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 int tcp_may_send_now(struct sock
*sk
, struct tcp_sock
*tp
)
516 struct sk_buff
*skb
= sk
->sk_send_head
;
519 tcp_snd_test(sk
, skb
, tcp_current_mss(sk
, 1),
520 (tcp_skb_is_last(sk
, skb
) ?
526 /* Send _single_ skb sitting at the send head. This function requires
527 * true push pending frames to setup probe timer etc.
529 void tcp_push_one(struct sock
*sk
, unsigned cur_mss
)
531 struct tcp_sock
*tp
= tcp_sk(sk
);
532 struct sk_buff
*skb
= sk
->sk_send_head
;
534 if (tcp_snd_test(sk
, skb
, cur_mss
, TCP_NAGLE_PUSH
)) {
535 /* Send it out now. */
536 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
537 tcp_tso_set_push(skb
);
538 if (!tcp_transmit_skb(sk
, skb_clone(skb
, sk
->sk_allocation
))) {
539 sk
->sk_send_head
= NULL
;
540 tp
->snd_nxt
= TCP_SKB_CB(skb
)->end_seq
;
541 tcp_packets_out_inc(sk
, tp
, skb
);
547 /* Function to create two new TCP segments. Shrinks the given segment
548 * to the specified size and appends a new segment with the rest of the
549 * packet to the list. This won't be called frequently, I hope.
550 * Remember, these are still headerless SKBs at this point.
552 static int tcp_fragment(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
554 struct tcp_sock
*tp
= tcp_sk(sk
);
555 struct sk_buff
*buff
;
559 nsize
= skb_headlen(skb
) - len
;
563 if (skb_cloned(skb
) &&
564 skb_is_nonlinear(skb
) &&
565 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
568 /* Get a new skb... force flag on. */
569 buff
= sk_stream_alloc_skb(sk
, nsize
, GFP_ATOMIC
);
571 return -ENOMEM
; /* We'll just try again later. */
572 sk_charge_skb(sk
, buff
);
574 /* Correct the sequence numbers. */
575 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
576 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
577 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
579 /* PSH and FIN should only be set in the second packet. */
580 flags
= TCP_SKB_CB(skb
)->flags
;
581 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
582 TCP_SKB_CB(buff
)->flags
= flags
;
583 TCP_SKB_CB(buff
)->sacked
=
584 (TCP_SKB_CB(skb
)->sacked
&
585 (TCPCB_LOST
| TCPCB_EVER_RETRANS
| TCPCB_AT_TAIL
));
586 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_AT_TAIL
;
588 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_HW
) {
589 /* Copy and checksum data tail into the new buffer. */
590 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
, skb_put(buff
, nsize
),
595 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
597 skb
->ip_summed
= CHECKSUM_HW
;
598 skb_split(skb
, buff
, len
);
601 buff
->ip_summed
= skb
->ip_summed
;
603 /* Looks stupid, but our code really uses when of
604 * skbs, which it never sent before. --ANK
606 TCP_SKB_CB(buff
)->when
= TCP_SKB_CB(skb
)->when
;
607 buff
->stamp
= skb
->stamp
;
609 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
) {
610 tp
->lost_out
-= tcp_skb_pcount(skb
);
611 tp
->left_out
-= tcp_skb_pcount(skb
);
614 /* Fix up tso_factor for both original and new SKB. */
615 tcp_set_skb_tso_segs(sk
, skb
);
616 tcp_set_skb_tso_segs(sk
, buff
);
618 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
) {
619 tp
->lost_out
+= tcp_skb_pcount(skb
);
620 tp
->left_out
+= tcp_skb_pcount(skb
);
623 if (TCP_SKB_CB(buff
)->sacked
&TCPCB_LOST
) {
624 tp
->lost_out
+= tcp_skb_pcount(buff
);
625 tp
->left_out
+= tcp_skb_pcount(buff
);
628 /* Link BUFF into the send queue. */
629 skb_header_release(buff
);
630 __skb_append(skb
, buff
);
635 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
636 * eventually). The difference is that pulled data not copied, but
637 * immediately discarded.
639 static unsigned char *__pskb_trim_head(struct sk_buff
*skb
, int len
)
645 for (i
=0; i
<skb_shinfo(skb
)->nr_frags
; i
++) {
646 if (skb_shinfo(skb
)->frags
[i
].size
<= eat
) {
647 put_page(skb_shinfo(skb
)->frags
[i
].page
);
648 eat
-= skb_shinfo(skb
)->frags
[i
].size
;
650 skb_shinfo(skb
)->frags
[k
] = skb_shinfo(skb
)->frags
[i
];
652 skb_shinfo(skb
)->frags
[k
].page_offset
+= eat
;
653 skb_shinfo(skb
)->frags
[k
].size
-= eat
;
659 skb_shinfo(skb
)->nr_frags
= k
;
661 skb
->tail
= skb
->data
;
662 skb
->data_len
-= len
;
663 skb
->len
= skb
->data_len
;
667 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
669 if (skb_cloned(skb
) &&
670 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
673 if (len
<= skb_headlen(skb
)) {
674 __skb_pull(skb
, len
);
676 if (__pskb_trim_head(skb
, len
-skb_headlen(skb
)) == NULL
)
680 TCP_SKB_CB(skb
)->seq
+= len
;
681 skb
->ip_summed
= CHECKSUM_HW
;
683 skb
->truesize
-= len
;
684 sk
->sk_wmem_queued
-= len
;
685 sk
->sk_forward_alloc
+= len
;
686 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
688 /* Any change of skb->len requires recalculation of tso
691 if (tcp_skb_pcount(skb
) > 1)
692 tcp_set_skb_tso_segs(sk
, skb
);
697 /* This function synchronize snd mss to current pmtu/exthdr set.
699 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
700 for TCP options, but includes only bare TCP header.
