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 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 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 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. */
253 static void tcp_build_and_update_options(__u32
*ptr
, struct tcp_sock
*tp
,
256 if (tp
->rx_opt
.tstamp_ok
) {
257 *ptr
++ = __constant_htonl((TCPOPT_NOP
<< 24) |
259 (TCPOPT_TIMESTAMP
<< 8) |
261 *ptr
++ = htonl(tstamp
);
262 *ptr
++ = htonl(tp
->rx_opt
.ts_recent
);
264 if (tp
->rx_opt
.eff_sacks
) {
265 struct tcp_sack_block
*sp
= tp
->rx_opt
.dsack
? tp
->duplicate_sack
: tp
->selective_acks
;
268 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
271 (TCPOLEN_SACK_BASE
+ (tp
->rx_opt
.eff_sacks
*
272 TCPOLEN_SACK_PERBLOCK
)));
273 for(this_sack
= 0; this_sack
< tp
->rx_opt
.eff_sacks
; this_sack
++) {
274 *ptr
++ = htonl(sp
[this_sack
].start_seq
);
275 *ptr
++ = htonl(sp
[this_sack
].end_seq
);
277 if (tp
->rx_opt
.dsack
) {
278 tp
->rx_opt
.dsack
= 0;
279 tp
->rx_opt
.eff_sacks
--;
284 /* Construct a tcp options header for a SYN or SYN_ACK packet.
285 * If this is every changed make sure to change the definition of
286 * MAX_SYN_SIZE to match the new maximum number of options that you
289 static void tcp_syn_build_options(__u32
*ptr
, int mss
, int ts
, int sack
,
290 int offer_wscale
, int wscale
, __u32 tstamp
,
293 /* We always get an MSS option.
294 * The option bytes which will be seen in normal data
295 * packets should timestamps be used, must be in the MSS
296 * advertised. But we subtract them from tp->mss_cache so
297 * that calculations in tcp_sendmsg are simpler etc.
298 * So account for this fact here if necessary. If we
299 * don't do this correctly, as a receiver we won't
300 * recognize data packets as being full sized when we
301 * should, and thus we won't abide by the delayed ACK
303 * SACKs don't matter, we never delay an ACK when we
304 * have any of those going out.
306 *ptr
++ = htonl((TCPOPT_MSS
<< 24) | (TCPOLEN_MSS
<< 16) | mss
);
309 *ptr
++ = __constant_htonl((TCPOPT_SACK_PERM
<< 24) | (TCPOLEN_SACK_PERM
<< 16) |
310 (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
);
312 *ptr
++ = __constant_htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16) |
313 (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
);
314 *ptr
++ = htonl(tstamp
); /* TSVAL */
315 *ptr
++ = htonl(ts_recent
); /* TSECR */
317 *ptr
++ = __constant_htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16) |
318 (TCPOPT_SACK_PERM
<< 8) | TCPOLEN_SACK_PERM
);
320 *ptr
++ = htonl((TCPOPT_NOP
<< 24) | (TCPOPT_WINDOW
<< 16) | (TCPOLEN_WINDOW
<< 8) | (wscale
));
323 /* This routine actually transmits TCP packets queued in by
324 * tcp_do_sendmsg(). This is used by both the initial
325 * transmission and possible later retransmissions.
326 * All SKB's seen here are completely headerless. It is our
327 * job to build the TCP header, and pass the packet down to
328 * IP so it can do the same plus pass the packet off to the
331 * We are working here with either a clone of the original
332 * SKB, or a fresh unique copy made by the retransmit engine.
334 static int tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int clone_it
, gfp_t gfp_mask
)
336 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
337 struct inet_sock
*inet
;
339 struct tcp_skb_cb
*tcb
;
345 BUG_ON(!skb
|| !tcp_skb_pcount(skb
));
347 /* If congestion control is doing timestamping, we must
348 * take such a timestamp before we potentially clone/copy.
350 if (icsk
->icsk_ca_ops
->rtt_sample
)
351 __net_timestamp(skb
);
353 if (likely(clone_it
)) {
354 if (unlikely(skb_cloned(skb
)))
355 skb
= pskb_copy(skb
, gfp_mask
);
357 skb
= skb_clone(skb
, gfp_mask
);
364 tcb
= TCP_SKB_CB(skb
);
365 tcp_header_size
= tp
->tcp_header_len
;
367 #define SYSCTL_FLAG_TSTAMPS 0x1
368 #define SYSCTL_FLAG_WSCALE 0x2
369 #define SYSCTL_FLAG_SACK 0x4
372 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
373 tcp_header_size
= sizeof(struct tcphdr
) + TCPOLEN_MSS
;
374 if(sysctl_tcp_timestamps
) {
375 tcp_header_size
+= TCPOLEN_TSTAMP_ALIGNED
;
376 sysctl_flags
|= SYSCTL_FLAG_TSTAMPS
;
378 if (sysctl_tcp_window_scaling
) {
379 tcp_header_size
+= TCPOLEN_WSCALE_ALIGNED
;
380 sysctl_flags
|= SYSCTL_FLAG_WSCALE
;
382 if (sysctl_tcp_sack
) {
383 sysctl_flags
|= SYSCTL_FLAG_SACK
;
384 if (!(sysctl_flags
& SYSCTL_FLAG_TSTAMPS
))
385 tcp_header_size
+= TCPOLEN_SACKPERM_ALIGNED
;
387 } else if (unlikely(tp
->rx_opt
.eff_sacks
)) {
388 /* A SACK is 2 pad bytes, a 2 byte header, plus
389 * 2 32-bit sequence numbers for each SACK block.
391 tcp_header_size
+= (TCPOLEN_SACK_BASE_ALIGNED
+
392 (tp
->rx_opt
.eff_sacks
*
393 TCPOLEN_SACK_PERBLOCK
));
396 if (tcp_packets_in_flight(tp
) == 0)
397 tcp_ca_event(sk
, CA_EVENT_TX_START
);
399 th
= (struct tcphdr
*) skb_push(skb
, tcp_header_size
);
401 skb_set_owner_w(skb
, sk
);
403 /* Build TCP header and checksum it. */
404 th
->source
= inet
->sport
;
405 th
->dest
= inet
->dport
;
406 th
->seq
= htonl(tcb
->seq
);
407 th
->ack_seq
= htonl(tp
->rcv_nxt
);
408 *(((__u16
*)th
) + 6) = htons(((tcp_header_size
>> 2) << 12) |
411 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
412 /* RFC1323: The window in SYN & SYN/ACK segments
415 th
->window
= htons(min(tp
->rcv_wnd
, 65535U));
417 th
->window
= htons(tcp_select_window(sk
));
422 if (unlikely(tp
->urg_mode
&&
423 between(tp
->snd_up
, tcb
->seq
+1, tcb
->seq
+0xFFFF))) {
424 th
->urg_ptr
= htons(tp
->snd_up
-tcb
->seq
);
428 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
429 tcp_syn_build_options((__u32
*)(th
+ 1),
430 tcp_advertise_mss(sk
),
431 (sysctl_flags
& SYSCTL_FLAG_TSTAMPS
),
432 (sysctl_flags
& SYSCTL_FLAG_SACK
),
433 (sysctl_flags
& SYSCTL_FLAG_WSCALE
),
434 tp
->rx_opt
.rcv_wscale
,
436 tp
->rx_opt
.ts_recent
);
438 tcp_build_and_update_options((__u32
*)(th
+ 1),
440 TCP_ECN_send(sk
, tp
, skb
, tcp_header_size
);
443 icsk
->icsk_af_ops
->send_check(sk
, skb
->len
, skb
);
445 if (likely(tcb
->flags
& TCPCB_FLAG_ACK
))
446 tcp_event_ack_sent(sk
, tcp_skb_pcount(skb
));
448 if (skb
->len
!= tcp_header_size
)
449 tcp_event_data_sent(tp
, skb
, sk
);
451 TCP_INC_STATS(TCP_MIB_OUTSEGS
);
453 err
= icsk
->icsk_af_ops
->queue_xmit(skb
, 0);
454 if (unlikely(err
<= 0))
459 /* NET_XMIT_CN is special. It does not guarantee,
460 * that this packet is lost. It tells that device
461 * is about to start to drop packets or already
462 * drops some packets of the same priority and
463 * invokes us to send less aggressively.
