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 inet_csk(sk)->icsk_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. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
635 are READ ONLY outside this function. --ANK (980731)
638 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
640 struct tcp_sock
*tp
= tcp_sk(sk
);
641 struct inet_connection_sock
*icsk
= inet_csk(sk
);
642 /* Calculate base mss without TCP options:
643 It is MMS_S - sizeof(tcphdr) of rfc1122
645 int mss_now
= (pmtu
- icsk
->icsk_af_ops
->net_header_len
-
646 sizeof(struct tcphdr
));
648 /* Clamp it (mss_clamp does not include tcp options) */
649 if (mss_now
> tp
->rx_opt
.mss_clamp
)
650 mss_now
= tp
->rx_opt
.mss_clamp
;
652 /* Now subtract optional transport overhead */
653 mss_now
-= icsk
->icsk_ext_hdr_len
;
655 /* Then reserve room for full set of TCP options and 8 bytes of data */
659 /* Now subtract TCP options size, not including SACKs */
660 mss_now
-= tp
->tcp_header_len
- sizeof(struct tcphdr
);
662 /* Bound mss with half of window */
663 if (tp
->max_window
&& mss_now
> (tp
->max_window
>>1))
664 mss_now
= max((tp
->max_window
>>1), 68U - tp
->tcp_header_len
);
666 /* And store cached results */
667 icsk
->icsk_pmtu_cookie
= pmtu
;
668 tp
->mss_cache
= mss_now
;
673 /* Compute the current effective MSS, taking SACKs and IP options,
674 * and even PMTU discovery events into account.
676 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
677 * cannot be large. However, taking into account rare use of URG, this
680 unsigned int tcp_current_mss(struct sock
*sk
, int large_allowed
)
682 struct tcp_sock
*tp
= tcp_sk(sk
);
683 struct dst_entry
*dst
= __sk_dst_get(sk
);
688 mss_now
= tp
->mss_cache
;
691 (sk
->sk_route_caps
& NETIF_F_TSO
) &&
696 u32 mtu
= dst_mtu(dst
);
697 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
698 mss_now
= tcp_sync_mss(sk
, mtu
);
701 if (tp
->rx_opt
.eff_sacks
)
702 mss_now
-= (TCPOLEN_SACK_BASE_ALIGNED
+
703 (tp
->rx_opt
.eff_sacks
* TCPOLEN_SACK_PERBLOCK
));
705 xmit_size_goal
= mss_now
;
708 xmit_size_goal
= (65535 -
709 inet_csk(sk
)->icsk_af_ops
->net_header_len
-
710 inet_csk(sk
)->icsk_ext_hdr_len
-
713 if (tp
->max_window
&&
714 (xmit_size_goal
> (tp
->max_window
>> 1)))
715 xmit_size_goal
= max((tp
->max_window
>> 1),
716 68U - tp
->tcp_header_len
);
718 xmit_size_goal
-= (xmit_size_goal
% mss_now
);
720 tp
->xmit_size_goal
= xmit_size_goal
;
725 /* Congestion window validation. (RFC2861) */
727 static inline void tcp_cwnd_validate(struct sock
*sk
, struct tcp_sock
*tp
)
729 __u32 packets_out
= tp
->packets_out
;
731 if (packets_out
>= tp
->snd_cwnd
) {
732 /* Network is feed fully. */
733 tp
->snd_cwnd_used
= 0;
734 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
736 /* Network starves. */
737 if (tp
->packets_out
> tp
->snd_cwnd_used
)
738 tp
->snd_cwnd_used
= tp
->packets_out
;
740 if ((s32
)(tcp_time_stamp
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
)
741 tcp_cwnd_application_limited(sk
);
745 static unsigned int tcp_window_allows(struct tcp_sock
*tp
, struct sk_buff
*skb
, unsigned int mss_now
, unsigned int cwnd
)
747 u32 window
, cwnd_len
;
749 window
= (tp
->snd_una
+ tp
->snd_wnd
- TCP_SKB_CB(skb
)->seq
);
750 cwnd_len
= mss_now
* cwnd
;
751 return min(window
, cwnd_len
);
754 /* Can at least one segment of SKB be sent right now, according to the
755 * congestion window rules? If so, return how many segments are allowed.
757 static inline unsigned int tcp_cwnd_test(struct tcp_sock
*tp
, struct sk_buff
*skb
)
761 /* Don't be strict about the congestion window for the final FIN. */
762 if (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
)
765 in_flight
= tcp_packets_in_flight(tp
);
767 if (in_flight
< cwnd
)
768 return (cwnd
- in_flight
);
773 /* This must be invoked the first time we consider transmitting
776 static inline int tcp_init_tso_segs(struct sock
*sk
, struct sk_buff
*skb
, unsigned int mss_now
)
778 int tso_segs
= tcp_skb_pcount(skb
);
782 skb_shinfo(skb
)->tso_size
!= mss_now
)) {
783 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
784 tso_segs
= tcp_skb_pcount(skb
);
789 static inline int tcp_minshall_check(const struct tcp_sock
*tp
)
791 return after(tp
->snd_sml
,tp
->snd_una
) &&
792 !after(tp
->snd_sml
, tp
->snd_nxt
);
795 /* Return 0, if packet can be sent now without violation Nagle's rules:
796 * 1. It is full sized.
797 * 2. Or it contains FIN. (already checked by caller)
798 * 3. Or TCP_NODELAY was set.
799 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
800 * With Minshall's modification: all sent small packets are ACKed.
803 static inline int tcp_nagle_check(const struct tcp_sock
*tp
,
804 const struct sk_buff
*skb
,
805 unsigned mss_now
, int nonagle
)
807 return (skb
->len
< mss_now
&&
808 ((nonagle
&TCP_NAGLE_CORK
) ||
811 tcp_minshall_check(tp
))));
814 /* Return non-zero if the Nagle test allows this packet to be
817 static inline int tcp_nagle_test(struct tcp_sock
*tp
, struct sk_buff
*skb
,
818 unsigned int cur_mss
, int nonagle
)
820 /* Nagle rule does not apply to frames, which sit in the middle of the
821 * write_queue (they have no chances to get new data).
