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/gfp.h>
41 #include <linux/module.h>
43 /* People can turn this off for buggy TCP's found in printers etc. */
44 int sysctl_tcp_retrans_collapse __read_mostly
= 1;
46 /* People can turn this on to work with those rare, broken TCPs that
47 * interpret the window field as a signed quantity.
49 int sysctl_tcp_workaround_signed_windows __read_mostly
= 0;
51 /* This limits the percentage of the congestion window which we
52 * will allow a single TSO frame to consume. Building TSO frames
53 * which are too large can cause TCP streams to be bursty.
55 int sysctl_tcp_tso_win_divisor __read_mostly
= 3;
57 int sysctl_tcp_mtu_probing __read_mostly
= 0;
58 int sysctl_tcp_base_mss __read_mostly
= 512;
60 /* By default, RFC2861 behavior. */
61 int sysctl_tcp_slow_start_after_idle __read_mostly
= 1;
63 int sysctl_tcp_cookie_size __read_mostly
= 0; /* TCP_COOKIE_MAX */
64 EXPORT_SYMBOL_GPL(sysctl_tcp_cookie_size
);
67 /* Account for new data that has been sent to the network. */
68 static void tcp_event_new_data_sent(struct sock
*sk
, struct sk_buff
*skb
)
70 struct tcp_sock
*tp
= tcp_sk(sk
);
71 unsigned int prior_packets
= tp
->packets_out
;
73 tcp_advance_send_head(sk
, skb
);
74 tp
->snd_nxt
= TCP_SKB_CB(skb
)->end_seq
;
76 /* Don't override Nagle indefinately with F-RTO */
77 if (tp
->frto_counter
== 2)
80 tp
->packets_out
+= tcp_skb_pcount(skb
);
82 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
83 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
86 /* SND.NXT, if window was not shrunk.
87 * If window has been shrunk, what should we make? It is not clear at all.
88 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
89 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
90 * invalid. OK, let's make this for now:
92 static inline __u32
tcp_acceptable_seq(struct sock
*sk
)
94 struct tcp_sock
*tp
= tcp_sk(sk
);
96 if (!before(tcp_wnd_end(tp
), tp
->snd_nxt
))
99 return tcp_wnd_end(tp
);
102 /* Calculate mss to advertise in SYN segment.
103 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
105 * 1. It is independent of path mtu.
106 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
107 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
108 * attached devices, because some buggy hosts are confused by
110 * 4. We do not make 3, we advertise MSS, calculated from first
111 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
112 * This may be overridden via information stored in routing table.
113 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
114 * probably even Jumbo".
116 static __u16
tcp_advertise_mss(struct sock
*sk
)
118 struct tcp_sock
*tp
= tcp_sk(sk
);
119 struct dst_entry
*dst
= __sk_dst_get(sk
);
120 int mss
= tp
->advmss
;
122 if (dst
&& dst_metric(dst
, RTAX_ADVMSS
) < mss
) {
123 mss
= dst_metric(dst
, RTAX_ADVMSS
);
130 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
131 * This is the first part of cwnd validation mechanism. */
132 static void tcp_cwnd_restart(struct sock
*sk
, struct dst_entry
*dst
)
134 struct tcp_sock
*tp
= tcp_sk(sk
);
135 s32 delta
= tcp_time_stamp
- tp
->lsndtime
;
136 u32 restart_cwnd
= tcp_init_cwnd(tp
, dst
);
137 u32 cwnd
= tp
->snd_cwnd
;
139 tcp_ca_event(sk
, CA_EVENT_CWND_RESTART
);
141 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
142 restart_cwnd
= min(restart_cwnd
, cwnd
);
144 while ((delta
-= inet_csk(sk
)->icsk_rto
) > 0 && cwnd
> restart_cwnd
)
146 tp
->snd_cwnd
= max(cwnd
, restart_cwnd
);
147 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
148 tp
->snd_cwnd_used
= 0;
151 /* Congestion state accounting after a packet has been sent. */
152 static void tcp_event_data_sent(struct tcp_sock
*tp
,
153 struct sk_buff
*skb
, struct sock
*sk
)
155 struct inet_connection_sock
*icsk
= inet_csk(sk
);
156 const u32 now
= tcp_time_stamp
;
158 if (sysctl_tcp_slow_start_after_idle
&&
159 (!tp
->packets_out
&& (s32
)(now
- tp
->lsndtime
) > icsk
->icsk_rto
))
160 tcp_cwnd_restart(sk
, __sk_dst_get(sk
));
164 /* If it is a reply for ato after last received
165 * packet, enter pingpong mode.
167 if ((u32
)(now
- icsk
->icsk_ack
.lrcvtime
) < icsk
->icsk_ack
.ato
)
168 icsk
->icsk_ack
.pingpong
= 1;
171 /* Account for an ACK we sent. */
172 static inline void tcp_event_ack_sent(struct sock
*sk
, unsigned int pkts
)
174 tcp_dec_quickack_mode(sk
, pkts
);
175 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_DACK
);
178 /* Determine a window scaling and initial window to offer.
179 * Based on the assumption that the given amount of space
180 * will be offered. Store the results in the tp structure.
181 * NOTE: for smooth operation initial space offering should
182 * be a multiple of mss if possible. We assume here that mss >= 1.
183 * This MUST be enforced by all callers.
185 void tcp_select_initial_window(int __space
, __u32 mss
,
186 __u32
*rcv_wnd
, __u32
*window_clamp
,
187 int wscale_ok
, __u8
*rcv_wscale
,
190 unsigned int space
= (__space
< 0 ? 0 : __space
);
192 /* If no clamp set the clamp to the max possible scaled window */
193 if (*window_clamp
== 0)
194 (*window_clamp
) = (65535 << 14);
195 space
= min(*window_clamp
, space
);
197 /* Quantize space offering to a multiple of mss if possible. */
199 space
= (space
/ mss
) * mss
;
201 /* NOTE: offering an initial window larger than 32767
202 * will break some buggy TCP stacks. If the admin tells us
203 * it is likely we could be speaking with such a buggy stack
204 * we will truncate our initial window offering to 32K-1
205 * unless the remote has sent us a window scaling option,
206 * which we interpret as a sign the remote TCP is not
207 * misinterpreting the window field as a signed quantity.
209 if (sysctl_tcp_workaround_signed_windows
)
210 (*rcv_wnd
) = min(space
, MAX_TCP_WINDOW
);
216 /* Set window scaling on max possible window
217 * See RFC1323 for an explanation of the limit to 14
219 space
= max_t(u32
, sysctl_tcp_rmem
[2], sysctl_rmem_max
);
220 space
= min_t(u32
, space
, *window_clamp
);
221 while (space
> 65535 && (*rcv_wscale
) < 14) {
227 /* Set initial window to value enough for senders,
228 * following RFC2414. Senders, not following this RFC,
229 * will be satisfied with 2.
231 if (mss
> (1 << *rcv_wscale
)) {
237 /* when initializing use the value from init_rcv_wnd
238 * rather than the default from above
241 *rcv_wnd
= min(*rcv_wnd
, init_rcv_wnd
* mss
);
243 *rcv_wnd
= min(*rcv_wnd
, init_cwnd
* mss
);
246 /* Set the clamp no higher than max representable value */
247 (*window_clamp
) = min(65535U << (*rcv_wscale
), *window_clamp
);
250 /* Chose a new window to advertise, update state in tcp_sock for the
251 * socket, and return result with RFC1323 scaling applied. The return
252 * value can be stuffed directly into th->window for an outgoing
255 static u16
tcp_select_window(struct sock
*sk
)
257 struct tcp_sock
*tp
= tcp_sk(sk
);
258 u32 cur_win
= tcp_receive_window(tp
);
259 u32 new_win
= __tcp_select_window(sk
);
261 /* Never shrink the offered window */
262 if (new_win
< cur_win
) {
263 /* Danger Will Robinson!
264 * Don't update rcv_wup/rcv_wnd here or else
265 * we will not be able to advertise a zero
266 * window in time. --DaveM
268 * Relax Will Robinson.
270 new_win
= ALIGN(cur_win
, 1 << tp
->rx_opt
.rcv_wscale
);
272 tp
->rcv_wnd
= new_win
;
273 tp
->rcv_wup
= tp
->rcv_nxt
;
275 /* Make sure we do not exceed the maximum possible
278 if (!tp
->rx_opt
.rcv_wscale
&& sysctl_tcp_workaround_signed_windows
)
279 new_win
= min(new_win
, MAX_TCP_WINDOW
);
281 new_win
= min(new_win
, (65535U << tp
->rx_opt
.rcv_wscale
));
283 /* RFC1323 scaling applied */
284 new_win
>>= tp
->rx_opt
.rcv_wscale
;
286 /* If we advertise zero window, disable fast path. */
293 /* Packet ECN state for a SYN-ACK */
294 static inline void TCP_ECN_send_synack(struct tcp_sock
*tp
, struct sk_buff
*skb
)
296 TCP_SKB_CB(skb
)->flags
&= ~TCPCB_FLAG_CWR
;
297 if (!(tp
->ecn_flags
& TCP_ECN_OK
))
298 TCP_SKB_CB(skb
)->flags
&= ~TCPCB_FLAG_ECE
;
301 /* Packet ECN state for a SYN. */
302 static inline void TCP_ECN_send_syn(struct sock
*sk
, struct sk_buff
*skb
)
304 struct tcp_sock
*tp
= tcp_sk(sk
);
307 if (sysctl_tcp_ecn
== 1) {
308 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_ECE
| TCPCB_FLAG_CWR
;
309 tp
->ecn_flags
= TCP_ECN_OK
;
313 static __inline__
void
314 TCP_ECN_make_synack(struct request_sock
*req
, struct tcphdr
*th
)
316 if (inet_rsk(req
)->ecn_ok
)
320 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
323 static inline void TCP_ECN_send(struct sock
*sk
, struct sk_buff
*skb
,
326 struct tcp_sock
*tp
= tcp_sk(sk
);
328 if (tp
->ecn_flags
& TCP_ECN_OK
) {
329 /* Not-retransmitted data segment: set ECT and inject CWR. */
330 if (skb
->len
!= tcp_header_len
&&
331 !before(TCP_SKB_CB(skb
)->seq
, tp
->snd_nxt
)) {
333 if (tp
->ecn_flags
& TCP_ECN_QUEUE_CWR
) {
334 tp
->ecn_flags
&= ~TCP_ECN_QUEUE_CWR
;
335 tcp_hdr(skb
)->cwr
= 1;
336 skb_shinfo(skb
)->gso_type
|= SKB_GSO_TCP_ECN
;
339 /* ACK or retransmitted segment: clear ECT|CE */
340 INET_ECN_dontxmit(sk
);
342 if (tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)
343 tcp_hdr(skb
)->ece
= 1;
347 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
348 * auto increment end seqno.
