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
> init_rcv_wnd
* mss
))
242 *rcv_wnd
= init_rcv_wnd
* mss
;
243 else if (*rcv_wnd
> init_cwnd
* mss
)
244 *rcv_wnd
= init_cwnd
* mss
;
247 /* Set the clamp no higher than max representable value */
248 (*window_clamp
) = min(65535U << (*rcv_wscale
), *window_clamp
);
251 /* Chose a new window to advertise, update state in tcp_sock for the
252 * socket, and return result with RFC1323 scaling applied. The return
253 * value can be stuffed directly into th->window for an outgoing
256 static u16
tcp_select_window(struct sock
*sk
)
258 struct tcp_sock
*tp
= tcp_sk(sk
);
259 u32 cur_win
= tcp_receive_window(tp
);
260 u32 new_win
= __tcp_select_window(sk
);
262 /* Never shrink the offered window */
263 if (new_win
< cur_win
) {
264 /* Danger Will Robinson!
265 * Don't update rcv_wup/rcv_wnd here or else
266 * we will not be able to advertise a zero
267 * window in time. --DaveM
269 * Relax Will Robinson.
271 new_win
= ALIGN(cur_win
, 1 << tp
->rx_opt
.rcv_wscale
);
273 tp
->rcv_wnd
= new_win
;
274 tp
->rcv_wup
= tp
->rcv_nxt
;
276 /* Make sure we do not exceed the maximum possible
279 if (!tp
->rx_opt
.rcv_wscale
&& sysctl_tcp_workaround_signed_windows
)
280 new_win
= min(new_win
, MAX_TCP_WINDOW
);
282 new_win
= min(new_win
, (65535U << tp
->rx_opt
.rcv_wscale
));
284 /* RFC1323 scaling applied */
285 new_win
>>= tp
->rx_opt
.rcv_wscale
;
287 /* If we advertise zero window, disable fast path. */
294 /* Packet ECN state for a SYN-ACK */
295 static inline void TCP_ECN_send_synack(struct tcp_sock
*tp
, struct sk_buff
*skb
)
297 TCP_SKB_CB(skb
)->flags
&= ~TCPCB_FLAG_CWR
;
298 if (!(tp
->ecn_flags
& TCP_ECN_OK
))
299 TCP_SKB_CB(skb
)->flags
&= ~TCPCB_FLAG_ECE
;
302 /* Packet ECN state for a SYN. */
303 static inline void TCP_ECN_send_syn(struct sock
*sk
, struct sk_buff
*skb
)
305 struct tcp_sock
*tp
= tcp_sk(sk
);
308 if (sysctl_tcp_ecn
== 1) {
309 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_ECE
| TCPCB_FLAG_CWR
;
310 tp
->ecn_flags
= TCP_ECN_OK
;
314 static __inline__
void
315 TCP_ECN_make_synack(struct request_sock
*req
, struct tcphdr
*th
)
317 if (inet_rsk(req
)->ecn_ok
)
321 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
324 static inline void TCP_ECN_send(struct sock
*sk
, struct sk_buff
*skb
,
327 struct tcp_sock
*tp
= tcp_sk(sk
);
329 if (tp
->ecn_flags
& TCP_ECN_OK
) {
330 /* Not-retransmitted data segment: set ECT and inject CWR. */
331 if (skb
->len
!= tcp_header_len
&&
332 !before(TCP_SKB_CB(skb
)->seq
, tp
->snd_nxt
)) {
334 if (tp
->ecn_flags
& TCP_ECN_QUEUE_CWR
) {
335 tp
->ecn_flags
&= ~TCP_ECN_QUEUE_CWR
;
336 tcp_hdr(skb
)->cwr
= 1;
337 skb_shinfo(skb
)->gso_type
|= SKB_GSO_TCP_ECN
;
340 /* ACK or retransmitted segment: clear ECT|CE */
341 INET_ECN_dontxmit(sk
);
343 if (tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)
344 tcp_hdr(skb
)->ece
= 1;
348 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
349 * auto increment end seqno.
351 static void tcp_init_nondata_skb(struct sk_buff
*skb
, u32 seq
, u8 flags
)
353 skb
->ip_summed
= CHECKSUM_PARTIAL
;
356 TCP_SKB_CB(skb
)->flags
= flags
;
357 TCP_SKB_CB(skb
)->sacked
= 0;
359 skb_shinfo(skb
)->gso_segs
= 1;
360 skb_shinfo(skb
)->gso_size
= 0;
361 skb_shinfo(skb
)->gso_type
= 0;
363 TCP_SKB_CB(skb
)->seq
= seq
;
364 if (flags
& (TCPCB_FLAG_SYN
| TCPCB_FLAG_FIN
))
366 TCP_SKB_CB(skb
)->end_seq
= seq
;
369 static inline int tcp_urg_mode(const struct tcp_sock
*tp
)
371 return tp
->snd_una
!= tp
->snd_up
;
374 #define OPTION_SACK_ADVERTISE (1 << 0)
375 #define OPTION_TS (1 << 1)
376 #define OPTION_MD5 (1 << 2)
377 #define OPTION_WSCALE (1 << 3)
378 #define OPTION_COOKIE_EXTENSION (1 << 4)
380 struct tcp_out_options
{
381 u8 options
; /* bit field of OPTION_* */
382 u8 ws
; /* window scale, 0 to disable */
383 u8 num_sack_blocks
; /* number of SACK blocks to include */
384 u8 hash_size
; /* bytes in hash_location */
385 u16 mss
; /* 0 to disable */
386 __u32 tsval
, tsecr
; /* need to include OPTION_TS */
387 __u8
*hash_location
; /* temporary pointer, overloaded */
390 /* The sysctl int routines are generic, so check consistency here.
392 static u8
tcp_cookie_size_check(u8 desired
)
395 /* previously specified */
398 if (sysctl_tcp_cookie_size
<= 0) {
399 /* no default specified */
402 if (sysctl_tcp_cookie_size
<= TCP_COOKIE_MIN
) {
403 /* value too small, specify minimum */
404 return TCP_COOKIE_MIN
;
406 if (sysctl_tcp_cookie_size
>= TCP_COOKIE_MAX
) {
407 /* value too large, specify maximum */
408 return TCP_COOKIE_MAX
;
410 if (0x1 & sysctl_tcp_cookie_size
) {
411 /* 8-bit multiple, illegal, fix it */
412 return (u8
)(sysctl_tcp_cookie_size
+ 0x1);
414 return (u8
)sysctl_tcp_cookie_size
;
417 /* Write previously computed TCP options to the packet.
419 * Beware: Something in the Internet is very sensitive to the ordering of
420 * TCP options, we learned this through the hard way, so be careful here.
421 * Luckily we can at least blame others for their non-compliance but from
422 * inter-operatibility perspective it seems that we're somewhat stuck with
423 * the ordering which we have been using if we want to keep working with
424 * those broken things (not that it currently hurts anybody as there isn't
425 * particular reason why the ordering would need to be changed).
427 * At least SACK_PERM as the first option is known to lead to a disaster
428 * (but it may well be that other scenarios fail similarly).
430 static void tcp_options_write(__be32
*ptr
, struct tcp_sock
*tp
,
431 struct tcp_out_options
*opts
)
433 u8 options
= opts
->options
; /* mungable copy */
435 /* Having both authentication and cookies for security is redundant,
436 * and there's certainly not enough room. Instead, the cookie-less
437 * extension variant is proposed.
439 * Consider the pessimal case with authentication. The options
441 * COOKIE|MD5(20) + MSS(4) + SACK|TS(12) + WSCALE(4) == 40
443 if (unlikely(OPTION_MD5
& options
)) {
444 if (unlikely(OPTION_COOKIE_EXTENSION
& options
)) {
445 *ptr
++ = htonl((TCPOPT_COOKIE
<< 24) |
446 (TCPOLEN_COOKIE_BASE
<< 16) |
447 (TCPOPT_MD5SIG
<< 8) |
450 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
452 (TCPOPT_MD5SIG
<< 8) |
455 options
&= ~OPTION_COOKIE_EXTENSION
;
456 /* overload cookie hash location */
457 opts
->hash_location
= (__u8
*)ptr
;
461 if (unlikely(opts
->mss
)) {
462 *ptr
++ = htonl((TCPOPT_MSS
<< 24) |
463 (TCPOLEN_MSS
<< 16) |
467 if (likely(OPTION_TS
& options
)) {
468 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
469 *ptr
++ = htonl((TCPOPT_SACK_PERM
<< 24) |
470 (TCPOLEN_SACK_PERM
<< 16) |
471 (TCPOPT_TIMESTAMP
<< 8) |
473 options
&= ~OPTION_SACK_ADVERTISE
;
475 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
477 (TCPOPT_TIMESTAMP
<< 8) |
480 *ptr
++ = htonl(opts
->tsval
);
481 *ptr
++ = htonl(opts
->tsecr
);
484 /* Specification requires after timestamp, so do it now.
486 * Consider the pessimal case without authentication. The options
488 * MSS(4) + SACK|TS(12) + COOKIE(20) + WSCALE(4) == 40
490 if (unlikely(OPTION_COOKIE_EXTENSION
& options
)) {
491 __u8
*cookie_copy
= opts
->hash_location
;
492 u8 cookie_size
= opts
->hash_size
;
494 /* 8-bit multiple handled in tcp_cookie_size_check() above,
497 if (0x2 & cookie_size
) {
498 __u8
*p
= (__u8
*)ptr
;
500 /* 16-bit multiple */
501 *p
++ = TCPOPT_COOKIE
;
502 *p
++ = TCPOLEN_COOKIE_BASE
+ cookie_size
;
503 *p
++ = *cookie_copy
++;
504 *p
++ = *cookie_copy
++;
508 /* 32-bit multiple */
509 *ptr
++ = htonl(((TCPOPT_NOP
<< 24) |
511 (TCPOPT_COOKIE
<< 8) |
512 TCPOLEN_COOKIE_BASE
) +
516 if (cookie_size
> 0) {
517 memcpy(ptr
, cookie_copy
, cookie_size
);
518 ptr
+= (cookie_size
/ 4);
522 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
523 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
525 (TCPOPT_SACK_PERM
<< 8) |
529 if (unlikely(OPTION_WSCALE
& options
)) {
530 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
531 (TCPOPT_WINDOW
<< 16) |
532 (TCPOLEN_WINDOW
<< 8) |
536 if (unlikely(opts
->num_sack_blocks
)) {
537 struct tcp_sack_block
*sp
= tp
->rx_opt
.dsack
?
