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(__u32
*ptr
, struct tcp_sock
*tp
,
275 if (tp
->rx_opt
.tstamp_ok
) {
276 *ptr
++ = __constant_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(__u32
*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
++ = __constant_htonl((TCPOPT_SACK_PERM
<< 24) | (TCPOLEN_SACK_PERM
<< 16) |
329 (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
);
331 *ptr
++ = __constant_htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16) |
332 (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
);
333 *ptr
++ = htonl(tstamp
); /* TSVAL */
334 *ptr
++ = htonl(ts_recent
); /* TSECR */
336 *ptr
++ = __constant_htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16) |
337 (TCPOPT_SACK_PERM
<< 8) | TCPOLEN_SACK_PERM
);
339 *ptr
++ = htonl((TCPOPT_NOP
<< 24) | (TCPOPT_WINDOW
<< 16) | (TCPOLEN_WINDOW
<< 8) | (wscale
));
342 /* This routine actually transmits TCP packets queued in by
343 * tcp_do_sendmsg(). This is used by both the initial
344 * transmission and possible later retransmissions.
345 * All SKB's seen here are completely headerless. It is our
346 * job to build the TCP header, and pass the packet down to
347 * IP so it can do the same plus pass the packet off to the
350 * We are working here with either a clone of the original
351 * SKB, or a fresh unique copy made by the retransmit engine.
353 static int tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int clone_it
, gfp_t gfp_mask
)
355 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
356 struct inet_sock
*inet
;
358 struct tcp_skb_cb
*tcb
;
364 BUG_ON(!skb
|| !tcp_skb_pcount(skb
));
366 /* If congestion control is doing timestamping, we must
367 * take such a timestamp before we potentially clone/copy.
369 if (icsk
->icsk_ca_ops
->rtt_sample
)
370 __net_timestamp(skb
);
372 if (likely(clone_it
)) {
373 if (unlikely(skb_cloned(skb
)))
374 skb
= pskb_copy(skb
, gfp_mask
);
376 skb
= skb_clone(skb
, gfp_mask
);
383 tcb
= TCP_SKB_CB(skb
);
384 tcp_header_size
= tp
->tcp_header_len
;
386 #define SYSCTL_FLAG_TSTAMPS 0x1
387 #define SYSCTL_FLAG_WSCALE 0x2
388 #define SYSCTL_FLAG_SACK 0x4
391 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
392 tcp_header_size
= sizeof(struct tcphdr
) + TCPOLEN_MSS
;
393 if(sysctl_tcp_timestamps
) {
394 tcp_header_size
+= TCPOLEN_TSTAMP_ALIGNED
;
395 sysctl_flags
|= SYSCTL_FLAG_TSTAMPS
;
397 if (sysctl_tcp_window_scaling
) {
398 tcp_header_size
+= TCPOLEN_WSCALE_ALIGNED
;
399 sysctl_flags
|= SYSCTL_FLAG_WSCALE
;
401 if (sysctl_tcp_sack
) {
402 sysctl_flags
|= SYSCTL_FLAG_SACK
;
403 if (!(sysctl_flags
& SYSCTL_FLAG_TSTAMPS
))
404 tcp_header_size
+= TCPOLEN_SACKPERM_ALIGNED
;
406 } else if (unlikely(tp
->rx_opt
.eff_sacks
)) {
407 /* A SACK is 2 pad bytes, a 2 byte header, plus
408 * 2 32-bit sequence numbers for each SACK block.
410 tcp_header_size
+= (TCPOLEN_SACK_BASE_ALIGNED
+
411 (tp
->rx_opt
.eff_sacks
*
412 TCPOLEN_SACK_PERBLOCK
));
415 if (tcp_packets_in_flight(tp
) == 0)
416 tcp_ca_event(sk
, CA_EVENT_TX_START
);
418 th
= (struct tcphdr
*) skb_push(skb
, tcp_header_size
);
420 skb_set_owner_w(skb
, sk
);
422 /* Build TCP header and checksum it. */
423 th
->source
= inet
->sport
;
424 th
->dest
= inet
->dport
;
425 th
->seq
= htonl(tcb
->seq
);
426 th
->ack_seq
= htonl(tp
->rcv_nxt
);
427 *(((__u16
*)th
) + 6) = htons(((tcp_header_size
>> 2) << 12) |
430 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
431 /* RFC1323: The window in SYN & SYN/ACK segments
434 th
->window
= htons(tp
->rcv_wnd
);
436 th
->window
= htons(tcp_select_window(sk
));
441 if (unlikely(tp
->urg_mode
&&
442 between(tp
->snd_up
, tcb
->seq
+1, tcb
->seq
+0xFFFF))) {
443 th
->urg_ptr
= htons(tp
->snd_up
-tcb
->seq
);
447 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
448 tcp_syn_build_options((__u32
*)(th
+ 1),
449 tcp_advertise_mss(sk
),
450 (sysctl_flags
& SYSCTL_FLAG_TSTAMPS
),
451 (sysctl_flags
& SYSCTL_FLAG_SACK
),
452 (sysctl_flags
& SYSCTL_FLAG_WSCALE
),
453 tp
->rx_opt
.rcv_wscale
,
455 tp
->rx_opt
.ts_recent
);
457 tcp_build_and_update_options((__u32
*)(th
+ 1),
459 TCP_ECN_send(sk
, tp
, skb
, tcp_header_size
);
462 icsk
->icsk_af_ops
->send_check(sk
, skb
->len
, skb
);
464 if (likely(tcb
->flags
& TCPCB_FLAG_ACK
))
465 tcp_event_ack_sent(sk
, tcp_skb_pcount(skb
));
467 if (skb
->len
!= tcp_header_size
)
468 tcp_event_data_sent(tp
, skb
, sk
);
470 if (after(tcb
->end_seq
, tp
->snd_nxt
) || tcb
->seq
== tcb
->end_seq
)
471 TCP_INC_STATS(TCP_MIB_OUTSEGS
);
473 err
= icsk
->icsk_af_ops
->queue_xmit(skb
, 0);
474 if (likely(err
<= 0))
479 /* NET_XMIT_CN is special. It does not guarantee,
480 * that this packet is lost. It tells that device
481 * is about to start to drop packets or already
482 * drops some packets of the same priority and
483 * invokes us to send less aggressively.
485 return err
== NET_XMIT_CN
? 0 : err
;
487 #undef SYSCTL_FLAG_TSTAMPS
488 #undef SYSCTL_FLAG_WSCALE
489 #undef SYSCTL_FLAG_SACK
493 /* This routine just queue's the buffer
495 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
496 * otherwise socket can stall.
498 static void tcp_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
500 struct tcp_sock
*tp
= tcp_sk(sk
);
502 /* Advance write_seq and place onto the write_queue. */
503 tp
->write_seq
= TCP_SKB_CB(skb
)->end_seq
;
504 skb_header_release(skb
);
505 __skb_queue_tail(&sk
->sk_write_queue
, skb
);
506 sk_charge_skb(sk
, skb
);
508 /* Queue it, remembering where we must start sending. */
509 if (sk
->sk_send_head
== NULL
)
510 sk
->sk_send_head
= skb
;
513 static void tcp_set_skb_tso_segs(struct sock
*sk
, struct sk_buff
*skb
, unsigned int mss_now
)
515 if (skb
->len
<= mss_now
|| !sk_can_gso(sk
)) {
516 /* Avoid the costly divide in the normal
519 skb_shinfo(skb
)->gso_segs
= 1;
520 skb_shinfo(skb
)->gso_size
= 0;
521 skb_shinfo(skb
)->gso_type
= 0;
525 factor
= skb
->len
+ (mss_now
- 1);
527 skb_shinfo(skb
)->gso_segs
= factor
;
528 skb_shinfo(skb
)->gso_size
= mss_now
;
529 skb_shinfo(skb
)->gso_type
= sk
->sk_gso_type
;
533 /* Function to create two new TCP segments. Shrinks the given segment
534 * to the specified size and appends a new segment with the rest of the
535 * packet to the list. This won't be called frequently, I hope.
536 * Remember, these are still headerless SKBs at this point.
