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
= 8;
54 static inline void update_send_head(struct sock
*sk
, struct tcp_sock
*tp
,
57 sk
->sk_send_head
= skb
->next
;
58 if (sk
->sk_send_head
== (struct sk_buff
*)&sk
->sk_write_queue
)
59 sk
->sk_send_head
= NULL
;
60 tp
->snd_nxt
= TCP_SKB_CB(skb
)->end_seq
;
61 tcp_packets_out_inc(sk
, tp
, skb
);
64 /* SND.NXT, if window was not shrunk.
65 * If window has been shrunk, what should we make? It is not clear at all.
66 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
67 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
68 * invalid. OK, let's make this for now:
70 static inline __u32
tcp_acceptable_seq(struct sock
*sk
, struct tcp_sock
*tp
)
72 if (!before(tp
->snd_una
+tp
->snd_wnd
, tp
->snd_nxt
))
75 return tp
->snd_una
+tp
->snd_wnd
;
78 /* Calculate mss to advertise in SYN segment.
79 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
81 * 1. It is independent of path mtu.
82 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
83 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
84 * attached devices, because some buggy hosts are confused by
86 * 4. We do not make 3, we advertise MSS, calculated from first
87 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
88 * This may be overridden via information stored in routing table.
89 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
90 * probably even Jumbo".
92 static __u16
tcp_advertise_mss(struct sock
*sk
)
94 struct tcp_sock
*tp
= tcp_sk(sk
);
95 struct dst_entry
*dst
= __sk_dst_get(sk
);
98 if (dst
&& dst_metric(dst
, RTAX_ADVMSS
) < mss
) {
99 mss
= dst_metric(dst
, RTAX_ADVMSS
);
106 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
107 * This is the first part of cwnd validation mechanism. */
108 static void tcp_cwnd_restart(struct tcp_sock
*tp
, struct dst_entry
*dst
)
110 s32 delta
= tcp_time_stamp
- tp
->lsndtime
;
111 u32 restart_cwnd
= tcp_init_cwnd(tp
, dst
);
112 u32 cwnd
= tp
->snd_cwnd
;
114 tcp_ca_event(tp
, CA_EVENT_CWND_RESTART
);
116 tp
->snd_ssthresh
= tcp_current_ssthresh(tp
);
117 restart_cwnd
= min(restart_cwnd
, cwnd
);
119 while ((delta
-= tp
->rto
) > 0 && cwnd
> restart_cwnd
)
121 tp
->snd_cwnd
= max(cwnd
, restart_cwnd
);
122 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
123 tp
->snd_cwnd_used
= 0;
126 static inline void tcp_event_data_sent(struct tcp_sock
*tp
,
127 struct sk_buff
*skb
, struct sock
*sk
)
129 u32 now
= tcp_time_stamp
;
131 if (!tp
->packets_out
&& (s32
)(now
- tp
->lsndtime
) > tp
->rto
)
132 tcp_cwnd_restart(tp
, __sk_dst_get(sk
));
136 /* If it is a reply for ato after last received
137 * packet, enter pingpong mode.
139 if ((u32
)(now
- tp
->ack
.lrcvtime
) < tp
->ack
.ato
)
140 tp
->ack
.pingpong
= 1;
143 static __inline__
void tcp_event_ack_sent(struct sock
*sk
, unsigned int pkts
)
145 struct tcp_sock
*tp
= tcp_sk(sk
);
147 tcp_dec_quickack_mode(tp
, pkts
);
148 tcp_clear_xmit_timer(sk
, TCP_TIME_DACK
);
151 /* Determine a window scaling and initial window to offer.
152 * Based on the assumption that the given amount of space
153 * will be offered. Store the results in the tp structure.
154 * NOTE: for smooth operation initial space offering should
155 * be a multiple of mss if possible. We assume here that mss >= 1.
156 * This MUST be enforced by all callers.
158 void tcp_select_initial_window(int __space
, __u32 mss
,
159 __u32
*rcv_wnd
, __u32
*window_clamp
,
160 int wscale_ok
, __u8
*rcv_wscale
)
162 unsigned int space
= (__space
< 0 ? 0 : __space
);
164 /* If no clamp set the clamp to the max possible scaled window */
165 if (*window_clamp
== 0)
166 (*window_clamp
) = (65535 << 14);
167 space
= min(*window_clamp
, space
);
169 /* Quantize space offering to a multiple of mss if possible. */
171 space
= (space
/ mss
) * mss
;
173 /* NOTE: offering an initial window larger than 32767
174 * will break some buggy TCP stacks. We try to be nice.
175 * If we are not window scaling, then this truncates
176 * our initial window offering to 32k. There should also
177 * be a sysctl option to stop being nice.
179 (*rcv_wnd
) = min(space
, MAX_TCP_WINDOW
);
182 /* Set window scaling on max possible window
183 * See RFC1323 for an explanation of the limit to 14
185 space
= max_t(u32
, sysctl_tcp_rmem
[2], sysctl_rmem_max
);
186 while (space
> 65535 && (*rcv_wscale
) < 14) {
192 /* Set initial window to value enough for senders,
193 * following RFC1414. Senders, not following this RFC,
194 * will be satisfied with 2.
196 if (mss
> (1<<*rcv_wscale
)) {
202 if (*rcv_wnd
> init_cwnd
*mss
)
203 *rcv_wnd
= init_cwnd
*mss
;
206 /* Set the clamp no higher than max representable value */
207 (*window_clamp
) = min(65535U << (*rcv_wscale
), *window_clamp
);
210 /* Chose a new window to advertise, update state in tcp_sock for the
211 * socket, and return result with RFC1323 scaling applied. The return
212 * value can be stuffed directly into th->window for an outgoing
215 static __inline__ u16
tcp_select_window(struct sock
*sk
)
217 struct tcp_sock
*tp
= tcp_sk(sk
);
218 u32 cur_win
= tcp_receive_window(tp
);
219 u32 new_win
= __tcp_select_window(sk
);
221 /* Never shrink the offered window */
222 if(new_win
< cur_win
) {
223 /* Danger Will Robinson!
224 * Don't update rcv_wup/rcv_wnd here or else
225 * we will not be able to advertise a zero
226 * window in time. --DaveM
228 * Relax Will Robinson.
232 tp
->rcv_wnd
= new_win
;
233 tp
->rcv_wup
= tp
->rcv_nxt
;
235 /* Make sure we do not exceed the maximum possible
238 if (!tp
->rx_opt
.rcv_wscale
)
239 new_win
= min(new_win
, MAX_TCP_WINDOW
);
241 new_win
= min(new_win
, (65535U << tp
->rx_opt
.rcv_wscale
));
243 /* RFC1323 scaling applied */
244 new_win
>>= tp
->rx_opt
.rcv_wscale
;
246 /* If we advertise zero window, disable fast path. */
254 /* This routine actually transmits TCP packets queued in by
255 * tcp_do_sendmsg(). This is used by both the initial
256 * transmission and possible later retransmissions.
257 * All SKB's seen here are completely headerless. It is our
258 * job to build the TCP header, and pass the packet down to
259 * IP so it can do the same plus pass the packet off to the
262 * We are working here with either a clone of the original
263 * SKB, or a fresh unique copy made by the retransmit engine.
