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 int sysctl_tcp_mtu_probing
= 0;
55 int sysctl_tcp_base_mss
= 512;
57 EXPORT_SYMBOL(sysctl_tcp_mtu_probing
);
58 EXPORT_SYMBOL(sysctl_tcp_base_mss
);
60 static void update_send_head(struct sock
*sk
, struct tcp_sock
*tp
,
63 sk
->sk_send_head
= skb
->next
;
64 if (sk
->sk_send_head
== (struct sk_buff
*)&sk
->sk_write_queue
)
65 sk
->sk_send_head
= NULL
;
66 tp
->snd_nxt
= TCP_SKB_CB(skb
)->end_seq
;
67 tcp_packets_out_inc(sk
, tp
, skb
);
70 /* SND.NXT, if window was not shrunk.
71 * If window has been shrunk, what should we make? It is not clear at all.
72 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
73 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
74 * invalid. OK, let's make this for now:
76 static inline __u32
tcp_acceptable_seq(struct sock
*sk
, struct tcp_sock
*tp
)
78 if (!before(tp
->snd_una
+tp
->snd_wnd
, tp
->snd_nxt
))
81 return tp
->snd_una
+tp
->snd_wnd
;
84 /* Calculate mss to advertise in SYN segment.
85 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
87 * 1. It is independent of path mtu.
88 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
89 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
90 * attached devices, because some buggy hosts are confused by
92 * 4. We do not make 3, we advertise MSS, calculated from first
93 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
94 * This may be overridden via information stored in routing table.
95 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
96 * probably even Jumbo".
98 static __u16
tcp_advertise_mss(struct sock
*sk
)
100 struct tcp_sock
*tp
= tcp_sk(sk
);
101 struct dst_entry
*dst
= __sk_dst_get(sk
);
102 int mss
= tp
->advmss
;
104 if (dst
&& dst_metric(dst
, RTAX_ADVMSS
) < mss
) {
105 mss
= dst_metric(dst
, RTAX_ADVMSS
);
112 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
113 * This is the first part of cwnd validation mechanism. */
114 static void tcp_cwnd_restart(struct sock
*sk
, struct dst_entry
*dst
)
116 struct tcp_sock
*tp
= tcp_sk(sk
);
117 s32 delta
= tcp_time_stamp
- tp
->lsndtime
;
118 u32 restart_cwnd
= tcp_init_cwnd(tp
, dst
);
119 u32 cwnd
= tp
->snd_cwnd
;
121 tcp_ca_event(sk
, CA_EVENT_CWND_RESTART
);
123 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
124 restart_cwnd
= min(restart_cwnd
, cwnd
);
126 while ((delta
-= inet_csk(sk
)->icsk_rto
) > 0 && cwnd
> restart_cwnd
)
128 tp
->snd_cwnd
= max(cwnd
, restart_cwnd
);
129 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
130 tp
->snd_cwnd_used
= 0;
133 static void tcp_event_data_sent(struct tcp_sock
*tp
,
134 struct sk_buff
*skb
, struct sock
*sk
)
136 struct inet_connection_sock
*icsk
= inet_csk(sk
);
137 const u32 now
= tcp_time_stamp
;
139 if (!tp
->packets_out
&& (s32
)(now
- tp
->lsndtime
) > icsk
->icsk_rto
)
140 tcp_cwnd_restart(sk
, __sk_dst_get(sk
));
144 /* If it is a reply for ato after last received
145 * packet, enter pingpong mode.
147 if ((u32
)(now
- icsk
->icsk_ack
.lrcvtime
) < icsk
->icsk_ack
.ato
)
148 icsk
->icsk_ack
.pingpong
= 1;
151 static inline void tcp_event_ack_sent(struct sock
*sk
, unsigned int pkts
)
153 tcp_dec_quickack_mode(sk
, pkts
);
154 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_DACK
);
157 /* Determine a window scaling and initial window to offer.
158 * Based on the assumption that the given amount of space
159 * will be offered. Store the results in the tp structure.
160 * NOTE: for smooth operation initial space offering should
161 * be a multiple of mss if possible. We assume here that mss >= 1.
162 * This MUST be enforced by all callers.
164 void tcp_select_initial_window(int __space
, __u32 mss
,
165 __u32
*rcv_wnd
, __u32
*window_clamp
,
166 int wscale_ok
, __u8
*rcv_wscale
)
168 unsigned int space
= (__space
< 0 ? 0 : __space
);
170 /* If no clamp set the clamp to the max possible scaled window */
171 if (*window_clamp
== 0)
172 (*window_clamp
) = (65535 << 14);
173 space
= min(*window_clamp
, space
);
175 /* Quantize space offering to a multiple of mss if possible. */
177 space
= (space
/ mss
) * mss
;
179 /* NOTE: offering an initial window larger than 32767
180 * will break some buggy TCP stacks. We try to be nice.
181 * If we are not window scaling, then this truncates
182 * our initial window offering to 32k. There should also
183 * be a sysctl option to stop being nice.
185 (*rcv_wnd
) = min(space
, MAX_TCP_WINDOW
);
188 /* Set window scaling on max possible window
189 * See RFC1323 for an explanation of the limit to 14
191 space
= max_t(u32
, sysctl_tcp_rmem
[2], sysctl_rmem_max
);
192 while (space
> 65535 && (*rcv_wscale
) < 14) {
198 /* Set initial window to value enough for senders,
199 * following RFC2414. Senders, not following this RFC,
200 * will be satisfied with 2.
202 if (mss
> (1<<*rcv_wscale
)) {
208 if (*rcv_wnd
> init_cwnd
*mss
)
209 *rcv_wnd
= init_cwnd
*mss
;
212 /* Set the clamp no higher than max representable value */
213 (*window_clamp
) = min(65535U << (*rcv_wscale
), *window_clamp
);
216 /* Chose a new window to advertise, update state in tcp_sock for the
217 * socket, and return result with RFC1323 scaling applied. The return
218 * value can be stuffed directly into th->window for an outgoing
221 static u16
tcp_select_window(struct sock
*sk
)
223 struct tcp_sock
*tp
= tcp_sk(sk
);
224 u32 cur_win
= tcp_receive_window(tp
);
225 u32 new_win
= __tcp_select_window(sk
);
227 /* Never shrink the offered window */
228 if(new_win
< cur_win
) {
229 /* Danger Will Robinson!
230 * Don't update rcv_wup/rcv_wnd here or else
231 * we will not be able to advertise a zero
232 * window in time. --DaveM
234 * Relax Will Robinson.
238 tp
->rcv_wnd
= new_win
;
239 tp
->rcv_wup
= tp
->rcv_nxt
;
241 /* Make sure we do not exceed the maximum possible
244 if (!tp
->rx_opt
.rcv_wscale
)
245 new_win
= min(new_win
, MAX_TCP_WINDOW
);
247 new_win
= min(new_win
, (65535U << tp
->rx_opt
.rcv_wscale
));
249 /* RFC1323 scaling applied */
250 new_win
>>= tp
->rx_opt
.rcv_wscale
;
252 /* If we advertise zero window, disable fast path. */
259 static void tcp_build_and_update_options(__u32
*ptr
, struct tcp_sock
*tp
,
262 if (tp
->rx_opt
.tstamp_ok
) {
263 *ptr
++ = __constant_htonl((TCPOPT_NOP
<< 24) |
265 (TCPOPT_TIMESTAMP
<< 8) |
267 *ptr
++ = htonl(tstamp
);
268 *ptr
++ = htonl(tp
->rx_opt
.ts_recent
);
270 if (tp
->rx_opt
.eff_sacks
) {
271 struct tcp_sack_block
*sp
= tp
->rx_opt
.dsack
? tp
->duplicate_sack
: tp
->selective_acks
;
274 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
277 (TCPOLEN_SACK_BASE
+ (tp
->rx_opt
.eff_sacks
*
278 TCPOLEN_SACK_PERBLOCK
)));
279 for(this_sack
= 0; this_sack
< tp
->rx_opt
.eff_sacks
; this_sack
++) {
280 *ptr
++ = htonl(sp
[this_sack
].start_seq
);
281 *ptr
++ = htonl(sp
[this_sack
].end_seq
);
283 if (tp
->rx_opt
.dsack
) {
284 tp
->rx_opt
.dsack
= 0;
285 tp
->rx_opt
.eff_sacks
--;
290 /* Construct a tcp options header for a SYN or SYN_ACK packet.
291 * If this is every changed make sure to change the definition of
292 * MAX_SYN_SIZE to match the new maximum number of options that you
295 static void tcp_syn_build_options(__u32
*ptr
, int mss
, int ts
, int sack
,
296 int offer_wscale
, int wscale
, __u32 tstamp
,
299 /* We always get an MSS option.
300 * The option bytes which will be seen in normal data
301 * packets should timestamps be used, must be in the MSS
302 * advertised. But we subtract them from tp->mss_cache so
303 * that calculations in tcp_sendmsg are simpler etc.
304 * So account for this fact here if necessary. If we
305 * don't do this correctly, as a receiver we won't
306 * recognize data packets as being full sized when we
307 * should, and thus we won't abide by the delayed ACK
309 * SACKs don't matter, we never delay an ACK when we
310 * have any of those going out.
