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 #define OPTION_SACK_ADVERTISE (1 << 0)
349 #define OPTION_TS (1 << 1)
350 #define OPTION_MD5 (1 << 2)
352 struct tcp_out_options
{
353 u8 options
; /* bit field of OPTION_* */
354 u8 ws
; /* window scale, 0 to disable */
355 u8 num_sack_blocks
; /* number of SACK blocks to include */
356 u16 mss
; /* 0 to disable */
357 __u32 tsval
, tsecr
; /* need to include OPTION_TS */
360 static void tcp_options_write(__be32
*ptr
, struct tcp_sock
*tp
,
361 const struct tcp_out_options
*opts
,
363 if (unlikely(OPTION_MD5
& opts
->options
)) {
364 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
366 (TCPOPT_MD5SIG
<< 8) |
368 *md5_hash
= (__u8
*)ptr
;
374 if (likely(OPTION_TS
& opts
->options
)) {
375 if (unlikely(OPTION_SACK_ADVERTISE
& opts
->options
)) {
376 *ptr
++ = htonl((TCPOPT_SACK_PERM
<< 24) |
377 (TCPOLEN_SACK_PERM
<< 16) |
378 (TCPOPT_TIMESTAMP
<< 8) |
381 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
383 (TCPOPT_TIMESTAMP
<< 8) |
386 *ptr
++ = htonl(opts
->tsval
);
387 *ptr
++ = htonl(opts
->tsecr
);
390 if (unlikely(opts
->mss
)) {
391 *ptr
++ = htonl((TCPOPT_MSS
<< 24) |
392 (TCPOLEN_MSS
<< 16) |
396 if (unlikely(OPTION_SACK_ADVERTISE
& opts
->options
&&
397 !(OPTION_TS
& opts
->options
))) {
398 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
400 (TCPOPT_SACK_PERM
<< 8) |
404 if (unlikely(opts
->ws
)) {
405 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
406 (TCPOPT_WINDOW
<< 16) |
407 (TCPOLEN_WINDOW
<< 8) |
411 if (unlikely(opts
->num_sack_blocks
)) {
412 struct tcp_sack_block
*sp
= tp
->rx_opt
.dsack
?
413 tp
->duplicate_sack
: tp
->selective_acks
;
416 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
419 (TCPOLEN_SACK_BASE
+ (opts
->num_sack_blocks
*
420 TCPOLEN_SACK_PERBLOCK
)));
422 for (this_sack
= 0; this_sack
< opts
->num_sack_blocks
;
424 *ptr
++ = htonl(sp
[this_sack
].start_seq
);
425 *ptr
++ = htonl(sp
[this_sack
].end_seq
);
428 if (tp
->rx_opt
.dsack
) {
429 tp
->rx_opt
.dsack
= 0;
430 tp
->rx_opt
.eff_sacks
--;
435 static unsigned tcp_syn_options(struct sock
*sk
, struct sk_buff
*skb
,
436 struct tcp_out_options
*opts
,
437 struct tcp_md5sig_key
**md5
) {
438 struct tcp_sock
*tp
= tcp_sk(sk
);
441 #ifdef CONFIG_TCP_MD5SIG
442 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
444 opts
->options
|= OPTION_MD5
;
445 size
+= TCPOLEN_MD5SIG_ALIGNED
;
451 /* We always get an MSS option. The option bytes which will be seen in
452 * normal data packets should timestamps be used, must be in the MSS
453 * advertised. But we subtract them from tp->mss_cache so that
454 * calculations in tcp_sendmsg are simpler etc. So account for this
455 * fact here if necessary. If we don't do this correctly, as a
456 * receiver we won't recognize data packets as being full sized when we
457 * should, and thus we won't abide by the delayed ACK rules correctly.
458 * SACKs don't matter, we never delay an ACK when we have any of those
460 opts
->mss
= tcp_advertise_mss(sk
);
461 size
+= TCPOLEN_MSS_ALIGNED
;
463 if (likely(sysctl_tcp_timestamps
&& *md5
== NULL
)) {
464 opts
->options
|= OPTION_TS
;
465 opts
->tsval
= TCP_SKB_CB(skb
)->when
;
466 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
467 size
+= TCPOLEN_TSTAMP_ALIGNED
;
469 if (likely(sysctl_tcp_window_scaling
)) {
470 opts
->ws
= tp
->rx_opt
.rcv_wscale
;
471 size
+= TCPOLEN_WSCALE_ALIGNED
;
473 if (likely(sysctl_tcp_sack
)) {
474 opts
->options
|= OPTION_SACK_ADVERTISE
;
475 if (unlikely(!(OPTION_TS
& opts
->options
)))
476 size
+= TCPOLEN_SACKPERM_ALIGNED
;
482 static unsigned tcp_synack_options(struct sock
*sk
,
483 struct request_sock
*req
,
484 unsigned mss
, struct sk_buff
*skb
,
485 struct tcp_out_options
*opts
,
486 struct tcp_md5sig_key
**md5
) {
488 struct inet_request_sock
*ireq
= inet_rsk(req
);
491 #ifdef CONFIG_TCP_MD5SIG
492 *md5
= tcp_rsk(req
)->af_specific
->md5_lookup(sk
, req
);
494 opts
->options
|= OPTION_MD5
;
495 size
+= TCPOLEN_MD5SIG_ALIGNED
;
501 /* we can't fit any SACK blocks in a packet with MD5 + TS
502 options. There was discussion about disabling SACK rather than TS in
503 order to fit in better with old, buggy kernels, but that was deemed
504 to be unnecessary. */
505 doing_ts
= ireq
->tstamp_ok
&& !(*md5
&& ireq
->sack_ok
);
508 size
+= TCPOLEN_MSS_ALIGNED
;
510 if (likely(ireq
->wscale_ok
)) {
511 opts
->ws
= ireq
->rcv_wscale
;
512 size
+= TCPOLEN_WSCALE_ALIGNED
;
514 if (likely(doing_ts
)) {
515 opts
->options
|= OPTION_TS
;
516 opts
->tsval
= TCP_SKB_CB(skb
)->when
;
517 opts
->tsecr
= req
->ts_recent
;
518 size
+= TCPOLEN_TSTAMP_ALIGNED
;
520 if (likely(ireq
->sack_ok
)) {
521 opts
->options
|= OPTION_SACK_ADVERTISE
;
522 if (unlikely(!doing_ts
))
523 size
+= TCPOLEN_SACKPERM_ALIGNED
;
529 static unsigned tcp_established_options(struct sock
*sk
, struct sk_buff
*skb
,
530 struct tcp_out_options
*opts
,
531 struct tcp_md5sig_key
**md5
) {
532 struct tcp_skb_cb
*tcb
= skb
? TCP_SKB_CB(skb
) : NULL
;
533 struct tcp_sock
*tp
= tcp_sk(sk
);
536 #ifdef CONFIG_TCP_MD5SIG
537 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
538 if (unlikely(*md5
)) {
539 opts
->options
|= OPTION_MD5
;
540 size
+= TCPOLEN_MD5SIG_ALIGNED
;
546 if (likely(tp
->rx_opt
.tstamp_ok
)) {
547 opts
->options
|= OPTION_TS
;
548 opts
->tsval
= tcb
? tcb
->when
: 0;
549 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
550 size
+= TCPOLEN_TSTAMP_ALIGNED
;
553 if (unlikely(tp
->rx_opt
.eff_sacks
)) {
554 const unsigned remaining
= MAX_TCP_OPTION_SPACE
- size
;
555 opts
->num_sack_blocks
=
556 min_t(unsigned, tp
->rx_opt
.eff_sacks
,
557 (remaining
- TCPOLEN_SACK_BASE_ALIGNED
) /
558 TCPOLEN_SACK_PERBLOCK
);
559 size
+= TCPOLEN_SACK_BASE_ALIGNED
+
560 opts
->num_sack_blocks
* TCPOLEN_SACK_PERBLOCK
;
566 /* This routine actually transmits TCP packets queued in by
567 * tcp_do_sendmsg(). This is used by both the initial
568 * transmission and possible later retransmissions.
569 * All SKB's seen here are completely headerless. It is our
570 * job to build the TCP header, and pass the packet down to
571 * IP so it can do the same plus pass the packet off to the
574 * We are working here with either a clone of the original
575 * SKB, or a fresh unique copy made by the retransmit engine.
577 static int tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int clone_it
,
580 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
581 struct inet_sock
*inet
;
583 struct tcp_skb_cb
*tcb
;
584 struct tcp_out_options opts
;
585 unsigned tcp_options_size
, tcp_header_size
;
586 struct tcp_md5sig_key
*md5
;
587 __u8
*md5_hash_location
;
591 BUG_ON(!skb
|| !tcp_skb_pcount(skb
));
593 /* If congestion control is doing timestamping, we must
594 * take such a timestamp before we potentially clone/copy.