538 int tcp_fragment(struct sock
*sk
, struct sk_buff
*skb
, u32 len
, unsigned int mss_now
)
540 struct tcp_sock
*tp
= tcp_sk(sk
);
541 struct sk_buff
*buff
;
542 int nsize
, old_factor
;
546 BUG_ON(len
> skb
->len
);
548 clear_all_retrans_hints(tp
);
549 nsize
= skb_headlen(skb
) - len
;
553 if (skb_cloned(skb
) &&
554 skb_is_nonlinear(skb
) &&
555 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
558 /* Get a new skb... force flag on. */
559 buff
= sk_stream_alloc_skb(sk
, nsize
, GFP_ATOMIC
);
561 return -ENOMEM
; /* We'll just try again later. */
563 sk_charge_skb(sk
, buff
);
564 nlen
= skb
->len
- len
- nsize
;
565 buff
->truesize
+= nlen
;
566 skb
->truesize
-= nlen
;
568 /* Correct the sequence numbers. */
569 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
570 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
571 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
573 /* PSH and FIN should only be set in the second packet. */
574 flags
= TCP_SKB_CB(skb
)->flags
;
575 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
576 TCP_SKB_CB(buff
)->flags
= flags
;
577 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
578 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_AT_TAIL
;
580 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_PARTIAL
) {
581 /* Copy and checksum data tail into the new buffer. */
582 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
, skb_put(buff
, nsize
),
587 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
589 skb
->ip_summed
= CHECKSUM_PARTIAL
;
590 skb_split(skb
, buff
, len
);
593 buff
->ip_summed
= skb
->ip_summed
;
595 /* Looks stupid, but our code really uses when of
596 * skbs, which it never sent before. --ANK
598 TCP_SKB_CB(buff
)->when
= TCP_SKB_CB(skb
)->when
;
599 buff
->tstamp
= skb
->tstamp
;
601 old_factor
= tcp_skb_pcount(skb
);
603 /* Fix up tso_factor for both original and new SKB. */
604 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
605 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
607 /* If this packet has been sent out already, we must
608 * adjust the various packet counters.
610 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
611 int diff
= old_factor
- tcp_skb_pcount(skb
) -
612 tcp_skb_pcount(buff
);
614 tp
->packets_out
-= diff
;
616 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
617 tp
->sacked_out
-= diff
;
618 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
619 tp
->retrans_out
-= diff
;
621 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
) {
622 tp
->lost_out
-= diff
;
623 tp
->left_out
-= diff
;
627 /* Adjust Reno SACK estimate. */
628 if (!tp
->rx_opt
.sack_ok
) {
629 tp
->sacked_out
-= diff
;
630 if ((int)tp
->sacked_out
< 0)
632 tcp_sync_left_out(tp
);
635 tp
->fackets_out
-= diff
;
636 if ((int)tp
->fackets_out
< 0)
641 /* Link BUFF into the send queue. */
642 skb_header_release(buff
);
643 __skb_append(skb
, buff
, &sk
->sk_write_queue
);
648 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
649 * eventually). The difference is that pulled data not copied, but
650 * immediately discarded.
652 static void __pskb_trim_head(struct sk_buff
*skb
, int len
)
658 for (i
=0; i
<skb_shinfo(skb
)->nr_frags
; i
++) {
659 if (skb_shinfo(skb
)->frags
[i
].size
<= eat
) {
660 put_page(skb_shinfo(skb
)->frags
[i
].page
);
661 eat
-= skb_shinfo(skb
)->frags
[i
].size
;
663 skb_shinfo(skb
)->frags
[k
] = skb_shinfo(skb
)->frags
[i
];
665 skb_shinfo(skb
)->frags
[k
].page_offset
+= eat
;
666 skb_shinfo(skb
)->frags
[k
].size
-= eat
;
672 skb_shinfo(skb
)->nr_frags
= k
;
674 skb
->tail
= skb
->data
;
675 skb
->data_len
-= len
;
676 skb
->len
= skb
->data_len
;
679 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
681 if (skb_cloned(skb
) &&
682 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
685 /* If len == headlen, we avoid __skb_pull to preserve alignment. */
686 if (unlikely(len
< skb_headlen(skb
)))
687 __skb_pull(skb
, len
);
689 __pskb_trim_head(skb
, len
- skb_headlen(skb
));
691 TCP_SKB_CB(skb
)->seq
+= len
;
692 skb
->ip_summed
= CHECKSUM_PARTIAL
;
694 skb
->truesize
-= len
;
695 sk
->sk_wmem_queued
-= len
;
696 sk
->sk_forward_alloc
+= len
;
697 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
699 /* Any change of skb->len requires recalculation of tso
702 if (tcp_skb_pcount(skb
) > 1)
703 tcp_set_skb_tso_segs(sk
, skb
, tcp_current_mss(sk
, 1));
708 /* Not accounting for SACKs here. */
709 int tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
711 struct tcp_sock
*tp
= tcp_sk(sk
);
712 struct inet_connection_sock
*icsk
= inet_csk(sk
);
715 /* Calculate base mss without TCP options:
716 It is MMS_S - sizeof(tcphdr) of rfc1122
718 mss_now
= pmtu
- icsk
->icsk_af_ops
->net_header_len
- sizeof(struct tcphdr
);
720 /* Clamp it (mss_clamp does not include tcp options) */
721 if (mss_now
> tp
->rx_opt
.mss_clamp
)
722 mss_now
= tp
->rx_opt
.mss_clamp
;
724 /* Now subtract optional transport overhead */
725 mss_now
-= icsk
->icsk_ext_hdr_len
;
727 /* Then reserve room for full set of TCP options and 8 bytes of data */
731 /* Now subtract TCP options size, not including SACKs */
732 mss_now
-= tp
->tcp_header_len
- sizeof(struct tcphdr
);
737 /* Inverse of above */
738 int tcp_mss_to_mtu(struct sock
*sk
, int mss
)
740 struct tcp_sock
*tp
= tcp_sk(sk
);
741 struct inet_connection_sock
*icsk
= inet_csk(sk
);
746 icsk
->icsk_ext_hdr_len
+
747 icsk
->icsk_af_ops
->net_header_len
;
752 void tcp_mtup_init(struct sock
*sk
)
754 struct tcp_sock
*tp
= tcp_sk(sk
);
755 struct inet_connection_sock
*icsk
= inet_csk(sk
);
757 icsk
->icsk_mtup
.enabled
= sysctl_tcp_mtu_probing
> 1;
758 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+ sizeof(struct tcphdr
) +
759 icsk
->icsk_af_ops
->net_header_len
;
760 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, sysctl_tcp_base_mss
);
761 icsk
->icsk_mtup
.probe_size
= 0;
764 /* This function synchronize snd mss to current pmtu/exthdr set.
766 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
767 for TCP options, but includes only bare TCP header.
769 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
770 It is minimum of user_mss and mss received with SYN.
771 It also does not include TCP options.
773 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
775 tp->mss_cache is current effective sending mss, including
776 all tcp options except for SACKs. It is evaluated,
777 taking into account current pmtu, but never exceeds
778 tp->rx_opt.mss_clamp.
780 NOTE1. rfc1122 clearly states that advertised MSS
781 DOES NOT include either tcp or ip options.
783 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
784 are READ ONLY outside this function. --ANK (980731)
787 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
789 struct tcp_sock
*tp
= tcp_sk(sk
);
790 struct inet_connection_sock
*icsk
= inet_csk(sk
);
793 if (icsk
->icsk_mtup
.search_high
> pmtu
)
794 icsk
->icsk_mtup
.search_high
= pmtu
;
796 mss_now
= tcp_mtu_to_mss(sk
, pmtu
);
798 /* Bound mss with half of window */
799 if (tp
->max_window
&& mss_now
> (tp
->max_window
>>1))
800 mss_now
= max((tp
->max_window
>>1), 68U - tp
->tcp_header_len
);
802 /* And store cached results */
803 icsk
->icsk_pmtu_cookie
= pmtu
;
804 if (icsk
->icsk_mtup
.enabled
)
805 mss_now
= min(mss_now
, tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
));
806 tp
->mss_cache
= mss_now
;
811 /* Compute the current effective MSS, taking SACKs and IP options,
812 * and even PMTU discovery events into account.
