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
39 #include <linux/compiler.h>
40 #include <linux/module.h>
42 /* People can turn this off for buggy TCP's found in printers etc. */
43 int sysctl_tcp_retrans_collapse __read_mostly
= 1;
45 /* People can turn this on to work with those rare, broken TCPs that
46 * interpret the window field as a signed quantity.
48 int sysctl_tcp_workaround_signed_windows __read_mostly
= 0;
50 /* This limits the percentage of the congestion window which we
51 * will allow a single TSO frame to consume. Building TSO frames
52 * which are too large can cause TCP streams to be bursty.
54 int sysctl_tcp_tso_win_divisor __read_mostly
= 3;
56 int sysctl_tcp_mtu_probing __read_mostly
= 0;
57 int sysctl_tcp_base_mss __read_mostly
= 512;
59 /* By default, RFC2861 behavior. */
60 int sysctl_tcp_slow_start_after_idle __read_mostly
= 1;
62 static void tcp_event_new_data_sent(struct sock
*sk
, struct sk_buff
*skb
)
64 struct tcp_sock
*tp
= tcp_sk(sk
);
65 unsigned int prior_packets
= tp
->packets_out
;
67 tcp_advance_send_head(sk
, skb
);
68 tp
->snd_nxt
= TCP_SKB_CB(skb
)->end_seq
;
70 /* Don't override Nagle indefinately with F-RTO */
71 if (tp
->frto_counter
== 2)
74 tp
->packets_out
+= tcp_skb_pcount(skb
);
76 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
77 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
80 /* SND.NXT, if window was not shrunk.
81 * If window has been shrunk, what should we make? It is not clear at all.
82 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
83 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
84 * invalid. OK, let's make this for now:
86 static inline __u32
tcp_acceptable_seq(struct sock
*sk
)
88 struct tcp_sock
*tp
= tcp_sk(sk
);
90 if (!before(tcp_wnd_end(tp
), tp
->snd_nxt
))
93 return tcp_wnd_end(tp
);
96 /* Calculate mss to advertise in SYN segment.
97 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
99 * 1. It is independent of path mtu.
100 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
101 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
102 * attached devices, because some buggy hosts are confused by
104 * 4. We do not make 3, we advertise MSS, calculated from first
105 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
106 * This may be overridden via information stored in routing table.
107 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
108 * probably even Jumbo".
110 static __u16
tcp_advertise_mss(struct sock
*sk
)
112 struct tcp_sock
*tp
= tcp_sk(sk
);
113 struct dst_entry
*dst
= __sk_dst_get(sk
);
114 int mss
= tp
->advmss
;
116 if (dst
&& dst_metric(dst
, RTAX_ADVMSS
) < mss
) {
117 mss
= dst_metric(dst
, RTAX_ADVMSS
);
124 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
125 * This is the first part of cwnd validation mechanism. */
126 static void tcp_cwnd_restart(struct sock
*sk
, struct dst_entry
*dst
)
128 struct tcp_sock
*tp
= tcp_sk(sk
);
129 s32 delta
= tcp_time_stamp
- tp
->lsndtime
;
130 u32 restart_cwnd
= tcp_init_cwnd(tp
, dst
);
131 u32 cwnd
= tp
->snd_cwnd
;
133 tcp_ca_event(sk
, CA_EVENT_CWND_RESTART
);
135 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
136 restart_cwnd
= min(restart_cwnd
, cwnd
);
138 while ((delta
-= inet_csk(sk
)->icsk_rto
) > 0 && cwnd
> restart_cwnd
)
140 tp
->snd_cwnd
= max(cwnd
, restart_cwnd
);
141 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
142 tp
->snd_cwnd_used
= 0;
145 static void tcp_event_data_sent(struct tcp_sock
*tp
,
146 struct sk_buff
*skb
, struct sock
*sk
)
148 struct inet_connection_sock
*icsk
= inet_csk(sk
);
149 const u32 now
= tcp_time_stamp
;
151 if (sysctl_tcp_slow_start_after_idle
&&
152 (!tp
->packets_out
&& (s32
)(now
- tp
->lsndtime
) > icsk
->icsk_rto
))
153 tcp_cwnd_restart(sk
, __sk_dst_get(sk
));
157 /* If it is a reply for ato after last received
158 * packet, enter pingpong mode.
160 if ((u32
)(now
- icsk
->icsk_ack
.lrcvtime
) < icsk
->icsk_ack
.ato
)
161 icsk
->icsk_ack
.pingpong
= 1;
164 static inline void tcp_event_ack_sent(struct sock
*sk
, unsigned int pkts
)
166 tcp_dec_quickack_mode(sk
, pkts
);
167 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_DACK
);
170 /* Determine a window scaling and initial window to offer.
171 * Based on the assumption that the given amount of space
172 * will be offered. Store the results in the tp structure.
173 * NOTE: for smooth operation initial space offering should
174 * be a multiple of mss if possible. We assume here that mss >= 1.
175 * This MUST be enforced by all callers.
177 void tcp_select_initial_window(int __space
, __u32 mss
,
178 __u32
*rcv_wnd
, __u32
*window_clamp
,
179 int wscale_ok
, __u8
*rcv_wscale
)
181 unsigned int space
= (__space
< 0 ? 0 : __space
);
183 /* If no clamp set the clamp to the max possible scaled window */
184 if (*window_clamp
== 0)
185 (*window_clamp
) = (65535 << 14);
186 space
= min(*window_clamp
, space
);
188 /* Quantize space offering to a multiple of mss if possible. */
190 space
= (space
/ mss
) * mss
;
192 /* NOTE: offering an initial window larger than 32767
193 * will break some buggy TCP stacks. If the admin tells us
194 * it is likely we could be speaking with such a buggy stack
195 * we will truncate our initial window offering to 32K-1
196 * unless the remote has sent us a window scaling option,
197 * which we interpret as a sign the remote TCP is not
198 * misinterpreting the window field as a signed quantity.
200 if (sysctl_tcp_workaround_signed_windows
)
201 (*rcv_wnd
) = min(space
, MAX_TCP_WINDOW
);
207 /* Set window scaling on max possible window
208 * See RFC1323 for an explanation of the limit to 14
210 space
= max_t(u32
, sysctl_tcp_rmem
[2], sysctl_rmem_max
);
211 space
= min_t(u32
, space
, *window_clamp
);
212 while (space
> 65535 && (*rcv_wscale
) < 14) {
218 /* Set initial window to value enough for senders,
219 * following RFC2414. Senders, not following this RFC,
220 * will be satisfied with 2.
222 if (mss
> (1 << *rcv_wscale
)) {
228 if (*rcv_wnd
> init_cwnd
* mss
)
229 *rcv_wnd
= init_cwnd
* mss
;
232 /* Set the clamp no higher than max representable value */
233 (*window_clamp
) = min(65535U << (*rcv_wscale
), *window_clamp
);
236 /* Chose a new window to advertise, update state in tcp_sock for the
237 * socket, and return result with RFC1323 scaling applied. The return
238 * value can be stuffed directly into th->window for an outgoing
241 static u16
tcp_select_window(struct sock
*sk
)
243 struct tcp_sock
*tp
= tcp_sk(sk
);
244 u32 cur_win
= tcp_receive_window(tp
);
245 u32 new_win
= __tcp_select_window(sk
);
247 /* Never shrink the offered window */
248 if (new_win
< cur_win
) {
249 /* Danger Will Robinson!
250 * Don't update rcv_wup/rcv_wnd here or else
251 * we will not be able to advertise a zero
252 * window in time. --DaveM
254 * Relax Will Robinson.
256 new_win
= ALIGN(cur_win
, 1 << tp
->rx_opt
.rcv_wscale
);
258 tp
->rcv_wnd
= new_win
;
259 tp
->rcv_wup
= tp
->rcv_nxt
;
261 /* Make sure we do not exceed the maximum possible
264 if (!tp
->rx_opt
.rcv_wscale
&& sysctl_tcp_workaround_signed_windows
)
265 new_win
= min(new_win
, MAX_TCP_WINDOW
);
267 new_win
= min(new_win
, (65535U << tp
->rx_opt
.rcv_wscale
));
269 /* RFC1323 scaling applied */
270 new_win
>>= tp
->rx_opt
.rcv_wscale
;
272 /* If we advertise zero window, disable fast path. */
279 static inline void TCP_ECN_send_synack(struct tcp_sock
*tp
, struct sk_buff
*skb
)
281 TCP_SKB_CB(skb
)->flags
&= ~TCPCB_FLAG_CWR
;
282 if (!(tp
->ecn_flags
& TCP_ECN_OK
))
283 TCP_SKB_CB(skb
)->flags
&= ~TCPCB_FLAG_ECE
;
286 static inline void TCP_ECN_send_syn(struct sock
*sk
, struct sk_buff
*skb
)
288 struct tcp_sock
*tp
= tcp_sk(sk
);
291 if (sysctl_tcp_ecn
) {
292 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_ECE
| TCPCB_FLAG_CWR
;
293 tp
->ecn_flags
= TCP_ECN_OK
;
297 static __inline__
void
298 TCP_ECN_make_synack(struct request_sock
*req
, struct tcphdr
*th
)
300 if (inet_rsk(req
)->ecn_ok
)
304 static inline void TCP_ECN_send(struct sock
*sk
, struct sk_buff
*skb
,
307 struct tcp_sock
*tp
= tcp_sk(sk
);
309 if (tp
->ecn_flags
& TCP_ECN_OK
) {
310 /* Not-retransmitted data segment: set ECT and inject CWR. */
311 if (skb
->len
!= tcp_header_len
&&
312 !before(TCP_SKB_CB(skb
)->seq
, tp
->snd_nxt
)) {
314 if (tp
->ecn_flags
& TCP_ECN_QUEUE_CWR
) {
315 tp
->ecn_flags
&= ~TCP_ECN_QUEUE_CWR
;
316 tcp_hdr(skb
)->cwr
= 1;
317 skb_shinfo(skb
)->gso_type
|= SKB_GSO_TCP_ECN
;
320 /* ACK or retransmitted segment: clear ECT|CE */
321 INET_ECN_dontxmit(sk
);
323 if (tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)
324 tcp_hdr(skb
)->ece
= 1;
328 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
329 * auto increment end seqno.
