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>
43 #include <linux/smp_lock.h>
45 /* People can turn this off for buggy TCP's found in printers etc. */
46 int sysctl_tcp_retrans_collapse
= 1;
48 /* This limits the percentage of the congestion window which we
49 * will allow a single TSO frame to consume. Building TSO frames
50 * which are too large can cause TCP streams to be bursty.
52 int sysctl_tcp_tso_win_divisor
= 3;
54 static inline void update_send_head(struct sock
*sk
, struct tcp_sock
*tp
,
57 sk
->sk_send_head
= skb
->next
;
58 if (sk
->sk_send_head
== (struct sk_buff
*)&sk
->sk_write_queue
)
59 sk
->sk_send_head
= NULL
;
60 tp
->snd_nxt
= TCP_SKB_CB(skb
)->end_seq
;
61 tcp_packets_out_inc(sk
, tp
, skb
);
64 /* SND.NXT, if window was not shrunk.
65 * If window has been shrunk, what should we make? It is not clear at all.
66 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
67 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
68 * invalid. OK, let's make this for now:
70 static inline __u32
tcp_acceptable_seq(struct sock
*sk
, struct tcp_sock
*tp
)
72 if (!before(tp
->snd_una
+tp
->snd_wnd
, tp
->snd_nxt
))
75 return tp
->snd_una
+tp
->snd_wnd
;
78 /* Calculate mss to advertise in SYN segment.
79 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
81 * 1. It is independent of path mtu.
82 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
83 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
84 * attached devices, because some buggy hosts are confused by
86 * 4. We do not make 3, we advertise MSS, calculated from first
87 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
88 * This may be overridden via information stored in routing table.
89 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
90 * probably even Jumbo".
92 static __u16
tcp_advertise_mss(struct sock
*sk
)
94 struct tcp_sock
*tp
= tcp_sk(sk
);
95 struct dst_entry
*dst
= __sk_dst_get(sk
);
98 if (dst
&& dst_metric(dst
, RTAX_ADVMSS
) < mss
) {
99 mss
= dst_metric(dst
, RTAX_ADVMSS
);
106 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
107 * This is the first part of cwnd validation mechanism. */
108 static void tcp_cwnd_restart(struct tcp_sock
*tp
, struct dst_entry
*dst
)
110 s32 delta
= tcp_time_stamp
- tp
->lsndtime
;
111 u32 restart_cwnd
= tcp_init_cwnd(tp
, dst
);
112 u32 cwnd
= tp
->snd_cwnd
;
114 tcp_ca_event(tp
, CA_EVENT_CWND_RESTART
);
116 tp
->snd_ssthresh
= tcp_current_ssthresh(tp
);
117 restart_cwnd
= min(restart_cwnd
, cwnd
);
119 while ((delta
-= tp
->rto
) > 0 && cwnd
> restart_cwnd
)
121 tp
->snd_cwnd
= max(cwnd
, restart_cwnd
);
122 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
123 tp
->snd_cwnd_used
= 0;
126 static inline void tcp_event_data_sent(struct tcp_sock
*tp
,
127 struct sk_buff
*skb
, struct sock
*sk
)
129 u32 now
= tcp_time_stamp
;
131 if (!tp
->packets_out
&& (s32
)(now
- tp
->lsndtime
) > tp
->rto
)
132 tcp_cwnd_restart(tp
, __sk_dst_get(sk
));
136 /* If it is a reply for ato after last received
137 * packet, enter pingpong mode.
139 if ((u32
)(now
- tp
->ack
.lrcvtime
) < tp
->ack
.ato
)
140 tp
->ack
.pingpong
= 1;
143 static __inline__
void tcp_event_ack_sent(struct sock
*sk
, unsigned int pkts
)
145 struct tcp_sock
*tp
= tcp_sk(sk
);
147 tcp_dec_quickack_mode(tp
, pkts
);
148 tcp_clear_xmit_timer(sk
, TCP_TIME_DACK
);
151 /* Determine a window scaling and initial window to offer.
152 * Based on the assumption that the given amount of space
153 * will be offered. Store the results in the tp structure.
154 * NOTE: for smooth operation initial space offering should
155 * be a multiple of mss if possible. We assume here that mss >= 1.
156 * This MUST be enforced by all callers.
158 void tcp_select_initial_window(int __space
, __u32 mss
,
159 __u32
*rcv_wnd
, __u32
*window_clamp
,
160 int wscale_ok
, __u8
*rcv_wscale
)
162 unsigned int space
= (__space
< 0 ? 0 : __space
);
164 /* If no clamp set the clamp to the max possible scaled window */
165 if (*window_clamp
== 0)
166 (*window_clamp
) = (65535 << 14);
167 space
= min(*window_clamp
, space
);
169 /* Quantize space offering to a multiple of mss if possible. */
171 space
= (space
/ mss
) * mss
;
173 /* NOTE: offering an initial window larger than 32767
174 * will break some buggy TCP stacks. We try to be nice.
175 * If we are not window scaling, then this truncates
176 * our initial window offering to 32k. There should also
177 * be a sysctl option to stop being nice.
179 (*rcv_wnd
) = min(space
, MAX_TCP_WINDOW
);
182 /* Set window scaling on max possible window
183 * See RFC1323 for an explanation of the limit to 14
185 space
= max_t(u32
, sysctl_tcp_rmem
[2], sysctl_rmem_max
);
186 while (space
> 65535 && (*rcv_wscale
) < 14) {
192 /* Set initial window to value enough for senders,
193 * following RFC1414. Senders, not following this RFC,
194 * will be satisfied with 2.
196 if (mss
> (1<<*rcv_wscale
)) {
202 if (*rcv_wnd
> init_cwnd
*mss
)
203 *rcv_wnd
= init_cwnd
*mss
;
206 /* Set the clamp no higher than max representable value */
207 (*window_clamp
) = min(65535U << (*rcv_wscale
), *window_clamp
);
210 /* Chose a new window to advertise, update state in tcp_sock for the
211 * socket, and return result with RFC1323 scaling applied. The return
212 * value can be stuffed directly into th->window for an outgoing
215 static __inline__ u16
tcp_select_window(struct sock
*sk
)
217 struct tcp_sock
*tp
= tcp_sk(sk
);
218 u32 cur_win
= tcp_receive_window(tp
);
219 u32 new_win
= __tcp_select_window(sk
);
221 /* Never shrink the offered window */
222 if(new_win
< cur_win
) {
223 /* Danger Will Robinson!
224 * Don't update rcv_wup/rcv_wnd here or else
225 * we will not be able to advertise a zero
226 * window in time. --DaveM
228 * Relax Will Robinson.
232 tp
->rcv_wnd
= new_win
;
233 tp
->rcv_wup
= tp
->rcv_nxt
;
235 /* Make sure we do not exceed the maximum possible
238 if (!tp
->rx_opt
.rcv_wscale
)
239 new_win
= min(new_win
, MAX_TCP_WINDOW
);
241 new_win
= min(new_win
, (65535U << tp
->rx_opt
.rcv_wscale
));
243 /* RFC1323 scaling applied */
244 new_win
>>= tp
->rx_opt
.rcv_wscale
;
246 /* If we advertise zero window, disable fast path. */
254 /* This routine actually transmits TCP packets queued in by
255 * tcp_do_sendmsg(). This is used by both the initial
256 * transmission and possible later retransmissions.
257 * All SKB's seen here are completely headerless. It is our
258 * job to build the TCP header, and pass the packet down to
259 * IP so it can do the same plus pass the packet off to the
262 * We are working here with either a clone of the original
263 * SKB, or a fresh unique copy made by the retransmit engine.
265 static int tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
268 struct inet_sock
*inet
= inet_sk(sk
);
269 struct tcp_sock
*tp
= tcp_sk(sk
);
270 struct tcp_skb_cb
*tcb
= TCP_SKB_CB(skb
);
271 int tcp_header_size
= tp
->tcp_header_len
;
276 BUG_ON(!tcp_skb_pcount(skb
));
278 #define SYSCTL_FLAG_TSTAMPS 0x1
279 #define SYSCTL_FLAG_WSCALE 0x2
280 #define SYSCTL_FLAG_SACK 0x4
282 /* If congestion control is doing timestamping */
283 if (tp
->ca_ops
->rtt_sample
)
284 do_gettimeofday(&skb
->stamp
);
287 if (tcb
->flags
& TCPCB_FLAG_SYN
) {
288 tcp_header_size
= sizeof(struct tcphdr
) + TCPOLEN_MSS
;
289 if(sysctl_tcp_timestamps
) {
290 tcp_header_size
+= TCPOLEN_TSTAMP_ALIGNED
;
291 sysctl_flags
|= SYSCTL_FLAG_TSTAMPS
;
293 if(sysctl_tcp_window_scaling
) {
294 tcp_header_size
+= TCPOLEN_WSCALE_ALIGNED
;
295 sysctl_flags
|= SYSCTL_FLAG_WSCALE
;
297 if(sysctl_tcp_sack
) {
298 sysctl_flags
|= SYSCTL_FLAG_SACK
;
299 if(!(sysctl_flags
& SYSCTL_FLAG_TSTAMPS
))
300 tcp_header_size
+= TCPOLEN_SACKPERM_ALIGNED
;
302 } else if (tp
->rx_opt
.eff_sacks
) {
303 /* A SACK is 2 pad bytes, a 2 byte header, plus
304 * 2 32-bit sequence numbers for each SACK block.