823 * This is implemented in the callers, where they modify the 'nonagle'
824 * argument based upon the location of SKB in the send queue.
826 if (nonagle
& TCP_NAGLE_PUSH
)
829 /* Don't use the nagle rule for urgent data (or for the final FIN). */
831 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
))
834 if (!tcp_nagle_check(tp
, skb
, cur_mss
, nonagle
))
840 /* Does at least the first segment of SKB fit into the send window? */
841 static inline int tcp_snd_wnd_test(struct tcp_sock
*tp
, struct sk_buff
*skb
, unsigned int cur_mss
)
843 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
845 if (skb
->len
> cur_mss
)
846 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
848 return !after(end_seq
, tp
->snd_una
+ tp
->snd_wnd
);
851 /* This checks if the data bearing packet SKB (usually sk->sk_send_head)
852 * should be put on the wire right now. If so, it returns the number of
853 * packets allowed by the congestion window.
855 static unsigned int tcp_snd_test(struct sock
*sk
, struct sk_buff
*skb
,
856 unsigned int cur_mss
, int nonagle
)
858 struct tcp_sock
*tp
= tcp_sk(sk
);
859 unsigned int cwnd_quota
;
861 tcp_init_tso_segs(sk
, skb
, cur_mss
);
863 if (!tcp_nagle_test(tp
, skb
, cur_mss
, nonagle
))
866 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
868 !tcp_snd_wnd_test(tp
, skb
, cur_mss
))
874 static inline int tcp_skb_is_last(const struct sock
*sk
,
875 const struct sk_buff
*skb
)
877 return skb
->next
== (struct sk_buff
*)&sk
->sk_write_queue
;
880 int tcp_may_send_now(struct sock
*sk
, struct tcp_sock
*tp
)
882 struct sk_buff
*skb
= sk
->sk_send_head
;
885 tcp_snd_test(sk
, skb
, tcp_current_mss(sk
, 1),
886 (tcp_skb_is_last(sk
, skb
) ?
891 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
892 * which is put after SKB on the list. It is very much like
893 * tcp_fragment() except that it may make several kinds of assumptions
894 * in order to speed up the splitting operation. In particular, we
895 * know that all the data is in scatter-gather pages, and that the
896 * packet has never been sent out before (and thus is not cloned).
898 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
, unsigned int mss_now
)
900 struct sk_buff
*buff
;
901 int nlen
= skb
->len
- len
;
904 /* All of a TSO frame must be composed of paged data. */
905 if (skb
->len
!= skb
->data_len
)
906 return tcp_fragment(sk
, skb
, len
, mss_now
);
908 buff
= sk_stream_alloc_pskb(sk
, 0, 0, GFP_ATOMIC
);
909 if (unlikely(buff
== NULL
))
912 buff
->truesize
= nlen
;
913 skb
->truesize
-= nlen
;
915 /* Correct the sequence numbers. */
916 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
917 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
918 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
920 /* PSH and FIN should only be set in the second packet. */
921 flags
= TCP_SKB_CB(skb
)->flags
;
922 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
923 TCP_SKB_CB(buff
)->flags
= flags
;
925 /* This packet was never sent out yet, so no SACK bits. */
926 TCP_SKB_CB(buff
)->sacked
= 0;
928 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_HW
;
929 skb_split(skb
, buff
, len
);
931 /* Fix up tso_factor for both original and new SKB. */
932 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
933 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
935 /* Link BUFF into the send queue. */
936 skb_header_release(buff
);
937 __skb_append(skb
, buff
, &sk
->sk_write_queue
);
942 /* Try to defer sending, if possible, in order to minimize the amount
943 * of TSO splitting we do. View it as a kind of TSO Nagle test.
945 * This algorithm is from John Heffner.
947 static int tcp_tso_should_defer(struct sock
*sk
, struct tcp_sock
*tp
, struct sk_buff
*skb
)
949 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
950 u32 send_win
, cong_win
, limit
, in_flight
;
952 if (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
)
955 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
958 in_flight
= tcp_packets_in_flight(tp
);
960 BUG_ON(tcp_skb_pcount(skb
) <= 1 ||
961 (tp
->snd_cwnd
<= in_flight
));
963 send_win
= (tp
->snd_una
+ tp
->snd_wnd
) - TCP_SKB_CB(skb
)->seq
;
965 /* From in_flight test above, we know that cwnd > in_flight. */
966 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
968 limit
= min(send_win
, cong_win
);
970 if (sysctl_tcp_tso_win_divisor
) {
971 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
973 /* If at least some fraction of a window is available,
976 chunk
/= sysctl_tcp_tso_win_divisor
;
980 /* Different approach, try not to defer past a single
981 * ACK. Receiver should ACK every other full sized
982 * frame, so if we have space for more than 3 frames
985 if (limit
> tcp_max_burst(tp
) * tp
->mss_cache
)
989 /* Ok, it looks like it is advisable to defer. */
993 /* This routine writes packets to the network. It advances the
994 * send_head. This happens as incoming acks open up the remote
997 * Returns 1, if no segments are in flight and we have queued segments, but
998 * cannot send anything now because of SWS or another problem.
1000 static int tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
)
1002 struct tcp_sock
*tp
= tcp_sk(sk
);
1003 struct sk_buff
*skb
;
1004 unsigned int tso_segs
, sent_pkts
;
1007 /* If we are closed, the bytes will have to remain here.
1008 * In time closedown will finish, we empty the write queue and all
1011 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
1015 while ((skb
= sk
->sk_send_head
)) {
1018 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1021 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1025 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
)))
1028 if (tso_segs
== 1) {
1029 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
1030 (tcp_skb_is_last(sk
, skb
) ?
