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).
8 * Version: $Id: tcp_output.c,v 1.146 2002/02/01 22:01:04 davem Exp $
11 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
12 * Mark Evans, <evansmp@uhura.aston.ac.uk>
13 * Corey Minyard <wf-rch!minyard@relay.EU.net>
14 * Florian La Roche, <flla@stud.uni-sb.de>
15 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
16 * Linus Torvalds, <torvalds@cs.helsinki.fi>
17 * Alan Cox, <gw4pts@gw4pts.ampr.org>
18 * Matthew Dillon, <dillon@apollo.west.oic.com>
19 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
20 * Jorge Cwik, <jorge@laser.satlink.net>
24 * Changes: Pedro Roque : Retransmit queue handled by TCP.
25 * : Fragmentation on mtu decrease
26 * : Segment collapse on retransmit
29 * Linus Torvalds : send_delayed_ack
30 * David S. Miller : Charge memory using the right skb
31 * during syn/ack processing.
32 * David S. Miller : Output engine completely rewritten.
33 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
34 * Cacophonix Gaul : draft-minshall-nagle-01
35 * J Hadi Salim : ECN support
41 #include <linux/compiler.h>
42 #include <linux/module.h>
44 /* People can turn this off for buggy TCP's found in printers etc. */
45 int sysctl_tcp_retrans_collapse __read_mostly
= 1;
47 /* People can turn this on to work with those rare, broken TCPs that
48 * interpret the window field as a signed quantity.
50 int sysctl_tcp_workaround_signed_windows __read_mostly
= 0;
52 /* This limits the percentage of the congestion window which we
53 * will allow a single TSO frame to consume. Building TSO frames
54 * which are too large can cause TCP streams to be bursty.
56 int sysctl_tcp_tso_win_divisor __read_mostly
= 3;
58 int sysctl_tcp_mtu_probing __read_mostly
= 0;
59 int sysctl_tcp_base_mss __read_mostly
= 512;
61 /* By default, RFC2861 behavior. */
62 int sysctl_tcp_slow_start_after_idle __read_mostly
= 1;
64 static void tcp_event_new_data_sent(struct sock
*sk
, struct sk_buff
*skb
)
66 struct tcp_sock
*tp
= tcp_sk(sk
);
67 unsigned int prior_packets
= tp
->packets_out
;
69 tcp_advance_send_head(sk
, skb
);
70 tp
->snd_nxt
= TCP_SKB_CB(skb
)->end_seq
;
72 /* Don't override Nagle indefinately with F-RTO */
73 if (tp
->frto_counter
== 2)
76 tp
->packets_out
+= tcp_skb_pcount(skb
);
78 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
79 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
82 /* SND.NXT, if window was not shrunk.
83 * If window has been shrunk, what should we make? It is not clear at all.
84 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
85 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
86 * invalid. OK, let's make this for now:
88 static inline __u32
tcp_acceptable_seq(struct sock
*sk
)
90 struct tcp_sock
*tp
= tcp_sk(sk
);
92 if (!before(tcp_wnd_end(tp
), tp
->snd_nxt
))
95 return tcp_wnd_end(tp
);
98 /* Calculate mss to advertise in SYN segment.
99 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
101 * 1. It is independent of path mtu.
102 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
103 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
104 * attached devices, because some buggy hosts are confused by
106 * 4. We do not make 3, we advertise MSS, calculated from first
107 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
108 * This may be overridden via information stored in routing table.
109 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
110 * probably even Jumbo".
112 static __u16
tcp_advertise_mss(struct sock
*sk
)
114 struct tcp_sock
*tp
= tcp_sk(sk
);
115 struct dst_entry
*dst
= __sk_dst_get(sk
);
116 int mss
= tp
->advmss
;
118 if (dst
&& dst_metric(dst
, RTAX_ADVMSS
) < mss
) {
119 mss
= dst_metric(dst
, RTAX_ADVMSS
);
126 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
127 * This is the first part of cwnd validation mechanism. */
128 static void tcp_cwnd_restart(struct sock
*sk
, struct dst_entry
*dst
)
130 struct tcp_sock
*tp
= tcp_sk(sk
);
131 s32 delta
= tcp_time_stamp
- tp
->lsndtime
;
132 u32 restart_cwnd
= tcp_init_cwnd(tp
, dst
);
133 u32 cwnd
= tp
->snd_cwnd
;
135 tcp_ca_event(sk
, CA_EVENT_CWND_RESTART
);
137 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
138 restart_cwnd
= min(restart_cwnd
, cwnd
);
140 while ((delta
-= inet_csk(sk
)->icsk_rto
) > 0 && cwnd
> restart_cwnd
)
142 tp
->snd_cwnd
= max(cwnd
, restart_cwnd
);
143 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
144 tp
->snd_cwnd_used
= 0;
147 static void tcp_event_data_sent(struct tcp_sock
*tp
,
148 struct sk_buff
*skb
, struct sock
*sk
)
150 struct inet_connection_sock
*icsk
= inet_csk(sk
);
151 const u32 now
= tcp_time_stamp
;
153 if (sysctl_tcp_slow_start_after_idle
&&
154 (!tp
->packets_out
&& (s32
)(now
- tp
->lsndtime
) > icsk
->icsk_rto
))
155 tcp_cwnd_restart(sk
, __sk_dst_get(sk
));
159 /* If it is a reply for ato after last received
160 * packet, enter pingpong mode.
162 if ((u32
)(now
- icsk
->icsk_ack
.lrcvtime
) < icsk
->icsk_ack
.ato
)
163 icsk
->icsk_ack
.pingpong
= 1;
166 static inline void tcp_event_ack_sent(struct sock
*sk
, unsigned int pkts
)
168 tcp_dec_quickack_mode(sk
, pkts
);
169 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_DACK
);
172 /* Determine a window scaling and initial window to offer.
173 * Based on the assumption that the given amount of space
174 * will be offered. Store the results in the tp structure.
175 * NOTE: for smooth operation initial space offering should
176 * be a multiple of mss if possible. We assume here that mss >= 1.
177 * This MUST be enforced by all callers.
179 void tcp_select_initial_window(int __space
, __u32 mss
,
180 __u32
*rcv_wnd
, __u32
*window_clamp
,
181 int wscale_ok
, __u8
*rcv_wscale
)
183 unsigned int space
= (__space
< 0 ? 0 : __space
);
185 /* If no clamp set the clamp to the max possible scaled window */
186 if (*window_clamp
== 0)
187 (*window_clamp
) = (65535 << 14);
188 space
= min(*window_clamp
, space
);
190 /* Quantize space offering to a multiple of mss if possible. */
192 space
= (space
/ mss
) * mss
;
194 /* NOTE: offering an initial window larger than 32767
195 * will break some buggy TCP stacks. If the admin tells us
196 * it is likely we could be speaking with such a buggy stack
197 * we will truncate our initial window offering to 32K-1
198 * unless the remote has sent us a window scaling option,
199 * which we interpret as a sign the remote TCP is not
200 * misinterpreting the window field as a signed quantity.
202 if (sysctl_tcp_workaround_signed_windows
)
203 (*rcv_wnd
) = min(space
, MAX_TCP_WINDOW
);
209 /* Set window scaling on max possible window
210 * See RFC1323 for an explanation of the limit to 14
212 space
= max_t(u32
, sysctl_tcp_rmem
[2], sysctl_rmem_max
);
213 space
= min_t(u32
, space
, *window_clamp
);
214 while (space
> 65535 && (*rcv_wscale
) < 14) {
220 /* Set initial window to value enough for senders,
221 * following RFC2414. Senders, not following this RFC,
222 * will be satisfied with 2.
224 if (mss
> (1 << *rcv_wscale
)) {
230 if (*rcv_wnd
> init_cwnd
* mss
)
231 *rcv_wnd
= init_cwnd
* mss
;
234 /* Set the clamp no higher than max representable value */
235 (*window_clamp
) = min(65535U << (*rcv_wscale
), *window_clamp
);
238 /* Chose a new window to advertise, update state in tcp_sock for the
239 * socket, and return result with RFC1323 scaling applied. The return
240 * value can be stuffed directly into th->window for an outgoing
243 static u16
tcp_select_window(struct sock
*sk
)
245 struct tcp_sock
*tp
= tcp_sk(sk
);
246 u32 cur_win
= tcp_receive_window(tp
);
247 u32 new_win
= __tcp_select_window(sk
);
249 /* Never shrink the offered window */
250 if (new_win
< cur_win
) {
251 /* Danger Will Robinson!
252 * Don't update rcv_wup/rcv_wnd here or else
253 * we will not be able to advertise a zero
254 * window in time. --DaveM
256 * Relax Will Robinson.