265 static int tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
268 struct inet_sock
*inet
= inet_sk(sk
);
269 struct tcp_sock
*tp
= tcp_sk(sk
);
270 struct tcp_skb_cb
*tcb
= TCP_SKB_CB(skb
);
271 int tcp_header_size
= tp
->tcp_header_len
;
276 BUG_ON(!tcp_skb_pcount(skb
));
278 #define SYSCTL_FLAG_TSTAMPS 0x1
279 #define SYSCTL_FLAG_WSCALE 0x2
280 #define SYSCTL_FLAG_SACK 0x4
282 /* If congestion control is doing timestamping */
283 if (tp
->ca_ops
->rtt_sample
)
284 do_gettimeofday(&skb
->stamp
);
287 if (tcb
->flags
& TCPCB_FLAG_SYN
) {
288 tcp_header_size
= sizeof(struct tcphdr
) + TCPOLEN_MSS
;
289 if(sysctl_tcp_timestamps
) {
290 tcp_header_size
+= TCPOLEN_TSTAMP_ALIGNED
;
291 sysctl_flags
|= SYSCTL_FLAG_TSTAMPS
;
293 if(sysctl_tcp_window_scaling
) {
294 tcp_header_size
+= TCPOLEN_WSCALE_ALIGNED
;
295 sysctl_flags
|= SYSCTL_FLAG_WSCALE
;
297 if(sysctl_tcp_sack
) {
298 sysctl_flags
|= SYSCTL_FLAG_SACK
;
299 if(!(sysctl_flags
& SYSCTL_FLAG_TSTAMPS
))
300 tcp_header_size
+= TCPOLEN_SACKPERM_ALIGNED
;
302 } else if (tp
->rx_opt
.eff_sacks
) {
303 /* A SACK is 2 pad bytes, a 2 byte header, plus
304 * 2 32-bit sequence numbers for each SACK block.
306 tcp_header_size
+= (TCPOLEN_SACK_BASE_ALIGNED
+
307 (tp
->rx_opt
.eff_sacks
* TCPOLEN_SACK_PERBLOCK
));
310 if (tcp_packets_in_flight(tp
) == 0)
311 tcp_ca_event(tp
, CA_EVENT_TX_START
);
313 th
= (struct tcphdr
*) skb_push(skb
, tcp_header_size
);
315 skb_set_owner_w(skb
, sk
);
317 /* Build TCP header and checksum it. */
318 th
->source
= inet
->sport
;
319 th
->dest
= inet
->dport
;
320 th
->seq
= htonl(tcb
->seq
);
321 th
->ack_seq
= htonl(tp
->rcv_nxt
);
322 *(((__u16
*)th
) + 6) = htons(((tcp_header_size
>> 2) << 12) | tcb
->flags
);
323 if (tcb
->flags
& TCPCB_FLAG_SYN
) {
324 /* RFC1323: The window in SYN & SYN/ACK segments
327 th
->window
= htons(tp
->rcv_wnd
);
329 th
->window
= htons(tcp_select_window(sk
));
335 between(tp
->snd_up
, tcb
->seq
+1, tcb
->seq
+0xFFFF)) {
336 th
->urg_ptr
= htons(tp
->snd_up
-tcb
->seq
);
340 if (tcb
->flags
& TCPCB_FLAG_SYN
) {
341 tcp_syn_build_options((__u32
*)(th
+ 1),
342 tcp_advertise_mss(sk
),
343 (sysctl_flags
& SYSCTL_FLAG_TSTAMPS
),
344 (sysctl_flags
& SYSCTL_FLAG_SACK
),
345 (sysctl_flags
& SYSCTL_FLAG_WSCALE
),
346 tp
->rx_opt
.rcv_wscale
,
348 tp
->rx_opt
.ts_recent
);
350 tcp_build_and_update_options((__u32
*)(th
+ 1),
353 TCP_ECN_send(sk
, tp
, skb
, tcp_header_size
);
355 tp
->af_specific
->send_check(sk
, th
, skb
->len
, skb
);
357 if (tcb
->flags
& TCPCB_FLAG_ACK
)
358 tcp_event_ack_sent(sk
, tcp_skb_pcount(skb
));
360 if (skb
->len
!= tcp_header_size
)
361 tcp_event_data_sent(tp
, skb
, sk
);
363 TCP_INC_STATS(TCP_MIB_OUTSEGS
);
365 err
= tp
->af_specific
->queue_xmit(skb
, 0);
371 /* NET_XMIT_CN is special. It does not guarantee,
372 * that this packet is lost. It tells that device
373 * is about to start to drop packets or already
374 * drops some packets of the same priority and
375 * invokes us to send less aggressively.
377 return err
== NET_XMIT_CN
? 0 : err
;
380 #undef SYSCTL_FLAG_TSTAMPS
381 #undef SYSCTL_FLAG_WSCALE
382 #undef SYSCTL_FLAG_SACK
386 /* This routine just queue's the buffer
388 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
389 * otherwise socket can stall.
391 static void tcp_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
393 struct tcp_sock
*tp
= tcp_sk(sk
);
395 /* Advance write_seq and place onto the write_queue. */
396 tp
->write_seq
= TCP_SKB_CB(skb
)->end_seq
;
397 skb_header_release(skb
);
398 __skb_queue_tail(&sk
->sk_write_queue
, skb
);
399 sk_charge_skb(sk
, skb
);
401 /* Queue it, remembering where we must start sending. */
402 if (sk
->sk_send_head
== NULL
)
403 sk
->sk_send_head
= skb
;
406 static inline void tcp_tso_set_push(struct sk_buff
*skb
)
408 /* Force push to be on for any TSO frames to workaround
409 * problems with busted implementations like Mac OS-X that
410 * hold off socket receive wakeups until push is seen.
412 if (tcp_skb_pcount(skb
) > 1)
413 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
416 static void tcp_set_skb_tso_segs(struct sock
*sk
, struct sk_buff
*skb
)
418 struct tcp_sock
*tp
= tcp_sk(sk
);
420 if (skb
->len
<= tp
->mss_cache_std
||
421 !(sk
->sk_route_caps
& NETIF_F_TSO
)) {
422 /* Avoid the costly divide in the normal
425 skb_shinfo(skb
)->tso_segs
= 1;
426 skb_shinfo(skb
)->tso_size
= 0;
430 factor
= skb
->len
+ (tp
->mss_cache_std
- 1);
431 factor
/= tp
->mss_cache_std
;
432 skb_shinfo(skb
)->tso_segs
= factor
;
433 skb_shinfo(skb
)->tso_size
= tp
->mss_cache_std
;
437 static inline int tcp_minshall_check(const struct tcp_sock
*tp
)
439 return after(tp
->snd_sml
,tp
->snd_una
) &&
440 !after(tp
->snd_sml
, tp
->snd_nxt
);
443 /* Return 0, if packet can be sent now without violation Nagle's rules:
444 * 1. It is full sized.
445 * 2. Or it contains FIN.
446 * 3. Or TCP_NODELAY was set.
447 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
448 * With Minshall's modification: all sent small packets are ACKed.
451 static inline int tcp_nagle_check(const struct tcp_sock
*tp
,
452 const struct sk_buff
*skb
,
453 unsigned mss_now
, int nonagle
)
455 return (skb
->len
< mss_now
&&
456 !(TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
) &&
457 ((nonagle
&TCP_NAGLE_CORK
) ||
460 tcp_minshall_check(tp
))));
463 /* This checks if the data bearing packet SKB (usually sk->sk_send_head)
464 * should be put on the wire right now.
466 static int tcp_snd_test(struct sock
*sk
, struct sk_buff
*skb
,
467 unsigned cur_mss
, int nonagle
)
469 struct tcp_sock
*tp
= tcp_sk(sk
);
470 int pkts
= tcp_skb_pcount(skb
);
473 tcp_set_skb_tso_segs(sk
, skb
);
474 pkts
= tcp_skb_pcount(skb
);
477 /* RFC 1122 - section 4.2.3.4
481 * a) The right edge of this frame exceeds the window
482 * b) There are packets in flight and we have a small segment
483 * [SWS avoidance and Nagle algorithm]
484 * (part of SWS is done on packetization)
485 * Minshall version sounds: there are no _small_
486 * segments in flight. (tcp_nagle_check)
487 * c) We have too many packets 'in flight'
489 * Don't use the nagle rule for urgent data (or
490 * for the final FIN -DaveM).
