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 int sysctl_tcp_cookie_size __read_mostly
= 0; /* TCP_COOKIE_MAX */
63 EXPORT_SYMBOL_GPL(sysctl_tcp_cookie_size
);
66 /* Account for new data that has been sent to the network. */
67 static void tcp_event_new_data_sent(struct sock
*sk
, struct sk_buff
*skb
)
69 struct tcp_sock
*tp
= tcp_sk(sk
);
70 unsigned int prior_packets
= tp
->packets_out
;
72 tcp_advance_send_head(sk
, skb
);
73 tp
->snd_nxt
= TCP_SKB_CB(skb
)->end_seq
;
75 /* Don't override Nagle indefinately with F-RTO */
76 if (tp
->frto_counter
== 2)
79 tp
->packets_out
+= tcp_skb_pcount(skb
);
81 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
82 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
85 /* SND.NXT, if window was not shrunk.
86 * If window has been shrunk, what should we make? It is not clear at all.
87 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
88 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
89 * invalid. OK, let's make this for now:
91 static inline __u32
tcp_acceptable_seq(struct sock
*sk
)
93 struct tcp_sock
*tp
= tcp_sk(sk
);
95 if (!before(tcp_wnd_end(tp
), tp
->snd_nxt
))
98 return tcp_wnd_end(tp
);
101 /* Calculate mss to advertise in SYN segment.
102 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
104 * 1. It is independent of path mtu.
105 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
106 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
107 * attached devices, because some buggy hosts are confused by
109 * 4. We do not make 3, we advertise MSS, calculated from first
110 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
111 * This may be overridden via information stored in routing table.
112 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
113 * probably even Jumbo".
115 static __u16
tcp_advertise_mss(struct sock
*sk
)
117 struct tcp_sock
*tp
= tcp_sk(sk
);
118 struct dst_entry
*dst
= __sk_dst_get(sk
);
119 int mss
= tp
->advmss
;
121 if (dst
&& dst_metric(dst
, RTAX_ADVMSS
) < mss
) {
122 mss
= dst_metric(dst
, RTAX_ADVMSS
);
129 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
130 * This is the first part of cwnd validation mechanism. */
131 static void tcp_cwnd_restart(struct sock
*sk
, struct dst_entry
*dst
)
133 struct tcp_sock
*tp
= tcp_sk(sk
);
134 s32 delta
= tcp_time_stamp
- tp
->lsndtime
;
135 u32 restart_cwnd
= tcp_init_cwnd(tp
, dst
);
136 u32 cwnd
= tp
->snd_cwnd
;
138 tcp_ca_event(sk
, CA_EVENT_CWND_RESTART
);
140 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
141 restart_cwnd
= min(restart_cwnd
, cwnd
);
143 while ((delta
-= inet_csk(sk
)->icsk_rto
) > 0 && cwnd
> restart_cwnd
)
145 tp
->snd_cwnd
= max(cwnd
, restart_cwnd
);
146 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
147 tp
->snd_cwnd_used
= 0;
150 /* Congestion state accounting after a packet has been sent. */
151 static void tcp_event_data_sent(struct tcp_sock
*tp
,
152 struct sk_buff
*skb
, struct sock
*sk
)
154 struct inet_connection_sock
*icsk
= inet_csk(sk
);
155 const u32 now
= tcp_time_stamp
;
157 if (sysctl_tcp_slow_start_after_idle
&&
158 (!tp
->packets_out
&& (s32
)(now
- tp
->lsndtime
) > icsk
->icsk_rto
))
159 tcp_cwnd_restart(sk
, __sk_dst_get(sk
));
163 /* If it is a reply for ato after last received
164 * packet, enter pingpong mode.
166 if ((u32
)(now
- icsk
->icsk_ack
.lrcvtime
) < icsk
->icsk_ack
.ato
)
167 icsk
->icsk_ack
.pingpong
= 1;
170 /* Account for an ACK we sent. */
171 static inline void tcp_event_ack_sent(struct sock
*sk
, unsigned int pkts
)
173 tcp_dec_quickack_mode(sk
, pkts
);
174 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_DACK
);
177 /* Determine a window scaling and initial window to offer.
178 * Based on the assumption that the given amount of space
179 * will be offered. Store the results in the tp structure.
180 * NOTE: for smooth operation initial space offering should
181 * be a multiple of mss if possible. We assume here that mss >= 1.
182 * This MUST be enforced by all callers.
184 void tcp_select_initial_window(int __space
, __u32 mss
,
185 __u32
*rcv_wnd
, __u32
*window_clamp
,
186 int wscale_ok
, __u8
*rcv_wscale
)
188 unsigned int space
= (__space
< 0 ? 0 : __space
);
190 /* If no clamp set the clamp to the max possible scaled window */
191 if (*window_clamp
== 0)
192 (*window_clamp
) = (65535 << 14);
193 space
= min(*window_clamp
, space
);
195 /* Quantize space offering to a multiple of mss if possible. */
197 space
= (space
/ mss
) * mss
;
199 /* NOTE: offering an initial window larger than 32767
200 * will break some buggy TCP stacks. If the admin tells us
201 * it is likely we could be speaking with such a buggy stack
202 * we will truncate our initial window offering to 32K-1
203 * unless the remote has sent us a window scaling option,
204 * which we interpret as a sign the remote TCP is not
205 * misinterpreting the window field as a signed quantity.
207 if (sysctl_tcp_workaround_signed_windows
)
208 (*rcv_wnd
) = min(space
, MAX_TCP_WINDOW
);
214 /* Set window scaling on max possible window
215 * See RFC1323 for an explanation of the limit to 14
217 space
= max_t(u32
, sysctl_tcp_rmem
[2], sysctl_rmem_max
);
218 space
= min_t(u32
, space
, *window_clamp
);
219 while (space
> 65535 && (*rcv_wscale
) < 14) {
225 /* Set initial window to value enough for senders,
226 * following RFC2414. Senders, not following this RFC,
227 * will be satisfied with 2.
229 if (mss
> (1 << *rcv_wscale
)) {
235 if (*rcv_wnd
> init_cwnd
* mss
)
236 *rcv_wnd
= init_cwnd
* mss
;
239 /* Set the clamp no higher than max representable value */
240 (*window_clamp
) = min(65535U << (*rcv_wscale
), *window_clamp
);
243 /* Chose a new window to advertise, update state in tcp_sock for the
244 * socket, and return result with RFC1323 scaling applied. The return
245 * value can be stuffed directly into th->window for an outgoing
248 static u16
tcp_select_window(struct sock
*sk
)
250 struct tcp_sock
*tp
= tcp_sk(sk
);
251 u32 cur_win
= tcp_receive_window(tp
);
252 u32 new_win
= __tcp_select_window(sk
);
254 /* Never shrink the offered window */
255 if (new_win
< cur_win
) {
256 /* Danger Will Robinson!
257 * Don't update rcv_wup/rcv_wnd here or else
258 * we will not be able to advertise a zero
259 * window in time. --DaveM
261 * Relax Will Robinson.
263 new_win
= ALIGN(cur_win
, 1 << tp
->rx_opt
.rcv_wscale
);
265 tp
->rcv_wnd
= new_win
;
266 tp
->rcv_wup
= tp
->rcv_nxt
;
268 /* Make sure we do not exceed the maximum possible
271 if (!tp
->rx_opt
.rcv_wscale
&& sysctl_tcp_workaround_signed_windows
)
272 new_win
= min(new_win
, MAX_TCP_WINDOW
);
274 new_win
= min(new_win
, (65535U << tp
->rx_opt
.rcv_wscale
));
276 /* RFC1323 scaling applied */
277 new_win
>>= tp
->rx_opt
.rcv_wscale
;
279 /* If we advertise zero window, disable fast path. */
286 /* Packet ECN state for a SYN-ACK */
287 static inline void TCP_ECN_send_synack(struct tcp_sock
*tp
, struct sk_buff
*skb
)
289 TCP_SKB_CB(skb
)->flags
&= ~TCPCB_FLAG_CWR
;
290 if (!(tp
->ecn_flags
& TCP_ECN_OK
))
291 TCP_SKB_CB(skb
)->flags
&= ~TCPCB_FLAG_ECE
;
294 /* Packet ECN state for a SYN. */
295 static inline void TCP_ECN_send_syn(struct sock
*sk
, struct sk_buff
*skb
)
297 struct tcp_sock
*tp
= tcp_sk(sk
);
300 if (sysctl_tcp_ecn
== 1) {
301 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_ECE
| TCPCB_FLAG_CWR
;
302 tp
->ecn_flags
= TCP_ECN_OK
;
306 static __inline__
void
307 TCP_ECN_make_synack(struct request_sock
*req
, struct tcphdr
*th
)
309 if (inet_rsk(req
)->ecn_ok
)
313 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
316 static inline void TCP_ECN_send(struct sock
*sk
, struct sk_buff
*skb
,
319 struct tcp_sock
*tp
= tcp_sk(sk
);
321 if (tp
->ecn_flags
& TCP_ECN_OK
) {
322 /* Not-retransmitted data segment: set ECT and inject CWR. */
323 if (skb
->len
!= tcp_header_len
&&
324 !before(TCP_SKB_CB(skb
)->seq
, tp
->snd_nxt
)) {
326 if (tp
->ecn_flags
& TCP_ECN_QUEUE_CWR
) {
327 tp
->ecn_flags
&= ~TCP_ECN_QUEUE_CWR
;
328 tcp_hdr(skb
)->cwr
= 1;
329 skb_shinfo(skb
)->gso_type
|= SKB_GSO_TCP_ECN
;
332 /* ACK or retransmitted segment: clear ECT|CE */
333 INET_ECN_dontxmit(sk
);
335 if (tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)
336 tcp_hdr(skb
)->ece
= 1;
340 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
341 * auto increment end seqno.
343 static void tcp_init_nondata_skb(struct sk_buff
*skb
, u32 seq
, u8 flags
)
347 TCP_SKB_CB(skb
)->flags
= flags
;
348 TCP_SKB_CB(skb
)->sacked
= 0;
350 skb_shinfo(skb
)->gso_segs
= 1;
351 skb_shinfo(skb
)->gso_size
= 0;
352 skb_shinfo(skb
)->gso_type
= 0;
354 TCP_SKB_CB(skb
)->seq
= seq
;
355 if (flags
& (TCPCB_FLAG_SYN
| TCPCB_FLAG_FIN
))
357 TCP_SKB_CB(skb
)->end_seq
= seq
;
360 static inline int tcp_urg_mode(const struct tcp_sock
*tp
)
362 return tp
->snd_una
!= tp
->snd_up
;
365 #define OPTION_SACK_ADVERTISE (1 << 0)
366 #define OPTION_TS (1 << 1)
367 #define OPTION_MD5 (1 << 2)
368 #define OPTION_WSCALE (1 << 3)
369 #define OPTION_COOKIE_EXTENSION (1 << 4)
371 struct tcp_out_options
{
372 u8 options
; /* bit field of OPTION_* */
373 u8 ws
; /* window scale, 0 to disable */
374 u8 num_sack_blocks
; /* number of SACK blocks to include */
375 u8 hash_size
; /* bytes in hash_location */
376 u16 mss
; /* 0 to disable */
377 __u32 tsval
, tsecr
; /* need to include OPTION_TS */
378 __u8
*hash_location
; /* temporary pointer, overloaded */
381 /* The sysctl int routines are generic, so check consistency here.