260 tp
->rcv_wnd
= new_win
;
261 tp
->rcv_wup
= tp
->rcv_nxt
;
263 /* Make sure we do not exceed the maximum possible
266 if (!tp
->rx_opt
.rcv_wscale
&& sysctl_tcp_workaround_signed_windows
)
267 new_win
= min(new_win
, MAX_TCP_WINDOW
);
269 new_win
= min(new_win
, (65535U << tp
->rx_opt
.rcv_wscale
));
271 /* RFC1323 scaling applied */
272 new_win
>>= tp
->rx_opt
.rcv_wscale
;
274 /* If we advertise zero window, disable fast path. */
281 static inline void TCP_ECN_send_synack(struct tcp_sock
*tp
, struct sk_buff
*skb
)
283 TCP_SKB_CB(skb
)->flags
&= ~TCPCB_FLAG_CWR
;
284 if (!(tp
->ecn_flags
& TCP_ECN_OK
))
285 TCP_SKB_CB(skb
)->flags
&= ~TCPCB_FLAG_ECE
;
288 static inline void TCP_ECN_send_syn(struct sock
*sk
, struct sk_buff
*skb
)
290 struct tcp_sock
*tp
= tcp_sk(sk
);
293 if (sysctl_tcp_ecn
) {
294 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_ECE
| TCPCB_FLAG_CWR
;
295 tp
->ecn_flags
= TCP_ECN_OK
;
299 static __inline__
void
300 TCP_ECN_make_synack(struct request_sock
*req
, struct tcphdr
*th
)
302 if (inet_rsk(req
)->ecn_ok
)
306 static inline void TCP_ECN_send(struct sock
*sk
, struct sk_buff
*skb
,
309 struct tcp_sock
*tp
= tcp_sk(sk
);
311 if (tp
->ecn_flags
& TCP_ECN_OK
) {
312 /* Not-retransmitted data segment: set ECT and inject CWR. */
313 if (skb
->len
!= tcp_header_len
&&
314 !before(TCP_SKB_CB(skb
)->seq
, tp
->snd_nxt
)) {
316 if (tp
->ecn_flags
& TCP_ECN_QUEUE_CWR
) {
317 tp
->ecn_flags
&= ~TCP_ECN_QUEUE_CWR
;
318 tcp_hdr(skb
)->cwr
= 1;
319 skb_shinfo(skb
)->gso_type
|= SKB_GSO_TCP_ECN
;
322 /* ACK or retransmitted segment: clear ECT|CE */
323 INET_ECN_dontxmit(sk
);
325 if (tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)
326 tcp_hdr(skb
)->ece
= 1;
330 static void tcp_build_and_update_options(__be32
*ptr
, struct tcp_sock
*tp
,
331 __u32 tstamp
, __u8
**md5_hash
)
333 if (tp
->rx_opt
.tstamp_ok
) {
334 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
336 (TCPOPT_TIMESTAMP
<< 8) |
338 *ptr
++ = htonl(tstamp
);
339 *ptr
++ = htonl(tp
->rx_opt
.ts_recent
);
341 if (tp
->rx_opt
.eff_sacks
) {
342 struct tcp_sack_block
*sp
= tp
->rx_opt
.dsack
? tp
->duplicate_sack
: tp
->selective_acks
;
345 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
348 (TCPOLEN_SACK_BASE
+ (tp
->rx_opt
.eff_sacks
*
349 TCPOLEN_SACK_PERBLOCK
)));
351 for (this_sack
= 0; this_sack
< tp
->rx_opt
.eff_sacks
; this_sack
++) {
352 *ptr
++ = htonl(sp
[this_sack
].start_seq
);
353 *ptr
++ = htonl(sp
[this_sack
].end_seq
);
356 if (tp
->rx_opt
.dsack
) {
357 tp
->rx_opt
.dsack
= 0;
358 tp
->rx_opt
.eff_sacks
--;
361 #ifdef CONFIG_TCP_MD5SIG
363 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
365 (TCPOPT_MD5SIG
<< 8) |
367 *md5_hash
= (__u8
*)ptr
;
372 /* Construct a tcp options header for a SYN or SYN_ACK packet.
373 * If this is every changed make sure to change the definition of
374 * MAX_SYN_SIZE to match the new maximum number of options that you
377 * Note - that with the RFC2385 TCP option, we make room for the
378 * 16 byte MD5 hash. This will be filled in later, so the pointer for the
379 * location to be filled is passed back up.
381 static void tcp_syn_build_options(__be32
*ptr
, int mss
, int ts
, int sack
,
382 int offer_wscale
, int wscale
, __u32 tstamp
,
383 __u32 ts_recent
, __u8
**md5_hash
)
385 /* We always get an MSS option.
386 * The option bytes which will be seen in normal data
387 * packets should timestamps be used, must be in the MSS
388 * advertised. But we subtract them from tp->mss_cache so
389 * that calculations in tcp_sendmsg are simpler etc.
390 * So account for this fact here if necessary. If we
391 * don't do this correctly, as a receiver we won't
392 * recognize data packets as being full sized when we
393 * should, and thus we won't abide by the delayed ACK
395 * SACKs don't matter, we never delay an ACK when we
396 * have any of those going out.
398 *ptr
++ = htonl((TCPOPT_MSS
<< 24) | (TCPOLEN_MSS
<< 16) | mss
);
401 *ptr
++ = htonl((TCPOPT_SACK_PERM
<< 24) |
402 (TCPOLEN_SACK_PERM
<< 16) |
403 (TCPOPT_TIMESTAMP
<< 8) |
406 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
408 (TCPOPT_TIMESTAMP
<< 8) |
410 *ptr
++ = htonl(tstamp
); /* TSVAL */
411 *ptr
++ = htonl(ts_recent
); /* TSECR */
413 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
415 (TCPOPT_SACK_PERM
<< 8) |
418 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
419 (TCPOPT_WINDOW
<< 16) |
420 (TCPOLEN_WINDOW
<< 8) |
422 #ifdef CONFIG_TCP_MD5SIG
424 * If MD5 is enabled, then we set the option, and include the size
425 * (always 18). The actual MD5 hash is added just before the
429 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
431 (TCPOPT_MD5SIG
<< 8) |
433 *md5_hash
= (__u8
*)ptr
;
438 /* This routine actually transmits TCP packets queued in by
439 * tcp_do_sendmsg(). This is used by both the initial
440 * transmission and possible later retransmissions.
441 * All SKB's seen here are completely headerless. It is our
442 * job to build the TCP header, and pass the packet down to
443 * IP so it can do the same plus pass the packet off to the
446 * We are working here with either a clone of the original
447 * SKB, or a fresh unique copy made by the retransmit engine.
449 static int tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int clone_it
,
452 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
453 struct inet_sock
*inet
;
455 struct tcp_skb_cb
*tcb
;
457 #ifdef CONFIG_TCP_MD5SIG
458 struct tcp_md5sig_key
*md5
;
459 __u8
*md5_hash_location
;
465 BUG_ON(!skb
|| !tcp_skb_pcount(skb
));
467 /* If congestion control is doing timestamping, we must
468 * take such a timestamp before we potentially clone/copy.
470 if (icsk
->icsk_ca_ops
->flags
& TCP_CONG_RTT_STAMP
)
471 __net_timestamp(skb
);
473 if (likely(clone_it
)) {
474 if (unlikely(skb_cloned(skb
)))
475 skb
= pskb_copy(skb
, gfp_mask
);
477 skb
= skb_clone(skb
, gfp_mask
);
484 tcb
= TCP_SKB_CB(skb
);
485 tcp_header_size
= tp
->tcp_header_len
;
487 #define SYSCTL_FLAG_TSTAMPS 0x1
488 #define SYSCTL_FLAG_WSCALE 0x2
489 #define SYSCTL_FLAG_SACK 0x4
492 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
493 tcp_header_size
= sizeof(struct tcphdr
) + TCPOLEN_MSS
;
494 if (sysctl_tcp_timestamps
) {
495 tcp_header_size
+= TCPOLEN_TSTAMP_ALIGNED
;
496 sysctl_flags
|= SYSCTL_FLAG_TSTAMPS
;
498 if (sysctl_tcp_window_scaling
) {
499 tcp_header_size
+= TCPOLEN_WSCALE_ALIGNED
;
500 sysctl_flags
|= SYSCTL_FLAG_WSCALE
;
502 if (sysctl_tcp_sack
) {
503 sysctl_flags
|= SYSCTL_FLAG_SACK
;
504 if (!(sysctl_flags
& SYSCTL_FLAG_TSTAMPS
))
505 tcp_header_size
+= TCPOLEN_SACKPERM_ALIGNED
;
507 } else if (unlikely(tp
->rx_opt
.eff_sacks
)) {
508 /* A SACK is 2 pad bytes, a 2 byte header, plus
509 * 2 32-bit sequence numbers for each SACK block.
511 tcp_header_size
+= (TCPOLEN_SACK_BASE_ALIGNED
+
512 (tp
->rx_opt
.eff_sacks
*
513 TCPOLEN_SACK_PERBLOCK
));
516 if (tcp_packets_in_flight(tp
) == 0)
517 tcp_ca_event(sk
, CA_EVENT_TX_START
);
519 #ifdef CONFIG_TCP_MD5SIG
521 * Are we doing MD5 on this segment? If so - make
524 md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
526 tcp_header_size
+= TCPOLEN_MD5SIG_ALIGNED
;
529 skb_push(skb
, tcp_header_size
);
530 skb_reset_transport_header(skb
);
531 skb_set_owner_w(skb
, sk
);
533 /* Build TCP header and checksum it. */
535 th
->source
= inet
->sport
;
536 th
->dest
= inet
->dport
;
537 th
->seq
= htonl(tcb
->seq
);
538 th
->ack_seq
= htonl(tp
->rcv_nxt
);
539 *(((__be16
*)th
) + 6) = htons(((tcp_header_size
>> 2) << 12) |
542 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
543 /* RFC1323: The window in SYN & SYN/ACK segments
546 th
->window
= htons(min(tp
->rcv_wnd
, 65535U));
548 th
->window
= htons(tcp_select_window(sk
));
553 if (unlikely(tp
->urg_mode
&&
554 between(tp
->snd_up
, tcb
->seq
+ 1, tcb
->seq
+ 0xFFFF))) {
555 th
->urg_ptr
= htons(tp
->snd_up
- tcb
->seq
);
559 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
560 tcp_syn_build_options((__be32
*)(th
+ 1),
561 tcp_advertise_mss(sk
),
562 (sysctl_flags
& SYSCTL_FLAG_TSTAMPS
),
563 (sysctl_flags
& SYSCTL_FLAG_SACK
),
564 (sysctl_flags
& SYSCTL_FLAG_WSCALE
),
565 tp
->rx_opt
.rcv_wscale
,
567 tp
->rx_opt
.ts_recent
,
569 #ifdef CONFIG_TCP_MD5SIG
570 md5
? &md5_hash_location
:
574 tcp_build_and_update_options((__be32
*)(th
+ 1),
576 #ifdef CONFIG_TCP_MD5SIG
577 md5
? &md5_hash_location
:
580 TCP_ECN_send(sk
, skb
, tcp_header_size
);
583 #ifdef CONFIG_TCP_MD5SIG
584 /* Calculate the MD5 hash, as we have all we need now */
586 tp
->af_specific
->calc_md5_hash(md5_hash_location
,
595 icsk
->icsk_af_ops
->send_check(sk
, skb
->len
, skb
);
597 if (likely(tcb
->flags
& TCPCB_FLAG_ACK
))
598 tcp_event_ack_sent(sk
, tcp_skb_pcount(skb
));
600 if (skb
->len
!= tcp_header_size
)
601 tcp_event_data_sent(tp
, skb
, sk
);
603 if (after(tcb
->end_seq
, tp
->snd_nxt
) || tcb
->seq
== tcb
->end_seq
)
604 TCP_INC_STATS(TCP_MIB_OUTSEGS
);
606 err
= icsk
->icsk_af_ops
->queue_xmit(skb
, 0);
607 if (likely(err
<= 0))
610 tcp_enter_cwr(sk
, 1);
612 return net_xmit_eval(err
);
614 #undef SYSCTL_FLAG_TSTAMPS
615 #undef SYSCTL_FLAG_WSCALE
616 #undef SYSCTL_FLAG_SACK
619 /* This routine just queue's the buffer
621 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
622 * otherwise socket can stall.
