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_charge_skb(sk
, skb
);
643 static void tcp_set_skb_tso_segs(struct sock
*sk
, struct sk_buff
*skb
, unsigned int mss_now
)
645 if (skb
->len
<= mss_now
|| !sk_can_gso(sk
)) {
646 /* Avoid the costly divide in the normal
649 skb_shinfo(skb
)->gso_segs
= 1;
650 skb_shinfo(skb
)->gso_size
= 0;
651 skb_shinfo(skb
)->gso_type
= 0;
653 skb_shinfo(skb
)->gso_segs
= DIV_ROUND_UP(skb
->len
, mss_now
);
654 skb_shinfo(skb
)->gso_size
= mss_now
;
655 skb_shinfo(skb
)->gso_type
= sk
->sk_gso_type
;
659 /* When a modification to fackets out becomes necessary, we need to check
660 * skb is counted to fackets_out or not.
662 static void tcp_adjust_fackets_out(struct sock
*sk
, struct sk_buff
*skb
,
665 struct tcp_sock
*tp
= tcp_sk(sk
);
667 if (!tp
->sacked_out
|| tcp_is_reno(tp
))
670 if (after(tcp_highest_sack_seq(tp
), TCP_SKB_CB(skb
)->seq
))
671 tp
->fackets_out
-= decr
;
674 /* Function to create two new TCP segments. Shrinks the given segment
675 * to the specified size and appends a new segment with the rest of the
676 * packet to the list. This won't be called frequently, I hope.
677 * Remember, these are still headerless SKBs at this point.
679 int tcp_fragment(struct sock
*sk
, struct sk_buff
*skb
, u32 len
, unsigned int mss_now
)
681 struct tcp_sock
*tp
= tcp_sk(sk
);
682 struct sk_buff
*buff
;
683 int nsize
, old_factor
;
687 BUG_ON(len
> skb
->len
);
689 tcp_clear_retrans_hints_partial(tp
);
690 nsize
= skb_headlen(skb
) - len
;
694 if (skb_cloned(skb
) &&
695 skb_is_nonlinear(skb
) &&
696 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
699 /* Get a new skb... force flag on. */
700 buff
= sk_stream_alloc_skb(sk
, nsize
, GFP_ATOMIC
);
702 return -ENOMEM
; /* We'll just try again later. */
704 sk_charge_skb(sk
, buff
);
705 nlen
= skb
->len
- len
- nsize
;
706 buff
->truesize
+= nlen
;
707 skb
->truesize
-= nlen
;
709 /* Correct the sequence numbers. */
710 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
711 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
712 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
714 /* PSH and FIN should only be set in the second packet. */
715 flags
= TCP_SKB_CB(skb
)->flags
;
716 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
717 TCP_SKB_CB(buff
)->flags
= flags
;
718 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
719 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_AT_TAIL
;
721 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_PARTIAL
) {
722 /* Copy and checksum data tail into the new buffer. */
723 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
, skb_put(buff
, nsize
),
728 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
730 skb
->ip_summed
= CHECKSUM_PARTIAL
;
731 skb_split(skb
, buff
, len
);
734 buff
->ip_summed
= skb
->ip_summed
;
736 /* Looks stupid, but our code really uses when of
737 * skbs, which it never sent before. --ANK
739 TCP_SKB_CB(buff
)->when
= TCP_SKB_CB(skb
)->when
;
740 buff
->tstamp
= skb
->tstamp
;
742 old_factor
= tcp_skb_pcount(skb
);
744 /* Fix up tso_factor for both original and new SKB. */
745 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
746 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
748 /* If this packet has been sent out already, we must
749 * adjust the various packet counters.
751 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
752 int diff
= old_factor
- tcp_skb_pcount(skb
) -
753 tcp_skb_pcount(buff
);
755 tp
->packets_out
-= diff
;
757 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
758 tp
->sacked_out
-= diff
;
759 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
760 tp
->retrans_out
-= diff
;
762 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
)
763 tp
->lost_out
-= diff
;
765 /* Adjust Reno SACK estimate. */
766 if (tcp_is_reno(tp
) && diff
> 0) {
767 tcp_dec_pcount_approx_int(&tp
->sacked_out
, diff
);
768 tcp_verify_left_out(tp
);
770 tcp_adjust_fackets_out(sk
, skb
, diff
);
773 /* Link BUFF into the send queue. */
774 skb_header_release(buff
);
775 tcp_insert_write_queue_after(skb
, buff
, sk
);
780 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
781 * eventually). The difference is that pulled data not copied, but
782 * immediately discarded.
784 static void __pskb_trim_head(struct sk_buff
*skb
, int len
)
790 for (i
=0; i
<skb_shinfo(skb
)->nr_frags
; i
++) {
791 if (skb_shinfo(skb
)->frags
[i
].size
<= eat
) {
792 put_page(skb_shinfo(skb
)->frags
[i
].page
);
793 eat
-= skb_shinfo(skb
)->frags
[i
].size
;
795 skb_shinfo(skb
)->frags
[k
] = skb_shinfo(skb
)->frags
[i
];
797 skb_shinfo(skb
)->frags
[k
].page_offset
+= eat
;
798 skb_shinfo(skb
)->frags
[k
].size
-= eat
;
804 skb_shinfo(skb
)->nr_frags
= k
;
806 skb_reset_tail_pointer(skb
);
807 skb
->data_len
-= len
;
808 skb
->len
= skb
->data_len
;
811 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
813 if (skb_cloned(skb
) &&
814 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
817 /* If len == headlen, we avoid __skb_pull to preserve alignment. */
818 if (unlikely(len
< skb_headlen(skb
)))
819 __skb_pull(skb
, len
);
821 __pskb_trim_head(skb
, len
- skb_headlen(skb
));
823 TCP_SKB_CB(skb
)->seq
+= len
;
824 skb
->ip_summed
= CHECKSUM_PARTIAL
;
826 skb
->truesize
-= len
;
827 sk
->sk_wmem_queued
-= len
;
828 sk
->sk_forward_alloc
+= len
;
829 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
831 /* Any change of skb->len requires recalculation of tso
834 if (tcp_skb_pcount(skb
) > 1)
835 tcp_set_skb_tso_segs(sk
, skb
, tcp_current_mss(sk
, 1));
840 /* Not accounting for SACKs here. */
841 int tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
843 struct tcp_sock
*tp
= tcp_sk(sk
);
844 struct inet_connection_sock
*icsk
= inet_csk(sk
);
847 /* Calculate base mss without TCP options:
848 It is MMS_S - sizeof(tcphdr) of rfc1122
850 mss_now
= pmtu
- icsk
->icsk_af_ops
->net_header_len
- sizeof(struct tcphdr
);
852 /* Clamp it (mss_clamp does not include tcp options) */
853 if (mss_now
> tp
->rx_opt
.mss_clamp
)
854 mss_now
= tp
->rx_opt
.mss_clamp
;
856 /* Now subtract optional transport overhead */
857 mss_now
-= icsk
->icsk_ext_hdr_len
;
859 /* Then reserve room for full set of TCP options and 8 bytes of data */
863 /* Now subtract TCP options size, not including SACKs */
864 mss_now
-= tp
->tcp_header_len
- sizeof(struct tcphdr
);
869 /* Inverse of above */
870 int tcp_mss_to_mtu(struct sock
*sk
, int mss
)
872 struct tcp_sock
*tp
= tcp_sk(sk
);
873 struct inet_connection_sock
*icsk
= inet_csk(sk
);
878 icsk
->icsk_ext_hdr_len
+
879 icsk
->icsk_af_ops
->net_header_len
;
884 void tcp_mtup_init(struct sock
*sk
)
886 struct tcp_sock
*tp
= tcp_sk(sk
);
887 struct inet_connection_sock
*icsk
= inet_csk(sk
);
889 icsk
->icsk_mtup
.enabled
= sysctl_tcp_mtu_probing
> 1;
890 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+ sizeof(struct tcphdr
) +
891 icsk
->icsk_af_ops
->net_header_len
;
892 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, sysctl_tcp_base_mss
);
893 icsk
->icsk_mtup
.probe_size
= 0;
896 /* This function synchronize snd mss to current pmtu/exthdr set.