492 * Also, Nagle rule does not apply to frames, which
493 * sit in the middle of queue (they have no chances
494 * to get new data) and if room at tail of skb is
495 * not enough to save something seriously (<32 for now).
498 /* Don't be strict about the congestion window for the
499 * final FIN frame. -DaveM
501 return (((nonagle
&TCP_NAGLE_PUSH
) || tp
->urg_mode
502 || !tcp_nagle_check(tp
, skb
, cur_mss
, nonagle
)) &&
503 (((tcp_packets_in_flight(tp
) + (pkts
-1)) < tp
->snd_cwnd
) ||
504 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
)) &&
505 !after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
+ tp
->snd_wnd
));
508 static inline int tcp_skb_is_last(const struct sock
*sk
,
509 const struct sk_buff
*skb
)
511 return skb
->next
== (struct sk_buff
*)&sk
->sk_write_queue
;
514 /* Push out any pending frames which were held back due to
515 * TCP_CORK or attempt at coalescing tiny packets.
516 * The socket must be locked by the caller.
518 void __tcp_push_pending_frames(struct sock
*sk
, struct tcp_sock
*tp
,
519 unsigned cur_mss
, int nonagle
)
521 struct sk_buff
*skb
= sk
->sk_send_head
;
524 if (!tcp_skb_is_last(sk
, skb
))
525 nonagle
= TCP_NAGLE_PUSH
;
526 if (!tcp_snd_test(sk
, skb
, cur_mss
, nonagle
) ||
527 tcp_write_xmit(sk
, nonagle
))
528 tcp_check_probe_timer(sk
, tp
);
530 tcp_cwnd_validate(sk
, tp
);
533 int tcp_may_send_now(struct sock
*sk
, struct tcp_sock
*tp
)
535 struct sk_buff
*skb
= sk
->sk_send_head
;
538 tcp_snd_test(sk
, skb
, tcp_current_mss(sk
, 1),
539 (tcp_skb_is_last(sk
, skb
) ?
545 /* Send _single_ skb sitting at the send head. This function requires
546 * true push pending frames to setup probe timer etc.
548 void tcp_push_one(struct sock
*sk
, unsigned cur_mss
)
550 struct tcp_sock
*tp
= tcp_sk(sk
);
551 struct sk_buff
*skb
= sk
->sk_send_head
;
553 if (tcp_snd_test(sk
, skb
, cur_mss
, TCP_NAGLE_PUSH
)) {
554 /* Send it out now. */
555 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
556 tcp_tso_set_push(skb
);
557 if (!tcp_transmit_skb(sk
, skb_clone(skb
, sk
->sk_allocation
))) {
558 sk
->sk_send_head
= NULL
;
559 tp
->snd_nxt
= TCP_SKB_CB(skb
)->end_seq
;
560 tcp_packets_out_inc(sk
, tp
, skb
);
566 /* Function to create two new TCP segments. Shrinks the given segment
567 * to the specified size and appends a new segment with the rest of the
568 * packet to the list. This won't be called frequently, I hope.
569 * Remember, these are still headerless SKBs at this point.
571 static int tcp_fragment(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
573 struct tcp_sock
*tp
= tcp_sk(sk
);
574 struct sk_buff
*buff
;
578 nsize
= skb_headlen(skb
) - len
;
582 if (skb_cloned(skb
) &&
583 skb_is_nonlinear(skb
) &&
584 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
587 /* Get a new skb... force flag on. */
588 buff
= sk_stream_alloc_skb(sk
, nsize
, GFP_ATOMIC
);
590 return -ENOMEM
; /* We'll just try again later. */
591 sk_charge_skb(sk
, buff
);
593 /* Correct the sequence numbers. */
594 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
595 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
596 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
598 /* PSH and FIN should only be set in the second packet. */
599 flags
= TCP_SKB_CB(skb
)->flags
;
600 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
601 TCP_SKB_CB(buff
)->flags
= flags
;
602 TCP_SKB_CB(buff
)->sacked
=
603 (TCP_SKB_CB(skb
)->sacked
&
604 (TCPCB_LOST
| TCPCB_EVER_RETRANS
| TCPCB_AT_TAIL
));
605 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_AT_TAIL
;
607 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_HW
) {
608 /* Copy and checksum data tail into the new buffer. */
609 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
, skb_put(buff
, nsize
),
614 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
616 skb
->ip_summed
= CHECKSUM_HW
;
617 skb_split(skb
, buff
, len
);
620 buff
->ip_summed
= skb
->ip_summed
;
622 /* Looks stupid, but our code really uses when of
623 * skbs, which it never sent before. --ANK
625 TCP_SKB_CB(buff
)->when
= TCP_SKB_CB(skb
)->when
;
626 buff
->stamp
= skb
->stamp
;
628 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
) {
629 tp
->lost_out
-= tcp_skb_pcount(skb
);
630 tp
->left_out
-= tcp_skb_pcount(skb
);
633 /* Fix up tso_factor for both original and new SKB. */
634 tcp_set_skb_tso_segs(sk
, skb
);
635 tcp_set_skb_tso_segs(sk
, buff
);
637 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
) {
638 tp
->lost_out
+= tcp_skb_pcount(skb
);
639 tp
->left_out
+= tcp_skb_pcount(skb
);
642 if (TCP_SKB_CB(buff
)->sacked
&TCPCB_LOST
) {
643 tp
->lost_out
+= tcp_skb_pcount(buff
);
644 tp
->left_out
+= tcp_skb_pcount(buff
);
647 /* Link BUFF into the send queue. */
648 __skb_append(skb
, buff
);
653 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
654 * eventually). The difference is that pulled data not copied, but
655 * immediately discarded.
657 static unsigned char *__pskb_trim_head(struct sk_buff
*skb
, int len
)
663 for (i
=0; i
<skb_shinfo(skb
)->nr_frags
; i
++) {
664 if (skb_shinfo(skb
)->frags
[i
].size
<= eat
) {
665 put_page(skb_shinfo(skb
)->frags
[i
].page
);
666 eat
-= skb_shinfo(skb
)->frags
[i
].size
;
668 skb_shinfo(skb
)->frags
[k
] = skb_shinfo(skb
)->frags
[i
];
670 skb_shinfo(skb
)->frags
[k
].page_offset
+= eat
;
671 skb_shinfo(skb
)->frags
[k
].size
-= eat
;
677 skb_shinfo(skb
)->nr_frags
= k
;
679 skb
->tail
= skb
->data
;
680 skb
->data_len
-= len
;
681 skb
->len
= skb
->data_len
;
685 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
687 if (skb_cloned(skb
) &&
688 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
691 if (len
<= skb_headlen(skb
)) {
692 __skb_pull(skb
, len
);
694 if (__pskb_trim_head(skb
, len
-skb_headlen(skb
)) == NULL
)
698 TCP_SKB_CB(skb
)->seq
+= len
;
699 skb
->ip_summed
= CHECKSUM_HW
;
701 skb
->truesize
-= len
;
702 sk
->sk_wmem_queued
-= len
;
703 sk
->sk_forward_alloc
+= len
;
704 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
706 /* Any change of skb->len requires recalculation of tso
709 if (tcp_skb_pcount(skb
) > 1)
710 tcp_set_skb_tso_segs(sk
, skb
);
715 /* This function synchronize snd mss to current pmtu/exthdr set.
717 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
718 for TCP options, but includes only bare TCP header.
720 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
721 It is minumum of user_mss and mss received with SYN.
722 It also does not include TCP options.
724 tp->pmtu_cookie is last pmtu, seen by this function.
726 tp->mss_cache is current effective sending mss, including
727 all tcp options except for SACKs. It is evaluated,
728 taking into account current pmtu, but never exceeds
729 tp->rx_opt.mss_clamp.
