNREFUSED
)
2944 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2945 * in order to make this packet get counted in tcpOutSegs.
2947 tp
->snd_nxt
= tp
->write_seq
;
2948 tp
->pushed_seq
= tp
->write_seq
;
2949 TCP_INC_STATS(sock_net(sk
), TCP_MIB_ACTIVEOPENS
);
2951 /* Timer for repeating the SYN until an answer. */
2952 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2953 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2956 EXPORT_SYMBOL(tcp_connect
);
2958 /* Send out a delayed ack, the caller does the policy checking
2959 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2962 void tcp_send_delayed_ack(struct sock
*sk
)
2964 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2965 int ato
= icsk
->icsk_ack
.ato
;
2966 unsigned long timeout
;
2968 if (ato
> TCP_DELACK_MIN
) {
2969 const struct tcp_sock
*tp
= tcp_sk(sk
);
2970 int max_ato
= HZ
/ 2;
2972 if (icsk
->icsk_ack
.pingpong
||
2973 (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
2974 max_ato
= TCP_DELACK_MAX
;
2976 /* Slow path, intersegment interval is "high". */
2978 /* If some rtt estimate is known, use it to bound delayed ack.
2979 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2983 int rtt
= max(tp
->srtt
>> 3, TCP_DELACK_MIN
);
2989 ato
= min(ato
, max_ato
);
2992 /* Stay within the limit we were given */
2993 timeout
= jiffies
+ ato
;
2995 /* Use new timeout only if there wasn't a older one earlier. */
2996 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
2997 /* If delack timer was blocked or is about to expire,
3000 if (icsk
->icsk_ack
.blocked
||
3001 time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
3006 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
3007 timeout
= icsk
->icsk_ack
.timeout
;
3009 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
3010 icsk
->icsk_ack
.timeout
= timeout
;
3011 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
3014 /* This routine sends an ack and also updates the window. */
3015 void tcp_send_ack(struct sock
*sk
)
3017 struct sk_buff
*buff
;
3019 /* If we have been reset, we may not send again. */
3020 if (sk
->sk_state
== TCP_CLOSE
)
3023 /* We are not putting this on the write queue, so
3024 * tcp_transmit_skb() will set the ownership to this
3027 buff
= alloc_skb(MAX_TCP_HEADER
, sk_gfp_atomic(sk
, GFP_ATOMIC
));
3029 inet_csk_schedule_ack(sk
);
3030 inet_csk(sk
)->icsk_ack
.ato
= TCP_ATO_MIN
;
3031 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
3032 TCP_DELACK_MAX
, TCP_RTO_MAX
);
3036 /* Reserve space for headers and prepare control bits. */
3037 skb_reserve(buff
, MAX_TCP_HEADER
);
3038 tcp_init_nondata_skb(buff
, tcp_acceptable_seq(sk
), TCPHDR_ACK
);
3040 /* Send it off, this clears delayed acks for us. */
3041 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
3042 tcp_transmit_skb(sk
, buff
, 0, sk_gfp_atomic(sk
, GFP_ATOMIC
));
3045 /* This routine sends a packet with an out of date sequence
3046 * number. It assumes the other end will try to ack it.
3048 * Question: what should we make while urgent mode?
3049 * 4.4BSD forces sending single byte of data. We cannot send
3050 * out of window data, because we have SND.NXT==SND.MAX...
3052 * Current solution: to send TWO zero-length segments in urgent mode:
3053 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3054 * out-of-date with SND.UNA-1 to probe window.
3056 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
)
3058 struct tcp_sock
*tp
= tcp_sk(sk
);
3059 struct sk_buff
*skb
;
3061 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3062 skb
= alloc_skb(MAX_TCP_HEADER
, sk_gfp_atomic(sk
, GFP_ATOMIC
));
3066 /* Reserve space for headers and set control bits. */
3067 skb_reserve(skb
, MAX_TCP_HEADER
);
3068 /* Use a previous sequence. This should cause the other
3069 * end to send an ack. Don't queue or clone SKB, just
3072 tcp_init_nondata_skb(skb
, tp
->snd_una
- !urgent
, TCPHDR_ACK
);
3073 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
3074 return tcp_transmit_skb(sk
, skb
, 0, GFP_ATOMIC
);
3077 void tcp_send_window_probe(struct sock
*sk
)
3079 if (sk
->sk_state
== TCP_ESTABLISHED
) {
3080 tcp_sk(sk
)->snd_wl1
= tcp_sk(sk
)->rcv_nxt
- 1;
3081 tcp_sk(sk
)->snd_nxt
= tcp_sk(sk
)->write_seq
;
3082 tcp_xmit_probe_skb(sk
, 0);
3086 /* Initiate keepalive or window probe from timer. */
3087 int tcp_write_wakeup(struct sock
*sk
)
3089 struct tcp_sock
*tp
= tcp_sk(sk
);
3090 struct sk_buff
*skb
;
3092 if (sk
->sk_state
== TCP_CLOSE
)
3095 if ((skb
= tcp_send_head(sk
)) != NULL
&&
3096 before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
))) {
3098 unsigned int mss
= tcp_current_mss(sk
);
3099 unsigned int seg_size
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
3101 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
3102 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
3104 /* We are probing the opening of a window
3105 * but the window size is != 0
3106 * must have been a result SWS avoidance ( sender )
3108 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
3110 seg_size
= min(seg_size
, mss
);
3111 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
3112 if (tcp_fragment(sk
, skb
, seg_size
, mss
))
3114 } else if (!tcp_skb_pcount(skb
))
3115 tcp_set_skb_tso_segs(sk
, skb
, mss
);
3117 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
3118 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
3119 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
3121 tcp_event_new_data_sent(sk
, skb
);
3124 if (between(tp
->snd_up
, tp
->snd_una
+ 1, tp
->snd_una
+ 0xFFFF))
3125 tcp_xmit_probe_skb(sk
, 1);
3126 return tcp_xmit_probe_skb(sk
, 0);
3130 /* A window probe timeout has occurred. If window is not closed send
3131 * a partial packet else a zero probe.
3133 void tcp_send_probe0(struct sock
*sk
)
3135 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3136 struct tcp_sock
*tp
= tcp_sk(sk
);
3139 err
= tcp_write_wakeup(sk
);
3141 if (tp
->packets_out
|| !tcp_send_head(sk
)) {
3142 /* Cancel probe timer, if it is not required. */
3143 icsk
->icsk_probes_out
= 0;
3144 icsk
->icsk_backoff
= 0;
3149 if (icsk
->icsk_backoff
< sysctl_tcp_retries2
)
3150 icsk
->icsk_backoff
++;
3151 icsk
->icsk_probes_out
++;
3152 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3153 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
3156 /* If packet was not sent due to local congestion,
3157 * do not backoff and do not remember icsk_probes_out.
3158 * Let local senders to fight for local resources.
3160 * Use accumulated backoff yet.
3162 if (!icsk
->icsk_probes_out
)
3163 icsk
->icsk_probes_out
= 1;
3164 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3165 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
,
3166 TCP_RESOURCE_PROBE_INTERVAL
),