1031 nonagle
: TCP_NAGLE_PUSH
))))
1034 if (tcp_tso_should_defer(sk
, tp
, skb
))
1040 limit
= tcp_window_allows(tp
, skb
,
1041 mss_now
, cwnd_quota
);
1043 if (skb
->len
< limit
) {
1044 unsigned int trim
= skb
->len
% mss_now
;
1047 limit
= skb
->len
- trim
;
1051 if (skb
->len
> limit
&&
1052 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
)))
1055 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1057 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
)))
1060 /* Advance the send_head. This one is sent out.
1061 * This call will increment packets_out.
1063 update_send_head(sk
, tp
, skb
);
1065 tcp_minshall_update(tp
, mss_now
, skb
);
1069 if (likely(sent_pkts
)) {
1070 tcp_cwnd_validate(sk
, tp
);
1073 return !tp
->packets_out
&& sk
->sk_send_head
;
1076 /* Push out any pending frames which were held back due to
1077 * TCP_CORK or attempt at coalescing tiny packets.
1078 * The socket must be locked by the caller.
1080 void __tcp_push_pending_frames(struct sock
*sk
, struct tcp_sock
*tp
,
1081 unsigned int cur_mss
, int nonagle
)
1083 struct sk_buff
*skb
= sk
->sk_send_head
;
1086 if (tcp_write_xmit(sk
, cur_mss
, nonagle
))
1087 tcp_check_probe_timer(sk
, tp
);
1091 /* Send _single_ skb sitting at the send head. This function requires
1092 * true push pending frames to setup probe timer etc.
1094 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
1096 struct tcp_sock
*tp
= tcp_sk(sk
);
1097 struct sk_buff
*skb
= sk
->sk_send_head
;
1098 unsigned int tso_segs
, cwnd_quota
;
1100 BUG_ON(!skb
|| skb
->len
< mss_now
);
1102 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1103 cwnd_quota
= tcp_snd_test(sk
, skb
, mss_now
, TCP_NAGLE_PUSH
);
1105 if (likely(cwnd_quota
)) {
1112 limit
= tcp_window_allows(tp
, skb
,
1113 mss_now
, cwnd_quota
);
1115 if (skb
->len
< limit
) {
1116 unsigned int trim
= skb
->len
% mss_now
;
1119 limit
= skb
->len
- trim
;
1123 if (skb
->len
> limit
&&
1124 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
)))
1127 /* Send it out now. */
1128 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1130 if (likely(!tcp_transmit_skb(sk
, skb
, 1, sk
->sk_allocation
))) {
1131 update_send_head(sk
, tp
, skb
);
1132 tcp_cwnd_validate(sk
, tp
);
1138 /* This function returns the amount that we can raise the
1139 * usable window based on the following constraints
1141 * 1. The window can never be shrunk once it is offered (RFC 793)
1142 * 2. We limit memory per socket
1145 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1146 * RECV.NEXT + RCV.WIN fixed until:
1147 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1149 * i.e. don't raise the right edge of the window until you can raise
1150 * it at least MSS bytes.
1152 * Unfortunately, the recommended algorithm breaks header prediction,
1153 * since header prediction assumes th->window stays fixed.
1155 * Strictly speaking, keeping th->window fixed violates the receiver
1156 * side SWS prevention criteria. The problem is that under this rule
1157 * a stream of single byte packets will cause the right side of the
1158 * window to always advance by a single byte.
1160 * Of course, if the sender implements sender side SWS prevention
1161 * then this will not be a problem.
1163 * BSD seems to make the following compromise:
1165 * If the free space is less than the 1/4 of the maximum
1166 * space available and the free space is less than 1/2 mss,
1167 * then set the window to 0.
1168 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1169 * Otherwise, just prevent the window from shrinking
1170 * and from being larger than the largest representable value.
1172 * This prevents incremental opening of the window in the regime
1173 * where TCP is limited by the speed of the reader side taking
1174 * data out of the TCP receive queue. It does nothing about
1175 * those cases where the window is constrained on the sender side
1176 * because the pipeline is full.
1178 * BSD also seems to "accidentally" limit itself to windows that are a
1179 * multiple of MSS, at least until the free space gets quite small.
1180 * This would appear to be a side effect of the mbuf implementation.
1181 * Combining these two algorithms results in the observed behavior
1182 * of having a fixed window size at almost all times.
1184 * Below we obtain similar behavior by forcing the offered window to
1185 * a multiple of the mss when it is feasible to do so.
1187 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1188 * Regular options like TIMESTAMP are taken into account.
1190 u32
__tcp_select_window(struct sock
*sk
)
1192 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1193 struct tcp_sock
*tp
= tcp_sk(sk
);
1194 /* MSS for the peer's data. Previous versions used mss_clamp
1195 * here. I don't know if the value based on our guesses
1196 * of peer's MSS is better for the performance. It's more correct
1197 * but may be worse for the performance because of rcv_mss
1198 * fluctuations. --SAW 1998/11/1
1200 int mss
= icsk
->icsk_ack
.rcv_mss
;
1201 int free_space
= tcp_space(sk
);
1202 int full_space
= min_t(int, tp
->window_clamp
, tcp_full_space(sk
));
1205 if (mss
> full_space
)
1208 if (free_space
< full_space
/2) {
1209 icsk
->icsk_ack
.quick
= 0;
1211 if (tcp_memory_pressure
)
1212 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U*tp
->advmss
);
1214 if (free_space
< mss
)
1218 if (free_space
> tp
->rcv_ssthresh
)
1219 free_space
= tp
->rcv_ssthresh
;
1221 /* Don't do rounding if we are using window scaling, since the
1222 * scaled window will not line up with the MSS boundary anyway.
1224 window
= tp
->rcv_wnd
;
1225 if (tp
->rx_opt
.rcv_wscale
) {
1226 window
= free_space
;
1228 /* Advertise enough space so that it won't get scaled away.
1229 * Import case: prevent zero window announcement if
1230 * 1<<rcv_wscale > mss.
1232 if (((window
>> tp
->rx_opt
.rcv_wscale
) << tp
->rx_opt
.rcv_wscale
) != window
)
1233 window
= (((window
>> tp
->rx_opt
.rcv_wscale
) + 1)
1234 << tp
->rx_opt
.rcv_wscale
);
1236 /* Get the largest window that is a nice multiple of mss.
