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
);
85 /* SND.NXT, if window was not shrunk.
86 * If window has been shrunk, what should we make? It is not clear at all.
87 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
88 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
89 * invalid. OK, let's make this for now:
91 static inline __u32
tcp_acceptable_seq(struct sock
*sk
)
93 struct tcp_sock
*tp
= tcp_sk(sk
);
95 if (!before(tp
->snd_una
+tp
->snd_wnd
, tp
->snd_nxt
))
98 return tp
->snd_una
+tp
->snd_wnd
;
101 /* Calculate mss to advertise in SYN segment.
102 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
104 * 1. It is independent of path mtu.
105 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
106 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
107 * attached devices, because some buggy hosts are confused by
109 * 4. We do not make 3, we advertise MSS, calculated from first
110 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
111 * This may be overridden via information stored in routing table.
112 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
113 * probably even Jumbo".
115 static __u16
tcp_advertise_mss(struct sock
*sk
)
117 struct tcp_sock
*tp
= tcp_sk(sk
);
118 struct dst_entry
*dst
= __sk_dst_get(sk
);
119 int mss
= tp
->advmss
;
121 if (dst
&& dst_metric(dst
, RTAX_ADVMSS
) < mss
) {
122 mss
= dst_metric(dst
, RTAX_ADVMSS
);
129 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
130 * This is the first part of cwnd validation mechanism. */
131 static void tcp_cwnd_restart(struct sock
*sk
, struct dst_entry
*dst
)
133 struct tcp_sock
*tp
= tcp_sk(sk
);
134 s32 delta
= tcp_time_stamp
- tp
->lsndtime
;
135 u32 restart_cwnd
= tcp_init_cwnd(tp
, dst
);
136 u32 cwnd
= tp
->snd_cwnd
;
138 tcp_ca_event(sk
, CA_EVENT_CWND_RESTART
);
140 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
141 restart_cwnd
= min(restart_cwnd
, cwnd
);
143 while ((delta
-= inet_csk(sk
)->icsk_rto
) > 0 && cwnd
> restart_cwnd
)
145 tp
->snd_cwnd
= max(cwnd
, restart_cwnd
);
146 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
147 tp
->snd_cwnd_used
= 0;
150 static void tcp_event_data_sent(struct tcp_sock
*tp
,
151 struct sk_buff
*skb
, struct sock
*sk
)
153 struct inet_connection_sock
*icsk
= inet_csk(sk
);
154 const u32 now
= tcp_time_stamp
;
156 if (sysctl_tcp_slow_start_after_idle
&&
157 (!tp
->packets_out
&& (s32
)(now
- tp
->lsndtime
) > icsk
->icsk_rto
))
158 tcp_cwnd_restart(sk
, __sk_dst_get(sk
));
162 /* If it is a reply for ato after last received
163 * packet, enter pingpong mode.
165 if ((u32
)(now
- icsk
->icsk_ack
.lrcvtime
) < icsk
->icsk_ack
.ato
)
166 icsk
->icsk_ack
.pingpong
= 1;
169 static inline void tcp_event_ack_sent(struct sock
*sk
, unsigned int pkts
)
171 tcp_dec_quickack_mode(sk
, pkts
);
172 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_DACK
);
175 /* Determine a window scaling and initial window to offer.
176 * Based on the assumption that the given amount of space
177 * will be offered. Store the results in the tp structure.
178 * NOTE: for smooth operation initial space offering should
179 * be a multiple of mss if possible. We assume here that mss >= 1.
180 * This MUST be enforced by all callers.
182 void tcp_select_initial_window(int __space
, __u32 mss
,
183 __u32
*rcv_wnd
, __u32
*window_clamp
,
184 int wscale_ok
, __u8
*rcv_wscale
)
186 unsigned int space
= (__space
< 0 ? 0 : __space
);
188 /* If no clamp set the clamp to the max possible scaled window */
189 if (*window_clamp
== 0)
190 (*window_clamp
) = (65535 << 14);
191 space
= min(*window_clamp
, space
);
193 /* Quantize space offering to a multiple of mss if possible. */
195 space
= (space
/ mss
) * mss
;
197 /* NOTE: offering an initial window larger than 32767
198 * will break some buggy TCP stacks. If the admin tells us
199 * it is likely we could be speaking with such a buggy stack
200 * we will truncate our initial window offering to 32K-1
201 * unless the remote has sent us a window scaling option,
202 * which we interpret as a sign the remote TCP is not
203 * misinterpreting the window field as a signed quantity.
205 if (sysctl_tcp_workaround_signed_windows
)
206 (*rcv_wnd
) = min(space
, MAX_TCP_WINDOW
);
212 /* Set window scaling on max possible window
213 * See RFC1323 for an explanation of the limit to 14
215 space
= max_t(u32
, sysctl_tcp_rmem
[2], sysctl_rmem_max
);
216 space
= min_t(u32
, space
, *window_clamp
);
217 while (space
> 65535 && (*rcv_wscale
) < 14) {
223 /* Set initial window to value enough for senders,
224 * following RFC2414. Senders, not following this RFC,
225 * will be satisfied with 2.
227 if (mss
> (1<<*rcv_wscale
)) {
233 if (*rcv_wnd
> init_cwnd
*mss
)
234 *rcv_wnd
= init_cwnd
*mss
;
237 /* Set the clamp no higher than max representable value */
238 (*window_clamp
) = min(65535U << (*rcv_wscale
), *window_clamp
);
241 /* Chose a new window to advertise, update state in tcp_sock for the
242 * socket, and return result with RFC1323 scaling applied. The return
243 * value can be stuffed directly into th->window for an outgoing
246 static u16
tcp_select_window(struct sock
*sk
)
248 struct tcp_sock
*tp
= tcp_sk(sk
);
249 u32 cur_win
= tcp_receive_window(tp
);
250 u32 new_win
= __tcp_select_window(sk
);
252 /* Never shrink the offered window */
253 if (new_win
< cur_win
) {
254 /* Danger Will Robinson!
255 * Don't update rcv_wup/rcv_wnd here or else
256 * we will not be able to advertise a zero
257 * window in time. --DaveM
259 * Relax Will Robinson.
263 tp
->rcv_wnd
= new_win
;
264 tp
->rcv_wup
= tp
->rcv_nxt
;
266 /* Make sure we do not exceed the maximum possible
269 if (!tp
->rx_opt
.rcv_wscale
&& sysctl_tcp_workaround_signed_windows
)
270 new_win
= min(new_win
, MAX_TCP_WINDOW
);
272 new_win
= min(new_win
, (65535U << tp
->rx_opt
.rcv_wscale
));
274 /* RFC1323 scaling applied */
275 new_win
>>= tp
->rx_opt
.rcv_wscale
;
277 /* If we advertise zero window, disable fast path. */
284 static inline void TCP_ECN_send_synack(struct tcp_sock
*tp
,
287 TCP_SKB_CB(skb
)->flags
&= ~TCPCB_FLAG_CWR
;
288 if (!(tp
->ecn_flags
&TCP_ECN_OK
))
289 TCP_SKB_CB(skb
)->flags
&= ~TCPCB_FLAG_ECE
;
292 static inline void TCP_ECN_send_syn(struct sock
*sk
, struct sk_buff
*skb
)
294 struct tcp_sock
*tp
= tcp_sk(sk
);
297 if (sysctl_tcp_ecn
) {
298 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_ECE
|TCPCB_FLAG_CWR
;
299 tp
->ecn_flags
= TCP_ECN_OK
;
303 static __inline__
void
304 TCP_ECN_make_synack(struct request_sock
*req
, struct tcphdr
*th
)
306 if (inet_rsk(req
)->ecn_ok
)
310 static inline void TCP_ECN_send(struct sock
*sk
, struct sk_buff
*skb
,
313 struct tcp_sock
*tp
= tcp_sk(sk
);
315 if (tp
->ecn_flags
& TCP_ECN_OK
) {
316 /* Not-retransmitted data segment: set ECT and inject CWR. */
317 if (skb
->len
!= tcp_header_len
&&
318 !before(TCP_SKB_CB(skb
)->seq
, tp
->snd_nxt
)) {
320 if (tp
->ecn_flags
&TCP_ECN_QUEUE_CWR
) {
321 tp
->ecn_flags
&= ~TCP_ECN_QUEUE_CWR
;
322 tcp_hdr(skb
)->cwr
= 1;
323 skb_shinfo(skb
)->gso_type
|= SKB_GSO_TCP_ECN
;
326 /* ACK or retransmitted segment: clear ECT|CE */
327 INET_ECN_dontxmit(sk
);
329 if (tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)
330 tcp_hdr(skb
)->ece
= 1;
334 static void tcp_build_and_update_options(__be32
*ptr
, struct tcp_sock
*tp
,
335 __u32 tstamp
, __u8
**md5_hash
)
337 if (tp
->rx_opt
.tstamp_ok
) {
338 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
340 (TCPOPT_TIMESTAMP
<< 8) |
342 *ptr
++ = htonl(tstamp
);
343 *ptr
++ = htonl(tp
->rx_opt
.ts_recent
);
345 if (tp
->rx_opt
.eff_sacks
) {
346 struct tcp_sack_block
*sp
= tp
->rx_opt
.dsack
? tp
->duplicate_sack
: tp
->selective_acks
;
349 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
352 (TCPOLEN_SACK_BASE
+ (tp
->rx_opt
.eff_sacks
*
353 TCPOLEN_SACK_PERBLOCK
)));
355 for (this_sack
= 0; this_sack
< tp
->rx_opt
.eff_sacks
; this_sack
++) {
356 *ptr
++ = htonl(sp
[this_sack
].start_seq
);
357 *ptr
++ = htonl(sp
[this_sack
].end_seq
);
360 if (tp
->rx_opt
.dsack
) {
361 tp
->rx_opt
.dsack
= 0;
362 tp
->rx_opt
.eff_sacks
--;
365 #ifdef CONFIG_TCP_MD5SIG
367 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
369 (TCPOPT_MD5SIG
<< 8) |
371 *md5_hash
= (__u8
*)ptr
;
376 /* Construct a tcp options header for a SYN or SYN_ACK packet.
