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 __read_mostly
= 1;
48 /* People can turn this on to work with those rare, broken TCPs that
49 * interpret the window field as a signed quantity.
51 int sysctl_tcp_workaround_signed_windows __read_mostly
= 0;
53 /* This limits the percentage of the congestion window which we
54 * will allow a single TSO frame to consume. Building TSO frames
55 * which are too large can cause TCP streams to be bursty.
57 int sysctl_tcp_tso_win_divisor __read_mostly
= 3;
59 int sysctl_tcp_mtu_probing __read_mostly
= 0;
60 int sysctl_tcp_base_mss __read_mostly
= 512;
62 /* By default, RFC2861 behavior. */
63 int sysctl_tcp_slow_start_after_idle __read_mostly
= 1;
65 static void update_send_head(struct sock
*sk
, struct tcp_sock
*tp
,
68 sk
->sk_send_head
= skb
->next
;
69 if (sk
->sk_send_head
== (struct sk_buff
*)&sk
->sk_write_queue
)
70 sk
->sk_send_head
= NULL
;
71 tp
->snd_nxt
= TCP_SKB_CB(skb
)->end_seq
;
72 tcp_packets_out_inc(sk
, tp
, skb
);
75 /* SND.NXT, if window was not shrunk.
76 * If window has been shrunk, what should we make? It is not clear at all.
77 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
78 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
79 * invalid. OK, let's make this for now:
81 static inline __u32
tcp_acceptable_seq(struct sock
*sk
, struct tcp_sock
*tp
)
83 if (!before(tp
->snd_una
+tp
->snd_wnd
, tp
->snd_nxt
))
86 return tp
->snd_una
+tp
->snd_wnd
;
89 /* Calculate mss to advertise in SYN segment.
90 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
92 * 1. It is independent of path mtu.
93 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
94 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
95 * attached devices, because some buggy hosts are confused by
97 * 4. We do not make 3, we advertise MSS, calculated from first
98 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
99 * This may be overridden via information stored in routing table.
100 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
101 * probably even Jumbo".
103 static __u16
tcp_advertise_mss(struct sock
*sk
)
105 struct tcp_sock
*tp
= tcp_sk(sk
);
106 struct dst_entry
*dst
= __sk_dst_get(sk
);
107 int mss
= tp
->advmss
;
109 if (dst
&& dst_metric(dst
, RTAX_ADVMSS
) < mss
) {
110 mss
= dst_metric(dst
, RTAX_ADVMSS
);
117 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
118 * This is the first part of cwnd validation mechanism. */
119 static void tcp_cwnd_restart(struct sock
*sk
, struct dst_entry
*dst
)
121 struct tcp_sock
*tp
= tcp_sk(sk
);
122 s32 delta
= tcp_time_stamp
- tp
->lsndtime
;
123 u32 restart_cwnd
= tcp_init_cwnd(tp
, dst
);
124 u32 cwnd
= tp
->snd_cwnd
;
126 tcp_ca_event(sk
, CA_EVENT_CWND_RESTART
);
128 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
129 restart_cwnd
= min(restart_cwnd
, cwnd
);
131 while ((delta
-= inet_csk(sk
)->icsk_rto
) > 0 && cwnd
> restart_cwnd
)
133 tp
->snd_cwnd
= max(cwnd
, restart_cwnd
);
134 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
135 tp
->snd_cwnd_used
= 0;
138 static void tcp_event_data_sent(struct tcp_sock
*tp
,
139 struct sk_buff
*skb
, struct sock
*sk
)
141 struct inet_connection_sock
*icsk
= inet_csk(sk
);
142 const u32 now
= tcp_time_stamp
;
144 if (sysctl_tcp_slow_start_after_idle
&&
145 (!tp
->packets_out
&& (s32
)(now
- tp
->lsndtime
) > icsk
->icsk_rto
))
146 tcp_cwnd_restart(sk
, __sk_dst_get(sk
));
150 /* If it is a reply for ato after last received
151 * packet, enter pingpong mode.
153 if ((u32
)(now
- icsk
->icsk_ack
.lrcvtime
) < icsk
->icsk_ack
.ato
)
154 icsk
->icsk_ack
.pingpong
= 1;
157 static inline void tcp_event_ack_sent(struct sock
*sk
, unsigned int pkts
)
159 tcp_dec_quickack_mode(sk
, pkts
);
160 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_DACK
);
163 /* Determine a window scaling and initial window to offer.
164 * Based on the assumption that the given amount of space
165 * will be offered. Store the results in the tp structure.
166 * NOTE: for smooth operation initial space offering should
167 * be a multiple of mss if possible. We assume here that mss >= 1.
168 * This MUST be enforced by all callers.
170 void tcp_select_initial_window(int __space
, __u32 mss
,
171 __u32
*rcv_wnd
, __u32
*window_clamp
,
172 int wscale_ok
, __u8
*rcv_wscale
)
174 unsigned int space
= (__space
< 0 ? 0 : __space
);
176 /* If no clamp set the clamp to the max possible scaled window */
177 if (*window_clamp
== 0)
178 (*window_clamp
) = (65535 << 14);
179 space
= min(*window_clamp
, space
);
181 /* Quantize space offering to a multiple of mss if possible. */
183 space
= (space
/ mss
) * mss
;
185 /* NOTE: offering an initial window larger than 32767
186 * will break some buggy TCP stacks. If the admin tells us
187 * it is likely we could be speaking with such a buggy stack
188 * we will truncate our initial window offering to 32K-1
189 * unless the remote has sent us a window scaling option,
190 * which we interpret as a sign the remote TCP is not
191 * misinterpreting the window field as a signed quantity.
193 if (sysctl_tcp_workaround_signed_windows
)
194 (*rcv_wnd
) = min(space
, MAX_TCP_WINDOW
);
200 /* Set window scaling on max possible window
201 * See RFC1323 for an explanation of the limit to 14
203 space
= max_t(u32
, sysctl_tcp_rmem
[2], sysctl_rmem_max
);
204 space
= min_t(u32
, space
, *window_clamp
);
205 while (space
> 65535 && (*rcv_wscale
) < 14) {
211 /* Set initial window to value enough for senders,
212 * following RFC2414. Senders, not following this RFC,
213 * will be satisfied with 2.
215 if (mss
> (1<<*rcv_wscale
)) {
221 if (*rcv_wnd
> init_cwnd
*mss
)
222 *rcv_wnd
= init_cwnd
*mss
;
225 /* Set the clamp no higher than max representable value */
226 (*window_clamp
) = min(65535U << (*rcv_wscale
), *window_clamp
);
229 /* Chose a new window to advertise, update state in tcp_sock for the
230 * socket, and return result with RFC1323 scaling applied. The return
231 * value can be stuffed directly into th->window for an outgoing
234 static u16
tcp_select_window(struct sock
*sk
)
236 struct tcp_sock
*tp
= tcp_sk(sk
);
237 u32 cur_win
= tcp_receive_window(tp
);
238 u32 new_win
= __tcp_select_window(sk
);
240 /* Never shrink the offered window */
241 if(new_win
< cur_win
) {
242 /* Danger Will Robinson!
243 * Don't update rcv_wup/rcv_wnd here or else
244 * we will not be able to advertise a zero
245 * window in time. --DaveM
247 * Relax Will Robinson.
251 tp
->rcv_wnd
= new_win
;
252 tp
->rcv_wup
= tp
->rcv_nxt
;
254 /* Make sure we do not exceed the maximum possible
257 if (!tp
->rx_opt
.rcv_wscale
&& sysctl_tcp_workaround_signed_windows
)
258 new_win
= min(new_win
, MAX_TCP_WINDOW
);
260 new_win
= min(new_win
, (65535U << tp
->rx_opt
.rcv_wscale
));
262 /* RFC1323 scaling applied */
263 new_win
>>= tp
->rx_opt
.rcv_wscale
;
265 /* If we advertise zero window, disable fast path. */
272 static void tcp_build_and_update_options(__be32
*ptr
, struct tcp_sock
*tp
,
275 if (tp
->rx_opt
.tstamp_ok
) {
276 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
278 (TCPOPT_TIMESTAMP
<< 8) |
280 *ptr
++ = htonl(tstamp
);
281 *ptr
++ = htonl(tp
->rx_opt
.ts_recent
);
283 if (tp
->rx_opt
.eff_sacks
) {
284 struct tcp_sack_block
*sp
= tp
->rx_opt
.dsack
? tp
->duplicate_sack
: tp
->selective_acks
;
287 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
290 (TCPOLEN_SACK_BASE
+ (tp
->rx_opt
.eff_sacks
*
291 TCPOLEN_SACK_PERBLOCK
)));
292 for(this_sack
= 0; this_sack
< tp
->rx_opt
.eff_sacks
; this_sack
++) {
293 *ptr
++ = htonl(sp
[this_sack
].start_seq
);
294 *ptr
++ = htonl(sp
[this_sack
].end_seq
);
296 if (tp
->rx_opt
.dsack
) {
297 tp
->rx_opt
.dsack
= 0;
298 tp
->rx_opt
.eff_sacks
--;
303 /* Construct a tcp options header for a SYN or SYN_ACK packet.
304 * If this is every changed make sure to change the definition of
305 * MAX_SYN_SIZE to match the new maximum number of options that you
308 static void tcp_syn_build_options(__be32
*ptr
, int mss
, int ts
, int sack
,
309 int offer_wscale
, int wscale
, __u32 tstamp
,
312 /* We always get an MSS option.
313 * The option bytes which will be seen in normal data
314 * packets should timestamps be used, must be in the MSS
315 * advertised. But we subtract them from tp->mss_cache so
316 * that calculations in tcp_sendmsg are simpler etc.