383 static u8
tcp_cookie_size_check(u8 desired
)
386 /* previously specified */
389 if (sysctl_tcp_cookie_size
<= 0) {
390 /* no default specified */
393 if (sysctl_tcp_cookie_size
<= TCP_COOKIE_MIN
) {
394 /* value too small, specify minimum */
395 return TCP_COOKIE_MIN
;
397 if (sysctl_tcp_cookie_size
>= TCP_COOKIE_MAX
) {
398 /* value too large, specify maximum */
399 return TCP_COOKIE_MAX
;
401 if (0x1 & sysctl_tcp_cookie_size
) {
402 /* 8-bit multiple, illegal, fix it */
403 return (u8
)(sysctl_tcp_cookie_size
+ 0x1);
405 return (u8
)sysctl_tcp_cookie_size
;
408 /* Write previously computed TCP options to the packet.
410 * Beware: Something in the Internet is very sensitive to the ordering of
411 * TCP options, we learned this through the hard way, so be careful here.
412 * Luckily we can at least blame others for their non-compliance but from
413 * inter-operatibility perspective it seems that we're somewhat stuck with
414 * the ordering which we have been using if we want to keep working with
415 * those broken things (not that it currently hurts anybody as there isn't
416 * particular reason why the ordering would need to be changed).
418 * At least SACK_PERM as the first option is known to lead to a disaster
419 * (but it may well be that other scenarios fail similarly).
421 static void tcp_options_write(__be32
*ptr
, struct tcp_sock
*tp
,
422 struct tcp_out_options
*opts
)
424 u8 options
= opts
->options
; /* mungable copy */
426 /* Having both authentication and cookies for security is redundant,
427 * and there's certainly not enough room. Instead, the cookie-less
428 * extension variant is proposed.
430 * Consider the pessimal case with authentication. The options
432 * COOKIE|MD5(20) + MSS(4) + SACK|TS(12) + WSCALE(4) == 40
434 if (unlikely(OPTION_MD5
& options
)) {
435 if (unlikely(OPTION_COOKIE_EXTENSION
& options
)) {
436 *ptr
++ = htonl((TCPOPT_COOKIE
<< 24) |
437 (TCPOLEN_COOKIE_BASE
<< 16) |
438 (TCPOPT_MD5SIG
<< 8) |
441 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
443 (TCPOPT_MD5SIG
<< 8) |
446 options
&= ~OPTION_COOKIE_EXTENSION
;
447 /* overload cookie hash location */
448 opts
->hash_location
= (__u8
*)ptr
;
452 if (unlikely(opts
->mss
)) {
453 *ptr
++ = htonl((TCPOPT_MSS
<< 24) |
454 (TCPOLEN_MSS
<< 16) |
458 if (likely(OPTION_TS
& options
)) {
459 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
460 *ptr
++ = htonl((TCPOPT_SACK_PERM
<< 24) |
461 (TCPOLEN_SACK_PERM
<< 16) |
462 (TCPOPT_TIMESTAMP
<< 8) |
464 options
&= ~OPTION_SACK_ADVERTISE
;
466 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
468 (TCPOPT_TIMESTAMP
<< 8) |
471 *ptr
++ = htonl(opts
->tsval
);
472 *ptr
++ = htonl(opts
->tsecr
);
475 /* Specification requires after timestamp, so do it now.
477 * Consider the pessimal case without authentication. The options
479 * MSS(4) + SACK|TS(12) + COOKIE(20) + WSCALE(4) == 40
481 if (unlikely(OPTION_COOKIE_EXTENSION
& options
)) {
482 __u8
*cookie_copy
= opts
->hash_location
;
483 u8 cookie_size
= opts
->hash_size
;
485 /* 8-bit multiple handled in tcp_cookie_size_check() above,
488 if (0x2 & cookie_size
) {
489 __u8
*p
= (__u8
*)ptr
;
491 /* 16-bit multiple */
492 *p
++ = TCPOPT_COOKIE
;
493 *p
++ = TCPOLEN_COOKIE_BASE
+ cookie_size
;
494 *p
++ = *cookie_copy
++;
495 *p
++ = *cookie_copy
++;
499 /* 32-bit multiple */
500 *ptr
++ = htonl(((TCPOPT_NOP
<< 24) |
502 (TCPOPT_COOKIE
<< 8) |
503 TCPOLEN_COOKIE_BASE
) +
507 if (cookie_size
> 0) {
508 memcpy(ptr
, cookie_copy
, cookie_size
);
509 ptr
+= (cookie_size
/ 4);
513 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
514 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
516 (TCPOPT_SACK_PERM
<< 8) |
520 if (unlikely(OPTION_WSCALE
& options
)) {
521 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
522 (TCPOPT_WINDOW
<< 16) |
523 (TCPOLEN_WINDOW
<< 8) |
527 if (unlikely(opts
->num_sack_blocks
)) {
528 struct tcp_sack_block
*sp
= tp
->rx_opt
.dsack
?
529 tp
->duplicate_sack
: tp
->selective_acks
;
532 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
535 (TCPOLEN_SACK_BASE
+ (opts
->num_sack_blocks
*
536 TCPOLEN_SACK_PERBLOCK
)));
538 for (this_sack
= 0; this_sack
< opts
->num_sack_blocks
;
540 *ptr
++ = htonl(sp
[this_sack
].start_seq
);
541 *ptr
++ = htonl(sp
[this_sack
].end_seq
);
544 tp
->rx_opt
.dsack
= 0;
548 /* Compute TCP options for SYN packets. This is not the final
549 * network wire format yet.
551 static unsigned tcp_syn_options(struct sock
*sk
, struct sk_buff
*skb
,
552 struct tcp_out_options
*opts
,
553 struct tcp_md5sig_key
**md5
) {
554 struct tcp_sock
*tp
= tcp_sk(sk
);
555 struct tcp_cookie_values
*cvp
= tp
->cookie_values
;
556 unsigned remaining
= MAX_TCP_OPTION_SPACE
;
557 u8 cookie_size
= (!tp
->rx_opt
.cookie_out_never
&& cvp
!= NULL
) ?
558 tcp_cookie_size_check(cvp
->cookie_desired
) :
561 #ifdef CONFIG_TCP_MD5SIG
562 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
564 opts
->options
|= OPTION_MD5
;
565 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
571 /* We always get an MSS option. The option bytes which will be seen in
572 * normal data packets should timestamps be used, must be in the MSS
573 * advertised. But we subtract them from tp->mss_cache so that
574 * calculations in tcp_sendmsg are simpler etc. So account for this
575 * fact here if necessary. If we don't do this correctly, as a
576 * receiver we won't recognize data packets as being full sized when we
577 * should, and thus we won't abide by the delayed ACK rules correctly.
578 * SACKs don't matter, we never delay an ACK when we have any of those
580 opts
->mss
= tcp_advertise_mss(sk
);
581 remaining
-= TCPOLEN_MSS_ALIGNED
;
583 if (likely(sysctl_tcp_timestamps
&& *md5
== NULL
)) {
584 opts
->options
|= OPTION_TS
;
585 opts
->tsval
= TCP_SKB_CB(skb
)->when
;
586 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
587 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
589 if (likely(sysctl_tcp_window_scaling
)) {
590 opts
->ws
= tp
->rx_opt
.rcv_wscale
;
591 opts
->options
|= OPTION_WSCALE
;
592 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
594 if (likely(sysctl_tcp_sack
)) {
595 opts
->options
|= OPTION_SACK_ADVERTISE
;
596 if (unlikely(!(OPTION_TS
& opts
->options
)))
597 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
600 /* Note that timestamps are required by the specification.
602 * Odd numbers of bytes are prohibited by the specification, ensuring
603 * that the cookie is 16-bit aligned, and the resulting cookie pair is
607 (OPTION_TS
& opts
->options
) &&
609 int need
= TCPOLEN_COOKIE_BASE
+ cookie_size
;
612 /* 32-bit multiple */
613 need
+= 2; /* NOPs */
615 if (need
> remaining
) {
616 /* try shrinking cookie to fit */
621 while (need
> remaining
&& TCP_COOKIE_MIN
<= cookie_size
) {
625 if (TCP_COOKIE_MIN
<= cookie_size
) {
626 opts
->options
|= OPTION_COOKIE_EXTENSION
;
627 opts
->hash_location
= (__u8
*)&cvp
->cookie_pair
[0];
628 opts
->hash_size
= cookie_size
;
630 /* Remember for future incarnations. */
631 cvp
->cookie_desired
= cookie_size
;
633 if (cvp
->cookie_desired
!= cvp
->cookie_pair_size
) {
634 /* Currently use random bytes as a nonce,
635 * assuming these are completely unpredictable
636 * by hostile users of the same system.
638 get_random_bytes(&cvp
->cookie_pair
[0],
640 cvp
->cookie_pair_size
= cookie_size
;
646 return MAX_TCP_OPTION_SPACE
- remaining
;
649 /* Set up TCP options for SYN-ACKs. */
650 static unsigned tcp_synack_options(struct sock
*sk
,
651 struct request_sock
*req
,
652 unsigned mss
, struct sk_buff
*skb
,
653 struct tcp_out_options
*opts
,
654 struct tcp_md5sig_key
**md5
,
655 struct tcp_extend_values
*xvp
)
657 struct inet_request_sock
*ireq
= inet_rsk(req
);
658 unsigned remaining
= MAX_TCP_OPTION_SPACE
;
659 u8 cookie_plus
= (xvp
!= NULL
&& !xvp
->cookie_out_never
) ?
662 bool doing_ts
= ireq
->tstamp_ok
;
664 #ifdef CONFIG_TCP_MD5SIG
665 *md5
= tcp_rsk(req
)->af_specific
->md5_lookup(sk
, req
);
667 opts
->options
|= OPTION_MD5
;
668 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
670 /* We can't fit any SACK blocks in a packet with MD5 + TS
671 * options. There was discussion about disabling SACK
672 * rather than TS in order to fit in better with old,
673 * buggy kernels, but that was deemed to be unnecessary.
675 doing_ts
&= !ireq
->sack_ok
;
681 /* We always send an MSS option. */
683 remaining
-= TCPOLEN_MSS_ALIGNED
;
685 if (likely(ireq
->wscale_ok
)) {
686 opts
->ws
= ireq
->rcv_wscale
;
687 opts
->options
|= OPTION_WSCALE
;
688 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
690 if (likely(doing_ts
)) {
691 opts
->options
|= OPTION_TS
;
692 opts
->tsval
= TCP_SKB_CB(skb
)->when
;
693 opts
->tsecr
= req
->ts_recent
;
694 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
696 if (likely(ireq
->sack_ok
)) {
697 opts
->options
|= OPTION_SACK_ADVERTISE
;
698 if (unlikely(!doing_ts
))
699 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
702 /* Similar rationale to tcp_syn_options() applies here, too.
703 * If the <SYN> options fit, the same options should fit now!