306 tcp_header_size
+= (TCPOLEN_SACK_BASE_ALIGNED
+
307 (tp
->rx_opt
.eff_sacks
* TCPOLEN_SACK_PERBLOCK
));
310 if (tcp_packets_in_flight(tp
) == 0)
311 tcp_ca_event(tp
, CA_EVENT_TX_START
);
313 th
= (struct tcphdr
*) skb_push(skb
, tcp_header_size
);
315 skb_set_owner_w(skb
, sk
);
317 /* Build TCP header and checksum it. */
318 th
->source
= inet
->sport
;
319 th
->dest
= inet
->dport
;
320 th
->seq
= htonl(tcb
->seq
);
321 th
->ack_seq
= htonl(tp
->rcv_nxt
);
322 *(((__u16
*)th
) + 6) = htons(((tcp_header_size
>> 2) << 12) | tcb
->flags
);
323 if (tcb
->flags
& TCPCB_FLAG_SYN
) {
324 /* RFC1323: The window in SYN & SYN/ACK segments
327 th
->window
= htons(tp
->rcv_wnd
);
329 th
->window
= htons(tcp_select_window(sk
));
335 between(tp
->snd_up
, tcb
->seq
+1, tcb
->seq
+0xFFFF)) {
336 th
->urg_ptr
= htons(tp
->snd_up
-tcb
->seq
);
340 if (tcb
->flags
& TCPCB_FLAG_SYN
) {
341 tcp_syn_build_options((__u32
*)(th
+ 1),
342 tcp_advertise_mss(sk
),
343 (sysctl_flags
& SYSCTL_FLAG_TSTAMPS
),
344 (sysctl_flags
& SYSCTL_FLAG_SACK
),
345 (sysctl_flags
& SYSCTL_FLAG_WSCALE
),
346 tp
->rx_opt
.rcv_wscale
,
348 tp
->rx_opt
.ts_recent
);
350 tcp_build_and_update_options((__u32
*)(th
+ 1),
353 TCP_ECN_send(sk
, tp
, skb
, tcp_header_size
);
355 tp
->af_specific
->send_check(sk
, th
, skb
->len
, skb
);
357 if (tcb
->flags
& TCPCB_FLAG_ACK
)
358 tcp_event_ack_sent(sk
, tcp_skb_pcount(skb
));
360 if (skb
->len
!= tcp_header_size
)
361 tcp_event_data_sent(tp
, skb
, sk
);
363 TCP_INC_STATS(TCP_MIB_OUTSEGS
);
365 err
= tp
->af_specific
->queue_xmit(skb
, 0);
371 /* NET_XMIT_CN is special. It does not guarantee,
372 * that this packet is lost. It tells that device
373 * is about to start to drop packets or already
374 * drops some packets of the same priority and
375 * invokes us to send less aggressively.
377 return err
== NET_XMIT_CN
? 0 : err
;
380 #undef SYSCTL_FLAG_TSTAMPS
381 #undef SYSCTL_FLAG_WSCALE
382 #undef SYSCTL_FLAG_SACK
386 /* This routine just queue's the buffer
388 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
389 * otherwise socket can stall.
391 static void tcp_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
393 struct tcp_sock
*tp
= tcp_sk(sk
);
395 /* Advance write_seq and place onto the write_queue. */
396 tp
->write_seq
= TCP_SKB_CB(skb
)->end_seq
;
397 skb_header_release(skb
);
398 __skb_queue_tail(&sk
->sk_write_queue
, skb
);
399 sk_charge_skb(sk
, skb
);
401 /* Queue it, remembering where we must start sending. */
402 if (sk
->sk_send_head
== NULL
)
403 sk
->sk_send_head
= skb
;
406 static void tcp_set_skb_tso_segs(struct sock
*sk
, struct sk_buff
*skb
)
408 struct tcp_sock
*tp
= tcp_sk(sk
);
410 if (skb
->len
<= tp
->mss_cache
||
411 !(sk
->sk_route_caps
& NETIF_F_TSO
)) {
412 /* Avoid the costly divide in the normal
415 skb_shinfo(skb
)->tso_segs
= 1;
416 skb_shinfo(skb
)->tso_size
= 0;
420 factor
= skb
->len
+ (tp
->mss_cache
- 1);
421 factor
/= tp
->mss_cache
;
422 skb_shinfo(skb
)->tso_segs
= factor
;
423 skb_shinfo(skb
)->tso_size
= tp
->mss_cache
;
427 /* Function to create two new TCP segments. Shrinks the given segment
428 * to the specified size and appends a new segment with the rest of the
429 * packet to the list. This won't be called frequently, I hope.
430 * Remember, these are still headerless SKBs at this point.
432 static int tcp_fragment(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
434 struct tcp_sock
*tp
= tcp_sk(sk
);
435 struct sk_buff
*buff
;
439 nsize
= skb_headlen(skb
) - len
;
443 if (skb_cloned(skb
) &&
444 skb_is_nonlinear(skb
) &&
445 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
448 /* Get a new skb... force flag on. */
449 buff
= sk_stream_alloc_skb(sk
, nsize
, GFP_ATOMIC
);
451 return -ENOMEM
; /* We'll just try again later. */
452 sk_charge_skb(sk
, buff
);
454 /* Correct the sequence numbers. */
455 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
456 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
457 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
459 /* PSH and FIN should only be set in the second packet. */
460 flags
= TCP_SKB_CB(skb
)->flags
;
461 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
462 TCP_SKB_CB(buff
)->flags
= flags
;
463 TCP_SKB_CB(buff
)->sacked
=
464 (TCP_SKB_CB(skb
)->sacked
&
465 (TCPCB_LOST
| TCPCB_EVER_RETRANS
| TCPCB_AT_TAIL
));
466 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_AT_TAIL
;
468 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_HW
) {
469 /* Copy and checksum data tail into the new buffer. */
470 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
, skb_put(buff
, nsize
),
475 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
477 skb
->ip_summed
= CHECKSUM_HW
;
478 skb_split(skb
, buff
, len
);
481 buff
->ip_summed
= skb
->ip_summed
;
483 /* Looks stupid, but our code really uses when of
484 * skbs, which it never sent before. --ANK
486 TCP_SKB_CB(buff
)->when
= TCP_SKB_CB(skb
)->when
;
487 buff
->stamp
= skb
->stamp
;
489 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
) {
490 tp
->lost_out
-= tcp_skb_pcount(skb
);
491 tp
->left_out
-= tcp_skb_pcount(skb
);
494 /* Fix up tso_factor for both original and new SKB. */
495 tcp_set_skb_tso_segs(sk
, skb
);
496 tcp_set_skb_tso_segs(sk
, buff
);
498 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
) {
499 tp
->lost_out
+= tcp_skb_pcount(skb
);
500 tp
->left_out
+= tcp_skb_pcount(skb
);
503 if (TCP_SKB_CB(buff
)->sacked
&TCPCB_LOST
) {
504 tp
->lost_out
+= tcp_skb_pcount(buff
);
505 tp
->left_out
+= tcp_skb_pcount(buff
);
508 /* Link BUFF into the send queue. */
509 skb_header_release(buff
);
510 __skb_append(skb
, buff
);
515 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
516 * eventually). The difference is that pulled data not copied, but
517 * immediately discarded.