1237 * Window clamp already applied above.
1238 * If our current window offering is within 1 mss of the
1239 * free space we just keep it. This prevents the divide
1240 * and multiply from happening most of the time.
1241 * We also don't do any window rounding when the free space
1244 if (window
<= free_space
- mss
|| window
> free_space
)
1245 window
= (free_space
/mss
)*mss
;
1251 /* Attempt to collapse two adjacent SKB's during retransmission. */
1252 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*skb
, int mss_now
)
1254 struct tcp_sock
*tp
= tcp_sk(sk
);
1255 struct sk_buff
*next_skb
= skb
->next
;
1257 /* The first test we must make is that neither of these two
1258 * SKB's are still referenced by someone else.
1260 if (!skb_cloned(skb
) && !skb_cloned(next_skb
)) {
1261 int skb_size
= skb
->len
, next_skb_size
= next_skb
->len
;
1262 u16 flags
= TCP_SKB_CB(skb
)->flags
;
1264 /* Also punt if next skb has been SACK'd. */
1265 if(TCP_SKB_CB(next_skb
)->sacked
& TCPCB_SACKED_ACKED
)
1268 /* Next skb is out of window. */
1269 if (after(TCP_SKB_CB(next_skb
)->end_seq
, tp
->snd_una
+tp
->snd_wnd
))
1272 /* Punt if not enough space exists in the first SKB for
1273 * the data in the second, or the total combined payload
1274 * would exceed the MSS.
1276 if ((next_skb_size
> skb_tailroom(skb
)) ||
1277 ((skb_size
+ next_skb_size
) > mss_now
))
1280 BUG_ON(tcp_skb_pcount(skb
) != 1 ||
1281 tcp_skb_pcount(next_skb
) != 1);
1283 /* changing transmit queue under us so clear hints */
1284 clear_all_retrans_hints(tp
);
1286 /* Ok. We will be able to collapse the packet. */
1287 __skb_unlink(next_skb
, &sk
->sk_write_queue
);
1289 memcpy(skb_put(skb
, next_skb_size
), next_skb
->data
, next_skb_size
);
1291 if (next_skb
->ip_summed
== CHECKSUM_HW
)
1292 skb
->ip_summed
= CHECKSUM_HW
;
1294 if (skb
->ip_summed
!= CHECKSUM_HW
)
1295 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
1297 /* Update sequence range on original skb. */
1298 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
1300 /* Merge over control information. */
1301 flags
|= TCP_SKB_CB(next_skb
)->flags
; /* This moves PSH/FIN etc. over */
1302 TCP_SKB_CB(skb
)->flags
= flags
;
1304 /* All done, get rid of second SKB and account for it so
1305 * packet counting does not break.
1307 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
&(TCPCB_EVER_RETRANS
|TCPCB_AT_TAIL
);
1308 if (TCP_SKB_CB(next_skb
)->sacked
&TCPCB_SACKED_RETRANS
)
1309 tp
->retrans_out
-= tcp_skb_pcount(next_skb
);
1310 if (TCP_SKB_CB(next_skb
)->sacked
&TCPCB_LOST
) {
1311 tp
->lost_out
-= tcp_skb_pcount(next_skb
);
1312 tp
->left_out
-= tcp_skb_pcount(next_skb
);
1314 /* Reno case is special. Sigh... */
1315 if (!tp
->rx_opt
.sack_ok
&& tp
->sacked_out
) {
1316 tcp_dec_pcount_approx(&tp
->sacked_out
, next_skb
);
1317 tp
->left_out
-= tcp_skb_pcount(next_skb
);
1320 /* Not quite right: it can be > snd.fack, but
1321 * it is better to underestimate fackets.
1323 tcp_dec_pcount_approx(&tp
->fackets_out
, next_skb
);
1324 tcp_packets_out_dec(tp
, next_skb
);
1325 sk_stream_free_skb(sk
, next_skb
);
1329 /* Do a simple retransmit without using the backoff mechanisms in
1330 * tcp_timer. This is used for path mtu discovery.
1331 * The socket is already locked here.
1333 void tcp_simple_retransmit(struct sock
*sk
)
1335 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1336 struct tcp_sock
*tp
= tcp_sk(sk
);
1337 struct sk_buff
*skb
;
1338 unsigned int mss
= tcp_current_mss(sk
, 0);
1341 sk_stream_for_retrans_queue(skb
, sk
) {
1342 if (skb
->len
> mss
&&
1343 !(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
)) {
1344 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) {
1345 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1346 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1348 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_LOST
)) {
1349 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1350 tp
->lost_out
+= tcp_skb_pcount(skb
);
1356 clear_all_retrans_hints(tp
);
1361 tcp_sync_left_out(tp
);
1363 /* Don't muck with the congestion window here.
1364 * Reason is that we do not increase amount of _data_
1365 * in network, but units changed and effective
1366 * cwnd/ssthresh really reduced now.
1368 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
1369 tp
->high_seq
= tp
->snd_nxt
;
1370 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
1371 tp
->prior_ssthresh
= 0;
1372 tp
->undo_marker
= 0;
1373 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1375 tcp_xmit_retransmit_queue(sk
);
1378 /* This retransmits one SKB. Policy decisions and retransmit queue
1379 * state updates are done by the caller. Returns non-zero if an
1380 * error occurred which prevented the send.
1382 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
1384 struct tcp_sock
*tp
= tcp_sk(sk
);
1385 unsigned int cur_mss
= tcp_current_mss(sk
, 0);
1388 /* Do not sent more than we queued. 1/4 is reserved for possible
1389 * copying overhead: fragmentation, tunneling, mangling etc.
1391 if (atomic_read(&sk
->sk_wmem_alloc
) >
1392 min(sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2), sk
->sk_sndbuf
))
1395 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
1396 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1398 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
1402 /* If receiver has shrunk his window, and skb is out of
1403 * new window, do not retransmit it. The exception is the
1404 * case, when window is shrunk to zero. In this case
1405 * our retransmit serves as a zero window probe.