814 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
815 * cannot be large. However, taking into account rare use of URG, this
818 unsigned int tcp_current_mss(struct sock
*sk
, int large_allowed
)
820 struct tcp_sock
*tp
= tcp_sk(sk
);
821 struct dst_entry
*dst
= __sk_dst_get(sk
);
826 mss_now
= tp
->mss_cache
;
828 if (large_allowed
&& sk_can_gso(sk
) && !tp
->urg_mode
)
832 u32 mtu
= dst_mtu(dst
);
833 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
834 mss_now
= tcp_sync_mss(sk
, mtu
);
837 if (tp
->rx_opt
.eff_sacks
)
838 mss_now
-= (TCPOLEN_SACK_BASE_ALIGNED
+
839 (tp
->rx_opt
.eff_sacks
* TCPOLEN_SACK_PERBLOCK
));
841 xmit_size_goal
= mss_now
;
844 xmit_size_goal
= (65535 -
845 inet_csk(sk
)->icsk_af_ops
->net_header_len
-
846 inet_csk(sk
)->icsk_ext_hdr_len
-
849 if (tp
->max_window
&&
850 (xmit_size_goal
> (tp
->max_window
>> 1)))
851 xmit_size_goal
= max((tp
->max_window
>> 1),
852 68U - tp
->tcp_header_len
);
854 xmit_size_goal
-= (xmit_size_goal
% mss_now
);
856 tp
->xmit_size_goal
= xmit_size_goal
;
861 /* Congestion window validation. (RFC2861) */
863 static void tcp_cwnd_validate(struct sock
*sk
, struct tcp_sock
*tp
)
865 __u32 packets_out
= tp
->packets_out
;
867 if (packets_out
>= tp
->snd_cwnd
) {
868 /* Network is feed fully. */
869 tp
->snd_cwnd_used
= 0;
870 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
872 /* Network starves. */
873 if (tp
->packets_out
> tp
->snd_cwnd_used
)
874 tp
->snd_cwnd_used
= tp
->packets_out
;
876 if ((s32
)(tcp_time_stamp
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
)
877 tcp_cwnd_application_limited(sk
);
881 static unsigned int tcp_window_allows(struct tcp_sock
*tp
, struct sk_buff
*skb
, unsigned int mss_now
, unsigned int cwnd
)
883 u32 window
, cwnd_len
;
885 window
= (tp
->snd_una
+ tp
->snd_wnd
- TCP_SKB_CB(skb
)->seq
);
886 cwnd_len
= mss_now
* cwnd
;
887 return min(window
, cwnd_len
);
890 /* Can at least one segment of SKB be sent right now, according to the
891 * congestion window rules? If so, return how many segments are allowed.
893 static inline unsigned int tcp_cwnd_test(struct tcp_sock
*tp
, struct sk_buff
*skb
)
897 /* Don't be strict about the congestion window for the final FIN. */
898 if (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
)
901 in_flight
= tcp_packets_in_flight(tp
);
903 if (in_flight
< cwnd
)
904 return (cwnd
- in_flight
);
909 /* This must be invoked the first time we consider transmitting
912 static int tcp_init_tso_segs(struct sock
*sk
, struct sk_buff
*skb
, unsigned int mss_now
)
914 int tso_segs
= tcp_skb_pcount(skb
);
918 tcp_skb_mss(skb
) != mss_now
)) {
919 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
920 tso_segs
= tcp_skb_pcount(skb
);
925 static inline int tcp_minshall_check(const struct tcp_sock
*tp
)
927 return after(tp
->snd_sml
,tp
->snd_una
) &&
928 !after(tp
->snd_sml
, tp
->snd_nxt
);
931 /* Return 0, if packet can be sent now without violation Nagle's rules:
932 * 1. It is full sized.
933 * 2. Or it contains FIN. (already checked by caller)
934 * 3. Or TCP_NODELAY was set.
935 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
936 * With Minshall's modification: all sent small packets are ACKed.
939 static inline int tcp_nagle_check(const struct tcp_sock
*tp
,
940 const struct sk_buff
*skb
,
941 unsigned mss_now
, int nonagle
)
943 return (skb
->len
< mss_now
&&
944 ((nonagle
&TCP_NAGLE_CORK
) ||
947 tcp_minshall_check(tp
))));
950 /* Return non-zero if the Nagle test allows this packet to be
953 static inline int tcp_nagle_test(struct tcp_sock
*tp
, struct sk_buff
*skb
,
954 unsigned int cur_mss
, int nonagle
)
956 /* Nagle rule does not apply to frames, which sit in the middle of the
957 * write_queue (they have no chances to get new data).
959 * This is implemented in the callers, where they modify the 'nonagle'
960 * argument based upon the location of SKB in the send queue.
962 if (nonagle
& TCP_NAGLE_PUSH
)
965 /* Don't use the nagle rule for urgent data (or for the final FIN). */
967 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
))
970 if (!tcp_nagle_check(tp
, skb
, cur_mss
, nonagle
))
976 /* Does at least the first segment of SKB fit into the send window? */
977 static inline int tcp_snd_wnd_test(struct tcp_sock
*tp
, struct sk_buff
*skb
, unsigned int cur_mss
)
979 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
981 if (skb
->len
> cur_mss
)
982 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
984 return !after(end_seq
, tp
->snd_una
+ tp
->snd_wnd
);
987 /* This checks if the data bearing packet SKB (usually sk->sk_send_head)
988 * should be put on the wire right now. If so, it returns the number of
989 * packets allowed by the congestion window.
991 static unsigned int tcp_snd_test(struct sock
*sk
, struct sk_buff
*skb
,
992 unsigned int cur_mss
, int nonagle
)
994 struct tcp_sock
*tp
= tcp_sk(sk
);
995 unsigned int cwnd_quota
;
997 tcp_init_tso_segs(sk
, skb
, cur_mss
);
999 if (!tcp_nagle_test(tp
, skb
, cur_mss
, nonagle
))
1002 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1004 !tcp_snd_wnd_test(tp
, skb
, cur_mss
))
1010 static inline int tcp_skb_is_last(const struct sock
*sk
,
1011 const struct sk_buff
*skb
)
1013 return skb
->next
== (struct sk_buff
*)&sk
->sk_write_queue
;
1016 int tcp_may_send_now(struct sock
*sk
, struct tcp_sock
*tp
)
1018 struct sk_buff
*skb
= sk
->sk_send_head
;
1021 tcp_snd_test(sk
, skb
, tcp_current_mss(sk
, 1),
1022 (tcp_skb_is_last(sk
, skb
) ?
1027 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1028 * which is put after SKB on the list. It is very much like
1029 * tcp_fragment() except that it may make several kinds of assumptions
1030 * in order to speed up the splitting operation. In particular, we
1031 * know that all the data is in scatter-gather pages, and that the
1032 * packet has never been sent out before (and thus is not cloned).
1034 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
, unsigned int mss_now
)
1036 struct sk_buff
*buff
;
1037 int nlen
= skb
->len
- len
;
1040 /* All of a TSO frame must be composed of paged data. */
1041 if (skb
->len
!= skb
->data_len
)
1042 return tcp_fragment(sk
, skb
, len
, mss_now
);
1044 buff
= sk_stream_alloc_pskb(sk
, 0, 0, GFP_ATOMIC
);
1045 if (unlikely(buff
== NULL
))
1048 sk_charge_skb(sk
, buff
);
1049 buff
->truesize
+= nlen
;
1050 skb
->truesize
-= nlen
;
1052 /* Correct the sequence numbers. */
1053 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1054 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1055 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1057 /* PSH and FIN should only be set in the second packet. */
1058 flags
= TCP_SKB_CB(skb
)->flags
;
1059 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
1060 TCP_SKB_CB(buff
)->flags
= flags
;
1062 /* This packet was never sent out yet, so no SACK bits. */
1063 TCP_SKB_CB(buff
)->sacked
= 0;
1065 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_PARTIAL
;
1066 skb_split(skb
, buff
, len
);
1068 /* Fix up tso_factor for both original and new SKB. */
1069 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1070 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1072 /* Link BUFF into the send queue. */
1073 skb_header_release(buff
);
1074 __skb_append(skb
, buff
, &sk
->sk_write_queue
);
1079 /* Try to defer sending, if possible, in order to minimize the amount
1080 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1082 * This algorithm is from John Heffner.
1084 static int tcp_tso_should_defer(struct sock
*sk
, struct tcp_sock
*tp
, struct sk_buff
*skb
)
1086 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1087 u32 send_win
, cong_win
, limit
, in_flight
;
1089 if (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
)
1092 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
1095 in_flight
= tcp_packets_in_flight(tp
);
1097 BUG_ON(tcp_skb_pcount(skb
) <= 1 ||
1098 (tp
->snd_cwnd
<= in_flight
));
1100 send_win
= (tp
->snd_una
+ tp
->snd_wnd
) - TCP_SKB_CB(skb
)->seq
;
1102 /* From in_flight test above, we know that cwnd > in_flight. */
1103 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
1105 limit
= min(send_win
, cong_win
);
1107 /* If a full-sized TSO skb can be sent, do it. */
1111 if (sysctl_tcp_tso_win_divisor
) {
1112 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
1114 /* If at least some fraction of a window is available,
1117 chunk
/= sysctl_tcp_tso_win_divisor
;
1121 /* Different approach, try not to defer past a single
1122 * ACK. Receiver should ACK every other full sized
1123 * frame, so if we have space for more than 3 frames
1126 if (limit
> tcp_max_burst(tp
) * tp
->mss_cache
)
1130 /* Ok, it looks like it is advisable to defer. */
1134 /* Create a new MTU probe if we are ready.
