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 /* People can turn this on to work with those rare, broken TCPs that
49 * interpret the window field as a signed quantity.
51 int sysctl_tcp_workaround_signed_windows
= 0;
53 /* This limits the percentage of the congestion window which we
54 * will allow a single TSO frame to consume. Building TSO frames
55 * which are too large can cause TCP streams to be bursty.
57 int sysctl_tcp_tso_win_divisor
= 3;
59 int sysctl_tcp_mtu_probing
= 0;
60 int sysctl_tcp_base_mss
= 512;
62 static void update_send_head(struct sock
*sk
, struct tcp_sock
*tp
,
65 sk
->sk_send_head
= skb
->next
;
66 if (sk
->sk_send_head
== (struct sk_buff
*)&sk
->sk_write_queue
)
67 sk
->sk_send_head
= NULL
;
68 tp
->snd_nxt
= TCP_SKB_CB(skb
)->end_seq
;
69 tcp_packets_out_inc(sk
, tp
, skb
);
72 /* SND.NXT, if window was not shrunk.
73 * If window has been shrunk, what should we make? It is not clear at all.
74 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
75 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
76 * invalid. OK, let's make this for now:
78 static inline __u32
tcp_acceptable_seq(struct sock
*sk
, struct tcp_sock
*tp
)
80 if (!before(tp
->snd_una
+tp
->snd_wnd
, tp
->snd_nxt
))
83 return tp
->snd_una
+tp
->snd_wnd
;
86 /* Calculate mss to advertise in SYN segment.
87 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
89 * 1. It is independent of path mtu.
90 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
91 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
92 * attached devices, because some buggy hosts are confused by
94 * 4. We do not make 3, we advertise MSS, calculated from first
95 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
96 * This may be overridden via information stored in routing table.
97 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
98 * probably even Jumbo".
100 static __u16
tcp_advertise_mss(struct sock
*sk
)
102 struct tcp_sock
*tp
= tcp_sk(sk
);
103 struct dst_entry
*dst
= __sk_dst_get(sk
);
104 int mss
= tp
->advmss
;
106 if (dst
&& dst_metric(dst
, RTAX_ADVMSS
) < mss
) {
107 mss
= dst_metric(dst
, RTAX_ADVMSS
);
114 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
115 * This is the first part of cwnd validation mechanism. */
116 static void tcp_cwnd_restart(struct sock
*sk
, struct dst_entry
*dst
)
118 struct tcp_sock
*tp
= tcp_sk(sk
);
119 s32 delta
= tcp_time_stamp
- tp
->lsndtime
;
120 u32 restart_cwnd
= tcp_init_cwnd(tp
, dst
);
121 u32 cwnd
= tp
->snd_cwnd
;
123 tcp_ca_event(sk
, CA_EVENT_CWND_RESTART
);
125 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
126 restart_cwnd
= min(restart_cwnd
, cwnd
);
128 while ((delta
-= inet_csk(sk
)->icsk_rto
) > 0 && cwnd
> restart_cwnd
)
130 tp
->snd_cwnd
= max(cwnd
, restart_cwnd
);
131 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
132 tp
->snd_cwnd_used
= 0;
135 static void tcp_event_data_sent(struct tcp_sock
*tp
,
136 struct sk_buff
*skb
, struct sock
*sk
)
138 struct inet_connection_sock
*icsk
= inet_csk(sk
);
139 const u32 now
= tcp_time_stamp
;
141 if (!tp
->packets_out
&& (s32
)(now
- tp
->lsndtime
) > icsk
->icsk_rto
)
142 tcp_cwnd_restart(sk
, __sk_dst_get(sk
));
146 /* If it is a reply for ato after last received
147 * packet, enter pingpong mode.
149 if ((u32
)(now
- icsk
->icsk_ack
.lrcvtime
) < icsk
->icsk_ack
.ato
)
150 icsk
->icsk_ack
.pingpong
= 1;
153 static inline void tcp_event_ack_sent(struct sock
*sk
, unsigned int pkts
)
155 tcp_dec_quickack_mode(sk
, pkts
);
156 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_DACK
);
159 /* Determine a window scaling and initial window to offer.
160 * Based on the assumption that the given amount of space
161 * will be offered. Store the results in the tp structure.
162 * NOTE: for smooth operation initial space offering should
163 * be a multiple of mss if possible. We assume here that mss >= 1.
164 * This MUST be enforced by all callers.
166 void tcp_select_initial_window(int __space
, __u32 mss
,
167 __u32
*rcv_wnd
, __u32
*window_clamp
,
168 int wscale_ok
, __u8
*rcv_wscale
)
170 unsigned int space
= (__space
< 0 ? 0 : __space
);
172 /* If no clamp set the clamp to the max possible scaled window */
173 if (*window_clamp
== 0)
174 (*window_clamp
) = (65535 << 14);
175 space
= min(*window_clamp
, space
);
177 /* Quantize space offering to a multiple of mss if possible. */
179 space
= (space
/ mss
) * mss
;
181 /* NOTE: offering an initial window larger than 32767
182 * will break some buggy TCP stacks. If the admin tells us
183 * it is likely we could be speaking with such a buggy stack
184 * we will truncate our initial window offering to 32K-1
185 * unless the remote has sent us a window scaling option,
186 * which we interpret as a sign the remote TCP is not
187 * misinterpreting the window field as a signed quantity.
189 if (sysctl_tcp_workaround_signed_windows
)
190 (*rcv_wnd
) = min(space
, MAX_TCP_WINDOW
);
196 /* Set window scaling on max possible window
197 * See RFC1323 for an explanation of the limit to 14
199 space
= max_t(u32
, sysctl_tcp_rmem
[2], sysctl_rmem_max
);
200 while (space
> 65535 && (*rcv_wscale
) < 14) {
206 /* Set initial window to value enough for senders,
207 * following RFC2414. Senders, not following this RFC,
208 * will be satisfied with 2.
210 if (mss
> (1<<*rcv_wscale
)) {
216 if (*rcv_wnd
> init_cwnd
*mss
)
217 *rcv_wnd
= init_cwnd
*mss
;
220 /* Set the clamp no higher than max representable value */
221 (*window_clamp
) = min(65535U << (*rcv_wscale
), *window_clamp
);
224 /* Chose a new window to advertise, update state in tcp_sock for the
225 * socket, and return result with RFC1323 scaling applied. The return
226 * value can be stuffed directly into th->window for an outgoing
229 static u16
tcp_select_window(struct sock
*sk
)
231 struct tcp_sock
*tp
= tcp_sk(sk
);
232 u32 cur_win
= tcp_receive_window(tp
);
233 u32 new_win
= __tcp_select_window(sk
);
235 /* Never shrink the offered window */
236 if(new_win
< cur_win
) {
237 /* Danger Will Robinson!
238 * Don't update rcv_wup/rcv_wnd here or else
239 * we will not be able to advertise a zero
240 * window in time. --DaveM
242 * Relax Will Robinson.
246 tp
->rcv_wnd
= new_win
;
247 tp
->rcv_wup
= tp
->rcv_nxt
;
249 /* Make sure we do not exceed the maximum possible
252 if (!tp
->rx_opt
.rcv_wscale
&& sysctl_tcp_workaround_signed_windows
)
253 new_win
= min(new_win
, MAX_TCP_WINDOW
);
255 new_win
= min(new_win
, (65535U << tp
->rx_opt
.rcv_wscale
));
257 /* RFC1323 scaling applied */
258 new_win
>>= tp
->rx_opt
.rcv_wscale
;
260 /* If we advertise zero window, disable fast path. */
267 static void tcp_build_and_update_options(__u32
*ptr
, struct tcp_sock
*tp
,
270 if (tp
->rx_opt
.tstamp_ok
) {
271 *ptr
++ = __constant_htonl((TCPOPT_NOP
<< 24) |
273 (TCPOPT_TIMESTAMP
<< 8) |
275 *ptr
++ = htonl(tstamp
);
276 *ptr
++ = htonl(tp
->rx_opt
.ts_recent
);
278 if (tp
->rx_opt
.eff_sacks
) {
279 struct tcp_sack_block
*sp
= tp
->rx_opt
.dsack
? tp
->duplicate_sack
: tp
->selective_acks
;
282 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
285 (TCPOLEN_SACK_BASE
+ (tp
->rx_opt
.eff_sacks
*
286 TCPOLEN_SACK_PERBLOCK
)));
287 for(this_sack
= 0; this_sack
< tp
->rx_opt
.eff_sacks
; this_sack
++) {
288 *ptr
++ = htonl(sp
[this_sack
].start_seq
);
289 *ptr
++ = htonl(sp
[this_sack
].end_seq
);
291 if (tp
->rx_opt
.dsack
) {
292 tp
->rx_opt
.dsack
= 0;
293 tp
->rx_opt
.eff_sacks
--;
298 /* Construct a tcp options header for a SYN or SYN_ACK packet.
299 * If this is every changed make sure to change the definition of
300 * MAX_SYN_SIZE to match the new maximum number of options that you
303 static void tcp_syn_build_options(__u32
*ptr
, int mss
, int ts
, int sack
,
304 int offer_wscale
, int wscale
, __u32 tstamp
,
307 /* We always get an MSS option.
308 * The option bytes which will be seen in normal data
309 * packets should timestamps be used, must be in the MSS
310 * advertised. But we subtract them from tp->mss_cache so
311 * that calculations in tcp_sendmsg are simpler etc.
312 * So account for this fact here if necessary. If we
313 * don't do this correctly, as a receiver we won't
314 * recognize data packets as being full sized when we
315 * should, and thus we won't abide by the delayed ACK
317 * SACKs don't matter, we never delay an ACK when we
318 * have any of those going out.
