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.
658 static void tcp_adjust_fackets_out(struct sock
*sk
, struct sk_buff
*skb
,
661 struct tcp_sock
*tp
= tcp_sk(sk
);
663 if (!tp
->sacked_out
|| tcp_is_reno(tp
))
666 if (!before(tcp_highest_sack_seq(tp
), TCP_SKB_CB(skb
)->seq
))
667 tp
->fackets_out
-= decr
;
670 /* Function to create two new TCP segments. Shrinks the given segment
671 * to the specified size and appends a new segment with the rest of the
672 * packet to the list. This won't be called frequently, I hope.
673 * Remember, these are still headerless SKBs at this point.
675 int tcp_fragment(struct sock
*sk
, struct sk_buff
*skb
, u32 len
, unsigned int mss_now
)
677 struct tcp_sock
*tp
= tcp_sk(sk
);
678 struct sk_buff
*buff
;
679 int nsize
, old_factor
;
683 BUG_ON(len
> skb
->len
);
685 tcp_clear_retrans_hints_partial(tp
);
686 nsize
= skb_headlen(skb
) - len
;
690 if (skb_cloned(skb
) &&
691 skb_is_nonlinear(skb
) &&
692 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
695 /* Get a new skb... force flag on. */
696 buff
= sk_stream_alloc_skb(sk
, nsize
, GFP_ATOMIC
);
698 return -ENOMEM
; /* We'll just try again later. */
700 sk_charge_skb(sk
, buff
);
701 nlen
= skb
->len
- len
- nsize
;
702 buff
->truesize
+= nlen
;
703 skb
->truesize
-= nlen
;
705 /* Correct the sequence numbers. */
706 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
707 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
708 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
710 if (tcp_is_sack(tp
) && tp
->sacked_out
&& (skb
== tp
->highest_sack
))
711 tp
->highest_sack
= buff
;
713 /* PSH and FIN should only be set in the second packet. */
714 flags
= TCP_SKB_CB(skb
)->flags
;
715 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
716 TCP_SKB_CB(buff
)->flags
= flags
;
717 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
718 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_AT_TAIL
;
720 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_PARTIAL
) {
721 /* Copy and checksum data tail into the new buffer. */
722 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
, skb_put(buff
, nsize
),
727 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
729 skb
->ip_summed
= CHECKSUM_PARTIAL
;
730 skb_split(skb
, buff
, len
);
733 buff
->ip_summed
= skb
->ip_summed
;
735 /* Looks stupid, but our code really uses when of
736 * skbs, which it never sent before. --ANK
738 TCP_SKB_CB(buff
)->when
= TCP_SKB_CB(skb
)->when
;
739 buff
->tstamp
= skb
->tstamp
;
741 old_factor
= tcp_skb_pcount(skb
);
743 /* Fix up tso_factor for both original and new SKB. */
744 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
745 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
747 /* If this packet has been sent out already, we must
748 * adjust the various packet counters.
750 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
751 int diff
= old_factor
- tcp_skb_pcount(skb
) -
752 tcp_skb_pcount(buff
);
754 tp
->packets_out
-= diff
;
756 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
757 tp
->sacked_out
-= diff
;
758 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
759 tp
->retrans_out
-= diff
;
761 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
)
762 tp
->lost_out
-= diff
;
764 /* Adjust Reno SACK estimate. */
765 if (tcp_is_reno(tp
) && diff
> 0) {
766 tcp_dec_pcount_approx_int(&tp
->sacked_out
, diff
);
767 tcp_verify_left_out(tp
);
769 tcp_adjust_fackets_out(sk
, skb
, diff
);
772 /* Link BUFF into the send queue. */
773 skb_header_release(buff
);
774 tcp_insert_write_queue_after(skb
, buff
, sk
);
779 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
780 * eventually). The difference is that pulled data not copied, but
781 * immediately discarded.
783 static void __pskb_trim_head(struct sk_buff
*skb
, int len
)
789 for (i
=0; i
<skb_shinfo(skb
)->nr_frags
; i
++) {
790 if (skb_shinfo(skb
)->frags
[i
].size
<= eat
) {
791 put_page(skb_shinfo(skb
)->frags
[i
].page
);
792 eat
-= skb_shinfo(skb
)->frags
[i
].size
;
794 skb_shinfo(skb
)->frags
[k
] = skb_shinfo(skb
)->frags
[i
];
796 skb_shinfo(skb
)->frags
[k
].page_offset
+= eat
;
797 skb_shinfo(skb
)->frags
[k
].size
-= eat
;
803 skb_shinfo(skb
)->nr_frags
= k
;
805 skb_reset_tail_pointer(skb
);
806 skb
->data_len
-= len
;
807 skb
->len
= skb
->data_len
;
810 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
812 if (skb_cloned(skb
) &&
813 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
816 /* If len == headlen, we avoid __skb_pull to preserve alignment. */
817 if (unlikely(len
< skb_headlen(skb
)))
818 __skb_pull(skb
, len
);
820 __pskb_trim_head(skb
, len
- skb_headlen(skb
));
822 TCP_SKB_CB(skb
)->seq
+= len
;
823 skb
->ip_summed
= CHECKSUM_PARTIAL
;
825 skb
->truesize
-= len
;
826 sk
->sk_wmem_queued
-= len
;
827 sk
->sk_forward_alloc
+= len
;
828 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
830 /* Any change of skb->len requires recalculation of tso
833 if (tcp_skb_pcount(skb
) > 1)
834 tcp_set_skb_tso_segs(sk
, skb
, tcp_current_mss(sk
, 1));
839 /* Not accounting for SACKs here. */
840 int tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
842 struct tcp_sock
*tp
= tcp_sk(sk
);
843 struct inet_connection_sock
*icsk
= inet_csk(sk
);
846 /* Calculate base mss without TCP options:
847 It is MMS_S - sizeof(tcphdr) of rfc1122
849 mss_now
= pmtu
- icsk
->icsk_af_ops
->net_header_len
- sizeof(struct tcphdr
);
851 /* Clamp it (mss_clamp does not include tcp options) */
852 if (mss_now
> tp
->rx_opt
.mss_clamp
)
853 mss_now
= tp
->rx_opt
.mss_clamp
;
855 /* Now subtract optional transport overhead */
856 mss_now
-= icsk
->icsk_ext_hdr_len
;
858 /* Then reserve room for full set of TCP options and 8 bytes of data */
862 /* Now subtract TCP options size, not including SACKs */
863 mss_now
-= tp
->tcp_header_len
- sizeof(struct tcphdr
);
868 /* Inverse of above */
869 int tcp_mss_to_mtu(struct sock
*sk
, int mss
)
871 struct tcp_sock
*tp
= tcp_sk(sk
);
872 struct inet_connection_sock
*icsk
= inet_csk(sk
);
877 icsk
->icsk_ext_hdr_len
+
878 icsk
->icsk_af_ops
->net_header_len
;
883 void tcp_mtup_init(struct sock
*sk
)
885 struct tcp_sock
*tp
= tcp_sk(sk
);
886 struct inet_connection_sock
*icsk
= inet_csk(sk
);
888 icsk
->icsk_mtup
.enabled
= sysctl_tcp_mtu_probing
> 1;
889 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+ sizeof(struct tcphdr
) +
890 icsk
->icsk_af_ops
->net_header_len
;
891 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, sysctl_tcp_base_mss
);
892 icsk
->icsk_mtup
.probe_size
= 0;
895 /* This function synchronize snd mss to current pmtu/exthdr set.