519 static unsigned char *__pskb_trim_head(struct sk_buff
*skb
, int len
)
525 for (i
=0; i
<skb_shinfo(skb
)->nr_frags
; i
++) {
526 if (skb_shinfo(skb
)->frags
[i
].size
<= eat
) {
527 put_page(skb_shinfo(skb
)->frags
[i
].page
);
528 eat
-= skb_shinfo(skb
)->frags
[i
].size
;
530 skb_shinfo(skb
)->frags
[k
] = skb_shinfo(skb
)->frags
[i
];
532 skb_shinfo(skb
)->frags
[k
].page_offset
+= eat
;
533 skb_shinfo(skb
)->frags
[k
].size
-= eat
;
539 skb_shinfo(skb
)->nr_frags
= k
;
541 skb
->tail
= skb
->data
;
542 skb
->data_len
-= len
;
543 skb
->len
= skb
->data_len
;
547 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
549 if (skb_cloned(skb
) &&
550 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
553 if (len
<= skb_headlen(skb
)) {
554 __skb_pull(skb
, len
);
556 if (__pskb_trim_head(skb
, len
-skb_headlen(skb
)) == NULL
)
560 TCP_SKB_CB(skb
)->seq
+= len
;
561 skb
->ip_summed
= CHECKSUM_HW
;
563 skb
->truesize
-= len
;
564 sk
->sk_wmem_queued
-= len
;
565 sk
->sk_forward_alloc
+= len
;
566 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
568 /* Any change of skb->len requires recalculation of tso
571 if (tcp_skb_pcount(skb
) > 1)
572 tcp_set_skb_tso_segs(sk
, skb
);
577 /* This function synchronize snd mss to current pmtu/exthdr set.
579 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
580 for TCP options, but includes only bare TCP header.
582 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
583 It is minumum of user_mss and mss received with SYN.
584 It also does not include TCP options.
586 tp->pmtu_cookie is last pmtu, seen by this function.
588 tp->mss_cache is current effective sending mss, including
589 all tcp options except for SACKs. It is evaluated,
590 taking into account current pmtu, but never exceeds
591 tp->rx_opt.mss_clamp.
593 NOTE1. rfc1122 clearly states that advertised MSS
594 DOES NOT include either tcp or ip options.
596 NOTE2. tp->pmtu_cookie and tp->mss_cache are READ ONLY outside
597 this function. --ANK (980731)
600 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
602 struct tcp_sock
*tp
= tcp_sk(sk
);
605 /* Calculate base mss without TCP options:
606 It is MMS_S - sizeof(tcphdr) of rfc1122
608 mss_now
= pmtu
- tp
->af_specific
->net_header_len
- sizeof(struct tcphdr
);
610 /* Clamp it (mss_clamp does not include tcp options) */
611 if (mss_now
> tp
->rx_opt
.mss_clamp
)
612 mss_now
= tp
->rx_opt
.mss_clamp
;
614 /* Now subtract optional transport overhead */
615 mss_now
-= tp
->ext_header_len
;
617 /* Then reserve room for full set of TCP options and 8 bytes of data */
621 /* Now subtract TCP options size, not including SACKs */
622 mss_now
-= tp
->tcp_header_len
- sizeof(struct tcphdr
);
624 /* Bound mss with half of window */
625 if (tp
->max_window
&& mss_now
> (tp
->max_window
>>1))
626 mss_now
= max((tp
->max_window
>>1), 68U - tp
->tcp_header_len
);
628 /* And store cached results */
629 tp
->pmtu_cookie
= pmtu
;
630 tp
->mss_cache
= mss_now
;
635 /* Compute the current effective MSS, taking SACKs and IP options,
636 * and even PMTU discovery events into account.
638 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
639 * cannot be large. However, taking into account rare use of URG, this
642 unsigned int tcp_current_mss(struct sock
*sk
, int large_allowed
)
644 struct tcp_sock
*tp
= tcp_sk(sk
);
645 struct dst_entry
*dst
= __sk_dst_get(sk
);
650 mss_now
= tp
->mss_cache
;
653 (sk
->sk_route_caps
& NETIF_F_TSO
) &&
658 u32 mtu
= dst_mtu(dst
);
659 if (mtu
!= tp
->pmtu_cookie
)
660 mss_now
= tcp_sync_mss(sk
, mtu
);
663 if (tp
->rx_opt
.eff_sacks
)
664 mss_now
-= (TCPOLEN_SACK_BASE_ALIGNED
+
665 (tp
->rx_opt
.eff_sacks
* TCPOLEN_SACK_PERBLOCK
));
667 xmit_size_goal
= mss_now
;
670 xmit_size_goal
= 65535 -
671 tp
->af_specific
->net_header_len
-
672 tp
->ext_header_len
- tp
->tcp_header_len
;
674 if (tp
->max_window
&&
675 (xmit_size_goal
> (tp
->max_window
>> 1)))
676 xmit_size_goal
= max((tp
->max_window
>> 1),
677 68U - tp
->tcp_header_len
);
679 xmit_size_goal
-= (xmit_size_goal
% mss_now
);
681 tp
->xmit_size_goal
= xmit_size_goal
;
686 /* Congestion window validation. (RFC2861) */
688 static inline void tcp_cwnd_validate(struct sock
*sk
, struct tcp_sock
*tp
)
690 __u32 packets_out
= tp
->packets_out
;
692 if (packets_out
>= tp
->snd_cwnd
) {
693 /* Network is feed fully. */
694 tp
->snd_cwnd_used
= 0;
695 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
697 /* Network starves. */
698 if (tp
->packets_out
> tp
->snd_cwnd_used
)
699 tp
->snd_cwnd_used
= tp
->packets_out
;
701 if ((s32
)(tcp_time_stamp
- tp
->snd_cwnd_stamp
) >= tp
->rto
)
702 tcp_cwnd_application_limited(sk
);
706 static unsigned int tcp_window_allows(struct tcp_sock
*tp
, struct sk_buff
*skb
, unsigned int mss_now
, unsigned int cwnd
)
708 u32 window
, cwnd_len
;
710 window
= (tp
->snd_una
+ tp
->snd_wnd
- TCP_SKB_CB(skb
)->seq
);
711 cwnd_len
= mss_now
* cwnd
;
712 return min(window
, cwnd_len
);
715 /* Can at least one segment of SKB be sent right now, according to the
716 * congestion window rules? If so, return how many segments are allowed.
718 static inline unsigned int tcp_cwnd_test(struct tcp_sock
*tp
, struct sk_buff
*skb
)
722 /* Don't be strict about the congestion window for the final FIN. */
723 if (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
)
726 in_flight
= tcp_packets_in_flight(tp
);
728 if (in_flight
< cwnd
)
729 return (cwnd
- in_flight
);
734 /* This must be invoked the first time we consider transmitting
737 static inline int tcp_init_tso_segs(struct sock
*sk
, struct sk_buff
*skb
)
739 int tso_segs
= tcp_skb_pcount(skb
);
742 tcp_set_skb_tso_segs(sk
, skb
);
743 tso_segs
= tcp_skb_pcount(skb
);
748 static inline int tcp_minshall_check(const struct tcp_sock
*tp
)
750 return after(tp
->snd_sml
,tp
->snd_una
) &&
751 !after(tp
->snd_sml
, tp
->snd_nxt
);
754 /* Return 0, if packet can be sent now without violation Nagle's rules:
755 * 1. It is full sized.
756 * 2. Or it contains FIN. (already checked by caller)
757 * 3. Or TCP_NODELAY was set.
758 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
759 * With Minshall's modification: all sent small packets are ACKed.
762 static inline int tcp_nagle_check(const struct tcp_sock
*tp
,
763 const struct sk_buff
*skb
,
764 unsigned mss_now
, int nonagle
)
766 return (skb
->len
< mss_now
&&
767 ((nonagle
&TCP_NAGLE_CORK
) ||
770 tcp_minshall_check(tp
))));
773 /* Return non-zero if the Nagle test allows this packet to be
776 static inline int tcp_nagle_test(struct tcp_sock
*tp
, struct sk_buff
*skb
,
777 unsigned int cur_mss
, int nonagle
)
779 /* Nagle rule does not apply to frames, which sit in the middle of the
780 * write_queue (they have no chances to get new data).
782 * This is implemented in the callers, where they modify the 'nonagle'
783 * argument based upon the location of SKB in the send queue.
785 if (nonagle
& TCP_NAGLE_PUSH
)
788 /* Don't use the nagle rule for urgent data (or for the final FIN). */
790 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
))
793 if (!tcp_nagle_check(tp
, skb
, cur_mss
, nonagle
))
799 /* Does at least the first segment of SKB fit into the send window? */
800 static inline int tcp_snd_wnd_test(struct tcp_sock
*tp
, struct sk_buff
*skb
, unsigned int cur_mss
)
802 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
804 if (skb
->len
> cur_mss
)
805 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
807 return !after(end_seq
, tp
->snd_una
+ tp
->snd_wnd
);
810 /* This checks if the data bearing packet SKB (usually sk->sk_send_head)
811 * should be put on the wire right now. If so, it returns the number of
812 * packets allowed by the congestion window.
