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 sock
*sk
, struct dst_entry
*dst
)
110 struct tcp_sock
*tp
= tcp_sk(sk
);
111 s32 delta
= tcp_time_stamp
- tp
->lsndtime
;
112 u32 restart_cwnd
= tcp_init_cwnd(tp
, dst
);
113 u32 cwnd
= tp
->snd_cwnd
;
115 tcp_ca_event(sk
, CA_EVENT_CWND_RESTART
);
117 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
118 restart_cwnd
= min(restart_cwnd
, cwnd
);
120 while ((delta
-= inet_csk(sk
)->icsk_rto
) > 0 && cwnd
> restart_cwnd
)
122 tp
->snd_cwnd
= max(cwnd
, restart_cwnd
);
123 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
124 tp
->snd_cwnd_used
= 0;
127 static inline void tcp_event_data_sent(struct tcp_sock
*tp
,
128 struct sk_buff
*skb
, struct sock
*sk
)
130 struct inet_connection_sock
*icsk
= inet_csk(sk
);
131 const u32 now
= tcp_time_stamp
;
133 if (!tp
->packets_out
&& (s32
)(now
- tp
->lsndtime
) > icsk
->icsk_rto
)
134 tcp_cwnd_restart(sk
, __sk_dst_get(sk
));
138 /* If it is a reply for ato after last received
139 * packet, enter pingpong mode.
141 if ((u32
)(now
- icsk
->icsk_ack
.lrcvtime
) < icsk
->icsk_ack
.ato
)
142 icsk
->icsk_ack
.pingpong
= 1;
145 static __inline__
void tcp_event_ack_sent(struct sock
*sk
, unsigned int pkts
)
147 tcp_dec_quickack_mode(sk
, pkts
);
148 inet_csk_clear_xmit_timer(sk
, ICSK_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 RFC2414. 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
, int clone_it
, gfp_t gfp_mask
)
267 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
268 struct inet_sock
*inet
;
270 struct tcp_skb_cb
*tcb
;
276 BUG_ON(!skb
|| !tcp_skb_pcount(skb
));
278 /* If congestion control is doing timestamping, we must
279 * take such a timestamp before we potentially clone/copy.
281 if (icsk
->icsk_ca_ops
->rtt_sample
)
282 __net_timestamp(skb
);
284 if (likely(clone_it
)) {
285 if (unlikely(skb_cloned(skb
)))
286 skb
= pskb_copy(skb
, gfp_mask
);
288 skb
= skb_clone(skb
, gfp_mask
);
295 tcb
= TCP_SKB_CB(skb
);
296 tcp_header_size
= tp
->tcp_header_len
;
298 #define SYSCTL_FLAG_TSTAMPS 0x1
299 #define SYSCTL_FLAG_WSCALE 0x2
300 #define SYSCTL_FLAG_SACK 0x4
303 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
304 tcp_header_size
= sizeof(struct tcphdr
) + TCPOLEN_MSS
;
305 if(sysctl_tcp_timestamps
) {
306 tcp_header_size
+= TCPOLEN_TSTAMP_ALIGNED
;
307 sysctl_flags
|= SYSCTL_FLAG_TSTAMPS
;
309 if (sysctl_tcp_window_scaling
) {
310 tcp_header_size
+= TCPOLEN_WSCALE_ALIGNED
;
311 sysctl_flags
|= SYSCTL_FLAG_WSCALE
;
313 if (sysctl_tcp_sack
) {
314 sysctl_flags
|= SYSCTL_FLAG_SACK
;
315 if (!(sysctl_flags
& SYSCTL_FLAG_TSTAMPS
))
316 tcp_header_size
+= TCPOLEN_SACKPERM_ALIGNED
;
318 } else if (unlikely(tp
->rx_opt
.eff_sacks
)) {
319 /* A SACK is 2 pad bytes, a 2 byte header, plus
320 * 2 32-bit sequence numbers for each SACK block.
322 tcp_header_size
+= (TCPOLEN_SACK_BASE_ALIGNED
+
323 (tp
->rx_opt
.eff_sacks
*
324 TCPOLEN_SACK_PERBLOCK
));
327 if (tcp_packets_in_flight(tp
) == 0)
328 tcp_ca_event(sk
, CA_EVENT_TX_START
);
330 th
= (struct tcphdr
*) skb_push(skb
, tcp_header_size
);
332 skb_set_owner_w(skb
, sk
);
334 /* Build TCP header and checksum it. */
335 th
->source
= inet
->sport
;
336 th
->dest
= inet
->dport
;
337 th
->seq
= htonl(tcb
->seq
);
338 th
->ack_seq
= htonl(tp
->rcv_nxt
);
339 *(((__u16
*)th
) + 6) = htons(((tcp_header_size
>> 2) << 12) |
342 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
343 /* RFC1323: The window in SYN & SYN/ACK segments
346 th
->window
= htons(tp
->rcv_wnd
);
348 th
->window
= htons(tcp_select_window(sk
));
353 if (unlikely(tp
->urg_mode
&&
354 between(tp
->snd_up
, tcb
->seq
+1, tcb
->seq
+0xFFFF))) {
355 th
->urg_ptr
= htons(tp
->snd_up
-tcb
->seq
);
359 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
360 tcp_syn_build_options((__u32
*)(th
+ 1),
361 tcp_advertise_mss(sk
),
362 (sysctl_flags
& SYSCTL_FLAG_TSTAMPS
),
363 (sysctl_flags
& SYSCTL_FLAG_SACK
),
364 (sysctl_flags
& SYSCTL_FLAG_WSCALE
),
365 tp
->rx_opt
.rcv_wscale
,
367 tp
->rx_opt
.ts_recent
);
369 tcp_build_and_update_options((__u32
*)(th
+ 1),
371 TCP_ECN_send(sk
, tp
, skb
, tcp_header_size
);
374 icsk
->icsk_af_ops
->send_check(sk
, skb
->len
, skb
);
376 if (likely(tcb
->flags
& TCPCB_FLAG_ACK
))
377 tcp_event_ack_sent(sk
, tcp_skb_pcount(skb
));
379 if (skb
->len
!= tcp_header_size
)
380 tcp_event_data_sent(tp
, skb
, sk
);
382 TCP_INC_STATS(TCP_MIB_OUTSEGS
);
384 err
= icsk
->icsk_af_ops
->queue_xmit(skb
, 0);
385 if (unlikely(err
<= 0))
390 /* NET_XMIT_CN is special. It does not guarantee,
391 * that this packet is lost. It tells that device
392 * is about to start to drop packets or already
393 * drops some packets of the same priority and
394 * invokes us to send less aggressively.
396 return err
== NET_XMIT_CN
? 0 : err
;
398 #undef SYSCTL_FLAG_TSTAMPS
399 #undef SYSCTL_FLAG_WSCALE
400 #undef SYSCTL_FLAG_SACK
404 /* This routine just queue's the buffer
406 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
407 * otherwise socket can stall.
409 static void tcp_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
411 struct tcp_sock
*tp
= tcp_sk(sk
);
413 /* Advance write_seq and place onto the write_queue. */
414 tp
->write_seq
= TCP_SKB_CB(skb
)->end_seq
;
415 skb_header_release(skb
);
416 __skb_queue_tail(&sk
->sk_write_queue
, skb
);
417 sk_charge_skb(sk
, skb
);
419 /* Queue it, remembering where we must start sending. */
420 if (sk
->sk_send_head
== NULL
)
421 sk
->sk_send_head
= skb
;
424 static void tcp_set_skb_tso_segs(struct sock
*sk
, struct sk_buff
*skb
, unsigned int mss_now
)
426 if (skb
->len
<= mss_now
||
427 !(sk
->sk_route_caps
& NETIF_F_TSO
)) {
428 /* Avoid the costly divide in the normal
431 skb_shinfo(skb
)->tso_segs
= 1;
432 skb_shinfo(skb
)->tso_size
= 0;
436 factor
= skb
->len
+ (mss_now
- 1);
438 skb_shinfo(skb
)->tso_segs
= factor
;
439 skb_shinfo(skb
)->tso_size
= mss_now
;
443 /* Function to create two new TCP segments. Shrinks the given segment
444 * to the specified size and appends a new segment with the rest of the
445 * packet to the list. This won't be called frequently, I hope.
