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;
651 factor
= skb
->len
+ (mss_now
- 1);
653 skb_shinfo(skb
)->gso_segs
= factor
;
654 skb_shinfo(skb
)->gso_size
= mss_now
;
655 skb_shinfo(skb
)->gso_type
= sk
->sk_gso_type
;
659 /* Function to create two new TCP segments. Shrinks the given segment
660 * to the specified size and appends a new segment with the rest of the
661 * packet to the list. This won't be called frequently, I hope.
662 * Remember, these are still headerless SKBs at this point.
664 int tcp_fragment(struct sock
*sk
, struct sk_buff
*skb
, u32 len
, unsigned int mss_now
)
666 struct tcp_sock
*tp
= tcp_sk(sk
);
667 struct sk_buff
*buff
;
668 int nsize
, old_factor
;
672 BUG_ON(len
> skb
->len
);
674 clear_all_retrans_hints(tp
);
675 nsize
= skb_headlen(skb
) - len
;
679 if (skb_cloned(skb
) &&
680 skb_is_nonlinear(skb
) &&
681 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
684 /* Get a new skb... force flag on. */
685 buff
= sk_stream_alloc_skb(sk
, nsize
, GFP_ATOMIC
);
687 return -ENOMEM
; /* We'll just try again later. */
689 sk_charge_skb(sk
, buff
);
690 nlen
= skb
->len
- len
- nsize
;
691 buff
->truesize
+= nlen
;
692 skb
->truesize
-= nlen
;
694 /* Correct the sequence numbers. */
695 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
696 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
697 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
699 /* PSH and FIN should only be set in the second packet. */
700 flags
= TCP_SKB_CB(skb
)->flags
;
701 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
702 TCP_SKB_CB(buff
)->flags
= flags
;
703 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
704 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_AT_TAIL
;
706 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_PARTIAL
) {
707 /* Copy and checksum data tail into the new buffer. */
708 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
, skb_put(buff
, nsize
),
713 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
715 skb
->ip_summed
= CHECKSUM_PARTIAL
;
716 skb_split(skb
, buff
, len
);
719 buff
->ip_summed
= skb
->ip_summed
;
721 /* Looks stupid, but our code really uses when of
722 * skbs, which it never sent before. --ANK
724 TCP_SKB_CB(buff
)->when
= TCP_SKB_CB(skb
)->when
;
725 buff
->tstamp
= skb
->tstamp
;
727 old_factor
= tcp_skb_pcount(skb
);
729 /* Fix up tso_factor for both original and new SKB. */
730 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
731 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
733 /* If this packet has been sent out already, we must
734 * adjust the various packet counters.
736 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
737 int diff
= old_factor
- tcp_skb_pcount(skb
) -
738 tcp_skb_pcount(buff
);
740 tp
->packets_out
-= diff
;
742 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
743 tp
->sacked_out
-= diff
;
744 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
745 tp
->retrans_out
-= diff
;
747 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
)
748 tp
->lost_out
-= diff
;
751 /* Adjust Reno SACK estimate. */
752 if (tcp_is_reno(tp
)) {
753 tcp_dec_pcount_approx_int(&tp
->sacked_out
, diff
);
754 tcp_verify_left_out(tp
);
757 tcp_dec_pcount_approx_int(&tp
->fackets_out
, diff
);
758 /* SACK fastpath might overwrite it unless dealt with */
759 if (tp
->fastpath_skb_hint
!= NULL
&&
760 after(TCP_SKB_CB(tp
->fastpath_skb_hint
)->seq
,
761 TCP_SKB_CB(skb
)->seq
)) {
762 tcp_dec_pcount_approx_int(&tp
->fastpath_cnt_hint
, diff
);
767 /* Link BUFF into the send queue. */
768 skb_header_release(buff
);
769 tcp_insert_write_queue_after(skb
, buff
, sk
);
774 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
775 * eventually). The difference is that pulled data not copied, but
776 * immediately discarded.
778 static void __pskb_trim_head(struct sk_buff
*skb
, int len
)
784 for (i
=0; i
<skb_shinfo(skb
)->nr_frags
; i
++) {
785 if (skb_shinfo(skb
)->frags
[i
].size
<= eat
) {
786 put_page(skb_shinfo(skb
)->frags
[i
].page
);
787 eat
-= skb_shinfo(skb
)->frags
[i
].size
;
789 skb_shinfo(skb
)->frags
[k
] = skb_shinfo(skb
)->frags
[i
];
791 skb_shinfo(skb
)->frags
[k
].page_offset
+= eat
;
792 skb_shinfo(skb
)->frags
[k
].size
-= eat
;
798 skb_shinfo(skb
)->nr_frags
= k
;
800 skb_reset_tail_pointer(skb
);
801 skb
->data_len
-= len
;
802 skb
->len
= skb
->data_len
;
805 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
807 if (skb_cloned(skb
) &&
808 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
811 /* If len == headlen, we avoid __skb_pull to preserve alignment. */
812 if (unlikely(len
< skb_headlen(skb
)))
813 __skb_pull(skb
, len
);
815 __pskb_trim_head(skb
, len
- skb_headlen(skb
));
817 TCP_SKB_CB(skb
)->seq
+= len
;
818 skb
->ip_summed
= CHECKSUM_PARTIAL
;
820 skb
->truesize
-= len
;
821 sk
->sk_wmem_queued
-= len
;
822 sk
->sk_forward_alloc
+= len
;
823 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
825 /* Any change of skb->len requires recalculation of tso
828 if (tcp_skb_pcount(skb
) > 1)
829 tcp_set_skb_tso_segs(sk
, skb
, tcp_current_mss(sk
, 1));
834 /* Not accounting for SACKs here. */
835 int tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
837 struct tcp_sock
*tp
= tcp_sk(sk
);
838 struct inet_connection_sock
*icsk
= inet_csk(sk
);
841 /* Calculate base mss without TCP options:
842 It is MMS_S - sizeof(tcphdr) of rfc1122
844 mss_now
= pmtu
- icsk
->icsk_af_ops
->net_header_len
- sizeof(struct tcphdr
);
846 /* Clamp it (mss_clamp does not include tcp options) */
847 if (mss_now
> tp
->rx_opt
.mss_clamp
)
848 mss_now
= tp
->rx_opt
.mss_clamp
;
850 /* Now subtract optional transport overhead */
851 mss_now
-= icsk
->icsk_ext_hdr_len
;
853 /* Then reserve room for full set of TCP options and 8 bytes of data */
857 /* Now subtract TCP options size, not including SACKs */
858 mss_now
-= tp
->tcp_header_len
- sizeof(struct tcphdr
);
863 /* Inverse of above */
864 int tcp_mss_to_mtu(struct sock
*sk
, int mss
)
866 struct tcp_sock
*tp
= tcp_sk(sk
);
867 struct inet_connection_sock
*icsk
= inet_csk(sk
);
872 icsk
->icsk_ext_hdr_len
+
873 icsk
->icsk_af_ops
->net_header_len
;
878 void tcp_mtup_init(struct sock
*sk
)
880 struct tcp_sock
*tp
= tcp_sk(sk
);
881 struct inet_connection_sock
*icsk
= inet_csk(sk
);
883 icsk
->icsk_mtup
.enabled
= sysctl_tcp_mtu_probing
> 1;
884 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+ sizeof(struct tcphdr
) +
885 icsk
->icsk_af_ops
->net_header_len
;
886 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, sysctl_tcp_base_mss
);
887 icsk
->icsk_mtup
.probe_size
= 0;
890 /* This function synchronize snd mss to current pmtu/exthdr set.
892 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
893 for TCP options, but includes only bare TCP header.
895 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
896 It is minimum of user_mss and mss received with SYN.
897 It also does not include TCP options.
899 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
901 tp->mss_cache is current effective sending mss, including
902 all tcp options except for SACKs. It is evaluated,
903 taking into account current pmtu, but never exceeds
904 tp->rx_opt.mss_clamp.