317 * So account for this fact here if necessary. If we
318 * don't do this correctly, as a receiver we won't
319 * recognize data packets as being full sized when we
320 * should, and thus we won't abide by the delayed ACK
322 * SACKs don't matter, we never delay an ACK when we
323 * have any of those going out.
325 *ptr
++ = htonl((TCPOPT_MSS
<< 24) | (TCPOLEN_MSS
<< 16) | mss
);
328 *ptr
++ = htonl((TCPOPT_SACK_PERM
<< 24) |
329 (TCPOLEN_SACK_PERM
<< 16) |
330 (TCPOPT_TIMESTAMP
<< 8) |
333 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
335 (TCPOPT_TIMESTAMP
<< 8) |
337 *ptr
++ = htonl(tstamp
); /* TSVAL */
338 *ptr
++ = htonl(ts_recent
); /* TSECR */
340 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
342 (TCPOPT_SACK_PERM
<< 8) |
345 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
346 (TCPOPT_WINDOW
<< 16) |
347 (TCPOLEN_WINDOW
<< 8) |
351 /* This routine actually transmits TCP packets queued in by
352 * tcp_do_sendmsg(). This is used by both the initial
353 * transmission and possible later retransmissions.
354 * All SKB's seen here are completely headerless. It is our
355 * job to build the TCP header, and pass the packet down to
356 * IP so it can do the same plus pass the packet off to the
359 * We are working here with either a clone of the original
360 * SKB, or a fresh unique copy made by the retransmit engine.
362 static int tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int clone_it
, gfp_t gfp_mask
)
364 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
365 struct inet_sock
*inet
;
367 struct tcp_skb_cb
*tcb
;
373 BUG_ON(!skb
|| !tcp_skb_pcount(skb
));
375 /* If congestion control is doing timestamping, we must
376 * take such a timestamp before we potentially clone/copy.
378 if (icsk
->icsk_ca_ops
->rtt_sample
)
379 __net_timestamp(skb
);
381 if (likely(clone_it
)) {
382 if (unlikely(skb_cloned(skb
)))
383 skb
= pskb_copy(skb
, gfp_mask
);
385 skb
= skb_clone(skb
, gfp_mask
);
392 tcb
= TCP_SKB_CB(skb
);
393 tcp_header_size
= tp
->tcp_header_len
;
395 #define SYSCTL_FLAG_TSTAMPS 0x1
396 #define SYSCTL_FLAG_WSCALE 0x2
397 #define SYSCTL_FLAG_SACK 0x4
400 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
401 tcp_header_size
= sizeof(struct tcphdr
) + TCPOLEN_MSS
;
402 if(sysctl_tcp_timestamps
) {
403 tcp_header_size
+= TCPOLEN_TSTAMP_ALIGNED
;
404 sysctl_flags
|= SYSCTL_FLAG_TSTAMPS
;
406 if (sysctl_tcp_window_scaling
) {
407 tcp_header_size
+= TCPOLEN_WSCALE_ALIGNED
;
408 sysctl_flags
|= SYSCTL_FLAG_WSCALE
;
410 if (sysctl_tcp_sack
) {
411 sysctl_flags
|= SYSCTL_FLAG_SACK
;
412 if (!(sysctl_flags
& SYSCTL_FLAG_TSTAMPS
))
413 tcp_header_size
+= TCPOLEN_SACKPERM_ALIGNED
;
415 } else if (unlikely(tp
->rx_opt
.eff_sacks
)) {
416 /* A SACK is 2 pad bytes, a 2 byte header, plus
417 * 2 32-bit sequence numbers for each SACK block.
419 tcp_header_size
+= (TCPOLEN_SACK_BASE_ALIGNED
+
420 (tp
->rx_opt
.eff_sacks
*
421 TCPOLEN_SACK_PERBLOCK
));
424 if (tcp_packets_in_flight(tp
) == 0)
425 tcp_ca_event(sk
, CA_EVENT_TX_START
);
427 th
= (struct tcphdr
*) skb_push(skb
, tcp_header_size
);
429 skb_set_owner_w(skb
, sk
);
431 /* Build TCP header and checksum it. */
432 th
->source
= inet
->sport
;
433 th
->dest
= inet
->dport
;
434 th
->seq
= htonl(tcb
->seq
);
435 th
->ack_seq
= htonl(tp
->rcv_nxt
);
436 *(((__be16
*)th
) + 6) = htons(((tcp_header_size
>> 2) << 12) |
439 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
440 /* RFC1323: The window in SYN & SYN/ACK segments
443 th
->window
= htons(tp
->rcv_wnd
);
445 th
->window
= htons(tcp_select_window(sk
));
450 if (unlikely(tp
->urg_mode
&&
451 between(tp
->snd_up
, tcb
->seq
+1, tcb
->seq
+0xFFFF))) {
452 th
->urg_ptr
= htons(tp
->snd_up
-tcb
->seq
);
456 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
457 tcp_syn_build_options((__be32
*)(th
+ 1),
458 tcp_advertise_mss(sk
),
459 (sysctl_flags
& SYSCTL_FLAG_TSTAMPS
),
460 (sysctl_flags
& SYSCTL_FLAG_SACK
),
461 (sysctl_flags
& SYSCTL_FLAG_WSCALE
),
462 tp
->rx_opt
.rcv_wscale
,
464 tp
->rx_opt
.ts_recent
);
466 tcp_build_and_update_options((__be32
*)(th
+ 1),
468 TCP_ECN_send(sk
, tp
, skb
, tcp_header_size
);
471 icsk
->icsk_af_ops
->send_check(sk
, skb
->len
, skb
);
473 if (likely(tcb
->flags
& TCPCB_FLAG_ACK
))
474 tcp_event_ack_sent(sk
, tcp_skb_pcount(skb
));
476 if (skb
->len
!= tcp_header_size
)
477 tcp_event_data_sent(tp
, skb
, sk
);
479 if (after(tcb
->end_seq
, tp
->snd_nxt
) || tcb
->seq
== tcb
->end_seq
)
480 TCP_INC_STATS(TCP_MIB_OUTSEGS
);
482 err
= icsk
->icsk_af_ops
->queue_xmit(skb
, 0);
483 if (likely(err
<= 0))
488 /* NET_XMIT_CN is special. It does not guarantee,
489 * that this packet is lost. It tells that device
490 * is about to start to drop packets or already
491 * drops some packets of the same priority and
492 * invokes us to send less aggressively.
494 return err
== NET_XMIT_CN
? 0 : err
;
496 #undef SYSCTL_FLAG_TSTAMPS
497 #undef SYSCTL_FLAG_WSCALE
498 #undef SYSCTL_FLAG_SACK
502 /* This routine just queue's the buffer
504 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
505 * otherwise socket can stall.
507 static void tcp_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
509 struct tcp_sock
*tp
= tcp_sk(sk
);
511 /* Advance write_seq and place onto the write_queue. */
512 tp
->write_seq
= TCP_SKB_CB(skb
)->end_seq
;
513 skb_header_release(skb
);
514 __skb_queue_tail(&sk
->sk_write_queue
, skb
);
515 sk_charge_skb(sk
, skb
);
517 /* Queue it, remembering where we must start sending. */
518 if (sk
->sk_send_head
== NULL
)
519 sk
->sk_send_head
= skb
;
522 static void tcp_set_skb_tso_segs(struct sock
*sk
, struct sk_buff
*skb
, unsigned int mss_now
)
524 if (skb
->len
<= mss_now
|| !sk_can_gso(sk
)) {
525 /* Avoid the costly divide in the normal
528 skb_shinfo(skb
)->gso_segs
= 1;
529 skb_shinfo(skb
)->gso_size
= 0;
530 skb_shinfo(skb
)->gso_type
= 0;
534 factor
= skb
->len
+ (mss_now
- 1);
536 skb_shinfo(skb
)->gso_segs
= factor
;
537 skb_shinfo(skb
)->gso_size
= mss_now
;
538 skb_shinfo(skb
)->gso_type
= sk
->sk_gso_type
;
542 /* Function to create two new TCP segments. Shrinks the given segment
543 * to the specified size and appends a new segment with the rest of the
544 * packet to the list. This won't be called frequently, I hope.
545 * Remember, these are still headerless SKBs at this point.
547 int tcp_fragment(struct sock
*sk
, struct sk_buff
*skb
, u32 len
, unsigned int mss_now
)
549 struct tcp_sock
*tp
= tcp_sk(sk
);
550 struct sk_buff
*buff
;
551 int nsize
, old_factor
;
555 BUG_ON(len
> skb
->len
);
557 clear_all_retrans_hints(tp
);
558 nsize
= skb_headlen(skb
) - len
;
562 if (skb_cloned(skb
) &&
563 skb_is_nonlinear(skb
) &&
564 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
567 /* Get a new skb... force flag on. */
568 buff
= sk_stream_alloc_skb(sk
, nsize
, GFP_ATOMIC
);
570 return -ENOMEM
; /* We'll just try again later. */
572 sk_charge_skb(sk
, buff
);
573 nlen
= skb
->len
- len
- nsize
;
574 buff
->truesize
+= nlen
;
575 skb
->truesize
-= nlen
;
577 /* Correct the sequence numbers. */
578 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
579 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
580 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
582 /* PSH and FIN should only be set in the second packet. */
583 flags
= TCP_SKB_CB(skb
)->flags
;
584 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
585 TCP_SKB_CB(buff
)->flags
= flags
;
586 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
587 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_AT_TAIL
;
589 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_PARTIAL
) {
590 /* Copy and checksum data tail into the new buffer. */
591 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
, skb_put(buff
, nsize
),
596 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
598 skb
->ip_summed
= CHECKSUM_PARTIAL
;
599 skb_split(skb
, buff
, len
);
602 buff
->ip_summed
= skb
->ip_summed
;
604 /* Looks stupid, but our code really uses when of
605 * skbs, which it never sent before. --ANK
607 TCP_SKB_CB(buff
)->when
= TCP_SKB_CB(skb
)->when
;
608 buff
->tstamp
= skb
->tstamp
;
610 old_factor
= tcp_skb_pcount(skb
);
612 /* Fix up tso_factor for both original and new SKB. */
613 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
614 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
616 /* If this packet has been sent out already, we must
617 * adjust the various packet counters.