1407 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
+tp
->snd_wnd
)
1408 && TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
1411 if (skb
->len
> cur_mss
) {
1412 if (tcp_fragment(sk
, skb
, cur_mss
, cur_mss
))
1413 return -ENOMEM
; /* We'll try again later. */
1416 /* Collapse two adjacent packets if worthwhile and we can. */
1417 if(!(TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_SYN
) &&
1418 (skb
->len
< (cur_mss
>> 1)) &&
1419 (skb
->next
!= sk
->sk_send_head
) &&
1420 (skb
->next
!= (struct sk_buff
*)&sk
->sk_write_queue
) &&
1421 (skb_shinfo(skb
)->nr_frags
== 0 && skb_shinfo(skb
->next
)->nr_frags
== 0) &&
1422 (tcp_skb_pcount(skb
) == 1 && tcp_skb_pcount(skb
->next
) == 1) &&
1423 (sysctl_tcp_retrans_collapse
!= 0))
1424 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
1426 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
1427 return -EHOSTUNREACH
; /* Routing failure or similar. */
1429 /* Some Solaris stacks overoptimize and ignore the FIN on a
1430 * retransmit when old data is attached. So strip it off
1431 * since it is cheap to do so and saves bytes on the network.
1434 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
) &&
1435 tp
->snd_una
== (TCP_SKB_CB(skb
)->end_seq
- 1)) {
1436 if (!pskb_trim(skb
, 0)) {
1437 TCP_SKB_CB(skb
)->seq
= TCP_SKB_CB(skb
)->end_seq
- 1;
1438 skb_shinfo(skb
)->tso_segs
= 1;
1439 skb_shinfo(skb
)->tso_size
= 0;
1440 skb
->ip_summed
= CHECKSUM_NONE
;
1445 /* Make a copy, if the first transmission SKB clone we made
1446 * is still in somebody's hands, else make a clone.
1448 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1450 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
1453 /* Update global TCP statistics. */
1454 TCP_INC_STATS(TCP_MIB_RETRANSSEGS
);
1456 tp
->total_retrans
++;
1458 #if FASTRETRANS_DEBUG > 0
1459 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) {
1460 if (net_ratelimit())
1461 printk(KERN_DEBUG
"retrans_out leaked.\n");
1464 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
1465 tp
->retrans_out
+= tcp_skb_pcount(skb
);
1467 /* Save stamp of the first retransmit. */
1468 if (!tp
->retrans_stamp
)
1469 tp
->retrans_stamp
= TCP_SKB_CB(skb
)->when
;
1473 /* snd_nxt is stored to detect loss of retransmitted segment,
1474 * see tcp_input.c tcp_sacktag_write_queue().
1476 TCP_SKB_CB(skb
)->ack_seq
= tp
->snd_nxt
;
1481 /* This gets called after a retransmit timeout, and the initially
1482 * retransmitted data is acknowledged. It tries to continue
1483 * resending the rest of the retransmit queue, until either
1484 * we've sent it all or the congestion window limit is reached.
1485 * If doing SACK, the first ACK which comes back for a timeout
1486 * based retransmit packet might feed us FACK information again.
1487 * If so, we use it to avoid unnecessarily retransmissions.
1489 void tcp_xmit_retransmit_queue(struct sock
*sk
)
1491 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1492 struct tcp_sock
*tp
= tcp_sk(sk
);
1493 struct sk_buff
*skb
;
1496 if (tp
->retransmit_skb_hint
) {
1497 skb
= tp
->retransmit_skb_hint
;
1498 packet_cnt
= tp
->retransmit_cnt_hint
;
1500 skb
= sk
->sk_write_queue
.next
;
1504 /* First pass: retransmit lost packets. */
1506 sk_stream_for_retrans_queue_from(skb
, sk
) {
1507 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
1509 /* we could do better than to assign each time */
1510 tp
->retransmit_skb_hint
= skb
;
1511 tp
->retransmit_cnt_hint
= packet_cnt
;
1513 /* Assume this retransmit will generate
1514 * only one packet for congestion window
1515 * calculation purposes. This works because
1516 * tcp_retransmit_skb() will chop up the
1517 * packet to be MSS sized and all the
1518 * packet counting works out.
1520 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
1523 if (sacked
& TCPCB_LOST
) {
1524 if (!(sacked
&(TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))) {
1525 if (tcp_retransmit_skb(sk
, skb
)) {
1526 tp
->retransmit_skb_hint
= NULL
;
1529 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
1530 NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS
);
1532 NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS
);
1535 skb_peek(&sk
->sk_write_queue
))
1536 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1537 inet_csk(sk
)->icsk_rto
,
1541 packet_cnt
+= tcp_skb_pcount(skb
);
1542 if (packet_cnt
>= tp
->lost_out
)
1548 /* OK, demanded retransmission is finished. */
1550 /* Forward retransmissions are possible only during Recovery. */
1551 if (icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
1554 /* No forward retransmissions in Reno are possible. */
1555 if (!tp
->rx_opt
.sack_ok
)
1558 /* Yeah, we have to make difficult choice between forward transmission
1559 * and retransmission... Both ways have their merits...
1561 * For now we do not retransmit anything, while we have some new
1565 if (tcp_may_send_now(sk
, tp
))
1568 if (tp
->forward_skb_hint
) {
1569 skb
= tp
->forward_skb_hint
;
1570 packet_cnt
= tp
->forward_cnt_hint
;
1572 skb
= sk
->sk_write_queue
.next
;
1576 sk_stream_for_retrans_queue_from(skb
, sk
) {
1577 tp
->forward_cnt_hint
= packet_cnt
;
1578 tp
->forward_skb_hint
= skb
;
1580 /* Similar to the retransmit loop above we
1581 * can pretend that the retransmitted SKB
1582 * we send out here will be composed of one
1583 * real MSS sized packet because tcp_retransmit_skb()
1584 * will fragment it if necessary.