702 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
703 It is minumum of user_mss and mss received with SYN.
704 It also does not include TCP options.
706 tp->pmtu_cookie is last pmtu, seen by this function.
708 tp->mss_cache is current effective sending mss, including
709 all tcp options except for SACKs. It is evaluated,
710 taking into account current pmtu, but never exceeds
711 tp->rx_opt.mss_clamp.
713 NOTE1. rfc1122 clearly states that advertised MSS
714 DOES NOT include either tcp or ip options.
716 NOTE2. tp->pmtu_cookie and tp->mss_cache are READ ONLY outside
717 this function. --ANK (980731)
720 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
722 struct tcp_sock
*tp
= tcp_sk(sk
);
725 /* Calculate base mss without TCP options:
726 It is MMS_S - sizeof(tcphdr) of rfc1122
728 mss_now
= pmtu
- tp
->af_specific
->net_header_len
- sizeof(struct tcphdr
);
730 /* Clamp it (mss_clamp does not include tcp options) */
731 if (mss_now
> tp
->rx_opt
.mss_clamp
)
732 mss_now
= tp
->rx_opt
.mss_clamp
;
734 /* Now subtract optional transport overhead */
735 mss_now
-= tp
->ext_header_len
;
737 /* Then reserve room for full set of TCP options and 8 bytes of data */
741 /* Now subtract TCP options size, not including SACKs */
742 mss_now
-= tp
->tcp_header_len
- sizeof(struct tcphdr
);
744 /* Bound mss with half of window */
745 if (tp
->max_window
&& mss_now
> (tp
->max_window
>>1))
746 mss_now
= max((tp
->max_window
>>1), 68U - tp
->tcp_header_len
);
748 /* And store cached results */
749 tp
->pmtu_cookie
= pmtu
;
750 tp
->mss_cache
= tp
->mss_cache_std
= mss_now
;
755 /* Compute the current effective MSS, taking SACKs and IP options,
756 * and even PMTU discovery events into account.
758 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
759 * cannot be large. However, taking into account rare use of URG, this
763 unsigned int tcp_current_mss(struct sock
*sk
, int large
)
765 struct tcp_sock
*tp
= tcp_sk(sk
);
766 struct dst_entry
*dst
= __sk_dst_get(sk
);
767 unsigned int do_large
, mss_now
;
769 mss_now
= tp
->mss_cache_std
;
771 u32 mtu
= dst_mtu(dst
);
772 if (mtu
!= tp
->pmtu_cookie
)
773 mss_now
= tcp_sync_mss(sk
, mtu
);
777 (sk
->sk_route_caps
& NETIF_F_TSO
) &&
781 unsigned int large_mss
, factor
, limit
;
783 large_mss
= 65535 - tp
->af_specific
->net_header_len
-
784 tp
->ext_header_len
- tp
->tcp_header_len
;
786 if (tp
->max_window
&& large_mss
> (tp
->max_window
>>1))
787 large_mss
= max((tp
->max_window
>>1),
788 68U - tp
->tcp_header_len
);
790 factor
= large_mss
/ mss_now
;
792 /* Always keep large mss multiple of real mss, but
793 * do not exceed 1/tso_win_divisor of the congestion window
794 * so we can keep the ACK clock ticking and minimize
797 limit
= tp
->snd_cwnd
;
798 if (sysctl_tcp_tso_win_divisor
)
799 limit
/= sysctl_tcp_tso_win_divisor
;
800 limit
= max(1U, limit
);
804 tp
->mss_cache
= mss_now
* factor
;
806 mss_now
= tp
->mss_cache
;
809 if (tp
->rx_opt
.eff_sacks
)
810 mss_now
-= (TCPOLEN_SACK_BASE_ALIGNED
+
811 (tp
->rx_opt
.eff_sacks
* TCPOLEN_SACK_PERBLOCK
));
815 /* Congestion window validation. (RFC2861) */
817 static inline void tcp_cwnd_validate(struct sock
*sk
, struct tcp_sock
*tp
)
819 __u32 packets_out
= tp
->packets_out
;
821 if (packets_out
>= tp
->snd_cwnd
) {
822 /* Network is feed fully. */
823 tp
->snd_cwnd_used
= 0;
824 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
826 /* Network starves. */
827 if (tp
->packets_out
> tp
->snd_cwnd_used
)
828 tp
->snd_cwnd_used
= tp
->packets_out
;
830 if ((s32
)(tcp_time_stamp
- tp
->snd_cwnd_stamp
) >= tp
->rto
)
831 tcp_cwnd_application_limited(sk
);
835 /* This routine writes packets to the network. It advances the
836 * send_head. This happens as incoming acks open up the remote
839 * Returns 1, if no segments are in flight and we have queued segments, but
840 * cannot send anything now because of SWS or another problem.
842 static int tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
)
844 struct tcp_sock
*tp
= tcp_sk(sk
);
848 /* If we are closed, the bytes will have to remain here.
849 * In time closedown will finish, we empty the write queue and all
852 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
856 while ((skb
= sk
->sk_send_head
) &&
857 tcp_snd_test(sk
, skb
, mss_now
,
858 tcp_skb_is_last(sk
, skb
) ? nonagle
:
860 if (skb
->len
> mss_now
) {
861 if (tcp_fragment(sk
, skb
, mss_now
))
865 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
866 tcp_tso_set_push(skb
);
867 if (tcp_transmit_skb(sk
, skb_clone(skb
, GFP_ATOMIC
)))
870 /* Advance the send_head. This one is sent out.
871 * This call will increment packets_out.
873 update_send_head(sk
, tp
, skb
);
875 tcp_minshall_update(tp
, mss_now
, skb
);
880 tcp_cwnd_validate(sk
, tp
);
884 return !tp
->packets_out
&& sk
->sk_send_head
;
887 /* Push out any pending frames which were held back due to
888 * TCP_CORK or attempt at coalescing tiny packets.