320 *ptr
++ = htonl((TCPOPT_MSS
<< 24) | (TCPOLEN_MSS
<< 16) | mss
);
323 *ptr
++ = __constant_htonl((TCPOPT_SACK_PERM
<< 24) | (TCPOLEN_SACK_PERM
<< 16) |
324 (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
);
326 *ptr
++ = __constant_htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16) |
327 (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
);
328 *ptr
++ = htonl(tstamp
); /* TSVAL */
329 *ptr
++ = htonl(ts_recent
); /* TSECR */
331 *ptr
++ = __constant_htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16) |
332 (TCPOPT_SACK_PERM
<< 8) | TCPOLEN_SACK_PERM
);
334 *ptr
++ = htonl((TCPOPT_NOP
<< 24) | (TCPOPT_WINDOW
<< 16) | (TCPOLEN_WINDOW
<< 8) | (wscale
));
337 /* This routine actually transmits TCP packets queued in by
338 * tcp_do_sendmsg(). This is used by both the initial
339 * transmission and possible later retransmissions.
340 * All SKB's seen here are completely headerless. It is our
341 * job to build the TCP header, and pass the packet down to
342 * IP so it can do the same plus pass the packet off to the
345 * We are working here with either a clone of the original
346 * SKB, or a fresh unique copy made by the retransmit engine.
348 static int tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int clone_it
, gfp_t gfp_mask
)
350 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
351 struct inet_sock
*inet
;
353 struct tcp_skb_cb
*tcb
;
359 BUG_ON(!skb
|| !tcp_skb_pcount(skb
));
361 /* If congestion control is doing timestamping, we must
362 * take such a timestamp before we potentially clone/copy.
364 if (icsk
->icsk_ca_ops
->rtt_sample
)
365 __net_timestamp(skb
);
367 if (likely(clone_it
)) {
368 if (unlikely(skb_cloned(skb
)))
369 skb
= pskb_copy(skb
, gfp_mask
);
371 skb
= skb_clone(skb
, gfp_mask
);
378 tcb
= TCP_SKB_CB(skb
);
379 tcp_header_size
= tp
->tcp_header_len
;
381 #define SYSCTL_FLAG_TSTAMPS 0x1
382 #define SYSCTL_FLAG_WSCALE 0x2
383 #define SYSCTL_FLAG_SACK 0x4
386 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
387 tcp_header_size
= sizeof(struct tcphdr
) + TCPOLEN_MSS
;
388 if(sysctl_tcp_timestamps
) {
389 tcp_header_size
+= TCPOLEN_TSTAMP_ALIGNED
;
390 sysctl_flags
|= SYSCTL_FLAG_TSTAMPS
;
392 if (sysctl_tcp_window_scaling
) {
393 tcp_header_size
+= TCPOLEN_WSCALE_ALIGNED
;
394 sysctl_flags
|= SYSCTL_FLAG_WSCALE
;
396 if (sysctl_tcp_sack
) {
397 sysctl_flags
|= SYSCTL_FLAG_SACK
;
398 if (!(sysctl_flags
& SYSCTL_FLAG_TSTAMPS
))
399 tcp_header_size
+= TCPOLEN_SACKPERM_ALIGNED
;
401 } else if (unlikely(tp
->rx_opt
.eff_sacks
)) {
402 /* A SACK is 2 pad bytes, a 2 byte header, plus
403 * 2 32-bit sequence numbers for each SACK block.
405 tcp_header_size
+= (TCPOLEN_SACK_BASE_ALIGNED
+
406 (tp
->rx_opt
.eff_sacks
*
407 TCPOLEN_SACK_PERBLOCK
));
410 if (tcp_packets_in_flight(tp
) == 0)
411 tcp_ca_event(sk
, CA_EVENT_TX_START
);
413 th
= (struct tcphdr
*) skb_push(skb
, tcp_header_size
);
415 skb_set_owner_w(skb
, sk
);
417 /* Build TCP header and checksum it. */
418 th
->source
= inet
->sport
;
419 th
->dest
= inet
->dport
;
420 th
->seq
= htonl(tcb
->seq
);
421 th
->ack_seq
= htonl(tp
->rcv_nxt
);
422 *(((__u16
*)th
) + 6) = htons(((tcp_header_size
>> 2) << 12) |
425 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
426 /* RFC1323: The window in SYN & SYN/ACK segments
429 th
->window
= htons(tp
->rcv_wnd
);
431 th
->window
= htons(tcp_select_window(sk
));
436 if (unlikely(tp
->urg_mode
&&
437 between(tp
->snd_up
, tcb
->seq
+1, tcb
->seq
+0xFFFF))) {
438 th
->urg_ptr
= htons(tp
->snd_up
-tcb
->seq
);
442 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
443 tcp_syn_build_options((__u32
*)(th
+ 1),
444 tcp_advertise_mss(sk
),
445 (sysctl_flags
& SYSCTL_FLAG_TSTAMPS
),
446 (sysctl_flags
& SYSCTL_FLAG_SACK
),
447 (sysctl_flags
& SYSCTL_FLAG_WSCALE
),
448 tp
->rx_opt
.rcv_wscale
,
450 tp
->rx_opt
.ts_recent
);
452 tcp_build_and_update_options((__u32
*)(th
+ 1),
454 TCP_ECN_send(sk
, tp
, skb
, tcp_header_size
);
457 icsk
->icsk_af_ops
->send_check(sk
, skb
->len
, skb
);
459 if (likely(tcb
->flags
& TCPCB_FLAG_ACK
))
460 tcp_event_ack_sent(sk
, tcp_skb_pcount(skb
));
462 if (skb
->len
!= tcp_header_size
)
463 tcp_event_data_sent(tp
, skb
, sk
);
465 TCP_INC_STATS(TCP_MIB_OUTSEGS
);
467 err
= icsk
->icsk_af_ops
->queue_xmit(skb
, 0);
468 if (unlikely(err
<= 0))
473 /* NET_XMIT_CN is special. It does not guarantee,
474 * that this packet is lost. It tells that device
475 * is about to start to drop packets or already
476 * drops some packets of the same priority and
477 * invokes us to send less aggressively.
479 return err
== NET_XMIT_CN
? 0 : err
;
481 #undef SYSCTL_FLAG_TSTAMPS
482 #undef SYSCTL_FLAG_WSCALE
483 #undef SYSCTL_FLAG_SACK
487 /* This routine just queue's the buffer
489 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
490 * otherwise socket can stall.
492 static void tcp_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
494 struct tcp_sock
*tp
= tcp_sk(sk
);
496 /* Advance write_seq and place onto the write_queue. */
497 tp
->write_seq
= TCP_SKB_CB(skb
)->end_seq
;
498 skb_header_release(skb
);
499 __skb_queue_tail(&sk
->sk_write_queue
, skb
);
500 sk_charge_skb(sk
, skb
);
502 /* Queue it, remembering where we must start sending. */
503 if (sk
->sk_send_head
== NULL
)
504 sk
->sk_send_head
= skb
;
507 static void tcp_set_skb_tso_segs(struct sock
*sk
, struct sk_buff
*skb
, unsigned int mss_now
)
509 if (skb
->len
<= mss_now
||
510 !(sk
->sk_route_caps
& NETIF_F_TSO
)) {
511 /* Avoid the costly divide in the normal
514 skb_shinfo(skb
)->tso_segs
= 1;
515 skb_shinfo(skb
)->tso_size
= 0;
519 factor
= skb
->len
+ (mss_now
- 1);
521 skb_shinfo(skb
)->tso_segs
= factor
;
522 skb_shinfo(skb
)->tso_size
= mss_now
;
526 /* Function to create two new TCP segments. Shrinks the given segment
527 * to the specified size and appends a new segment with the rest of the
528 * packet to the list. This won't be called frequently, I hope.
529 * Remember, these are still headerless SKBs at this point.
531 int tcp_fragment(struct sock
*sk
, struct sk_buff
*skb
, u32 len
, unsigned int mss_now
)
533 struct tcp_sock
*tp
= tcp_sk(sk
);
534 struct sk_buff
*buff
;
535 int nsize
, old_factor
;
538 BUG_ON(len
> skb
->len
);
540 clear_all_retrans_hints(tp
);
541 nsize
= skb_headlen(skb
) - len
;
545 if (skb_cloned(skb
) &&
546 skb_is_nonlinear(skb
) &&
547 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
550 /* Get a new skb... force flag on. */
551 buff
= sk_stream_alloc_skb(sk
, nsize
, GFP_ATOMIC
);
553 return -ENOMEM
; /* We'll just try again later. */
555 buff
->truesize
= skb
->len
- len
;
556 skb
->truesize
-= buff
->truesize
;
558 /* Correct the sequence numbers. */
559 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
560 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
561 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
563 /* PSH and FIN should only be set in the second packet. */
564 flags
= TCP_SKB_CB(skb
)->flags
;
565 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
566 TCP_SKB_CB(buff
)->flags
= flags
;
567 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
568 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_AT_TAIL
;
570 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_HW
) {
571 /* Copy and checksum data tail into the new buffer. */
572 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
, skb_put(buff
, nsize
),
577 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
579 skb
->ip_summed
= CHECKSUM_HW
;
580 skb_split(skb
, buff
, len
);
583 buff
->ip_summed
= skb
->ip_summed
;
585 /* Looks stupid, but our code really uses when of
586 * skbs, which it never sent before. --ANK
588 TCP_SKB_CB(buff
)->when
= TCP_SKB_CB(skb
)->when
;
589 buff
->tstamp
= skb
->tstamp
;
591 old_factor
= tcp_skb_pcount(skb
);
593 /* Fix up tso_factor for both original and new SKB. */
594 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
595 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
597 /* If this packet has been sent out already, we must
598 * adjust the various packet counters.