624 static void tcp_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
626 struct tcp_sock
*tp
= tcp_sk(sk
);
628 /* Advance write_seq and place onto the write_queue. */
629 tp
->write_seq
= TCP_SKB_CB(skb
)->end_seq
;
630 skb_header_release(skb
);
631 tcp_add_write_queue_tail(sk
, skb
);
632 sk
->sk_wmem_queued
+= skb
->truesize
;
633 sk_mem_charge(sk
, skb
->truesize
);
636 static void tcp_set_skb_tso_segs(struct sock
*sk
, struct sk_buff
*skb
,
637 unsigned int mss_now
)
639 if (skb
->len
<= mss_now
|| !sk_can_gso(sk
)) {
640 /* Avoid the costly divide in the normal
643 skb_shinfo(skb
)->gso_segs
= 1;
644 skb_shinfo(skb
)->gso_size
= 0;
645 skb_shinfo(skb
)->gso_type
= 0;
647 skb_shinfo(skb
)->gso_segs
= DIV_ROUND_UP(skb
->len
, mss_now
);
648 skb_shinfo(skb
)->gso_size
= mss_now
;
649 skb_shinfo(skb
)->gso_type
= sk
->sk_gso_type
;
653 /* When a modification to fackets out becomes necessary, we need to check
654 * skb is counted to fackets_out or not.
656 static void tcp_adjust_fackets_out(struct sock
*sk
, struct sk_buff
*skb
,
659 struct tcp_sock
*tp
= tcp_sk(sk
);
661 if (!tp
->sacked_out
|| tcp_is_reno(tp
))
664 if (after(tcp_highest_sack_seq(tp
), TCP_SKB_CB(skb
)->seq
))
665 tp
->fackets_out
-= decr
;
668 /* Function to create two new TCP segments. Shrinks the given segment
669 * to the specified size and appends a new segment with the rest of the
670 * packet to the list. This won't be called frequently, I hope.
671 * Remember, these are still headerless SKBs at this point.
673 int tcp_fragment(struct sock
*sk
, struct sk_buff
*skb
, u32 len
,
674 unsigned int mss_now
)
676 struct tcp_sock
*tp
= tcp_sk(sk
);
677 struct sk_buff
*buff
;
678 int nsize
, old_factor
;
682 BUG_ON(len
> skb
->len
);
684 tcp_clear_retrans_hints_partial(tp
);
685 nsize
= skb_headlen(skb
) - len
;
689 if (skb_cloned(skb
) &&
690 skb_is_nonlinear(skb
) &&
691 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
694 /* Get a new skb... force flag on. */
695 buff
= sk_stream_alloc_skb(sk
, nsize
, GFP_ATOMIC
);
697 return -ENOMEM
; /* We'll just try again later. */
699 sk
->sk_wmem_queued
+= buff
->truesize
;
700 sk_mem_charge(sk
, buff
->truesize
);
701 nlen
= skb
->len
- len
- nsize
;
702 buff
->truesize
+= nlen
;
703 skb
->truesize
-= nlen
;
705 /* Correct the sequence numbers. */
706 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
707 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
708 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
710 /* PSH and FIN should only be set in the second packet. */
711 flags
= TCP_SKB_CB(skb
)->flags
;
712 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
| TCPCB_FLAG_PSH
);
713 TCP_SKB_CB(buff
)->flags
= flags
;
714 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
716 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_PARTIAL
) {
717 /* Copy and checksum data tail into the new buffer. */
718 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
,
719 skb_put(buff
, nsize
),
724 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
726 skb
->ip_summed
= CHECKSUM_PARTIAL
;
727 skb_split(skb
, buff
, len
);
730 buff
->ip_summed
= skb
->ip_summed
;
732 /* Looks stupid, but our code really uses when of
733 * skbs, which it never sent before. --ANK
735 TCP_SKB_CB(buff
)->when
= TCP_SKB_CB(skb
)->when
;
736 buff
->tstamp
= skb
->tstamp
;
738 old_factor
= tcp_skb_pcount(skb
);
740 /* Fix up tso_factor for both original and new SKB. */
741 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
742 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
744 /* If this packet has been sent out already, we must
745 * adjust the various packet counters.
747 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
748 int diff
= old_factor
- tcp_skb_pcount(skb
) -
749 tcp_skb_pcount(buff
);
751 tp
->packets_out
-= diff
;
753 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
754 tp
->sacked_out
-= diff
;
755 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
756 tp
->retrans_out
-= diff
;
758 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
)
759 tp
->lost_out
-= diff
;
761 /* Adjust Reno SACK estimate. */
762 if (tcp_is_reno(tp
) && diff
> 0) {
763 tcp_dec_pcount_approx_int(&tp
->sacked_out
, diff
);
764 tcp_verify_left_out(tp
);
766 tcp_adjust_fackets_out(sk
, skb
, diff
);
769 /* Link BUFF into the send queue. */
770 skb_header_release(buff
);
771 tcp_insert_write_queue_after(skb
, buff
, sk
);
776 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
777 * eventually). The difference is that pulled data not copied, but
778 * immediately discarded.
780 static void __pskb_trim_head(struct sk_buff
*skb
, int len
)
786 for (i
= 0; i
< skb_shinfo(skb
)->nr_frags
; i
++) {
787 if (skb_shinfo(skb
)->frags
[i
].size
<= eat
) {
788 put_page(skb_shinfo(skb
)->frags
[i
].page
);
789 eat
-= skb_shinfo(skb
)->frags
[i
].size
;
791 skb_shinfo(skb
)->frags
[k
] = skb_shinfo(skb
)->frags
[i
];
793 skb_shinfo(skb
)->frags
[k
].page_offset
+= eat
;
794 skb_shinfo(skb
)->frags
[k
].size
-= eat
;
800 skb_shinfo(skb
)->nr_frags
= k
;
802 skb_reset_tail_pointer(skb
);
803 skb
->data_len
-= len
;
804 skb
->len
= skb
->data_len
;
807 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
809 if (skb_cloned(skb
) && pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
812 /* If len == headlen, we avoid __skb_pull to preserve alignment. */
813 if (unlikely(len
< skb_headlen(skb
)))
814 __skb_pull(skb
, len
);
816 __pskb_trim_head(skb
, len
- skb_headlen(skb
));
818 TCP_SKB_CB(skb
)->seq
+= len
;
819 skb
->ip_summed
= CHECKSUM_PARTIAL
;
821 skb
->truesize
-= len
;
822 sk
->sk_wmem_queued
-= len
;
823 sk_mem_uncharge(sk
, len
);
824 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
826 /* Any change of skb->len requires recalculation of tso
829 if (tcp_skb_pcount(skb
) > 1)
830 tcp_set_skb_tso_segs(sk
, skb
, tcp_current_mss(sk
, 1));
835 /* Not accounting for SACKs here. */
836 int tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
838 struct tcp_sock
*tp
= tcp_sk(sk
);
839 struct inet_connection_sock
*icsk
= inet_csk(sk
);
842 /* Calculate base mss without TCP options:
843 It is MMS_S - sizeof(tcphdr) of rfc1122
845 mss_now
= pmtu
- icsk
->icsk_af_ops
->net_header_len
- sizeof(struct tcphdr
);
847 /* Clamp it (mss_clamp does not include tcp options) */
848 if (mss_now
> tp
->rx_opt
.mss_clamp
)
849 mss_now
= tp
->rx_opt
.mss_clamp
;
851 /* Now subtract optional transport overhead */
852 mss_now
-= icsk
->icsk_ext_hdr_len
;
854 /* Then reserve room for full set of TCP options and 8 bytes of data */
858 /* Now subtract TCP options size, not including SACKs */
859 mss_now
-= tp
->tcp_header_len
- sizeof(struct tcphdr
);
864 /* Inverse of above */
865 int tcp_mss_to_mtu(struct sock
*sk
, int mss
)
867 struct tcp_sock
*tp
= tcp_sk(sk
);
868 struct inet_connection_sock
*icsk
= inet_csk(sk
);
873 icsk
->icsk_ext_hdr_len
+
874 icsk
->icsk_af_ops
->net_header_len
;
879 void tcp_mtup_init(struct sock
*sk
)
881 struct tcp_sock
*tp
= tcp_sk(sk
);
882 struct inet_connection_sock
*icsk
= inet_csk(sk
);
884 icsk
->icsk_mtup
.enabled
= sysctl_tcp_mtu_probing
> 1;
885 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+ sizeof(struct tcphdr
) +
886 icsk
->icsk_af_ops
->net_header_len
;
887 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, sysctl_tcp_base_mss
);
888 icsk
->icsk_mtup
.probe_size
= 0;
891 /* Bound MSS / TSO packet size with the half of the window */
892 static int tcp_bound_to_half_wnd(struct tcp_sock
*tp
, int pktsize
)
894 if (tp
->max_window
&& pktsize
> (tp
->max_window
>> 1))
895 return max(tp
->max_window
>> 1, 68U - tp
->tcp_header_len
);
900 /* This function synchronize snd mss to current pmtu/exthdr set.
902 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
903 for TCP options, but includes only bare TCP header.
905 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
906 It is minimum of user_mss and mss received with SYN.
907 It also does not include TCP options.
909 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
911 tp->mss_cache is current effective sending mss, including
912 all tcp options except for SACKs. It is evaluated,
913 taking into account current pmtu, but never exceeds
914 tp->rx_opt.mss_clamp.
916 NOTE1. rfc1122 clearly states that advertised MSS
917 DOES NOT include either tcp or ip options.
919 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
920 are READ ONLY outside this function. --ANK (980731)
922 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
924 struct tcp_sock
*tp
= tcp_sk(sk
);
925 struct inet_connection_sock
*icsk
= inet_csk(sk
);
928 if (icsk
->icsk_mtup
.search_high
> pmtu
)
929 icsk
->icsk_mtup
.search_high
= pmtu
;
931 mss_now
= tcp_mtu_to_mss(sk
, pmtu
);
932 mss_now
= tcp_bound_to_half_wnd(tp
, mss_now
);
934 /* And store cached results */
935 icsk
->icsk_pmtu_cookie
= pmtu
;
936 if (icsk
->icsk_mtup
.enabled
)
937 mss_now
= min(mss_now
, tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
));
938 tp
->mss_cache
= mss_now
;
943 /* Compute the current effective MSS, taking SACKs and IP options,
944 * and even PMTU discovery events into account.