350 static void tcp_init_nondata_skb(struct sk_buff
*skb
, u32 seq
, u8 flags
)
352 skb
->ip_summed
= CHECKSUM_PARTIAL
;
355 TCP_SKB_CB(skb
)->flags
= flags
;
356 TCP_SKB_CB(skb
)->sacked
= 0;
358 skb_shinfo(skb
)->gso_segs
= 1;
359 skb_shinfo(skb
)->gso_size
= 0;
360 skb_shinfo(skb
)->gso_type
= 0;
362 TCP_SKB_CB(skb
)->seq
= seq
;
363 if (flags
& (TCPCB_FLAG_SYN
| TCPCB_FLAG_FIN
))
365 TCP_SKB_CB(skb
)->end_seq
= seq
;
368 static inline int tcp_urg_mode(const struct tcp_sock
*tp
)
370 return tp
->snd_una
!= tp
->snd_up
;
373 #define OPTION_SACK_ADVERTISE (1 << 0)
374 #define OPTION_TS (1 << 1)
375 #define OPTION_MD5 (1 << 2)
376 #define OPTION_WSCALE (1 << 3)
377 #define OPTION_COOKIE_EXTENSION (1 << 4)
379 struct tcp_out_options
{
380 u8 options
; /* bit field of OPTION_* */
381 u8 ws
; /* window scale, 0 to disable */
382 u8 num_sack_blocks
; /* number of SACK blocks to include */
383 u8 hash_size
; /* bytes in hash_location */
384 u16 mss
; /* 0 to disable */
385 __u32 tsval
, tsecr
; /* need to include OPTION_TS */
386 __u8
*hash_location
; /* temporary pointer, overloaded */
389 /* The sysctl int routines are generic, so check consistency here.
391 static u8
tcp_cookie_size_check(u8 desired
)
396 /* previously specified */
399 cookie_size
= ACCESS_ONCE(sysctl_tcp_cookie_size
);
400 if (cookie_size
<= 0)
401 /* no default specified */
404 if (cookie_size
<= TCP_COOKIE_MIN
)
405 /* value too small, specify minimum */
406 return TCP_COOKIE_MIN
;
408 if (cookie_size
>= TCP_COOKIE_MAX
)
409 /* value too large, specify maximum */
410 return TCP_COOKIE_MAX
;
413 /* 8-bit multiple, illegal, fix it */
416 return (u8
)cookie_size
;
419 /* Write previously computed TCP options to the packet.
421 * Beware: Something in the Internet is very sensitive to the ordering of
422 * TCP options, we learned this through the hard way, so be careful here.
423 * Luckily we can at least blame others for their non-compliance but from
424 * inter-operatibility perspective it seems that we're somewhat stuck with
425 * the ordering which we have been using if we want to keep working with
426 * those broken things (not that it currently hurts anybody as there isn't
427 * particular reason why the ordering would need to be changed).
429 * At least SACK_PERM as the first option is known to lead to a disaster
430 * (but it may well be that other scenarios fail similarly).
432 static void tcp_options_write(__be32
*ptr
, struct tcp_sock
*tp
,
433 struct tcp_out_options
*opts
)
435 u8 options
= opts
->options
; /* mungable copy */
437 /* Having both authentication and cookies for security is redundant,
438 * and there's certainly not enough room. Instead, the cookie-less
439 * extension variant is proposed.
441 * Consider the pessimal case with authentication. The options
443 * COOKIE|MD5(20) + MSS(4) + SACK|TS(12) + WSCALE(4) == 40
445 if (unlikely(OPTION_MD5
& options
)) {
446 if (unlikely(OPTION_COOKIE_EXTENSION
& options
)) {
447 *ptr
++ = htonl((TCPOPT_COOKIE
<< 24) |
448 (TCPOLEN_COOKIE_BASE
<< 16) |
449 (TCPOPT_MD5SIG
<< 8) |
452 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
454 (TCPOPT_MD5SIG
<< 8) |
457 options
&= ~OPTION_COOKIE_EXTENSION
;
458 /* overload cookie hash location */
459 opts
->hash_location
= (__u8
*)ptr
;
463 if (unlikely(opts
->mss
)) {
464 *ptr
++ = htonl((TCPOPT_MSS
<< 24) |
465 (TCPOLEN_MSS
<< 16) |
469 if (likely(OPTION_TS
& options
)) {
470 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
471 *ptr
++ = htonl((TCPOPT_SACK_PERM
<< 24) |
472 (TCPOLEN_SACK_PERM
<< 16) |
473 (TCPOPT_TIMESTAMP
<< 8) |
475 options
&= ~OPTION_SACK_ADVERTISE
;
477 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
479 (TCPOPT_TIMESTAMP
<< 8) |
482 *ptr
++ = htonl(opts
->tsval
);
483 *ptr
++ = htonl(opts
->tsecr
);
486 /* Specification requires after timestamp, so do it now.
488 * Consider the pessimal case without authentication. The options
490 * MSS(4) + SACK|TS(12) + COOKIE(20) + WSCALE(4) == 40
492 if (unlikely(OPTION_COOKIE_EXTENSION
& options
)) {
493 __u8
*cookie_copy
= opts
->hash_location
;
494 u8 cookie_size
= opts
->hash_size
;
496 /* 8-bit multiple handled in tcp_cookie_size_check() above,
499 if (0x2 & cookie_size
) {
500 __u8
*p
= (__u8
*)ptr
;
502 /* 16-bit multiple */
503 *p
++ = TCPOPT_COOKIE
;
504 *p
++ = TCPOLEN_COOKIE_BASE
+ cookie_size
;
505 *p
++ = *cookie_copy
++;
506 *p
++ = *cookie_copy
++;
510 /* 32-bit multiple */
511 *ptr
++ = htonl(((TCPOPT_NOP
<< 24) |
513 (TCPOPT_COOKIE
<< 8) |
514 TCPOLEN_COOKIE_BASE
) +
518 if (cookie_size
> 0) {
519 memcpy(ptr
, cookie_copy
, cookie_size
);
520 ptr
+= (cookie_size
/ 4);
524 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
525 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
527 (TCPOPT_SACK_PERM
<< 8) |
531 if (unlikely(OPTION_WSCALE
& options
)) {
532 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
533 (TCPOPT_WINDOW
<< 16) |
534 (TCPOLEN_WINDOW
<< 8) |
538 if (unlikely(opts
->num_sack_blocks
)) {
539 struct tcp_sack_block
*sp
= tp
->rx_opt
.dsack
?
540 tp
->duplicate_sack
: tp
->selective_acks
;
543 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
546 (TCPOLEN_SACK_BASE
+ (opts
->num_sack_blocks
*
547 TCPOLEN_SACK_PERBLOCK
)));
549 for (this_sack
= 0; this_sack
< opts
->num_sack_blocks
;
551 *ptr
++ = htonl(sp
[this_sack
].start_seq
);
552 *ptr
++ = htonl(sp
[this_sack
].end_seq
);
555 tp
->rx_opt
.dsack
= 0;
559 /* Compute TCP options for SYN packets. This is not the final
560 * network wire format yet.
562 static unsigned tcp_syn_options(struct sock
*sk
, struct sk_buff
*skb
,
563 struct tcp_out_options
*opts
,
564 struct tcp_md5sig_key
**md5
) {
565 struct tcp_sock
*tp
= tcp_sk(sk
);
566 struct tcp_cookie_values
*cvp
= tp
->cookie_values
;
567 unsigned remaining
= MAX_TCP_OPTION_SPACE
;
568 u8 cookie_size
= (!tp
->rx_opt
.cookie_out_never
&& cvp
!= NULL
) ?
569 tcp_cookie_size_check(cvp
->cookie_desired
) :
572 #ifdef CONFIG_TCP_MD5SIG
573 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
575 opts
->options
|= OPTION_MD5
;
576 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
582 /* We always get an MSS option. The option bytes which will be seen in
583 * normal data packets should timestamps be used, must be in the MSS
584 * advertised. But we subtract them from tp->mss_cache so that
585 * calculations in tcp_sendmsg are simpler etc. So account for this
586 * fact here if necessary. If we don't do this correctly, as a
587 * receiver we won't recognize data packets as being full sized when we
588 * should, and thus we won't abide by the delayed ACK rules correctly.
589 * SACKs don't matter, we never delay an ACK when we have any of those
591 opts
->mss
= tcp_advertise_mss(sk
);
592 remaining
-= TCPOLEN_MSS_ALIGNED
;
594 if (likely(sysctl_tcp_timestamps
&& *md5
== NULL
)) {
595 opts
->options
|= OPTION_TS
;
596 opts
->tsval
= TCP_SKB_CB(skb
)->when
;
597 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
598 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
600 if (likely(sysctl_tcp_window_scaling
)) {
601 opts
->ws
= tp
->rx_opt
.rcv_wscale
;
602 opts
->options
|= OPTION_WSCALE
;
603 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
605 if (likely(sysctl_tcp_sack
)) {
606 opts
->options
|= OPTION_SACK_ADVERTISE
;
607 if (unlikely(!(OPTION_TS
& opts
->options
)))
608 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
611 /* Note that timestamps are required by the specification.
613 * Odd numbers of bytes are prohibited by the specification, ensuring
614 * that the cookie is 16-bit aligned, and the resulting cookie pair is
618 (OPTION_TS
& opts
->options
) &&
620 int need
= TCPOLEN_COOKIE_BASE
+ cookie_size
;
623 /* 32-bit multiple */
624 need
+= 2; /* NOPs */
626 if (need
> remaining
) {
627 /* try shrinking cookie to fit */
632 while (need
> remaining
&& TCP_COOKIE_MIN
<= cookie_size
) {
636 if (TCP_COOKIE_MIN
<= cookie_size
) {
637 opts
->options
|= OPTION_COOKIE_EXTENSION
;
638 opts
->hash_location
= (__u8
*)&cvp
->cookie_pair
[0];
639 opts
->hash_size
= cookie_size
;
641 /* Remember for future incarnations. */
642 cvp
->cookie_desired
= cookie_size
;
644 if (cvp
->cookie_desired
!= cvp
->cookie_pair_size
) {
645 /* Currently use random bytes as a nonce,
646 * assuming these are completely unpredictable
647 * by hostile users of the same system.
649 get_random_bytes(&cvp
->cookie_pair
[0],
651 cvp
->cookie_pair_size
= cookie_size
;
657 return MAX_TCP_OPTION_SPACE
- remaining
;
660 /* Set up TCP options for SYN-ACKs. */
661 static unsigned tcp_synack_options(struct sock
*sk
,
662 struct request_sock
*req
,
663 unsigned mss
, struct sk_buff
*skb
,
664 struct tcp_out_options
*opts
,
665 struct tcp_md5sig_key
**md5
,
666 struct tcp_extend_values
*xvp
)
668 struct inet_request_sock
*ireq
= inet_rsk(req
);
669 unsigned remaining
= MAX_TCP_OPTION_SPACE
;
670 u8 cookie_plus
= (xvp
!= NULL
&& !xvp
->cookie_out_never
) ?
674 #ifdef CONFIG_TCP_MD5SIG
675 *md5
= tcp_rsk(req
)->af_specific
->md5_lookup(sk
, req
);
677 opts
->options
|= OPTION_MD5
;
678 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
680 /* We can't fit any SACK blocks in a packet with MD5 + TS
681 * options. There was discussion about disabling SACK
682 * rather than TS in order to fit in better with old,
683 * buggy kernels, but that was deemed to be unnecessary.
685 ireq
->tstamp_ok
&= !ireq
->sack_ok
;
691 /* We always send an MSS option. */
693 remaining
-= TCPOLEN_MSS_ALIGNED
;
695 if (likely(ireq
->wscale_ok
)) {
696 opts
->ws
= ireq
->rcv_wscale
;
697 opts
->options
|= OPTION_WSCALE
;
698 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
700 if (likely(ireq
->tstamp_ok
)) {
701 opts
->options
|= OPTION_TS
;
702 opts
->tsval
= TCP_SKB_CB(skb
)->when
;
703 opts
->tsecr
= req
->ts_recent
;
704 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
706 if (likely(ireq
->sack_ok
)) {
707 opts
->options
|= OPTION_SACK_ADVERTISE
;
708 if (unlikely(!ireq
->tstamp_ok
))
709 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
712 /* Similar rationale to tcp_syn_options() applies here, too.
