2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
8 * Version: $Id: tcp_output.c,v 1.146 2002/02/01 22:01:04 davem Exp $
11 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
12 * Mark Evans, <evansmp@uhura.aston.ac.uk>
13 * Corey Minyard <wf-rch!minyard@relay.EU.net>
14 * Florian La Roche, <flla@stud.uni-sb.de>
15 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
16 * Linus Torvalds, <torvalds@cs.helsinki.fi>
17 * Alan Cox, <gw4pts@gw4pts.ampr.org>
18 * Matthew Dillon, <dillon@apollo.west.oic.com>
19 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
20 * Jorge Cwik, <jorge@laser.satlink.net>
24 * Changes: Pedro Roque : Retransmit queue handled by TCP.
25 * : Fragmentation on mtu decrease
26 * : Segment collapse on retransmit
29 * Linus Torvalds : send_delayed_ack
30 * David S. Miller : Charge memory using the right skb
31 * during syn/ack processing.
32 * David S. Miller : Output engine completely rewritten.
33 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
34 * Cacophonix Gaul : draft-minshall-nagle-01
35 * J Hadi Salim : ECN support
41 #include <linux/compiler.h>
42 #include <linux/module.h>
43 #include <linux/smp_lock.h>
45 /* People can turn this off for buggy TCP's found in printers etc. */
46 int sysctl_tcp_retrans_collapse
= 1;
48 /* People can turn this on to work with those rare, broken TCPs that
49 * interpret the window field as a signed quantity.
51 int sysctl_tcp_workaround_signed_windows
= 0;
53 /* This limits the percentage of the congestion window which we
54 * will allow a single TSO frame to consume. Building TSO frames
55 * which are too large can cause TCP streams to be bursty.
57 int sysctl_tcp_tso_win_divisor
= 3;
59 int sysctl_tcp_mtu_probing
= 0;
60 int sysctl_tcp_base_mss
= 512;
62 /* By default, RFC2861 behavior. */
63 int sysctl_tcp_slow_start_after_idle
= 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 while (space
> 65535 && (*rcv_wscale
) < 14) {
210 /* Set initial window to value enough for senders,
211 * following RFC2414. Senders, not following this RFC,
212 * will be satisfied with 2.
214 if (mss
> (1<<*rcv_wscale
)) {
220 if (*rcv_wnd
> init_cwnd
*mss
)
221 *rcv_wnd
= init_cwnd
*mss
;
224 /* Set the clamp no higher than max representable value */
225 (*window_clamp
) = min(65535U << (*rcv_wscale
), *window_clamp
);
228 /* Chose a new window to advertise, update state in tcp_sock for the
229 * socket, and return result with RFC1323 scaling applied. The return
230 * value can be stuffed directly into th->window for an outgoing
233 static u16
tcp_select_window(struct sock
*sk
)
235 struct tcp_sock
*tp
= tcp_sk(sk
);
236 u32 cur_win
= tcp_receive_window(tp
);
237 u32 new_win
= __tcp_select_window(sk
);
239 /* Never shrink the offered window */
240 if(new_win
< cur_win
) {
241 /* Danger Will Robinson!
242 * Don't update rcv_wup/rcv_wnd here or else
243 * we will not be able to advertise a zero
244 * window in time. --DaveM
246 * Relax Will Robinson.
250 tp
->rcv_wnd
= new_win
;
251 tp
->rcv_wup
= tp
->rcv_nxt
;
253 /* Make sure we do not exceed the maximum possible
256 if (!tp
->rx_opt
.rcv_wscale
&& sysctl_tcp_workaround_signed_windows
)
257 new_win
= min(new_win
, MAX_TCP_WINDOW
);
259 new_win
= min(new_win
, (65535U << tp
->rx_opt
.rcv_wscale
));
261 /* RFC1323 scaling applied */
262 new_win
>>= tp
->rx_opt
.rcv_wscale
;
264 /* If we advertise zero window, disable fast path. */
271 static void tcp_build_and_update_options(__u32
*ptr
, struct tcp_sock
*tp
,
274 if (tp
->rx_opt
.tstamp_ok
) {
275 *ptr
++ = __constant_htonl((TCPOPT_NOP
<< 24) |
277 (TCPOPT_TIMESTAMP
<< 8) |
279 *ptr
++ = htonl(tstamp
);
280 *ptr
++ = htonl(tp
->rx_opt
.ts_recent
);
282 if (tp
->rx_opt
.eff_sacks
) {
283 struct tcp_sack_block
*sp
= tp
->rx_opt
.dsack
? tp
->duplicate_sack
: tp
->selective_acks
;
286 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
289 (TCPOLEN_SACK_BASE
+ (tp
->rx_opt
.eff_sacks
*
290 TCPOLEN_SACK_PERBLOCK
)));
291 for(this_sack
= 0; this_sack
< tp
->rx_opt
.eff_sacks
; this_sack
++) {
292 *ptr
++ = htonl(sp
[this_sack
].start_seq
);
293 *ptr
++ = htonl(sp
[this_sack
].end_seq
);
295 if (tp
->rx_opt
.dsack
) {
296 tp
->rx_opt
.dsack
= 0;
297 tp
->rx_opt
.eff_sacks
--;
302 /* Construct a tcp options header for a SYN or SYN_ACK packet.
303 * If this is every changed make sure to change the definition of
304 * MAX_SYN_SIZE to match the new maximum number of options that you
307 static void tcp_syn_build_options(__u32
*ptr
, int mss
, int ts
, int sack
,
308 int offer_wscale
, int wscale
, __u32 tstamp
,
311 /* We always get an MSS option.
312 * The option bytes which will be seen in normal data
313 * packets should timestamps be used, must be in the MSS
314 * advertised. But we subtract them from tp->mss_cache so
315 * that calculations in tcp_sendmsg are simpler etc.
316 * So account for this fact here if necessary. If we
317 * don't do this correctly, as a receiver we won't
318 * recognize data packets as being full sized when we
319 * should, and thus we won't abide by the delayed ACK
321 * SACKs don't matter, we never delay an ACK when we
322 * have any of those going out.
324 *ptr
++ = htonl((TCPOPT_MSS
<< 24) | (TCPOLEN_MSS
<< 16) | mss
);
327 *ptr
++ = __constant_htonl((TCPOPT_SACK_PERM
<< 24) | (TCPOLEN_SACK_PERM
<< 16) |
328 (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
);
330 *ptr
++ = __constant_htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16) |
331 (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
);
332 *ptr
++ = htonl(tstamp
); /* TSVAL */
333 *ptr
++ = htonl(ts_recent
); /* TSECR */
335 *ptr
++ = __constant_htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16) |
336 (TCPOPT_SACK_PERM
<< 8) | TCPOLEN_SACK_PERM
);
338 *ptr
++ = htonl((TCPOPT_NOP
<< 24) | (TCPOPT_WINDOW
<< 16) | (TCPOLEN_WINDOW
<< 8) | (wscale
));
341 /* This routine actually transmits TCP packets queued in by
342 * tcp_do_sendmsg(). This is used by both the initial
343 * transmission and possible later retransmissions.
344 * All SKB's seen here are completely headerless. It is our
345 * job to build the TCP header, and pass the packet down to
346 * IP so it can do the same plus pass the packet off to the
349 * We are working here with either a clone of the original
350 * SKB, or a fresh unique copy made by the retransmit engine.
352 static int tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int clone_it
, gfp_t gfp_mask
)
354 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
355 struct inet_sock
*inet
;
357 struct tcp_skb_cb
*tcb
;
363 BUG_ON(!skb
|| !tcp_skb_pcount(skb
));
365 /* If congestion control is doing timestamping, we must
366 * take such a timestamp before we potentially clone/copy.
