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>
44 /* People can turn this off for buggy TCP's found in printers etc. */
45 int sysctl_tcp_retrans_collapse __read_mostly
= 1;
47 /* People can turn this on to work with those rare, broken TCPs that
48 * interpret the window field as a signed quantity.
50 int sysctl_tcp_workaround_signed_windows __read_mostly
= 0;
52 /* This limits the percentage of the congestion window which we
53 * will allow a single TSO frame to consume. Building TSO frames
54 * which are too large can cause TCP streams to be bursty.
56 int sysctl_tcp_tso_win_divisor __read_mostly
= 3;
58 int sysctl_tcp_mtu_probing __read_mostly
= 0;
59 int sysctl_tcp_base_mss __read_mostly
= 512;
61 /* By default, RFC2861 behavior. */
62 int sysctl_tcp_slow_start_after_idle __read_mostly
= 1;
64 static inline void tcp_packets_out_inc(struct sock
*sk
,
65 const struct sk_buff
*skb
)
67 struct tcp_sock
*tp
= tcp_sk(sk
);
68 int orig
= tp
->packets_out
;
70 tp
->packets_out
+= tcp_skb_pcount(skb
);
72 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
73 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
76 static void update_send_head(struct sock
*sk
, struct sk_buff
*skb
)
78 struct tcp_sock
*tp
= tcp_sk(sk
);
80 tcp_advance_send_head(sk
, skb
);
81 tp
->snd_nxt
= TCP_SKB_CB(skb
)->end_seq
;
82 tcp_packets_out_inc(sk
, skb
);
84 /* Don't override Nagle indefinately with F-RTO */
85 if (tp
->frto_counter
== 2)
89 /* SND.NXT, if window was not shrunk.
90 * If window has been shrunk, what should we make? It is not clear at all.
91 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
92 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
93 * invalid. OK, let's make this for now:
95 static inline __u32
tcp_acceptable_seq(struct sock
*sk
)
97 struct tcp_sock
*tp
= tcp_sk(sk
);
99 if (!before(tp
->snd_una
+tp
->snd_wnd
, tp
->snd_nxt
))
102 return tp
->snd_una
+tp
->snd_wnd
;
105 /* Calculate mss to advertise in SYN segment.
106 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
108 * 1. It is independent of path mtu.
109 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
110 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
111 * attached devices, because some buggy hosts are confused by
113 * 4. We do not make 3, we advertise MSS, calculated from first
114 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
115 * This may be overridden via information stored in routing table.
116 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
117 * probably even Jumbo".
119 static __u16
tcp_advertise_mss(struct sock
*sk
)
121 struct tcp_sock
*tp
= tcp_sk(sk
);
122 struct dst_entry
*dst
= __sk_dst_get(sk
);
123 int mss
= tp
->advmss
;
125 if (dst
&& dst_metric(dst
, RTAX_ADVMSS
) < mss
) {
126 mss
= dst_metric(dst
, RTAX_ADVMSS
);
133 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
134 * This is the first part of cwnd validation mechanism. */
135 static void tcp_cwnd_restart(struct sock
*sk
, struct dst_entry
*dst
)
137 struct tcp_sock
*tp
= tcp_sk(sk
);
138 s32 delta
= tcp_time_stamp
- tp
->lsndtime
;
139 u32 restart_cwnd
= tcp_init_cwnd(tp
, dst
);
140 u32 cwnd
= tp
->snd_cwnd
;
142 tcp_ca_event(sk
, CA_EVENT_CWND_RESTART
);
144 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
145 restart_cwnd
= min(restart_cwnd
, cwnd
);
147 while ((delta
-= inet_csk(sk
)->icsk_rto
) > 0 && cwnd
> restart_cwnd
)
149 tp
->snd_cwnd
= max(cwnd
, restart_cwnd
);
150 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
151 tp
->snd_cwnd_used
= 0;
154 static void tcp_event_data_sent(struct tcp_sock
*tp
,
155 struct sk_buff
*skb
, struct sock
*sk
)
157 struct inet_connection_sock
*icsk
= inet_csk(sk
);
158 const u32 now
= tcp_time_stamp
;
160 if (sysctl_tcp_slow_start_after_idle
&&
161 (!tp
->packets_out
&& (s32
)(now
- tp
->lsndtime
) > icsk
->icsk_rto
))
162 tcp_cwnd_restart(sk
, __sk_dst_get(sk
));
166 /* If it is a reply for ato after last received
167 * packet, enter pingpong mode.
169 if ((u32
)(now
- icsk
->icsk_ack
.lrcvtime
) < icsk
->icsk_ack
.ato
)
170 icsk
->icsk_ack
.pingpong
= 1;
173 static inline void tcp_event_ack_sent(struct sock
*sk
, unsigned int pkts
)
175 tcp_dec_quickack_mode(sk
, pkts
);
176 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_DACK
);
179 /* Determine a window scaling and initial window to offer.
180 * Based on the assumption that the given amount of space
181 * will be offered. Store the results in the tp structure.
182 * NOTE: for smooth operation initial space offering should
183 * be a multiple of mss if possible. We assume here that mss >= 1.
184 * This MUST be enforced by all callers.
186 void tcp_select_initial_window(int __space
, __u32 mss
,
187 __u32
*rcv_wnd
, __u32
*window_clamp
,
188 int wscale_ok
, __u8
*rcv_wscale
)
190 unsigned int space
= (__space
< 0 ? 0 : __space
);
192 /* If no clamp set the clamp to the max possible scaled window */
193 if (*window_clamp
== 0)
194 (*window_clamp
) = (65535 << 14);
195 space
= min(*window_clamp
, space
);
197 /* Quantize space offering to a multiple of mss if possible. */
199 space
= (space
/ mss
) * mss
;
201 /* NOTE: offering an initial window larger than 32767
202 * will break some buggy TCP stacks. If the admin tells us
203 * it is likely we could be speaking with such a buggy stack
204 * we will truncate our initial window offering to 32K-1
205 * unless the remote has sent us a window scaling option,
206 * which we interpret as a sign the remote TCP is not
207 * misinterpreting the window field as a signed quantity.
209 if (sysctl_tcp_workaround_signed_windows
)
210 (*rcv_wnd
) = min(space
, MAX_TCP_WINDOW
);
216 /* Set window scaling on max possible window
217 * See RFC1323 for an explanation of the limit to 14
219 space
= max_t(u32
, sysctl_tcp_rmem
[2], sysctl_rmem_max
);
220 space
= min_t(u32
, space
, *window_clamp
);
221 while (space
> 65535 && (*rcv_wscale
) < 14) {
227 /* Set initial window to value enough for senders,
228 * following RFC2414. Senders, not following this RFC,
229 * will be satisfied with 2.
231 if (mss
> (1<<*rcv_wscale
)) {
237 if (*rcv_wnd
> init_cwnd
*mss
)
238 *rcv_wnd
= init_cwnd
*mss
;
241 /* Set the clamp no higher than max representable value */
242 (*window_clamp
) = min(65535U << (*rcv_wscale
), *window_clamp
);
245 /* Chose a new window to advertise, update state in tcp_sock for the
246 * socket, and return result with RFC1323 scaling applied. The return
247 * value can be stuffed directly into th->window for an outgoing
250 static u16
tcp_select_window(struct sock
*sk
)
252 struct tcp_sock
*tp
= tcp_sk(sk
);
253 u32 cur_win
= tcp_receive_window(tp
);
254 u32 new_win
= __tcp_select_window(sk
);
256 /* Never shrink the offered window */
257 if (new_win
< cur_win
) {
258 /* Danger Will Robinson!
259 * Don't update rcv_wup/rcv_wnd here or else
260 * we will not be able to advertise a zero
261 * window in time. --DaveM
263 * Relax Will Robinson.
267 tp
->rcv_wnd
= new_win
;
268 tp
->rcv_wup
= tp
->rcv_nxt
;
270 /* Make sure we do not exceed the maximum possible
273 if (!tp
->rx_opt
.rcv_wscale
&& sysctl_tcp_workaround_signed_windows
)
274 new_win
= min(new_win
, MAX_TCP_WINDOW
);
276 new_win
= min(new_win
, (65535U << tp
->rx_opt
.rcv_wscale
));
278 /* RFC1323 scaling applied */
279 new_win
>>= tp
->rx_opt
.rcv_wscale
;
281 /* If we advertise zero window, disable fast path. */
288 static inline void TCP_ECN_send_synack(struct tcp_sock
*tp
,
291 TCP_SKB_CB(skb
)->flags
&= ~TCPCB_FLAG_CWR
;
292 if (!(tp
->ecn_flags
&TCP_ECN_OK
))
293 TCP_SKB_CB(skb
)->flags
&= ~TCPCB_FLAG_ECE
;
296 static inline void TCP_ECN_send_syn(struct sock
*sk
, struct sk_buff
*skb
)
298 struct tcp_sock
*tp
= tcp_sk(sk
);
301 if (sysctl_tcp_ecn
) {
302 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_ECE
|TCPCB_FLAG_CWR
;
303 tp
->ecn_flags
= TCP_ECN_OK
;
307 static __inline__
void
308 TCP_ECN_make_synack(struct request_sock
*req
, struct tcphdr
*th
)
310 if (inet_rsk(req
)->ecn_ok
)
314 static inline void TCP_ECN_send(struct sock
*sk
, struct sk_buff
*skb
,
317 struct tcp_sock
*tp
= tcp_sk(sk
);
319 if (tp
->ecn_flags
& TCP_ECN_OK
) {
320 /* Not-retransmitted data segment: set ECT and inject CWR. */
321 if (skb
->len
!= tcp_header_len
&&
322 !before(TCP_SKB_CB(skb
)->seq
, tp
->snd_nxt
)) {
324 if (tp
->ecn_flags
&TCP_ECN_QUEUE_CWR
) {
325 tp
->ecn_flags
&= ~TCP_ECN_QUEUE_CWR
;
326 tcp_hdr(skb
)->cwr
= 1;
327 skb_shinfo(skb
)->gso_type
|= SKB_GSO_TCP_ECN
;
330 /* ACK or retransmitted segment: clear ECT|CE */
331 INET_ECN_dontxmit(sk
);
333 if (tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)
334 tcp_hdr(skb
)->ece
= 1;
338 static void tcp_build_and_update_options(__be32
*ptr
, struct tcp_sock
*tp
,
339 __u32 tstamp
, __u8
**md5_hash
)
341 if (tp
->rx_opt
.tstamp_ok
) {
342 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
344 (TCPOPT_TIMESTAMP
<< 8) |
346 *ptr
++ = htonl(tstamp
);
347 *ptr
++ = htonl(tp
->rx_opt
.ts_recent
);
349 if (tp
->rx_opt
.eff_sacks
) {
350 struct tcp_sack_block
*sp
= tp
->rx_opt
.dsack
? tp
->duplicate_sack
: tp
->selective_acks
;
353 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
356 (TCPOLEN_SACK_BASE
+ (tp
->rx_opt
.eff_sacks
*
357 TCPOLEN_SACK_PERBLOCK
)));
359 for (this_sack
= 0; this_sack
< tp
->rx_opt
.eff_sacks
; this_sack
++) {
360 *ptr
++ = htonl(sp
[this_sack
].start_seq
);
361 *ptr
++ = htonl(sp
[this_sack
].end_seq
);
364 if (tp
->rx_opt
.dsack
) {
365 tp
->rx_opt
.dsack
= 0;
366 tp
->rx_opt
.eff_sacks
--;
369 #ifdef CONFIG_TCP_MD5SIG
371 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
373 (TCPOPT_MD5SIG
<< 8) |
375 *md5_hash
= (__u8
*)ptr
;
380 /* Construct a tcp options header for a SYN or SYN_ACK packet.