619 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
620 int diff
= old_factor
- tcp_skb_pcount(skb
) -
621 tcp_skb_pcount(buff
);
623 tp
->packets_out
-= diff
;
625 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
626 tp
->sacked_out
-= diff
;
627 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
628 tp
->retrans_out
-= diff
;
630 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
) {
631 tp
->lost_out
-= diff
;
632 tp
->left_out
-= diff
;
636 /* Adjust Reno SACK estimate. */
637 if (!tp
->rx_opt
.sack_ok
) {
638 tp
->sacked_out
-= diff
;
639 if ((int)tp
->sacked_out
< 0)
641 tcp_sync_left_out(tp
);
644 tp
->fackets_out
-= diff
;
645 if ((int)tp
->fackets_out
< 0)
650 /* Link BUFF into the send queue. */
651 skb_header_release(buff
);
652 __skb_append(skb
, buff
, &sk
->sk_write_queue
);
657 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
658 * eventually). The difference is that pulled data not copied, but
659 * immediately discarded.
661 static void __pskb_trim_head(struct sk_buff
*skb
, int len
)
667 for (i
=0; i
<skb_shinfo(skb
)->nr_frags
; i
++) {
668 if (skb_shinfo(skb
)->frags
[i
].size
<= eat
) {
669 put_page(skb_shinfo(skb
)->frags
[i
].page
);
670 eat
-= skb_shinfo(skb
)->frags
[i
].size
;
672 skb_shinfo(skb
)->frags
[k
] = skb_shinfo(skb
)->frags
[i
];
674 skb_shinfo(skb
)->frags
[k
].page_offset
+= eat
;
675 skb_shinfo(skb
)->frags
[k
].size
-= eat
;
681 skb_shinfo(skb
)->nr_frags
= k
;
683 skb
->tail
= skb
->data
;
684 skb
->data_len
-= len
;
685 skb
->len
= skb
->data_len
;
688 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
690 if (skb_cloned(skb
) &&
691 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
694 /* If len == headlen, we avoid __skb_pull to preserve alignment. */
695 if (unlikely(len
< skb_headlen(skb
)))
696 __skb_pull(skb
, len
);
698 __pskb_trim_head(skb
, len
- skb_headlen(skb
));
700 TCP_SKB_CB(skb
)->seq
+= len
;
701 skb
->ip_summed
= CHECKSUM_PARTIAL
;
703 skb
->truesize
-= len
;
704 sk
->sk_wmem_queued
-= len
;
705 sk
->sk_forward_alloc
+= len
;
706 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
708 /* Any change of skb->len requires recalculation of tso
711 if (tcp_skb_pcount(skb
) > 1)
712 tcp_set_skb_tso_segs(sk
, skb
, tcp_current_mss(sk
, 1));
717 /* Not accounting for SACKs here. */
718 int tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
720 struct tcp_sock
*tp
= tcp_sk(sk
);
721 struct inet_connection_sock
*icsk
= inet_csk(sk
);
724 /* Calculate base mss without TCP options:
725 It is MMS_S - sizeof(tcphdr) of rfc1122
727 mss_now
= pmtu
- icsk
->icsk_af_ops
->net_header_len
- sizeof(struct tcphdr
);
729 /* Clamp it (mss_clamp does not include tcp options) */
730 if (mss_now
> tp
->rx_opt
.mss_clamp
)
731 mss_now
= tp
->rx_opt
.mss_clamp
;
733 /* Now subtract optional transport overhead */
734 mss_now
-= icsk
->icsk_ext_hdr_len
;
736 /* Then reserve room for full set of TCP options and 8 bytes of data */
740 /* Now subtract TCP options size, not including SACKs */
741 mss_now
-= tp
->tcp_header_len
- sizeof(struct tcphdr
);
746 /* Inverse of above */
747 int tcp_mss_to_mtu(struct sock
*sk
, int mss
)
749 struct tcp_sock
*tp
= tcp_sk(sk
);
750 struct inet_connection_sock
*icsk
= inet_csk(sk
);
755 icsk
->icsk_ext_hdr_len
+
756 icsk
->icsk_af_ops
->net_header_len
;
761 void tcp_mtup_init(struct sock
*sk
)
763 struct tcp_sock
*tp
= tcp_sk(sk
);
764 struct inet_connection_sock
*icsk
= inet_csk(sk
);
766 icsk
->icsk_mtup
.enabled
= sysctl_tcp_mtu_probing
> 1;
767 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+ sizeof(struct tcphdr
) +
768 icsk
->icsk_af_ops
->net_header_len
;
769 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, sysctl_tcp_base_mss
);
770 icsk
->icsk_mtup
.probe_size
= 0;
773 /* This function synchronize snd mss to current pmtu/exthdr set.
775 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
776 for TCP options, but includes only bare TCP header.
778 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
779 It is minimum of user_mss and mss received with SYN.
780 It also does not include TCP options.
782 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
784 tp->mss_cache is current effective sending mss, including
785 all tcp options except for SACKs. It is evaluated,
786 taking into account current pmtu, but never exceeds
787 tp->rx_opt.mss_clamp.
789 NOTE1. rfc1122 clearly states that advertised MSS
790 DOES NOT include either tcp or ip options.
792 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
793 are READ ONLY outside this function. --ANK (980731)
796 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
798 struct tcp_sock
*tp
= tcp_sk(sk
);
799 struct inet_connection_sock
*icsk
= inet_csk(sk
);
802 if (icsk
->icsk_mtup
.search_high
> pmtu
)
803 icsk
->icsk_mtup
.search_high
= pmtu
;
805 mss_now
= tcp_mtu_to_mss(sk
, pmtu
);
807 /* Bound mss with half of window */
808 if (tp
->max_window
&& mss_now
> (tp
->max_window
>>1))
809 mss_now
= max((tp
->max_window
>>1), 68U - tp
->tcp_header_len
);
811 /* And store cached results */
812 icsk
->icsk_pmtu_cookie
= pmtu
;
813 if (icsk
->icsk_mtup
.enabled
)
814 mss_now
= min(mss_now
, tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
));
815 tp
->mss_cache
= mss_now
;
820 /* Compute the current effective MSS, taking SACKs and IP options,
821 * and even PMTU discovery events into account.
823 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
824 * cannot be large. However, taking into account rare use of URG, this
827 unsigned int tcp_current_mss(struct sock
*sk
, int large_allowed
)
829 struct tcp_sock
*tp
= tcp_sk(sk
);
830 struct dst_entry
*dst
= __sk_dst_get(sk
);
835 mss_now
= tp
->mss_cache
;
837 if (large_allowed
&& sk_can_gso(sk
) && !tp
->urg_mode
)
841 u32 mtu
= dst_mtu(dst
);
842 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
843 mss_now
= tcp_sync_mss(sk
, mtu
);
846 if (tp
->rx_opt
.eff_sacks
)
847 mss_now
-= (TCPOLEN_SACK_BASE_ALIGNED
+
848 (tp
->rx_opt
.eff_sacks
* TCPOLEN_SACK_PERBLOCK
));
850 xmit_size_goal
= mss_now
;
853 xmit_size_goal
= (65535 -
854 inet_csk(sk
)->icsk_af_ops
->net_header_len
-
855 inet_csk(sk
)->icsk_ext_hdr_len
-
858 if (tp
->max_window
&&
859 (xmit_size_goal
> (tp
->max_window
>> 1)))
860 xmit_size_goal
= max((tp
->max_window
>> 1),
861 68U - tp
->tcp_header_len
);
863 xmit_size_goal
-= (xmit_size_goal
% mss_now
);
865 tp
->xmit_size_goal
= xmit_size_goal
;
870 /* Congestion window validation. (RFC2861) */
872 static void tcp_cwnd_validate(struct sock
*sk
, struct tcp_sock
*tp
)
874 __u32 packets_out
= tp
->packets_out
;
876 if (packets_out
>= tp
->snd_cwnd
) {
877 /* Network is feed fully. */
878 tp
->snd_cwnd_used
= 0;
879 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
881 /* Network starves. */
882 if (tp
->packets_out
> tp
->snd_cwnd_used
)
883 tp
->snd_cwnd_used
= tp
->packets_out
;
885 if ((s32
)(tcp_time_stamp
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
)
886 tcp_cwnd_application_limited(sk
);
890 static unsigned int tcp_window_allows(struct tcp_sock
*tp
, struct sk_buff
*skb
, unsigned int mss_now
, unsigned int cwnd
)
892 u32 window
, cwnd_len
;
894 window
= (tp
->snd_una
+ tp
->snd_wnd
- TCP_SKB_CB(skb
)->seq
);
895 cwnd_len
= mss_now
* cwnd
;
896 return min(window
, cwnd_len
);
899 /* Can at least one segment of SKB be sent right now, according to the
900 * congestion window rules? If so, return how many segments are allowed.