897 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
898 for TCP options, but includes only bare TCP header.
900 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
901 It is minimum of user_mss and mss received with SYN.
902 It also does not include TCP options.
904 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
906 tp->mss_cache is current effective sending mss, including
907 all tcp options except for SACKs. It is evaluated,
908 taking into account current pmtu, but never exceeds
909 tp->rx_opt.mss_clamp.
911 NOTE1. rfc1122 clearly states that advertised MSS
912 DOES NOT include either tcp or ip options.
914 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
915 are READ ONLY outside this function. --ANK (980731)
918 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
920 struct tcp_sock
*tp
= tcp_sk(sk
);
921 struct inet_connection_sock
*icsk
= inet_csk(sk
);
924 if (icsk
->icsk_mtup
.search_high
> pmtu
)
925 icsk
->icsk_mtup
.search_high
= pmtu
;
927 mss_now
= tcp_mtu_to_mss(sk
, pmtu
);
929 /* Bound mss with half of window */
930 if (tp
->max_window
&& mss_now
> (tp
->max_window
>>1))
931 mss_now
= max((tp
->max_window
>>1), 68U - tp
->tcp_header_len
);
933 /* And store cached results */
934 icsk
->icsk_pmtu_cookie
= pmtu
;
935 if (icsk
->icsk_mtup
.enabled
)
936 mss_now
= min(mss_now
, tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
));
937 tp
->mss_cache
= mss_now
;
942 /* Compute the current effective MSS, taking SACKs and IP options,
943 * and even PMTU discovery events into account.
945 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
946 * cannot be large. However, taking into account rare use of URG, this
949 unsigned int tcp_current_mss(struct sock
*sk
, int large_allowed
)
951 struct tcp_sock
*tp
= tcp_sk(sk
);
952 struct dst_entry
*dst
= __sk_dst_get(sk
);
957 mss_now
= tp
->mss_cache
;
959 if (large_allowed
&& sk_can_gso(sk
) && !tp
->urg_mode
)
963 u32 mtu
= dst_mtu(dst
);
964 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
965 mss_now
= tcp_sync_mss(sk
, mtu
);
968 if (tp
->rx_opt
.eff_sacks
)
969 mss_now
-= (TCPOLEN_SACK_BASE_ALIGNED
+
970 (tp
->rx_opt
.eff_sacks
* TCPOLEN_SACK_PERBLOCK
));
972 #ifdef CONFIG_TCP_MD5SIG
973 if (tp
->af_specific
->md5_lookup(sk
, sk
))
974 mss_now
-= TCPOLEN_MD5SIG_ALIGNED
;
977 xmit_size_goal
= mss_now
;
980 xmit_size_goal
= (65535 -
981 inet_csk(sk
)->icsk_af_ops
->net_header_len
-
982 inet_csk(sk
)->icsk_ext_hdr_len
-
985 if (tp
->max_window
&&
986 (xmit_size_goal
> (tp
->max_window
>> 1)))
987 xmit_size_goal
= max((tp
->max_window
>> 1),
988 68U - tp
->tcp_header_len
);
990 xmit_size_goal
-= (xmit_size_goal
% mss_now
);
992 tp
->xmit_size_goal
= xmit_size_goal
;
997 /* Congestion window validation. (RFC2861) */
999 static void tcp_cwnd_validate(struct sock
*sk
)
1001 struct tcp_sock
*tp
= tcp_sk(sk
);
1002 __u32 packets_out
= tp
->packets_out
;
1004 if (packets_out
>= tp
->snd_cwnd
) {
1005 /* Network is feed fully. */
1006 tp
->snd_cwnd_used
= 0;
1007 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1009 /* Network starves. */
1010 if (tp
->packets_out
> tp
->snd_cwnd_used
)
1011 tp
->snd_cwnd_used
= tp
->packets_out
;
1013 if (sysctl_tcp_slow_start_after_idle
&&
1014 (s32
)(tcp_time_stamp
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
)
1015 tcp_cwnd_application_limited(sk
);
1019 static unsigned int tcp_window_allows(struct tcp_sock
*tp
, struct sk_buff
*skb
, unsigned int mss_now
, unsigned int cwnd
)
1021 u32 window
, cwnd_len
;
1023 window
= (tp
->snd_una
+ tp
->snd_wnd
- TCP_SKB_CB(skb
)->seq
);
1024 cwnd_len
= mss_now
* cwnd
;
1025 return min(window
, cwnd_len
);
1028 /* Can at least one segment of SKB be sent right now, according to the
1029 * congestion window rules? If so, return how many segments are allowed.
1031 static inline unsigned int tcp_cwnd_test(struct tcp_sock
*tp
, struct sk_buff
*skb
)
1033 u32 in_flight
, cwnd
;
1035 /* Don't be strict about the congestion window for the final FIN. */
1036 if ((TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
) &&
1037 tcp_skb_pcount(skb
) == 1)
1040 in_flight
= tcp_packets_in_flight(tp
);
1041 cwnd
= tp
->snd_cwnd
;
1042 if (in_flight
< cwnd
)
1043 return (cwnd
- in_flight
);
1048 /* This must be invoked the first time we consider transmitting
1049 * SKB onto the wire.
1051 static int tcp_init_tso_segs(struct sock
*sk
, struct sk_buff
*skb
, unsigned int mss_now
)
1053 int tso_segs
= tcp_skb_pcount(skb
);
1057 tcp_skb_mss(skb
) != mss_now
)) {
1058 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1059 tso_segs
= tcp_skb_pcount(skb
);
1064 static inline int tcp_minshall_check(const struct tcp_sock
*tp
)
1066 return after(tp
->snd_sml
,tp
->snd_una
) &&
1067 !after(tp
->snd_sml
, tp
->snd_nxt
);
1070 /* Return 0, if packet can be sent now without violation Nagle's rules:
1071 * 1. It is full sized.