596 if (icsk
->icsk_ca_ops
->flags
& TCP_CONG_RTT_STAMP
)
597 __net_timestamp(skb
);
599 if (likely(clone_it
)) {
600 if (unlikely(skb_cloned(skb
)))
601 skb
= pskb_copy(skb
, gfp_mask
);
603 skb
= skb_clone(skb
, gfp_mask
);
610 tcb
= TCP_SKB_CB(skb
);
611 memset(&opts
, 0, sizeof(opts
));
613 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
))
614 tcp_options_size
= tcp_syn_options(sk
, skb
, &opts
, &md5
);
616 tcp_options_size
= tcp_established_options(sk
, skb
, &opts
,
618 tcp_header_size
= tcp_options_size
+ sizeof(struct tcphdr
);
620 if (tcp_packets_in_flight(tp
) == 0)
621 tcp_ca_event(sk
, CA_EVENT_TX_START
);
623 skb_push(skb
, tcp_header_size
);
624 skb_reset_transport_header(skb
);
625 skb_set_owner_w(skb
, sk
);
627 /* Build TCP header and checksum it. */
629 th
->source
= inet
->sport
;
630 th
->dest
= inet
->dport
;
631 th
->seq
= htonl(tcb
->seq
);
632 th
->ack_seq
= htonl(tp
->rcv_nxt
);
633 *(((__be16
*)th
) + 6) = htons(((tcp_header_size
>> 2) << 12) |
636 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
637 /* RFC1323: The window in SYN & SYN/ACK segments
640 th
->window
= htons(min(tp
->rcv_wnd
, 65535U));
642 th
->window
= htons(tcp_select_window(sk
));
647 if (unlikely(tp
->urg_mode
&&
648 between(tp
->snd_up
, tcb
->seq
+ 1, tcb
->seq
+ 0xFFFF))) {
649 th
->urg_ptr
= htons(tp
->snd_up
- tcb
->seq
);
653 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
, &md5_hash_location
);
654 if (likely((tcb
->flags
& TCPCB_FLAG_SYN
) == 0))
655 TCP_ECN_send(sk
, skb
, tcp_header_size
);
657 #ifdef CONFIG_TCP_MD5SIG
658 /* Calculate the MD5 hash, as we have all we need now */
660 sk
->sk_route_caps
&= ~NETIF_F_GSO_MASK
;
661 tp
->af_specific
->calc_md5_hash(md5_hash_location
,
666 icsk
->icsk_af_ops
->send_check(sk
, skb
->len
, skb
);
668 if (likely(tcb
->flags
& TCPCB_FLAG_ACK
))
669 tcp_event_ack_sent(sk
, tcp_skb_pcount(skb
));
671 if (skb
->len
!= tcp_header_size
)
672 tcp_event_data_sent(tp
, skb
, sk
);
674 if (after(tcb
->end_seq
, tp
->snd_nxt
) || tcb
->seq
== tcb
->end_seq
)
675 TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
);
677 err
= icsk
->icsk_af_ops
->queue_xmit(skb
, 0);
678 if (likely(err
<= 0))
681 tcp_enter_cwr(sk
, 1);
683 return net_xmit_eval(err
);
686 /* This routine just queue's the buffer
688 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
689 * otherwise socket can stall.
691 static void tcp_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
693 struct tcp_sock
*tp
= tcp_sk(sk
);
695 /* Advance write_seq and place onto the write_queue. */
696 tp
->write_seq
= TCP_SKB_CB(skb
)->end_seq
;
697 skb_header_release(skb
);
698 tcp_add_write_queue_tail(sk
, skb
);
699 sk
->sk_wmem_queued
+= skb
->truesize
;
700 sk_mem_charge(sk
, skb
->truesize
);
703 static void tcp_set_skb_tso_segs(struct sock
*sk
, struct sk_buff
*skb
,
704 unsigned int mss_now
)
706 if (skb
->len
<= mss_now
|| !sk_can_gso(sk
)) {
707 /* Avoid the costly divide in the normal
710 skb_shinfo(skb
)->gso_segs
= 1;
711 skb_shinfo(skb
)->gso_size
= 0;
712 skb_shinfo(skb
)->gso_type
= 0;
714 skb_shinfo(skb
)->gso_segs
= DIV_ROUND_UP(skb
->len
, mss_now
);
715 skb_shinfo(skb
)->gso_size
= mss_now
;
716 skb_shinfo(skb
)->gso_type
= sk
->sk_gso_type
;
720 /* When a modification to fackets out becomes necessary, we need to check
721 * skb is counted to fackets_out or not.
723 static void tcp_adjust_fackets_out(struct sock
*sk
, struct sk_buff
*skb
,
726 struct tcp_sock
*tp
= tcp_sk(sk
);
728 if (!tp
->sacked_out
|| tcp_is_reno(tp
))
731 if (after(tcp_highest_sack_seq(tp
), TCP_SKB_CB(skb
)->seq
))
732 tp
->fackets_out
-= decr
;
735 /* Function to create two new TCP segments. Shrinks the given segment
736 * to the specified size and appends a new segment with the rest of the
737 * packet to the list. This won't be called frequently, I hope.
738 * Remember, these are still headerless SKBs at this point.
740 int tcp_fragment(struct sock
*sk
, struct sk_buff
*skb
, u32 len
,
741 unsigned int mss_now
)
743 struct tcp_sock
*tp
= tcp_sk(sk
);
744 struct sk_buff
*buff
;
745 int nsize
, old_factor
;
749 BUG_ON(len
> skb
->len
);
751 tcp_clear_retrans_hints_partial(tp
);
752 nsize
= skb_headlen(skb
) - len
;
756 if (skb_cloned(skb
) &&
757 skb_is_nonlinear(skb
) &&
758 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
761 /* Get a new skb... force flag on. */
762 buff
= sk_stream_alloc_skb(sk
, nsize
, GFP_ATOMIC
);
764 return -ENOMEM
; /* We'll just try again later. */
766 sk
->sk_wmem_queued
+= buff
->truesize
;
767 sk_mem_charge(sk
, buff
->truesize
);
768 nlen
= skb
->len
- len
- nsize
;
769 buff
->truesize
+= nlen
;
770 skb
->truesize
-= nlen
;
772 /* Correct the sequence numbers. */
773 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
774 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
775 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
777 /* PSH and FIN should only be set in the second packet. */
778 flags
= TCP_SKB_CB(skb
)->flags
;
779 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
| TCPCB_FLAG_PSH
);
780 TCP_SKB_CB(buff
)->flags
= flags
;
781 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
783 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_PARTIAL
) {
784 /* Copy and checksum data tail into the new buffer. */
785 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
,
786 skb_put(buff
, nsize
),
791 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
793 skb
->ip_summed
= CHECKSUM_PARTIAL
;
794 skb_split(skb
, buff
, len
);
797 buff
->ip_summed
= skb
->ip_summed
;
799 /* Looks stupid, but our code really uses when of
800 * skbs, which it never sent before. --ANK
802 TCP_SKB_CB(buff
)->when
= TCP_SKB_CB(skb
)->when
;
803 buff
->tstamp
= skb
->tstamp
;
805 old_factor
= tcp_skb_pcount(skb
);
807 /* Fix up tso_factor for both original and new SKB. */
808 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
809 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
811 /* If this packet has been sent out already, we must
812 * adjust the various packet counters.
814 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
815 int diff
= old_factor
- tcp_skb_pcount(skb
) -
816 tcp_skb_pcount(buff
);
818 tp
->packets_out
-= diff
;
820 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
821 tp
->sacked_out
-= diff
;
822 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
823 tp
->retrans_out
-= diff
;
825 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
)
826 tp
->lost_out
-= diff
;
828 /* Adjust Reno SACK estimate. */
829 if (tcp_is_reno(tp
) && diff
> 0) {
830 tcp_dec_pcount_approx_int(&tp
->sacked_out
, diff
);
831 tcp_verify_left_out(tp
);
833 tcp_adjust_fackets_out(sk
, skb
, diff
);
836 /* Link BUFF into the send queue. */
837 skb_header_release(buff
);
838 tcp_insert_write_queue_after(skb
, buff
, sk
);
843 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
844 * eventually). The difference is that pulled data not copied, but
845 * immediately discarded.