906 NOTE1. rfc1122 clearly states that advertised MSS
907 DOES NOT include either tcp or ip options.
909 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
910 are READ ONLY outside this function. --ANK (980731)
913 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
915 struct tcp_sock
*tp
= tcp_sk(sk
);
916 struct inet_connection_sock
*icsk
= inet_csk(sk
);
919 if (icsk
->icsk_mtup
.search_high
> pmtu
)
920 icsk
->icsk_mtup
.search_high
= pmtu
;
922 mss_now
= tcp_mtu_to_mss(sk
, pmtu
);
924 /* Bound mss with half of window */
925 if (tp
->max_window
&& mss_now
> (tp
->max_window
>>1))
926 mss_now
= max((tp
->max_window
>>1), 68U - tp
->tcp_header_len
);
928 /* And store cached results */
929 icsk
->icsk_pmtu_cookie
= pmtu
;
930 if (icsk
->icsk_mtup
.enabled
)
931 mss_now
= min(mss_now
, tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
));
932 tp
->mss_cache
= mss_now
;
937 /* Compute the current effective MSS, taking SACKs and IP options,
938 * and even PMTU discovery events into account.
940 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
941 * cannot be large. However, taking into account rare use of URG, this
944 unsigned int tcp_current_mss(struct sock
*sk
, int large_allowed
)
946 struct tcp_sock
*tp
= tcp_sk(sk
);
947 struct dst_entry
*dst
= __sk_dst_get(sk
);
952 mss_now
= tp
->mss_cache
;
954 if (large_allowed
&& sk_can_gso(sk
) && !tp
->urg_mode
)
958 u32 mtu
= dst_mtu(dst
);
959 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
960 mss_now
= tcp_sync_mss(sk
, mtu
);
963 if (tp
->rx_opt
.eff_sacks
)
964 mss_now
-= (TCPOLEN_SACK_BASE_ALIGNED
+
965 (tp
->rx_opt
.eff_sacks
* TCPOLEN_SACK_PERBLOCK
));
967 #ifdef CONFIG_TCP_MD5SIG
968 if (tp
->af_specific
->md5_lookup(sk
, sk
))
969 mss_now
-= TCPOLEN_MD5SIG_ALIGNED
;
972 xmit_size_goal
= mss_now
;
975 xmit_size_goal
= (65535 -
976 inet_csk(sk
)->icsk_af_ops
->net_header_len
-
977 inet_csk(sk
)->icsk_ext_hdr_len
-
980 if (tp
->max_window
&&
981 (xmit_size_goal
> (tp
->max_window
>> 1)))
982 xmit_size_goal
= max((tp
->max_window
>> 1),
983 68U - tp
->tcp_header_len
);
985 xmit_size_goal
-= (xmit_size_goal
% mss_now
);
987 tp
->xmit_size_goal
= xmit_size_goal
;
992 /* Congestion window validation. (RFC2861) */
994 static void tcp_cwnd_validate(struct sock
*sk
)
996 struct tcp_sock
*tp
= tcp_sk(sk
);
997 __u32 packets_out
= tp
->packets_out
;
999 if (packets_out
>= tp
->snd_cwnd
) {
1000 /* Network is feed fully. */
1001 tp
->snd_cwnd_used
= 0;
1002 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1004 /* Network starves. */
1005 if (tp
->packets_out
> tp
->snd_cwnd_used
)
1006 tp
->snd_cwnd_used
= tp
->packets_out
;
1008 if (sysctl_tcp_slow_start_after_idle
&&
1009 (s32
)(tcp_time_stamp
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
)
1010 tcp_cwnd_application_limited(sk
);
1014 static unsigned int tcp_window_allows(struct tcp_sock
*tp
, struct sk_buff
*skb
, unsigned int mss_now
, unsigned int cwnd
)
1016 u32 window
, cwnd_len
;
1018 window
= (tp
->snd_una
+ tp
->snd_wnd
- TCP_SKB_CB(skb
)->seq
);
1019 cwnd_len
= mss_now
* cwnd
;
1020 return min(window
, cwnd_len
);
1023 /* Can at least one segment of SKB be sent right now, according to the
1024 * congestion window rules? If so, return how many segments are allowed.
1026 static inline unsigned int tcp_cwnd_test(struct tcp_sock
*tp
, struct sk_buff
*skb
)
1028 u32 in_flight
, cwnd
;
1030 /* Don't be strict about the congestion window for the final FIN. */
1031 if ((TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
) &&
1032 tcp_skb_pcount(skb
) == 1)
1035 in_flight
= tcp_packets_in_flight(tp
);
1036 cwnd
= tp
->snd_cwnd
;
1037 if (in_flight
< cwnd
)
1038 return (cwnd
- in_flight
);
1043 /* This must be invoked the first time we consider transmitting
1044 * SKB onto the wire.
1046 static int tcp_init_tso_segs(struct sock
*sk
, struct sk_buff
*skb
, unsigned int mss_now
)
1048 int tso_segs
= tcp_skb_pcount(skb
);
1052 tcp_skb_mss(skb
) != mss_now
)) {
1053 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1054 tso_segs
= tcp_skb_pcount(skb
);
1059 static inline int tcp_minshall_check(const struct tcp_sock
*tp
)
1061 return after(tp
->snd_sml
,tp
->snd_una
) &&
1062 !after(tp
->snd_sml
, tp
->snd_nxt
);
1065 /* Return 0, if packet can be sent now without violation Nagle's rules:
1066 * 1. It is full sized.
1067 * 2. Or it contains FIN. (already checked by caller)
1068 * 3. Or TCP_NODELAY was set.
1069 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1070 * With Minshall's modification: all sent small packets are ACKed.
1073 static inline int tcp_nagle_check(const struct tcp_sock
*tp
,
1074 const struct sk_buff
*skb
,
1075 unsigned mss_now
, int nonagle
)
1077 return (skb
->len
< mss_now
&&
1078 ((nonagle
&TCP_NAGLE_CORK
) ||
1081 tcp_minshall_check(tp
))));
1084 /* Return non-zero if the Nagle test allows this packet to be
1087 static inline int tcp_nagle_test(struct tcp_sock
*tp
, struct sk_buff
*skb
,
1088 unsigned int cur_mss
, int nonagle
)
1090 /* Nagle rule does not apply to frames, which sit in the middle of the
1091 * write_queue (they have no chances to get new data).
1093 * This is implemented in the callers, where they modify the 'nonagle'
1094 * argument based upon the location of SKB in the send queue.
1096 if (nonagle
& TCP_NAGLE_PUSH
)
1099 /* Don't use the nagle rule for urgent data (or for the final FIN).
1100 * Nagle can be ignored during F-RTO too (see RFC4138).
1102 if (tp
->urg_mode
|| (tp
->frto_counter
== 2) ||
1103 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
))
1106 if (!tcp_nagle_check(tp
, skb
, cur_mss
, nonagle
))
1112 /* Does at least the first segment of SKB fit into the send window? */
1113 static inline int tcp_snd_wnd_test(struct tcp_sock
*tp
, struct sk_buff
*skb
, unsigned int cur_mss
)
1115 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1117 if (skb
->len
> cur_mss
)
1118 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
1120 return !after(end_seq
, tp
->snd_una
+ tp
->snd_wnd
);
1123 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1124 * should be put on the wire right now. If so, it returns the number of
1125 * packets allowed by the congestion window.
