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 unsigned int sysctl_tcp_notsent_lowat __read_mostly
= UINT_MAX
;
69 EXPORT_SYMBOL(sysctl_tcp_notsent_lowat
);
71 static bool tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
,
72 int push_one
, gfp_t gfp
);
74 /* Account for new data that has been sent to the network. */
75 static void tcp_event_new_data_sent(struct sock
*sk
, const struct sk_buff
*skb
)
77 struct inet_connection_sock
*icsk
= inet_csk(sk
);
78 struct tcp_sock
*tp
= tcp_sk(sk
);
79 unsigned int prior_packets
= tp
->packets_out
;
81 tcp_advance_send_head(sk
, skb
);
82 tp
->snd_nxt
= TCP_SKB_CB(skb
)->end_seq
;
84 tp
->packets_out
+= tcp_skb_pcount(skb
);
85 if (!prior_packets
|| icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
||
86 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
;
166 const struct dst_entry
*dst
= __sk_dst_get(sk
);
168 if (sysctl_tcp_slow_start_after_idle
&&
169 (!tp
->packets_out
&& (s32
)(now
- tp
->lsndtime
) > icsk
->icsk_rto
))
170 tcp_cwnd_restart(sk
, __sk_dst_get(sk
));
174 /* If it is a reply for ato after last received
175 * packet, enter pingpong mode.
177 if ((u32
)(now
- icsk
->icsk_ack
.lrcvtime
) < icsk
->icsk_ack
.ato
&&
178 (!dst
|| !dst_metric(dst
, RTAX_QUICKACK
)))
179 icsk
->icsk_ack
.pingpong
= 1;
182 /* Account for an ACK we sent. */
183 static inline void tcp_event_ack_sent(struct sock
*sk
, unsigned int pkts
)
185 tcp_dec_quickack_mode(sk
, pkts
);
186 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_DACK
);
190 u32
tcp_default_init_rwnd(u32 mss
)
192 /* Initial receive window should be twice of TCP_INIT_CWND to
193 * enable proper sending of new unsent data during fast recovery
194 * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a
195 * limit when mss is larger than 1460.
197 u32 init_rwnd
= TCP_INIT_CWND
* 2;
200 init_rwnd
= max((1460 * init_rwnd
) / mss
, 2U);
204 /* Determine a window scaling and initial window to offer.
205 * Based on the assumption that the given amount of space
206 * will be offered. Store the results in the tp structure.
207 * NOTE: for smooth operation initial space offering should
208 * be a multiple of mss if possible. We assume here that mss >= 1.
209 * This MUST be enforced by all callers.
211 void tcp_select_initial_window(int __space
, __u32 mss
,
212 __u32
*rcv_wnd
, __u32
*window_clamp
,
213 int wscale_ok
, __u8
*rcv_wscale
,
216 unsigned int space
= (__space
< 0 ? 0 : __space
);
218 /* If no clamp set the clamp to the max possible scaled window */
219 if (*window_clamp
== 0)
220 (*window_clamp
) = (65535 << 14);
221 space
= min(*window_clamp
, space
);
223 /* Quantize space offering to a multiple of mss if possible. */
225 space
= (space
/ mss
) * mss
;
227 /* NOTE: offering an initial window larger than 32767
228 * will break some buggy TCP stacks. If the admin tells us
229 * it is likely we could be speaking with such a buggy stack
230 * we will truncate our initial window offering to 32K-1
231 * unless the remote has sent us a window scaling option,
232 * which we interpret as a sign the remote TCP is not
233 * misinterpreting the window field as a signed quantity.
235 if (sysctl_tcp_workaround_signed_windows
)
236 (*rcv_wnd
) = min(space
, MAX_TCP_WINDOW
);
242 /* Set window scaling on max possible window
243 * See RFC1323 for an explanation of the limit to 14
245 space
= max_t(u32
, sysctl_tcp_rmem
[2], sysctl_rmem_max
);
246 space
= min_t(u32
, space
, *window_clamp
);
247 while (space
> 65535 && (*rcv_wscale
) < 14) {
253 if (mss
> (1 << *rcv_wscale
)) {
254 if (!init_rcv_wnd
) /* Use default unless specified otherwise */
255 init_rcv_wnd
= tcp_default_init_rwnd(mss
);
256 *rcv_wnd
= min(*rcv_wnd
, init_rcv_wnd
* mss
);
259 /* Set the clamp no higher than max representable value */
260 (*window_clamp
) = min(65535U << (*rcv_wscale
), *window_clamp
);
262 EXPORT_SYMBOL(tcp_select_initial_window
);
264 /* Chose a new window to advertise, update state in tcp_sock for the
265 * socket, and return result with RFC1323 scaling applied. The return
266 * value can be stuffed directly into th->window for an outgoing
269 static u16
tcp_select_window(struct sock
*sk
)
271 struct tcp_sock
*tp
= tcp_sk(sk
);
272 u32 cur_win
= tcp_receive_window(tp
);
273 u32 new_win
= __tcp_select_window(sk
);
275 /* Never shrink the offered window */
276 if (new_win
< cur_win
) {
277 /* Danger Will Robinson!
278 * Don't update rcv_wup/rcv_wnd here or else
279 * we will not be able to advertise a zero
280 * window in time. --DaveM
282 * Relax Will Robinson.
284 new_win
= ALIGN(cur_win
, 1 << tp
->rx_opt
.rcv_wscale
);
286 tp
->rcv_wnd
= new_win
;
287 tp
->rcv_wup
= tp
->rcv_nxt
;
289 /* Make sure we do not exceed the maximum possible
292 if (!tp
->rx_opt
.rcv_wscale
&& sysctl_tcp_workaround_signed_windows
)
293 new_win
= min(new_win
, MAX_TCP_WINDOW
);
295 new_win
= min(new_win
, (65535U << tp
->rx_opt
.rcv_wscale
));
297 /* RFC1323 scaling applied */
298 new_win
>>= tp
->rx_opt
.rcv_wscale
;
300 /* If we advertise zero window, disable fast path. */
307 /* Packet ECN state for a SYN-ACK */
308 static inline void TCP_ECN_send_synack(const struct tcp_sock
*tp
, struct sk_buff
*skb
)
310 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_CWR
;
311 if (!(tp
->ecn_flags
& TCP_ECN_OK
))
312 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_ECE
;
315 /* Packet ECN state for a SYN. */
316 static inline void TCP_ECN_send_syn(struct sock
*sk
, struct sk_buff
*skb
)
318 struct tcp_sock
*tp
= tcp_sk(sk
);
321 if (sock_net(sk
)->ipv4
.sysctl_tcp_ecn
== 1) {
322 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_ECE
| TCPHDR_CWR
;
323 tp
->ecn_flags
= TCP_ECN_OK
;
327 static __inline__
void
328 TCP_ECN_make_synack(const struct request_sock
*req
, struct tcphdr
*th
)
330 if (inet_rsk(req
)->ecn_ok
)
334 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
337 static inline void TCP_ECN_send(struct sock
*sk
, struct sk_buff
*skb
,
340 struct tcp_sock
*tp
= tcp_sk(sk
);
342 if (tp
->ecn_flags
& TCP_ECN_OK
) {
343 /* Not-retransmitted data segment: set ECT and inject CWR. */
344 if (skb
->len
!= tcp_header_len
&&
345 !before(TCP_SKB_CB(skb
)->seq
, tp
->snd_nxt
)) {
347 if (tp
->ecn_flags
& TCP_ECN_QUEUE_CWR
) {
348 tp
->ecn_flags
&= ~TCP_ECN_QUEUE_CWR
;
349 tcp_hdr(skb
)->cwr
= 1;
350 skb_shinfo(skb
)->gso_type
|= SKB_GSO_TCP_ECN
;
353 /* ACK or retransmitted segment: clear ECT|CE */
354 INET_ECN_dontxmit(sk
);
356 if (tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)
357 tcp_hdr(skb
)->ece
= 1;
361 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
362 * auto increment end seqno.
364 static void tcp_init_nondata_skb(struct sk_buff
*skb
, u32 seq
, u8 flags
)
366 struct skb_shared_info
*shinfo
= skb_shinfo(skb
);
368 skb
->ip_summed
= CHECKSUM_PARTIAL
;
371 TCP_SKB_CB(skb
)->tcp_flags
= flags
;
372 TCP_SKB_CB(skb
)->sacked
= 0;
374 shinfo
->gso_segs
= 1;
375 shinfo
->gso_size
= 0;
376 shinfo
->gso_type
= 0;
378 TCP_SKB_CB(skb
)->seq
= seq
;
379 if (flags
& (TCPHDR_SYN
| TCPHDR_FIN
))
381 TCP_SKB_CB(skb
)->end_seq
= seq
;
384 static inline bool tcp_urg_mode(const struct tcp_sock
*tp
)
386 return tp
->snd_una
!= tp
->snd_up
;
389 #define OPTION_SACK_ADVERTISE (1 << 0)
390 #define OPTION_TS (1 << 1)
391 #define OPTION_MD5 (1 << 2)
392 #define OPTION_WSCALE (1 << 3)
393 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
395 struct tcp_out_options
{
396 u16 options
; /* bit field of OPTION_* */
397 u16 mss
; /* 0 to disable */
398 u8 ws
; /* window scale, 0 to disable */
399 u8 num_sack_blocks
; /* number of SACK blocks to include */
400 u8 hash_size
; /* bytes in hash_location */
401 __u8
*hash_location
; /* temporary pointer, overloaded */
402 __u32 tsval
, tsecr
; /* need to include OPTION_TS */
403 struct tcp_fastopen_cookie
*fastopen_cookie
; /* Fast open cookie */
406 /* Write previously computed TCP options to the packet.
408 * Beware: Something in the Internet is very sensitive to the ordering of
409 * TCP options, we learned this through the hard way, so be careful here.
410 * Luckily we can at least blame others for their non-compliance but from
411 * inter-operability perspective it seems that we're somewhat stuck with
412 * the ordering which we have been using if we want to keep working with
413 * those broken things (not that it currently hurts anybody as there isn't
414 * particular reason why the ordering would need to be changed).
416 * At least SACK_PERM as the first option is known to lead to a disaster
417 * (but it may well be that other scenarios fail similarly).
419 static void tcp_options_write(__be32
*ptr
, struct tcp_sock
*tp
,
420 struct tcp_out_options
*opts
)
422 u16 options
= opts
->options
; /* mungable copy */
424 if (unlikely(OPTION_MD5
& options
)) {
425 *ptr
++ = htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16) |
426 (TCPOPT_MD5SIG
<< 8) | TCPOLEN_MD5SIG
);
427 /* overload cookie hash location */
428 opts
->hash_location
= (__u8
*)ptr
;
432 if (unlikely(opts
->mss
)) {
433 *ptr
++ = htonl((TCPOPT_MSS
<< 24) |
434 (TCPOLEN_MSS
<< 16) |
438 if (likely(OPTION_TS
& options
)) {
439 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
440 *ptr
++ = htonl((TCPOPT_SACK_PERM
<< 24) |
441 (TCPOLEN_SACK_PERM
<< 16) |
442 (TCPOPT_TIMESTAMP
<< 8) |
444 options
&= ~OPTION_SACK_ADVERTISE
;
446 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
448 (TCPOPT_TIMESTAMP
<< 8) |
451 *ptr
++ = htonl(opts
->tsval
);
452 *ptr
++ = htonl(opts
->tsecr
);
455 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
456 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
458 (TCPOPT_SACK_PERM
<< 8) |
462 if (unlikely(OPTION_WSCALE
& options
)) {
463 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
464 (TCPOPT_WINDOW
<< 16) |
465 (TCPOLEN_WINDOW
<< 8) |
469 if (unlikely(opts
->num_sack_blocks
)) {
470 struct tcp_sack_block
*sp
= tp
->rx_opt
.dsack
?
471 tp
->duplicate_sack
: tp
->selective_acks
;
474 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
477 (TCPOLEN_SACK_BASE
+ (opts
->num_sack_blocks
*
478 TCPOLEN_SACK_PERBLOCK
)));
480 for (this_sack
= 0; this_sack
< opts
->num_sack_blocks
;
482 *ptr
++ = htonl(sp
[this_sack
].start_seq
);
483 *ptr
++ = htonl(sp
[this_sack
].end_seq
);
486 tp
->rx_opt
.dsack
= 0;
489 if (unlikely(OPTION_FAST_OPEN_COOKIE
& options
)) {
490 struct tcp_fastopen_cookie
*foc
= opts
->fastopen_cookie
;
492 *ptr
++ = htonl((TCPOPT_EXP
<< 24) |
493 ((TCPOLEN_EXP_FASTOPEN_BASE
+ foc
->len
) << 16) |
494 TCPOPT_FASTOPEN_MAGIC
);
496 memcpy(ptr
, foc
->val
, foc
->len
);
497 if ((foc
->len
& 3) == 2) {
498 u8
*align
= ((u8
*)ptr
) + foc
->len
;
499 align
[0] = align
[1] = TCPOPT_NOP
;
501 ptr
+= (foc
->len
+ 3) >> 2;
505 /* Compute TCP options for SYN packets. This is not the final
506 * network wire format yet.