368 if (icsk
->icsk_ca_ops
->rtt_sample
)
369 __net_timestamp(skb
);
371 if (likely(clone_it
)) {
372 if (unlikely(skb_cloned(skb
)))
373 skb
= pskb_copy(skb
, gfp_mask
);
375 skb
= skb_clone(skb
, gfp_mask
);
382 tcb
= TCP_SKB_CB(skb
);
383 tcp_header_size
= tp
->tcp_header_len
;
385 #define SYSCTL_FLAG_TSTAMPS 0x1
386 #define SYSCTL_FLAG_WSCALE 0x2
387 #define SYSCTL_FLAG_SACK 0x4
390 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
391 tcp_header_size
= sizeof(struct tcphdr
) + TCPOLEN_MSS
;
392 if(sysctl_tcp_timestamps
) {
393 tcp_header_size
+= TCPOLEN_TSTAMP_ALIGNED
;
394 sysctl_flags
|= SYSCTL_FLAG_TSTAMPS
;
396 if (sysctl_tcp_window_scaling
) {
397 tcp_header_size
+= TCPOLEN_WSCALE_ALIGNED
;
398 sysctl_flags
|= SYSCTL_FLAG_WSCALE
;
400 if (sysctl_tcp_sack
) {
401 sysctl_flags
|= SYSCTL_FLAG_SACK
;
402 if (!(sysctl_flags
& SYSCTL_FLAG_TSTAMPS
))
403 tcp_header_size
+= TCPOLEN_SACKPERM_ALIGNED
;
405 } else if (unlikely(tp
->rx_opt
.eff_sacks
)) {
406 /* A SACK is 2 pad bytes, a 2 byte header, plus
407 * 2 32-bit sequence numbers for each SACK block.
409 tcp_header_size
+= (TCPOLEN_SACK_BASE_ALIGNED
+
410 (tp
->rx_opt
.eff_sacks
*
411 TCPOLEN_SACK_PERBLOCK
));
414 if (tcp_packets_in_flight(tp
) == 0)
415 tcp_ca_event(sk
, CA_EVENT_TX_START
);
417 th
= (struct tcphdr
*) skb_push(skb
, tcp_header_size
);
419 skb_set_owner_w(skb
, sk
);
421 /* Build TCP header and checksum it. */
422 th
->source
= inet
->sport
;
423 th
->dest
= inet
->dport
;
424 th
->seq
= htonl(tcb
->seq
);
425 th
->ack_seq
= htonl(tp
->rcv_nxt
);
426 *(((__u16
*)th
) + 6) = htons(((tcp_header_size
>> 2) << 12) |
429 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
430 /* RFC1323: The window in SYN & SYN/ACK segments
433 th
->window
= htons(tp
->rcv_wnd
);
435 th
->window
= htons(tcp_select_window(sk
));
440 if (unlikely(tp
->urg_mode
&&
441 between(tp
->snd_up
, tcb
->seq
+1, tcb
->seq
+0xFFFF))) {
442 th
->urg_ptr
= htons(tp
->snd_up
-tcb
->seq
);
446 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
447 tcp_syn_build_options((__u32
*)(th
+ 1),
448 tcp_advertise_mss(sk
),
449 (sysctl_flags
& SYSCTL_FLAG_TSTAMPS
),
450 (sysctl_flags
& SYSCTL_FLAG_SACK
),
451 (sysctl_flags
& SYSCTL_FLAG_WSCALE
),
452 tp
->rx_opt
.rcv_wscale
,
454 tp
->rx_opt
.ts_recent
);
456 tcp_build_and_update_options((__u32
*)(th
+ 1),
458 TCP_ECN_send(sk
, tp
, skb
, tcp_header_size
);
461 icsk
->icsk_af_ops
->send_check(sk
, skb
->len
, skb
);
463 if (likely(tcb
->flags
& TCPCB_FLAG_ACK
))
464 tcp_event_ack_sent(sk
, tcp_skb_pcount(skb
));
466 if (skb
->len
!= tcp_header_size
)
467 tcp_event_data_sent(tp
, skb
, sk
);
469 TCP_INC_STATS(TCP_MIB_OUTSEGS
);
471 err
= icsk
->icsk_af_ops
->queue_xmit(skb
, 0);
472 if (likely(err
<= 0))
477 /* NET_XMIT_CN is special. It does not guarantee,
478 * that this packet is lost. It tells that device
479 * is about to start to drop packets or already
480 * drops some packets of the same priority and
481 * invokes us to send less aggressively.
483 return err
== NET_XMIT_CN
? 0 : err
;
485 #undef SYSCTL_FLAG_TSTAMPS
486 #undef SYSCTL_FLAG_WSCALE
487 #undef SYSCTL_FLAG_SACK
491 /* This routine just queue's the buffer
493 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
494 * otherwise socket can stall.
496 static void tcp_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
498 struct tcp_sock
*tp
= tcp_sk(sk
);
500 /* Advance write_seq and place onto the write_queue. */
501 tp
->write_seq
= TCP_SKB_CB(skb
)->end_seq
;
502 skb_header_release(skb
);
503 __skb_queue_tail(&sk
->sk_write_queue
, skb
);
504 sk_charge_skb(sk
, skb
);
506 /* Queue it, remembering where we must start sending. */
507 if (sk
->sk_send_head
== NULL
)
508 sk
->sk_send_head
= skb
;
511 static void tcp_set_skb_tso_segs(struct sock
*sk
, struct sk_buff
*skb
, unsigned int mss_now
)
513 if (skb
->len
<= mss_now
||
514 !(sk
->sk_route_caps
& NETIF_F_TSO
)) {
515 /* Avoid the costly divide in the normal
518 skb_shinfo(skb
)->gso_segs
= 1;
519 skb_shinfo(skb
)->gso_size
= 0;
520 skb_shinfo(skb
)->gso_type
= 0;
524 factor
= skb
->len
+ (mss_now
- 1);
526 skb_shinfo(skb
)->gso_segs
= factor
;
527 skb_shinfo(skb
)->gso_size
= mss_now
;
528 skb_shinfo(skb
)->gso_type
= SKB_GSO_TCPV4
;
532 /* Function to create two new TCP segments. Shrinks the given segment
533 * to the specified size and appends a new segment with the rest of the
534 * packet to the list. This won't be called frequently, I hope.
535 * Remember, these are still headerless SKBs at this point.
537 int tcp_fragment(struct sock
*sk
, struct sk_buff
*skb
, u32 len
, unsigned int mss_now
)
539 struct tcp_sock
*tp
= tcp_sk(sk
);
540 struct sk_buff
*buff
;
541 int nsize
, old_factor
;
545 BUG_ON(len
> skb
->len
);
547 clear_all_retrans_hints(tp
);
548 nsize
= skb_headlen(skb
) - len
;
552 if (skb_cloned(skb
) &&
553 skb_is_nonlinear(skb
) &&
554 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
557 /* Get a new skb... force flag on. */
558 buff
= sk_stream_alloc_skb(sk
, nsize
, GFP_ATOMIC
);
560 return -ENOMEM
; /* We'll just try again later. */
562 sk_charge_skb(sk
, buff
);
563 nlen
= skb
->len
- len
- nsize
;
564 buff
->truesize
+= nlen
;
565 skb
->truesize
-= nlen
;
567 /* Correct the sequence numbers. */
568 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
569 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
570 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
572 /* PSH and FIN should only be set in the second packet. */
573 flags
= TCP_SKB_CB(skb
)->flags
;
574 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
575 TCP_SKB_CB(buff
)->flags
= flags
;
576 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
577 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_AT_TAIL
;
579 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_HW
) {
580 /* Copy and checksum data tail into the new buffer. */
581 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
, skb_put(buff
, nsize
),
586 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
588 skb
->ip_summed
= CHECKSUM_HW
;
589 skb_split(skb
, buff
, len
);
592 buff
->ip_summed
= skb
->ip_summed
;
594 /* Looks stupid, but our code really uses when of
595 * skbs, which it never sent before. --ANK
597 TCP_SKB_CB(buff
)->when
= TCP_SKB_CB(skb
)->when
;
598 buff
->tstamp
= skb
->tstamp
;
600 old_factor
= tcp_skb_pcount(skb
);
602 /* Fix up tso_factor for both original and new SKB. */
603 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
604 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
606 /* If this packet has been sent out already, we must
607 * adjust the various packet counters.
609 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
610 int diff
= old_factor
- tcp_skb_pcount(skb
) -
611 tcp_skb_pcount(buff
);
613 tp
->packets_out
-= diff
;
615 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
616 tp
->sacked_out
-= diff
;
617 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
618 tp
->retrans_out
-= diff
;
620 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
) {
621 tp
->lost_out
-= diff
;
622 tp
->left_out
-= diff
;
626 /* Adjust Reno SACK estimate. */
627 if (!tp
->rx_opt
.sack_ok
) {
628 tp
->sacked_out
-= diff
;
629 if ((int)tp
->sacked_out
< 0)
631 tcp_sync_left_out(tp
);
634 tp
->fackets_out
-= diff
;
635 if ((int)tp
->fackets_out
< 0)
640 /* Link BUFF into the send queue. */
641 skb_header_release(buff
);
642 __skb_append(skb
, buff
, &sk
->sk_write_queue
);
647 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
648 * eventually). The difference is that pulled data not copied, but
649 * immediately discarded.