1127 static unsigned int tcp_snd_test(struct sock
*sk
, struct sk_buff
*skb
,
1128 unsigned int cur_mss
, int nonagle
)
1130 struct tcp_sock
*tp
= tcp_sk(sk
);
1131 unsigned int cwnd_quota
;
1133 tcp_init_tso_segs(sk
, skb
, cur_mss
);
1135 if (!tcp_nagle_test(tp
, skb
, cur_mss
, nonagle
))
1138 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1140 !tcp_snd_wnd_test(tp
, skb
, cur_mss
))
1146 int tcp_may_send_now(struct sock
*sk
)
1148 struct tcp_sock
*tp
= tcp_sk(sk
);
1149 struct sk_buff
*skb
= tcp_send_head(sk
);
1152 tcp_snd_test(sk
, skb
, tcp_current_mss(sk
, 1),
1153 (tcp_skb_is_last(sk
, skb
) ?
1158 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1159 * which is put after SKB on the list. It is very much like
1160 * tcp_fragment() except that it may make several kinds of assumptions
1161 * in order to speed up the splitting operation. In particular, we
1162 * know that all the data is in scatter-gather pages, and that the
1163 * packet has never been sent out before (and thus is not cloned).
1165 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
, unsigned int mss_now
)
1167 struct sk_buff
*buff
;
1168 int nlen
= skb
->len
- len
;
1171 /* All of a TSO frame must be composed of paged data. */
1172 if (skb
->len
!= skb
->data_len
)
1173 return tcp_fragment(sk
, skb
, len
, mss_now
);
1175 buff
= sk_stream_alloc_pskb(sk
, 0, 0, GFP_ATOMIC
);
1176 if (unlikely(buff
== NULL
))
1179 sk_charge_skb(sk
, buff
);
1180 buff
->truesize
+= nlen
;
1181 skb
->truesize
-= nlen
;
1183 /* Correct the sequence numbers. */
1184 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1185 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1186 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1188 /* PSH and FIN should only be set in the second packet. */
1189 flags
= TCP_SKB_CB(skb
)->flags
;
1190 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
1191 TCP_SKB_CB(buff
)->flags
= flags
;
1193 /* This packet was never sent out yet, so no SACK bits. */
1194 TCP_SKB_CB(buff
)->sacked
= 0;
1196 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_PARTIAL
;
1197 skb_split(skb
, buff
, len
);
1199 /* Fix up tso_factor for both original and new SKB. */
1200 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1201 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1203 /* Link BUFF into the send queue. */
1204 skb_header_release(buff
);
1205 tcp_insert_write_queue_after(skb
, buff
, sk
);
1210 /* Try to defer sending, if possible, in order to minimize the amount
1211 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1213 * This algorithm is from John Heffner.
1215 static int tcp_tso_should_defer(struct sock
*sk
, struct sk_buff
*skb
)
1217 struct tcp_sock
*tp
= tcp_sk(sk
);
1218 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1219 u32 send_win
, cong_win
, limit
, in_flight
;
1221 if (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
)
1224 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
1227 /* Defer for less than two clock ticks. */
1228 if (!tp
->tso_deferred
&& ((jiffies
<<1)>>1) - (tp
->tso_deferred
>>1) > 1)
1231 in_flight
= tcp_packets_in_flight(tp
);
1233 BUG_ON(tcp_skb_pcount(skb
) <= 1 ||
1234 (tp
->snd_cwnd
<= in_flight
));
1236 send_win
= (tp
->snd_una
+ tp
->snd_wnd
) - TCP_SKB_CB(skb
)->seq
;
1238 /* From in_flight test above, we know that cwnd > in_flight. */
1239 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
1241 limit
= min(send_win
, cong_win
);
1243 /* If a full-sized TSO skb can be sent, do it. */
1247 if (sysctl_tcp_tso_win_divisor
) {
1248 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
1250 /* If at least some fraction of a window is available,
1253 chunk
/= sysctl_tcp_tso_win_divisor
;
1257 /* Different approach, try not to defer past a single
1258 * ACK. Receiver should ACK every other full sized
1259 * frame, so if we have space for more than 3 frames
1262 if (limit
> tcp_max_burst(tp
) * tp
->mss_cache
)
1266 /* Ok, it looks like it is advisable to defer. */
1267 tp
->tso_deferred
= 1 | (jiffies
<<1);
1272 tp
->tso_deferred
= 0;
1276 /* Create a new MTU probe if we are ready.
1277 * Returns 0 if we should wait to probe (no cwnd available),
1278 * 1 if a probe was sent,
1280 static int tcp_mtu_probe(struct sock
*sk
)
1282 struct tcp_sock
*tp
= tcp_sk(sk
);
1283 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1284 struct sk_buff
*skb
, *nskb
, *next
;
1291 /* Not currently probing/verifying,
1293 * have enough cwnd, and
1294 * not SACKing (the variable headers throw things off) */
1295 if (!icsk
->icsk_mtup
.enabled
||
1296 icsk
->icsk_mtup
.probe_size
||
1297 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
||
1298 tp
->snd_cwnd
< 11 ||
1299 tp
->rx_opt
.eff_sacks
)
1302 /* Very simple search strategy: just double the MSS. */
1303 mss_now
= tcp_current_mss(sk
, 0);
1304 probe_size
= 2*tp
->mss_cache
;
1305 if (probe_size
> tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_high
)) {
1306 /* TODO: set timer for probe_converge_event */
1310 /* Have enough data in the send queue to probe? */
1312 if ((skb
= tcp_send_head(sk
)) == NULL
)
1314 while ((len
+= skb
->len
) < probe_size
&& !tcp_skb_is_last(sk
, skb
))
1315 skb
= tcp_write_queue_next(sk
, skb
);
1316 if (len
< probe_size
)
1319 /* Receive window check. */
1320 if (after(TCP_SKB_CB(skb
)->seq
+ probe_size
, tp
->snd_una
+ tp
->snd_wnd
)) {
1321 if (tp
->snd_wnd
< probe_size
)
1327 /* Do we need to wait to drain cwnd? */
1328 pif
= tcp_packets_in_flight(tp
);
1329 if (pif
+ 2 > tp
->snd_cwnd
) {
1330 /* With no packets in flight, don't stall. */
1337 /* We're allowed to probe. Build it now. */
1338 if ((nskb
= sk_stream_alloc_skb(sk
, probe_size
, GFP_ATOMIC
)) == NULL
)
1340 sk_charge_skb(sk
, nskb
);
1342 skb
= tcp_send_head(sk
);
1343 tcp_insert_write_queue_before(nskb
, skb
, sk
);
1344 tcp_advance_send_head(sk
, skb
);
1346 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(skb
)->seq
;
1347 TCP_SKB_CB(nskb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ probe_size
;
1348 TCP_SKB_CB(nskb
)->flags
= TCPCB_FLAG_ACK
;
1349 TCP_SKB_CB(nskb
)->sacked
= 0;
1351 nskb
->ip_summed
= skb
->ip_summed
;
1354 while (len
< probe_size
) {
1355 next
= tcp_write_queue_next(sk
, skb
);
1357 copy
= min_t(int, skb
->len
, probe_size
- len
);
1358 if (nskb
->ip_summed
)
1359 skb_copy_bits(skb
, 0, skb_put(nskb
, copy
), copy
);
1361 nskb
->csum
= skb_copy_and_csum_bits(skb
, 0,
1362 skb_put(nskb
, copy
), copy
, nskb
->csum
);
1364 if (skb
->len
<= copy
) {
1365 /* We've eaten all the data from this skb.