508 static unsigned int tcp_syn_options(struct sock
*sk
, struct sk_buff
*skb
,
509 struct tcp_out_options
*opts
,
510 struct tcp_md5sig_key
**md5
)
512 struct tcp_sock
*tp
= tcp_sk(sk
);
513 unsigned int remaining
= MAX_TCP_OPTION_SPACE
;
514 struct tcp_fastopen_request
*fastopen
= tp
->fastopen_req
;
516 #ifdef CONFIG_TCP_MD5SIG
517 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
519 opts
->options
|= OPTION_MD5
;
520 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
526 /* We always get an MSS option. The option bytes which will be seen in
527 * normal data packets should timestamps be used, must be in the MSS
528 * advertised. But we subtract them from tp->mss_cache so that
529 * calculations in tcp_sendmsg are simpler etc. So account for this
530 * fact here if necessary. If we don't do this correctly, as a
531 * receiver we won't recognize data packets as being full sized when we
532 * should, and thus we won't abide by the delayed ACK rules correctly.
533 * SACKs don't matter, we never delay an ACK when we have any of those
535 opts
->mss
= tcp_advertise_mss(sk
);
536 remaining
-= TCPOLEN_MSS_ALIGNED
;
538 if (likely(sysctl_tcp_timestamps
&& *md5
== NULL
)) {
539 opts
->options
|= OPTION_TS
;
540 opts
->tsval
= TCP_SKB_CB(skb
)->when
+ tp
->tsoffset
;
541 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
542 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
544 if (likely(sysctl_tcp_window_scaling
)) {
545 opts
->ws
= tp
->rx_opt
.rcv_wscale
;
546 opts
->options
|= OPTION_WSCALE
;
547 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
549 if (likely(sysctl_tcp_sack
)) {
550 opts
->options
|= OPTION_SACK_ADVERTISE
;
551 if (unlikely(!(OPTION_TS
& opts
->options
)))
552 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
555 if (fastopen
&& fastopen
->cookie
.len
>= 0) {
556 u32 need
= TCPOLEN_EXP_FASTOPEN_BASE
+ fastopen
->cookie
.len
;
557 need
= (need
+ 3) & ~3U; /* Align to 32 bits */
558 if (remaining
>= need
) {
559 opts
->options
|= OPTION_FAST_OPEN_COOKIE
;
560 opts
->fastopen_cookie
= &fastopen
->cookie
;
562 tp
->syn_fastopen
= 1;
566 return MAX_TCP_OPTION_SPACE
- remaining
;
569 /* Set up TCP options for SYN-ACKs. */
570 static unsigned int tcp_synack_options(struct sock
*sk
,
571 struct request_sock
*req
,
572 unsigned int mss
, struct sk_buff
*skb
,
573 struct tcp_out_options
*opts
,
574 struct tcp_md5sig_key
**md5
,
575 struct tcp_fastopen_cookie
*foc
)
577 struct inet_request_sock
*ireq
= inet_rsk(req
);
578 unsigned int remaining
= MAX_TCP_OPTION_SPACE
;
580 #ifdef CONFIG_TCP_MD5SIG
581 *md5
= tcp_rsk(req
)->af_specific
->md5_lookup(sk
, req
);
583 opts
->options
|= OPTION_MD5
;
584 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
586 /* We can't fit any SACK blocks in a packet with MD5 + TS
587 * options. There was discussion about disabling SACK
588 * rather than TS in order to fit in better with old,
589 * buggy kernels, but that was deemed to be unnecessary.
591 ireq
->tstamp_ok
&= !ireq
->sack_ok
;
597 /* We always send an MSS option. */
599 remaining
-= TCPOLEN_MSS_ALIGNED
;
601 if (likely(ireq
->wscale_ok
)) {
602 opts
->ws
= ireq
->rcv_wscale
;
603 opts
->options
|= OPTION_WSCALE
;
604 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
606 if (likely(ireq
->tstamp_ok
)) {
607 opts
->options
|= OPTION_TS
;
608 opts
->tsval
= TCP_SKB_CB(skb
)->when
;
609 opts
->tsecr
= req
->ts_recent
;
610 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
612 if (likely(ireq
->sack_ok
)) {
613 opts
->options
|= OPTION_SACK_ADVERTISE
;
614 if (unlikely(!ireq
->tstamp_ok
))
615 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
618 u32 need
= TCPOLEN_EXP_FASTOPEN_BASE
+ foc
->len
;
619 need
= (need
+ 3) & ~3U; /* Align to 32 bits */
620 if (remaining
>= need
) {
621 opts
->options
|= OPTION_FAST_OPEN_COOKIE
;
622 opts
->fastopen_cookie
= foc
;
627 return MAX_TCP_OPTION_SPACE
- remaining
;
630 /* Compute TCP options for ESTABLISHED sockets. This is not the
631 * final wire format yet.
633 static unsigned int tcp_established_options(struct sock
*sk
, struct sk_buff
*skb
,
634 struct tcp_out_options
*opts
,
635 struct tcp_md5sig_key
**md5
)
637 struct tcp_skb_cb
*tcb
= skb
? TCP_SKB_CB(skb
) : NULL
;
638 struct tcp_sock
*tp
= tcp_sk(sk
);
639 unsigned int size
= 0;
640 unsigned int eff_sacks
;
644 #ifdef CONFIG_TCP_MD5SIG
645 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
646 if (unlikely(*md5
)) {
647 opts
->options
|= OPTION_MD5
;
648 size
+= TCPOLEN_MD5SIG_ALIGNED
;
654 if (likely(tp
->rx_opt
.tstamp_ok
)) {
655 opts
->options
|= OPTION_TS
;
656 opts
->tsval
= tcb
? tcb
->when
+ tp
->tsoffset
: 0;
657 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
658 size
+= TCPOLEN_TSTAMP_ALIGNED
;
661 eff_sacks
= tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
;
662 if (unlikely(eff_sacks
)) {
663 const unsigned int remaining
= MAX_TCP_OPTION_SPACE
- size
;
664 opts
->num_sack_blocks
=
665 min_t(unsigned int, eff_sacks
,
666 (remaining
- TCPOLEN_SACK_BASE_ALIGNED
) /
667 TCPOLEN_SACK_PERBLOCK
);
668 size
+= TCPOLEN_SACK_BASE_ALIGNED
+
669 opts
->num_sack_blocks
* TCPOLEN_SACK_PERBLOCK
;
676 /* TCP SMALL QUEUES (TSQ)
678 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
679 * to reduce RTT and bufferbloat.
680 * We do this using a special skb destructor (tcp_wfree).
682 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
683 * needs to be reallocated in a driver.
684 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
686 * Since transmit from skb destructor is forbidden, we use a tasklet
687 * to process all sockets that eventually need to send more skbs.
688 * We use one tasklet per cpu, with its own queue of sockets.
691 struct tasklet_struct tasklet
;
692 struct list_head head
; /* queue of tcp sockets */
694 static DEFINE_PER_CPU(struct tsq_tasklet
, tsq_tasklet
);
696 static void tcp_tsq_handler(struct sock
*sk
)
698 if ((1 << sk
->sk_state
) &
699 (TCPF_ESTABLISHED
| TCPF_FIN_WAIT1
| TCPF_CLOSING
|
700 TCPF_CLOSE_WAIT
| TCPF_LAST_ACK
))
701 tcp_write_xmit(sk
, tcp_current_mss(sk
), 0, 0, GFP_ATOMIC
);
704 * One tasklet per cpu tries to send more skbs.
705 * We run in tasklet context but need to disable irqs when
706 * transferring tsq->head because tcp_wfree() might
707 * interrupt us (non NAPI drivers)
709 static void tcp_tasklet_func(unsigned long data
)
711 struct tsq_tasklet
*tsq
= (struct tsq_tasklet
*)data
;
714 struct list_head
*q
, *n
;
718 local_irq_save(flags
);
719 list_splice_init(&tsq
->head
, &list
);
720 local_irq_restore(flags
);
722 list_for_each_safe(q
, n
, &list
) {
723 tp
= list_entry(q
, struct tcp_sock
, tsq_node
);
724 list_del(&tp
->tsq_node
);
726 sk
= (struct sock
*)tp
;
729 if (!sock_owned_by_user(sk
)) {
732 /* defer the work to tcp_release_cb() */
733 set_bit(TCP_TSQ_DEFERRED
, &tp
->tsq_flags
);
737 clear_bit(TSQ_QUEUED
, &tp
->tsq_flags
);
742 #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \
743 (1UL << TCP_WRITE_TIMER_DEFERRED) | \
744 (1UL << TCP_DELACK_TIMER_DEFERRED) | \
745 (1UL << TCP_MTU_REDUCED_DEFERRED))
747 * tcp_release_cb - tcp release_sock() callback
750 * called from release_sock() to perform protocol dependent
751 * actions before socket release.
753 void tcp_release_cb(struct sock
*sk
)
755 struct tcp_sock
*tp
= tcp_sk(sk
);
756 unsigned long flags
, nflags
;
758 /* perform an atomic operation only if at least one flag is set */
760 flags
= tp
->tsq_flags
;
761 if (!(flags
& TCP_DEFERRED_ALL
))
763 nflags
= flags
& ~TCP_DEFERRED_ALL
;
764 } while (cmpxchg(&tp
->tsq_flags
, flags
, nflags
) != flags
);
766 if (flags
& (1UL << TCP_TSQ_DEFERRED
))
769 if (flags
& (1UL << TCP_WRITE_TIMER_DEFERRED
)) {
770 tcp_write_timer_handler(sk
);
773 if (flags
& (1UL << TCP_DELACK_TIMER_DEFERRED
)) {
774 tcp_delack_timer_handler(sk
);
777 if (flags
& (1UL << TCP_MTU_REDUCED_DEFERRED
)) {
778 sk
->sk_prot
->mtu_reduced(sk
);
782 EXPORT_SYMBOL(tcp_release_cb
);
784 void __init
tcp_tasklet_init(void)
788 for_each_possible_cpu(i
) {
789 struct tsq_tasklet
*tsq
= &per_cpu(tsq_tasklet
, i
);
791 INIT_LIST_HEAD(&tsq
->head
);
792 tasklet_init(&tsq
->tasklet
,
799 * Write buffer destructor automatically called from kfree_skb.
800 * We can't xmit new skbs from this context, as we might already
803 void tcp_wfree(struct sk_buff
*skb
)
805 struct sock
*sk
= skb
->sk
;
806 struct tcp_sock
*tp
= tcp_sk(sk
);
808 if (test_and_clear_bit(TSQ_THROTTLED
, &tp
->tsq_flags
) &&
809 !test_and_set_bit(TSQ_QUEUED
, &tp
->tsq_flags
)) {
811 struct tsq_tasklet
*tsq
;
813 /* Keep a ref on socket.
814 * This last ref will be released in tcp_tasklet_func()
816 atomic_sub(skb
->truesize
- 1, &sk
->sk_wmem_alloc
);
818 /* queue this socket to tasklet queue */
819 local_irq_save(flags
);
820 tsq
= &__get_cpu_var(tsq_tasklet
);
821 list_add(&tp
->tsq_node
, &tsq
->head
);
822 tasklet_schedule(&tsq
->tasklet
);
823 local_irq_restore(flags
);
829 /* This routine actually transmits TCP packets queued in by
830 * tcp_do_sendmsg(). This is used by both the initial
831 * transmission and possible later retransmissions.
832 * All SKB's seen here are completely headerless. It is our
833 * job to build the TCP header, and pass the packet down to
834 * IP so it can do the same plus pass the packet off to the
837 * We are working here with either a clone of the original
838 * SKB, or a fresh unique copy made by the retransmit engine.
840 static int tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int clone_it
,
843 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
844 struct inet_sock
*inet
;
846 struct tcp_skb_cb
*tcb
;
847 struct tcp_out_options opts
;
848 unsigned int tcp_options_size
, tcp_header_size
;
849 struct tcp_md5sig_key
*md5
;
853 BUG_ON(!skb
|| !tcp_skb_pcount(skb
));
856 const struct sk_buff
*fclone
= skb
+ 1;
858 /* If congestion control is doing timestamping, we must
859 * take such a timestamp before we potentially clone/copy.
