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
= TCP_BASE_MSS
;
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
;
123 unsigned int metric
= dst_metric_advmss(dst
);
134 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
135 * This is the first part of cwnd validation mechanism. */
136 static void tcp_cwnd_restart(struct sock
*sk
, struct dst_entry
*dst
)
138 struct tcp_sock
*tp
= tcp_sk(sk
);
139 s32 delta
= tcp_time_stamp
- tp
->lsndtime
;
140 u32 restart_cwnd
= tcp_init_cwnd(tp
, dst
);
141 u32 cwnd
= tp
->snd_cwnd
;
143 tcp_ca_event(sk
, CA_EVENT_CWND_RESTART
);
145 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
146 restart_cwnd
= min(restart_cwnd
, cwnd
);
148 while ((delta
-= inet_csk(sk
)->icsk_rto
) > 0 && cwnd
> restart_cwnd
)
150 tp
->snd_cwnd
= max(cwnd
, restart_cwnd
);
151 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
152 tp
->snd_cwnd_used
= 0;
155 /* Congestion state accounting after a packet has been sent. */
156 static void tcp_event_data_sent(struct tcp_sock
*tp
,
157 struct sk_buff
*skb
, struct sock
*sk
)
159 struct inet_connection_sock
*icsk
= inet_csk(sk
);
160 const u32 now
= tcp_time_stamp
;
162 if (sysctl_tcp_slow_start_after_idle
&&
163 (!tp
->packets_out
&& (s32
)(now
- tp
->lsndtime
) > icsk
->icsk_rto
))
164 tcp_cwnd_restart(sk
, __sk_dst_get(sk
));
168 /* If it is a reply for ato after last received
169 * packet, enter pingpong mode.
171 if ((u32
)(now
- icsk
->icsk_ack
.lrcvtime
) < icsk
->icsk_ack
.ato
)
172 icsk
->icsk_ack
.pingpong
= 1;
175 /* Account for an ACK we sent. */
176 static inline void tcp_event_ack_sent(struct sock
*sk
, unsigned int pkts
)
178 tcp_dec_quickack_mode(sk
, pkts
);
179 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_DACK
);
182 /* Determine a window scaling and initial window to offer.
183 * Based on the assumption that the given amount of space
184 * will be offered. Store the results in the tp structure.
185 * NOTE: for smooth operation initial space offering should
186 * be a multiple of mss if possible. We assume here that mss >= 1.
187 * This MUST be enforced by all callers.
189 void tcp_select_initial_window(int __space
, __u32 mss
,
190 __u32
*rcv_wnd
, __u32
*window_clamp
,
191 int wscale_ok
, __u8
*rcv_wscale
,
194 unsigned int space
= (__space
< 0 ? 0 : __space
);
196 /* If no clamp set the clamp to the max possible scaled window */
197 if (*window_clamp
== 0)
198 (*window_clamp
) = (65535 << 14);
199 space
= min(*window_clamp
, space
);
201 /* Quantize space offering to a multiple of mss if possible. */
203 space
= (space
/ mss
) * mss
;
205 /* NOTE: offering an initial window larger than 32767
206 * will break some buggy TCP stacks. If the admin tells us
207 * it is likely we could be speaking with such a buggy stack
208 * we will truncate our initial window offering to 32K-1
209 * unless the remote has sent us a window scaling option,
210 * which we interpret as a sign the remote TCP is not
211 * misinterpreting the window field as a signed quantity.
213 if (sysctl_tcp_workaround_signed_windows
)
214 (*rcv_wnd
) = min(space
, MAX_TCP_WINDOW
);
220 /* Set window scaling on max possible window
221 * See RFC1323 for an explanation of the limit to 14
223 space
= max_t(u32
, sysctl_tcp_rmem
[2], sysctl_rmem_max
);
224 space
= min_t(u32
, space
, *window_clamp
);
225 while (space
> 65535 && (*rcv_wscale
) < 14) {
231 /* Set initial window to value enough for senders, following RFC5681. */
232 if (mss
> (1 << *rcv_wscale
)) {
233 int init_cwnd
= rfc3390_bytes_to_packets(mss
);
235 /* when initializing use the value from init_rcv_wnd
236 * rather than the default from above
239 *rcv_wnd
= min(*rcv_wnd
, init_rcv_wnd
* mss
);
241 *rcv_wnd
= min(*rcv_wnd
, init_cwnd
* mss
);
244 /* Set the clamp no higher than max representable value */
245 (*window_clamp
) = min(65535U << (*rcv_wscale
), *window_clamp
);
247 EXPORT_SYMBOL(tcp_select_initial_window
);
249 /* Chose a new window to advertise, update state in tcp_sock for the
250 * socket, and return result with RFC1323 scaling applied. The return
251 * value can be stuffed directly into th->window for an outgoing
254 static u16
tcp_select_window(struct sock
*sk
)
256 struct tcp_sock
*tp
= tcp_sk(sk
);
257 u32 cur_win
= tcp_receive_window(tp
);
258 u32 new_win
= __tcp_select_window(sk
);
260 /* Never shrink the offered window */
261 if (new_win
< cur_win
) {
262 /* Danger Will Robinson!
263 * Don't update rcv_wup/rcv_wnd here or else
264 * we will not be able to advertise a zero
265 * window in time. --DaveM
267 * Relax Will Robinson.
269 new_win
= ALIGN(cur_win
, 1 << tp
->rx_opt
.rcv_wscale
);
271 tp
->rcv_wnd
= new_win
;
272 tp
->rcv_wup
= tp
->rcv_nxt
;
274 /* Make sure we do not exceed the maximum possible
277 if (!tp
->rx_opt
.rcv_wscale
&& sysctl_tcp_workaround_signed_windows
)
278 new_win
= min(new_win
, MAX_TCP_WINDOW
);
280 new_win
= min(new_win
, (65535U << tp
->rx_opt
.rcv_wscale
));
282 /* RFC1323 scaling applied */
283 new_win
>>= tp
->rx_opt
.rcv_wscale
;
285 /* If we advertise zero window, disable fast path. */
292 /* Packet ECN state for a SYN-ACK */
293 static inline void TCP_ECN_send_synack(struct tcp_sock
*tp
, struct sk_buff
*skb
)
295 TCP_SKB_CB(skb
)->flags
&= ~TCPHDR_CWR
;
296 if (!(tp
->ecn_flags
& TCP_ECN_OK
))
297 TCP_SKB_CB(skb
)->flags
&= ~TCPHDR_ECE
;
300 /* Packet ECN state for a SYN. */
301 static inline void TCP_ECN_send_syn(struct sock
*sk
, struct sk_buff
*skb
)
303 struct tcp_sock
*tp
= tcp_sk(sk
);
306 if (sysctl_tcp_ecn
== 1) {
307 TCP_SKB_CB(skb
)->flags
|= TCPHDR_ECE
| TCPHDR_CWR
;
308 tp
->ecn_flags
= TCP_ECN_OK
;
312 static __inline__
void
313 TCP_ECN_make_synack(struct request_sock
*req
, struct tcphdr
*th
)
315 if (inet_rsk(req
)->ecn_ok
)
319 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
322 static inline void TCP_ECN_send(struct sock
*sk
, struct sk_buff
*skb
,
325 struct tcp_sock
*tp
= tcp_sk(sk
);
327 if (tp
->ecn_flags
& TCP_ECN_OK
) {
328 /* Not-retransmitted data segment: set ECT and inject CWR. */
329 if (skb
->len
!= tcp_header_len
&&
330 !before(TCP_SKB_CB(skb
)->seq
, tp
->snd_nxt
)) {
332 if (tp
->ecn_flags
& TCP_ECN_QUEUE_CWR
) {
333 tp
->ecn_flags
&= ~TCP_ECN_QUEUE_CWR
;
334 tcp_hdr(skb
)->cwr
= 1;
335 skb_shinfo(skb
)->gso_type
|= SKB_GSO_TCP_ECN
;
338 /* ACK or retransmitted segment: clear ECT|CE */
339 INET_ECN_dontxmit(sk
);
341 if (tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)
342 tcp_hdr(skb
)->ece
= 1;
346 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
347 * auto increment end seqno.
349 static void tcp_init_nondata_skb(struct sk_buff
*skb
, u32 seq
, u8 flags
)
351 skb
->ip_summed
= CHECKSUM_PARTIAL
;
354 TCP_SKB_CB(skb
)->flags
= flags
;
355 TCP_SKB_CB(skb
)->sacked
= 0;
357 skb_shinfo(skb
)->gso_segs
= 1;
358 skb_shinfo(skb
)->gso_size
= 0;
359 skb_shinfo(skb
)->gso_type
= 0;
361 TCP_SKB_CB(skb
)->seq
= seq
;
362 if (flags
& (TCPHDR_SYN
| TCPHDR_FIN
))
364 TCP_SKB_CB(skb
)->end_seq
= seq
;
367 static inline int tcp_urg_mode(const struct tcp_sock
*tp
)
369 return tp
->snd_una
!= tp
->snd_up
;
372 #define OPTION_SACK_ADVERTISE (1 << 0)
373 #define OPTION_TS (1 << 1)
374 #define OPTION_MD5 (1 << 2)
375 #define OPTION_WSCALE (1 << 3)
376 #define OPTION_COOKIE_EXTENSION (1 << 4)
378 struct tcp_out_options
{
379 u8 options
; /* bit field of OPTION_* */
380 u8 ws
; /* window scale, 0 to disable */
381 u8 num_sack_blocks
; /* number of SACK blocks to include */
382 u8 hash_size
; /* bytes in hash_location */
383 u16 mss
; /* 0 to disable */
384 __u32 tsval
, tsecr
; /* need to include OPTION_TS */
385 __u8
*hash_location
; /* temporary pointer, overloaded */
388 /* The sysctl int routines are generic, so check consistency here.
390 static u8
tcp_cookie_size_check(u8 desired
)
395 /* previously specified */
398 cookie_size
= ACCESS_ONCE(sysctl_tcp_cookie_size
);
399 if (cookie_size
<= 0)
400 /* no default specified */
403 if (cookie_size
<= TCP_COOKIE_MIN
)
404 /* value too small, specify minimum */
405 return TCP_COOKIE_MIN
;
407 if (cookie_size
>= TCP_COOKIE_MAX
)
408 /* value too large, specify maximum */
409 return TCP_COOKIE_MAX
;
412 /* 8-bit multiple, illegal, fix it */
415 return (u8
)cookie_size
;
418 /* Write previously computed TCP options to the packet.
420 * Beware: Something in the Internet is very sensitive to the ordering of
421 * TCP options, we learned this through the hard way, so be careful here.
422 * Luckily we can at least blame others for their non-compliance but from
423 * inter-operatibility perspective it seems that we're somewhat stuck with
424 * the ordering which we have been using if we want to keep working with
425 * those broken things (not that it currently hurts anybody as there isn't
426 * particular reason why the ordering would need to be changed).
428 * At least SACK_PERM as the first option is known to lead to a disaster
429 * (but it may well be that other scenarios fail similarly).
431 static void tcp_options_write(__be32
*ptr
, struct tcp_sock
*tp
,
432 struct tcp_out_options
*opts
)
434 u8 options
= opts
->options
; /* mungable copy */
436 /* Having both authentication and cookies for security is redundant,
437 * and there's certainly not enough room. Instead, the cookie-less
438 * extension variant is proposed.