713 * If the <SYN> options fit, the same options should fit now!
717 cookie_plus
> TCPOLEN_COOKIE_BASE
) {
718 int need
= cookie_plus
; /* has TCPOLEN_COOKIE_BASE */
721 /* 32-bit multiple */
722 need
+= 2; /* NOPs */
724 if (need
<= remaining
) {
725 opts
->options
|= OPTION_COOKIE_EXTENSION
;
726 opts
->hash_size
= cookie_plus
- TCPOLEN_COOKIE_BASE
;
729 /* There's no error return, so flag it. */
730 xvp
->cookie_out_never
= 1; /* true */
734 return MAX_TCP_OPTION_SPACE
- remaining
;
737 /* Compute TCP options for ESTABLISHED sockets. This is not the
738 * final wire format yet.
740 static unsigned tcp_established_options(struct sock
*sk
, struct sk_buff
*skb
,
741 struct tcp_out_options
*opts
,
742 struct tcp_md5sig_key
**md5
) {
743 struct tcp_skb_cb
*tcb
= skb
? TCP_SKB_CB(skb
) : NULL
;
744 struct tcp_sock
*tp
= tcp_sk(sk
);
746 unsigned int eff_sacks
;
748 #ifdef CONFIG_TCP_MD5SIG
749 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
750 if (unlikely(*md5
)) {
751 opts
->options
|= OPTION_MD5
;
752 size
+= TCPOLEN_MD5SIG_ALIGNED
;
758 if (likely(tp
->rx_opt
.tstamp_ok
)) {
759 opts
->options
|= OPTION_TS
;
760 opts
->tsval
= tcb
? tcb
->when
: 0;
761 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
762 size
+= TCPOLEN_TSTAMP_ALIGNED
;
765 eff_sacks
= tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
;
766 if (unlikely(eff_sacks
)) {
767 const unsigned remaining
= MAX_TCP_OPTION_SPACE
- size
;
768 opts
->num_sack_blocks
=
769 min_t(unsigned, eff_sacks
,
770 (remaining
- TCPOLEN_SACK_BASE_ALIGNED
) /
771 TCPOLEN_SACK_PERBLOCK
);
772 size
+= TCPOLEN_SACK_BASE_ALIGNED
+
773 opts
->num_sack_blocks
* TCPOLEN_SACK_PERBLOCK
;
779 /* This routine actually transmits TCP packets queued in by
780 * tcp_do_sendmsg(). This is used by both the initial
781 * transmission and possible later retransmissions.
782 * All SKB's seen here are completely headerless. It is our
783 * job to build the TCP header, and pass the packet down to
784 * IP so it can do the same plus pass the packet off to the
787 * We are working here with either a clone of the original
788 * SKB, or a fresh unique copy made by the retransmit engine.
790 static int tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int clone_it
,
793 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
794 struct inet_sock
*inet
;
796 struct tcp_skb_cb
*tcb
;
797 struct tcp_out_options opts
;
798 unsigned tcp_options_size
, tcp_header_size
;
799 struct tcp_md5sig_key
*md5
;
803 BUG_ON(!skb
|| !tcp_skb_pcount(skb
));
805 /* If congestion control is doing timestamping, we must
806 * take such a timestamp before we potentially clone/copy.
808 if (icsk
->icsk_ca_ops
->flags
& TCP_CONG_RTT_STAMP
)
809 __net_timestamp(skb
);
811 if (likely(clone_it
)) {
812 if (unlikely(skb_cloned(skb
)))
813 skb
= pskb_copy(skb
, gfp_mask
);
815 skb
= skb_clone(skb
, gfp_mask
);
822 tcb
= TCP_SKB_CB(skb
);
823 memset(&opts
, 0, sizeof(opts
));
825 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
))
826 tcp_options_size
= tcp_syn_options(sk
, skb
, &opts
, &md5
);
828 tcp_options_size
= tcp_established_options(sk
, skb
, &opts
,
830 tcp_header_size
= tcp_options_size
+ sizeof(struct tcphdr
);
832 if (tcp_packets_in_flight(tp
) == 0)
833 tcp_ca_event(sk
, CA_EVENT_TX_START
);
835 skb_push(skb
, tcp_header_size
);
836 skb_reset_transport_header(skb
);
837 skb_set_owner_w(skb
, sk
);
839 /* Build TCP header and checksum it. */
841 th
->source
= inet
->inet_sport
;
842 th
->dest
= inet
->inet_dport
;
843 th
->seq
= htonl(tcb
->seq
);
844 th
->ack_seq
= htonl(tp
->rcv_nxt
);
845 *(((__be16
*)th
) + 6) = htons(((tcp_header_size
>> 2) << 12) |
848 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
849 /* RFC1323: The window in SYN & SYN/ACK segments
852 th
->window
= htons(min(tp
->rcv_wnd
, 65535U));
854 th
->window
= htons(tcp_select_window(sk
));
859 /* The urg_mode check is necessary during a below snd_una win probe */
860 if (unlikely(tcp_urg_mode(tp
) && before(tcb
->seq
, tp
->snd_up
))) {
861 if (before(tp
->snd_up
, tcb
->seq
+ 0x10000)) {
862 th
->urg_ptr
= htons(tp
->snd_up
- tcb
->seq
);
864 } else if (after(tcb
->seq
+ 0xFFFF, tp
->snd_nxt
)) {
865 th
->urg_ptr
= htons(0xFFFF);
870 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
);
871 if (likely((tcb
->flags
& TCPCB_FLAG_SYN
) == 0))
872 TCP_ECN_send(sk
, skb
, tcp_header_size
);
874 #ifdef CONFIG_TCP_MD5SIG
875 /* Calculate the MD5 hash, as we have all we need now */
877 sk_nocaps_add(sk
, NETIF_F_GSO_MASK
);
878 tp
->af_specific
->calc_md5_hash(opts
.hash_location
,
883 icsk
->icsk_af_ops
->send_check(sk
, skb
);
885 if (likely(tcb
->flags
& TCPCB_FLAG_ACK
))
886 tcp_event_ack_sent(sk
, tcp_skb_pcount(skb
));
888 if (skb
->len
!= tcp_header_size
)
889 tcp_event_data_sent(tp
, skb
, sk
);
891 if (after(tcb
->end_seq
, tp
->snd_nxt
) || tcb
->seq
== tcb
->end_seq
)
892 TCP_ADD_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
,
893 tcp_skb_pcount(skb
));
895 err
= icsk
->icsk_af_ops
->queue_xmit(skb
);
896 if (likely(err
<= 0))
899 tcp_enter_cwr(sk
, 1);
901 return net_xmit_eval(err
);
904 /* This routine just queues the buffer for sending.
906 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
907 * otherwise socket can stall.
909 static void tcp_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
911 struct tcp_sock
*tp
= tcp_sk(sk
);
913 /* Advance write_seq and place onto the write_queue. */
914 tp
->write_seq
= TCP_SKB_CB(skb
)->end_seq
;
915 skb_header_release(skb
);
916 tcp_add_write_queue_tail(sk
, skb
);
917 sk
->sk_wmem_queued
+= skb
->truesize
;
918 sk_mem_charge(sk
, skb
->truesize
);
921 /* Initialize TSO segments for a packet. */
922 static void tcp_set_skb_tso_segs(struct sock
*sk
, struct sk_buff
*skb
,
923 unsigned int mss_now
)
925 if (skb
->len
<= mss_now
|| !sk_can_gso(sk
) ||
926 skb
->ip_summed
== CHECKSUM_NONE
) {
927 /* Avoid the costly divide in the normal
930 skb_shinfo(skb
)->gso_segs
= 1;
931 skb_shinfo(skb
)->gso_size
= 0;
932 skb_shinfo(skb
)->gso_type
= 0;
934 skb_shinfo(skb
)->gso_segs
= DIV_ROUND_UP(skb
->len
, mss_now
);
935 skb_shinfo(skb
)->gso_size
= mss_now
;
936 skb_shinfo(skb
)->gso_type
= sk
->sk_gso_type
;
940 /* When a modification to fackets out becomes necessary, we need to check
941 * skb is counted to fackets_out or not.
943 static void tcp_adjust_fackets_out(struct sock
*sk
, struct sk_buff
*skb
,
946 struct tcp_sock
*tp
= tcp_sk(sk
);
948 if (!tp
->sacked_out
|| tcp_is_reno(tp
))
951 if (after(tcp_highest_sack_seq(tp
), TCP_SKB_CB(skb
)->seq
))
952 tp
->fackets_out
-= decr
;
955 /* Pcount in the middle of the write queue got changed, we need to do various
956 * tweaks to fix counters
958 static void tcp_adjust_pcount(struct sock
*sk
, struct sk_buff
*skb
, int decr
)
960 struct tcp_sock
*tp
= tcp_sk(sk
);
962 tp
->packets_out
-= decr
;
964 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
965 tp
->sacked_out
-= decr
;
966 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
967 tp
->retrans_out
-= decr
;
968 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
)
969 tp
->lost_out
-= decr
;
971 /* Reno case is special. Sigh... */
972 if (tcp_is_reno(tp
) && decr
> 0)
973 tp
->sacked_out
-= min_t(u32
, tp
->sacked_out
, decr
);
975 tcp_adjust_fackets_out(sk
, skb
, decr
);
977 if (tp
->lost_skb_hint
&&
978 before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
) &&
979 (tcp_is_fack(tp
) || (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)))
980 tp
->lost_cnt_hint
-= decr
;
982 tcp_verify_left_out(tp
);
985 /* Function to create two new TCP segments. Shrinks the given segment
986 * to the specified size and appends a new segment with the rest of the
987 * packet to the list. This won't be called frequently, I hope.
988 * Remember, these are still headerless SKBs at this point.
990 int tcp_fragment(struct sock
*sk
, struct sk_buff
*skb
, u32 len
,
991 unsigned int mss_now
)
993 struct tcp_sock
*tp
= tcp_sk(sk
);
994 struct sk_buff
*buff
;
995 int nsize
, old_factor
;
999 BUG_ON(len
> skb
->len
);
1001 nsize
= skb_headlen(skb
) - len
;
1005 if (skb_cloned(skb
) &&
1006 skb_is_nonlinear(skb
) &&
1007 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
1010 /* Get a new skb... force flag on. */
1011 buff
= sk_stream_alloc_skb(sk
, nsize
, GFP_ATOMIC
);
1013 return -ENOMEM
; /* We'll just try again later. */
1015 sk
->sk_wmem_queued
+= buff
->truesize
;
1016 sk_mem_charge(sk
, buff
->truesize
);
1017 nlen
= skb
->len
- len
- nsize
;
1018 buff
->truesize
+= nlen
;
1019 skb
->truesize
-= nlen
;
1021 /* Correct the sequence numbers. */
1022 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1023 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1024 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1026 /* PSH and FIN should only be set in the second packet. */
1027 flags
= TCP_SKB_CB(skb
)->flags
;
1028 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
| TCPCB_FLAG_PSH
);
1029 TCP_SKB_CB(buff
)->flags
= flags
;
1030 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
1032 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_PARTIAL
) {
1033 /* Copy and checksum data tail into the new buffer. */
1034 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
,
1035 skb_put(buff
, nsize
),
1040 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
1042 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1043 skb_split(skb
, buff
, len
);
1046 buff
->ip_summed
= skb
->ip_summed
;
1048 /* Looks stupid, but our code really uses when of
1049 * skbs, which it never sent before. --ANK
1051 TCP_SKB_CB(buff
)->when
= TCP_SKB_CB(skb
)->when
;
1052 buff
->tstamp
= skb
->tstamp
;
1054 old_factor
= tcp_skb_pcount(skb
);
1056 /* Fix up tso_factor for both original and new SKB. */
1057 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1058 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1060 /* If this packet has been sent out already, we must
1061 * adjust the various packet counters.