1586 if (++packet_cnt
> tp
->fackets_out
)
1589 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
1592 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
)
1595 /* Ok, retransmit it. */
1596 if (tcp_retransmit_skb(sk
, skb
)) {
1597 tp
->forward_skb_hint
= NULL
;
1601 if (skb
== skb_peek(&sk
->sk_write_queue
))
1602 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1603 inet_csk(sk
)->icsk_rto
,
1606 NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS
);
1611 /* Send a fin. The caller locks the socket for us. This cannot be
1612 * allowed to fail queueing a FIN frame under any circumstances.
1614 void tcp_send_fin(struct sock
*sk
)
1616 struct tcp_sock
*tp
= tcp_sk(sk
);
1617 struct sk_buff
*skb
= skb_peek_tail(&sk
->sk_write_queue
);
1620 /* Optimization, tack on the FIN if we have a queue of
1621 * unsent frames. But be careful about outgoing SACKS
1624 mss_now
= tcp_current_mss(sk
, 1);
1626 if (sk
->sk_send_head
!= NULL
) {
1627 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_FIN
;
1628 TCP_SKB_CB(skb
)->end_seq
++;
1631 /* Socket is locked, keep trying until memory is available. */
1633 skb
= alloc_skb_fclone(MAX_TCP_HEADER
, GFP_KERNEL
);
1639 /* Reserve space for headers and prepare control bits. */
1640 skb_reserve(skb
, MAX_TCP_HEADER
);
1642 TCP_SKB_CB(skb
)->flags
= (TCPCB_FLAG_ACK
| TCPCB_FLAG_FIN
);
1643 TCP_SKB_CB(skb
)->sacked
= 0;
1644 skb_shinfo(skb
)->tso_segs
= 1;
1645 skb_shinfo(skb
)->tso_size
= 0;
1647 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
1648 TCP_SKB_CB(skb
)->seq
= tp
->write_seq
;
1649 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ 1;
1650 tcp_queue_skb(sk
, skb
);
1652 __tcp_push_pending_frames(sk
, tp
, mss_now
, TCP_NAGLE_OFF
);
1655 /* We get here when a process closes a file descriptor (either due to
1656 * an explicit close() or as a byproduct of exit()'ing) and there
1657 * was unread data in the receive queue. This behavior is recommended
1658 * by draft-ietf-tcpimpl-prob-03.txt section 3.10. -DaveM
1660 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
1662 struct tcp_sock
*tp
= tcp_sk(sk
);
1663 struct sk_buff
*skb
;
1665 /* NOTE: No TCP options attached and we never retransmit this. */
1666 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
1668 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED
);
1672 /* Reserve space for headers and prepare control bits. */
1673 skb_reserve(skb
, MAX_TCP_HEADER
);
1675 TCP_SKB_CB(skb
)->flags
= (TCPCB_FLAG_ACK
| TCPCB_FLAG_RST
);
1676 TCP_SKB_CB(skb
)->sacked
= 0;
1677 skb_shinfo(skb
)->tso_segs
= 1;
1678 skb_shinfo(skb
)->tso_size
= 0;
1681 TCP_SKB_CB(skb
)->seq
= tcp_acceptable_seq(sk
, tp
);
1682 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
;
1683 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1684 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
1685 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED
);
1688 /* WARNING: This routine must only be called when we have already sent
1689 * a SYN packet that crossed the incoming SYN that caused this routine
1690 * to get called. If this assumption fails then the initial rcv_wnd
1691 * and rcv_wscale values will not be correct.
1693 int tcp_send_synack(struct sock
*sk
)
1695 struct sk_buff
* skb
;
1697 skb
= skb_peek(&sk
->sk_write_queue
);
1698 if (skb
== NULL
|| !(TCP_SKB_CB(skb
)->flags
&TCPCB_FLAG_SYN
)) {
1699 printk(KERN_DEBUG
"tcp_send_synack: wrong queue state\n");
1702 if (!(TCP_SKB_CB(skb
)->flags
&TCPCB_FLAG_ACK
)) {
1703 if (skb_cloned(skb
)) {
1704 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
1707 __skb_unlink(skb
, &sk
->sk_write_queue
);
1708 skb_header_release(nskb
);
1709 __skb_queue_head(&sk
->sk_write_queue
, nskb
);
1710 sk_stream_free_skb(sk
, skb
);
1711 sk_charge_skb(sk
, nskb
);
1715 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_ACK
;
1716 TCP_ECN_send_synack(tcp_sk(sk
), skb
);
1718 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1719 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
1723 * Prepare a SYN-ACK.