1135 * Returns 0 if we should wait to probe (no cwnd available),
1136 * 1 if a probe was sent,
1138 static int tcp_mtu_probe(struct sock
*sk
)
1140 struct tcp_sock
*tp
= tcp_sk(sk
);
1141 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1142 struct sk_buff
*skb
, *nskb
, *next
;
1149 /* Not currently probing/verifying,
1151 * have enough cwnd, and
1152 * not SACKing (the variable headers throw things off) */
1153 if (!icsk
->icsk_mtup
.enabled
||
1154 icsk
->icsk_mtup
.probe_size
||
1155 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
||
1156 tp
->snd_cwnd
< 11 ||
1157 tp
->rx_opt
.eff_sacks
)
1160 /* Very simple search strategy: just double the MSS. */
1161 mss_now
= tcp_current_mss(sk
, 0);
1162 probe_size
= 2*tp
->mss_cache
;
1163 if (probe_size
> tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_high
)) {
1164 /* TODO: set timer for probe_converge_event */
1168 /* Have enough data in the send queue to probe? */
1170 if ((skb
= sk
->sk_send_head
) == NULL
)
1172 while ((len
+= skb
->len
) < probe_size
&& !tcp_skb_is_last(sk
, skb
))
1174 if (len
< probe_size
)
1177 /* Receive window check. */
1178 if (after(TCP_SKB_CB(skb
)->seq
+ probe_size
, tp
->snd_una
+ tp
->snd_wnd
)) {
1179 if (tp
->snd_wnd
< probe_size
)
1185 /* Do we need to wait to drain cwnd? */
1186 pif
= tcp_packets_in_flight(tp
);
1187 if (pif
+ 2 > tp
->snd_cwnd
) {
1188 /* With no packets in flight, don't stall. */
1195 /* We're allowed to probe. Build it now. */
1196 if ((nskb
= sk_stream_alloc_skb(sk
, probe_size
, GFP_ATOMIC
)) == NULL
)
1198 sk_charge_skb(sk
, nskb
);
1200 skb
= sk
->sk_send_head
;
1201 __skb_insert(nskb
, skb
->prev
, skb
, &sk
->sk_write_queue
);
1202 sk
->sk_send_head
= nskb
;
1204 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(skb
)->seq
;
1205 TCP_SKB_CB(nskb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ probe_size
;
1206 TCP_SKB_CB(nskb
)->flags
= TCPCB_FLAG_ACK
;
1207 TCP_SKB_CB(nskb
)->sacked
= 0;
1209 nskb
->ip_summed
= skb
->ip_summed
;
1212 while (len
< probe_size
) {
1215 copy
= min_t(int, skb
->len
, probe_size
- len
);
1216 if (nskb
->ip_summed
)
1217 skb_copy_bits(skb
, 0, skb_put(nskb
, copy
), copy
);
1219 nskb
->csum
= skb_copy_and_csum_bits(skb
, 0,
1220 skb_put(nskb
, copy
), copy
, nskb
->csum
);
1222 if (skb
->len
<= copy
) {
1223 /* We've eaten all the data from this skb.
1225 TCP_SKB_CB(nskb
)->flags
|= TCP_SKB_CB(skb
)->flags
;
1226 __skb_unlink(skb
, &sk
->sk_write_queue
);
1227 sk_stream_free_skb(sk
, skb
);
1229 TCP_SKB_CB(nskb
)->flags
|= TCP_SKB_CB(skb
)->flags
&
1230 ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
1231 if (!skb_shinfo(skb
)->nr_frags
) {
1232 skb_pull(skb
, copy
);
1233 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1234 skb
->csum
= csum_partial(skb
->data
, skb
->len
, 0);
1236 __pskb_trim_head(skb
, copy
);
1237 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1239 TCP_SKB_CB(skb
)->seq
+= copy
;
1245 tcp_init_tso_segs(sk
, nskb
, nskb
->len
);
1247 /* We're ready to send. If this fails, the probe will
1248 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1249 TCP_SKB_CB(nskb
)->when
= tcp_time_stamp
;
1250 if (!tcp_transmit_skb(sk
, nskb
, 1, GFP_ATOMIC
)) {
1251 /* Decrement cwnd here because we are sending
1252 * effectively two packets. */
1254 update_send_head(sk
, tp
, nskb
);
1256 icsk
->icsk_mtup
.probe_size
= tcp_mss_to_mtu(sk
, nskb
->len
);
1257 tp
->mtu_probe
.probe_seq_start
= TCP_SKB_CB(nskb
)->seq
;
1258 tp
->mtu_probe
.probe_seq_end
= TCP_SKB_CB(nskb
)->end_seq
;
1267 /* This routine writes packets to the network. It advances the
1268 * send_head. This happens as incoming acks open up the remote
1271 * Returns 1, if no segments are in flight and we have queued segments, but
1272 * cannot send anything now because of SWS or another problem.
1274 static int tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
)
1276 struct tcp_sock
*tp
= tcp_sk(sk
);
1277 struct sk_buff
*skb
;
1278 unsigned int tso_segs
, sent_pkts
;
1282 /* If we are closed, the bytes will have to remain here.
1283 * In time closedown will finish, we empty the write queue and all
1286 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
1291 /* Do MTU probing. */
1292 if ((result
= tcp_mtu_probe(sk
)) == 0) {
1294 } else if (result
> 0) {
1298 while ((skb
= sk
->sk_send_head
)) {
1301 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1304 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1308 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
)))
1311 if (tso_segs
== 1) {
1312 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
1313 (tcp_skb_is_last(sk
, skb
) ?
1314 nonagle
: TCP_NAGLE_PUSH
))))
1317 if (tcp_tso_should_defer(sk
, tp
, skb
))
1323 limit
= tcp_window_allows(tp
, skb
,
1324 mss_now
, cwnd_quota
);
1326 if (skb
->len
< limit
) {
1327 unsigned int trim
= skb
->len
% mss_now
;
1330 limit
= skb
->len
- trim
;
1334 if (skb
->len
> limit
&&
1335 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
)))
1338 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1340 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
)))
1343 /* Advance the send_head. This one is sent out.
1344 * This call will increment packets_out.
1346 update_send_head(sk
, tp
, skb
);
1348 tcp_minshall_update(tp
, mss_now
, skb
);
1352 if (likely(sent_pkts
)) {
1353 tcp_cwnd_validate(sk
, tp
);
1356 return !tp
->packets_out
&& sk
->sk_send_head
;
1359 /* Push out any pending frames which were held back due to
1360 * TCP_CORK or attempt at coalescing tiny packets.
1361 * The socket must be locked by the caller.
1363 void __tcp_push_pending_frames(struct sock
*sk
, struct tcp_sock
*tp
,
1364 unsigned int cur_mss
, int nonagle
)
1366 struct sk_buff
*skb
= sk
->sk_send_head
;
1369 if (tcp_write_xmit(sk
, cur_mss
, nonagle
))
1370 tcp_check_probe_timer(sk
, tp
);
1374 /* Send _single_ skb sitting at the send head. This function requires
1375 * true push pending frames to setup probe timer etc.
1377 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
1379 struct tcp_sock
*tp
= tcp_sk(sk
);
1380 struct sk_buff
*skb
= sk
->sk_send_head
;
1381 unsigned int tso_segs
, cwnd_quota
;
1383 BUG_ON(!skb
|| skb
->len
< mss_now
);
1385 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1386 cwnd_quota
= tcp_snd_test(sk
, skb
, mss_now
, TCP_NAGLE_PUSH
);
1388 if (likely(cwnd_quota
)) {
1395 limit
= tcp_window_allows(tp
, skb
,
1396 mss_now
, cwnd_quota
);
1398 if (skb
->len
< limit
) {
1399 unsigned int trim
= skb
->len
% mss_now
;
1402 limit
= skb
->len
- trim
;
1406 if (skb
->len
> limit
&&
1407 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
)))
1410 /* Send it out now. */
1411 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1413 if (likely(!tcp_transmit_skb(sk
, skb
, 1, sk
->sk_allocation
))) {
1414 update_send_head(sk
, tp
, skb
);
1415 tcp_cwnd_validate(sk
, tp
);
1421 /* This function returns the amount that we can raise the
1422 * usable window based on the following constraints
1424 * 1. The window can never be shrunk once it is offered (RFC 793)
1425 * 2. We limit memory per socket
1428 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1429 * RECV.NEXT + RCV.WIN fixed until:
1430 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1432 * i.e. don't raise the right edge of the window until you can raise
1433 * it at least MSS bytes.