446 * Remember, these are still headerless SKBs at this point.
448 int tcp_fragment(struct sock
*sk
, struct sk_buff
*skb
, u32 len
, unsigned int mss_now
)
450 struct tcp_sock
*tp
= tcp_sk(sk
);
451 struct sk_buff
*buff
;
452 int nsize
, old_factor
;
455 BUG_ON(len
> skb
->len
);
457 clear_all_retrans_hints(tp
);
458 nsize
= skb_headlen(skb
) - len
;
462 if (skb_cloned(skb
) &&
463 skb_is_nonlinear(skb
) &&
464 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
467 /* Get a new skb... force flag on. */
468 buff
= sk_stream_alloc_skb(sk
, nsize
, GFP_ATOMIC
);
470 return -ENOMEM
; /* We'll just try again later. */
471 sk_charge_skb(sk
, buff
);
473 /* Correct the sequence numbers. */
474 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
475 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
476 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
478 /* PSH and FIN should only be set in the second packet. */
479 flags
= TCP_SKB_CB(skb
)->flags
;
480 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
481 TCP_SKB_CB(buff
)->flags
= flags
;
482 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
483 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_AT_TAIL
;
485 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_HW
) {
486 /* Copy and checksum data tail into the new buffer. */
487 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
, skb_put(buff
, nsize
),
492 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
494 skb
->ip_summed
= CHECKSUM_HW
;
495 skb_split(skb
, buff
, len
);
498 buff
->ip_summed
= skb
->ip_summed
;
500 /* Looks stupid, but our code really uses when of
501 * skbs, which it never sent before. --ANK
503 TCP_SKB_CB(buff
)->when
= TCP_SKB_CB(skb
)->when
;
504 buff
->tstamp
= skb
->tstamp
;
506 old_factor
= tcp_skb_pcount(skb
);
508 /* Fix up tso_factor for both original and new SKB. */
509 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
510 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
512 /* If this packet has been sent out already, we must
513 * adjust the various packet counters.
515 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
516 int diff
= old_factor
- tcp_skb_pcount(skb
) -
517 tcp_skb_pcount(buff
);
519 tp
->packets_out
-= diff
;
521 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
522 tp
->sacked_out
-= diff
;
523 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
524 tp
->retrans_out
-= diff
;
526 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
) {
527 tp
->lost_out
-= diff
;
528 tp
->left_out
-= diff
;
532 /* Adjust Reno SACK estimate. */
533 if (!tp
->rx_opt
.sack_ok
) {
534 tp
->sacked_out
-= diff
;
535 if ((int)tp
->sacked_out
< 0)
537 tcp_sync_left_out(tp
);
540 tp
->fackets_out
-= diff
;
541 if ((int)tp
->fackets_out
< 0)
546 /* Link BUFF into the send queue. */
547 skb_header_release(buff
);
548 __skb_append(skb
, buff
, &sk
->sk_write_queue
);
553 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
554 * eventually). The difference is that pulled data not copied, but
555 * immediately discarded.
557 static unsigned char *__pskb_trim_head(struct sk_buff
*skb
, int len
)
563 for (i
=0; i
<skb_shinfo(skb
)->nr_frags
; i
++) {
564 if (skb_shinfo(skb
)->frags
[i
].size
<= eat
) {
565 put_page(skb_shinfo(skb
)->frags
[i
].page
);
566 eat
-= skb_shinfo(skb
)->frags
[i
].size
;
568 skb_shinfo(skb
)->frags
[k
] = skb_shinfo(skb
)->frags
[i
];
570 skb_shinfo(skb
)->frags
[k
].page_offset
+= eat
;
571 skb_shinfo(skb
)->frags
[k
].size
-= eat
;
577 skb_shinfo(skb
)->nr_frags
= k
;
579 skb
->tail
= skb
->data
;
580 skb
->data_len
-= len
;
581 skb
->len
= skb
->data_len
;
585 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
587 if (skb_cloned(skb
) &&
588 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
591 if (len
<= skb_headlen(skb
)) {
592 __skb_pull(skb
, len
);
594 if (__pskb_trim_head(skb
, len
-skb_headlen(skb
)) == NULL
)
598 TCP_SKB_CB(skb
)->seq
+= len
;
599 skb
->ip_summed
= CHECKSUM_HW
;
601 skb
->truesize
-= len
;
602 sk
->sk_wmem_queued
-= len
;
603 sk
->sk_forward_alloc
+= len
;
604 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
606 /* Any change of skb->len requires recalculation of tso
609 if (tcp_skb_pcount(skb
) > 1)
610 tcp_set_skb_tso_segs(sk
, skb
, tcp_current_mss(sk
, 1));
615 /* This function synchronize snd mss to current pmtu/exthdr set.
617 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
618 for TCP options, but includes only bare TCP header.
620 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
621 It is minimum of user_mss and mss received with SYN.
622 It also does not include TCP options.
624 tp->pmtu_cookie is last pmtu, seen by this function.
626 tp->mss_cache is current effective sending mss, including
627 all tcp options except for SACKs. It is evaluated,
628 taking into account current pmtu, but never exceeds
629 tp->rx_opt.mss_clamp.
631 NOTE1. rfc1122 clearly states that advertised MSS
632 DOES NOT include either tcp or ip options.
634 NOTE2. tp->pmtu_cookie and tp->mss_cache are READ ONLY outside
635 this function. --ANK (980731)
638 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
640 struct tcp_sock
*tp
= tcp_sk(sk
);
641 /* Calculate base mss without TCP options:
642 It is MMS_S - sizeof(tcphdr) of rfc1122
644 int mss_now
= (pmtu
- inet_csk(sk
)->icsk_af_ops
->net_header_len
-
645 sizeof(struct tcphdr
));
647 /* Clamp it (mss_clamp does not include tcp options) */
648 if (mss_now
> tp
->rx_opt
.mss_clamp
)
649 mss_now
= tp
->rx_opt
.mss_clamp
;
651 /* Now subtract optional transport overhead */
652 mss_now
-= tp
->ext_header_len
;
654 /* Then reserve room for full set of TCP options and 8 bytes of data */
658 /* Now subtract TCP options size, not including SACKs */
659 mss_now
-= tp
->tcp_header_len
- sizeof(struct tcphdr
);
661 /* Bound mss with half of window */
662 if (tp
->max_window
&& mss_now
> (tp
->max_window
>>1))
663 mss_now
= max((tp
->max_window
>>1), 68U - tp
->tcp_header_len
);
665 /* And store cached results */
666 tp
->pmtu_cookie
= pmtu
;
667 tp
->mss_cache
= mss_now
;
672 /* Compute the current effective MSS, taking SACKs and IP options,
673 * and even PMTU discovery events into account.