651 static void __pskb_trim_head(struct sk_buff
*skb
, int len
)
657 for (i
=0; i
<skb_shinfo(skb
)->nr_frags
; i
++) {
658 if (skb_shinfo(skb
)->frags
[i
].size
<= eat
) {
659 put_page(skb_shinfo(skb
)->frags
[i
].page
);
660 eat
-= skb_shinfo(skb
)->frags
[i
].size
;
662 skb_shinfo(skb
)->frags
[k
] = skb_shinfo(skb
)->frags
[i
];
664 skb_shinfo(skb
)->frags
[k
].page_offset
+= eat
;
665 skb_shinfo(skb
)->frags
[k
].size
-= eat
;
671 skb_shinfo(skb
)->nr_frags
= k
;
673 skb
->tail
= skb
->data
;
674 skb
->data_len
-= len
;
675 skb
->len
= skb
->data_len
;
678 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
680 if (skb_cloned(skb
) &&
681 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
684 /* If len == headlen, we avoid __skb_pull to preserve alignment. */
685 if (unlikely(len
< skb_headlen(skb
)))
686 __skb_pull(skb
, len
);
688 __pskb_trim_head(skb
, len
- skb_headlen(skb
));
690 TCP_SKB_CB(skb
)->seq
+= len
;
691 skb
->ip_summed
= CHECKSUM_HW
;
693 skb
->truesize
-= len
;
694 sk
->sk_wmem_queued
-= len
;
695 sk
->sk_forward_alloc
+= len
;
696 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
698 /* Any change of skb->len requires recalculation of tso
701 if (tcp_skb_pcount(skb
) > 1)
702 tcp_set_skb_tso_segs(sk
, skb
, tcp_current_mss(sk
, 1));
707 /* Not accounting for SACKs here. */
708 int tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
710 struct tcp_sock
*tp
= tcp_sk(sk
);
711 struct inet_connection_sock
*icsk
= inet_csk(sk
);
714 /* Calculate base mss without TCP options:
715 It is MMS_S - sizeof(tcphdr) of rfc1122
717 mss_now
= pmtu
- icsk
->icsk_af_ops
->net_header_len
- sizeof(struct tcphdr
);
719 /* Clamp it (mss_clamp does not include tcp options) */
720 if (mss_now
> tp
->rx_opt
.mss_clamp
)
721 mss_now
= tp
->rx_opt
.mss_clamp
;
723 /* Now subtract optional transport overhead */
724 mss_now
-= icsk
->icsk_ext_hdr_len
;
726 /* Then reserve room for full set of TCP options and 8 bytes of data */
730 /* Now subtract TCP options size, not including SACKs */
731 mss_now
-= tp
->tcp_header_len
- sizeof(struct tcphdr
);
736 /* Inverse of above */
737 int tcp_mss_to_mtu(struct sock
*sk
, int mss
)
739 struct tcp_sock
*tp
= tcp_sk(sk
);
740 struct inet_connection_sock
*icsk
= inet_csk(sk
);
745 icsk
->icsk_ext_hdr_len
+
746 icsk
->icsk_af_ops
->net_header_len
;
751 void tcp_mtup_init(struct sock
*sk
)
753 struct tcp_sock
*tp
= tcp_sk(sk
);
754 struct inet_connection_sock
*icsk
= inet_csk(sk
);
756 icsk
->icsk_mtup
.enabled
= sysctl_tcp_mtu_probing
> 1;
757 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+ sizeof(struct tcphdr
) +
758 icsk
->icsk_af_ops
->net_header_len
;
759 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, sysctl_tcp_base_mss
);
760 icsk
->icsk_mtup
.probe_size
= 0;
763 /* This function synchronize snd mss to current pmtu/exthdr set.
765 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
766 for TCP options, but includes only bare TCP header.
768 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
769 It is minimum of user_mss and mss received with SYN.
770 It also does not include TCP options.
772 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
774 tp->mss_cache is current effective sending mss, including
775 all tcp options except for SACKs. It is evaluated,
776 taking into account current pmtu, but never exceeds
777 tp->rx_opt.mss_clamp.
779 NOTE1. rfc1122 clearly states that advertised MSS
780 DOES NOT include either tcp or ip options.
782 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
783 are READ ONLY outside this function. --ANK (980731)
786 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
788 struct tcp_sock
*tp
= tcp_sk(sk
);
789 struct inet_connection_sock
*icsk
= inet_csk(sk
);
792 if (icsk
->icsk_mtup
.search_high
> pmtu
)
793 icsk
->icsk_mtup
.search_high
= pmtu
;
795 mss_now
= tcp_mtu_to_mss(sk
, pmtu
);
797 /* Bound mss with half of window */
798 if (tp
->max_window
&& mss_now
> (tp
->max_window
>>1))
799 mss_now
= max((tp
->max_window
>>1), 68U - tp
->tcp_header_len
);
801 /* And store cached results */
802 icsk
->icsk_pmtu_cookie
= pmtu
;
803 if (icsk
->icsk_mtup
.enabled
)
804 mss_now
= min(mss_now
, tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
));
805 tp
->mss_cache
= mss_now
;
810 /* Compute the current effective MSS, taking SACKs and IP options,
811 * and even PMTU discovery events into account.
813 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
814 * cannot be large. However, taking into account rare use of URG, this
817 unsigned int tcp_current_mss(struct sock
*sk
, int large_allowed
)
819 struct tcp_sock
*tp
= tcp_sk(sk
);
820 struct dst_entry
*dst
= __sk_dst_get(sk
);
825 mss_now
= tp
->mss_cache
;
828 (sk
->sk_route_caps
& NETIF_F_TSO
) &&
833 u32 mtu
= dst_mtu(dst
);
834 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
835 mss_now
= tcp_sync_mss(sk
, mtu
);
838 if (tp
->rx_opt
.eff_sacks
)
839 mss_now
-= (TCPOLEN_SACK_BASE_ALIGNED
+
840 (tp
->rx_opt
.eff_sacks
* TCPOLEN_SACK_PERBLOCK
));
842 xmit_size_goal
= mss_now
;
845 xmit_size_goal
= (65535 -
846 inet_csk(sk
)->icsk_af_ops
->net_header_len
-
847 inet_csk(sk
)->icsk_ext_hdr_len
-
850 if (tp
->max_window
&&
851 (xmit_size_goal
> (tp
->max_window
>> 1)))
852 xmit_size_goal
= max((tp
->max_window
>> 1),
853 68U - tp
->tcp_header_len
);
855 xmit_size_goal
-= (xmit_size_goal
% mss_now
);
857 tp
->xmit_size_goal
= xmit_size_goal
;
862 /* Congestion window validation. (RFC2861) */
864 static void tcp_cwnd_validate(struct sock
*sk
, struct tcp_sock
*tp
)
866 __u32 packets_out
= tp
->packets_out
;
868 if (packets_out
>= tp
->snd_cwnd
) {
869 /* Network is feed fully. */
870 tp
->snd_cwnd_used
= 0;
871 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
873 /* Network starves. */
874 if (tp
->packets_out
> tp
->snd_cwnd_used
)
875 tp
->snd_cwnd_used
= tp
->packets_out
;
877 if ((s32
)(tcp_time_stamp
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
)
878 tcp_cwnd_application_limited(sk
);
882 static unsigned int tcp_window_allows(struct tcp_sock
*tp
, struct sk_buff
*skb
, unsigned int mss_now
, unsigned int cwnd
)
884 u32 window
, cwnd_len
;
886 window
= (tp
->snd_una
+ tp
->snd_wnd
- TCP_SKB_CB(skb
)->seq
);
887 cwnd_len
= mss_now
* cwnd
;
888 return min(window
, cwnd_len
);
891 /* Can at least one segment of SKB be sent right now, according to the
892 * congestion window rules? If so, return how many segments are allowed.