377 * If this is every changed make sure to change the definition of
378 * MAX_SYN_SIZE to match the new maximum number of options that you
381 * Note - that with the RFC2385 TCP option, we make room for the
382 * 16 byte MD5 hash. This will be filled in later, so the pointer for the
383 * location to be filled is passed back up.
385 static void tcp_syn_build_options(__be32
*ptr
, int mss
, int ts
, int sack
,
386 int offer_wscale
, int wscale
, __u32 tstamp
,
387 __u32 ts_recent
, __u8
**md5_hash
)
389 /* We always get an MSS option.
390 * The option bytes which will be seen in normal data
391 * packets should timestamps be used, must be in the MSS
392 * advertised. But we subtract them from tp->mss_cache so
393 * that calculations in tcp_sendmsg are simpler etc.
394 * So account for this fact here if necessary. If we
395 * don't do this correctly, as a receiver we won't
396 * recognize data packets as being full sized when we
397 * should, and thus we won't abide by the delayed ACK
399 * SACKs don't matter, we never delay an ACK when we
400 * have any of those going out.
402 *ptr
++ = htonl((TCPOPT_MSS
<< 24) | (TCPOLEN_MSS
<< 16) | mss
);
405 *ptr
++ = htonl((TCPOPT_SACK_PERM
<< 24) |
406 (TCPOLEN_SACK_PERM
<< 16) |
407 (TCPOPT_TIMESTAMP
<< 8) |
410 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
412 (TCPOPT_TIMESTAMP
<< 8) |
414 *ptr
++ = htonl(tstamp
); /* TSVAL */
415 *ptr
++ = htonl(ts_recent
); /* TSECR */
417 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
419 (TCPOPT_SACK_PERM
<< 8) |
422 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
423 (TCPOPT_WINDOW
<< 16) |
424 (TCPOLEN_WINDOW
<< 8) |
426 #ifdef CONFIG_TCP_MD5SIG
428 * If MD5 is enabled, then we set the option, and include the size
429 * (always 18). The actual MD5 hash is added just before the
433 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
435 (TCPOPT_MD5SIG
<< 8) |
437 *md5_hash
= (__u8
*) ptr
;
442 /* This routine actually transmits TCP packets queued in by
443 * tcp_do_sendmsg(). This is used by both the initial
444 * transmission and possible later retransmissions.
445 * All SKB's seen here are completely headerless. It is our
446 * job to build the TCP header, and pass the packet down to
447 * IP so it can do the same plus pass the packet off to the
450 * We are working here with either a clone of the original
451 * SKB, or a fresh unique copy made by the retransmit engine.
453 static int tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int clone_it
, gfp_t gfp_mask
)
455 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
456 struct inet_sock
*inet
;
458 struct tcp_skb_cb
*tcb
;
460 #ifdef CONFIG_TCP_MD5SIG
461 struct tcp_md5sig_key
*md5
;
462 __u8
*md5_hash_location
;
468 BUG_ON(!skb
|| !tcp_skb_pcount(skb
));
470 /* If congestion control is doing timestamping, we must
471 * take such a timestamp before we potentially clone/copy.
473 if (icsk
->icsk_ca_ops
->flags
& TCP_CONG_RTT_STAMP
)
474 __net_timestamp(skb
);
476 if (likely(clone_it
)) {
477 if (unlikely(skb_cloned(skb
)))
478 skb
= pskb_copy(skb
, gfp_mask
);
480 skb
= skb_clone(skb
, gfp_mask
);
487 tcb
= TCP_SKB_CB(skb
);
488 tcp_header_size
= tp
->tcp_header_len
;
490 #define SYSCTL_FLAG_TSTAMPS 0x1
491 #define SYSCTL_FLAG_WSCALE 0x2
492 #define SYSCTL_FLAG_SACK 0x4
495 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
496 tcp_header_size
= sizeof(struct tcphdr
) + TCPOLEN_MSS
;
497 if (sysctl_tcp_timestamps
) {
498 tcp_header_size
+= TCPOLEN_TSTAMP_ALIGNED
;
499 sysctl_flags
|= SYSCTL_FLAG_TSTAMPS
;
501 if (sysctl_tcp_window_scaling
) {
502 tcp_header_size
+= TCPOLEN_WSCALE_ALIGNED
;
503 sysctl_flags
|= SYSCTL_FLAG_WSCALE
;
505 if (sysctl_tcp_sack
) {
506 sysctl_flags
|= SYSCTL_FLAG_SACK
;
507 if (!(sysctl_flags
& SYSCTL_FLAG_TSTAMPS
))
508 tcp_header_size
+= TCPOLEN_SACKPERM_ALIGNED
;
510 } else if (unlikely(tp
->rx_opt
.eff_sacks
)) {
511 /* A SACK is 2 pad bytes, a 2 byte header, plus
512 * 2 32-bit sequence numbers for each SACK block.
514 tcp_header_size
+= (TCPOLEN_SACK_BASE_ALIGNED
+
515 (tp
->rx_opt
.eff_sacks
*
516 TCPOLEN_SACK_PERBLOCK
));
519 if (tcp_packets_in_flight(tp
) == 0)
520 tcp_ca_event(sk
, CA_EVENT_TX_START
);
522 #ifdef CONFIG_TCP_MD5SIG
524 * Are we doing MD5 on this segment? If so - make
527 md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
529 tcp_header_size
+= TCPOLEN_MD5SIG_ALIGNED
;
532 skb_push(skb
, tcp_header_size
);
533 skb_reset_transport_header(skb
);
534 skb_set_owner_w(skb
, sk
);
536 /* Build TCP header and checksum it. */
538 th
->source
= inet
->sport
;
539 th
->dest
= inet
->dport
;
540 th
->seq
= htonl(tcb
->seq
);
541 th
->ack_seq
= htonl(tp
->rcv_nxt
);
542 *(((__be16
*)th
) + 6) = htons(((tcp_header_size
>> 2) << 12) |
545 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
546 /* RFC1323: The window in SYN & SYN/ACK segments
549 th
->window
= htons(min(tp
->rcv_wnd
, 65535U));
551 th
->window
= htons(tcp_select_window(sk
));
556 if (unlikely(tp
->urg_mode
&&
557 between(tp
->snd_up
, tcb
->seq
+1, tcb
->seq
+0xFFFF))) {
558 th
->urg_ptr
= htons(tp
->snd_up
-tcb
->seq
);
562 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
563 tcp_syn_build_options((__be32
*)(th
+ 1),
564 tcp_advertise_mss(sk
),
565 (sysctl_flags
& SYSCTL_FLAG_TSTAMPS
),
566 (sysctl_flags
& SYSCTL_FLAG_SACK
),
567 (sysctl_flags
& SYSCTL_FLAG_WSCALE
),
568 tp
->rx_opt
.rcv_wscale
,
570 tp
->rx_opt
.ts_recent
,
572 #ifdef CONFIG_TCP_MD5SIG
573 md5
? &md5_hash_location
:
577 tcp_build_and_update_options((__be32
*)(th
+ 1),
579 #ifdef CONFIG_TCP_MD5SIG
580 md5
? &md5_hash_location
:
583 TCP_ECN_send(sk
, skb
, tcp_header_size
);
586 #ifdef CONFIG_TCP_MD5SIG
587 /* Calculate the MD5 hash, as we have all we need now */
589 tp
->af_specific
->calc_md5_hash(md5_hash_location
,
598 icsk
->icsk_af_ops
->send_check(sk
, skb
->len
, skb
);
600 if (likely(tcb
->flags
& TCPCB_FLAG_ACK
))
601 tcp_event_ack_sent(sk
, tcp_skb_pcount(skb
));
603 if (skb
->len
!= tcp_header_size
)
604 tcp_event_data_sent(tp
, skb
, sk
);
606 if (after(tcb
->end_seq
, tp
->snd_nxt
) || tcb
->seq
== tcb
->end_seq
)
607 TCP_INC_STATS(TCP_MIB_OUTSEGS
);
609 err
= icsk
->icsk_af_ops
->queue_xmit(skb
, 0);
610 if (likely(err
<= 0))
613 tcp_enter_cwr(sk
, 1);
615 return net_xmit_eval(err
);
617 #undef SYSCTL_FLAG_TSTAMPS
618 #undef SYSCTL_FLAG_WSCALE
619 #undef SYSCTL_FLAG_SACK
623 /* This routine just queue's the buffer
625 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
626 * otherwise socket can stall.
628 static void tcp_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
630 struct tcp_sock
*tp
= tcp_sk(sk
);
632 /* Advance write_seq and place onto the write_queue. */
633 tp
->write_seq
= TCP_SKB_CB(skb
)->end_seq
;
634 skb_header_release(skb
);
635 tcp_add_write_queue_tail(sk
, skb
);
636 sk_charge_skb(sk
, skb
);
639 static void tcp_set_skb_tso_segs(struct sock
*sk
, struct sk_buff
*skb
, unsigned int mss_now
)
641 if (skb
->len
<= mss_now
|| !sk_can_gso(sk
)) {
642 /* Avoid the costly divide in the normal
645 skb_shinfo(skb
)->gso_segs
= 1;
646 skb_shinfo(skb
)->gso_size
= 0;
647 skb_shinfo(skb
)->gso_type
= 0;
649 skb_shinfo(skb
)->gso_segs
= DIV_ROUND_UP(skb
->len
, mss_now
);
650 skb_shinfo(skb
)->gso_size
= mss_now
;
651 skb_shinfo(skb
)->gso_type
= sk
->sk_gso_type
;
655 /* When a modification to fackets out becomes necessary, we need to check
656 * skb is counted to fackets_out or not. Another important thing is to
657 * tweak SACK fastpath hint too as it would overwrite all changes unless
658 * hint is also changed.