465 return err
== NET_XMIT_CN
? 0 : err
;
467 #undef SYSCTL_FLAG_TSTAMPS
468 #undef SYSCTL_FLAG_WSCALE
469 #undef SYSCTL_FLAG_SACK
473 /* This routine just queue's the buffer
475 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
476 * otherwise socket can stall.
478 static void tcp_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
480 struct tcp_sock
*tp
= tcp_sk(sk
);
482 /* Advance write_seq and place onto the write_queue. */
483 tp
->write_seq
= TCP_SKB_CB(skb
)->end_seq
;
484 skb_header_release(skb
);
485 __skb_queue_tail(&sk
->sk_write_queue
, skb
);
486 sk_charge_skb(sk
, skb
);
488 /* Queue it, remembering where we must start sending. */
489 if (sk
->sk_send_head
== NULL
)
490 sk
->sk_send_head
= skb
;
493 static void tcp_set_skb_tso_segs(struct sock
*sk
, struct sk_buff
*skb
, unsigned int mss_now
)
495 if (skb
->len
<= mss_now
||
496 !(sk
->sk_route_caps
& NETIF_F_TSO
)) {
497 /* Avoid the costly divide in the normal
500 skb_shinfo(skb
)->tso_segs
= 1;
501 skb_shinfo(skb
)->tso_size
= 0;
505 factor
= skb
->len
+ (mss_now
- 1);
507 skb_shinfo(skb
)->tso_segs
= factor
;
508 skb_shinfo(skb
)->tso_size
= mss_now
;
512 /* Function to create two new TCP segments. Shrinks the given segment
513 * to the specified size and appends a new segment with the rest of the
514 * packet to the list. This won't be called frequently, I hope.
515 * Remember, these are still headerless SKBs at this point.
517 int tcp_fragment(struct sock
*sk
, struct sk_buff
*skb
, u32 len
, unsigned int mss_now
)
519 struct tcp_sock
*tp
= tcp_sk(sk
);
520 struct sk_buff
*buff
;
521 int nsize
, old_factor
;
524 BUG_ON(len
> skb
->len
);
526 clear_all_retrans_hints(tp
);
527 nsize
= skb_headlen(skb
) - len
;
531 if (skb_cloned(skb
) &&
532 skb_is_nonlinear(skb
) &&
533 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
536 /* Get a new skb... force flag on. */
537 buff
= sk_stream_alloc_skb(sk
, nsize
, GFP_ATOMIC
);
539 return -ENOMEM
; /* We'll just try again later. */
541 buff
->truesize
= skb
->len
- len
;
542 skb
->truesize
-= buff
->truesize
;
544 /* Correct the sequence numbers. */
545 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
546 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
547 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
549 /* PSH and FIN should only be set in the second packet. */
550 flags
= TCP_SKB_CB(skb
)->flags
;
551 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
552 TCP_SKB_CB(buff
)->flags
= flags
;
553 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
554 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_AT_TAIL
;
556 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_HW
) {
557 /* Copy and checksum data tail into the new buffer. */
558 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
, skb_put(buff
, nsize
),
563 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
565 skb
->ip_summed
= CHECKSUM_HW
;
566 skb_split(skb
, buff
, len
);
569 buff
->ip_summed
= skb
->ip_summed
;
571 /* Looks stupid, but our code really uses when of
572 * skbs, which it never sent before. --ANK
574 TCP_SKB_CB(buff
)->when
= TCP_SKB_CB(skb
)->when
;
575 buff
->tstamp
= skb
->tstamp
;
577 old_factor
= tcp_skb_pcount(skb
);
579 /* Fix up tso_factor for both original and new SKB. */
580 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
581 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
583 /* If this packet has been sent out already, we must
584 * adjust the various packet counters.
586 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
587 int diff
= old_factor
- tcp_skb_pcount(skb
) -
588 tcp_skb_pcount(buff
);
590 tp
->packets_out
-= diff
;
592 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
593 tp
->sacked_out
-= diff
;
594 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
595 tp
->retrans_out
-= diff
;
597 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
) {
598 tp
->lost_out
-= diff
;
599 tp
->left_out
-= diff
;
603 /* Adjust Reno SACK estimate. */
604 if (!tp
->rx_opt
.sack_ok
) {
605 tp
->sacked_out
-= diff
;
606 if ((int)tp
->sacked_out
< 0)
608 tcp_sync_left_out(tp
);
611 tp
->fackets_out
-= diff
;
612 if ((int)tp
->fackets_out
< 0)
617 /* Link BUFF into the send queue. */
618 skb_header_release(buff
);
619 __skb_append(skb
, buff
, &sk
->sk_write_queue
);
624 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
625 * eventually). The difference is that pulled data not copied, but
626 * immediately discarded.
628 static unsigned char *__pskb_trim_head(struct sk_buff
*skb
, int len
)
634 for (i
=0; i
<skb_shinfo(skb
)->nr_frags
; i
++) {
635 if (skb_shinfo(skb
)->frags
[i
].size
<= eat
) {
636 put_page(skb_shinfo(skb
)->frags
[i
].page
);
637 eat
-= skb_shinfo(skb
)->frags
[i
].size
;
639 skb_shinfo(skb
)->frags
[k
] = skb_shinfo(skb
)->frags
[i
];
641 skb_shinfo(skb
)->frags
[k
].page_offset
+= eat
;
642 skb_shinfo(skb
)->frags
[k
].size
-= eat
;
648 skb_shinfo(skb
)->nr_frags
= k
;
650 skb
->tail
= skb
->data
;
651 skb
->data_len
-= len
;
652 skb
->len
= skb
->data_len
;
656 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
658 if (skb_cloned(skb
) &&
659 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
662 if (len
<= skb_headlen(skb
)) {
663 __skb_pull(skb
, len
);
665 if (__pskb_trim_head(skb
, len
-skb_headlen(skb
)) == NULL
)
669 TCP_SKB_CB(skb
)->seq
+= len
;
670 skb
->ip_summed
= CHECKSUM_HW
;
672 skb
->truesize
-= len
;
673 sk
->sk_wmem_queued
-= len
;
674 sk
->sk_forward_alloc
+= len
;
675 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
677 /* Any change of skb->len requires recalculation of tso
680 if (tcp_skb_pcount(skb
) > 1)
681 tcp_set_skb_tso_segs(sk
, skb
, tcp_current_mss(sk
, 1));
686 /* This function synchronize snd mss to current pmtu/exthdr set.
688 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
689 for TCP options, but includes only bare TCP header.
691 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
692 It is minimum of user_mss and mss received with SYN.
693 It also does not include TCP options.
695 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
697 tp->mss_cache is current effective sending mss, including
698 all tcp options except for SACKs. It is evaluated,
699 taking into account current pmtu, but never exceeds
700 tp->rx_opt.mss_clamp.