814 static unsigned int tcp_snd_test(struct sock
*sk
, struct sk_buff
*skb
,
815 unsigned int cur_mss
, int nonagle
)
817 struct tcp_sock
*tp
= tcp_sk(sk
);
818 unsigned int cwnd_quota
;
820 tcp_init_tso_segs(sk
, skb
);
822 if (!tcp_nagle_test(tp
, skb
, cur_mss
, nonagle
))
825 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
827 !tcp_snd_wnd_test(tp
, skb
, cur_mss
))
833 static inline int tcp_skb_is_last(const struct sock
*sk
,
834 const struct sk_buff
*skb
)
836 return skb
->next
== (struct sk_buff
*)&sk
->sk_write_queue
;
839 int tcp_may_send_now(struct sock
*sk
, struct tcp_sock
*tp
)
841 struct sk_buff
*skb
= sk
->sk_send_head
;
844 tcp_snd_test(sk
, skb
, tcp_current_mss(sk
, 1),
845 (tcp_skb_is_last(sk
, skb
) ?
850 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
851 * which is put after SKB on the list. It is very much like
852 * tcp_fragment() except that it may make several kinds of assumptions
853 * in order to speed up the splitting operation. In particular, we
854 * know that all the data is in scatter-gather pages, and that the
855 * packet has never been sent out before (and thus is not cloned).
857 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
)
859 struct sk_buff
*buff
;
860 int nlen
= skb
->len
- len
;
863 /* All of a TSO frame must be composed of paged data. */
864 BUG_ON(skb
->len
!= skb
->data_len
);
866 buff
= sk_stream_alloc_pskb(sk
, 0, 0, GFP_ATOMIC
);
867 if (unlikely(buff
== NULL
))
870 buff
->truesize
= nlen
;
871 skb
->truesize
-= nlen
;
873 /* Correct the sequence numbers. */
874 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
875 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
876 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
878 /* PSH and FIN should only be set in the second packet. */
879 flags
= TCP_SKB_CB(skb
)->flags
;
880 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
881 TCP_SKB_CB(buff
)->flags
= flags
;
883 /* This packet was never sent out yet, so no SACK bits. */
884 TCP_SKB_CB(buff
)->sacked
= 0;
886 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_HW
;
887 skb_split(skb
, buff
, len
);
889 /* Fix up tso_factor for both original and new SKB. */
890 tcp_set_skb_tso_segs(sk
, skb
);
891 tcp_set_skb_tso_segs(sk
, buff
);
893 /* Link BUFF into the send queue. */
894 skb_header_release(buff
);
895 __skb_append(skb
, buff
);
900 /* Try to defer sending, if possible, in order to minimize the amount
901 * of TSO splitting we do. View it as a kind of TSO Nagle test.
903 * This algorithm is from John Heffner.
905 static int tcp_tso_should_defer(struct sock
*sk
, struct tcp_sock
*tp
, struct sk_buff
*skb
)
907 u32 send_win
, cong_win
, limit
, in_flight
;
909 if (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
)
912 if (tp
->ca_state
!= TCP_CA_Open
)
915 in_flight
= tcp_packets_in_flight(tp
);
917 BUG_ON(tcp_skb_pcount(skb
) <= 1 ||
918 (tp
->snd_cwnd
<= in_flight
));
920 send_win
= (tp
->snd_una
+ tp
->snd_wnd
) - TCP_SKB_CB(skb
)->seq
;
922 /* From in_flight test above, we know that cwnd > in_flight. */
923 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
925 limit
= min(send_win
, cong_win
);
927 /* If sk_send_head can be sent fully now, just do it. */
928 if (skb
->len
<= limit
)
931 if (sysctl_tcp_tso_win_divisor
) {
932 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
934 /* If at least some fraction of a window is available,
937 chunk
/= sysctl_tcp_tso_win_divisor
;
941 /* Different approach, try not to defer past a single
942 * ACK. Receiver should ACK every other full sized
943 * frame, so if we have space for more than 3 frames
946 if (limit
> tcp_max_burst(tp
) * tp
->mss_cache
)
950 /* Ok, it looks like it is advisable to defer. */
954 /* This routine writes packets to the network. It advances the
955 * send_head. This happens as incoming acks open up the remote
958 * Returns 1, if no segments are in flight and we have queued segments, but
959 * cannot send anything now because of SWS or another problem.
961 static int tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
)
963 struct tcp_sock
*tp
= tcp_sk(sk
);
965 unsigned int tso_segs
, sent_pkts
;
968 /* If we are closed, the bytes will have to remain here.
969 * In time closedown will finish, we empty the write queue and all
972 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
975 skb
= sk
->sk_send_head
;
979 tso_segs
= tcp_init_tso_segs(sk
, skb
);
980 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
981 if (unlikely(!cwnd_quota
))
985 while (likely(tcp_snd_wnd_test(tp
, skb
, mss_now
))) {
989 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
990 (tcp_skb_is_last(sk
, skb
) ?
991 nonagle
: TCP_NAGLE_PUSH
))))
994 if (tcp_tso_should_defer(sk
, tp
, skb
))
999 u32 limit
= tcp_window_allows(tp
, skb
,
1000 mss_now
, cwnd_quota
);
1002 if (skb
->len
< limit
) {
1003 unsigned int trim
= skb
->len
% mss_now
;
1006 limit
= skb
->len
- trim
;
1008 if (skb
->len
> limit
) {
1009 if (tso_fragment(sk
, skb
, limit
))
1012 } else if (unlikely(skb
->len
> mss_now
)) {
1013 if (unlikely(tcp_fragment(sk
, skb
, mss_now
)))
1017 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1019 if (unlikely(tcp_transmit_skb(sk
, skb_clone(skb
, GFP_ATOMIC
))))
1022 /* Advance the send_head. This one is sent out.
1023 * This call will increment packets_out.
1025 update_send_head(sk
, tp
, skb
);
1027 tcp_minshall_update(tp
, mss_now
, skb
);
1030 /* Do not optimize this to use tso_segs. If we chopped up
1031 * the packet above, tso_segs will no longer be valid.
1033 cwnd_quota
-= tcp_skb_pcount(skb
);
1035 BUG_ON(cwnd_quota
< 0);
1039 skb
= sk
->sk_send_head
;
1042 tso_segs
= tcp_init_tso_segs(sk
, skb
);
1045 if (likely(sent_pkts
)) {
1046 tcp_cwnd_validate(sk
, tp
);
1050 return !tp
->packets_out
&& sk
->sk_send_head
;
1053 /* Push out any pending frames which were held back due to
1054 * TCP_CORK or attempt at coalescing tiny packets.
1055 * The socket must be locked by the caller.
1057 void __tcp_push_pending_frames(struct sock
*sk
, struct tcp_sock
*tp
,
1058 unsigned int cur_mss
, int nonagle
)
1060 struct sk_buff
*skb
= sk
->sk_send_head
;
1063 if (tcp_write_xmit(sk
, cur_mss
, nonagle
))
1064 tcp_check_probe_timer(sk
, tp
);
1068 /* Send _single_ skb sitting at the send head. This function requires
1069 * true push pending frames to setup probe timer etc.
1071 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
1073 struct tcp_sock
*tp
= tcp_sk(sk
);
1074 struct sk_buff
*skb
= sk
->sk_send_head
;
1075 unsigned int tso_segs
, cwnd_quota
;
1077 BUG_ON(!skb
|| skb
->len
< mss_now
);
1079 tso_segs
= tcp_init_tso_segs(sk
, skb
);
1080 cwnd_quota
= tcp_snd_test(sk
, skb
, mss_now
, TCP_NAGLE_PUSH
);
1082 if (likely(cwnd_quota
)) {
1086 u32 limit
= tcp_window_allows(tp
, skb
,
1087 mss_now
, cwnd_quota
);
1089 if (skb
->len
< limit
) {
1090 unsigned int trim
= skb
->len
% mss_now
;
1093 limit
= skb
->len
- trim
;
1095 if (skb
->len
> limit
) {
1096 if (unlikely(tso_fragment(sk
, skb
, limit
)))
1099 } else if (unlikely(skb
->len
> mss_now
)) {
1100 if (unlikely(tcp_fragment(sk
, skb
, mss_now
)))
1104 /* Send it out now. */
1105 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1107 if (likely(!tcp_transmit_skb(sk
, skb_clone(skb
, sk
->sk_allocation
)))) {
1108 update_send_head(sk
, tp
, skb
);
1109 tcp_cwnd_validate(sk
, tp
);
1115 /* This function returns the amount that we can raise the
1116 * usable window based on the following constraints
1118 * 1. The window can never be shrunk once it is offered (RFC 793)
1119 * 2. We limit memory per socket
1122 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1123 * RECV.NEXT + RCV.WIN fixed until:
1124 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1126 * i.e. don't raise the right edge of the window until you can raise
1127 * it at least MSS bytes.