898 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
899 for TCP options, but includes only bare TCP header.
901 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
902 It is minimum of user_mss and mss received with SYN.
903 It also does not include TCP options.
905 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
907 tp->mss_cache is current effective sending mss, including
908 all tcp options except for SACKs. It is evaluated,
909 taking into account current pmtu, but never exceeds
910 tp->rx_opt.mss_clamp.
912 NOTE1. rfc1122 clearly states that advertised MSS
913 DOES NOT include either tcp or ip options.
915 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
916 are READ ONLY outside this function. --ANK (980731)
919 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
921 struct tcp_sock
*tp
= tcp_sk(sk
);
922 struct inet_connection_sock
*icsk
= inet_csk(sk
);
925 if (icsk
->icsk_mtup
.search_high
> pmtu
)
926 icsk
->icsk_mtup
.search_high
= pmtu
;
928 mss_now
= tcp_mtu_to_mss(sk
, pmtu
);
930 /* Bound mss with half of window */
931 if (tp
->max_window
&& mss_now
> (tp
->max_window
>>1))
932 mss_now
= max((tp
->max_window
>>1), 68U - tp
->tcp_header_len
);
934 /* And store cached results */
935 icsk
->icsk_pmtu_cookie
= pmtu
;
936 if (icsk
->icsk_mtup
.enabled
)
937 mss_now
= min(mss_now
, tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
));
938 tp
->mss_cache
= mss_now
;
943 /* Compute the current effective MSS, taking SACKs and IP options,
944 * and even PMTU discovery events into account.
946 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
947 * cannot be large. However, taking into account rare use of URG, this
950 unsigned int tcp_current_mss(struct sock
*sk
, int large_allowed
)
952 struct tcp_sock
*tp
= tcp_sk(sk
);
953 struct dst_entry
*dst
= __sk_dst_get(sk
);
958 mss_now
= tp
->mss_cache
;
960 if (large_allowed
&& sk_can_gso(sk
) && !tp
->urg_mode
)
964 u32 mtu
= dst_mtu(dst
);
965 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
966 mss_now
= tcp_sync_mss(sk
, mtu
);
969 if (tp
->rx_opt
.eff_sacks
)
970 mss_now
-= (TCPOLEN_SACK_BASE_ALIGNED
+
971 (tp
->rx_opt
.eff_sacks
* TCPOLEN_SACK_PERBLOCK
));
973 #ifdef CONFIG_TCP_MD5SIG
974 if (tp
->af_specific
->md5_lookup(sk
, sk
))
975 mss_now
-= TCPOLEN_MD5SIG_ALIGNED
;
978 xmit_size_goal
= mss_now
;
981 xmit_size_goal
= (65535 -
982 inet_csk(sk
)->icsk_af_ops
->net_header_len
-
983 inet_csk(sk
)->icsk_ext_hdr_len
-
986 if (tp
->max_window
&&
987 (xmit_size_goal
> (tp
->max_window
>> 1)))
988 xmit_size_goal
= max((tp
->max_window
>> 1),
989 68U - tp
->tcp_header_len
);
991 xmit_size_goal
-= (xmit_size_goal
% mss_now
);
993 tp
->xmit_size_goal
= xmit_size_goal
;
998 /* Congestion window validation. (RFC2861) */
1000 static void tcp_cwnd_validate(struct sock
*sk
)
1002 struct tcp_sock
*tp
= tcp_sk(sk
);
1003 __u32 packets_out
= tp
->packets_out
;
1005 if (packets_out
>= tp
->snd_cwnd
) {
1006 /* Network is feed fully. */
1007 tp
->snd_cwnd_used
= 0;
1008 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1010 /* Network starves. */
1011 if (tp
->packets_out
> tp
->snd_cwnd_used
)
1012 tp
->snd_cwnd_used
= tp
->packets_out
;
1014 if (sysctl_tcp_slow_start_after_idle
&&
1015 (s32
)(tcp_time_stamp
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
)
1016 tcp_cwnd_application_limited(sk
);
1020 static unsigned int tcp_window_allows(struct tcp_sock
*tp
, struct sk_buff
*skb
, unsigned int mss_now
, unsigned int cwnd
)
1022 u32 window
, cwnd_len
;
1024 window
= (tp
->snd_una
+ tp
->snd_wnd
- TCP_SKB_CB(skb
)->seq
);
1025 cwnd_len
= mss_now
* cwnd
;
1026 return min(window
, cwnd_len
);
1029 /* Can at least one segment of SKB be sent right now, according to the
1030 * congestion window rules? If so, return how many segments are allowed.
1032 static inline unsigned int tcp_cwnd_test(struct tcp_sock
*tp
, struct sk_buff
*skb
)
1034 u32 in_flight
, cwnd
;
1036 /* Don't be strict about the congestion window for the final FIN. */
1037 if ((TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
) &&
1038 tcp_skb_pcount(skb
) == 1)
1041 in_flight
= tcp_packets_in_flight(tp
);
1042 cwnd
= tp
->snd_cwnd
;
1043 if (in_flight
< cwnd
)
1044 return (cwnd
- in_flight
);
1049 /* This must be invoked the first time we consider transmitting
1050 * SKB onto the wire.
1052 static int tcp_init_tso_segs(struct sock
*sk
, struct sk_buff
*skb
, unsigned int mss_now
)
1054 int tso_segs
= tcp_skb_pcount(skb
);
1058 tcp_skb_mss(skb
) != mss_now
)) {
1059 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1060 tso_segs
= tcp_skb_pcount(skb
);
1065 static inline int tcp_minshall_check(const struct tcp_sock
*tp
)
1067 return after(tp
->snd_sml
,tp
->snd_una
) &&
1068 !after(tp
->snd_sml
, tp
->snd_nxt
);
1071 /* Return 0, if packet can be sent now without violation Nagle's rules:
1072 * 1. It is full sized.
1073 * 2. Or it contains FIN. (already checked by caller)
1074 * 3. Or TCP_NODELAY was set.
1075 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1076 * With Minshall's modification: all sent small packets are ACKed.
1079 static inline int tcp_nagle_check(const struct tcp_sock
*tp
,
1080 const struct sk_buff
*skb
,
1081 unsigned mss_now
, int nonagle
)
1083 return (skb
->len
< mss_now
&&
1084 ((nonagle
&TCP_NAGLE_CORK
) ||
1087 tcp_minshall_check(tp
))));
1090 /* Return non-zero if the Nagle test allows this packet to be
1093 static inline int tcp_nagle_test(struct tcp_sock
*tp
, struct sk_buff
*skb
,
1094 unsigned int cur_mss
, int nonagle
)
1096 /* Nagle rule does not apply to frames, which sit in the middle of the
1097 * write_queue (they have no chances to get new data).
1099 * This is implemented in the callers, where they modify the 'nonagle'
1100 * argument based upon the location of SKB in the send queue.
1102 if (nonagle
& TCP_NAGLE_PUSH
)
1105 /* Don't use the nagle rule for urgent data (or for the final FIN).
1106 * Nagle can be ignored during F-RTO too (see RFC4138).
1108 if (tp
->urg_mode
|| (tp
->frto_counter
== 2) ||
1109 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
))
1112 if (!tcp_nagle_check(tp
, skb
, cur_mss
, nonagle
))
1118 /* Does at least the first segment of SKB fit into the send window? */
1119 static inline int tcp_snd_wnd_test(struct tcp_sock
*tp
, struct sk_buff
*skb
, unsigned int cur_mss
)
1121 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1123 if (skb
->len
> cur_mss
)
1124 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
1126 return !after(end_seq
, tp
->snd_una
+ tp
->snd_wnd
);
1129 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1130 * should be put on the wire right now. If so, it returns the number of
1131 * packets allowed by the congestion window.