600 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
601 int diff
= old_factor
- tcp_skb_pcount(skb
) -
602 tcp_skb_pcount(buff
);
604 tp
->packets_out
-= diff
;
606 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
607 tp
->sacked_out
-= diff
;
608 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
609 tp
->retrans_out
-= diff
;
611 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
) {
612 tp
->lost_out
-= diff
;
613 tp
->left_out
-= diff
;
617 /* Adjust Reno SACK estimate. */
618 if (!tp
->rx_opt
.sack_ok
) {
619 tp
->sacked_out
-= diff
;
620 if ((int)tp
->sacked_out
< 0)
622 tcp_sync_left_out(tp
);
625 tp
->fackets_out
-= diff
;
626 if ((int)tp
->fackets_out
< 0)
631 /* Link BUFF into the send queue. */
632 skb_header_release(buff
);
633 __skb_append(skb
, buff
, &sk
->sk_write_queue
);
638 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
639 * eventually). The difference is that pulled data not copied, but
640 * immediately discarded.
642 static unsigned char *__pskb_trim_head(struct sk_buff
*skb
, int len
)
648 for (i
=0; i
<skb_shinfo(skb
)->nr_frags
; i
++) {
649 if (skb_shinfo(skb
)->frags
[i
].size
<= eat
) {
650 put_page(skb_shinfo(skb
)->frags
[i
].page
);
651 eat
-= skb_shinfo(skb
)->frags
[i
].size
;
653 skb_shinfo(skb
)->frags
[k
] = skb_shinfo(skb
)->frags
[i
];
655 skb_shinfo(skb
)->frags
[k
].page_offset
+= eat
;
656 skb_shinfo(skb
)->frags
[k
].size
-= eat
;
662 skb_shinfo(skb
)->nr_frags
= k
;
664 skb
->tail
= skb
->data
;
665 skb
->data_len
-= len
;
666 skb
->len
= skb
->data_len
;
670 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
672 if (skb_cloned(skb
) &&
673 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
676 if (len
<= skb_headlen(skb
)) {
677 __skb_pull(skb
, len
);
679 if (__pskb_trim_head(skb
, len
-skb_headlen(skb
)) == NULL
)
683 TCP_SKB_CB(skb
)->seq
+= len
;
684 skb
->ip_summed
= CHECKSUM_HW
;
686 skb
->truesize
-= len
;
687 sk
->sk_wmem_queued
-= len
;
688 sk
->sk_forward_alloc
+= len
;
689 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
691 /* Any change of skb->len requires recalculation of tso
694 if (tcp_skb_pcount(skb
) > 1)
695 tcp_set_skb_tso_segs(sk
, skb
, tcp_current_mss(sk
, 1));
700 /* Not accounting for SACKs here. */
701 int tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
703 struct tcp_sock
*tp
= tcp_sk(sk
);
704 struct inet_connection_sock
*icsk
= inet_csk(sk
);
707 /* Calculate base mss without TCP options:
708 It is MMS_S - sizeof(tcphdr) of rfc1122
710 mss_now
= pmtu
- icsk
->icsk_af_ops
->net_header_len
- sizeof(struct tcphdr
);
712 /* Clamp it (mss_clamp does not include tcp options) */
713 if (mss_now
> tp
->rx_opt
.mss_clamp
)
714 mss_now
= tp
->rx_opt
.mss_clamp
;
716 /* Now subtract optional transport overhead */
717 mss_now
-= icsk
->icsk_ext_hdr_len
;
719 /* Then reserve room for full set of TCP options and 8 bytes of data */
723 /* Now subtract TCP options size, not including SACKs */
724 mss_now
-= tp
->tcp_header_len
- sizeof(struct tcphdr
);
729 /* Inverse of above */
730 int tcp_mss_to_mtu(struct sock
*sk
, int mss
)
732 struct tcp_sock
*tp
= tcp_sk(sk
);
733 struct inet_connection_sock
*icsk
= inet_csk(sk
);
738 icsk
->icsk_ext_hdr_len
+
739 icsk
->icsk_af_ops
->net_header_len
;
744 void tcp_mtup_init(struct sock
*sk
)
746 struct tcp_sock
*tp
= tcp_sk(sk
);
747 struct inet_connection_sock
*icsk
= inet_csk(sk
);
749 icsk
->icsk_mtup
.enabled
= sysctl_tcp_mtu_probing
> 1;
750 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+ sizeof(struct tcphdr
) +
751 icsk
->icsk_af_ops
->net_header_len
;
752 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, sysctl_tcp_base_mss
);
753 icsk
->icsk_mtup
.probe_size
= 0;
756 /* This function synchronize snd mss to current pmtu/exthdr set.
758 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
759 for TCP options, but includes only bare TCP header.
761 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
762 It is minimum of user_mss and mss received with SYN.
763 It also does not include TCP options.
765 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
767 tp->mss_cache is current effective sending mss, including
768 all tcp options except for SACKs. It is evaluated,
769 taking into account current pmtu, but never exceeds
770 tp->rx_opt.mss_clamp.
772 NOTE1. rfc1122 clearly states that advertised MSS
773 DOES NOT include either tcp or ip options.
775 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
776 are READ ONLY outside this function. --ANK (980731)
779 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
781 struct tcp_sock
*tp
= tcp_sk(sk
);
782 struct inet_connection_sock
*icsk
= inet_csk(sk
);
785 if (icsk
->icsk_mtup
.search_high
> pmtu
)
786 icsk
->icsk_mtup
.search_high
= pmtu
;
788 mss_now
= tcp_mtu_to_mss(sk
, pmtu
);
790 /* Bound mss with half of window */
791 if (tp
->max_window
&& mss_now
> (tp
->max_window
>>1))
792 mss_now
= max((tp
->max_window
>>1), 68U - tp
->tcp_header_len
);
794 /* And store cached results */
795 icsk
->icsk_pmtu_cookie
= pmtu
;
796 if (icsk
->icsk_mtup
.enabled
)
797 mss_now
= min(mss_now
, tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
));
798 tp
->mss_cache
= mss_now
;
803 /* Compute the current effective MSS, taking SACKs and IP options,
804 * and even PMTU discovery events into account.
806 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
807 * cannot be large. However, taking into account rare use of URG, this
810 unsigned int tcp_current_mss(struct sock
*sk
, int large_allowed
)
812 struct tcp_sock
*tp
= tcp_sk(sk
);
813 struct dst_entry
*dst
= __sk_dst_get(sk
);
818 mss_now
= tp
->mss_cache
;
821 (sk
->sk_route_caps
& NETIF_F_TSO
) &&
826 u32 mtu
= dst_mtu(dst
);
827 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
828 mss_now
= tcp_sync_mss(sk
, mtu
);
831 if (tp
->rx_opt
.eff_sacks
)
832 mss_now
-= (TCPOLEN_SACK_BASE_ALIGNED
+
833 (tp
->rx_opt
.eff_sacks
* TCPOLEN_SACK_PERBLOCK
));
835 xmit_size_goal
= mss_now
;
838 xmit_size_goal
= (65535 -
839 inet_csk(sk
)->icsk_af_ops
->net_header_len
-
840 inet_csk(sk
)->icsk_ext_hdr_len
-
843 if (tp
->max_window
&&
844 (xmit_size_goal
> (tp
->max_window
>> 1)))
845 xmit_size_goal
= max((tp
->max_window
>> 1),
846 68U - tp
->tcp_header_len
);
848 xmit_size_goal
-= (xmit_size_goal
% mss_now
);
850 tp
->xmit_size_goal
= xmit_size_goal
;
855 /* Congestion window validation. (RFC2861) */
857 static void tcp_cwnd_validate(struct sock
*sk
, struct tcp_sock
*tp
)
859 __u32 packets_out
= tp
->packets_out
;
861 if (packets_out
>= tp
->snd_cwnd
) {
862 /* Network is feed fully. */
863 tp
->snd_cwnd_used
= 0;
864 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
866 /* Network starves. */
867 if (tp
->packets_out
> tp
->snd_cwnd_used
)
868 tp
->snd_cwnd_used
= tp
->packets_out
;
870 if ((s32
)(tcp_time_stamp
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
)
871 tcp_cwnd_application_limited(sk
);
875 static unsigned int tcp_window_allows(struct tcp_sock
*tp
, struct sk_buff
*skb
, unsigned int mss_now
, unsigned int cwnd
)
877 u32 window
, cwnd_len
;
879 window
= (tp
->snd_una
+ tp
->snd_wnd
- TCP_SKB_CB(skb
)->seq
);
880 cwnd_len
= mss_now
* cwnd
;
881 return min(window
, cwnd_len
);
884 /* Can at least one segment of SKB be sent right now, according to the
885 * congestion window rules? If so, return how many segments are allowed.
887 static inline unsigned int tcp_cwnd_test(struct tcp_sock
*tp
, struct sk_buff
*skb
)
891 /* Don't be strict about the congestion window for the final FIN. */
892 if (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
)
895 in_flight
= tcp_packets_in_flight(tp
);
897 if (in_flight
< cwnd
)
898 return (cwnd
- in_flight
);
903 /* This must be invoked the first time we consider transmitting
906 static int tcp_init_tso_segs(struct sock
*sk
, struct sk_buff
*skb
, unsigned int mss_now
)
908 int tso_segs
= tcp_skb_pcount(skb
);
912 skb_shinfo(skb
)->tso_size
!= mss_now
)) {
913 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
914 tso_segs
= tcp_skb_pcount(skb
);
919 static inline int tcp_minshall_check(const struct tcp_sock
*tp
)
921 return after(tp
->snd_sml
,tp
->snd_una
) &&
922 !after(tp
->snd_sml
, tp
->snd_nxt
);
925 /* Return 0, if packet can be sent now without violation Nagle's rules:
926 * 1. It is full sized.