381 * If this is every changed make sure to change the definition of
382 * MAX_SYN_SIZE to match the new maximum number of options that you
385 * Note - that with the RFC2385 TCP option, we make room for the
386 * 16 byte MD5 hash. This will be filled in later, so the pointer for the
387 * location to be filled is passed back up.
389 static void tcp_syn_build_options(__be32
*ptr
, int mss
, int ts
, int sack
,
390 int offer_wscale
, int wscale
, __u32 tstamp
,
391 __u32 ts_recent
, __u8
**md5_hash
)
393 /* We always get an MSS option.
394 * The option bytes which will be seen in normal data
395 * packets should timestamps be used, must be in the MSS
396 * advertised. But we subtract them from tp->mss_cache so
397 * that calculations in tcp_sendmsg are simpler etc.
398 * So account for this fact here if necessary. If we
399 * don't do this correctly, as a receiver we won't
400 * recognize data packets as being full sized when we
401 * should, and thus we won't abide by the delayed ACK
403 * SACKs don't matter, we never delay an ACK when we
404 * have any of those going out.
406 *ptr
++ = htonl((TCPOPT_MSS
<< 24) | (TCPOLEN_MSS
<< 16) | mss
);
409 *ptr
++ = htonl((TCPOPT_SACK_PERM
<< 24) |
410 (TCPOLEN_SACK_PERM
<< 16) |
411 (TCPOPT_TIMESTAMP
<< 8) |
414 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
416 (TCPOPT_TIMESTAMP
<< 8) |
418 *ptr
++ = htonl(tstamp
); /* TSVAL */
419 *ptr
++ = htonl(ts_recent
); /* TSECR */
421 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
423 (TCPOPT_SACK_PERM
<< 8) |
426 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
427 (TCPOPT_WINDOW
<< 16) |
428 (TCPOLEN_WINDOW
<< 8) |
430 #ifdef CONFIG_TCP_MD5SIG
432 * If MD5 is enabled, then we set the option, and include the size
433 * (always 18). The actual MD5 hash is added just before the
437 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
439 (TCPOPT_MD5SIG
<< 8) |
441 *md5_hash
= (__u8
*) ptr
;
446 /* This routine actually transmits TCP packets queued in by
447 * tcp_do_sendmsg(). This is used by both the initial
448 * transmission and possible later retransmissions.
449 * All SKB's seen here are completely headerless. It is our
450 * job to build the TCP header, and pass the packet down to
451 * IP so it can do the same plus pass the packet off to the
454 * We are working here with either a clone of the original
455 * SKB, or a fresh unique copy made by the retransmit engine.
457 static int tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int clone_it
, gfp_t gfp_mask
)
459 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
460 struct inet_sock
*inet
;
462 struct tcp_skb_cb
*tcb
;
464 #ifdef CONFIG_TCP_MD5SIG
465 struct tcp_md5sig_key
*md5
;
466 __u8
*md5_hash_location
;
472 BUG_ON(!skb
|| !tcp_skb_pcount(skb
));
474 /* If congestion control is doing timestamping, we must
475 * take such a timestamp before we potentially clone/copy.
477 if (icsk
->icsk_ca_ops
->flags
& TCP_CONG_RTT_STAMP
)
478 __net_timestamp(skb
);
480 if (likely(clone_it
)) {
481 if (unlikely(skb_cloned(skb
)))
482 skb
= pskb_copy(skb
, gfp_mask
);
484 skb
= skb_clone(skb
, gfp_mask
);
491 tcb
= TCP_SKB_CB(skb
);
492 tcp_header_size
= tp
->tcp_header_len
;
494 #define SYSCTL_FLAG_TSTAMPS 0x1
495 #define SYSCTL_FLAG_WSCALE 0x2
496 #define SYSCTL_FLAG_SACK 0x4
499 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
500 tcp_header_size
= sizeof(struct tcphdr
) + TCPOLEN_MSS
;
501 if (sysctl_tcp_timestamps
) {
502 tcp_header_size
+= TCPOLEN_TSTAMP_ALIGNED
;
503 sysctl_flags
|= SYSCTL_FLAG_TSTAMPS
;
505 if (sysctl_tcp_window_scaling
) {
506 tcp_header_size
+= TCPOLEN_WSCALE_ALIGNED
;
507 sysctl_flags
|= SYSCTL_FLAG_WSCALE
;
509 if (sysctl_tcp_sack
) {
510 sysctl_flags
|= SYSCTL_FLAG_SACK
;
511 if (!(sysctl_flags
& SYSCTL_FLAG_TSTAMPS
))
512 tcp_header_size
+= TCPOLEN_SACKPERM_ALIGNED
;
514 } else if (unlikely(tp
->rx_opt
.eff_sacks
)) {
515 /* A SACK is 2 pad bytes, a 2 byte header, plus
516 * 2 32-bit sequence numbers for each SACK block.
518 tcp_header_size
+= (TCPOLEN_SACK_BASE_ALIGNED
+
519 (tp
->rx_opt
.eff_sacks
*
520 TCPOLEN_SACK_PERBLOCK
));
523 if (tcp_packets_in_flight(tp
) == 0)
524 tcp_ca_event(sk
, CA_EVENT_TX_START
);
526 #ifdef CONFIG_TCP_MD5SIG
528 * Are we doing MD5 on this segment? If so - make
531 md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
533 tcp_header_size
+= TCPOLEN_MD5SIG_ALIGNED
;
536 skb_push(skb
, tcp_header_size
);
537 skb_reset_transport_header(skb
);
538 skb_set_owner_w(skb
, sk
);
540 /* Build TCP header and checksum it. */
542 th
->source
= inet
->sport
;
543 th
->dest
= inet
->dport
;
544 th
->seq
= htonl(tcb
->seq
);
545 th
->ack_seq
= htonl(tp
->rcv_nxt
);
546 *(((__be16
*)th
) + 6) = htons(((tcp_header_size
>> 2) << 12) |
549 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
550 /* RFC1323: The window in SYN & SYN/ACK segments
553 th
->window
= htons(min(tp
->rcv_wnd
, 65535U));
555 th
->window
= htons(tcp_select_window(sk
));
560 if (unlikely(tp
->urg_mode
&&
561 between(tp
->snd_up
, tcb
->seq
+1, tcb
->seq
+0xFFFF))) {
562 th
->urg_ptr
= htons(tp
->snd_up
-tcb
->seq
);
566 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
567 tcp_syn_build_options((__be32
*)(th
+ 1),
568 tcp_advertise_mss(sk
),
569 (sysctl_flags
& SYSCTL_FLAG_TSTAMPS
),
570 (sysctl_flags
& SYSCTL_FLAG_SACK
),
571 (sysctl_flags
& SYSCTL_FLAG_WSCALE
),
572 tp
->rx_opt
.rcv_wscale
,
574 tp
->rx_opt
.ts_recent
,
576 #ifdef CONFIG_TCP_MD5SIG
577 md5
? &md5_hash_location
:
581 tcp_build_and_update_options((__be32
*)(th
+ 1),
583 #ifdef CONFIG_TCP_MD5SIG
584 md5
? &md5_hash_location
:
587 TCP_ECN_send(sk
, skb
, tcp_header_size
);
590 #ifdef CONFIG_TCP_MD5SIG
591 /* Calculate the MD5 hash, as we have all we need now */
593 tp
->af_specific
->calc_md5_hash(md5_hash_location
,
602 icsk
->icsk_af_ops
->send_check(sk
, skb
->len
, skb
);
604 if (likely(tcb
->flags
& TCPCB_FLAG_ACK
))
605 tcp_event_ack_sent(sk
, tcp_skb_pcount(skb
));
607 if (skb
->len
!= tcp_header_size
)
608 tcp_event_data_sent(tp
, skb
, sk
);
610 if (after(tcb
->end_seq
, tp
->snd_nxt
) || tcb
->seq
== tcb
->end_seq
)
611 TCP_INC_STATS(TCP_MIB_OUTSEGS
);
613 err
= icsk
->icsk_af_ops
->queue_xmit(skb
, 0);
614 if (likely(err
<= 0))
617 tcp_enter_cwr(sk
, 1);
619 return net_xmit_eval(err
);
621 #undef SYSCTL_FLAG_TSTAMPS
622 #undef SYSCTL_FLAG_WSCALE
623 #undef SYSCTL_FLAG_SACK
627 /* This routine just queue's the buffer
629 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
630 * otherwise socket can stall.
632 static void tcp_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
634 struct tcp_sock
*tp
= tcp_sk(sk
);
636 /* Advance write_seq and place onto the write_queue. */
637 tp
->write_seq
= TCP_SKB_CB(skb
)->end_seq
;
638 skb_header_release(skb
);
639 tcp_add_write_queue_tail(sk
, skb
);
640 sk
->sk_wmem_queued
+= skb
->truesize
;
641 sk_mem_charge(sk
, skb
->truesize
);
644 static void tcp_set_skb_tso_segs(struct sock
*sk
, struct sk_buff
*skb
, unsigned int mss_now
)
646 if (skb
->len
<= mss_now
|| !sk_can_gso(sk
)) {
647 /* Avoid the costly divide in the normal
650 skb_shinfo(skb
)->gso_segs
= 1;
651 skb_shinfo(skb
)->gso_size
= 0;
652 skb_shinfo(skb
)->gso_type
= 0;
654 skb_shinfo(skb
)->gso_segs
= DIV_ROUND_UP(skb
->len
, mss_now
);
655 skb_shinfo(skb
)->gso_size
= mss_now
;
656 skb_shinfo(skb
)->gso_type
= sk
->sk_gso_type
;
660 /* When a modification to fackets out becomes necessary, we need to check
661 * skb is counted to fackets_out or not.