1133 static unsigned int tcp_snd_test(struct sock
*sk
, struct sk_buff
*skb
,
1134 unsigned int cur_mss
, int nonagle
)
1136 struct tcp_sock
*tp
= tcp_sk(sk
);
1137 unsigned int cwnd_quota
;
1139 tcp_init_tso_segs(sk
, skb
, cur_mss
);
1141 if (!tcp_nagle_test(tp
, skb
, cur_mss
, nonagle
))
1144 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1146 !tcp_snd_wnd_test(tp
, skb
, cur_mss
))
1152 int tcp_may_send_now(struct sock
*sk
)
1154 struct tcp_sock
*tp
= tcp_sk(sk
);
1155 struct sk_buff
*skb
= tcp_send_head(sk
);
1158 tcp_snd_test(sk
, skb
, tcp_current_mss(sk
, 1),
1159 (tcp_skb_is_last(sk
, skb
) ?
1160 tp
->nonagle
: TCP_NAGLE_PUSH
)));
1163 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1164 * which is put after SKB on the list. It is very much like
1165 * tcp_fragment() except that it may make several kinds of assumptions
1166 * in order to speed up the splitting operation. In particular, we
1167 * know that all the data is in scatter-gather pages, and that the
1168 * packet has never been sent out before (and thus is not cloned).
1170 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
, unsigned int mss_now
)
1172 struct sk_buff
*buff
;
1173 int nlen
= skb
->len
- len
;
1176 /* All of a TSO frame must be composed of paged data. */
1177 if (skb
->len
!= skb
->data_len
)
1178 return tcp_fragment(sk
, skb
, len
, mss_now
);
1180 buff
= sk_stream_alloc_skb(sk
, 0, GFP_ATOMIC
);
1181 if (unlikely(buff
== NULL
))
1184 sk_charge_skb(sk
, buff
);
1185 buff
->truesize
+= nlen
;
1186 skb
->truesize
-= nlen
;
1188 /* Correct the sequence numbers. */
1189 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1190 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1191 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1193 /* PSH and FIN should only be set in the second packet. */
1194 flags
= TCP_SKB_CB(skb
)->flags
;
1195 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
1196 TCP_SKB_CB(buff
)->flags
= flags
;
1198 /* This packet was never sent out yet, so no SACK bits. */
1199 TCP_SKB_CB(buff
)->sacked
= 0;
1201 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_PARTIAL
;
1202 skb_split(skb
, buff
, len
);
1204 /* Fix up tso_factor for both original and new SKB. */
1205 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1206 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1208 /* Link BUFF into the send queue. */
1209 skb_header_release(buff
);
1210 tcp_insert_write_queue_after(skb
, buff
, sk
);
1215 /* Try to defer sending, if possible, in order to minimize the amount
1216 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1218 * This algorithm is from John Heffner.
1220 static int tcp_tso_should_defer(struct sock
*sk
, struct sk_buff
*skb
)
1222 struct tcp_sock
*tp
= tcp_sk(sk
);
1223 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1224 u32 send_win
, cong_win
, limit
, in_flight
;
1226 if (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
)
1229 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
1232 /* Defer for less than two clock ticks. */
1233 if (tp
->tso_deferred
&&
1234 ((jiffies
<< 1) >> 1) - (tp
->tso_deferred
>> 1) > 1)
1237 in_flight
= tcp_packets_in_flight(tp
);
1239 BUG_ON(tcp_skb_pcount(skb
) <= 1 ||
1240 (tp
->snd_cwnd
<= in_flight
));
1242 send_win
= (tp
->snd_una
+ tp
->snd_wnd
) - TCP_SKB_CB(skb
)->seq
;
1244 /* From in_flight test above, we know that cwnd > in_flight. */
1245 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
1247 limit
= min(send_win
, cong_win
);
1249 /* If a full-sized TSO skb can be sent, do it. */
1253 if (sysctl_tcp_tso_win_divisor
) {
1254 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
1256 /* If at least some fraction of a window is available,
1259 chunk
/= sysctl_tcp_tso_win_divisor
;
1263 /* Different approach, try not to defer past a single
1264 * ACK. Receiver should ACK every other full sized
1265 * frame, so if we have space for more than 3 frames
1268 if (limit
> tcp_max_burst(tp
) * tp
->mss_cache
)
1272 /* Ok, it looks like it is advisable to defer. */
1273 tp
->tso_deferred
= 1 | (jiffies
<<1);
1278 tp
->tso_deferred
= 0;
1282 /* Create a new MTU probe if we are ready.
1283 * Returns 0 if we should wait to probe (no cwnd available),
1284 * 1 if a probe was sent,
1286 static int tcp_mtu_probe(struct sock
*sk
)
1288 struct tcp_sock
*tp
= tcp_sk(sk
);
1289 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1290 struct sk_buff
*skb
, *nskb
, *next
;
1297 /* Not currently probing/verifying,
1299 * have enough cwnd, and
1300 * not SACKing (the variable headers throw things off) */
1301 if (!icsk
->icsk_mtup
.enabled
||
1302 icsk
->icsk_mtup
.probe_size
||
1303 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
||
1304 tp
->snd_cwnd
< 11 ||
1305 tp
->rx_opt
.eff_sacks
)
1308 /* Very simple search strategy: just double the MSS. */
1309 mss_now
= tcp_current_mss(sk
, 0);
1310 probe_size
= 2*tp
->mss_cache
;
1311 size_needed
= probe_size
+ (tp
->reordering
+ 1) * tp
->mss_cache
;
1312 if (probe_size
> tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_high
)) {
1313 /* TODO: set timer for probe_converge_event */
1317 /* Have enough data in the send queue to probe? */
1318 if (tp
->write_seq
- tp
->snd_nxt
< size_needed
)
1321 if (tp
->snd_wnd
< size_needed
)
1323 if (after(tp
->snd_nxt
+ size_needed
, tp
->snd_una
+ tp
->snd_wnd
))
1326 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1327 if (tcp_packets_in_flight(tp
) + 2 > tp
->snd_cwnd
) {
1328 if (!tcp_packets_in_flight(tp
))
1334 /* We're allowed to probe. Build it now. */
1335 if ((nskb
= sk_stream_alloc_skb(sk
, probe_size
, GFP_ATOMIC
)) == NULL
)
1337 sk_charge_skb(sk
, nskb
);
1339 skb
= tcp_send_head(sk
);
1341 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(skb
)->seq
;
1342 TCP_SKB_CB(nskb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ probe_size
;
1343 TCP_SKB_CB(nskb
)->flags
= TCPCB_FLAG_ACK
;
1344 TCP_SKB_CB(nskb
)->sacked
= 0;
1346 nskb
->ip_summed
= skb
->ip_summed
;
1348 tcp_insert_write_queue_before(nskb
, skb
, sk
);
1351 tcp_for_write_queue_from_safe(skb
, next
, sk
) {
1352 copy
= min_t(int, skb
->len
, probe_size
- len
);
1353 if (nskb
->ip_summed
)
1354 skb_copy_bits(skb
, 0, skb_put(nskb
, copy
), copy
);
1356 nskb
->csum
= skb_copy_and_csum_bits(skb
, 0,
1357 skb_put(nskb
, copy
), copy
, nskb
->csum
);
1359 if (skb
->len
<= copy
) {
1360 /* We've eaten all the data from this skb.