1072 * 2. Or it contains FIN. (already checked by caller)
1073 * 3. Or TCP_NODELAY was set.
1074 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1075 * With Minshall's modification: all sent small packets are ACKed.
1078 static inline int tcp_nagle_check(const struct tcp_sock
*tp
,
1079 const struct sk_buff
*skb
,
1080 unsigned mss_now
, int nonagle
)
1082 return (skb
->len
< mss_now
&&
1083 ((nonagle
&TCP_NAGLE_CORK
) ||
1086 tcp_minshall_check(tp
))));
1089 /* Return non-zero if the Nagle test allows this packet to be
1092 static inline int tcp_nagle_test(struct tcp_sock
*tp
, struct sk_buff
*skb
,
1093 unsigned int cur_mss
, int nonagle
)
1095 /* Nagle rule does not apply to frames, which sit in the middle of the
1096 * write_queue (they have no chances to get new data).
1098 * This is implemented in the callers, where they modify the 'nonagle'
1099 * argument based upon the location of SKB in the send queue.
1101 if (nonagle
& TCP_NAGLE_PUSH
)
1104 /* Don't use the nagle rule for urgent data (or for the final FIN).
1105 * Nagle can be ignored during F-RTO too (see RFC4138).
1107 if (tp
->urg_mode
|| (tp
->frto_counter
== 2) ||
1108 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
))
1111 if (!tcp_nagle_check(tp
, skb
, cur_mss
, nonagle
))
1117 /* Does at least the first segment of SKB fit into the send window? */
1118 static inline int tcp_snd_wnd_test(struct tcp_sock
*tp
, struct sk_buff
*skb
, unsigned int cur_mss
)
1120 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1122 if (skb
->len
> cur_mss
)
1123 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
1125 return !after(end_seq
, tp
->snd_una
+ tp
->snd_wnd
);
1128 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1129 * should be put on the wire right now. If so, it returns the number of
1130 * packets allowed by the congestion window.
1132 static unsigned int tcp_snd_test(struct sock
*sk
, struct sk_buff
*skb
,
1133 unsigned int cur_mss
, int nonagle
)
1135 struct tcp_sock
*tp
= tcp_sk(sk
);
1136 unsigned int cwnd_quota
;
1138 tcp_init_tso_segs(sk
, skb
, cur_mss
);
1140 if (!tcp_nagle_test(tp
, skb
, cur_mss
, nonagle
))
1143 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1145 !tcp_snd_wnd_test(tp
, skb
, cur_mss
))
1151 int tcp_may_send_now(struct sock
*sk
)
1153 struct tcp_sock
*tp
= tcp_sk(sk
);
1154 struct sk_buff
*skb
= tcp_send_head(sk
);
1157 tcp_snd_test(sk
, skb
, tcp_current_mss(sk
, 1),
1158 (tcp_skb_is_last(sk
, skb
) ?
1159 tp
->nonagle
: TCP_NAGLE_PUSH
)));
1162 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1163 * which is put after SKB on the list. It is very much like
1164 * tcp_fragment() except that it may make several kinds of assumptions
1165 * in order to speed up the splitting operation. In particular, we
1166 * know that all the data is in scatter-gather pages, and that the
1167 * packet has never been sent out before (and thus is not cloned).
1169 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
, unsigned int mss_now
)
1171 struct sk_buff
*buff
;
1172 int nlen
= skb
->len
- len
;
1175 /* All of a TSO frame must be composed of paged data. */
1176 if (skb
->len
!= skb
->data_len
)
1177 return tcp_fragment(sk
, skb
, len
, mss_now
);
1179 buff
= sk_stream_alloc_pskb(sk
, 0, 0, GFP_ATOMIC
);
1180 if (unlikely(buff
== NULL
))
1183 sk_charge_skb(sk
, buff
);
1184 buff
->truesize
+= nlen
;
1185 skb
->truesize
-= nlen
;
1187 /* Correct the sequence numbers. */
1188 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1189 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1190 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1192 /* PSH and FIN should only be set in the second packet. */
1193 flags
= TCP_SKB_CB(skb
)->flags
;
1194 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
1195 TCP_SKB_CB(buff
)->flags
= flags
;
1197 /* This packet was never sent out yet, so no SACK bits. */
1198 TCP_SKB_CB(buff
)->sacked
= 0;
1200 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_PARTIAL
;
1201 skb_split(skb
, buff
, len
);
1203 /* Fix up tso_factor for both original and new SKB. */
1204 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1205 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1207 /* Link BUFF into the send queue. */
1208 skb_header_release(buff
);
1209 tcp_insert_write_queue_after(skb
, buff
, sk
);
1214 /* Try to defer sending, if possible, in order to minimize the amount
1215 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1217 * This algorithm is from John Heffner.
1219 static int tcp_tso_should_defer(struct sock
*sk
, struct sk_buff
*skb
)
1221 struct tcp_sock
*tp
= tcp_sk(sk
);
1222 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1223 u32 send_win
, cong_win
, limit
, in_flight
;
1225 if (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
)
1228 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
1231 /* Defer for less than two clock ticks. */
1232 if (!tp
->tso_deferred
&& ((jiffies
<<1)>>1) - (tp
->tso_deferred
>>1) > 1)
1235 in_flight
= tcp_packets_in_flight(tp
);
1237 BUG_ON(tcp_skb_pcount(skb
) <= 1 ||
1238 (tp
->snd_cwnd
<= in_flight
));
1240 send_win
= (tp
->snd_una
+ tp
->snd_wnd
) - TCP_SKB_CB(skb
)->seq
;
1242 /* From in_flight test above, we know that cwnd > in_flight. */
1243 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
1245 limit
= min(send_win
, cong_win
);
1247 /* If a full-sized TSO skb can be sent, do it. */
1251 if (sysctl_tcp_tso_win_divisor
) {
1252 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
1254 /* If at least some fraction of a window is available,
1257 chunk
/= sysctl_tcp_tso_win_divisor
;
1261 /* Different approach, try not to defer past a single
1262 * ACK. Receiver should ACK every other full sized
1263 * frame, so if we have space for more than 3 frames
1266 if (limit
> tcp_max_burst(tp
) * tp
->mss_cache
)
1270 /* Ok, it looks like it is advisable to defer. */
1271 tp
->tso_deferred
= 1 | (jiffies
<<1);
1276 tp
->tso_deferred
= 0;
1280 /* Create a new MTU probe if we are ready.