1435 * Unfortunately, the recommended algorithm breaks header prediction,
1436 * since header prediction assumes th->window stays fixed.
1438 * Strictly speaking, keeping th->window fixed violates the receiver
1439 * side SWS prevention criteria. The problem is that under this rule
1440 * a stream of single byte packets will cause the right side of the
1441 * window to always advance by a single byte.
1443 * Of course, if the sender implements sender side SWS prevention
1444 * then this will not be a problem.
1446 * BSD seems to make the following compromise:
1448 * If the free space is less than the 1/4 of the maximum
1449 * space available and the free space is less than 1/2 mss,
1450 * then set the window to 0.
1451 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1452 * Otherwise, just prevent the window from shrinking
1453 * and from being larger than the largest representable value.
1455 * This prevents incremental opening of the window in the regime
1456 * where TCP is limited by the speed of the reader side taking
1457 * data out of the TCP receive queue. It does nothing about
1458 * those cases where the window is constrained on the sender side
1459 * because the pipeline is full.
1461 * BSD also seems to "accidentally" limit itself to windows that are a
1462 * multiple of MSS, at least until the free space gets quite small.
1463 * This would appear to be a side effect of the mbuf implementation.
1464 * Combining these two algorithms results in the observed behavior
1465 * of having a fixed window size at almost all times.
1467 * Below we obtain similar behavior by forcing the offered window to
1468 * a multiple of the mss when it is feasible to do so.
1470 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1471 * Regular options like TIMESTAMP are taken into account.
1473 u32
__tcp_select_window(struct sock
*sk
)
1475 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1476 struct tcp_sock
*tp
= tcp_sk(sk
);
1477 /* MSS for the peer's data. Previous versions used mss_clamp
1478 * here. I don't know if the value based on our guesses
1479 * of peer's MSS is better for the performance. It's more correct
1480 * but may be worse for the performance because of rcv_mss
1481 * fluctuations. --SAW 1998/11/1
1483 int mss
= icsk
->icsk_ack
.rcv_mss
;
1484 int free_space
= tcp_space(sk
);
1485 int full_space
= min_t(int, tp
->window_clamp
, tcp_full_space(sk
));
1488 if (mss
> full_space
)
1491 if (free_space
< full_space
/2) {
1492 icsk
->icsk_ack
.quick
= 0;
1494 if (tcp_memory_pressure
)
1495 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U*tp
->advmss
);
1497 if (free_space
< mss
)
1501 if (free_space
> tp
->rcv_ssthresh
)
1502 free_space
= tp
->rcv_ssthresh
;
1504 /* Don't do rounding if we are using window scaling, since the
1505 * scaled window will not line up with the MSS boundary anyway.
1507 window
= tp
->rcv_wnd
;
1508 if (tp
->rx_opt
.rcv_wscale
) {
1509 window
= free_space
;
1511 /* Advertise enough space so that it won't get scaled away.
1512 * Import case: prevent zero window announcement if
1513 * 1<<rcv_wscale > mss.
1515 if (((window
>> tp
->rx_opt
.rcv_wscale
) << tp
->rx_opt
.rcv_wscale
) != window
)
1516 window
= (((window
>> tp
->rx_opt
.rcv_wscale
) + 1)
1517 << tp
->rx_opt
.rcv_wscale
);
1519 /* Get the largest window that is a nice multiple of mss.
1520 * Window clamp already applied above.
1521 * If our current window offering is within 1 mss of the
1522 * free space we just keep it. This prevents the divide
1523 * and multiply from happening most of the time.
1524 * We also don't do any window rounding when the free space
1527 if (window
<= free_space
- mss
|| window
> free_space
)
1528 window
= (free_space
/mss
)*mss
;
1534 /* Attempt to collapse two adjacent SKB's during retransmission. */
1535 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*skb
, int mss_now
)
1537 struct tcp_sock
*tp
= tcp_sk(sk
);
1538 struct sk_buff
*next_skb
= skb
->next
;
1540 /* The first test we must make is that neither of these two
1541 * SKB's are still referenced by someone else.
1543 if (!skb_cloned(skb
) && !skb_cloned(next_skb
)) {
1544 int skb_size
= skb
->len
, next_skb_size
= next_skb
->len
;
1545 u16 flags
= TCP_SKB_CB(skb
)->flags
;
1547 /* Also punt if next skb has been SACK'd. */
1548 if(TCP_SKB_CB(next_skb
)->sacked
& TCPCB_SACKED_ACKED
)
1551 /* Next skb is out of window. */
1552 if (after(TCP_SKB_CB(next_skb
)->end_seq
, tp
->snd_una
+tp
->snd_wnd
))
1555 /* Punt if not enough space exists in the first SKB for
1556 * the data in the second, or the total combined payload
1557 * would exceed the MSS.
1559 if ((next_skb_size
> skb_tailroom(skb
)) ||
1560 ((skb_size
+ next_skb_size
) > mss_now
))
1563 BUG_ON(tcp_skb_pcount(skb
) != 1 ||
1564 tcp_skb_pcount(next_skb
) != 1);
1566 /* changing transmit queue under us so clear hints */
1567 clear_all_retrans_hints(tp
);
1569 /* Ok. We will be able to collapse the packet. */
1570 __skb_unlink(next_skb
, &sk
->sk_write_queue
);
1572 memcpy(skb_put(skb
, next_skb_size
), next_skb
->data
, next_skb_size
);
1574 skb
->ip_summed
= next_skb
->ip_summed
;
1576 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1577 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
1579 /* Update sequence range on original skb. */
1580 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
1582 /* Merge over control information. */
1583 flags
|= TCP_SKB_CB(next_skb
)->flags
; /* This moves PSH/FIN etc. over */
1584 TCP_SKB_CB(skb
)->flags
= flags
;
1586 /* All done, get rid of second SKB and account for it so
1587 * packet counting does not break.
1589 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
&(TCPCB_EVER_RETRANS
|TCPCB_AT_TAIL
);
1590 if (TCP_SKB_CB(next_skb
)->sacked
&TCPCB_SACKED_RETRANS
)
1591 tp
->retrans_out
-= tcp_skb_pcount(next_skb
);
1592 if (TCP_SKB_CB(next_skb
)->sacked
&TCPCB_LOST
) {
1593 tp
->lost_out
-= tcp_skb_pcount(next_skb
);
1594 tp
->left_out
-= tcp_skb_pcount(next_skb
);
1596 /* Reno case is special. Sigh... */
1597 if (!tp
->rx_opt
.sack_ok
&& tp
->sacked_out
) {
1598 tcp_dec_pcount_approx(&tp
->sacked_out
, next_skb
);
1599 tp
->left_out
-= tcp_skb_pcount(next_skb
);
1602 /* Not quite right: it can be > snd.fack, but
1603 * it is better to underestimate fackets.
1605 tcp_dec_pcount_approx(&tp
->fackets_out
, next_skb
);
1606 tcp_packets_out_dec(tp
, next_skb
);
1607 sk_stream_free_skb(sk
, next_skb
);
1611 /* Do a simple retransmit without using the backoff mechanisms in
1612 * tcp_timer. This is used for path mtu discovery.
1613 * The socket is already locked here.
1615 void tcp_simple_retransmit(struct sock
*sk
)
1617 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1618 struct tcp_sock
*tp
= tcp_sk(sk
);
1619 struct sk_buff
*skb
;
1620 unsigned int mss
= tcp_current_mss(sk
, 0);
1623 sk_stream_for_retrans_queue(skb
, sk
) {
1624 if (skb
->len
> mss
&&
1625 !(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
)) {
1626 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) {
1627 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1628 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1630 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_LOST
)) {
1631 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1632 tp
->lost_out
+= tcp_skb_pcount(skb
);
1638 clear_all_retrans_hints(tp
);
1643 tcp_sync_left_out(tp
);
1645 /* Don't muck with the congestion window here.
1646 * Reason is that we do not increase amount of _data_
1647 * in network, but units changed and effective
1648 * cwnd/ssthresh really reduced now.