702 NOTE1. rfc1122 clearly states that advertised MSS
703 DOES NOT include either tcp or ip options.
705 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
706 are READ ONLY outside this function. --ANK (980731)
709 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
711 struct tcp_sock
*tp
= tcp_sk(sk
);
712 struct inet_connection_sock
*icsk
= inet_csk(sk
);
713 /* Calculate base mss without TCP options:
714 It is MMS_S - sizeof(tcphdr) of rfc1122
716 int mss_now
= (pmtu
- icsk
->icsk_af_ops
->net_header_len
-
717 sizeof(struct tcphdr
));
719 /* Clamp it (mss_clamp does not include tcp options) */
720 if (mss_now
> tp
->rx_opt
.mss_clamp
)
721 mss_now
= tp
->rx_opt
.mss_clamp
;
723 /* Now subtract optional transport overhead */
724 mss_now
-= icsk
->icsk_ext_hdr_len
;
726 /* Then reserve room for full set of TCP options and 8 bytes of data */
730 /* Now subtract TCP options size, not including SACKs */
731 mss_now
-= tp
->tcp_header_len
- sizeof(struct tcphdr
);
733 /* Bound mss with half of window */
734 if (tp
->max_window
&& mss_now
> (tp
->max_window
>>1))
735 mss_now
= max((tp
->max_window
>>1), 68U - tp
->tcp_header_len
);
737 /* And store cached results */
738 icsk
->icsk_pmtu_cookie
= pmtu
;
739 tp
->mss_cache
= mss_now
;
744 /* Compute the current effective MSS, taking SACKs and IP options,
745 * and even PMTU discovery events into account.
747 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
748 * cannot be large. However, taking into account rare use of URG, this
751 unsigned int tcp_current_mss(struct sock
*sk
, int large_allowed
)
753 struct tcp_sock
*tp
= tcp_sk(sk
);
754 struct dst_entry
*dst
= __sk_dst_get(sk
);
759 mss_now
= tp
->mss_cache
;
762 (sk
->sk_route_caps
& NETIF_F_TSO
) &&
767 u32 mtu
= dst_mtu(dst
);
768 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
769 mss_now
= tcp_sync_mss(sk
, mtu
);
772 if (tp
->rx_opt
.eff_sacks
)
773 mss_now
-= (TCPOLEN_SACK_BASE_ALIGNED
+
774 (tp
->rx_opt
.eff_sacks
* TCPOLEN_SACK_PERBLOCK
));
776 xmit_size_goal
= mss_now
;
779 xmit_size_goal
= (65535 -
780 inet_csk(sk
)->icsk_af_ops
->net_header_len
-
781 inet_csk(sk
)->icsk_ext_hdr_len
-
784 if (tp
->max_window
&&
785 (xmit_size_goal
> (tp
->max_window
>> 1)))
786 xmit_size_goal
= max((tp
->max_window
>> 1),
787 68U - tp
->tcp_header_len
);
789 xmit_size_goal
-= (xmit_size_goal
% mss_now
);
791 tp
->xmit_size_goal
= xmit_size_goal
;
796 /* Congestion window validation. (RFC2861) */
798 static void tcp_cwnd_validate(struct sock
*sk
, struct tcp_sock
*tp
)
800 __u32 packets_out
= tp
->packets_out
;
802 if (packets_out
>= tp
->snd_cwnd
) {
803 /* Network is feed fully. */
804 tp
->snd_cwnd_used
= 0;
805 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
807 /* Network starves. */
808 if (tp
->packets_out
> tp
->snd_cwnd_used
)
809 tp
->snd_cwnd_used
= tp
->packets_out
;
811 if ((s32
)(tcp_time_stamp
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
)
812 tcp_cwnd_application_limited(sk
);
816 static unsigned int tcp_window_allows(struct tcp_sock
*tp
, struct sk_buff
*skb
, unsigned int mss_now
, unsigned int cwnd
)
818 u32 window
, cwnd_len
;
820 window
= (tp
->snd_una
+ tp
->snd_wnd
- TCP_SKB_CB(skb
)->seq
);
821 cwnd_len
= mss_now
* cwnd
;
822 return min(window
, cwnd_len
);
825 /* Can at least one segment of SKB be sent right now, according to the
826 * congestion window rules? If so, return how many segments are allowed.
828 static inline unsigned int tcp_cwnd_test(struct tcp_sock
*tp
, struct sk_buff
*skb
)
832 /* Don't be strict about the congestion window for the final FIN. */
833 if ((TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
) &&
834 tcp_skb_pcount(skb
) == 1)
837 in_flight
= tcp_packets_in_flight(tp
);
839 if (in_flight
< cwnd
)
840 return (cwnd
- in_flight
);
845 /* This must be invoked the first time we consider transmitting
848 static int tcp_init_tso_segs(struct sock
*sk
, struct sk_buff
*skb
, unsigned int mss_now
)
850 int tso_segs
= tcp_skb_pcount(skb
);
854 skb_shinfo(skb
)->tso_size
!= mss_now
)) {
855 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
856 tso_segs
= tcp_skb_pcount(skb
);
861 static inline int tcp_minshall_check(const struct tcp_sock
*tp
)
863 return after(tp
->snd_sml
,tp
->snd_una
) &&
864 !after(tp
->snd_sml
, tp
->snd_nxt
);
867 /* Return 0, if packet can be sent now without violation Nagle's rules:
868 * 1. It is full sized.
869 * 2. Or it contains FIN. (already checked by caller)
870 * 3. Or TCP_NODELAY was set.
871 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
872 * With Minshall's modification: all sent small packets are ACKed.
875 static inline int tcp_nagle_check(const struct tcp_sock
*tp
,
876 const struct sk_buff
*skb
,
877 unsigned mss_now
, int nonagle
)
879 return (skb
->len
< mss_now
&&
880 ((nonagle
&TCP_NAGLE_CORK
) ||
883 tcp_minshall_check(tp
))));
886 /* Return non-zero if the Nagle test allows this packet to be
889 static inline int tcp_nagle_test(struct tcp_sock
*tp
, struct sk_buff
*skb
,
890 unsigned int cur_mss
, int nonagle
)
892 /* Nagle rule does not apply to frames, which sit in the middle of the
893 * write_queue (they have no chances to get new data).
895 * This is implemented in the callers, where they modify the 'nonagle'
896 * argument based upon the location of SKB in the send queue.
898 if (nonagle
& TCP_NAGLE_PUSH
)
901 /* Don't use the nagle rule for urgent data (or for the final FIN). */
903 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
))
906 if (!tcp_nagle_check(tp
, skb
, cur_mss
, nonagle
))
912 /* Does at least the first segment of SKB fit into the send window? */
913 static inline int tcp_snd_wnd_test(struct tcp_sock
*tp
, struct sk_buff
*skb
, unsigned int cur_mss
)
915 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
917 if (skb
->len
> cur_mss
)
918 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
920 return !after(end_seq
, tp
->snd_una
+ tp
->snd_wnd
);
923 /* This checks if the data bearing packet SKB (usually sk->sk_send_head)
924 * should be put on the wire right now. If so, it returns the number of
925 * packets allowed by the congestion window.
927 static unsigned int tcp_snd_test(struct sock
*sk
, struct sk_buff
*skb
,
928 unsigned int cur_mss
, int nonagle
)
930 struct tcp_sock
*tp
= tcp_sk(sk
);
931 unsigned int cwnd_quota
;
933 tcp_init_tso_segs(sk
, skb
, cur_mss
);
935 if (!tcp_nagle_test(tp
, skb
, cur_mss
, nonagle
))
938 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
940 !tcp_snd_wnd_test(tp
, skb
, cur_mss
))
946 static inline int tcp_skb_is_last(const struct sock
*sk
,
947 const struct sk_buff
*skb
)
949 return skb
->next
== (struct sk_buff
*)&sk
->sk_write_queue
;
952 int tcp_may_send_now(struct sock
*sk
, struct tcp_sock
*tp
)
954 struct sk_buff
*skb
= sk
->sk_send_head
;
957 tcp_snd_test(sk
, skb
, tcp_current_mss(sk
, 1),
958 (tcp_skb_is_last(sk
, skb
) ?
963 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
964 * which is put after SKB on the list. It is very much like
965 * tcp_fragment() except that it may make several kinds of assumptions
966 * in order to speed up the splitting operation. In particular, we
967 * know that all the data is in scatter-gather pages, and that the
968 * packet has never been sent out before (and thus is not cloned).