440 * Consider the pessimal case with authentication. The options
442 * COOKIE|MD5(20) + MSS(4) + SACK|TS(12) + WSCALE(4) == 40
444 if (unlikely(OPTION_MD5
& options
)) {
445 if (unlikely(OPTION_COOKIE_EXTENSION
& options
)) {
446 *ptr
++ = htonl((TCPOPT_COOKIE
<< 24) |
447 (TCPOLEN_COOKIE_BASE
<< 16) |
448 (TCPOPT_MD5SIG
<< 8) |
451 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
453 (TCPOPT_MD5SIG
<< 8) |
456 options
&= ~OPTION_COOKIE_EXTENSION
;
457 /* overload cookie hash location */
458 opts
->hash_location
= (__u8
*)ptr
;
462 if (unlikely(opts
->mss
)) {
463 *ptr
++ = htonl((TCPOPT_MSS
<< 24) |
464 (TCPOLEN_MSS
<< 16) |
468 if (likely(OPTION_TS
& options
)) {
469 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
470 *ptr
++ = htonl((TCPOPT_SACK_PERM
<< 24) |
471 (TCPOLEN_SACK_PERM
<< 16) |
472 (TCPOPT_TIMESTAMP
<< 8) |
474 options
&= ~OPTION_SACK_ADVERTISE
;
476 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
478 (TCPOPT_TIMESTAMP
<< 8) |
481 *ptr
++ = htonl(opts
->tsval
);
482 *ptr
++ = htonl(opts
->tsecr
);
485 /* Specification requires after timestamp, so do it now.
487 * Consider the pessimal case without authentication. The options
489 * MSS(4) + SACK|TS(12) + COOKIE(20) + WSCALE(4) == 40
491 if (unlikely(OPTION_COOKIE_EXTENSION
& options
)) {
492 __u8
*cookie_copy
= opts
->hash_location
;
493 u8 cookie_size
= opts
->hash_size
;
495 /* 8-bit multiple handled in tcp_cookie_size_check() above,
498 if (0x2 & cookie_size
) {
499 __u8
*p
= (__u8
*)ptr
;
501 /* 16-bit multiple */
502 *p
++ = TCPOPT_COOKIE
;
503 *p
++ = TCPOLEN_COOKIE_BASE
+ cookie_size
;
504 *p
++ = *cookie_copy
++;
505 *p
++ = *cookie_copy
++;
509 /* 32-bit multiple */
510 *ptr
++ = htonl(((TCPOPT_NOP
<< 24) |
512 (TCPOPT_COOKIE
<< 8) |
513 TCPOLEN_COOKIE_BASE
) +
517 if (cookie_size
> 0) {
518 memcpy(ptr
, cookie_copy
, cookie_size
);
519 ptr
+= (cookie_size
/ 4);
523 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
524 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
526 (TCPOPT_SACK_PERM
<< 8) |
530 if (unlikely(OPTION_WSCALE
& options
)) {
531 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
532 (TCPOPT_WINDOW
<< 16) |
533 (TCPOLEN_WINDOW
<< 8) |
537 if (unlikely(opts
->num_sack_blocks
)) {
538 struct tcp_sack_block
*sp
= tp
->rx_opt
.dsack
?
539 tp
->duplicate_sack
: tp
->selective_acks
;
542 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
545 (TCPOLEN_SACK_BASE
+ (opts
->num_sack_blocks
*
546 TCPOLEN_SACK_PERBLOCK
)));
548 for (this_sack
= 0; this_sack
< opts
->num_sack_blocks
;
550 *ptr
++ = htonl(sp
[this_sack
].start_seq
);
551 *ptr
++ = htonl(sp
[this_sack
].end_seq
);
554 tp
->rx_opt
.dsack
= 0;
558 /* Compute TCP options for SYN packets. This is not the final
559 * network wire format yet.
561 static unsigned tcp_syn_options(struct sock
*sk
, struct sk_buff
*skb
,
562 struct tcp_out_options
*opts
,
563 struct tcp_md5sig_key
**md5
) {
564 struct tcp_sock
*tp
= tcp_sk(sk
);
565 struct tcp_cookie_values
*cvp
= tp
->cookie_values
;
566 unsigned remaining
= MAX_TCP_OPTION_SPACE
;
567 u8 cookie_size
= (!tp
->rx_opt
.cookie_out_never
&& cvp
!= NULL
) ?
568 tcp_cookie_size_check(cvp
->cookie_desired
) :
571 #ifdef CONFIG_TCP_MD5SIG
572 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
574 opts
->options
|= OPTION_MD5
;
575 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
581 /* We always get an MSS option. The option bytes which will be seen in
582 * normal data packets should timestamps be used, must be in the MSS
583 * advertised. But we subtract them from tp->mss_cache so that
584 * calculations in tcp_sendmsg are simpler etc. So account for this
585 * fact here if necessary. If we don't do this correctly, as a
586 * receiver we won't recognize data packets as being full sized when we
587 * should, and thus we won't abide by the delayed ACK rules correctly.
588 * SACKs don't matter, we never delay an ACK when we have any of those
590 opts
->mss
= tcp_advertise_mss(sk
);
591 remaining
-= TCPOLEN_MSS_ALIGNED
;
593 if (likely(sysctl_tcp_timestamps
&& *md5
== NULL
)) {
594 opts
->options
|= OPTION_TS
;
595 opts
->tsval
= TCP_SKB_CB(skb
)->when
;
596 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
597 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
599 if (likely(sysctl_tcp_window_scaling
)) {
600 opts
->ws
= tp
->rx_opt
.rcv_wscale
;
601 opts
->options
|= OPTION_WSCALE
;
602 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
604 if (likely(sysctl_tcp_sack
)) {
605 opts
->options
|= OPTION_SACK_ADVERTISE
;
606 if (unlikely(!(OPTION_TS
& opts
->options
)))
607 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
610 /* Note that timestamps are required by the specification.
612 * Odd numbers of bytes are prohibited by the specification, ensuring
613 * that the cookie is 16-bit aligned, and the resulting cookie pair is
617 (OPTION_TS
& opts
->options
) &&
619 int need
= TCPOLEN_COOKIE_BASE
+ cookie_size
;
622 /* 32-bit multiple */
623 need
+= 2; /* NOPs */
625 if (need
> remaining
) {
626 /* try shrinking cookie to fit */
631 while (need
> remaining
&& TCP_COOKIE_MIN
<= cookie_size
) {
635 if (TCP_COOKIE_MIN
<= cookie_size
) {
636 opts
->options
|= OPTION_COOKIE_EXTENSION
;
637 opts
->hash_location
= (__u8
*)&cvp
->cookie_pair
[0];
638 opts
->hash_size
= cookie_size
;
640 /* Remember for future incarnations. */
641 cvp
->cookie_desired
= cookie_size
;
643 if (cvp
->cookie_desired
!= cvp
->cookie_pair_size
) {
644 /* Currently use random bytes as a nonce,
645 * assuming these are completely unpredictable
646 * by hostile users of the same system.
648 get_random_bytes(&cvp
->cookie_pair
[0],
650 cvp
->cookie_pair_size
= cookie_size
;
656 return MAX_TCP_OPTION_SPACE
- remaining
;
659 /* Set up TCP options for SYN-ACKs. */
660 static unsigned tcp_synack_options(struct sock
*sk
,
661 struct request_sock
*req
,
662 unsigned mss
, struct sk_buff
*skb
,
663 struct tcp_out_options
*opts
,
664 struct tcp_md5sig_key
**md5
,
665 struct tcp_extend_values
*xvp
)
667 struct inet_request_sock
*ireq
= inet_rsk(req
);
668 unsigned remaining
= MAX_TCP_OPTION_SPACE
;
669 u8 cookie_plus
= (xvp
!= NULL
&& !xvp
->cookie_out_never
) ?
673 #ifdef CONFIG_TCP_MD5SIG
674 *md5
= tcp_rsk(req
)->af_specific
->md5_lookup(sk
, req
);
676 opts
->options
|= OPTION_MD5
;
677 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
679 /* We can't fit any SACK blocks in a packet with MD5 + TS
680 * options. There was discussion about disabling SACK
681 * rather than TS in order to fit in better with old,
682 * buggy kernels, but that was deemed to be unnecessary.
684 ireq
->tstamp_ok
&= !ireq
->sack_ok
;
690 /* We always send an MSS option. */
692 remaining
-= TCPOLEN_MSS_ALIGNED
;
694 if (likely(ireq
->wscale_ok
)) {
695 opts
->ws
= ireq
->rcv_wscale
;
696 opts
->options
|= OPTION_WSCALE
;
697 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
699 if (likely(ireq
->tstamp_ok
)) {
700 opts
->options
|= OPTION_TS
;
701 opts
->tsval
= TCP_SKB_CB(skb
)->when
;
702 opts
->tsecr
= req
->ts_recent
;
703 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
705 if (likely(ireq
->sack_ok
)) {
706 opts
->options
|= OPTION_SACK_ADVERTISE
;
707 if (unlikely(!ireq
->tstamp_ok
))
708 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
711 /* Similar rationale to tcp_syn_options() applies here, too.
712 * If the <SYN> options fit, the same options should fit now!
716 cookie_plus
> TCPOLEN_COOKIE_BASE
) {
717 int need
= cookie_plus
; /* has TCPOLEN_COOKIE_BASE */
720 /* 32-bit multiple */
721 need
+= 2; /* NOPs */
723 if (need
<= remaining
) {
724 opts
->options
|= OPTION_COOKIE_EXTENSION
;
725 opts
->hash_size
= cookie_plus
- TCPOLEN_COOKIE_BASE
;
728 /* There's no error return, so flag it. */
729 xvp
->cookie_out_never
= 1; /* true */
733 return MAX_TCP_OPTION_SPACE
- remaining
;
736 /* Compute TCP options for ESTABLISHED sockets. This is not the
737 * final wire format yet.
739 static unsigned tcp_established_options(struct sock
*sk
, struct sk_buff
*skb
,
740 struct tcp_out_options
*opts
,
741 struct tcp_md5sig_key
**md5
) {
742 struct tcp_skb_cb
*tcb
= skb
? TCP_SKB_CB(skb
) : NULL
;
743 struct tcp_sock
*tp
= tcp_sk(sk
);
745 unsigned int eff_sacks
;
747 #ifdef CONFIG_TCP_MD5SIG
748 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
749 if (unlikely(*md5
)) {
750 opts
->options
|= OPTION_MD5
;
751 size
+= TCPOLEN_MD5SIG_ALIGNED
;
757 if (likely(tp
->rx_opt
.tstamp_ok
)) {
758 opts
->options
|= OPTION_TS
;
759 opts
->tsval
= tcb
? tcb
->when
: 0;
760 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
761 size
+= TCPOLEN_TSTAMP_ALIGNED
;
764 eff_sacks
= tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
;
765 if (unlikely(eff_sacks
)) {
766 const unsigned remaining
= MAX_TCP_OPTION_SPACE
- size
;
767 opts
->num_sack_blocks
=
768 min_t(unsigned, eff_sacks
,
769 (remaining
- TCPOLEN_SACK_BASE_ALIGNED
) /
770 TCPOLEN_SACK_PERBLOCK
);
771 size
+= TCPOLEN_SACK_BASE_ALIGNED
+
772 opts
->num_sack_blocks
* TCPOLEN_SACK_PERBLOCK
;
778 /* This routine actually transmits TCP packets queued in by
779 * tcp_do_sendmsg(). This is used by both the initial
780 * transmission and possible later retransmissions.