1063 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
1064 int diff
= old_factor
- tcp_skb_pcount(skb
) -
1065 tcp_skb_pcount(buff
);
1068 tcp_adjust_pcount(sk
, skb
, diff
);
1071 /* Link BUFF into the send queue. */
1072 skb_header_release(buff
);
1073 tcp_insert_write_queue_after(skb
, buff
, sk
);
1078 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1079 * eventually). The difference is that pulled data not copied, but
1080 * immediately discarded.
1082 static void __pskb_trim_head(struct sk_buff
*skb
, int len
)
1088 for (i
= 0; i
< skb_shinfo(skb
)->nr_frags
; i
++) {
1089 if (skb_shinfo(skb
)->frags
[i
].size
<= eat
) {
1090 put_page(skb_shinfo(skb
)->frags
[i
].page
);
1091 eat
-= skb_shinfo(skb
)->frags
[i
].size
;
1093 skb_shinfo(skb
)->frags
[k
] = skb_shinfo(skb
)->frags
[i
];
1095 skb_shinfo(skb
)->frags
[k
].page_offset
+= eat
;
1096 skb_shinfo(skb
)->frags
[k
].size
-= eat
;
1102 skb_shinfo(skb
)->nr_frags
= k
;
1104 skb_reset_tail_pointer(skb
);
1105 skb
->data_len
-= len
;
1106 skb
->len
= skb
->data_len
;
1109 /* Remove acked data from a packet in the transmit queue. */
1110 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
1112 if (skb_cloned(skb
) && pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
1115 /* If len == headlen, we avoid __skb_pull to preserve alignment. */
1116 if (unlikely(len
< skb_headlen(skb
)))
1117 __skb_pull(skb
, len
);
1119 __pskb_trim_head(skb
, len
- skb_headlen(skb
));
1121 TCP_SKB_CB(skb
)->seq
+= len
;
1122 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1124 skb
->truesize
-= len
;
1125 sk
->sk_wmem_queued
-= len
;
1126 sk_mem_uncharge(sk
, len
);
1127 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
1129 /* Any change of skb->len requires recalculation of tso
1132 if (tcp_skb_pcount(skb
) > 1)
1133 tcp_set_skb_tso_segs(sk
, skb
, tcp_current_mss(sk
));
1138 /* Calculate MSS. Not accounting for SACKs here. */
1139 int tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1141 struct tcp_sock
*tp
= tcp_sk(sk
);
1142 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1145 /* Calculate base mss without TCP options:
1146 It is MMS_S - sizeof(tcphdr) of rfc1122
1148 mss_now
= pmtu
- icsk
->icsk_af_ops
->net_header_len
- sizeof(struct tcphdr
);
1150 /* Clamp it (mss_clamp does not include tcp options) */
1151 if (mss_now
> tp
->rx_opt
.mss_clamp
)
1152 mss_now
= tp
->rx_opt
.mss_clamp
;
1154 /* Now subtract optional transport overhead */
1155 mss_now
-= icsk
->icsk_ext_hdr_len
;
1157 /* Then reserve room for full set of TCP options and 8 bytes of data */
1161 /* Now subtract TCP options size, not including SACKs */
1162 mss_now
-= tp
->tcp_header_len
- sizeof(struct tcphdr
);
1167 /* Inverse of above */
1168 int tcp_mss_to_mtu(struct sock
*sk
, int mss
)
1170 struct tcp_sock
*tp
= tcp_sk(sk
);
1171 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1175 tp
->tcp_header_len
+
1176 icsk
->icsk_ext_hdr_len
+
1177 icsk
->icsk_af_ops
->net_header_len
;
1182 /* MTU probing init per socket */
1183 void tcp_mtup_init(struct sock
*sk
)
1185 struct tcp_sock
*tp
= tcp_sk(sk
);
1186 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1188 icsk
->icsk_mtup
.enabled
= sysctl_tcp_mtu_probing
> 1;
1189 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+ sizeof(struct tcphdr
) +
1190 icsk
->icsk_af_ops
->net_header_len
;
1191 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, sysctl_tcp_base_mss
);
1192 icsk
->icsk_mtup
.probe_size
= 0;
1195 /* This function synchronize snd mss to current pmtu/exthdr set.
1197 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1198 for TCP options, but includes only bare TCP header.
1200 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1201 It is minimum of user_mss and mss received with SYN.
1202 It also does not include TCP options.
1204 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1206 tp->mss_cache is current effective sending mss, including
1207 all tcp options except for SACKs. It is evaluated,
1208 taking into account current pmtu, but never exceeds
1209 tp->rx_opt.mss_clamp.
1211 NOTE1. rfc1122 clearly states that advertised MSS
1212 DOES NOT include either tcp or ip options.
1214 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1215 are READ ONLY outside this function. --ANK (980731)
1217 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
1219 struct tcp_sock
*tp
= tcp_sk(sk
);
1220 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1223 if (icsk
->icsk_mtup
.search_high
> pmtu
)
1224 icsk
->icsk_mtup
.search_high
= pmtu
;
1226 mss_now
= tcp_mtu_to_mss(sk
, pmtu
);
1227 mss_now
= tcp_bound_to_half_wnd(tp
, mss_now
);
1229 /* And store cached results */
1230 icsk
->icsk_pmtu_cookie
= pmtu
;
1231 if (icsk
->icsk_mtup
.enabled
)
1232 mss_now
= min(mss_now
, tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
));
1233 tp
->mss_cache
= mss_now
;
1238 /* Compute the current effective MSS, taking SACKs and IP options,
1239 * and even PMTU discovery events into account.
1241 unsigned int tcp_current_mss(struct sock
*sk
)
1243 struct tcp_sock
*tp
= tcp_sk(sk
);
1244 struct dst_entry
*dst
= __sk_dst_get(sk
);
1246 unsigned header_len
;
1247 struct tcp_out_options opts
;
1248 struct tcp_md5sig_key
*md5
;
1250 mss_now
= tp
->mss_cache
;
1253 u32 mtu
= dst_mtu(dst
);
1254 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
1255 mss_now
= tcp_sync_mss(sk
, mtu
);
1258 header_len
= tcp_established_options(sk
, NULL
, &opts
, &md5
) +
1259 sizeof(struct tcphdr
);
1260 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1261 * some common options. If this is an odd packet (because we have SACK
1262 * blocks etc) then our calculated header_len will be different, and
1263 * we have to adjust mss_now correspondingly */
1264 if (header_len
!= tp
->tcp_header_len
) {
1265 int delta
= (int) header_len
- tp
->tcp_header_len
;
1272 /* Congestion window validation. (RFC2861) */
1273 static void tcp_cwnd_validate(struct sock
*sk
)
1275 struct tcp_sock
*tp
= tcp_sk(sk
);
1277 if (tp
->packets_out
>= tp
->snd_cwnd
) {
1278 /* Network is feed fully. */
1279 tp
->snd_cwnd_used
= 0;
1280 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1282 /* Network starves. */
1283 if (tp
->packets_out
> tp
->snd_cwnd_used
)
1284 tp
->snd_cwnd_used
= tp
->packets_out
;
1286 if (sysctl_tcp_slow_start_after_idle
&&
1287 (s32
)(tcp_time_stamp
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
)
1288 tcp_cwnd_application_limited(sk
);
1292 /* Returns the portion of skb which can be sent right away without
1293 * introducing MSS oddities to segment boundaries. In rare cases where
1294 * mss_now != mss_cache, we will request caller to create a small skb
1295 * per input skb which could be mostly avoided here (if desired).
1297 * We explicitly want to create a request for splitting write queue tail
1298 * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1299 * thus all the complexity (cwnd_len is always MSS multiple which we
1300 * return whenever allowed by the other factors). Basically we need the
1301 * modulo only when the receiver window alone is the limiting factor or
1302 * when we would be allowed to send the split-due-to-Nagle skb fully.
1304 static unsigned int tcp_mss_split_point(struct sock
*sk
, struct sk_buff
*skb
,
1305 unsigned int mss_now
, unsigned int cwnd
)
1307 struct tcp_sock
*tp
= tcp_sk(sk
);
1308 u32 needed
, window
, cwnd_len
;
1310 window
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1311 cwnd_len
= mss_now
* cwnd
;
1313 if (likely(cwnd_len
<= window
&& skb
!= tcp_write_queue_tail(sk
)))
1316 needed
= min(skb
->len
, window
);
1318 if (cwnd_len
<= needed
)
1321 return needed
- needed
% mss_now
;
1324 /* Can at least one segment of SKB be sent right now, according to the
1325 * congestion window rules? If so, return how many segments are allowed.
1327 static inline unsigned int tcp_cwnd_test(struct tcp_sock
*tp
,
1328 struct sk_buff
*skb
)
1330 u32 in_flight
, cwnd
;
1332 /* Don't be strict about the congestion window for the final FIN. */
1333 if ((TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
) &&
1334 tcp_skb_pcount(skb
) == 1)
1337 in_flight
= tcp_packets_in_flight(tp
);
1338 cwnd
= tp
->snd_cwnd
;
1339 if (in_flight
< cwnd
)
1340 return (cwnd
- in_flight
);
1345 /* Intialize TSO state of a skb.
1346 * This must be invoked the first time we consider transmitting
1347 * SKB onto the wire.
1349 static int tcp_init_tso_segs(struct sock
*sk
, struct sk_buff
*skb
,
1350 unsigned int mss_now
)
1352 int tso_segs
= tcp_skb_pcount(skb
);
1354 if (!tso_segs
|| (tso_segs
> 1 && tcp_skb_mss(skb
) != mss_now
)) {
1355 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1356 tso_segs
= tcp_skb_pcount(skb
);
1361 /* Minshall's variant of the Nagle send check. */
1362 static inline int tcp_minshall_check(const struct tcp_sock
*tp
)
1364 return after(tp
->snd_sml
, tp
->snd_una
) &&
1365 !after(tp
->snd_sml
, tp
->snd_nxt
);
1368 /* Return 0, if packet can be sent now without violation Nagle's rules:
1369 * 1. It is full sized.
1370 * 2. Or it contains FIN. (already checked by caller)
1371 * 3. Or TCP_NODELAY was set.
1372 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1373 * With Minshall's modification: all sent small packets are ACKed.
1375 static inline int tcp_nagle_check(const struct tcp_sock
*tp
,
1376 const struct sk_buff
*skb
,
1377 unsigned mss_now
, int nonagle
)
1379 return (skb
->len
< mss_now
&&
1380 ((nonagle
& TCP_NAGLE_CORK
) ||
1381 (!nonagle
&& tp
->packets_out
&& tcp_minshall_check(tp
))));
1384 /* Return non-zero if the Nagle test allows this packet to be
1387 static inline int tcp_nagle_test(struct tcp_sock
*tp
, struct sk_buff
*skb
,
1388 unsigned int cur_mss
, int nonagle
)
1390 /* Nagle rule does not apply to frames, which sit in the middle of the
1391 * write_queue (they have no chances to get new data).