663 static void tcp_adjust_fackets_out(struct sock
*sk
, struct sk_buff
*skb
,
666 struct tcp_sock
*tp
= tcp_sk(sk
);
668 if (!tp
->sacked_out
|| tcp_is_reno(tp
))
671 if (after(tcp_highest_sack_seq(tp
), TCP_SKB_CB(skb
)->seq
))
672 tp
->fackets_out
-= decr
;
675 /* Function to create two new TCP segments. Shrinks the given segment
676 * to the specified size and appends a new segment with the rest of the
677 * packet to the list. This won't be called frequently, I hope.
678 * Remember, these are still headerless SKBs at this point.
680 int tcp_fragment(struct sock
*sk
, struct sk_buff
*skb
, u32 len
, unsigned int mss_now
)
682 struct tcp_sock
*tp
= tcp_sk(sk
);
683 struct sk_buff
*buff
;
684 int nsize
, old_factor
;
688 BUG_ON(len
> skb
->len
);
690 tcp_clear_retrans_hints_partial(tp
);
691 nsize
= skb_headlen(skb
) - len
;
695 if (skb_cloned(skb
) &&
696 skb_is_nonlinear(skb
) &&
697 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
700 /* Get a new skb... force flag on. */
701 buff
= sk_stream_alloc_skb(sk
, nsize
, GFP_ATOMIC
);
703 return -ENOMEM
; /* We'll just try again later. */
705 sk
->sk_wmem_queued
+= buff
->truesize
;
706 sk_mem_charge(sk
, buff
->truesize
);
707 nlen
= skb
->len
- len
- nsize
;
708 buff
->truesize
+= nlen
;
709 skb
->truesize
-= nlen
;
711 /* Correct the sequence numbers. */
712 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
713 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
714 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
716 /* PSH and FIN should only be set in the second packet. */
717 flags
= TCP_SKB_CB(skb
)->flags
;
718 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
719 TCP_SKB_CB(buff
)->flags
= flags
;
720 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
721 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_AT_TAIL
;
723 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_PARTIAL
) {
724 /* Copy and checksum data tail into the new buffer. */
725 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
, skb_put(buff
, nsize
),
730 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
732 skb
->ip_summed
= CHECKSUM_PARTIAL
;
733 skb_split(skb
, buff
, len
);
736 buff
->ip_summed
= skb
->ip_summed
;
738 /* Looks stupid, but our code really uses when of
739 * skbs, which it never sent before. --ANK
741 TCP_SKB_CB(buff
)->when
= TCP_SKB_CB(skb
)->when
;
742 buff
->tstamp
= skb
->tstamp
;
744 old_factor
= tcp_skb_pcount(skb
);
746 /* Fix up tso_factor for both original and new SKB. */
747 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
748 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
750 /* If this packet has been sent out already, we must
751 * adjust the various packet counters.
753 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
754 int diff
= old_factor
- tcp_skb_pcount(skb
) -
755 tcp_skb_pcount(buff
);
757 tp
->packets_out
-= diff
;
759 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
760 tp
->sacked_out
-= diff
;
761 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
762 tp
->retrans_out
-= diff
;
764 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
)
765 tp
->lost_out
-= diff
;
767 /* Adjust Reno SACK estimate. */
768 if (tcp_is_reno(tp
) && diff
> 0) {
769 tcp_dec_pcount_approx_int(&tp
->sacked_out
, diff
);
770 tcp_verify_left_out(tp
);
772 tcp_adjust_fackets_out(sk
, skb
, diff
);
775 /* Link BUFF into the send queue. */
776 skb_header_release(buff
);
777 tcp_insert_write_queue_after(skb
, buff
, sk
);
782 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
783 * eventually). The difference is that pulled data not copied, but
784 * immediately discarded.
786 static void __pskb_trim_head(struct sk_buff
*skb
, int len
)
792 for (i
=0; i
<skb_shinfo(skb
)->nr_frags
; i
++) {
793 if (skb_shinfo(skb
)->frags
[i
].size
<= eat
) {
794 put_page(skb_shinfo(skb
)->frags
[i
].page
);
795 eat
-= skb_shinfo(skb
)->frags
[i
].size
;
797 skb_shinfo(skb
)->frags
[k
] = skb_shinfo(skb
)->frags
[i
];
799 skb_shinfo(skb
)->frags
[k
].page_offset
+= eat
;
800 skb_shinfo(skb
)->frags
[k
].size
-= eat
;
806 skb_shinfo(skb
)->nr_frags
= k
;
808 skb_reset_tail_pointer(skb
);
809 skb
->data_len
-= len
;
810 skb
->len
= skb
->data_len
;
813 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
815 if (skb_cloned(skb
) &&
816 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
819 /* If len == headlen, we avoid __skb_pull to preserve alignment. */
820 if (unlikely(len
< skb_headlen(skb
)))
821 __skb_pull(skb
, len
);
823 __pskb_trim_head(skb
, len
- skb_headlen(skb
));
825 TCP_SKB_CB(skb
)->seq
+= len
;
826 skb
->ip_summed
= CHECKSUM_PARTIAL
;
828 skb
->truesize
-= len
;
829 sk
->sk_wmem_queued
-= len
;
830 sk_mem_uncharge(sk
, len
);
831 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
833 /* Any change of skb->len requires recalculation of tso
836 if (tcp_skb_pcount(skb
) > 1)
837 tcp_set_skb_tso_segs(sk
, skb
, tcp_current_mss(sk
, 1));
842 /* Not accounting for SACKs here. */
843 int tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
845 struct tcp_sock
*tp
= tcp_sk(sk
);
846 struct inet_connection_sock
*icsk
= inet_csk(sk
);
849 /* Calculate base mss without TCP options:
850 It is MMS_S - sizeof(tcphdr) of rfc1122
852 mss_now
= pmtu
- icsk
->icsk_af_ops
->net_header_len
- sizeof(struct tcphdr
);
854 /* Clamp it (mss_clamp does not include tcp options) */
855 if (mss_now
> tp
->rx_opt
.mss_clamp
)
856 mss_now
= tp
->rx_opt
.mss_clamp
;
858 /* Now subtract optional transport overhead */
859 mss_now
-= icsk
->icsk_ext_hdr_len
;
861 /* Then reserve room for full set of TCP options and 8 bytes of data */
865 /* Now subtract TCP options size, not including SACKs */
866 mss_now
-= tp
->tcp_header_len
- sizeof(struct tcphdr
);
871 /* Inverse of above */
872 int tcp_mss_to_mtu(struct sock
*sk
, int mss
)
874 struct tcp_sock
*tp
= tcp_sk(sk
);
875 struct inet_connection_sock
*icsk
= inet_csk(sk
);
880 icsk
->icsk_ext_hdr_len
+
881 icsk
->icsk_af_ops
->net_header_len
;
886 void tcp_mtup_init(struct sock
*sk
)
888 struct tcp_sock
*tp
= tcp_sk(sk
);
889 struct inet_connection_sock
*icsk
= inet_csk(sk
);
891 icsk
->icsk_mtup
.enabled
= sysctl_tcp_mtu_probing
> 1;
892 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+ sizeof(struct tcphdr
) +
893 icsk
->icsk_af_ops
->net_header_len
;
894 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, sysctl_tcp_base_mss
);
895 icsk
->icsk_mtup
.probe_size
= 0;
898 /* This function synchronize snd mss to current pmtu/exthdr set.
900 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
901 for TCP options, but includes only bare TCP header.
903 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
904 It is minimum of user_mss and mss received with SYN.
905 It also does not include TCP options.
907 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
909 tp->mss_cache is current effective sending mss, including
910 all tcp options except for SACKs. It is evaluated,
911 taking into account current pmtu, but never exceeds
912 tp->rx_opt.mss_clamp.
914 NOTE1. rfc1122 clearly states that advertised MSS
915 DOES NOT include either tcp or ip options.
917 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
918 are READ ONLY outside this function. --ANK (980731)
921 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
923 struct tcp_sock
*tp
= tcp_sk(sk
);
924 struct inet_connection_sock
*icsk
= inet_csk(sk
);
927 if (icsk
->icsk_mtup
.search_high
> pmtu
)
928 icsk
->icsk_mtup
.search_high
= pmtu
;
930 mss_now
= tcp_mtu_to_mss(sk
, pmtu
);
932 /* Bound mss with half of window */
933 if (tp
->max_window
&& mss_now
> (tp
->max_window
>>1))
934 mss_now
= max((tp
->max_window
>>1), 68U - tp
->tcp_header_len
);
936 /* And store cached results */
937 icsk
->icsk_pmtu_cookie
= pmtu
;
938 if (icsk
->icsk_mtup
.enabled
)
939 mss_now
= min(mss_now
, tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
));
940 tp
->mss_cache
= mss_now
;
945 /* Compute the current effective MSS, taking SACKs and IP options,
946 * and even PMTU discovery events into account.