1129 * Unfortunately, the recommended algorithm breaks header prediction,
1130 * since header prediction assumes th->window stays fixed.
1132 * Strictly speaking, keeping th->window fixed violates the receiver
1133 * side SWS prevention criteria. The problem is that under this rule
1134 * a stream of single byte packets will cause the right side of the
1135 * window to always advance by a single byte.
1137 * Of course, if the sender implements sender side SWS prevention
1138 * then this will not be a problem.
1140 * BSD seems to make the following compromise:
1142 * If the free space is less than the 1/4 of the maximum
1143 * space available and the free space is less than 1/2 mss,
1144 * then set the window to 0.
1145 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1146 * Otherwise, just prevent the window from shrinking
1147 * and from being larger than the largest representable value.
1149 * This prevents incremental opening of the window in the regime
1150 * where TCP is limited by the speed of the reader side taking
1151 * data out of the TCP receive queue. It does nothing about
1152 * those cases where the window is constrained on the sender side
1153 * because the pipeline is full.
1155 * BSD also seems to "accidentally" limit itself to windows that are a
1156 * multiple of MSS, at least until the free space gets quite small.
1157 * This would appear to be a side effect of the mbuf implementation.
1158 * Combining these two algorithms results in the observed behavior
1159 * of having a fixed window size at almost all times.
1161 * Below we obtain similar behavior by forcing the offered window to
1162 * a multiple of the mss when it is feasible to do so.
1164 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1165 * Regular options like TIMESTAMP are taken into account.
1167 u32
__tcp_select_window(struct sock
*sk
)
1169 struct tcp_sock
*tp
= tcp_sk(sk
);
1170 /* MSS for the peer's data. Previous verions used mss_clamp
1171 * here. I don't know if the value based on our guesses
1172 * of peer's MSS is better for the performance. It's more correct
1173 * but may be worse for the performance because of rcv_mss
1174 * fluctuations. --SAW 1998/11/1
1176 int mss
= tp
->ack
.rcv_mss
;
1177 int free_space
= tcp_space(sk
);
1178 int full_space
= min_t(int, tp
->window_clamp
, tcp_full_space(sk
));
1181 if (mss
> full_space
)
1184 if (free_space
< full_space
/2) {
1187 if (tcp_memory_pressure
)
1188 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U*tp
->advmss
);
1190 if (free_space
< mss
)
1194 if (free_space
> tp
->rcv_ssthresh
)
1195 free_space
= tp
->rcv_ssthresh
;
1197 /* Don't do rounding if we are using window scaling, since the
1198 * scaled window will not line up with the MSS boundary anyway.
1200 window
= tp
->rcv_wnd
;
1201 if (tp
->rx_opt
.rcv_wscale
) {
1202 window
= free_space
;
1204 /* Advertise enough space so that it won't get scaled away.
1205 * Import case: prevent zero window announcement if
1206 * 1<<rcv_wscale > mss.
1208 if (((window
>> tp
->rx_opt
.rcv_wscale
) << tp
->rx_opt
.rcv_wscale
) != window
)
1209 window
= (((window
>> tp
->rx_opt
.rcv_wscale
) + 1)
1210 << tp
->rx_opt
.rcv_wscale
);
1212 /* Get the largest window that is a nice multiple of mss.
1213 * Window clamp already applied above.
1214 * If our current window offering is within 1 mss of the
1215 * free space we just keep it. This prevents the divide
1216 * and multiply from happening most of the time.
1217 * We also don't do any window rounding when the free space
1220 if (window
<= free_space
- mss
|| window
> free_space
)
1221 window
= (free_space
/mss
)*mss
;
1227 /* Attempt to collapse two adjacent SKB's during retransmission. */
1228 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*skb
, int mss_now
)
1230 struct tcp_sock
*tp
= tcp_sk(sk
);
1231 struct sk_buff
*next_skb
= skb
->next
;
1233 /* The first test we must make is that neither of these two
1234 * SKB's are still referenced by someone else.
1236 if (!skb_cloned(skb
) && !skb_cloned(next_skb
)) {
1237 int skb_size
= skb
->len
, next_skb_size
= next_skb
->len
;
1238 u16 flags
= TCP_SKB_CB(skb
)->flags
;
1240 /* Also punt if next skb has been SACK'd. */
1241 if(TCP_SKB_CB(next_skb
)->sacked
& TCPCB_SACKED_ACKED
)
1244 /* Next skb is out of window. */
1245 if (after(TCP_SKB_CB(next_skb
)->end_seq
, tp
->snd_una
+tp
->snd_wnd
))
1248 /* Punt if not enough space exists in the first SKB for
1249 * the data in the second, or the total combined payload
1250 * would exceed the MSS.
1252 if ((next_skb_size
> skb_tailroom(skb
)) ||
1253 ((skb_size
+ next_skb_size
) > mss_now
))
1256 BUG_ON(tcp_skb_pcount(skb
) != 1 ||
1257 tcp_skb_pcount(next_skb
) != 1);
1259 /* Ok. We will be able to collapse the packet. */
1260 __skb_unlink(next_skb
, next_skb
->list
);
1262 memcpy(skb_put(skb
, next_skb_size
), next_skb
->data
, next_skb_size
);
1264 if (next_skb
->ip_summed
== CHECKSUM_HW
)
1265 skb
->ip_summed
= CHECKSUM_HW
;
1267 if (skb
->ip_summed
!= CHECKSUM_HW
)
1268 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
1270 /* Update sequence range on original skb. */
1271 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
1273 /* Merge over control information. */
1274 flags
|= TCP_SKB_CB(next_skb
)->flags
; /* This moves PSH/FIN etc. over */
1275 TCP_SKB_CB(skb
)->flags
= flags
;
1277 /* All done, get rid of second SKB and account for it so
1278 * packet counting does not break.
1280 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
&(TCPCB_EVER_RETRANS
|TCPCB_AT_TAIL
);
1281 if (TCP_SKB_CB(next_skb
)->sacked
&TCPCB_SACKED_RETRANS
)
1282 tp
->retrans_out
-= tcp_skb_pcount(next_skb
);
1283 if (TCP_SKB_CB(next_skb
)->sacked
&TCPCB_LOST
) {
1284 tp
->lost_out
-= tcp_skb_pcount(next_skb
);
1285 tp
->left_out
-= tcp_skb_pcount(next_skb
);
1287 /* Reno case is special. Sigh... */
1288 if (!tp
->rx_opt
.sack_ok
&& tp
->sacked_out
) {
1289 tcp_dec_pcount_approx(&tp
->sacked_out
, next_skb
);
1290 tp
->left_out
-= tcp_skb_pcount(next_skb
);
1293 /* Not quite right: it can be > snd.fack, but
1294 * it is better to underestimate fackets.
1296 tcp_dec_pcount_approx(&tp
->fackets_out
, next_skb
);
1297 tcp_packets_out_dec(tp
, next_skb
);
1298 sk_stream_free_skb(sk
, next_skb
);
1302 /* Do a simple retransmit without using the backoff mechanisms in
1303 * tcp_timer. This is used for path mtu discovery.
1304 * The socket is already locked here.
1306 void tcp_simple_retransmit(struct sock
*sk
)
1308 struct tcp_sock
*tp
= tcp_sk(sk
);
1309 struct sk_buff
*skb
;
1310 unsigned int mss
= tcp_current_mss(sk
, 0);
1313 sk_stream_for_retrans_queue(skb
, sk
) {
1314 if (skb
->len
> mss
&&
1315 !(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
)) {
1316 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) {
1317 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1318 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1320 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_LOST
)) {
1321 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1322 tp
->lost_out
+= tcp_skb_pcount(skb
);
1331 tcp_sync_left_out(tp
);
1333 /* Don't muck with the congestion window here.
1334 * Reason is that we do not increase amount of _data_
1335 * in network, but units changed and effective
1336 * cwnd/ssthresh really reduced now.
1338 if (tp
->ca_state
!= TCP_CA_Loss
) {
1339 tp
->high_seq
= tp
->snd_nxt
;
1340 tp
->snd_ssthresh
= tcp_current_ssthresh(tp
);
1341 tp
->prior_ssthresh
= 0;
1342 tp
->undo_marker
= 0;
1343 tcp_set_ca_state(tp
, TCP_CA_Loss
);
1345 tcp_xmit_retransmit_queue(sk
);
1348 /* This retransmits one SKB. Policy decisions and retransmit queue
1349 * state updates are done by the caller. Returns non-zero if an
1350 * error occurred which prevented the send.