312 *ptr
++ = htonl((TCPOPT_MSS
<< 24) | (TCPOLEN_MSS
<< 16) | mss
);
315 *ptr
++ = __constant_htonl((TCPOPT_SACK_PERM
<< 24) | (TCPOLEN_SACK_PERM
<< 16) |
316 (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
);
318 *ptr
++ = __constant_htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16) |
319 (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
);
320 *ptr
++ = htonl(tstamp
); /* TSVAL */
321 *ptr
++ = htonl(ts_recent
); /* TSECR */
323 *ptr
++ = __constant_htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16) |
324 (TCPOPT_SACK_PERM
<< 8) | TCPOLEN_SACK_PERM
);
326 *ptr
++ = htonl((TCPOPT_NOP
<< 24) | (TCPOPT_WINDOW
<< 16) | (TCPOLEN_WINDOW
<< 8) | (wscale
));
329 /* This routine actually transmits TCP packets queued in by
330 * tcp_do_sendmsg(). This is used by both the initial
331 * transmission and possible later retransmissions.
332 * All SKB's seen here are completely headerless. It is our
333 * job to build the TCP header, and pass the packet down to
334 * IP so it can do the same plus pass the packet off to the
337 * We are working here with either a clone of the original
338 * SKB, or a fresh unique copy made by the retransmit engine.
340 static int tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int clone_it
, gfp_t gfp_mask
)
342 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
343 struct inet_sock
*inet
;
345 struct tcp_skb_cb
*tcb
;
351 BUG_ON(!skb
|| !tcp_skb_pcount(skb
));
353 /* If congestion control is doing timestamping, we must
354 * take such a timestamp before we potentially clone/copy.
356 if (icsk
->icsk_ca_ops
->rtt_sample
)
357 __net_timestamp(skb
);
359 if (likely(clone_it
)) {
360 if (unlikely(skb_cloned(skb
)))
361 skb
= pskb_copy(skb
, gfp_mask
);
363 skb
= skb_clone(skb
, gfp_mask
);
370 tcb
= TCP_SKB_CB(skb
);
371 tcp_header_size
= tp
->tcp_header_len
;
373 #define SYSCTL_FLAG_TSTAMPS 0x1
374 #define SYSCTL_FLAG_WSCALE 0x2
375 #define SYSCTL_FLAG_SACK 0x4
378 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
379 tcp_header_size
= sizeof(struct tcphdr
) + TCPOLEN_MSS
;
380 if(sysctl_tcp_timestamps
) {
381 tcp_header_size
+= TCPOLEN_TSTAMP_ALIGNED
;
382 sysctl_flags
|= SYSCTL_FLAG_TSTAMPS
;
384 if (sysctl_tcp_window_scaling
) {
385 tcp_header_size
+= TCPOLEN_WSCALE_ALIGNED
;
386 sysctl_flags
|= SYSCTL_FLAG_WSCALE
;
388 if (sysctl_tcp_sack
) {
389 sysctl_flags
|= SYSCTL_FLAG_SACK
;
390 if (!(sysctl_flags
& SYSCTL_FLAG_TSTAMPS
))
391 tcp_header_size
+= TCPOLEN_SACKPERM_ALIGNED
;
393 } else if (unlikely(tp
->rx_opt
.eff_sacks
)) {
394 /* A SACK is 2 pad bytes, a 2 byte header, plus
395 * 2 32-bit sequence numbers for each SACK block.
397 tcp_header_size
+= (TCPOLEN_SACK_BASE_ALIGNED
+
398 (tp
->rx_opt
.eff_sacks
*
399 TCPOLEN_SACK_PERBLOCK
));
402 if (tcp_packets_in_flight(tp
) == 0)
403 tcp_ca_event(sk
, CA_EVENT_TX_START
);
405 th
= (struct tcphdr
*) skb_push(skb
, tcp_header_size
);
407 skb_set_owner_w(skb
, sk
);
409 /* Build TCP header and checksum it. */
410 th
->source
= inet
->sport
;
411 th
->dest
= inet
->dport
;
412 th
->seq
= htonl(tcb
->seq
);
413 th
->ack_seq
= htonl(tp
->rcv_nxt
);
414 *(((__u16
*)th
) + 6) = htons(((tcp_header_size
>> 2) << 12) |
417 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
418 /* RFC1323: The window in SYN & SYN/ACK segments
421 th
->window
= htons(tp
->rcv_wnd
);
423 th
->window
= htons(tcp_select_window(sk
));
428 if (unlikely(tp
->urg_mode
&&
429 between(tp
->snd_up
, tcb
->seq
+1, tcb
->seq
+0xFFFF))) {
430 th
->urg_ptr
= htons(tp
->snd_up
-tcb
->seq
);
434 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
435 tcp_syn_build_options((__u32
*)(th
+ 1),
436 tcp_advertise_mss(sk
),
437 (sysctl_flags
& SYSCTL_FLAG_TSTAMPS
),
438 (sysctl_flags
& SYSCTL_FLAG_SACK
),
439 (sysctl_flags
& SYSCTL_FLAG_WSCALE
),
440 tp
->rx_opt
.rcv_wscale
,
442 tp
->rx_opt
.ts_recent
);
444 tcp_build_and_update_options((__u32
*)(th
+ 1),
446 TCP_ECN_send(sk
, tp
, skb
, tcp_header_size
);
449 icsk
->icsk_af_ops
->send_check(sk
, skb
->len
, skb
);
451 if (likely(tcb
->flags
& TCPCB_FLAG_ACK
))
452 tcp_event_ack_sent(sk
, tcp_skb_pcount(skb
));
454 if (skb
->len
!= tcp_header_size
)
455 tcp_event_data_sent(tp
, skb
, sk
);
457 TCP_INC_STATS(TCP_MIB_OUTSEGS
);
459 err
= icsk
->icsk_af_ops
->queue_xmit(skb
, 0);
460 if (unlikely(err
<= 0))
465 /* NET_XMIT_CN is special. It does not guarantee,
466 * that this packet is lost. It tells that device
467 * is about to start to drop packets or already
468 * drops some packets of the same priority and
469 * invokes us to send less aggressively.
471 return err
== NET_XMIT_CN
? 0 : err
;
473 #undef SYSCTL_FLAG_TSTAMPS
474 #undef SYSCTL_FLAG_WSCALE
475 #undef SYSCTL_FLAG_SACK
479 /* This routine just queue's the buffer
481 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
482 * otherwise socket can stall.
484 static void tcp_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
486 struct tcp_sock
*tp
= tcp_sk(sk
);
488 /* Advance write_seq and place onto the write_queue. */
489 tp
->write_seq
= TCP_SKB_CB(skb
)->end_seq
;
490 skb_header_release(skb
);
491 __skb_queue_tail(&sk
->sk_write_queue
, skb
);
492 sk_charge_skb(sk
, skb
);
494 /* Queue it, remembering where we must start sending. */
495 if (sk
->sk_send_head
== NULL
)
496 sk
->sk_send_head
= skb
;
499 static void tcp_set_skb_tso_segs(struct sock
*sk
, struct sk_buff
*skb
, unsigned int mss_now
)
501 if (skb
->len
<= mss_now
||
502 !(sk
->sk_route_caps
& NETIF_F_TSO
)) {
503 /* Avoid the costly divide in the normal
506 skb_shinfo(skb
)->tso_segs
= 1;
507 skb_shinfo(skb
)->tso_size
= 0;
511 factor
= skb
->len
+ (mss_now
- 1);
513 skb_shinfo(skb
)->tso_segs
= factor
;
514 skb_shinfo(skb
)->tso_size
= mss_now
;
518 /* Function to create two new TCP segments. Shrinks the given segment
519 * to the specified size and appends a new segment with the rest of the
520 * packet to the list. This won't be called frequently, I hope.
521 * Remember, these are still headerless SKBs at this point.
523 int tcp_fragment(struct sock
*sk
, struct sk_buff
*skb
, u32 len
, unsigned int mss_now
)
525 struct tcp_sock
*tp
= tcp_sk(sk
);
526 struct sk_buff
*buff
;
527 int nsize
, old_factor
;
530 BUG_ON(len
> skb
->len
);
532 clear_all_retrans_hints(tp
);
533 nsize
= skb_headlen(skb
) - len
;
537 if (skb_cloned(skb
) &&
538 skb_is_nonlinear(skb
) &&
539 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
542 /* Get a new skb... force flag on. */
543 buff
= sk_stream_alloc_skb(sk
, nsize
, GFP_ATOMIC
);
545 return -ENOMEM
; /* We'll just try again later. */
546 sk_charge_skb(sk
, buff
);
548 /* Correct the sequence numbers. */
549 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
550 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
551 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
553 /* PSH and FIN should only be set in the second packet. */
554 flags
= TCP_SKB_CB(skb
)->flags
;
555 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
556 TCP_SKB_CB(buff
)->flags
= flags
;
557 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
558 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_AT_TAIL
;
560 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_HW
) {
561 /* Copy and checksum data tail into the new buffer. */
562 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
, skb_put(buff
, nsize
),
567 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
569 skb
->ip_summed
= CHECKSUM_HW
;
570 skb_split(skb
, buff
, len
);
573 buff
->ip_summed
= skb
->ip_summed
;
575 /* Looks stupid, but our code really uses when of
576 * skbs, which it never sent before. --ANK
578 TCP_SKB_CB(buff
)->when
= TCP_SKB_CB(skb
)->when
;
579 buff
->tstamp
= skb
->tstamp
;
581 old_factor
= tcp_skb_pcount(skb
);
583 /* Fix up tso_factor for both original and new SKB. */
584 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
585 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
587 /* If this packet has been sent out already, we must
588 * adjust the various packet counters.