861 if (icsk
->icsk_ca_ops
->flags
& TCP_CONG_RTT_STAMP
)
862 __net_timestamp(skb
);
864 if (unlikely(skb
->fclone
== SKB_FCLONE_ORIG
&&
865 fclone
->fclone
== SKB_FCLONE_CLONE
))
866 NET_INC_STATS_BH(sock_net(sk
),
867 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES
);
869 if (unlikely(skb_cloned(skb
)))
870 skb
= pskb_copy(skb
, gfp_mask
);
872 skb
= skb_clone(skb
, gfp_mask
);
879 tcb
= TCP_SKB_CB(skb
);
880 memset(&opts
, 0, sizeof(opts
));
882 if (unlikely(tcb
->tcp_flags
& TCPHDR_SYN
))
883 tcp_options_size
= tcp_syn_options(sk
, skb
, &opts
, &md5
);
885 tcp_options_size
= tcp_established_options(sk
, skb
, &opts
,
887 tcp_header_size
= tcp_options_size
+ sizeof(struct tcphdr
);
889 if (tcp_packets_in_flight(tp
) == 0)
890 tcp_ca_event(sk
, CA_EVENT_TX_START
);
892 /* if no packet is in qdisc/device queue, then allow XPS to select
895 skb
->ooo_okay
= sk_wmem_alloc_get(sk
) == 0;
897 skb_push(skb
, tcp_header_size
);
898 skb_reset_transport_header(skb
);
902 skb
->destructor
= tcp_wfree
;
903 atomic_add(skb
->truesize
, &sk
->sk_wmem_alloc
);
905 /* Build TCP header and checksum it. */
907 th
->source
= inet
->inet_sport
;
908 th
->dest
= inet
->inet_dport
;
909 th
->seq
= htonl(tcb
->seq
);
910 th
->ack_seq
= htonl(tp
->rcv_nxt
);
911 *(((__be16
*)th
) + 6) = htons(((tcp_header_size
>> 2) << 12) |
914 if (unlikely(tcb
->tcp_flags
& TCPHDR_SYN
)) {
915 /* RFC1323: The window in SYN & SYN/ACK segments
918 th
->window
= htons(min(tp
->rcv_wnd
, 65535U));
920 th
->window
= htons(tcp_select_window(sk
));
925 /* The urg_mode check is necessary during a below snd_una win probe */
926 if (unlikely(tcp_urg_mode(tp
) && before(tcb
->seq
, tp
->snd_up
))) {
927 if (before(tp
->snd_up
, tcb
->seq
+ 0x10000)) {
928 th
->urg_ptr
= htons(tp
->snd_up
- tcb
->seq
);
930 } else if (after(tcb
->seq
+ 0xFFFF, tp
->snd_nxt
)) {
931 th
->urg_ptr
= htons(0xFFFF);
936 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
);
937 if (likely((tcb
->tcp_flags
& TCPHDR_SYN
) == 0))
938 TCP_ECN_send(sk
, skb
, tcp_header_size
);
940 #ifdef CONFIG_TCP_MD5SIG
941 /* Calculate the MD5 hash, as we have all we need now */
943 sk_nocaps_add(sk
, NETIF_F_GSO_MASK
);
944 tp
->af_specific
->calc_md5_hash(opts
.hash_location
,
949 icsk
->icsk_af_ops
->send_check(sk
, skb
);
951 if (likely(tcb
->tcp_flags
& TCPHDR_ACK
))
952 tcp_event_ack_sent(sk
, tcp_skb_pcount(skb
));
954 if (skb
->len
!= tcp_header_size
)
955 tcp_event_data_sent(tp
, sk
);
957 if (after(tcb
->end_seq
, tp
->snd_nxt
) || tcb
->seq
== tcb
->end_seq
)
958 TCP_ADD_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
,
959 tcp_skb_pcount(skb
));
961 err
= icsk
->icsk_af_ops
->queue_xmit(skb
, &inet
->cork
.fl
);
962 if (likely(err
<= 0))
965 tcp_enter_cwr(sk
, 1);
967 return net_xmit_eval(err
);
970 /* This routine just queues the buffer for sending.
972 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
973 * otherwise socket can stall.
975 static void tcp_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
977 struct tcp_sock
*tp
= tcp_sk(sk
);
979 /* Advance write_seq and place onto the write_queue. */
980 tp
->write_seq
= TCP_SKB_CB(skb
)->end_seq
;
981 skb_header_release(skb
);
982 tcp_add_write_queue_tail(sk
, skb
);
983 sk
->sk_wmem_queued
+= skb
->truesize
;
984 sk_mem_charge(sk
, skb
->truesize
);
987 /* Initialize TSO segments for a packet. */
988 static void tcp_set_skb_tso_segs(const struct sock
*sk
, struct sk_buff
*skb
,
989 unsigned int mss_now
)
991 struct skb_shared_info
*shinfo
= skb_shinfo(skb
);
993 /* Make sure we own this skb before messing gso_size/gso_segs */
994 WARN_ON_ONCE(skb_cloned(skb
));
996 if (skb
->len
<= mss_now
|| skb
->ip_summed
== CHECKSUM_NONE
) {
997 /* Avoid the costly divide in the normal
1000 shinfo
->gso_segs
= 1;
1001 shinfo
->gso_size
= 0;
1002 shinfo
->gso_type
= 0;
1004 shinfo
->gso_segs
= DIV_ROUND_UP(skb
->len
, mss_now
);
1005 shinfo
->gso_size
= mss_now
;
1006 shinfo
->gso_type
= sk
->sk_gso_type
;
1010 /* When a modification to fackets out becomes necessary, we need to check
1011 * skb is counted to fackets_out or not.
1013 static void tcp_adjust_fackets_out(struct sock
*sk
, const struct sk_buff
*skb
,
1016 struct tcp_sock
*tp
= tcp_sk(sk
);
1018 if (!tp
->sacked_out
|| tcp_is_reno(tp
))
1021 if (after(tcp_highest_sack_seq(tp
), TCP_SKB_CB(skb
)->seq
))
1022 tp
->fackets_out
-= decr
;
1025 /* Pcount in the middle of the write queue got changed, we need to do various
1026 * tweaks to fix counters
1028 static void tcp_adjust_pcount(struct sock
*sk
, const struct sk_buff
*skb
, int decr
)
1030 struct tcp_sock
*tp
= tcp_sk(sk
);
1032 tp
->packets_out
-= decr
;
1034 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
1035 tp
->sacked_out
-= decr
;
1036 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
1037 tp
->retrans_out
-= decr
;
1038 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
)
1039 tp
->lost_out
-= decr
;
1041 /* Reno case is special. Sigh... */
1042 if (tcp_is_reno(tp
) && decr
> 0)
1043 tp
->sacked_out
-= min_t(u32
, tp
->sacked_out
, decr
);
1045 tcp_adjust_fackets_out(sk
, skb
, decr
);
1047 if (tp
->lost_skb_hint
&&
1048 before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
) &&
1049 (tcp_is_fack(tp
) || (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)))
1050 tp
->lost_cnt_hint
-= decr
;
1052 tcp_verify_left_out(tp
);
1055 /* Function to create two new TCP segments. Shrinks the given segment
1056 * to the specified size and appends a new segment with the rest of the
1057 * packet to the list. This won't be called frequently, I hope.
1058 * Remember, these are still headerless SKBs at this point.
1060 int tcp_fragment(struct sock
*sk
, struct sk_buff
*skb
, u32 len
,
1061 unsigned int mss_now
)
1063 struct tcp_sock
*tp
= tcp_sk(sk
);
1064 struct sk_buff
*buff
;
1065 int nsize
, old_factor
;
1069 if (WARN_ON(len
> skb
->len
))
1072 nsize
= skb_headlen(skb
) - len
;
1076 if (skb_unclone(skb
, GFP_ATOMIC
))
1079 /* Get a new skb... force flag on. */
1080 buff
= sk_stream_alloc_skb(sk
, nsize
, GFP_ATOMIC
);
1082 return -ENOMEM
; /* We'll just try again later. */
1084 sk
->sk_wmem_queued
+= buff
->truesize
;
1085 sk_mem_charge(sk
, buff
->truesize
);
1086 nlen
= skb
->len
- len
- nsize
;
1087 buff
->truesize
+= nlen
;
1088 skb
->truesize
-= nlen
;
1090 /* Correct the sequence numbers. */
1091 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1092 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1093 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1095 /* PSH and FIN should only be set in the second packet. */
1096 flags
= TCP_SKB_CB(skb
)->tcp_flags
;
1097 TCP_SKB_CB(skb
)->tcp_flags
= flags
& ~(TCPHDR_FIN
| TCPHDR_PSH
);
1098 TCP_SKB_CB(buff
)->tcp_flags
= flags
;
1099 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
1101 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_PARTIAL
) {
1102 /* Copy and checksum data tail into the new buffer. */
1103 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
,
1104 skb_put(buff
, nsize
),
1109 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
1111 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1112 skb_split(skb
, buff
, len
);
1115 buff
->ip_summed
= skb
->ip_summed
;
1117 /* Looks stupid, but our code really uses when of
1118 * skbs, which it never sent before. --ANK
1120 TCP_SKB_CB(buff
)->when
= TCP_SKB_CB(skb
)->when
;
1121 buff
->tstamp
= skb
->tstamp
;
1123 old_factor
= tcp_skb_pcount(skb
);
1125 /* Fix up tso_factor for both original and new SKB. */
1126 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1127 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1129 /* If this packet has been sent out already, we must
1130 * adjust the various packet counters.
1132 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
1133 int diff
= old_factor
- tcp_skb_pcount(skb
) -
1134 tcp_skb_pcount(buff
);
1137 tcp_adjust_pcount(sk
, skb
, diff
);
1140 /* Link BUFF into the send queue. */
1141 skb_header_release(buff
);
1142 tcp_insert_write_queue_after(skb
, buff
, sk
);
1147 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1148 * eventually). The difference is that pulled data not copied, but
1149 * immediately discarded.
1151 static void __pskb_trim_head(struct sk_buff
*skb
, int len
)
1153 struct skb_shared_info
*shinfo
;
1156 eat
= min_t(int, len
, skb_headlen(skb
));
1158 __skb_pull(skb
, eat
);
1165 shinfo
= skb_shinfo(skb
);
1166 for (i
= 0; i
< shinfo
->nr_frags
; i
++) {
1167 int size
= skb_frag_size(&shinfo
->frags
[i
]);
1170 skb_frag_unref(skb
, i
);
1173 shinfo
->frags
[k
] = shinfo
->frags
[i
];
1175 shinfo
->frags
[k
].page_offset
+= eat
;
1176 skb_frag_size_sub(&shinfo
->frags
[k
], eat
);
1182 shinfo
->nr_frags
= k
;
1184 skb_reset_tail_pointer(skb
);
1185 skb
->data_len
-= len
;
1186 skb
->len
= skb
->data_len
;
1189 /* Remove acked data from a packet in the transmit queue. */
1190 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
1192 if (skb_unclone(skb
, GFP_ATOMIC
))
1195 __pskb_trim_head(skb
, len
);
1197 TCP_SKB_CB(skb
)->seq
+= len
;
1198 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1200 skb
->truesize
-= len
;
1201 sk
->sk_wmem_queued
-= len
;
1202 sk_mem_uncharge(sk
, len
);
1203 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
1205 /* Any change of skb->len requires recalculation of tso factor. */
1206 if (tcp_skb_pcount(skb
) > 1)
1207 tcp_set_skb_tso_segs(sk
, skb
, tcp_skb_mss(skb
));
1212 /* Calculate MSS not accounting any TCP options. */
1213 static inline int __tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1215 const struct tcp_sock
*tp
= tcp_sk(sk
);
1216 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1219 /* Calculate base mss without TCP options:
1220 It is MMS_S - sizeof(tcphdr) of rfc1122
1222 mss_now
= pmtu
- icsk
->icsk_af_ops
->net_header_len
- sizeof(struct tcphdr
);
1224 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1225 if (icsk
->icsk_af_ops
->net_frag_header_len
) {
1226 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1228 if (dst
&& dst_allfrag(dst
))
1229 mss_now
-= icsk
->icsk_af_ops
->net_frag_header_len
;
1232 /* Clamp it (mss_clamp does not include tcp options) */
1233 if (mss_now
> tp
->rx_opt
.mss_clamp
)
1234 mss_now
= tp
->rx_opt
.mss_clamp
;
1236 /* Now subtract optional transport overhead */
1237 mss_now
-= icsk
->icsk_ext_hdr_len
;
1239 /* Then reserve room for full set of TCP options and 8 bytes of data */
1245 /* Calculate MSS. Not accounting for SACKs here. */
1246 int tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1248 /* Subtract TCP options size, not including SACKs */
1249 return __tcp_mtu_to_mss(sk
, pmtu
) -
1250 (tcp_sk(sk
)->tcp_header_len
- sizeof(struct tcphdr
));
1253 /* Inverse of above */
1254 int tcp_mss_to_mtu(struct sock
*sk
, int mss
)
1256 const struct tcp_sock
*tp
= tcp_sk(sk
);
1257 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1261 tp
->tcp_header_len
+
1262 icsk
->icsk_ext_hdr_len
+
1263 icsk
->icsk_af_ops
->net_header_len
;
1265 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1266 if (icsk
->icsk_af_ops
->net_frag_header_len
) {
1267 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1269 if (dst
&& dst_allfrag(dst
))
1270 mtu
+= icsk
->icsk_af_ops
->net_frag_header_len
;
1275 /* MTU probing init per socket */
1276 void tcp_mtup_init(struct sock
*sk
)
1278 struct tcp_sock
*tp
= tcp_sk(sk
);
1279 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1281 icsk
->icsk_mtup
.enabled
= sysctl_tcp_mtu_probing
> 1;
1282 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+ sizeof(struct tcphdr
) +
1283 icsk
->icsk_af_ops
->net_header_len
;
1284 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, sysctl_tcp_base_mss
);
1285 icsk
->icsk_mtup
.probe_size
= 0;
1287 EXPORT_SYMBOL(tcp_mtup_init
);
1289 /* This function synchronize snd mss to current pmtu/exthdr set.