1650 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
1651 tp
->high_seq
= tp
->snd_nxt
;
1652 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
1653 tp
->prior_ssthresh
= 0;
1654 tp
->undo_marker
= 0;
1655 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1657 tcp_xmit_retransmit_queue(sk
);
1660 /* This retransmits one SKB. Policy decisions and retransmit queue
1661 * state updates are done by the caller. Returns non-zero if an
1662 * error occurred which prevented the send.
1664 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
1666 struct tcp_sock
*tp
= tcp_sk(sk
);
1667 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1668 unsigned int cur_mss
= tcp_current_mss(sk
, 0);
1671 /* Inconslusive MTU probe */
1672 if (icsk
->icsk_mtup
.probe_size
) {
1673 icsk
->icsk_mtup
.probe_size
= 0;
1676 /* Do not sent more than we queued. 1/4 is reserved for possible
1677 * copying overhead: fragmentation, tunneling, mangling etc.
1679 if (atomic_read(&sk
->sk_wmem_alloc
) >
1680 min(sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2), sk
->sk_sndbuf
))
1683 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
1684 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1686 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
1690 /* If receiver has shrunk his window, and skb is out of
1691 * new window, do not retransmit it. The exception is the
1692 * case, when window is shrunk to zero. In this case
1693 * our retransmit serves as a zero window probe.
1695 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
+tp
->snd_wnd
)
1696 && TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
1699 if (skb
->len
> cur_mss
) {
1700 if (tcp_fragment(sk
, skb
, cur_mss
, cur_mss
))
1701 return -ENOMEM
; /* We'll try again later. */
1704 /* Collapse two adjacent packets if worthwhile and we can. */
1705 if(!(TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_SYN
) &&
1706 (skb
->len
< (cur_mss
>> 1)) &&
1707 (skb
->next
!= sk
->sk_send_head
) &&
1708 (skb
->next
!= (struct sk_buff
*)&sk
->sk_write_queue
) &&
1709 (skb_shinfo(skb
)->nr_frags
== 0 && skb_shinfo(skb
->next
)->nr_frags
== 0) &&
1710 (tcp_skb_pcount(skb
) == 1 && tcp_skb_pcount(skb
->next
) == 1) &&
1711 (sysctl_tcp_retrans_collapse
!= 0))
1712 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
1714 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
1715 return -EHOSTUNREACH
; /* Routing failure or similar. */
1717 /* Some Solaris stacks overoptimize and ignore the FIN on a
1718 * retransmit when old data is attached. So strip it off
1719 * since it is cheap to do so and saves bytes on the network.
1722 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
) &&
1723 tp
->snd_una
== (TCP_SKB_CB(skb
)->end_seq
- 1)) {
1724 if (!pskb_trim(skb
, 0)) {
1725 TCP_SKB_CB(skb
)->seq
= TCP_SKB_CB(skb
)->end_seq
- 1;
1726 skb_shinfo(skb
)->gso_segs
= 1;
1727 skb_shinfo(skb
)->gso_size
= 0;
1728 skb_shinfo(skb
)->gso_type
= 0;
1729 skb
->ip_summed
= CHECKSUM_NONE
;
1734 /* Make a copy, if the first transmission SKB clone we made
1735 * is still in somebody's hands, else make a clone.
1737 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1739 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
1742 /* Update global TCP statistics. */
1743 TCP_INC_STATS(TCP_MIB_RETRANSSEGS
);
1745 tp
->total_retrans
++;
1747 #if FASTRETRANS_DEBUG > 0
1748 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) {
1749 if (net_ratelimit())
1750 printk(KERN_DEBUG
"retrans_out leaked.\n");
1753 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
1754 tp
->retrans_out
+= tcp_skb_pcount(skb
);
1756 /* Save stamp of the first retransmit. */
1757 if (!tp
->retrans_stamp
)
1758 tp
->retrans_stamp
= TCP_SKB_CB(skb
)->when
;
1762 /* snd_nxt is stored to detect loss of retransmitted segment,
1763 * see tcp_input.c tcp_sacktag_write_queue().
1765 TCP_SKB_CB(skb
)->ack_seq
= tp
->snd_nxt
;
1770 /* This gets called after a retransmit timeout, and the initially
1771 * retransmitted data is acknowledged. It tries to continue
1772 * resending the rest of the retransmit queue, until either
1773 * we've sent it all or the congestion window limit is reached.
1774 * If doing SACK, the first ACK which comes back for a timeout
1775 * based retransmit packet might feed us FACK information again.
1776 * If so, we use it to avoid unnecessarily retransmissions.
1778 void tcp_xmit_retransmit_queue(struct sock
*sk
)
1780 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1781 struct tcp_sock
*tp
= tcp_sk(sk
);
1782 struct sk_buff
*skb
;
1785 if (tp
->retransmit_skb_hint
) {
1786 skb
= tp
->retransmit_skb_hint
;
1787 packet_cnt
= tp
->retransmit_cnt_hint
;
1789 skb
= sk
->sk_write_queue
.next
;
1793 /* First pass: retransmit lost packets. */
1795 sk_stream_for_retrans_queue_from(skb
, sk
) {
1796 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
1798 /* we could do better than to assign each time */
1799 tp
->retransmit_skb_hint
= skb
;
1800 tp
->retransmit_cnt_hint
= packet_cnt
;
1802 /* Assume this retransmit will generate
1803 * only one packet for congestion window
1804 * calculation purposes. This works because
1805 * tcp_retransmit_skb() will chop up the
1806 * packet to be MSS sized and all the
1807 * packet counting works out.
1809 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
1812 if (sacked
& TCPCB_LOST
) {
1813 if (!(sacked
&(TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))) {
1814 if (tcp_retransmit_skb(sk
, skb
)) {
1815 tp
->retransmit_skb_hint
= NULL
;
1818 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
1819 NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS
);
1821 NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS
);
1824 skb_peek(&sk
->sk_write_queue
))
1825 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1826 inet_csk(sk
)->icsk_rto
,
1830 packet_cnt
+= tcp_skb_pcount(skb
);
1831 if (packet_cnt
>= tp
->lost_out
)
1837 /* OK, demanded retransmission is finished. */
1839 /* Forward retransmissions are possible only during Recovery. */
1840 if (icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
1843 /* No forward retransmissions in Reno are possible. */
1844 if (!tp
->rx_opt
.sack_ok
)
1847 /* Yeah, we have to make difficult choice between forward transmission
1848 * and retransmission... Both ways have their merits...
1850 * For now we do not retransmit anything, while we have some new
1854 if (tcp_may_send_now(sk
, tp
))
1857 if (tp
->forward_skb_hint
) {
1858 skb
= tp
->forward_skb_hint
;
1859 packet_cnt
= tp
->forward_cnt_hint
;
1861 skb
= sk
->sk_write_queue
.next
;
1865 sk_stream_for_retrans_queue_from(skb
, sk
) {
1866 tp
->forward_cnt_hint
= packet_cnt
;
1867 tp
->forward_skb_hint
= skb
;
1869 /* Similar to the retransmit loop above we
1870 * can pretend that the retransmitted SKB
1871 * we send out here will be composed of one
1872 * real MSS sized packet because tcp_retransmit_skb()
1873 * will fragment it if necessary.
1875 if (++packet_cnt
> tp
->fackets_out
)
1878 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
1881 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
)
1884 /* Ok, retransmit it. */
1885 if (tcp_retransmit_skb(sk
, skb
)) {
1886 tp
->forward_skb_hint
= NULL
;
1890 if (skb
== skb_peek(&sk
->sk_write_queue
))
1891 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1892 inet_csk(sk
)->icsk_rto
,
1895 NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS
);
1900 /* Send a fin. The caller locks the socket for us. This cannot be
1901 * allowed to fail queueing a FIN frame under any circumstances.