948 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
949 * cannot be large. However, taking into account rare use of URG, this
952 unsigned int tcp_current_mss(struct sock
*sk
, int large_allowed
)
954 struct tcp_sock
*tp
= tcp_sk(sk
);
955 struct dst_entry
*dst
= __sk_dst_get(sk
);
960 mss_now
= tp
->mss_cache
;
962 if (large_allowed
&& sk_can_gso(sk
) && !tp
->urg_mode
)
966 u32 mtu
= dst_mtu(dst
);
967 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
968 mss_now
= tcp_sync_mss(sk
, mtu
);
971 if (tp
->rx_opt
.eff_sacks
)
972 mss_now
-= (TCPOLEN_SACK_BASE_ALIGNED
+
973 (tp
->rx_opt
.eff_sacks
* TCPOLEN_SACK_PERBLOCK
));
975 #ifdef CONFIG_TCP_MD5SIG
976 if (tp
->af_specific
->md5_lookup(sk
, sk
))
977 mss_now
-= TCPOLEN_MD5SIG_ALIGNED
;
980 xmit_size_goal
= mss_now
;
983 xmit_size_goal
= (65535 -
984 inet_csk(sk
)->icsk_af_ops
->net_header_len
-
985 inet_csk(sk
)->icsk_ext_hdr_len
-
988 if (tp
->max_window
&&
989 (xmit_size_goal
> (tp
->max_window
>> 1)))
990 xmit_size_goal
= max((tp
->max_window
>> 1),
991 68U - tp
->tcp_header_len
);
993 xmit_size_goal
-= (xmit_size_goal
% mss_now
);
995 tp
->xmit_size_goal
= xmit_size_goal
;
1000 /* Congestion window validation. (RFC2861) */
1002 static void tcp_cwnd_validate(struct sock
*sk
)
1004 struct tcp_sock
*tp
= tcp_sk(sk
);
1005 __u32 packets_out
= tp
->packets_out
;
1007 if (packets_out
>= tp
->snd_cwnd
) {
1008 /* Network is feed fully. */
1009 tp
->snd_cwnd_used
= 0;
1010 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1012 /* Network starves. */
1013 if (tp
->packets_out
> tp
->snd_cwnd_used
)
1014 tp
->snd_cwnd_used
= tp
->packets_out
;
1016 if (sysctl_tcp_slow_start_after_idle
&&
1017 (s32
)(tcp_time_stamp
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
)
1018 tcp_cwnd_application_limited(sk
);
1022 /* Returns the portion of skb which can be sent right away without
1023 * introducing MSS oddities to segment boundaries. In rare cases where
1024 * mss_now != mss_cache, we will request caller to create a small skb
1025 * per input skb which could be mostly avoided here (if desired).
1027 static unsigned int tcp_mss_split_point(struct sock
*sk
, struct sk_buff
*skb
,
1028 unsigned int mss_now
,
1031 struct tcp_sock
*tp
= tcp_sk(sk
);
1032 u32 needed
, window
, cwnd_len
;
1034 window
= (tp
->snd_una
+ tp
->snd_wnd
- TCP_SKB_CB(skb
)->seq
);
1035 cwnd_len
= mss_now
* cwnd
;
1037 if (likely(cwnd_len
<= window
&& skb
!= tcp_write_queue_tail(sk
)))
1040 if (skb
== tcp_write_queue_tail(sk
) && cwnd_len
<= skb
->len
)
1043 needed
= min(skb
->len
, window
);
1044 return needed
- needed
% mss_now
;
1047 /* Can at least one segment of SKB be sent right now, according to the
1048 * congestion window rules? If so, return how many segments are allowed.
1050 static inline unsigned int tcp_cwnd_test(struct tcp_sock
*tp
, struct sk_buff
*skb
)
1052 u32 in_flight
, cwnd
;
1054 /* Don't be strict about the congestion window for the final FIN. */
1055 if ((TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
) &&
1056 tcp_skb_pcount(skb
) == 1)
1059 in_flight
= tcp_packets_in_flight(tp
);
1060 cwnd
= tp
->snd_cwnd
;
1061 if (in_flight
< cwnd
)
1062 return (cwnd
- in_flight
);
1067 /* This must be invoked the first time we consider transmitting
1068 * SKB onto the wire.
1070 static int tcp_init_tso_segs(struct sock
*sk
, struct sk_buff
*skb
, unsigned int mss_now
)
1072 int tso_segs
= tcp_skb_pcount(skb
);
1076 tcp_skb_mss(skb
) != mss_now
)) {
1077 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1078 tso_segs
= tcp_skb_pcount(skb
);
1083 static inline int tcp_minshall_check(const struct tcp_sock
*tp
)
1085 return after(tp
->snd_sml
,tp
->snd_una
) &&
1086 !after(tp
->snd_sml
, tp
->snd_nxt
);
1089 /* Return 0, if packet can be sent now without violation Nagle's rules:
1090 * 1. It is full sized.
1091 * 2. Or it contains FIN. (already checked by caller)
1092 * 3. Or TCP_NODELAY was set.
1093 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1094 * With Minshall's modification: all sent small packets are ACKed.
1097 static inline int tcp_nagle_check(const struct tcp_sock
*tp
,
1098 const struct sk_buff
*skb
,
1099 unsigned mss_now
, int nonagle
)
1101 return (skb
->len
< mss_now
&&
1102 ((nonagle
&TCP_NAGLE_CORK
) ||
1105 tcp_minshall_check(tp
))));
1108 /* Return non-zero if the Nagle test allows this packet to be
1111 static inline int tcp_nagle_test(struct tcp_sock
*tp
, struct sk_buff
*skb
,
1112 unsigned int cur_mss
, int nonagle
)
1114 /* Nagle rule does not apply to frames, which sit in the middle of the
1115 * write_queue (they have no chances to get new data).
1117 * This is implemented in the callers, where they modify the 'nonagle'
1118 * argument based upon the location of SKB in the send queue.
1120 if (nonagle
& TCP_NAGLE_PUSH
)
1123 /* Don't use the nagle rule for urgent data (or for the final FIN).
1124 * Nagle can be ignored during F-RTO too (see RFC4138).
1126 if (tp
->urg_mode
|| (tp
->frto_counter
== 2) ||
1127 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
))
1130 if (!tcp_nagle_check(tp
, skb
, cur_mss
, nonagle
))
1136 /* Does at least the first segment of SKB fit into the send window? */
1137 static inline int tcp_snd_wnd_test(struct tcp_sock
*tp
, struct sk_buff
*skb
, unsigned int cur_mss
)
1139 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1141 if (skb
->len
> cur_mss
)
1142 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
1144 return !after(end_seq
, tp
->snd_una
+ tp
->snd_wnd
);
1147 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1148 * should be put on the wire right now. If so, it returns the number of
1149 * packets allowed by the congestion window.
1151 static unsigned int tcp_snd_test(struct sock
*sk
, struct sk_buff
*skb
,
1152 unsigned int cur_mss
, int nonagle
)
1154 struct tcp_sock
*tp
= tcp_sk(sk
);
1155 unsigned int cwnd_quota
;
1157 tcp_init_tso_segs(sk
, skb
, cur_mss
);
1159 if (!tcp_nagle_test(tp
, skb
, cur_mss
, nonagle
))
1162 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1164 !tcp_snd_wnd_test(tp
, skb
, cur_mss
))
1170 int tcp_may_send_now(struct sock
*sk
)
1172 struct tcp_sock
*tp
= tcp_sk(sk
);
1173 struct sk_buff
*skb
= tcp_send_head(sk
);
1176 tcp_snd_test(sk
, skb
, tcp_current_mss(sk
, 1),
1177 (tcp_skb_is_last(sk
, skb
) ?
1178 tp
->nonagle
: TCP_NAGLE_PUSH
)));
1181 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1182 * which is put after SKB on the list. It is very much like
1183 * tcp_fragment() except that it may make several kinds of assumptions
1184 * in order to speed up the splitting operation. In particular, we
1185 * know that all the data is in scatter-gather pages, and that the
1186 * packet has never been sent out before (and thus is not cloned).
1188 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
, unsigned int mss_now
)
1190 struct sk_buff
*buff
;
1191 int nlen
= skb
->len
- len
;
1194 /* All of a TSO frame must be composed of paged data. */
1195 if (skb
->len
!= skb
->data_len
)
1196 return tcp_fragment(sk
, skb
, len
, mss_now
);
1198 buff
= sk_stream_alloc_skb(sk
, 0, GFP_ATOMIC
);
1199 if (unlikely(buff
== NULL
))
1202 sk
->sk_wmem_queued
+= buff
->truesize
;
1203 sk_mem_charge(sk
, buff
->truesize
);
1204 buff
->truesize
+= nlen
;
1205 skb
->truesize
-= nlen
;
1207 /* Correct the sequence numbers. */
1208 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1209 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1210 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1212 /* PSH and FIN should only be set in the second packet. */
1213 flags
= TCP_SKB_CB(skb
)->flags
;
1214 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
1215 TCP_SKB_CB(buff
)->flags
= flags
;
1217 /* This packet was never sent out yet, so no SACK bits. */
1218 TCP_SKB_CB(buff
)->sacked
= 0;
1220 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_PARTIAL
;
1221 skb_split(skb
, buff
, len
);
1223 /* Fix up tso_factor for both original and new SKB. */
1224 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1225 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1227 /* Link BUFF into the send queue. */
1228 skb_header_release(buff
);
1229 tcp_insert_write_queue_after(skb
, buff
, sk
);
1234 /* Try to defer sending, if possible, in order to minimize the amount
1235 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1237 * This algorithm is from John Heffner.
1239 static int tcp_tso_should_defer(struct sock
*sk
, struct sk_buff
*skb
)
1241 struct tcp_sock
*tp
= tcp_sk(sk
);
1242 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1243 u32 send_win
, cong_win
, limit
, in_flight
;
1245 if (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
)
1248 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
1251 /* Defer for less than two clock ticks. */
1252 if (tp
->tso_deferred
&&
1253 ((jiffies
<< 1) >> 1) - (tp
->tso_deferred
>> 1) > 1)
1256 in_flight
= tcp_packets_in_flight(tp
);
1258 BUG_ON(tcp_skb_pcount(skb
) <= 1 ||
1259 (tp
->snd_cwnd
<= in_flight
));
1261 send_win
= (tp
->snd_una
+ tp
->snd_wnd
) - TCP_SKB_CB(skb
)->seq
;
1263 /* From in_flight test above, we know that cwnd > in_flight. */
1264 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
1266 limit
= min(send_win
, cong_win
);
1268 /* If a full-sized TSO skb can be sent, do it. */
1272 if (sysctl_tcp_tso_win_divisor
) {
1273 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
1275 /* If at least some fraction of a window is available,
1278 chunk
/= sysctl_tcp_tso_win_divisor
;
1282 /* Different approach, try not to defer past a single
1283 * ACK. Receiver should ACK every other full sized
1284 * frame, so if we have space for more than 3 frames
1287 if (limit
> tcp_max_burst(tp
) * tp
->mss_cache
)
1291 /* Ok, it looks like it is advisable to defer. */
1292 tp
->tso_deferred
= 1 | (jiffies
<<1);
1297 tp
->tso_deferred
= 0;
1301 /* Create a new MTU probe if we are ready.