927 * 2. Or it contains FIN. (already checked by caller)
928 * 3. Or TCP_NODELAY was set.
929 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
930 * With Minshall's modification: all sent small packets are ACKed.
933 static inline int tcp_nagle_check(const struct tcp_sock
*tp
,
934 const struct sk_buff
*skb
,
935 unsigned mss_now
, int nonagle
)
937 return (skb
->len
< mss_now
&&
938 ((nonagle
&TCP_NAGLE_CORK
) ||
941 tcp_minshall_check(tp
))));
944 /* Return non-zero if the Nagle test allows this packet to be
947 static inline int tcp_nagle_test(struct tcp_sock
*tp
, struct sk_buff
*skb
,
948 unsigned int cur_mss
, int nonagle
)
950 /* Nagle rule does not apply to frames, which sit in the middle of the
951 * write_queue (they have no chances to get new data).
953 * This is implemented in the callers, where they modify the 'nonagle'
954 * argument based upon the location of SKB in the send queue.
956 if (nonagle
& TCP_NAGLE_PUSH
)
959 /* Don't use the nagle rule for urgent data (or for the final FIN). */
961 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
))
964 if (!tcp_nagle_check(tp
, skb
, cur_mss
, nonagle
))
970 /* Does at least the first segment of SKB fit into the send window? */
971 static inline int tcp_snd_wnd_test(struct tcp_sock
*tp
, struct sk_buff
*skb
, unsigned int cur_mss
)
973 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
975 if (skb
->len
> cur_mss
)
976 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
978 return !after(end_seq
, tp
->snd_una
+ tp
->snd_wnd
);
981 /* This checks if the data bearing packet SKB (usually sk->sk_send_head)
982 * should be put on the wire right now. If so, it returns the number of
983 * packets allowed by the congestion window.
985 static unsigned int tcp_snd_test(struct sock
*sk
, struct sk_buff
*skb
,
986 unsigned int cur_mss
, int nonagle
)
988 struct tcp_sock
*tp
= tcp_sk(sk
);
989 unsigned int cwnd_quota
;
991 tcp_init_tso_segs(sk
, skb
, cur_mss
);
993 if (!tcp_nagle_test(tp
, skb
, cur_mss
, nonagle
))
996 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
998 !tcp_snd_wnd_test(tp
, skb
, cur_mss
))
1004 static inline int tcp_skb_is_last(const struct sock
*sk
,
1005 const struct sk_buff
*skb
)
1007 return skb
->next
== (struct sk_buff
*)&sk
->sk_write_queue
;
1010 int tcp_may_send_now(struct sock
*sk
, struct tcp_sock
*tp
)
1012 struct sk_buff
*skb
= sk
->sk_send_head
;
1015 tcp_snd_test(sk
, skb
, tcp_current_mss(sk
, 1),
1016 (tcp_skb_is_last(sk
, skb
) ?
1021 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1022 * which is put after SKB on the list. It is very much like
1023 * tcp_fragment() except that it may make several kinds of assumptions
1024 * in order to speed up the splitting operation. In particular, we
1025 * know that all the data is in scatter-gather pages, and that the
1026 * packet has never been sent out before (and thus is not cloned).
1028 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
, unsigned int mss_now
)
1030 struct sk_buff
*buff
;
1031 int nlen
= skb
->len
- len
;
1034 /* All of a TSO frame must be composed of paged data. */
1035 if (skb
->len
!= skb
->data_len
)
1036 return tcp_fragment(sk
, skb
, len
, mss_now
);
1038 buff
= sk_stream_alloc_pskb(sk
, 0, 0, GFP_ATOMIC
);
1039 if (unlikely(buff
== NULL
))
1042 buff
->truesize
= nlen
;
1043 skb
->truesize
-= nlen
;
1045 /* Correct the sequence numbers. */
1046 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1047 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1048 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1050 /* PSH and FIN should only be set in the second packet. */
1051 flags
= TCP_SKB_CB(skb
)->flags
;
1052 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
1053 TCP_SKB_CB(buff
)->flags
= flags
;
1055 /* This packet was never sent out yet, so no SACK bits. */
1056 TCP_SKB_CB(buff
)->sacked
= 0;
1058 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_HW
;
1059 skb_split(skb
, buff
, len
);
1061 /* Fix up tso_factor for both original and new SKB. */
1062 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1063 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1065 /* Link BUFF into the send queue. */
1066 skb_header_release(buff
);
1067 __skb_append(skb
, buff
, &sk
->sk_write_queue
);
1072 /* Try to defer sending, if possible, in order to minimize the amount
1073 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1075 * This algorithm is from John Heffner.
1077 static int tcp_tso_should_defer(struct sock
*sk
, struct tcp_sock
*tp
, struct sk_buff
*skb
)
1079 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1080 u32 send_win
, cong_win
, limit
, in_flight
;
1082 if (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
)
1085 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
1088 in_flight
= tcp_packets_in_flight(tp
);
1090 BUG_ON(tcp_skb_pcount(skb
) <= 1 ||
1091 (tp
->snd_cwnd
<= in_flight
));
1093 send_win
= (tp
->snd_una
+ tp
->snd_wnd
) - TCP_SKB_CB(skb
)->seq
;
1095 /* From in_flight test above, we know that cwnd > in_flight. */
1096 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
1098 limit
= min(send_win
, cong_win
);
1100 /* If a full-sized TSO skb can be sent, do it. */
1104 if (sysctl_tcp_tso_win_divisor
) {
1105 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
1107 /* If at least some fraction of a window is available,
1110 chunk
/= sysctl_tcp_tso_win_divisor
;
1114 /* Different approach, try not to defer past a single
1115 * ACK. Receiver should ACK every other full sized
1116 * frame, so if we have space for more than 3 frames
1119 if (limit
> tcp_max_burst(tp
) * tp
->mss_cache
)
1123 /* Ok, it looks like it is advisable to defer. */
1127 /* Create a new MTU probe if we are ready.
1128 * Returns 0 if we should wait to probe (no cwnd available),
1129 * 1 if a probe was sent,
1131 static int tcp_mtu_probe(struct sock
*sk
)
1133 struct tcp_sock
*tp
= tcp_sk(sk
);
1134 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1135 struct sk_buff
*skb
, *nskb
, *next
;
1142 /* Not currently probing/verifying,
1144 * have enough cwnd, and
1145 * not SACKing (the variable headers throw things off) */
1146 if (!icsk
->icsk_mtup
.enabled
||
1147 icsk
->icsk_mtup
.probe_size
||
1148 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
||
1149 tp
->snd_cwnd
< 11 ||
1150 tp
->rx_opt
.eff_sacks
)
1153 /* Very simple search strategy: just double the MSS. */
1154 mss_now
= tcp_current_mss(sk
, 0);
1155 probe_size
= 2*tp
->mss_cache
;
1156 if (probe_size
> tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_high
)) {
1157 /* TODO: set timer for probe_converge_event */
1161 /* Have enough data in the send queue to probe? */
1163 if ((skb
= sk
->sk_send_head
) == NULL
)
1165 while ((len
+= skb
->len
) < probe_size
&& !tcp_skb_is_last(sk
, skb
))
1167 if (len
< probe_size
)
1170 /* Receive window check. */
1171 if (after(TCP_SKB_CB(skb
)->seq
+ probe_size
, tp
->snd_una
+ tp
->snd_wnd
)) {
1172 if (tp
->snd_wnd
< probe_size
)
1178 /* Do we need to wait to drain cwnd? */
1179 pif
= tcp_packets_in_flight(tp
);
1180 if (pif
+ 2 > tp
->snd_cwnd
) {
1181 /* With no packets in flight, don't stall. */
1188 /* We're allowed to probe. Build it now. */
1189 if ((nskb
= sk_stream_alloc_skb(sk
, probe_size
, GFP_ATOMIC
)) == NULL
)
1191 sk_charge_skb(sk
, nskb
);
1193 skb
= sk
->sk_send_head
;
1194 __skb_insert(nskb
, skb
->prev
, skb
, &sk
->sk_write_queue
);
1195 sk
->sk_send_head
= nskb
;
1197 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(skb
)->seq
;
1198 TCP_SKB_CB(nskb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ probe_size
;
1199 TCP_SKB_CB(nskb
)->flags
= TCPCB_FLAG_ACK
;
1200 TCP_SKB_CB(nskb
)->sacked
= 0;
1202 if (skb
->ip_summed
== CHECKSUM_HW
)
1203 nskb
->ip_summed
= CHECKSUM_HW
;
1206 while (len
< probe_size
) {
1209 copy
= min_t(int, skb
->len
, probe_size
- len
);
1210 if (nskb
->ip_summed
)
1211 skb_copy_bits(skb
, 0, skb_put(nskb
, copy
), copy
);
1213 nskb
->csum
= skb_copy_and_csum_bits(skb
, 0,
1214 skb_put(nskb
, copy
), copy
, nskb
->csum
);
1216 if (skb
->len
<= copy
) {
1217 /* We've eaten all the data from this skb.