889 * The socket must be locked by the caller.
891 void __tcp_push_pending_frames(struct sock
*sk
, struct tcp_sock
*tp
,
892 unsigned int cur_mss
, int nonagle
)
894 struct sk_buff
*skb
= sk
->sk_send_head
;
897 if (tcp_write_xmit(sk
, cur_mss
, nonagle
))
898 tcp_check_probe_timer(sk
, tp
);
902 /* This function returns the amount that we can raise the
903 * usable window based on the following constraints
905 * 1. The window can never be shrunk once it is offered (RFC 793)
906 * 2. We limit memory per socket
909 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
910 * RECV.NEXT + RCV.WIN fixed until:
911 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
913 * i.e. don't raise the right edge of the window until you can raise
914 * it at least MSS bytes.
916 * Unfortunately, the recommended algorithm breaks header prediction,
917 * since header prediction assumes th->window stays fixed.
919 * Strictly speaking, keeping th->window fixed violates the receiver
920 * side SWS prevention criteria. The problem is that under this rule
921 * a stream of single byte packets will cause the right side of the
922 * window to always advance by a single byte.
924 * Of course, if the sender implements sender side SWS prevention
925 * then this will not be a problem.
927 * BSD seems to make the following compromise:
929 * If the free space is less than the 1/4 of the maximum
930 * space available and the free space is less than 1/2 mss,
931 * then set the window to 0.
932 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
933 * Otherwise, just prevent the window from shrinking
934 * and from being larger than the largest representable value.
936 * This prevents incremental opening of the window in the regime
937 * where TCP is limited by the speed of the reader side taking
938 * data out of the TCP receive queue. It does nothing about
939 * those cases where the window is constrained on the sender side
940 * because the pipeline is full.
942 * BSD also seems to "accidentally" limit itself to windows that are a
943 * multiple of MSS, at least until the free space gets quite small.
944 * This would appear to be a side effect of the mbuf implementation.
945 * Combining these two algorithms results in the observed behavior
946 * of having a fixed window size at almost all times.
948 * Below we obtain similar behavior by forcing the offered window to
949 * a multiple of the mss when it is feasible to do so.
951 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
952 * Regular options like TIMESTAMP are taken into account.
954 u32
__tcp_select_window(struct sock
*sk
)
956 struct tcp_sock
*tp
= tcp_sk(sk
);
957 /* MSS for the peer's data. Previous verions used mss_clamp
958 * here. I don't know if the value based on our guesses
959 * of peer's MSS is better for the performance. It's more correct
960 * but may be worse for the performance because of rcv_mss
961 * fluctuations. --SAW 1998/11/1
963 int mss
= tp
->ack
.rcv_mss
;
964 int free_space
= tcp_space(sk
);
965 int full_space
= min_t(int, tp
->window_clamp
, tcp_full_space(sk
));
968 if (mss
> full_space
)
971 if (free_space
< full_space
/2) {
974 if (tcp_memory_pressure
)
975 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U*tp
->advmss
);
977 if (free_space
< mss
)
981 if (free_space
> tp
->rcv_ssthresh
)
982 free_space
= tp
->rcv_ssthresh
;
984 /* Don't do rounding if we are using window scaling, since the
985 * scaled window will not line up with the MSS boundary anyway.
987 window
= tp
->rcv_wnd
;
988 if (tp
->rx_opt
.rcv_wscale
) {
991 /* Advertise enough space so that it won't get scaled away.
992 * Import case: prevent zero window announcement if
993 * 1<<rcv_wscale > mss.
995 if (((window
>> tp
->rx_opt
.rcv_wscale
) << tp
->rx_opt
.rcv_wscale
) != window
)
996 window
= (((window
>> tp
->rx_opt
.rcv_wscale
) + 1)
997 << tp
->rx_opt
.rcv_wscale
);
999 /* Get the largest window that is a nice multiple of mss.
1000 * Window clamp already applied above.
1001 * If our current window offering is within 1 mss of the
1002 * free space we just keep it. This prevents the divide
1003 * and multiply from happening most of the time.
1004 * We also don't do any window rounding when the free space
1007 if (window
<= free_space
- mss
|| window
> free_space
)
1008 window
= (free_space
/mss
)*mss
;
1014 /* Attempt to collapse two adjacent SKB's during retransmission. */
1015 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*skb
, int mss_now
)
1017 struct tcp_sock
*tp
= tcp_sk(sk
);
1018 struct sk_buff
*next_skb
= skb
->next
;
1020 /* The first test we must make is that neither of these two
1021 * SKB's are still referenced by someone else.
1023 if (!skb_cloned(skb
) && !skb_cloned(next_skb
)) {
1024 int skb_size
= skb
->len
, next_skb_size
= next_skb
->len
;
1025 u16 flags
= TCP_SKB_CB(skb
)->flags
;
1027 /* Also punt if next skb has been SACK'd. */
1028 if(TCP_SKB_CB(next_skb
)->sacked
& TCPCB_SACKED_ACKED
)
1031 /* Next skb is out of window. */
1032 if (after(TCP_SKB_CB(next_skb
)->end_seq
, tp
->snd_una
+tp
->snd_wnd
))
1035 /* Punt if not enough space exists in the first SKB for
1036 * the data in the second, or the total combined payload
1037 * would exceed the MSS.