707 cookie_plus
> TCPOLEN_COOKIE_BASE
) {
708 int need
= cookie_plus
; /* has TCPOLEN_COOKIE_BASE */
711 /* 32-bit multiple */
712 need
+= 2; /* NOPs */
714 if (need
<= remaining
) {
715 opts
->options
|= OPTION_COOKIE_EXTENSION
;
716 opts
->hash_size
= cookie_plus
- TCPOLEN_COOKIE_BASE
;
719 /* There's no error return, so flag it. */
720 xvp
->cookie_out_never
= 1; /* true */
724 return MAX_TCP_OPTION_SPACE
- remaining
;
727 /* Compute TCP options for ESTABLISHED sockets. This is not the
728 * final wire format yet.
730 static unsigned tcp_established_options(struct sock
*sk
, struct sk_buff
*skb
,
731 struct tcp_out_options
*opts
,
732 struct tcp_md5sig_key
**md5
) {
733 struct tcp_skb_cb
*tcb
= skb
? TCP_SKB_CB(skb
) : NULL
;
734 struct tcp_sock
*tp
= tcp_sk(sk
);
736 unsigned int eff_sacks
;
738 #ifdef CONFIG_TCP_MD5SIG
739 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
740 if (unlikely(*md5
)) {
741 opts
->options
|= OPTION_MD5
;
742 size
+= TCPOLEN_MD5SIG_ALIGNED
;
748 if (likely(tp
->rx_opt
.tstamp_ok
)) {
749 opts
->options
|= OPTION_TS
;
750 opts
->tsval
= tcb
? tcb
->when
: 0;
751 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
752 size
+= TCPOLEN_TSTAMP_ALIGNED
;
755 eff_sacks
= tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
;
756 if (unlikely(eff_sacks
)) {
757 const unsigned remaining
= MAX_TCP_OPTION_SPACE
- size
;
758 opts
->num_sack_blocks
=
759 min_t(unsigned, eff_sacks
,
760 (remaining
- TCPOLEN_SACK_BASE_ALIGNED
) /
761 TCPOLEN_SACK_PERBLOCK
);
762 size
+= TCPOLEN_SACK_BASE_ALIGNED
+
763 opts
->num_sack_blocks
* TCPOLEN_SACK_PERBLOCK
;
769 /* This routine actually transmits TCP packets queued in by
770 * tcp_do_sendmsg(). This is used by both the initial
771 * transmission and possible later retransmissions.
772 * All SKB's seen here are completely headerless. It is our
773 * job to build the TCP header, and pass the packet down to
774 * IP so it can do the same plus pass the packet off to the
777 * We are working here with either a clone of the original
778 * SKB, or a fresh unique copy made by the retransmit engine.
780 static int tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int clone_it
,
783 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
784 struct inet_sock
*inet
;
786 struct tcp_skb_cb
*tcb
;
787 struct tcp_out_options opts
;
788 unsigned tcp_options_size
, tcp_header_size
;
789 struct tcp_md5sig_key
*md5
;
793 BUG_ON(!skb
|| !tcp_skb_pcount(skb
));
795 /* If congestion control is doing timestamping, we must
796 * take such a timestamp before we potentially clone/copy.
798 if (icsk
->icsk_ca_ops
->flags
& TCP_CONG_RTT_STAMP
)
799 __net_timestamp(skb
);
801 if (likely(clone_it
)) {
802 if (unlikely(skb_cloned(skb
)))
803 skb
= pskb_copy(skb
, gfp_mask
);
805 skb
= skb_clone(skb
, gfp_mask
);
812 tcb
= TCP_SKB_CB(skb
);
813 memset(&opts
, 0, sizeof(opts
));
815 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
))
816 tcp_options_size
= tcp_syn_options(sk
, skb
, &opts
, &md5
);
818 tcp_options_size
= tcp_established_options(sk
, skb
, &opts
,
820 tcp_header_size
= tcp_options_size
+ sizeof(struct tcphdr
);
822 if (tcp_packets_in_flight(tp
) == 0)
823 tcp_ca_event(sk
, CA_EVENT_TX_START
);
825 skb_push(skb
, tcp_header_size
);
826 skb_reset_transport_header(skb
);
827 skb_set_owner_w(skb
, sk
);
829 /* Build TCP header and checksum it. */
831 th
->source
= inet
->inet_sport
;
832 th
->dest
= inet
->inet_dport
;
833 th
->seq
= htonl(tcb
->seq
);
834 th
->ack_seq
= htonl(tp
->rcv_nxt
);
835 *(((__be16
*)th
) + 6) = htons(((tcp_header_size
>> 2) << 12) |
838 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
839 /* RFC1323: The window in SYN & SYN/ACK segments
842 th
->window
= htons(min(tp
->rcv_wnd
, 65535U));
844 th
->window
= htons(tcp_select_window(sk
));
849 /* The urg_mode check is necessary during a below snd_una win probe */
850 if (unlikely(tcp_urg_mode(tp
) && before(tcb
->seq
, tp
->snd_up
))) {
851 if (before(tp
->snd_up
, tcb
->seq
+ 0x10000)) {
852 th
->urg_ptr
= htons(tp
->snd_up
- tcb
->seq
);
854 } else if (after(tcb
->seq
+ 0xFFFF, tp
->snd_nxt
)) {
855 th
->urg_ptr
= 0xFFFF;
860 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
);
861 if (likely((tcb
->flags
& TCPCB_FLAG_SYN
) == 0))
862 TCP_ECN_send(sk
, skb
, tcp_header_size
);
864 #ifdef CONFIG_TCP_MD5SIG
865 /* Calculate the MD5 hash, as we have all we need now */
867 sk
->sk_route_caps
&= ~NETIF_F_GSO_MASK
;
868 tp
->af_specific
->calc_md5_hash(opts
.hash_location
,
873 icsk
->icsk_af_ops
->send_check(sk
, skb
->len
, skb
);
875 if (likely(tcb
->flags
& TCPCB_FLAG_ACK
))
876 tcp_event_ack_sent(sk
, tcp_skb_pcount(skb
));
878 if (skb
->len
!= tcp_header_size
)
879 tcp_event_data_sent(tp
, skb
, sk
);
881 if (after(tcb
->end_seq
, tp
->snd_nxt
) || tcb
->seq
== tcb
->end_seq
)
882 TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
);
884 err
= icsk
->icsk_af_ops
->queue_xmit(skb
, 0);
885 if (likely(err
<= 0))
888 tcp_enter_cwr(sk
, 1);
890 return net_xmit_eval(err
);
893 /* This routine just queues the buffer for sending.
895 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
896 * otherwise socket can stall.
898 static void tcp_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
900 struct tcp_sock
*tp
= tcp_sk(sk
);
902 /* Advance write_seq and place onto the write_queue. */
903 tp
->write_seq
= TCP_SKB_CB(skb
)->end_seq
;
904 skb_header_release(skb
);
905 tcp_add_write_queue_tail(sk
, skb
);
906 sk
->sk_wmem_queued
+= skb
->truesize
;
907 sk_mem_charge(sk
, skb
->truesize
);
910 /* Initialize TSO segments for a packet. */
911 static void tcp_set_skb_tso_segs(struct sock
*sk
, struct sk_buff
*skb
,
912 unsigned int mss_now
)
914 if (skb
->len
<= mss_now
|| !sk_can_gso(sk
) ||
915 skb
->ip_summed
== CHECKSUM_NONE
) {
916 /* Avoid the costly divide in the normal
919 skb_shinfo(skb
)->gso_segs
= 1;
920 skb_shinfo(skb
)->gso_size
= 0;
921 skb_shinfo(skb
)->gso_type
= 0;
923 skb_shinfo(skb
)->gso_segs
= DIV_ROUND_UP(skb
->len
, mss_now
);
924 skb_shinfo(skb
)->gso_size
= mss_now
;
925 skb_shinfo(skb
)->gso_type
= sk
->sk_gso_type
;
929 /* When a modification to fackets out becomes necessary, we need to check
930 * skb is counted to fackets_out or not.
932 static void tcp_adjust_fackets_out(struct sock
*sk
, struct sk_buff
*skb
,
935 struct tcp_sock
*tp
= tcp_sk(sk
);
937 if (!tp
->sacked_out
|| tcp_is_reno(tp
))
940 if (after(tcp_highest_sack_seq(tp
), TCP_SKB_CB(skb
)->seq
))
941 tp
->fackets_out
-= decr
;
944 /* Pcount in the middle of the write queue got changed, we need to do various
945 * tweaks to fix counters
947 static void tcp_adjust_pcount(struct sock
*sk
, struct sk_buff
*skb
, int decr
)
949 struct tcp_sock
*tp
= tcp_sk(sk
);
951 tp
->packets_out
-= decr
;
953 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
954 tp
->sacked_out
-= decr
;
955 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
956 tp
->retrans_out
-= decr
;
957 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
)
958 tp
->lost_out
-= decr
;
960 /* Reno case is special. Sigh... */
961 if (tcp_is_reno(tp
) && decr
> 0)
962 tp
->sacked_out
-= min_t(u32
, tp
->sacked_out
, decr
);
964 tcp_adjust_fackets_out(sk
, skb
, decr
);
966 if (tp
->lost_skb_hint
&&
967 before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
) &&
968 (tcp_is_fack(tp
) || (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)))
969 tp
->lost_cnt_hint
-= decr
;
971 tcp_verify_left_out(tp
);
974 /* Function to create two new TCP segments. Shrinks the given segment
975 * to the specified size and appends a new segment with the rest of the
976 * packet to the list. This won't be called frequently, I hope.
977 * Remember, these are still headerless SKBs at this point.
979 int tcp_fragment(struct sock
*sk
, struct sk_buff
*skb
, u32 len
,
980 unsigned int mss_now
)
982 struct tcp_sock
*tp
= tcp_sk(sk
);
983 struct sk_buff
*buff
;
984 int nsize
, old_factor
;
988 BUG_ON(len
> skb
->len
);
990 nsize
= skb_headlen(skb
) - len
;
994 if (skb_cloned(skb
) &&
995 skb_is_nonlinear(skb
) &&
996 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
999 /* Get a new skb... force flag on. */
1000 buff
= sk_stream_alloc_skb(sk
, nsize
, GFP_ATOMIC
);
1002 return -ENOMEM
; /* We'll just try again later. */
1004 sk
->sk_wmem_queued
+= buff
->truesize
;
1005 sk_mem_charge(sk
, buff
->truesize
);
1006 nlen
= skb
->len
- len
- nsize
;
1007 buff
->truesize
+= nlen
;
1008 skb
->truesize
-= nlen
;
1010 /* Correct the sequence numbers. */
1011 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1012 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1013 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1015 /* PSH and FIN should only be set in the second packet. */
1016 flags
= TCP_SKB_CB(skb
)->flags
;
1017 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
| TCPCB_FLAG_PSH
);
1018 TCP_SKB_CB(buff
)->flags
= flags
;
1019 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
1021 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_PARTIAL
) {
1022 /* Copy and checksum data tail into the new buffer. */
1023 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
,
1024 skb_put(buff
, nsize
),
1029 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
1031 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1032 skb_split(skb
, buff
, len
);
1035 buff
->ip_summed
= skb
->ip_summed
;
1037 /* Looks stupid, but our code really uses when of
1038 * skbs, which it never sent before. --ANK
1040 TCP_SKB_CB(buff
)->when
= TCP_SKB_CB(skb
)->when
;
1041 buff
->tstamp
= skb
->tstamp
;
1043 old_factor
= tcp_skb_pcount(skb
);
1045 /* Fix up tso_factor for both original and new SKB. */
1046 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1047 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1049 /* If this packet has been sent out already, we must
1050 * adjust the various packet counters.