1219 TCP_SKB_CB(nskb
)->flags
|= TCP_SKB_CB(skb
)->flags
;
1220 __skb_unlink(skb
, &sk
->sk_write_queue
);
1221 sk_stream_free_skb(sk
, skb
);
1223 TCP_SKB_CB(nskb
)->flags
|= TCP_SKB_CB(skb
)->flags
&
1224 ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
1225 if (!skb_shinfo(skb
)->nr_frags
) {
1226 skb_pull(skb
, copy
);
1227 if (skb
->ip_summed
!= CHECKSUM_HW
)
1228 skb
->csum
= csum_partial(skb
->data
, skb
->len
, 0);
1230 __pskb_trim_head(skb
, copy
);
1231 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1233 TCP_SKB_CB(skb
)->seq
+= copy
;
1239 tcp_init_tso_segs(sk
, nskb
, nskb
->len
);
1241 /* We're ready to send. If this fails, the probe will
1242 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1243 TCP_SKB_CB(nskb
)->when
= tcp_time_stamp
;
1244 if (!tcp_transmit_skb(sk
, nskb
, 1, GFP_ATOMIC
)) {
1245 /* Decrement cwnd here because we are sending
1246 * effectively two packets. */
1248 update_send_head(sk
, tp
, nskb
);
1250 icsk
->icsk_mtup
.probe_size
= tcp_mss_to_mtu(sk
, nskb
->len
);
1251 tp
->mtu_probe
.probe_seq_start
= TCP_SKB_CB(nskb
)->seq
;
1252 tp
->mtu_probe
.probe_seq_end
= TCP_SKB_CB(nskb
)->end_seq
;
1261 /* This routine writes packets to the network. It advances the
1262 * send_head. This happens as incoming acks open up the remote
1265 * Returns 1, if no segments are in flight and we have queued segments, but
1266 * cannot send anything now because of SWS or another problem.
1268 static int tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
)
1270 struct tcp_sock
*tp
= tcp_sk(sk
);
1271 struct sk_buff
*skb
;
1272 unsigned int tso_segs
, sent_pkts
;
1276 /* If we are closed, the bytes will have to remain here.
1277 * In time closedown will finish, we empty the write queue and all
1280 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
1285 /* Do MTU probing. */
1286 if ((result
= tcp_mtu_probe(sk
)) == 0) {
1288 } else if (result
> 0) {
1292 while ((skb
= sk
->sk_send_head
)) {
1295 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1298 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1302 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
)))
1305 if (tso_segs
== 1) {
1306 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
1307 (tcp_skb_is_last(sk
, skb
) ?
1308 nonagle
: TCP_NAGLE_PUSH
))))
1311 if (tcp_tso_should_defer(sk
, tp
, skb
))
1317 limit
= tcp_window_allows(tp
, skb
,
1318 mss_now
, cwnd_quota
);
1320 if (skb
->len
< limit
) {
1321 unsigned int trim
= skb
->len
% mss_now
;
1324 limit
= skb
->len
- trim
;
1328 if (skb
->len
> limit
&&
1329 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
)))
1332 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1334 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
)))
1337 /* Advance the send_head. This one is sent out.
1338 * This call will increment packets_out.
1340 update_send_head(sk
, tp
, skb
);
1342 tcp_minshall_update(tp
, mss_now
, skb
);
1346 if (likely(sent_pkts
)) {
1347 tcp_cwnd_validate(sk
, tp
);
1350 return !tp
->packets_out
&& sk
->sk_send_head
;
1353 /* Push out any pending frames which were held back due to
1354 * TCP_CORK or attempt at coalescing tiny packets.
1355 * The socket must be locked by the caller.
1357 void __tcp_push_pending_frames(struct sock
*sk
, struct tcp_sock
*tp
,
1358 unsigned int cur_mss
, int nonagle
)
1360 struct sk_buff
*skb
= sk
->sk_send_head
;
1363 if (tcp_write_xmit(sk
, cur_mss
, nonagle
))
1364 tcp_check_probe_timer(sk
, tp
);
1368 /* Send _single_ skb sitting at the send head. This function requires
1369 * true push pending frames to setup probe timer etc.
1371 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
1373 struct tcp_sock
*tp
= tcp_sk(sk
);
1374 struct sk_buff
*skb
= sk
->sk_send_head
;
1375 unsigned int tso_segs
, cwnd_quota
;
1377 BUG_ON(!skb
|| skb
->len
< mss_now
);
1379 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1380 cwnd_quota
= tcp_snd_test(sk
, skb
, mss_now
, TCP_NAGLE_PUSH
);
1382 if (likely(cwnd_quota
)) {
1389 limit
= tcp_window_allows(tp
, skb
,
1390 mss_now
, cwnd_quota
);
1392 if (skb
->len
< limit
) {
1393 unsigned int trim
= skb
->len
% mss_now
;
1396 limit
= skb
->len
- trim
;
1400 if (skb
->len
> limit
&&
1401 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
)))
1404 /* Send it out now. */
1405 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1407 if (likely(!tcp_transmit_skb(sk
, skb
, 1, sk
->sk_allocation
))) {
1408 update_send_head(sk
, tp
, skb
);
1409 tcp_cwnd_validate(sk
, tp
);
1415 /* This function returns the amount that we can raise the
1416 * usable window based on the following constraints
1418 * 1. The window can never be shrunk once it is offered (RFC 793)
1419 * 2. We limit memory per socket
1422 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1423 * RECV.NEXT + RCV.WIN fixed until:
1424 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1426 * i.e. don't raise the right edge of the window until you can raise
1427 * it at least MSS bytes.
1429 * Unfortunately, the recommended algorithm breaks header prediction,
1430 * since header prediction assumes th->window stays fixed.
1432 * Strictly speaking, keeping th->window fixed violates the receiver
1433 * side SWS prevention criteria. The problem is that under this rule
1434 * a stream of single byte packets will cause the right side of the
1435 * window to always advance by a single byte.
1437 * Of course, if the sender implements sender side SWS prevention
1438 * then this will not be a problem.
1440 * BSD seems to make the following compromise:
1442 * If the free space is less than the 1/4 of the maximum
1443 * space available and the free space is less than 1/2 mss,
1444 * then set the window to 0.
1445 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1446 * Otherwise, just prevent the window from shrinking
1447 * and from being larger than the largest representable value.
1449 * This prevents incremental opening of the window in the regime
1450 * where TCP is limited by the speed of the reader side taking
1451 * data out of the TCP receive queue. It does nothing about
1452 * those cases where the window is constrained on the sender side
1453 * because the pipeline is full.
1455 * BSD also seems to "accidentally" limit itself to windows that are a
1456 * multiple of MSS, at least until the free space gets quite small.
1457 * This would appear to be a side effect of the mbuf implementation.
1458 * Combining these two algorithms results in the observed behavior
1459 * of having a fixed window size at almost all times.
1461 * Below we obtain similar behavior by forcing the offered window to
1462 * a multiple of the mss when it is feasible to do so.
1464 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1465 * Regular options like TIMESTAMP are taken into account.
1467 u32
__tcp_select_window(struct sock
*sk
)
1469 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1470 struct tcp_sock
*tp
= tcp_sk(sk
);
1471 /* MSS for the peer's data. Previous versions used mss_clamp
1472 * here. I don't know if the value based on our guesses
1473 * of peer's MSS is better for the performance. It's more correct
1474 * but may be worse for the performance because of rcv_mss
1475 * fluctuations. --SAW 1998/11/1
1477 int mss
= icsk
->icsk_ack
.rcv_mss
;
1478 int free_space
= tcp_space(sk
);
1479 int full_space
= min_t(int, tp
->window_clamp
, tcp_full_space(sk
));
1482 if (mss
> full_space
)
1485 if (free_space
< full_space
/2) {
1486 icsk
->icsk_ack
.quick
= 0;
1488 if (tcp_memory_pressure
)
1489 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U*tp
->advmss
);
1491 if (free_space
< mss
)
1495 if (free_space
> tp
->rcv_ssthresh
)
1496 free_space
= tp
->rcv_ssthresh
;
1498 /* Don't do rounding if we are using window scaling, since the
1499 * scaled window will not line up with the MSS boundary anyway.
1501 window
= tp
->rcv_wnd
;
1502 if (tp
->rx_opt
.rcv_wscale
) {
1503 window
= free_space
;
1505 /* Advertise enough space so that it won't get scaled away.
1506 * Import case: prevent zero window announcement if
1507 * 1<<rcv_wscale > mss.
1509 if (((window
>> tp
->rx_opt
.rcv_wscale
) << tp
->rx_opt
.rcv_wscale
) != window
)
1510 window
= (((window
>> tp
->rx_opt
.rcv_wscale
) + 1)
1511 << tp
->rx_opt
.rcv_wscale
);
1513 /* Get the largest window that is a nice multiple of mss.
1514 * Window clamp already applied above.
1515 * If our current window offering is within 1 mss of the
1516 * free space we just keep it. This prevents the divide
1517 * and multiply from happening most of the time.
1518 * We also don't do any window rounding when the free space
1521 if (window
<= free_space
- mss
|| window
> free_space
)
1522 window
= (free_space
/mss
)*mss
;
1528 /* Attempt to collapse two adjacent SKB's during retransmission. */
1529 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*skb
, int mss_now
)
1531 struct tcp_sock
*tp
= tcp_sk(sk
);
1532 struct sk_buff
*next_skb
= skb
->next
;
1534 /* The first test we must make is that neither of these two
1535 * SKB's are still referenced by someone else.