1725 struct sk_buff
* tcp_make_synack(struct sock
*sk
, struct dst_entry
*dst
,
1726 struct request_sock
*req
)
1728 struct inet_request_sock
*ireq
= inet_rsk(req
);
1729 struct tcp_sock
*tp
= tcp_sk(sk
);
1731 int tcp_header_size
;
1732 struct sk_buff
*skb
;
1734 skb
= sock_wmalloc(sk
, MAX_TCP_HEADER
+ 15, 1, GFP_ATOMIC
);
1738 /* Reserve space for headers. */
1739 skb_reserve(skb
, MAX_TCP_HEADER
);
1741 skb
->dst
= dst_clone(dst
);
1743 tcp_header_size
= (sizeof(struct tcphdr
) + TCPOLEN_MSS
+
1744 (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0) +
1745 (ireq
->wscale_ok
? TCPOLEN_WSCALE_ALIGNED
: 0) +
1746 /* SACK_PERM is in the place of NOP NOP of TS */
1747 ((ireq
->sack_ok
&& !ireq
->tstamp_ok
) ? TCPOLEN_SACKPERM_ALIGNED
: 0));
1748 skb
->h
.th
= th
= (struct tcphdr
*) skb_push(skb
, tcp_header_size
);
1750 memset(th
, 0, sizeof(struct tcphdr
));
1753 if (dst
->dev
->features
&NETIF_F_TSO
)
1755 TCP_ECN_make_synack(req
, th
);
1756 th
->source
= inet_sk(sk
)->sport
;
1757 th
->dest
= ireq
->rmt_port
;
1758 TCP_SKB_CB(skb
)->seq
= tcp_rsk(req
)->snt_isn
;
1759 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ 1;
1760 TCP_SKB_CB(skb
)->sacked
= 0;
1761 skb_shinfo(skb
)->tso_segs
= 1;
1762 skb_shinfo(skb
)->tso_size
= 0;
1763 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
1764 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_isn
+ 1);
1765 if (req
->rcv_wnd
== 0) { /* ignored for retransmitted syns */
1767 /* Set this up on the first call only */
1768 req
->window_clamp
= tp
->window_clamp
? : dst_metric(dst
, RTAX_WINDOW
);
1769 /* tcp_full_space because it is guaranteed to be the first packet */
1770 tcp_select_initial_window(tcp_full_space(sk
),
1771 dst_metric(dst
, RTAX_ADVMSS
) - (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0),
1776 ireq
->rcv_wscale
= rcv_wscale
;
1779 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
1780 th
->window
= htons(req
->rcv_wnd
);
1782 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1783 tcp_syn_build_options((__u32
*)(th
+ 1), dst_metric(dst
, RTAX_ADVMSS
), ireq
->tstamp_ok
,
1784 ireq
->sack_ok
, ireq
->wscale_ok
, ireq
->rcv_wscale
,
1785 TCP_SKB_CB(skb
)->when
,
1789 th
->doff
= (tcp_header_size
>> 2);
1790 TCP_INC_STATS(TCP_MIB_OUTSEGS
);
1795 * Do all connect socket setups that can be done AF independent.
1797 static inline void tcp_connect_init(struct sock
*sk
)
1799 struct dst_entry
*dst
= __sk_dst_get(sk
);
1800 struct tcp_sock
*tp
= tcp_sk(sk
);
1803 /* We'll fix this up when we get a response from the other end.
1804 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
1806 tp
->tcp_header_len
= sizeof(struct tcphdr
) +
1807 (sysctl_tcp_timestamps
? TCPOLEN_TSTAMP_ALIGNED
: 0);
1809 /* If user gave his TCP_MAXSEG, record it to clamp */
1810 if (tp
->rx_opt
.user_mss
)
1811 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
1813 tcp_sync_mss(sk
, dst_mtu(dst
));
1815 if (!tp
->window_clamp
)
1816 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
1817 tp
->advmss
= dst_metric(dst
, RTAX_ADVMSS
);
1818 tcp_initialize_rcv_mss(sk
);
1820 tcp_select_initial_window(tcp_full_space(sk
),
1821 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
1824 sysctl_tcp_window_scaling
,
1827 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
1828 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
1831 sock_reset_flag(sk
, SOCK_DONE
);
1833 tcp_init_wl(tp
, tp
->write_seq
, 0);
1834 tp
->snd_una
= tp
->write_seq
;
1835 tp
->snd_sml
= tp
->write_seq
;
1840 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
1841 inet_csk(sk
)->icsk_retransmits
= 0;
1842 tcp_clear_retrans(tp
);
1846 * Build a SYN and send it off.
1848 int tcp_connect(struct sock
*sk
)
1850 struct tcp_sock
*tp
= tcp_sk(sk
);
1851 struct sk_buff
*buff
;
1853 tcp_connect_init(sk
);
1855 buff
= alloc_skb_fclone(MAX_TCP_HEADER
+ 15, sk
->sk_allocation
);
1856 if (unlikely(buff
== NULL
))
1859 /* Reserve space for headers. */
1860 skb_reserve(buff
, MAX_TCP_HEADER
);
1862 TCP_SKB_CB(buff
)->flags
= TCPCB_FLAG_SYN
;
1863 TCP_ECN_send_syn(sk
, tp
, buff
);
1864 TCP_SKB_CB(buff
)->sacked
= 0;
1865 skb_shinfo(buff
)->tso_segs
= 1;
1866 skb_shinfo(buff
)->tso_size
= 0;
1868 TCP_SKB_CB(buff
)->seq
= tp
->write_seq
++;
1869 TCP_SKB_CB(buff
)->end_seq
= tp
->write_seq
;
1870 tp
->snd_nxt
= tp
->write_seq
;
1871 tp
->pushed_seq
= tp
->write_seq
;
1874 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
1875 tp
->retrans_stamp
= TCP_SKB_CB(buff
)->when
;
1876 skb_header_release(buff
);
1877 __skb_queue_tail(&sk
->sk_write_queue
, buff
);
1878 sk_charge_skb(sk
, buff
);
1879 tp
->packets_out
+= tcp_skb_pcount(buff
);
1880 tcp_transmit_skb(sk
, buff
, 1, GFP_KERNEL
);
1881 TCP_INC_STATS(TCP_MIB_ACTIVEOPENS
);
1883 /* Timer for repeating the SYN until an answer. */
1884 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1885 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
1889 /* Send out a delayed ack, the caller does the policy checking
1890 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
1893 void tcp_send_delayed_ack(struct sock
*sk
)
1895 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1896 int ato
= icsk
->icsk_ack
.ato
;
1897 unsigned long timeout
;
1899 if (ato
> TCP_DELACK_MIN
) {
1900 const struct tcp_sock
*tp
= tcp_sk(sk
);
1903 if (icsk
->icsk_ack
.pingpong
|| (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
1904 max_ato
= TCP_DELACK_MAX
;
1906 /* Slow path, intersegment interval is "high". */
1908 /* If some rtt estimate is known, use it to bound delayed ack.
