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
)
268 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
269 struct inet_sock
*inet
= inet_sk(sk
);
270 struct tcp_sock
*tp
= tcp_sk(sk
);
271 struct tcp_skb_cb
*tcb
= TCP_SKB_CB(skb
);
272 int tcp_header_size
= tp
->tcp_header_len
;
277 BUG_ON(!tcp_skb_pcount(skb
));
279 #define SYSCTL_FLAG_TSTAMPS 0x1
280 #define SYSCTL_FLAG_WSCALE 0x2
281 #define SYSCTL_FLAG_SACK 0x4
283 /* If congestion control is doing timestamping */
284 if (icsk
->icsk_ca_ops
->rtt_sample
)
285 __net_timestamp(skb
);
288 if (tcb
->flags
& TCPCB_FLAG_SYN
) {
289 tcp_header_size
= sizeof(struct tcphdr
) + TCPOLEN_MSS
;
290 if(sysctl_tcp_timestamps
) {
291 tcp_header_size
+= TCPOLEN_TSTAMP_ALIGNED
;
292 sysctl_flags
|= SYSCTL_FLAG_TSTAMPS
;
294 if(sysctl_tcp_window_scaling
) {
295 tcp_header_size
+= TCPOLEN_WSCALE_ALIGNED
;
296 sysctl_flags
|= SYSCTL_FLAG_WSCALE
;
298 if(sysctl_tcp_sack
) {
299 sysctl_flags
|= SYSCTL_FLAG_SACK
;
300 if(!(sysctl_flags
& SYSCTL_FLAG_TSTAMPS
))
301 tcp_header_size
+= TCPOLEN_SACKPERM_ALIGNED
;
303 } else if (tp
->rx_opt
.eff_sacks
) {
304 /* A SACK is 2 pad bytes, a 2 byte header, plus
305 * 2 32-bit sequence numbers for each SACK block.
307 tcp_header_size
+= (TCPOLEN_SACK_BASE_ALIGNED
+
308 (tp
->rx_opt
.eff_sacks
* TCPOLEN_SACK_PERBLOCK
));
311 if (tcp_packets_in_flight(tp
) == 0)
312 tcp_ca_event(sk
, CA_EVENT_TX_START
);
314 th
= (struct tcphdr
*) skb_push(skb
, tcp_header_size
);
316 skb_set_owner_w(skb
, sk
);
318 /* Build TCP header and checksum it. */
319 th
->source
= inet
->sport
;
320 th
->dest
= inet
->dport
;
321 th
->seq
= htonl(tcb
->seq
);
322 th
->ack_seq
= htonl(tp
->rcv_nxt
);
323 *(((__u16
*)th
) + 6) = htons(((tcp_header_size
>> 2) << 12) | tcb
->flags
);
324 if (tcb
->flags
& TCPCB_FLAG_SYN
) {
325 /* RFC1323: The window in SYN & SYN/ACK segments
328 th
->window
= htons(tp
->rcv_wnd
);
330 th
->window
= htons(tcp_select_window(sk
));
336 between(tp
->snd_up
, tcb
->seq
+1, tcb
->seq
+0xFFFF)) {
337 th
->urg_ptr
= htons(tp
->snd_up
-tcb
->seq
);
341 if (tcb
->flags
& TCPCB_FLAG_SYN
) {
342 tcp_syn_build_options((__u32
*)(th
+ 1),
343 tcp_advertise_mss(sk
),
344 (sysctl_flags
& SYSCTL_FLAG_TSTAMPS
),
345 (sysctl_flags
& SYSCTL_FLAG_SACK
),
346 (sysctl_flags
& SYSCTL_FLAG_WSCALE
),
347 tp
->rx_opt
.rcv_wscale
,
349 tp
->rx_opt
.ts_recent
);
351 tcp_build_and_update_options((__u32
*)(th
+ 1),
354 TCP_ECN_send(sk
, tp
, skb
, tcp_header_size
);
356 tp
->af_specific
->send_check(sk
, th
, skb
->len
, skb
);
358 if (tcb
->flags
& TCPCB_FLAG_ACK
)
359 tcp_event_ack_sent(sk
, tcp_skb_pcount(skb
));
361 if (skb
->len
!= tcp_header_size
)
362 tcp_event_data_sent(tp
, skb
, sk
);
364 TCP_INC_STATS(TCP_MIB_OUTSEGS
);
366 err
= tp
->af_specific
->queue_xmit(skb
, 0);
372 /* NET_XMIT_CN is special. It does not guarantee,
373 * that this packet is lost. It tells that device
374 * is about to start to drop packets or already
375 * drops some packets of the same priority and
376 * invokes us to send less aggressively.
378 return err
== NET_XMIT_CN
? 0 : err
;
381 #undef SYSCTL_FLAG_TSTAMPS
382 #undef SYSCTL_FLAG_WSCALE
383 #undef SYSCTL_FLAG_SACK
387 /* This routine just queue's the buffer
389 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
390 * otherwise socket can stall.
392 static void tcp_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
394 struct tcp_sock
*tp
= tcp_sk(sk
);
396 /* Advance write_seq and place onto the write_queue. */
397 tp
->write_seq
= TCP_SKB_CB(skb
)->end_seq
;
398 skb_header_release(skb
);
399 __skb_queue_tail(&sk
->sk_write_queue
, skb
);
400 sk_charge_skb(sk
, skb
);
402 /* Queue it, remembering where we must start sending. */
403 if (sk
->sk_send_head
== NULL
)
404 sk
->sk_send_head
= skb
;
407 static void tcp_set_skb_tso_segs(struct sock
*sk
, struct sk_buff
*skb
, unsigned int mss_now
)
409 if (skb
->len
<= mss_now
||
410 !(sk
->sk_route_caps
& NETIF_F_TSO
)) {
411 /* Avoid the costly divide in the normal
414 skb_shinfo(skb
)->tso_segs
= 1;
415 skb_shinfo(skb
)->tso_size
= 0;
419 factor
= skb
->len
+ (mss_now
- 1);
421 skb_shinfo(skb
)->tso_segs
= factor
;
422 skb_shinfo(skb
)->tso_size
= mss_now
;
426 /* Function to create two new TCP segments. Shrinks the given segment
427 * to the specified size and appends a new segment with the rest of the
428 * packet to the list. This won't be called frequently, I hope.
429 * Remember, these are still headerless SKBs at this point.