1537 if (!skb_cloned(skb
) && !skb_cloned(next_skb
)) {
1538 int skb_size
= skb
->len
, next_skb_size
= next_skb
->len
;
1539 u16 flags
= TCP_SKB_CB(skb
)->flags
;
1541 /* Also punt if next skb has been SACK'd. */
1542 if(TCP_SKB_CB(next_skb
)->sacked
& TCPCB_SACKED_ACKED
)
1545 /* Next skb is out of window. */
1546 if (after(TCP_SKB_CB(next_skb
)->end_seq
, tp
->snd_una
+tp
->snd_wnd
))
1549 /* Punt if not enough space exists in the first SKB for
1550 * the data in the second, or the total combined payload
1551 * would exceed the MSS.
1553 if ((next_skb_size
> skb_tailroom(skb
)) ||
1554 ((skb_size
+ next_skb_size
) > mss_now
))
1557 BUG_ON(tcp_skb_pcount(skb
) != 1 ||
1558 tcp_skb_pcount(next_skb
) != 1);
1560 /* changing transmit queue under us so clear hints */
1561 clear_all_retrans_hints(tp
);
1563 /* Ok. We will be able to collapse the packet. */
1564 __skb_unlink(next_skb
, &sk
->sk_write_queue
);
1566 memcpy(skb_put(skb
, next_skb_size
), next_skb
->data
, next_skb_size
);
1568 if (next_skb
->ip_summed
== CHECKSUM_HW
)
1569 skb
->ip_summed
= CHECKSUM_HW
;
1571 if (skb
->ip_summed
!= CHECKSUM_HW
)
1572 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
1574 /* Update sequence range on original skb. */
1575 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
1577 /* Merge over control information. */
1578 flags
|= TCP_SKB_CB(next_skb
)->flags
; /* This moves PSH/FIN etc. over */
1579 TCP_SKB_CB(skb
)->flags
= flags
;
1581 /* All done, get rid of second SKB and account for it so
1582 * packet counting does not break.
1584 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
&(TCPCB_EVER_RETRANS
|TCPCB_AT_TAIL
);
1585 if (TCP_SKB_CB(next_skb
)->sacked
&TCPCB_SACKED_RETRANS
)
1586 tp
->retrans_out
-= tcp_skb_pcount(next_skb
);
1587 if (TCP_SKB_CB(next_skb
)->sacked
&TCPCB_LOST
) {
1588 tp
->lost_out
-= tcp_skb_pcount(next_skb
);
1589 tp
->left_out
-= tcp_skb_pcount(next_skb
);
1591 /* Reno case is special. Sigh... */
1592 if (!tp
->rx_opt
.sack_ok
&& tp
->sacked_out
) {
1593 tcp_dec_pcount_approx(&tp
->sacked_out
, next_skb
);
1594 tp
->left_out
-= tcp_skb_pcount(next_skb
);
1597 /* Not quite right: it can be > snd.fack, but
1598 * it is better to underestimate fackets.
1600 tcp_dec_pcount_approx(&tp
->fackets_out
, next_skb
);
1601 tcp_packets_out_dec(tp
, next_skb
);
1602 sk_stream_free_skb(sk
, next_skb
);
1606 /* Do a simple retransmit without using the backoff mechanisms in
1607 * tcp_timer. This is used for path mtu discovery.
1608 * The socket is already locked here.
1610 void tcp_simple_retransmit(struct sock
*sk
)
1612 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1613 struct tcp_sock
*tp
= tcp_sk(sk
);
1614 struct sk_buff
*skb
;
1615 unsigned int mss
= tcp_current_mss(sk
, 0);
1618 sk_stream_for_retrans_queue(skb
, sk
) {
1619 if (skb
->len
> mss
&&
1620 !(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
)) {
1621 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) {
1622 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1623 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1625 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_LOST
)) {
1626 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1627 tp
->lost_out
+= tcp_skb_pcount(skb
);
1633 clear_all_retrans_hints(tp
);
1638 tcp_sync_left_out(tp
);
1640 /* Don't muck with the congestion window here.
1641 * Reason is that we do not increase amount of _data_
1642 * in network, but units changed and effective
1643 * cwnd/ssthresh really reduced now.
1645 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
1646 tp
->high_seq
= tp
->snd_nxt
;
1647 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
1648 tp
->prior_ssthresh
= 0;
1649 tp
->undo_marker
= 0;
1650 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1652 tcp_xmit_retransmit_queue(sk
);
1655 /* This retransmits one SKB. Policy decisions and retransmit queue
1656 * state updates are done by the caller. Returns non-zero if an
1657 * error occurred which prevented the send.
1659 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
1661 struct tcp_sock
*tp
= tcp_sk(sk
);
1662 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1663 unsigned int cur_mss
= tcp_current_mss(sk
, 0);
1666 /* Inconslusive MTU probe */
1667 if (icsk
->icsk_mtup
.probe_size
) {
1668 icsk
->icsk_mtup
.probe_size
= 0;
1671 /* Do not sent more than we queued. 1/4 is reserved for possible
1672 * copying overhead: fragmentation, tunneling, mangling etc.
1674 if (atomic_read(&sk
->sk_wmem_alloc
) >
1675 min(sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2), sk
->sk_sndbuf
))
1678 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
1679 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1681 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
1685 /* If receiver has shrunk his window, and skb is out of
1686 * new window, do not retransmit it. The exception is the
1687 * case, when window is shrunk to zero. In this case
1688 * our retransmit serves as a zero window probe.
1690 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
+tp
->snd_wnd
)
1691 && TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
1694 if (skb
->len
> cur_mss
) {
1695 if (tcp_fragment(sk
, skb
, cur_mss
, cur_mss
))
1696 return -ENOMEM
; /* We'll try again later. */
1699 /* Collapse two adjacent packets if worthwhile and we can. */
1700 if(!(TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_SYN
) &&
1701 (skb
->len
< (cur_mss
>> 1)) &&
1702 (skb
->next
!= sk
->sk_send_head
) &&
1703 (skb
->next
!= (struct sk_buff
*)&sk
->sk_write_queue
) &&
1704 (skb_shinfo(skb
)->nr_frags
== 0 && skb_shinfo(skb
->next
)->nr_frags
== 0) &&
1705 (tcp_skb_pcount(skb
) == 1 && tcp_skb_pcount(skb
->next
) == 1) &&
1706 (sysctl_tcp_retrans_collapse
!= 0))
1707 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
1709 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
1710 return -EHOSTUNREACH
; /* Routing failure or similar. */
1712 /* Some Solaris stacks overoptimize and ignore the FIN on a
1713 * retransmit when old data is attached. So strip it off
1714 * since it is cheap to do so and saves bytes on the network.
1717 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
) &&
1718 tp
->snd_una
== (TCP_SKB_CB(skb
)->end_seq
- 1)) {
1719 if (!pskb_trim(skb
, 0)) {
1720 TCP_SKB_CB(skb
)->seq
= TCP_SKB_CB(skb
)->end_seq
- 1;
1721 skb_shinfo(skb
)->tso_segs
= 1;
1722 skb_shinfo(skb
)->tso_size
= 0;
1723 skb
->ip_summed
= CHECKSUM_NONE
;
1728 /* Make a copy, if the first transmission SKB clone we made
1729 * is still in somebody's hands, else make a clone.
1731 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1733 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
1736 /* Update global TCP statistics. */
1737 TCP_INC_STATS(TCP_MIB_RETRANSSEGS
);
1739 tp
->total_retrans
++;
1741 #if FASTRETRANS_DEBUG > 0
1742 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) {
1743 if (net_ratelimit())
1744 printk(KERN_DEBUG
"retrans_out leaked.\n");
1747 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
1748 tp
->retrans_out
+= tcp_skb_pcount(skb
);
1750 /* Save stamp of the first retransmit. */
1751 if (!tp
->retrans_stamp
)
1752 tp
->retrans_stamp
= TCP_SKB_CB(skb
)->when
;
1756 /* snd_nxt is stored to detect loss of retransmitted segment,
1757 * see tcp_input.c tcp_sacktag_write_queue().
1759 TCP_SKB_CB(skb
)->ack_seq
= tp
->snd_nxt
;
1764 /* This gets called after a retransmit timeout, and the initially
1765 * retransmitted data is acknowledged. It tries to continue
1766 * resending the rest of the retransmit queue, until either
1767 * we've sent it all or the congestion window limit is reached.
1768 * If doing SACK, the first ACK which comes back for a timeout
1769 * based retransmit packet might feed us FACK information again.
1770 * If so, we use it to avoid unnecessarily retransmissions.
1772 void tcp_xmit_retransmit_queue(struct sock
*sk
)
1774 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1775 struct tcp_sock
*tp
= tcp_sk(sk
);
1776 struct sk_buff
*skb
;
1779 if (tp
->retransmit_skb_hint
) {
1780 skb
= tp
->retransmit_skb_hint
;
1781 packet_cnt
= tp
->retransmit_cnt_hint
;
1783 skb
= sk
->sk_write_queue
.next
;
1787 /* First pass: retransmit lost packets. */
1789 sk_stream_for_retrans_queue_from(skb
, sk
) {
1790 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
1792 /* we could do better than to assign each time */
1793 tp
->retransmit_skb_hint
= skb
;
1794 tp
->retransmit_cnt_hint
= packet_cnt
;
1796 /* Assume this retransmit will generate
1797 * only one packet for congestion window
1798 * calculation purposes. This works because
1799 * tcp_retransmit_skb() will chop up the
1800 * packet to be MSS sized and all the
1801 * packet counting works out.