538 tp
->duplicate_sack
: tp
->selective_acks
;
541 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
544 (TCPOLEN_SACK_BASE
+ (opts
->num_sack_blocks
*
545 TCPOLEN_SACK_PERBLOCK
)));
547 for (this_sack
= 0; this_sack
< opts
->num_sack_blocks
;
549 *ptr
++ = htonl(sp
[this_sack
].start_seq
);
550 *ptr
++ = htonl(sp
[this_sack
].end_seq
);
553 tp
->rx_opt
.dsack
= 0;
557 /* Compute TCP options for SYN packets. This is not the final
558 * network wire format yet.
560 static unsigned tcp_syn_options(struct sock
*sk
, struct sk_buff
*skb
,
561 struct tcp_out_options
*opts
,
562 struct tcp_md5sig_key
**md5
) {
563 struct tcp_sock
*tp
= tcp_sk(sk
);
564 struct tcp_cookie_values
*cvp
= tp
->cookie_values
;
565 unsigned remaining
= MAX_TCP_OPTION_SPACE
;
566 u8 cookie_size
= (!tp
->rx_opt
.cookie_out_never
&& cvp
!= NULL
) ?
567 tcp_cookie_size_check(cvp
->cookie_desired
) :
570 #ifdef CONFIG_TCP_MD5SIG
571 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
573 opts
->options
|= OPTION_MD5
;
574 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
580 /* We always get an MSS option. The option bytes which will be seen in
581 * normal data packets should timestamps be used, must be in the MSS
582 * advertised. But we subtract them from tp->mss_cache so that
583 * calculations in tcp_sendmsg are simpler etc. So account for this
584 * fact here if necessary. If we don't do this correctly, as a
585 * receiver we won't recognize data packets as being full sized when we
586 * should, and thus we won't abide by the delayed ACK rules correctly.
587 * SACKs don't matter, we never delay an ACK when we have any of those
589 opts
->mss
= tcp_advertise_mss(sk
);
590 remaining
-= TCPOLEN_MSS_ALIGNED
;
592 if (likely(sysctl_tcp_timestamps
&& *md5
== NULL
)) {
593 opts
->options
|= OPTION_TS
;
594 opts
->tsval
= TCP_SKB_CB(skb
)->when
;
595 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
596 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
598 if (likely(sysctl_tcp_window_scaling
)) {
599 opts
->ws
= tp
->rx_opt
.rcv_wscale
;
600 opts
->options
|= OPTION_WSCALE
;
601 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
603 if (likely(sysctl_tcp_sack
)) {
604 opts
->options
|= OPTION_SACK_ADVERTISE
;
605 if (unlikely(!(OPTION_TS
& opts
->options
)))
606 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
609 /* Note that timestamps are required by the specification.
611 * Odd numbers of bytes are prohibited by the specification, ensuring
612 * that the cookie is 16-bit aligned, and the resulting cookie pair is
616 (OPTION_TS
& opts
->options
) &&
618 int need
= TCPOLEN_COOKIE_BASE
+ cookie_size
;
621 /* 32-bit multiple */
622 need
+= 2; /* NOPs */
624 if (need
> remaining
) {
625 /* try shrinking cookie to fit */
630 while (need
> remaining
&& TCP_COOKIE_MIN
<= cookie_size
) {
634 if (TCP_COOKIE_MIN
<= cookie_size
) {
635 opts
->options
|= OPTION_COOKIE_EXTENSION
;
636 opts
->hash_location
= (__u8
*)&cvp
->cookie_pair
[0];
637 opts
->hash_size
= cookie_size
;
639 /* Remember for future incarnations. */
640 cvp
->cookie_desired
= cookie_size
;
642 if (cvp
->cookie_desired
!= cvp
->cookie_pair_size
) {
643 /* Currently use random bytes as a nonce,
644 * assuming these are completely unpredictable
645 * by hostile users of the same system.
647 get_random_bytes(&cvp
->cookie_pair
[0],
649 cvp
->cookie_pair_size
= cookie_size
;
655 return MAX_TCP_OPTION_SPACE
- remaining
;
658 /* Set up TCP options for SYN-ACKs. */
659 static unsigned tcp_synack_options(struct sock
*sk
,
660 struct request_sock
*req
,
661 unsigned mss
, struct sk_buff
*skb
,
662 struct tcp_out_options
*opts
,
663 struct tcp_md5sig_key
**md5
,
664 struct tcp_extend_values
*xvp
)
666 struct inet_request_sock
*ireq
= inet_rsk(req
);
667 unsigned remaining
= MAX_TCP_OPTION_SPACE
;
668 u8 cookie_plus
= (xvp
!= NULL
&& !xvp
->cookie_out_never
) ?
671 bool doing_ts
= ireq
->tstamp_ok
;
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 doing_ts
&= !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(doing_ts
)) {
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(!doing_ts
))
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
& TCPCB_FLAG_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
);
834 skb_push(skb
, tcp_header_size
);
835 skb_reset_transport_header(skb
);
836 skb_set_owner_w(skb
, sk
);
838 /* Build TCP header and checksum it. */
840 th
->source
= inet
->inet_sport
;
841 th
->dest
= inet
->inet_dport
;
842 th
->seq
= htonl(tcb
->seq
);
843 th
->ack_seq
= htonl(tp
->rcv_nxt
);
844 *(((__be16
*)th
) + 6) = htons(((tcp_header_size
>> 2) << 12) |
847 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
848 /* RFC1323: The window in SYN & SYN/ACK segments
851 th
->window
= htons(min(tp
->rcv_wnd
, 65535U));
853 th
->window
= htons(tcp_select_window(sk
));
858 /* The urg_mode check is necessary during a below snd_una win probe */
859 if (unlikely(tcp_urg_mode(tp
) && before(tcb
->seq
, tp
->snd_up
))) {
860 if (before(tp
->snd_up
, tcb
->seq
+ 0x10000)) {
861 th
->urg_ptr
= htons(tp
->snd_up
- tcb
->seq
);
863 } else if (after(tcb
->seq
+ 0xFFFF, tp
->snd_nxt
)) {
864 th
->urg_ptr
= htons(0xFFFF);
869 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
);
870 if (likely((tcb
->flags
& TCPCB_FLAG_SYN
) == 0))
871 TCP_ECN_send(sk
, skb
, tcp_header_size
);
873 #ifdef CONFIG_TCP_MD5SIG
874 /* Calculate the MD5 hash, as we have all we need now */
876 sk
->sk_route_caps
&= ~NETIF_F_GSO_MASK
;
877 tp
->af_specific
->calc_md5_hash(opts
.hash_location
,
882 icsk
->icsk_af_ops
->send_check(sk
, skb
);
884 if (likely(tcb
->flags
& TCPCB_FLAG_ACK
))
885 tcp_event_ack_sent(sk
, tcp_skb_pcount(skb
));
887 if (skb
->len
!= tcp_header_size
)
888 tcp_event_data_sent(tp
, skb
, sk
);
890 if (after(tcb
->end_seq
, tp
->snd_nxt
) || tcb
->seq
== tcb
->end_seq
)
891 TCP_ADD_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
,
892 tcp_skb_pcount(skb
));
894 err
= icsk
->icsk_af_ops
->queue_xmit(skb
);
895 if (likely(err
<= 0))
898 tcp_enter_cwr(sk
, 1);
900 return net_xmit_eval(err
);
903 /* This routine just queues the buffer for sending.
905 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
906 * otherwise socket can stall.
908 static void tcp_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
910 struct tcp_sock
*tp
= tcp_sk(sk
);
912 /* Advance write_seq and place onto the write_queue. */
913 tp
->write_seq
= TCP_SKB_CB(skb
)->end_seq
;
914 skb_header_release(skb
);
915 tcp_add_write_queue_tail(sk
, skb
);
916 sk
->sk_wmem_queued
+= skb
->truesize
;
917 sk_mem_charge(sk
, skb
->truesize
);
920 /* Initialize TSO segments for a packet. */
921 static void tcp_set_skb_tso_segs(struct sock
*sk
, struct sk_buff
*skb
,
922 unsigned int mss_now
)
924 if (skb
->len
<= mss_now
|| !sk_can_gso(sk
) ||
925 skb
->ip_summed
== CHECKSUM_NONE
) {
926 /* Avoid the costly divide in the normal
929 skb_shinfo(skb
)->gso_segs
= 1;
930 skb_shinfo(skb
)->gso_size
= 0;
931 skb_shinfo(skb
)->gso_type
= 0;
933 skb_shinfo(skb
)->gso_segs
= DIV_ROUND_UP(skb
->len
, mss_now
);
934 skb_shinfo(skb
)->gso_size
= mss_now
;
935 skb_shinfo(skb
)->gso_type
= sk
->sk_gso_type
;
939 /* When a modification to fackets out becomes necessary, we need to check
940 * skb is counted to fackets_out or not.