431 int tcp_fragment(struct sock
*sk
, struct sk_buff
*skb
, u32 len
, unsigned int mss_now
)
433 struct tcp_sock
*tp
= tcp_sk(sk
);
434 struct sk_buff
*buff
;
435 int nsize
, old_factor
;
438 if (unlikely(len
>= skb
->len
)) {
439 if (net_ratelimit()) {
440 printk(KERN_DEBUG
"TCP: seg_size=%u, mss=%u, seq=%u, "
441 "end_seq=%u, skb->len=%u.\n", len
, mss_now
,
442 TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
,
449 nsize
= skb_headlen(skb
) - len
;
453 if (skb_cloned(skb
) &&
454 skb_is_nonlinear(skb
) &&
455 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
458 /* Get a new skb... force flag on. */
459 buff
= sk_stream_alloc_skb(sk
, nsize
, GFP_ATOMIC
);
461 return -ENOMEM
; /* We'll just try again later. */
462 sk_charge_skb(sk
, buff
);
464 /* Correct the sequence numbers. */
465 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
466 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
467 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
469 /* PSH and FIN should only be set in the second packet. */
470 flags
= TCP_SKB_CB(skb
)->flags
;
471 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
472 TCP_SKB_CB(buff
)->flags
= flags
;
473 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
474 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_AT_TAIL
;
476 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_HW
) {
477 /* Copy and checksum data tail into the new buffer. */
478 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
, skb_put(buff
, nsize
),
483 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
485 skb
->ip_summed
= CHECKSUM_HW
;
486 skb_split(skb
, buff
, len
);
489 buff
->ip_summed
= skb
->ip_summed
;
491 /* Looks stupid, but our code really uses when of
492 * skbs, which it never sent before. --ANK
494 TCP_SKB_CB(buff
)->when
= TCP_SKB_CB(skb
)->when
;
495 buff
->tstamp
= skb
->tstamp
;
497 old_factor
= tcp_skb_pcount(skb
);
499 /* Fix up tso_factor for both original and new SKB. */
500 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
501 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
503 /* If this packet has been sent out already, we must
504 * adjust the various packet counters.
506 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
507 int diff
= old_factor
- tcp_skb_pcount(skb
) -
508 tcp_skb_pcount(buff
);
510 tp
->packets_out
-= diff
;
512 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
513 tp
->sacked_out
-= diff
;
514 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
515 tp
->retrans_out
-= diff
;
517 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
) {
518 tp
->lost_out
-= diff
;
519 tp
->left_out
-= diff
;
523 /* Adjust Reno SACK estimate. */
524 if (!tp
->rx_opt
.sack_ok
) {
525 tp
->sacked_out
-= diff
;
526 if ((int)tp
->sacked_out
< 0)
528 tcp_sync_left_out(tp
);
531 tp
->fackets_out
-= diff
;
532 if ((int)tp
->fackets_out
< 0)
537 /* Link BUFF into the send queue. */
538 skb_header_release(buff
);
539 __skb_append(skb
, buff
, &sk
->sk_write_queue
);
544 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
545 * eventually). The difference is that pulled data not copied, but
546 * immediately discarded.
548 static unsigned char *__pskb_trim_head(struct sk_buff
*skb
, int len
)
554 for (i
=0; i
<skb_shinfo(skb
)->nr_frags
; i
++) {
555 if (skb_shinfo(skb
)->frags
[i
].size
<= eat
) {
556 put_page(skb_shinfo(skb
)->frags
[i
].page
);
557 eat
-= skb_shinfo(skb
)->frags
[i
].size
;
559 skb_shinfo(skb
)->frags
[k
] = skb_shinfo(skb
)->frags
[i
];
561 skb_shinfo(skb
)->frags
[k
].page_offset
+= eat
;
562 skb_shinfo(skb
)->frags
[k
].size
-= eat
;
568 skb_shinfo(skb
)->nr_frags
= k
;
570 skb
->tail
= skb
->data
;
571 skb
->data_len
-= len
;
572 skb
->len
= skb
->data_len
;
576 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
578 if (skb_cloned(skb
) &&
579 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
582 if (len
<= skb_headlen(skb
)) {
583 __skb_pull(skb
, len
);
585 if (__pskb_trim_head(skb
, len
-skb_headlen(skb
)) == NULL
)
589 TCP_SKB_CB(skb
)->seq
+= len
;
590 skb
->ip_summed
= CHECKSUM_HW
;
592 skb
->truesize
-= len
;
593 sk
->sk_wmem_queued
-= len
;
594 sk
->sk_forward_alloc
+= len
;
595 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
597 /* Any change of skb->len requires recalculation of tso
600 if (tcp_skb_pcount(skb
) > 1)
601 tcp_set_skb_tso_segs(sk
, skb
, tcp_current_mss(sk
, 1));
606 /* This function synchronize snd mss to current pmtu/exthdr set.
608 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
609 for TCP options, but includes only bare TCP header.
611 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
612 It is minumum of user_mss and mss received with SYN.
613 It also does not include TCP options.
615 tp->pmtu_cookie is last pmtu, seen by this function.
617 tp->mss_cache is current effective sending mss, including
618 all tcp options except for SACKs. It is evaluated,
619 taking into account current pmtu, but never exceeds
620 tp->rx_opt.mss_clamp.
622 NOTE1. rfc1122 clearly states that advertised MSS
623 DOES NOT include either tcp or ip options.
625 NOTE2. tp->pmtu_cookie and tp->mss_cache are READ ONLY outside
626 this function. --ANK (980731)
629 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
631 struct tcp_sock
*tp
= tcp_sk(sk
);
634 /* Calculate base mss without TCP options:
635 It is MMS_S - sizeof(tcphdr) of rfc1122
637 mss_now
= pmtu
- tp
->af_specific
->net_header_len
- sizeof(struct tcphdr
);
639 /* Clamp it (mss_clamp does not include tcp options) */
640 if (mss_now
> tp
->rx_opt
.mss_clamp
)
641 mss_now
= tp
->rx_opt
.mss_clamp
;
643 /* Now subtract optional transport overhead */
644 mss_now
-= tp
->ext_header_len
;
646 /* Then reserve room for full set of TCP options and 8 bytes of data */
650 /* Now subtract TCP options size, not including SACKs */
651 mss_now
-= tp
->tcp_header_len
- sizeof(struct tcphdr
);
653 /* Bound mss with half of window */
654 if (tp
->max_window
&& mss_now
> (tp
->max_window
>>1))
655 mss_now
= max((tp
->max_window
>>1), 68U - tp
->tcp_header_len
);
657 /* And store cached results */
658 tp
->pmtu_cookie
= pmtu
;
659 tp
->mss_cache
= mss_now
;
664 /* Compute the current effective MSS, taking SACKs and IP options,
665 * and even PMTU discovery events into account.
667 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
668 * cannot be large. However, taking into account rare use of URG, this
671 unsigned int tcp_current_mss(struct sock
*sk
, int large_allowed
)
673 struct tcp_sock
*tp
= tcp_sk(sk
);
674 struct dst_entry
*dst
= __sk_dst_get(sk
);
679 mss_now
= tp
->mss_cache
;
682 (sk
->sk_route_caps
& NETIF_F_TSO
) &&
687 u32 mtu
= dst_mtu(dst
);
688 if (mtu
!= tp
->pmtu_cookie
)
689 mss_now
= tcp_sync_mss(sk
, mtu
);
692 if (tp
->rx_opt
.eff_sacks
)
693 mss_now
-= (TCPOLEN_SACK_BASE_ALIGNED
+
694 (tp
->rx_opt
.eff_sacks
* TCPOLEN_SACK_PERBLOCK
));
696 xmit_size_goal
= mss_now
;
699 xmit_size_goal
= 65535 -
700 tp
->af_specific
->net_header_len
-
701 tp
->ext_header_len
- tp
->tcp_header_len
;
703 if (tp
->max_window
&&
704 (xmit_size_goal
> (tp
->max_window
>> 1)))
705 xmit_size_goal
= max((tp
->max_window
>> 1),
706 68U - tp
->tcp_header_len
);
708 xmit_size_goal
-= (xmit_size_goal
% mss_now
);
710 tp
->xmit_size_goal
= xmit_size_goal
;
715 /* Congestion window validation. (RFC2861) */
717 static inline void tcp_cwnd_validate(struct sock
*sk
, struct tcp_sock
*tp
)
719 __u32 packets_out
= tp
->packets_out
;
721 if (packets_out
>= tp
->snd_cwnd
) {
722 /* Network is feed fully. */
723 tp
->snd_cwnd_used
= 0;
724 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
726 /* Network starves. */
727 if (tp
->packets_out
> tp
->snd_cwnd_used
)
728 tp
->snd_cwnd_used
= tp
->packets_out
;
730 if ((s32
)(tcp_time_stamp
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
)
731 tcp_cwnd_application_limited(sk
);
735 static unsigned int tcp_window_allows(struct tcp_sock
*tp
, struct sk_buff
*skb
, unsigned int mss_now
, unsigned int cwnd
)
737 u32 window
, cwnd_len
;
739 window
= (tp
->snd_una
+ tp
->snd_wnd
- TCP_SKB_CB(skb
)->seq
);
740 cwnd_len
= mss_now
* cwnd
;
741 return min(window
, cwnd_len
);
744 /* Can at least one segment of SKB be sent right now, according to the
745 * congestion window rules? If so, return how many segments are allowed.