847 static void __pskb_trim_head(struct sk_buff
*skb
, int len
)
853 for (i
= 0; i
< skb_shinfo(skb
)->nr_frags
; i
++) {
854 if (skb_shinfo(skb
)->frags
[i
].size
<= eat
) {
855 put_page(skb_shinfo(skb
)->frags
[i
].page
);
856 eat
-= skb_shinfo(skb
)->frags
[i
].size
;
858 skb_shinfo(skb
)->frags
[k
] = skb_shinfo(skb
)->frags
[i
];
860 skb_shinfo(skb
)->frags
[k
].page_offset
+= eat
;
861 skb_shinfo(skb
)->frags
[k
].size
-= eat
;
867 skb_shinfo(skb
)->nr_frags
= k
;
869 skb_reset_tail_pointer(skb
);
870 skb
->data_len
-= len
;
871 skb
->len
= skb
->data_len
;
874 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
876 if (skb_cloned(skb
) && pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
879 /* If len == headlen, we avoid __skb_pull to preserve alignment. */
880 if (unlikely(len
< skb_headlen(skb
)))
881 __skb_pull(skb
, len
);
883 __pskb_trim_head(skb
, len
- skb_headlen(skb
));
885 TCP_SKB_CB(skb
)->seq
+= len
;
886 skb
->ip_summed
= CHECKSUM_PARTIAL
;
888 skb
->truesize
-= len
;
889 sk
->sk_wmem_queued
-= len
;
890 sk_mem_uncharge(sk
, len
);
891 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
893 /* Any change of skb->len requires recalculation of tso
896 if (tcp_skb_pcount(skb
) > 1)
897 tcp_set_skb_tso_segs(sk
, skb
, tcp_current_mss(sk
, 1));
902 /* Not accounting for SACKs here. */
903 int tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
905 struct tcp_sock
*tp
= tcp_sk(sk
);
906 struct inet_connection_sock
*icsk
= inet_csk(sk
);
909 /* Calculate base mss without TCP options:
910 It is MMS_S - sizeof(tcphdr) of rfc1122
912 mss_now
= pmtu
- icsk
->icsk_af_ops
->net_header_len
- sizeof(struct tcphdr
);
914 /* Clamp it (mss_clamp does not include tcp options) */
915 if (mss_now
> tp
->rx_opt
.mss_clamp
)
916 mss_now
= tp
->rx_opt
.mss_clamp
;
918 /* Now subtract optional transport overhead */
919 mss_now
-= icsk
->icsk_ext_hdr_len
;
921 /* Then reserve room for full set of TCP options and 8 bytes of data */
925 /* Now subtract TCP options size, not including SACKs */
926 mss_now
-= tp
->tcp_header_len
- sizeof(struct tcphdr
);
931 /* Inverse of above */
932 int tcp_mss_to_mtu(struct sock
*sk
, int mss
)
934 struct tcp_sock
*tp
= tcp_sk(sk
);
935 struct inet_connection_sock
*icsk
= inet_csk(sk
);
940 icsk
->icsk_ext_hdr_len
+
941 icsk
->icsk_af_ops
->net_header_len
;
946 void tcp_mtup_init(struct sock
*sk
)
948 struct tcp_sock
*tp
= tcp_sk(sk
);
949 struct inet_connection_sock
*icsk
= inet_csk(sk
);
951 icsk
->icsk_mtup
.enabled
= sysctl_tcp_mtu_probing
> 1;
952 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+ sizeof(struct tcphdr
) +
953 icsk
->icsk_af_ops
->net_header_len
;
954 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, sysctl_tcp_base_mss
);
955 icsk
->icsk_mtup
.probe_size
= 0;
958 /* Bound MSS / TSO packet size with the half of the window */
959 static int tcp_bound_to_half_wnd(struct tcp_sock
*tp
, int pktsize
)
961 if (tp
->max_window
&& pktsize
> (tp
->max_window
>> 1))
962 return max(tp
->max_window
>> 1, 68U - tp
->tcp_header_len
);
967 /* This function synchronize snd mss to current pmtu/exthdr set.
969 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
970 for TCP options, but includes only bare TCP header.
972 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
973 It is minimum of user_mss and mss received with SYN.
974 It also does not include TCP options.
976 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
978 tp->mss_cache is current effective sending mss, including
979 all tcp options except for SACKs. It is evaluated,
980 taking into account current pmtu, but never exceeds
981 tp->rx_opt.mss_clamp.
983 NOTE1. rfc1122 clearly states that advertised MSS
984 DOES NOT include either tcp or ip options.
986 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
987 are READ ONLY outside this function. --ANK (980731)
989 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
991 struct tcp_sock
*tp
= tcp_sk(sk
);
992 struct inet_connection_sock
*icsk
= inet_csk(sk
);
995 if (icsk
->icsk_mtup
.search_high
> pmtu
)
996 icsk
->icsk_mtup
.search_high
= pmtu
;
998 mss_now
= tcp_mtu_to_mss(sk
, pmtu
);
999 mss_now
= tcp_bound_to_half_wnd(tp
, mss_now
);
1001 /* And store cached results */
1002 icsk
->icsk_pmtu_cookie
= pmtu
;
1003 if (icsk
->icsk_mtup
.enabled
)
1004 mss_now
= min(mss_now
, tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
));
1005 tp
->mss_cache
= mss_now
;
1010 /* Compute the current effective MSS, taking SACKs and IP options,
1011 * and even PMTU discovery events into account.
1013 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
1014 * cannot be large. However, taking into account rare use of URG, this
1015 * is not a big flaw.
1017 unsigned int tcp_current_mss(struct sock
*sk
, int large_allowed
)
1019 struct tcp_sock
*tp
= tcp_sk(sk
);
1020 struct dst_entry
*dst
= __sk_dst_get(sk
);
1024 unsigned header_len
;
1025 struct tcp_out_options opts
;
1026 struct tcp_md5sig_key
*md5
;
1028 mss_now
= tp
->mss_cache
;
1030 if (large_allowed
&& sk_can_gso(sk
) && !tp
->urg_mode
)
1034 u32 mtu
= dst_mtu(dst
);
1035 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
1036 mss_now
= tcp_sync_mss(sk
, mtu
);
1039 header_len
= tcp_established_options(sk
, NULL
, &opts
, &md5
) +
1040 sizeof(struct tcphdr
);
1041 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1042 * some common options. If this is an odd packet (because we have SACK
1043 * blocks etc) then our calculated header_len will be different, and
1044 * we have to adjust mss_now correspondingly */
1045 if (header_len
!= tp
->tcp_header_len
) {
1046 int delta
= (int) header_len
- tp
->tcp_header_len
;
1050 xmit_size_goal
= mss_now
;
1053 xmit_size_goal
= ((sk
->sk_gso_max_size
- 1) -
1054 inet_csk(sk
)->icsk_af_ops
->net_header_len
-
1055 inet_csk(sk
)->icsk_ext_hdr_len
-
1056 tp
->tcp_header_len
);
1058 xmit_size_goal
= tcp_bound_to_half_wnd(tp
, xmit_size_goal
);
1059 xmit_size_goal
-= (xmit_size_goal
% mss_now
);
1061 tp
->xmit_size_goal
= xmit_size_goal
;
1066 /* Congestion window validation. (RFC2861) */
1067 static void tcp_cwnd_validate(struct sock
*sk
)
1069 struct tcp_sock
*tp
= tcp_sk(sk
);
1071 if (tp
->packets_out
>= tp
->snd_cwnd
) {
1072 /* Network is feed fully. */
1073 tp
->snd_cwnd_used
= 0;
1074 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1076 /* Network starves. */
1077 if (tp
->packets_out
> tp
->snd_cwnd_used
)
1078 tp
->snd_cwnd_used
= tp
->packets_out
;
1080 if (sysctl_tcp_slow_start_after_idle
&&
1081 (s32
)(tcp_time_stamp
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
)
1082 tcp_cwnd_application_limited(sk
);
1086 /* Returns the portion of skb which can be sent right away without
1087 * introducing MSS oddities to segment boundaries. In rare cases where
1088 * mss_now != mss_cache, we will request caller to create a small skb
1089 * per input skb which could be mostly avoided here (if desired).
1091 * We explicitly want to create a request for splitting write queue tail
1092 * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1093 * thus all the complexity (cwnd_len is always MSS multiple which we
1094 * return whenever allowed by the other factors). Basically we need the
1095 * modulo only when the receiver window alone is the limiting factor or
1096 * when we would be allowed to send the split-due-to-Nagle skb fully.
1098 static unsigned int tcp_mss_split_point(struct sock
*sk
, struct sk_buff
*skb
,
1099 unsigned int mss_now
, unsigned int cwnd
)
1101 struct tcp_sock
*tp
= tcp_sk(sk
);
1102 u32 needed
, window
, cwnd_len
;
1104 window
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1105 cwnd_len
= mss_now
* cwnd
;
1107 if (likely(cwnd_len
<= window
&& skb
!= tcp_write_queue_tail(sk
)))
1110 needed
= min(skb
->len
, window
);
1112 if (cwnd_len
<= needed
)
1115 return needed
- needed
% mss_now
;
1118 /* Can at least one segment of SKB be sent right now, according to the
1119 * congestion window rules? If so, return how many segments are allowed.
1121 static inline unsigned int tcp_cwnd_test(struct tcp_sock
*tp
,
1122 struct sk_buff
*skb
)
1124 u32 in_flight
, cwnd
;
1126 /* Don't be strict about the congestion window for the final FIN. */
1127 if ((TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
) &&
1128 tcp_skb_pcount(skb
) == 1)
1131 in_flight
= tcp_packets_in_flight(tp
);
1132 cwnd
= tp
->snd_cwnd
;
1133 if (in_flight
< cwnd
)
1134 return (cwnd
- in_flight
);
1139 /* This must be invoked the first time we consider transmitting
1140 * SKB onto the wire.
1142 static int tcp_init_tso_segs(struct sock
*sk
, struct sk_buff
*skb
,
1143 unsigned int mss_now
)
1145 int tso_segs
= tcp_skb_pcount(skb
);
1147 if (!tso_segs
|| (tso_segs
> 1 && tcp_skb_mss(skb
) != mss_now
)) {
1148 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1149 tso_segs
= tcp_skb_pcount(skb
);
1154 static inline int tcp_minshall_check(const struct tcp_sock
*tp
)
1156 return after(tp
->snd_sml
,tp
->snd_una
) &&
1157 !after(tp
->snd_sml
, tp
->snd_nxt
);
1160 /* Return 0, if packet can be sent now without violation Nagle's rules:
1161 * 1. It is full sized.