942 static void tcp_adjust_fackets_out(struct sock
*sk
, struct sk_buff
*skb
,
945 struct tcp_sock
*tp
= tcp_sk(sk
);
947 if (!tp
->sacked_out
|| tcp_is_reno(tp
))
950 if (after(tcp_highest_sack_seq(tp
), TCP_SKB_CB(skb
)->seq
))
951 tp
->fackets_out
-= decr
;
954 /* Pcount in the middle of the write queue got changed, we need to do various
955 * tweaks to fix counters
957 static void tcp_adjust_pcount(struct sock
*sk
, struct sk_buff
*skb
, int decr
)
959 struct tcp_sock
*tp
= tcp_sk(sk
);
961 tp
->packets_out
-= decr
;
963 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
964 tp
->sacked_out
-= decr
;
965 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
966 tp
->retrans_out
-= decr
;
967 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
)
968 tp
->lost_out
-= decr
;
970 /* Reno case is special. Sigh... */
971 if (tcp_is_reno(tp
) && decr
> 0)
972 tp
->sacked_out
-= min_t(u32
, tp
->sacked_out
, decr
);
974 tcp_adjust_fackets_out(sk
, skb
, decr
);
976 if (tp
->lost_skb_hint
&&
977 before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
) &&
978 (tcp_is_fack(tp
) || (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)))
979 tp
->lost_cnt_hint
-= decr
;
981 tcp_verify_left_out(tp
);
984 /* Function to create two new TCP segments. Shrinks the given segment
985 * to the specified size and appends a new segment with the rest of the
986 * packet to the list. This won't be called frequently, I hope.
987 * Remember, these are still headerless SKBs at this point.
989 int tcp_fragment(struct sock
*sk
, struct sk_buff
*skb
, u32 len
,
990 unsigned int mss_now
)
992 struct tcp_sock
*tp
= tcp_sk(sk
);
993 struct sk_buff
*buff
;
994 int nsize
, old_factor
;
998 BUG_ON(len
> skb
->len
);
1000 nsize
= skb_headlen(skb
) - len
;
1004 if (skb_cloned(skb
) &&
1005 skb_is_nonlinear(skb
) &&
1006 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
1009 /* Get a new skb... force flag on. */
1010 buff
= sk_stream_alloc_skb(sk
, nsize
, GFP_ATOMIC
);
1012 return -ENOMEM
; /* We'll just try again later. */
1014 sk
->sk_wmem_queued
+= buff
->truesize
;
1015 sk_mem_charge(sk
, buff
->truesize
);
1016 nlen
= skb
->len
- len
- nsize
;
1017 buff
->truesize
+= nlen
;
1018 skb
->truesize
-= nlen
;
1020 /* Correct the sequence numbers. */
1021 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1022 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1023 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1025 /* PSH and FIN should only be set in the second packet. */
1026 flags
= TCP_SKB_CB(skb
)->flags
;
1027 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
| TCPCB_FLAG_PSH
);
1028 TCP_SKB_CB(buff
)->flags
= flags
;
1029 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
1031 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_PARTIAL
) {
1032 /* Copy and checksum data tail into the new buffer. */
1033 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
,
1034 skb_put(buff
, nsize
),
1039 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
1041 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1042 skb_split(skb
, buff
, len
);
1045 buff
->ip_summed
= skb
->ip_summed
;
1047 /* Looks stupid, but our code really uses when of
1048 * skbs, which it never sent before. --ANK
1050 TCP_SKB_CB(buff
)->when
= TCP_SKB_CB(skb
)->when
;
1051 buff
->tstamp
= skb
->tstamp
;
1053 old_factor
= tcp_skb_pcount(skb
);
1055 /* Fix up tso_factor for both original and new SKB. */
1056 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1057 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1059 /* If this packet has been sent out already, we must
1060 * adjust the various packet counters.
1062 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
1063 int diff
= old_factor
- tcp_skb_pcount(skb
) -
1064 tcp_skb_pcount(buff
);
1067 tcp_adjust_pcount(sk
, skb
, diff
);
1070 /* Link BUFF into the send queue. */
1071 skb_header_release(buff
);
1072 tcp_insert_write_queue_after(skb
, buff
, sk
);
1077 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1078 * eventually). The difference is that pulled data not copied, but
1079 * immediately discarded.
1081 static void __pskb_trim_head(struct sk_buff
*skb
, int len
)
1087 for (i
= 0; i
< skb_shinfo(skb
)->nr_frags
; i
++) {
1088 if (skb_shinfo(skb
)->frags
[i
].size
<= eat
) {
1089 put_page(skb_shinfo(skb
)->frags
[i
].page
);
1090 eat
-= skb_shinfo(skb
)->frags
[i
].size
;
1092 skb_shinfo(skb
)->frags
[k
] = skb_shinfo(skb
)->frags
[i
];
1094 skb_shinfo(skb
)->frags
[k
].page_offset
+= eat
;
1095 skb_shinfo(skb
)->frags
[k
].size
-= eat
;
1101 skb_shinfo(skb
)->nr_frags
= k
;
1103 skb_reset_tail_pointer(skb
);
1104 skb
->data_len
-= len
;
1105 skb
->len
= skb
->data_len
;
1108 /* Remove acked data from a packet in the transmit queue. */
1109 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
1111 if (skb_cloned(skb
) && pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
1114 /* If len == headlen, we avoid __skb_pull to preserve alignment. */
1115 if (unlikely(len
< skb_headlen(skb
)))
1116 __skb_pull(skb
, len
);
1118 __pskb_trim_head(skb
, len
- skb_headlen(skb
));
1120 TCP_SKB_CB(skb
)->seq
+= len
;
1121 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1123 skb
->truesize
-= len
;
1124 sk
->sk_wmem_queued
-= len
;
1125 sk_mem_uncharge(sk
, len
);
1126 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
1128 /* Any change of skb->len requires recalculation of tso
1131 if (tcp_skb_pcount(skb
) > 1)
1132 tcp_set_skb_tso_segs(sk
, skb
, tcp_current_mss(sk
));
1137 /* Calculate MSS. Not accounting for SACKs here. */
1138 int tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1140 struct tcp_sock
*tp
= tcp_sk(sk
);
1141 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1144 /* Calculate base mss without TCP options:
1145 It is MMS_S - sizeof(tcphdr) of rfc1122
1147 mss_now
= pmtu
- icsk
->icsk_af_ops
->net_header_len
- sizeof(struct tcphdr
);
1149 /* Clamp it (mss_clamp does not include tcp options) */
1150 if (mss_now
> tp
->rx_opt
.mss_clamp
)
1151 mss_now
= tp
->rx_opt
.mss_clamp
;
1153 /* Now subtract optional transport overhead */
1154 mss_now
-= icsk
->icsk_ext_hdr_len
;
1156 /* Then reserve room for full set of TCP options and 8 bytes of data */
1160 /* Now subtract TCP options size, not including SACKs */
1161 mss_now
-= tp
->tcp_header_len
- sizeof(struct tcphdr
);
1166 /* Inverse of above */
1167 int tcp_mss_to_mtu(struct sock
*sk
, int mss
)
1169 struct tcp_sock
*tp
= tcp_sk(sk
);
1170 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1174 tp
->tcp_header_len
+
1175 icsk
->icsk_ext_hdr_len
+
1176 icsk
->icsk_af_ops
->net_header_len
;
1181 /* MTU probing init per socket */
1182 void tcp_mtup_init(struct sock
*sk
)
1184 struct tcp_sock
*tp
= tcp_sk(sk
);
1185 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1187 icsk
->icsk_mtup
.enabled
= sysctl_tcp_mtu_probing
> 1;
1188 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+ sizeof(struct tcphdr
) +
1189 icsk
->icsk_af_ops
->net_header_len
;
1190 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, sysctl_tcp_base_mss
);
1191 icsk
->icsk_mtup
.probe_size
= 0;
1194 /* This function synchronize snd mss to current pmtu/exthdr set.
1196 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1197 for TCP options, but includes only bare TCP header.
1199 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1200 It is minimum of user_mss and mss received with SYN.
1201 It also does not include TCP options.
1203 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1205 tp->mss_cache is current effective sending mss, including
1206 all tcp options except for SACKs. It is evaluated,
1207 taking into account current pmtu, but never exceeds
1208 tp->rx_opt.mss_clamp.
1210 NOTE1. rfc1122 clearly states that advertised MSS
1211 DOES NOT include either tcp or ip options.
1213 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1214 are READ ONLY outside this function. --ANK (980731)
1216 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
1218 struct tcp_sock
*tp
= tcp_sk(sk
);
1219 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1222 if (icsk
->icsk_mtup
.search_high
> pmtu
)
1223 icsk
->icsk_mtup
.search_high
= pmtu
;
1225 mss_now
= tcp_mtu_to_mss(sk
, pmtu
);
1226 mss_now
= tcp_bound_to_half_wnd(tp
, mss_now
);
1228 /* And store cached results */
1229 icsk
->icsk_pmtu_cookie
= pmtu
;
1230 if (icsk
->icsk_mtup
.enabled
)
1231 mss_now
= min(mss_now
, tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
));
1232 tp
->mss_cache
= mss_now
;
1237 /* Compute the current effective MSS, taking SACKs and IP options,
1238 * and even PMTU discovery events into account.
1240 unsigned int tcp_current_mss(struct sock
*sk
)
1242 struct tcp_sock
*tp
= tcp_sk(sk
);
1243 struct dst_entry
*dst
= __sk_dst_get(sk
);
1245 unsigned header_len
;
1246 struct tcp_out_options opts
;
1247 struct tcp_md5sig_key
*md5
;
1249 mss_now
= tp
->mss_cache
;
1252 u32 mtu
= dst_mtu(dst
);
1253 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
1254 mss_now
= tcp_sync_mss(sk
, mtu
);
1257 header_len
= tcp_established_options(sk
, NULL
, &opts
, &md5
) +
1258 sizeof(struct tcphdr
);
1259 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1260 * some common options. If this is an odd packet (because we have SACK
1261 * blocks etc) then our calculated header_len will be different, and
1262 * we have to adjust mss_now correspondingly */
1263 if (header_len
!= tp
->tcp_header_len
) {
1264 int delta
= (int) header_len
- tp
->tcp_header_len
;
1271 /* Congestion window validation. (RFC2861) */
1272 static void tcp_cwnd_validate(struct sock
*sk
)
1274 struct tcp_sock
*tp
= tcp_sk(sk
);
1276 if (tp
->packets_out
>= tp
->snd_cwnd
) {
1277 /* Network is feed fully. */
1278 tp
->snd_cwnd_used
= 0;
1279 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1281 /* Network starves. */
1282 if (tp
->packets_out
> tp
->snd_cwnd_used
)
1283 tp
->snd_cwnd_used
= tp
->packets_out
;
1285 if (sysctl_tcp_slow_start_after_idle
&&
1286 (s32
)(tcp_time_stamp
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
)
1287 tcp_cwnd_application_limited(sk
);
1291 /* Returns the portion of skb which can be sent right away without
1292 * introducing MSS oddities to segment boundaries. In rare cases where
1293 * mss_now != mss_cache, we will request caller to create a small skb
1294 * per input skb which could be mostly avoided here (if desired).