660 static void tcp_adjust_fackets_out(struct tcp_sock
*tp
, struct sk_buff
*skb
,
663 if (!tp
->sacked_out
|| tcp_is_reno(tp
))
666 if (!before(tp
->highest_sack
, TCP_SKB_CB(skb
)->seq
))
667 tp
->fackets_out
-= decr
;
669 /* cnt_hint is "off-by-one" compared with fackets_out (see sacktag) */
670 if (tp
->fastpath_skb_hint
!= NULL
&&
671 after(TCP_SKB_CB(tp
->fastpath_skb_hint
)->seq
, TCP_SKB_CB(skb
)->seq
))
672 tp
->fastpath_cnt_hint
-= 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_charge_skb(sk
, buff
);
706 nlen
= skb
->len
- len
- nsize
;
707 buff
->truesize
+= nlen
;
708 skb
->truesize
-= nlen
;
710 /* Correct the sequence numbers. */
711 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
712 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
713 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
715 if (tcp_is_sack(tp
) && tp
->sacked_out
&&
716 (TCP_SKB_CB(skb
)->seq
== tp
->highest_sack
))
717 tp
->highest_sack
= TCP_SKB_CB(buff
)->seq
;
719 /* PSH and FIN should only be set in the second packet. */
720 flags
= TCP_SKB_CB(skb
)->flags
;
721 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
722 TCP_SKB_CB(buff
)->flags
= flags
;
723 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
724 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_AT_TAIL
;
726 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_PARTIAL
) {
727 /* Copy and checksum data tail into the new buffer. */
728 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
, skb_put(buff
, nsize
),
733 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
735 skb
->ip_summed
= CHECKSUM_PARTIAL
;
736 skb_split(skb
, buff
, len
);
739 buff
->ip_summed
= skb
->ip_summed
;
741 /* Looks stupid, but our code really uses when of
742 * skbs, which it never sent before. --ANK
744 TCP_SKB_CB(buff
)->when
= TCP_SKB_CB(skb
)->when
;
745 buff
->tstamp
= skb
->tstamp
;
747 old_factor
= tcp_skb_pcount(skb
);
749 /* Fix up tso_factor for both original and new SKB. */
750 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
751 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
753 /* If this packet has been sent out already, we must
754 * adjust the various packet counters.
756 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
757 int diff
= old_factor
- tcp_skb_pcount(skb
) -
758 tcp_skb_pcount(buff
);
760 tp
->packets_out
-= diff
;
762 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
763 tp
->sacked_out
-= diff
;
764 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
765 tp
->retrans_out
-= diff
;
767 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
)
768 tp
->lost_out
-= diff
;
770 /* Adjust Reno SACK estimate. */
771 if (tcp_is_reno(tp
) && diff
> 0) {
772 tcp_dec_pcount_approx_int(&tp
->sacked_out
, diff
);
773 tcp_verify_left_out(tp
);
775 tcp_adjust_fackets_out(tp
, skb
, diff
);
778 /* Link BUFF into the send queue. */
779 skb_header_release(buff
);
780 tcp_insert_write_queue_after(skb
, buff
, sk
);
785 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
786 * eventually). The difference is that pulled data not copied, but
787 * immediately discarded.
789 static void __pskb_trim_head(struct sk_buff
*skb
, int len
)
795 for (i
=0; i
<skb_shinfo(skb
)->nr_frags
; i
++) {
796 if (skb_shinfo(skb
)->frags
[i
].size
<= eat
) {
797 put_page(skb_shinfo(skb
)->frags
[i
].page
);
798 eat
-= skb_shinfo(skb
)->frags
[i
].size
;
800 skb_shinfo(skb
)->frags
[k
] = skb_shinfo(skb
)->frags
[i
];
802 skb_shinfo(skb
)->frags
[k
].page_offset
+= eat
;
803 skb_shinfo(skb
)->frags
[k
].size
-= eat
;
809 skb_shinfo(skb
)->nr_frags
= k
;
811 skb_reset_tail_pointer(skb
);
812 skb
->data_len
-= len
;
813 skb
->len
= skb
->data_len
;
816 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
818 if (skb_cloned(skb
) &&
819 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
822 /* If len == headlen, we avoid __skb_pull to preserve alignment. */
823 if (unlikely(len
< skb_headlen(skb
)))
824 __skb_pull(skb
, len
);
826 __pskb_trim_head(skb
, len
- skb_headlen(skb
));
828 TCP_SKB_CB(skb
)->seq
+= len
;
829 skb
->ip_summed
= CHECKSUM_PARTIAL
;
831 skb
->truesize
-= len
;
832 sk
->sk_wmem_queued
-= len
;
833 sk
->sk_forward_alloc
+= len
;
834 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
836 /* Any change of skb->len requires recalculation of tso
839 if (tcp_skb_pcount(skb
) > 1)
840 tcp_set_skb_tso_segs(sk
, skb
, tcp_current_mss(sk
, 1));
845 /* Not accounting for SACKs here. */
846 int tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
848 struct tcp_sock
*tp
= tcp_sk(sk
);
849 struct inet_connection_sock
*icsk
= inet_csk(sk
);
852 /* Calculate base mss without TCP options:
853 It is MMS_S - sizeof(tcphdr) of rfc1122
855 mss_now
= pmtu
- icsk
->icsk_af_ops
->net_header_len
- sizeof(struct tcphdr
);
857 /* Clamp it (mss_clamp does not include tcp options) */
858 if (mss_now
> tp
->rx_opt
.mss_clamp
)
859 mss_now
= tp
->rx_opt
.mss_clamp
;
861 /* Now subtract optional transport overhead */
862 mss_now
-= icsk
->icsk_ext_hdr_len
;
864 /* Then reserve room for full set of TCP options and 8 bytes of data */
868 /* Now subtract TCP options size, not including SACKs */
869 mss_now
-= tp
->tcp_header_len
- sizeof(struct tcphdr
);
874 /* Inverse of above */
875 int tcp_mss_to_mtu(struct sock
*sk
, int mss
)
877 struct tcp_sock
*tp
= tcp_sk(sk
);
878 struct inet_connection_sock
*icsk
= inet_csk(sk
);
883 icsk
->icsk_ext_hdr_len
+
884 icsk
->icsk_af_ops
->net_header_len
;
889 void tcp_mtup_init(struct sock
*sk
)
891 struct tcp_sock
*tp
= tcp_sk(sk
);
892 struct inet_connection_sock
*icsk
= inet_csk(sk
);
894 icsk
->icsk_mtup
.enabled
= sysctl_tcp_mtu_probing
> 1;
895 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+ sizeof(struct tcphdr
) +
896 icsk
->icsk_af_ops
->net_header_len
;
897 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, sysctl_tcp_base_mss
);
898 icsk
->icsk_mtup
.probe_size
= 0;
901 /* This function synchronize snd mss to current pmtu/exthdr set.
903 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
904 for TCP options, but includes only bare TCP header.
906 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
907 It is minimum of user_mss and mss received with SYN.
908 It also does not include TCP options.
910 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
912 tp->mss_cache is current effective sending mss, including
913 all tcp options except for SACKs. It is evaluated,
914 taking into account current pmtu, but never exceeds
915 tp->rx_opt.mss_clamp.
917 NOTE1. rfc1122 clearly states that advertised MSS
918 DOES NOT include either tcp or ip options.
920 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
921 are READ ONLY outside this function. --ANK (980731)
924 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
926 struct tcp_sock
*tp
= tcp_sk(sk
);
927 struct inet_connection_sock
*icsk
= inet_csk(sk
);
930 if (icsk
->icsk_mtup
.search_high
> pmtu
)
931 icsk
->icsk_mtup
.search_high
= pmtu
;
933 mss_now
= tcp_mtu_to_mss(sk
, pmtu
);
935 /* Bound mss with half of window */
936 if (tp
->max_window
&& mss_now
> (tp
->max_window
>>1))
937 mss_now
= max((tp
->max_window
>>1), 68U - tp
->tcp_header_len
);
939 /* And store cached results */
940 icsk
->icsk_pmtu_cookie
= pmtu
;
941 if (icsk
->icsk_mtup
.enabled
)
942 mss_now
= min(mss_now
, tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
));
943 tp
->mss_cache
= mss_now
;
948 /* Compute the current effective MSS, taking SACKs and IP options,
949 * and even PMTU discovery events into account.
951 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
952 * cannot be large. However, taking into account rare use of URG, this
955 unsigned int tcp_current_mss(struct sock
*sk
, int large_allowed
)
957 struct tcp_sock
*tp
= tcp_sk(sk
);
958 struct dst_entry
*dst
= __sk_dst_get(sk
);
963 mss_now
= tp
->mss_cache
;
965 if (large_allowed
&& sk_can_gso(sk
) && !tp
->urg_mode
)
969 u32 mtu
= dst_mtu(dst
);
970 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
971 mss_now
= tcp_sync_mss(sk
, mtu
);
974 if (tp
->rx_opt
.eff_sacks
)
975 mss_now
-= (TCPOLEN_SACK_BASE_ALIGNED
+
976 (tp
->rx_opt
.eff_sacks
* TCPOLEN_SACK_PERBLOCK
));
978 #ifdef CONFIG_TCP_MD5SIG
979 if (tp
->af_specific
->md5_lookup(sk
, sk
))
980 mss_now
-= TCPOLEN_MD5SIG_ALIGNED
;
983 xmit_size_goal
= mss_now
;
986 xmit_size_goal
= (65535 -
987 inet_csk(sk
)->icsk_af_ops
->net_header_len
-
988 inet_csk(sk
)->icsk_ext_hdr_len
-
991 if (tp
->max_window
&&
992 (xmit_size_goal
> (tp
->max_window
>> 1)))
993 xmit_size_goal
= max((tp
->max_window
>> 1),
994 68U - tp
->tcp_header_len
);
996 xmit_size_goal
-= (xmit_size_goal
% mss_now
);
998 tp
->xmit_size_goal
= xmit_size_goal
;
1003 /* Congestion window validation. (RFC2861) */
1005 static void tcp_cwnd_validate(struct sock
*sk
)
1007 struct tcp_sock
*tp
= tcp_sk(sk
);
1008 __u32 packets_out
= tp
->packets_out
;
1010 if (packets_out
>= tp
->snd_cwnd
) {
1011 /* Network is feed fully. */
1012 tp
->snd_cwnd_used
= 0;
1013 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1015 /* Network starves. */
1016 if (tp
->packets_out
> tp
->snd_cwnd_used
)
1017 tp
->snd_cwnd_used
= tp
->packets_out
;
1019 if (sysctl_tcp_slow_start_after_idle
&&
1020 (s32
)(tcp_time_stamp
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
)
1021 tcp_cwnd_application_limited(sk
);
1025 static unsigned int tcp_window_allows(struct tcp_sock
*tp
, struct sk_buff
*skb
, unsigned int mss_now
, unsigned int cwnd
)
1027 u32 window
, cwnd_len
;
1029 window
= (tp
->snd_una
+ tp
->snd_wnd
- TCP_SKB_CB(skb
)->seq
);
1030 cwnd_len
= mss_now
* cwnd
;
1031 return min(window
, cwnd_len
);
1034 /* Can at least one segment of SKB be sent right now, according to the
1035 * congestion window rules? If so, return how many segments are allowed.