781 * All SKB's seen here are completely headerless. It is our
782 * job to build the TCP header, and pass the packet down to
783 * IP so it can do the same plus pass the packet off to the
786 * We are working here with either a clone of the original
787 * SKB, or a fresh unique copy made by the retransmit engine.
789 static int tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int clone_it
,
792 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
793 struct inet_sock
*inet
;
795 struct tcp_skb_cb
*tcb
;
796 struct tcp_out_options opts
;
797 unsigned tcp_options_size
, tcp_header_size
;
798 struct tcp_md5sig_key
*md5
;
802 BUG_ON(!skb
|| !tcp_skb_pcount(skb
));
804 /* If congestion control is doing timestamping, we must
805 * take such a timestamp before we potentially clone/copy.
807 if (icsk
->icsk_ca_ops
->flags
& TCP_CONG_RTT_STAMP
)
808 __net_timestamp(skb
);
810 if (likely(clone_it
)) {
811 if (unlikely(skb_cloned(skb
)))
812 skb
= pskb_copy(skb
, gfp_mask
);
814 skb
= skb_clone(skb
, gfp_mask
);
821 tcb
= TCP_SKB_CB(skb
);
822 memset(&opts
, 0, sizeof(opts
));
824 if (unlikely(tcb
->flags
& TCPHDR_SYN
))
825 tcp_options_size
= tcp_syn_options(sk
, skb
, &opts
, &md5
);
827 tcp_options_size
= tcp_established_options(sk
, skb
, &opts
,
829 tcp_header_size
= tcp_options_size
+ sizeof(struct tcphdr
);
831 if (tcp_packets_in_flight(tp
) == 0) {
832 tcp_ca_event(sk
, CA_EVENT_TX_START
);
837 skb_push(skb
, tcp_header_size
);
838 skb_reset_transport_header(skb
);
839 skb_set_owner_w(skb
, sk
);
841 /* Build TCP header and checksum it. */
843 th
->source
= inet
->inet_sport
;
844 th
->dest
= inet
->inet_dport
;
845 th
->seq
= htonl(tcb
->seq
);
846 th
->ack_seq
= htonl(tp
->rcv_nxt
);
847 *(((__be16
*)th
) + 6) = htons(((tcp_header_size
>> 2) << 12) |
850 if (unlikely(tcb
->flags
& TCPHDR_SYN
)) {
851 /* RFC1323: The window in SYN & SYN/ACK segments
854 th
->window
= htons(min(tp
->rcv_wnd
, 65535U));
856 th
->window
= htons(tcp_select_window(sk
));
861 /* The urg_mode check is necessary during a below snd_una win probe */
862 if (unlikely(tcp_urg_mode(tp
) && before(tcb
->seq
, tp
->snd_up
))) {
863 if (before(tp
->snd_up
, tcb
->seq
+ 0x10000)) {
864 th
->urg_ptr
= htons(tp
->snd_up
- tcb
->seq
);
866 } else if (after(tcb
->seq
+ 0xFFFF, tp
->snd_nxt
)) {
867 th
->urg_ptr
= htons(0xFFFF);
872 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
);
873 if (likely((tcb
->flags
& TCPHDR_SYN
) == 0))
874 TCP_ECN_send(sk
, skb
, tcp_header_size
);
876 #ifdef CONFIG_TCP_MD5SIG
877 /* Calculate the MD5 hash, as we have all we need now */
879 sk_nocaps_add(sk
, NETIF_F_GSO_MASK
);
880 tp
->af_specific
->calc_md5_hash(opts
.hash_location
,
885 icsk
->icsk_af_ops
->send_check(sk
, skb
);
887 if (likely(tcb
->flags
& TCPHDR_ACK
))
888 tcp_event_ack_sent(sk
, tcp_skb_pcount(skb
));
890 if (skb
->len
!= tcp_header_size
)
891 tcp_event_data_sent(tp
, skb
, sk
);
893 if (after(tcb
->end_seq
, tp
->snd_nxt
) || tcb
->seq
== tcb
->end_seq
)
894 TCP_ADD_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
,
895 tcp_skb_pcount(skb
));
897 err
= icsk
->icsk_af_ops
->queue_xmit(skb
);
898 if (likely(err
<= 0))
901 tcp_enter_cwr(sk
, 1);
903 return net_xmit_eval(err
);
906 /* This routine just queues the buffer for sending.
908 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
909 * otherwise socket can stall.
911 static void tcp_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
913 struct tcp_sock
*tp
= tcp_sk(sk
);
915 /* Advance write_seq and place onto the write_queue. */
916 tp
->write_seq
= TCP_SKB_CB(skb
)->end_seq
;
917 skb_header_release(skb
);
918 tcp_add_write_queue_tail(sk
, skb
);
919 sk
->sk_wmem_queued
+= skb
->truesize
;
920 sk_mem_charge(sk
, skb
->truesize
);
923 /* Initialize TSO segments for a packet. */
924 static void tcp_set_skb_tso_segs(struct sock
*sk
, struct sk_buff
*skb
,
925 unsigned int mss_now
)
927 if (skb
->len
<= mss_now
|| !sk_can_gso(sk
) ||
928 skb
->ip_summed
== CHECKSUM_NONE
) {
929 /* Avoid the costly divide in the normal
932 skb_shinfo(skb
)->gso_segs
= 1;
933 skb_shinfo(skb
)->gso_size
= 0;
934 skb_shinfo(skb
)->gso_type
= 0;
936 skb_shinfo(skb
)->gso_segs
= DIV_ROUND_UP(skb
->len
, mss_now
);
937 skb_shinfo(skb
)->gso_size
= mss_now
;
938 skb_shinfo(skb
)->gso_type
= sk
->sk_gso_type
;
942 /* When a modification to fackets out becomes necessary, we need to check
943 * skb is counted to fackets_out or not.
945 static void tcp_adjust_fackets_out(struct sock
*sk
, struct sk_buff
*skb
,
948 struct tcp_sock
*tp
= tcp_sk(sk
);
950 if (!tp
->sacked_out
|| tcp_is_reno(tp
))
953 if (after(tcp_highest_sack_seq(tp
), TCP_SKB_CB(skb
)->seq
))
954 tp
->fackets_out
-= decr
;
957 /* Pcount in the middle of the write queue got changed, we need to do various
958 * tweaks to fix counters
960 static void tcp_adjust_pcount(struct sock
*sk
, struct sk_buff
*skb
, int decr
)
962 struct tcp_sock
*tp
= tcp_sk(sk
);
964 tp
->packets_out
-= decr
;
966 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
967 tp
->sacked_out
-= decr
;
968 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
969 tp
->retrans_out
-= decr
;
970 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
)
971 tp
->lost_out
-= decr
;
973 /* Reno case is special. Sigh... */
974 if (tcp_is_reno(tp
) && decr
> 0)
975 tp
->sacked_out
-= min_t(u32
, tp
->sacked_out
, decr
);
977 tcp_adjust_fackets_out(sk
, skb
, decr
);
979 if (tp
->lost_skb_hint
&&
980 before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
) &&
981 (tcp_is_fack(tp
) || (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)))
982 tp
->lost_cnt_hint
-= decr
;
984 tcp_verify_left_out(tp
);
987 /* Function to create two new TCP segments. Shrinks the given segment
988 * to the specified size and appends a new segment with the rest of the
989 * packet to the list. This won't be called frequently, I hope.
990 * Remember, these are still headerless SKBs at this point.
992 int tcp_fragment(struct sock
*sk
, struct sk_buff
*skb
, u32 len
,
993 unsigned int mss_now
)
995 struct tcp_sock
*tp
= tcp_sk(sk
);
996 struct sk_buff
*buff
;
997 int nsize
, old_factor
;
1001 BUG_ON(len
> skb
->len
);
1003 nsize
= skb_headlen(skb
) - len
;
1007 if (skb_cloned(skb
) &&
1008 skb_is_nonlinear(skb
) &&
1009 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
1012 /* Get a new skb... force flag on. */
1013 buff
= sk_stream_alloc_skb(sk
, nsize
, GFP_ATOMIC
);
1015 return -ENOMEM
; /* We'll just try again later. */
1017 sk
->sk_wmem_queued
+= buff
->truesize
;
1018 sk_mem_charge(sk
, buff
->truesize
);
1019 nlen
= skb
->len
- len
- nsize
;
1020 buff
->truesize
+= nlen
;
1021 skb
->truesize
-= nlen
;
1023 /* Correct the sequence numbers. */
1024 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1025 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1026 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1028 /* PSH and FIN should only be set in the second packet. */
1029 flags
= TCP_SKB_CB(skb
)->flags
;
1030 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPHDR_FIN
| TCPHDR_PSH
);
1031 TCP_SKB_CB(buff
)->flags
= flags
;
1032 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
1034 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_PARTIAL
) {
1035 /* Copy and checksum data tail into the new buffer. */
1036 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
,
1037 skb_put(buff
, nsize
),
1042 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
1044 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1045 skb_split(skb
, buff
, len
);
1048 buff
->ip_summed
= skb
->ip_summed
;
1050 /* Looks stupid, but our code really uses when of
1051 * skbs, which it never sent before. --ANK
1053 TCP_SKB_CB(buff
)->when
= TCP_SKB_CB(skb
)->when
;
1054 buff
->tstamp
= skb
->tstamp
;
1056 old_factor
= tcp_skb_pcount(skb
);
1058 /* Fix up tso_factor for both original and new SKB. */
1059 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1060 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1062 /* If this packet has been sent out already, we must
1063 * adjust the various packet counters.
1065 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
1066 int diff
= old_factor
- tcp_skb_pcount(skb
) -
1067 tcp_skb_pcount(buff
);
1070 tcp_adjust_pcount(sk
, skb
, diff
);
1073 /* Link BUFF into the send queue. */
1074 skb_header_release(buff
);
1075 tcp_insert_write_queue_after(skb
, buff
, sk
);
1080 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1081 * eventually). The difference is that pulled data not copied, but
1082 * immediately discarded.