894 static inline unsigned int tcp_cwnd_test(struct tcp_sock
*tp
, struct sk_buff
*skb
)
898 /* Don't be strict about the congestion window for the final FIN. */
899 if (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
)
902 in_flight
= tcp_packets_in_flight(tp
);
904 if (in_flight
< cwnd
)
905 return (cwnd
- in_flight
);
910 /* This must be invoked the first time we consider transmitting
913 static int tcp_init_tso_segs(struct sock
*sk
, struct sk_buff
*skb
, unsigned int mss_now
)
915 int tso_segs
= tcp_skb_pcount(skb
);
919 tcp_skb_mss(skb
) != mss_now
)) {
920 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
921 tso_segs
= tcp_skb_pcount(skb
);
926 static inline int tcp_minshall_check(const struct tcp_sock
*tp
)
928 return after(tp
->snd_sml
,tp
->snd_una
) &&
929 !after(tp
->snd_sml
, tp
->snd_nxt
);
932 /* Return 0, if packet can be sent now without violation Nagle's rules:
933 * 1. It is full sized.
934 * 2. Or it contains FIN. (already checked by caller)
935 * 3. Or TCP_NODELAY was set.
936 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
937 * With Minshall's modification: all sent small packets are ACKed.
940 static inline int tcp_nagle_check(const struct tcp_sock
*tp
,
941 const struct sk_buff
*skb
,
942 unsigned mss_now
, int nonagle
)
944 return (skb
->len
< mss_now
&&
945 ((nonagle
&TCP_NAGLE_CORK
) ||
948 tcp_minshall_check(tp
))));
951 /* Return non-zero if the Nagle test allows this packet to be
954 static inline int tcp_nagle_test(struct tcp_sock
*tp
, struct sk_buff
*skb
,
955 unsigned int cur_mss
, int nonagle
)
957 /* Nagle rule does not apply to frames, which sit in the middle of the
958 * write_queue (they have no chances to get new data).
960 * This is implemented in the callers, where they modify the 'nonagle'
961 * argument based upon the location of SKB in the send queue.
963 if (nonagle
& TCP_NAGLE_PUSH
)
966 /* Don't use the nagle rule for urgent data (or for the final FIN). */
968 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
))
971 if (!tcp_nagle_check(tp
, skb
, cur_mss
, nonagle
))
977 /* Does at least the first segment of SKB fit into the send window? */
978 static inline int tcp_snd_wnd_test(struct tcp_sock
*tp
, struct sk_buff
*skb
, unsigned int cur_mss
)
980 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
982 if (skb
->len
> cur_mss
)
983 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
985 return !after(end_seq
, tp
->snd_una
+ tp
->snd_wnd
);
988 /* This checks if the data bearing packet SKB (usually sk->sk_send_head)
989 * should be put on the wire right now. If so, it returns the number of
990 * packets allowed by the congestion window.
992 static unsigned int tcp_snd_test(struct sock
*sk
, struct sk_buff
*skb
,
993 unsigned int cur_mss
, int nonagle
)
995 struct tcp_sock
*tp
= tcp_sk(sk
);
996 unsigned int cwnd_quota
;
998 tcp_init_tso_segs(sk
, skb
, cur_mss
);
1000 if (!tcp_nagle_test(tp
, skb
, cur_mss
, nonagle
))
1003 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1005 !tcp_snd_wnd_test(tp
, skb
, cur_mss
))
1011 static inline int tcp_skb_is_last(const struct sock
*sk
,
1012 const struct sk_buff
*skb
)
1014 return skb
->next
== (struct sk_buff
*)&sk
->sk_write_queue
;
1017 int tcp_may_send_now(struct sock
*sk
, struct tcp_sock
*tp
)
1019 struct sk_buff
*skb
= sk
->sk_send_head
;
1022 tcp_snd_test(sk
, skb
, tcp_current_mss(sk
, 1),
1023 (tcp_skb_is_last(sk
, skb
) ?
1028 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1029 * which is put after SKB on the list. It is very much like
1030 * tcp_fragment() except that it may make several kinds of assumptions
1031 * in order to speed up the splitting operation. In particular, we
1032 * know that all the data is in scatter-gather pages, and that the
1033 * packet has never been sent out before (and thus is not cloned).
1035 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
, unsigned int mss_now
)
1037 struct sk_buff
*buff
;
1038 int nlen
= skb
->len
- len
;
1041 /* All of a TSO frame must be composed of paged data. */
1042 if (skb
->len
!= skb
->data_len
)
1043 return tcp_fragment(sk
, skb
, len
, mss_now
);
1045 buff
= sk_stream_alloc_pskb(sk
, 0, 0, GFP_ATOMIC
);
1046 if (unlikely(buff
== NULL
))
1049 sk_charge_skb(sk
, buff
);
1050 buff
->truesize
+= nlen
;
1051 skb
->truesize
-= nlen
;
1053 /* Correct the sequence numbers. */
1054 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1055 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1056 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1058 /* PSH and FIN should only be set in the second packet. */
1059 flags
= TCP_SKB_CB(skb
)->flags
;
1060 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
1061 TCP_SKB_CB(buff
)->flags
= flags
;
1063 /* This packet was never sent out yet, so no SACK bits. */
1064 TCP_SKB_CB(buff
)->sacked
= 0;
1066 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_HW
;
1067 skb_split(skb
, buff
, len
);
1069 /* Fix up tso_factor for both original and new SKB. */
1070 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1071 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1073 /* Link BUFF into the send queue. */
1074 skb_header_release(buff
);
1075 __skb_append(skb
, buff
, &sk
->sk_write_queue
);
1080 /* Try to defer sending, if possible, in order to minimize the amount
1081 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1083 * This algorithm is from John Heffner.
1085 static int tcp_tso_should_defer(struct sock
*sk
, struct tcp_sock
*tp
, struct sk_buff
*skb
)
1087 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1088 u32 send_win
, cong_win
, limit
, in_flight
;
1090 if (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
)
1093 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
1096 in_flight
= tcp_packets_in_flight(tp
);
1098 BUG_ON(tcp_skb_pcount(skb
) <= 1 ||
1099 (tp
->snd_cwnd
<= in_flight
));
1101 send_win
= (tp
->snd_una
+ tp
->snd_wnd
) - TCP_SKB_CB(skb
)->seq
;
1103 /* From in_flight test above, we know that cwnd > in_flight. */
1104 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
1106 limit
= min(send_win
, cong_win
);
1108 /* If a full-sized TSO skb can be sent, do it. */
1112 if (sysctl_tcp_tso_win_divisor
) {
1113 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
1115 /* If at least some fraction of a window is available,
1118 chunk
/= sysctl_tcp_tso_win_divisor
;
1122 /* Different approach, try not to defer past a single
1123 * ACK. Receiver should ACK every other full sized
1124 * frame, so if we have space for more than 3 frames
1127 if (limit
> tcp_max_burst(tp
) * tp
->mss_cache
)
1131 /* Ok, it looks like it is advisable to defer. */
1135 /* Create a new MTU probe if we are ready.