1393 * This is implemented in the callers, where they modify the 'nonagle'
1394 * argument based upon the location of SKB in the send queue.
1396 if (nonagle
& TCP_NAGLE_PUSH
)
1399 /* Don't use the nagle rule for urgent data (or for the final FIN).
1400 * Nagle can be ignored during F-RTO too (see RFC4138).
1402 if (tcp_urg_mode(tp
) || (tp
->frto_counter
== 2) ||
1403 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
))
1406 if (!tcp_nagle_check(tp
, skb
, cur_mss
, nonagle
))
1412 /* Does at least the first segment of SKB fit into the send window? */
1413 static inline int tcp_snd_wnd_test(struct tcp_sock
*tp
, struct sk_buff
*skb
,
1414 unsigned int cur_mss
)
1416 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1418 if (skb
->len
> cur_mss
)
1419 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
1421 return !after(end_seq
, tcp_wnd_end(tp
));
1424 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1425 * should be put on the wire right now. If so, it returns the number of
1426 * packets allowed by the congestion window.
1428 static unsigned int tcp_snd_test(struct sock
*sk
, struct sk_buff
*skb
,
1429 unsigned int cur_mss
, int nonagle
)
1431 struct tcp_sock
*tp
= tcp_sk(sk
);
1432 unsigned int cwnd_quota
;
1434 tcp_init_tso_segs(sk
, skb
, cur_mss
);
1436 if (!tcp_nagle_test(tp
, skb
, cur_mss
, nonagle
))
1439 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1440 if (cwnd_quota
&& !tcp_snd_wnd_test(tp
, skb
, cur_mss
))
1446 /* Test if sending is allowed right now. */
1447 int tcp_may_send_now(struct sock
*sk
)
1449 struct tcp_sock
*tp
= tcp_sk(sk
);
1450 struct sk_buff
*skb
= tcp_send_head(sk
);
1453 tcp_snd_test(sk
, skb
, tcp_current_mss(sk
),
1454 (tcp_skb_is_last(sk
, skb
) ?
1455 tp
->nonagle
: TCP_NAGLE_PUSH
)));
1458 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1459 * which is put after SKB on the list. It is very much like
1460 * tcp_fragment() except that it may make several kinds of assumptions
1461 * in order to speed up the splitting operation. In particular, we
1462 * know that all the data is in scatter-gather pages, and that the
1463 * packet has never been sent out before (and thus is not cloned).
1465 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
,
1466 unsigned int mss_now
)
1468 struct sk_buff
*buff
;
1469 int nlen
= skb
->len
- len
;
1472 /* All of a TSO frame must be composed of paged data. */
1473 if (skb
->len
!= skb
->data_len
)
1474 return tcp_fragment(sk
, skb
, len
, mss_now
);
1476 buff
= sk_stream_alloc_skb(sk
, 0, GFP_ATOMIC
);
1477 if (unlikely(buff
== NULL
))
1480 sk
->sk_wmem_queued
+= buff
->truesize
;
1481 sk_mem_charge(sk
, buff
->truesize
);
1482 buff
->truesize
+= nlen
;
1483 skb
->truesize
-= nlen
;
1485 /* Correct the sequence numbers. */
1486 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1487 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1488 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1490 /* PSH and FIN should only be set in the second packet. */
1491 flags
= TCP_SKB_CB(skb
)->flags
;
1492 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
| TCPCB_FLAG_PSH
);
1493 TCP_SKB_CB(buff
)->flags
= flags
;
1495 /* This packet was never sent out yet, so no SACK bits. */
1496 TCP_SKB_CB(buff
)->sacked
= 0;
1498 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_PARTIAL
;
1499 skb_split(skb
, buff
, len
);
1501 /* Fix up tso_factor for both original and new SKB. */
1502 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1503 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1505 /* Link BUFF into the send queue. */
1506 skb_header_release(buff
);
1507 tcp_insert_write_queue_after(skb
, buff
, sk
);
1512 /* Try to defer sending, if possible, in order to minimize the amount
1513 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1515 * This algorithm is from John Heffner.
1517 static int tcp_tso_should_defer(struct sock
*sk
, struct sk_buff
*skb
)
1519 struct tcp_sock
*tp
= tcp_sk(sk
);
1520 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1521 u32 send_win
, cong_win
, limit
, in_flight
;
1524 if (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
)
1527 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
1530 /* Defer for less than two clock ticks. */
1531 if (tp
->tso_deferred
&&
1532 (((u32
)jiffies
<< 1) >> 1) - (tp
->tso_deferred
>> 1) > 1)
1535 in_flight
= tcp_packets_in_flight(tp
);
1537 BUG_ON(tcp_skb_pcount(skb
) <= 1 || (tp
->snd_cwnd
<= in_flight
));
1539 send_win
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1541 /* From in_flight test above, we know that cwnd > in_flight. */
1542 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
1544 limit
= min(send_win
, cong_win
);
1546 /* If a full-sized TSO skb can be sent, do it. */
1547 if (limit
>= sk
->sk_gso_max_size
)
1550 /* Middle in queue won't get any more data, full sendable already? */
1551 if ((skb
!= tcp_write_queue_tail(sk
)) && (limit
>= skb
->len
))
1554 win_divisor
= ACCESS_ONCE(sysctl_tcp_tso_win_divisor
);
1556 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
1558 /* If at least some fraction of a window is available,
1561 chunk
/= win_divisor
;
1565 /* Different approach, try not to defer past a single
1566 * ACK. Receiver should ACK every other full sized
1567 * frame, so if we have space for more than 3 frames
1570 if (limit
> tcp_max_burst(tp
) * tp
->mss_cache
)
1574 /* Ok, it looks like it is advisable to defer. */
1575 tp
->tso_deferred
= 1 | (jiffies
<< 1);
1580 tp
->tso_deferred
= 0;
1584 /* Create a new MTU probe if we are ready.
1585 * MTU probe is regularly attempting to increase the path MTU by
1586 * deliberately sending larger packets. This discovers routing
1587 * changes resulting in larger path MTUs.
1589 * Returns 0 if we should wait to probe (no cwnd available),
1590 * 1 if a probe was sent,
1593 static int tcp_mtu_probe(struct sock
*sk
)
1595 struct tcp_sock
*tp
= tcp_sk(sk
);
1596 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1597 struct sk_buff
*skb
, *nskb
, *next
;
1604 /* Not currently probing/verifying,
1606 * have enough cwnd, and
1607 * not SACKing (the variable headers throw things off) */
1608 if (!icsk
->icsk_mtup
.enabled
||
1609 icsk
->icsk_mtup
.probe_size
||
1610 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
||
1611 tp
->snd_cwnd
< 11 ||
1612 tp
->rx_opt
.num_sacks
|| tp
->rx_opt
.dsack
)
1615 /* Very simple search strategy: just double the MSS. */
1616 mss_now
= tcp_current_mss(sk
);
1617 probe_size
= 2 * tp
->mss_cache
;
1618 size_needed
= probe_size
+ (tp
->reordering
+ 1) * tp
->mss_cache
;
1619 if (probe_size
> tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_high
)) {
1620 /* TODO: set timer for probe_converge_event */
1624 /* Have enough data in the send queue to probe? */
1625 if (tp
->write_seq
- tp
->snd_nxt
< size_needed
)
1628 if (tp
->snd_wnd
< size_needed
)
1630 if (after(tp
->snd_nxt
+ size_needed
, tcp_wnd_end(tp
)))
1633 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1634 if (tcp_packets_in_flight(tp
) + 2 > tp
->snd_cwnd
) {
1635 if (!tcp_packets_in_flight(tp
))
1641 /* We're allowed to probe. Build it now. */
1642 if ((nskb
= sk_stream_alloc_skb(sk
, probe_size
, GFP_ATOMIC
)) == NULL
)
1644 sk
->sk_wmem_queued
+= nskb
->truesize
;
1645 sk_mem_charge(sk
, nskb
->truesize
);
1647 skb
= tcp_send_head(sk
);
1649 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(skb
)->seq
;
1650 TCP_SKB_CB(nskb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ probe_size
;
1651 TCP_SKB_CB(nskb
)->flags
= TCPCB_FLAG_ACK
;
1652 TCP_SKB_CB(nskb
)->sacked
= 0;
1654 nskb
->ip_summed
= skb
->ip_summed
;
1656 tcp_insert_write_queue_before(nskb
, skb
, sk
);
1659 tcp_for_write_queue_from_safe(skb
, next
, sk
) {
1660 copy
= min_t(int, skb
->len
, probe_size
- len
);
1661 if (nskb
->ip_summed
)
1662 skb_copy_bits(skb
, 0, skb_put(nskb
, copy
), copy
);
1664 nskb
->csum
= skb_copy_and_csum_bits(skb
, 0,
1665 skb_put(nskb
, copy
),
1668 if (skb
->len
<= copy
) {
1669 /* We've eaten all the data from this skb.
1671 TCP_SKB_CB(nskb
)->flags
|= TCP_SKB_CB(skb
)->flags
;
1672 tcp_unlink_write_queue(skb
, sk
);
1673 sk_wmem_free_skb(sk
, skb
);
1675 TCP_SKB_CB(nskb
)->flags
|= TCP_SKB_CB(skb
)->flags
&
1676 ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
1677 if (!skb_shinfo(skb
)->nr_frags
) {
1678 skb_pull(skb
, copy
);
1679 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1680 skb
->csum
= csum_partial(skb
->data
,
1683 __pskb_trim_head(skb
, copy
);
1684 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1686 TCP_SKB_CB(skb
)->seq
+= copy
;
1691 if (len
>= probe_size
)
1694 tcp_init_tso_segs(sk
, nskb
, nskb
->len
);
1696 /* We're ready to send. If this fails, the probe will
1697 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1698 TCP_SKB_CB(nskb
)->when
= tcp_time_stamp
;
1699 if (!tcp_transmit_skb(sk
, nskb
, 1, GFP_ATOMIC
)) {
1700 /* Decrement cwnd here because we are sending
1701 * effectively two packets. */
1703 tcp_event_new_data_sent(sk
, nskb
);
1705 icsk
->icsk_mtup
.probe_size
= tcp_mss_to_mtu(sk
, nskb
->len
);
1706 tp
->mtu_probe
.probe_seq_start
= TCP_SKB_CB(nskb
)->seq
;
1707 tp
->mtu_probe
.probe_seq_end
= TCP_SKB_CB(nskb
)->end_seq
;
1715 /* This routine writes packets to the network. It advances the
1716 * send_head. This happens as incoming acks open up the remote
1719 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1720 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1721 * account rare use of URG, this is not a big flaw.
1723 * Returns 1, if no segments are in flight and we have queued segments, but
1724 * cannot send anything now because of SWS or another problem.
1726 static int tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
,
1727 int push_one
, gfp_t gfp
)
1729 struct tcp_sock
*tp
= tcp_sk(sk
);
1730 struct sk_buff
*skb
;
1731 unsigned int tso_segs
, sent_pkts
;
1738 /* Do MTU probing. */
1739 result
= tcp_mtu_probe(sk
);
1742 } else if (result
> 0) {
1747 while ((skb
= tcp_send_head(sk
))) {
1750 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1753 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1757 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
)))
1760 if (tso_segs
== 1) {
1761 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
1762 (tcp_skb_is_last(sk
, skb
) ?