747 static inline unsigned int tcp_cwnd_test(struct tcp_sock
*tp
, struct sk_buff
*skb
)
751 /* Don't be strict about the congestion window for the final FIN. */
752 if (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
)
755 in_flight
= tcp_packets_in_flight(tp
);
757 if (in_flight
< cwnd
)
758 return (cwnd
- in_flight
);
763 /* This must be invoked the first time we consider transmitting
766 static inline int tcp_init_tso_segs(struct sock
*sk
, struct sk_buff
*skb
, unsigned int mss_now
)
768 int tso_segs
= tcp_skb_pcount(skb
);
772 skb_shinfo(skb
)->tso_size
!= mss_now
)) {
773 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
774 tso_segs
= tcp_skb_pcount(skb
);
779 static inline int tcp_minshall_check(const struct tcp_sock
*tp
)
781 return after(tp
->snd_sml
,tp
->snd_una
) &&
782 !after(tp
->snd_sml
, tp
->snd_nxt
);
785 /* Return 0, if packet can be sent now without violation Nagle's rules:
786 * 1. It is full sized.
787 * 2. Or it contains FIN. (already checked by caller)
788 * 3. Or TCP_NODELAY was set.
789 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
790 * With Minshall's modification: all sent small packets are ACKed.
793 static inline int tcp_nagle_check(const struct tcp_sock
*tp
,
794 const struct sk_buff
*skb
,
795 unsigned mss_now
, int nonagle
)
797 return (skb
->len
< mss_now
&&
798 ((nonagle
&TCP_NAGLE_CORK
) ||
801 tcp_minshall_check(tp
))));
804 /* Return non-zero if the Nagle test allows this packet to be
807 static inline int tcp_nagle_test(struct tcp_sock
*tp
, struct sk_buff
*skb
,
808 unsigned int cur_mss
, int nonagle
)
810 /* Nagle rule does not apply to frames, which sit in the middle of the
811 * write_queue (they have no chances to get new data).
813 * This is implemented in the callers, where they modify the 'nonagle'
814 * argument based upon the location of SKB in the send queue.
816 if (nonagle
& TCP_NAGLE_PUSH
)
819 /* Don't use the nagle rule for urgent data (or for the final FIN). */
821 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
))
824 if (!tcp_nagle_check(tp
, skb
, cur_mss
, nonagle
))
830 /* Does at least the first segment of SKB fit into the send window? */
831 static inline int tcp_snd_wnd_test(struct tcp_sock
*tp
, struct sk_buff
*skb
, unsigned int cur_mss
)
833 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
835 if (skb
->len
> cur_mss
)
836 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
838 return !after(end_seq
, tp
->snd_una
+ tp
->snd_wnd
);
841 /* This checks if the data bearing packet SKB (usually sk->sk_send_head)
842 * should be put on the wire right now. If so, it returns the number of
843 * packets allowed by the congestion window.
845 static unsigned int tcp_snd_test(struct sock
*sk
, struct sk_buff
*skb
,
846 unsigned int cur_mss
, int nonagle
)
848 struct tcp_sock
*tp
= tcp_sk(sk
);
849 unsigned int cwnd_quota
;
851 tcp_init_tso_segs(sk
, skb
, cur_mss
);
853 if (!tcp_nagle_test(tp
, skb
, cur_mss
, nonagle
))
856 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
858 !tcp_snd_wnd_test(tp
, skb
, cur_mss
))
864 static inline int tcp_skb_is_last(const struct sock
*sk
,
865 const struct sk_buff
*skb
)
867 return skb
->next
== (struct sk_buff
*)&sk
->sk_write_queue
;
870 int tcp_may_send_now(struct sock
*sk
, struct tcp_sock
*tp
)
872 struct sk_buff
*skb
= sk
->sk_send_head
;
875 tcp_snd_test(sk
, skb
, tcp_current_mss(sk
, 1),
876 (tcp_skb_is_last(sk
, skb
) ?
881 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
882 * which is put after SKB on the list. It is very much like
883 * tcp_fragment() except that it may make several kinds of assumptions
884 * in order to speed up the splitting operation. In particular, we
885 * know that all the data is in scatter-gather pages, and that the
886 * packet has never been sent out before (and thus is not cloned).
888 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
, unsigned int mss_now
)
890 struct sk_buff
*buff
;
891 int nlen
= skb
->len
- len
;
894 /* All of a TSO frame must be composed of paged data. */
895 if (skb
->len
!= skb
->data_len
)
896 return tcp_fragment(sk
, skb
, len
, mss_now
);
898 buff
= sk_stream_alloc_pskb(sk
, 0, 0, GFP_ATOMIC
);
899 if (unlikely(buff
== NULL
))
902 buff
->truesize
= nlen
;
903 skb
->truesize
-= nlen
;
905 /* Correct the sequence numbers. */
906 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
907 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
908 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
910 /* PSH and FIN should only be set in the second packet. */
911 flags
= TCP_SKB_CB(skb
)->flags
;
912 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
913 TCP_SKB_CB(buff
)->flags
= flags
;
915 /* This packet was never sent out yet, so no SACK bits. */
916 TCP_SKB_CB(buff
)->sacked
= 0;
918 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_HW
;
919 skb_split(skb
, buff
, len
);
921 /* Fix up tso_factor for both original and new SKB. */
922 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
923 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
925 /* Link BUFF into the send queue. */
926 skb_header_release(buff
);
927 __skb_append(skb
, buff
, &sk
->sk_write_queue
);
932 /* Try to defer sending, if possible, in order to minimize the amount
933 * of TSO splitting we do. View it as a kind of TSO Nagle test.
935 * This algorithm is from John Heffner.