1362 TCP_SKB_CB(nskb
)->flags
|= TCP_SKB_CB(skb
)->flags
;
1363 tcp_unlink_write_queue(skb
, sk
);
1364 sk_stream_free_skb(sk
, skb
);
1366 TCP_SKB_CB(nskb
)->flags
|= TCP_SKB_CB(skb
)->flags
&
1367 ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
1368 if (!skb_shinfo(skb
)->nr_frags
) {
1369 skb_pull(skb
, copy
);
1370 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1371 skb
->csum
= csum_partial(skb
->data
, skb
->len
, 0);
1373 __pskb_trim_head(skb
, copy
);
1374 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1376 TCP_SKB_CB(skb
)->seq
+= copy
;
1381 if (len
>= probe_size
)
1384 tcp_init_tso_segs(sk
, nskb
, nskb
->len
);
1386 /* We're ready to send. If this fails, the probe will
1387 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1388 TCP_SKB_CB(nskb
)->when
= tcp_time_stamp
;
1389 if (!tcp_transmit_skb(sk
, nskb
, 1, GFP_ATOMIC
)) {
1390 /* Decrement cwnd here because we are sending
1391 * effectively two packets. */
1393 update_send_head(sk
, nskb
);
1395 icsk
->icsk_mtup
.probe_size
= tcp_mss_to_mtu(sk
, nskb
->len
);
1396 tp
->mtu_probe
.probe_seq_start
= TCP_SKB_CB(nskb
)->seq
;
1397 tp
->mtu_probe
.probe_seq_end
= TCP_SKB_CB(nskb
)->end_seq
;
1406 /* This routine writes packets to the network. It advances the
1407 * send_head. This happens as incoming acks open up the remote
1410 * Returns 1, if no segments are in flight and we have queued segments, but
1411 * cannot send anything now because of SWS or another problem.
1413 static int tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
)
1415 struct tcp_sock
*tp
= tcp_sk(sk
);
1416 struct sk_buff
*skb
;
1417 unsigned int tso_segs
, sent_pkts
;
1421 /* If we are closed, the bytes will have to remain here.
1422 * In time closedown will finish, we empty the write queue and all
1425 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
1430 /* Do MTU probing. */
1431 if ((result
= tcp_mtu_probe(sk
)) == 0) {
1433 } else if (result
> 0) {
1437 while ((skb
= tcp_send_head(sk
))) {
1440 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1443 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1447 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
)))
1450 if (tso_segs
== 1) {
1451 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
1452 (tcp_skb_is_last(sk
, skb
) ?
1453 nonagle
: TCP_NAGLE_PUSH
))))
1456 if (tcp_tso_should_defer(sk
, skb
))
1462 limit
= tcp_window_allows(tp
, skb
,
1463 mss_now
, cwnd_quota
);
1465 if (skb
->len
< limit
) {
1466 unsigned int trim
= skb
->len
% mss_now
;
1469 limit
= skb
->len
- trim
;
1473 if (skb
->len
> limit
&&
1474 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
)))
1477 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1479 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
)))
1482 /* Advance the send_head. This one is sent out.
1483 * This call will increment packets_out.
1485 update_send_head(sk
, skb
);
1487 tcp_minshall_update(tp
, mss_now
, skb
);
1491 if (likely(sent_pkts
)) {
1492 tcp_cwnd_validate(sk
);
1495 return !tp
->packets_out
&& tcp_send_head(sk
);
1498 /* Push out any pending frames which were held back due to
1499 * TCP_CORK or attempt at coalescing tiny packets.
1500 * The socket must be locked by the caller.
1502 void __tcp_push_pending_frames(struct sock
*sk
, unsigned int cur_mss
,
1505 struct sk_buff
*skb
= tcp_send_head(sk
);
1508 if (tcp_write_xmit(sk
, cur_mss
, nonagle
))
1509 tcp_check_probe_timer(sk
);
1513 /* Send _single_ skb sitting at the send head. This function requires
1514 * true push pending frames to setup probe timer etc.
1516 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
1518 struct tcp_sock
*tp
= tcp_sk(sk
);
1519 struct sk_buff
*skb
= tcp_send_head(sk
);
1520 unsigned int tso_segs
, cwnd_quota
;
1522 BUG_ON(!skb
|| skb
->len
< mss_now
);
1524 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1525 cwnd_quota
= tcp_snd_test(sk
, skb
, mss_now
, TCP_NAGLE_PUSH
);
1527 if (likely(cwnd_quota
)) {
1534 limit
= tcp_window_allows(tp
, skb
,
1535 mss_now
, cwnd_quota
);
1537 if (skb
->len
< limit
) {
1538 unsigned int trim
= skb
->len
% mss_now
;
1541 limit
= skb
->len
- trim
;
1545 if (skb
->len
> limit
&&
1546 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
)))
1549 /* Send it out now. */
1550 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1552 if (likely(!tcp_transmit_skb(sk
, skb
, 1, sk
->sk_allocation
))) {
1553 update_send_head(sk
, skb
);
1554 tcp_cwnd_validate(sk
);
1560 /* This function returns the amount that we can raise the
1561 * usable window based on the following constraints
1563 * 1. The window can never be shrunk once it is offered (RFC 793)
1564 * 2. We limit memory per socket
1567 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1568 * RECV.NEXT + RCV.WIN fixed until:
1569 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1571 * i.e. don't raise the right edge of the window until you can raise
1572 * it at least MSS bytes.
1574 * Unfortunately, the recommended algorithm breaks header prediction,
1575 * since header prediction assumes th->window stays fixed.
1577 * Strictly speaking, keeping th->window fixed violates the receiver
1578 * side SWS prevention criteria. The problem is that under this rule
1579 * a stream of single byte packets will cause the right side of the
1580 * window to always advance by a single byte.
1582 * Of course, if the sender implements sender side SWS prevention
1583 * then this will not be a problem.
1585 * BSD seems to make the following compromise:
1587 * If the free space is less than the 1/4 of the maximum
1588 * space available and the free space is less than 1/2 mss,
1589 * then set the window to 0.
1590 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1591 * Otherwise, just prevent the window from shrinking
1592 * and from being larger than the largest representable value.
1594 * This prevents incremental opening of the window in the regime
1595 * where TCP is limited by the speed of the reader side taking
1596 * data out of the TCP receive queue. It does nothing about
1597 * those cases where the window is constrained on the sender side
1598 * because the pipeline is full.
1600 * BSD also seems to "accidentally" limit itself to windows that are a
1601 * multiple of MSS, at least until the free space gets quite small.
1602 * This would appear to be a side effect of the mbuf implementation.
1603 * Combining these two algorithms results in the observed behavior
1604 * of having a fixed window size at almost all times.
1606 * Below we obtain similar behavior by forcing the offered window to
1607 * a multiple of the mss when it is feasible to do so.
1609 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1610 * Regular options like TIMESTAMP are taken into account.
1612 u32
__tcp_select_window(struct sock
*sk
)
1614 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1615 struct tcp_sock
*tp
= tcp_sk(sk
);
1616 /* MSS for the peer's data. Previous versions used mss_clamp
1617 * here. I don't know if the value based on our guesses
1618 * of peer's MSS is better for the performance. It's more correct
1619 * but may be worse for the performance because of rcv_mss
1620 * fluctuations. --SAW 1998/11/1
1622 int mss
= icsk
->icsk_ack
.rcv_mss
;
1623 int free_space
= tcp_space(sk
);
1624 int full_space
= min_t(int, tp
->window_clamp
, tcp_full_space(sk
));
1627 if (mss
> full_space
)
1630 if (free_space
< (full_space
>> 1)) {
1631 icsk
->icsk_ack
.quick
= 0;
1633 if (tcp_memory_pressure
)
1634 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U*tp
->advmss
);
1636 if (free_space
< mss
)
1640 if (free_space
> tp
->rcv_ssthresh
)
1641 free_space
= tp
->rcv_ssthresh
;
1643 /* Don't do rounding if we are using window scaling, since the
1644 * scaled window will not line up with the MSS boundary anyway.
1646 window
= tp
->rcv_wnd
;
1647 if (tp
->rx_opt
.rcv_wscale
) {
1648 window
= free_space
;
1650 /* Advertise enough space so that it won't get scaled away.