1037 static inline unsigned int tcp_cwnd_test(struct tcp_sock
*tp
, struct sk_buff
*skb
)
1039 u32 in_flight
, cwnd
;
1041 /* Don't be strict about the congestion window for the final FIN. */
1042 if ((TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
) &&
1043 tcp_skb_pcount(skb
) == 1)
1046 in_flight
= tcp_packets_in_flight(tp
);
1047 cwnd
= tp
->snd_cwnd
;
1048 if (in_flight
< cwnd
)
1049 return (cwnd
- in_flight
);
1054 /* This must be invoked the first time we consider transmitting
1055 * SKB onto the wire.
1057 static int tcp_init_tso_segs(struct sock
*sk
, struct sk_buff
*skb
, unsigned int mss_now
)
1059 int tso_segs
= tcp_skb_pcount(skb
);
1063 tcp_skb_mss(skb
) != mss_now
)) {
1064 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1065 tso_segs
= tcp_skb_pcount(skb
);
1070 static inline int tcp_minshall_check(const struct tcp_sock
*tp
)
1072 return after(tp
->snd_sml
,tp
->snd_una
) &&
1073 !after(tp
->snd_sml
, tp
->snd_nxt
);
1076 /* Return 0, if packet can be sent now without violation Nagle's rules:
1077 * 1. It is full sized.
1078 * 2. Or it contains FIN. (already checked by caller)
1079 * 3. Or TCP_NODELAY was set.
1080 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1081 * With Minshall's modification: all sent small packets are ACKed.
1084 static inline int tcp_nagle_check(const struct tcp_sock
*tp
,
1085 const struct sk_buff
*skb
,
1086 unsigned mss_now
, int nonagle
)
1088 return (skb
->len
< mss_now
&&
1089 ((nonagle
&TCP_NAGLE_CORK
) ||
1092 tcp_minshall_check(tp
))));
1095 /* Return non-zero if the Nagle test allows this packet to be
1098 static inline int tcp_nagle_test(struct tcp_sock
*tp
, struct sk_buff
*skb
,
1099 unsigned int cur_mss
, int nonagle
)
1101 /* Nagle rule does not apply to frames, which sit in the middle of the
1102 * write_queue (they have no chances to get new data).
1104 * This is implemented in the callers, where they modify the 'nonagle'
1105 * argument based upon the location of SKB in the send queue.
1107 if (nonagle
& TCP_NAGLE_PUSH
)
1110 /* Don't use the nagle rule for urgent data (or for the final FIN).
1111 * Nagle can be ignored during F-RTO too (see RFC4138).
1113 if (tp
->urg_mode
|| (tp
->frto_counter
== 2) ||
1114 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
))
1117 if (!tcp_nagle_check(tp
, skb
, cur_mss
, nonagle
))
1123 /* Does at least the first segment of SKB fit into the send window? */
1124 static inline int tcp_snd_wnd_test(struct tcp_sock
*tp
, struct sk_buff
*skb
, unsigned int cur_mss
)
1126 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1128 if (skb
->len
> cur_mss
)
1129 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
1131 return !after(end_seq
, tp
->snd_una
+ tp
->snd_wnd
);
1134 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1135 * should be put on the wire right now. If so, it returns the number of
1136 * packets allowed by the congestion window.
1138 static unsigned int tcp_snd_test(struct sock
*sk
, struct sk_buff
*skb
,
1139 unsigned int cur_mss
, int nonagle
)
1141 struct tcp_sock
*tp
= tcp_sk(sk
);
1142 unsigned int cwnd_quota
;
1144 tcp_init_tso_segs(sk
, skb
, cur_mss
);
1146 if (!tcp_nagle_test(tp
, skb
, cur_mss
, nonagle
))
1149 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1151 !tcp_snd_wnd_test(tp
, skb
, cur_mss
))
1157 int tcp_may_send_now(struct sock
*sk
)
1159 struct tcp_sock
*tp
= tcp_sk(sk
);
1160 struct sk_buff
*skb
= tcp_send_head(sk
);
1163 tcp_snd_test(sk
, skb
, tcp_current_mss(sk
, 1),
1164 (tcp_skb_is_last(sk
, skb
) ?
1169 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1170 * which is put after SKB on the list. It is very much like
1171 * tcp_fragment() except that it may make several kinds of assumptions
1172 * in order to speed up the splitting operation. In particular, we
1173 * know that all the data is in scatter-gather pages, and that the
1174 * packet has never been sent out before (and thus is not cloned).
1176 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
, unsigned int mss_now
)
1178 struct sk_buff
*buff
;
1179 int nlen
= skb
->len
- len
;
1182 /* All of a TSO frame must be composed of paged data. */
1183 if (skb
->len
!= skb
->data_len
)
1184 return tcp_fragment(sk
, skb
, len
, mss_now
);
1186 buff
= sk_stream_alloc_pskb(sk
, 0, 0, GFP_ATOMIC
);
1187 if (unlikely(buff
== NULL
))
1190 sk_charge_skb(sk
, buff
);
1191 buff
->truesize
+= nlen
;
1192 skb
->truesize
-= nlen
;
1194 /* Correct the sequence numbers. */
1195 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1196 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1197 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1199 /* PSH and FIN should only be set in the second packet. */
1200 flags
= TCP_SKB_CB(skb
)->flags
;
1201 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
1202 TCP_SKB_CB(buff
)->flags
= flags
;
1204 /* This packet was never sent out yet, so no SACK bits. */
1205 TCP_SKB_CB(buff
)->sacked
= 0;
1207 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_PARTIAL
;
1208 skb_split(skb
, buff
, len
);
1210 /* Fix up tso_factor for both original and new SKB. */
1211 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1212 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1214 /* Link BUFF into the send queue. */
1215 skb_header_release(buff
);
1216 tcp_insert_write_queue_after(skb
, buff
, sk
);
1221 /* Try to defer sending, if possible, in order to minimize the amount
1222 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1224 * This algorithm is from John Heffner.
1226 static int tcp_tso_should_defer(struct sock
*sk
, struct sk_buff
*skb
)
1228 struct tcp_sock
*tp
= tcp_sk(sk
);
1229 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1230 u32 send_win
, cong_win
, limit
, in_flight
;
1232 if (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
)
1235 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
1238 /* Defer for less than two clock ticks. */
1239 if (!tp
->tso_deferred
&& ((jiffies
<<1)>>1) - (tp
->tso_deferred
>>1) > 1)
1242 in_flight
= tcp_packets_in_flight(tp
);
1244 BUG_ON(tcp_skb_pcount(skb
) <= 1 ||
1245 (tp
->snd_cwnd
<= in_flight
));
1247 send_win
= (tp
->snd_una
+ tp
->snd_wnd
) - TCP_SKB_CB(skb
)->seq
;
1249 /* From in_flight test above, we know that cwnd > in_flight. */
1250 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
1252 limit
= min(send_win
, cong_win
);
1254 /* If a full-sized TSO skb can be sent, do it. */
1258 if (sysctl_tcp_tso_win_divisor
) {
1259 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
1261 /* If at least some fraction of a window is available,
1264 chunk
/= sysctl_tcp_tso_win_divisor
;
1268 /* Different approach, try not to defer past a single
1269 * ACK. Receiver should ACK every other full sized
1270 * frame, so if we have space for more than 3 frames
1273 if (limit
> tcp_max_burst(tp
) * tp
->mss_cache
)
1277 /* Ok, it looks like it is advisable to defer. */
1278 tp
->tso_deferred
= 1 | (jiffies
<<1);
1283 tp
->tso_deferred
= 0;
1287 /* Create a new MTU probe if we are ready.
1288 * Returns 0 if we should wait to probe (no cwnd available),
1289 * 1 if a probe was sent,
1291 static int tcp_mtu_probe(struct sock
*sk
)
1293 struct tcp_sock
*tp
= tcp_sk(sk
);
1294 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1295 struct sk_buff
*skb
, *nskb
, *next
;
1303 /* Not currently probing/verifying,
1305 * have enough cwnd, and
1306 * not SACKing (the variable headers throw things off) */
1307 if (!icsk
->icsk_mtup
.enabled
||
1308 icsk
->icsk_mtup
.probe_size
||
1309 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
||
1310 tp
->snd_cwnd
< 11 ||
1311 tp
->rx_opt
.eff_sacks
)
1314 /* Very simple search strategy: just double the MSS. */
1315 mss_now
= tcp_current_mss(sk
, 0);
1316 probe_size
= 2*tp
->mss_cache
;
1317 size_needed
= probe_size
+ (tp
->reordering
+ 1) * tp
->mss_cache
;
1318 if (probe_size
> tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_high
)) {
1319 /* TODO: set timer for probe_converge_event */
1323 /* Have enough data in the send queue to probe? */
1324 if (tp
->write_seq
- tp
->snd_nxt
< size_needed
)
1327 if (tp
->snd_wnd
< size_needed
)
1329 if (after(tp
->snd_nxt
+ size_needed
, tp
->snd_una
+ tp
->snd_wnd
))
1332 /* Do we need to wait to drain cwnd? */
1333 pif
= tcp_packets_in_flight(tp
);
1334 if (pif
+ 2 > tp
->snd_cwnd
) {
1335 /* With no packets in flight, don't stall. */
1342 /* We're allowed to probe. Build it now. */
1343 if ((nskb
= sk_stream_alloc_skb(sk
, probe_size
, GFP_ATOMIC
)) == NULL
)
1345 sk_charge_skb(sk
, nskb
);
1347 skb
= tcp_send_head(sk
);
1348 tcp_insert_write_queue_before(nskb
, skb
, sk
);
1350 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(skb
)->seq
;
1351 TCP_SKB_CB(nskb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ probe_size
;
1352 TCP_SKB_CB(nskb
)->flags
= TCPCB_FLAG_ACK
;
1353 TCP_SKB_CB(nskb
)->sacked
= 0;
1355 nskb
->ip_summed
= skb
->ip_summed
;
1358 while (len
< probe_size
) {
1359 next
= tcp_write_queue_next(sk
, skb
);
1361 copy
= min_t(int, skb
->len
, probe_size
- len
);
1362 if (nskb
->ip_summed
)
1363 skb_copy_bits(skb
, 0, skb_put(nskb
, copy
), copy
);
1365 nskb
->csum
= skb_copy_and_csum_bits(skb
, 0,
1366 skb_put(nskb
, copy
), copy
, nskb
->csum
);
1368 if (skb
->len
<= copy
) {
1369 /* We've eaten all the data from this skb.