902 static inline unsigned int tcp_cwnd_test(struct tcp_sock
*tp
, struct sk_buff
*skb
)
906 /* Don't be strict about the congestion window for the final FIN. */
907 if (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
)
910 in_flight
= tcp_packets_in_flight(tp
);
912 if (in_flight
< cwnd
)
913 return (cwnd
- in_flight
);
918 /* This must be invoked the first time we consider transmitting
921 static int tcp_init_tso_segs(struct sock
*sk
, struct sk_buff
*skb
, unsigned int mss_now
)
923 int tso_segs
= tcp_skb_pcount(skb
);
927 tcp_skb_mss(skb
) != mss_now
)) {
928 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
929 tso_segs
= tcp_skb_pcount(skb
);
934 static inline int tcp_minshall_check(const struct tcp_sock
*tp
)
936 return after(tp
->snd_sml
,tp
->snd_una
) &&
937 !after(tp
->snd_sml
, tp
->snd_nxt
);
940 /* Return 0, if packet can be sent now without violation Nagle's rules:
941 * 1. It is full sized.
942 * 2. Or it contains FIN. (already checked by caller)
943 * 3. Or TCP_NODELAY was set.
944 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
945 * With Minshall's modification: all sent small packets are ACKed.
948 static inline int tcp_nagle_check(const struct tcp_sock
*tp
,
949 const struct sk_buff
*skb
,
950 unsigned mss_now
, int nonagle
)
952 return (skb
->len
< mss_now
&&
953 ((nonagle
&TCP_NAGLE_CORK
) ||
956 tcp_minshall_check(tp
))));
959 /* Return non-zero if the Nagle test allows this packet to be
962 static inline int tcp_nagle_test(struct tcp_sock
*tp
, struct sk_buff
*skb
,
963 unsigned int cur_mss
, int nonagle
)
965 /* Nagle rule does not apply to frames, which sit in the middle of the
966 * write_queue (they have no chances to get new data).
968 * This is implemented in the callers, where they modify the 'nonagle'
969 * argument based upon the location of SKB in the send queue.
971 if (nonagle
& TCP_NAGLE_PUSH
)
974 /* Don't use the nagle rule for urgent data (or for the final FIN). */
976 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
))
979 if (!tcp_nagle_check(tp
, skb
, cur_mss
, nonagle
))
985 /* Does at least the first segment of SKB fit into the send window? */
986 static inline int tcp_snd_wnd_test(struct tcp_sock
*tp
, struct sk_buff
*skb
, unsigned int cur_mss
)
988 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
990 if (skb
->len
> cur_mss
)
991 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
993 return !after(end_seq
, tp
->snd_una
+ tp
->snd_wnd
);
996 /* This checks if the data bearing packet SKB (usually sk->sk_send_head)
997 * should be put on the wire right now. If so, it returns the number of
998 * packets allowed by the congestion window.
1000 static unsigned int tcp_snd_test(struct sock
*sk
, struct sk_buff
*skb
,
1001 unsigned int cur_mss
, int nonagle
)
1003 struct tcp_sock
*tp
= tcp_sk(sk
);
1004 unsigned int cwnd_quota
;
1006 tcp_init_tso_segs(sk
, skb
, cur_mss
);
1008 if (!tcp_nagle_test(tp
, skb
, cur_mss
, nonagle
))
1011 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1013 !tcp_snd_wnd_test(tp
, skb
, cur_mss
))
1019 static inline int tcp_skb_is_last(const struct sock
*sk
,
1020 const struct sk_buff
*skb
)
1022 return skb
->next
== (struct sk_buff
*)&sk
->sk_write_queue
;
1025 int tcp_may_send_now(struct sock
*sk
, struct tcp_sock
*tp
)
1027 struct sk_buff
*skb
= sk
->sk_send_head
;
1030 tcp_snd_test(sk
, skb
, tcp_current_mss(sk
, 1),
1031 (tcp_skb_is_last(sk
, skb
) ?
1036 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1037 * which is put after SKB on the list. It is very much like
1038 * tcp_fragment() except that it may make several kinds of assumptions
1039 * in order to speed up the splitting operation. In particular, we
1040 * know that all the data is in scatter-gather pages, and that the
1041 * packet has never been sent out before (and thus is not cloned).
1043 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
, unsigned int mss_now
)
1045 struct sk_buff
*buff
;
1046 int nlen
= skb
->len
- len
;
1049 /* All of a TSO frame must be composed of paged data. */
1050 if (skb
->len
!= skb
->data_len
)
1051 return tcp_fragment(sk
, skb
, len
, mss_now
);
1053 buff
= sk_stream_alloc_pskb(sk
, 0, 0, GFP_ATOMIC
);
1054 if (unlikely(buff
== NULL
))
1057 sk_charge_skb(sk
, buff
);
1058 buff
->truesize
+= nlen
;
1059 skb
->truesize
-= nlen
;
1061 /* Correct the sequence numbers. */
1062 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1063 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1064 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1066 /* PSH and FIN should only be set in the second packet. */
1067 flags
= TCP_SKB_CB(skb
)->flags
;
1068 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
1069 TCP_SKB_CB(buff
)->flags
= flags
;
1071 /* This packet was never sent out yet, so no SACK bits. */
1072 TCP_SKB_CB(buff
)->sacked
= 0;
1074 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_PARTIAL
;
1075 skb_split(skb
, buff
, len
);
1077 /* Fix up tso_factor for both original and new SKB. */
1078 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1079 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1081 /* Link BUFF into the send queue. */
1082 skb_header_release(buff
);
1083 __skb_append(skb
, buff
, &sk
->sk_write_queue
);
1088 /* Try to defer sending, if possible, in order to minimize the amount
1089 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1091 * This algorithm is from John Heffner.
1093 static int tcp_tso_should_defer(struct sock
*sk
, struct tcp_sock
*tp
, struct sk_buff
*skb
)
1095 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1096 u32 send_win
, cong_win
, limit
, in_flight
;
1098 if (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
)
1101 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
1104 /* Defer for less than two clock ticks. */
1105 if (!tp
->tso_deferred
&& ((jiffies
<<1)>>1) - (tp
->tso_deferred
>>1) > 1)
1108 in_flight
= tcp_packets_in_flight(tp
);
1110 BUG_ON(tcp_skb_pcount(skb
) <= 1 ||
1111 (tp
->snd_cwnd
<= in_flight
));
1113 send_win
= (tp
->snd_una
+ tp
->snd_wnd
) - TCP_SKB_CB(skb
)->seq
;
1115 /* From in_flight test above, we know that cwnd > in_flight. */
1116 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
1118 limit
= min(send_win
, cong_win
);
1120 /* If a full-sized TSO skb can be sent, do it. */
1124 if (sysctl_tcp_tso_win_divisor
) {
1125 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
1127 /* If at least some fraction of a window is available,
1130 chunk
/= sysctl_tcp_tso_win_divisor
;
1134 /* Different approach, try not to defer past a single
1135 * ACK. Receiver should ACK every other full sized
1136 * frame, so if we have space for more than 3 frames
1139 if (limit
> tcp_max_burst(tp
) * tp
->mss_cache
)
1143 /* Ok, it looks like it is advisable to defer. */
1144 tp
->tso_deferred
= 1 | (jiffies
<<1);
1149 tp
->tso_deferred
= 0;
1153 /* Create a new MTU probe if we are ready.
1154 * Returns 0 if we should wait to probe (no cwnd available),
1155 * 1 if a probe was sent,
1157 static int tcp_mtu_probe(struct sock
*sk
)
1159 struct tcp_sock
*tp
= tcp_sk(sk
);
1160 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1161 struct sk_buff
*skb
, *nskb
, *next
;
1168 /* Not currently probing/verifying,
1170 * have enough cwnd, and
1171 * not SACKing (the variable headers throw things off) */
1172 if (!icsk
->icsk_mtup
.enabled
||
1173 icsk
->icsk_mtup
.probe_size
||
1174 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
||
1175 tp
->snd_cwnd
< 11 ||
1176 tp
->rx_opt
.eff_sacks
)
1179 /* Very simple search strategy: just double the MSS. */
1180 mss_now
= tcp_current_mss(sk
, 0);
1181 probe_size
= 2*tp
->mss_cache
;
1182 if (probe_size
> tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_high
)) {
1183 /* TODO: set timer for probe_converge_event */
1187 /* Have enough data in the send queue to probe? */
1189 if ((skb
= sk
->sk_send_head
) == NULL
)
1191 while ((len
+= skb
->len
) < probe_size
&& !tcp_skb_is_last(sk
, skb
))
1193 if (len
< probe_size
)
1196 /* Receive window check. */
1197 if (after(TCP_SKB_CB(skb
)->seq
+ probe_size
, tp
->snd_una
+ tp
->snd_wnd
)) {
1198 if (tp
->snd_wnd
< probe_size
)
1204 /* Do we need to wait to drain cwnd? */
1205 pif
= tcp_packets_in_flight(tp
);
1206 if (pif
+ 2 > tp
->snd_cwnd
) {
1207 /* With no packets in flight, don't stall. */
1214 /* We're allowed to probe. Build it now. */
1215 if ((nskb
= sk_stream_alloc_skb(sk
, probe_size
, GFP_ATOMIC
)) == NULL
)
1217 sk_charge_skb(sk
, nskb
);
1219 skb
= sk
->sk_send_head
;
1220 __skb_insert(nskb
, skb
->prev
, skb
, &sk
->sk_write_queue
);
1221 sk
->sk_send_head
= nskb
;
1223 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(skb
)->seq
;
1224 TCP_SKB_CB(nskb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ probe_size
;
1225 TCP_SKB_CB(nskb
)->flags
= TCPCB_FLAG_ACK
;
1226 TCP_SKB_CB(nskb
)->sacked
= 0;
1228 nskb
->ip_summed
= skb
->ip_summed
;
1231 while (len
< probe_size
) {
1234 copy
= min_t(int, skb
->len
, probe_size
- len
);
1235 if (nskb
->ip_summed
)
1236 skb_copy_bits(skb
, 0, skb_put(nskb
, copy
), copy
);
1238 nskb
->csum
= skb_copy_and_csum_bits(skb
, 0,
1239 skb_put(nskb
, copy
), copy
, nskb
->csum
);
1241 if (skb
->len
<= copy
) {
1242 /* We've eaten all the data from this skb.