590 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
591 int diff
= old_factor
- tcp_skb_pcount(skb
) -
592 tcp_skb_pcount(buff
);
594 tp
->packets_out
-= diff
;
596 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
597 tp
->sacked_out
-= diff
;
598 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
599 tp
->retrans_out
-= diff
;
601 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
) {
602 tp
->lost_out
-= diff
;
603 tp
->left_out
-= diff
;
607 /* Adjust Reno SACK estimate. */
608 if (!tp
->rx_opt
.sack_ok
) {
609 tp
->sacked_out
-= diff
;
610 if ((int)tp
->sacked_out
< 0)
612 tcp_sync_left_out(tp
);
615 tp
->fackets_out
-= diff
;
616 if ((int)tp
->fackets_out
< 0)
621 /* Link BUFF into the send queue. */
622 skb_header_release(buff
);
623 __skb_append(skb
, buff
, &sk
->sk_write_queue
);
628 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
629 * eventually). The difference is that pulled data not copied, but
630 * immediately discarded.
632 static unsigned char *__pskb_trim_head(struct sk_buff
*skb
, int len
)
638 for (i
=0; i
<skb_shinfo(skb
)->nr_frags
; i
++) {
639 if (skb_shinfo(skb
)->frags
[i
].size
<= eat
) {
640 put_page(skb_shinfo(skb
)->frags
[i
].page
);
641 eat
-= skb_shinfo(skb
)->frags
[i
].size
;
643 skb_shinfo(skb
)->frags
[k
] = skb_shinfo(skb
)->frags
[i
];
645 skb_shinfo(skb
)->frags
[k
].page_offset
+= eat
;
646 skb_shinfo(skb
)->frags
[k
].size
-= eat
;
652 skb_shinfo(skb
)->nr_frags
= k
;
654 skb
->tail
= skb
->data
;
655 skb
->data_len
-= len
;
656 skb
->len
= skb
->data_len
;
660 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
662 if (skb_cloned(skb
) &&
663 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
666 if (len
<= skb_headlen(skb
)) {
667 __skb_pull(skb
, len
);
669 if (__pskb_trim_head(skb
, len
-skb_headlen(skb
)) == NULL
)
673 TCP_SKB_CB(skb
)->seq
+= len
;
674 skb
->ip_summed
= CHECKSUM_HW
;
676 skb
->truesize
-= len
;
677 sk
->sk_wmem_queued
-= len
;
678 sk
->sk_forward_alloc
+= len
;
679 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
681 /* Any change of skb->len requires recalculation of tso
684 if (tcp_skb_pcount(skb
) > 1)
685 tcp_set_skb_tso_segs(sk
, skb
, tcp_current_mss(sk
, 1));
690 /* Not accounting for SACKs here. */
691 int tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
693 struct tcp_sock
*tp
= tcp_sk(sk
);
694 struct inet_connection_sock
*icsk
= inet_csk(sk
);
697 /* Calculate base mss without TCP options:
698 It is MMS_S - sizeof(tcphdr) of rfc1122
700 mss_now
= pmtu
- icsk
->icsk_af_ops
->net_header_len
- sizeof(struct tcphdr
);
702 /* Clamp it (mss_clamp does not include tcp options) */
703 if (mss_now
> tp
->rx_opt
.mss_clamp
)
704 mss_now
= tp
->rx_opt
.mss_clamp
;
706 /* Now subtract optional transport overhead */
707 mss_now
-= icsk
->icsk_ext_hdr_len
;
709 /* Then reserve room for full set of TCP options and 8 bytes of data */
713 /* Now subtract TCP options size, not including SACKs */
714 mss_now
-= tp
->tcp_header_len
- sizeof(struct tcphdr
);
719 /* Inverse of above */
720 int tcp_mss_to_mtu(struct sock
*sk
, int mss
)
722 struct tcp_sock
*tp
= tcp_sk(sk
);
723 struct inet_connection_sock
*icsk
= inet_csk(sk
);
728 icsk
->icsk_ext_hdr_len
+
729 icsk
->icsk_af_ops
->net_header_len
;
734 void tcp_mtup_init(struct sock
*sk
)
736 struct tcp_sock
*tp
= tcp_sk(sk
);
737 struct inet_connection_sock
*icsk
= inet_csk(sk
);
739 icsk
->icsk_mtup
.enabled
= sysctl_tcp_mtu_probing
> 1;
740 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+ sizeof(struct tcphdr
) +
741 icsk
->icsk_af_ops
->net_header_len
;
742 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, sysctl_tcp_base_mss
);
743 icsk
->icsk_mtup
.probe_size
= 0;
746 /* This function synchronize snd mss to current pmtu/exthdr set.
748 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
749 for TCP options, but includes only bare TCP header.
751 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
752 It is minimum of user_mss and mss received with SYN.
753 It also does not include TCP options.
755 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
757 tp->mss_cache is current effective sending mss, including
758 all tcp options except for SACKs. It is evaluated,
759 taking into account current pmtu, but never exceeds
760 tp->rx_opt.mss_clamp.
762 NOTE1. rfc1122 clearly states that advertised MSS
763 DOES NOT include either tcp or ip options.
765 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
766 are READ ONLY outside this function. --ANK (980731)
769 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
771 struct tcp_sock
*tp
= tcp_sk(sk
);
772 struct inet_connection_sock
*icsk
= inet_csk(sk
);
775 if (icsk
->icsk_mtup
.search_high
> pmtu
)
776 icsk
->icsk_mtup
.search_high
= pmtu
;
778 mss_now
= tcp_mtu_to_mss(sk
, pmtu
);
780 /* Bound mss with half of window */
781 if (tp
->max_window
&& mss_now
> (tp
->max_window
>>1))
782 mss_now
= max((tp
->max_window
>>1), 68U - tp
->tcp_header_len
);
784 /* And store cached results */
785 icsk
->icsk_pmtu_cookie
= pmtu
;
786 if (icsk
->icsk_mtup
.enabled
)
787 mss_now
= min(mss_now
, tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
));
788 tp
->mss_cache
= mss_now
;
793 /* Compute the current effective MSS, taking SACKs and IP options,
794 * and even PMTU discovery events into account.
796 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
797 * cannot be large. However, taking into account rare use of URG, this
800 unsigned int tcp_current_mss(struct sock
*sk
, int large_allowed
)
802 struct tcp_sock
*tp
= tcp_sk(sk
);
803 struct dst_entry
*dst
= __sk_dst_get(sk
);
808 mss_now
= tp
->mss_cache
;
811 (sk
->sk_route_caps
& NETIF_F_TSO
) &&
816 u32 mtu
= dst_mtu(dst
);
817 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
818 mss_now
= tcp_sync_mss(sk
, mtu
);
821 if (tp
->rx_opt
.eff_sacks
)
822 mss_now
-= (TCPOLEN_SACK_BASE_ALIGNED
+
823 (tp
->rx_opt
.eff_sacks
* TCPOLEN_SACK_PERBLOCK
));
825 xmit_size_goal
= mss_now
;
828 xmit_size_goal
= (65535 -
829 inet_csk(sk
)->icsk_af_ops
->net_header_len
-
830 inet_csk(sk
)->icsk_ext_hdr_len
-
833 if (tp
->max_window
&&
834 (xmit_size_goal
> (tp
->max_window
>> 1)))
835 xmit_size_goal
= max((tp
->max_window
>> 1),
836 68U - tp
->tcp_header_len
);
838 xmit_size_goal
-= (xmit_size_goal
% mss_now
);
840 tp
->xmit_size_goal
= xmit_size_goal
;
845 /* Congestion window validation. (RFC2861) */
847 static void tcp_cwnd_validate(struct sock
*sk
, struct tcp_sock
*tp
)
849 __u32 packets_out
= tp
->packets_out
;
851 if (packets_out
>= tp
->snd_cwnd
) {
852 /* Network is feed fully. */
853 tp
->snd_cwnd_used
= 0;
854 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
856 /* Network starves. */
857 if (tp
->packets_out
> tp
->snd_cwnd_used
)
858 tp
->snd_cwnd_used
= tp
->packets_out
;
860 if ((s32
)(tcp_time_stamp
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
)
861 tcp_cwnd_application_limited(sk
);
865 static unsigned int tcp_window_allows(struct tcp_sock
*tp
, struct sk_buff
*skb
, unsigned int mss_now
, unsigned int cwnd
)
867 u32 window
, cwnd_len
;
869 window
= (tp
->snd_una
+ tp
->snd_wnd
- TCP_SKB_CB(skb
)->seq
);
870 cwnd_len
= mss_now
* cwnd
;
871 return min(window
, cwnd_len
);
874 /* Can at least one segment of SKB be sent right now, according to the
875 * congestion window rules? If so, return how many segments are allowed.
877 static inline unsigned int tcp_cwnd_test(struct tcp_sock
*tp
, struct sk_buff
*skb
)
881 /* Don't be strict about the congestion window for the final FIN. */
882 if (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
)
885 in_flight
= tcp_packets_in_flight(tp
);
887 if (in_flight
< cwnd
)
888 return (cwnd
- in_flight
);
893 /* This must be invoked the first time we consider transmitting
896 static int tcp_init_tso_segs(struct sock
*sk
, struct sk_buff
*skb
, unsigned int mss_now
)
898 int tso_segs
= tcp_skb_pcount(skb
);
902 skb_shinfo(skb
)->tso_size
!= mss_now
)) {
903 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
904 tso_segs
= tcp_skb_pcount(skb
);
909 static inline int tcp_minshall_check(const struct tcp_sock
*tp
)
911 return after(tp
->snd_sml
,tp
->snd_una
) &&
912 !after(tp
->snd_sml
, tp
->snd_nxt
);
915 /* Return 0, if packet can be sent now without violation Nagle's rules:
916 * 1. It is full sized.