946 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
947 * cannot be large. However, taking into account rare use of URG, this
950 unsigned int tcp_current_mss(struct sock
*sk
, int large_allowed
)
952 struct tcp_sock
*tp
= tcp_sk(sk
);
953 struct dst_entry
*dst
= __sk_dst_get(sk
);
958 mss_now
= tp
->mss_cache
;
960 if (large_allowed
&& sk_can_gso(sk
) && !tp
->urg_mode
)
964 u32 mtu
= dst_mtu(dst
);
965 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
966 mss_now
= tcp_sync_mss(sk
, mtu
);
969 if (tp
->rx_opt
.eff_sacks
)
970 mss_now
-= (TCPOLEN_SACK_BASE_ALIGNED
+
971 (tp
->rx_opt
.eff_sacks
* TCPOLEN_SACK_PERBLOCK
));
973 #ifdef CONFIG_TCP_MD5SIG
974 if (tp
->af_specific
->md5_lookup(sk
, sk
))
975 mss_now
-= TCPOLEN_MD5SIG_ALIGNED
;
978 xmit_size_goal
= mss_now
;
981 xmit_size_goal
= (65535 -
982 inet_csk(sk
)->icsk_af_ops
->net_header_len
-
983 inet_csk(sk
)->icsk_ext_hdr_len
-
986 xmit_size_goal
= tcp_bound_to_half_wnd(tp
, xmit_size_goal
);
987 xmit_size_goal
-= (xmit_size_goal
% mss_now
);
989 tp
->xmit_size_goal
= xmit_size_goal
;
994 /* Congestion window validation. (RFC2861) */
995 static void tcp_cwnd_validate(struct sock
*sk
)
997 struct tcp_sock
*tp
= tcp_sk(sk
);
999 if (tp
->packets_out
>= tp
->snd_cwnd
) {
1000 /* Network is feed fully. */
1001 tp
->snd_cwnd_used
= 0;
1002 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1004 /* Network starves. */
1005 if (tp
->packets_out
> tp
->snd_cwnd_used
)
1006 tp
->snd_cwnd_used
= tp
->packets_out
;
1008 if (sysctl_tcp_slow_start_after_idle
&&
1009 (s32
)(tcp_time_stamp
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
)
1010 tcp_cwnd_application_limited(sk
);
1014 /* Returns the portion of skb which can be sent right away without
1015 * introducing MSS oddities to segment boundaries. In rare cases where
1016 * mss_now != mss_cache, we will request caller to create a small skb
1017 * per input skb which could be mostly avoided here (if desired).
1019 static unsigned int tcp_mss_split_point(struct sock
*sk
, struct sk_buff
*skb
,
1020 unsigned int mss_now
, unsigned int cwnd
)
1022 struct tcp_sock
*tp
= tcp_sk(sk
);
1023 u32 needed
, window
, cwnd_len
;
1025 window
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1026 cwnd_len
= mss_now
* cwnd
;
1028 if (likely(cwnd_len
<= window
&& skb
!= tcp_write_queue_tail(sk
)))
1031 if (skb
== tcp_write_queue_tail(sk
) && cwnd_len
<= skb
->len
)
1034 needed
= min(skb
->len
, window
);
1035 return needed
- needed
% mss_now
;
1038 /* Can at least one segment of SKB be sent right now, according to the
1039 * congestion window rules? If so, return how many segments are allowed.
1041 static inline unsigned int tcp_cwnd_test(struct tcp_sock
*tp
,
1042 struct sk_buff
*skb
)
1044 u32 in_flight
, cwnd
;
1046 /* Don't be strict about the congestion window for the final FIN. */
1047 if ((TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
) &&
1048 tcp_skb_pcount(skb
) == 1)
1051 in_flight
= tcp_packets_in_flight(tp
);
1052 cwnd
= tp
->snd_cwnd
;
1053 if (in_flight
< cwnd
)
1054 return (cwnd
- in_flight
);
1059 /* This must be invoked the first time we consider transmitting
1060 * SKB onto the wire.
1062 static int tcp_init_tso_segs(struct sock
*sk
, struct sk_buff
*skb
,
1063 unsigned int mss_now
)
1065 int tso_segs
= tcp_skb_pcount(skb
);
1067 if (!tso_segs
|| (tso_segs
> 1 && tcp_skb_mss(skb
) != mss_now
)) {
1068 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1069 tso_segs
= tcp_skb_pcount(skb
);
1074 static inline int tcp_minshall_check(const struct tcp_sock
*tp
)
1076 return after(tp
->snd_sml
,tp
->snd_una
) &&
1077 !after(tp
->snd_sml
, tp
->snd_nxt
);
1080 /* Return 0, if packet can be sent now without violation Nagle's rules:
1081 * 1. It is full sized.
1082 * 2. Or it contains FIN. (already checked by caller)
1083 * 3. Or TCP_NODELAY was set.
1084 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1085 * With Minshall's modification: all sent small packets are ACKed.
1087 static inline int tcp_nagle_check(const struct tcp_sock
*tp
,
1088 const struct sk_buff
*skb
,
1089 unsigned mss_now
, int nonagle
)
1091 return (skb
->len
< mss_now
&&
1092 ((nonagle
& TCP_NAGLE_CORK
) ||
1093 (!nonagle
&& tp
->packets_out
&& tcp_minshall_check(tp
))));
1096 /* Return non-zero if the Nagle test allows this packet to be
1099 static inline int tcp_nagle_test(struct tcp_sock
*tp
, struct sk_buff
*skb
,
1100 unsigned int cur_mss
, int nonagle
)
1102 /* Nagle rule does not apply to frames, which sit in the middle of the
1103 * write_queue (they have no chances to get new data).
1105 * This is implemented in the callers, where they modify the 'nonagle'
1106 * argument based upon the location of SKB in the send queue.
1108 if (nonagle
& TCP_NAGLE_PUSH
)
1111 /* Don't use the nagle rule for urgent data (or for the final FIN).
1112 * Nagle can be ignored during F-RTO too (see RFC4138).
1114 if (tp
->urg_mode
|| (tp
->frto_counter
== 2) ||
1115 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
))
1118 if (!tcp_nagle_check(tp
, skb
, cur_mss
, nonagle
))
1124 /* Does at least the first segment of SKB fit into the send window? */
1125 static inline int tcp_snd_wnd_test(struct tcp_sock
*tp
, struct sk_buff
*skb
,
1126 unsigned int cur_mss
)
1128 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1130 if (skb
->len
> cur_mss
)
1131 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
1133 return !after(end_seq
, tcp_wnd_end(tp
));
1136 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1137 * should be put on the wire right now. If so, it returns the number of
1138 * packets allowed by the congestion window.
1140 static unsigned int tcp_snd_test(struct sock
*sk
, struct sk_buff
*skb
,
1141 unsigned int cur_mss
, int nonagle
)
1143 struct tcp_sock
*tp
= tcp_sk(sk
);
1144 unsigned int cwnd_quota
;
1146 tcp_init_tso_segs(sk
, skb
, cur_mss
);
1148 if (!tcp_nagle_test(tp
, skb
, cur_mss
, nonagle
))
1151 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1152 if (cwnd_quota
&& !tcp_snd_wnd_test(tp
, skb
, cur_mss
))
1158 int tcp_may_send_now(struct sock
*sk
)
1160 struct tcp_sock
*tp
= tcp_sk(sk
);
1161 struct sk_buff
*skb
= tcp_send_head(sk
);
1164 tcp_snd_test(sk
, skb
, tcp_current_mss(sk
, 1),
1165 (tcp_skb_is_last(sk
, skb
) ?
1166 tp
->nonagle
: TCP_NAGLE_PUSH
)));
1169 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1170 * which is put after SKB on the list. It is very much like
1171 * tcp_fragment() except that it may make several kinds of assumptions
1172 * in order to speed up the splitting operation. In particular, we
1173 * know that all the data is in scatter-gather pages, and that the
1174 * packet has never been sent out before (and thus is not cloned).
1176 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
,
1177 unsigned int mss_now
)
1179 struct sk_buff
*buff
;
1180 int nlen
= skb
->len
- len
;
1183 /* All of a TSO frame must be composed of paged data. */
1184 if (skb
->len
!= skb
->data_len
)
1185 return tcp_fragment(sk
, skb
, len
, mss_now
);
1187 buff
= sk_stream_alloc_skb(sk
, 0, GFP_ATOMIC
);
1188 if (unlikely(buff
== NULL
))
1191 sk
->sk_wmem_queued
+= buff
->truesize
;
1192 sk_mem_charge(sk
, buff
->truesize
);
1193 buff
->truesize
+= nlen
;
1194 skb
->truesize
-= nlen
;
1196 /* Correct the sequence numbers. */
1197 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1198 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1199 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1201 /* PSH and FIN should only be set in the second packet. */
1202 flags
= TCP_SKB_CB(skb
)->flags
;
1203 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
| TCPCB_FLAG_PSH
);
1204 TCP_SKB_CB(buff
)->flags
= flags
;
1206 /* This packet was never sent out yet, so no SACK bits. */
1207 TCP_SKB_CB(buff
)->sacked
= 0;
1209 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_PARTIAL
;
1210 skb_split(skb
, buff
, len
);
1212 /* Fix up tso_factor for both original and new SKB. */
1213 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1214 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1216 /* Link BUFF into the send queue. */
1217 skb_header_release(buff
);
1218 tcp_insert_write_queue_after(skb
, buff
, sk
);
1223 /* Try to defer sending, if possible, in order to minimize the amount
1224 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1226 * This algorithm is from John Heffner.
1228 static int tcp_tso_should_defer(struct sock
*sk
, struct sk_buff
*skb
)
1230 struct tcp_sock
*tp
= tcp_sk(sk
);
1231 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1232 u32 send_win
, cong_win
, limit
, in_flight
;
1234 if (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
)
1237 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
1240 /* Defer for less than two clock ticks. */
1241 if (tp
->tso_deferred
&&
1242 ((jiffies
<< 1) >> 1) - (tp
->tso_deferred
>> 1) > 1)
1245 in_flight
= tcp_packets_in_flight(tp
);
1247 BUG_ON(tcp_skb_pcount(skb
) <= 1 || (tp
->snd_cwnd
<= in_flight
));
1249 send_win
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1251 /* From in_flight test above, we know that cwnd > in_flight. */
1252 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
1254 limit
= min(send_win
, cong_win
);
1256 /* If a full-sized TSO skb can be sent, do it. */
1260 if (sysctl_tcp_tso_win_divisor
) {
1261 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
1263 /* If at least some fraction of a window is available,
1266 chunk
/= sysctl_tcp_tso_win_divisor
;
1270 /* Different approach, try not to defer past a single
1271 * ACK. Receiver should ACK every other full sized
1272 * frame, so if we have space for more than 3 frames
1275 if (limit
> tcp_max_burst(tp
) * tp
->mss_cache
)
1279 /* Ok, it looks like it is advisable to defer. */
1280 tp
->tso_deferred
= 1 | (jiffies
<< 1);
1285 tp
->tso_deferred
= 0;
1289 /* Create a new MTU probe if we are ready.