1136 * Returns 0 if we should wait to probe (no cwnd available),
1137 * 1 if a probe was sent,
1139 static int tcp_mtu_probe(struct sock
*sk
)
1141 struct tcp_sock
*tp
= tcp_sk(sk
);
1142 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1143 struct sk_buff
*skb
, *nskb
, *next
;
1150 /* Not currently probing/verifying,
1152 * have enough cwnd, and
1153 * not SACKing (the variable headers throw things off) */
1154 if (!icsk
->icsk_mtup
.enabled
||
1155 icsk
->icsk_mtup
.probe_size
||
1156 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
||
1157 tp
->snd_cwnd
< 11 ||
1158 tp
->rx_opt
.eff_sacks
)
1161 /* Very simple search strategy: just double the MSS. */
1162 mss_now
= tcp_current_mss(sk
, 0);
1163 probe_size
= 2*tp
->mss_cache
;
1164 if (probe_size
> tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_high
)) {
1165 /* TODO: set timer for probe_converge_event */
1169 /* Have enough data in the send queue to probe? */
1171 if ((skb
= sk
->sk_send_head
) == NULL
)
1173 while ((len
+= skb
->len
) < probe_size
&& !tcp_skb_is_last(sk
, skb
))
1175 if (len
< probe_size
)
1178 /* Receive window check. */
1179 if (after(TCP_SKB_CB(skb
)->seq
+ probe_size
, tp
->snd_una
+ tp
->snd_wnd
)) {
1180 if (tp
->snd_wnd
< probe_size
)
1186 /* Do we need to wait to drain cwnd? */
1187 pif
= tcp_packets_in_flight(tp
);
1188 if (pif
+ 2 > tp
->snd_cwnd
) {
1189 /* With no packets in flight, don't stall. */
1196 /* We're allowed to probe. Build it now. */
1197 if ((nskb
= sk_stream_alloc_skb(sk
, probe_size
, GFP_ATOMIC
)) == NULL
)
1199 sk_charge_skb(sk
, nskb
);
1201 skb
= sk
->sk_send_head
;
1202 __skb_insert(nskb
, skb
->prev
, skb
, &sk
->sk_write_queue
);
1203 sk
->sk_send_head
= nskb
;
1205 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(skb
)->seq
;
1206 TCP_SKB_CB(nskb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ probe_size
;
1207 TCP_SKB_CB(nskb
)->flags
= TCPCB_FLAG_ACK
;
1208 TCP_SKB_CB(nskb
)->sacked
= 0;
1210 if (skb
->ip_summed
== CHECKSUM_HW
)
1211 nskb
->ip_summed
= CHECKSUM_HW
;
1214 while (len
< probe_size
) {
1217 copy
= min_t(int, skb
->len
, probe_size
- len
);
1218 if (nskb
->ip_summed
)
1219 skb_copy_bits(skb
, 0, skb_put(nskb
, copy
), copy
);
1221 nskb
->csum
= skb_copy_and_csum_bits(skb
, 0,
1222 skb_put(nskb
, copy
), copy
, nskb
->csum
);
1224 if (skb
->len
<= copy
) {
1225 /* We've eaten all the data from this skb.
1227 TCP_SKB_CB(nskb
)->flags
|= TCP_SKB_CB(skb
)->flags
;
1228 __skb_unlink(skb
, &sk
->sk_write_queue
);
1229 sk_stream_free_skb(sk
, skb
);
1231 TCP_SKB_CB(nskb
)->flags
|= TCP_SKB_CB(skb
)->flags
&
1232 ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
1233 if (!skb_shinfo(skb
)->nr_frags
) {
1234 skb_pull(skb
, copy
);
1235 if (skb
->ip_summed
!= CHECKSUM_HW
)
1236 skb
->csum
= csum_partial(skb
->data
, skb
->len
, 0);
1238 __pskb_trim_head(skb
, copy
);
1239 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1241 TCP_SKB_CB(skb
)->seq
+= copy
;
1247 tcp_init_tso_segs(sk
, nskb
, nskb
->len
);
1249 /* We're ready to send. If this fails, the probe will
1250 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1251 TCP_SKB_CB(nskb
)->when
= tcp_time_stamp
;
1252 if (!tcp_transmit_skb(sk
, nskb
, 1, GFP_ATOMIC
)) {
1253 /* Decrement cwnd here because we are sending
1254 * effectively two packets. */
1256 update_send_head(sk
, tp
, nskb
);
1258 icsk
->icsk_mtup
.probe_size
= tcp_mss_to_mtu(sk
, nskb
->len
);
1259 tp
->mtu_probe
.probe_seq_start
= TCP_SKB_CB(nskb
)->seq
;
1260 tp
->mtu_probe
.probe_seq_end
= TCP_SKB_CB(nskb
)->end_seq
;
1269 /* This routine writes packets to the network. It advances the
1270 * send_head. This happens as incoming acks open up the remote
1273 * Returns 1, if no segments are in flight and we have queued segments, but
1274 * cannot send anything now because of SWS or another problem.
1276 static int tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
)
1278 struct tcp_sock
*tp
= tcp_sk(sk
);
1279 struct sk_buff
*skb
;
1280 unsigned int tso_segs
, sent_pkts
;
1284 /* If we are closed, the bytes will have to remain here.
1285 * In time closedown will finish, we empty the write queue and all
1288 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
1293 /* Do MTU probing. */
1294 if ((result
= tcp_mtu_probe(sk
)) == 0) {
1296 } else if (result
> 0) {
1300 while ((skb
= sk
->sk_send_head
)) {
1303 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1306 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1310 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
)))
1313 if (tso_segs
== 1) {
1314 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
1315 (tcp_skb_is_last(sk
, skb
) ?
1316 nonagle
: TCP_NAGLE_PUSH
))))
1319 if (tcp_tso_should_defer(sk
, tp
, skb
))
1325 limit
= tcp_window_allows(tp
, skb
,
1326 mss_now
, cwnd_quota
);
1328 if (skb
->len
< limit
) {
1329 unsigned int trim
= skb
->len
% mss_now
;
1332 limit
= skb
->len
- trim
;
1336 if (skb
->len
> limit
&&
1337 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
)))
1340 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1342 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
)))
1345 /* Advance the send_head. This one is sent out.
1346 * This call will increment packets_out.
1348 update_send_head(sk
, tp
, skb
);
1350 tcp_minshall_update(tp
, mss_now
, skb
);
1354 if (likely(sent_pkts
)) {
1355 tcp_cwnd_validate(sk
, tp
);
1358 return !tp
->packets_out
&& sk
->sk_send_head
;
1361 /* Push out any pending frames which were held back due to
1362 * TCP_CORK or attempt at coalescing tiny packets.
1363 * The socket must be locked by the caller.
1365 void __tcp_push_pending_frames(struct sock
*sk
, struct tcp_sock
*tp
,
1366 unsigned int cur_mss
, int nonagle
)
1368 struct sk_buff
*skb
= sk
->sk_send_head
;
1371 if (tcp_write_xmit(sk
, cur_mss
, nonagle
))
1372 tcp_check_probe_timer(sk
, tp
);
1376 /* Send _single_ skb sitting at the send head. This function requires
1377 * true push pending frames to setup probe timer etc.
1379 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
1381 struct tcp_sock
*tp
= tcp_sk(sk
);
1382 struct sk_buff
*skb
= sk
->sk_send_head
;
1383 unsigned int tso_segs
, cwnd_quota
;
1385 BUG_ON(!skb
|| skb
->len
< mss_now
);
1387 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1388 cwnd_quota
= tcp_snd_test(sk
, skb
, mss_now
, TCP_NAGLE_PUSH
);
1390 if (likely(cwnd_quota
)) {
1397 limit
= tcp_window_allows(tp
, skb
,
1398 mss_now
, cwnd_quota
);
1400 if (skb
->len
< limit
) {
1401 unsigned int trim
= skb
->len
% mss_now
;
1404 limit
= skb
->len
- trim
;
1408 if (skb
->len
> limit
&&
1409 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
)))
1412 /* Send it out now. */
1413 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1415 if (likely(!tcp_transmit_skb(sk
, skb
, 1, sk
->sk_allocation
))) {
1416 update_send_head(sk
, tp
, skb
);
1417 tcp_cwnd_validate(sk
, tp
);
1423 /* This function returns the amount that we can raise the
1424 * usable window based on the following constraints
1426 * 1. The window can never be shrunk once it is offered (RFC 793)
1427 * 2. We limit memory per socket
1430 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1431 * RECV.NEXT + RCV.WIN fixed until:
1432 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1434 * i.e. don't raise the right edge of the window until you can raise
1435 * it at least MSS bytes.
1437 * Unfortunately, the recommended algorithm breaks header prediction,
1438 * since header prediction assumes th->window stays fixed.
1440 * Strictly speaking, keeping th->window fixed violates the receiver
1441 * side SWS prevention criteria. The problem is that under this rule
1442 * a stream of single byte packets will cause the right side of the
1443 * window to always advance by a single byte.
1445 * Of course, if the sender implements sender side SWS prevention
1446 * then this will not be a problem.
1448 * BSD seems to make the following compromise:
1450 * If the free space is less than the 1/4 of the maximum
1451 * space available and the free space is less than 1/2 mss,
1452 * then set the window to 0.
1453 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1454 * Otherwise, just prevent the window from shrinking
1455 * and from being larger than the largest representable value.
1457 * This prevents incremental opening of the window in the regime
1458 * where TCP is limited by the speed of the reader side taking
1459 * data out of the TCP receive queue. It does nothing about
1460 * those cases where the window is constrained on the sender side
1461 * because the pipeline is full.
1463 * BSD also seems to "accidentally" limit itself to windows that are a
1464 * multiple of MSS, at least until the free space gets quite small.
1465 * This would appear to be a side effect of the mbuf implementation.