1367 TCP_SKB_CB(nskb
)->flags
|= TCP_SKB_CB(skb
)->flags
;
1368 tcp_unlink_write_queue(skb
, sk
);
1369 sk_stream_free_skb(sk
, skb
);
1371 TCP_SKB_CB(nskb
)->flags
|= TCP_SKB_CB(skb
)->flags
&
1372 ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
1373 if (!skb_shinfo(skb
)->nr_frags
) {
1374 skb_pull(skb
, copy
);
1375 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1376 skb
->csum
= csum_partial(skb
->data
, skb
->len
, 0);
1378 __pskb_trim_head(skb
, copy
);
1379 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1381 TCP_SKB_CB(skb
)->seq
+= copy
;
1387 tcp_init_tso_segs(sk
, nskb
, nskb
->len
);
1389 /* We're ready to send. If this fails, the probe will
1390 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1391 TCP_SKB_CB(nskb
)->when
= tcp_time_stamp
;
1392 if (!tcp_transmit_skb(sk
, nskb
, 1, GFP_ATOMIC
)) {
1393 /* Decrement cwnd here because we are sending
1394 * effectively two packets. */
1396 update_send_head(sk
, nskb
);
1398 icsk
->icsk_mtup
.probe_size
= tcp_mss_to_mtu(sk
, nskb
->len
);
1399 tp
->mtu_probe
.probe_seq_start
= TCP_SKB_CB(nskb
)->seq
;
1400 tp
->mtu_probe
.probe_seq_end
= TCP_SKB_CB(nskb
)->end_seq
;
1409 /* This routine writes packets to the network. It advances the
1410 * send_head. This happens as incoming acks open up the remote
1413 * Returns 1, if no segments are in flight and we have queued segments, but
1414 * cannot send anything now because of SWS or another problem.
1416 static int tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
)
1418 struct tcp_sock
*tp
= tcp_sk(sk
);
1419 struct sk_buff
*skb
;
1420 unsigned int tso_segs
, sent_pkts
;
1424 /* If we are closed, the bytes will have to remain here.
1425 * In time closedown will finish, we empty the write queue and all
1428 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
1433 /* Do MTU probing. */
1434 if ((result
= tcp_mtu_probe(sk
)) == 0) {
1436 } else if (result
> 0) {
1440 while ((skb
= tcp_send_head(sk
))) {
1443 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1446 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1450 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
)))
1453 if (tso_segs
== 1) {
1454 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
1455 (tcp_skb_is_last(sk
, skb
) ?
1456 nonagle
: TCP_NAGLE_PUSH
))))
1459 if (tcp_tso_should_defer(sk
, skb
))
1465 limit
= tcp_window_allows(tp
, skb
,
1466 mss_now
, cwnd_quota
);
1468 if (skb
->len
< limit
) {
1469 unsigned int trim
= skb
->len
% mss_now
;
1472 limit
= skb
->len
- trim
;
1476 if (skb
->len
> limit
&&
1477 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
)))
1480 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1482 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
)))
1485 /* Advance the send_head. This one is sent out.
1486 * This call will increment packets_out.
1488 update_send_head(sk
, skb
);
1490 tcp_minshall_update(tp
, mss_now
, skb
);
1494 if (likely(sent_pkts
)) {
1495 tcp_cwnd_validate(sk
);
1498 return !tp
->packets_out
&& tcp_send_head(sk
);
1501 /* Push out any pending frames which were held back due to
1502 * TCP_CORK or attempt at coalescing tiny packets.
1503 * The socket must be locked by the caller.
1505 void __tcp_push_pending_frames(struct sock
*sk
, unsigned int cur_mss
,
1508 struct sk_buff
*skb
= tcp_send_head(sk
);
1511 if (tcp_write_xmit(sk
, cur_mss
, nonagle
))
1512 tcp_check_probe_timer(sk
);
1516 /* Send _single_ skb sitting at the send head. This function requires
1517 * true push pending frames to setup probe timer etc.
1519 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
1521 struct tcp_sock
*tp
= tcp_sk(sk
);
1522 struct sk_buff
*skb
= tcp_send_head(sk
);
1523 unsigned int tso_segs
, cwnd_quota
;
1525 BUG_ON(!skb
|| skb
->len
< mss_now
);
1527 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1528 cwnd_quota
= tcp_snd_test(sk
, skb
, mss_now
, TCP_NAGLE_PUSH
);
1530 if (likely(cwnd_quota
)) {
1537 limit
= tcp_window_allows(tp
, skb
,
1538 mss_now
, cwnd_quota
);
1540 if (skb
->len
< limit
) {
1541 unsigned int trim
= skb
->len
% mss_now
;
1544 limit
= skb
->len
- trim
;
1548 if (skb
->len
> limit
&&
1549 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
)))
1552 /* Send it out now. */
1553 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1555 if (likely(!tcp_transmit_skb(sk
, skb
, 1, sk
->sk_allocation
))) {
1556 update_send_head(sk
, skb
);
1557 tcp_cwnd_validate(sk
);
1563 /* This function returns the amount that we can raise the
1564 * usable window based on the following constraints
1566 * 1. The window can never be shrunk once it is offered (RFC 793)
1567 * 2. We limit memory per socket
1570 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1571 * RECV.NEXT + RCV.WIN fixed until:
1572 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1574 * i.e. don't raise the right edge of the window until you can raise
1575 * it at least MSS bytes.
1577 * Unfortunately, the recommended algorithm breaks header prediction,
1578 * since header prediction assumes th->window stays fixed.
1580 * Strictly speaking, keeping th->window fixed violates the receiver
1581 * side SWS prevention criteria. The problem is that under this rule
1582 * a stream of single byte packets will cause the right side of the
1583 * window to always advance by a single byte.
1585 * Of course, if the sender implements sender side SWS prevention
1586 * then this will not be a problem.
1588 * BSD seems to make the following compromise:
1590 * If the free space is less than the 1/4 of the maximum
1591 * space available and the free space is less than 1/2 mss,
1592 * then set the window to 0.
1593 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1594 * Otherwise, just prevent the window from shrinking
1595 * and from being larger than the largest representable value.
1597 * This prevents incremental opening of the window in the regime
1598 * where TCP is limited by the speed of the reader side taking
1599 * data out of the TCP receive queue. It does nothing about
1600 * those cases where the window is constrained on the sender side
1601 * because the pipeline is full.
1603 * BSD also seems to "accidentally" limit itself to windows that are a
1604 * multiple of MSS, at least until the free space gets quite small.
1605 * This would appear to be a side effect of the mbuf implementation.
1606 * Combining these two algorithms results in the observed behavior
1607 * of having a fixed window size at almost all times.
1609 * Below we obtain similar behavior by forcing the offered window to
1610 * a multiple of the mss when it is feasible to do so.
1612 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1613 * Regular options like TIMESTAMP are taken into account.
1615 u32
__tcp_select_window(struct sock
*sk
)
1617 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1618 struct tcp_sock
*tp
= tcp_sk(sk
);
1619 /* MSS for the peer's data. Previous versions used mss_clamp
1620 * here. I don't know if the value based on our guesses
1621 * of peer's MSS is better for the performance. It's more correct
1622 * but may be worse for the performance because of rcv_mss
1623 * fluctuations. --SAW 1998/11/1
1625 int mss
= icsk
->icsk_ack
.rcv_mss
;
1626 int free_space
= tcp_space(sk
);
1627 int full_space
= min_t(int, tp
->window_clamp
, tcp_full_space(sk
));
1630 if (mss
> full_space
)
1633 if (free_space
< full_space
/2) {
1634 icsk
->icsk_ack
.quick
= 0;
1636 if (tcp_memory_pressure
)
1637 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U*tp
->advmss
);
1639 if (free_space
< mss
)
1643 if (free_space
> tp
->rcv_ssthresh
)
1644 free_space
= tp
->rcv_ssthresh
;
1646 /* Don't do rounding if we are using window scaling, since the
1647 * scaled window will not line up with the MSS boundary anyway.
1649 window
= tp
->rcv_wnd
;
1650 if (tp
->rx_opt
.rcv_wscale
) {
1651 window
= free_space
;
1653 /* Advertise enough space so that it won't get scaled away.
1654 * Import case: prevent zero window announcement if
1655 * 1<<rcv_wscale > mss.