331 static void tcp_init_nondata_skb(struct sk_buff
*skb
, u32 seq
, u8 flags
)
335 TCP_SKB_CB(skb
)->flags
= flags
;
336 TCP_SKB_CB(skb
)->sacked
= 0;
338 skb_shinfo(skb
)->gso_segs
= 1;
339 skb_shinfo(skb
)->gso_size
= 0;
340 skb_shinfo(skb
)->gso_type
= 0;
342 TCP_SKB_CB(skb
)->seq
= seq
;
343 if (flags
& (TCPCB_FLAG_SYN
| TCPCB_FLAG_FIN
))
345 TCP_SKB_CB(skb
)->end_seq
= seq
;
348 #define OPTION_SACK_ADVERTISE (1 << 0)
349 #define OPTION_TS (1 << 1)
350 #define OPTION_MD5 (1 << 2)
352 struct tcp_out_options
{
353 u8 options
; /* bit field of OPTION_* */
354 u8 ws
; /* window scale, 0 to disable */
355 u8 num_sack_blocks
; /* number of SACK blocks to include */
356 u16 mss
; /* 0 to disable */
357 __u32 tsval
, tsecr
; /* need to include OPTION_TS */
360 /* Beware: Something in the Internet is very sensitive to the ordering of
361 * TCP options, we learned this through the hard way, so be careful here.
362 * Luckily we can at least blame others for their non-compliance but from
363 * inter-operatibility perspective it seems that we're somewhat stuck with
364 * the ordering which we have been using if we want to keep working with
365 * those broken things (not that it currently hurts anybody as there isn't
366 * particular reason why the ordering would need to be changed).
368 * At least SACK_PERM as the first option is known to lead to a disaster
369 * (but it may well be that other scenarios fail similarly).
371 static void tcp_options_write(__be32
*ptr
, struct tcp_sock
*tp
,
372 const struct tcp_out_options
*opts
,
374 if (unlikely(OPTION_MD5
& opts
->options
)) {
375 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
377 (TCPOPT_MD5SIG
<< 8) |
379 *md5_hash
= (__u8
*)ptr
;
385 if (unlikely(opts
->mss
)) {
386 *ptr
++ = htonl((TCPOPT_MSS
<< 24) |
387 (TCPOLEN_MSS
<< 16) |
391 if (likely(OPTION_TS
& opts
->options
)) {
392 if (unlikely(OPTION_SACK_ADVERTISE
& opts
->options
)) {
393 *ptr
++ = htonl((TCPOPT_SACK_PERM
<< 24) |
394 (TCPOLEN_SACK_PERM
<< 16) |
395 (TCPOPT_TIMESTAMP
<< 8) |
398 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
400 (TCPOPT_TIMESTAMP
<< 8) |
403 *ptr
++ = htonl(opts
->tsval
);
404 *ptr
++ = htonl(opts
->tsecr
);
407 if (unlikely(OPTION_SACK_ADVERTISE
& opts
->options
&&
408 !(OPTION_TS
& opts
->options
))) {
409 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
411 (TCPOPT_SACK_PERM
<< 8) |
415 if (unlikely(opts
->ws
)) {
416 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
417 (TCPOPT_WINDOW
<< 16) |
418 (TCPOLEN_WINDOW
<< 8) |
422 if (unlikely(opts
->num_sack_blocks
)) {
423 struct tcp_sack_block
*sp
= tp
->rx_opt
.dsack
?
424 tp
->duplicate_sack
: tp
->selective_acks
;
427 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
430 (TCPOLEN_SACK_BASE
+ (opts
->num_sack_blocks
*
431 TCPOLEN_SACK_PERBLOCK
)));
433 for (this_sack
= 0; this_sack
< opts
->num_sack_blocks
;
435 *ptr
++ = htonl(sp
[this_sack
].start_seq
);
436 *ptr
++ = htonl(sp
[this_sack
].end_seq
);
439 if (tp
->rx_opt
.dsack
) {
440 tp
->rx_opt
.dsack
= 0;
441 tp
->rx_opt
.eff_sacks
--;
446 static unsigned tcp_syn_options(struct sock
*sk
, struct sk_buff
*skb
,
447 struct tcp_out_options
*opts
,
448 struct tcp_md5sig_key
**md5
) {
449 struct tcp_sock
*tp
= tcp_sk(sk
);
452 #ifdef CONFIG_TCP_MD5SIG
453 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
455 opts
->options
|= OPTION_MD5
;
456 size
+= TCPOLEN_MD5SIG_ALIGNED
;
462 /* We always get an MSS option. The option bytes which will be seen in
463 * normal data packets should timestamps be used, must be in the MSS
464 * advertised. But we subtract them from tp->mss_cache so that
465 * calculations in tcp_sendmsg are simpler etc. So account for this
466 * fact here if necessary. If we don't do this correctly, as a
467 * receiver we won't recognize data packets as being full sized when we
468 * should, and thus we won't abide by the delayed ACK rules correctly.
469 * SACKs don't matter, we never delay an ACK when we have any of those
471 opts
->mss
= tcp_advertise_mss(sk
);
472 size
+= TCPOLEN_MSS_ALIGNED
;
474 if (likely(sysctl_tcp_timestamps
&& *md5
== NULL
)) {
475 opts
->options
|= OPTION_TS
;
476 opts
->tsval
= TCP_SKB_CB(skb
)->when
;
477 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
478 size
+= TCPOLEN_TSTAMP_ALIGNED
;
480 if (likely(sysctl_tcp_window_scaling
)) {
481 opts
->ws
= tp
->rx_opt
.rcv_wscale
;
483 size
+= TCPOLEN_WSCALE_ALIGNED
;
485 if (likely(sysctl_tcp_sack
)) {
486 opts
->options
|= OPTION_SACK_ADVERTISE
;
487 if (unlikely(!(OPTION_TS
& opts
->options
)))
488 size
+= TCPOLEN_SACKPERM_ALIGNED
;
494 static unsigned tcp_synack_options(struct sock
*sk
,
495 struct request_sock
*req
,
496 unsigned mss
, struct sk_buff
*skb
,
497 struct tcp_out_options
*opts
,
498 struct tcp_md5sig_key
**md5
) {
500 struct inet_request_sock
*ireq
= inet_rsk(req
);
503 #ifdef CONFIG_TCP_MD5SIG
504 *md5
= tcp_rsk(req
)->af_specific
->md5_lookup(sk
, req
);
506 opts
->options
|= OPTION_MD5
;
507 size
+= TCPOLEN_MD5SIG_ALIGNED
;
513 /* we can't fit any SACK blocks in a packet with MD5 + TS
514 options. There was discussion about disabling SACK rather than TS in
515 order to fit in better with old, buggy kernels, but that was deemed
516 to be unnecessary. */
517 doing_ts
= ireq
->tstamp_ok
&& !(*md5
&& ireq
->sack_ok
);
520 size
+= TCPOLEN_MSS_ALIGNED
;
522 if (likely(ireq
->wscale_ok
)) {
523 opts
->ws
= ireq
->rcv_wscale
;
525 size
+= TCPOLEN_WSCALE_ALIGNED
;
527 if (likely(doing_ts
)) {
528 opts
->options
|= OPTION_TS
;
529 opts
->tsval
= TCP_SKB_CB(skb
)->when
;
530 opts
->tsecr
= req
->ts_recent
;
531 size
+= TCPOLEN_TSTAMP_ALIGNED
;
533 if (likely(ireq
->sack_ok
)) {
534 opts
->options
|= OPTION_SACK_ADVERTISE
;
535 if (unlikely(!doing_ts
))
536 size
+= TCPOLEN_SACKPERM_ALIGNED
;
542 static unsigned tcp_established_options(struct sock
*sk
, struct sk_buff
*skb
,
543 struct tcp_out_options
*opts
,
544 struct tcp_md5sig_key
**md5
) {
545 struct tcp_skb_cb
*tcb
= skb
? TCP_SKB_CB(skb
) : NULL
;
546 struct tcp_sock
*tp
= tcp_sk(sk
);
549 #ifdef CONFIG_TCP_MD5SIG
550 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
551 if (unlikely(*md5
)) {
552 opts
->options
|= OPTION_MD5
;
553 size
+= TCPOLEN_MD5SIG_ALIGNED
;
559 if (likely(tp
->rx_opt
.tstamp_ok
)) {
560 opts
->options
|= OPTION_TS
;
561 opts
->tsval
= tcb
? tcb
->when
: 0;
562 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
563 size
+= TCPOLEN_TSTAMP_ALIGNED
;
566 if (unlikely(tp
->rx_opt
.eff_sacks
)) {
567 const unsigned remaining
= MAX_TCP_OPTION_SPACE
- size
;
568 opts
->num_sack_blocks
=
569 min_t(unsigned, tp
->rx_opt
.eff_sacks
,
570 (remaining
- TCPOLEN_SACK_BASE_ALIGNED
) /
571 TCPOLEN_SACK_PERBLOCK
);
572 size
+= TCPOLEN_SACK_BASE_ALIGNED
+
573 opts
->num_sack_blocks
* TCPOLEN_SACK_PERBLOCK
;
579 /* This routine actually transmits TCP packets queued in by
580 * tcp_do_sendmsg(). This is used by both the initial
581 * transmission and possible later retransmissions.
582 * All SKB's seen here are completely headerless. It is our
583 * job to build the TCP header, and pass the packet down to
584 * IP so it can do the same plus pass the packet off to the
587 * We are working here with either a clone of the original
588 * SKB, or a fresh unique copy made by the retransmit engine.
590 static int tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int clone_it
,
593 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
594 struct inet_sock
*inet
;
596 struct tcp_skb_cb
*tcb
;
597 struct tcp_out_options opts
;
598 unsigned tcp_options_size
, tcp_header_size
;
599 struct tcp_md5sig_key
*md5
;
600 __u8
*md5_hash_location
;
604 BUG_ON(!skb
|| !tcp_skb_pcount(skb
));
606 /* If congestion control is doing timestamping, we must
607 * take such a timestamp before we potentially clone/copy.