731 NOTE1. rfc1122 clearly states that advertised MSS
732 DOES NOT include either tcp or ip options.
734 NOTE2. tp->pmtu_cookie and tp->mss_cache are READ ONLY outside
735 this function. --ANK (980731)
738 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
740 struct tcp_sock
*tp
= tcp_sk(sk
);
743 /* Calculate base mss without TCP options:
744 It is MMS_S - sizeof(tcphdr) of rfc1122
746 mss_now
= pmtu
- tp
->af_specific
->net_header_len
- sizeof(struct tcphdr
);
748 /* Clamp it (mss_clamp does not include tcp options) */
749 if (mss_now
> tp
->rx_opt
.mss_clamp
)
750 mss_now
= tp
->rx_opt
.mss_clamp
;
752 /* Now subtract optional transport overhead */
753 mss_now
-= tp
->ext_header_len
;
755 /* Then reserve room for full set of TCP options and 8 bytes of data */
759 /* Now subtract TCP options size, not including SACKs */
760 mss_now
-= tp
->tcp_header_len
- sizeof(struct tcphdr
);
762 /* Bound mss with half of window */
763 if (tp
->max_window
&& mss_now
> (tp
->max_window
>>1))
764 mss_now
= max((tp
->max_window
>>1), 68U - tp
->tcp_header_len
);
766 /* And store cached results */
767 tp
->pmtu_cookie
= pmtu
;
768 tp
->mss_cache
= tp
->mss_cache_std
= mss_now
;
773 /* Compute the current effective MSS, taking SACKs and IP options,
774 * and even PMTU discovery events into account.
776 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
777 * cannot be large. However, taking into account rare use of URG, this
781 unsigned int tcp_current_mss(struct sock
*sk
, int large
)
783 struct tcp_sock
*tp
= tcp_sk(sk
);
784 struct dst_entry
*dst
= __sk_dst_get(sk
);
785 unsigned int do_large
, mss_now
;
787 mss_now
= tp
->mss_cache_std
;
789 u32 mtu
= dst_mtu(dst
);
790 if (mtu
!= tp
->pmtu_cookie
)
791 mss_now
= tcp_sync_mss(sk
, mtu
);
795 (sk
->sk_route_caps
& NETIF_F_TSO
) &&
799 unsigned int large_mss
, factor
, limit
;
801 large_mss
= 65535 - tp
->af_specific
->net_header_len
-
802 tp
->ext_header_len
- tp
->tcp_header_len
;
804 if (tp
->max_window
&& large_mss
> (tp
->max_window
>>1))
805 large_mss
= max((tp
->max_window
>>1),
806 68U - tp
->tcp_header_len
);
808 factor
= large_mss
/ mss_now
;
810 /* Always keep large mss multiple of real mss, but
811 * do not exceed 1/tso_win_divisor of the congestion window
812 * so we can keep the ACK clock ticking and minimize
815 limit
= tp
->snd_cwnd
;
816 if (sysctl_tcp_tso_win_divisor
)
817 limit
/= sysctl_tcp_tso_win_divisor
;
818 limit
= max(1U, limit
);
822 tp
->mss_cache
= mss_now
* factor
;
824 mss_now
= tp
->mss_cache
;
827 if (tp
->rx_opt
.eff_sacks
)
828 mss_now
-= (TCPOLEN_SACK_BASE_ALIGNED
+
829 (tp
->rx_opt
.eff_sacks
* TCPOLEN_SACK_PERBLOCK
));
833 /* This routine writes packets to the network. It advances the
834 * send_head. This happens as incoming acks open up the remote
837 * Returns 1, if no segments are in flight and we have queued segments, but
838 * cannot send anything now because of SWS or another problem.
840 int tcp_write_xmit(struct sock
*sk
, int nonagle
)
842 struct tcp_sock
*tp
= tcp_sk(sk
);
843 unsigned int mss_now
;
845 /* If we are closed, the bytes will have to remain here.
846 * In time closedown will finish, we empty the write queue and all
849 if (sk
->sk_state
!= TCP_CLOSE
) {
853 /* Account for SACKS, we may need to fragment due to this.
854 * It is just like the real MSS changing on us midstream.
855 * We also handle things correctly when the user adds some
856 * IP options mid-stream. Silly to do, but cover it.
858 mss_now
= tcp_current_mss(sk
, 1);
860 while ((skb
= sk
->sk_send_head
) &&
861 tcp_snd_test(sk
, skb
, mss_now
,
862 tcp_skb_is_last(sk
, skb
) ? nonagle
:
864 if (skb
->len
> mss_now
) {
865 if (tcp_fragment(sk
, skb
, mss_now
))
869 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
870 tcp_tso_set_push(skb
);
871 if (tcp_transmit_skb(sk
, skb_clone(skb
, GFP_ATOMIC
)))
874 /* Advance the send_head. This one is sent out.
875 * This call will increment packets_out.
877 update_send_head(sk
, tp
, skb
);
879 tcp_minshall_update(tp
, mss_now
, skb
);
884 tcp_cwnd_validate(sk
, tp
);
888 return !tp
->packets_out
&& sk
->sk_send_head
;
893 /* This function returns the amount that we can raise the
894 * usable window based on the following constraints
896 * 1. The window can never be shrunk once it is offered (RFC 793)
897 * 2. We limit memory per socket
900 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
901 * RECV.NEXT + RCV.WIN fixed until:
902 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
904 * i.e. don't raise the right edge of the window until you can raise
905 * it at least MSS bytes.
907 * Unfortunately, the recommended algorithm breaks header prediction,
908 * since header prediction assumes th->window stays fixed.
910 * Strictly speaking, keeping th->window fixed violates the receiver
911 * side SWS prevention criteria. The problem is that under this rule
912 * a stream of single byte packets will cause the right side of the
913 * window to always advance by a single byte.
915 * Of course, if the sender implements sender side SWS prevention
916 * then this will not be a problem.
918 * BSD seems to make the following compromise:
920 * If the free space is less than the 1/4 of the maximum
921 * space available and the free space is less than 1/2 mss,
922 * then set the window to 0.
923 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
924 * Otherwise, just prevent the window from shrinking
925 * and from being larger than the largest representable value.
927 * This prevents incremental opening of the window in the regime
928 * where TCP is limited by the speed of the reader side taking
929 * data out of the TCP receive queue. It does nothing about
930 * those cases where the window is constrained on the sender side
931 * because the pipeline is full.
933 * BSD also seems to "accidentally" limit itself to windows that are a
934 * multiple of MSS, at least until the free space gets quite small.
935 * This would appear to be a side effect of the mbuf implementation.
936 * Combining these two algorithms results in the observed behavior
937 * of having a fixed window size at almost all times.
939 * Below we obtain similar behavior by forcing the offered window to
940 * a multiple of the mss when it is feasible to do so.
942 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
943 * Regular options like TIMESTAMP are taken into account.
945 u32
__tcp_select_window(struct sock
*sk
)
947 struct tcp_sock
*tp
= tcp_sk(sk
);
948 /* MSS for the peer's data. Previous verions used mss_clamp
949 * here. I don't know if the value based on our guesses
950 * of peer's MSS is better for the performance. It's more correct
951 * but may be worse for the performance because of rcv_mss
952 * fluctuations. --SAW 1998/11/1
954 int mss
= tp
->ack
.rcv_mss
;
955 int free_space
= tcp_space(sk
);
956 int full_space
= min_t(int, tp
->window_clamp
, tcp_full_space(sk
));
959 if (mss
> full_space
)
962 if (free_space
< full_space
/2) {
965 if (tcp_memory_pressure
)
966 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U*tp
->advmss
);
968 if (free_space
< mss
)
972 if (free_space
> tp
->rcv_ssthresh
)
973 free_space
= tp
->rcv_ssthresh
;
975 /* Don't do rounding if we are using window scaling, since the
976 * scaled window will not line up with the MSS boundary anyway.
978 window
= tp
->rcv_wnd
;
979 if (tp
->rx_opt
.rcv_wscale
) {
982 /* Advertise enough space so that it won't get scaled away.