1903 void tcp_send_fin(struct sock
*sk
)
1905 struct tcp_sock
*tp
= tcp_sk(sk
);
1906 struct sk_buff
*skb
= skb_peek_tail(&sk
->sk_write_queue
);
1909 /* Optimization, tack on the FIN if we have a queue of
1910 * unsent frames. But be careful about outgoing SACKS
1913 mss_now
= tcp_current_mss(sk
, 1);
1915 if (sk
->sk_send_head
!= NULL
) {
1916 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_FIN
;
1917 TCP_SKB_CB(skb
)->end_seq
++;
1920 /* Socket is locked, keep trying until memory is available. */
1922 skb
= alloc_skb_fclone(MAX_TCP_HEADER
, GFP_KERNEL
);
1928 /* Reserve space for headers and prepare control bits. */
1929 skb_reserve(skb
, MAX_TCP_HEADER
);
1931 TCP_SKB_CB(skb
)->flags
= (TCPCB_FLAG_ACK
| TCPCB_FLAG_FIN
);
1932 TCP_SKB_CB(skb
)->sacked
= 0;
1933 skb_shinfo(skb
)->gso_segs
= 1;
1934 skb_shinfo(skb
)->gso_size
= 0;
1935 skb_shinfo(skb
)->gso_type
= 0;
1937 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
1938 TCP_SKB_CB(skb
)->seq
= tp
->write_seq
;
1939 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ 1;
1940 tcp_queue_skb(sk
, skb
);
1942 __tcp_push_pending_frames(sk
, tp
, mss_now
, TCP_NAGLE_OFF
);
1945 /* We get here when a process closes a file descriptor (either due to
1946 * an explicit close() or as a byproduct of exit()'ing) and there
1947 * was unread data in the receive queue. This behavior is recommended
1948 * by draft-ietf-tcpimpl-prob-03.txt section 3.10. -DaveM
1950 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
1952 struct tcp_sock
*tp
= tcp_sk(sk
);
1953 struct sk_buff
*skb
;
1955 /* NOTE: No TCP options attached and we never retransmit this. */
1956 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
1958 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED
);
1962 /* Reserve space for headers and prepare control bits. */
1963 skb_reserve(skb
, MAX_TCP_HEADER
);
1965 TCP_SKB_CB(skb
)->flags
= (TCPCB_FLAG_ACK
| TCPCB_FLAG_RST
);
1966 TCP_SKB_CB(skb
)->sacked
= 0;
1967 skb_shinfo(skb
)->gso_segs
= 1;
1968 skb_shinfo(skb
)->gso_size
= 0;
1969 skb_shinfo(skb
)->gso_type
= 0;
1972 TCP_SKB_CB(skb
)->seq
= tcp_acceptable_seq(sk
, tp
);
1973 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
;
1974 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1975 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
1976 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED
);
1979 /* WARNING: This routine must only be called when we have already sent
1980 * a SYN packet that crossed the incoming SYN that caused this routine
1981 * to get called. If this assumption fails then the initial rcv_wnd
1982 * and rcv_wscale values will not be correct.
1984 int tcp_send_synack(struct sock
*sk
)
1986 struct sk_buff
* skb
;
1988 skb
= skb_peek(&sk
->sk_write_queue
);
1989 if (skb
== NULL
|| !(TCP_SKB_CB(skb
)->flags
&TCPCB_FLAG_SYN
)) {
1990 printk(KERN_DEBUG
"tcp_send_synack: wrong queue state\n");
1993 if (!(TCP_SKB_CB(skb
)->flags
&TCPCB_FLAG_ACK
)) {
1994 if (skb_cloned(skb
)) {
1995 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
1998 __skb_unlink(skb
, &sk
->sk_write_queue
);
1999 skb_header_release(nskb
);
2000 __skb_queue_head(&sk
->sk_write_queue
, nskb
);
2001 sk_stream_free_skb(sk
, skb
);
2002 sk_charge_skb(sk
, nskb
);
2006 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_ACK
;
2007 TCP_ECN_send_synack(tcp_sk(sk
), skb
);
2009 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2010 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2014 * Prepare a SYN-ACK.
2016 struct sk_buff
* tcp_make_synack(struct sock
*sk
, struct dst_entry
*dst
,
2017 struct request_sock
*req
)
2019 struct inet_request_sock
*ireq
= inet_rsk(req
);
2020 struct tcp_sock
*tp
= tcp_sk(sk
);
2022 int tcp_header_size
;
2023 struct sk_buff
*skb
;
2025 skb
= sock_wmalloc(sk
, MAX_TCP_HEADER
+ 15, 1, GFP_ATOMIC
);
2029 /* Reserve space for headers. */
2030 skb_reserve(skb
, MAX_TCP_HEADER
);
2032 skb
->dst
= dst_clone(dst
);
2034 tcp_header_size
= (sizeof(struct tcphdr
) + TCPOLEN_MSS
+
2035 (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0) +
2036 (ireq
->wscale_ok
? TCPOLEN_WSCALE_ALIGNED
: 0) +
2037 /* SACK_PERM is in the place of NOP NOP of TS */
2038 ((ireq
->sack_ok
&& !ireq
->tstamp_ok
) ? TCPOLEN_SACKPERM_ALIGNED
: 0));
2039 skb
->h
.th
= th
= (struct tcphdr
*) skb_push(skb
, tcp_header_size
);
2041 memset(th
, 0, sizeof(struct tcphdr
));
2044 TCP_ECN_make_synack(req
, th
);
2045 th
->source
= inet_sk(sk
)->sport
;
2046 th
->dest
= ireq
->rmt_port
;
2047 TCP_SKB_CB(skb
)->seq
= tcp_rsk(req
)->snt_isn
;
2048 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ 1;
2049 TCP_SKB_CB(skb
)->sacked
= 0;
2050 skb_shinfo(skb
)->gso_segs
= 1;
2051 skb_shinfo(skb
)->gso_size
= 0;
2052 skb_shinfo(skb
)->gso_type
= 0;
2053 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
2054 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_isn
+ 1);
2055 if (req
->rcv_wnd
== 0) { /* ignored for retransmitted syns */
2057 /* Set this up on the first call only */
2058 req
->window_clamp
= tp
->window_clamp
? : dst_metric(dst
, RTAX_WINDOW
);
2059 /* tcp_full_space because it is guaranteed to be the first packet */
2060 tcp_select_initial_window(tcp_full_space(sk
),
2061 dst_metric(dst
, RTAX_ADVMSS
) - (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0),
2066 ireq
->rcv_wscale
= rcv_wscale
;
2069 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2070 th
->window
= htons(req
->rcv_wnd
);
2072 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2073 tcp_syn_build_options((__u32
*)(th
+ 1), dst_metric(dst
, RTAX_ADVMSS
), ireq
->tstamp_ok
,
2074 ireq
->sack_ok
, ireq
->wscale_ok
, ireq
->rcv_wscale
,
2075 TCP_SKB_CB(skb
)->when
,
2079 th
->doff
= (tcp_header_size
>> 2);
2080 TCP_INC_STATS(TCP_MIB_OUTSEGS
);
2085 * Do all connect socket setups that can be done AF independent.
2087 static void tcp_connect_init(struct sock
*sk
)
2089 struct dst_entry
*dst
= __sk_dst_get(sk
);
2090 struct tcp_sock
*tp
= tcp_sk(sk
);
2093 /* We'll fix this up when we get a response from the other end.
2094 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2096 tp
->tcp_header_len
= sizeof(struct tcphdr
) +
2097 (sysctl_tcp_timestamps
? TCPOLEN_TSTAMP_ALIGNED
: 0);
2099 /* If user gave his TCP_MAXSEG, record it to clamp */
2100 if (tp
->rx_opt
.user_mss
)
2101 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
2104 tcp_sync_mss(sk
, dst_mtu(dst
));
2106 if (!tp
->window_clamp
)
2107 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
2108 tp
->advmss
= dst_metric(dst
, RTAX_ADVMSS
);
2109 tcp_initialize_rcv_mss(sk
);
2111 tcp_select_initial_window(tcp_full_space(sk
),
2112 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
2115 sysctl_tcp_window_scaling
,
2118 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
2119 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
2122 sock_reset_flag(sk
, SOCK_DONE
);
2124 tcp_init_wl(tp
, tp
->write_seq
, 0);
2125 tp
->snd_una
= tp
->write_seq
;
2126 tp
->snd_sml
= tp
->write_seq
;
2131 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
2132 inet_csk(sk
)->icsk_retransmits
= 0;
2133 tcp_clear_retrans(tp
);
2137 * Build a SYN and send it off.