1291 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1292 for TCP options, but includes only bare TCP header.
1294 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1295 It is minimum of user_mss and mss received with SYN.
1296 It also does not include TCP options.
1298 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1300 tp->mss_cache is current effective sending mss, including
1301 all tcp options except for SACKs. It is evaluated,
1302 taking into account current pmtu, but never exceeds
1303 tp->rx_opt.mss_clamp.
1305 NOTE1. rfc1122 clearly states that advertised MSS
1306 DOES NOT include either tcp or ip options.
1308 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1309 are READ ONLY outside this function. --ANK (980731)
1311 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
1313 struct tcp_sock
*tp
= tcp_sk(sk
);
1314 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1317 if (icsk
->icsk_mtup
.search_high
> pmtu
)
1318 icsk
->icsk_mtup
.search_high
= pmtu
;
1320 mss_now
= tcp_mtu_to_mss(sk
, pmtu
);
1321 mss_now
= tcp_bound_to_half_wnd(tp
, mss_now
);
1323 /* And store cached results */
1324 icsk
->icsk_pmtu_cookie
= pmtu
;
1325 if (icsk
->icsk_mtup
.enabled
)
1326 mss_now
= min(mss_now
, tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
));
1327 tp
->mss_cache
= mss_now
;
1331 EXPORT_SYMBOL(tcp_sync_mss
);
1333 /* Compute the current effective MSS, taking SACKs and IP options,
1334 * and even PMTU discovery events into account.
1336 unsigned int tcp_current_mss(struct sock
*sk
)
1338 const struct tcp_sock
*tp
= tcp_sk(sk
);
1339 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1341 unsigned int header_len
;
1342 struct tcp_out_options opts
;
1343 struct tcp_md5sig_key
*md5
;
1345 mss_now
= tp
->mss_cache
;
1348 u32 mtu
= dst_mtu(dst
);
1349 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
1350 mss_now
= tcp_sync_mss(sk
, mtu
);
1353 header_len
= tcp_established_options(sk
, NULL
, &opts
, &md5
) +
1354 sizeof(struct tcphdr
);
1355 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1356 * some common options. If this is an odd packet (because we have SACK
1357 * blocks etc) then our calculated header_len will be different, and
1358 * we have to adjust mss_now correspondingly */
1359 if (header_len
!= tp
->tcp_header_len
) {
1360 int delta
= (int) header_len
- tp
->tcp_header_len
;
1367 /* Congestion window validation. (RFC2861) */
1368 static void tcp_cwnd_validate(struct sock
*sk
)
1370 struct tcp_sock
*tp
= tcp_sk(sk
);
1372 if (tp
->packets_out
>= tp
->snd_cwnd
) {
1373 /* Network is feed fully. */
1374 tp
->snd_cwnd_used
= 0;
1375 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1377 /* Network starves. */
1378 if (tp
->packets_out
> tp
->snd_cwnd_used
)
1379 tp
->snd_cwnd_used
= tp
->packets_out
;
1381 if (sysctl_tcp_slow_start_after_idle
&&
1382 (s32
)(tcp_time_stamp
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
)
1383 tcp_cwnd_application_limited(sk
);
1387 /* Minshall's variant of the Nagle send check. */
1388 static bool tcp_minshall_check(const struct tcp_sock
*tp
)
1390 return after(tp
->snd_sml
, tp
->snd_una
) &&
1391 !after(tp
->snd_sml
, tp
->snd_nxt
);
1394 /* Update snd_sml if this skb is under mss
1395 * Note that a TSO packet might end with a sub-mss segment
1396 * The test is really :
1397 * if ((skb->len % mss) != 0)
1398 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1399 * But we can avoid doing the divide again given we already have
1400 * skb_pcount = skb->len / mss_now
1402 static void tcp_minshall_update(struct tcp_sock
*tp
, unsigned int mss_now
,
1403 const struct sk_buff
*skb
)
1405 if (skb
->len
< tcp_skb_pcount(skb
) * mss_now
)
1406 tp
->snd_sml
= TCP_SKB_CB(skb
)->end_seq
;
1409 /* Return false, if packet can be sent now without violation Nagle's rules:
1410 * 1. It is full sized. (provided by caller in %partial bool)
1411 * 2. Or it contains FIN. (already checked by caller)
1412 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1413 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1414 * With Minshall's modification: all sent small packets are ACKed.
1416 static bool tcp_nagle_check(bool partial
, const struct tcp_sock
*tp
,
1417 unsigned int mss_now
, int nonagle
)
1420 ((nonagle
& TCP_NAGLE_CORK
) ||
1421 (!nonagle
&& tp
->packets_out
&& tcp_minshall_check(tp
)));
1423 /* Returns the portion of skb which can be sent right away */
1424 static unsigned int tcp_mss_split_point(const struct sock
*sk
,
1425 const struct sk_buff
*skb
,
1426 unsigned int mss_now
,
1427 unsigned int max_segs
,
1430 const struct tcp_sock
*tp
= tcp_sk(sk
);
1431 u32 partial
, needed
, window
, max_len
;
1433 window
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1434 max_len
= mss_now
* max_segs
;
1436 if (likely(max_len
<= window
&& skb
!= tcp_write_queue_tail(sk
)))
1439 needed
= min(skb
->len
, window
);
1441 if (max_len
<= needed
)
1444 partial
= needed
% mss_now
;
1445 /* If last segment is not a full MSS, check if Nagle rules allow us
1446 * to include this last segment in this skb.
1447 * Otherwise, we'll split the skb at last MSS boundary
1449 if (tcp_nagle_check(partial
!= 0, tp
, mss_now
, nonagle
))
1450 return needed
- partial
;
1455 /* Can at least one segment of SKB be sent right now, according to the
1456 * congestion window rules? If so, return how many segments are allowed.
1458 static inline unsigned int tcp_cwnd_test(const struct tcp_sock
*tp
,
1459 const struct sk_buff
*skb
)
1461 u32 in_flight
, cwnd
;
1463 /* Don't be strict about the congestion window for the final FIN. */
1464 if ((TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
) &&
1465 tcp_skb_pcount(skb
) == 1)
1468 in_flight
= tcp_packets_in_flight(tp
);
1469 cwnd
= tp
->snd_cwnd
;
1470 if (in_flight
< cwnd
)
1471 return (cwnd
- in_flight
);
1476 /* Initialize TSO state of a skb.
1477 * This must be invoked the first time we consider transmitting
1478 * SKB onto the wire.
1480 static int tcp_init_tso_segs(const struct sock
*sk
, struct sk_buff
*skb
,
1481 unsigned int mss_now
)
1483 int tso_segs
= tcp_skb_pcount(skb
);
1485 if (!tso_segs
|| (tso_segs
> 1 && tcp_skb_mss(skb
) != mss_now
)) {
1486 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1487 tso_segs
= tcp_skb_pcount(skb
);
1493 /* Return true if the Nagle test allows this packet to be
1496 static inline bool tcp_nagle_test(const struct tcp_sock
*tp
, const struct sk_buff
*skb
,
1497 unsigned int cur_mss
, int nonagle
)
1499 /* Nagle rule does not apply to frames, which sit in the middle of the
1500 * write_queue (they have no chances to get new data).
1502 * This is implemented in the callers, where they modify the 'nonagle'
1503 * argument based upon the location of SKB in the send queue.
1505 if (nonagle
& TCP_NAGLE_PUSH
)
1508 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1509 if (tcp_urg_mode(tp
) || (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
))
1512 if (!tcp_nagle_check(skb
->len
< cur_mss
, tp
, cur_mss
, nonagle
))
1518 /* Does at least the first segment of SKB fit into the send window? */
1519 static bool tcp_snd_wnd_test(const struct tcp_sock
*tp
,
1520 const struct sk_buff
*skb
,
1521 unsigned int cur_mss
)
1523 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1525 if (skb
->len
> cur_mss
)
1526 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
1528 return !after(end_seq
, tcp_wnd_end(tp
));
1531 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1532 * should be put on the wire right now. If so, it returns the number of
1533 * packets allowed by the congestion window.
1535 static unsigned int tcp_snd_test(const struct sock
*sk
, struct sk_buff
*skb
,
1536 unsigned int cur_mss
, int nonagle
)
1538 const struct tcp_sock
*tp
= tcp_sk(sk
);
1539 unsigned int cwnd_quota
;
1541 tcp_init_tso_segs(sk
, skb
, cur_mss
);
1543 if (!tcp_nagle_test(tp
, skb
, cur_mss
, nonagle
))
1546 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1547 if (cwnd_quota
&& !tcp_snd_wnd_test(tp
, skb
, cur_mss
))
1553 /* Test if sending is allowed right now. */
1554 bool tcp_may_send_now(struct sock
*sk
)
1556 const struct tcp_sock
*tp
= tcp_sk(sk
);
1557 struct sk_buff
*skb
= tcp_send_head(sk
);
1560 tcp_snd_test(sk
, skb
, tcp_current_mss(sk
),
1561 (tcp_skb_is_last(sk
, skb
) ?
1562 tp
->nonagle
: TCP_NAGLE_PUSH
));
1565 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1566 * which is put after SKB on the list. It is very much like
1567 * tcp_fragment() except that it may make several kinds of assumptions
1568 * in order to speed up the splitting operation. In particular, we
1569 * know that all the data is in scatter-gather pages, and that the
1570 * packet has never been sent out before (and thus is not cloned).
1572 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
,
1573 unsigned int mss_now
, gfp_t gfp
)
1575 struct sk_buff
*buff
;
1576 int nlen
= skb
->len
- len
;
1579 /* All of a TSO frame must be composed of paged data. */
1580 if (skb
->len
!= skb
->data_len
)
1581 return tcp_fragment(sk
, skb
, len
, mss_now
);
1583 buff
= sk_stream_alloc_skb(sk
, 0, gfp
);
1584 if (unlikely(buff
== NULL
))
1587 sk
->sk_wmem_queued
+= buff
->truesize
;
1588 sk_mem_charge(sk
, buff
->truesize
);
1589 buff
->truesize
+= nlen
;
1590 skb
->truesize
-= nlen
;
1592 /* Correct the sequence numbers. */
1593 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1594 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1595 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1597 /* PSH and FIN should only be set in the second packet. */
1598 flags
= TCP_SKB_CB(skb
)->tcp_flags
;
1599 TCP_SKB_CB(skb
)->tcp_flags
= flags
& ~(TCPHDR_FIN
| TCPHDR_PSH
);
1600 TCP_SKB_CB(buff
)->tcp_flags
= flags
;
1602 /* This packet was never sent out yet, so no SACK bits. */
1603 TCP_SKB_CB(buff
)->sacked
= 0;
1605 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_PARTIAL
;
1606 skb_split(skb
, buff
, len
);
1608 /* Fix up tso_factor for both original and new SKB. */
1609 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1610 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1612 /* Link BUFF into the send queue. */
1613 skb_header_release(buff
);
1614 tcp_insert_write_queue_after(skb
, buff
, sk
);
1619 /* Try to defer sending, if possible, in order to minimize the amount
1620 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1622 * This algorithm is from John Heffner.