937 static int tcp_tso_should_defer(struct sock
*sk
, struct tcp_sock
*tp
, struct sk_buff
*skb
)
939 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
940 u32 send_win
, cong_win
, limit
, in_flight
;
942 if (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
)
945 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
948 in_flight
= tcp_packets_in_flight(tp
);
950 BUG_ON(tcp_skb_pcount(skb
) <= 1 ||
951 (tp
->snd_cwnd
<= in_flight
));
953 send_win
= (tp
->snd_una
+ tp
->snd_wnd
) - TCP_SKB_CB(skb
)->seq
;
955 /* From in_flight test above, we know that cwnd > in_flight. */
956 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
958 limit
= min(send_win
, cong_win
);
960 if (sysctl_tcp_tso_win_divisor
) {
961 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
963 /* If at least some fraction of a window is available,
966 chunk
/= sysctl_tcp_tso_win_divisor
;
970 /* Different approach, try not to defer past a single
971 * ACK. Receiver should ACK every other full sized
972 * frame, so if we have space for more than 3 frames
975 if (limit
> tcp_max_burst(tp
) * tp
->mss_cache
)
979 /* Ok, it looks like it is advisable to defer. */
983 /* This routine writes packets to the network. It advances the
984 * send_head. This happens as incoming acks open up the remote
987 * Returns 1, if no segments are in flight and we have queued segments, but
988 * cannot send anything now because of SWS or another problem.
990 static int tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
)
992 struct tcp_sock
*tp
= tcp_sk(sk
);
994 unsigned int tso_segs
, sent_pkts
;
997 /* If we are closed, the bytes will have to remain here.
998 * In time closedown will finish, we empty the write queue and all
1001 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
1005 while ((skb
= sk
->sk_send_head
)) {
1008 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1011 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1015 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
)))
1018 if (tso_segs
== 1) {
1019 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
1020 (tcp_skb_is_last(sk
, skb
) ?
1021 nonagle
: TCP_NAGLE_PUSH
))))
1024 if (tcp_tso_should_defer(sk
, tp
, skb
))
1030 limit
= tcp_window_allows(tp
, skb
,
1031 mss_now
, cwnd_quota
);
1033 if (skb
->len
< limit
) {
1034 unsigned int trim
= skb
->len
% mss_now
;
1037 limit
= skb
->len
- trim
;
1041 if (skb
->len
> limit
&&
1042 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
)))
1045 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1047 if (unlikely(tcp_transmit_skb(sk
, skb_clone(skb
, GFP_ATOMIC
))))
1050 /* Advance the send_head. This one is sent out.
1051 * This call will increment packets_out.
1053 update_send_head(sk
, tp
, skb
);
1055 tcp_minshall_update(tp
, mss_now
, skb
);
1059 if (likely(sent_pkts
)) {
1060 tcp_cwnd_validate(sk
, tp
);
1063 return !tp
->packets_out
&& sk
->sk_send_head
;
1066 /* Push out any pending frames which were held back due to
1067 * TCP_CORK or attempt at coalescing tiny packets.
1068 * The socket must be locked by the caller.
1070 void __tcp_push_pending_frames(struct sock
*sk
, struct tcp_sock
*tp
,
1071 unsigned int cur_mss
, int nonagle
)
1073 struct sk_buff
*skb
= sk
->sk_send_head
;
1076 if (tcp_write_xmit(sk
, cur_mss
, nonagle
))
1077 tcp_check_probe_timer(sk
, tp
);
1081 /* Send _single_ skb sitting at the send head. This function requires
1082 * true push pending frames to setup probe timer etc.
1084 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
1086 struct tcp_sock
*tp
= tcp_sk(sk
);
1087 struct sk_buff
*skb
= sk
->sk_send_head
;
1088 unsigned int tso_segs
, cwnd_quota
;
1090 BUG_ON(!skb
|| skb
->len
< mss_now
);
1092 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1093 cwnd_quota
= tcp_snd_test(sk
, skb
, mss_now
, TCP_NAGLE_PUSH
);
1095 if (likely(cwnd_quota
)) {
1102 limit
= tcp_window_allows(tp
, skb
,
1103 mss_now
, cwnd_quota
);
1105 if (skb
->len
< limit
) {
1106 unsigned int trim
= skb
->len
% mss_now
;
1109 limit
= skb
->len
- trim
;
1113 if (skb
->len
> limit
&&
1114 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
)))
1117 /* Send it out now. */
1118 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1120 if (likely(!tcp_transmit_skb(sk
, skb_clone(skb
, sk
->sk_allocation
)))) {
1121 update_send_head(sk
, tp
, skb
);
1122 tcp_cwnd_validate(sk
, tp
);
1128 /* This function returns the amount that we can raise the
1129 * usable window based on the following constraints
1131 * 1. The window can never be shrunk once it is offered (RFC 793)
1132 * 2. We limit memory per socket
1135 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1136 * RECV.NEXT + RCV.WIN fixed until:
1137 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1139 * i.e. don't raise the right edge of the window until you can raise
1140 * it at least MSS bytes.
1142 * Unfortunately, the recommended algorithm breaks header prediction,
1143 * since header prediction assumes th->window stays fixed.
1145 * Strictly speaking, keeping th->window fixed violates the receiver
1146 * side SWS prevention criteria. The problem is that under this rule
1147 * a stream of single byte packets will cause the right side of the
1148 * window to always advance by a single byte.
1150 * Of course, if the sender implements sender side SWS prevention
1151 * then this will not be a problem.
1153 * BSD seems to make the following compromise:
1155 * If the free space is less than the 1/4 of the maximum
1156 * space available and the free space is less than 1/2 mss,
1157 * then set the window to 0.
1158 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1159 * Otherwise, just prevent the window from shrinking
1160 * and from being larger than the largest representable value.
1162 * This prevents incremental opening of the window in the regime
1163 * where TCP is limited by the speed of the reader side taking
1164 * data out of the TCP receive queue. It does nothing about
1165 * those cases where the window is constrained on the sender side
1166 * because the pipeline is full.
1168 * BSD also seems to "accidentally" limit itself to windows that are a
1169 * multiple of MSS, at least until the free space gets quite small.
1170 * This would appear to be a side effect of the mbuf implementation.
1171 * Combining these two algorithms results in the observed behavior
1172 * of having a fixed window size at almost all times.
1174 * Below we obtain similar behavior by forcing the offered window to
1175 * a multiple of the mss when it is feasible to do so.
1177 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1178 * Regular options like TIMESTAMP are taken into account.
1180 u32
__tcp_select_window(struct sock
*sk
)
1182 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1183 struct tcp_sock
*tp
= tcp_sk(sk
);
1184 /* MSS for the peer's data. Previous verions used mss_clamp
1185 * here. I don't know if the value based on our guesses
1186 * of peer's MSS is better for the performance. It's more correct
1187 * but may be worse for the performance because of rcv_mss
1188 * fluctuations. --SAW 1998/11/1
1190 int mss
= icsk
->icsk_ack
.rcv_mss
;
1191 int free_space
= tcp_space(sk
);
1192 int full_space
= min_t(int, tp
->window_clamp
, tcp_full_space(sk
));
1195 if (mss
> full_space
)
1198 if (free_space
< full_space
/2) {
1199 icsk
->icsk_ack
.quick
= 0;
1201 if (tcp_memory_pressure
)
1202 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U*tp
->advmss
);
1204 if (free_space
< mss
)
1208 if (free_space
> tp
->rcv_ssthresh
)
1209 free_space
= tp
->rcv_ssthresh
;
1211 /* Don't do rounding if we are using window scaling, since the
1212 * scaled window will not line up with the MSS boundary anyway.
1214 window
= tp
->rcv_wnd
;
1215 if (tp
->rx_opt
.rcv_wscale
) {
1216 window
= free_space
;
1218 /* Advertise enough space so that it won't get scaled away.
1219 * Import case: prevent zero window announcement if
1220 * 1<<rcv_wscale > mss.