1084 static void __pskb_trim_head(struct sk_buff
*skb
, int len
)
1090 for (i
= 0; i
< skb_shinfo(skb
)->nr_frags
; i
++) {
1091 if (skb_shinfo(skb
)->frags
[i
].size
<= eat
) {
1092 put_page(skb_shinfo(skb
)->frags
[i
].page
);
1093 eat
-= skb_shinfo(skb
)->frags
[i
].size
;
1095 skb_shinfo(skb
)->frags
[k
] = skb_shinfo(skb
)->frags
[i
];
1097 skb_shinfo(skb
)->frags
[k
].page_offset
+= eat
;
1098 skb_shinfo(skb
)->frags
[k
].size
-= eat
;
1104 skb_shinfo(skb
)->nr_frags
= k
;
1106 skb_reset_tail_pointer(skb
);
1107 skb
->data_len
-= len
;
1108 skb
->len
= skb
->data_len
;
1111 /* Remove acked data from a packet in the transmit queue. */
1112 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
1114 if (skb_cloned(skb
) && pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
1117 /* If len == headlen, we avoid __skb_pull to preserve alignment. */
1118 if (unlikely(len
< skb_headlen(skb
)))
1119 __skb_pull(skb
, len
);
1121 __pskb_trim_head(skb
, len
- skb_headlen(skb
));
1123 TCP_SKB_CB(skb
)->seq
+= len
;
1124 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1126 skb
->truesize
-= len
;
1127 sk
->sk_wmem_queued
-= len
;
1128 sk_mem_uncharge(sk
, len
);
1129 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
1131 /* Any change of skb->len requires recalculation of tso
1134 if (tcp_skb_pcount(skb
) > 1)
1135 tcp_set_skb_tso_segs(sk
, skb
, tcp_current_mss(sk
));
1140 /* Calculate MSS. Not accounting for SACKs here. */
1141 int tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1143 struct tcp_sock
*tp
= tcp_sk(sk
);
1144 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1147 /* Calculate base mss without TCP options:
1148 It is MMS_S - sizeof(tcphdr) of rfc1122
1150 mss_now
= pmtu
- icsk
->icsk_af_ops
->net_header_len
- sizeof(struct tcphdr
);
1152 /* Clamp it (mss_clamp does not include tcp options) */
1153 if (mss_now
> tp
->rx_opt
.mss_clamp
)
1154 mss_now
= tp
->rx_opt
.mss_clamp
;
1156 /* Now subtract optional transport overhead */
1157 mss_now
-= icsk
->icsk_ext_hdr_len
;
1159 /* Then reserve room for full set of TCP options and 8 bytes of data */
1163 /* Now subtract TCP options size, not including SACKs */
1164 mss_now
-= tp
->tcp_header_len
- sizeof(struct tcphdr
);
1169 /* Inverse of above */
1170 int tcp_mss_to_mtu(struct sock
*sk
, int mss
)
1172 struct tcp_sock
*tp
= tcp_sk(sk
);
1173 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1177 tp
->tcp_header_len
+
1178 icsk
->icsk_ext_hdr_len
+
1179 icsk
->icsk_af_ops
->net_header_len
;
1184 /* MTU probing init per socket */
1185 void tcp_mtup_init(struct sock
*sk
)
1187 struct tcp_sock
*tp
= tcp_sk(sk
);
1188 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1190 icsk
->icsk_mtup
.enabled
= sysctl_tcp_mtu_probing
> 1;
1191 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+ sizeof(struct tcphdr
) +
1192 icsk
->icsk_af_ops
->net_header_len
;
1193 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, sysctl_tcp_base_mss
);
1194 icsk
->icsk_mtup
.probe_size
= 0;
1196 EXPORT_SYMBOL(tcp_mtup_init
);
1198 /* This function synchronize snd mss to current pmtu/exthdr set.
1200 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1201 for TCP options, but includes only bare TCP header.
1203 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1204 It is minimum of user_mss and mss received with SYN.
1205 It also does not include TCP options.
1207 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1209 tp->mss_cache is current effective sending mss, including
1210 all tcp options except for SACKs. It is evaluated,
1211 taking into account current pmtu, but never exceeds
1212 tp->rx_opt.mss_clamp.
1214 NOTE1. rfc1122 clearly states that advertised MSS
1215 DOES NOT include either tcp or ip options.
1217 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1218 are READ ONLY outside this function. --ANK (980731)
1220 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
1222 struct tcp_sock
*tp
= tcp_sk(sk
);
1223 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1226 if (icsk
->icsk_mtup
.search_high
> pmtu
)
1227 icsk
->icsk_mtup
.search_high
= pmtu
;
1229 mss_now
= tcp_mtu_to_mss(sk
, pmtu
);
1230 mss_now
= tcp_bound_to_half_wnd(tp
, mss_now
);
1232 /* And store cached results */
1233 icsk
->icsk_pmtu_cookie
= pmtu
;
1234 if (icsk
->icsk_mtup
.enabled
)
1235 mss_now
= min(mss_now
, tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
));
1236 tp
->mss_cache
= mss_now
;
1240 EXPORT_SYMBOL(tcp_sync_mss
);
1242 /* Compute the current effective MSS, taking SACKs and IP options,
1243 * and even PMTU discovery events into account.
1245 unsigned int tcp_current_mss(struct sock
*sk
)
1247 struct tcp_sock
*tp
= tcp_sk(sk
);
1248 struct dst_entry
*dst
= __sk_dst_get(sk
);
1250 unsigned header_len
;
1251 struct tcp_out_options opts
;
1252 struct tcp_md5sig_key
*md5
;
1254 mss_now
= tp
->mss_cache
;
1257 u32 mtu
= dst_mtu(dst
);
1258 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
1259 mss_now
= tcp_sync_mss(sk
, mtu
);
1262 header_len
= tcp_established_options(sk
, NULL
, &opts
, &md5
) +
1263 sizeof(struct tcphdr
);
1264 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1265 * some common options. If this is an odd packet (because we have SACK
1266 * blocks etc) then our calculated header_len will be different, and
1267 * we have to adjust mss_now correspondingly */
1268 if (header_len
!= tp
->tcp_header_len
) {
1269 int delta
= (int) header_len
- tp
->tcp_header_len
;
1276 /* Congestion window validation. (RFC2861) */
1277 static void tcp_cwnd_validate(struct sock
*sk
)
1279 struct tcp_sock
*tp
= tcp_sk(sk
);
1281 if (tp
->packets_out
>= tp
->snd_cwnd
) {
1282 /* Network is feed fully. */
1283 tp
->snd_cwnd_used
= 0;
1284 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1286 /* Network starves. */
1287 if (tp
->packets_out
> tp
->snd_cwnd_used
)
1288 tp
->snd_cwnd_used
= tp
->packets_out
;
1290 if (sysctl_tcp_slow_start_after_idle
&&
1291 (s32
)(tcp_time_stamp
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
)
1292 tcp_cwnd_application_limited(sk
);
1296 /* Returns the portion of skb which can be sent right away without
1297 * introducing MSS oddities to segment boundaries. In rare cases where
1298 * mss_now != mss_cache, we will request caller to create a small skb
1299 * per input skb which could be mostly avoided here (if desired).
1301 * We explicitly want to create a request for splitting write queue tail
1302 * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1303 * thus all the complexity (cwnd_len is always MSS multiple which we
1304 * return whenever allowed by the other factors). Basically we need the
1305 * modulo only when the receiver window alone is the limiting factor or
1306 * when we would be allowed to send the split-due-to-Nagle skb fully.
1308 static unsigned int tcp_mss_split_point(struct sock
*sk
, struct sk_buff
*skb
,
1309 unsigned int mss_now
, unsigned int cwnd
)
1311 struct tcp_sock
*tp
= tcp_sk(sk
);
1312 u32 needed
, window
, cwnd_len
;
1314 window
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1315 cwnd_len
= mss_now
* cwnd
;
1317 if (likely(cwnd_len
<= window
&& skb
!= tcp_write_queue_tail(sk
)))
1320 needed
= min(skb
->len
, window
);
1322 if (cwnd_len
<= needed
)
1325 return needed
- needed
% mss_now
;
1328 /* Can at least one segment of SKB be sent right now, according to the
1329 * congestion window rules? If so, return how many segments are allowed.
1331 static inline unsigned int tcp_cwnd_test(struct tcp_sock
*tp
,
1332 struct sk_buff
*skb
)
1334 u32 in_flight
, cwnd
;
1336 /* Don't be strict about the congestion window for the final FIN. */
1337 if ((TCP_SKB_CB(skb
)->flags
& TCPHDR_FIN
) && tcp_skb_pcount(skb
) == 1)
1340 in_flight
= tcp_packets_in_flight(tp
);
1341 cwnd
= tp
->snd_cwnd
;
1342 if (in_flight
< cwnd
)
1343 return (cwnd
- in_flight
);
1348 /* Intialize TSO state of a skb.
1349 * This must be invoked the first time we consider transmitting
1350 * SKB onto the wire.
1352 static int tcp_init_tso_segs(struct sock
*sk
, struct sk_buff
*skb
,
1353 unsigned int mss_now
)
1355 int tso_segs
= tcp_skb_pcount(skb
);
1357 if (!tso_segs
|| (tso_segs
> 1 && tcp_skb_mss(skb
) != mss_now
)) {
1358 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1359 tso_segs
= tcp_skb_pcount(skb
);
1364 /* Minshall's variant of the Nagle send check. */
1365 static inline int tcp_minshall_check(const struct tcp_sock
*tp
)
1367 return after(tp
->snd_sml
, tp
->snd_una
) &&
1368 !after(tp
->snd_sml
, tp
->snd_nxt
);
1371 /* Return 0, if packet can be sent now without violation Nagle's rules:
1372 * 1. It is full sized.
1373 * 2. Or it contains FIN. (already checked by caller)
1374 * 3. Or TCP_NODELAY was set.
1375 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1376 * With Minshall's modification: all sent small packets are ACKed.
1378 static inline int tcp_nagle_check(const struct tcp_sock
*tp
,
1379 const struct sk_buff
*skb
,
1380 unsigned mss_now
, int nonagle
)
1382 return skb
->len
< mss_now
&&
1383 ((nonagle
& TCP_NAGLE_CORK
) ||
1384 (!nonagle
&& tp
->packets_out
&& tcp_minshall_check(tp
)));
1387 /* Return non-zero if the Nagle test allows this packet to be
1390 static inline int tcp_nagle_test(struct tcp_sock
*tp
, struct sk_buff
*skb
,
1391 unsigned int cur_mss
, int nonagle
)
1393 /* Nagle rule does not apply to frames, which sit in the middle of the
1394 * write_queue (they have no chances to get new data).
1396 * This is implemented in the callers, where they modify the 'nonagle'
1397 * argument based upon the location of SKB in the send queue.
1399 if (nonagle
& TCP_NAGLE_PUSH
)
1402 /* Don't use the nagle rule for urgent data (or for the final FIN).
1403 * Nagle can be ignored during F-RTO too (see RFC4138).
1405 if (tcp_urg_mode(tp
) || (tp
->frto_counter
== 2) ||
1406 (TCP_SKB_CB(skb
)->flags
& TCPHDR_FIN
))
1409 if (!tcp_nagle_check(tp
, skb
, cur_mss
, nonagle
))
1415 /* Does at least the first segment of SKB fit into the send window? */
1416 static inline int tcp_snd_wnd_test(struct tcp_sock
*tp
, struct sk_buff
*skb
,
1417 unsigned int cur_mss
)
1419 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1421 if (skb
->len
> cur_mss
)
1422 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
1424 return !after(end_seq
, tcp_wnd_end(tp
));
1427 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1428 * should be put on the wire right now. If so, it returns the number of
1429 * packets allowed by the congestion window.