1281 * Returns 0 if we should wait to probe (no cwnd available),
1282 * 1 if a probe was sent,
1284 static int tcp_mtu_probe(struct sock
*sk
)
1286 struct tcp_sock
*tp
= tcp_sk(sk
);
1287 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1288 struct sk_buff
*skb
, *nskb
, *next
;
1296 /* Not currently probing/verifying,
1298 * have enough cwnd, and
1299 * not SACKing (the variable headers throw things off) */
1300 if (!icsk
->icsk_mtup
.enabled
||
1301 icsk
->icsk_mtup
.probe_size
||
1302 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
||
1303 tp
->snd_cwnd
< 11 ||
1304 tp
->rx_opt
.eff_sacks
)
1307 /* Very simple search strategy: just double the MSS. */
1308 mss_now
= tcp_current_mss(sk
, 0);
1309 probe_size
= 2*tp
->mss_cache
;
1310 size_needed
= probe_size
+ (tp
->reordering
+ 1) * tp
->mss_cache
;
1311 if (probe_size
> tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_high
)) {
1312 /* TODO: set timer for probe_converge_event */
1316 /* Have enough data in the send queue to probe? */
1317 if (tp
->write_seq
- tp
->snd_nxt
< size_needed
)
1320 if (tp
->snd_wnd
< size_needed
)
1322 if (after(tp
->snd_nxt
+ size_needed
, tp
->snd_una
+ tp
->snd_wnd
))
1325 /* Do we need to wait to drain cwnd? */
1326 pif
= tcp_packets_in_flight(tp
);
1327 if (pif
+ 2 > tp
->snd_cwnd
) {
1328 /* With no packets in flight, don't stall. */
1335 /* We're allowed to probe. Build it now. */
1336 if ((nskb
= sk_stream_alloc_skb(sk
, probe_size
, GFP_ATOMIC
)) == NULL
)
1338 sk_charge_skb(sk
, nskb
);
1340 skb
= tcp_send_head(sk
);
1341 tcp_insert_write_queue_before(nskb
, skb
, sk
);
1343 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(skb
)->seq
;
1344 TCP_SKB_CB(nskb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ probe_size
;
1345 TCP_SKB_CB(nskb
)->flags
= TCPCB_FLAG_ACK
;
1346 TCP_SKB_CB(nskb
)->sacked
= 0;
1348 nskb
->ip_summed
= skb
->ip_summed
;
1351 while (len
< probe_size
) {
1352 next
= tcp_write_queue_next(sk
, skb
);
1354 copy
= min_t(int, skb
->len
, probe_size
- len
);
1355 if (nskb
->ip_summed
)
1356 skb_copy_bits(skb
, 0, skb_put(nskb
, copy
), copy
);
1358 nskb
->csum
= skb_copy_and_csum_bits(skb
, 0,
1359 skb_put(nskb
, copy
), copy
, nskb
->csum
);
1361 if (skb
->len
<= copy
) {
1362 /* We've eaten all the data from this skb.
1364 TCP_SKB_CB(nskb
)->flags
|= TCP_SKB_CB(skb
)->flags
;
1365 tcp_unlink_write_queue(skb
, sk
);
1366 sk_stream_free_skb(sk
, skb
);
1368 TCP_SKB_CB(nskb
)->flags
|= TCP_SKB_CB(skb
)->flags
&
1369 ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
1370 if (!skb_shinfo(skb
)->nr_frags
) {
1371 skb_pull(skb
, copy
);
1372 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1373 skb
->csum
= csum_partial(skb
->data
, skb
->len
, 0);
1375 __pskb_trim_head(skb
, copy
);
1376 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1378 TCP_SKB_CB(skb
)->seq
+= copy
;
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
/2) {
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
/2)
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 if (WARN_ON(tcp_is_sack(tp
) && tp
->sacked_out
&&
1709 (next_skb
== tp
->highest_sack
)))
1712 /* Ok. We will be able to collapse the packet. */
1713 tcp_unlink_write_queue(next_skb
, sk
);
1715 skb_copy_from_linear_data(next_skb
,
1716 skb_put(skb
, next_skb_size
),
1719 if (next_skb
->ip_summed
== CHECKSUM_PARTIAL
)
1720 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1722 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1723 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
1725 /* Update sequence range on original skb. */
1726 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
1728 /* Merge over control information. */
1729 flags
|= TCP_SKB_CB(next_skb
)->flags
; /* This moves PSH/FIN etc. over */
1730 TCP_SKB_CB(skb
)->flags
= flags
;
1732 /* All done, get rid of second SKB and account for it so
1733 * packet counting does not break.
1735 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
&(TCPCB_EVER_RETRANS
|TCPCB_AT_TAIL
);
1736 if (TCP_SKB_CB(next_skb
)->sacked
&TCPCB_SACKED_RETRANS
)
1737 tp
->retrans_out
-= tcp_skb_pcount(next_skb
);
1738 if (TCP_SKB_CB(next_skb
)->sacked
&TCPCB_LOST
)
1739 tp
->lost_out
-= tcp_skb_pcount(next_skb
);
1740 /* Reno case is special. Sigh... */
1741 if (tcp_is_reno(tp
) && tp
->sacked_out
)
1742 tcp_dec_pcount_approx(&tp
->sacked_out
, next_skb
);
1744 tcp_adjust_fackets_out(sk
, next_skb
, tcp_skb_pcount(next_skb
));
1745 tp
->packets_out
-= tcp_skb_pcount(next_skb
);
1747 /* changed transmit queue under us so clear hints */
1748 tcp_clear_retrans_hints_partial(tp
);
1750 sk_stream_free_skb(sk
, next_skb
);
1754 /* Do a simple retransmit without using the backoff mechanisms in
1755 * tcp_timer. This is used for path mtu discovery.
1756 * The socket is already locked here.
1758 void tcp_simple_retransmit(struct sock
*sk
)
1760 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1761 struct tcp_sock
*tp
= tcp_sk(sk
);
1762 struct sk_buff
*skb
;
1763 unsigned int mss
= tcp_current_mss(sk
, 0);
1766 tcp_for_write_queue(skb
, sk
) {
1767 if (skb
== tcp_send_head(sk
))
1769 if (skb
->len
> mss
&&
1770 !(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
)) {
1771 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) {
1772 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1773 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1775 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_LOST
)) {
1776 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1777 tp
->lost_out
+= tcp_skb_pcount(skb
);
1783 tcp_clear_all_retrans_hints(tp
);
1788 tcp_verify_left_out(tp
);
1790 /* Don't muck with the congestion window here.
1791 * Reason is that we do not increase amount of _data_
1792 * in network, but units changed and effective
1793 * cwnd/ssthresh really reduced now.