1222 if (((window
>> tp
->rx_opt
.rcv_wscale
) << tp
->rx_opt
.rcv_wscale
) != window
)
1223 window
= (((window
>> tp
->rx_opt
.rcv_wscale
) + 1)
1224 << tp
->rx_opt
.rcv_wscale
);
1226 /* Get the largest window that is a nice multiple of mss.
1227 * Window clamp already applied above.
1228 * If our current window offering is within 1 mss of the
1229 * free space we just keep it. This prevents the divide
1230 * and multiply from happening most of the time.
1231 * We also don't do any window rounding when the free space
1234 if (window
<= free_space
- mss
|| window
> free_space
)
1235 window
= (free_space
/mss
)*mss
;
1241 /* Attempt to collapse two adjacent SKB's during retransmission. */
1242 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*skb
, int mss_now
)
1244 struct tcp_sock
*tp
= tcp_sk(sk
);
1245 struct sk_buff
*next_skb
= skb
->next
;
1247 /* The first test we must make is that neither of these two
1248 * SKB's are still referenced by someone else.
1250 if (!skb_cloned(skb
) && !skb_cloned(next_skb
)) {
1251 int skb_size
= skb
->len
, next_skb_size
= next_skb
->len
;
1252 u16 flags
= TCP_SKB_CB(skb
)->flags
;
1254 /* Also punt if next skb has been SACK'd. */
1255 if(TCP_SKB_CB(next_skb
)->sacked
& TCPCB_SACKED_ACKED
)
1258 /* Next skb is out of window. */
1259 if (after(TCP_SKB_CB(next_skb
)->end_seq
, tp
->snd_una
+tp
->snd_wnd
))
1262 /* Punt if not enough space exists in the first SKB for
1263 * the data in the second, or the total combined payload
1264 * would exceed the MSS.
1266 if ((next_skb_size
> skb_tailroom(skb
)) ||
1267 ((skb_size
+ next_skb_size
) > mss_now
))
1270 BUG_ON(tcp_skb_pcount(skb
) != 1 ||
1271 tcp_skb_pcount(next_skb
) != 1);
1273 /* Ok. We will be able to collapse the packet. */
1274 __skb_unlink(next_skb
, &sk
->sk_write_queue
);
1276 memcpy(skb_put(skb
, next_skb_size
), next_skb
->data
, next_skb_size
);
1278 if (next_skb
->ip_summed
== CHECKSUM_HW
)
1279 skb
->ip_summed
= CHECKSUM_HW
;
1281 if (skb
->ip_summed
!= CHECKSUM_HW
)
1282 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
1284 /* Update sequence range on original skb. */
1285 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
1287 /* Merge over control information. */
1288 flags
|= TCP_SKB_CB(next_skb
)->flags
; /* This moves PSH/FIN etc. over */
1289 TCP_SKB_CB(skb
)->flags
= flags
;
1291 /* All done, get rid of second SKB and account for it so
1292 * packet counting does not break.
1294 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
&(TCPCB_EVER_RETRANS
|TCPCB_AT_TAIL
);
1295 if (TCP_SKB_CB(next_skb
)->sacked
&TCPCB_SACKED_RETRANS
)
1296 tp
->retrans_out
-= tcp_skb_pcount(next_skb
);
1297 if (TCP_SKB_CB(next_skb
)->sacked
&TCPCB_LOST
) {
1298 tp
->lost_out
-= tcp_skb_pcount(next_skb
);
1299 tp
->left_out
-= tcp_skb_pcount(next_skb
);
1301 /* Reno case is special. Sigh... */
1302 if (!tp
->rx_opt
.sack_ok
&& tp
->sacked_out
) {
1303 tcp_dec_pcount_approx(&tp
->sacked_out
, next_skb
);
1304 tp
->left_out
-= tcp_skb_pcount(next_skb
);
1307 /* Not quite right: it can be > snd.fack, but
1308 * it is better to underestimate fackets.
1310 tcp_dec_pcount_approx(&tp
->fackets_out
, next_skb
);
1311 tcp_packets_out_dec(tp
, next_skb
);
1312 sk_stream_free_skb(sk
, next_skb
);
1316 /* Do a simple retransmit without using the backoff mechanisms in
1317 * tcp_timer. This is used for path mtu discovery.
1318 * The socket is already locked here.
1320 void tcp_simple_retransmit(struct sock
*sk
)
1322 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1323 struct tcp_sock
*tp
= tcp_sk(sk
);
1324 struct sk_buff
*skb
;
1325 unsigned int mss
= tcp_current_mss(sk
, 0);
1328 sk_stream_for_retrans_queue(skb
, sk
) {
1329 if (skb
->len
> mss
&&
1330 !(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
)) {
1331 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) {
1332 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1333 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1335 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_LOST
)) {
1336 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1337 tp
->lost_out
+= tcp_skb_pcount(skb
);
1346 tcp_sync_left_out(tp
);
1348 /* Don't muck with the congestion window here.
1349 * Reason is that we do not increase amount of _data_
1350 * in network, but units changed and effective
1351 * cwnd/ssthresh really reduced now.
1353 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
1354 tp
->high_seq
= tp
->snd_nxt
;
1355 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
1356 tp
->prior_ssthresh
= 0;
1357 tp
->undo_marker
= 0;
1358 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1360 tcp_xmit_retransmit_queue(sk
);
1363 /* This retransmits one SKB. Policy decisions and retransmit queue
1364 * state updates are done by the caller. Returns non-zero if an
1365 * error occurred which prevented the send.
1367 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
1369 struct tcp_sock
*tp
= tcp_sk(sk
);
1370 unsigned int cur_mss
= tcp_current_mss(sk
, 0);
1373 /* Do not sent more than we queued. 1/4 is reserved for possible
1374 * copying overhead: frgagmentation, tunneling, mangling etc.
1376 if (atomic_read(&sk
->sk_wmem_alloc
) >
1377 min(sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2), sk
->sk_sndbuf
))
1380 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
1381 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1383 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
1387 /* If receiver has shrunk his window, and skb is out of
1388 * new window, do not retransmit it. The exception is the
1389 * case, when window is shrunk to zero. In this case
1390 * our retransmit serves as a zero window probe.
1392 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
+tp
->snd_wnd
)
1393 && TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
1396 if (skb
->len
> cur_mss
) {
1397 if (tcp_fragment(sk
, skb
, cur_mss
, cur_mss
))
1398 return -ENOMEM
; /* We'll try again later. */
1401 /* Collapse two adjacent packets if worthwhile and we can. */
1402 if(!(TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_SYN
) &&
1403 (skb
->len
< (cur_mss
>> 1)) &&
1404 (skb
->next
!= sk
->sk_send_head
) &&
1405 (skb
->next
!= (struct sk_buff
*)&sk
->sk_write_queue
) &&
1406 (skb_shinfo(skb
)->nr_frags
== 0 && skb_shinfo(skb
->next
)->nr_frags
== 0) &&
1407 (tcp_skb_pcount(skb
) == 1 && tcp_skb_pcount(skb
->next
) == 1) &&
1408 (sysctl_tcp_retrans_collapse
!= 0))
1409 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
1411 if(tp
->af_specific
->rebuild_header(sk
))
1412 return -EHOSTUNREACH
; /* Routing failure or similar. */
1414 /* Some Solaris stacks overoptimize and ignore the FIN on a
1415 * retransmit when old data is attached. So strip it off
1416 * since it is cheap to do so and saves bytes on the network.