1431 static unsigned int tcp_snd_test(struct sock
*sk
, struct sk_buff
*skb
,
1432 unsigned int cur_mss
, int nonagle
)
1434 struct tcp_sock
*tp
= tcp_sk(sk
);
1435 unsigned int cwnd_quota
;
1437 tcp_init_tso_segs(sk
, skb
, cur_mss
);
1439 if (!tcp_nagle_test(tp
, skb
, cur_mss
, nonagle
))
1442 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1443 if (cwnd_quota
&& !tcp_snd_wnd_test(tp
, skb
, cur_mss
))
1449 /* Test if sending is allowed right now. */
1450 int tcp_may_send_now(struct sock
*sk
)
1452 struct tcp_sock
*tp
= tcp_sk(sk
);
1453 struct sk_buff
*skb
= tcp_send_head(sk
);
1456 tcp_snd_test(sk
, skb
, tcp_current_mss(sk
),
1457 (tcp_skb_is_last(sk
, skb
) ?
1458 tp
->nonagle
: TCP_NAGLE_PUSH
));
1461 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1462 * which is put after SKB on the list. It is very much like
1463 * tcp_fragment() except that it may make several kinds of assumptions
1464 * in order to speed up the splitting operation. In particular, we
1465 * know that all the data is in scatter-gather pages, and that the
1466 * packet has never been sent out before (and thus is not cloned).
1468 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
,
1469 unsigned int mss_now
, gfp_t gfp
)
1471 struct sk_buff
*buff
;
1472 int nlen
= skb
->len
- len
;
1475 /* All of a TSO frame must be composed of paged data. */
1476 if (skb
->len
!= skb
->data_len
)
1477 return tcp_fragment(sk
, skb
, len
, mss_now
);
1479 buff
= sk_stream_alloc_skb(sk
, 0, gfp
);
1480 if (unlikely(buff
== NULL
))
1483 sk
->sk_wmem_queued
+= buff
->truesize
;
1484 sk_mem_charge(sk
, buff
->truesize
);
1485 buff
->truesize
+= nlen
;
1486 skb
->truesize
-= nlen
;
1488 /* Correct the sequence numbers. */
1489 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1490 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1491 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1493 /* PSH and FIN should only be set in the second packet. */
1494 flags
= TCP_SKB_CB(skb
)->flags
;
1495 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPHDR_FIN
| TCPHDR_PSH
);
1496 TCP_SKB_CB(buff
)->flags
= flags
;
1498 /* This packet was never sent out yet, so no SACK bits. */
1499 TCP_SKB_CB(buff
)->sacked
= 0;
1501 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_PARTIAL
;
1502 skb_split(skb
, buff
, len
);
1504 /* Fix up tso_factor for both original and new SKB. */
1505 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1506 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1508 /* Link BUFF into the send queue. */
1509 skb_header_release(buff
);
1510 tcp_insert_write_queue_after(skb
, buff
, sk
);
1515 /* Try to defer sending, if possible, in order to minimize the amount
1516 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1518 * This algorithm is from John Heffner.
1520 static int tcp_tso_should_defer(struct sock
*sk
, struct sk_buff
*skb
)
1522 struct tcp_sock
*tp
= tcp_sk(sk
);
1523 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1524 u32 send_win
, cong_win
, limit
, in_flight
;
1527 if (TCP_SKB_CB(skb
)->flags
& TCPHDR_FIN
)
1530 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
1533 /* Defer for less than two clock ticks. */
1534 if (tp
->tso_deferred
&&
1535 (((u32
)jiffies
<< 1) >> 1) - (tp
->tso_deferred
>> 1) > 1)
1538 in_flight
= tcp_packets_in_flight(tp
);
1540 BUG_ON(tcp_skb_pcount(skb
) <= 1 || (tp
->snd_cwnd
<= in_flight
));
1542 send_win
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1544 /* From in_flight test above, we know that cwnd > in_flight. */
1545 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
1547 limit
= min(send_win
, cong_win
);
1549 /* If a full-sized TSO skb can be sent, do it. */
1550 if (limit
>= sk
->sk_gso_max_size
)
1553 /* Middle in queue won't get any more data, full sendable already? */
1554 if ((skb
!= tcp_write_queue_tail(sk
)) && (limit
>= skb
->len
))
1557 win_divisor
= ACCESS_ONCE(sysctl_tcp_tso_win_divisor
);
1559 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
1561 /* If at least some fraction of a window is available,
1564 chunk
/= win_divisor
;
1568 /* Different approach, try not to defer past a single
1569 * ACK. Receiver should ACK every other full sized
1570 * frame, so if we have space for more than 3 frames
1573 if (limit
> tcp_max_burst(tp
) * tp
->mss_cache
)
1577 /* Ok, it looks like it is advisable to defer. */
1578 tp
->tso_deferred
= 1 | (jiffies
<< 1);
1583 tp
->tso_deferred
= 0;
1587 /* Create a new MTU probe if we are ready.
1588 * MTU probe is regularly attempting to increase the path MTU by
1589 * deliberately sending larger packets. This discovers routing
1590 * changes resulting in larger path MTUs.
1592 * Returns 0 if we should wait to probe (no cwnd available),
1593 * 1 if a probe was sent,
1596 static int tcp_mtu_probe(struct sock
*sk
)
1598 struct tcp_sock
*tp
= tcp_sk(sk
);
1599 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1600 struct sk_buff
*skb
, *nskb
, *next
;
1607 /* Not currently probing/verifying,
1609 * have enough cwnd, and
1610 * not SACKing (the variable headers throw things off) */
1611 if (!icsk
->icsk_mtup
.enabled
||
1612 icsk
->icsk_mtup
.probe_size
||
1613 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
||
1614 tp
->snd_cwnd
< 11 ||
1615 tp
->rx_opt
.num_sacks
|| tp
->rx_opt
.dsack
)
1618 /* Very simple search strategy: just double the MSS. */
1619 mss_now
= tcp_current_mss(sk
);
1620 probe_size
= 2 * tp
->mss_cache
;
1621 size_needed
= probe_size
+ (tp
->reordering
+ 1) * tp
->mss_cache
;
1622 if (probe_size
> tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_high
)) {
1623 /* TODO: set timer for probe_converge_event */
1627 /* Have enough data in the send queue to probe? */
1628 if (tp
->write_seq
- tp
->snd_nxt
< size_needed
)
1631 if (tp
->snd_wnd
< size_needed
)
1633 if (after(tp
->snd_nxt
+ size_needed
, tcp_wnd_end(tp
)))
1636 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1637 if (tcp_packets_in_flight(tp
) + 2 > tp
->snd_cwnd
) {
1638 if (!tcp_packets_in_flight(tp
))
1644 /* We're allowed to probe. Build it now. */
1645 if ((nskb
= sk_stream_alloc_skb(sk
, probe_size
, GFP_ATOMIC
)) == NULL
)
1647 sk
->sk_wmem_queued
+= nskb
->truesize
;
1648 sk_mem_charge(sk
, nskb
->truesize
);
1650 skb
= tcp_send_head(sk
);
1652 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(skb
)->seq
;
1653 TCP_SKB_CB(nskb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ probe_size
;
1654 TCP_SKB_CB(nskb
)->flags
= TCPHDR_ACK
;
1655 TCP_SKB_CB(nskb
)->sacked
= 0;
1657 nskb
->ip_summed
= skb
->ip_summed
;
1659 tcp_insert_write_queue_before(nskb
, skb
, sk
);
1662 tcp_for_write_queue_from_safe(skb
, next
, sk
) {
1663 copy
= min_t(int, skb
->len
, probe_size
- len
);
1664 if (nskb
->ip_summed
)
1665 skb_copy_bits(skb
, 0, skb_put(nskb
, copy
), copy
);
1667 nskb
->csum
= skb_copy_and_csum_bits(skb
, 0,
1668 skb_put(nskb
, copy
),
1671 if (skb
->len
<= copy
) {
1672 /* We've eaten all the data from this skb.
1674 TCP_SKB_CB(nskb
)->flags
|= TCP_SKB_CB(skb
)->flags
;
1675 tcp_unlink_write_queue(skb
, sk
);
1676 sk_wmem_free_skb(sk
, skb
);
1678 TCP_SKB_CB(nskb
)->flags
|= TCP_SKB_CB(skb
)->flags
&
1679 ~(TCPHDR_FIN
|TCPHDR_PSH
);
1680 if (!skb_shinfo(skb
)->nr_frags
) {
1681 skb_pull(skb
, copy
);
1682 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1683 skb
->csum
= csum_partial(skb
->data
,
1686 __pskb_trim_head(skb
, copy
);
1687 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1689 TCP_SKB_CB(skb
)->seq
+= copy
;
1694 if (len
>= probe_size
)
1697 tcp_init_tso_segs(sk
, nskb
, nskb
->len
);
1699 /* We're ready to send. If this fails, the probe will
1700 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1701 TCP_SKB_CB(nskb
)->when
= tcp_time_stamp
;
1702 if (!tcp_transmit_skb(sk
, nskb
, 1, GFP_ATOMIC
)) {
1703 /* Decrement cwnd here because we are sending
1704 * effectively two packets. */
1706 tcp_event_new_data_sent(sk
, nskb
);
1708 icsk
->icsk_mtup
.probe_size
= tcp_mss_to_mtu(sk
, nskb
->len
);
1709 tp
->mtu_probe
.probe_seq_start
= TCP_SKB_CB(nskb
)->seq
;
1710 tp
->mtu_probe
.probe_seq_end
= TCP_SKB_CB(nskb
)->end_seq
;
1718 /* This routine writes packets to the network. It advances the
1719 * send_head. This happens as incoming acks open up the remote
1722 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1723 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1724 * account rare use of URG, this is not a big flaw.
1726 * Returns 1, if no segments are in flight and we have queued segments, but
1727 * cannot send anything now because of SWS or another problem.
1729 static int tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
,
1730 int push_one
, gfp_t gfp
)
1732 struct tcp_sock
*tp
= tcp_sk(sk
);
1733 struct sk_buff
*skb
;
1734 unsigned int tso_segs
, sent_pkts
;
1741 /* Do MTU probing. */
1742 result
= tcp_mtu_probe(sk
);
1745 } else if (result
> 0) {
1750 while ((skb
= tcp_send_head(sk
))) {
1753 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1756 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1760 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
)))
1763 if (tso_segs
== 1) {
1764 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
1765 (tcp_skb_is_last(sk
, skb
) ?
1766 nonagle
: TCP_NAGLE_PUSH
))))
1769 if (!push_one
&& tcp_tso_should_defer(sk
, skb
))
1774 if (tso_segs
> 1 && !tcp_urg_mode(tp
))
1775 limit
= tcp_mss_split_point(sk
, skb
, mss_now
,
1778 if (skb
->len
> limit
&&
1779 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
, gfp
)))
1782 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1784 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, gfp
)))
1787 /* Advance the send_head. This one is sent out.
1788 * This call will increment packets_out.
1790 tcp_event_new_data_sent(sk
, skb
);
1792 tcp_minshall_update(tp
, mss_now
, skb
);
1799 if (likely(sent_pkts
)) {
1800 tcp_cwnd_validate(sk
);
1803 return !tp
->packets_out
&& tcp_send_head(sk
);
1806 /* Push out any pending frames which were held back due to
1807 * TCP_CORK or attempt at coalescing tiny packets.
1808 * The socket must be locked by the caller.
1810 void __tcp_push_pending_frames(struct sock
*sk
, unsigned int cur_mss
,
1813 /* If we are closed, the bytes will have to remain here.
1814 * In time closedown will finish, we empty the write queue and
1815 * all will be happy.