1371 TCP_SKB_CB(nskb
)->flags
|= TCP_SKB_CB(skb
)->flags
;
1372 tcp_unlink_write_queue(skb
, sk
);
1373 sk_stream_free_skb(sk
, skb
);
1375 TCP_SKB_CB(nskb
)->flags
|= TCP_SKB_CB(skb
)->flags
&
1376 ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
1377 if (!skb_shinfo(skb
)->nr_frags
) {
1378 skb_pull(skb
, copy
);
1379 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1380 skb
->csum
= csum_partial(skb
->data
, skb
->len
, 0);
1382 __pskb_trim_head(skb
, copy
);
1383 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1385 TCP_SKB_CB(skb
)->seq
+= copy
;
1391 tcp_init_tso_segs(sk
, nskb
, nskb
->len
);
1393 /* We're ready to send. If this fails, the probe will
1394 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1395 TCP_SKB_CB(nskb
)->when
= tcp_time_stamp
;
1396 if (!tcp_transmit_skb(sk
, nskb
, 1, GFP_ATOMIC
)) {
1397 /* Decrement cwnd here because we are sending
1398 * effectively two packets. */
1400 update_send_head(sk
, nskb
);
1402 icsk
->icsk_mtup
.probe_size
= tcp_mss_to_mtu(sk
, nskb
->len
);
1403 tp
->mtu_probe
.probe_seq_start
= TCP_SKB_CB(nskb
)->seq
;
1404 tp
->mtu_probe
.probe_seq_end
= TCP_SKB_CB(nskb
)->end_seq
;
1413 /* This routine writes packets to the network. It advances the
1414 * send_head. This happens as incoming acks open up the remote
1417 * Returns 1, if no segments are in flight and we have queued segments, but
1418 * cannot send anything now because of SWS or another problem.
1420 static int tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
)
1422 struct tcp_sock
*tp
= tcp_sk(sk
);
1423 struct sk_buff
*skb
;
1424 unsigned int tso_segs
, sent_pkts
;
1428 /* If we are closed, the bytes will have to remain here.
1429 * In time closedown will finish, we empty the write queue and all
1432 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
1437 /* Do MTU probing. */
1438 if ((result
= tcp_mtu_probe(sk
)) == 0) {
1440 } else if (result
> 0) {
1444 while ((skb
= tcp_send_head(sk
))) {
1447 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1450 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1454 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
)))
1457 if (tso_segs
== 1) {
1458 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
1459 (tcp_skb_is_last(sk
, skb
) ?
1460 nonagle
: TCP_NAGLE_PUSH
))))
1463 if (tcp_tso_should_defer(sk
, skb
))
1469 limit
= tcp_window_allows(tp
, skb
,
1470 mss_now
, cwnd_quota
);
1472 if (skb
->len
< limit
) {
1473 unsigned int trim
= skb
->len
% mss_now
;
1476 limit
= skb
->len
- trim
;
1480 if (skb
->len
> limit
&&
1481 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
)))
1484 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1486 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
)))
1489 /* Advance the send_head. This one is sent out.
1490 * This call will increment packets_out.
1492 update_send_head(sk
, skb
);
1494 tcp_minshall_update(tp
, mss_now
, skb
);
1498 if (likely(sent_pkts
)) {
1499 tcp_cwnd_validate(sk
);
1502 return !tp
->packets_out
&& tcp_send_head(sk
);
1505 /* Push out any pending frames which were held back due to
1506 * TCP_CORK or attempt at coalescing tiny packets.
1507 * The socket must be locked by the caller.
1509 void __tcp_push_pending_frames(struct sock
*sk
, unsigned int cur_mss
,
1512 struct sk_buff
*skb
= tcp_send_head(sk
);
1515 if (tcp_write_xmit(sk
, cur_mss
, nonagle
))
1516 tcp_check_probe_timer(sk
);
1520 /* Send _single_ skb sitting at the send head. This function requires
1521 * true push pending frames to setup probe timer etc.
1523 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
1525 struct tcp_sock
*tp
= tcp_sk(sk
);
1526 struct sk_buff
*skb
= tcp_send_head(sk
);
1527 unsigned int tso_segs
, cwnd_quota
;
1529 BUG_ON(!skb
|| skb
->len
< mss_now
);
1531 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1532 cwnd_quota
= tcp_snd_test(sk
, skb
, mss_now
, TCP_NAGLE_PUSH
);
1534 if (likely(cwnd_quota
)) {
1541 limit
= tcp_window_allows(tp
, skb
,
1542 mss_now
, cwnd_quota
);
1544 if (skb
->len
< limit
) {
1545 unsigned int trim
= skb
->len
% mss_now
;
1548 limit
= skb
->len
- trim
;
1552 if (skb
->len
> limit
&&
1553 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
)))
1556 /* Send it out now. */
1557 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1559 if (likely(!tcp_transmit_skb(sk
, skb
, 1, sk
->sk_allocation
))) {
1560 update_send_head(sk
, skb
);
1561 tcp_cwnd_validate(sk
);
1567 /* This function returns the amount that we can raise the
1568 * usable window based on the following constraints
1570 * 1. The window can never be shrunk once it is offered (RFC 793)
1571 * 2. We limit memory per socket
1574 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1575 * RECV.NEXT + RCV.WIN fixed until:
1576 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1578 * i.e. don't raise the right edge of the window until you can raise
1579 * it at least MSS bytes.
1581 * Unfortunately, the recommended algorithm breaks header prediction,
1582 * since header prediction assumes th->window stays fixed.
1584 * Strictly speaking, keeping th->window fixed violates the receiver
1585 * side SWS prevention criteria. The problem is that under this rule
1586 * a stream of single byte packets will cause the right side of the
1587 * window to always advance by a single byte.
1589 * Of course, if the sender implements sender side SWS prevention
1590 * then this will not be a problem.
1592 * BSD seems to make the following compromise:
1594 * If the free space is less than the 1/4 of the maximum
1595 * space available and the free space is less than 1/2 mss,
1596 * then set the window to 0.
1597 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1598 * Otherwise, just prevent the window from shrinking
1599 * and from being larger than the largest representable value.
1601 * This prevents incremental opening of the window in the regime
1602 * where TCP is limited by the speed of the reader side taking
1603 * data out of the TCP receive queue. It does nothing about
1604 * those cases where the window is constrained on the sender side
1605 * because the pipeline is full.
1607 * BSD also seems to "accidentally" limit itself to windows that are a
1608 * multiple of MSS, at least until the free space gets quite small.
1609 * This would appear to be a side effect of the mbuf implementation.
1610 * Combining these two algorithms results in the observed behavior
1611 * of having a fixed window size at almost all times.
1613 * Below we obtain similar behavior by forcing the offered window to
1614 * a multiple of the mss when it is feasible to do so.
1616 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1617 * Regular options like TIMESTAMP are taken into account.
1619 u32
__tcp_select_window(struct sock
*sk
)
1621 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1622 struct tcp_sock
*tp
= tcp_sk(sk
);
1623 /* MSS for the peer's data. Previous versions used mss_clamp
1624 * here. I don't know if the value based on our guesses
1625 * of peer's MSS is better for the performance. It's more correct
1626 * but may be worse for the performance because of rcv_mss
1627 * fluctuations. --SAW 1998/11/1
1629 int mss
= icsk
->icsk_ack
.rcv_mss
;
1630 int free_space
= tcp_space(sk
);
1631 int full_space
= min_t(int, tp
->window_clamp
, tcp_full_space(sk
));
1634 if (mss
> full_space
)
1637 if (free_space
< full_space
/2) {
1638 icsk
->icsk_ack
.quick
= 0;
1640 if (tcp_memory_pressure
)
1641 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U*tp
->advmss
);
1643 if (free_space
< mss
)
1647 if (free_space
> tp
->rcv_ssthresh
)
1648 free_space
= tp
->rcv_ssthresh
;
1650 /* Don't do rounding if we are using window scaling, since the
1651 * scaled window will not line up with the MSS boundary anyway.
1653 window
= tp
->rcv_wnd
;
1654 if (tp
->rx_opt
.rcv_wscale
) {
1655 window
= free_space
;
1657 /* Advertise enough space so that it won't get scaled away.
1658 * Import case: prevent zero window announcement if
1659 * 1<<rcv_wscale > mss.
1661 if (((window
>> tp
->rx_opt
.rcv_wscale
) << tp
->rx_opt
.rcv_wscale
) != window
)
1662 window
= (((window
>> tp
->rx_opt
.rcv_wscale
) + 1)
1663 << tp
->rx_opt
.rcv_wscale
);
1665 /* Get the largest window that is a nice multiple of mss.
1666 * Window clamp already applied above.
1667 * If our current window offering is within 1 mss of the
1668 * free space we just keep it. This prevents the divide
1669 * and multiply from happening most of the time.
1670 * We also don't do any window rounding when the free space
1673 if (window
<= free_space
- mss
|| window
> free_space
)
1674 window
= (free_space
/mss
)*mss
;
1675 else if (mss
== full_space
&&
1676 free_space
> window
+ full_space
/2)
1677 window
= free_space
;
1683 /* Attempt to collapse two adjacent SKB's during retransmission. */
1684 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*skb
, int mss_now
)
1686 struct tcp_sock
*tp
= tcp_sk(sk
);
1687 struct sk_buff
*next_skb
= tcp_write_queue_next(sk
, skb
);
1689 /* The first test we must make is that neither of these two
1690 * SKB's are still referenced by someone else.