917 * 2. Or it contains FIN. (already checked by caller)
918 * 3. Or TCP_NODELAY was set.
919 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
920 * With Minshall's modification: all sent small packets are ACKed.
923 static inline int tcp_nagle_check(const struct tcp_sock
*tp
,
924 const struct sk_buff
*skb
,
925 unsigned mss_now
, int nonagle
)
927 return (skb
->len
< mss_now
&&
928 ((nonagle
&TCP_NAGLE_CORK
) ||
931 tcp_minshall_check(tp
))));
934 /* Return non-zero if the Nagle test allows this packet to be
937 static inline int tcp_nagle_test(struct tcp_sock
*tp
, struct sk_buff
*skb
,
938 unsigned int cur_mss
, int nonagle
)
940 /* Nagle rule does not apply to frames, which sit in the middle of the
941 * write_queue (they have no chances to get new data).
943 * This is implemented in the callers, where they modify the 'nonagle'
944 * argument based upon the location of SKB in the send queue.
946 if (nonagle
& TCP_NAGLE_PUSH
)
949 /* Don't use the nagle rule for urgent data (or for the final FIN). */
951 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
))
954 if (!tcp_nagle_check(tp
, skb
, cur_mss
, nonagle
))
960 /* Does at least the first segment of SKB fit into the send window? */
961 static inline int tcp_snd_wnd_test(struct tcp_sock
*tp
, struct sk_buff
*skb
, unsigned int cur_mss
)
963 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
965 if (skb
->len
> cur_mss
)
966 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
968 return !after(end_seq
, tp
->snd_una
+ tp
->snd_wnd
);
971 /* This checks if the data bearing packet SKB (usually sk->sk_send_head)
972 * should be put on the wire right now. If so, it returns the number of
973 * packets allowed by the congestion window.
975 static unsigned int tcp_snd_test(struct sock
*sk
, struct sk_buff
*skb
,
976 unsigned int cur_mss
, int nonagle
)
978 struct tcp_sock
*tp
= tcp_sk(sk
);
979 unsigned int cwnd_quota
;
981 tcp_init_tso_segs(sk
, skb
, cur_mss
);
983 if (!tcp_nagle_test(tp
, skb
, cur_mss
, nonagle
))
986 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
988 !tcp_snd_wnd_test(tp
, skb
, cur_mss
))
994 static inline int tcp_skb_is_last(const struct sock
*sk
,
995 const struct sk_buff
*skb
)
997 return skb
->next
== (struct sk_buff
*)&sk
->sk_write_queue
;
1000 int tcp_may_send_now(struct sock
*sk
, struct tcp_sock
*tp
)
1002 struct sk_buff
*skb
= sk
->sk_send_head
;
1005 tcp_snd_test(sk
, skb
, tcp_current_mss(sk
, 1),
1006 (tcp_skb_is_last(sk
, skb
) ?
1011 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1012 * which is put after SKB on the list. It is very much like
1013 * tcp_fragment() except that it may make several kinds of assumptions
1014 * in order to speed up the splitting operation. In particular, we
1015 * know that all the data is in scatter-gather pages, and that the
1016 * packet has never been sent out before (and thus is not cloned).
1018 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
, unsigned int mss_now
)
1020 struct sk_buff
*buff
;
1021 int nlen
= skb
->len
- len
;
1024 /* All of a TSO frame must be composed of paged data. */
1025 if (skb
->len
!= skb
->data_len
)
1026 return tcp_fragment(sk
, skb
, len
, mss_now
);
1028 buff
= sk_stream_alloc_pskb(sk
, 0, 0, GFP_ATOMIC
);
1029 if (unlikely(buff
== NULL
))
1032 buff
->truesize
= nlen
;
1033 skb
->truesize
-= nlen
;
1035 /* Correct the sequence numbers. */
1036 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1037 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1038 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1040 /* PSH and FIN should only be set in the second packet. */
1041 flags
= TCP_SKB_CB(skb
)->flags
;
1042 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
1043 TCP_SKB_CB(buff
)->flags
= flags
;
1045 /* This packet was never sent out yet, so no SACK bits. */
1046 TCP_SKB_CB(buff
)->sacked
= 0;
1048 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_HW
;
1049 skb_split(skb
, buff
, len
);
1051 /* Fix up tso_factor for both original and new SKB. */
1052 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1053 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1055 /* Link BUFF into the send queue. */
1056 skb_header_release(buff
);
1057 __skb_append(skb
, buff
, &sk
->sk_write_queue
);
1062 /* Try to defer sending, if possible, in order to minimize the amount
1063 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1065 * This algorithm is from John Heffner.
1067 static int tcp_tso_should_defer(struct sock
*sk
, struct tcp_sock
*tp
, struct sk_buff
*skb
)
1069 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1070 u32 send_win
, cong_win
, limit
, in_flight
;
1072 if (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
)
1075 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
1078 in_flight
= tcp_packets_in_flight(tp
);
1080 BUG_ON(tcp_skb_pcount(skb
) <= 1 ||
1081 (tp
->snd_cwnd
<= in_flight
));
1083 send_win
= (tp
->snd_una
+ tp
->snd_wnd
) - TCP_SKB_CB(skb
)->seq
;
1085 /* From in_flight test above, we know that cwnd > in_flight. */
1086 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
1088 limit
= min(send_win
, cong_win
);
1090 /* If a full-sized TSO skb can be sent, do it. */
1094 if (sysctl_tcp_tso_win_divisor
) {
1095 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
1097 /* If at least some fraction of a window is available,
1100 chunk
/= sysctl_tcp_tso_win_divisor
;
1104 /* Different approach, try not to defer past a single
1105 * ACK. Receiver should ACK every other full sized
1106 * frame, so if we have space for more than 3 frames
1109 if (limit
> tcp_max_burst(tp
) * tp
->mss_cache
)
1113 /* Ok, it looks like it is advisable to defer. */
1117 /* Create a new MTU probe if we are ready.
1118 * Returns 0 if we should wait to probe (no cwnd available),
1119 * 1 if a probe was sent,
1121 static int tcp_mtu_probe(struct sock
*sk
)
1123 struct tcp_sock
*tp
= tcp_sk(sk
);
1124 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1125 struct sk_buff
*skb
, *nskb
, *next
;
1132 /* Not currently probing/verifying,
1134 * have enough cwnd, and
1135 * not SACKing (the variable headers throw things off) */
1136 if (!icsk
->icsk_mtup
.enabled
||
1137 icsk
->icsk_mtup
.probe_size
||
1138 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
||
1139 tp
->snd_cwnd
< 11 ||
1140 tp
->rx_opt
.eff_sacks
)
1143 /* Very simple search strategy: just double the MSS. */
1144 mss_now
= tcp_current_mss(sk
, 0);
1145 probe_size
= 2*tp
->mss_cache
;
1146 if (probe_size
> tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_high
)) {
1147 /* TODO: set timer for probe_converge_event */
1151 /* Have enough data in the send queue to probe? */
1153 if ((skb
= sk
->sk_send_head
) == NULL
)
1155 while ((len
+= skb
->len
) < probe_size
&& !tcp_skb_is_last(sk
, skb
))
1157 if (len
< probe_size
)
1160 /* Receive window check. */
1161 if (after(TCP_SKB_CB(skb
)->seq
+ probe_size
, tp
->snd_una
+ tp
->snd_wnd
)) {
1162 if (tp
->snd_wnd
< probe_size
)
1168 /* Do we need to wait to drain cwnd? */
1169 pif
= tcp_packets_in_flight(tp
);
1170 if (pif
+ 2 > tp
->snd_cwnd
) {
1171 /* With no packets in flight, don't stall. */
1178 /* We're allowed to probe. Build it now. */
1179 if ((nskb
= sk_stream_alloc_skb(sk
, probe_size
, GFP_ATOMIC
)) == NULL
)
1181 sk_charge_skb(sk
, nskb
);
1183 skb
= sk
->sk_send_head
;
1184 __skb_insert(nskb
, skb
->prev
, skb
, &sk
->sk_write_queue
);
1185 sk
->sk_send_head
= nskb
;
1187 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(skb
)->seq
;
1188 TCP_SKB_CB(nskb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ probe_size
;
1189 TCP_SKB_CB(nskb
)->flags
= TCPCB_FLAG_ACK
;
1190 TCP_SKB_CB(nskb
)->sacked
= 0;
1192 if (skb
->ip_summed
== CHECKSUM_HW
)
1193 nskb
->ip_summed
= CHECKSUM_HW
;
1196 while (len
< probe_size
) {
1199 copy
= min_t(int, skb
->len
, probe_size
- len
);
1200 if (nskb
->ip_summed
)
1201 skb_copy_bits(skb
, 0, skb_put(nskb
, copy
), copy
);
1203 nskb
->csum
= skb_copy_and_csum_bits(skb
, 0,
1204 skb_put(nskb
, copy
), copy
, nskb
->csum
);
1206 if (skb
->len
<= copy
) {
1207 /* We've eaten all the data from this skb.