1795 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
1796 tp
->high_seq
= tp
->snd_nxt
;
1797 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
1798 tp
->prior_ssthresh
= 0;
1799 tp
->undo_marker
= 0;
1800 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1802 tcp_xmit_retransmit_queue(sk
);
1805 /* This retransmits one SKB. Policy decisions and retransmit queue
1806 * state updates are done by the caller. Returns non-zero if an
1807 * error occurred which prevented the send.
1809 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
1811 struct tcp_sock
*tp
= tcp_sk(sk
);
1812 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1813 unsigned int cur_mss
= tcp_current_mss(sk
, 0);
1816 /* Inconslusive MTU probe */
1817 if (icsk
->icsk_mtup
.probe_size
) {
1818 icsk
->icsk_mtup
.probe_size
= 0;
1821 /* Do not sent more than we queued. 1/4 is reserved for possible
1822 * copying overhead: fragmentation, tunneling, mangling etc.
1824 if (atomic_read(&sk
->sk_wmem_alloc
) >
1825 min(sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2), sk
->sk_sndbuf
))
1828 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
1829 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1831 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
1835 /* If receiver has shrunk his window, and skb is out of
1836 * new window, do not retransmit it. The exception is the
1837 * case, when window is shrunk to zero. In this case
1838 * our retransmit serves as a zero window probe.
1840 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
+tp
->snd_wnd
)
1841 && TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
1844 if (skb
->len
> cur_mss
) {
1845 if (tcp_fragment(sk
, skb
, cur_mss
, cur_mss
))
1846 return -ENOMEM
; /* We'll try again later. */
1849 /* Collapse two adjacent packets if worthwhile and we can. */
1850 if (!(TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_SYN
) &&
1851 (skb
->len
< (cur_mss
>> 1)) &&
1852 (tcp_write_queue_next(sk
, skb
) != tcp_send_head(sk
)) &&
1853 (!tcp_skb_is_last(sk
, skb
)) &&
1854 (skb_shinfo(skb
)->nr_frags
== 0 && skb_shinfo(tcp_write_queue_next(sk
, skb
))->nr_frags
== 0) &&
1855 (tcp_skb_pcount(skb
) == 1 && tcp_skb_pcount(tcp_write_queue_next(sk
, skb
)) == 1) &&
1856 (sysctl_tcp_retrans_collapse
!= 0))
1857 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
1859 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
1860 return -EHOSTUNREACH
; /* Routing failure or similar. */
1862 /* Some Solaris stacks overoptimize and ignore the FIN on a
1863 * retransmit when old data is attached. So strip it off
1864 * since it is cheap to do so and saves bytes on the network.
1867 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
) &&
1868 tp
->snd_una
== (TCP_SKB_CB(skb
)->end_seq
- 1)) {
1869 if (!pskb_trim(skb
, 0)) {
1870 TCP_SKB_CB(skb
)->seq
= TCP_SKB_CB(skb
)->end_seq
- 1;
1871 skb_shinfo(skb
)->gso_segs
= 1;
1872 skb_shinfo(skb
)->gso_size
= 0;
1873 skb_shinfo(skb
)->gso_type
= 0;
1874 skb
->ip_summed
= CHECKSUM_NONE
;
1879 /* Make a copy, if the first transmission SKB clone we made
1880 * is still in somebody's hands, else make a clone.
1882 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1884 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
1887 /* Update global TCP statistics. */
1888 TCP_INC_STATS(TCP_MIB_RETRANSSEGS
);
1890 tp
->total_retrans
++;
1892 #if FASTRETRANS_DEBUG > 0
1893 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) {
1894 if (net_ratelimit())
1895 printk(KERN_DEBUG
"retrans_out leaked.\n");
1898 if (!tp
->retrans_out
)
1899 tp
->lost_retrans_low
= tp
->snd_nxt
;
1900 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
1901 tp
->retrans_out
+= tcp_skb_pcount(skb
);
1903 /* Save stamp of the first retransmit. */
1904 if (!tp
->retrans_stamp
)
1905 tp
->retrans_stamp
= TCP_SKB_CB(skb
)->when
;
1909 /* snd_nxt is stored to detect loss of retransmitted segment,
1910 * see tcp_input.c tcp_sacktag_write_queue().
1912 TCP_SKB_CB(skb
)->ack_seq
= tp
->snd_nxt
;
1917 /* This gets called after a retransmit timeout, and the initially
1918 * retransmitted data is acknowledged. It tries to continue
1919 * resending the rest of the retransmit queue, until either
1920 * we've sent it all or the congestion window limit is reached.
1921 * If doing SACK, the first ACK which comes back for a timeout
1922 * based retransmit packet might feed us FACK information again.
1923 * If so, we use it to avoid unnecessarily retransmissions.
1925 void tcp_xmit_retransmit_queue(struct sock
*sk
)
1927 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1928 struct tcp_sock
*tp
= tcp_sk(sk
);
1929 struct sk_buff
*skb
;
1932 if (tp
->retransmit_skb_hint
) {
1933 skb
= tp
->retransmit_skb_hint
;
1934 packet_cnt
= tp
->retransmit_cnt_hint
;
1936 skb
= tcp_write_queue_head(sk
);
1940 /* First pass: retransmit lost packets. */
1942 tcp_for_write_queue_from(skb
, sk
) {
1943 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
1945 if (skb
== tcp_send_head(sk
))
1947 /* we could do better than to assign each time */
1948 tp
->retransmit_skb_hint
= skb
;
1949 tp
->retransmit_cnt_hint
= packet_cnt
;
1951 /* Assume this retransmit will generate
1952 * only one packet for congestion window
1953 * calculation purposes. This works because
1954 * tcp_retransmit_skb() will chop up the
1955 * packet to be MSS sized and all the
1956 * packet counting works out.
1958 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
1961 if (sacked
& TCPCB_LOST
) {
1962 if (!(sacked
&(TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))) {
1963 if (tcp_retransmit_skb(sk
, skb
)) {
1964 tp
->retransmit_skb_hint
= NULL
;
1967 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
1968 NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS
);
1970 NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS
);
1972 if (skb
== tcp_write_queue_head(sk
))
1973 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1974 inet_csk(sk
)->icsk_rto
,
1978 packet_cnt
+= tcp_skb_pcount(skb
);
1979 if (packet_cnt
>= tp
->lost_out
)
1985 /* OK, demanded retransmission is finished. */
1987 /* Forward retransmissions are possible only during Recovery. */
1988 if (icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
1991 /* No forward retransmissions in Reno are possible. */
1992 if (tcp_is_reno(tp
))
1995 /* Yeah, we have to make difficult choice between forward transmission
1996 * and retransmission... Both ways have their merits...