1039 if ((next_skb_size
> skb_tailroom(skb
)) ||
1040 ((skb_size
+ next_skb_size
) > mss_now
))
1043 BUG_ON(tcp_skb_pcount(skb
) != 1 ||
1044 tcp_skb_pcount(next_skb
) != 1);
1046 /* Ok. We will be able to collapse the packet. */
1047 __skb_unlink(next_skb
, next_skb
->list
);
1049 memcpy(skb_put(skb
, next_skb_size
), next_skb
->data
, next_skb_size
);
1051 if (next_skb
->ip_summed
== CHECKSUM_HW
)
1052 skb
->ip_summed
= CHECKSUM_HW
;
1054 if (skb
->ip_summed
!= CHECKSUM_HW
)
1055 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
1057 /* Update sequence range on original skb. */
1058 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
1060 /* Merge over control information. */
1061 flags
|= TCP_SKB_CB(next_skb
)->flags
; /* This moves PSH/FIN etc. over */
1062 TCP_SKB_CB(skb
)->flags
= flags
;
1064 /* All done, get rid of second SKB and account for it so
1065 * packet counting does not break.
1067 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
&(TCPCB_EVER_RETRANS
|TCPCB_AT_TAIL
);
1068 if (TCP_SKB_CB(next_skb
)->sacked
&TCPCB_SACKED_RETRANS
)
1069 tp
->retrans_out
-= tcp_skb_pcount(next_skb
);
1070 if (TCP_SKB_CB(next_skb
)->sacked
&TCPCB_LOST
) {
1071 tp
->lost_out
-= tcp_skb_pcount(next_skb
);
1072 tp
->left_out
-= tcp_skb_pcount(next_skb
);
1074 /* Reno case is special. Sigh... */
1075 if (!tp
->rx_opt
.sack_ok
&& tp
->sacked_out
) {
1076 tcp_dec_pcount_approx(&tp
->sacked_out
, next_skb
);
1077 tp
->left_out
-= tcp_skb_pcount(next_skb
);
1080 /* Not quite right: it can be > snd.fack, but
1081 * it is better to underestimate fackets.
1083 tcp_dec_pcount_approx(&tp
->fackets_out
, next_skb
);
1084 tcp_packets_out_dec(tp
, next_skb
);
1085 sk_stream_free_skb(sk
, next_skb
);
1089 /* Do a simple retransmit without using the backoff mechanisms in
1090 * tcp_timer. This is used for path mtu discovery.
1091 * The socket is already locked here.
1093 void tcp_simple_retransmit(struct sock
*sk
)
1095 struct tcp_sock
*tp
= tcp_sk(sk
);
1096 struct sk_buff
*skb
;
1097 unsigned int mss
= tcp_current_mss(sk
, 0);
1100 sk_stream_for_retrans_queue(skb
, sk
) {
1101 if (skb
->len
> mss
&&
1102 !(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
)) {
1103 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) {
1104 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1105 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1107 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_LOST
)) {
1108 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1109 tp
->lost_out
+= tcp_skb_pcount(skb
);
1118 tcp_sync_left_out(tp
);
1120 /* Don't muck with the congestion window here.
1121 * Reason is that we do not increase amount of _data_
1122 * in network, but units changed and effective
1123 * cwnd/ssthresh really reduced now.
1125 if (tp
->ca_state
!= TCP_CA_Loss
) {
1126 tp
->high_seq
= tp
->snd_nxt
;
1127 tp
->snd_ssthresh
= tcp_current_ssthresh(tp
);
1128 tp
->prior_ssthresh
= 0;
1129 tp
->undo_marker
= 0;
1130 tcp_set_ca_state(tp
, TCP_CA_Loss
);
1132 tcp_xmit_retransmit_queue(sk
);
1135 /* This retransmits one SKB. Policy decisions and retransmit queue
1136 * state updates are done by the caller. Returns non-zero if an
1137 * error occurred which prevented the send.
1139 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
1141 struct tcp_sock
*tp
= tcp_sk(sk
);
1142 unsigned int cur_mss
= tcp_current_mss(sk
, 0);
1145 /* Do not sent more than we queued. 1/4 is reserved for possible
1146 * copying overhead: frgagmentation, tunneling, mangling etc.
1148 if (atomic_read(&sk
->sk_wmem_alloc
) >
1149 min(sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2), sk
->sk_sndbuf
))
1152 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
1153 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1156 if (sk
->sk_route_caps
& NETIF_F_TSO
) {
1157 sk
->sk_route_caps
&= ~NETIF_F_TSO
;
1158 sock_set_flag(sk
, SOCK_NO_LARGESEND
);
1159 tp
->mss_cache
= tp
->mss_cache_std
;
1162 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
1166 /* If receiver has shrunk his window, and skb is out of
1167 * new window, do not retransmit it. The exception is the
1168 * case, when window is shrunk to zero. In this case
1169 * our retransmit serves as a zero window probe.
1171 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
+tp
->snd_wnd
)
1172 && TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
1175 if (skb
->len
> cur_mss
) {
1176 int old_factor
= tcp_skb_pcount(skb
);
1179 if (tcp_fragment(sk
, skb
, cur_mss
))
1180 return -ENOMEM
; /* We'll try again later. */
1182 /* New SKB created, account for it. */
1183 new_factor
= tcp_skb_pcount(skb
);
1184 tp
->packets_out
-= old_factor
- new_factor
;
1185 tp
->packets_out
+= tcp_skb_pcount(skb
->next
);
1188 /* Collapse two adjacent packets if worthwhile and we can. */
1189 if(!(TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_SYN
) &&
1190 (skb
->len
< (cur_mss
>> 1)) &&
1191 (skb
->next
!= sk
->sk_send_head
) &&
1192 (skb
->next
!= (struct sk_buff
*)&sk
->sk_write_queue
) &&
1193 (skb_shinfo(skb
)->nr_frags
== 0 && skb_shinfo(skb
->next
)->nr_frags
== 0) &&
1194 (tcp_skb_pcount(skb
) == 1 && tcp_skb_pcount(skb
->next
) == 1) &&
1195 (sysctl_tcp_retrans_collapse
!= 0))
1196 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
1198 if(tp
->af_specific
->rebuild_header(sk
))
1199 return -EHOSTUNREACH
; /* Routing failure or similar. */
1201 /* Some Solaris stacks overoptimize and ignore the FIN on a
1202 * retransmit when old data is attached. So strip it off
1203 * since it is cheap to do so and saves bytes on the network.