1302 * Returns 0 if we should wait to probe (no cwnd available),
1303 * 1 if a probe was sent,
1305 static int tcp_mtu_probe(struct sock
*sk
)
1307 struct tcp_sock
*tp
= tcp_sk(sk
);
1308 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1309 struct sk_buff
*skb
, *nskb
, *next
;
1316 /* Not currently probing/verifying,
1318 * have enough cwnd, and
1319 * not SACKing (the variable headers throw things off) */
1320 if (!icsk
->icsk_mtup
.enabled
||
1321 icsk
->icsk_mtup
.probe_size
||
1322 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
||
1323 tp
->snd_cwnd
< 11 ||
1324 tp
->rx_opt
.eff_sacks
)
1327 /* Very simple search strategy: just double the MSS. */
1328 mss_now
= tcp_current_mss(sk
, 0);
1329 probe_size
= 2*tp
->mss_cache
;
1330 size_needed
= probe_size
+ (tp
->reordering
+ 1) * tp
->mss_cache
;
1331 if (probe_size
> tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_high
)) {
1332 /* TODO: set timer for probe_converge_event */
1336 /* Have enough data in the send queue to probe? */
1337 if (tp
->write_seq
- tp
->snd_nxt
< size_needed
)
1340 if (tp
->snd_wnd
< size_needed
)
1342 if (after(tp
->snd_nxt
+ size_needed
, tp
->snd_una
+ tp
->snd_wnd
))
1345 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1346 if (tcp_packets_in_flight(tp
) + 2 > tp
->snd_cwnd
) {
1347 if (!tcp_packets_in_flight(tp
))
1353 /* We're allowed to probe. Build it now. */
1354 if ((nskb
= sk_stream_alloc_skb(sk
, probe_size
, GFP_ATOMIC
)) == NULL
)
1356 sk
->sk_wmem_queued
+= nskb
->truesize
;
1357 sk_mem_charge(sk
, nskb
->truesize
);
1359 skb
= tcp_send_head(sk
);
1361 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(skb
)->seq
;
1362 TCP_SKB_CB(nskb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ probe_size
;
1363 TCP_SKB_CB(nskb
)->flags
= TCPCB_FLAG_ACK
;
1364 TCP_SKB_CB(nskb
)->sacked
= 0;
1366 nskb
->ip_summed
= skb
->ip_summed
;
1368 tcp_insert_write_queue_before(nskb
, skb
, sk
);
1371 tcp_for_write_queue_from_safe(skb
, next
, sk
) {
1372 copy
= min_t(int, skb
->len
, probe_size
- len
);
1373 if (nskb
->ip_summed
)
1374 skb_copy_bits(skb
, 0, skb_put(nskb
, copy
), copy
);
1376 nskb
->csum
= skb_copy_and_csum_bits(skb
, 0,
1377 skb_put(nskb
, copy
), copy
, nskb
->csum
);
1379 if (skb
->len
<= copy
) {
1380 /* We've eaten all the data from this skb.
1382 TCP_SKB_CB(nskb
)->flags
|= TCP_SKB_CB(skb
)->flags
;
1383 tcp_unlink_write_queue(skb
, sk
);
1384 sk_wmem_free_skb(sk
, skb
);
1386 TCP_SKB_CB(nskb
)->flags
|= TCP_SKB_CB(skb
)->flags
&
1387 ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
1388 if (!skb_shinfo(skb
)->nr_frags
) {
1389 skb_pull(skb
, copy
);
1390 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1391 skb
->csum
= csum_partial(skb
->data
, skb
->len
, 0);
1393 __pskb_trim_head(skb
, copy
);
1394 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1396 TCP_SKB_CB(skb
)->seq
+= copy
;
1401 if (len
>= probe_size
)
1404 tcp_init_tso_segs(sk
, nskb
, nskb
->len
);
1406 /* We're ready to send. If this fails, the probe will
1407 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1408 TCP_SKB_CB(nskb
)->when
= tcp_time_stamp
;
1409 if (!tcp_transmit_skb(sk
, nskb
, 1, GFP_ATOMIC
)) {
1410 /* Decrement cwnd here because we are sending
1411 * effectively two packets. */
1413 update_send_head(sk
, nskb
);
1415 icsk
->icsk_mtup
.probe_size
= tcp_mss_to_mtu(sk
, nskb
->len
);
1416 tp
->mtu_probe
.probe_seq_start
= TCP_SKB_CB(nskb
)->seq
;
1417 tp
->mtu_probe
.probe_seq_end
= TCP_SKB_CB(nskb
)->end_seq
;
1426 /* This routine writes packets to the network. It advances the
1427 * send_head. This happens as incoming acks open up the remote
1430 * Returns 1, if no segments are in flight and we have queued segments, but
1431 * cannot send anything now because of SWS or another problem.
1433 static int tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
)
1435 struct tcp_sock
*tp
= tcp_sk(sk
);
1436 struct sk_buff
*skb
;
1437 unsigned int tso_segs
, sent_pkts
;
1441 /* If we are closed, the bytes will have to remain here.
1442 * In time closedown will finish, we empty the write queue and all
1445 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
1450 /* Do MTU probing. */
1451 if ((result
= tcp_mtu_probe(sk
)) == 0) {
1453 } else if (result
> 0) {
1457 while ((skb
= tcp_send_head(sk
))) {
1460 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1463 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1467 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
)))
1470 if (tso_segs
== 1) {
1471 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
1472 (tcp_skb_is_last(sk
, skb
) ?
1473 nonagle
: TCP_NAGLE_PUSH
))))
1476 if (tcp_tso_should_defer(sk
, skb
))
1482 limit
= tcp_mss_split_point(sk
, skb
, mss_now
,
1485 if (skb
->len
> limit
&&
1486 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
)))
1489 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1491 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
)))
1494 /* Advance the send_head. This one is sent out.
1495 * This call will increment packets_out.
1497 update_send_head(sk
, skb
);
1499 tcp_minshall_update(tp
, mss_now
, skb
);
1503 if (likely(sent_pkts
)) {
1504 tcp_cwnd_validate(sk
);
1507 return !tp
->packets_out
&& tcp_send_head(sk
);
1510 /* Push out any pending frames which were held back due to
1511 * TCP_CORK or attempt at coalescing tiny packets.
1512 * The socket must be locked by the caller.
1514 void __tcp_push_pending_frames(struct sock
*sk
, unsigned int cur_mss
,
1517 struct sk_buff
*skb
= tcp_send_head(sk
);
1520 if (tcp_write_xmit(sk
, cur_mss
, nonagle
))
1521 tcp_check_probe_timer(sk
);
1525 /* Send _single_ skb sitting at the send head. This function requires
1526 * true push pending frames to setup probe timer etc.
1528 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
1530 struct sk_buff
*skb
= tcp_send_head(sk
);
1531 unsigned int tso_segs
, cwnd_quota
;
1533 BUG_ON(!skb
|| skb
->len
< mss_now
);
1535 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1536 cwnd_quota
= tcp_snd_test(sk
, skb
, mss_now
, TCP_NAGLE_PUSH
);
1538 if (likely(cwnd_quota
)) {
1545 limit
= tcp_mss_split_point(sk
, skb
, mss_now
,
1548 if (skb
->len
> limit
&&
1549 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
)))
1552 /* Send it out now. */
1553 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1555 if (likely(!tcp_transmit_skb(sk
, skb
, 1, sk
->sk_allocation
))) {
1556 update_send_head(sk
, skb
);
1557 tcp_cwnd_validate(sk
);
1563 /* This function returns the amount that we can raise the
1564 * usable window based on the following constraints
1566 * 1. The window can never be shrunk once it is offered (RFC 793)
1567 * 2. We limit memory per socket
1570 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1571 * RECV.NEXT + RCV.WIN fixed until:
1572 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1574 * i.e. don't raise the right edge of the window until you can raise
1575 * it at least MSS bytes.
1577 * Unfortunately, the recommended algorithm breaks header prediction,
1578 * since header prediction assumes th->window stays fixed.
1580 * Strictly speaking, keeping th->window fixed violates the receiver
1581 * side SWS prevention criteria. The problem is that under this rule
1582 * a stream of single byte packets will cause the right side of the
1583 * window to always advance by a single byte.
1585 * Of course, if the sender implements sender side SWS prevention
1586 * then this will not be a problem.
1588 * BSD seems to make the following compromise:
1590 * If the free space is less than the 1/4 of the maximum
1591 * space available and the free space is less than 1/2 mss,
1592 * then set the window to 0.
1593 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1594 * Otherwise, just prevent the window from shrinking
1595 * and from being larger than the largest representable value.
1597 * This prevents incremental opening of the window in the regime
1598 * where TCP is limited by the speed of the reader side taking
1599 * data out of the TCP receive queue. It does nothing about
1600 * those cases where the window is constrained on the sender side
1601 * because the pipeline is full.
1603 * BSD also seems to "accidentally" limit itself to windows that are a
1604 * multiple of MSS, at least until the free space gets quite small.
1605 * This would appear to be a side effect of the mbuf implementation.
1606 * Combining these two algorithms results in the observed behavior
1607 * of having a fixed window size at almost all times.
1609 * Below we obtain similar behavior by forcing the offered window to
1610 * a multiple of the mss when it is feasible to do so.
1612 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1613 * Regular options like TIMESTAMP are taken into account.
1615 u32
__tcp_select_window(struct sock
*sk
)
1617 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1618 struct tcp_sock
*tp
= tcp_sk(sk
);
1619 /* MSS for the peer's data. Previous versions used mss_clamp
1620 * here. I don't know if the value based on our guesses
1621 * of peer's MSS is better for the performance. It's more correct
1622 * but may be worse for the performance because of rcv_mss
1623 * fluctuations. --SAW 1998/11/1
1625 int mss
= icsk
->icsk_ack
.rcv_mss
;
1626 int free_space
= tcp_space(sk
);
1627 int full_space
= min_t(int, tp
->window_clamp
, tcp_full_space(sk
));
1630 if (mss
> full_space
)
1633 if (free_space
< (full_space
>> 1)) {
1634 icsk
->icsk_ack
.quick
= 0;
1636 if (tcp_memory_pressure
)
1637 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U*tp
->advmss
);
1639 if (free_space
< mss
)
1643 if (free_space
> tp
->rcv_ssthresh
)
1644 free_space
= tp
->rcv_ssthresh
;
1646 /* Don't do rounding if we are using window scaling, since the
1647 * scaled window will not line up with the MSS boundary anyway.
1649 window
= tp
->rcv_wnd
;
1650 if (tp
->rx_opt
.rcv_wscale
) {
1651 window
= free_space
;
1653 /* Advertise enough space so that it won't get scaled away.
1654 * Import case: prevent zero window announcement if
1655 * 1<<rcv_wscale > mss.
