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).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
22 * Changes: Pedro Roque : Retransmit queue handled by TCP.
23 * : Fragmentation on mtu decrease
24 * : Segment collapse on retransmit
27 * Linus Torvalds : send_delayed_ack
28 * David S. Miller : Charge memory using the right skb
29 * during syn/ack processing.
30 * David S. Miller : Output engine completely rewritten.
31 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
32 * Cacophonix Gaul : draft-minshall-nagle-01
33 * J Hadi Salim : ECN support
39 #include <linux/compiler.h>
40 #include <linux/module.h>
42 /* People can turn this off for buggy TCP's found in printers etc. */
43 int sysctl_tcp_retrans_collapse __read_mostly
= 1;
45 /* People can turn this on to work with those rare, broken TCPs that
46 * interpret the window field as a signed quantity.
48 int sysctl_tcp_workaround_signed_windows __read_mostly
= 0;
50 /* This limits the percentage of the congestion window which we
51 * will allow a single TSO frame to consume. Building TSO frames
52 * which are too large can cause TCP streams to be bursty.
54 int sysctl_tcp_tso_win_divisor __read_mostly
= 3;
56 int sysctl_tcp_mtu_probing __read_mostly
= 0;
57 int sysctl_tcp_base_mss __read_mostly
= 512;
59 /* By default, RFC2861 behavior. */
60 int sysctl_tcp_slow_start_after_idle __read_mostly
= 1;
62 static void tcp_event_new_data_sent(struct sock
*sk
, struct sk_buff
*skb
)
64 struct tcp_sock
*tp
= tcp_sk(sk
);
65 unsigned int prior_packets
= tp
->packets_out
;
67 tcp_advance_send_head(sk
, skb
);
68 tp
->snd_nxt
= TCP_SKB_CB(skb
)->end_seq
;
70 /* Don't override Nagle indefinately with F-RTO */
71 if (tp
->frto_counter
== 2)
74 tp
->packets_out
+= tcp_skb_pcount(skb
);
76 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
77 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
80 /* SND.NXT, if window was not shrunk.
81 * If window has been shrunk, what should we make? It is not clear at all.
82 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
83 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
84 * invalid. OK, let's make this for now:
86 static inline __u32
tcp_acceptable_seq(struct sock
*sk
)
88 struct tcp_sock
*tp
= tcp_sk(sk
);
90 if (!before(tcp_wnd_end(tp
), tp
->snd_nxt
))
93 return tcp_wnd_end(tp
);
96 /* Calculate mss to advertise in SYN segment.
97 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
99 * 1. It is independent of path mtu.
100 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
101 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
102 * attached devices, because some buggy hosts are confused by
104 * 4. We do not make 3, we advertise MSS, calculated from first
105 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
106 * This may be overridden via information stored in routing table.
107 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
108 * probably even Jumbo".
110 static __u16
tcp_advertise_mss(struct sock
*sk
)
112 struct tcp_sock
*tp
= tcp_sk(sk
);
113 struct dst_entry
*dst
= __sk_dst_get(sk
);
114 int mss
= tp
->advmss
;
116 if (dst
&& dst_metric(dst
, RTAX_ADVMSS
) < mss
) {
117 mss
= dst_metric(dst
, RTAX_ADVMSS
);
124 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
125 * This is the first part of cwnd validation mechanism. */
126 static void tcp_cwnd_restart(struct sock
*sk
, struct dst_entry
*dst
)
128 struct tcp_sock
*tp
= tcp_sk(sk
);
129 s32 delta
= tcp_time_stamp
- tp
->lsndtime
;
130 u32 restart_cwnd
= tcp_init_cwnd(tp
, dst
);
131 u32 cwnd
= tp
->snd_cwnd
;
133 tcp_ca_event(sk
, CA_EVENT_CWND_RESTART
);
135 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
136 restart_cwnd
= min(restart_cwnd
, cwnd
);
138 while ((delta
-= inet_csk(sk
)->icsk_rto
) > 0 && cwnd
> restart_cwnd
)
140 tp
->snd_cwnd
= max(cwnd
, restart_cwnd
);
141 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
142 tp
->snd_cwnd_used
= 0;
145 static void tcp_event_data_sent(struct tcp_sock
*tp
,
146 struct sk_buff
*skb
, struct sock
*sk
)
148 struct inet_connection_sock
*icsk
= inet_csk(sk
);
149 const u32 now
= tcp_time_stamp
;
151 if (sysctl_tcp_slow_start_after_idle
&&
152 (!tp
->packets_out
&& (s32
)(now
- tp
->lsndtime
) > icsk
->icsk_rto
))
153 tcp_cwnd_restart(sk
, __sk_dst_get(sk
));
157 /* If it is a reply for ato after last received
158 * packet, enter pingpong mode.
160 if ((u32
)(now
- icsk
->icsk_ack
.lrcvtime
) < icsk
->icsk_ack
.ato
)
161 icsk
->icsk_ack
.pingpong
= 1;
164 static inline void tcp_event_ack_sent(struct sock
*sk
, unsigned int pkts
)
166 tcp_dec_quickack_mode(sk
, pkts
);
167 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_DACK
);
170 /* Determine a window scaling and initial window to offer.
171 * Based on the assumption that the given amount of space
172 * will be offered. Store the results in the tp structure.
173 * NOTE: for smooth operation initial space offering should
174 * be a multiple of mss if possible. We assume here that mss >= 1.
175 * This MUST be enforced by all callers.
177 void tcp_select_initial_window(int __space
, __u32 mss
,
178 __u32
*rcv_wnd
, __u32
*window_clamp
,
179 int wscale_ok
, __u8
*rcv_wscale
)
181 unsigned int space
= (__space
< 0 ? 0 : __space
);
183 /* If no clamp set the clamp to the max possible scaled window */
184 if (*window_clamp
== 0)
185 (*window_clamp
) = (65535 << 14);
186 space
= min(*window_clamp
, space
);
188 /* Quantize space offering to a multiple of mss if possible. */
190 space
= (space
/ mss
) * mss
;
192 /* NOTE: offering an initial window larger than 32767
193 * will break some buggy TCP stacks. If the admin tells us
194 * it is likely we could be speaking with such a buggy stack
195 * we will truncate our initial window offering to 32K-1
196 * unless the remote has sent us a window scaling option,
197 * which we interpret as a sign the remote TCP is not
198 * misinterpreting the window field as a signed quantity.
200 if (sysctl_tcp_workaround_signed_windows
)
201 (*rcv_wnd
) = min(space
, MAX_TCP_WINDOW
);
207 /* Set window scaling on max possible window
208 * See RFC1323 for an explanation of the limit to 14
210 space
= max_t(u32
, sysctl_tcp_rmem
[2], sysctl_rmem_max
);
211 space
= min_t(u32
, space
, *window_clamp
);
212 while (space
> 65535 && (*rcv_wscale
) < 14) {
218 /* Set initial window to value enough for senders,
219 * following RFC2414. Senders, not following this RFC,
220 * will be satisfied with 2.
222 if (mss
> (1 << *rcv_wscale
)) {
228 if (*rcv_wnd
> init_cwnd
* mss
)
229 *rcv_wnd
= init_cwnd
* mss
;
232 /* Set the clamp no higher than max representable value */
233 (*window_clamp
) = min(65535U << (*rcv_wscale
), *window_clamp
);
236 /* Chose a new window to advertise, update state in tcp_sock for the
237 * socket, and return result with RFC1323 scaling applied. The return
238 * value can be stuffed directly into th->window for an outgoing
241 static u16
tcp_select_window(struct sock
*sk
)
243 struct tcp_sock
*tp
= tcp_sk(sk
);
244 u32 cur_win
= tcp_receive_window(tp
);
245 u32 new_win
= __tcp_select_window(sk
);
247 /* Never shrink the offered window */
248 if (new_win
< cur_win
) {
249 /* Danger Will Robinson!
250 * Don't update rcv_wup/rcv_wnd here or else
251 * we will not be able to advertise a zero
252 * window in time. --DaveM
254 * Relax Will Robinson.
256 new_win
= ALIGN(cur_win
, 1 << tp
->rx_opt
.rcv_wscale
);
258 tp
->rcv_wnd
= new_win
;
259 tp
->rcv_wup
= tp
->rcv_nxt
;
261 /* Make sure we do not exceed the maximum possible
264 if (!tp
->rx_opt
.rcv_wscale
&& sysctl_tcp_workaround_signed_windows
)
265 new_win
= min(new_win
, MAX_TCP_WINDOW
);
267 new_win
= min(new_win
, (65535U << tp
->rx_opt
.rcv_wscale
));
269 /* RFC1323 scaling applied */
270 new_win
>>= tp
->rx_opt
.rcv_wscale
;
272 /* If we advertise zero window, disable fast path. */
279 static inline void TCP_ECN_send_synack(struct tcp_sock
*tp
, struct sk_buff
*skb
)
281 TCP_SKB_CB(skb
)->flags
&= ~TCPCB_FLAG_CWR
;
282 if (!(tp
->ecn_flags
& TCP_ECN_OK
))
283 TCP_SKB_CB(skb
)->flags
&= ~TCPCB_FLAG_ECE
;
286 static inline void TCP_ECN_send_syn(struct sock
*sk
, struct sk_buff
*skb
)
288 struct tcp_sock
*tp
= tcp_sk(sk
);
291 if (sysctl_tcp_ecn
) {
292 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_ECE
| TCPCB_FLAG_CWR
;
293 tp
->ecn_flags
= TCP_ECN_OK
;
297 static __inline__
void
298 TCP_ECN_make_synack(struct request_sock
*req
, struct tcphdr
*th
)
300 if (inet_rsk(req
)->ecn_ok
)
304 static inline void TCP_ECN_send(struct sock
*sk
, struct sk_buff
*skb
,
307 struct tcp_sock
*tp
= tcp_sk(sk
);
309 if (tp
->ecn_flags
& TCP_ECN_OK
) {
310 /* Not-retransmitted data segment: set ECT and inject CWR. */
311 if (skb
->len
!= tcp_header_len
&&
312 !before(TCP_SKB_CB(skb
)->seq
, tp
->snd_nxt
)) {
314 if (tp
->ecn_flags
& TCP_ECN_QUEUE_CWR
) {
315 tp
->ecn_flags
&= ~TCP_ECN_QUEUE_CWR
;
316 tcp_hdr(skb
)->cwr
= 1;
317 skb_shinfo(skb
)->gso_type
|= SKB_GSO_TCP_ECN
;
320 /* ACK or retransmitted segment: clear ECT|CE */
321 INET_ECN_dontxmit(sk
);
323 if (tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)
324 tcp_hdr(skb
)->ece
= 1;
328 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
329 * auto increment end seqno.