1657 if (((window
>> tp
->rx_opt
.rcv_wscale
) << tp
->rx_opt
.rcv_wscale
) != window
)
1658 window
= (((window
>> tp
->rx_opt
.rcv_wscale
) + 1)
1659 << tp
->rx_opt
.rcv_wscale
);
1661 /* Get the largest window that is a nice multiple of mss.
1662 * Window clamp already applied above.
1663 * If our current window offering is within 1 mss of the
1664 * free space we just keep it. This prevents the divide
1665 * and multiply from happening most of the time.
1666 * We also don't do any window rounding when the free space
1669 if (window
<= free_space
- mss
|| window
> free_space
)
1670 window
= (free_space
/mss
)*mss
;
1671 else if (mss
== full_space
&&
1672 free_space
> window
+ full_space
/2)
1673 window
= free_space
;
1679 /* Attempt to collapse two adjacent SKB's during retransmission. */
1680 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*skb
, int mss_now
)
1682 struct tcp_sock
*tp
= tcp_sk(sk
);
1683 struct sk_buff
*next_skb
= tcp_write_queue_next(sk
, skb
);
1685 /* The first test we must make is that neither of these two
1686 * SKB's are still referenced by someone else.
1688 if (!skb_cloned(skb
) && !skb_cloned(next_skb
)) {
1689 int skb_size
= skb
->len
, next_skb_size
= next_skb
->len
;
1690 u16 flags
= TCP_SKB_CB(skb
)->flags
;
1692 /* Also punt if next skb has been SACK'd. */
1693 if (TCP_SKB_CB(next_skb
)->sacked
& TCPCB_SACKED_ACKED
)
1696 /* Next skb is out of window. */
1697 if (after(TCP_SKB_CB(next_skb
)->end_seq
, tp
->snd_una
+tp
->snd_wnd
))
1700 /* Punt if not enough space exists in the first SKB for
1701 * the data in the second, or the total combined payload
1702 * would exceed the MSS.
1704 if ((next_skb_size
> skb_tailroom(skb
)) ||
1705 ((skb_size
+ next_skb_size
) > mss_now
))
1708 BUG_ON(tcp_skb_pcount(skb
) != 1 ||
1709 tcp_skb_pcount(next_skb
) != 1);
1711 /* changing transmit queue under us so clear hints */
1712 clear_all_retrans_hints(tp
);
1714 /* Ok. We will be able to collapse the packet. */
1715 tcp_unlink_write_queue(next_skb
, sk
);
1717 skb_copy_from_linear_data(next_skb
,
1718 skb_put(skb
, next_skb_size
),
1721 if (next_skb
->ip_summed
== CHECKSUM_PARTIAL
)
1722 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1724 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1725 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
1727 /* Update sequence range on original skb. */
1728 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
1730 /* Merge over control information. */
1731 flags
|= TCP_SKB_CB(next_skb
)->flags
; /* This moves PSH/FIN etc. over */
1732 TCP_SKB_CB(skb
)->flags
= flags
;
1734 /* All done, get rid of second SKB and account for it so
1735 * packet counting does not break.
1737 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
&(TCPCB_EVER_RETRANS
|TCPCB_AT_TAIL
);
1738 if (TCP_SKB_CB(next_skb
)->sacked
&TCPCB_SACKED_RETRANS
)
1739 tp
->retrans_out
-= tcp_skb_pcount(next_skb
);
1740 if (TCP_SKB_CB(next_skb
)->sacked
&TCPCB_LOST
)
1741 tp
->lost_out
-= tcp_skb_pcount(next_skb
);
1742 /* Reno case is special. Sigh... */
1743 if (tcp_is_reno(tp
) && tp
->sacked_out
)
1744 tcp_dec_pcount_approx(&tp
->sacked_out
, next_skb
);
1746 /* Not quite right: it can be > snd.fack, but
1747 * it is better to underestimate fackets.
1749 tcp_dec_pcount_approx(&tp
->fackets_out
, next_skb
);
1750 tp
->packets_out
-= tcp_skb_pcount(next_skb
);
1751 sk_stream_free_skb(sk
, next_skb
);
1755 /* Do a simple retransmit without using the backoff mechanisms in
1756 * tcp_timer. This is used for path mtu discovery.
1757 * The socket is already locked here.
1759 void tcp_simple_retransmit(struct sock
*sk
)
1761 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1762 struct tcp_sock
*tp
= tcp_sk(sk
);
1763 struct sk_buff
*skb
;
1764 unsigned int mss
= tcp_current_mss(sk
, 0);
1767 tcp_for_write_queue(skb
, sk
) {
1768 if (skb
== tcp_send_head(sk
))
1770 if (skb
->len
> mss
&&
1771 !(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
)) {
1772 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) {
1773 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1774 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1776 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_LOST
)) {
1777 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1778 tp
->lost_out
+= tcp_skb_pcount(skb
);
1784 clear_all_retrans_hints(tp
);
1789 tcp_verify_left_out(tp
);
1791 /* Don't muck with the congestion window here.
1792 * Reason is that we do not increase amount of _data_
1793 * in network, but units changed and effective
1794 * cwnd/ssthresh really reduced now.
1796 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
1797 tp
->high_seq
= tp
->snd_nxt
;
1798 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
1799 tp
->prior_ssthresh
= 0;
1800 tp
->undo_marker
= 0;
1801 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1803 tcp_xmit_retransmit_queue(sk
);
1806 /* This retransmits one SKB. Policy decisions and retransmit queue
1807 * state updates are done by the caller. Returns non-zero if an
1808 * error occurred which prevented the send.
1810 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
1812 struct tcp_sock
*tp
= tcp_sk(sk
);
1813 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1814 unsigned int cur_mss
= tcp_current_mss(sk
, 0);
1817 /* Inconslusive MTU probe */
1818 if (icsk
->icsk_mtup
.probe_size
) {
1819 icsk
->icsk_mtup
.probe_size
= 0;
1822 /* Do not sent more than we queued. 1/4 is reserved for possible
1823 * copying overhead: fragmentation, tunneling, mangling etc.
1825 if (atomic_read(&sk
->sk_wmem_alloc
) >
1826 min(sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2), sk
->sk_sndbuf
))
1829 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
1830 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1832 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
1836 /* If receiver has shrunk his window, and skb is out of
1837 * new window, do not retransmit it. The exception is the
1838 * case, when window is shrunk to zero. In this case
1839 * our retransmit serves as a zero window probe.
1841 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
+tp
->snd_wnd
)
1842 && TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
1845 if (skb
->len
> cur_mss
) {
1846 if (tcp_fragment(sk
, skb
, cur_mss
, cur_mss
))
1847 return -ENOMEM
; /* We'll try again later. */
1850 /* Collapse two adjacent packets if worthwhile and we can. */
1851 if (!(TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_SYN
) &&
1852 (skb
->len
< (cur_mss
>> 1)) &&
1853 (tcp_write_queue_next(sk
, skb
) != tcp_send_head(sk
)) &&
1854 (!tcp_skb_is_last(sk
, skb
)) &&
1855 (skb_shinfo(skb
)->nr_frags
== 0 && skb_shinfo(tcp_write_queue_next(sk
, skb
))->nr_frags
== 0) &&
1856 (tcp_skb_pcount(skb
) == 1 && tcp_skb_pcount(tcp_write_queue_next(sk
, skb
)) == 1) &&
1857 (sysctl_tcp_retrans_collapse
!= 0))
1858 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
1860 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
1861 return -EHOSTUNREACH
; /* Routing failure or similar. */
1863 /* Some Solaris stacks overoptimize and ignore the FIN on a
1864 * retransmit when old data is attached. So strip it off
1865 * since it is cheap to do so and saves bytes on the network.