1209 TCP_SKB_CB(nskb
)->flags
|= TCP_SKB_CB(skb
)->flags
;
1210 __skb_unlink(skb
, &sk
->sk_write_queue
);
1211 sk_stream_free_skb(sk
, skb
);
1213 TCP_SKB_CB(nskb
)->flags
|= TCP_SKB_CB(skb
)->flags
&
1214 ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
1215 if (!skb_shinfo(skb
)->nr_frags
) {
1216 skb_pull(skb
, copy
);
1217 if (skb
->ip_summed
!= CHECKSUM_HW
)
1218 skb
->csum
= csum_partial(skb
->data
, skb
->len
, 0);
1220 __pskb_trim_head(skb
, copy
);
1221 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1223 TCP_SKB_CB(skb
)->seq
+= copy
;
1229 tcp_init_tso_segs(sk
, nskb
, nskb
->len
);
1231 /* We're ready to send. If this fails, the probe will
1232 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1233 TCP_SKB_CB(nskb
)->when
= tcp_time_stamp
;
1234 if (!tcp_transmit_skb(sk
, nskb
, 1, GFP_ATOMIC
)) {
1235 /* Decrement cwnd here because we are sending
1236 * effectively two packets. */
1238 update_send_head(sk
, tp
, nskb
);
1240 icsk
->icsk_mtup
.probe_size
= tcp_mss_to_mtu(sk
, nskb
->len
);
1241 icsk
->icsk_mtup
.probe_seq_start
= TCP_SKB_CB(nskb
)->seq
;
1242 icsk
->icsk_mtup
.probe_seq_end
= TCP_SKB_CB(nskb
)->end_seq
;
1251 /* This routine writes packets to the network. It advances the
1252 * send_head. This happens as incoming acks open up the remote
1255 * Returns 1, if no segments are in flight and we have queued segments, but
1256 * cannot send anything now because of SWS or another problem.
1258 static int tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
)
1260 struct tcp_sock
*tp
= tcp_sk(sk
);
1261 struct sk_buff
*skb
;
1262 unsigned int tso_segs
, sent_pkts
;
1266 /* If we are closed, the bytes will have to remain here.
1267 * In time closedown will finish, we empty the write queue and all
1270 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
1275 /* Do MTU probing. */
1276 if ((result
= tcp_mtu_probe(sk
)) == 0) {
1278 } else if (result
> 0) {
1282 while ((skb
= sk
->sk_send_head
)) {
1285 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1288 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1292 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
)))
1295 if (tso_segs
== 1) {
1296 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
1297 (tcp_skb_is_last(sk
, skb
) ?
1298 nonagle
: TCP_NAGLE_PUSH
))))
1301 if (tcp_tso_should_defer(sk
, tp
, skb
))
1307 limit
= tcp_window_allows(tp
, skb
,
1308 mss_now
, cwnd_quota
);
1310 if (skb
->len
< limit
) {
1311 unsigned int trim
= skb
->len
% mss_now
;
1314 limit
= skb
->len
- trim
;
1318 if (skb
->len
> limit
&&
1319 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
)))
1322 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1324 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
)))
1327 /* Advance the send_head. This one is sent out.
1328 * This call will increment packets_out.
1330 update_send_head(sk
, tp
, skb
);
1332 tcp_minshall_update(tp
, mss_now
, skb
);
1336 if (likely(sent_pkts
)) {
1337 tcp_cwnd_validate(sk
, tp
);
1340 return !tp
->packets_out
&& sk
->sk_send_head
;
1343 /* Push out any pending frames which were held back due to
1344 * TCP_CORK or attempt at coalescing tiny packets.
1345 * The socket must be locked by the caller.
1347 void __tcp_push_pending_frames(struct sock
*sk
, struct tcp_sock
*tp
,
1348 unsigned int cur_mss
, int nonagle
)
1350 struct sk_buff
*skb
= sk
->sk_send_head
;
1353 if (tcp_write_xmit(sk
, cur_mss
, nonagle
))
1354 tcp_check_probe_timer(sk
, tp
);
1358 /* Send _single_ skb sitting at the send head. This function requires
1359 * true push pending frames to setup probe timer etc.
1361 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
1363 struct tcp_sock
*tp
= tcp_sk(sk
);
1364 struct sk_buff
*skb
= sk
->sk_send_head
;
1365 unsigned int tso_segs
, cwnd_quota
;
1367 BUG_ON(!skb
|| skb
->len
< mss_now
);
1369 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1370 cwnd_quota
= tcp_snd_test(sk
, skb
, mss_now
, TCP_NAGLE_PUSH
);
1372 if (likely(cwnd_quota
)) {
1379 limit
= tcp_window_allows(tp
, skb
,
1380 mss_now
, cwnd_quota
);
1382 if (skb
->len
< limit
) {
1383 unsigned int trim
= skb
->len
% mss_now
;
1386 limit
= skb
->len
- trim
;
1390 if (skb
->len
> limit
&&
1391 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
)))
1394 /* Send it out now. */
1395 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1397 if (likely(!tcp_transmit_skb(sk
, skb
, 1, sk
->sk_allocation
))) {
1398 update_send_head(sk
, tp
, skb
);
1399 tcp_cwnd_validate(sk
, tp
);
1405 /* This function returns the amount that we can raise the
1406 * usable window based on the following constraints
1408 * 1. The window can never be shrunk once it is offered (RFC 793)
1409 * 2. We limit memory per socket
1412 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1413 * RECV.NEXT + RCV.WIN fixed until:
1414 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1416 * i.e. don't raise the right edge of the window until you can raise
1417 * it at least MSS bytes.
1419 * Unfortunately, the recommended algorithm breaks header prediction,
1420 * since header prediction assumes th->window stays fixed.
1422 * Strictly speaking, keeping th->window fixed violates the receiver
1423 * side SWS prevention criteria. The problem is that under this rule
1424 * a stream of single byte packets will cause the right side of the
1425 * window to always advance by a single byte.
1427 * Of course, if the sender implements sender side SWS prevention
1428 * then this will not be a problem.
1430 * BSD seems to make the following compromise:
1432 * If the free space is less than the 1/4 of the maximum
1433 * space available and the free space is less than 1/2 mss,
1434 * then set the window to 0.
1435 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1436 * Otherwise, just prevent the window from shrinking
1437 * and from being larger than the largest representable value.
1439 * This prevents incremental opening of the window in the regime
1440 * where TCP is limited by the speed of the reader side taking
1441 * data out of the TCP receive queue. It does nothing about
1442 * those cases where the window is constrained on the sender side
1443 * because the pipeline is full.
1445 * BSD also seems to "accidentally" limit itself to windows that are a
1446 * multiple of MSS, at least until the free space gets quite small.
1447 * This would appear to be a side effect of the mbuf implementation.
1448 * Combining these two algorithms results in the observed behavior
1449 * of having a fixed window size at almost all times.
1451 * Below we obtain similar behavior by forcing the offered window to
1452 * a multiple of the mss when it is feasible to do so.
1454 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1455 * Regular options like TIMESTAMP are taken into account.
1457 u32
__tcp_select_window(struct sock
*sk
)
1459 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1460 struct tcp_sock
*tp
= tcp_sk(sk
);
1461 /* MSS for the peer's data. Previous versions used mss_clamp
1462 * here. I don't know if the value based on our guesses
1463 * of peer's MSS is better for the performance. It's more correct
1464 * but may be worse for the performance because of rcv_mss
1465 * fluctuations. --SAW 1998/11/1
1467 int mss
= icsk
->icsk_ack
.rcv_mss
;
1468 int free_space
= tcp_space(sk
);
1469 int full_space
= min_t(int, tp
->window_clamp
, tcp_full_space(sk
));
1472 if (mss
> full_space
)
1475 if (free_space
< full_space
/2) {
1476 icsk
->icsk_ack
.quick
= 0;
1478 if (tcp_memory_pressure
)
1479 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U*tp
->advmss
);
1481 if (free_space
< mss
)
1485 if (free_space
> tp
->rcv_ssthresh
)
1486 free_space
= tp
->rcv_ssthresh
;
1488 /* Don't do rounding if we are using window scaling, since the
1489 * scaled window will not line up with the MSS boundary anyway.
1491 window
= tp
->rcv_wnd
;
1492 if (tp
->rx_opt
.rcv_wscale
) {
1493 window
= free_space
;
1495 /* Advertise enough space so that it won't get scaled away.
1496 * Import case: prevent zero window announcement if
1497 * 1<<rcv_wscale > mss.
1499 if (((window
>> tp
->rx_opt
.rcv_wscale
) << tp
->rx_opt
.rcv_wscale
) != window
)
1500 window
= (((window
>> tp
->rx_opt
.rcv_wscale
) + 1)
1501 << tp
->rx_opt
.rcv_wscale
);
1503 /* Get the largest window that is a nice multiple of mss.
1504 * Window clamp already applied above.
1505 * If our current window offering is within 1 mss of the
1506 * free space we just keep it. This prevents the divide
1507 * and multiply from happening most of the time.
1508 * We also don't do any window rounding when the free space
1511 if (window
<= free_space
- mss
|| window
> free_space
)
1512 window
= (free_space
/mss
)*mss
;
1518 /* Attempt to collapse two adjacent SKB's during retransmission. */
1519 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*skb
, int mss_now
)
1521 struct tcp_sock
*tp
= tcp_sk(sk
);
1522 struct sk_buff
*next_skb
= skb
->next
;
1524 /* The first test we must make is that neither of these two
1525 * SKB's are still referenced by someone else.
1527 if (!skb_cloned(skb
) && !skb_cloned(next_skb
)) {
1528 int skb_size
= skb
->len
, next_skb_size
= next_skb
->len
;
1529 u16 flags
= TCP_SKB_CB(skb
)->flags
;
1531 /* Also punt if next skb has been SACK'd. */
1532 if(TCP_SKB_CB(next_skb
)->sacked
& TCPCB_SACKED_ACKED
)
1535 /* Next skb is out of window. */
1536 if (after(TCP_SKB_CB(next_skb
)->end_seq
, tp
->snd_una
+tp
->snd_wnd
))
1539 /* Punt if not enough space exists in the first SKB for
1540 * the data in the second, or the total combined payload
1541 * would exceed the MSS.