1296 * We explicitly want to create a request for splitting write queue tail
1297 * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1298 * thus all the complexity (cwnd_len is always MSS multiple which we
1299 * return whenever allowed by the other factors). Basically we need the
1300 * modulo only when the receiver window alone is the limiting factor or
1301 * when we would be allowed to send the split-due-to-Nagle skb fully.
1303 static unsigned int tcp_mss_split_point(struct sock
*sk
, struct sk_buff
*skb
,
1304 unsigned int mss_now
, unsigned int cwnd
)
1306 struct tcp_sock
*tp
= tcp_sk(sk
);
1307 u32 needed
, window
, cwnd_len
;
1309 window
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1310 cwnd_len
= mss_now
* cwnd
;
1312 if (likely(cwnd_len
<= window
&& skb
!= tcp_write_queue_tail(sk
)))
1315 needed
= min(skb
->len
, window
);
1317 if (cwnd_len
<= needed
)
1320 return needed
- needed
% mss_now
;
1323 /* Can at least one segment of SKB be sent right now, according to the
1324 * congestion window rules? If so, return how many segments are allowed.
1326 static inline unsigned int tcp_cwnd_test(struct tcp_sock
*tp
,
1327 struct sk_buff
*skb
)
1329 u32 in_flight
, cwnd
;
1331 /* Don't be strict about the congestion window for the final FIN. */
1332 if ((TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
) &&
1333 tcp_skb_pcount(skb
) == 1)
1336 in_flight
= tcp_packets_in_flight(tp
);
1337 cwnd
= tp
->snd_cwnd
;
1338 if (in_flight
< cwnd
)
1339 return (cwnd
- in_flight
);
1344 /* Intialize TSO state of a skb.
1345 * This must be invoked the first time we consider transmitting
1346 * SKB onto the wire.
1348 static int tcp_init_tso_segs(struct sock
*sk
, struct sk_buff
*skb
,
1349 unsigned int mss_now
)
1351 int tso_segs
= tcp_skb_pcount(skb
);
1353 if (!tso_segs
|| (tso_segs
> 1 && tcp_skb_mss(skb
) != mss_now
)) {
1354 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1355 tso_segs
= tcp_skb_pcount(skb
);
1360 /* Minshall's variant of the Nagle send check. */
1361 static inline int tcp_minshall_check(const struct tcp_sock
*tp
)
1363 return after(tp
->snd_sml
, tp
->snd_una
) &&
1364 !after(tp
->snd_sml
, tp
->snd_nxt
);
1367 /* Return 0, if packet can be sent now without violation Nagle's rules:
1368 * 1. It is full sized.
1369 * 2. Or it contains FIN. (already checked by caller)
1370 * 3. Or TCP_NODELAY was set.
1371 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1372 * With Minshall's modification: all sent small packets are ACKed.
1374 static inline int tcp_nagle_check(const struct tcp_sock
*tp
,
1375 const struct sk_buff
*skb
,
1376 unsigned mss_now
, int nonagle
)
1378 return (skb
->len
< mss_now
&&
1379 ((nonagle
& TCP_NAGLE_CORK
) ||
1380 (!nonagle
&& tp
->packets_out
&& tcp_minshall_check(tp
))));
1383 /* Return non-zero if the Nagle test allows this packet to be
1386 static inline int tcp_nagle_test(struct tcp_sock
*tp
, struct sk_buff
*skb
,
1387 unsigned int cur_mss
, int nonagle
)
1389 /* Nagle rule does not apply to frames, which sit in the middle of the
1390 * write_queue (they have no chances to get new data).
1392 * This is implemented in the callers, where they modify the 'nonagle'
1393 * argument based upon the location of SKB in the send queue.
1395 if (nonagle
& TCP_NAGLE_PUSH
)
1398 /* Don't use the nagle rule for urgent data (or for the final FIN).
1399 * Nagle can be ignored during F-RTO too (see RFC4138).
1401 if (tcp_urg_mode(tp
) || (tp
->frto_counter
== 2) ||
1402 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
))
1405 if (!tcp_nagle_check(tp
, skb
, cur_mss
, nonagle
))
1411 /* Does at least the first segment of SKB fit into the send window? */
1412 static inline int tcp_snd_wnd_test(struct tcp_sock
*tp
, struct sk_buff
*skb
,
1413 unsigned int cur_mss
)
1415 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1417 if (skb
->len
> cur_mss
)
1418 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
1420 return !after(end_seq
, tcp_wnd_end(tp
));
1423 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1424 * should be put on the wire right now. If so, it returns the number of
1425 * packets allowed by the congestion window.
1427 static unsigned int tcp_snd_test(struct sock
*sk
, struct sk_buff
*skb
,
1428 unsigned int cur_mss
, int nonagle
)
1430 struct tcp_sock
*tp
= tcp_sk(sk
);
1431 unsigned int cwnd_quota
;
1433 tcp_init_tso_segs(sk
, skb
, cur_mss
);
1435 if (!tcp_nagle_test(tp
, skb
, cur_mss
, nonagle
))
1438 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1439 if (cwnd_quota
&& !tcp_snd_wnd_test(tp
, skb
, cur_mss
))
1445 /* Test if sending is allowed right now. */
1446 int tcp_may_send_now(struct sock
*sk
)
1448 struct tcp_sock
*tp
= tcp_sk(sk
);
1449 struct sk_buff
*skb
= tcp_send_head(sk
);
1452 tcp_snd_test(sk
, skb
, tcp_current_mss(sk
),
1453 (tcp_skb_is_last(sk
, skb
) ?
1454 tp
->nonagle
: TCP_NAGLE_PUSH
)));
1457 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1458 * which is put after SKB on the list. It is very much like
1459 * tcp_fragment() except that it may make several kinds of assumptions
1460 * in order to speed up the splitting operation. In particular, we
1461 * know that all the data is in scatter-gather pages, and that the
1462 * packet has never been sent out before (and thus is not cloned).
1464 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
,
1465 unsigned int mss_now
)
1467 struct sk_buff
*buff
;
1468 int nlen
= skb
->len
- len
;
1471 /* All of a TSO frame must be composed of paged data. */
1472 if (skb
->len
!= skb
->data_len
)
1473 return tcp_fragment(sk
, skb
, len
, mss_now
);
1475 buff
= sk_stream_alloc_skb(sk
, 0, GFP_ATOMIC
);
1476 if (unlikely(buff
== NULL
))
1479 sk
->sk_wmem_queued
+= buff
->truesize
;
1480 sk_mem_charge(sk
, buff
->truesize
);
1481 buff
->truesize
+= nlen
;
1482 skb
->truesize
-= nlen
;
1484 /* Correct the sequence numbers. */
1485 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1486 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1487 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1489 /* PSH and FIN should only be set in the second packet. */
1490 flags
= TCP_SKB_CB(skb
)->flags
;
1491 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
| TCPCB_FLAG_PSH
);
1492 TCP_SKB_CB(buff
)->flags
= flags
;
1494 /* This packet was never sent out yet, so no SACK bits. */
1495 TCP_SKB_CB(buff
)->sacked
= 0;
1497 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_PARTIAL
;
1498 skb_split(skb
, buff
, len
);
1500 /* Fix up tso_factor for both original and new SKB. */
1501 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1502 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1504 /* Link BUFF into the send queue. */
1505 skb_header_release(buff
);
1506 tcp_insert_write_queue_after(skb
, buff
, sk
);
1511 /* Try to defer sending, if possible, in order to minimize the amount
1512 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1514 * This algorithm is from John Heffner.
1516 static int tcp_tso_should_defer(struct sock
*sk
, struct sk_buff
*skb
)
1518 struct tcp_sock
*tp
= tcp_sk(sk
);
1519 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1520 u32 send_win
, cong_win
, limit
, in_flight
;
1522 if (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
)
1525 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
1528 /* Defer for less than two clock ticks. */
1529 if (tp
->tso_deferred
&&
1530 (((u32
)jiffies
<< 1) >> 1) - (tp
->tso_deferred
>> 1) > 1)
1533 in_flight
= tcp_packets_in_flight(tp
);
1535 BUG_ON(tcp_skb_pcount(skb
) <= 1 || (tp
->snd_cwnd
<= in_flight
));
1537 send_win
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1539 /* From in_flight test above, we know that cwnd > in_flight. */
1540 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
1542 limit
= min(send_win
, cong_win
);
1544 /* If a full-sized TSO skb can be sent, do it. */
1545 if (limit
>= sk
->sk_gso_max_size
)
1548 /* Middle in queue won't get any more data, full sendable already? */
1549 if ((skb
!= tcp_write_queue_tail(sk
)) && (limit
>= skb
->len
))
1552 if (sysctl_tcp_tso_win_divisor
) {
1553 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
1555 /* If at least some fraction of a window is available,
1558 chunk
/= sysctl_tcp_tso_win_divisor
;
1562 /* Different approach, try not to defer past a single
1563 * ACK. Receiver should ACK every other full sized
1564 * frame, so if we have space for more than 3 frames
1567 if (limit
> tcp_max_burst(tp
) * tp
->mss_cache
)
1571 /* Ok, it looks like it is advisable to defer. */
1572 tp
->tso_deferred
= 1 | (jiffies
<< 1);
1577 tp
->tso_deferred
= 0;
1581 /* Create a new MTU probe if we are ready.