1244 TCP_SKB_CB(nskb
)->flags
|= TCP_SKB_CB(skb
)->flags
;
1245 __skb_unlink(skb
, &sk
->sk_write_queue
);
1246 sk_stream_free_skb(sk
, skb
);
1248 TCP_SKB_CB(nskb
)->flags
|= TCP_SKB_CB(skb
)->flags
&
1249 ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
1250 if (!skb_shinfo(skb
)->nr_frags
) {
1251 skb_pull(skb
, copy
);
1252 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1253 skb
->csum
= csum_partial(skb
->data
, skb
->len
, 0);
1255 __pskb_trim_head(skb
, copy
);
1256 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1258 TCP_SKB_CB(skb
)->seq
+= copy
;
1264 tcp_init_tso_segs(sk
, nskb
, nskb
->len
);
1266 /* We're ready to send. If this fails, the probe will
1267 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1268 TCP_SKB_CB(nskb
)->when
= tcp_time_stamp
;
1269 if (!tcp_transmit_skb(sk
, nskb
, 1, GFP_ATOMIC
)) {
1270 /* Decrement cwnd here because we are sending
1271 * effectively two packets. */
1273 update_send_head(sk
, tp
, nskb
);
1275 icsk
->icsk_mtup
.probe_size
= tcp_mss_to_mtu(sk
, nskb
->len
);
1276 tp
->mtu_probe
.probe_seq_start
= TCP_SKB_CB(nskb
)->seq
;
1277 tp
->mtu_probe
.probe_seq_end
= TCP_SKB_CB(nskb
)->end_seq
;
1286 /* This routine writes packets to the network. It advances the
1287 * send_head. This happens as incoming acks open up the remote
1290 * Returns 1, if no segments are in flight and we have queued segments, but
1291 * cannot send anything now because of SWS or another problem.
1293 static int tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
)
1295 struct tcp_sock
*tp
= tcp_sk(sk
);
1296 struct sk_buff
*skb
;
1297 unsigned int tso_segs
, sent_pkts
;
1301 /* If we are closed, the bytes will have to remain here.
1302 * In time closedown will finish, we empty the write queue and all
1305 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
1310 /* Do MTU probing. */
1311 if ((result
= tcp_mtu_probe(sk
)) == 0) {
1313 } else if (result
> 0) {
1317 while ((skb
= sk
->sk_send_head
)) {
1320 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1323 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1327 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
)))
1330 if (tso_segs
== 1) {
1331 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
1332 (tcp_skb_is_last(sk
, skb
) ?
1333 nonagle
: TCP_NAGLE_PUSH
))))
1336 if (tcp_tso_should_defer(sk
, tp
, skb
))
1342 limit
= tcp_window_allows(tp
, skb
,
1343 mss_now
, cwnd_quota
);
1345 if (skb
->len
< limit
) {
1346 unsigned int trim
= skb
->len
% mss_now
;
1349 limit
= skb
->len
- trim
;
1353 if (skb
->len
> limit
&&
1354 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
)))
1357 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1359 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
)))
1362 /* Advance the send_head. This one is sent out.
1363 * This call will increment packets_out.
1365 update_send_head(sk
, tp
, skb
);
1367 tcp_minshall_update(tp
, mss_now
, skb
);
1371 if (likely(sent_pkts
)) {
1372 tcp_cwnd_validate(sk
, tp
);
1375 return !tp
->packets_out
&& sk
->sk_send_head
;
1378 /* Push out any pending frames which were held back due to
1379 * TCP_CORK or attempt at coalescing tiny packets.
1380 * The socket must be locked by the caller.
1382 void __tcp_push_pending_frames(struct sock
*sk
, struct tcp_sock
*tp
,
1383 unsigned int cur_mss
, int nonagle
)
1385 struct sk_buff
*skb
= sk
->sk_send_head
;
1388 if (tcp_write_xmit(sk
, cur_mss
, nonagle
))
1389 tcp_check_probe_timer(sk
, tp
);
1393 /* Send _single_ skb sitting at the send head. This function requires
1394 * true push pending frames to setup probe timer etc.
1396 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
1398 struct tcp_sock
*tp
= tcp_sk(sk
);
1399 struct sk_buff
*skb
= sk
->sk_send_head
;
1400 unsigned int tso_segs
, cwnd_quota
;
1402 BUG_ON(!skb
|| skb
->len
< mss_now
);
1404 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1405 cwnd_quota
= tcp_snd_test(sk
, skb
, mss_now
, TCP_NAGLE_PUSH
);
1407 if (likely(cwnd_quota
)) {
1414 limit
= tcp_window_allows(tp
, skb
,
1415 mss_now
, cwnd_quota
);
1417 if (skb
->len
< limit
) {
1418 unsigned int trim
= skb
->len
% mss_now
;
1421 limit
= skb
->len
- trim
;
1425 if (skb
->len
> limit
&&
1426 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
)))
1429 /* Send it out now. */
1430 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1432 if (likely(!tcp_transmit_skb(sk
, skb
, 1, sk
->sk_allocation
))) {
1433 update_send_head(sk
, tp
, skb
);
1434 tcp_cwnd_validate(sk
, tp
);
1440 /* This function returns the amount that we can raise the
1441 * usable window based on the following constraints
1443 * 1. The window can never be shrunk once it is offered (RFC 793)
1444 * 2. We limit memory per socket
1447 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1448 * RECV.NEXT + RCV.WIN fixed until:
1449 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1451 * i.e. don't raise the right edge of the window until you can raise
1452 * it at least MSS bytes.
1454 * Unfortunately, the recommended algorithm breaks header prediction,
1455 * since header prediction assumes th->window stays fixed.
1457 * Strictly speaking, keeping th->window fixed violates the receiver
1458 * side SWS prevention criteria. The problem is that under this rule
1459 * a stream of single byte packets will cause the right side of the
1460 * window to always advance by a single byte.
1462 * Of course, if the sender implements sender side SWS prevention
1463 * then this will not be a problem.
1465 * BSD seems to make the following compromise:
1467 * If the free space is less than the 1/4 of the maximum
1468 * space available and the free space is less than 1/2 mss,
1469 * then set the window to 0.
1470 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1471 * Otherwise, just prevent the window from shrinking
1472 * and from being larger than the largest representable value.
1474 * This prevents incremental opening of the window in the regime
1475 * where TCP is limited by the speed of the reader side taking
1476 * data out of the TCP receive queue. It does nothing about
1477 * those cases where the window is constrained on the sender side
1478 * because the pipeline is full.
1480 * BSD also seems to "accidentally" limit itself to windows that are a
1481 * multiple of MSS, at least until the free space gets quite small.
1482 * This would appear to be a side effect of the mbuf implementation.
1483 * Combining these two algorithms results in the observed behavior
1484 * of having a fixed window size at almost all times.
1486 * Below we obtain similar behavior by forcing the offered window to
1487 * a multiple of the mss when it is feasible to do so.
1489 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1490 * Regular options like TIMESTAMP are taken into account.
1492 u32
__tcp_select_window(struct sock
*sk
)
1494 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1495 struct tcp_sock
*tp
= tcp_sk(sk
);
1496 /* MSS for the peer's data. Previous versions used mss_clamp
1497 * here. I don't know if the value based on our guesses
1498 * of peer's MSS is better for the performance. It's more correct
1499 * but may be worse for the performance because of rcv_mss
1500 * fluctuations. --SAW 1998/11/1
1502 int mss
= icsk
->icsk_ack
.rcv_mss
;
1503 int free_space
= tcp_space(sk
);
1504 int full_space
= min_t(int, tp
->window_clamp
, tcp_full_space(sk
));
1507 if (mss
> full_space
)
1510 if (free_space
< full_space
/2) {
1511 icsk
->icsk_ack
.quick
= 0;
1513 if (tcp_memory_pressure
)
1514 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U*tp
->advmss
);
1516 if (free_space
< mss
)
1520 if (free_space
> tp
->rcv_ssthresh
)
1521 free_space
= tp
->rcv_ssthresh
;
1523 /* Don't do rounding if we are using window scaling, since the
1524 * scaled window will not line up with the MSS boundary anyway.
1526 window
= tp
->rcv_wnd
;
1527 if (tp
->rx_opt
.rcv_wscale
) {
1528 window
= free_space
;
1530 /* Advertise enough space so that it won't get scaled away.
1531 * Import case: prevent zero window announcement if
1532 * 1<<rcv_wscale > mss.
1534 if (((window
>> tp
->rx_opt
.rcv_wscale
) << tp
->rx_opt
.rcv_wscale
) != window
)
1535 window
= (((window
>> tp
->rx_opt
.rcv_wscale
) + 1)
1536 << tp
->rx_opt
.rcv_wscale
);
1538 /* Get the largest window that is a nice multiple of mss.
1539 * Window clamp already applied above.
1540 * If our current window offering is within 1 mss of the
1541 * free space we just keep it. This prevents the divide
1542 * and multiply from happening most of the time.
1543 * We also don't do any window rounding when the free space
1546 if (window
<= free_space
- mss
|| window
> free_space
)
1547 window
= (free_space
/mss
)*mss
;
1553 /* Attempt to collapse two adjacent SKB's during retransmission. */
1554 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*skb
, int mss_now
)
1556 struct tcp_sock
*tp
= tcp_sk(sk
);
1557 struct sk_buff
*next_skb
= skb
->next
;
1559 /* The first test we must make is that neither of these two
1560 * SKB's are still referenced by someone else.