970 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
, unsigned int mss_now
)
972 struct sk_buff
*buff
;
973 int nlen
= skb
->len
- len
;
976 /* All of a TSO frame must be composed of paged data. */
977 if (skb
->len
!= skb
->data_len
)
978 return tcp_fragment(sk
, skb
, len
, mss_now
);
980 buff
= sk_stream_alloc_pskb(sk
, 0, 0, GFP_ATOMIC
);
981 if (unlikely(buff
== NULL
))
984 buff
->truesize
= nlen
;
985 skb
->truesize
-= nlen
;
987 /* Correct the sequence numbers. */
988 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
989 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
990 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
992 /* PSH and FIN should only be set in the second packet. */
993 flags
= TCP_SKB_CB(skb
)->flags
;
994 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
995 TCP_SKB_CB(buff
)->flags
= flags
;
997 /* This packet was never sent out yet, so no SACK bits. */
998 TCP_SKB_CB(buff
)->sacked
= 0;
1000 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_HW
;
1001 skb_split(skb
, buff
, len
);
1003 /* Fix up tso_factor for both original and new SKB. */
1004 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1005 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1007 /* Link BUFF into the send queue. */
1008 skb_header_release(buff
);
1009 __skb_append(skb
, buff
, &sk
->sk_write_queue
);
1014 /* Try to defer sending, if possible, in order to minimize the amount
1015 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1017 * This algorithm is from John Heffner.
1019 static int tcp_tso_should_defer(struct sock
*sk
, struct tcp_sock
*tp
, struct sk_buff
*skb
)
1021 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1022 u32 send_win
, cong_win
, limit
, in_flight
;
1024 if (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
)
1027 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
1030 in_flight
= tcp_packets_in_flight(tp
);
1032 BUG_ON(tcp_skb_pcount(skb
) <= 1 ||
1033 (tp
->snd_cwnd
<= in_flight
));
1035 send_win
= (tp
->snd_una
+ tp
->snd_wnd
) - TCP_SKB_CB(skb
)->seq
;
1037 /* From in_flight test above, we know that cwnd > in_flight. */
1038 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
1040 limit
= min(send_win
, cong_win
);
1042 /* If a full-sized TSO skb can be sent, do it. */
1046 if (sysctl_tcp_tso_win_divisor
) {
1047 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
1049 /* If at least some fraction of a window is available,
1052 chunk
/= sysctl_tcp_tso_win_divisor
;
1056 /* Different approach, try not to defer past a single
1057 * ACK. Receiver should ACK every other full sized
1058 * frame, so if we have space for more than 3 frames
1061 if (limit
> tcp_max_burst(tp
) * tp
->mss_cache
)
1065 /* Ok, it looks like it is advisable to defer. */
1069 /* This routine writes packets to the network. It advances the
1070 * send_head. This happens as incoming acks open up the remote
1073 * Returns 1, if no segments are in flight and we have queued segments, but
1074 * cannot send anything now because of SWS or another problem.
1076 static int tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
)
1078 struct tcp_sock
*tp
= tcp_sk(sk
);
1079 struct sk_buff
*skb
;
1080 unsigned int tso_segs
, sent_pkts
;
1083 /* If we are closed, the bytes will have to remain here.
1084 * In time closedown will finish, we empty the write queue and all
1087 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
1091 while ((skb
= sk
->sk_send_head
)) {
1094 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1097 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1101 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
)))
1104 if (tso_segs
== 1) {
1105 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
1106 (tcp_skb_is_last(sk
, skb
) ?
1107 nonagle
: TCP_NAGLE_PUSH
))))
1110 if (tcp_tso_should_defer(sk
, tp
, skb
))
1116 limit
= tcp_window_allows(tp
, skb
,
1117 mss_now
, cwnd_quota
);
1119 if (skb
->len
< limit
) {
1120 unsigned int trim
= skb
->len
% mss_now
;
1123 limit
= skb
->len
- trim
;
1127 if (skb
->len
> limit
&&
1128 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
)))
1131 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1133 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
)))
1136 /* Advance the send_head. This one is sent out.
1137 * This call will increment packets_out.
1139 update_send_head(sk
, tp
, skb
);
1141 tcp_minshall_update(tp
, mss_now
, skb
);
1145 if (likely(sent_pkts
)) {
1146 tcp_cwnd_validate(sk
, tp
);
1149 return !tp
->packets_out
&& sk
->sk_send_head
;
1152 /* Push out any pending frames which were held back due to
1153 * TCP_CORK or attempt at coalescing tiny packets.
1154 * The socket must be locked by the caller.
1156 void __tcp_push_pending_frames(struct sock
*sk
, struct tcp_sock
*tp
,
1157 unsigned int cur_mss
, int nonagle
)
1159 struct sk_buff
*skb
= sk
->sk_send_head
;
1162 if (tcp_write_xmit(sk
, cur_mss
, nonagle
))
1163 tcp_check_probe_timer(sk
, tp
);
1167 /* Send _single_ skb sitting at the send head. This function requires
1168 * true push pending frames to setup probe timer etc.
1170 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
1172 struct tcp_sock
*tp
= tcp_sk(sk
);
1173 struct sk_buff
*skb
= sk
->sk_send_head
;
1174 unsigned int tso_segs
, cwnd_quota
;
1176 BUG_ON(!skb
|| skb
->len
< mss_now
);
1178 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1179 cwnd_quota
= tcp_snd_test(sk
, skb
, mss_now
, TCP_NAGLE_PUSH
);
1181 if (likely(cwnd_quota
)) {
1188 limit
= tcp_window_allows(tp
, skb
,
1189 mss_now
, cwnd_quota
);
1191 if (skb
->len
< limit
) {
1192 unsigned int trim
= skb
->len
% mss_now
;
1195 limit
= skb
->len
- trim
;
1199 if (skb
->len
> limit
&&
1200 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
)))
1203 /* Send it out now. */
1204 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1206 if (likely(!tcp_transmit_skb(sk
, skb
, 1, sk
->sk_allocation
))) {
1207 update_send_head(sk
, tp
, skb
);
1208 tcp_cwnd_validate(sk
, tp
);
1214 /* This function returns the amount that we can raise the
1215 * usable window based on the following constraints
1217 * 1. The window can never be shrunk once it is offered (RFC 793)
1218 * 2. We limit memory per socket
1221 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1222 * RECV.NEXT + RCV.WIN fixed until:
1223 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1225 * i.e. don't raise the right edge of the window until you can raise
1226 * it at least MSS bytes.
1228 * Unfortunately, the recommended algorithm breaks header prediction,
1229 * since header prediction assumes th->window stays fixed.
1231 * Strictly speaking, keeping th->window fixed violates the receiver
1232 * side SWS prevention criteria. The problem is that under this rule
1233 * a stream of single byte packets will cause the right side of the
1234 * window to always advance by a single byte.
1236 * Of course, if the sender implements sender side SWS prevention
1237 * then this will not be a problem.
1239 * BSD seems to make the following compromise:
1241 * If the free space is less than the 1/4 of the maximum
1242 * space available and the free space is less than 1/2 mss,
1243 * then set the window to 0.
1244 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1245 * Otherwise, just prevent the window from shrinking
1246 * and from being larger than the largest representable value.
1248 * This prevents incremental opening of the window in the regime
1249 * where TCP is limited by the speed of the reader side taking
1250 * data out of the TCP receive queue. It does nothing about
1251 * those cases where the window is constrained on the sender side
1252 * because the pipeline is full.
1254 * BSD also seems to "accidentally" limit itself to windows that are a
1255 * multiple of MSS, at least until the free space gets quite small.
1256 * This would appear to be a side effect of the mbuf implementation.
1257 * Combining these two algorithms results in the observed behavior
1258 * of having a fixed window size at almost all times.
1260 * Below we obtain similar behavior by forcing the offered window to
1261 * a multiple of the mss when it is feasible to do so.
1263 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1264 * Regular options like TIMESTAMP are taken into account.