609 if (icsk
->icsk_ca_ops
->flags
& TCP_CONG_RTT_STAMP
)
610 __net_timestamp(skb
);
612 if (likely(clone_it
)) {
613 if (unlikely(skb_cloned(skb
)))
614 skb
= pskb_copy(skb
, gfp_mask
);
616 skb
= skb_clone(skb
, gfp_mask
);
623 tcb
= TCP_SKB_CB(skb
);
624 memset(&opts
, 0, sizeof(opts
));
626 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
))
627 tcp_options_size
= tcp_syn_options(sk
, skb
, &opts
, &md5
);
629 tcp_options_size
= tcp_established_options(sk
, skb
, &opts
,
631 tcp_header_size
= tcp_options_size
+ sizeof(struct tcphdr
);
633 if (tcp_packets_in_flight(tp
) == 0)
634 tcp_ca_event(sk
, CA_EVENT_TX_START
);
636 skb_push(skb
, tcp_header_size
);
637 skb_reset_transport_header(skb
);
638 skb_set_owner_w(skb
, sk
);
640 /* Build TCP header and checksum it. */
642 th
->source
= inet
->sport
;
643 th
->dest
= inet
->dport
;
644 th
->seq
= htonl(tcb
->seq
);
645 th
->ack_seq
= htonl(tp
->rcv_nxt
);
646 *(((__be16
*)th
) + 6) = htons(((tcp_header_size
>> 2) << 12) |
649 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
650 /* RFC1323: The window in SYN & SYN/ACK segments
653 th
->window
= htons(min(tp
->rcv_wnd
, 65535U));
655 th
->window
= htons(tcp_select_window(sk
));
660 if (unlikely(tp
->urg_mode
&&
661 between(tp
->snd_up
, tcb
->seq
+ 1, tcb
->seq
+ 0xFFFF))) {
662 th
->urg_ptr
= htons(tp
->snd_up
- tcb
->seq
);
666 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
, &md5_hash_location
);
667 if (likely((tcb
->flags
& TCPCB_FLAG_SYN
) == 0))
668 TCP_ECN_send(sk
, skb
, tcp_header_size
);
670 #ifdef CONFIG_TCP_MD5SIG
671 /* Calculate the MD5 hash, as we have all we need now */
673 sk
->sk_route_caps
&= ~NETIF_F_GSO_MASK
;
674 tp
->af_specific
->calc_md5_hash(md5_hash_location
,
679 icsk
->icsk_af_ops
->send_check(sk
, skb
->len
, skb
);
681 if (likely(tcb
->flags
& TCPCB_FLAG_ACK
))
682 tcp_event_ack_sent(sk
, tcp_skb_pcount(skb
));
684 if (skb
->len
!= tcp_header_size
)
685 tcp_event_data_sent(tp
, skb
, sk
);
687 if (after(tcb
->end_seq
, tp
->snd_nxt
) || tcb
->seq
== tcb
->end_seq
)
688 TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
);
690 err
= icsk
->icsk_af_ops
->queue_xmit(skb
, 0);
691 if (likely(err
<= 0))
694 tcp_enter_cwr(sk
, 1);
696 return net_xmit_eval(err
);
699 /* This routine just queue's the buffer
701 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
702 * otherwise socket can stall.
704 static void tcp_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
706 struct tcp_sock
*tp
= tcp_sk(sk
);
708 /* Advance write_seq and place onto the write_queue. */
709 tp
->write_seq
= TCP_SKB_CB(skb
)->end_seq
;
710 skb_header_release(skb
);
711 tcp_add_write_queue_tail(sk
, skb
);
712 sk
->sk_wmem_queued
+= skb
->truesize
;
713 sk_mem_charge(sk
, skb
->truesize
);
716 static void tcp_set_skb_tso_segs(struct sock
*sk
, struct sk_buff
*skb
,
717 unsigned int mss_now
)
719 if (skb
->len
<= mss_now
|| !sk_can_gso(sk
)) {
720 /* Avoid the costly divide in the normal
723 skb_shinfo(skb
)->gso_segs
= 1;
724 skb_shinfo(skb
)->gso_size
= 0;
725 skb_shinfo(skb
)->gso_type
= 0;
727 skb_shinfo(skb
)->gso_segs
= DIV_ROUND_UP(skb
->len
, mss_now
);
728 skb_shinfo(skb
)->gso_size
= mss_now
;
729 skb_shinfo(skb
)->gso_type
= sk
->sk_gso_type
;
733 /* When a modification to fackets out becomes necessary, we need to check
734 * skb is counted to fackets_out or not.
736 static void tcp_adjust_fackets_out(struct sock
*sk
, struct sk_buff
*skb
,
739 struct tcp_sock
*tp
= tcp_sk(sk
);
741 if (!tp
->sacked_out
|| tcp_is_reno(tp
))
744 if (after(tcp_highest_sack_seq(tp
), TCP_SKB_CB(skb
)->seq
))
745 tp
->fackets_out
-= decr
;
748 /* Function to create two new TCP segments. Shrinks the given segment
749 * to the specified size and appends a new segment with the rest of the
750 * packet to the list. This won't be called frequently, I hope.
751 * Remember, these are still headerless SKBs at this point.
753 int tcp_fragment(struct sock
*sk
, struct sk_buff
*skb
, u32 len
,
754 unsigned int mss_now
)
756 struct tcp_sock
*tp
= tcp_sk(sk
);
757 struct sk_buff
*buff
;
758 int nsize
, old_factor
;
762 BUG_ON(len
> skb
->len
);
764 tcp_clear_retrans_hints_partial(tp
);
765 nsize
= skb_headlen(skb
) - len
;
769 if (skb_cloned(skb
) &&
770 skb_is_nonlinear(skb
) &&
771 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
774 /* Get a new skb... force flag on. */
775 buff
= sk_stream_alloc_skb(sk
, nsize
, GFP_ATOMIC
);
777 return -ENOMEM
; /* We'll just try again later. */
779 sk
->sk_wmem_queued
+= buff
->truesize
;
780 sk_mem_charge(sk
, buff
->truesize
);
781 nlen
= skb
->len
- len
- nsize
;
782 buff
->truesize
+= nlen
;
783 skb
->truesize
-= nlen
;
785 /* Correct the sequence numbers. */
786 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
787 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
788 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
790 /* PSH and FIN should only be set in the second packet. */
791 flags
= TCP_SKB_CB(skb
)->flags
;
792 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
| TCPCB_FLAG_PSH
);
793 TCP_SKB_CB(buff
)->flags
= flags
;
794 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
796 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_PARTIAL
) {
797 /* Copy and checksum data tail into the new buffer. */
798 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
,
799 skb_put(buff
, nsize
),
804 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
806 skb
->ip_summed
= CHECKSUM_PARTIAL
;
807 skb_split(skb
, buff
, len
);
810 buff
->ip_summed
= skb
->ip_summed
;
812 /* Looks stupid, but our code really uses when of
813 * skbs, which it never sent before. --ANK
815 TCP_SKB_CB(buff
)->when
= TCP_SKB_CB(skb
)->when
;
816 buff
->tstamp
= skb
->tstamp
;
818 old_factor
= tcp_skb_pcount(skb
);
820 /* Fix up tso_factor for both original and new SKB. */
821 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
822 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
824 /* If this packet has been sent out already, we must
825 * adjust the various packet counters.
827 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
828 int diff
= old_factor
- tcp_skb_pcount(skb
) -
829 tcp_skb_pcount(buff
);
831 tp
->packets_out
-= diff
;
833 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
834 tp
->sacked_out
-= diff
;
835 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
836 tp
->retrans_out
-= diff
;
838 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
)
839 tp
->lost_out
-= diff
;
841 /* Adjust Reno SACK estimate. */
842 if (tcp_is_reno(tp
) && diff
> 0) {
843 tcp_dec_pcount_approx_int(&tp
->sacked_out
, diff
);
844 tcp_verify_left_out(tp
);
846 tcp_adjust_fackets_out(sk
, skb
, diff
);
849 /* Link BUFF into the send queue. */
850 skb_header_release(buff
);
851 tcp_insert_write_queue_after(skb
, buff
, sk
);
856 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
857 * eventually). The difference is that pulled data not copied, but
858 * immediately discarded.
860 static void __pskb_trim_head(struct sk_buff
*skb
, int len
)
866 for (i
= 0; i
< skb_shinfo(skb
)->nr_frags
; i
++) {
867 if (skb_shinfo(skb
)->frags
[i
].size
<= eat
) {
868 put_page(skb_shinfo(skb
)->frags
[i
].page
);
869 eat
-= skb_shinfo(skb
)->frags
[i
].size
;
871 skb_shinfo(skb
)->frags
[k
] = skb_shinfo(skb
)->frags
[i
];
873 skb_shinfo(skb
)->frags
[k
].page_offset
+= eat
;
874 skb_shinfo(skb
)->frags
[k
].size
-= eat
;
880 skb_shinfo(skb
)->nr_frags
= k
;
882 skb_reset_tail_pointer(skb
);
883 skb
->data_len
-= len
;
884 skb
->len
= skb
->data_len
;
887 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
889 if (skb_cloned(skb
) && pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
892 /* If len == headlen, we avoid __skb_pull to preserve alignment. */
893 if (unlikely(len
< skb_headlen(skb
)))
894 __skb_pull(skb
, len
);
896 __pskb_trim_head(skb
, len
- skb_headlen(skb
));
898 TCP_SKB_CB(skb
)->seq
+= len
;
899 skb
->ip_summed
= CHECKSUM_PARTIAL
;
901 skb
->truesize
-= len
;
902 sk
->sk_wmem_queued
-= len
;
903 sk_mem_uncharge(sk
, len
);
904 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
906 /* Any change of skb->len requires recalculation of tso
909 if (tcp_skb_pcount(skb
) > 1)
910 tcp_set_skb_tso_segs(sk
, skb
, tcp_current_mss(sk
, 1));
915 /* Not accounting for SACKs here. */
916 int tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
918 struct tcp_sock
*tp
= tcp_sk(sk
);
919 struct inet_connection_sock
*icsk
= inet_csk(sk
);
922 /* Calculate base mss without TCP options:
923 It is MMS_S - sizeof(tcphdr) of rfc1122
925 mss_now
= pmtu
- icsk
->icsk_af_ops
->net_header_len
- sizeof(struct tcphdr
);
927 /* Clamp it (mss_clamp does not include tcp options) */
928 if (mss_now
> tp
->rx_opt
.mss_clamp
)
929 mss_now
= tp
->rx_opt
.mss_clamp
;
931 /* Now subtract optional transport overhead */
932 mss_now
-= icsk
->icsk_ext_hdr_len
;
934 /* Then reserve room for full set of TCP options and 8 bytes of data */
938 /* Now subtract TCP options size, not including SACKs */
939 mss_now
-= tp
->tcp_header_len
- sizeof(struct tcphdr
);
944 /* Inverse of above */
945 int tcp_mss_to_mtu(struct sock
*sk
, int mss
)
947 struct tcp_sock
*tp
= tcp_sk(sk
);
948 struct inet_connection_sock
*icsk
= inet_csk(sk
);
953 icsk
->icsk_ext_hdr_len
+
954 icsk
->icsk_af_ops
->net_header_len
;
959 void tcp_mtup_init(struct sock
*sk
)
961 struct tcp_sock
*tp
= tcp_sk(sk
);
962 struct inet_connection_sock
*icsk
= inet_csk(sk
);
964 icsk
->icsk_mtup
.enabled
= sysctl_tcp_mtu_probing
> 1;
965 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+ sizeof(struct tcphdr
) +
966 icsk
->icsk_af_ops
->net_header_len
;
967 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, sysctl_tcp_base_mss
);
968 icsk
->icsk_mtup
.probe_size
= 0;
971 /* Bound MSS / TSO packet size with the half of the window */
972 static int tcp_bound_to_half_wnd(struct tcp_sock
*tp
, int pktsize
)
974 if (tp
->max_window
&& pktsize
> (tp
->max_window
>> 1))
975 return max(tp
->max_window
>> 1, 68U - tp
->tcp_header_len
);
980 /* This function synchronize snd mss to current pmtu/exthdr set.