1290 * Returns 0 if we should wait to probe (no cwnd available),
1291 * 1 if a probe was sent,
1294 static int tcp_mtu_probe(struct sock
*sk
)
1296 struct tcp_sock
*tp
= tcp_sk(sk
);
1297 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1298 struct sk_buff
*skb
, *nskb
, *next
;
1305 /* Not currently probing/verifying,
1307 * have enough cwnd, and
1308 * not SACKing (the variable headers throw things off) */
1309 if (!icsk
->icsk_mtup
.enabled
||
1310 icsk
->icsk_mtup
.probe_size
||
1311 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
||
1312 tp
->snd_cwnd
< 11 ||
1313 tp
->rx_opt
.eff_sacks
)
1316 /* Very simple search strategy: just double the MSS. */
1317 mss_now
= tcp_current_mss(sk
, 0);
1318 probe_size
= 2 * tp
->mss_cache
;
1319 size_needed
= probe_size
+ (tp
->reordering
+ 1) * tp
->mss_cache
;
1320 if (probe_size
> tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_high
)) {
1321 /* TODO: set timer for probe_converge_event */
1325 /* Have enough data in the send queue to probe? */
1326 if (tp
->write_seq
- tp
->snd_nxt
< size_needed
)
1329 if (tp
->snd_wnd
< size_needed
)
1331 if (after(tp
->snd_nxt
+ size_needed
, tcp_wnd_end(tp
)))
1334 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1335 if (tcp_packets_in_flight(tp
) + 2 > tp
->snd_cwnd
) {
1336 if (!tcp_packets_in_flight(tp
))
1342 /* We're allowed to probe. Build it now. */
1343 if ((nskb
= sk_stream_alloc_skb(sk
, probe_size
, GFP_ATOMIC
)) == NULL
)
1345 sk
->sk_wmem_queued
+= nskb
->truesize
;
1346 sk_mem_charge(sk
, nskb
->truesize
);
1348 skb
= tcp_send_head(sk
);
1350 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(skb
)->seq
;
1351 TCP_SKB_CB(nskb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ probe_size
;
1352 TCP_SKB_CB(nskb
)->flags
= TCPCB_FLAG_ACK
;
1353 TCP_SKB_CB(nskb
)->sacked
= 0;
1355 nskb
->ip_summed
= skb
->ip_summed
;
1357 tcp_insert_write_queue_before(nskb
, skb
, sk
);
1360 tcp_for_write_queue_from_safe(skb
, next
, sk
) {
1361 copy
= min_t(int, skb
->len
, probe_size
- len
);
1362 if (nskb
->ip_summed
)
1363 skb_copy_bits(skb
, 0, skb_put(nskb
, copy
), copy
);
1365 nskb
->csum
= skb_copy_and_csum_bits(skb
, 0,
1366 skb_put(nskb
, copy
),
1369 if (skb
->len
<= copy
) {
1370 /* We've eaten all the data from this skb.
1372 TCP_SKB_CB(nskb
)->flags
|= TCP_SKB_CB(skb
)->flags
;
1373 tcp_unlink_write_queue(skb
, sk
);
1374 sk_wmem_free_skb(sk
, skb
);
1376 TCP_SKB_CB(nskb
)->flags
|= TCP_SKB_CB(skb
)->flags
&
1377 ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
1378 if (!skb_shinfo(skb
)->nr_frags
) {
1379 skb_pull(skb
, copy
);
1380 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1381 skb
->csum
= csum_partial(skb
->data
,
1384 __pskb_trim_head(skb
, copy
);
1385 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1387 TCP_SKB_CB(skb
)->seq
+= copy
;
1392 if (len
>= probe_size
)
1395 tcp_init_tso_segs(sk
, nskb
, nskb
->len
);
1397 /* We're ready to send. If this fails, the probe will
1398 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1399 TCP_SKB_CB(nskb
)->when
= tcp_time_stamp
;
1400 if (!tcp_transmit_skb(sk
, nskb
, 1, GFP_ATOMIC
)) {
1401 /* Decrement cwnd here because we are sending
1402 * effectively two packets. */
1404 tcp_event_new_data_sent(sk
, nskb
);
1406 icsk
->icsk_mtup
.probe_size
= tcp_mss_to_mtu(sk
, nskb
->len
);
1407 tp
->mtu_probe
.probe_seq_start
= TCP_SKB_CB(nskb
)->seq
;
1408 tp
->mtu_probe
.probe_seq_end
= TCP_SKB_CB(nskb
)->end_seq
;
1416 /* This routine writes packets to the network. It advances the
1417 * send_head. This happens as incoming acks open up the remote
1420 * Returns 1, if no segments are in flight and we have queued segments, but
1421 * cannot send anything now because of SWS or another problem.
1423 static int tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
)
1425 struct tcp_sock
*tp
= tcp_sk(sk
);
1426 struct sk_buff
*skb
;
1427 unsigned int tso_segs
, sent_pkts
;
1431 /* If we are closed, the bytes will have to remain here.
1432 * In time closedown will finish, we empty the write queue and all
1435 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
1440 /* Do MTU probing. */
1441 if ((result
= tcp_mtu_probe(sk
)) == 0) {
1443 } else if (result
> 0) {
1447 while ((skb
= tcp_send_head(sk
))) {
1450 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1453 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1457 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
)))
1460 if (tso_segs
== 1) {
1461 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
1462 (tcp_skb_is_last(sk
, skb
) ?
1463 nonagle
: TCP_NAGLE_PUSH
))))
1466 if (tcp_tso_should_defer(sk
, skb
))
1472 limit
= tcp_mss_split_point(sk
, skb
, mss_now
,
1475 if (skb
->len
> limit
&&
1476 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
)))
1479 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1481 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
)))
1484 /* Advance the send_head. This one is sent out.
1485 * This call will increment packets_out.
1487 tcp_event_new_data_sent(sk
, skb
);
1489 tcp_minshall_update(tp
, mss_now
, skb
);
1493 if (likely(sent_pkts
)) {
1494 tcp_cwnd_validate(sk
);
1497 return !tp
->packets_out
&& tcp_send_head(sk
);
1500 /* Push out any pending frames which were held back due to
1501 * TCP_CORK or attempt at coalescing tiny packets.
1502 * The socket must be locked by the caller.
1504 void __tcp_push_pending_frames(struct sock
*sk
, unsigned int cur_mss
,
1507 struct sk_buff
*skb
= tcp_send_head(sk
);
1510 if (tcp_write_xmit(sk
, cur_mss
, nonagle
))
1511 tcp_check_probe_timer(sk
);
1515 /* Send _single_ skb sitting at the send head. This function requires
1516 * true push pending frames to setup probe timer etc.
1518 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
1520 struct sk_buff
*skb
= tcp_send_head(sk
);
1521 unsigned int tso_segs
, cwnd_quota
;
1523 BUG_ON(!skb
|| skb
->len
< mss_now
);
1525 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1526 cwnd_quota
= tcp_snd_test(sk
, skb
, mss_now
, TCP_NAGLE_PUSH
);
1528 if (likely(cwnd_quota
)) {
1535 limit
= tcp_mss_split_point(sk
, skb
, mss_now
,
1538 if (skb
->len
> limit
&&
1539 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
)))
1542 /* Send it out now. */
1543 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1545 if (likely(!tcp_transmit_skb(sk
, skb
, 1, sk
->sk_allocation
))) {
1546 tcp_event_new_data_sent(sk
, skb
);
1547 tcp_cwnd_validate(sk
);
1553 /* This function returns the amount that we can raise the
1554 * usable window based on the following constraints
1556 * 1. The window can never be shrunk once it is offered (RFC 793)
1557 * 2. We limit memory per socket
1560 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1561 * RECV.NEXT + RCV.WIN fixed until:
1562 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1564 * i.e. don't raise the right edge of the window until you can raise
1565 * it at least MSS bytes.
1567 * Unfortunately, the recommended algorithm breaks header prediction,
1568 * since header prediction assumes th->window stays fixed.
1570 * Strictly speaking, keeping th->window fixed violates the receiver
1571 * side SWS prevention criteria. The problem is that under this rule
1572 * a stream of single byte packets will cause the right side of the
1573 * window to always advance by a single byte.
1575 * Of course, if the sender implements sender side SWS prevention
1576 * then this will not be a problem.
1578 * BSD seems to make the following compromise:
1580 * If the free space is less than the 1/4 of the maximum
1581 * space available and the free space is less than 1/2 mss,
1582 * then set the window to 0.
1583 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1584 * Otherwise, just prevent the window from shrinking
1585 * and from being larger than the largest representable value.
1587 * This prevents incremental opening of the window in the regime
1588 * where TCP is limited by the speed of the reader side taking
1589 * data out of the TCP receive queue. It does nothing about
1590 * those cases where the window is constrained on the sender side
1591 * because the pipeline is full.
1593 * BSD also seems to "accidentally" limit itself to windows that are a
1594 * multiple of MSS, at least until the free space gets quite small.
1595 * This would appear to be a side effect of the mbuf implementation.
1596 * Combining these two algorithms results in the observed behavior
1597 * of having a fixed window size at almost all times.
1599 * Below we obtain similar behavior by forcing the offered window to
1600 * a multiple of the mss when it is feasible to do so.
1602 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1603 * Regular options like TIMESTAMP are taken into account.
1605 u32
__tcp_select_window(struct sock
*sk
)
1607 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1608 struct tcp_sock
*tp
= tcp_sk(sk
);
1609 /* MSS for the peer's data. Previous versions used mss_clamp
1610 * here. I don't know if the value based on our guesses
1611 * of peer's MSS is better for the performance. It's more correct
1612 * but may be worse for the performance because of rcv_mss
1613 * fluctuations. --SAW 1998/11/1
1615 int mss
= icsk
->icsk_ack
.rcv_mss
;
1616 int free_space
= tcp_space(sk
);
1617 int full_space
= min_t(int, tp
->window_clamp
, tcp_full_space(sk
));
1620 if (mss
> full_space
)
1623 if (free_space
< (full_space
>> 1)) {
1624 icsk
->icsk_ack
.quick
= 0;
1626 if (tcp_memory_pressure
)
1627 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
,
1630 if (free_space
< mss
)
1634 if (free_space
> tp
->rcv_ssthresh
)
1635 free_space
= tp
->rcv_ssthresh
;
1637 /* Don't do rounding if we are using window scaling, since the
1638 * scaled window will not line up with the MSS boundary anyway.