1419 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
) &&
1420 tp
->snd_una
== (TCP_SKB_CB(skb
)->end_seq
- 1)) {
1421 if (!pskb_trim(skb
, 0)) {
1422 TCP_SKB_CB(skb
)->seq
= TCP_SKB_CB(skb
)->end_seq
- 1;
1423 skb_shinfo(skb
)->tso_segs
= 1;
1424 skb_shinfo(skb
)->tso_size
= 0;
1425 skb
->ip_summed
= CHECKSUM_NONE
;
1430 /* Make a copy, if the first transmission SKB clone we made
1431 * is still in somebody's hands, else make a clone.
1433 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1435 err
= tcp_transmit_skb(sk
, (skb_cloned(skb
) ?
1436 pskb_copy(skb
, GFP_ATOMIC
):
1437 skb_clone(skb
, GFP_ATOMIC
)));
1440 /* Update global TCP statistics. */
1441 TCP_INC_STATS(TCP_MIB_RETRANSSEGS
);
1443 tp
->total_retrans
++;
1445 #if FASTRETRANS_DEBUG > 0
1446 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) {
1447 if (net_ratelimit())
1448 printk(KERN_DEBUG
"retrans_out leaked.\n");
1451 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
1452 tp
->retrans_out
+= tcp_skb_pcount(skb
);
1454 /* Save stamp of the first retransmit. */
1455 if (!tp
->retrans_stamp
)
1456 tp
->retrans_stamp
= TCP_SKB_CB(skb
)->when
;
1460 /* snd_nxt is stored to detect loss of retransmitted segment,
1461 * see tcp_input.c tcp_sacktag_write_queue().
1463 TCP_SKB_CB(skb
)->ack_seq
= tp
->snd_nxt
;
1468 /* This gets called after a retransmit timeout, and the initially
1469 * retransmitted data is acknowledged. It tries to continue
1470 * resending the rest of the retransmit queue, until either
1471 * we've sent it all or the congestion window limit is reached.
1472 * If doing SACK, the first ACK which comes back for a timeout
1473 * based retransmit packet might feed us FACK information again.
1474 * If so, we use it to avoid unnecessarily retransmissions.
1476 void tcp_xmit_retransmit_queue(struct sock
*sk
)
1478 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1479 struct tcp_sock
*tp
= tcp_sk(sk
);
1480 struct sk_buff
*skb
;
1481 int packet_cnt
= tp
->lost_out
;
1483 /* First pass: retransmit lost packets. */
1485 sk_stream_for_retrans_queue(skb
, sk
) {
1486 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
1488 /* Assume this retransmit will generate
1489 * only one packet for congestion window
1490 * calculation purposes. This works because
1491 * tcp_retransmit_skb() will chop up the
1492 * packet to be MSS sized and all the
1493 * packet counting works out.
1495 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
1498 if (sacked
&TCPCB_LOST
) {
1499 if (!(sacked
&(TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))) {
1500 if (tcp_retransmit_skb(sk
, skb
))
1502 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
1503 NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS
);
1505 NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS
);
1508 skb_peek(&sk
->sk_write_queue
))
1509 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1510 inet_csk(sk
)->icsk_rto
,
1514 packet_cnt
-= tcp_skb_pcount(skb
);
1515 if (packet_cnt
<= 0)
1521 /* OK, demanded retransmission is finished. */
1523 /* Forward retransmissions are possible only during Recovery. */
1524 if (icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
1527 /* No forward retransmissions in Reno are possible. */
1528 if (!tp
->rx_opt
.sack_ok
)
1531 /* Yeah, we have to make difficult choice between forward transmission
1532 * and retransmission... Both ways have their merits...
1534 * For now we do not retransmit anything, while we have some new
1538 if (tcp_may_send_now(sk
, tp
))
1543 sk_stream_for_retrans_queue(skb
, sk
) {
1544 /* Similar to the retransmit loop above we
1545 * can pretend that the retransmitted SKB
1546 * we send out here will be composed of one
1547 * real MSS sized packet because tcp_retransmit_skb()
1548 * will fragment it if necessary.
1550 if (++packet_cnt
> tp
->fackets_out
)
1553 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
1556 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
)
1559 /* Ok, retransmit it. */
1560 if (tcp_retransmit_skb(sk
, skb
))
1563 if (skb
== skb_peek(&sk
->sk_write_queue
))
1564 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1565 inet_csk(sk
)->icsk_rto
,
1568 NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS
);
1573 /* Send a fin. The caller locks the socket for us. This cannot be
1574 * allowed to fail queueing a FIN frame under any circumstances.
1576 void tcp_send_fin(struct sock
*sk
)
1578 struct tcp_sock
*tp
= tcp_sk(sk
);
1579 struct sk_buff
*skb
= skb_peek_tail(&sk
->sk_write_queue
);
1582 /* Optimization, tack on the FIN if we have a queue of
1583 * unsent frames. But be careful about outgoing SACKS
1586 mss_now
= tcp_current_mss(sk
, 1);
1588 if (sk
->sk_send_head
!= NULL
) {
1589 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_FIN
;
1590 TCP_SKB_CB(skb
)->end_seq
++;
1593 /* Socket is locked, keep trying until memory is available. */
1595 skb
= alloc_skb_fclone(MAX_TCP_HEADER
, GFP_KERNEL
);
1601 /* Reserve space for headers and prepare control bits. */
1602 skb_reserve(skb
, MAX_TCP_HEADER
);
1604 TCP_SKB_CB(skb
)->flags
= (TCPCB_FLAG_ACK
| TCPCB_FLAG_FIN
);
1605 TCP_SKB_CB(skb
)->sacked
= 0;
1606 skb_shinfo(skb
)->tso_segs
= 1;
1607 skb_shinfo(skb
)->tso_size
= 0;
1609 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
1610 TCP_SKB_CB(skb
)->seq
= tp
->write_seq
;
1611 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ 1;
1612 tcp_queue_skb(sk
, skb
);
1614 __tcp_push_pending_frames(sk
, tp
, mss_now
, TCP_NAGLE_OFF
);
1617 /* We get here when a process closes a file descriptor (either due to
1618 * an explicit close() or as a byproduct of exit()'ing) and there
1619 * was unread data in the receive queue. This behavior is recommended
1620 * by draft-ietf-tcpimpl-prob-03.txt section 3.10. -DaveM
1622 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
1624 struct tcp_sock
*tp
= tcp_sk(sk
);
1625 struct sk_buff
*skb
;
1627 /* NOTE: No TCP options attached and we never retransmit this. */
1628 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
1630 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED
);
1634 /* Reserve space for headers and prepare control bits. */
1635 skb_reserve(skb
, MAX_TCP_HEADER
);
1637 TCP_SKB_CB(skb
)->flags
= (TCPCB_FLAG_ACK
| TCPCB_FLAG_RST
);
1638 TCP_SKB_CB(skb
)->sacked
= 0;
1639 skb_shinfo(skb
)->tso_segs
= 1;
1640 skb_shinfo(skb
)->tso_size
= 0;
1643 TCP_SKB_CB(skb
)->seq
= tcp_acceptable_seq(sk
, tp
);
1644 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
;
1645 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1646 if (tcp_transmit_skb(sk
, skb
))
1647 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED
);
1650 /* WARNING: This routine must only be called when we have already sent
1651 * a SYN packet that crossed the incoming SYN that caused this routine
1652 * to get called. If this assumption fails then the initial rcv_wnd
1653 * and rcv_wscale values will not be correct.