1543 if ((next_skb_size
> skb_tailroom(skb
)) ||
1544 ((skb_size
+ next_skb_size
) > mss_now
))
1547 BUG_ON(tcp_skb_pcount(skb
) != 1 ||
1548 tcp_skb_pcount(next_skb
) != 1);
1550 /* changing transmit queue under us so clear hints */
1551 clear_all_retrans_hints(tp
);
1553 /* Ok. We will be able to collapse the packet. */
1554 __skb_unlink(next_skb
, &sk
->sk_write_queue
);
1556 memcpy(skb_put(skb
, next_skb_size
), next_skb
->data
, next_skb_size
);
1558 if (next_skb
->ip_summed
== CHECKSUM_HW
)
1559 skb
->ip_summed
= CHECKSUM_HW
;
1561 if (skb
->ip_summed
!= CHECKSUM_HW
)
1562 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
1564 /* Update sequence range on original skb. */
1565 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
1567 /* Merge over control information. */
1568 flags
|= TCP_SKB_CB(next_skb
)->flags
; /* This moves PSH/FIN etc. over */
1569 TCP_SKB_CB(skb
)->flags
= flags
;
1571 /* All done, get rid of second SKB and account for it so
1572 * packet counting does not break.
1574 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
&(TCPCB_EVER_RETRANS
|TCPCB_AT_TAIL
);
1575 if (TCP_SKB_CB(next_skb
)->sacked
&TCPCB_SACKED_RETRANS
)
1576 tp
->retrans_out
-= tcp_skb_pcount(next_skb
);
1577 if (TCP_SKB_CB(next_skb
)->sacked
&TCPCB_LOST
) {
1578 tp
->lost_out
-= tcp_skb_pcount(next_skb
);
1579 tp
->left_out
-= tcp_skb_pcount(next_skb
);
1581 /* Reno case is special. Sigh... */
1582 if (!tp
->rx_opt
.sack_ok
&& tp
->sacked_out
) {
1583 tcp_dec_pcount_approx(&tp
->sacked_out
, next_skb
);
1584 tp
->left_out
-= tcp_skb_pcount(next_skb
);
1587 /* Not quite right: it can be > snd.fack, but
1588 * it is better to underestimate fackets.
1590 tcp_dec_pcount_approx(&tp
->fackets_out
, next_skb
);
1591 tcp_packets_out_dec(tp
, next_skb
);
1592 sk_stream_free_skb(sk
, next_skb
);
1596 /* Do a simple retransmit without using the backoff mechanisms in
1597 * tcp_timer. This is used for path mtu discovery.
1598 * The socket is already locked here.
1600 void tcp_simple_retransmit(struct sock
*sk
)
1602 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1603 struct tcp_sock
*tp
= tcp_sk(sk
);
1604 struct sk_buff
*skb
;
1605 unsigned int mss
= tcp_current_mss(sk
, 0);
1608 sk_stream_for_retrans_queue(skb
, sk
) {
1609 if (skb
->len
> mss
&&
1610 !(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
)) {
1611 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) {
1612 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1613 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1615 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_LOST
)) {
1616 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1617 tp
->lost_out
+= tcp_skb_pcount(skb
);
1623 clear_all_retrans_hints(tp
);
1628 tcp_sync_left_out(tp
);
1630 /* Don't muck with the congestion window here.
1631 * Reason is that we do not increase amount of _data_
1632 * in network, but units changed and effective
1633 * cwnd/ssthresh really reduced now.
1635 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
1636 tp
->high_seq
= tp
->snd_nxt
;
1637 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
1638 tp
->prior_ssthresh
= 0;
1639 tp
->undo_marker
= 0;
1640 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1642 tcp_xmit_retransmit_queue(sk
);
1645 /* This retransmits one SKB. Policy decisions and retransmit queue
1646 * state updates are done by the caller. Returns non-zero if an
1647 * error occurred which prevented the send.
1649 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
1651 struct tcp_sock
*tp
= tcp_sk(sk
);
1652 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1653 unsigned int cur_mss
= tcp_current_mss(sk
, 0);
1656 /* Inconslusive MTU probe */
1657 if (icsk
->icsk_mtup
.probe_size
) {
1658 icsk
->icsk_mtup
.probe_size
= 0;
1661 /* Do not sent more than we queued. 1/4 is reserved for possible
1662 * copying overhead: fragmentation, tunneling, mangling etc.
1664 if (atomic_read(&sk
->sk_wmem_alloc
) >
1665 min(sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2), sk
->sk_sndbuf
))
1668 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
1669 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1671 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
1675 /* If receiver has shrunk his window, and skb is out of
1676 * new window, do not retransmit it. The exception is the
1677 * case, when window is shrunk to zero. In this case
1678 * our retransmit serves as a zero window probe.
1680 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
+tp
->snd_wnd
)
1681 && TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
1684 if (skb
->len
> cur_mss
) {
1685 if (tcp_fragment(sk
, skb
, cur_mss
, cur_mss
))
1686 return -ENOMEM
; /* We'll try again later. */
1689 /* Collapse two adjacent packets if worthwhile and we can. */
1690 if(!(TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_SYN
) &&
1691 (skb
->len
< (cur_mss
>> 1)) &&
1692 (skb
->next
!= sk
->sk_send_head
) &&
1693 (skb
->next
!= (struct sk_buff
*)&sk
->sk_write_queue
) &&
1694 (skb_shinfo(skb
)->nr_frags
== 0 && skb_shinfo(skb
->next
)->nr_frags
== 0) &&
1695 (tcp_skb_pcount(skb
) == 1 && tcp_skb_pcount(skb
->next
) == 1) &&
1696 (sysctl_tcp_retrans_collapse
!= 0))
1697 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
1699 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
1700 return -EHOSTUNREACH
; /* Routing failure or similar. */
1702 /* Some Solaris stacks overoptimize and ignore the FIN on a
1703 * retransmit when old data is attached. So strip it off
1704 * since it is cheap to do so and saves bytes on the network.
1707 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
) &&
1708 tp
->snd_una
== (TCP_SKB_CB(skb
)->end_seq
- 1)) {
1709 if (!pskb_trim(skb
, 0)) {
1710 TCP_SKB_CB(skb
)->seq
= TCP_SKB_CB(skb
)->end_seq
- 1;
1711 skb_shinfo(skb
)->tso_segs
= 1;
1712 skb_shinfo(skb
)->tso_size
= 0;
1713 skb
->ip_summed
= CHECKSUM_NONE
;
1718 /* Make a copy, if the first transmission SKB clone we made
1719 * is still in somebody's hands, else make a clone.
1721 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1723 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
1726 /* Update global TCP statistics. */
1727 TCP_INC_STATS(TCP_MIB_RETRANSSEGS
);
1729 tp
->total_retrans
++;
1731 #if FASTRETRANS_DEBUG > 0
1732 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) {
1733 if (net_ratelimit())
1734 printk(KERN_DEBUG
"retrans_out leaked.\n");
1737 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
1738 tp
->retrans_out
+= tcp_skb_pcount(skb
);
1740 /* Save stamp of the first retransmit. */
1741 if (!tp
->retrans_stamp
)
1742 tp
->retrans_stamp
= TCP_SKB_CB(skb
)->when
;
1746 /* snd_nxt is stored to detect loss of retransmitted segment,
1747 * see tcp_input.c tcp_sacktag_write_queue().
1749 TCP_SKB_CB(skb
)->ack_seq
= tp
->snd_nxt
;
1754 /* This gets called after a retransmit timeout, and the initially
1755 * retransmitted data is acknowledged. It tries to continue
1756 * resending the rest of the retransmit queue, until either
1757 * we've sent it all or the congestion window limit is reached.
1758 * If doing SACK, the first ACK which comes back for a timeout
1759 * based retransmit packet might feed us FACK information again.
1760 * If so, we use it to avoid unnecessarily retransmissions.
1762 void tcp_xmit_retransmit_queue(struct sock
*sk
)
1764 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1765 struct tcp_sock
*tp
= tcp_sk(sk
);
1766 struct sk_buff
*skb
;
1769 if (tp
->retransmit_skb_hint
) {
1770 skb
= tp
->retransmit_skb_hint
;
1771 packet_cnt
= tp
->retransmit_cnt_hint
;
1773 skb
= sk
->sk_write_queue
.next
;
1777 /* First pass: retransmit lost packets. */
1779 sk_stream_for_retrans_queue_from(skb
, sk
) {
1780 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
1782 /* we could do better than to assign each time */
1783 tp
->retransmit_skb_hint
= skb
;
1784 tp
->retransmit_cnt_hint
= packet_cnt
;
1786 /* Assume this retransmit will generate
1787 * only one packet for congestion window
1788 * calculation purposes. This works because
1789 * tcp_retransmit_skb() will chop up the
1790 * packet to be MSS sized and all the
1791 * packet counting works out.