1692 if (!skb_cloned(skb
) && !skb_cloned(next_skb
)) {
1693 int skb_size
= skb
->len
, next_skb_size
= next_skb
->len
;
1694 u16 flags
= TCP_SKB_CB(skb
)->flags
;
1696 /* Also punt if next skb has been SACK'd. */
1697 if (TCP_SKB_CB(next_skb
)->sacked
& TCPCB_SACKED_ACKED
)
1700 /* Next skb is out of window. */
1701 if (after(TCP_SKB_CB(next_skb
)->end_seq
, tp
->snd_una
+tp
->snd_wnd
))
1704 /* Punt if not enough space exists in the first SKB for
1705 * the data in the second, or the total combined payload
1706 * would exceed the MSS.
1708 if ((next_skb_size
> skb_tailroom(skb
)) ||
1709 ((skb_size
+ next_skb_size
) > mss_now
))
1712 BUG_ON(tcp_skb_pcount(skb
) != 1 ||
1713 tcp_skb_pcount(next_skb
) != 1);
1715 if (WARN_ON(tcp_is_sack(tp
) && tp
->sacked_out
&&
1716 (TCP_SKB_CB(next_skb
)->seq
== tp
->highest_sack
)))
1719 /* Ok. We will be able to collapse the packet. */
1720 tcp_unlink_write_queue(next_skb
, sk
);
1722 skb_copy_from_linear_data(next_skb
,
1723 skb_put(skb
, next_skb_size
),
1726 if (next_skb
->ip_summed
== CHECKSUM_PARTIAL
)
1727 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1729 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1730 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
1732 /* Update sequence range on original skb. */
1733 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
1735 /* Merge over control information. */
1736 flags
|= TCP_SKB_CB(next_skb
)->flags
; /* This moves PSH/FIN etc. over */
1737 TCP_SKB_CB(skb
)->flags
= flags
;
1739 /* All done, get rid of second SKB and account for it so
1740 * packet counting does not break.
1742 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
&(TCPCB_EVER_RETRANS
|TCPCB_AT_TAIL
);
1743 if (TCP_SKB_CB(next_skb
)->sacked
&TCPCB_SACKED_RETRANS
)
1744 tp
->retrans_out
-= tcp_skb_pcount(next_skb
);
1745 if (TCP_SKB_CB(next_skb
)->sacked
&TCPCB_LOST
)
1746 tp
->lost_out
-= tcp_skb_pcount(next_skb
);
1747 /* Reno case is special. Sigh... */
1748 if (tcp_is_reno(tp
) && tp
->sacked_out
)
1749 tcp_dec_pcount_approx(&tp
->sacked_out
, next_skb
);
1751 tcp_adjust_fackets_out(tp
, next_skb
, tcp_skb_pcount(next_skb
));
1752 tp
->packets_out
-= tcp_skb_pcount(next_skb
);
1754 /* changed transmit queue under us so clear hints */
1755 tcp_clear_retrans_hints_partial(tp
);
1756 /* manually tune sacktag skb hint */
1757 if (tp
->fastpath_skb_hint
== next_skb
) {
1758 tp
->fastpath_skb_hint
= skb
;
1759 tp
->fastpath_cnt_hint
-= tcp_skb_pcount(skb
);
1762 sk_stream_free_skb(sk
, next_skb
);
1766 /* Do a simple retransmit without using the backoff mechanisms in
1767 * tcp_timer. This is used for path mtu discovery.
1768 * The socket is already locked here.
1770 void tcp_simple_retransmit(struct sock
*sk
)
1772 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1773 struct tcp_sock
*tp
= tcp_sk(sk
);
1774 struct sk_buff
*skb
;
1775 unsigned int mss
= tcp_current_mss(sk
, 0);
1778 tcp_for_write_queue(skb
, sk
) {
1779 if (skb
== tcp_send_head(sk
))
1781 if (skb
->len
> mss
&&
1782 !(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
)) {
1783 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) {
1784 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1785 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1787 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_LOST
)) {
1788 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1789 tp
->lost_out
+= tcp_skb_pcount(skb
);
1795 tcp_clear_all_retrans_hints(tp
);
1800 tcp_verify_left_out(tp
);
1802 /* Don't muck with the congestion window here.
1803 * Reason is that we do not increase amount of _data_
1804 * in network, but units changed and effective
1805 * cwnd/ssthresh really reduced now.
1807 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
1808 tp
->high_seq
= tp
->snd_nxt
;
1809 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
1810 tp
->prior_ssthresh
= 0;
1811 tp
->undo_marker
= 0;
1812 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1814 tcp_xmit_retransmit_queue(sk
);
1817 /* This retransmits one SKB. Policy decisions and retransmit queue
1818 * state updates are done by the caller. Returns non-zero if an
1819 * error occurred which prevented the send.
1821 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
1823 struct tcp_sock
*tp
= tcp_sk(sk
);
1824 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1825 unsigned int cur_mss
= tcp_current_mss(sk
, 0);
1828 /* Inconslusive MTU probe */
1829 if (icsk
->icsk_mtup
.probe_size
) {
1830 icsk
->icsk_mtup
.probe_size
= 0;
1833 /* Do not sent more than we queued. 1/4 is reserved for possible
1834 * copying overhead: fragmentation, tunneling, mangling etc.
1836 if (atomic_read(&sk
->sk_wmem_alloc
) >
1837 min(sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2), sk
->sk_sndbuf
))
1840 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
1841 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1843 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
1847 /* If receiver has shrunk his window, and skb is out of
1848 * new window, do not retransmit it. The exception is the
1849 * case, when window is shrunk to zero. In this case
1850 * our retransmit serves as a zero window probe.
1852 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
+tp
->snd_wnd
)
1853 && TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
1856 if (skb
->len
> cur_mss
) {
1857 if (tcp_fragment(sk
, skb
, cur_mss
, cur_mss
))
1858 return -ENOMEM
; /* We'll try again later. */
1861 /* Collapse two adjacent packets if worthwhile and we can. */
1862 if (!(TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_SYN
) &&
1863 (skb
->len
< (cur_mss
>> 1)) &&
1864 (tcp_write_queue_next(sk
, skb
) != tcp_send_head(sk
)) &&
1865 (!tcp_skb_is_last(sk
, skb
)) &&
1866 (skb_shinfo(skb
)->nr_frags
== 0 && skb_shinfo(tcp_write_queue_next(sk
, skb
))->nr_frags
== 0) &&
1867 (tcp_skb_pcount(skb
) == 1 && tcp_skb_pcount(tcp_write_queue_next(sk
, skb
)) == 1) &&
1868 (sysctl_tcp_retrans_collapse
!= 0))
1869 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
1871 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
1872 return -EHOSTUNREACH
; /* Routing failure or similar. */
1874 /* Some Solaris stacks overoptimize and ignore the FIN on a
1875 * retransmit when old data is attached. So strip it off
1876 * since it is cheap to do so and saves bytes on the network.
1879 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
) &&
1880 tp
->snd_una
== (TCP_SKB_CB(skb
)->end_seq
- 1)) {
1881 if (!pskb_trim(skb
, 0)) {
1882 TCP_SKB_CB(skb
)->seq
= TCP_SKB_CB(skb
)->end_seq
- 1;
1883 skb_shinfo(skb
)->gso_segs
= 1;
1884 skb_shinfo(skb
)->gso_size
= 0;
1885 skb_shinfo(skb
)->gso_type
= 0;
1886 skb
->ip_summed
= CHECKSUM_NONE
;
1891 /* Make a copy, if the first transmission SKB clone we made
1892 * is still in somebody's hands, else make a clone.
1894 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1896 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
1899 /* Update global TCP statistics. */
1900 TCP_INC_STATS(TCP_MIB_RETRANSSEGS
);
1902 tp
->total_retrans
++;
1904 #if FASTRETRANS_DEBUG > 0
1905 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) {
1906 if (net_ratelimit())
1907 printk(KERN_DEBUG
"retrans_out leaked.\n");
1910 if (!tp
->retrans_out
)
1911 tp
->lost_retrans_low
= tp
->snd_nxt
;
1912 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
1913 tp
->retrans_out
+= tcp_skb_pcount(skb
);
1915 /* Save stamp of the first retransmit. */
1916 if (!tp
->retrans_stamp
)
1917 tp
->retrans_stamp
= TCP_SKB_CB(skb
)->when
;
1921 /* snd_nxt is stored to detect loss of retransmitted segment,
1922 * see tcp_input.c tcp_sacktag_write_queue().
1924 TCP_SKB_CB(skb
)->ack_seq
= tp
->snd_nxt
;
1929 /* This gets called after a retransmit timeout, and the initially
1930 * retransmitted data is acknowledged. It tries to continue
1931 * resending the rest of the retransmit queue, until either
1932 * we've sent it all or the congestion window limit is reached.
1933 * If doing SACK, the first ACK which comes back for a timeout
1934 * based retransmit packet might feed us FACK information again.
1935 * If so, we use it to avoid unnecessarily retransmissions.
1937 void tcp_xmit_retransmit_queue(struct sock
*sk
)
1939 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1940 struct tcp_sock
*tp
= tcp_sk(sk
);
1941 struct sk_buff
*skb
;
1944 if (tp
->retransmit_skb_hint
) {
1945 skb
= tp
->retransmit_skb_hint
;
1946 packet_cnt
= tp
->retransmit_cnt_hint
;
1948 skb
= tcp_write_queue_head(sk
);
1952 /* First pass: retransmit lost packets. */
1954 tcp_for_write_queue_from(skb
, sk
) {
1955 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
1957 if (skb
== tcp_send_head(sk
))
1959 /* we could do better than to assign each time */
1960 tp
->retransmit_skb_hint
= skb
;
1961 tp
->retransmit_cnt_hint
= packet_cnt
;
1963 /* Assume this retransmit will generate
1964 * only one packet for congestion window
1965 * calculation purposes. This works because
1966 * tcp_retransmit_skb() will chop up the
1967 * packet to be MSS sized and all the
1968 * packet counting works out.