1466 * Combining these two algorithms results in the observed behavior
1467 * of having a fixed window size at almost all times.
1469 * Below we obtain similar behavior by forcing the offered window to
1470 * a multiple of the mss when it is feasible to do so.
1472 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1473 * Regular options like TIMESTAMP are taken into account.
1475 u32
__tcp_select_window(struct sock
*sk
)
1477 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1478 struct tcp_sock
*tp
= tcp_sk(sk
);
1479 /* MSS for the peer's data. Previous versions used mss_clamp
1480 * here. I don't know if the value based on our guesses
1481 * of peer's MSS is better for the performance. It's more correct
1482 * but may be worse for the performance because of rcv_mss
1483 * fluctuations. --SAW 1998/11/1
1485 int mss
= icsk
->icsk_ack
.rcv_mss
;
1486 int free_space
= tcp_space(sk
);
1487 int full_space
= min_t(int, tp
->window_clamp
, tcp_full_space(sk
));
1490 if (mss
> full_space
)
1493 if (free_space
< full_space
/2) {
1494 icsk
->icsk_ack
.quick
= 0;
1496 if (tcp_memory_pressure
)
1497 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U*tp
->advmss
);
1499 if (free_space
< mss
)
1503 if (free_space
> tp
->rcv_ssthresh
)
1504 free_space
= tp
->rcv_ssthresh
;
1506 /* Don't do rounding if we are using window scaling, since the
1507 * scaled window will not line up with the MSS boundary anyway.
1509 window
= tp
->rcv_wnd
;
1510 if (tp
->rx_opt
.rcv_wscale
) {
1511 window
= free_space
;
1513 /* Advertise enough space so that it won't get scaled away.
1514 * Import case: prevent zero window announcement if
1515 * 1<<rcv_wscale > mss.
1517 if (((window
>> tp
->rx_opt
.rcv_wscale
) << tp
->rx_opt
.rcv_wscale
) != window
)
1518 window
= (((window
>> tp
->rx_opt
.rcv_wscale
) + 1)
1519 << tp
->rx_opt
.rcv_wscale
);
1521 /* Get the largest window that is a nice multiple of mss.
1522 * Window clamp already applied above.
1523 * If our current window offering is within 1 mss of the
1524 * free space we just keep it. This prevents the divide
1525 * and multiply from happening most of the time.
1526 * We also don't do any window rounding when the free space
1529 if (window
<= free_space
- mss
|| window
> free_space
)
1530 window
= (free_space
/mss
)*mss
;
1536 /* Attempt to collapse two adjacent SKB's during retransmission. */
1537 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*skb
, int mss_now
)
1539 struct tcp_sock
*tp
= tcp_sk(sk
);
1540 struct sk_buff
*next_skb
= skb
->next
;
1542 /* The first test we must make is that neither of these two
1543 * SKB's are still referenced by someone else.
1545 if (!skb_cloned(skb
) && !skb_cloned(next_skb
)) {
1546 int skb_size
= skb
->len
, next_skb_size
= next_skb
->len
;
1547 u16 flags
= TCP_SKB_CB(skb
)->flags
;
1549 /* Also punt if next skb has been SACK'd. */
1550 if(TCP_SKB_CB(next_skb
)->sacked
& TCPCB_SACKED_ACKED
)
1553 /* Next skb is out of window. */
1554 if (after(TCP_SKB_CB(next_skb
)->end_seq
, tp
->snd_una
+tp
->snd_wnd
))
1557 /* Punt if not enough space exists in the first SKB for
1558 * the data in the second, or the total combined payload
1559 * would exceed the MSS.
1561 if ((next_skb_size
> skb_tailroom(skb
)) ||
1562 ((skb_size
+ next_skb_size
) > mss_now
))
1565 BUG_ON(tcp_skb_pcount(skb
) != 1 ||
1566 tcp_skb_pcount(next_skb
) != 1);
1568 /* changing transmit queue under us so clear hints */
1569 clear_all_retrans_hints(tp
);
1571 /* Ok. We will be able to collapse the packet. */
1572 __skb_unlink(next_skb
, &sk
->sk_write_queue
);
1574 memcpy(skb_put(skb
, next_skb_size
), next_skb
->data
, next_skb_size
);
1576 if (next_skb
->ip_summed
== CHECKSUM_HW
)
1577 skb
->ip_summed
= CHECKSUM_HW
;
1579 if (skb
->ip_summed
!= CHECKSUM_HW
)
1580 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
1582 /* Update sequence range on original skb. */
1583 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
1585 /* Merge over control information. */
1586 flags
|= TCP_SKB_CB(next_skb
)->flags
; /* This moves PSH/FIN etc. over */
1587 TCP_SKB_CB(skb
)->flags
= flags
;
1589 /* All done, get rid of second SKB and account for it so
1590 * packet counting does not break.
1592 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
&(TCPCB_EVER_RETRANS
|TCPCB_AT_TAIL
);
1593 if (TCP_SKB_CB(next_skb
)->sacked
&TCPCB_SACKED_RETRANS
)
1594 tp
->retrans_out
-= tcp_skb_pcount(next_skb
);
1595 if (TCP_SKB_CB(next_skb
)->sacked
&TCPCB_LOST
) {
1596 tp
->lost_out
-= tcp_skb_pcount(next_skb
);
1597 tp
->left_out
-= tcp_skb_pcount(next_skb
);
1599 /* Reno case is special. Sigh... */
1600 if (!tp
->rx_opt
.sack_ok
&& tp
->sacked_out
) {
1601 tcp_dec_pcount_approx(&tp
->sacked_out
, next_skb
);
1602 tp
->left_out
-= tcp_skb_pcount(next_skb
);
1605 /* Not quite right: it can be > snd.fack, but
1606 * it is better to underestimate fackets.
1608 tcp_dec_pcount_approx(&tp
->fackets_out
, next_skb
);
1609 tcp_packets_out_dec(tp
, next_skb
);
1610 sk_stream_free_skb(sk
, next_skb
);
1614 /* Do a simple retransmit without using the backoff mechanisms in
1615 * tcp_timer. This is used for path mtu discovery.
1616 * The socket is already locked here.
1618 void tcp_simple_retransmit(struct sock
*sk
)
1620 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1621 struct tcp_sock
*tp
= tcp_sk(sk
);
1622 struct sk_buff
*skb
;
1623 unsigned int mss
= tcp_current_mss(sk
, 0);
1626 sk_stream_for_retrans_queue(skb
, sk
) {
1627 if (skb
->len
> mss
&&
1628 !(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
)) {
1629 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) {
1630 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1631 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1633 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_LOST
)) {
1634 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1635 tp
->lost_out
+= tcp_skb_pcount(skb
);
1641 clear_all_retrans_hints(tp
);
1646 tcp_sync_left_out(tp
);
1648 /* Don't muck with the congestion window here.
1649 * Reason is that we do not increase amount of _data_
1650 * in network, but units changed and effective
1651 * cwnd/ssthresh really reduced now.
1653 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
1654 tp
->high_seq
= tp
->snd_nxt
;
1655 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
1656 tp
->prior_ssthresh
= 0;
1657 tp
->undo_marker
= 0;
1658 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1660 tcp_xmit_retransmit_queue(sk
);
1663 /* This retransmits one SKB. Policy decisions and retransmit queue
1664 * state updates are done by the caller. Returns non-zero if an
1665 * error occurred which prevented the send.
1667 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
1669 struct tcp_sock
*tp
= tcp_sk(sk
);
1670 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1671 unsigned int cur_mss
= tcp_current_mss(sk
, 0);
1674 /* Inconslusive MTU probe */
1675 if (icsk
->icsk_mtup
.probe_size
) {
1676 icsk
->icsk_mtup
.probe_size
= 0;
1679 /* Do not sent more than we queued. 1/4 is reserved for possible
1680 * copying overhead: fragmentation, tunneling, mangling etc.
1682 if (atomic_read(&sk
->sk_wmem_alloc
) >
1683 min(sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2), sk
->sk_sndbuf
))
1686 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
1687 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1689 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
1693 /* If receiver has shrunk his window, and skb is out of
1694 * new window, do not retransmit it. The exception is the
1695 * case, when window is shrunk to zero. In this case
1696 * our retransmit serves as a zero window probe.