1657 if (((window
>> tp
->rx_opt
.rcv_wscale
) << tp
->rx_opt
.rcv_wscale
) != window
)
1658 window
= (((window
>> tp
->rx_opt
.rcv_wscale
) + 1)
1659 << tp
->rx_opt
.rcv_wscale
);
1661 /* Get the largest window that is a nice multiple of mss.
1662 * Window clamp already applied above.
1663 * If our current window offering is within 1 mss of the
1664 * free space we just keep it. This prevents the divide
1665 * and multiply from happening most of the time.
1666 * We also don't do any window rounding when the free space
1669 if (window
<= free_space
- mss
|| window
> free_space
)
1670 window
= (free_space
/mss
)*mss
;
1671 else if (mss
== full_space
&&
1672 free_space
> window
+ (full_space
>> 1))
1673 window
= free_space
;
1679 /* Attempt to collapse two adjacent SKB's during retransmission. */
1680 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*skb
, int mss_now
)
1682 struct tcp_sock
*tp
= tcp_sk(sk
);
1683 struct sk_buff
*next_skb
= tcp_write_queue_next(sk
, skb
);
1685 /* The first test we must make is that neither of these two
1686 * SKB's are still referenced by someone else.
1688 if (!skb_cloned(skb
) && !skb_cloned(next_skb
)) {
1689 int skb_size
= skb
->len
, next_skb_size
= next_skb
->len
;
1690 u16 flags
= TCP_SKB_CB(skb
)->flags
;
1692 /* Also punt if next skb has been SACK'd. */
1693 if (TCP_SKB_CB(next_skb
)->sacked
& TCPCB_SACKED_ACKED
)
1696 /* Next skb is out of window. */
1697 if (after(TCP_SKB_CB(next_skb
)->end_seq
, tp
->snd_una
+tp
->snd_wnd
))
1700 /* Punt if not enough space exists in the first SKB for
1701 * the data in the second, or the total combined payload
1702 * would exceed the MSS.
1704 if ((next_skb_size
> skb_tailroom(skb
)) ||
1705 ((skb_size
+ next_skb_size
) > mss_now
))
1708 BUG_ON(tcp_skb_pcount(skb
) != 1 ||
1709 tcp_skb_pcount(next_skb
) != 1);
1711 tcp_highest_sack_combine(sk
, next_skb
, skb
);
1713 /* Ok. We will be able to collapse the packet. */
1714 tcp_unlink_write_queue(next_skb
, sk
);
1716 skb_copy_from_linear_data(next_skb
,
1717 skb_put(skb
, next_skb_size
),
1720 if (next_skb
->ip_summed
== CHECKSUM_PARTIAL
)
1721 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1723 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1724 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
1726 /* Update sequence range on original skb. */
1727 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
1729 /* Merge over control information. */
1730 flags
|= TCP_SKB_CB(next_skb
)->flags
; /* This moves PSH/FIN etc. over */
1731 TCP_SKB_CB(skb
)->flags
= flags
;
1733 /* All done, get rid of second SKB and account for it so
1734 * packet counting does not break.
1736 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
&(TCPCB_EVER_RETRANS
|TCPCB_AT_TAIL
);
1737 if (TCP_SKB_CB(next_skb
)->sacked
&TCPCB_SACKED_RETRANS
)
1738 tp
->retrans_out
-= tcp_skb_pcount(next_skb
);
1739 if (TCP_SKB_CB(next_skb
)->sacked
&TCPCB_LOST
)
1740 tp
->lost_out
-= tcp_skb_pcount(next_skb
);
1741 /* Reno case is special. Sigh... */
1742 if (tcp_is_reno(tp
) && tp
->sacked_out
)
1743 tcp_dec_pcount_approx(&tp
->sacked_out
, next_skb
);
1745 tcp_adjust_fackets_out(sk
, next_skb
, tcp_skb_pcount(next_skb
));
1746 tp
->packets_out
-= tcp_skb_pcount(next_skb
);
1748 /* changed transmit queue under us so clear hints */
1749 tcp_clear_retrans_hints_partial(tp
);
1751 sk_wmem_free_skb(sk
, next_skb
);
1755 /* Do a simple retransmit without using the backoff mechanisms in
1756 * tcp_timer. This is used for path mtu discovery.
1757 * The socket is already locked here.
1759 void tcp_simple_retransmit(struct sock
*sk
)
1761 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1762 struct tcp_sock
*tp
= tcp_sk(sk
);
1763 struct sk_buff
*skb
;
1764 unsigned int mss
= tcp_current_mss(sk
, 0);
1767 tcp_for_write_queue(skb
, sk
) {
1768 if (skb
== tcp_send_head(sk
))
1770 if (skb
->len
> mss
&&
1771 !(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
)) {
1772 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) {
1773 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1774 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1776 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_LOST
)) {
1777 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1778 tp
->lost_out
+= tcp_skb_pcount(skb
);
1784 tcp_clear_all_retrans_hints(tp
);
1789 tcp_verify_left_out(tp
);
1791 /* Don't muck with the congestion window here.
1792 * Reason is that we do not increase amount of _data_
1793 * in network, but units changed and effective
1794 * cwnd/ssthresh really reduced now.
1796 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
1797 tp
->high_seq
= tp
->snd_nxt
;
1798 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
1799 tp
->prior_ssthresh
= 0;
1800 tp
->undo_marker
= 0;
1801 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1803 tcp_xmit_retransmit_queue(sk
);
1806 /* This retransmits one SKB. Policy decisions and retransmit queue
1807 * state updates are done by the caller. Returns non-zero if an
1808 * error occurred which prevented the send.
1810 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
1812 struct tcp_sock
*tp
= tcp_sk(sk
);
1813 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1814 unsigned int cur_mss
= tcp_current_mss(sk
, 0);
1817 /* Inconslusive MTU probe */
1818 if (icsk
->icsk_mtup
.probe_size
) {
1819 icsk
->icsk_mtup
.probe_size
= 0;
1822 /* Do not sent more than we queued. 1/4 is reserved for possible
1823 * copying overhead: fragmentation, tunneling, mangling etc.
1825 if (atomic_read(&sk
->sk_wmem_alloc
) >
1826 min(sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2), sk
->sk_sndbuf
))
1829 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
1830 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1832 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
1836 /* If receiver has shrunk his window, and skb is out of
1837 * new window, do not retransmit it. The exception is the
1838 * case, when window is shrunk to zero. In this case
1839 * our retransmit serves as a zero window probe.
1841 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
+tp
->snd_wnd
)
1842 && TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
1845 if (skb
->len
> cur_mss
) {
1846 if (tcp_fragment(sk
, skb
, cur_mss
, cur_mss
))
1847 return -ENOMEM
; /* We'll try again later. */
1850 /* Collapse two adjacent packets if worthwhile and we can. */
1851 if (!(TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_SYN
) &&
1852 (skb
->len
< (cur_mss
>> 1)) &&
1853 (tcp_write_queue_next(sk
, skb
) != tcp_send_head(sk
)) &&
1854 (!tcp_skb_is_last(sk
, skb
)) &&
1855 (skb_shinfo(skb
)->nr_frags
== 0 && skb_shinfo(tcp_write_queue_next(sk
, skb
))->nr_frags
== 0) &&
1856 (tcp_skb_pcount(skb
) == 1 && tcp_skb_pcount(tcp_write_queue_next(sk
, skb
)) == 1) &&
1857 (sysctl_tcp_retrans_collapse
!= 0))
1858 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
1860 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
1861 return -EHOSTUNREACH
; /* Routing failure or similar. */
1863 /* Some Solaris stacks overoptimize and ignore the FIN on a
1864 * retransmit when old data is attached. So strip it off
1865 * since it is cheap to do so and saves bytes on the network.
1868 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
) &&
1869 tp
->snd_una
== (TCP_SKB_CB(skb
)->end_seq
- 1)) {
1870 if (!pskb_trim(skb
, 0)) {
1871 TCP_SKB_CB(skb
)->seq
= TCP_SKB_CB(skb
)->end_seq
- 1;
1872 skb_shinfo(skb
)->gso_segs
= 1;
1873 skb_shinfo(skb
)->gso_size
= 0;
1874 skb_shinfo(skb
)->gso_type
= 0;
1875 skb
->ip_summed
= CHECKSUM_NONE
;
1880 /* Make a copy, if the first transmission SKB clone we made
1881 * is still in somebody's hands, else make a clone.
1883 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1885 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
1888 /* Update global TCP statistics. */
1889 TCP_INC_STATS(TCP_MIB_RETRANSSEGS
);
1891 tp
->total_retrans
++;
1893 #if FASTRETRANS_DEBUG > 0
1894 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) {
1895 if (net_ratelimit())
1896 printk(KERN_DEBUG
"retrans_out leaked.\n");
1899 if (!tp
->retrans_out
)
1900 tp
->lost_retrans_low
= tp
->snd_nxt
;
1901 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
1902 tp
->retrans_out
+= tcp_skb_pcount(skb
);
1904 /* Save stamp of the first retransmit. */
1905 if (!tp
->retrans_stamp
)
1906 tp
->retrans_stamp
= TCP_SKB_CB(skb
)->when
;
1910 /* snd_nxt is stored to detect loss of retransmitted segment,
1911 * see tcp_input.c tcp_sacktag_write_queue().
1913 TCP_SKB_CB(skb
)->ack_seq
= tp
->snd_nxt
;
1918 /* This gets called after a retransmit timeout, and the initially
1919 * retransmitted data is acknowledged. It tries to continue
1920 * resending the rest of the retransmit queue, until either
1921 * we've sent it all or the congestion window limit is reached.
1922 * If doing SACK, the first ACK which comes back for a timeout
1923 * based retransmit packet might feed us FACK information again.
1924 * If so, we use it to avoid unnecessarily retransmissions.
1926 void tcp_xmit_retransmit_queue(struct sock
*sk
)
1928 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1929 struct tcp_sock
*tp
= tcp_sk(sk
);
1930 struct sk_buff
*skb
;
1933 if (tp
->retransmit_skb_hint
) {
1934 skb
= tp
->retransmit_skb_hint
;
1935 packet_cnt
= tp
->retransmit_cnt_hint
;
1937 skb
= tcp_write_queue_head(sk
);
1941 /* First pass: retransmit lost packets. */
1943 tcp_for_write_queue_from(skb
, sk
) {
1944 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
1946 if (skb
== tcp_send_head(sk
))
1948 /* we could do better than to assign each time */
1949 tp
->retransmit_skb_hint
= skb
;
1950 tp
->retransmit_cnt_hint
= packet_cnt
;
1952 /* Assume this retransmit will generate
1953 * only one packet for congestion window
1954 * calculation purposes. This works because
1955 * tcp_retransmit_skb() will chop up the
1956 * packet to be MSS sized and all the
1957 * packet counting works out.