1162 * 2. Or it contains FIN. (already checked by caller)
1163 * 3. Or TCP_NODELAY was set.
1164 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1165 * With Minshall's modification: all sent small packets are ACKed.
1167 static inline int tcp_nagle_check(const struct tcp_sock
*tp
,
1168 const struct sk_buff
*skb
,
1169 unsigned mss_now
, int nonagle
)
1171 return (skb
->len
< mss_now
&&
1172 ((nonagle
& TCP_NAGLE_CORK
) ||
1173 (!nonagle
&& tp
->packets_out
&& tcp_minshall_check(tp
))));
1176 /* Return non-zero if the Nagle test allows this packet to be
1179 static inline int tcp_nagle_test(struct tcp_sock
*tp
, struct sk_buff
*skb
,
1180 unsigned int cur_mss
, int nonagle
)
1182 /* Nagle rule does not apply to frames, which sit in the middle of the
1183 * write_queue (they have no chances to get new data).
1185 * This is implemented in the callers, where they modify the 'nonagle'
1186 * argument based upon the location of SKB in the send queue.
1188 if (nonagle
& TCP_NAGLE_PUSH
)
1191 /* Don't use the nagle rule for urgent data (or for the final FIN).
1192 * Nagle can be ignored during F-RTO too (see RFC4138).
1194 if (tp
->urg_mode
|| (tp
->frto_counter
== 2) ||
1195 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
))
1198 if (!tcp_nagle_check(tp
, skb
, cur_mss
, nonagle
))
1204 /* Does at least the first segment of SKB fit into the send window? */
1205 static inline int tcp_snd_wnd_test(struct tcp_sock
*tp
, struct sk_buff
*skb
,
1206 unsigned int cur_mss
)
1208 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1210 if (skb
->len
> cur_mss
)
1211 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
1213 return !after(end_seq
, tcp_wnd_end(tp
));
1216 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1217 * should be put on the wire right now. If so, it returns the number of
1218 * packets allowed by the congestion window.
1220 static unsigned int tcp_snd_test(struct sock
*sk
, struct sk_buff
*skb
,
1221 unsigned int cur_mss
, int nonagle
)
1223 struct tcp_sock
*tp
= tcp_sk(sk
);
1224 unsigned int cwnd_quota
;
1226 tcp_init_tso_segs(sk
, skb
, cur_mss
);
1228 if (!tcp_nagle_test(tp
, skb
, cur_mss
, nonagle
))
1231 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1232 if (cwnd_quota
&& !tcp_snd_wnd_test(tp
, skb
, cur_mss
))
1238 int tcp_may_send_now(struct sock
*sk
)
1240 struct tcp_sock
*tp
= tcp_sk(sk
);
1241 struct sk_buff
*skb
= tcp_send_head(sk
);
1244 tcp_snd_test(sk
, skb
, tcp_current_mss(sk
, 1),
1245 (tcp_skb_is_last(sk
, skb
) ?
1246 tp
->nonagle
: TCP_NAGLE_PUSH
)));
1249 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1250 * which is put after SKB on the list. It is very much like
1251 * tcp_fragment() except that it may make several kinds of assumptions
1252 * in order to speed up the splitting operation. In particular, we
1253 * know that all the data is in scatter-gather pages, and that the
1254 * packet has never been sent out before (and thus is not cloned).
1256 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
,
1257 unsigned int mss_now
)
1259 struct sk_buff
*buff
;
1260 int nlen
= skb
->len
- len
;
1263 /* All of a TSO frame must be composed of paged data. */
1264 if (skb
->len
!= skb
->data_len
)
1265 return tcp_fragment(sk
, skb
, len
, mss_now
);
1267 buff
= sk_stream_alloc_skb(sk
, 0, GFP_ATOMIC
);
1268 if (unlikely(buff
== NULL
))
1271 sk
->sk_wmem_queued
+= buff
->truesize
;
1272 sk_mem_charge(sk
, buff
->truesize
);
1273 buff
->truesize
+= nlen
;
1274 skb
->truesize
-= nlen
;
1276 /* Correct the sequence numbers. */
1277 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1278 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1279 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1281 /* PSH and FIN should only be set in the second packet. */
1282 flags
= TCP_SKB_CB(skb
)->flags
;
1283 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
| TCPCB_FLAG_PSH
);
1284 TCP_SKB_CB(buff
)->flags
= flags
;
1286 /* This packet was never sent out yet, so no SACK bits. */
1287 TCP_SKB_CB(buff
)->sacked
= 0;
1289 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_PARTIAL
;
1290 skb_split(skb
, buff
, len
);
1292 /* Fix up tso_factor for both original and new SKB. */
1293 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1294 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1296 /* Link BUFF into the send queue. */
1297 skb_header_release(buff
);
1298 tcp_insert_write_queue_after(skb
, buff
, sk
);
1303 /* Try to defer sending, if possible, in order to minimize the amount
1304 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1306 * This algorithm is from John Heffner.
1308 static int tcp_tso_should_defer(struct sock
*sk
, struct sk_buff
*skb
)
1310 struct tcp_sock
*tp
= tcp_sk(sk
);
1311 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1312 u32 send_win
, cong_win
, limit
, in_flight
;
1314 if (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
)
1317 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
1320 /* Defer for less than two clock ticks. */
1321 if (tp
->tso_deferred
&&
1322 ((jiffies
<< 1) >> 1) - (tp
->tso_deferred
>> 1) > 1)
1325 in_flight
= tcp_packets_in_flight(tp
);
1327 BUG_ON(tcp_skb_pcount(skb
) <= 1 || (tp
->snd_cwnd
<= in_flight
));
1329 send_win
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1331 /* From in_flight test above, we know that cwnd > in_flight. */
1332 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
1334 limit
= min(send_win
, cong_win
);
1336 /* If a full-sized TSO skb can be sent, do it. */
1337 if (limit
>= sk
->sk_gso_max_size
)
1340 if (sysctl_tcp_tso_win_divisor
) {
1341 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
1343 /* If at least some fraction of a window is available,
1346 chunk
/= sysctl_tcp_tso_win_divisor
;
1350 /* Different approach, try not to defer past a single
1351 * ACK. Receiver should ACK every other full sized
1352 * frame, so if we have space for more than 3 frames
1355 if (limit
> tcp_max_burst(tp
) * tp
->mss_cache
)
1359 /* Ok, it looks like it is advisable to defer. */
1360 tp
->tso_deferred
= 1 | (jiffies
<< 1);
1365 tp
->tso_deferred
= 0;
1369 /* Create a new MTU probe if we are ready.
1370 * Returns 0 if we should wait to probe (no cwnd available),
1371 * 1 if a probe was sent,
1374 static int tcp_mtu_probe(struct sock
*sk
)
1376 struct tcp_sock
*tp
= tcp_sk(sk
);
1377 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1378 struct sk_buff
*skb
, *nskb
, *next
;
1385 /* Not currently probing/verifying,
1387 * have enough cwnd, and
1388 * not SACKing (the variable headers throw things off) */
1389 if (!icsk
->icsk_mtup
.enabled
||
1390 icsk
->icsk_mtup
.probe_size
||
1391 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
||
1392 tp
->snd_cwnd
< 11 ||
1393 tp
->rx_opt
.eff_sacks
)
1396 /* Very simple search strategy: just double the MSS. */
1397 mss_now
= tcp_current_mss(sk
, 0);
1398 probe_size
= 2 * tp
->mss_cache
;
1399 size_needed
= probe_size
+ (tp
->reordering
+ 1) * tp
->mss_cache
;
1400 if (probe_size
> tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_high
)) {
1401 /* TODO: set timer for probe_converge_event */
1405 /* Have enough data in the send queue to probe? */
1406 if (tp
->write_seq
- tp
->snd_nxt
< size_needed
)
1409 if (tp
->snd_wnd
< size_needed
)
1411 if (after(tp
->snd_nxt
+ size_needed
, tcp_wnd_end(tp
)))
1414 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1415 if (tcp_packets_in_flight(tp
) + 2 > tp
->snd_cwnd
) {
1416 if (!tcp_packets_in_flight(tp
))
1422 /* We're allowed to probe. Build it now. */
1423 if ((nskb
= sk_stream_alloc_skb(sk
, probe_size
, GFP_ATOMIC
)) == NULL
)
1425 sk
->sk_wmem_queued
+= nskb
->truesize
;
1426 sk_mem_charge(sk
, nskb
->truesize
);
1428 skb
= tcp_send_head(sk
);
1430 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(skb
)->seq
;
1431 TCP_SKB_CB(nskb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ probe_size
;
1432 TCP_SKB_CB(nskb
)->flags
= TCPCB_FLAG_ACK
;
1433 TCP_SKB_CB(nskb
)->sacked
= 0;
1435 nskb
->ip_summed
= skb
->ip_summed
;
1437 tcp_insert_write_queue_before(nskb
, skb
, sk
);
1440 tcp_for_write_queue_from_safe(skb
, next
, sk
) {
1441 copy
= min_t(int, skb
->len
, probe_size
- len
);
1442 if (nskb
->ip_summed
)
1443 skb_copy_bits(skb
, 0, skb_put(nskb
, copy
), copy
);
1445 nskb
->csum
= skb_copy_and_csum_bits(skb
, 0,
1446 skb_put(nskb
, copy
),
1449 if (skb
->len
<= copy
) {
1450 /* We've eaten all the data from this skb.