1655 int tcp_send_synack(struct sock
*sk
)
1657 struct sk_buff
* skb
;
1659 skb
= skb_peek(&sk
->sk_write_queue
);
1660 if (skb
== NULL
|| !(TCP_SKB_CB(skb
)->flags
&TCPCB_FLAG_SYN
)) {
1661 printk(KERN_DEBUG
"tcp_send_synack: wrong queue state\n");
1664 if (!(TCP_SKB_CB(skb
)->flags
&TCPCB_FLAG_ACK
)) {
1665 if (skb_cloned(skb
)) {
1666 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
1669 __skb_unlink(skb
, &sk
->sk_write_queue
);
1670 skb_header_release(nskb
);
1671 __skb_queue_head(&sk
->sk_write_queue
, nskb
);
1672 sk_stream_free_skb(sk
, skb
);
1673 sk_charge_skb(sk
, nskb
);
1677 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_ACK
;
1678 TCP_ECN_send_synack(tcp_sk(sk
), skb
);
1680 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1681 return tcp_transmit_skb(sk
, skb_clone(skb
, GFP_ATOMIC
));
1685 * Prepare a SYN-ACK.
1687 struct sk_buff
* tcp_make_synack(struct sock
*sk
, struct dst_entry
*dst
,
1688 struct request_sock
*req
)
1690 struct inet_request_sock
*ireq
= inet_rsk(req
);
1691 struct tcp_sock
*tp
= tcp_sk(sk
);
1693 int tcp_header_size
;
1694 struct sk_buff
*skb
;
1696 skb
= sock_wmalloc(sk
, MAX_TCP_HEADER
+ 15, 1, GFP_ATOMIC
);
1700 /* Reserve space for headers. */
1701 skb_reserve(skb
, MAX_TCP_HEADER
);
1703 skb
->dst
= dst_clone(dst
);
1705 tcp_header_size
= (sizeof(struct tcphdr
) + TCPOLEN_MSS
+
1706 (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0) +
1707 (ireq
->wscale_ok
? TCPOLEN_WSCALE_ALIGNED
: 0) +
1708 /* SACK_PERM is in the place of NOP NOP of TS */
1709 ((ireq
->sack_ok
&& !ireq
->tstamp_ok
) ? TCPOLEN_SACKPERM_ALIGNED
: 0));
1710 skb
->h
.th
= th
= (struct tcphdr
*) skb_push(skb
, tcp_header_size
);
1712 memset(th
, 0, sizeof(struct tcphdr
));
1715 if (dst
->dev
->features
&NETIF_F_TSO
)
1717 TCP_ECN_make_synack(req
, th
);
1718 th
->source
= inet_sk(sk
)->sport
;
1719 th
->dest
= ireq
->rmt_port
;
1720 TCP_SKB_CB(skb
)->seq
= tcp_rsk(req
)->snt_isn
;
1721 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ 1;
1722 TCP_SKB_CB(skb
)->sacked
= 0;
1723 skb_shinfo(skb
)->tso_segs
= 1;
1724 skb_shinfo(skb
)->tso_size
= 0;
1725 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
1726 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_isn
+ 1);
1727 if (req
->rcv_wnd
== 0) { /* ignored for retransmitted syns */
1729 /* Set this up on the first call only */
1730 req
->window_clamp
= tp
->window_clamp
? : dst_metric(dst
, RTAX_WINDOW
);
1731 /* tcp_full_space because it is guaranteed to be the first packet */
1732 tcp_select_initial_window(tcp_full_space(sk
),
1733 dst_metric(dst
, RTAX_ADVMSS
) - (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0),
1738 ireq
->rcv_wscale
= rcv_wscale
;
1741 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
1742 th
->window
= htons(req
->rcv_wnd
);
1744 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1745 tcp_syn_build_options((__u32
*)(th
+ 1), dst_metric(dst
, RTAX_ADVMSS
), ireq
->tstamp_ok
,
1746 ireq
->sack_ok
, ireq
->wscale_ok
, ireq
->rcv_wscale
,
1747 TCP_SKB_CB(skb
)->when
,
1751 th
->doff
= (tcp_header_size
>> 2);
1752 TCP_INC_STATS(TCP_MIB_OUTSEGS
);
1757 * Do all connect socket setups that can be done AF independent.
1759 static inline void tcp_connect_init(struct sock
*sk
)
1761 struct dst_entry
*dst
= __sk_dst_get(sk
);
1762 struct tcp_sock
*tp
= tcp_sk(sk
);
1765 /* We'll fix this up when we get a response from the other end.
1766 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
1768 tp
->tcp_header_len
= sizeof(struct tcphdr
) +
1769 (sysctl_tcp_timestamps
? TCPOLEN_TSTAMP_ALIGNED
: 0);
1771 /* If user gave his TCP_MAXSEG, record it to clamp */
1772 if (tp
->rx_opt
.user_mss
)
1773 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
1775 tcp_sync_mss(sk
, dst_mtu(dst
));
1777 if (!tp
->window_clamp
)
1778 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
1779 tp
->advmss
= dst_metric(dst
, RTAX_ADVMSS
);
1780 tcp_initialize_rcv_mss(sk
);
1782 tcp_select_initial_window(tcp_full_space(sk
),
1783 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
1786 sysctl_tcp_window_scaling
,
1789 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
1790 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
1793 sock_reset_flag(sk
, SOCK_DONE
);
1795 tcp_init_wl(tp
, tp
->write_seq
, 0);
1796 tp
->snd_una
= tp
->write_seq
;
1797 tp
->snd_sml
= tp
->write_seq
;
1802 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
1803 inet_csk(sk
)->icsk_retransmits
= 0;
1804 tcp_clear_retrans(tp
);
1808 * Build a SYN and send it off.
1810 int tcp_connect(struct sock
*sk
)
1812 struct tcp_sock
*tp
= tcp_sk(sk
);
1813 struct sk_buff
*buff
;
1815 tcp_connect_init(sk
);
1817 buff
= alloc_skb_fclone(MAX_TCP_HEADER
+ 15, sk
->sk_allocation
);
1818 if (unlikely(buff
== NULL
))
1821 /* Reserve space for headers. */
1822 skb_reserve(buff
, MAX_TCP_HEADER
);
1824 TCP_SKB_CB(buff
)->flags
= TCPCB_FLAG_SYN
;
1825 TCP_ECN_send_syn(sk
, tp
, buff
);
1826 TCP_SKB_CB(buff
)->sacked
= 0;
1827 skb_shinfo(buff
)->tso_segs
= 1;
1828 skb_shinfo(buff
)->tso_size
= 0;
1830 TCP_SKB_CB(buff
)->seq
= tp
->write_seq
++;
1831 TCP_SKB_CB(buff
)->end_seq
= tp
->write_seq
;
1832 tp
->snd_nxt
= tp
->write_seq
;
1833 tp
->pushed_seq
= tp
->write_seq
;
1836 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
1837 tp
->retrans_stamp
= TCP_SKB_CB(buff
)->when
;
1838 skb_header_release(buff
);
1839 __skb_queue_tail(&sk
->sk_write_queue
, buff
);
1840 sk_charge_skb(sk
, buff
);
1841 tp
->packets_out
+= tcp_skb_pcount(buff
);
1842 tcp_transmit_skb(sk
, skb_clone(buff
, GFP_KERNEL
));
1843 TCP_INC_STATS(TCP_MIB_ACTIVEOPENS
);
1845 /* Timer for repeating the SYN until an answer. */
1846 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1847 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
1851 /* Send out a delayed ack, the caller does the policy checking
1852 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
1855 void tcp_send_delayed_ack(struct sock
*sk
)
1857 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1858 int ato
= icsk
->icsk_ack
.ato
;
1859 unsigned long timeout
;
1861 if (ato
> TCP_DELACK_MIN
) {
1862 const struct tcp_sock
*tp
= tcp_sk(sk
);
1865 if (icsk
->icsk_ack
.pingpong
|| (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
1866 max_ato
= TCP_DELACK_MAX
;
1868 /* Slow path, intersegment interval is "high". */
1870 /* If some rtt estimate is known, use it to bound delayed ack.