675 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
676 * cannot be large. However, taking into account rare use of URG, this
679 unsigned int tcp_current_mss(struct sock
*sk
, int large_allowed
)
681 struct tcp_sock
*tp
= tcp_sk(sk
);
682 struct dst_entry
*dst
= __sk_dst_get(sk
);
687 mss_now
= tp
->mss_cache
;
690 (sk
->sk_route_caps
& NETIF_F_TSO
) &&
695 u32 mtu
= dst_mtu(dst
);
696 if (mtu
!= tp
->pmtu_cookie
)
697 mss_now
= tcp_sync_mss(sk
, mtu
);
700 if (tp
->rx_opt
.eff_sacks
)
701 mss_now
-= (TCPOLEN_SACK_BASE_ALIGNED
+
702 (tp
->rx_opt
.eff_sacks
* TCPOLEN_SACK_PERBLOCK
));
704 xmit_size_goal
= mss_now
;
707 xmit_size_goal
= (65535 -
708 inet_csk(sk
)->icsk_af_ops
->net_header_len
-
709 tp
->ext_header_len
- tp
->tcp_header_len
);
711 if (tp
->max_window
&&
712 (xmit_size_goal
> (tp
->max_window
>> 1)))
713 xmit_size_goal
= max((tp
->max_window
>> 1),
714 68U - tp
->tcp_header_len
);
716 xmit_size_goal
-= (xmit_size_goal
% mss_now
);
718 tp
->xmit_size_goal
= xmit_size_goal
;
723 /* Congestion window validation. (RFC2861) */
725 static inline void tcp_cwnd_validate(struct sock
*sk
, struct tcp_sock
*tp
)
727 __u32 packets_out
= tp
->packets_out
;
729 if (packets_out
>= tp
->snd_cwnd
) {
730 /* Network is feed fully. */
731 tp
->snd_cwnd_used
= 0;
732 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
734 /* Network starves. */
735 if (tp
->packets_out
> tp
->snd_cwnd_used
)
736 tp
->snd_cwnd_used
= tp
->packets_out
;
738 if ((s32
)(tcp_time_stamp
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
)
739 tcp_cwnd_application_limited(sk
);
743 static unsigned int tcp_window_allows(struct tcp_sock
*tp
, struct sk_buff
*skb
, unsigned int mss_now
, unsigned int cwnd
)
745 u32 window
, cwnd_len
;
747 window
= (tp
->snd_una
+ tp
->snd_wnd
- TCP_SKB_CB(skb
)->seq
);
748 cwnd_len
= mss_now
* cwnd
;
749 return min(window
, cwnd_len
);
752 /* Can at least one segment of SKB be sent right now, according to the
753 * congestion window rules? If so, return how many segments are allowed.
755 static inline unsigned int tcp_cwnd_test(struct tcp_sock
*tp
, struct sk_buff
*skb
)
759 /* Don't be strict about the congestion window for the final FIN. */
760 if (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
)
763 in_flight
= tcp_packets_in_flight(tp
);
765 if (in_flight
< cwnd
)
766 return (cwnd
- in_flight
);
771 /* This must be invoked the first time we consider transmitting
774 static inline int tcp_init_tso_segs(struct sock
*sk
, struct sk_buff
*skb
, unsigned int mss_now
)
776 int tso_segs
= tcp_skb_pcount(skb
);
780 skb_shinfo(skb
)->tso_size
!= mss_now
)) {
781 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
782 tso_segs
= tcp_skb_pcount(skb
);
787 static inline int tcp_minshall_check(const struct tcp_sock
*tp
)
789 return after(tp
->snd_sml
,tp
->snd_una
) &&
790 !after(tp
->snd_sml
, tp
->snd_nxt
);
793 /* Return 0, if packet can be sent now without violation Nagle's rules:
794 * 1. It is full sized.
795 * 2. Or it contains FIN. (already checked by caller)
796 * 3. Or TCP_NODELAY was set.
797 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
798 * With Minshall's modification: all sent small packets are ACKed.
801 static inline int tcp_nagle_check(const struct tcp_sock
*tp
,
802 const struct sk_buff
*skb
,
803 unsigned mss_now
, int nonagle
)
805 return (skb
->len
< mss_now
&&
806 ((nonagle
&TCP_NAGLE_CORK
) ||
809 tcp_minshall_check(tp
))));
812 /* Return non-zero if the Nagle test allows this packet to be
815 static inline int tcp_nagle_test(struct tcp_sock
*tp
, struct sk_buff
*skb
,
816 unsigned int cur_mss
, int nonagle
)
818 /* Nagle rule does not apply to frames, which sit in the middle of the
819 * write_queue (they have no chances to get new data).
821 * This is implemented in the callers, where they modify the 'nonagle'
822 * argument based upon the location of SKB in the send queue.
824 if (nonagle
& TCP_NAGLE_PUSH
)
827 /* Don't use the nagle rule for urgent data (or for the final FIN). */
829 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
))
832 if (!tcp_nagle_check(tp
, skb
, cur_mss
, nonagle
))
838 /* Does at least the first segment of SKB fit into the send window? */
839 static inline int tcp_snd_wnd_test(struct tcp_sock
*tp
, struct sk_buff
*skb
, unsigned int cur_mss
)
841 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
843 if (skb
->len
> cur_mss
)
844 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
846 return !after(end_seq
, tp
->snd_una
+ tp
->snd_wnd
);
849 /* This checks if the data bearing packet SKB (usually sk->sk_send_head)
850 * should be put on the wire right now. If so, it returns the number of
851 * packets allowed by the congestion window.
853 static unsigned int tcp_snd_test(struct sock
*sk
, struct sk_buff
*skb
,
854 unsigned int cur_mss
, int nonagle
)
856 struct tcp_sock
*tp
= tcp_sk(sk
);
857 unsigned int cwnd_quota
;
859 tcp_init_tso_segs(sk
, skb
, cur_mss
);
861 if (!tcp_nagle_test(tp
, skb
, cur_mss
, nonagle
))
864 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
866 !tcp_snd_wnd_test(tp
, skb
, cur_mss
))
872 static inline int tcp_skb_is_last(const struct sock
*sk
,
873 const struct sk_buff
*skb
)
875 return skb
->next
== (struct sk_buff
*)&sk
->sk_write_queue
;
878 int tcp_may_send_now(struct sock
*sk
, struct tcp_sock
*tp
)
880 struct sk_buff
*skb
= sk
->sk_send_head
;
883 tcp_snd_test(sk
, skb
, tcp_current_mss(sk
, 1),
884 (tcp_skb_is_last(sk
, skb
) ?
889 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
890 * which is put after SKB on the list. It is very much like
891 * tcp_fragment() except that it may make several kinds of assumptions
892 * in order to speed up the splitting operation. In particular, we
893 * know that all the data is in scatter-gather pages, and that the
894 * packet has never been sent out before (and thus is not cloned).
896 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
, unsigned int mss_now
)
898 struct sk_buff
*buff
;
899 int nlen
= skb
->len
- len
;
902 /* All of a TSO frame must be composed of paged data. */
903 if (skb
->len
!= skb
->data_len
)
904 return tcp_fragment(sk
, skb
, len
, mss_now
);
906 buff
= sk_stream_alloc_pskb(sk
, 0, 0, GFP_ATOMIC
);
907 if (unlikely(buff
== NULL
))
910 buff
->truesize
= nlen
;
911 skb
->truesize
-= nlen
;
913 /* Correct the sequence numbers. */
914 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
915 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
916 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
918 /* PSH and FIN should only be set in the second packet. */
919 flags
= TCP_SKB_CB(skb
)->flags
;
920 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
921 TCP_SKB_CB(buff
)->flags
= flags
;
923 /* This packet was never sent out yet, so no SACK bits. */
924 TCP_SKB_CB(buff
)->sacked
= 0;
926 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_HW
;
927 skb_split(skb
, buff
, len
);
929 /* Fix up tso_factor for both original and new SKB. */
930 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
931 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
933 /* Link BUFF into the send queue. */
934 skb_header_release(buff
);
935 __skb_append(skb
, buff
, &sk
->sk_write_queue
);
940 /* Try to defer sending, if possible, in order to minimize the amount
941 * of TSO splitting we do. View it as a kind of TSO Nagle test.