1698 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
+tp
->snd_wnd
)
1699 && TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
1702 if (skb
->len
> cur_mss
) {
1703 if (tcp_fragment(sk
, skb
, cur_mss
, cur_mss
))
1704 return -ENOMEM
; /* We'll try again later. */
1707 /* Collapse two adjacent packets if worthwhile and we can. */
1708 if(!(TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_SYN
) &&
1709 (skb
->len
< (cur_mss
>> 1)) &&
1710 (skb
->next
!= sk
->sk_send_head
) &&
1711 (skb
->next
!= (struct sk_buff
*)&sk
->sk_write_queue
) &&
1712 (skb_shinfo(skb
)->nr_frags
== 0 && skb_shinfo(skb
->next
)->nr_frags
== 0) &&
1713 (tcp_skb_pcount(skb
) == 1 && tcp_skb_pcount(skb
->next
) == 1) &&
1714 (sysctl_tcp_retrans_collapse
!= 0))
1715 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
1717 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
1718 return -EHOSTUNREACH
; /* Routing failure or similar. */
1720 /* Some Solaris stacks overoptimize and ignore the FIN on a
1721 * retransmit when old data is attached. So strip it off
1722 * since it is cheap to do so and saves bytes on the network.
1725 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
) &&
1726 tp
->snd_una
== (TCP_SKB_CB(skb
)->end_seq
- 1)) {
1727 if (!pskb_trim(skb
, 0)) {
1728 TCP_SKB_CB(skb
)->seq
= TCP_SKB_CB(skb
)->end_seq
- 1;
1729 skb_shinfo(skb
)->gso_segs
= 1;
1730 skb_shinfo(skb
)->gso_size
= 0;
1731 skb_shinfo(skb
)->gso_type
= 0;
1732 skb
->ip_summed
= CHECKSUM_NONE
;
1737 /* Make a copy, if the first transmission SKB clone we made
1738 * is still in somebody's hands, else make a clone.
1740 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1742 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
1745 /* Update global TCP statistics. */
1746 TCP_INC_STATS(TCP_MIB_RETRANSSEGS
);
1748 tp
->total_retrans
++;
1750 #if FASTRETRANS_DEBUG > 0
1751 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) {
1752 if (net_ratelimit())
1753 printk(KERN_DEBUG
"retrans_out leaked.\n");
1756 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
1757 tp
->retrans_out
+= tcp_skb_pcount(skb
);
1759 /* Save stamp of the first retransmit. */
1760 if (!tp
->retrans_stamp
)
1761 tp
->retrans_stamp
= TCP_SKB_CB(skb
)->when
;
1765 /* snd_nxt is stored to detect loss of retransmitted segment,
1766 * see tcp_input.c tcp_sacktag_write_queue().
1768 TCP_SKB_CB(skb
)->ack_seq
= tp
->snd_nxt
;
1773 /* This gets called after a retransmit timeout, and the initially
1774 * retransmitted data is acknowledged. It tries to continue
1775 * resending the rest of the retransmit queue, until either
1776 * we've sent it all or the congestion window limit is reached.
1777 * If doing SACK, the first ACK which comes back for a timeout
1778 * based retransmit packet might feed us FACK information again.
1779 * If so, we use it to avoid unnecessarily retransmissions.
1781 void tcp_xmit_retransmit_queue(struct sock
*sk
)
1783 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1784 struct tcp_sock
*tp
= tcp_sk(sk
);
1785 struct sk_buff
*skb
;
1788 if (tp
->retransmit_skb_hint
) {
1789 skb
= tp
->retransmit_skb_hint
;
1790 packet_cnt
= tp
->retransmit_cnt_hint
;
1792 skb
= sk
->sk_write_queue
.next
;
1796 /* First pass: retransmit lost packets. */
1798 sk_stream_for_retrans_queue_from(skb
, sk
) {
1799 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
1801 /* we could do better than to assign each time */
1802 tp
->retransmit_skb_hint
= skb
;
1803 tp
->retransmit_cnt_hint
= packet_cnt
;
1805 /* Assume this retransmit will generate
1806 * only one packet for congestion window
1807 * calculation purposes. This works because
1808 * tcp_retransmit_skb() will chop up the
1809 * packet to be MSS sized and all the
1810 * packet counting works out.
1812 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
1815 if (sacked
& TCPCB_LOST
) {
1816 if (!(sacked
&(TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))) {
1817 if (tcp_retransmit_skb(sk
, skb
)) {
1818 tp
->retransmit_skb_hint
= NULL
;
1821 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
1822 NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS
);
1824 NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS
);
1827 skb_peek(&sk
->sk_write_queue
))
1828 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1829 inet_csk(sk
)->icsk_rto
,
1833 packet_cnt
+= tcp_skb_pcount(skb
);
1834 if (packet_cnt
>= tp
->lost_out
)
1840 /* OK, demanded retransmission is finished. */
1842 /* Forward retransmissions are possible only during Recovery. */
1843 if (icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
1846 /* No forward retransmissions in Reno are possible. */
1847 if (!tp
->rx_opt
.sack_ok
)
1850 /* Yeah, we have to make difficult choice between forward transmission
1851 * and retransmission... Both ways have their merits...
1853 * For now we do not retransmit anything, while we have some new
1857 if (tcp_may_send_now(sk
, tp
))
1860 if (tp
->forward_skb_hint
) {
1861 skb
= tp
->forward_skb_hint
;
1862 packet_cnt
= tp
->forward_cnt_hint
;
1864 skb
= sk
->sk_write_queue
.next
;
1868 sk_stream_for_retrans_queue_from(skb
, sk
) {
1869 tp
->forward_cnt_hint
= packet_cnt
;
1870 tp
->forward_skb_hint
= skb
;
1872 /* Similar to the retransmit loop above we
1873 * can pretend that the retransmitted SKB
1874 * we send out here will be composed of one
1875 * real MSS sized packet because tcp_retransmit_skb()
1876 * will fragment it if necessary.
1878 if (++packet_cnt
> tp
->fackets_out
)
1881 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
1884 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
)
1887 /* Ok, retransmit it. */
1888 if (tcp_retransmit_skb(sk
, skb
)) {
1889 tp
->forward_skb_hint
= NULL
;
1893 if (skb
== skb_peek(&sk
->sk_write_queue
))
1894 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1895 inet_csk(sk
)->icsk_rto
,
1898 NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS
);
1903 /* Send a fin. The caller locks the socket for us. This cannot be
1904 * allowed to fail queueing a FIN frame under any circumstances.
1906 void tcp_send_fin(struct sock
*sk
)
1908 struct tcp_sock
*tp
= tcp_sk(sk
);
1909 struct sk_buff
*skb
= skb_peek_tail(&sk
->sk_write_queue
);
1912 /* Optimization, tack on the FIN if we have a queue of
1913 * unsent frames. But be careful about outgoing SACKS
1916 mss_now
= tcp_current_mss(sk
, 1);
1918 if (sk
->sk_send_head
!= NULL
) {
1919 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_FIN
;
1920 TCP_SKB_CB(skb
)->end_seq
++;
1923 /* Socket is locked, keep trying until memory is available. */
1925 skb
= alloc_skb_fclone(MAX_TCP_HEADER
, GFP_KERNEL
);
1931 /* Reserve space for headers and prepare control bits. */
1932 skb_reserve(skb
, MAX_TCP_HEADER
);
1934 TCP_SKB_CB(skb
)->flags
= (TCPCB_FLAG_ACK
| TCPCB_FLAG_FIN
);
1935 TCP_SKB_CB(skb
)->sacked
= 0;
1936 skb_shinfo(skb
)->gso_segs
= 1;
1937 skb_shinfo(skb
)->gso_size
= 0;
1938 skb_shinfo(skb
)->gso_type
= 0;
1940 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
1941 TCP_SKB_CB(skb
)->seq
= tp
->write_seq
;
1942 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ 1;
1943 tcp_queue_skb(sk
, skb
);
1945 __tcp_push_pending_frames(sk
, tp
, mss_now
, TCP_NAGLE_OFF
);
1948 /* We get here when a process closes a file descriptor (either due to
1949 * an explicit close() or as a byproduct of exit()'ing) and there
1950 * was unread data in the receive queue. This behavior is recommended
1951 * by draft-ietf-tcpimpl-prob-03.txt section 3.10. -DaveM
1953 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
1955 struct tcp_sock
*tp
= tcp_sk(sk
);
1956 struct sk_buff
*skb
;
1958 /* NOTE: No TCP options attached and we never retransmit this. */
1959 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
1961 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED
);
1965 /* Reserve space for headers and prepare control bits. */
1966 skb_reserve(skb
, MAX_TCP_HEADER
);
1968 TCP_SKB_CB(skb
)->flags
= (TCPCB_FLAG_ACK
| TCPCB_FLAG_RST
);
1969 TCP_SKB_CB(skb
)->sacked
= 0;
1970 skb_shinfo(skb
)->gso_segs
= 1;
1971 skb_shinfo(skb
)->gso_size
= 0;
1972 skb_shinfo(skb
)->gso_type
= 0;
1975 TCP_SKB_CB(skb
)->seq
= tcp_acceptable_seq(sk
, tp
);
1976 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
;
1977 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1978 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
1979 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED
);
1982 /* WARNING: This routine must only be called when we have already sent
1983 * a SYN packet that crossed the incoming SYN that caused this routine
1984 * to get called. If this assumption fails then the initial rcv_wnd
1985 * and rcv_wscale values will not be correct.