331 static void tcp_init_nondata_skb(struct sk_buff
*skb
, u32 seq
, u8 flags
)
335 TCP_SKB_CB(skb
)->flags
= flags
;
336 TCP_SKB_CB(skb
)->sacked
= 0;
338 skb_shinfo(skb
)->gso_segs
= 1;
339 skb_shinfo(skb
)->gso_size
= 0;
340 skb_shinfo(skb
)->gso_type
= 0;
342 TCP_SKB_CB(skb
)->seq
= seq
;
343 if (flags
& (TCPCB_FLAG_SYN
| TCPCB_FLAG_FIN
))
345 TCP_SKB_CB(skb
)->end_seq
= seq
;
348 static inline int tcp_urg_mode(const struct tcp_sock
*tp
)
350 return tp
->snd_una
!= tp
->snd_up
;
353 #define OPTION_SACK_ADVERTISE (1 << 0)
354 #define OPTION_TS (1 << 1)
355 #define OPTION_MD5 (1 << 2)
357 struct tcp_out_options
{
358 u8 options
; /* bit field of OPTION_* */
359 u8 ws
; /* window scale, 0 to disable */
360 u8 num_sack_blocks
; /* number of SACK blocks to include */
361 u16 mss
; /* 0 to disable */
362 __u32 tsval
, tsecr
; /* need to include OPTION_TS */
365 /* Beware: Something in the Internet is very sensitive to the ordering of
366 * TCP options, we learned this through the hard way, so be careful here.
367 * Luckily we can at least blame others for their non-compliance but from
368 * inter-operatibility perspective it seems that we're somewhat stuck with
369 * the ordering which we have been using if we want to keep working with
370 * those broken things (not that it currently hurts anybody as there isn't
371 * particular reason why the ordering would need to be changed).
373 * At least SACK_PERM as the first option is known to lead to a disaster
374 * (but it may well be that other scenarios fail similarly).
376 static void tcp_options_write(__be32
*ptr
, struct tcp_sock
*tp
,
377 const struct tcp_out_options
*opts
,
379 if (unlikely(OPTION_MD5
& opts
->options
)) {
380 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
382 (TCPOPT_MD5SIG
<< 8) |
384 *md5_hash
= (__u8
*)ptr
;
390 if (unlikely(opts
->mss
)) {
391 *ptr
++ = htonl((TCPOPT_MSS
<< 24) |
392 (TCPOLEN_MSS
<< 16) |
396 if (likely(OPTION_TS
& opts
->options
)) {
397 if (unlikely(OPTION_SACK_ADVERTISE
& opts
->options
)) {
398 *ptr
++ = htonl((TCPOPT_SACK_PERM
<< 24) |
399 (TCPOLEN_SACK_PERM
<< 16) |
400 (TCPOPT_TIMESTAMP
<< 8) |
403 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
405 (TCPOPT_TIMESTAMP
<< 8) |
408 *ptr
++ = htonl(opts
->tsval
);
409 *ptr
++ = htonl(opts
->tsecr
);
412 if (unlikely(OPTION_SACK_ADVERTISE
& opts
->options
&&
413 !(OPTION_TS
& opts
->options
))) {
414 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
416 (TCPOPT_SACK_PERM
<< 8) |
420 if (unlikely(opts
->ws
)) {
421 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
422 (TCPOPT_WINDOW
<< 16) |
423 (TCPOLEN_WINDOW
<< 8) |
427 if (unlikely(opts
->num_sack_blocks
)) {
428 struct tcp_sack_block
*sp
= tp
->rx_opt
.dsack
?
429 tp
->duplicate_sack
: tp
->selective_acks
;
432 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
435 (TCPOLEN_SACK_BASE
+ (opts
->num_sack_blocks
*
436 TCPOLEN_SACK_PERBLOCK
)));
438 for (this_sack
= 0; this_sack
< opts
->num_sack_blocks
;
440 *ptr
++ = htonl(sp
[this_sack
].start_seq
);
441 *ptr
++ = htonl(sp
[this_sack
].end_seq
);
444 if (tp
->rx_opt
.dsack
) {
445 tp
->rx_opt
.dsack
= 0;
446 tp
->rx_opt
.eff_sacks
= tp
->rx_opt
.num_sacks
;
451 static unsigned tcp_syn_options(struct sock
*sk
, struct sk_buff
*skb
,
452 struct tcp_out_options
*opts
,
453 struct tcp_md5sig_key
**md5
) {
454 struct tcp_sock
*tp
= tcp_sk(sk
);
457 #ifdef CONFIG_TCP_MD5SIG
458 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
460 opts
->options
|= OPTION_MD5
;
461 size
+= TCPOLEN_MD5SIG_ALIGNED
;
467 /* We always get an MSS option. The option bytes which will be seen in
468 * normal data packets should timestamps be used, must be in the MSS
469 * advertised. But we subtract them from tp->mss_cache so that
470 * calculations in tcp_sendmsg are simpler etc. So account for this
471 * fact here if necessary. If we don't do this correctly, as a
472 * receiver we won't recognize data packets as being full sized when we
473 * should, and thus we won't abide by the delayed ACK rules correctly.
474 * SACKs don't matter, we never delay an ACK when we have any of those
476 opts
->mss
= tcp_advertise_mss(sk
);
477 size
+= TCPOLEN_MSS_ALIGNED
;
479 if (likely(sysctl_tcp_timestamps
&& *md5
== NULL
)) {
480 opts
->options
|= OPTION_TS
;
481 opts
->tsval
= TCP_SKB_CB(skb
)->when
;
482 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
483 size
+= TCPOLEN_TSTAMP_ALIGNED
;
485 if (likely(sysctl_tcp_window_scaling
)) {
486 opts
->ws
= tp
->rx_opt
.rcv_wscale
;
487 if (likely(opts
->ws
))
488 size
+= TCPOLEN_WSCALE_ALIGNED
;
490 if (likely(sysctl_tcp_sack
)) {
491 opts
->options
|= OPTION_SACK_ADVERTISE
;
492 if (unlikely(!(OPTION_TS
& opts
->options
)))
493 size
+= TCPOLEN_SACKPERM_ALIGNED
;
499 static unsigned tcp_synack_options(struct sock
*sk
,
500 struct request_sock
*req
,
501 unsigned mss
, struct sk_buff
*skb
,
502 struct tcp_out_options
*opts
,
503 struct tcp_md5sig_key
**md5
) {
505 struct inet_request_sock
*ireq
= inet_rsk(req
);
508 #ifdef CONFIG_TCP_MD5SIG
509 *md5
= tcp_rsk(req
)->af_specific
->md5_lookup(sk
, req
);
511 opts
->options
|= OPTION_MD5
;
512 size
+= TCPOLEN_MD5SIG_ALIGNED
;
518 /* we can't fit any SACK blocks in a packet with MD5 + TS
519 options. There was discussion about disabling SACK rather than TS in
520 order to fit in better with old, buggy kernels, but that was deemed
521 to be unnecessary. */
522 doing_ts
= ireq
->tstamp_ok
&& !(*md5
&& ireq
->sack_ok
);
525 size
+= TCPOLEN_MSS_ALIGNED
;
527 if (likely(ireq
->wscale_ok
)) {
528 opts
->ws
= ireq
->rcv_wscale
;
529 if (likely(opts
->ws
))
530 size
+= TCPOLEN_WSCALE_ALIGNED
;
532 if (likely(doing_ts
)) {
533 opts
->options
|= OPTION_TS
;
534 opts
->tsval
= TCP_SKB_CB(skb
)->when
;
535 opts
->tsecr
= req
->ts_recent
;
536 size
+= TCPOLEN_TSTAMP_ALIGNED
;
538 if (likely(ireq
->sack_ok
)) {
539 opts
->options
|= OPTION_SACK_ADVERTISE
;
540 if (unlikely(!doing_ts
))
541 size
+= TCPOLEN_SACKPERM_ALIGNED
;
547 static unsigned tcp_established_options(struct sock
*sk
, struct sk_buff
*skb
,
548 struct tcp_out_options
*opts
,
549 struct tcp_md5sig_key
**md5
) {
550 struct tcp_skb_cb
*tcb
= skb
? TCP_SKB_CB(skb
) : NULL
;
551 struct tcp_sock
*tp
= tcp_sk(sk
);
554 #ifdef CONFIG_TCP_MD5SIG
555 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
556 if (unlikely(*md5
)) {
557 opts
->options
|= OPTION_MD5
;
558 size
+= TCPOLEN_MD5SIG_ALIGNED
;
564 if (likely(tp
->rx_opt
.tstamp_ok
)) {
565 opts
->options
|= OPTION_TS
;
566 opts
->tsval
= tcb
? tcb
->when
: 0;
567 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
568 size
+= TCPOLEN_TSTAMP_ALIGNED
;
571 if (unlikely(tp
->rx_opt
.eff_sacks
)) {
572 const unsigned remaining
= MAX_TCP_OPTION_SPACE
- size
;
573 opts
->num_sack_blocks
=
574 min_t(unsigned, tp
->rx_opt
.eff_sacks
,
575 (remaining
- TCPOLEN_SACK_BASE_ALIGNED
) /
576 TCPOLEN_SACK_PERBLOCK
);
577 size
+= TCPOLEN_SACK_BASE_ALIGNED
+
578 opts
->num_sack_blocks
* TCPOLEN_SACK_PERBLOCK
;
584 /* This routine actually transmits TCP packets queued in by
585 * tcp_do_sendmsg(). This is used by both the initial
586 * transmission and possible later retransmissions.
587 * All SKB's seen here are completely headerless. It is our
588 * job to build the TCP header, and pass the packet down to
589 * IP so it can do the same plus pass the packet off to the
592 * We are working here with either a clone of the original
593 * SKB, or a fresh unique copy made by the retransmit engine.
595 static int tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int clone_it
,
598 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
599 struct inet_sock
*inet
;
601 struct tcp_skb_cb
*tcb
;
602 struct tcp_out_options opts
;
603 unsigned tcp_options_size
, tcp_header_size
;
604 struct tcp_md5sig_key
*md5
;
605 __u8
*md5_hash_location
;
609 BUG_ON(!skb
|| !tcp_skb_pcount(skb
));
611 /* If congestion control is doing timestamping, we must
612 * take such a timestamp before we potentially clone/copy.