1582 * MTU probe is regularly attempting to increase the path MTU by
1583 * deliberately sending larger packets. This discovers routing
1584 * changes resulting in larger path MTUs.
1586 * Returns 0 if we should wait to probe (no cwnd available),
1587 * 1 if a probe was sent,
1590 static int tcp_mtu_probe(struct sock
*sk
)
1592 struct tcp_sock
*tp
= tcp_sk(sk
);
1593 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1594 struct sk_buff
*skb
, *nskb
, *next
;
1601 /* Not currently probing/verifying,
1603 * have enough cwnd, and
1604 * not SACKing (the variable headers throw things off) */
1605 if (!icsk
->icsk_mtup
.enabled
||
1606 icsk
->icsk_mtup
.probe_size
||
1607 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
||
1608 tp
->snd_cwnd
< 11 ||
1609 tp
->rx_opt
.num_sacks
|| tp
->rx_opt
.dsack
)
1612 /* Very simple search strategy: just double the MSS. */
1613 mss_now
= tcp_current_mss(sk
);
1614 probe_size
= 2 * tp
->mss_cache
;
1615 size_needed
= probe_size
+ (tp
->reordering
+ 1) * tp
->mss_cache
;
1616 if (probe_size
> tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_high
)) {
1617 /* TODO: set timer for probe_converge_event */
1621 /* Have enough data in the send queue to probe? */
1622 if (tp
->write_seq
- tp
->snd_nxt
< size_needed
)
1625 if (tp
->snd_wnd
< size_needed
)
1627 if (after(tp
->snd_nxt
+ size_needed
, tcp_wnd_end(tp
)))
1630 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1631 if (tcp_packets_in_flight(tp
) + 2 > tp
->snd_cwnd
) {
1632 if (!tcp_packets_in_flight(tp
))
1638 /* We're allowed to probe. Build it now. */
1639 if ((nskb
= sk_stream_alloc_skb(sk
, probe_size
, GFP_ATOMIC
)) == NULL
)
1641 sk
->sk_wmem_queued
+= nskb
->truesize
;
1642 sk_mem_charge(sk
, nskb
->truesize
);
1644 skb
= tcp_send_head(sk
);
1646 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(skb
)->seq
;
1647 TCP_SKB_CB(nskb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ probe_size
;
1648 TCP_SKB_CB(nskb
)->flags
= TCPCB_FLAG_ACK
;
1649 TCP_SKB_CB(nskb
)->sacked
= 0;
1651 nskb
->ip_summed
= skb
->ip_summed
;
1653 tcp_insert_write_queue_before(nskb
, skb
, sk
);
1656 tcp_for_write_queue_from_safe(skb
, next
, sk
) {
1657 copy
= min_t(int, skb
->len
, probe_size
- len
);
1658 if (nskb
->ip_summed
)
1659 skb_copy_bits(skb
, 0, skb_put(nskb
, copy
), copy
);
1661 nskb
->csum
= skb_copy_and_csum_bits(skb
, 0,
1662 skb_put(nskb
, copy
),
1665 if (skb
->len
<= copy
) {
1666 /* We've eaten all the data from this skb.
1668 TCP_SKB_CB(nskb
)->flags
|= TCP_SKB_CB(skb
)->flags
;
1669 tcp_unlink_write_queue(skb
, sk
);
1670 sk_wmem_free_skb(sk
, skb
);
1672 TCP_SKB_CB(nskb
)->flags
|= TCP_SKB_CB(skb
)->flags
&
1673 ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
1674 if (!skb_shinfo(skb
)->nr_frags
) {
1675 skb_pull(skb
, copy
);
1676 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1677 skb
->csum
= csum_partial(skb
->data
,
1680 __pskb_trim_head(skb
, copy
);
1681 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1683 TCP_SKB_CB(skb
)->seq
+= copy
;
1688 if (len
>= probe_size
)
1691 tcp_init_tso_segs(sk
, nskb
, nskb
->len
);
1693 /* We're ready to send. If this fails, the probe will
1694 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1695 TCP_SKB_CB(nskb
)->when
= tcp_time_stamp
;
1696 if (!tcp_transmit_skb(sk
, nskb
, 1, GFP_ATOMIC
)) {
1697 /* Decrement cwnd here because we are sending
1698 * effectively two packets. */
1700 tcp_event_new_data_sent(sk
, nskb
);
1702 icsk
->icsk_mtup
.probe_size
= tcp_mss_to_mtu(sk
, nskb
->len
);
1703 tp
->mtu_probe
.probe_seq_start
= TCP_SKB_CB(nskb
)->seq
;
1704 tp
->mtu_probe
.probe_seq_end
= TCP_SKB_CB(nskb
)->end_seq
;
1712 /* This routine writes packets to the network. It advances the
1713 * send_head. This happens as incoming acks open up the remote
1716 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1717 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1718 * account rare use of URG, this is not a big flaw.
1720 * Returns 1, if no segments are in flight and we have queued segments, but
1721 * cannot send anything now because of SWS or another problem.
1723 static int tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
,
1724 int push_one
, gfp_t gfp
)
1726 struct tcp_sock
*tp
= tcp_sk(sk
);
1727 struct sk_buff
*skb
;
1728 unsigned int tso_segs
, sent_pkts
;
1735 /* Do MTU probing. */
1736 result
= tcp_mtu_probe(sk
);
1739 } else if (result
> 0) {
1744 while ((skb
= tcp_send_head(sk
))) {
1747 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1750 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1754 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
)))
1757 if (tso_segs
== 1) {
1758 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
1759 (tcp_skb_is_last(sk
, skb
) ?
1760 nonagle
: TCP_NAGLE_PUSH
))))
1763 if (!push_one
&& tcp_tso_should_defer(sk
, skb
))
1768 if (tso_segs
> 1 && !tcp_urg_mode(tp
))
1769 limit
= tcp_mss_split_point(sk
, skb
, mss_now
,
1772 if (skb
->len
> limit
&&
1773 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
)))
1776 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1778 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, gfp
)))
1781 /* Advance the send_head. This one is sent out.
1782 * This call will increment packets_out.
1784 tcp_event_new_data_sent(sk
, skb
);
1786 tcp_minshall_update(tp
, mss_now
, skb
);
1793 if (likely(sent_pkts
)) {
1794 tcp_cwnd_validate(sk
);
1797 return !tp
->packets_out
&& tcp_send_head(sk
);
1800 /* Push out any pending frames which were held back due to
1801 * TCP_CORK or attempt at coalescing tiny packets.
1802 * The socket must be locked by the caller.
1804 void __tcp_push_pending_frames(struct sock
*sk
, unsigned int cur_mss
,
1807 /* If we are closed, the bytes will have to remain here.
1808 * In time closedown will finish, we empty the write queue and
1809 * all will be happy.
1811 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
1814 if (tcp_write_xmit(sk
, cur_mss
, nonagle
, 0, GFP_ATOMIC
))
1815 tcp_check_probe_timer(sk
);
1818 /* Send _single_ skb sitting at the send head. This function requires
1819 * true push pending frames to setup probe timer etc.
1821 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
1823 struct sk_buff
*skb
= tcp_send_head(sk
);
1825 BUG_ON(!skb
|| skb
->len
< mss_now
);
1827 tcp_write_xmit(sk
, mss_now
, TCP_NAGLE_PUSH
, 1, sk
->sk_allocation
);
1830 /* This function returns the amount that we can raise the
1831 * usable window based on the following constraints
1833 * 1. The window can never be shrunk once it is offered (RFC 793)
1834 * 2. We limit memory per socket
1837 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1838 * RECV.NEXT + RCV.WIN fixed until:
1839 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1841 * i.e. don't raise the right edge of the window until you can raise
1842 * it at least MSS bytes.
1844 * Unfortunately, the recommended algorithm breaks header prediction,
1845 * since header prediction assumes th->window stays fixed.
1847 * Strictly speaking, keeping th->window fixed violates the receiver
1848 * side SWS prevention criteria. The problem is that under this rule
1849 * a stream of single byte packets will cause the right side of the
1850 * window to always advance by a single byte.
1852 * Of course, if the sender implements sender side SWS prevention
1853 * then this will not be a problem.
1855 * BSD seems to make the following compromise:
1857 * If the free space is less than the 1/4 of the maximum
1858 * space available and the free space is less than 1/2 mss,
1859 * then set the window to 0.
1860 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1861 * Otherwise, just prevent the window from shrinking
1862 * and from being larger than the largest representable value.
1864 * This prevents incremental opening of the window in the regime
1865 * where TCP is limited by the speed of the reader side taking
1866 * data out of the TCP receive queue. It does nothing about
1867 * those cases where the window is constrained on the sender side
1868 * because the pipeline is full.
1870 * BSD also seems to "accidentally" limit itself to windows that are a
1871 * multiple of MSS, at least until the free space gets quite small.
1872 * This would appear to be a side effect of the mbuf implementation.
1873 * Combining these two algorithms results in the observed behavior
1874 * of having a fixed window size at almost all times.
1876 * Below we obtain similar behavior by forcing the offered window to
1877 * a multiple of the mss when it is feasible to do so.
1879 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1880 * Regular options like TIMESTAMP are taken into account.
1882 u32
__tcp_select_window(struct sock
*sk
)
1884 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1885 struct tcp_sock
*tp
= tcp_sk(sk
);
1886 /* MSS for the peer's data. Previous versions used mss_clamp
1887 * here. I don't know if the value based on our guesses
1888 * of peer's MSS is better for the performance. It's more correct
1889 * but may be worse for the performance because of rcv_mss
1890 * fluctuations. --SAW 1998/11/1
1892 int mss
= icsk
->icsk_ack
.rcv_mss
;
1893 int free_space
= tcp_space(sk
);
1894 int full_space
= min_t(int, tp
->window_clamp
, tcp_full_space(sk
));
1897 if (mss
> full_space
)
1900 if (free_space
< (full_space
>> 1)) {
1901 icsk
->icsk_ack
.quick
= 0;
1903 if (tcp_memory_pressure
)
1904 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
,
1907 if (free_space
< mss
)
1911 if (free_space
> tp
->rcv_ssthresh
)
1912 free_space
= tp
->rcv_ssthresh
;
1914 /* Don't do rounding if we are using window scaling, since the
1915 * scaled window will not line up with the MSS boundary anyway.