1817 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
1820 if (tcp_write_xmit(sk
, cur_mss
, nonagle
, 0, GFP_ATOMIC
))
1821 tcp_check_probe_timer(sk
);
1824 /* Send _single_ skb sitting at the send head. This function requires
1825 * true push pending frames to setup probe timer etc.
1827 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
1829 struct sk_buff
*skb
= tcp_send_head(sk
);
1831 BUG_ON(!skb
|| skb
->len
< mss_now
);
1833 tcp_write_xmit(sk
, mss_now
, TCP_NAGLE_PUSH
, 1, sk
->sk_allocation
);
1836 /* This function returns the amount that we can raise the
1837 * usable window based on the following constraints
1839 * 1. The window can never be shrunk once it is offered (RFC 793)
1840 * 2. We limit memory per socket
1843 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1844 * RECV.NEXT + RCV.WIN fixed until:
1845 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1847 * i.e. don't raise the right edge of the window until you can raise
1848 * it at least MSS bytes.
1850 * Unfortunately, the recommended algorithm breaks header prediction,
1851 * since header prediction assumes th->window stays fixed.
1853 * Strictly speaking, keeping th->window fixed violates the receiver
1854 * side SWS prevention criteria. The problem is that under this rule
1855 * a stream of single byte packets will cause the right side of the
1856 * window to always advance by a single byte.
1858 * Of course, if the sender implements sender side SWS prevention
1859 * then this will not be a problem.
1861 * BSD seems to make the following compromise:
1863 * If the free space is less than the 1/4 of the maximum
1864 * space available and the free space is less than 1/2 mss,
1865 * then set the window to 0.
1866 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1867 * Otherwise, just prevent the window from shrinking
1868 * and from being larger than the largest representable value.
1870 * This prevents incremental opening of the window in the regime
1871 * where TCP is limited by the speed of the reader side taking
1872 * data out of the TCP receive queue. It does nothing about
1873 * those cases where the window is constrained on the sender side
1874 * because the pipeline is full.
1876 * BSD also seems to "accidentally" limit itself to windows that are a
1877 * multiple of MSS, at least until the free space gets quite small.
1878 * This would appear to be a side effect of the mbuf implementation.
1879 * Combining these two algorithms results in the observed behavior
1880 * of having a fixed window size at almost all times.
1882 * Below we obtain similar behavior by forcing the offered window to
1883 * a multiple of the mss when it is feasible to do so.
1885 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1886 * Regular options like TIMESTAMP are taken into account.
1888 u32
__tcp_select_window(struct sock
*sk
)
1890 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1891 struct tcp_sock
*tp
= tcp_sk(sk
);
1892 /* MSS for the peer's data. Previous versions used mss_clamp
1893 * here. I don't know if the value based on our guesses
1894 * of peer's MSS is better for the performance. It's more correct
1895 * but may be worse for the performance because of rcv_mss
1896 * fluctuations. --SAW 1998/11/1
1898 int mss
= icsk
->icsk_ack
.rcv_mss
;
1899 int free_space
= tcp_space(sk
);
1900 int full_space
= min_t(int, tp
->window_clamp
, tcp_full_space(sk
));
1903 if (mss
> full_space
)
1906 if (free_space
< (full_space
>> 1)) {
1907 icsk
->icsk_ack
.quick
= 0;
1909 if (tcp_memory_pressure
)
1910 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
,
1913 if (free_space
< mss
)
1917 if (free_space
> tp
->rcv_ssthresh
)
1918 free_space
= tp
->rcv_ssthresh
;
1920 /* Don't do rounding if we are using window scaling, since the
1921 * scaled window will not line up with the MSS boundary anyway.
1923 window
= tp
->rcv_wnd
;
1924 if (tp
->rx_opt
.rcv_wscale
) {
1925 window
= free_space
;
1927 /* Advertise enough space so that it won't get scaled away.
1928 * Import case: prevent zero window announcement if
1929 * 1<<rcv_wscale > mss.
1931 if (((window
>> tp
->rx_opt
.rcv_wscale
) << tp
->rx_opt
.rcv_wscale
) != window
)
1932 window
= (((window
>> tp
->rx_opt
.rcv_wscale
) + 1)
1933 << tp
->rx_opt
.rcv_wscale
);
1935 /* Get the largest window that is a nice multiple of mss.
1936 * Window clamp already applied above.
1937 * If our current window offering is within 1 mss of the
1938 * free space we just keep it. This prevents the divide
1939 * and multiply from happening most of the time.
1940 * We also don't do any window rounding when the free space
1943 if (window
<= free_space
- mss
|| window
> free_space
)
1944 window
= (free_space
/ mss
) * mss
;
1945 else if (mss
== full_space
&&
1946 free_space
> window
+ (full_space
>> 1))
1947 window
= free_space
;
1953 /* Collapses two adjacent SKB's during retransmission. */
1954 static void tcp_collapse_retrans(struct sock
*sk
, struct sk_buff
*skb
)
1956 struct tcp_sock
*tp
= tcp_sk(sk
);
1957 struct sk_buff
*next_skb
= tcp_write_queue_next(sk
, skb
);
1958 int skb_size
, next_skb_size
;
1960 skb_size
= skb
->len
;
1961 next_skb_size
= next_skb
->len
;
1963 BUG_ON(tcp_skb_pcount(skb
) != 1 || tcp_skb_pcount(next_skb
) != 1);
1965 tcp_highest_sack_combine(sk
, next_skb
, skb
);
1967 tcp_unlink_write_queue(next_skb
, sk
);
1969 skb_copy_from_linear_data(next_skb
, skb_put(skb
, next_skb_size
),
1972 if (next_skb
->ip_summed
== CHECKSUM_PARTIAL
)
1973 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1975 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1976 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
1978 /* Update sequence range on original skb. */
1979 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
1981 /* Merge over control information. This moves PSH/FIN etc. over */
1982 TCP_SKB_CB(skb
)->flags
|= TCP_SKB_CB(next_skb
)->flags
;
1984 /* All done, get rid of second SKB and account for it so
1985 * packet counting does not break.
1987 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
& TCPCB_EVER_RETRANS
;
1989 /* changed transmit queue under us so clear hints */
1990 tcp_clear_retrans_hints_partial(tp
);
1991 if (next_skb
== tp
->retransmit_skb_hint
)
1992 tp
->retransmit_skb_hint
= skb
;
1994 tcp_adjust_pcount(sk
, next_skb
, tcp_skb_pcount(next_skb
));
1996 sk_wmem_free_skb(sk
, next_skb
);
1999 /* Check if coalescing SKBs is legal. */
2000 static int tcp_can_collapse(struct sock
*sk
, struct sk_buff
*skb
)
2002 if (tcp_skb_pcount(skb
) > 1)
2004 /* TODO: SACK collapsing could be used to remove this condition */
2005 if (skb_shinfo(skb
)->nr_frags
!= 0)
2007 if (skb_cloned(skb
))
2009 if (skb
== tcp_send_head(sk
))
2011 /* Some heurestics for collapsing over SACK'd could be invented */
2012 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
2018 /* Collapse packets in the retransmit queue to make to create
2019 * less packets on the wire. This is only done on retransmission.
2021 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*to
,
2024 struct tcp_sock
*tp
= tcp_sk(sk
);
2025 struct sk_buff
*skb
= to
, *tmp
;
2028 if (!sysctl_tcp_retrans_collapse
)
2030 if (TCP_SKB_CB(skb
)->flags
& TCPHDR_SYN
)
2033 tcp_for_write_queue_from_safe(skb
, tmp
, sk
) {
2034 if (!tcp_can_collapse(sk
, skb
))
2046 /* Punt if not enough space exists in the first SKB for
2047 * the data in the second
2049 if (skb
->len
> skb_tailroom(to
))
2052 if (after(TCP_SKB_CB(skb
)->end_seq
, tcp_wnd_end(tp
)))
2055 tcp_collapse_retrans(sk
, to
);
2059 /* This retransmits one SKB. Policy decisions and retransmit queue
2060 * state updates are done by the caller. Returns non-zero if an
2061 * error occurred which prevented the send.
2063 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
2065 struct tcp_sock
*tp
= tcp_sk(sk
);
2066 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2067 unsigned int cur_mss
;
2070 /* Inconslusive MTU probe */
2071 if (icsk
->icsk_mtup
.probe_size
) {
2072 icsk
->icsk_mtup
.probe_size
= 0;
2075 /* Do not sent more than we queued. 1/4 is reserved for possible
2076 * copying overhead: fragmentation, tunneling, mangling etc.
2078 if (atomic_read(&sk
->sk_wmem_alloc
) >
2079 min(sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2), sk
->sk_sndbuf
))
2082 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
2083 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
2085 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
2089 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
2090 return -EHOSTUNREACH
; /* Routing failure or similar. */
2092 cur_mss
= tcp_current_mss(sk
);
2094 /* If receiver has shrunk his window, and skb is out of
2095 * new window, do not retransmit it. The exception is the
2096 * case, when window is shrunk to zero. In this case
2097 * our retransmit serves as a zero window probe.
2099 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
)) &&
2100 TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
2103 if (skb
->len
> cur_mss
) {
2104 if (tcp_fragment(sk
, skb
, cur_mss
, cur_mss
))
2105 return -ENOMEM
; /* We'll try again later. */
2107 int oldpcount
= tcp_skb_pcount(skb
);
2109 if (unlikely(oldpcount
> 1)) {
2110 tcp_init_tso_segs(sk
, skb
, cur_mss
);
2111 tcp_adjust_pcount(sk
, skb
, oldpcount
- tcp_skb_pcount(skb
));
2115 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
2117 /* Some Solaris stacks overoptimize and ignore the FIN on a
2118 * retransmit when old data is attached. So strip it off
2119 * since it is cheap to do so and saves bytes on the network.
2122 (TCP_SKB_CB(skb
)->flags
& TCPHDR_FIN
) &&
2123 tp
->snd_una
== (TCP_SKB_CB(skb
)->end_seq
- 1)) {
2124 if (!pskb_trim(skb
, 0)) {
2125 /* Reuse, even though it does some unnecessary work */
2126 tcp_init_nondata_skb(skb
, TCP_SKB_CB(skb
)->end_seq
- 1,
2127 TCP_SKB_CB(skb
)->flags
);
2128 skb
->ip_summed
= CHECKSUM_NONE
;
2132 /* Make a copy, if the first transmission SKB clone we made
2133 * is still in somebody's hands, else make a clone.
2135 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2137 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2140 /* Update global TCP statistics. */
2141 TCP_INC_STATS(sock_net(sk
), TCP_MIB_RETRANSSEGS
);
2143 tp
->total_retrans
++;
2145 #if FASTRETRANS_DEBUG > 0
2146 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2147 if (net_ratelimit())
2148 printk(KERN_DEBUG
"retrans_out leaked.\n");
2151 if (!tp
->retrans_out
)
2152 tp
->lost_retrans_low
= tp
->snd_nxt
;
2153 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
2154 tp
->retrans_out
+= tcp_skb_pcount(skb
);
2156 /* Save stamp of the first retransmit. */
2157 if (!tp
->retrans_stamp
)
2158 tp
->retrans_stamp
= TCP_SKB_CB(skb
)->when
;
2162 /* snd_nxt is stored to detect loss of retransmitted segment,
2163 * see tcp_input.c tcp_sacktag_write_queue().