614 if (icsk
->icsk_ca_ops
->flags
& TCP_CONG_RTT_STAMP
)
615 __net_timestamp(skb
);
617 if (likely(clone_it
)) {
618 if (unlikely(skb_cloned(skb
)))
619 skb
= pskb_copy(skb
, gfp_mask
);
621 skb
= skb_clone(skb
, gfp_mask
);
628 tcb
= TCP_SKB_CB(skb
);
629 memset(&opts
, 0, sizeof(opts
));
631 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
))
632 tcp_options_size
= tcp_syn_options(sk
, skb
, &opts
, &md5
);
634 tcp_options_size
= tcp_established_options(sk
, skb
, &opts
,
636 tcp_header_size
= tcp_options_size
+ sizeof(struct tcphdr
);
638 if (tcp_packets_in_flight(tp
) == 0)
639 tcp_ca_event(sk
, CA_EVENT_TX_START
);
641 skb_push(skb
, tcp_header_size
);
642 skb_reset_transport_header(skb
);
643 skb_set_owner_w(skb
, sk
);
645 /* Build TCP header and checksum it. */
647 th
->source
= inet
->sport
;
648 th
->dest
= inet
->dport
;
649 th
->seq
= htonl(tcb
->seq
);
650 th
->ack_seq
= htonl(tp
->rcv_nxt
);
651 *(((__be16
*)th
) + 6) = htons(((tcp_header_size
>> 2) << 12) |
654 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
655 /* RFC1323: The window in SYN & SYN/ACK segments
658 th
->window
= htons(min(tp
->rcv_wnd
, 65535U));
660 th
->window
= htons(tcp_select_window(sk
));
665 /* The urg_mode check is necessary during a below snd_una win probe */
666 if (unlikely(tcp_urg_mode(tp
))) {
667 if (between(tp
->snd_up
, tcb
->seq
+ 1, tcb
->seq
+ 0xFFFF)) {
668 th
->urg_ptr
= htons(tp
->snd_up
- tcb
->seq
);
670 } else if (after(tcb
->seq
+ 0xFFFF, tp
->snd_nxt
)) {
671 th
->urg_ptr
= 0xFFFF;
676 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
, &md5_hash_location
);
677 if (likely((tcb
->flags
& TCPCB_FLAG_SYN
) == 0))
678 TCP_ECN_send(sk
, skb
, tcp_header_size
);
680 #ifdef CONFIG_TCP_MD5SIG
681 /* Calculate the MD5 hash, as we have all we need now */
683 sk
->sk_route_caps
&= ~NETIF_F_GSO_MASK
;
684 tp
->af_specific
->calc_md5_hash(md5_hash_location
,
689 icsk
->icsk_af_ops
->send_check(sk
, skb
->len
, skb
);
691 if (likely(tcb
->flags
& TCPCB_FLAG_ACK
))
692 tcp_event_ack_sent(sk
, tcp_skb_pcount(skb
));
694 if (skb
->len
!= tcp_header_size
)
695 tcp_event_data_sent(tp
, skb
, sk
);
697 if (after(tcb
->end_seq
, tp
->snd_nxt
) || tcb
->seq
== tcb
->end_seq
)
698 TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
);
700 err
= icsk
->icsk_af_ops
->queue_xmit(skb
, 0);
701 if (likely(err
<= 0))
704 tcp_enter_cwr(sk
, 1);
706 return net_xmit_eval(err
);
709 /* This routine just queue's the buffer
711 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
712 * otherwise socket can stall.
714 static void tcp_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
716 struct tcp_sock
*tp
= tcp_sk(sk
);
718 /* Advance write_seq and place onto the write_queue. */
719 tp
->write_seq
= TCP_SKB_CB(skb
)->end_seq
;
720 skb_header_release(skb
);
721 tcp_add_write_queue_tail(sk
, skb
);
722 sk
->sk_wmem_queued
+= skb
->truesize
;
723 sk_mem_charge(sk
, skb
->truesize
);
726 static void tcp_set_skb_tso_segs(struct sock
*sk
, struct sk_buff
*skb
,
727 unsigned int mss_now
)
729 if (skb
->len
<= mss_now
|| !sk_can_gso(sk
)) {
730 /* Avoid the costly divide in the normal
733 skb_shinfo(skb
)->gso_segs
= 1;
734 skb_shinfo(skb
)->gso_size
= 0;
735 skb_shinfo(skb
)->gso_type
= 0;
737 skb_shinfo(skb
)->gso_segs
= DIV_ROUND_UP(skb
->len
, mss_now
);
738 skb_shinfo(skb
)->gso_size
= mss_now
;
739 skb_shinfo(skb
)->gso_type
= sk
->sk_gso_type
;
743 /* When a modification to fackets out becomes necessary, we need to check
744 * skb is counted to fackets_out or not.
746 static void tcp_adjust_fackets_out(struct sock
*sk
, struct sk_buff
*skb
,
749 struct tcp_sock
*tp
= tcp_sk(sk
);
751 if (!tp
->sacked_out
|| tcp_is_reno(tp
))
754 if (after(tcp_highest_sack_seq(tp
), TCP_SKB_CB(skb
)->seq
))
755 tp
->fackets_out
-= decr
;
758 /* Function to create two new TCP segments. Shrinks the given segment
759 * to the specified size and appends a new segment with the rest of the
760 * packet to the list. This won't be called frequently, I hope.
761 * Remember, these are still headerless SKBs at this point.
763 int tcp_fragment(struct sock
*sk
, struct sk_buff
*skb
, u32 len
,
764 unsigned int mss_now
)
766 struct tcp_sock
*tp
= tcp_sk(sk
);
767 struct sk_buff
*buff
;
768 int nsize
, old_factor
;
772 BUG_ON(len
> skb
->len
);
774 tcp_clear_retrans_hints_partial(tp
);
775 nsize
= skb_headlen(skb
) - len
;
779 if (skb_cloned(skb
) &&
780 skb_is_nonlinear(skb
) &&
781 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
784 /* Get a new skb... force flag on. */
785 buff
= sk_stream_alloc_skb(sk
, nsize
, GFP_ATOMIC
);
787 return -ENOMEM
; /* We'll just try again later. */
789 sk
->sk_wmem_queued
+= buff
->truesize
;
790 sk_mem_charge(sk
, buff
->truesize
);
791 nlen
= skb
->len
- len
- nsize
;
792 buff
->truesize
+= nlen
;
793 skb
->truesize
-= nlen
;
795 /* Correct the sequence numbers. */
796 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
797 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
798 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
800 /* PSH and FIN should only be set in the second packet. */
801 flags
= TCP_SKB_CB(skb
)->flags
;
802 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
| TCPCB_FLAG_PSH
);
803 TCP_SKB_CB(buff
)->flags
= flags
;
804 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
806 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_PARTIAL
) {
807 /* Copy and checksum data tail into the new buffer. */
808 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
,
809 skb_put(buff
, nsize
),
814 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
816 skb
->ip_summed
= CHECKSUM_PARTIAL
;
817 skb_split(skb
, buff
, len
);
820 buff
->ip_summed
= skb
->ip_summed
;
822 /* Looks stupid, but our code really uses when of
823 * skbs, which it never sent before. --ANK
825 TCP_SKB_CB(buff
)->when
= TCP_SKB_CB(skb
)->when
;
826 buff
->tstamp
= skb
->tstamp
;
828 old_factor
= tcp_skb_pcount(skb
);
830 /* Fix up tso_factor for both original and new SKB. */
831 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
832 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
834 /* If this packet has been sent out already, we must
835 * adjust the various packet counters.
837 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
838 int diff
= old_factor
- tcp_skb_pcount(skb
) -
839 tcp_skb_pcount(buff
);
841 tp
->packets_out
-= diff
;
843 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
844 tp
->sacked_out
-= diff
;
845 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
846 tp
->retrans_out
-= diff
;
848 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
)
849 tp
->lost_out
-= diff
;
851 /* Adjust Reno SACK estimate. */
852 if (tcp_is_reno(tp
) && diff
> 0) {
853 tcp_dec_pcount_approx_int(&tp
->sacked_out
, diff
);
854 tcp_verify_left_out(tp
);
856 tcp_adjust_fackets_out(sk
, skb
, diff
);
859 /* Link BUFF into the send queue. */
860 skb_header_release(buff
);
861 tcp_insert_write_queue_after(skb
, buff
, sk
);
866 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
867 * eventually). The difference is that pulled data not copied, but
868 * immediately discarded.
870 static void __pskb_trim_head(struct sk_buff
*skb
, int len
)
876 for (i
= 0; i
< skb_shinfo(skb
)->nr_frags
; i
++) {
877 if (skb_shinfo(skb
)->frags
[i
].size
<= eat
) {
878 put_page(skb_shinfo(skb
)->frags
[i
].page
);
879 eat
-= skb_shinfo(skb
)->frags
[i
].size
;
881 skb_shinfo(skb
)->frags
[k
] = skb_shinfo(skb
)->frags
[i
];
883 skb_shinfo(skb
)->frags
[k
].page_offset
+= eat
;
884 skb_shinfo(skb
)->frags
[k
].size
-= eat
;
890 skb_shinfo(skb
)->nr_frags
= k
;
892 skb_reset_tail_pointer(skb
);
893 skb
->data_len
-= len
;
894 skb
->len
= skb
->data_len
;
897 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
899 if (skb_cloned(skb
) && pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
902 /* If len == headlen, we avoid __skb_pull to preserve alignment. */
903 if (unlikely(len
< skb_headlen(skb
)))
904 __skb_pull(skb
, len
);
906 __pskb_trim_head(skb
, len
- skb_headlen(skb
));
908 TCP_SKB_CB(skb
)->seq
+= len
;
909 skb
->ip_summed
= CHECKSUM_PARTIAL
;
911 skb
->truesize
-= len
;
912 sk
->sk_wmem_queued
-= len
;
913 sk_mem_uncharge(sk
, len
);
914 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
916 /* Any change of skb->len requires recalculation of tso
919 if (tcp_skb_pcount(skb
) > 1)
920 tcp_set_skb_tso_segs(sk
, skb
, tcp_current_mss(sk
, 1));
925 /* Not accounting for SACKs here. */
926 int tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
928 struct tcp_sock
*tp
= tcp_sk(sk
);
929 struct inet_connection_sock
*icsk
= inet_csk(sk
);
932 /* Calculate base mss without TCP options:
933 It is MMS_S - sizeof(tcphdr) of rfc1122
935 mss_now
= pmtu
- icsk
->icsk_af_ops
->net_header_len
- sizeof(struct tcphdr
);
937 /* Clamp it (mss_clamp does not include tcp options) */
938 if (mss_now
> tp
->rx_opt
.mss_clamp
)
939 mss_now
= tp
->rx_opt
.mss_clamp
;
941 /* Now subtract optional transport overhead */
942 mss_now
-= icsk
->icsk_ext_hdr_len
;
944 /* Then reserve room for full set of TCP options and 8 bytes of data */
948 /* Now subtract TCP options size, not including SACKs */
949 mss_now
-= tp
->tcp_header_len
- sizeof(struct tcphdr
);
954 /* Inverse of above */
955 int tcp_mss_to_mtu(struct sock
*sk
, int mss
)
957 struct tcp_sock
*tp
= tcp_sk(sk
);
958 struct inet_connection_sock
*icsk
= inet_csk(sk
);
963 icsk
->icsk_ext_hdr_len
+
964 icsk
->icsk_af_ops
->net_header_len
;
969 void tcp_mtup_init(struct sock
*sk
)
971 struct tcp_sock
*tp
= tcp_sk(sk
);
972 struct inet_connection_sock
*icsk
= inet_csk(sk
);
974 icsk
->icsk_mtup
.enabled
= sysctl_tcp_mtu_probing
> 1;
975 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+ sizeof(struct tcphdr
) +
976 icsk
->icsk_af_ops
->net_header_len
;
977 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, sysctl_tcp_base_mss
);
978 icsk
->icsk_mtup
.probe_size
= 0;
981 /* Bound MSS / TSO packet size with the half of the window */
982 static int tcp_bound_to_half_wnd(struct tcp_sock
*tp
, int pktsize
)
984 if (tp
->max_window
&& pktsize
> (tp
->max_window
>> 1))
985 return max(tp
->max_window
>> 1, 68U - tp
->tcp_header_len
);
990 /* This function synchronize snd mss to current pmtu/exthdr set.