1917 window
= tp
->rcv_wnd
;
1918 if (tp
->rx_opt
.rcv_wscale
) {
1919 window
= free_space
;
1921 /* Advertise enough space so that it won't get scaled away.
1922 * Import case: prevent zero window announcement if
1923 * 1<<rcv_wscale > mss.
1925 if (((window
>> tp
->rx_opt
.rcv_wscale
) << tp
->rx_opt
.rcv_wscale
) != window
)
1926 window
= (((window
>> tp
->rx_opt
.rcv_wscale
) + 1)
1927 << tp
->rx_opt
.rcv_wscale
);
1929 /* Get the largest window that is a nice multiple of mss.
1930 * Window clamp already applied above.
1931 * If our current window offering is within 1 mss of the
1932 * free space we just keep it. This prevents the divide
1933 * and multiply from happening most of the time.
1934 * We also don't do any window rounding when the free space
1937 if (window
<= free_space
- mss
|| window
> free_space
)
1938 window
= (free_space
/ mss
) * mss
;
1939 else if (mss
== full_space
&&
1940 free_space
> window
+ (full_space
>> 1))
1941 window
= free_space
;
1947 /* Collapses two adjacent SKB's during retransmission. */
1948 static void tcp_collapse_retrans(struct sock
*sk
, struct sk_buff
*skb
)
1950 struct tcp_sock
*tp
= tcp_sk(sk
);
1951 struct sk_buff
*next_skb
= tcp_write_queue_next(sk
, skb
);
1952 int skb_size
, next_skb_size
;
1954 skb_size
= skb
->len
;
1955 next_skb_size
= next_skb
->len
;
1957 BUG_ON(tcp_skb_pcount(skb
) != 1 || tcp_skb_pcount(next_skb
) != 1);
1959 tcp_highest_sack_combine(sk
, next_skb
, skb
);
1961 tcp_unlink_write_queue(next_skb
, sk
);
1963 skb_copy_from_linear_data(next_skb
, skb_put(skb
, next_skb_size
),
1966 if (next_skb
->ip_summed
== CHECKSUM_PARTIAL
)
1967 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1969 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1970 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
1972 /* Update sequence range on original skb. */
1973 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
1975 /* Merge over control information. This moves PSH/FIN etc. over */
1976 TCP_SKB_CB(skb
)->flags
|= TCP_SKB_CB(next_skb
)->flags
;
1978 /* All done, get rid of second SKB and account for it so
1979 * packet counting does not break.
1981 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
& TCPCB_EVER_RETRANS
;
1983 /* changed transmit queue under us so clear hints */
1984 tcp_clear_retrans_hints_partial(tp
);
1985 if (next_skb
== tp
->retransmit_skb_hint
)
1986 tp
->retransmit_skb_hint
= skb
;
1988 tcp_adjust_pcount(sk
, next_skb
, tcp_skb_pcount(next_skb
));
1990 sk_wmem_free_skb(sk
, next_skb
);
1993 /* Check if coalescing SKBs is legal. */
1994 static int tcp_can_collapse(struct sock
*sk
, struct sk_buff
*skb
)
1996 if (tcp_skb_pcount(skb
) > 1)
1998 /* TODO: SACK collapsing could be used to remove this condition */
1999 if (skb_shinfo(skb
)->nr_frags
!= 0)
2001 if (skb_cloned(skb
))
2003 if (skb
== tcp_send_head(sk
))
2005 /* Some heurestics for collapsing over SACK'd could be invented */
2006 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
2012 /* Collapse packets in the retransmit queue to make to create
2013 * less packets on the wire. This is only done on retransmission.
2015 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*to
,
2018 struct tcp_sock
*tp
= tcp_sk(sk
);
2019 struct sk_buff
*skb
= to
, *tmp
;
2022 if (!sysctl_tcp_retrans_collapse
)
2024 if (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_SYN
)
2027 tcp_for_write_queue_from_safe(skb
, tmp
, sk
) {
2028 if (!tcp_can_collapse(sk
, skb
))
2040 /* Punt if not enough space exists in the first SKB for
2041 * the data in the second
2043 if (skb
->len
> skb_tailroom(to
))
2046 if (after(TCP_SKB_CB(skb
)->end_seq
, tcp_wnd_end(tp
)))
2049 tcp_collapse_retrans(sk
, to
);
2053 /* This retransmits one SKB. Policy decisions and retransmit queue
2054 * state updates are done by the caller. Returns non-zero if an
2055 * error occurred which prevented the send.
2057 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
2059 struct tcp_sock
*tp
= tcp_sk(sk
);
2060 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2061 unsigned int cur_mss
;
2064 /* Inconslusive MTU probe */
2065 if (icsk
->icsk_mtup
.probe_size
) {
2066 icsk
->icsk_mtup
.probe_size
= 0;
2069 /* Do not sent more than we queued. 1/4 is reserved for possible
2070 * copying overhead: fragmentation, tunneling, mangling etc.
2072 if (atomic_read(&sk
->sk_wmem_alloc
) >
2073 min(sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2), sk
->sk_sndbuf
))
2076 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
2077 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
2079 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
2083 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
2084 return -EHOSTUNREACH
; /* Routing failure or similar. */
2086 cur_mss
= tcp_current_mss(sk
);
2088 /* If receiver has shrunk his window, and skb is out of
2089 * new window, do not retransmit it. The exception is the
2090 * case, when window is shrunk to zero. In this case
2091 * our retransmit serves as a zero window probe.
2093 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
)) &&
2094 TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
2097 if (skb
->len
> cur_mss
) {
2098 if (tcp_fragment(sk
, skb
, cur_mss
, cur_mss
))
2099 return -ENOMEM
; /* We'll try again later. */
2101 int oldpcount
= tcp_skb_pcount(skb
);
2103 if (unlikely(oldpcount
> 1)) {
2104 tcp_init_tso_segs(sk
, skb
, cur_mss
);
2105 tcp_adjust_pcount(sk
, skb
, oldpcount
- tcp_skb_pcount(skb
));
2109 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
2111 /* Some Solaris stacks overoptimize and ignore the FIN on a
2112 * retransmit when old data is attached. So strip it off
2113 * since it is cheap to do so and saves bytes on the network.
2116 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
) &&
2117 tp
->snd_una
== (TCP_SKB_CB(skb
)->end_seq
- 1)) {
2118 if (!pskb_trim(skb
, 0)) {
2119 /* Reuse, even though it does some unnecessary work */
2120 tcp_init_nondata_skb(skb
, TCP_SKB_CB(skb
)->end_seq
- 1,
2121 TCP_SKB_CB(skb
)->flags
);
2122 skb
->ip_summed
= CHECKSUM_NONE
;
2126 /* Make a copy, if the first transmission SKB clone we made
2127 * is still in somebody's hands, else make a clone.
2129 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2131 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2134 /* Update global TCP statistics. */
2135 TCP_INC_STATS(sock_net(sk
), TCP_MIB_RETRANSSEGS
);
2137 tp
->total_retrans
++;
2139 #if FASTRETRANS_DEBUG > 0
2140 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2141 if (net_ratelimit())
2142 printk(KERN_DEBUG
"retrans_out leaked.\n");
2145 if (!tp
->retrans_out
)
2146 tp
->lost_retrans_low
= tp
->snd_nxt
;
2147 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
2148 tp
->retrans_out
+= tcp_skb_pcount(skb
);
2150 /* Save stamp of the first retransmit. */
2151 if (!tp
->retrans_stamp
)
2152 tp
->retrans_stamp
= TCP_SKB_CB(skb
)->when
;
2156 /* snd_nxt is stored to detect loss of retransmitted segment,
2157 * see tcp_input.c tcp_sacktag_write_queue().
2159 TCP_SKB_CB(skb
)->ack_seq
= tp
->snd_nxt
;
2164 /* Check if we forward retransmits are possible in the current
2165 * window/congestion state.
2167 static int tcp_can_forward_retransmit(struct sock
*sk
)
2169 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2170 struct tcp_sock
*tp
= tcp_sk(sk
);
2172 /* Forward retransmissions are possible only during Recovery. */
2173 if (icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
2176 /* No forward retransmissions in Reno are possible. */
2177 if (tcp_is_reno(tp
))
2180 /* Yeah, we have to make difficult choice between forward transmission
2181 * and retransmission... Both ways have their merits...
2183 * For now we do not retransmit anything, while we have some new
2184 * segments to send. In the other cases, follow rule 3 for
2185 * NextSeg() specified in RFC3517.
2188 if (tcp_may_send_now(sk
))
2194 /* This gets called after a retransmit timeout, and the initially
2195 * retransmitted data is acknowledged. It tries to continue
2196 * resending the rest of the retransmit queue, until either
2197 * we've sent it all or the congestion window limit is reached.
2198 * If doing SACK, the first ACK which comes back for a timeout
2199 * based retransmit packet might feed us FACK information again.
2200 * If so, we use it to avoid unnecessarily retransmissions.
2202 void tcp_xmit_retransmit_queue(struct sock
*sk
)
2204 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2205 struct tcp_sock
*tp
= tcp_sk(sk
);
2206 struct sk_buff
*skb
;
2207 struct sk_buff
*hole
= NULL
;
2210 int fwd_rexmitting
= 0;
2213 tp
->retransmit_high
= tp
->snd_una
;
2215 if (tp
->retransmit_skb_hint
) {
2216 skb
= tp
->retransmit_skb_hint
;
2217 last_lost
= TCP_SKB_CB(skb
)->end_seq
;
2218 if (after(last_lost
, tp
->retransmit_high
))
2219 last_lost
= tp
->retransmit_high
;
2221 skb
= tcp_write_queue_head(sk
);
2222 last_lost
= tp
->snd_una
;
2225 tcp_for_write_queue_from(skb
, sk
) {
2226 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
2228 if (skb
== tcp_send_head(sk
))
2230 /* we could do better than to assign each time */
2232 tp
->retransmit_skb_hint
= skb
;
2234 /* Assume this retransmit will generate
2235 * only one packet for congestion window
2236 * calculation purposes. This works because
2237 * tcp_retransmit_skb() will chop up the
2238 * packet to be MSS sized and all the
2239 * packet counting works out.