1868 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
) &&
1869 tp
->snd_una
== (TCP_SKB_CB(skb
)->end_seq
- 1)) {
1870 if (!pskb_trim(skb
, 0)) {
1871 TCP_SKB_CB(skb
)->seq
= TCP_SKB_CB(skb
)->end_seq
- 1;
1872 skb_shinfo(skb
)->gso_segs
= 1;
1873 skb_shinfo(skb
)->gso_size
= 0;
1874 skb_shinfo(skb
)->gso_type
= 0;
1875 skb
->ip_summed
= CHECKSUM_NONE
;
1880 /* Make a copy, if the first transmission SKB clone we made
1881 * is still in somebody's hands, else make a clone.
1883 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1885 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
1888 /* Update global TCP statistics. */
1889 TCP_INC_STATS(TCP_MIB_RETRANSSEGS
);
1891 tp
->total_retrans
++;
1893 #if FASTRETRANS_DEBUG > 0
1894 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) {
1895 if (net_ratelimit())
1896 printk(KERN_DEBUG
"retrans_out leaked.\n");
1899 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
1900 tp
->retrans_out
+= tcp_skb_pcount(skb
);
1902 /* Save stamp of the first retransmit. */
1903 if (!tp
->retrans_stamp
)
1904 tp
->retrans_stamp
= TCP_SKB_CB(skb
)->when
;
1908 /* snd_nxt is stored to detect loss of retransmitted segment,
1909 * see tcp_input.c tcp_sacktag_write_queue().
1911 TCP_SKB_CB(skb
)->ack_seq
= tp
->snd_nxt
;
1916 /* This gets called after a retransmit timeout, and the initially
1917 * retransmitted data is acknowledged. It tries to continue
1918 * resending the rest of the retransmit queue, until either
1919 * we've sent it all or the congestion window limit is reached.
1920 * If doing SACK, the first ACK which comes back for a timeout
1921 * based retransmit packet might feed us FACK information again.
1922 * If so, we use it to avoid unnecessarily retransmissions.
1924 void tcp_xmit_retransmit_queue(struct sock
*sk
)
1926 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1927 struct tcp_sock
*tp
= tcp_sk(sk
);
1928 struct sk_buff
*skb
;
1931 if (tp
->retransmit_skb_hint
) {
1932 skb
= tp
->retransmit_skb_hint
;
1933 packet_cnt
= tp
->retransmit_cnt_hint
;
1935 skb
= tcp_write_queue_head(sk
);
1939 /* First pass: retransmit lost packets. */
1941 tcp_for_write_queue_from(skb
, sk
) {
1942 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
1944 if (skb
== tcp_send_head(sk
))
1946 /* we could do better than to assign each time */
1947 tp
->retransmit_skb_hint
= skb
;
1948 tp
->retransmit_cnt_hint
= packet_cnt
;
1950 /* Assume this retransmit will generate
1951 * only one packet for congestion window
1952 * calculation purposes. This works because
1953 * tcp_retransmit_skb() will chop up the
1954 * packet to be MSS sized and all the
1955 * packet counting works out.
1957 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
1960 if (sacked
& TCPCB_LOST
) {
1961 if (!(sacked
&(TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))) {
1962 if (tcp_retransmit_skb(sk
, skb
)) {
1963 tp
->retransmit_skb_hint
= NULL
;
1966 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
1967 NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS
);
1969 NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS
);
1971 if (skb
== tcp_write_queue_head(sk
))
1972 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1973 inet_csk(sk
)->icsk_rto
,
1977 packet_cnt
+= tcp_skb_pcount(skb
);
1978 if (packet_cnt
>= tp
->lost_out
)
1984 /* OK, demanded retransmission is finished. */
1986 /* Forward retransmissions are possible only during Recovery. */
1987 if (icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
1990 /* No forward retransmissions in Reno are possible. */
1991 if (tcp_is_reno(tp
))
1994 /* Yeah, we have to make difficult choice between forward transmission
1995 * and retransmission... Both ways have their merits...
1997 * For now we do not retransmit anything, while we have some new
1998 * segments to send. In the other cases, follow rule 3 for
1999 * NextSeg() specified in RFC3517.
2002 if (tcp_may_send_now(sk
))
2005 /* If nothing is SACKed, highest_sack in the loop won't be valid */
2006 if (!tp
->sacked_out
)
2009 if (tp
->forward_skb_hint
)
2010 skb
= tp
->forward_skb_hint
;
2012 skb
= tcp_write_queue_head(sk
);
2014 tcp_for_write_queue_from(skb
, sk
) {
2015 if (skb
== tcp_send_head(sk
))
2017 tp
->forward_skb_hint
= skb
;
2019 if (after(TCP_SKB_CB(skb
)->seq
, tp
->highest_sack
))
2022 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
2025 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
)
2028 /* Ok, retransmit it. */
2029 if (tcp_retransmit_skb(sk
, skb
)) {
2030 tp
->forward_skb_hint
= NULL
;
2034 if (skb
== tcp_write_queue_head(sk
))
2035 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2036 inet_csk(sk
)->icsk_rto
,
2039 NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS
);
2044 /* Send a fin. The caller locks the socket for us. This cannot be
2045 * allowed to fail queueing a FIN frame under any circumstances.
2047 void tcp_send_fin(struct sock
*sk
)
2049 struct tcp_sock
*tp
= tcp_sk(sk
);
2050 struct sk_buff
*skb
= tcp_write_queue_tail(sk
);
2053 /* Optimization, tack on the FIN if we have a queue of
2054 * unsent frames. But be careful about outgoing SACKS
2057 mss_now
= tcp_current_mss(sk
, 1);
2059 if (tcp_send_head(sk
) != NULL
) {
2060 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_FIN
;
2061 TCP_SKB_CB(skb
)->end_seq
++;
2064 /* Socket is locked, keep trying until memory is available. */
2066 skb
= alloc_skb_fclone(MAX_TCP_HEADER
, GFP_KERNEL
);
2072 /* Reserve space for headers and prepare control bits. */
2073 skb_reserve(skb
, MAX_TCP_HEADER
);
2075 TCP_SKB_CB(skb
)->flags
= (TCPCB_FLAG_ACK
| TCPCB_FLAG_FIN
);
2076 TCP_SKB_CB(skb
)->sacked
= 0;
2077 skb_shinfo(skb
)->gso_segs
= 1;
2078 skb_shinfo(skb
)->gso_size
= 0;
2079 skb_shinfo(skb
)->gso_type
= 0;
2081 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2082 TCP_SKB_CB(skb
)->seq
= tp
->write_seq
;
2083 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ 1;
2084 tcp_queue_skb(sk
, skb
);
2086 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_OFF
);
2089 /* We get here when a process closes a file descriptor (either due to
2090 * an explicit close() or as a byproduct of exit()'ing) and there
2091 * was unread data in the receive queue. This behavior is recommended
2092 * by RFC 2525, section 2.17. -DaveM
2094 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
2096 struct sk_buff
*skb
;
2098 /* NOTE: No TCP options attached and we never retransmit this. */
2099 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
2101 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED
);
2105 /* Reserve space for headers and prepare control bits. */
2106 skb_reserve(skb
, MAX_TCP_HEADER
);
2108 TCP_SKB_CB(skb
)->flags
= (TCPCB_FLAG_ACK
| TCPCB_FLAG_RST
);
2109 TCP_SKB_CB(skb
)->sacked
= 0;
2110 skb_shinfo(skb
)->gso_segs
= 1;
2111 skb_shinfo(skb
)->gso_size
= 0;
2112 skb_shinfo(skb
)->gso_type
= 0;
2115 TCP_SKB_CB(skb
)->seq
= tcp_acceptable_seq(sk
);
2116 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
;
2117 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2118 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
2119 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED
);
2122 /* WARNING: This routine must only be called when we have already sent
2123 * a SYN packet that crossed the incoming SYN that caused this routine
2124 * to get called. If this assumption fails then the initial rcv_wnd
2125 * and rcv_wscale values will not be correct.