1452 TCP_SKB_CB(nskb
)->flags
|= TCP_SKB_CB(skb
)->flags
;
1453 tcp_unlink_write_queue(skb
, sk
);
1454 sk_wmem_free_skb(sk
, skb
);
1456 TCP_SKB_CB(nskb
)->flags
|= TCP_SKB_CB(skb
)->flags
&
1457 ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
1458 if (!skb_shinfo(skb
)->nr_frags
) {
1459 skb_pull(skb
, copy
);
1460 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1461 skb
->csum
= csum_partial(skb
->data
,
1464 __pskb_trim_head(skb
, copy
);
1465 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1467 TCP_SKB_CB(skb
)->seq
+= copy
;
1472 if (len
>= probe_size
)
1475 tcp_init_tso_segs(sk
, nskb
, nskb
->len
);
1477 /* We're ready to send. If this fails, the probe will
1478 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1479 TCP_SKB_CB(nskb
)->when
= tcp_time_stamp
;
1480 if (!tcp_transmit_skb(sk
, nskb
, 1, GFP_ATOMIC
)) {
1481 /* Decrement cwnd here because we are sending
1482 * effectively two packets. */
1484 tcp_event_new_data_sent(sk
, nskb
);
1486 icsk
->icsk_mtup
.probe_size
= tcp_mss_to_mtu(sk
, nskb
->len
);
1487 tp
->mtu_probe
.probe_seq_start
= TCP_SKB_CB(nskb
)->seq
;
1488 tp
->mtu_probe
.probe_seq_end
= TCP_SKB_CB(nskb
)->end_seq
;
1496 /* This routine writes packets to the network. It advances the
1497 * send_head. This happens as incoming acks open up the remote
1500 * Returns 1, if no segments are in flight and we have queued segments, but
1501 * cannot send anything now because of SWS or another problem.
1503 static int tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
)
1505 struct tcp_sock
*tp
= tcp_sk(sk
);
1506 struct sk_buff
*skb
;
1507 unsigned int tso_segs
, sent_pkts
;
1511 /* If we are closed, the bytes will have to remain here.
1512 * In time closedown will finish, we empty the write queue and all
1515 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
1520 /* Do MTU probing. */
1521 if ((result
= tcp_mtu_probe(sk
)) == 0) {
1523 } else if (result
> 0) {
1527 while ((skb
= tcp_send_head(sk
))) {
1530 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1533 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1537 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
)))
1540 if (tso_segs
== 1) {
1541 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
1542 (tcp_skb_is_last(sk
, skb
) ?
1543 nonagle
: TCP_NAGLE_PUSH
))))
1546 if (tcp_tso_should_defer(sk
, skb
))
1552 limit
= tcp_mss_split_point(sk
, skb
, mss_now
,
1555 if (skb
->len
> limit
&&
1556 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
)))
1559 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1561 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
)))
1564 /* Advance the send_head. This one is sent out.
1565 * This call will increment packets_out.
1567 tcp_event_new_data_sent(sk
, skb
);
1569 tcp_minshall_update(tp
, mss_now
, skb
);
1573 if (likely(sent_pkts
)) {
1574 tcp_cwnd_validate(sk
);
1577 return !tp
->packets_out
&& tcp_send_head(sk
);
1580 /* Push out any pending frames which were held back due to
1581 * TCP_CORK or attempt at coalescing tiny packets.
1582 * The socket must be locked by the caller.
1584 void __tcp_push_pending_frames(struct sock
*sk
, unsigned int cur_mss
,
1587 struct sk_buff
*skb
= tcp_send_head(sk
);
1590 if (tcp_write_xmit(sk
, cur_mss
, nonagle
))
1591 tcp_check_probe_timer(sk
);
1595 /* Send _single_ skb sitting at the send head. This function requires
1596 * true push pending frames to setup probe timer etc.
1598 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
1600 struct sk_buff
*skb
= tcp_send_head(sk
);
1601 unsigned int tso_segs
, cwnd_quota
;
1603 BUG_ON(!skb
|| skb
->len
< mss_now
);
1605 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1606 cwnd_quota
= tcp_snd_test(sk
, skb
, mss_now
, TCP_NAGLE_PUSH
);
1608 if (likely(cwnd_quota
)) {
1615 limit
= tcp_mss_split_point(sk
, skb
, mss_now
,
1618 if (skb
->len
> limit
&&
1619 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
)))
1622 /* Send it out now. */
1623 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1625 if (likely(!tcp_transmit_skb(sk
, skb
, 1, sk
->sk_allocation
))) {
1626 tcp_event_new_data_sent(sk
, skb
);
1627 tcp_cwnd_validate(sk
);
1633 /* This function returns the amount that we can raise the
1634 * usable window based on the following constraints
1636 * 1. The window can never be shrunk once it is offered (RFC 793)
1637 * 2. We limit memory per socket
1640 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1641 * RECV.NEXT + RCV.WIN fixed until:
1642 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1644 * i.e. don't raise the right edge of the window until you can raise
1645 * it at least MSS bytes.
1647 * Unfortunately, the recommended algorithm breaks header prediction,
1648 * since header prediction assumes th->window stays fixed.
1650 * Strictly speaking, keeping th->window fixed violates the receiver
1651 * side SWS prevention criteria. The problem is that under this rule
1652 * a stream of single byte packets will cause the right side of the
1653 * window to always advance by a single byte.
1655 * Of course, if the sender implements sender side SWS prevention
1656 * then this will not be a problem.
1658 * BSD seems to make the following compromise:
1660 * If the free space is less than the 1/4 of the maximum
1661 * space available and the free space is less than 1/2 mss,
1662 * then set the window to 0.
1663 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1664 * Otherwise, just prevent the window from shrinking
1665 * and from being larger than the largest representable value.
1667 * This prevents incremental opening of the window in the regime
1668 * where TCP is limited by the speed of the reader side taking
1669 * data out of the TCP receive queue. It does nothing about
1670 * those cases where the window is constrained on the sender side
1671 * because the pipeline is full.
1673 * BSD also seems to "accidentally" limit itself to windows that are a
1674 * multiple of MSS, at least until the free space gets quite small.
1675 * This would appear to be a side effect of the mbuf implementation.
1676 * Combining these two algorithms results in the observed behavior
1677 * of having a fixed window size at almost all times.
1679 * Below we obtain similar behavior by forcing the offered window to
1680 * a multiple of the mss when it is feasible to do so.
1682 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1683 * Regular options like TIMESTAMP are taken into account.
1685 u32
__tcp_select_window(struct sock
*sk
)
1687 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1688 struct tcp_sock
*tp
= tcp_sk(sk
);
1689 /* MSS for the peer's data. Previous versions used mss_clamp
1690 * here. I don't know if the value based on our guesses
1691 * of peer's MSS is better for the performance. It's more correct
1692 * but may be worse for the performance because of rcv_mss
1693 * fluctuations. --SAW 1998/11/1
1695 int mss
= icsk
->icsk_ack
.rcv_mss
;
1696 int free_space
= tcp_space(sk
);
1697 int full_space
= min_t(int, tp
->window_clamp
, tcp_full_space(sk
));
1700 if (mss
> full_space
)
1703 if (free_space
< (full_space
>> 1)) {
1704 icsk
->icsk_ack
.quick
= 0;
1706 if (tcp_memory_pressure
)
1707 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
,
1710 if (free_space
< mss
)
1714 if (free_space
> tp
->rcv_ssthresh
)
1715 free_space
= tp
->rcv_ssthresh
;
1717 /* Don't do rounding if we are using window scaling, since the
1718 * scaled window will not line up with the MSS boundary anyway.
1720 window
= tp
->rcv_wnd
;
1721 if (tp
->rx_opt
.rcv_wscale
) {
1722 window
= free_space
;
1724 /* Advertise enough space so that it won't get scaled away.
1725 * Import case: prevent zero window announcement if
1726 * 1<<rcv_wscale > mss.
1728 if (((window
>> tp
->rx_opt
.rcv_wscale
) << tp
->rx_opt
.rcv_wscale
) != window
)
1729 window
= (((window
>> tp
->rx_opt
.rcv_wscale
) + 1)
1730 << tp
->rx_opt
.rcv_wscale
);
1732 /* Get the largest window that is a nice multiple of mss.
1733 * Window clamp already applied above.
1734 * If our current window offering is within 1 mss of the
1735 * free space we just keep it. This prevents the divide
1736 * and multiply from happening most of the time.
1737 * We also don't do any window rounding when the free space
1740 if (window
<= free_space
- mss
|| window
> free_space
)
1741 window
= (free_space
/ mss
) * mss
;
1742 else if (mss
== full_space
&&
1743 free_space
> window
+ (full_space
>> 1))
1744 window
= free_space
;
1750 /* Attempt to collapse two adjacent SKB's during retransmission. */
1751 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*skb
,
1754 struct tcp_sock
*tp
= tcp_sk(sk
);
1755 struct sk_buff
*next_skb
= tcp_write_queue_next(sk
, skb
);
1756 int skb_size
, next_skb_size
;
1759 /* The first test we must make is that neither of these two
1760 * SKB's are still referenced by someone else.