1640 window
= tp
->rcv_wnd
;
1641 if (tp
->rx_opt
.rcv_wscale
) {
1642 window
= free_space
;
1644 /* Advertise enough space so that it won't get scaled away.
1645 * Import case: prevent zero window announcement if
1646 * 1<<rcv_wscale > mss.
1648 if (((window
>> tp
->rx_opt
.rcv_wscale
) << tp
->rx_opt
.rcv_wscale
) != window
)
1649 window
= (((window
>> tp
->rx_opt
.rcv_wscale
) + 1)
1650 << tp
->rx_opt
.rcv_wscale
);
1652 /* Get the largest window that is a nice multiple of mss.
1653 * Window clamp already applied above.
1654 * If our current window offering is within 1 mss of the
1655 * free space we just keep it. This prevents the divide
1656 * and multiply from happening most of the time.
1657 * We also don't do any window rounding when the free space
1660 if (window
<= free_space
- mss
|| window
> free_space
)
1661 window
= (free_space
/ mss
) * mss
;
1662 else if (mss
== full_space
&&
1663 free_space
> window
+ (full_space
>> 1))
1664 window
= free_space
;
1670 /* Attempt to collapse two adjacent SKB's during retransmission. */
1671 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*skb
,
1674 struct tcp_sock
*tp
= tcp_sk(sk
);
1675 struct sk_buff
*next_skb
= tcp_write_queue_next(sk
, skb
);
1676 int skb_size
, next_skb_size
;
1679 /* The first test we must make is that neither of these two
1680 * SKB's are still referenced by someone else.
1682 if (skb_cloned(skb
) || skb_cloned(next_skb
))
1685 skb_size
= skb
->len
;
1686 next_skb_size
= next_skb
->len
;
1687 flags
= TCP_SKB_CB(skb
)->flags
;
1689 /* Also punt if next skb has been SACK'd. */
1690 if (TCP_SKB_CB(next_skb
)->sacked
& TCPCB_SACKED_ACKED
)
1693 /* Next skb is out of window. */
1694 if (after(TCP_SKB_CB(next_skb
)->end_seq
, tcp_wnd_end(tp
)))
1697 /* Punt if not enough space exists in the first SKB for
1698 * the data in the second, or the total combined payload
1699 * would exceed the MSS.
1701 if ((next_skb_size
> skb_tailroom(skb
)) ||
1702 ((skb_size
+ next_skb_size
) > mss_now
))
1705 BUG_ON(tcp_skb_pcount(skb
) != 1 || tcp_skb_pcount(next_skb
) != 1);
1707 tcp_highest_sack_combine(sk
, next_skb
, skb
);
1709 /* Ok. We will be able to collapse the packet. */
1710 tcp_unlink_write_queue(next_skb
, sk
);
1712 skb_copy_from_linear_data(next_skb
, skb_put(skb
, next_skb_size
),
1715 if (next_skb
->ip_summed
== CHECKSUM_PARTIAL
)
1716 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1718 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1719 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
1721 /* Update sequence range on original skb. */
1722 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
1724 /* Merge over control information. */
1725 flags
|= TCP_SKB_CB(next_skb
)->flags
; /* This moves PSH/FIN etc. over */
1726 TCP_SKB_CB(skb
)->flags
= flags
;
1728 /* All done, get rid of second SKB and account for it so
1729 * packet counting does not break.
1731 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
& TCPCB_EVER_RETRANS
;
1732 if (TCP_SKB_CB(next_skb
)->sacked
& TCPCB_SACKED_RETRANS
)
1733 tp
->retrans_out
-= tcp_skb_pcount(next_skb
);
1734 if (TCP_SKB_CB(next_skb
)->sacked
& TCPCB_LOST
)
1735 tp
->lost_out
-= tcp_skb_pcount(next_skb
);
1736 /* Reno case is special. Sigh... */
1737 if (tcp_is_reno(tp
) && tp
->sacked_out
)
1738 tcp_dec_pcount_approx(&tp
->sacked_out
, next_skb
);
1740 tcp_adjust_fackets_out(sk
, next_skb
, tcp_skb_pcount(next_skb
));
1741 tp
->packets_out
-= tcp_skb_pcount(next_skb
);
1743 /* changed transmit queue under us so clear hints */
1744 tcp_clear_retrans_hints_partial(tp
);
1746 sk_wmem_free_skb(sk
, next_skb
);
1749 /* Do a simple retransmit without using the backoff mechanisms in
1750 * tcp_timer. This is used for path mtu discovery.
1751 * The socket is already locked here.
1753 void tcp_simple_retransmit(struct sock
*sk
)
1755 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1756 struct tcp_sock
*tp
= tcp_sk(sk
);
1757 struct sk_buff
*skb
;
1758 unsigned int mss
= tcp_current_mss(sk
, 0);
1761 tcp_for_write_queue(skb
, sk
) {
1762 if (skb
== tcp_send_head(sk
))
1764 if (skb
->len
> mss
&&
1765 !(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
1766 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
1767 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1768 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1770 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
)) {
1771 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1772 tp
->lost_out
+= tcp_skb_pcount(skb
);
1778 tcp_clear_all_retrans_hints(tp
);
1783 tcp_verify_left_out(tp
);
1785 /* Don't muck with the congestion window here.
1786 * Reason is that we do not increase amount of _data_
1787 * in network, but units changed and effective
1788 * cwnd/ssthresh really reduced now.
1790 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
1791 tp
->high_seq
= tp
->snd_nxt
;
1792 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
1793 tp
->prior_ssthresh
= 0;
1794 tp
->undo_marker
= 0;
1795 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1797 tcp_xmit_retransmit_queue(sk
);
1800 /* This retransmits one SKB. Policy decisions and retransmit queue
1801 * state updates are done by the caller. Returns non-zero if an
1802 * error occurred which prevented the send.
1804 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
1806 struct tcp_sock
*tp
= tcp_sk(sk
);
1807 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1808 unsigned int cur_mss
= tcp_current_mss(sk
, 0);
1811 /* Inconslusive MTU probe */
1812 if (icsk
->icsk_mtup
.probe_size
) {
1813 icsk
->icsk_mtup
.probe_size
= 0;
1816 /* Do not sent more than we queued. 1/4 is reserved for possible
1817 * copying overhead: fragmentation, tunneling, mangling etc.
1819 if (atomic_read(&sk
->sk_wmem_alloc
) >
1820 min(sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2), sk
->sk_sndbuf
))
1823 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
1824 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1826 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
1830 /* If receiver has shrunk his window, and skb is out of
1831 * new window, do not retransmit it. The exception is the
1832 * case, when window is shrunk to zero. In this case
1833 * our retransmit serves as a zero window probe.
1835 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
))
1836 && TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
1839 if (skb
->len
> cur_mss
) {
1840 if (tcp_fragment(sk
, skb
, cur_mss
, cur_mss
))
1841 return -ENOMEM
; /* We'll try again later. */
1844 /* Collapse two adjacent packets if worthwhile and we can. */
1845 if (!(TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_SYN
) &&
1846 (skb
->len
< (cur_mss
>> 1)) &&
1847 (tcp_write_queue_next(sk
, skb
) != tcp_send_head(sk
)) &&
1848 (!tcp_skb_is_last(sk
, skb
)) &&
1849 (skb_shinfo(skb
)->nr_frags
== 0 &&
1850 skb_shinfo(tcp_write_queue_next(sk
, skb
))->nr_frags
== 0) &&
1851 (tcp_skb_pcount(skb
) == 1 &&
1852 tcp_skb_pcount(tcp_write_queue_next(sk
, skb
)) == 1) &&
1853 (sysctl_tcp_retrans_collapse
!= 0))
1854 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
1856 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
1857 return -EHOSTUNREACH
; /* Routing failure or similar. */
1859 /* Some Solaris stacks overoptimize and ignore the FIN on a
1860 * retransmit when old data is attached. So strip it off
1861 * since it is cheap to do so and saves bytes on the network.
1864 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
) &&
1865 tp
->snd_una
== (TCP_SKB_CB(skb
)->end_seq
- 1)) {
1866 if (!pskb_trim(skb
, 0)) {
1867 TCP_SKB_CB(skb
)->seq
= TCP_SKB_CB(skb
)->end_seq
- 1;
1868 skb_shinfo(skb
)->gso_segs
= 1;
1869 skb_shinfo(skb
)->gso_size
= 0;
1870 skb_shinfo(skb
)->gso_type
= 0;
1871 skb
->ip_summed
= CHECKSUM_NONE
;
1876 /* Make a copy, if the first transmission SKB clone we made
1877 * is still in somebody's hands, else make a clone.
1879 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1881 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
1884 /* Update global TCP statistics. */
1885 TCP_INC_STATS(TCP_MIB_RETRANSSEGS
);
1887 tp
->total_retrans
++;
1889 #if FASTRETRANS_DEBUG > 0
1890 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
1891 if (net_ratelimit())
1892 printk(KERN_DEBUG
"retrans_out leaked.\n");
1895 if (!tp
->retrans_out
)
1896 tp
->lost_retrans_low
= tp
->snd_nxt
;
1897 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
1898 tp
->retrans_out
+= tcp_skb_pcount(skb
);
1900 /* Save stamp of the first retransmit. */
1901 if (!tp
->retrans_stamp
)
1902 tp
->retrans_stamp
= TCP_SKB_CB(skb
)->when
;
1906 /* snd_nxt is stored to detect loss of retransmitted segment,
1907 * see tcp_input.c tcp_sacktag_write_queue().
1909 TCP_SKB_CB(skb
)->ack_seq
= tp
->snd_nxt
;
1914 /* This gets called after a retransmit timeout, and the initially
1915 * retransmitted data is acknowledged. It tries to continue
1916 * resending the rest of the retransmit queue, until either
1917 * we've sent it all or the congestion window limit is reached.
1918 * If doing SACK, the first ACK which comes back for a timeout
1919 * based retransmit packet might feed us FACK information again.
1920 * If so, we use it to avoid unnecessarily retransmissions.