983 * Import case: prevent zero window announcement if
984 * 1<<rcv_wscale > mss.
986 if (((window
>> tp
->rx_opt
.rcv_wscale
) << tp
->rx_opt
.rcv_wscale
) != window
)
987 window
= (((window
>> tp
->rx_opt
.rcv_wscale
) + 1)
988 << tp
->rx_opt
.rcv_wscale
);
990 /* Get the largest window that is a nice multiple of mss.
991 * Window clamp already applied above.
992 * If our current window offering is within 1 mss of the
993 * free space we just keep it. This prevents the divide
994 * and multiply from happening most of the time.
995 * We also don't do any window rounding when the free space
998 if (window
<= free_space
- mss
|| window
> free_space
)
999 window
= (free_space
/mss
)*mss
;
1005 /* Attempt to collapse two adjacent SKB's during retransmission. */
1006 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*skb
, int mss_now
)
1008 struct tcp_sock
*tp
= tcp_sk(sk
);
1009 struct sk_buff
*next_skb
= skb
->next
;
1011 /* The first test we must make is that neither of these two
1012 * SKB's are still referenced by someone else.
1014 if (!skb_cloned(skb
) && !skb_cloned(next_skb
)) {
1015 int skb_size
= skb
->len
, next_skb_size
= next_skb
->len
;
1016 u16 flags
= TCP_SKB_CB(skb
)->flags
;
1018 /* Also punt if next skb has been SACK'd. */
1019 if(TCP_SKB_CB(next_skb
)->sacked
& TCPCB_SACKED_ACKED
)
1022 /* Next skb is out of window. */
1023 if (after(TCP_SKB_CB(next_skb
)->end_seq
, tp
->snd_una
+tp
->snd_wnd
))
1026 /* Punt if not enough space exists in the first SKB for
1027 * the data in the second, or the total combined payload
1028 * would exceed the MSS.
1030 if ((next_skb_size
> skb_tailroom(skb
)) ||
1031 ((skb_size
+ next_skb_size
) > mss_now
))
1034 BUG_ON(tcp_skb_pcount(skb
) != 1 ||
1035 tcp_skb_pcount(next_skb
) != 1);
1037 /* Ok. We will be able to collapse the packet. */
1038 __skb_unlink(next_skb
, next_skb
->list
);
1040 memcpy(skb_put(skb
, next_skb_size
), next_skb
->data
, next_skb_size
);
1042 if (next_skb
->ip_summed
== CHECKSUM_HW
)
1043 skb
->ip_summed
= CHECKSUM_HW
;
1045 if (skb
->ip_summed
!= CHECKSUM_HW
)
1046 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
1048 /* Update sequence range on original skb. */
1049 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
1051 /* Merge over control information. */
1052 flags
|= TCP_SKB_CB(next_skb
)->flags
; /* This moves PSH/FIN etc. over */
1053 TCP_SKB_CB(skb
)->flags
= flags
;
1055 /* All done, get rid of second SKB and account for it so
1056 * packet counting does not break.
1058 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
&(TCPCB_EVER_RETRANS
|TCPCB_AT_TAIL
);
1059 if (TCP_SKB_CB(next_skb
)->sacked
&TCPCB_SACKED_RETRANS
)
1060 tp
->retrans_out
-= tcp_skb_pcount(next_skb
);
1061 if (TCP_SKB_CB(next_skb
)->sacked
&TCPCB_LOST
) {
1062 tp
->lost_out
-= tcp_skb_pcount(next_skb
);
1063 tp
->left_out
-= tcp_skb_pcount(next_skb
);
1065 /* Reno case is special. Sigh... */
1066 if (!tp
->rx_opt
.sack_ok
&& tp
->sacked_out
) {
1067 tcp_dec_pcount_approx(&tp
->sacked_out
, next_skb
);
1068 tp
->left_out
-= tcp_skb_pcount(next_skb
);
1071 /* Not quite right: it can be > snd.fack, but
1072 * it is better to underestimate fackets.
1074 tcp_dec_pcount_approx(&tp
->fackets_out
, next_skb
);
1075 tcp_packets_out_dec(tp
, next_skb
);
1076 sk_stream_free_skb(sk
, next_skb
);
1080 /* Do a simple retransmit without using the backoff mechanisms in
1081 * tcp_timer. This is used for path mtu discovery.
1082 * The socket is already locked here.
1084 void tcp_simple_retransmit(struct sock
*sk
)
1086 struct tcp_sock
*tp
= tcp_sk(sk
);
1087 struct sk_buff
*skb
;
1088 unsigned int mss
= tcp_current_mss(sk
, 0);
1091 sk_stream_for_retrans_queue(skb
, sk
) {
1092 if (skb
->len
> mss
&&
1093 !(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
)) {
1094 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) {
1095 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1096 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1098 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_LOST
)) {
1099 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1100 tp
->lost_out
+= tcp_skb_pcount(skb
);
1109 tcp_sync_left_out(tp
);
1111 /* Don't muck with the congestion window here.
1112 * Reason is that we do not increase amount of _data_
1113 * in network, but units changed and effective
1114 * cwnd/ssthresh really reduced now.
1116 if (tp
->ca_state
!= TCP_CA_Loss
) {
1117 tp
->high_seq
= tp
->snd_nxt
;
1118 tp
->snd_ssthresh
= tcp_current_ssthresh(tp
);
1119 tp
->prior_ssthresh
= 0;
1120 tp
->undo_marker
= 0;
1121 tcp_set_ca_state(tp
, TCP_CA_Loss
);
1123 tcp_xmit_retransmit_queue(sk
);
1126 /* This retransmits one SKB. Policy decisions and retransmit queue
1127 * state updates are done by the caller. Returns non-zero if an
1128 * error occurred which prevented the send.
1130 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
1132 struct tcp_sock
*tp
= tcp_sk(sk
);
1133 unsigned int cur_mss
= tcp_current_mss(sk
, 0);
1136 /* Do not sent more than we queued. 1/4 is reserved for possible
1137 * copying overhead: frgagmentation, tunneling, mangling etc.
1139 if (atomic_read(&sk
->sk_wmem_alloc
) >
1140 min(sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2), sk
->sk_sndbuf
))
1143 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
1144 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1147 if (sk
->sk_route_caps
& NETIF_F_TSO
) {
1148 sk
->sk_route_caps
&= ~NETIF_F_TSO
;
1149 sock_set_flag(sk
, SOCK_NO_LARGESEND
);
1150 tp
->mss_cache
= tp
->mss_cache_std
;
1153 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
1157 /* If receiver has shrunk his window, and skb is out of
1158 * new window, do not retransmit it. The exception is the
1159 * case, when window is shrunk to zero. In this case
1160 * our retransmit serves as a zero window probe.
1162 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
+tp
->snd_wnd
)
1163 && TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
1166 if (skb
->len
> cur_mss
) {
1167 int old_factor
= tcp_skb_pcount(skb
);
1170 if (tcp_fragment(sk
, skb
, cur_mss
))
1171 return -ENOMEM
; /* We'll try again later. */
1173 /* New SKB created, account for it. */
1174 new_factor
= tcp_skb_pcount(skb
);
1175 tp
->packets_out
-= old_factor
- new_factor
;
1176 tp
->packets_out
+= tcp_skb_pcount(skb
->next
);
1179 /* Collapse two adjacent packets if worthwhile and we can. */
1180 if(!(TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_SYN
) &&
1181 (skb
->len
< (cur_mss
>> 1)) &&
1182 (skb
->next
!= sk
->sk_send_head
) &&
1183 (skb
->next
!= (struct sk_buff
*)&sk
->sk_write_queue
) &&
1184 (skb_shinfo(skb
)->nr_frags
== 0 && skb_shinfo(skb
->next
)->nr_frags
== 0) &&
1185 (tcp_skb_pcount(skb
) == 1 && tcp_skb_pcount(skb
->next
) == 1) &&
1186 (sysctl_tcp_retrans_collapse
!= 0))
1187 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
1189 if(tp
->af_specific
->rebuild_header(sk
))
1190 return -EHOSTUNREACH
; /* Routing failure or similar. */
1192 /* Some Solaris stacks overoptimize and ignore the FIN on a
1193 * retransmit when old data is attached. So strip it off
1194 * since it is cheap to do so and saves bytes on the network.