943 * This algorithm is from John Heffner.
945 static int tcp_tso_should_defer(struct sock
*sk
, struct tcp_sock
*tp
, struct sk_buff
*skb
)
947 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
948 u32 send_win
, cong_win
, limit
, in_flight
;
950 if (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
)
953 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
956 in_flight
= tcp_packets_in_flight(tp
);
958 BUG_ON(tcp_skb_pcount(skb
) <= 1 ||
959 (tp
->snd_cwnd
<= in_flight
));
961 send_win
= (tp
->snd_una
+ tp
->snd_wnd
) - TCP_SKB_CB(skb
)->seq
;
963 /* From in_flight test above, we know that cwnd > in_flight. */
964 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
966 limit
= min(send_win
, cong_win
);
968 if (sysctl_tcp_tso_win_divisor
) {
969 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
971 /* If at least some fraction of a window is available,
974 chunk
/= sysctl_tcp_tso_win_divisor
;
978 /* Different approach, try not to defer past a single
979 * ACK. Receiver should ACK every other full sized
980 * frame, so if we have space for more than 3 frames
983 if (limit
> tcp_max_burst(tp
) * tp
->mss_cache
)
987 /* Ok, it looks like it is advisable to defer. */
991 /* This routine writes packets to the network. It advances the
992 * send_head. This happens as incoming acks open up the remote
995 * Returns 1, if no segments are in flight and we have queued segments, but
996 * cannot send anything now because of SWS or another problem.
998 static int tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
)
1000 struct tcp_sock
*tp
= tcp_sk(sk
);
1001 struct sk_buff
*skb
;
1002 unsigned int tso_segs
, sent_pkts
;
1005 /* If we are closed, the bytes will have to remain here.
1006 * In time closedown will finish, we empty the write queue and all
1009 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
1013 while ((skb
= sk
->sk_send_head
)) {
1016 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1019 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1023 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
)))
1026 if (tso_segs
== 1) {
1027 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
1028 (tcp_skb_is_last(sk
, skb
) ?
1029 nonagle
: TCP_NAGLE_PUSH
))))
1032 if (tcp_tso_should_defer(sk
, tp
, skb
))
1038 limit
= tcp_window_allows(tp
, skb
,
1039 mss_now
, cwnd_quota
);
1041 if (skb
->len
< limit
) {
1042 unsigned int trim
= skb
->len
% mss_now
;
1045 limit
= skb
->len
- trim
;
1049 if (skb
->len
> limit
&&
1050 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
)))
1053 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1055 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
)))
1058 /* Advance the send_head. This one is sent out.
1059 * This call will increment packets_out.
1061 update_send_head(sk
, tp
, skb
);
1063 tcp_minshall_update(tp
, mss_now
, skb
);
1067 if (likely(sent_pkts
)) {
1068 tcp_cwnd_validate(sk
, tp
);
1071 return !tp
->packets_out
&& sk
->sk_send_head
;
1074 /* Push out any pending frames which were held back due to
1075 * TCP_CORK or attempt at coalescing tiny packets.
1076 * The socket must be locked by the caller.
1078 void __tcp_push_pending_frames(struct sock
*sk
, struct tcp_sock
*tp
,
1079 unsigned int cur_mss
, int nonagle
)
1081 struct sk_buff
*skb
= sk
->sk_send_head
;
1084 if (tcp_write_xmit(sk
, cur_mss
, nonagle
))
1085 tcp_check_probe_timer(sk
, tp
);
1089 /* Send _single_ skb sitting at the send head. This function requires
1090 * true push pending frames to setup probe timer etc.
1092 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
1094 struct tcp_sock
*tp
= tcp_sk(sk
);
1095 struct sk_buff
*skb
= sk
->sk_send_head
;
1096 unsigned int tso_segs
, cwnd_quota
;
1098 BUG_ON(!skb
|| skb
->len
< mss_now
);
1100 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1101 cwnd_quota
= tcp_snd_test(sk
, skb
, mss_now
, TCP_NAGLE_PUSH
);
1103 if (likely(cwnd_quota
)) {
1110 limit
= tcp_window_allows(tp
, skb
,
1111 mss_now
, cwnd_quota
);
1113 if (skb
->len
< limit
) {
1114 unsigned int trim
= skb
->len
% mss_now
;
1117 limit
= skb
->len
- trim
;
1121 if (skb
->len
> limit
&&
1122 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
)))
1125 /* Send it out now. */
1126 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1128 if (likely(!tcp_transmit_skb(sk
, skb
, 1, sk
->sk_allocation
))) {
1129 update_send_head(sk
, tp
, skb
);
1130 tcp_cwnd_validate(sk
, tp
);
1136 /* This function returns the amount that we can raise the
1137 * usable window based on the following constraints
1139 * 1. The window can never be shrunk once it is offered (RFC 793)
1140 * 2. We limit memory per socket
1143 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1144 * RECV.NEXT + RCV.WIN fixed until:
1145 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1147 * i.e. don't raise the right edge of the window until you can raise
1148 * it at least MSS bytes.
1150 * Unfortunately, the recommended algorithm breaks header prediction,
1151 * since header prediction assumes th->window stays fixed.
1153 * Strictly speaking, keeping th->window fixed violates the receiver
1154 * side SWS prevention criteria. The problem is that under this rule
1155 * a stream of single byte packets will cause the right side of the
1156 * window to always advance by a single byte.
1158 * Of course, if the sender implements sender side SWS prevention
1159 * then this will not be a problem.
1161 * BSD seems to make the following compromise:
1163 * If the free space is less than the 1/4 of the maximum
1164 * space available and the free space is less than 1/2 mss,
1165 * then set the window to 0.
1166 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1167 * Otherwise, just prevent the window from shrinking
1168 * and from being larger than the largest representable value.
1170 * This prevents incremental opening of the window in the regime
1171 * where TCP is limited by the speed of the reader side taking
1172 * data out of the TCP receive queue. It does nothing about
1173 * those cases where the window is constrained on the sender side
1174 * because the pipeline is full.
1176 * BSD also seems to "accidentally" limit itself to windows that are a
1177 * multiple of MSS, at least until the free space gets quite small.
1178 * This would appear to be a side effect of the mbuf implementation.
1179 * Combining these two algorithms results in the observed behavior
1180 * of having a fixed window size at almost all times.
1182 * Below we obtain similar behavior by forcing the offered window to
1183 * a multiple of the mss when it is feasible to do so.
1185 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1186 * Regular options like TIMESTAMP are taken into account.
1188 u32
__tcp_select_window(struct sock
*sk
)
1190 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1191 struct tcp_sock
*tp
= tcp_sk(sk
);
1192 /* MSS for the peer's data. Previous versions used mss_clamp
1193 * here. I don't know if the value based on our guesses
1194 * of peer's MSS is better for the performance. It's more correct
1195 * but may be worse for the performance because of rcv_mss
1196 * fluctuations. --SAW 1998/11/1
1198 int mss
= icsk
->icsk_ack
.rcv_mss
;
1199 int free_space
= tcp_space(sk
);
1200 int full_space
= min_t(int, tp
->window_clamp
, tcp_full_space(sk
));
1203 if (mss
> full_space
)
1206 if (free_space
< full_space
/2) {
1207 icsk
->icsk_ack
.quick
= 0;
1209 if (tcp_memory_pressure
)
1210 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U*tp
->advmss
);
1212 if (free_space
< mss
)
1216 if (free_space
> tp
->rcv_ssthresh
)
1217 free_space
= tp
->rcv_ssthresh
;
1219 /* Don't do rounding if we are using window scaling, since the
1220 * scaled window will not line up with the MSS boundary anyway.