1562 if (!skb_cloned(skb
) && !skb_cloned(next_skb
)) {
1563 int skb_size
= skb
->len
, next_skb_size
= next_skb
->len
;
1564 u16 flags
= TCP_SKB_CB(skb
)->flags
;
1566 /* Also punt if next skb has been SACK'd. */
1567 if(TCP_SKB_CB(next_skb
)->sacked
& TCPCB_SACKED_ACKED
)
1570 /* Next skb is out of window. */
1571 if (after(TCP_SKB_CB(next_skb
)->end_seq
, tp
->snd_una
+tp
->snd_wnd
))
1574 /* Punt if not enough space exists in the first SKB for
1575 * the data in the second, or the total combined payload
1576 * would exceed the MSS.
1578 if ((next_skb_size
> skb_tailroom(skb
)) ||
1579 ((skb_size
+ next_skb_size
) > mss_now
))
1582 BUG_ON(tcp_skb_pcount(skb
) != 1 ||
1583 tcp_skb_pcount(next_skb
) != 1);
1585 /* changing transmit queue under us so clear hints */
1586 clear_all_retrans_hints(tp
);
1588 /* Ok. We will be able to collapse the packet. */
1589 __skb_unlink(next_skb
, &sk
->sk_write_queue
);
1591 memcpy(skb_put(skb
, next_skb_size
), next_skb
->data
, next_skb_size
);
1593 skb
->ip_summed
= next_skb
->ip_summed
;
1595 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1596 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
1598 /* Update sequence range on original skb. */
1599 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
1601 /* Merge over control information. */
1602 flags
|= TCP_SKB_CB(next_skb
)->flags
; /* This moves PSH/FIN etc. over */
1603 TCP_SKB_CB(skb
)->flags
= flags
;
1605 /* All done, get rid of second SKB and account for it so
1606 * packet counting does not break.
1608 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
&(TCPCB_EVER_RETRANS
|TCPCB_AT_TAIL
);
1609 if (TCP_SKB_CB(next_skb
)->sacked
&TCPCB_SACKED_RETRANS
)
1610 tp
->retrans_out
-= tcp_skb_pcount(next_skb
);
1611 if (TCP_SKB_CB(next_skb
)->sacked
&TCPCB_LOST
) {
1612 tp
->lost_out
-= tcp_skb_pcount(next_skb
);
1613 tp
->left_out
-= tcp_skb_pcount(next_skb
);
1615 /* Reno case is special. Sigh... */
1616 if (!tp
->rx_opt
.sack_ok
&& tp
->sacked_out
) {
1617 tcp_dec_pcount_approx(&tp
->sacked_out
, next_skb
);
1618 tp
->left_out
-= tcp_skb_pcount(next_skb
);
1621 /* Not quite right: it can be > snd.fack, but
1622 * it is better to underestimate fackets.
1624 tcp_dec_pcount_approx(&tp
->fackets_out
, next_skb
);
1625 tcp_packets_out_dec(tp
, next_skb
);
1626 sk_stream_free_skb(sk
, next_skb
);
1630 /* Do a simple retransmit without using the backoff mechanisms in
1631 * tcp_timer. This is used for path mtu discovery.
1632 * The socket is already locked here.
1634 void tcp_simple_retransmit(struct sock
*sk
)
1636 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1637 struct tcp_sock
*tp
= tcp_sk(sk
);
1638 struct sk_buff
*skb
;
1639 unsigned int mss
= tcp_current_mss(sk
, 0);
1642 sk_stream_for_retrans_queue(skb
, sk
) {
1643 if (skb
->len
> mss
&&
1644 !(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
)) {
1645 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) {
1646 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1647 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1649 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_LOST
)) {
1650 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1651 tp
->lost_out
+= tcp_skb_pcount(skb
);
1657 clear_all_retrans_hints(tp
);
1662 tcp_sync_left_out(tp
);
1664 /* Don't muck with the congestion window here.
1665 * Reason is that we do not increase amount of _data_
1666 * in network, but units changed and effective
1667 * cwnd/ssthresh really reduced now.
1669 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
1670 tp
->high_seq
= tp
->snd_nxt
;
1671 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
1672 tp
->prior_ssthresh
= 0;
1673 tp
->undo_marker
= 0;
1674 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1676 tcp_xmit_retransmit_queue(sk
);
1679 /* This retransmits one SKB. Policy decisions and retransmit queue
1680 * state updates are done by the caller. Returns non-zero if an
1681 * error occurred which prevented the send.
1683 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
1685 struct tcp_sock
*tp
= tcp_sk(sk
);
1686 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1687 unsigned int cur_mss
= tcp_current_mss(sk
, 0);
1690 /* Inconslusive MTU probe */
1691 if (icsk
->icsk_mtup
.probe_size
) {
1692 icsk
->icsk_mtup
.probe_size
= 0;
1695 /* Do not sent more than we queued. 1/4 is reserved for possible
1696 * copying overhead: fragmentation, tunneling, mangling etc.
1698 if (atomic_read(&sk
->sk_wmem_alloc
) >
1699 min(sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2), sk
->sk_sndbuf
))
1702 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
1703 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1705 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
1709 /* If receiver has shrunk his window, and skb is out of
1710 * new window, do not retransmit it. The exception is the
1711 * case, when window is shrunk to zero. In this case
1712 * our retransmit serves as a zero window probe.
1714 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
+tp
->snd_wnd
)
1715 && TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
1718 if (skb
->len
> cur_mss
) {
1719 if (tcp_fragment(sk
, skb
, cur_mss
, cur_mss
))
1720 return -ENOMEM
; /* We'll try again later. */
1723 /* Collapse two adjacent packets if worthwhile and we can. */
1724 if(!(TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_SYN
) &&
1725 (skb
->len
< (cur_mss
>> 1)) &&
1726 (skb
->next
!= sk
->sk_send_head
) &&
1727 (skb
->next
!= (struct sk_buff
*)&sk
->sk_write_queue
) &&
1728 (skb_shinfo(skb
)->nr_frags
== 0 && skb_shinfo(skb
->next
)->nr_frags
== 0) &&
1729 (tcp_skb_pcount(skb
) == 1 && tcp_skb_pcount(skb
->next
) == 1) &&
1730 (sysctl_tcp_retrans_collapse
!= 0))
1731 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
1733 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
1734 return -EHOSTUNREACH
; /* Routing failure or similar. */
1736 /* Some Solaris stacks overoptimize and ignore the FIN on a
1737 * retransmit when old data is attached. So strip it off
1738 * since it is cheap to do so and saves bytes on the network.
1741 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
) &&
1742 tp
->snd_una
== (TCP_SKB_CB(skb
)->end_seq
- 1)) {
1743 if (!pskb_trim(skb
, 0)) {
1744 TCP_SKB_CB(skb
)->seq
= TCP_SKB_CB(skb
)->end_seq
- 1;
1745 skb_shinfo(skb
)->gso_segs
= 1;
1746 skb_shinfo(skb
)->gso_size
= 0;
1747 skb_shinfo(skb
)->gso_type
= 0;
1748 skb
->ip_summed
= CHECKSUM_NONE
;
1753 /* Make a copy, if the first transmission SKB clone we made
1754 * is still in somebody's hands, else make a clone.
1756 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1758 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
1761 /* Update global TCP statistics. */
1762 TCP_INC_STATS(TCP_MIB_RETRANSSEGS
);
1764 tp
->total_retrans
++;
1766 #if FASTRETRANS_DEBUG > 0
1767 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) {
1768 if (net_ratelimit())
1769 printk(KERN_DEBUG
"retrans_out leaked.\n");
1772 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
1773 tp
->retrans_out
+= tcp_skb_pcount(skb
);
1775 /* Save stamp of the first retransmit. */
1776 if (!tp
->retrans_stamp
)
1777 tp
->retrans_stamp
= TCP_SKB_CB(skb
)->when
;
1781 /* snd_nxt is stored to detect loss of retransmitted segment,
1782 * see tcp_input.c tcp_sacktag_write_queue().
1784 TCP_SKB_CB(skb
)->ack_seq
= tp
->snd_nxt
;
1789 /* This gets called after a retransmit timeout, and the initially
1790 * retransmitted data is acknowledged. It tries to continue
1791 * resending the rest of the retransmit queue, until either
1792 * we've sent it all or the congestion window limit is reached.
1793 * If doing SACK, the first ACK which comes back for a timeout
1794 * based retransmit packet might feed us FACK information again.
1795 * If so, we use it to avoid unnecessarily retransmissions.
1797 void tcp_xmit_retransmit_queue(struct sock
*sk
)
1799 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1800 struct tcp_sock
*tp
= tcp_sk(sk
);
1801 struct sk_buff
*skb
;
1804 if (tp
->retransmit_skb_hint
) {
1805 skb
= tp
->retransmit_skb_hint
;
1806 packet_cnt
= tp
->retransmit_cnt_hint
;
1808 skb
= sk
->sk_write_queue
.next
;
1812 /* First pass: retransmit lost packets. */
1814 sk_stream_for_retrans_queue_from(skb
, sk
) {
1815 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
1817 /* we could do better than to assign each time */
1818 tp
->retransmit_skb_hint
= skb
;
1819 tp
->retransmit_cnt_hint
= packet_cnt
;
1821 /* Assume this retransmit will generate
1822 * only one packet for congestion window
1823 * calculation purposes. This works because
1824 * tcp_retransmit_skb() will chop up the
1825 * packet to be MSS sized and all the
1826 * packet counting works out.