1197 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
) &&
1198 tp
->snd_una
== (TCP_SKB_CB(skb
)->end_seq
- 1)) {
1199 if (!pskb_trim(skb
, 0)) {
1200 TCP_SKB_CB(skb
)->seq
= TCP_SKB_CB(skb
)->end_seq
- 1;
1201 skb_shinfo(skb
)->tso_segs
= 1;
1202 skb_shinfo(skb
)->tso_size
= 0;
1203 skb
->ip_summed
= CHECKSUM_NONE
;
1208 /* Make a copy, if the first transmission SKB clone we made
1209 * is still in somebody's hands, else make a clone.
1211 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1212 tcp_tso_set_push(skb
);
1214 err
= tcp_transmit_skb(sk
, (skb_cloned(skb
) ?
1215 pskb_copy(skb
, GFP_ATOMIC
):
1216 skb_clone(skb
, GFP_ATOMIC
)));
1219 /* Update global TCP statistics. */
1220 TCP_INC_STATS(TCP_MIB_RETRANSSEGS
);
1222 tp
->total_retrans
++;
1224 #if FASTRETRANS_DEBUG > 0
1225 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) {
1226 if (net_ratelimit())
1227 printk(KERN_DEBUG
"retrans_out leaked.\n");
1230 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
1231 tp
->retrans_out
+= tcp_skb_pcount(skb
);
1233 /* Save stamp of the first retransmit. */
1234 if (!tp
->retrans_stamp
)
1235 tp
->retrans_stamp
= TCP_SKB_CB(skb
)->when
;
1239 /* snd_nxt is stored to detect loss of retransmitted segment,
1240 * see tcp_input.c tcp_sacktag_write_queue().
1242 TCP_SKB_CB(skb
)->ack_seq
= tp
->snd_nxt
;
1247 /* This gets called after a retransmit timeout, and the initially
1248 * retransmitted data is acknowledged. It tries to continue
1249 * resending the rest of the retransmit queue, until either
1250 * we've sent it all or the congestion window limit is reached.
1251 * If doing SACK, the first ACK which comes back for a timeout
1252 * based retransmit packet might feed us FACK information again.
1253 * If so, we use it to avoid unnecessarily retransmissions.
1255 void tcp_xmit_retransmit_queue(struct sock
*sk
)
1257 struct tcp_sock
*tp
= tcp_sk(sk
);
1258 struct sk_buff
*skb
;
1259 int packet_cnt
= tp
->lost_out
;
1261 /* First pass: retransmit lost packets. */
1263 sk_stream_for_retrans_queue(skb
, sk
) {
1264 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
1266 /* Assume this retransmit will generate
1267 * only one packet for congestion window
1268 * calculation purposes. This works because
1269 * tcp_retransmit_skb() will chop up the
1270 * packet to be MSS sized and all the
1271 * packet counting works out.
1273 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
1276 if (sacked
&TCPCB_LOST
) {
1277 if (!(sacked
&(TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))) {
1278 if (tcp_retransmit_skb(sk
, skb
))
1280 if (tp
->ca_state
!= TCP_CA_Loss
)
1281 NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS
);
1283 NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS
);
1286 skb_peek(&sk
->sk_write_queue
))
1287 tcp_reset_xmit_timer(sk
, TCP_TIME_RETRANS
, tp
->rto
);
1290 packet_cnt
-= tcp_skb_pcount(skb
);
1291 if (packet_cnt
<= 0)
1297 /* OK, demanded retransmission is finished. */
1299 /* Forward retransmissions are possible only during Recovery. */
1300 if (tp
->ca_state
!= TCP_CA_Recovery
)
1303 /* No forward retransmissions in Reno are possible. */
1304 if (!tp
->rx_opt
.sack_ok
)
1307 /* Yeah, we have to make difficult choice between forward transmission
1308 * and retransmission... Both ways have their merits...
1310 * For now we do not retransmit anything, while we have some new
1314 if (tcp_may_send_now(sk
, tp
))
1319 sk_stream_for_retrans_queue(skb
, sk
) {
1320 /* Similar to the retransmit loop above we
1321 * can pretend that the retransmitted SKB
1322 * we send out here will be composed of one
1323 * real MSS sized packet because tcp_retransmit_skb()
1324 * will fragment it if necessary.
1326 if (++packet_cnt
> tp
->fackets_out
)
1329 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
1332 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
)
1335 /* Ok, retransmit it. */
1336 if (tcp_retransmit_skb(sk
, skb
))
1339 if (skb
== skb_peek(&sk
->sk_write_queue
))
1340 tcp_reset_xmit_timer(sk
, TCP_TIME_RETRANS
, tp
->rto
);
1342 NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS
);
1347 /* Send a fin. The caller locks the socket for us. This cannot be
1348 * allowed to fail queueing a FIN frame under any circumstances.
1350 void tcp_send_fin(struct sock
*sk
)
1352 struct tcp_sock
*tp
= tcp_sk(sk
);
1353 struct sk_buff
*skb
= skb_peek_tail(&sk
->sk_write_queue
);
1356 /* Optimization, tack on the FIN if we have a queue of
1357 * unsent frames. But be careful about outgoing SACKS
1360 mss_now
= tcp_current_mss(sk
, 1);
1362 if (sk
->sk_send_head
!= NULL
) {
1363 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_FIN
;
1364 TCP_SKB_CB(skb
)->end_seq
++;
1367 /* Socket is locked, keep trying until memory is available. */
1369 skb
= alloc_skb(MAX_TCP_HEADER
, GFP_KERNEL
);
1375 /* Reserve space for headers and prepare control bits. */
1376 skb_reserve(skb
, MAX_TCP_HEADER
);
1378 TCP_SKB_CB(skb
)->flags
= (TCPCB_FLAG_ACK
| TCPCB_FLAG_FIN
);
1379 TCP_SKB_CB(skb
)->sacked
= 0;
1380 skb_shinfo(skb
)->tso_segs
= 1;
1381 skb_shinfo(skb
)->tso_size
= 0;
1383 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
1384 TCP_SKB_CB(skb
)->seq
= tp
->write_seq
;
1385 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ 1;
1386 tcp_queue_skb(sk
, skb
);
1388 __tcp_push_pending_frames(sk
, tp
, mss_now
, TCP_NAGLE_OFF
);
1391 /* We get here when a process closes a file descriptor (either due to
1392 * an explicit close() or as a byproduct of exit()'ing) and there
1393 * was unread data in the receive queue. This behavior is recommended
1394 * by draft-ietf-tcpimpl-prob-03.txt section 3.10. -DaveM
1396 void tcp_send_active_reset(struct sock
*sk
, int priority
)
1398 struct tcp_sock
*tp
= tcp_sk(sk
);
1399 struct sk_buff
*skb
;
1401 /* NOTE: No TCP options attached and we never retransmit this. */
1402 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
1404 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED
);
1408 /* Reserve space for headers and prepare control bits. */
1409 skb_reserve(skb
, MAX_TCP_HEADER
);
1411 TCP_SKB_CB(skb
)->flags
= (TCPCB_FLAG_ACK
| TCPCB_FLAG_RST
);
1412 TCP_SKB_CB(skb
)->sacked
= 0;
1413 skb_shinfo(skb
)->tso_segs
= 1;
1414 skb_shinfo(skb
)->tso_size
= 0;
1417 TCP_SKB_CB(skb
)->seq
= tcp_acceptable_seq(sk
, tp
);
1418 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
;
1419 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1420 if (tcp_transmit_skb(sk
, skb
))
1421 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED
);
1424 /* WARNING: This routine must only be called when we have already sent
1425 * a SYN packet that crossed the incoming SYN that caused this routine
1426 * to get called. If this assumption fails then the initial rcv_wnd
1427 * and rcv_wscale values will not be correct.