1762 if (skb_cloned(skb
) || skb_cloned(next_skb
))
1765 skb_size
= skb
->len
;
1766 next_skb_size
= next_skb
->len
;
1767 flags
= TCP_SKB_CB(skb
)->flags
;
1769 /* Also punt if next skb has been SACK'd. */
1770 if (TCP_SKB_CB(next_skb
)->sacked
& TCPCB_SACKED_ACKED
)
1773 /* Next skb is out of window. */
1774 if (after(TCP_SKB_CB(next_skb
)->end_seq
, tcp_wnd_end(tp
)))
1777 /* Punt if not enough space exists in the first SKB for
1778 * the data in the second, or the total combined payload
1779 * would exceed the MSS.
1781 if ((next_skb_size
> skb_tailroom(skb
)) ||
1782 ((skb_size
+ next_skb_size
) > mss_now
))
1785 BUG_ON(tcp_skb_pcount(skb
) != 1 || tcp_skb_pcount(next_skb
) != 1);
1787 tcp_highest_sack_combine(sk
, next_skb
, skb
);
1789 /* Ok. We will be able to collapse the packet. */
1790 tcp_unlink_write_queue(next_skb
, sk
);
1792 skb_copy_from_linear_data(next_skb
, skb_put(skb
, next_skb_size
),
1795 if (next_skb
->ip_summed
== CHECKSUM_PARTIAL
)
1796 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1798 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1799 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
1801 /* Update sequence range on original skb. */
1802 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
1804 /* Merge over control information. */
1805 flags
|= TCP_SKB_CB(next_skb
)->flags
; /* This moves PSH/FIN etc. over */
1806 TCP_SKB_CB(skb
)->flags
= flags
;
1808 /* All done, get rid of second SKB and account for it so
1809 * packet counting does not break.
1811 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
& TCPCB_EVER_RETRANS
;
1812 if (TCP_SKB_CB(next_skb
)->sacked
& TCPCB_SACKED_RETRANS
)
1813 tp
->retrans_out
-= tcp_skb_pcount(next_skb
);
1814 if (TCP_SKB_CB(next_skb
)->sacked
& TCPCB_LOST
)
1815 tp
->lost_out
-= tcp_skb_pcount(next_skb
);
1816 /* Reno case is special. Sigh... */
1817 if (tcp_is_reno(tp
) && tp
->sacked_out
)
1818 tcp_dec_pcount_approx(&tp
->sacked_out
, next_skb
);
1820 tcp_adjust_fackets_out(sk
, next_skb
, tcp_skb_pcount(next_skb
));
1821 tp
->packets_out
-= tcp_skb_pcount(next_skb
);
1823 /* changed transmit queue under us so clear hints */
1824 tcp_clear_retrans_hints_partial(tp
);
1826 sk_wmem_free_skb(sk
, next_skb
);
1829 /* Do a simple retransmit without using the backoff mechanisms in
1830 * tcp_timer. This is used for path mtu discovery.
1831 * The socket is already locked here.
1833 void tcp_simple_retransmit(struct sock
*sk
)
1835 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1836 struct tcp_sock
*tp
= tcp_sk(sk
);
1837 struct sk_buff
*skb
;
1838 unsigned int mss
= tcp_current_mss(sk
, 0);
1841 tcp_for_write_queue(skb
, sk
) {
1842 if (skb
== tcp_send_head(sk
))
1844 if (skb
->len
> mss
&&
1845 !(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
1846 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
1847 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1848 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1850 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
)) {
1851 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1852 tp
->lost_out
+= tcp_skb_pcount(skb
);
1858 tcp_clear_all_retrans_hints(tp
);
1863 if (tcp_is_reno(tp
))
1864 tcp_limit_reno_sacked(tp
);
1866 tcp_verify_left_out(tp
);
1868 /* Don't muck with the congestion window here.
1869 * Reason is that we do not increase amount of _data_
1870 * in network, but units changed and effective
1871 * cwnd/ssthresh really reduced now.
1873 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
1874 tp
->high_seq
= tp
->snd_nxt
;
1875 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
1876 tp
->prior_ssthresh
= 0;
1877 tp
->undo_marker
= 0;
1878 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1880 tcp_xmit_retransmit_queue(sk
);
1883 /* This retransmits one SKB. Policy decisions and retransmit queue
1884 * state updates are done by the caller. Returns non-zero if an
1885 * error occurred which prevented the send.
1887 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
1889 struct tcp_sock
*tp
= tcp_sk(sk
);
1890 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1891 unsigned int cur_mss
;
1894 /* Inconslusive MTU probe */
1895 if (icsk
->icsk_mtup
.probe_size
) {
1896 icsk
->icsk_mtup
.probe_size
= 0;
1899 /* Do not sent more than we queued. 1/4 is reserved for possible
1900 * copying overhead: fragmentation, tunneling, mangling etc.
1902 if (atomic_read(&sk
->sk_wmem_alloc
) >
1903 min(sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2), sk
->sk_sndbuf
))
1906 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
1907 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1909 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
1913 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
1914 return -EHOSTUNREACH
; /* Routing failure or similar. */
1916 cur_mss
= tcp_current_mss(sk
, 0);
1918 /* If receiver has shrunk his window, and skb is out of
1919 * new window, do not retransmit it. The exception is the
1920 * case, when window is shrunk to zero. In this case
1921 * our retransmit serves as a zero window probe.
1923 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
))
1924 && TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
1927 if (skb
->len
> cur_mss
) {
1928 if (tcp_fragment(sk
, skb
, cur_mss
, cur_mss
))
1929 return -ENOMEM
; /* We'll try again later. */
1932 /* Collapse two adjacent packets if worthwhile and we can. */
1933 if (!(TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_SYN
) &&
1934 (skb
->len
< (cur_mss
>> 1)) &&
1935 (tcp_write_queue_next(sk
, skb
) != tcp_send_head(sk
)) &&
1936 (!tcp_skb_is_last(sk
, skb
)) &&
1937 (skb_shinfo(skb
)->nr_frags
== 0 &&
1938 skb_shinfo(tcp_write_queue_next(sk
, skb
))->nr_frags
== 0) &&
1939 (tcp_skb_pcount(skb
) == 1 &&
1940 tcp_skb_pcount(tcp_write_queue_next(sk
, skb
)) == 1) &&
1941 (sysctl_tcp_retrans_collapse
!= 0))
1942 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
1944 /* Some Solaris stacks overoptimize and ignore the FIN on a
1945 * retransmit when old data is attached. So strip it off
1946 * since it is cheap to do so and saves bytes on the network.
1949 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
) &&
1950 tp
->snd_una
== (TCP_SKB_CB(skb
)->end_seq
- 1)) {
1951 if (!pskb_trim(skb
, 0)) {
1952 /* Reuse, even though it does some unnecessary work */
1953 tcp_init_nondata_skb(skb
, TCP_SKB_CB(skb
)->end_seq
- 1,
1954 TCP_SKB_CB(skb
)->flags
);
1955 skb
->ip_summed
= CHECKSUM_NONE
;
1959 /* Make a copy, if the first transmission SKB clone we made
1960 * is still in somebody's hands, else make a clone.
1962 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1964 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
1967 /* Update global TCP statistics. */
1968 TCP_INC_STATS(sock_net(sk
), TCP_MIB_RETRANSSEGS
);
1970 tp
->total_retrans
++;
1972 #if FASTRETRANS_DEBUG > 0
1973 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
1974 if (net_ratelimit())
1975 printk(KERN_DEBUG
"retrans_out leaked.\n");
1978 if (!tp
->retrans_out
)
1979 tp
->lost_retrans_low
= tp
->snd_nxt
;
1980 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
1981 tp
->retrans_out
+= tcp_skb_pcount(skb
);
1983 /* Save stamp of the first retransmit. */
1984 if (!tp
->retrans_stamp
)
1985 tp
->retrans_stamp
= TCP_SKB_CB(skb
)->when
;
1989 /* snd_nxt is stored to detect loss of retransmitted segment,
1990 * see tcp_input.c tcp_sacktag_write_queue().
1992 TCP_SKB_CB(skb
)->ack_seq
= tp
->snd_nxt
;
1997 /* This gets called after a retransmit timeout, and the initially
1998 * retransmitted data is acknowledged. It tries to continue
1999 * resending the rest of the retransmit queue, until either
2000 * we've sent it all or the congestion window limit is reached.
2001 * If doing SACK, the first ACK which comes back for a timeout
2002 * based retransmit packet might feed us FACK information again.
2003 * If so, we use it to avoid unnecessarily retransmissions.
2005 void tcp_xmit_retransmit_queue(struct sock
*sk
)
2007 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2008 struct tcp_sock
*tp
= tcp_sk(sk
);
2009 struct sk_buff
*skb
;
2012 if (tp
->retransmit_skb_hint
) {
2013 skb
= tp
->retransmit_skb_hint
;
2014 packet_cnt
= tp
->retransmit_cnt_hint
;
2016 skb
= tcp_write_queue_head(sk
);
2020 /* First pass: retransmit lost packets. */
2022 tcp_for_write_queue_from(skb
, sk
) {
2023 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
2025 if (skb
== tcp_send_head(sk
))
2027 /* we could do better than to assign each time */
2028 tp
->retransmit_skb_hint
= skb
;
2029 tp
->retransmit_cnt_hint
= packet_cnt
;
2031 /* Assume this retransmit will generate
2032 * only one packet for congestion window
2033 * calculation purposes. This works because
2034 * tcp_retransmit_skb() will chop up the
2035 * packet to be MSS sized and all the
2036 * packet counting works out.