1998 * For now we do not retransmit anything, while we have some new
1999 * segments to send. In the other cases, follow rule 3 for
2000 * NextSeg() specified in RFC3517.
2003 if (tcp_may_send_now(sk
))
2006 /* If nothing is SACKed, highest_sack in the loop won't be valid */
2007 if (!tp
->sacked_out
)
2010 if (tp
->forward_skb_hint
)
2011 skb
= tp
->forward_skb_hint
;
2013 skb
= tcp_write_queue_head(sk
);
2015 tcp_for_write_queue_from(skb
, sk
) {
2016 if (skb
== tcp_send_head(sk
))
2018 tp
->forward_skb_hint
= skb
;
2020 if (after(TCP_SKB_CB(skb
)->seq
, tcp_highest_sack_seq(tp
)))
2023 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
2026 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
)
2029 /* Ok, retransmit it. */
2030 if (tcp_retransmit_skb(sk
, skb
)) {
2031 tp
->forward_skb_hint
= NULL
;
2035 if (skb
== tcp_write_queue_head(sk
))
2036 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2037 inet_csk(sk
)->icsk_rto
,
2040 NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS
);
2045 /* Send a fin. The caller locks the socket for us. This cannot be
2046 * allowed to fail queueing a FIN frame under any circumstances.
2048 void tcp_send_fin(struct sock
*sk
)
2050 struct tcp_sock
*tp
= tcp_sk(sk
);
2051 struct sk_buff
*skb
= tcp_write_queue_tail(sk
);
2054 /* Optimization, tack on the FIN if we have a queue of
2055 * unsent frames. But be careful about outgoing SACKS
2058 mss_now
= tcp_current_mss(sk
, 1);
2060 if (tcp_send_head(sk
) != NULL
) {
2061 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_FIN
;
2062 TCP_SKB_CB(skb
)->end_seq
++;
2065 /* Socket is locked, keep trying until memory is available. */
2067 skb
= alloc_skb_fclone(MAX_TCP_HEADER
, GFP_KERNEL
);
2073 /* Reserve space for headers and prepare control bits. */
2074 skb_reserve(skb
, MAX_TCP_HEADER
);
2076 TCP_SKB_CB(skb
)->flags
= (TCPCB_FLAG_ACK
| TCPCB_FLAG_FIN
);
2077 TCP_SKB_CB(skb
)->sacked
= 0;
2078 skb_shinfo(skb
)->gso_segs
= 1;
2079 skb_shinfo(skb
)->gso_size
= 0;
2080 skb_shinfo(skb
)->gso_type
= 0;
2082 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2083 TCP_SKB_CB(skb
)->seq
= tp
->write_seq
;
2084 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ 1;
2085 tcp_queue_skb(sk
, skb
);
2087 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_OFF
);
2090 /* We get here when a process closes a file descriptor (either due to
2091 * an explicit close() or as a byproduct of exit()'ing) and there
2092 * was unread data in the receive queue. This behavior is recommended
2093 * by RFC 2525, section 2.17. -DaveM
2095 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
2097 struct sk_buff
*skb
;
2099 /* NOTE: No TCP options attached and we never retransmit this. */
2100 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
2102 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED
);
2106 /* Reserve space for headers and prepare control bits. */
2107 skb_reserve(skb
, MAX_TCP_HEADER
);
2109 TCP_SKB_CB(skb
)->flags
= (TCPCB_FLAG_ACK
| TCPCB_FLAG_RST
);
2110 TCP_SKB_CB(skb
)->sacked
= 0;
2111 skb_shinfo(skb
)->gso_segs
= 1;
2112 skb_shinfo(skb
)->gso_size
= 0;
2113 skb_shinfo(skb
)->gso_type
= 0;
2116 TCP_SKB_CB(skb
)->seq
= tcp_acceptable_seq(sk
);
2117 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
;
2118 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2119 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
2120 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED
);
2123 /* WARNING: This routine must only be called when we have already sent
2124 * a SYN packet that crossed the incoming SYN that caused this routine
2125 * to get called. If this assumption fails then the initial rcv_wnd
2126 * and rcv_wscale values will not be correct.
2128 int tcp_send_synack(struct sock
*sk
)
2130 struct sk_buff
* skb
;
2132 skb
= tcp_write_queue_head(sk
);
2133 if (skb
== NULL
|| !(TCP_SKB_CB(skb
)->flags
&TCPCB_FLAG_SYN
)) {
2134 printk(KERN_DEBUG
"tcp_send_synack: wrong queue state\n");
2137 if (!(TCP_SKB_CB(skb
)->flags
&TCPCB_FLAG_ACK
)) {
2138 if (skb_cloned(skb
)) {
2139 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
2142 tcp_unlink_write_queue(skb
, sk
);
2143 skb_header_release(nskb
);
2144 __tcp_add_write_queue_head(sk
, nskb
);
2145 sk_stream_free_skb(sk
, skb
);
2146 sk_charge_skb(sk
, nskb
);
2150 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_ACK
;
2151 TCP_ECN_send_synack(tcp_sk(sk
), skb
);
2153 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2154 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2158 * Prepare a SYN-ACK.