982 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
983 for TCP options, but includes only bare TCP header.
985 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
986 It is minimum of user_mss and mss received with SYN.
987 It also does not include TCP options.
989 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
991 tp->mss_cache is current effective sending mss, including
992 all tcp options except for SACKs. It is evaluated,
993 taking into account current pmtu, but never exceeds
994 tp->rx_opt.mss_clamp.
996 NOTE1. rfc1122 clearly states that advertised MSS
997 DOES NOT include either tcp or ip options.
999 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1000 are READ ONLY outside this function. --ANK (980731)
1002 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
1004 struct tcp_sock
*tp
= tcp_sk(sk
);
1005 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1008 if (icsk
->icsk_mtup
.search_high
> pmtu
)
1009 icsk
->icsk_mtup
.search_high
= pmtu
;
1011 mss_now
= tcp_mtu_to_mss(sk
, pmtu
);
1012 mss_now
= tcp_bound_to_half_wnd(tp
, mss_now
);
1014 /* And store cached results */
1015 icsk
->icsk_pmtu_cookie
= pmtu
;
1016 if (icsk
->icsk_mtup
.enabled
)
1017 mss_now
= min(mss_now
, tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
));
1018 tp
->mss_cache
= mss_now
;
1023 /* Compute the current effective MSS, taking SACKs and IP options,
1024 * and even PMTU discovery events into account.
1026 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
1027 * cannot be large. However, taking into account rare use of URG, this
1028 * is not a big flaw.
1030 unsigned int tcp_current_mss(struct sock
*sk
, int large_allowed
)
1032 struct tcp_sock
*tp
= tcp_sk(sk
);
1033 struct dst_entry
*dst
= __sk_dst_get(sk
);
1037 unsigned header_len
;
1038 struct tcp_out_options opts
;
1039 struct tcp_md5sig_key
*md5
;
1041 mss_now
= tp
->mss_cache
;
1043 if (large_allowed
&& sk_can_gso(sk
) && !tp
->urg_mode
)
1047 u32 mtu
= dst_mtu(dst
);
1048 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
1049 mss_now
= tcp_sync_mss(sk
, mtu
);
1052 header_len
= tcp_established_options(sk
, NULL
, &opts
, &md5
) +
1053 sizeof(struct tcphdr
);
1054 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1055 * some common options. If this is an odd packet (because we have SACK
1056 * blocks etc) then our calculated header_len will be different, and
1057 * we have to adjust mss_now correspondingly */
1058 if (header_len
!= tp
->tcp_header_len
) {
1059 int delta
= (int) header_len
- tp
->tcp_header_len
;
1063 xmit_size_goal
= mss_now
;
1066 xmit_size_goal
= ((sk
->sk_gso_max_size
- 1) -
1067 inet_csk(sk
)->icsk_af_ops
->net_header_len
-
1068 inet_csk(sk
)->icsk_ext_hdr_len
-
1069 tp
->tcp_header_len
);
1071 xmit_size_goal
= tcp_bound_to_half_wnd(tp
, xmit_size_goal
);
1072 xmit_size_goal
-= (xmit_size_goal
% mss_now
);
1074 tp
->xmit_size_goal
= xmit_size_goal
;
1079 /* Congestion window validation. (RFC2861) */
1080 static void tcp_cwnd_validate(struct sock
*sk
)
1082 struct tcp_sock
*tp
= tcp_sk(sk
);
1084 if (tp
->packets_out
>= tp
->snd_cwnd
) {
1085 /* Network is feed fully. */
1086 tp
->snd_cwnd_used
= 0;
1087 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1089 /* Network starves. */
1090 if (tp
->packets_out
> tp
->snd_cwnd_used
)
1091 tp
->snd_cwnd_used
= tp
->packets_out
;
1093 if (sysctl_tcp_slow_start_after_idle
&&
1094 (s32
)(tcp_time_stamp
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
)
1095 tcp_cwnd_application_limited(sk
);
1099 /* Returns the portion of skb which can be sent right away without
1100 * introducing MSS oddities to segment boundaries. In rare cases where
1101 * mss_now != mss_cache, we will request caller to create a small skb
1102 * per input skb which could be mostly avoided here (if desired).
1104 * We explicitly want to create a request for splitting write queue tail
1105 * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1106 * thus all the complexity (cwnd_len is always MSS multiple which we
1107 * return whenever allowed by the other factors). Basically we need the
1108 * modulo only when the receiver window alone is the limiting factor or
1109 * when we would be allowed to send the split-due-to-Nagle skb fully.
1111 static unsigned int tcp_mss_split_point(struct sock
*sk
, struct sk_buff
*skb
,
1112 unsigned int mss_now
, unsigned int cwnd
)
1114 struct tcp_sock
*tp
= tcp_sk(sk
);
1115 u32 needed
, window
, cwnd_len
;
1117 window
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1118 cwnd_len
= mss_now
* cwnd
;
1120 if (likely(cwnd_len
<= window
&& skb
!= tcp_write_queue_tail(sk
)))
1123 needed
= min(skb
->len
, window
);
1125 if (cwnd_len
<= needed
)
1128 return needed
- needed
% mss_now
;
1131 /* Can at least one segment of SKB be sent right now, according to the
1132 * congestion window rules? If so, return how many segments are allowed.
1134 static inline unsigned int tcp_cwnd_test(struct tcp_sock
*tp
,
1135 struct sk_buff
*skb
)
1137 u32 in_flight
, cwnd
;
1139 /* Don't be strict about the congestion window for the final FIN. */
1140 if ((TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
) &&
1141 tcp_skb_pcount(skb
) == 1)
1144 in_flight
= tcp_packets_in_flight(tp
);
1145 cwnd
= tp
->snd_cwnd
;
1146 if (in_flight
< cwnd
)
1147 return (cwnd
- in_flight
);
1152 /* This must be invoked the first time we consider transmitting
1153 * SKB onto the wire.
1155 static int tcp_init_tso_segs(struct sock
*sk
, struct sk_buff
*skb
,
1156 unsigned int mss_now
)
1158 int tso_segs
= tcp_skb_pcount(skb
);
1160 if (!tso_segs
|| (tso_segs
> 1 && tcp_skb_mss(skb
) != mss_now
)) {
1161 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1162 tso_segs
= tcp_skb_pcount(skb
);
1167 static inline int tcp_minshall_check(const struct tcp_sock
*tp
)
1169 return after(tp
->snd_sml
,tp
->snd_una
) &&
1170 !after(tp
->snd_sml
, tp
->snd_nxt
);
1173 /* Return 0, if packet can be sent now without violation Nagle's rules:
1174 * 1. It is full sized.
1175 * 2. Or it contains FIN. (already checked by caller)
1176 * 3. Or TCP_NODELAY was set.
1177 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1178 * With Minshall's modification: all sent small packets are ACKed.
1180 static inline int tcp_nagle_check(const struct tcp_sock
*tp
,
1181 const struct sk_buff
*skb
,
1182 unsigned mss_now
, int nonagle
)
1184 return (skb
->len
< mss_now
&&
1185 ((nonagle
& TCP_NAGLE_CORK
) ||
1186 (!nonagle
&& tp
->packets_out
&& tcp_minshall_check(tp
))));
1189 /* Return non-zero if the Nagle test allows this packet to be
1192 static inline int tcp_nagle_test(struct tcp_sock
*tp
, struct sk_buff
*skb
,
1193 unsigned int cur_mss
, int nonagle
)
1195 /* Nagle rule does not apply to frames, which sit in the middle of the
1196 * write_queue (they have no chances to get new data).
1198 * This is implemented in the callers, where they modify the 'nonagle'
1199 * argument based upon the location of SKB in the send queue.
1201 if (nonagle
& TCP_NAGLE_PUSH
)
1204 /* Don't use the nagle rule for urgent data (or for the final FIN).
1205 * Nagle can be ignored during F-RTO too (see RFC4138).
1207 if (tp
->urg_mode
|| (tp
->frto_counter
== 2) ||
1208 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
))
1211 if (!tcp_nagle_check(tp
, skb
, cur_mss
, nonagle
))
1217 /* Does at least the first segment of SKB fit into the send window? */
1218 static inline int tcp_snd_wnd_test(struct tcp_sock
*tp
, struct sk_buff
*skb
,
1219 unsigned int cur_mss
)
1221 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1223 if (skb
->len
> cur_mss
)
1224 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
1226 return !after(end_seq
, tcp_wnd_end(tp
));
1229 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1230 * should be put on the wire right now. If so, it returns the number of
1231 * packets allowed by the congestion window.
1233 static unsigned int tcp_snd_test(struct sock
*sk
, struct sk_buff
*skb
,
1234 unsigned int cur_mss
, int nonagle
)
1236 struct tcp_sock
*tp
= tcp_sk(sk
);
1237 unsigned int cwnd_quota
;
1239 tcp_init_tso_segs(sk
, skb
, cur_mss
);
1241 if (!tcp_nagle_test(tp
, skb
, cur_mss
, nonagle
))
1244 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1245 if (cwnd_quota
&& !tcp_snd_wnd_test(tp
, skb
, cur_mss
))
1251 int tcp_may_send_now(struct sock
*sk
)
1253 struct tcp_sock
*tp
= tcp_sk(sk
);
1254 struct sk_buff
*skb
= tcp_send_head(sk
);
1257 tcp_snd_test(sk
, skb
, tcp_current_mss(sk
, 1),
1258 (tcp_skb_is_last(sk
, skb
) ?
1259 tp
->nonagle
: TCP_NAGLE_PUSH
)));
1262 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1263 * which is put after SKB on the list. It is very much like
1264 * tcp_fragment() except that it may make several kinds of assumptions
1265 * in order to speed up the splitting operation. In particular, we
1266 * know that all the data is in scatter-gather pages, and that the
1267 * packet has never been sent out before (and thus is not cloned).