1959 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
1962 if (sacked
& TCPCB_LOST
) {
1963 if (!(sacked
&(TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))) {
1964 if (tcp_retransmit_skb(sk
, skb
)) {
1965 tp
->retransmit_skb_hint
= NULL
;
1968 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
1969 NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS
);
1971 NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS
);
1973 if (skb
== tcp_write_queue_head(sk
))
1974 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1975 inet_csk(sk
)->icsk_rto
,
1979 packet_cnt
+= tcp_skb_pcount(skb
);
1980 if (packet_cnt
>= tp
->lost_out
)
1986 /* OK, demanded retransmission is finished. */
1988 /* Forward retransmissions are possible only during Recovery. */
1989 if (icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
1992 /* No forward retransmissions in Reno are possible. */
1993 if (tcp_is_reno(tp
))
1996 /* Yeah, we have to make difficult choice between forward transmission
1997 * and retransmission... Both ways have their merits...
1999 * For now we do not retransmit anything, while we have some new
2000 * segments to send. In the other cases, follow rule 3 for
2001 * NextSeg() specified in RFC3517.
2004 if (tcp_may_send_now(sk
))
2007 /* If nothing is SACKed, highest_sack in the loop won't be valid */
2008 if (!tp
->sacked_out
)
2011 if (tp
->forward_skb_hint
)
2012 skb
= tp
->forward_skb_hint
;
2014 skb
= tcp_write_queue_head(sk
);
2016 tcp_for_write_queue_from(skb
, sk
) {
2017 if (skb
== tcp_send_head(sk
))
2019 tp
->forward_skb_hint
= skb
;
2021 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_highest_sack_seq(tp
)))
2024 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
2027 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
)
2030 /* Ok, retransmit it. */
2031 if (tcp_retransmit_skb(sk
, skb
)) {
2032 tp
->forward_skb_hint
= NULL
;
2036 if (skb
== tcp_write_queue_head(sk
))
2037 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2038 inet_csk(sk
)->icsk_rto
,
2041 NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS
);
2046 /* Send a fin. The caller locks the socket for us. This cannot be
2047 * allowed to fail queueing a FIN frame under any circumstances.
2049 void tcp_send_fin(struct sock
*sk
)
2051 struct tcp_sock
*tp
= tcp_sk(sk
);
2052 struct sk_buff
*skb
= tcp_write_queue_tail(sk
);
2055 /* Optimization, tack on the FIN if we have a queue of
2056 * unsent frames. But be careful about outgoing SACKS
2059 mss_now
= tcp_current_mss(sk
, 1);
2061 if (tcp_send_head(sk
) != NULL
) {
2062 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_FIN
;
2063 TCP_SKB_CB(skb
)->end_seq
++;
2066 /* Socket is locked, keep trying until memory is available. */
2068 skb
= alloc_skb_fclone(MAX_TCP_HEADER
, GFP_KERNEL
);
2074 /* Reserve space for headers and prepare control bits. */
2075 skb_reserve(skb
, MAX_TCP_HEADER
);
2077 TCP_SKB_CB(skb
)->flags
= (TCPCB_FLAG_ACK
| TCPCB_FLAG_FIN
);
2078 TCP_SKB_CB(skb
)->sacked
= 0;
2079 skb_shinfo(skb
)->gso_segs
= 1;
2080 skb_shinfo(skb
)->gso_size
= 0;
2081 skb_shinfo(skb
)->gso_type
= 0;
2083 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2084 TCP_SKB_CB(skb
)->seq
= tp
->write_seq
;
2085 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ 1;
2086 tcp_queue_skb(sk
, skb
);
2088 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_OFF
);
2091 /* We get here when a process closes a file descriptor (either due to
2092 * an explicit close() or as a byproduct of exit()'ing) and there
2093 * was unread data in the receive queue. This behavior is recommended
2094 * by RFC 2525, section 2.17. -DaveM
2096 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
2098 struct sk_buff
*skb
;
2100 /* NOTE: No TCP options attached and we never retransmit this. */
2101 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
2103 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED
);
2107 /* Reserve space for headers and prepare control bits. */
2108 skb_reserve(skb
, MAX_TCP_HEADER
);
2110 TCP_SKB_CB(skb
)->flags
= (TCPCB_FLAG_ACK
| TCPCB_FLAG_RST
);
2111 TCP_SKB_CB(skb
)->sacked
= 0;
2112 skb_shinfo(skb
)->gso_segs
= 1;
2113 skb_shinfo(skb
)->gso_size
= 0;
2114 skb_shinfo(skb
)->gso_type
= 0;
2117 TCP_SKB_CB(skb
)->seq
= tcp_acceptable_seq(sk
);
2118 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
;
2119 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2120 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
2121 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED
);
2124 /* WARNING: This routine must only be called when we have already sent
2125 * a SYN packet that crossed the incoming SYN that caused this routine
2126 * to get called. If this assumption fails then the initial rcv_wnd
2127 * and rcv_wscale values will not be correct.
2129 int tcp_send_synack(struct sock
*sk
)
2131 struct sk_buff
* skb
;
2133 skb
= tcp_write_queue_head(sk
);
2134 if (skb
== NULL
|| !(TCP_SKB_CB(skb
)->flags
&TCPCB_FLAG_SYN
)) {
2135 printk(KERN_DEBUG
"tcp_send_synack: wrong queue state\n");
2138 if (!(TCP_SKB_CB(skb
)->flags
&TCPCB_FLAG_ACK
)) {
2139 if (skb_cloned(skb
)) {
2140 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
2143 tcp_unlink_write_queue(skb
, sk
);
2144 skb_header_release(nskb
);
2145 __tcp_add_write_queue_head(sk
, nskb
);
2146 sk_wmem_free_skb(sk
, skb
);
2147 sk
->sk_wmem_queued
+= nskb
->truesize
;
2148 sk_mem_charge(sk
, nskb
->truesize
);
2152 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_ACK
;
2153 TCP_ECN_send_synack(tcp_sk(sk
), skb
);
2155 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2156 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2160 * Prepare a SYN-ACK.
2162 struct sk_buff
* tcp_make_synack(struct sock
*sk
, struct dst_entry
*dst
,
2163 struct request_sock
*req
)
2165 struct inet_request_sock
*ireq
= inet_rsk(req
);
2166 struct tcp_sock
*tp
= tcp_sk(sk
);
2168 int tcp_header_size
;
2169 struct sk_buff
*skb
;
2170 #ifdef CONFIG_TCP_MD5SIG
2171 struct tcp_md5sig_key
*md5
;
2172 __u8
*md5_hash_location
;
2175 skb
= sock_wmalloc(sk
, MAX_TCP_HEADER
+ 15, 1, GFP_ATOMIC
);
2179 /* Reserve space for headers. */
2180 skb_reserve(skb
, MAX_TCP_HEADER
);
2182 skb
->dst
= dst_clone(dst
);
2184 tcp_header_size
= (sizeof(struct tcphdr
) + TCPOLEN_MSS
+
2185 (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0) +
2186 (ireq
->wscale_ok
? TCPOLEN_WSCALE_ALIGNED
: 0) +
2187 /* SACK_PERM is in the place of NOP NOP of TS */
2188 ((ireq
->sack_ok
&& !ireq
->tstamp_ok
) ? TCPOLEN_SACKPERM_ALIGNED
: 0));
2190 #ifdef CONFIG_TCP_MD5SIG
2191 /* Are we doing MD5 on this segment? If so - make room for it */
2192 md5
= tcp_rsk(req
)->af_specific
->md5_lookup(sk
, req
);
2194 tcp_header_size
+= TCPOLEN_MD5SIG_ALIGNED
;
2196 skb_push(skb
, tcp_header_size
);
2197 skb_reset_transport_header(skb
);
2200 memset(th
, 0, sizeof(struct tcphdr
));
2203 TCP_ECN_make_synack(req
, th
);
2204 th
->source
= inet_sk(sk
)->sport
;
2205 th
->dest
= ireq
->rmt_port
;
2206 TCP_SKB_CB(skb
)->seq
= tcp_rsk(req
)->snt_isn
;
2207 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ 1;
2208 TCP_SKB_CB(skb
)->sacked
= 0;
2209 skb_shinfo(skb
)->gso_segs
= 1;
2210 skb_shinfo(skb
)->gso_size
= 0;
2211 skb_shinfo(skb
)->gso_type
= 0;
2212 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
2213 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_isn
+ 1);
2214 if (req
->rcv_wnd
== 0) { /* ignored for retransmitted syns */
2216 /* Set this up on the first call only */
2217 req
->window_clamp
= tp
->window_clamp
? : dst_metric(dst
, RTAX_WINDOW
);
2218 /* tcp_full_space because it is guaranteed to be the first packet */
2219 tcp_select_initial_window(tcp_full_space(sk
),
2220 dst_metric(dst
, RTAX_ADVMSS
) - (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0),
2225 ireq
->rcv_wscale
= rcv_wscale
;
2228 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2229 th
->window
= htons(min(req
->rcv_wnd
, 65535U));
2231 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2232 tcp_syn_build_options((__be32
*)(th
+ 1), dst_metric(dst
, RTAX_ADVMSS
), ireq
->tstamp_ok
,
2233 ireq
->sack_ok
, ireq
->wscale_ok
, ireq
->rcv_wscale
,
2234 TCP_SKB_CB(skb
)->when
,
2237 #ifdef CONFIG_TCP_MD5SIG
2238 md5
? &md5_hash_location
:
2244 th
->doff
= (tcp_header_size
>> 2);
2245 TCP_INC_STATS(TCP_MIB_OUTSEGS
);
2247 #ifdef CONFIG_TCP_MD5SIG
2248 /* Okay, we have all we need - do the md5 hash if needed */
2250 tp
->af_specific
->calc_md5_hash(md5_hash_location
,
2253 tcp_hdr(skb
), sk
->sk_protocol
,
2262 * Do all connect socket setups that can be done AF independent.
2264 static void tcp_connect_init(struct sock
*sk
)
2266 struct dst_entry
*dst
= __sk_dst_get(sk
);
2267 struct tcp_sock
*tp
= tcp_sk(sk
);
2270 /* We'll fix this up when we get a response from the other end.