1266 u32
__tcp_select_window(struct sock
*sk
)
1268 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1269 struct tcp_sock
*tp
= tcp_sk(sk
);
1270 /* MSS for the peer's data. Previous versions used mss_clamp
1271 * here. I don't know if the value based on our guesses
1272 * of peer's MSS is better for the performance. It's more correct
1273 * but may be worse for the performance because of rcv_mss
1274 * fluctuations. --SAW 1998/11/1
1276 int mss
= icsk
->icsk_ack
.rcv_mss
;
1277 int free_space
= tcp_space(sk
);
1278 int full_space
= min_t(int, tp
->window_clamp
, tcp_full_space(sk
));
1281 if (mss
> full_space
)
1284 if (free_space
< full_space
/2) {
1285 icsk
->icsk_ack
.quick
= 0;
1287 if (tcp_memory_pressure
)
1288 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U*tp
->advmss
);
1290 if (free_space
< mss
)
1294 if (free_space
> tp
->rcv_ssthresh
)
1295 free_space
= tp
->rcv_ssthresh
;
1297 /* Don't do rounding if we are using window scaling, since the
1298 * scaled window will not line up with the MSS boundary anyway.
1300 window
= tp
->rcv_wnd
;
1301 if (tp
->rx_opt
.rcv_wscale
) {
1302 window
= free_space
;
1304 /* Advertise enough space so that it won't get scaled away.
1305 * Import case: prevent zero window announcement if
1306 * 1<<rcv_wscale > mss.
1308 if (((window
>> tp
->rx_opt
.rcv_wscale
) << tp
->rx_opt
.rcv_wscale
) != window
)
1309 window
= (((window
>> tp
->rx_opt
.rcv_wscale
) + 1)
1310 << tp
->rx_opt
.rcv_wscale
);
1312 /* Get the largest window that is a nice multiple of mss.
1313 * Window clamp already applied above.
1314 * If our current window offering is within 1 mss of the
1315 * free space we just keep it. This prevents the divide
1316 * and multiply from happening most of the time.
1317 * We also don't do any window rounding when the free space
1320 if (window
<= free_space
- mss
|| window
> free_space
)
1321 window
= (free_space
/mss
)*mss
;
1322 else if (mss
== full_space
&&
1323 free_space
> window
+ full_space
/2)
1324 window
= free_space
;
1330 /* Attempt to collapse two adjacent SKB's during retransmission. */
1331 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*skb
, int mss_now
)
1333 struct tcp_sock
*tp
= tcp_sk(sk
);
1334 struct sk_buff
*next_skb
= skb
->next
;
1336 /* The first test we must make is that neither of these two
1337 * SKB's are still referenced by someone else.
1339 if (!skb_cloned(skb
) && !skb_cloned(next_skb
)) {
1340 int skb_size
= skb
->len
, next_skb_size
= next_skb
->len
;
1341 u16 flags
= TCP_SKB_CB(skb
)->flags
;
1343 /* Also punt if next skb has been SACK'd. */
1344 if(TCP_SKB_CB(next_skb
)->sacked
& TCPCB_SACKED_ACKED
)
1347 /* Next skb is out of window. */
1348 if (after(TCP_SKB_CB(next_skb
)->end_seq
, tp
->snd_una
+tp
->snd_wnd
))
1351 /* Punt if not enough space exists in the first SKB for
1352 * the data in the second, or the total combined payload
1353 * would exceed the MSS.
1355 if ((next_skb_size
> skb_tailroom(skb
)) ||
1356 ((skb_size
+ next_skb_size
) > mss_now
))
1359 BUG_ON(tcp_skb_pcount(skb
) != 1 ||
1360 tcp_skb_pcount(next_skb
) != 1);
1362 /* changing transmit queue under us so clear hints */
1363 clear_all_retrans_hints(tp
);
1365 /* Ok. We will be able to collapse the packet. */
1366 __skb_unlink(next_skb
, &sk
->sk_write_queue
);
1368 memcpy(skb_put(skb
, next_skb_size
), next_skb
->data
, next_skb_size
);
1370 if (next_skb
->ip_summed
== CHECKSUM_HW
)
1371 skb
->ip_summed
= CHECKSUM_HW
;
1373 if (skb
->ip_summed
!= CHECKSUM_HW
)
1374 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
1376 /* Update sequence range on original skb. */
1377 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
1379 /* Merge over control information. */
1380 flags
|= TCP_SKB_CB(next_skb
)->flags
; /* This moves PSH/FIN etc. over */
1381 TCP_SKB_CB(skb
)->flags
= flags
;
1383 /* All done, get rid of second SKB and account for it so
1384 * packet counting does not break.
1386 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
&(TCPCB_EVER_RETRANS
|TCPCB_AT_TAIL
);
1387 if (TCP_SKB_CB(next_skb
)->sacked
&TCPCB_SACKED_RETRANS
)
1388 tp
->retrans_out
-= tcp_skb_pcount(next_skb
);
1389 if (TCP_SKB_CB(next_skb
)->sacked
&TCPCB_LOST
) {
1390 tp
->lost_out
-= tcp_skb_pcount(next_skb
);
1391 tp
->left_out
-= tcp_skb_pcount(next_skb
);
1393 /* Reno case is special. Sigh... */
1394 if (!tp
->rx_opt
.sack_ok
&& tp
->sacked_out
) {
1395 tcp_dec_pcount_approx(&tp
->sacked_out
, next_skb
);
1396 tp
->left_out
-= tcp_skb_pcount(next_skb
);
1399 /* Not quite right: it can be > snd.fack, but
1400 * it is better to underestimate fackets.
1402 tcp_dec_pcount_approx(&tp
->fackets_out
, next_skb
);
1403 tcp_packets_out_dec(tp
, next_skb
);
1404 sk_stream_free_skb(sk
, next_skb
);
1408 /* Do a simple retransmit without using the backoff mechanisms in
1409 * tcp_timer. This is used for path mtu discovery.
1410 * The socket is already locked here.
1412 void tcp_simple_retransmit(struct sock
*sk
)
1414 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1415 struct tcp_sock
*tp
= tcp_sk(sk
);
1416 struct sk_buff
*skb
;
1417 unsigned int mss
= tcp_current_mss(sk
, 0);
1420 sk_stream_for_retrans_queue(skb
, sk
) {
1421 if (skb
->len
> mss
&&
1422 !(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
)) {
1423 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) {
1424 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1425 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1427 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_LOST
)) {
1428 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1429 tp
->lost_out
+= tcp_skb_pcount(skb
);
1435 clear_all_retrans_hints(tp
);
1440 tcp_sync_left_out(tp
);
1442 /* Don't muck with the congestion window here.
1443 * Reason is that we do not increase amount of _data_
1444 * in network, but units changed and effective
1445 * cwnd/ssthresh really reduced now.
1447 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
1448 tp
->high_seq
= tp
->snd_nxt
;
1449 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
1450 tp
->prior_ssthresh
= 0;
1451 tp
->undo_marker
= 0;
1452 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1454 tcp_xmit_retransmit_queue(sk
);
1457 /* This retransmits one SKB. Policy decisions and retransmit queue
1458 * state updates are done by the caller. Returns non-zero if an
1459 * error occurred which prevented the send.
1461 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
1463 struct tcp_sock
*tp
= tcp_sk(sk
);
1464 unsigned int cur_mss
= tcp_current_mss(sk
, 0);
1467 /* Do not sent more than we queued. 1/4 is reserved for possible
1468 * copying overhead: fragmentation, tunneling, mangling etc.
1470 if (atomic_read(&sk
->sk_wmem_alloc
) >
1471 min(sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2), sk
->sk_sndbuf
))
1474 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
1475 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1477 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
1481 /* If receiver has shrunk his window, and skb is out of
1482 * new window, do not retransmit it. The exception is the
1483 * case, when window is shrunk to zero. In this case
1484 * our retransmit serves as a zero window probe.