2241 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
2244 if (fwd_rexmitting
) {
2246 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_highest_sack_seq(tp
)))
2248 mib_idx
= LINUX_MIB_TCPFORWARDRETRANS
;
2250 } else if (!before(TCP_SKB_CB(skb
)->seq
, tp
->retransmit_high
)) {
2251 tp
->retransmit_high
= last_lost
;
2252 if (!tcp_can_forward_retransmit(sk
))
2254 /* Backtrack if necessary to non-L'ed skb */
2262 } else if (!(sacked
& TCPCB_LOST
)) {
2263 if (hole
== NULL
&& !(sacked
& (TCPCB_SACKED_RETRANS
|TCPCB_SACKED_ACKED
)))
2268 last_lost
= TCP_SKB_CB(skb
)->end_seq
;
2269 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
2270 mib_idx
= LINUX_MIB_TCPFASTRETRANS
;
2272 mib_idx
= LINUX_MIB_TCPSLOWSTARTRETRANS
;
2275 if (sacked
& (TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))
2278 if (tcp_retransmit_skb(sk
, skb
))
2280 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2282 if (skb
== tcp_write_queue_head(sk
))
2283 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2284 inet_csk(sk
)->icsk_rto
,
2289 /* Send a fin. The caller locks the socket for us. This cannot be
2290 * allowed to fail queueing a FIN frame under any circumstances.
2292 void tcp_send_fin(struct sock
*sk
)
2294 struct tcp_sock
*tp
= tcp_sk(sk
);
2295 struct sk_buff
*skb
= tcp_write_queue_tail(sk
);
2298 /* Optimization, tack on the FIN if we have a queue of
2299 * unsent frames. But be careful about outgoing SACKS
2302 mss_now
= tcp_current_mss(sk
);
2304 if (tcp_send_head(sk
) != NULL
) {
2305 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_FIN
;
2306 TCP_SKB_CB(skb
)->end_seq
++;
2309 /* Socket is locked, keep trying until memory is available. */
2311 skb
= alloc_skb_fclone(MAX_TCP_HEADER
,
2318 /* Reserve space for headers and prepare control bits. */
2319 skb_reserve(skb
, MAX_TCP_HEADER
);
2320 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2321 tcp_init_nondata_skb(skb
, tp
->write_seq
,
2322 TCPCB_FLAG_ACK
| TCPCB_FLAG_FIN
);
2323 tcp_queue_skb(sk
, skb
);
2325 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_OFF
);
2328 /* We get here when a process closes a file descriptor (either due to
2329 * an explicit close() or as a byproduct of exit()'ing) and there
2330 * was unread data in the receive queue. This behavior is recommended
2331 * by RFC 2525, section 2.17. -DaveM
2333 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
2335 struct sk_buff
*skb
;
2337 /* NOTE: No TCP options attached and we never retransmit this. */
2338 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
2340 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
2344 /* Reserve space for headers and prepare control bits. */
2345 skb_reserve(skb
, MAX_TCP_HEADER
);
2346 tcp_init_nondata_skb(skb
, tcp_acceptable_seq(sk
),
2347 TCPCB_FLAG_ACK
| TCPCB_FLAG_RST
);
2349 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2350 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
2351 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
2353 TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTRSTS
);
2356 /* Send a crossed SYN-ACK during socket establishment.
2357 * WARNING: This routine must only be called when we have already sent
2358 * a SYN packet that crossed the incoming SYN that caused this routine
2359 * to get called. If this assumption fails then the initial rcv_wnd
2360 * and rcv_wscale values will not be correct.
2362 int tcp_send_synack(struct sock
*sk
)
2364 struct sk_buff
*skb
;
2366 skb
= tcp_write_queue_head(sk
);
2367 if (skb
== NULL
|| !(TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_SYN
)) {
2368 printk(KERN_DEBUG
"tcp_send_synack: wrong queue state\n");
2371 if (!(TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_ACK
)) {
2372 if (skb_cloned(skb
)) {
2373 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
2376 tcp_unlink_write_queue(skb
, sk
);
2377 skb_header_release(nskb
);
2378 __tcp_add_write_queue_head(sk
, nskb
);
2379 sk_wmem_free_skb(sk
, skb
);
2380 sk
->sk_wmem_queued
+= nskb
->truesize
;
2381 sk_mem_charge(sk
, nskb
->truesize
);
2385 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_ACK
;
2386 TCP_ECN_send_synack(tcp_sk(sk
), skb
);
2388 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2389 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2392 /* Prepare a SYN-ACK. */
2393 struct sk_buff
*tcp_make_synack(struct sock
*sk
, struct dst_entry
*dst
,
2394 struct request_sock
*req
,
2395 struct request_values
*rvp
)
2397 struct tcp_out_options opts
;
2398 struct tcp_extend_values
*xvp
= tcp_xv(rvp
);
2399 struct inet_request_sock
*ireq
= inet_rsk(req
);
2400 struct tcp_sock
*tp
= tcp_sk(sk
);
2401 const struct tcp_cookie_values
*cvp
= tp
->cookie_values
;
2403 struct sk_buff
*skb
;
2404 struct tcp_md5sig_key
*md5
;
2405 int tcp_header_size
;
2407 int s_data_desired
= 0;
2409 if (cvp
!= NULL
&& cvp
->s_data_constant
&& cvp
->s_data_desired
)
2410 s_data_desired
= cvp
->s_data_desired
;
2411 skb
= sock_wmalloc(sk
, MAX_TCP_HEADER
+ 15 + s_data_desired
, 1, GFP_ATOMIC
);
2415 /* Reserve space for headers. */
2416 skb_reserve(skb
, MAX_TCP_HEADER
);
2418 skb_dst_set(skb
, dst_clone(dst
));
2420 mss
= dst_metric(dst
, RTAX_ADVMSS
);
2421 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< mss
)
2422 mss
= tp
->rx_opt
.user_mss
;
2424 if (req
->rcv_wnd
== 0) { /* ignored for retransmitted syns */
2426 /* Set this up on the first call only */
2427 req
->window_clamp
= tp
->window_clamp
? : dst_metric(dst
, RTAX_WINDOW
);
2428 /* tcp_full_space because it is guaranteed to be the first packet */
2429 tcp_select_initial_window(tcp_full_space(sk
),
2430 mss
- (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0),
2435 dst_metric(dst
, RTAX_INITRWND
));
2436 ireq
->rcv_wscale
= rcv_wscale
;
2439 memset(&opts
, 0, sizeof(opts
));
2440 #ifdef CONFIG_SYN_COOKIES
2441 if (unlikely(req
->cookie_ts
))
2442 TCP_SKB_CB(skb
)->when
= cookie_init_timestamp(req
);
2445 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2446 tcp_header_size
= tcp_synack_options(sk
, req
, mss
,
2447 skb
, &opts
, &md5
, xvp
)
2450 skb_push(skb
, tcp_header_size
);
2451 skb_reset_transport_header(skb
);
2454 memset(th
, 0, sizeof(struct tcphdr
));
2457 TCP_ECN_make_synack(req
, th
);
2458 th
->source
= ireq
->loc_port
;
2459 th
->dest
= ireq
->rmt_port
;
2460 /* Setting of flags are superfluous here for callers (and ECE is
2461 * not even correctly set)
2463 tcp_init_nondata_skb(skb
, tcp_rsk(req
)->snt_isn
,
2464 TCPCB_FLAG_SYN
| TCPCB_FLAG_ACK
);
2466 if (OPTION_COOKIE_EXTENSION
& opts
.options
) {
2467 if (s_data_desired
) {
2468 u8
*buf
= skb_put(skb
, s_data_desired
);
2470 /* copy data directly from the listening socket. */
2471 memcpy(buf
, cvp
->s_data_payload
, s_data_desired
);
2472 TCP_SKB_CB(skb
)->end_seq
+= s_data_desired
;
2475 if (opts
.hash_size
> 0) {
2476 __u32 workspace
[SHA_WORKSPACE_WORDS
];
2477 u32
*mess
= &xvp
->cookie_bakery
[COOKIE_DIGEST_WORDS
];
2478 u32
*tail
= &mess
[COOKIE_MESSAGE_WORDS
-1];
2480 /* Secret recipe depends on the Timestamp, (future)
2481 * Sequence and Acknowledgment Numbers, Initiator
2482 * Cookie, and others handled by IP variant caller.
2484 *tail
-- ^= opts
.tsval
;
2485 *tail
-- ^= tcp_rsk(req
)->rcv_isn
+ 1;
2486 *tail
-- ^= TCP_SKB_CB(skb
)->seq
+ 1;
2489 *tail
-- ^= (((__force u32
)th
->dest
<< 16) | (__force u32
)th
->source
);
2490 *tail
-- ^= (u32
)(unsigned long)cvp
; /* per sockopt */
2492 sha_transform((__u32
*)&xvp
->cookie_bakery
[0],
2495 opts
.hash_location
=
2496 (__u8
*)&xvp
->cookie_bakery
[0];
2500 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
2501 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_isn
+ 1);
2503 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2504 th
->window
= htons(min(req
->rcv_wnd
, 65535U));
2505 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
);
2506 th
->doff
= (tcp_header_size
>> 2);
2507 TCP_ADD_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
, tcp_skb_pcount(skb
));
2509 #ifdef CONFIG_TCP_MD5SIG
2510 /* Okay, we have all we need - do the md5 hash if needed */
2512 tcp_rsk(req
)->af_specific
->calc_md5_hash(opts
.hash_location
,
2513 md5
, NULL
, req
, skb
);
2520 /* Do all connect socket setups that can be done AF independent. */
2521 static void tcp_connect_init(struct sock
*sk
)
2523 struct dst_entry
*dst
= __sk_dst_get(sk
);
2524 struct tcp_sock
*tp
= tcp_sk(sk
);
2527 /* We'll fix this up when we get a response from the other end.