1651 * Import case: prevent zero window announcement if
1652 * 1<<rcv_wscale > mss.
1654 if (((window
>> tp
->rx_opt
.rcv_wscale
) << tp
->rx_opt
.rcv_wscale
) != window
)
1655 window
= (((window
>> tp
->rx_opt
.rcv_wscale
) + 1)
1656 << tp
->rx_opt
.rcv_wscale
);
1658 /* Get the largest window that is a nice multiple of mss.
1659 * Window clamp already applied above.
1660 * If our current window offering is within 1 mss of the
1661 * free space we just keep it. This prevents the divide
1662 * and multiply from happening most of the time.
1663 * We also don't do any window rounding when the free space
1666 if (window
<= free_space
- mss
|| window
> free_space
)
1667 window
= (free_space
/mss
)*mss
;
1668 else if (mss
== full_space
&&
1669 free_space
> window
+ (full_space
>> 1))
1670 window
= free_space
;
1676 /* Attempt to collapse two adjacent SKB's during retransmission. */
1677 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*skb
, int mss_now
)
1679 struct tcp_sock
*tp
= tcp_sk(sk
);
1680 struct sk_buff
*next_skb
= tcp_write_queue_next(sk
, skb
);
1682 /* The first test we must make is that neither of these two
1683 * SKB's are still referenced by someone else.
1685 if (!skb_cloned(skb
) && !skb_cloned(next_skb
)) {
1686 int skb_size
= skb
->len
, next_skb_size
= next_skb
->len
;
1687 u16 flags
= TCP_SKB_CB(skb
)->flags
;
1689 /* Also punt if next skb has been SACK'd. */
1690 if (TCP_SKB_CB(next_skb
)->sacked
& TCPCB_SACKED_ACKED
)
1693 /* Next skb is out of window. */
1694 if (after(TCP_SKB_CB(next_skb
)->end_seq
, tp
->snd_una
+tp
->snd_wnd
))
1697 /* Punt if not enough space exists in the first SKB for
1698 * the data in the second, or the total combined payload
1699 * would exceed the MSS.
1701 if ((next_skb_size
> skb_tailroom(skb
)) ||
1702 ((skb_size
+ next_skb_size
) > mss_now
))
1705 BUG_ON(tcp_skb_pcount(skb
) != 1 ||
1706 tcp_skb_pcount(next_skb
) != 1);
1708 tcp_highest_sack_combine(sk
, next_skb
, skb
);
1710 /* Ok. We will be able to collapse the packet. */
1711 tcp_unlink_write_queue(next_skb
, sk
);
1713 skb_copy_from_linear_data(next_skb
,
1714 skb_put(skb
, next_skb_size
),
1717 if (next_skb
->ip_summed
== CHECKSUM_PARTIAL
)
1718 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1720 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1721 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
1723 /* Update sequence range on original skb. */
1724 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
1726 /* Merge over control information. */
1727 flags
|= TCP_SKB_CB(next_skb
)->flags
; /* This moves PSH/FIN etc. over */
1728 TCP_SKB_CB(skb
)->flags
= flags
;
1730 /* All done, get rid of second SKB and account for it so
1731 * packet counting does not break.
1733 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
&(TCPCB_EVER_RETRANS
|TCPCB_AT_TAIL
);
1734 if (TCP_SKB_CB(next_skb
)->sacked
&TCPCB_SACKED_RETRANS
)
1735 tp
->retrans_out
-= tcp_skb_pcount(next_skb
);
1736 if (TCP_SKB_CB(next_skb
)->sacked
&TCPCB_LOST
)
1737 tp
->lost_out
-= tcp_skb_pcount(next_skb
);
1738 /* Reno case is special. Sigh... */
1739 if (tcp_is_reno(tp
) && tp
->sacked_out
)
1740 tcp_dec_pcount_approx(&tp
->sacked_out
, next_skb
);
1742 tcp_adjust_fackets_out(sk
, next_skb
, tcp_skb_pcount(next_skb
));
1743 tp
->packets_out
-= tcp_skb_pcount(next_skb
);
1745 /* changed transmit queue under us so clear hints */
1746 tcp_clear_retrans_hints_partial(tp
);
1748 sk_stream_free_skb(sk
, next_skb
);
1752 /* Do a simple retransmit without using the backoff mechanisms in
1753 * tcp_timer. This is used for path mtu discovery.
1754 * The socket is already locked here.
1756 void tcp_simple_retransmit(struct sock
*sk
)
1758 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1759 struct tcp_sock
*tp
= tcp_sk(sk
);
1760 struct sk_buff
*skb
;
1761 unsigned int mss
= tcp_current_mss(sk
, 0);
1764 tcp_for_write_queue(skb
, sk
) {
1765 if (skb
== tcp_send_head(sk
))
1767 if (skb
->len
> mss
&&
1768 !(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
)) {
1769 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) {
1770 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1771 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1773 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_LOST
)) {
1774 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1775 tp
->lost_out
+= tcp_skb_pcount(skb
);
1781 tcp_clear_all_retrans_hints(tp
);
1786 tcp_verify_left_out(tp
);
1788 /* Don't muck with the congestion window here.
1789 * Reason is that we do not increase amount of _data_
1790 * in network, but units changed and effective
1791 * cwnd/ssthresh really reduced now.
1793 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
1794 tp
->high_seq
= tp
->snd_nxt
;
1795 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
1796 tp
->prior_ssthresh
= 0;
1797 tp
->undo_marker
= 0;
1798 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1800 tcp_xmit_retransmit_queue(sk
);
1803 /* This retransmits one SKB. Policy decisions and retransmit queue
1804 * state updates are done by the caller. Returns non-zero if an
1805 * error occurred which prevented the send.
1807 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
1809 struct tcp_sock
*tp
= tcp_sk(sk
);
1810 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1811 unsigned int cur_mss
= tcp_current_mss(sk
, 0);
1814 /* Inconslusive MTU probe */
1815 if (icsk
->icsk_mtup
.probe_size
) {
1816 icsk
->icsk_mtup
.probe_size
= 0;
1819 /* Do not sent more than we queued. 1/4 is reserved for possible
1820 * copying overhead: fragmentation, tunneling, mangling etc.
1822 if (atomic_read(&sk
->sk_wmem_alloc
) >
1823 min(sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2), sk
->sk_sndbuf
))
1826 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
1827 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1829 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
1833 /* If receiver has shrunk his window, and skb is out of
1834 * new window, do not retransmit it. The exception is the
1835 * case, when window is shrunk to zero. In this case
1836 * our retransmit serves as a zero window probe.
1838 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
+tp
->snd_wnd
)
1839 && TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
1842 if (skb
->len
> cur_mss
) {
1843 if (tcp_fragment(sk
, skb
, cur_mss
, cur_mss
))
1844 return -ENOMEM
; /* We'll try again later. */
1847 /* Collapse two adjacent packets if worthwhile and we can. */
1848 if (!(TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_SYN
) &&
1849 (skb
->len
< (cur_mss
>> 1)) &&
1850 (tcp_write_queue_next(sk
, skb
) != tcp_send_head(sk
)) &&
1851 (!tcp_skb_is_last(sk
, skb
)) &&
1852 (skb_shinfo(skb
)->nr_frags
== 0 && skb_shinfo(tcp_write_queue_next(sk
, skb
))->nr_frags
== 0) &&
1853 (tcp_skb_pcount(skb
) == 1 && tcp_skb_pcount(tcp_write_queue_next(sk
, skb
)) == 1) &&
1854 (sysctl_tcp_retrans_collapse
!= 0))
1855 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
1857 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
1858 return -EHOSTUNREACH
; /* Routing failure or similar. */
1860 /* Some Solaris stacks overoptimize and ignore the FIN on a
1861 * retransmit when old data is attached. So strip it off
1862 * since it is cheap to do so and saves bytes on the network.