1222 window
= tp
->rcv_wnd
;
1223 if (tp
->rx_opt
.rcv_wscale
) {
1224 window
= free_space
;
1226 /* Advertise enough space so that it won't get scaled away.
1227 * Import case: prevent zero window announcement if
1228 * 1<<rcv_wscale > mss.
1230 if (((window
>> tp
->rx_opt
.rcv_wscale
) << tp
->rx_opt
.rcv_wscale
) != window
)
1231 window
= (((window
>> tp
->rx_opt
.rcv_wscale
) + 1)
1232 << tp
->rx_opt
.rcv_wscale
);
1234 /* Get the largest window that is a nice multiple of mss.
1235 * Window clamp already applied above.
1236 * If our current window offering is within 1 mss of the
1237 * free space we just keep it. This prevents the divide
1238 * and multiply from happening most of the time.
1239 * We also don't do any window rounding when the free space
1242 if (window
<= free_space
- mss
|| window
> free_space
)
1243 window
= (free_space
/mss
)*mss
;
1249 /* Attempt to collapse two adjacent SKB's during retransmission. */
1250 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*skb
, int mss_now
)
1252 struct tcp_sock
*tp
= tcp_sk(sk
);
1253 struct sk_buff
*next_skb
= skb
->next
;
1255 /* The first test we must make is that neither of these two
1256 * SKB's are still referenced by someone else.
1258 if (!skb_cloned(skb
) && !skb_cloned(next_skb
)) {
1259 int skb_size
= skb
->len
, next_skb_size
= next_skb
->len
;
1260 u16 flags
= TCP_SKB_CB(skb
)->flags
;
1262 /* Also punt if next skb has been SACK'd. */
1263 if(TCP_SKB_CB(next_skb
)->sacked
& TCPCB_SACKED_ACKED
)
1266 /* Next skb is out of window. */
1267 if (after(TCP_SKB_CB(next_skb
)->end_seq
, tp
->snd_una
+tp
->snd_wnd
))
1270 /* Punt if not enough space exists in the first SKB for
1271 * the data in the second, or the total combined payload
1272 * would exceed the MSS.
1274 if ((next_skb_size
> skb_tailroom(skb
)) ||
1275 ((skb_size
+ next_skb_size
) > mss_now
))
1278 BUG_ON(tcp_skb_pcount(skb
) != 1 ||
1279 tcp_skb_pcount(next_skb
) != 1);
1281 /* changing transmit queue under us so clear hints */
1282 clear_all_retrans_hints(tp
);
1284 /* Ok. We will be able to collapse the packet. */
1285 __skb_unlink(next_skb
, &sk
->sk_write_queue
);
1287 memcpy(skb_put(skb
, next_skb_size
), next_skb
->data
, next_skb_size
);
1289 if (next_skb
->ip_summed
== CHECKSUM_HW
)
1290 skb
->ip_summed
= CHECKSUM_HW
;
1292 if (skb
->ip_summed
!= CHECKSUM_HW
)
1293 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
1295 /* Update sequence range on original skb. */
1296 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
1298 /* Merge over control information. */
1299 flags
|= TCP_SKB_CB(next_skb
)->flags
; /* This moves PSH/FIN etc. over */
1300 TCP_SKB_CB(skb
)->flags
= flags
;
1302 /* All done, get rid of second SKB and account for it so
1303 * packet counting does not break.
1305 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
&(TCPCB_EVER_RETRANS
|TCPCB_AT_TAIL
);
1306 if (TCP_SKB_CB(next_skb
)->sacked
&TCPCB_SACKED_RETRANS
)
1307 tp
->retrans_out
-= tcp_skb_pcount(next_skb
);
1308 if (TCP_SKB_CB(next_skb
)->sacked
&TCPCB_LOST
) {
1309 tp
->lost_out
-= tcp_skb_pcount(next_skb
);
1310 tp
->left_out
-= tcp_skb_pcount(next_skb
);
1312 /* Reno case is special. Sigh... */
1313 if (!tp
->rx_opt
.sack_ok
&& tp
->sacked_out
) {
1314 tcp_dec_pcount_approx(&tp
->sacked_out
, next_skb
);
1315 tp
->left_out
-= tcp_skb_pcount(next_skb
);
1318 /* Not quite right: it can be > snd.fack, but
1319 * it is better to underestimate fackets.
1321 tcp_dec_pcount_approx(&tp
->fackets_out
, next_skb
);
1322 tcp_packets_out_dec(tp
, next_skb
);
1323 sk_stream_free_skb(sk
, next_skb
);
1327 /* Do a simple retransmit without using the backoff mechanisms in
1328 * tcp_timer. This is used for path mtu discovery.
1329 * The socket is already locked here.
1331 void tcp_simple_retransmit(struct sock
*sk
)
1333 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1334 struct tcp_sock
*tp
= tcp_sk(sk
);
1335 struct sk_buff
*skb
;
1336 unsigned int mss
= tcp_current_mss(sk
, 0);
1339 sk_stream_for_retrans_queue(skb
, sk
) {
1340 if (skb
->len
> mss
&&
1341 !(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
)) {
1342 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) {
1343 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1344 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1346 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_LOST
)) {
1347 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1348 tp
->lost_out
+= tcp_skb_pcount(skb
);
1354 clear_all_retrans_hints(tp
);
1359 tcp_sync_left_out(tp
);
1361 /* Don't muck with the congestion window here.
1362 * Reason is that we do not increase amount of _data_
1363 * in network, but units changed and effective
1364 * cwnd/ssthresh really reduced now.
1366 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
1367 tp
->high_seq
= tp
->snd_nxt
;
1368 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
1369 tp
->prior_ssthresh
= 0;
1370 tp
->undo_marker
= 0;
1371 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1373 tcp_xmit_retransmit_queue(sk
);
1376 /* This retransmits one SKB. Policy decisions and retransmit queue
1377 * state updates are done by the caller. Returns non-zero if an
1378 * error occurred which prevented the send.
1380 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
1382 struct tcp_sock
*tp
= tcp_sk(sk
);
1383 unsigned int cur_mss
= tcp_current_mss(sk
, 0);
1386 /* Do not sent more than we queued. 1/4 is reserved for possible
1387 * copying overhead: fragmentation, tunneling, mangling etc.
1389 if (atomic_read(&sk
->sk_wmem_alloc
) >
1390 min(sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2), sk
->sk_sndbuf
))
1393 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
1394 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1396 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
1400 /* If receiver has shrunk his window, and skb is out of
1401 * new window, do not retransmit it. The exception is the
1402 * case, when window is shrunk to zero. In this case
1403 * our retransmit serves as a zero window probe.
1405 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
+tp
->snd_wnd
)
1406 && TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
1409 if (skb
->len
> cur_mss
) {
1410 if (tcp_fragment(sk
, skb
, cur_mss
, cur_mss
))
1411 return -ENOMEM
; /* We'll try again later. */
1414 /* Collapse two adjacent packets if worthwhile and we can. */
1415 if(!(TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_SYN
) &&
1416 (skb
->len
< (cur_mss
>> 1)) &&
1417 (skb
->next
!= sk
->sk_send_head
) &&
1418 (skb
->next
!= (struct sk_buff
*)&sk
->sk_write_queue
) &&
1419 (skb_shinfo(skb
)->nr_frags
== 0 && skb_shinfo(skb
->next
)->nr_frags
== 0) &&
1420 (tcp_skb_pcount(skb
) == 1 && tcp_skb_pcount(skb
->next
) == 1) &&
1421 (sysctl_tcp_retrans_collapse
!= 0))
1422 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
1424 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
1425 return -EHOSTUNREACH
; /* Routing failure or similar. */
1427 /* Some Solaris stacks overoptimize and ignore the FIN on a
1428 * retransmit when old data is attached. So strip it off
1429 * since it is cheap to do so and saves bytes on the network.