2127 int tcp_send_synack(struct sock
*sk
)
2129 struct sk_buff
* skb
;
2131 skb
= tcp_write_queue_head(sk
);
2132 if (skb
== NULL
|| !(TCP_SKB_CB(skb
)->flags
&TCPCB_FLAG_SYN
)) {
2133 printk(KERN_DEBUG
"tcp_send_synack: wrong queue state\n");
2136 if (!(TCP_SKB_CB(skb
)->flags
&TCPCB_FLAG_ACK
)) {
2137 if (skb_cloned(skb
)) {
2138 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
2141 tcp_unlink_write_queue(skb
, sk
);
2142 skb_header_release(nskb
);
2143 __tcp_add_write_queue_head(sk
, nskb
);
2144 sk_stream_free_skb(sk
, skb
);
2145 sk_charge_skb(sk
, nskb
);
2149 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_ACK
;
2150 TCP_ECN_send_synack(tcp_sk(sk
), skb
);
2152 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2153 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2157 * Prepare a SYN-ACK.
2159 struct sk_buff
* tcp_make_synack(struct sock
*sk
, struct dst_entry
*dst
,
2160 struct request_sock
*req
)
2162 struct inet_request_sock
*ireq
= inet_rsk(req
);
2163 struct tcp_sock
*tp
= tcp_sk(sk
);
2165 int tcp_header_size
;
2166 struct sk_buff
*skb
;
2167 #ifdef CONFIG_TCP_MD5SIG
2168 struct tcp_md5sig_key
*md5
;
2169 __u8
*md5_hash_location
;
2172 skb
= sock_wmalloc(sk
, MAX_TCP_HEADER
+ 15, 1, GFP_ATOMIC
);
2176 /* Reserve space for headers. */
2177 skb_reserve(skb
, MAX_TCP_HEADER
);
2179 skb
->dst
= dst_clone(dst
);
2181 tcp_header_size
= (sizeof(struct tcphdr
) + TCPOLEN_MSS
+
2182 (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0) +
2183 (ireq
->wscale_ok
? TCPOLEN_WSCALE_ALIGNED
: 0) +
2184 /* SACK_PERM is in the place of NOP NOP of TS */
2185 ((ireq
->sack_ok
&& !ireq
->tstamp_ok
) ? TCPOLEN_SACKPERM_ALIGNED
: 0));
2187 #ifdef CONFIG_TCP_MD5SIG
2188 /* Are we doing MD5 on this segment? If so - make room for it */
2189 md5
= tcp_rsk(req
)->af_specific
->md5_lookup(sk
, req
);
2191 tcp_header_size
+= TCPOLEN_MD5SIG_ALIGNED
;
2193 skb_push(skb
, tcp_header_size
);
2194 skb_reset_transport_header(skb
);
2197 memset(th
, 0, sizeof(struct tcphdr
));
2200 TCP_ECN_make_synack(req
, th
);
2201 th
->source
= inet_sk(sk
)->sport
;
2202 th
->dest
= ireq
->rmt_port
;
2203 TCP_SKB_CB(skb
)->seq
= tcp_rsk(req
)->snt_isn
;
2204 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ 1;
2205 TCP_SKB_CB(skb
)->sacked
= 0;
2206 skb_shinfo(skb
)->gso_segs
= 1;
2207 skb_shinfo(skb
)->gso_size
= 0;
2208 skb_shinfo(skb
)->gso_type
= 0;
2209 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
2210 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_isn
+ 1);
2211 if (req
->rcv_wnd
== 0) { /* ignored for retransmitted syns */
2213 /* Set this up on the first call only */
2214 req
->window_clamp
= tp
->window_clamp
? : dst_metric(dst
, RTAX_WINDOW
);
2215 /* tcp_full_space because it is guaranteed to be the first packet */
2216 tcp_select_initial_window(tcp_full_space(sk
),
2217 dst_metric(dst
, RTAX_ADVMSS
) - (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0),
2222 ireq
->rcv_wscale
= rcv_wscale
;
2225 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2226 th
->window
= htons(min(req
->rcv_wnd
, 65535U));
2228 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2229 tcp_syn_build_options((__be32
*)(th
+ 1), dst_metric(dst
, RTAX_ADVMSS
), ireq
->tstamp_ok
,
2230 ireq
->sack_ok
, ireq
->wscale_ok
, ireq
->rcv_wscale
,
2231 TCP_SKB_CB(skb
)->when
,
2234 #ifdef CONFIG_TCP_MD5SIG
2235 md5
? &md5_hash_location
:
2241 th
->doff
= (tcp_header_size
>> 2);
2242 TCP_INC_STATS(TCP_MIB_OUTSEGS
);
2244 #ifdef CONFIG_TCP_MD5SIG
2245 /* Okay, we have all we need - do the md5 hash if needed */
2247 tp
->af_specific
->calc_md5_hash(md5_hash_location
,
2250 tcp_hdr(skb
), sk
->sk_protocol
,
2259 * Do all connect socket setups that can be done AF independent.
2261 static void tcp_connect_init(struct sock
*sk
)
2263 struct dst_entry
*dst
= __sk_dst_get(sk
);
2264 struct tcp_sock
*tp
= tcp_sk(sk
);
2267 /* We'll fix this up when we get a response from the other end.
2268 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2270 tp
->tcp_header_len
= sizeof(struct tcphdr
) +
2271 (sysctl_tcp_timestamps
? TCPOLEN_TSTAMP_ALIGNED
: 0);
2273 #ifdef CONFIG_TCP_MD5SIG
2274 if (tp
->af_specific
->md5_lookup(sk
, sk
) != NULL
)
2275 tp
->tcp_header_len
+= TCPOLEN_MD5SIG_ALIGNED
;
2278 /* If user gave his TCP_MAXSEG, record it to clamp */
2279 if (tp
->rx_opt
.user_mss
)
2280 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
2283 tcp_sync_mss(sk
, dst_mtu(dst
));
2285 if (!tp
->window_clamp
)
2286 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
2287 tp
->advmss
= dst_metric(dst
, RTAX_ADVMSS
);
2288 tcp_initialize_rcv_mss(sk
);
2290 tcp_select_initial_window(tcp_full_space(sk
),
2291 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
2294 sysctl_tcp_window_scaling
,
2297 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
2298 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
2301 sock_reset_flag(sk
, SOCK_DONE
);
2303 tcp_init_wl(tp
, tp
->write_seq
, 0);
2304 tp
->snd_una
= tp
->write_seq
;
2305 tp
->snd_sml
= tp
->write_seq
;
2310 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
2311 inet_csk(sk
)->icsk_retransmits
= 0;
2312 tcp_clear_retrans(tp
);
2316 * Build a SYN and send it off.
2318 int tcp_connect(struct sock
*sk
)
2320 struct tcp_sock
*tp
= tcp_sk(sk
);
2321 struct sk_buff
*buff
;
2323 tcp_connect_init(sk
);
2325 buff
= alloc_skb_fclone(MAX_TCP_HEADER
+ 15, sk
->sk_allocation
);
2326 if (unlikely(buff
== NULL
))
2329 /* Reserve space for headers. */
2330 skb_reserve(buff
, MAX_TCP_HEADER
);
2332 TCP_SKB_CB(buff
)->flags
= TCPCB_FLAG_SYN
;
2333 TCP_ECN_send_syn(sk
, buff
);
2334 TCP_SKB_CB(buff
)->sacked
= 0;
2335 skb_shinfo(buff
)->gso_segs
= 1;
2336 skb_shinfo(buff
)->gso_size
= 0;
2337 skb_shinfo(buff
)->gso_type
= 0;
2339 tp
->snd_nxt
= tp
->write_seq
;
2340 TCP_SKB_CB(buff
)->seq
= tp
->write_seq
++;
2341 TCP_SKB_CB(buff
)->end_seq
= tp
->write_seq
;
2344 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2345 tp
->retrans_stamp
= TCP_SKB_CB(buff
)->when
;
2346 skb_header_release(buff
);
2347 __tcp_add_write_queue_tail(sk
, buff
);
2348 sk_charge_skb(sk
, buff
);
2349 tp
->packets_out
+= tcp_skb_pcount(buff
);
2350 tcp_transmit_skb(sk
, buff
, 1, GFP_KERNEL
);
2352 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2353 * in order to make this packet get counted in tcpOutSegs.