1269 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
,
1270 unsigned int mss_now
)
1272 struct sk_buff
*buff
;
1273 int nlen
= skb
->len
- len
;
1276 /* All of a TSO frame must be composed of paged data. */
1277 if (skb
->len
!= skb
->data_len
)
1278 return tcp_fragment(sk
, skb
, len
, mss_now
);
1280 buff
= sk_stream_alloc_skb(sk
, 0, GFP_ATOMIC
);
1281 if (unlikely(buff
== NULL
))
1284 sk
->sk_wmem_queued
+= buff
->truesize
;
1285 sk_mem_charge(sk
, buff
->truesize
);
1286 buff
->truesize
+= nlen
;
1287 skb
->truesize
-= nlen
;
1289 /* Correct the sequence numbers. */
1290 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1291 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1292 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1294 /* PSH and FIN should only be set in the second packet. */
1295 flags
= TCP_SKB_CB(skb
)->flags
;
1296 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
| TCPCB_FLAG_PSH
);
1297 TCP_SKB_CB(buff
)->flags
= flags
;
1299 /* This packet was never sent out yet, so no SACK bits. */
1300 TCP_SKB_CB(buff
)->sacked
= 0;
1302 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_PARTIAL
;
1303 skb_split(skb
, buff
, len
);
1305 /* Fix up tso_factor for both original and new SKB. */
1306 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1307 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1309 /* Link BUFF into the send queue. */
1310 skb_header_release(buff
);
1311 tcp_insert_write_queue_after(skb
, buff
, sk
);
1316 /* Try to defer sending, if possible, in order to minimize the amount
1317 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1319 * This algorithm is from John Heffner.
1321 static int tcp_tso_should_defer(struct sock
*sk
, struct sk_buff
*skb
)
1323 struct tcp_sock
*tp
= tcp_sk(sk
);
1324 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1325 u32 send_win
, cong_win
, limit
, in_flight
;
1327 if (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
)
1330 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
1333 /* Defer for less than two clock ticks. */
1334 if (tp
->tso_deferred
&&
1335 ((jiffies
<< 1) >> 1) - (tp
->tso_deferred
>> 1) > 1)
1338 in_flight
= tcp_packets_in_flight(tp
);
1340 BUG_ON(tcp_skb_pcount(skb
) <= 1 || (tp
->snd_cwnd
<= in_flight
));
1342 send_win
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1344 /* From in_flight test above, we know that cwnd > in_flight. */
1345 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
1347 limit
= min(send_win
, cong_win
);
1349 /* If a full-sized TSO skb can be sent, do it. */
1350 if (limit
>= sk
->sk_gso_max_size
)
1353 if (sysctl_tcp_tso_win_divisor
) {
1354 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
1356 /* If at least some fraction of a window is available,
1359 chunk
/= sysctl_tcp_tso_win_divisor
;
1363 /* Different approach, try not to defer past a single
1364 * ACK. Receiver should ACK every other full sized
1365 * frame, so if we have space for more than 3 frames
1368 if (limit
> tcp_max_burst(tp
) * tp
->mss_cache
)
1372 /* Ok, it looks like it is advisable to defer. */
1373 tp
->tso_deferred
= 1 | (jiffies
<< 1);
1378 tp
->tso_deferred
= 0;
1382 /* Create a new MTU probe if we are ready.
1383 * Returns 0 if we should wait to probe (no cwnd available),
1384 * 1 if a probe was sent,
1387 static int tcp_mtu_probe(struct sock
*sk
)
1389 struct tcp_sock
*tp
= tcp_sk(sk
);
1390 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1391 struct sk_buff
*skb
, *nskb
, *next
;
1398 /* Not currently probing/verifying,
1400 * have enough cwnd, and
1401 * not SACKing (the variable headers throw things off) */
1402 if (!icsk
->icsk_mtup
.enabled
||
1403 icsk
->icsk_mtup
.probe_size
||
1404 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
||
1405 tp
->snd_cwnd
< 11 ||
1406 tp
->rx_opt
.eff_sacks
)
1409 /* Very simple search strategy: just double the MSS. */
1410 mss_now
= tcp_current_mss(sk
, 0);
1411 probe_size
= 2 * tp
->mss_cache
;
1412 size_needed
= probe_size
+ (tp
->reordering
+ 1) * tp
->mss_cache
;
1413 if (probe_size
> tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_high
)) {
1414 /* TODO: set timer for probe_converge_event */
1418 /* Have enough data in the send queue to probe? */
1419 if (tp
->write_seq
- tp
->snd_nxt
< size_needed
)
1422 if (tp
->snd_wnd
< size_needed
)
1424 if (after(tp
->snd_nxt
+ size_needed
, tcp_wnd_end(tp
)))
1427 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1428 if (tcp_packets_in_flight(tp
) + 2 > tp
->snd_cwnd
) {
1429 if (!tcp_packets_in_flight(tp
))
1435 /* We're allowed to probe. Build it now. */
1436 if ((nskb
= sk_stream_alloc_skb(sk
, probe_size
, GFP_ATOMIC
)) == NULL
)
1438 sk
->sk_wmem_queued
+= nskb
->truesize
;
1439 sk_mem_charge(sk
, nskb
->truesize
);
1441 skb
= tcp_send_head(sk
);
1443 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(skb
)->seq
;
1444 TCP_SKB_CB(nskb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ probe_size
;
1445 TCP_SKB_CB(nskb
)->flags
= TCPCB_FLAG_ACK
;
1446 TCP_SKB_CB(nskb
)->sacked
= 0;
1448 nskb
->ip_summed
= skb
->ip_summed
;
1450 tcp_insert_write_queue_before(nskb
, skb
, sk
);
1453 tcp_for_write_queue_from_safe(skb
, next
, sk
) {
1454 copy
= min_t(int, skb
->len
, probe_size
- len
);
1455 if (nskb
->ip_summed
)
1456 skb_copy_bits(skb
, 0, skb_put(nskb
, copy
), copy
);
1458 nskb
->csum
= skb_copy_and_csum_bits(skb
, 0,
1459 skb_put(nskb
, copy
),
1462 if (skb
->len
<= copy
) {
1463 /* We've eaten all the data from this skb.
1465 TCP_SKB_CB(nskb
)->flags
|= TCP_SKB_CB(skb
)->flags
;
1466 tcp_unlink_write_queue(skb
, sk
);
1467 sk_wmem_free_skb(sk
, skb
);
1469 TCP_SKB_CB(nskb
)->flags
|= TCP_SKB_CB(skb
)->flags
&
1470 ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
1471 if (!skb_shinfo(skb
)->nr_frags
) {
1472 skb_pull(skb
, copy
);
1473 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1474 skb
->csum
= csum_partial(skb
->data
,
1477 __pskb_trim_head(skb
, copy
);
1478 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1480 TCP_SKB_CB(skb
)->seq
+= copy
;
1485 if (len
>= probe_size
)
1488 tcp_init_tso_segs(sk
, nskb
, nskb
->len
);
1490 /* We're ready to send. If this fails, the probe will
1491 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1492 TCP_SKB_CB(nskb
)->when
= tcp_time_stamp
;
1493 if (!tcp_transmit_skb(sk
, nskb
, 1, GFP_ATOMIC
)) {
1494 /* Decrement cwnd here because we are sending
1495 * effectively two packets. */
1497 tcp_event_new_data_sent(sk
, nskb
);
1499 icsk
->icsk_mtup
.probe_size
= tcp_mss_to_mtu(sk
, nskb
->len
);
1500 tp
->mtu_probe
.probe_seq_start
= TCP_SKB_CB(nskb
)->seq
;
1501 tp
->mtu_probe
.probe_seq_end
= TCP_SKB_CB(nskb
)->end_seq
;
1509 /* This routine writes packets to the network. It advances the
1510 * send_head. This happens as incoming acks open up the remote
1513 * Returns 1, if no segments are in flight and we have queued segments, but
1514 * cannot send anything now because of SWS or another problem.
1516 static int tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
)
1518 struct tcp_sock
*tp
= tcp_sk(sk
);
1519 struct sk_buff
*skb
;
1520 unsigned int tso_segs
, sent_pkts
;
1524 /* If we are closed, the bytes will have to remain here.
1525 * In time closedown will finish, we empty the write queue and all
1528 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
1533 /* Do MTU probing. */
1534 if ((result
= tcp_mtu_probe(sk
)) == 0) {
1536 } else if (result
> 0) {
1540 while ((skb
= tcp_send_head(sk
))) {
1543 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1546 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1550 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
)))
1553 if (tso_segs
== 1) {
1554 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
1555 (tcp_skb_is_last(sk
, skb
) ?
1556 nonagle
: TCP_NAGLE_PUSH
))))
1559 if (tcp_tso_should_defer(sk
, skb
))
1565 limit
= tcp_mss_split_point(sk
, skb
, mss_now
,
1568 if (skb
->len
> limit
&&
1569 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
)))
1572 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1574 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
)))
1577 /* Advance the send_head. This one is sent out.
1578 * This call will increment packets_out.
1580 tcp_event_new_data_sent(sk
, skb
);
1582 tcp_minshall_update(tp
, mss_now
, skb
);
1586 if (likely(sent_pkts
)) {
1587 tcp_cwnd_validate(sk
);
1590 return !tp
->packets_out
&& tcp_send_head(sk
);
1593 /* Push out any pending frames which were held back due to
1594 * TCP_CORK or attempt at coalescing tiny packets.
1595 * The socket must be locked by the caller.
1597 void __tcp_push_pending_frames(struct sock
*sk
, unsigned int cur_mss
,
1600 struct sk_buff
*skb
= tcp_send_head(sk
);
1603 if (tcp_write_xmit(sk
, cur_mss
, nonagle
))
1604 tcp_check_probe_timer(sk
);
1608 /* Send _single_ skb sitting at the send head. This function requires
1609 * true push pending frames to setup probe timer etc.
1611 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
1613 struct sk_buff
*skb
= tcp_send_head(sk
);
1614 unsigned int tso_segs
, cwnd_quota
;
1616 BUG_ON(!skb
|| skb
->len
< mss_now
);
1618 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1619 cwnd_quota
= tcp_snd_test(sk
, skb
, mss_now
, TCP_NAGLE_PUSH
);
1621 if (likely(cwnd_quota
)) {
1628 limit
= tcp_mss_split_point(sk
, skb
, mss_now
,
1631 if (skb
->len
> limit
&&
1632 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
)))
1635 /* Send it out now. */
1636 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1638 if (likely(!tcp_transmit_skb(sk
, skb
, 1, sk
->sk_allocation
))) {
1639 tcp_event_new_data_sent(sk
, skb
);
1640 tcp_cwnd_validate(sk
);
1646 /* This function returns the amount that we can raise the
1647 * usable window based on the following constraints
1649 * 1. The window can never be shrunk once it is offered (RFC 793)
1650 * 2. We limit memory per socket
1653 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1654 * RECV.NEXT + RCV.WIN fixed until:
1655 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1657 * i.e. don't raise the right edge of the window until you can raise
1658 * it at least MSS bytes.