1865 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
) &&
1866 tp
->snd_una
== (TCP_SKB_CB(skb
)->end_seq
- 1)) {
1867 if (!pskb_trim(skb
, 0)) {
1868 TCP_SKB_CB(skb
)->seq
= TCP_SKB_CB(skb
)->end_seq
- 1;
1869 skb_shinfo(skb
)->gso_segs
= 1;
1870 skb_shinfo(skb
)->gso_size
= 0;
1871 skb_shinfo(skb
)->gso_type
= 0;
1872 skb
->ip_summed
= CHECKSUM_NONE
;
1877 /* Make a copy, if the first transmission SKB clone we made
1878 * is still in somebody's hands, else make a clone.
1880 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1882 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
1885 /* Update global TCP statistics. */
1886 TCP_INC_STATS(TCP_MIB_RETRANSSEGS
);
1888 tp
->total_retrans
++;
1890 #if FASTRETRANS_DEBUG > 0
1891 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) {
1892 if (net_ratelimit())
1893 printk(KERN_DEBUG
"retrans_out leaked.\n");
1896 if (!tp
->retrans_out
)
1897 tp
->lost_retrans_low
= tp
->snd_nxt
;
1898 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
1899 tp
->retrans_out
+= tcp_skb_pcount(skb
);
1901 /* Save stamp of the first retransmit. */
1902 if (!tp
->retrans_stamp
)
1903 tp
->retrans_stamp
= TCP_SKB_CB(skb
)->when
;
1907 /* snd_nxt is stored to detect loss of retransmitted segment,
1908 * see tcp_input.c tcp_sacktag_write_queue().
1910 TCP_SKB_CB(skb
)->ack_seq
= tp
->snd_nxt
;
1915 /* This gets called after a retransmit timeout, and the initially
1916 * retransmitted data is acknowledged. It tries to continue
1917 * resending the rest of the retransmit queue, until either
1918 * we've sent it all or the congestion window limit is reached.
1919 * If doing SACK, the first ACK which comes back for a timeout
1920 * based retransmit packet might feed us FACK information again.
1921 * If so, we use it to avoid unnecessarily retransmissions.
1923 void tcp_xmit_retransmit_queue(struct sock
*sk
)
1925 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1926 struct tcp_sock
*tp
= tcp_sk(sk
);
1927 struct sk_buff
*skb
;
1930 if (tp
->retransmit_skb_hint
) {
1931 skb
= tp
->retransmit_skb_hint
;
1932 packet_cnt
= tp
->retransmit_cnt_hint
;
1934 skb
= tcp_write_queue_head(sk
);
1938 /* First pass: retransmit lost packets. */
1940 tcp_for_write_queue_from(skb
, sk
) {
1941 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
1943 if (skb
== tcp_send_head(sk
))
1945 /* we could do better than to assign each time */
1946 tp
->retransmit_skb_hint
= skb
;
1947 tp
->retransmit_cnt_hint
= packet_cnt
;
1949 /* Assume this retransmit will generate
1950 * only one packet for congestion window
1951 * calculation purposes. This works because
1952 * tcp_retransmit_skb() will chop up the
1953 * packet to be MSS sized and all the
1954 * packet counting works out.
1956 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
1959 if (sacked
& TCPCB_LOST
) {
1960 if (!(sacked
&(TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))) {
1961 if (tcp_retransmit_skb(sk
, skb
)) {
1962 tp
->retransmit_skb_hint
= NULL
;
1965 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
1966 NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS
);
1968 NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS
);
1970 if (skb
== tcp_write_queue_head(sk
))
1971 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1972 inet_csk(sk
)->icsk_rto
,
1976 packet_cnt
+= tcp_skb_pcount(skb
);
1977 if (packet_cnt
>= tp
->lost_out
)
1983 /* OK, demanded retransmission is finished. */
1985 /* Forward retransmissions are possible only during Recovery. */
1986 if (icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
1989 /* No forward retransmissions in Reno are possible. */
1990 if (tcp_is_reno(tp
))
1993 /* Yeah, we have to make difficult choice between forward transmission
1994 * and retransmission... Both ways have their merits...
1996 * For now we do not retransmit anything, while we have some new
1997 * segments to send. In the other cases, follow rule 3 for
1998 * NextSeg() specified in RFC3517.
2001 if (tcp_may_send_now(sk
))
2004 /* If nothing is SACKed, highest_sack in the loop won't be valid */
2005 if (!tp
->sacked_out
)
2008 if (tp
->forward_skb_hint
)
2009 skb
= tp
->forward_skb_hint
;
2011 skb
= tcp_write_queue_head(sk
);
2013 tcp_for_write_queue_from(skb
, sk
) {
2014 if (skb
== tcp_send_head(sk
))
2016 tp
->forward_skb_hint
= skb
;
2018 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_highest_sack_seq(tp
)))
2021 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
2024 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
)
2027 /* Ok, retransmit it. */
2028 if (tcp_retransmit_skb(sk
, skb
)) {
2029 tp
->forward_skb_hint
= NULL
;
2033 if (skb
== tcp_write_queue_head(sk
))
2034 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2035 inet_csk(sk
)->icsk_rto
,
2038 NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS
);
2043 /* Send a fin. The caller locks the socket for us. This cannot be
2044 * allowed to fail queueing a FIN frame under any circumstances.
2046 void tcp_send_fin(struct sock
*sk
)
2048 struct tcp_sock
*tp
= tcp_sk(sk
);
2049 struct sk_buff
*skb
= tcp_write_queue_tail(sk
);
2052 /* Optimization, tack on the FIN if we have a queue of
2053 * unsent frames. But be careful about outgoing SACKS
2056 mss_now
= tcp_current_mss(sk
, 1);
2058 if (tcp_send_head(sk
) != NULL
) {
2059 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_FIN
;
2060 TCP_SKB_CB(skb
)->end_seq
++;
2063 /* Socket is locked, keep trying until memory is available. */
2065 skb
= alloc_skb_fclone(MAX_TCP_HEADER
, GFP_KERNEL
);
2071 /* Reserve space for headers and prepare control bits. */
2072 skb_reserve(skb
, MAX_TCP_HEADER
);
2074 TCP_SKB_CB(skb
)->flags
= (TCPCB_FLAG_ACK
| TCPCB_FLAG_FIN
);
2075 TCP_SKB_CB(skb
)->sacked
= 0;
2076 skb_shinfo(skb
)->gso_segs
= 1;
2077 skb_shinfo(skb
)->gso_size
= 0;
2078 skb_shinfo(skb
)->gso_type
= 0;
2080 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2081 TCP_SKB_CB(skb
)->seq
= tp
->write_seq
;
2082 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ 1;
2083 tcp_queue_skb(sk
, skb
);
2085 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_OFF
);
2088 /* We get here when a process closes a file descriptor (either due to
2089 * an explicit close() or as a byproduct of exit()'ing) and there
2090 * was unread data in the receive queue. This behavior is recommended
2091 * by RFC 2525, section 2.17. -DaveM
2093 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
2095 struct sk_buff
*skb
;
2097 /* NOTE: No TCP options attached and we never retransmit this. */
2098 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
2100 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED
);
2104 /* Reserve space for headers and prepare control bits. */
2105 skb_reserve(skb
, MAX_TCP_HEADER
);
2107 TCP_SKB_CB(skb
)->flags
= (TCPCB_FLAG_ACK
| TCPCB_FLAG_RST
);
2108 TCP_SKB_CB(skb
)->sacked
= 0;
2109 skb_shinfo(skb
)->gso_segs
= 1;
2110 skb_shinfo(skb
)->gso_size
= 0;
2111 skb_shinfo(skb
)->gso_type
= 0;
2114 TCP_SKB_CB(skb
)->seq
= tcp_acceptable_seq(sk
);
2115 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
;
2116 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2117 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
2118 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED
);
2121 /* WARNING: This routine must only be called when we have already sent
2122 * a SYN packet that crossed the incoming SYN that caused this routine
2123 * to get called. If this assumption fails then the initial rcv_wnd
2124 * and rcv_wscale values will not be correct.