1206 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
) &&
1207 tp
->snd_una
== (TCP_SKB_CB(skb
)->end_seq
- 1)) {
1208 if (!pskb_trim(skb
, 0)) {
1209 TCP_SKB_CB(skb
)->seq
= TCP_SKB_CB(skb
)->end_seq
- 1;
1210 skb_shinfo(skb
)->tso_segs
= 1;
1211 skb_shinfo(skb
)->tso_size
= 0;
1212 skb
->ip_summed
= CHECKSUM_NONE
;
1217 /* Make a copy, if the first transmission SKB clone we made
1218 * is still in somebody's hands, else make a clone.
1220 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1221 tcp_tso_set_push(skb
);
1223 err
= tcp_transmit_skb(sk
, (skb_cloned(skb
) ?
1224 pskb_copy(skb
, GFP_ATOMIC
):
1225 skb_clone(skb
, GFP_ATOMIC
)));
1228 /* Update global TCP statistics. */
1229 TCP_INC_STATS(TCP_MIB_RETRANSSEGS
);
1231 tp
->total_retrans
++;
1233 #if FASTRETRANS_DEBUG > 0
1234 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) {
1235 if (net_ratelimit())
1236 printk(KERN_DEBUG
"retrans_out leaked.\n");
1239 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
1240 tp
->retrans_out
+= tcp_skb_pcount(skb
);
1242 /* Save stamp of the first retransmit. */
1243 if (!tp
->retrans_stamp
)
1244 tp
->retrans_stamp
= TCP_SKB_CB(skb
)->when
;
1248 /* snd_nxt is stored to detect loss of retransmitted segment,
1249 * see tcp_input.c tcp_sacktag_write_queue().
1251 TCP_SKB_CB(skb
)->ack_seq
= tp
->snd_nxt
;
1256 /* This gets called after a retransmit timeout, and the initially
1257 * retransmitted data is acknowledged. It tries to continue
1258 * resending the rest of the retransmit queue, until either
1259 * we've sent it all or the congestion window limit is reached.
1260 * If doing SACK, the first ACK which comes back for a timeout
1261 * based retransmit packet might feed us FACK information again.
1262 * If so, we use it to avoid unnecessarily retransmissions.
1264 void tcp_xmit_retransmit_queue(struct sock
*sk
)
1266 struct tcp_sock
*tp
= tcp_sk(sk
);
1267 struct sk_buff
*skb
;
1268 int packet_cnt
= tp
->lost_out
;
1270 /* First pass: retransmit lost packets. */
1272 sk_stream_for_retrans_queue(skb
, sk
) {
1273 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
1275 /* Assume this retransmit will generate
1276 * only one packet for congestion window
1277 * calculation purposes. This works because
1278 * tcp_retransmit_skb() will chop up the
1279 * packet to be MSS sized and all the
1280 * packet counting works out.
1282 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
1285 if (sacked
&TCPCB_LOST
) {
1286 if (!(sacked
&(TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))) {
1287 if (tcp_retransmit_skb(sk
, skb
))
1289 if (tp
->ca_state
!= TCP_CA_Loss
)
1290 NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS
);
1292 NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS
);
1295 skb_peek(&sk
->sk_write_queue
))
1296 tcp_reset_xmit_timer(sk
, TCP_TIME_RETRANS
, tp
->rto
);
1299 packet_cnt
-= tcp_skb_pcount(skb
);
1300 if (packet_cnt
<= 0)
1306 /* OK, demanded retransmission is finished. */
1308 /* Forward retransmissions are possible only during Recovery. */
1309 if (tp
->ca_state
!= TCP_CA_Recovery
)
1312 /* No forward retransmissions in Reno are possible. */
1313 if (!tp
->rx_opt
.sack_ok
)
1316 /* Yeah, we have to make difficult choice between forward transmission
1317 * and retransmission... Both ways have their merits...
1319 * For now we do not retransmit anything, while we have some new
1323 if (tcp_may_send_now(sk
, tp
))
1328 sk_stream_for_retrans_queue(skb
, sk
) {
1329 /* Similar to the retransmit loop above we
1330 * can pretend that the retransmitted SKB
1331 * we send out here will be composed of one
1332 * real MSS sized packet because tcp_retransmit_skb()
1333 * will fragment it if necessary.
1335 if (++packet_cnt
> tp
->fackets_out
)
1338 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
1341 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
)
1344 /* Ok, retransmit it. */
1345 if (tcp_retransmit_skb(sk
, skb
))
1348 if (skb
== skb_peek(&sk
->sk_write_queue
))
1349 tcp_reset_xmit_timer(sk
, TCP_TIME_RETRANS
, tp
->rto
);
1351 NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS
);
1356 /* Send a fin. The caller locks the socket for us. This cannot be
1357 * allowed to fail queueing a FIN frame under any circumstances.