1660 * Unfortunately, the recommended algorithm breaks header prediction,
1661 * since header prediction assumes th->window stays fixed.
1663 * Strictly speaking, keeping th->window fixed violates the receiver
1664 * side SWS prevention criteria. The problem is that under this rule
1665 * a stream of single byte packets will cause the right side of the
1666 * window to always advance by a single byte.
1668 * Of course, if the sender implements sender side SWS prevention
1669 * then this will not be a problem.
1671 * BSD seems to make the following compromise:
1673 * If the free space is less than the 1/4 of the maximum
1674 * space available and the free space is less than 1/2 mss,
1675 * then set the window to 0.
1676 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1677 * Otherwise, just prevent the window from shrinking
1678 * and from being larger than the largest representable value.
1680 * This prevents incremental opening of the window in the regime
1681 * where TCP is limited by the speed of the reader side taking
1682 * data out of the TCP receive queue. It does nothing about
1683 * those cases where the window is constrained on the sender side
1684 * because the pipeline is full.
1686 * BSD also seems to "accidentally" limit itself to windows that are a
1687 * multiple of MSS, at least until the free space gets quite small.
1688 * This would appear to be a side effect of the mbuf implementation.
1689 * Combining these two algorithms results in the observed behavior
1690 * of having a fixed window size at almost all times.
1692 * Below we obtain similar behavior by forcing the offered window to
1693 * a multiple of the mss when it is feasible to do so.
1695 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1696 * Regular options like TIMESTAMP are taken into account.
1698 u32
__tcp_select_window(struct sock
*sk
)
1700 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1701 struct tcp_sock
*tp
= tcp_sk(sk
);
1702 /* MSS for the peer's data. Previous versions used mss_clamp
1703 * here. I don't know if the value based on our guesses
1704 * of peer's MSS is better for the performance. It's more correct
1705 * but may be worse for the performance because of rcv_mss
1706 * fluctuations. --SAW 1998/11/1
1708 int mss
= icsk
->icsk_ack
.rcv_mss
;
1709 int free_space
= tcp_space(sk
);
1710 int full_space
= min_t(int, tp
->window_clamp
, tcp_full_space(sk
));
1713 if (mss
> full_space
)
1716 if (free_space
< (full_space
>> 1)) {
1717 icsk
->icsk_ack
.quick
= 0;
1719 if (tcp_memory_pressure
)
1720 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
,
1723 if (free_space
< mss
)
1727 if (free_space
> tp
->rcv_ssthresh
)
1728 free_space
= tp
->rcv_ssthresh
;
1730 /* Don't do rounding if we are using window scaling, since the
1731 * scaled window will not line up with the MSS boundary anyway.
1733 window
= tp
->rcv_wnd
;
1734 if (tp
->rx_opt
.rcv_wscale
) {
1735 window
= free_space
;
1737 /* Advertise enough space so that it won't get scaled away.
1738 * Import case: prevent zero window announcement if
1739 * 1<<rcv_wscale > mss.
1741 if (((window
>> tp
->rx_opt
.rcv_wscale
) << tp
->rx_opt
.rcv_wscale
) != window
)
1742 window
= (((window
>> tp
->rx_opt
.rcv_wscale
) + 1)
1743 << tp
->rx_opt
.rcv_wscale
);
1745 /* Get the largest window that is a nice multiple of mss.
1746 * Window clamp already applied above.
1747 * If our current window offering is within 1 mss of the
1748 * free space we just keep it. This prevents the divide
1749 * and multiply from happening most of the time.
1750 * We also don't do any window rounding when the free space
1753 if (window
<= free_space
- mss
|| window
> free_space
)
1754 window
= (free_space
/ mss
) * mss
;
1755 else if (mss
== full_space
&&
1756 free_space
> window
+ (full_space
>> 1))
1757 window
= free_space
;
1763 /* Attempt to collapse two adjacent SKB's during retransmission. */
1764 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*skb
,
1767 struct tcp_sock
*tp
= tcp_sk(sk
);
1768 struct sk_buff
*next_skb
= tcp_write_queue_next(sk
, skb
);
1769 int skb_size
, next_skb_size
;
1772 /* The first test we must make is that neither of these two
1773 * SKB's are still referenced by someone else.
1775 if (skb_cloned(skb
) || skb_cloned(next_skb
))
1778 skb_size
= skb
->len
;
1779 next_skb_size
= next_skb
->len
;
1780 flags
= TCP_SKB_CB(skb
)->flags
;
1782 /* Also punt if next skb has been SACK'd. */
1783 if (TCP_SKB_CB(next_skb
)->sacked
& TCPCB_SACKED_ACKED
)
1786 /* Next skb is out of window. */
1787 if (after(TCP_SKB_CB(next_skb
)->end_seq
, tcp_wnd_end(tp
)))
1790 /* Punt if not enough space exists in the first SKB for
1791 * the data in the second, or the total combined payload
1792 * would exceed the MSS.
1794 if ((next_skb_size
> skb_tailroom(skb
)) ||
1795 ((skb_size
+ next_skb_size
) > mss_now
))
1798 BUG_ON(tcp_skb_pcount(skb
) != 1 || tcp_skb_pcount(next_skb
) != 1);
1800 tcp_highest_sack_combine(sk
, next_skb
, skb
);
1802 /* Ok. We will be able to collapse the packet. */
1803 tcp_unlink_write_queue(next_skb
, sk
);
1805 skb_copy_from_linear_data(next_skb
, skb_put(skb
, next_skb_size
),
1808 if (next_skb
->ip_summed
== CHECKSUM_PARTIAL
)
1809 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1811 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1812 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
1814 /* Update sequence range on original skb. */
1815 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
1817 /* Merge over control information. */
1818 flags
|= TCP_SKB_CB(next_skb
)->flags
; /* This moves PSH/FIN etc. over */
1819 TCP_SKB_CB(skb
)->flags
= flags
;
1821 /* All done, get rid of second SKB and account for it so
1822 * packet counting does not break.
1824 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
& TCPCB_EVER_RETRANS
;
1825 if (TCP_SKB_CB(next_skb
)->sacked
& TCPCB_SACKED_RETRANS
)
1826 tp
->retrans_out
-= tcp_skb_pcount(next_skb
);
1827 if (TCP_SKB_CB(next_skb
)->sacked
& TCPCB_LOST
)
1828 tp
->lost_out
-= tcp_skb_pcount(next_skb
);
1829 /* Reno case is special. Sigh... */
1830 if (tcp_is_reno(tp
) && tp
->sacked_out
)
1831 tcp_dec_pcount_approx(&tp
->sacked_out
, next_skb
);
1833 tcp_adjust_fackets_out(sk
, next_skb
, tcp_skb_pcount(next_skb
));
1834 tp
->packets_out
-= tcp_skb_pcount(next_skb
);
1836 /* changed transmit queue under us so clear hints */
1837 tcp_clear_retrans_hints_partial(tp
);
1839 sk_wmem_free_skb(sk
, next_skb
);
1842 /* Do a simple retransmit without using the backoff mechanisms in
1843 * tcp_timer. This is used for path mtu discovery.
1844 * The socket is already locked here.
1846 void tcp_simple_retransmit(struct sock
*sk
)
1848 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1849 struct tcp_sock
*tp
= tcp_sk(sk
);
1850 struct sk_buff
*skb
;
1851 unsigned int mss
= tcp_current_mss(sk
, 0);
1854 tcp_for_write_queue(skb
, sk
) {
1855 if (skb
== tcp_send_head(sk
))
1857 if (skb
->len
> mss
&&
1858 !(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
1859 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
1860 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1861 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1863 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
)) {
1864 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1865 tp
->lost_out
+= tcp_skb_pcount(skb
);
1871 tcp_clear_all_retrans_hints(tp
);
1876 if (tcp_is_reno(tp
))
1877 tcp_limit_reno_sacked(tp
);
1879 tcp_verify_left_out(tp
);
1881 /* Don't muck with the congestion window here.
1882 * Reason is that we do not increase amount of _data_
1883 * in network, but units changed and effective
1884 * cwnd/ssthresh really reduced now.
1886 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
1887 tp
->high_seq
= tp
->snd_nxt
;
1888 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
1889 tp
->prior_ssthresh
= 0;
1890 tp
->undo_marker
= 0;
1891 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1893 tcp_xmit_retransmit_queue(sk
);
1896 /* This retransmits one SKB. Policy decisions and retransmit queue
1897 * state updates are done by the caller. Returns non-zero if an
1898 * error occurred which prevented the send.
1900 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
1902 struct tcp_sock
*tp
= tcp_sk(sk
);
1903 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1904 unsigned int cur_mss
;
1907 /* Inconslusive MTU probe */
1908 if (icsk
->icsk_mtup
.probe_size
) {
1909 icsk
->icsk_mtup
.probe_size
= 0;
1912 /* Do not sent more than we queued. 1/4 is reserved for possible
1913 * copying overhead: fragmentation, tunneling, mangling etc.
1915 if (atomic_read(&sk
->sk_wmem_alloc
) >
1916 min(sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2), sk
->sk_sndbuf
))
1919 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
1920 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1922 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
1926 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
1927 return -EHOSTUNREACH
; /* Routing failure or similar. */
1929 cur_mss
= tcp_current_mss(sk
, 0);
1931 /* If receiver has shrunk his window, and skb is out of
1932 * new window, do not retransmit it. The exception is the
1933 * case, when window is shrunk to zero. In this case
1934 * our retransmit serves as a zero window probe.
1936 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
))
1937 && TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
1940 if (skb
->len
> cur_mss
) {
1941 if (tcp_fragment(sk
, skb
, cur_mss
, cur_mss
))
1942 return -ENOMEM
; /* We'll try again later. */
1945 /* Collapse two adjacent packets if worthwhile and we can. */
1946 if (!(TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_SYN
) &&
1947 (skb
->len
< (cur_mss
>> 1)) &&
1948 (tcp_write_queue_next(sk
, skb
) != tcp_send_head(sk
)) &&
1949 (!tcp_skb_is_last(sk
, skb
)) &&
1950 (skb_shinfo(skb
)->nr_frags
== 0 &&
1951 skb_shinfo(tcp_write_queue_next(sk
, skb
))->nr_frags
== 0) &&
1952 (tcp_skb_pcount(skb
) == 1 &&
1953 tcp_skb_pcount(tcp_write_queue_next(sk
, skb
)) == 1) &&
1954 (sysctl_tcp_retrans_collapse
!= 0))
1955 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
1957 /* Some Solaris stacks overoptimize and ignore the FIN on a
1958 * retransmit when old data is attached. So strip it off
1959 * since it is cheap to do so and saves bytes on the network.