1922 void tcp_xmit_retransmit_queue(struct sock
*sk
)
1924 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1925 struct tcp_sock
*tp
= tcp_sk(sk
);
1926 struct sk_buff
*skb
;
1929 if (tp
->retransmit_skb_hint
) {
1930 skb
= tp
->retransmit_skb_hint
;
1931 packet_cnt
= tp
->retransmit_cnt_hint
;
1933 skb
= tcp_write_queue_head(sk
);
1937 /* First pass: retransmit lost packets. */
1939 tcp_for_write_queue_from(skb
, sk
) {
1940 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
1942 if (skb
== tcp_send_head(sk
))
1944 /* we could do better than to assign each time */
1945 tp
->retransmit_skb_hint
= skb
;
1946 tp
->retransmit_cnt_hint
= packet_cnt
;
1948 /* Assume this retransmit will generate
1949 * only one packet for congestion window
1950 * calculation purposes. This works because
1951 * tcp_retransmit_skb() will chop up the
1952 * packet to be MSS sized and all the
1953 * packet counting works out.
1955 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
1958 if (sacked
& TCPCB_LOST
) {
1959 if (!(sacked
& (TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))) {
1960 if (tcp_retransmit_skb(sk
, skb
)) {
1961 tp
->retransmit_skb_hint
= NULL
;
1964 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
1965 NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS
);
1967 NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS
);
1969 if (skb
== tcp_write_queue_head(sk
))
1970 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1971 inet_csk(sk
)->icsk_rto
,
1975 packet_cnt
+= tcp_skb_pcount(skb
);
1976 if (packet_cnt
>= tp
->lost_out
)
1982 /* OK, demanded retransmission is finished. */
1984 /* Forward retransmissions are possible only during Recovery. */
1985 if (icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
1988 /* No forward retransmissions in Reno are possible. */
1989 if (tcp_is_reno(tp
))
1992 /* Yeah, we have to make difficult choice between forward transmission
1993 * and retransmission... Both ways have their merits...
1995 * For now we do not retransmit anything, while we have some new
1996 * segments to send. In the other cases, follow rule 3 for
1997 * NextSeg() specified in RFC3517.
2000 if (tcp_may_send_now(sk
))
2003 /* If nothing is SACKed, highest_sack in the loop won't be valid */
2004 if (!tp
->sacked_out
)
2007 if (tp
->forward_skb_hint
)
2008 skb
= tp
->forward_skb_hint
;
2010 skb
= tcp_write_queue_head(sk
);
2012 tcp_for_write_queue_from(skb
, sk
) {
2013 if (skb
== tcp_send_head(sk
))
2015 tp
->forward_skb_hint
= skb
;
2017 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_highest_sack_seq(tp
)))
2020 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
2023 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
)
2026 /* Ok, retransmit it. */
2027 if (tcp_retransmit_skb(sk
, skb
)) {
2028 tp
->forward_skb_hint
= NULL
;
2032 if (skb
== tcp_write_queue_head(sk
))
2033 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2034 inet_csk(sk
)->icsk_rto
,
2037 NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS
);
2041 /* Send a fin. The caller locks the socket for us. This cannot be
2042 * allowed to fail queueing a FIN frame under any circumstances.
2044 void tcp_send_fin(struct sock
*sk
)
2046 struct tcp_sock
*tp
= tcp_sk(sk
);
2047 struct sk_buff
*skb
= tcp_write_queue_tail(sk
);
2050 /* Optimization, tack on the FIN if we have a queue of
2051 * unsent frames. But be careful about outgoing SACKS
2054 mss_now
= tcp_current_mss(sk
, 1);
2056 if (tcp_send_head(sk
) != NULL
) {
2057 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_FIN
;
2058 TCP_SKB_CB(skb
)->end_seq
++;
2061 /* Socket is locked, keep trying until memory is available. */
2063 skb
= alloc_skb_fclone(MAX_TCP_HEADER
, GFP_KERNEL
);
2069 /* Reserve space for headers and prepare control bits. */
2070 skb_reserve(skb
, MAX_TCP_HEADER
);
2072 TCP_SKB_CB(skb
)->flags
= (TCPCB_FLAG_ACK
| TCPCB_FLAG_FIN
);
2073 TCP_SKB_CB(skb
)->sacked
= 0;
2074 skb_shinfo(skb
)->gso_segs
= 1;
2075 skb_shinfo(skb
)->gso_size
= 0;
2076 skb_shinfo(skb
)->gso_type
= 0;
2078 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2079 TCP_SKB_CB(skb
)->seq
= tp
->write_seq
;
2080 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ 1;
2081 tcp_queue_skb(sk
, skb
);
2083 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_OFF
);
2086 /* We get here when a process closes a file descriptor (either due to
2087 * an explicit close() or as a byproduct of exit()'ing) and there
2088 * was unread data in the receive queue. This behavior is recommended
2089 * by RFC 2525, section 2.17. -DaveM
2091 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
2093 struct sk_buff
*skb
;
2095 /* NOTE: No TCP options attached and we never retransmit this. */
2096 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
2098 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED
);
2102 /* Reserve space for headers and prepare control bits. */
2103 skb_reserve(skb
, MAX_TCP_HEADER
);
2105 TCP_SKB_CB(skb
)->flags
= (TCPCB_FLAG_ACK
| TCPCB_FLAG_RST
);
2106 TCP_SKB_CB(skb
)->sacked
= 0;
2107 skb_shinfo(skb
)->gso_segs
= 1;
2108 skb_shinfo(skb
)->gso_size
= 0;
2109 skb_shinfo(skb
)->gso_type
= 0;
2112 TCP_SKB_CB(skb
)->seq
= tcp_acceptable_seq(sk
);
2113 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
;
2114 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2115 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
2116 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED
);
2119 /* WARNING: This routine must only be called when we have already sent
2120 * a SYN packet that crossed the incoming SYN that caused this routine
2121 * to get called. If this assumption fails then the initial rcv_wnd
2122 * and rcv_wscale values will not be correct.
2124 int tcp_send_synack(struct sock
*sk
)
2126 struct sk_buff
*skb
;
2128 skb
= tcp_write_queue_head(sk
);
2129 if (skb
== NULL
|| !(TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_SYN
)) {
2130 printk(KERN_DEBUG
"tcp_send_synack: wrong queue state\n");
2133 if (!(TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_ACK
)) {
2134 if (skb_cloned(skb
)) {
2135 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
2138 tcp_unlink_write_queue(skb
, sk
);
2139 skb_header_release(nskb
);
2140 __tcp_add_write_queue_head(sk
, nskb
);
2141 sk_wmem_free_skb(sk
, skb
);
2142 sk
->sk_wmem_queued
+= nskb
->truesize
;
2143 sk_mem_charge(sk
, nskb
->truesize
);
2147 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_ACK
;
2148 TCP_ECN_send_synack(tcp_sk(sk
), skb
);
2150 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2151 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2155 * Prepare a SYN-ACK.
2157 struct sk_buff
*tcp_make_synack(struct sock
*sk
, struct dst_entry
*dst
,
2158 struct request_sock
*req
)
2160 struct inet_request_sock
*ireq
= inet_rsk(req
);
2161 struct tcp_sock
*tp
= tcp_sk(sk
);
2163 int tcp_header_size
;
2164 struct sk_buff
*skb
;
2165 #ifdef CONFIG_TCP_MD5SIG
2166 struct tcp_md5sig_key
*md5
;
2167 __u8
*md5_hash_location
;
2170 skb
= sock_wmalloc(sk
, MAX_TCP_HEADER
+ 15, 1, GFP_ATOMIC
);
2174 /* Reserve space for headers. */
2175 skb_reserve(skb
, MAX_TCP_HEADER
);
2177 skb
->dst
= dst_clone(dst
);
2179 tcp_header_size
= (sizeof(struct tcphdr
) + TCPOLEN_MSS
+
2180 (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0) +
2181 (ireq
->wscale_ok
? TCPOLEN_WSCALE_ALIGNED
: 0) +
2182 /* SACK_PERM is in the place of NOP NOP of TS */
2183 ((ireq
->sack_ok
&& !ireq
->tstamp_ok
) ? TCPOLEN_SACKPERM_ALIGNED
: 0));
2185 #ifdef CONFIG_TCP_MD5SIG
2186 /* Are we doing MD5 on this segment? If so - make room for it */
2187 md5
= tcp_rsk(req
)->af_specific
->md5_lookup(sk
, req
);
2189 tcp_header_size
+= TCPOLEN_MD5SIG_ALIGNED
;
2191 skb_push(skb
, tcp_header_size
);
2192 skb_reset_transport_header(skb
);
2195 memset(th
, 0, sizeof(struct tcphdr
));
2198 TCP_ECN_make_synack(req
, th
);
2199 th
->source
= inet_sk(sk
)->sport
;
2200 th
->dest
= ireq
->rmt_port
;
2201 TCP_SKB_CB(skb
)->seq
= tcp_rsk(req
)->snt_isn
;
2202 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ 1;
2203 TCP_SKB_CB(skb
)->sacked
= 0;
2204 skb_shinfo(skb
)->gso_segs
= 1;
2205 skb_shinfo(skb
)->gso_size
= 0;
2206 skb_shinfo(skb
)->gso_type
= 0;
2207 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
2208 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_isn
+ 1);
2209 if (req
->rcv_wnd
== 0) { /* ignored for retransmitted syns */
2211 /* Set this up on the first call only */
2212 req
->window_clamp
= tp
->window_clamp
? : dst_metric(dst
, RTAX_WINDOW
);
2213 /* tcp_full_space because it is guaranteed to be the first packet */
2214 tcp_select_initial_window(tcp_full_space(sk
),
2215 dst_metric(dst
, RTAX_ADVMSS
) - (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0),
2220 ireq
->rcv_wscale
= rcv_wscale
;
2223 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2224 th
->window
= htons(min(req
->rcv_wnd
, 65535U));
2226 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2227 tcp_syn_build_options((__be32
*)(th
+ 1), dst_metric(dst
, RTAX_ADVMSS
), ireq
->tstamp_ok
,
2228 ireq
->sack_ok
, ireq
->wscale_ok
, ireq
->rcv_wscale
,
2229 TCP_SKB_CB(skb
)->when
,
2232 #ifdef CONFIG_TCP_MD5SIG
2233 md5
? &md5_hash_location
:
2239 th
->doff
= (tcp_header_size
>> 2);
2240 TCP_INC_STATS(TCP_MIB_OUTSEGS
);
2242 #ifdef CONFIG_TCP_MD5SIG
2243 /* Okay, we have all we need - do the md5 hash if needed */
2245 tp
->af_specific
->calc_md5_hash(md5_hash_location
,
2248 tcp_hdr(skb
), sk
->sk_protocol
,
2257 * Do all connect socket setups that can be done AF independent.