1763 nonagle
: TCP_NAGLE_PUSH
))))
1766 if (!push_one
&& tcp_tso_should_defer(sk
, skb
))
1771 if (tso_segs
> 1 && !tcp_urg_mode(tp
))
1772 limit
= tcp_mss_split_point(sk
, skb
, mss_now
,
1775 if (skb
->len
> limit
&&
1776 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
)))
1779 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1781 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, gfp
)))
1784 /* Advance the send_head. This one is sent out.
1785 * This call will increment packets_out.
1787 tcp_event_new_data_sent(sk
, skb
);
1789 tcp_minshall_update(tp
, mss_now
, skb
);
1796 if (likely(sent_pkts
)) {
1797 tcp_cwnd_validate(sk
);
1800 return !tp
->packets_out
&& tcp_send_head(sk
);
1803 /* Push out any pending frames which were held back due to
1804 * TCP_CORK or attempt at coalescing tiny packets.
1805 * The socket must be locked by the caller.
1807 void __tcp_push_pending_frames(struct sock
*sk
, unsigned int cur_mss
,
1810 /* If we are closed, the bytes will have to remain here.
1811 * In time closedown will finish, we empty the write queue and
1812 * all will be happy.
1814 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
1817 if (tcp_write_xmit(sk
, cur_mss
, nonagle
, 0, GFP_ATOMIC
))
1818 tcp_check_probe_timer(sk
);
1821 /* Send _single_ skb sitting at the send head. This function requires
1822 * true push pending frames to setup probe timer etc.
1824 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
1826 struct sk_buff
*skb
= tcp_send_head(sk
);
1828 BUG_ON(!skb
|| skb
->len
< mss_now
);
1830 tcp_write_xmit(sk
, mss_now
, TCP_NAGLE_PUSH
, 1, sk
->sk_allocation
);
1833 /* This function returns the amount that we can raise the
1834 * usable window based on the following constraints
1836 * 1. The window can never be shrunk once it is offered (RFC 793)
1837 * 2. We limit memory per socket
1840 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1841 * RECV.NEXT + RCV.WIN fixed until:
1842 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1844 * i.e. don't raise the right edge of the window until you can raise
1845 * it at least MSS bytes.
1847 * Unfortunately, the recommended algorithm breaks header prediction,
1848 * since header prediction assumes th->window stays fixed.
1850 * Strictly speaking, keeping th->window fixed violates the receiver
1851 * side SWS prevention criteria. The problem is that under this rule
1852 * a stream of single byte packets will cause the right side of the
1853 * window to always advance by a single byte.
1855 * Of course, if the sender implements sender side SWS prevention
1856 * then this will not be a problem.
1858 * BSD seems to make the following compromise:
1860 * If the free space is less than the 1/4 of the maximum
1861 * space available and the free space is less than 1/2 mss,
1862 * then set the window to 0.
1863 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1864 * Otherwise, just prevent the window from shrinking
1865 * and from being larger than the largest representable value.
1867 * This prevents incremental opening of the window in the regime
1868 * where TCP is limited by the speed of the reader side taking
1869 * data out of the TCP receive queue. It does nothing about
1870 * those cases where the window is constrained on the sender side
1871 * because the pipeline is full.
1873 * BSD also seems to "accidentally" limit itself to windows that are a
1874 * multiple of MSS, at least until the free space gets quite small.
1875 * This would appear to be a side effect of the mbuf implementation.
1876 * Combining these two algorithms results in the observed behavior
1877 * of having a fixed window size at almost all times.
1879 * Below we obtain similar behavior by forcing the offered window to
1880 * a multiple of the mss when it is feasible to do so.
1882 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1883 * Regular options like TIMESTAMP are taken into account.
1885 u32
__tcp_select_window(struct sock
*sk
)
1887 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1888 struct tcp_sock
*tp
= tcp_sk(sk
);
1889 /* MSS for the peer's data. Previous versions used mss_clamp
1890 * here. I don't know if the value based on our guesses
1891 * of peer's MSS is better for the performance. It's more correct
1892 * but may be worse for the performance because of rcv_mss
1893 * fluctuations. --SAW 1998/11/1
1895 int mss
= icsk
->icsk_ack
.rcv_mss
;
1896 int free_space
= tcp_space(sk
);
1897 int full_space
= min_t(int, tp
->window_clamp
, tcp_full_space(sk
));
1900 if (mss
> full_space
)
1903 if (free_space
< (full_space
>> 1)) {
1904 icsk
->icsk_ack
.quick
= 0;
1906 if (tcp_memory_pressure
)
1907 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
,
1910 if (free_space
< mss
)
1914 if (free_space
> tp
->rcv_ssthresh
)
1915 free_space
= tp
->rcv_ssthresh
;
1917 /* Don't do rounding if we are using window scaling, since the
1918 * scaled window will not line up with the MSS boundary anyway.
1920 window
= tp
->rcv_wnd
;
1921 if (tp
->rx_opt
.rcv_wscale
) {
1922 window
= free_space
;
1924 /* Advertise enough space so that it won't get scaled away.
1925 * Import case: prevent zero window announcement if
1926 * 1<<rcv_wscale > mss.
1928 if (((window
>> tp
->rx_opt
.rcv_wscale
) << tp
->rx_opt
.rcv_wscale
) != window
)
1929 window
= (((window
>> tp
->rx_opt
.rcv_wscale
) + 1)
1930 << tp
->rx_opt
.rcv_wscale
);
1932 /* Get the largest window that is a nice multiple of mss.
1933 * Window clamp already applied above.
1934 * If our current window offering is within 1 mss of the
1935 * free space we just keep it. This prevents the divide
1936 * and multiply from happening most of the time.
1937 * We also don't do any window rounding when the free space
1940 if (window
<= free_space
- mss
|| window
> free_space
)
1941 window
= (free_space
/ mss
) * mss
;
1942 else if (mss
== full_space
&&
1943 free_space
> window
+ (full_space
>> 1))
1944 window
= free_space
;
1950 /* Collapses two adjacent SKB's during retransmission. */
1951 static void tcp_collapse_retrans(struct sock
*sk
, struct sk_buff
*skb
)
1953 struct tcp_sock
*tp
= tcp_sk(sk
);
1954 struct sk_buff
*next_skb
= tcp_write_queue_next(sk
, skb
);
1955 int skb_size
, next_skb_size
;
1957 skb_size
= skb
->len
;
1958 next_skb_size
= next_skb
->len
;
1960 BUG_ON(tcp_skb_pcount(skb
) != 1 || tcp_skb_pcount(next_skb
) != 1);
1962 tcp_highest_sack_combine(sk
, next_skb
, skb
);
1964 tcp_unlink_write_queue(next_skb
, sk
);
1966 skb_copy_from_linear_data(next_skb
, skb_put(skb
, next_skb_size
),
1969 if (next_skb
->ip_summed
== CHECKSUM_PARTIAL
)
1970 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1972 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1973 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
1975 /* Update sequence range on original skb. */
1976 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
1978 /* Merge over control information. This moves PSH/FIN etc. over */
1979 TCP_SKB_CB(skb
)->flags
|= TCP_SKB_CB(next_skb
)->flags
;
1981 /* All done, get rid of second SKB and account for it so
1982 * packet counting does not break.
1984 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
& TCPCB_EVER_RETRANS
;
1986 /* changed transmit queue under us so clear hints */
1987 tcp_clear_retrans_hints_partial(tp
);
1988 if (next_skb
== tp
->retransmit_skb_hint
)
1989 tp
->retransmit_skb_hint
= skb
;
1991 tcp_adjust_pcount(sk
, next_skb
, tcp_skb_pcount(next_skb
));
1993 sk_wmem_free_skb(sk
, next_skb
);
1996 /* Check if coalescing SKBs is legal. */
1997 static int tcp_can_collapse(struct sock
*sk
, struct sk_buff
*skb
)
1999 if (tcp_skb_pcount(skb
) > 1)
2001 /* TODO: SACK collapsing could be used to remove this condition */
2002 if (skb_shinfo(skb
)->nr_frags
!= 0)
2004 if (skb_cloned(skb
))
2006 if (skb
== tcp_send_head(sk
))
2008 /* Some heurestics for collapsing over SACK'd could be invented */
2009 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
2015 /* Collapse packets in the retransmit queue to make to create
2016 * less packets on the wire. This is only done on retransmission.
2018 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*to
,
2021 struct tcp_sock
*tp
= tcp_sk(sk
);
2022 struct sk_buff
*skb
= to
, *tmp
;
2025 if (!sysctl_tcp_retrans_collapse
)
2027 if (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_SYN
)
2030 tcp_for_write_queue_from_safe(skb
, tmp
, sk
) {
2031 if (!tcp_can_collapse(sk
, skb
))
2043 /* Punt if not enough space exists in the first SKB for
2044 * the data in the second
2046 if (skb
->len
> skb_tailroom(to
))
2049 if (after(TCP_SKB_CB(skb
)->end_seq
, tcp_wnd_end(tp
)))
2052 tcp_collapse_retrans(sk
, to
);
2056 /* This retransmits one SKB. Policy decisions and retransmit queue
2057 * state updates are done by the caller. Returns non-zero if an
2058 * error occurred which prevented the send.
2060 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
2062 struct tcp_sock
*tp
= tcp_sk(sk
);
2063 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2064 unsigned int cur_mss
;
2067 /* Inconslusive MTU probe */
2068 if (icsk
->icsk_mtup
.probe_size
) {
2069 icsk
->icsk_mtup
.probe_size
= 0;
2072 /* Do not sent more than we queued. 1/4 is reserved for possible
2073 * copying overhead: fragmentation, tunneling, mangling etc.
2075 if (atomic_read(&sk
->sk_wmem_alloc
) >
2076 min(sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2), sk
->sk_sndbuf
))
2079 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
2080 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
2082 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
2086 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
2087 return -EHOSTUNREACH
; /* Routing failure or similar. */
2089 cur_mss
= tcp_current_mss(sk
);
2091 /* If receiver has shrunk his window, and skb is out of
2092 * new window, do not retransmit it. The exception is the
2093 * case, when window is shrunk to zero. In this case
2094 * our retransmit serves as a zero window probe.
2096 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
)) &&
2097 TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
2100 if (skb
->len
> cur_mss
) {
2101 if (tcp_fragment(sk
, skb
, cur_mss
, cur_mss
))
2102 return -ENOMEM
; /* We'll try again later. */
2104 int oldpcount
= tcp_skb_pcount(skb
);
2106 if (unlikely(oldpcount
> 1)) {
2107 tcp_init_tso_segs(sk
, skb
, cur_mss
);
2108 tcp_adjust_pcount(sk
, skb
, oldpcount
- tcp_skb_pcount(skb
));
2112 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
2114 /* Some Solaris stacks overoptimize and ignore the FIN on a
2115 * retransmit when old data is attached. So strip it off
2116 * since it is cheap to do so and saves bytes on the network.
2119 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
) &&
2120 tp
->snd_una
== (TCP_SKB_CB(skb
)->end_seq
- 1)) {
2121 if (!pskb_trim(skb
, 0)) {
2122 /* Reuse, even though it does some unnecessary work */
2123 tcp_init_nondata_skb(skb
, TCP_SKB_CB(skb
)->end_seq
- 1,
2124 TCP_SKB_CB(skb
)->flags
);
2125 skb
->ip_summed
= CHECKSUM_NONE
;
2129 /* Make a copy, if the first transmission SKB clone we made
2130 * is still in somebody's hands, else make a clone.