992 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
993 for TCP options, but includes only bare TCP header.
995 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
996 It is minimum of user_mss and mss received with SYN.
997 It also does not include TCP options.
999 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1001 tp->mss_cache is current effective sending mss, including
1002 all tcp options except for SACKs. It is evaluated,
1003 taking into account current pmtu, but never exceeds
1004 tp->rx_opt.mss_clamp.
1006 NOTE1. rfc1122 clearly states that advertised MSS
1007 DOES NOT include either tcp or ip options.
1009 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1010 are READ ONLY outside this function. --ANK (980731)
1012 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
1014 struct tcp_sock
*tp
= tcp_sk(sk
);
1015 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1018 if (icsk
->icsk_mtup
.search_high
> pmtu
)
1019 icsk
->icsk_mtup
.search_high
= pmtu
;
1021 mss_now
= tcp_mtu_to_mss(sk
, pmtu
);
1022 mss_now
= tcp_bound_to_half_wnd(tp
, mss_now
);
1024 /* And store cached results */
1025 icsk
->icsk_pmtu_cookie
= pmtu
;
1026 if (icsk
->icsk_mtup
.enabled
)
1027 mss_now
= min(mss_now
, tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
));
1028 tp
->mss_cache
= mss_now
;
1033 /* Compute the current effective MSS, taking SACKs and IP options,
1034 * and even PMTU discovery events into account.
1036 unsigned int tcp_current_mss(struct sock
*sk
, int large_allowed
)
1038 struct tcp_sock
*tp
= tcp_sk(sk
);
1039 struct dst_entry
*dst
= __sk_dst_get(sk
);
1043 unsigned header_len
;
1044 struct tcp_out_options opts
;
1045 struct tcp_md5sig_key
*md5
;
1047 mss_now
= tp
->mss_cache
;
1049 if (large_allowed
&& sk_can_gso(sk
))
1053 u32 mtu
= dst_mtu(dst
);
1054 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
1055 mss_now
= tcp_sync_mss(sk
, mtu
);
1058 header_len
= tcp_established_options(sk
, NULL
, &opts
, &md5
) +
1059 sizeof(struct tcphdr
);
1060 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1061 * some common options. If this is an odd packet (because we have SACK
1062 * blocks etc) then our calculated header_len will be different, and
1063 * we have to adjust mss_now correspondingly */
1064 if (header_len
!= tp
->tcp_header_len
) {
1065 int delta
= (int) header_len
- tp
->tcp_header_len
;
1069 xmit_size_goal
= mss_now
;
1072 xmit_size_goal
= ((sk
->sk_gso_max_size
- 1) -
1073 inet_csk(sk
)->icsk_af_ops
->net_header_len
-
1074 inet_csk(sk
)->icsk_ext_hdr_len
-
1075 tp
->tcp_header_len
);
1077 xmit_size_goal
= tcp_bound_to_half_wnd(tp
, xmit_size_goal
);
1078 xmit_size_goal
-= (xmit_size_goal
% mss_now
);
1080 tp
->xmit_size_goal
= xmit_size_goal
;
1085 /* Congestion window validation. (RFC2861) */
1086 static void tcp_cwnd_validate(struct sock
*sk
)
1088 struct tcp_sock
*tp
= tcp_sk(sk
);
1090 if (tp
->packets_out
>= tp
->snd_cwnd
) {
1091 /* Network is feed fully. */
1092 tp
->snd_cwnd_used
= 0;
1093 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1095 /* Network starves. */
1096 if (tp
->packets_out
> tp
->snd_cwnd_used
)
1097 tp
->snd_cwnd_used
= tp
->packets_out
;
1099 if (sysctl_tcp_slow_start_after_idle
&&
1100 (s32
)(tcp_time_stamp
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
)
1101 tcp_cwnd_application_limited(sk
);
1105 /* Returns the portion of skb which can be sent right away without
1106 * introducing MSS oddities to segment boundaries. In rare cases where
1107 * mss_now != mss_cache, we will request caller to create a small skb
1108 * per input skb which could be mostly avoided here (if desired).
1110 * We explicitly want to create a request for splitting write queue tail
1111 * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1112 * thus all the complexity (cwnd_len is always MSS multiple which we
1113 * return whenever allowed by the other factors). Basically we need the
1114 * modulo only when the receiver window alone is the limiting factor or
1115 * when we would be allowed to send the split-due-to-Nagle skb fully.
1117 static unsigned int tcp_mss_split_point(struct sock
*sk
, struct sk_buff
*skb
,
1118 unsigned int mss_now
, unsigned int cwnd
)
1120 struct tcp_sock
*tp
= tcp_sk(sk
);
1121 u32 needed
, window
, cwnd_len
;
1123 window
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1124 cwnd_len
= mss_now
* cwnd
;
1126 if (likely(cwnd_len
<= window
&& skb
!= tcp_write_queue_tail(sk
)))
1129 needed
= min(skb
->len
, window
);
1131 if (cwnd_len
<= needed
)
1134 return needed
- needed
% mss_now
;
1137 /* Can at least one segment of SKB be sent right now, according to the
1138 * congestion window rules? If so, return how many segments are allowed.
1140 static inline unsigned int tcp_cwnd_test(struct tcp_sock
*tp
,
1141 struct sk_buff
*skb
)
1143 u32 in_flight
, cwnd
;
1145 /* Don't be strict about the congestion window for the final FIN. */
1146 if ((TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
) &&
1147 tcp_skb_pcount(skb
) == 1)
1150 in_flight
= tcp_packets_in_flight(tp
);
1151 cwnd
= tp
->snd_cwnd
;
1152 if (in_flight
< cwnd
)
1153 return (cwnd
- in_flight
);
1158 /* This must be invoked the first time we consider transmitting
1159 * SKB onto the wire.
1161 static int tcp_init_tso_segs(struct sock
*sk
, struct sk_buff
*skb
,
1162 unsigned int mss_now
)
1164 int tso_segs
= tcp_skb_pcount(skb
);
1166 if (!tso_segs
|| (tso_segs
> 1 && tcp_skb_mss(skb
) != mss_now
)) {
1167 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1168 tso_segs
= tcp_skb_pcount(skb
);
1173 static inline int tcp_minshall_check(const struct tcp_sock
*tp
)
1175 return after(tp
->snd_sml
, tp
->snd_una
) &&
1176 !after(tp
->snd_sml
, tp
->snd_nxt
);
1179 /* Return 0, if packet can be sent now without violation Nagle's rules:
1180 * 1. It is full sized.
1181 * 2. Or it contains FIN. (already checked by caller)
1182 * 3. Or TCP_NODELAY was set.
1183 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1184 * With Minshall's modification: all sent small packets are ACKed.
1186 static inline int tcp_nagle_check(const struct tcp_sock
*tp
,
1187 const struct sk_buff
*skb
,
1188 unsigned mss_now
, int nonagle
)
1190 return (skb
->len
< mss_now
&&
1191 ((nonagle
& TCP_NAGLE_CORK
) ||
1192 (!nonagle
&& tp
->packets_out
&& tcp_minshall_check(tp
))));
1195 /* Return non-zero if the Nagle test allows this packet to be
1198 static inline int tcp_nagle_test(struct tcp_sock
*tp
, struct sk_buff
*skb
,
1199 unsigned int cur_mss
, int nonagle
)
1201 /* Nagle rule does not apply to frames, which sit in the middle of the
1202 * write_queue (they have no chances to get new data).
1204 * This is implemented in the callers, where they modify the 'nonagle'
1205 * argument based upon the location of SKB in the send queue.
1207 if (nonagle
& TCP_NAGLE_PUSH
)
1210 /* Don't use the nagle rule for urgent data (or for the final FIN).
1211 * Nagle can be ignored during F-RTO too (see RFC4138).
1213 if (tcp_urg_mode(tp
) || (tp
->frto_counter
== 2) ||
1214 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
))
1217 if (!tcp_nagle_check(tp
, skb
, cur_mss
, nonagle
))
1223 /* Does at least the first segment of SKB fit into the send window? */
1224 static inline int tcp_snd_wnd_test(struct tcp_sock
*tp
, struct sk_buff
*skb
,
1225 unsigned int cur_mss
)
1227 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1229 if (skb
->len
> cur_mss
)
1230 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
1232 return !after(end_seq
, tcp_wnd_end(tp
));
1235 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1236 * should be put on the wire right now. If so, it returns the number of
1237 * packets allowed by the congestion window.