1793 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
1796 if (sacked
& TCPCB_LOST
) {
1797 if (!(sacked
&(TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))) {
1798 if (tcp_retransmit_skb(sk
, skb
)) {
1799 tp
->retransmit_skb_hint
= NULL
;
1802 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
1803 NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS
);
1805 NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS
);
1808 skb_peek(&sk
->sk_write_queue
))
1809 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1810 inet_csk(sk
)->icsk_rto
,
1814 packet_cnt
+= tcp_skb_pcount(skb
);
1815 if (packet_cnt
>= tp
->lost_out
)
1821 /* OK, demanded retransmission is finished. */
1823 /* Forward retransmissions are possible only during Recovery. */
1824 if (icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
1827 /* No forward retransmissions in Reno are possible. */
1828 if (!tp
->rx_opt
.sack_ok
)
1831 /* Yeah, we have to make difficult choice between forward transmission
1832 * and retransmission... Both ways have their merits...
1834 * For now we do not retransmit anything, while we have some new
1838 if (tcp_may_send_now(sk
, tp
))
1841 if (tp
->forward_skb_hint
) {
1842 skb
= tp
->forward_skb_hint
;
1843 packet_cnt
= tp
->forward_cnt_hint
;
1845 skb
= sk
->sk_write_queue
.next
;
1849 sk_stream_for_retrans_queue_from(skb
, sk
) {
1850 tp
->forward_cnt_hint
= packet_cnt
;
1851 tp
->forward_skb_hint
= skb
;
1853 /* Similar to the retransmit loop above we
1854 * can pretend that the retransmitted SKB
1855 * we send out here will be composed of one
1856 * real MSS sized packet because tcp_retransmit_skb()
1857 * will fragment it if necessary.
1859 if (++packet_cnt
> tp
->fackets_out
)
1862 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
1865 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
)
1868 /* Ok, retransmit it. */
1869 if (tcp_retransmit_skb(sk
, skb
)) {
1870 tp
->forward_skb_hint
= NULL
;
1874 if (skb
== skb_peek(&sk
->sk_write_queue
))
1875 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1876 inet_csk(sk
)->icsk_rto
,
1879 NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS
);
1884 /* Send a fin. The caller locks the socket for us. This cannot be
1885 * allowed to fail queueing a FIN frame under any circumstances.
1887 void tcp_send_fin(struct sock
*sk
)
1889 struct tcp_sock
*tp
= tcp_sk(sk
);
1890 struct sk_buff
*skb
= skb_peek_tail(&sk
->sk_write_queue
);
1893 /* Optimization, tack on the FIN if we have a queue of
1894 * unsent frames. But be careful about outgoing SACKS
1897 mss_now
= tcp_current_mss(sk
, 1);
1899 if (sk
->sk_send_head
!= NULL
) {
1900 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_FIN
;
1901 TCP_SKB_CB(skb
)->end_seq
++;
1904 /* Socket is locked, keep trying until memory is available. */
1906 skb
= alloc_skb_fclone(MAX_TCP_HEADER
, GFP_KERNEL
);
1912 /* Reserve space for headers and prepare control bits. */
1913 skb_reserve(skb
, MAX_TCP_HEADER
);
1915 TCP_SKB_CB(skb
)->flags
= (TCPCB_FLAG_ACK
| TCPCB_FLAG_FIN
);
1916 TCP_SKB_CB(skb
)->sacked
= 0;
1917 skb_shinfo(skb
)->tso_segs
= 1;
1918 skb_shinfo(skb
)->tso_size
= 0;
1920 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
1921 TCP_SKB_CB(skb
)->seq
= tp
->write_seq
;
1922 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ 1;
1923 tcp_queue_skb(sk
, skb
);
1925 __tcp_push_pending_frames(sk
, tp
, mss_now
, TCP_NAGLE_OFF
);
1928 /* We get here when a process closes a file descriptor (either due to
1929 * an explicit close() or as a byproduct of exit()'ing) and there
1930 * was unread data in the receive queue. This behavior is recommended
1931 * by draft-ietf-tcpimpl-prob-03.txt section 3.10. -DaveM
1933 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
1935 struct tcp_sock
*tp
= tcp_sk(sk
);
1936 struct sk_buff
*skb
;
1938 /* NOTE: No TCP options attached and we never retransmit this. */
1939 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
1941 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED
);
1945 /* Reserve space for headers and prepare control bits. */
1946 skb_reserve(skb
, MAX_TCP_HEADER
);
1948 TCP_SKB_CB(skb
)->flags
= (TCPCB_FLAG_ACK
| TCPCB_FLAG_RST
);
1949 TCP_SKB_CB(skb
)->sacked
= 0;
1950 skb_shinfo(skb
)->tso_segs
= 1;
1951 skb_shinfo(skb
)->tso_size
= 0;
1954 TCP_SKB_CB(skb
)->seq
= tcp_acceptable_seq(sk
, tp
);
1955 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
;
1956 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1957 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
1958 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED
);
1961 /* WARNING: This routine must only be called when we have already sent
1962 * a SYN packet that crossed the incoming SYN that caused this routine
1963 * to get called. If this assumption fails then the initial rcv_wnd
1964 * and rcv_wscale values will not be correct.
1966 int tcp_send_synack(struct sock
*sk
)
1968 struct sk_buff
* skb
;
1970 skb
= skb_peek(&sk
->sk_write_queue
);
1971 if (skb
== NULL
|| !(TCP_SKB_CB(skb
)->flags
&TCPCB_FLAG_SYN
)) {
1972 printk(KERN_DEBUG
"tcp_send_synack: wrong queue state\n");
1975 if (!(TCP_SKB_CB(skb
)->flags
&TCPCB_FLAG_ACK
)) {
1976 if (skb_cloned(skb
)) {
1977 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
1980 __skb_unlink(skb
, &sk
->sk_write_queue
);
1981 skb_header_release(nskb
);
1982 __skb_queue_head(&sk
->sk_write_queue
, nskb
);
1983 sk_stream_free_skb(sk
, skb
);
1984 sk_charge_skb(sk
, nskb
);
1988 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_ACK
;
1989 TCP_ECN_send_synack(tcp_sk(sk
), skb
);
1991 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1992 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
1996 * Prepare a SYN-ACK.
1998 struct sk_buff
* tcp_make_synack(struct sock
*sk
, struct dst_entry
*dst
,
1999 struct request_sock
*req
)
2001 struct inet_request_sock
*ireq
= inet_rsk(req
);
2002 struct tcp_sock
*tp
= tcp_sk(sk
);
2004 int tcp_header_size
;
2005 struct sk_buff
*skb
;
2007 skb
= sock_wmalloc(sk
, MAX_TCP_HEADER
+ 15, 1, GFP_ATOMIC
);
2011 /* Reserve space for headers. */
2012 skb_reserve(skb
, MAX_TCP_HEADER
);
2014 skb
->dst
= dst_clone(dst
);
2016 tcp_header_size
= (sizeof(struct tcphdr
) + TCPOLEN_MSS
+
2017 (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0) +
2018 (ireq
->wscale_ok
? TCPOLEN_WSCALE_ALIGNED
: 0) +
2019 /* SACK_PERM is in the place of NOP NOP of TS */
2020 ((ireq
->sack_ok
&& !ireq
->tstamp_ok
) ? TCPOLEN_SACKPERM_ALIGNED
: 0));
2021 skb
->h
.th
= th
= (struct tcphdr
*) skb_push(skb
, tcp_header_size
);
2023 memset(th
, 0, sizeof(struct tcphdr
));
2026 if (dst
->dev
->features
&NETIF_F_TSO
)
2028 TCP_ECN_make_synack(req
, th
);
2029 th
->source
= inet_sk(sk
)->sport
;
2030 th
->dest
= ireq
->rmt_port
;
2031 TCP_SKB_CB(skb
)->seq
= tcp_rsk(req
)->snt_isn
;
2032 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ 1;
2033 TCP_SKB_CB(skb
)->sacked
= 0;
2034 skb_shinfo(skb
)->tso_segs
= 1;
2035 skb_shinfo(skb
)->tso_size
= 0;
2036 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
2037 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_isn
+ 1);
2038 if (req
->rcv_wnd
== 0) { /* ignored for retransmitted syns */
2040 /* Set this up on the first call only */
2041 req
->window_clamp
= tp
->window_clamp
? : dst_metric(dst
, RTAX_WINDOW
);
2042 /* tcp_full_space because it is guaranteed to be the first packet */
2043 tcp_select_initial_window(tcp_full_space(sk
),
2044 dst_metric(dst
, RTAX_ADVMSS
) - (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0),
2049 ireq
->rcv_wscale
= rcv_wscale
;
2052 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2053 th
->window
= htons(req
->rcv_wnd
);
2055 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2056 tcp_syn_build_options((__u32
*)(th
+ 1), dst_metric(dst
, RTAX_ADVMSS
), ireq
->tstamp_ok
,
2057 ireq
->sack_ok
, ireq
->wscale_ok
, ireq
->rcv_wscale
,
2058 TCP_SKB_CB(skb
)->when
,
2062 th
->doff
= (tcp_header_size
>> 2);
2063 TCP_INC_STATS(TCP_MIB_OUTSEGS
);
2068 * Do all connect socket setups that can be done AF independent.
2070 static void tcp_connect_init(struct sock
*sk
)
2072 struct dst_entry
*dst
= __sk_dst_get(sk
);
2073 struct tcp_sock
*tp
= tcp_sk(sk
);
2076 /* We'll fix this up when we get a response from the other end.