2126 int tcp_send_synack(struct sock
*sk
)
2128 struct sk_buff
* skb
;
2130 skb
= tcp_write_queue_head(sk
);
2131 if (skb
== NULL
|| !(TCP_SKB_CB(skb
)->flags
&TCPCB_FLAG_SYN
)) {
2132 printk(KERN_DEBUG
"tcp_send_synack: wrong queue state\n");
2135 if (!(TCP_SKB_CB(skb
)->flags
&TCPCB_FLAG_ACK
)) {
2136 if (skb_cloned(skb
)) {
2137 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
2140 tcp_unlink_write_queue(skb
, sk
);
2141 skb_header_release(nskb
);
2142 __tcp_add_write_queue_head(sk
, nskb
);
2143 sk_stream_free_skb(sk
, skb
);
2144 sk_charge_skb(sk
, nskb
);
2148 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_ACK
;
2149 TCP_ECN_send_synack(tcp_sk(sk
), skb
);
2151 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2152 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2156 * Prepare a SYN-ACK.
2158 struct sk_buff
* tcp_make_synack(struct sock
*sk
, struct dst_entry
*dst
,
2159 struct request_sock
*req
)
2161 struct inet_request_sock
*ireq
= inet_rsk(req
);
2162 struct tcp_sock
*tp
= tcp_sk(sk
);
2164 int tcp_header_size
;
2165 struct sk_buff
*skb
;
2166 #ifdef CONFIG_TCP_MD5SIG
2167 struct tcp_md5sig_key
*md5
;
2168 __u8
*md5_hash_location
;
2171 skb
= sock_wmalloc(sk
, MAX_TCP_HEADER
+ 15, 1, GFP_ATOMIC
);
2175 /* Reserve space for headers. */
2176 skb_reserve(skb
, MAX_TCP_HEADER
);
2178 skb
->dst
= dst_clone(dst
);
2180 tcp_header_size
= (sizeof(struct tcphdr
) + TCPOLEN_MSS
+
2181 (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0) +
2182 (ireq
->wscale_ok
? TCPOLEN_WSCALE_ALIGNED
: 0) +
2183 /* SACK_PERM is in the place of NOP NOP of TS */
2184 ((ireq
->sack_ok
&& !ireq
->tstamp_ok
) ? TCPOLEN_SACKPERM_ALIGNED
: 0));
2186 #ifdef CONFIG_TCP_MD5SIG
2187 /* Are we doing MD5 on this segment? If so - make room for it */
2188 md5
= tcp_rsk(req
)->af_specific
->md5_lookup(sk
, req
);
2190 tcp_header_size
+= TCPOLEN_MD5SIG_ALIGNED
;
2192 skb_push(skb
, tcp_header_size
);
2193 skb_reset_transport_header(skb
);
2196 memset(th
, 0, sizeof(struct tcphdr
));
2199 TCP_ECN_make_synack(req
, th
);
2200 th
->source
= inet_sk(sk
)->sport
;
2201 th
->dest
= ireq
->rmt_port
;
2202 TCP_SKB_CB(skb
)->seq
= tcp_rsk(req
)->snt_isn
;
2203 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ 1;
2204 TCP_SKB_CB(skb
)->sacked
= 0;
2205 skb_shinfo(skb
)->gso_segs
= 1;
2206 skb_shinfo(skb
)->gso_size
= 0;
2207 skb_shinfo(skb
)->gso_type
= 0;
2208 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
2209 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_isn
+ 1);
2210 if (req
->rcv_wnd
== 0) { /* ignored for retransmitted syns */
2212 /* Set this up on the first call only */
2213 req
->window_clamp
= tp
->window_clamp
? : dst_metric(dst
, RTAX_WINDOW
);
2214 /* tcp_full_space because it is guaranteed to be the first packet */
2215 tcp_select_initial_window(tcp_full_space(sk
),
2216 dst_metric(dst
, RTAX_ADVMSS
) - (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0),
2221 ireq
->rcv_wscale
= rcv_wscale
;
2224 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2225 th
->window
= htons(min(req
->rcv_wnd
, 65535U));
2227 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2228 tcp_syn_build_options((__be32
*)(th
+ 1), dst_metric(dst
, RTAX_ADVMSS
), ireq
->tstamp_ok
,
2229 ireq
->sack_ok
, ireq
->wscale_ok
, ireq
->rcv_wscale
,
2230 TCP_SKB_CB(skb
)->when
,
2233 #ifdef CONFIG_TCP_MD5SIG
2234 md5
? &md5_hash_location
:
2240 th
->doff
= (tcp_header_size
>> 2);
2241 TCP_INC_STATS(TCP_MIB_OUTSEGS
);
2243 #ifdef CONFIG_TCP_MD5SIG
2244 /* Okay, we have all we need - do the md5 hash if needed */
2246 tp
->af_specific
->calc_md5_hash(md5_hash_location
,
2249 tcp_hdr(skb
), sk
->sk_protocol
,
2258 * Do all connect socket setups that can be done AF independent.
2260 static void tcp_connect_init(struct sock
*sk
)
2262 struct dst_entry
*dst
= __sk_dst_get(sk
);
2263 struct tcp_sock
*tp
= tcp_sk(sk
);
2266 /* We'll fix this up when we get a response from the other end.
2267 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2269 tp
->tcp_header_len
= sizeof(struct tcphdr
) +
2270 (sysctl_tcp_timestamps
? TCPOLEN_TSTAMP_ALIGNED
: 0);
2272 #ifdef CONFIG_TCP_MD5SIG
2273 if (tp
->af_specific
->md5_lookup(sk
, sk
) != NULL
)
2274 tp
->tcp_header_len
+= TCPOLEN_MD5SIG_ALIGNED
;
2277 /* If user gave his TCP_MAXSEG, record it to clamp */
2278 if (tp
->rx_opt
.user_mss
)
2279 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
2282 tcp_sync_mss(sk
, dst_mtu(dst
));
2284 if (!tp
->window_clamp
)
2285 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
2286 tp
->advmss
= dst_metric(dst
, RTAX_ADVMSS
);
2287 tcp_initialize_rcv_mss(sk
);
2289 tcp_select_initial_window(tcp_full_space(sk
),
2290 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
2293 sysctl_tcp_window_scaling
,
2296 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
2297 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
2300 sock_reset_flag(sk
, SOCK_DONE
);
2302 tcp_init_wl(tp
, tp
->write_seq
, 0);
2303 tp
->snd_una
= tp
->write_seq
;
2304 tp
->snd_sml
= tp
->write_seq
;
2309 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
2310 inet_csk(sk
)->icsk_retransmits
= 0;
2311 tcp_clear_retrans(tp
);
2315 * Build a SYN and send it off.
2317 int tcp_connect(struct sock
*sk
)
2319 struct tcp_sock
*tp
= tcp_sk(sk
);
2320 struct sk_buff
*buff
;
2322 tcp_connect_init(sk
);
2324 buff
= alloc_skb_fclone(MAX_TCP_HEADER
+ 15, sk
->sk_allocation
);
2325 if (unlikely(buff
== NULL
))
2328 /* Reserve space for headers. */
2329 skb_reserve(buff
, MAX_TCP_HEADER
);
2331 TCP_SKB_CB(buff
)->flags
= TCPCB_FLAG_SYN
;
2332 TCP_ECN_send_syn(sk
, buff
);
2333 TCP_SKB_CB(buff
)->sacked
= 0;
2334 skb_shinfo(buff
)->gso_segs
= 1;
2335 skb_shinfo(buff
)->gso_size
= 0;
2336 skb_shinfo(buff
)->gso_type
= 0;
2338 tp
->snd_nxt
= tp
->write_seq
;
2339 TCP_SKB_CB(buff
)->seq
= tp
->write_seq
++;
2340 TCP_SKB_CB(buff
)->end_seq
= tp
->write_seq
;
2343 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2344 tp
->retrans_stamp
= TCP_SKB_CB(buff
)->when
;
2345 skb_header_release(buff
);
2346 __tcp_add_write_queue_tail(sk
, buff
);
2347 sk_charge_skb(sk
, buff
);
2348 tp
->packets_out
+= tcp_skb_pcount(buff
);
2349 tcp_transmit_skb(sk
, buff
, 1, GFP_KERNEL
);
2351 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2352 * in order to make this packet get counted in tcpOutSegs.