1239 static unsigned int tcp_snd_test(struct sock
*sk
, struct sk_buff
*skb
,
1240 unsigned int cur_mss
, int nonagle
)
1242 struct tcp_sock
*tp
= tcp_sk(sk
);
1243 unsigned int cwnd_quota
;
1245 tcp_init_tso_segs(sk
, skb
, cur_mss
);
1247 if (!tcp_nagle_test(tp
, skb
, cur_mss
, nonagle
))
1250 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1251 if (cwnd_quota
&& !tcp_snd_wnd_test(tp
, skb
, cur_mss
))
1257 int tcp_may_send_now(struct sock
*sk
)
1259 struct tcp_sock
*tp
= tcp_sk(sk
);
1260 struct sk_buff
*skb
= tcp_send_head(sk
);
1263 tcp_snd_test(sk
, skb
, tcp_current_mss(sk
, 1),
1264 (tcp_skb_is_last(sk
, skb
) ?
1265 tp
->nonagle
: TCP_NAGLE_PUSH
)));
1268 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1269 * which is put after SKB on the list. It is very much like
1270 * tcp_fragment() except that it may make several kinds of assumptions
1271 * in order to speed up the splitting operation. In particular, we
1272 * know that all the data is in scatter-gather pages, and that the
1273 * packet has never been sent out before (and thus is not cloned).
1275 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
,
1276 unsigned int mss_now
)
1278 struct sk_buff
*buff
;
1279 int nlen
= skb
->len
- len
;
1282 /* All of a TSO frame must be composed of paged data. */
1283 if (skb
->len
!= skb
->data_len
)
1284 return tcp_fragment(sk
, skb
, len
, mss_now
);
1286 buff
= sk_stream_alloc_skb(sk
, 0, GFP_ATOMIC
);
1287 if (unlikely(buff
== NULL
))
1290 sk
->sk_wmem_queued
+= buff
->truesize
;
1291 sk_mem_charge(sk
, buff
->truesize
);
1292 buff
->truesize
+= nlen
;
1293 skb
->truesize
-= nlen
;
1295 /* Correct the sequence numbers. */
1296 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1297 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1298 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1300 /* PSH and FIN should only be set in the second packet. */
1301 flags
= TCP_SKB_CB(skb
)->flags
;
1302 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
| TCPCB_FLAG_PSH
);
1303 TCP_SKB_CB(buff
)->flags
= flags
;
1305 /* This packet was never sent out yet, so no SACK bits. */
1306 TCP_SKB_CB(buff
)->sacked
= 0;
1308 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_PARTIAL
;
1309 skb_split(skb
, buff
, len
);
1311 /* Fix up tso_factor for both original and new SKB. */
1312 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1313 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1315 /* Link BUFF into the send queue. */
1316 skb_header_release(buff
);
1317 tcp_insert_write_queue_after(skb
, buff
, sk
);
1322 /* Try to defer sending, if possible, in order to minimize the amount
1323 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1325 * This algorithm is from John Heffner.
1327 static int tcp_tso_should_defer(struct sock
*sk
, struct sk_buff
*skb
)
1329 struct tcp_sock
*tp
= tcp_sk(sk
);
1330 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1331 u32 send_win
, cong_win
, limit
, in_flight
;
1333 if (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
)
1336 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
1339 /* Defer for less than two clock ticks. */
1340 if (tp
->tso_deferred
&&
1341 (((u32
)jiffies
<< 1) >> 1) - (tp
->tso_deferred
>> 1) > 1)
1344 in_flight
= tcp_packets_in_flight(tp
);
1346 BUG_ON(tcp_skb_pcount(skb
) <= 1 || (tp
->snd_cwnd
<= in_flight
));
1348 send_win
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1350 /* From in_flight test above, we know that cwnd > in_flight. */
1351 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
1353 limit
= min(send_win
, cong_win
);
1355 /* If a full-sized TSO skb can be sent, do it. */
1356 if (limit
>= sk
->sk_gso_max_size
)
1359 if (sysctl_tcp_tso_win_divisor
) {
1360 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
1362 /* If at least some fraction of a window is available,
1365 chunk
/= sysctl_tcp_tso_win_divisor
;
1369 /* Different approach, try not to defer past a single
1370 * ACK. Receiver should ACK every other full sized
1371 * frame, so if we have space for more than 3 frames
1374 if (limit
> tcp_max_burst(tp
) * tp
->mss_cache
)
1378 /* Ok, it looks like it is advisable to defer. */
1379 tp
->tso_deferred
= 1 | (jiffies
<< 1);
1384 tp
->tso_deferred
= 0;
1388 /* Create a new MTU probe if we are ready.
1389 * Returns 0 if we should wait to probe (no cwnd available),
1390 * 1 if a probe was sent,
1393 static int tcp_mtu_probe(struct sock
*sk
)
1395 struct tcp_sock
*tp
= tcp_sk(sk
);
1396 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1397 struct sk_buff
*skb
, *nskb
, *next
;
1404 /* Not currently probing/verifying,
1406 * have enough cwnd, and
1407 * not SACKing (the variable headers throw things off) */
1408 if (!icsk
->icsk_mtup
.enabled
||
1409 icsk
->icsk_mtup
.probe_size
||
1410 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
||
1411 tp
->snd_cwnd
< 11 ||
1412 tp
->rx_opt
.eff_sacks
)
1415 /* Very simple search strategy: just double the MSS. */
1416 mss_now
= tcp_current_mss(sk
, 0);
1417 probe_size
= 2 * tp
->mss_cache
;
1418 size_needed
= probe_size
+ (tp
->reordering
+ 1) * tp
->mss_cache
;
1419 if (probe_size
> tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_high
)) {
1420 /* TODO: set timer for probe_converge_event */
1424 /* Have enough data in the send queue to probe? */
1425 if (tp
->write_seq
- tp
->snd_nxt
< size_needed
)
1428 if (tp
->snd_wnd
< size_needed
)
1430 if (after(tp
->snd_nxt
+ size_needed
, tcp_wnd_end(tp
)))
1433 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1434 if (tcp_packets_in_flight(tp
) + 2 > tp
->snd_cwnd
) {
1435 if (!tcp_packets_in_flight(tp
))
1441 /* We're allowed to probe. Build it now. */
1442 if ((nskb
= sk_stream_alloc_skb(sk
, probe_size
, GFP_ATOMIC
)) == NULL
)
1444 sk
->sk_wmem_queued
+= nskb
->truesize
;
1445 sk_mem_charge(sk
, nskb
->truesize
);
1447 skb
= tcp_send_head(sk
);
1449 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(skb
)->seq
;
1450 TCP_SKB_CB(nskb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ probe_size
;
1451 TCP_SKB_CB(nskb
)->flags
= TCPCB_FLAG_ACK
;
1452 TCP_SKB_CB(nskb
)->sacked
= 0;
1454 nskb
->ip_summed
= skb
->ip_summed
;
1456 tcp_insert_write_queue_before(nskb
, skb
, sk
);
1459 tcp_for_write_queue_from_safe(skb
, next
, sk
) {
1460 copy
= min_t(int, skb
->len
, probe_size
- len
);
1461 if (nskb
->ip_summed
)
1462 skb_copy_bits(skb
, 0, skb_put(nskb
, copy
), copy
);
1464 nskb
->csum
= skb_copy_and_csum_bits(skb
, 0,
1465 skb_put(nskb
, copy
),
1468 if (skb
->len
<= copy
) {
1469 /* We've eaten all the data from this skb.
1471 TCP_SKB_CB(nskb
)->flags
|= TCP_SKB_CB(skb
)->flags
;
1472 tcp_unlink_write_queue(skb
, sk
);
1473 sk_wmem_free_skb(sk
, skb
);
1475 TCP_SKB_CB(nskb
)->flags
|= TCP_SKB_CB(skb
)->flags
&
1476 ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
1477 if (!skb_shinfo(skb
)->nr_frags
) {
1478 skb_pull(skb
, copy
);
1479 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1480 skb
->csum
= csum_partial(skb
->data
,
1483 __pskb_trim_head(skb
, copy
);
1484 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1486 TCP_SKB_CB(skb
)->seq
+= copy
;
1491 if (len
>= probe_size
)
1494 tcp_init_tso_segs(sk
, nskb
, nskb
->len
);
1496 /* We're ready to send. If this fails, the probe will
1497 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1498 TCP_SKB_CB(nskb
)->when
= tcp_time_stamp
;
1499 if (!tcp_transmit_skb(sk
, nskb
, 1, GFP_ATOMIC
)) {
1500 /* Decrement cwnd here because we are sending
1501 * effectively two packets. */
1503 tcp_event_new_data_sent(sk
, nskb
);
1505 icsk
->icsk_mtup
.probe_size
= tcp_mss_to_mtu(sk
, nskb
->len
);
1506 tp
->mtu_probe
.probe_seq_start
= TCP_SKB_CB(nskb
)->seq
;
1507 tp
->mtu_probe
.probe_seq_end
= TCP_SKB_CB(nskb
)->end_seq
;
1515 /* This routine writes packets to the network. It advances the
1516 * send_head. This happens as incoming acks open up the remote
1519 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1520 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1521 * account rare use of URG, this is not a big flaw.
1523 * Returns 1, if no segments are in flight and we have queued segments, but
1524 * cannot send anything now because of SWS or another problem.
1526 static int tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
,
1527 int push_one
, gfp_t gfp
)
1529 struct tcp_sock
*tp
= tcp_sk(sk
);
1530 struct sk_buff
*skb
;
1531 unsigned int tso_segs
, sent_pkts
;
1538 /* Do MTU probing. */
1539 result
= tcp_mtu_probe(sk
);
1542 } else if (result
> 0) {
1547 while ((skb
= tcp_send_head(sk
))) {
1550 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1553 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1557 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
)))
1560 if (tso_segs
== 1) {
1561 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
1562 (tcp_skb_is_last(sk
, skb
) ?
1563 nonagle
: TCP_NAGLE_PUSH
))))
1566 if (!push_one
&& tcp_tso_should_defer(sk
, skb
))
1571 if (tso_segs
> 1 && !tcp_urg_mode(tp
))
1572 limit
= tcp_mss_split_point(sk
, skb
, mss_now
,
1575 if (skb
->len
> limit
&&
1576 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
)))
1579 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1581 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, gfp
)))
1584 /* Advance the send_head. This one is sent out.
1585 * This call will increment packets_out.
1587 tcp_event_new_data_sent(sk
, skb
);
1589 tcp_minshall_update(tp
, mss_now
, skb
);
1596 if (likely(sent_pkts
)) {
1597 tcp_cwnd_validate(sk
);
1600 return !tp
->packets_out
&& tcp_send_head(sk
);
1603 /* Push out any pending frames which were held back due to
1604 * TCP_CORK or attempt at coalescing tiny packets.