1486 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
+tp
->snd_wnd
)
1487 && TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
1490 if (skb
->len
> cur_mss
) {
1491 if (tcp_fragment(sk
, skb
, cur_mss
, cur_mss
))
1492 return -ENOMEM
; /* We'll try again later. */
1495 /* Collapse two adjacent packets if worthwhile and we can. */
1496 if(!(TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_SYN
) &&
1497 (skb
->len
< (cur_mss
>> 1)) &&
1498 (skb
->next
!= sk
->sk_send_head
) &&
1499 (skb
->next
!= (struct sk_buff
*)&sk
->sk_write_queue
) &&
1500 (skb_shinfo(skb
)->nr_frags
== 0 && skb_shinfo(skb
->next
)->nr_frags
== 0) &&
1501 (tcp_skb_pcount(skb
) == 1 && tcp_skb_pcount(skb
->next
) == 1) &&
1502 (sysctl_tcp_retrans_collapse
!= 0))
1503 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
1505 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
1506 return -EHOSTUNREACH
; /* Routing failure or similar. */
1508 /* Some Solaris stacks overoptimize and ignore the FIN on a
1509 * retransmit when old data is attached. So strip it off
1510 * since it is cheap to do so and saves bytes on the network.
1513 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
) &&
1514 tp
->snd_una
== (TCP_SKB_CB(skb
)->end_seq
- 1)) {
1515 if (!pskb_trim(skb
, 0)) {
1516 TCP_SKB_CB(skb
)->seq
= TCP_SKB_CB(skb
)->end_seq
- 1;
1517 skb_shinfo(skb
)->tso_segs
= 1;
1518 skb_shinfo(skb
)->tso_size
= 0;
1519 skb
->ip_summed
= CHECKSUM_NONE
;
1524 /* Make a copy, if the first transmission SKB clone we made
1525 * is still in somebody's hands, else make a clone.
1527 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1529 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
1532 /* Update global TCP statistics. */
1533 TCP_INC_STATS(TCP_MIB_RETRANSSEGS
);
1535 tp
->total_retrans
++;
1537 #if FASTRETRANS_DEBUG > 0
1538 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) {
1539 if (net_ratelimit())
1540 printk(KERN_DEBUG
"retrans_out leaked.\n");
1543 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
1544 tp
->retrans_out
+= tcp_skb_pcount(skb
);
1546 /* Save stamp of the first retransmit. */
1547 if (!tp
->retrans_stamp
)
1548 tp
->retrans_stamp
= TCP_SKB_CB(skb
)->when
;
1552 /* snd_nxt is stored to detect loss of retransmitted segment,
1553 * see tcp_input.c tcp_sacktag_write_queue().
1555 TCP_SKB_CB(skb
)->ack_seq
= tp
->snd_nxt
;
1560 /* This gets called after a retransmit timeout, and the initially
1561 * retransmitted data is acknowledged. It tries to continue
1562 * resending the rest of the retransmit queue, until either
1563 * we've sent it all or the congestion window limit is reached.
1564 * If doing SACK, the first ACK which comes back for a timeout
1565 * based retransmit packet might feed us FACK information again.
1566 * If so, we use it to avoid unnecessarily retransmissions.
1568 void tcp_xmit_retransmit_queue(struct sock
*sk
)
1570 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1571 struct tcp_sock
*tp
= tcp_sk(sk
);
1572 struct sk_buff
*skb
;
1575 if (tp
->retransmit_skb_hint
) {
1576 skb
= tp
->retransmit_skb_hint
;
1577 packet_cnt
= tp
->retransmit_cnt_hint
;
1579 skb
= sk
->sk_write_queue
.next
;
1583 /* First pass: retransmit lost packets. */
1585 sk_stream_for_retrans_queue_from(skb
, sk
) {
1586 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
1588 /* we could do better than to assign each time */
1589 tp
->retransmit_skb_hint
= skb
;
1590 tp
->retransmit_cnt_hint
= packet_cnt
;
1592 /* Assume this retransmit will generate
1593 * only one packet for congestion window
1594 * calculation purposes. This works because
1595 * tcp_retransmit_skb() will chop up the
1596 * packet to be MSS sized and all the
1597 * packet counting works out.
1599 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
1602 if (sacked
& TCPCB_LOST
) {
1603 if (!(sacked
&(TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))) {
1604 if (tcp_retransmit_skb(sk
, skb
)) {
1605 tp
->retransmit_skb_hint
= NULL
;
1608 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
1609 NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS
);
1611 NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS
);
1614 skb_peek(&sk
->sk_write_queue
))
1615 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1616 inet_csk(sk
)->icsk_rto
,
1620 packet_cnt
+= tcp_skb_pcount(skb
);
1621 if (packet_cnt
>= tp
->lost_out
)
1627 /* OK, demanded retransmission is finished. */
1629 /* Forward retransmissions are possible only during Recovery. */
1630 if (icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
1633 /* No forward retransmissions in Reno are possible. */
1634 if (!tp
->rx_opt
.sack_ok
)
1637 /* Yeah, we have to make difficult choice between forward transmission
1638 * and retransmission... Both ways have their merits...
1640 * For now we do not retransmit anything, while we have some new
1644 if (tcp_may_send_now(sk
, tp
))
1647 if (tp
->forward_skb_hint
) {
1648 skb
= tp
->forward_skb_hint
;
1649 packet_cnt
= tp
->forward_cnt_hint
;
1651 skb
= sk
->sk_write_queue
.next
;
1655 sk_stream_for_retrans_queue_from(skb
, sk
) {
1656 tp
->forward_cnt_hint
= packet_cnt
;
1657 tp
->forward_skb_hint
= skb
;
1659 /* Similar to the retransmit loop above we
1660 * can pretend that the retransmitted SKB
1661 * we send out here will be composed of one
1662 * real MSS sized packet because tcp_retransmit_skb()
1663 * will fragment it if necessary.
1665 if (++packet_cnt
> tp
->fackets_out
)
1668 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
1671 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
)
1674 /* Ok, retransmit it. */
1675 if (tcp_retransmit_skb(sk
, skb
)) {
1676 tp
->forward_skb_hint
= NULL
;
1680 if (skb
== skb_peek(&sk
->sk_write_queue
))
1681 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1682 inet_csk(sk
)->icsk_rto
,
1685 NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS
);
1690 /* Send a fin. The caller locks the socket for us. This cannot be
1691 * allowed to fail queueing a FIN frame under any circumstances.
1693 void tcp_send_fin(struct sock
*sk
)
1695 struct tcp_sock
*tp
= tcp_sk(sk
);
1696 struct sk_buff
*skb
= skb_peek_tail(&sk
->sk_write_queue
);
1699 /* Optimization, tack on the FIN if we have a queue of
1700 * unsent frames. But be careful about outgoing SACKS
1703 mss_now
= tcp_current_mss(sk
, 1);
1705 if (sk
->sk_send_head
!= NULL
) {
1706 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_FIN
;
1707 TCP_SKB_CB(skb
)->end_seq
++;
1710 /* Socket is locked, keep trying until memory is available. */
1712 skb
= alloc_skb_fclone(MAX_TCP_HEADER
, GFP_KERNEL
);
1718 /* Reserve space for headers and prepare control bits. */
1719 skb_reserve(skb
, MAX_TCP_HEADER
);
1721 TCP_SKB_CB(skb
)->flags
= (TCPCB_FLAG_ACK
| TCPCB_FLAG_FIN
);
1722 TCP_SKB_CB(skb
)->sacked
= 0;
1723 skb_shinfo(skb
)->tso_segs
= 1;
1724 skb_shinfo(skb
)->tso_size
= 0;
1726 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
1727 TCP_SKB_CB(skb
)->seq
= tp
->write_seq
;
1728 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ 1;
1729 tcp_queue_skb(sk
, skb
);
1731 __tcp_push_pending_frames(sk
, tp
, mss_now
, TCP_NAGLE_OFF
);
1734 /* We get here when a process closes a file descriptor (either due to
1735 * an explicit close() or as a byproduct of exit()'ing) and there
1736 * was unread data in the receive queue. This behavior is recommended
1737 * by draft-ietf-tcpimpl-prob-03.txt section 3.10. -DaveM
1739 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
1741 struct tcp_sock
*tp
= tcp_sk(sk
);
1742 struct sk_buff
*skb
;
1744 /* NOTE: No TCP options attached and we never retransmit this. */
1745 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
1747 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED
);
1751 /* Reserve space for headers and prepare control bits. */
1752 skb_reserve(skb
, MAX_TCP_HEADER
);
1754 TCP_SKB_CB(skb
)->flags
= (TCPCB_FLAG_ACK
| TCPCB_FLAG_RST
);
1755 TCP_SKB_CB(skb
)->sacked
= 0;
1756 skb_shinfo(skb
)->tso_segs
= 1;
1757 skb_shinfo(skb
)->tso_size
= 0;
1760 TCP_SKB_CB(skb
)->seq
= tcp_acceptable_seq(sk
, tp
);
1761 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
;
1762 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1763 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
1764 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED
);
1767 /* WARNING: This routine must only be called when we have already sent
1768 * a SYN packet that crossed the incoming SYN that caused this routine
1769 * to get called. If this assumption fails then the initial rcv_wnd
1770 * and rcv_wscale values will not be correct.