1352 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
1354 struct tcp_sock
*tp
= tcp_sk(sk
);
1355 unsigned int cur_mss
= tcp_current_mss(sk
, 0);
1358 /* Do not sent more than we queued. 1/4 is reserved for possible
1359 * copying overhead: frgagmentation, tunneling, mangling etc.
1361 if (atomic_read(&sk
->sk_wmem_alloc
) >
1362 min(sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2), sk
->sk_sndbuf
))
1365 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
1366 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1369 if (sk
->sk_route_caps
& NETIF_F_TSO
) {
1370 sk
->sk_route_caps
&= ~NETIF_F_TSO
;
1371 sock_set_flag(sk
, SOCK_NO_LARGESEND
);
1374 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
1378 /* If receiver has shrunk his window, and skb is out of
1379 * new window, do not retransmit it. The exception is the
1380 * case, when window is shrunk to zero. In this case
1381 * our retransmit serves as a zero window probe.
1383 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
+tp
->snd_wnd
)
1384 && TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
1387 if (skb
->len
> cur_mss
) {
1388 int old_factor
= tcp_skb_pcount(skb
);
1391 if (tcp_fragment(sk
, skb
, cur_mss
))
1392 return -ENOMEM
; /* We'll try again later. */
1394 /* New SKB created, account for it. */
1395 new_factor
= tcp_skb_pcount(skb
);
1396 tp
->packets_out
-= old_factor
- new_factor
;
1397 tp
->packets_out
+= tcp_skb_pcount(skb
->next
);
1400 /* Collapse two adjacent packets if worthwhile and we can. */
1401 if(!(TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_SYN
) &&
1402 (skb
->len
< (cur_mss
>> 1)) &&
1403 (skb
->next
!= sk
->sk_send_head
) &&
1404 (skb
->next
!= (struct sk_buff
*)&sk
->sk_write_queue
) &&
1405 (skb_shinfo(skb
)->nr_frags
== 0 && skb_shinfo(skb
->next
)->nr_frags
== 0) &&
1406 (tcp_skb_pcount(skb
) == 1 && tcp_skb_pcount(skb
->next
) == 1) &&
1407 (sysctl_tcp_retrans_collapse
!= 0))
1408 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
1410 if(tp
->af_specific
->rebuild_header(sk
))
1411 return -EHOSTUNREACH
; /* Routing failure or similar. */
1413 /* Some Solaris stacks overoptimize and ignore the FIN on a
1414 * retransmit when old data is attached. So strip it off
1415 * since it is cheap to do so and saves bytes on the network.
1418 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
) &&
1419 tp
->snd_una
== (TCP_SKB_CB(skb
)->end_seq
- 1)) {
1420 if (!pskb_trim(skb
, 0)) {
1421 TCP_SKB_CB(skb
)->seq
= TCP_SKB_CB(skb
)->end_seq
- 1;
1422 skb_shinfo(skb
)->tso_segs
= 1;
1423 skb_shinfo(skb
)->tso_size
= 0;
1424 skb
->ip_summed
= CHECKSUM_NONE
;
1429 /* Make a copy, if the first transmission SKB clone we made
1430 * is still in somebody's hands, else make a clone.
1432 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1434 err
= tcp_transmit_skb(sk
, (skb_cloned(skb
) ?
1435 pskb_copy(skb
, GFP_ATOMIC
):
1436 skb_clone(skb
, GFP_ATOMIC
)));
1439 /* Update global TCP statistics. */
1440 TCP_INC_STATS(TCP_MIB_RETRANSSEGS
);
1442 tp
->total_retrans
++;
1444 #if FASTRETRANS_DEBUG > 0
1445 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) {
1446 if (net_ratelimit())
1447 printk(KERN_DEBUG
"retrans_out leaked.\n");
1450 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
1451 tp
->retrans_out
+= tcp_skb_pcount(skb
);
1453 /* Save stamp of the first retransmit. */
1454 if (!tp
->retrans_stamp
)
1455 tp
->retrans_stamp
= TCP_SKB_CB(skb
)->when
;
1459 /* snd_nxt is stored to detect loss of retransmitted segment,
1460 * see tcp_input.c tcp_sacktag_write_queue().
1462 TCP_SKB_CB(skb
)->ack_seq
= tp
->snd_nxt
;
1467 /* This gets called after a retransmit timeout, and the initially
1468 * retransmitted data is acknowledged. It tries to continue
1469 * resending the rest of the retransmit queue, until either
1470 * we've sent it all or the congestion window limit is reached.
1471 * If doing SACK, the first ACK which comes back for a timeout
1472 * based retransmit packet might feed us FACK information again.
1473 * If so, we use it to avoid unnecessarily retransmissions.
1475 void tcp_xmit_retransmit_queue(struct sock
*sk
)
1477 struct tcp_sock
*tp
= tcp_sk(sk
);
1478 struct sk_buff
*skb
;
1479 int packet_cnt
= tp
->lost_out
;
1481 /* First pass: retransmit lost packets. */
1483 sk_stream_for_retrans_queue(skb
, sk
) {
1484 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
1486 /* Assume this retransmit will generate
1487 * only one packet for congestion window
1488 * calculation purposes. This works because
1489 * tcp_retransmit_skb() will chop up the
1490 * packet to be MSS sized and all the
1491 * packet counting works out.
1493 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
1496 if (sacked
&TCPCB_LOST
) {
1497 if (!(sacked
&(TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))) {
1498 if (tcp_retransmit_skb(sk
, skb
))
1500 if (tp
->ca_state
!= TCP_CA_Loss
)
1501 NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS
);
1503 NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS
);
1506 skb_peek(&sk
->sk_write_queue
))
1507 tcp_reset_xmit_timer(sk
, TCP_TIME_RETRANS
, tp
->rto
);
1510 packet_cnt
-= tcp_skb_pcount(skb
);
1511 if (packet_cnt
<= 0)
1517 /* OK, demanded retransmission is finished. */
1519 /* Forward retransmissions are possible only during Recovery. */
1520 if (tp
->ca_state
!= TCP_CA_Recovery
)
1523 /* No forward retransmissions in Reno are possible. */
1524 if (!tp
->rx_opt
.sack_ok
)
1527 /* Yeah, we have to make difficult choice between forward transmission
1528 * and retransmission... Both ways have their merits...
1530 * For now we do not retransmit anything, while we have some new
1534 if (tcp_may_send_now(sk
, tp
))
1539 sk_stream_for_retrans_queue(skb
, sk
) {
1540 /* Similar to the retransmit loop above we
1541 * can pretend that the retransmitted SKB
1542 * we send out here will be composed of one
1543 * real MSS sized packet because tcp_retransmit_skb()
1544 * will fragment it if necessary.
1546 if (++packet_cnt
> tp
->fackets_out
)
1549 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
1552 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
)
1555 /* Ok, retransmit it. */
1556 if (tcp_retransmit_skb(sk
, skb
))
1559 if (skb
== skb_peek(&sk
->sk_write_queue
))
1560 tcp_reset_xmit_timer(sk
, TCP_TIME_RETRANS
, tp
->rto
);
1562 NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS
);
1567 /* Send a fin. The caller locks the socket for us. This cannot be
1568 * allowed to fail queueing a FIN frame under any circumstances.