1828 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
1831 if (sacked
& TCPCB_LOST
) {
1832 if (!(sacked
&(TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))) {
1833 if (tcp_retransmit_skb(sk
, skb
)) {
1834 tp
->retransmit_skb_hint
= NULL
;
1837 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
1838 NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS
);
1840 NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS
);
1843 skb_peek(&sk
->sk_write_queue
))
1844 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1845 inet_csk(sk
)->icsk_rto
,
1849 packet_cnt
+= tcp_skb_pcount(skb
);
1850 if (packet_cnt
>= tp
->lost_out
)
1856 /* OK, demanded retransmission is finished. */
1858 /* Forward retransmissions are possible only during Recovery. */
1859 if (icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
1862 /* No forward retransmissions in Reno are possible. */
1863 if (!tp
->rx_opt
.sack_ok
)
1866 /* Yeah, we have to make difficult choice between forward transmission
1867 * and retransmission... Both ways have their merits...
1869 * For now we do not retransmit anything, while we have some new
1873 if (tcp_may_send_now(sk
, tp
))
1876 if (tp
->forward_skb_hint
) {
1877 skb
= tp
->forward_skb_hint
;
1878 packet_cnt
= tp
->forward_cnt_hint
;
1880 skb
= sk
->sk_write_queue
.next
;
1884 sk_stream_for_retrans_queue_from(skb
, sk
) {
1885 tp
->forward_cnt_hint
= packet_cnt
;
1886 tp
->forward_skb_hint
= skb
;
1888 /* Similar to the retransmit loop above we
1889 * can pretend that the retransmitted SKB
1890 * we send out here will be composed of one
1891 * real MSS sized packet because tcp_retransmit_skb()
1892 * will fragment it if necessary.
1894 if (++packet_cnt
> tp
->fackets_out
)
1897 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
1900 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
)
1903 /* Ok, retransmit it. */
1904 if (tcp_retransmit_skb(sk
, skb
)) {
1905 tp
->forward_skb_hint
= NULL
;
1909 if (skb
== skb_peek(&sk
->sk_write_queue
))
1910 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1911 inet_csk(sk
)->icsk_rto
,
1914 NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS
);
1919 /* Send a fin. The caller locks the socket for us. This cannot be
1920 * allowed to fail queueing a FIN frame under any circumstances.
1922 void tcp_send_fin(struct sock
*sk
)
1924 struct tcp_sock
*tp
= tcp_sk(sk
);
1925 struct sk_buff
*skb
= skb_peek_tail(&sk
->sk_write_queue
);
1928 /* Optimization, tack on the FIN if we have a queue of
1929 * unsent frames. But be careful about outgoing SACKS
1932 mss_now
= tcp_current_mss(sk
, 1);
1934 if (sk
->sk_send_head
!= NULL
) {
1935 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_FIN
;
1936 TCP_SKB_CB(skb
)->end_seq
++;
1939 /* Socket is locked, keep trying until memory is available. */
1941 skb
= alloc_skb_fclone(MAX_TCP_HEADER
, GFP_KERNEL
);
1947 /* Reserve space for headers and prepare control bits. */
1948 skb_reserve(skb
, MAX_TCP_HEADER
);
1950 TCP_SKB_CB(skb
)->flags
= (TCPCB_FLAG_ACK
| TCPCB_FLAG_FIN
);
1951 TCP_SKB_CB(skb
)->sacked
= 0;
1952 skb_shinfo(skb
)->gso_segs
= 1;
1953 skb_shinfo(skb
)->gso_size
= 0;
1954 skb_shinfo(skb
)->gso_type
= 0;
1956 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
1957 TCP_SKB_CB(skb
)->seq
= tp
->write_seq
;
1958 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ 1;
1959 tcp_queue_skb(sk
, skb
);
1961 __tcp_push_pending_frames(sk
, tp
, mss_now
, TCP_NAGLE_OFF
);
1964 /* We get here when a process closes a file descriptor (either due to
1965 * an explicit close() or as a byproduct of exit()'ing) and there
1966 * was unread data in the receive queue. This behavior is recommended
1967 * by draft-ietf-tcpimpl-prob-03.txt section 3.10. -DaveM
1969 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
1971 struct tcp_sock
*tp
= tcp_sk(sk
);
1972 struct sk_buff
*skb
;
1974 /* NOTE: No TCP options attached and we never retransmit this. */
1975 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
1977 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED
);
1981 /* Reserve space for headers and prepare control bits. */
1982 skb_reserve(skb
, MAX_TCP_HEADER
);
1984 TCP_SKB_CB(skb
)->flags
= (TCPCB_FLAG_ACK
| TCPCB_FLAG_RST
);
1985 TCP_SKB_CB(skb
)->sacked
= 0;
1986 skb_shinfo(skb
)->gso_segs
= 1;
1987 skb_shinfo(skb
)->gso_size
= 0;
1988 skb_shinfo(skb
)->gso_type
= 0;
1991 TCP_SKB_CB(skb
)->seq
= tcp_acceptable_seq(sk
, tp
);
1992 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
;
1993 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1994 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
1995 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED
);
1998 /* WARNING: This routine must only be called when we have already sent
1999 * a SYN packet that crossed the incoming SYN that caused this routine
2000 * to get called. If this assumption fails then the initial rcv_wnd
2001 * and rcv_wscale values will not be correct.
2003 int tcp_send_synack(struct sock
*sk
)
2005 struct sk_buff
* skb
;
2007 skb
= skb_peek(&sk
->sk_write_queue
);
2008 if (skb
== NULL
|| !(TCP_SKB_CB(skb
)->flags
&TCPCB_FLAG_SYN
)) {
2009 printk(KERN_DEBUG
"tcp_send_synack: wrong queue state\n");
2012 if (!(TCP_SKB_CB(skb
)->flags
&TCPCB_FLAG_ACK
)) {
2013 if (skb_cloned(skb
)) {
2014 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
2017 __skb_unlink(skb
, &sk
->sk_write_queue
);
2018 skb_header_release(nskb
);
2019 __skb_queue_head(&sk
->sk_write_queue
, nskb
);
2020 sk_stream_free_skb(sk
, skb
);
2021 sk_charge_skb(sk
, nskb
);
2025 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_ACK
;
2026 TCP_ECN_send_synack(tcp_sk(sk
), skb
);
2028 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2029 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2033 * Prepare a SYN-ACK.
2035 struct sk_buff
* tcp_make_synack(struct sock
*sk
, struct dst_entry
*dst
,
2036 struct request_sock
*req
)
2038 struct inet_request_sock
*ireq
= inet_rsk(req
);
2039 struct tcp_sock
*tp
= tcp_sk(sk
);
2041 int tcp_header_size
;
2042 struct sk_buff
*skb
;
2044 skb
= sock_wmalloc(sk
, MAX_TCP_HEADER
+ 15, 1, GFP_ATOMIC
);
2048 /* Reserve space for headers. */
2049 skb_reserve(skb
, MAX_TCP_HEADER
);
2051 skb
->dst
= dst_clone(dst
);
2053 tcp_header_size
= (sizeof(struct tcphdr
) + TCPOLEN_MSS
+
2054 (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0) +
2055 (ireq
->wscale_ok
? TCPOLEN_WSCALE_ALIGNED
: 0) +
2056 /* SACK_PERM is in the place of NOP NOP of TS */
2057 ((ireq
->sack_ok
&& !ireq
->tstamp_ok
) ? TCPOLEN_SACKPERM_ALIGNED
: 0));
2058 skb
->h
.th
= th
= (struct tcphdr
*) skb_push(skb
, tcp_header_size
);
2060 memset(th
, 0, sizeof(struct tcphdr
));
2063 TCP_ECN_make_synack(req
, th
);
2064 th
->source
= inet_sk(sk
)->sport
;
2065 th
->dest
= ireq
->rmt_port
;
2066 TCP_SKB_CB(skb
)->seq
= tcp_rsk(req
)->snt_isn
;
2067 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ 1;
2068 TCP_SKB_CB(skb
)->sacked
= 0;
2069 skb_shinfo(skb
)->gso_segs
= 1;
2070 skb_shinfo(skb
)->gso_size
= 0;
2071 skb_shinfo(skb
)->gso_type
= 0;
2072 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
2073 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_isn
+ 1);
2074 if (req
->rcv_wnd
== 0) { /* ignored for retransmitted syns */
2076 /* Set this up on the first call only */
2077 req
->window_clamp
= tp
->window_clamp
? : dst_metric(dst
, RTAX_WINDOW
);
2078 /* tcp_full_space because it is guaranteed to be the first packet */
2079 tcp_select_initial_window(tcp_full_space(sk
),
2080 dst_metric(dst
, RTAX_ADVMSS
) - (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0),
2085 ireq
->rcv_wscale
= rcv_wscale
;
2088 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2089 th
->window
= htons(req
->rcv_wnd
);
2091 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2092 tcp_syn_build_options((__be32
*)(th
+ 1), dst_metric(dst
, RTAX_ADVMSS
), ireq
->tstamp_ok
,
2093 ireq
->sack_ok
, ireq
->wscale_ok
, ireq
->rcv_wscale
,
2094 TCP_SKB_CB(skb
)->when
,
2098 th
->doff
= (tcp_header_size
>> 2);
2099 TCP_INC_STATS(TCP_MIB_OUTSEGS
);
2104 * Do all connect socket setups that can be done AF independent.
2106 static void tcp_connect_init(struct sock
*sk
)
2108 struct dst_entry
*dst
= __sk_dst_get(sk
);
2109 struct tcp_sock
*tp
= tcp_sk(sk
);
2112 /* We'll fix this up when we get a response from the other end.