1605 * The socket must be locked by the caller.
1607 void __tcp_push_pending_frames(struct sock
*sk
, unsigned int cur_mss
,
1610 struct sk_buff
*skb
= tcp_send_head(sk
);
1615 /* If we are closed, the bytes will have to remain here.
1616 * In time closedown will finish, we empty the write queue and
1617 * all will be happy.
1619 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
1622 if (tcp_write_xmit(sk
, cur_mss
, nonagle
, 0, GFP_ATOMIC
))
1623 tcp_check_probe_timer(sk
);
1626 /* Send _single_ skb sitting at the send head. This function requires
1627 * true push pending frames to setup probe timer etc.
1629 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
1631 struct sk_buff
*skb
= tcp_send_head(sk
);
1633 BUG_ON(!skb
|| skb
->len
< mss_now
);
1635 tcp_write_xmit(sk
, mss_now
, TCP_NAGLE_PUSH
, 1, sk
->sk_allocation
);
1638 /* This function returns the amount that we can raise the
1639 * usable window based on the following constraints
1641 * 1. The window can never be shrunk once it is offered (RFC 793)
1642 * 2. We limit memory per socket
1645 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1646 * RECV.NEXT + RCV.WIN fixed until:
1647 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1649 * i.e. don't raise the right edge of the window until you can raise
1650 * it at least MSS bytes.
1652 * Unfortunately, the recommended algorithm breaks header prediction,
1653 * since header prediction assumes th->window stays fixed.
1655 * Strictly speaking, keeping th->window fixed violates the receiver
1656 * side SWS prevention criteria. The problem is that under this rule
1657 * a stream of single byte packets will cause the right side of the
1658 * window to always advance by a single byte.
1660 * Of course, if the sender implements sender side SWS prevention
1661 * then this will not be a problem.
1663 * BSD seems to make the following compromise:
1665 * If the free space is less than the 1/4 of the maximum
1666 * space available and the free space is less than 1/2 mss,
1667 * then set the window to 0.
1668 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1669 * Otherwise, just prevent the window from shrinking
1670 * and from being larger than the largest representable value.
1672 * This prevents incremental opening of the window in the regime
1673 * where TCP is limited by the speed of the reader side taking
1674 * data out of the TCP receive queue. It does nothing about
1675 * those cases where the window is constrained on the sender side
1676 * because the pipeline is full.
1678 * BSD also seems to "accidentally" limit itself to windows that are a
1679 * multiple of MSS, at least until the free space gets quite small.
1680 * This would appear to be a side effect of the mbuf implementation.
1681 * Combining these two algorithms results in the observed behavior
1682 * of having a fixed window size at almost all times.
1684 * Below we obtain similar behavior by forcing the offered window to
1685 * a multiple of the mss when it is feasible to do so.
1687 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1688 * Regular options like TIMESTAMP are taken into account.
1690 u32
__tcp_select_window(struct sock
*sk
)
1692 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1693 struct tcp_sock
*tp
= tcp_sk(sk
);
1694 /* MSS for the peer's data. Previous versions used mss_clamp
1695 * here. I don't know if the value based on our guesses
1696 * of peer's MSS is better for the performance. It's more correct
1697 * but may be worse for the performance because of rcv_mss
1698 * fluctuations. --SAW 1998/11/1
1700 int mss
= icsk
->icsk_ack
.rcv_mss
;
1701 int free_space
= tcp_space(sk
);
1702 int full_space
= min_t(int, tp
->window_clamp
, tcp_full_space(sk
));
1705 if (mss
> full_space
)
1708 if (free_space
< (full_space
>> 1)) {
1709 icsk
->icsk_ack
.quick
= 0;
1711 if (tcp_memory_pressure
)
1712 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
,
1715 if (free_space
< mss
)
1719 if (free_space
> tp
->rcv_ssthresh
)
1720 free_space
= tp
->rcv_ssthresh
;
1722 /* Don't do rounding if we are using window scaling, since the
1723 * scaled window will not line up with the MSS boundary anyway.
1725 window
= tp
->rcv_wnd
;
1726 if (tp
->rx_opt
.rcv_wscale
) {
1727 window
= free_space
;
1729 /* Advertise enough space so that it won't get scaled away.
1730 * Import case: prevent zero window announcement if
1731 * 1<<rcv_wscale > mss.
1733 if (((window
>> tp
->rx_opt
.rcv_wscale
) << tp
->rx_opt
.rcv_wscale
) != window
)
1734 window
= (((window
>> tp
->rx_opt
.rcv_wscale
) + 1)
1735 << tp
->rx_opt
.rcv_wscale
);
1737 /* Get the largest window that is a nice multiple of mss.
1738 * Window clamp already applied above.
1739 * If our current window offering is within 1 mss of the
1740 * free space we just keep it. This prevents the divide
1741 * and multiply from happening most of the time.
1742 * We also don't do any window rounding when the free space
1745 if (window
<= free_space
- mss
|| window
> free_space
)
1746 window
= (free_space
/ mss
) * mss
;
1747 else if (mss
== full_space
&&
1748 free_space
> window
+ (full_space
>> 1))
1749 window
= free_space
;
1755 /* Collapses two adjacent SKB's during retransmission. */
1756 static void tcp_collapse_retrans(struct sock
*sk
, struct sk_buff
*skb
)
1758 struct tcp_sock
*tp
= tcp_sk(sk
);
1759 struct sk_buff
*next_skb
= tcp_write_queue_next(sk
, skb
);
1760 int skb_size
, next_skb_size
;
1763 skb_size
= skb
->len
;
1764 next_skb_size
= next_skb
->len
;
1765 flags
= TCP_SKB_CB(skb
)->flags
;
1767 BUG_ON(tcp_skb_pcount(skb
) != 1 || tcp_skb_pcount(next_skb
) != 1);
1769 tcp_highest_sack_combine(sk
, next_skb
, skb
);
1771 tcp_unlink_write_queue(next_skb
, sk
);
1773 skb_copy_from_linear_data(next_skb
, skb_put(skb
, next_skb_size
),
1776 if (next_skb
->ip_summed
== CHECKSUM_PARTIAL
)
1777 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1779 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1780 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
1782 /* Update sequence range on original skb. */
1783 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
1785 /* Merge over control information. */
1786 flags
|= TCP_SKB_CB(next_skb
)->flags
; /* This moves PSH/FIN etc. over */
1787 TCP_SKB_CB(skb
)->flags
= flags
;
1789 /* All done, get rid of second SKB and account for it so
1790 * packet counting does not break.
1792 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
& TCPCB_EVER_RETRANS
;
1793 if (TCP_SKB_CB(next_skb
)->sacked
& TCPCB_SACKED_RETRANS
)
1794 tp
->retrans_out
-= tcp_skb_pcount(next_skb
);
1795 if (TCP_SKB_CB(next_skb
)->sacked
& TCPCB_LOST
)
1796 tp
->lost_out
-= tcp_skb_pcount(next_skb
);
1797 /* Reno case is special. Sigh... */
1798 if (tcp_is_reno(tp
) && tp
->sacked_out
)
1799 tcp_dec_pcount_approx(&tp
->sacked_out
, next_skb
);
1801 tcp_adjust_fackets_out(sk
, next_skb
, tcp_skb_pcount(next_skb
));
1802 tp
->packets_out
-= tcp_skb_pcount(next_skb
);
1804 /* changed transmit queue under us so clear hints */
1805 tcp_clear_retrans_hints_partial(tp
);
1806 if (next_skb
== tp
->retransmit_skb_hint
)
1807 tp
->retransmit_skb_hint
= skb
;
1809 sk_wmem_free_skb(sk
, next_skb
);
1812 static int tcp_can_collapse(struct sock
*sk
, struct sk_buff
*skb
)
1814 if (tcp_skb_pcount(skb
) > 1)
1816 /* TODO: SACK collapsing could be used to remove this condition */
1817 if (skb_shinfo(skb
)->nr_frags
!= 0)
1819 if (skb_cloned(skb
))
1821 if (skb
== tcp_send_head(sk
))
1823 /* Some heurestics for collapsing over SACK'd could be invented */
1824 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
1830 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*to
,
1833 struct tcp_sock
*tp
= tcp_sk(sk
);
1834 struct sk_buff
*skb
= to
, *tmp
;
1837 if (!sysctl_tcp_retrans_collapse
)
1839 if (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_SYN
)
1842 tcp_for_write_queue_from_safe(skb
, tmp
, sk
) {
1843 if (!tcp_can_collapse(sk
, skb
))
1855 /* Punt if not enough space exists in the first SKB for
1856 * the data in the second
1858 if (skb
->len
> skb_tailroom(to
))
1861 if (after(TCP_SKB_CB(skb
)->end_seq
, tcp_wnd_end(tp
)))
1864 tcp_collapse_retrans(sk
, to
);
1868 /* This retransmits one SKB. Policy decisions and retransmit queue
1869 * state updates are done by the caller. Returns non-zero if an
1870 * error occurred which prevented the send.
1872 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
1874 struct tcp_sock
*tp
= tcp_sk(sk
);
1875 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1876 unsigned int cur_mss
;
1879 /* Inconslusive MTU probe */
1880 if (icsk
->icsk_mtup
.probe_size
) {
1881 icsk
->icsk_mtup
.probe_size
= 0;
1884 /* Do not sent more than we queued. 1/4 is reserved for possible
1885 * copying overhead: fragmentation, tunneling, mangling etc.
1887 if (atomic_read(&sk
->sk_wmem_alloc
) >
1888 min(sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2), sk
->sk_sndbuf
))
1891 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
1892 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1894 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
1898 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
1899 return -EHOSTUNREACH
; /* Routing failure or similar. */
1901 cur_mss
= tcp_current_mss(sk
, 0);
1903 /* If receiver has shrunk his window, and skb is out of
1904 * new window, do not retransmit it. The exception is the
1905 * case, when window is shrunk to zero. In this case
1906 * our retransmit serves as a zero window probe.
1908 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
))
1909 && TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
1912 if (skb
->len
> cur_mss
) {
1913 if (tcp_fragment(sk
, skb
, cur_mss
, cur_mss
))
1914 return -ENOMEM
; /* We'll try again later. */
1917 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
1919 /* Some Solaris stacks overoptimize and ignore the FIN on a
1920 * retransmit when old data is attached. So strip it off
1921 * since it is cheap to do so and saves bytes on the network.