2038 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
2041 if (sacked
& TCPCB_LOST
) {
2042 if (!(sacked
& (TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))) {
2045 if (tcp_retransmit_skb(sk
, skb
)) {
2046 tp
->retransmit_skb_hint
= NULL
;
2049 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
2050 mib_idx
= LINUX_MIB_TCPFASTRETRANS
;
2052 mib_idx
= LINUX_MIB_TCPSLOWSTARTRETRANS
;
2053 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2055 if (skb
== tcp_write_queue_head(sk
))
2056 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2057 inet_csk(sk
)->icsk_rto
,
2061 packet_cnt
+= tcp_skb_pcount(skb
);
2062 if (packet_cnt
>= tp
->lost_out
)
2068 /* OK, demanded retransmission is finished. */
2070 /* Forward retransmissions are possible only during Recovery. */
2071 if (icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
2074 /* No forward retransmissions in Reno are possible. */
2075 if (tcp_is_reno(tp
))
2078 /* Yeah, we have to make difficult choice between forward transmission
2079 * and retransmission... Both ways have their merits...
2081 * For now we do not retransmit anything, while we have some new
2082 * segments to send. In the other cases, follow rule 3 for
2083 * NextSeg() specified in RFC3517.
2086 if (tcp_may_send_now(sk
))
2089 /* If nothing is SACKed, highest_sack in the loop won't be valid */
2090 if (!tp
->sacked_out
)
2093 if (tp
->forward_skb_hint
)
2094 skb
= tp
->forward_skb_hint
;
2096 skb
= tcp_write_queue_head(sk
);
2098 tcp_for_write_queue_from(skb
, sk
) {
2099 if (skb
== tcp_send_head(sk
))
2101 tp
->forward_skb_hint
= skb
;
2103 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_highest_sack_seq(tp
)))
2106 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
2109 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
)
2112 /* Ok, retransmit it. */
2113 if (tcp_retransmit_skb(sk
, skb
)) {
2114 tp
->forward_skb_hint
= NULL
;
2118 if (skb
== tcp_write_queue_head(sk
))
2119 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2120 inet_csk(sk
)->icsk_rto
,
2123 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPFORWARDRETRANS
);
2127 /* Send a fin. The caller locks the socket for us. This cannot be
2128 * allowed to fail queueing a FIN frame under any circumstances.
2130 void tcp_send_fin(struct sock
*sk
)
2132 struct tcp_sock
*tp
= tcp_sk(sk
);
2133 struct sk_buff
*skb
= tcp_write_queue_tail(sk
);
2136 /* Optimization, tack on the FIN if we have a queue of
2137 * unsent frames. But be careful about outgoing SACKS
2140 mss_now
= tcp_current_mss(sk
, 1);
2142 if (tcp_send_head(sk
) != NULL
) {
2143 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_FIN
;
2144 TCP_SKB_CB(skb
)->end_seq
++;
2147 /* Socket is locked, keep trying until memory is available. */
2149 skb
= alloc_skb_fclone(MAX_TCP_HEADER
, GFP_KERNEL
);
2155 /* Reserve space for headers and prepare control bits. */
2156 skb_reserve(skb
, MAX_TCP_HEADER
);
2157 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2158 tcp_init_nondata_skb(skb
, tp
->write_seq
,
2159 TCPCB_FLAG_ACK
| TCPCB_FLAG_FIN
);
2160 tcp_queue_skb(sk
, skb
);
2162 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_OFF
);
2165 /* We get here when a process closes a file descriptor (either due to
2166 * an explicit close() or as a byproduct of exit()'ing) and there
2167 * was unread data in the receive queue. This behavior is recommended
2168 * by RFC 2525, section 2.17. -DaveM
2170 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
2172 struct sk_buff
*skb
;
2174 /* NOTE: No TCP options attached and we never retransmit this. */
2175 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
2177 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
2181 /* Reserve space for headers and prepare control bits. */
2182 skb_reserve(skb
, MAX_TCP_HEADER
);
2183 tcp_init_nondata_skb(skb
, tcp_acceptable_seq(sk
),
2184 TCPCB_FLAG_ACK
| TCPCB_FLAG_RST
);
2186 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2187 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
2188 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
2190 TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTRSTS
);
2193 /* WARNING: This routine must only be called when we have already sent
2194 * a SYN packet that crossed the incoming SYN that caused this routine
2195 * to get called. If this assumption fails then the initial rcv_wnd
2196 * and rcv_wscale values will not be correct.
2198 int tcp_send_synack(struct sock
*sk
)
2200 struct sk_buff
*skb
;
2202 skb
= tcp_write_queue_head(sk
);
2203 if (skb
== NULL
|| !(TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_SYN
)) {
2204 printk(KERN_DEBUG
"tcp_send_synack: wrong queue state\n");
2207 if (!(TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_ACK
)) {
2208 if (skb_cloned(skb
)) {
2209 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
2212 tcp_unlink_write_queue(skb
, sk
);
2213 skb_header_release(nskb
);
2214 __tcp_add_write_queue_head(sk
, nskb
);
2215 sk_wmem_free_skb(sk
, skb
);
2216 sk
->sk_wmem_queued
+= nskb
->truesize
;
2217 sk_mem_charge(sk
, nskb
->truesize
);
2221 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_ACK
;
2222 TCP_ECN_send_synack(tcp_sk(sk
), skb
);
2224 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2225 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2229 * Prepare a SYN-ACK.
2231 struct sk_buff
*tcp_make_synack(struct sock
*sk
, struct dst_entry
*dst
,
2232 struct request_sock
*req
)
2234 struct inet_request_sock
*ireq
= inet_rsk(req
);
2235 struct tcp_sock
*tp
= tcp_sk(sk
);
2237 int tcp_header_size
;
2238 struct tcp_out_options opts
;
2239 struct sk_buff
*skb
;
2240 struct tcp_md5sig_key
*md5
;
2241 __u8
*md5_hash_location
;
2243 skb
= sock_wmalloc(sk
, MAX_TCP_HEADER
+ 15, 1, GFP_ATOMIC
);
2247 /* Reserve space for headers. */
2248 skb_reserve(skb
, MAX_TCP_HEADER
);
2250 skb
->dst
= dst_clone(dst
);
2252 if (req
->rcv_wnd
== 0) { /* ignored for retransmitted syns */
2254 /* Set this up on the first call only */
2255 req
->window_clamp
= tp
->window_clamp
? : dst_metric(dst
, RTAX_WINDOW
);
2256 /* tcp_full_space because it is guaranteed to be the first packet */
2257 tcp_select_initial_window(tcp_full_space(sk
),
2258 dst_metric(dst
, RTAX_ADVMSS
) - (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0),
2263 ireq
->rcv_wscale
= rcv_wscale
;
2266 memset(&opts
, 0, sizeof(opts
));
2267 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2268 tcp_header_size
= tcp_synack_options(sk
, req
,
2269 dst_metric(dst
, RTAX_ADVMSS
),
2271 sizeof(struct tcphdr
);
2273 skb_push(skb
, tcp_header_size
);
2274 skb_reset_transport_header(skb
);
2277 memset(th
, 0, sizeof(struct tcphdr
));
2280 TCP_ECN_make_synack(req
, th
);
2281 th
->source
= inet_sk(sk
)->sport
;
2282 th
->dest
= ireq
->rmt_port
;
2283 /* Setting of flags are superfluous here for callers (and ECE is
2284 * not even correctly set)
2286 tcp_init_nondata_skb(skb
, tcp_rsk(req
)->snt_isn
,
2287 TCPCB_FLAG_SYN
| TCPCB_FLAG_ACK
);
2288 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
2289 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_isn
+ 1);
2291 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2292 th
->window
= htons(min(req
->rcv_wnd
, 65535U));
2293 #ifdef CONFIG_SYN_COOKIES
2294 if (unlikely(req
->cookie_ts
))
2295 TCP_SKB_CB(skb
)->when
= cookie_init_timestamp(req
);
2298 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
, &md5_hash_location
);
2299 th
->doff
= (tcp_header_size
>> 2);
2300 TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
);
2302 #ifdef CONFIG_TCP_MD5SIG
2303 /* Okay, we have all we need - do the md5 hash if needed */
2305 tp
->af_specific
->calc_md5_hash(md5_hash_location
,
2306 md5
, NULL
, req
, skb
);
2314 * Do all connect socket setups that can be done AF independent.
2316 static void tcp_connect_init(struct sock
*sk
)
2318 struct dst_entry
*dst
= __sk_dst_get(sk
);
2319 struct tcp_sock
*tp
= tcp_sk(sk
);
2322 /* We'll fix this up when we get a response from the other end.