2113 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2115 tp
->tcp_header_len
= sizeof(struct tcphdr
) +
2116 (sysctl_tcp_timestamps
? TCPOLEN_TSTAMP_ALIGNED
: 0);
2118 /* If user gave his TCP_MAXSEG, record it to clamp */
2119 if (tp
->rx_opt
.user_mss
)
2120 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
2123 tcp_sync_mss(sk
, dst_mtu(dst
));
2125 if (!tp
->window_clamp
)
2126 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
2127 tp
->advmss
= dst_metric(dst
, RTAX_ADVMSS
);
2128 tcp_initialize_rcv_mss(sk
);
2130 tcp_select_initial_window(tcp_full_space(sk
),
2131 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
2134 sysctl_tcp_window_scaling
,
2137 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
2138 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
2141 sock_reset_flag(sk
, SOCK_DONE
);
2143 tcp_init_wl(tp
, tp
->write_seq
, 0);
2144 tp
->snd_una
= tp
->write_seq
;
2145 tp
->snd_sml
= tp
->write_seq
;
2150 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
2151 inet_csk(sk
)->icsk_retransmits
= 0;
2152 tcp_clear_retrans(tp
);
2156 * Build a SYN and send it off.
2158 int tcp_connect(struct sock
*sk
)
2160 struct tcp_sock
*tp
= tcp_sk(sk
);
2161 struct sk_buff
*buff
;
2163 tcp_connect_init(sk
);
2165 buff
= alloc_skb_fclone(MAX_TCP_HEADER
+ 15, sk
->sk_allocation
);
2166 if (unlikely(buff
== NULL
))
2169 /* Reserve space for headers. */
2170 skb_reserve(buff
, MAX_TCP_HEADER
);
2172 TCP_SKB_CB(buff
)->flags
= TCPCB_FLAG_SYN
;
2173 TCP_ECN_send_syn(sk
, tp
, buff
);
2174 TCP_SKB_CB(buff
)->sacked
= 0;
2175 skb_shinfo(buff
)->gso_segs
= 1;
2176 skb_shinfo(buff
)->gso_size
= 0;
2177 skb_shinfo(buff
)->gso_type
= 0;
2179 tp
->snd_nxt
= tp
->write_seq
;
2180 TCP_SKB_CB(buff
)->seq
= tp
->write_seq
++;
2181 TCP_SKB_CB(buff
)->end_seq
= tp
->write_seq
;
2184 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2185 tp
->retrans_stamp
= TCP_SKB_CB(buff
)->when
;
2186 skb_header_release(buff
);
2187 __skb_queue_tail(&sk
->sk_write_queue
, buff
);
2188 sk_charge_skb(sk
, buff
);
2189 tp
->packets_out
+= tcp_skb_pcount(buff
);
2190 tcp_transmit_skb(sk
, buff
, 1, GFP_KERNEL
);
2192 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2193 * in order to make this packet get counted in tcpOutSegs.
2195 tp
->snd_nxt
= tp
->write_seq
;
2196 tp
->pushed_seq
= tp
->write_seq
;
2197 TCP_INC_STATS(TCP_MIB_ACTIVEOPENS
);
2199 /* Timer for repeating the SYN until an answer. */
2200 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2201 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2205 /* Send out a delayed ack, the caller does the policy checking
2206 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2209 void tcp_send_delayed_ack(struct sock
*sk
)
2211 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2212 int ato
= icsk
->icsk_ack
.ato
;
2213 unsigned long timeout
;
2215 if (ato
> TCP_DELACK_MIN
) {
2216 const struct tcp_sock
*tp
= tcp_sk(sk
);
2219 if (icsk
->icsk_ack
.pingpong
|| (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
2220 max_ato
= TCP_DELACK_MAX
;
2222 /* Slow path, intersegment interval is "high". */
2224 /* If some rtt estimate is known, use it to bound delayed ack.
2225 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2229 int rtt
= max(tp
->srtt
>>3, TCP_DELACK_MIN
);
2235 ato
= min(ato
, max_ato
);
2238 /* Stay within the limit we were given */
2239 timeout
= jiffies
+ ato
;
2241 /* Use new timeout only if there wasn't a older one earlier. */
2242 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
2243 /* If delack timer was blocked or is about to expire,
2246 if (icsk
->icsk_ack
.blocked
||
2247 time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
2252 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
2253 timeout
= icsk
->icsk_ack
.timeout
;
2255 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
2256 icsk
->icsk_ack
.timeout
= timeout
;
2257 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
2260 /* This routine sends an ack and also updates the window. */
2261 void tcp_send_ack(struct sock
*sk
)
2263 /* If we have been reset, we may not send again. */
2264 if (sk
->sk_state
!= TCP_CLOSE
) {
2265 struct tcp_sock
*tp
= tcp_sk(sk
);
2266 struct sk_buff
*buff
;
2268 /* We are not putting this on the write queue, so
2269 * tcp_transmit_skb() will set the ownership to this
2272 buff
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2274 inet_csk_schedule_ack(sk
);
2275 inet_csk(sk
)->icsk_ack
.ato
= TCP_ATO_MIN
;
2276 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
2277 TCP_DELACK_MAX
, TCP_RTO_MAX
);
2281 /* Reserve space for headers and prepare control bits. */
2282 skb_reserve(buff
, MAX_TCP_HEADER
);
2284 TCP_SKB_CB(buff
)->flags
= TCPCB_FLAG_ACK
;
2285 TCP_SKB_CB(buff
)->sacked
= 0;
2286 skb_shinfo(buff
)->gso_segs
= 1;
2287 skb_shinfo(buff
)->gso_size
= 0;
2288 skb_shinfo(buff
)->gso_type
= 0;
2290 /* Send it off, this clears delayed acks for us. */
2291 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(buff
)->end_seq
= tcp_acceptable_seq(sk
, tp
);
2292 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2293 tcp_transmit_skb(sk
, buff
, 0, GFP_ATOMIC
);
2297 /* This routine sends a packet with an out of date sequence
2298 * number. It assumes the other end will try to ack it.
2300 * Question: what should we make while urgent mode?
2301 * 4.4BSD forces sending single byte of data. We cannot send
2302 * out of window data, because we have SND.NXT==SND.MAX...
2304 * Current solution: to send TWO zero-length segments in urgent mode:
2305 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2306 * out-of-date with SND.UNA-1 to probe window.
2308 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
)
2310 struct tcp_sock
*tp
= tcp_sk(sk
);
2311 struct sk_buff
*skb
;
2313 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2314 skb
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2318 /* Reserve space for headers and set control bits. */
2319 skb_reserve(skb
, MAX_TCP_HEADER
);
2321 TCP_SKB_CB(skb
)->flags
= TCPCB_FLAG_ACK
;
2322 TCP_SKB_CB(skb
)->sacked
= urgent
;
2323 skb_shinfo(skb
)->gso_segs
= 1;
2324 skb_shinfo(skb
)->gso_size
= 0;
2325 skb_shinfo(skb
)->gso_type
= 0;
2327 /* Use a previous sequence. This should cause the other
2328 * end to send an ack. Don't queue or clone SKB, just
2331 TCP_SKB_CB(skb
)->seq
= urgent
? tp
->snd_una
: tp
->snd_una
- 1;
2332 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
;
2333 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2334 return tcp_transmit_skb(sk
, skb
, 0, GFP_ATOMIC
);
2337 int tcp_write_wakeup(struct sock
*sk
)
2339 if (sk
->sk_state
!= TCP_CLOSE
) {
2340 struct tcp_sock
*tp
= tcp_sk(sk
);
2341 struct sk_buff
*skb
;
2343 if ((skb
= sk
->sk_send_head
) != NULL
&&
2344 before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
+tp
->snd_wnd
)) {
2346 unsigned int mss
= tcp_current_mss(sk
, 0);
2347 unsigned int seg_size
= tp
->snd_una
+tp
->snd_wnd
-TCP_SKB_CB(skb
)->seq
;
2349 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
2350 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
2352 /* We are probing the opening of a window
2353 * but the window size is != 0
2354 * must have been a result SWS avoidance ( sender )
2356 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
2358 seg_size
= min(seg_size
, mss
);
2359 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2360 if (tcp_fragment(sk
, skb
, seg_size
, mss
))
2362 } else if (!tcp_skb_pcount(skb
))
2363 tcp_set_skb_tso_segs(sk
, skb
, mss
);
2365 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2366 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2367 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2369 update_send_head(sk
, tp
, skb
);
2374 between(tp
->snd_up
, tp
->snd_una
+1, tp
->snd_una
+0xFFFF))
2375 tcp_xmit_probe_skb(sk
, TCPCB_URG
);
2376 return tcp_xmit_probe_skb(sk
, 0);
2382 /* A window probe timeout has occurred. If window is not closed send
2383 * a partial packet else a zero probe.
2385 void tcp_send_probe0(struct sock
*sk
)
2387 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2388 struct tcp_sock
*tp
= tcp_sk(sk
);
2391 err
= tcp_write_wakeup(sk
);
2393 if (tp
->packets_out
|| !sk
->sk_send_head
) {
2394 /* Cancel probe timer, if it is not required. */
2395 icsk
->icsk_probes_out
= 0;
2396 icsk
->icsk_backoff
= 0;
2401 if (icsk
->icsk_backoff
< sysctl_tcp_retries2
)
2402 icsk
->icsk_backoff
++;
2403 icsk
->icsk_probes_out
++;
2404 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2405 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
2408 /* If packet was not sent due to local congestion,
2409 * do not backoff and do not remember icsk_probes_out.
2410 * Let local senders to fight for local resources.
2412 * Use accumulated backoff yet.
2414 if (!icsk
->icsk_probes_out
)
2415 icsk
->icsk_probes_out
= 1;
2416 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2417 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
,
2418 TCP_RESOURCE_PROBE_INTERVAL
),
2423 EXPORT_SYMBOL(tcp_connect
);
2424 EXPORT_SYMBOL(tcp_make_synack
);
2425 EXPORT_SYMBOL(tcp_simple_retransmit
);
2426 EXPORT_SYMBOL(tcp_sync_mss
);
2427 EXPORT_SYMBOL(sysctl_tcp_tso_win_divisor
);
2428 EXPORT_SYMBOL(tcp_mtup_init
);