1871 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
1875 int rtt
= max(tp
->srtt
>>3, TCP_DELACK_MIN
);
1881 ato
= min(ato
, max_ato
);
1884 /* Stay within the limit we were given */
1885 timeout
= jiffies
+ ato
;
1887 /* Use new timeout only if there wasn't a older one earlier. */
1888 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
1889 /* If delack timer was blocked or is about to expire,
1892 if (icsk
->icsk_ack
.blocked
||
1893 time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
1898 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
1899 timeout
= icsk
->icsk_ack
.timeout
;
1901 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
1902 icsk
->icsk_ack
.timeout
= timeout
;
1903 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
1906 /* This routine sends an ack and also updates the window. */
1907 void tcp_send_ack(struct sock
*sk
)
1909 /* If we have been reset, we may not send again. */
1910 if (sk
->sk_state
!= TCP_CLOSE
) {
1911 struct tcp_sock
*tp
= tcp_sk(sk
);
1912 struct sk_buff
*buff
;
1914 /* We are not putting this on the write queue, so
1915 * tcp_transmit_skb() will set the ownership to this
1918 buff
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
1920 inet_csk_schedule_ack(sk
);
1921 inet_csk(sk
)->icsk_ack
.ato
= TCP_ATO_MIN
;
1922 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
1923 TCP_DELACK_MAX
, TCP_RTO_MAX
);
1927 /* Reserve space for headers and prepare control bits. */
1928 skb_reserve(buff
, MAX_TCP_HEADER
);
1930 TCP_SKB_CB(buff
)->flags
= TCPCB_FLAG_ACK
;
1931 TCP_SKB_CB(buff
)->sacked
= 0;
1932 skb_shinfo(buff
)->tso_segs
= 1;
1933 skb_shinfo(buff
)->tso_size
= 0;
1935 /* Send it off, this clears delayed acks for us. */
1936 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(buff
)->end_seq
= tcp_acceptable_seq(sk
, tp
);
1937 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
1938 tcp_transmit_skb(sk
, buff
);
1942 /* This routine sends a packet with an out of date sequence
1943 * number. It assumes the other end will try to ack it.
1945 * Question: what should we make while urgent mode?
1946 * 4.4BSD forces sending single byte of data. We cannot send
1947 * out of window data, because we have SND.NXT==SND.MAX...
1949 * Current solution: to send TWO zero-length segments in urgent mode:
1950 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
1951 * out-of-date with SND.UNA-1 to probe window.
1953 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
)
1955 struct tcp_sock
*tp
= tcp_sk(sk
);
1956 struct sk_buff
*skb
;
1958 /* We don't queue it, tcp_transmit_skb() sets ownership. */
1959 skb
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
1963 /* Reserve space for headers and set control bits. */
1964 skb_reserve(skb
, MAX_TCP_HEADER
);
1966 TCP_SKB_CB(skb
)->flags
= TCPCB_FLAG_ACK
;
1967 TCP_SKB_CB(skb
)->sacked
= urgent
;
1968 skb_shinfo(skb
)->tso_segs
= 1;
1969 skb_shinfo(skb
)->tso_size
= 0;
1971 /* Use a previous sequence. This should cause the other
1972 * end to send an ack. Don't queue or clone SKB, just
1975 TCP_SKB_CB(skb
)->seq
= urgent
? tp
->snd_una
: tp
->snd_una
- 1;
1976 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
;
1977 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1978 return tcp_transmit_skb(sk
, skb
);
1981 int tcp_write_wakeup(struct sock
*sk
)
1983 if (sk
->sk_state
!= TCP_CLOSE
) {
1984 struct tcp_sock
*tp
= tcp_sk(sk
);
1985 struct sk_buff
*skb
;
1987 if ((skb
= sk
->sk_send_head
) != NULL
&&
1988 before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
+tp
->snd_wnd
)) {
1990 unsigned int mss
= tcp_current_mss(sk
, 0);
1991 unsigned int seg_size
= tp
->snd_una
+tp
->snd_wnd
-TCP_SKB_CB(skb
)->seq
;
1993 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
1994 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
1996 /* We are probing the opening of a window
1997 * but the window size is != 0
1998 * must have been a result SWS avoidance ( sender )
2000 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
2002 seg_size
= min(seg_size
, mss
);
2003 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2004 if (tcp_fragment(sk
, skb
, seg_size
, mss
))
2006 } else if (!tcp_skb_pcount(skb
))
2007 tcp_set_skb_tso_segs(sk
, skb
, mss
);
2009 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2010 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2011 err
= tcp_transmit_skb(sk
, skb_clone(skb
, GFP_ATOMIC
));
2013 update_send_head(sk
, tp
, skb
);
2018 between(tp
->snd_up
, tp
->snd_una
+1, tp
->snd_una
+0xFFFF))
2019 tcp_xmit_probe_skb(sk
, TCPCB_URG
);
2020 return tcp_xmit_probe_skb(sk
, 0);
2026 /* A window probe timeout has occurred. If window is not closed send
2027 * a partial packet else a zero probe.
2029 void tcp_send_probe0(struct sock
*sk
)
2031 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2032 struct tcp_sock
*tp
= tcp_sk(sk
);
2035 err
= tcp_write_wakeup(sk
);
2037 if (tp
->packets_out
|| !sk
->sk_send_head
) {
2038 /* Cancel probe timer, if it is not required. */
2039 icsk
->icsk_probes_out
= 0;
2040 icsk
->icsk_backoff
= 0;
2045 if (icsk
->icsk_backoff
< sysctl_tcp_retries2
)
2046 icsk
->icsk_backoff
++;
2047 icsk
->icsk_probes_out
++;
2048 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2049 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
2052 /* If packet was not sent due to local congestion,
2053 * do not backoff and do not remember icsk_probes_out.
2054 * Let local senders to fight for local resources.
2056 * Use accumulated backoff yet.
2058 if (!icsk
->icsk_probes_out
)
2059 icsk
->icsk_probes_out
= 1;
2060 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2061 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
,
2062 TCP_RESOURCE_PROBE_INTERVAL
),
2067 EXPORT_SYMBOL(tcp_connect
);
2068 EXPORT_SYMBOL(tcp_make_synack
);
2069 EXPORT_SYMBOL(tcp_simple_retransmit
);
2070 EXPORT_SYMBOL(tcp_sync_mss
);