1987 int tcp_send_synack(struct sock
*sk
)
1989 struct sk_buff
* skb
;
1991 skb
= skb_peek(&sk
->sk_write_queue
);
1992 if (skb
== NULL
|| !(TCP_SKB_CB(skb
)->flags
&TCPCB_FLAG_SYN
)) {
1993 printk(KERN_DEBUG
"tcp_send_synack: wrong queue state\n");
1996 if (!(TCP_SKB_CB(skb
)->flags
&TCPCB_FLAG_ACK
)) {
1997 if (skb_cloned(skb
)) {
1998 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
2001 __skb_unlink(skb
, &sk
->sk_write_queue
);
2002 skb_header_release(nskb
);
2003 __skb_queue_head(&sk
->sk_write_queue
, nskb
);
2004 sk_stream_free_skb(sk
, skb
);
2005 sk_charge_skb(sk
, nskb
);
2009 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_ACK
;
2010 TCP_ECN_send_synack(tcp_sk(sk
), skb
);
2012 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2013 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2017 * Prepare a SYN-ACK.
2019 struct sk_buff
* tcp_make_synack(struct sock
*sk
, struct dst_entry
*dst
,
2020 struct request_sock
*req
)
2022 struct inet_request_sock
*ireq
= inet_rsk(req
);
2023 struct tcp_sock
*tp
= tcp_sk(sk
);
2025 int tcp_header_size
;
2026 struct sk_buff
*skb
;
2028 skb
= sock_wmalloc(sk
, MAX_TCP_HEADER
+ 15, 1, GFP_ATOMIC
);
2032 /* Reserve space for headers. */
2033 skb_reserve(skb
, MAX_TCP_HEADER
);
2035 skb
->dst
= dst_clone(dst
);
2037 tcp_header_size
= (sizeof(struct tcphdr
) + TCPOLEN_MSS
+
2038 (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0) +
2039 (ireq
->wscale_ok
? TCPOLEN_WSCALE_ALIGNED
: 0) +
2040 /* SACK_PERM is in the place of NOP NOP of TS */
2041 ((ireq
->sack_ok
&& !ireq
->tstamp_ok
) ? TCPOLEN_SACKPERM_ALIGNED
: 0));
2042 skb
->h
.th
= th
= (struct tcphdr
*) skb_push(skb
, tcp_header_size
);
2044 memset(th
, 0, sizeof(struct tcphdr
));
2047 if (dst
->dev
->features
&NETIF_F_TSO
)
2049 TCP_ECN_make_synack(req
, th
);
2050 th
->source
= inet_sk(sk
)->sport
;
2051 th
->dest
= ireq
->rmt_port
;
2052 TCP_SKB_CB(skb
)->seq
= tcp_rsk(req
)->snt_isn
;
2053 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ 1;
2054 TCP_SKB_CB(skb
)->sacked
= 0;
2055 skb_shinfo(skb
)->gso_segs
= 1;
2056 skb_shinfo(skb
)->gso_size
= 0;
2057 skb_shinfo(skb
)->gso_type
= 0;
2058 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
2059 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_isn
+ 1);
2060 if (req
->rcv_wnd
== 0) { /* ignored for retransmitted syns */
2062 /* Set this up on the first call only */
2063 req
->window_clamp
= tp
->window_clamp
? : dst_metric(dst
, RTAX_WINDOW
);
2064 /* tcp_full_space because it is guaranteed to be the first packet */
2065 tcp_select_initial_window(tcp_full_space(sk
),
2066 dst_metric(dst
, RTAX_ADVMSS
) - (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0),
2071 ireq
->rcv_wscale
= rcv_wscale
;
2074 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2075 th
->window
= htons(req
->rcv_wnd
);
2077 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2078 tcp_syn_build_options((__u32
*)(th
+ 1), dst_metric(dst
, RTAX_ADVMSS
), ireq
->tstamp_ok
,
2079 ireq
->sack_ok
, ireq
->wscale_ok
, ireq
->rcv_wscale
,
2080 TCP_SKB_CB(skb
)->when
,
2084 th
->doff
= (tcp_header_size
>> 2);
2085 TCP_INC_STATS(TCP_MIB_OUTSEGS
);
2090 * Do all connect socket setups that can be done AF independent.
2092 static void tcp_connect_init(struct sock
*sk
)
2094 struct dst_entry
*dst
= __sk_dst_get(sk
);
2095 struct tcp_sock
*tp
= tcp_sk(sk
);
2098 /* We'll fix this up when we get a response from the other end.
2099 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2101 tp
->tcp_header_len
= sizeof(struct tcphdr
) +
2102 (sysctl_tcp_timestamps
? TCPOLEN_TSTAMP_ALIGNED
: 0);
2104 /* If user gave his TCP_MAXSEG, record it to clamp */
2105 if (tp
->rx_opt
.user_mss
)
2106 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
2109 tcp_sync_mss(sk
, dst_mtu(dst
));
2111 if (!tp
->window_clamp
)
2112 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
2113 tp
->advmss
= dst_metric(dst
, RTAX_ADVMSS
);
2114 tcp_initialize_rcv_mss(sk
);
2116 tcp_select_initial_window(tcp_full_space(sk
),
2117 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
2120 sysctl_tcp_window_scaling
,
2123 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
2124 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
2127 sock_reset_flag(sk
, SOCK_DONE
);
2129 tcp_init_wl(tp
, tp
->write_seq
, 0);
2130 tp
->snd_una
= tp
->write_seq
;
2131 tp
->snd_sml
= tp
->write_seq
;
2136 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
2137 inet_csk(sk
)->icsk_retransmits
= 0;
2138 tcp_clear_retrans(tp
);
2142 * Build a SYN and send it off.
2144 int tcp_connect(struct sock
*sk
)
2146 struct tcp_sock
*tp
= tcp_sk(sk
);
2147 struct sk_buff
*buff
;
2149 tcp_connect_init(sk
);
2151 buff
= alloc_skb_fclone(MAX_TCP_HEADER
+ 15, sk
->sk_allocation
);
2152 if (unlikely(buff
== NULL
))
2155 /* Reserve space for headers. */
2156 skb_reserve(buff
, MAX_TCP_HEADER
);
2158 TCP_SKB_CB(buff
)->flags
= TCPCB_FLAG_SYN
;
2159 TCP_ECN_send_syn(sk
, tp
, buff
);
2160 TCP_SKB_CB(buff
)->sacked
= 0;
2161 skb_shinfo(buff
)->gso_segs
= 1;
2162 skb_shinfo(buff
)->gso_size
= 0;
2163 skb_shinfo(buff
)->gso_type
= 0;
2165 TCP_SKB_CB(buff
)->seq
= tp
->write_seq
++;
2166 TCP_SKB_CB(buff
)->end_seq
= tp
->write_seq
;
2167 tp
->snd_nxt
= tp
->write_seq
;
2168 tp
->pushed_seq
= tp
->write_seq
;
2171 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2172 tp
->retrans_stamp
= TCP_SKB_CB(buff
)->when
;
2173 skb_header_release(buff
);
2174 __skb_queue_tail(&sk
->sk_write_queue
, buff
);
2175 sk_charge_skb(sk
, buff
);
2176 tp
->packets_out
+= tcp_skb_pcount(buff
);
2177 tcp_transmit_skb(sk
, buff
, 1, GFP_KERNEL
);
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
);