1803 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
1806 if (sacked
& TCPCB_LOST
) {
1807 if (!(sacked
&(TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))) {
1808 if (tcp_retransmit_skb(sk
, skb
)) {
1809 tp
->retransmit_skb_hint
= NULL
;
1812 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
1813 NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS
);
1815 NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS
);
1818 skb_peek(&sk
->sk_write_queue
))
1819 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1820 inet_csk(sk
)->icsk_rto
,
1824 packet_cnt
+= tcp_skb_pcount(skb
);
1825 if (packet_cnt
>= tp
->lost_out
)
1831 /* OK, demanded retransmission is finished. */
1833 /* Forward retransmissions are possible only during Recovery. */
1834 if (icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
1837 /* No forward retransmissions in Reno are possible. */
1838 if (!tp
->rx_opt
.sack_ok
)
1841 /* Yeah, we have to make difficult choice between forward transmission
1842 * and retransmission... Both ways have their merits...
1844 * For now we do not retransmit anything, while we have some new
1848 if (tcp_may_send_now(sk
, tp
))
1851 if (tp
->forward_skb_hint
) {
1852 skb
= tp
->forward_skb_hint
;
1853 packet_cnt
= tp
->forward_cnt_hint
;
1855 skb
= sk
->sk_write_queue
.next
;
1859 sk_stream_for_retrans_queue_from(skb
, sk
) {
1860 tp
->forward_cnt_hint
= packet_cnt
;
1861 tp
->forward_skb_hint
= skb
;
1863 /* Similar to the retransmit loop above we
1864 * can pretend that the retransmitted SKB
1865 * we send out here will be composed of one
1866 * real MSS sized packet because tcp_retransmit_skb()
1867 * will fragment it if necessary.
1869 if (++packet_cnt
> tp
->fackets_out
)
1872 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
1875 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
)
1878 /* Ok, retransmit it. */
1879 if (tcp_retransmit_skb(sk
, skb
)) {
1880 tp
->forward_skb_hint
= NULL
;
1884 if (skb
== skb_peek(&sk
->sk_write_queue
))
1885 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1886 inet_csk(sk
)->icsk_rto
,
1889 NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS
);
1894 /* Send a fin. The caller locks the socket for us. This cannot be
1895 * allowed to fail queueing a FIN frame under any circumstances.
1897 void tcp_send_fin(struct sock
*sk
)
1899 struct tcp_sock
*tp
= tcp_sk(sk
);
1900 struct sk_buff
*skb
= skb_peek_tail(&sk
->sk_write_queue
);
1903 /* Optimization, tack on the FIN if we have a queue of
1904 * unsent frames. But be careful about outgoing SACKS
1907 mss_now
= tcp_current_mss(sk
, 1);
1909 if (sk
->sk_send_head
!= NULL
) {
1910 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_FIN
;
1911 TCP_SKB_CB(skb
)->end_seq
++;
1914 /* Socket is locked, keep trying until memory is available. */
1916 skb
= alloc_skb_fclone(MAX_TCP_HEADER
, GFP_KERNEL
);
1922 /* Reserve space for headers and prepare control bits. */
1923 skb_reserve(skb
, MAX_TCP_HEADER
);
1925 TCP_SKB_CB(skb
)->flags
= (TCPCB_FLAG_ACK
| TCPCB_FLAG_FIN
);
1926 TCP_SKB_CB(skb
)->sacked
= 0;
1927 skb_shinfo(skb
)->tso_segs
= 1;
1928 skb_shinfo(skb
)->tso_size
= 0;
1930 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
1931 TCP_SKB_CB(skb
)->seq
= tp
->write_seq
;
1932 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ 1;
1933 tcp_queue_skb(sk
, skb
);
1935 __tcp_push_pending_frames(sk
, tp
, mss_now
, TCP_NAGLE_OFF
);
1938 /* We get here when a process closes a file descriptor (either due to
1939 * an explicit close() or as a byproduct of exit()'ing) and there
1940 * was unread data in the receive queue. This behavior is recommended
1941 * by draft-ietf-tcpimpl-prob-03.txt section 3.10. -DaveM
1943 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
1945 struct tcp_sock
*tp
= tcp_sk(sk
);
1946 struct sk_buff
*skb
;
1948 /* NOTE: No TCP options attached and we never retransmit this. */
1949 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
1951 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED
);
1955 /* Reserve space for headers and prepare control bits. */
1956 skb_reserve(skb
, MAX_TCP_HEADER
);
1958 TCP_SKB_CB(skb
)->flags
= (TCPCB_FLAG_ACK
| TCPCB_FLAG_RST
);
1959 TCP_SKB_CB(skb
)->sacked
= 0;
1960 skb_shinfo(skb
)->tso_segs
= 1;
1961 skb_shinfo(skb
)->tso_size
= 0;
1964 TCP_SKB_CB(skb
)->seq
= tcp_acceptable_seq(sk
, tp
);
1965 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
;
1966 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1967 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
1968 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED
);
1971 /* WARNING: This routine must only be called when we have already sent
1972 * a SYN packet that crossed the incoming SYN that caused this routine
1973 * to get called. If this assumption fails then the initial rcv_wnd
1974 * and rcv_wscale values will not be correct.
1976 int tcp_send_synack(struct sock
*sk
)
1978 struct sk_buff
* skb
;
1980 skb
= skb_peek(&sk
->sk_write_queue
);
1981 if (skb
== NULL
|| !(TCP_SKB_CB(skb
)->flags
&TCPCB_FLAG_SYN
)) {
1982 printk(KERN_DEBUG
"tcp_send_synack: wrong queue state\n");
1985 if (!(TCP_SKB_CB(skb
)->flags
&TCPCB_FLAG_ACK
)) {
1986 if (skb_cloned(skb
)) {
1987 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
1990 __skb_unlink(skb
, &sk
->sk_write_queue
);
1991 skb_header_release(nskb
);
1992 __skb_queue_head(&sk
->sk_write_queue
, nskb
);
1993 sk_stream_free_skb(sk
, skb
);
1994 sk_charge_skb(sk
, nskb
);
1998 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_ACK
;
1999 TCP_ECN_send_synack(tcp_sk(sk
), skb
);
2001 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2002 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2006 * Prepare a SYN-ACK.
2008 struct sk_buff
* tcp_make_synack(struct sock
*sk
, struct dst_entry
*dst
,
2009 struct request_sock
*req
)
2011 struct inet_request_sock
*ireq
= inet_rsk(req
);
2012 struct tcp_sock
*tp
= tcp_sk(sk
);
2014 int tcp_header_size
;
2015 struct sk_buff
*skb
;
2017 skb
= sock_wmalloc(sk
, MAX_TCP_HEADER
+ 15, 1, GFP_ATOMIC
);
2021 /* Reserve space for headers. */
2022 skb_reserve(skb
, MAX_TCP_HEADER
);
2024 skb
->dst
= dst_clone(dst
);
2026 tcp_header_size
= (sizeof(struct tcphdr
) + TCPOLEN_MSS
+
2027 (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0) +
2028 (ireq
->wscale_ok
? TCPOLEN_WSCALE_ALIGNED
: 0) +
2029 /* SACK_PERM is in the place of NOP NOP of TS */
2030 ((ireq
->sack_ok
&& !ireq
->tstamp_ok
) ? TCPOLEN_SACKPERM_ALIGNED
: 0));
2031 skb
->h
.th
= th
= (struct tcphdr
*) skb_push(skb
, tcp_header_size
);
2033 memset(th
, 0, sizeof(struct tcphdr
));
2036 if (dst
->dev
->features
&NETIF_F_TSO
)
2038 TCP_ECN_make_synack(req
, th
);
2039 th
->source
= inet_sk(sk
)->sport
;
2040 th
->dest
= ireq
->rmt_port
;
2041 TCP_SKB_CB(skb
)->seq
= tcp_rsk(req
)->snt_isn
;
2042 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ 1;
2043 TCP_SKB_CB(skb
)->sacked
= 0;
2044 skb_shinfo(skb
)->tso_segs
= 1;
2045 skb_shinfo(skb
)->tso_size
= 0;
2046 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
2047 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_isn
+ 1);
2048 if (req
->rcv_wnd
== 0) { /* ignored for retransmitted syns */
2050 /* Set this up on the first call only */
2051 req
->window_clamp
= tp
->window_clamp
? : dst_metric(dst
, RTAX_WINDOW
);
2052 /* tcp_full_space because it is guaranteed to be the first packet */
2053 tcp_select_initial_window(tcp_full_space(sk
),
2054 dst_metric(dst
, RTAX_ADVMSS
) - (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0),
2059 ireq
->rcv_wscale
= rcv_wscale
;
2062 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2063 th
->window
= htons(req
->rcv_wnd
);
2065 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2066 tcp_syn_build_options((__u32
*)(th
+ 1), dst_metric(dst
, RTAX_ADVMSS
), ireq
->tstamp_ok
,
2067 ireq
->sack_ok
, ireq
->wscale_ok
, ireq
->rcv_wscale
,
2068 TCP_SKB_CB(skb
)->when
,
2072 th
->doff
= (tcp_header_size
>> 2);
2073 TCP_INC_STATS(TCP_MIB_OUTSEGS
);
2078 * Do all connect socket setups that can be done AF independent.
2080 static void tcp_connect_init(struct sock
*sk
)
2082 struct dst_entry
*dst
= __sk_dst_get(sk
);
2083 struct tcp_sock
*tp
= tcp_sk(sk
);
2086 /* We'll fix this up when we get a response from the other end.