1909 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
1913 int rtt
= max(tp
->srtt
>>3, TCP_DELACK_MIN
);
1919 ato
= min(ato
, max_ato
);
1922 /* Stay within the limit we were given */
1923 timeout
= jiffies
+ ato
;
1925 /* Use new timeout only if there wasn't a older one earlier. */
1926 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
1927 /* If delack timer was blocked or is about to expire,
1930 if (icsk
->icsk_ack
.blocked
||
1931 time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
1936 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
1937 timeout
= icsk
->icsk_ack
.timeout
;
1939 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
1940 icsk
->icsk_ack
.timeout
= timeout
;
1941 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
1944 /* This routine sends an ack and also updates the window. */
1945 void tcp_send_ack(struct sock
*sk
)
1947 /* If we have been reset, we may not send again. */
1948 if (sk
->sk_state
!= TCP_CLOSE
) {
1949 struct tcp_sock
*tp
= tcp_sk(sk
);
1950 struct sk_buff
*buff
;
1952 /* We are not putting this on the write queue, so
1953 * tcp_transmit_skb() will set the ownership to this
1956 buff
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
1958 inet_csk_schedule_ack(sk
);
1959 inet_csk(sk
)->icsk_ack
.ato
= TCP_ATO_MIN
;
1960 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
1961 TCP_DELACK_MAX
, TCP_RTO_MAX
);
1965 /* Reserve space for headers and prepare control bits. */
1966 skb_reserve(buff
, MAX_TCP_HEADER
);
1968 TCP_SKB_CB(buff
)->flags
= TCPCB_FLAG_ACK
;
1969 TCP_SKB_CB(buff
)->sacked
= 0;
1970 skb_shinfo(buff
)->tso_segs
= 1;
1971 skb_shinfo(buff
)->tso_size
= 0;
1973 /* Send it off, this clears delayed acks for us. */
1974 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(buff
)->end_seq
= tcp_acceptable_seq(sk
, tp
);
1975 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
1976 tcp_transmit_skb(sk
, buff
, 0, GFP_ATOMIC
);
1980 /* This routine sends a packet with an out of date sequence
1981 * number. It assumes the other end will try to ack it.
1983 * Question: what should we make while urgent mode?
1984 * 4.4BSD forces sending single byte of data. We cannot send
1985 * out of window data, because we have SND.NXT==SND.MAX...
1987 * Current solution: to send TWO zero-length segments in urgent mode:
1988 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
1989 * out-of-date with SND.UNA-1 to probe window.
1991 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
)
1993 struct tcp_sock
*tp
= tcp_sk(sk
);
1994 struct sk_buff
*skb
;
1996 /* We don't queue it, tcp_transmit_skb() sets ownership. */
1997 skb
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2001 /* Reserve space for headers and set control bits. */
2002 skb_reserve(skb
, MAX_TCP_HEADER
);
2004 TCP_SKB_CB(skb
)->flags
= TCPCB_FLAG_ACK
;
2005 TCP_SKB_CB(skb
)->sacked
= urgent
;
2006 skb_shinfo(skb
)->tso_segs
= 1;
2007 skb_shinfo(skb
)->tso_size
= 0;
2009 /* Use a previous sequence. This should cause the other
2010 * end to send an ack. Don't queue or clone SKB, just
2013 TCP_SKB_CB(skb
)->seq
= urgent
? tp
->snd_una
: tp
->snd_una
- 1;
2014 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
;
2015 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2016 return tcp_transmit_skb(sk
, skb
, 0, GFP_ATOMIC
);
2019 int tcp_write_wakeup(struct sock
*sk
)
2021 if (sk
->sk_state
!= TCP_CLOSE
) {
2022 struct tcp_sock
*tp
= tcp_sk(sk
);
2023 struct sk_buff
*skb
;
2025 if ((skb
= sk
->sk_send_head
) != NULL
&&
2026 before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
+tp
->snd_wnd
)) {
2028 unsigned int mss
= tcp_current_mss(sk
, 0);
2029 unsigned int seg_size
= tp
->snd_una
+tp
->snd_wnd
-TCP_SKB_CB(skb
)->seq
;
2031 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
2032 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
2034 /* We are probing the opening of a window
2035 * but the window size is != 0
2036 * must have been a result SWS avoidance ( sender )
2038 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
2040 seg_size
= min(seg_size
, mss
);
2041 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2042 if (tcp_fragment(sk
, skb
, seg_size
, mss
))
2044 } else if (!tcp_skb_pcount(skb
))
2045 tcp_set_skb_tso_segs(sk
, skb
, mss
);
2047 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2048 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2049 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2051 update_send_head(sk
, tp
, skb
);
2056 between(tp
->snd_up
, tp
->snd_una
+1, tp
->snd_una
+0xFFFF))
2057 tcp_xmit_probe_skb(sk
, TCPCB_URG
);
2058 return tcp_xmit_probe_skb(sk
, 0);
2064 /* A window probe timeout has occurred. If window is not closed send
2065 * a partial packet else a zero probe.
2067 void tcp_send_probe0(struct sock
*sk
)
2069 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2070 struct tcp_sock
*tp
= tcp_sk(sk
);
2073 err
= tcp_write_wakeup(sk
);
2075 if (tp
->packets_out
|| !sk
->sk_send_head
) {
2076 /* Cancel probe timer, if it is not required. */
2077 icsk
->icsk_probes_out
= 0;
2078 icsk
->icsk_backoff
= 0;
2083 if (icsk
->icsk_backoff
< sysctl_tcp_retries2
)
2084 icsk
->icsk_backoff
++;
2085 icsk
->icsk_probes_out
++;
2086 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2087 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
2090 /* If packet was not sent due to local congestion,
2091 * do not backoff and do not remember icsk_probes_out.
2092 * Let local senders to fight for local resources.
2094 * Use accumulated backoff yet.
2096 if (!icsk
->icsk_probes_out
)
2097 icsk
->icsk_probes_out
= 1;
2098 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2099 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
,
2100 TCP_RESOURCE_PROBE_INTERVAL
),
2105 EXPORT_SYMBOL(tcp_connect
);
2106 EXPORT_SYMBOL(tcp_make_synack
);
2107 EXPORT_SYMBOL(tcp_simple_retransmit
);
2108 EXPORT_SYMBOL(tcp_sync_mss
);
2109 EXPORT_SYMBOL(sysctl_tcp_tso_win_divisor
);