1624 static bool tcp_tso_should_defer(struct sock
*sk
, struct sk_buff
*skb
)
1626 struct tcp_sock
*tp
= tcp_sk(sk
);
1627 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1628 u32 send_win
, cong_win
, limit
, in_flight
;
1631 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
1634 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
1637 /* Defer for less than two clock ticks. */
1638 if (tp
->tso_deferred
&&
1639 (((u32
)jiffies
<< 1) >> 1) - (tp
->tso_deferred
>> 1) > 1)
1642 in_flight
= tcp_packets_in_flight(tp
);
1644 BUG_ON(tcp_skb_pcount(skb
) <= 1 || (tp
->snd_cwnd
<= in_flight
));
1646 send_win
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1648 /* From in_flight test above, we know that cwnd > in_flight. */
1649 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
1651 limit
= min(send_win
, cong_win
);
1653 /* If a full-sized TSO skb can be sent, do it. */
1654 if (limit
>= min_t(unsigned int, sk
->sk_gso_max_size
,
1655 tp
->xmit_size_goal_segs
* tp
->mss_cache
))
1658 /* Middle in queue won't get any more data, full sendable already? */
1659 if ((skb
!= tcp_write_queue_tail(sk
)) && (limit
>= skb
->len
))
1662 win_divisor
= ACCESS_ONCE(sysctl_tcp_tso_win_divisor
);
1664 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
1666 /* If at least some fraction of a window is available,
1669 chunk
/= win_divisor
;
1673 /* Different approach, try not to defer past a single
1674 * ACK. Receiver should ACK every other full sized
1675 * frame, so if we have space for more than 3 frames
1678 if (limit
> tcp_max_tso_deferred_mss(tp
) * tp
->mss_cache
)
1682 /* Ok, it looks like it is advisable to defer.
1683 * Do not rearm the timer if already set to not break TCP ACK clocking.
1685 if (!tp
->tso_deferred
)
1686 tp
->tso_deferred
= 1 | (jiffies
<< 1);
1691 tp
->tso_deferred
= 0;
1695 /* Create a new MTU probe if we are ready.
1696 * MTU probe is regularly attempting to increase the path MTU by
1697 * deliberately sending larger packets. This discovers routing
1698 * changes resulting in larger path MTUs.
1700 * Returns 0 if we should wait to probe (no cwnd available),
1701 * 1 if a probe was sent,
1704 static int tcp_mtu_probe(struct sock
*sk
)
1706 struct tcp_sock
*tp
= tcp_sk(sk
);
1707 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1708 struct sk_buff
*skb
, *nskb
, *next
;
1715 /* Not currently probing/verifying,
1717 * have enough cwnd, and
1718 * not SACKing (the variable headers throw things off) */
1719 if (!icsk
->icsk_mtup
.enabled
||
1720 icsk
->icsk_mtup
.probe_size
||
1721 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
||
1722 tp
->snd_cwnd
< 11 ||
1723 tp
->rx_opt
.num_sacks
|| tp
->rx_opt
.dsack
)
1726 /* Very simple search strategy: just double the MSS. */
1727 mss_now
= tcp_current_mss(sk
);
1728 probe_size
= 2 * tp
->mss_cache
;
1729 size_needed
= probe_size
+ (tp
->reordering
+ 1) * tp
->mss_cache
;
1730 if (probe_size
> tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_high
)) {
1731 /* TODO: set timer for probe_converge_event */
1735 /* Have enough data in the send queue to probe? */
1736 if (tp
->write_seq
- tp
->snd_nxt
< size_needed
)
1739 if (tp
->snd_wnd
< size_needed
)
1741 if (after(tp
->snd_nxt
+ size_needed
, tcp_wnd_end(tp
)))
1744 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1745 if (tcp_packets_in_flight(tp
) + 2 > tp
->snd_cwnd
) {
1746 if (!tcp_packets_in_flight(tp
))
1752 /* We're allowed to probe. Build it now. */
1753 if ((nskb
= sk_stream_alloc_skb(sk
, probe_size
, GFP_ATOMIC
)) == NULL
)
1755 sk
->sk_wmem_queued
+= nskb
->truesize
;
1756 sk_mem_charge(sk
, nskb
->truesize
);
1758 skb
= tcp_send_head(sk
);
1760 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(skb
)->seq
;
1761 TCP_SKB_CB(nskb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ probe_size
;
1762 TCP_SKB_CB(nskb
)->tcp_flags
= TCPHDR_ACK
;
1763 TCP_SKB_CB(nskb
)->sacked
= 0;
1765 nskb
->ip_summed
= skb
->ip_summed
;
1767 tcp_insert_write_queue_before(nskb
, skb
, sk
);
1770 tcp_for_write_queue_from_safe(skb
, next
, sk
) {
1771 copy
= min_t(int, skb
->len
, probe_size
- len
);
1772 if (nskb
->ip_summed
)
1773 skb_copy_bits(skb
, 0, skb_put(nskb
, copy
), copy
);
1775 nskb
->csum
= skb_copy_and_csum_bits(skb
, 0,
1776 skb_put(nskb
, copy
),
1779 if (skb
->len
<= copy
) {
1780 /* We've eaten all the data from this skb.
1782 TCP_SKB_CB(nskb
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
;
1783 tcp_unlink_write_queue(skb
, sk
);
1784 sk_wmem_free_skb(sk
, skb
);
1786 TCP_SKB_CB(nskb
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
&
1787 ~(TCPHDR_FIN
|TCPHDR_PSH
);
1788 if (!skb_shinfo(skb
)->nr_frags
) {
1789 skb_pull(skb
, copy
);
1790 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1791 skb
->csum
= csum_partial(skb
->data
,
1794 __pskb_trim_head(skb
, copy
);
1795 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1797 TCP_SKB_CB(skb
)->seq
+= copy
;
1802 if (len
>= probe_size
)
1805 tcp_init_tso_segs(sk
, nskb
, nskb
->len
);
1807 /* We're ready to send. If this fails, the probe will
1808 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1809 TCP_SKB_CB(nskb
)->when
= tcp_time_stamp
;
1810 if (!tcp_transmit_skb(sk
, nskb
, 1, GFP_ATOMIC
)) {
1811 /* Decrement cwnd here because we are sending
1812 * effectively two packets. */
1814 tcp_event_new_data_sent(sk
, nskb
);
1816 icsk
->icsk_mtup
.probe_size
= tcp_mss_to_mtu(sk
, nskb
->len
);
1817 tp
->mtu_probe
.probe_seq_start
= TCP_SKB_CB(nskb
)->seq
;
1818 tp
->mtu_probe
.probe_seq_end
= TCP_SKB_CB(nskb
)->end_seq
;
1826 /* This routine writes packets to the network. It advances the
1827 * send_head. This happens as incoming acks open up the remote
1830 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1831 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1832 * account rare use of URG, this is not a big flaw.
1834 * Send at most one packet when push_one > 0. Temporarily ignore
1835 * cwnd limit to force at most one packet out when push_one == 2.
1837 * Returns true, if no segments are in flight and we have queued segments,
1838 * but cannot send anything now because of SWS or another problem.
1840 static bool tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
,
1841 int push_one
, gfp_t gfp
)
1843 struct tcp_sock
*tp
= tcp_sk(sk
);
1844 struct sk_buff
*skb
;
1845 unsigned int tso_segs
, sent_pkts
;
1852 /* Do MTU probing. */
1853 result
= tcp_mtu_probe(sk
);
1856 } else if (result
> 0) {
1861 while ((skb
= tcp_send_head(sk
))) {
1864 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1867 if (unlikely(tp
->repair
) && tp
->repair_queue
== TCP_SEND_QUEUE
)
1868 goto repair
; /* Skip network transmission */
1870 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1873 /* Force out a loss probe pkt. */
1879 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
)))
1882 if (tso_segs
== 1) {
1883 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
1884 (tcp_skb_is_last(sk
, skb
) ?
1885 nonagle
: TCP_NAGLE_PUSH
))))
1888 if (!push_one
&& tcp_tso_should_defer(sk
, skb
))
1892 /* TCP Small Queues :
1893 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
1895 * - better RTT estimation and ACK scheduling
1898 * Alas, some drivers / subsystems require a fair amount
1899 * of queued bytes to ensure line rate.
1900 * One example is wifi aggregation (802.11 AMPDU)
1902 limit
= max_t(unsigned int, sysctl_tcp_limit_output_bytes
,
1903 sk
->sk_pacing_rate
>> 10);
1905 if (atomic_read(&sk
->sk_wmem_alloc
) > limit
) {
1906 set_bit(TSQ_THROTTLED
, &tp
->tsq_flags
);
1911 if (tso_segs
> 1 && !tcp_urg_mode(tp
))
1912 limit
= tcp_mss_split_point(sk
, skb
, mss_now
,
1915 sk
->sk_gso_max_segs
),
1918 if (skb
->len
> limit
&&
1919 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
, gfp
)))
1922 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1924 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, gfp
)))
1928 /* Advance the send_head. This one is sent out.
1929 * This call will increment packets_out.
1931 tcp_event_new_data_sent(sk
, skb
);
1933 tcp_minshall_update(tp
, mss_now
, skb
);
1934 sent_pkts
+= tcp_skb_pcount(skb
);
1940 if (likely(sent_pkts
)) {
1941 if (tcp_in_cwnd_reduction(sk
))
1942 tp
->prr_out
+= sent_pkts
;
1944 /* Send one loss probe per tail loss episode. */
1946 tcp_schedule_loss_probe(sk
);
1947 tcp_cwnd_validate(sk
);
1950 return (push_one
== 2) || (!tp
->packets_out
&& tcp_send_head(sk
));
1953 bool tcp_schedule_loss_probe(struct sock
*sk
)
1955 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1956 struct tcp_sock
*tp
= tcp_sk(sk
);
1957 u32 timeout
, tlp_time_stamp
, rto_time_stamp
;
1958 u32 rtt
= tp
->srtt
>> 3;
1960 if (WARN_ON(icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
))
1962 /* No consecutive loss probes. */
1963 if (WARN_ON(icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
)) {
1967 /* Don't do any loss probe on a Fast Open connection before 3WHS
1970 if (sk
->sk_state
== TCP_SYN_RECV
)
1973 /* TLP is only scheduled when next timer event is RTO. */
1974 if (icsk
->icsk_pending
!= ICSK_TIME_RETRANS
)
1977 /* Schedule a loss probe in 2*RTT for SACK capable connections
1978 * in Open state, that are either limited by cwnd or application.
1980 if (sysctl_tcp_early_retrans
< 3 || !rtt
|| !tp
->packets_out
||
1981 !tcp_is_sack(tp
) || inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
)
1984 if ((tp
->snd_cwnd
> tcp_packets_in_flight(tp
)) &&
1988 /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
1989 * for delayed ack when there's one outstanding packet.
1992 if (tp
->packets_out
== 1)
1993 timeout
= max_t(u32
, timeout
,
1994 (rtt
+ (rtt
>> 1) + TCP_DELACK_MAX
));
1995 timeout
= max_t(u32
, timeout
, msecs_to_jiffies(10));
1997 /* If RTO is shorter, just schedule TLP in its place. */
1998 tlp_time_stamp
= tcp_time_stamp
+ timeout
;
1999 rto_time_stamp
= (u32
)inet_csk(sk
)->icsk_timeout
;
2000 if ((s32
)(tlp_time_stamp
- rto_time_stamp
) > 0) {
2001 s32 delta
= rto_time_stamp
- tcp_time_stamp
;
2006 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_LOSS_PROBE
, timeout
,
2011 /* When probe timeout (PTO) fires, send a new segment if one exists, else
2012 * retransmit the last segment.
2014 void tcp_send_loss_probe(struct sock
*sk
)
2016 struct tcp_sock
*tp
= tcp_sk(sk
);
2017 struct sk_buff
*skb
;
2019 int mss
= tcp_current_mss(sk
);
2022 if (tcp_send_head(sk
) != NULL
) {
2023 err
= tcp_write_xmit(sk
, mss
, TCP_NAGLE_OFF
, 2, GFP_ATOMIC
);
2027 /* At most one outstanding TLP retransmission. */
2028 if (tp
->tlp_high_seq
)
2031 /* Retransmit last segment. */
2032 skb
= tcp_write_queue_tail(sk
);
2036 pcount
= tcp_skb_pcount(skb
);
2037 if (WARN_ON(!pcount
))
2040 if ((pcount
> 1) && (skb
->len
> (pcount
- 1) * mss
)) {
2041 if (unlikely(tcp_fragment(sk
, skb
, (pcount
- 1) * mss
, mss
)))
2043 skb
= tcp_write_queue_tail(sk
);
2046 if (WARN_ON(!skb
|| !tcp_skb_pcount(skb
)))
2049 /* Probe with zero data doesn't trigger fast recovery. */
2051 err
= __tcp_retransmit_skb(sk
, skb
);
2053 /* Record snd_nxt for loss detection. */
2055 tp
->tlp_high_seq
= tp
->snd_nxt
;
2058 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2059 inet_csk(sk
)->icsk_rto
,
2063 NET_INC_STATS_BH(sock_net(sk
),
2064 LINUX_MIB_TCPLOSSPROBES
);
2068 /* Push out any pending frames which were held back due to
2069 * TCP_CORK or attempt at coalescing tiny packets.