1429 int tcp_send_synack(struct sock
*sk
)
1431 struct sk_buff
* skb
;
1433 skb
= skb_peek(&sk
->sk_write_queue
);
1434 if (skb
== NULL
|| !(TCP_SKB_CB(skb
)->flags
&TCPCB_FLAG_SYN
)) {
1435 printk(KERN_DEBUG
"tcp_send_synack: wrong queue state\n");
1438 if (!(TCP_SKB_CB(skb
)->flags
&TCPCB_FLAG_ACK
)) {
1439 if (skb_cloned(skb
)) {
1440 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
1443 __skb_unlink(skb
, &sk
->sk_write_queue
);
1444 skb_header_release(nskb
);
1445 __skb_queue_head(&sk
->sk_write_queue
, nskb
);
1446 sk_stream_free_skb(sk
, skb
);
1447 sk_charge_skb(sk
, nskb
);
1451 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_ACK
;
1452 TCP_ECN_send_synack(tcp_sk(sk
), skb
);
1454 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1455 return tcp_transmit_skb(sk
, skb_clone(skb
, GFP_ATOMIC
));
1459 * Prepare a SYN-ACK.
1461 struct sk_buff
* tcp_make_synack(struct sock
*sk
, struct dst_entry
*dst
,
1462 struct request_sock
*req
)
1464 struct inet_request_sock
*ireq
= inet_rsk(req
);
1465 struct tcp_sock
*tp
= tcp_sk(sk
);
1467 int tcp_header_size
;
1468 struct sk_buff
*skb
;
1470 skb
= sock_wmalloc(sk
, MAX_TCP_HEADER
+ 15, 1, GFP_ATOMIC
);
1474 /* Reserve space for headers. */
1475 skb_reserve(skb
, MAX_TCP_HEADER
);
1477 skb
->dst
= dst_clone(dst
);
1479 tcp_header_size
= (sizeof(struct tcphdr
) + TCPOLEN_MSS
+
1480 (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0) +
1481 (ireq
->wscale_ok
? TCPOLEN_WSCALE_ALIGNED
: 0) +
1482 /* SACK_PERM is in the place of NOP NOP of TS */
1483 ((ireq
->sack_ok
&& !ireq
->tstamp_ok
) ? TCPOLEN_SACKPERM_ALIGNED
: 0));
1484 skb
->h
.th
= th
= (struct tcphdr
*) skb_push(skb
, tcp_header_size
);
1486 memset(th
, 0, sizeof(struct tcphdr
));
1489 if (dst
->dev
->features
&NETIF_F_TSO
)
1491 TCP_ECN_make_synack(req
, th
);
1492 th
->source
= inet_sk(sk
)->sport
;
1493 th
->dest
= ireq
->rmt_port
;
1494 TCP_SKB_CB(skb
)->seq
= tcp_rsk(req
)->snt_isn
;
1495 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ 1;
1496 TCP_SKB_CB(skb
)->sacked
= 0;
1497 skb_shinfo(skb
)->tso_segs
= 1;
1498 skb_shinfo(skb
)->tso_size
= 0;
1499 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
1500 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_isn
+ 1);
1501 if (req
->rcv_wnd
== 0) { /* ignored for retransmitted syns */
1503 /* Set this up on the first call only */
1504 req
->window_clamp
= tp
->window_clamp
? : dst_metric(dst
, RTAX_WINDOW
);
1505 /* tcp_full_space because it is guaranteed to be the first packet */
1506 tcp_select_initial_window(tcp_full_space(sk
),
1507 dst_metric(dst
, RTAX_ADVMSS
) - (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0),
1512 ireq
->rcv_wscale
= rcv_wscale
;
1515 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
1516 th
->window
= htons(req
->rcv_wnd
);
1518 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1519 tcp_syn_build_options((__u32
*)(th
+ 1), dst_metric(dst
, RTAX_ADVMSS
), ireq
->tstamp_ok
,
1520 ireq
->sack_ok
, ireq
->wscale_ok
, ireq
->rcv_wscale
,
1521 TCP_SKB_CB(skb
)->when
,
1525 th
->doff
= (tcp_header_size
>> 2);
1526 TCP_INC_STATS(TCP_MIB_OUTSEGS
);
1531 * Do all connect socket setups that can be done AF independent.
1533 static inline void tcp_connect_init(struct sock
*sk
)
1535 struct dst_entry
*dst
= __sk_dst_get(sk
);
1536 struct tcp_sock
*tp
= tcp_sk(sk
);
1539 /* We'll fix this up when we get a response from the other end.
1540 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
1542 tp
->tcp_header_len
= sizeof(struct tcphdr
) +
1543 (sysctl_tcp_timestamps
? TCPOLEN_TSTAMP_ALIGNED
: 0);
1545 /* If user gave his TCP_MAXSEG, record it to clamp */
1546 if (tp
->rx_opt
.user_mss
)
1547 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
1549 tcp_sync_mss(sk
, dst_mtu(dst
));
1551 if (!tp
->window_clamp
)
1552 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
1553 tp
->advmss
= dst_metric(dst
, RTAX_ADVMSS
);
1554 tcp_initialize_rcv_mss(sk
);
1556 tcp_select_initial_window(tcp_full_space(sk
),
1557 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
1560 sysctl_tcp_window_scaling
,
1563 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
1564 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
1567 sock_reset_flag(sk
, SOCK_DONE
);
1569 tcp_init_wl(tp
, tp
->write_seq
, 0);
1570 tp
->snd_una
= tp
->write_seq
;
1571 tp
->snd_sml
= tp
->write_seq
;
1576 tp
->rto
= TCP_TIMEOUT_INIT
;
1577 tp
->retransmits
= 0;
1578 tcp_clear_retrans(tp
);
1582 * Build a SYN and send it off.
1584 int tcp_connect(struct sock
*sk
)
1586 struct tcp_sock
*tp
= tcp_sk(sk
);
1587 struct sk_buff
*buff
;
1589 tcp_connect_init(sk
);
1591 buff
= alloc_skb(MAX_TCP_HEADER
+ 15, sk
->sk_allocation
);
1592 if (unlikely(buff
== NULL
))
1595 /* Reserve space for headers. */
1596 skb_reserve(buff
, MAX_TCP_HEADER
);
1598 TCP_SKB_CB(buff
)->flags
= TCPCB_FLAG_SYN
;
1599 TCP_ECN_send_syn(sk
, tp
, buff
);
1600 TCP_SKB_CB(buff
)->sacked
= 0;
1601 skb_shinfo(buff
)->tso_segs
= 1;
1602 skb_shinfo(buff
)->tso_size
= 0;
1604 TCP_SKB_CB(buff
)->seq
= tp
->write_seq
++;
1605 TCP_SKB_CB(buff
)->end_seq
= tp
->write_seq
;
1606 tp
->snd_nxt
= tp
->write_seq
;
1607 tp
->pushed_seq
= tp
->write_seq
;
1610 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
1611 tp
->retrans_stamp
= TCP_SKB_CB(buff
)->when
;
1612 skb_header_release(buff
);
1613 __skb_queue_tail(&sk
->sk_write_queue
, buff
);
1614 sk_charge_skb(sk
, buff
);
1615 tp
->packets_out
+= tcp_skb_pcount(buff
);
1616 tcp_transmit_skb(sk
, skb_clone(buff
, GFP_KERNEL
));
1617 TCP_INC_STATS(TCP_MIB_ACTIVEOPENS
);
1619 /* Timer for repeating the SYN until an answer. */
1620 tcp_reset_xmit_timer(sk
, TCP_TIME_RETRANS
, tp
->rto
);
1624 /* Send out a delayed ack, the caller does the policy checking
1625 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
1628 void tcp_send_delayed_ack(struct sock
*sk
)
1630 struct tcp_sock
*tp
= tcp_sk(sk
);
1631 int ato
= tp
->ack
.ato
;
1632 unsigned long timeout
;
1634 if (ato
> TCP_DELACK_MIN
) {
1637 if (tp
->ack
.pingpong
|| (tp
->ack
.pending
&TCP_ACK_PUSHED
))
1638 max_ato
= TCP_DELACK_MAX
;
1640 /* Slow path, intersegment interval is "high". */
1642 /* If some rtt estimate is known, use it to bound delayed ack.