1970 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
1973 if (sacked
& TCPCB_LOST
) {
1974 if (!(sacked
&(TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))) {
1975 if (tcp_retransmit_skb(sk
, skb
)) {
1976 tp
->retransmit_skb_hint
= NULL
;
1979 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
1980 NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS
);
1982 NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS
);
1984 if (skb
== tcp_write_queue_head(sk
))
1985 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1986 inet_csk(sk
)->icsk_rto
,
1990 packet_cnt
+= tcp_skb_pcount(skb
);
1991 if (packet_cnt
>= tp
->lost_out
)
1997 /* OK, demanded retransmission is finished. */
1999 /* Forward retransmissions are possible only during Recovery. */
2000 if (icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
2003 /* No forward retransmissions in Reno are possible. */
2004 if (tcp_is_reno(tp
))
2007 /* Yeah, we have to make difficult choice between forward transmission
2008 * and retransmission... Both ways have their merits...
2010 * For now we do not retransmit anything, while we have some new
2011 * segments to send. In the other cases, follow rule 3 for
2012 * NextSeg() specified in RFC3517.
2015 if (tcp_may_send_now(sk
))
2018 /* If nothing is SACKed, highest_sack in the loop won't be valid */
2019 if (!tp
->sacked_out
)
2022 if (tp
->forward_skb_hint
)
2023 skb
= tp
->forward_skb_hint
;
2025 skb
= tcp_write_queue_head(sk
);
2027 tcp_for_write_queue_from(skb
, sk
) {
2028 if (skb
== tcp_send_head(sk
))
2030 tp
->forward_skb_hint
= skb
;
2032 if (after(TCP_SKB_CB(skb
)->seq
, tp
->highest_sack
))
2035 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
2038 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
)
2041 /* Ok, retransmit it. */
2042 if (tcp_retransmit_skb(sk
, skb
)) {
2043 tp
->forward_skb_hint
= NULL
;
2047 if (skb
== tcp_write_queue_head(sk
))
2048 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2049 inet_csk(sk
)->icsk_rto
,
2052 NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS
);
2057 /* Send a fin. The caller locks the socket for us. This cannot be
2058 * allowed to fail queueing a FIN frame under any circumstances.
2060 void tcp_send_fin(struct sock
*sk
)
2062 struct tcp_sock
*tp
= tcp_sk(sk
);
2063 struct sk_buff
*skb
= tcp_write_queue_tail(sk
);
2066 /* Optimization, tack on the FIN if we have a queue of
2067 * unsent frames. But be careful about outgoing SACKS
2070 mss_now
= tcp_current_mss(sk
, 1);
2072 if (tcp_send_head(sk
) != NULL
) {
2073 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_FIN
;
2074 TCP_SKB_CB(skb
)->end_seq
++;
2077 /* Socket is locked, keep trying until memory is available. */
2079 skb
= alloc_skb_fclone(MAX_TCP_HEADER
, GFP_KERNEL
);
2085 /* Reserve space for headers and prepare control bits. */
2086 skb_reserve(skb
, MAX_TCP_HEADER
);
2088 TCP_SKB_CB(skb
)->flags
= (TCPCB_FLAG_ACK
| TCPCB_FLAG_FIN
);
2089 TCP_SKB_CB(skb
)->sacked
= 0;
2090 skb_shinfo(skb
)->gso_segs
= 1;
2091 skb_shinfo(skb
)->gso_size
= 0;
2092 skb_shinfo(skb
)->gso_type
= 0;
2094 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2095 TCP_SKB_CB(skb
)->seq
= tp
->write_seq
;
2096 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ 1;
2097 tcp_queue_skb(sk
, skb
);
2099 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_OFF
);
2102 /* We get here when a process closes a file descriptor (either due to
2103 * an explicit close() or as a byproduct of exit()'ing) and there
2104 * was unread data in the receive queue. This behavior is recommended
2105 * by RFC 2525, section 2.17. -DaveM
2107 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
2109 struct sk_buff
*skb
;
2111 /* NOTE: No TCP options attached and we never retransmit this. */
2112 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
2114 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED
);
2118 /* Reserve space for headers and prepare control bits. */
2119 skb_reserve(skb
, MAX_TCP_HEADER
);
2121 TCP_SKB_CB(skb
)->flags
= (TCPCB_FLAG_ACK
| TCPCB_FLAG_RST
);
2122 TCP_SKB_CB(skb
)->sacked
= 0;
2123 skb_shinfo(skb
)->gso_segs
= 1;
2124 skb_shinfo(skb
)->gso_size
= 0;
2125 skb_shinfo(skb
)->gso_type
= 0;
2128 TCP_SKB_CB(skb
)->seq
= tcp_acceptable_seq(sk
);
2129 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
;
2130 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2131 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
2132 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED
);
2135 /* WARNING: This routine must only be called when we have already sent
2136 * a SYN packet that crossed the incoming SYN that caused this routine
2137 * to get called. If this assumption fails then the initial rcv_wnd
2138 * and rcv_wscale values will not be correct.
2140 int tcp_send_synack(struct sock
*sk
)
2142 struct sk_buff
* skb
;
2144 skb
= tcp_write_queue_head(sk
);
2145 if (skb
== NULL
|| !(TCP_SKB_CB(skb
)->flags
&TCPCB_FLAG_SYN
)) {
2146 printk(KERN_DEBUG
"tcp_send_synack: wrong queue state\n");
2149 if (!(TCP_SKB_CB(skb
)->flags
&TCPCB_FLAG_ACK
)) {
2150 if (skb_cloned(skb
)) {
2151 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
2154 tcp_unlink_write_queue(skb
, sk
);
2155 skb_header_release(nskb
);
2156 __tcp_add_write_queue_head(sk
, nskb
);
2157 sk_stream_free_skb(sk
, skb
);
2158 sk_charge_skb(sk
, nskb
);
2162 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_ACK
;
2163 TCP_ECN_send_synack(tcp_sk(sk
), skb
);
2165 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2166 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2170 * Prepare a SYN-ACK.
2172 struct sk_buff
* tcp_make_synack(struct sock
*sk
, struct dst_entry
*dst
,
2173 struct request_sock
*req
)
2175 struct inet_request_sock
*ireq
= inet_rsk(req
);
2176 struct tcp_sock
*tp
= tcp_sk(sk
);
2178 int tcp_header_size
;
2179 struct sk_buff
*skb
;
2180 #ifdef CONFIG_TCP_MD5SIG
2181 struct tcp_md5sig_key
*md5
;
2182 __u8
*md5_hash_location
;
2185 skb
= sock_wmalloc(sk
, MAX_TCP_HEADER
+ 15, 1, GFP_ATOMIC
);
2189 /* Reserve space for headers. */
2190 skb_reserve(skb
, MAX_TCP_HEADER
);
2192 skb
->dst
= dst_clone(dst
);
2194 tcp_header_size
= (sizeof(struct tcphdr
) + TCPOLEN_MSS
+
2195 (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0) +
2196 (ireq
->wscale_ok
? TCPOLEN_WSCALE_ALIGNED
: 0) +
2197 /* SACK_PERM is in the place of NOP NOP of TS */
2198 ((ireq
->sack_ok
&& !ireq
->tstamp_ok
) ? TCPOLEN_SACKPERM_ALIGNED
: 0));
2200 #ifdef CONFIG_TCP_MD5SIG
2201 /* Are we doing MD5 on this segment? If so - make room for it */
2202 md5
= tcp_rsk(req
)->af_specific
->md5_lookup(sk
, req
);
2204 tcp_header_size
+= TCPOLEN_MD5SIG_ALIGNED
;
2206 skb_push(skb
, tcp_header_size
);
2207 skb_reset_transport_header(skb
);
2210 memset(th
, 0, sizeof(struct tcphdr
));
2213 TCP_ECN_make_synack(req
, th
);
2214 th
->source
= inet_sk(sk
)->sport
;
2215 th
->dest
= ireq
->rmt_port
;
2216 TCP_SKB_CB(skb
)->seq
= tcp_rsk(req
)->snt_isn
;
2217 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ 1;
2218 TCP_SKB_CB(skb
)->sacked
= 0;
2219 skb_shinfo(skb
)->gso_segs
= 1;
2220 skb_shinfo(skb
)->gso_size
= 0;
2221 skb_shinfo(skb
)->gso_type
= 0;
2222 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
2223 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_isn
+ 1);
2224 if (req
->rcv_wnd
== 0) { /* ignored for retransmitted syns */
2226 /* Set this up on the first call only */
2227 req
->window_clamp
= tp
->window_clamp
? : dst_metric(dst
, RTAX_WINDOW
);
2228 /* tcp_full_space because it is guaranteed to be the first packet */
2229 tcp_select_initial_window(tcp_full_space(sk
),
2230 dst_metric(dst
, RTAX_ADVMSS
) - (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0),
2235 ireq
->rcv_wscale
= rcv_wscale
;
2238 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2239 th
->window
= htons(min(req
->rcv_wnd
, 65535U));
2241 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2242 tcp_syn_build_options((__be32
*)(th
+ 1), dst_metric(dst
, RTAX_ADVMSS
), ireq
->tstamp_ok
,
2243 ireq
->sack_ok
, ireq
->wscale_ok
, ireq
->rcv_wscale
,
2244 TCP_SKB_CB(skb
)->when
,
2247 #ifdef CONFIG_TCP_MD5SIG
2248 md5
? &md5_hash_location
:
2254 th
->doff
= (tcp_header_size
>> 2);
2255 TCP_INC_STATS(TCP_MIB_OUTSEGS
);
2257 #ifdef CONFIG_TCP_MD5SIG
2258 /* Okay, we have all we need - do the md5 hash if needed */
2260 tp
->af_specific
->calc_md5_hash(md5_hash_location
,
2263 tcp_hdr(skb
), sk
->sk_protocol
,
2272 * Do all connect socket setups that can be done AF independent.
2274 static void tcp_connect_init(struct sock
*sk
)
2276 struct dst_entry
*dst
= __sk_dst_get(sk
);
2277 struct tcp_sock
*tp
= tcp_sk(sk
);
2280 /* We'll fix this up when we get a response from the other end.