1052 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
1053 int diff
= old_factor
- tcp_skb_pcount(skb
) -
1054 tcp_skb_pcount(buff
);
1057 tcp_adjust_pcount(sk
, skb
, diff
);
1060 /* Link BUFF into the send queue. */
1061 skb_header_release(buff
);
1062 tcp_insert_write_queue_after(skb
, buff
, sk
);
1067 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1068 * eventually). The difference is that pulled data not copied, but
1069 * immediately discarded.
1071 static void __pskb_trim_head(struct sk_buff
*skb
, int len
)
1077 for (i
= 0; i
< skb_shinfo(skb
)->nr_frags
; i
++) {
1078 if (skb_shinfo(skb
)->frags
[i
].size
<= eat
) {
1079 put_page(skb_shinfo(skb
)->frags
[i
].page
);
1080 eat
-= skb_shinfo(skb
)->frags
[i
].size
;
1082 skb_shinfo(skb
)->frags
[k
] = skb_shinfo(skb
)->frags
[i
];
1084 skb_shinfo(skb
)->frags
[k
].page_offset
+= eat
;
1085 skb_shinfo(skb
)->frags
[k
].size
-= eat
;
1091 skb_shinfo(skb
)->nr_frags
= k
;
1093 skb_reset_tail_pointer(skb
);
1094 skb
->data_len
-= len
;
1095 skb
->len
= skb
->data_len
;
1098 /* Remove acked data from a packet in the transmit queue. */
1099 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
1101 if (skb_cloned(skb
) && pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
1104 /* If len == headlen, we avoid __skb_pull to preserve alignment. */
1105 if (unlikely(len
< skb_headlen(skb
)))
1106 __skb_pull(skb
, len
);
1108 __pskb_trim_head(skb
, len
- skb_headlen(skb
));
1110 TCP_SKB_CB(skb
)->seq
+= len
;
1111 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1113 skb
->truesize
-= len
;
1114 sk
->sk_wmem_queued
-= len
;
1115 sk_mem_uncharge(sk
, len
);
1116 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
1118 /* Any change of skb->len requires recalculation of tso
1121 if (tcp_skb_pcount(skb
) > 1)
1122 tcp_set_skb_tso_segs(sk
, skb
, tcp_current_mss(sk
));
1127 /* Calculate MSS. Not accounting for SACKs here. */
1128 int tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1130 struct tcp_sock
*tp
= tcp_sk(sk
);
1131 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1134 /* Calculate base mss without TCP options:
1135 It is MMS_S - sizeof(tcphdr) of rfc1122
1137 mss_now
= pmtu
- icsk
->icsk_af_ops
->net_header_len
- sizeof(struct tcphdr
);
1139 /* Clamp it (mss_clamp does not include tcp options) */
1140 if (mss_now
> tp
->rx_opt
.mss_clamp
)
1141 mss_now
= tp
->rx_opt
.mss_clamp
;
1143 /* Now subtract optional transport overhead */
1144 mss_now
-= icsk
->icsk_ext_hdr_len
;
1146 /* Then reserve room for full set of TCP options and 8 bytes of data */
1150 /* Now subtract TCP options size, not including SACKs */
1151 mss_now
-= tp
->tcp_header_len
- sizeof(struct tcphdr
);
1156 /* Inverse of above */
1157 int tcp_mss_to_mtu(struct sock
*sk
, int mss
)
1159 struct tcp_sock
*tp
= tcp_sk(sk
);
1160 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1164 tp
->tcp_header_len
+
1165 icsk
->icsk_ext_hdr_len
+
1166 icsk
->icsk_af_ops
->net_header_len
;
1171 /* MTU probing init per socket */
1172 void tcp_mtup_init(struct sock
*sk
)
1174 struct tcp_sock
*tp
= tcp_sk(sk
);
1175 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1177 icsk
->icsk_mtup
.enabled
= sysctl_tcp_mtu_probing
> 1;
1178 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+ sizeof(struct tcphdr
) +
1179 icsk
->icsk_af_ops
->net_header_len
;
1180 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, sysctl_tcp_base_mss
);
1181 icsk
->icsk_mtup
.probe_size
= 0;
1184 /* This function synchronize snd mss to current pmtu/exthdr set.
1186 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1187 for TCP options, but includes only bare TCP header.
1189 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1190 It is minimum of user_mss and mss received with SYN.
1191 It also does not include TCP options.
1193 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1195 tp->mss_cache is current effective sending mss, including
1196 all tcp options except for SACKs. It is evaluated,
1197 taking into account current pmtu, but never exceeds
1198 tp->rx_opt.mss_clamp.
1200 NOTE1. rfc1122 clearly states that advertised MSS
1201 DOES NOT include either tcp or ip options.
1203 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1204 are READ ONLY outside this function. --ANK (980731)
1206 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
1208 struct tcp_sock
*tp
= tcp_sk(sk
);
1209 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1212 if (icsk
->icsk_mtup
.search_high
> pmtu
)
1213 icsk
->icsk_mtup
.search_high
= pmtu
;
1215 mss_now
= tcp_mtu_to_mss(sk
, pmtu
);
1216 mss_now
= tcp_bound_to_half_wnd(tp
, mss_now
);
1218 /* And store cached results */
1219 icsk
->icsk_pmtu_cookie
= pmtu
;
1220 if (icsk
->icsk_mtup
.enabled
)
1221 mss_now
= min(mss_now
, tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
));
1222 tp
->mss_cache
= mss_now
;
1227 /* Compute the current effective MSS, taking SACKs and IP options,
1228 * and even PMTU discovery events into account.
1230 unsigned int tcp_current_mss(struct sock
*sk
)
1232 struct tcp_sock
*tp
= tcp_sk(sk
);
1233 struct dst_entry
*dst
= __sk_dst_get(sk
);
1235 unsigned header_len
;
1236 struct tcp_out_options opts
;
1237 struct tcp_md5sig_key
*md5
;
1239 mss_now
= tp
->mss_cache
;
1242 u32 mtu
= dst_mtu(dst
);
1243 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
1244 mss_now
= tcp_sync_mss(sk
, mtu
);
1247 header_len
= tcp_established_options(sk
, NULL
, &opts
, &md5
) +
1248 sizeof(struct tcphdr
);
1249 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1250 * some common options. If this is an odd packet (because we have SACK
1251 * blocks etc) then our calculated header_len will be different, and
1252 * we have to adjust mss_now correspondingly */
1253 if (header_len
!= tp
->tcp_header_len
) {
1254 int delta
= (int) header_len
- tp
->tcp_header_len
;
1261 /* Congestion window validation. (RFC2861) */
1262 static void tcp_cwnd_validate(struct sock
*sk
)
1264 struct tcp_sock
*tp
= tcp_sk(sk
);
1266 if (tp
->packets_out
>= tp
->snd_cwnd
) {
1267 /* Network is feed fully. */
1268 tp
->snd_cwnd_used
= 0;
1269 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1271 /* Network starves. */
1272 if (tp
->packets_out
> tp
->snd_cwnd_used
)
1273 tp
->snd_cwnd_used
= tp
->packets_out
;
1275 if (sysctl_tcp_slow_start_after_idle
&&
1276 (s32
)(tcp_time_stamp
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
)
1277 tcp_cwnd_application_limited(sk
);
1281 /* Returns the portion of skb which can be sent right away without
1282 * introducing MSS oddities to segment boundaries. In rare cases where
1283 * mss_now != mss_cache, we will request caller to create a small skb
1284 * per input skb which could be mostly avoided here (if desired).
1286 * We explicitly want to create a request for splitting write queue tail
1287 * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1288 * thus all the complexity (cwnd_len is always MSS multiple which we
1289 * return whenever allowed by the other factors). Basically we need the
1290 * modulo only when the receiver window alone is the limiting factor or
1291 * when we would be allowed to send the split-due-to-Nagle skb fully.
1293 static unsigned int tcp_mss_split_point(struct sock
*sk
, struct sk_buff
*skb
,
1294 unsigned int mss_now
, unsigned int cwnd
)
1296 struct tcp_sock
*tp
= tcp_sk(sk
);
1297 u32 needed
, window
, cwnd_len
;
1299 window
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1300 cwnd_len
= mss_now
* cwnd
;
1302 if (likely(cwnd_len
<= window
&& skb
!= tcp_write_queue_tail(sk
)))
1305 needed
= min(skb
->len
, window
);
1307 if (cwnd_len
<= needed
)
1310 return needed
- needed
% mss_now
;
1313 /* Can at least one segment of SKB be sent right now, according to the
1314 * congestion window rules? If so, return how many segments are allowed.
1316 static inline unsigned int tcp_cwnd_test(struct tcp_sock
*tp
,
1317 struct sk_buff
*skb
)
1319 u32 in_flight
, cwnd
;
1321 /* Don't be strict about the congestion window for the final FIN. */
1322 if ((TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
) &&
1323 tcp_skb_pcount(skb
) == 1)
1326 in_flight
= tcp_packets_in_flight(tp
);
1327 cwnd
= tp
->snd_cwnd
;
1328 if (in_flight
< cwnd
)
1329 return (cwnd
- in_flight
);
1334 /* Intialize TSO state of a skb.
1335 * This must be invoked the first time we consider transmitting
1336 * SKB onto the wire.
1338 static int tcp_init_tso_segs(struct sock
*sk
, struct sk_buff
*skb
,
1339 unsigned int mss_now
)
1341 int tso_segs
= tcp_skb_pcount(skb
);
1343 if (!tso_segs
|| (tso_segs
> 1 && tcp_skb_mss(skb
) != mss_now
)) {
1344 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1345 tso_segs
= tcp_skb_pcount(skb
);
1350 /* Minshall's variant of the Nagle send check. */
1351 static inline int tcp_minshall_check(const struct tcp_sock
*tp
)
1353 return after(tp
->snd_sml
, tp
->snd_una
) &&
1354 !after(tp
->snd_sml
, tp
->snd_nxt
);
1357 /* Return 0, if packet can be sent now without violation Nagle's rules:
1358 * 1. It is full sized.
1359 * 2. Or it contains FIN. (already checked by caller)
1360 * 3. Or TCP_NODELAY was set.
1361 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1362 * With Minshall's modification: all sent small packets are ACKed.
1364 static inline int tcp_nagle_check(const struct tcp_sock
*tp
,
1365 const struct sk_buff
*skb
,
1366 unsigned mss_now
, int nonagle
)
1368 return (skb
->len
< mss_now
&&
1369 ((nonagle
& TCP_NAGLE_CORK
) ||
1370 (!nonagle
&& tp
->packets_out
&& tcp_minshall_check(tp
))));
1373 /* Return non-zero if the Nagle test allows this packet to be
1376 static inline int tcp_nagle_test(struct tcp_sock
*tp
, struct sk_buff
*skb
,
1377 unsigned int cur_mss
, int nonagle
)
1379 /* Nagle rule does not apply to frames, which sit in the middle of the
1380 * write_queue (they have no chances to get new data).
1382 * This is implemented in the callers, where they modify the 'nonagle'
1383 * argument based upon the location of SKB in the send queue.