2070 * The socket must be locked by the caller.
2072 void __tcp_push_pending_frames(struct sock
*sk
, unsigned int cur_mss
,
2075 /* If we are closed, the bytes will have to remain here.
2076 * In time closedown will finish, we empty the write queue and
2077 * all will be happy.
2079 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
2082 if (tcp_write_xmit(sk
, cur_mss
, nonagle
, 0,
2083 sk_gfp_atomic(sk
, GFP_ATOMIC
)))
2084 tcp_check_probe_timer(sk
);
2087 /* Send _single_ skb sitting at the send head. This function requires
2088 * true push pending frames to setup probe timer etc.
2090 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
2092 struct sk_buff
*skb
= tcp_send_head(sk
);
2094 BUG_ON(!skb
|| skb
->len
< mss_now
);
2096 tcp_write_xmit(sk
, mss_now
, TCP_NAGLE_PUSH
, 1, sk
->sk_allocation
);
2099 /* This function returns the amount that we can raise the
2100 * usable window based on the following constraints
2102 * 1. The window can never be shrunk once it is offered (RFC 793)
2103 * 2. We limit memory per socket
2106 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2107 * RECV.NEXT + RCV.WIN fixed until:
2108 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2110 * i.e. don't raise the right edge of the window until you can raise
2111 * it at least MSS bytes.
2113 * Unfortunately, the recommended algorithm breaks header prediction,
2114 * since header prediction assumes th->window stays fixed.
2116 * Strictly speaking, keeping th->window fixed violates the receiver
2117 * side SWS prevention criteria. The problem is that under this rule
2118 * a stream of single byte packets will cause the right side of the
2119 * window to always advance by a single byte.
2121 * Of course, if the sender implements sender side SWS prevention
2122 * then this will not be a problem.
2124 * BSD seems to make the following compromise:
2126 * If the free space is less than the 1/4 of the maximum
2127 * space available and the free space is less than 1/2 mss,
2128 * then set the window to 0.
2129 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2130 * Otherwise, just prevent the window from shrinking
2131 * and from being larger than the largest representable value.
2133 * This prevents incremental opening of the window in the regime
2134 * where TCP is limited by the speed of the reader side taking
2135 * data out of the TCP receive queue. It does nothing about
2136 * those cases where the window is constrained on the sender side
2137 * because the pipeline is full.
2139 * BSD also seems to "accidentally" limit itself to windows that are a
2140 * multiple of MSS, at least until the free space gets quite small.
2141 * This would appear to be a side effect of the mbuf implementation.
2142 * Combining these two algorithms results in the observed behavior
2143 * of having a fixed window size at almost all times.
2145 * Below we obtain similar behavior by forcing the offered window to
2146 * a multiple of the mss when it is feasible to do so.
2148 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2149 * Regular options like TIMESTAMP are taken into account.
2151 u32
__tcp_select_window(struct sock
*sk
)
2153 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2154 struct tcp_sock
*tp
= tcp_sk(sk
);
2155 /* MSS for the peer's data. Previous versions used mss_clamp
2156 * here. I don't know if the value based on our guesses
2157 * of peer's MSS is better for the performance. It's more correct
2158 * but may be worse for the performance because of rcv_mss
2159 * fluctuations. --SAW 1998/11/1
2161 int mss
= icsk
->icsk_ack
.rcv_mss
;
2162 int free_space
= tcp_space(sk
);
2163 int full_space
= min_t(int, tp
->window_clamp
, tcp_full_space(sk
));
2166 if (mss
> full_space
)
2169 if (free_space
< (full_space
>> 1)) {
2170 icsk
->icsk_ack
.quick
= 0;
2172 if (sk_under_memory_pressure(sk
))
2173 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
,
2176 if (free_space
< mss
)
2180 if (free_space
> tp
->rcv_ssthresh
)
2181 free_space
= tp
->rcv_ssthresh
;
2183 /* Don't do rounding if we are using window scaling, since the
2184 * scaled window will not line up with the MSS boundary anyway.
2186 window
= tp
->rcv_wnd
;
2187 if (tp
->rx_opt
.rcv_wscale
) {
2188 window
= free_space
;
2190 /* Advertise enough space so that it won't get scaled away.
2191 * Import case: prevent zero window announcement if
2192 * 1<<rcv_wscale > mss.
2194 if (((window
>> tp
->rx_opt
.rcv_wscale
) << tp
->rx_opt
.rcv_wscale
) != window
)
2195 window
= (((window
>> tp
->rx_opt
.rcv_wscale
) + 1)
2196 << tp
->rx_opt
.rcv_wscale
);
2198 /* Get the largest window that is a nice multiple of mss.
2199 * Window clamp already applied above.
2200 * If our current window offering is within 1 mss of the
2201 * free space we just keep it. This prevents the divide
2202 * and multiply from happening most of the time.
2203 * We also don't do any window rounding when the free space
2206 if (window
<= free_space
- mss
|| window
> free_space
)
2207 window
= (free_space
/ mss
) * mss
;
2208 else if (mss
== full_space
&&
2209 free_space
> window
+ (full_space
>> 1))
2210 window
= free_space
;
2216 /* Collapses two adjacent SKB's during retransmission. */
2217 static void tcp_collapse_retrans(struct sock
*sk
, struct sk_buff
*skb
)
2219 struct tcp_sock
*tp
= tcp_sk(sk
);
2220 struct sk_buff
*next_skb
= tcp_write_queue_next(sk
, skb
);
2221 int skb_size
, next_skb_size
;
2223 skb_size
= skb
->len
;
2224 next_skb_size
= next_skb
->len
;
2226 BUG_ON(tcp_skb_pcount(skb
) != 1 || tcp_skb_pcount(next_skb
) != 1);
2228 tcp_highest_sack_combine(sk
, next_skb
, skb
);
2230 tcp_unlink_write_queue(next_skb
, sk
);
2232 skb_copy_from_linear_data(next_skb
, skb_put(skb
, next_skb_size
),
2235 if (next_skb
->ip_summed
== CHECKSUM_PARTIAL
)
2236 skb
->ip_summed
= CHECKSUM_PARTIAL
;
2238 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
2239 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
2241 /* Update sequence range on original skb. */
2242 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
2244 /* Merge over control information. This moves PSH/FIN etc. over */
2245 TCP_SKB_CB(skb
)->tcp_flags
|= TCP_SKB_CB(next_skb
)->tcp_flags
;
2247 /* All done, get rid of second SKB and account for it so
2248 * packet counting does not break.
2250 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
& TCPCB_EVER_RETRANS
;
2252 /* changed transmit queue under us so clear hints */
2253 tcp_clear_retrans_hints_partial(tp
);
2254 if (next_skb
== tp
->retransmit_skb_hint
)
2255 tp
->retransmit_skb_hint
= skb
;
2257 tcp_adjust_pcount(sk
, next_skb
, tcp_skb_pcount(next_skb
));
2259 sk_wmem_free_skb(sk
, next_skb
);
2262 /* Check if coalescing SKBs is legal. */
2263 static bool tcp_can_collapse(const struct sock
*sk
, const struct sk_buff
*skb
)
2265 if (tcp_skb_pcount(skb
) > 1)
2267 /* TODO: SACK collapsing could be used to remove this condition */
2268 if (skb_shinfo(skb
)->nr_frags
!= 0)
2270 if (skb_cloned(skb
))
2272 if (skb
== tcp_send_head(sk
))
2274 /* Some heurestics for collapsing over SACK'd could be invented */
2275 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
2281 /* Collapse packets in the retransmit queue to make to create
2282 * less packets on the wire. This is only done on retransmission.
2284 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*to
,
2287 struct tcp_sock
*tp
= tcp_sk(sk
);
2288 struct sk_buff
*skb
= to
, *tmp
;
2291 if (!sysctl_tcp_retrans_collapse
)
2293 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)
2296 tcp_for_write_queue_from_safe(skb
, tmp
, sk
) {
2297 if (!tcp_can_collapse(sk
, skb
))
2309 /* Punt if not enough space exists in the first SKB for
2310 * the data in the second
2312 if (skb
->len
> skb_availroom(to
))
2315 if (after(TCP_SKB_CB(skb
)->end_seq
, tcp_wnd_end(tp
)))
2318 tcp_collapse_retrans(sk
, to
);
2322 /* This retransmits one SKB. Policy decisions and retransmit queue
2323 * state updates are done by the caller. Returns non-zero if an
2324 * error occurred which prevented the send.
2326 int __tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
2328 struct tcp_sock
*tp
= tcp_sk(sk
);
2329 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2330 unsigned int cur_mss
;
2332 /* Inconslusive MTU probe */
2333 if (icsk
->icsk_mtup
.probe_size
) {
2334 icsk
->icsk_mtup
.probe_size
= 0;
2337 /* Do not sent more than we queued. 1/4 is reserved for possible
2338 * copying overhead: fragmentation, tunneling, mangling etc.
2340 if (atomic_read(&sk
->sk_wmem_alloc
) >
2341 min(sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2), sk
->sk_sndbuf
))
2344 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
2345 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
2347 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
2351 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
2352 return -EHOSTUNREACH
; /* Routing failure or similar. */
2354 cur_mss
= tcp_current_mss(sk
);
2356 /* If receiver has shrunk his window, and skb is out of
2357 * new window, do not retransmit it. The exception is the
2358 * case, when window is shrunk to zero. In this case
2359 * our retransmit serves as a zero window probe.
2361 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
)) &&
2362 TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
2365 if (skb
->len
> cur_mss
) {
2366 if (tcp_fragment(sk
, skb
, cur_mss
, cur_mss
))
2367 return -ENOMEM
; /* We'll try again later. */
2369 int oldpcount
= tcp_skb_pcount(skb
);
2371 if (unlikely(oldpcount
> 1)) {
2372 if (skb_unclone(skb
, GFP_ATOMIC
))
2374 tcp_init_tso_segs(sk
, skb
, cur_mss
);
2375 tcp_adjust_pcount(sk
, skb
, oldpcount
- tcp_skb_pcount(skb
));
2379 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
2381 /* Make a copy, if the first transmission SKB clone we made
2382 * is still in somebody's hands, else make a clone.
2384 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2386 /* make sure skb->data is aligned on arches that require it
2387 * and check if ack-trimming & collapsing extended the headroom
2388 * beyond what csum_start can cover.
2390 if (unlikely((NET_IP_ALIGN
&& ((unsigned long)skb
->data
& 3)) ||
2391 skb_headroom(skb
) >= 0xFFFF)) {
2392 struct sk_buff
*nskb
= __pskb_copy(skb
, MAX_TCP_HEADER
,
2394 return nskb
? tcp_transmit_skb(sk
, nskb
, 0, GFP_ATOMIC
) :
2397 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2401 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
2403 struct tcp_sock
*tp
= tcp_sk(sk
);
2404 int err
= __tcp_retransmit_skb(sk
, skb
);
2407 /* Update global TCP statistics. */
2408 TCP_INC_STATS(sock_net(sk
), TCP_MIB_RETRANSSEGS
);
2410 tp
->total_retrans
++;
2412 #if FASTRETRANS_DEBUG > 0
2413 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2414 net_dbg_ratelimited("retrans_out leaked\n");
2417 if (!tp
->retrans_out
)
2418 tp
->lost_retrans_low
= tp
->snd_nxt
;
2419 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
2420 tp
->retrans_out
+= tcp_skb_pcount(skb
);
2422 /* Save stamp of the first retransmit. */
2423 if (!tp
->retrans_stamp
)
2424 tp
->retrans_stamp
= TCP_SKB_CB(skb
)->when
;
2426 tp
->undo_retrans
+= tcp_skb_pcount(skb
);
2428 /* snd_nxt is stored to detect loss of retransmitted segment,
2429 * see tcp_input.c tcp_sacktag_write_queue().
2431 TCP_SKB_CB(skb
)->ack_seq
= tp
->snd_nxt
;
2433 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRETRANSFAIL
);
2438 /* Check if we forward retransmits are possible in the current
2439 * window/congestion state.