1359 void tcp_send_fin(struct sock
*sk
)
1361 struct tcp_sock
*tp
= tcp_sk(sk
);
1362 struct sk_buff
*skb
= skb_peek_tail(&sk
->sk_write_queue
);
1365 /* Optimization, tack on the FIN if we have a queue of
1366 * unsent frames. But be careful about outgoing SACKS
1369 mss_now
= tcp_current_mss(sk
, 1);
1371 if (sk
->sk_send_head
!= NULL
) {
1372 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_FIN
;
1373 TCP_SKB_CB(skb
)->end_seq
++;
1376 /* Socket is locked, keep trying until memory is available. */
1378 skb
= alloc_skb(MAX_TCP_HEADER
, GFP_KERNEL
);
1384 /* Reserve space for headers and prepare control bits. */
1385 skb_reserve(skb
, MAX_TCP_HEADER
);
1387 TCP_SKB_CB(skb
)->flags
= (TCPCB_FLAG_ACK
| TCPCB_FLAG_FIN
);
1388 TCP_SKB_CB(skb
)->sacked
= 0;
1389 skb_shinfo(skb
)->tso_segs
= 1;
1390 skb_shinfo(skb
)->tso_size
= 0;
1392 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
1393 TCP_SKB_CB(skb
)->seq
= tp
->write_seq
;
1394 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ 1;
1395 tcp_queue_skb(sk
, skb
);
1397 __tcp_push_pending_frames(sk
, tp
, mss_now
, TCP_NAGLE_OFF
);
1400 /* We get here when a process closes a file descriptor (either due to
1401 * an explicit close() or as a byproduct of exit()'ing) and there
1402 * was unread data in the receive queue. This behavior is recommended
1403 * by draft-ietf-tcpimpl-prob-03.txt section 3.10. -DaveM
1405 void tcp_send_active_reset(struct sock
*sk
, int priority
)
1407 struct tcp_sock
*tp
= tcp_sk(sk
);
1408 struct sk_buff
*skb
;
1410 /* NOTE: No TCP options attached and we never retransmit this. */
1411 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
1413 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED
);
1417 /* Reserve space for headers and prepare control bits. */
1418 skb_reserve(skb
, MAX_TCP_HEADER
);
1420 TCP_SKB_CB(skb
)->flags
= (TCPCB_FLAG_ACK
| TCPCB_FLAG_RST
);
1421 TCP_SKB_CB(skb
)->sacked
= 0;
1422 skb_shinfo(skb
)->tso_segs
= 1;
1423 skb_shinfo(skb
)->tso_size
= 0;
1426 TCP_SKB_CB(skb
)->seq
= tcp_acceptable_seq(sk
, tp
);
1427 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
;
1428 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1429 if (tcp_transmit_skb(sk
, skb
))
1430 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED
);
1433 /* WARNING: This routine must only be called when we have already sent
1434 * a SYN packet that crossed the incoming SYN that caused this routine
1435 * to get called. If this assumption fails then the initial rcv_wnd
1436 * and rcv_wscale values will not be correct.
1438 int tcp_send_synack(struct sock
*sk
)
1440 struct sk_buff
* skb
;
1442 skb
= skb_peek(&sk
->sk_write_queue
);
1443 if (skb
== NULL
|| !(TCP_SKB_CB(skb
)->flags
&TCPCB_FLAG_SYN
)) {
1444 printk(KERN_DEBUG
"tcp_send_synack: wrong queue state\n");
1447 if (!(TCP_SKB_CB(skb
)->flags
&TCPCB_FLAG_ACK
)) {
1448 if (skb_cloned(skb
)) {
1449 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
1452 __skb_unlink(skb
, &sk
->sk_write_queue
);
1453 skb_header_release(nskb
);
1454 __skb_queue_head(&sk
->sk_write_queue
, nskb
);
1455 sk_stream_free_skb(sk
, skb
);
1456 sk_charge_skb(sk
, nskb
);
1460 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_ACK
;
1461 TCP_ECN_send_synack(tcp_sk(sk
), skb
);
1463 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1464 return tcp_transmit_skb(sk
, skb_clone(skb
, GFP_ATOMIC
));
1468 * Prepare a SYN-ACK.
1470 struct sk_buff
* tcp_make_synack(struct sock
*sk
, struct dst_entry
*dst
,
1471 struct request_sock
*req
)
1473 struct inet_request_sock
*ireq
= inet_rsk(req
);
1474 struct tcp_sock
*tp
= tcp_sk(sk
);
1476 int tcp_header_size
;
1477 struct sk_buff
*skb
;
1479 skb
= sock_wmalloc(sk
, MAX_TCP_HEADER
+ 15, 1, GFP_ATOMIC
);
1483 /* Reserve space for headers. */
1484 skb_reserve(skb
, MAX_TCP_HEADER
);
1486 skb
->dst
= dst_clone(dst
);
1488 tcp_header_size
= (sizeof(struct tcphdr
) + TCPOLEN_MSS
+
1489 (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0) +
1490 (ireq
->wscale_ok
? TCPOLEN_WSCALE_ALIGNED
: 0) +
1491 /* SACK_PERM is in the place of NOP NOP of TS */
1492 ((ireq
->sack_ok
&& !ireq
->tstamp_ok
) ? TCPOLEN_SACKPERM_ALIGNED
: 0));
1493 skb
->h
.th
= th
= (struct tcphdr
*) skb_push(skb
, tcp_header_size
);
1495 memset(th
, 0, sizeof(struct tcphdr
));
1498 if (dst
->dev
->features
&NETIF_F_TSO
)
1500 TCP_ECN_make_synack(req
, th
);
1501 th
->source
= inet_sk(sk
)->sport
;
1502 th
->dest
= ireq
->rmt_port
;
1503 TCP_SKB_CB(skb
)->seq
= tcp_rsk(req
)->snt_isn
;
1504 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ 1;
1505 TCP_SKB_CB(skb
)->sacked
= 0;
1506 skb_shinfo(skb
)->tso_segs
= 1;
1507 skb_shinfo(skb
)->tso_size
= 0;
1508 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
1509 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_isn
+ 1);
1510 if (req
->rcv_wnd
== 0) { /* ignored for retransmitted syns */
1512 /* Set this up on the first call only */
1513 req
->window_clamp
= tp
->window_clamp
? : dst_metric(dst
, RTAX_WINDOW
);
1514 /* tcp_full_space because it is guaranteed to be the first packet */
1515 tcp_select_initial_window(tcp_full_space(sk
),
1516 dst_metric(dst
, RTAX_ADVMSS
) - (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0),
1521 ireq
->rcv_wscale
= rcv_wscale
;
1524 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
1525 th
->window
= htons(req
->rcv_wnd
);
1527 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1528 tcp_syn_build_options((__u32
*)(th
+ 1), dst_metric(dst
, RTAX_ADVMSS
), ireq
->tstamp_ok
,
1529 ireq
->sack_ok
, ireq
->wscale_ok
, ireq
->rcv_wscale
,
1530 TCP_SKB_CB(skb
)->when
,
1534 th
->doff
= (tcp_header_size
>> 2);
1535 TCP_INC_STATS(TCP_MIB_OUTSEGS
);
1540 * Do all connect socket setups that can be done AF independent.