1924 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
) &&
1925 tp
->snd_una
== (TCP_SKB_CB(skb
)->end_seq
- 1)) {
1926 if (!pskb_trim(skb
, 0)) {
1927 /* Reuse, even though it does some unnecessary work */
1928 tcp_init_nondata_skb(skb
, TCP_SKB_CB(skb
)->end_seq
- 1,
1929 TCP_SKB_CB(skb
)->flags
);
1930 skb
->ip_summed
= CHECKSUM_NONE
;
1934 /* Make a copy, if the first transmission SKB clone we made
1935 * is still in somebody's hands, else make a clone.
1937 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1939 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
1942 /* Update global TCP statistics. */
1943 TCP_INC_STATS(sock_net(sk
), TCP_MIB_RETRANSSEGS
);
1945 tp
->total_retrans
++;
1947 #if FASTRETRANS_DEBUG > 0
1948 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
1949 if (net_ratelimit())
1950 printk(KERN_DEBUG
"retrans_out leaked.\n");
1953 if (!tp
->retrans_out
)
1954 tp
->lost_retrans_low
= tp
->snd_nxt
;
1955 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
1956 tp
->retrans_out
+= tcp_skb_pcount(skb
);
1958 /* Save stamp of the first retransmit. */
1959 if (!tp
->retrans_stamp
)
1960 tp
->retrans_stamp
= TCP_SKB_CB(skb
)->when
;
1964 /* snd_nxt is stored to detect loss of retransmitted segment,
1965 * see tcp_input.c tcp_sacktag_write_queue().
1967 TCP_SKB_CB(skb
)->ack_seq
= tp
->snd_nxt
;
1972 static int tcp_can_forward_retransmit(struct sock
*sk
)
1974 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1975 struct tcp_sock
*tp
= tcp_sk(sk
);
1977 /* Forward retransmissions are possible only during Recovery. */
1978 if (icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
1981 /* No forward retransmissions in Reno are possible. */
1982 if (tcp_is_reno(tp
))
1985 /* Yeah, we have to make difficult choice between forward transmission
1986 * and retransmission... Both ways have their merits...
1988 * For now we do not retransmit anything, while we have some new
1989 * segments to send. In the other cases, follow rule 3 for
1990 * NextSeg() specified in RFC3517.
1993 if (tcp_may_send_now(sk
))
1999 /* This gets called after a retransmit timeout, and the initially
2000 * retransmitted data is acknowledged. It tries to continue
2001 * resending the rest of the retransmit queue, until either
2002 * we've sent it all or the congestion window limit is reached.
2003 * If doing SACK, the first ACK which comes back for a timeout
2004 * based retransmit packet might feed us FACK information again.
2005 * If so, we use it to avoid unnecessarily retransmissions.
2007 void tcp_xmit_retransmit_queue(struct sock
*sk
)
2009 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2010 struct tcp_sock
*tp
= tcp_sk(sk
);
2011 struct sk_buff
*skb
;
2012 struct sk_buff
*hole
= NULL
;
2015 int fwd_rexmitting
= 0;
2018 tp
->retransmit_high
= tp
->snd_una
;
2020 if (tp
->retransmit_skb_hint
) {
2021 skb
= tp
->retransmit_skb_hint
;
2022 last_lost
= TCP_SKB_CB(skb
)->end_seq
;
2023 if (after(last_lost
, tp
->retransmit_high
))
2024 last_lost
= tp
->retransmit_high
;
2026 skb
= tcp_write_queue_head(sk
);
2027 last_lost
= tp
->snd_una
;
2030 /* First pass: retransmit lost packets. */
2031 tcp_for_write_queue_from(skb
, sk
) {
2032 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
2034 if (skb
== tcp_send_head(sk
))
2036 /* we could do better than to assign each time */
2038 tp
->retransmit_skb_hint
= skb
;
2040 /* Assume this retransmit will generate
2041 * only one packet for congestion window
2042 * calculation purposes. This works because
2043 * tcp_retransmit_skb() will chop up the
2044 * packet to be MSS sized and all the
2045 * packet counting works out.
2047 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
2050 if (fwd_rexmitting
) {
2052 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_highest_sack_seq(tp
)))
2054 mib_idx
= LINUX_MIB_TCPFORWARDRETRANS
;
2056 } else if (!before(TCP_SKB_CB(skb
)->seq
, tp
->retransmit_high
)) {
2057 tp
->retransmit_high
= last_lost
;
2058 if (!tcp_can_forward_retransmit(sk
))
2060 /* Backtrack if necessary to non-L'ed skb */
2068 } else if (!(sacked
& TCPCB_LOST
)) {
2069 if (hole
== NULL
&& !(sacked
& TCPCB_SACKED_RETRANS
))
2074 last_lost
= TCP_SKB_CB(skb
)->end_seq
;
2075 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
2076 mib_idx
= LINUX_MIB_TCPFASTRETRANS
;
2078 mib_idx
= LINUX_MIB_TCPSLOWSTARTRETRANS
;
2081 if (sacked
& (TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))
2084 if (tcp_retransmit_skb(sk
, skb
))
2086 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2088 if (skb
== tcp_write_queue_head(sk
))
2089 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2090 inet_csk(sk
)->icsk_rto
,
2095 /* Send a fin. The caller locks the socket for us. This cannot be
2096 * allowed to fail queueing a FIN frame under any circumstances.
2098 void tcp_send_fin(struct sock
*sk
)
2100 struct tcp_sock
*tp
= tcp_sk(sk
);
2101 struct sk_buff
*skb
= tcp_write_queue_tail(sk
);
2104 /* Optimization, tack on the FIN if we have a queue of
2105 * unsent frames. But be careful about outgoing SACKS
2108 mss_now
= tcp_current_mss(sk
, 1);
2110 if (tcp_send_head(sk
) != NULL
) {
2111 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_FIN
;
2112 TCP_SKB_CB(skb
)->end_seq
++;
2115 /* Socket is locked, keep trying until memory is available. */
2117 skb
= alloc_skb_fclone(MAX_TCP_HEADER
, GFP_KERNEL
);
2123 /* Reserve space for headers and prepare control bits. */
2124 skb_reserve(skb
, MAX_TCP_HEADER
);
2125 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2126 tcp_init_nondata_skb(skb
, tp
->write_seq
,
2127 TCPCB_FLAG_ACK
| TCPCB_FLAG_FIN
);
2128 tcp_queue_skb(sk
, skb
);
2130 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_OFF
);
2133 /* We get here when a process closes a file descriptor (either due to
2134 * an explicit close() or as a byproduct of exit()'ing) and there
2135 * was unread data in the receive queue. This behavior is recommended
2136 * by RFC 2525, section 2.17. -DaveM
2138 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
2140 struct sk_buff
*skb
;
2142 /* NOTE: No TCP options attached and we never retransmit this. */
2143 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
2145 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
2149 /* Reserve space for headers and prepare control bits. */
2150 skb_reserve(skb
, MAX_TCP_HEADER
);
2151 tcp_init_nondata_skb(skb
, tcp_acceptable_seq(sk
),
2152 TCPCB_FLAG_ACK
| TCPCB_FLAG_RST
);
2154 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2155 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
2156 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
2158 TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTRSTS
);
2161 /* WARNING: This routine must only be called when we have already sent
2162 * a SYN packet that crossed the incoming SYN that caused this routine
2163 * to get called. If this assumption fails then the initial rcv_wnd
2164 * and rcv_wscale values will not be correct.
2166 int tcp_send_synack(struct sock
*sk
)
2168 struct sk_buff
*skb
;
2170 skb
= tcp_write_queue_head(sk
);
2171 if (skb
== NULL
|| !(TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_SYN
)) {
2172 printk(KERN_DEBUG
"tcp_send_synack: wrong queue state\n");
2175 if (!(TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_ACK
)) {
2176 if (skb_cloned(skb
)) {
2177 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
2180 tcp_unlink_write_queue(skb
, sk
);
2181 skb_header_release(nskb
);
2182 __tcp_add_write_queue_head(sk
, nskb
);
2183 sk_wmem_free_skb(sk
, skb
);
2184 sk
->sk_wmem_queued
+= nskb
->truesize
;
2185 sk_mem_charge(sk
, nskb
->truesize
);
2189 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_ACK
;
2190 TCP_ECN_send_synack(tcp_sk(sk
), skb
);
2192 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2193 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2197 * Prepare a SYN-ACK.
2199 struct sk_buff
*tcp_make_synack(struct sock
*sk
, struct dst_entry
*dst
,
2200 struct request_sock
*req
)
2202 struct inet_request_sock
*ireq
= inet_rsk(req
);
2203 struct tcp_sock
*tp
= tcp_sk(sk
);
2205 int tcp_header_size
;
2206 struct tcp_out_options opts
;
2207 struct sk_buff
*skb
;
2208 struct tcp_md5sig_key
*md5
;
2209 __u8
*md5_hash_location
;
2212 skb
= sock_wmalloc(sk
, MAX_TCP_HEADER
+ 15, 1, GFP_ATOMIC
);
2216 /* Reserve space for headers. */
2217 skb_reserve(skb
, MAX_TCP_HEADER
);
2219 skb
->dst
= dst_clone(dst
);
2221 mss
= dst_metric(dst
, RTAX_ADVMSS
);
2222 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< mss
)
2223 mss
= tp
->rx_opt
.user_mss
;
2225 if (req
->rcv_wnd
== 0) { /* ignored for retransmitted syns */
2227 /* Set this up on the first call only */
2228 req
->window_clamp
= tp
->window_clamp
? : dst_metric(dst
, RTAX_WINDOW
);
2229 /* tcp_full_space because it is guaranteed to be the first packet */
2230 tcp_select_initial_window(tcp_full_space(sk
),
2231 mss
- (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0),
2236 ireq
->rcv_wscale
= rcv_wscale
;
2239 memset(&opts
, 0, sizeof(opts
));
2240 #ifdef CONFIG_SYN_COOKIES
2241 if (unlikely(req
->cookie_ts
))
2242 TCP_SKB_CB(skb
)->when
= cookie_init_timestamp(req
);
2245 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2246 tcp_header_size
= tcp_synack_options(sk
, req
, mss
,
2248 sizeof(struct tcphdr
);
2250 skb_push(skb
, tcp_header_size
);
2251 skb_reset_transport_header(skb
);
2254 memset(th
, 0, sizeof(struct tcphdr
));
2257 TCP_ECN_make_synack(req
, th
);
2258 th
->source
= ireq
->loc_port
;
2259 th
->dest
= ireq
->rmt_port
;
2260 /* Setting of flags are superfluous here for callers (and ECE is
2261 * not even correctly set)
2263 tcp_init_nondata_skb(skb
, tcp_rsk(req
)->snt_isn
,
2264 TCPCB_FLAG_SYN
| TCPCB_FLAG_ACK
);
2265 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
2266 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_isn
+ 1);
2268 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2269 th
->window
= htons(min(req
->rcv_wnd
, 65535U));
2270 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
, &md5_hash_location
);
2271 th
->doff
= (tcp_header_size
>> 2);
2272 TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
);
2274 #ifdef CONFIG_TCP_MD5SIG
2275 /* Okay, we have all we need - do the md5 hash if needed */
2277 tp
->af_specific
->calc_md5_hash(md5_hash_location
,
2278 md5
, NULL
, req
, skb
);
2286 * Do all connect socket setups that can be done AF independent.