1772 int tcp_send_synack(struct sock
*sk
)
1774 struct sk_buff
* skb
;
1776 skb
= skb_peek(&sk
->sk_write_queue
);
1777 if (skb
== NULL
|| !(TCP_SKB_CB(skb
)->flags
&TCPCB_FLAG_SYN
)) {
1778 printk(KERN_DEBUG
"tcp_send_synack: wrong queue state\n");
1781 if (!(TCP_SKB_CB(skb
)->flags
&TCPCB_FLAG_ACK
)) {
1782 if (skb_cloned(skb
)) {
1783 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
1786 __skb_unlink(skb
, &sk
->sk_write_queue
);
1787 skb_header_release(nskb
);
1788 __skb_queue_head(&sk
->sk_write_queue
, nskb
);
1789 sk_stream_free_skb(sk
, skb
);
1790 sk_charge_skb(sk
, nskb
);
1794 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_ACK
;
1795 TCP_ECN_send_synack(tcp_sk(sk
), skb
);
1797 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1798 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
1802 * Prepare a SYN-ACK.
1804 struct sk_buff
* tcp_make_synack(struct sock
*sk
, struct dst_entry
*dst
,
1805 struct request_sock
*req
)
1807 struct inet_request_sock
*ireq
= inet_rsk(req
);
1808 struct tcp_sock
*tp
= tcp_sk(sk
);
1810 int tcp_header_size
;
1811 struct sk_buff
*skb
;
1813 skb
= sock_wmalloc(sk
, MAX_TCP_HEADER
+ 15, 1, GFP_ATOMIC
);
1817 /* Reserve space for headers. */
1818 skb_reserve(skb
, MAX_TCP_HEADER
);
1820 skb
->dst
= dst_clone(dst
);
1822 tcp_header_size
= (sizeof(struct tcphdr
) + TCPOLEN_MSS
+
1823 (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0) +
1824 (ireq
->wscale_ok
? TCPOLEN_WSCALE_ALIGNED
: 0) +
1825 /* SACK_PERM is in the place of NOP NOP of TS */
1826 ((ireq
->sack_ok
&& !ireq
->tstamp_ok
) ? TCPOLEN_SACKPERM_ALIGNED
: 0));
1827 skb
->h
.th
= th
= (struct tcphdr
*) skb_push(skb
, tcp_header_size
);
1829 memset(th
, 0, sizeof(struct tcphdr
));
1832 if (dst
->dev
->features
&NETIF_F_TSO
)
1834 TCP_ECN_make_synack(req
, th
);
1835 th
->source
= inet_sk(sk
)->sport
;
1836 th
->dest
= ireq
->rmt_port
;
1837 TCP_SKB_CB(skb
)->seq
= tcp_rsk(req
)->snt_isn
;
1838 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ 1;
1839 TCP_SKB_CB(skb
)->sacked
= 0;
1840 skb_shinfo(skb
)->tso_segs
= 1;
1841 skb_shinfo(skb
)->tso_size
= 0;
1842 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
1843 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_isn
+ 1);
1844 if (req
->rcv_wnd
== 0) { /* ignored for retransmitted syns */
1846 /* Set this up on the first call only */
1847 req
->window_clamp
= tp
->window_clamp
? : dst_metric(dst
, RTAX_WINDOW
);
1848 /* tcp_full_space because it is guaranteed to be the first packet */
1849 tcp_select_initial_window(tcp_full_space(sk
),
1850 dst_metric(dst
, RTAX_ADVMSS
) - (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0),
1855 ireq
->rcv_wscale
= rcv_wscale
;
1858 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
1859 th
->window
= htons(min(req
->rcv_wnd
, 65535U));
1861 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1862 tcp_syn_build_options((__u32
*)(th
+ 1), dst_metric(dst
, RTAX_ADVMSS
), ireq
->tstamp_ok
,
1863 ireq
->sack_ok
, ireq
->wscale_ok
, ireq
->rcv_wscale
,
1864 TCP_SKB_CB(skb
)->when
,
1868 th
->doff
= (tcp_header_size
>> 2);
1869 TCP_INC_STATS(TCP_MIB_OUTSEGS
);
1874 * Do all connect socket setups that can be done AF independent.
1876 static void tcp_connect_init(struct sock
*sk
)
1878 struct dst_entry
*dst
= __sk_dst_get(sk
);
1879 struct tcp_sock
*tp
= tcp_sk(sk
);
1882 /* We'll fix this up when we get a response from the other end.
1883 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
1885 tp
->tcp_header_len
= sizeof(struct tcphdr
) +
1886 (sysctl_tcp_timestamps
? TCPOLEN_TSTAMP_ALIGNED
: 0);
1888 /* If user gave his TCP_MAXSEG, record it to clamp */
1889 if (tp
->rx_opt
.user_mss
)
1890 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
1892 tcp_sync_mss(sk
, dst_mtu(dst
));
1894 if (!tp
->window_clamp
)
1895 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
1896 tp
->advmss
= dst_metric(dst
, RTAX_ADVMSS
);
1897 tcp_initialize_rcv_mss(sk
);
1899 tcp_select_initial_window(tcp_full_space(sk
),
1900 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
1903 sysctl_tcp_window_scaling
,
1906 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
1907 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
1910 sock_reset_flag(sk
, SOCK_DONE
);
1912 tcp_init_wl(tp
, tp
->write_seq
, 0);
1913 tp
->snd_una
= tp
->write_seq
;
1914 tp
->snd_sml
= tp
->write_seq
;
1919 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
1920 inet_csk(sk
)->icsk_retransmits
= 0;
1921 tcp_clear_retrans(tp
);
1925 * Build a SYN and send it off.