2165 TCP_SKB_CB(skb
)->ack_seq
= tp
->snd_nxt
;
2170 /* Check if we forward retransmits are possible in the current
2171 * window/congestion state.
2173 static int tcp_can_forward_retransmit(struct sock
*sk
)
2175 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2176 struct tcp_sock
*tp
= tcp_sk(sk
);
2178 /* Forward retransmissions are possible only during Recovery. */
2179 if (icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
2182 /* No forward retransmissions in Reno are possible. */
2183 if (tcp_is_reno(tp
))
2186 /* Yeah, we have to make difficult choice between forward transmission
2187 * and retransmission... Both ways have their merits...
2189 * For now we do not retransmit anything, while we have some new
2190 * segments to send. In the other cases, follow rule 3 for
2191 * NextSeg() specified in RFC3517.
2194 if (tcp_may_send_now(sk
))
2200 /* This gets called after a retransmit timeout, and the initially
2201 * retransmitted data is acknowledged. It tries to continue
2202 * resending the rest of the retransmit queue, until either
2203 * we've sent it all or the congestion window limit is reached.
2204 * If doing SACK, the first ACK which comes back for a timeout
2205 * based retransmit packet might feed us FACK information again.
2206 * If so, we use it to avoid unnecessarily retransmissions.
2208 void tcp_xmit_retransmit_queue(struct sock
*sk
)
2210 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2211 struct tcp_sock
*tp
= tcp_sk(sk
);
2212 struct sk_buff
*skb
;
2213 struct sk_buff
*hole
= NULL
;
2216 int fwd_rexmitting
= 0;
2218 if (!tp
->packets_out
)
2222 tp
->retransmit_high
= tp
->snd_una
;
2224 if (tp
->retransmit_skb_hint
) {
2225 skb
= tp
->retransmit_skb_hint
;
2226 last_lost
= TCP_SKB_CB(skb
)->end_seq
;
2227 if (after(last_lost
, tp
->retransmit_high
))
2228 last_lost
= tp
->retransmit_high
;
2230 skb
= tcp_write_queue_head(sk
);
2231 last_lost
= tp
->snd_una
;
2234 tcp_for_write_queue_from(skb
, sk
) {
2235 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
2237 if (skb
== tcp_send_head(sk
))
2239 /* we could do better than to assign each time */
2241 tp
->retransmit_skb_hint
= skb
;
2243 /* Assume this retransmit will generate
2244 * only one packet for congestion window
2245 * calculation purposes. This works because
2246 * tcp_retransmit_skb() will chop up the
2247 * packet to be MSS sized and all the
2248 * packet counting works out.
2250 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
2253 if (fwd_rexmitting
) {
2255 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_highest_sack_seq(tp
)))
2257 mib_idx
= LINUX_MIB_TCPFORWARDRETRANS
;
2259 } else if (!before(TCP_SKB_CB(skb
)->seq
, tp
->retransmit_high
)) {
2260 tp
->retransmit_high
= last_lost
;
2261 if (!tcp_can_forward_retransmit(sk
))
2263 /* Backtrack if necessary to non-L'ed skb */
2271 } else if (!(sacked
& TCPCB_LOST
)) {
2272 if (hole
== NULL
&& !(sacked
& (TCPCB_SACKED_RETRANS
|TCPCB_SACKED_ACKED
)))
2277 last_lost
= TCP_SKB_CB(skb
)->end_seq
;
2278 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
2279 mib_idx
= LINUX_MIB_TCPFASTRETRANS
;
2281 mib_idx
= LINUX_MIB_TCPSLOWSTARTRETRANS
;
2284 if (sacked
& (TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))
2287 if (tcp_retransmit_skb(sk
, skb
))
2289 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2291 if (skb
== tcp_write_queue_head(sk
))
2292 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2293 inet_csk(sk
)->icsk_rto
,
2298 /* Send a fin. The caller locks the socket for us. This cannot be
2299 * allowed to fail queueing a FIN frame under any circumstances.
2301 void tcp_send_fin(struct sock
*sk
)
2303 struct tcp_sock
*tp
= tcp_sk(sk
);
2304 struct sk_buff
*skb
= tcp_write_queue_tail(sk
);
2307 /* Optimization, tack on the FIN if we have a queue of
2308 * unsent frames. But be careful about outgoing SACKS
2311 mss_now
= tcp_current_mss(sk
);
2313 if (tcp_send_head(sk
) != NULL
) {
2314 TCP_SKB_CB(skb
)->flags
|= TCPHDR_FIN
;
2315 TCP_SKB_CB(skb
)->end_seq
++;
2318 /* Socket is locked, keep trying until memory is available. */
2320 skb
= alloc_skb_fclone(MAX_TCP_HEADER
,
2327 /* Reserve space for headers and prepare control bits. */
2328 skb_reserve(skb
, MAX_TCP_HEADER
);
2329 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2330 tcp_init_nondata_skb(skb
, tp
->write_seq
,
2331 TCPHDR_ACK
| TCPHDR_FIN
);
2332 tcp_queue_skb(sk
, skb
);
2334 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_OFF
);
2337 /* We get here when a process closes a file descriptor (either due to
2338 * an explicit close() or as a byproduct of exit()'ing) and there
2339 * was unread data in the receive queue. This behavior is recommended
2340 * by RFC 2525, section 2.17. -DaveM
2342 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
2344 struct sk_buff
*skb
;
2346 /* NOTE: No TCP options attached and we never retransmit this. */
2347 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
2349 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
2353 /* Reserve space for headers and prepare control bits. */
2354 skb_reserve(skb
, MAX_TCP_HEADER
);
2355 tcp_init_nondata_skb(skb
, tcp_acceptable_seq(sk
),
2356 TCPHDR_ACK
| TCPHDR_RST
);
2358 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2359 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
2360 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
2362 TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTRSTS
);
2365 /* Send a crossed SYN-ACK during socket establishment.
2366 * WARNING: This routine must only be called when we have already sent
2367 * a SYN packet that crossed the incoming SYN that caused this routine
2368 * to get called. If this assumption fails then the initial rcv_wnd
2369 * and rcv_wscale values will not be correct.
2371 int tcp_send_synack(struct sock
*sk
)
2373 struct sk_buff
*skb
;
2375 skb
= tcp_write_queue_head(sk
);
2376 if (skb
== NULL
|| !(TCP_SKB_CB(skb
)->flags
& TCPHDR_SYN
)) {
2377 printk(KERN_DEBUG
"tcp_send_synack: wrong queue state\n");
2380 if (!(TCP_SKB_CB(skb
)->flags
& TCPHDR_ACK
)) {
2381 if (skb_cloned(skb
)) {
2382 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
2385 tcp_unlink_write_queue(skb
, sk
);
2386 skb_header_release(nskb
);
2387 __tcp_add_write_queue_head(sk
, nskb
);
2388 sk_wmem_free_skb(sk
, skb
);
2389 sk
->sk_wmem_queued
+= nskb
->truesize
;
2390 sk_mem_charge(sk
, nskb
->truesize
);
2394 TCP_SKB_CB(skb
)->flags
|= TCPHDR_ACK
;
2395 TCP_ECN_send_synack(tcp_sk(sk
), skb
);
2397 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2398 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2401 /* Prepare a SYN-ACK. */
2402 struct sk_buff
*tcp_make_synack(struct sock
*sk
, struct dst_entry
*dst
,
2403 struct request_sock
*req
,
2404 struct request_values
*rvp
)
2406 struct tcp_out_options opts
;
2407 struct tcp_extend_values
*xvp
= tcp_xv(rvp
);
2408 struct inet_request_sock
*ireq
= inet_rsk(req
);
2409 struct tcp_sock
*tp
= tcp_sk(sk
);
2410 const struct tcp_cookie_values
*cvp
= tp
->cookie_values
;
2412 struct sk_buff
*skb
;
2413 struct tcp_md5sig_key
*md5
;
2414 int tcp_header_size
;
2416 int s_data_desired
= 0;
2418 if (cvp
!= NULL
&& cvp
->s_data_constant
&& cvp
->s_data_desired
)
2419 s_data_desired
= cvp
->s_data_desired
;
2420 skb
= sock_wmalloc(sk
, MAX_TCP_HEADER
+ 15 + s_data_desired
, 1, GFP_ATOMIC
);
2424 /* Reserve space for headers. */
2425 skb_reserve(skb
, MAX_TCP_HEADER
);
2427 skb_dst_set(skb
, dst_clone(dst
));
2429 mss
= dst_metric_advmss(dst
);
2430 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< mss
)
2431 mss
= tp
->rx_opt
.user_mss
;
2433 if (req
->rcv_wnd
== 0) { /* ignored for retransmitted syns */
2435 /* Set this up on the first call only */
2436 req
->window_clamp
= tp
->window_clamp
? : dst_metric(dst
, RTAX_WINDOW
);
2438 /* limit the window selection if the user enforce a smaller rx buffer */
2439 if (sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
&&
2440 (req
->window_clamp
> tcp_full_space(sk
) || req
->window_clamp
== 0))
2441 req
->window_clamp
= tcp_full_space(sk
);
2443 /* tcp_full_space because it is guaranteed to be the first packet */
2444 tcp_select_initial_window(tcp_full_space(sk
),
2445 mss
- (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0),
2450 dst_metric(dst
, RTAX_INITRWND
));
2451 ireq
->rcv_wscale
= rcv_wscale
;
2454 memset(&opts
, 0, sizeof(opts
));
2455 #ifdef CONFIG_SYN_COOKIES
2456 if (unlikely(req
->cookie_ts
))
2457 TCP_SKB_CB(skb
)->when
= cookie_init_timestamp(req
);
2460 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2461 tcp_header_size
= tcp_synack_options(sk
, req
, mss
,
2462 skb
, &opts
, &md5
, xvp
)
2465 skb_push(skb
, tcp_header_size
);
2466 skb_reset_transport_header(skb
);
2469 memset(th
, 0, sizeof(struct tcphdr
));
2472 TCP_ECN_make_synack(req
, th
);
2473 th
->source
= ireq
->loc_port
;
2474 th
->dest
= ireq
->rmt_port
;
2475 /* Setting of flags are superfluous here for callers (and ECE is
2476 * not even correctly set)
2478 tcp_init_nondata_skb(skb
, tcp_rsk(req
)->snt_isn
,
2479 TCPHDR_SYN
| TCPHDR_ACK
);
2481 if (OPTION_COOKIE_EXTENSION
& opts
.options
) {
2482 if (s_data_desired
) {
2483 u8
*buf
= skb_put(skb
, s_data_desired
);
2485 /* copy data directly from the listening socket. */
2486 memcpy(buf
, cvp
->s_data_payload
, s_data_desired
);
2487 TCP_SKB_CB(skb
)->end_seq
+= s_data_desired
;
2490 if (opts
.hash_size
> 0) {
2491 __u32 workspace
[SHA_WORKSPACE_WORDS
];
2492 u32
*mess
= &xvp
->cookie_bakery
[COOKIE_DIGEST_WORDS
];
2493 u32
*tail
= &mess
[COOKIE_MESSAGE_WORDS
-1];
2495 /* Secret recipe depends on the Timestamp, (future)
2496 * Sequence and Acknowledgment Numbers, Initiator
2497 * Cookie, and others handled by IP variant caller.