1385 if (nonagle
& TCP_NAGLE_PUSH
)
1388 /* Don't use the nagle rule for urgent data (or for the final FIN).
1389 * Nagle can be ignored during F-RTO too (see RFC4138).
1391 if (tcp_urg_mode(tp
) || (tp
->frto_counter
== 2) ||
1392 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
))
1395 if (!tcp_nagle_check(tp
, skb
, cur_mss
, nonagle
))
1401 /* Does at least the first segment of SKB fit into the send window? */
1402 static inline int tcp_snd_wnd_test(struct tcp_sock
*tp
, struct sk_buff
*skb
,
1403 unsigned int cur_mss
)
1405 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1407 if (skb
->len
> cur_mss
)
1408 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
1410 return !after(end_seq
, tcp_wnd_end(tp
));
1413 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1414 * should be put on the wire right now. If so, it returns the number of
1415 * packets allowed by the congestion window.
1417 static unsigned int tcp_snd_test(struct sock
*sk
, struct sk_buff
*skb
,
1418 unsigned int cur_mss
, int nonagle
)
1420 struct tcp_sock
*tp
= tcp_sk(sk
);
1421 unsigned int cwnd_quota
;
1423 tcp_init_tso_segs(sk
, skb
, cur_mss
);
1425 if (!tcp_nagle_test(tp
, skb
, cur_mss
, nonagle
))
1428 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1429 if (cwnd_quota
&& !tcp_snd_wnd_test(tp
, skb
, cur_mss
))
1435 /* Test if sending is allowed right now. */
1436 int tcp_may_send_now(struct sock
*sk
)
1438 struct tcp_sock
*tp
= tcp_sk(sk
);
1439 struct sk_buff
*skb
= tcp_send_head(sk
);
1442 tcp_snd_test(sk
, skb
, tcp_current_mss(sk
),
1443 (tcp_skb_is_last(sk
, skb
) ?
1444 tp
->nonagle
: TCP_NAGLE_PUSH
)));
1447 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1448 * which is put after SKB on the list. It is very much like
1449 * tcp_fragment() except that it may make several kinds of assumptions
1450 * in order to speed up the splitting operation. In particular, we
1451 * know that all the data is in scatter-gather pages, and that the
1452 * packet has never been sent out before (and thus is not cloned).
1454 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
,
1455 unsigned int mss_now
)
1457 struct sk_buff
*buff
;
1458 int nlen
= skb
->len
- len
;
1461 /* All of a TSO frame must be composed of paged data. */
1462 if (skb
->len
!= skb
->data_len
)
1463 return tcp_fragment(sk
, skb
, len
, mss_now
);
1465 buff
= sk_stream_alloc_skb(sk
, 0, GFP_ATOMIC
);
1466 if (unlikely(buff
== NULL
))
1469 sk
->sk_wmem_queued
+= buff
->truesize
;
1470 sk_mem_charge(sk
, buff
->truesize
);
1471 buff
->truesize
+= nlen
;
1472 skb
->truesize
-= nlen
;
1474 /* Correct the sequence numbers. */
1475 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1476 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1477 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1479 /* PSH and FIN should only be set in the second packet. */
1480 flags
= TCP_SKB_CB(skb
)->flags
;
1481 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
| TCPCB_FLAG_PSH
);
1482 TCP_SKB_CB(buff
)->flags
= flags
;
1484 /* This packet was never sent out yet, so no SACK bits. */
1485 TCP_SKB_CB(buff
)->sacked
= 0;
1487 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_PARTIAL
;
1488 skb_split(skb
, buff
, len
);
1490 /* Fix up tso_factor for both original and new SKB. */
1491 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1492 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1494 /* Link BUFF into the send queue. */
1495 skb_header_release(buff
);
1496 tcp_insert_write_queue_after(skb
, buff
, sk
);
1501 /* Try to defer sending, if possible, in order to minimize the amount
1502 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1504 * This algorithm is from John Heffner.
1506 static int tcp_tso_should_defer(struct sock
*sk
, struct sk_buff
*skb
)
1508 struct tcp_sock
*tp
= tcp_sk(sk
);
1509 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1510 u32 send_win
, cong_win
, limit
, in_flight
;
1512 if (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
)
1515 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
1518 /* Defer for less than two clock ticks. */
1519 if (tp
->tso_deferred
&&
1520 (((u32
)jiffies
<< 1) >> 1) - (tp
->tso_deferred
>> 1) > 1)
1523 in_flight
= tcp_packets_in_flight(tp
);
1525 BUG_ON(tcp_skb_pcount(skb
) <= 1 || (tp
->snd_cwnd
<= in_flight
));
1527 send_win
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1529 /* From in_flight test above, we know that cwnd > in_flight. */
1530 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
1532 limit
= min(send_win
, cong_win
);
1534 /* If a full-sized TSO skb can be sent, do it. */
1535 if (limit
>= sk
->sk_gso_max_size
)
1538 /* Middle in queue won't get any more data, full sendable already? */
1539 if ((skb
!= tcp_write_queue_tail(sk
)) && (limit
>= skb
->len
))
1542 if (sysctl_tcp_tso_win_divisor
) {
1543 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
1545 /* If at least some fraction of a window is available,
1548 chunk
/= sysctl_tcp_tso_win_divisor
;
1552 /* Different approach, try not to defer past a single
1553 * ACK. Receiver should ACK every other full sized
1554 * frame, so if we have space for more than 3 frames
1557 if (limit
> tcp_max_burst(tp
) * tp
->mss_cache
)
1561 /* Ok, it looks like it is advisable to defer. */
1562 tp
->tso_deferred
= 1 | (jiffies
<< 1);
1567 tp
->tso_deferred
= 0;
1571 /* Create a new MTU probe if we are ready.
1572 * MTU probe is regularly attempting to increase the path MTU by
1573 * deliberately sending larger packets. This discovers routing
1574 * changes resulting in larger path MTUs.
1576 * Returns 0 if we should wait to probe (no cwnd available),
1577 * 1 if a probe was sent,
1580 static int tcp_mtu_probe(struct sock
*sk
)
1582 struct tcp_sock
*tp
= tcp_sk(sk
);
1583 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1584 struct sk_buff
*skb
, *nskb
, *next
;
1591 /* Not currently probing/verifying,
1593 * have enough cwnd, and
1594 * not SACKing (the variable headers throw things off) */
1595 if (!icsk
->icsk_mtup
.enabled
||
1596 icsk
->icsk_mtup
.probe_size
||
1597 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
||
1598 tp
->snd_cwnd
< 11 ||
1599 tp
->rx_opt
.num_sacks
|| tp
->rx_opt
.dsack
)
1602 /* Very simple search strategy: just double the MSS. */
1603 mss_now
= tcp_current_mss(sk
);
1604 probe_size
= 2 * tp
->mss_cache
;
1605 size_needed
= probe_size
+ (tp
->reordering
+ 1) * tp
->mss_cache
;
1606 if (probe_size
> tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_high
)) {
1607 /* TODO: set timer for probe_converge_event */
1611 /* Have enough data in the send queue to probe? */
1612 if (tp
->write_seq
- tp
->snd_nxt
< size_needed
)
1615 if (tp
->snd_wnd
< size_needed
)
1617 if (after(tp
->snd_nxt
+ size_needed
, tcp_wnd_end(tp
)))
1620 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1621 if (tcp_packets_in_flight(tp
) + 2 > tp
->snd_cwnd
) {
1622 if (!tcp_packets_in_flight(tp
))
1628 /* We're allowed to probe. Build it now. */
1629 if ((nskb
= sk_stream_alloc_skb(sk
, probe_size
, GFP_ATOMIC
)) == NULL
)
1631 sk
->sk_wmem_queued
+= nskb
->truesize
;
1632 sk_mem_charge(sk
, nskb
->truesize
);
1634 skb
= tcp_send_head(sk
);
1636 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(skb
)->seq
;
1637 TCP_SKB_CB(nskb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ probe_size
;
1638 TCP_SKB_CB(nskb
)->flags
= TCPCB_FLAG_ACK
;
1639 TCP_SKB_CB(nskb
)->sacked
= 0;
1641 nskb
->ip_summed
= skb
->ip_summed
;
1643 tcp_insert_write_queue_before(nskb
, skb
, sk
);
1646 tcp_for_write_queue_from_safe(skb
, next
, sk
) {
1647 copy
= min_t(int, skb
->len
, probe_size
- len
);
1648 if (nskb
->ip_summed
)
1649 skb_copy_bits(skb
, 0, skb_put(nskb
, copy
), copy
);
1651 nskb
->csum
= skb_copy_and_csum_bits(skb
, 0,
1652 skb_put(nskb
, copy
),
1655 if (skb
->len
<= copy
) {
1656 /* We've eaten all the data from this skb.
1658 TCP_SKB_CB(nskb
)->flags
|= TCP_SKB_CB(skb
)->flags
;
1659 tcp_unlink_write_queue(skb
, sk
);
1660 sk_wmem_free_skb(sk
, skb
);
1662 TCP_SKB_CB(nskb
)->flags
|= TCP_SKB_CB(skb
)->flags
&
1663 ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
1664 if (!skb_shinfo(skb
)->nr_frags
) {
1665 skb_pull(skb
, copy
);
1666 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1667 skb
->csum
= csum_partial(skb
->data
,
1670 __pskb_trim_head(skb
, copy
);
1671 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1673 TCP_SKB_CB(skb
)->seq
+= copy
;
1678 if (len
>= probe_size
)
1681 tcp_init_tso_segs(sk
, nskb
, nskb
->len
);
1683 /* We're ready to send. If this fails, the probe will
1684 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1685 TCP_SKB_CB(nskb
)->when
= tcp_time_stamp
;
1686 if (!tcp_transmit_skb(sk
, nskb
, 1, GFP_ATOMIC
)) {
1687 /* Decrement cwnd here because we are sending
1688 * effectively two packets. */
1690 tcp_event_new_data_sent(sk
, nskb
);
1692 icsk
->icsk_mtup
.probe_size
= tcp_mss_to_mtu(sk
, nskb
->len
);
1693 tp
->mtu_probe
.probe_seq_start
= TCP_SKB_CB(nskb
)->seq
;
1694 tp
->mtu_probe
.probe_seq_end
= TCP_SKB_CB(nskb
)->end_seq
;
1702 /* This routine writes packets to the network. It advances the
1703 * send_head. This happens as incoming acks open up the remote
1706 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1707 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1708 * account rare use of URG, this is not a big flaw.
1710 * Returns 1, if no segments are in flight and we have queued segments, but
1711 * cannot send anything now because of SWS or another problem.
1713 static int tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
,
1714 int push_one
, gfp_t gfp
)
1716 struct tcp_sock
*tp
= tcp_sk(sk
);
1717 struct sk_buff
*skb
;
1718 unsigned int tso_segs
, sent_pkts
;
1725 /* Do MTU probing. */
1726 result
= tcp_mtu_probe(sk
);
1729 } else if (result
> 0) {
1734 while ((skb
= tcp_send_head(sk
))) {
1737 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1740 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1744 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
)))
1747 if (tso_segs
== 1) {
1748 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
1749 (tcp_skb_is_last(sk
, skb
) ?