2160 struct sk_buff
* tcp_make_synack(struct sock
*sk
, struct dst_entry
*dst
,
2161 struct request_sock
*req
)
2163 struct inet_request_sock
*ireq
= inet_rsk(req
);
2164 struct tcp_sock
*tp
= tcp_sk(sk
);
2166 int tcp_header_size
;
2167 struct sk_buff
*skb
;
2168 #ifdef CONFIG_TCP_MD5SIG
2169 struct tcp_md5sig_key
*md5
;
2170 __u8
*md5_hash_location
;
2173 skb
= sock_wmalloc(sk
, MAX_TCP_HEADER
+ 15, 1, GFP_ATOMIC
);
2177 /* Reserve space for headers. */
2178 skb_reserve(skb
, MAX_TCP_HEADER
);
2180 skb
->dst
= dst_clone(dst
);
2182 tcp_header_size
= (sizeof(struct tcphdr
) + TCPOLEN_MSS
+
2183 (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0) +
2184 (ireq
->wscale_ok
? TCPOLEN_WSCALE_ALIGNED
: 0) +
2185 /* SACK_PERM is in the place of NOP NOP of TS */
2186 ((ireq
->sack_ok
&& !ireq
->tstamp_ok
) ? TCPOLEN_SACKPERM_ALIGNED
: 0));
2188 #ifdef CONFIG_TCP_MD5SIG
2189 /* Are we doing MD5 on this segment? If so - make room for it */
2190 md5
= tcp_rsk(req
)->af_specific
->md5_lookup(sk
, req
);
2192 tcp_header_size
+= TCPOLEN_MD5SIG_ALIGNED
;
2194 skb_push(skb
, tcp_header_size
);
2195 skb_reset_transport_header(skb
);
2198 memset(th
, 0, sizeof(struct tcphdr
));
2201 TCP_ECN_make_synack(req
, th
);
2202 th
->source
= inet_sk(sk
)->sport
;
2203 th
->dest
= ireq
->rmt_port
;
2204 TCP_SKB_CB(skb
)->seq
= tcp_rsk(req
)->snt_isn
;
2205 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ 1;
2206 TCP_SKB_CB(skb
)->sacked
= 0;
2207 skb_shinfo(skb
)->gso_segs
= 1;
2208 skb_shinfo(skb
)->gso_size
= 0;
2209 skb_shinfo(skb
)->gso_type
= 0;
2210 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
2211 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_isn
+ 1);
2212 if (req
->rcv_wnd
== 0) { /* ignored for retransmitted syns */
2214 /* Set this up on the first call only */
2215 req
->window_clamp
= tp
->window_clamp
? : dst_metric(dst
, RTAX_WINDOW
);
2216 /* tcp_full_space because it is guaranteed to be the first packet */
2217 tcp_select_initial_window(tcp_full_space(sk
),
2218 dst_metric(dst
, RTAX_ADVMSS
) - (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0),
2223 ireq
->rcv_wscale
= rcv_wscale
;
2226 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2227 th
->window
= htons(min(req
->rcv_wnd
, 65535U));
2229 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2230 tcp_syn_build_options((__be32
*)(th
+ 1), dst_metric(dst
, RTAX_ADVMSS
), ireq
->tstamp_ok
,
2231 ireq
->sack_ok
, ireq
->wscale_ok
, ireq
->rcv_wscale
,
2232 TCP_SKB_CB(skb
)->when
,
2235 #ifdef CONFIG_TCP_MD5SIG
2236 md5
? &md5_hash_location
:
2242 th
->doff
= (tcp_header_size
>> 2);
2243 TCP_INC_STATS(TCP_MIB_OUTSEGS
);
2245 #ifdef CONFIG_TCP_MD5SIG
2246 /* Okay, we have all we need - do the md5 hash if needed */
2248 tp
->af_specific
->calc_md5_hash(md5_hash_location
,
2251 tcp_hdr(skb
), sk
->sk_protocol
,
2260 * Do all connect socket setups that can be done AF independent.
2262 static void tcp_connect_init(struct sock
*sk
)
2264 struct dst_entry
*dst
= __sk_dst_get(sk
);
2265 struct tcp_sock
*tp
= tcp_sk(sk
);
2268 /* We'll fix this up when we get a response from the other end.
2269 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2271 tp
->tcp_header_len
= sizeof(struct tcphdr
) +
2272 (sysctl_tcp_timestamps
? TCPOLEN_TSTAMP_ALIGNED
: 0);
2274 #ifdef CONFIG_TCP_MD5SIG
2275 if (tp
->af_specific
->md5_lookup(sk
, sk
) != NULL
)
2276 tp
->tcp_header_len
+= TCPOLEN_MD5SIG_ALIGNED
;
2279 /* If user gave his TCP_MAXSEG, record it to clamp */
2280 if (tp
->rx_opt
.user_mss
)
2281 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
2284 tcp_sync_mss(sk
, dst_mtu(dst
));
2286 if (!tp
->window_clamp
)
2287 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
2288 tp
->advmss
= dst_metric(dst
, RTAX_ADVMSS
);
2289 tcp_initialize_rcv_mss(sk
);
2291 tcp_select_initial_window(tcp_full_space(sk
),
2292 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
2295 sysctl_tcp_window_scaling
,
2298 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
2299 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
2302 sock_reset_flag(sk
, SOCK_DONE
);
2304 tcp_init_wl(tp
, tp
->write_seq
, 0);
2305 tp
->snd_una
= tp
->write_seq
;
2306 tp
->snd_sml
= tp
->write_seq
;
2311 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
2312 inet_csk(sk
)->icsk_retransmits
= 0;
2313 tcp_clear_retrans(tp
);
2317 * Build a SYN and send it off.
2319 int tcp_connect(struct sock
*sk
)
2321 struct tcp_sock
*tp
= tcp_sk(sk
);
2322 struct sk_buff
*buff
;
2324 tcp_connect_init(sk
);
2326 buff
= alloc_skb_fclone(MAX_TCP_HEADER
+ 15, sk
->sk_allocation
);
2327 if (unlikely(buff
== NULL
))
2330 /* Reserve space for headers. */
2331 skb_reserve(buff
, MAX_TCP_HEADER
);
2333 TCP_SKB_CB(buff
)->flags
= TCPCB_FLAG_SYN
;
2334 TCP_ECN_send_syn(sk
, buff
);
2335 TCP_SKB_CB(buff
)->sacked
= 0;
2336 skb_shinfo(buff
)->gso_segs
= 1;
2337 skb_shinfo(buff
)->gso_size
= 0;
2338 skb_shinfo(buff
)->gso_type
= 0;
2340 tp
->snd_nxt
= tp
->write_seq
;
2341 TCP_SKB_CB(buff
)->seq
= tp
->write_seq
++;
2342 TCP_SKB_CB(buff
)->end_seq
= tp
->write_seq
;
2345 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2346 tp
->retrans_stamp
= TCP_SKB_CB(buff
)->when
;
2347 skb_header_release(buff
);
2348 __tcp_add_write_queue_tail(sk
, buff
);
2349 sk_charge_skb(sk
, buff
);
2350 tp
->packets_out
+= tcp_skb_pcount(buff
);
2351 tcp_transmit_skb(sk
, buff
, 1, GFP_KERNEL
);
2353 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2354 * in order to make this packet get counted in tcpOutSegs.
2356 tp
->snd_nxt
= tp
->write_seq
;
2357 tp
->pushed_seq
= tp
->write_seq
;
2358 TCP_INC_STATS(TCP_MIB_ACTIVEOPENS
);
2360 /* Timer for repeating the SYN until an answer. */
2361 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2362 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2366 /* Send out a delayed ack, the caller does the policy checking
2367 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2370 void tcp_send_delayed_ack(struct sock
*sk
)
2372 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2373 int ato
= icsk
->icsk_ack
.ato
;
2374 unsigned long timeout
;
2376 if (ato
> TCP_DELACK_MIN
) {
2377 const struct tcp_sock
*tp
= tcp_sk(sk
);
2380 if (icsk
->icsk_ack
.pingpong
|| (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
2381 max_ato
= TCP_DELACK_MAX
;
2383 /* Slow path, intersegment interval is "high". */
2385 /* If some rtt estimate is known, use it to bound delayed ack.