1432 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
) &&
1433 tp
->snd_una
== (TCP_SKB_CB(skb
)->end_seq
- 1)) {
1434 if (!pskb_trim(skb
, 0)) {
1435 TCP_SKB_CB(skb
)->seq
= TCP_SKB_CB(skb
)->end_seq
- 1;
1436 skb_shinfo(skb
)->tso_segs
= 1;
1437 skb_shinfo(skb
)->tso_size
= 0;
1438 skb
->ip_summed
= CHECKSUM_NONE
;
1443 /* Make a copy, if the first transmission SKB clone we made
1444 * is still in somebody's hands, else make a clone.
1446 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1448 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
1451 /* Update global TCP statistics. */
1452 TCP_INC_STATS(TCP_MIB_RETRANSSEGS
);
1454 tp
->total_retrans
++;
1456 #if FASTRETRANS_DEBUG > 0
1457 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) {
1458 if (net_ratelimit())
1459 printk(KERN_DEBUG
"retrans_out leaked.\n");
1462 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
1463 tp
->retrans_out
+= tcp_skb_pcount(skb
);
1465 /* Save stamp of the first retransmit. */
1466 if (!tp
->retrans_stamp
)
1467 tp
->retrans_stamp
= TCP_SKB_CB(skb
)->when
;
1471 /* snd_nxt is stored to detect loss of retransmitted segment,
1472 * see tcp_input.c tcp_sacktag_write_queue().
1474 TCP_SKB_CB(skb
)->ack_seq
= tp
->snd_nxt
;
1479 /* This gets called after a retransmit timeout, and the initially
1480 * retransmitted data is acknowledged. It tries to continue
1481 * resending the rest of the retransmit queue, until either
1482 * we've sent it all or the congestion window limit is reached.
1483 * If doing SACK, the first ACK which comes back for a timeout
1484 * based retransmit packet might feed us FACK information again.
1485 * If so, we use it to avoid unnecessarily retransmissions.
1487 void tcp_xmit_retransmit_queue(struct sock
*sk
)
1489 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1490 struct tcp_sock
*tp
= tcp_sk(sk
);
1491 struct sk_buff
*skb
;
1494 if (tp
->retransmit_skb_hint
) {
1495 skb
= tp
->retransmit_skb_hint
;
1496 packet_cnt
= tp
->retransmit_cnt_hint
;
1498 skb
= sk
->sk_write_queue
.next
;
1502 /* First pass: retransmit lost packets. */
1504 sk_stream_for_retrans_queue_from(skb
, sk
) {
1505 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
1507 /* we could do better than to assign each time */
1508 tp
->retransmit_skb_hint
= skb
;
1509 tp
->retransmit_cnt_hint
= packet_cnt
;
1511 /* Assume this retransmit will generate
1512 * only one packet for congestion window
1513 * calculation purposes. This works because
1514 * tcp_retransmit_skb() will chop up the
1515 * packet to be MSS sized and all the
1516 * packet counting works out.
1518 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
1521 if (sacked
& TCPCB_LOST
) {
1522 if (!(sacked
&(TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))) {
1523 if (tcp_retransmit_skb(sk
, skb
)) {
1524 tp
->retransmit_skb_hint
= NULL
;
1527 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
1528 NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS
);
1530 NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS
);
1533 skb_peek(&sk
->sk_write_queue
))
1534 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1535 inet_csk(sk
)->icsk_rto
,
1539 packet_cnt
+= tcp_skb_pcount(skb
);
1540 if (packet_cnt
>= tp
->lost_out
)
1546 /* OK, demanded retransmission is finished. */
1548 /* Forward retransmissions are possible only during Recovery. */
1549 if (icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
1552 /* No forward retransmissions in Reno are possible. */
1553 if (!tp
->rx_opt
.sack_ok
)
1556 /* Yeah, we have to make difficult choice between forward transmission
1557 * and retransmission... Both ways have their merits...
1559 * For now we do not retransmit anything, while we have some new
1563 if (tcp_may_send_now(sk
, tp
))
1566 if (tp
->forward_skb_hint
) {
1567 skb
= tp
->forward_skb_hint
;
1568 packet_cnt
= tp
->forward_cnt_hint
;
1570 skb
= sk
->sk_write_queue
.next
;
1574 sk_stream_for_retrans_queue_from(skb
, sk
) {
1575 tp
->forward_cnt_hint
= packet_cnt
;
1576 tp
->forward_skb_hint
= skb
;
1578 /* Similar to the retransmit loop above we
1579 * can pretend that the retransmitted SKB
1580 * we send out here will be composed of one
1581 * real MSS sized packet because tcp_retransmit_skb()
1582 * will fragment it if necessary.
1584 if (++packet_cnt
> tp
->fackets_out
)
1587 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
1590 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
)
1593 /* Ok, retransmit it. */
1594 if (tcp_retransmit_skb(sk
, skb
)) {
1595 tp
->forward_skb_hint
= NULL
;
1599 if (skb
== skb_peek(&sk
->sk_write_queue
))
1600 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1601 inet_csk(sk
)->icsk_rto
,
1604 NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS
);
1609 /* Send a fin. The caller locks the socket for us. This cannot be
1610 * allowed to fail queueing a FIN frame under any circumstances.
1612 void tcp_send_fin(struct sock
*sk
)
1614 struct tcp_sock
*tp
= tcp_sk(sk
);
1615 struct sk_buff
*skb
= skb_peek_tail(&sk
->sk_write_queue
);
1618 /* Optimization, tack on the FIN if we have a queue of
1619 * unsent frames. But be careful about outgoing SACKS
1622 mss_now
= tcp_current_mss(sk
, 1);
1624 if (sk
->sk_send_head
!= NULL
) {
1625 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_FIN
;
1626 TCP_SKB_CB(skb
)->end_seq
++;
1629 /* Socket is locked, keep trying until memory is available. */
1631 skb
= alloc_skb_fclone(MAX_TCP_HEADER
, GFP_KERNEL
);
1637 /* Reserve space for headers and prepare control bits. */
1638 skb_reserve(skb
, MAX_TCP_HEADER
);
1640 TCP_SKB_CB(skb
)->flags
= (TCPCB_FLAG_ACK
| TCPCB_FLAG_FIN
);
1641 TCP_SKB_CB(skb
)->sacked
= 0;
1642 skb_shinfo(skb
)->tso_segs
= 1;
1643 skb_shinfo(skb
)->tso_size
= 0;
1645 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
1646 TCP_SKB_CB(skb
)->seq
= tp
->write_seq
;
1647 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ 1;
1648 tcp_queue_skb(sk
, skb
);
1650 __tcp_push_pending_frames(sk
, tp
, mss_now
, TCP_NAGLE_OFF
);
1653 /* We get here when a process closes a file descriptor (either due to
1654 * an explicit close() or as a byproduct of exit()'ing) and there
1655 * was unread data in the receive queue. This behavior is recommended
1656 * by draft-ietf-tcpimpl-prob-03.txt section 3.10. -DaveM
1658 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
1660 struct tcp_sock
*tp
= tcp_sk(sk
);
1661 struct sk_buff
*skb
;
1663 /* NOTE: No TCP options attached and we never retransmit this. */
1664 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
1666 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED
);
1670 /* Reserve space for headers and prepare control bits. */
1671 skb_reserve(skb
, MAX_TCP_HEADER
);
1673 TCP_SKB_CB(skb
)->flags
= (TCPCB_FLAG_ACK
| TCPCB_FLAG_RST
);
1674 TCP_SKB_CB(skb
)->sacked
= 0;
1675 skb_shinfo(skb
)->tso_segs
= 1;
1676 skb_shinfo(skb
)->tso_size
= 0;
1679 TCP_SKB_CB(skb
)->seq
= tcp_acceptable_seq(sk
, tp
);
1680 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
;
1681 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1682 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
1683 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED
);
1686 /* WARNING: This routine must only be called when we have already sent
1687 * a SYN packet that crossed the incoming SYN that caused this routine
1688 * to get called. If this assumption fails then the initial rcv_wnd
1689 * and rcv_wscale values will not be correct.