1750 nonagle
: TCP_NAGLE_PUSH
))))
1753 if (!push_one
&& tcp_tso_should_defer(sk
, skb
))
1758 if (tso_segs
> 1 && !tcp_urg_mode(tp
))
1759 limit
= tcp_mss_split_point(sk
, skb
, mss_now
,
1762 if (skb
->len
> limit
&&
1763 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
)))
1766 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1768 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, gfp
)))
1771 /* Advance the send_head. This one is sent out.
1772 * This call will increment packets_out.
1774 tcp_event_new_data_sent(sk
, skb
);
1776 tcp_minshall_update(tp
, mss_now
, skb
);
1783 if (likely(sent_pkts
)) {
1784 tcp_cwnd_validate(sk
);
1787 return !tp
->packets_out
&& tcp_send_head(sk
);
1790 /* Push out any pending frames which were held back due to
1791 * TCP_CORK or attempt at coalescing tiny packets.
1792 * The socket must be locked by the caller.
1794 void __tcp_push_pending_frames(struct sock
*sk
, unsigned int cur_mss
,
1797 struct sk_buff
*skb
= tcp_send_head(sk
);
1802 /* If we are closed, the bytes will have to remain here.
1803 * In time closedown will finish, we empty the write queue and
1804 * all will be happy.
1806 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
1809 if (tcp_write_xmit(sk
, cur_mss
, nonagle
, 0, GFP_ATOMIC
))
1810 tcp_check_probe_timer(sk
);
1813 /* Send _single_ skb sitting at the send head. This function requires
1814 * true push pending frames to setup probe timer etc.
1816 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
1818 struct sk_buff
*skb
= tcp_send_head(sk
);
1820 BUG_ON(!skb
|| skb
->len
< mss_now
);
1822 tcp_write_xmit(sk
, mss_now
, TCP_NAGLE_PUSH
, 1, sk
->sk_allocation
);
1825 /* This function returns the amount that we can raise the
1826 * usable window based on the following constraints
1828 * 1. The window can never be shrunk once it is offered (RFC 793)
1829 * 2. We limit memory per socket
1832 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1833 * RECV.NEXT + RCV.WIN fixed until:
1834 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1836 * i.e. don't raise the right edge of the window until you can raise
1837 * it at least MSS bytes.
1839 * Unfortunately, the recommended algorithm breaks header prediction,
1840 * since header prediction assumes th->window stays fixed.
1842 * Strictly speaking, keeping th->window fixed violates the receiver
1843 * side SWS prevention criteria. The problem is that under this rule
1844 * a stream of single byte packets will cause the right side of the
1845 * window to always advance by a single byte.
1847 * Of course, if the sender implements sender side SWS prevention
1848 * then this will not be a problem.
1850 * BSD seems to make the following compromise:
1852 * If the free space is less than the 1/4 of the maximum
1853 * space available and the free space is less than 1/2 mss,
1854 * then set the window to 0.
1855 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1856 * Otherwise, just prevent the window from shrinking
1857 * and from being larger than the largest representable value.
1859 * This prevents incremental opening of the window in the regime
1860 * where TCP is limited by the speed of the reader side taking
1861 * data out of the TCP receive queue. It does nothing about
1862 * those cases where the window is constrained on the sender side
1863 * because the pipeline is full.
1865 * BSD also seems to "accidentally" limit itself to windows that are a
1866 * multiple of MSS, at least until the free space gets quite small.
1867 * This would appear to be a side effect of the mbuf implementation.
1868 * Combining these two algorithms results in the observed behavior
1869 * of having a fixed window size at almost all times.
1871 * Below we obtain similar behavior by forcing the offered window to
1872 * a multiple of the mss when it is feasible to do so.
1874 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1875 * Regular options like TIMESTAMP are taken into account.
1877 u32
__tcp_select_window(struct sock
*sk
)
1879 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1880 struct tcp_sock
*tp
= tcp_sk(sk
);
1881 /* MSS for the peer's data. Previous versions used mss_clamp
1882 * here. I don't know if the value based on our guesses
1883 * of peer's MSS is better for the performance. It's more correct
1884 * but may be worse for the performance because of rcv_mss
1885 * fluctuations. --SAW 1998/11/1
1887 int mss
= icsk
->icsk_ack
.rcv_mss
;
1888 int free_space
= tcp_space(sk
);
1889 int full_space
= min_t(int, tp
->window_clamp
, tcp_full_space(sk
));
1892 if (mss
> full_space
)
1895 if (free_space
< (full_space
>> 1)) {
1896 icsk
->icsk_ack
.quick
= 0;
1898 if (tcp_memory_pressure
)
1899 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
,
1902 if (free_space
< mss
)
1906 if (free_space
> tp
->rcv_ssthresh
)
1907 free_space
= tp
->rcv_ssthresh
;
1909 /* Don't do rounding if we are using window scaling, since the
1910 * scaled window will not line up with the MSS boundary anyway.
1912 window
= tp
->rcv_wnd
;
1913 if (tp
->rx_opt
.rcv_wscale
) {
1914 window
= free_space
;
1916 /* Advertise enough space so that it won't get scaled away.
1917 * Import case: prevent zero window announcement if
1918 * 1<<rcv_wscale > mss.
1920 if (((window
>> tp
->rx_opt
.rcv_wscale
) << tp
->rx_opt
.rcv_wscale
) != window
)
1921 window
= (((window
>> tp
->rx_opt
.rcv_wscale
) + 1)
1922 << tp
->rx_opt
.rcv_wscale
);
1924 /* Get the largest window that is a nice multiple of mss.
1925 * Window clamp already applied above.
1926 * If our current window offering is within 1 mss of the
1927 * free space we just keep it. This prevents the divide
1928 * and multiply from happening most of the time.
1929 * We also don't do any window rounding when the free space
1932 if (window
<= free_space
- mss
|| window
> free_space
)
1933 window
= (free_space
/ mss
) * mss
;
1934 else if (mss
== full_space
&&
1935 free_space
> window
+ (full_space
>> 1))
1936 window
= free_space
;
1942 /* Collapses two adjacent SKB's during retransmission. */
1943 static void tcp_collapse_retrans(struct sock
*sk
, struct sk_buff
*skb
)
1945 struct tcp_sock
*tp
= tcp_sk(sk
);
1946 struct sk_buff
*next_skb
= tcp_write_queue_next(sk
, skb
);
1947 int skb_size
, next_skb_size
;
1949 skb_size
= skb
->len
;
1950 next_skb_size
= next_skb
->len
;
1952 BUG_ON(tcp_skb_pcount(skb
) != 1 || tcp_skb_pcount(next_skb
) != 1);
1954 tcp_highest_sack_combine(sk
, next_skb
, skb
);
1956 tcp_unlink_write_queue(next_skb
, sk
);
1958 skb_copy_from_linear_data(next_skb
, skb_put(skb
, next_skb_size
),
1961 if (next_skb
->ip_summed
== CHECKSUM_PARTIAL
)
1962 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1964 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1965 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
1967 /* Update sequence range on original skb. */
1968 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
1970 /* Merge over control information. This moves PSH/FIN etc. over */
1971 TCP_SKB_CB(skb
)->flags
|= TCP_SKB_CB(next_skb
)->flags
;
1973 /* All done, get rid of second SKB and account for it so
1974 * packet counting does not break.
1976 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
& TCPCB_EVER_RETRANS
;
1978 /* changed transmit queue under us so clear hints */
1979 tcp_clear_retrans_hints_partial(tp
);
1980 if (next_skb
== tp
->retransmit_skb_hint
)
1981 tp
->retransmit_skb_hint
= skb
;
1983 tcp_adjust_pcount(sk
, next_skb
, tcp_skb_pcount(next_skb
));
1985 sk_wmem_free_skb(sk
, next_skb
);
1988 /* Check if coalescing SKBs is legal. */
1989 static int tcp_can_collapse(struct sock
*sk
, struct sk_buff
*skb
)
1991 if (tcp_skb_pcount(skb
) > 1)
1993 /* TODO: SACK collapsing could be used to remove this condition */
1994 if (skb_shinfo(skb
)->nr_frags
!= 0)
1996 if (skb_cloned(skb
))
1998 if (skb
== tcp_send_head(sk
))
2000 /* Some heurestics for collapsing over SACK'd could be invented */
2001 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
2007 /* Collapse packets in the retransmit queue to make to create
2008 * less packets on the wire. This is only done on retransmission.
2010 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*to
,
2013 struct tcp_sock
*tp
= tcp_sk(sk
);
2014 struct sk_buff
*skb
= to
, *tmp
;
2017 if (!sysctl_tcp_retrans_collapse
)
2019 if (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_SYN
)
2022 tcp_for_write_queue_from_safe(skb
, tmp
, sk
) {
2023 if (!tcp_can_collapse(sk
, skb
))
2035 /* Punt if not enough space exists in the first SKB for
2036 * the data in the second
2038 if (skb
->len
> skb_tailroom(to
))
2041 if (after(TCP_SKB_CB(skb
)->end_seq
, tcp_wnd_end(tp
)))
2044 tcp_collapse_retrans(sk
, to
);
2048 /* This retransmits one SKB. Policy decisions and retransmit queue
2049 * state updates are done by the caller. Returns non-zero if an
2050 * error occurred which prevented the send.
2052 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
2054 struct tcp_sock
*tp
= tcp_sk(sk
);
2055 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2056 unsigned int cur_mss
;
2059 /* Inconslusive MTU probe */
2060 if (icsk
->icsk_mtup
.probe_size
) {
2061 icsk
->icsk_mtup
.probe_size
= 0;
2064 /* Do not sent more than we queued. 1/4 is reserved for possible
2065 * copying overhead: fragmentation, tunneling, mangling etc.
2067 if (atomic_read(&sk
->sk_wmem_alloc
) >
2068 min(sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2), sk
->sk_sndbuf
))
2071 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
2072 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
2074 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
2078 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
2079 return -EHOSTUNREACH
; /* Routing failure or similar. */
2081 cur_mss
= tcp_current_mss(sk
);
2083 /* If receiver has shrunk his window, and skb is out of
2084 * new window, do not retransmit it. The exception is the
2085 * case, when window is shrunk to zero. In this case
2086 * our retransmit serves as a zero window probe.
2088 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
)) &&
2089 TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
2092 if (skb
->len
> cur_mss
) {
2093 if (tcp_fragment(sk
, skb
, cur_mss
, cur_mss
))
2094 return -ENOMEM
; /* We'll try again later. */
2096 int oldpcount
= tcp_skb_pcount(skb
);
2098 if (unlikely(oldpcount
> 1)) {
2099 tcp_init_tso_segs(sk
, skb
, cur_mss
);
2100 tcp_adjust_pcount(sk
, skb
, oldpcount
- tcp_skb_pcount(skb
));
2104 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
2106 /* Some Solaris stacks overoptimize and ignore the FIN on a
2107 * retransmit when old data is attached. So strip it off
2108 * since it is cheap to do so and saves bytes on the network.
2111 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
) &&
2112 tp
->snd_una
== (TCP_SKB_CB(skb
)->end_seq
- 1)) {
2113 if (!pskb_trim(skb
, 0)) {
2114 /* Reuse, even though it does some unnecessary work */
2115 tcp_init_nondata_skb(skb
, TCP_SKB_CB(skb
)->end_seq
- 1,
2116 TCP_SKB_CB(skb
)->flags
);
2117 skb
->ip_summed
= CHECKSUM_NONE
;
2121 /* Make a copy, if the first transmission SKB clone we made
2122 * is still in somebody's hands, else make a clone.