2386 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2390 int rtt
= max(tp
->srtt
>>3, TCP_DELACK_MIN
);
2396 ato
= min(ato
, max_ato
);
2399 /* Stay within the limit we were given */
2400 timeout
= jiffies
+ ato
;
2402 /* Use new timeout only if there wasn't a older one earlier. */
2403 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
2404 /* If delack timer was blocked or is about to expire,
2407 if (icsk
->icsk_ack
.blocked
||
2408 time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
2413 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
2414 timeout
= icsk
->icsk_ack
.timeout
;
2416 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
2417 icsk
->icsk_ack
.timeout
= timeout
;
2418 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
2421 /* This routine sends an ack and also updates the window. */
2422 void tcp_send_ack(struct sock
*sk
)
2424 /* If we have been reset, we may not send again. */
2425 if (sk
->sk_state
!= TCP_CLOSE
) {
2426 struct sk_buff
*buff
;
2428 /* We are not putting this on the write queue, so
2429 * tcp_transmit_skb() will set the ownership to this
2432 buff
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2434 inet_csk_schedule_ack(sk
);
2435 inet_csk(sk
)->icsk_ack
.ato
= TCP_ATO_MIN
;
2436 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
2437 TCP_DELACK_MAX
, TCP_RTO_MAX
);
2441 /* Reserve space for headers and prepare control bits. */
2442 skb_reserve(buff
, MAX_TCP_HEADER
);
2444 TCP_SKB_CB(buff
)->flags
= TCPCB_FLAG_ACK
;
2445 TCP_SKB_CB(buff
)->sacked
= 0;
2446 skb_shinfo(buff
)->gso_segs
= 1;
2447 skb_shinfo(buff
)->gso_size
= 0;
2448 skb_shinfo(buff
)->gso_type
= 0;
2450 /* Send it off, this clears delayed acks for us. */
2451 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(buff
)->end_seq
= tcp_acceptable_seq(sk
);
2452 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2453 tcp_transmit_skb(sk
, buff
, 0, GFP_ATOMIC
);
2457 /* This routine sends a packet with an out of date sequence
2458 * number. It assumes the other end will try to ack it.
2460 * Question: what should we make while urgent mode?
2461 * 4.4BSD forces sending single byte of data. We cannot send
2462 * out of window data, because we have SND.NXT==SND.MAX...
2464 * Current solution: to send TWO zero-length segments in urgent mode:
2465 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2466 * out-of-date with SND.UNA-1 to probe window.
2468 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
)
2470 struct tcp_sock
*tp
= tcp_sk(sk
);
2471 struct sk_buff
*skb
;
2473 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2474 skb
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2478 /* Reserve space for headers and set control bits. */
2479 skb_reserve(skb
, MAX_TCP_HEADER
);
2481 TCP_SKB_CB(skb
)->flags
= TCPCB_FLAG_ACK
;
2482 TCP_SKB_CB(skb
)->sacked
= urgent
;
2483 skb_shinfo(skb
)->gso_segs
= 1;
2484 skb_shinfo(skb
)->gso_size
= 0;
2485 skb_shinfo(skb
)->gso_type
= 0;
2487 /* Use a previous sequence. This should cause the other
2488 * end to send an ack. Don't queue or clone SKB, just
2491 TCP_SKB_CB(skb
)->seq
= urgent
? tp
->snd_una
: tp
->snd_una
- 1;
2492 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
;
2493 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2494 return tcp_transmit_skb(sk
, skb
, 0, GFP_ATOMIC
);
2497 int tcp_write_wakeup(struct sock
*sk
)
2499 if (sk
->sk_state
!= TCP_CLOSE
) {
2500 struct tcp_sock
*tp
= tcp_sk(sk
);
2501 struct sk_buff
*skb
;
2503 if ((skb
= tcp_send_head(sk
)) != NULL
&&
2504 before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
+tp
->snd_wnd
)) {
2506 unsigned int mss
= tcp_current_mss(sk
, 0);
2507 unsigned int seg_size
= tp
->snd_una
+tp
->snd_wnd
-TCP_SKB_CB(skb
)->seq
;
2509 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
2510 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
2512 /* We are probing the opening of a window
2513 * but the window size is != 0
2514 * must have been a result SWS avoidance ( sender )
2516 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
2518 seg_size
= min(seg_size
, mss
);
2519 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2520 if (tcp_fragment(sk
, skb
, seg_size
, mss
))
2522 } else if (!tcp_skb_pcount(skb
))
2523 tcp_set_skb_tso_segs(sk
, skb
, mss
);
2525 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2526 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2527 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2529 update_send_head(sk
, skb
);
2534 between(tp
->snd_up
, tp
->snd_una
+1, tp
->snd_una
+0xFFFF))
2535 tcp_xmit_probe_skb(sk
, TCPCB_URG
);
2536 return tcp_xmit_probe_skb(sk
, 0);
2542 /* A window probe timeout has occurred. If window is not closed send
2543 * a partial packet else a zero probe.
2545 void tcp_send_probe0(struct sock
*sk
)
2547 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2548 struct tcp_sock
*tp
= tcp_sk(sk
);
2551 err
= tcp_write_wakeup(sk
);
2553 if (tp
->packets_out
|| !tcp_send_head(sk
)) {
2554 /* Cancel probe timer, if it is not required. */
2555 icsk
->icsk_probes_out
= 0;
2556 icsk
->icsk_backoff
= 0;
2561 if (icsk
->icsk_backoff
< sysctl_tcp_retries2
)
2562 icsk
->icsk_backoff
++;
2563 icsk
->icsk_probes_out
++;
2564 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2565 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
2568 /* If packet was not sent due to local congestion,
2569 * do not backoff and do not remember icsk_probes_out.
2570 * Let local senders to fight for local resources.
2572 * Use accumulated backoff yet.
2574 if (!icsk
->icsk_probes_out
)
2575 icsk
->icsk_probes_out
= 1;
2576 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2577 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
,
2578 TCP_RESOURCE_PROBE_INTERVAL
),
2583 EXPORT_SYMBOL(tcp_connect
);
2584 EXPORT_SYMBOL(tcp_make_synack
);
2585 EXPORT_SYMBOL(tcp_simple_retransmit
);
2586 EXPORT_SYMBOL(tcp_sync_mss
);
2587 EXPORT_SYMBOL(sysctl_tcp_tso_win_divisor
);
2588 EXPORT_SYMBOL(tcp_mtup_init
);