2288 static void tcp_connect_init(struct sock
*sk
)
2290 struct dst_entry
*dst
= __sk_dst_get(sk
);
2291 struct tcp_sock
*tp
= tcp_sk(sk
);
2294 /* We'll fix this up when we get a response from the other end.
2295 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2297 tp
->tcp_header_len
= sizeof(struct tcphdr
) +
2298 (sysctl_tcp_timestamps
? TCPOLEN_TSTAMP_ALIGNED
: 0);
2300 #ifdef CONFIG_TCP_MD5SIG
2301 if (tp
->af_specific
->md5_lookup(sk
, sk
) != NULL
)
2302 tp
->tcp_header_len
+= TCPOLEN_MD5SIG_ALIGNED
;
2305 /* If user gave his TCP_MAXSEG, record it to clamp */
2306 if (tp
->rx_opt
.user_mss
)
2307 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
2310 tcp_sync_mss(sk
, dst_mtu(dst
));
2312 if (!tp
->window_clamp
)
2313 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
2314 tp
->advmss
= dst_metric(dst
, RTAX_ADVMSS
);
2315 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< tp
->advmss
)
2316 tp
->advmss
= tp
->rx_opt
.user_mss
;
2318 tcp_initialize_rcv_mss(sk
);
2320 tcp_select_initial_window(tcp_full_space(sk
),
2321 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
2324 sysctl_tcp_window_scaling
,
2327 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
2328 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
2331 sock_reset_flag(sk
, SOCK_DONE
);
2333 tcp_init_wl(tp
, tp
->write_seq
, 0);
2334 tp
->snd_una
= tp
->write_seq
;
2335 tp
->snd_sml
= tp
->write_seq
;
2336 tp
->snd_up
= tp
->write_seq
;
2341 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
2342 inet_csk(sk
)->icsk_retransmits
= 0;
2343 tcp_clear_retrans(tp
);
2347 * Build a SYN and send it off.
2349 int tcp_connect(struct sock
*sk
)
2351 struct tcp_sock
*tp
= tcp_sk(sk
);
2352 struct sk_buff
*buff
;
2354 tcp_connect_init(sk
);
2356 buff
= alloc_skb_fclone(MAX_TCP_HEADER
+ 15, sk
->sk_allocation
);
2357 if (unlikely(buff
== NULL
))
2360 /* Reserve space for headers. */
2361 skb_reserve(buff
, MAX_TCP_HEADER
);
2363 tp
->snd_nxt
= tp
->write_seq
;
2364 tcp_init_nondata_skb(buff
, tp
->write_seq
++, TCPCB_FLAG_SYN
);
2365 TCP_ECN_send_syn(sk
, buff
);
2368 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2369 tp
->retrans_stamp
= TCP_SKB_CB(buff
)->when
;
2370 skb_header_release(buff
);
2371 __tcp_add_write_queue_tail(sk
, buff
);
2372 sk
->sk_wmem_queued
+= buff
->truesize
;
2373 sk_mem_charge(sk
, buff
->truesize
);
2374 tp
->packets_out
+= tcp_skb_pcount(buff
);
2375 tcp_transmit_skb(sk
, buff
, 1, GFP_KERNEL
);
2377 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2378 * in order to make this packet get counted in tcpOutSegs.
2380 tp
->snd_nxt
= tp
->write_seq
;
2381 tp
->pushed_seq
= tp
->write_seq
;
2382 TCP_INC_STATS(sock_net(sk
), TCP_MIB_ACTIVEOPENS
);
2384 /* Timer for repeating the SYN until an answer. */
2385 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2386 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2390 /* Send out a delayed ack, the caller does the policy checking
2391 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2394 void tcp_send_delayed_ack(struct sock
*sk
)
2396 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2397 int ato
= icsk
->icsk_ack
.ato
;
2398 unsigned long timeout
;
2400 if (ato
> TCP_DELACK_MIN
) {
2401 const struct tcp_sock
*tp
= tcp_sk(sk
);
2402 int max_ato
= HZ
/ 2;
2404 if (icsk
->icsk_ack
.pingpong
||
2405 (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
2406 max_ato
= TCP_DELACK_MAX
;
2408 /* Slow path, intersegment interval is "high". */
2410 /* If some rtt estimate is known, use it to bound delayed ack.
2411 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2415 int rtt
= max(tp
->srtt
>> 3, TCP_DELACK_MIN
);
2421 ato
= min(ato
, max_ato
);
2424 /* Stay within the limit we were given */
2425 timeout
= jiffies
+ ato
;
2427 /* Use new timeout only if there wasn't a older one earlier. */
2428 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
2429 /* If delack timer was blocked or is about to expire,
2432 if (icsk
->icsk_ack
.blocked
||
2433 time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
2438 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
2439 timeout
= icsk
->icsk_ack
.timeout
;
2441 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
2442 icsk
->icsk_ack
.timeout
= timeout
;
2443 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
2446 /* This routine sends an ack and also updates the window. */
2447 void tcp_send_ack(struct sock
*sk
)
2449 struct sk_buff
*buff
;
2451 /* If we have been reset, we may not send again. */
2452 if (sk
->sk_state
== TCP_CLOSE
)
2455 /* We are not putting this on the write queue, so
2456 * tcp_transmit_skb() will set the ownership to this
2459 buff
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2461 inet_csk_schedule_ack(sk
);
2462 inet_csk(sk
)->icsk_ack
.ato
= TCP_ATO_MIN
;
2463 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
2464 TCP_DELACK_MAX
, TCP_RTO_MAX
);
2468 /* Reserve space for headers and prepare control bits. */
2469 skb_reserve(buff
, MAX_TCP_HEADER
);
2470 tcp_init_nondata_skb(buff
, tcp_acceptable_seq(sk
), TCPCB_FLAG_ACK
);
2472 /* Send it off, this clears delayed acks for us. */
2473 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2474 tcp_transmit_skb(sk
, buff
, 0, GFP_ATOMIC
);
2477 /* This routine sends a packet with an out of date sequence
2478 * number. It assumes the other end will try to ack it.
2480 * Question: what should we make while urgent mode?
2481 * 4.4BSD forces sending single byte of data. We cannot send
2482 * out of window data, because we have SND.NXT==SND.MAX...
2484 * Current solution: to send TWO zero-length segments in urgent mode:
2485 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2486 * out-of-date with SND.UNA-1 to probe window.
2488 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
)
2490 struct tcp_sock
*tp
= tcp_sk(sk
);
2491 struct sk_buff
*skb
;
2493 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2494 skb
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2498 /* Reserve space for headers and set control bits. */
2499 skb_reserve(skb
, MAX_TCP_HEADER
);
2500 /* Use a previous sequence. This should cause the other
2501 * end to send an ack. Don't queue or clone SKB, just
2504 tcp_init_nondata_skb(skb
, tp
->snd_una
- !urgent
, TCPCB_FLAG_ACK
);
2505 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2506 return tcp_transmit_skb(sk
, skb
, 0, GFP_ATOMIC
);
2509 int tcp_write_wakeup(struct sock
*sk
)
2511 struct tcp_sock
*tp
= tcp_sk(sk
);
2512 struct sk_buff
*skb
;
2514 if (sk
->sk_state
== TCP_CLOSE
)
2517 if ((skb
= tcp_send_head(sk
)) != NULL
&&
2518 before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
))) {
2520 unsigned int mss
= tcp_current_mss(sk
, 0);
2521 unsigned int seg_size
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
2523 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
2524 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
2526 /* We are probing the opening of a window
2527 * but the window size is != 0
2528 * must have been a result SWS avoidance ( sender )
2530 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
2532 seg_size
= min(seg_size
, mss
);
2533 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2534 if (tcp_fragment(sk
, skb
, seg_size
, mss
))
2536 } else if (!tcp_skb_pcount(skb
))
2537 tcp_set_skb_tso_segs(sk
, skb
, mss
);
2539 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2540 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2541 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2543 tcp_event_new_data_sent(sk
, skb
);
2546 if (between(tp
->snd_up
, tp
->snd_una
+ 1, tp
->snd_una
+ 0xFFFF))
2547 tcp_xmit_probe_skb(sk
, 1);
2548 return tcp_xmit_probe_skb(sk
, 0);
2552 /* A window probe timeout has occurred. If window is not closed send
2553 * a partial packet else a zero probe.
2555 void tcp_send_probe0(struct sock
*sk
)
2557 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2558 struct tcp_sock
*tp
= tcp_sk(sk
);
2561 err
= tcp_write_wakeup(sk
);
2563 if (tp
->packets_out
|| !tcp_send_head(sk
)) {
2564 /* Cancel probe timer, if it is not required. */
2565 icsk
->icsk_probes_out
= 0;
2566 icsk
->icsk_backoff
= 0;
2571 if (icsk
->icsk_backoff
< sysctl_tcp_retries2
)
2572 icsk
->icsk_backoff
++;
2573 icsk
->icsk_probes_out
++;
2574 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2575 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
2578 /* If packet was not sent due to local congestion,
2579 * do not backoff and do not remember icsk_probes_out.
2580 * Let local senders to fight for local resources.
2582 * Use accumulated backoff yet.
2584 if (!icsk
->icsk_probes_out
)
2585 icsk
->icsk_probes_out
= 1;
2586 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2587 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
,
2588 TCP_RESOURCE_PROBE_INTERVAL
),
2593 EXPORT_SYMBOL(tcp_select_initial_window
);
2594 EXPORT_SYMBOL(tcp_connect
);
2595 EXPORT_SYMBOL(tcp_make_synack
);
2596 EXPORT_SYMBOL(tcp_simple_retransmit
);
2597 EXPORT_SYMBOL(tcp_sync_mss
);
2598 EXPORT_SYMBOL(tcp_mtup_init
);