2077 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2079 tp
->tcp_header_len
= sizeof(struct tcphdr
) +
2080 (sysctl_tcp_timestamps
? TCPOLEN_TSTAMP_ALIGNED
: 0);
2082 /* If user gave his TCP_MAXSEG, record it to clamp */
2083 if (tp
->rx_opt
.user_mss
)
2084 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
2087 tcp_sync_mss(sk
, dst_mtu(dst
));
2089 if (!tp
->window_clamp
)
2090 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
2091 tp
->advmss
= dst_metric(dst
, RTAX_ADVMSS
);
2092 tcp_initialize_rcv_mss(sk
);
2094 tcp_select_initial_window(tcp_full_space(sk
),
2095 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
2098 sysctl_tcp_window_scaling
,
2101 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
2102 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
2105 sock_reset_flag(sk
, SOCK_DONE
);
2107 tcp_init_wl(tp
, tp
->write_seq
, 0);
2108 tp
->snd_una
= tp
->write_seq
;
2109 tp
->snd_sml
= tp
->write_seq
;
2114 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
2115 inet_csk(sk
)->icsk_retransmits
= 0;
2116 tcp_clear_retrans(tp
);
2120 * Build a SYN and send it off.
2122 int tcp_connect(struct sock
*sk
)
2124 struct tcp_sock
*tp
= tcp_sk(sk
);
2125 struct sk_buff
*buff
;
2127 tcp_connect_init(sk
);
2129 buff
= alloc_skb_fclone(MAX_TCP_HEADER
+ 15, sk
->sk_allocation
);
2130 if (unlikely(buff
== NULL
))
2133 /* Reserve space for headers. */
2134 skb_reserve(buff
, MAX_TCP_HEADER
);
2136 TCP_SKB_CB(buff
)->flags
= TCPCB_FLAG_SYN
;
2137 TCP_ECN_send_syn(sk
, tp
, buff
);
2138 TCP_SKB_CB(buff
)->sacked
= 0;
2139 skb_shinfo(buff
)->tso_segs
= 1;
2140 skb_shinfo(buff
)->tso_size
= 0;
2142 TCP_SKB_CB(buff
)->seq
= tp
->write_seq
++;
2143 TCP_SKB_CB(buff
)->end_seq
= tp
->write_seq
;
2144 tp
->snd_nxt
= tp
->write_seq
;
2145 tp
->pushed_seq
= tp
->write_seq
;
2148 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2149 tp
->retrans_stamp
= TCP_SKB_CB(buff
)->when
;
2150 skb_header_release(buff
);
2151 __skb_queue_tail(&sk
->sk_write_queue
, buff
);
2152 sk_charge_skb(sk
, buff
);
2153 tp
->packets_out
+= tcp_skb_pcount(buff
);
2154 tcp_transmit_skb(sk
, buff
, 1, GFP_KERNEL
);
2155 TCP_INC_STATS(TCP_MIB_ACTIVEOPENS
);
2157 /* Timer for repeating the SYN until an answer. */
2158 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2159 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2163 /* Send out a delayed ack, the caller does the policy checking
2164 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2167 void tcp_send_delayed_ack(struct sock
*sk
)
2169 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2170 int ato
= icsk
->icsk_ack
.ato
;
2171 unsigned long timeout
;
2173 if (ato
> TCP_DELACK_MIN
) {
2174 const struct tcp_sock
*tp
= tcp_sk(sk
);
2177 if (icsk
->icsk_ack
.pingpong
|| (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
2178 max_ato
= TCP_DELACK_MAX
;
2180 /* Slow path, intersegment interval is "high". */
2182 /* If some rtt estimate is known, use it to bound delayed ack.
2183 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2187 int rtt
= max(tp
->srtt
>>3, TCP_DELACK_MIN
);
2193 ato
= min(ato
, max_ato
);
2196 /* Stay within the limit we were given */
2197 timeout
= jiffies
+ ato
;
2199 /* Use new timeout only if there wasn't a older one earlier. */
2200 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
2201 /* If delack timer was blocked or is about to expire,
2204 if (icsk
->icsk_ack
.blocked
||
2205 time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
2210 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
2211 timeout
= icsk
->icsk_ack
.timeout
;
2213 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
2214 icsk
->icsk_ack
.timeout
= timeout
;
2215 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
2218 /* This routine sends an ack and also updates the window. */
2219 void tcp_send_ack(struct sock
*sk
)
2221 /* If we have been reset, we may not send again. */
2222 if (sk
->sk_state
!= TCP_CLOSE
) {
2223 struct tcp_sock
*tp
= tcp_sk(sk
);
2224 struct sk_buff
*buff
;
2226 /* We are not putting this on the write queue, so
2227 * tcp_transmit_skb() will set the ownership to this
2230 buff
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2232 inet_csk_schedule_ack(sk
);
2233 inet_csk(sk
)->icsk_ack
.ato
= TCP_ATO_MIN
;
2234 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
2235 TCP_DELACK_MAX
, TCP_RTO_MAX
);
2239 /* Reserve space for headers and prepare control bits. */
2240 skb_reserve(buff
, MAX_TCP_HEADER
);
2242 TCP_SKB_CB(buff
)->flags
= TCPCB_FLAG_ACK
;
2243 TCP_SKB_CB(buff
)->sacked
= 0;
2244 skb_shinfo(buff
)->tso_segs
= 1;
2245 skb_shinfo(buff
)->tso_size
= 0;
2247 /* Send it off, this clears delayed acks for us. */
2248 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(buff
)->end_seq
= tcp_acceptable_seq(sk
, tp
);
2249 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2250 tcp_transmit_skb(sk
, buff
, 0, GFP_ATOMIC
);
2254 /* This routine sends a packet with an out of date sequence
2255 * number. It assumes the other end will try to ack it.
2257 * Question: what should we make while urgent mode?
2258 * 4.4BSD forces sending single byte of data. We cannot send
2259 * out of window data, because we have SND.NXT==SND.MAX...
2261 * Current solution: to send TWO zero-length segments in urgent mode:
2262 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2263 * out-of-date with SND.UNA-1 to probe window.
2265 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
)
2267 struct tcp_sock
*tp
= tcp_sk(sk
);
2268 struct sk_buff
*skb
;
2270 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2271 skb
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2275 /* Reserve space for headers and set control bits. */
2276 skb_reserve(skb
, MAX_TCP_HEADER
);
2278 TCP_SKB_CB(skb
)->flags
= TCPCB_FLAG_ACK
;
2279 TCP_SKB_CB(skb
)->sacked
= urgent
;
2280 skb_shinfo(skb
)->tso_segs
= 1;
2281 skb_shinfo(skb
)->tso_size
= 0;
2283 /* Use a previous sequence. This should cause the other
2284 * end to send an ack. Don't queue or clone SKB, just
2287 TCP_SKB_CB(skb
)->seq
= urgent
? tp
->snd_una
: tp
->snd_una
- 1;
2288 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
;
2289 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2290 return tcp_transmit_skb(sk
, skb
, 0, GFP_ATOMIC
);
2293 int tcp_write_wakeup(struct sock
*sk
)
2295 if (sk
->sk_state
!= TCP_CLOSE
) {
2296 struct tcp_sock
*tp
= tcp_sk(sk
);
2297 struct sk_buff
*skb
;
2299 if ((skb
= sk
->sk_send_head
) != NULL
&&
2300 before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
+tp
->snd_wnd
)) {
2302 unsigned int mss
= tcp_current_mss(sk
, 0);
2303 unsigned int seg_size
= tp
->snd_una
+tp
->snd_wnd
-TCP_SKB_CB(skb
)->seq
;
2305 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
2306 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
2308 /* We are probing the opening of a window
2309 * but the window size is != 0
2310 * must have been a result SWS avoidance ( sender )
2312 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
2314 seg_size
= min(seg_size
, mss
);
2315 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2316 if (tcp_fragment(sk
, skb
, seg_size
, mss
))
2318 } else if (!tcp_skb_pcount(skb
))
2319 tcp_set_skb_tso_segs(sk
, skb
, mss
);
2321 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2322 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2323 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2325 update_send_head(sk
, tp
, skb
);
2330 between(tp
->snd_up
, tp
->snd_una
+1, tp
->snd_una
+0xFFFF))
2331 tcp_xmit_probe_skb(sk
, TCPCB_URG
);
2332 return tcp_xmit_probe_skb(sk
, 0);
2338 /* A window probe timeout has occurred. If window is not closed send
2339 * a partial packet else a zero probe.
2341 void tcp_send_probe0(struct sock
*sk
)
2343 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2344 struct tcp_sock
*tp
= tcp_sk(sk
);
2347 err
= tcp_write_wakeup(sk
);
2349 if (tp
->packets_out
|| !sk
->sk_send_head
) {
2350 /* Cancel probe timer, if it is not required. */
2351 icsk
->icsk_probes_out
= 0;
2352 icsk
->icsk_backoff
= 0;
2357 if (icsk
->icsk_backoff
< sysctl_tcp_retries2
)
2358 icsk
->icsk_backoff
++;
2359 icsk
->icsk_probes_out
++;
2360 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2361 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
2364 /* If packet was not sent due to local congestion,
2365 * do not backoff and do not remember icsk_probes_out.
2366 * Let local senders to fight for local resources.
2368 * Use accumulated backoff yet.
2370 if (!icsk
->icsk_probes_out
)
2371 icsk
->icsk_probes_out
= 1;
2372 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2373 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
,
2374 TCP_RESOURCE_PROBE_INTERVAL
),
2379 EXPORT_SYMBOL(tcp_connect
);
2380 EXPORT_SYMBOL(tcp_make_synack
);
2381 EXPORT_SYMBOL(tcp_simple_retransmit
);
2382 EXPORT_SYMBOL(tcp_sync_mss
);
2383 EXPORT_SYMBOL(sysctl_tcp_tso_win_divisor
);
2384 EXPORT_SYMBOL(tcp_mtup_init
);