2259 static void tcp_connect_init(struct sock
*sk
)
2261 struct dst_entry
*dst
= __sk_dst_get(sk
);
2262 struct tcp_sock
*tp
= tcp_sk(sk
);
2265 /* We'll fix this up when we get a response from the other end.
2266 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2268 tp
->tcp_header_len
= sizeof(struct tcphdr
) +
2269 (sysctl_tcp_timestamps
? TCPOLEN_TSTAMP_ALIGNED
: 0);
2271 #ifdef CONFIG_TCP_MD5SIG
2272 if (tp
->af_specific
->md5_lookup(sk
, sk
) != NULL
)
2273 tp
->tcp_header_len
+= TCPOLEN_MD5SIG_ALIGNED
;
2276 /* If user gave his TCP_MAXSEG, record it to clamp */
2277 if (tp
->rx_opt
.user_mss
)
2278 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
2281 tcp_sync_mss(sk
, dst_mtu(dst
));
2283 if (!tp
->window_clamp
)
2284 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
2285 tp
->advmss
= dst_metric(dst
, RTAX_ADVMSS
);
2286 tcp_initialize_rcv_mss(sk
);
2288 tcp_select_initial_window(tcp_full_space(sk
),
2289 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
2292 sysctl_tcp_window_scaling
,
2295 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
2296 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
2299 sock_reset_flag(sk
, SOCK_DONE
);
2301 tcp_init_wl(tp
, tp
->write_seq
, 0);
2302 tp
->snd_una
= tp
->write_seq
;
2303 tp
->snd_sml
= tp
->write_seq
;
2308 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
2309 inet_csk(sk
)->icsk_retransmits
= 0;
2310 tcp_clear_retrans(tp
);
2314 * Build a SYN and send it off.
2316 int tcp_connect(struct sock
*sk
)
2318 struct tcp_sock
*tp
= tcp_sk(sk
);
2319 struct sk_buff
*buff
;
2321 tcp_connect_init(sk
);
2323 buff
= alloc_skb_fclone(MAX_TCP_HEADER
+ 15, sk
->sk_allocation
);
2324 if (unlikely(buff
== NULL
))
2327 /* Reserve space for headers. */
2328 skb_reserve(buff
, MAX_TCP_HEADER
);
2330 TCP_SKB_CB(buff
)->flags
= TCPCB_FLAG_SYN
;
2331 TCP_ECN_send_syn(sk
, buff
);
2332 TCP_SKB_CB(buff
)->sacked
= 0;
2333 skb_shinfo(buff
)->gso_segs
= 1;
2334 skb_shinfo(buff
)->gso_size
= 0;
2335 skb_shinfo(buff
)->gso_type
= 0;
2337 tp
->snd_nxt
= tp
->write_seq
;
2338 TCP_SKB_CB(buff
)->seq
= tp
->write_seq
++;
2339 TCP_SKB_CB(buff
)->end_seq
= tp
->write_seq
;
2342 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2343 tp
->retrans_stamp
= TCP_SKB_CB(buff
)->when
;
2344 skb_header_release(buff
);
2345 __tcp_add_write_queue_tail(sk
, buff
);
2346 sk
->sk_wmem_queued
+= buff
->truesize
;
2347 sk_mem_charge(sk
, buff
->truesize
);
2348 tp
->packets_out
+= tcp_skb_pcount(buff
);
2349 tcp_transmit_skb(sk
, buff
, 1, GFP_KERNEL
);
2351 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2352 * in order to make this packet get counted in tcpOutSegs.
2354 tp
->snd_nxt
= tp
->write_seq
;
2355 tp
->pushed_seq
= tp
->write_seq
;
2356 TCP_INC_STATS(TCP_MIB_ACTIVEOPENS
);
2358 /* Timer for repeating the SYN until an answer. */
2359 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2360 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2364 /* Send out a delayed ack, the caller does the policy checking
2365 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2368 void tcp_send_delayed_ack(struct sock
*sk
)
2370 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2371 int ato
= icsk
->icsk_ack
.ato
;
2372 unsigned long timeout
;
2374 if (ato
> TCP_DELACK_MIN
) {
2375 const struct tcp_sock
*tp
= tcp_sk(sk
);
2376 int max_ato
= HZ
/ 2;
2378 if (icsk
->icsk_ack
.pingpong
||
2379 (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
2380 max_ato
= TCP_DELACK_MAX
;
2382 /* Slow path, intersegment interval is "high". */
2384 /* If some rtt estimate is known, use it to bound delayed ack.
2385 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2389 int rtt
= max(tp
->srtt
>> 3, TCP_DELACK_MIN
);
2395 ato
= min(ato
, max_ato
);
2398 /* Stay within the limit we were given */
2399 timeout
= jiffies
+ ato
;
2401 /* Use new timeout only if there wasn't a older one earlier. */
2402 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
2403 /* If delack timer was blocked or is about to expire,
2406 if (icsk
->icsk_ack
.blocked
||
2407 time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
2412 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
2413 timeout
= icsk
->icsk_ack
.timeout
;
2415 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
2416 icsk
->icsk_ack
.timeout
= timeout
;
2417 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
2420 /* This routine sends an ack and also updates the window. */
2421 void tcp_send_ack(struct sock
*sk
)
2423 struct sk_buff
*buff
;
2425 /* If we have been reset, we may not send again. */
2426 if (sk
->sk_state
== TCP_CLOSE
)
2429 /* We are not putting this on the write queue, so
2430 * tcp_transmit_skb() will set the ownership to this
2433 buff
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2435 inet_csk_schedule_ack(sk
);
2436 inet_csk(sk
)->icsk_ack
.ato
= TCP_ATO_MIN
;
2437 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
2438 TCP_DELACK_MAX
, TCP_RTO_MAX
);
2442 /* Reserve space for headers and prepare control bits. */
2443 skb_reserve(buff
, MAX_TCP_HEADER
);
2445 TCP_SKB_CB(buff
)->flags
= TCPCB_FLAG_ACK
;
2446 TCP_SKB_CB(buff
)->sacked
= 0;
2447 skb_shinfo(buff
)->gso_segs
= 1;
2448 skb_shinfo(buff
)->gso_size
= 0;
2449 skb_shinfo(buff
)->gso_type
= 0;
2451 /* Send it off, this clears delayed acks for us. */
2452 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(buff
)->end_seq
= tcp_acceptable_seq(sk
);
2453 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2454 tcp_transmit_skb(sk
, buff
, 0, GFP_ATOMIC
);
2457 /* This routine sends a packet with an out of date sequence
2458 * number. It assumes the other end will try to ack it.
2460 * Question: what should we make while urgent mode?
2461 * 4.4BSD forces sending single byte of data. We cannot send
2462 * out of window data, because we have SND.NXT==SND.MAX...
2464 * Current solution: to send TWO zero-length segments in urgent mode:
2465 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2466 * out-of-date with SND.UNA-1 to probe window.
2468 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
)
2470 struct tcp_sock
*tp
= tcp_sk(sk
);
2471 struct sk_buff
*skb
;
2473 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2474 skb
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2478 /* Reserve space for headers and set control bits. */
2479 skb_reserve(skb
, MAX_TCP_HEADER
);
2481 TCP_SKB_CB(skb
)->flags
= TCPCB_FLAG_ACK
;
2482 TCP_SKB_CB(skb
)->sacked
= 0;
2483 skb_shinfo(skb
)->gso_segs
= 1;
2484 skb_shinfo(skb
)->gso_size
= 0;
2485 skb_shinfo(skb
)->gso_type
= 0;
2487 /* Use a previous sequence. This should cause the other
2488 * end to send an ack. Don't queue or clone SKB, just
2491 TCP_SKB_CB(skb
)->seq
= urgent
? tp
->snd_una
: tp
->snd_una
- 1;
2492 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
;
2493 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2494 return tcp_transmit_skb(sk
, skb
, 0, GFP_ATOMIC
);
2497 int tcp_write_wakeup(struct sock
*sk
)
2499 struct tcp_sock
*tp
= tcp_sk(sk
);
2500 struct sk_buff
*skb
;
2502 if (sk
->sk_state
== TCP_CLOSE
)
2505 if ((skb
= tcp_send_head(sk
)) != NULL
&&
2506 before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
))) {
2508 unsigned int mss
= tcp_current_mss(sk
, 0);
2509 unsigned int seg_size
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
2511 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
2512 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
2514 /* We are probing the opening of a window
2515 * but the window size is != 0
2516 * must have been a result SWS avoidance ( sender )
2518 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
2520 seg_size
= min(seg_size
, mss
);
2521 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2522 if (tcp_fragment(sk
, skb
, seg_size
, mss
))
2524 } else if (!tcp_skb_pcount(skb
))
2525 tcp_set_skb_tso_segs(sk
, skb
, mss
);
2527 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2528 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2529 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2531 tcp_event_new_data_sent(sk
, skb
);
2535 between(tp
->snd_up
, tp
->snd_una
+ 1, tp
->snd_una
+ 0xFFFF))
2536 tcp_xmit_probe_skb(sk
, 1);
2537 return tcp_xmit_probe_skb(sk
, 0);
2541 /* A window probe timeout has occurred. If window is not closed send
2542 * a partial packet else a zero probe.
2544 void tcp_send_probe0(struct sock
*sk
)
2546 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2547 struct tcp_sock
*tp
= tcp_sk(sk
);
2550 err
= tcp_write_wakeup(sk
);
2552 if (tp
->packets_out
|| !tcp_send_head(sk
)) {
2553 /* Cancel probe timer, if it is not required. */
2554 icsk
->icsk_probes_out
= 0;
2555 icsk
->icsk_backoff
= 0;
2560 if (icsk
->icsk_backoff
< sysctl_tcp_retries2
)
2561 icsk
->icsk_backoff
++;
2562 icsk
->icsk_probes_out
++;
2563 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2564 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
2567 /* If packet was not sent due to local congestion,
2568 * do not backoff and do not remember icsk_probes_out.
2569 * Let local senders to fight for local resources.
2571 * Use accumulated backoff yet.
2573 if (!icsk
->icsk_probes_out
)
2574 icsk
->icsk_probes_out
= 1;
2575 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2576 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
,
2577 TCP_RESOURCE_PROBE_INTERVAL
),
2582 EXPORT_SYMBOL(tcp_connect
);
2583 EXPORT_SYMBOL(tcp_make_synack
);
2584 EXPORT_SYMBOL(tcp_simple_retransmit
);
2585 EXPORT_SYMBOL(tcp_sync_mss
);
2586 EXPORT_SYMBOL(sysctl_tcp_tso_win_divisor
);
2587 EXPORT_SYMBOL(tcp_mtup_init
);