1691 int tcp_send_synack(struct sock
*sk
)
1693 struct sk_buff
* skb
;
1695 skb
= skb_peek(&sk
->sk_write_queue
);
1696 if (skb
== NULL
|| !(TCP_SKB_CB(skb
)->flags
&TCPCB_FLAG_SYN
)) {
1697 printk(KERN_DEBUG
"tcp_send_synack: wrong queue state\n");
1700 if (!(TCP_SKB_CB(skb
)->flags
&TCPCB_FLAG_ACK
)) {
1701 if (skb_cloned(skb
)) {
1702 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
1705 __skb_unlink(skb
, &sk
->sk_write_queue
);
1706 skb_header_release(nskb
);
1707 __skb_queue_head(&sk
->sk_write_queue
, nskb
);
1708 sk_stream_free_skb(sk
, skb
);
1709 sk_charge_skb(sk
, nskb
);
1713 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_ACK
;
1714 TCP_ECN_send_synack(tcp_sk(sk
), skb
);
1716 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1717 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
1721 * Prepare a SYN-ACK.
1723 struct sk_buff
* tcp_make_synack(struct sock
*sk
, struct dst_entry
*dst
,
1724 struct request_sock
*req
)
1726 struct inet_request_sock
*ireq
= inet_rsk(req
);
1727 struct tcp_sock
*tp
= tcp_sk(sk
);
1729 int tcp_header_size
;
1730 struct sk_buff
*skb
;
1732 skb
= sock_wmalloc(sk
, MAX_TCP_HEADER
+ 15, 1, GFP_ATOMIC
);
1736 /* Reserve space for headers. */
1737 skb_reserve(skb
, MAX_TCP_HEADER
);
1739 skb
->dst
= dst_clone(dst
);
1741 tcp_header_size
= (sizeof(struct tcphdr
) + TCPOLEN_MSS
+
1742 (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0) +
1743 (ireq
->wscale_ok
? TCPOLEN_WSCALE_ALIGNED
: 0) +
1744 /* SACK_PERM is in the place of NOP NOP of TS */
1745 ((ireq
->sack_ok
&& !ireq
->tstamp_ok
) ? TCPOLEN_SACKPERM_ALIGNED
: 0));
1746 skb
->h
.th
= th
= (struct tcphdr
*) skb_push(skb
, tcp_header_size
);
1748 memset(th
, 0, sizeof(struct tcphdr
));
1751 if (dst
->dev
->features
&NETIF_F_TSO
)
1753 TCP_ECN_make_synack(req
, th
);
1754 th
->source
= inet_sk(sk
)->sport
;
1755 th
->dest
= ireq
->rmt_port
;
1756 TCP_SKB_CB(skb
)->seq
= tcp_rsk(req
)->snt_isn
;
1757 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ 1;
1758 TCP_SKB_CB(skb
)->sacked
= 0;
1759 skb_shinfo(skb
)->tso_segs
= 1;
1760 skb_shinfo(skb
)->tso_size
= 0;
1761 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
1762 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_isn
+ 1);
1763 if (req
->rcv_wnd
== 0) { /* ignored for retransmitted syns */
1765 /* Set this up on the first call only */
1766 req
->window_clamp
= tp
->window_clamp
? : dst_metric(dst
, RTAX_WINDOW
);
1767 /* tcp_full_space because it is guaranteed to be the first packet */
1768 tcp_select_initial_window(tcp_full_space(sk
),
1769 dst_metric(dst
, RTAX_ADVMSS
) - (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0),
1774 ireq
->rcv_wscale
= rcv_wscale
;
1777 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
1778 th
->window
= htons(req
->rcv_wnd
);
1780 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1781 tcp_syn_build_options((__u32
*)(th
+ 1), dst_metric(dst
, RTAX_ADVMSS
), ireq
->tstamp_ok
,
1782 ireq
->sack_ok
, ireq
->wscale_ok
, ireq
->rcv_wscale
,
1783 TCP_SKB_CB(skb
)->when
,
1787 th
->doff
= (tcp_header_size
>> 2);
1788 TCP_INC_STATS(TCP_MIB_OUTSEGS
);
1793 * Do all connect socket setups that can be done AF independent.
1795 static inline void tcp_connect_init(struct sock
*sk
)
1797 struct dst_entry
*dst
= __sk_dst_get(sk
);
1798 struct tcp_sock
*tp
= tcp_sk(sk
);
1801 /* We'll fix this up when we get a response from the other end.
1802 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
1804 tp
->tcp_header_len
= sizeof(struct tcphdr
) +
1805 (sysctl_tcp_timestamps
? TCPOLEN_TSTAMP_ALIGNED
: 0);
1807 /* If user gave his TCP_MAXSEG, record it to clamp */
1808 if (tp
->rx_opt
.user_mss
)
1809 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
1811 tcp_sync_mss(sk
, dst_mtu(dst
));
1813 if (!tp
->window_clamp
)
1814 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
1815 tp
->advmss
= dst_metric(dst
, RTAX_ADVMSS
);
1816 tcp_initialize_rcv_mss(sk
);
1818 tcp_select_initial_window(tcp_full_space(sk
),
1819 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
1822 sysctl_tcp_window_scaling
,
1825 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
1826 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
1829 sock_reset_flag(sk
, SOCK_DONE
);
1831 tcp_init_wl(tp
, tp
->write_seq
, 0);
1832 tp
->snd_una
= tp
->write_seq
;
1833 tp
->snd_sml
= tp
->write_seq
;
1838 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
1839 inet_csk(sk
)->icsk_retransmits
= 0;
1840 tcp_clear_retrans(tp
);
1844 * Build a SYN and send it off.