1570 void tcp_send_fin(struct sock
*sk
)
1572 struct tcp_sock
*tp
= tcp_sk(sk
);
1573 struct sk_buff
*skb
= skb_peek_tail(&sk
->sk_write_queue
);
1576 /* Optimization, tack on the FIN if we have a queue of
1577 * unsent frames. But be careful about outgoing SACKS
1580 mss_now
= tcp_current_mss(sk
, 1);
1582 if (sk
->sk_send_head
!= NULL
) {
1583 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_FIN
;
1584 TCP_SKB_CB(skb
)->end_seq
++;
1587 /* Socket is locked, keep trying until memory is available. */
1589 skb
= alloc_skb(MAX_TCP_HEADER
, GFP_KERNEL
);
1595 /* Reserve space for headers and prepare control bits. */
1596 skb_reserve(skb
, MAX_TCP_HEADER
);
1598 TCP_SKB_CB(skb
)->flags
= (TCPCB_FLAG_ACK
| TCPCB_FLAG_FIN
);
1599 TCP_SKB_CB(skb
)->sacked
= 0;
1600 skb_shinfo(skb
)->tso_segs
= 1;
1601 skb_shinfo(skb
)->tso_size
= 0;
1603 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
1604 TCP_SKB_CB(skb
)->seq
= tp
->write_seq
;
1605 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ 1;
1606 tcp_queue_skb(sk
, skb
);
1608 __tcp_push_pending_frames(sk
, tp
, mss_now
, TCP_NAGLE_OFF
);
1611 /* We get here when a process closes a file descriptor (either due to
1612 * an explicit close() or as a byproduct of exit()'ing) and there
1613 * was unread data in the receive queue. This behavior is recommended
1614 * by draft-ietf-tcpimpl-prob-03.txt section 3.10. -DaveM
1616 void tcp_send_active_reset(struct sock
*sk
, unsigned int __nocast priority
)
1618 struct tcp_sock
*tp
= tcp_sk(sk
);
1619 struct sk_buff
*skb
;
1621 /* NOTE: No TCP options attached and we never retransmit this. */
1622 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
1624 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED
);
1628 /* Reserve space for headers and prepare control bits. */
1629 skb_reserve(skb
, MAX_TCP_HEADER
);
1631 TCP_SKB_CB(skb
)->flags
= (TCPCB_FLAG_ACK
| TCPCB_FLAG_RST
);
1632 TCP_SKB_CB(skb
)->sacked
= 0;
1633 skb_shinfo(skb
)->tso_segs
= 1;
1634 skb_shinfo(skb
)->tso_size
= 0;
1637 TCP_SKB_CB(skb
)->seq
= tcp_acceptable_seq(sk
, tp
);
1638 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
;
1639 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1640 if (tcp_transmit_skb(sk
, skb
))
1641 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED
);
1644 /* WARNING: This routine must only be called when we have already sent
1645 * a SYN packet that crossed the incoming SYN that caused this routine
1646 * to get called. If this assumption fails then the initial rcv_wnd
1647 * and rcv_wscale values will not be correct.
1649 int tcp_send_synack(struct sock
*sk
)
1651 struct sk_buff
* skb
;
1653 skb
= skb_peek(&sk
->sk_write_queue
);
1654 if (skb
== NULL
|| !(TCP_SKB_CB(skb
)->flags
&TCPCB_FLAG_SYN
)) {
1655 printk(KERN_DEBUG
"tcp_send_synack: wrong queue state\n");
1658 if (!(TCP_SKB_CB(skb
)->flags
&TCPCB_FLAG_ACK
)) {
1659 if (skb_cloned(skb
)) {
1660 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
1663 __skb_unlink(skb
, &sk
->sk_write_queue
);
1664 skb_header_release(nskb
);
1665 __skb_queue_head(&sk
->sk_write_queue
, nskb
);
1666 sk_stream_free_skb(sk
, skb
);
1667 sk_charge_skb(sk
, nskb
);
1671 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_ACK
;
1672 TCP_ECN_send_synack(tcp_sk(sk
), skb
);
1674 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1675 return tcp_transmit_skb(sk
, skb_clone(skb
, GFP_ATOMIC
));
1679 * Prepare a SYN-ACK.
1681 struct sk_buff
* tcp_make_synack(struct sock
*sk
, struct dst_entry
*dst
,
1682 struct request_sock
*req
)
1684 struct inet_request_sock
*ireq
= inet_rsk(req
);
1685 struct tcp_sock
*tp
= tcp_sk(sk
);
1687 int tcp_header_size
;
1688 struct sk_buff
*skb
;
1690 skb
= sock_wmalloc(sk
, MAX_TCP_HEADER
+ 15, 1, GFP_ATOMIC
);
1694 /* Reserve space for headers. */
1695 skb_reserve(skb
, MAX_TCP_HEADER
);
1697 skb
->dst
= dst_clone(dst
);
1699 tcp_header_size
= (sizeof(struct tcphdr
) + TCPOLEN_MSS
+
1700 (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0) +
1701 (ireq
->wscale_ok
? TCPOLEN_WSCALE_ALIGNED
: 0) +
1702 /* SACK_PERM is in the place of NOP NOP of TS */
1703 ((ireq
->sack_ok
&& !ireq
->tstamp_ok
) ? TCPOLEN_SACKPERM_ALIGNED
: 0));
1704 skb
->h
.th
= th
= (struct tcphdr
*) skb_push(skb
, tcp_header_size
);
1706 memset(th
, 0, sizeof(struct tcphdr
));
1709 if (dst
->dev
->features
&NETIF_F_TSO
)
1711 TCP_ECN_make_synack(req
, th
);
1712 th
->source
= inet_sk(sk
)->sport
;
1713 th
->dest
= ireq
->rmt_port
;
1714 TCP_SKB_CB(skb
)->seq
= tcp_rsk(req
)->snt_isn
;
1715 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ 1;
1716 TCP_SKB_CB(skb
)->sacked
= 0;
1717 skb_shinfo(skb
)->tso_segs
= 1;
1718 skb_shinfo(skb
)->tso_size
= 0;
1719 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
1720 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_isn
+ 1);
1721 if (req
->rcv_wnd
== 0) { /* ignored for retransmitted syns */
1723 /* Set this up on the first call only */
1724 req
->window_clamp
= tp
->window_clamp
? : dst_metric(dst
, RTAX_WINDOW
);
1725 /* tcp_full_space because it is guaranteed to be the first packet */
1726 tcp_select_initial_window(tcp_full_space(sk
),
1727 dst_metric(dst
, RTAX_ADVMSS
) - (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0),
1732 ireq
->rcv_wscale
= rcv_wscale
;
1735 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
1736 th
->window
= htons(req
->rcv_wnd
);
1738 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1739 tcp_syn_build_options((__u32
*)(th
+ 1), dst_metric(dst
, RTAX_ADVMSS
), ireq
->tstamp_ok
,
1740 ireq
->sack_ok
, ireq
->wscale_ok
, ireq
->rcv_wscale
,
1741 TCP_SKB_CB(skb
)->when
,
1745 th
->doff
= (tcp_header_size
>> 2);
1746 TCP_INC_STATS(TCP_MIB_OUTSEGS
);
1751 * Do all connect socket setups that can be done AF independent.
1753 static inline void tcp_connect_init(struct sock
*sk
)
1755 struct dst_entry
*dst
= __sk_dst_get(sk
);
1756 struct tcp_sock
*tp
= tcp_sk(sk
);
1759 /* We'll fix this up when we get a response from the other end.
1760 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
1762 tp
->tcp_header_len
= sizeof(struct tcphdr
) +
1763 (sysctl_tcp_timestamps
? TCPOLEN_TSTAMP_ALIGNED
: 0);
1765 /* If user gave his TCP_MAXSEG, record it to clamp */
1766 if (tp
->rx_opt
.user_mss
)
1767 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
1769 tcp_sync_mss(sk
, dst_mtu(dst
));
1771 if (!tp
->window_clamp
)
1772 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
1773 tp
->advmss
= dst_metric(dst
, RTAX_ADVMSS
);
1774 tcp_initialize_rcv_mss(sk
);
1776 tcp_select_initial_window(tcp_full_space(sk
),
1777 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
1780 sysctl_tcp_window_scaling
,
1783 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
1784 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
1787 sock_reset_flag(sk
, SOCK_DONE
);
1789 tcp_init_wl(tp
, tp
->write_seq
, 0);
1790 tp
->snd_una
= tp
->write_seq
;
1791 tp
->snd_sml
= tp
->write_seq
;
1796 tp
->rto
= TCP_TIMEOUT_INIT
;
1797 tp
->retransmits
= 0;
1798 tcp_clear_retrans(tp
);
1802 * Build a SYN and send it off.