1643 * Do not use tp->rto here, use results of rtt measurements
1647 int rtt
= max(tp
->srtt
>>3, TCP_DELACK_MIN
);
1653 ato
= min(ato
, max_ato
);
1656 /* Stay within the limit we were given */
1657 timeout
= jiffies
+ ato
;
1659 /* Use new timeout only if there wasn't a older one earlier. */
1660 if (tp
->ack
.pending
&TCP_ACK_TIMER
) {
1661 /* If delack timer was blocked or is about to expire,
1664 if (tp
->ack
.blocked
|| time_before_eq(tp
->ack
.timeout
, jiffies
+(ato
>>2))) {
1669 if (!time_before(timeout
, tp
->ack
.timeout
))
1670 timeout
= tp
->ack
.timeout
;
1672 tp
->ack
.pending
|= TCP_ACK_SCHED
|TCP_ACK_TIMER
;
1673 tp
->ack
.timeout
= timeout
;
1674 sk_reset_timer(sk
, &tp
->delack_timer
, timeout
);
1677 /* This routine sends an ack and also updates the window. */
1678 void tcp_send_ack(struct sock
*sk
)
1680 /* If we have been reset, we may not send again. */
1681 if (sk
->sk_state
!= TCP_CLOSE
) {
1682 struct tcp_sock
*tp
= tcp_sk(sk
);
1683 struct sk_buff
*buff
;
1685 /* We are not putting this on the write queue, so
1686 * tcp_transmit_skb() will set the ownership to this
1689 buff
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
1691 tcp_schedule_ack(tp
);
1692 tp
->ack
.ato
= TCP_ATO_MIN
;
1693 tcp_reset_xmit_timer(sk
, TCP_TIME_DACK
, TCP_DELACK_MAX
);
1697 /* Reserve space for headers and prepare control bits. */
1698 skb_reserve(buff
, MAX_TCP_HEADER
);
1700 TCP_SKB_CB(buff
)->flags
= TCPCB_FLAG_ACK
;
1701 TCP_SKB_CB(buff
)->sacked
= 0;
1702 skb_shinfo(buff
)->tso_segs
= 1;
1703 skb_shinfo(buff
)->tso_size
= 0;
1705 /* Send it off, this clears delayed acks for us. */
1706 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(buff
)->end_seq
= tcp_acceptable_seq(sk
, tp
);
1707 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
1708 tcp_transmit_skb(sk
, buff
);
1712 /* This routine sends a packet with an out of date sequence
1713 * number. It assumes the other end will try to ack it.
1715 * Question: what should we make while urgent mode?
1716 * 4.4BSD forces sending single byte of data. We cannot send
1717 * out of window data, because we have SND.NXT==SND.MAX...
1719 * Current solution: to send TWO zero-length segments in urgent mode:
1720 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
1721 * out-of-date with SND.UNA-1 to probe window.
1723 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
)
1725 struct tcp_sock
*tp
= tcp_sk(sk
);
1726 struct sk_buff
*skb
;
1728 /* We don't queue it, tcp_transmit_skb() sets ownership. */
1729 skb
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
1733 /* Reserve space for headers and set control bits. */
1734 skb_reserve(skb
, MAX_TCP_HEADER
);
1736 TCP_SKB_CB(skb
)->flags
= TCPCB_FLAG_ACK
;
1737 TCP_SKB_CB(skb
)->sacked
= urgent
;
1738 skb_shinfo(skb
)->tso_segs
= 1;
1739 skb_shinfo(skb
)->tso_size
= 0;
1741 /* Use a previous sequence. This should cause the other
1742 * end to send an ack. Don't queue or clone SKB, just
1745 TCP_SKB_CB(skb
)->seq
= urgent
? tp
->snd_una
: tp
->snd_una
- 1;
1746 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
;
1747 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1748 return tcp_transmit_skb(sk
, skb
);
1751 int tcp_write_wakeup(struct sock
*sk
)
1753 if (sk
->sk_state
!= TCP_CLOSE
) {
1754 struct tcp_sock
*tp
= tcp_sk(sk
);
1755 struct sk_buff
*skb
;
1757 if ((skb
= sk
->sk_send_head
) != NULL
&&
1758 before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
+tp
->snd_wnd
)) {
1760 unsigned int mss
= tcp_current_mss(sk
, 0);
1761 unsigned int seg_size
= tp
->snd_una
+tp
->snd_wnd
-TCP_SKB_CB(skb
)->seq
;
1763 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
1764 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
1766 /* We are probing the opening of a window
1767 * but the window size is != 0
1768 * must have been a result SWS avoidance ( sender )
1770 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
1772 seg_size
= min(seg_size
, mss
);
1773 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
1774 if (tcp_fragment(sk
, skb
, seg_size
))
1776 /* SWS override triggered forced fragmentation.
1777 * Disable TSO, the connection is too sick. */
1778 if (sk
->sk_route_caps
& NETIF_F_TSO
) {
1779 sock_set_flag(sk
, SOCK_NO_LARGESEND
);
1780 sk
->sk_route_caps
&= ~NETIF_F_TSO
;
1781 tp
->mss_cache
= tp
->mss_cache_std
;
1783 } else if (!tcp_skb_pcount(skb
))
1784 tcp_set_skb_tso_segs(sk
, skb
);
1786 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
1787 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1788 tcp_tso_set_push(skb
);
1789 err
= tcp_transmit_skb(sk
, skb_clone(skb
, GFP_ATOMIC
));
1791 update_send_head(sk
, tp
, skb
);
1796 between(tp
->snd_up
, tp
->snd_una
+1, tp
->snd_una
+0xFFFF))
1797 tcp_xmit_probe_skb(sk
, TCPCB_URG
);
1798 return tcp_xmit_probe_skb(sk
, 0);
1804 /* A window probe timeout has occurred. If window is not closed send
1805 * a partial packet else a zero probe.
1807 void tcp_send_probe0(struct sock
*sk
)
1809 struct tcp_sock
*tp
= tcp_sk(sk
);
1812 err
= tcp_write_wakeup(sk
);
1814 if (tp
->packets_out
|| !sk
->sk_send_head
) {
1815 /* Cancel probe timer, if it is not required. */
1822 if (tp
->backoff
< sysctl_tcp_retries2
)
1825 tcp_reset_xmit_timer (sk
, TCP_TIME_PROBE0
,
1826 min(tp
->rto
<< tp
->backoff
, TCP_RTO_MAX
));
1828 /* If packet was not sent due to local congestion,
1829 * do not backoff and do not remember probes_out.
1830 * Let local senders to fight for local resources.
1832 * Use accumulated backoff yet.
1834 if (!tp
->probes_out
)
1836 tcp_reset_xmit_timer (sk
, TCP_TIME_PROBE0
,
1837 min(tp
->rto
<< tp
->backoff
, TCP_RESOURCE_PROBE_INTERVAL
));
1841 EXPORT_SYMBOL(tcp_connect
);
1842 EXPORT_SYMBOL(tcp_make_synack
);
1843 EXPORT_SYMBOL(tcp_simple_retransmit
);
1844 EXPORT_SYMBOL(tcp_sync_mss
);