2124 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2126 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2129 /* Update global TCP statistics. */
2130 TCP_INC_STATS(sock_net(sk
), TCP_MIB_RETRANSSEGS
);
2132 tp
->total_retrans
++;
2134 #if FASTRETRANS_DEBUG > 0
2135 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2136 if (net_ratelimit())
2137 printk(KERN_DEBUG
"retrans_out leaked.\n");
2140 if (!tp
->retrans_out
)
2141 tp
->lost_retrans_low
= tp
->snd_nxt
;
2142 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
2143 tp
->retrans_out
+= tcp_skb_pcount(skb
);
2145 /* Save stamp of the first retransmit. */
2146 if (!tp
->retrans_stamp
)
2147 tp
->retrans_stamp
= TCP_SKB_CB(skb
)->when
;
2151 /* snd_nxt is stored to detect loss of retransmitted segment,
2152 * see tcp_input.c tcp_sacktag_write_queue().
2154 TCP_SKB_CB(skb
)->ack_seq
= tp
->snd_nxt
;
2159 /* Check if we forward retransmits are possible in the current
2160 * window/congestion state.
2162 static int tcp_can_forward_retransmit(struct sock
*sk
)
2164 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2165 struct tcp_sock
*tp
= tcp_sk(sk
);
2167 /* Forward retransmissions are possible only during Recovery. */
2168 if (icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
2171 /* No forward retransmissions in Reno are possible. */
2172 if (tcp_is_reno(tp
))
2175 /* Yeah, we have to make difficult choice between forward transmission
2176 * and retransmission... Both ways have their merits...
2178 * For now we do not retransmit anything, while we have some new
2179 * segments to send. In the other cases, follow rule 3 for
2180 * NextSeg() specified in RFC3517.
2183 if (tcp_may_send_now(sk
))
2189 /* This gets called after a retransmit timeout, and the initially
2190 * retransmitted data is acknowledged. It tries to continue
2191 * resending the rest of the retransmit queue, until either
2192 * we've sent it all or the congestion window limit is reached.
2193 * If doing SACK, the first ACK which comes back for a timeout
2194 * based retransmit packet might feed us FACK information again.
2195 * If so, we use it to avoid unnecessarily retransmissions.
2197 void tcp_xmit_retransmit_queue(struct sock
*sk
)
2199 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2200 struct tcp_sock
*tp
= tcp_sk(sk
);
2201 struct sk_buff
*skb
;
2202 struct sk_buff
*hole
= NULL
;
2205 int fwd_rexmitting
= 0;
2208 tp
->retransmit_high
= tp
->snd_una
;
2210 if (tp
->retransmit_skb_hint
) {
2211 skb
= tp
->retransmit_skb_hint
;
2212 last_lost
= TCP_SKB_CB(skb
)->end_seq
;
2213 if (after(last_lost
, tp
->retransmit_high
))
2214 last_lost
= tp
->retransmit_high
;
2216 skb
= tcp_write_queue_head(sk
);
2217 last_lost
= tp
->snd_una
;
2220 tcp_for_write_queue_from(skb
, sk
) {
2221 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
2223 if (skb
== tcp_send_head(sk
))
2225 /* we could do better than to assign each time */
2227 tp
->retransmit_skb_hint
= skb
;
2229 /* Assume this retransmit will generate
2230 * only one packet for congestion window
2231 * calculation purposes. This works because
2232 * tcp_retransmit_skb() will chop up the
2233 * packet to be MSS sized and all the
2234 * packet counting works out.
2236 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
2239 if (fwd_rexmitting
) {
2241 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_highest_sack_seq(tp
)))
2243 mib_idx
= LINUX_MIB_TCPFORWARDRETRANS
;
2245 } else if (!before(TCP_SKB_CB(skb
)->seq
, tp
->retransmit_high
)) {
2246 tp
->retransmit_high
= last_lost
;
2247 if (!tcp_can_forward_retransmit(sk
))
2249 /* Backtrack if necessary to non-L'ed skb */
2257 } else if (!(sacked
& TCPCB_LOST
)) {
2258 if (hole
== NULL
&& !(sacked
& (TCPCB_SACKED_RETRANS
|TCPCB_SACKED_ACKED
)))
2263 last_lost
= TCP_SKB_CB(skb
)->end_seq
;
2264 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
2265 mib_idx
= LINUX_MIB_TCPFASTRETRANS
;
2267 mib_idx
= LINUX_MIB_TCPSLOWSTARTRETRANS
;
2270 if (sacked
& (TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))
2273 if (tcp_retransmit_skb(sk
, skb
))
2275 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2277 if (skb
== tcp_write_queue_head(sk
))
2278 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2279 inet_csk(sk
)->icsk_rto
,
2284 /* Send a fin. The caller locks the socket for us. This cannot be
2285 * allowed to fail queueing a FIN frame under any circumstances.
2287 void tcp_send_fin(struct sock
*sk
)
2289 struct tcp_sock
*tp
= tcp_sk(sk
);
2290 struct sk_buff
*skb
= tcp_write_queue_tail(sk
);
2293 /* Optimization, tack on the FIN if we have a queue of
2294 * unsent frames. But be careful about outgoing SACKS
2297 mss_now
= tcp_current_mss(sk
);
2299 if (tcp_send_head(sk
) != NULL
) {
2300 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_FIN
;
2301 TCP_SKB_CB(skb
)->end_seq
++;
2304 /* Socket is locked, keep trying until memory is available. */
2306 skb
= alloc_skb_fclone(MAX_TCP_HEADER
,
2313 /* Reserve space for headers and prepare control bits. */
2314 skb_reserve(skb
, MAX_TCP_HEADER
);
2315 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2316 tcp_init_nondata_skb(skb
, tp
->write_seq
,
2317 TCPCB_FLAG_ACK
| TCPCB_FLAG_FIN
);
2318 tcp_queue_skb(sk
, skb
);
2320 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_OFF
);
2323 /* We get here when a process closes a file descriptor (either due to
2324 * an explicit close() or as a byproduct of exit()'ing) and there
2325 * was unread data in the receive queue. This behavior is recommended
2326 * by RFC 2525, section 2.17. -DaveM
2328 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
2330 struct sk_buff
*skb
;
2332 /* NOTE: No TCP options attached and we never retransmit this. */
2333 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
2335 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
2339 /* Reserve space for headers and prepare control bits. */
2340 skb_reserve(skb
, MAX_TCP_HEADER
);
2341 tcp_init_nondata_skb(skb
, tcp_acceptable_seq(sk
),
2342 TCPCB_FLAG_ACK
| TCPCB_FLAG_RST
);
2344 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2345 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
2346 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
2348 TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTRSTS
);
2351 /* Send a crossed SYN-ACK during socket establishment.
2352 * WARNING: This routine must only be called when we have already sent
2353 * a SYN packet that crossed the incoming SYN that caused this routine
2354 * to get called. If this assumption fails then the initial rcv_wnd
2355 * and rcv_wscale values will not be correct.
2357 int tcp_send_synack(struct sock
*sk
)
2359 struct sk_buff
*skb
;
2361 skb
= tcp_write_queue_head(sk
);
2362 if (skb
== NULL
|| !(TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_SYN
)) {
2363 printk(KERN_DEBUG
"tcp_send_synack: wrong queue state\n");
2366 if (!(TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_ACK
)) {
2367 if (skb_cloned(skb
)) {
2368 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
2371 tcp_unlink_write_queue(skb
, sk
);
2372 skb_header_release(nskb
);
2373 __tcp_add_write_queue_head(sk
, nskb
);
2374 sk_wmem_free_skb(sk
, skb
);
2375 sk
->sk_wmem_queued
+= nskb
->truesize
;
2376 sk_mem_charge(sk
, nskb
->truesize
);
2380 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_ACK
;
2381 TCP_ECN_send_synack(tcp_sk(sk
), skb
);
2383 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2384 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2387 /* Prepare a SYN-ACK. */
2388 struct sk_buff
*tcp_make_synack(struct sock
*sk
, struct dst_entry
*dst
,
2389 struct request_sock
*req
,
2390 struct request_values
*rvp
)
2392 struct tcp_out_options opts
;
2393 struct tcp_extend_values
*xvp
= tcp_xv(rvp
);
2394 struct inet_request_sock
*ireq
= inet_rsk(req
);
2395 struct tcp_sock
*tp
= tcp_sk(sk
);
2397 struct sk_buff
*skb
;
2398 struct tcp_md5sig_key
*md5
;
2399 int tcp_header_size
;
2402 skb
= sock_wmalloc(sk
, MAX_TCP_HEADER
+ 15, 1, GFP_ATOMIC
);
2406 /* Reserve space for headers. */
2407 skb_reserve(skb
, MAX_TCP_HEADER
);
2409 skb_dst_set(skb
, dst_clone(dst
));
2411 mss
= dst_metric(dst
, RTAX_ADVMSS
);
2412 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< mss
)
2413 mss
= tp
->rx_opt
.user_mss
;
2415 if (req
->rcv_wnd
== 0) { /* ignored for retransmitted syns */
2417 /* Set this up on the first call only */
2418 req
->window_clamp
= tp
->window_clamp
? : dst_metric(dst
, RTAX_WINDOW
);
2419 /* tcp_full_space because it is guaranteed to be the first packet */
2420 tcp_select_initial_window(tcp_full_space(sk
),
2421 mss
- (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0),
2426 ireq
->rcv_wscale
= rcv_wscale
;
2429 memset(&opts
, 0, sizeof(opts
));
2430 #ifdef CONFIG_SYN_COOKIES
2431 if (unlikely(req
->cookie_ts
))
2432 TCP_SKB_CB(skb
)->when
= cookie_init_timestamp(req
);
2435 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2436 tcp_header_size
= tcp_synack_options(sk
, req
, mss
,
2437 skb
, &opts
, &md5
, xvp
)
2440 skb_push(skb
, tcp_header_size
);
2441 skb_reset_transport_header(skb
);
2444 memset(th
, 0, sizeof(struct tcphdr
));
2447 TCP_ECN_make_synack(req
, th
);
2448 th
->source
= ireq
->loc_port
;
2449 th
->dest
= ireq
->rmt_port
;
2450 /* Setting of flags are superfluous here for callers (and ECE is
2451 * not even correctly set)
2453 tcp_init_nondata_skb(skb
, tcp_rsk(req
)->snt_isn
,
2454 TCPCB_FLAG_SYN
| TCPCB_FLAG_ACK
);
2456 if (OPTION_COOKIE_EXTENSION
& opts
.options
) {
2457 const struct tcp_cookie_values
*cvp
= tp
->cookie_values
;
2460 cvp
->s_data_constant
&&
2461 cvp
->s_data_desired
> 0) {
2462 u8
*buf
= skb_put(skb
, cvp
->s_data_desired
);
2464 /* copy data directly from the listening socket. */
2465 memcpy(buf
, cvp
->s_data_payload
, cvp
->s_data_desired
);
2466 TCP_SKB_CB(skb
)->end_seq
+= cvp
->s_data_desired
;
2469 if (opts
.hash_size
> 0) {
2470 __u32 workspace
[SHA_WORKSPACE_WORDS
];
2471 u32
*mess
= &xvp
->cookie_bakery
[COOKIE_DIGEST_WORDS
];
2472 u32
*tail
= &mess
[COOKIE_MESSAGE_WORDS
-1];
2474 /* Secret recipe depends on the Timestamp, (future)
2475 * Sequence and Acknowledgment Numbers, Initiator
2476 * Cookie, and others handled by IP variant caller.