2441 static bool tcp_can_forward_retransmit(struct sock
*sk
)
2443 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2444 const struct tcp_sock
*tp
= tcp_sk(sk
);
2446 /* Forward retransmissions are possible only during Recovery. */
2447 if (icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
2450 /* No forward retransmissions in Reno are possible. */
2451 if (tcp_is_reno(tp
))
2454 /* Yeah, we have to make difficult choice between forward transmission
2455 * and retransmission... Both ways have their merits...
2457 * For now we do not retransmit anything, while we have some new
2458 * segments to send. In the other cases, follow rule 3 for
2459 * NextSeg() specified in RFC3517.
2462 if (tcp_may_send_now(sk
))
2468 /* This gets called after a retransmit timeout, and the initially
2469 * retransmitted data is acknowledged. It tries to continue
2470 * resending the rest of the retransmit queue, until either
2471 * we've sent it all or the congestion window limit is reached.
2472 * If doing SACK, the first ACK which comes back for a timeout
2473 * based retransmit packet might feed us FACK information again.
2474 * If so, we use it to avoid unnecessarily retransmissions.
2476 void tcp_xmit_retransmit_queue(struct sock
*sk
)
2478 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2479 struct tcp_sock
*tp
= tcp_sk(sk
);
2480 struct sk_buff
*skb
;
2481 struct sk_buff
*hole
= NULL
;
2484 int fwd_rexmitting
= 0;
2486 if (!tp
->packets_out
)
2490 tp
->retransmit_high
= tp
->snd_una
;
2492 if (tp
->retransmit_skb_hint
) {
2493 skb
= tp
->retransmit_skb_hint
;
2494 last_lost
= TCP_SKB_CB(skb
)->end_seq
;
2495 if (after(last_lost
, tp
->retransmit_high
))
2496 last_lost
= tp
->retransmit_high
;
2498 skb
= tcp_write_queue_head(sk
);
2499 last_lost
= tp
->snd_una
;
2502 tcp_for_write_queue_from(skb
, sk
) {
2503 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
2505 if (skb
== tcp_send_head(sk
))
2507 /* we could do better than to assign each time */
2509 tp
->retransmit_skb_hint
= skb
;
2511 /* Assume this retransmit will generate
2512 * only one packet for congestion window
2513 * calculation purposes. This works because
2514 * tcp_retransmit_skb() will chop up the
2515 * packet to be MSS sized and all the
2516 * packet counting works out.
2518 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
2521 if (fwd_rexmitting
) {
2523 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_highest_sack_seq(tp
)))
2525 mib_idx
= LINUX_MIB_TCPFORWARDRETRANS
;
2527 } else if (!before(TCP_SKB_CB(skb
)->seq
, tp
->retransmit_high
)) {
2528 tp
->retransmit_high
= last_lost
;
2529 if (!tcp_can_forward_retransmit(sk
))
2531 /* Backtrack if necessary to non-L'ed skb */
2539 } else if (!(sacked
& TCPCB_LOST
)) {
2540 if (hole
== NULL
&& !(sacked
& (TCPCB_SACKED_RETRANS
|TCPCB_SACKED_ACKED
)))
2545 last_lost
= TCP_SKB_CB(skb
)->end_seq
;
2546 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
2547 mib_idx
= LINUX_MIB_TCPFASTRETRANS
;
2549 mib_idx
= LINUX_MIB_TCPSLOWSTARTRETRANS
;
2552 if (sacked
& (TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))
2555 if (tcp_retransmit_skb(sk
, skb
))
2558 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2560 if (tcp_in_cwnd_reduction(sk
))
2561 tp
->prr_out
+= tcp_skb_pcount(skb
);
2563 if (skb
== tcp_write_queue_head(sk
))
2564 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2565 inet_csk(sk
)->icsk_rto
,
2570 /* Send a fin. The caller locks the socket for us. This cannot be
2571 * allowed to fail queueing a FIN frame under any circumstances.
2573 void tcp_send_fin(struct sock
*sk
)
2575 struct tcp_sock
*tp
= tcp_sk(sk
);
2576 struct sk_buff
*skb
= tcp_write_queue_tail(sk
);
2579 /* Optimization, tack on the FIN if we have a queue of
2580 * unsent frames. But be careful about outgoing SACKS
2583 mss_now
= tcp_current_mss(sk
);
2585 if (tcp_send_head(sk
) != NULL
) {
2586 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_FIN
;
2587 TCP_SKB_CB(skb
)->end_seq
++;
2590 /* Socket is locked, keep trying until memory is available. */
2592 skb
= alloc_skb_fclone(MAX_TCP_HEADER
,
2599 /* Reserve space for headers and prepare control bits. */
2600 skb_reserve(skb
, MAX_TCP_HEADER
);
2601 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2602 tcp_init_nondata_skb(skb
, tp
->write_seq
,
2603 TCPHDR_ACK
| TCPHDR_FIN
);
2604 tcp_queue_skb(sk
, skb
);
2606 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_OFF
);
2609 /* We get here when a process closes a file descriptor (either due to
2610 * an explicit close() or as a byproduct of exit()'ing) and there
2611 * was unread data in the receive queue. This behavior is recommended
2612 * by RFC 2525, section 2.17. -DaveM
2614 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
2616 struct sk_buff
*skb
;
2618 /* NOTE: No TCP options attached and we never retransmit this. */
2619 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
2621 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
2625 /* Reserve space for headers and prepare control bits. */
2626 skb_reserve(skb
, MAX_TCP_HEADER
);
2627 tcp_init_nondata_skb(skb
, tcp_acceptable_seq(sk
),
2628 TCPHDR_ACK
| TCPHDR_RST
);
2630 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2631 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
2632 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
2634 TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTRSTS
);
2637 /* Send a crossed SYN-ACK during socket establishment.
2638 * WARNING: This routine must only be called when we have already sent
2639 * a SYN packet that crossed the incoming SYN that caused this routine
2640 * to get called. If this assumption fails then the initial rcv_wnd
2641 * and rcv_wscale values will not be correct.
2643 int tcp_send_synack(struct sock
*sk
)
2645 struct sk_buff
*skb
;
2647 skb
= tcp_write_queue_head(sk
);
2648 if (skb
== NULL
|| !(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)) {
2649 pr_debug("%s: wrong queue state\n", __func__
);
2652 if (!(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_ACK
)) {
2653 if (skb_cloned(skb
)) {
2654 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
2657 tcp_unlink_write_queue(skb
, sk
);
2658 skb_header_release(nskb
);
2659 __tcp_add_write_queue_head(sk
, nskb
);
2660 sk_wmem_free_skb(sk
, skb
);
2661 sk
->sk_wmem_queued
+= nskb
->truesize
;
2662 sk_mem_charge(sk
, nskb
->truesize
);
2666 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_ACK
;
2667 TCP_ECN_send_synack(tcp_sk(sk
), skb
);
2669 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2670 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2674 * tcp_make_synack - Prepare a SYN-ACK.
2675 * sk: listener socket
2676 * dst: dst entry attached to the SYNACK
2677 * req: request_sock pointer
2679 * Allocate one skb and build a SYNACK packet.
2680 * @dst is consumed : Caller should not use it again.
2682 struct sk_buff
*tcp_make_synack(struct sock
*sk
, struct dst_entry
*dst
,
2683 struct request_sock
*req
,
2684 struct tcp_fastopen_cookie
*foc
)
2686 struct tcp_out_options opts
;
2687 struct inet_request_sock
*ireq
= inet_rsk(req
);
2688 struct tcp_sock
*tp
= tcp_sk(sk
);
2690 struct sk_buff
*skb
;
2691 struct tcp_md5sig_key
*md5
;
2692 int tcp_header_size
;
2695 skb
= sock_wmalloc(sk
, MAX_TCP_HEADER
+ 15, 1, GFP_ATOMIC
);
2696 if (unlikely(!skb
)) {
2700 /* Reserve space for headers. */
2701 skb_reserve(skb
, MAX_TCP_HEADER
);
2703 skb_dst_set(skb
, dst
);
2704 security_skb_owned_by(skb
, sk
);
2706 mss
= dst_metric_advmss(dst
);
2707 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< mss
)
2708 mss
= tp
->rx_opt
.user_mss
;
2710 if (req
->rcv_wnd
== 0) { /* ignored for retransmitted syns */
2712 /* Set this up on the first call only */
2713 req
->window_clamp
= tp
->window_clamp
? : dst_metric(dst
, RTAX_WINDOW
);
2715 /* limit the window selection if the user enforce a smaller rx buffer */
2716 if (sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
&&
2717 (req
->window_clamp
> tcp_full_space(sk
) || req
->window_clamp
== 0))
2718 req
->window_clamp
= tcp_full_space(sk
);
2720 /* tcp_full_space because it is guaranteed to be the first packet */
2721 tcp_select_initial_window(tcp_full_space(sk
),
2722 mss
- (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0),
2727 dst_metric(dst
, RTAX_INITRWND
));
2728 ireq
->rcv_wscale
= rcv_wscale
;
2731 memset(&opts
, 0, sizeof(opts
));
2732 #ifdef CONFIG_SYN_COOKIES
2733 if (unlikely(req
->cookie_ts
))
2734 TCP_SKB_CB(skb
)->when
= cookie_init_timestamp(req
);
2737 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2738 tcp_header_size
= tcp_synack_options(sk
, req
, mss
, skb
, &opts
, &md5
,
2741 skb_push(skb
, tcp_header_size
);
2742 skb_reset_transport_header(skb
);
2745 memset(th
, 0, sizeof(struct tcphdr
));
2748 TCP_ECN_make_synack(req
, th
);
2749 th
->source
= htons(ireq
->ir_num
);
2750 th
->dest
= ireq
->ir_rmt_port
;
2751 /* Setting of flags are superfluous here for callers (and ECE is
2752 * not even correctly set)
2754 tcp_init_nondata_skb(skb
, tcp_rsk(req
)->snt_isn
,
2755 TCPHDR_SYN
| TCPHDR_ACK
);
2757 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
2758 /* XXX data is queued and acked as is. No buffer/window check */
2759 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_nxt
);
2761 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2762 th
->window
= htons(min(req
->rcv_wnd
, 65535U));
2763 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
);
2764 th
->doff
= (tcp_header_size
>> 2);
2765 TCP_ADD_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
, tcp_skb_pcount(skb
));
2767 #ifdef CONFIG_TCP_MD5SIG
2768 /* Okay, we have all we need - do the md5 hash if needed */
2770 tcp_rsk(req
)->af_specific
->calc_md5_hash(opts
.hash_location
,
2771 md5
, NULL
, req
, skb
);
2777 EXPORT_SYMBOL(tcp_make_synack
);
2779 /* Do all connect socket setups that can be done AF independent. */
2780 static void tcp_connect_init(struct sock
*sk
)
2782 const struct dst_entry
*dst
= __sk_dst_get(sk
);
2783 struct tcp_sock
*tp
= tcp_sk(sk
);
2786 /* We'll fix this up when we get a response from the other end.
2787 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2789 tp
->tcp_header_len
= sizeof(struct tcphdr
) +
2790 (sysctl_tcp_timestamps
? TCPOLEN_TSTAMP_ALIGNED
: 0);
2792 #ifdef CONFIG_TCP_MD5SIG
2793 if (tp
->af_specific
->md5_lookup(sk
, sk
) != NULL
)
2794 tp
->tcp_header_len
+= TCPOLEN_MD5SIG_ALIGNED
;
2797 /* If user gave his TCP_MAXSEG, record it to clamp */
2798 if (tp
->rx_opt
.user_mss
)
2799 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
2802 tcp_sync_mss(sk
, dst_mtu(dst
));
2804 if (!tp
->window_clamp
)
2805 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
2806 tp
->advmss
= dst_metric_advmss(dst
);
2807 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< tp
->advmss
)
2808 tp
->advmss
= tp
->rx_opt
.user_mss
;
2810 tcp_initialize_rcv_mss(sk
);
2812 /* limit the window selection if the user enforce a smaller rx buffer */
2813 if (sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
&&
2814 (tp
->window_clamp
> tcp_full_space(sk
) || tp
->window_clamp
== 0))
2815 tp
->window_clamp
= tcp_full_space(sk
);
2817 tcp_select_initial_window(tcp_full_space(sk
),
2818 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
2821 sysctl_tcp_window_scaling
,
2823 dst_metric(dst
, RTAX_INITRWND
));
2825 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
2826 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
2829 sock_reset_flag(sk
, SOCK_DONE
);
2832 tp
->snd_una
= tp
->write_seq
;
2833 tp
->snd_sml
= tp
->write_seq
;
2834 tp
->snd_up
= tp
->write_seq
;
2835 tp
->snd_nxt
= tp
->write_seq
;
2837 if (likely(!tp
->repair
))
2840 tp
->rcv_tstamp
= tcp_time_stamp
;
2841 tp
->rcv_wup
= tp
->rcv_nxt
;
2842 tp
->copied_seq
= tp
->rcv_nxt
;
2844 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
2845 inet_csk(sk
)->icsk_retransmits
= 0;
2846 tcp_clear_retrans(tp
);
2849 static void tcp_connect_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
2851 struct tcp_sock
*tp
= tcp_sk(sk
);
2852 struct tcp_skb_cb
*tcb
= TCP_SKB_CB(skb
);
2854 tcb
->end_seq
+= skb
->len
;
2855 skb_header_release(skb
);
2856 __tcp_add_write_queue_tail(sk
, skb
);
2857 sk
->sk_wmem_queued
+= skb
->truesize
;
2858 sk_mem_charge(sk
, skb
->truesize
);
2859 tp
->write_seq
= tcb
->end_seq
;
2860 tp
->packets_out
+= tcp_skb_pcount(skb
);
2863 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
2864 * queue a data-only packet after the regular SYN, such that regular SYNs
2865 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
2866 * only the SYN sequence, the data are retransmitted in the first ACK.