1962 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
) &&
1963 tp
->snd_una
== (TCP_SKB_CB(skb
)->end_seq
- 1)) {
1964 if (!pskb_trim(skb
, 0)) {
1965 /* Reuse, even though it does some unnecessary work */
1966 tcp_init_nondata_skb(skb
, TCP_SKB_CB(skb
)->end_seq
- 1,
1967 TCP_SKB_CB(skb
)->flags
);
1968 skb
->ip_summed
= CHECKSUM_NONE
;
1972 /* Make a copy, if the first transmission SKB clone we made
1973 * is still in somebody's hands, else make a clone.
1975 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1977 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
1980 /* Update global TCP statistics. */
1981 TCP_INC_STATS(sock_net(sk
), TCP_MIB_RETRANSSEGS
);
1983 tp
->total_retrans
++;
1985 #if FASTRETRANS_DEBUG > 0
1986 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
1987 if (net_ratelimit())
1988 printk(KERN_DEBUG
"retrans_out leaked.\n");
1991 if (!tp
->retrans_out
)
1992 tp
->lost_retrans_low
= tp
->snd_nxt
;
1993 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
1994 tp
->retrans_out
+= tcp_skb_pcount(skb
);
1996 /* Save stamp of the first retransmit. */
1997 if (!tp
->retrans_stamp
)
1998 tp
->retrans_stamp
= TCP_SKB_CB(skb
)->when
;
2002 /* snd_nxt is stored to detect loss of retransmitted segment,
2003 * see tcp_input.c tcp_sacktag_write_queue().
2005 TCP_SKB_CB(skb
)->ack_seq
= tp
->snd_nxt
;
2010 /* This gets called after a retransmit timeout, and the initially
2011 * retransmitted data is acknowledged. It tries to continue
2012 * resending the rest of the retransmit queue, until either
2013 * we've sent it all or the congestion window limit is reached.
2014 * If doing SACK, the first ACK which comes back for a timeout
2015 * based retransmit packet might feed us FACK information again.
2016 * If so, we use it to avoid unnecessarily retransmissions.
2018 void tcp_xmit_retransmit_queue(struct sock
*sk
)
2020 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2021 struct tcp_sock
*tp
= tcp_sk(sk
);
2022 struct sk_buff
*skb
;
2025 if (tp
->retransmit_skb_hint
) {
2026 skb
= tp
->retransmit_skb_hint
;
2027 packet_cnt
= tp
->retransmit_cnt_hint
;
2029 skb
= tcp_write_queue_head(sk
);
2033 /* First pass: retransmit lost packets. */
2035 tcp_for_write_queue_from(skb
, sk
) {
2036 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
2038 if (skb
== tcp_send_head(sk
))
2040 /* we could do better than to assign each time */
2041 tp
->retransmit_skb_hint
= skb
;
2042 tp
->retransmit_cnt_hint
= packet_cnt
;
2044 /* Assume this retransmit will generate
2045 * only one packet for congestion window
2046 * calculation purposes. This works because
2047 * tcp_retransmit_skb() will chop up the
2048 * packet to be MSS sized and all the
2049 * packet counting works out.
2051 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
2054 if (sacked
& TCPCB_LOST
) {
2055 if (!(sacked
& (TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))) {
2058 if (tcp_retransmit_skb(sk
, skb
)) {
2059 tp
->retransmit_skb_hint
= NULL
;
2062 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
2063 mib_idx
= LINUX_MIB_TCPFASTRETRANS
;
2065 mib_idx
= LINUX_MIB_TCPSLOWSTARTRETRANS
;
2066 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2068 if (skb
== tcp_write_queue_head(sk
))
2069 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2070 inet_csk(sk
)->icsk_rto
,
2074 packet_cnt
+= tcp_skb_pcount(skb
);
2075 if (packet_cnt
>= tp
->lost_out
)
2081 /* OK, demanded retransmission is finished. */
2083 /* Forward retransmissions are possible only during Recovery. */
2084 if (icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
2087 /* No forward retransmissions in Reno are possible. */
2088 if (tcp_is_reno(tp
))
2091 /* Yeah, we have to make difficult choice between forward transmission
2092 * and retransmission... Both ways have their merits...
2094 * For now we do not retransmit anything, while we have some new
2095 * segments to send. In the other cases, follow rule 3 for
2096 * NextSeg() specified in RFC3517.
2099 if (tcp_may_send_now(sk
))
2102 /* If nothing is SACKed, highest_sack in the loop won't be valid */
2103 if (!tp
->sacked_out
)
2106 if (tp
->forward_skb_hint
)
2107 skb
= tp
->forward_skb_hint
;
2109 skb
= tcp_write_queue_head(sk
);
2111 tcp_for_write_queue_from(skb
, sk
) {
2112 if (skb
== tcp_send_head(sk
))
2114 tp
->forward_skb_hint
= skb
;
2116 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_highest_sack_seq(tp
)))
2119 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
2122 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
)
2125 /* Ok, retransmit it. */
2126 if (tcp_retransmit_skb(sk
, skb
)) {
2127 tp
->forward_skb_hint
= NULL
;
2131 if (skb
== tcp_write_queue_head(sk
))
2132 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2133 inet_csk(sk
)->icsk_rto
,
2136 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPFORWARDRETRANS
);
2140 /* Send a fin. The caller locks the socket for us. This cannot be
2141 * allowed to fail queueing a FIN frame under any circumstances.
2143 void tcp_send_fin(struct sock
*sk
)
2145 struct tcp_sock
*tp
= tcp_sk(sk
);
2146 struct sk_buff
*skb
= tcp_write_queue_tail(sk
);
2149 /* Optimization, tack on the FIN if we have a queue of
2150 * unsent frames. But be careful about outgoing SACKS
2153 mss_now
= tcp_current_mss(sk
, 1);
2155 if (tcp_send_head(sk
) != NULL
) {
2156 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_FIN
;
2157 TCP_SKB_CB(skb
)->end_seq
++;
2160 /* Socket is locked, keep trying until memory is available. */
2162 skb
= alloc_skb_fclone(MAX_TCP_HEADER
, GFP_KERNEL
);
2168 /* Reserve space for headers and prepare control bits. */
2169 skb_reserve(skb
, MAX_TCP_HEADER
);
2170 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2171 tcp_init_nondata_skb(skb
, tp
->write_seq
,
2172 TCPCB_FLAG_ACK
| TCPCB_FLAG_FIN
);
2173 tcp_queue_skb(sk
, skb
);
2175 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_OFF
);
2178 /* We get here when a process closes a file descriptor (either due to
2179 * an explicit close() or as a byproduct of exit()'ing) and there
2180 * was unread data in the receive queue. This behavior is recommended
2181 * by RFC 2525, section 2.17. -DaveM
2183 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
2185 struct sk_buff
*skb
;
2187 /* NOTE: No TCP options attached and we never retransmit this. */
2188 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
2190 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
2194 /* Reserve space for headers and prepare control bits. */
2195 skb_reserve(skb
, MAX_TCP_HEADER
);
2196 tcp_init_nondata_skb(skb
, tcp_acceptable_seq(sk
),
2197 TCPCB_FLAG_ACK
| TCPCB_FLAG_RST
);
2199 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2200 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
2201 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
2203 TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTRSTS
);
2206 /* WARNING: This routine must only be called when we have already sent
2207 * a SYN packet that crossed the incoming SYN that caused this routine
2208 * to get called. If this assumption fails then the initial rcv_wnd
2209 * and rcv_wscale values will not be correct.
2211 int tcp_send_synack(struct sock
*sk
)
2213 struct sk_buff
*skb
;
2215 skb
= tcp_write_queue_head(sk
);
2216 if (skb
== NULL
|| !(TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_SYN
)) {
2217 printk(KERN_DEBUG
"tcp_send_synack: wrong queue state\n");
2220 if (!(TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_ACK
)) {
2221 if (skb_cloned(skb
)) {
2222 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
2225 tcp_unlink_write_queue(skb
, sk
);
2226 skb_header_release(nskb
);
2227 __tcp_add_write_queue_head(sk
, nskb
);
2228 sk_wmem_free_skb(sk
, skb
);
2229 sk
->sk_wmem_queued
+= nskb
->truesize
;
2230 sk_mem_charge(sk
, nskb
->truesize
);
2234 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_ACK
;
2235 TCP_ECN_send_synack(tcp_sk(sk
), skb
);
2237 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2238 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2242 * Prepare a SYN-ACK.
2244 struct sk_buff
*tcp_make_synack(struct sock
*sk
, struct dst_entry
*dst
,
2245 struct request_sock
*req
)
2247 struct inet_request_sock
*ireq
= inet_rsk(req
);
2248 struct tcp_sock
*tp
= tcp_sk(sk
);
2250 int tcp_header_size
;
2251 struct tcp_out_options opts
;
2252 struct sk_buff
*skb
;
2253 struct tcp_md5sig_key
*md5
;
2254 __u8
*md5_hash_location
;
2257 skb
= sock_wmalloc(sk
, MAX_TCP_HEADER
+ 15, 1, GFP_ATOMIC
);
2261 /* Reserve space for headers. */
2262 skb_reserve(skb
, MAX_TCP_HEADER
);
2264 skb
->dst
= dst_clone(dst
);
2266 mss
= dst_metric(dst
, RTAX_ADVMSS
);
2267 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< mss
)
2268 mss
= tp
->rx_opt
.user_mss
;
2270 if (req
->rcv_wnd
== 0) { /* ignored for retransmitted syns */
2272 /* Set this up on the first call only */
2273 req
->window_clamp
= tp
->window_clamp
? : dst_metric(dst
, RTAX_WINDOW
);
2274 /* tcp_full_space because it is guaranteed to be the first packet */
2275 tcp_select_initial_window(tcp_full_space(sk
),
2276 mss
- (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0),
2281 ireq
->rcv_wscale
= rcv_wscale
;
2284 memset(&opts
, 0, sizeof(opts
));
2285 #ifdef CONFIG_SYN_COOKIES
2286 if (unlikely(req
->cookie_ts
))
2287 TCP_SKB_CB(skb
)->when
= cookie_init_timestamp(req
);
2290 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2291 tcp_header_size
= tcp_synack_options(sk
, req
, mss
,
2293 sizeof(struct tcphdr
);
2295 skb_push(skb
, tcp_header_size
);
2296 skb_reset_transport_header(skb
);
2299 memset(th
, 0, sizeof(struct tcphdr
));
2302 TCP_ECN_make_synack(req
, th
);
2303 th
->source
= inet_sk(sk
)->sport
;
2304 th
->dest
= ireq
->rmt_port
;
2305 /* Setting of flags are superfluous here for callers (and ECE is
2306 * not even correctly set)
2308 tcp_init_nondata_skb(skb
, tcp_rsk(req
)->snt_isn
,
2309 TCPCB_FLAG_SYN
| TCPCB_FLAG_ACK
);
2310 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
2311 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_isn
+ 1);
2313 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2314 th
->window
= htons(min(req
->rcv_wnd
, 65535U));
2315 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
, &md5_hash_location
);
2316 th
->doff
= (tcp_header_size
>> 2);
2317 TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
);
2319 #ifdef CONFIG_TCP_MD5SIG
2320 /* Okay, we have all we need - do the md5 hash if needed */
2322 tp
->af_specific
->calc_md5_hash(md5_hash_location
,
2323 md5
, NULL
, req
, skb
);
2331 * Do all connect socket setups that can be done AF independent.