1542 static inline void tcp_connect_init(struct sock
*sk
)
1544 struct dst_entry
*dst
= __sk_dst_get(sk
);
1545 struct tcp_sock
*tp
= tcp_sk(sk
);
1548 /* We'll fix this up when we get a response from the other end.
1549 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
1551 tp
->tcp_header_len
= sizeof(struct tcphdr
) +
1552 (sysctl_tcp_timestamps
? TCPOLEN_TSTAMP_ALIGNED
: 0);
1554 /* If user gave his TCP_MAXSEG, record it to clamp */
1555 if (tp
->rx_opt
.user_mss
)
1556 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
1558 tcp_sync_mss(sk
, dst_mtu(dst
));
1560 if (!tp
->window_clamp
)
1561 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
1562 tp
->advmss
= dst_metric(dst
, RTAX_ADVMSS
);
1563 tcp_initialize_rcv_mss(sk
);
1565 tcp_select_initial_window(tcp_full_space(sk
),
1566 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
1569 sysctl_tcp_window_scaling
,
1572 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
1573 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
1576 sock_reset_flag(sk
, SOCK_DONE
);
1578 tcp_init_wl(tp
, tp
->write_seq
, 0);
1579 tp
->snd_una
= tp
->write_seq
;
1580 tp
->snd_sml
= tp
->write_seq
;
1585 tp
->rto
= TCP_TIMEOUT_INIT
;
1586 tp
->retransmits
= 0;
1587 tcp_clear_retrans(tp
);
1591 * Build a SYN and send it off.
1593 int tcp_connect(struct sock
*sk
)
1595 struct tcp_sock
*tp
= tcp_sk(sk
);
1596 struct sk_buff
*buff
;
1598 tcp_connect_init(sk
);
1600 buff
= alloc_skb(MAX_TCP_HEADER
+ 15, sk
->sk_allocation
);
1601 if (unlikely(buff
== NULL
))
1604 /* Reserve space for headers. */
1605 skb_reserve(buff
, MAX_TCP_HEADER
);
1607 TCP_SKB_CB(buff
)->flags
= TCPCB_FLAG_SYN
;
1608 TCP_ECN_send_syn(sk
, tp
, buff
);
1609 TCP_SKB_CB(buff
)->sacked
= 0;
1610 skb_shinfo(buff
)->tso_segs
= 1;
1611 skb_shinfo(buff
)->tso_size
= 0;
1613 TCP_SKB_CB(buff
)->seq
= tp
->write_seq
++;
1614 TCP_SKB_CB(buff
)->end_seq
= tp
->write_seq
;
1615 tp
->snd_nxt
= tp
->write_seq
;
1616 tp
->pushed_seq
= tp
->write_seq
;
1619 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
1620 tp
->retrans_stamp
= TCP_SKB_CB(buff
)->when
;
1621 skb_header_release(buff
);
1622 __skb_queue_tail(&sk
->sk_write_queue
, buff
);
1623 sk_charge_skb(sk
, buff
);
1624 tp
->packets_out
+= tcp_skb_pcount(buff
);
1625 tcp_transmit_skb(sk
, skb_clone(buff
, GFP_KERNEL
));
1626 TCP_INC_STATS(TCP_MIB_ACTIVEOPENS
);
1628 /* Timer for repeating the SYN until an answer. */
1629 tcp_reset_xmit_timer(sk
, TCP_TIME_RETRANS
, tp
->rto
);
1633 /* Send out a delayed ack, the caller does the policy checking
1634 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
1637 void tcp_send_delayed_ack(struct sock
*sk
)
1639 struct tcp_sock
*tp
= tcp_sk(sk
);
1640 int ato
= tp
->ack
.ato
;
1641 unsigned long timeout
;
1643 if (ato
> TCP_DELACK_MIN
) {
1646 if (tp
->ack
.pingpong
|| (tp
->ack
.pending
&TCP_ACK_PUSHED
))
1647 max_ato
= TCP_DELACK_MAX
;
1649 /* Slow path, intersegment interval is "high". */
1651 /* If some rtt estimate is known, use it to bound delayed ack.
1652 * Do not use tp->rto here, use results of rtt measurements
1656 int rtt
= max(tp
->srtt
>>3, TCP_DELACK_MIN
);
1662 ato
= min(ato
, max_ato
);
1665 /* Stay within the limit we were given */
1666 timeout
= jiffies
+ ato
;
1668 /* Use new timeout only if there wasn't a older one earlier. */
1669 if (tp
->ack
.pending
&TCP_ACK_TIMER
) {
1670 /* If delack timer was blocked or is about to expire,
1673 if (tp
->ack
.blocked
|| time_before_eq(tp
->ack
.timeout
, jiffies
+(ato
>>2))) {
1678 if (!time_before(timeout
, tp
->ack
.timeout
))
1679 timeout
= tp
->ack
.timeout
;
1681 tp
->ack
.pending
|= TCP_ACK_SCHED
|TCP_ACK_TIMER
;
1682 tp
->ack
.timeout
= timeout
;
1683 sk_reset_timer(sk
, &tp
->delack_timer
, timeout
);
1686 /* This routine sends an ack and also updates the window. */
1687 void tcp_send_ack(struct sock
*sk
)
1689 /* If we have been reset, we may not send again. */
1690 if (sk
->sk_state
!= TCP_CLOSE
) {
1691 struct tcp_sock
*tp
= tcp_sk(sk
);
1692 struct sk_buff
*buff
;
1694 /* We are not putting this on the write queue, so
1695 * tcp_transmit_skb() will set the ownership to this
1698 buff
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
1700 tcp_schedule_ack(tp
);
1701 tp
->ack
.ato
= TCP_ATO_MIN
;
1702 tcp_reset_xmit_timer(sk
, TCP_TIME_DACK
, TCP_DELACK_MAX
);
1706 /* Reserve space for headers and prepare control bits. */
1707 skb_reserve(buff
, MAX_TCP_HEADER
);
1709 TCP_SKB_CB(buff
)->flags
= TCPCB_FLAG_ACK
;
1710 TCP_SKB_CB(buff
)->sacked
= 0;
1711 skb_shinfo(buff
)->tso_segs
= 1;
1712 skb_shinfo(buff
)->tso_size
= 0;
1714 /* Send it off, this clears delayed acks for us. */
1715 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(buff
)->end_seq
= tcp_acceptable_seq(sk
, tp
);
1716 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
1717 tcp_transmit_skb(sk
, buff
);
1721 /* This routine sends a packet with an out of date sequence
1722 * number. It assumes the other end will try to ack it.