2499 *tail
-- ^= opts
.tsval
;
2500 *tail
-- ^= tcp_rsk(req
)->rcv_isn
+ 1;
2501 *tail
-- ^= TCP_SKB_CB(skb
)->seq
+ 1;
2504 *tail
-- ^= (((__force u32
)th
->dest
<< 16) | (__force u32
)th
->source
);
2505 *tail
-- ^= (u32
)(unsigned long)cvp
; /* per sockopt */
2507 sha_transform((__u32
*)&xvp
->cookie_bakery
[0],
2510 opts
.hash_location
=
2511 (__u8
*)&xvp
->cookie_bakery
[0];
2515 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
2516 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_isn
+ 1);
2518 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2519 th
->window
= htons(min(req
->rcv_wnd
, 65535U));
2520 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
);
2521 th
->doff
= (tcp_header_size
>> 2);
2522 TCP_ADD_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
, tcp_skb_pcount(skb
));
2524 #ifdef CONFIG_TCP_MD5SIG
2525 /* Okay, we have all we need - do the md5 hash if needed */
2527 tcp_rsk(req
)->af_specific
->calc_md5_hash(opts
.hash_location
,
2528 md5
, NULL
, req
, skb
);
2534 EXPORT_SYMBOL(tcp_make_synack
);
2536 /* Do all connect socket setups that can be done AF independent. */
2537 static void tcp_connect_init(struct sock
*sk
)
2539 struct dst_entry
*dst
= __sk_dst_get(sk
);
2540 struct tcp_sock
*tp
= tcp_sk(sk
);
2543 /* We'll fix this up when we get a response from the other end.
2544 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2546 tp
->tcp_header_len
= sizeof(struct tcphdr
) +
2547 (sysctl_tcp_timestamps
? TCPOLEN_TSTAMP_ALIGNED
: 0);
2549 #ifdef CONFIG_TCP_MD5SIG
2550 if (tp
->af_specific
->md5_lookup(sk
, sk
) != NULL
)
2551 tp
->tcp_header_len
+= TCPOLEN_MD5SIG_ALIGNED
;
2554 /* If user gave his TCP_MAXSEG, record it to clamp */
2555 if (tp
->rx_opt
.user_mss
)
2556 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
2559 tcp_sync_mss(sk
, dst_mtu(dst
));
2561 if (!tp
->window_clamp
)
2562 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
2563 tp
->advmss
= dst_metric_advmss(dst
);
2564 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< tp
->advmss
)
2565 tp
->advmss
= tp
->rx_opt
.user_mss
;
2567 tcp_initialize_rcv_mss(sk
);
2569 /* limit the window selection if the user enforce a smaller rx buffer */
2570 if (sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
&&
2571 (tp
->window_clamp
> tcp_full_space(sk
) || tp
->window_clamp
== 0))
2572 tp
->window_clamp
= tcp_full_space(sk
);
2574 tcp_select_initial_window(tcp_full_space(sk
),
2575 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
2578 sysctl_tcp_window_scaling
,
2580 dst_metric(dst
, RTAX_INITRWND
));
2582 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
2583 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
2586 sock_reset_flag(sk
, SOCK_DONE
);
2589 tp
->snd_una
= tp
->write_seq
;
2590 tp
->snd_sml
= tp
->write_seq
;
2591 tp
->snd_up
= tp
->write_seq
;
2596 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
2597 inet_csk(sk
)->icsk_retransmits
= 0;
2598 tcp_clear_retrans(tp
);
2601 /* Build a SYN and send it off. */
2602 int tcp_connect(struct sock
*sk
)
2604 struct tcp_sock
*tp
= tcp_sk(sk
);
2605 struct sk_buff
*buff
;
2608 tcp_connect_init(sk
);
2610 buff
= alloc_skb_fclone(MAX_TCP_HEADER
+ 15, sk
->sk_allocation
);
2611 if (unlikely(buff
== NULL
))
2614 /* Reserve space for headers. */
2615 skb_reserve(buff
, MAX_TCP_HEADER
);
2617 tp
->snd_nxt
= tp
->write_seq
;
2618 tcp_init_nondata_skb(buff
, tp
->write_seq
++, TCPHDR_SYN
);
2619 TCP_ECN_send_syn(sk
, buff
);
2622 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2623 tp
->retrans_stamp
= TCP_SKB_CB(buff
)->when
;
2624 skb_header_release(buff
);
2625 __tcp_add_write_queue_tail(sk
, buff
);
2626 sk
->sk_wmem_queued
+= buff
->truesize
;
2627 sk_mem_charge(sk
, buff
->truesize
);
2628 tp
->packets_out
+= tcp_skb_pcount(buff
);
2629 err
= tcp_transmit_skb(sk
, buff
, 1, sk
->sk_allocation
);
2630 if (err
== -ECONNREFUSED
)
2633 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2634 * in order to make this packet get counted in tcpOutSegs.
2636 tp
->snd_nxt
= tp
->write_seq
;
2637 tp
->pushed_seq
= tp
->write_seq
;
2638 TCP_INC_STATS(sock_net(sk
), TCP_MIB_ACTIVEOPENS
);
2640 /* Timer for repeating the SYN until an answer. */
2641 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2642 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2645 EXPORT_SYMBOL(tcp_connect
);
2647 /* Send out a delayed ack, the caller does the policy checking
2648 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2651 void tcp_send_delayed_ack(struct sock
*sk
)
2653 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2654 int ato
= icsk
->icsk_ack
.ato
;
2655 unsigned long timeout
;
2657 if (ato
> TCP_DELACK_MIN
) {
2658 const struct tcp_sock
*tp
= tcp_sk(sk
);
2659 int max_ato
= HZ
/ 2;
2661 if (icsk
->icsk_ack
.pingpong
||
2662 (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
2663 max_ato
= TCP_DELACK_MAX
;
2665 /* Slow path, intersegment interval is "high". */
2667 /* If some rtt estimate is known, use it to bound delayed ack.
2668 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2672 int rtt
= max(tp
->srtt
>> 3, TCP_DELACK_MIN
);
2678 ato
= min(ato
, max_ato
);
2681 /* Stay within the limit we were given */
2682 timeout
= jiffies
+ ato
;
2684 /* Use new timeout only if there wasn't a older one earlier. */
2685 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
2686 /* If delack timer was blocked or is about to expire,
2689 if (icsk
->icsk_ack
.blocked
||
2690 time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
2695 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
2696 timeout
= icsk
->icsk_ack
.timeout
;
2698 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
2699 icsk
->icsk_ack
.timeout
= timeout
;
2700 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
2703 /* This routine sends an ack and also updates the window. */
2704 void tcp_send_ack(struct sock
*sk
)
2706 struct sk_buff
*buff
;
2708 /* If we have been reset, we may not send again. */
2709 if (sk
->sk_state
== TCP_CLOSE
)
2712 /* We are not putting this on the write queue, so
2713 * tcp_transmit_skb() will set the ownership to this
2716 buff
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2718 inet_csk_schedule_ack(sk
);
2719 inet_csk(sk
)->icsk_ack
.ato
= TCP_ATO_MIN
;
2720 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
2721 TCP_DELACK_MAX
, TCP_RTO_MAX
);
2725 /* Reserve space for headers and prepare control bits. */
2726 skb_reserve(buff
, MAX_TCP_HEADER
);
2727 tcp_init_nondata_skb(buff
, tcp_acceptable_seq(sk
), TCPHDR_ACK
);
2729 /* Send it off, this clears delayed acks for us. */
2730 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2731 tcp_transmit_skb(sk
, buff
, 0, GFP_ATOMIC
);
2734 /* This routine sends a packet with an out of date sequence
2735 * number. It assumes the other end will try to ack it.
2737 * Question: what should we make while urgent mode?
2738 * 4.4BSD forces sending single byte of data. We cannot send
2739 * out of window data, because we have SND.NXT==SND.MAX...
2741 * Current solution: to send TWO zero-length segments in urgent mode:
2742 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2743 * out-of-date with SND.UNA-1 to probe window.
2745 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
)
2747 struct tcp_sock
*tp
= tcp_sk(sk
);
2748 struct sk_buff
*skb
;
2750 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2751 skb
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2755 /* Reserve space for headers and set control bits. */
2756 skb_reserve(skb
, MAX_TCP_HEADER
);
2757 /* Use a previous sequence. This should cause the other
2758 * end to send an ack. Don't queue or clone SKB, just
2761 tcp_init_nondata_skb(skb
, tp
->snd_una
- !urgent
, TCPHDR_ACK
);
2762 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2763 return tcp_transmit_skb(sk
, skb
, 0, GFP_ATOMIC
);
2766 /* Initiate keepalive or window probe from timer. */
2767 int tcp_write_wakeup(struct sock
*sk
)
2769 struct tcp_sock
*tp
= tcp_sk(sk
);
2770 struct sk_buff
*skb
;
2772 if (sk
->sk_state
== TCP_CLOSE
)
2775 if ((skb
= tcp_send_head(sk
)) != NULL
&&
2776 before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
))) {
2778 unsigned int mss
= tcp_current_mss(sk
);
2779 unsigned int seg_size
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
2781 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
2782 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
2784 /* We are probing the opening of a window
2785 * but the window size is != 0
2786 * must have been a result SWS avoidance ( sender )
2788 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
2790 seg_size
= min(seg_size
, mss
);
2791 TCP_SKB_CB(skb
)->flags
|= TCPHDR_PSH
;
2792 if (tcp_fragment(sk
, skb
, seg_size
, mss
))
2794 } else if (!tcp_skb_pcount(skb
))
2795 tcp_set_skb_tso_segs(sk
, skb
, mss
);
2797 TCP_SKB_CB(skb
)->flags
|= TCPHDR_PSH
;
2798 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2799 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2801 tcp_event_new_data_sent(sk
, skb
);
2804 if (between(tp
->snd_up
, tp
->snd_una
+ 1, tp
->snd_una
+ 0xFFFF))
2805 tcp_xmit_probe_skb(sk
, 1);
2806 return tcp_xmit_probe_skb(sk
, 0);
2810 /* A window probe timeout has occurred. If window is not closed send
2811 * a partial packet else a zero probe.
2813 void tcp_send_probe0(struct sock
*sk
)
2815 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2816 struct tcp_sock
*tp
= tcp_sk(sk
);
2819 err
= tcp_write_wakeup(sk
);
2821 if (tp
->packets_out
|| !tcp_send_head(sk
)) {
2822 /* Cancel probe timer, if it is not required. */
2823 icsk
->icsk_probes_out
= 0;
2824 icsk
->icsk_backoff
= 0;
2829 if (icsk
->icsk_backoff
< sysctl_tcp_retries2
)
2830 icsk
->icsk_backoff
++;
2831 icsk
->icsk_probes_out
++;
2832 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2833 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
2836 /* If packet was not sent due to local congestion,
2837 * do not backoff and do not remember icsk_probes_out.
2838 * Let local senders to fight for local resources.
2840 * Use accumulated backoff yet.
2842 if (!icsk
->icsk_probes_out
)
2843 icsk
->icsk_probes_out
= 1;
2844 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2845 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
,
2846 TCP_RESOURCE_PROBE_INTERVAL
),