2087 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2089 tp
->tcp_header_len
= sizeof(struct tcphdr
) +
2090 (sysctl_tcp_timestamps
? TCPOLEN_TSTAMP_ALIGNED
: 0);
2092 /* If user gave his TCP_MAXSEG, record it to clamp */
2093 if (tp
->rx_opt
.user_mss
)
2094 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
2097 tcp_sync_mss(sk
, dst_mtu(dst
));
2099 if (!tp
->window_clamp
)
2100 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
2101 tp
->advmss
= dst_metric(dst
, RTAX_ADVMSS
);
2102 tcp_initialize_rcv_mss(sk
);
2104 tcp_select_initial_window(tcp_full_space(sk
),
2105 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
2108 sysctl_tcp_window_scaling
,
2111 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
2112 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
2115 sock_reset_flag(sk
, SOCK_DONE
);
2117 tcp_init_wl(tp
, tp
->write_seq
, 0);
2118 tp
->snd_una
= tp
->write_seq
;
2119 tp
->snd_sml
= tp
->write_seq
;
2124 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
2125 inet_csk(sk
)->icsk_retransmits
= 0;
2126 tcp_clear_retrans(tp
);
2130 * Build a SYN and send it off.
2132 int tcp_connect(struct sock
*sk
)
2134 struct tcp_sock
*tp
= tcp_sk(sk
);
2135 struct sk_buff
*buff
;
2137 tcp_connect_init(sk
);
2139 buff
= alloc_skb_fclone(MAX_TCP_HEADER
+ 15, sk
->sk_allocation
);
2140 if (unlikely(buff
== NULL
))
2143 /* Reserve space for headers. */
2144 skb_reserve(buff
, MAX_TCP_HEADER
);
2146 TCP_SKB_CB(buff
)->flags
= TCPCB_FLAG_SYN
;
2147 TCP_ECN_send_syn(sk
, tp
, buff
);
2148 TCP_SKB_CB(buff
)->sacked
= 0;
2149 skb_shinfo(buff
)->tso_segs
= 1;
2150 skb_shinfo(buff
)->tso_size
= 0;
2152 TCP_SKB_CB(buff
)->seq
= tp
->write_seq
++;
2153 TCP_SKB_CB(buff
)->end_seq
= tp
->write_seq
;
2154 tp
->snd_nxt
= tp
->write_seq
;
2155 tp
->pushed_seq
= tp
->write_seq
;
2158 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2159 tp
->retrans_stamp
= TCP_SKB_CB(buff
)->when
;
2160 skb_header_release(buff
);
2161 __skb_queue_tail(&sk
->sk_write_queue
, buff
);
2162 sk_charge_skb(sk
, buff
);
2163 tp
->packets_out
+= tcp_skb_pcount(buff
);
2164 tcp_transmit_skb(sk
, buff
, 1, GFP_KERNEL
);
2165 TCP_INC_STATS(TCP_MIB_ACTIVEOPENS
);
2167 /* Timer for repeating the SYN until an answer. */
2168 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2169 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2173 /* Send out a delayed ack, the caller does the policy checking
2174 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2177 void tcp_send_delayed_ack(struct sock
*sk
)
2179 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2180 int ato
= icsk
->icsk_ack
.ato
;
2181 unsigned long timeout
;
2183 if (ato
> TCP_DELACK_MIN
) {
2184 const struct tcp_sock
*tp
= tcp_sk(sk
);
2187 if (icsk
->icsk_ack
.pingpong
|| (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
2188 max_ato
= TCP_DELACK_MAX
;
2190 /* Slow path, intersegment interval is "high". */
2192 /* If some rtt estimate is known, use it to bound delayed ack.
2193 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2197 int rtt
= max(tp
->srtt
>>3, TCP_DELACK_MIN
);
2203 ato
= min(ato
, max_ato
);
2206 /* Stay within the limit we were given */
2207 timeout
= jiffies
+ ato
;
2209 /* Use new timeout only if there wasn't a older one earlier. */
2210 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
2211 /* If delack timer was blocked or is about to expire,
2214 if (icsk
->icsk_ack
.blocked
||
2215 time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
2220 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
2221 timeout
= icsk
->icsk_ack
.timeout
;
2223 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
2224 icsk
->icsk_ack
.timeout
= timeout
;
2225 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
2228 /* This routine sends an ack and also updates the window. */
2229 void tcp_send_ack(struct sock
*sk
)
2231 /* If we have been reset, we may not send again. */
2232 if (sk
->sk_state
!= TCP_CLOSE
) {
2233 struct tcp_sock
*tp
= tcp_sk(sk
);
2234 struct sk_buff
*buff
;
2236 /* We are not putting this on the write queue, so
2237 * tcp_transmit_skb() will set the ownership to this
2240 buff
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2242 inet_csk_schedule_ack(sk
);
2243 inet_csk(sk
)->icsk_ack
.ato
= TCP_ATO_MIN
;
2244 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
2245 TCP_DELACK_MAX
, TCP_RTO_MAX
);
2249 /* Reserve space for headers and prepare control bits. */
2250 skb_reserve(buff
, MAX_TCP_HEADER
);
2252 TCP_SKB_CB(buff
)->flags
= TCPCB_FLAG_ACK
;
2253 TCP_SKB_CB(buff
)->sacked
= 0;
2254 skb_shinfo(buff
)->tso_segs
= 1;
2255 skb_shinfo(buff
)->tso_size
= 0;
2257 /* Send it off, this clears delayed acks for us. */
2258 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(buff
)->end_seq
= tcp_acceptable_seq(sk
, tp
);
2259 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2260 tcp_transmit_skb(sk
, buff
, 0, GFP_ATOMIC
);
2264 /* This routine sends a packet with an out of date sequence
2265 * number. It assumes the other end will try to ack it.
2267 * Question: what should we make while urgent mode?
2268 * 4.4BSD forces sending single byte of data. We cannot send
2269 * out of window data, because we have SND.NXT==SND.MAX...
2271 * Current solution: to send TWO zero-length segments in urgent mode:
2272 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2273 * out-of-date with SND.UNA-1 to probe window.
2275 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
)
2277 struct tcp_sock
*tp
= tcp_sk(sk
);
2278 struct sk_buff
*skb
;
2280 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2281 skb
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2285 /* Reserve space for headers and set control bits. */
2286 skb_reserve(skb
, MAX_TCP_HEADER
);
2288 TCP_SKB_CB(skb
)->flags
= TCPCB_FLAG_ACK
;
2289 TCP_SKB_CB(skb
)->sacked
= urgent
;
2290 skb_shinfo(skb
)->tso_segs
= 1;
2291 skb_shinfo(skb
)->tso_size
= 0;
2293 /* Use a previous sequence. This should cause the other
2294 * end to send an ack. Don't queue or clone SKB, just
2297 TCP_SKB_CB(skb
)->seq
= urgent
? tp
->snd_una
: tp
->snd_una
- 1;
2298 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
;
2299 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2300 return tcp_transmit_skb(sk
, skb
, 0, GFP_ATOMIC
);
2303 int tcp_write_wakeup(struct sock
*sk
)
2305 if (sk
->sk_state
!= TCP_CLOSE
) {
2306 struct tcp_sock
*tp
= tcp_sk(sk
);
2307 struct sk_buff
*skb
;
2309 if ((skb
= sk
->sk_send_head
) != NULL
&&
2310 before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
+tp
->snd_wnd
)) {
2312 unsigned int mss
= tcp_current_mss(sk
, 0);
2313 unsigned int seg_size
= tp
->snd_una
+tp
->snd_wnd
-TCP_SKB_CB(skb
)->seq
;
2315 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
2316 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
2318 /* We are probing the opening of a window
2319 * but the window size is != 0
2320 * must have been a result SWS avoidance ( sender )
2322 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
2324 seg_size
= min(seg_size
, mss
);
2325 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2326 if (tcp_fragment(sk
, skb
, seg_size
, mss
))
2328 } else if (!tcp_skb_pcount(skb
))
2329 tcp_set_skb_tso_segs(sk
, skb
, mss
);
2331 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2332 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2333 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2335 update_send_head(sk
, tp
, skb
);
2340 between(tp
->snd_up
, tp
->snd_una
+1, tp
->snd_una
+0xFFFF))
2341 tcp_xmit_probe_skb(sk
, TCPCB_URG
);
2342 return tcp_xmit_probe_skb(sk
, 0);
2348 /* A window probe timeout has occurred. If window is not closed send
2349 * a partial packet else a zero probe.
2351 void tcp_send_probe0(struct sock
*sk
)
2353 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2354 struct tcp_sock
*tp
= tcp_sk(sk
);
2357 err
= tcp_write_wakeup(sk
);
2359 if (tp
->packets_out
|| !sk
->sk_send_head
) {
2360 /* Cancel probe timer, if it is not required. */
2361 icsk
->icsk_probes_out
= 0;
2362 icsk
->icsk_backoff
= 0;
2367 if (icsk
->icsk_backoff
< sysctl_tcp_retries2
)
2368 icsk
->icsk_backoff
++;
2369 icsk
->icsk_probes_out
++;
2370 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2371 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
2374 /* If packet was not sent due to local congestion,
2375 * do not backoff and do not remember icsk_probes_out.
2376 * Let local senders to fight for local resources.
2378 * Use accumulated backoff yet.
2380 if (!icsk
->icsk_probes_out
)
2381 icsk
->icsk_probes_out
= 1;
2382 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2383 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
,
2384 TCP_RESOURCE_PROBE_INTERVAL
),
2389 EXPORT_SYMBOL(tcp_connect
);
2390 EXPORT_SYMBOL(tcp_make_synack
);
2391 EXPORT_SYMBOL(tcp_simple_retransmit
);
2392 EXPORT_SYMBOL(tcp_sync_mss
);
2393 EXPORT_SYMBOL(sysctl_tcp_tso_win_divisor
);
2394 EXPORT_SYMBOL(tcp_mtup_init
);