2132 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2134 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2137 /* Update global TCP statistics. */
2138 TCP_INC_STATS(sock_net(sk
), TCP_MIB_RETRANSSEGS
);
2140 tp
->total_retrans
++;
2142 #if FASTRETRANS_DEBUG > 0
2143 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2144 if (net_ratelimit())
2145 printk(KERN_DEBUG
"retrans_out leaked.\n");
2148 if (!tp
->retrans_out
)
2149 tp
->lost_retrans_low
= tp
->snd_nxt
;
2150 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
2151 tp
->retrans_out
+= tcp_skb_pcount(skb
);
2153 /* Save stamp of the first retransmit. */
2154 if (!tp
->retrans_stamp
)
2155 tp
->retrans_stamp
= TCP_SKB_CB(skb
)->when
;
2159 /* snd_nxt is stored to detect loss of retransmitted segment,
2160 * see tcp_input.c tcp_sacktag_write_queue().
2162 TCP_SKB_CB(skb
)->ack_seq
= tp
->snd_nxt
;
2167 /* Check if we forward retransmits are possible in the current
2168 * window/congestion state.
2170 static int tcp_can_forward_retransmit(struct sock
*sk
)
2172 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2173 struct tcp_sock
*tp
= tcp_sk(sk
);
2175 /* Forward retransmissions are possible only during Recovery. */
2176 if (icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
2179 /* No forward retransmissions in Reno are possible. */
2180 if (tcp_is_reno(tp
))
2183 /* Yeah, we have to make difficult choice between forward transmission
2184 * and retransmission... Both ways have their merits...
2186 * For now we do not retransmit anything, while we have some new
2187 * segments to send. In the other cases, follow rule 3 for
2188 * NextSeg() specified in RFC3517.
2191 if (tcp_may_send_now(sk
))
2197 /* This gets called after a retransmit timeout, and the initially
2198 * retransmitted data is acknowledged. It tries to continue
2199 * resending the rest of the retransmit queue, until either
2200 * we've sent it all or the congestion window limit is reached.
2201 * If doing SACK, the first ACK which comes back for a timeout
2202 * based retransmit packet might feed us FACK information again.
2203 * If so, we use it to avoid unnecessarily retransmissions.
2205 void tcp_xmit_retransmit_queue(struct sock
*sk
)
2207 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2208 struct tcp_sock
*tp
= tcp_sk(sk
);
2209 struct sk_buff
*skb
;
2210 struct sk_buff
*hole
= NULL
;
2213 int fwd_rexmitting
= 0;
2215 if (!tp
->packets_out
)
2219 tp
->retransmit_high
= tp
->snd_una
;
2221 if (tp
->retransmit_skb_hint
) {
2222 skb
= tp
->retransmit_skb_hint
;
2223 last_lost
= TCP_SKB_CB(skb
)->end_seq
;
2224 if (after(last_lost
, tp
->retransmit_high
))
2225 last_lost
= tp
->retransmit_high
;
2227 skb
= tcp_write_queue_head(sk
);
2228 last_lost
= tp
->snd_una
;
2231 tcp_for_write_queue_from(skb
, sk
) {
2232 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
2234 if (skb
== tcp_send_head(sk
))
2236 /* we could do better than to assign each time */
2238 tp
->retransmit_skb_hint
= skb
;
2240 /* Assume this retransmit will generate
2241 * only one packet for congestion window
2242 * calculation purposes. This works because
2243 * tcp_retransmit_skb() will chop up the
2244 * packet to be MSS sized and all the
2245 * packet counting works out.
2247 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
2250 if (fwd_rexmitting
) {
2252 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_highest_sack_seq(tp
)))
2254 mib_idx
= LINUX_MIB_TCPFORWARDRETRANS
;
2256 } else if (!before(TCP_SKB_CB(skb
)->seq
, tp
->retransmit_high
)) {
2257 tp
->retransmit_high
= last_lost
;
2258 if (!tcp_can_forward_retransmit(sk
))
2260 /* Backtrack if necessary to non-L'ed skb */
2268 } else if (!(sacked
& TCPCB_LOST
)) {
2269 if (hole
== NULL
&& !(sacked
& (TCPCB_SACKED_RETRANS
|TCPCB_SACKED_ACKED
)))
2274 last_lost
= TCP_SKB_CB(skb
)->end_seq
;
2275 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
2276 mib_idx
= LINUX_MIB_TCPFASTRETRANS
;
2278 mib_idx
= LINUX_MIB_TCPSLOWSTARTRETRANS
;
2281 if (sacked
& (TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))
2284 if (tcp_retransmit_skb(sk
, skb
))
2286 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2288 if (skb
== tcp_write_queue_head(sk
))
2289 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2290 inet_csk(sk
)->icsk_rto
,
2295 /* Send a fin. The caller locks the socket for us. This cannot be
2296 * allowed to fail queueing a FIN frame under any circumstances.
2298 void tcp_send_fin(struct sock
*sk
)
2300 struct tcp_sock
*tp
= tcp_sk(sk
);
2301 struct sk_buff
*skb
= tcp_write_queue_tail(sk
);
2304 /* Optimization, tack on the FIN if we have a queue of
2305 * unsent frames. But be careful about outgoing SACKS
2308 mss_now
= tcp_current_mss(sk
);
2310 if (tcp_send_head(sk
) != NULL
) {
2311 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_FIN
;
2312 TCP_SKB_CB(skb
)->end_seq
++;
2315 /* Socket is locked, keep trying until memory is available. */
2317 skb
= alloc_skb_fclone(MAX_TCP_HEADER
,
2324 /* Reserve space for headers and prepare control bits. */
2325 skb_reserve(skb
, MAX_TCP_HEADER
);
2326 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2327 tcp_init_nondata_skb(skb
, tp
->write_seq
,
2328 TCPCB_FLAG_ACK
| TCPCB_FLAG_FIN
);
2329 tcp_queue_skb(sk
, skb
);
2331 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_OFF
);
2334 /* We get here when a process closes a file descriptor (either due to
2335 * an explicit close() or as a byproduct of exit()'ing) and there
2336 * was unread data in the receive queue. This behavior is recommended
2337 * by RFC 2525, section 2.17. -DaveM
2339 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
2341 struct sk_buff
*skb
;
2343 /* NOTE: No TCP options attached and we never retransmit this. */
2344 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
2346 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
2350 /* Reserve space for headers and prepare control bits. */
2351 skb_reserve(skb
, MAX_TCP_HEADER
);
2352 tcp_init_nondata_skb(skb
, tcp_acceptable_seq(sk
),
2353 TCPCB_FLAG_ACK
| TCPCB_FLAG_RST
);
2355 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2356 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
2357 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
2359 TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTRSTS
);
2362 /* Send a crossed SYN-ACK during socket establishment.
2363 * WARNING: This routine must only be called when we have already sent
2364 * a SYN packet that crossed the incoming SYN that caused this routine
2365 * to get called. If this assumption fails then the initial rcv_wnd
2366 * and rcv_wscale values will not be correct.
2368 int tcp_send_synack(struct sock
*sk
)
2370 struct sk_buff
*skb
;
2372 skb
= tcp_write_queue_head(sk
);
2373 if (skb
== NULL
|| !(TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_SYN
)) {
2374 printk(KERN_DEBUG
"tcp_send_synack: wrong queue state\n");
2377 if (!(TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_ACK
)) {
2378 if (skb_cloned(skb
)) {
2379 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
2382 tcp_unlink_write_queue(skb
, sk
);
2383 skb_header_release(nskb
);
2384 __tcp_add_write_queue_head(sk
, nskb
);
2385 sk_wmem_free_skb(sk
, skb
);
2386 sk
->sk_wmem_queued
+= nskb
->truesize
;
2387 sk_mem_charge(sk
, nskb
->truesize
);
2391 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_ACK
;
2392 TCP_ECN_send_synack(tcp_sk(sk
), skb
);
2394 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2395 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2398 /* Prepare a SYN-ACK. */
2399 struct sk_buff
*tcp_make_synack(struct sock
*sk
, struct dst_entry
*dst
,
2400 struct request_sock
*req
,
2401 struct request_values
*rvp
)
2403 struct tcp_out_options opts
;
2404 struct tcp_extend_values
*xvp
= tcp_xv(rvp
);
2405 struct inet_request_sock
*ireq
= inet_rsk(req
);
2406 struct tcp_sock
*tp
= tcp_sk(sk
);
2407 const struct tcp_cookie_values
*cvp
= tp
->cookie_values
;
2409 struct sk_buff
*skb
;
2410 struct tcp_md5sig_key
*md5
;
2411 int tcp_header_size
;
2413 int s_data_desired
= 0;
2415 if (cvp
!= NULL
&& cvp
->s_data_constant
&& cvp
->s_data_desired
)
2416 s_data_desired
= cvp
->s_data_desired
;
2417 skb
= sock_wmalloc(sk
, MAX_TCP_HEADER
+ 15 + s_data_desired
, 1, GFP_ATOMIC
);
2421 /* Reserve space for headers. */
2422 skb_reserve(skb
, MAX_TCP_HEADER
);
2424 skb_dst_set(skb
, dst_clone(dst
));
2426 mss
= dst_metric(dst
, RTAX_ADVMSS
);
2427 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< mss
)
2428 mss
= tp
->rx_opt
.user_mss
;
2430 if (req
->rcv_wnd
== 0) { /* ignored for retransmitted syns */
2432 /* Set this up on the first call only */
2433 req
->window_clamp
= tp
->window_clamp
? : dst_metric(dst
, RTAX_WINDOW
);
2434 /* tcp_full_space because it is guaranteed to be the first packet */
2435 tcp_select_initial_window(tcp_full_space(sk
),
2436 mss
- (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0),
2441 dst_metric(dst
, RTAX_INITRWND
));
2442 ireq
->rcv_wscale
= rcv_wscale
;
2445 memset(&opts
, 0, sizeof(opts
));
2446 #ifdef CONFIG_SYN_COOKIES
2447 if (unlikely(req
->cookie_ts
))
2448 TCP_SKB_CB(skb
)->when
= cookie_init_timestamp(req
);
2451 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2452 tcp_header_size
= tcp_synack_options(sk
, req
, mss
,
2453 skb
, &opts
, &md5
, xvp
)
2456 skb_push(skb
, tcp_header_size
);
2457 skb_reset_transport_header(skb
);
2460 memset(th
, 0, sizeof(struct tcphdr
));
2463 TCP_ECN_make_synack(req
, th
);
2464 th
->source
= ireq
->loc_port
;
2465 th
->dest
= ireq
->rmt_port
;
2466 /* Setting of flags are superfluous here for callers (and ECE is
2467 * not even correctly set)
2469 tcp_init_nondata_skb(skb
, tcp_rsk(req
)->snt_isn
,
2470 TCPCB_FLAG_SYN
| TCPCB_FLAG_ACK
);
2472 if (OPTION_COOKIE_EXTENSION
& opts
.options
) {
2473 if (s_data_desired
) {
2474 u8
*buf
= skb_put(skb
, s_data_desired
);
2476 /* copy data directly from the listening socket. */
2477 memcpy(buf
, cvp
->s_data_payload
, s_data_desired
);
2478 TCP_SKB_CB(skb
)->end_seq
+= s_data_desired
;
2481 if (opts
.hash_size
> 0) {
2482 __u32 workspace
[SHA_WORKSPACE_WORDS
];
2483 u32
*mess
= &xvp
->cookie_bakery
[COOKIE_DIGEST_WORDS
];
2484 u32
*tail
= &mess
[COOKIE_MESSAGE_WORDS
-1];
2486 /* Secret recipe depends on the Timestamp, (future)
2487 * Sequence and Acknowledgment Numbers, Initiator
2488 * Cookie, and others handled by IP variant caller.