2867 * If cookie is not cached or other error occurs, falls back to send a
2868 * regular SYN with Fast Open cookie request option.
2870 static int tcp_send_syn_data(struct sock
*sk
, struct sk_buff
*syn
)
2872 struct tcp_sock
*tp
= tcp_sk(sk
);
2873 struct tcp_fastopen_request
*fo
= tp
->fastopen_req
;
2874 int syn_loss
= 0, space
, i
, err
= 0, iovlen
= fo
->data
->msg_iovlen
;
2875 struct sk_buff
*syn_data
= NULL
, *data
;
2876 unsigned long last_syn_loss
= 0;
2878 tp
->rx_opt
.mss_clamp
= tp
->advmss
; /* If MSS is not cached */
2879 tcp_fastopen_cache_get(sk
, &tp
->rx_opt
.mss_clamp
, &fo
->cookie
,
2880 &syn_loss
, &last_syn_loss
);
2881 /* Recurring FO SYN losses: revert to regular handshake temporarily */
2883 time_before(jiffies
, last_syn_loss
+ (60*HZ
<< syn_loss
))) {
2884 fo
->cookie
.len
= -1;
2888 if (sysctl_tcp_fastopen
& TFO_CLIENT_NO_COOKIE
)
2889 fo
->cookie
.len
= -1;
2890 else if (fo
->cookie
.len
<= 0)
2893 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
2894 * user-MSS. Reserve maximum option space for middleboxes that add
2895 * private TCP options. The cost is reduced data space in SYN :(
2897 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< tp
->rx_opt
.mss_clamp
)
2898 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
2899 space
= __tcp_mtu_to_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
) -
2900 MAX_TCP_OPTION_SPACE
;
2902 syn_data
= skb_copy_expand(syn
, skb_headroom(syn
), space
,
2904 if (syn_data
== NULL
)
2907 for (i
= 0; i
< iovlen
&& syn_data
->len
< space
; ++i
) {
2908 struct iovec
*iov
= &fo
->data
->msg_iov
[i
];
2909 unsigned char __user
*from
= iov
->iov_base
;
2910 int len
= iov
->iov_len
;
2912 if (syn_data
->len
+ len
> space
)
2913 len
= space
- syn_data
->len
;
2914 else if (i
+ 1 == iovlen
)
2915 /* No more data pending in inet_wait_for_connect() */
2918 if (skb_add_data(syn_data
, from
, len
))
2922 /* Queue a data-only packet after the regular SYN for retransmission */
2923 data
= pskb_copy(syn_data
, sk
->sk_allocation
);
2926 TCP_SKB_CB(data
)->seq
++;
2927 TCP_SKB_CB(data
)->tcp_flags
&= ~TCPHDR_SYN
;
2928 TCP_SKB_CB(data
)->tcp_flags
= (TCPHDR_ACK
|TCPHDR_PSH
);
2929 tcp_connect_queue_skb(sk
, data
);
2930 fo
->copied
= data
->len
;
2932 if (tcp_transmit_skb(sk
, syn_data
, 0, sk
->sk_allocation
) == 0) {
2933 tp
->syn_data
= (fo
->copied
> 0);
2934 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPFASTOPENACTIVE
);
2940 /* Send a regular SYN with Fast Open cookie request option */
2941 if (fo
->cookie
.len
> 0)
2943 err
= tcp_transmit_skb(sk
, syn
, 1, sk
->sk_allocation
);
2945 tp
->syn_fastopen
= 0;
2946 kfree_skb(syn_data
);
2948 fo
->cookie
.len
= -1; /* Exclude Fast Open option for SYN retries */
2952 /* Build a SYN and send it off. */
2953 int tcp_connect(struct sock
*sk
)
2955 struct tcp_sock
*tp
= tcp_sk(sk
);
2956 struct sk_buff
*buff
;
2959 tcp_connect_init(sk
);
2961 if (unlikely(tp
->repair
)) {
2962 tcp_finish_connect(sk
, NULL
);
2966 buff
= alloc_skb_fclone(MAX_TCP_HEADER
+ 15, sk
->sk_allocation
);
2967 if (unlikely(buff
== NULL
))
2970 /* Reserve space for headers. */
2971 skb_reserve(buff
, MAX_TCP_HEADER
);
2973 tcp_init_nondata_skb(buff
, tp
->write_seq
++, TCPHDR_SYN
);
2974 tp
->retrans_stamp
= TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2975 tcp_connect_queue_skb(sk
, buff
);
2976 TCP_ECN_send_syn(sk
, buff
);
2978 /* Send off SYN; include data in Fast Open. */
2979 err
= tp
->fastopen_req
? tcp_send_syn_data(sk
, buff
) :
2980 tcp_transmit_skb(sk
, buff
, 1, sk
->sk_allocation
);
2981 if (err
== -ECONNREFUSED
)
2984 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2985 * in order to make this packet get counted in tcpOutSegs.
2987 tp
->snd_nxt
= tp
->write_seq
;
2988 tp
->pushed_seq
= tp
->write_seq
;
2989 TCP_INC_STATS(sock_net(sk
), TCP_MIB_ACTIVEOPENS
);
2991 /* Timer for repeating the SYN until an answer. */
2992 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2993 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2996 EXPORT_SYMBOL(tcp_connect
);
2998 /* Send out a delayed ack, the caller does the policy checking
2999 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3002 void tcp_send_delayed_ack(struct sock
*sk
)
3004 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3005 int ato
= icsk
->icsk_ack
.ato
;
3006 unsigned long timeout
;
3008 if (ato
> TCP_DELACK_MIN
) {
3009 const struct tcp_sock
*tp
= tcp_sk(sk
);
3010 int max_ato
= HZ
/ 2;
3012 if (icsk
->icsk_ack
.pingpong
||
3013 (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
3014 max_ato
= TCP_DELACK_MAX
;
3016 /* Slow path, intersegment interval is "high". */
3018 /* If some rtt estimate is known, use it to bound delayed ack.
3019 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3023 int rtt
= max(tp
->srtt
>> 3, TCP_DELACK_MIN
);
3029 ato
= min(ato
, max_ato
);
3032 /* Stay within the limit we were given */
3033 timeout
= jiffies
+ ato
;
3035 /* Use new timeout only if there wasn't a older one earlier. */
3036 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
3037 /* If delack timer was blocked or is about to expire,
3040 if (icsk
->icsk_ack
.blocked
||
3041 time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
3046 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
3047 timeout
= icsk
->icsk_ack
.timeout
;
3049 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
3050 icsk
->icsk_ack
.timeout
= timeout
;
3051 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
3054 /* This routine sends an ack and also updates the window. */
3055 void tcp_send_ack(struct sock
*sk
)
3057 struct sk_buff
*buff
;
3059 /* If we have been reset, we may not send again. */
3060 if (sk
->sk_state
== TCP_CLOSE
)
3063 /* We are not putting this on the write queue, so
3064 * tcp_transmit_skb() will set the ownership to this
3067 buff
= alloc_skb(MAX_TCP_HEADER
, sk_gfp_atomic(sk
, GFP_ATOMIC
));
3069 inet_csk_schedule_ack(sk
);
3070 inet_csk(sk
)->icsk_ack
.ato
= TCP_ATO_MIN
;
3071 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
3072 TCP_DELACK_MAX
, TCP_RTO_MAX
);
3076 /* Reserve space for headers and prepare control bits. */
3077 skb_reserve(buff
, MAX_TCP_HEADER
);
3078 tcp_init_nondata_skb(buff
, tcp_acceptable_seq(sk
), TCPHDR_ACK
);
3080 /* Send it off, this clears delayed acks for us. */
3081 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
3082 tcp_transmit_skb(sk
, buff
, 0, sk_gfp_atomic(sk
, GFP_ATOMIC
));
3085 /* This routine sends a packet with an out of date sequence
3086 * number. It assumes the other end will try to ack it.
3088 * Question: what should we make while urgent mode?
3089 * 4.4BSD forces sending single byte of data. We cannot send
3090 * out of window data, because we have SND.NXT==SND.MAX...
3092 * Current solution: to send TWO zero-length segments in urgent mode:
3093 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3094 * out-of-date with SND.UNA-1 to probe window.
3096 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
)
3098 struct tcp_sock
*tp
= tcp_sk(sk
);
3099 struct sk_buff
*skb
;
3101 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3102 skb
= alloc_skb(MAX_TCP_HEADER
, sk_gfp_atomic(sk
, GFP_ATOMIC
));
3106 /* Reserve space for headers and set control bits. */
3107 skb_reserve(skb
, MAX_TCP_HEADER
);
3108 /* Use a previous sequence. This should cause the other
3109 * end to send an ack. Don't queue or clone SKB, just
3112 tcp_init_nondata_skb(skb
, tp
->snd_una
- !urgent
, TCPHDR_ACK
);
3113 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
3114 return tcp_transmit_skb(sk
, skb
, 0, GFP_ATOMIC
);
3117 void tcp_send_window_probe(struct sock
*sk
)
3119 if (sk
->sk_state
== TCP_ESTABLISHED
) {
3120 tcp_sk(sk
)->snd_wl1
= tcp_sk(sk
)->rcv_nxt
- 1;
3121 tcp_xmit_probe_skb(sk
, 0);
3125 /* Initiate keepalive or window probe from timer. */
3126 int tcp_write_wakeup(struct sock
*sk
)
3128 struct tcp_sock
*tp
= tcp_sk(sk
);
3129 struct sk_buff
*skb
;
3131 if (sk
->sk_state
== TCP_CLOSE
)
3134 if ((skb
= tcp_send_head(sk
)) != NULL
&&
3135 before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
))) {
3137 unsigned int mss
= tcp_current_mss(sk
);
3138 unsigned int seg_size
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
3140 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
3141 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
3143 /* We are probing the opening of a window
3144 * but the window size is != 0
3145 * must have been a result SWS avoidance ( sender )
3147 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
3149 seg_size
= min(seg_size
, mss
);
3150 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
3151 if (tcp_fragment(sk
, skb
, seg_size
, mss
))
3153 } else if (!tcp_skb_pcount(skb
))
3154 tcp_set_skb_tso_segs(sk
, skb
, mss
);
3156 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
3157 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
3158 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
3160 tcp_event_new_data_sent(sk
, skb
);
3163 if (between(tp
->snd_up
, tp
->snd_una
+ 1, tp
->snd_una
+ 0xFFFF))
3164 tcp_xmit_probe_skb(sk
, 1);
3165 return tcp_xmit_probe_skb(sk
, 0);
3169 /* A window probe timeout has occurred. If window is not closed send
3170 * a partial packet else a zero probe.
3172 void tcp_send_probe0(struct sock
*sk
)
3174 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3175 struct tcp_sock
*tp
= tcp_sk(sk
);
3178 err
= tcp_write_wakeup(sk
);
3180 if (tp
->packets_out
|| !tcp_send_head(sk
)) {
3181 /* Cancel probe timer, if it is not required. */
3182 icsk
->icsk_probes_out
= 0;
3183 icsk
->icsk_backoff
= 0;
3188 if (icsk
->icsk_backoff
< sysctl_tcp_retries2
)
3189 icsk
->icsk_backoff
++;
3190 icsk
->icsk_probes_out
++;
3191 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3192 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
3195 /* If packet was not sent due to local congestion,
3196 * do not backoff and do not remember icsk_probes_out.
3197 * Let local senders to fight for local resources.
3199 * Use accumulated backoff yet.
3201 if (!icsk
->icsk_probes_out
)
3202 icsk
->icsk_probes_out
= 1;
3203 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3204 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
,
3205 TCP_RESOURCE_PROBE_INTERVAL
),