2355 tp
->snd_nxt
= tp
->write_seq
;
2356 tp
->pushed_seq
= tp
->write_seq
;
2357 TCP_INC_STATS(TCP_MIB_ACTIVEOPENS
);
2359 /* Timer for repeating the SYN until an answer. */
2360 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2361 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2365 /* Send out a delayed ack, the caller does the policy checking
2366 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2369 void tcp_send_delayed_ack(struct sock
*sk
)
2371 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2372 int ato
= icsk
->icsk_ack
.ato
;
2373 unsigned long timeout
;
2375 if (ato
> TCP_DELACK_MIN
) {
2376 const struct tcp_sock
*tp
= tcp_sk(sk
);
2379 if (icsk
->icsk_ack
.pingpong
|| (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
2380 max_ato
= TCP_DELACK_MAX
;
2382 /* Slow path, intersegment interval is "high". */
2384 /* If some rtt estimate is known, use it to bound delayed ack.
2385 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2389 int rtt
= max(tp
->srtt
>>3, TCP_DELACK_MIN
);
2395 ato
= min(ato
, max_ato
);
2398 /* Stay within the limit we were given */
2399 timeout
= jiffies
+ ato
;
2401 /* Use new timeout only if there wasn't a older one earlier. */
2402 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
2403 /* If delack timer was blocked or is about to expire,
2406 if (icsk
->icsk_ack
.blocked
||
2407 time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
2412 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
2413 timeout
= icsk
->icsk_ack
.timeout
;
2415 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
2416 icsk
->icsk_ack
.timeout
= timeout
;
2417 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
2420 /* This routine sends an ack and also updates the window. */
2421 void tcp_send_ack(struct sock
*sk
)
2423 /* If we have been reset, we may not send again. */
2424 if (sk
->sk_state
!= TCP_CLOSE
) {
2425 struct sk_buff
*buff
;
2427 /* We are not putting this on the write queue, so
2428 * tcp_transmit_skb() will set the ownership to this
2431 buff
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2433 inet_csk_schedule_ack(sk
);
2434 inet_csk(sk
)->icsk_ack
.ato
= TCP_ATO_MIN
;
2435 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
2436 TCP_DELACK_MAX
, TCP_RTO_MAX
);
2440 /* Reserve space for headers and prepare control bits. */
2441 skb_reserve(buff
, MAX_TCP_HEADER
);
2443 TCP_SKB_CB(buff
)->flags
= TCPCB_FLAG_ACK
;
2444 TCP_SKB_CB(buff
)->sacked
= 0;
2445 skb_shinfo(buff
)->gso_segs
= 1;
2446 skb_shinfo(buff
)->gso_size
= 0;
2447 skb_shinfo(buff
)->gso_type
= 0;
2449 /* Send it off, this clears delayed acks for us. */
2450 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(buff
)->end_seq
= tcp_acceptable_seq(sk
);
2451 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2452 tcp_transmit_skb(sk
, buff
, 0, GFP_ATOMIC
);
2456 /* This routine sends a packet with an out of date sequence
2457 * number. It assumes the other end will try to ack it.
2459 * Question: what should we make while urgent mode?
2460 * 4.4BSD forces sending single byte of data. We cannot send
2461 * out of window data, because we have SND.NXT==SND.MAX...
2463 * Current solution: to send TWO zero-length segments in urgent mode:
2464 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2465 * out-of-date with SND.UNA-1 to probe window.
2467 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
)
2469 struct tcp_sock
*tp
= tcp_sk(sk
);
2470 struct sk_buff
*skb
;
2472 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2473 skb
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2477 /* Reserve space for headers and set control bits. */
2478 skb_reserve(skb
, MAX_TCP_HEADER
);
2480 TCP_SKB_CB(skb
)->flags
= TCPCB_FLAG_ACK
;
2481 TCP_SKB_CB(skb
)->sacked
= urgent
;
2482 skb_shinfo(skb
)->gso_segs
= 1;
2483 skb_shinfo(skb
)->gso_size
= 0;
2484 skb_shinfo(skb
)->gso_type
= 0;
2486 /* Use a previous sequence. This should cause the other
2487 * end to send an ack. Don't queue or clone SKB, just
2490 TCP_SKB_CB(skb
)->seq
= urgent
? tp
->snd_una
: tp
->snd_una
- 1;
2491 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
;
2492 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2493 return tcp_transmit_skb(sk
, skb
, 0, GFP_ATOMIC
);
2496 int tcp_write_wakeup(struct sock
*sk
)
2498 if (sk
->sk_state
!= TCP_CLOSE
) {
2499 struct tcp_sock
*tp
= tcp_sk(sk
);
2500 struct sk_buff
*skb
;
2502 if ((skb
= tcp_send_head(sk
)) != NULL
&&
2503 before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
+tp
->snd_wnd
)) {
2505 unsigned int mss
= tcp_current_mss(sk
, 0);
2506 unsigned int seg_size
= tp
->snd_una
+tp
->snd_wnd
-TCP_SKB_CB(skb
)->seq
;
2508 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
2509 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
2511 /* We are probing the opening of a window
2512 * but the window size is != 0
2513 * must have been a result SWS avoidance ( sender )
2515 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
2517 seg_size
= min(seg_size
, mss
);
2518 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2519 if (tcp_fragment(sk
, skb
, seg_size
, mss
))
2521 } else if (!tcp_skb_pcount(skb
))
2522 tcp_set_skb_tso_segs(sk
, skb
, mss
);
2524 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2525 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2526 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2528 update_send_head(sk
, skb
);
2533 between(tp
->snd_up
, tp
->snd_una
+1, tp
->snd_una
+0xFFFF))
2534 tcp_xmit_probe_skb(sk
, TCPCB_URG
);
2535 return tcp_xmit_probe_skb(sk
, 0);
2541 /* A window probe timeout has occurred. If window is not closed send
2542 * a partial packet else a zero probe.
2544 void tcp_send_probe0(struct sock
*sk
)
2546 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2547 struct tcp_sock
*tp
= tcp_sk(sk
);
2550 err
= tcp_write_wakeup(sk
);
2552 if (tp
->packets_out
|| !tcp_send_head(sk
)) {
2553 /* Cancel probe timer, if it is not required. */
2554 icsk
->icsk_probes_out
= 0;
2555 icsk
->icsk_backoff
= 0;
2560 if (icsk
->icsk_backoff
< sysctl_tcp_retries2
)
2561 icsk
->icsk_backoff
++;
2562 icsk
->icsk_probes_out
++;
2563 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2564 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
2567 /* If packet was not sent due to local congestion,
2568 * do not backoff and do not remember icsk_probes_out.
2569 * Let local senders to fight for local resources.
2571 * Use accumulated backoff yet.
2573 if (!icsk
->icsk_probes_out
)
2574 icsk
->icsk_probes_out
= 1;
2575 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2576 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
,
2577 TCP_RESOURCE_PROBE_INTERVAL
),
2582 EXPORT_SYMBOL(tcp_connect
);
2583 EXPORT_SYMBOL(tcp_make_synack
);
2584 EXPORT_SYMBOL(tcp_simple_retransmit
);
2585 EXPORT_SYMBOL(tcp_sync_mss
);
2586 EXPORT_SYMBOL(sysctl_tcp_tso_win_divisor
);
2587 EXPORT_SYMBOL(tcp_mtup_init
);