1724 * Question: what should we make while urgent mode?
1725 * 4.4BSD forces sending single byte of data. We cannot send
1726 * out of window data, because we have SND.NXT==SND.MAX...
1728 * Current solution: to send TWO zero-length segments in urgent mode:
1729 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
1730 * out-of-date with SND.UNA-1 to probe window.
1732 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
)
1734 struct tcp_sock
*tp
= tcp_sk(sk
);
1735 struct sk_buff
*skb
;
1737 /* We don't queue it, tcp_transmit_skb() sets ownership. */
1738 skb
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
1742 /* Reserve space for headers and set control bits. */
1743 skb_reserve(skb
, MAX_TCP_HEADER
);
1745 TCP_SKB_CB(skb
)->flags
= TCPCB_FLAG_ACK
;
1746 TCP_SKB_CB(skb
)->sacked
= urgent
;
1747 skb_shinfo(skb
)->tso_segs
= 1;
1748 skb_shinfo(skb
)->tso_size
= 0;
1750 /* Use a previous sequence. This should cause the other
1751 * end to send an ack. Don't queue or clone SKB, just
1754 TCP_SKB_CB(skb
)->seq
= urgent
? tp
->snd_una
: tp
->snd_una
- 1;
1755 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
;
1756 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1757 return tcp_transmit_skb(sk
, skb
);
1760 int tcp_write_wakeup(struct sock
*sk
)
1762 if (sk
->sk_state
!= TCP_CLOSE
) {
1763 struct tcp_sock
*tp
= tcp_sk(sk
);
1764 struct sk_buff
*skb
;
1766 if ((skb
= sk
->sk_send_head
) != NULL
&&
1767 before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
+tp
->snd_wnd
)) {
1769 unsigned int mss
= tcp_current_mss(sk
, 0);
1770 unsigned int seg_size
= tp
->snd_una
+tp
->snd_wnd
-TCP_SKB_CB(skb
)->seq
;
1772 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
1773 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
1775 /* We are probing the opening of a window
1776 * but the window size is != 0
1777 * must have been a result SWS avoidance ( sender )
1779 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
1781 seg_size
= min(seg_size
, mss
);
1782 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
1783 if (tcp_fragment(sk
, skb
, seg_size
))
1785 /* SWS override triggered forced fragmentation.
1786 * Disable TSO, the connection is too sick. */
1787 if (sk
->sk_route_caps
& NETIF_F_TSO
) {
1788 sock_set_flag(sk
, SOCK_NO_LARGESEND
);
1789 sk
->sk_route_caps
&= ~NETIF_F_TSO
;
1790 tp
->mss_cache
= tp
->mss_cache_std
;
1792 } else if (!tcp_skb_pcount(skb
))
1793 tcp_set_skb_tso_segs(sk
, skb
);
1795 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
1796 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1797 tcp_tso_set_push(skb
);
1798 err
= tcp_transmit_skb(sk
, skb_clone(skb
, GFP_ATOMIC
));
1800 update_send_head(sk
, tp
, skb
);
1805 between(tp
->snd_up
, tp
->snd_una
+1, tp
->snd_una
+0xFFFF))
1806 tcp_xmit_probe_skb(sk
, TCPCB_URG
);
1807 return tcp_xmit_probe_skb(sk
, 0);
1813 /* A window probe timeout has occurred. If window is not closed send
1814 * a partial packet else a zero probe.
1816 void tcp_send_probe0(struct sock
*sk
)
1818 struct tcp_sock
*tp
= tcp_sk(sk
);
1821 err
= tcp_write_wakeup(sk
);
1823 if (tp
->packets_out
|| !sk
->sk_send_head
) {
1824 /* Cancel probe timer, if it is not required. */
1831 if (tp
->backoff
< sysctl_tcp_retries2
)
1834 tcp_reset_xmit_timer (sk
, TCP_TIME_PROBE0
,
1835 min(tp
->rto
<< tp
->backoff
, TCP_RTO_MAX
));
1837 /* If packet was not sent due to local congestion,
1838 * do not backoff and do not remember probes_out.
1839 * Let local senders to fight for local resources.
1841 * Use accumulated backoff yet.
1843 if (!tp
->probes_out
)
1845 tcp_reset_xmit_timer (sk
, TCP_TIME_PROBE0
,
1846 min(tp
->rto
<< tp
->backoff
, TCP_RESOURCE_PROBE_INTERVAL
));
1850 EXPORT_SYMBOL(tcp_connect
);
1851 EXPORT_SYMBOL(tcp_make_synack
);
1852 EXPORT_SYMBOL(tcp_simple_retransmit
);
1853 EXPORT_SYMBOL(tcp_sync_mss
);