1804 int tcp_connect(struct sock
*sk
)
1806 struct tcp_sock
*tp
= tcp_sk(sk
);
1807 struct sk_buff
*buff
;
1809 tcp_connect_init(sk
);
1811 buff
= alloc_skb(MAX_TCP_HEADER
+ 15, sk
->sk_allocation
);
1812 if (unlikely(buff
== NULL
))
1815 /* Reserve space for headers. */
1816 skb_reserve(buff
, MAX_TCP_HEADER
);
1818 TCP_SKB_CB(buff
)->flags
= TCPCB_FLAG_SYN
;
1819 TCP_ECN_send_syn(sk
, tp
, buff
);
1820 TCP_SKB_CB(buff
)->sacked
= 0;
1821 skb_shinfo(buff
)->tso_segs
= 1;
1822 skb_shinfo(buff
)->tso_size
= 0;
1824 TCP_SKB_CB(buff
)->seq
= tp
->write_seq
++;
1825 TCP_SKB_CB(buff
)->end_seq
= tp
->write_seq
;
1826 tp
->snd_nxt
= tp
->write_seq
;
1827 tp
->pushed_seq
= tp
->write_seq
;
1830 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
1831 tp
->retrans_stamp
= TCP_SKB_CB(buff
)->when
;
1832 skb_header_release(buff
);
1833 __skb_queue_tail(&sk
->sk_write_queue
, buff
);
1834 sk_charge_skb(sk
, buff
);
1835 tp
->packets_out
+= tcp_skb_pcount(buff
);
1836 tcp_transmit_skb(sk
, skb_clone(buff
, GFP_KERNEL
));
1837 TCP_INC_STATS(TCP_MIB_ACTIVEOPENS
);
1839 /* Timer for repeating the SYN until an answer. */
1840 tcp_reset_xmit_timer(sk
, TCP_TIME_RETRANS
, tp
->rto
);
1844 /* Send out a delayed ack, the caller does the policy checking
1845 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
1848 void tcp_send_delayed_ack(struct sock
*sk
)
1850 struct tcp_sock
*tp
= tcp_sk(sk
);
1851 int ato
= tp
->ack
.ato
;
1852 unsigned long timeout
;
1854 if (ato
> TCP_DELACK_MIN
) {
1857 if (tp
->ack
.pingpong
|| (tp
->ack
.pending
&TCP_ACK_PUSHED
))
1858 max_ato
= TCP_DELACK_MAX
;
1860 /* Slow path, intersegment interval is "high". */
1862 /* If some rtt estimate is known, use it to bound delayed ack.
1863 * Do not use tp->rto here, use results of rtt measurements
1867 int rtt
= max(tp
->srtt
>>3, TCP_DELACK_MIN
);
1873 ato
= min(ato
, max_ato
);
1876 /* Stay within the limit we were given */
1877 timeout
= jiffies
+ ato
;
1879 /* Use new timeout only if there wasn't a older one earlier. */
1880 if (tp
->ack
.pending
&TCP_ACK_TIMER
) {
1881 /* If delack timer was blocked or is about to expire,
1884 if (tp
->ack
.blocked
|| time_before_eq(tp
->ack
.timeout
, jiffies
+(ato
>>2))) {
1889 if (!time_before(timeout
, tp
->ack
.timeout
))
1890 timeout
= tp
->ack
.timeout
;
1892 tp
->ack
.pending
|= TCP_ACK_SCHED
|TCP_ACK_TIMER
;
1893 tp
->ack
.timeout
= timeout
;
1894 sk_reset_timer(sk
, &tp
->delack_timer
, timeout
);
1897 /* This routine sends an ack and also updates the window. */
1898 void tcp_send_ack(struct sock
*sk
)
1900 /* If we have been reset, we may not send again. */
1901 if (sk
->sk_state
!= TCP_CLOSE
) {
1902 struct tcp_sock
*tp
= tcp_sk(sk
);
1903 struct sk_buff
*buff
;
1905 /* We are not putting this on the write queue, so
1906 * tcp_transmit_skb() will set the ownership to this
1909 buff
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
1911 tcp_schedule_ack(tp
);
1912 tp
->ack
.ato
= TCP_ATO_MIN
;
1913 tcp_reset_xmit_timer(sk
, TCP_TIME_DACK
, TCP_DELACK_MAX
);
1917 /* Reserve space for headers and prepare control bits. */
1918 skb_reserve(buff
, MAX_TCP_HEADER
);
1920 TCP_SKB_CB(buff
)->flags
= TCPCB_FLAG_ACK
;
1921 TCP_SKB_CB(buff
)->sacked
= 0;
1922 skb_shinfo(buff
)->tso_segs
= 1;
1923 skb_shinfo(buff
)->tso_size
= 0;
1925 /* Send it off, this clears delayed acks for us. */
1926 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(buff
)->end_seq
= tcp_acceptable_seq(sk
, tp
);
1927 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
1928 tcp_transmit_skb(sk
, buff
);
1932 /* This routine sends a packet with an out of date sequence
1933 * number. It assumes the other end will try to ack it.
1935 * Question: what should we make while urgent mode?
1936 * 4.4BSD forces sending single byte of data. We cannot send
1937 * out of window data, because we have SND.NXT==SND.MAX...
1939 * Current solution: to send TWO zero-length segments in urgent mode:
1940 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
1941 * out-of-date with SND.UNA-1 to probe window.
1943 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
)
1945 struct tcp_sock
*tp
= tcp_sk(sk
);
1946 struct sk_buff
*skb
;
1948 /* We don't queue it, tcp_transmit_skb() sets ownership. */
1949 skb
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
1953 /* Reserve space for headers and set control bits. */
1954 skb_reserve(skb
, MAX_TCP_HEADER
);
1956 TCP_SKB_CB(skb
)->flags
= TCPCB_FLAG_ACK
;
1957 TCP_SKB_CB(skb
)->sacked
= urgent
;
1958 skb_shinfo(skb
)->tso_segs
= 1;
1959 skb_shinfo(skb
)->tso_size
= 0;
1961 /* Use a previous sequence. This should cause the other
1962 * end to send an ack. Don't queue or clone SKB, just
1965 TCP_SKB_CB(skb
)->seq
= urgent
? tp
->snd_una
: tp
->snd_una
- 1;
1966 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
;
1967 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1968 return tcp_transmit_skb(sk
, skb
);
1971 int tcp_write_wakeup(struct sock
*sk
)
1973 if (sk
->sk_state
!= TCP_CLOSE
) {
1974 struct tcp_sock
*tp
= tcp_sk(sk
);
1975 struct sk_buff
*skb
;
1977 if ((skb
= sk
->sk_send_head
) != NULL
&&
1978 before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
+tp
->snd_wnd
)) {
1980 unsigned int mss
= tcp_current_mss(sk
, 0);
1981 unsigned int seg_size
= tp
->snd_una
+tp
->snd_wnd
-TCP_SKB_CB(skb
)->seq
;
1983 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
1984 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
1986 /* We are probing the opening of a window
1987 * but the window size is != 0
1988 * must have been a result SWS avoidance ( sender )
1990 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
1992 seg_size
= min(seg_size
, mss
);
1993 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
1994 if (tcp_fragment(sk
, skb
, seg_size
))
1996 /* SWS override triggered forced fragmentation.
1997 * Disable TSO, the connection is too sick. */
1998 if (sk
->sk_route_caps
& NETIF_F_TSO
) {
1999 sock_set_flag(sk
, SOCK_NO_LARGESEND
);
2000 sk
->sk_route_caps
&= ~NETIF_F_TSO
;
2002 } else if (!tcp_skb_pcount(skb
))
2003 tcp_set_skb_tso_segs(sk
, skb
);
2005 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2006 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2007 err
= tcp_transmit_skb(sk
, skb_clone(skb
, GFP_ATOMIC
));
2009 update_send_head(sk
, tp
, skb
);
2014 between(tp
->snd_up
, tp
->snd_una
+1, tp
->snd_una
+0xFFFF))
2015 tcp_xmit_probe_skb(sk
, TCPCB_URG
);
2016 return tcp_xmit_probe_skb(sk
, 0);
2022 /* A window probe timeout has occurred. If window is not closed send
2023 * a partial packet else a zero probe.
2025 void tcp_send_probe0(struct sock
*sk
)
2027 struct tcp_sock
*tp
= tcp_sk(sk
);
2030 err
= tcp_write_wakeup(sk
);
2032 if (tp
->packets_out
|| !sk
->sk_send_head
) {
2033 /* Cancel probe timer, if it is not required. */
2040 if (tp
->backoff
< sysctl_tcp_retries2
)
2043 tcp_reset_xmit_timer (sk
, TCP_TIME_PROBE0
,
2044 min(tp
->rto
<< tp
->backoff
, TCP_RTO_MAX
));
2046 /* If packet was not sent due to local congestion,
2047 * do not backoff and do not remember probes_out.
2048 * Let local senders to fight for local resources.
2050 * Use accumulated backoff yet.
2052 if (!tp
->probes_out
)
2054 tcp_reset_xmit_timer (sk
, TCP_TIME_PROBE0
,
2055 min(tp
->rto
<< tp
->backoff
, TCP_RESOURCE_PROBE_INTERVAL
));
2059 EXPORT_SYMBOL(tcp_connect
);
2060 EXPORT_SYMBOL(tcp_make_synack
);
2061 EXPORT_SYMBOL(tcp_simple_retransmit
);
2062 EXPORT_SYMBOL(tcp_sync_mss
);