2281 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2283 tp
->tcp_header_len
= sizeof(struct tcphdr
) +
2284 (sysctl_tcp_timestamps
? TCPOLEN_TSTAMP_ALIGNED
: 0);
2286 #ifdef CONFIG_TCP_MD5SIG
2287 if (tp
->af_specific
->md5_lookup(sk
, sk
) != NULL
)
2288 tp
->tcp_header_len
+= TCPOLEN_MD5SIG_ALIGNED
;
2291 /* If user gave his TCP_MAXSEG, record it to clamp */
2292 if (tp
->rx_opt
.user_mss
)
2293 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
2296 tcp_sync_mss(sk
, dst_mtu(dst
));
2298 if (!tp
->window_clamp
)
2299 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
2300 tp
->advmss
= dst_metric(dst
, RTAX_ADVMSS
);
2301 tcp_initialize_rcv_mss(sk
);
2303 tcp_select_initial_window(tcp_full_space(sk
),
2304 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
2307 sysctl_tcp_window_scaling
,
2310 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
2311 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
2314 sock_reset_flag(sk
, SOCK_DONE
);
2316 tcp_init_wl(tp
, tp
->write_seq
, 0);
2317 tp
->snd_una
= tp
->write_seq
;
2318 tp
->snd_sml
= tp
->write_seq
;
2323 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
2324 inet_csk(sk
)->icsk_retransmits
= 0;
2325 tcp_clear_retrans(tp
);
2329 * Build a SYN and send it off.
2331 int tcp_connect(struct sock
*sk
)
2333 struct tcp_sock
*tp
= tcp_sk(sk
);
2334 struct sk_buff
*buff
;
2336 tcp_connect_init(sk
);
2338 buff
= alloc_skb_fclone(MAX_TCP_HEADER
+ 15, sk
->sk_allocation
);
2339 if (unlikely(buff
== NULL
))
2342 /* Reserve space for headers. */
2343 skb_reserve(buff
, MAX_TCP_HEADER
);
2345 TCP_SKB_CB(buff
)->flags
= TCPCB_FLAG_SYN
;
2346 TCP_ECN_send_syn(sk
, buff
);
2347 TCP_SKB_CB(buff
)->sacked
= 0;
2348 skb_shinfo(buff
)->gso_segs
= 1;
2349 skb_shinfo(buff
)->gso_size
= 0;
2350 skb_shinfo(buff
)->gso_type
= 0;
2352 tp
->snd_nxt
= tp
->write_seq
;
2353 TCP_SKB_CB(buff
)->seq
= tp
->write_seq
++;
2354 TCP_SKB_CB(buff
)->end_seq
= tp
->write_seq
;
2357 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2358 tp
->retrans_stamp
= TCP_SKB_CB(buff
)->when
;
2359 skb_header_release(buff
);
2360 __tcp_add_write_queue_tail(sk
, buff
);
2361 sk_charge_skb(sk
, buff
);
2362 tp
->packets_out
+= tcp_skb_pcount(buff
);
2363 tcp_transmit_skb(sk
, buff
, 1, GFP_KERNEL
);
2365 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2366 * in order to make this packet get counted in tcpOutSegs.
2368 tp
->snd_nxt
= tp
->write_seq
;
2369 tp
->pushed_seq
= tp
->write_seq
;
2370 TCP_INC_STATS(TCP_MIB_ACTIVEOPENS
);
2372 /* Timer for repeating the SYN until an answer. */
2373 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2374 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2378 /* Send out a delayed ack, the caller does the policy checking
2379 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2382 void tcp_send_delayed_ack(struct sock
*sk
)
2384 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2385 int ato
= icsk
->icsk_ack
.ato
;
2386 unsigned long timeout
;
2388 if (ato
> TCP_DELACK_MIN
) {
2389 const struct tcp_sock
*tp
= tcp_sk(sk
);
2392 if (icsk
->icsk_ack
.pingpong
|| (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
2393 max_ato
= TCP_DELACK_MAX
;
2395 /* Slow path, intersegment interval is "high". */
2397 /* If some rtt estimate is known, use it to bound delayed ack.
2398 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2402 int rtt
= max(tp
->srtt
>>3, TCP_DELACK_MIN
);
2408 ato
= min(ato
, max_ato
);
2411 /* Stay within the limit we were given */
2412 timeout
= jiffies
+ ato
;
2414 /* Use new timeout only if there wasn't a older one earlier. */
2415 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
2416 /* If delack timer was blocked or is about to expire,
2419 if (icsk
->icsk_ack
.blocked
||
2420 time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
2425 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
2426 timeout
= icsk
->icsk_ack
.timeout
;
2428 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
2429 icsk
->icsk_ack
.timeout
= timeout
;
2430 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
2433 /* This routine sends an ack and also updates the window. */
2434 void tcp_send_ack(struct sock
*sk
)
2436 /* If we have been reset, we may not send again. */
2437 if (sk
->sk_state
!= TCP_CLOSE
) {
2438 struct sk_buff
*buff
;
2440 /* We are not putting this on the write queue, so
2441 * tcp_transmit_skb() will set the ownership to this
2444 buff
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2446 inet_csk_schedule_ack(sk
);
2447 inet_csk(sk
)->icsk_ack
.ato
= TCP_ATO_MIN
;
2448 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
2449 TCP_DELACK_MAX
, TCP_RTO_MAX
);
2453 /* Reserve space for headers and prepare control bits. */
2454 skb_reserve(buff
, MAX_TCP_HEADER
);
2456 TCP_SKB_CB(buff
)->flags
= TCPCB_FLAG_ACK
;
2457 TCP_SKB_CB(buff
)->sacked
= 0;
2458 skb_shinfo(buff
)->gso_segs
= 1;
2459 skb_shinfo(buff
)->gso_size
= 0;
2460 skb_shinfo(buff
)->gso_type
= 0;
2462 /* Send it off, this clears delayed acks for us. */
2463 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(buff
)->end_seq
= tcp_acceptable_seq(sk
);
2464 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2465 tcp_transmit_skb(sk
, buff
, 0, GFP_ATOMIC
);
2469 /* This routine sends a packet with an out of date sequence
2470 * number. It assumes the other end will try to ack it.
2472 * Question: what should we make while urgent mode?
2473 * 4.4BSD forces sending single byte of data. We cannot send
2474 * out of window data, because we have SND.NXT==SND.MAX...
2476 * Current solution: to send TWO zero-length segments in urgent mode:
2477 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2478 * out-of-date with SND.UNA-1 to probe window.
2480 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
)
2482 struct tcp_sock
*tp
= tcp_sk(sk
);
2483 struct sk_buff
*skb
;
2485 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2486 skb
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2490 /* Reserve space for headers and set control bits. */
2491 skb_reserve(skb
, MAX_TCP_HEADER
);
2493 TCP_SKB_CB(skb
)->flags
= TCPCB_FLAG_ACK
;
2494 TCP_SKB_CB(skb
)->sacked
= urgent
;
2495 skb_shinfo(skb
)->gso_segs
= 1;
2496 skb_shinfo(skb
)->gso_size
= 0;
2497 skb_shinfo(skb
)->gso_type
= 0;
2499 /* Use a previous sequence. This should cause the other
2500 * end to send an ack. Don't queue or clone SKB, just
2503 TCP_SKB_CB(skb
)->seq
= urgent
? tp
->snd_una
: tp
->snd_una
- 1;
2504 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
;
2505 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2506 return tcp_transmit_skb(sk
, skb
, 0, GFP_ATOMIC
);
2509 int tcp_write_wakeup(struct sock
*sk
)
2511 if (sk
->sk_state
!= TCP_CLOSE
) {
2512 struct tcp_sock
*tp
= tcp_sk(sk
);
2513 struct sk_buff
*skb
;
2515 if ((skb
= tcp_send_head(sk
)) != NULL
&&
2516 before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
+tp
->snd_wnd
)) {
2518 unsigned int mss
= tcp_current_mss(sk
, 0);
2519 unsigned int seg_size
= tp
->snd_una
+tp
->snd_wnd
-TCP_SKB_CB(skb
)->seq
;
2521 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
2522 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
2524 /* We are probing the opening of a window
2525 * but the window size is != 0
2526 * must have been a result SWS avoidance ( sender )
2528 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
2530 seg_size
= min(seg_size
, mss
);
2531 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2532 if (tcp_fragment(sk
, skb
, seg_size
, mss
))
2534 } else if (!tcp_skb_pcount(skb
))
2535 tcp_set_skb_tso_segs(sk
, skb
, mss
);
2537 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2538 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2539 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2541 update_send_head(sk
, skb
);
2546 between(tp
->snd_up
, tp
->snd_una
+1, tp
->snd_una
+0xFFFF))
2547 tcp_xmit_probe_skb(sk
, TCPCB_URG
);
2548 return tcp_xmit_probe_skb(sk
, 0);
2554 /* A window probe timeout has occurred. If window is not closed send
2555 * a partial packet else a zero probe.
2557 void tcp_send_probe0(struct sock
*sk
)
2559 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2560 struct tcp_sock
*tp
= tcp_sk(sk
);
2563 err
= tcp_write_wakeup(sk
);
2565 if (tp
->packets_out
|| !tcp_send_head(sk
)) {
2566 /* Cancel probe timer, if it is not required. */
2567 icsk
->icsk_probes_out
= 0;
2568 icsk
->icsk_backoff
= 0;
2573 if (icsk
->icsk_backoff
< sysctl_tcp_retries2
)
2574 icsk
->icsk_backoff
++;
2575 icsk
->icsk_probes_out
++;
2576 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2577 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
2580 /* If packet was not sent due to local congestion,
2581 * do not backoff and do not remember icsk_probes_out.
2582 * Let local senders to fight for local resources.
2584 * Use accumulated backoff yet.
2586 if (!icsk
->icsk_probes_out
)
2587 icsk
->icsk_probes_out
= 1;
2588 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2589 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
,
2590 TCP_RESOURCE_PROBE_INTERVAL
),
2595 EXPORT_SYMBOL(tcp_connect
);
2596 EXPORT_SYMBOL(tcp_make_synack
);
2597 EXPORT_SYMBOL(tcp_simple_retransmit
);
2598 EXPORT_SYMBOL(tcp_sync_mss
);
2599 EXPORT_SYMBOL(sysctl_tcp_tso_win_divisor
);
2600 EXPORT_SYMBOL(tcp_mtup_init
);