2271 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2273 tp
->tcp_header_len
= sizeof(struct tcphdr
) +
2274 (sysctl_tcp_timestamps
? TCPOLEN_TSTAMP_ALIGNED
: 0);
2276 #ifdef CONFIG_TCP_MD5SIG
2277 if (tp
->af_specific
->md5_lookup(sk
, sk
) != NULL
)
2278 tp
->tcp_header_len
+= TCPOLEN_MD5SIG_ALIGNED
;
2281 /* If user gave his TCP_MAXSEG, record it to clamp */
2282 if (tp
->rx_opt
.user_mss
)
2283 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
2286 tcp_sync_mss(sk
, dst_mtu(dst
));
2288 if (!tp
->window_clamp
)
2289 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
2290 tp
->advmss
= dst_metric(dst
, RTAX_ADVMSS
);
2291 tcp_initialize_rcv_mss(sk
);
2293 tcp_select_initial_window(tcp_full_space(sk
),
2294 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
2297 sysctl_tcp_window_scaling
,
2300 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
2301 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
2304 sock_reset_flag(sk
, SOCK_DONE
);
2306 tcp_init_wl(tp
, tp
->write_seq
, 0);
2307 tp
->snd_una
= tp
->write_seq
;
2308 tp
->snd_sml
= tp
->write_seq
;
2313 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
2314 inet_csk(sk
)->icsk_retransmits
= 0;
2315 tcp_clear_retrans(tp
);
2319 * Build a SYN and send it off.
2321 int tcp_connect(struct sock
*sk
)
2323 struct tcp_sock
*tp
= tcp_sk(sk
);
2324 struct sk_buff
*buff
;
2326 tcp_connect_init(sk
);
2328 buff
= alloc_skb_fclone(MAX_TCP_HEADER
+ 15, sk
->sk_allocation
);
2329 if (unlikely(buff
== NULL
))
2332 /* Reserve space for headers. */
2333 skb_reserve(buff
, MAX_TCP_HEADER
);
2335 TCP_SKB_CB(buff
)->flags
= TCPCB_FLAG_SYN
;
2336 TCP_ECN_send_syn(sk
, buff
);
2337 TCP_SKB_CB(buff
)->sacked
= 0;
2338 skb_shinfo(buff
)->gso_segs
= 1;
2339 skb_shinfo(buff
)->gso_size
= 0;
2340 skb_shinfo(buff
)->gso_type
= 0;
2342 tp
->snd_nxt
= tp
->write_seq
;
2343 TCP_SKB_CB(buff
)->seq
= tp
->write_seq
++;
2344 TCP_SKB_CB(buff
)->end_seq
= tp
->write_seq
;
2347 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2348 tp
->retrans_stamp
= TCP_SKB_CB(buff
)->when
;
2349 skb_header_release(buff
);
2350 __tcp_add_write_queue_tail(sk
, buff
);
2351 sk
->sk_wmem_queued
+= buff
->truesize
;
2352 sk_mem_charge(sk
, buff
->truesize
);
2353 tp
->packets_out
+= tcp_skb_pcount(buff
);
2354 tcp_transmit_skb(sk
, buff
, 1, GFP_KERNEL
);
2356 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2357 * in order to make this packet get counted in tcpOutSegs.
2359 tp
->snd_nxt
= tp
->write_seq
;
2360 tp
->pushed_seq
= tp
->write_seq
;
2361 TCP_INC_STATS(TCP_MIB_ACTIVEOPENS
);
2363 /* Timer for repeating the SYN until an answer. */
2364 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2365 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2369 /* Send out a delayed ack, the caller does the policy checking
2370 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2373 void tcp_send_delayed_ack(struct sock
*sk
)
2375 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2376 int ato
= icsk
->icsk_ack
.ato
;
2377 unsigned long timeout
;
2379 if (ato
> TCP_DELACK_MIN
) {
2380 const struct tcp_sock
*tp
= tcp_sk(sk
);
2383 if (icsk
->icsk_ack
.pingpong
|| (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
2384 max_ato
= TCP_DELACK_MAX
;
2386 /* Slow path, intersegment interval is "high". */
2388 /* If some rtt estimate is known, use it to bound delayed ack.
2389 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2393 int rtt
= max(tp
->srtt
>>3, TCP_DELACK_MIN
);
2399 ato
= min(ato
, max_ato
);
2402 /* Stay within the limit we were given */
2403 timeout
= jiffies
+ ato
;
2405 /* Use new timeout only if there wasn't a older one earlier. */
2406 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
2407 /* If delack timer was blocked or is about to expire,
2410 if (icsk
->icsk_ack
.blocked
||
2411 time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
2416 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
2417 timeout
= icsk
->icsk_ack
.timeout
;
2419 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
2420 icsk
->icsk_ack
.timeout
= timeout
;
2421 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
2424 /* This routine sends an ack and also updates the window. */
2425 void tcp_send_ack(struct sock
*sk
)
2427 /* If we have been reset, we may not send again. */
2428 if (sk
->sk_state
!= TCP_CLOSE
) {
2429 struct sk_buff
*buff
;
2431 /* We are not putting this on the write queue, so
2432 * tcp_transmit_skb() will set the ownership to this
2435 buff
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2437 inet_csk_schedule_ack(sk
);
2438 inet_csk(sk
)->icsk_ack
.ato
= TCP_ATO_MIN
;
2439 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
2440 TCP_DELACK_MAX
, TCP_RTO_MAX
);
2444 /* Reserve space for headers and prepare control bits. */
2445 skb_reserve(buff
, MAX_TCP_HEADER
);
2447 TCP_SKB_CB(buff
)->flags
= TCPCB_FLAG_ACK
;
2448 TCP_SKB_CB(buff
)->sacked
= 0;
2449 skb_shinfo(buff
)->gso_segs
= 1;
2450 skb_shinfo(buff
)->gso_size
= 0;
2451 skb_shinfo(buff
)->gso_type
= 0;
2453 /* Send it off, this clears delayed acks for us. */
2454 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(buff
)->end_seq
= tcp_acceptable_seq(sk
);
2455 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2456 tcp_transmit_skb(sk
, buff
, 0, GFP_ATOMIC
);
2460 /* This routine sends a packet with an out of date sequence
2461 * number. It assumes the other end will try to ack it.
2463 * Question: what should we make while urgent mode?
2464 * 4.4BSD forces sending single byte of data. We cannot send
2465 * out of window data, because we have SND.NXT==SND.MAX...
2467 * Current solution: to send TWO zero-length segments in urgent mode:
2468 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2469 * out-of-date with SND.UNA-1 to probe window.
2471 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
)
2473 struct tcp_sock
*tp
= tcp_sk(sk
);
2474 struct sk_buff
*skb
;
2476 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2477 skb
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2481 /* Reserve space for headers and set control bits. */
2482 skb_reserve(skb
, MAX_TCP_HEADER
);
2484 TCP_SKB_CB(skb
)->flags
= TCPCB_FLAG_ACK
;
2485 TCP_SKB_CB(skb
)->sacked
= urgent
;
2486 skb_shinfo(skb
)->gso_segs
= 1;
2487 skb_shinfo(skb
)->gso_size
= 0;
2488 skb_shinfo(skb
)->gso_type
= 0;
2490 /* Use a previous sequence. This should cause the other
2491 * end to send an ack. Don't queue or clone SKB, just
2494 TCP_SKB_CB(skb
)->seq
= urgent
? tp
->snd_una
: tp
->snd_una
- 1;
2495 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
;
2496 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2497 return tcp_transmit_skb(sk
, skb
, 0, GFP_ATOMIC
);
2500 int tcp_write_wakeup(struct sock
*sk
)
2502 if (sk
->sk_state
!= TCP_CLOSE
) {
2503 struct tcp_sock
*tp
= tcp_sk(sk
);
2504 struct sk_buff
*skb
;
2506 if ((skb
= tcp_send_head(sk
)) != NULL
&&
2507 before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
+tp
->snd_wnd
)) {
2509 unsigned int mss
= tcp_current_mss(sk
, 0);
2510 unsigned int seg_size
= tp
->snd_una
+tp
->snd_wnd
-TCP_SKB_CB(skb
)->seq
;
2512 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
2513 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
2515 /* We are probing the opening of a window
2516 * but the window size is != 0
2517 * must have been a result SWS avoidance ( sender )
2519 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
2521 seg_size
= min(seg_size
, mss
);
2522 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2523 if (tcp_fragment(sk
, skb
, seg_size
, mss
))
2525 } else if (!tcp_skb_pcount(skb
))
2526 tcp_set_skb_tso_segs(sk
, skb
, mss
);
2528 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2529 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2530 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2532 update_send_head(sk
, skb
);
2537 between(tp
->snd_up
, tp
->snd_una
+1, tp
->snd_una
+0xFFFF))
2538 tcp_xmit_probe_skb(sk
, TCPCB_URG
);
2539 return tcp_xmit_probe_skb(sk
, 0);
2545 /* A window probe timeout has occurred. If window is not closed send
2546 * a partial packet else a zero probe.
2548 void tcp_send_probe0(struct sock
*sk
)
2550 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2551 struct tcp_sock
*tp
= tcp_sk(sk
);
2554 err
= tcp_write_wakeup(sk
);
2556 if (tp
->packets_out
|| !tcp_send_head(sk
)) {
2557 /* Cancel probe timer, if it is not required. */
2558 icsk
->icsk_probes_out
= 0;
2559 icsk
->icsk_backoff
= 0;
2564 if (icsk
->icsk_backoff
< sysctl_tcp_retries2
)
2565 icsk
->icsk_backoff
++;
2566 icsk
->icsk_probes_out
++;
2567 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2568 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
2571 /* If packet was not sent due to local congestion,
2572 * do not backoff and do not remember icsk_probes_out.
2573 * Let local senders to fight for local resources.
2575 * Use accumulated backoff yet.
2577 if (!icsk
->icsk_probes_out
)
2578 icsk
->icsk_probes_out
= 1;
2579 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2580 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
,
2581 TCP_RESOURCE_PROBE_INTERVAL
),
2586 EXPORT_SYMBOL(tcp_connect
);
2587 EXPORT_SYMBOL(tcp_make_synack
);
2588 EXPORT_SYMBOL(tcp_simple_retransmit
);
2589 EXPORT_SYMBOL(tcp_sync_mss
);
2590 EXPORT_SYMBOL(sysctl_tcp_tso_win_divisor
);
2591 EXPORT_SYMBOL(tcp_mtup_init
);