2490 *tail
-- ^= opts
.tsval
;
2491 *tail
-- ^= tcp_rsk(req
)->rcv_isn
+ 1;
2492 *tail
-- ^= TCP_SKB_CB(skb
)->seq
+ 1;
2495 *tail
-- ^= (((__force u32
)th
->dest
<< 16) | (__force u32
)th
->source
);
2496 *tail
-- ^= (u32
)(unsigned long)cvp
; /* per sockopt */
2498 sha_transform((__u32
*)&xvp
->cookie_bakery
[0],
2501 opts
.hash_location
=
2502 (__u8
*)&xvp
->cookie_bakery
[0];
2506 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
2507 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_isn
+ 1);
2509 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2510 th
->window
= htons(min(req
->rcv_wnd
, 65535U));
2511 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
);
2512 th
->doff
= (tcp_header_size
>> 2);
2513 TCP_ADD_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
, tcp_skb_pcount(skb
));
2515 #ifdef CONFIG_TCP_MD5SIG
2516 /* Okay, we have all we need - do the md5 hash if needed */
2518 tcp_rsk(req
)->af_specific
->calc_md5_hash(opts
.hash_location
,
2519 md5
, NULL
, req
, skb
);
2526 /* Do all connect socket setups that can be done AF independent. */
2527 static void tcp_connect_init(struct sock
*sk
)
2529 struct dst_entry
*dst
= __sk_dst_get(sk
);
2530 struct tcp_sock
*tp
= tcp_sk(sk
);
2533 /* We'll fix this up when we get a response from the other end.
2534 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2536 tp
->tcp_header_len
= sizeof(struct tcphdr
) +
2537 (sysctl_tcp_timestamps
? TCPOLEN_TSTAMP_ALIGNED
: 0);
2539 #ifdef CONFIG_TCP_MD5SIG
2540 if (tp
->af_specific
->md5_lookup(sk
, sk
) != NULL
)
2541 tp
->tcp_header_len
+= TCPOLEN_MD5SIG_ALIGNED
;
2544 /* If user gave his TCP_MAXSEG, record it to clamp */
2545 if (tp
->rx_opt
.user_mss
)
2546 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
2549 tcp_sync_mss(sk
, dst_mtu(dst
));
2551 if (!tp
->window_clamp
)
2552 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
2553 tp
->advmss
= dst_metric(dst
, RTAX_ADVMSS
);
2554 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< tp
->advmss
)
2555 tp
->advmss
= tp
->rx_opt
.user_mss
;
2557 tcp_initialize_rcv_mss(sk
);
2559 tcp_select_initial_window(tcp_full_space(sk
),
2560 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
2563 sysctl_tcp_window_scaling
,
2565 dst_metric(dst
, RTAX_INITRWND
));
2567 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
2568 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
2571 sock_reset_flag(sk
, SOCK_DONE
);
2574 tp
->snd_una
= tp
->write_seq
;
2575 tp
->snd_sml
= tp
->write_seq
;
2576 tp
->snd_up
= tp
->write_seq
;
2581 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
2582 inet_csk(sk
)->icsk_retransmits
= 0;
2583 tcp_clear_retrans(tp
);
2586 /* Build a SYN and send it off. */
2587 int tcp_connect(struct sock
*sk
)
2589 struct tcp_sock
*tp
= tcp_sk(sk
);
2590 struct sk_buff
*buff
;
2592 tcp_connect_init(sk
);
2594 buff
= alloc_skb_fclone(MAX_TCP_HEADER
+ 15, sk
->sk_allocation
);
2595 if (unlikely(buff
== NULL
))
2598 /* Reserve space for headers. */
2599 skb_reserve(buff
, MAX_TCP_HEADER
);
2601 tp
->snd_nxt
= tp
->write_seq
;
2602 tcp_init_nondata_skb(buff
, tp
->write_seq
++, TCPCB_FLAG_SYN
);
2603 TCP_ECN_send_syn(sk
, buff
);
2606 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2607 tp
->retrans_stamp
= TCP_SKB_CB(buff
)->when
;
2608 skb_header_release(buff
);
2609 __tcp_add_write_queue_tail(sk
, buff
);
2610 sk
->sk_wmem_queued
+= buff
->truesize
;
2611 sk_mem_charge(sk
, buff
->truesize
);
2612 tp
->packets_out
+= tcp_skb_pcount(buff
);
2613 tcp_transmit_skb(sk
, buff
, 1, sk
->sk_allocation
);
2615 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2616 * in order to make this packet get counted in tcpOutSegs.
2618 tp
->snd_nxt
= tp
->write_seq
;
2619 tp
->pushed_seq
= tp
->write_seq
;
2620 TCP_INC_STATS(sock_net(sk
), TCP_MIB_ACTIVEOPENS
);
2622 /* Timer for repeating the SYN until an answer. */
2623 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2624 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2628 /* Send out a delayed ack, the caller does the policy checking
2629 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2632 void tcp_send_delayed_ack(struct sock
*sk
)
2634 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2635 int ato
= icsk
->icsk_ack
.ato
;
2636 unsigned long timeout
;
2638 if (ato
> TCP_DELACK_MIN
) {
2639 const struct tcp_sock
*tp
= tcp_sk(sk
);
2640 int max_ato
= HZ
/ 2;
2642 if (icsk
->icsk_ack
.pingpong
||
2643 (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
2644 max_ato
= TCP_DELACK_MAX
;
2646 /* Slow path, intersegment interval is "high". */
2648 /* If some rtt estimate is known, use it to bound delayed ack.
2649 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2653 int rtt
= max(tp
->srtt
>> 3, TCP_DELACK_MIN
);
2659 ato
= min(ato
, max_ato
);
2662 /* Stay within the limit we were given */
2663 timeout
= jiffies
+ ato
;
2665 /* Use new timeout only if there wasn't a older one earlier. */
2666 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
2667 /* If delack timer was blocked or is about to expire,
2670 if (icsk
->icsk_ack
.blocked
||
2671 time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
2676 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
2677 timeout
= icsk
->icsk_ack
.timeout
;
2679 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
2680 icsk
->icsk_ack
.timeout
= timeout
;
2681 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
2684 /* This routine sends an ack and also updates the window. */
2685 void tcp_send_ack(struct sock
*sk
)
2687 struct sk_buff
*buff
;
2689 /* If we have been reset, we may not send again. */
2690 if (sk
->sk_state
== TCP_CLOSE
)
2693 /* We are not putting this on the write queue, so
2694 * tcp_transmit_skb() will set the ownership to this
2697 buff
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2699 inet_csk_schedule_ack(sk
);
2700 inet_csk(sk
)->icsk_ack
.ato
= TCP_ATO_MIN
;
2701 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
2702 TCP_DELACK_MAX
, TCP_RTO_MAX
);
2706 /* Reserve space for headers and prepare control bits. */
2707 skb_reserve(buff
, MAX_TCP_HEADER
);
2708 tcp_init_nondata_skb(buff
, tcp_acceptable_seq(sk
), TCPCB_FLAG_ACK
);
2710 /* Send it off, this clears delayed acks for us. */
2711 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2712 tcp_transmit_skb(sk
, buff
, 0, GFP_ATOMIC
);
2715 /* This routine sends a packet with an out of date sequence
2716 * number. It assumes the other end will try to ack it.
2718 * Question: what should we make while urgent mode?
2719 * 4.4BSD forces sending single byte of data. We cannot send
2720 * out of window data, because we have SND.NXT==SND.MAX...
2722 * Current solution: to send TWO zero-length segments in urgent mode:
2723 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2724 * out-of-date with SND.UNA-1 to probe window.
2726 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
)
2728 struct tcp_sock
*tp
= tcp_sk(sk
);
2729 struct sk_buff
*skb
;
2731 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2732 skb
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2736 /* Reserve space for headers and set control bits. */
2737 skb_reserve(skb
, MAX_TCP_HEADER
);
2738 /* Use a previous sequence. This should cause the other
2739 * end to send an ack. Don't queue or clone SKB, just
2742 tcp_init_nondata_skb(skb
, tp
->snd_una
- !urgent
, TCPCB_FLAG_ACK
);
2743 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2744 return tcp_transmit_skb(sk
, skb
, 0, GFP_ATOMIC
);
2747 /* Initiate keepalive or window probe from timer. */
2748 int tcp_write_wakeup(struct sock
*sk
)
2750 struct tcp_sock
*tp
= tcp_sk(sk
);
2751 struct sk_buff
*skb
;
2753 if (sk
->sk_state
== TCP_CLOSE
)
2756 if ((skb
= tcp_send_head(sk
)) != NULL
&&
2757 before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
))) {
2759 unsigned int mss
= tcp_current_mss(sk
);
2760 unsigned int seg_size
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
2762 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
2763 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
2765 /* We are probing the opening of a window
2766 * but the window size is != 0
2767 * must have been a result SWS avoidance ( sender )
2769 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
2771 seg_size
= min(seg_size
, mss
);
2772 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2773 if (tcp_fragment(sk
, skb
, seg_size
, mss
))
2775 } else if (!tcp_skb_pcount(skb
))
2776 tcp_set_skb_tso_segs(sk
, skb
, mss
);
2778 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2779 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2780 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2782 tcp_event_new_data_sent(sk
, skb
);
2785 if (between(tp
->snd_up
, tp
->snd_una
+ 1, tp
->snd_una
+ 0xFFFF))
2786 tcp_xmit_probe_skb(sk
, 1);
2787 return tcp_xmit_probe_skb(sk
, 0);
2791 /* A window probe timeout has occurred. If window is not closed send
2792 * a partial packet else a zero probe.
2794 void tcp_send_probe0(struct sock
*sk
)
2796 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2797 struct tcp_sock
*tp
= tcp_sk(sk
);
2800 err
= tcp_write_wakeup(sk
);
2802 if (tp
->packets_out
|| !tcp_send_head(sk
)) {
2803 /* Cancel probe timer, if it is not required. */
2804 icsk
->icsk_probes_out
= 0;
2805 icsk
->icsk_backoff
= 0;
2810 if (icsk
->icsk_backoff
< sysctl_tcp_retries2
)
2811 icsk
->icsk_backoff
++;
2812 icsk
->icsk_probes_out
++;
2813 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2814 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
2817 /* If packet was not sent due to local congestion,
2818 * do not backoff and do not remember icsk_probes_out.
2819 * Let local senders to fight for local resources.
2821 * Use accumulated backoff yet.
2823 if (!icsk
->icsk_probes_out
)
2824 icsk
->icsk_probes_out
= 1;
2825 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2826 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
,
2827 TCP_RESOURCE_PROBE_INTERVAL
),
2832 EXPORT_SYMBOL(tcp_select_initial_window
);
2833 EXPORT_SYMBOL(tcp_connect
);
2834 EXPORT_SYMBOL(tcp_make_synack
);
2835 EXPORT_SYMBOL(tcp_simple_retransmit
);
2836 EXPORT_SYMBOL(tcp_sync_mss
);
2837 EXPORT_SYMBOL(tcp_mtup_init
);