1927 int tcp_connect(struct sock
*sk
)
1929 struct tcp_sock
*tp
= tcp_sk(sk
);
1930 struct sk_buff
*buff
;
1932 tcp_connect_init(sk
);
1934 buff
= alloc_skb_fclone(MAX_TCP_HEADER
+ 15, sk
->sk_allocation
);
1935 if (unlikely(buff
== NULL
))
1938 /* Reserve space for headers. */
1939 skb_reserve(buff
, MAX_TCP_HEADER
);
1941 TCP_SKB_CB(buff
)->flags
= TCPCB_FLAG_SYN
;
1942 TCP_ECN_send_syn(sk
, tp
, buff
);
1943 TCP_SKB_CB(buff
)->sacked
= 0;
1944 skb_shinfo(buff
)->tso_segs
= 1;
1945 skb_shinfo(buff
)->tso_size
= 0;
1947 TCP_SKB_CB(buff
)->seq
= tp
->write_seq
++;
1948 TCP_SKB_CB(buff
)->end_seq
= tp
->write_seq
;
1949 tp
->snd_nxt
= tp
->write_seq
;
1950 tp
->pushed_seq
= tp
->write_seq
;
1953 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
1954 tp
->retrans_stamp
= TCP_SKB_CB(buff
)->when
;
1955 skb_header_release(buff
);
1956 __skb_queue_tail(&sk
->sk_write_queue
, buff
);
1957 sk_charge_skb(sk
, buff
);
1958 tp
->packets_out
+= tcp_skb_pcount(buff
);
1959 tcp_transmit_skb(sk
, buff
, 1, GFP_KERNEL
);
1960 TCP_INC_STATS(TCP_MIB_ACTIVEOPENS
);
1962 /* Timer for repeating the SYN until an answer. */
1963 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1964 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
1968 /* Send out a delayed ack, the caller does the policy checking
1969 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
1972 void tcp_send_delayed_ack(struct sock
*sk
)
1974 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1975 int ato
= icsk
->icsk_ack
.ato
;
1976 unsigned long timeout
;
1978 if (ato
> TCP_DELACK_MIN
) {
1979 const struct tcp_sock
*tp
= tcp_sk(sk
);
1982 if (icsk
->icsk_ack
.pingpong
|| (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
1983 max_ato
= TCP_DELACK_MAX
;
1985 /* Slow path, intersegment interval is "high". */
1987 /* If some rtt estimate is known, use it to bound delayed ack.
1988 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
1992 int rtt
= max(tp
->srtt
>>3, TCP_DELACK_MIN
);
1998 ato
= min(ato
, max_ato
);
2001 /* Stay within the limit we were given */
2002 timeout
= jiffies
+ ato
;
2004 /* Use new timeout only if there wasn't a older one earlier. */
2005 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
2006 /* If delack timer was blocked or is about to expire,
2009 if (icsk
->icsk_ack
.blocked
||
2010 time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
2015 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
2016 timeout
= icsk
->icsk_ack
.timeout
;
2018 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
2019 icsk
->icsk_ack
.timeout
= timeout
;
2020 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
2023 /* This routine sends an ack and also updates the window. */
2024 void tcp_send_ack(struct sock
*sk
)
2026 /* If we have been reset, we may not send again. */
2027 if (sk
->sk_state
!= TCP_CLOSE
) {
2028 struct tcp_sock
*tp
= tcp_sk(sk
);
2029 struct sk_buff
*buff
;
2031 /* We are not putting this on the write queue, so
2032 * tcp_transmit_skb() will set the ownership to this
2035 buff
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2037 inet_csk_schedule_ack(sk
);
2038 inet_csk(sk
)->icsk_ack
.ato
= TCP_ATO_MIN
;
2039 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
2040 TCP_DELACK_MAX
, TCP_RTO_MAX
);
2044 /* Reserve space for headers and prepare control bits. */
2045 skb_reserve(buff
, MAX_TCP_HEADER
);
2047 TCP_SKB_CB(buff
)->flags
= TCPCB_FLAG_ACK
;
2048 TCP_SKB_CB(buff
)->sacked
= 0;
2049 skb_shinfo(buff
)->tso_segs
= 1;
2050 skb_shinfo(buff
)->tso_size
= 0;
2052 /* Send it off, this clears delayed acks for us. */
2053 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(buff
)->end_seq
= tcp_acceptable_seq(sk
, tp
);
2054 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2055 tcp_transmit_skb(sk
, buff
, 0, GFP_ATOMIC
);
2059 /* This routine sends a packet with an out of date sequence
2060 * number. It assumes the other end will try to ack it.
2062 * Question: what should we make while urgent mode?
2063 * 4.4BSD forces sending single byte of data. We cannot send
2064 * out of window data, because we have SND.NXT==SND.MAX...
2066 * Current solution: to send TWO zero-length segments in urgent mode:
2067 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2068 * out-of-date with SND.UNA-1 to probe window.
2070 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
)
2072 struct tcp_sock
*tp
= tcp_sk(sk
);
2073 struct sk_buff
*skb
;
2075 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2076 skb
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2080 /* Reserve space for headers and set control bits. */
2081 skb_reserve(skb
, MAX_TCP_HEADER
);
2083 TCP_SKB_CB(skb
)->flags
= TCPCB_FLAG_ACK
;
2084 TCP_SKB_CB(skb
)->sacked
= urgent
;
2085 skb_shinfo(skb
)->tso_segs
= 1;
2086 skb_shinfo(skb
)->tso_size
= 0;
2088 /* Use a previous sequence. This should cause the other
2089 * end to send an ack. Don't queue or clone SKB, just
2092 TCP_SKB_CB(skb
)->seq
= urgent
? tp
->snd_una
: tp
->snd_una
- 1;
2093 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
;
2094 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2095 return tcp_transmit_skb(sk
, skb
, 0, GFP_ATOMIC
);
2098 int tcp_write_wakeup(struct sock
*sk
)
2100 if (sk
->sk_state
!= TCP_CLOSE
) {
2101 struct tcp_sock
*tp
= tcp_sk(sk
);
2102 struct sk_buff
*skb
;
2104 if ((skb
= sk
->sk_send_head
) != NULL
&&
2105 before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
+tp
->snd_wnd
)) {
2107 unsigned int mss
= tcp_current_mss(sk
, 0);
2108 unsigned int seg_size
= tp
->snd_una
+tp
->snd_wnd
-TCP_SKB_CB(skb
)->seq
;
2110 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
2111 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
2113 /* We are probing the opening of a window
2114 * but the window size is != 0
2115 * must have been a result SWS avoidance ( sender )
2117 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
2119 seg_size
= min(seg_size
, mss
);
2120 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2121 if (tcp_fragment(sk
, skb
, seg_size
, mss
))
2123 } else if (!tcp_skb_pcount(skb
))
2124 tcp_set_skb_tso_segs(sk
, skb
, mss
);
2126 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2127 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2128 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2130 update_send_head(sk
, tp
, skb
);
2135 between(tp
->snd_up
, tp
->snd_una
+1, tp
->snd_una
+0xFFFF))
2136 tcp_xmit_probe_skb(sk
, TCPCB_URG
);
2137 return tcp_xmit_probe_skb(sk
, 0);
2143 /* A window probe timeout has occurred. If window is not closed send
2144 * a partial packet else a zero probe.
2146 void tcp_send_probe0(struct sock
*sk
)
2148 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2149 struct tcp_sock
*tp
= tcp_sk(sk
);
2152 err
= tcp_write_wakeup(sk
);
2154 if (tp
->packets_out
|| !sk
->sk_send_head
) {
2155 /* Cancel probe timer, if it is not required. */
2156 icsk
->icsk_probes_out
= 0;
2157 icsk
->icsk_backoff
= 0;
2162 if (icsk
->icsk_backoff
< sysctl_tcp_retries2
)
2163 icsk
->icsk_backoff
++;
2164 icsk
->icsk_probes_out
++;
2165 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2166 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
2169 /* If packet was not sent due to local congestion,
2170 * do not backoff and do not remember icsk_probes_out.
2171 * Let local senders to fight for local resources.
2173 * Use accumulated backoff yet.
2175 if (!icsk
->icsk_probes_out
)
2176 icsk
->icsk_probes_out
= 1;
2177 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2178 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
,
2179 TCP_RESOURCE_PROBE_INTERVAL
),
2184 EXPORT_SYMBOL(tcp_connect
);
2185 EXPORT_SYMBOL(tcp_make_synack
);
2186 EXPORT_SYMBOL(tcp_simple_retransmit
);
2187 EXPORT_SYMBOL(tcp_sync_mss
);
2188 EXPORT_SYMBOL(sysctl_tcp_tso_win_divisor
);