1846 int tcp_connect(struct sock
*sk
)
1848 struct tcp_sock
*tp
= tcp_sk(sk
);
1849 struct sk_buff
*buff
;
1851 tcp_connect_init(sk
);
1853 buff
= alloc_skb_fclone(MAX_TCP_HEADER
+ 15, sk
->sk_allocation
);
1854 if (unlikely(buff
== NULL
))
1857 /* Reserve space for headers. */
1858 skb_reserve(buff
, MAX_TCP_HEADER
);
1860 TCP_SKB_CB(buff
)->flags
= TCPCB_FLAG_SYN
;
1861 TCP_ECN_send_syn(sk
, tp
, buff
);
1862 TCP_SKB_CB(buff
)->sacked
= 0;
1863 skb_shinfo(buff
)->tso_segs
= 1;
1864 skb_shinfo(buff
)->tso_size
= 0;
1866 TCP_SKB_CB(buff
)->seq
= tp
->write_seq
++;
1867 TCP_SKB_CB(buff
)->end_seq
= tp
->write_seq
;
1868 tp
->snd_nxt
= tp
->write_seq
;
1869 tp
->pushed_seq
= tp
->write_seq
;
1872 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
1873 tp
->retrans_stamp
= TCP_SKB_CB(buff
)->when
;
1874 skb_header_release(buff
);
1875 __skb_queue_tail(&sk
->sk_write_queue
, buff
);
1876 sk_charge_skb(sk
, buff
);
1877 tp
->packets_out
+= tcp_skb_pcount(buff
);
1878 tcp_transmit_skb(sk
, buff
, 1, GFP_KERNEL
);
1879 TCP_INC_STATS(TCP_MIB_ACTIVEOPENS
);
1881 /* Timer for repeating the SYN until an answer. */
1882 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1883 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
1887 /* Send out a delayed ack, the caller does the policy checking
1888 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
1891 void tcp_send_delayed_ack(struct sock
*sk
)
1893 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1894 int ato
= icsk
->icsk_ack
.ato
;
1895 unsigned long timeout
;
1897 if (ato
> TCP_DELACK_MIN
) {
1898 const struct tcp_sock
*tp
= tcp_sk(sk
);
1901 if (icsk
->icsk_ack
.pingpong
|| (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
1902 max_ato
= TCP_DELACK_MAX
;
1904 /* Slow path, intersegment interval is "high". */
1906 /* If some rtt estimate is known, use it to bound delayed ack.
1907 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
1911 int rtt
= max(tp
->srtt
>>3, TCP_DELACK_MIN
);
1917 ato
= min(ato
, max_ato
);
1920 /* Stay within the limit we were given */
1921 timeout
= jiffies
+ ato
;
1923 /* Use new timeout only if there wasn't a older one earlier. */
1924 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
1925 /* If delack timer was blocked or is about to expire,
1928 if (icsk
->icsk_ack
.blocked
||
1929 time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
1934 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
1935 timeout
= icsk
->icsk_ack
.timeout
;
1937 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
1938 icsk
->icsk_ack
.timeout
= timeout
;
1939 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
1942 /* This routine sends an ack and also updates the window. */
1943 void tcp_send_ack(struct sock
*sk
)
1945 /* If we have been reset, we may not send again. */
1946 if (sk
->sk_state
!= TCP_CLOSE
) {
1947 struct tcp_sock
*tp
= tcp_sk(sk
);
1948 struct sk_buff
*buff
;
1950 /* We are not putting this on the write queue, so
1951 * tcp_transmit_skb() will set the ownership to this
1954 buff
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
1956 inet_csk_schedule_ack(sk
);
1957 inet_csk(sk
)->icsk_ack
.ato
= TCP_ATO_MIN
;
1958 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
1959 TCP_DELACK_MAX
, TCP_RTO_MAX
);
1963 /* Reserve space for headers and prepare control bits. */
1964 skb_reserve(buff
, MAX_TCP_HEADER
);
1966 TCP_SKB_CB(buff
)->flags
= TCPCB_FLAG_ACK
;
1967 TCP_SKB_CB(buff
)->sacked
= 0;
1968 skb_shinfo(buff
)->tso_segs
= 1;
1969 skb_shinfo(buff
)->tso_size
= 0;
1971 /* Send it off, this clears delayed acks for us. */
1972 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(buff
)->end_seq
= tcp_acceptable_seq(sk
, tp
);
1973 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
1974 tcp_transmit_skb(sk
, buff
, 0, GFP_ATOMIC
);
1978 /* This routine sends a packet with an out of date sequence
1979 * number. It assumes the other end will try to ack it.
1981 * Question: what should we make while urgent mode?
1982 * 4.4BSD forces sending single byte of data. We cannot send
1983 * out of window data, because we have SND.NXT==SND.MAX...
1985 * Current solution: to send TWO zero-length segments in urgent mode:
1986 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
1987 * out-of-date with SND.UNA-1 to probe window.
1989 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
)
1991 struct tcp_sock
*tp
= tcp_sk(sk
);
1992 struct sk_buff
*skb
;
1994 /* We don't queue it, tcp_transmit_skb() sets ownership. */
1995 skb
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
1999 /* Reserve space for headers and set control bits. */
2000 skb_reserve(skb
, MAX_TCP_HEADER
);
2002 TCP_SKB_CB(skb
)->flags
= TCPCB_FLAG_ACK
;
2003 TCP_SKB_CB(skb
)->sacked
= urgent
;
2004 skb_shinfo(skb
)->tso_segs
= 1;
2005 skb_shinfo(skb
)->tso_size
= 0;
2007 /* Use a previous sequence. This should cause the other
2008 * end to send an ack. Don't queue or clone SKB, just
2011 TCP_SKB_CB(skb
)->seq
= urgent
? tp
->snd_una
: tp
->snd_una
- 1;
2012 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
;
2013 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2014 return tcp_transmit_skb(sk
, skb
, 0, GFP_ATOMIC
);
2017 int tcp_write_wakeup(struct sock
*sk
)
2019 if (sk
->sk_state
!= TCP_CLOSE
) {
2020 struct tcp_sock
*tp
= tcp_sk(sk
);
2021 struct sk_buff
*skb
;
2023 if ((skb
= sk
->sk_send_head
) != NULL
&&
2024 before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
+tp
->snd_wnd
)) {
2026 unsigned int mss
= tcp_current_mss(sk
, 0);
2027 unsigned int seg_size
= tp
->snd_una
+tp
->snd_wnd
-TCP_SKB_CB(skb
)->seq
;
2029 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
2030 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
2032 /* We are probing the opening of a window
2033 * but the window size is != 0
2034 * must have been a result SWS avoidance ( sender )
2036 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
2038 seg_size
= min(seg_size
, mss
);
2039 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2040 if (tcp_fragment(sk
, skb
, seg_size
, mss
))
2042 } else if (!tcp_skb_pcount(skb
))
2043 tcp_set_skb_tso_segs(sk
, skb
, mss
);
2045 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2046 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2047 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2049 update_send_head(sk
, tp
, skb
);
2054 between(tp
->snd_up
, tp
->snd_una
+1, tp
->snd_una
+0xFFFF))
2055 tcp_xmit_probe_skb(sk
, TCPCB_URG
);
2056 return tcp_xmit_probe_skb(sk
, 0);
2062 /* A window probe timeout has occurred. If window is not closed send
2063 * a partial packet else a zero probe.
2065 void tcp_send_probe0(struct sock
*sk
)
2067 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2068 struct tcp_sock
*tp
= tcp_sk(sk
);
2071 err
= tcp_write_wakeup(sk
);
2073 if (tp
->packets_out
|| !sk
->sk_send_head
) {
2074 /* Cancel probe timer, if it is not required. */
2075 icsk
->icsk_probes_out
= 0;
2076 icsk
->icsk_backoff
= 0;
2081 if (icsk
->icsk_backoff
< sysctl_tcp_retries2
)
2082 icsk
->icsk_backoff
++;
2083 icsk
->icsk_probes_out
++;
2084 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2085 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
2088 /* If packet was not sent due to local congestion,
2089 * do not backoff and do not remember icsk_probes_out.
2090 * Let local senders to fight for local resources.
2092 * Use accumulated backoff yet.
2094 if (!icsk
->icsk_probes_out
)
2095 icsk
->icsk_probes_out
= 1;
2096 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2097 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
,
2098 TCP_RESOURCE_PROBE_INTERVAL
),
2103 EXPORT_SYMBOL(tcp_connect
);
2104 EXPORT_SYMBOL(tcp_make_synack
);
2105 EXPORT_SYMBOL(tcp_simple_retransmit
);
2106 EXPORT_SYMBOL(tcp_sync_mss
);
2107 EXPORT_SYMBOL(sysctl_tcp_tso_win_divisor
);