2323 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2325 tp
->tcp_header_len
= sizeof(struct tcphdr
) +
2326 (sysctl_tcp_timestamps
? TCPOLEN_TSTAMP_ALIGNED
: 0);
2328 #ifdef CONFIG_TCP_MD5SIG
2329 if (tp
->af_specific
->md5_lookup(sk
, sk
) != NULL
)
2330 tp
->tcp_header_len
+= TCPOLEN_MD5SIG_ALIGNED
;
2333 /* If user gave his TCP_MAXSEG, record it to clamp */
2334 if (tp
->rx_opt
.user_mss
)
2335 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
2338 tcp_sync_mss(sk
, dst_mtu(dst
));
2340 if (!tp
->window_clamp
)
2341 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
2342 tp
->advmss
= dst_metric(dst
, RTAX_ADVMSS
);
2343 tcp_initialize_rcv_mss(sk
);
2345 tcp_select_initial_window(tcp_full_space(sk
),
2346 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
2349 sysctl_tcp_window_scaling
,
2352 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
2353 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
2356 sock_reset_flag(sk
, SOCK_DONE
);
2358 tcp_init_wl(tp
, tp
->write_seq
, 0);
2359 tp
->snd_una
= tp
->write_seq
;
2360 tp
->snd_sml
= tp
->write_seq
;
2365 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
2366 inet_csk(sk
)->icsk_retransmits
= 0;
2367 tcp_clear_retrans(tp
);
2371 * Build a SYN and send it off.
2373 int tcp_connect(struct sock
*sk
)
2375 struct tcp_sock
*tp
= tcp_sk(sk
);
2376 struct sk_buff
*buff
;
2378 tcp_connect_init(sk
);
2380 buff
= alloc_skb_fclone(MAX_TCP_HEADER
+ 15, sk
->sk_allocation
);
2381 if (unlikely(buff
== NULL
))
2384 /* Reserve space for headers. */
2385 skb_reserve(buff
, MAX_TCP_HEADER
);
2387 tp
->snd_nxt
= tp
->write_seq
;
2388 tcp_init_nondata_skb(buff
, tp
->write_seq
++, TCPCB_FLAG_SYN
);
2389 TCP_ECN_send_syn(sk
, buff
);
2392 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2393 tp
->retrans_stamp
= TCP_SKB_CB(buff
)->when
;
2394 skb_header_release(buff
);
2395 __tcp_add_write_queue_tail(sk
, buff
);
2396 sk
->sk_wmem_queued
+= buff
->truesize
;
2397 sk_mem_charge(sk
, buff
->truesize
);
2398 tp
->packets_out
+= tcp_skb_pcount(buff
);
2399 tcp_transmit_skb(sk
, buff
, 1, GFP_KERNEL
);
2401 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2402 * in order to make this packet get counted in tcpOutSegs.
2404 tp
->snd_nxt
= tp
->write_seq
;
2405 tp
->pushed_seq
= tp
->write_seq
;
2406 TCP_INC_STATS(sock_net(sk
), TCP_MIB_ACTIVEOPENS
);
2408 /* Timer for repeating the SYN until an answer. */
2409 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2410 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2414 /* Send out a delayed ack, the caller does the policy checking
2415 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2418 void tcp_send_delayed_ack(struct sock
*sk
)
2420 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2421 int ato
= icsk
->icsk_ack
.ato
;
2422 unsigned long timeout
;
2424 if (ato
> TCP_DELACK_MIN
) {
2425 const struct tcp_sock
*tp
= tcp_sk(sk
);
2426 int max_ato
= HZ
/ 2;
2428 if (icsk
->icsk_ack
.pingpong
||
2429 (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
2430 max_ato
= TCP_DELACK_MAX
;
2432 /* Slow path, intersegment interval is "high". */
2434 /* If some rtt estimate is known, use it to bound delayed ack.
2435 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2439 int rtt
= max(tp
->srtt
>> 3, TCP_DELACK_MIN
);
2445 ato
= min(ato
, max_ato
);
2448 /* Stay within the limit we were given */
2449 timeout
= jiffies
+ ato
;
2451 /* Use new timeout only if there wasn't a older one earlier. */
2452 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
2453 /* If delack timer was blocked or is about to expire,
2456 if (icsk
->icsk_ack
.blocked
||
2457 time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
2462 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
2463 timeout
= icsk
->icsk_ack
.timeout
;
2465 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
2466 icsk
->icsk_ack
.timeout
= timeout
;
2467 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
2470 /* This routine sends an ack and also updates the window. */
2471 void tcp_send_ack(struct sock
*sk
)
2473 struct sk_buff
*buff
;
2475 /* If we have been reset, we may not send again. */
2476 if (sk
->sk_state
== TCP_CLOSE
)
2479 /* We are not putting this on the write queue, so
2480 * tcp_transmit_skb() will set the ownership to this
2483 buff
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2485 inet_csk_schedule_ack(sk
);
2486 inet_csk(sk
)->icsk_ack
.ato
= TCP_ATO_MIN
;
2487 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
2488 TCP_DELACK_MAX
, TCP_RTO_MAX
);
2492 /* Reserve space for headers and prepare control bits. */
2493 skb_reserve(buff
, MAX_TCP_HEADER
);
2494 tcp_init_nondata_skb(buff
, tcp_acceptable_seq(sk
), TCPCB_FLAG_ACK
);
2496 /* Send it off, this clears delayed acks for us. */
2497 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2498 tcp_transmit_skb(sk
, buff
, 0, GFP_ATOMIC
);
2501 /* This routine sends a packet with an out of date sequence
2502 * number. It assumes the other end will try to ack it.
2504 * Question: what should we make while urgent mode?
2505 * 4.4BSD forces sending single byte of data. We cannot send
2506 * out of window data, because we have SND.NXT==SND.MAX...
2508 * Current solution: to send TWO zero-length segments in urgent mode:
2509 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2510 * out-of-date with SND.UNA-1 to probe window.
2512 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
)
2514 struct tcp_sock
*tp
= tcp_sk(sk
);
2515 struct sk_buff
*skb
;
2517 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2518 skb
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2522 /* Reserve space for headers and set control bits. */
2523 skb_reserve(skb
, MAX_TCP_HEADER
);
2524 /* Use a previous sequence. This should cause the other
2525 * end to send an ack. Don't queue or clone SKB, just
2528 tcp_init_nondata_skb(skb
, tp
->snd_una
- !urgent
, TCPCB_FLAG_ACK
);
2529 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2530 return tcp_transmit_skb(sk
, skb
, 0, GFP_ATOMIC
);
2533 int tcp_write_wakeup(struct sock
*sk
)
2535 struct tcp_sock
*tp
= tcp_sk(sk
);
2536 struct sk_buff
*skb
;
2538 if (sk
->sk_state
== TCP_CLOSE
)
2541 if ((skb
= tcp_send_head(sk
)) != NULL
&&
2542 before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
))) {
2544 unsigned int mss
= tcp_current_mss(sk
, 0);
2545 unsigned int seg_size
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
2547 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
2548 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
2550 /* We are probing the opening of a window
2551 * but the window size is != 0
2552 * must have been a result SWS avoidance ( sender )
2554 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
2556 seg_size
= min(seg_size
, mss
);
2557 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2558 if (tcp_fragment(sk
, skb
, seg_size
, mss
))
2560 } else if (!tcp_skb_pcount(skb
))
2561 tcp_set_skb_tso_segs(sk
, skb
, mss
);
2563 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2564 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2565 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2567 tcp_event_new_data_sent(sk
, skb
);
2571 between(tp
->snd_up
, tp
->snd_una
+ 1, tp
->snd_una
+ 0xFFFF))
2572 tcp_xmit_probe_skb(sk
, 1);
2573 return tcp_xmit_probe_skb(sk
, 0);
2577 /* A window probe timeout has occurred. If window is not closed send
2578 * a partial packet else a zero probe.
2580 void tcp_send_probe0(struct sock
*sk
)
2582 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2583 struct tcp_sock
*tp
= tcp_sk(sk
);
2586 err
= tcp_write_wakeup(sk
);
2588 if (tp
->packets_out
|| !tcp_send_head(sk
)) {
2589 /* Cancel probe timer, if it is not required. */
2590 icsk
->icsk_probes_out
= 0;
2591 icsk
->icsk_backoff
= 0;
2596 if (icsk
->icsk_backoff
< sysctl_tcp_retries2
)
2597 icsk
->icsk_backoff
++;
2598 icsk
->icsk_probes_out
++;
2599 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2600 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
2603 /* If packet was not sent due to local congestion,
2604 * do not backoff and do not remember icsk_probes_out.
2605 * Let local senders to fight for local resources.
2607 * Use accumulated backoff yet.
2609 if (!icsk
->icsk_probes_out
)
2610 icsk
->icsk_probes_out
= 1;
2611 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2612 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
,
2613 TCP_RESOURCE_PROBE_INTERVAL
),
2618 EXPORT_SYMBOL(tcp_select_initial_window
);
2619 EXPORT_SYMBOL(tcp_connect
);
2620 EXPORT_SYMBOL(tcp_make_synack
);
2621 EXPORT_SYMBOL(tcp_simple_retransmit
);
2622 EXPORT_SYMBOL(tcp_sync_mss
);
2623 EXPORT_SYMBOL(tcp_mtup_init
);