2478 *tail
-- ^= opts
.tsval
;
2479 *tail
-- ^= tcp_rsk(req
)->rcv_isn
+ 1;
2480 *tail
-- ^= TCP_SKB_CB(skb
)->seq
+ 1;
2483 *tail
-- ^= ((th
->dest
<< 16) | th
->source
);
2484 *tail
-- ^= (u32
)(unsigned long)cvp
; /* per sockopt */
2486 sha_transform((__u32
*)&xvp
->cookie_bakery
[0],
2489 opts
.hash_location
=
2490 (__u8
*)&xvp
->cookie_bakery
[0];
2494 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
2495 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_isn
+ 1);
2497 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2498 th
->window
= htons(min(req
->rcv_wnd
, 65535U));
2499 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
);
2500 th
->doff
= (tcp_header_size
>> 2);
2501 TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
);
2503 #ifdef CONFIG_TCP_MD5SIG
2504 /* Okay, we have all we need - do the md5 hash if needed */
2506 tcp_rsk(req
)->af_specific
->calc_md5_hash(opts
.hash_location
,
2507 md5
, NULL
, req
, skb
);
2514 /* Do all connect socket setups that can be done AF independent. */
2515 static void tcp_connect_init(struct sock
*sk
)
2517 struct dst_entry
*dst
= __sk_dst_get(sk
);
2518 struct tcp_sock
*tp
= tcp_sk(sk
);
2521 /* We'll fix this up when we get a response from the other end.
2522 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2524 tp
->tcp_header_len
= sizeof(struct tcphdr
) +
2525 (sysctl_tcp_timestamps
? TCPOLEN_TSTAMP_ALIGNED
: 0);
2527 #ifdef CONFIG_TCP_MD5SIG
2528 if (tp
->af_specific
->md5_lookup(sk
, sk
) != NULL
)
2529 tp
->tcp_header_len
+= TCPOLEN_MD5SIG_ALIGNED
;
2532 /* If user gave his TCP_MAXSEG, record it to clamp */
2533 if (tp
->rx_opt
.user_mss
)
2534 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
2537 tcp_sync_mss(sk
, dst_mtu(dst
));
2539 if (!tp
->window_clamp
)
2540 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
2541 tp
->advmss
= dst_metric(dst
, RTAX_ADVMSS
);
2542 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< tp
->advmss
)
2543 tp
->advmss
= tp
->rx_opt
.user_mss
;
2545 tcp_initialize_rcv_mss(sk
);
2547 tcp_select_initial_window(tcp_full_space(sk
),
2548 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
2551 sysctl_tcp_window_scaling
,
2554 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
2555 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
2558 sock_reset_flag(sk
, SOCK_DONE
);
2561 tp
->snd_una
= tp
->write_seq
;
2562 tp
->snd_sml
= tp
->write_seq
;
2563 tp
->snd_up
= tp
->write_seq
;
2568 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
2569 inet_csk(sk
)->icsk_retransmits
= 0;
2570 tcp_clear_retrans(tp
);
2573 /* Build a SYN and send it off. */
2574 int tcp_connect(struct sock
*sk
)
2576 struct tcp_sock
*tp
= tcp_sk(sk
);
2577 struct sk_buff
*buff
;
2579 tcp_connect_init(sk
);
2581 buff
= alloc_skb_fclone(MAX_TCP_HEADER
+ 15, sk
->sk_allocation
);
2582 if (unlikely(buff
== NULL
))
2585 /* Reserve space for headers. */
2586 skb_reserve(buff
, MAX_TCP_HEADER
);
2588 tp
->snd_nxt
= tp
->write_seq
;
2589 tcp_init_nondata_skb(buff
, tp
->write_seq
++, TCPCB_FLAG_SYN
);
2590 TCP_ECN_send_syn(sk
, buff
);
2593 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2594 tp
->retrans_stamp
= TCP_SKB_CB(buff
)->when
;
2595 skb_header_release(buff
);
2596 __tcp_add_write_queue_tail(sk
, buff
);
2597 sk
->sk_wmem_queued
+= buff
->truesize
;
2598 sk_mem_charge(sk
, buff
->truesize
);
2599 tp
->packets_out
+= tcp_skb_pcount(buff
);
2600 tcp_transmit_skb(sk
, buff
, 1, sk
->sk_allocation
);
2602 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2603 * in order to make this packet get counted in tcpOutSegs.
2605 tp
->snd_nxt
= tp
->write_seq
;
2606 tp
->pushed_seq
= tp
->write_seq
;
2607 TCP_INC_STATS(sock_net(sk
), TCP_MIB_ACTIVEOPENS
);
2609 /* Timer for repeating the SYN until an answer. */
2610 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2611 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2615 /* Send out a delayed ack, the caller does the policy checking
2616 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2619 void tcp_send_delayed_ack(struct sock
*sk
)
2621 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2622 int ato
= icsk
->icsk_ack
.ato
;
2623 unsigned long timeout
;
2625 if (ato
> TCP_DELACK_MIN
) {
2626 const struct tcp_sock
*tp
= tcp_sk(sk
);
2627 int max_ato
= HZ
/ 2;
2629 if (icsk
->icsk_ack
.pingpong
||
2630 (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
2631 max_ato
= TCP_DELACK_MAX
;
2633 /* Slow path, intersegment interval is "high". */
2635 /* If some rtt estimate is known, use it to bound delayed ack.
2636 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2640 int rtt
= max(tp
->srtt
>> 3, TCP_DELACK_MIN
);
2646 ato
= min(ato
, max_ato
);
2649 /* Stay within the limit we were given */
2650 timeout
= jiffies
+ ato
;
2652 /* Use new timeout only if there wasn't a older one earlier. */
2653 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
2654 /* If delack timer was blocked or is about to expire,
2657 if (icsk
->icsk_ack
.blocked
||
2658 time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
2663 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
2664 timeout
= icsk
->icsk_ack
.timeout
;
2666 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
2667 icsk
->icsk_ack
.timeout
= timeout
;
2668 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
2671 /* This routine sends an ack and also updates the window. */
2672 void tcp_send_ack(struct sock
*sk
)
2674 struct sk_buff
*buff
;
2676 /* If we have been reset, we may not send again. */
2677 if (sk
->sk_state
== TCP_CLOSE
)
2680 /* We are not putting this on the write queue, so
2681 * tcp_transmit_skb() will set the ownership to this
2684 buff
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2686 inet_csk_schedule_ack(sk
);
2687 inet_csk(sk
)->icsk_ack
.ato
= TCP_ATO_MIN
;
2688 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
2689 TCP_DELACK_MAX
, TCP_RTO_MAX
);
2693 /* Reserve space for headers and prepare control bits. */
2694 skb_reserve(buff
, MAX_TCP_HEADER
);
2695 tcp_init_nondata_skb(buff
, tcp_acceptable_seq(sk
), TCPCB_FLAG_ACK
);
2697 /* Send it off, this clears delayed acks for us. */
2698 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2699 tcp_transmit_skb(sk
, buff
, 0, GFP_ATOMIC
);
2702 /* This routine sends a packet with an out of date sequence
2703 * number. It assumes the other end will try to ack it.
2705 * Question: what should we make while urgent mode?
2706 * 4.4BSD forces sending single byte of data. We cannot send
2707 * out of window data, because we have SND.NXT==SND.MAX...
2709 * Current solution: to send TWO zero-length segments in urgent mode:
2710 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2711 * out-of-date with SND.UNA-1 to probe window.
2713 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
)
2715 struct tcp_sock
*tp
= tcp_sk(sk
);
2716 struct sk_buff
*skb
;
2718 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2719 skb
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2723 /* Reserve space for headers and set control bits. */
2724 skb_reserve(skb
, MAX_TCP_HEADER
);
2725 /* Use a previous sequence. This should cause the other
2726 * end to send an ack. Don't queue or clone SKB, just
2729 tcp_init_nondata_skb(skb
, tp
->snd_una
- !urgent
, TCPCB_FLAG_ACK
);
2730 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2731 return tcp_transmit_skb(sk
, skb
, 0, GFP_ATOMIC
);
2734 /* Initiate keepalive or window probe from timer. */
2735 int tcp_write_wakeup(struct sock
*sk
)
2737 struct tcp_sock
*tp
= tcp_sk(sk
);
2738 struct sk_buff
*skb
;
2740 if (sk
->sk_state
== TCP_CLOSE
)
2743 if ((skb
= tcp_send_head(sk
)) != NULL
&&
2744 before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
))) {
2746 unsigned int mss
= tcp_current_mss(sk
);
2747 unsigned int seg_size
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
2749 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
2750 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
2752 /* We are probing the opening of a window
2753 * but the window size is != 0
2754 * must have been a result SWS avoidance ( sender )
2756 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
2758 seg_size
= min(seg_size
, mss
);
2759 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2760 if (tcp_fragment(sk
, skb
, seg_size
, mss
))
2762 } else if (!tcp_skb_pcount(skb
))
2763 tcp_set_skb_tso_segs(sk
, skb
, mss
);
2765 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2766 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2767 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2769 tcp_event_new_data_sent(sk
, skb
);
2772 if (between(tp
->snd_up
, tp
->snd_una
+ 1, tp
->snd_una
+ 0xFFFF))
2773 tcp_xmit_probe_skb(sk
, 1);
2774 return tcp_xmit_probe_skb(sk
, 0);
2778 /* A window probe timeout has occurred. If window is not closed send
2779 * a partial packet else a zero probe.
2781 void tcp_send_probe0(struct sock
*sk
)
2783 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2784 struct tcp_sock
*tp
= tcp_sk(sk
);
2787 err
= tcp_write_wakeup(sk
);
2789 if (tp
->packets_out
|| !tcp_send_head(sk
)) {
2790 /* Cancel probe timer, if it is not required. */
2791 icsk
->icsk_probes_out
= 0;
2792 icsk
->icsk_backoff
= 0;
2797 if (icsk
->icsk_backoff
< sysctl_tcp_retries2
)
2798 icsk
->icsk_backoff
++;
2799 icsk
->icsk_probes_out
++;
2800 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2801 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
2804 /* If packet was not sent due to local congestion,
2805 * do not backoff and do not remember icsk_probes_out.
2806 * Let local senders to fight for local resources.
2808 * Use accumulated backoff yet.
2810 if (!icsk
->icsk_probes_out
)
2811 icsk
->icsk_probes_out
= 1;
2812 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2813 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
,
2814 TCP_RESOURCE_PROBE_INTERVAL
),
2819 EXPORT_SYMBOL(tcp_select_initial_window
);
2820 EXPORT_SYMBOL(tcp_connect
);
2821 EXPORT_SYMBOL(tcp_make_synack
);
2822 EXPORT_SYMBOL(tcp_simple_retransmit
);
2823 EXPORT_SYMBOL(tcp_sync_mss
);
2824 EXPORT_SYMBOL(tcp_mtup_init
);