2354 tp
->snd_nxt
= tp
->write_seq
;
2355 tp
->pushed_seq
= tp
->write_seq
;
2356 TCP_INC_STATS(TCP_MIB_ACTIVEOPENS
);
2358 /* Timer for repeating the SYN until an answer. */
2359 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2360 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2364 /* Send out a delayed ack, the caller does the policy checking
2365 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2368 void tcp_send_delayed_ack(struct sock
*sk
)
2370 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2371 int ato
= icsk
->icsk_ack
.ato
;
2372 unsigned long timeout
;
2374 if (ato
> TCP_DELACK_MIN
) {
2375 const struct tcp_sock
*tp
= tcp_sk(sk
);
2378 if (icsk
->icsk_ack
.pingpong
|| (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
2379 max_ato
= TCP_DELACK_MAX
;
2381 /* Slow path, intersegment interval is "high". */
2383 /* If some rtt estimate is known, use it to bound delayed ack.
2384 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2388 int rtt
= max(tp
->srtt
>>3, TCP_DELACK_MIN
);
2394 ato
= min(ato
, max_ato
);
2397 /* Stay within the limit we were given */
2398 timeout
= jiffies
+ ato
;
2400 /* Use new timeout only if there wasn't a older one earlier. */
2401 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
2402 /* If delack timer was blocked or is about to expire,
2405 if (icsk
->icsk_ack
.blocked
||
2406 time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
2411 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
2412 timeout
= icsk
->icsk_ack
.timeout
;
2414 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
2415 icsk
->icsk_ack
.timeout
= timeout
;
2416 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
2419 /* This routine sends an ack and also updates the window. */
2420 void tcp_send_ack(struct sock
*sk
)
2422 /* If we have been reset, we may not send again. */
2423 if (sk
->sk_state
!= TCP_CLOSE
) {
2424 struct sk_buff
*buff
;
2426 /* We are not putting this on the write queue, so
2427 * tcp_transmit_skb() will set the ownership to this
2430 buff
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2432 inet_csk_schedule_ack(sk
);
2433 inet_csk(sk
)->icsk_ack
.ato
= TCP_ATO_MIN
;
2434 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
2435 TCP_DELACK_MAX
, TCP_RTO_MAX
);
2439 /* Reserve space for headers and prepare control bits. */
2440 skb_reserve(buff
, MAX_TCP_HEADER
);
2442 TCP_SKB_CB(buff
)->flags
= TCPCB_FLAG_ACK
;
2443 TCP_SKB_CB(buff
)->sacked
= 0;
2444 skb_shinfo(buff
)->gso_segs
= 1;
2445 skb_shinfo(buff
)->gso_size
= 0;
2446 skb_shinfo(buff
)->gso_type
= 0;
2448 /* Send it off, this clears delayed acks for us. */
2449 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(buff
)->end_seq
= tcp_acceptable_seq(sk
);
2450 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2451 tcp_transmit_skb(sk
, buff
, 0, GFP_ATOMIC
);
2455 /* This routine sends a packet with an out of date sequence
2456 * number. It assumes the other end will try to ack it.
2458 * Question: what should we make while urgent mode?
2459 * 4.4BSD forces sending single byte of data. We cannot send
2460 * out of window data, because we have SND.NXT==SND.MAX...
2462 * Current solution: to send TWO zero-length segments in urgent mode:
2463 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2464 * out-of-date with SND.UNA-1 to probe window.
2466 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
)
2468 struct tcp_sock
*tp
= tcp_sk(sk
);
2469 struct sk_buff
*skb
;
2471 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2472 skb
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2476 /* Reserve space for headers and set control bits. */
2477 skb_reserve(skb
, MAX_TCP_HEADER
);
2479 TCP_SKB_CB(skb
)->flags
= TCPCB_FLAG_ACK
;
2480 TCP_SKB_CB(skb
)->sacked
= urgent
;
2481 skb_shinfo(skb
)->gso_segs
= 1;
2482 skb_shinfo(skb
)->gso_size
= 0;
2483 skb_shinfo(skb
)->gso_type
= 0;
2485 /* Use a previous sequence. This should cause the other
2486 * end to send an ack. Don't queue or clone SKB, just
2489 TCP_SKB_CB(skb
)->seq
= urgent
? tp
->snd_una
: tp
->snd_una
- 1;
2490 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
;
2491 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2492 return tcp_transmit_skb(sk
, skb
, 0, GFP_ATOMIC
);
2495 int tcp_write_wakeup(struct sock
*sk
)
2497 if (sk
->sk_state
!= TCP_CLOSE
) {
2498 struct tcp_sock
*tp
= tcp_sk(sk
);
2499 struct sk_buff
*skb
;
2501 if ((skb
= tcp_send_head(sk
)) != NULL
&&
2502 before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
+tp
->snd_wnd
)) {
2504 unsigned int mss
= tcp_current_mss(sk
, 0);
2505 unsigned int seg_size
= tp
->snd_una
+tp
->snd_wnd
-TCP_SKB_CB(skb
)->seq
;
2507 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
2508 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
2510 /* We are probing the opening of a window
2511 * but the window size is != 0
2512 * must have been a result SWS avoidance ( sender )
2514 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
2516 seg_size
= min(seg_size
, mss
);
2517 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2518 if (tcp_fragment(sk
, skb
, seg_size
, mss
))
2520 } else if (!tcp_skb_pcount(skb
))
2521 tcp_set_skb_tso_segs(sk
, skb
, mss
);
2523 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2524 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2525 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2527 update_send_head(sk
, skb
);
2532 between(tp
->snd_up
, tp
->snd_una
+1, tp
->snd_una
+0xFFFF))
2533 tcp_xmit_probe_skb(sk
, TCPCB_URG
);
2534 return tcp_xmit_probe_skb(sk
, 0);
2540 /* A window probe timeout has occurred. If window is not closed send
2541 * a partial packet else a zero probe.
2543 void tcp_send_probe0(struct sock
*sk
)
2545 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2546 struct tcp_sock
*tp
= tcp_sk(sk
);
2549 err
= tcp_write_wakeup(sk
);
2551 if (tp
->packets_out
|| !tcp_send_head(sk
)) {
2552 /* Cancel probe timer, if it is not required. */
2553 icsk
->icsk_probes_out
= 0;
2554 icsk
->icsk_backoff
= 0;
2559 if (icsk
->icsk_backoff
< sysctl_tcp_retries2
)
2560 icsk
->icsk_backoff
++;
2561 icsk
->icsk_probes_out
++;
2562 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2563 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
2566 /* If packet was not sent due to local congestion,
2567 * do not backoff and do not remember icsk_probes_out.
2568 * Let local senders to fight for local resources.
2570 * Use accumulated backoff yet.
2572 if (!icsk
->icsk_probes_out
)
2573 icsk
->icsk_probes_out
= 1;
2574 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2575 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
,
2576 TCP_RESOURCE_PROBE_INTERVAL
),
2581 EXPORT_SYMBOL(tcp_connect
);
2582 EXPORT_SYMBOL(tcp_make_synack
);
2583 EXPORT_SYMBOL(tcp_simple_retransmit
);
2584 EXPORT_SYMBOL(tcp_sync_mss
);
2585 EXPORT_SYMBOL(sysctl_tcp_tso_win_divisor
);
2586 EXPORT_SYMBOL(tcp_mtup_init
);