2528 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2530 tp
->tcp_header_len
= sizeof(struct tcphdr
) +
2531 (sysctl_tcp_timestamps
? TCPOLEN_TSTAMP_ALIGNED
: 0);
2533 #ifdef CONFIG_TCP_MD5SIG
2534 if (tp
->af_specific
->md5_lookup(sk
, sk
) != NULL
)
2535 tp
->tcp_header_len
+= TCPOLEN_MD5SIG_ALIGNED
;
2538 /* If user gave his TCP_MAXSEG, record it to clamp */
2539 if (tp
->rx_opt
.user_mss
)
2540 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
2543 tcp_sync_mss(sk
, dst_mtu(dst
));
2545 if (!tp
->window_clamp
)
2546 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
2547 tp
->advmss
= dst_metric(dst
, RTAX_ADVMSS
);
2548 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< tp
->advmss
)
2549 tp
->advmss
= tp
->rx_opt
.user_mss
;
2551 tcp_initialize_rcv_mss(sk
);
2553 tcp_select_initial_window(tcp_full_space(sk
),
2554 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
2557 sysctl_tcp_window_scaling
,
2559 dst_metric(dst
, RTAX_INITRWND
));
2561 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
2562 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
2565 sock_reset_flag(sk
, SOCK_DONE
);
2568 tp
->snd_una
= tp
->write_seq
;
2569 tp
->snd_sml
= tp
->write_seq
;
2570 tp
->snd_up
= tp
->write_seq
;
2575 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
2576 inet_csk(sk
)->icsk_retransmits
= 0;
2577 tcp_clear_retrans(tp
);
2580 /* Build a SYN and send it off. */
2581 int tcp_connect(struct sock
*sk
)
2583 struct tcp_sock
*tp
= tcp_sk(sk
);
2584 struct sk_buff
*buff
;
2586 tcp_connect_init(sk
);
2588 buff
= alloc_skb_fclone(MAX_TCP_HEADER
+ 15, sk
->sk_allocation
);
2589 if (unlikely(buff
== NULL
))
2592 /* Reserve space for headers. */
2593 skb_reserve(buff
, MAX_TCP_HEADER
);
2595 tp
->snd_nxt
= tp
->write_seq
;
2596 tcp_init_nondata_skb(buff
, tp
->write_seq
++, TCPCB_FLAG_SYN
);
2597 TCP_ECN_send_syn(sk
, buff
);
2600 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2601 tp
->retrans_stamp
= TCP_SKB_CB(buff
)->when
;
2602 skb_header_release(buff
);
2603 __tcp_add_write_queue_tail(sk
, buff
);
2604 sk
->sk_wmem_queued
+= buff
->truesize
;
2605 sk_mem_charge(sk
, buff
->truesize
);
2606 tp
->packets_out
+= tcp_skb_pcount(buff
);
2607 tcp_transmit_skb(sk
, buff
, 1, sk
->sk_allocation
);
2609 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2610 * in order to make this packet get counted in tcpOutSegs.
2612 tp
->snd_nxt
= tp
->write_seq
;
2613 tp
->pushed_seq
= tp
->write_seq
;
2614 TCP_INC_STATS(sock_net(sk
), TCP_MIB_ACTIVEOPENS
);
2616 /* Timer for repeating the SYN until an answer. */
2617 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2618 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2622 /* Send out a delayed ack, the caller does the policy checking
2623 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2626 void tcp_send_delayed_ack(struct sock
*sk
)
2628 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2629 int ato
= icsk
->icsk_ack
.ato
;
2630 unsigned long timeout
;
2632 if (ato
> TCP_DELACK_MIN
) {
2633 const struct tcp_sock
*tp
= tcp_sk(sk
);
2634 int max_ato
= HZ
/ 2;
2636 if (icsk
->icsk_ack
.pingpong
||
2637 (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
2638 max_ato
= TCP_DELACK_MAX
;
2640 /* Slow path, intersegment interval is "high". */
2642 /* If some rtt estimate is known, use it to bound delayed ack.
2643 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2647 int rtt
= max(tp
->srtt
>> 3, TCP_DELACK_MIN
);
2653 ato
= min(ato
, max_ato
);
2656 /* Stay within the limit we were given */
2657 timeout
= jiffies
+ ato
;
2659 /* Use new timeout only if there wasn't a older one earlier. */
2660 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
2661 /* If delack timer was blocked or is about to expire,
2664 if (icsk
->icsk_ack
.blocked
||
2665 time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
2670 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
2671 timeout
= icsk
->icsk_ack
.timeout
;
2673 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
2674 icsk
->icsk_ack
.timeout
= timeout
;
2675 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
2678 /* This routine sends an ack and also updates the window. */
2679 void tcp_send_ack(struct sock
*sk
)
2681 struct sk_buff
*buff
;
2683 /* If we have been reset, we may not send again. */
2684 if (sk
->sk_state
== TCP_CLOSE
)
2687 /* We are not putting this on the write queue, so
2688 * tcp_transmit_skb() will set the ownership to this
2691 buff
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2693 inet_csk_schedule_ack(sk
);
2694 inet_csk(sk
)->icsk_ack
.ato
= TCP_ATO_MIN
;
2695 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
2696 TCP_DELACK_MAX
, TCP_RTO_MAX
);
2700 /* Reserve space for headers and prepare control bits. */
2701 skb_reserve(buff
, MAX_TCP_HEADER
);
2702 tcp_init_nondata_skb(buff
, tcp_acceptable_seq(sk
), TCPCB_FLAG_ACK
);
2704 /* Send it off, this clears delayed acks for us. */
2705 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2706 tcp_transmit_skb(sk
, buff
, 0, GFP_ATOMIC
);
2709 /* This routine sends a packet with an out of date sequence
2710 * number. It assumes the other end will try to ack it.
2712 * Question: what should we make while urgent mode?
2713 * 4.4BSD forces sending single byte of data. We cannot send
2714 * out of window data, because we have SND.NXT==SND.MAX...
2716 * Current solution: to send TWO zero-length segments in urgent mode:
2717 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2718 * out-of-date with SND.UNA-1 to probe window.
2720 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
)
2722 struct tcp_sock
*tp
= tcp_sk(sk
);
2723 struct sk_buff
*skb
;
2725 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2726 skb
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2730 /* Reserve space for headers and set control bits. */
2731 skb_reserve(skb
, MAX_TCP_HEADER
);
2732 /* Use a previous sequence. This should cause the other
2733 * end to send an ack. Don't queue or clone SKB, just
2736 tcp_init_nondata_skb(skb
, tp
->snd_una
- !urgent
, TCPCB_FLAG_ACK
);
2737 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2738 return tcp_transmit_skb(sk
, skb
, 0, GFP_ATOMIC
);
2741 /* Initiate keepalive or window probe from timer. */
2742 int tcp_write_wakeup(struct sock
*sk
)
2744 struct tcp_sock
*tp
= tcp_sk(sk
);
2745 struct sk_buff
*skb
;
2747 if (sk
->sk_state
== TCP_CLOSE
)
2750 if ((skb
= tcp_send_head(sk
)) != NULL
&&
2751 before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
))) {
2753 unsigned int mss
= tcp_current_mss(sk
);
2754 unsigned int seg_size
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
2756 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
2757 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
2759 /* We are probing the opening of a window
2760 * but the window size is != 0
2761 * must have been a result SWS avoidance ( sender )
2763 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
2765 seg_size
= min(seg_size
, mss
);
2766 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2767 if (tcp_fragment(sk
, skb
, seg_size
, mss
))
2769 } else if (!tcp_skb_pcount(skb
))
2770 tcp_set_skb_tso_segs(sk
, skb
, mss
);
2772 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2773 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2774 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2776 tcp_event_new_data_sent(sk
, skb
);
2779 if (between(tp
->snd_up
, tp
->snd_una
+ 1, tp
->snd_una
+ 0xFFFF))
2780 tcp_xmit_probe_skb(sk
, 1);
2781 return tcp_xmit_probe_skb(sk
, 0);
2785 /* A window probe timeout has occurred. If window is not closed send
2786 * a partial packet else a zero probe.
2788 void tcp_send_probe0(struct sock
*sk
)
2790 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2791 struct tcp_sock
*tp
= tcp_sk(sk
);
2794 err
= tcp_write_wakeup(sk
);
2796 if (tp
->packets_out
|| !tcp_send_head(sk
)) {
2797 /* Cancel probe timer, if it is not required. */
2798 icsk
->icsk_probes_out
= 0;
2799 icsk
->icsk_backoff
= 0;
2804 if (icsk
->icsk_backoff
< sysctl_tcp_retries2
)
2805 icsk
->icsk_backoff
++;
2806 icsk
->icsk_probes_out
++;
2807 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2808 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
2811 /* If packet was not sent due to local congestion,
2812 * do not backoff and do not remember icsk_probes_out.
2813 * Let local senders to fight for local resources.
2815 * Use accumulated backoff yet.
2817 if (!icsk
->icsk_probes_out
)
2818 icsk
->icsk_probes_out
= 1;
2819 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2820 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
,
2821 TCP_RESOURCE_PROBE_INTERVAL
),
2826 EXPORT_SYMBOL(tcp_select_initial_window
);
2827 EXPORT_SYMBOL(tcp_connect
);
2828 EXPORT_SYMBOL(tcp_make_synack
);
2829 EXPORT_SYMBOL(tcp_simple_retransmit
);
2830 EXPORT_SYMBOL(tcp_sync_mss
);
2831 EXPORT_SYMBOL(tcp_mtup_init
);