2333 static void tcp_connect_init(struct sock
*sk
)
2335 struct dst_entry
*dst
= __sk_dst_get(sk
);
2336 struct tcp_sock
*tp
= tcp_sk(sk
);
2339 /* We'll fix this up when we get a response from the other end.
2340 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2342 tp
->tcp_header_len
= sizeof(struct tcphdr
) +
2343 (sysctl_tcp_timestamps
? TCPOLEN_TSTAMP_ALIGNED
: 0);
2345 #ifdef CONFIG_TCP_MD5SIG
2346 if (tp
->af_specific
->md5_lookup(sk
, sk
) != NULL
)
2347 tp
->tcp_header_len
+= TCPOLEN_MD5SIG_ALIGNED
;
2350 /* If user gave his TCP_MAXSEG, record it to clamp */
2351 if (tp
->rx_opt
.user_mss
)
2352 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
2355 tcp_sync_mss(sk
, dst_mtu(dst
));
2357 if (!tp
->window_clamp
)
2358 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
2359 tp
->advmss
= dst_metric(dst
, RTAX_ADVMSS
);
2360 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< tp
->advmss
)
2361 tp
->advmss
= tp
->rx_opt
.user_mss
;
2363 tcp_initialize_rcv_mss(sk
);
2365 tcp_select_initial_window(tcp_full_space(sk
),
2366 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
2369 sysctl_tcp_window_scaling
,
2372 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
2373 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
2376 sock_reset_flag(sk
, SOCK_DONE
);
2378 tcp_init_wl(tp
, tp
->write_seq
, 0);
2379 tp
->snd_una
= tp
->write_seq
;
2380 tp
->snd_sml
= tp
->write_seq
;
2385 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
2386 inet_csk(sk
)->icsk_retransmits
= 0;
2387 tcp_clear_retrans(tp
);
2391 * Build a SYN and send it off.
2393 int tcp_connect(struct sock
*sk
)
2395 struct tcp_sock
*tp
= tcp_sk(sk
);
2396 struct sk_buff
*buff
;
2398 tcp_connect_init(sk
);
2400 buff
= alloc_skb_fclone(MAX_TCP_HEADER
+ 15, sk
->sk_allocation
);
2401 if (unlikely(buff
== NULL
))
2404 /* Reserve space for headers. */
2405 skb_reserve(buff
, MAX_TCP_HEADER
);
2407 tp
->snd_nxt
= tp
->write_seq
;
2408 tcp_init_nondata_skb(buff
, tp
->write_seq
++, TCPCB_FLAG_SYN
);
2409 TCP_ECN_send_syn(sk
, buff
);
2412 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2413 tp
->retrans_stamp
= TCP_SKB_CB(buff
)->when
;
2414 skb_header_release(buff
);
2415 __tcp_add_write_queue_tail(sk
, buff
);
2416 sk
->sk_wmem_queued
+= buff
->truesize
;
2417 sk_mem_charge(sk
, buff
->truesize
);
2418 tp
->packets_out
+= tcp_skb_pcount(buff
);
2419 tcp_transmit_skb(sk
, buff
, 1, GFP_KERNEL
);
2421 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2422 * in order to make this packet get counted in tcpOutSegs.
2424 tp
->snd_nxt
= tp
->write_seq
;
2425 tp
->pushed_seq
= tp
->write_seq
;
2426 TCP_INC_STATS(sock_net(sk
), TCP_MIB_ACTIVEOPENS
);
2428 /* Timer for repeating the SYN until an answer. */
2429 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2430 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2434 /* Send out a delayed ack, the caller does the policy checking
2435 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2438 void tcp_send_delayed_ack(struct sock
*sk
)
2440 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2441 int ato
= icsk
->icsk_ack
.ato
;
2442 unsigned long timeout
;
2444 if (ato
> TCP_DELACK_MIN
) {
2445 const struct tcp_sock
*tp
= tcp_sk(sk
);
2446 int max_ato
= HZ
/ 2;
2448 if (icsk
->icsk_ack
.pingpong
||
2449 (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
2450 max_ato
= TCP_DELACK_MAX
;
2452 /* Slow path, intersegment interval is "high". */
2454 /* If some rtt estimate is known, use it to bound delayed ack.
2455 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2459 int rtt
= max(tp
->srtt
>> 3, TCP_DELACK_MIN
);
2465 ato
= min(ato
, max_ato
);
2468 /* Stay within the limit we were given */
2469 timeout
= jiffies
+ ato
;
2471 /* Use new timeout only if there wasn't a older one earlier. */
2472 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
2473 /* If delack timer was blocked or is about to expire,
2476 if (icsk
->icsk_ack
.blocked
||
2477 time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
2482 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
2483 timeout
= icsk
->icsk_ack
.timeout
;
2485 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
2486 icsk
->icsk_ack
.timeout
= timeout
;
2487 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
2490 /* This routine sends an ack and also updates the window. */
2491 void tcp_send_ack(struct sock
*sk
)
2493 struct sk_buff
*buff
;
2495 /* If we have been reset, we may not send again. */
2496 if (sk
->sk_state
== TCP_CLOSE
)
2499 /* We are not putting this on the write queue, so
2500 * tcp_transmit_skb() will set the ownership to this
2503 buff
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2505 inet_csk_schedule_ack(sk
);
2506 inet_csk(sk
)->icsk_ack
.ato
= TCP_ATO_MIN
;
2507 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
2508 TCP_DELACK_MAX
, TCP_RTO_MAX
);
2512 /* Reserve space for headers and prepare control bits. */
2513 skb_reserve(buff
, MAX_TCP_HEADER
);
2514 tcp_init_nondata_skb(buff
, tcp_acceptable_seq(sk
), TCPCB_FLAG_ACK
);
2516 /* Send it off, this clears delayed acks for us. */
2517 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2518 tcp_transmit_skb(sk
, buff
, 0, GFP_ATOMIC
);
2521 /* This routine sends a packet with an out of date sequence
2522 * number. It assumes the other end will try to ack it.
2524 * Question: what should we make while urgent mode?
2525 * 4.4BSD forces sending single byte of data. We cannot send
2526 * out of window data, because we have SND.NXT==SND.MAX...
2528 * Current solution: to send TWO zero-length segments in urgent mode:
2529 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2530 * out-of-date with SND.UNA-1 to probe window.
2532 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
)
2534 struct tcp_sock
*tp
= tcp_sk(sk
);
2535 struct sk_buff
*skb
;
2537 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2538 skb
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2542 /* Reserve space for headers and set control bits. */
2543 skb_reserve(skb
, MAX_TCP_HEADER
);
2544 /* Use a previous sequence. This should cause the other
2545 * end to send an ack. Don't queue or clone SKB, just
2548 tcp_init_nondata_skb(skb
, tp
->snd_una
- !urgent
, TCPCB_FLAG_ACK
);
2549 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2550 return tcp_transmit_skb(sk
, skb
, 0, GFP_ATOMIC
);
2553 int tcp_write_wakeup(struct sock
*sk
)
2555 struct tcp_sock
*tp
= tcp_sk(sk
);
2556 struct sk_buff
*skb
;
2558 if (sk
->sk_state
== TCP_CLOSE
)
2561 if ((skb
= tcp_send_head(sk
)) != NULL
&&
2562 before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
))) {
2564 unsigned int mss
= tcp_current_mss(sk
, 0);
2565 unsigned int seg_size
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
2567 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
2568 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
2570 /* We are probing the opening of a window
2571 * but the window size is != 0
2572 * must have been a result SWS avoidance ( sender )
2574 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
2576 seg_size
= min(seg_size
, mss
);
2577 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2578 if (tcp_fragment(sk
, skb
, seg_size
, mss
))
2580 } else if (!tcp_skb_pcount(skb
))
2581 tcp_set_skb_tso_segs(sk
, skb
, mss
);
2583 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2584 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2585 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2587 tcp_event_new_data_sent(sk
, skb
);
2591 between(tp
->snd_up
, tp
->snd_una
+ 1, tp
->snd_una
+ 0xFFFF))
2592 tcp_xmit_probe_skb(sk
, 1);
2593 return tcp_xmit_probe_skb(sk
, 0);
2597 /* A window probe timeout has occurred. If window is not closed send
2598 * a partial packet else a zero probe.
2600 void tcp_send_probe0(struct sock
*sk
)
2602 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2603 struct tcp_sock
*tp
= tcp_sk(sk
);
2606 err
= tcp_write_wakeup(sk
);
2608 if (tp
->packets_out
|| !tcp_send_head(sk
)) {
2609 /* Cancel probe timer, if it is not required. */
2610 icsk
->icsk_probes_out
= 0;
2611 icsk
->icsk_backoff
= 0;
2616 if (icsk
->icsk_backoff
< sysctl_tcp_retries2
)
2617 icsk
->icsk_backoff
++;
2618 icsk
->icsk_probes_out
++;
2619 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2620 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
2623 /* If packet was not sent due to local congestion,
2624 * do not backoff and do not remember icsk_probes_out.
2625 * Let local senders to fight for local resources.
2627 * Use accumulated backoff yet.
2629 if (!icsk
->icsk_probes_out
)
2630 icsk
->icsk_probes_out
= 1;
2631 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2632 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
,
2633 TCP_RESOURCE_PROBE_INTERVAL
),
2638 EXPORT_SYMBOL(tcp_select_initial_window
);
2639 EXPORT_SYMBOL(tcp_connect
);
2640 EXPORT_SYMBOL(tcp_make_synack
);
2641 EXPORT_SYMBOL(tcp_simple_retransmit
);
2642 EXPORT_SYMBOL(tcp_sync_mss
);
2643 EXPORT_SYMBOL(tcp_mtup_init
);