2139 int tcp_connect(struct sock
*sk
)
2141 struct tcp_sock
*tp
= tcp_sk(sk
);
2142 struct sk_buff
*buff
;
2144 tcp_connect_init(sk
);
2146 buff
= alloc_skb_fclone(MAX_TCP_HEADER
+ 15, sk
->sk_allocation
);
2147 if (unlikely(buff
== NULL
))
2150 /* Reserve space for headers. */
2151 skb_reserve(buff
, MAX_TCP_HEADER
);
2153 TCP_SKB_CB(buff
)->flags
= TCPCB_FLAG_SYN
;
2154 TCP_ECN_send_syn(sk
, tp
, buff
);
2155 TCP_SKB_CB(buff
)->sacked
= 0;
2156 skb_shinfo(buff
)->gso_segs
= 1;
2157 skb_shinfo(buff
)->gso_size
= 0;
2158 skb_shinfo(buff
)->gso_type
= 0;
2160 tp
->snd_nxt
= tp
->write_seq
;
2161 TCP_SKB_CB(buff
)->seq
= tp
->write_seq
++;
2162 TCP_SKB_CB(buff
)->end_seq
= tp
->write_seq
;
2165 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2166 tp
->retrans_stamp
= TCP_SKB_CB(buff
)->when
;
2167 skb_header_release(buff
);
2168 __skb_queue_tail(&sk
->sk_write_queue
, buff
);
2169 sk_charge_skb(sk
, buff
);
2170 tp
->packets_out
+= tcp_skb_pcount(buff
);
2171 tcp_transmit_skb(sk
, buff
, 1, GFP_KERNEL
);
2173 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2174 * in order to make this packet get counted in tcpOutSegs.
2176 tp
->snd_nxt
= tp
->write_seq
;
2177 tp
->pushed_seq
= tp
->write_seq
;
2178 TCP_INC_STATS(TCP_MIB_ACTIVEOPENS
);
2180 /* Timer for repeating the SYN until an answer. */
2181 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2182 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2186 /* Send out a delayed ack, the caller does the policy checking
2187 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2190 void tcp_send_delayed_ack(struct sock
*sk
)
2192 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2193 int ato
= icsk
->icsk_ack
.ato
;
2194 unsigned long timeout
;
2196 if (ato
> TCP_DELACK_MIN
) {
2197 const struct tcp_sock
*tp
= tcp_sk(sk
);
2200 if (icsk
->icsk_ack
.pingpong
|| (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
2201 max_ato
= TCP_DELACK_MAX
;
2203 /* Slow path, intersegment interval is "high". */
2205 /* If some rtt estimate is known, use it to bound delayed ack.
2206 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2210 int rtt
= max(tp
->srtt
>>3, TCP_DELACK_MIN
);
2216 ato
= min(ato
, max_ato
);
2219 /* Stay within the limit we were given */
2220 timeout
= jiffies
+ ato
;
2222 /* Use new timeout only if there wasn't a older one earlier. */
2223 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
2224 /* If delack timer was blocked or is about to expire,
2227 if (icsk
->icsk_ack
.blocked
||
2228 time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
2233 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
2234 timeout
= icsk
->icsk_ack
.timeout
;
2236 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
2237 icsk
->icsk_ack
.timeout
= timeout
;
2238 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
2241 /* This routine sends an ack and also updates the window. */
2242 void tcp_send_ack(struct sock
*sk
)
2244 /* If we have been reset, we may not send again. */
2245 if (sk
->sk_state
!= TCP_CLOSE
) {
2246 struct tcp_sock
*tp
= tcp_sk(sk
);
2247 struct sk_buff
*buff
;
2249 /* We are not putting this on the write queue, so
2250 * tcp_transmit_skb() will set the ownership to this
2253 buff
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2255 inet_csk_schedule_ack(sk
);
2256 inet_csk(sk
)->icsk_ack
.ato
= TCP_ATO_MIN
;
2257 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
2258 TCP_DELACK_MAX
, TCP_RTO_MAX
);
2262 /* Reserve space for headers and prepare control bits. */
2263 skb_reserve(buff
, MAX_TCP_HEADER
);
2265 TCP_SKB_CB(buff
)->flags
= TCPCB_FLAG_ACK
;
2266 TCP_SKB_CB(buff
)->sacked
= 0;
2267 skb_shinfo(buff
)->gso_segs
= 1;
2268 skb_shinfo(buff
)->gso_size
= 0;
2269 skb_shinfo(buff
)->gso_type
= 0;
2271 /* Send it off, this clears delayed acks for us. */
2272 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(buff
)->end_seq
= tcp_acceptable_seq(sk
, tp
);
2273 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2274 tcp_transmit_skb(sk
, buff
, 0, GFP_ATOMIC
);
2278 /* This routine sends a packet with an out of date sequence
2279 * number. It assumes the other end will try to ack it.
2281 * Question: what should we make while urgent mode?
2282 * 4.4BSD forces sending single byte of data. We cannot send
2283 * out of window data, because we have SND.NXT==SND.MAX...
2285 * Current solution: to send TWO zero-length segments in urgent mode:
2286 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2287 * out-of-date with SND.UNA-1 to probe window.
2289 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
)
2291 struct tcp_sock
*tp
= tcp_sk(sk
);
2292 struct sk_buff
*skb
;
2294 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2295 skb
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2299 /* Reserve space for headers and set control bits. */
2300 skb_reserve(skb
, MAX_TCP_HEADER
);
2302 TCP_SKB_CB(skb
)->flags
= TCPCB_FLAG_ACK
;
2303 TCP_SKB_CB(skb
)->sacked
= urgent
;
2304 skb_shinfo(skb
)->gso_segs
= 1;
2305 skb_shinfo(skb
)->gso_size
= 0;
2306 skb_shinfo(skb
)->gso_type
= 0;
2308 /* Use a previous sequence. This should cause the other
2309 * end to send an ack. Don't queue or clone SKB, just
2312 TCP_SKB_CB(skb
)->seq
= urgent
? tp
->snd_una
: tp
->snd_una
- 1;
2313 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
;
2314 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2315 return tcp_transmit_skb(sk
, skb
, 0, GFP_ATOMIC
);
2318 int tcp_write_wakeup(struct sock
*sk
)
2320 if (sk
->sk_state
!= TCP_CLOSE
) {
2321 struct tcp_sock
*tp
= tcp_sk(sk
);
2322 struct sk_buff
*skb
;
2324 if ((skb
= sk
->sk_send_head
) != NULL
&&
2325 before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
+tp
->snd_wnd
)) {
2327 unsigned int mss
= tcp_current_mss(sk
, 0);
2328 unsigned int seg_size
= tp
->snd_una
+tp
->snd_wnd
-TCP_SKB_CB(skb
)->seq
;
2330 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
2331 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
2333 /* We are probing the opening of a window
2334 * but the window size is != 0
2335 * must have been a result SWS avoidance ( sender )
2337 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
2339 seg_size
= min(seg_size
, mss
);
2340 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2341 if (tcp_fragment(sk
, skb
, seg_size
, mss
))
2343 } else if (!tcp_skb_pcount(skb
))
2344 tcp_set_skb_tso_segs(sk
, skb
, mss
);
2346 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2347 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2348 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2350 update_send_head(sk
, tp
, skb
);
2355 between(tp
->snd_up
, tp
->snd_una
+1, tp
->snd_una
+0xFFFF))
2356 tcp_xmit_probe_skb(sk
, TCPCB_URG
);
2357 return tcp_xmit_probe_skb(sk
, 0);
2363 /* A window probe timeout has occurred. If window is not closed send
2364 * a partial packet else a zero probe.
2366 void tcp_send_probe0(struct sock
*sk
)
2368 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2369 struct tcp_sock
*tp
= tcp_sk(sk
);
2372 err
= tcp_write_wakeup(sk
);
2374 if (tp
->packets_out
|| !sk
->sk_send_head
) {
2375 /* Cancel probe timer, if it is not required. */
2376 icsk
->icsk_probes_out
= 0;
2377 icsk
->icsk_backoff
= 0;
2382 if (icsk
->icsk_backoff
< sysctl_tcp_retries2
)
2383 icsk
->icsk_backoff
++;
2384 icsk
->icsk_probes_out
++;
2385 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2386 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
2389 /* If packet was not sent due to local congestion,
2390 * do not backoff and do not remember icsk_probes_out.
2391 * Let local senders to fight for local resources.
2393 * Use accumulated backoff yet.
2395 if (!icsk
->icsk_probes_out
)
2396 icsk
->icsk_probes_out
= 1;
2397 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2398 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
,
2399 TCP_RESOURCE_PROBE_INTERVAL
),
2404 EXPORT_SYMBOL(tcp_connect
);
2405 EXPORT_SYMBOL(tcp_make_synack
);
2406 EXPORT_SYMBOL(tcp_simple_retransmit
);
2407 EXPORT_SYMBOL(tcp_sync_mss
);
2408 EXPORT_SYMBOL(sysctl_tcp_tso_win_divisor
);
2409 EXPORT_SYMBOL(tcp_mtup_init
);